diff --git a/.gitattributes b/.gitattributes index 82162c364b65749ac6ea4ad7213f9794f6f51c13..5be54c12713972d4233f5cfe068dfad82aa9af46 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1525,3 +1525,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Occluded[[:space:]]Gait[[:space:]]Recognition[[:space:]]with[[:space:]]Mixture[[:space:]]of[[:space:]]Experts_[[:space:]]An[[:space:]]Action[[:space:]]Detection[[:space:]]Perspective/47e82503-1eee-4ea3-a983-f317bb28bfec_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Occlusion[[:space:]]Handling[[:space:]]in[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation[[:space:]]with[[:space:]]Perturbed[[:space:]]Positional[[:space:]]Encoding/6bd87a98-c63f-4d0c-b044-51bccc6fdd33_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Occlusion-Aware[[:space:]]Seamless[[:space:]]Segmentation/eeb35d34-dc12-4092-9b80-16b448e64afe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Occupancy[[:space:]]as[[:space:]]Set[[:space:]]of[[:space:]]Points/e88bbcd3-f64f-4fcb-a277-94c463373078_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Octopus_[[:space:]]Embodied[[:space:]]Vision-Language[[:space:]]Programmer[[:space:]]from[[:space:]]Environmental[[:space:]]Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Omni-Recon_[[:space:]]Harnessing[[:space:]]Image-based[[:space:]]Rendering[[:space:]]for[[:space:]]General-Purpose[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/0d53a505-587b-4461-974a-55e35a489a73_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Omni6DPose_[[:space:]]A[[:space:]]Benchmark[[:space:]]and[[:space:]]Model[[:space:]]for[[:space:]]Universal[[:space:]]6D[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation[[:space:]]and[[:space:]]Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Omni6D_[[:space:]]Large-Vocabulary[[:space:]]3D[[:space:]]Object[[:space:]]Dataset[[:space:]]for[[:space:]]Category-Level[[:space:]]6D[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OmniACT_[[:space:]]A[[:space:]]Dataset[[:space:]]and[[:space:]]Benchmark[[:space:]]for[[:space:]]Enabling[[:space:]]Multimodal[[:space:]]Generalist[[:space:]]Autonomous[[:space:]]Agents[[:space:]]for[[:space:]]Desktop[[:space:]]and[[:space:]]Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OmniNOCS_[[:space:]]A[[:space:]]unified[[:space:]]NOCS[[:space:]]dataset[[:space:]]and[[:space:]]model[[:space:]]for[[:space:]]3D[[:space:]]lifting[[:space:]]of[[:space:]]2D[[:space:]]objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OmniSSR_[[:space:]]Zero-shot[[:space:]]Omnidirectional[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]using[[:space:]]Stable[[:space:]]Diffusion[[:space:]]Model/7452c846-9323-4781-952d-a6062713a9d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OmniSat_[[:space:]]Self-Supervised[[:space:]]Modality[[:space:]]Fusion[[:space:]]for[[:space:]]Earth[[:space:]]Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Omniview-Tuning_[[:space:]]Boosting[[:space:]]Viewpoint[[:space:]]Invariance[[:space:]]of[[:space:]]Vision-Language[[:space:]]Pre-training[[:space:]]Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]Calibration[[:space:]]of[[:space:]]Object[[:space:]]Detectors_[[:space:]]Pitfalls,[[:space:]]Evaluation[[:space:]]and[[:space:]]Baselines/be195471-6153-4286-905c-575a7ab7761b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]Learning[[:space:]]Discriminative[[:space:]]Features[[:space:]]from[[:space:]]Synthesized[[:space:]]Data[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Fine-Grained[[:space:]]Visual[[:space:]]Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]Pretraining[[:space:]]Data[[:space:]]Diversity[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]Spectral[[:space:]]Properties[[:space:]]of[[:space:]]Gradient-based[[:space:]]Explanation[[:space:]]Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]the[[:space:]]Approximation[[:space:]]Risk[[:space:]]of[[:space:]]Few-Shot[[:space:]]Class-Incremental[[:space:]]Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]the[[:space:]]Error[[:space:]]Analysis[[:space:]]of[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]and[[:space:]]an[[:space:]]Optimal[[:space:]]Projection[[:space:]]Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]the[[:space:]]Evaluation[[:space:]]Consistency[[:space:]]of[[:space:]]Attribution-based[[:space:]]Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]the[[:space:]]Topology[[:space:]]Awareness[[:space:]]and[[:space:]]Generalization[[:space:]]Performance[[:space:]]of[[:space:]]Graph[[:space:]]Neural[[:space:]]Networks/a30a461c-3f3b-47ab-8426-e61337113303_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]the[[:space:]]Utility[[:space:]]of[[:space:]]3D[[:space:]]Hand[[:space:]]Poses[[:space:]]for[[:space:]]Action[[:space:]]Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]the[[:space:]]Viability[[:space:]]of[[:space:]]Monocular[[:space:]]Depth[[:space:]]Pre-training[[:space:]]for[[:space:]]Semantic[[:space:]]Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On[[:space:]]the[[:space:]]Vulnerability[[:space:]]of[[:space:]]Skip[[:space:]]Connections[[:space:]]to[[:space:]]Model[[:space:]]Inversion[[:space:]]Attacks/ef086c62-5a84-4279-b346-33af023d90c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/On-the-fly[[:space:]]Category[[:space:]]Discovery[[:space:]]for[[:space:]]LiDAR[[:space:]]Semantic[[:space:]]Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/One-Shot[[:space:]]Diffusion[[:space:]]Mimicker[[:space:]]for[[:space:]]Handwritten[[:space:]]Text[[:space:]]Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/One-stage[[:space:]]Prompt-based[[:space:]]Continual[[:space:]]Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OneRestore_[[:space:]]A[[:space:]]Universal[[:space:]]Restoration[[:space:]]Framework[[:space:]]for[[:space:]]Composite[[:space:]]Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OneTrack_[[:space:]]Demystifying[[:space:]]the[[:space:]]Conflict[[:space:]]Between[[:space:]]Detection[[:space:]]and[[:space:]]Tracking[[:space:]]in[[:space:]]End-to-End[[:space:]]3D[[:space:]]Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OneVOS_[[:space:]]Unifying[[:space:]]Video[[:space:]]Object[[:space:]]Segmentation[[:space:]]with[[:space:]]All-in-One[[:space:]]Transformer[[:space:]]Framework/746728e7-d87a-494d-8a5a-df2168951889_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Online[[:space:]]Continuous[[:space:]]Generalized[[:space:]]Category[[:space:]]Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Online[[:space:]]Temporal[[:space:]]Action[[:space:]]Localization[[:space:]]with[[:space:]]Memory-Augmented[[:space:]]Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Online[[:space:]]Vectorized[[:space:]]HD[[:space:]]Map[[:space:]]Construction[[:space:]]using[[:space:]]Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Online[[:space:]]Video[[:space:]]Quality[[:space:]]Enhancement[[:space:]]with[[:space:]]Spatial-Temporal[[:space:]]Look-up[[:space:]]Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Online[[:space:]]Zero-Shot[[:space:]]Classification[[:space:]]with[[:space:]]CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open[[:space:]]Panoramic[[:space:]]Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open[[:space:]]Vocabulary[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding[[:space:]]via[[:space:]]Geometry[[:space:]]Guided[[:space:]]Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open[[:space:]]Vocabulary[[:space:]]Multi-Label[[:space:]]Video[[:space:]]Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-Set[[:space:]]Biometrics_[[:space:]]Beyond[[:space:]]Good[[:space:]]Closed-Set[[:space:]]Models/5d8636e8-0041-40e7-8842-7b710a232989_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-Set[[:space:]]Recognition[[:space:]]in[[:space:]]the[[:space:]]Age[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-Vocabulary[[:space:]]3D[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]with[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-Vocabulary[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-Vocabulary[[:space:]]RGB-Thermal[[:space:]]Semantic[[:space:]]Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-Vocabulary[[:space:]]SAM_[[:space:]]Segment[[:space:]]and[[:space:]]Recognize[[:space:]]Twenty-thousand[[:space:]]Classes[[:space:]]Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-World[[:space:]]Dynamic[[:space:]]Prompt[[:space:]]and[[:space:]]Continual[[:space:]]Visual[[:space:]]Representation[[:space:]]Learning/5cd14218-379f-4209-a524-2be87a9eba2f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Open-set[[:space:]]Domain[[:space:]]Adaptation[[:space:]]via[[:space:]]Joint[[:space:]]Error[[:space:]]based[[:space:]]Multi-class[[:space:]]Positive[[:space:]]and[[:space:]]Unlabeled[[:space:]]Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OpenIns3D_[[:space:]]Snap[[:space:]]and[[:space:]]Lookup[[:space:]]for[[:space:]]3D[[:space:]]Open-vocabulary[[:space:]]Instance[[:space:]]Segmentation/f206041e-1c46-4539-919a-ce8570607515_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OpenKD_[[:space:]]Opening[[:space:]]Prompt[[:space:]]Diversity[[:space:]]for[[:space:]]Zero-[[:space:]]and[[:space:]]Few-shot[[:space:]]Keypoint[[:space:]]Detection/dd93a63c-860b-4bc5-8581-032ed835a163_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OpenPSG_[[:space:]]Open-set[[:space:]]Panoptic[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]via[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OpenSight_[[:space:]]A[[:space:]]Simple[[:space:]]Open-Vocabulary[[:space:]]Framework[[:space:]]for[[:space:]]LiDAR-Based[[:space:]]Object[[:space:]]Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Operational[[:space:]]Open-Set[[:space:]]Recognition[[:space:]]and[[:space:]]PostMax[[:space:]]Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OphNet_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Video[[:space:]]Benchmark[[:space:]]for[[:space:]]Ophthalmic[[:space:]]Surgical[[:space:]]Workflow[[:space:]]Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Optimal[[:space:]]Transport[[:space:]]of[[:space:]]Diverse[[:space:]]Unsupervised[[:space:]]Tasks[[:space:]]for[[:space:]]Robust[[:space:]]Learning[[:space:]]from[[:space:]]Noisy[[:space:]]Few-Shot[[:space:]]Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Optimization-based[[:space:]]Uncertainty[[:space:]]Attribution[[:space:]]Via[[:space:]]Learning[[:space:]]Informative[[:space:]]Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Optimizing[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Joint[[:space:]]Trajectory[[:space:]]Prediction[[:space:]]and[[:space:]]Controllable[[:space:]]Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Optimizing[[:space:]]Factorized[[:space:]]Encoder[[:space:]]Models_[[:space:]]Time[[:space:]]and[[:space:]]Memory[[:space:]]Reduction[[:space:]]for[[:space:]]Scalable[[:space:]]and[[:space:]]Efficient[[:space:]]Action[[:space:]]Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Optimizing[[:space:]]Illuminant[[:space:]]Estimation[[:space:]]in[[:space:]]Dual-Exposure[[:space:]]HDR[[:space:]]Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Osmosis_[[:space:]]RGBD[[:space:]]Diffusion[[:space:]]Prior[[:space:]]for[[:space:]]Underwater[[:space:]]Image[[:space:]]Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Oulu[[:space:]]Remote-photoplethysmography[[:space:]]Physical[[:space:]]Domain[[:space:]]Attacks[[:space:]]Database[[:space:]](ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Out-of-Bounding-Box[[:space:]]Triggers_[[:space:]]A[[:space:]]Stealthy[[:space:]]Approach[[:space:]]to[[:space:]]Cheat[[:space:]]Object[[:space:]]Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/OvSW_[[:space:]]Overcoming[[:space:]]Silent[[:space:]]Weights[[:space:]]for[[:space:]]Accurate[[:space:]]Binary[[:space:]]Neural[[:space:]]Networks/43eb14f0-154a-457a-9941-4861ded25366_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Overcome[[:space:]]Modal[[:space:]]Bias[[:space:]]in[[:space:]]Multi-modal[[:space:]]Federated[[:space:]]Learning[[:space:]]via[[:space:]]Balanced[[:space:]]Modality[[:space:]]Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Overcoming[[:space:]]Distribution[[:space:]]Mismatch[[:space:]]in[[:space:]]Quantizing[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PACE_[[:space:]]Pose[[:space:]]Annotations[[:space:]]in[[:space:]]Cluttered[[:space:]]Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PALM_[[:space:]]Predicting[[:space:]]Actions[[:space:]]through[[:space:]]Language[[:space:]]Models/26a4ce29-0668-4f26-8004-06463aacfbd9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PARE-Net_[[:space:]]Position-Aware[[:space:]]Rotation-Equivariant[[:space:]]Networks[[:space:]]for[[:space:]]Robust[[:space:]]Point[[:space:]]Cloud[[:space:]]Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PARIS3D_[[:space:]]Reasoning-based[[:space:]]3D[[:space:]]Part[[:space:]]Segmentation[[:space:]]Using[[:space:]]Large[[:space:]]Multimodal[[:space:]]Model/ee219343-d433-449b-a58a-eb679a19a121_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_content_list.json b/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c6e9f93ac6ac7a0dc90a58c721cff7914cbf10d7 --- /dev/null +++ b/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_content_list.json @@ -0,0 +1,1595 @@ +[ + { + "type": "text", + "text": "Occupancy as Set of Points", + "text_level": 1, + "bbox": [ + 346, + 141, + 656, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiang Shi $^{1,\\star}$ , Tianheng Cheng $^{1,\\star}$ , Qian Zhang $^{2}$ , Wenyu Liu $^{1}$ , and Xinggang Wang $^{1,\\text{忍}}$", + "bbox": [ + 328, + 188, + 673, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of EIC, Huazhong University of Science & Technology", + "bbox": [ + 285, + 231, + 710, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Horizon Robotics", + "bbox": [ + 434, + 246, + 566, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we explore a novel point representation for 3D occupancy prediction from multi-view images, which is named Occupancy as Set of Points. Existing camera-based methods tend to exploit dense volume-based representation to predict the occupancy of the whole scene, making it hard to focus on the special areas or areas out of the perception range. In comparison, we present the Points of Interest (PoIs) to represent the scene and propose OSP, a novel framework for point-based 3D occupancy prediction. Owing to the inherent flexibility of the point-based representation, OSP achieves strong performance compared with existing methods and excels in terms of training and inference adaptability. It extends beyond traditional perception boundaries and can be seamlessly integrated with volume-based methods to significantly enhance their effectiveness. Experiments on the Occ3D-nuScenes occupancy benchmark show that OSP has strong performance and flexibility. Code and models are available at https://github.com/hustvl/osp.", + "bbox": [ + 261, + 306, + 743, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Occupancy Prediction $\\cdot$ Autonomous Vehicles $\\cdot$ Multi-view 3D Perception", + "bbox": [ + 259, + 526, + 740, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 578, + 375, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Holistic 3D scene understanding is crucial for autonomous driving systems, directly affecting the efficiency and accuracy of subsequent tasks. Considering the cost-effectiveness and ease of deployment of cameras compared to other sensors, developing visual-based methods for 3D scene understanding has become a significant and widely researched challenge.", + "bbox": [ + 212, + 607, + 785, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this challenge, 3D Semantic Scene Completion (SSC) [22] has been proposed and widely studied to jointly infer the geometry and semantics information of the scene from limited observations. The SSC task requires the model to accurately predict the visible locations and complete the information for the invisible locations. Recently, Occ3D [25] introduces a new task definition called 3D occupancy prediction. The main difference between this task and SSC is that 3D occupancy prediction only focuses on the visible areas and is tailored for dynamic scenes.", + "bbox": [ + 212, + 683, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author: Xinggang Wang (xgwang@hust.edu.cn).", + "bbox": [ + 230, + 811, + 782, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2db89918190d37d13c18068154ca7442a3a75eef145310947f2ffd263686312.jpg", + "image_caption": [ + "Fig. 1: Comparison between volume-based methods and our method. The volume-based methods, represented by BEVFormer, infers every region within the scene and gets standard occupancy as shown in (a). Our method uses a point-based decoder as shown in (b). Thus it infers the Points of Interest including standard, adaptively sampled, and manually sampled grids as shown in (c)." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 570, + 361 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cecde533082c1cb6a3ea0177789a83d58650a9cc44368b218888afd5ac117637.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 148, + 779, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing 3D occupancy prediction methods are mostly based on dense BEV methods, e.g., BEVFormer [15], BEVDet [7]. These methods integrate a BEV encoder with an occupancy head to generate the output and enhance BEV perception capabilities for better results. However, they share some common drawbacks. (1) Uniform Sampling: BEV-based methods fail to differentiate between different areas within the same scene, treating them equally. This leads to coarse sampling and hinders dynamic or multi-resolution sampling capabilities. (2) Limited Inference Flexibility: During inference, these methods can only process the entire scene at once. They lack the ability to infer different parts of the scene based on varying downstream tasks or specific practical needs.", + "bbox": [ + 212, + 486, + 787, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Those limitations highlight the need for more flexible 3D occupancy prediction methods that can handle complex scenes while adapting to different inference requirements. In this paper, we propose a novel point-based representation for 3D occupancy prediction. Instead of dividing the scene into uniform grids applied by existing volume-based methods, we propose Points of Interest (PoIs) to view the scene as a collection of points that help in flexibly sampling the scene during both training and inference stages. Fig. 1 compares volume-based and point-based representations. Compared to volume-based representations, our point-based representation has the following advantages: (1) it can accept inputs of any scale and position to make occupancy predictions, including manually designed and adaptively designed input, offering flexibility; (2) it can pay extra attention to certain areas rather than treating all areas equally, enhancing the model's perceptual capabilities.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 271, + 114, + 611, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Occupancy as Set of Points (OSP), a novel and flexible point-based framework, which is built upon the foundational concept of Points of Interest (PoIs). OSP excels in 3D occupancy prediction and is composed of an image backbone, a 3D positioning encoder, and a decoder, as illustrated in Fig. 2. Central to our methodology is the innovative use of PoIs, which we have categorized into three distinct types to meet diverse needs, thereby significantly enhancing various aspects of our model's performance. Detailed descriptions of these PoIs will be provided in Sec. 3. Notably, PoIs can be designed as needed beyond the three types initially proposed by us.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method stands out for its strong performance and flexibility. The flexibility enables it to process any arbitrary local scene without necessitating retraining. Additionally, our method can serve as an augmentative plugin module for existing volume-based methods by adaptively resampling areas of low confidence to yield more accurate occupancy predictions. It is also adept at predicting areas beyond the scene. The key contributions of our approach can be summarized as follows:", + "bbox": [ + 212, + 282, + 787, + 387 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel point-based occupancy representation, established by interacting point queries with 2D image features, enables a comprehensive understanding of 3D scenes.", + "- A flexible framework that allows for inference at any area of interest without retraining or sacrificing accuracy and predicts areas beyond the scene.", + "- A plugin module that enhances the performance of the volume-based baseline significantly.", + "- OSP has obtained strong experiment results, i.e., 39.4 mIoU in the 3D occupancy prediction task on the Occ3D-nuScenes benchmark." + ], + "bbox": [ + 223, + 395, + 782, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 549, + 387, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 214, + 578, + 473, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D occupancy prediction, a concept recently defined by Occ3D [25], exhibits notable parallels with Occupancy Grid Mapping (OGM) [20, 24, 27] used in robotics. This task aims to predict the state of each voxel grid in a scene based on a series of sensor inputs. Occ3D establishes two benchmarks leveraging the Waymo Open Dataset [23] and the nuScenes Dataset [2] to facilitate this. In vision-based 3D Occupancy prediction, Occ3D implements camera visibility estimation and creates visibility masks to ensure evaluations are confined to visible areas. It also evaluates various SSC methodologies on its benchmarks, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32], and BEV-Former [15].", + "bbox": [ + 212, + 601, + 787, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 3D Semantic Scene Completion", + "text_level": 1, + "bbox": [ + 214, + 772, + 521, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Semantic Completion (SSC) represents a task closely associated with 3D occupancy prediction. The concept of SSC is initially presented in SSC-Net [22], with a focus on predicting the comprehensive semantic information", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of a scene based on its partially visible regions. Over recent years, the study of SSC has expanded significantly, particularly in the context of small indoor scenes [3,5,11-13,16,30,31]. In recent times, the study of Scene Semantic Completion (SSC) for expansive outdoor environments has gained momentum, particularly following the introduction of the SemanticKITTI dataset [1]. Notably, MonoScene [4] emerges as the first method to apply monocular pure vision-based SSC. In a parallel advancement, OccDepth [19] enhances 2D to 3D feature transformation by incorporating depth data from stereo input. TPVFormer [9] argues against the limitations of single-plane modeling in capturing intricate details, hence it adopts a tri-perspective view (TPV) approach, combining a Bird's Eye View (BEV) with two additional vertical planes. Additionally, Symphonies [10] highlights the significance of instance representation in SSC tasks. While SSC methods can be directly applied to 3D occupancy prediction, two primary distinctions exist: (1) SSC primarily aims to infer the occupancy of non-visible areas based on visible regions, in contrast to 3D occupancy prediction which focuses on visible areas; (2) SSC methods usually target static scenes, whereas 3D occupancy prediction methods are often designed to handle dynamic scenes.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most existing volume-based SSC and 3D occupancy prediction methods are characterized by their dense nature, encompassing inputs and outputs that span the entire scene. Consider the BEVFormer baseline as an example: it segments the scene into uniform BEV grids, failing to distinguish between grids in varying areas. This uniformity restricts the ability of volume-based methods like BEVFormer to sample areas of interest for better performance during training. Besides, if we want to focus on a specific area in the inference stage, volume-based methods are limited to infer the entire scene and then perform post-processing, inevitably leading to increased and unnecessary costs. Moreover, as scene size and voxel resolution increase, the computational demands skyrocket exponentially. In stark contrast, our point-based model introduces much-needed flexibility by focusing on PoIs. Our method facilitates direct inference in specific areas, eliminating the need for post-processing and avoiding additional computational burdens.", + "bbox": [ + 212, + 404, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A point-related SSC method is PointOcc [34], a point cloud-based SSC prediction method using three complementary view planes for efficient point cloud feature modeling and an efficient 2D backbone for processing to reduce computational load, while our method focuses on the flexibility of training and inference.", + "bbox": [ + 212, + 616, + 787, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Camera-based 3D Detection", + "text_level": 1, + "bbox": [ + 215, + 704, + 495, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera-based 3D perception tasks have gained substantial attention in recent research, largely due to the convenience and cost-effectiveness of cameras as data collection sensors. Initial efforts, such as FCOS3D [26] and DETR3D [28], explore the transition from 2D to 3D predictions. BEVFormer [15] represents a significant advancement in this area, transforming images captured from vehicle-mounted cameras into a bird's eye view (BEV) representation. This technique not only enhances a vehicle's environmental understanding but also finds appli", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 271, + 114, + 611, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cations in various downstream tasks like BEVStereo [14] and BEVDet [7] and extends to 3D occupancy prediction.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D detection tasks are important in camera-based 3D perception and also have a high similarity to 3D occupancy prediction. The primary objective of 3D detection involves estimating the position and dimensions of objects in 3D space. DETR3D [28], drawing inspiration from DETR [33], innovatively combines 3D object queries with image features, incorporating camera intrinsic and extrinsic parameters. PETR [17] and its successor, PETRv2 [18], further refine this approach by addressing the accuracy of reference point sampling and the inclusion of global information, enhancing 3D detection through historical data integration. Compared to object detection, 3D occupancy prediction offers a finer granularity, which is crucial for navigating irregular obstacles or overhanging objects. It is imperative to explore how insights from BEV representation and 3D object detection can inspire innovative solutions tailored to the specific requirements of 3D occupancy prediction.", + "bbox": [ + 212, + 178, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 404, + 367, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem setup. We aim to provide occupancy predictions around the ego-vehicle, given only $N$ surround-view RGB images. More specifically, we use as input current images denoted by $\\mathbf{I}_i = \\{I_0,I_1,\\dots,I_N\\}$ , and use as output an occupancy prediction $\\mathbf{Y}_i\\in \\{c_0,c_1,\\dots,c_M\\}^{H\\times W\\times Z}$ defined in the coordinate of ego-vehicle, where each occupancy prediction is either empty (denoted by $c_{0}$ ) or occupied by a certain semantic class in $\\{c_{1},c_{m},\\dots,c_{M}\\}$ . We assume known camera intrinsic parameters $\\{K_i\\}$ and extrinsic parameters $\\{[R_i|t_i]\\}$ in each frame. We assume to know whether each area is visible or not by applying a camera visibility mask.", + "bbox": [ + 212, + 441, + 787, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Points of Interest. In our model, we innovatively introduce the concept of Points of Interest (PoIs), which is a set of sparse points to represent the 3D scene. PoIs can flexibly represent objects or regions that require extra attention such as pedestrians or regions near ego-vehicles and can be designed as needed in both training and inference phases. We use three types of PoIs and the definitions and functions are introduced as follows:", + "bbox": [ + 212, + 580, + 787, + 669 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Standard Grids: By sampling center points of grids in the inference stage and making predictions of standard 3D occupancy grids, our model makes a fair comparison with existing methods and achieves good results.", + "(2) Adaptively Sampling: During the training stage, our model adaptively samples points and oversamples points around them to enhance accuracy. Recognizing that volume-based methods uniformly treat all locations, our point-based approach allows for resampling in either areas of special interest or those that are challenging to learn. This adaptively resampling strategy is also used to augment the performance of volume-based methods. Consequently, our method can function as a versatile plugin, seamlessly integrating with and enhancing existing volume-based approaches." + ], + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(3) Manually Sampling: Our model excels in its flexibility to sample any area, specifically catering to the unique demands of various downstream tasks. Our model can make predictions of areas beyond the standard perception range by setting PoIs to areas outside the scene manually, e.g. 200 meters away from ego-vehicle, which is a feat unattainable by traditional volume-based methods. This extension not only broadens the scope of inference but also introduces a new dimension to scene understanding. These PoIs are the foundation of our method, offering precision and high flexibility.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/449a3e8aed3cf8a8fcbc855e54bd5977137f27256f6b3f0d14dcfd7ea7ae6ea4.jpg", + "image_caption": [ + "Fig. 2: Overall framework of Occupancy as Set of Points. OSP leverages the Transformer architecture to derive 3D point features from 2D images to make 3D occupancy predictions. Initially, we extract 2D features from multi-view images. Following this, we employ a set of 3D point queries to index these 2D features. The selection of these 3D point queries depends on the Points of Interest (PoIs)." + ], + "image_footnote": [], + "bbox": [ + 225, + 297, + 782, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 672, + 330, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Overall Architecture", + "text_level": 1, + "bbox": [ + 215, + 705, + 431, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 2 shows the overall architecture of the proposed Occupancy as Set of Points. Given the images $\\mathbf{I} = \\{I_i, i = 1, 2, \\dots, N\\}$ from $\\mathbf{N}$ views, We feed these images into ResNet [6] to obtain their features.", + "bbox": [ + 212, + 733, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OSP begins with sampling a set of 3D points in the space, which are the initial PoIs. In our experiments, we use center points of grids as the initial PoIs to make a fair comparison with traditional volume-based methods and provide a reliable baseline for performance comparison and evaluation. We sample $K(K = 8000)$", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 269, + 114, + 611, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "points within the camera's visible region and introduce random perturbations to these points.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then we normalize the 3D points. To these normalized coordinates, we apply sine and cosine functions as a form of positional encoding. This encoded positional information is then utilized to create query position embeddings. In the training phase, within each decoder layer, the query position corresponding to each individual query remains consistent. The 3D coordinates, along with the camera's intrinsic and extrinsic parameters, are used to map these points onto the pixel plane. This mapping process yields corresponding key and value pairs. Subsequently, we employ point cross-attention mechanisms to compute the output. Then we adaptively oversample a group of $M$ points ( $M = 8000$ ) whose coordinates are calculated by a linear layer and employ group point cross-attention to fuse the features of the additional sampling points.", + "bbox": [ + 212, + 176, + 787, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 3D Position Encoder", + "text_level": 1, + "bbox": [ + 215, + 364, + 433, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After obtaining 3D points by applying PoIs, we first normalize coordinate points using equations as follows:", + "bbox": [ + 212, + 391, + 782, + 421 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx, y, z = \\frac {x - x _ {\\operatorname* {m i n}}}{x _ {\\operatorname* {m a x}} - x _ {\\operatorname* {m i n}}}, \\frac {y - y _ {\\operatorname* {m i n}}}{y _ {\\operatorname* {m a x}} - y _ {\\operatorname* {m i n}}}, \\frac {z - z _ {\\operatorname* {m i n}}}{z _ {\\operatorname* {m a x}} - z _ {\\operatorname* {m i n}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 433, + 785, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $[x_{\\min}, y_{\\min}, z_{\\min}, x_{\\max}, y_{\\max}, z_{\\max}] = [-40m, -40m, -1m, 40m, 40m, 5.4m]$ , are the preset boundaries of the scene. Through sine and cosine functions, we encode the normalized coordinates into high-dimensional positional information. Then we use a small MLP containing two linear layers and one ReLU layer to transform the high-dimensional positional information into learnable embeddings.", + "bbox": [ + 212, + 472, + 797, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Point Decoder", + "text_level": 1, + "bbox": [ + 215, + 585, + 382, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We used three decoder layers to build our point decoder. Each layer of the decoder contains point cross-attention (PCA) and group point cross-attention (GPCA).", + "bbox": [ + 212, + 609, + 782, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The purpose of PCA and GPCA is to integrate position embedding with image features, thereby facilitating a more cohesive representation. Owing to the high efficiency of deformable attention [15], our point decoder mechanisms employ this technique. Consequently, each query within our framework can be updated as follows:", + "bbox": [ + 212, + 656, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D A} (\\mathbf {q}, \\mathbf {p}, \\mathbf {F}) = \\sum_ {s = 1} ^ {N _ {s}} \\mathbf {A} _ {s} \\mathbf {W} _ {s} \\mathbf {F} ^ {\\mathbf {2 D}} (\\mathbf {p} + \\Delta \\mathbf {p} _ {s}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 743, + 785, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N_{s}$ represents the number of sampling offsets, $\\mathbf{A}_s$ represents the learnable attention weights, $\\mathbf{F}^{2\\mathbf{D}}(\\mathbf{p} + \\varDelta \\mathbf{p}_s)$ represents the image features collected at location $\\mathbf{p} + \\varDelta \\mathbf{p}_s$ in which $\\varDelta\\mathbf{p}_s$ represents the offset apply to position $\\mathbf{p}$ .", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point cross-attention. The point cross-attention mechanism begins by taking point queries, initially set to zero, and combines them with point position encoding derived from our 3D position encoder. This combination forms the input queries. Subsequently, these queries undergo deformable cross-attention with 2D image features. Given that not all 3D points project onto the image plane, particularly in the context of the six surround views provided by the nuScenes Dataset. Each point is likely to map onto only one or two of these views. We utilize the camera's intrinsic and extrinsic parameters to ascertain which images a given point can map to. This approach ensures that for any specific point, we only consider the features of the image(s) it maps onto, significantly reducing memory consumption. $\\mathbf{F}^{2D} = \\{\\mathbf{F}_t^{2D},\\mathbf{F}_{t - 1}^{2D},\\ldots \\}$ represents the mapped 2D image features where $t$ indexes the images. Therefore, the formula of the point cross-attention can be described as follows:", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P C A} (\\mathbf {q}, \\mathbf {F} ^ {2 D}) = \\frac {1}{| \\mathcal {V} _ {t} |} \\sum_ {t \\in \\mathcal {V} _ {t}} \\mathrm {D A} (\\mathbf {q}, \\mathcal {P} (\\mathbf {p}, t), \\mathbf {F} _ {t} ^ {2 D}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 345, + 785, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{V}_t$ represents the hit image, $t$ indexes the images, $\\mathcal{P}(\\mathbf{p}, t)$ represents the projection function of input position $\\mathbf{p}$ .", + "bbox": [ + 214, + 383, + 785, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Group point cross-attention. The group point cross-attention mechanism is aimed at solving the lack of local context in PCA since each point independently interacts with the image features. We adaptively oversample a group of $M$ ( $M = 8000$ ) points around our PoIs whose coordinates are calculated by a linear layer. We use the attention obtained from PCA and the 2D image features mapped by our group of points. Therefore, the formula of the group point cross-attention can be described as follows:", + "bbox": [ + 214, + 424, + 787, + 529 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {G P C A} \\left(\\mathbf {q} _ {\\mathbf {g}}, \\mathbf {F} ^ {2 D}\\right) = \\frac {1}{\\left| \\mathcal {V} _ {t} \\right|} \\sum_ {t \\in \\mathcal {V} _ {t}} \\mathrm {D A} \\left(\\mathbf {q} _ {\\mathbf {g}}, \\mathcal {P} \\left(\\mathbf {p} _ {\\mathbf {g}}, t\\right), \\mathbf {F} _ {t} ^ {2 D}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 532, + 785, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{q_g}$ represents attention calculated by PCA and $\\mathbf{p_g}$ represents the input position of the oversampled group of points.", + "bbox": [ + 214, + 570, + 785, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 618, + 377, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We apply class-wise cross-entropy loss and dice loss to this task. The ground truth $\\hat{\\mathbf{G}}_t \\in \\{c_0, c_1, \\dots, c_M\\}$ represent semantic information of a group of spatial points. The class-wise cross-entropy loss can be computed by:", + "bbox": [ + 214, + 640, + 785, + 685 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = - \\sum_ {n = 1} ^ {N} \\sum_ {c = c _ {0}} ^ {c _ {M}} w _ {c} \\hat {g} _ {n, c} \\log \\left(\\frac {e ^ {g _ {n , c}}}{\\sum_ {c} e ^ {g _ {n , c}}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 691, + 785, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n$ is the index of points, $N$ is the number of selected points, $c$ indexes class, $g_{n,c}$ is the predicted logits for the $n$ -th point belonging to class $c$ , $w_{c}$ is a weight for each class according to the inverse of the class frequency as in [21]. The dice loss can be computed by:", + "bbox": [ + 214, + 739, + 787, + 799 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i c e} = 1 - \\frac {2 \\sum_ {i = 1} ^ {N} p _ {i} g _ {i}}{\\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} g _ {i} ^ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 805, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 271, + 114, + 611, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_i$ is the predicted probability for the point and $g_i$ is the ground truth binary label for the point. Our final loss is the sum of the two losses:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {d i c e}} + \\mathcal {L} _ {\\text {c e}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 191, + 785, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 231, + 366, + 248 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 265, + 428, + 281 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We perform our experiments on the nuScenes Dataset with annotation provided by Occ3D. Occ3D-nuScenes benchmark is interested in a volume of $80.0\\mathrm{m}$ to the front and back of the car, $80.0\\mathrm{m}$ to the left and right side, and $6.4\\mathrm{m}$ in height. Each sample is divided as a group of 3D voxel grids with a dimension of [200, 200, 16] since each voxel has a size of $[0.4m, 0.4m, 0.4m]$ . There are 18 categories of occupancy semantics, of which the 18th category is empty and does not participate in the evaluation. Occ3D-nuScenes provides ground-truth semantical voxel grids by aggregate points of the static scenes and moving objects, respectively. Furthermore, Occ3D-nuScenes utilizes ray-casting-based methods to estimate camera visibility and provides camera visibility masks.", + "bbox": [ + 212, + 292, + 787, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metric. For this task, we use mIoU as the evaluation metric, which can be formulated as follows:", + "bbox": [ + 212, + 444, + 785, + 473 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nm I o U = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\frac {T P _ {c}}{T P _ {c} + F P _ {c} + F N _ {c}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 484, + 785, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following the definition of 3D occupancy prediction, we only evaluate our results in visible regions.", + "bbox": [ + 212, + 536, + 785, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. We input six images of the original size into ResNet10 to obtain image features. Our image backbone is pre-trained on the FCOS3D [26], which is the same backbone used in the BEVFormer baseline. Then the features will be taken by FPN to produce multi-scale feature maps. The dimension of features is $d = 256$ . We stacked three decoder layers, with point cross-attention and group point cross-attention in each decoder layer. There is a small MLP as our point-based occupancy predictor that projects the 256 feature dimension to the number of classes. We train our model on 8 NVIDIA 3090 GPUs with 24 epochs. We use the AdamW optimizer with a learning rate of $2 \\times 10^{-4}$ and a weight decay of 0.01. The learning rate of the backbone is 10 times smaller. Besides, we conducted a separate experiment in which we trained our decoder with a frozen BEVFormer baseline backbone using the same setting above and this decoder will be used as a plugin module for the BEVFormer baseline.", + "bbox": [ + 212, + 566, + 789, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baseline methods. We compare Occupancy as Set of Points with existing methods replicated by Occ3D on Occ3D-nuScenes benchmark, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32] and BEVFormer [15]. To make a fair comparison with the baseline, we provide results using BEVFormer by our implementation using the camera visible mask during the training.", + "bbox": [ + 212, + 763, + 805, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/50c385c2004b4f13d970befbd1a6e55ef338c1968a1530ba8150724e2a76613d.jpg", + "table_caption": [ + "Table 1: 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. * means the performance is achieved by our implementation using the camera mask during training. † means the performance is achieved with a frozen BEVFormer baseline backbone." + ], + "table_footnote": [], + "table_body": "
MethodothersbarrierbicyclebuscarCons. Vehmotorcyclepedestriantraffic conetrailertruckDri. Surother flatsidewalkterrainmammade vegetationmIoU
MonoScene [4]1.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.656.06
TPVFormer [9]7.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.7827.83
BEVDet [7]4.3930.310.2332.2634.4712.9710.3410.366.268.9323.6552.2724.6126.0622.3115.0415.1019.38
OccFormer [32]5.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.9721.93
BEVFormer [15]5.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.3028.0620.0417.6926.88
CTF-Occ [25]8.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.028.53
SurroundOcc [29]8.739.219.741.446.218.720.626.423.327.032.578.038.346.649.636.731.634.4
OpenOccupancy [27]10.445.723.642.449.314.824.627.727.827.633.379.239.847.150.537.731.836.1
BEVDet-depth [8]6.641.27.042.748.418.412.922.018.228.233.280.139.749.152.139.933.833.7
BEVDet-stereo [8]8.645.914.346.051.223.818.924.122.333.637.981.540.552.655.946.941.238.0
BEVFormer* [15]8.5446.2420.2847.4653.0419.3325.3926.1625.133.4537.7581.1738.6449.3252.5442.7336.1337.84
OSP11.247.2527.0647.5753.6623.2129.3729.6828.4132.3937.9479.3541.4650.3153.2340.5235.3939.41
\\( OSP^† \\)8.8746.3321.3247.5153.1419.626.1226.8426.6833.6737.9481.2139.1349.4852.7642.7336.1238.20
BEVFormer w/ \\( OSP^† \\)10.9549.027.6850.2455.9922.9631.0230.9130.2535.641.2382.0942.5951.955.144.8238.1741.21
", + "bbox": [ + 222, + 212, + 785, + 381 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Performance", + "text_level": 1, + "bbox": [ + 215, + 414, + 366, + 428 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation strategy. Our evaluation strategy for the method comprehensively utilized the three types of Points of Interest (PoIs) previously outlined. This approach entailed: (1) Standard Grids. We conducted sampling at the centers of grids, enabling us to benchmark our method's mIoU against traditional volume-based methods. This comparison provided a direct assessment of our method's performance in a standard scenario. (2) Adaptively Sampling. We adaptively oversample in our training phase to enhance our performance. Furthermore, we integrated our method as a complementary tool with volume-based approaches. By adaptively selecting PoIs that need refinement, our method enhances and augments existing volume-based techniques as shown in Fig. 3. (3) Manually Sampling. We manually select points out of the standard perception range and test out the model's perceptual ability.", + "bbox": [ + 212, + 444, + 787, + 626 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Through these varied and thorough evaluation techniques, we are able to comprehensively assess the performance and flexibility of our method across different scenarios and use cases.", + "bbox": [ + 212, + 627, + 785, + 671 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Standard grids. In Tab. 1, we benchmarked our method against existing camera-based 3D occupancy prediction techniques on the Occ3D-nuScenes benchmark by setting PoIs to standard grids. Our point-based approach achieved a notable improvement, obtains 1.57 mIoU performance gain compared to our implementation of the Bevformer baseline. Notably, our method outperformed the BEVFormer baseline across almost all category metrics, demonstrating a particularly clear advantage in detecting small targets, such as the bicycle $(20.28 \\rightarrow 27.06)$ , motorcycle $(25.39 \\rightarrow 29.37)$ , pedestrian $(26.16 \\rightarrow 29.37)$ and traffic cone $(25.1 \\rightarrow 28.31)$ . This advantage is mainly because the direct sampling of spatial points is beneficial for feature extraction and mapping of small objects.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 271, + 113, + 609, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/36c0684c0386e0388fcbe44d86f3de89e6b0d20b734825c9b09116904f526eff.jpg", + "image_caption": [ + "Fig. 3: Pipeline of refining volume-based methods with OSP. Given RGB images, 2D features are extracted by the frozen image backbone of the volume-based method which in our case is the BEVFormer baseline. We use the volume-based decoder to infer the entire scene, the point decoder to infer our selected 3D points, and combine the results of both using a weighted sum method." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 779, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adaptively Sampling. (1) The results shown in Tab. 4 demonstrate that our adaptively oversampling strategy improved the results from 38.01 to 38.48. (2) In the adaptive refinement experiment, we use our own implementation of the BEVFormer baseline as a representative of volume-based methods and enhance it with our method using different scales of PoIs by adaptively selecting locations with low confidence. Our model is trained with the frozen BEVFormer baseline backbone. The experimental results, as shown in Tab. 3, confirm that our plugin model significantly enhances BEVformer's performance. Notably, the BEVFormer w/ OSP results surpass those obtained from the two models operating independently. Furthermore, as we increase the size of the refined scene, the mIoU scores of BEVFormer w/ OSP also exhibit marked improvements, indicating that our refinement approach is very effective across various areas.", + "bbox": [ + 212, + 507, + 787, + 690 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Manually Sampling. The Occ3D-nuScenes dataset offers annotations across an $[80m\\times 80m]$ range. We manually set the points and annotations within a smaller $[60m\\times 60m]$ area and set PoIs to points out of it. To comprehensively evaluate our method, we conducted assessments across three distinct ranges: the standard $[60m\\times 60m]$ range, the full dataset range of $[80m\\times 80m]$ , and an intermediate-range spanning from $60m$ to $80m$ . Results shown in Tab. 2 demonstrate the capability of our method to make predictions at manually selected locations which is outside the predefined range in this case.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8bbfee116e7cf4b097259fec1ed671207cf93a6613fd5e32ce76c6115962dc49.jpg", + "table_caption": [ + "Table 2: Predict regions beyond the scene. The model was trained on scenes within a $60m \\times 60m$ range, and evaluations were conducted on the training scenes, the entire size of the scene, and the distance from $60m$ to $80m$ ." + ], + "table_footnote": [], + "table_body": "
MethodSceneRangemIoU
OSPStandard0m ~ 60m42.29
OSPLarge0m ~ 80m29.00
OSPBeyond60m ~ 80m23.00
", + "bbox": [ + 228, + 198, + 771, + 272 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/aab992bdbd0e93a61c94cc6b7ba69f52a63779f218261d2763b6a00628f32ccd.jpg", + "table_caption": [ + "Table 3: Refine BEVFormer with our decoder. $\\dagger$ means our model is trained with a frozen BEVFormer baseline backbone. Refine scale refers to an area defined by the length multiplied by the width, centered around the self-vehicle. The entire scene spans a scale of $80m\\times 80m$ . As the refine scale increases, the mIoU also gradually increases." + ], + "table_footnote": [], + "table_body": "
MethodRefine ScalemIoU
BEVFormer-37.84
\\( OSP^† \\)-38.20
BEVFormer w/ OSP20m × 20m38.20
BEVFormer w/ OSP40m × 40m39.35
BEVFormer w/ OSP50m × 50m40.32
BEVFormer w/ OSP80m × 80m41.21
", + "bbox": [ + 230, + 383, + 772, + 503 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation studies", + "text_level": 1, + "bbox": [ + 215, + 540, + 397, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Model architecture. We conducted ablation experiments on the model structure, particularly exploring variations in the number of transformer layers and the adaptively oversampling strategy. The results of these experiments are presented in Tab. 4. Even without an oversampling strategy and with just a 3-layer transformer, our method's baseline already surpasses the BEVFormer baseline, advancing from 37.84 to 38.01 in mIoU. The oversampling strategy further amplifies this improvement by enhancing the connections between spatial points, culminating in an increased mIoU of 38.48.", + "bbox": [ + 212, + 573, + 787, + 695 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/395d7db5240c3b60a7adf2004cc391f5761fb87d6923110e80fef5c8501d010e.jpg", + "table_caption": [ + "Table 4: Model architecture. To make a fair comparison with the baseline, we set the number of layers to three. The oversampling strategy brings $0.47\\mathrm{mIoU}$ ." + ], + "table_footnote": [], + "table_body": "
Layer NumOversamplemIoU
1-36.76
3-38.01
3✓38.48
", + "bbox": [ + 228, + 763, + 771, + 837 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 271, + 114, + 611, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "2D image features. The performance of our method is heavily reliant on the quality of image features. To elucidate this dependency, we conducted ablation studies focusing on the output from the model's 'neck' component. These experiments demonstrated the substantial impact of using a multi-scale output from the FPN. This architectural choice boosts the model's performance, as evidenced by the increase in mIoU from 38.67 to 39.41 as shown in Tab. 5.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Grid center sampling method. Our ablation experiments also investigated the technique for the grid center sampling method during training. While initially using grid center points as coordinates, we introduced a variation by adding random perturbations. The results, as detailed in Tab. 5, indicate the introduction of this disturbance to the grid center points elevates the mIoU to 38.67.", + "bbox": [ + 212, + 237, + 787, + 313 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e2ff541d2001225b1ba0d73a3841ac34d69ad963bde736e3d933586e8f5b2dd6.jpg", + "table_caption": [ + "Table 5: 2D image features and point perturbation. Increasing the multi-scale feature maps of images from 2 to 4 brings $0.74\\mathrm{mIoU}$ . We randomly applied perturbations of $0.1m$ to the coordinates of each point in three directions. This strategy brings $0.19\\mathrm{mIoU}$" + ], + "table_footnote": [], + "table_body": "
Multi-scale Images FeaturesPoint PerturbationmIoU
2-38.48
2✓38.67
4✓39.41
", + "bbox": [ + 228, + 406, + 771, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss design. The task of 3D occupancy prediction bears similarities to segmentation tasks. In the context of the Occ3D-nuScenes dataset, there is a notable issue of class imbalance. To address this challenge, we opted to incorporate dice loss into our optimization strategy. The effectiveness of this choice is evident in the results presented in Tab. 6, which demonstrate that the integration of dice loss benefits our method.", + "bbox": [ + 212, + 520, + 787, + 609 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9dc705170cae3f24b6e9187006045f42d77cba9a825d974a7bf4d6f2240e600e.jpg", + "table_caption": [ + "Table 6: Loss design. Incorporating dice loss for optimization brings $0.53\\mathrm{mIoU}$ ." + ], + "table_footnote": [], + "table_body": "
Dice LossmIoU
-38.88
✓39.41
", + "bbox": [ + 228, + 660, + 771, + 720 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Adaptive point sampling. During the inference process, we apply adaptive sampling (adaptively select points with high uncertainty) in BEVFormer w/ OSP and OSP. Adaptive sampling can reduce computational burden while maintaining good performance as shown in Tab. 7. BEVFormer w/ OSP using adaptive sampling means we refine BEVFormer by adaptively selecting points with high uncertainty (around $20\\%$ ). OSP using adaptive sampling means we only forward", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2a3dbc240ab4eda7f604e8326eb92905785f7c7805511f3d6f5385d74a92d5d2.jpg", + "table_caption": [ + "Table 7: Adaptability of OSP. 'rel' denotes relative." + ], + "table_footnote": [], + "table_body": "
MethodAdaptive SamplingRel. ComputationmIoU
BEVFormer--37.84
BEVFormer w/ OSP-1.0×39.35
BEVFormer w/ OSP✓0.8×39.22
OSP-1.0×39.41
OSP✓0.93×39.08
", + "bbox": [ + 228, + 171, + 769, + 277 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "points with low confidence to the next decoder layer, while directly outputting the results for points with high confidence in the decoder. Points of high uncertainty are defined as those with a confidence score less than a threshold of 0.9 after softmax.", + "bbox": [ + 212, + 308, + 784, + 368 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2daec395d6a850edda3f9b6943bfd379555023b28f066d76d77de987a8c88d00.jpg", + "image_caption": [ + "Fig. 4: Visualization of our results. Our visualization results have voids compared to the ground truth ego-vehicle position, which is due to the surrounding view having gaps near the ego-vehicle." + ], + "image_footnote": [], + "bbox": [ + 223, + 397, + 782, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 680, + 359, + 696 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present a novel perspective on 3D scene representation, viewing it through a set of points. We introduce the innovative concept of Points of Interest (PoIs), which significantly advances the flexibility in scene representation. Building upon the foundation laid by PoIs, we develop a highly adaptable, point-based 3D occupancy prediction framework, named OSP. We validate the strong performance and flexibility of OSP on the Occ3D-nuScenes benchmark. Our work not only contributes to the field of 3D occupancy prediction but also paves the way for more dynamic and adaptable methods in 3D scene analysis.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 269, + 113, + 611, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (No. 62376102) and the Agricultural Science and Technology Independent Innovation Foundation of Jiangsu Province, China, No. CX(22)2031.", + "bbox": [ + 215, + 178, + 784, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 248, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019) 4", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) 3", + "3. Cai, Y., Chen, X., Zhang, C., Lin, K.Y., Wang, X., Li, H.: Semantic scene completion via integrating instances and scene in-the-loop. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 324-333 (2021) 4", + "4. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3991-4001 (2022) 3, 4, 9, 10", + "5. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4193-4202 (2020) 4", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 6", + "7. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) 2, 3, 5, 9, 10", + "8. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR (2021) 10", + "9. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR. pp. 9223-9232 (2023) 3, 4, 9, 10", + "0. Jiang, H., Cheng, T., Gao, N., Zhang, H., Liu, W., Wang, X.: Symphonize 3d semantic scene completion with contextual instance queries. CoRR (2023) 4", + "1. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3351-3359 (2020) 4", + "2. Li, J., Liu, Y., Gong, D., Shi, Q., Yuan, X., Zhao, C., Reid, I.D.: RGBD based dimensional decomposition residual network for 3d semantic scene completion. In: CVPR. pp. 7693-7702 (2019) 4", + "3. Li, J., Liu, Y., Yuan, X., Zhao, C., Siegwart, R., Reid, I., Cadena, C.: Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters 5(1), 219-226 (2019) 4" + ], + "bbox": [ + 225, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI. pp. 1486-1494 (2023) 5", + "15. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Yu, Q., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270 (2022) 2, 3, 4, 7, 9, 10", + "16. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. In: Advances in Neural Information Processing Systems. vol. 31 (2018) 4", + "17. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625 (2022) 5", + "18. Liu, Y., Yan, J., Jia, F., Li, S., Gao, Q., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022) 5", + "19. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023) 4", + "20. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE international conference on robotics and automation. vol. 2, pp. 116-121. IEEE (1985) 3", + "21. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 2020 International Conference on 3D Vision (3DV). pp. 111-119. IEEE (2020) 8", + "22. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1746-1754 (2017) 1, 3", + "23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) 3", + "24. Thrun, S.: Probabilistic robotics. Communications of the ACM 45(3), 52-57 (2002) 3", + "25. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving (2023) 1, 3, 10", + "26. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. arXiv preprint arXiv:2104.10956 (2021) 4, 9", + "27. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023) 3, 10", + "28. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022) 4, 5", + "29. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023) 10", + "30. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 733-749 (2018) 4", + "31. Zhang, P., Liu, W., Lei, Y., Lu, H., Yang, X.: Cascaded context pyramid for full-resolution 3d semantic scene completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7801-7810 (2019) 4" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang", + "bbox": [ + 271, + 114, + 609, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316 (2023) 3, 9, 10", + "33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2020) 5", + "34. Zuo, S., Zheng, W., Huang, Y., Zhou, J., Lu, J.: Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction (2023) 4" + ], + "bbox": [ + 212, + 146, + 787, + 244 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Occupancy as Set of Points", + "bbox": [ + 547, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_model.json b/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e461f9c5b137d557871fedb540ece64fcf74ef64 --- /dev/null +++ b/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_model.json @@ -0,0 +1,2060 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.347, + 0.142, + 0.658, + 0.163 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.189, + 0.674, + 0.221 + ], + "angle": 0, + "content": "Yiang Shi\\(^{1,\\star}\\), Tianheng Cheng\\(^{1,\\star}\\), Qian Zhang\\(^{2}\\), Wenyu Liu\\(^{1}\\), and Xinggang Wang\\(^{1,\\text{忍}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.232, + 0.712, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\) School of EIC, Huazhong University of Science & Technology" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.247, + 0.568, + 0.259 + ], + "angle": 0, + "content": "2 Horizon Robotics" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.307, + 0.744, + 0.515 + ], + "angle": 0, + "content": "Abstract. In this paper, we explore a novel point representation for 3D occupancy prediction from multi-view images, which is named Occupancy as Set of Points. Existing camera-based methods tend to exploit dense volume-based representation to predict the occupancy of the whole scene, making it hard to focus on the special areas or areas out of the perception range. In comparison, we present the Points of Interest (PoIs) to represent the scene and propose OSP, a novel framework for point-based 3D occupancy prediction. Owing to the inherent flexibility of the point-based representation, OSP achieves strong performance compared with existing methods and excels in terms of training and inference adaptability. It extends beyond traditional perception boundaries and can be seamlessly integrated with volume-based methods to significantly enhance their effectiveness. Experiments on the Occ3D-nuScenes occupancy benchmark show that OSP has strong performance and flexibility. Code and models are available at https://github.com/hustvl/osp." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.527, + 0.741, + 0.556 + ], + "angle": 0, + "content": "Keywords: 3D Occupancy Prediction \\(\\cdot\\) Autonomous Vehicles \\(\\cdot\\) Multi-view 3D Perception" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.579, + 0.377, + 0.595 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.787, + 0.683 + ], + "angle": 0, + "content": "Holistic 3D scene understanding is crucial for autonomous driving systems, directly affecting the efficiency and accuracy of subsequent tasks. Considering the cost-effectiveness and ease of deployment of cameras compared to other sensors, developing visual-based methods for 3D scene understanding has become a significant and widely researched challenge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.804 + ], + "angle": 0, + "content": "To tackle this challenge, 3D Semantic Scene Completion (SSC) [22] has been proposed and widely studied to jointly infer the geometry and semantics information of the scene from limited observations. The SSC task requires the model to accurately predict the visible locations and complete the information for the invisible locations. Recently, Occ3D [25] introduces a new task definition called 3D occupancy prediction. The main difference between this task and SSC is that 3D occupancy prediction only focuses on the visible areas and is tailored for dynamic scenes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.784, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author: Xinggang Wang (xgwang@hust.edu.cn)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.612, + 0.13 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.15, + 0.571, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.15, + 0.781, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Fig. 1: Comparison between volume-based methods and our method. The volume-based methods, represented by BEVFormer, infers every region within the scene and gets standard occupancy as shown in (a). Our method uses a point-based decoder as shown in (b). Thus it infers the Points of Interest including standard, adaptively sampled, and manually sampled grids as shown in (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.639 + ], + "angle": 0, + "content": "Existing 3D occupancy prediction methods are mostly based on dense BEV methods, e.g., BEVFormer [15], BEVDet [7]. These methods integrate a BEV encoder with an occupancy head to generate the output and enhance BEV perception capabilities for better results. However, they share some common drawbacks. (1) Uniform Sampling: BEV-based methods fail to differentiate between different areas within the same scene, treating them equally. This leads to coarse sampling and hinders dynamic or multi-resolution sampling capabilities. (2) Limited Inference Flexibility: During inference, these methods can only process the entire scene at once. They lack the ability to infer different parts of the scene based on varying downstream tasks or specific practical needs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Those limitations highlight the need for more flexible 3D occupancy prediction methods that can handle complex scenes while adapting to different inference requirements. In this paper, we propose a novel point-based representation for 3D occupancy prediction. Instead of dividing the scene into uniform grids applied by existing volume-based methods, we propose Points of Interest (PoIs) to view the scene as a collection of points that help in flexibly sampling the scene during both training and inference stages. Fig. 1 compares volume-based and point-based representations. Compared to volume-based representations, our point-based representation has the following advantages: (1) it can accept inputs of any scale and position to make occupancy predictions, including manually designed and adaptively designed input, offering flexibility; (2) it can pay extra attention to certain areas rather than treating all areas equally, enhancing the model's perceptual capabilities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "We introduce Occupancy as Set of Points (OSP), a novel and flexible point-based framework, which is built upon the foundational concept of Points of Interest (PoIs). OSP excels in 3D occupancy prediction and is composed of an image backbone, a 3D positioning encoder, and a decoder, as illustrated in Fig. 2. Central to our methodology is the innovative use of PoIs, which we have categorized into three distinct types to meet diverse needs, thereby significantly enhancing various aspects of our model's performance. Detailed descriptions of these PoIs will be provided in Sec. 3. Notably, PoIs can be designed as needed beyond the three types initially proposed by us." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Our method stands out for its strong performance and flexibility. The flexibility enables it to process any arbitrary local scene without necessitating retraining. Additionally, our method can serve as an augmentative plugin module for existing volume-based methods by adaptively resampling areas of low confidence to yield more accurate occupancy predictions. It is also adept at predicting areas beyond the scene. The key contributions of our approach can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.396, + 0.784, + 0.437 + ], + "angle": 0, + "content": "- A novel point-based occupancy representation, established by interacting point queries with 2D image features, enables a comprehensive understanding of 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.44, + 0.784, + 0.468 + ], + "angle": 0, + "content": "- A flexible framework that allows for inference at any area of interest without retraining or sacrificing accuracy and predicts areas beyond the scene." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.47, + 0.784, + 0.498 + ], + "angle": 0, + "content": "- A plugin module that enhances the performance of the volume-based baseline significantly." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.499, + 0.784, + 0.527 + ], + "angle": 0, + "content": "- OSP has obtained strong experiment results, i.e., 39.4 mIoU in the 3D occupancy prediction task on the Occ3D-nuScenes benchmark." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.396, + 0.784, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.55, + 0.388, + 0.566 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.579, + 0.475, + 0.594 + ], + "angle": 0, + "content": "2.1 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.753 + ], + "angle": 0, + "content": "3D occupancy prediction, a concept recently defined by Occ3D [25], exhibits notable parallels with Occupancy Grid Mapping (OGM) [20, 24, 27] used in robotics. This task aims to predict the state of each voxel grid in a scene based on a series of sensor inputs. Occ3D establishes two benchmarks leveraging the Waymo Open Dataset [23] and the nuScenes Dataset [2] to facilitate this. In vision-based 3D Occupancy prediction, Occ3D implements camera visibility estimation and creates visibility masks to ensure evaluations are confined to visible areas. It also evaluates various SSC methodologies on its benchmarks, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32], and BEV-Former [15]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.773, + 0.522, + 0.789 + ], + "angle": 0, + "content": "2.2 3D Semantic Scene Completion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Scene Semantic Completion (SSC) represents a task closely associated with 3D occupancy prediction. The concept of SSC is initially presented in SSC-Net [22], with a focus on predicting the comprehensive semantic information" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.612, + 0.129 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.403 + ], + "angle": 0, + "content": "of a scene based on its partially visible regions. Over recent years, the study of SSC has expanded significantly, particularly in the context of small indoor scenes [3,5,11-13,16,30,31]. In recent times, the study of Scene Semantic Completion (SSC) for expansive outdoor environments has gained momentum, particularly following the introduction of the SemanticKITTI dataset [1]. Notably, MonoScene [4] emerges as the first method to apply monocular pure vision-based SSC. In a parallel advancement, OccDepth [19] enhances 2D to 3D feature transformation by incorporating depth data from stereo input. TPVFormer [9] argues against the limitations of single-plane modeling in capturing intricate details, hence it adopts a tri-perspective view (TPV) approach, combining a Bird's Eye View (BEV) with two additional vertical planes. Additionally, Symphonies [10] highlights the significance of instance representation in SSC tasks. While SSC methods can be directly applied to 3D occupancy prediction, two primary distinctions exist: (1) SSC primarily aims to infer the occupancy of non-visible areas based on visible regions, in contrast to 3D occupancy prediction which focuses on visible areas; (2) SSC methods usually target static scenes, whereas 3D occupancy prediction methods are often designed to handle dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Most existing volume-based SSC and 3D occupancy prediction methods are characterized by their dense nature, encompassing inputs and outputs that span the entire scene. Consider the BEVFormer baseline as an example: it segments the scene into uniform BEV grids, failing to distinguish between grids in varying areas. This uniformity restricts the ability of volume-based methods like BEVFormer to sample areas of interest for better performance during training. Besides, if we want to focus on a specific area in the inference stage, volume-based methods are limited to infer the entire scene and then perform post-processing, inevitably leading to increased and unnecessary costs. Moreover, as scene size and voxel resolution increase, the computational demands skyrocket exponentially. In stark contrast, our point-based model introduces much-needed flexibility by focusing on PoIs. Our method facilitates direct inference in specific areas, eliminating the need for post-processing and avoiding additional computational burdens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.678 + ], + "angle": 0, + "content": "A point-related SSC method is PointOcc [34], a point cloud-based SSC prediction method using three complementary view planes for efficient point cloud feature modeling and an efficient 2D backbone for processing to reduce computational load, while our method focuses on the flexibility of training and inference." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.705, + 0.496, + 0.719 + ], + "angle": 0, + "content": "2.3 Camera-based 3D Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Camera-based 3D perception tasks have gained substantial attention in recent research, largely due to the convenience and cost-effectiveness of cameras as data collection sensors. Initial efforts, such as FCOS3D [26] and DETR3D [28], explore the transition from 2D to 3D predictions. BEVFormer [15] represents a significant advancement in this area, transforming images captured from vehicle-mounted cameras into a bird's eye view (BEV) representation. This technique not only enhances a vehicle's environmental understanding but also finds appli" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "cations in various downstream tasks like BEVStereo [14] and BEVDet [7] and extends to 3D occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.376 + ], + "angle": 0, + "content": "3D detection tasks are important in camera-based 3D perception and also have a high similarity to 3D occupancy prediction. The primary objective of 3D detection involves estimating the position and dimensions of objects in 3D space. DETR3D [28], drawing inspiration from DETR [33], innovatively combines 3D object queries with image features, incorporating camera intrinsic and extrinsic parameters. PETR [17] and its successor, PETRv2 [18], further refine this approach by addressing the accuracy of reference point sampling and the inclusion of global information, enhancing 3D detection through historical data integration. Compared to object detection, 3D occupancy prediction offers a finer granularity, which is crucial for navigating irregular obstacles or overhanging objects. It is imperative to explore how insights from BEV representation and 3D object detection can inspire innovative solutions tailored to the specific requirements of 3D occupancy prediction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.405, + 0.369, + 0.423 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.579 + ], + "angle": 0, + "content": "Problem setup. We aim to provide occupancy predictions around the ego-vehicle, given only \\(N\\) surround-view RGB images. More specifically, we use as input current images denoted by \\(\\mathbf{I}_i = \\{I_0,I_1,\\dots,I_N\\}\\), and use as output an occupancy prediction \\(\\mathbf{Y}_i\\in \\{c_0,c_1,\\dots,c_M\\}^{H\\times W\\times Z}\\) defined in the coordinate of ego-vehicle, where each occupancy prediction is either empty (denoted by \\(c_{0}\\)) or occupied by a certain semantic class in \\(\\{c_{1},c_{m},\\dots,c_{M}\\}\\). We assume known camera intrinsic parameters \\(\\{K_i\\}\\) and extrinsic parameters \\(\\{[R_i|t_i]\\}\\) in each frame. We assume to know whether each area is visible or not by applying a camera visibility mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.67 + ], + "angle": 0, + "content": "Points of Interest. In our model, we innovatively introduce the concept of Points of Interest (PoIs), which is a set of sparse points to represent the 3D scene. PoIs can flexibly represent objects or regions that require extra attention such as pedestrians or regions near ego-vehicles and can be designed as needed in both training and inference phases. We use three types of PoIs and the definitions and functions are introduced as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.717 + ], + "angle": 0, + "content": "(1) Standard Grids: By sampling center points of grids in the inference stage and making predictions of standard 3D occupancy grids, our model makes a fair comparison with existing methods and achieves good results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "(2) Adaptively Sampling: During the training stage, our model adaptively samples points and oversamples points around them to enhance accuracy. Recognizing that volume-based methods uniformly treat all locations, our point-based approach allows for resampling in either areas of special interest or those that are challenging to learn. This adaptively resampling strategy is also used to augment the performance of volume-based methods. Consequently, our method can function as a versatile plugin, seamlessly integrating with and enhancing existing volume-based approaches." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.612, + 0.13 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "(3) Manually Sampling: Our model excels in its flexibility to sample any area, specifically catering to the unique demands of various downstream tasks. Our model can make predictions of areas beyond the standard perception range by setting PoIs to areas outside the scene manually, e.g. 200 meters away from ego-vehicle, which is a feat unattainable by traditional volume-based methods. This extension not only broadens the scope of inference but also introduces a new dimension to scene understanding. These PoIs are the foundation of our method, offering precision and high flexibility." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.298, + 0.784, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Fig. 2: Overall framework of Occupancy as Set of Points. OSP leverages the Transformer architecture to derive 3D point features from 2D images to make 3D occupancy predictions. Initially, we extract 2D features from multi-view images. Following this, we employ a set of 3D point queries to index these 2D features. The selection of these 3D point queries depends on the Points of Interest (PoIs)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.673, + 0.331, + 0.687 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.433, + 0.721 + ], + "angle": 0, + "content": "4.1 Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Fig. 2 shows the overall architecture of the proposed Occupancy as Set of Points. Given the images \\(\\mathbf{I} = \\{I_i, i = 1, 2, \\dots, N\\}\\) from \\(\\mathbf{N}\\) views, We feed these images into ResNet [6] to obtain their features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "OSP begins with sampling a set of 3D points in the space, which are the initial PoIs. In our experiments, we use center points of grids as the initial PoIs to make a fair comparison with traditional volume-based methods and provide a reliable baseline for performance comparison and evaluation. We sample \\( K(K = 8000) \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "points within the camera's visible region and introduce random perturbations to these points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Then we normalize the 3D points. To these normalized coordinates, we apply sine and cosine functions as a form of positional encoding. This encoded positional information is then utilized to create query position embeddings. In the training phase, within each decoder layer, the query position corresponding to each individual query remains consistent. The 3D coordinates, along with the camera's intrinsic and extrinsic parameters, are used to map these points onto the pixel plane. This mapping process yields corresponding key and value pairs. Subsequently, we employ point cross-attention mechanisms to compute the output. Then we adaptively oversample a group of \\( M \\) points (\\( M = 8000 \\)) whose coordinates are calculated by a linear layer and employ group point cross-attention to fuse the features of the additional sampling points." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.366, + 0.434, + 0.38 + ], + "angle": 0, + "content": "4.2 3D Position Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.784, + 0.422 + ], + "angle": 0, + "content": "After obtaining 3D points by applying PoIs, we first normalize coordinate points using equations as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.434, + 0.786, + 0.463 + ], + "angle": 0, + "content": "\\[\nx, y, z = \\frac {x - x _ {\\operatorname* {m i n}}}{x _ {\\operatorname* {m a x}} - x _ {\\operatorname* {m i n}}}, \\frac {y - y _ {\\operatorname* {m i n}}}{y _ {\\operatorname* {m a x}} - y _ {\\operatorname* {m i n}}}, \\frac {z - z _ {\\operatorname* {m i n}}}{z _ {\\operatorname* {m a x}} - z _ {\\operatorname* {m i n}}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.798, + 0.564 + ], + "angle": 0, + "content": "where \\([x_{\\min}, y_{\\min}, z_{\\min}, x_{\\max}, y_{\\max}, z_{\\max}] = [-40m, -40m, -1m, 40m, 40m, 5.4m]\\), are the preset boundaries of the scene. Through sine and cosine functions, we encode the normalized coordinates into high-dimensional positional information. Then we use a small MLP containing two linear layers and one ReLU layer to transform the high-dimensional positional information into learnable embeddings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.383, + 0.601 + ], + "angle": 0, + "content": "4.3 Point Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.784, + 0.656 + ], + "angle": 0, + "content": "We used three decoder layers to build our point decoder. Each layer of the decoder contains point cross-attention (PCA) and group point cross-attention (GPCA)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.735 + ], + "angle": 0, + "content": "The purpose of PCA and GPCA is to integrate position embedding with image features, thereby facilitating a more cohesive representation. Owing to the high efficiency of deformable attention [15], our point decoder mechanisms employ this technique. Consequently, each query within our framework can be updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.744, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D A} (\\mathbf {q}, \\mathbf {p}, \\mathbf {F}) = \\sum_ {s = 1} ^ {N _ {s}} \\mathbf {A} _ {s} \\mathbf {W} _ {s} \\mathbf {F} ^ {\\mathbf {2 D}} (\\mathbf {p} + \\Delta \\mathbf {p} _ {s}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(N_{s}\\) represents the number of sampling offsets, \\(\\mathbf{A}_s\\) represents the learnable attention weights, \\(\\mathbf{F}^{2\\mathbf{D}}(\\mathbf{p} + \\varDelta \\mathbf{p}_s)\\) represents the image features collected at location \\(\\mathbf{p} + \\varDelta \\mathbf{p}_s\\) in which \\(\\varDelta\\mathbf{p}_s\\) represents the offset apply to position \\(\\mathbf{p}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.612, + 0.13 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Point cross-attention. The point cross-attention mechanism begins by taking point queries, initially set to zero, and combines them with point position encoding derived from our 3D position encoder. This combination forms the input queries. Subsequently, these queries undergo deformable cross-attention with 2D image features. Given that not all 3D points project onto the image plane, particularly in the context of the six surround views provided by the nuScenes Dataset. Each point is likely to map onto only one or two of these views. We utilize the camera's intrinsic and extrinsic parameters to ascertain which images a given point can map to. This approach ensures that for any specific point, we only consider the features of the image(s) it maps onto, significantly reducing memory consumption. \\(\\mathbf{F}^{2D} = \\{\\mathbf{F}_t^{2D},\\mathbf{F}_{t - 1}^{2D},\\ldots \\}\\) represents the mapped 2D image features where \\(t\\) indexes the images. Therefore, the formula of the point cross-attention can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.346, + 0.786, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P C A} (\\mathbf {q}, \\mathbf {F} ^ {2 D}) = \\frac {1}{| \\mathcal {V} _ {t} |} \\sum_ {t \\in \\mathcal {V} _ {t}} \\mathrm {D A} (\\mathbf {q}, \\mathcal {P} (\\mathbf {p}, t), \\mathbf {F} _ {t} ^ {2 D}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.384, + 0.787, + 0.415 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}_t\\) represents the hit image, \\(t\\) indexes the images, \\(\\mathcal{P}(\\mathbf{p}, t)\\) represents the projection function of input position \\(\\mathbf{p}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.425, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Group point cross-attention. The group point cross-attention mechanism is aimed at solving the lack of local context in PCA since each point independently interacts with the image features. We adaptively oversample a group of \\( M \\) (\\( M = 8000 \\)) points around our PoIs whose coordinates are calculated by a linear layer. We use the attention obtained from PCA and the 2D image features mapped by our group of points. Therefore, the formula of the group point cross-attention can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.533, + 0.786, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\operatorname {G P C A} \\left(\\mathbf {q} _ {\\mathbf {g}}, \\mathbf {F} ^ {2 D}\\right) = \\frac {1}{\\left| \\mathcal {V} _ {t} \\right|} \\sum_ {t \\in \\mathcal {V} _ {t}} \\mathrm {D A} \\left(\\mathbf {q} _ {\\mathbf {g}}, \\mathcal {P} \\left(\\mathbf {p} _ {\\mathbf {g}}, t\\right), \\mathbf {F} _ {t} ^ {2 D}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.571, + 0.787, + 0.601 + ], + "angle": 0, + "content": "where \\(\\mathbf{q_g}\\) represents attention calculated by PCA and \\(\\mathbf{p_g}\\) represents the input position of the oversampled group of points." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.378, + 0.633 + ], + "angle": 0, + "content": "4.4 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.641, + 0.787, + 0.686 + ], + "angle": 0, + "content": "We apply class-wise cross-entropy loss and dice loss to this task. The ground truth \\(\\hat{\\mathbf{G}}_t \\in \\{c_0, c_1, \\dots, c_M\\}\\) represent semantic information of a group of spatial points. The class-wise cross-entropy loss can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.693, + 0.786, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = - \\sum_ {n = 1} ^ {N} \\sum_ {c = c _ {0}} ^ {c _ {M}} w _ {c} \\hat {g} _ {n, c} \\log \\left(\\frac {e ^ {g _ {n , c}}}{\\sum_ {c} e ^ {g _ {n , c}}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.74, + 0.788, + 0.8 + ], + "angle": 0, + "content": "where \\( n \\) is the index of points, \\( N \\) is the number of selected points, \\( c \\) indexes class, \\( g_{n,c} \\) is the predicted logits for the \\( n \\)-th point belonging to class \\( c \\), \\( w_{c} \\) is a weight for each class according to the inverse of the class frequency as in [21]. The dice loss can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.806, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i c e} = 1 - \\frac {2 \\sum_ {i = 1} ^ {N} p _ {i} g _ {i}}{\\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} g _ {i} ^ {2}}, \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "where \\( p_i \\) is the predicted probability for the point and \\( g_i \\) is the ground truth binary label for the point. Our final loss is the sum of the two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.192, + 0.786, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {d i c e}} + \\mathcal {L} _ {\\text {c e}}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.232, + 0.367, + 0.249 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.266, + 0.429, + 0.282 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Dataset. We perform our experiments on the nuScenes Dataset with annotation provided by Occ3D. Occ3D-nuScenes benchmark is interested in a volume of \\(80.0\\mathrm{m}\\) to the front and back of the car, \\(80.0\\mathrm{m}\\) to the left and right side, and \\(6.4\\mathrm{m}\\) in height. Each sample is divided as a group of 3D voxel grids with a dimension of [200, 200, 16] since each voxel has a size of \\([0.4m, 0.4m, 0.4m]\\). There are 18 categories of occupancy semantics, of which the 18th category is empty and does not participate in the evaluation. Occ3D-nuScenes provides ground-truth semantical voxel grids by aggregate points of the static scenes and moving objects, respectively. Furthermore, Occ3D-nuScenes utilizes ray-casting-based methods to estimate camera visibility and provides camera visibility masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.787, + 0.474 + ], + "angle": 0, + "content": "Evaluation metric. For this task, we use mIoU as the evaluation metric, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.486, + 0.786, + 0.526 + ], + "angle": 0, + "content": "\\[\nm I o U = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\frac {T P _ {c}}{T P _ {c} + F P _ {c} + F N _ {c}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.786, + 0.568 + ], + "angle": 0, + "content": "Following the definition of 3D occupancy prediction, we only evaluate our results in visible regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.79, + 0.765 + ], + "angle": 0, + "content": "Implementation details. We input six images of the original size into ResNet10 to obtain image features. Our image backbone is pre-trained on the FCOS3D [26], which is the same backbone used in the BEVFormer baseline. Then the features will be taken by FPN to produce multi-scale feature maps. The dimension of features is \\( d = 256 \\). We stacked three decoder layers, with point cross-attention and group point cross-attention in each decoder layer. There is a small MLP as our point-based occupancy predictor that projects the 256 feature dimension to the number of classes. We train our model on 8 NVIDIA 3090 GPUs with 24 epochs. We use the AdamW optimizer with a learning rate of \\( 2 \\times 10^{-4} \\) and a weight decay of 0.01. The learning rate of the backbone is 10 times smaller. Besides, we conducted a separate experiment in which we trained our decoder with a frozen BEVFormer baseline backbone using the same setting above and this decoder will be used as a plugin module for the BEVFormer baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.807, + 0.842 + ], + "angle": 0, + "content": "Baseline methods. We compare Occupancy as Set of Points with existing methods replicated by Occ3D on Occ3D-nuScenes benchmark, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32] and BEVFormer [15]. To make a fair comparison with the baseline, we provide results using BEVFormer by our implementation using the camera visible mask during the training." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.611, + 0.13 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. * means the performance is achieved by our implementation using the camera mask during training. † means the performance is achieved with a frozen BEVFormer baseline backbone." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.213, + 0.787, + 0.382 + ], + "angle": 0, + "content": "
MethodothersbarrierbicyclebuscarCons. Vehmotorcyclepedestriantraffic conetrailertruckDri. Surother flatsidewalkterrainmammade vegetationmIoU
MonoScene [4]1.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.656.06
TPVFormer [9]7.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.7827.83
BEVDet [7]4.3930.310.2332.2634.4712.9710.3410.366.268.9323.6552.2724.6126.0622.3115.0415.1019.38
OccFormer [32]5.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.9721.93
BEVFormer [15]5.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.3028.0620.0417.6926.88
CTF-Occ [25]8.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.028.53
SurroundOcc [29]8.739.219.741.446.218.720.626.423.327.032.578.038.346.649.636.731.634.4
OpenOccupancy [27]10.445.723.642.449.314.824.627.727.827.633.379.239.847.150.537.731.836.1
BEVDet-depth [8]6.641.27.042.748.418.412.922.018.228.233.280.139.749.152.139.933.833.7
BEVDet-stereo [8]8.645.914.346.051.223.818.924.122.333.637.981.540.552.655.946.941.238.0
BEVFormer* [15]8.5446.2420.2847.4653.0419.3325.3926.1625.133.4537.7581.1738.6449.3252.5442.7336.1337.84
OSP11.247.2527.0647.5753.6623.2129.3729.6828.4132.3937.9479.3541.4650.3153.2340.5235.3939.41
\\( OSP^† \\)8.8746.3321.3247.5153.1419.626.1226.8426.6833.6737.9481.2139.1349.4852.7642.7336.1238.20
BEVFormer w/ \\( OSP^† \\)10.9549.027.6850.2455.9922.9631.0230.9130.2535.641.2382.0942.5951.955.144.8238.1741.21
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.415, + 0.367, + 0.429 + ], + "angle": 0, + "content": "5.2 Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Evaluation strategy. Our evaluation strategy for the method comprehensively utilized the three types of Points of Interest (PoIs) previously outlined. This approach entailed: (1) Standard Grids. We conducted sampling at the centers of grids, enabling us to benchmark our method's mIoU against traditional volume-based methods. This comparison provided a direct assessment of our method's performance in a standard scenario. (2) Adaptively Sampling. We adaptively oversample in our training phase to enhance our performance. Furthermore, we integrated our method as a complementary tool with volume-based approaches. By adaptively selecting PoIs that need refinement, our method enhances and augments existing volume-based techniques as shown in Fig. 3. (3) Manually Sampling. We manually select points out of the standard perception range and test out the model's perceptual ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.672 + ], + "angle": 0, + "content": "Through these varied and thorough evaluation techniques, we are able to comprehensively assess the performance and flexibility of our method across different scenarios and use cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Standard grids. In Tab. 1, we benchmarked our method against existing camera-based 3D occupancy prediction techniques on the Occ3D-nuScenes benchmark by setting PoIs to standard grids. Our point-based approach achieved a notable improvement, obtains 1.57 mIoU performance gain compared to our implementation of the Bevformer baseline. Notably, our method outperformed the BEVFormer baseline across almost all category metrics, demonstrating a particularly clear advantage in detecting small targets, such as the bicycle \\((20.28 \\rightarrow 27.06)\\), motorcycle \\((25.39 \\rightarrow 29.37)\\), pedestrian \\((26.16 \\rightarrow 29.37)\\) and traffic cone \\((25.1 \\rightarrow 28.31)\\). This advantage is mainly because the direct sampling of spatial points is beneficial for feature extraction and mapping of small objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.78, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Fig. 3: Pipeline of refining volume-based methods with OSP. Given RGB images, 2D features are extracted by the frozen image backbone of the volume-based method which in our case is the BEVFormer baseline. We use the volume-based decoder to infer the entire scene, the point decoder to infer our selected 3D points, and combine the results of both using a weighted sum method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Adaptively Sampling. (1) The results shown in Tab. 4 demonstrate that our adaptively oversampling strategy improved the results from 38.01 to 38.48. (2) In the adaptive refinement experiment, we use our own implementation of the BEVFormer baseline as a representative of volume-based methods and enhance it with our method using different scales of PoIs by adaptively selecting locations with low confidence. Our model is trained with the frozen BEVFormer baseline backbone. The experimental results, as shown in Tab. 3, confirm that our plugin model significantly enhances BEVformer's performance. Notably, the BEVFormer w/ OSP results surpass those obtained from the two models operating independently. Furthermore, as we increase the size of the refined scene, the mIoU scores of BEVFormer w/ OSP also exhibit marked improvements, indicating that our refinement approach is very effective across various areas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Manually Sampling. The Occ3D-nuScenes dataset offers annotations across an \\([80m\\times 80m]\\) range. We manually set the points and annotations within a smaller \\([60m\\times 60m]\\) area and set PoIs to points out of it. To comprehensively evaluate our method, we conducted assessments across three distinct ranges: the standard \\([60m\\times 60m]\\) range, the full dataset range of \\([80m\\times 80m]\\), and an intermediate-range spanning from \\(60m\\) to \\(80m\\). Results shown in Tab. 2 demonstrate the capability of our method to make predictions at manually selected locations which is outside the predefined range in this case." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.612, + 0.13 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Predict regions beyond the scene. The model was trained on scenes within a \\(60m \\times 60m\\) range, and evaluations were conducted on the training scenes, the entire size of the scene, and the distance from \\(60m\\) to \\(80m\\)." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.199, + 0.772, + 0.273 + ], + "angle": 0, + "content": "
MethodSceneRangemIoU
OSPStandard0m ~ 60m42.29
OSPLarge0m ~ 80m29.00
OSPBeyond60m ~ 80m23.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.302, + 0.787, + 0.37 + ], + "angle": 0, + "content": "Table 3: Refine BEVFormer with our decoder. \\(\\dagger\\) means our model is trained with a frozen BEVFormer baseline backbone. Refine scale refers to an area defined by the length multiplied by the width, centered around the self-vehicle. The entire scene spans a scale of \\(80m\\times 80m\\). As the refine scale increases, the mIoU also gradually increases." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.384, + 0.773, + 0.505 + ], + "angle": 0, + "content": "
MethodRefine ScalemIoU
BEVFormer-37.84
\\( OSP^† \\)-38.20
BEVFormer w/ OSP20m × 20m38.20
BEVFormer w/ OSP40m × 40m39.35
BEVFormer w/ OSP50m × 50m40.32
BEVFormer w/ OSP80m × 80m41.21
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.541, + 0.398, + 0.555 + ], + "angle": 0, + "content": "5.3 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.696 + ], + "angle": 0, + "content": "Model architecture. We conducted ablation experiments on the model structure, particularly exploring variations in the number of transformer layers and the adaptively oversampling strategy. The results of these experiments are presented in Tab. 4. Even without an oversampling strategy and with just a 3-layer transformer, our method's baseline already surpasses the BEVFormer baseline, advancing from 37.84 to 38.01 in mIoU. The oversampling strategy further amplifies this improvement by enhancing the connections between spatial points, culminating in an increased mIoU of 38.48." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.725, + 0.787, + 0.753 + ], + "angle": 0, + "content": "Table 4: Model architecture. To make a fair comparison with the baseline, we set the number of layers to three. The oversampling strategy brings \\(0.47\\mathrm{mIoU}\\)." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.765, + 0.772, + 0.838 + ], + "angle": 0, + "content": "
Layer NumOversamplemIoU
1-36.76
3-38.01
3✓38.48
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "2D image features. The performance of our method is heavily reliant on the quality of image features. To elucidate this dependency, we conducted ablation studies focusing on the output from the model's 'neck' component. These experiments demonstrated the substantial impact of using a multi-scale output from the FPN. This architectural choice boosts the model's performance, as evidenced by the increase in mIoU from 38.67 to 39.41 as shown in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Grid center sampling method. Our ablation experiments also investigated the technique for the grid center sampling method during training. While initially using grid center points as coordinates, we introduced a variation by adding random perturbations. The results, as detailed in Tab. 5, indicate the introduction of this disturbance to the grid center points elevates the mIoU to 38.67." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Table 5: 2D image features and point perturbation. Increasing the multi-scale feature maps of images from 2 to 4 brings \\(0.74\\mathrm{mIoU}\\). We randomly applied perturbations of \\(0.1m\\) to the coordinates of each point in three directions. This strategy brings \\(0.19\\mathrm{mIoU}\\)" + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.407, + 0.772, + 0.495 + ], + "angle": 0, + "content": "
Multi-scale Images FeaturesPoint PerturbationmIoU
2-38.48
2✓38.67
4✓39.41
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Loss design. The task of 3D occupancy prediction bears similarities to segmentation tasks. In the context of the Occ3D-nuScenes dataset, there is a notable issue of class imbalance. To address this challenge, we opted to incorporate dice loss into our optimization strategy. The effectiveness of this choice is evident in the results presented in Tab. 6, which demonstrate that the integration of dice loss benefits our method." + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.634, + 0.776, + 0.649 + ], + "angle": 0, + "content": "Table 6: Loss design. Incorporating dice loss for optimization brings \\(0.53\\mathrm{mIoU}\\)." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.661, + 0.772, + 0.721 + ], + "angle": 0, + "content": "
Dice LossmIoU
-38.88
✓39.41
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Adaptive point sampling. During the inference process, we apply adaptive sampling (adaptively select points with high uncertainty) in BEVFormer w/ OSP and OSP. Adaptive sampling can reduce computational burden while maintaining good performance as shown in Tab. 7. BEVFormer w/ OSP using adaptive sampling means we refine BEVFormer by adaptively selecting points with high uncertainty (around \\(20\\%\\)). OSP using adaptive sampling means we only forward" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.612, + 0.13 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.145, + 0.687, + 0.159 + ], + "angle": 0, + "content": "Table 7: Adaptability of OSP. 'rel' denotes relative." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.172, + 0.77, + 0.279 + ], + "angle": 0, + "content": "
MethodAdaptive SamplingRel. ComputationmIoU
BEVFormer--37.84
BEVFormer w/ OSP-1.0×39.35
BEVFormer w/ OSP✓0.8×39.22
OSP-1.0×39.41
OSP✓0.93×39.08
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.785, + 0.369 + ], + "angle": 0, + "content": "points with low confidence to the next decoder layer, while directly outputting the results for points with high confidence in the decoder. Points of high uncertainty are defined as those with a confidence score less than a threshold of 0.9 after softmax." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.398, + 0.784, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Fig. 4: Visualization of our results. Our visualization results have voids compared to the ground truth ego-vehicle position, which is due to the surrounding view having gaps near the ego-vehicle." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.36, + 0.697 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we present a novel perspective on 3D scene representation, viewing it through a set of points. We introduce the innovative concept of Points of Interest (PoIs), which significantly advances the flexibility in scene representation. Building upon the foundation laid by PoIs, we develop a highly adaptable, point-based 3D occupancy prediction framework, named OSP. We validate the strong performance and flexibility of OSP on the Occ3D-nuScenes benchmark. Our work not only contributes to the field of 3D occupancy prediction but also paves the way for more dynamic and adaptable methods in 3D scene analysis." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.224 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (No. 62376102) and the Agricultural Science and Technology Independent Innovation Foundation of Jiangsu Province, China, No. CX(22)2031." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.249, + 0.323, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.393 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.449 + ], + "angle": 0, + "content": "3. Cai, Y., Chen, X., Zhang, C., Lin, K.Y., Wang, X., Li, H.: Semantic scene completion via integrating instances and scene in-the-loop. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 324-333 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.785, + 0.491 + ], + "angle": 0, + "content": "4. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3991-4001 (2022) 3, 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.785, + 0.533 + ], + "angle": 0, + "content": "5. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4193-4202 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.575 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.785, + 0.616 + ], + "angle": 0, + "content": "7. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) 2, 3, 5, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.645 + ], + "angle": 0, + "content": "8. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "9. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR. pp. 9223-9232 (2023) 3, 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.714 + ], + "angle": 0, + "content": "0. Jiang, H., Cheng, T., Gao, N., Zhang, H., Liu, W., Wang, X.: Symphonize 3d semantic scene completion with contextual instance queries. CoRR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.757 + ], + "angle": 0, + "content": "1. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3351-3359 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "2. Li, J., Liu, Y., Gong, D., Shi, Q., Yuan, X., Zhao, C., Reid, I.D.: RGBD based dimensional decomposition residual network for 3d semantic scene completion. In: CVPR. pp. 7693-7702 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Li, J., Liu, Y., Yuan, X., Zhao, C., Siegwart, R., Reid, I., Cadena, C.: Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters 5(1), 219-226 (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.611, + 0.129 + ], + "angle": 0, + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI. pp. 1486-1494 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "15. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Yu, Q., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270 (2022) 2, 3, 4, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "16. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. In: Advances in Neural Information Processing Systems. vol. 31 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "17. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "18. Liu, Y., Yan, J., Jia, F., Li, S., Gao, Q., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.785, + 0.366 + ], + "angle": 0, + "content": "19. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.785, + 0.405 + ], + "angle": 0, + "content": "20. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE international conference on robotics and automation. vol. 2, pp. 116-121. IEEE (1985) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.405, + 0.785, + 0.447 + ], + "angle": 0, + "content": "21. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 2020 International Conference on 3D Vision (3DV). pp. 111-119. IEEE (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "22. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1746-1754 (2017) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.542 + ], + "angle": 0, + "content": "23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.785, + 0.567 + ], + "angle": 0, + "content": "24. Thrun, S.: Probabilistic robotics. Communications of the ACM 45(3), 52-57 (2002) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.568, + 0.785, + 0.61 + ], + "angle": 0, + "content": "25. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving (2023) 1, 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "26. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. arXiv preprint arXiv:2104.10956 (2021) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "27. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "28. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "29. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "30. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 733-749 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Zhang, P., Liu, W., Lei, Y., Lu, H., Yang, X.: Cascaded context pyramid for full-resolution 3d semantic scene completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7801-7810 (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Occupancy as Set of Points" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "32. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316 (2023) 3, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.217 + ], + "angle": 0, + "content": "33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.246 + ], + "angle": 0, + "content": "34. Zuo, S., Zheng, W., Huang, Y., Zhou, J., Lu, J.: Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.246 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_origin.pdf b/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..406d060bcb2c1c13416dd2e1b858ba99f126ed0c --- /dev/null +++ b/2024/Occupancy as Set of Points/e88bbcd3-f64f-4fcb-a277-94c463373078_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28d1fd16b249a03c33b4bdbbb97118b327e954811203069938d78835acf4180b +size 6317318 diff --git a/2024/Occupancy as Set of Points/full.md b/2024/Occupancy as Set of Points/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50e1ea361380788b9e153bc1f54fd7b01765caf0 --- /dev/null +++ b/2024/Occupancy as Set of Points/full.md @@ -0,0 +1,268 @@ +# Occupancy as Set of Points + +Yiang Shi $^{1,\star}$ , Tianheng Cheng $^{1,\star}$ , Qian Zhang $^{2}$ , Wenyu Liu $^{1}$ , and Xinggang Wang $^{1,\text{忍}}$ + +$^{1}$ School of EIC, Huazhong University of Science & Technology + +2 Horizon Robotics + +Abstract. In this paper, we explore a novel point representation for 3D occupancy prediction from multi-view images, which is named Occupancy as Set of Points. Existing camera-based methods tend to exploit dense volume-based representation to predict the occupancy of the whole scene, making it hard to focus on the special areas or areas out of the perception range. In comparison, we present the Points of Interest (PoIs) to represent the scene and propose OSP, a novel framework for point-based 3D occupancy prediction. Owing to the inherent flexibility of the point-based representation, OSP achieves strong performance compared with existing methods and excels in terms of training and inference adaptability. It extends beyond traditional perception boundaries and can be seamlessly integrated with volume-based methods to significantly enhance their effectiveness. Experiments on the Occ3D-nuScenes occupancy benchmark show that OSP has strong performance and flexibility. Code and models are available at https://github.com/hustvl/osp. + +Keywords: 3D Occupancy Prediction $\cdot$ Autonomous Vehicles $\cdot$ Multi-view 3D Perception + +# 1 Introduction + +Holistic 3D scene understanding is crucial for autonomous driving systems, directly affecting the efficiency and accuracy of subsequent tasks. Considering the cost-effectiveness and ease of deployment of cameras compared to other sensors, developing visual-based methods for 3D scene understanding has become a significant and widely researched challenge. + +To tackle this challenge, 3D Semantic Scene Completion (SSC) [22] has been proposed and widely studied to jointly infer the geometry and semantics information of the scene from limited observations. The SSC task requires the model to accurately predict the visible locations and complete the information for the invisible locations. Recently, Occ3D [25] introduces a new task definition called 3D occupancy prediction. The main difference between this task and SSC is that 3D occupancy prediction only focuses on the visible areas and is tailored for dynamic scenes. + +![](images/c2db89918190d37d13c18068154ca7442a3a75eef145310947f2ffd263686312.jpg) +Fig. 1: Comparison between volume-based methods and our method. The volume-based methods, represented by BEVFormer, infers every region within the scene and gets standard occupancy as shown in (a). Our method uses a point-based decoder as shown in (b). Thus it infers the Points of Interest including standard, adaptively sampled, and manually sampled grids as shown in (c). + +![](images/cecde533082c1cb6a3ea0177789a83d58650a9cc44368b218888afd5ac117637.jpg) + +Existing 3D occupancy prediction methods are mostly based on dense BEV methods, e.g., BEVFormer [15], BEVDet [7]. These methods integrate a BEV encoder with an occupancy head to generate the output and enhance BEV perception capabilities for better results. However, they share some common drawbacks. (1) Uniform Sampling: BEV-based methods fail to differentiate between different areas within the same scene, treating them equally. This leads to coarse sampling and hinders dynamic or multi-resolution sampling capabilities. (2) Limited Inference Flexibility: During inference, these methods can only process the entire scene at once. They lack the ability to infer different parts of the scene based on varying downstream tasks or specific practical needs. + +Those limitations highlight the need for more flexible 3D occupancy prediction methods that can handle complex scenes while adapting to different inference requirements. In this paper, we propose a novel point-based representation for 3D occupancy prediction. Instead of dividing the scene into uniform grids applied by existing volume-based methods, we propose Points of Interest (PoIs) to view the scene as a collection of points that help in flexibly sampling the scene during both training and inference stages. Fig. 1 compares volume-based and point-based representations. Compared to volume-based representations, our point-based representation has the following advantages: (1) it can accept inputs of any scale and position to make occupancy predictions, including manually designed and adaptively designed input, offering flexibility; (2) it can pay extra attention to certain areas rather than treating all areas equally, enhancing the model's perceptual capabilities. + +We introduce Occupancy as Set of Points (OSP), a novel and flexible point-based framework, which is built upon the foundational concept of Points of Interest (PoIs). OSP excels in 3D occupancy prediction and is composed of an image backbone, a 3D positioning encoder, and a decoder, as illustrated in Fig. 2. Central to our methodology is the innovative use of PoIs, which we have categorized into three distinct types to meet diverse needs, thereby significantly enhancing various aspects of our model's performance. Detailed descriptions of these PoIs will be provided in Sec. 3. Notably, PoIs can be designed as needed beyond the three types initially proposed by us. + +Our method stands out for its strong performance and flexibility. The flexibility enables it to process any arbitrary local scene without necessitating retraining. Additionally, our method can serve as an augmentative plugin module for existing volume-based methods by adaptively resampling areas of low confidence to yield more accurate occupancy predictions. It is also adept at predicting areas beyond the scene. The key contributions of our approach can be summarized as follows: + +- A novel point-based occupancy representation, established by interacting point queries with 2D image features, enables a comprehensive understanding of 3D scenes. +- A flexible framework that allows for inference at any area of interest without retraining or sacrificing accuracy and predicts areas beyond the scene. +- A plugin module that enhances the performance of the volume-based baseline significantly. +- OSP has obtained strong experiment results, i.e., 39.4 mIoU in the 3D occupancy prediction task on the Occ3D-nuScenes benchmark. + +# 2 Related Work + +# 2.1 3D Occupancy Prediction + +3D occupancy prediction, a concept recently defined by Occ3D [25], exhibits notable parallels with Occupancy Grid Mapping (OGM) [20, 24, 27] used in robotics. This task aims to predict the state of each voxel grid in a scene based on a series of sensor inputs. Occ3D establishes two benchmarks leveraging the Waymo Open Dataset [23] and the nuScenes Dataset [2] to facilitate this. In vision-based 3D Occupancy prediction, Occ3D implements camera visibility estimation and creates visibility masks to ensure evaluations are confined to visible areas. It also evaluates various SSC methodologies on its benchmarks, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32], and BEV-Former [15]. + +# 2.2 3D Semantic Scene Completion + +Scene Semantic Completion (SSC) represents a task closely associated with 3D occupancy prediction. The concept of SSC is initially presented in SSC-Net [22], with a focus on predicting the comprehensive semantic information + +of a scene based on its partially visible regions. Over recent years, the study of SSC has expanded significantly, particularly in the context of small indoor scenes [3,5,11-13,16,30,31]. In recent times, the study of Scene Semantic Completion (SSC) for expansive outdoor environments has gained momentum, particularly following the introduction of the SemanticKITTI dataset [1]. Notably, MonoScene [4] emerges as the first method to apply monocular pure vision-based SSC. In a parallel advancement, OccDepth [19] enhances 2D to 3D feature transformation by incorporating depth data from stereo input. TPVFormer [9] argues against the limitations of single-plane modeling in capturing intricate details, hence it adopts a tri-perspective view (TPV) approach, combining a Bird's Eye View (BEV) with two additional vertical planes. Additionally, Symphonies [10] highlights the significance of instance representation in SSC tasks. While SSC methods can be directly applied to 3D occupancy prediction, two primary distinctions exist: (1) SSC primarily aims to infer the occupancy of non-visible areas based on visible regions, in contrast to 3D occupancy prediction which focuses on visible areas; (2) SSC methods usually target static scenes, whereas 3D occupancy prediction methods are often designed to handle dynamic scenes. + +Most existing volume-based SSC and 3D occupancy prediction methods are characterized by their dense nature, encompassing inputs and outputs that span the entire scene. Consider the BEVFormer baseline as an example: it segments the scene into uniform BEV grids, failing to distinguish between grids in varying areas. This uniformity restricts the ability of volume-based methods like BEVFormer to sample areas of interest for better performance during training. Besides, if we want to focus on a specific area in the inference stage, volume-based methods are limited to infer the entire scene and then perform post-processing, inevitably leading to increased and unnecessary costs. Moreover, as scene size and voxel resolution increase, the computational demands skyrocket exponentially. In stark contrast, our point-based model introduces much-needed flexibility by focusing on PoIs. Our method facilitates direct inference in specific areas, eliminating the need for post-processing and avoiding additional computational burdens. + +A point-related SSC method is PointOcc [34], a point cloud-based SSC prediction method using three complementary view planes for efficient point cloud feature modeling and an efficient 2D backbone for processing to reduce computational load, while our method focuses on the flexibility of training and inference. + +# 2.3 Camera-based 3D Detection + +Camera-based 3D perception tasks have gained substantial attention in recent research, largely due to the convenience and cost-effectiveness of cameras as data collection sensors. Initial efforts, such as FCOS3D [26] and DETR3D [28], explore the transition from 2D to 3D predictions. BEVFormer [15] represents a significant advancement in this area, transforming images captured from vehicle-mounted cameras into a bird's eye view (BEV) representation. This technique not only enhances a vehicle's environmental understanding but also finds appli + +cations in various downstream tasks like BEVStereo [14] and BEVDet [7] and extends to 3D occupancy prediction. + +3D detection tasks are important in camera-based 3D perception and also have a high similarity to 3D occupancy prediction. The primary objective of 3D detection involves estimating the position and dimensions of objects in 3D space. DETR3D [28], drawing inspiration from DETR [33], innovatively combines 3D object queries with image features, incorporating camera intrinsic and extrinsic parameters. PETR [17] and its successor, PETRv2 [18], further refine this approach by addressing the accuracy of reference point sampling and the inclusion of global information, enhancing 3D detection through historical data integration. Compared to object detection, 3D occupancy prediction offers a finer granularity, which is crucial for navigating irregular obstacles or overhanging objects. It is imperative to explore how insights from BEV representation and 3D object detection can inspire innovative solutions tailored to the specific requirements of 3D occupancy prediction. + +# 3 Preliminary + +Problem setup. We aim to provide occupancy predictions around the ego-vehicle, given only $N$ surround-view RGB images. More specifically, we use as input current images denoted by $\mathbf{I}_i = \{I_0,I_1,\dots,I_N\}$ , and use as output an occupancy prediction $\mathbf{Y}_i\in \{c_0,c_1,\dots,c_M\}^{H\times W\times Z}$ defined in the coordinate of ego-vehicle, where each occupancy prediction is either empty (denoted by $c_{0}$ ) or occupied by a certain semantic class in $\{c_{1},c_{m},\dots,c_{M}\}$ . We assume known camera intrinsic parameters $\{K_i\}$ and extrinsic parameters $\{[R_i|t_i]\}$ in each frame. We assume to know whether each area is visible or not by applying a camera visibility mask. + +Points of Interest. In our model, we innovatively introduce the concept of Points of Interest (PoIs), which is a set of sparse points to represent the 3D scene. PoIs can flexibly represent objects or regions that require extra attention such as pedestrians or regions near ego-vehicles and can be designed as needed in both training and inference phases. We use three types of PoIs and the definitions and functions are introduced as follows: + +(1) Standard Grids: By sampling center points of grids in the inference stage and making predictions of standard 3D occupancy grids, our model makes a fair comparison with existing methods and achieves good results. +(2) Adaptively Sampling: During the training stage, our model adaptively samples points and oversamples points around them to enhance accuracy. Recognizing that volume-based methods uniformly treat all locations, our point-based approach allows for resampling in either areas of special interest or those that are challenging to learn. This adaptively resampling strategy is also used to augment the performance of volume-based methods. Consequently, our method can function as a versatile plugin, seamlessly integrating with and enhancing existing volume-based approaches. + +(3) Manually Sampling: Our model excels in its flexibility to sample any area, specifically catering to the unique demands of various downstream tasks. Our model can make predictions of areas beyond the standard perception range by setting PoIs to areas outside the scene manually, e.g. 200 meters away from ego-vehicle, which is a feat unattainable by traditional volume-based methods. This extension not only broadens the scope of inference but also introduces a new dimension to scene understanding. These PoIs are the foundation of our method, offering precision and high flexibility. + +![](images/449a3e8aed3cf8a8fcbc855e54bd5977137f27256f6b3f0d14dcfd7ea7ae6ea4.jpg) +Fig. 2: Overall framework of Occupancy as Set of Points. OSP leverages the Transformer architecture to derive 3D point features from 2D images to make 3D occupancy predictions. Initially, we extract 2D features from multi-view images. Following this, we employ a set of 3D point queries to index these 2D features. The selection of these 3D point queries depends on the Points of Interest (PoIs). + +# 4 Method + +# 4.1 Overall Architecture + +Fig. 2 shows the overall architecture of the proposed Occupancy as Set of Points. Given the images $\mathbf{I} = \{I_i, i = 1, 2, \dots, N\}$ from $\mathbf{N}$ views, We feed these images into ResNet [6] to obtain their features. + +OSP begins with sampling a set of 3D points in the space, which are the initial PoIs. In our experiments, we use center points of grids as the initial PoIs to make a fair comparison with traditional volume-based methods and provide a reliable baseline for performance comparison and evaluation. We sample $K(K = 8000)$ + +points within the camera's visible region and introduce random perturbations to these points. + +Then we normalize the 3D points. To these normalized coordinates, we apply sine and cosine functions as a form of positional encoding. This encoded positional information is then utilized to create query position embeddings. In the training phase, within each decoder layer, the query position corresponding to each individual query remains consistent. The 3D coordinates, along with the camera's intrinsic and extrinsic parameters, are used to map these points onto the pixel plane. This mapping process yields corresponding key and value pairs. Subsequently, we employ point cross-attention mechanisms to compute the output. Then we adaptively oversample a group of $M$ points ( $M = 8000$ ) whose coordinates are calculated by a linear layer and employ group point cross-attention to fuse the features of the additional sampling points. + +# 4.2 3D Position Encoder + +After obtaining 3D points by applying PoIs, we first normalize coordinate points using equations as follows: + +$$ +x, y, z = \frac {x - x _ {\operatorname* {m i n}}}{x _ {\operatorname* {m a x}} - x _ {\operatorname* {m i n}}}, \frac {y - y _ {\operatorname* {m i n}}}{y _ {\operatorname* {m a x}} - y _ {\operatorname* {m i n}}}, \frac {z - z _ {\operatorname* {m i n}}}{z _ {\operatorname* {m a x}} - z _ {\operatorname* {m i n}}}, \tag {1} +$$ + +where $[x_{\min}, y_{\min}, z_{\min}, x_{\max}, y_{\max}, z_{\max}] = [-40m, -40m, -1m, 40m, 40m, 5.4m]$ , are the preset boundaries of the scene. Through sine and cosine functions, we encode the normalized coordinates into high-dimensional positional information. Then we use a small MLP containing two linear layers and one ReLU layer to transform the high-dimensional positional information into learnable embeddings. + +# 4.3 Point Decoder + +We used three decoder layers to build our point decoder. Each layer of the decoder contains point cross-attention (PCA) and group point cross-attention (GPCA). + +The purpose of PCA and GPCA is to integrate position embedding with image features, thereby facilitating a more cohesive representation. Owing to the high efficiency of deformable attention [15], our point decoder mechanisms employ this technique. Consequently, each query within our framework can be updated as follows: + +$$ +\mathrm {D A} (\mathbf {q}, \mathbf {p}, \mathbf {F}) = \sum_ {s = 1} ^ {N _ {s}} \mathbf {A} _ {s} \mathbf {W} _ {s} \mathbf {F} ^ {\mathbf {2 D}} (\mathbf {p} + \Delta \mathbf {p} _ {s}), \tag {2} +$$ + +where $N_{s}$ represents the number of sampling offsets, $\mathbf{A}_s$ represents the learnable attention weights, $\mathbf{F}^{2\mathbf{D}}(\mathbf{p} + \varDelta \mathbf{p}_s)$ represents the image features collected at location $\mathbf{p} + \varDelta \mathbf{p}_s$ in which $\varDelta\mathbf{p}_s$ represents the offset apply to position $\mathbf{p}$ . + +Point cross-attention. The point cross-attention mechanism begins by taking point queries, initially set to zero, and combines them with point position encoding derived from our 3D position encoder. This combination forms the input queries. Subsequently, these queries undergo deformable cross-attention with 2D image features. Given that not all 3D points project onto the image plane, particularly in the context of the six surround views provided by the nuScenes Dataset. Each point is likely to map onto only one or two of these views. We utilize the camera's intrinsic and extrinsic parameters to ascertain which images a given point can map to. This approach ensures that for any specific point, we only consider the features of the image(s) it maps onto, significantly reducing memory consumption. $\mathbf{F}^{2D} = \{\mathbf{F}_t^{2D},\mathbf{F}_{t - 1}^{2D},\ldots \}$ represents the mapped 2D image features where $t$ indexes the images. Therefore, the formula of the point cross-attention can be described as follows: + +$$ +\operatorname {P C A} (\mathbf {q}, \mathbf {F} ^ {2 D}) = \frac {1}{| \mathcal {V} _ {t} |} \sum_ {t \in \mathcal {V} _ {t}} \mathrm {D A} (\mathbf {q}, \mathcal {P} (\mathbf {p}, t), \mathbf {F} _ {t} ^ {2 D}), \tag {3} +$$ + +where $\mathcal{V}_t$ represents the hit image, $t$ indexes the images, $\mathcal{P}(\mathbf{p}, t)$ represents the projection function of input position $\mathbf{p}$ . + +Group point cross-attention. The group point cross-attention mechanism is aimed at solving the lack of local context in PCA since each point independently interacts with the image features. We adaptively oversample a group of $M$ ( $M = 8000$ ) points around our PoIs whose coordinates are calculated by a linear layer. We use the attention obtained from PCA and the 2D image features mapped by our group of points. Therefore, the formula of the group point cross-attention can be described as follows: + +$$ +\operatorname {G P C A} \left(\mathbf {q} _ {\mathbf {g}}, \mathbf {F} ^ {2 D}\right) = \frac {1}{\left| \mathcal {V} _ {t} \right|} \sum_ {t \in \mathcal {V} _ {t}} \mathrm {D A} \left(\mathbf {q} _ {\mathbf {g}}, \mathcal {P} \left(\mathbf {p} _ {\mathbf {g}}, t\right), \mathbf {F} _ {t} ^ {2 D}\right), \tag {4} +$$ + +where $\mathbf{q_g}$ represents attention calculated by PCA and $\mathbf{p_g}$ represents the input position of the oversampled group of points. + +# 4.4 Loss Function + +We apply class-wise cross-entropy loss and dice loss to this task. The ground truth $\hat{\mathbf{G}}_t \in \{c_0, c_1, \dots, c_M\}$ represent semantic information of a group of spatial points. The class-wise cross-entropy loss can be computed by: + +$$ +\mathcal {L} _ {c e} = - \sum_ {n = 1} ^ {N} \sum_ {c = c _ {0}} ^ {c _ {M}} w _ {c} \hat {g} _ {n, c} \log \left(\frac {e ^ {g _ {n , c}}}{\sum_ {c} e ^ {g _ {n , c}}}\right), \tag {5} +$$ + +where $n$ is the index of points, $N$ is the number of selected points, $c$ indexes class, $g_{n,c}$ is the predicted logits for the $n$ -th point belonging to class $c$ , $w_{c}$ is a weight for each class according to the inverse of the class frequency as in [21]. The dice loss can be computed by: + +$$ +\mathcal {L} _ {d i c e} = 1 - \frac {2 \sum_ {i = 1} ^ {N} p _ {i} g _ {i}}{\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \sum_ {i = 1} ^ {N} g _ {i} ^ {2}}, \tag {6} +$$ + +where $p_i$ is the predicted probability for the point and $g_i$ is the ground truth binary label for the point. Our final loss is the sum of the two losses: + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {d i c e}} + \mathcal {L} _ {\text {c e}}. \tag {7} +$$ + +# 5 Experiment + +# 5.1 Experimental Setup + +Dataset. We perform our experiments on the nuScenes Dataset with annotation provided by Occ3D. Occ3D-nuScenes benchmark is interested in a volume of $80.0\mathrm{m}$ to the front and back of the car, $80.0\mathrm{m}$ to the left and right side, and $6.4\mathrm{m}$ in height. Each sample is divided as a group of 3D voxel grids with a dimension of [200, 200, 16] since each voxel has a size of $[0.4m, 0.4m, 0.4m]$ . There are 18 categories of occupancy semantics, of which the 18th category is empty and does not participate in the evaluation. Occ3D-nuScenes provides ground-truth semantical voxel grids by aggregate points of the static scenes and moving objects, respectively. Furthermore, Occ3D-nuScenes utilizes ray-casting-based methods to estimate camera visibility and provides camera visibility masks. + +Evaluation metric. For this task, we use mIoU as the evaluation metric, which can be formulated as follows: + +$$ +m I o U = \frac {1}{C} \sum_ {c = 1} ^ {C} \frac {T P _ {c}}{T P _ {c} + F P _ {c} + F N _ {c}}. \tag {8} +$$ + +Following the definition of 3D occupancy prediction, we only evaluate our results in visible regions. + +Implementation details. We input six images of the original size into ResNet10 to obtain image features. Our image backbone is pre-trained on the FCOS3D [26], which is the same backbone used in the BEVFormer baseline. Then the features will be taken by FPN to produce multi-scale feature maps. The dimension of features is $d = 256$ . We stacked three decoder layers, with point cross-attention and group point cross-attention in each decoder layer. There is a small MLP as our point-based occupancy predictor that projects the 256 feature dimension to the number of classes. We train our model on 8 NVIDIA 3090 GPUs with 24 epochs. We use the AdamW optimizer with a learning rate of $2 \times 10^{-4}$ and a weight decay of 0.01. The learning rate of the backbone is 10 times smaller. Besides, we conducted a separate experiment in which we trained our decoder with a frozen BEVFormer baseline backbone using the same setting above and this decoder will be used as a plugin module for the BEVFormer baseline. + +Baseline methods. We compare Occupancy as Set of Points with existing methods replicated by Occ3D on Occ3D-nuScenes benchmark, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32] and BEVFormer [15]. To make a fair comparison with the baseline, we provide results using BEVFormer by our implementation using the camera visible mask during the training. + +Table 1: 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. * means the performance is achieved by our implementation using the camera mask during training. † means the performance is achieved with a frozen BEVFormer baseline backbone. + +
MethodothersbarrierbicyclebuscarCons. Vehmotorcyclepedestriantraffic conetrailertruckDri. Surother flatsidewalkterrainmammade vegetationmIoU
MonoScene [4]1.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.656.06
TPVFormer [9]7.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.7827.83
BEVDet [7]4.3930.310.2332.2634.4712.9710.3410.366.268.9323.6552.2724.6126.0622.3115.0415.1019.38
OccFormer [32]5.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.9721.93
BEVFormer [15]5.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.3028.0620.0417.6926.88
CTF-Occ [25]8.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.028.53
SurroundOcc [29]8.739.219.741.446.218.720.626.423.327.032.578.038.346.649.636.731.634.4
OpenOccupancy [27]10.445.723.642.449.314.824.627.727.827.633.379.239.847.150.537.731.836.1
BEVDet-depth [8]6.641.27.042.748.418.412.922.018.228.233.280.139.749.152.139.933.833.7
BEVDet-stereo [8]8.645.914.346.051.223.818.924.122.333.637.981.540.552.655.946.941.238.0
BEVFormer* [15]8.5446.2420.2847.4653.0419.3325.3926.1625.133.4537.7581.1738.6449.3252.5442.7336.1337.84
OSP11.247.2527.0647.5753.6623.2129.3729.6828.4132.3937.9479.3541.4650.3153.2340.5235.3939.41
\( OSP^† \)8.8746.3321.3247.5153.1419.626.1226.8426.6833.6737.9481.2139.1349.4852.7642.7336.1238.20
BEVFormer w/ \( OSP^† \)10.9549.027.6850.2455.9922.9631.0230.9130.2535.641.2382.0942.5951.955.144.8238.1741.21
+ +# 5.2 Performance + +Evaluation strategy. Our evaluation strategy for the method comprehensively utilized the three types of Points of Interest (PoIs) previously outlined. This approach entailed: (1) Standard Grids. We conducted sampling at the centers of grids, enabling us to benchmark our method's mIoU against traditional volume-based methods. This comparison provided a direct assessment of our method's performance in a standard scenario. (2) Adaptively Sampling. We adaptively oversample in our training phase to enhance our performance. Furthermore, we integrated our method as a complementary tool with volume-based approaches. By adaptively selecting PoIs that need refinement, our method enhances and augments existing volume-based techniques as shown in Fig. 3. (3) Manually Sampling. We manually select points out of the standard perception range and test out the model's perceptual ability. + +Through these varied and thorough evaluation techniques, we are able to comprehensively assess the performance and flexibility of our method across different scenarios and use cases. + +Standard grids. In Tab. 1, we benchmarked our method against existing camera-based 3D occupancy prediction techniques on the Occ3D-nuScenes benchmark by setting PoIs to standard grids. Our point-based approach achieved a notable improvement, obtains 1.57 mIoU performance gain compared to our implementation of the Bevformer baseline. Notably, our method outperformed the BEVFormer baseline across almost all category metrics, demonstrating a particularly clear advantage in detecting small targets, such as the bicycle $(20.28 \rightarrow 27.06)$ , motorcycle $(25.39 \rightarrow 29.37)$ , pedestrian $(26.16 \rightarrow 29.37)$ and traffic cone $(25.1 \rightarrow 28.31)$ . This advantage is mainly because the direct sampling of spatial points is beneficial for feature extraction and mapping of small objects. + +![](images/36c0684c0386e0388fcbe44d86f3de89e6b0d20b734825c9b09116904f526eff.jpg) +Fig. 3: Pipeline of refining volume-based methods with OSP. Given RGB images, 2D features are extracted by the frozen image backbone of the volume-based method which in our case is the BEVFormer baseline. We use the volume-based decoder to infer the entire scene, the point decoder to infer our selected 3D points, and combine the results of both using a weighted sum method. + +Adaptively Sampling. (1) The results shown in Tab. 4 demonstrate that our adaptively oversampling strategy improved the results from 38.01 to 38.48. (2) In the adaptive refinement experiment, we use our own implementation of the BEVFormer baseline as a representative of volume-based methods and enhance it with our method using different scales of PoIs by adaptively selecting locations with low confidence. Our model is trained with the frozen BEVFormer baseline backbone. The experimental results, as shown in Tab. 3, confirm that our plugin model significantly enhances BEVformer's performance. Notably, the BEVFormer w/ OSP results surpass those obtained from the two models operating independently. Furthermore, as we increase the size of the refined scene, the mIoU scores of BEVFormer w/ OSP also exhibit marked improvements, indicating that our refinement approach is very effective across various areas. + +Manually Sampling. The Occ3D-nuScenes dataset offers annotations across an $[80m\times 80m]$ range. We manually set the points and annotations within a smaller $[60m\times 60m]$ area and set PoIs to points out of it. To comprehensively evaluate our method, we conducted assessments across three distinct ranges: the standard $[60m\times 60m]$ range, the full dataset range of $[80m\times 80m]$ , and an intermediate-range spanning from $60m$ to $80m$ . Results shown in Tab. 2 demonstrate the capability of our method to make predictions at manually selected locations which is outside the predefined range in this case. + +Table 2: Predict regions beyond the scene. The model was trained on scenes within a $60m \times 60m$ range, and evaluations were conducted on the training scenes, the entire size of the scene, and the distance from $60m$ to $80m$ . + +
MethodSceneRangemIoU
OSPStandard0m ~ 60m42.29
OSPLarge0m ~ 80m29.00
OSPBeyond60m ~ 80m23.00
+ +Table 3: Refine BEVFormer with our decoder. $\dagger$ means our model is trained with a frozen BEVFormer baseline backbone. Refine scale refers to an area defined by the length multiplied by the width, centered around the self-vehicle. The entire scene spans a scale of $80m\times 80m$ . As the refine scale increases, the mIoU also gradually increases. + +
MethodRefine ScalemIoU
BEVFormer-37.84
\( OSP^† \)-38.20
BEVFormer w/ OSP20m × 20m38.20
BEVFormer w/ OSP40m × 40m39.35
BEVFormer w/ OSP50m × 50m40.32
BEVFormer w/ OSP80m × 80m41.21
+ +# 5.3 Ablation studies + +Model architecture. We conducted ablation experiments on the model structure, particularly exploring variations in the number of transformer layers and the adaptively oversampling strategy. The results of these experiments are presented in Tab. 4. Even without an oversampling strategy and with just a 3-layer transformer, our method's baseline already surpasses the BEVFormer baseline, advancing from 37.84 to 38.01 in mIoU. The oversampling strategy further amplifies this improvement by enhancing the connections between spatial points, culminating in an increased mIoU of 38.48. + +Table 4: Model architecture. To make a fair comparison with the baseline, we set the number of layers to three. The oversampling strategy brings $0.47\mathrm{mIoU}$ . + +
Layer NumOversamplemIoU
1-36.76
3-38.01
3✓38.48
+ +2D image features. The performance of our method is heavily reliant on the quality of image features. To elucidate this dependency, we conducted ablation studies focusing on the output from the model's 'neck' component. These experiments demonstrated the substantial impact of using a multi-scale output from the FPN. This architectural choice boosts the model's performance, as evidenced by the increase in mIoU from 38.67 to 39.41 as shown in Tab. 5. + +Grid center sampling method. Our ablation experiments also investigated the technique for the grid center sampling method during training. While initially using grid center points as coordinates, we introduced a variation by adding random perturbations. The results, as detailed in Tab. 5, indicate the introduction of this disturbance to the grid center points elevates the mIoU to 38.67. + +Table 5: 2D image features and point perturbation. Increasing the multi-scale feature maps of images from 2 to 4 brings $0.74\mathrm{mIoU}$ . We randomly applied perturbations of $0.1m$ to the coordinates of each point in three directions. This strategy brings $0.19\mathrm{mIoU}$ + +
Multi-scale Images FeaturesPoint PerturbationmIoU
2-38.48
2✓38.67
4✓39.41
+ +Loss design. The task of 3D occupancy prediction bears similarities to segmentation tasks. In the context of the Occ3D-nuScenes dataset, there is a notable issue of class imbalance. To address this challenge, we opted to incorporate dice loss into our optimization strategy. The effectiveness of this choice is evident in the results presented in Tab. 6, which demonstrate that the integration of dice loss benefits our method. + +Table 6: Loss design. Incorporating dice loss for optimization brings $0.53\mathrm{mIoU}$ . + +
Dice LossmIoU
-38.88
✓39.41
+ +Adaptive point sampling. During the inference process, we apply adaptive sampling (adaptively select points with high uncertainty) in BEVFormer w/ OSP and OSP. Adaptive sampling can reduce computational burden while maintaining good performance as shown in Tab. 7. BEVFormer w/ OSP using adaptive sampling means we refine BEVFormer by adaptively selecting points with high uncertainty (around $20\%$ ). OSP using adaptive sampling means we only forward + +Table 7: Adaptability of OSP. 'rel' denotes relative. + +
MethodAdaptive SamplingRel. ComputationmIoU
BEVFormer--37.84
BEVFormer w/ OSP-1.0×39.35
BEVFormer w/ OSP✓0.8×39.22
OSP-1.0×39.41
OSP✓0.93×39.08
+ +points with low confidence to the next decoder layer, while directly outputting the results for points with high confidence in the decoder. Points of high uncertainty are defined as those with a confidence score less than a threshold of 0.9 after softmax. + +![](images/2daec395d6a850edda3f9b6943bfd379555023b28f066d76d77de987a8c88d00.jpg) +Fig. 4: Visualization of our results. Our visualization results have voids compared to the ground truth ego-vehicle position, which is due to the surrounding view having gaps near the ego-vehicle. + +# 6 Conclusion + +In this work, we present a novel perspective on 3D scene representation, viewing it through a set of points. We introduce the innovative concept of Points of Interest (PoIs), which significantly advances the flexibility in scene representation. Building upon the foundation laid by PoIs, we develop a highly adaptable, point-based 3D occupancy prediction framework, named OSP. We validate the strong performance and flexibility of OSP on the Occ3D-nuScenes benchmark. Our work not only contributes to the field of 3D occupancy prediction but also paves the way for more dynamic and adaptable methods in 3D scene analysis. + +# Acknowledgements + +This work was supported by National Natural Science Foundation of China (No. 62376102) and the Agricultural Science and Technology Independent Innovation Foundation of Jiangsu Province, China, No. CX(22)2031. + +# References + +1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019) 4 +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) 3 +3. Cai, Y., Chen, X., Zhang, C., Lin, K.Y., Wang, X., Li, H.: Semantic scene completion via integrating instances and scene in-the-loop. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 324-333 (2021) 4 +4. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3991-4001 (2022) 3, 4, 9, 10 +5. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4193-4202 (2020) 4 +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 6 +7. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) 2, 3, 5, 9, 10 +8. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR (2021) 10 +9. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR. pp. 9223-9232 (2023) 3, 4, 9, 10 +0. Jiang, H., Cheng, T., Gao, N., Zhang, H., Liu, W., Wang, X.: Symphonize 3d semantic scene completion with contextual instance queries. CoRR (2023) 4 +1. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3351-3359 (2020) 4 +2. Li, J., Liu, Y., Gong, D., Shi, Q., Yuan, X., Zhao, C., Reid, I.D.: RGBD based dimensional decomposition residual network for 3d semantic scene completion. In: CVPR. pp. 7693-7702 (2019) 4 +3. Li, J., Liu, Y., Yuan, X., Zhao, C., Siegwart, R., Reid, I., Cadena, C.: Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters 5(1), 219-226 (2019) 4 + +14. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI. pp. 1486-1494 (2023) 5 +15. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Yu, Q., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270 (2022) 2, 3, 4, 7, 9, 10 +16. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. In: Advances in Neural Information Processing Systems. vol. 31 (2018) 4 +17. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625 (2022) 5 +18. Liu, Y., Yan, J., Jia, F., Li, S., Gao, Q., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022) 5 +19. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023) 4 +20. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE international conference on robotics and automation. vol. 2, pp. 116-121. IEEE (1985) 3 +21. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 2020 International Conference on 3D Vision (3DV). pp. 111-119. IEEE (2020) 8 +22. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1746-1754 (2017) 1, 3 +23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) 3 +24. Thrun, S.: Probabilistic robotics. Communications of the ACM 45(3), 52-57 (2002) 3 +25. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving (2023) 1, 3, 10 +26. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. arXiv preprint arXiv:2104.10956 (2021) 4, 9 +27. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023) 3, 10 +28. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022) 4, 5 +29. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023) 10 +30. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 733-749 (2018) 4 +31. Zhang, P., Liu, W., Lei, Y., Lu, H., Yang, X.: Cascaded context pyramid for full-resolution 3d semantic scene completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7801-7810 (2019) 4 + +32. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316 (2023) 3, 9, 10 +33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2020) 5 +34. Zuo, S., Zheng, W., Huang, Y., Zhou, J., Lu, J.: Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction (2023) 4 \ No newline at end of file diff --git a/2024/Occupancy as Set of Points/images.zip b/2024/Occupancy as Set of Points/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..135a443d8e42a3a179fed6332ad4095a30b55569 --- /dev/null +++ b/2024/Occupancy as Set of Points/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:134f20e325d1fc340c13bf460b25d27e9e5176a211b65aba2cf95ba688cbec8c +size 427941 diff --git a/2024/Occupancy as Set of Points/layout.json b/2024/Occupancy as Set of Points/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2c51f4da4c9df0fc08552bc09aec8f9f2743aa17 --- /dev/null +++ b/2024/Occupancy as Set of Points/layout.json @@ -0,0 +1,7769 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 212, + 112, + 402, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 112, + 402, + 129 + ], + "spans": [ + { + "bbox": [ + 212, + 112, + 402, + 129 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "spans": [ + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "text", + "content": "Yiang Shi" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\star}" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "text", + "content": ", Tianheng Cheng" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\star}" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "text", + "content": ", Qian Zhang" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "text", + "content": ", Wenyu Liu" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "text", + "content": ", and Xinggang Wang" + }, + { + "bbox": [ + 201, + 149, + 412, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\text{忍}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 183, + 435, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 183, + 435, + 195 + ], + "spans": [ + { + "bbox": [ + 175, + 183, + 435, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 183, + 435, + 195 + ], + "type": "text", + "content": " School of EIC, Huazhong University of Science & Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 195, + 347, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 195, + 347, + 205 + ], + "spans": [ + { + "bbox": [ + 266, + 195, + 347, + 205 + ], + "type": "text", + "content": "2 Horizon Robotics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 243, + 455, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 455, + 407 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 455, + 407 + ], + "type": "text", + "content": "Abstract. In this paper, we explore a novel point representation for 3D occupancy prediction from multi-view images, which is named Occupancy as Set of Points. Existing camera-based methods tend to exploit dense volume-based representation to predict the occupancy of the whole scene, making it hard to focus on the special areas or areas out of the perception range. In comparison, we present the Points of Interest (PoIs) to represent the scene and propose OSP, a novel framework for point-based 3D occupancy prediction. Owing to the inherent flexibility of the point-based representation, OSP achieves strong performance compared with existing methods and excels in terms of training and inference adaptability. It extends beyond traditional perception boundaries and can be seamlessly integrated with volume-based methods to significantly enhance their effectiveness. Experiments on the Occ3D-nuScenes occupancy benchmark show that OSP has strong performance and flexibility. Code and models are available at https://github.com/hustvl/osp." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 417, + 453, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 417, + 453, + 440 + ], + "spans": [ + { + "bbox": [ + 159, + 417, + 453, + 440 + ], + "type": "text", + "content": "Keywords: 3D Occupancy Prediction " + }, + { + "bbox": [ + 159, + 417, + 453, + 440 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 417, + 453, + 440 + ], + "type": "text", + "content": " Autonomous Vehicles " + }, + { + "bbox": [ + 159, + 417, + 453, + 440 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 417, + 453, + 440 + ], + "type": "text", + "content": " Multi-view 3D Perception" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 458, + 230, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 230, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 230, + 471 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 481, + 540 + ], + "type": "text", + "content": "Holistic 3D scene understanding is crucial for autonomous driving systems, directly affecting the efficiency and accuracy of subsequent tasks. Considering the cost-effectiveness and ease of deployment of cameras compared to other sensors, developing visual-based methods for 3D scene understanding has become a significant and widely researched challenge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 541, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 636 + ], + "type": "text", + "content": "To tackle this challenge, 3D Semantic Scene Completion (SSC) [22] has been proposed and widely studied to jointly infer the geometry and semantics information of the scene from limited observations. The SSC task requires the model to accurately predict the visible locations and complete the information for the invisible locations. Recently, Occ3D [25] introduces a new task definition called 3D occupancy prediction. The main difference between this task and SSC is that 3D occupancy prediction only focuses on the visible areas and is tailored for dynamic scenes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 479, + 665 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author: Xinggang Wang (xgwang@hust.edu.cn)." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 349, + 286 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 349, + 286 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 349, + 286 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 349, + 286 + ], + "type": "image", + "image_path": "c2db89918190d37d13c18068154ca7442a3a75eef145310947f2ffd263686312.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 354 + ], + "type": "text", + "content": "Fig. 1: Comparison between volume-based methods and our method. The volume-based methods, represented by BEVFormer, infers every region within the scene and gets standard occupancy as shown in (a). Our method uses a point-based decoder as shown in (b). Thus it infers the Points of Interest including standard, adaptively sampled, and manually sampled grids as shown in (c)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 352, + 118, + 477, + 286 + ], + "blocks": [ + { + "bbox": [ + 352, + 118, + 477, + 286 + ], + "lines": [ + { + "bbox": [ + 352, + 118, + 477, + 286 + ], + "spans": [ + { + "bbox": [ + 352, + 118, + 477, + 286 + ], + "type": "image", + "image_path": "cecde533082c1cb6a3ea0177789a83d58650a9cc44368b218888afd5ac117637.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 385, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 506 + ], + "type": "text", + "content": "Existing 3D occupancy prediction methods are mostly based on dense BEV methods, e.g., BEVFormer [15], BEVDet [7]. These methods integrate a BEV encoder with an occupancy head to generate the output and enhance BEV perception capabilities for better results. However, they share some common drawbacks. (1) Uniform Sampling: BEV-based methods fail to differentiate between different areas within the same scene, treating them equally. This leads to coarse sampling and hinders dynamic or multi-resolution sampling capabilities. (2) Limited Inference Flexibility: During inference, these methods can only process the entire scene at once. They lack the ability to infer different parts of the scene based on varying downstream tasks or specific practical needs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Those limitations highlight the need for more flexible 3D occupancy prediction methods that can handle complex scenes while adapting to different inference requirements. In this paper, we propose a novel point-based representation for 3D occupancy prediction. Instead of dividing the scene into uniform grids applied by existing volume-based methods, we propose Points of Interest (PoIs) to view the scene as a collection of points that help in flexibly sampling the scene during both training and inference stages. Fig. 1 compares volume-based and point-based representations. Compared to volume-based representations, our point-based representation has the following advantages: (1) it can accept inputs of any scale and position to make occupancy predictions, including manually designed and adaptively designed input, offering flexibility; (2) it can pay extra attention to certain areas rather than treating all areas equally, enhancing the model's perceptual capabilities." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "We introduce Occupancy as Set of Points (OSP), a novel and flexible point-based framework, which is built upon the foundational concept of Points of Interest (PoIs). OSP excels in 3D occupancy prediction and is composed of an image backbone, a 3D positioning encoder, and a decoder, as illustrated in Fig. 2. Central to our methodology is the innovative use of PoIs, which we have categorized into three distinct types to meet diverse needs, thereby significantly enhancing various aspects of our model's performance. Detailed descriptions of these PoIs will be provided in Sec. 3. Notably, PoIs can be designed as needed beyond the three types initially proposed by us." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 307 + ], + "type": "text", + "content": "Our method stands out for its strong performance and flexibility. The flexibility enables it to process any arbitrary local scene without necessitating retraining. Additionally, our method can serve as an augmentative plugin module for existing volume-based methods by adaptively resampling areas of low confidence to yield more accurate occupancy predictions. It is also adept at predicting areas beyond the scene. The key contributions of our approach can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 313, + 479, + 417 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 313, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 479, + 346 + ], + "type": "text", + "content": "- A novel point-based occupancy representation, established by interacting point queries with 2D image features, enables a comprehensive understanding of 3D scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 348, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 348, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 137, + 348, + 479, + 370 + ], + "type": "text", + "content": "- A flexible framework that allows for inference at any area of interest without retraining or sacrificing accuracy and predicts areas beyond the scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 372, + 479, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 372, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 137, + 372, + 479, + 394 + ], + "type": "text", + "content": "- A plugin module that enhances the performance of the volume-based baseline significantly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 395, + 479, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 395, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 137, + 395, + 479, + 417 + ], + "type": "text", + "content": "- OSP has obtained strong experiment results, i.e., 39.4 mIoU in the 3D occupancy prediction task on the Occ3D-nuScenes benchmark." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 435, + 237, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 237, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 237, + 448 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 458, + 290, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 290, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 290, + 470 + ], + "type": "text", + "content": "2.1 3D Occupancy Prediction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": "3D occupancy prediction, a concept recently defined by Occ3D [25], exhibits notable parallels with Occupancy Grid Mapping (OGM) [20, 24, 27] used in robotics. This task aims to predict the state of each voxel grid in a scene based on a series of sensor inputs. Occ3D establishes two benchmarks leveraging the Waymo Open Dataset [23] and the nuScenes Dataset [2] to facilitate this. In vision-based 3D Occupancy prediction, Occ3D implements camera visibility estimation and creates visibility masks to ensure evaluations are confined to visible areas. It also evaluates various SSC methodologies on its benchmarks, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32], and BEV-Former [15]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 612, + 319, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 319, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 319, + 624 + ], + "type": "text", + "content": "2.2 3D Semantic Scene Completion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Scene Semantic Completion (SSC) represents a task closely associated with 3D occupancy prediction. The concept of SSC is initially presented in SSC-Net [22], with a focus on predicting the comprehensive semantic information" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "of a scene based on its partially visible regions. Over recent years, the study of SSC has expanded significantly, particularly in the context of small indoor scenes [3,5,11-13,16,30,31]. In recent times, the study of Scene Semantic Completion (SSC) for expansive outdoor environments has gained momentum, particularly following the introduction of the SemanticKITTI dataset [1]. Notably, MonoScene [4] emerges as the first method to apply monocular pure vision-based SSC. In a parallel advancement, OccDepth [19] enhances 2D to 3D feature transformation by incorporating depth data from stereo input. TPVFormer [9] argues against the limitations of single-plane modeling in capturing intricate details, hence it adopts a tri-perspective view (TPV) approach, combining a Bird's Eye View (BEV) with two additional vertical planes. Additionally, Symphonies [10] highlights the significance of instance representation in SSC tasks. While SSC methods can be directly applied to 3D occupancy prediction, two primary distinctions exist: (1) SSC primarily aims to infer the occupancy of non-visible areas based on visible regions, in contrast to 3D occupancy prediction which focuses on visible areas; (2) SSC methods usually target static scenes, whereas 3D occupancy prediction methods are often designed to handle dynamic scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 320, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 486 + ], + "type": "text", + "content": "Most existing volume-based SSC and 3D occupancy prediction methods are characterized by their dense nature, encompassing inputs and outputs that span the entire scene. Consider the BEVFormer baseline as an example: it segments the scene into uniform BEV grids, failing to distinguish between grids in varying areas. This uniformity restricts the ability of volume-based methods like BEVFormer to sample areas of interest for better performance during training. Besides, if we want to focus on a specific area in the inference stage, volume-based methods are limited to infer the entire scene and then perform post-processing, inevitably leading to increased and unnecessary costs. Moreover, as scene size and voxel resolution increase, the computational demands skyrocket exponentially. In stark contrast, our point-based model introduces much-needed flexibility by focusing on PoIs. Our method facilitates direct inference in specific areas, eliminating the need for post-processing and avoiding additional computational burdens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 488, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 536 + ], + "type": "text", + "content": "A point-related SSC method is PointOcc [34], a point cloud-based SSC prediction method using three complementary view planes for efficient point cloud feature modeling and an efficient 2D backbone for processing to reduce computational load, while our method focuses on the flexibility of training and inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 558, + 303, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 303, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 303, + 569 + ], + "type": "text", + "content": "2.3 Camera-based 3D Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Camera-based 3D perception tasks have gained substantial attention in recent research, largely due to the convenience and cost-effectiveness of cameras as data collection sensors. Initial efforts, such as FCOS3D [26] and DETR3D [28], explore the transition from 2D to 3D predictions. BEVFormer [15] represents a significant advancement in this area, transforming images captured from vehicle-mounted cameras into a bird's eye view (BEV) representation. This technique not only enhances a vehicle's environmental understanding but also finds appli" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "cations in various downstream tasks like BEVStereo [14] and BEVDet [7] and extends to 3D occupancy prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "type": "text", + "content": "3D detection tasks are important in camera-based 3D perception and also have a high similarity to 3D occupancy prediction. The primary objective of 3D detection involves estimating the position and dimensions of objects in 3D space. DETR3D [28], drawing inspiration from DETR [33], innovatively combines 3D object queries with image features, incorporating camera intrinsic and extrinsic parameters. PETR [17] and its successor, PETRv2 [18], further refine this approach by addressing the accuracy of reference point sampling and the inclusion of global information, enhancing 3D detection through historical data integration. Compared to object detection, 3D occupancy prediction offers a finer granularity, which is crucial for navigating irregular obstacles or overhanging objects. It is imperative to explore how insights from BEV representation and 3D object detection can inspire innovative solutions tailored to the specific requirements of 3D occupancy prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 320, + 225, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 225, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 225, + 335 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": "Problem setup. We aim to provide occupancy predictions around the ego-vehicle, given only " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": " surround-view RGB images. More specifically, we use as input current images denoted by " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i = \\{I_0,I_1,\\dots,I_N\\}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": ", and use as output an occupancy prediction " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_i\\in \\{c_0,c_1,\\dots,c_M\\}^{H\\times W\\times Z}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": " defined in the coordinate of ego-vehicle, where each occupancy prediction is either empty (denoted by " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": ") or occupied by a certain semantic class in " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\{c_{1},c_{m},\\dots,c_{M}\\}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": ". We assume known camera intrinsic parameters " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\{K_i\\}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": " and extrinsic parameters " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\{[R_i|t_i]\\}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": " in each frame. We assume to know whether each area is visible or not by applying a camera visibility mask." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 530 + ], + "type": "text", + "content": "Points of Interest. In our model, we innovatively introduce the concept of Points of Interest (PoIs), which is a set of sparse points to represent the 3D scene. PoIs can flexibly represent objects or regions that require extra attention such as pedestrians or regions near ego-vehicles and can be designed as needed in both training and inference phases. We use three types of PoIs and the definitions and functions are introduced as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 533, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 567 + ], + "type": "text", + "content": "(1) Standard Grids: By sampling center points of grids in the inference stage and making predictions of standard 3D occupancy grids, our model makes a fair comparison with existing methods and achieves good results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "(2) Adaptively Sampling: During the training stage, our model adaptively samples points and oversamples points around them to enhance accuracy. Recognizing that volume-based methods uniformly treat all locations, our point-based approach allows for resampling in either areas of special interest or those that are challenging to learn. This adaptively resampling strategy is also used to augment the performance of volume-based methods. Consequently, our method can function as a versatile plugin, seamlessly integrating with and enhancing existing volume-based approaches." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "(3) Manually Sampling: Our model excels in its flexibility to sample any area, specifically catering to the unique demands of various downstream tasks. Our model can make predictions of areas beyond the standard perception range by setting PoIs to areas outside the scene manually, e.g. 200 meters away from ego-vehicle, which is a feat unattainable by traditional volume-based methods. This extension not only broadens the scope of inference but also introduces a new dimension to scene understanding. These PoIs are the foundation of our method, offering precision and high flexibility." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 236, + 479, + 422 + ], + "blocks": [ + { + "bbox": [ + 138, + 236, + 479, + 422 + ], + "lines": [ + { + "bbox": [ + 138, + 236, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 479, + 422 + ], + "type": "image", + "image_path": "449a3e8aed3cf8a8fcbc855e54bd5977137f27256f6b3f0d14dcfd7ea7ae6ea4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 432, + 482, + 489 + ], + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 489 + ], + "type": "text", + "content": "Fig. 2: Overall framework of Occupancy as Set of Points. OSP leverages the Transformer architecture to derive 3D point features from 2D images to make 3D occupancy predictions. Initially, we extract 2D features from multi-view images. Following this, we employ a set of 3D point queries to index these 2D features. The selection of these 3D point queries depends on the Points of Interest (PoIs)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 533, + 202, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 202, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 202, + 544 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 559, + 264, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 264, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 264, + 571 + ], + "type": "text", + "content": "4.1 Overall Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "text", + "content": "Fig. 2 shows the overall architecture of the proposed Occupancy as Set of Points. Given the images " + }, + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{I_i, i = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "text", + "content": " views, We feed these images into ResNet [6] to obtain their features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "OSP begins with sampling a set of 3D points in the space, which are the initial PoIs. In our experiments, we use center points of grids as the initial PoIs to make a fair comparison with traditional volume-based methods and provide a reliable baseline for performance comparison and evaluation. We sample " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "K(K = 8000)" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 374, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 374, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 374, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "points within the camera's visible region and introduce random perturbations to these points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": "Then we normalize the 3D points. To these normalized coordinates, we apply sine and cosine functions as a form of positional encoding. This encoded positional information is then utilized to create query position embeddings. In the training phase, within each decoder layer, the query position corresponding to each individual query remains consistent. The 3D coordinates, along with the camera's intrinsic and extrinsic parameters, are used to map these points onto the pixel plane. This mapping process yields corresponding key and value pairs. Subsequently, we employ point cross-attention mechanisms to compute the output. Then we adaptively oversample a group of " + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": " points (" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "inline_equation", + "content": "M = 8000" + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": ") whose coordinates are calculated by a linear layer and employ group point cross-attention to fuse the features of the additional sampling points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 289, + 265, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 265, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 265, + 300 + ], + "type": "text", + "content": "4.2 3D Position Encoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 479, + 334 + ], + "type": "text", + "content": "After obtaining 3D points by applying PoIs, we first normalize coordinate points using equations as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 343, + 481, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 343, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 201, + 343, + 481, + 366 + ], + "type": "interline_equation", + "content": "x, y, z = \\frac {x - x _ {\\operatorname* {m i n}}}{x _ {\\operatorname* {m a x}} - x _ {\\operatorname* {m i n}}}, \\frac {y - y _ {\\operatorname* {m i n}}}{y _ {\\operatorname* {m a x}} - y _ {\\operatorname* {m i n}}}, \\frac {z - z _ {\\operatorname* {m i n}}}{z _ {\\operatorname* {m a x}} - z _ {\\operatorname* {m i n}}}, \\tag {1}", + "image_path": "b8e1c2d78a34a826823752d80a5a970193c927193808fda785862c7c53323836.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 374, + 488, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 488, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 488, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 374, + 488, + 446 + ], + "type": "inline_equation", + "content": "[x_{\\min}, y_{\\min}, z_{\\min}, x_{\\max}, y_{\\max}, z_{\\max}] = [-40m, -40m, -1m, 40m, 40m, 5.4m]" + }, + { + "bbox": [ + 130, + 374, + 488, + 446 + ], + "type": "text", + "content": ", are the preset boundaries of the scene. Through sine and cosine functions, we encode the normalized coordinates into high-dimensional positional information. Then we use a small MLP containing two linear layers and one ReLU layer to transform the high-dimensional positional information into learnable embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 464, + 234, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 234, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 234, + 475 + ], + "type": "text", + "content": "4.3 Point Decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 483, + 479, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 479, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 479, + 519 + ], + "type": "text", + "content": "We used three decoder layers to build our point decoder. Each layer of the decoder contains point cross-attention (PCA) and group point cross-attention (GPCA)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 520, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 582 + ], + "type": "text", + "content": "The purpose of PCA and GPCA is to integrate position embedding with image features, thereby facilitating a more cohesive representation. Owing to the high efficiency of deformable attention [15], our point decoder mechanisms employ this technique. Consequently, each query within our framework can be updated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 220, + 589, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 220, + 589, + 481, + 622 + ], + "type": "interline_equation", + "content": "\\mathrm {D A} (\\mathbf {q}, \\mathbf {p}, \\mathbf {F}) = \\sum_ {s = 1} ^ {N _ {s}} \\mathbf {A} _ {s} \\mathbf {W} _ {s} \\mathbf {F} ^ {\\mathbf {2 D}} (\\mathbf {p} + \\Delta \\mathbf {p} _ {s}), \\tag {2}", + "image_path": "86ae191c576922428e4c421d767eef230daf0c598698e88e8fd0946fa1a4db6b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " represents the number of sampling offsets, " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " represents the learnable attention weights, " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{2\\mathbf{D}}(\\mathbf{p} + \\varDelta \\mathbf{p}_s)" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " represents the image features collected at location " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{p} + \\varDelta \\mathbf{p}_s" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\varDelta\\mathbf{p}_s" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " represents the offset apply to position " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "Point cross-attention. The point cross-attention mechanism begins by taking point queries, initially set to zero, and combines them with point position encoding derived from our 3D position encoder. This combination forms the input queries. Subsequently, these queries undergo deformable cross-attention with 2D image features. Given that not all 3D points project onto the image plane, particularly in the context of the six surround views provided by the nuScenes Dataset. Each point is likely to map onto only one or two of these views. We utilize the camera's intrinsic and extrinsic parameters to ascertain which images a given point can map to. This approach ensures that for any specific point, we only consider the features of the image(s) it maps onto, significantly reducing memory consumption. " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{2D} = \\{\\mathbf{F}_t^{2D},\\mathbf{F}_{t - 1}^{2D},\\ldots \\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " represents the mapped 2D image features where " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " indexes the images. Therefore, the formula of the point cross-attention can be described as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 274, + 481, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 274, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 219, + 274, + 481, + 299 + ], + "type": "interline_equation", + "content": "\\operatorname {P C A} (\\mathbf {q}, \\mathbf {F} ^ {2 D}) = \\frac {1}{| \\mathcal {V} _ {t} |} \\sum_ {t \\in \\mathcal {V} _ {t}} \\mathrm {D A} (\\mathbf {q}, \\mathcal {P} (\\mathbf {p}, t), \\mathbf {F} _ {t} ^ {2 D}), \\tag {3}", + "image_path": "07c05e849a850faca656343afd054ab9b794fa147d5f9a9a99c848e5d7649ad2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_t" + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "text", + "content": " represents the hit image, " + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "text", + "content": " indexes the images, " + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathbf{p}, t)" + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "text", + "content": " represents the projection function of input position " + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 304, + 481, + 328 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 336, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 482, + 419 + ], + "type": "text", + "content": "Group point cross-attention. The group point cross-attention mechanism is aimed at solving the lack of local context in PCA since each point independently interacts with the image features. We adaptively oversample a group of " + }, + { + "bbox": [ + 131, + 336, + 482, + 419 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 336, + 482, + 419 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 131, + 336, + 482, + 419 + ], + "type": "inline_equation", + "content": "M = 8000" + }, + { + "bbox": [ + 131, + 336, + 482, + 419 + ], + "type": "text", + "content": ") points around our PoIs whose coordinates are calculated by a linear layer. We use the attention obtained from PCA and the 2D image features mapped by our group of points. Therefore, the formula of the group point cross-attention can be described as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 422, + 481, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 422, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 211, + 422, + 481, + 447 + ], + "type": "interline_equation", + "content": "\\operatorname {G P C A} \\left(\\mathbf {q} _ {\\mathbf {g}}, \\mathbf {F} ^ {2 D}\\right) = \\frac {1}{\\left| \\mathcal {V} _ {t} \\right|} \\sum_ {t \\in \\mathcal {V} _ {t}} \\mathrm {D A} \\left(\\mathbf {q} _ {\\mathbf {g}}, \\mathcal {P} \\left(\\mathbf {p} _ {\\mathbf {g}}, t\\right), \\mathbf {F} _ {t} ^ {2 D}\\right), \\tag {4}", + "image_path": "82db58513e41867e2f40ade8c6029aa7addc199c6ef6d21432cfe1a57dbce77c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 452, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 452, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 452, + 481, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 452, + 481, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{q_g}" + }, + { + "bbox": [ + 131, + 452, + 481, + 475 + ], + "type": "text", + "content": " represents attention calculated by PCA and " + }, + { + "bbox": [ + 131, + 452, + 481, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{p_g}" + }, + { + "bbox": [ + 131, + 452, + 481, + 475 + ], + "type": "text", + "content": " represents the input position of the oversampled group of points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 490, + 231, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 231, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 231, + 501 + ], + "type": "text", + "content": "4.4 Loss Function" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 507, + 481, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 507, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 131, + 507, + 481, + 543 + ], + "type": "text", + "content": "We apply class-wise cross-entropy loss and dice loss to this task. The ground truth " + }, + { + "bbox": [ + 131, + 507, + 481, + 543 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}_t \\in \\{c_0, c_1, \\dots, c_M\\}" + }, + { + "bbox": [ + 131, + 507, + 481, + 543 + ], + "type": "text", + "content": " represent semantic information of a group of spatial points. The class-wise cross-entropy loss can be computed by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 548, + 481, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 548, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 225, + 548, + 481, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = - \\sum_ {n = 1} ^ {N} \\sum_ {c = c _ {0}} ^ {c _ {M}} w _ {c} \\hat {g} _ {n, c} \\log \\left(\\frac {e ^ {g _ {n , c}}}{\\sum_ {c} e ^ {g _ {n , c}}}\\right), \\tag {5}", + "image_path": "45f61db54b36ded09b6826cd7008acecf29831b2e2dd32921a9074ba6c045e50.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": " is the index of points, " + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": " is the number of selected points, " + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": " indexes class, " + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "inline_equation", + "content": "g_{n,c}" + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": " is the predicted logits for the " + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": "-th point belonging to class " + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 131, + 586, + 482, + 633 + ], + "type": "text", + "content": " is a weight for each class according to the inverse of the class frequency as in [21]. The dice loss can be computed by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 236, + 638, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 638, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 236, + 638, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i c e} = 1 - \\frac {2 \\sum_ {i = 1} ^ {N} p _ {i} g _ {i}}{\\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} g _ {i} ^ {2}}, \\tag {6}", + "image_path": "a0d9b55eeba5a52c03923e6cec1e66b064f3fe8fab58bacdf79cd6730c0a232d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the predicted probability for the point and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the ground truth binary label for the point. Our final loss is the sum of the two losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 152, + 481, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 152, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 266, + 152, + 481, + 163 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {d i c e}} + \\mathcal {L} _ {\\text {c e}}. \\tag {7}", + "image_path": "b79c2fe58cd6af0f4f10752eb8f21878cffe8e64191d06eed1ef57826fa73920.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 224, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 224, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 224, + 197 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 210, + 262, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 210, + 262, + 223 + ], + "spans": [ + { + "bbox": [ + 132, + 210, + 262, + 223 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "text", + "content": "Dataset. We perform our experiments on the nuScenes Dataset with annotation provided by Occ3D. Occ3D-nuScenes benchmark is interested in a volume of " + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "inline_equation", + "content": "80.0\\mathrm{m}" + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "text", + "content": " to the front and back of the car, " + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "inline_equation", + "content": "80.0\\mathrm{m}" + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "text", + "content": " to the left and right side, and " + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "inline_equation", + "content": "6.4\\mathrm{m}" + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "text", + "content": " in height. Each sample is divided as a group of 3D voxel grids with a dimension of [200, 200, 16] since each voxel has a size of " + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "inline_equation", + "content": "[0.4m, 0.4m, 0.4m]" + }, + { + "bbox": [ + 130, + 232, + 482, + 352 + ], + "type": "text", + "content": ". There are 18 categories of occupancy semantics, of which the 18th category is empty and does not participate in the evaluation. Occ3D-nuScenes provides ground-truth semantical voxel grids by aggregate points of the static scenes and moving objects, respectively. Furthermore, Occ3D-nuScenes utilizes ray-casting-based methods to estimate camera visibility and provides camera visibility masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 352, + 481, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 375 + ], + "type": "text", + "content": "Evaluation metric. For this task, we use mIoU as the evaluation metric, which can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 384, + 481, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 384, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 230, + 384, + 481, + 416 + ], + "type": "interline_equation", + "content": "m I o U = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\frac {T P _ {c}}{T P _ {c} + F P _ {c} + F N _ {c}}. \\tag {8}", + "image_path": "ff1bf28cbf2689fb5e9fa33e52032d553baa5245df91564beff34bba0de0f4e3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 425, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 481, + 449 + ], + "type": "text", + "content": "Following the definition of 3D occupancy prediction, we only evaluate our results in visible regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 483, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 483, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 483, + 605 + ], + "type": "text", + "content": "Implementation details. We input six images of the original size into ResNet10 to obtain image features. Our image backbone is pre-trained on the FCOS3D [26], which is the same backbone used in the BEVFormer baseline. Then the features will be taken by FPN to produce multi-scale feature maps. The dimension of features is " + }, + { + "bbox": [ + 130, + 449, + 483, + 605 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 130, + 449, + 483, + 605 + ], + "type": "text", + "content": ". We stacked three decoder layers, with point cross-attention and group point cross-attention in each decoder layer. There is a small MLP as our point-based occupancy predictor that projects the 256 feature dimension to the number of classes. We train our model on 8 NVIDIA 3090 GPUs with 24 epochs. We use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 449, + 483, + 605 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 449, + 483, + 605 + ], + "type": "text", + "content": " and a weight decay of 0.01. The learning rate of the backbone is 10 times smaller. Besides, we conducted a separate experiment in which we trained our decoder with a frozen BEVFormer baseline backbone using the same setting above and this decoder will be used as a plugin module for the BEVFormer baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 493, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 493, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 493, + 666 + ], + "type": "text", + "content": "Baseline methods. We compare Occupancy as Set of Points with existing methods replicated by Occ3D on Occ3D-nuScenes benchmark, including MonoScene [4], TPVFormer [9], BEVDet [7], OccFormer [32] and BEVFormer [15]. To make a fair comparison with the baseline, we provide results using BEVFormer by our implementation using the camera visible mask during the training." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 168, + 481, + 302 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. * means the performance is achieved by our implementation using the camera mask during training. † means the performance is achieved with a frozen BEVFormer baseline backbone." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 168, + 481, + 302 + ], + "lines": [ + { + "bbox": [ + 136, + 168, + 481, + 302 + ], + "spans": [ + { + "bbox": [ + 136, + 168, + 481, + 302 + ], + "type": "table", + "html": "
MethodothersbarrierbicyclebuscarCons. Vehmotorcyclepedestriantraffic conetrailertruckDri. Surother flatsidewalkterrainmammade vegetationmIoU
MonoScene [4]1.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.656.06
TPVFormer [9]7.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.7827.83
BEVDet [7]4.3930.310.2332.2634.4712.9710.3410.366.268.9323.6552.2724.6126.0622.3115.0415.1019.38
OccFormer [32]5.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.9721.93
BEVFormer [15]5.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.3028.0620.0417.6926.88
CTF-Occ [25]8.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.028.53
SurroundOcc [29]8.739.219.741.446.218.720.626.423.327.032.578.038.346.649.636.731.634.4
OpenOccupancy [27]10.445.723.642.449.314.824.627.727.827.633.379.239.847.150.537.731.836.1
BEVDet-depth [8]6.641.27.042.748.418.412.922.018.228.233.280.139.749.152.139.933.833.7
BEVDet-stereo [8]8.645.914.346.051.223.818.924.122.333.637.981.540.552.655.946.941.238.0
BEVFormer* [15]8.5446.2420.2847.4653.0419.3325.3926.1625.133.4537.7581.1738.6449.3252.5442.7336.1337.84
OSP11.247.2527.0647.5753.6623.2129.3729.6828.4132.3937.9479.3541.4650.3153.2340.5235.3939.41
\\( OSP^† \\)8.8746.3321.3247.5153.1419.626.1226.8426.6833.6737.9481.2139.1349.4852.7642.7336.1238.20
BEVFormer w/ \\( OSP^† \\)10.9549.027.6850.2455.9922.9631.0230.9130.2535.641.2382.0942.5951.955.144.8238.1741.21
", + "image_path": "50c385c2004b4f13d970befbd1a6e55ef338c1968a1530ba8150724e2a76613d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 328, + 224, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 328, + 224, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 224, + 339 + ], + "type": "text", + "content": "5.2 Performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 496 + ], + "type": "text", + "content": "Evaluation strategy. Our evaluation strategy for the method comprehensively utilized the three types of Points of Interest (PoIs) previously outlined. This approach entailed: (1) Standard Grids. We conducted sampling at the centers of grids, enabling us to benchmark our method's mIoU against traditional volume-based methods. This comparison provided a direct assessment of our method's performance in a standard scenario. (2) Adaptively Sampling. We adaptively oversample in our training phase to enhance our performance. Furthermore, we integrated our method as a complementary tool with volume-based approaches. By adaptively selecting PoIs that need refinement, our method enhances and augments existing volume-based techniques as shown in Fig. 3. (3) Manually Sampling. We manually select points out of the standard perception range and test out the model's perceptual ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 532 + ], + "type": "text", + "content": "Through these varied and thorough evaluation techniques, we are able to comprehensively assess the performance and flexibility of our method across different scenarios and use cases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Standard grids. In Tab. 1, we benchmarked our method against existing camera-based 3D occupancy prediction techniques on the Occ3D-nuScenes benchmark by setting PoIs to standard grids. Our point-based approach achieved a notable improvement, obtains 1.57 mIoU performance gain compared to our implementation of the Bevformer baseline. Notably, our method outperformed the BEVFormer baseline across almost all category metrics, demonstrating a particularly clear advantage in detecting small targets, such as the bicycle " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "(20.28 \\rightarrow 27.06)" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": ", motorcycle " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "(25.39 \\rightarrow 29.37)" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": ", pedestrian " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "(26.16 \\rightarrow 29.37)" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": " and traffic cone " + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "inline_equation", + "content": "(25.1 \\rightarrow 28.31)" + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": ". This advantage is mainly because the direct sampling of spatial points is beneficial for feature extraction and mapping of small objects." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 373, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 373, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 373, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 477, + 266 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 477, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 477, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 477, + 266 + ], + "type": "image", + "image_path": "36c0684c0386e0388fcbe44d86f3de89e6b0d20b734825c9b09116904f526eff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "type": "text", + "content": "Fig. 3: Pipeline of refining volume-based methods with OSP. Given RGB images, 2D features are extracted by the frozen image backbone of the volume-based method which in our case is the BEVFormer baseline. We use the volume-based decoder to infer the entire scene, the point decoder to infer our selected 3D points, and combine the results of both using a weighted sum method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 402, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 547 + ], + "type": "text", + "content": "Adaptively Sampling. (1) The results shown in Tab. 4 demonstrate that our adaptively oversampling strategy improved the results from 38.01 to 38.48. (2) In the adaptive refinement experiment, we use our own implementation of the BEVFormer baseline as a representative of volume-based methods and enhance it with our method using different scales of PoIs by adaptively selecting locations with low confidence. Our model is trained with the frozen BEVFormer baseline backbone. The experimental results, as shown in Tab. 3, confirm that our plugin model significantly enhances BEVformer's performance. Notably, the BEVFormer w/ OSP results surpass those obtained from the two models operating independently. Furthermore, as we increase the size of the refined scene, the mIoU scores of BEVFormer w/ OSP also exhibit marked improvements, indicating that our refinement approach is very effective across various areas." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Manually Sampling. The Occ3D-nuScenes dataset offers annotations across an " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "[80m\\times 80m]" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " range. We manually set the points and annotations within a smaller " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "[60m\\times 60m]" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " area and set PoIs to points out of it. To comprehensively evaluate our method, we conducted assessments across three distinct ranges: the standard " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "[60m\\times 60m]" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " range, the full dataset range of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "[80m\\times 80m]" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", and an intermediate-range spanning from " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "60m" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "80m" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". Results shown in Tab. 2 demonstrate the capability of our method to make predictions at manually selected locations which is outside the predefined range in this case." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 157, + 472, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Predict regions beyond the scene. The model was trained on scenes within a " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "60m \\times 60m" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " range, and evaluations were conducted on the training scenes, the entire size of the scene, and the distance from " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "60m" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "80m" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 157, + 472, + 216 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 472, + 216 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 472, + 216 + ], + "type": "table", + "html": "
MethodSceneRangemIoU
OSPStandard0m ~ 60m42.29
OSPLarge0m ~ 80m29.00
OSPBeyond60m ~ 80m23.00
", + "image_path": "8bbfee116e7cf4b097259fec1ed671207cf93a6613fd5e32ce76c6115962dc49.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 141, + 304, + 473, + 399 + ], + "blocks": [ + { + "bbox": [ + 132, + 239, + 481, + 293 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 293 + ], + "type": "text", + "content": "Table 3: Refine BEVFormer with our decoder. " + }, + { + "bbox": [ + 132, + 239, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 239, + 481, + 293 + ], + "type": "text", + "content": " means our model is trained with a frozen BEVFormer baseline backbone. Refine scale refers to an area defined by the length multiplied by the width, centered around the self-vehicle. The entire scene spans a scale of " + }, + { + "bbox": [ + 132, + 239, + 481, + 293 + ], + "type": "inline_equation", + "content": "80m\\times 80m" + }, + { + "bbox": [ + 132, + 239, + 481, + 293 + ], + "type": "text", + "content": ". As the refine scale increases, the mIoU also gradually increases." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 304, + 473, + 399 + ], + "lines": [ + { + "bbox": [ + 141, + 304, + 473, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 473, + 399 + ], + "type": "table", + "html": "
MethodRefine ScalemIoU
BEVFormer-37.84
\\( OSP^† \\)-38.20
BEVFormer w/ OSP20m × 20m38.20
BEVFormer w/ OSP40m × 40m39.35
BEVFormer w/ OSP50m × 50m40.32
BEVFormer w/ OSP80m × 80m41.21
", + "image_path": "aab992bdbd0e93a61c94cc6b7ba69f52a63779f218261d2763b6a00628f32ccd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 428, + 243, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 243, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 243, + 439 + ], + "type": "text", + "content": "5.3 Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 454, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 551 + ], + "type": "text", + "content": "Model architecture. We conducted ablation experiments on the model structure, particularly exploring variations in the number of transformer layers and the adaptively oversampling strategy. The results of these experiments are presented in Tab. 4. Even without an oversampling strategy and with just a 3-layer transformer, our method's baseline already surpasses the BEVFormer baseline, advancing from 37.84 to 38.01 in mIoU. The oversampling strategy further amplifies this improvement by enhancing the connections between spatial points, culminating in an increased mIoU of 38.48." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 140, + 605, + 472, + 663 + ], + "blocks": [ + { + "bbox": [ + 132, + 574, + 481, + 596 + ], + "lines": [ + { + "bbox": [ + 132, + 574, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 481, + 596 + ], + "type": "text", + "content": "Table 4: Model architecture. To make a fair comparison with the baseline, we set the number of layers to three. The oversampling strategy brings " + }, + { + "bbox": [ + 132, + 574, + 481, + 596 + ], + "type": "inline_equation", + "content": "0.47\\mathrm{mIoU}" + }, + { + "bbox": [ + 132, + 574, + 481, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 605, + 472, + 663 + ], + "lines": [ + { + "bbox": [ + 140, + 605, + 472, + 663 + ], + "spans": [ + { + "bbox": [ + 140, + 605, + 472, + 663 + ], + "type": "table", + "html": "
Layer NumOversamplemIoU
1-36.76
3-38.01
3✓38.48
", + "image_path": "395d7db5240c3b60a7adf2004cc391f5761fb87d6923110e80fef5c8501d010e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 374, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "2D image features. The performance of our method is heavily reliant on the quality of image features. To elucidate this dependency, we conducted ablation studies focusing on the output from the model's 'neck' component. These experiments demonstrated the substantial impact of using a multi-scale output from the FPN. This architectural choice boosts the model's performance, as evidenced by the increase in mIoU from 38.67 to 39.41 as shown in Tab. 5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 248 + ], + "type": "text", + "content": "Grid center sampling method. Our ablation experiments also investigated the technique for the grid center sampling method during training. While initially using grid center points as coordinates, we introduced a variation by adding random perturbations. The results, as detailed in Tab. 5, indicate the introduction of this disturbance to the grid center points elevates the mIoU to 38.67." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 322, + 472, + 392 + ], + "blocks": [ + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "text", + "content": "Table 5: 2D image features and point perturbation. Increasing the multi-scale feature maps of images from 2 to 4 brings " + }, + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "inline_equation", + "content": "0.74\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "text", + "content": ". We randomly applied perturbations of " + }, + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "text", + "content": " to the coordinates of each point in three directions. This strategy brings " + }, + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "inline_equation", + "content": "0.19\\mathrm{mIoU}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 322, + 472, + 392 + ], + "lines": [ + { + "bbox": [ + 140, + 322, + 472, + 392 + ], + "spans": [ + { + "bbox": [ + 140, + 322, + 472, + 392 + ], + "type": "table", + "html": "
Multi-scale Images FeaturesPoint PerturbationmIoU
2-38.48
2✓38.67
4✓39.41
", + "image_path": "e2ff541d2001225b1ba0d73a3841ac34d69ad963bde736e3d933586e8f5b2dd6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 483 + ], + "type": "text", + "content": "Loss design. The task of 3D occupancy prediction bears similarities to segmentation tasks. In the context of the Occ3D-nuScenes dataset, there is a notable issue of class imbalance. To address this challenge, we opted to incorporate dice loss into our optimization strategy. The effectiveness of this choice is evident in the results presented in Tab. 6, which demonstrate that the integration of dice loss benefits our method." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 140, + 523, + 472, + 571 + ], + "blocks": [ + { + "bbox": [ + 138, + 502, + 474, + 514 + ], + "lines": [ + { + "bbox": [ + 138, + 502, + 474, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 474, + 514 + ], + "type": "text", + "content": "Table 6: Loss design. Incorporating dice loss for optimization brings " + }, + { + "bbox": [ + 138, + 502, + 474, + 514 + ], + "type": "inline_equation", + "content": "0.53\\mathrm{mIoU}" + }, + { + "bbox": [ + 138, + 502, + 474, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 523, + 472, + 571 + ], + "lines": [ + { + "bbox": [ + 140, + 523, + 472, + 571 + ], + "spans": [ + { + "bbox": [ + 140, + 523, + 472, + 571 + ], + "type": "table", + "html": "
Dice LossmIoU
-38.88
✓39.41
", + "image_path": "9dc705170cae3f24b6e9187006045f42d77cba9a825d974a7bf4d6f2240e600e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Adaptive point sampling. During the inference process, we apply adaptive sampling (adaptively select points with high uncertainty) in BEVFormer w/ OSP and OSP. Adaptive sampling can reduce computational burden while maintaining good performance as shown in Tab. 7. BEVFormer w/ OSP using adaptive sampling means we refine BEVFormer by adaptively selecting points with high uncertainty (around " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "). OSP using adaptive sampling means we only forward" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 136, + 471, + 220 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 420, + 125 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 420, + 125 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 420, + 125 + ], + "type": "text", + "content": "Table 7: Adaptability of OSP. 'rel' denotes relative." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 136, + 471, + 220 + ], + "lines": [ + { + "bbox": [ + 140, + 136, + 471, + 220 + ], + "spans": [ + { + "bbox": [ + 140, + 136, + 471, + 220 + ], + "type": "table", + "html": "
MethodAdaptive SamplingRel. ComputationmIoU
BEVFormer--37.84
BEVFormer w/ OSP-1.0×39.35
BEVFormer w/ OSP✓0.8×39.22
OSP-1.0×39.41
OSP✓0.93×39.08
", + "image_path": "2a3dbc240ab4eda7f604e8326eb92905785f7c7805511f3d6f5385d74a92d5d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 480, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 480, + 292 + ], + "type": "text", + "content": "points with low confidence to the next decoder layer, while directly outputting the results for points with high confidence in the decoder. Points of high uncertainty are defined as those with a confidence score less than a threshold of 0.9 after softmax." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 315, + 479, + 445 + ], + "blocks": [ + { + "bbox": [ + 137, + 315, + 479, + 445 + ], + "lines": [ + { + "bbox": [ + 137, + 315, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 137, + 315, + 479, + 445 + ], + "type": "image", + "image_path": "2daec395d6a850edda3f9b6943bfd379555023b28f066d76d77de987a8c88d00.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 455, + 482, + 489 + ], + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 489 + ], + "type": "text", + "content": "Fig. 4: Visualization of our results. Our visualization results have voids compared to the ground truth ego-vehicle position, which is due to the surrounding view having gaps near the ego-vehicle." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 539, + 220, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 220, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 220, + 552 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this work, we present a novel perspective on 3D scene representation, viewing it through a set of points. We introduce the innovative concept of Points of Interest (PoIs), which significantly advances the flexibility in scene representation. Building upon the foundation laid by PoIs, we develop a highly adaptable, point-based 3D occupancy prediction framework, named OSP. We validate the strong performance and flexibility of OSP on the Occ3D-nuScenes benchmark. Our work not only contributes to the field of 3D occupancy prediction but also paves the way for more dynamic and adaptable methods in 3D scene analysis." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 374, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 374, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 374, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 177 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (No. 62376102) and the Agricultural Science and Technology Independent Innovation Foundation of Jiangsu Province, China, No. CX(22)2031." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 267 + ], + "type": "text", + "content": "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 311 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 355 + ], + "type": "text", + "content": "3. Cai, Y., Chen, X., Zhang, C., Lin, K.Y., Wang, X., Li, H.: Semantic scene completion via integrating instances and scene in-the-loop. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 324-333 (2021) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "type": "text", + "content": "4. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3991-4001 (2022) 3, 4, 9, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 389, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 480, + 422 + ], + "type": "text", + "content": "5. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4193-4202 (2020) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "type": "text", + "content": "7. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) 2, 3, 5, 9, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "text", + "content": "8. Huang, J., Huang, G., Zhu, Z., Du, D.: Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR (2021) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "text", + "content": "9. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR. pp. 9223-9232 (2023) 3, 4, 9, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "text", + "content": "0. Jiang, H., Cheng, T., Gao, N., Zhang, H., Liu, W., Wang, X.: Symphonize 3d semantic scene completion with contextual instance queries. CoRR (2023) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 599 + ], + "type": "text", + "content": "1. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3351-3359 (2020) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "2. Li, J., Liu, Y., Gong, D., Shi, Q., Yuan, X., Zhao, C., Reid, I.D.: RGBD based dimensional decomposition residual network for 3d semantic scene completion. In: CVPR. pp. 7693-7702 (2019) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "3. Li, J., Liu, Y., Yuan, X., Zhao, C., Siegwart, R., Reid, I., Cadena, C.: Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters 5(1), 219-226 (2019) 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI. pp. 1486-1494 (2023) 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "15. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Yu, Q., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270 (2022) 2, 3, 4, 7, 9, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "16. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. In: Advances in Neural Information Processing Systems. vol. 31 (2018) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "type": "text", + "content": "17. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625 (2022) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "text", + "content": "18. Liu, Y., Yan, J., Jia, F., Li, S., Gao, Q., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "type": "text", + "content": "19. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 320 + ], + "type": "text", + "content": "20. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE international conference on robotics and automation. vol. 2, pp. 116-121. IEEE (1985) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 320, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 320, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 480, + 354 + ], + "type": "text", + "content": "21. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 2020 International Conference on 3D Vision (3DV). pp. 111-119. IEEE (2020) 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 480, + 386 + ], + "type": "text", + "content": "22. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1746-1754 (2017) 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 386, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 429 + ], + "type": "text", + "content": "23. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 480, + 449 + ], + "type": "text", + "content": "24. Thrun, S.: Probabilistic robotics. Communications of the ACM 45(3), 52-57 (2002) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 449, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 480, + 483 + ], + "type": "text", + "content": "25. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving (2023) 1, 3, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 480, + 504 + ], + "type": "text", + "content": "26. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. arXiv preprint arXiv:2104.10956 (2021) 4, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "type": "text", + "content": "27. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991 (2023) 3, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "text", + "content": "28. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022) 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "type": "text", + "content": "29. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551 (2023) 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "text", + "content": "30. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 733-749 (2018) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "31. Zhang, P., Liu, W., Lei, Y., Lu, H., Yang, X.: Cascaded context pyramid for full-resolution 3d semantic scene completion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7801-7810 (2019) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 373, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 373, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 373, + 102 + ], + "type": "text", + "content": "Y. Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 194 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "32. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316 (2023) 3, 9, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "text", + "content": "33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2020) 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 194 + ], + "type": "text", + "content": "34. Zuo, S., Zheng, W., Huang, Y., Zhou, J., Lu, J.: Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction (2023) 4" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 102 + ], + "type": "text", + "content": "Occupancy as Set of Points" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_content_list.json b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..511765ddf01a428517e6ef4ebf0313995a19350e --- /dev/null +++ b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_content_list.json @@ -0,0 +1,1784 @@ +[ + { + "type": "text", + "text": "Octopus: Embodied Vision-Language Programmer from Environmental Feedback", + "text_level": 1, + "bbox": [ + 254, + 141, + 750, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingkang Yang\\*,1, Yuhao Dong\\*,2,3, Shuai Liu\\*,2,4, Bo Li\\*,1, Ziyue Wang\\*,1, Haoran Tan\\*,4, Chencheng Jiang\\*,5, Jiamu Kang\\*,3, Yuanhan Zhang\\*, Kaiyang Zhou\\* and and Ziwei Liu\\*", + "bbox": [ + 256, + 210, + 743, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai AI Laboratory $^{3}$ Tsinghua University $^{4}$ BUPT $^{5}$ XJTU $^{6}$ Hong Kong Baptist University $\\{\\text{jingkang001, ziwei.liu}\\}@\\text{ntu.edu.sg}$", + "bbox": [ + 215, + 268, + 785, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e1fffbae234e11e03cf9b934aa2f24e859d1db42aff5e8724d5471c83bc1950.jpg", + "image_caption": [ + "Fig. 1: Illustration of our vision-language programmer, Octopus, complete a task in GTA environment. Given a task in the form of natural language, Octopus relies on its egocentric vision to generate plans and the corresponding executable code." + ], + "image_footnote": [], + "bbox": [ + 218, + 319, + 782, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large vision-language models (VLMs) have achieved substantial progress in multimodal perception and reasoning. When integrated into an embodied agent, existing embodied VLM works either output detailed action sequences at the manipulation level or only provide plans at an abstract level, leaving a gap between high-level planning and real-world manipulation. To bridge this gap, we introduce Octopus, an embodied vision-language programmer that uses executable code generation as a medium to connect planning and manipulation. Octopus is designed to 1) proficiently comprehend an agent's visual and textual task objectives, 2) formulate intricate action sequences, and 3) generate executable code. To facilitate Octopus model development, we introduce OctoVerse: a suite of environments tailored for benchmarking", + "bbox": [ + 259, + 648, + 743, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution, †Equal engineering contribution, ⌿ Corresponding author.", + "bbox": [ + 217, + 824, + 751, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vision-based code generators on a wide spectrum of tasks, ranging from mundane daily chores in simulators to sophisticated interactions in complex video games such as Grand Theft Auto (GTA) and Minecraft. To train Octopus, we leverage GPT-4 to control an explorative agent that generates training data, i.e., action blueprints and corresponding executable code. We also collect feedback that enables an enhanced training scheme called Reinforcement Learning with Environmental Feedback (RLEF). Through a series of experiments, we demonstrate Octopus's functionality and present compelling results, showing that the proposed RLEF refines the agent's decision-making. By open-sourcing our simulation environments, dataset, and model architecture, we aspire to ignite further innovation and foster collaborative applications within the broader embodied AI community. The project page is available at https://choiszt.github.io/0ctopus/.", + "bbox": [ + 259, + 147, + 743, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 368, + 375, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rise of large language models (LLMs) [11,15,42,46,54] led to a surge in vision-language models (VLMs) [4,5,31,33,35-37], enabling tasks such as image/video-based descriptions [33], reasoning [14,38,60], and conversations [16,31]. In the realm of embodied AI, notable efforts [10,18,33] have trained agents to process visual input and relay motor control commands.", + "bbox": [ + 212, + 401, + 785, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another approach to interacting with the environment focuses on task execution through code invocations, mirroring the human System-I stimulation [19,29] (automatic, intuitive actions) with predefined code, and leaving the System-II processes [19,29] (planning and reasoning) for large models. For example, referring to Fig. 1, planning a car ride with a pet might entail a subconscious checklist (e.g., getOutOf() the house, open() the car door), each action could be implemented using specific techniques [8,24] such as imitation learning [22,28]. This programmatic paradigm has been, although not in vision, leveraged by works [25, 50, 52, 53] using LLMs to craft programs and trigger APIs. Game-centric models like Voyager [55] have similarly employed GPT for function calls within game engines, though they often parse data directly from their environments.", + "bbox": [ + 212, + 477, + 787, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, when incorporating visual perception, the programming paradigms are largely unexplored. Primary initiatives [48,58] can only output plans, which anchor their strategies only in initial environmental states or employ dynamic scene graphs for LLM inputs, respectively. Despite their innovations, the overreliance on pre-trained vision models to convert vision content into language can occasionally hinder the LLM's planning performance. The conversion from plans into real-world actions is still missing. While EmbodiedGPT [41] addresses the problem by integrating vision-language modeling for planning and then transitioning to manipulation using policy mapping, the capability of embodied vision-language models to generate executable programs remains largely uncharted.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our exploration aims to bridge this gap. An embodied vision-language programmer should integrate visual perspective with textual objectives to devise", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "action plans and executable code (Fig. 1). However, existing simulator environments often lack the carefully designed functions necessary to support such models effectively. These functions should balance usefulness and complexity to avoid hindering the development of genuine embodied vision-language programmers. For example, the **explore_until()** function in Minecraft, which can lead the player directly to specific blocks without relying on vision information, may not be suitable for training these models.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To meet the requirement, we carefully design and develop OctoVerse, a suite of environments consisting of diverse simulators, including (i) OctoGibson, built upon the photorealistic OmniGibson [32], (ii) OctoMC, developed on the infinitely creative, pixel-style Minecraft platform [2], and (iii) OctoGTA, adapted from the highly interactive and immersive Grand Theft Auto V (GTA-V) [1]. These environments enable the training and benchmarking of our embodied vision-language programming model in a wide range of scenarios, from daily household tasks to complex urban navigation and open-world exploration, while the function calls are tailored to be vision-dependent.", + "bbox": [ + 212, + 252, + 784, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the OctoVerse environment, we train Octopus by leveraging GPT-4 to collect data. We provide GPT-4 with system messages, environmental cues, and objectives, enabling it to formulate action strategies and code. Simultaneously, the agent captures visual perspectives, forming the image-code pair for Octopus training. During data collection, the agent receives simulator feedback, distinguishing successful moves from unsuccessful ones. We incorporate this feedback using Reinforcement Learning with Environmental Feedback (RLEF) and fine-tune Octopus using Proximal Policy Optimization (PPO) [51]. Empirically, Octopus demonstrates strong adaptability in various scenarios, outperforming existing models in task planning, code generation, and execution. The integration of RLEF further enhances Octopus's performance, showcasing the effectiveness of this training approach. In sum, our key contributions include:", + "bbox": [ + 212, + 388, + 784, + 568 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A Novel Vision-Language Programming Benchmark: Three diverse embodied environments with designed tasks: (i) OctoGibson, which is developed upon OmniGibson [32], (ii) OctoMC that developed on Minecraft [2], and (iii) OctoGTA, which is adapted from GTA-V [1].", + "- A New Vision-Language Programming Model: An embodied vision-language planner and programmer trained with Reinforcement Learning with Environmental Feedback (RLEF), demonstrates compelling results.", + "- Insights on Vision-Language Programming: We extensively explore Octopus and share useful insights facilitating future research on visual planning and programming." + ], + "bbox": [ + 225, + 578, + 785, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 752, + 387, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Embodied AI Simulators", + "text_level": 1, + "bbox": [ + 215, + 784, + 467, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Embodied AI has advanced significantly with the development of diverse simulation environments, enabling research tasks such as visual exploration [47],", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c5542a1160ed522d15337fac866fb91a9da23c4f1225a38aa8ecc7ceb793c5d6.jpg", + "table_caption": [ + "Table 1: Related Work for OctoVerse - Overview of Embodied AI Environments. We select three environments into OctoVerse and carefully design executable tasks and vision-dependent function calls (VC), in comparison to undesigned standard function calls (C)." + ], + "table_footnote": [], + "table_body": "
SimulatorKinematicsContinuous \nExtended \nStatesFlexible \nMaterialsDeformable \nBodiesRealistic \nAction \nExecutionGame- or \nWorld-BasedFormulated \nTasksFunction \nCall \nType
OpenAIGym [9]✓×××✓G×C
Matterport3D [12]×××××W××
AI2THOR [30]✓××××G×C
VirtualHome [44]×××××G××
House3D [57]×××××W××
Habitat 1.0 [49]✓×××✓W×C
Robosuite [66]✓×××✓W×C
RFUniverse [21]✓×✓✓✓W×C
Minecraft [2]✓×✓×✓G×C
OctoMC✓×✓×✓G✓VC
GTA [1]✓✓✓✓✓G×C
OctoGTA✓✓✓✓✓G✓VC
OmniGibson [32]✓✓✓✓✓W×C
OctoGibson✓✓✓✓✓W✓VC
", + "bbox": [ + 217, + 203, + 785, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "navigation [56], and question-answering [17]. Several simulators, including AI2-THOR [30], VirtualHome [44], Habitat-Sim [49], SAPIEN [59], and Omnigibson [32], provide realistic representations of the world for investigating embodied AI challenges. OmniGibson [32] stands out for its high-fidelity simulation of diverse indoor and outdoor environments. OctoGibson environment further enhances OmniGibson with carefully designed function calls and formulated tasks, making it well-suited for vision-language programming.", + "bbox": [ + 212, + 385, + 785, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Game-related simulators like Arade [7], CHALET [61], and VRKitchen [23] also contribute significantly to embodied AI. Minecraft [2] has gained attention in reinforcement learning and game agents [6,20,39,55,63,64] but lacks the necessary structure for vision-language programming. OctoMC addresses this by providing designed function calls and formulated tasks. In contrast to Minecraft's voxel-based representations that limit transferability to real-world environments, GTA-V [1] offers a highly realistic environment. In this work, we introduce OctoGTA as a new setting, leveraging GTA-V's rich, open-world environment with incorporated tasks and function calls, extending this platform for embodied AI study.", + "bbox": [ + 212, + 491, + 785, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Embodied AI with Large Models", + "text_level": 1, + "bbox": [ + 215, + 664, + 532, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The recent wave of research focuses on merging LLMs with embodied AI tasks [11, 42,46,54]. For instance, VoxPoser addresses robotic manipulation problems through unsupervised methods [27]. A group of projects, namely SayCan [3], Palm-E [18], RT-2 [10], and EmbodiedGPT [41], effectively integrate visual or linguistic cues with robot manipulation data. Outside the domain of robotic manipulation, initiatives like Voyager [55] and Smallville [43] harness the capabilities of GPT to interface with game functions, relying on preset functions to manage intricate manipulations. In a parallel vein, VisProg [25] leverages GPT-3 language prompts to craft Python programs, opening the door to a multitude of fascinating applications. While the proposed Octopus model also formulates plans", + "bbox": [ + 212, + 688, + 799, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/cf730938763d2c71668e4f6703713e8b17853f2b0bf5c5ec1b5df819afe6c5e4.jpg", + "table_caption": [ + "Table 2: Related Work for Octopus - Overviewing Embodied AI Models. The proposed Octopus distinguishes itself from other models as a unified vision-language model for both plan and code generation." + ], + "table_footnote": [], + "table_body": "
ModelsRelease DateSupported EnvironmentVision ModelCode GeneratorAction w/ FeedbackLLM Training Enabled
Text2Motion [34]Mar. 2023Sim×✓×
Instruct2Act [26]May 2023Sim×✓×
Lang2Rewards [62]Jun. 2023Sim×✓×
VoxPoser [27]Jul. 2023Sim✓××
SayCan [3]Apr. 2022Real✓××
PALM-E [18]Mar. 2023Sim, Real✓×✓
RT-2 [10]Jul. 2023Real✓×✓
SayPlan [48]Jun. 2023Real×××
EmbodiedGPT [41]May 2023Sim✓×✓
TaPA [58]Jul. 2023Sim××✓
Voyager [55]May 2023Game×✓×
Steve-Eye [64]Dec 2023Game✓×✓
RoboScript [13]Feb 2024Sim, Real×✓×
Octopus-Sim, Game✓✓✓
", + "bbox": [ + 218, + 191, + 787, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and code, its distinguishing feature is the seamless integration of visual input in program and code generation. This also stands in contrast to other embodied planners like TAPA [58] and SayPlan [48], which deploy separate vision modules to translate visual data into linguistic inputs for LLMs. Octopus excels as a cohesive vision-language model, delivering not just plans but also code.", + "bbox": [ + 212, + 422, + 782, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More discussion on additional related works (Vision Language Model, and Feedback in Large Language Models) is included in the supplementary materials.", + "bbox": [ + 214, + 500, + 782, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The OctoVerse Environment", + "text_level": 1, + "bbox": [ + 215, + 558, + 527, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce three simulator environments designed to train and evaluate the Octopus model. For each environment, we will describe their overall information, the special design considerations that ensure the tasks are well-formulated and the callable functions are vision-dependent.", + "bbox": [ + 212, + 595, + 784, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "OctoGibson We built the environment on the foundation of OmniGibson [32], an existing simulation framework that supports 1,000 daily activities across 50 scenes, featuring over 5,000 meticulously annotated objects. We incorporated 16 functions that the robot can execute, such as moveBot() and easyGrasp(). Within this environment, we meticulously crafted 476 tasks1, each with a well-defined initial state and a definitive termination state, allowing for a straightforward assessment of task completion. Among these tasks, 367 are routine tasks—simple and direct actions like \"place a glass in a trash can,\" marked as Follow. Conversely, the remaining 109 are reasoning tasks that necessitate deeper comprehension. An example is \"buy a chocolate,\" where the agent needs", + "bbox": [ + 212, + 661, + 787, + 813 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 The full list of tasks and their categories are listed in the supplementary material.", + "bbox": [ + 217, + 824, + 777, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to know to pick a chocolate bar from the shelf and then place it, along with money, on the checkout counter, denoted as Reason tasks.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure our tasks are vision-aware, we deliberately constrain the usage of certain functions, such as moveBot(object), which moves the agent in front of the given object. To avoid making the task too easy and vision-agnostic, we limit the given parameter to a predefined set of large objects, such as tables and cupboards, rather than small items like cups and glasses. In this case, if the robot wants to pick up a cup, it needs to recognize whether the cup is on the table or in the cupboard. A simple moveBot(object) call with an inappropriate parameter would cause a runtime error. The full list of the functions is in the Appendix.", + "bbox": [ + 212, + 178, + 787, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OctoMC The OctoMC environment is built on Minecraft [2], a popular platform for reinforcement learning and game agents. We integrated 6 functional actions and crafted 40 tasks $^{2}$ , each designed to facilitate comprehensive observations and executions by the agent. These tasks are distributed across 10 different biomes, including indoor, outdoor, and underground settings, under varying weather conditions.", + "bbox": [ + 212, + 320, + 787, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, existing Minecraft environments often lack vision-formulated tasks and the necessary structure for vision-language programming. For example, in Voyager [55], the exploreUntil() function allows the player to navigate directly to specific blocks without relying on vision information. This function works by randomly exploring the environment within a certain range until the desired object is found, at which point it returns a true value, enabling the agent to interact with the located object, such as gold blocks or trees. While effective, this approach is entirely automated and does not utilize visual information, making it unsuitable for our vision-based objectives.", + "bbox": [ + 212, + 412, + 787, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this limitation, we crafted a vision-dependent exploration function, teleport(yaw, distance), which operates within the robot's perceptual range. This function ensures that the agent's operations are vision-dependent and require active perception and navigation within the environment. For instance, to locate a tree block, the agent must actively navigate towards the direction of the forest, relying on visual cues to guide its exploration.", + "bbox": [ + 212, + 549, + 787, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OctoGTA The OctoGTA environment is built on GTA-V [1] with the help of the active GTA modding community. We integrated 19 functions and methodically crafted 25 tasks3, such as \"help NPC drive their boat back to shore\" and \"mediate a fight between two NPCs,\" spanning across 5 groups (shown in Fig. 2 (b)). Each task is assigned to 5 different locations within the game world.", + "bbox": [ + 212, + 645, + 787, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have implemented a set of functions that enable the agent to interact with the game world in a visually-aware manner. Similar to the design in Minecraft, we get rid of functions like walkTo location) that might trivialize the task of reaching a particular building or landmark. Instead, we provide functions such", + "bbox": [ + 212, + 723, + 787, + 782 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2 Detailed tasks are listed in the supplementary material.", + "bbox": [ + 217, + 796, + 602, + 811 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3 We meticulously designed tasks to be friendly, ensuring they exclude any inappropriate or violent behaviors.", + "bbox": [ + 218, + 811, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76f8463ebcdb002f43f190e5dcf9ce9a13839fdab317d9567151f0cc97fd2181.jpg", + "image_caption": [ + "Fig. 2: The Statistics of the OctoVerse Environment with Function Designs." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 781, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as goForward(distance) and turnPlayer(degree) (Fig. 2 (b)). For tasks like mediating a fight between two NPCs, the essential function stopFight() only works when the player is within 5 meters of the fighting NPCs. These design choices ensure that the agent's operations are vision-dependent and require active perception and navigation within the environment.", + "bbox": [ + 212, + 337, + 784, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Octopus: The Embodied Vision-Language Programmer", + "text_level": 1, + "bbox": [ + 215, + 431, + 781, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present the procedure for training Octopus. Starting from collecting training data within the OctoVerse environment, Octopus builds upon a VLM architecture of Otter [31] and includes specialized RLEF modules to handle vision-language programming tasks. Fig. 4 illustrates the entire Octopus training pipeline.", + "bbox": [ + 212, + 460, + 784, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Training Data Collection", + "text_level": 1, + "bbox": [ + 215, + 556, + 465, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the automatic training data collection pipeline described here for OctoGibson and OctoMC, with the latter using customized prompts inspired by Voyager [55]. For OctoGTA, we rely on human labor to hand-craft the training dataset due to the difficulty of obtaining textual environment messages in the GTA environment. In the following parts, we use OctoGibson as the primary example to illustrate the data collection pipeline. Note that the primary task in organizing training data is to form a succinct pairing: \"vision input + current/historical states $\\rightarrow$ next step plan + executable code\".", + "bbox": [ + 212, + 577, + 784, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Environment Info Collection As shown in Fig. 4 (a) and Fig. 3, we format an environment message for each state, encompassing attributes like Observed Objects, Observed Relations, Inventory, and more. Specifically, the simulator can provide us with an exact scene graph at each state, shaping the content for the first two parts. The inventory info is also accessible. The task, e.g., \"cooking bacon\" in Fig. 3, is represented by the Task Goal.", + "bbox": [ + 212, + 700, + 784, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Automation with GPT-4 After preparing the environment message, we crafted a structured system message to ensure that the robot not only understands its input but also maintains a consistent output format. A detailed", + "bbox": [ + 212, + 795, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6cdd976d7b1070035e63f9f0eb63fef5fd12684a6160ac0c712a8b05aaf85f52.jpg", + "image_caption": [ + "Fig.3: Data Collection Example for \"Cook a Bacon\" Task. GPT-4 perceives the environment through the environmental message and produces anticipated plans and code following the detailed system message. This code is subsequently executed in the simulator, directing the agent to the subsequent state. For each state, we gather the environmental message, wherein observed objects and relations are substituted by egocentric images to serve as the training input. The response from GPT-4 acts as the training output. Environmental feedback, specifically the determination of whether each target state is met, is documented for RLEF training." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 779, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "examination of this prompt can be found in the appendix. Experiments have shown that a well-articulated prompt enables GPT-4 to effectively generate executable code. It is important to note that the combined length of the system and environment messages can be extremely long, which may cause standard GPT-4 8K models to struggle with producing meaningful outputs. To address this issue, we employ the more robust GPT-4 32K model. As illustrated in Fig. 3, when GPT-4 receives a consistent system and environment message, it generates comprehensive outputs that include current scenario analysis, planning, and actionable code, supporting the training process in Section 4.3.", + "bbox": [ + 212, + 513, + 787, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Error Management Notably, GPT-4 collects training data under the main task of guiding the agent to complete tasks. However, GPT-4 is not infallible. Errors can manifest in multiple ways, ranging from syntax errors to physical challenges in the simulator. For instance, in Fig. 3, between states $\\#5$ and $\\#6$ , the action failed due to the long distance between the agent (holding bacon) and the pan. Such setbacks reset the task to its previous state. If a main task remains incomplete after 10 steps, it is deemed unsuccessful, and we terminate this task for budget concerns. However, all data pairs without syntax errors, regardless of the task's completion status, are valuable for refining instructions and improving the model's performance.", + "bbox": [ + 212, + 654, + 787, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Environmental Feedback GPT-4's continual trial-and-error approach while guiding the agent toward task completion serves a dual purpose: collecting vision-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7e886cf6a620a9c98069660b102aace976d3cf7ca9f6b7a9e3ddf03c6685308.jpg", + "image_caption": [ + "(a) Data Collection Pipeline" + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 454, + 215 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d0ff0b0f34aadd20b48a71ac7fdbaec50a85b8bdbcf92072cdeb9a257db8ea84.jpg", + "image_caption": [ + "(b) Octopus Training Pipeline" + ], + "image_footnote": [], + "bbox": [ + 230, + 236, + 455, + 272 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2b87a40202ac8da67bda6e6d35469e9fe011aa49fa4927b481cae1065f0d6289.jpg", + "image_caption": [ + "(c) Octopus (Otter model)" + ], + "image_footnote": [], + "bbox": [ + 464, + 146, + 635, + 267 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/41f47cd6d088c512b2b154383379f89ea6c4edf682b2cf99beebf2161af803f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 146, + 784, + 202 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/65b407683e8cb947939a9954f9b85688b2bf07e5c7b7867d83395650594fc61d.jpg", + "image_caption": [ + "Step 2: Training Reward Model" + ], + "image_footnote": [], + "bbox": [ + 640, + 213, + 779, + 246 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d0606d05feeaff5577056bd7d0673252451a64569f06890ba05703f7fdf75430.jpg", + "image_caption": [ + "(d) RLEF Training", + "Fig. 4: How to train Octopus: data collection and training pipeline." + ], + "image_footnote": [], + "bbox": [ + 647, + 247, + 779, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "output pairs and generating a rich set of feedback data. The automatic annotation of this feedback focuses on two levels: step-level and task-level judgments. Step-level judgment assesses the alignment of post-execution states with their target states. For instance, in Fig. 3, steps color-coded in green lead to positive feedback. One can visualize the action sequence for task completion as a tree, where each node indicates a step (subtask), encapsulating an action code. Accompanying each step is a binary value that denotes success or failure, giving preference to the successful branch over its counterpart. Task-level judgment, on the other hand, gauges the successful execution of the overall task. If the task is not completed as intended, every state within that task is labeled as negative, regardless of the status of the subtasks. This collated feedback data serves as a foundation for our Reinforcement Learning with Environmental Feedback (RLEF) methodology, which we discuss in greater detail in Section 4.4.", + "bbox": [ + 212, + 316, + 787, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Model Architecture", + "text_level": 1, + "bbox": [ + 215, + 535, + 424, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Octopus architecture (shown in Fig. 4 (c)) is heavily inspired by the Otter model [31], integrates the MPT-7B Language Decoder [40] and CLIP VIT-L/14 Vision Encoder [45]. Adopting design principles from Flamingo [4], Octopus employs the Perceiver Resampler and Cross-Gated Attention modules to enhance vision-language synergy. This architecture enables Octopus to excel in tasks requiring understanding of both visual and textual data. The Octopus is also compatible with other VLMs such as LLaVA [37].", + "bbox": [ + 212, + 560, + 787, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 SFT: Supervised Finetuning with Instructions", + "text_level": 1, + "bbox": [ + 215, + 689, + 642, + 704 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train the Octopus model on our collected dataset from OctoVerse $\\mathcal{D}_{\\mathrm{E}} = \\{(\\mathbf{X}_v,\\mathbf{T}_i,\\mathbf{T}_r)\\}$ using token-level supervised fine-tuning (SFT) [42,54]. The Perceiver Resampler transforms images $\\mathbf{X}_v$ into visual tokens that condition subsequent layers via Cross-Gated Attention modules. The training objective is next-token prediction, modeling the likelihood of a targeted response $\\mathbf{T}_r$ as:", + "bbox": [ + 212, + 714, + 785, + 791 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {T} _ {r} \\mid \\mathbf {T} _ {i}, \\mathbf {X} _ {v}\\right) = \\prod_ {l = 1} ^ {L} p \\left(t _ {l} \\mid \\mathbf {X} _ {v}, \\mathbf {T} _ {i}, \\mathbf {T} _ {r, < l}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 801, + 785, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that $\\mathbf{T}_i$ denotes the instruction tokens and $\\mathbf{T}_{r, < l}$ denotes the response tokens before the current predicted token $t_l$ . During inference, tokens are converted into natural language via the language decoder's text tokenizer.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Visual observations $\\mathbf{X}_v = \\{x_F^0,\\dots ,x_F^7,x_B^0,x_B^1\\}$ consist of 8 first-person view (FPV) images and two bird's-eye view (BEV) images for OctoGibson and OctoGTA. OctoMC only takes 4 FPV images. The FPV captures the agent's direct observations, while the BEV provides a holistic understanding of the environment. The eight FPV images are captured every 45 degrees, ensuring a complete 360-degree perspective.", + "bbox": [ + 212, + 191, + 784, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 RLEF: Reinforcement Learning with Environmental Feedback", + "text_level": 1, + "bbox": [ + 214, + 301, + 772, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In OctoVerse, task progression can be visualized as a tree (Fig. 4 (d)), where each node represents a sub-task with a binary value indicating success (1) or failure (0). If a node (or sub-task) has a value of 1, it is a step in the correct direction toward our end goal.", + "bbox": [ + 212, + 323, + 782, + 382 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tree-based Task Representation According to the environmental feedback part in Sec. 4.1, environmental reward datasets $\\mathcal{D}_{\\mathbb{R}} = (\\mathbf{X}_v^*,\\mathbf{T}_i^*,\\mathbf{T}_r^i,\\mathbf{T}_r^j,c)$ are organized, where $\\mathbf{T}_r^i$ and $\\mathbf{T}_r^j$ are responses sharing the same parent task $\\mathbf{T}_i^*$ , and $c$ indicates the preferred response leading to task completion. This ensures the reward mechanism favors the successfully executed branch. Note that even if a parental node does not have multiple responses, we can still assign feedback according to the rule in Sec. 4.1.", + "bbox": [ + 212, + 383, + 784, + 488 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Reward Model Configuration A single-modal CodeLLaMA-7B model with an additional value head is fine-tuned on $\\mathcal{D}_{\\mathrm{R}}$ as the reward model $r_{\\phi}$ . This text-based model assesses state transitions ( $\\mathbf{T}_i^* \\rightarrow \\mathbf{T}_r^{i,j}$ ) to determine high-reward transitions, assisting the agent in task execution and completion. The rationale for using CodeLLaMA as the reward model is that evaluating rewards can be purely dependent on the textual output. Furthermore, CodeLLaMA's strong programming skills make it well-suited for assessing the quality and effectiveness of the generated code in the context of task completion.", + "bbox": [ + 212, + 489, + 784, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Policy Model Development The supervised fine-tuned model serves as the initial policy model $\\pi^{\\mathrm{INIT}}$ with fixed parameters. A duplicate model, $\\pi_{\\theta}^{\\mathrm{RL}}$ , is initialized and trained using Proximal Policy Optimization (PPO) [51] to maximize response rewards. The loss function is:", + "bbox": [ + 212, + 609, + 784, + 670 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\pi_ {\\theta} ^ {\\mathrm {R L}}\\right) = - \\mathbb {E} _ {\\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right) \\in \\mathcal {D} _ {\\mathrm {R}}, \\mathbf {T} _ {r} \\sim \\pi^ {\\mathrm {R L}}} \\left[ r _ {\\phi} \\left(\\mathbf {T} _ {i} ^ {*}, \\mathbf {T} _ {r}\\right) - \\beta \\cdot \\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} ^ {\\mathrm {R L}} \\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right) \\| \\pi^ {\\mathrm {I N I T}} \\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right)\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 676, + 821, + 708 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\beta$ acts as a hyper-parameter to regulate the magnitude of the Kullback-Leibler (KL) penalty.", + "bbox": [ + 212, + 709, + 782, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 758, + 375, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Main Results on OctoGibson", + "text_level": 1, + "bbox": [ + 214, + 787, + 501, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Setup We first set up the OctoGibson to evaluate the performance of Octopus and other related models. Specifically, we are utilizing the", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a4ad0b6a9dc17f8ac4c7399acb28f0bcdf7615a6395a1567d34ae72bec4be958.jpg", + "table_caption": [ + "Table 3: Main Results on OctoGibson. We compare various models: standalone language models, adapted vision-language planners, and our Octopus models, across different evaluation settings. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators. GT denotes that the model input is directly parsed from the simulator, with information on objects (O) or relations (R). Octopus shows consistent advance in task completion." + ], + "table_footnote": [], + "table_body": "
ModelVision ModelLanguage ModelEntire Goal Task
Seen EnvUnseen EnvFollowReasonAll
GPT-4--0.42 / 0.690.46 / 0.670.49 / 0.780.27 / 0.400.43 / 0.68
GPT-4V--0.40 / 0.620.60 / 0.670.42 / 0.670.53 / 0.530.45 / 0.63
LLaMAGT (O+R)LLaMA2-7B0.07 / 0.110.13 / 0.130.11 / 0.160.00 / 0.000.08 / 0.12
CodeLLaMAGT (O+R)CodeLLaMA-7B0.09 / 0.200.20 / 0.400.16 / 0.310.00 / 0.070.12 / 0.25
TAPA (task-level)QvD GT (O)CodeLLaMA-7B0.09 / 0.360.13 / 0.330.11 / 0.360.06 / 0.330.10 / 0.35
TAPA (step-level)QvD GT (O)CodeLLaMA-7B0.16 / 0.420.13 / 0.270.18 / 0.380.07 / 0.400.15 / 0.38
EmbodiedGPTCLIP-ViTMPT-7B0.04 / 0.360.27 / 0.530.13 / 0.380.00 / 0.400.10 / 0.40
Octopus (SFT Only)CLIP-ViTMPT-7B0.11 / 0.330.27 / 0.470.16 / 0.380.13 / 0.330.15 / 0.37
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.13 / 0.380.33 / 0.530.18 / 0.400.20 / 0.530.18 / 0.42
", + "bbox": [ + 217, + 241, + 784, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "metrics of goal task completion score to check whether the task is completed in the simulator and the plan score from human evaluation. We have 60 evaluation tasks, with 45 from the seen environment, and 15 that are unseen during training. We also have 45 routine tasks and 15 require reasoning. Please note that models like Octopus might not always accurately identify specific object names as they appear in the simulator (e.g., \"water_bottle_189\"). To address this, we implement a post-processing step for the generated code, substituting generic object references with their exact names from the simulator with simple string similarity matching. If multiple objects, we select the one closest to the agent.", + "bbox": [ + 212, + 381, + 784, + 517 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For Blind LLMs, we provide them with all the environment information in a textual format. Referring to Figure 3, we hope the Blind LLMs could perform as GPT-4 but internalize the system message. For TAPA utilizes the open-vocabulary detection (OVD) technique [65] to recognize objects within images and parse them into textual environmental messages, we still provide it with ground-truth environmental messages as an oracle setting.", + "bbox": [ + 212, + 517, + 785, + 608 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CodeLLaMA Improves Coding but not Planning. The first two rows in Table 3 highlight the suboptimal task completion rate of the blind LLMs. Among them, CodeLLaMA boasts pre-training on a large programming dataset, resulting in a notable enhancement in code execution from our observation, with $92\\%$ of the written code being successfully executed compared to LLaMA's $24\\%$ . However, its prowess in planning remains limited. In contrast, the proposed Octopus MPT-7B model displays superior planning and task completion metrics while maintaining commendable coding abilities ( $72\\%$ of the written code can be executed). We surmise that the coding requirements within the OctoGibson environment might not be exceedingly intricate, rendering an advanced programming language model, like CodeLLaMA, less crucial, albeit beneficial. For more insight, although not shown in the table, our efforts to replace the MPT model with CodeLLaMA encountered challenges of generating non-sense outputs, suggesting that more refined code, or image-code paired data might be necessary for a successful Octopus-CodeLLaMA integration.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a68d0cb8ddc8c569626e79a3f1dea21e19a76649253c2fd8619d3bc8a989ce56.jpg", + "image_caption": [ + "(a) TAPA (step) + CodeLLaMA" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 388, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bf2d68a0a852df651c24935ae658d8b2430478c74c364fde7e43907caebd633c.jpg", + "image_caption": [ + "(b) Octopus (SFT only)" + ], + "image_footnote": [], + "bbox": [ + 395, + 145, + 578, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3c1e63dfccfb810c30e5665f109edf1a15ea91cbcda2b41185f1c6239dd2cad4.jpg", + "image_caption": [ + "(c) Octopus (SFT + RLEF)" + ], + "image_footnote": [], + "bbox": [ + 586, + 145, + 782, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f71f405984cf4e8393b9c1e753f545beaca6a3878d55a4d6b61507c55dadcd07.jpg", + "image_caption": [ + "(d) Vision Input for Vision-Language Models (e.g., b, c, e)" + ], + "image_footnote": [], + "bbox": [ + 218, + 297, + 565, + 435 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/454c045c5ac0ea802e3826daf5c56f2251fb041a482a63198e9ed0dc3ce774fb.jpg", + "image_caption": [ + "(e) GPT-4V Output", + "Fig. 5: Qualitative Results on the task of find a carboy in OctoGibson environment. We show that the models shown can write executable code, but the proposed Octopus has stronger planning ability, especially after RLEF. We also explore the performance of GPT-4V on the specific task." + ], + "image_footnote": [], + "bbox": [ + 571, + 297, + 781, + 435 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Blind LLMs Struggle with Extended Input Content. Our observations indicate that the step-level TAPA model, when supplied with a ground-truth object list, achieves a notable enhancement in planning. The primary distinction between it and the blind CodeLLaMA lies in the input length; the latter deals with protracted, pairwise relation content, complicating the language model's ability to extract crucial data from the environment message. This scenario highlights the inherent limitation of blind LLMs: relying on language alone to convey the entirety of environmental data can result in less informative input.", + "bbox": [ + 212, + 508, + 787, + 631 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Octopus Demonstrates Superior Task Generalization. Table 3 underscores Octopus's strong performance, evidencing its consistent edge over standalone language models in task completion. Its adeptness in adapting to previously unencumbered environments underlines the inherent advantages of vision-language models. A more detailed ablation analysis is provided later.", + "bbox": [ + 212, + 637, + 787, + 713 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "RLEF Enhances Octopus's Planning Strategy. Table 3 shows Octopus's strong reasoning capabilities after the RLEF finetuning. An example can be observed in Fig. 5(b-c), where, after refinement via RLEF, Octopus astutely navigates to the cabinet housing the carboy instead of attempting a direct yet distant capture. Quantitatively, Octopus exhibits enhanced adaptability to previously unseen reasoning tasks, reinforcing its prowess in logical task resolution. When juxtaposed with other strategies, such as the embodied queries employed by EmbodiedGPT, RLEF emerges as the more efficacious approach.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/721900636f8ea71ee414022a79cb72aa5446b63aee809211ccdb846fe3fbce34.jpg", + "image_caption": [ + "(a) On Tuning Different Components" + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 406, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/57ce72ccfe3522c85a01290dbffdf88e7ffd2cd05813e6e487154389caab0f2d.jpg", + "image_caption": [ + "(b) On SFT & RLEF and Model Size", + "(c) On Necessity of Vision Input", + "Fig.6: Ablation Study on model components, model size, and vision input. For bars with different colors, the upper bar denotes the number of successful reasoning tasks, and the lower is routine tasks." + ], + "image_footnote": [], + "bbox": [ + 433, + 147, + 635, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/961aa896bd9576199f781b54b4eac79e45a7bc0e813ffd32b68085f9dc26eda2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 148, + 767, + 241 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 297, + 388, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tunning Different Components Fig. 6 (a) demonstrates that solely adjusting the connector (marked \"fire\" in Fig. 4 (a)) leads to success for merely 4 out of 60 tasks. Conversely, finetuning both the connector and language decoder nudges the success rate slightly higher, with 5 tasks being accomplished.", + "bbox": [ + 212, + 320, + 784, + 381 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7B v.s. 3B Model Size We embarked on experiments centered on model size to discern the influence of the total parameter count on the efficacy of vision-language models. As illustrated in Fig. 6 (b), downsizing the model manifests in a noticeable performance drop. The congruency of results across both the SFT and RLEF models underscores the importance of an apt model size when sculpting vision-language models.", + "bbox": [ + 212, + 383, + 784, + 474 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Significance of Visual Inputs in Task Performance In our standard configuration, the vision component processes a sequence of image inputs, consisting of eight circularly captured first-person view (FPV) images, complemented by two bird's-eye view (BEV) images. With the intent to investigate the impact of visual inputs on task performance, we initiated an ablation study. In a modified setup, the sequence of these visual inputs was deliberately randomized, aiming to attenuate the strength of the visual signals. As illustrated in Fig. 6 (c), this intentional disruption in visual input consistency led to a pronounced decline in task performance. This result highlights the crucial role that clear and structured visual inputs play in the Octopus model, emphasizing that it significantly leverages visual cues for effective planning and task execution.", + "bbox": [ + 212, + 478, + 785, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Results on Minecraft and GTA Tasks", + "text_level": 1, + "bbox": [ + 215, + 662, + 570, + 678 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results on OctoMC According to the OctoMC part in Sec. 3, we designed 40 tasks, each task is operated on 2 locations so a total of 80 tasks. We set aside 10 tasks as unseen tasks and 10 tasks as seen tasks, getting 60 training tasks and 30 testing tasks in OctoMC. Similar to OctoGibson, the training data for OctoMC is collected using GPT-4. The agent, guided by GPT-4, explores the Minecraft environment and generates action plans and corresponding code based on the provided system messages, environmental cues, and objectives.", + "bbox": [ + 212, + 688, + 784, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4 shows that the SFT model trained on OctoMC can complete most tasks in both seen and unseen scenarios, demonstrating better performance compared to OctoGTA. However, upon analyzing the failure cases, we find that the", + "bbox": [ + 212, + 795, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/42d0fb4a90c2701ba28200c150a89d74222db3bb3b5dea5d2adaeffc7b6c5aed.jpg", + "table_caption": [ + "Table 4: Main Results for GTA and Minecraft Tasks. Despite limited training data, the Octopus still shows its good ability in task completion. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators." + ], + "table_footnote": [], + "table_body": "
ModelVision ModelLanguage ModelOctoMCOctoGTA
Seen TaskUnseen TaskAllSeen TaskUnseen TaskAll
GPT-4--0.70 / 0.850.60 / 0.800.65 / 0.830.55 / 0.800.50 / 0.600.54 / 0.76
GPT-4V--0.75 / 0.850.70 / 0.850.73 / 0.850.58 / 0.850.50 / 0.700.56 / 0.82
EmbodiedGPTCLIP-ViTMPT-7B0.30 / 0.550.20 / 0.600.25 / 0.580.15 / 0.380.20 / 0.600.16 / 0.42
Octopus (SFT Only)CLIP-ViTMPT-7B0.30 / 0.600.20 / 0.700.25 / 0.650.18 / 0.480.20 / 0.600.18 / 0.50
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.40 / 0.600.20 / 0.700.30 / 0.650.18 / 0.530.30 / 0.700.20 / 0.56
", + "bbox": [ + 217, + 215, + 785, + 308 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "model sometimes struggles with tasks requiring precise spatial reasoning. For instance, when tasked with killing a pig, the agent may have difficulty finding the creature with the exact angle and distance. While it shows that the agent relies on visual information to navigate and interact with the environment, it also means that even with correct planning, imprecise actions can lead to failure.", + "bbox": [ + 212, + 311, + 782, + 387 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results on OctoGTA According to the OctoGTA part in Sec. 3, we designed 25 tasks. We set aside 5 tasks in the boat-related group (e.g., boat retrieval and shore return) as unseen tasks for testing only, using 2 different locations. We replicate the remaining 20 tasks for both training (8 different locations) and testing (2 locations). As a result, we have 160 training tasks and 50 testing tasks in OctoGTA. Unlike the training procedure for OctoGibson and OctoMC, the training data for OctoGTA is entirely created by the authors, as it is challenging to gather textual environmental messages in the GTA environment.", + "bbox": [ + 212, + 390, + 787, + 511 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 4 shows that, despite having only 160 training tasks, the SFT model can complete some tasks in both seen and unseen scenarios, and RLEF also outperforms. However, upon careful examination of the failure cases, we find that the model struggles with tasks that are not straightforward. For instance, when a wall separates the player from the car, the player still finds it difficult to decide to climb the wall, even if similar cases exist in the training data. Similar to OctoMC, as illustrated in Sec. 3, when approaching certain locations, the code involves functions like turnPlayer() and goForward() rather than a simple walkTo location). Consequently, even with correct planning, imprecise actual actions can still lead to task failure.", + "bbox": [ + 212, + 513, + 787, + 664 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 702, + 359, + 719 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces Octopus, an embodied vision-language programmer designed to bridge the gap between high-level planning and real-world manipulation with programming. By open-sourcing our OctoVerse environments, dataset, and Octopus architecture, we aim to foster collaboration and innovation within the research community, paving the way for future developments in embodied vision-language programming.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments and Disclosure of Funding", + "text_level": 1, + "bbox": [ + 215, + 143, + 651, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221- 0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 212, + 176, + 787, + 237 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Grand theft auto v (2014)", + "2. Minecraft (2023)", + "3. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Ho, D., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E., Ruano, R.J., Jeffrey, K., Jesmonth, S., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Lee, K.H., Levine, S., Lu, Y., Luu, L., Parada, C., Pastor, P., Quiambao, J., Rao, K., Rettinghouse, J., Reyes, D., Sermanet, P., Sievers, N., Tan, C., Toshev, A., Vanhoucke, V., Xia, F., Xiao, T., Xu, P., Xu, S., Yan, M., Zeng, A.: Do as i can and not as i say: Grounding language in robotic affordances. In: arXiv preprint arXiv:2204.01691 (2022)", + "4. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "6. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos (2022)", + "7. Bellemare, M.G., Naddaf, Y., Veness, J., Bowling, M.: The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research 47, 253-279 (Jun 2013). https://doi.org/10.1613/jair.3912, http://dx.doi.org/10.1613/jair.3912", + "8. Billard, A., Kragic, D.: Trends and challenges in robot manipulation. Science 364(6446), eaat8414 (2019)", + "9. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., Zaremba, W.: Openai gym. arXiv preprint arXiv:1606.01540 (2016)", + "10. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023)", + "11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "12. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)" + ], + "bbox": [ + 218, + 287, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Chen, J., Mu, Y., Yu, Q., Wei, T., Wu, S., Yuan, Z., Liang, Z., Yang, C., Zhang, K., Shao, W., Qiao, Y., Xu, H., Ding, M., Luo, P.: Roboscript: Code generation for free-form manipulation tasks across real and simulation (2024)", + "14. Chen, L., Li, B., Shen, S., Yang, J., Li, C., Keutzer, K., Darrell, T., Liu, Z.: Language models are visual reasoning coordinators. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023)", + "15. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)", + "16. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)", + "17. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering (2017)", + "18. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model. In: arXiv preprint arXiv:2303.03378 (2023)", + "19. Evans, J.S.B.: Dual-processing accounts of reasoning, judgment, and social cognition. Annu. Rev. Psychol. 59, 255-278 (2008)", + "20. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge (2022)", + "21. Fu, H., Xu, W., Ye, R., Xue, H., Yu, Z., Tang, T., Li, Y., Du, W., Zhang, J., Lu, C.: Rfuniverse: A multiphysics simulation platform for embodied ai (2023)", + "22. Fu, Z., Zhao, T.Z., Finn, C.: Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In: arXiv (2024)", + "23. Gao, X., Gong, R., Shu, T., Xie, X., Wang, S., Zhu, S.C.: Vrkitchen: an interactive 3d virtual environment for task-oriented learning (2019)", + "24. Gu, S., Holly, E., Lillicrap, T., Levine, S.: Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 3389-3396. IEEE (2017)", + "25. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023)", + "26. Huang, S., Jiang, Z., Dong, H., Qiao, Y., Gao, P., Li, H.: Instruct2act: Mapping multi-modality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176 (2023)", + "27. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023)", + "28. Hussein, A., Gaber, M.M., Elyan, E., Jayne, C.: Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR) 50(2), 1-35 (2017)", + "29. Kahneman, D.: Thinking, fast and slow. macmillan (2011)", + "30. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)", + "31. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Conference on Robot Learning. pp. 80–93. PMLR (2023)", + "33. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "34. Lin, K., Agia, C., Migimatsu, T., Pavone, M., Bohg, J.: Text2motion: From natural language instructions to feasible plans. arXiv preprint arXiv:2303.12153 (2023)", + "35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)", + "36. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/", + "37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)", + "38. Liu, Z., Dong, Y., Rao, Y., Zhou, J., Lu, J.: Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966 (2024)", + "39. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition (2021)", + "40. MosaicML: Mpt-7b (2023), https://www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-23", + "41. Mu, Y., Zhang, Q., Hu, M., Wang, W., Ding, M., Jin, J., Wang, B., Dai, J., Qiao, Y., Luo, P.: Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021 (2023)", + "42. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)", + "43. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023)", + "44. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtual-home: Simulating household activities via programs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8494-8502 (2018)", + "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "46. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "47. Ramakrishnan, S.K., Jayaraman, D., Grauman, K.: An exploration of embodied visual exploration (2020)", + "48. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable task planning. arXiv preprint arXiv:2307.06135 (2023)", + "49. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Octopus: Embodied VLM from Environmental Feedback", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023)", + "51. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)", + "52. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)", + "53. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023)", + "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "55. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models (2023)", + "56. Wang, H., Liang, W., Gool, L.V., Wang, W.: Towards versatile embodied navigation (2022)", + "57. Wu, Y., Wu, Y., Gkioxari, G., Tian, Y.: Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209 (2018)", + "58. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023)", + "59. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: Sapien: A simulated part-based interactive environment (2020)", + "60. Xie, B., Zhang, S., Zhou, Z., Li, B., Zhang, Y., Hessel, J., Yang, J., Liu, Z.: Funqa: Towards surprising video comprehension. arXiv preprint arXiv:2306.14899 (2023)", + "61. Yan, C., Misra, D., Bennnett, A., Walsman, A., Bisk, Y., Artzi, Y.: Chalet: Cornell house agent learning environment (2019)", + "62. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Arenas, M.G., Chiang, H.T.L., Erez, T., Hasenclever, L., Humplik, J., et al.: Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647 (2023)", + "63. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Skill reinforcement learning and planning for open-world long-horizon tasks (2023)", + "64. Zheng, S., Liu, J., Feng, Y., Lu, Z.: Steve-eye: Equipping llm-based embodied agents with visual perception in open worlds (2023)", + "65. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)", + "66. Zhu, Y., Wong, J., Mandlekar, A., Martin-Martín, R., Joshi, A., Nasiriany, S., Zhu, Y.: robosite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 729 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_model.json b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae15f261c32cc30f4dae4d24e3e11f82aa1098d2 --- /dev/null +++ b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_model.json @@ -0,0 +1,2656 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.142, + 0.75, + 0.186 + ], + "angle": 0, + "content": "Octopus: Embodied Vision-Language Programmer from Environmental Feedback" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.212, + 0.744, + 0.259 + ], + "angle": 0, + "content": "Jingkang Yang\\*,1, Yuhao Dong\\*,2,3, Shuai Liu\\*,2,4, Bo Li\\*,1, Ziyue Wang\\*,1, Haoran Tan\\*,4, Chencheng Jiang\\*,5, Jiamu Kang\\*,3, Yuanhan Zhang\\*, Kaiyang Zhou\\* and and Ziwei Liu\\*" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.269, + 0.786, + 0.313 + ], + "angle": 0, + "content": "\\(^{1}\\) S-Lab, Nanyang Technological University \\(^{2}\\) Shanghai AI Laboratory \\(^{3}\\) Tsinghua University \\(^{4}\\) BUPT \\(^{5}\\) XJTU \\(^{6}\\) Hong Kong Baptist University \\(\\{\\text{jingkang001, ziwei.liu}\\}@\\text{ntu.edu.sg}\\)" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.32, + 0.784, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Fig. 1: Illustration of our vision-language programmer, Octopus, complete a task in GTA environment. Given a task in the form of natural language, Octopus relies on its egocentric vision to generate plans and the corresponding executable code." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.65, + 0.744, + 0.817 + ], + "angle": 0, + "content": "Abstract. Large vision-language models (VLMs) have achieved substantial progress in multimodal perception and reasoning. When integrated into an embodied agent, existing embodied VLM works either output detailed action sequences at the manipulation level or only provide plans at an abstract level, leaving a gap between high-level planning and real-world manipulation. To bridge this gap, we introduce Octopus, an embodied vision-language programmer that uses executable code generation as a medium to connect planning and manipulation. Octopus is designed to 1) proficiently comprehend an agent's visual and textual task objectives, 2) formulate intricate action sequences, and 3) generate executable code. To facilitate Octopus model development, we introduce OctoVerse: a suite of environments tailored for benchmarking" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.753, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution, †Equal engineering contribution, ⌿ Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.343 + ], + "angle": 0, + "content": "vision-based code generators on a wide spectrum of tasks, ranging from mundane daily chores in simulators to sophisticated interactions in complex video games such as Grand Theft Auto (GTA) and Minecraft. To train Octopus, we leverage GPT-4 to control an explorative agent that generates training data, i.e., action blueprints and corresponding executable code. We also collect feedback that enables an enhanced training scheme called Reinforcement Learning with Environmental Feedback (RLEF). Through a series of experiments, we demonstrate Octopus's functionality and present compelling results, showing that the proposed RLEF refines the agent's decision-making. By open-sourcing our simulation environments, dataset, and model architecture, we aspire to ignite further innovation and foster collaborative applications within the broader embodied AI community. The project page is available at https://choiszt.github.io/0ctopus/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.369, + 0.377, + 0.386 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.787, + 0.476 + ], + "angle": 0, + "content": "The rise of large language models (LLMs) [11,15,42,46,54] led to a surge in vision-language models (VLMs) [4,5,31,33,35-37], enabling tasks such as image/video-based descriptions [33], reasoning [14,38,60], and conversations [16,31]. In the realm of embodied AI, notable efforts [10,18,33] have trained agents to process visual input and relay motor control commands." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Another approach to interacting with the environment focuses on task execution through code invocations, mirroring the human System-I stimulation [19,29] (automatic, intuitive actions) with predefined code, and leaving the System-II processes [19,29] (planning and reasoning) for large models. For example, referring to Fig. 1, planning a car ride with a pet might entail a subconscious checklist (e.g., getOutOf() the house, open() the car door), each action could be implemented using specific techniques [8,24] such as imitation learning [22,28]. This programmatic paradigm has been, although not in vision, leveraged by works [25, 50, 52, 53] using LLMs to craft programs and trigger APIs. Game-centric models like Voyager [55] have similarly employed GPT for function calls within game engines, though they often parse data directly from their environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "However, when incorporating visual perception, the programming paradigms are largely unexplored. Primary initiatives [48,58] can only output plans, which anchor their strategies only in initial environmental states or employ dynamic scene graphs for LLM inputs, respectively. Despite their innovations, the overreliance on pre-trained vision models to convert vision content into language can occasionally hinder the LLM's planning performance. The conversion from plans into real-world actions is still missing. While EmbodiedGPT [41] addresses the problem by integrating vision-language modeling for planning and then transitioning to manipulation using policy mapping, the capability of embodied vision-language models to generate executable programs remains largely uncharted." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Our exploration aims to bridge this gap. An embodied vision-language programmer should integrate visual perspective with textual objectives to devise" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "action plans and executable code (Fig. 1). However, existing simulator environments often lack the carefully designed functions necessary to support such models effectively. These functions should balance usefulness and complexity to avoid hindering the development of genuine embodied vision-language programmers. For example, the **explore_until()** function in Minecraft, which can lead the player directly to specific blocks without relying on vision information, may not be suitable for training these models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.388 + ], + "angle": 0, + "content": "To meet the requirement, we carefully design and develop OctoVerse, a suite of environments consisting of diverse simulators, including (i) OctoGibson, built upon the photorealistic OmniGibson [32], (ii) OctoMC, developed on the infinitely creative, pixel-style Minecraft platform [2], and (iii) OctoGTA, adapted from the highly interactive and immersive Grand Theft Auto V (GTA-V) [1]. These environments enable the training and benchmarking of our embodied vision-language programming model in a wide range of scenarios, from daily household tasks to complex urban navigation and open-world exploration, while the function calls are tailored to be vision-dependent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.785, + 0.569 + ], + "angle": 0, + "content": "Using the OctoVerse environment, we train Octopus by leveraging GPT-4 to collect data. We provide GPT-4 with system messages, environmental cues, and objectives, enabling it to formulate action strategies and code. Simultaneously, the agent captures visual perspectives, forming the image-code pair for Octopus training. During data collection, the agent receives simulator feedback, distinguishing successful moves from unsuccessful ones. We incorporate this feedback using Reinforcement Learning with Environmental Feedback (RLEF) and fine-tune Octopus using Proximal Policy Optimization (PPO) [51]. Empirically, Octopus demonstrates strong adaptability in various scenarios, outperforming existing models in task planning, code generation, and execution. The integration of RLEF further enhances Octopus's performance, showcasing the effectiveness of this training approach. In sum, our key contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.579, + 0.787, + 0.64 + ], + "angle": 0, + "content": "- A Novel Vision-Language Programming Benchmark: Three diverse embodied environments with designed tasks: (i) OctoGibson, which is developed upon OmniGibson [32], (ii) OctoMC that developed on Minecraft [2], and (iii) OctoGTA, which is adapted from GTA-V [1]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.64, + 0.787, + 0.685 + ], + "angle": 0, + "content": "- A New Vision-Language Programming Model: An embodied vision-language planner and programmer trained with Reinforcement Learning with Environmental Feedback (RLEF), demonstrates compelling results." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.685, + 0.787, + 0.731 + ], + "angle": 0, + "content": "- Insights on Vision-Language Programming: We extensively explore Octopus and share useful insights facilitating future research on visual planning and programming." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.579, + 0.787, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.388, + 0.769 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.468, + 0.799 + ], + "angle": 0, + "content": "2.1 Embodied AI Simulators" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Embodied AI has advanced significantly with the development of diverse simulation environments, enabling research tasks such as visual exploration [47]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.201 + ], + "angle": 0, + "content": "Table 1: Related Work for OctoVerse - Overview of Embodied AI Environments. We select three environments into OctoVerse and carefully design executable tasks and vision-dependent function calls (VC), in comparison to undesigned standard function calls (C)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.381 + ], + "angle": 0, + "content": "
SimulatorKinematicsContinuous \nExtended \nStatesFlexible \nMaterialsDeformable \nBodiesRealistic \nAction \nExecutionGame- or \nWorld-BasedFormulated \nTasksFunction \nCall \nType
OpenAIGym [9]✓×××✓G×C
Matterport3D [12]×××××W××
AI2THOR [30]✓××××G×C
VirtualHome [44]×××××G××
House3D [57]×××××W××
Habitat 1.0 [49]✓×××✓W×C
Robosuite [66]✓×××✓W×C
RFUniverse [21]✓×✓✓✓W×C
Minecraft [2]✓×✓×✓G×C
OctoMC✓×✓×✓G✓VC
GTA [1]✓✓✓✓✓G×C
OctoGTA✓✓✓✓✓G✓VC
OmniGibson [32]✓✓✓✓✓W×C
OctoGibson✓✓✓✓✓W✓VC
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.787, + 0.492 + ], + "angle": 0, + "content": "navigation [56], and question-answering [17]. Several simulators, including AI2-THOR [30], VirtualHome [44], Habitat-Sim [49], SAPIEN [59], and Omnigibson [32], provide realistic representations of the world for investigating embodied AI challenges. OmniGibson [32] stands out for its high-fidelity simulation of diverse indoor and outdoor environments. OctoGibson environment further enhances OmniGibson with carefully designed function calls and formulated tasks, making it well-suited for vision-language programming." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.645 + ], + "angle": 0, + "content": "Game-related simulators like Arade [7], CHALET [61], and VRKitchen [23] also contribute significantly to embodied AI. Minecraft [2] has gained attention in reinforcement learning and game agents [6,20,39,55,63,64] but lacks the necessary structure for vision-language programming. OctoMC addresses this by providing designed function calls and formulated tasks. In contrast to Minecraft's voxel-based representations that limit transferability to real-world environments, GTA-V [1] offers a highly realistic environment. In this work, we introduce OctoGTA as a new setting, leveraging GTA-V's rich, open-world environment with incorporated tasks and function calls, extending this platform for embodied AI study." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.534, + 0.68 + ], + "angle": 0, + "content": "2.2 Embodied AI with Large Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.8, + 0.842 + ], + "angle": 0, + "content": "The recent wave of research focuses on merging LLMs with embodied AI tasks [11, 42,46,54]. For instance, VoxPoser addresses robotic manipulation problems through unsupervised methods [27]. A group of projects, namely SayCan [3], Palm-E [18], RT-2 [10], and EmbodiedGPT [41], effectively integrate visual or linguistic cues with robot manipulation data. Outside the domain of robotic manipulation, initiatives like Voyager [55] and Smallville [43] harness the capabilities of GPT to interface with game functions, relying on preset functions to manage intricate manipulations. In a parallel vein, VisProg [25] leverages GPT-3 language prompts to craft Python programs, opening the door to a multitude of fascinating applications. While the proposed Octopus model also formulates plans" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Related Work for Octopus - Overviewing Embodied AI Models. The proposed Octopus distinguishes itself from other models as a unified vision-language model for both plan and code generation." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.193, + 0.789, + 0.406 + ], + "angle": 0, + "content": "
ModelsRelease DateSupported EnvironmentVision ModelCode GeneratorAction w/ FeedbackLLM Training Enabled
Text2Motion [34]Mar. 2023Sim×✓×
Instruct2Act [26]May 2023Sim×✓×
Lang2Rewards [62]Jun. 2023Sim×✓×
VoxPoser [27]Jul. 2023Sim✓××
SayCan [3]Apr. 2022Real✓××
PALM-E [18]Mar. 2023Sim, Real✓×✓
RT-2 [10]Jul. 2023Real✓×✓
SayPlan [48]Jun. 2023Real×××
EmbodiedGPT [41]May 2023Sim✓×✓
TaPA [58]Jul. 2023Sim××✓
Voyager [55]May 2023Game×✓×
Steve-Eye [64]Dec 2023Game✓×✓
RoboScript [13]Feb 2024Sim, Real×✓×
Octopus-Sim, Game✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.784, + 0.5 + ], + "angle": 0, + "content": "and code, its distinguishing feature is the seamless integration of visual input in program and code generation. This also stands in contrast to other embodied planners like TAPA [58] and SayPlan [48], which deploy separate vision modules to translate visual data into linguistic inputs for LLMs. Octopus excels as a cohesive vision-language model, delivering not just plans but also code." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.501, + 0.784, + 0.531 + ], + "angle": 0, + "content": "More discussion on additional related works (Vision Language Model, and Feedback in Large Language Models) is included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.559, + 0.528, + 0.575 + ], + "angle": 0, + "content": "3 The OctoVerse Environment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.785, + 0.657 + ], + "angle": 0, + "content": "In this section, we introduce three simulator environments designed to train and evaluate the Octopus model. For each environment, we will describe their overall information, the special design considerations that ensure the tasks are well-formulated and the callable functions are vision-dependent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.788, + 0.814 + ], + "angle": 0, + "content": "OctoGibson We built the environment on the foundation of OmniGibson [32], an existing simulation framework that supports 1,000 daily activities across 50 scenes, featuring over 5,000 meticulously annotated objects. We incorporated 16 functions that the robot can execute, such as moveBot() and easyGrasp(). Within this environment, we meticulously crafted 476 tasks1, each with a well-defined initial state and a definitive termination state, allowing for a straightforward assessment of task completion. Among these tasks, 367 are routine tasks—simple and direct actions like \"place a glass in a trash can,\" marked as Follow. Conversely, the remaining 109 are reasoning tasks that necessitate deeper comprehension. An example is \"buy a chocolate,\" where the agent needs" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.778, + 0.84 + ], + "angle": 0, + "content": "1 The full list of tasks and their categories are listed in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "to know to pick a chocolate bar from the shelf and then place it, along with money, on the checkout counter, denoted as Reason tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.315 + ], + "angle": 0, + "content": "To ensure our tasks are vision-aware, we deliberately constrain the usage of certain functions, such as moveBot(object), which moves the agent in front of the given object. To avoid making the task too easy and vision-agnostic, we limit the given parameter to a predefined set of large objects, such as tables and cupboards, rather than small items like cups and glasses. In this case, if the robot wants to pick up a cup, it needs to recognize whether the cup is on the table or in the cupboard. A simple moveBot(object) call with an inappropriate parameter would cause a runtime error. The full list of the functions is in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.41 + ], + "angle": 0, + "content": "OctoMC The OctoMC environment is built on Minecraft [2], a popular platform for reinforcement learning and game agents. We integrated 6 functional actions and crafted 40 tasks\\(^{2}\\), each designed to facilitate comprehensive observations and executions by the agent. These tasks are distributed across 10 different biomes, including indoor, outdoor, and underground settings, under varying weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.548 + ], + "angle": 0, + "content": "However, existing Minecraft environments often lack vision-formulated tasks and the necessary structure for vision-language programming. For example, in Voyager [55], the exploreUntil() function allows the player to navigate directly to specific blocks without relying on vision information. This function works by randomly exploring the environment within a certain range until the desired object is found, at which point it returns a true value, enabling the agent to interact with the located object, such as gold blocks or trees. While effective, this approach is entirely automated and does not utilize visual information, making it unsuitable for our vision-based objectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.641 + ], + "angle": 0, + "content": "To address this limitation, we crafted a vision-dependent exploration function, teleport(yaw, distance), which operates within the robot's perceptual range. This function ensures that the agent's operations are vision-dependent and require active perception and navigation within the environment. For instance, to locate a tree block, the agent must actively navigate towards the direction of the forest, relying on visual cues to guide its exploration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.722 + ], + "angle": 0, + "content": "OctoGTA The OctoGTA environment is built on GTA-V [1] with the help of the active GTA modding community. We integrated 19 functions and methodically crafted 25 tasks3, such as \"help NPC drive their boat back to shore\" and \"mediate a fight between two NPCs,\" spanning across 5 groups (shown in Fig. 2 (b)). Each task is assigned to 5 different locations within the game world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.784 + ], + "angle": 0, + "content": "We have implemented a set of functions that enable the agent to interact with the game world in a visually-aware manner. Similar to the design in Minecraft, we get rid of functions like walkTo location) that might trivialize the task of reaching a particular building or landmark. Instead, we provide functions such" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.604, + 0.812 + ], + "angle": 0, + "content": "2 Detailed tasks are listed in the supplementary material." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3 We meticulously designed tasks to be friendly, ensuring they exclude any inappropriate or violent behaviors." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.782, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.31, + 0.785, + 0.324 + ], + "angle": 0, + "content": "Fig. 2: The Statistics of the OctoVerse Environment with Function Designs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.785, + 0.414 + ], + "angle": 0, + "content": "as goForward(distance) and turnPlayer(degree) (Fig. 2 (b)). For tasks like mediating a fight between two NPCs, the essential function stopFight() only works when the player is within 5 meters of the fighting NPCs. These design choices ensure that the agent's operations are vision-dependent and require active perception and navigation within the environment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.782, + 0.451 + ], + "angle": 0, + "content": "4 Octopus: The Embodied Vision-Language Programmer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.537 + ], + "angle": 0, + "content": "In this section, we present the procedure for training Octopus. Starting from collecting training data within the OctoVerse environment, Octopus builds upon a VLM architecture of Otter [31] and includes specialized RLEF modules to handle vision-language programming tasks. Fig. 4 illustrates the entire Octopus training pipeline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.467, + 0.571 + ], + "angle": 0, + "content": "4.1 Training Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.785, + 0.699 + ], + "angle": 0, + "content": "We use the automatic training data collection pipeline described here for OctoGibson and OctoMC, with the latter using customized prompts inspired by Voyager [55]. For OctoGTA, we rely on human labor to hand-craft the training dataset due to the difficulty of obtaining textual environment messages in the GTA environment. In the following parts, we use OctoGibson as the primary example to illustrate the data collection pipeline. Note that the primary task in organizing training data is to form a succinct pairing: \"vision input + current/historical states \\(\\rightarrow\\) next step plan + executable code\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.785, + 0.792 + ], + "angle": 0, + "content": "Environment Info Collection As shown in Fig. 4 (a) and Fig. 3, we format an environment message for each state, encompassing attributes like Observed Objects, Observed Relations, Inventory, and more. Specifically, the simulator can provide us with an exact scene graph at each state, shaping the content for the first two parts. The inventory info is also accessible. The task, e.g., \"cooking bacon\" in Fig. 3, is represented by the Task Goal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Automation with GPT-4 After preparing the environment message, we crafted a structured system message to ensure that the robot not only understands its input but also maintains a consistent output format. A detailed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.145, + 0.781, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Fig.3: Data Collection Example for \"Cook a Bacon\" Task. GPT-4 perceives the environment through the environmental message and produces anticipated plans and code following the detailed system message. This code is subsequently executed in the simulator, directing the agent to the subsequent state. For each state, we gather the environmental message, wherein observed objects and relations are substituted by egocentric images to serve as the training input. The response from GPT-4 acts as the training output. Environmental feedback, specifically the determination of whether each target state is met, is documented for RLEF training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.651 + ], + "angle": 0, + "content": "examination of this prompt can be found in the appendix. Experiments have shown that a well-articulated prompt enables GPT-4 to effectively generate executable code. It is important to note that the combined length of the system and environment messages can be extremely long, which may cause standard GPT-4 8K models to struggle with producing meaningful outputs. To address this issue, we employ the more robust GPT-4 32K model. As illustrated in Fig. 3, when GPT-4 receives a consistent system and environment message, it generates comprehensive outputs that include current scenario analysis, planning, and actionable code, supporting the training process in Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Error Management Notably, GPT-4 collects training data under the main task of guiding the agent to complete tasks. However, GPT-4 is not infallible. Errors can manifest in multiple ways, ranging from syntax errors to physical challenges in the simulator. For instance, in Fig. 3, between states \\(\\#5\\) and \\(\\#6\\), the action failed due to the long distance between the agent (holding bacon) and the pan. Such setbacks reset the task to its previous state. If a main task remains incomplete after 10 steps, it is deemed unsuccessful, and we terminate this task for budget concerns. However, all data pairs without syntax errors, regardless of the task's completion status, are valuable for refining instructions and improving the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Environmental Feedback GPT-4's continual trial-and-error approach while guiding the agent toward task completion serves a dual purpose: collecting vision-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.149, + 0.455, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.217, + 0.41, + 0.228 + ], + "angle": 0, + "content": "(a) Data Collection Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.237, + 0.457, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.274, + 0.415, + 0.285 + ], + "angle": 0, + "content": "(b) Octopus Training Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.147, + 0.636, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.274, + 0.619, + 0.285 + ], + "angle": 0, + "content": "(c) Octopus (Otter model)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.147, + 0.785, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.205, + 0.756, + 0.214 + ], + "angle": 0, + "content": "Step 2: Training Reward Model" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.214, + 0.78, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.248, + 0.78, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.276, + 0.769, + 0.286 + ], + "angle": 0, + "content": "(d) RLEF Training" + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.292, + 0.737, + 0.306 + ], + "angle": 0, + "content": "Fig. 4: How to train Octopus: data collection and training pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.514 + ], + "angle": 0, + "content": "output pairs and generating a rich set of feedback data. The automatic annotation of this feedback focuses on two levels: step-level and task-level judgments. Step-level judgment assesses the alignment of post-execution states with their target states. For instance, in Fig. 3, steps color-coded in green lead to positive feedback. One can visualize the action sequence for task completion as a tree, where each node indicates a step (subtask), encapsulating an action code. Accompanying each step is a binary value that denotes success or failure, giving preference to the successful branch over its counterpart. Task-level judgment, on the other hand, gauges the successful execution of the overall task. If the task is not completed as intended, every state within that task is labeled as negative, regardless of the status of the subtasks. This collated feedback data serves as a foundation for our Reinforcement Learning with Environmental Feedback (RLEF) methodology, which we discuss in greater detail in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.536, + 0.426, + 0.549 + ], + "angle": 0, + "content": "4.2 Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.669 + ], + "angle": 0, + "content": "The Octopus architecture (shown in Fig. 4 (c)) is heavily inspired by the Otter model [31], integrates the MPT-7B Language Decoder [40] and CLIP VIT-L/14 Vision Encoder [45]. Adopting design principles from Flamingo [4], Octopus employs the Perceiver Resampler and Cross-Gated Attention modules to enhance vision-language synergy. This architecture enables Octopus to excel in tasks requiring understanding of both visual and textual data. The Octopus is also compatible with other VLMs such as LLaVA [37]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.643, + 0.705 + ], + "angle": 0, + "content": "4.3 SFT: Supervised Finetuning with Instructions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.787, + 0.792 + ], + "angle": 0, + "content": "We train the Octopus model on our collected dataset from OctoVerse \\(\\mathcal{D}_{\\mathrm{E}} = \\{(\\mathbf{X}_v,\\mathbf{T}_i,\\mathbf{T}_r)\\}\\) using token-level supervised fine-tuning (SFT) [42,54]. The Perceiver Resampler transforms images \\(\\mathbf{X}_v\\) into visual tokens that condition subsequent layers via Cross-Gated Attention modules. The training objective is next-token prediction, modeling the likelihood of a targeted response \\(\\mathbf{T}_r\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.802, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {T} _ {r} \\mid \\mathbf {T} _ {i}, \\mathbf {X} _ {v}\\right) = \\prod_ {l = 1} ^ {L} p \\left(t _ {l} \\mid \\mathbf {X} _ {v}, \\mathbf {T} _ {i}, \\mathbf {T} _ {r, < l}\\right). \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "Note that \\(\\mathbf{T}_i\\) denotes the instruction tokens and \\(\\mathbf{T}_{r, < l}\\) denotes the response tokens before the current predicted token \\(t_l\\). During inference, tokens are converted into natural language via the language decoder's text tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.283 + ], + "angle": 0, + "content": "Visual observations \\(\\mathbf{X}_v = \\{x_F^0,\\dots ,x_F^7,x_B^0,x_B^1\\}\\) consist of 8 first-person view (FPV) images and two bird's-eye view (BEV) images for OctoGibson and OctoGTA. OctoMC only takes 4 FPV images. The FPV captures the agent's direct observations, while the BEV provides a holistic understanding of the environment. The eight FPV images are captured every 45 degrees, ensuring a complete 360-degree perspective." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.302, + 0.773, + 0.318 + ], + "angle": 0, + "content": "4.4 RLEF: Reinforcement Learning with Environmental Feedback" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.784, + 0.383 + ], + "angle": 0, + "content": "In OctoVerse, task progression can be visualized as a tree (Fig. 4 (d)), where each node represents a sub-task with a binary value indicating success (1) or failure (0). If a node (or sub-task) has a value of 1, it is a step in the correct direction toward our end goal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.785, + 0.489 + ], + "angle": 0, + "content": "Tree-based Task Representation According to the environmental feedback part in Sec. 4.1, environmental reward datasets \\(\\mathcal{D}_{\\mathbb{R}} = (\\mathbf{X}_v^*,\\mathbf{T}_i^*,\\mathbf{T}_r^i,\\mathbf{T}_r^j,c)\\) are organized, where \\(\\mathbf{T}_r^i\\) and \\(\\mathbf{T}_r^j\\) are responses sharing the same parent task \\(\\mathbf{T}_i^*\\), and \\(c\\) indicates the preferred response leading to task completion. This ensures the reward mechanism favors the successfully executed branch. Note that even if a parental node does not have multiple responses, we can still assign feedback according to the rule in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.61 + ], + "angle": 0, + "content": "Reward Model Configuration A single-modal CodeLLaMA-7B model with an additional value head is fine-tuned on \\(\\mathcal{D}_{\\mathrm{R}}\\) as the reward model \\(r_{\\phi}\\). This text-based model assesses state transitions (\\(\\mathbf{T}_i^* \\rightarrow \\mathbf{T}_r^{i,j}\\)) to determine high-reward transitions, assisting the agent in task execution and completion. The rationale for using CodeLLaMA as the reward model is that evaluating rewards can be purely dependent on the textual output. Furthermore, CodeLLaMA's strong programming skills make it well-suited for assessing the quality and effectiveness of the generated code in the context of task completion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.671 + ], + "angle": 0, + "content": "Policy Model Development The supervised fine-tuned model serves as the initial policy model \\(\\pi^{\\mathrm{INIT}}\\) with fixed parameters. A duplicate model, \\(\\pi_{\\theta}^{\\mathrm{RL}}\\), is initialized and trained using Proximal Policy Optimization (PPO) [51] to maximize response rewards. The loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.677, + 0.823, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\pi_ {\\theta} ^ {\\mathrm {R L}}\\right) = - \\mathbb {E} _ {\\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right) \\in \\mathcal {D} _ {\\mathrm {R}}, \\mathbf {T} _ {r} \\sim \\pi^ {\\mathrm {R L}}} \\left[ r _ {\\phi} \\left(\\mathbf {T} _ {i} ^ {*}, \\mathbf {T} _ {r}\\right) - \\beta \\cdot \\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} ^ {\\mathrm {R L}} \\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right) \\| \\pi^ {\\mathrm {I N I T}} \\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right)\\right) \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.784, + 0.74 + ], + "angle": 0, + "content": "where \\(\\beta\\) acts as a hyper-parameter to regulate the magnitude of the Kullback-Leibler (KL) penalty." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.76, + 0.376, + 0.777 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.502, + 0.803 + ], + "angle": 0, + "content": "5.1 Main Results on OctoGibson" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Experimental Setup We first set up the OctoGibson to evaluate the performance of Octopus and other related models. Specifically, we are utilizing the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.242 + ], + "angle": 0, + "content": "Table 3: Main Results on OctoGibson. We compare various models: standalone language models, adapted vision-language planners, and our Octopus models, across different evaluation settings. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators. GT denotes that the model input is directly parsed from the simulator, with information on objects (O) or relations (R). Octopus shows consistent advance in task completion." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.242, + 0.785, + 0.376 + ], + "angle": 0, + "content": "
ModelVision ModelLanguage ModelEntire Goal Task
Seen EnvUnseen EnvFollowReasonAll
GPT-4--0.42 / 0.690.46 / 0.670.49 / 0.780.27 / 0.400.43 / 0.68
GPT-4V--0.40 / 0.620.60 / 0.670.42 / 0.670.53 / 0.530.45 / 0.63
LLaMAGT (O+R)LLaMA2-7B0.07 / 0.110.13 / 0.130.11 / 0.160.00 / 0.000.08 / 0.12
CodeLLaMAGT (O+R)CodeLLaMA-7B0.09 / 0.200.20 / 0.400.16 / 0.310.00 / 0.070.12 / 0.25
TAPA (task-level)QvD GT (O)CodeLLaMA-7B0.09 / 0.360.13 / 0.330.11 / 0.360.06 / 0.330.10 / 0.35
TAPA (step-level)QvD GT (O)CodeLLaMA-7B0.16 / 0.420.13 / 0.270.18 / 0.380.07 / 0.400.15 / 0.38
EmbodiedGPTCLIP-ViTMPT-7B0.04 / 0.360.27 / 0.530.13 / 0.380.00 / 0.400.10 / 0.40
Octopus (SFT Only)CLIP-ViTMPT-7B0.11 / 0.330.27 / 0.470.16 / 0.380.13 / 0.330.15 / 0.37
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.13 / 0.380.33 / 0.530.18 / 0.400.20 / 0.530.18 / 0.42
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.518 + ], + "angle": 0, + "content": "metrics of goal task completion score to check whether the task is completed in the simulator and the plan score from human evaluation. We have 60 evaluation tasks, with 45 from the seen environment, and 15 that are unseen during training. We also have 45 routine tasks and 15 require reasoning. Please note that models like Octopus might not always accurately identify specific object names as they appear in the simulator (e.g., \"water_bottle_189\"). To address this, we implement a post-processing step for the generated code, substituting generic object references with their exact names from the simulator with simple string similarity matching. If multiple objects, we select the one closest to the agent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.787, + 0.609 + ], + "angle": 0, + "content": "For Blind LLMs, we provide them with all the environment information in a textual format. Referring to Figure 3, we hope the Blind LLMs could perform as GPT-4 but internalize the system message. For TAPA utilizes the open-vocabulary detection (OVD) technique [65] to recognize objects within images and parse them into textual environmental messages, we still provide it with ground-truth environmental messages as an oracle setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "CodeLLaMA Improves Coding but not Planning. The first two rows in Table 3 highlight the suboptimal task completion rate of the blind LLMs. Among them, CodeLLaMA boasts pre-training on a large programming dataset, resulting in a notable enhancement in code execution from our observation, with \\(92\\%\\) of the written code being successfully executed compared to LLaMA's \\(24\\%\\). However, its prowess in planning remains limited. In contrast, the proposed Octopus MPT-7B model displays superior planning and task completion metrics while maintaining commendable coding abilities (\\(72\\%\\) of the written code can be executed). We surmise that the coding requirements within the OctoGibson environment might not be exceedingly intricate, rendering an advanced programming language model, like CodeLLaMA, less crucial, albeit beneficial. For more insight, although not shown in the table, our efforts to replace the MPT model with CodeLLaMA encountered challenges of generating non-sense outputs, suggesting that more refined code, or image-code paired data might be necessary for a successful Octopus-CodeLLaMA integration." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.389, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.146, + 0.58, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.146, + 0.783, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.285, + 0.381, + 0.296 + ], + "angle": 0, + "content": "(a) TAPA (step) + CodeLLaMA" + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.285, + 0.545, + 0.296 + ], + "angle": 0, + "content": "(b) Octopus (SFT only)" + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.285, + 0.759, + 0.296 + ], + "angle": 0, + "content": "(c) Octopus (SFT + RLEF)" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.298, + 0.566, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.299, + 0.782, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.437, + 0.541, + 0.448 + ], + "angle": 0, + "content": "(d) Vision Input for Vision-Language Models (e.g., b, c, e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.437, + 0.742, + 0.448 + ], + "angle": 0, + "content": "(e) GPT-4V Output" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.448, + 0.787, + 0.503 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Results on the task of find a carboy in OctoGibson environment. We show that the models shown can write executable code, but the proposed Octopus has stronger planning ability, especially after RLEF. We also explore the performance of GPT-4V on the specific task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.632 + ], + "angle": 0, + "content": "Blind LLMs Struggle with Extended Input Content. Our observations indicate that the step-level TAPA model, when supplied with a ground-truth object list, achieves a notable enhancement in planning. The primary distinction between it and the blind CodeLLaMA lies in the input length; the latter deals with protracted, pairwise relation content, complicating the language model's ability to extract crucial data from the environment message. This scenario highlights the inherent limitation of blind LLMs: relying on language alone to convey the entirety of environmental data can result in less informative input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.714 + ], + "angle": 0, + "content": "Octopus Demonstrates Superior Task Generalization. Table 3 underscores Octopus's strong performance, evidencing its consistent edge over standalone language models in task completion. Its adeptness in adapting to previously unencumbered environments underlines the inherent advantages of vision-language models. A more detailed ablation analysis is provided later." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "RLEF Enhances Octopus's Planning Strategy. Table 3 shows Octopus's strong reasoning capabilities after the RLEF finetuning. An example can be observed in Fig. 5(b-c), where, after refinement via RLEF, Octopus astutely navigates to the cabinet housing the carboy instead of attempting a direct yet distant capture. Quantitatively, Octopus exhibits enhanced adaptability to previously unseen reasoning tasks, reinforcing its prowess in logical task resolution. When juxtaposed with other strategies, such as the embodied queries employed by EmbodiedGPT, RLEF emerges as the more efficacious approach." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.408, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.246, + 0.407, + 0.256 + ], + "angle": 0, + "content": "(a) On Tuning Different Components" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.148, + 0.637, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.246, + 0.618, + 0.256 + ], + "angle": 0, + "content": "(b) On SFT & RLEF and Model Size" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.149, + 0.769, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.246, + 0.78, + 0.256 + ], + "angle": 0, + "content": "(c) On Necessity of Vision Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.785, + 0.295 + ], + "angle": 0, + "content": "Fig.6: Ablation Study on model components, model size, and vision input. For bars with different colors, the upper bar denotes the number of successful reasoning tasks, and the lower is routine tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.299, + 0.389, + 0.314 + ], + "angle": 0, + "content": "5.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.785, + 0.382 + ], + "angle": 0, + "content": "Tunning Different Components Fig. 6 (a) demonstrates that solely adjusting the connector (marked \"fire\" in Fig. 4 (a)) leads to success for merely 4 out of 60 tasks. Conversely, finetuning both the connector and language decoder nudges the success rate slightly higher, with 5 tasks being accomplished." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.785, + 0.476 + ], + "angle": 0, + "content": "7B v.s. 3B Model Size We embarked on experiments centered on model size to discern the influence of the total parameter count on the efficacy of vision-language models. As illustrated in Fig. 6 (b), downsizing the model manifests in a noticeable performance drop. The congruency of results across both the SFT and RLEF models underscores the importance of an apt model size when sculpting vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.787, + 0.645 + ], + "angle": 0, + "content": "Significance of Visual Inputs in Task Performance In our standard configuration, the vision component processes a sequence of image inputs, consisting of eight circularly captured first-person view (FPV) images, complemented by two bird's-eye view (BEV) images. With the intent to investigate the impact of visual inputs on task performance, we initiated an ablation study. In a modified setup, the sequence of these visual inputs was deliberately randomized, aiming to attenuate the strength of the visual signals. As illustrated in Fig. 6 (c), this intentional disruption in visual input consistency led to a pronounced decline in task performance. This result highlights the crucial role that clear and structured visual inputs play in the Octopus model, emphasizing that it significantly leverages visual cues for effective planning and task execution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.571, + 0.679 + ], + "angle": 0, + "content": "5.3 Results on Minecraft and GTA Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.795 + ], + "angle": 0, + "content": "Results on OctoMC According to the OctoMC part in Sec. 3, we designed 40 tasks, each task is operated on 2 locations so a total of 80 tasks. We set aside 10 tasks as unseen tasks and 10 tasks as seen tasks, getting 60 training tasks and 30 testing tasks in OctoMC. Similar to OctoGibson, the training data for OctoMC is collected using GPT-4. The agent, guided by GPT-4, explores the Minecraft environment and generates action plans and corresponding code based on the provided system messages, environmental cues, and objectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Table 4 shows that the SFT model trained on OctoMC can complete most tasks in both seen and unseen scenarios, demonstrating better performance compared to OctoGTA. However, upon analyzing the failure cases, we find that the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: Main Results for GTA and Minecraft Tasks. Despite limited training data, the Octopus still shows its good ability in task completion. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.309 + ], + "angle": 0, + "content": "
ModelVision ModelLanguage ModelOctoMCOctoGTA
Seen TaskUnseen TaskAllSeen TaskUnseen TaskAll
GPT-4--0.70 / 0.850.60 / 0.800.65 / 0.830.55 / 0.800.50 / 0.600.54 / 0.76
GPT-4V--0.75 / 0.850.70 / 0.850.73 / 0.850.58 / 0.850.50 / 0.700.56 / 0.82
EmbodiedGPTCLIP-ViTMPT-7B0.30 / 0.550.20 / 0.600.25 / 0.580.15 / 0.380.20 / 0.600.16 / 0.42
Octopus (SFT Only)CLIP-ViTMPT-7B0.30 / 0.600.20 / 0.700.25 / 0.650.18 / 0.480.20 / 0.600.18 / 0.50
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.40 / 0.600.20 / 0.700.30 / 0.650.18 / 0.530.30 / 0.700.20 / 0.56
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.784, + 0.388 + ], + "angle": 0, + "content": "model sometimes struggles with tasks requiring precise spatial reasoning. For instance, when tasked with killing a pig, the agent may have difficulty finding the creature with the exact angle and distance. While it shows that the agent relies on visual information to navigate and interact with the environment, it also means that even with correct planning, imprecise actions can lead to failure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Results on OctoGTA According to the OctoGTA part in Sec. 3, we designed 25 tasks. We set aside 5 tasks in the boat-related group (e.g., boat retrieval and shore return) as unseen tasks for testing only, using 2 different locations. We replicate the remaining 20 tasks for both training (8 different locations) and testing (2 locations). As a result, we have 160 training tasks and 50 testing tasks in OctoGTA. Unlike the training procedure for OctoGibson and OctoMC, the training data for OctoGTA is entirely created by the authors, as it is challenging to gather textual environmental messages in the GTA environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.789, + 0.665 + ], + "angle": 0, + "content": "Table 4 shows that, despite having only 160 training tasks, the SFT model can complete some tasks in both seen and unseen scenarios, and RLEF also outperforms. However, upon careful examination of the failure cases, we find that the model struggles with tasks that are not straightforward. For instance, when a wall separates the player from the car, the player still finds it difficult to decide to climb the wall, even if similar cases exist in the training data. Similar to OctoMC, as illustrated in Sec. 3, when approaching certain locations, the code involves functions like turnPlayer() and goForward() rather than a simple walkTo location). Consequently, even with correct planning, imprecise actual actions can still lead to task failure." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.36, + 0.72 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper introduces Octopus, an embodied vision-language programmer designed to bridge the gap between high-level planning and real-world manipulation with programming. By open-sourcing our OctoVerse environments, dataset, and Octopus architecture, we aim to foster collaboration and innovation within the research community, paving the way for future developments in embodied vision-language programming." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.652, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments and Disclosure of Funding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.238 + ], + "angle": 0, + "content": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221- 0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.259, + 0.323, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.42, + 0.303 + ], + "angle": 0, + "content": "1. Grand theft auto v (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.357, + 0.316 + ], + "angle": 0, + "content": "2. Minecraft (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.786, + 0.427 + ], + "angle": 0, + "content": "3. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Ho, D., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E., Ruano, R.J., Jeffrey, K., Jesmonth, S., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Lee, K.H., Levine, S., Lu, Y., Luu, L., Parada, C., Pastor, P., Quiambao, J., Rao, K., Rettinghouse, J., Reyes, D., Sermanet, P., Sievers, N., Tan, C., Toshev, A., Vanhoucke, V., Xia, F., Xiao, T., Xu, P., Xu, S., Yan, M., Zeng, A.: Do as i can and not as i say: Grounding language in robotic affordances. In: arXiv preprint arXiv:2204.01691 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.428, + 0.786, + 0.482 + ], + "angle": 0, + "content": "4. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.786, + 0.55 + ], + "angle": 0, + "content": "5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.786, + 0.592 + ], + "angle": 0, + "content": "6. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.786, + 0.648 + ], + "angle": 0, + "content": "7. Bellemare, M.G., Naddaf, Y., Veness, J., Bowling, M.: The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research 47, 253-279 (Jun 2013). https://doi.org/10.1613/jair.3912, http://dx.doi.org/10.1613/jair.3912" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.786, + 0.674 + ], + "angle": 0, + "content": "8. Billard, A., Kragic, D.: Trends and challenges in robot manipulation. Science 364(6446), eaat8414 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.786, + 0.702 + ], + "angle": 0, + "content": "9. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., Zaremba, W.: Openai gym. arXiv preprint arXiv:1606.01540 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.786, + 0.758 + ], + "angle": 0, + "content": "10. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "12. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Chen, J., Mu, Y., Yu, Q., Wei, T., Wu, S., Yuan, Z., Liang, Z., Yang, C., Zhang, K., Shao, W., Qiao, Y., Xu, H., Ding, M., Luo, P.: Roboscript: Code generation for free-form manipulation tasks across real and simulation (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "14. Chen, L., Li, B., Shen, S., Yang, J., Li, C., Keutzer, K., Darrell, T., Liu, Z.: Language models are visual reasoning coordinators. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "15. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.785, + 0.326 + ], + "angle": 0, + "content": "16. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.326, + 0.785, + 0.353 + ], + "angle": 0, + "content": "17. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.353, + 0.785, + 0.421 + ], + "angle": 0, + "content": "18. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model. In: arXiv preprint arXiv:2303.03378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.421, + 0.785, + 0.448 + ], + "angle": 0, + "content": "19. Evans, J.S.B.: Dual-processing accounts of reasoning, judgment, and social cognition. Annu. Rev. Psychol. 59, 255-278 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.448, + 0.785, + 0.489 + ], + "angle": 0, + "content": "20. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.489, + 0.785, + 0.516 + ], + "angle": 0, + "content": "21. Fu, H., Xu, W., Ye, R., Xue, H., Yu, Z., Tang, T., Li, Y., Du, W., Zhang, J., Lu, C.: Rfuniverse: A multiphysics simulation platform for embodied ai (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.516, + 0.785, + 0.543 + ], + "angle": 0, + "content": "22. Fu, Z., Zhao, T.Z., Finn, C.: Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In: arXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.543, + 0.785, + 0.569 + ], + "angle": 0, + "content": "23. Gao, X., Gong, R., Shu, T., Xie, X., Wang, S., Zhu, S.C.: Vrkitchen: an interactive 3d virtual environment for task-oriented learning (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.611 + ], + "angle": 0, + "content": "24. Gu, S., Holly, E., Lillicrap, T., Levine, S.: Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 3389-3396. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.611, + 0.785, + 0.651 + ], + "angle": 0, + "content": "25. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.785, + 0.692 + ], + "angle": 0, + "content": "26. Huang, S., Jiang, Z., Dong, H., Qiao, Y., Gao, P., Li, H.: Instruct2act: Mapping multi-modality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.692, + 0.785, + 0.733 + ], + "angle": 0, + "content": "27. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "28. Hussein, A., Gaber, M.M., Elyan, E., Jayne, C.: Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR) 50(2), 1-35 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.632, + 0.773 + ], + "angle": 0, + "content": "29. Kahneman, D.: Thinking, fast and slow. macmillan (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.773, + 0.785, + 0.813 + ], + "angle": 0, + "content": "30. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Octopus: Embodied VLM from Environmental Feedback" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "32. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Conference on Robot Learning. pp. 80–93. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "33. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.247, + 0.786, + 0.275 + ], + "angle": 0, + "content": "34. Lin, K., Agia, C., Migimatsu, T., Pavone, M., Bohg, J.: Text2motion: From natural language instructions to feasible plans. arXiv preprint arXiv:2303.12153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.276, + 0.787, + 0.302 + ], + "angle": 0, + "content": "35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "36. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.346, + 0.69, + 0.361 + ], + "angle": 0, + "content": "37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.362, + 0.787, + 0.389 + ], + "angle": 0, + "content": "38. Liu, Z., Dong, Y., Rao, Y., Zhou, J., Lu, J.: Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.39, + 0.787, + 0.417 + ], + "angle": 0, + "content": "39. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.418, + 0.787, + 0.445 + ], + "angle": 0, + "content": "40. MosaicML: Mpt-7b (2023), https://www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-23" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.446, + 0.787, + 0.488 + ], + "angle": 0, + "content": "41. Mu, Y., Zhang, Q., Hu, M., Wang, W., Ding, M., Jin, J., Wang, B., Dai, J., Qiao, Y., Luo, P.: Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.489, + 0.787, + 0.543 + ], + "angle": 0, + "content": "42. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.545, + 0.787, + 0.587 + ], + "angle": 0, + "content": "43. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.588, + 0.787, + 0.629 + ], + "angle": 0, + "content": "44. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtual-home: Simulating household activities via programs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8494-8502 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.63, + 0.787, + 0.685 + ], + "angle": 0, + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.686, + 0.787, + 0.714 + ], + "angle": 0, + "content": "46. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.715, + 0.787, + 0.742 + ], + "angle": 0, + "content": "47. Ramakrishnan, S.K., Jayaraman, D., Grauman, K.: An exploration of embodied visual exploration (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.743, + 0.787, + 0.785 + ], + "angle": 0, + "content": "48. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable task planning. arXiv preprint arXiv:2307.06135 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.787, + 0.841 + ], + "angle": 0, + "content": "49. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "50. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "51. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "52. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.287 + ], + "angle": 0, + "content": "53. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "55. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "56. Wang, H., Liang, W., Gool, L.V., Wang, W.: Towards versatile embodied navigation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.425 + ], + "angle": 0, + "content": "57. Wu, Y., Wu, Y., Gkioxari, G., Tian, Y.: Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "58. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.785, + 0.494 + ], + "angle": 0, + "content": "59. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: Sapien: A simulated part-based interactive environment (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "60. Xie, B., Zhang, S., Zhou, Z., Li, B., Zhang, Y., Hessel, J., Yang, J., Liu, Z.: Funqa: Towards surprising video comprehension. arXiv preprint arXiv:2306.14899 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "61. Yan, C., Misra, D., Bennnett, A., Walsman, A., Bisk, Y., Artzi, Y.: Chalet: Cornell house agent learning environment (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "62. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Arenas, M.G., Chiang, H.T.L., Erez, T., Hasenclever, L., Humplik, J., et al.: Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "63. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Skill reinforcement learning and planning for open-world long-horizon tasks (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.647 + ], + "angle": 0, + "content": "64. Zheng, S., Liu, J., Feng, Y., Lu, Z.: Steve-eye: Equipping llm-based embodied agents with visual perception in open worlds (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "65. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "66. Zhu, Y., Wong, J., Mandlekar, A., Martin-Martín, R., Joshi, A., Nasiriany, S., Zhu, Y.: robosite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_origin.pdf b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b87ff262f94ea609e0965c7d1e4f484b073548a8 --- /dev/null +++ b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/e3940d59-4ec7-4868-b252-33e7cda42131_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f0245542e0e1d486172a8a41ad0d2efa007b589c62194d43fd5e4f268f7e035 +size 6990601 diff --git a/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/full.md b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a0a9a7dce15bb8fee487c9015372019a3e36612 --- /dev/null +++ b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/full.md @@ -0,0 +1,318 @@ +# Octopus: Embodied Vision-Language Programmer from Environmental Feedback + +Jingkang Yang\*,1, Yuhao Dong\*,2,3, Shuai Liu\*,2,4, Bo Li\*,1, Ziyue Wang\*,1, Haoran Tan\*,4, Chencheng Jiang\*,5, Jiamu Kang\*,3, Yuanhan Zhang\*, Kaiyang Zhou\* and and Ziwei Liu\* + +$^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai AI Laboratory $^{3}$ Tsinghua University $^{4}$ BUPT $^{5}$ XJTU $^{6}$ Hong Kong Baptist University $\{\text{jingkang001, ziwei.liu}\}@\text{ntu.edu.sg}$ + +![](images/6e1fffbae234e11e03cf9b934aa2f24e859d1db42aff5e8724d5471c83bc1950.jpg) +Fig. 1: Illustration of our vision-language programmer, Octopus, complete a task in GTA environment. Given a task in the form of natural language, Octopus relies on its egocentric vision to generate plans and the corresponding executable code. + +Abstract. Large vision-language models (VLMs) have achieved substantial progress in multimodal perception and reasoning. When integrated into an embodied agent, existing embodied VLM works either output detailed action sequences at the manipulation level or only provide plans at an abstract level, leaving a gap between high-level planning and real-world manipulation. To bridge this gap, we introduce Octopus, an embodied vision-language programmer that uses executable code generation as a medium to connect planning and manipulation. Octopus is designed to 1) proficiently comprehend an agent's visual and textual task objectives, 2) formulate intricate action sequences, and 3) generate executable code. To facilitate Octopus model development, we introduce OctoVerse: a suite of environments tailored for benchmarking + +vision-based code generators on a wide spectrum of tasks, ranging from mundane daily chores in simulators to sophisticated interactions in complex video games such as Grand Theft Auto (GTA) and Minecraft. To train Octopus, we leverage GPT-4 to control an explorative agent that generates training data, i.e., action blueprints and corresponding executable code. We also collect feedback that enables an enhanced training scheme called Reinforcement Learning with Environmental Feedback (RLEF). Through a series of experiments, we demonstrate Octopus's functionality and present compelling results, showing that the proposed RLEF refines the agent's decision-making. By open-sourcing our simulation environments, dataset, and model architecture, we aspire to ignite further innovation and foster collaborative applications within the broader embodied AI community. The project page is available at https://choiszt.github.io/0ctopus/. + +# 1 Introduction + +The rise of large language models (LLMs) [11,15,42,46,54] led to a surge in vision-language models (VLMs) [4,5,31,33,35-37], enabling tasks such as image/video-based descriptions [33], reasoning [14,38,60], and conversations [16,31]. In the realm of embodied AI, notable efforts [10,18,33] have trained agents to process visual input and relay motor control commands. + +Another approach to interacting with the environment focuses on task execution through code invocations, mirroring the human System-I stimulation [19,29] (automatic, intuitive actions) with predefined code, and leaving the System-II processes [19,29] (planning and reasoning) for large models. For example, referring to Fig. 1, planning a car ride with a pet might entail a subconscious checklist (e.g., getOutOf() the house, open() the car door), each action could be implemented using specific techniques [8,24] such as imitation learning [22,28]. This programmatic paradigm has been, although not in vision, leveraged by works [25, 50, 52, 53] using LLMs to craft programs and trigger APIs. Game-centric models like Voyager [55] have similarly employed GPT for function calls within game engines, though they often parse data directly from their environments. + +However, when incorporating visual perception, the programming paradigms are largely unexplored. Primary initiatives [48,58] can only output plans, which anchor their strategies only in initial environmental states or employ dynamic scene graphs for LLM inputs, respectively. Despite their innovations, the overreliance on pre-trained vision models to convert vision content into language can occasionally hinder the LLM's planning performance. The conversion from plans into real-world actions is still missing. While EmbodiedGPT [41] addresses the problem by integrating vision-language modeling for planning and then transitioning to manipulation using policy mapping, the capability of embodied vision-language models to generate executable programs remains largely uncharted. + +Our exploration aims to bridge this gap. An embodied vision-language programmer should integrate visual perspective with textual objectives to devise + +action plans and executable code (Fig. 1). However, existing simulator environments often lack the carefully designed functions necessary to support such models effectively. These functions should balance usefulness and complexity to avoid hindering the development of genuine embodied vision-language programmers. For example, the **explore_until()** function in Minecraft, which can lead the player directly to specific blocks without relying on vision information, may not be suitable for training these models. + +To meet the requirement, we carefully design and develop OctoVerse, a suite of environments consisting of diverse simulators, including (i) OctoGibson, built upon the photorealistic OmniGibson [32], (ii) OctoMC, developed on the infinitely creative, pixel-style Minecraft platform [2], and (iii) OctoGTA, adapted from the highly interactive and immersive Grand Theft Auto V (GTA-V) [1]. These environments enable the training and benchmarking of our embodied vision-language programming model in a wide range of scenarios, from daily household tasks to complex urban navigation and open-world exploration, while the function calls are tailored to be vision-dependent. + +Using the OctoVerse environment, we train Octopus by leveraging GPT-4 to collect data. We provide GPT-4 with system messages, environmental cues, and objectives, enabling it to formulate action strategies and code. Simultaneously, the agent captures visual perspectives, forming the image-code pair for Octopus training. During data collection, the agent receives simulator feedback, distinguishing successful moves from unsuccessful ones. We incorporate this feedback using Reinforcement Learning with Environmental Feedback (RLEF) and fine-tune Octopus using Proximal Policy Optimization (PPO) [51]. Empirically, Octopus demonstrates strong adaptability in various scenarios, outperforming existing models in task planning, code generation, and execution. The integration of RLEF further enhances Octopus's performance, showcasing the effectiveness of this training approach. In sum, our key contributions include: + +- A Novel Vision-Language Programming Benchmark: Three diverse embodied environments with designed tasks: (i) OctoGibson, which is developed upon OmniGibson [32], (ii) OctoMC that developed on Minecraft [2], and (iii) OctoGTA, which is adapted from GTA-V [1]. +- A New Vision-Language Programming Model: An embodied vision-language planner and programmer trained with Reinforcement Learning with Environmental Feedback (RLEF), demonstrates compelling results. +- Insights on Vision-Language Programming: We extensively explore Octopus and share useful insights facilitating future research on visual planning and programming. + +# 2 Related Work + +# 2.1 Embodied AI Simulators + +Embodied AI has advanced significantly with the development of diverse simulation environments, enabling research tasks such as visual exploration [47], + +Table 1: Related Work for OctoVerse - Overview of Embodied AI Environments. We select three environments into OctoVerse and carefully design executable tasks and vision-dependent function calls (VC), in comparison to undesigned standard function calls (C). + +
SimulatorKinematicsContinuous +Extended +StatesFlexible +MaterialsDeformable +BodiesRealistic +Action +ExecutionGame- or +World-BasedFormulated +TasksFunction +Call +Type
OpenAIGym [9]✓×××✓G×C
Matterport3D [12]×××××W××
AI2THOR [30]✓××××G×C
VirtualHome [44]×××××G××
House3D [57]×××××W××
Habitat 1.0 [49]✓×××✓W×C
Robosuite [66]✓×××✓W×C
RFUniverse [21]✓×✓✓✓W×C
Minecraft [2]✓×✓×✓G×C
OctoMC✓×✓×✓G✓VC
GTA [1]✓✓✓✓✓G×C
OctoGTA✓✓✓✓✓G✓VC
OmniGibson [32]✓✓✓✓✓W×C
OctoGibson✓✓✓✓✓W✓VC
+ +navigation [56], and question-answering [17]. Several simulators, including AI2-THOR [30], VirtualHome [44], Habitat-Sim [49], SAPIEN [59], and Omnigibson [32], provide realistic representations of the world for investigating embodied AI challenges. OmniGibson [32] stands out for its high-fidelity simulation of diverse indoor and outdoor environments. OctoGibson environment further enhances OmniGibson with carefully designed function calls and formulated tasks, making it well-suited for vision-language programming. + +Game-related simulators like Arade [7], CHALET [61], and VRKitchen [23] also contribute significantly to embodied AI. Minecraft [2] has gained attention in reinforcement learning and game agents [6,20,39,55,63,64] but lacks the necessary structure for vision-language programming. OctoMC addresses this by providing designed function calls and formulated tasks. In contrast to Minecraft's voxel-based representations that limit transferability to real-world environments, GTA-V [1] offers a highly realistic environment. In this work, we introduce OctoGTA as a new setting, leveraging GTA-V's rich, open-world environment with incorporated tasks and function calls, extending this platform for embodied AI study. + +# 2.2 Embodied AI with Large Models + +The recent wave of research focuses on merging LLMs with embodied AI tasks [11, 42,46,54]. For instance, VoxPoser addresses robotic manipulation problems through unsupervised methods [27]. A group of projects, namely SayCan [3], Palm-E [18], RT-2 [10], and EmbodiedGPT [41], effectively integrate visual or linguistic cues with robot manipulation data. Outside the domain of robotic manipulation, initiatives like Voyager [55] and Smallville [43] harness the capabilities of GPT to interface with game functions, relying on preset functions to manage intricate manipulations. In a parallel vein, VisProg [25] leverages GPT-3 language prompts to craft Python programs, opening the door to a multitude of fascinating applications. While the proposed Octopus model also formulates plans + +Table 2: Related Work for Octopus - Overviewing Embodied AI Models. The proposed Octopus distinguishes itself from other models as a unified vision-language model for both plan and code generation. + +
ModelsRelease DateSupported EnvironmentVision ModelCode GeneratorAction w/ FeedbackLLM Training Enabled
Text2Motion [34]Mar. 2023Sim×✓×
Instruct2Act [26]May 2023Sim×✓×
Lang2Rewards [62]Jun. 2023Sim×✓×
VoxPoser [27]Jul. 2023Sim✓××
SayCan [3]Apr. 2022Real✓××
PALM-E [18]Mar. 2023Sim, Real✓×✓
RT-2 [10]Jul. 2023Real✓×✓
SayPlan [48]Jun. 2023Real×××
EmbodiedGPT [41]May 2023Sim✓×✓
TaPA [58]Jul. 2023Sim××✓
Voyager [55]May 2023Game×✓×
Steve-Eye [64]Dec 2023Game✓×✓
RoboScript [13]Feb 2024Sim, Real×✓×
Octopus-Sim, Game✓✓✓
+ +and code, its distinguishing feature is the seamless integration of visual input in program and code generation. This also stands in contrast to other embodied planners like TAPA [58] and SayPlan [48], which deploy separate vision modules to translate visual data into linguistic inputs for LLMs. Octopus excels as a cohesive vision-language model, delivering not just plans but also code. + +More discussion on additional related works (Vision Language Model, and Feedback in Large Language Models) is included in the supplementary materials. + +# 3 The OctoVerse Environment + +In this section, we introduce three simulator environments designed to train and evaluate the Octopus model. For each environment, we will describe their overall information, the special design considerations that ensure the tasks are well-formulated and the callable functions are vision-dependent. + +OctoGibson We built the environment on the foundation of OmniGibson [32], an existing simulation framework that supports 1,000 daily activities across 50 scenes, featuring over 5,000 meticulously annotated objects. We incorporated 16 functions that the robot can execute, such as moveBot() and easyGrasp(). Within this environment, we meticulously crafted 476 tasks1, each with a well-defined initial state and a definitive termination state, allowing for a straightforward assessment of task completion. Among these tasks, 367 are routine tasks—simple and direct actions like "place a glass in a trash can," marked as Follow. Conversely, the remaining 109 are reasoning tasks that necessitate deeper comprehension. An example is "buy a chocolate," where the agent needs + +to know to pick a chocolate bar from the shelf and then place it, along with money, on the checkout counter, denoted as Reason tasks. + +To ensure our tasks are vision-aware, we deliberately constrain the usage of certain functions, such as moveBot(object), which moves the agent in front of the given object. To avoid making the task too easy and vision-agnostic, we limit the given parameter to a predefined set of large objects, such as tables and cupboards, rather than small items like cups and glasses. In this case, if the robot wants to pick up a cup, it needs to recognize whether the cup is on the table or in the cupboard. A simple moveBot(object) call with an inappropriate parameter would cause a runtime error. The full list of the functions is in the Appendix. + +OctoMC The OctoMC environment is built on Minecraft [2], a popular platform for reinforcement learning and game agents. We integrated 6 functional actions and crafted 40 tasks $^{2}$ , each designed to facilitate comprehensive observations and executions by the agent. These tasks are distributed across 10 different biomes, including indoor, outdoor, and underground settings, under varying weather conditions. + +However, existing Minecraft environments often lack vision-formulated tasks and the necessary structure for vision-language programming. For example, in Voyager [55], the exploreUntil() function allows the player to navigate directly to specific blocks without relying on vision information. This function works by randomly exploring the environment within a certain range until the desired object is found, at which point it returns a true value, enabling the agent to interact with the located object, such as gold blocks or trees. While effective, this approach is entirely automated and does not utilize visual information, making it unsuitable for our vision-based objectives. + +To address this limitation, we crafted a vision-dependent exploration function, teleport(yaw, distance), which operates within the robot's perceptual range. This function ensures that the agent's operations are vision-dependent and require active perception and navigation within the environment. For instance, to locate a tree block, the agent must actively navigate towards the direction of the forest, relying on visual cues to guide its exploration. + +OctoGTA The OctoGTA environment is built on GTA-V [1] with the help of the active GTA modding community. We integrated 19 functions and methodically crafted 25 tasks3, such as "help NPC drive their boat back to shore" and "mediate a fight between two NPCs," spanning across 5 groups (shown in Fig. 2 (b)). Each task is assigned to 5 different locations within the game world. + +We have implemented a set of functions that enable the agent to interact with the game world in a visually-aware manner. Similar to the design in Minecraft, we get rid of functions like walkTo location) that might trivialize the task of reaching a particular building or landmark. Instead, we provide functions such + +![](images/76f8463ebcdb002f43f190e5dcf9ce9a13839fdab317d9567151f0cc97fd2181.jpg) +Fig. 2: The Statistics of the OctoVerse Environment with Function Designs. + +as goForward(distance) and turnPlayer(degree) (Fig. 2 (b)). For tasks like mediating a fight between two NPCs, the essential function stopFight() only works when the player is within 5 meters of the fighting NPCs. These design choices ensure that the agent's operations are vision-dependent and require active perception and navigation within the environment. + +# 4 Octopus: The Embodied Vision-Language Programmer + +In this section, we present the procedure for training Octopus. Starting from collecting training data within the OctoVerse environment, Octopus builds upon a VLM architecture of Otter [31] and includes specialized RLEF modules to handle vision-language programming tasks. Fig. 4 illustrates the entire Octopus training pipeline. + +# 4.1 Training Data Collection + +We use the automatic training data collection pipeline described here for OctoGibson and OctoMC, with the latter using customized prompts inspired by Voyager [55]. For OctoGTA, we rely on human labor to hand-craft the training dataset due to the difficulty of obtaining textual environment messages in the GTA environment. In the following parts, we use OctoGibson as the primary example to illustrate the data collection pipeline. Note that the primary task in organizing training data is to form a succinct pairing: "vision input + current/historical states $\rightarrow$ next step plan + executable code". + +Environment Info Collection As shown in Fig. 4 (a) and Fig. 3, we format an environment message for each state, encompassing attributes like Observed Objects, Observed Relations, Inventory, and more. Specifically, the simulator can provide us with an exact scene graph at each state, shaping the content for the first two parts. The inventory info is also accessible. The task, e.g., "cooking bacon" in Fig. 3, is represented by the Task Goal. + +Automation with GPT-4 After preparing the environment message, we crafted a structured system message to ensure that the robot not only understands its input but also maintains a consistent output format. A detailed + +![](images/6cdd976d7b1070035e63f9f0eb63fef5fd12684a6160ac0c712a8b05aaf85f52.jpg) +Fig.3: Data Collection Example for "Cook a Bacon" Task. GPT-4 perceives the environment through the environmental message and produces anticipated plans and code following the detailed system message. This code is subsequently executed in the simulator, directing the agent to the subsequent state. For each state, we gather the environmental message, wherein observed objects and relations are substituted by egocentric images to serve as the training input. The response from GPT-4 acts as the training output. Environmental feedback, specifically the determination of whether each target state is met, is documented for RLEF training. + +examination of this prompt can be found in the appendix. Experiments have shown that a well-articulated prompt enables GPT-4 to effectively generate executable code. It is important to note that the combined length of the system and environment messages can be extremely long, which may cause standard GPT-4 8K models to struggle with producing meaningful outputs. To address this issue, we employ the more robust GPT-4 32K model. As illustrated in Fig. 3, when GPT-4 receives a consistent system and environment message, it generates comprehensive outputs that include current scenario analysis, planning, and actionable code, supporting the training process in Section 4.3. + +Error Management Notably, GPT-4 collects training data under the main task of guiding the agent to complete tasks. However, GPT-4 is not infallible. Errors can manifest in multiple ways, ranging from syntax errors to physical challenges in the simulator. For instance, in Fig. 3, between states $\#5$ and $\#6$ , the action failed due to the long distance between the agent (holding bacon) and the pan. Such setbacks reset the task to its previous state. If a main task remains incomplete after 10 steps, it is deemed unsuccessful, and we terminate this task for budget concerns. However, all data pairs without syntax errors, regardless of the task's completion status, are valuable for refining instructions and improving the model's performance. + +Environmental Feedback GPT-4's continual trial-and-error approach while guiding the agent toward task completion serves a dual purpose: collecting vision- + +![](images/e7e886cf6a620a9c98069660b102aace976d3cf7ca9f6b7a9e3ddf03c6685308.jpg) +(a) Data Collection Pipeline + +![](images/d0ff0b0f34aadd20b48a71ac7fdbaec50a85b8bdbcf92072cdeb9a257db8ea84.jpg) +(b) Octopus Training Pipeline + +![](images/2b87a40202ac8da67bda6e6d35469e9fe011aa49fa4927b481cae1065f0d6289.jpg) +(c) Octopus (Otter model) + +![](images/41f47cd6d088c512b2b154383379f89ea6c4edf682b2cf99beebf2161af803f9.jpg) + +![](images/65b407683e8cb947939a9954f9b85688b2bf07e5c7b7867d83395650594fc61d.jpg) +Step 2: Training Reward Model + +![](images/d0606d05feeaff5577056bd7d0673252451a64569f06890ba05703f7fdf75430.jpg) +(d) RLEF Training +Fig. 4: How to train Octopus: data collection and training pipeline. + +output pairs and generating a rich set of feedback data. The automatic annotation of this feedback focuses on two levels: step-level and task-level judgments. Step-level judgment assesses the alignment of post-execution states with their target states. For instance, in Fig. 3, steps color-coded in green lead to positive feedback. One can visualize the action sequence for task completion as a tree, where each node indicates a step (subtask), encapsulating an action code. Accompanying each step is a binary value that denotes success or failure, giving preference to the successful branch over its counterpart. Task-level judgment, on the other hand, gauges the successful execution of the overall task. If the task is not completed as intended, every state within that task is labeled as negative, regardless of the status of the subtasks. This collated feedback data serves as a foundation for our Reinforcement Learning with Environmental Feedback (RLEF) methodology, which we discuss in greater detail in Section 4.4. + +# 4.2 Model Architecture + +The Octopus architecture (shown in Fig. 4 (c)) is heavily inspired by the Otter model [31], integrates the MPT-7B Language Decoder [40] and CLIP VIT-L/14 Vision Encoder [45]. Adopting design principles from Flamingo [4], Octopus employs the Perceiver Resampler and Cross-Gated Attention modules to enhance vision-language synergy. This architecture enables Octopus to excel in tasks requiring understanding of both visual and textual data. The Octopus is also compatible with other VLMs such as LLaVA [37]. + +# 4.3 SFT: Supervised Finetuning with Instructions + +We train the Octopus model on our collected dataset from OctoVerse $\mathcal{D}_{\mathrm{E}} = \{(\mathbf{X}_v,\mathbf{T}_i,\mathbf{T}_r)\}$ using token-level supervised fine-tuning (SFT) [42,54]. The Perceiver Resampler transforms images $\mathbf{X}_v$ into visual tokens that condition subsequent layers via Cross-Gated Attention modules. The training objective is next-token prediction, modeling the likelihood of a targeted response $\mathbf{T}_r$ as: + +$$ +p \left(\mathbf {T} _ {r} \mid \mathbf {T} _ {i}, \mathbf {X} _ {v}\right) = \prod_ {l = 1} ^ {L} p \left(t _ {l} \mid \mathbf {X} _ {v}, \mathbf {T} _ {i}, \mathbf {T} _ {r, < l}\right). \tag {1} +$$ + +Note that $\mathbf{T}_i$ denotes the instruction tokens and $\mathbf{T}_{r, < l}$ denotes the response tokens before the current predicted token $t_l$ . During inference, tokens are converted into natural language via the language decoder's text tokenizer. + +Visual observations $\mathbf{X}_v = \{x_F^0,\dots ,x_F^7,x_B^0,x_B^1\}$ consist of 8 first-person view (FPV) images and two bird's-eye view (BEV) images for OctoGibson and OctoGTA. OctoMC only takes 4 FPV images. The FPV captures the agent's direct observations, while the BEV provides a holistic understanding of the environment. The eight FPV images are captured every 45 degrees, ensuring a complete 360-degree perspective. + +# 4.4 RLEF: Reinforcement Learning with Environmental Feedback + +In OctoVerse, task progression can be visualized as a tree (Fig. 4 (d)), where each node represents a sub-task with a binary value indicating success (1) or failure (0). If a node (or sub-task) has a value of 1, it is a step in the correct direction toward our end goal. + +Tree-based Task Representation According to the environmental feedback part in Sec. 4.1, environmental reward datasets $\mathcal{D}_{\mathbb{R}} = (\mathbf{X}_v^*,\mathbf{T}_i^*,\mathbf{T}_r^i,\mathbf{T}_r^j,c)$ are organized, where $\mathbf{T}_r^i$ and $\mathbf{T}_r^j$ are responses sharing the same parent task $\mathbf{T}_i^*$ , and $c$ indicates the preferred response leading to task completion. This ensures the reward mechanism favors the successfully executed branch. Note that even if a parental node does not have multiple responses, we can still assign feedback according to the rule in Sec. 4.1. + +Reward Model Configuration A single-modal CodeLLaMA-7B model with an additional value head is fine-tuned on $\mathcal{D}_{\mathrm{R}}$ as the reward model $r_{\phi}$ . This text-based model assesses state transitions ( $\mathbf{T}_i^* \rightarrow \mathbf{T}_r^{i,j}$ ) to determine high-reward transitions, assisting the agent in task execution and completion. The rationale for using CodeLLaMA as the reward model is that evaluating rewards can be purely dependent on the textual output. Furthermore, CodeLLaMA's strong programming skills make it well-suited for assessing the quality and effectiveness of the generated code in the context of task completion. + +Policy Model Development The supervised fine-tuned model serves as the initial policy model $\pi^{\mathrm{INIT}}$ with fixed parameters. A duplicate model, $\pi_{\theta}^{\mathrm{RL}}$ , is initialized and trained using Proximal Policy Optimization (PPO) [51] to maximize response rewards. The loss function is: + +$$ +\mathcal {L} \left(\pi_ {\theta} ^ {\mathrm {R L}}\right) = - \mathbb {E} _ {\left(\mathbf {X} _ {v} ^ {*}, \mathbf {T} _ {i} ^ {*}\right) \in \mathcal {D} _ {\mathrm {R}}, \mathbf {T} _ {r} \sim \pi^ {\mathrm {R L}}} \left[ r _ {\phi} \left(\mathbf {T} _ {i} ^ {*}, \mathbf {T} _ {r}\right) - \beta \cdot \mathbb {D} _ {\mathrm {K L}} \left(\pi_ {\theta} ^ {\mathrm {R L}} \left(\mathbf {X} _ {v} ^ {*}, \mathbf {T} _ {i} ^ {*}\right) \| \pi^ {\mathrm {I N I T}} \left(\mathbf {X} _ {v} ^ {*}, \mathbf {T} _ {i} ^ {*}\right)\right) \right], +$$ + +where $\beta$ acts as a hyper-parameter to regulate the magnitude of the Kullback-Leibler (KL) penalty. + +# 5 Experiments + +# 5.1 Main Results on OctoGibson + +Experimental Setup We first set up the OctoGibson to evaluate the performance of Octopus and other related models. Specifically, we are utilizing the + +Table 3: Main Results on OctoGibson. We compare various models: standalone language models, adapted vision-language planners, and our Octopus models, across different evaluation settings. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators. GT denotes that the model input is directly parsed from the simulator, with information on objects (O) or relations (R). Octopus shows consistent advance in task completion. + +
ModelVision ModelLanguage ModelEntire Goal Task
Seen EnvUnseen EnvFollowReasonAll
GPT-4--0.42 / 0.690.46 / 0.670.49 / 0.780.27 / 0.400.43 / 0.68
GPT-4V--0.40 / 0.620.60 / 0.670.42 / 0.670.53 / 0.530.45 / 0.63
LLaMAGT (O+R)LLaMA2-7B0.07 / 0.110.13 / 0.130.11 / 0.160.00 / 0.000.08 / 0.12
CodeLLaMAGT (O+R)CodeLLaMA-7B0.09 / 0.200.20 / 0.400.16 / 0.310.00 / 0.070.12 / 0.25
TAPA (task-level)QvD GT (O)CodeLLaMA-7B0.09 / 0.360.13 / 0.330.11 / 0.360.06 / 0.330.10 / 0.35
TAPA (step-level)QvD GT (O)CodeLLaMA-7B0.16 / 0.420.13 / 0.270.18 / 0.380.07 / 0.400.15 / 0.38
EmbodiedGPTCLIP-ViTMPT-7B0.04 / 0.360.27 / 0.530.13 / 0.380.00 / 0.400.10 / 0.40
Octopus (SFT Only)CLIP-ViTMPT-7B0.11 / 0.330.27 / 0.470.16 / 0.380.13 / 0.330.15 / 0.37
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.13 / 0.380.33 / 0.530.18 / 0.400.20 / 0.530.18 / 0.42
+ +metrics of goal task completion score to check whether the task is completed in the simulator and the plan score from human evaluation. We have 60 evaluation tasks, with 45 from the seen environment, and 15 that are unseen during training. We also have 45 routine tasks and 15 require reasoning. Please note that models like Octopus might not always accurately identify specific object names as they appear in the simulator (e.g., "water_bottle_189"). To address this, we implement a post-processing step for the generated code, substituting generic object references with their exact names from the simulator with simple string similarity matching. If multiple objects, we select the one closest to the agent. + +For Blind LLMs, we provide them with all the environment information in a textual format. Referring to Figure 3, we hope the Blind LLMs could perform as GPT-4 but internalize the system message. For TAPA utilizes the open-vocabulary detection (OVD) technique [65] to recognize objects within images and parse them into textual environmental messages, we still provide it with ground-truth environmental messages as an oracle setting. + +CodeLLaMA Improves Coding but not Planning. The first two rows in Table 3 highlight the suboptimal task completion rate of the blind LLMs. Among them, CodeLLaMA boasts pre-training on a large programming dataset, resulting in a notable enhancement in code execution from our observation, with $92\%$ of the written code being successfully executed compared to LLaMA's $24\%$ . However, its prowess in planning remains limited. In contrast, the proposed Octopus MPT-7B model displays superior planning and task completion metrics while maintaining commendable coding abilities ( $72\%$ of the written code can be executed). We surmise that the coding requirements within the OctoGibson environment might not be exceedingly intricate, rendering an advanced programming language model, like CodeLLaMA, less crucial, albeit beneficial. For more insight, although not shown in the table, our efforts to replace the MPT model with CodeLLaMA encountered challenges of generating non-sense outputs, suggesting that more refined code, or image-code paired data might be necessary for a successful Octopus-CodeLLaMA integration. + +![](images/a68d0cb8ddc8c569626e79a3f1dea21e19a76649253c2fd8619d3bc8a989ce56.jpg) +(a) TAPA (step) + CodeLLaMA + +![](images/bf2d68a0a852df651c24935ae658d8b2430478c74c364fde7e43907caebd633c.jpg) +(b) Octopus (SFT only) + +![](images/3c1e63dfccfb810c30e5665f109edf1a15ea91cbcda2b41185f1c6239dd2cad4.jpg) +(c) Octopus (SFT + RLEF) + +![](images/f71f405984cf4e8393b9c1e753f545beaca6a3878d55a4d6b61507c55dadcd07.jpg) +(d) Vision Input for Vision-Language Models (e.g., b, c, e) + +![](images/454c045c5ac0ea802e3826daf5c56f2251fb041a482a63198e9ed0dc3ce774fb.jpg) +(e) GPT-4V Output +Fig. 5: Qualitative Results on the task of find a carboy in OctoGibson environment. We show that the models shown can write executable code, but the proposed Octopus has stronger planning ability, especially after RLEF. We also explore the performance of GPT-4V on the specific task. + +Blind LLMs Struggle with Extended Input Content. Our observations indicate that the step-level TAPA model, when supplied with a ground-truth object list, achieves a notable enhancement in planning. The primary distinction between it and the blind CodeLLaMA lies in the input length; the latter deals with protracted, pairwise relation content, complicating the language model's ability to extract crucial data from the environment message. This scenario highlights the inherent limitation of blind LLMs: relying on language alone to convey the entirety of environmental data can result in less informative input. + +Octopus Demonstrates Superior Task Generalization. Table 3 underscores Octopus's strong performance, evidencing its consistent edge over standalone language models in task completion. Its adeptness in adapting to previously unencumbered environments underlines the inherent advantages of vision-language models. A more detailed ablation analysis is provided later. + +RLEF Enhances Octopus's Planning Strategy. Table 3 shows Octopus's strong reasoning capabilities after the RLEF finetuning. An example can be observed in Fig. 5(b-c), where, after refinement via RLEF, Octopus astutely navigates to the cabinet housing the carboy instead of attempting a direct yet distant capture. Quantitatively, Octopus exhibits enhanced adaptability to previously unseen reasoning tasks, reinforcing its prowess in logical task resolution. When juxtaposed with other strategies, such as the embodied queries employed by EmbodiedGPT, RLEF emerges as the more efficacious approach. + +![](images/721900636f8ea71ee414022a79cb72aa5446b63aee809211ccdb846fe3fbce34.jpg) +(a) On Tuning Different Components + +![](images/57ce72ccfe3522c85a01290dbffdf88e7ffd2cd05813e6e487154389caab0f2d.jpg) +(b) On SFT & RLEF and Model Size +(c) On Necessity of Vision Input +Fig.6: Ablation Study on model components, model size, and vision input. For bars with different colors, the upper bar denotes the number of successful reasoning tasks, and the lower is routine tasks. + +![](images/961aa896bd9576199f781b54b4eac79e45a7bc0e813ffd32b68085f9dc26eda2.jpg) + +# 5.2 Ablation Study + +Tunning Different Components Fig. 6 (a) demonstrates that solely adjusting the connector (marked "fire" in Fig. 4 (a)) leads to success for merely 4 out of 60 tasks. Conversely, finetuning both the connector and language decoder nudges the success rate slightly higher, with 5 tasks being accomplished. + +7B v.s. 3B Model Size We embarked on experiments centered on model size to discern the influence of the total parameter count on the efficacy of vision-language models. As illustrated in Fig. 6 (b), downsizing the model manifests in a noticeable performance drop. The congruency of results across both the SFT and RLEF models underscores the importance of an apt model size when sculpting vision-language models. + +Significance of Visual Inputs in Task Performance In our standard configuration, the vision component processes a sequence of image inputs, consisting of eight circularly captured first-person view (FPV) images, complemented by two bird's-eye view (BEV) images. With the intent to investigate the impact of visual inputs on task performance, we initiated an ablation study. In a modified setup, the sequence of these visual inputs was deliberately randomized, aiming to attenuate the strength of the visual signals. As illustrated in Fig. 6 (c), this intentional disruption in visual input consistency led to a pronounced decline in task performance. This result highlights the crucial role that clear and structured visual inputs play in the Octopus model, emphasizing that it significantly leverages visual cues for effective planning and task execution. + +# 5.3 Results on Minecraft and GTA Tasks + +Results on OctoMC According to the OctoMC part in Sec. 3, we designed 40 tasks, each task is operated on 2 locations so a total of 80 tasks. We set aside 10 tasks as unseen tasks and 10 tasks as seen tasks, getting 60 training tasks and 30 testing tasks in OctoMC. Similar to OctoGibson, the training data for OctoMC is collected using GPT-4. The agent, guided by GPT-4, explores the Minecraft environment and generates action plans and corresponding code based on the provided system messages, environmental cues, and objectives. + +Table 4 shows that the SFT model trained on OctoMC can complete most tasks in both seen and unseen scenarios, demonstrating better performance compared to OctoGTA. However, upon analyzing the failure cases, we find that the + +Table 4: Main Results for GTA and Minecraft Tasks. Despite limited training data, the Octopus still shows its good ability in task completion. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators. + +
ModelVision ModelLanguage ModelOctoMCOctoGTA
Seen TaskUnseen TaskAllSeen TaskUnseen TaskAll
GPT-4--0.70 / 0.850.60 / 0.800.65 / 0.830.55 / 0.800.50 / 0.600.54 / 0.76
GPT-4V--0.75 / 0.850.70 / 0.850.73 / 0.850.58 / 0.850.50 / 0.700.56 / 0.82
EmbodiedGPTCLIP-ViTMPT-7B0.30 / 0.550.20 / 0.600.25 / 0.580.15 / 0.380.20 / 0.600.16 / 0.42
Octopus (SFT Only)CLIP-ViTMPT-7B0.30 / 0.600.20 / 0.700.25 / 0.650.18 / 0.480.20 / 0.600.18 / 0.50
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.40 / 0.600.20 / 0.700.30 / 0.650.18 / 0.530.30 / 0.700.20 / 0.56
+ +model sometimes struggles with tasks requiring precise spatial reasoning. For instance, when tasked with killing a pig, the agent may have difficulty finding the creature with the exact angle and distance. While it shows that the agent relies on visual information to navigate and interact with the environment, it also means that even with correct planning, imprecise actions can lead to failure. + +Results on OctoGTA According to the OctoGTA part in Sec. 3, we designed 25 tasks. We set aside 5 tasks in the boat-related group (e.g., boat retrieval and shore return) as unseen tasks for testing only, using 2 different locations. We replicate the remaining 20 tasks for both training (8 different locations) and testing (2 locations). As a result, we have 160 training tasks and 50 testing tasks in OctoGTA. Unlike the training procedure for OctoGibson and OctoMC, the training data for OctoGTA is entirely created by the authors, as it is challenging to gather textual environmental messages in the GTA environment. + +Table 4 shows that, despite having only 160 training tasks, the SFT model can complete some tasks in both seen and unseen scenarios, and RLEF also outperforms. However, upon careful examination of the failure cases, we find that the model struggles with tasks that are not straightforward. For instance, when a wall separates the player from the car, the player still finds it difficult to decide to climb the wall, even if similar cases exist in the training data. Similar to OctoMC, as illustrated in Sec. 3, when approaching certain locations, the code involves functions like turnPlayer() and goForward() rather than a simple walkTo location). Consequently, even with correct planning, imprecise actual actions can still lead to task failure. + +# 6 Conclusion + +This paper introduces Octopus, an embodied vision-language programmer designed to bridge the gap between high-level planning and real-world manipulation with programming. By open-sourcing our OctoVerse environments, dataset, and Octopus architecture, we aim to foster collaboration and innovation within the research community, paving the way for future developments in embodied vision-language programming. + +# Acknowledgments and Disclosure of Funding + +This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221- 0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +1. Grand theft auto v (2014) +2. Minecraft (2023) +3. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Ho, D., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E., Ruano, R.J., Jeffrey, K., Jesmonth, S., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Lee, K.H., Levine, S., Lu, Y., Luu, L., Parada, C., Pastor, P., Quiambao, J., Rao, K., Rettinghouse, J., Reyes, D., Sermanet, P., Sievers, N., Tan, C., Toshev, A., Vanhoucke, V., Xia, F., Xiao, T., Xu, P., Xu, S., Yan, M., Zeng, A.: Do as i can and not as i say: Grounding language in robotic affordances. In: arXiv preprint arXiv:2204.01691 (2022) +4. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +6. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos (2022) +7. Bellemare, M.G., Naddaf, Y., Veness, J., Bowling, M.: The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research 47, 253-279 (Jun 2013). https://doi.org/10.1613/jair.3912, http://dx.doi.org/10.1613/jair.3912 +8. Billard, A., Kragic, D.: Trends and challenges in robot manipulation. Science 364(6446), eaat8414 (2019) +9. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., Zaremba, W.: Openai gym. arXiv preprint arXiv:1606.01540 (2016) +10. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023) +11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +12. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017) + +13. Chen, J., Mu, Y., Yu, Q., Wei, T., Wu, S., Yuan, Z., Liang, Z., Yang, C., Zhang, K., Shao, W., Qiao, Y., Xu, H., Ding, M., Luo, P.: Roboscript: Code generation for free-form manipulation tasks across real and simulation (2024) +14. Chen, L., Li, B., Shen, S., Yang, J., Li, C., Keutzer, K., Darrell, T., Liu, Z.: Language models are visual reasoning coordinators. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023) +15. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) +16. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) +17. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering (2017) +18. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model. In: arXiv preprint arXiv:2303.03378 (2023) +19. Evans, J.S.B.: Dual-processing accounts of reasoning, judgment, and social cognition. Annu. Rev. Psychol. 59, 255-278 (2008) +20. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge (2022) +21. Fu, H., Xu, W., Ye, R., Xue, H., Yu, Z., Tang, T., Li, Y., Du, W., Zhang, J., Lu, C.: Rfuniverse: A multiphysics simulation platform for embodied ai (2023) +22. Fu, Z., Zhao, T.Z., Finn, C.: Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In: arXiv (2024) +23. Gao, X., Gong, R., Shu, T., Xie, X., Wang, S., Zhu, S.C.: Vrkitchen: an interactive 3d virtual environment for task-oriented learning (2019) +24. Gu, S., Holly, E., Lillicrap, T., Levine, S.: Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 3389-3396. IEEE (2017) +25. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023) +26. Huang, S., Jiang, Z., Dong, H., Qiao, Y., Gao, P., Li, H.: Instruct2act: Mapping multi-modality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176 (2023) +27. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023) +28. Hussein, A., Gaber, M.M., Elyan, E., Jayne, C.: Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR) 50(2), 1-35 (2017) +29. Kahneman, D.: Thinking, fast and slow. macmillan (2011) +30. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017) +31. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) + +32. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Conference on Robot Learning. pp. 80–93. PMLR (2023) +33. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +34. Lin, K., Agia, C., Migimatsu, T., Pavone, M., Bohg, J.: Text2motion: From natural language instructions to feasible plans. arXiv preprint arXiv:2303.12153 (2023) +35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023) +36. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/ +37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023) +38. Liu, Z., Dong, Y., Rao, Y., Zhou, J., Lu, J.: Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966 (2024) +39. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition (2021) +40. MosaicML: Mpt-7b (2023), https://www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-23 +41. Mu, Y., Zhang, Q., Hu, M., Wang, W., Ding, M., Jin, J., Wang, B., Dai, J., Qiao, Y., Luo, P.: Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021 (2023) +42. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) +43. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023) +44. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtual-home: Simulating household activities via programs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8494-8502 (2018) +45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +46. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +47. Ramakrishnan, S.K., Jayaraman, D., Grauman, K.: An exploration of embodied visual exploration (2020) +48. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable task planning. arXiv preprint arXiv:2307.06135 (2023) +49. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019) + +50. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023) +51. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) +52. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023) +53. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023) +54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +55. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models (2023) +56. Wang, H., Liang, W., Gool, L.V., Wang, W.: Towards versatile embodied navigation (2022) +57. Wu, Y., Wu, Y., Gkioxari, G., Tian, Y.: Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209 (2018) +58. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023) +59. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: Sapien: A simulated part-based interactive environment (2020) +60. Xie, B., Zhang, S., Zhou, Z., Li, B., Zhang, Y., Hessel, J., Yang, J., Liu, Z.: Funqa: Towards surprising video comprehension. arXiv preprint arXiv:2306.14899 (2023) +61. Yan, C., Misra, D., Bennnett, A., Walsman, A., Bisk, Y., Artzi, Y.: Chalet: Cornell house agent learning environment (2019) +62. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Arenas, M.G., Chiang, H.T.L., Erez, T., Hasenclever, L., Humplik, J., et al.: Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647 (2023) +63. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Skill reinforcement learning and planning for open-world long-horizon tasks (2023) +64. Zheng, S., Liu, J., Feng, Y., Lu, Z.: Steve-eye: Equipping llm-based embodied agents with visual perception in open worlds (2023) +65. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022) +66. Zhu, Y., Wong, J., Mandlekar, A., Martin-Martín, R., Joshi, A., Nasiriany, S., Zhu, Y.: robosite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293 (2020) \ No newline at end of file diff --git a/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/images.zip b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6997670b5f233255bad6ba4f57d89d7105dc057b --- /dev/null +++ b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29bb1a86974d712337bb4d010c74c3b518940e9591e887bea6470a1690a916c4 +size 692938 diff --git a/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/layout.json b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fae9caedb166e915a299a93cd5a92812768c1694 --- /dev/null +++ b/2024/Octopus_ Embodied Vision-Language Programmer from Environmental Feedback/layout.json @@ -0,0 +1,8828 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 112, + 459, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 112, + 459, + 147 + ], + "spans": [ + { + "bbox": [ + 156, + 112, + 459, + 147 + ], + "type": "text", + "content": "Octopus: Embodied Vision-Language Programmer from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 167, + 455, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 455, + 205 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 455, + 205 + ], + "type": "text", + "content": "Jingkang Yang\\*,1, Yuhao Dong\\*,2,3, Shuai Liu\\*,2,4, Bo Li\\*,1, Ziyue Wang\\*,1, Haoran Tan\\*,4, Chencheng Jiang\\*,5, Jiamu Kang\\*,3, Yuanhan Zhang\\*, Kaiyang Zhou\\* and and Ziwei Liu\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "content": " S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "content": " Shanghai AI Laboratory " + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "content": " BUPT " + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "content": " XJTU " + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "content": " Hong Kong Baptist University " + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\{\\text{jingkang001, ziwei.liu}\\}@\\text{ntu.edu.sg}" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 253, + 479, + 459 + ], + "blocks": [ + { + "bbox": [ + 134, + 253, + 479, + 459 + ], + "lines": [ + { + "bbox": [ + 134, + 253, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 134, + 253, + 479, + 459 + ], + "type": "image", + "image_path": "6e1fffbae234e11e03cf9b934aa2f24e859d1db42aff5e8724d5471c83bc1950.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 463, + 482, + 497 + ], + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 497 + ], + "type": "text", + "content": "Fig. 1: Illustration of our vision-language programmer, Octopus, complete a task in GTA environment. Given a task in the form of natural language, Octopus relies on its egocentric vision to generate plans and the corresponding executable code." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 514, + 455, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 514, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 514, + 455, + 647 + ], + "type": "text", + "content": "Abstract. Large vision-language models (VLMs) have achieved substantial progress in multimodal perception and reasoning. When integrated into an embodied agent, existing embodied VLM works either output detailed action sequences at the manipulation level or only provide plans at an abstract level, leaving a gap between high-level planning and real-world manipulation. To bridge this gap, we introduce Octopus, an embodied vision-language programmer that uses executable code generation as a medium to connect planning and manipulation. Octopus is designed to 1) proficiently comprehend an agent's visual and textual task objectives, 2) formulate intricate action sequences, and 3) generate executable code. To facilitate Octopus model development, we introduce OctoVerse: a suite of environments tailored for benchmarking" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 460, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 460, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 460, + 666 + ], + "type": "text", + "content": "* Equal contribution, †Equal engineering contribution, ⌿ Corresponding author." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 271 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 271 + ], + "type": "text", + "content": "vision-based code generators on a wide spectrum of tasks, ranging from mundane daily chores in simulators to sophisticated interactions in complex video games such as Grand Theft Auto (GTA) and Minecraft. To train Octopus, we leverage GPT-4 to control an explorative agent that generates training data, i.e., action blueprints and corresponding executable code. We also collect feedback that enables an enhanced training scheme called Reinforcement Learning with Environmental Feedback (RLEF). Through a series of experiments, we demonstrate Octopus's functionality and present compelling results, showing that the proposed RLEF refines the agent's decision-making. By open-sourcing our simulation environments, dataset, and model architecture, we aspire to ignite further innovation and foster collaborative applications within the broader embodied AI community. The project page is available at https://choiszt.github.io/0ctopus/." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 292, + 230, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 230, + 305 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 230, + 305 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 481, + 376 + ], + "type": "text", + "content": "The rise of large language models (LLMs) [11,15,42,46,54] led to a surge in vision-language models (VLMs) [4,5,31,33,35-37], enabling tasks such as image/video-based descriptions [33], reasoning [14,38,60], and conversations [16,31]. In the realm of embodied AI, notable efforts [10,18,33] have trained agents to process visual input and relay motor control commands." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 520 + ], + "type": "text", + "content": "Another approach to interacting with the environment focuses on task execution through code invocations, mirroring the human System-I stimulation [19,29] (automatic, intuitive actions) with predefined code, and leaving the System-II processes [19,29] (planning and reasoning) for large models. For example, referring to Fig. 1, planning a car ride with a pet might entail a subconscious checklist (e.g., getOutOf() the house, open() the car door), each action could be implemented using specific techniques [8,24] such as imitation learning [22,28]. This programmatic paradigm has been, although not in vision, leveraged by works [25, 50, 52, 53] using LLMs to craft programs and trigger APIs. Game-centric models like Voyager [55] have similarly employed GPT for function calls within game engines, though they often parse data directly from their environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "However, when incorporating visual perception, the programming paradigms are largely unexplored. Primary initiatives [48,58] can only output plans, which anchor their strategies only in initial environmental states or employ dynamic scene graphs for LLM inputs, respectively. Despite their innovations, the overreliance on pre-trained vision models to convert vision content into language can occasionally hinder the LLM's planning performance. The conversion from plans into real-world actions is still missing. While EmbodiedGPT [41] addresses the problem by integrating vision-language modeling for planning and then transitioning to manipulation using policy mapping, the capability of embodied vision-language models to generate executable programs remains largely uncharted." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Our exploration aims to bridge this gap. An embodied vision-language programmer should integrate visual perspective with textual objectives to devise" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "action plans and executable code (Fig. 1). However, existing simulator environments often lack the carefully designed functions necessary to support such models effectively. These functions should balance usefulness and complexity to avoid hindering the development of genuine embodied vision-language programmers. For example, the **explore_until()** function in Minecraft, which can lead the player directly to specific blocks without relying on vision information, may not be suitable for training these models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 307 + ], + "type": "text", + "content": "To meet the requirement, we carefully design and develop OctoVerse, a suite of environments consisting of diverse simulators, including (i) OctoGibson, built upon the photorealistic OmniGibson [32], (ii) OctoMC, developed on the infinitely creative, pixel-style Minecraft platform [2], and (iii) OctoGTA, adapted from the highly interactive and immersive Grand Theft Auto V (GTA-V) [1]. These environments enable the training and benchmarking of our embodied vision-language programming model in a wide range of scenarios, from daily household tasks to complex urban navigation and open-world exploration, while the function calls are tailored to be vision-dependent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 450 + ], + "type": "text", + "content": "Using the OctoVerse environment, we train Octopus by leveraging GPT-4 to collect data. We provide GPT-4 with system messages, environmental cues, and objectives, enabling it to formulate action strategies and code. Simultaneously, the agent captures visual perspectives, forming the image-code pair for Octopus training. During data collection, the agent receives simulator feedback, distinguishing successful moves from unsuccessful ones. We incorporate this feedback using Reinforcement Learning with Environmental Feedback (RLEF) and fine-tune Octopus using Proximal Policy Optimization (PPO) [51]. Empirically, Octopus demonstrates strong adaptability in various scenarios, outperforming existing models in task planning, code generation, and execution. The integration of RLEF further enhances Octopus's performance, showcasing the effectiveness of this training approach. In sum, our key contributions include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 458, + 481, + 578 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 458, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 506 + ], + "type": "text", + "content": "- A Novel Vision-Language Programming Benchmark: Three diverse embodied environments with designed tasks: (i) OctoGibson, which is developed upon OmniGibson [32], (ii) OctoMC that developed on Minecraft [2], and (iii) OctoGTA, which is adapted from GTA-V [1]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 506, + 481, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 506, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 506, + 481, + 542 + ], + "type": "text", + "content": "- A New Vision-Language Programming Model: An embodied vision-language planner and programmer trained with Reinforcement Learning with Environmental Feedback (RLEF), demonstrates compelling results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 542, + 481, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 542, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 542, + 481, + 578 + ], + "type": "text", + "content": "- Insights on Vision-Language Programming: We extensively explore Octopus and share useful insights facilitating future research on visual planning and programming." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 596, + 237, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 237, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 237, + 609 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 621, + 286, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 286, + 632 + ], + "type": "text", + "content": "2.1 Embodied AI Simulators" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "Embodied AI has advanced significantly with the development of diverse simulation environments, enabling research tasks such as visual exploration [47]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 161, + 481, + 301 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 159 + ], + "type": "text", + "content": "Table 1: Related Work for OctoVerse - Overview of Embodied AI Environments. We select three environments into OctoVerse and carefully design executable tasks and vision-dependent function calls (VC), in comparison to undesigned standard function calls (C)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 161, + 481, + 301 + ], + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 301 + ], + "type": "table", + "html": "
SimulatorKinematicsContinuous \nExtended \nStatesFlexible \nMaterialsDeformable \nBodiesRealistic \nAction \nExecutionGame- or \nWorld-BasedFormulated \nTasksFunction \nCall \nType
OpenAIGym [9]✓×××✓G×C
Matterport3D [12]×××××W××
AI2THOR [30]✓××××G×C
VirtualHome [44]×××××G××
House3D [57]×××××W××
Habitat 1.0 [49]✓×××✓W×C
Robosuite [66]✓×××✓W×C
RFUniverse [21]✓×✓✓✓W×C
Minecraft [2]✓×✓×✓G×C
OctoMC✓×✓×✓G✓VC
GTA [1]✓✓✓✓✓G×C
OctoGTA✓✓✓✓✓G✓VC
OmniGibson [32]✓✓✓✓✓W×C
OctoGibson✓✓✓✓✓W✓VC
", + "image_path": "c5542a1160ed522d15337fac866fb91a9da23c4f1225a38aa8ecc7ceb793c5d6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 305, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 481, + 389 + ], + "type": "text", + "content": "navigation [56], and question-answering [17]. Several simulators, including AI2-THOR [30], VirtualHome [44], Habitat-Sim [49], SAPIEN [59], and Omnigibson [32], provide realistic representations of the world for investigating embodied AI challenges. OmniGibson [32] stands out for its high-fidelity simulation of diverse indoor and outdoor environments. OctoGibson environment further enhances OmniGibson with carefully designed function calls and formulated tasks, making it well-suited for vision-language programming." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 510 + ], + "type": "text", + "content": "Game-related simulators like Arade [7], CHALET [61], and VRKitchen [23] also contribute significantly to embodied AI. Minecraft [2] has gained attention in reinforcement learning and game agents [6,20,39,55,63,64] but lacks the necessary structure for vision-language programming. OctoMC addresses this by providing designed function calls and formulated tasks. In contrast to Minecraft's voxel-based representations that limit transferability to real-world environments, GTA-V [1] offers a highly realistic environment. In this work, we introduce OctoGTA as a new setting, leveraging GTA-V's rich, open-world environment with incorporated tasks and function calls, extending this platform for embodied AI study." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 526, + 326, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 326, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 326, + 538 + ], + "type": "text", + "content": "2.2 Embodied AI with Large Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 489, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 489, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 489, + 666 + ], + "type": "text", + "content": "The recent wave of research focuses on merging LLMs with embodied AI tasks [11, 42,46,54]. For instance, VoxPoser addresses robotic manipulation problems through unsupervised methods [27]. A group of projects, namely SayCan [3], Palm-E [18], RT-2 [10], and EmbodiedGPT [41], effectively integrate visual or linguistic cues with robot manipulation data. Outside the domain of robotic manipulation, initiatives like Voyager [55] and Smallville [43] harness the capabilities of GPT to interface with game functions, relying on preset functions to manage intricate manipulations. In a parallel vein, VisProg [25] leverages GPT-3 language prompts to craft Python programs, opening the door to a multitude of fascinating applications. While the proposed Octopus model also formulates plans" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 152, + 482, + 321 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Related Work for Octopus - Overviewing Embodied AI Models. The proposed Octopus distinguishes itself from other models as a unified vision-language model for both plan and code generation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 152, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 134, + 152, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 134, + 152, + 482, + 321 + ], + "type": "table", + "html": "
ModelsRelease DateSupported EnvironmentVision ModelCode GeneratorAction w/ FeedbackLLM Training Enabled
Text2Motion [34]Mar. 2023Sim×✓×
Instruct2Act [26]May 2023Sim×✓×
Lang2Rewards [62]Jun. 2023Sim×✓×
VoxPoser [27]Jul. 2023Sim✓××
SayCan [3]Apr. 2022Real✓××
PALM-E [18]Mar. 2023Sim, Real✓×✓
RT-2 [10]Jul. 2023Real✓×✓
SayPlan [48]Jun. 2023Real×××
EmbodiedGPT [41]May 2023Sim✓×✓
TaPA [58]Jul. 2023Sim××✓
Voyager [55]May 2023Game×✓×
Steve-Eye [64]Dec 2023Game✓×✓
RoboScript [13]Feb 2024Sim, Real×✓×
Octopus-Sim, Game✓✓✓
", + "image_path": "cf730938763d2c71668e4f6703713e8b17853f2b0bf5c5ec1b5df819afe6c5e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": "and code, its distinguishing feature is the seamless integration of visual input in program and code generation. This also stands in contrast to other embodied planners like TAPA [58] and SayPlan [48], which deploy separate vision modules to translate visual data into linguistic inputs for LLMs. Octopus excels as a cohesive vision-language model, delivering not just plans but also code." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 396, + 479, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 479, + 420 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 479, + 420 + ], + "type": "text", + "content": "More discussion on additional related works (Vision Language Model, and Feedback in Large Language Models) is included in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 442, + 323, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 323, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 323, + 455 + ], + "type": "text", + "content": "3 The OctoVerse Environment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 520 + ], + "type": "text", + "content": "In this section, we introduce three simulator environments designed to train and evaluate the Octopus model. For each environment, we will describe their overall information, the special design considerations that ensure the tasks are well-formulated and the callable functions are vision-dependent." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 524, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 482, + 644 + ], + "type": "text", + "content": "OctoGibson We built the environment on the foundation of OmniGibson [32], an existing simulation framework that supports 1,000 daily activities across 50 scenes, featuring over 5,000 meticulously annotated objects. We incorporated 16 functions that the robot can execute, such as moveBot() and easyGrasp(). Within this environment, we meticulously crafted 476 tasks1, each with a well-defined initial state and a definitive termination state, allowing for a straightforward assessment of task completion. Among these tasks, 367 are routine tasks—simple and direct actions like \"place a glass in a trash can,\" marked as Follow. Conversely, the remaining 109 are reasoning tasks that necessitate deeper comprehension. An example is \"buy a chocolate,\" where the agent needs" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 476, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 476, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 476, + 665 + ], + "type": "text", + "content": "1 The full list of tasks and their categories are listed in the supplementary material." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "to know to pick a chocolate bar from the shelf and then place it, along with money, on the checkout counter, denoted as Reason tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 249 + ], + "type": "text", + "content": "To ensure our tasks are vision-aware, we deliberately constrain the usage of certain functions, such as moveBot(object), which moves the agent in front of the given object. To avoid making the task too easy and vision-agnostic, we limit the given parameter to a predefined set of large objects, such as tables and cupboards, rather than small items like cups and glasses. In this case, if the robot wants to pick up a cup, it needs to recognize whether the cup is on the table or in the cupboard. A simple moveBot(object) call with an inappropriate parameter would cause a runtime error. The full list of the functions is in the Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 254, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 324 + ], + "type": "text", + "content": "OctoMC The OctoMC environment is built on Minecraft [2], a popular platform for reinforcement learning and game agents. We integrated 6 functional actions and crafted 40 tasks" + }, + { + "bbox": [ + 130, + 254, + 482, + 324 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 254, + 482, + 324 + ], + "type": "text", + "content": ", each designed to facilitate comprehensive observations and executions by the agent. These tasks are distributed across 10 different biomes, including indoor, outdoor, and underground settings, under varying weather conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 434 + ], + "type": "text", + "content": "However, existing Minecraft environments often lack vision-formulated tasks and the necessary structure for vision-language programming. For example, in Voyager [55], the exploreUntil() function allows the player to navigate directly to specific blocks without relying on vision information. This function works by randomly exploring the environment within a certain range until the desired object is found, at which point it returns a true value, enabling the agent to interact with the located object, such as gold blocks or trees. While effective, this approach is entirely automated and does not utilize visual information, making it unsuitable for our vision-based objectives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 435, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 507 + ], + "type": "text", + "content": "To address this limitation, we crafted a vision-dependent exploration function, teleport(yaw, distance), which operates within the robot's perceptual range. This function ensures that the agent's operations are vision-dependent and require active perception and navigation within the environment. For instance, to locate a tree block, the agent must actively navigate towards the direction of the forest, relying on visual cues to guide its exploration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 511, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 571 + ], + "type": "text", + "content": "OctoGTA The OctoGTA environment is built on GTA-V [1] with the help of the active GTA modding community. We integrated 19 functions and methodically crafted 25 tasks3, such as \"help NPC drive their boat back to shore\" and \"mediate a fight between two NPCs,\" spanning across 5 groups (shown in Fig. 2 (b)). Each task is assigned to 5 different locations within the game world." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 573, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 620 + ], + "type": "text", + "content": "We have implemented a set of functions that enable the agent to interact with the game world in a visually-aware manner. Similar to the design in Minecraft, we get rid of functions like walkTo location) that might trivialize the task of reaching a particular building or landmark. Instead, we provide functions such" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 369, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 369, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 369, + 643 + ], + "type": "text", + "content": "2 Detailed tasks are listed in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "3 We meticulously designed tasks to be friendly, ensuring they exclude any inappropriate or violent behaviors." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 478, + 243 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 243 + ], + "type": "image", + "image_path": "76f8463ebcdb002f43f190e5dcf9ce9a13839fdab317d9567151f0cc97fd2181.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 245, + 480, + 256 + ], + "lines": [ + { + "bbox": [ + 132, + 245, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 480, + 256 + ], + "type": "text", + "content": "Fig. 2: The Statistics of the OctoVerse Environment with Function Designs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 267, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 480, + 327 + ], + "type": "text", + "content": "as goForward(distance) and turnPlayer(degree) (Fig. 2 (b)). For tasks like mediating a fight between two NPCs, the essential function stopFight() only works when the player is within 5 meters of the fighting NPCs. These design choices ensure that the agent's operations are vision-dependent and require active perception and navigation within the environment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 478, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 478, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 478, + 357 + ], + "type": "text", + "content": "4 Octopus: The Embodied Vision-Language Programmer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 425 + ], + "type": "text", + "content": "In this section, we present the procedure for training Octopus. Starting from collecting training data within the OctoVerse environment, Octopus builds upon a VLM architecture of Otter [31] and includes specialized RLEF modules to handle vision-language programming tasks. Fig. 4 illustrates the entire Octopus training pipeline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 441, + 285, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 285, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 285, + 452 + ], + "type": "text", + "content": "4.1 Training Data Collection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 457, + 480, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 480, + 553 + ], + "type": "text", + "content": "We use the automatic training data collection pipeline described here for OctoGibson and OctoMC, with the latter using customized prompts inspired by Voyager [55]. For OctoGTA, we rely on human labor to hand-craft the training dataset due to the difficulty of obtaining textual environment messages in the GTA environment. In the following parts, we use OctoGibson as the primary example to illustrate the data collection pipeline. Note that the primary task in organizing training data is to form a succinct pairing: \"vision input + current/historical states " + }, + { + "bbox": [ + 130, + 457, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 457, + 480, + 553 + ], + "type": "text", + "content": " next step plan + executable code\"." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 555, + 480, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 480, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 480, + 627 + ], + "type": "text", + "content": "Environment Info Collection As shown in Fig. 4 (a) and Fig. 3, we format an environment message for each state, encompassing attributes like Observed Objects, Observed Relations, Inventory, and more. Specifically, the simulator can provide us with an exact scene graph at each state, shaping the content for the first two parts. The inventory info is also accessible. The task, e.g., \"cooking bacon\" in Fig. 3, is represented by the Task Goal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": "Automation with GPT-4 After preparing the environment message, we crafted a structured system message to ensure that the robot not only understands its input but also maintains a consistent output format. A detailed" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 477, + 304 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 477, + 304 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 477, + 304 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 477, + 304 + ], + "type": "image", + "image_path": "6cdd976d7b1070035e63f9f0eb63fef5fd12684a6160ac0c712a8b05aaf85f52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 392 + ], + "type": "text", + "content": "Fig.3: Data Collection Example for \"Cook a Bacon\" Task. GPT-4 perceives the environment through the environmental message and produces anticipated plans and code following the detailed system message. This code is subsequently executed in the simulator, directing the agent to the subsequent state. For each state, we gather the environmental message, wherein observed objects and relations are substituted by egocentric images to serve as the training input. The response from GPT-4 acts as the training output. Environmental feedback, specifically the determination of whether each target state is met, is documented for RLEF training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 407, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 515 + ], + "type": "text", + "content": "examination of this prompt can be found in the appendix. Experiments have shown that a well-articulated prompt enables GPT-4 to effectively generate executable code. It is important to note that the combined length of the system and environment messages can be extremely long, which may cause standard GPT-4 8K models to struggle with producing meaningful outputs. To address this issue, we employ the more robust GPT-4 32K model. As illustrated in Fig. 3, when GPT-4 receives a consistent system and environment message, it generates comprehensive outputs that include current scenario analysis, planning, and actionable code, supporting the training process in Section 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 518, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 638 + ], + "type": "text", + "content": "Error Management Notably, GPT-4 collects training data under the main task of guiding the agent to complete tasks. However, GPT-4 is not infallible. Errors can manifest in multiple ways, ranging from syntax errors to physical challenges in the simulator. For instance, in Fig. 3, between states " + }, + { + "bbox": [ + 130, + 518, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\#5" + }, + { + "bbox": [ + 130, + 518, + 482, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 518, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\#6" + }, + { + "bbox": [ + 130, + 518, + 482, + 638 + ], + "type": "text", + "content": ", the action failed due to the long distance between the agent (holding bacon) and the pan. Such setbacks reset the task to its previous state. If a main task remains incomplete after 10 steps, it is deemed unsuccessful, and we terminate this task for budget concerns. However, all data pairs without syntax errors, regardless of the task's completion status, are valuable for refining instructions and improving the model's performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Environmental Feedback GPT-4's continual trial-and-error approach while guiding the agent toward task completion serves a dual purpose: collecting vision-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 278, + 171 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 278, + 171 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 278, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 278, + 171 + ], + "type": "image", + "image_path": "e7e886cf6a620a9c98069660b102aace976d3cf7ca9f6b7a9e3ddf03c6685308.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 171, + 250, + 180 + ], + "lines": [ + { + "bbox": [ + 162, + 171, + 250, + 180 + ], + "spans": [ + { + "bbox": [ + 162, + 171, + 250, + 180 + ], + "type": "text", + "content": "(a) Data Collection Pipeline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 187, + 279, + 216 + ], + "blocks": [ + { + "bbox": [ + 141, + 187, + 279, + 216 + ], + "lines": [ + { + "bbox": [ + 141, + 187, + 279, + 216 + ], + "spans": [ + { + "bbox": [ + 141, + 187, + 279, + 216 + ], + "type": "image", + "image_path": "d0ff0b0f34aadd20b48a71ac7fdbaec50a85b8bdbcf92072cdeb9a257db8ea84.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 217, + 253, + 225 + ], + "lines": [ + { + "bbox": [ + 158, + 217, + 253, + 225 + ], + "spans": [ + { + "bbox": [ + 158, + 217, + 253, + 225 + ], + "type": "text", + "content": "(b) Octopus Training Pipeline" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 284, + 116, + 389, + 212 + ], + "blocks": [ + { + "bbox": [ + 284, + 116, + 389, + 212 + ], + "lines": [ + { + "bbox": [ + 284, + 116, + 389, + 212 + ], + "spans": [ + { + "bbox": [ + 284, + 116, + 389, + 212 + ], + "type": "image", + "image_path": "2b87a40202ac8da67bda6e6d35469e9fe011aa49fa4927b481cae1065f0d6289.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 217, + 378, + 225 + ], + "lines": [ + { + "bbox": [ + 296, + 217, + 378, + 225 + ], + "spans": [ + { + "bbox": [ + 296, + 217, + 378, + 225 + ], + "type": "text", + "content": "(c) Octopus (Otter model)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 116, + 480, + 160 + ], + "blocks": [ + { + "bbox": [ + 391, + 116, + 480, + 160 + ], + "lines": [ + { + "bbox": [ + 391, + 116, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 391, + 116, + 480, + 160 + ], + "type": "image", + "image_path": "41f47cd6d088c512b2b154383379f89ea6c4edf682b2cf99beebf2161af803f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 169, + 477, + 195 + ], + "blocks": [ + { + "bbox": [ + 392, + 162, + 462, + 169 + ], + "lines": [ + { + "bbox": [ + 392, + 162, + 462, + 169 + ], + "spans": [ + { + "bbox": [ + 392, + 162, + 462, + 169 + ], + "type": "text", + "content": "Step 2: Training Reward Model" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 169, + 477, + 195 + ], + "lines": [ + { + "bbox": [ + 392, + 169, + 477, + 195 + ], + "spans": [ + { + "bbox": [ + 392, + 169, + 477, + 195 + ], + "type": "image", + "image_path": "65b407683e8cb947939a9954f9b85688b2bf07e5c7b7867d83395650594fc61d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 396, + 196, + 477, + 217 + ], + "blocks": [ + { + "bbox": [ + 396, + 196, + 477, + 217 + ], + "lines": [ + { + "bbox": [ + 396, + 196, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 396, + 196, + 477, + 217 + ], + "type": "image", + "image_path": "d0606d05feeaff5577056bd7d0673252451a64569f06890ba05703f7fdf75430.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 218, + 470, + 226 + ], + "lines": [ + { + "bbox": [ + 409, + 218, + 470, + 226 + ], + "spans": [ + { + "bbox": [ + 409, + 218, + 470, + 226 + ], + "type": "text", + "content": "(d) RLEF Training" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 231, + 451, + 242 + ], + "lines": [ + { + "bbox": [ + 160, + 231, + 451, + 242 + ], + "spans": [ + { + "bbox": [ + 160, + 231, + 451, + 242 + ], + "type": "text", + "content": "Fig. 4: How to train Octopus: data collection and training pipeline." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 251, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 407 + ], + "type": "text", + "content": "output pairs and generating a rich set of feedback data. The automatic annotation of this feedback focuses on two levels: step-level and task-level judgments. Step-level judgment assesses the alignment of post-execution states with their target states. For instance, in Fig. 3, steps color-coded in green lead to positive feedback. One can visualize the action sequence for task completion as a tree, where each node indicates a step (subtask), encapsulating an action code. Accompanying each step is a binary value that denotes success or failure, giving preference to the successful branch over its counterpart. Task-level judgment, on the other hand, gauges the successful execution of the overall task. If the task is not completed as intended, every state within that task is labeled as negative, regardless of the status of the subtasks. This collated feedback data serves as a foundation for our Reinforcement Learning with Environmental Feedback (RLEF) methodology, which we discuss in greater detail in Section 4.4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 424, + 260, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 260, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 260, + 434 + ], + "type": "text", + "content": "4.2 Model Architecture" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 444, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 529 + ], + "type": "text", + "content": "The Octopus architecture (shown in Fig. 4 (c)) is heavily inspired by the Otter model [31], integrates the MPT-7B Language Decoder [40] and CLIP VIT-L/14 Vision Encoder [45]. Adopting design principles from Flamingo [4], Octopus employs the Perceiver Resampler and Cross-Gated Attention modules to enhance vision-language synergy. This architecture enables Octopus to excel in tasks requiring understanding of both visual and textual data. The Octopus is also compatible with other VLMs such as LLaVA [37]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 546, + 393, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 393, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 393, + 558 + ], + "type": "text", + "content": "4.3 SFT: Supervised Finetuning with Instructions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "text", + "content": "We train the Octopus model on our collected dataset from OctoVerse " + }, + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{E}} = \\{(\\mathbf{X}_v,\\mathbf{T}_i,\\mathbf{T}_r)\\}" + }, + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "text", + "content": " using token-level supervised fine-tuning (SFT) [42,54]. The Perceiver Resampler transforms images " + }, + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v" + }, + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "text", + "content": " into visual tokens that condition subsequent layers via Cross-Gated Attention modules. The training objective is next-token prediction, modeling the likelihood of a targeted response " + }, + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_r" + }, + { + "bbox": [ + 130, + 566, + 481, + 627 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 214, + 635, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 635, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 214, + 635, + 481, + 667 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {T} _ {r} \\mid \\mathbf {T} _ {i}, \\mathbf {X} _ {v}\\right) = \\prod_ {l = 1} ^ {L} p \\left(t _ {l} \\mid \\mathbf {X} _ {v}, \\mathbf {T} _ {i}, \\mathbf {T} _ {r, < l}\\right). \\tag {1}", + "image_path": "8825d6d37162a34c50775c52c8407bce9c2488e4fdb23e92ec30d5ec649e13af.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " denotes the instruction tokens and " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{r, < l}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " denotes the response tokens before the current predicted token " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "t_l" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": ". During inference, tokens are converted into natural language via the language decoder's text tokenizer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "Visual observations " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v = \\{x_F^0,\\dots ,x_F^7,x_B^0,x_B^1\\}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " consist of 8 first-person view (FPV) images and two bird's-eye view (BEV) images for OctoGibson and OctoGTA. OctoMC only takes 4 FPV images. The FPV captures the agent's direct observations, while the BEV provides a holistic understanding of the environment. The eight FPV images are captured every 45 degrees, ensuring a complete 360-degree perspective." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 239, + 473, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 239, + 473, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 473, + 251 + ], + "type": "text", + "content": "4.4 RLEF: Reinforcement Learning with Environmental Feedback" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 256, + 479, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 479, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 479, + 303 + ], + "type": "text", + "content": "In OctoVerse, task progression can be visualized as a tree (Fig. 4 (d)), where each node represents a sub-task with a binary value indicating success (1) or failure (0). If a node (or sub-task) has a value of 1, it is a step in the correct direction toward our end goal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "text", + "content": "Tree-based Task Representation According to the environmental feedback part in Sec. 4.1, environmental reward datasets " + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbb{R}} = (\\mathbf{X}_v^*,\\mathbf{T}_i^*,\\mathbf{T}_r^i,\\mathbf{T}_r^j,c)" + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "text", + "content": " are organized, where " + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_r^i" + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_r^j" + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "text", + "content": " are responses sharing the same parent task " + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i^*" + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 304, + 480, + 387 + ], + "type": "text", + "content": " indicates the preferred response leading to task completion. This ensures the reward mechanism favors the successfully executed branch. Note that even if a parental node does not have multiple responses, we can still assign feedback according to the rule in Sec. 4.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "text", + "content": "Reward Model Configuration A single-modal CodeLLaMA-7B model with an additional value head is fine-tuned on " + }, + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{R}}" + }, + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "text", + "content": " as the reward model " + }, + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "inline_equation", + "content": "r_{\\phi}" + }, + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "text", + "content": ". This text-based model assesses state transitions (" + }, + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i^* \\rightarrow \\mathbf{T}_r^{i,j}" + }, + { + "bbox": [ + 130, + 388, + 480, + 483 + ], + "type": "text", + "content": ") to determine high-reward transitions, assisting the agent in task execution and completion. The rationale for using CodeLLaMA as the reward model is that evaluating rewards can be purely dependent on the textual output. Furthermore, CodeLLaMA's strong programming skills make it well-suited for assessing the quality and effectiveness of the generated code in the context of task completion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 480, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 531 + ], + "type": "text", + "content": "Policy Model Development The supervised fine-tuned model serves as the initial policy model " + }, + { + "bbox": [ + 130, + 483, + 480, + 531 + ], + "type": "inline_equation", + "content": "\\pi^{\\mathrm{INIT}}" + }, + { + "bbox": [ + 130, + 483, + 480, + 531 + ], + "type": "text", + "content": " with fixed parameters. A duplicate model, " + }, + { + "bbox": [ + 130, + 483, + 480, + 531 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}^{\\mathrm{RL}}" + }, + { + "bbox": [ + 130, + 483, + 480, + 531 + ], + "type": "text", + "content": ", is initialized and trained using Proximal Policy Optimization (PPO) [51] to maximize response rewards. The loss function is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 536, + 503, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 503, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 503, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\pi_ {\\theta} ^ {\\mathrm {R L}}\\right) = - \\mathbb {E} _ {\\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right) \\in \\mathcal {D} _ {\\mathrm {R}}, \\mathbf {T} _ {r} \\sim \\pi^ {\\mathrm {R L}}} \\left[ r _ {\\phi} \\left(\\mathbf {T} _ {i} ^ {*}, \\mathbf {T} _ {r}\\right) - \\beta \\cdot \\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} ^ {\\mathrm {R L}} \\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right) \\| \\pi^ {\\mathrm {I N I T}} \\left(\\mathbf {X} _ {v} ^ {*}, \\mathbf {T} _ {i} ^ {*}\\right)\\right) \\right],", + "image_path": "0cebec29510e37d7991340298e43b0e77cf9769697365f2d3569ab2c3c719d9d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "type": "text", + "content": " acts as a hyper-parameter to regulate the magnitude of the Kullback-Leibler (KL) penalty." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 601, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 230, + 615 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 624, + 307, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 307, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 307, + 635 + ], + "type": "text", + "content": "5.1 Main Results on OctoGibson" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Experimental Setup We first set up the OctoGibson to evaluate the performance of Octopus and other related models. Specifically, we are utilizing the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 191, + 480, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 191 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 191 + ], + "type": "text", + "content": "Table 3: Main Results on OctoGibson. We compare various models: standalone language models, adapted vision-language planners, and our Octopus models, across different evaluation settings. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators. GT denotes that the model input is directly parsed from the simulator, with information on objects (O) or relations (R). Octopus shows consistent advance in task completion." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 191, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 191, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 480, + 297 + ], + "type": "table", + "html": "
ModelVision ModelLanguage ModelEntire Goal Task
Seen EnvUnseen EnvFollowReasonAll
GPT-4--0.42 / 0.690.46 / 0.670.49 / 0.780.27 / 0.400.43 / 0.68
GPT-4V--0.40 / 0.620.60 / 0.670.42 / 0.670.53 / 0.530.45 / 0.63
LLaMAGT (O+R)LLaMA2-7B0.07 / 0.110.13 / 0.130.11 / 0.160.00 / 0.000.08 / 0.12
CodeLLaMAGT (O+R)CodeLLaMA-7B0.09 / 0.200.20 / 0.400.16 / 0.310.00 / 0.070.12 / 0.25
TAPA (task-level)QvD GT (O)CodeLLaMA-7B0.09 / 0.360.13 / 0.330.11 / 0.360.06 / 0.330.10 / 0.35
TAPA (step-level)QvD GT (O)CodeLLaMA-7B0.16 / 0.420.13 / 0.270.18 / 0.380.07 / 0.400.15 / 0.38
EmbodiedGPTCLIP-ViTMPT-7B0.04 / 0.360.27 / 0.530.13 / 0.380.00 / 0.400.10 / 0.40
Octopus (SFT Only)CLIP-ViTMPT-7B0.11 / 0.330.27 / 0.470.16 / 0.380.13 / 0.330.15 / 0.37
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.13 / 0.380.33 / 0.530.18 / 0.400.20 / 0.530.18 / 0.42
", + "image_path": "a4ad0b6a9dc17f8ac4c7399acb28f0bcdf7615a6395a1567d34ae72bec4be958.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 302, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 410 + ], + "type": "text", + "content": "metrics of goal task completion score to check whether the task is completed in the simulator and the plan score from human evaluation. We have 60 evaluation tasks, with 45 from the seen environment, and 15 that are unseen during training. We also have 45 routine tasks and 15 require reasoning. Please note that models like Octopus might not always accurately identify specific object names as they appear in the simulator (e.g., \"water_bottle_189\"). To address this, we implement a post-processing step for the generated code, substituting generic object references with their exact names from the simulator with simple string similarity matching. If multiple objects, we select the one closest to the agent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 410, + 481, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 481, + 482 + ], + "type": "text", + "content": "For Blind LLMs, we provide them with all the environment information in a textual format. Referring to Figure 3, we hope the Blind LLMs could perform as GPT-4 but internalize the system message. For TAPA utilizes the open-vocabulary detection (OVD) technique [65] to recognize objects within images and parse them into textual environmental messages, we still provide it with ground-truth environmental messages as an oracle setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "CodeLLaMA Improves Coding but not Planning. The first two rows in Table 3 highlight the suboptimal task completion rate of the blind LLMs. Among them, CodeLLaMA boasts pre-training on a large programming dataset, resulting in a notable enhancement in code execution from our observation, with " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " of the written code being successfully executed compared to LLaMA's " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". However, its prowess in planning remains limited. In contrast, the proposed Octopus MPT-7B model displays superior planning and task completion metrics while maintaining commendable coding abilities (" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " of the written code can be executed). We surmise that the coding requirements within the OctoGibson environment might not be exceedingly intricate, rendering an advanced programming language model, like CodeLLaMA, less crucial, albeit beneficial. For more insight, although not shown in the table, our efforts to replace the MPT model with CodeLLaMA encountered challenges of generating non-sense outputs, suggesting that more refined code, or image-code paired data might be necessary for a successful Octopus-CodeLLaMA integration." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 238, + 224 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 238, + 224 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 238, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 238, + 224 + ], + "type": "image", + "image_path": "a68d0cb8ddc8c569626e79a3f1dea21e19a76649253c2fd8619d3bc8a989ce56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 225, + 233, + 234 + ], + "lines": [ + { + "bbox": [ + 138, + 225, + 233, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 225, + 233, + 234 + ], + "type": "text", + "content": "(a) TAPA (step) + CodeLLaMA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 115, + 354, + 224 + ], + "blocks": [ + { + "bbox": [ + 242, + 115, + 354, + 224 + ], + "lines": [ + { + "bbox": [ + 242, + 115, + 354, + 224 + ], + "spans": [ + { + "bbox": [ + 242, + 115, + 354, + 224 + ], + "type": "image", + "image_path": "bf2d68a0a852df651c24935ae658d8b2430478c74c364fde7e43907caebd633c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 225, + 333, + 234 + ], + "lines": [ + { + "bbox": [ + 264, + 225, + 333, + 234 + ], + "spans": [ + { + "bbox": [ + 264, + 225, + 333, + 234 + ], + "type": "text", + "content": "(b) Octopus (SFT only)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 359, + 115, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 359, + 115, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 359, + 115, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 359, + 115, + 479, + 224 + ], + "type": "image", + "image_path": "3c1e63dfccfb810c30e5665f109edf1a15ea91cbcda2b41185f1c6239dd2cad4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 225, + 464, + 234 + ], + "lines": [ + { + "bbox": [ + 385, + 225, + 464, + 234 + ], + "spans": [ + { + "bbox": [ + 385, + 225, + 464, + 234 + ], + "type": "text", + "content": "(c) Octopus (SFT + RLEF)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 236, + 346, + 345 + ], + "blocks": [ + { + "bbox": [ + 134, + 236, + 346, + 345 + ], + "lines": [ + { + "bbox": [ + 134, + 236, + 346, + 345 + ], + "spans": [ + { + "bbox": [ + 134, + 236, + 346, + 345 + ], + "type": "image", + "image_path": "f71f405984cf4e8393b9c1e753f545beaca6a3878d55a4d6b61507c55dadcd07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 346, + 331, + 354 + ], + "lines": [ + { + "bbox": [ + 160, + 346, + 331, + 354 + ], + "spans": [ + { + "bbox": [ + 160, + 346, + 331, + 354 + ], + "type": "text", + "content": "(d) Vision Input for Vision-Language Models (e.g., b, c, e)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 350, + 236, + 478, + 345 + ], + "blocks": [ + { + "bbox": [ + 350, + 236, + 478, + 345 + ], + "lines": [ + { + "bbox": [ + 350, + 236, + 478, + 345 + ], + "spans": [ + { + "bbox": [ + 350, + 236, + 478, + 345 + ], + "type": "image", + "image_path": "454c045c5ac0ea802e3826daf5c56f2251fb041a482a63198e9ed0dc3ce774fb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 346, + 454, + 354 + ], + "lines": [ + { + "bbox": [ + 396, + 346, + 454, + 354 + ], + "spans": [ + { + "bbox": [ + 396, + 346, + 454, + 354 + ], + "type": "text", + "content": "(e) GPT-4V Output" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 354, + 481, + 398 + ], + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 398 + ], + "type": "text", + "content": "Fig. 5: Qualitative Results on the task of find a carboy in OctoGibson environment. We show that the models shown can write executable code, but the proposed Octopus has stronger planning ability, especially after RLEF. We also explore the performance of GPT-4V on the specific task." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "text", + "content": "Blind LLMs Struggle with Extended Input Content. Our observations indicate that the step-level TAPA model, when supplied with a ground-truth object list, achieves a notable enhancement in planning. The primary distinction between it and the blind CodeLLaMA lies in the input length; the latter deals with protracted, pairwise relation content, complicating the language model's ability to extract crucial data from the environment message. This scenario highlights the inherent limitation of blind LLMs: relying on language alone to convey the entirety of environmental data can result in less informative input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 505, + 482, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 565 + ], + "type": "text", + "content": "Octopus Demonstrates Superior Task Generalization. Table 3 underscores Octopus's strong performance, evidencing its consistent edge over standalone language models in task completion. Its adeptness in adapting to previously unencumbered environments underlines the inherent advantages of vision-language models. A more detailed ablation analysis is provided later." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "RLEF Enhances Octopus's Planning Strategy. Table 3 shows Octopus's strong reasoning capabilities after the RLEF finetuning. An example can be observed in Fig. 5(b-c), where, after refinement via RLEF, Octopus astutely navigates to the cabinet housing the carboy instead of attempting a direct yet distant capture. Quantitatively, Octopus exhibits enhanced adaptability to previously unseen reasoning tasks, reinforcing its prowess in logical task resolution. When juxtaposed with other strategies, such as the embodied queries employed by EmbodiedGPT, RLEF emerges as the more efficacious approach." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 249, + 193 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 249, + 193 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 249, + 193 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 249, + 193 + ], + "type": "image", + "image_path": "721900636f8ea71ee414022a79cb72aa5446b63aee809211ccdb846fe3fbce34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 194, + 249, + 202 + ], + "lines": [ + { + "bbox": [ + 144, + 194, + 249, + 202 + ], + "spans": [ + { + "bbox": [ + 144, + 194, + 249, + 202 + ], + "type": "text", + "content": "(a) On Tuning Different Components" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 117, + 389, + 191 + ], + "blocks": [ + { + "bbox": [ + 265, + 117, + 389, + 191 + ], + "lines": [ + { + "bbox": [ + 265, + 117, + 389, + 191 + ], + "spans": [ + { + "bbox": [ + 265, + 117, + 389, + 191 + ], + "type": "image", + "image_path": "57ce72ccfe3522c85a01290dbffdf88e7ffd2cd05813e6e487154389caab0f2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 194, + 378, + 202 + ], + "lines": [ + { + "bbox": [ + 274, + 194, + 378, + 202 + ], + "spans": [ + { + "bbox": [ + 274, + 194, + 378, + 202 + ], + "type": "text", + "content": "(b) On SFT & RLEF and Model Size" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 194, + 477, + 202 + ], + "lines": [ + { + "bbox": [ + 388, + 194, + 477, + 202 + ], + "spans": [ + { + "bbox": [ + 388, + 194, + 477, + 202 + ], + "type": "text", + "content": "(c) On Necessity of Vision Input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 202, + 480, + 233 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 480, + 233 + ], + "type": "text", + "content": "Fig.6: Ablation Study on model components, model size, and vision input. For bars with different colors, the upper bar denotes the number of successful reasoning tasks, and the lower is routine tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 402, + 118, + 470, + 191 + ], + "blocks": [ + { + "bbox": [ + 402, + 118, + 470, + 191 + ], + "lines": [ + { + "bbox": [ + 402, + 118, + 470, + 191 + ], + "spans": [ + { + "bbox": [ + 402, + 118, + 470, + 191 + ], + "type": "image", + "image_path": "961aa896bd9576199f781b54b4eac79e45a7bc0e813ffd32b68085f9dc26eda2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 238, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 238, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 238, + 248 + ], + "type": "text", + "content": "5.2 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 254, + 480, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 480, + 302 + ], + "type": "text", + "content": "Tunning Different Components Fig. 6 (a) demonstrates that solely adjusting the connector (marked \"fire\" in Fig. 4 (a)) leads to success for merely 4 out of 60 tasks. Conversely, finetuning both the connector and language decoder nudges the success rate slightly higher, with 5 tasks being accomplished." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 304, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 376 + ], + "type": "text", + "content": "7B v.s. 3B Model Size We embarked on experiments centered on model size to discern the influence of the total parameter count on the efficacy of vision-language models. As illustrated in Fig. 6 (b), downsizing the model manifests in a noticeable performance drop. The congruency of results across both the SFT and RLEF models underscores the importance of an apt model size when sculpting vision-language models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 379, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 510 + ], + "type": "text", + "content": "Significance of Visual Inputs in Task Performance In our standard configuration, the vision component processes a sequence of image inputs, consisting of eight circularly captured first-person view (FPV) images, complemented by two bird's-eye view (BEV) images. With the intent to investigate the impact of visual inputs on task performance, we initiated an ablation study. In a modified setup, the sequence of these visual inputs was deliberately randomized, aiming to attenuate the strength of the visual signals. As illustrated in Fig. 6 (c), this intentional disruption in visual input consistency led to a pronounced decline in task performance. This result highlights the crucial role that clear and structured visual inputs play in the Octopus model, emphasizing that it significantly leverages visual cues for effective planning and task execution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 525, + 349, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 349, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 349, + 537 + ], + "type": "text", + "content": "5.3 Results on Minecraft and GTA Tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 629 + ], + "type": "text", + "content": "Results on OctoMC According to the OctoMC part in Sec. 3, we designed 40 tasks, each task is operated on 2 locations so a total of 80 tasks. We set aside 10 tasks as unseen tasks and 10 tasks as seen tasks, getting 60 training tasks and 30 testing tasks in OctoMC. Similar to OctoGibson, the training data for OctoMC is collected using GPT-4. The agent, guided by GPT-4, explores the Minecraft environment and generates action plans and corresponding code based on the provided system messages, environmental cues, and objectives." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": "Table 4 shows that the SFT model trained on OctoMC can complete most tasks in both seen and unseen scenarios, demonstrating better performance compared to OctoGTA. However, upon analyzing the failure cases, we find that the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 171, + 481, + 244 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: Main Results for GTA and Minecraft Tasks. Despite limited training data, the Octopus still shows its good ability in task completion. In cells displaying two values, the first represents the task completion rate across the target validation task sets, while the second assesses the conceptual accuracy of the model's planning as judged by human evaluators." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 171, + 481, + 244 + ], + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 244 + ], + "type": "table", + "html": "
ModelVision ModelLanguage ModelOctoMCOctoGTA
Seen TaskUnseen TaskAllSeen TaskUnseen TaskAll
GPT-4--0.70 / 0.850.60 / 0.800.65 / 0.830.55 / 0.800.50 / 0.600.54 / 0.76
GPT-4V--0.75 / 0.850.70 / 0.850.73 / 0.850.58 / 0.850.50 / 0.700.56 / 0.82
EmbodiedGPTCLIP-ViTMPT-7B0.30 / 0.550.20 / 0.600.25 / 0.580.15 / 0.380.20 / 0.600.16 / 0.42
Octopus (SFT Only)CLIP-ViTMPT-7B0.30 / 0.600.20 / 0.700.25 / 0.650.18 / 0.480.20 / 0.600.18 / 0.50
Octopus (SFT + RLEF)CLIP-ViTMPT-7B0.40 / 0.600.20 / 0.700.30 / 0.650.18 / 0.530.30 / 0.700.20 / 0.56
", + "image_path": "42d0fb4a90c2701ba28200c150a89d74222db3bb3b5dea5d2adaeffc7b6c5aed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 307 + ], + "type": "text", + "content": "model sometimes struggles with tasks requiring precise spatial reasoning. For instance, when tasked with killing a pig, the agent may have difficulty finding the creature with the exact angle and distance. While it shows that the agent relies on visual information to navigate and interact with the environment, it also means that even with correct planning, imprecise actions can lead to failure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 405 + ], + "type": "text", + "content": "Results on OctoGTA According to the OctoGTA part in Sec. 3, we designed 25 tasks. We set aside 5 tasks in the boat-related group (e.g., boat retrieval and shore return) as unseen tasks for testing only, using 2 different locations. We replicate the remaining 20 tasks for both training (8 different locations) and testing (2 locations). As a result, we have 160 training tasks and 50 testing tasks in OctoGTA. Unlike the training procedure for OctoGibson and OctoMC, the training data for OctoGTA is entirely created by the authors, as it is challenging to gather textual environmental messages in the GTA environment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 526 + ], + "type": "text", + "content": "Table 4 shows that, despite having only 160 training tasks, the SFT model can complete some tasks in both seen and unseen scenarios, and RLEF also outperforms. However, upon careful examination of the failure cases, we find that the model struggles with tasks that are not straightforward. For instance, when a wall separates the player from the car, the player still finds it difficult to decide to climb the wall, even if similar cases exist in the training data. Similar to OctoMC, as illustrated in Sec. 3, when approaching certain locations, the code involves functions like turnPlayer() and goForward() rather than a simple walkTo location). Consequently, even with correct planning, imprecise actual actions can still lead to task failure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 556, + 220, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 220, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 220, + 570 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "This paper introduces Octopus, an embodied vision-language programmer designed to bridge the gap between high-level planning and real-world manipulation with programming. By open-sourcing our OctoVerse environments, dataset, and Octopus architecture, we aim to foster collaboration and innovation within the research community, paving the way for future developments in embodied vision-language programming." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 399, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 399, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 399, + 129 + ], + "type": "text", + "content": "Acknowledgments and Disclosure of Funding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "type": "text", + "content": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221- 0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 228, + 257, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 257, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 257, + 239 + ], + "type": "text", + "content": "1. Grand theft auto v (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 218, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 218, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 218, + 250 + ], + "type": "text", + "content": "2. Minecraft (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 251, + 481, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 338 + ], + "type": "text", + "content": "3. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Ho, D., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E., Ruano, R.J., Jeffrey, K., Jesmonth, S., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Lee, K.H., Levine, S., Lu, Y., Luu, L., Parada, C., Pastor, P., Quiambao, J., Rao, K., Rettinghouse, J., Reyes, D., Sermanet, P., Sievers, N., Tan, C., Toshev, A., Vanhoucke, V., Xia, F., Xiao, T., Xu, P., Xu, S., Yan, M., Zeng, A.: Do as i can and not as i say: Grounding language in robotic affordances. In: arXiv preprint arXiv:2204.01691 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 338, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 381 + ], + "type": "text", + "content": "4. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 382, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 481, + 435 + ], + "type": "text", + "content": "5. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "text", + "content": "6. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (vpt): Learning to act by watching unlabeled online videos (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 469, + 481, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 481, + 513 + ], + "type": "text", + "content": "7. Bellemare, M.G., Naddaf, Y., Veness, J., Bowling, M.: The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research 47, 253-279 (Jun 2013). https://doi.org/10.1613/jair.3912, http://dx.doi.org/10.1613/jair.3912" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 514, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 533 + ], + "type": "text", + "content": "8. Billard, A., Kragic, D.: Trends and challenges in robot manipulation. Science 364(6446), eaat8414 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 555 + ], + "type": "text", + "content": "9. Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., Zaremba, W.: Openai gym. arXiv preprint arXiv:1606.01540 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 556, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 481, + 600 + ], + "type": "text", + "content": "10. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 632 + ], + "type": "text", + "content": "11. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "12. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. Chen, J., Mu, Y., Yu, Q., Wei, T., Wu, S., Yuan, Z., Liang, Z., Yang, C., Zhang, K., Shao, W., Qiao, Y., Xu, H., Ding, M., Luo, P.: Roboscript: Code generation for free-form manipulation tasks across real and simulation (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "14. Chen, L., Li, B., Shen, S., Yang, J., Li, C., Keutzer, K., Darrell, T., Liu, Z.: Language models are visual reasoning coordinators. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "text", + "content": "15. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 480, + 258 + ], + "type": "text", + "content": "16. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 258, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 258, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 480, + 279 + ], + "type": "text", + "content": "17. Das, A., Datta, S., Gkioxari, G., Lee, S., Parikh, D., Batra, D.: Embodied question answering (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 279, + 480, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 279, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 133, + 279, + 480, + 333 + ], + "type": "text", + "content": "18. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model. In: arXiv preprint arXiv:2303.03378 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 333, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 333, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 333, + 480, + 354 + ], + "type": "text", + "content": "19. Evans, J.S.B.: Dual-processing accounts of reasoning, judgment, and social cognition. Annu. Rev. Psychol. 59, 255-278 (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 354, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 480, + 387 + ], + "type": "text", + "content": "20. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 387, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 387, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 133, + 387, + 480, + 408 + ], + "type": "text", + "content": "21. Fu, H., Xu, W., Ye, R., Xue, H., Yu, Z., Tang, T., Li, Y., Du, W., Zhang, J., Lu, C.: Rfuniverse: A multiphysics simulation platform for embodied ai (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 408, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 408, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 408, + 480, + 430 + ], + "type": "text", + "content": "22. Fu, Z., Zhao, T.Z., Finn, C.: Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In: arXiv (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 430, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 430, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 430, + 480, + 450 + ], + "type": "text", + "content": "23. Gao, X., Gong, R., Shu, T., Xie, X., Wang, S., Zhu, S.C.: Vrkitchen: an interactive 3d virtual environment for task-oriented learning (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 483 + ], + "type": "text", + "content": "24. Gu, S., Holly, E., Lillicrap, T., Levine, S.: Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In: 2017 IEEE international conference on robotics and automation (ICRA). pp. 3389-3396. IEEE (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "type": "text", + "content": "25. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 515, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 480, + 548 + ], + "type": "text", + "content": "26. Huang, S., Jiang, Z., Dong, H., Qiao, Y., Gao, P., Li, H.: Instruct2act: Mapping multi-modality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 548, + 480, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 548, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 133, + 548, + 480, + 580 + ], + "type": "text", + "content": "27. Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "type": "text", + "content": "28. Hussein, A., Gaber, M.M., Elyan, E., Jayne, C.: Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR) 50(2), 1-35 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 601, + 386, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 386, + 612 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 386, + 612 + ], + "type": "text", + "content": "29. Kahneman, D.: Thinking, fast and slow. macmillan (2011)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "text", + "content": "30. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "31. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "text", + "content": "32. Li, C., Zhang, R., Wong, J., Gokmen, C., Srivastava, S., Martin-Martin, R., Wang, C., Levine, G., Lingelbach, M., Sun, J., et al.: Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In: Conference on Robot Learning. pp. 80–93. PMLR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 195 + ], + "type": "text", + "content": "33. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 217 + ], + "type": "text", + "content": "34. Lin, K., Agia, C., Migimatsu, T., Pavone, M., Bohg, J.: Text2motion: From natural language instructions to feasible plans. arXiv preprint arXiv:2303.12153 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "35. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "36. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 422, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 422, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 422, + 285 + ], + "type": "text", + "content": "37. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 286, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 481, + 308 + ], + "type": "text", + "content": "38. Liu, Z., Dong, Y., Rao, Y., Zhou, J., Lu, J.: Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 330 + ], + "type": "text", + "content": "39. Mao, H., Wang, C., Hao, X., Mao, Y., Lu, Y., Wu, C., Hao, J., Li, D., Tang, P.: Seihai: A sample-efficient hierarchical ai for the minerl competition (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "type": "text", + "content": "40. MosaicML: Mpt-7b (2023), https://www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-23" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "text", + "content": "41. Mu, Y., Zhang, Q., Hu, M., Wang, W., Ding, M., Jin, J., Wang, B., Dai, J., Qiao, Y., Luo, P.: Embodiedgpt: Vision-language pre-training via embodied chain of thought. arXiv preprint arXiv:2305.15021 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 387, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 481, + 430 + ], + "type": "text", + "content": "42. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 431, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 464 + ], + "type": "text", + "content": "43. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 481, + 498 + ], + "type": "text", + "content": "44. Puig, X., Ra, K., Boben, M., Li, J., Wang, T., Fidler, S., Torralba, A.: Virtual-home: Simulating household activities via programs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8494-8502 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "text", + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 565 + ], + "type": "text", + "content": "46. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 566, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 587 + ], + "type": "text", + "content": "47. Ramakrishnan, S.K., Jayaraman, D., Grauman, K.: An exploration of embodied visual exploration (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "text", + "content": "48. Rana, K., Haviland, J., Garg, S., Abou-Chakra, J., Reid, I., Suenderhauf, N.: Say-plan: Grounding large language models using 3d scene graphs for scalable task planning. arXiv preprint arXiv:2307.06135 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 666 + ], + "type": "text", + "content": "49. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339-9347 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Octopus: Embodied VLM from Environmental Feedback" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 578 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "50. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "51. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "52. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 227 + ], + "type": "text", + "content": "53. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "55. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "56. Wang, H., Liang, W., Gool, L.V., Wang, W.: Towards versatile embodied navigation (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 314, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 336 + ], + "type": "text", + "content": "57. Wu, Y., Wu, Y., Gkioxari, G., Tian, Y.: Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "58. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 357, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 480, + 391 + ], + "type": "text", + "content": "59. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: Sapien: A simulated part-based interactive environment (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "60. Xie, B., Zhang, S., Zhou, Z., Li, B., Zhang, Y., Hessel, J., Yang, J., Liu, Z.: Funqa: Towards surprising video comprehension. arXiv preprint arXiv:2306.14899 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "61. Yan, C., Misra, D., Bennnett, A., Walsman, A., Bisk, Y., Artzi, Y.: Chalet: Cornell house agent learning environment (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "62. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Arenas, M.G., Chiang, H.T.L., Erez, T., Hasenclever, L., Humplik, J., et al.: Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "63. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Skill reinforcement learning and planning for open-world long-horizon tasks (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "text", + "content": "64. Zheng, S., Liu, J., Feng, Y., Lu, Z.: Steve-eye: Equipping llm-based embodied agents with visual perception in open worlds (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "65. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "66. Zhu, Y., Wong, J., Mandlekar, A., Martin-Martín, R., Joshi, A., Nasiriany, S., Zhu, Y.: robosite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_content_list.json b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ebf17f3a6f7644140279383ff34bf058ab1994d6 --- /dev/null +++ b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_content_list.json @@ -0,0 +1,1835 @@ +[ + { + "type": "text", + "text": "Omni-Recon: Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yonggan Fu, Huaizhi Qu, Zhifan Ye, Chaojian Li, Kevin Zhao, and Yingyan (Celine) Lin", + "bbox": [ + 320, + 213, + 679, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Georgia Institute of Technology {yfu314, zye327, cli851, kzhao14, celine.lin}@gatech.edu", + "bbox": [ + 299, + 255, + 702, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent breakthroughs in Neural Radiance Fields (NeRFs) have sparked significant demand for their integration into real-world 3D applications. However, the varied functionalities required by different 3D applications often necessitate diverse NeRF models with various pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. Drawing inspiration from the generalization capability and adaptability of emerging foundation models, our work aims to develop one general-purpose NeRF for handling diverse 3D tasks. We achieve this by proposing a framework called Omni-Recon, which is capable of (1) generalizable 3D reconstruction and zero-shot multitask scene understanding, and (2) adaptability to diverse downstream 3D applications such as real-time rendering and scene editing. Our key insight is that an image-based rendering pipeline, with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our Omni-Recon features a general-purpose NeRF model using image-based rendering with two decoupled branches: one complex transformer-based branch that progressively fuses geometry and appearance features for accurate geometry estimation, and one lightweight branch for predicting blending weights of source views. This design achieves state-of-the-art (SOTA) generalizable 3D surface reconstruction quality with blending weights reusable across diverse tasks for zero-shot multitask scene understanding. In addition, it can enable real-time rendering after baking the complex geometry branch into meshes, swift adaptation to achieve SOTA generalizable 3D understanding performance, and seamless integration with 2D diffusion models for text-guided 3D editing. Our code is available at: https://github.com/GATECH-EIC/Omni-Recon.", + "bbox": [ + 261, + 327, + 743, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 743, + 375, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs) [41] have garnered significant attention and hold promise to become crucial enablers for emerging 3D applications. However, deploying NeRFs in real-world scenarios still requires substantial effort, as different 3D applications often demand distinct NeRF models with varied", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. For instance, while both cross-scene generalization and real-time rendering are desirable in the real world, generalizable NeRFs for instant reconstruction [10,38,77] and real-time NeRFs using mesh-based rasterization [12,65,85] typically feature diverse pipelines, making it hard to simultaneously satisfy the two requirements. Additionally, understanding new scene properties [36,90,93] necessitates training new NeRF models, which is not scalable given the increasing number of scene properties of interest.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In parallel, the emergence of foundation models has driven transformative progress in real-world artificial intelligence applications. Inspired by this exciting trend, there has been a growing interest in integrating NeRFs into this realm by developing general-purpose NeRFs for handling various 3D tasks. Specifically, similar to the generalization capability and adaptability of foundation models, we expect general-purpose NeRFs to possess three essential capabilities, including (1) generalizable 3D reconstruction for instant surface reconstruction and novel view synthesis, (2) zero-shot multitask scene understanding of various scene properties, e.g., semantics and edges, and (3) adaptability to diverse downstream tasks such as real-time rendering and scene editing. Achieving these capabilities is promising to enable users to effortlessly start from these general-purpose NeRFs when facing a new 3D application, significantly reducing the need for training and experimenting with various NeRF pipelines.", + "bbox": [ + 212, + 268, + 787, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the promise, existing NeRF pipelines have not yet realized the mentioned capabilities. Firstly, per-scene optimized NeRFs [2, 3, 22, 25, 27, 42, 85] rely on costly per-scene training, thus precluding generalization capabilities. Secondly, although generalizable NeRF designs [10, 38, 54, 68, 75, 77, 82, 88] can achieve cross-scene generalization, their huge computational complexity limits their suitability for application scenarios where real-time rendering is crucial. Some generalizable surface reconstruction methods [34, 55] can support mesh extraction, but an expensive volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views, leaving the demand for real-time rendering unmet. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored.", + "bbox": [ + 212, + 464, + 787, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the significant potential and associated challenges, our work aims to advance the integration of NeRFs into the realm of foundation models by developing general-purpose NeRFs. To achieve this goal, we propose a framework called Omni-Recon, which is an image-based rendering pipeline that realizes all the mentioned capabilities. The key insight leveraged by our Omni-Recon is that an image-based rendering pipeline, when equipped with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our contributions are summarized as follows:", + "bbox": [ + 212, + 630, + 787, + 767 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "- We introduce Omni-Recon, a general-purpose NeRF pipeline, which possesses both decent generalization capabilities, including generalizable 3D reconstruction and zero-shot multitask scene understanding, and adaptability to diverse downstream 3D applications.", + "bbox": [ + 223, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our Omni-Recon features a general-purpose NeRF model backbone using image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and appearance features to accurately predict a signed distance function (SDF), and one lightweight appearance branch for predicting blending weights of source views from the target scene. The advantages of this design include: (1) ensuring accurate geometry estimation for high-quality 3D reconstruction, and (2) enabling real-time rendering after baking the complex geometry branch into meshes and using the lightweight appearance branch as a Shader.", + "- We intriguingly find that a properly trained image-based rendering pipeline can directly lift 2D monocular priors to 3D novel views by reusing the blending weights predicted by the appearance branch and leverage this to enable generalizable, zero-shot, and multitask scene understanding.", + "- We demonstrate the adaptability of Omni-Recon by extending it to diverse downstream tasks, including: (1) a Nvdiffrast-based [31] real-time rendering pipeline using the extracted mesh, (2) parameter-efficient tuning (PET) of Omni-Recon on downstream generalizable scene understanding tasks, and (3) a new text-guided scene editing scheme by iteratively editing and reconstructing the source images from a 3D scene using 2D diffusion models.", + "- Extensive experiments demonstrate that our Omni-Recon can achieve state-of-the-art (SOTA) generalizable 3D surface reconstruction and scene understanding performance. Additionally, it can support real-time rendering of a new scene after rapid adaptation, along with easy-to-use 3D scene editing." + ], + "bbox": [ + 223, + 146, + 792, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Omni-Recon underscores the potentially wide usage of image-based rendering pipelines in diverse real-world 3D applications, which may have been overlooked by previous works. We believe that the insights we provide could ignite future innovations in more advanced generalizable rendering pipelines.", + "bbox": [ + 212, + 507, + 785, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 585, + 387, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields. NeRFs [42], which represent the target scene as a continuous volume with density and view-dependent color, have garnered increasing attention thanks to their SOTA rendering quality when compared to prior approaches using explicit neural representations [24,39,60,66]. Subsequent works further improve NeRFs from various angles, such as enhancing their rendering quality [2,3,11,72], training NeRFs with sparse views [18,28,46,69,80,81], extending NeRFs to large-scale scenes [64,70,78,79,91], or applying NeRFs to other tasks, such as generative modeling [7,47,58,74], dynamic scenes [6,33,49, 50], or scene understanding [21,30,73,93].", + "bbox": [ + 212, + 612, + 785, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalizable NeRFs. Vanilla NeRFs rely on per-scene optimization, which limits their ability to perform instant reconstruction. To address this limitation, generalizable NeRF designs [10, 13, 29, 38, 54, 68, 75, 77, 82, 88] achieve cross-scene generalization by conditioning vanilla NeRFs on the source views from the target new scene. Specifically, [68, 75, 77, 88] take the extracted scene features from the source views as inputs to reconstruct the radiance field via a one-shot forward", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pass. Subsequent works further improve cross-scene generalization capabilities by enhancing scene geometry estimation [10, 29, 38] and by introducing attention mechanisms and mixture-of-experts into NeRF backbone design [15, 36, 75]. Despite the promise of generalizable NeRFs, they still fall short in terms of rendering efficiency due to their huge computational complexity, which entails costly inference for each sampled point along the ray. While some generalizable surface reconstruction methods [34, 55] can support mesh extraction, a costly volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views. This poses a hindrance to their deployment in 3D applications that require real-time rendering. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient scene representations. Various scene representations have been proposed to accelerate NeRFs' training and inference process. To speed up NeRF training, [8,9,20,43,62] integrate explicit neural representations with volumetric rendering. In parallel, to speed up NeRF inference/rendering, [35,52,53] attempt to reduce the complexity of the multi-layer perceptron (MLP) model in NeRF. In addition, [22, 25, 27, 37, 53, 59, 87] explore the free space in a target 3D scene to improve the sampling efficiency and thus reduce unnecessary MLP inference. Motivated by the insight of deferred rendering [67], [12, 25, 65, 85] pre-compute scene features offline and thus require less computation at run-time. In particular, [12, 65, 85] represents NeRF as a set of textured polygons, where the textures can be pre-computed and thus the NeRF pipeline can be efficiently executed in modern graphics hardware. However, these efficient scene representations still necessitate costly per-scene optimization, lacking generalization capabilities.", + "bbox": [ + 212, + 313, + 787, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Omni-Recon: The General-Purpose NeRF Backbone", + "text_level": 1, + "bbox": [ + 214, + 518, + 759, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To develop general-purpose NeRFs for diverse 3D tasks, in this section, we design a general-purpose NeRF backbone and demonstrate its SOTA generalizable 3D reconstruction capability. We will leverage this backbone to enable zero-shot multitask scene understanding in Sec. 4 and adapt it to diverse downstream 3D tasks, including real-time rendering, PET for 3D scene understanding, and text-guided scene editing in Sec. 5.", + "bbox": [ + 212, + 542, + 787, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Backbone Design Overview", + "text_level": 1, + "bbox": [ + 214, + 642, + 486, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Design principles. Our design is driven by the following principles: (1) Similar to foundation models, the ability to instantly reconstruct a new scene with generalizable capabilities is essential. Therefore, we leverage image-based rendering [10,38,77], which lifts 2D image features into their 3D counterparts, as our backbone design and uncover their decent capabilities in diverse 3D applications, which may have been overlooked by previous works. (2) Considering a wide range of 3D applications that require real-time rendering, it is highly desirable that our backbone design can be seamlessly converted into efficient 3D representations, particularly meshes, to enable real-time rendering.", + "bbox": [ + 212, + 657, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our backbone design. As shown in Fig. 1, the backbone model in OmniRecon leverages image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5556299a659e0430e950b9d7ef37408a3da658d642f050271fb4e782aae15406.jpg", + "image_caption": [ + "Fig. 1: An overview of our Omni-Recon framework." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 777, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "appearance features to accurately predict an SDF, and one lightweight appearance branch for estimating blending weights of source views. Specifically, the progressive feature fusion in the geometry branch empowers accurate geometry estimation for high-quality 3D reconstruction, and the decoupling between geometry and appearance enables real-time rendering after baking the complex geometry branch into meshes, with the lightweight appearance branch serving as a shader. Additionally, the nature of image-based rendering, when equipped with accurate geometry and appearance estimation, can enable a wide range of downstream 3D applications, as shown in Sec. 4 and Sec. 5.", + "bbox": [ + 212, + 415, + 787, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 The Overall Image-Based Rendering Pipeline", + "text_level": 1, + "bbox": [ + 215, + 569, + 635, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following generalizable NeRFs [10,38,77,88], our NeRF backbone is conditioned on source views from the target scene. This is achieved by estimating the geometry and appearance based on the extracted features of the source views from the target scene, which is elaborated as follows.", + "bbox": [ + 212, + 595, + 785, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acquire geometry features. Given $N$ source views $\\{I_i\\}_{i=1}^N \\in \\mathbb{R}^{H \\times W \\times 3}$ from the target scene, we utilize a CNN encoder [34, 55] to extract source features $\\{\\mathbf{F}_i\\}_{i=1}^N \\in \\mathbb{R}^{H \\times W \\times C}$ from the source views. Next, we construct a 3D feature volume $V \\in \\mathbb{R}^{M \\times M \\times M \\times C}$ , which serves similar roles as cost volumes for multiview stereo [10, 83], to aggregate geometry information from different source views, following [44, 55, 63]. Specifically, we divide the bounding volume of the scene into $M^3$ voxels and project each voxel center onto $N$ source views, where the mean and variance of $N$ source features are computed and concatenated as the corresponding voxel feature. We then leverage a 3D U-Net [14] to further enhance the feature volume to acquire the final geometry feature volume $V$ .", + "bbox": [ + 212, + 656, + 787, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acquire appearance features. For each sampled 3D point $\\mathbf{p}(t) = \\mathbf{o} + t\\mathbf{d}$ along the ray emitted from the target pixel on the novel view with camera origin $\\mathbf{o}$", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and direction $\\mathbf{d}$ , we project it on the source views to acquire the projected colors $\\{\\mathbf{c}_i\\}_{i=1}^N$ and appearance features $\\{\\mathbf{f}_i\\}_{i=1}^N$ using bilinear interpolation.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometry and appearance estimation. For more accurate reconstruction of the scene geometry and mitigation of the intrinsic color-density ambiguity [76], we estimate the per-point SDF and model the density as a function of SDF, following [40,55,76]. For appearance, we estimate the per-point radiance through color blending, i.e., predicting the blending weights of the projected colors on the source views to compose the per-point color, following [38,77,88]. The geometry and appearance are modeled using separate branches and can be formulated as:", + "bbox": [ + 212, + 178, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns = \\mathbf {M} _ {s d f} (\\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}, V), \\quad \\{\\omega_ {i} \\} _ {i = 1} ^ {N} = \\mathbf {M} _ {c o l o r} (\\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}, \\mathbf {d}), \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 289, + 785, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s$ is the SDF, $\\{\\omega_i\\}_{i=1}^N$ are the blending weights, and $\\mathbf{M}_{sdf}$ and $\\mathbf{M}_{color}$ are the geometry and appearance branches, respectively. The per-point radiance can be derived as the weighted sum $\\hat{\\mathbf{c}} = \\sum_{i=1}^{N} \\omega_i \\mathbf{c}_i$ across $N$ source views.", + "bbox": [ + 212, + 313, + 782, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For SDF-based volumetric rendering, we adopt the formulation in NeuS [76]. Specifically, the color is accumulated along the $K$ sampled points along each ray:", + "bbox": [ + 212, + 359, + 782, + 388 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} = \\sum_ {j = 1} ^ {K} T _ {j} \\alpha_ {j} \\hat {\\mathbf {c}} _ {j}, \\quad \\alpha_ {j} = 1 - \\exp \\left(- \\int_ {t _ {j}} ^ {t _ {j + 1}} \\rho (t) d t\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 390, + 785, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T_{j} = \\prod_{n=1}^{j-1}(1 - \\alpha_{n})$ is the accumulated transmittance and $\\rho(t)$ is the density determined by the estimated SDF, following the definition in NeuS [76]. Complexities of the two branches. When predicting the radiance of a sampled 3D point, the colors of valid projections on the source views are already close to the radiance of this 3D point. In light of this, we employ a lightweight $\\mathbf{M}_{color}$ , consisting of three MLP layers, for efficient blending weight prediction. In contrast, the geometry branch should be allocated with sufficient complexity for accurately estimating the occlusion effects among different sampled points along the same ray as well as that among the sampled points and the source views. We introduce the design of the geometry branch in the next subsection.", + "bbox": [ + 212, + 429, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 The Proposed Transformer-based Geometry Branch", + "text_level": 1, + "bbox": [ + 215, + 594, + 689, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview. Our geometry branch leverages transformer modules [71] to progressively fuse geometry and appearance features, thus properly handling the two occlusion effects mentioned above. Specifically, for the $K$ sampled 3D points along the ray, we acquire their geometry features $\\{\\mathbf{v}_k\\}_{k=1}^K$ from the geometry feature volume $V$ , which are sequentially processed through $B$ blocks ( $B = 2$ in our design), each consisting of a geometry transformer, an appearance transformer, and an occlusion transformer, to produce the per-point SDF.", + "bbox": [ + 212, + 616, + 787, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The geometry transformer. The geometry transformer fuses the geometry features into its inputs by using $\\{\\mathbf{v}_k\\}_{k=1}^K$ as the key and value in a cross-attention scheme [71] and performs attention across the sampled points along the same ray to model their occlusion effects. This process can be formulated as follows:", + "bbox": [ + 212, + 722, + 787, + 782 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {s d f} ^ {g e o} (\\mathbf {x}, \\{\\mathbf {v} _ {k} \\} _ {k = 1} ^ {K}) = C r o s s A t t e n t i o n (\\mathbf {q} = \\mathbf {x}, \\mathbf {k} = \\mathbf {v} = \\{\\mathbf {v} _ {k} \\} _ {k = 1} ^ {K}), \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 795, + 785, + 815 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is the input to the geometry transformer, which is exactly the geometry features themselves for the geometry transformer in the first block, and $\\mathbf{q}$ , $\\mathbf{k}$ , and $\\mathbf{v}$ are the query, key, and value of the attention scheme [71], respectively.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The appearance transformer. The appearance transformer integrates the appearance features $\\{\\mathbf{f}_i\\}_{i=1}^N$ into the outputs of the geometry transformer, considering the potential occlusions among the sampled points and the source views that may cause invalid projections. This integration is achieved using subtraction attention [75,92], which is more effective for geometric relationship reasoning [75,92], with $\\{\\mathbf{f}_i\\}_{i=1}^N$ as the key and value. Specifically, subtraction attention computes the attention scores between the input query features and each source view in the key features, which are then used to aggregate different source views in the value features into new output features with the same dimension as the input query features. This process can be formulated as follows:", + "bbox": [ + 212, + 191, + 784, + 342 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {s d f} ^ {a p p r} (\\mathbf {x}, \\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}) = S u b A t t e n t i o n (\\mathbf {q} = \\mathbf {x}, \\mathbf {k} = \\mathbf {v} = \\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 349, + 784, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is the output of $\\mathbf{M}_{sdf}^{geo}$ . The detailed formulation of subtraction attention is provided in the supplementary materials.", + "bbox": [ + 212, + 376, + 782, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The occlusion transformer. We further integrate an occlusion transformer that performs self-attention across sampled points along the same ray to model their occlusion effects explicitly:", + "bbox": [ + 212, + 407, + 782, + 452 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {s d f} ^ {\\text {o c c}} (\\mathbf {x}) = \\text {S e l f A t t e n t i o n} (\\mathbf {q} = \\mathbf {k} = \\mathbf {v} = \\mathbf {x}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 462, + 784, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is the output of $\\mathbf{M}_{sdf}^{app}$ . We apply another MLP on top of the output features of the last $M_{sdf}^{occ}$ module to predict the per-point SDF $s$ in Eq. 1.", + "bbox": [ + 212, + 486, + 782, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Training Objectives", + "text_level": 1, + "bbox": [ + 214, + 530, + 423, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our NeRF backbone is trained using two objectives: (1) a photometric loss $\\mathcal{L}_{rgb}$ between the rendered images and the ground-truth images, and (2) a depth loss $\\mathcal{L}_{depth}$ between the rendered depth $\\hat{\\mathbf{D}} = \\sum_{j=1}^{K} T_j \\alpha_j t$ and the ground-truth depth. The overall objective can be formulated as follows:", + "bbox": [ + 212, + 551, + 782, + 613 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\beta \\mathcal {L} _ {\\text {d e p t h}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 625, + 784, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{rgb} = \\frac{1}{R}\\sum_{\\mathbf{r} = 1}^{R}\\left\\| \\hat{\\mathbf{C}}_{\\mathbf{r}} - \\mathbf{C}_{\\mathbf{r}}\\right\\|_{2}^{2}$ with $R$ denoting the number of rendered pixels, $\\mathcal{L}_{depth} = \\frac{1}{R}\\sum_{\\mathbf{r} = 1}^{R}\\left\\| \\hat{\\mathbf{D}}_{\\mathbf{r}} - \\mathbf{D}_{\\mathbf{r}}\\right\\|_{2}^{2}$ , and $\\beta$ is set as 1.", + "bbox": [ + 212, + 646, + 782, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Evaluation: Generalizable 3D Reconstruction", + "text_level": 1, + "bbox": [ + 214, + 712, + 633, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup. We assess the backbone of Omni-Recon by evaluating its generalizable 3D reconstruction capability. Datasets: Following [10, 34, 40, 55], we adopt the DTU dataset [1], which comprises 124 different scenes and 7 distinct lighting conditions. For testing, 15 scenes are utilized, while the remaining scenes are allocated for training, following the dataset split in [34, 40, 55]. Training settings: During training, we employ $N = 4$ source views with a resolution of $640 \\times 512$ . The ray number per batch and batch size are set to 1024 and 2, respectively,", + "bbox": [ + 212, + 733, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/392c5243bf0bcf854f9c5821518b5bda36908a739af619ada6a432a7161008d9.jpg", + "table_caption": [ + "Table 1: Benchmark the quantitative performance in sparse view mesh reconstruction in terms of Chamfer distance (the lower, the better) on 15 testing scenes on DTU. The best scores are in bold, and the second-best are underlined." + ], + "table_footnote": [], + "table_body": "
MethodMean243740556365698397105106110114118122
COLMAP [57]1.520.902.891.631.082.181.941.611.302.341.281.101.420.761.171.14
MVSNet [83]1.221.052.521.711.041.451.520.881.291.381.050.910.660.611.081.16
IDR [86]3.394.016.403.521.913.962.364.851.626.375.971.234.730.911.721.26
VolSDF [84]3.414.034.216.120.918.241.732.741.825.143.092.084.810.603.512.18
UNISURF [48]4.395.087.183.965.304.612.243.943.145.633.405.096.382.984.052.81
NeuS [76]4.004.574.493.974.324.631.954.683.834.152.501.526.471.265.576.11
PixelNeRF [88]6.185.138.075.854.407.114.645.686.769.056.113.955.926.266.896.93
IBRNet [77]2.322.293.702.661.833.022.831.772.282.731.961.872.131.582.052.09
MVSNeRF [10]2.091.963.272.541.932.572.711.821.722.291.751.721.471.292.092.26
SparseNeuS [40]1.962.173.292.741.672.692.421.581.861.941.351.501.450.981.861.87
VolRecon [55]1.381.202.591.561.081.431.921.111.481.421.051.191.380.741.231.27
ReTR [34]1.171.052.311.440.981.181.520.881.351.300.871.070.770.591.051.12
Omni-Recon (Ours)1.130.912.131.520.931.091.700.841.291.200.831.040.810.551.051.05
", + "bbox": [ + 223, + 186, + 776, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "following [34,55]. Mesh extraction: We utilize TSDF fusion [45,63] to reconstruct scene meshes from predicted source view depths, following [34,55]. More detailed model/training settings can be found in the supplementary materials.", + "bbox": [ + 212, + 366, + 784, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We primarily benchmark against SOTA generalizable neural implicit reconstruction methods [16, 34, 55], along with generalizable neural rendering methods [10, 77, 88], per-scene optimization methods [48, 76, 84, 86], and multiview stereo methods [57, 83]. We directly report their official results in [34, 55].", + "bbox": [ + 212, + 412, + 784, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse view reconstruction. We benchmark the performance of sparse view reconstruction using three views from each testing scene in DTU, following the setting in [16, 34, 55]. As shown in Tab. 1, we can observe that (1) Our Omni-Recon can achieve new SOTA reconstruction performance, with the lowest Chamfer distance averaged over all scenes. Additionally, as visualized in Fig. 2 (Row 1), our reconstructed mesh can exhibit smooth surfaces and correctly maintain the structure, e.g., the hole in the mouth. (2) Our method achieves the best reconstruction performance in 10 out of 15 scenes, indicating its general effectiveness across diverse scenes.", + "bbox": [ + 212, + 474, + 787, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to more diverse scenes. To demonstrate the generalization capability on more diverse scenes with larger domain shifts compared to the training scenes, we further benchmark the mesh reconstruction quality on the NeRF-Synthetic dataset [41] and NSVF-Synthetic [37] under a full-view reconstruction setting, i.e., the depth of each source view from each scene's training set is estimated and fused into a mesh using TSDF fusion [45, 63]. As visualized in Fig. 2 (Rows 2-4), we observe that our Omni-Recon can deliver notably higher-quality meshes with better-maintained structures, fewer holes, and smoother surfaces compared to the strongest baselines ReTR [34] and VolRecon [55]. The decent generalization capability of our method is attributed to the proper decoupling and fusion of geometry and appearance features in our backbone. More visualizations and ablation studies, such as the contributions of each backbone component, are provided in the supplementary materials.", + "bbox": [ + 212, + 611, + 787, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Rendering quality. We further benchmark the performance of generalizable novel view rendering on the test scenes from DTU. As shown in Tab. 2, our", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/253e7b23dc30b8c24932e32989d56897591d241547df22e0ec08fea8de4aceb3.jpg", + "image_caption": [ + "Fig. 2: Visualize the reconstructed mesh of our Omni-Recon and the two strongest baselines [34, 55]. Row 1: Scan65 from the test scenes of DTU; Rows 2-4: Scenes from NeRF-Synthetic [41] and NSVF-Synthetic [37], which present relatively challenging cases due to domain shifts w.r.t. the training scenes from DTU." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 759, + 452 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9ff302180515419a10fe20e0b50cdca2ecaeee10f37c1cfedf4bacf84920425f.jpg", + "table_caption": [ + "Table 2: Benchmark the quality of novel view rendering in terms of PSNR on DTU." + ], + "table_footnote": [], + "table_body": "
MethodMean.243740556365698397105106110114118122
VolRecon [55]24.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]25.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Omni-Recon26.3224.7722.3323.9225.5626.3724.7525.1924.9424.9225.0628.3928.6328.0131.4930.5
", + "bbox": [ + 222, + 529, + 777, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "method achieves the highest PSNR, e.g., a $+0.73$ and $+1.74$ PSNR improvement over ReTR and VolRecon, respectively. As will be demonstrated in Sec. 5.4, this PSNR can be notably boosted with super-fast Nvdiffrast-based finetuning [31].", + "bbox": [ + 212, + 590, + 787, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Omni-Recon: Enable Zero-shot Scene Understanding", + "text_level": 1, + "bbox": [ + 212, + 652, + 763, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To enable the understanding of various 3D scene properties, one approach is to jointly train the NeRF backbone on all target tasks. However, this approach is not scalable due to the diversity of scene properties, such as semantic segmentation, edge detection, and keypoint prediction, making it costly to cover new properties of interest. Therefore, it is highly desirable to enable zero-shot scene understanding for new tasks that were not seen during training.", + "bbox": [ + 212, + 676, + 787, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Our Strategy: The Predict-then-Blend Strategy", + "text_level": 1, + "bbox": [ + 212, + 773, + 653, + 789 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our hypothesis. We hypothesize that with accurate geometry and appearance estimation, the blending weights initially learned for radiance can be repurposed for other scene understanding tasks. In essence, just as blending source colors", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2499968a06e76450b64793bc71a72b1a6d2e791f72f49d34b6b6f2244b7a3f6b.jpg", + "image_caption": [ + "Fig. 3: Benchmark our predict-then-blend strategy with the render-then-predict baseline across different scene understanding tasks on Replica [61] and ScanNet [17]." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 774, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "yields radiance for each sampled 3D point, blending the properties of source views using the same weights can yield other properties. This hypothesis stems from the observation that scene appearance is closely related to other scene properties and thus regions with similar appearance are likely to share similar scene properties, such as semantics and edges. As such, the blending weights learned for appearance can also indicate the weighting for other scene properties.", + "bbox": [ + 212, + 542, + 787, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our method. Motivated by this hypothesis, our Omni-Recon employs a predict-then-blend strategy to enable zero-shot multitask scene understanding by lifting 2D monocular priors to 3D novel views. Specifically, considering the prevalence of 2D monocular vision models, our method first utilizes pretrained 2D models to generate predictions (e.g., logits in semantic segmentation) on each source view, denoted as $\\{\\mathbf{P}_i\\}_{i=1}^N$ . Then, the property of each sampled point can be obtained by blending their projected predictions on source views $\\{\\mathbf{p}_i\\}_{i=1}^N$ , reusing the RGB blending weights, i.e., $\\hat{\\mathbf{p}} = \\sum_{i=1}^{N} \\omega_i \\mathbf{p}_i$ . The final pixel-wise prediction can be derived using volumetric rendering similar to RGB reconstruction in Eq. 2.", + "bbox": [ + 212, + 635, + 787, + 773 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation: Zero-shot Multitask Scene Understanding", + "text_level": 1, + "bbox": [ + 214, + 786, + 705, + 803 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Setup. To evaluate the zero-shot scene understanding capabilities of our model, we adopt two indoor 3D scene datasets: Replica [61] and ScanNet [17], following", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the dataset splits in [93] and [36], respectively. Specifically, we focus on four scene understanding tasks: semantic segmentation, edge detection, and two keypoint detection tasks defined in [89]. For these tasks, the 2D monocular priors are provided by [51], [5], and [56], respectively. We employ mIoU for the semantic segmentation task and $\\ell_1$ error for the other three tasks as default metrics.", + "bbox": [ + 212, + 146, + 784, + 222 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baseline. An intuitive way to achieve zero-shot 3D scene understanding is to leverage existing 2D monocular models to predict the scene properties on top of the rendered images, a process referred to as a render-then-predict pipeline. We benchmark this strategy against our proposed predict-then-blend strategy.", + "bbox": [ + 212, + 222, + 785, + 281 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zero-shot scene understanding benchmark. As shown in Tab. 3 and Fig. 3, we observe that (1) Our predict-then-blend strategy achieves decent scene understanding performance on unseen scenes and unseen tasks, which echoes our hypothesis in Sec. 4.1 that 2D monocular priors can be elevated to 3D novel views through", + "bbox": [ + 212, + 282, + 486, + 417 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a5a803a16789e034a88f4166eb0abbf59d457244b624b5dce79c7549634176b2.jpg", + "table_caption": [ + "Table 3: Benchmark the two zero-shot scene understanding strategies across tasks." + ], + "table_footnote": [], + "table_body": "
DatasetMethod↑ Semantic↓ Edge↓ Key Point↓ Key Point 3D
Replica [61]Render-then -Predict15.640.04560.11010.0470
Predict-then -Blend (Ours)32.110.04120.07740.0176
ScanNet [17]Render-then -Predict41.320.04710.05680.0412
Predict-then -Blend (Ours)61.110.04340.04240.0197
", + "bbox": [ + 501, + 310, + 784, + 415 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "blending weight reuse, thanks to the high correlation between scene appearance and other properties; (2) Compared to the render-then-predict scheme, our method achieves notably higher quantitative accuracy, e.g., a $19.79\\%$ higher mIoU on ScanNet, and consistently better visual effects across all tasks. We attribute this improvement to two aspects: (1) Our predict-then-blend strategy can avoid feeding noisy inputs caused by rendering errors to the 2D monocular models, thus reducing reliance on the robustness of monocular models and ensuring more accurate monocular priors; (2) The multiview information gathered by our strategy can contribute to a more accurate understanding of objects with limited observations under a monocular setting. More experiments on other vision models like CLIP-LSeg [32] are in the supplementary materials.", + "bbox": [ + 212, + 417, + 787, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Omni-Recon: Adaptability to Downstream 3D Tasks", + "text_level": 1, + "bbox": [ + 214, + 590, + 756, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition to the generalization capabilities, we further demonstrate the adaptability of our Omni-Recon by extending it to three different types of downstream 3D applications. These demonstrations highlight the potentially broad utility of image-based rendering pipelines, which may have been overlooked by previous works, and can be generalized to a wider range of real-world 3D applications.", + "bbox": [ + 212, + 616, + 785, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Application 1: A New Real-time Rendering Pipeline", + "text_level": 1, + "bbox": [ + 214, + 698, + 692, + 714 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To rapidly enable real-time rendering using Omni-Recon's pretrained backbone, we build a real-time rendering pipeline based on Nvdiffrast [31]. This involves baking the complex geometry branch into meshes using TSDF fusion [45,63] as discussed in Sec. 3.5, and using the lightweight appearance branch as a Shader.", + "bbox": [ + 212, + 719, + 785, + 780 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our real-time rendering pipeline. As illustrated in Fig. 1 (bottom left), for any novel view of a new scene, we first transform its camera pose to clip space as required by Nvdiffrast [31]. Then, we perform a rasterization process to obtain the intersection point of each camera ray with the extracted mesh. Next, these", + "bbox": [ + 212, + 780, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "intersection points are input into the shader to predict the color by blending their projected colors on the source views as described in Sec. 3.2. Thanks to the blending scheme, where valid projections can provide strong appearance clues, the rendering quality is often satisfactory even without finetuning.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Joint mesh and Shader finetuning. To further enhance the rendering quality, we can jointly finetune the mesh and the Shader in a gradient-based manner using differentiable rendering enabled by Nvdiffrast [31]. Specifically, we finetune the location of each vertex in the mesh and the parameters of the lightweight Shader. Additionally, we periodically prune redundant mesh faces that cannot intersect with any camera rays in one training epoch. More details are provided in the supplementary materials.", + "bbox": [ + 212, + 207, + 784, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Application 2: Parameter-Efficient Tuning for 3D Understanding", + "text_level": 1, + "bbox": [ + 214, + 321, + 782, + 338 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We explore the possibility of finetuning Omni-Recon for various downstream 3D tasks, utilizing a series of 3D scene understanding tasks as demonstrations. Specifically, we conduct PET [19] on top of Omni-Recon's pretrained backbone by incorporating low-rank adapters [26] into the linear layers within the transformer modules introduced in Sec. 3.3, and only tuning these adapters while keeping the pretrained backbone frozen. It is worth noting that when evaluating the performance of the tuned model, we assess it on new scenes that have never been encountered during both training and tuning, representing a generalizable scene understanding scenario [36].", + "bbox": [ + 212, + 342, + 784, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Application 3: Text-guided 3D Scene Editing", + "text_level": 1, + "bbox": [ + 214, + 488, + 632, + 503 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We target text-guided 3D scene editing, which involves editing a 3D scene based on textual instructions. To achieve this, our Omni-Recon capitalizes on the nature of image-based rendering, allowing edits made to 2D source views to be propagated to 3D novel views. Specifically, leveraging the availability of powerful 2D diffusion models, we use publicly available ones to edit the 2D source views. However, doing so naively may result in inconsistent source views, as the same textual instruction may lead to different visualizations, thus degrading reconstruction quality. Inspired by [23], which iteratively edits the training images and retransmits a new NeRF to achieve scene editing, we address the inconsistent editing issue by proposing an iterative editing and reconstruction pipeline.", + "bbox": [ + 212, + 508, + 784, + 659 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The iterative editing and reconstruction pipeline. To ensure consistent 3D editing, we iterate between source view editing and reconstruction as follows: First, we edit each source view using the diffusion model InstructPix2Pix [4], which is guided by both textual instructions and the original image to maintain image fidelity. While the edited source views can faithfully follow the textual instructions, they may exhibit inconsistencies in the edited region. Next, we reconstruct each source view by conditioning on their nearby source views, as described in Sec. 3.2. This process improves the 3D consistency among source views at the expense of reduced image fidelity and less adherence to instructions. We then iterate through the editing and reconstruction steps to progressively enhance image fidelity, 3D consistency, and adherence to instructions. The resulting source views are finally used to reconstruct the entire scene.", + "bbox": [ + 212, + 660, + 784, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7959c899f57713872cbe4f7112f394692f483946af3bdf56dfa96c6ffd985863.jpg", + "table_caption": [ + "Table 4: Benchmark the rendering speed and rendering quality of our real-time rendering pipeline and the baselines [34,55] on DTU. The best PSNR is in bold, and the shortest finetuning time that our pipelines surpass the strongest baseline is underlined." + ], + "table_footnote": [], + "table_body": "
MethodFPSMean243740556365698397105106110114118122
VolRecon [55]0.02924.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]0.02425.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Ours w/o ft.71.3(40.82)22.9620.1219.7122.2722.7824.5521.7721.5126.7222.3324.4922.5222.9324.6823.8424.12
Ours (ft. 10s)25.6821.4221.6824.0624.1228.1924.1023.9531.6524.4128.1525.6325.8526.1526.8226.89
Ours (ft. 20s)27.2122.6322.9225.1225.4230.0325.9526.1633.1926.2230.3627.4427.0426.9329.1529.65
Ours (ft. 30s)27.7823.223.2625.5425.730.5926.8326.9633.6626.4730.7328.1427.7027.130.1730.65
Ours (ft. 1min)28.3424.6924.1125.7626.0530.9327.6627.4933.6827.0730.9728.5128.5427.3531.1731.54
Ours (ft. 3min)28.9525.224.3225.9426.1632.1528.9927.8834.9427.3531.6228.9328.9727.5631.7032.49
Ours (ft. 5min)29.0225.3424.3625.6326.2132.1629.3327.8134.9427.3231.7429.0429.0527.6931.7432.89
", + "bbox": [ + 220, + 188, + 782, + 292 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Evaluation: Effectiveness on Different Downstream 3D Tasks", + "text_level": 1, + "bbox": [ + 214, + 305, + 761, + 319 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Real-time rendering. We evaluate our real-time rendering pipeline in Sec. 5.1 in terms of both rendering quality and rendering speed, measured on an NVIDIA A5000 GPU, using the DTU dataset. Specifically, we finetune the scene meshes, extracted by TSDF fusion as mentioned in Sec. 3.5, and the Shader in no more than 5 minutes and record the rendering quality at different time steps.", + "bbox": [ + 212, + 325, + 782, + 400 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Observations and analysis: As shown in Tab. 4, we can observe that (1) Without any finetuning, the rendering pipeline already exhibits certain rendering capabilities, despite the model backbone being trained using volumetric rendering with more than one sampled point; (2) Our rendering pipeline can be swiftly finetuned to significantly boost the PSNR. For instance, with just a 10-second finetuning, our rendering pipeline can match the average rendering quality of the strongest baseline, ReTR [34]; With 1-minute/2-minute finetuning, it can surpass ReTR by a +2.75 and +3.36 PSNR improvement on average, respectively; (3) Our rendering pipeline enables real-time rendering with 71.3 FPS, i.e., a >2458× speed-up over the baselines. Here, the assumption is that the feature extraction of each source view from a target scene can be done in a one-time effort and does not account for throughput. If we still consider this feature extraction latency for each rendering, the resulting 40.82 FPS still meets the real-time requirement.", + "bbox": [ + 212, + 402, + 784, + 598 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "These experiments indicate that given a new scene, our Omni-Recon is promising to enable both instant and accurate reconstruction and real-time rendering.", + "bbox": [ + 212, + 598, + 789, + 628 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PET for 3D Understanding. To evaluate the adaptability of our Omni-Recon, we apply PET to it for various scene understanding tasks on ScanNet [17]. Specifically, we finetune a LoRA adaptor [26] for each task on a set of training scenes and evaluate on", + "bbox": [ + 214, + 628, + 459, + 748 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c63928ecbcfdd75abcb62989ff79d2000c5ab9e24cabcd3753af41f87d89e7bb.jpg", + "table_caption": [ + "Table 5: Evaluate the effectiveness of PET on Omni-Recon and benchmark with SRay [36]." + ], + "table_footnote": [], + "table_body": "
Method↑ Sem. mIoU↑ Sem. Total Acc↑ Sem. Avg Acc↓ Edge↓ Key Point↓ Key Point 3D
SRay [36]57.1578.2462.55---
Ours (Zero-shot)61.1180.8069.170.04340.04240.0197
Ours (PET + Zero-shot)62.3581.8469.890.03420.02070.0132
", + "bbox": [ + 473, + 659, + 782, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "non-overlapping test scenes, representing a generalizable 3D understanding setting, following the dataset split described in [36]. We benchmark against the SOTA generalizable semantic segmentation work, SRay [36].", + "bbox": [ + 212, + 750, + 784, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Observations and analysis: As shown in Tab. 5, we can observe that (1) After PET, our model outperforms SRay in terms of generalizable semantic segmentation with a $+5.20\\%$ mIoU and a $+7.34\\%$ accuracy improvement; (2) Despite", + "bbox": [ + 212, + 795, + 785, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6c7be7d29fa88ab9cad67e961ec0abd8829608f9dfec4b17ab9ee1cd19f8a98c.jpg", + "image_caption": [ + "Fig. 4: Visualize the scene understanding performance on hard cases before/after PET." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 718, + 345 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/33fe245c26fdbe0ab18443afadf59a49d34d5e8ccb4aa221368daa5744f241c1.jpg", + "image_caption": [ + "Fig. 5: Three text-guided scene editing examples using our pipeline in Sec. 5.3." + ], + "image_footnote": [], + "bbox": [ + 250, + 364, + 754, + 518 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "having limited trainable parameters, PET can enhance the scene understanding performance across tasks, especially on challenging cases visualized in Fig. 4, indicating the adaptability of Omni-Recon to diverse downstream tasks.", + "bbox": [ + 212, + 542, + 785, + 587 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Text-guided scene editing. We visualize the edited examples in Fig. 5, showcasing the ability to ensure both instruction-following and 3D consistency.", + "bbox": [ + 212, + 588, + 785, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 633, + 359, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Motivated by the significant demand for supporting various functionalities required by different 3D applications through a unified NeRF model, our work aims to develop a general-purpose NeRF capable of handling a broad spectrum of 3D tasks. We propose a framework called Omni-Recon, which features a general-purpose NeRF model using image-based rendering with two separate branches, and demonstrate its integration with various 3D tasks. Specifically, our Omni-Recon can achieve SOTA generalizable 3D reconstruction quality, enable zero-shot multitask scene understanding, achieve SOTA scene understanding performance and real-time rendering after rapid adaptation, and support scene editing. Omni-Recon underscores the potential of image-based rendering pipelines in diverse real-world 3D applications, which may have been underestimated by prior studies, thus sparking advancements in future rendering pipelines.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 393, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The work is supported by the National Science Foundation (NSF) through the SCH program (Award number: 1838873) and CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA.", + "bbox": [ + 212, + 175, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. International Journal of Computer Vision 120, 153-168 (2016)", + "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "5. Canny, J.: A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence (6), 679-698 (1986)", + "6. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023)", + "7. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. CVPR (2021)", + "8. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517 (2022)", + "9. Chen, A., Xu, Z., Wei, X., Tang, S., Su, H., Geiger, A.: Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226 (2023)", + "10. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021)", + "11. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022)", + "12. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022)", + "13. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016)", + "15. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3193-3204 (2023)", + "16. Croce, F., Hein, M.: Sparse and imperceivable adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4724-4732 (2019)", + "17. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)", + "18. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022)", + "19. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence 5(3), 220-235 (2023)", + "20. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "21. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nef: 3d-to-2d label transfer for panoptic urban scene segmentation. arXiv preprint arXiv:2203.15224 (2022)", + "22. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)", + "23. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789 (2023)", + "24. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (2018)", + "25. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)", + "26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "27. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)", + "28. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021)", + "29. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)", + "31. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)", + "32. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN", + "33. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. CVPR (2021)", + "34. Liang, Y., He, H., Chen, Y.: Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems 36 (2024)", + "35. Lindell, D.B., Martel, J.N., Wetzstein, G.: Autoint: Automatic integration for fast neural rendering. CVPR (2021)", + "36. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023)", + "37. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022)", + "39. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. SIGGRAPH (2019)", + "40. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)", + "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)", + "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022)", + "44. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16. pp. 414-431. Springer (2020)", + "45. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)", + "46. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs." + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)", + "47. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)", + "48. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)", + "49. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. CVPR (2021)", + "50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)", + "51. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. ArXiv preprint (2021)", + "52. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: DeRF: Decomposed radiance fields. CVPR (2021)", + "53. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)", + "54. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021)", + "55. Ren, Y., Zhang, T., Pollefeys, M., Süsstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023)", + "56. Sax, A., Emi, B., Zamir, A.R., Guibas, L.J., Savarese, S., Malik, J.: Mid-level visual representations improve generalization and sample efficiency for learning visuomotor policies. (2018)", + "57. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016)", + "58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3D-aware image synthesis. NeurIPS (2020)", + "59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. arXiv preprint arXiv:2206.07695 (2022)", + "60. Seitz, S.M., Dyer, C.R.: Photorealistic scene reconstruction by voxel coloring. IJCV (1999)", + "61. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)", + "62. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)", + "63. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)", + "65. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)", + "66. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (2019)", + "67. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)", + "68. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15182-15192 (2021)", + "69. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. arXiv preprint arXiv:2211.11738 (2022)", + "70. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022)", + "71. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "72. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)", + "73. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021)", + "74. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022)", + "75. Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z., et al.: Is attention all nerf needs? arXiv preprint arXiv:2207.13298 (2022)", + "76. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)", + "77. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4690-4699 (2021)", + "78. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504 (2021)", + "79. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 106-122. Springer (2022)", + "80. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Omni-Recon", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "81. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 $\\{\\backslash \\mathrm{deg}\\}$ views. arXiv preprint arXiv:2211.16431 (2022)", + "82. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)", + "83. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)", + "84. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)", + "85. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)", + "86. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020)", + "87. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021)", + "88. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021)", + "89. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2018)", + "90. Zhang, M., Zheng, S., Bao, Z., Hebert, M., Wang, Y.X.: Beyond rgb: Scene-property synthesis with neural radiance fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 795-805 (2023)", + "91. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5449-5458 (2022)", + "92. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021)", + "93. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 659 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Y. Fu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_model.json b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89e63f3c6af386ded9d0ad3f6b30489513b9ab5b --- /dev/null +++ b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_model.json @@ -0,0 +1,2880 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.187 + ], + "angle": 0, + "content": "Omni-Recon: Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.214, + 0.681, + 0.245 + ], + "angle": 0, + "content": "Yonggan Fu, Huaizhi Qu, Zhifan Ye, Chaojian Li, Kevin Zhao, and Yingyan (Celine) Lin" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.256, + 0.703, + 0.285 + ], + "angle": 0, + "content": "Georgia Institute of Technology {yfu314, zye327, cli851, kzhao14, celine.lin}@gatech.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.328, + 0.744, + 0.717 + ], + "angle": 0, + "content": "Abstract. Recent breakthroughs in Neural Radiance Fields (NeRFs) have sparked significant demand for their integration into real-world 3D applications. However, the varied functionalities required by different 3D applications often necessitate diverse NeRF models with various pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. Drawing inspiration from the generalization capability and adaptability of emerging foundation models, our work aims to develop one general-purpose NeRF for handling diverse 3D tasks. We achieve this by proposing a framework called Omni-Recon, which is capable of (1) generalizable 3D reconstruction and zero-shot multitask scene understanding, and (2) adaptability to diverse downstream 3D applications such as real-time rendering and scene editing. Our key insight is that an image-based rendering pipeline, with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our Omni-Recon features a general-purpose NeRF model using image-based rendering with two decoupled branches: one complex transformer-based branch that progressively fuses geometry and appearance features for accurate geometry estimation, and one lightweight branch for predicting blending weights of source views. This design achieves state-of-the-art (SOTA) generalizable 3D surface reconstruction quality with blending weights reusable across diverse tasks for zero-shot multitask scene understanding. In addition, it can enable real-time rendering after baking the complex geometry branch into meshes, swift adaptation to achieve SOTA generalizable 3D understanding performance, and seamless integration with 2D diffusion models for text-guided 3D editing. Our code is available at: https://github.com/GATECH-EIC/Omni-Recon." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.744, + 0.377, + 0.76 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs) [41] have garnered significant attention and hold promise to become crucial enablers for emerging 3D applications. However, deploying NeRFs in real-world scenarios still requires substantial effort, as different 3D applications often demand distinct NeRF models with varied" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. For instance, while both cross-scene generalization and real-time rendering are desirable in the real world, generalizable NeRFs for instant reconstruction [10,38,77] and real-time NeRFs using mesh-based rasterization [12,65,85] typically feature diverse pipelines, making it hard to simultaneously satisfy the two requirements. Additionally, understanding new scene properties [36,90,93] necessitates training new NeRF models, which is not scalable given the increasing number of scene properties of interest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.465 + ], + "angle": 0, + "content": "In parallel, the emergence of foundation models has driven transformative progress in real-world artificial intelligence applications. Inspired by this exciting trend, there has been a growing interest in integrating NeRFs into this realm by developing general-purpose NeRFs for handling various 3D tasks. Specifically, similar to the generalization capability and adaptability of foundation models, we expect general-purpose NeRFs to possess three essential capabilities, including (1) generalizable 3D reconstruction for instant surface reconstruction and novel view synthesis, (2) zero-shot multitask scene understanding of various scene properties, e.g., semantics and edges, and (3) adaptability to diverse downstream tasks such as real-time rendering and scene editing. Achieving these capabilities is promising to enable users to effortlessly start from these general-purpose NeRFs when facing a new 3D application, significantly reducing the need for training and experimenting with various NeRF pipelines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Despite the promise, existing NeRF pipelines have not yet realized the mentioned capabilities. Firstly, per-scene optimized NeRFs [2, 3, 22, 25, 27, 42, 85] rely on costly per-scene training, thus precluding generalization capabilities. Secondly, although generalizable NeRF designs [10, 38, 54, 68, 75, 77, 82, 88] can achieve cross-scene generalization, their huge computational complexity limits their suitability for application scenarios where real-time rendering is crucial. Some generalizable surface reconstruction methods [34, 55] can support mesh extraction, but an expensive volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views, leaving the demand for real-time rendering unmet. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Given the significant potential and associated challenges, our work aims to advance the integration of NeRFs into the realm of foundation models by developing general-purpose NeRFs. To achieve this goal, we propose a framework called Omni-Recon, which is an image-based rendering pipeline that realizes all the mentioned capabilities. The key insight leveraged by our Omni-Recon is that an image-based rendering pipeline, when equipped with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our contributions are summarized as follows:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.225, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- We introduce Omni-Recon, a general-purpose NeRF pipeline, which possesses both decent generalization capabilities, including generalizable 3D reconstruction and zero-shot multitask scene understanding, and adaptability to diverse downstream 3D applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.794, + 0.282 + ], + "angle": 0, + "content": "- Our Omni-Recon features a general-purpose NeRF model backbone using image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and appearance features to accurately predict a signed distance function (SDF), and one lightweight appearance branch for predicting blending weights of source views from the target scene. The advantages of this design include: (1) ensuring accurate geometry estimation for high-quality 3D reconstruction, and (2) enabling real-time rendering after baking the complex geometry branch into meshes and using the lightweight appearance branch as a Shader." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.284, + 0.787, + 0.342 + ], + "angle": 0, + "content": "- We intriguingly find that a properly trained image-based rendering pipeline can directly lift 2D monocular priors to 3D novel views by reusing the blending weights predicted by the appearance branch and leverage this to enable generalizable, zero-shot, and multitask scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.344, + 0.787, + 0.433 + ], + "angle": 0, + "content": "- We demonstrate the adaptability of Omni-Recon by extending it to diverse downstream tasks, including: (1) a Nvdiffrast-based [31] real-time rendering pipeline using the extracted mesh, (2) parameter-efficient tuning (PET) of Omni-Recon on downstream generalizable scene understanding tasks, and (3) a new text-guided scene editing scheme by iteratively editing and reconstructing the source images from a 3D scene using 2D diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.435, + 0.787, + 0.495 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our Omni-Recon can achieve state-of-the-art (SOTA) generalizable 3D surface reconstruction and scene understanding performance. Additionally, it can support real-time rendering of a new scene after rapid adaptation, along with easy-to-use 3D scene editing." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.794, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.57 + ], + "angle": 0, + "content": "Our Omni-Recon underscores the potentially wide usage of image-based rendering pipelines in diverse real-world 3D applications, which may have been overlooked by previous works. We believe that the insights we provide could ignite future innovations in more advanced generalizable rendering pipelines." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.388, + 0.603 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.75 + ], + "angle": 0, + "content": "Neural Radiance Fields. NeRFs [42], which represent the target scene as a continuous volume with density and view-dependent color, have garnered increasing attention thanks to their SOTA rendering quality when compared to prior approaches using explicit neural representations [24,39,60,66]. Subsequent works further improve NeRFs from various angles, such as enhancing their rendering quality [2,3,11,72], training NeRFs with sparse views [18,28,46,69,80,81], extending NeRFs to large-scale scenes [64,70,78,79,91], or applying NeRFs to other tasks, such as generative modeling [7,47,58,74], dynamic scenes [6,33,49, 50], or scene understanding [21,30,73,93]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Generalizable NeRFs. Vanilla NeRFs rely on per-scene optimization, which limits their ability to perform instant reconstruction. To address this limitation, generalizable NeRF designs [10, 13, 29, 38, 54, 68, 75, 77, 82, 88] achieve cross-scene generalization by conditioning vanilla NeRFs on the source views from the target new scene. Specifically, [68, 75, 77, 88] take the extracted scene features from the source views as inputs to reconstruct the radiance field via a one-shot forward" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "pass. Subsequent works further improve cross-scene generalization capabilities by enhancing scene geometry estimation [10, 29, 38] and by introducing attention mechanisms and mixture-of-experts into NeRF backbone design [15, 36, 75]. Despite the promise of generalizable NeRFs, they still fall short in terms of rendering efficiency due to their huge computational complexity, which entails costly inference for each sampled point along the ray. While some generalizable surface reconstruction methods [34, 55] can support mesh extraction, a costly volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views. This poses a hindrance to their deployment in 3D applications that require real-time rendering. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Efficient scene representations. Various scene representations have been proposed to accelerate NeRFs' training and inference process. To speed up NeRF training, [8,9,20,43,62] integrate explicit neural representations with volumetric rendering. In parallel, to speed up NeRF inference/rendering, [35,52,53] attempt to reduce the complexity of the multi-layer perceptron (MLP) model in NeRF. In addition, [22, 25, 27, 37, 53, 59, 87] explore the free space in a target 3D scene to improve the sampling efficiency and thus reduce unnecessary MLP inference. Motivated by the insight of deferred rendering [67], [12, 25, 65, 85] pre-compute scene features offline and thus require less computation at run-time. In particular, [12, 65, 85] represents NeRF as a set of textured polygons, where the textures can be pre-computed and thus the NeRF pipeline can be efficiently executed in modern graphics hardware. However, these efficient scene representations still necessitate costly per-scene optimization, lacking generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.519, + 0.76, + 0.537 + ], + "angle": 0, + "content": "3 Omni-Recon: The General-Purpose NeRF Backbone" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.634 + ], + "angle": 0, + "content": "To develop general-purpose NeRFs for diverse 3D tasks, in this section, we design a general-purpose NeRF backbone and demonstrate its SOTA generalizable 3D reconstruction capability. We will leverage this backbone to enable zero-shot multitask scene understanding in Sec. 4 and adapt it to diverse downstream 3D tasks, including real-time rendering, PET for 3D scene understanding, and text-guided scene editing in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.643, + 0.488, + 0.657 + ], + "angle": 0, + "content": "3.1 Backbone Design Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Design principles. Our design is driven by the following principles: (1) Similar to foundation models, the ability to instantly reconstruct a new scene with generalizable capabilities is essential. Therefore, we leverage image-based rendering [10,38,77], which lifts 2D image features into their 3D counterparts, as our backbone design and uncover their decent capabilities in diverse 3D applications, which may have been overlooked by previous works. (2) Considering a wide range of 3D applications that require real-time rendering, it is highly desirable that our backbone design can be seamlessly converted into efficient 3D representations, particularly meshes, to enable real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our backbone design. As shown in Fig. 1, the backbone model in OmniRecon leverages image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.778, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.393, + 0.673, + 0.406 + ], + "angle": 0, + "content": "Fig. 1: An overview of our Omni-Recon framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.552 + ], + "angle": 0, + "content": "appearance features to accurately predict an SDF, and one lightweight appearance branch for estimating blending weights of source views. Specifically, the progressive feature fusion in the geometry branch empowers accurate geometry estimation for high-quality 3D reconstruction, and the decoupling between geometry and appearance enables real-time rendering after baking the complex geometry branch into meshes, with the lightweight appearance branch serving as a shader. Additionally, the nature of image-based rendering, when equipped with accurate geometry and appearance estimation, can enable a wide range of downstream 3D applications, as shown in Sec. 4 and Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.636, + 0.586 + ], + "angle": 0, + "content": "3.2 The Overall Image-Based Rendering Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.656 + ], + "angle": 0, + "content": "Following generalizable NeRFs [10,38,77,88], our NeRF backbone is conditioned on source views from the target scene. This is achieved by estimating the geometry and appearance based on the extracted features of the source views from the target scene, which is elaborated as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Acquire geometry features. Given \\(N\\) source views \\(\\{I_i\\}_{i=1}^N \\in \\mathbb{R}^{H \\times W \\times 3}\\) from the target scene, we utilize a CNN encoder [34, 55] to extract source features \\(\\{\\mathbf{F}_i\\}_{i=1}^N \\in \\mathbb{R}^{H \\times W \\times C}\\) from the source views. Next, we construct a 3D feature volume \\(V \\in \\mathbb{R}^{M \\times M \\times M \\times C}\\), which serves similar roles as cost volumes for multiview stereo [10, 83], to aggregate geometry information from different source views, following [44, 55, 63]. Specifically, we divide the bounding volume of the scene into \\(M^3\\) voxels and project each voxel center onto \\(N\\) source views, where the mean and variance of \\(N\\) source features are computed and concatenated as the corresponding voxel feature. We then leverage a 3D U-Net [14] to further enhance the feature volume to acquire the final geometry feature volume \\(V\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acquire appearance features. For each sampled 3D point \\(\\mathbf{p}(t) = \\mathbf{o} + t\\mathbf{d}\\) along the ray emitted from the target pixel on the novel view with camera origin \\(\\mathbf{o}\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "and direction \\(\\mathbf{d}\\), we project it on the source views to acquire the projected colors \\(\\{\\mathbf{c}_i\\}_{i=1}^N\\) and appearance features \\(\\{\\mathbf{f}_i\\}_{i=1}^N\\) using bilinear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Geometry and appearance estimation. For more accurate reconstruction of the scene geometry and mitigation of the intrinsic color-density ambiguity [76], we estimate the per-point SDF and model the density as a function of SDF, following [40,55,76]. For appearance, we estimate the per-point radiance through color blending, i.e., predicting the blending weights of the projected colors on the source views to compose the per-point color, following [38,77,88]. The geometry and appearance are modeled using separate branches and can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.29, + 0.786, + 0.309 + ], + "angle": 0, + "content": "\\[\ns = \\mathbf {M} _ {s d f} (\\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}, V), \\quad \\{\\omega_ {i} \\} _ {i = 1} ^ {N} = \\mathbf {M} _ {c o l o r} (\\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}, \\mathbf {d}), \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.359 + ], + "angle": 0, + "content": "where \\(s\\) is the SDF, \\(\\{\\omega_i\\}_{i=1}^N\\) are the blending weights, and \\(\\mathbf{M}_{sdf}\\) and \\(\\mathbf{M}_{color}\\) are the geometry and appearance branches, respectively. The per-point radiance can be derived as the weighted sum \\(\\hat{\\mathbf{c}} = \\sum_{i=1}^{N} \\omega_i \\mathbf{c}_i\\) across \\(N\\) source views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.784, + 0.39 + ], + "angle": 0, + "content": "For SDF-based volumetric rendering, we adopt the formulation in NeuS [76]. Specifically, the color is accumulated along the \\(K\\) sampled points along each ray:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.391, + 0.786, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} = \\sum_ {j = 1} ^ {K} T _ {j} \\alpha_ {j} \\hat {\\mathbf {c}} _ {j}, \\quad \\alpha_ {j} = 1 - \\exp \\left(- \\int_ {t _ {j}} ^ {t _ {j + 1}} \\rho (t) d t\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.582 + ], + "angle": 0, + "content": "where \\( T_{j} = \\prod_{n=1}^{j-1}(1 - \\alpha_{n}) \\) is the accumulated transmittance and \\( \\rho(t) \\) is the density determined by the estimated SDF, following the definition in NeuS [76]. Complexities of the two branches. When predicting the radiance of a sampled 3D point, the colors of valid projections on the source views are already close to the radiance of this 3D point. In light of this, we employ a lightweight \\( \\mathbf{M}_{color} \\), consisting of three MLP layers, for efficient blending weight prediction. In contrast, the geometry branch should be allocated with sufficient complexity for accurately estimating the occlusion effects among different sampled points along the same ray as well as that among the sampled points and the source views. We introduce the design of the geometry branch in the next subsection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.691, + 0.611 + ], + "angle": 0, + "content": "3.3 The Proposed Transformer-based Geometry Branch" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.722 + ], + "angle": 0, + "content": "Overview. Our geometry branch leverages transformer modules [71] to progressively fuse geometry and appearance features, thus properly handling the two occlusion effects mentioned above. Specifically, for the \\(K\\) sampled 3D points along the ray, we acquire their geometry features \\(\\{\\mathbf{v}_k\\}_{k=1}^K\\) from the geometry feature volume \\(V\\), which are sequentially processed through \\(B\\) blocks (\\(B = 2\\) in our design), each consisting of a geometry transformer, an appearance transformer, and an occlusion transformer, to produce the per-point SDF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.784 + ], + "angle": 0, + "content": "The geometry transformer. The geometry transformer fuses the geometry features into its inputs by using \\(\\{\\mathbf{v}_k\\}_{k=1}^K\\) as the key and value in a cross-attention scheme [71] and performs attention across the sampled points along the same ray to model their occlusion effects. This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.275, + 0.796, + 0.786, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {s d f} ^ {g e o} (\\mathbf {x}, \\{\\mathbf {v} _ {k} \\} _ {k = 1} ^ {K}) = C r o s s A t t e n t i o n (\\mathbf {q} = \\mathbf {x}, \\mathbf {k} = \\mathbf {v} = \\{\\mathbf {v} _ {k} \\} _ {k = 1} ^ {K}), \\qquad (3)\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is the input to the geometry transformer, which is exactly the geometry features themselves for the geometry transformer in the first block, and \\(\\mathbf{q}\\), \\(\\mathbf{k}\\), and \\(\\mathbf{v}\\) are the query, key, and value of the attention scheme [71], respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.343 + ], + "angle": 0, + "content": "The appearance transformer. The appearance transformer integrates the appearance features \\(\\{\\mathbf{f}_i\\}_{i=1}^N\\) into the outputs of the geometry transformer, considering the potential occlusions among the sampled points and the source views that may cause invalid projections. This integration is achieved using subtraction attention [75,92], which is more effective for geometric relationship reasoning [75,92], with \\(\\{\\mathbf{f}_i\\}_{i=1}^N\\) as the key and value. Specifically, subtraction attention computes the attention scores between the input query features and each source view in the key features, which are then used to aggregate different source views in the value features into new output features with the same dimension as the input query features. This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.351, + 0.785, + 0.371 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {s d f} ^ {a p p r} (\\mathbf {x}, \\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}) = S u b A t t e n t i o n (\\mathbf {q} = \\mathbf {x}, \\mathbf {k} = \\mathbf {v} = \\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.408 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is the output of \\(\\mathbf{M}_{sdf}^{geo}\\). The detailed formulation of subtraction attention is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.784, + 0.453 + ], + "angle": 0, + "content": "The occlusion transformer. We further integrate an occlusion transformer that performs self-attention across sampled points along the same ray to model their occlusion effects explicitly:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.463, + 0.785, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {s d f} ^ {\\text {o c c}} (\\mathbf {x}) = \\text {S e l f A t t e n t i o n} (\\mathbf {q} = \\mathbf {k} = \\mathbf {v} = \\mathbf {x}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.784, + 0.52 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is the output of \\(\\mathbf{M}_{sdf}^{app}\\). We apply another MLP on top of the output features of the last \\(M_{sdf}^{occ}\\) module to predict the per-point SDF \\(s\\) in Eq. 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.424, + 0.547 + ], + "angle": 0, + "content": "3.4 Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.784, + 0.614 + ], + "angle": 0, + "content": "Our NeRF backbone is trained using two objectives: (1) a photometric loss \\(\\mathcal{L}_{rgb}\\) between the rendered images and the ground-truth images, and (2) a depth loss \\(\\mathcal{L}_{depth}\\) between the rendered depth \\(\\hat{\\mathbf{D}} = \\sum_{j=1}^{K} T_j \\alpha_j t\\) and the ground-truth depth. The overall objective can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.625, + 0.785, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\beta \\mathcal {L} _ {\\text {d e p t h}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.784, + 0.702 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{rgb} = \\frac{1}{R}\\sum_{\\mathbf{r} = 1}^{R}\\left\\| \\hat{\\mathbf{C}}_{\\mathbf{r}} - \\mathbf{C}_{\\mathbf{r}}\\right\\|_{2}^{2}\\) with \\(R\\) denoting the number of rendered pixels, \\(\\mathcal{L}_{depth} = \\frac{1}{R}\\sum_{\\mathbf{r} = 1}^{R}\\left\\| \\hat{\\mathbf{D}}_{\\mathbf{r}} - \\mathbf{D}_{\\mathbf{r}}\\right\\|_{2}^{2}\\), and \\(\\beta\\) is set as 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.713, + 0.634, + 0.727 + ], + "angle": 0, + "content": "3.5 Evaluation: Generalizable 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Setup. We assess the backbone of Omni-Recon by evaluating its generalizable 3D reconstruction capability. Datasets: Following [10, 34, 40, 55], we adopt the DTU dataset [1], which comprises 124 different scenes and 7 distinct lighting conditions. For testing, 15 scenes are utilized, while the remaining scenes are allocated for training, following the dataset split in [34, 40, 55]. Training settings: During training, we employ \\( N = 4 \\) source views with a resolution of \\( 640 \\times 512 \\). The ray number per batch and batch size are set to 1024 and 2, respectively," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.186 + ], + "angle": 0, + "content": "Table 1: Benchmark the quantitative performance in sparse view mesh reconstruction in terms of Chamfer distance (the lower, the better) on 15 testing scenes on DTU. The best scores are in bold, and the second-best are underlined." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.188, + 0.777, + 0.359 + ], + "angle": 0, + "content": "
MethodMean243740556365698397105106110114118122
COLMAP [57]1.520.902.891.631.082.181.941.611.302.341.281.101.420.761.171.14
MVSNet [83]1.221.052.521.711.041.451.520.881.291.381.050.910.660.611.081.16
IDR [86]3.394.016.403.521.913.962.364.851.626.375.971.234.730.911.721.26
VolSDF [84]3.414.034.216.120.918.241.732.741.825.143.092.084.810.603.512.18
UNISURF [48]4.395.087.183.965.304.612.243.943.145.633.405.096.382.984.052.81
NeuS [76]4.004.574.493.974.324.631.954.683.834.152.501.526.471.265.576.11
PixelNeRF [88]6.185.138.075.854.407.114.645.686.769.056.113.955.926.266.896.93
IBRNet [77]2.322.293.702.661.833.022.831.772.282.731.961.872.131.582.052.09
MVSNeRF [10]2.091.963.272.541.932.572.711.821.722.291.751.721.471.292.092.26
SparseNeuS [40]1.962.173.292.741.672.692.421.581.861.941.351.501.450.981.861.87
VolRecon [55]1.381.202.591.561.081.431.921.111.481.421.051.191.380.741.231.27
ReTR [34]1.171.052.311.440.981.181.520.881.351.300.871.070.770.591.051.12
Omni-Recon (Ours)1.130.912.131.520.931.091.700.841.291.200.831.040.810.551.051.05
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.412 + ], + "angle": 0, + "content": "following [34,55]. Mesh extraction: We utilize TSDF fusion [45,63] to reconstruct scene meshes from predicted source view depths, following [34,55]. More detailed model/training settings can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.473 + ], + "angle": 0, + "content": "Baselines. We primarily benchmark against SOTA generalizable neural implicit reconstruction methods [16, 34, 55], along with generalizable neural rendering methods [10, 77, 88], per-scene optimization methods [48, 76, 84, 86], and multiview stereo methods [57, 83]. We directly report their official results in [34, 55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Sparse view reconstruction. We benchmark the performance of sparse view reconstruction using three views from each testing scene in DTU, following the setting in [16, 34, 55]. As shown in Tab. 1, we can observe that (1) Our Omni-Recon can achieve new SOTA reconstruction performance, with the lowest Chamfer distance averaged over all scenes. Additionally, as visualized in Fig. 2 (Row 1), our reconstructed mesh can exhibit smooth surfaces and correctly maintain the structure, e.g., the hole in the mouth. (2) Our method achieves the best reconstruction performance in 10 out of 15 scenes, indicating its general effectiveness across diverse scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Generalization to more diverse scenes. To demonstrate the generalization capability on more diverse scenes with larger domain shifts compared to the training scenes, we further benchmark the mesh reconstruction quality on the NeRF-Synthetic dataset [41] and NSVF-Synthetic [37] under a full-view reconstruction setting, i.e., the depth of each source view from each scene's training set is estimated and fused into a mesh using TSDF fusion [45, 63]. As visualized in Fig. 2 (Rows 2-4), we observe that our Omni-Recon can deliver notably higher-quality meshes with better-maintained structures, fewer holes, and smoother surfaces compared to the strongest baselines ReTR [34] and VolRecon [55]. The decent generalization capability of our method is attributed to the proper decoupling and fusion of geometry and appearance features in our backbone. More visualizations and ablation studies, such as the contributions of each backbone component, are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Rendering quality. We further benchmark the performance of generalizable novel view rendering on the test scenes from DTU. As shown in Tab. 2, our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.761, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Fig. 2: Visualize the reconstructed mesh of our Omni-Recon and the two strongest baselines [34, 55]. Row 1: Scan65 from the test scenes of DTU; Rows 2-4: Scenes from NeRF-Synthetic [41] and NSVF-Synthetic [37], which present relatively challenging cases due to domain shifts w.r.t. the training scenes from DTU." + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.516, + 0.784, + 0.529 + ], + "angle": 0, + "content": "Table 2: Benchmark the quality of novel view rendering in terms of PSNR on DTU." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.53, + 0.778, + 0.581 + ], + "angle": 0, + "content": "
MethodMean.243740556365698397105106110114118122
VolRecon [55]24.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]25.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Omni-Recon26.3224.7722.3323.9225.5626.3724.7525.1924.9424.9225.0628.3928.6328.0131.4930.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.639 + ], + "angle": 0, + "content": "method achieves the highest PSNR, e.g., a \\(+0.73\\) and \\(+1.74\\) PSNR improvement over ReTR and VolRecon, respectively. As will be demonstrated in Sec. 5.4, this PSNR can be notably boosted with super-fast Nvdiffrast-based finetuning [31]." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.653, + 0.764, + 0.672 + ], + "angle": 0, + "content": "4 Omni-Recon: Enable Zero-shot Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.77 + ], + "angle": 0, + "content": "To enable the understanding of various 3D scene properties, one approach is to jointly train the NeRF backbone on all target tasks. However, this approach is not scalable due to the diversity of scene properties, such as semantic segmentation, edge detection, and keypoint prediction, making it costly to cover new properties of interest. Therefore, it is highly desirable to enable zero-shot scene understanding for new tasks that were not seen during training." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.775, + 0.655, + 0.79 + ], + "angle": 0, + "content": "4.1 Our Strategy: The Predict-then-Blend Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our hypothesis. We hypothesize that with accurate geometry and appearance estimation, the blending weights initially learned for radiance can be repurposed for other scene understanding tasks. In essence, just as blending source colors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.775, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.512, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Fig. 3: Benchmark our predict-then-blend strategy with the render-then-predict baseline across different scene understanding tasks on Replica [61] and ScanNet [17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.635 + ], + "angle": 0, + "content": "yields radiance for each sampled 3D point, blending the properties of source views using the same weights can yield other properties. This hypothesis stems from the observation that scene appearance is closely related to other scene properties and thus regions with similar appearance are likely to share similar scene properties, such as semantics and edges. As such, the blending weights learned for appearance can also indicate the weighting for other scene properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Our method. Motivated by this hypothesis, our Omni-Recon employs a predict-then-blend strategy to enable zero-shot multitask scene understanding by lifting 2D monocular priors to 3D novel views. Specifically, considering the prevalence of 2D monocular vision models, our method first utilizes pretrained 2D models to generate predictions (e.g., logits in semantic segmentation) on each source view, denoted as \\(\\{\\mathbf{P}_i\\}_{i=1}^N\\). Then, the property of each sampled point can be obtained by blending their projected predictions on source views \\(\\{\\mathbf{p}_i\\}_{i=1}^N\\), reusing the RGB blending weights, i.e., \\(\\hat{\\mathbf{p}} = \\sum_{i=1}^{N} \\omega_i \\mathbf{p}_i\\). The final pixel-wise prediction can be derived using volumetric rendering similar to RGB reconstruction in Eq. 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.707, + 0.804 + ], + "angle": 0, + "content": "4.2 Evaluation: Zero-shot Multitask Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Setup. To evaluate the zero-shot scene understanding capabilities of our model, we adopt two indoor 3D scene datasets: Replica [61] and ScanNet [17], following" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.223 + ], + "angle": 0, + "content": "the dataset splits in [93] and [36], respectively. Specifically, we focus on four scene understanding tasks: semantic segmentation, edge detection, and two keypoint detection tasks defined in [89]. For these tasks, the 2D monocular priors are provided by [51], [5], and [56], respectively. We employ mIoU for the semantic segmentation task and \\(\\ell_1\\) error for the other three tasks as default metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.787, + 0.282 + ], + "angle": 0, + "content": "Baseline. An intuitive way to achieve zero-shot 3D scene understanding is to leverage existing 2D monocular models to predict the scene properties on top of the rendered images, a process referred to as a render-then-predict pipeline. We benchmark this strategy against our proposed predict-then-blend strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.487, + 0.418 + ], + "angle": 0, + "content": "Zero-shot scene understanding benchmark. As shown in Tab. 3 and Fig. 3, we observe that (1) Our predict-then-blend strategy achieves decent scene understanding performance on unseen scenes and unseen tasks, which echoes our hypothesis in Sec. 4.1 that 2D monocular priors can be elevated to 3D novel views through" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.283, + 0.786, + 0.31 + ], + "angle": 0, + "content": "Table 3: Benchmark the two zero-shot scene understanding strategies across tasks." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.311, + 0.785, + 0.416 + ], + "angle": 0, + "content": "
DatasetMethod↑ Semantic↓ Edge↓ Key Point↓ Key Point 3D
Replica [61]Render-then -Predict15.640.04560.11010.0470
Predict-then -Blend (Ours)32.110.04120.07740.0176
ScanNet [17]Render-then -Predict41.320.04710.05680.0412
Predict-then -Blend (Ours)61.110.04340.04240.0197
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.586 + ], + "angle": 0, + "content": "blending weight reuse, thanks to the high correlation between scene appearance and other properties; (2) Compared to the render-then-predict scheme, our method achieves notably higher quantitative accuracy, e.g., a \\(19.79\\%\\) higher mIoU on ScanNet, and consistently better visual effects across all tasks. We attribute this improvement to two aspects: (1) Our predict-then-blend strategy can avoid feeding noisy inputs caused by rendering errors to the 2D monocular models, thus reducing reliance on the robustness of monocular models and ensuring more accurate monocular priors; (2) The multiview information gathered by our strategy can contribute to a more accurate understanding of objects with limited observations under a monocular setting. More experiments on other vision models like CLIP-LSeg [32] are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.757, + 0.61 + ], + "angle": 0, + "content": "5 Omni-Recon: Adaptability to Downstream 3D Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.787, + 0.693 + ], + "angle": 0, + "content": "In addition to the generalization capabilities, we further demonstrate the adaptability of our Omni-Recon by extending it to three different types of downstream 3D applications. These demonstrations highlight the potentially broad utility of image-based rendering pipelines, which may have been overlooked by previous works, and can be generalized to a wider range of real-world 3D applications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.699, + 0.694, + 0.715 + ], + "angle": 0, + "content": "5.1 Application 1: A New Real-time Rendering Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.781 + ], + "angle": 0, + "content": "To rapidly enable real-time rendering using Omni-Recon's pretrained backbone, we build a real-time rendering pipeline based on Nvdiffrast [31]. This involves baking the complex geometry branch into meshes using TSDF fusion [45,63] as discussed in Sec. 3.5, and using the lightweight appearance branch as a Shader." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our real-time rendering pipeline. As illustrated in Fig. 1 (bottom left), for any novel view of a new scene, we first transform its camera pose to clip space as required by Nvdiffrast [31]. Then, we perform a rasterization process to obtain the intersection point of each camera ray with the extracted mesh. Next, these" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "intersection points are input into the shader to predict the color by blending their projected colors on the source views as described in Sec. 3.2. Thanks to the blending scheme, where valid projections can provide strong appearance clues, the rendering quality is often satisfactory even without finetuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.313 + ], + "angle": 0, + "content": "Joint mesh and Shader finetuning. To further enhance the rendering quality, we can jointly finetune the mesh and the Shader in a gradient-based manner using differentiable rendering enabled by Nvdiffrast [31]. Specifically, we finetune the location of each vertex in the mesh and the parameters of the lightweight Shader. Additionally, we periodically prune redundant mesh faces that cannot intersect with any camera rays in one training epoch. More details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.323, + 0.784, + 0.339 + ], + "angle": 0, + "content": "5.2 Application 2: Parameter-Efficient Tuning for 3D Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.785, + 0.48 + ], + "angle": 0, + "content": "We explore the possibility of finetuning Omni-Recon for various downstream 3D tasks, utilizing a series of 3D scene understanding tasks as demonstrations. Specifically, we conduct PET [19] on top of Omni-Recon's pretrained backbone by incorporating low-rank adapters [26] into the linear layers within the transformer modules introduced in Sec. 3.3, and only tuning these adapters while keeping the pretrained backbone frozen. It is worth noting that when evaluating the performance of the tuned model, we assess it on new scenes that have never been encountered during both training and tuning, representing a generalizable scene understanding scenario [36]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.489, + 0.633, + 0.505 + ], + "angle": 0, + "content": "5.3 Application 3: Text-guided 3D Scene Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.785, + 0.66 + ], + "angle": 0, + "content": "We target text-guided 3D scene editing, which involves editing a 3D scene based on textual instructions. To achieve this, our Omni-Recon capitalizes on the nature of image-based rendering, allowing edits made to 2D source views to be propagated to 3D novel views. Specifically, leveraging the availability of powerful 2D diffusion models, we use publicly available ones to edit the 2D source views. However, doing so naively may result in inconsistent source views, as the same textual instruction may lead to different visualizations, thus degrading reconstruction quality. Inspired by [23], which iteratively edits the training images and retransmits a new NeRF to achieve scene editing, we address the inconsistent editing issue by proposing an iterative editing and reconstruction pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.785, + 0.841 + ], + "angle": 0, + "content": "The iterative editing and reconstruction pipeline. To ensure consistent 3D editing, we iterate between source view editing and reconstruction as follows: First, we edit each source view using the diffusion model InstructPix2Pix [4], which is guided by both textual instructions and the original image to maintain image fidelity. While the edited source views can faithfully follow the textual instructions, they may exhibit inconsistencies in the edited region. Next, we reconstruct each source view by conditioning on their nearby source views, as described in Sec. 3.2. This process improves the 3D consistency among source views at the expense of reduced image fidelity and less adherence to instructions. We then iterate through the editing and reconstruction steps to progressively enhance image fidelity, 3D consistency, and adherence to instructions. The resulting source views are finally used to reconstruct the entire scene." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.139, + 0.788, + 0.182 + ], + "angle": 0, + "content": "Table 4: Benchmark the rendering speed and rendering quality of our real-time rendering pipeline and the baselines [34,55] on DTU. The best PSNR is in bold, and the shortest finetuning time that our pipelines surpass the strongest baseline is underlined." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.189, + 0.784, + 0.293 + ], + "angle": 0, + "content": "
MethodFPSMean243740556365698397105106110114118122
VolRecon [55]0.02924.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]0.02425.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Ours w/o ft.71.3(40.82)22.9620.1219.7122.2722.7824.5521.7721.5126.7222.3324.4922.5222.9324.6823.8424.12
Ours (ft. 10s)25.6821.4221.6824.0624.1228.1924.1023.9531.6524.4128.1525.6325.8526.1526.8226.89
Ours (ft. 20s)27.2122.6322.9225.1225.4230.0325.9526.1633.1926.2230.3627.4427.0426.9329.1529.65
Ours (ft. 30s)27.7823.223.2625.5425.730.5926.8326.9633.6626.4730.7328.1427.7027.130.1730.65
Ours (ft. 1min)28.3424.6924.1125.7626.0530.9327.6627.4933.6827.0730.9728.5128.5427.3531.1731.54
Ours (ft. 3min)28.9525.224.3225.9426.1632.1528.9927.8834.9427.3531.6228.9328.9727.5631.7032.49
Ours (ft. 5min)29.0225.3424.3625.6326.2132.1629.3327.8134.9427.3231.7429.0429.0527.6931.7432.89
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.306, + 0.762, + 0.32 + ], + "angle": 0, + "content": "5.4 Evaluation: Effectiveness on Different Downstream 3D Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.401 + ], + "angle": 0, + "content": "Real-time rendering. We evaluate our real-time rendering pipeline in Sec. 5.1 in terms of both rendering quality and rendering speed, measured on an NVIDIA A5000 GPU, using the DTU dataset. Specifically, we finetune the scene meshes, extracted by TSDF fusion as mentioned in Sec. 3.5, and the Shader in no more than 5 minutes and record the rendering quality at different time steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.599 + ], + "angle": 0, + "content": "Observations and analysis: As shown in Tab. 4, we can observe that (1) Without any finetuning, the rendering pipeline already exhibits certain rendering capabilities, despite the model backbone being trained using volumetric rendering with more than one sampled point; (2) Our rendering pipeline can be swiftly finetuned to significantly boost the PSNR. For instance, with just a 10-second finetuning, our rendering pipeline can match the average rendering quality of the strongest baseline, ReTR [34]; With 1-minute/2-minute finetuning, it can surpass ReTR by a +2.75 and +3.36 PSNR improvement on average, respectively; (3) Our rendering pipeline enables real-time rendering with 71.3 FPS, i.e., a >2458× speed-up over the baselines. Here, the assumption is that the feature extraction of each source view from a target scene can be done in a one-time effort and does not account for throughput. If we still consider this feature extraction latency for each rendering, the resulting 40.82 FPS still meets the real-time requirement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.79, + 0.629 + ], + "angle": 0, + "content": "These experiments indicate that given a new scene, our Omni-Recon is promising to enable both instant and accurate reconstruction and real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.63, + 0.46, + 0.749 + ], + "angle": 0, + "content": "PET for 3D Understanding. To evaluate the adaptability of our Omni-Recon, we apply PET to it for various scene understanding tasks on ScanNet [17]. Specifically, we finetune a LoRA adaptor [26] for each task on a set of training scenes and evaluate on" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.631, + 0.787, + 0.659 + ], + "angle": 0, + "content": "Table 5: Evaluate the effectiveness of PET on Omni-Recon and benchmark with SRay [36]." + }, + { + "type": "table", + "bbox": [ + 0.474, + 0.66, + 0.784, + 0.741 + ], + "angle": 0, + "content": "
Method↑ Sem. mIoU↑ Sem. Total Acc↑ Sem. Avg Acc↓ Edge↓ Key Point↓ Key Point 3D
SRay [36]57.1578.2462.55---
Ours (Zero-shot)61.1180.8069.170.04340.04240.0197
Ours (PET + Zero-shot)62.3581.8469.890.03420.02070.0132
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.795 + ], + "angle": 0, + "content": "non-overlapping test scenes, representing a generalizable 3D understanding setting, following the dataset split described in [36]. We benchmark against the SOTA generalizable semantic segmentation work, SRay [36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Observations and analysis: As shown in Tab. 5, we can observe that (1) After PET, our model outperforms SRay in terms of generalizable semantic segmentation with a \\(+5.20\\%\\) mIoU and a \\(+7.34\\%\\) accuracy improvement; (2) Despite" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.72, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.348, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Fig. 4: Visualize the scene understanding performance on hard cases before/after PET." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.366, + 0.756, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.524, + 0.765, + 0.538 + ], + "angle": 0, + "content": "Fig. 5: Three text-guided scene editing examples using our pipeline in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.787, + 0.588 + ], + "angle": 0, + "content": "having limited trainable parameters, PET can enhance the scene understanding performance across tasks, especially on challenging cases visualized in Fig. 4, indicating the adaptability of Omni-Recon to diverse downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.787, + 0.62 + ], + "angle": 0, + "content": "Text-guided scene editing. We visualize the edited examples in Fig. 5, showcasing the ability to ensure both instruction-following and 3D consistency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.36, + 0.65 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Motivated by the significant demand for supporting various functionalities required by different 3D applications through a unified NeRF model, our work aims to develop a general-purpose NeRF capable of handling a broad spectrum of 3D tasks. We propose a framework called Omni-Recon, which features a general-purpose NeRF model using image-based rendering with two separate branches, and demonstrate its integration with various 3D tasks. Specifically, our Omni-Recon can achieve SOTA generalizable 3D reconstruction quality, enable zero-shot multitask scene understanding, achieve SOTA scene understanding performance and real-time rendering after rapid adaptation, and support scene editing. Omni-Recon underscores the potential of image-based rendering pipelines in diverse real-world 3D applications, which may have been underestimated by prior studies, thus sparking advancements in future rendering pipelines." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.395, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.237 + ], + "angle": 0, + "content": "The work is supported by the National Science Foundation (NSF) through the SCH program (Award number: 1838873) and CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. International Journal of Computer Vision 120, 153-168 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.386 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.785, + 0.442 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.483 + ], + "angle": 0, + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "5. Canny, J.: A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence (6), 679-698 (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.538 + ], + "angle": 0, + "content": "6. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.579 + ], + "angle": 0, + "content": "7. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.606 + ], + "angle": 0, + "content": "8. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.634 + ], + "angle": 0, + "content": "9. Chen, A., Xu, Z., Wei, X., Tang, S., Su, H., Geiger, A.: Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.635, + 0.785, + 0.689 + ], + "angle": 0, + "content": "10. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.689, + 0.785, + 0.744 + ], + "angle": 0, + "content": "11. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "12. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.217 + ], + "angle": 0, + "content": "14. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "15. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3193-3204 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "16. Croce, F., Hein, M.: Sparse and imperceivable adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4724-4732 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "17. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "18. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "19. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence 5(3), 220-235 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "20. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "21. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nef: 3d-to-2d label transfer for panoptic urban scene segmentation. arXiv preprint arXiv:2203.15224 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.564 + ], + "angle": 0, + "content": "22. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.592 + ], + "angle": 0, + "content": "23. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "24. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "25. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "27. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "28. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "30. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "31. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "32. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "33. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "34. Liang, Y., He, H., Chen, Y.: Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.384 + ], + "angle": 0, + "content": "35. Lindell, D.B., Martel, J.N., Wetzstein, G.: Autoint: Automatic integration for fast neural rendering. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.787, + 0.425 + ], + "angle": 0, + "content": "36. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.786, + 0.452 + ], + "angle": 0, + "content": "37. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "39. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. SIGGRAPH (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.787, + 0.591 + ], + "angle": 0, + "content": "40. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.674 + ], + "angle": 0, + "content": "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.787, + 0.757 + ], + "angle": 0, + "content": "44. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16. pp. 414-431. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "45. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "47. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "48. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "49. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "51. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. ArXiv preprint (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "52. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: DeRF: Decomposed radiance fields. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "53. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.48 + ], + "angle": 0, + "content": "54. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "55. Ren, Y., Zhang, T., Pollefeys, M., Süsstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "56. Sax, A., Emi, B., Zamir, A.R., Guibas, L.J., Savarese, S., Malik, J.: Mid-level visual representations improve generalization and sample efficiency for learning visuomotor policies. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "57. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3D-aware image synthesis. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. arXiv preprint arXiv:2206.07695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "60. Seitz, S.M., Dyer, C.R.: Photorealistic scene reconstruction by voxel coloring. IJCV (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "61. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "62. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni-Recon" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "64. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "65. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "66. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "67. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "68. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15182-15192 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "69. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. arXiv preprint arXiv:2211.11738 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "70. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "71. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "72. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "73. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "74. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "75. Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z., et al.: Is attention all nerf needs? arXiv preprint arXiv:2207.13298 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "76. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.66, + 0.787, + 0.716 + ], + "angle": 0, + "content": "77. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4690-4699 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.716, + 0.787, + 0.744 + ], + "angle": 0, + "content": "78. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "79. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 106-122. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "80. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Y. Fu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "81. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 \\(\\{\\backslash \\mathrm{deg}\\}\\) views. arXiv preprint arXiv:2211.16431 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "82. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "83. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "84. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "85. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.384 + ], + "angle": 0, + "content": "86. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "87. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "88. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "89. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.536 + ], + "angle": 0, + "content": "90. Zhang, M., Zheng, S., Bao, Z., Hebert, M., Wang, Y.X.: Beyond rgb: Scene-property synthesis with neural radiance fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 795-805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.578 + ], + "angle": 0, + "content": "91. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5449-5458 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "92. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "93. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.66 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_origin.pdf b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a4b5007a14d0408d3d95f62fe0cbaa385b9d970 --- /dev/null +++ b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/0d53a505-587b-4461-974a-55e35a489a73_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd3778f8ec8ff72cf22ddcfd4e26eebe8a2be30a7ed255a2d0b066fb655f0294 +size 20958177 diff --git a/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/full.md b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7315eea21208652cd3217c70042cd493f8a871cf --- /dev/null +++ b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/full.md @@ -0,0 +1,339 @@ +# Omni-Recon: Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields + +Yonggan Fu, Huaizhi Qu, Zhifan Ye, Chaojian Li, Kevin Zhao, and Yingyan (Celine) Lin + +Georgia Institute of Technology {yfu314, zye327, cli851, kzhao14, celine.lin}@gatech.edu + +Abstract. Recent breakthroughs in Neural Radiance Fields (NeRFs) have sparked significant demand for their integration into real-world 3D applications. However, the varied functionalities required by different 3D applications often necessitate diverse NeRF models with various pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. Drawing inspiration from the generalization capability and adaptability of emerging foundation models, our work aims to develop one general-purpose NeRF for handling diverse 3D tasks. We achieve this by proposing a framework called Omni-Recon, which is capable of (1) generalizable 3D reconstruction and zero-shot multitask scene understanding, and (2) adaptability to diverse downstream 3D applications such as real-time rendering and scene editing. Our key insight is that an image-based rendering pipeline, with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our Omni-Recon features a general-purpose NeRF model using image-based rendering with two decoupled branches: one complex transformer-based branch that progressively fuses geometry and appearance features for accurate geometry estimation, and one lightweight branch for predicting blending weights of source views. This design achieves state-of-the-art (SOTA) generalizable 3D surface reconstruction quality with blending weights reusable across diverse tasks for zero-shot multitask scene understanding. In addition, it can enable real-time rendering after baking the complex geometry branch into meshes, swift adaptation to achieve SOTA generalizable 3D understanding performance, and seamless integration with 2D diffusion models for text-guided 3D editing. Our code is available at: https://github.com/GATECH-EIC/Omni-Recon. + +# 1 Introduction + +Neural Radiance Fields (NeRFs) [41] have garnered significant attention and hold promise to become crucial enablers for emerging 3D applications. However, deploying NeRFs in real-world scenarios still requires substantial effort, as different 3D applications often demand distinct NeRF models with varied + +pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. For instance, while both cross-scene generalization and real-time rendering are desirable in the real world, generalizable NeRFs for instant reconstruction [10,38,77] and real-time NeRFs using mesh-based rasterization [12,65,85] typically feature diverse pipelines, making it hard to simultaneously satisfy the two requirements. Additionally, understanding new scene properties [36,90,93] necessitates training new NeRF models, which is not scalable given the increasing number of scene properties of interest. + +In parallel, the emergence of foundation models has driven transformative progress in real-world artificial intelligence applications. Inspired by this exciting trend, there has been a growing interest in integrating NeRFs into this realm by developing general-purpose NeRFs for handling various 3D tasks. Specifically, similar to the generalization capability and adaptability of foundation models, we expect general-purpose NeRFs to possess three essential capabilities, including (1) generalizable 3D reconstruction for instant surface reconstruction and novel view synthesis, (2) zero-shot multitask scene understanding of various scene properties, e.g., semantics and edges, and (3) adaptability to diverse downstream tasks such as real-time rendering and scene editing. Achieving these capabilities is promising to enable users to effortlessly start from these general-purpose NeRFs when facing a new 3D application, significantly reducing the need for training and experimenting with various NeRF pipelines. + +Despite the promise, existing NeRF pipelines have not yet realized the mentioned capabilities. Firstly, per-scene optimized NeRFs [2, 3, 22, 25, 27, 42, 85] rely on costly per-scene training, thus precluding generalization capabilities. Secondly, although generalizable NeRF designs [10, 38, 54, 68, 75, 77, 82, 88] can achieve cross-scene generalization, their huge computational complexity limits their suitability for application scenarios where real-time rendering is crucial. Some generalizable surface reconstruction methods [34, 55] can support mesh extraction, but an expensive volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views, leaving the demand for real-time rendering unmet. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored. + +Given the significant potential and associated challenges, our work aims to advance the integration of NeRFs into the realm of foundation models by developing general-purpose NeRFs. To achieve this goal, we propose a framework called Omni-Recon, which is an image-based rendering pipeline that realizes all the mentioned capabilities. The key insight leveraged by our Omni-Recon is that an image-based rendering pipeline, when equipped with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our contributions are summarized as follows: + +- Our Omni-Recon features a general-purpose NeRF model backbone using image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and appearance features to accurately predict a signed distance function (SDF), and one lightweight appearance branch for predicting blending weights of source views from the target scene. The advantages of this design include: (1) ensuring accurate geometry estimation for high-quality 3D reconstruction, and (2) enabling real-time rendering after baking the complex geometry branch into meshes and using the lightweight appearance branch as a Shader. +- We intriguingly find that a properly trained image-based rendering pipeline can directly lift 2D monocular priors to 3D novel views by reusing the blending weights predicted by the appearance branch and leverage this to enable generalizable, zero-shot, and multitask scene understanding. +- We demonstrate the adaptability of Omni-Recon by extending it to diverse downstream tasks, including: (1) a Nvdiffrast-based [31] real-time rendering pipeline using the extracted mesh, (2) parameter-efficient tuning (PET) of Omni-Recon on downstream generalizable scene understanding tasks, and (3) a new text-guided scene editing scheme by iteratively editing and reconstructing the source images from a 3D scene using 2D diffusion models. +- Extensive experiments demonstrate that our Omni-Recon can achieve state-of-the-art (SOTA) generalizable 3D surface reconstruction and scene understanding performance. Additionally, it can support real-time rendering of a new scene after rapid adaptation, along with easy-to-use 3D scene editing. + +Our Omni-Recon underscores the potentially wide usage of image-based rendering pipelines in diverse real-world 3D applications, which may have been overlooked by previous works. We believe that the insights we provide could ignite future innovations in more advanced generalizable rendering pipelines. + +# 2 Related Work + +Neural Radiance Fields. NeRFs [42], which represent the target scene as a continuous volume with density and view-dependent color, have garnered increasing attention thanks to their SOTA rendering quality when compared to prior approaches using explicit neural representations [24,39,60,66]. Subsequent works further improve NeRFs from various angles, such as enhancing their rendering quality [2,3,11,72], training NeRFs with sparse views [18,28,46,69,80,81], extending NeRFs to large-scale scenes [64,70,78,79,91], or applying NeRFs to other tasks, such as generative modeling [7,47,58,74], dynamic scenes [6,33,49, 50], or scene understanding [21,30,73,93]. + +Generalizable NeRFs. Vanilla NeRFs rely on per-scene optimization, which limits their ability to perform instant reconstruction. To address this limitation, generalizable NeRF designs [10, 13, 29, 38, 54, 68, 75, 77, 82, 88] achieve cross-scene generalization by conditioning vanilla NeRFs on the source views from the target new scene. Specifically, [68, 75, 77, 88] take the extracted scene features from the source views as inputs to reconstruct the radiance field via a one-shot forward + +pass. Subsequent works further improve cross-scene generalization capabilities by enhancing scene geometry estimation [10, 29, 38] and by introducing attention mechanisms and mixture-of-experts into NeRF backbone design [15, 36, 75]. Despite the promise of generalizable NeRFs, they still fall short in terms of rendering efficiency due to their huge computational complexity, which entails costly inference for each sampled point along the ray. While some generalizable surface reconstruction methods [34, 55] can support mesh extraction, a costly volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views. This poses a hindrance to their deployment in 3D applications that require real-time rendering. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored. + +Efficient scene representations. Various scene representations have been proposed to accelerate NeRFs' training and inference process. To speed up NeRF training, [8,9,20,43,62] integrate explicit neural representations with volumetric rendering. In parallel, to speed up NeRF inference/rendering, [35,52,53] attempt to reduce the complexity of the multi-layer perceptron (MLP) model in NeRF. In addition, [22, 25, 27, 37, 53, 59, 87] explore the free space in a target 3D scene to improve the sampling efficiency and thus reduce unnecessary MLP inference. Motivated by the insight of deferred rendering [67], [12, 25, 65, 85] pre-compute scene features offline and thus require less computation at run-time. In particular, [12, 65, 85] represents NeRF as a set of textured polygons, where the textures can be pre-computed and thus the NeRF pipeline can be efficiently executed in modern graphics hardware. However, these efficient scene representations still necessitate costly per-scene optimization, lacking generalization capabilities. + +# 3 Omni-Recon: The General-Purpose NeRF Backbone + +To develop general-purpose NeRFs for diverse 3D tasks, in this section, we design a general-purpose NeRF backbone and demonstrate its SOTA generalizable 3D reconstruction capability. We will leverage this backbone to enable zero-shot multitask scene understanding in Sec. 4 and adapt it to diverse downstream 3D tasks, including real-time rendering, PET for 3D scene understanding, and text-guided scene editing in Sec. 5. + +# 3.1 Backbone Design Overview + +Design principles. Our design is driven by the following principles: (1) Similar to foundation models, the ability to instantly reconstruct a new scene with generalizable capabilities is essential. Therefore, we leverage image-based rendering [10,38,77], which lifts 2D image features into their 3D counterparts, as our backbone design and uncover their decent capabilities in diverse 3D applications, which may have been overlooked by previous works. (2) Considering a wide range of 3D applications that require real-time rendering, it is highly desirable that our backbone design can be seamlessly converted into efficient 3D representations, particularly meshes, to enable real-time rendering. + +Our backbone design. As shown in Fig. 1, the backbone model in OmniRecon leverages image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and + +![](images/5556299a659e0430e950b9d7ef37408a3da658d642f050271fb4e782aae15406.jpg) +Fig. 1: An overview of our Omni-Recon framework. + +appearance features to accurately predict an SDF, and one lightweight appearance branch for estimating blending weights of source views. Specifically, the progressive feature fusion in the geometry branch empowers accurate geometry estimation for high-quality 3D reconstruction, and the decoupling between geometry and appearance enables real-time rendering after baking the complex geometry branch into meshes, with the lightweight appearance branch serving as a shader. Additionally, the nature of image-based rendering, when equipped with accurate geometry and appearance estimation, can enable a wide range of downstream 3D applications, as shown in Sec. 4 and Sec. 5. + +# 3.2 The Overall Image-Based Rendering Pipeline + +Following generalizable NeRFs [10,38,77,88], our NeRF backbone is conditioned on source views from the target scene. This is achieved by estimating the geometry and appearance based on the extracted features of the source views from the target scene, which is elaborated as follows. + +Acquire geometry features. Given $N$ source views $\{I_i\}_{i=1}^N \in \mathbb{R}^{H \times W \times 3}$ from the target scene, we utilize a CNN encoder [34, 55] to extract source features $\{\mathbf{F}_i\}_{i=1}^N \in \mathbb{R}^{H \times W \times C}$ from the source views. Next, we construct a 3D feature volume $V \in \mathbb{R}^{M \times M \times M \times C}$ , which serves similar roles as cost volumes for multiview stereo [10, 83], to aggregate geometry information from different source views, following [44, 55, 63]. Specifically, we divide the bounding volume of the scene into $M^3$ voxels and project each voxel center onto $N$ source views, where the mean and variance of $N$ source features are computed and concatenated as the corresponding voxel feature. We then leverage a 3D U-Net [14] to further enhance the feature volume to acquire the final geometry feature volume $V$ . + +Acquire appearance features. For each sampled 3D point $\mathbf{p}(t) = \mathbf{o} + t\mathbf{d}$ along the ray emitted from the target pixel on the novel view with camera origin $\mathbf{o}$ + +and direction $\mathbf{d}$ , we project it on the source views to acquire the projected colors $\{\mathbf{c}_i\}_{i=1}^N$ and appearance features $\{\mathbf{f}_i\}_{i=1}^N$ using bilinear interpolation. + +Geometry and appearance estimation. For more accurate reconstruction of the scene geometry and mitigation of the intrinsic color-density ambiguity [76], we estimate the per-point SDF and model the density as a function of SDF, following [40,55,76]. For appearance, we estimate the per-point radiance through color blending, i.e., predicting the blending weights of the projected colors on the source views to compose the per-point color, following [38,77,88]. The geometry and appearance are modeled using separate branches and can be formulated as: + +$$ +s = \mathbf {M} _ {s d f} (\{\mathbf {f} _ {i} \} _ {i = 1} ^ {N}, V), \quad \{\omega_ {i} \} _ {i = 1} ^ {N} = \mathbf {M} _ {c o l o r} (\{\mathbf {f} _ {i} \} _ {i = 1} ^ {N}, \mathbf {d}), \qquad (1) +$$ + +where $s$ is the SDF, $\{\omega_i\}_{i=1}^N$ are the blending weights, and $\mathbf{M}_{sdf}$ and $\mathbf{M}_{color}$ are the geometry and appearance branches, respectively. The per-point radiance can be derived as the weighted sum $\hat{\mathbf{c}} = \sum_{i=1}^{N} \omega_i \mathbf{c}_i$ across $N$ source views. + +For SDF-based volumetric rendering, we adopt the formulation in NeuS [76]. Specifically, the color is accumulated along the $K$ sampled points along each ray: + +$$ +\hat {\mathbf {C}} = \sum_ {j = 1} ^ {K} T _ {j} \alpha_ {j} \hat {\mathbf {c}} _ {j}, \quad \alpha_ {j} = 1 - \exp \left(- \int_ {t _ {j}} ^ {t _ {j + 1}} \rho (t) d t\right), \tag {2} +$$ + +where $T_{j} = \prod_{n=1}^{j-1}(1 - \alpha_{n})$ is the accumulated transmittance and $\rho(t)$ is the density determined by the estimated SDF, following the definition in NeuS [76]. Complexities of the two branches. When predicting the radiance of a sampled 3D point, the colors of valid projections on the source views are already close to the radiance of this 3D point. In light of this, we employ a lightweight $\mathbf{M}_{color}$ , consisting of three MLP layers, for efficient blending weight prediction. In contrast, the geometry branch should be allocated with sufficient complexity for accurately estimating the occlusion effects among different sampled points along the same ray as well as that among the sampled points and the source views. We introduce the design of the geometry branch in the next subsection. + +# 3.3 The Proposed Transformer-based Geometry Branch + +Overview. Our geometry branch leverages transformer modules [71] to progressively fuse geometry and appearance features, thus properly handling the two occlusion effects mentioned above. Specifically, for the $K$ sampled 3D points along the ray, we acquire their geometry features $\{\mathbf{v}_k\}_{k=1}^K$ from the geometry feature volume $V$ , which are sequentially processed through $B$ blocks ( $B = 2$ in our design), each consisting of a geometry transformer, an appearance transformer, and an occlusion transformer, to produce the per-point SDF. + +The geometry transformer. The geometry transformer fuses the geometry features into its inputs by using $\{\mathbf{v}_k\}_{k=1}^K$ as the key and value in a cross-attention scheme [71] and performs attention across the sampled points along the same ray to model their occlusion effects. This process can be formulated as follows: + +$$ +\mathbf {M} _ {s d f} ^ {g e o} (\mathbf {x}, \{\mathbf {v} _ {k} \} _ {k = 1} ^ {K}) = C r o s s A t t e n t i o n (\mathbf {q} = \mathbf {x}, \mathbf {k} = \mathbf {v} = \{\mathbf {v} _ {k} \} _ {k = 1} ^ {K}), \qquad (3) +$$ + +where $\mathbf{x}$ is the input to the geometry transformer, which is exactly the geometry features themselves for the geometry transformer in the first block, and $\mathbf{q}$ , $\mathbf{k}$ , and $\mathbf{v}$ are the query, key, and value of the attention scheme [71], respectively. + +The appearance transformer. The appearance transformer integrates the appearance features $\{\mathbf{f}_i\}_{i=1}^N$ into the outputs of the geometry transformer, considering the potential occlusions among the sampled points and the source views that may cause invalid projections. This integration is achieved using subtraction attention [75,92], which is more effective for geometric relationship reasoning [75,92], with $\{\mathbf{f}_i\}_{i=1}^N$ as the key and value. Specifically, subtraction attention computes the attention scores between the input query features and each source view in the key features, which are then used to aggregate different source views in the value features into new output features with the same dimension as the input query features. This process can be formulated as follows: + +$$ +\mathbf {M} _ {s d f} ^ {a p p r} (\mathbf {x}, \{\mathbf {f} _ {i} \} _ {i = 1} ^ {N}) = S u b A t t e n t i o n (\mathbf {q} = \mathbf {x}, \mathbf {k} = \mathbf {v} = \{\mathbf {f} _ {i} \} _ {i = 1} ^ {N}), \tag {4} +$$ + +where $\mathbf{x}$ is the output of $\mathbf{M}_{sdf}^{geo}$ . The detailed formulation of subtraction attention is provided in the supplementary materials. + +The occlusion transformer. We further integrate an occlusion transformer that performs self-attention across sampled points along the same ray to model their occlusion effects explicitly: + +$$ +\mathbf {M} _ {s d f} ^ {\text {o c c}} (\mathbf {x}) = \text {S e l f A t t e n t i o n} (\mathbf {q} = \mathbf {k} = \mathbf {v} = \mathbf {x}), \tag {5} +$$ + +where $\mathbf{x}$ is the output of $\mathbf{M}_{sdf}^{app}$ . We apply another MLP on top of the output features of the last $M_{sdf}^{occ}$ module to predict the per-point SDF $s$ in Eq. 1. + +# 3.4 Training Objectives + +Our NeRF backbone is trained using two objectives: (1) a photometric loss $\mathcal{L}_{rgb}$ between the rendered images and the ground-truth images, and (2) a depth loss $\mathcal{L}_{depth}$ between the rendered depth $\hat{\mathbf{D}} = \sum_{j=1}^{K} T_j \alpha_j t$ and the ground-truth depth. The overall objective can be formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o l o r}} + \beta \mathcal {L} _ {\text {d e p t h}}, \tag {6} +$$ + +where $\mathcal{L}_{rgb} = \frac{1}{R}\sum_{\mathbf{r} = 1}^{R}\left\| \hat{\mathbf{C}}_{\mathbf{r}} - \mathbf{C}_{\mathbf{r}}\right\|_{2}^{2}$ with $R$ denoting the number of rendered pixels, $\mathcal{L}_{depth} = \frac{1}{R}\sum_{\mathbf{r} = 1}^{R}\left\| \hat{\mathbf{D}}_{\mathbf{r}} - \mathbf{D}_{\mathbf{r}}\right\|_{2}^{2}$ , and $\beta$ is set as 1. + +# 3.5 Evaluation: Generalizable 3D Reconstruction + +Setup. We assess the backbone of Omni-Recon by evaluating its generalizable 3D reconstruction capability. Datasets: Following [10, 34, 40, 55], we adopt the DTU dataset [1], which comprises 124 different scenes and 7 distinct lighting conditions. For testing, 15 scenes are utilized, while the remaining scenes are allocated for training, following the dataset split in [34, 40, 55]. Training settings: During training, we employ $N = 4$ source views with a resolution of $640 \times 512$ . The ray number per batch and batch size are set to 1024 and 2, respectively, + +Table 1: Benchmark the quantitative performance in sparse view mesh reconstruction in terms of Chamfer distance (the lower, the better) on 15 testing scenes on DTU. The best scores are in bold, and the second-best are underlined. + +
MethodMean243740556365698397105106110114118122
COLMAP [57]1.520.902.891.631.082.181.941.611.302.341.281.101.420.761.171.14
MVSNet [83]1.221.052.521.711.041.451.520.881.291.381.050.910.660.611.081.16
IDR [86]3.394.016.403.521.913.962.364.851.626.375.971.234.730.911.721.26
VolSDF [84]3.414.034.216.120.918.241.732.741.825.143.092.084.810.603.512.18
UNISURF [48]4.395.087.183.965.304.612.243.943.145.633.405.096.382.984.052.81
NeuS [76]4.004.574.493.974.324.631.954.683.834.152.501.526.471.265.576.11
PixelNeRF [88]6.185.138.075.854.407.114.645.686.769.056.113.955.926.266.896.93
IBRNet [77]2.322.293.702.661.833.022.831.772.282.731.961.872.131.582.052.09
MVSNeRF [10]2.091.963.272.541.932.572.711.821.722.291.751.721.471.292.092.26
SparseNeuS [40]1.962.173.292.741.672.692.421.581.861.941.351.501.450.981.861.87
VolRecon [55]1.381.202.591.561.081.431.921.111.481.421.051.191.380.741.231.27
ReTR [34]1.171.052.311.440.981.181.520.881.351.300.871.070.770.591.051.12
Omni-Recon (Ours)1.130.912.131.520.931.091.700.841.291.200.831.040.810.551.051.05
+ +following [34,55]. Mesh extraction: We utilize TSDF fusion [45,63] to reconstruct scene meshes from predicted source view depths, following [34,55]. More detailed model/training settings can be found in the supplementary materials. + +Baselines. We primarily benchmark against SOTA generalizable neural implicit reconstruction methods [16, 34, 55], along with generalizable neural rendering methods [10, 77, 88], per-scene optimization methods [48, 76, 84, 86], and multiview stereo methods [57, 83]. We directly report their official results in [34, 55]. + +Sparse view reconstruction. We benchmark the performance of sparse view reconstruction using three views from each testing scene in DTU, following the setting in [16, 34, 55]. As shown in Tab. 1, we can observe that (1) Our Omni-Recon can achieve new SOTA reconstruction performance, with the lowest Chamfer distance averaged over all scenes. Additionally, as visualized in Fig. 2 (Row 1), our reconstructed mesh can exhibit smooth surfaces and correctly maintain the structure, e.g., the hole in the mouth. (2) Our method achieves the best reconstruction performance in 10 out of 15 scenes, indicating its general effectiveness across diverse scenes. + +Generalization to more diverse scenes. To demonstrate the generalization capability on more diverse scenes with larger domain shifts compared to the training scenes, we further benchmark the mesh reconstruction quality on the NeRF-Synthetic dataset [41] and NSVF-Synthetic [37] under a full-view reconstruction setting, i.e., the depth of each source view from each scene's training set is estimated and fused into a mesh using TSDF fusion [45, 63]. As visualized in Fig. 2 (Rows 2-4), we observe that our Omni-Recon can deliver notably higher-quality meshes with better-maintained structures, fewer holes, and smoother surfaces compared to the strongest baselines ReTR [34] and VolRecon [55]. The decent generalization capability of our method is attributed to the proper decoupling and fusion of geometry and appearance features in our backbone. More visualizations and ablation studies, such as the contributions of each backbone component, are provided in the supplementary materials. + +Rendering quality. We further benchmark the performance of generalizable novel view rendering on the test scenes from DTU. As shown in Tab. 2, our + +![](images/253e7b23dc30b8c24932e32989d56897591d241547df22e0ec08fea8de4aceb3.jpg) +Fig. 2: Visualize the reconstructed mesh of our Omni-Recon and the two strongest baselines [34, 55]. Row 1: Scan65 from the test scenes of DTU; Rows 2-4: Scenes from NeRF-Synthetic [41] and NSVF-Synthetic [37], which present relatively challenging cases due to domain shifts w.r.t. the training scenes from DTU. + +Table 2: Benchmark the quality of novel view rendering in terms of PSNR on DTU. + +
MethodMean.243740556365698397105106110114118122
VolRecon [55]24.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]25.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Omni-Recon26.3224.7722.3323.9225.5626.3724.7525.1924.9424.9225.0628.3928.6328.0131.4930.5
+ +method achieves the highest PSNR, e.g., a $+0.73$ and $+1.74$ PSNR improvement over ReTR and VolRecon, respectively. As will be demonstrated in Sec. 5.4, this PSNR can be notably boosted with super-fast Nvdiffrast-based finetuning [31]. + +# 4 Omni-Recon: Enable Zero-shot Scene Understanding + +To enable the understanding of various 3D scene properties, one approach is to jointly train the NeRF backbone on all target tasks. However, this approach is not scalable due to the diversity of scene properties, such as semantic segmentation, edge detection, and keypoint prediction, making it costly to cover new properties of interest. Therefore, it is highly desirable to enable zero-shot scene understanding for new tasks that were not seen during training. + +# 4.1 Our Strategy: The Predict-then-Blend Strategy + +Our hypothesis. We hypothesize that with accurate geometry and appearance estimation, the blending weights initially learned for radiance can be repurposed for other scene understanding tasks. In essence, just as blending source colors + +![](images/2499968a06e76450b64793bc71a72b1a6d2e791f72f49d34b6b6f2244b7a3f6b.jpg) +Fig. 3: Benchmark our predict-then-blend strategy with the render-then-predict baseline across different scene understanding tasks on Replica [61] and ScanNet [17]. + +yields radiance for each sampled 3D point, blending the properties of source views using the same weights can yield other properties. This hypothesis stems from the observation that scene appearance is closely related to other scene properties and thus regions with similar appearance are likely to share similar scene properties, such as semantics and edges. As such, the blending weights learned for appearance can also indicate the weighting for other scene properties. + +Our method. Motivated by this hypothesis, our Omni-Recon employs a predict-then-blend strategy to enable zero-shot multitask scene understanding by lifting 2D monocular priors to 3D novel views. Specifically, considering the prevalence of 2D monocular vision models, our method first utilizes pretrained 2D models to generate predictions (e.g., logits in semantic segmentation) on each source view, denoted as $\{\mathbf{P}_i\}_{i=1}^N$ . Then, the property of each sampled point can be obtained by blending their projected predictions on source views $\{\mathbf{p}_i\}_{i=1}^N$ , reusing the RGB blending weights, i.e., $\hat{\mathbf{p}} = \sum_{i=1}^{N} \omega_i \mathbf{p}_i$ . The final pixel-wise prediction can be derived using volumetric rendering similar to RGB reconstruction in Eq. 2. + +# 4.2 Evaluation: Zero-shot Multitask Scene Understanding + +Setup. To evaluate the zero-shot scene understanding capabilities of our model, we adopt two indoor 3D scene datasets: Replica [61] and ScanNet [17], following + +the dataset splits in [93] and [36], respectively. Specifically, we focus on four scene understanding tasks: semantic segmentation, edge detection, and two keypoint detection tasks defined in [89]. For these tasks, the 2D monocular priors are provided by [51], [5], and [56], respectively. We employ mIoU for the semantic segmentation task and $\ell_1$ error for the other three tasks as default metrics. + +Baseline. An intuitive way to achieve zero-shot 3D scene understanding is to leverage existing 2D monocular models to predict the scene properties on top of the rendered images, a process referred to as a render-then-predict pipeline. We benchmark this strategy against our proposed predict-then-blend strategy. + +Zero-shot scene understanding benchmark. As shown in Tab. 3 and Fig. 3, we observe that (1) Our predict-then-blend strategy achieves decent scene understanding performance on unseen scenes and unseen tasks, which echoes our hypothesis in Sec. 4.1 that 2D monocular priors can be elevated to 3D novel views through + +Table 3: Benchmark the two zero-shot scene understanding strategies across tasks. + +
DatasetMethod↑ Semantic↓ Edge↓ Key Point↓ Key Point 3D
Replica [61]Render-then -Predict15.640.04560.11010.0470
Predict-then -Blend (Ours)32.110.04120.07740.0176
ScanNet [17]Render-then -Predict41.320.04710.05680.0412
Predict-then -Blend (Ours)61.110.04340.04240.0197
+ +blending weight reuse, thanks to the high correlation between scene appearance and other properties; (2) Compared to the render-then-predict scheme, our method achieves notably higher quantitative accuracy, e.g., a $19.79\%$ higher mIoU on ScanNet, and consistently better visual effects across all tasks. We attribute this improvement to two aspects: (1) Our predict-then-blend strategy can avoid feeding noisy inputs caused by rendering errors to the 2D monocular models, thus reducing reliance on the robustness of monocular models and ensuring more accurate monocular priors; (2) The multiview information gathered by our strategy can contribute to a more accurate understanding of objects with limited observations under a monocular setting. More experiments on other vision models like CLIP-LSeg [32] are in the supplementary materials. + +# 5 Omni-Recon: Adaptability to Downstream 3D Tasks + +In addition to the generalization capabilities, we further demonstrate the adaptability of our Omni-Recon by extending it to three different types of downstream 3D applications. These demonstrations highlight the potentially broad utility of image-based rendering pipelines, which may have been overlooked by previous works, and can be generalized to a wider range of real-world 3D applications. + +# 5.1 Application 1: A New Real-time Rendering Pipeline + +To rapidly enable real-time rendering using Omni-Recon's pretrained backbone, we build a real-time rendering pipeline based on Nvdiffrast [31]. This involves baking the complex geometry branch into meshes using TSDF fusion [45,63] as discussed in Sec. 3.5, and using the lightweight appearance branch as a Shader. + +Our real-time rendering pipeline. As illustrated in Fig. 1 (bottom left), for any novel view of a new scene, we first transform its camera pose to clip space as required by Nvdiffrast [31]. Then, we perform a rasterization process to obtain the intersection point of each camera ray with the extracted mesh. Next, these + +intersection points are input into the shader to predict the color by blending their projected colors on the source views as described in Sec. 3.2. Thanks to the blending scheme, where valid projections can provide strong appearance clues, the rendering quality is often satisfactory even without finetuning. + +Joint mesh and Shader finetuning. To further enhance the rendering quality, we can jointly finetune the mesh and the Shader in a gradient-based manner using differentiable rendering enabled by Nvdiffrast [31]. Specifically, we finetune the location of each vertex in the mesh and the parameters of the lightweight Shader. Additionally, we periodically prune redundant mesh faces that cannot intersect with any camera rays in one training epoch. More details are provided in the supplementary materials. + +# 5.2 Application 2: Parameter-Efficient Tuning for 3D Understanding + +We explore the possibility of finetuning Omni-Recon for various downstream 3D tasks, utilizing a series of 3D scene understanding tasks as demonstrations. Specifically, we conduct PET [19] on top of Omni-Recon's pretrained backbone by incorporating low-rank adapters [26] into the linear layers within the transformer modules introduced in Sec. 3.3, and only tuning these adapters while keeping the pretrained backbone frozen. It is worth noting that when evaluating the performance of the tuned model, we assess it on new scenes that have never been encountered during both training and tuning, representing a generalizable scene understanding scenario [36]. + +# 5.3 Application 3: Text-guided 3D Scene Editing + +We target text-guided 3D scene editing, which involves editing a 3D scene based on textual instructions. To achieve this, our Omni-Recon capitalizes on the nature of image-based rendering, allowing edits made to 2D source views to be propagated to 3D novel views. Specifically, leveraging the availability of powerful 2D diffusion models, we use publicly available ones to edit the 2D source views. However, doing so naively may result in inconsistent source views, as the same textual instruction may lead to different visualizations, thus degrading reconstruction quality. Inspired by [23], which iteratively edits the training images and retransmits a new NeRF to achieve scene editing, we address the inconsistent editing issue by proposing an iterative editing and reconstruction pipeline. + +The iterative editing and reconstruction pipeline. To ensure consistent 3D editing, we iterate between source view editing and reconstruction as follows: First, we edit each source view using the diffusion model InstructPix2Pix [4], which is guided by both textual instructions and the original image to maintain image fidelity. While the edited source views can faithfully follow the textual instructions, they may exhibit inconsistencies in the edited region. Next, we reconstruct each source view by conditioning on their nearby source views, as described in Sec. 3.2. This process improves the 3D consistency among source views at the expense of reduced image fidelity and less adherence to instructions. We then iterate through the editing and reconstruction steps to progressively enhance image fidelity, 3D consistency, and adherence to instructions. The resulting source views are finally used to reconstruct the entire scene. + +Table 4: Benchmark the rendering speed and rendering quality of our real-time rendering pipeline and the baselines [34,55] on DTU. The best PSNR is in bold, and the shortest finetuning time that our pipelines surpass the strongest baseline is underlined. + +
MethodFPSMean243740556365698397105106110114118122
VolRecon [55]0.02924.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]0.02425.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Ours w/o ft.71.3(40.82)22.9620.1219.7122.2722.7824.5521.7721.5126.7222.3324.4922.5222.9324.6823.8424.12
Ours (ft. 10s)25.6821.4221.6824.0624.1228.1924.1023.9531.6524.4128.1525.6325.8526.1526.8226.89
Ours (ft. 20s)27.2122.6322.9225.1225.4230.0325.9526.1633.1926.2230.3627.4427.0426.9329.1529.65
Ours (ft. 30s)27.7823.223.2625.5425.730.5926.8326.9633.6626.4730.7328.1427.7027.130.1730.65
Ours (ft. 1min)28.3424.6924.1125.7626.0530.9327.6627.4933.6827.0730.9728.5128.5427.3531.1731.54
Ours (ft. 3min)28.9525.224.3225.9426.1632.1528.9927.8834.9427.3531.6228.9328.9727.5631.7032.49
Ours (ft. 5min)29.0225.3424.3625.6326.2132.1629.3327.8134.9427.3231.7429.0429.0527.6931.7432.89
+ +# 5.4 Evaluation: Effectiveness on Different Downstream 3D Tasks + +Real-time rendering. We evaluate our real-time rendering pipeline in Sec. 5.1 in terms of both rendering quality and rendering speed, measured on an NVIDIA A5000 GPU, using the DTU dataset. Specifically, we finetune the scene meshes, extracted by TSDF fusion as mentioned in Sec. 3.5, and the Shader in no more than 5 minutes and record the rendering quality at different time steps. + +Observations and analysis: As shown in Tab. 4, we can observe that (1) Without any finetuning, the rendering pipeline already exhibits certain rendering capabilities, despite the model backbone being trained using volumetric rendering with more than one sampled point; (2) Our rendering pipeline can be swiftly finetuned to significantly boost the PSNR. For instance, with just a 10-second finetuning, our rendering pipeline can match the average rendering quality of the strongest baseline, ReTR [34]; With 1-minute/2-minute finetuning, it can surpass ReTR by a +2.75 and +3.36 PSNR improvement on average, respectively; (3) Our rendering pipeline enables real-time rendering with 71.3 FPS, i.e., a >2458× speed-up over the baselines. Here, the assumption is that the feature extraction of each source view from a target scene can be done in a one-time effort and does not account for throughput. If we still consider this feature extraction latency for each rendering, the resulting 40.82 FPS still meets the real-time requirement. + +These experiments indicate that given a new scene, our Omni-Recon is promising to enable both instant and accurate reconstruction and real-time rendering. + +PET for 3D Understanding. To evaluate the adaptability of our Omni-Recon, we apply PET to it for various scene understanding tasks on ScanNet [17]. Specifically, we finetune a LoRA adaptor [26] for each task on a set of training scenes and evaluate on + +Table 5: Evaluate the effectiveness of PET on Omni-Recon and benchmark with SRay [36]. + +
Method↑ Sem. mIoU↑ Sem. Total Acc↑ Sem. Avg Acc↓ Edge↓ Key Point↓ Key Point 3D
SRay [36]57.1578.2462.55---
Ours (Zero-shot)61.1180.8069.170.04340.04240.0197
Ours (PET + Zero-shot)62.3581.8469.890.03420.02070.0132
+ +non-overlapping test scenes, representing a generalizable 3D understanding setting, following the dataset split described in [36]. We benchmark against the SOTA generalizable semantic segmentation work, SRay [36]. + +Observations and analysis: As shown in Tab. 5, we can observe that (1) After PET, our model outperforms SRay in terms of generalizable semantic segmentation with a $+5.20\%$ mIoU and a $+7.34\%$ accuracy improvement; (2) Despite + +![](images/6c7be7d29fa88ab9cad67e961ec0abd8829608f9dfec4b17ab9ee1cd19f8a98c.jpg) +Fig. 4: Visualize the scene understanding performance on hard cases before/after PET. + +![](images/33fe245c26fdbe0ab18443afadf59a49d34d5e8ccb4aa221368daa5744f241c1.jpg) +Fig. 5: Three text-guided scene editing examples using our pipeline in Sec. 5.3. + +having limited trainable parameters, PET can enhance the scene understanding performance across tasks, especially on challenging cases visualized in Fig. 4, indicating the adaptability of Omni-Recon to diverse downstream tasks. + +Text-guided scene editing. We visualize the edited examples in Fig. 5, showcasing the ability to ensure both instruction-following and 3D consistency. + +# 6 Conclusion + +Motivated by the significant demand for supporting various functionalities required by different 3D applications through a unified NeRF model, our work aims to develop a general-purpose NeRF capable of handling a broad spectrum of 3D tasks. We propose a framework called Omni-Recon, which features a general-purpose NeRF model using image-based rendering with two separate branches, and demonstrate its integration with various 3D tasks. Specifically, our Omni-Recon can achieve SOTA generalizable 3D reconstruction quality, enable zero-shot multitask scene understanding, achieve SOTA scene understanding performance and real-time rendering after rapid adaptation, and support scene editing. Omni-Recon underscores the potential of image-based rendering pipelines in diverse real-world 3D applications, which may have been underestimated by prior studies, thus sparking advancements in future rendering pipelines. + +# Acknowledgement + +The work is supported by the National Science Foundation (NSF) through the SCH program (Award number: 1838873) and CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA. + +# References + +1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. International Journal of Computer Vision 120, 153-168 (2016) +2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +5. Canny, J.: A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence (6), 679-698 (1986) +6. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023) +7. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. CVPR (2021) +8. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517 (2022) +9. Chen, A., Xu, Z., Wei, X., Tang, S., Su, H., Geiger, A.: Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226 (2023) +10. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021) +11. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022) +12. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022) +13. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021) + +14. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016) +15. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3193-3204 (2023) +16. Croce, F., Hein, M.: Sparse and imperceivable adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4724-4732 (2019) +17. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) +18. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022) +19. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence 5(3), 220-235 (2023) +20. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +21. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nef: 3d-to-2d label transfer for panoptic urban scene segmentation. arXiv preprint arXiv:2203.15224 (2022) +22. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021) +23. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789 (2023) +24. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (2018) +25. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021) +26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +27. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022) +28. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021) +29. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022) + +30. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022) +31. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020) +32. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN +33. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. CVPR (2021) +34. Liang, Y., He, H., Chen, Y.: Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems 36 (2024) +35. Lindell, D.B., Martel, J.N., Wetzstein, G.: Autoint: Automatic integration for fast neural rendering. CVPR (2021) +36. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023) +37. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022) +39. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. SIGGRAPH (2019) +40. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022) +41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020) +42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022) +44. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16. pp. 414-431. Springer (2020) +45. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011) +46. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. + +In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022) +47. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021) +48. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021) +49. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. CVPR (2021) +50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021) +51. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. ArXiv preprint (2021) +52. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: DeRF: Decomposed radiance fields. CVPR (2021) +53. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021) +54. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021) +55. Ren, Y., Zhang, T., Pollefeys, M., Süsstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023) +56. Sax, A., Emi, B., Zamir, A.R., Guibas, L.J., Savarese, S., Malik, J.: Mid-level visual representations improve generalization and sample efficiency for learning visuomotor policies. (2018) +57. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016) +58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3D-aware image synthesis. NeurIPS (2020) +59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. arXiv preprint arXiv:2206.07695 (2022) +60. Seitz, S.M., Dyer, C.R.: Photorealistic scene reconstruction by voxel coloring. IJCV (1999) +61. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) +62. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022) +63. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021) + +64. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022) +65. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023) +66. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (2019) +67. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019) +68. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15182-15192 (2021) +69. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. arXiv preprint arXiv:2211.11738 (2022) +70. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022) +71. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +72. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022) +73. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021) +74. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022) +75. Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z., et al.: Is attention all nerf needs? arXiv preprint arXiv:2207.13298 (2022) +76. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021) +77. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4690-4699 (2021) +78. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504 (2021) +79. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 106-122. Springer (2022) +80. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928 (2022) + +81. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 $\{\backslash \mathrm{deg}\}$ views. arXiv preprint arXiv:2211.16431 (2022) +82. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) +83. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018) +84. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) +85. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023) +86. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020) +87. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021) +88. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021) +89. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2018) +90. Zhang, M., Zheng, S., Bao, Z., Hebert, M., Wang, Y.X.: Beyond rgb: Scene-property synthesis with neural radiance fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 795-805 (2023) +91. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5449-5458 (2022) +92. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021) +93. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021) \ No newline at end of file diff --git a/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/images.zip b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f5e79d55316b246d08bf6c5a81a15c793a8a0d6e --- /dev/null +++ b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce27cc69793b6f50ce12989dba255e360da0060676aa68fcf6db0ea4ae4a6a0a +size 617159 diff --git a/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/layout.json b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dd90692fade759dd8c44d2be794ddbb4138aaaca --- /dev/null +++ b/2024/Omni-Recon_ Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields/layout.json @@ -0,0 +1,9927 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "text", + "content": "Omni-Recon: Harnessing Image-based Rendering for General-Purpose Neural Radiance Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 169, + 416, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 169, + 416, + 194 + ], + "spans": [ + { + "bbox": [ + 196, + 169, + 416, + 194 + ], + "type": "text", + "content": "Yonggan Fu, Huaizhi Qu, Zhifan Ye, Chaojian Li, Kevin Zhao, and Yingyan (Celine) Lin" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 183, + 202, + 430, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 202, + 430, + 225 + ], + "spans": [ + { + "bbox": [ + 183, + 202, + 430, + 225 + ], + "type": "text", + "content": "Georgia Institute of Technology {yfu314, zye327, cli851, kzhao14, celine.lin}@gatech.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 259, + 455, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 259, + 455, + 567 + ], + "spans": [ + { + "bbox": [ + 160, + 259, + 455, + 567 + ], + "type": "text", + "content": "Abstract. Recent breakthroughs in Neural Radiance Fields (NeRFs) have sparked significant demand for their integration into real-world 3D applications. However, the varied functionalities required by different 3D applications often necessitate diverse NeRF models with various pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. Drawing inspiration from the generalization capability and adaptability of emerging foundation models, our work aims to develop one general-purpose NeRF for handling diverse 3D tasks. We achieve this by proposing a framework called Omni-Recon, which is capable of (1) generalizable 3D reconstruction and zero-shot multitask scene understanding, and (2) adaptability to diverse downstream 3D applications such as real-time rendering and scene editing. Our key insight is that an image-based rendering pipeline, with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our Omni-Recon features a general-purpose NeRF model using image-based rendering with two decoupled branches: one complex transformer-based branch that progressively fuses geometry and appearance features for accurate geometry estimation, and one lightweight branch for predicting blending weights of source views. This design achieves state-of-the-art (SOTA) generalizable 3D surface reconstruction quality with blending weights reusable across diverse tasks for zero-shot multitask scene understanding. In addition, it can enable real-time rendering after baking the complex geometry branch into meshes, swift adaptation to achieve SOTA generalizable 3D understanding performance, and seamless integration with 2D diffusion models for text-guided 3D editing. Our code is available at: https://github.com/GATECH-EIC/Omni-Recon." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 589, + 230, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 230, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 230, + 601 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs) [41] have garnered significant attention and hold promise to become crucial enablers for emerging 3D applications. However, deploying NeRFs in real-world scenarios still requires substantial effort, as different 3D applications often demand distinct NeRF models with varied" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "pipelines, leading to tedious NeRF training for each target task and cumbersome trial-and-error experiments. For instance, while both cross-scene generalization and real-time rendering are desirable in the real world, generalizable NeRFs for instant reconstruction [10,38,77] and real-time NeRFs using mesh-based rasterization [12,65,85] typically feature diverse pipelines, making it hard to simultaneously satisfy the two requirements. Additionally, understanding new scene properties [36,90,93] necessitates training new NeRF models, which is not scalable given the increasing number of scene properties of interest." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 368 + ], + "type": "text", + "content": "In parallel, the emergence of foundation models has driven transformative progress in real-world artificial intelligence applications. Inspired by this exciting trend, there has been a growing interest in integrating NeRFs into this realm by developing general-purpose NeRFs for handling various 3D tasks. Specifically, similar to the generalization capability and adaptability of foundation models, we expect general-purpose NeRFs to possess three essential capabilities, including (1) generalizable 3D reconstruction for instant surface reconstruction and novel view synthesis, (2) zero-shot multitask scene understanding of various scene properties, e.g., semantics and edges, and (3) adaptability to diverse downstream tasks such as real-time rendering and scene editing. Achieving these capabilities is promising to enable users to effortlessly start from these general-purpose NeRFs when facing a new 3D application, significantly reducing the need for training and experimenting with various NeRF pipelines." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 368, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 499 + ], + "type": "text", + "content": "Despite the promise, existing NeRF pipelines have not yet realized the mentioned capabilities. Firstly, per-scene optimized NeRFs [2, 3, 22, 25, 27, 42, 85] rely on costly per-scene training, thus precluding generalization capabilities. Secondly, although generalizable NeRF designs [10, 38, 54, 68, 75, 77, 82, 88] can achieve cross-scene generalization, their huge computational complexity limits their suitability for application scenarios where real-time rendering is crucial. Some generalizable surface reconstruction methods [34, 55] can support mesh extraction, but an expensive volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views, leaving the demand for real-time rendering unmet. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 499, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 608 + ], + "type": "text", + "content": "Given the significant potential and associated challenges, our work aims to advance the integration of NeRFs into the realm of foundation models by developing general-purpose NeRFs. To achieve this goal, we propose a framework called Omni-Recon, which is an image-based rendering pipeline that realizes all the mentioned capabilities. The key insight leveraged by our Omni-Recon is that an image-based rendering pipeline, when equipped with accurate geometry and appearance estimation, can lift 2D image features into their 3D counterparts, thus extending widely explored 2D tasks to the 3D world in a generalizable manner. Specifically, our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 137, + 617, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 482, + 666 + ], + "type": "text", + "content": "- We introduce Omni-Recon, a general-purpose NeRF pipeline, which possesses both decent generalization capabilities, including generalizable 3D reconstruction and zero-shot multitask scene understanding, and adaptability to diverse downstream 3D applications." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 485, + 392 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 116, + 485, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 485, + 223 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 485, + 223 + ], + "type": "text", + "content": "- Our Omni-Recon features a general-purpose NeRF model backbone using image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and appearance features to accurately predict a signed distance function (SDF), and one lightweight appearance branch for predicting blending weights of source views from the target scene. The advantages of this design include: (1) ensuring accurate geometry estimation for high-quality 3D reconstruction, and (2) enabling real-time rendering after baking the complex geometry branch into meshes and using the lightweight appearance branch as a Shader." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 224, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 224, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 481, + 270 + ], + "type": "text", + "content": "- We intriguingly find that a properly trained image-based rendering pipeline can directly lift 2D monocular priors to 3D novel views by reusing the blending weights predicted by the appearance branch and leverage this to enable generalizable, zero-shot, and multitask scene understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 272, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 481, + 342 + ], + "type": "text", + "content": "- We demonstrate the adaptability of Omni-Recon by extending it to diverse downstream tasks, including: (1) a Nvdiffrast-based [31] real-time rendering pipeline using the extracted mesh, (2) parameter-efficient tuning (PET) of Omni-Recon on downstream generalizable scene understanding tasks, and (3) a new text-guided scene editing scheme by iteratively editing and reconstructing the source images from a 3D scene using 2D diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 344, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 481, + 392 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our Omni-Recon can achieve state-of-the-art (SOTA) generalizable 3D surface reconstruction and scene understanding performance. Additionally, it can support real-time rendering of a new scene after rapid adaptation, along with easy-to-use 3D scene editing." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 402, + 481, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 451 + ], + "type": "text", + "content": "Our Omni-Recon underscores the potentially wide usage of image-based rendering pipelines in diverse real-world 3D applications, which may have been overlooked by previous works. We believe that the insights we provide could ignite future innovations in more advanced generalizable rendering pipelines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 594 + ], + "type": "text", + "content": "Neural Radiance Fields. NeRFs [42], which represent the target scene as a continuous volume with density and view-dependent color, have garnered increasing attention thanks to their SOTA rendering quality when compared to prior approaches using explicit neural representations [24,39,60,66]. Subsequent works further improve NeRFs from various angles, such as enhancing their rendering quality [2,3,11,72], training NeRFs with sparse views [18,28,46,69,80,81], extending NeRFs to large-scale scenes [64,70,78,79,91], or applying NeRFs to other tasks, such as generative modeling [7,47,58,74], dynamic scenes [6,33,49, 50], or scene understanding [21,30,73,93]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Generalizable NeRFs. Vanilla NeRFs rely on per-scene optimization, which limits their ability to perform instant reconstruction. To address this limitation, generalizable NeRF designs [10, 13, 29, 38, 54, 68, 75, 77, 82, 88] achieve cross-scene generalization by conditioning vanilla NeRFs on the source views from the target new scene. Specifically, [68, 75, 77, 88] take the extracted scene features from the source views as inputs to reconstruct the radiance field via a one-shot forward" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "pass. Subsequent works further improve cross-scene generalization capabilities by enhancing scene geometry estimation [10, 29, 38] and by introducing attention mechanisms and mixture-of-experts into NeRF backbone design [15, 36, 75]. Despite the promise of generalizable NeRFs, they still fall short in terms of rendering efficiency due to their huge computational complexity, which entails costly inference for each sampled point along the ray. While some generalizable surface reconstruction methods [34, 55] can support mesh extraction, a costly volumetric rendering process, independent of the extracted mesh, is still necessary for rendering novel views. This poses a hindrance to their deployment in 3D applications that require real-time rendering. Furthermore, their potential for zero-shot scene understanding and scene editing remains unexplored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "type": "text", + "content": "Efficient scene representations. Various scene representations have been proposed to accelerate NeRFs' training and inference process. To speed up NeRF training, [8,9,20,43,62] integrate explicit neural representations with volumetric rendering. In parallel, to speed up NeRF inference/rendering, [35,52,53] attempt to reduce the complexity of the multi-layer perceptron (MLP) model in NeRF. In addition, [22, 25, 27, 37, 53, 59, 87] explore the free space in a target 3D scene to improve the sampling efficiency and thus reduce unnecessary MLP inference. Motivated by the insight of deferred rendering [67], [12, 25, 65, 85] pre-compute scene features offline and thus require less computation at run-time. In particular, [12, 65, 85] represents NeRF as a set of textured polygons, where the textures can be pre-computed and thus the NeRF pipeline can be efficiently executed in modern graphics hardware. However, these efficient scene representations still necessitate costly per-scene optimization, lacking generalization capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 411, + 465, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 465, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 465, + 425 + ], + "type": "text", + "content": "3 Omni-Recon: The General-Purpose NeRF Backbone" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "content": "To develop general-purpose NeRFs for diverse 3D tasks, in this section, we design a general-purpose NeRF backbone and demonstrate its SOTA generalizable 3D reconstruction capability. We will leverage this backbone to enable zero-shot multitask scene understanding in Sec. 4 and adapt it to diverse downstream 3D tasks, including real-time rendering, PET for 3D scene understanding, and text-guided scene editing in Sec. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 509, + 298, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 298, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 298, + 520 + ], + "type": "text", + "content": "3.1 Backbone Design Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": "Design principles. Our design is driven by the following principles: (1) Similar to foundation models, the ability to instantly reconstruct a new scene with generalizable capabilities is essential. Therefore, we leverage image-based rendering [10,38,77], which lifts 2D image features into their 3D counterparts, as our backbone design and uncover their decent capabilities in diverse 3D applications, which may have been overlooked by previous works. (2) Considering a wide range of 3D applications that require real-time rendering, it is highly desirable that our backbone design can be seamlessly converted into efficient 3D representations, particularly meshes, to enable real-time rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Our backbone design. As shown in Fig. 1, the backbone model in OmniRecon leverages image-based rendering with two decoupled branches: one complex transformer-based geometry branch that progressively fuses geometry and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 476, + 308 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 476, + 308 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 476, + 308 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 476, + 308 + ], + "type": "image", + "image_path": "5556299a659e0430e950b9d7ef37408a3da658d642f050271fb4e782aae15406.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 311, + 411, + 321 + ], + "lines": [ + { + "bbox": [ + 200, + 311, + 411, + 321 + ], + "spans": [ + { + "bbox": [ + 200, + 311, + 411, + 321 + ], + "type": "text", + "content": "Fig. 1: An overview of our Omni-Recon framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "type": "text", + "content": "appearance features to accurately predict an SDF, and one lightweight appearance branch for estimating blending weights of source views. Specifically, the progressive feature fusion in the geometry branch empowers accurate geometry estimation for high-quality 3D reconstruction, and the decoupling between geometry and appearance enables real-time rendering after baking the complex geometry branch into meshes, with the lightweight appearance branch serving as a shader. Additionally, the nature of image-based rendering, when equipped with accurate geometry and appearance estimation, can enable a wide range of downstream 3D applications, as shown in Sec. 4 and Sec. 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 451, + 389, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 389, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 389, + 464 + ], + "type": "text", + "content": "3.2 The Overall Image-Based Rendering Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 481, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 519 + ], + "type": "text", + "content": "Following generalizable NeRFs [10,38,77,88], our NeRF backbone is conditioned on source views from the target scene. This is achieved by estimating the geometry and appearance based on the extracted features of the source views from the target scene, which is elaborated as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": "Acquire geometry features. Given " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": " source views " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": " from the target scene, we utilize a CNN encoder [34, 55] to extract source features " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i\\}_{i=1}^N \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": " from the source views. Next, we construct a 3D feature volume " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{M \\times M \\times M \\times C}" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": ", which serves similar roles as cost volumes for multiview stereo [10, 83], to aggregate geometry information from different source views, following [44, 55, 63]. Specifically, we divide the bounding volume of the scene into " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "M^3" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": " voxels and project each voxel center onto " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": " source views, where the mean and variance of " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": " source features are computed and concatenated as the corresponding voxel feature. We then leverage a 3D U-Net [14] to further enhance the feature volume to acquire the final geometry feature volume " + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 520, + 482, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Acquire appearance features. For each sampled 3D point " + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{p}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": " along the ray emitted from the target pixel on the novel view with camera origin " + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "and direction " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": ", we project it on the source views to acquire the projected colors " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": " and appearance features " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": " using bilinear interpolation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "content": "Geometry and appearance estimation. For more accurate reconstruction of the scene geometry and mitigation of the intrinsic color-density ambiguity [76], we estimate the per-point SDF and model the density as a function of SDF, following [40,55,76]. For appearance, we estimate the per-point radiance through color blending, i.e., predicting the blending weights of the projected colors on the source views to compose the per-point color, following [38,77,88]. The geometry and appearance are modeled using separate branches and can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 229, + 481, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 229, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 192, + 229, + 481, + 244 + ], + "type": "interline_equation", + "content": "s = \\mathbf {M} _ {s d f} (\\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}, V), \\quad \\{\\omega_ {i} \\} _ {i = 1} ^ {N} = \\mathbf {M} _ {c o l o r} (\\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}, \\mathbf {d}), \\qquad (1)", + "image_path": "f002678730610f34b4d6f27022543e34a89696dd1bbf709bf281507564aadada.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " is the SDF, " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "\\{\\omega_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " are the blending weights, and " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sdf}" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{color}" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " are the geometry and appearance branches, respectively. The per-point radiance can be derived as the weighted sum " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}} = \\sum_{i=1}^{N} \\omega_i \\mathbf{c}_i" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " source views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "content": "For SDF-based volumetric rendering, we adopt the formulation in NeuS [76]. Specifically, the color is accumulated along the " + }, + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "content": " sampled points along each ray:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 202, + 309, + 481, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 309, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 202, + 309, + 481, + 341 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} = \\sum_ {j = 1} ^ {K} T _ {j} \\alpha_ {j} \\hat {\\mathbf {c}} _ {j}, \\quad \\alpha_ {j} = 1 - \\exp \\left(- \\int_ {t _ {j}} ^ {t _ {j + 1}} \\rho (t) d t\\right), \\tag {2}", + "image_path": "89af800bc8c16cce0b2069d16325e6c55938848d70893d9bc3cf9dd6b5a50c9d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "inline_equation", + "content": "T_{j} = \\prod_{n=1}^{j-1}(1 - \\alpha_{n})" + }, + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "text", + "content": " is the accumulated transmittance and " + }, + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\rho(t)" + }, + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "text", + "content": " is the density determined by the estimated SDF, following the definition in NeuS [76]. Complexities of the two branches. When predicting the radiance of a sampled 3D point, the colors of valid projections on the source views are already close to the radiance of this 3D point. In light of this, we employ a lightweight " + }, + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{color}" + }, + { + "bbox": [ + 130, + 340, + 482, + 460 + ], + "type": "text", + "content": ", consisting of three MLP layers, for efficient blending weight prediction. In contrast, the geometry branch should be allocated with sufficient complexity for accurately estimating the occlusion effects among different sampled points along the same ray as well as that among the sampled points and the source views. We introduce the design of the geometry branch in the next subsection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 471, + 422, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 422, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 422, + 483 + ], + "type": "text", + "content": "3.3 The Proposed Transformer-based Geometry Branch" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "text", + "content": "Overview. Our geometry branch leverages transformer modules [71] to progressively fuse geometry and appearance features, thus properly handling the two occlusion effects mentioned above. Specifically, for the " + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "text", + "content": " sampled 3D points along the ray, we acquire their geometry features " + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "text", + "content": " from the geometry feature volume " + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "text", + "content": ", which are sequentially processed through " + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "text", + "content": " blocks (" + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "inline_equation", + "content": "B = 2" + }, + { + "bbox": [ + 130, + 488, + 482, + 571 + ], + "type": "text", + "content": " in our design), each consisting of a geometry transformer, an appearance transformer, and an occlusion transformer, to produce the per-point SDF." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 572, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 620 + ], + "type": "text", + "content": "The geometry transformer. The geometry transformer fuses the geometry features into its inputs by using " + }, + { + "bbox": [ + 130, + 572, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 130, + 572, + 482, + 620 + ], + "type": "text", + "content": " as the key and value in a cross-attention scheme [71] and performs attention across the sampled points along the same ray to model their occlusion effects. This process can be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 168, + 630, + 481, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 630, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 168, + 630, + 481, + 646 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {s d f} ^ {g e o} (\\mathbf {x}, \\{\\mathbf {v} _ {k} \\} _ {k = 1} ^ {K}) = C r o s s A t t e n t i o n (\\mathbf {q} = \\mathbf {x}, \\mathbf {k} = \\mathbf {v} = \\{\\mathbf {v} _ {k} \\} _ {k = 1} ^ {K}), \\qquad (3)", + "image_path": "686cf39c51162e36583db88d599701a9a646f0294aa3baa98a12291bf34f9863.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " is the input to the geometry transformer, which is exactly the geometry features themselves for the geometry transformer in the first block, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " are the query, key, and value of the attention scheme [71], respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": "The appearance transformer. The appearance transformer integrates the appearance features " + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": " into the outputs of the geometry transformer, considering the potential occlusions among the sampled points and the source views that may cause invalid projections. This integration is achieved using subtraction attention [75,92], which is more effective for geometric relationship reasoning [75,92], with " + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 480, + 271 + ], + "type": "text", + "content": " as the key and value. Specifically, subtraction attention computes the attention scores between the input query features and each source view in the key features, which are then used to aggregate different source views in the value features into new output features with the same dimension as the input query features. This process can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 277, + 480, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 277, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 176, + 277, + 480, + 293 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {s d f} ^ {a p p r} (\\mathbf {x}, \\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}) = S u b A t t e n t i o n (\\mathbf {q} = \\mathbf {x}, \\mathbf {k} = \\mathbf {v} = \\{\\mathbf {f} _ {i} \\} _ {i = 1} ^ {N}), \\tag {4}", + "image_path": "348866d3c85a62166cbc5b32ec0cd3b04ee2ccd054e8aa45bd5d13350bb2aedc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 298, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 298, + 479, + 323 + ], + "type": "text", + "content": " is the output of " + }, + { + "bbox": [ + 130, + 298, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sdf}^{geo}" + }, + { + "bbox": [ + 130, + 298, + 479, + 323 + ], + "type": "text", + "content": ". The detailed formulation of subtraction attention is provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 479, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 358 + ], + "type": "text", + "content": "The occlusion transformer. We further integrate an occlusion transformer that performs self-attention across sampled points along the same ray to model their occlusion effects explicitly:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 212, + 366, + 480, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 366, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 212, + 366, + 480, + 379 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {s d f} ^ {\\text {o c c}} (\\mathbf {x}) = \\text {S e l f A t t e n t i o n} (\\mathbf {q} = \\mathbf {k} = \\mathbf {v} = \\mathbf {x}), \\tag {5}", + "image_path": "0ff04062f0518509e0e03d4630b6e2f3d0bdf196dd7846af0d39a637e34a77be.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "text", + "content": " is the output of " + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sdf}^{app}" + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "text", + "content": ". We apply another MLP on top of the output features of the last " + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "inline_equation", + "content": "M_{sdf}^{occ}" + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "text", + "content": " module to predict the per-point SDF " + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 385, + 479, + 411 + ], + "type": "text", + "content": " in Eq. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 420, + 259, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 259, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 259, + 433 + ], + "type": "text", + "content": "3.4 Training Objectives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "text", + "content": "Our NeRF backbone is trained using two objectives: (1) a photometric loss " + }, + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "text", + "content": " between the rendered images and the ground-truth images, and (2) a depth loss " + }, + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth}" + }, + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "text", + "content": " between the rendered depth " + }, + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}} = \\sum_{j=1}^{K} T_j \\alpha_j t" + }, + { + "bbox": [ + 130, + 437, + 479, + 486 + ], + "type": "text", + "content": " and the ground-truth depth. The overall objective can be formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 260, + 495, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 495, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 260, + 495, + 480, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\beta \\mathcal {L} _ {\\text {d e p t h}}, \\tag {6}", + "image_path": "380dba171b80b647a64b43a59d139055c21d9c3b3bf341b6a64d397bfd7036c1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb} = \\frac{1}{R}\\sum_{\\mathbf{r} = 1}^{R}\\left\\| \\hat{\\mathbf{C}}_{\\mathbf{r}} - \\mathbf{C}_{\\mathbf{r}}\\right\\|_{2}^{2}" + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "text", + "content": " denoting the number of rendered pixels, " + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth} = \\frac{1}{R}\\sum_{\\mathbf{r} = 1}^{R}\\left\\| \\hat{\\mathbf{D}}_{\\mathbf{r}} - \\mathbf{D}_{\\mathbf{r}}\\right\\|_{2}^{2}" + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 512, + 479, + 555 + ], + "type": "text", + "content": " is set as 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 564, + 388, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 564, + 388, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 564, + 388, + 575 + ], + "type": "text", + "content": "3.5 Evaluation: Generalizable 3D Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": "Setup. We assess the backbone of Omni-Recon by evaluating its generalizable 3D reconstruction capability. Datasets: Following [10, 34, 40, 55], we adopt the DTU dataset [1], which comprises 124 different scenes and 7 distinct lighting conditions. For testing, 15 scenes are utilized, while the remaining scenes are allocated for training, following the dataset split in [34, 40, 55]. Training settings: During training, we employ " + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": " source views with a resolution of " + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "inline_equation", + "content": "640 \\times 512" + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": ". The ray number per batch and batch size are set to 1024 and 2, respectively," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 148, + 475, + 284 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 147 + ], + "type": "text", + "content": "Table 1: Benchmark the quantitative performance in sparse view mesh reconstruction in terms of Chamfer distance (the lower, the better) on 15 testing scenes on DTU. The best scores are in bold, and the second-best are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 148, + 475, + 284 + ], + "lines": [ + { + "bbox": [ + 137, + 148, + 475, + 284 + ], + "spans": [ + { + "bbox": [ + 137, + 148, + 475, + 284 + ], + "type": "table", + "html": "
MethodMean243740556365698397105106110114118122
COLMAP [57]1.520.902.891.631.082.181.941.611.302.341.281.101.420.761.171.14
MVSNet [83]1.221.052.521.711.041.451.520.881.291.381.050.910.660.611.081.16
IDR [86]3.394.016.403.521.913.962.364.851.626.375.971.234.730.911.721.26
VolSDF [84]3.414.034.216.120.918.241.732.741.825.143.092.084.810.603.512.18
UNISURF [48]4.395.087.183.965.304.612.243.943.145.633.405.096.382.984.052.81
NeuS [76]4.004.574.493.974.324.631.954.683.834.152.501.526.471.265.576.11
PixelNeRF [88]6.185.138.075.854.407.114.645.686.769.056.113.955.926.266.896.93
IBRNet [77]2.322.293.702.661.833.022.831.772.282.731.961.872.131.582.052.09
MVSNeRF [10]2.091.963.272.541.932.572.711.821.722.291.751.721.471.292.092.26
SparseNeuS [40]1.962.173.292.741.672.692.421.581.861.941.351.501.450.981.861.87
VolRecon [55]1.381.202.591.561.081.431.921.111.481.421.051.191.380.741.231.27
ReTR [34]1.171.052.311.440.981.181.520.881.351.300.871.070.770.591.051.12
Omni-Recon (Ours)1.130.912.131.520.931.091.700.841.291.200.831.040.810.551.051.05
", + "image_path": "392c5243bf0bcf854f9c5821518b5bda36908a739af619ada6a432a7161008d9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "text", + "content": "following [34,55]. Mesh extraction: We utilize TSDF fusion [45,63] to reconstruct scene meshes from predicted source view depths, following [34,55]. More detailed model/training settings can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 480, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 374 + ], + "type": "text", + "content": "Baselines. We primarily benchmark against SOTA generalizable neural implicit reconstruction methods [16, 34, 55], along with generalizable neural rendering methods [10, 77, 88], per-scene optimization methods [48, 76, 84, 86], and multiview stereo methods [57, 83]. We directly report their official results in [34, 55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "type": "text", + "content": "Sparse view reconstruction. We benchmark the performance of sparse view reconstruction using three views from each testing scene in DTU, following the setting in [16, 34, 55]. As shown in Tab. 1, we can observe that (1) Our Omni-Recon can achieve new SOTA reconstruction performance, with the lowest Chamfer distance averaged over all scenes. Additionally, as visualized in Fig. 2 (Row 1), our reconstructed mesh can exhibit smooth surfaces and correctly maintain the structure, e.g., the hole in the mouth. (2) Our method achieves the best reconstruction performance in 10 out of 15 scenes, indicating its general effectiveness across diverse scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "type": "text", + "content": "Generalization to more diverse scenes. To demonstrate the generalization capability on more diverse scenes with larger domain shifts compared to the training scenes, we further benchmark the mesh reconstruction quality on the NeRF-Synthetic dataset [41] and NSVF-Synthetic [37] under a full-view reconstruction setting, i.e., the depth of each source view from each scene's training set is estimated and fused into a mesh using TSDF fusion [45, 63]. As visualized in Fig. 2 (Rows 2-4), we observe that our Omni-Recon can deliver notably higher-quality meshes with better-maintained structures, fewer holes, and smoother surfaces compared to the strongest baselines ReTR [34] and VolRecon [55]. The decent generalization capability of our method is attributed to the proper decoupling and fusion of geometry and appearance features in our backbone. More visualizations and ablation studies, such as the contributions of each backbone component, are provided in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Rendering quality. We further benchmark the performance of generalizable novel view rendering on the test scenes from DTU. As shown in Tab. 2, our" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 465, + 358 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 465, + 358 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 465, + 358 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 465, + 358 + ], + "type": "image", + "image_path": "253e7b23dc30b8c24932e32989d56897591d241547df22e0ec08fea8de4aceb3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 362, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 407 + ], + "type": "text", + "content": "Fig. 2: Visualize the reconstructed mesh of our Omni-Recon and the two strongest baselines [34, 55]. Row 1: Scan65 from the test scenes of DTU; Rows 2-4: Scenes from NeRF-Synthetic [41] and NSVF-Synthetic [37], which present relatively challenging cases due to domain shifts w.r.t. the training scenes from DTU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 136, + 419, + 476, + 460 + ], + "blocks": [ + { + "bbox": [ + 134, + 408, + 479, + 418 + ], + "lines": [ + { + "bbox": [ + 134, + 408, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 134, + 408, + 479, + 418 + ], + "type": "text", + "content": "Table 2: Benchmark the quality of novel view rendering in terms of PSNR on DTU." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 419, + 476, + 460 + ], + "lines": [ + { + "bbox": [ + 136, + 419, + 476, + 460 + ], + "spans": [ + { + "bbox": [ + 136, + 419, + 476, + 460 + ], + "type": "table", + "html": "
MethodMean.243740556365698397105106110114118122
VolRecon [55]24.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]25.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Omni-Recon26.3224.7722.3323.9225.5626.3724.7525.1924.9424.9225.0628.3928.6328.0131.4930.5
", + "image_path": "9ff302180515419a10fe20e0b50cdca2ecaeee10f37c1cfedf4bacf84920425f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "content": "method achieves the highest PSNR, e.g., a " + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "inline_equation", + "content": "+0.73" + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "inline_equation", + "content": "+1.74" + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "content": " PSNR improvement over ReTR and VolRecon, respectively. As will be demonstrated in Sec. 5.4, this PSNR can be notably boosted with super-fast Nvdiffrast-based finetuning [31]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 467, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 467, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 467, + 532 + ], + "type": "text", + "content": "4 Omni-Recon: Enable Zero-shot Scene Understanding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": "To enable the understanding of various 3D scene properties, one approach is to jointly train the NeRF backbone on all target tasks. However, this approach is not scalable due to the diversity of scene properties, such as semantic segmentation, edge detection, and keypoint prediction, making it costly to cover new properties of interest. Therefore, it is highly desirable to enable zero-shot scene understanding for new tasks that were not seen during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 613, + 400, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 400, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 400, + 625 + ], + "type": "text", + "content": "4.1 Our Strategy: The Predict-then-Blend Strategy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Our hypothesis. We hypothesize that with accurate geometry and appearance estimation, the blending weights initially learned for radiance can be repurposed for other scene understanding tasks. In essence, just as blending source colors" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 474, + 403 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 474, + 403 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 474, + 403 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 474, + 403 + ], + "type": "image", + "image_path": "2499968a06e76450b64793bc71a72b1a6d2e791f72f49d34b6b6f2244b7a3f6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 405, + 480, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 405, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 480, + 429 + ], + "type": "text", + "content": "Fig. 3: Benchmark our predict-then-blend strategy with the render-then-predict baseline across different scene understanding tasks on Replica [61] and ScanNet [17]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "content": "yields radiance for each sampled 3D point, blending the properties of source views using the same weights can yield other properties. This hypothesis stems from the observation that scene appearance is closely related to other scene properties and thus regions with similar appearance are likely to share similar scene properties, such as semantics and edges. As such, the blending weights learned for appearance can also indicate the weighting for other scene properties." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "text", + "content": "Our method. Motivated by this hypothesis, our Omni-Recon employs a predict-then-blend strategy to enable zero-shot multitask scene understanding by lifting 2D monocular priors to 3D novel views. Specifically, considering the prevalence of 2D monocular vision models, our method first utilizes pretrained 2D models to generate predictions (e.g., logits in semantic segmentation) on each source view, denoted as " + }, + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "text", + "content": ". Then, the property of each sampled point can be obtained by blending their projected predictions on source views " + }, + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "text", + "content": ", reusing the RGB blending weights, i.e., " + }, + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}} = \\sum_{i=1}^{N} \\omega_i \\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 503, + 482, + 613 + ], + "type": "text", + "content": ". The final pixel-wise prediction can be derived using volumetric rendering similar to RGB reconstruction in Eq. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 623, + 432, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 432, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 432, + 636 + ], + "type": "text", + "content": "4.2 Evaluation: Zero-shot Multitask Scene Understanding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Setup. To evaluate the zero-shot scene understanding capabilities of our model, we adopt two indoor 3D scene datasets: Replica [61] and ScanNet [17], following" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "content": "the dataset splits in [93] and [36], respectively. Specifically, we focus on four scene understanding tasks: semantic segmentation, edge detection, and two keypoint detection tasks defined in [89]. For these tasks, the 2D monocular priors are provided by [51], [5], and [56], respectively. We employ mIoU for the semantic segmentation task and " + }, + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "content": " error for the other three tasks as default metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 481, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 481, + 223 + ], + "type": "text", + "content": "Baseline. An intuitive way to achieve zero-shot 3D scene understanding is to leverage existing 2D monocular models to predict the scene properties on top of the rendered images, a process referred to as a render-then-predict pipeline. We benchmark this strategy against our proposed predict-then-blend strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 298, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 298, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 298, + 331 + ], + "type": "text", + "content": "Zero-shot scene understanding benchmark. As shown in Tab. 3 and Fig. 3, we observe that (1) Our predict-then-blend strategy achieves decent scene understanding performance on unseen scenes and unseen tasks, which echoes our hypothesis in Sec. 4.1 that 2D monocular priors can be elevated to 3D novel views through" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 246, + 480, + 329 + ], + "blocks": [ + { + "bbox": [ + 305, + 224, + 481, + 245 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 481, + 245 + ], + "type": "text", + "content": "Table 3: Benchmark the two zero-shot scene understanding strategies across tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 246, + 480, + 329 + ], + "lines": [ + { + "bbox": [ + 307, + 246, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 480, + 329 + ], + "type": "table", + "html": "
DatasetMethod↑ Semantic↓ Edge↓ Key Point↓ Key Point 3D
Replica [61]Render-then -Predict15.640.04560.11010.0470
Predict-then -Blend (Ours)32.110.04120.07740.0176
ScanNet [17]Render-then -Predict41.320.04710.05680.0412
Predict-then -Blend (Ours)61.110.04340.04240.0197
", + "image_path": "a5a803a16789e034a88f4166eb0abbf59d457244b624b5dce79c7549634176b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 464 + ], + "type": "text", + "content": "blending weight reuse, thanks to the high correlation between scene appearance and other properties; (2) Compared to the render-then-predict scheme, our method achieves notably higher quantitative accuracy, e.g., a " + }, + { + "bbox": [ + 130, + 331, + 482, + 464 + ], + "type": "inline_equation", + "content": "19.79\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 464 + ], + "type": "text", + "content": " higher mIoU on ScanNet, and consistently better visual effects across all tasks. We attribute this improvement to two aspects: (1) Our predict-then-blend strategy can avoid feeding noisy inputs caused by rendering errors to the 2D monocular models, thus reducing reliance on the robustness of monocular models and ensuring more accurate monocular priors; (2) The multiview information gathered by our strategy can contribute to a more accurate understanding of objects with limited observations under a monocular setting. More experiments on other vision models like CLIP-LSeg [32] are in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 468, + 463, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 463, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 463, + 483 + ], + "type": "text", + "content": "5 Omni-Recon: Adaptability to Downstream 3D Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 488, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 481, + 548 + ], + "type": "text", + "content": "In addition to the generalization capabilities, we further demonstrate the adaptability of our Omni-Recon by extending it to three different types of downstream 3D applications. These demonstrations highlight the potentially broad utility of image-based rendering pipelines, which may have been overlooked by previous works, and can be generalized to a wider range of real-world 3D applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 553, + 424, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 553, + 424, + 566 + ], + "spans": [ + { + "bbox": [ + 131, + 553, + 424, + 566 + ], + "type": "text", + "content": "5.1 Application 1: A New Real-time Rendering Pipeline" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 481, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 618 + ], + "type": "text", + "content": "To rapidly enable real-time rendering using Omni-Recon's pretrained backbone, we build a real-time rendering pipeline based on Nvdiffrast [31]. This involves baking the complex geometry branch into meshes using TSDF fusion [45,63] as discussed in Sec. 3.5, and using the lightweight appearance branch as a Shader." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "content": "Our real-time rendering pipeline. As illustrated in Fig. 1 (bottom left), for any novel view of a new scene, we first transform its camera pose to clip space as required by Nvdiffrast [31]. Then, we perform a rasterization process to obtain the intersection point of each camera ray with the extracted mesh. Next, these" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "intersection points are input into the shader to predict the color by blending their projected colors on the source views as described in Sec. 3.2. Thanks to the blending scheme, where valid projections can provide strong appearance clues, the rendering quality is often satisfactory even without finetuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "type": "text", + "content": "Joint mesh and Shader finetuning. To further enhance the rendering quality, we can jointly finetune the mesh and the Shader in a gradient-based manner using differentiable rendering enabled by Nvdiffrast [31]. Specifically, we finetune the location of each vertex in the mesh and the parameters of the lightweight Shader. Additionally, we periodically prune redundant mesh faces that cannot intersect with any camera rays in one training epoch. More details are provided in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 255, + 479, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 479, + 268 + ], + "type": "text", + "content": "5.2 Application 2: Parameter-Efficient Tuning for 3D Understanding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 480, + 380 + ], + "type": "text", + "content": "We explore the possibility of finetuning Omni-Recon for various downstream 3D tasks, utilizing a series of 3D scene understanding tasks as demonstrations. Specifically, we conduct PET [19] on top of Omni-Recon's pretrained backbone by incorporating low-rank adapters [26] into the linear layers within the transformer modules introduced in Sec. 3.3, and only tuning these adapters while keeping the pretrained backbone frozen. It is worth noting that when evaluating the performance of the tuned model, we assess it on new scenes that have never been encountered during both training and tuning, representing a generalizable scene understanding scenario [36]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 387, + 387, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 387, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 387, + 399 + ], + "type": "text", + "content": "5.3 Application 3: Text-guided 3D Scene Editing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 480, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 522 + ], + "type": "text", + "content": "We target text-guided 3D scene editing, which involves editing a 3D scene based on textual instructions. To achieve this, our Omni-Recon capitalizes on the nature of image-based rendering, allowing edits made to 2D source views to be propagated to 3D novel views. Specifically, leveraging the availability of powerful 2D diffusion models, we use publicly available ones to edit the 2D source views. However, doing so naively may result in inconsistent source views, as the same textual instruction may lead to different visualizations, thus degrading reconstruction quality. Inspired by [23], which iteratively edits the training images and retransmits a new NeRF to achieve scene editing, we address the inconsistent editing issue by proposing an iterative editing and reconstruction pipeline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 523, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 480, + 666 + ], + "type": "text", + "content": "The iterative editing and reconstruction pipeline. To ensure consistent 3D editing, we iterate between source view editing and reconstruction as follows: First, we edit each source view using the diffusion model InstructPix2Pix [4], which is guided by both textual instructions and the original image to maintain image fidelity. While the edited source views can faithfully follow the textual instructions, they may exhibit inconsistencies in the edited region. Next, we reconstruct each source view by conditioning on their nearby source views, as described in Sec. 3.2. This process improves the 3D consistency among source views at the expense of reduced image fidelity and less adherence to instructions. We then iterate through the editing and reconstruction steps to progressively enhance image fidelity, 3D consistency, and adherence to instructions. The resulting source views are finally used to reconstruct the entire scene." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 149, + 479, + 232 + ], + "blocks": [ + { + "bbox": [ + 131, + 110, + 482, + 144 + ], + "lines": [ + { + "bbox": [ + 131, + 110, + 482, + 144 + ], + "spans": [ + { + "bbox": [ + 131, + 110, + 482, + 144 + ], + "type": "text", + "content": "Table 4: Benchmark the rendering speed and rendering quality of our real-time rendering pipeline and the baselines [34,55] on DTU. The best PSNR is in bold, and the shortest finetuning time that our pipelines surpass the strongest baseline is underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 149, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 135, + 149, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 149, + 479, + 232 + ], + "type": "table", + "html": "
MethodFPSMean243740556365698397105106110114118122
VolRecon [55]0.02924.5822.3320.5921.5323.7224.223.6524.4722.7723.5422.6226.8927.4425.7630.1429.19
ReTR [34]0.02425.5924.3221.8423.424.5626.3124.524.6324.324.5823.8527.8427.9726.7630.0328.96
Ours w/o ft.71.3(40.82)22.9620.1219.7122.2722.7824.5521.7721.5126.7222.3324.4922.5222.9324.6823.8424.12
Ours (ft. 10s)25.6821.4221.6824.0624.1228.1924.1023.9531.6524.4128.1525.6325.8526.1526.8226.89
Ours (ft. 20s)27.2122.6322.9225.1225.4230.0325.9526.1633.1926.2230.3627.4427.0426.9329.1529.65
Ours (ft. 30s)27.7823.223.2625.5425.730.5926.8326.9633.6626.4730.7328.1427.7027.130.1730.65
Ours (ft. 1min)28.3424.6924.1125.7626.0530.9327.6627.4933.6827.0730.9728.5128.5427.3531.1731.54
Ours (ft. 3min)28.9525.224.3225.9426.1632.1528.9927.8834.9427.3531.6228.9328.9727.5631.7032.49
Ours (ft. 5min)29.0225.3424.3625.6326.2132.1629.3327.8134.9427.3231.7429.0429.0527.6931.7432.89
", + "image_path": "7959c899f57713872cbe4f7112f394692f483946af3bdf56dfa96c6ffd985863.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 242, + 466, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 466, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 466, + 253 + ], + "type": "text", + "content": "5.4 Evaluation: Effectiveness on Different Downstream 3D Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 317 + ], + "type": "text", + "content": "Real-time rendering. We evaluate our real-time rendering pipeline in Sec. 5.1 in terms of both rendering quality and rendering speed, measured on an NVIDIA A5000 GPU, using the DTU dataset. Specifically, we finetune the scene meshes, extracted by TSDF fusion as mentioned in Sec. 3.5, and the Shader in no more than 5 minutes and record the rendering quality at different time steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 474 + ], + "type": "text", + "content": "Observations and analysis: As shown in Tab. 4, we can observe that (1) Without any finetuning, the rendering pipeline already exhibits certain rendering capabilities, despite the model backbone being trained using volumetric rendering with more than one sampled point; (2) Our rendering pipeline can be swiftly finetuned to significantly boost the PSNR. For instance, with just a 10-second finetuning, our rendering pipeline can match the average rendering quality of the strongest baseline, ReTR [34]; With 1-minute/2-minute finetuning, it can surpass ReTR by a +2.75 and +3.36 PSNR improvement on average, respectively; (3) Our rendering pipeline enables real-time rendering with 71.3 FPS, i.e., a >2458× speed-up over the baselines. Here, the assumption is that the feature extraction of each source view from a target scene can be done in a one-time effort and does not account for throughput. If we still consider this feature extraction latency for each rendering, the resulting 40.82 FPS still meets the real-time requirement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 483, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 483, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 483, + 498 + ], + "type": "text", + "content": "These experiments indicate that given a new scene, our Omni-Recon is promising to enable both instant and accurate reconstruction and real-time rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 498, + 281, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 281, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 281, + 593 + ], + "type": "text", + "content": "PET for 3D Understanding. To evaluate the adaptability of our Omni-Recon, we apply PET to it for various scene understanding tasks on ScanNet [17]. Specifically, we finetune a LoRA adaptor [26] for each task on a set of training scenes and evaluate on" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 290, + 522, + 479, + 586 + ], + "blocks": [ + { + "bbox": [ + 287, + 499, + 481, + 521 + ], + "lines": [ + { + "bbox": [ + 287, + 499, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 287, + 499, + 481, + 521 + ], + "type": "text", + "content": "Table 5: Evaluate the effectiveness of PET on Omni-Recon and benchmark with SRay [36]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 522, + 479, + 586 + ], + "lines": [ + { + "bbox": [ + 290, + 522, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 290, + 522, + 479, + 586 + ], + "type": "table", + "html": "
Method↑ Sem. mIoU↑ Sem. Total Acc↑ Sem. Avg Acc↓ Edge↓ Key Point↓ Key Point 3D
SRay [36]57.1578.2462.55---
Ours (Zero-shot)61.1180.8069.170.04340.04240.0197
Ours (PET + Zero-shot)62.3581.8469.890.03420.02070.0132
", + "image_path": "c63928ecbcfdd75abcb62989ff79d2000c5ab9e24cabcd3753af41f87d89e7bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": "non-overlapping test scenes, representing a generalizable 3D understanding setting, following the dataset split described in [36]. We benchmark against the SOTA generalizable semantic segmentation work, SRay [36]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": "Observations and analysis: As shown in Tab. 5, we can observe that (1) After PET, our model outperforms SRay in terms of generalizable semantic segmentation with a " + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "inline_equation", + "content": "+5.20\\%" + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": " mIoU and a " + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "inline_equation", + "content": "+7.34\\%" + }, + { + "bbox": [ + 130, + 630, + 481, + 667 + ], + "type": "text", + "content": " accuracy improvement; (2) Despite" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 440, + 274 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 440, + 274 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 440, + 274 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 440, + 274 + ], + "type": "image", + "image_path": "6c7be7d29fa88ab9cad67e961ec0abd8829608f9dfec4b17ab9ee1cd19f8a98c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 275, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 132, + 275, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 480, + 286 + ], + "type": "text", + "content": "Fig. 4: Visualize the scene understanding performance on hard cases before/after PET." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 153, + 289, + 462, + 411 + ], + "blocks": [ + { + "bbox": [ + 153, + 289, + 462, + 411 + ], + "lines": [ + { + "bbox": [ + 153, + 289, + 462, + 411 + ], + "spans": [ + { + "bbox": [ + 153, + 289, + 462, + 411 + ], + "type": "image", + "image_path": "33fe245c26fdbe0ab18443afadf59a49d34d5e8ccb4aa221368daa5744f241c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 415, + 468, + 426 + ], + "lines": [ + { + "bbox": [ + 144, + 415, + 468, + 426 + ], + "spans": [ + { + "bbox": [ + 144, + 415, + 468, + 426 + ], + "type": "text", + "content": "Fig. 5: Three text-guided scene editing examples using our pipeline in Sec. 5.3." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 481, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 465 + ], + "type": "text", + "content": "having limited trainable parameters, PET can enhance the scene understanding performance across tasks, especially on challenging cases visualized in Fig. 4, indicating the adaptability of Omni-Recon to diverse downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 466, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 481, + 491 + ], + "type": "text", + "content": "Text-guided scene editing. We visualize the edited examples in Fig. 5, showcasing the ability to ensure both instruction-following and 3D consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Motivated by the significant demand for supporting various functionalities required by different 3D applications through a unified NeRF model, our work aims to develop a general-purpose NeRF capable of handling a broad spectrum of 3D tasks. We propose a framework called Omni-Recon, which features a general-purpose NeRF model using image-based rendering with two separate branches, and demonstrate its integration with various 3D tasks. Specifically, our Omni-Recon can achieve SOTA generalizable 3D reconstruction quality, enable zero-shot multitask scene understanding, achieve SOTA scene understanding performance and real-time rendering after rapid adaptation, and support scene editing. Omni-Recon underscores the potential of image-based rendering pipelines in diverse real-world 3D applications, which may have been underestimated by prior studies, thus sparking advancements in future rendering pipelines." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "content": "The work is supported by the National Science Foundation (NSF) through the SCH program (Award number: 1838873) and CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. International Journal of Computer Vision 120, 153-168 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 306, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 480, + 350 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 350, + 480, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 382 + ], + "type": "text", + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "text", + "content": "5. Canny, J.: A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence (6), 679-698 (1986)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "text", + "content": "6. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "text", + "content": "7. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis. CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "text", + "content": "8. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 480, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 502 + ], + "type": "text", + "content": "9. Chen, A., Xu, Z., Wei, X., Tang, S., Su, H., Geiger, A.: Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 502, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 502, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 134, + 502, + 480, + 545 + ], + "type": "text", + "content": "10. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 545, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 480, + 589 + ], + "type": "text", + "content": "11. Chen, T., Wang, P., Fan, Z., Wang, Z.: Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15191-15202 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "text", + "content": "12. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "13. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "text", + "content": "14. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "text", + "content": "15. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3193-3204 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "16. Croce, F., Hein, M.: Sparse and imperceivable adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4724-4732 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "17. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "18. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12882-12891 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "19. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence 5(3), 220-235 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "20. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "21. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhu, L., Zhou, X., Geiger, A., Liao, Y.: Panoptic nef: 3d-to-2d label transfer for panoptic urban scene segmentation. arXiv preprint arXiv:2203.15224 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "text", + "content": "22. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "text", + "content": "23. Haque, A., Tancik, M., Efros, A.A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "24. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "25. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "27. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "28. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5885-5894 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "29. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "30. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "type": "text", + "content": "31. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "32. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "33. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "34. Liang, Y., He, H., Chen, Y.: Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "35. Lindell, D.B., Martel, J.N., Wetzstein, G.: Autoint: Automatic integration for fast neural rendering. CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "36. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "text", + "content": "37. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "39. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. SIGGRAPH (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "text", + "content": "40. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "text", + "content": "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "43. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 556, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 599 + ], + "type": "text", + "content": "44. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16. pp. 414-431. Springer (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "45. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "46. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5480-5490 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "47. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "48. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "text", + "content": "49. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "51. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. ArXiv preprint (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "52. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: DeRF: Decomposed radiance fields. CVPR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "53. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 380 + ], + "type": "text", + "content": "54. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "text", + "content": "55. Ren, Y., Zhang, T., Pollefeys, M., Süsstrunk, S., Wang, F.: Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16685-16695 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "56. Sax, A., Emi, B., Zamir, A.R., Guibas, L.J., Savarese, S., Malik, J.: Mid-level visual representations improve generalization and sample efficiency for learning visuomotor policies. (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "text", + "content": "57. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 501-518. Springer (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "58. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: GRAF: Generative radiance fields for 3D-aware image synthesis. NeurIPS (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "59. Schwarz, K., Sauer, A., Niemeyer, M., Liao, Y., Geiger, A.: Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. arXiv preprint arXiv:2206.07695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "text", + "content": "60. Seitz, S.M., Dyer, C.R.: Photorealistic scene reconstruction by voxel coloring. IJCV (1999)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "61. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "62. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Sun, J., Xie, Y., Chen, L., Zhou, X., Bao, H.: Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15598-15607 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "64. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "65. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "66. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "67. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "68. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15182-15192 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "69. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. arXiv preprint arXiv:2211.11738 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "70. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12922–12931 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "71. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "72. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "73. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "74. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "75. Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z., et al.: Is attention all nerf needs? arXiv preprint arXiv:2207.13298 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "76. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 567 + ], + "type": "text", + "content": "77. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4690-4699 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "text", + "content": "78. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "79. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 106-122. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "80. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni-Recon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 522 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "81. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 " + }, + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "inline_equation", + "content": "\\{\\backslash \\mathrm{deg}\\}" + }, + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": " views. arXiv preprint arXiv:2211.16431 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "82. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "83. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "84. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "85. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "text", + "content": "86. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "87. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "88. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "89. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "text", + "content": "90. Zhang, M., Zheng, S., Bao, Z., Hebert, M., Wang, Y.X.: Beyond rgb: Scene-property synthesis with neural radiance fields. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 795-805 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "text", + "content": "91. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5449-5458 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "92. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "93. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Y. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_content_list.json b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..08fdf803f804fb1f491450114979b29efc740811 --- /dev/null +++ b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_content_list.json @@ -0,0 +1,1959 @@ +[ + { + "type": "text", + "text": "Omni6DPose: A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking", + "text_level": 1, + "bbox": [ + 256, + 140, + 750, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiyao Zhang $^{1,2,3*}$ , Weiyao Huang $^{1*}$ , Bo Peng $^{1*}$ , Mingdong Wu $^{1,2,3}$ , Fei Hu $^{1,2,3}$ , Zijian Chen $^{4}$ , Bo Zhao $^{5,6}$ , and Hao Dong $^{1,2,3\\dagger}$", + "bbox": [ + 217, + 233, + 782, + 267 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Center on Frontiers of Computing Studies, School of Computer Science, Peking University $^{2}$ PKU-Agibot Lab, School of Computer Science, Peking University", + "$^{3}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University $^{4}$ Waseda University $^{5}$ Shanghai Jiao Tong University $^{6}$ BAAI {jiyaozhang, sshwy, bo.peng, hofee}@stu.pku.edu.cn zijianchenwaseda@akane.waseda.jp bo.zhao@sjtu.edu.cn {wmingd, hao.dong}@pku.edu.cn" + ], + "bbox": [ + 222, + 276, + 777, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://jiyao06.github.io/Omni6DPose/", + "bbox": [ + 356, + 345, + 643, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 6D object pose estimation is crucial in the field of computer vision. However, it suffers from a significant lack of large-scale and diverse datasets, impeding comprehensive model evaluation and curtailing downstream applications. To address these issues, this paper introduces Omni6DPose, a substantial benchmark featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is divided into three main components: ROPE (Real 6D Object Pose Estimation Dataset), which includes 332K images annotated with over 1.5M annotations across 581 instances in 149 categories; SOPE(Simulated 6D Object Pose Estimation Dataset), a simulated training set created by mixed reality and physics-based depth simulation; and PAM(Pose Aligned 3D Models), the manually aligned real scanned objects used in ROPE and SOPE. Omni6DPose is inherently challenging due to the substantial variations and ambiguities. To address this issue, we introduce GenPose++, an enhanced version of the SOTA category-level 6D object pose estimation framework, incorporating two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation. Moreover, we provide a comprehensive benchmarking analysis to evaluate the performance of previous methods on this new large-scale dataset in the realms of 6D object pose estimation and pose tracking.", + "bbox": [ + 259, + 372, + 738, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Benchmark $\\cdot$ object pose estimation $\\cdot$ object pose tracking", + "bbox": [ + 261, + 652, + 733, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 691, + 375, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6D object pose estimation [16,28,45] and pose tracking [15,21] from single images is an essential task in computer vision, holding immense potential for applications in robotics [1] and augmented reality/virtual reality (AR/VR) [19]. Over recent decades, the domain has experienced significant advancements, primarily dominated by data-driven learning approaches. Analogous to the pivotal role of data in learning-based 2D foundation tasks, high-quality, comprehensive datasets are paramount in the context of 6D object pose estimation and tracking.", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: equal contributions, †: corresponding author", + "bbox": [ + 230, + 825, + 545, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e30f9fb1ef9101e65707ac7f168841fdf416464f146a4f76a392dff6c0d76ea.jpg", + "image_caption": [ + "Fig. 1: We introduce Omni6DPose for universal 6D object pose estimation and tracking, which comprises three key components: (1) ROPE (right), a large-scale real-world dataset for evaluation. (2) SOPE (left), a vast synthetic dataset generated through mixed reality and physics-based simulation for training. (3) PAM (center), a collection of pose-aligned 3D object models. Omni6DPose is featured by its diversity in object categories, large scale, and variety in object materials." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Today 6D object pose estimation is studied under two lenses: instance-level and category-level. In instance-level settings, datasets such as Linemod [10], YCB-Video [36], and T-LESS [11] have gained widespread acceptance as benchmarks. These datasets are distinguished by their focus on detailed, individual object instances, thereby enabling algorithms to precisely learn and predict the poses of specific items. On the other hand, category-level pose estimation emphasizes generalization across different items within a particular object category. The NOCS [28] dataset stands out as the most widely used in the category-level object pose estimation field, providing a simulated dataset for training and a small-scale real-world dataset for evaluation. Despite their contributions to advancing the field, these datasets present limitations due to their small scale in terms of instances or categories. This results in two significant challenges:", + "bbox": [ + 212, + 440, + 787, + 623 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. It hampers comprehensive evaluation of different models' performance, limiting the development of research in this field.", + "2. It restricts the applicability of research findings across diverse domains, due to the limited variety of object instances or categories represented." + ], + "bbox": [ + 220, + 632, + 784, + 693 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges and drive advancements in this field, this paper introduces Omni6DPose, a universal 6D object pose estimation dataset featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is segmented into three principal components: (1) ROPE (Real 6D Object Pose Estimation Dataset), which encompasses $332\\mathrm{K}$ images annotated with over 1.5M annotations across 581 instances in 149 categories; (2) SOPE (Simulated 6D Object Pose Estimation Dataset), comprising $475\\mathrm{K}$ images generated in a mixed reality setting with depth simulation, furnished with over 5M annotations spanning 4162 instances in the same 149", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 113, + 761, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1a60111597c270923e5a9fe2d9337ff4c25a52767811667ea1190b636c2b590f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 146, + 331, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e30d062643149c774e69ec1aee782198766bc6e4587e0098a9fa42d0a9e1cac6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 195, + 331, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43275bc3b8beb29183c70ae36d9adae17d40d064a09a49bb097a67b6f133b781.jpg", + "image_caption": [ + "Fig. 2: ROPE dataset visualization. In the figure, bounding boxes are colored according to the coordinates in the object's coordinate system." + ], + "image_footnote": [], + "bbox": [ + 222, + 244, + 331, + 294 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44ff07bbb3552f46ea54d5355b86627c6238da0b82a64dc8ea16bf8033463a1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 146, + 444, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6468a5f50a56759c8b1bbf03e1aa1b915eb71628c75cdcf150bdc344d9cccb97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 196, + 444, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/57fea1551a53410447966c9d8b602271c459fccd8967a553a84fabb67a4623ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 244, + 444, + 294 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b8808cb50f283d7c2f4884717cd265866e36cf74011e315d21131af851534fce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 146, + 555, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/172fd19562b74c51f5ab7fa2c961c534695d45e685e2ea5611bb5ebf4f21b840.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 196, + 555, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b423517e481a3550e627b5e8ada15cd45c58163646cad585cb6f5cab3e6e0f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 244, + 555, + 294 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cfc55b1d4839de861b119a85310341fb2d3146b2b9b25b746da1388d41fadb75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 146, + 669, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2996098182c76da3d4a14b8593baa467e1c4312c27a59b8b6e9fdb2fbb194d5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 196, + 669, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f89108a76ee285ad48ac784e5c47e30101301b38ef746d7f5687b12476d37d37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 244, + 669, + 294 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9b4943baaa6d025c9201974d559431e8719b9631d0841cc21f6dd49a6c98717b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 146, + 779, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7511dc9f5d88aada149176f9573656101fa178c98e0e1b8986d4678451f1b6cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 196, + 779, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5eeb4a4d92a5496ce3e870202d7eb7a7e2a8f2723c6af499c5d3209a846479de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 244, + 779, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "categories. The mixed reality bridges the semantic sim2real gap, while the depth sensor simulation closes the geometric sim2real gap; (3) PAM(Pose Aligned 3D Models), the manually aligned, real scanned objects utilized in both ROPE and SOPE, enabling the generation of diverse downstream task data.", + "bbox": [ + 212, + 337, + 782, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Omni6DPose poses inherent challenges due to its considerable variations, diverse materials, and inherent ambiguities, which reflect the complexities encountered in real-world applications. Figure 2 illustrates some examples from ROPE. To tackle these issues, we introduce GenPose++, which incorporates GenPose [43] with two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation, tailored specifically to the nuances of the Omni6DPose in question. Furthermore, as a Universal 6D object pose estimation dataset, this paper also offers a comprehensive benchmarking analysis to assess the performance of existing methods on category-level 6D object pose estimation and pose tracking. We summarize our contributions as follows:", + "bbox": [ + 212, + 398, + 784, + 547 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present Omni6DPose, a comprehensive 6D object pose estimation dataset with extensive categories, instance diversity, and material variety.", + "2. We propose a real data collection pipeline and a simulation framework for generating synthetic data with low semantic and geometry sim2real gaps.", + "3. We introduce GenPose++ for category-level 6D object pose estimation and tracking, demonstrating SOTA performance on Omni6DPose." + ], + "bbox": [ + 220, + 554, + 795, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 214, + 655, + 382, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For 6D object pose estimation, there are two main branches: instance-level and category-level. Instance-level estimation is tested on seen objects or on unseen objects with reference, while category-level estimation is tested on unseen instances of known categories without CAD models. In this section, we review and compare existing datasets to our large-scale category-level dataset and review relevant algorithms for category-level pose estimation and tracking.", + "bbox": [ + 212, + 676, + 784, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 6D Object Pose Estimation Datasets", + "text_level": 1, + "bbox": [ + 214, + 773, + 563, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the outlined branches of 6D object pose estimation, we have reviewed datasets corresponding to both instance-level and category-level 6D object pose estimation. A comparative analysis of these datasets is provided in Table 1.", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/960f233612f00747db8565b437de6da3ff9cbceec3543f6b71811531f6c0578f.jpg", + "table_caption": [ + "Table 1: This table compares datasets for 6D object pose estimation, focusing on object category count, reality of the data, data modalities (RGB, Depth, IR), and object attributes such as quantity, CAD model availability, and inclusion of transparent and specular objects. It also details video characteristics by number and marker presence, along with image and annotation counts. 'Wild6D' refers specifically to the test split of the Wild6D dataset, as the training data does not provide annotations. The symbol '-' indicates the absence of a particular feature within the dataset." + ], + "table_footnote": [], + "table_body": "
DatasetCat. RealModalityObjectMarker-free Vid.Img.Anno.
RGBDepth IRIRNum. CAD Trans. Spec.
CAMERA [28]6X✓✓X1085✓XX✓-300K
SOPE(Ours)149X✓✓✓4162✓✓✓✓-475K
YCB-Video [36]-✓✓✓X21✓XX✓92133K
T-LESS [11]-✓✓✓X30✓XXX2048K
Linemod [10]-✓✓✓X15✓XXX-18K
StereoOBJ-1M [17]-✓✓XX18✓✓✓X182393K
REAL275 [28]6✓✓✓X42✓XXX188K
PhoCal [30]8✓✓✓X60✓✓✓✓243.9K
HouseCat6D [12]10✓✓✓X194✓✓✓✓4123.5K
Wild6D* [9]5✓✓✓X162X(✓)X✓48610K
PACE [41]44✓✓✓X576✓XX✓30055K
ROPE(Ours)149✓✓✓✓581✓✓✓✓363332K
", + "bbox": [ + 225, + 242, + 776, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-level 6D object pose estimation dataset. LineMod [10] is one of the most used datasets, providing non-temporal RGB-D images and ground truth pose annotations. YCB-Video [36] provides RGB-D videos and annotations, enabling pose-tracking approaches. T-LESS [11] features texture-less objects with symmetries and mutual similarities. StereoOBJ-1M [17] achieves a leap in dataset scale and features transparent and reflective objects. While these datasets are extensive in terms of image and annotation count, they are limited in the diversity of instances they cover. For example, StereoOBJ-1M comprises 339K frames and 1.5M annotations, yet it includes only 18 unique object instances.", + "bbox": [ + 212, + 440, + 787, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category-level 6D object pose estimation dataset. NOCS [28] provides the first benchmark in category-level pose estimation, while Wild6D [9] addresses the scalability issue of datasets by leveraging unlabeled and synthetic data. PhoCal [30] focuses on photometrically challenging objects and HouseCat6D [12] offers diverse scenes, viewpoints, and grasping annotations. However, these datasets cover only a limited number of categories, even the most extensive dataset, HouseCat6D, includes merely ten categories. PACE [41], a concurrent work, focusing on object pose estimation in complex scenes. In contrast, our datasets, SOPE and ROPE, set new benchmarks by offering the widest range of categories and featuring objects with diverse materials, thereby enhancing dataset diversity and realism for pose estimation research.", + "bbox": [ + 212, + 582, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 6D Object Pose Estimation and Tracking Algorithm", + "text_level": 1, + "bbox": [ + 214, + 757, + 692, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category-level 6D object pose estimation. Category-level 6D object pose estimation [28,39,40] aims to estimate unseen instance poses within the same category. NOCS [28] introduces a normalized object coordinate space for pose prediction without CAD models, while SPD [29] and SGPA [3] utilize category-level", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 113, + 761, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prriors for enhanced estimation. HS-Pose [44] extends 3D-GC [14] for improved feature extraction from point clouds. IST-Net [16] transforms features between camera and world spaces implicitly, without relying on 3D priors, surpassing previous methods. However, these techniques, mainly regression-based, require ad-hoc designs for symmetric objects. GenPose [43] addresses it by generatively modeling the pose distribution, eliminating the need for symmetry considerations. Yet, GenPose neglects RGB semantic information, which is increasingly vital as object category scales grow. Additionally, its energy-based aggregation algorithm fails with discontinuous multimodal distributions. We introduce GenPose++, which incorporates a 2D foundation model to leverage RGB semantics, improving generalization, and introduces an aggregation module to handle discrete multimodal distributions, addressing the limitations of GenPose.", + "bbox": [ + 212, + 145, + 787, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6D object pose tracking. This paper is situated within the domain of category-level 6D object pose tracking and model-free object tracking. BundleTrack [31] pioneers model-free tracking by leveraging multi-view feature detection for tracking unseen objects without 3D models. CAPTRA [32] enhances articulated pose tracking through recursive updates for better temporal consistency. CATRE [18] aligns partially observed point clouds to abstract shape priors for relative transformations and pose estimation. GenPose, adopting a generative approach, effectively addresses the challenge of pose ambiguities in symmetric objects. Together, these approaches underscore the evolving landscape of object pose tracking, highlighting both the progress and the diversity of strategies being explored.", + "bbox": [ + 212, + 335, + 789, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Omni6DPose Dataset", + "text_level": 1, + "bbox": [ + 214, + 497, + 460, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This paper introduces a rich variety of object categories, a large-scale, and materially diverse dataset for real 6D object pose estimation, named ROPE. And, a simulated dataset, SOPE, synthesizing with mixed reality and featuring depth simulation, is provided for training. Section 3.1 will discuss the collection and alignment of 3D objects. Section 3.2 will cover the acquisition and labeling of the ROPE dataset. Section 3.3 will explain the generation of the SOPE dataset.", + "bbox": [ + 212, + 515, + 787, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 3D Object Collection and Alignment", + "text_level": 1, + "bbox": [ + 214, + 628, + 566, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Universal 6D object pose estimation relies on a comprehensive set of objects. We selected 149 categories of everyday objects, all reconstructed with high-precision scanners, and categorized them into two sets: SOPE for simulated data and ROPE for real-world scenes. SOPE primarily includes objects from sources like OmniObject3D [35], PhoCal [30], and GoogleScan [7], alongside a subset from our scans, totaling 5,000 instances. ROPE consists of 580 instances we reconstructed using industrial scanners. Importantly, while most SOPE objects are from public datasets, manual category-level pose alignment is necessary. For object reconstruction, as shown in Figure 3, we use two professional scanners, EinScan H2 $^{7}$ and Revopoint POP 3 $^{8}$ for objects in different scales. The scan", + "bbox": [ + 212, + 651, + 787, + 803 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{7}$ https://www.einscan.com/", + "bbox": [ + 217, + 810, + 424, + 824 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "8 https://www_REVopoint3d.com/", + "bbox": [ + 222, + 825, + 452, + 839 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee72ddeb9dd614b8107981b5074fea0e9e9d96b84d2dc124a17f7cf47d4d842b.jpg", + "image_caption": [ + "Fig. 3: ROPE dataset collection and annotation. (1) Object scanning, where high-precision industrial scanners are used to acquire the CAD models of objects; (2) Object canonicalization, involving the alignment of each object category to the canonical space; (3) Video capture, capturing video sequences in varied scenarios with a depth camera; and (4) Pose annotation, calculating camera poses through Structure from Motion (SFM), further utilizing Farthest Point Sampling (FPS) to select keyframes for keypoint annotation, and performing bundle adjustment to derive initial object pose values, which are then manually refined to obtain more precise annotations." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 465, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fdeea2904a14f8872961b09ecaedf36c35b31eb8abc2fe5c48f6bd60ff54a8c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 143, + 785, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ning time depends on object features: it took about 15 minutes for small, simple, Lambertian items like a mouse, and up to an hour for complex, large, or non-Lambertian items like a transparent mug. Finally, we constructed a specialized annotation tool for manually aligning objects in the same category to the category-level canonical space, with each alignment taking roughly one minute.", + "bbox": [ + 212, + 405, + 782, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 ROPE Acquisition and Annotation", + "text_level": 1, + "bbox": [ + 215, + 489, + 549, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ROPE dataset was systematically acquired utilizing the RealSense D415 imaging device, encompassing scenarios with 2 to 6 distinct objects and video sequences extending from 762 to 1,349 frames. The integrity and utility of the dataset are underpinned by the precision of object pose annotations, which present notable challenges, chiefly:", + "bbox": [ + 212, + 513, + 784, + 590 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Derivation of relative camera poses, denoted as $T_{c} = \\{(R_{i},t_{i})\\}_{i = 1}^{n}$ , where each pair $(R_i,t_i)\\in \\mathrm{SE}(3)$ , signifying the transformation from the camera space to the world space for the $i^{th}$ frame, with $n$ symbolizing the aggregate frame count within the video sequence.", + "2. Procurement of high-accuracy object poses, represented as $T_{o} = \\{(R,t)\\}$ , where the pair $(R,t) \\in \\mathrm{SE}(3)$ , delineating the transformation from the object space the camera space for any selected frame." + ], + "bbox": [ + 220, + 601, + 785, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With $T_{c}$ and $T_{o}$ known, it is possible to automate the generation of all annotations within the dataset. Addressing these challenges, we propose a marker-free annotation system. Previous datasets for calculating relative camera pose rely on markers, like NOCS [28], or external robot arms for indirect calculation, such as PhoCal [30], limiting scene diversity. As shown in Figure 3, to enable marker-free annotation in open scenes, we consider it a structure-from-motion (SfM) problem, utilizing intrinsic scene features to optimize camera poses $T_{c}$ using bundle adjustment techniques. This approach aims to solve the optimization problem:", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 113, + 761, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee6825a0c6cd87d43cf531eb473786b2d5b4fc4606b3b301b1180875f4d4c284.jpg", + "image_caption": [ + "Fig. 4: SOPE synthesis, utilizing mixed reality to bridge the RGB sim2real gap and physical-based depth sensor simulation to minimize the geometric sim2real gap." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 455, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6208e949c4c1e03ff6fa2b7e8f0a535277ba4674ccd460c5d24793674b3bd92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 143, + 781, + 253 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {T _ {c}} \\sum_ {i = 1} ^ {n} \\sum_ {j \\in \\mathcal {P} _ {i}} \\| \\pi \\left(R _ {i} X _ {j} + t _ {i}\\right) - x _ {i j} \\| ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 303, + 784, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pi$ denotes the camera projection function, $X_{j}$ represents the 3D points in the world space, and $x_{ij}$ corresponds to the 2D projection of $X_{j}$ in the $i^{th}$ camera frame, with $\\mathcal{P}_i$ being the set of point correspondences in frame $i$ .", + "bbox": [ + 212, + 352, + 782, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For object pose $T_{o}$ annotations, previous methods consider only a single frame, leading to inaccuracies due to the lack of multi-viewpoint constraints. To overcome this, we introduce a two-stage object pose annotation process: in the first stage, keyframes are sampled from SfM results using Farthest Point Sampling (FPS), and 2D-3D keypoint pairs on these keyframes are annotated, providing initial object pose by minimizing the reprojection error:", + "bbox": [ + 212, + 398, + 782, + 488 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {R, t} \\sum_ {k \\in \\mathcal {K}} \\| \\pi \\left(R X _ {k} + t\\right) - x _ {k} \\| ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 502, + 784, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $X_{k}$ are 3D keypoints of object, $x_{k}$ are corresponding 2D annotations in the image, and $\\mathcal{K}$ is the set of all keypoint correspondences. The initial poses are manually fine-tuned based on the object's projection across all keyframes.", + "bbox": [ + 212, + 542, + 782, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 SOPE Synthesis", + "text_level": 1, + "bbox": [ + 214, + 599, + 395, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ROPE represents a comprehensive benchmark in category-level object pose estimation by scaling up the diversity and number of object categories to unprecedented levels, encompassing a wide range of materials. This diversity presents new challenges for network training data due to the higher demands on the dataset's scale and diversity. Collecting a larger real-world dataset would be prohibitively expensive and unlikely to ensure sufficient diversity.", + "bbox": [ + 212, + 628, + 784, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To bridge the sim2real gap, which is pronounced when using synthetic data, either in RGB or geometry, this paper proposes a novel method based on mixed reality with depth simulation for synthetic data generation. Specifically, as demonstrated in Figure 4, we employ mixed reality [28] techniques to generate RGB data, thereby reducing the RGB sim2real gap. In parallel, we simulate the mechanism of structured light depth sensors within blender [6]. This involves rendering infrared (IR) images and applying stereo matching to produce synthetic depth maps, effectively narrowing the geometry's sim2real gap.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b24d7864e681bc47c2a299d2a477da0236290482111c0d79cb050ffe4dbd8173.jpg", + "image_caption": [ + "Fig. 5: Omni6DPose statistics, showcasing the dataset distribution. Left: Category distribution, highlighting 149 categories and diverse materials. Right: Object size distribution across 5000 objects, illustrating diversity in shapes." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 488, + 300 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8960cd64b7036bd10197f1ff8903d7c074e3fd6859f78cfb271cb3f671c0fc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 146, + 777, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During data generation, we implement domain randomization for illumination and object materials to further enhance the dataset's diversity. All the background images are sourced from public datasets, including 19,658 images from MatterPort3D [2], 2,572 from ScanNet++ [38], and 540 from IKEA [28]. To the best of our knowledge, this is the first simulated dataset that uses a context-aware mixed reality approach combined with physical-based depth sensor simulation for object pose estimation tasks.", + "bbox": [ + 212, + 366, + 787, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Dataset Statistics", + "text_level": 1, + "bbox": [ + 214, + 484, + 408, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Category Statistics The comprehensive distribution of object category and size are both demonstrated in Figure 5. Most of the categories possess $\\geq 25\\mathrm{K}$ pose annotations in the SOPE dataset, providing sufficient training opportunities. Categories containing objects with diverse and challenging material options (e.g., transparent or specular materials) are equipped with apparently more data generation, such as dishes, cups, bottles, bowls, mugs, etc.", + "bbox": [ + 212, + 506, + 787, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Size Statistics As shown in Figure 5, the objects in our dataset span a wide range of sizes. The majority of the objects are approximately 0.1 meters in length along the diagonal of their bounding boxes, with the largest objects exceeding 1 meter.", + "bbox": [ + 212, + 609, + 787, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Category-level 6D Pose Estimation Method", + "text_level": 1, + "bbox": [ + 214, + 684, + 671, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the Omni6DPose, one naturally ponders the optimal technical approach for large-scale category-level pose estimation. The recently introduced state-of-the-art category-level 6D pose estimation technique, GenPose [43], offers a promising avenue by employing a diffusion-based probabilistic method [23, 24]. Moreover, the diffusion model has demonstrated remarkable efficacy across various high-dimensional domains with extensive training data [5, 25, 34, 42]. Expanding on this groundwork, our study delves further into the probabilistic approach, presenting an enhanced iteration of GenPose, named GenPose++.", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 113, + 761, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd6667f41f34f9d181debef2094ec10cadf3934d738b094a7bbef93d5a365440.jpg", + "image_caption": [ + "Fig. 6: Overview of GenPose++. We employ segmented point clouds and cropped RGB images as inputs. PointNet++ extracts geometric features, while DINO v2, a pre-trained 2D foundation model, extracts semantic features. These features are fused to condition a diffusion model, which generates object pose candidates and their corresponding energy. Finally, clustering is applied to address aggregation issues in the multimodal distribution of poses for objects with non-continuous symmetry, such as boxes, effectively resolving the pose estimation challenge." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 782, + 315 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\mathrm{GenPose + + }$ integrates two crucial enhancements: Semantic-aware feature extraction (see Fig.6 (a)) and Clustering-based aggregation (as shown in Fig.6 (c)). The subsequent sections will detail the three primary stages of the GenPose++ pipeline. Moreover, given the estimated 6D pose, $\\mathrm{GenPose + + }$ provide a additional regression network to predict the 3D scale of the object.", + "bbox": [ + 214, + 439, + 787, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Training Semantic-aware Score and Energy Networks", + "text_level": 1, + "bbox": [ + 215, + 523, + 702, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The learning agent is trained on our paired dataset $\\mathcal{D} = \\{(p_k,O_k,I_k)\\}_{k = 1}^n$ , where $\\pmb {p}_k\\in \\mathrm{SE}(3)$ , $O_{k}\\in \\mathbb{R}^{3\\times N}$ , and $I_{k}\\in \\mathbb{R}^{3\\times H\\times W}$ denote a 6D pose, a partially observed 3D point cloud with $N$ points, and a cropped RGB image with $H\\times W$ resolution, respectively. Given an unseen object with point cloud $O^{*}$ and RGB image $I^{*}$ , the goal is to recover the corresponding ground-truth pose $\\pmb{p}^{*}$ .", + "bbox": [ + 212, + 547, + 787, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following GenPose, we initially train a score network $\\varPhi_{\\theta}:\\mathbb{R}^{|\\mathcal{P}|}\\times \\mathbb{R}^{1}\\times \\mathbb{R}^{3\\times N}\\times$ $\\mathbb{R}^{3\\times H\\times W}\\to \\mathbb{R}^{|P|}$ and an energy network $\\varPsi_{\\phi}:\\mathbb{R}^{|\\mathcal{P}|}\\times \\mathbb{R}^{1}\\times \\mathbb{R}^{3\\times N}\\times \\mathbb{R}^{3\\times H\\times W}\\to \\mathbb{R}^{1}$ from the dataset $\\mathcal{D}$ using the denoising score-matching objective [26]:", + "bbox": [ + 214, + 623, + 785, + 669 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t \\sim \\mathcal {U} (\\epsilon , 1)} \\left\\{\\lambda (t) \\mathbb {E} _ {\\substack {\\boldsymbol {p} (0) \\sim p _ {\\text {data}} (\\boldsymbol {p} (0) | O, I), \\\\ \\boldsymbol {p} (t) \\sim \\mathcal {N} (\\boldsymbol {p} (t); \\boldsymbol {p} (0), \\sigma^ {2} (t) \\mathbf {I})}} \\left[ \\left\\| \\boldsymbol {s} (\\boldsymbol {p} (t), t | O, I) - \\frac {\\boldsymbol {p} (0) - \\boldsymbol {p} (t)}{\\sigma (t) ^ {2}} \\right\\| _ {2} ^ {2} \\right] \\right\\} \\tag{3}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 678, + 785, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The training loss of the score and energy network can be obtained by replacing $s(\\pmb{p}(t), t|O, I)$ in Eq. 3 with $\\pmb{\\varPhi}_{\\theta}(\\pmb{p}(t), t|O, I)$ and $\\nabla_{\\pmb{p}}\\pmb{\\Psi}_{\\phi}^{*}(\\pmb{p}(t), t|O, I)$ , respectively.", + "bbox": [ + 214, + 728, + 787, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Unlike GenPose, our score and energy network are semantic-aware, as both networks are conditioned on an RGB image to incorporate semantic cues for pose estimation. To fuse the features extracted from the image and point cloud, we encode the RGB image img and point cloud obj using the pre-trained feature extractors from DINOv2 [20] and PointNet++ [22], respectively. Then, we concatenate these features together in a pointwise manner similar to [27].", + "bbox": [ + 212, + 760, + 787, + 849 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Candidates Generation and Outlier Removal", + "text_level": 1, + "bbox": [ + 215, + 146, + 630, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following GenPose, we subsequently sample pose candidates $\\{\\hat{p}_i\\}_{i=1}^K$ by solving the following Probability Flow (PF) ODE [24] constructed by the score network $\\Phi_\\theta$ from $t = 1$ to $t = \\epsilon$ :", + "bbox": [ + 212, + 167, + 787, + 213 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d \\boldsymbol {p}}{d t} = - \\sigma (t) \\dot {\\sigma} (t) \\boldsymbol {\\Phi} _ {\\theta} (\\boldsymbol {p} (t), t | O, I) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 220, + 785, + 247 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\pmb{p}(1) \\sim \\mathcal{N}(\\mathbf{0}, \\sigma_{\\max}^2 \\mathbf{I})$ , $\\sigma(t) = \\sigma_{\\min} \\left( \\frac{\\sigma_{\\max}}{\\sigma_{\\min}} \\right)^t$ , $\\sigma_{\\min} = 0.01$ and $\\sigma_{\\max} = 50$ .", + "bbox": [ + 212, + 253, + 754, + 270 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To remove the outliers in candidates, we sort the candidates into a sequence $\\hat{\\pmb{p}}_{\\tau_1} \\succ \\hat{\\pmb{p}}_{\\tau_2} \\dots \\succ \\hat{\\pmb{p}}_{\\tau_K}$ where:", + "bbox": [ + 212, + 270, + 782, + 301 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} _ {\\tau_ {i}} \\succ \\hat {\\boldsymbol {p}} _ {\\tau_ {j}} \\Longleftrightarrow \\boldsymbol {\\Psi} _ {\\phi} (\\hat {\\boldsymbol {p}} _ {\\tau_ {i}}, \\epsilon | O) > \\boldsymbol {\\Psi} _ {\\phi} (\\hat {\\boldsymbol {p}} _ {\\tau_ {j}}, \\epsilon | O) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 310, + 785, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Then, we filter out the last $1 - \\delta \\%$ candidates and obtain $\\hat{\\pmb{p}}_{\\tau_1} \\succ \\hat{\\pmb{p}}_{\\tau_2} \\dots \\succ \\hat{\\pmb{p}}_{\\tau_M}$ where $\\delta = 40\\%$ is a hyper parameter and $M = \\lfloor \\delta \\cdot K \\rfloor$ .", + "bbox": [ + 214, + 332, + 782, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Clustering-based Aggregation", + "text_level": 1, + "bbox": [ + 214, + 375, + 506, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we aggregate the remaining candidates $\\{\\hat{\\pmb{p}}_{\\tau_i} = (\\hat{T}_{\\tau_i},\\hat{R}_{\\tau_i})\\}_{i = 1}^M$ to obtain the final results. GenPose achieves this by simply mean-pooling the filtered candidates. However, this strategy will encounter a severe mean-mode issue when the object possesses discrete symmetrical properties. As illustrated in Fig. 6, a ping pong paddle has two symmetric ground truth poses (modes). Since the score network has encountered both modes during training, an optimally trained score network will likely output candidates around both modes. Simply mean-pooling these candidates will yield the average of the two modes, known as the 'mean mode', which will deviate from both modes.", + "bbox": [ + 212, + 388, + 787, + 525 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To mitigate this issue, we introduce a clustering-based aggregation mechanism. We employ DBSCAN [8] to cluster the candidates. It identify dense regions in the data space, forming clusters based on the density of data points and effectively separating noise from meaningful patterns, without the need to specify the number of clusters. This is achieved through dynamic determination of cluster quantities based on distance threshold $(\\varepsilon)$ and density threshold (MinPts). For instance, in our empirical setting, we set $\\varepsilon \\approx 0.45$ rad and MinPts $= 5$ . After clustering the candidates, we select the cluster with the largest number of objects and get the mean-pooling result as the final estimation following GenPose.", + "bbox": [ + 212, + 526, + 787, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 675, + 375, + 691 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Category-Level 6D Object Pose Estimation", + "text_level": 1, + "bbox": [ + 214, + 696, + 617, + 713 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metric In prior studies, metrics such as the mean average precision (mAP) for 3D bounding box IoU and the mean average precision (mAP) for objects with translation errors less than $m$ cm and rotation errors less than $n^{\\circ}$ have been commonly used [28]. This evaluation typically involves two steps: instance-level object segmentation and pose estimation from these detections. However, these metrics are influenced by the detection model's performance. To focus solely on evaluating the precision of pose estimation, we assume ground truth instance segmentation is known and propose the following two metrics:", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 113, + 761, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a7d38b6fc40da845b9125db99ea6d63455fa730cdeca388d79157a8f82bbf78b.jpg", + "table_caption": [ + "Table 2: Quantitative comparison of category-level pose estimation on ROPE. \\(\\uparrow\\) represents a higher value indicating better performance, while \\(\\downarrow\\) represents a lower value indicating better performance. Prior-free indicates whether the method requires category prior. \\{'-'\\) indicates that GenPose does not predict the object scale." + ], + "table_footnote": [], + "table_body": "
MethodInputPrior-freeAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
DeterministicNOCS [28]RGBD✓0.00.00.00.00.00.00.0
SGPA [3]RGBD✗10.52.00.04.36.79.315.0
IST-Net [16]RGBD✓28.710.60.52.03.45.38.8
HS-Pose [44]D✓31.613.61.13.55.38.412.7
ProbabilisticGenPose [43]D✓---6.69.613.119.3
GenPose++(Ours)RGBD✓39.019.12.010.015.119.529.4
", + "bbox": [ + 230, + 205, + 777, + 323 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- AUC@IoU $_n$ : This metric assesses the accuracy of predicted 3D bounding boxes, calculated via the Area Under the Curve (AUC) from various Intersection over Union (IoU) thresholds starting at $n$ . In our study, we utilize AUC@IoU $_{25}$ , AUC@IoU $_{50}$ , and AUC@IoU $_{75}$ as the benchmarks.", + "- VUS@ $n^{\\circ}mcm$ : This metric offers a detailed analysis of 6D pose estimation accuracy, derived from the Volume Under Surface (VUS) across ranges of rotational (up to $n^{\\circ}$ ) and translational (up to $mcm$ ) error thresholds. VUS aggregates the accuracy of pose predictions within set boundaries. In this paper, we apply VUS@ $5^{\\circ}2cm$ , VUS@ $5^{\\circ}5cm$ , VUS@ $10^{\\circ}2cm$ , and VUS@ $10^{\\circ}5cm$ for comprehensive performance assessment." + ], + "bbox": [ + 225, + 339, + 784, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Additionally, to demonstrate the robustness of our method to segmentation masks, we present the pose estimation performance of GenPose++ using SAM [13] segmentation results in the supplementary material.", + "bbox": [ + 215, + 503, + 784, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines We evaluate five category-level pose estimation methods: NOCS [28], SGPA [3], HS-Pose [44], IST-Net [16] and GenPose [43]. Except for NOCS, which conducts both object detection and pose estimation as a whole, all methods are equipped with ground truth detection results. For SGPA, the prior point cloud of each category is constructed by randomly selecting an object from the training dataset. Considering that previous methods' augmentations for symmetry properties are only applicable to specific object categories within the NOCS dataset and not suitable for all object categories in Omni6Dpose, data augmentation for object symmetry is disabled during training for all baseline methods. All methods are trained on SOPE and directly tested on ROPE.", + "bbox": [ + 212, + 558, + 785, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results and Analysis. In Table 2, we present the quantitative evaluation results of previous methods compared to GenPose++ on the ROPE dataset. Overall, generative methods continue to dominate in the performance evaluation on ROPE. The VUS surface depicted in Figure 7 provides a more detailed reflection of the performance of each model. Unlike deterministic approaches, generative methods can handle ambiguity without any specialized design requirements. Moreover, these methods directly generate the distribution of object poses, eliminating the need for depth map-based pose fitting. This approach is particularly", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8577d558e9ccc200a47273e9044677f58ad2053039a13285e0607224c4d64cc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 146, + 390, + 246 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7b2051dcc5293d9a7fe0796028210d31c06e74c9b9890724cda7329cb842a01e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 146, + 571, + 244 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c077af05b498de02cdc8eab0c596a7d8c21efe75484d6d0273e29d17bcfd9d10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 146, + 787, + 244 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fec35d4ed07e2909d968f20e09d0c1a7fd9731396d3a649319ce5e8374292dad.jpg", + "image_caption": [ + "Fig. 7: Qualitative comparison with baselines on ROPE dataset." + ], + "image_footnote": [], + "bbox": [ + 215, + 248, + 388, + 344 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/03ce1c66280d01f8a3c7411edbf255b53f6b004a790450d12fe12ac5dee565f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 248, + 571, + 344 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2525cd83110824542c6fb1a551052ff2447259b4ec7dd52b968d177337a488e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 248, + 785, + 345 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "advantageous for challenging material types, such as transparent or reflective objects, where structured-light depth cameras tend to introduce significant noise, severely impacting pose fitting accuracy. Furthermore, the NOCS method does not demonstrate effective performance on the ROPE dataset, leading to the supposition that methods relying solely on RGB information to predict the shape of an object in the canonical space become less robust as the scale of category diversity increases. Compared to GenPose, GenPose++ achieves a significant lead by leveraging the powerful perception capabilities of the 2D foundation model, along with the robustness of clustering towards discrete symmetric properties. You can find qualitative visualizations in Figure 8.", + "bbox": [ + 212, + 375, + 787, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 534, + 388, + 549 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In order to validate the design decisions of our approach, we performed a series of ablation experiments on our method:", + "bbox": [ + 212, + 559, + 785, + 590 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- w/o clustering. Without clustering directly take the average of all remaining pose candidates after outlier removal as the pose estimation output.", + "- w/o scale prediction. Use the estimated pose to transform the observed point cloud into object space, then take the bounding box length as the maximum projection from the point cloud to each axis.", + "- w/o simulated depth. Use perfect depth for training.", + "- w/o point-wise feature fusion. In the feature extraction stage, separately extract the RGB feature and geometric feature, and then concatenation." + ], + "bbox": [ + 223, + 601, + 784, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3 illustrates the contribution of each component of GenPose++ to its performance. The introduction of the clustering module allows GenPose++ to effectively aggregate the multimodal distributions caused by discrete symmetries, leading to higher performance. The scale prediction in GenPose++ significantly surpasses direct calculations from the object's point cloud due to ambiguities from partial observations and errors from point cloud noise, particularly in transparent and reflective objects. Training with simulated depth data results in", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 113, + 761, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3afd96b96af6e85dcbc6fd6490af7edbf579f44495b34ee8436c5e695e886d30.jpg", + "table_caption": [ + "Table 3: Ablation study on category-level 6D object pose estimation" + ], + "table_footnote": [], + "table_body": "
AblationAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
Full39.019.12.010.015.119.529.4
w/o scale prediction13.83.40.210.015.119.529.4
w/o clustering38.618.71.99.414.118.427.8
w/o simulated depth35.316.51.77.311.914.624.3
w/o point-wise fusion34.215.41.47.110.614.621.9
", + "bbox": [ + 259, + 164, + 738, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0adc22a2ab6a498e57403c9ae654072fee93fcf84548c0e1cdf8b1a97d1e38dd.jpg", + "image_caption": [ + "Fig. 8: Qualitative comparison with baselines on ROPE dataset." + ], + "image_footnote": [], + "bbox": [ + 220, + 301, + 784, + 462 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "better performance than training with perfect point clouds, as the physics-based depth camera simulation substantially reduces the sim2real gap for depth data. Point-wise fusion outperforms global fusion as it retains more of the object's local geometric features, which are crucial for accurate object pose prediction.", + "bbox": [ + 212, + 494, + 784, + 555 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Category-Level 6D Object Pose Tracking", + "text_level": 1, + "bbox": [ + 214, + 563, + 601, + 579 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Metric. We report the following metrics for object pose tracking evaluation:", + "bbox": [ + 214, + 585, + 771, + 602 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- FPS: Frames Per Second, which indicates the speed of pose tracking.", + "- VUS@5°5cm: Volume Under Surface, assessing pose estimation accuracy for rotation errors within 0 to $5^{\\circ}$ and translation errors within 0 to 5cm.", + "- mIoU: Mean Intersection over Union, representing the average 3D overlap between the ground truth and the predicted bounding boxes.", + "- $\\mathbf{Rerr}(\\circ)$ : Average rotation error in degrees.", + "- $\\mathbf{Terr}(\\mathbf{cm})$ : Average translation error in centimeters." + ], + "bbox": [ + 223, + 609, + 784, + 712 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Baselines. This paper employs BundleTrack, CATRE, and GenPose as baselines for object pose tracking. BundleTrack is a training-free approach that utilizes multi-view feature point detection and matching for tracking the pose of unseen objects. CATRE aligns partially observed point clouds to abstract shape priors to estimate relative transformations, enhancing pose accuracy. Conversely, GenPose utilizes a generative approach to effectively resolve pose ambiguities, notably in symmetric objects. Following GenPose, we have adapted GenPose++ to object pose tracking with minor modifications.", + "bbox": [ + 212, + 728, + 787, + 851 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results and Analysis. Table 4 presents the results of category-level object pose tracking algorithms CATRE, GenPose, and our method, as well as the results for the unseen object pose tracking algorithm BundleTrack. The category-level pose estimation methods appear to achieve relatively better outcomes since they benefit from learning the category-level canonical space of objects within the SOPE dataset. However, for the training-free BundleTrack, reliance on RGB information for keypoint detection and matching poses challenges, often failing on objects with weak textures. Additionally, its dependency on depth values for global optimization renders it less effective in handling instances with substantial depth noise, such as transparent or specular objects. Our method, without undergoing specialized design, has achieved results comparable to state-of-the-art approaches. Although the inference speed of our method is lower than that of CATRE and GenPose, the achieved 17.8 FPS is sufficient for certain downstream tasks, such as robotic manipulation. Furthermore, recent rapid developments in research on fast samplers are beneficial to our approach, potentially enhancing its performance and applicability in real-time scenarios.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/f634bc31b432a88ca62eefc05fecd81ccbed471dff7b48076fe4609e385c2e08.jpg", + "table_caption": [ + "Table 4: Results of category-level object pose tracking on ROPE. The results are averaged over all 149 categories. GT. Pert. denotes that a perturbed ground truth pose is utilized as the initial object pose." + ], + "table_footnote": [], + "table_body": "
MethodsInputInit.Speed(FPS)↑5°5cm↑mIoU↑Rerr(°)↓Terr(cm)↓
BundleTrack [31]RGBDGT.12.41.33.946.923.5
CATRE [18]DGT. Pert.38.515.955.421.32.6
GenPose [43]DGT. Pert.26.313.3-19.31.2
GenPose++(Ours)RGBDGT. Pert.17.815.953.417.61.2
", + "bbox": [ + 220, + 449, + 777, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions and Discussion", + "text_level": 1, + "bbox": [ + 215, + 553, + 517, + 569 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduce Omni6DPose, a comprehensive dataset for 6D object pose estimation, featuring extensive scale, diversity, and material variety. Through thorough experimentation, our findings suggest that the probabilistic framework holds promise for category-level 6D object estimation, leveraging semantic information provided by RGB images to address large-scale pose estimation challenges. However, the performance of GenPose++ on Omni6DPose reveals significant room for improvement, with the model still hampered by slow inference speeds resulting from the iterative refinement nature inherent in the diffusion model. Future works could focus on addressing these challenges and integrating the universal 6D pose estimation module trained on Omni6DPose into a broader range of downstream tasks [4,33,37,42].", + "bbox": [ + 212, + 583, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Acknowledgment", + "text_level": 1, + "bbox": [ + 215, + 763, + 416, + 781 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China - General Program (Project ID: 62376006), the National Youth Talent Support Program (Project ID: 8200800081), and NSFC-62306046.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 113, + 761, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. An, B., Geng, Y., Chen, K., Li, X., Dou, Q., Dong, H.: RGBmanip: Monocular image-based robotic manipulation through active object pose estimation (2023)", + "2. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)", + "3. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021)", + "4. Cheng, J., Wu, M., Zhang, R., Zhan, G., Wu, C., Dong, H.: Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220 (2023)", + "5. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4800-4810 (2023)", + "6. Dai, Q., Zhang, J., Li, Q., Wu, T., Dong, H., Liu, Z., Tan, P., Wang, H.: Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and transparent objects. In: European Conference on Computer Vision. pp. 374-391. Springer (2022)", + "7. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)", + "8. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: kdd. vol. 96, pp. 226-231 (1996)", + "9. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022)", + "10. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. vol. 7724 (10 2012). https://doi.org/10.1007/978-3-642-33885-4_60", + "11. Hodań, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017)", + "12. Jung, H., Zhai, G., Wu, S.C., Ruhkamp, P., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., Roth, D., Navab, N., Busam, B.: Housecat6d - a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios (2023)", + "13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)", + "14. Lin, Z.H., Huang, S.Y., Wang, Y.C.F.: Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "15. Liu, J., Sun, W., Liu, C., Zhang, X., Fan, S., Wu, W.: Hff6d: Hierarchical feature fusion network for robust 6d object pose tracking. IEEE Transactions on Circuits" + ], + "bbox": [ + 218, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Systems for Video Technology pp. 1-13 (06 2022). https://doi.org/10.1109/TCSVT.2022.3181597", + "16. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. arXiv preprint arXiv:2303.13479 (2023)", + "17. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: ICCV (2021)", + "18. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: European Conference on Computer Vision (ECCV) (October 2022)", + "19. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics 22(12), 2633-2651 (2016). https://doi.org/10.1109/TVCG.2015.2513408", + "20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "21. Piga, N.A., Onyshchuk, Y., Pasquale, G., Pattacini, U., Natale, L.: Roft: Realtime optical flow-aided 6d object pose and velocity tracking. IEEE Robotics and Automation Letters 7(1), 159–166 (2022). https://doi.org/10.1109/LRA.2021.3119379", + "22. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "23. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems 32 (2019)", + "24. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "25. Tian, Y., Zhang, J., Huang, G., Wang, B., Wang, P., Pang, J., Dong, H.: Robokeygen: Robot pose and joint angles estimation via diffusion-based 3d keypoint generation. arXiv preprint arXiv:2403.18259 (2024)", + "26. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011)", + "27. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019)", + "28. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "29. Wang, J., Chen, K., Dou, Q.: Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4807-4814. IEEE (2021)", + "30. Wang, P., Jung, H., Li, Y., Shen, S., Srikanth, R.P., Garattoni, L., Meier, S., Navab, N., Busam, B.: Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21222-21231 (2022)", + "31. Wen, B., Bekris, K.E.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong", + "bbox": [ + 269, + 114, + 761, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 13209-13218 (October 2021)", + "33. Wu, M., Zhong, F., Xia, Y., Dong, H.: Targf: Learning target gradient field to rearrange objects without explicit goal specification. Advances in Neural Information Processing Systems 35, 31986-31999 (2022)", + "34. Wu, T., Wu, M., Zhang, J., Gan, Y., Dong, H.: Learning score-based grasping primitive for human-assisting dexterous grasping. Advances in Neural Information Processing Systems 36 (2024)", + "35. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)", + "36. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018)", + "37. Xue, T., Wu, M., Lu, L., Wang, H., Dong, H., Chen, B.: Learning gradient fields for scalable and generalizable irregular packing. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)", + "38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023)", + "39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)", + "40. You, Y., Shi, R., Wang, W., Lu, C.: CPPf: Towards robust category-level 9d pose estimation in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6866-6875 (2022)", + "41. You, Y., Xiong, K., Yang, Z., Huang, Z., Zhou, J., Shi, R., Fang, Z., Harley, A.W., Lu, C.: Pace: Pose annotations in cluttered environments. arXiv preprint arXiv:2312.15130 (2023)", + "42. Zeng, Y., Wu, M., Yang, L., Zhang, J., Ding, H., Cheng, H., Dong, H.: Distilling functional rearrangement priors from large models. arXiv preprint arXiv:2312.01474 (2023)", + "43. Zhang, J., Wu, M., Dong, H.: Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems 36 (2024)", + "44. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17163-17173 (June 2023)", + "45. Zhou, L., Liu, Z., Gan, R., Wang, H., au2, M.H.A.J.: Dr-pose: A two-stage deformation-and-registration pipeline for category-level 6d object pose estimation (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 742 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Omni6DPose", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_model.json b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd7b3fe6b80f5f26e387520fa9a74b16f6197661 --- /dev/null +++ b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_model.json @@ -0,0 +1,2654 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.257, + 0.141, + 0.75, + 0.21 + ], + "angle": 0, + "content": "Omni6DPose: A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.234, + 0.784, + 0.268 + ], + "angle": 0, + "content": "Jiyao Zhang\\(^{1,2,3*}\\), Weiyao Huang\\(^{1*}\\), Bo Peng\\(^{1*}\\), Mingdong Wu\\(^{1,2,3}\\), Fei Hu\\(^{1,2,3}\\), Zijian Chen\\(^{4}\\), Bo Zhao\\(^{5,6}\\), and Hao Dong\\(^{1,2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.277, + 0.768, + 0.3 + ], + "angle": 0, + "content": "\\(^{1}\\) Center on Frontiers of Computing Studies, School of Computer Science, Peking University \\(^{2}\\) PKU-Agibot Lab, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.3, + 0.779, + 0.342 + ], + "angle": 0, + "content": "\\(^{3}\\) National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University \\(^{4}\\) Waseda University \\(^{5}\\) Shanghai Jiao Tong University \\(^{6}\\) BAAI {jiyaozhang, sshwy, bo.peng, hofee}@stu.pku.edu.cn zijianchenwaseda@akane.waseda.jp bo.zhao@sjtu.edu.cn {wmingd, hao.dong}@pku.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.277, + 0.779, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.346, + 0.645, + 0.361 + ], + "angle": 0, + "content": "https://jiyao06.github.io/Omni6DPose/" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.373, + 0.74, + 0.65 + ], + "angle": 0, + "content": "Abstract. 6D object pose estimation is crucial in the field of computer vision. However, it suffers from a significant lack of large-scale and diverse datasets, impeding comprehensive model evaluation and curtailing downstream applications. To address these issues, this paper introduces Omni6DPose, a substantial benchmark featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is divided into three main components: ROPE (Real 6D Object Pose Estimation Dataset), which includes 332K images annotated with over 1.5M annotations across 581 instances in 149 categories; SOPE(Simulated 6D Object Pose Estimation Dataset), a simulated training set created by mixed reality and physics-based depth simulation; and PAM(Pose Aligned 3D Models), the manually aligned real scanned objects used in ROPE and SOPE. Omni6DPose is inherently challenging due to the substantial variations and ambiguities. To address this issue, we introduce GenPose++, an enhanced version of the SOTA category-level 6D object pose estimation framework, incorporating two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation. Moreover, we provide a comprehensive benchmarking analysis to evaluate the performance of previous methods on this new large-scale dataset in the realms of 6D object pose estimation and pose tracking." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.653, + 0.735, + 0.668 + ], + "angle": 0, + "content": "Keywords: Benchmark \\(\\cdot\\) object pose estimation \\(\\cdot\\) object pose tracking" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.376, + 0.707 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.817 + ], + "angle": 0, + "content": "6D object pose estimation [16,28,45] and pose tracking [15,21] from single images is an essential task in computer vision, holding immense potential for applications in robotics [1] and augmented reality/virtual reality (AR/VR) [19]. Over recent decades, the domain has experienced significant advancements, primarily dominated by data-driven learning approaches. Analogous to the pivotal role of data in learning-based 2D foundation tasks, high-quality, comprehensive datasets are paramount in the context of 6D object pose estimation and tracking." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.547, + 0.84 + ], + "angle": 0, + "content": "*: equal contributions, †: corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 1: We introduce Omni6DPose for universal 6D object pose estimation and tracking, which comprises three key components: (1) ROPE (right), a large-scale real-world dataset for evaluation. (2) SOPE (left), a vast synthetic dataset generated through mixed reality and physics-based simulation for training. (3) PAM (center), a collection of pose-aligned 3D object models. Omni6DPose is featured by its diversity in object categories, large scale, and variety in object materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.624 + ], + "angle": 0, + "content": "Today 6D object pose estimation is studied under two lenses: instance-level and category-level. In instance-level settings, datasets such as Linemod [10], YCB-Video [36], and T-LESS [11] have gained widespread acceptance as benchmarks. These datasets are distinguished by their focus on detailed, individual object instances, thereby enabling algorithms to precisely learn and predict the poses of specific items. On the other hand, category-level pose estimation emphasizes generalization across different items within a particular object category. The NOCS [28] dataset stands out as the most widely used in the category-level object pose estimation field, providing a simulated dataset for training and a small-scale real-world dataset for evaluation. Despite their contributions to advancing the field, these datasets present limitations due to their small scale in terms of instances or categories. This results in two significant challenges:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.633, + 0.785, + 0.661 + ], + "angle": 0, + "content": "1. It hampers comprehensive evaluation of different models' performance, limiting the development of research in this field." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.663, + 0.785, + 0.694 + ], + "angle": 0, + "content": "2. It restricts the applicability of research findings across diverse domains, due to the limited variety of object instances or categories represented." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.633, + 0.785, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address the aforementioned challenges and drive advancements in this field, this paper introduces Omni6DPose, a universal 6D object pose estimation dataset featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is segmented into three principal components: (1) ROPE (Real 6D Object Pose Estimation Dataset), which encompasses \\(332\\mathrm{K}\\) images annotated with over 1.5M annotations across 581 instances in 149 categories; (2) SOPE (Simulated 6D Object Pose Estimation Dataset), comprising \\(475\\mathrm{K}\\) images generated in a mixed reality setting with depth simulation, furnished with over 5M annotations spanning 4162 instances in the same 149" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.333, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.196, + 0.333, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.245, + 0.333, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.147, + 0.445, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.197, + 0.446, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.245, + 0.445, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.147, + 0.557, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.197, + 0.557, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.245, + 0.557, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.147, + 0.67, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.197, + 0.67, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.245, + 0.67, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.147, + 0.781, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.197, + 0.781, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.245, + 0.781, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.302, + 0.784, + 0.331 + ], + "angle": 0, + "content": "Fig. 2: ROPE dataset visualization. In the figure, bounding boxes are colored according to the coordinates in the object's coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.784, + 0.398 + ], + "angle": 0, + "content": "categories. The mixed reality bridges the semantic sim2real gap, while the depth sensor simulation closes the geometric sim2real gap; (3) PAM(Pose Aligned 3D Models), the manually aligned, real scanned objects utilized in both ROPE and SOPE, enabling the generation of diverse downstream task data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.549 + ], + "angle": 0, + "content": "Omni6DPose poses inherent challenges due to its considerable variations, diverse materials, and inherent ambiguities, which reflect the complexities encountered in real-world applications. Figure 2 illustrates some examples from ROPE. To tackle these issues, we introduce GenPose++, which incorporates GenPose [43] with two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation, tailored specifically to the nuances of the Omni6DPose in question. Furthermore, as a Universal 6D object pose estimation dataset, this paper also offers a comprehensive benchmarking analysis to assess the performance of existing methods on category-level 6D object pose estimation and pose tracking. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.555, + 0.797, + 0.584 + ], + "angle": 0, + "content": "1. We present Omni6DPose, a comprehensive 6D object pose estimation dataset with extensive categories, instance diversity, and material variety." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.584, + 0.785, + 0.614 + ], + "angle": 0, + "content": "2. We propose a real data collection pipeline and a simulation framework for generating synthetic data with low semantic and geometry sim2real gaps." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.614, + 0.785, + 0.643 + ], + "angle": 0, + "content": "3. We introduce GenPose++ for category-level 6D object pose estimation and tracking, demonstrating SOTA performance on Omni6DPose." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.555, + 0.797, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.656, + 0.383, + 0.672 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.785, + 0.768 + ], + "angle": 0, + "content": "For 6D object pose estimation, there are two main branches: instance-level and category-level. Instance-level estimation is tested on seen objects or on unseen objects with reference, while category-level estimation is tested on unseen instances of known categories without CAD models. In this section, we review and compare existing datasets to our large-scale category-level dataset and review relevant algorithms for category-level pose estimation and tracking." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.774, + 0.565, + 0.789 + ], + "angle": 0, + "content": "2.1 6D Object Pose Estimation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Following the outlined branches of 6D object pose estimation, we have reviewed datasets corresponding to both instance-level and category-level 6D object pose estimation. A comparative analysis of these datasets is provided in Table 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.241 + ], + "angle": 0, + "content": "Table 1: This table compares datasets for 6D object pose estimation, focusing on object category count, reality of the data, data modalities (RGB, Depth, IR), and object attributes such as quantity, CAD model availability, and inclusion of transparent and specular objects. It also details video characteristics by number and marker presence, along with image and annotation counts. 'Wild6D' refers specifically to the test split of the Wild6D dataset, as the training data does not provide annotations. The symbol '-' indicates the absence of a particular feature within the dataset." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.243, + 0.777, + 0.433 + ], + "angle": 0, + "content": "
DatasetCat. RealModalityObjectMarker-free Vid.Img.Anno.
RGBDepth IRIRNum. CAD Trans. Spec.
CAMERA [28]6X✓✓X1085✓XX✓-300K
SOPE(Ours)149X✓✓✓4162✓✓✓✓-475K
YCB-Video [36]-✓✓✓X21✓XX✓92133K
T-LESS [11]-✓✓✓X30✓XXX2048K
Linemod [10]-✓✓✓X15✓XXX-18K
StereoOBJ-1M [17]-✓✓XX18✓✓✓X182393K
REAL275 [28]6✓✓✓X42✓XXX188K
PhoCal [30]8✓✓✓X60✓✓✓✓243.9K
HouseCat6D [12]10✓✓✓X194✓✓✓✓4123.5K
Wild6D* [9]5✓✓✓X162X(✓)X✓48610K
PACE [41]44✓✓✓X576✓XX✓30055K
ROPE(Ours)149✓✓✓✓581✓✓✓✓363332K
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Instance-level 6D object pose estimation dataset. LineMod [10] is one of the most used datasets, providing non-temporal RGB-D images and ground truth pose annotations. YCB-Video [36] provides RGB-D videos and annotations, enabling pose-tracking approaches. T-LESS [11] features texture-less objects with symmetries and mutual similarities. StereoOBJ-1M [17] achieves a leap in dataset scale and features transparent and reflective objects. While these datasets are extensive in terms of image and annotation count, they are limited in the diversity of instances they cover. For example, StereoOBJ-1M comprises 339K frames and 1.5M annotations, yet it includes only 18 unique object instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Category-level 6D object pose estimation dataset. NOCS [28] provides the first benchmark in category-level pose estimation, while Wild6D [9] addresses the scalability issue of datasets by leveraging unlabeled and synthetic data. PhoCal [30] focuses on photometrically challenging objects and HouseCat6D [12] offers diverse scenes, viewpoints, and grasping annotations. However, these datasets cover only a limited number of categories, even the most extensive dataset, HouseCat6D, includes merely ten categories. PACE [41], a concurrent work, focusing on object pose estimation in complex scenes. In contrast, our datasets, SOPE and ROPE, set new benchmarks by offering the widest range of categories and featuring objects with diverse materials, thereby enhancing dataset diversity and realism for pose estimation research." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.758, + 0.694, + 0.774 + ], + "angle": 0, + "content": "2.2 6D Object Pose Estimation and Tracking Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Category-level 6D object pose estimation. Category-level 6D object pose estimation [28,39,40] aims to estimate unseen instance poses within the same category. NOCS [28] introduces a normalized object coordinate space for pose prediction without CAD models, while SPD [29] and SGPA [3] utilize category-level" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.329 + ], + "angle": 0, + "content": "prriors for enhanced estimation. HS-Pose [44] extends 3D-GC [14] for improved feature extraction from point clouds. IST-Net [16] transforms features between camera and world spaces implicitly, without relying on 3D priors, surpassing previous methods. However, these techniques, mainly regression-based, require ad-hoc designs for symmetric objects. GenPose [43] addresses it by generatively modeling the pose distribution, eliminating the need for symmetry considerations. Yet, GenPose neglects RGB semantic information, which is increasingly vital as object category scales grow. Additionally, its energy-based aggregation algorithm fails with discontinuous multimodal distributions. We introduce GenPose++, which incorporates a 2D foundation model to leverage RGB semantics, improving generalization, and introduces an aggregation module to handle discrete multimodal distributions, addressing the limitations of GenPose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.79, + 0.489 + ], + "angle": 0, + "content": "6D object pose tracking. This paper is situated within the domain of category-level 6D object pose tracking and model-free object tracking. BundleTrack [31] pioneers model-free tracking by leveraging multi-view feature detection for tracking unseen objects without 3D models. CAPTRA [32] enhances articulated pose tracking through recursive updates for better temporal consistency. CATRE [18] aligns partially observed point clouds to abstract shape priors for relative transformations and pose estimation. GenPose, adopting a generative approach, effectively addresses the challenge of pose ambiguities in symmetric objects. Together, these approaches underscore the evolving landscape of object pose tracking, highlighting both the progress and the diversity of strategies being explored." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.461, + 0.514 + ], + "angle": 0, + "content": "3 Omni6DPose Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.609 + ], + "angle": 0, + "content": "This paper introduces a rich variety of object categories, a large-scale, and materially diverse dataset for real 6D object pose estimation, named ROPE. And, a simulated dataset, SOPE, synthesizing with mixed reality and featuring depth simulation, is provided for training. Section 3.1 will discuss the collection and alignment of 3D objects. Section 3.2 will cover the acquisition and labeling of the ROPE dataset. Section 3.3 will explain the generation of the SOPE dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.629, + 0.567, + 0.645 + ], + "angle": 0, + "content": "3.1 3D Object Collection and Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Universal 6D object pose estimation relies on a comprehensive set of objects. We selected 149 categories of everyday objects, all reconstructed with high-precision scanners, and categorized them into two sets: SOPE for simulated data and ROPE for real-world scenes. SOPE primarily includes objects from sources like OmniObject3D [35], PhoCal [30], and GoogleScan [7], alongside a subset from our scans, totaling 5,000 instances. ROPE consists of 580 instances we reconstructed using industrial scanners. Importantly, while most SOPE objects are from public datasets, manual category-level pose alignment is necessary. For object reconstruction, as shown in Figure 3, we use two professional scanners, EinScan H2\\(^{7}\\) and Revopoint POP 3\\(^{8}\\) for objects in different scales. The scan" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.426, + 0.825 + ], + "angle": 0, + "content": "\\(^{7}\\) https://www.einscan.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.826, + 0.453, + 0.84 + ], + "angle": 0, + "content": "8 https://www_REVopoint3d.com/" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.453, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.27, + 0.114, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.466, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.145, + 0.787, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 3: ROPE dataset collection and annotation. (1) Object scanning, where high-precision industrial scanners are used to acquire the CAD models of objects; (2) Object canonicalization, involving the alignment of each object category to the canonical space; (3) Video capture, capturing video sequences in varied scenarios with a depth camera; and (4) Pose annotation, calculating camera poses through Structure from Motion (SFM), further utilizing Farthest Point Sampling (FPS) to select keyframes for keypoint annotation, and performing bundle adjustment to derive initial object pose values, which are then manually refined to obtain more precise annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.482 + ], + "angle": 0, + "content": "ning time depends on object features: it took about 15 minutes for small, simple, Lambertian items like a mouse, and up to an hour for complex, large, or non-Lambertian items like a transparent mug. Finally, we constructed a specialized annotation tool for manually aligning objects in the same category to the category-level canonical space, with each alignment taking roughly one minute." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.55, + 0.506 + ], + "angle": 0, + "content": "3.2 ROPE Acquisition and Annotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.591 + ], + "angle": 0, + "content": "The ROPE dataset was systematically acquired utilizing the RealSense D415 imaging device, encompassing scenarios with 2 to 6 distinct objects and video sequences extending from 762 to 1,349 frames. The integrity and utility of the dataset are underpinned by the precision of object pose annotations, which present notable challenges, chiefly:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.602, + 0.787, + 0.663 + ], + "angle": 0, + "content": "1. Derivation of relative camera poses, denoted as \\( T_{c} = \\{(R_{i},t_{i})\\}_{i = 1}^{n} \\), where each pair \\( (R_i,t_i)\\in \\mathrm{SE}(3) \\), signifying the transformation from the camera space to the world space for the \\( i^{th} \\) frame, with \\( n \\) symbolizing the aggregate frame count within the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.662, + 0.787, + 0.709 + ], + "angle": 0, + "content": "2. Procurement of high-accuracy object poses, represented as \\( T_{o} = \\{(R,t)\\} \\), where the pair \\( (R,t) \\in \\mathrm{SE}(3) \\), delineating the transformation from the object space the camera space for any selected frame." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.602, + 0.787, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "With \\( T_{c} \\) and \\( T_{o} \\) known, it is possible to automate the generation of all annotations within the dataset. Addressing these challenges, we propose a marker-free annotation system. Previous datasets for calculating relative camera pose rely on markers, like NOCS [28], or external robot arms for indirect calculation, such as PhoCal [30], limiting scene diversity. As shown in Figure 3, to enable marker-free annotation in open scenes, we consider it a structure-from-motion (SfM) problem, utilizing intrinsic scene features to optimize camera poses \\( T_{c} \\) using bundle adjustment techniques. This approach aims to solve the optimization problem:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.457, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.145, + 0.782, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.26, + 0.785, + 0.289 + ], + "angle": 0, + "content": "Fig. 4: SOPE synthesis, utilizing mixed reality to bridge the RGB sim2real gap and physical-based depth sensor simulation to minimize the geometric sim2real gap." + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.304, + 0.785, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\min _ {T _ {c}} \\sum_ {i = 1} ^ {n} \\sum_ {j \\in \\mathcal {P} _ {i}} \\| \\pi \\left(R _ {i} X _ {j} + t _ {i}\\right) - x _ {i j} \\| ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.398 + ], + "angle": 0, + "content": "where \\(\\pi\\) denotes the camera projection function, \\(X_{j}\\) represents the 3D points in the world space, and \\(x_{ij}\\) corresponds to the 2D projection of \\(X_{j}\\) in the \\(i^{th}\\) camera frame, with \\(\\mathcal{P}_i\\) being the set of point correspondences in frame \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.489 + ], + "angle": 0, + "content": "For object pose \\( T_{o} \\) annotations, previous methods consider only a single frame, leading to inaccuracies due to the lack of multi-viewpoint constraints. To overcome this, we introduce a two-stage object pose annotation process: in the first stage, keyframes are sampled from SfM results using Farthest Point Sampling (FPS), and 2D-3D keypoint pairs on these keyframes are annotated, providing initial object pose by minimizing the reprojection error:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.503, + 0.785, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\min _ {R, t} \\sum_ {k \\in \\mathcal {K}} \\| \\pi \\left(R X _ {k} + t\\right) - x _ {k} \\| ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.784, + 0.588 + ], + "angle": 0, + "content": "where \\( X_{k} \\) are 3D keypoints of object, \\( x_{k} \\) are corresponding 2D annotations in the image, and \\( \\mathcal{K} \\) is the set of all keypoint correspondences. The initial poses are manually fine-tuned based on the object's projection across all keyframes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.601, + 0.396, + 0.617 + ], + "angle": 0, + "content": "3.3 SOPE Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.719 + ], + "angle": 0, + "content": "ROPE represents a comprehensive benchmark in category-level object pose estimation by scaling up the diversity and number of object categories to unprecedented levels, encompassing a wide range of materials. This diversity presents new challenges for network training data due to the higher demands on the dataset's scale and diversity. Collecting a larger real-world dataset would be prohibitively expensive and unlikely to ensure sufficient diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To bridge the sim2real gap, which is pronounced when using synthetic data, either in RGB or geometry, this paper proposes a novel method based on mixed reality with depth simulation for synthetic data generation. Specifically, as demonstrated in Figure 4, we employ mixed reality [28] techniques to generate RGB data, thereby reducing the RGB sim2real gap. In parallel, we simulate the mechanism of structured light depth sensors within blender [6]. This involves rendering infrared (IR) images and applying stereo matching to produce synthetic depth maps, effectively narrowing the geometry's sim2real gap." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.27, + 0.114, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.49, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.147, + 0.779, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 5: Omni6DPose statistics, showcasing the dataset distribution. Left: Category distribution, highlighting 149 categories and diverse materials. Right: Object size distribution across 5000 objects, illustrating diversity in shapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.473 + ], + "angle": 0, + "content": "During data generation, we implement domain randomization for illumination and object materials to further enhance the dataset's diversity. All the background images are sourced from public datasets, including 19,658 images from MatterPort3D [2], 2,572 from ScanNet++ [38], and 540 from IKEA [28]. To the best of our knowledge, this is the first simulated dataset that uses a context-aware mixed reality approach combined with physical-based depth sensor simulation for object pose estimation tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.485, + 0.409, + 0.499 + ], + "angle": 0, + "content": "3.4 Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Object Category Statistics The comprehensive distribution of object category and size are both demonstrated in Figure 5. Most of the categories possess \\(\\geq 25\\mathrm{K}\\) pose annotations in the SOPE dataset, providing sufficient training opportunities. Categories containing objects with diverse and challenging material options (e.g., transparent or specular materials) are equipped with apparently more data generation, such as dishes, cups, bottles, bowls, mugs, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Object Size Statistics As shown in Figure 5, the objects in our dataset span a wide range of sizes. The majority of the objects are approximately 0.1 meters in length along the diagonal of their bounding boxes, with the largest objects exceeding 1 meter." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.685, + 0.673, + 0.703 + ], + "angle": 0, + "content": "4 Category-level 6D Pose Estimation Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Given the Omni6DPose, one naturally ponders the optimal technical approach for large-scale category-level pose estimation. The recently introduced state-of-the-art category-level 6D pose estimation technique, GenPose [43], offers a promising avenue by employing a diffusion-based probabilistic method [23, 24]. Moreover, the diffusion model has demonstrated remarkable efficacy across various high-dimensional domains with extensive training data [5, 25, 34, 42]. Expanding on this groundwork, our study delves further into the probabilistic approach, presenting an enhanced iteration of GenPose, named GenPose++." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.784, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 6: Overview of GenPose++. We employ segmented point clouds and cropped RGB images as inputs. PointNet++ extracts geometric features, while DINO v2, a pre-trained 2D foundation model, extracts semantic features. These features are fused to condition a diffusion model, which generates object pose candidates and their corresponding energy. Finally, clustering is applied to address aggregation issues in the multimodal distribution of poses for objects with non-continuous symmetry, such as boxes, effectively resolving the pose estimation challenge." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.44, + 0.788, + 0.515 + ], + "angle": 0, + "content": "\\(\\mathrm{GenPose + + }\\) integrates two crucial enhancements: Semantic-aware feature extraction (see Fig.6 (a)) and Clustering-based aggregation (as shown in Fig.6 (c)). The subsequent sections will detail the three primary stages of the GenPose++ pipeline. Moreover, given the estimated 6D pose, \\(\\mathrm{GenPose + + }\\) provide a additional regression network to predict the 3D scale of the object." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.525, + 0.704, + 0.54 + ], + "angle": 0, + "content": "4.1 Training Semantic-aware Score and Energy Networks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.623 + ], + "angle": 0, + "content": "The learning agent is trained on our paired dataset \\(\\mathcal{D} = \\{(p_k,O_k,I_k)\\}_{k = 1}^n\\), where \\(\\pmb {p}_k\\in \\mathrm{SE}(3)\\), \\(O_{k}\\in \\mathbb{R}^{3\\times N}\\), and \\(I_{k}\\in \\mathbb{R}^{3\\times H\\times W}\\) denote a 6D pose, a partially observed 3D point cloud with \\(N\\) points, and a cropped RGB image with \\(H\\times W\\) resolution, respectively. Given an unseen object with point cloud \\(O^{*}\\) and RGB image \\(I^{*}\\), the goal is to recover the corresponding ground-truth pose \\(\\pmb{p}^{*}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.787, + 0.67 + ], + "angle": 0, + "content": "Following GenPose, we initially train a score network \\(\\varPhi_{\\theta}:\\mathbb{R}^{|\\mathcal{P}|}\\times \\mathbb{R}^{1}\\times \\mathbb{R}^{3\\times N}\\times\\) \\(\\mathbb{R}^{3\\times H\\times W}\\to \\mathbb{R}^{|P|}\\) and an energy network \\(\\varPsi_{\\phi}:\\mathbb{R}^{|\\mathcal{P}|}\\times \\mathbb{R}^{1}\\times \\mathbb{R}^{3\\times N}\\times \\mathbb{R}^{3\\times H\\times W}\\to \\mathbb{R}^{1}\\) from the dataset \\(\\mathcal{D}\\) using the denoising score-matching objective [26]:" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.679, + 0.787, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t \\sim \\mathcal {U} (\\epsilon , 1)} \\left\\{\\lambda (t) \\mathbb {E} _ {\\substack {\\boldsymbol {p} (0) \\sim p _ {\\text {data}} (\\boldsymbol {p} (0) | O, I), \\\\ \\boldsymbol {p} (t) \\sim \\mathcal {N} (\\boldsymbol {p} (t); \\boldsymbol {p} (0), \\sigma^ {2} (t) \\mathbf {I})}} \\left[ \\left\\| \\boldsymbol {s} (\\boldsymbol {p} (t), t | O, I) - \\frac {\\boldsymbol {p} (0) - \\boldsymbol {p} (t)}{\\sigma (t) ^ {2}} \\right\\| _ {2} ^ {2} \\right] \\right\\} \\tag{3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.729, + 0.788, + 0.761 + ], + "angle": 0, + "content": "The training loss of the score and energy network can be obtained by replacing \\( s(\\pmb{p}(t), t|O, I) \\) in Eq. 3 with \\( \\pmb{\\varPhi}_{\\theta}(\\pmb{p}(t), t|O, I) \\) and \\( \\nabla_{\\pmb{p}}\\pmb{\\Psi}_{\\phi}^{*}(\\pmb{p}(t), t|O, I) \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.788, + 0.851 + ], + "angle": 0, + "content": "Unlike GenPose, our score and energy network are semantic-aware, as both networks are conditioned on an RGB image to incorporate semantic cues for pose estimation. To fuse the features extracted from the image and point cloud, we encode the RGB image img and point cloud obj using the pre-trained feature extractors from DINOv2 [20] and PointNet++ [22], respectively. Then, we concatenate these features together in a pointwise manner similar to [27]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.27, + 0.114, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.631, + 0.162 + ], + "angle": 0, + "content": "4.2 Candidates Generation and Outlier Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Following GenPose, we subsequently sample pose candidates \\(\\{\\hat{p}_i\\}_{i=1}^K\\) by solving the following Probability Flow (PF) ODE [24] constructed by the score network \\(\\Phi_\\theta\\) from \\(t = 1\\) to \\(t = \\epsilon\\):" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.221, + 0.786, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\frac {d \\boldsymbol {p}}{d t} = - \\sigma (t) \\dot {\\sigma} (t) \\boldsymbol {\\Phi} _ {\\theta} (\\boldsymbol {p} (t), t | O, I) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.756, + 0.271 + ], + "angle": 0, + "content": "where \\(\\pmb{p}(1) \\sim \\mathcal{N}(\\mathbf{0}, \\sigma_{\\max}^2 \\mathbf{I})\\), \\(\\sigma(t) = \\sigma_{\\min} \\left( \\frac{\\sigma_{\\max}}{\\sigma_{\\min}} \\right)^t\\), \\(\\sigma_{\\min} = 0.01\\) and \\(\\sigma_{\\max} = 50\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.784, + 0.302 + ], + "angle": 0, + "content": "To remove the outliers in candidates, we sort the candidates into a sequence \\(\\hat{\\pmb{p}}_{\\tau_1} \\succ \\hat{\\pmb{p}}_{\\tau_2} \\dots \\succ \\hat{\\pmb{p}}_{\\tau_K}\\) where:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.311, + 0.786, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} _ {\\tau_ {i}} \\succ \\hat {\\boldsymbol {p}} _ {\\tau_ {j}} \\Longleftrightarrow \\boldsymbol {\\Psi} _ {\\phi} (\\hat {\\boldsymbol {p}} _ {\\tau_ {i}}, \\epsilon | O) > \\boldsymbol {\\Psi} _ {\\phi} (\\hat {\\boldsymbol {p}} _ {\\tau_ {j}}, \\epsilon | O) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.333, + 0.784, + 0.364 + ], + "angle": 0, + "content": "Then, we filter out the last \\(1 - \\delta \\%\\) candidates and obtain \\(\\hat{\\pmb{p}}_{\\tau_1} \\succ \\hat{\\pmb{p}}_{\\tau_2} \\dots \\succ \\hat{\\pmb{p}}_{\\tau_M}\\) where \\(\\delta = 40\\%\\) is a hyper parameter and \\(M = \\lfloor \\delta \\cdot K \\rfloor\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.376, + 0.507, + 0.39 + ], + "angle": 0, + "content": "4.3 Clustering-based Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.526 + ], + "angle": 0, + "content": "In this section, we aggregate the remaining candidates \\(\\{\\hat{\\pmb{p}}_{\\tau_i} = (\\hat{T}_{\\tau_i},\\hat{R}_{\\tau_i})\\}_{i = 1}^M\\) to obtain the final results. GenPose achieves this by simply mean-pooling the filtered candidates. However, this strategy will encounter a severe mean-mode issue when the object possesses discrete symmetrical properties. As illustrated in Fig. 6, a ping pong paddle has two symmetric ground truth poses (modes). Since the score network has encountered both modes during training, an optimally trained score network will likely output candidates around both modes. Simply mean-pooling these candidates will yield the average of the two modes, known as the 'mean mode', which will deviate from both modes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.664 + ], + "angle": 0, + "content": "To mitigate this issue, we introduce a clustering-based aggregation mechanism. We employ DBSCAN [8] to cluster the candidates. It identify dense regions in the data space, forming clusters based on the density of data points and effectively separating noise from meaningful patterns, without the need to specify the number of clusters. This is achieved through dynamic determination of cluster quantities based on distance threshold \\((\\varepsilon)\\) and density threshold (MinPts). For instance, in our empirical setting, we set \\(\\varepsilon \\approx 0.45\\) rad and MinPts \\(= 5\\). After clustering the candidates, we select the cluster with the largest number of objects and get the mean-pooling result as the final estimation following GenPose." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.376, + 0.693 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.618, + 0.714 + ], + "angle": 0, + "content": "5.1 Category-Level 6D Object Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Metric In prior studies, metrics such as the mean average precision (mAP) for 3D bounding box IoU and the mean average precision (mAP) for objects with translation errors less than \\( m \\) cm and rotation errors less than \\( n^{\\circ} \\) have been commonly used [28]. This evaluation typically involves two steps: instance-level object segmentation and pose estimation from these detections. However, these metrics are influenced by the detection model's performance. To focus solely on evaluating the precision of pose estimation, we assume ground truth instance segmentation is known and propose the following two metrics:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison of category-level pose estimation on ROPE. \\(\\uparrow\\) represents a higher value indicating better performance, while \\(\\downarrow\\) represents a lower value indicating better performance. Prior-free indicates whether the method requires category prior. \\{'-'\\) indicates that GenPose does not predict the object scale." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.207, + 0.778, + 0.324 + ], + "angle": 0, + "content": "
MethodInputPrior-freeAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
DeterministicNOCS [28]RGBD✓0.00.00.00.00.00.00.0
SGPA [3]RGBD✗10.52.00.04.36.79.315.0
IST-Net [16]RGBD✓28.710.60.52.03.45.38.8
HS-Pose [44]D✓31.613.61.13.55.38.412.7
ProbabilisticGenPose [43]D✓---6.69.613.119.3
GenPose++(Ours)RGBD✓39.019.12.010.015.119.529.4
" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.34, + 0.784, + 0.4 + ], + "angle": 0, + "content": "- AUC@IoU\\(_n\\): This metric assesses the accuracy of predicted 3D bounding boxes, calculated via the Area Under the Curve (AUC) from various Intersection over Union (IoU) thresholds starting at \\( n \\). In our study, we utilize AUC@IoU\\(_{25}\\), AUC@IoU\\(_{50}\\), and AUC@IoU\\(_{75}\\) as the benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.492 + ], + "angle": 0, + "content": "- VUS@\\(n^{\\circ}mcm\\): This metric offers a detailed analysis of 6D pose estimation accuracy, derived from the Volume Under Surface (VUS) across ranges of rotational (up to \\(n^{\\circ}\\)) and translational (up to \\(mcm\\)) error thresholds. VUS aggregates the accuracy of pose predictions within set boundaries. In this paper, we apply VUS@\\(5^{\\circ}2cm\\), VUS@\\(5^{\\circ}5cm\\), VUS@\\(10^{\\circ}2cm\\), and VUS@\\(10^{\\circ}5cm\\) for comprehensive performance assessment." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.504, + 0.785, + 0.55 + ], + "angle": 0, + "content": "Additionally, to demonstrate the robustness of our method to segmentation masks, we present the pose estimation performance of GenPose++ using SAM [13] segmentation results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.787, + 0.71 + ], + "angle": 0, + "content": "Baselines We evaluate five category-level pose estimation methods: NOCS [28], SGPA [3], HS-Pose [44], IST-Net [16] and GenPose [43]. Except for NOCS, which conducts both object detection and pose estimation as a whole, all methods are equipped with ground truth detection results. For SGPA, the prior point cloud of each category is constructed by randomly selecting an object from the training dataset. Considering that previous methods' augmentations for symmetry properties are only applicable to specific object categories within the NOCS dataset and not suitable for all object categories in Omni6Dpose, data augmentation for object symmetry is disabled during training for all baseline methods. All methods are trained on SOPE and directly tested on ROPE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results and Analysis. In Table 2, we present the quantitative evaluation results of previous methods compared to GenPose++ on the ROPE dataset. Overall, generative methods continue to dominate in the performance evaluation on ROPE. The VUS surface depicted in Figure 7 provides a more detailed reflection of the performance of each model. Unlike deterministic approaches, generative methods can handle ambiguity without any specialized design requirements. Moreover, these methods directly generate the distribution of object poses, eliminating the need for depth map-based pose fitting. This approach is particularly" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.391, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.147, + 0.572, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.147, + 0.788, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.249, + 0.39, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.249, + 0.573, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.249, + 0.787, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.352, + 0.744, + 0.367 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparison with baselines on ROPE dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.526 + ], + "angle": 0, + "content": "advantageous for challenging material types, such as transparent or reflective objects, where structured-light depth cameras tend to introduce significant noise, severely impacting pose fitting accuracy. Furthermore, the NOCS method does not demonstrate effective performance on the ROPE dataset, leading to the supposition that methods relying solely on RGB information to predict the shape of an object in the canonical space become less robust as the scale of category diversity increases. Compared to GenPose, GenPose++ achieves a significant lead by leveraging the powerful perception capabilities of the 2D foundation model, along with the robustness of clustering towards discrete symmetric properties. You can find qualitative visualizations in Figure 8." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.535, + 0.39, + 0.55 + ], + "angle": 0, + "content": "5.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.787, + 0.591 + ], + "angle": 0, + "content": "In order to validate the design decisions of our approach, we performed a series of ablation experiments on our method:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.632 + ], + "angle": 0, + "content": "- w/o clustering. Without clustering directly take the average of all remaining pose candidates after outlier removal as the pose estimation output." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.633, + 0.785, + 0.677 + ], + "angle": 0, + "content": "- w/o scale prediction. Use the estimated pose to transform the observed point cloud into object space, then take the bounding box length as the maximum projection from the point cloud to each axis." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.678, + 0.646, + 0.693 + ], + "angle": 0, + "content": "- w/o simulated depth. Use perfect depth for training." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.693, + 0.785, + 0.723 + ], + "angle": 0, + "content": "- w/o point-wise feature fusion. In the feature extraction stage, separately extract the RGB feature and geometric feature, and then concatenation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.602, + 0.785, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Table 3 illustrates the contribution of each component of GenPose++ to its performance. The introduction of the clustering module allows GenPose++ to effectively aggregate the multimodal distributions caused by discrete symmetries, leading to higher performance. The scale prediction in GenPose++ significantly surpasses direct calculations from the object's point cloud due to ambiguities from partial observations and errors from point cloud noise, particularly in transparent and reflective objects. Training with simulated depth data results in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.145, + 0.763, + 0.16 + ], + "angle": 0, + "content": "Table 3: Ablation study on category-level 6D object pose estimation" + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.165, + 0.739, + 0.289 + ], + "angle": 0, + "content": "
AblationAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
Full39.019.12.010.015.119.529.4
w/o scale prediction13.83.40.210.015.119.529.4
w/o clustering38.618.71.99.414.118.427.8
w/o simulated depth35.316.51.77.311.914.624.3
w/o point-wise fusion34.215.41.47.110.614.621.9
" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.302, + 0.785, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.468, + 0.744, + 0.482 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparison with baselines on ROPE dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.556 + ], + "angle": 0, + "content": "better performance than training with perfect point clouds, as the physics-based depth camera simulation substantially reduces the sim2real gap for depth data. Point-wise fusion outperforms global fusion as it retains more of the object's local geometric features, which are crucial for accurate object pose prediction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.564, + 0.602, + 0.58 + ], + "angle": 0, + "content": "5.3 Category-Level 6D Object Pose Tracking" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.587, + 0.772, + 0.603 + ], + "angle": 0, + "content": "Metric. We report the following metrics for object pose tracking evaluation:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.61, + 0.744, + 0.624 + ], + "angle": 0, + "content": "- FPS: Frames Per Second, which indicates the speed of pose tracking." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.625, + 0.785, + 0.653 + ], + "angle": 0, + "content": "- VUS@5°5cm: Volume Under Surface, assessing pose estimation accuracy for rotation errors within 0 to \\(5^{\\circ}\\) and translation errors within 0 to 5cm." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.654, + 0.785, + 0.683 + ], + "angle": 0, + "content": "- mIoU: Mean Intersection over Union, representing the average 3D overlap between the ground truth and the predicted bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.684, + 0.558, + 0.699 + ], + "angle": 0, + "content": "- \\(\\mathbf{Rerr}(\\circ)\\): Average rotation error in degrees." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.699, + 0.619, + 0.713 + ], + "angle": 0, + "content": "- \\(\\mathbf{Terr}(\\mathbf{cm})\\): Average translation error in centimeters." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.61, + 0.785, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.788, + 0.852 + ], + "angle": 0, + "content": "Baselines. This paper employs BundleTrack, CATRE, and GenPose as baselines for object pose tracking. BundleTrack is a training-free approach that utilizes multi-view feature point detection and matching for tracking the pose of unseen objects. CATRE aligns partially observed point clouds to abstract shape priors to estimate relative transformations, enhancing pose accuracy. Conversely, GenPose utilizes a generative approach to effectively resolve pose ambiguities, notably in symmetric objects. Following GenPose, we have adapted GenPose++ to object pose tracking with minor modifications." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.114, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Results and Analysis. Table 4 presents the results of category-level object pose tracking algorithms CATRE, GenPose, and our method, as well as the results for the unseen object pose tracking algorithm BundleTrack. The category-level pose estimation methods appear to achieve relatively better outcomes since they benefit from learning the category-level canonical space of objects within the SOPE dataset. However, for the training-free BundleTrack, reliance on RGB information for keypoint detection and matching poses challenges, often failing on objects with weak textures. Additionally, its dependency on depth values for global optimization renders it less effective in handling instances with substantial depth noise, such as transparent or specular objects. Our method, without undergoing specialized design, has achieved results comparable to state-of-the-art approaches. Although the inference speed of our method is lower than that of CATRE and GenPose, the achieved 17.8 FPS is sufficient for certain downstream tasks, such as robotic manipulation. Furthermore, recent rapid developments in research on fast samplers are beneficial to our approach, potentially enhancing its performance and applicability in real-time scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Table 4: Results of category-level object pose tracking on ROPE. The results are averaged over all 149 categories. GT. Pert. denotes that a perturbed ground truth pose is utilized as the initial object pose." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.45, + 0.779, + 0.535 + ], + "angle": 0, + "content": "
MethodsInputInit.Speed(FPS)↑5°5cm↑mIoU↑Rerr(°)↓Terr(cm)↓
BundleTrack [31]RGBDGT.12.41.33.946.923.5
CATRE [18]DGT. Pert.38.515.955.421.32.6
GenPose [43]DGT. Pert.26.313.3-19.31.2
GenPose++(Ours)RGBDGT. Pert.17.815.953.417.61.2
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.519, + 0.57 + ], + "angle": 0, + "content": "6 Conclusions and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.751 + ], + "angle": 0, + "content": "In this study, we introduce Omni6DPose, a comprehensive dataset for 6D object pose estimation, featuring extensive scale, diversity, and material variety. Through thorough experimentation, our findings suggest that the probabilistic framework holds promise for category-level 6D object estimation, leveraging semantic information provided by RGB images to address large-scale pose estimation challenges. However, the performance of GenPose++ on Omni6DPose reveals significant room for improvement, with the model still hampered by slow inference speeds resulting from the iterative refinement nature inherent in the diffusion model. Future works could focus on addressing these challenges and integrating the universal 6D pose estimation module trained on Omni6DPose into a broader range of downstream tasks [4,33,37,42]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.765, + 0.418, + 0.782 + ], + "angle": 0, + "content": "7 Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China - General Program (Project ID: 62376006), the National Youth Talent Support Program (Project ID: 8200800081), and NSFC-62306046." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. An, B., Geng, Y., Chen, K., Li, X., Dou, Q., Dong, H.: RGBmanip: Monocular image-based robotic manipulation through active object pose estimation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "4. Cheng, J., Wu, M., Zhang, R., Zhan, G., Wu, C., Dong, H.: Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.356 + ], + "angle": 0, + "content": "5. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4800-4810 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.412 + ], + "angle": 0, + "content": "6. Dai, Q., Zhang, J., Li, Q., Wu, T., Dong, H., Liu, Z., Tan, P., Wang, H.: Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and transparent objects. In: European Conference on Computer Vision. pp. 374-391. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.467 + ], + "angle": 0, + "content": "7. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.508 + ], + "angle": 0, + "content": "8. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: kdd. vol. 96, pp. 226-231 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.551, + 0.785, + 0.605 + ], + "angle": 0, + "content": "10. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. vol. 7724 (10 2012). https://doi.org/10.1007/978-3-642-33885-4_60" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "11. Hodań, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.648, + 0.785, + 0.702 + ], + "angle": 0, + "content": "12. Jung, H., Zhai, G., Wu, S.C., Ruhkamp, P., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., Roth, D., Navab, N., Busam, B.: Housecat6d - a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.757 + ], + "angle": 0, + "content": "13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.813 + ], + "angle": 0, + "content": "14. Lin, Z.H., Huang, S.Y., Wang, Y.C.F.: Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Liu, J., Sun, W., Liu, C., Zhang, X., Fan, S., Wu, W.: Hff6d: Hierarchical feature fusion network for robust 6d object pose tracking. IEEE Transactions on Circuits" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.763, + 0.129 + ], + "angle": 0, + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.174 + ], + "angle": 0, + "content": "and Systems for Video Technology pp. 1-13 (06 2022). https://doi.org/10.1109/TCSVT.2022.3181597" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.786, + 0.202 + ], + "angle": 0, + "content": "16. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. arXiv preprint arXiv:2303.13479 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.786, + 0.228 + ], + "angle": 0, + "content": "17. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.229, + 0.786, + 0.27 + ], + "angle": 0, + "content": "18. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: European Conference on Computer Vision (ECCV) (October 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.786, + 0.311 + ], + "angle": 0, + "content": "19. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics 22(12), 2633-2651 (2016). https://doi.org/10.1109/TVCG.2015.2513408" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.786, + 0.351 + ], + "angle": 0, + "content": "20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.786, + 0.404 + ], + "angle": 0, + "content": "21. Piga, N.A., Onyshchuk, Y., Pasquale, G., Pattacini, U., Natale, L.: Roft: Realtime optical flow-aided 6d object pose and velocity tracking. IEEE Robotics and Automation Letters 7(1), 159–166 (2022). https://doi.org/10.1109/LRA.2021.3119379" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.786, + 0.446 + ], + "angle": 0, + "content": "22. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.786, + 0.473 + ], + "angle": 0, + "content": "23. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.786, + 0.514 + ], + "angle": 0, + "content": "24. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.786, + 0.554 + ], + "angle": 0, + "content": "25. Tian, Y., Zhang, J., Huang, G., Wang, B., Wang, P., Pang, J., Dong, H.: Robokeygen: Robot pose and joint angles estimation via diffusion-based 3d keypoint generation. arXiv preprint arXiv:2403.18259 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.581 + ], + "angle": 0, + "content": "26. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.786, + 0.635 + ], + "angle": 0, + "content": "27. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.786, + 0.689 + ], + "angle": 0, + "content": "28. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.786, + 0.731 + ], + "angle": 0, + "content": "29. Wang, J., Chen, K., Dou, Q.: Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4807-4814. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.786, + 0.799 + ], + "angle": 0, + "content": "30. Wang, P., Jung, H., Li, Y., Shen, S., Srikanth, R.P., Garattoni, L., Meier, S., Navab, N., Busam, B.: Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21222-21231 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "31. Wen, B., Bekris, K.E.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6DPose" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "32. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 13209-13218 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "33. Wu, M., Zhong, F., Xia, Y., Dong, H.: Targf: Learning target gradient field to rearrange objects without explicit goal specification. Advances in Neural Information Processing Systems 35, 31986-31999 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "34. Wu, T., Wu, M., Zhang, J., Gan, Y., Dong, H.: Learning score-based grasping primitive for human-assisting dexterous grasping. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "35. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.786, + 0.37 + ], + "angle": 0, + "content": "36. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "37. Xue, T., Wu, M., Lu, L., Wang, H., Dong, H., Chen, B.: Learning gradient fields for scalable and generalizable irregular packing. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.494, + 0.787, + 0.535 + ], + "angle": 0, + "content": "40. You, Y., Shi, R., Wang, W., Lu, C.: CPPf: Towards robust category-level 9d pose estimation in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6866-6875 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "41. You, Y., Xiong, K., Yang, Z., Huang, Z., Zhou, J., Shi, R., Fang, Z., Harley, A.W., Lu, C.: Pace: Pose annotations in cluttered environments. arXiv preprint arXiv:2312.15130 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "42. Zeng, Y., Wu, M., Yang, L., Zhang, J., Ding, H., Cheng, H., Dong, H.: Distilling functional rearrangement priors from large models. arXiv preprint arXiv:2312.01474 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.619, + 0.787, + 0.646 + ], + "angle": 0, + "content": "43. Zhang, J., Wu, M., Dong, H.: Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.702 + ], + "angle": 0, + "content": "44. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17163-17173 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "45. Zhou, L., Liu, Z., Gan, R., Wang, H., au2, M.H.A.J.: Dr-pose: A two-stage deformation-and-registration pipeline for category-level 6d object pose estimation (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_origin.pdf b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd4a0a781a6069fa311680a59e048113720d4014 --- /dev/null +++ b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/cf5501e2-725b-4f12-bc02-1c29948756c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bf0bae7c18f6f40926d4d763177fc22a2e0badd54ed6516e6ef5115c124d5ed +size 22570725 diff --git a/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/full.md b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..782b0603ce2d442e2a3be8b50174fc51c6b5a66e --- /dev/null +++ b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/full.md @@ -0,0 +1,329 @@ +# Omni6DPose: A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking + +Jiyao Zhang $^{1,2,3*}$ , Weiyao Huang $^{1*}$ , Bo Peng $^{1*}$ , Mingdong Wu $^{1,2,3}$ , Fei Hu $^{1,2,3}$ , Zijian Chen $^{4}$ , Bo Zhao $^{5,6}$ , and Hao Dong $^{1,2,3\dagger}$ + +$^{1}$ Center on Frontiers of Computing Studies, School of Computer Science, Peking University $^{2}$ PKU-Agibot Lab, School of Computer Science, Peking University +$^{3}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University $^{4}$ Waseda University $^{5}$ Shanghai Jiao Tong University $^{6}$ BAAI {jiyaozhang, sshwy, bo.peng, hofee}@stu.pku.edu.cn zijianchenwaseda@akane.waseda.jp bo.zhao@sjtu.edu.cn {wmingd, hao.dong}@pku.edu.cn + +https://jiyao06.github.io/Omni6DPose/ + +Abstract. 6D object pose estimation is crucial in the field of computer vision. However, it suffers from a significant lack of large-scale and diverse datasets, impeding comprehensive model evaluation and curtailing downstream applications. To address these issues, this paper introduces Omni6DPose, a substantial benchmark featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is divided into three main components: ROPE (Real 6D Object Pose Estimation Dataset), which includes 332K images annotated with over 1.5M annotations across 581 instances in 149 categories; SOPE(Simulated 6D Object Pose Estimation Dataset), a simulated training set created by mixed reality and physics-based depth simulation; and PAM(Pose Aligned 3D Models), the manually aligned real scanned objects used in ROPE and SOPE. Omni6DPose is inherently challenging due to the substantial variations and ambiguities. To address this issue, we introduce GenPose++, an enhanced version of the SOTA category-level 6D object pose estimation framework, incorporating two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation. Moreover, we provide a comprehensive benchmarking analysis to evaluate the performance of previous methods on this new large-scale dataset in the realms of 6D object pose estimation and pose tracking. + +Keywords: Benchmark $\cdot$ object pose estimation $\cdot$ object pose tracking + +# 1 Introduction + +6D object pose estimation [16,28,45] and pose tracking [15,21] from single images is an essential task in computer vision, holding immense potential for applications in robotics [1] and augmented reality/virtual reality (AR/VR) [19]. Over recent decades, the domain has experienced significant advancements, primarily dominated by data-driven learning approaches. Analogous to the pivotal role of data in learning-based 2D foundation tasks, high-quality, comprehensive datasets are paramount in the context of 6D object pose estimation and tracking. + +![](images/6e30f9fb1ef9101e65707ac7f168841fdf416464f146a4f76a392dff6c0d76ea.jpg) +Fig. 1: We introduce Omni6DPose for universal 6D object pose estimation and tracking, which comprises three key components: (1) ROPE (right), a large-scale real-world dataset for evaluation. (2) SOPE (left), a vast synthetic dataset generated through mixed reality and physics-based simulation for training. (3) PAM (center), a collection of pose-aligned 3D object models. Omni6DPose is featured by its diversity in object categories, large scale, and variety in object materials. + +Today 6D object pose estimation is studied under two lenses: instance-level and category-level. In instance-level settings, datasets such as Linemod [10], YCB-Video [36], and T-LESS [11] have gained widespread acceptance as benchmarks. These datasets are distinguished by their focus on detailed, individual object instances, thereby enabling algorithms to precisely learn and predict the poses of specific items. On the other hand, category-level pose estimation emphasizes generalization across different items within a particular object category. The NOCS [28] dataset stands out as the most widely used in the category-level object pose estimation field, providing a simulated dataset for training and a small-scale real-world dataset for evaluation. Despite their contributions to advancing the field, these datasets present limitations due to their small scale in terms of instances or categories. This results in two significant challenges: + +1. It hampers comprehensive evaluation of different models' performance, limiting the development of research in this field. +2. It restricts the applicability of research findings across diverse domains, due to the limited variety of object instances or categories represented. + +To address the aforementioned challenges and drive advancements in this field, this paper introduces Omni6DPose, a universal 6D object pose estimation dataset featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is segmented into three principal components: (1) ROPE (Real 6D Object Pose Estimation Dataset), which encompasses $332\mathrm{K}$ images annotated with over 1.5M annotations across 581 instances in 149 categories; (2) SOPE (Simulated 6D Object Pose Estimation Dataset), comprising $475\mathrm{K}$ images generated in a mixed reality setting with depth simulation, furnished with over 5M annotations spanning 4162 instances in the same 149 + +![](images/1a60111597c270923e5a9fe2d9337ff4c25a52767811667ea1190b636c2b590f.jpg) + +![](images/e30d062643149c774e69ec1aee782198766bc6e4587e0098a9fa42d0a9e1cac6.jpg) + +![](images/43275bc3b8beb29183c70ae36d9adae17d40d064a09a49bb097a67b6f133b781.jpg) +Fig. 2: ROPE dataset visualization. In the figure, bounding boxes are colored according to the coordinates in the object's coordinate system. + +![](images/44ff07bbb3552f46ea54d5355b86627c6238da0b82a64dc8ea16bf8033463a1d.jpg) + +![](images/6468a5f50a56759c8b1bbf03e1aa1b915eb71628c75cdcf150bdc344d9cccb97.jpg) + +![](images/57fea1551a53410447966c9d8b602271c459fccd8967a553a84fabb67a4623ed.jpg) + +![](images/b8808cb50f283d7c2f4884717cd265866e36cf74011e315d21131af851534fce.jpg) + +![](images/172fd19562b74c51f5ab7fa2c961c534695d45e685e2ea5611bb5ebf4f21b840.jpg) + +![](images/5b423517e481a3550e627b5e8ada15cd45c58163646cad585cb6f5cab3e6e0f8.jpg) + +![](images/cfc55b1d4839de861b119a85310341fb2d3146b2b9b25b746da1388d41fadb75.jpg) + +![](images/2996098182c76da3d4a14b8593baa467e1c4312c27a59b8b6e9fdb2fbb194d5d.jpg) + +![](images/f89108a76ee285ad48ac784e5c47e30101301b38ef746d7f5687b12476d37d37.jpg) + +![](images/9b4943baaa6d025c9201974d559431e8719b9631d0841cc21f6dd49a6c98717b.jpg) + +![](images/7511dc9f5d88aada149176f9573656101fa178c98e0e1b8986d4678451f1b6cf.jpg) + +![](images/5eeb4a4d92a5496ce3e870202d7eb7a7e2a8f2723c6af499c5d3209a846479de.jpg) + +categories. The mixed reality bridges the semantic sim2real gap, while the depth sensor simulation closes the geometric sim2real gap; (3) PAM(Pose Aligned 3D Models), the manually aligned, real scanned objects utilized in both ROPE and SOPE, enabling the generation of diverse downstream task data. + +Omni6DPose poses inherent challenges due to its considerable variations, diverse materials, and inherent ambiguities, which reflect the complexities encountered in real-world applications. Figure 2 illustrates some examples from ROPE. To tackle these issues, we introduce GenPose++, which incorporates GenPose [43] with two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation, tailored specifically to the nuances of the Omni6DPose in question. Furthermore, as a Universal 6D object pose estimation dataset, this paper also offers a comprehensive benchmarking analysis to assess the performance of existing methods on category-level 6D object pose estimation and pose tracking. We summarize our contributions as follows: + +1. We present Omni6DPose, a comprehensive 6D object pose estimation dataset with extensive categories, instance diversity, and material variety. +2. We propose a real data collection pipeline and a simulation framework for generating synthetic data with low semantic and geometry sim2real gaps. +3. We introduce GenPose++ for category-level 6D object pose estimation and tracking, demonstrating SOTA performance on Omni6DPose. + +# 2 Related work + +For 6D object pose estimation, there are two main branches: instance-level and category-level. Instance-level estimation is tested on seen objects or on unseen objects with reference, while category-level estimation is tested on unseen instances of known categories without CAD models. In this section, we review and compare existing datasets to our large-scale category-level dataset and review relevant algorithms for category-level pose estimation and tracking. + +# 2.1 6D Object Pose Estimation Datasets + +Following the outlined branches of 6D object pose estimation, we have reviewed datasets corresponding to both instance-level and category-level 6D object pose estimation. A comparative analysis of these datasets is provided in Table 1. + +Table 1: This table compares datasets for 6D object pose estimation, focusing on object category count, reality of the data, data modalities (RGB, Depth, IR), and object attributes such as quantity, CAD model availability, and inclusion of transparent and specular objects. It also details video characteristics by number and marker presence, along with image and annotation counts. 'Wild6D' refers specifically to the test split of the Wild6D dataset, as the training data does not provide annotations. The symbol '-' indicates the absence of a particular feature within the dataset. + +
DatasetCat. RealModalityObjectMarker-free Vid.Img.Anno.
RGBDepth IRIRNum. CAD Trans. Spec.
CAMERA [28]6X✓✓X1085✓XX✓-300K
SOPE(Ours)149X✓✓✓4162✓✓✓✓-475K
YCB-Video [36]-✓✓✓X21✓XX✓92133K
T-LESS [11]-✓✓✓X30✓XXX2048K
Linemod [10]-✓✓✓X15✓XXX-18K
StereoOBJ-1M [17]-✓✓XX18✓✓✓X182393K
REAL275 [28]6✓✓✓X42✓XXX188K
PhoCal [30]8✓✓✓X60✓✓✓✓243.9K
HouseCat6D [12]10✓✓✓X194✓✓✓✓4123.5K
Wild6D* [9]5✓✓✓X162X(✓)X✓48610K
PACE [41]44✓✓✓X576✓XX✓30055K
ROPE(Ours)149✓✓✓✓581✓✓✓✓363332K
+ +Instance-level 6D object pose estimation dataset. LineMod [10] is one of the most used datasets, providing non-temporal RGB-D images and ground truth pose annotations. YCB-Video [36] provides RGB-D videos and annotations, enabling pose-tracking approaches. T-LESS [11] features texture-less objects with symmetries and mutual similarities. StereoOBJ-1M [17] achieves a leap in dataset scale and features transparent and reflective objects. While these datasets are extensive in terms of image and annotation count, they are limited in the diversity of instances they cover. For example, StereoOBJ-1M comprises 339K frames and 1.5M annotations, yet it includes only 18 unique object instances. + +Category-level 6D object pose estimation dataset. NOCS [28] provides the first benchmark in category-level pose estimation, while Wild6D [9] addresses the scalability issue of datasets by leveraging unlabeled and synthetic data. PhoCal [30] focuses on photometrically challenging objects and HouseCat6D [12] offers diverse scenes, viewpoints, and grasping annotations. However, these datasets cover only a limited number of categories, even the most extensive dataset, HouseCat6D, includes merely ten categories. PACE [41], a concurrent work, focusing on object pose estimation in complex scenes. In contrast, our datasets, SOPE and ROPE, set new benchmarks by offering the widest range of categories and featuring objects with diverse materials, thereby enhancing dataset diversity and realism for pose estimation research. + +# 2.2 6D Object Pose Estimation and Tracking Algorithm + +Category-level 6D object pose estimation. Category-level 6D object pose estimation [28,39,40] aims to estimate unseen instance poses within the same category. NOCS [28] introduces a normalized object coordinate space for pose prediction without CAD models, while SPD [29] and SGPA [3] utilize category-level + +prriors for enhanced estimation. HS-Pose [44] extends 3D-GC [14] for improved feature extraction from point clouds. IST-Net [16] transforms features between camera and world spaces implicitly, without relying on 3D priors, surpassing previous methods. However, these techniques, mainly regression-based, require ad-hoc designs for symmetric objects. GenPose [43] addresses it by generatively modeling the pose distribution, eliminating the need for symmetry considerations. Yet, GenPose neglects RGB semantic information, which is increasingly vital as object category scales grow. Additionally, its energy-based aggregation algorithm fails with discontinuous multimodal distributions. We introduce GenPose++, which incorporates a 2D foundation model to leverage RGB semantics, improving generalization, and introduces an aggregation module to handle discrete multimodal distributions, addressing the limitations of GenPose. + +6D object pose tracking. This paper is situated within the domain of category-level 6D object pose tracking and model-free object tracking. BundleTrack [31] pioneers model-free tracking by leveraging multi-view feature detection for tracking unseen objects without 3D models. CAPTRA [32] enhances articulated pose tracking through recursive updates for better temporal consistency. CATRE [18] aligns partially observed point clouds to abstract shape priors for relative transformations and pose estimation. GenPose, adopting a generative approach, effectively addresses the challenge of pose ambiguities in symmetric objects. Together, these approaches underscore the evolving landscape of object pose tracking, highlighting both the progress and the diversity of strategies being explored. + +# 3 Omni6DPose Dataset + +This paper introduces a rich variety of object categories, a large-scale, and materially diverse dataset for real 6D object pose estimation, named ROPE. And, a simulated dataset, SOPE, synthesizing with mixed reality and featuring depth simulation, is provided for training. Section 3.1 will discuss the collection and alignment of 3D objects. Section 3.2 will cover the acquisition and labeling of the ROPE dataset. Section 3.3 will explain the generation of the SOPE dataset. + +# 3.1 3D Object Collection and Alignment + +Universal 6D object pose estimation relies on a comprehensive set of objects. We selected 149 categories of everyday objects, all reconstructed with high-precision scanners, and categorized them into two sets: SOPE for simulated data and ROPE for real-world scenes. SOPE primarily includes objects from sources like OmniObject3D [35], PhoCal [30], and GoogleScan [7], alongside a subset from our scans, totaling 5,000 instances. ROPE consists of 580 instances we reconstructed using industrial scanners. Importantly, while most SOPE objects are from public datasets, manual category-level pose alignment is necessary. For object reconstruction, as shown in Figure 3, we use two professional scanners, EinScan H2 $^{7}$ and Revopoint POP 3 $^{8}$ for objects in different scales. The scan + +![](images/ee72ddeb9dd614b8107981b5074fea0e9e9d96b84d2dc124a17f7cf47d4d842b.jpg) +Fig. 3: ROPE dataset collection and annotation. (1) Object scanning, where high-precision industrial scanners are used to acquire the CAD models of objects; (2) Object canonicalization, involving the alignment of each object category to the canonical space; (3) Video capture, capturing video sequences in varied scenarios with a depth camera; and (4) Pose annotation, calculating camera poses through Structure from Motion (SFM), further utilizing Farthest Point Sampling (FPS) to select keyframes for keypoint annotation, and performing bundle adjustment to derive initial object pose values, which are then manually refined to obtain more precise annotations. + +![](images/fdeea2904a14f8872961b09ecaedf36c35b31eb8abc2fe5c48f6bd60ff54a8c1.jpg) + +ning time depends on object features: it took about 15 minutes for small, simple, Lambertian items like a mouse, and up to an hour for complex, large, or non-Lambertian items like a transparent mug. Finally, we constructed a specialized annotation tool for manually aligning objects in the same category to the category-level canonical space, with each alignment taking roughly one minute. + +# 3.2 ROPE Acquisition and Annotation + +The ROPE dataset was systematically acquired utilizing the RealSense D415 imaging device, encompassing scenarios with 2 to 6 distinct objects and video sequences extending from 762 to 1,349 frames. The integrity and utility of the dataset are underpinned by the precision of object pose annotations, which present notable challenges, chiefly: + +1. Derivation of relative camera poses, denoted as $T_{c} = \{(R_{i},t_{i})\}_{i = 1}^{n}$ , where each pair $(R_i,t_i)\in \mathrm{SE}(3)$ , signifying the transformation from the camera space to the world space for the $i^{th}$ frame, with $n$ symbolizing the aggregate frame count within the video sequence. +2. Procurement of high-accuracy object poses, represented as $T_{o} = \{(R,t)\}$ , where the pair $(R,t) \in \mathrm{SE}(3)$ , delineating the transformation from the object space the camera space for any selected frame. + +With $T_{c}$ and $T_{o}$ known, it is possible to automate the generation of all annotations within the dataset. Addressing these challenges, we propose a marker-free annotation system. Previous datasets for calculating relative camera pose rely on markers, like NOCS [28], or external robot arms for indirect calculation, such as PhoCal [30], limiting scene diversity. As shown in Figure 3, to enable marker-free annotation in open scenes, we consider it a structure-from-motion (SfM) problem, utilizing intrinsic scene features to optimize camera poses $T_{c}$ using bundle adjustment techniques. This approach aims to solve the optimization problem: + +![](images/ee6825a0c6cd87d43cf531eb473786b2d5b4fc4606b3b301b1180875f4d4c284.jpg) +Fig. 4: SOPE synthesis, utilizing mixed reality to bridge the RGB sim2real gap and physical-based depth sensor simulation to minimize the geometric sim2real gap. + +![](images/b6208e949c4c1e03ff6fa2b7e8f0a535277ba4674ccd460c5d24793674b3bd92.jpg) + +$$ +\min _ {T _ {c}} \sum_ {i = 1} ^ {n} \sum_ {j \in \mathcal {P} _ {i}} \| \pi \left(R _ {i} X _ {j} + t _ {i}\right) - x _ {i j} \| ^ {2} \tag {1} +$$ + +where $\pi$ denotes the camera projection function, $X_{j}$ represents the 3D points in the world space, and $x_{ij}$ corresponds to the 2D projection of $X_{j}$ in the $i^{th}$ camera frame, with $\mathcal{P}_i$ being the set of point correspondences in frame $i$ . + +For object pose $T_{o}$ annotations, previous methods consider only a single frame, leading to inaccuracies due to the lack of multi-viewpoint constraints. To overcome this, we introduce a two-stage object pose annotation process: in the first stage, keyframes are sampled from SfM results using Farthest Point Sampling (FPS), and 2D-3D keypoint pairs on these keyframes are annotated, providing initial object pose by minimizing the reprojection error: + +$$ +\min _ {R, t} \sum_ {k \in \mathcal {K}} \| \pi \left(R X _ {k} + t\right) - x _ {k} \| ^ {2} \tag {2} +$$ + +where $X_{k}$ are 3D keypoints of object, $x_{k}$ are corresponding 2D annotations in the image, and $\mathcal{K}$ is the set of all keypoint correspondences. The initial poses are manually fine-tuned based on the object's projection across all keyframes. + +# 3.3 SOPE Synthesis + +ROPE represents a comprehensive benchmark in category-level object pose estimation by scaling up the diversity and number of object categories to unprecedented levels, encompassing a wide range of materials. This diversity presents new challenges for network training data due to the higher demands on the dataset's scale and diversity. Collecting a larger real-world dataset would be prohibitively expensive and unlikely to ensure sufficient diversity. + +To bridge the sim2real gap, which is pronounced when using synthetic data, either in RGB or geometry, this paper proposes a novel method based on mixed reality with depth simulation for synthetic data generation. Specifically, as demonstrated in Figure 4, we employ mixed reality [28] techniques to generate RGB data, thereby reducing the RGB sim2real gap. In parallel, we simulate the mechanism of structured light depth sensors within blender [6]. This involves rendering infrared (IR) images and applying stereo matching to produce synthetic depth maps, effectively narrowing the geometry's sim2real gap. + +![](images/b24d7864e681bc47c2a299d2a477da0236290482111c0d79cb050ffe4dbd8173.jpg) +Fig. 5: Omni6DPose statistics, showcasing the dataset distribution. Left: Category distribution, highlighting 149 categories and diverse materials. Right: Object size distribution across 5000 objects, illustrating diversity in shapes. + +![](images/a8960cd64b7036bd10197f1ff8903d7c074e3fd6859f78cfb271cb3f671c0fc7.jpg) + +During data generation, we implement domain randomization for illumination and object materials to further enhance the dataset's diversity. All the background images are sourced from public datasets, including 19,658 images from MatterPort3D [2], 2,572 from ScanNet++ [38], and 540 from IKEA [28]. To the best of our knowledge, this is the first simulated dataset that uses a context-aware mixed reality approach combined with physical-based depth sensor simulation for object pose estimation tasks. + +# 3.4 Dataset Statistics + +Object Category Statistics The comprehensive distribution of object category and size are both demonstrated in Figure 5. Most of the categories possess $\geq 25\mathrm{K}$ pose annotations in the SOPE dataset, providing sufficient training opportunities. Categories containing objects with diverse and challenging material options (e.g., transparent or specular materials) are equipped with apparently more data generation, such as dishes, cups, bottles, bowls, mugs, etc. + +Object Size Statistics As shown in Figure 5, the objects in our dataset span a wide range of sizes. The majority of the objects are approximately 0.1 meters in length along the diagonal of their bounding boxes, with the largest objects exceeding 1 meter. + +# 4 Category-level 6D Pose Estimation Method + +Given the Omni6DPose, one naturally ponders the optimal technical approach for large-scale category-level pose estimation. The recently introduced state-of-the-art category-level 6D pose estimation technique, GenPose [43], offers a promising avenue by employing a diffusion-based probabilistic method [23, 24]. Moreover, the diffusion model has demonstrated remarkable efficacy across various high-dimensional domains with extensive training data [5, 25, 34, 42]. Expanding on this groundwork, our study delves further into the probabilistic approach, presenting an enhanced iteration of GenPose, named GenPose++. + +![](images/bd6667f41f34f9d181debef2094ec10cadf3934d738b094a7bbef93d5a365440.jpg) +Fig. 6: Overview of GenPose++. We employ segmented point clouds and cropped RGB images as inputs. PointNet++ extracts geometric features, while DINO v2, a pre-trained 2D foundation model, extracts semantic features. These features are fused to condition a diffusion model, which generates object pose candidates and their corresponding energy. Finally, clustering is applied to address aggregation issues in the multimodal distribution of poses for objects with non-continuous symmetry, such as boxes, effectively resolving the pose estimation challenge. + +$\mathrm{GenPose + + }$ integrates two crucial enhancements: Semantic-aware feature extraction (see Fig.6 (a)) and Clustering-based aggregation (as shown in Fig.6 (c)). The subsequent sections will detail the three primary stages of the GenPose++ pipeline. Moreover, given the estimated 6D pose, $\mathrm{GenPose + + }$ provide a additional regression network to predict the 3D scale of the object. + +# 4.1 Training Semantic-aware Score and Energy Networks + +The learning agent is trained on our paired dataset $\mathcal{D} = \{(p_k,O_k,I_k)\}_{k = 1}^n$ , where $\pmb {p}_k\in \mathrm{SE}(3)$ , $O_{k}\in \mathbb{R}^{3\times N}$ , and $I_{k}\in \mathbb{R}^{3\times H\times W}$ denote a 6D pose, a partially observed 3D point cloud with $N$ points, and a cropped RGB image with $H\times W$ resolution, respectively. Given an unseen object with point cloud $O^{*}$ and RGB image $I^{*}$ , the goal is to recover the corresponding ground-truth pose $\pmb{p}^{*}$ . + +Following GenPose, we initially train a score network $\varPhi_{\theta}:\mathbb{R}^{|\mathcal{P}|}\times \mathbb{R}^{1}\times \mathbb{R}^{3\times N}\times$ $\mathbb{R}^{3\times H\times W}\to \mathbb{R}^{|P|}$ and an energy network $\varPsi_{\phi}:\mathbb{R}^{|\mathcal{P}|}\times \mathbb{R}^{1}\times \mathbb{R}^{3\times N}\times \mathbb{R}^{3\times H\times W}\to \mathbb{R}^{1}$ from the dataset $\mathcal{D}$ using the denoising score-matching objective [26]: + +$$ +\mathbb {E} _ {t \sim \mathcal {U} (\epsilon , 1)} \left\{\lambda (t) \mathbb {E} _ {\substack {\boldsymbol {p} (0) \sim p _ {\text {data}} (\boldsymbol {p} (0) | O, I), \\ \boldsymbol {p} (t) \sim \mathcal {N} (\boldsymbol {p} (t); \boldsymbol {p} (0), \sigma^ {2} (t) \mathbf {I})}} \left[ \left\| \boldsymbol {s} (\boldsymbol {p} (t), t | O, I) - \frac {\boldsymbol {p} (0) - \boldsymbol {p} (t)}{\sigma (t) ^ {2}} \right\| _ {2} ^ {2} \right] \right\} \tag{3} +$$ + +The training loss of the score and energy network can be obtained by replacing $s(\pmb{p}(t), t|O, I)$ in Eq. 3 with $\pmb{\varPhi}_{\theta}(\pmb{p}(t), t|O, I)$ and $\nabla_{\pmb{p}}\pmb{\Psi}_{\phi}^{*}(\pmb{p}(t), t|O, I)$ , respectively. + +Unlike GenPose, our score and energy network are semantic-aware, as both networks are conditioned on an RGB image to incorporate semantic cues for pose estimation. To fuse the features extracted from the image and point cloud, we encode the RGB image img and point cloud obj using the pre-trained feature extractors from DINOv2 [20] and PointNet++ [22], respectively. Then, we concatenate these features together in a pointwise manner similar to [27]. + +# 4.2 Candidates Generation and Outlier Removal + +Following GenPose, we subsequently sample pose candidates $\{\hat{p}_i\}_{i=1}^K$ by solving the following Probability Flow (PF) ODE [24] constructed by the score network $\Phi_\theta$ from $t = 1$ to $t = \epsilon$ : + +$$ +\frac {d \boldsymbol {p}}{d t} = - \sigma (t) \dot {\sigma} (t) \boldsymbol {\Phi} _ {\theta} (\boldsymbol {p} (t), t | O, I) \tag {4} +$$ + +where $\pmb{p}(1) \sim \mathcal{N}(\mathbf{0}, \sigma_{\max}^2 \mathbf{I})$ , $\sigma(t) = \sigma_{\min} \left( \frac{\sigma_{\max}}{\sigma_{\min}} \right)^t$ , $\sigma_{\min} = 0.01$ and $\sigma_{\max} = 50$ . + +To remove the outliers in candidates, we sort the candidates into a sequence $\hat{\pmb{p}}_{\tau_1} \succ \hat{\pmb{p}}_{\tau_2} \dots \succ \hat{\pmb{p}}_{\tau_K}$ where: + +$$ +\hat {\boldsymbol {p}} _ {\tau_ {i}} \succ \hat {\boldsymbol {p}} _ {\tau_ {j}} \Longleftrightarrow \boldsymbol {\Psi} _ {\phi} (\hat {\boldsymbol {p}} _ {\tau_ {i}}, \epsilon | O) > \boldsymbol {\Psi} _ {\phi} (\hat {\boldsymbol {p}} _ {\tau_ {j}}, \epsilon | O) \tag {5} +$$ + +Then, we filter out the last $1 - \delta \%$ candidates and obtain $\hat{\pmb{p}}_{\tau_1} \succ \hat{\pmb{p}}_{\tau_2} \dots \succ \hat{\pmb{p}}_{\tau_M}$ where $\delta = 40\%$ is a hyper parameter and $M = \lfloor \delta \cdot K \rfloor$ . + +# 4.3 Clustering-based Aggregation + +In this section, we aggregate the remaining candidates $\{\hat{\pmb{p}}_{\tau_i} = (\hat{T}_{\tau_i},\hat{R}_{\tau_i})\}_{i = 1}^M$ to obtain the final results. GenPose achieves this by simply mean-pooling the filtered candidates. However, this strategy will encounter a severe mean-mode issue when the object possesses discrete symmetrical properties. As illustrated in Fig. 6, a ping pong paddle has two symmetric ground truth poses (modes). Since the score network has encountered both modes during training, an optimally trained score network will likely output candidates around both modes. Simply mean-pooling these candidates will yield the average of the two modes, known as the 'mean mode', which will deviate from both modes. + +To mitigate this issue, we introduce a clustering-based aggregation mechanism. We employ DBSCAN [8] to cluster the candidates. It identify dense regions in the data space, forming clusters based on the density of data points and effectively separating noise from meaningful patterns, without the need to specify the number of clusters. This is achieved through dynamic determination of cluster quantities based on distance threshold $(\varepsilon)$ and density threshold (MinPts). For instance, in our empirical setting, we set $\varepsilon \approx 0.45$ rad and MinPts $= 5$ . After clustering the candidates, we select the cluster with the largest number of objects and get the mean-pooling result as the final estimation following GenPose. + +# 5 Experiments + +# 5.1 Category-Level 6D Object Pose Estimation + +Metric In prior studies, metrics such as the mean average precision (mAP) for 3D bounding box IoU and the mean average precision (mAP) for objects with translation errors less than $m$ cm and rotation errors less than $n^{\circ}$ have been commonly used [28]. This evaluation typically involves two steps: instance-level object segmentation and pose estimation from these detections. However, these metrics are influenced by the detection model's performance. To focus solely on evaluating the precision of pose estimation, we assume ground truth instance segmentation is known and propose the following two metrics: + +Table 2: Quantitative comparison of category-level pose estimation on ROPE. \(\uparrow\) represents a higher value indicating better performance, while \(\downarrow\) represents a lower value indicating better performance. Prior-free indicates whether the method requires category prior. \{'-'\) indicates that GenPose does not predict the object scale. + +
MethodInputPrior-freeAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
DeterministicNOCS [28]RGBD✓0.00.00.00.00.00.00.0
SGPA [3]RGBD✗10.52.00.04.36.79.315.0
IST-Net [16]RGBD✓28.710.60.52.03.45.38.8
HS-Pose [44]D✓31.613.61.13.55.38.412.7
ProbabilisticGenPose [43]D✓---6.69.613.119.3
GenPose++(Ours)RGBD✓39.019.12.010.015.119.529.4
+ +- AUC@IoU $_n$ : This metric assesses the accuracy of predicted 3D bounding boxes, calculated via the Area Under the Curve (AUC) from various Intersection over Union (IoU) thresholds starting at $n$ . In our study, we utilize AUC@IoU $_{25}$ , AUC@IoU $_{50}$ , and AUC@IoU $_{75}$ as the benchmarks. +- VUS@ $n^{\circ}mcm$ : This metric offers a detailed analysis of 6D pose estimation accuracy, derived from the Volume Under Surface (VUS) across ranges of rotational (up to $n^{\circ}$ ) and translational (up to $mcm$ ) error thresholds. VUS aggregates the accuracy of pose predictions within set boundaries. In this paper, we apply VUS@ $5^{\circ}2cm$ , VUS@ $5^{\circ}5cm$ , VUS@ $10^{\circ}2cm$ , and VUS@ $10^{\circ}5cm$ for comprehensive performance assessment. + +Additionally, to demonstrate the robustness of our method to segmentation masks, we present the pose estimation performance of GenPose++ using SAM [13] segmentation results in the supplementary material. + +Baselines We evaluate five category-level pose estimation methods: NOCS [28], SGPA [3], HS-Pose [44], IST-Net [16] and GenPose [43]. Except for NOCS, which conducts both object detection and pose estimation as a whole, all methods are equipped with ground truth detection results. For SGPA, the prior point cloud of each category is constructed by randomly selecting an object from the training dataset. Considering that previous methods' augmentations for symmetry properties are only applicable to specific object categories within the NOCS dataset and not suitable for all object categories in Omni6Dpose, data augmentation for object symmetry is disabled during training for all baseline methods. All methods are trained on SOPE and directly tested on ROPE. + +Results and Analysis. In Table 2, we present the quantitative evaluation results of previous methods compared to GenPose++ on the ROPE dataset. Overall, generative methods continue to dominate in the performance evaluation on ROPE. The VUS surface depicted in Figure 7 provides a more detailed reflection of the performance of each model. Unlike deterministic approaches, generative methods can handle ambiguity without any specialized design requirements. Moreover, these methods directly generate the distribution of object poses, eliminating the need for depth map-based pose fitting. This approach is particularly + +![](images/8577d558e9ccc200a47273e9044677f58ad2053039a13285e0607224c4d64cc9.jpg) + +![](images/7b2051dcc5293d9a7fe0796028210d31c06e74c9b9890724cda7329cb842a01e.jpg) + +![](images/c077af05b498de02cdc8eab0c596a7d8c21efe75484d6d0273e29d17bcfd9d10.jpg) + +![](images/fec35d4ed07e2909d968f20e09d0c1a7fd9731396d3a649319ce5e8374292dad.jpg) +Fig. 7: Qualitative comparison with baselines on ROPE dataset. + +![](images/03ce1c66280d01f8a3c7411edbf255b53f6b004a790450d12fe12ac5dee565f6.jpg) + +![](images/2525cd83110824542c6fb1a551052ff2447259b4ec7dd52b968d177337a488e9.jpg) + +advantageous for challenging material types, such as transparent or reflective objects, where structured-light depth cameras tend to introduce significant noise, severely impacting pose fitting accuracy. Furthermore, the NOCS method does not demonstrate effective performance on the ROPE dataset, leading to the supposition that methods relying solely on RGB information to predict the shape of an object in the canonical space become less robust as the scale of category diversity increases. Compared to GenPose, GenPose++ achieves a significant lead by leveraging the powerful perception capabilities of the 2D foundation model, along with the robustness of clustering towards discrete symmetric properties. You can find qualitative visualizations in Figure 8. + +# 5.2 Ablation Study + +In order to validate the design decisions of our approach, we performed a series of ablation experiments on our method: + +- w/o clustering. Without clustering directly take the average of all remaining pose candidates after outlier removal as the pose estimation output. +- w/o scale prediction. Use the estimated pose to transform the observed point cloud into object space, then take the bounding box length as the maximum projection from the point cloud to each axis. +- w/o simulated depth. Use perfect depth for training. +- w/o point-wise feature fusion. In the feature extraction stage, separately extract the RGB feature and geometric feature, and then concatenation. + +Table 3 illustrates the contribution of each component of GenPose++ to its performance. The introduction of the clustering module allows GenPose++ to effectively aggregate the multimodal distributions caused by discrete symmetries, leading to higher performance. The scale prediction in GenPose++ significantly surpasses direct calculations from the object's point cloud due to ambiguities from partial observations and errors from point cloud noise, particularly in transparent and reflective objects. Training with simulated depth data results in + +Table 3: Ablation study on category-level 6D object pose estimation + +
AblationAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
Full39.019.12.010.015.119.529.4
w/o scale prediction13.83.40.210.015.119.529.4
w/o clustering38.618.71.99.414.118.427.8
w/o simulated depth35.316.51.77.311.914.624.3
w/o point-wise fusion34.215.41.47.110.614.621.9
+ +![](images/0adc22a2ab6a498e57403c9ae654072fee93fcf84548c0e1cdf8b1a97d1e38dd.jpg) +Fig. 8: Qualitative comparison with baselines on ROPE dataset. + +better performance than training with perfect point clouds, as the physics-based depth camera simulation substantially reduces the sim2real gap for depth data. Point-wise fusion outperforms global fusion as it retains more of the object's local geometric features, which are crucial for accurate object pose prediction. + +# 5.3 Category-Level 6D Object Pose Tracking + +Metric. We report the following metrics for object pose tracking evaluation: + +- FPS: Frames Per Second, which indicates the speed of pose tracking. +- VUS@5°5cm: Volume Under Surface, assessing pose estimation accuracy for rotation errors within 0 to $5^{\circ}$ and translation errors within 0 to 5cm. +- mIoU: Mean Intersection over Union, representing the average 3D overlap between the ground truth and the predicted bounding boxes. +- $\mathbf{Rerr}(\circ)$ : Average rotation error in degrees. +- $\mathbf{Terr}(\mathbf{cm})$ : Average translation error in centimeters. + +Baselines. This paper employs BundleTrack, CATRE, and GenPose as baselines for object pose tracking. BundleTrack is a training-free approach that utilizes multi-view feature point detection and matching for tracking the pose of unseen objects. CATRE aligns partially observed point clouds to abstract shape priors to estimate relative transformations, enhancing pose accuracy. Conversely, GenPose utilizes a generative approach to effectively resolve pose ambiguities, notably in symmetric objects. Following GenPose, we have adapted GenPose++ to object pose tracking with minor modifications. + +Results and Analysis. Table 4 presents the results of category-level object pose tracking algorithms CATRE, GenPose, and our method, as well as the results for the unseen object pose tracking algorithm BundleTrack. The category-level pose estimation methods appear to achieve relatively better outcomes since they benefit from learning the category-level canonical space of objects within the SOPE dataset. However, for the training-free BundleTrack, reliance on RGB information for keypoint detection and matching poses challenges, often failing on objects with weak textures. Additionally, its dependency on depth values for global optimization renders it less effective in handling instances with substantial depth noise, such as transparent or specular objects. Our method, without undergoing specialized design, has achieved results comparable to state-of-the-art approaches. Although the inference speed of our method is lower than that of CATRE and GenPose, the achieved 17.8 FPS is sufficient for certain downstream tasks, such as robotic manipulation. Furthermore, recent rapid developments in research on fast samplers are beneficial to our approach, potentially enhancing its performance and applicability in real-time scenarios. + +Table 4: Results of category-level object pose tracking on ROPE. The results are averaged over all 149 categories. GT. Pert. denotes that a perturbed ground truth pose is utilized as the initial object pose. + +
MethodsInputInit.Speed(FPS)↑5°5cm↑mIoU↑Rerr(°)↓Terr(cm)↓
BundleTrack [31]RGBDGT.12.41.33.946.923.5
CATRE [18]DGT. Pert.38.515.955.421.32.6
GenPose [43]DGT. Pert.26.313.3-19.31.2
GenPose++(Ours)RGBDGT. Pert.17.815.953.417.61.2
+ +# 6 Conclusions and Discussion + +In this study, we introduce Omni6DPose, a comprehensive dataset for 6D object pose estimation, featuring extensive scale, diversity, and material variety. Through thorough experimentation, our findings suggest that the probabilistic framework holds promise for category-level 6D object estimation, leveraging semantic information provided by RGB images to address large-scale pose estimation challenges. However, the performance of GenPose++ on Omni6DPose reveals significant room for improvement, with the model still hampered by slow inference speeds resulting from the iterative refinement nature inherent in the diffusion model. Future works could focus on addressing these challenges and integrating the universal 6D pose estimation module trained on Omni6DPose into a broader range of downstream tasks [4,33,37,42]. + +# 7 Acknowledgment + +This work is supported by the National Natural Science Foundation of China - General Program (Project ID: 62376006), the National Youth Talent Support Program (Project ID: 8200800081), and NSFC-62306046. + +# References + +1. An, B., Geng, Y., Chen, K., Li, X., Dou, Q., Dong, H.: RGBmanip: Monocular image-based robotic manipulation through active object pose estimation (2023) +2. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017) +3. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021) +4. Cheng, J., Wu, M., Zhang, R., Zhan, G., Wu, C., Dong, H.: Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220 (2023) +5. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4800-4810 (2023) +6. Dai, Q., Zhang, J., Li, Q., Wu, T., Dong, H., Liu, Z., Tan, P., Wang, H.: Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and transparent objects. In: European Conference on Computer Vision. pp. 374-391. Springer (2022) +7. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022) +8. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: kdd. vol. 96, pp. 226-231 (1996) +9. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022) +10. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. vol. 7724 (10 2012). https://doi.org/10.1007/978-3-642-33885-4_60 +11. Hodań, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017) +12. Jung, H., Zhai, G., Wu, S.C., Ruhkamp, P., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., Roth, D., Navab, N., Busam, B.: Housecat6d - a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios (2023) +13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023) +14. Lin, Z.H., Huang, S.Y., Wang, Y.C.F.: Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +15. Liu, J., Sun, W., Liu, C., Zhang, X., Fan, S., Wu, W.: Hff6d: Hierarchical feature fusion network for robust 6d object pose tracking. IEEE Transactions on Circuits + +and Systems for Video Technology pp. 1-13 (06 2022). https://doi.org/10.1109/TCSVT.2022.3181597 +16. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. arXiv preprint arXiv:2303.13479 (2023) +17. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: ICCV (2021) +18. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: European Conference on Computer Vision (ECCV) (October 2022) +19. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics 22(12), 2633-2651 (2016). https://doi.org/10.1109/TVCG.2015.2513408 +20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +21. Piga, N.A., Onyshchuk, Y., Pasquale, G., Pattacini, U., Natale, L.: Roft: Realtime optical flow-aided 6d object pose and velocity tracking. IEEE Robotics and Automation Letters 7(1), 159–166 (2022). https://doi.org/10.1109/LRA.2021.3119379 +22. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +23. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems 32 (2019) +24. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +25. Tian, Y., Zhang, J., Huang, G., Wang, B., Wang, P., Pang, J., Dong, H.: Robokeygen: Robot pose and joint angles estimation via diffusion-based 3d keypoint generation. arXiv preprint arXiv:2403.18259 (2024) +26. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011) +27. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019) +28. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +29. Wang, J., Chen, K., Dou, Q.: Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4807-4814. IEEE (2021) +30. Wang, P., Jung, H., Li, Y., Shen, S., Srikanth, R.P., Garattoni, L., Meier, S., Navab, N., Busam, B.: Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21222-21231 (2022) +31. Wen, B., Bekris, K.E.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (2021) + +32. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 13209-13218 (October 2021) +33. Wu, M., Zhong, F., Xia, Y., Dong, H.: Targf: Learning target gradient field to rearrange objects without explicit goal specification. Advances in Neural Information Processing Systems 35, 31986-31999 (2022) +34. Wu, T., Wu, M., Zhang, J., Gan, Y., Dong, H.: Learning score-based grasping primitive for human-assisting dexterous grasping. Advances in Neural Information Processing Systems 36 (2024) +35. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023) +36. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018) +37. Xue, T., Wu, M., Lu, L., Wang, H., Dong, H., Chen, B.: Learning gradient fields for scalable and generalizable irregular packing. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023) +38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023) +39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) +40. You, Y., Shi, R., Wang, W., Lu, C.: CPPf: Towards robust category-level 9d pose estimation in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6866-6875 (2022) +41. You, Y., Xiong, K., Yang, Z., Huang, Z., Zhou, J., Shi, R., Fang, Z., Harley, A.W., Lu, C.: Pace: Pose annotations in cluttered environments. arXiv preprint arXiv:2312.15130 (2023) +42. Zeng, Y., Wu, M., Yang, L., Zhang, J., Ding, H., Cheng, H., Dong, H.: Distilling functional rearrangement priors from large models. arXiv preprint arXiv:2312.01474 (2023) +43. Zhang, J., Wu, M., Dong, H.: Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems 36 (2024) +44. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17163-17173 (June 2023) +45. Zhou, L., Liu, Z., Gan, R., Wang, H., au2, M.H.A.J.: Dr-pose: A two-stage deformation-and-registration pipeline for category-level 6d object pose estimation (2023) \ No newline at end of file diff --git a/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/images.zip b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e8f43b0ae90015d7b23a147701c5f50b7865a59 --- /dev/null +++ b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd3ac23fd8cda9b9d6d4ae4a933cc92ce71d8988e6e1e90e7b7ec5341b4ce67e +size 777252 diff --git a/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/layout.json b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6698da4cf5ba05ba7184f5b4e10fb8abe4d66747 --- /dev/null +++ b/2024/Omni6DPose_ A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking/layout.json @@ -0,0 +1,10061 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 459, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 459, + 166 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 459, + 166 + ], + "type": "text", + "content": "Omni6DPose: A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": "Jiyao Zhang" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": ", Weiyao Huang" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": ", Bo Peng" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": ", Mingdong Wu" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": ", Fei Hu" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": ", Zijian Chen" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": ", Bo Zhao" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{5,6}" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "text", + "content": ", and Hao Dong" + }, + { + "bbox": [ + 133, + 185, + 479, + 212 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 219, + 476, + 270 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 142, + 219, + 470, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 219, + 470, + 237 + ], + "spans": [ + { + "bbox": [ + 142, + 219, + 470, + 237 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 219, + 470, + 237 + ], + "type": "text", + "content": " Center on Frontiers of Computing Studies, School of Computer Science, Peking University " + }, + { + "bbox": [ + 142, + 219, + 470, + 237 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 219, + 470, + 237 + ], + "type": "text", + "content": " PKU-Agibot Lab, School of Computer Science, Peking University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "spans": [ + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "text", + "content": " National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University " + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "text", + "content": " Waseda University " + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "text", + "content": " Shanghai Jiao Tong University " + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 136, + 237, + 476, + 270 + ], + "type": "text", + "content": " BAAI {jiyaozhang, sshwy, bo.peng, hofee}@stu.pku.edu.cn zijianchenwaseda@akane.waseda.jp bo.zhao@sjtu.edu.cn {wmingd, hao.dong}@pku.edu.cn" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 218, + 274, + 394, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 274, + 394, + 285 + ], + "spans": [ + { + "bbox": [ + 218, + 274, + 394, + 285 + ], + "type": "text", + "content": "https://jiyao06.github.io/Omni6DPose/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 295, + 452, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 295, + 452, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 295, + 452, + 514 + ], + "type": "text", + "content": "Abstract. 6D object pose estimation is crucial in the field of computer vision. However, it suffers from a significant lack of large-scale and diverse datasets, impeding comprehensive model evaluation and curtailing downstream applications. To address these issues, this paper introduces Omni6DPose, a substantial benchmark featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is divided into three main components: ROPE (Real 6D Object Pose Estimation Dataset), which includes 332K images annotated with over 1.5M annotations across 581 instances in 149 categories; SOPE(Simulated 6D Object Pose Estimation Dataset), a simulated training set created by mixed reality and physics-based depth simulation; and PAM(Pose Aligned 3D Models), the manually aligned real scanned objects used in ROPE and SOPE. Omni6DPose is inherently challenging due to the substantial variations and ambiguities. To address this issue, we introduce GenPose++, an enhanced version of the SOTA category-level 6D object pose estimation framework, incorporating two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation. Moreover, we provide a comprehensive benchmarking analysis to evaluate the performance of previous methods on this new large-scale dataset in the realms of 6D object pose estimation and pose tracking." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 517, + 449, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 517, + 449, + 529 + ], + "spans": [ + { + "bbox": [ + 160, + 517, + 449, + 529 + ], + "type": "text", + "content": "Keywords: Benchmark " + }, + { + "bbox": [ + 160, + 517, + 449, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 517, + 449, + 529 + ], + "type": "text", + "content": " object pose estimation " + }, + { + "bbox": [ + 160, + 517, + 449, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 517, + 449, + 529 + ], + "type": "text", + "content": " object pose tracking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 548, + 230, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 230, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 230, + 559 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "6D object pose estimation [16,28,45] and pose tracking [15,21] from single images is an essential task in computer vision, holding immense potential for applications in robotics [1] and augmented reality/virtual reality (AR/VR) [19]. Over recent decades, the domain has experienced significant advancements, primarily dominated by data-driven learning approaches. Analogous to the pivotal role of data in learning-based 2D foundation tasks, high-quality, comprehensive datasets are paramount in the context of 6D object pose estimation and tracking." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 334, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 334, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 334, + 665 + ], + "type": "text", + "content": "*: equal contributions, †: corresponding author" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 268 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 268 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 268 + ], + "type": "image", + "image_path": "6e30f9fb1ef9101e65707ac7f168841fdf416464f146a4f76a392dff6c0d76ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 271, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: We introduce Omni6DPose for universal 6D object pose estimation and tracking, which comprises three key components: (1) ROPE (right), a large-scale real-world dataset for evaluation. (2) SOPE (left), a vast synthetic dataset generated through mixed reality and physics-based simulation for training. (3) PAM (center), a collection of pose-aligned 3D object models. Omni6DPose is featured by its diversity in object categories, large scale, and variety in object materials." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 349, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 494 + ], + "type": "text", + "content": "Today 6D object pose estimation is studied under two lenses: instance-level and category-level. In instance-level settings, datasets such as Linemod [10], YCB-Video [36], and T-LESS [11] have gained widespread acceptance as benchmarks. These datasets are distinguished by their focus on detailed, individual object instances, thereby enabling algorithms to precisely learn and predict the poses of specific items. On the other hand, category-level pose estimation emphasizes generalization across different items within a particular object category. The NOCS [28] dataset stands out as the most widely used in the category-level object pose estimation field, providing a simulated dataset for training and a small-scale real-world dataset for evaluation. Despite their contributions to advancing the field, these datasets present limitations due to their small scale in terms of instances or categories. This results in two significant challenges:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 501, + 480, + 549 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 501, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 501, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 136, + 501, + 480, + 523 + ], + "type": "text", + "content": "1. It hampers comprehensive evaluation of different models' performance, limiting the development of research in this field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 525, + 480, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 525, + 480, + 549 + ], + "spans": [ + { + "bbox": [ + 135, + 525, + 480, + 549 + ], + "type": "text", + "content": "2. It restricts the applicability of research findings across diverse domains, due to the limited variety of object instances or categories represented." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "To address the aforementioned challenges and drive advancements in this field, this paper introduces Omni6DPose, a universal 6D object pose estimation dataset featured by its diversity in object categories, large scale, and variety in object materials. Omni6DPose is segmented into three principal components: (1) ROPE (Real 6D Object Pose Estimation Dataset), which encompasses " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "332\\mathrm{K}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " images annotated with over 1.5M annotations across 581 instances in 149 categories; (2) SOPE (Simulated 6D Object Pose Estimation Dataset), comprising " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "475\\mathrm{K}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " images generated in a mixed reality setting with depth simulation, furnished with over 5M annotations spanning 4162 instances in the same 149" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 203, + 153 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 203, + 153 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 203, + 153 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 203, + 153 + ], + "type": "image", + "image_path": "1a60111597c270923e5a9fe2d9337ff4c25a52767811667ea1190b636c2b590f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 155, + 203, + 194 + ], + "blocks": [ + { + "bbox": [ + 136, + 155, + 203, + 194 + ], + "lines": [ + { + "bbox": [ + 136, + 155, + 203, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 155, + 203, + 194 + ], + "type": "image", + "image_path": "e30d062643149c774e69ec1aee782198766bc6e4587e0098a9fa42d0a9e1cac6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 194, + 203, + 233 + ], + "blocks": [ + { + "bbox": [ + 136, + 194, + 203, + 233 + ], + "lines": [ + { + "bbox": [ + 136, + 194, + 203, + 233 + ], + "spans": [ + { + "bbox": [ + 136, + 194, + 203, + 233 + ], + "type": "image", + "image_path": "43275bc3b8beb29183c70ae36d9adae17d40d064a09a49bb097a67b6f133b781.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 479, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 479, + 262 + ], + "type": "text", + "content": "Fig. 2: ROPE dataset visualization. In the figure, bounding boxes are colored according to the coordinates in the object's coordinate system." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 205, + 116, + 272, + 155 + ], + "blocks": [ + { + "bbox": [ + 205, + 116, + 272, + 155 + ], + "lines": [ + { + "bbox": [ + 205, + 116, + 272, + 155 + ], + "spans": [ + { + "bbox": [ + 205, + 116, + 272, + 155 + ], + "type": "image", + "image_path": "44ff07bbb3552f46ea54d5355b86627c6238da0b82a64dc8ea16bf8033463a1d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 156, + 272, + 194 + ], + "blocks": [ + { + "bbox": [ + 205, + 156, + 272, + 194 + ], + "lines": [ + { + "bbox": [ + 205, + 156, + 272, + 194 + ], + "spans": [ + { + "bbox": [ + 205, + 156, + 272, + 194 + ], + "type": "image", + "image_path": "6468a5f50a56759c8b1bbf03e1aa1b915eb71628c75cdcf150bdc344d9cccb97.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 205, + 194, + 272, + 233 + ], + "blocks": [ + { + "bbox": [ + 205, + 194, + 272, + 233 + ], + "lines": [ + { + "bbox": [ + 205, + 194, + 272, + 233 + ], + "spans": [ + { + "bbox": [ + 205, + 194, + 272, + 233 + ], + "type": "image", + "image_path": "57fea1551a53410447966c9d8b602271c459fccd8967a553a84fabb67a4623ed.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 274, + 116, + 340, + 155 + ], + "blocks": [ + { + "bbox": [ + 274, + 116, + 340, + 155 + ], + "lines": [ + { + "bbox": [ + 274, + 116, + 340, + 155 + ], + "spans": [ + { + "bbox": [ + 274, + 116, + 340, + 155 + ], + "type": "image", + "image_path": "b8808cb50f283d7c2f4884717cd265866e36cf74011e315d21131af851534fce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 274, + 156, + 340, + 194 + ], + "blocks": [ + { + "bbox": [ + 274, + 156, + 340, + 194 + ], + "lines": [ + { + "bbox": [ + 274, + 156, + 340, + 194 + ], + "spans": [ + { + "bbox": [ + 274, + 156, + 340, + 194 + ], + "type": "image", + "image_path": "172fd19562b74c51f5ab7fa2c961c534695d45e685e2ea5611bb5ebf4f21b840.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 274, + 194, + 340, + 233 + ], + "blocks": [ + { + "bbox": [ + 274, + 194, + 340, + 233 + ], + "lines": [ + { + "bbox": [ + 274, + 194, + 340, + 233 + ], + "spans": [ + { + "bbox": [ + 274, + 194, + 340, + 233 + ], + "type": "image", + "image_path": "5b423517e481a3550e627b5e8ada15cd45c58163646cad585cb6f5cab3e6e0f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 342, + 116, + 410, + 155 + ], + "blocks": [ + { + "bbox": [ + 342, + 116, + 410, + 155 + ], + "lines": [ + { + "bbox": [ + 342, + 116, + 410, + 155 + ], + "spans": [ + { + "bbox": [ + 342, + 116, + 410, + 155 + ], + "type": "image", + "image_path": "cfc55b1d4839de861b119a85310341fb2d3146b2b9b25b746da1388d41fadb75.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 342, + 156, + 410, + 194 + ], + "blocks": [ + { + "bbox": [ + 342, + 156, + 410, + 194 + ], + "lines": [ + { + "bbox": [ + 342, + 156, + 410, + 194 + ], + "spans": [ + { + "bbox": [ + 342, + 156, + 410, + 194 + ], + "type": "image", + "image_path": "2996098182c76da3d4a14b8593baa467e1c4312c27a59b8b6e9fdb2fbb194d5d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 342, + 194, + 410, + 233 + ], + "blocks": [ + { + "bbox": [ + 342, + 194, + 410, + 233 + ], + "lines": [ + { + "bbox": [ + 342, + 194, + 410, + 233 + ], + "spans": [ + { + "bbox": [ + 342, + 194, + 410, + 233 + ], + "type": "image", + "image_path": "f89108a76ee285ad48ac784e5c47e30101301b38ef746d7f5687b12476d37d37.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 410, + 116, + 477, + 155 + ], + "blocks": [ + { + "bbox": [ + 410, + 116, + 477, + 155 + ], + "lines": [ + { + "bbox": [ + 410, + 116, + 477, + 155 + ], + "spans": [ + { + "bbox": [ + 410, + 116, + 477, + 155 + ], + "type": "image", + "image_path": "9b4943baaa6d025c9201974d559431e8719b9631d0841cc21f6dd49a6c98717b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 410, + 156, + 477, + 194 + ], + "blocks": [ + { + "bbox": [ + 410, + 156, + 477, + 194 + ], + "lines": [ + { + "bbox": [ + 410, + 156, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 410, + 156, + 477, + 194 + ], + "type": "image", + "image_path": "7511dc9f5d88aada149176f9573656101fa178c98e0e1b8986d4678451f1b6cf.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 410, + 194, + 477, + 233 + ], + "blocks": [ + { + "bbox": [ + 410, + 194, + 477, + 233 + ], + "lines": [ + { + "bbox": [ + 410, + 194, + 477, + 233 + ], + "spans": [ + { + "bbox": [ + 410, + 194, + 477, + 233 + ], + "type": "image", + "image_path": "5eeb4a4d92a5496ce3e870202d7eb7a7e2a8f2723c6af499c5d3209a846479de.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 267, + 479, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 479, + 315 + ], + "type": "text", + "content": "categories. The mixed reality bridges the semantic sim2real gap, while the depth sensor simulation closes the geometric sim2real gap; (3) PAM(Pose Aligned 3D Models), the manually aligned, real scanned objects utilized in both ROPE and SOPE, enabling the generation of diverse downstream task data." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 316, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 434 + ], + "type": "text", + "content": "Omni6DPose poses inherent challenges due to its considerable variations, diverse materials, and inherent ambiguities, which reflect the complexities encountered in real-world applications. Figure 2 illustrates some examples from ROPE. To tackle these issues, we introduce GenPose++, which incorporates GenPose [43] with two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation, tailored specifically to the nuances of the Omni6DPose in question. Furthermore, as a Universal 6D object pose estimation dataset, this paper also offers a comprehensive benchmarking analysis to assess the performance of existing methods on category-level 6D object pose estimation and pose tracking. We summarize our contributions as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 135, + 439, + 487, + 509 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 135, + 439, + 487, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 439, + 487, + 462 + ], + "spans": [ + { + "bbox": [ + 135, + 439, + 487, + 462 + ], + "type": "text", + "content": "1. We present Omni6DPose, a comprehensive 6D object pose estimation dataset with extensive categories, instance diversity, and material variety." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 135, + 462, + 480, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 462, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 135, + 462, + 480, + 486 + ], + "type": "text", + "content": "2. We propose a real data collection pipeline and a simulation framework for generating synthetic data with low semantic and geometry sim2real gaps." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 135, + 486, + 480, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 486, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 135, + 486, + 480, + 509 + ], + "type": "text", + "content": "3. We introduce GenPose++ for category-level 6D object pose estimation and tracking, demonstrating SOTA performance on Omni6DPose." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 519, + 234, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 519, + 234, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 234, + 532 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 536, + 480, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 608 + ], + "type": "text", + "content": "For 6D object pose estimation, there are two main branches: instance-level and category-level. Instance-level estimation is tested on seen objects or on unseen objects with reference, while category-level estimation is tested on unseen instances of known categories without CAD models. In this section, we review and compare existing datasets to our large-scale category-level dataset and review relevant algorithms for category-level pose estimation and tracking." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 131, + 613, + 345, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 345, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 345, + 624 + ], + "type": "text", + "content": "2.1 6D Object Pose Estimation Datasets" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Following the outlined branches of 6D object pose estimation, we have reviewed datasets corresponding to both instance-level and category-level 6D object pose estimation. A comparative analysis of these datasets is provided in Table 1." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 192, + 475, + 342 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 190 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 190 + ], + "type": "text", + "content": "Table 1: This table compares datasets for 6D object pose estimation, focusing on object category count, reality of the data, data modalities (RGB, Depth, IR), and object attributes such as quantity, CAD model availability, and inclusion of transparent and specular objects. It also details video characteristics by number and marker presence, along with image and annotation counts. 'Wild6D' refers specifically to the test split of the Wild6D dataset, as the training data does not provide annotations. The symbol '-' indicates the absence of a particular feature within the dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 192, + 475, + 342 + ], + "lines": [ + { + "bbox": [ + 138, + 192, + 475, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 475, + 342 + ], + "type": "table", + "html": "
DatasetCat. RealModalityObjectMarker-free Vid.Img.Anno.
RGBDepth IRIRNum. CAD Trans. Spec.
CAMERA [28]6X✓✓X1085✓XX✓-300K
SOPE(Ours)149X✓✓✓4162✓✓✓✓-475K
YCB-Video [36]-✓✓✓X21✓XX✓92133K
T-LESS [11]-✓✓✓X30✓XXX2048K
Linemod [10]-✓✓✓X15✓XXX-18K
StereoOBJ-1M [17]-✓✓XX18✓✓✓X182393K
REAL275 [28]6✓✓✓X42✓XXX188K
PhoCal [30]8✓✓✓X60✓✓✓✓243.9K
HouseCat6D [12]10✓✓✓X194✓✓✓✓4123.5K
Wild6D* [9]5✓✓✓X162X(✓)X✓48610K
PACE [41]44✓✓✓X576✓XX✓30055K
ROPE(Ours)149✓✓✓✓581✓✓✓✓363332K
", + "image_path": "960f233612f00747db8565b437de6da3ff9cbceec3543f6b71811531f6c0578f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 349, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 456 + ], + "type": "text", + "content": "Instance-level 6D object pose estimation dataset. LineMod [10] is one of the most used datasets, providing non-temporal RGB-D images and ground truth pose annotations. YCB-Video [36] provides RGB-D videos and annotations, enabling pose-tracking approaches. T-LESS [11] features texture-less objects with symmetries and mutual similarities. StereoOBJ-1M [17] achieves a leap in dataset scale and features transparent and reflective objects. While these datasets are extensive in terms of image and annotation count, they are limited in the diversity of instances they cover. For example, StereoOBJ-1M comprises 339K frames and 1.5M annotations, yet it includes only 18 unique object instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": "Category-level 6D object pose estimation dataset. NOCS [28] provides the first benchmark in category-level pose estimation, while Wild6D [9] addresses the scalability issue of datasets by leveraging unlabeled and synthetic data. PhoCal [30] focuses on photometrically challenging objects and HouseCat6D [12] offers diverse scenes, viewpoints, and grasping annotations. However, these datasets cover only a limited number of categories, even the most extensive dataset, HouseCat6D, includes merely ten categories. PACE [41], a concurrent work, focusing on object pose estimation in complex scenes. In contrast, our datasets, SOPE and ROPE, set new benchmarks by offering the widest range of categories and featuring objects with diverse materials, thereby enhancing dataset diversity and realism for pose estimation research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 600, + 424, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 600, + 424, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 600, + 424, + 613 + ], + "type": "text", + "content": "2.2 6D Object Pose Estimation and Tracking Algorithm" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Category-level 6D object pose estimation. Category-level 6D object pose estimation [28,39,40] aims to estimate unseen instance poses within the same category. NOCS [28] introduces a normalized object coordinate space for pose prediction without CAD models, while SPD [29] and SGPA [3] utilize category-level" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": "prriors for enhanced estimation. HS-Pose [44] extends 3D-GC [14] for improved feature extraction from point clouds. IST-Net [16] transforms features between camera and world spaces implicitly, without relying on 3D priors, surpassing previous methods. However, these techniques, mainly regression-based, require ad-hoc designs for symmetric objects. GenPose [43] addresses it by generatively modeling the pose distribution, eliminating the need for symmetry considerations. Yet, GenPose neglects RGB semantic information, which is increasingly vital as object category scales grow. Additionally, its energy-based aggregation algorithm fails with discontinuous multimodal distributions. We introduce GenPose++, which incorporates a 2D foundation model to leverage RGB semantics, improving generalization, and introduces an aggregation module to handle discrete multimodal distributions, addressing the limitations of GenPose." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 483, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 483, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 483, + 387 + ], + "type": "text", + "content": "6D object pose tracking. This paper is situated within the domain of category-level 6D object pose tracking and model-free object tracking. BundleTrack [31] pioneers model-free tracking by leveraging multi-view feature detection for tracking unseen objects without 3D models. CAPTRA [32] enhances articulated pose tracking through recursive updates for better temporal consistency. CATRE [18] aligns partially observed point clouds to abstract shape priors for relative transformations and pose estimation. GenPose, adopting a generative approach, effectively addresses the challenge of pose ambiguities in symmetric objects. Together, these approaches underscore the evolving landscape of object pose tracking, highlighting both the progress and the diversity of strategies being explored." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 394, + 282, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 282, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 282, + 407 + ], + "type": "text", + "content": "3 Omni6DPose Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 482 + ], + "type": "text", + "content": "This paper introduces a rich variety of object categories, a large-scale, and materially diverse dataset for real 6D object pose estimation, named ROPE. And, a simulated dataset, SOPE, synthesizing with mixed reality and featuring depth simulation, is provided for training. Section 3.1 will discuss the collection and alignment of 3D objects. Section 3.2 will cover the acquisition and labeling of the ROPE dataset. Section 3.3 will explain the generation of the SOPE dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 498, + 347, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 347, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 347, + 510 + ], + "type": "text", + "content": "3.1 3D Object Collection and Alignment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 516, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 636 + ], + "type": "text", + "content": "Universal 6D object pose estimation relies on a comprehensive set of objects. We selected 149 categories of everyday objects, all reconstructed with high-precision scanners, and categorized them into two sets: SOPE for simulated data and ROPE for real-world scenes. SOPE primarily includes objects from sources like OmniObject3D [35], PhoCal [30], and GoogleScan [7], alongside a subset from our scans, totaling 5,000 instances. ROPE consists of 580 instances we reconstructed using industrial scanners. Importantly, while most SOPE objects are from public datasets, manual category-level pose alignment is necessary. For object reconstruction, as shown in Figure 3, we use two professional scanners, EinScan H2" + }, + { + "bbox": [ + 130, + 516, + 482, + 636 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 130, + 516, + 482, + 636 + ], + "type": "text", + "content": " and Revopoint POP 3" + }, + { + "bbox": [ + 130, + 516, + 482, + 636 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 130, + 516, + 482, + 636 + ], + "type": "text", + "content": " for objects in different scales. The scan" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 260, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 260, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 260, + 653 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 133, + 642, + 260, + 653 + ], + "type": "text", + "content": " https://www.einscan.com/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 654, + 277, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 277, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 277, + 665 + ], + "type": "text", + "content": "8 https://www_REVopoint3d.com/" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 285, + 217 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 285, + 217 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 285, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 285, + 217 + ], + "type": "image", + "image_path": "ee72ddeb9dd614b8107981b5074fea0e9e9d96b84d2dc124a17f7cf47d4d842b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 221, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 310 + ], + "type": "text", + "content": "Fig. 3: ROPE dataset collection and annotation. (1) Object scanning, where high-precision industrial scanners are used to acquire the CAD models of objects; (2) Object canonicalization, involving the alignment of each object category to the canonical space; (3) Video capture, capturing video sequences in varied scenarios with a depth camera; and (4) Pose annotation, calculating camera poses through Structure from Motion (SFM), further utilizing Farthest Point Sampling (FPS) to select keyframes for keypoint annotation, and performing bundle adjustment to derive initial object pose values, which are then manually refined to obtain more precise annotations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 114, + 481, + 217 + ], + "blocks": [ + { + "bbox": [ + 286, + 114, + 481, + 217 + ], + "lines": [ + { + "bbox": [ + 286, + 114, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 286, + 114, + 481, + 217 + ], + "type": "image", + "image_path": "fdeea2904a14f8872961b09ecaedf36c35b31eb8abc2fe5c48f6bd60ff54a8c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 321, + 479, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 381 + ], + "type": "text", + "content": "ning time depends on object features: it took about 15 minutes for small, simple, Lambertian items like a mouse, and up to an hour for complex, large, or non-Lambertian items like a transparent mug. Finally, we constructed a specialized annotation tool for manually aligning objects in the same category to the category-level canonical space, with each alignment taking roughly one minute." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 388, + 336, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 336, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 336, + 400 + ], + "type": "text", + "content": "3.2 ROPE Acquisition and Annotation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "content": "The ROPE dataset was systematically acquired utilizing the RealSense D415 imaging device, encompassing scenarios with 2 to 6 distinct objects and video sequences extending from 762 to 1,349 frames. The integrity and utility of the dataset are underpinned by the precision of object pose annotations, which present notable challenges, chiefly:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 476, + 481, + 561 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "text", + "content": "1. Derivation of relative camera poses, denoted as " + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "inline_equation", + "content": "T_{c} = \\{(R_{i},t_{i})\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "text", + "content": ", where each pair " + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "inline_equation", + "content": "(R_i,t_i)\\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "text", + "content": ", signifying the transformation from the camera space to the world space for the " + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "text", + "content": " frame, with " + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 136, + 476, + 481, + 525 + ], + "type": "text", + "content": " symbolizing the aggregate frame count within the video sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 524, + 481, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 524, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 135, + 524, + 481, + 561 + ], + "type": "text", + "content": "2. Procurement of high-accuracy object poses, represented as " + }, + { + "bbox": [ + 135, + 524, + 481, + 561 + ], + "type": "inline_equation", + "content": "T_{o} = \\{(R,t)\\}" + }, + { + "bbox": [ + 135, + 524, + 481, + 561 + ], + "type": "text", + "content": ", where the pair " + }, + { + "bbox": [ + 135, + 524, + 481, + 561 + ], + "type": "inline_equation", + "content": "(R,t) \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 135, + 524, + 481, + 561 + ], + "type": "text", + "content": ", delineating the transformation from the object space the camera space for any selected frame." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " known, it is possible to automate the generation of all annotations within the dataset. Addressing these challenges, we propose a marker-free annotation system. Previous datasets for calculating relative camera pose rely on markers, like NOCS [28], or external robot arms for indirect calculation, such as PhoCal [30], limiting scene diversity. As shown in Figure 3, to enable marker-free annotation in open scenes, we consider it a structure-from-motion (SfM) problem, utilizing intrinsic scene features to optimize camera poses " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " using bundle adjustment techniques. This approach aims to solve the optimization problem:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 279, + 201 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 279, + 201 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 279, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 279, + 201 + ], + "type": "image", + "image_path": "ee6825a0c6cd87d43cf531eb473786b2d5b4fc4606b3b301b1180875f4d4c284.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 480, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 228 + ], + "type": "text", + "content": "Fig. 4: SOPE synthesis, utilizing mixed reality to bridge the RGB sim2real gap and physical-based depth sensor simulation to minimize the geometric sim2real gap." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 281, + 114, + 478, + 201 + ], + "blocks": [ + { + "bbox": [ + 281, + 114, + 478, + 201 + ], + "lines": [ + { + "bbox": [ + 281, + 114, + 478, + 201 + ], + "spans": [ + { + "bbox": [ + 281, + 114, + 478, + 201 + ], + "type": "image", + "image_path": "b6208e949c4c1e03ff6fa2b7e8f0a535277ba4674ccd460c5d24793674b3bd92.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 240, + 480, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 240, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 229, + 240, + 480, + 273 + ], + "type": "interline_equation", + "content": "\\min _ {T _ {c}} \\sum_ {i = 1} ^ {n} \\sum_ {j \\in \\mathcal {P} _ {i}} \\| \\pi \\left(R _ {i} X _ {j} + t _ {i}\\right) - x _ {i j} \\| ^ {2} \\tag {1}", + "image_path": "154fa2ba6669ccf893ef33724e762f5b259ffe1376ee4d30184fe2bf789f4d6e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": " denotes the camera projection function, " + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": " represents the 3D points in the world space, and " + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "inline_equation", + "content": "x_{ij}" + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": " corresponds to the 2D projection of " + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": " camera frame, with " + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": " being the set of point correspondences in frame " + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 279, + 479, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "type": "text", + "content": "For object pose " + }, + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "type": "text", + "content": " annotations, previous methods consider only a single frame, leading to inaccuracies due to the lack of multi-viewpoint constraints. To overcome this, we introduce a two-stage object pose annotation process: in the first stage, keyframes are sampled from SfM results using Farthest Point Sampling (FPS), and 2D-3D keypoint pairs on these keyframes are annotated, providing initial object pose by minimizing the reprojection error:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 398, + 480, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 398, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 242, + 398, + 480, + 423 + ], + "type": "interline_equation", + "content": "\\min _ {R, t} \\sum_ {k \\in \\mathcal {K}} \\| \\pi \\left(R X _ {k} + t\\right) - x _ {k} \\| ^ {2} \\tag {2}", + "image_path": "4169d18c72ea1cd92a08904a6a600414146a4211acb27894ccaa9bd0a24c733a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "inline_equation", + "content": "X_{k}" + }, + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "text", + "content": " are 3D keypoints of object, " + }, + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "text", + "content": " are corresponding 2D annotations in the image, and " + }, + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 130, + 430, + 479, + 465 + ], + "type": "text", + "content": " is the set of all keypoint correspondences. The initial poses are manually fine-tuned based on the object's projection across all keyframes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 475, + 242, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 242, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 242, + 488 + ], + "type": "text", + "content": "3.3 SOPE Synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "content": "ROPE represents a comprehensive benchmark in category-level object pose estimation by scaling up the diversity and number of object categories to unprecedented levels, encompassing a wide range of materials. This diversity presents new challenges for network training data due to the higher demands on the dataset's scale and diversity. Collecting a larger real-world dataset would be prohibitively expensive and unlikely to ensure sufficient diversity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To bridge the sim2real gap, which is pronounced when using synthetic data, either in RGB or geometry, this paper proposes a novel method based on mixed reality with depth simulation for synthetic data generation. Specifically, as demonstrated in Figure 4, we employ mixed reality [28] techniques to generate RGB data, thereby reducing the RGB sim2real gap. In parallel, we simulate the mechanism of structured light depth sensors within blender [6]. This involves rendering infrared (IR) images and applying stereo matching to produce synthetic depth maps, effectively narrowing the geometry's sim2real gap." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 299, + 238 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 299, + 238 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 299, + 238 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 299, + 238 + ], + "type": "image", + "image_path": "b24d7864e681bc47c2a299d2a477da0236290482111c0d79cb050ffe4dbd8173.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 278 + ], + "type": "text", + "content": "Fig. 5: Omni6DPose statistics, showcasing the dataset distribution. Left: Category distribution, highlighting 149 categories and diverse materials. Right: Object size distribution across 5000 objects, illustrating diversity in shapes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 116, + 476, + 238 + ], + "blocks": [ + { + "bbox": [ + 304, + 116, + 476, + 238 + ], + "lines": [ + { + "bbox": [ + 304, + 116, + 476, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 476, + 238 + ], + "type": "image", + "image_path": "a8960cd64b7036bd10197f1ff8903d7c074e3fd6859f78cfb271cb3f671c0fc7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "content": "During data generation, we implement domain randomization for illumination and object materials to further enhance the dataset's diversity. All the background images are sourced from public datasets, including 19,658 images from MatterPort3D [2], 2,572 from ScanNet++ [38], and 540 from IKEA [28]. To the best of our knowledge, this is the first simulated dataset that uses a context-aware mixed reality approach combined with physical-based depth sensor simulation for object pose estimation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 384, + 250, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 250, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 250, + 395 + ], + "type": "text", + "content": "3.4 Dataset Statistics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 474 + ], + "type": "text", + "content": "Object Category Statistics The comprehensive distribution of object category and size are both demonstrated in Figure 5. Most of the categories possess " + }, + { + "bbox": [ + 130, + 401, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\geq 25\\mathrm{K}" + }, + { + "bbox": [ + 130, + 401, + 482, + 474 + ], + "type": "text", + "content": " pose annotations in the SOPE dataset, providing sufficient training opportunities. Categories containing objects with diverse and challenging material options (e.g., transparent or specular materials) are equipped with apparently more data generation, such as dishes, cups, bottles, bowls, mugs, etc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 533 + ], + "type": "text", + "content": "Object Size Statistics As shown in Figure 5, the objects in our dataset span a wide range of sizes. The majority of the objects are approximately 0.1 meters in length along the diagonal of their bounding boxes, with the largest objects exceeding 1 meter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 542, + 411, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 542, + 411, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 542, + 411, + 556 + ], + "type": "text", + "content": "4 Category-level 6D Pose Estimation Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Given the Omni6DPose, one naturally ponders the optimal technical approach for large-scale category-level pose estimation. The recently introduced state-of-the-art category-level 6D pose estimation technique, GenPose [43], offers a promising avenue by employing a diffusion-based probabilistic method [23, 24]. Moreover, the diffusion model has demonstrated remarkable efficacy across various high-dimensional domains with extensive training data [5, 25, 34, 42]. Expanding on this groundwork, our study delves further into the probabilistic approach, presenting an enhanced iteration of GenPose, named GenPose++." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 479, + 250 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 479, + 250 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 479, + 250 + ], + "type": "image", + "image_path": "bd6667f41f34f9d181debef2094ec10cadf3934d738b094a7bbef93d5a365440.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 338 + ], + "type": "text", + "content": "Fig. 6: Overview of GenPose++. We employ segmented point clouds and cropped RGB images as inputs. PointNet++ extracts geometric features, while DINO v2, a pre-trained 2D foundation model, extracts semantic features. These features are fused to condition a diffusion model, which generates object pose candidates and their corresponding energy. Finally, clustering is applied to address aggregation issues in the multimodal distribution of poses for objects with non-continuous symmetry, such as boxes, effectively resolving the pose estimation challenge." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 348, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{GenPose + + }" + }, + { + "bbox": [ + 131, + 348, + 482, + 407 + ], + "type": "text", + "content": " integrates two crucial enhancements: Semantic-aware feature extraction (see Fig.6 (a)) and Clustering-based aggregation (as shown in Fig.6 (c)). The subsequent sections will detail the three primary stages of the GenPose++ pipeline. Moreover, given the estimated 6D pose, " + }, + { + "bbox": [ + 131, + 348, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{GenPose + + }" + }, + { + "bbox": [ + 131, + 348, + 482, + 407 + ], + "type": "text", + "content": " provide a additional regression network to predict the 3D scale of the object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 415, + 430, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 430, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 430, + 427 + ], + "type": "text", + "content": "4.1 Training Semantic-aware Score and Energy Networks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": "The learning agent is trained on our paired dataset " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(p_k,O_k,I_k)\\}_{k = 1}^n" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\pmb {p}_k\\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "O_{k}\\in \\mathbb{R}^{3\\times N}" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "I_{k}\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": " denote a 6D pose, a partially observed 3D point cloud with " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": " points, and a cropped RGB image with " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": " resolution, respectively. Given an unseen object with point cloud " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "O^{*}" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": " and RGB image " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": ", the goal is to recover the corresponding ground-truth pose " + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{*}" + }, + { + "bbox": [ + 130, + 434, + 482, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "text", + "content": "Following GenPose, we initially train a score network " + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}:\\mathbb{R}^{|\\mathcal{P}|}\\times \\mathbb{R}^{1}\\times \\mathbb{R}^{3\\times N}\\times" + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{3\\times H\\times W}\\to \\mathbb{R}^{|P|}" + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "text", + "content": " and an energy network " + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\varPsi_{\\phi}:\\mathbb{R}^{|\\mathcal{P}|}\\times \\mathbb{R}^{1}\\times \\mathbb{R}^{3\\times N}\\times \\mathbb{R}^{3\\times H\\times W}\\to \\mathbb{R}^{1}" + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "text", + "content": " from the dataset " + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 494, + 481, + 530 + ], + "type": "text", + "content": " using the denoising score-matching objective [26]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 537, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 537, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 143, + 537, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t \\sim \\mathcal {U} (\\epsilon , 1)} \\left\\{\\lambda (t) \\mathbb {E} _ {\\substack {\\boldsymbol {p} (0) \\sim p _ {\\text {data}} (\\boldsymbol {p} (0) | O, I), \\\\ \\boldsymbol {p} (t) \\sim \\mathcal {N} (\\boldsymbol {p} (t); \\boldsymbol {p} (0), \\sigma^ {2} (t) \\mathbf {I})}} \\left[ \\left\\| \\boldsymbol {s} (\\boldsymbol {p} (t), t | O, I) - \\frac {\\boldsymbol {p} (0) - \\boldsymbol {p} (t)}{\\sigma (t) ^ {2}} \\right\\| _ {2} ^ {2} \\right] \\right\\} \\tag{3}", + "image_path": "4cba938b314231d923b9757ebbe2eca0d23fcff5e408841a87c8187a5c660efe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "text", + "content": "The training loss of the score and energy network can be obtained by replacing " + }, + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "inline_equation", + "content": "s(\\pmb{p}(t), t|O, I)" + }, + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "text", + "content": " in Eq. 3 with " + }, + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{\\varPhi}_{\\theta}(\\pmb{p}(t), t|O, I)" + }, + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{p}}\\pmb{\\Psi}_{\\phi}^{*}(\\pmb{p}(t), t|O, I)" + }, + { + "bbox": [ + 131, + 577, + 482, + 602 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 602, + 482, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 482, + 673 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 482, + 673 + ], + "type": "text", + "content": "Unlike GenPose, our score and energy network are semantic-aware, as both networks are conditioned on an RGB image to incorporate semantic cues for pose estimation. To fuse the features extracted from the image and point cloud, we encode the RGB image img and point cloud obj using the pre-trained feature extractors from DINOv2 [20] and PointNet++ [22], respectively. Then, we concatenate these features together in a pointwise manner similar to [27]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 386, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 386, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 386, + 128 + ], + "type": "text", + "content": "4.2 Candidates Generation and Outlier Removal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "text", + "content": "Following GenPose, we subsequently sample pose candidates " + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\{\\hat{p}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "text", + "content": " by solving the following Probability Flow (PF) ODE [24] constructed by the score network " + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\Phi_\\theta" + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "inline_equation", + "content": "t = \\epsilon" + }, + { + "bbox": [ + 130, + 133, + 482, + 169 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 240, + 175, + 481, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 175, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 240, + 175, + 481, + 196 + ], + "type": "interline_equation", + "content": "\\frac {d \\boldsymbol {p}}{d t} = - \\sigma (t) \\dot {\\sigma} (t) \\boldsymbol {\\Phi} _ {\\theta} (\\boldsymbol {p} (t), t | O, I) \\tag {4}", + "image_path": "c5b66c59496d988c837d5752f6aacb864f7e73db96a45a2bb1a386578d3b5526.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "inline_equation", + "content": "\\pmb{p}(1) \\sim \\mathcal{N}(\\mathbf{0}, \\sigma_{\\max}^2 \\mathbf{I})" + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "inline_equation", + "content": "\\sigma(t) = \\sigma_{\\min} \\left( \\frac{\\sigma_{\\max}}{\\sigma_{\\min}} \\right)^t" + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "inline_equation", + "content": "\\sigma_{\\min} = 0.01" + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "inline_equation", + "content": "\\sigma_{\\max} = 50" + }, + { + "bbox": [ + 130, + 201, + 462, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 214, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 479, + 239 + ], + "type": "text", + "content": "To remove the outliers in candidates, we sort the candidates into a sequence " + }, + { + "bbox": [ + 130, + 214, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_{\\tau_1} \\succ \\hat{\\pmb{p}}_{\\tau_2} \\dots \\succ \\hat{\\pmb{p}}_{\\tau_K}" + }, + { + "bbox": [ + 130, + 214, + 479, + 239 + ], + "type": "text", + "content": " where:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 246, + 481, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 246, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 211, + 246, + 481, + 259 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} _ {\\tau_ {i}} \\succ \\hat {\\boldsymbol {p}} _ {\\tau_ {j}} \\Longleftrightarrow \\boldsymbol {\\Psi} _ {\\phi} (\\hat {\\boldsymbol {p}} _ {\\tau_ {i}}, \\epsilon | O) > \\boldsymbol {\\Psi} _ {\\phi} (\\hat {\\boldsymbol {p}} _ {\\tau_ {j}}, \\epsilon | O) \\tag {5}", + "image_path": "9db332348143993f19555ff3db871f8e3b10efc57d7031f0e7576d78019dec62.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "text", + "content": "Then, we filter out the last " + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "inline_equation", + "content": "1 - \\delta \\%" + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "text", + "content": " candidates and obtain " + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_{\\tau_1} \\succ \\hat{\\pmb{p}}_{\\tau_2} \\dots \\succ \\hat{\\pmb{p}}_{\\tau_M}" + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "inline_equation", + "content": "\\delta = 40\\%" + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "text", + "content": " is a hyper parameter and " + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "inline_equation", + "content": "M = \\lfloor \\delta \\cdot K \\rfloor" + }, + { + "bbox": [ + 131, + 263, + 479, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 297, + 310, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 297, + 310, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 310, + 308 + ], + "type": "text", + "content": "4.3 Clustering-based Aggregation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "type": "text", + "content": "In this section, we aggregate the remaining candidates " + }, + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\pmb{p}}_{\\tau_i} = (\\hat{T}_{\\tau_i},\\hat{R}_{\\tau_i})\\}_{i = 1}^M" + }, + { + "bbox": [ + 130, + 308, + 482, + 416 + ], + "type": "text", + "content": " to obtain the final results. GenPose achieves this by simply mean-pooling the filtered candidates. However, this strategy will encounter a severe mean-mode issue when the object possesses discrete symmetrical properties. As illustrated in Fig. 6, a ping pong paddle has two symmetric ground truth poses (modes). Since the score network has encountered both modes during training, an optimally trained score network will likely output candidates around both modes. Simply mean-pooling these candidates will yield the average of the two modes, known as the 'mean mode', which will deviate from both modes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": "To mitigate this issue, we introduce a clustering-based aggregation mechanism. We employ DBSCAN [8] to cluster the candidates. It identify dense regions in the data space, forming clusters based on the density of data points and effectively separating noise from meaningful patterns, without the need to specify the number of clusters. This is achieved through dynamic determination of cluster quantities based on distance threshold " + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "inline_equation", + "content": "(\\varepsilon)" + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": " and density threshold (MinPts). For instance, in our empirical setting, we set " + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\varepsilon \\approx 0.45" + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": " rad and MinPts " + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "inline_equation", + "content": "= 5" + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": ". After clustering the candidates, we select the cluster with the largest number of objects and get the mean-pooling result as the final estimation following GenPose." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 535, + 230, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 230, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 230, + 548 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 552, + 378, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 378, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 378, + 565 + ], + "type": "text", + "content": "5.1 Category-Level 6D Object Pose Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Metric In prior studies, metrics such as the mean average precision (mAP) for 3D bounding box IoU and the mean average precision (mAP) for objects with translation errors less than " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " cm and rotation errors less than " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "n^{\\circ}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " have been commonly used [28]. This evaluation typically involves two steps: instance-level object segmentation and pose estimation from these detections. However, these metrics are influenced by the detection model's performance. To focus solely on evaluating the precision of pose estimation, we assume ground truth instance segmentation is known and propose the following two metrics:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 163, + 476, + 256 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Quantitative comparison of category-level pose estimation on ROPE. \\(\\uparrow\\) represents a higher value indicating better performance, while \\(\\downarrow\\) represents a lower value indicating better performance. Prior-free indicates whether the method requires category prior. \\{'-'\\) indicates that GenPose does not predict the object scale." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 163, + 476, + 256 + ], + "lines": [ + { + "bbox": [ + 141, + 163, + 476, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 163, + 476, + 256 + ], + "type": "table", + "html": "
MethodInputPrior-freeAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
DeterministicNOCS [28]RGBD✓0.00.00.00.00.00.00.0
SGPA [3]RGBD✗10.52.00.04.36.79.315.0
IST-Net [16]RGBD✓28.710.60.52.03.45.38.8
HS-Pose [44]D✓31.613.61.13.55.38.412.7
ProbabilisticGenPose [43]D✓---6.69.613.119.3
GenPose++(Ours)RGBD✓39.019.12.010.015.119.529.4
", + "image_path": "a7d38b6fc40da845b9125db99ea6d63455fa730cdeca388d79157a8f82bbf78b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 269, + 480, + 389 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "text", + "content": "- AUC@IoU" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "inline_equation", + "content": "_n" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "text", + "content": ": This metric assesses the accuracy of predicted 3D bounding boxes, calculated via the Area Under the Curve (AUC) from various Intersection over Union (IoU) thresholds starting at " + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "text", + "content": ". In our study, we utilize AUC@IoU" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "inline_equation", + "content": "_{25}" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "text", + "content": ", AUC@IoU" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "inline_equation", + "content": "_{50}" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "text", + "content": ", and AUC@IoU" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "inline_equation", + "content": "_{75}" + }, + { + "bbox": [ + 138, + 269, + 479, + 316 + ], + "type": "text", + "content": " as the benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": "- VUS@" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "inline_equation", + "content": "n^{\\circ}mcm" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": ": This metric offers a detailed analysis of 6D pose estimation accuracy, derived from the Volume Under Surface (VUS) across ranges of rotational (up to " + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "inline_equation", + "content": "n^{\\circ}" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": ") and translational (up to " + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "inline_equation", + "content": "mcm" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": ") error thresholds. VUS aggregates the accuracy of pose predictions within set boundaries. In this paper, we apply VUS@" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "inline_equation", + "content": "5^{\\circ}2cm" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": ", VUS@" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "inline_equation", + "content": "5^{\\circ}5cm" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": ", VUS@" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "inline_equation", + "content": "10^{\\circ}2cm" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": ", and VUS@" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "inline_equation", + "content": "10^{\\circ}5cm" + }, + { + "bbox": [ + 138, + 318, + 480, + 389 + ], + "type": "text", + "content": " for comprehensive performance assessment." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "type": "text", + "content": "Additionally, to demonstrate the robustness of our method to segmentation masks, we present the pose estimation performance of GenPose++ using SAM [13] segmentation results in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 442, + 481, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 481, + 562 + ], + "type": "text", + "content": "Baselines We evaluate five category-level pose estimation methods: NOCS [28], SGPA [3], HS-Pose [44], IST-Net [16] and GenPose [43]. Except for NOCS, which conducts both object detection and pose estimation as a whole, all methods are equipped with ground truth detection results. For SGPA, the prior point cloud of each category is constructed by randomly selecting an object from the training dataset. Considering that previous methods' augmentations for symmetry properties are only applicable to specific object categories within the NOCS dataset and not suitable for all object categories in Omni6Dpose, data augmentation for object symmetry is disabled during training for all baseline methods. All methods are trained on SOPE and directly tested on ROPE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Results and Analysis. In Table 2, we present the quantitative evaluation results of previous methods compared to GenPose++ on the ROPE dataset. Overall, generative methods continue to dominate in the performance evaluation on ROPE. The VUS surface depicted in Figure 7 provides a more detailed reflection of the performance of each model. Unlike deterministic approaches, generative methods can handle ambiguity without any specialized design requirements. Moreover, these methods directly generate the distribution of object poses, eliminating the need for depth map-based pose fitting. This approach is particularly" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 239, + 195 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 239, + 195 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 239, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 239, + 195 + ], + "type": "image", + "image_path": "8577d558e9ccc200a47273e9044677f58ad2053039a13285e0607224c4d64cc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 116, + 350, + 194 + ], + "blocks": [ + { + "bbox": [ + 246, + 116, + 350, + 194 + ], + "lines": [ + { + "bbox": [ + 246, + 116, + 350, + 194 + ], + "spans": [ + { + "bbox": [ + 246, + 116, + 350, + 194 + ], + "type": "image", + "image_path": "7b2051dcc5293d9a7fe0796028210d31c06e74c9b9890724cda7329cb842a01e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 357, + 116, + 482, + 194 + ], + "blocks": [ + { + "bbox": [ + 357, + 116, + 482, + 194 + ], + "lines": [ + { + "bbox": [ + 357, + 116, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 357, + 116, + 482, + 194 + ], + "type": "image", + "image_path": "c077af05b498de02cdc8eab0c596a7d8c21efe75484d6d0273e29d17bcfd9d10.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 197, + 238, + 273 + ], + "blocks": [ + { + "bbox": [ + 132, + 197, + 238, + 273 + ], + "lines": [ + { + "bbox": [ + 132, + 197, + 238, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 238, + 273 + ], + "type": "image", + "image_path": "fec35d4ed07e2909d968f20e09d0c1a7fd9731396d3a649319ce5e8374292dad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 278, + 455, + 290 + ], + "lines": [ + { + "bbox": [ + 156, + 278, + 455, + 290 + ], + "spans": [ + { + "bbox": [ + 156, + 278, + 455, + 290 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparison with baselines on ROPE dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 246, + 197, + 350, + 273 + ], + "blocks": [ + { + "bbox": [ + 246, + 197, + 350, + 273 + ], + "lines": [ + { + "bbox": [ + 246, + 197, + 350, + 273 + ], + "spans": [ + { + "bbox": [ + 246, + 197, + 350, + 273 + ], + "type": "image", + "image_path": "03ce1c66280d01f8a3c7411edbf255b53f6b004a790450d12fe12ac5dee565f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 358, + 197, + 481, + 274 + ], + "blocks": [ + { + "bbox": [ + 358, + 197, + 481, + 274 + ], + "lines": [ + { + "bbox": [ + 358, + 197, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 358, + 197, + 481, + 274 + ], + "type": "image", + "image_path": "2525cd83110824542c6fb1a551052ff2447259b4ec7dd52b968d177337a488e9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": "advantageous for challenging material types, such as transparent or reflective objects, where structured-light depth cameras tend to introduce significant noise, severely impacting pose fitting accuracy. Furthermore, the NOCS method does not demonstrate effective performance on the ROPE dataset, leading to the supposition that methods relying solely on RGB information to predict the shape of an object in the canonical space become less robust as the scale of category diversity increases. Compared to GenPose, GenPose++ achieves a significant lead by leveraging the powerful perception capabilities of the 2D foundation model, along with the robustness of clustering towards discrete symmetric properties. You can find qualitative visualizations in Figure 8." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 423, + 238, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 238, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 238, + 435 + ], + "type": "text", + "content": "5.2 Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 443, + 481, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 481, + 468 + ], + "type": "text", + "content": "In order to validate the design decisions of our approach, we performed a series of ablation experiments on our method:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 476, + 480, + 572 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 476, + 480, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 500 + ], + "type": "text", + "content": "- w/o clustering. Without clustering directly take the average of all remaining pose candidates after outlier removal as the pose estimation output." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 501, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 501, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 137, + 501, + 480, + 536 + ], + "type": "text", + "content": "- w/o scale prediction. Use the estimated pose to transform the observed point cloud into object space, then take the bounding box length as the maximum projection from the point cloud to each axis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 395, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 395, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 395, + 548 + ], + "type": "text", + "content": "- w/o simulated depth. Use perfect depth for training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 548, + 480, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 572 + ], + "type": "text", + "content": "- w/o point-wise feature fusion. In the feature extraction stage, separately extract the RGB feature and geometric feature, and then concatenation." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Table 3 illustrates the contribution of each component of GenPose++ to its performance. The introduction of the clustering module allows GenPose++ to effectively aggregate the multimodal distributions caused by discrete symmetries, leading to higher performance. The scale prediction in GenPose++ significantly surpasses direct calculations from the object's point cloud due to ambiguities from partial observations and errors from point cloud noise, particularly in transparent and reflective objects. Training with simulated depth data results in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 130, + 452, + 228 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 466, + 126 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 466, + 126 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 466, + 126 + ], + "type": "text", + "content": "Table 3: Ablation study on category-level 6D object pose estimation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 130, + 452, + 228 + ], + "lines": [ + { + "bbox": [ + 159, + 130, + 452, + 228 + ], + "spans": [ + { + "bbox": [ + 159, + 130, + 452, + 228 + ], + "type": "table", + "html": "
AblationAUC ↑VUS ↑
IoU25IoU50IoU755°2cm5°5cm10°2cm10°5cm
Full39.019.12.010.015.119.529.4
w/o scale prediction13.83.40.210.015.119.529.4
w/o clustering38.618.71.99.414.118.427.8
w/o simulated depth35.316.51.77.311.914.624.3
w/o point-wise fusion34.215.41.47.110.614.621.9
", + "image_path": "3afd96b96af6e85dcbc6fd6490af7edbf579f44495b34ee8436c5e695e886d30.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 239, + 480, + 366 + ], + "blocks": [ + { + "bbox": [ + 135, + 239, + 480, + 366 + ], + "lines": [ + { + "bbox": [ + 135, + 239, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 135, + 239, + 480, + 366 + ], + "type": "image", + "image_path": "0adc22a2ab6a498e57403c9ae654072fee93fcf84548c0e1cdf8b1a97d1e38dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 370, + 455, + 381 + ], + "lines": [ + { + "bbox": [ + 156, + 370, + 455, + 381 + ], + "spans": [ + { + "bbox": [ + 156, + 370, + 455, + 381 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparison with baselines on ROPE dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 480, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 440 + ], + "type": "text", + "content": "better performance than training with perfect point clouds, as the physics-based depth camera simulation substantially reduces the sim2real gap for depth data. Point-wise fusion outperforms global fusion as it retains more of the object's local geometric features, which are crucial for accurate object pose prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 446, + 368, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 446, + 368, + 459 + ], + "spans": [ + { + "bbox": [ + 131, + 446, + 368, + 459 + ], + "type": "text", + "content": "5.3 Category-Level 6D Object Pose Tracking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 464, + 472, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 472, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 472, + 477 + ], + "type": "text", + "content": "Metric. We report the following metrics for object pose tracking evaluation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 483, + 480, + 564 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 137, + 483, + 455, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 483, + 455, + 494 + ], + "spans": [ + { + "bbox": [ + 137, + 483, + 455, + 494 + ], + "type": "text", + "content": "- FPS: Frames Per Second, which indicates the speed of pose tracking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 495, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 495, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 495, + 480, + 517 + ], + "type": "text", + "content": "- VUS@5°5cm: Volume Under Surface, assessing pose estimation accuracy for rotation errors within 0 to " + }, + { + "bbox": [ + 138, + 495, + 480, + 517 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 138, + 495, + 480, + 517 + ], + "type": "text", + "content": " and translation errors within 0 to 5cm." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 517, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 517, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 480, + 540 + ], + "type": "text", + "content": "- mIoU: Mean Intersection over Union, representing the average 3D overlap between the ground truth and the predicted bounding boxes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 541, + 341, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 541, + 341, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 541, + 341, + 553 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 541, + 341, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{Rerr}(\\circ)" + }, + { + "bbox": [ + 138, + 541, + 341, + 553 + ], + "type": "text", + "content": ": Average rotation error in degrees." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 553, + 378, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 553, + 378, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 553, + 378, + 564 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 553, + 378, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{Terr}(\\mathbf{cm})" + }, + { + "bbox": [ + 138, + 553, + 378, + 564 + ], + "type": "text", + "content": ": Average translation error in centimeters." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 577, + 482, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 482, + 674 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 482, + 674 + ], + "type": "text", + "content": "Baselines. This paper employs BundleTrack, CATRE, and GenPose as baselines for object pose tracking. BundleTrack is a training-free approach that utilizes multi-view feature point detection and matching for tracking the pose of unseen objects. CATRE aligns partially observed point clouds to abstract shape priors to estimate relative transformations, enhancing pose accuracy. Conversely, GenPose utilizes a generative approach to effectively resolve pose ambiguities, notably in symmetric objects. Following GenPose, we have adapted GenPose++ to object pose tracking with minor modifications." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "Results and Analysis. Table 4 presents the results of category-level object pose tracking algorithms CATRE, GenPose, and our method, as well as the results for the unseen object pose tracking algorithm BundleTrack. The category-level pose estimation methods appear to achieve relatively better outcomes since they benefit from learning the category-level canonical space of objects within the SOPE dataset. However, for the training-free BundleTrack, reliance on RGB information for keypoint detection and matching poses challenges, often failing on objects with weak textures. Additionally, its dependency on depth values for global optimization renders it less effective in handling instances with substantial depth noise, such as transparent or specular objects. Our method, without undergoing specialized design, has achieved results comparable to state-of-the-art approaches. Although the inference speed of our method is lower than that of CATRE and GenPose, the achieved 17.8 FPS is sufficient for certain downstream tasks, such as robotic manipulation. Furthermore, recent rapid developments in research on fast samplers are beneficial to our approach, potentially enhancing its performance and applicability in real-time scenarios." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 356, + 476, + 423 + ], + "blocks": [ + { + "bbox": [ + 130, + 317, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 352 + ], + "type": "text", + "content": "Table 4: Results of category-level object pose tracking on ROPE. The results are averaged over all 149 categories. GT. Pert. denotes that a perturbed ground truth pose is utilized as the initial object pose." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 356, + 476, + 423 + ], + "lines": [ + { + "bbox": [ + 135, + 356, + 476, + 423 + ], + "spans": [ + { + "bbox": [ + 135, + 356, + 476, + 423 + ], + "type": "table", + "html": "
MethodsInputInit.Speed(FPS)↑5°5cm↑mIoU↑Rerr(°)↓Terr(cm)↓
BundleTrack [31]RGBDGT.12.41.33.946.923.5
CATRE [18]DGT. Pert.38.515.955.421.32.6
GenPose [43]DGT. Pert.26.313.3-19.31.2
GenPose++(Ours)RGBDGT. Pert.17.815.953.417.61.2
", + "image_path": "f634bc31b432a88ca62eefc05fecd81ccbed471dff7b48076fe4609e385c2e08.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 438, + 317, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 317, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 317, + 451 + ], + "type": "text", + "content": "6 Conclusions and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 594 + ], + "type": "text", + "content": "In this study, we introduce Omni6DPose, a comprehensive dataset for 6D object pose estimation, featuring extensive scale, diversity, and material variety. Through thorough experimentation, our findings suggest that the probabilistic framework holds promise for category-level 6D object estimation, leveraging semantic information provided by RGB images to address large-scale pose estimation challenges. However, the performance of GenPose++ on Omni6DPose reveals significant room for improvement, with the model still hampered by slow inference speeds resulting from the iterative refinement nature inherent in the diffusion model. Future works could focus on addressing these challenges and integrating the universal 6D pose estimation module trained on Omni6DPose into a broader range of downstream tasks [4,33,37,42]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 605, + 255, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 255, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 255, + 619 + ], + "type": "text", + "content": "7 Acknowledgment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China - General Program (Project ID: 62376006), the National Youth Talent Support Program (Project ID: 8200800081), and NSFC-62306046." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "text", + "content": "1. An, B., Geng, Y., Chen, K., Li, X., Dou, Q., Dong, H.: RGBmanip: Monocular image-based robotic manipulation through active object pose estimation (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "text", + "content": "4. Cheng, J., Wu, M., Zhang, R., Zhan, G., Wu, C., Dong, H.: Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 250, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 281 + ], + "type": "text", + "content": "5. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4800-4810 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 326 + ], + "type": "text", + "content": "6. Dai, Q., Zhang, J., Li, Q., Wu, T., Dong, H., Liu, Z., Tan, P., Wang, H.: Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and transparent objects. In: European Conference on Computer Vision. pp. 374-391. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 327, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 369 + ], + "type": "text", + "content": "7. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "type": "text", + "content": "8. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: kdd. vol. 96, pp. 226-231 (1996)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "9. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 436, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 436, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 134, + 436, + 480, + 479 + ], + "type": "text", + "content": "10. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. vol. 7724 (10 2012). https://doi.org/10.1007/978-3-642-33885-4_60" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 480, + 512 + ], + "type": "text", + "content": "11. Hodań, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X.: T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. IEEE Winter Conference on Applications of Computer Vision (WACV) (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 513, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 513, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 513, + 480, + 555 + ], + "type": "text", + "content": "12. Jung, H., Zhai, G., Wu, S.C., Ruhkamp, P., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., Roth, D., Navab, N., Busam, B.: Housecat6d - a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 556, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 599 + ], + "type": "text", + "content": "13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "type": "text", + "content": "14. Lin, Z.H., Huang, S.Y., Wang, Y.C.F.: Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "15. Liu, J., Sun, W., Liu, C., Zhang, X., Fan, S., Wu, W.: Hff6d: Hierarchical feature fusion network for robust 6d object pose tracking. IEEE Transactions on Circuits" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 137 + ], + "type": "text", + "content": "and Systems for Video Technology pp. 1-13 (06 2022). https://doi.org/10.1109/TCSVT.2022.3181597" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 159 + ], + "type": "text", + "content": "16. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. arXiv preprint arXiv:2303.13479 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 180 + ], + "type": "text", + "content": "17. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 181, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 181, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 481, + 213 + ], + "type": "text", + "content": "18. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: European Conference on Computer Vision (ECCV) (October 2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "text", + "content": "19. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics 22(12), 2633-2651 (2016). https://doi.org/10.1109/TVCG.2015.2513408" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 277 + ], + "type": "text", + "content": "20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 319 + ], + "type": "text", + "content": "21. Piga, N.A., Onyshchuk, Y., Pasquale, G., Pattacini, U., Natale, L.: Roft: Realtime optical flow-aided 6d object pose and velocity tracking. IEEE Robotics and Automation Letters 7(1), 159–166 (2022). https://doi.org/10.1109/LRA.2021.3119379" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 353 + ], + "type": "text", + "content": "22. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "type": "text", + "content": "23. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "24. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "text", + "content": "25. Tian, Y., Zhang, J., Huang, G., Wang, B., Wang, P., Pang, J., Dong, H.: Robokeygen: Robot pose and joint angles estimation via diffusion-based 3d keypoint generation. arXiv preprint arXiv:2403.18259 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "text", + "content": "26. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 460, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 502 + ], + "type": "text", + "content": "27. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 503, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 481, + 545 + ], + "type": "text", + "content": "28. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "text", + "content": "29. Wang, J., Chen, K., Dou, Q.: Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4807-4814. IEEE (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 632 + ], + "type": "text", + "content": "30. Wang, P., Jung, H., Li, Y., Shen, S., Srikanth, R.P., Garattoni, L., Meier, S., Navab, N., Busam, B.: Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21222-21231 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "31. Wen, B., Bekris, K.E.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 466, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 466, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 466, + 102 + ], + "type": "text", + "content": "J. Zhang, W. Huang, B. Peng, M. Wu, F. Hu, Z. Chen, B. Zhao, H. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 588 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "32. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 13209-13218 (October 2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "33. Wu, M., Zhong, F., Xia, Y., Dong, H.: Targf: Learning target gradient field to rearrange objects without explicit goal specification. Advances in Neural Information Processing Systems 35, 31986-31999 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "34. Wu, T., Wu, M., Zhang, J., Gan, Y., Dong, H.: Learning score-based grasping primitive for human-assisting dexterous grasping. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "35. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., et al.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 803-814 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "36. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "37. Xue, T., Wu, M., Lu, L., Wang, H., Dong, H., Chen, B.: Learning gradient fields for scalable and generalizable irregular packing. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12-22 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 423 + ], + "type": "text", + "content": "40. You, Y., Shi, R., Wang, W., Lu, C.: CPPf: Towards robust category-level 9d pose estimation in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6866-6875 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "41. You, Y., Xiong, K., Yang, Z., Huang, Z., Zhou, J., Shi, R., Fang, Z., Harley, A.W., Lu, C.: Pace: Pose annotations in cluttered environments. arXiv preprint arXiv:2312.15130 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "42. Zeng, Y., Wu, M., Yang, L., Zhang, J., Ding, H., Cheng, H., Dong, H.: Distilling functional rearrangement priors from large models. arXiv preprint arXiv:2312.01474 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "text", + "content": "43. Zhang, J., Wu, M., Dong, H.: Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "text", + "content": "44. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17163-17173 (June 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "45. Zhou, L., Liu, Z., Gan, R., Wang, H., au2, M.H.A.J.: Dr-pose: A two-stage deformation-and-registration pipeline for category-level 6d object pose estimation (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6DPose" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_content_list.json b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd580114c11d8bac1aa9e8c442105b829e96612b --- /dev/null +++ b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_content_list.json @@ -0,0 +1,1730 @@ +[ + { + "type": "text", + "text": "Omni6D: Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation", + "text_level": 1, + "bbox": [ + 232, + 141, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mengchen Zhang $^{1,2}$ , Tong Wu $^{3}$ , Tai Wang $^{2}$ , Tengfei Wang $^{2}$ , Ziwei Liu $^{4}$ , and Dahua Lin $^{2,3}$", + "bbox": [ + 227, + 210, + 776, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Zhejiang University, Zhejiang, China", + "bbox": [ + 369, + 253, + 630, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai Artificial Intelligence Laboratory, Shanghai, China", + "bbox": [ + 292, + 268, + 709, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The Chinese University of Hong Kong, Hong Kong SAR", + "bbox": [ + 305, + 281, + 694, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Nanyang Technological University, Singapore", + "bbox": [ + 341, + 296, + 658, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangmengchen, wangtai, wangtengfei}@pjlab.org.cn,", + "bbox": [ + 308, + 311, + 689, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wt020,dhlin}@ie.cuhk.edu.hk,ziwei.liu@ntu.edu.sg", + "bbox": [ + 307, + 325, + 694, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55b5a6610150eb955d988781bccb266b91dce88136f8d16961fa3019f16d46c2.jpg", + "image_caption": [ + "Fig.1: Omni6D is a dataset for 6D object pose and size estimation with large vocabulary categories and rich annotations. (a) showcases ground truth of RGB image, depth map and NOCS map. (b) presents shape priors derived from a variational autoencoder [5] with adjusted canonical poses. (c) provides examples of the rotational symmetry of objects we have annotated, indicating the multiples of angles by which the shape remains unchanged when rotated around the xyz axes." + ], + "image_footnote": [], + "bbox": [ + 223, + 364, + 782, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 6D object pose estimation aims at determining an object's translation, rotation, and scale, typically from a single RGBD image. Recent advancements have expanded this estimation from instance-level to category-level, allowing models to generalize across unseen instances within the same category. However, this generalization is limited by the narrow range of categories covered by existing datasets, such as NOCS, which also tend to overlook common real-world challenges like occlusion. To tackle these challenges, we introduce Omni6D, a comprehensive RGBD dataset featuring a wide range of categories and varied backgrounds, elevating the task to a more realistic context. 1) The dataset comprises an extensive spectrum of 166 categories, 4688 instances adjusted to the canonical pose, and over 0.8 million captures, significantly broadening the scope for evaluation. 2) We introduce a symmetry-aware", + "bbox": [ + 259, + 659, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "metric and conduct systematic benchmarks of existing algorithms on Omni6D, offering a thorough exploration of new challenges and insights. 3) Additionally, we propose an effective fine-tuning approach that adapts models from previous datasets to our extensive vocabulary setting. We believe this initiative will pave the way for new insights and substantial progress in both the industrial and academic fields, pushing forward the boundaries of general 6D pose estimation.", + "bbox": [ + 259, + 147, + 743, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: 6DoF Pose Estimation $\\cdot$ Large Vocabulary Dataset $\\cdot$ Metrics and Benchmarks", + "bbox": [ + 261, + 258, + 743, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 316, + 375, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "6D pose estimation aims at predicting the position, orientation, and size of objects in a 3D space using RGB (D) images, enabling various applications such as augmented/virtual reality [26,33], robot manipulation [11,35], and scene understanding [15,28].", + "bbox": [ + 212, + 353, + 787, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early instance-level pose estimation approaches [32, 38, 39, 42, 43] typically involve providing instance CAD models and predicting poses of instances that were seen during training, restricting the generalization to unseen objects. In contrast, recent research has shifted towards category-level 6D object pose estimation [6-8, 10, 16, 17, 20, 24, 25, 29, 34, 37, 40, 44-47], which learns category prior from a large number of instances within a category, allowing for pose estimation of new instances within the same category without the need for CAD models. By learning on a diverse range of categories, category-level approaches could be a more versatile solution for 6D pose estimation in real-world scenarios.", + "bbox": [ + 212, + 415, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, most existing datasets [22, 40, 44] are limited to a small number of object categories, typically less than 10, as shown in Tab. 1, hindering their practical applicability to complex scenes.", + "bbox": [ + 212, + 551, + 787, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations in previous category-level 6D pose estimation datasets, such as limited category numbers, lack of instance diversity within categories, and overly simplistic scenes, this paper presents a novel category-level dataset dubbed Omni6D for 6D pose estimation. Omni6D significantly extends the number of object categories to 166, and includes 4,688 real-scanned and well-annotated instance objects with a diverse range of shapes, sizes, and textures. The constructed benchmark includes 0.8M images featuring complex scenes with various occlusions, changing lighting conditions, complex backgrounds, and varying viewpoints. For each scene, we provide the rendered image, depth map, NOCS map, and instance mask. Also, considering the widespread rotational symmetry in objects, we examine three types of rotational invariance where an object maintains its original shape under following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we introduce a symmetry-aware metric to specifically address rotational invariance. Every object in Omni6D is adjusted to the canonical pose and annotated with rotational symmetry around three axes.", + "bbox": [ + 212, + 599, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/c956b8a1bf6ec64f44009d2c61c9b4f21b46619d7aea3acb8888e6b17a4644e6.jpg", + "table_caption": [ + "Table 1: Comparisons between Omni6D and existing datasets. Omni6D significantly extends the range of everyday object categories and instances." + ], + "table_footnote": [], + "table_body": "
DatasetsModeRealism# Categories # Instances# Images
ShapeNet-SRNCars [22]RGBSynthetic13514
Sim2Real Cars [22]RGBReal110
CAMERA [40]RGBDSynthetic61085
REAL [40]RGBDReal642
Wild6D [44]RGBDReal51722
Omni6DRGBDReal-Scanned1664,688
", + "bbox": [ + 259, + 184, + 741, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Including a broader range of categories, our dataset offers a more comprehensive and challenging evaluation benchmark for category-level 6D object pose estimation. Utilizing Omni6D, we train and analyze existing algorithms, initiating a profound exploration of the challenges and vital elements involved in category-level estimation within large-vocabulary categories. Additionally, we assess these algorithms' capability to generalize across categories, and carry out a category-wise analysis. Experiments show that our dataset presents a more challenging benchmark for 6D pose estimation, highlighting the need for more robust and generalized pose estimation approaches. As an initial attempt, we present a finetuning strategy that assists in broadening the scope of existing approaches from a limited range of categories to a broader vocabulary. Moreover, we conduct an analysis of the domain gap between our dataset and real-world dataset, emphasizing the benefits of their combined use.", + "bbox": [ + 217, + 305, + 785, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our dataset will be publicly available to the research community, which will foster future research on more practical and robust 6D pose estimation algorithms and pave the way for broader applications.", + "bbox": [ + 215, + 501, + 784, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 566, + 385, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing work on category-level 6D object pose estimation can be generally divided into two types. After extracting features from images or point clouds, they compute Rotation, Translation, and Size (RTS) either through implicit point correspondence or explicit regression.", + "bbox": [ + 215, + 598, + 784, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing Datasets. The most commonly used dataset for category-level 6D object pose estimation is NOCS [40], comprising both the synthetic CAMERA dataset and the real-world REAL dataset. CAMERA includes 300k RGBD images of 31 indoor scenes with 1,085 object instances across 6 categories, while REAL mirrors the categories in CAMERA and includes 8k RGBD images capturing 42 instances in 18 real scenes. Wild6D [44] consists of 5,166 videos with 1.1 million images over 1,722 object instances in 5 categories. ShapeNet-SRN Cars dataset and Sim2Real Cars dataset proposed in iNerf [22] both exclusively include a single car category. The former includes 3,514 instances derived from ShapeNet cars, while the latter is extracted from videos capturing 10 distinct unseen car models. These datasets are limited by their narrow range of categories, hindering their ability to generalize broadly. Additionally, most training images", + "bbox": [ + 217, + 659, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are synthetic and lack realism, and their scenes are overly simplified, failing to account for common real-world challenges like occlusions.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implicit Methods. Implicit methods are based on point correspondence [6, 20, 24, 34, 37, 40, 44, 46]. NOCS [40], one of the pioneering works in this area, introduced the concept of Normalized Object Coordinate Space (NOCS). The final pose and size of the object are obtained by matching the predicted NOCS map with the observed depth input using the Umeyama algorithm [36] and RANSAC algorithm [12].", + "bbox": [ + 212, + 178, + 785, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequent algorithms such as DualPoseNet, RBP-Net and RePoNet [20,44, 46] have continued to develop along the vein of NOCS, implicitly solving for pose after predicting the NOCS map. SPD [34] proposed a category-level shape prior, subsequently deforming this shape prior (i.e., average shape) to fit observed point cloud. SGPA, RePoNet, and CATRE [6,24,44] continue to develop along SPD's category-level shape prior approach. Algorithms like 6-PACK and SGPA [6,37] extract low-rank structure points, i.e., keypoints, from dense observed point clouds. 6-PACK [37] predicts interframe motion of target instances through keypoint matching, while SGPA [6] employs keypoints for more effective incorporation of sparse structural information during prior adaptation. These methods rely heavily on the RANSAC process to eliminate outliers, making them non-differentiable and time-consuming.", + "bbox": [ + 212, + 270, + 787, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explicit Methods. Explicit methods are based on direct pose regression [7, 10, 20, 24, 46, 47]. DualPoseNet and RBP-Net [20, 46] conduct both explicit and implicit training, where one parallel pose decoder explicitly regresses the pose. CATRE [24], recognizes the inherent difference between estimations of rotation and translation/size, explicitly regressing their residuals and carrying out an iterative pose estimation process. FS-Net [7] designs an autoencoder with 3D Graphic Convolution for latent feature extraction and separates the predictions for rotation and translation/size into two distinct networks: one estimates translation/size through two residuals, while the other handles rotation prediction by estimating deflections on two orthogonal axes. GPV-Pose and HS-Pose [10, 47] utilize the same foundational mechanism introduced by FS-Net [7]. GPV-Pose [10] proposes a decoupled confidence-driven rotation representation that facilitates geometrically-aware recovery of correlated rotation matrices and introduces a new geometry-guided point-by-point voting paradigm for robust retrieval of 3D object bounding boxes. Meanwhile, HS-Pose [47] extends 3D-GC to extract mixed-range latent features from point cloud data through a simple network structure known as the HS layer.", + "bbox": [ + 212, + 452, + 787, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Omni6D Dataset", + "text_level": 1, + "bbox": [ + 215, + 739, + 415, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Construction", + "text_level": 1, + "bbox": [ + 215, + 777, + 370, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset Collection. As shown in Tab. 1, Omni6D comprises 4,688 instances across an impressive span of 166 categories. Each instance is a high-resolution", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "textured mesh, obtained using Shining 3D scanner $^{1}$ and Artec Eva 3D scanner $^{2}$ , collected from OmniObject3D [41]. We normalize object models to fit within a $(-1,1)^{3}(m^{3})$ three-dimensional space, and align objects within each category to a consistent canonical pose. In the latest dataset, Omni6D-xl builds upon and extends Omni6D, comprising 13,161 instances across an impressive span of 339 categories. For more details, please refer to Appendix Section C.", + "bbox": [ + 212, + 145, + 782, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Rendering.** We employ stratified sampling to split instances within each category, subsequently dividing them into training, validation, and test sets in a 7:2:1 ratio. In the construction of our dataset, we utilize 9 room models from the Replica dataset as backdrops. For each scenery setup, we randomly select a room model to act as the background, along with 6 - 8 object instance models, which are allowed to perform free-fall motion within the room model, resulting in random scattering in a specific section of the room. Each object model is scaled by a random factor ranging from 0.8 to 1.2 as part of our data augmentation strategy. Considering the attention center of the combined instance models as the origin point, the camera randomly selects ten positions within a radius of $8 - 9m$ and an elevation angle range between $30 - 90^{\\circ}$ . The camera then performs rendering at these selected positions while facing towards the attention center.", + "bbox": [ + 212, + 237, + 784, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setting. We utilize BlenderProc 2.5.0 [9] to implement the aforementioned rendering process. The intrinsic parameters of the camera are set to [577.5, 577.5, 319.5, 239.5], with an image size specified as $640 \\times 480$ . Our approach ensures the diversity and breadth of the dataset, making it suitable for rigorous testing and yielding accurate results.", + "bbox": [ + 212, + 417, + 784, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Data Annotations", + "text_level": 1, + "bbox": [ + 214, + 518, + 408, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rich Annotations. Each rendered output includes the ground truth class label, instance mask, NOCS mapping [40], depth map, as well as 6D pose and size. Fig. 1 exhibits a selection of rendered outputs. To reduce the storage size of the dataset, we encode high-precision depth maps into RGB images by multiplying depth by 10,000, rounding to nearest integer, and converting to base 256. The resulting three digits represent RGB channels.", + "bbox": [ + 212, + 544, + 782, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rotational Invariance. Rotational invariance implies that a symmetric object can retain its original shape after rotation by certain angles. Many common objects have this property. As shown in Fig. 6, we define the coordinate system as a right-handed system with the x-axis pointing outwards and the y-axis oriented upwards. We contemplate three cases of rotational invariance where an object maintains its original shape after following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we denote the case of no rotational invariance around the axis as Sym-0. According to these definitions, all objects in Omni6D are annotated for their rotational symmetry around the xyz-axes. It's worth noting that symmetry attributes may differ among instances within the same category, requiring instance-level rather", + "bbox": [ + 212, + 635, + 784, + 800 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 https://www.einscan.com/", + "bbox": [ + 217, + 810, + 411, + 825 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www.artec3d.cn/", + "bbox": [ + 218, + 825, + 401, + 839 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21a06c09d06bd4bceb684939b8b7eea0d18a22f25c84e14e1f783f15922c5b0f.jpg", + "image_caption": [ + "Fig. 2: Symmetry statistics. The figure demonstrates different symmetry cases using object instances and provides a quantitative representation of the occurrence frequency for various combinations of distinct symmetry cases across the xyz-axes." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "than category-level annotations. Fig. 6 illustrates all kinds of symmetry cases using object instances and quantifies their occurrence frequency. Fig. 1 selects several examples to provide a more visual explanation of rotational invariance. These considerations are then integrated into our evaluation protocols in Sec. 4.2.", + "bbox": [ + 212, + 463, + 787, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Dataset Statistics", + "text_level": 1, + "bbox": [ + 215, + 544, + 408, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Statistics. Omni6D aims to overcome challenges in estimating poses for occluded object instances. Fig. 3a and Fig. 3b shows the spatial distribution of point clouds and objects by projecting their centroids on the XY-plane (top) and XZ-plane (bottom) [4]. Fig. 3c depicts the relative object size distribution, defined as the square root of the object-to-image area ratio. We observe that the spatial distribution of Omni6D is similar to that of CAMERA and REAL, with a greater resemblance to CAMERA despite having a closer depth range. However, a more pronounced discrepancy between the spatial distribution of point clouds and objects is evident in Omni6D compared to CAMERA and REAL. This observation suggests a higher occurrence of occlusion scenes in Omni6D, highlighting the intricate challenges it presents to 6D object pose estimation. Nonetheless, as depicted in Fig. 4a, algorithms trained on Omni6D demonstrate their robustness in tackling these complexities.", + "bbox": [ + 212, + 568, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Angular Deviation. Omni6D enables accurate pose estimation using only the lower half or bottom appearance of objects. Fig. 3d depicts the density of angular deviations from the upward direction, i.e. y-axis. Our dataset displays a more uniform distribution of object angles relative to the upward axis and exhibits greater deviation from the canonical pose angles. Unlike NOCS, which primarily", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8fee4c399f65e261ce78ad239b6c70672ec07711d2cd00d69f74712914c2c832.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 145, + 364, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39c29b7165ee116496b5b46d2b01809f3ea22361cecf798569bb9851a1b529a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 146, + 517, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d37c58de0e61aa6e6d242ce52c589c3410f4c955e4e848f5f7bba27e754f4a4.jpg", + "image_caption": [ + "(c) Relative 2D object size" + ], + "image_footnote": [], + "bbox": [ + 223, + 220, + 370, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d2a6dc09651eda420e55261b4d686340073f719d92e6cbff552b2ab33ce5d30.jpg", + "image_caption": [ + "(d) Angular deviation" + ], + "image_footnote": [], + "bbox": [ + 390, + 220, + 521, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c18e3c527805bd2f8dd6ab3bc6734b5033be4bee52a798a040a219a804cd5ad0.jpg", + "image_caption": [ + "(e) Clustering results", + "Fig. 3: Omni6D analysis. (a) distribution of point cloud centroids, (b) distribution of object centroids on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (c) density of relative 2D object size, (d) density of angular deviation from the upward direction, (e) Omni6D dataset clustering results. The angle of each sector in the chart reflects the relative size of the instance count within that category." + ], + "image_footnote": [], + "bbox": [ + 534, + 146, + 736, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "uses upright object placement, Omni6D utilizes physical simulations for freefall object positioning [9]. As a result, it presents more challenging and diverse pose estimation scenes. Training on Omni6D enhances algorithms' robustness to object rotation angles, as evidenced by the image in Fig. 4b.", + "bbox": [ + 212, + 430, + 784, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shape Priors. We obtain the mean latent embedding and shape prior for each category from the variational autoencoder [5]. Fig. 1 showcases categorical shape priors, each displaying unique characteristics, facilitating an intuitive association between point cloud shapes and corresponding real-world entities. Meanwhile, Fig. 3e explains clustering results based on categorical latent embeddings, where we employ agglomerative clustering [27] to group categories into 20 clusters. It highlights the geometric coherence among semantically identical objects (especially man-made ones) in Omni6D dataset and further confirms that these categorical shape priors can effectively leverage the wealth of shape information from numerous similar objects to elucidate category features. These insights provide a theoretical basis for applications of category-level 6D object pose estimation using our Omni6D dataset.", + "bbox": [ + 212, + 491, + 785, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Evaluation and Analysis", + "text_level": 1, + "bbox": [ + 215, + 694, + 485, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 724, + 426, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. Our experimentation utilized two datasets, namely Omni6D and Omni6Dout. Omni6D are partitioned into training, validation, and test sets in a 7:2:1 ratio, denoted as Omni6Dtrain, Omni6Dval and Omni6Dtest respectively. These sets are further subdivided into subsets with increasing category sizes of 3, 6, 12, 24, and 48. We denote the subset containing $n$ categories as c1sn. Each subset includes all classes present in the previous subset with additional", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71ab0bbfbe8f9e114c0dfa45c2482f0cda0dc4ae1eca47cd529806e901147a56.jpg", + "image_caption": [ + "(a) Challenges from occluded object" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 370, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a884f1b76d3e22b47a9e9ae129f19ed947e2be829c487c81fbde988623259513.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 143, + 500, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f6e3fcb3935ac8383fdd002bbc432fd4255be8def1274eebab5495446190d70.jpg", + "image_caption": [ + "Fig. 4: Challenges from Omni6D. (a) Algorithms trained on Omni6D can overcome challenges in estimating poses for occluded object instances. The left shows an occluded object instance at the edge of the image, while the right image shows an object instance obstructed by other objects. (b) Algorithm trained on Omni6D can accurately estimate poses with only the lower half or bottom appearance of an object. The green and red colors respectively denote the ground truth and predicted 3D bounding boxes. The blue and orange lines on the boxes separately highlight the intersecting lines of the frontal face and the top face of the two 3D bounding boxes, while the darker lines indicate the bottom of the bounding boxes." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 630, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/586f974ef75a54088c7f57fd9a14b7105d4364da27f6377bab913ac916f10a06.jpg", + "image_caption": [ + "(b) Challenges from bottom views" + ], + "image_footnote": [], + "bbox": [ + 630, + 143, + 759, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "classes included to meet the desired total. Fig. 6a presents the specific categories included in c1sn and their respective sizes relative to each other. Omni6Dout is utilized as an additional test set to measure our algorithm's inter-category generalization. This dataset, constructed similarly to Omni6D, encompasses 52 models spanning 17 categories unseen in Omni6D, along with 4762 images. For additional details on datasets, please refer to the appendix.", + "bbox": [ + 212, + 422, + 787, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Details. All experiments are carried out on a server equipped with an Intel(R) Xeon(R) Gold 6248R CPU @ 3.00GHz and an NVIDIA A100-SXM4-80GB GPU. We maintain consistency in parameters and strategies throughout training, ensuring uniformity in our experiment environment. Given the challenges of semantic classification with a large vocabulary, we use ground truth masks to mitigate the impact of low-quality classification on pose estimation results.", + "bbox": [ + 212, + 513, + 787, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Symmetry-Aware Evaluation", + "text_level": 1, + "bbox": [ + 215, + 630, + 500, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Basic Evaluation Metrics. We utilize the average accuracy of Intersection over 3D Union (IoU) [14] in object detection, and $n^{\\circ}m$ cm in pose estimation. We further decompose $n^{\\circ}m$ cm [19,31] to individually evaluate the model's predictive error $n^{\\circ}$ for pose and $m$ cm for translation. For these three types of errors, the thresholds considered are $\\{50\\%, 75\\%\\}$ , $\\{5^{\\circ}, 10^{\\circ}\\}$ and $\\{2cm,5cm\\}$ [3, 30, 42]. Additionally, we set a detection threshold for objects requiring at least a $10\\%$ overlap between predicted and ground-truth bounding boxes.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our Symmetry-Aware Metrics. Due to NOCS's limited categories, traditional algorithms mainly handle basic symmetry cases, such as rotational symmetry around the y-axis. However, Omni6D has a wider range of objects with different rotational invariances across multiple axes. Fig. 6 provides symmetry statistics for Omni6D objects. To alleviate this issue, we propose a symmetry-", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Compute Our Symmetry-Aware Metric $L_{s}$" + ], + "code_body": "1: procedure SYMMETRIC_METRIC(L, R, $n_x$ , $n_y$ , $n_z$ ) \n2: $\\Theta_0 = \\{0^\\circ\\}$ \n3: $\\Theta_2 = \\{0^\\circ, 90^\\circ, 180^\\circ, 270^\\circ\\}$ \n4: $\\Theta_3 = \\{0^\\circ, 180^\\circ\\}$ // Rotations around Sym-1 axis need not be considered. \n5: $c =$ count(1 occurrences in $\\{n_x, n_y, n_z\\}$ ) \n6: if $c \\geq 2$ then // The object is a sphere. \n7: $L_s = L(R^*, R)$ \n8: else if $c == 1$ then // Rotations around Sym-1 axis can be disregarded. \n9: Without loss of generality, assume $n_x == 1$ . \n10: $L_s = \\min_{\\theta_y \\in \\Theta_{ny}, \\theta_z \\in \\Theta_{nz}} L(R_{\\theta_y, \\theta_z}^*, R)$ \n11: else if $c == 0$ then // Simply enumerate all cases. \n12: $L_s = \\min_{\\theta_x \\in \\Theta_{nx}, \\theta_y \\in \\Theta_{ny}, \\theta_z \\in \\Theta_{nz}} L(R_{\\theta_x, \\theta_y, \\theta_z}^*, R)$ \n13: end if \n14: return $L_s$ \n15: end procedure", + "bbox": [ + 218, + 164, + 787, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "aware metric. Unlike prior works focusing solely on the y-axis, our method considers rotation symmetry around all three axes.", + "bbox": [ + 212, + 357, + 782, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We define the relevant variables as follows: $L_{s}$ denotes our symmetry-aware metric, $L$ denotes the original metric. $R$ stands for the ground truth rotation matrix, while $R^{*}$ represents the predicted rotation matrix. $R_{\\theta_x,\\theta_y,\\theta_z}^*$ corresponds to the predicted rotation matrix after sequentially rotating by $\\theta_{x},\\theta_{y}$ , and $\\theta_{z}$ degrees around the xyz axes. The rotational invariance cases around the x, y, and z axes are denoted as Sym- $n_{x}$ , Sym- $n_{y}$ , and Sym- $n_{z}$ , where $n_x,n_y$ , and $n_z$ are the respective rotation parameters. Objects that align with Sym- $n$ around an axis maintain their original shape when rotated by an angle from $\\Theta_{n}$ .", + "bbox": [ + 212, + 387, + 784, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since the Euler angles are compact [13], the most straightforward approach is to determine the category of rotational invariance for each axis $\\{\\mathbf{x},\\mathbf{y},\\mathbf{z}\\}$ sequentially, as mentioned in 3.2. To simplify computations, we set $\\Theta_0 = \\{0^\\circ \\}$ , $\\Theta_{1} = \\{0^{\\circ},1^{\\circ},\\dots,359^{\\circ}\\}$ , $\\Theta_{2} = \\{0^{\\circ},90^{\\circ},180^{\\circ},270^{\\circ}\\}$ , $\\Theta_{3} = \\{0^{\\circ},180^{\\circ}\\}$ . We can define $L_{s}$ as $L_{s} = \\min_{\\theta_{x}\\in \\Theta_{n_{x}},\\theta_{y}\\in \\Theta_{n_{y}},\\theta_{z}\\in \\Theta_{n_{z}}}L(R_{\\theta_{x},\\theta_{y},\\theta_{z}}^{*},R)$ .", + "bbox": [ + 212, + 508, + 784, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "However, due to the singularity of Euler angles [13], we can simplify the above rotation transformation. The pseudo-code implementation of our Symmetry-Aware Evaluation is provided in Algorithm 1. It allows us to simplify what was originally at most $360^{3}$ computations to a maximum of only $4^{3}$ computations.", + "bbox": [ + 212, + 585, + 784, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Large-Vocabulary 6D Pose and Size Estimation", + "text_level": 1, + "bbox": [ + 214, + 665, + 651, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Performance on Omni6D. We present results of algorithms [6,10,34,46,47] trained on Omni6Dtrain and tested on Omni6Dtest. We compare their quantitative results in Tab. 2 and their qualitative results in Fig. S10 in Appendix. The performance disparity among algorithms for category-level 6D object pose estimation becomes markedly pronounced when applied to large-vocabulary datasets, in contrast to the more consistent performance previously observed on the Real and CAMERA datasets [40]. This highlights the inherent strengths and weaknesses across various model structures.", + "bbox": [ + 212, + 688, + 785, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This observation suggests the potential importance of our large-vocabulary dataset in uncovering the relative performance of different models. It appears", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e6f45eb75e29e5ebdaca5cad65b230e7f225bfcab51f337ae31cac986691cce6.jpg", + "table_caption": [ + "Table 2: Category-level performance on Omni6D dataset. Models are trained on Omni6Dtrain and tested on Omni6Dtest. Instances within each category in the test set are unseen during training, substantiating the algorithms' capacity to generalize within individual categories under large-vocabulary settings. Bold and underlined results indicate the best and second-best performers." + ], + "table_footnote": [], + "table_body": "
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit44.5620.377.559.5614.7619.2310.6821.0237.4970.09
SGPA [6]implicit36.3414.444.786.8410.1315.038.4917.7325.5759.18
DualPoseNet [20]hybrid58.8425.498.289.3017.2619.059.3819.1873.8296.37
RBP-Pose [46]hybrid35.924.660.370.600.530.800.750.9639.7383.55
GPV-Pose [10]explicit15.280.260.100.700.140.962.252.965.3133.70
HS-Pose [47]explicit62.6523.024.264.8510.4911.614.9611.7580.9397.78
", + "bbox": [ + 217, + 226, + 784, + 314 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/585bedaa562e7b9201e8a889ed9f401c61069c0c42d843d52194ac43db91ff16.jpg", + "table_caption": [ + "Table 3: Category-level performance on unseen categories. Models are trained on Omni6Dtrain and tested on Omni6Dout. Categories in the test set never appear in the training set, validating the algorithms' ability to generalize across categories." + ], + "table_footnote": [], + "table_body": "
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit7.560.950.180.400.801.650.652.368.8840.59
SGPA [6]implicit7.050.600.070.280.190.820.531.693.8728.28
DualPoseNet [20]hybrid36.8512.063.243.378.048.513.398.6478.0098.60
RBP-Pose [46]hybrid26.181.950.010.020.020.030.020.0316.7443.06
GPV-Pose [10]explicit10.970.140.030.180.120.570.301.077.1441.30
HS-Pose [47]explicit36.758.921.541.664.675.161.755.3879.9598.27
", + "bbox": [ + 217, + 383, + 784, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "that the increased complexity of the dataset could push model architectures to their theoretical limits, possibly revealing intrinsic characteristics otherwise obscured in less complex scenarios. For example, SPD, SGPA is particularly proficient in predicting rotation, and SPD achieves the highest score in $n^{\\circ}mcm$ . This could be due to its implicit network's propensity for generating more reliable rotational forecasts. Meanwhile, DualPoseNet and HS-Pose provide more accurate predictions for translation and score higher in IoU. This could be associated with the characteristic of models with explicit networks to produce better translations and size estimates.", + "bbox": [ + 212, + 503, + 787, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our large-vocabulary dataset, encompassing a broad spectrum of shapes and appearances, enables a comprehensive evaluation of diverse category-level pose estimation methods. This serves not only as a robust test of an algorithm's generalizability but also as a valuable tool in understanding the advantages offered by different algorithmic structures.", + "bbox": [ + 212, + 641, + 787, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Generalization Performance. We evaluate algorithms on $\\mathrm{Omni6D}_{out}$ to assess their inter-category generalization capabilities. The outcomes are presented in Tab. 3. Notably, DualPoseNet and HS-Pose emerged as superior performers, outclassing others across all metrics, thereby demonstrating excellent generalization abilities. Contrastingly, implicit methods including SPD and SPGA exhibited marked limitations. Qualitative results are shown in Fig. S11 in Appendix.", + "bbox": [ + 212, + 717, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Drawing parallels with the observations from Tab. 2, we found that metrics such as translation and IoU were relatively easier to excel in, suggesting superior", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4570cd25a0b82c7cbb1192ae3e87459036b14737d54672c088e5deb7b36383b5.jpg", + "image_caption": [ + "Fig. 5: Category-Wise Performance on Omni6D Dataset. The x-axis, moving from left to right, sequentially represents: the number of objects within a category (Semantic Category), the number of objects within a cluster clustered based on shape priors (Shape Category) and the diversity of instances within a category. The y-axis depicts category or clustered group results for IoU75 and 5°2 cm metrics. Each plotted point illustrates the algorithm's result for a specific category or cluster, while the line showcases the trend of the linear fit for the scattered points." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 769, + 250 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "generalization abilities in translation and size prediction. Conversely, the generalization of rotation emerges as a considerable challenge in category-level 6D object pose estimation, especially within large-vocabulary scenes.", + "bbox": [ + 212, + 419, + 784, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Category-wise Analysis. Based on the $\\mathrm{IoU}_{75}$ and $5^{\\circ}2$ cm metrics, we conducted a detailed category-wise analysis of the results from Tab. 2. Left columns in Fig. 5 illustrate the correlation between category-level 6D pose estimation performance and the number of instances within each category in Omni6Dtrain. Middle columns in Fig. 5 analyze the correlation between cluster-level average performance and cluster size based on the clustering results described in Fig. 3e. We found that the performance of pose estimation for each category is more strongly correlated with the number of instances within clusters than with semantic categories, showing a positive correlation. This suggests that shape categories have a greater impact on training than semantic categories do. Notably, algorithms like SPD, SGPA, and RBP-Pose that utilize shape prior structures are particularly sensitive to this influence.", + "bbox": [ + 212, + 478, + 787, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Right columns in Fig. 5 reveal the correlation of pose estimation performance relative to instance diversity within each category in the training set. We measured instance diversity by calculating the mean chamfer distance [1] among all pairs of instances in each category. The results show that as diversity within a category increases, pose estimation performance tends to improve. This observation aligns with the assertion made by [23]: The key to the success of prior-based methods lies in the deformation modules, which learns to synthesize world-space target objects and explicitly builds the correspondence between camera and world-space. As the number of instances increases and the diversity within a shape category expands, the model's capacity to learn deformation from priors to actual instance shapes is strengthened, leading to improved results.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c85a544023dec18ffe03f680575a37364787a74cbda9b5523875a42459d5637a.jpg", + "image_caption": [ + "(a) Category inventory of clsn" + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 444, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e3ef99cf02c9d254bf358461b49c61e658fc66d4a7ae50b861cceaee9b6b7f9d.jpg", + "image_caption": [ + "(b) Our finetune strategy", + "Fig.6: Our finetune strategy. (a) Category inventory of c1sn within Omni6D dataset. The angle of each sector in the chart reflects the relative size of the instance count within that category. (b) In each fine-tuning step, we double the category count, copying trained global features and old category parameters into the new network while initializing the new category parameters. An observable deepening of color is indicative of the escalating count of training iterations." + ], + "image_footnote": [], + "bbox": [ + 457, + 152, + 746, + 282 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Fine-Tuning from Limited Categories", + "text_level": 1, + "bbox": [ + 214, + 422, + 570, + 439 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We propose a finetuning strategy that helps extend methods from a limited set of categories to large-vocabulary. We take SPD [34], DualPoseNet [20], and HS-Pose [47] as examples which belong to three different network architectures and show good performance on Omni6D $_{test}$ . We respectively take their best models on CAMERA as our pre-trained models.", + "bbox": [ + 212, + 446, + 787, + 521 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Initiating the fine-tuning process, we utilize three categories: bottle, bowl, and cup, which are concurrently present in both Omni6D and CAMERA datasets, aligning with the cls3 category. By facilitating the training on Omni6D-cls3, we enable a transfer of the model from CAMERA to Omni6D. Following the method illustrated in Fig. 6b, we engage in an iterative fine-tuning process on a progressively expanded category dataset until it reaches our desired number. In our experiments, we set this target number to be 48 categories.", + "bbox": [ + 212, + 523, + 787, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In parallel, we conduct training from scratch separately on cls3, cls6, ..., and cls48 as a comparison, employing the same number of training iterations. As shown in Fig. 7, even with an exponential increase in the number of categories, pre-trained models remain pivotal in our fine-tuning strategy. The performance of fine-tuning consistently outperforms that of training from scratch.", + "bbox": [ + 212, + 628, + 787, + 703 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "However, regardless of whether the training approach is finetuning or training from scratch, a decline in performance is observed as the number of categories increases. The decline rates for SPD and DualPoseNet are slower, coupled with an initial augmentation in performance due to increased training data and iterations. In contrast, HS-Pose experiences a more rapid decline, with fine-tuned $5^{\\circ}2$ cm results dropping from initial $62.52\\%$ to $14.42\\%$ . Models that excel in tasks involving a limited number of categories may not necessarily maintain their superiority in large-vocabulary tasks, they might be surpassed by models that are more robust and easier to train.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ba9ac324d8671bccc8c28c4b9fcfa9dcde22967901e28bfec40073491d65830a.jpg", + "image_caption": [ + "Fig. 7: Finetuned results. Each figure's x-axis represents the number of categories in the training and test set, while the y-axis displays the outcomes of $5^{\\circ}2$ cm, $5^{\\circ}$ and $2cm$ metrics. Each row, from top to bottom, sequentially employs three methods: SPD [34], DualPoseNet [20], and HS-Pose [47]. The figures depict the outcomes derived from two training strategies as the number of training categories increases, accompanied by the gradual expansion of corresponding test sets." + ], + "image_footnote": [], + "bbox": [ + 246, + 146, + 754, + 320 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Visual Realism", + "text_level": 1, + "bbox": [ + 215, + 455, + 385, + 469 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Due to the complexity of collecting and annotating real-world data, contemporary datasets like NOCS [40] are composed of a large amount of synthetic data and a small portion of real-world data. While collecting real data is relatively straightforward when the number of categories is limited, gathering well-annotated real-world data for pose estimation tasks involving large vocabulary categories becomes a monumental task.", + "bbox": [ + 212, + 497, + 782, + 587 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our Omni6D dataset, which includes large vocabulary objects, is also derived from rendering. However, the incorporation of real-scanned objects significantly enhances the realism of the rendered images. As depicted in Fig. 8, Omni6D receives a score of $2.69 \\pm 0.39$ , surpassing the results obtained by CAMERA.", + "bbox": [ + 212, + 590, + 784, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Given these significant advantages, our dataset excels not only in large-vocabulary scenarios but also in real-world scenes. As depicted in Tab. 4, We use DualPoseNet [20] to train on the common categories in REAL [40] and Omni6D, namely bottles, bowls, and mugs. We train separately on the two datasets and their mix. The results show that Omni6D models perform well on REAL275, and training on the mixed dataset outperforms using REAL or Omni6D datasets alone. This demonstrates that our dataset enables the direct transfer of models to real-world scenes. Moreover, it seamlessly supplements the existing real-world dataset, enabling joint training of models on our dataset and the real-world data.", + "bbox": [ + 212, + 655, + 787, + 791 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To further validate the sim2real capability of models trained with Omni6D, we constructed a real-world dataset, Omni6D-Real, comprising 30 scenes, 39 categories, 73 instances, and 1k images. We captured RGBD images with Azure", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/38cafe713d1ef714db0a8789a5bbedf739e0033b567d6dbaa80f835060781305.jpg", + "image_caption": [ + "Omni6D $2.69\\pm 0.39$" + ], + "image_footnote": [], + "bbox": [ + 245, + 143, + 388, + 227 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/55df81ccc0ce4ee128d7ec79f7423764222c93e70de8753013ec6fbeee0fe8d8.jpg", + "image_caption": [ + "CAMERA $1.55\\pm 0.08$", + "Fig. 8: Comparison of Visual Realism. We evaluated the visual realism of Omni6D in comparison to other datasets through a survey involving 70 human subjects. We randomly selected 10 images from each dataset and introduced noise by blending in 5 images from COCO [21], which included captured photos, and $\\mathrm{SKETCH}^4$ , which comprised rendered images. Subjects were asked to rate the realism of sampled images on a scale from 1 (least realistic) to 5 (most realistic). We report the mean and standard deviation and include a sampled image from the study." + ], + "image_footnote": [], + "bbox": [ + 388, + 143, + 531, + 227 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7b9cffb67d1dbd3750ea2af15078b90bead68e931b3c33ee7a8e542584b0f8ab.jpg", + "image_caption": [ + "REAL $3.53\\pm 0.28$" + ], + "image_footnote": [], + "bbox": [ + 531, + 143, + 674, + 227 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f37437a1c1f431216714574d7ee1872c7966b598f3e50fea098885e97b7140f2.jpg", + "image_caption": [ + "WILD6D $4.38\\pm 0.23$" + ], + "image_footnote": [], + "bbox": [ + 674, + 143, + 756, + 227 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c8adf5872a9f5b977cc30b55b749aaba40ae431fc9353a762329178fda12eee1.jpg", + "table_caption": [ + "Table 4: Performance on REAL275 with Different Training Sets. It compares how different training sets influence DualPoseNet's performance on REAL275 [40], providing insights into the model's ability to generalize in real-world tasks using Omni6D." + ], + "table_footnote": [], + "table_body": "
Train dataRealismIoU50IoU755°2cm5°5cm10°2cm10°5cm5°2cm
Omni6DReal-Scanned78.7632.696.558.8015.0021.3811.2049.54
REAL [40]Real84.5143.438.7610.4021.2425.3913.0169.46
REAL+Omni6DMixed85.2858.5914.1017.8330.1038.9720.9671.00
", + "bbox": [ + 241, + 435, + 756, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Kinect $\\mathrm{DK}^3$ and preprocessed them using SAM [18] for object masks and ICP [2] for point cloud registration. Details are provided in Appendix Section D.", + "bbox": [ + 212, + 511, + 784, + 544 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 564, + 359, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, this paper introduces Omni6D, a novel 6D pose estimation dataset with large-vocabulary categories and intricate scenes. We evaluate existing category-level 6D object pose estimation methods on this benchmark, analyze its challenges, and propose a fine-tuning strategy for large-vocabulary scenarios. Limitations. Our dataset, though more complex, doesn't fully encompass all real-world challenges. Additionally, our fine-tuning strategy effectively extends methods from a small set to a larger one, but its efficacy may decrease with growing category diversity.", + "bbox": [ + 212, + 594, + 787, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future Work. Our study paves the way for diverse research avenues. An immediate next step is expanding the Omni6D dataset with more object types and scenes for comprehensive coverage. Additionally, annotating videos for scanned objects will validate algorithms' large-vocab pose estimation in real-world scenarios. Designing new training strategies for coping with increasing category diversity presents an intriguing challenge.", + "bbox": [ + 212, + 715, + 787, + 806 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "3 https://learn.microsoft.com/azure/kinect-dk/", + "bbox": [ + 217, + 809, + 535, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "4 https://sketchfab.com/", + "bbox": [ + 220, + 825, + 388, + 839 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported by Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative.", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 292, + 325, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barrow, H.G., Tenenbaum, J.M., Bolles, R.C., Wolf, H.C.: Parametric correspondence and chamfer matching: Two new techniques for image matching. In: IJCAI. pp. 659-663. William Kaufmann (1977)", + "2. Besl, P.J., McKay, N.D.: A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell. 14(2), 239-256 (1992)", + "3. Brachmann, E., Michel, F., Krull, A., Yang, M.Y., Gumhold, S., Rother, C.: Uncertainty-driven 6d pose estimation of objects and scenes from a single RGB image. In: CVPR. pp. 3364-3372 (2016)", + "4. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3d: A large benchmark and model for 3d object detection in the wild. In: CVPR. pp. 13154-13164 (2023)", + "5. Chen, D., Li, J., Wang, Z., Xu, K.: Learning canonical shape space for category-level 6d object pose and size estimation. In: CVPR. pp. 11970-11979. Computer Vision Foundation / IEEE (2020)", + "6. Chen, K., Dou, Q.: SGPA: structure-guided prior adaptation for category-level 6d object pose estimation. In: ICCV. pp. 2753-2762 (2021)", + "7. Chen, W., Jia, X., Chang, H.J., Duan, J., Shen, L., Leonardis, A.: Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In: CVPR. pp. 1581-1590 (2021)", + "8. Chen, X., Dong, Z., Song, J., Geiger, A., Hilliges, O.: Category level object pose estimation via neural analysis-by-synthesis. In: ECCV (26). pp. 139-156 (2020)", + "9. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. J. Open Source Softw. 8(83), 4901 (2023)", + "0. Di, Y., Zhang, R., Lou, Z., Manhardt, F., Ji, X., Navab, N., Tombari, F.: Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In: CVPR. pp. 6771-6781 (2022)", + "1. Du, G., Wang, K., Lian, S., Zhao, K.: Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artif. Intell. Rev. 54(3), 1677-1734 (2021)", + "2. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981)", + "3. Gao, X., Zhang, T.: Introduction to Visual SLAM - From Theory to Practice. Springer (2021)", + "4. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013)" + ], + "bbox": [ + 225, + 324, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Huang, S., Qi, S., Xiao, Y., Zhu, Y., Wu, Y.N., Zhu, S.: Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. In: NeurIPS. pp. 206-217 (2018)", + "16. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: ICRA. pp. 10632-10640. IEEE (2022)", + "17. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape, appearance, and pose optimization. In: ECCV (2). Springer (2022)", + "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "19. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis. 128(3), 657-678 (2020)", + "20. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: ICCV. pp. 3540-3549 (2021)", + "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755 (2014)", + "22. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS. pp. 1323-1330 (2021)", + "23. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. CoRR abs/2303.13479 (2023)", + "24. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: ECCV (2). pp. 499-516 (2022)", + "25. Lunayach, M., Zakharov, S., Chen, D., Ambrus, R., Kira, Z., Irshad, M.Z.: FSD: fast self-supervised single RGB-D to categorical 3d objects. CoRR abs/2310.12974 (2023)", + "26. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph. 22(12), 2633-2651 (2016)", + "27. Murtagh, F., Legendre, P.: Ward's hierarchical agglomerative clustering method: Which algorithms implement ward's criterion? J. Classif. 31(3), 274-295 (2014)", + "28. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: CVPR. pp. 52-61 (2020)", + "29. Peng, W., Yan, J., Wen, H., Sun, Y.: Self-supervised category-level 6d object pose estimation with deep implicit shape representation. In: AAAI. pp. 2082-2090. AAAI Press (2022)", + "30. Rad, M., Lepetit, V.: BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: ICCV. pp. 3848-3856 (2017)", + "31. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.W.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR. pp. 2930-2937 (2013)", + "32. Song, C., Song, J., Huang, Q.: Hybridpose: 6d object pose estimation under hybrid representations. In: CVPR. pp. 428-437 (2020)", + "33. Su, Y., Rambach, J.R., Minaskan, N., Lesur, P., Pagani, A., Stricker, D.: Deep multi-state object pose estimation for augmented reality assembly. In: ISMAR Adjunct. pp. 222-227. IEEE (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Tian, M., Ang, M.H., Lee, G.H.: Shape prior deformation for categorical 6d object pose and size estimation. In: ECCV (21). pp. 530-546 (2020)", + "35. Tremblay, J., To, T., Sundaralingam, B., Xiang, Y., Fox, D., Birchfield, S.: Deep object pose estimation for semantic robotic grasping of household objects. In: CoRL. pp. 306-316 (2018)", + "36. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4), 376-380 (1991)", + "37. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: ICRA. pp. 10059-10066 (2020)", + "38. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. CoRR abs/2203.10339 (2022)", + "39. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: CVPR. pp. 16611-16621 (2021)", + "40. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR. pp. 2642-2651 (2019)", + "41. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR. pp. 803-814 (2023)", + "42. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. In: Robotics: Science and Systems (2018)", + "43. Zakharov, S., Shugurov, I., Ilic, S.: DPOD: dense 6d pose object detector in RGB images. CoRR abs/1902.11020 (2019)", + "44. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semisupervised learning approach and A new dataset. In: NeurIPS (2022)", + "45. Zhang, K., Fu, Y., Borse, S., Cai, H., Porikli, F., Wang, X.: Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In: ICLR. OpenReview.net (2023)", + "46. Zhang, R., Di, Y., Lou, Z., Manhardt, F., Tombari, F., Ji, X.: Rbp-posed: Residual bounding box projection for category-level pose estimation. In: ECCV (1). pp. 655-672 (2022)", + "47. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: CVPR. pp. 17163-17173 (2023)" + ], + "bbox": [ + 212, + 146, + 784, + 659 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Omni6D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_model.json b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_model.json new file mode 100644 index 0000000000000000000000000000000000000000..36c431e106b5f2a4f1a3a7e91ec2b5c66a818269 --- /dev/null +++ b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_model.json @@ -0,0 +1,2401 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.142, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Omni6D: Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.212, + 0.777, + 0.243 + ], + "angle": 0, + "content": "Mengchen Zhang\\(^{1,2}\\), Tong Wu\\(^{3}\\), Tai Wang\\(^{2}\\), Tengfei Wang\\(^{2}\\), Ziwei Liu\\(^{4}\\), and Dahua Lin\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.255, + 0.632, + 0.27 + ], + "angle": 0, + "content": "1 Zhejiang University, Zhejiang, China" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.269, + 0.71, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) Shanghai Artificial Intelligence Laboratory, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.282, + 0.696, + 0.298 + ], + "angle": 0, + "content": "3 The Chinese University of Hong Kong, Hong Kong SAR" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.297, + 0.66, + 0.312 + ], + "angle": 0, + "content": "\\(^{4}\\) Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.312, + 0.69, + 0.325 + ], + "angle": 0, + "content": "{zhangmengchen, wangtai, wangtengfei}@pjlab.org.cn," + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.326, + 0.696, + 0.34 + ], + "angle": 0, + "content": "{wt020,dhlin}@ie.cuhk.edu.hk,ziwei.liu@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.365, + 0.784, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.637 + ], + "angle": 0, + "content": "Fig.1: Omni6D is a dataset for 6D object pose and size estimation with large vocabulary categories and rich annotations. (a) showcases ground truth of RGB image, depth map and NOCS map. (b) presents shape priors derived from a variational autoencoder [5] with adjusted canonical poses. (c) provides examples of the rotational symmetry of objects we have annotated, indicating the multiples of angles by which the shape remains unchanged when rotated around the xyz axes." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.66, + 0.744, + 0.842 + ], + "angle": 0, + "content": "Abstract. 6D object pose estimation aims at determining an object's translation, rotation, and scale, typically from a single RGBD image. Recent advancements have expanded this estimation from instance-level to category-level, allowing models to generalize across unseen instances within the same category. However, this generalization is limited by the narrow range of categories covered by existing datasets, such as NOCS, which also tend to overlook common real-world challenges like occlusion. To tackle these challenges, we introduce Omni6D, a comprehensive RGBD dataset featuring a wide range of categories and varied backgrounds, elevating the task to a more realistic context. 1) The dataset comprises an extensive spectrum of 166 categories, 4688 instances adjusted to the canonical pose, and over 0.8 million captures, significantly broadening the scope for evaluation. 2) We introduce a symmetry-aware" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.248 + ], + "angle": 0, + "content": "metric and conduct systematic benchmarks of existing algorithms on Omni6D, offering a thorough exploration of new challenges and insights. 3) Additionally, we propose an effective fine-tuning approach that adapts models from previous datasets to our extensive vocabulary setting. We believe this initiative will pave the way for new insights and substantial progress in both the industrial and academic fields, pushing forward the boundaries of general 6D pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.259, + 0.744, + 0.288 + ], + "angle": 0, + "content": "Keywords: 6DoF Pose Estimation \\(\\cdot\\) Large Vocabulary Dataset \\(\\cdot\\) Metrics and Benchmarks" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.318, + 0.377, + 0.334 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.414 + ], + "angle": 0, + "content": "6D pose estimation aims at predicting the position, orientation, and size of objects in a 3D space using RGB (D) images, enabling various applications such as augmented/virtual reality [26,33], robot manipulation [11,35], and scene understanding [15,28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Early instance-level pose estimation approaches [32, 38, 39, 42, 43] typically involve providing instance CAD models and predicting poses of instances that were seen during training, restricting the generalization to unseen objects. In contrast, recent research has shifted towards category-level 6D object pose estimation [6-8, 10, 16, 17, 20, 24, 25, 29, 34, 37, 40, 44-47], which learns category prior from a large number of instances within a category, allowing for pose estimation of new instances within the same category without the need for CAD models. By learning on a diverse range of categories, category-level approaches could be a more versatile solution for 6D pose estimation in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.598 + ], + "angle": 0, + "content": "However, most existing datasets [22, 40, 44] are limited to a small number of object categories, typically less than 10, as shown in Tab. 1, hindering their practical applicability to complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To overcome the limitations in previous category-level 6D pose estimation datasets, such as limited category numbers, lack of instance diversity within categories, and overly simplistic scenes, this paper presents a novel category-level dataset dubbed Omni6D for 6D pose estimation. Omni6D significantly extends the number of object categories to 166, and includes 4,688 real-scanned and well-annotated instance objects with a diverse range of shapes, sizes, and textures. The constructed benchmark includes 0.8M images featuring complex scenes with various occlusions, changing lighting conditions, complex backgrounds, and varying viewpoints. For each scene, we provide the rendered image, depth map, NOCS map, and instance mask. Also, considering the widespread rotational symmetry in objects, we examine three types of rotational invariance where an object maintains its original shape under following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we introduce a symmetry-aware metric to specifically address rotational invariance. Every object in Omni6D is adjusted to the canonical pose and annotated with rotational symmetry around three axes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Comparisons between Omni6D and existing datasets. Omni6D significantly extends the range of everyday object categories and instances." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.185, + 0.742, + 0.28 + ], + "angle": 0, + "content": "
DatasetsModeRealism# Categories # Instances# Images
ShapeNet-SRNCars [22]RGBSynthetic13514
Sim2Real Cars [22]RGBReal110
CAMERA [40]RGBDSynthetic61085
REAL [40]RGBDReal642
Wild6D [44]RGBDReal51722
Omni6DRGBDReal-Scanned1664,688
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.306, + 0.786, + 0.501 + ], + "angle": 0, + "content": "Including a broader range of categories, our dataset offers a more comprehensive and challenging evaluation benchmark for category-level 6D object pose estimation. Utilizing Omni6D, we train and analyze existing algorithms, initiating a profound exploration of the challenges and vital elements involved in category-level estimation within large-vocabulary categories. Additionally, we assess these algorithms' capability to generalize across categories, and carry out a category-wise analysis. Experiments show that our dataset presents a more challenging benchmark for 6D pose estimation, highlighting the need for more robust and generalized pose estimation approaches. As an initial attempt, we present a finetuning strategy that assists in broadening the scope of existing approaches from a limited range of categories to a broader vocabulary. Moreover, we conduct an analysis of the domain gap between our dataset and real-world dataset, emphasizing the benefits of their combined use." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.547 + ], + "angle": 0, + "content": "Our dataset will be publicly available to the research community, which will foster future research on more practical and robust 6D pose estimation algorithms and pave the way for broader applications." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.568, + 0.387, + 0.584 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.599, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Existing work on category-level 6D object pose estimation can be generally divided into two types. After extracting features from images or point clouds, they compute Rotation, Translation, and Size (RTS) either through implicit point correspondence or explicit regression." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.66, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Existing Datasets. The most commonly used dataset for category-level 6D object pose estimation is NOCS [40], comprising both the synthetic CAMERA dataset and the real-world REAL dataset. CAMERA includes 300k RGBD images of 31 indoor scenes with 1,085 object instances across 6 categories, while REAL mirrors the categories in CAMERA and includes 8k RGBD images capturing 42 instances in 18 real scenes. Wild6D [44] consists of 5,166 videos with 1.1 million images over 1,722 object instances in 5 categories. ShapeNet-SRN Cars dataset and Sim2Real Cars dataset proposed in iNerf [22] both exclusively include a single car category. The former includes 3,514 instances derived from ShapeNet cars, while the latter is extracted from videos capturing 10 distinct unseen car models. These datasets are limited by their narrow range of categories, hindering their ability to generalize broadly. Additionally, most training images" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "are synthetic and lack realism, and their scenes are overly simplified, failing to account for common real-world challenges like occlusions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.787, + 0.269 + ], + "angle": 0, + "content": "Implicit Methods. Implicit methods are based on point correspondence [6, 20, 24, 34, 37, 40, 44, 46]. NOCS [40], one of the pioneering works in this area, introduced the concept of Normalized Object Coordinate Space (NOCS). The final pose and size of the object are obtained by matching the predicted NOCS map with the observed depth input using the Umeyama algorithm [36] and RANSAC algorithm [12]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Subsequent algorithms such as DualPoseNet, RBP-Net and RePoNet [20,44, 46] have continued to develop along the vein of NOCS, implicitly solving for pose after predicting the NOCS map. SPD [34] proposed a category-level shape prior, subsequently deforming this shape prior (i.e., average shape) to fit observed point cloud. SGPA, RePoNet, and CATRE [6,24,44] continue to develop along SPD's category-level shape prior approach. Algorithms like 6-PACK and SGPA [6,37] extract low-rank structure points, i.e., keypoints, from dense observed point clouds. 6-PACK [37] predicts interframe motion of target instances through keypoint matching, while SGPA [6] employs keypoints for more effective incorporation of sparse structural information during prior adaptation. These methods rely heavily on the RANSAC process to eliminate outliers, making them non-differentiable and time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Explicit Methods. Explicit methods are based on direct pose regression [7, 10, 20, 24, 46, 47]. DualPoseNet and RBP-Net [20, 46] conduct both explicit and implicit training, where one parallel pose decoder explicitly regresses the pose. CATRE [24], recognizes the inherent difference between estimations of rotation and translation/size, explicitly regressing their residuals and carrying out an iterative pose estimation process. FS-Net [7] designs an autoencoder with 3D Graphic Convolution for latent feature extraction and separates the predictions for rotation and translation/size into two distinct networks: one estimates translation/size through two residuals, while the other handles rotation prediction by estimating deflections on two orthogonal axes. GPV-Pose and HS-Pose [10, 47] utilize the same foundational mechanism introduced by FS-Net [7]. GPV-Pose [10] proposes a decoupled confidence-driven rotation representation that facilitates geometrically-aware recovery of correlated rotation matrices and introduces a new geometry-guided point-by-point voting paradigm for robust retrieval of 3D object bounding boxes. Meanwhile, HS-Pose [47] extends 3D-GC to extract mixed-range latent features from point cloud data through a simple network structure known as the HS layer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.416, + 0.756 + ], + "angle": 0, + "content": "3 Omni6D Dataset" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.371, + 0.793 + ], + "angle": 0, + "content": "3.1 Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Dataset Collection. As shown in Tab. 1, Omni6D comprises 4,688 instances across an impressive span of 166 categories. Each instance is a high-resolution" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.236 + ], + "angle": 0, + "content": "textured mesh, obtained using Shining 3D scanner\\(^{1}\\) and Artec Eva 3D scanner\\(^{2}\\), collected from OmniObject3D [41]. We normalize object models to fit within a \\((-1,1)^{3}(m^{3})\\) three-dimensional space, and align objects within each category to a consistent canonical pose. In the latest dataset, Omni6D-xl builds upon and extends Omni6D, comprising 13,161 instances across an impressive span of 339 categories. For more details, please refer to Appendix Section C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.418 + ], + "angle": 0, + "content": "**Rendering.** We employ stratified sampling to split instances within each category, subsequently dividing them into training, validation, and test sets in a 7:2:1 ratio. In the construction of our dataset, we utilize 9 room models from the Replica dataset as backdrops. For each scenery setup, we randomly select a room model to act as the background, along with 6 - 8 object instance models, which are allowed to perform free-fall motion within the room model, resulting in random scattering in a specific section of the room. Each object model is scaled by a random factor ranging from 0.8 to 1.2 as part of our data augmentation strategy. Considering the attention center of the combined instance models as the origin point, the camera randomly selects ten positions within a radius of \\(8 - 9m\\) and an elevation angle range between \\(30 - 90^{\\circ}\\). The camera then performs rendering at these selected positions while facing towards the attention center." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.785, + 0.495 + ], + "angle": 0, + "content": "Setting. We utilize BlenderProc 2.5.0 [9] to implement the aforementioned rendering process. The intrinsic parameters of the camera are set to [577.5, 577.5, 319.5, 239.5], with an image size specified as \\(640 \\times 480\\). Our approach ensures the diversity and breadth of the dataset, making it suitable for rigorous testing and yielding accurate results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.519, + 0.41, + 0.533 + ], + "angle": 0, + "content": "3.2 Data Annotations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.784, + 0.634 + ], + "angle": 0, + "content": "Rich Annotations. Each rendered output includes the ground truth class label, instance mask, NOCS mapping [40], depth map, as well as 6D pose and size. Fig. 1 exhibits a selection of rendered outputs. To reduce the storage size of the dataset, we encode high-precision depth maps into RGB images by multiplying depth by 10,000, rounding to nearest integer, and converting to base 256. The resulting three digits represent RGB channels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.785, + 0.801 + ], + "angle": 0, + "content": "Rotational Invariance. Rotational invariance implies that a symmetric object can retain its original shape after rotation by certain angles. Many common objects have this property. As shown in Fig. 6, we define the coordinate system as a right-handed system with the x-axis pointing outwards and the y-axis oriented upwards. We contemplate three cases of rotational invariance where an object maintains its original shape after following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we denote the case of no rotational invariance around the axis as Sym-0. According to these definitions, all objects in Omni6D are annotated for their rotational symmetry around the xyz-axes. It's worth noting that symmetry attributes may differ among instances within the same category, requiring instance-level rather" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.413, + 0.826 + ], + "angle": 0, + "content": "1 https://www.einscan.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.403, + 0.84 + ], + "angle": 0, + "content": "\\(^{2}\\) https://www.artec3d.cn/" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.413, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.437 + ], + "angle": 0, + "content": "Fig. 2: Symmetry statistics. The figure demonstrates different symmetry cases using object instances and provides a quantitative representation of the occurrence frequency for various combinations of distinct symmetry cases across the xyz-axes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.526 + ], + "angle": 0, + "content": "than category-level annotations. Fig. 6 illustrates all kinds of symmetry cases using object instances and quantifies their occurrence frequency. Fig. 1 selects several examples to provide a more visual explanation of rotational invariance. These considerations are then integrated into our evaluation protocols in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.409, + 0.56 + ], + "angle": 0, + "content": "3.3 Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Spatial Statistics. Omni6D aims to overcome challenges in estimating poses for occluded object instances. Fig. 3a and Fig. 3b shows the spatial distribution of point clouds and objects by projecting their centroids on the XY-plane (top) and XZ-plane (bottom) [4]. Fig. 3c depicts the relative object size distribution, defined as the square root of the object-to-image area ratio. We observe that the spatial distribution of Omni6D is similar to that of CAMERA and REAL, with a greater resemblance to CAMERA despite having a closer depth range. However, a more pronounced discrepancy between the spatial distribution of point clouds and objects is evident in Omni6D compared to CAMERA and REAL. This observation suggests a higher occurrence of occlusion scenes in Omni6D, highlighting the intricate challenges it presents to 6D object pose estimation. Nonetheless, as depicted in Fig. 4a, algorithms trained on Omni6D demonstrate their robustness in tackling these complexities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Angular Deviation. Omni6D enables accurate pose estimation using only the lower half or bottom appearance of objects. Fig. 3d depicts the density of angular deviations from the upward direction, i.e. y-axis. Our dataset displays a more uniform distribution of object angles relative to the upward axis and exhibits greater deviation from the canonical pose angles. Unlike NOCS, which primarily" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.146, + 0.365, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.147, + 0.518, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.221, + 0.371, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.309, + 0.374, + 0.32 + ], + "angle": 0, + "content": "(c) Relative 2D object size" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.221, + 0.522, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.309, + 0.517, + 0.32 + ], + "angle": 0, + "content": "(d) Angular deviation" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.147, + 0.737, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.308, + 0.7, + 0.319 + ], + "angle": 0, + "content": "(e) Clustering results" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Fig. 3: Omni6D analysis. (a) distribution of point cloud centroids, (b) distribution of object centroids on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (c) density of relative 2D object size, (d) density of angular deviation from the upward direction, (e) Omni6D dataset clustering results. The angle of each sector in the chart reflects the relative size of the instance count within that category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.785, + 0.492 + ], + "angle": 0, + "content": "uses upright object placement, Omni6D utilizes physical simulations for freefall object positioning [9]. As a result, it presents more challenging and diverse pose estimation scenes. Training on Omni6D enhances algorithms' robustness to object rotation angles, as evidenced by the image in Fig. 4b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.786, + 0.673 + ], + "angle": 0, + "content": "Shape Priors. We obtain the mean latent embedding and shape prior for each category from the variational autoencoder [5]. Fig. 1 showcases categorical shape priors, each displaying unique characteristics, facilitating an intuitive association between point cloud shapes and corresponding real-world entities. Meanwhile, Fig. 3e explains clustering results based on categorical latent embeddings, where we employ agglomerative clustering [27] to group categories into 20 clusters. It highlights the geometric coherence among semantically identical objects (especially man-made ones) in Omni6D dataset and further confirms that these categorical shape priors can effectively leverage the wealth of shape information from numerous similar objects to elucidate category features. These insights provide a theoretical basis for applications of category-level 6D object pose estimation using our Omni6D dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.486, + 0.713 + ], + "angle": 0, + "content": "4 Evaluation and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.427, + 0.741 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Datasets. Our experimentation utilized two datasets, namely Omni6D and Omni6Dout. Omni6D are partitioned into training, validation, and test sets in a 7:2:1 ratio, denoted as Omni6Dtrain, Omni6Dval and Omni6Dtest respectively. These sets are further subdivided into subsets with increasing category sizes of 3, 6, 12, 24, and 48. We denote the subset containing \\( n \\) categories as c1sn. Each subset includes all classes present in the previous subset with additional" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.144, + 0.371, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.244, + 0.478, + 0.256 + ], + "angle": 0, + "content": "(a) Challenges from occluded object" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.144, + 0.501, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.631, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.144, + 0.761, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.244, + 0.732, + 0.257 + ], + "angle": 0, + "content": "(b) Challenges from bottom views" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 4: Challenges from Omni6D. (a) Algorithms trained on Omni6D can overcome challenges in estimating poses for occluded object instances. The left shows an occluded object instance at the edge of the image, while the right image shows an object instance obstructed by other objects. (b) Algorithm trained on Omni6D can accurately estimate poses with only the lower half or bottom appearance of an object. The green and red colors respectively denote the ground truth and predicted 3D bounding boxes. The blue and orange lines on the boxes separately highlight the intersecting lines of the frontal face and the top face of the two 3D bounding boxes, while the darker lines indicate the bottom of the bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.514 + ], + "angle": 0, + "content": "classes included to meet the desired total. Fig. 6a presents the specific categories included in c1sn and their respective sizes relative to each other. Omni6Dout is utilized as an additional test set to measure our algorithm's inter-category generalization. This dataset, constructed similarly to Omni6D, encompasses 52 models spanning 17 categories unseen in Omni6D, along with 4762 images. For additional details on datasets, please refer to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.606 + ], + "angle": 0, + "content": "Details. All experiments are carried out on a server equipped with an Intel(R) Xeon(R) Gold 6248R CPU @ 3.00GHz and an NVIDIA A100-SXM4-80GB GPU. We maintain consistency in parameters and strategies throughout training, ensuring uniformity in our experiment environment. Given the challenges of semantic classification with a large vocabulary, we use ground truth masks to mitigate the impact of low-quality classification on pose estimation results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.5, + 0.646 + ], + "angle": 0, + "content": "4.2 Symmetry-Aware Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Basic Evaluation Metrics. We utilize the average accuracy of Intersection over 3D Union (IoU) [14] in object detection, and \\( n^{\\circ}m \\) cm in pose estimation. We further decompose \\( n^{\\circ}m \\) cm [19,31] to individually evaluate the model's predictive error \\( n^{\\circ} \\) for pose and \\( m \\) cm for translation. For these three types of errors, the thresholds considered are \\( \\{50\\%, 75\\%\\} \\), \\( \\{5^{\\circ}, 10^{\\circ}\\} \\) and \\( \\{2cm,5cm\\} \\) [3, 30, 42]. Additionally, we set a detection threshold for objects requiring at least a \\( 10\\% \\) overlap between predicted and ground-truth bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our Symmetry-Aware Metrics. Due to NOCS's limited categories, traditional algorithms mainly handle basic symmetry cases, such as rotational symmetry around the y-axis. However, Omni6D has a wider range of objects with different rotational invariances across multiple axes. Fig. 6 provides symmetry statistics for Omni6D objects. To alleviate this issue, we propose a symmetry-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.629, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Compute Our Symmetry-Aware Metric \\(L_{s}\\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.219, + 0.165, + 0.788, + 0.332 + ], + "angle": 0, + "content": "1: procedure SYMMETRIC_METRIC(L, R, \\(n_x\\), \\(n_y\\), \\(n_z\\)) \n2: \\(\\Theta_0 = \\{0^\\circ\\}\\) \n3: \\(\\Theta_2 = \\{0^\\circ, 90^\\circ, 180^\\circ, 270^\\circ\\}\\) \n4: \\(\\Theta_3 = \\{0^\\circ, 180^\\circ\\}\\) // Rotations around Sym-1 axis need not be considered. \n5: \\(c =\\) count(1 occurrences in \\(\\{n_x, n_y, n_z\\}\\)) \n6: if \\(c \\geq 2\\) then // The object is a sphere. \n7: \\(L_s = L(R^*, R)\\) \n8: else if \\(c == 1\\) then // Rotations around Sym-1 axis can be disregarded. \n9: Without loss of generality, assume \\(n_x == 1\\). \n10: \\(L_s = \\min_{\\theta_y \\in \\Theta_{ny}, \\theta_z \\in \\Theta_{nz}} L(R_{\\theta_y, \\theta_z}^*, R)\\) \n11: else if \\(c == 0\\) then // Simply enumerate all cases. \n12: \\(L_s = \\min_{\\theta_x \\in \\Theta_{nx}, \\theta_y \\in \\Theta_{ny}, \\theta_z \\in \\Theta_{nz}} L(R_{\\theta_x, \\theta_y, \\theta_z}^*, R)\\) \n13: end if \n14: return \\(L_s\\) \n15: end procedure" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.784, + 0.387 + ], + "angle": 0, + "content": "aware metric. Unlike prior works focusing solely on the y-axis, our method considers rotation symmetry around all three axes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.508 + ], + "angle": 0, + "content": "We define the relevant variables as follows: \\( L_{s} \\) denotes our symmetry-aware metric, \\( L \\) denotes the original metric. \\( R \\) stands for the ground truth rotation matrix, while \\( R^{*} \\) represents the predicted rotation matrix. \\( R_{\\theta_x,\\theta_y,\\theta_z}^* \\) corresponds to the predicted rotation matrix after sequentially rotating by \\( \\theta_{x},\\theta_{y} \\), and \\( \\theta_{z} \\) degrees around the xyz axes. The rotational invariance cases around the x, y, and z axes are denoted as Sym- \\( n_{x} \\), Sym- \\( n_{y} \\), and Sym- \\( n_{z} \\), where \\( n_x,n_y \\), and \\( n_z \\) are the respective rotation parameters. Objects that align with Sym- \\( n \\) around an axis maintain their original shape when rotated by an angle from \\( \\Theta_{n} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.785, + 0.586 + ], + "angle": 0, + "content": "Since the Euler angles are compact [13], the most straightforward approach is to determine the category of rotational invariance for each axis \\(\\{\\mathbf{x},\\mathbf{y},\\mathbf{z}\\}\\) sequentially, as mentioned in 3.2. To simplify computations, we set \\(\\Theta_0 = \\{0^\\circ \\}\\), \\(\\Theta_{1} = \\{0^{\\circ},1^{\\circ},\\dots,359^{\\circ}\\}\\), \\(\\Theta_{2} = \\{0^{\\circ},90^{\\circ},180^{\\circ},270^{\\circ}\\}\\), \\(\\Theta_{3} = \\{0^{\\circ},180^{\\circ}\\}\\). We can define \\(L_{s}\\) as \\(L_{s} = \\min_{\\theta_{x}\\in \\Theta_{n_{x}},\\theta_{y}\\in \\Theta_{n_{y}},\\theta_{z}\\in \\Theta_{n_{z}}}L(R_{\\theta_{x},\\theta_{y},\\theta_{z}}^{*},R)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.785, + 0.646 + ], + "angle": 0, + "content": "However, due to the singularity of Euler angles [13], we can simplify the above rotation transformation. The pseudo-code implementation of our Symmetry-Aware Evaluation is provided in Algorithm 1. It allows us to simplify what was originally at most \\(360^{3}\\) computations to a maximum of only \\(4^{3}\\) computations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.652, + 0.682 + ], + "angle": 0, + "content": "4.3 Large-Vocabulary 6D Pose and Size Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Performance on Omni6D. We present results of algorithms [6,10,34,46,47] trained on Omni6Dtrain and tested on Omni6Dtest. We compare their quantitative results in Tab. 2 and their qualitative results in Fig. S10 in Appendix. The performance disparity among algorithms for category-level 6D object pose estimation becomes markedly pronounced when applied to large-vocabulary datasets, in contrast to the more consistent performance previously observed on the Real and CAMERA datasets [40]. This highlights the inherent strengths and weaknesses across various model structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "This observation suggests the potential importance of our large-vocabulary dataset in uncovering the relative performance of different models. It appears" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Category-level performance on Omni6D dataset. Models are trained on Omni6Dtrain and tested on Omni6Dtest. Instances within each category in the test set are unseen during training, substantiating the algorithms' capacity to generalize within individual categories under large-vocabulary settings. Bold and underlined results indicate the best and second-best performers." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.227, + 0.785, + 0.315 + ], + "angle": 0, + "content": "
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit44.5620.377.559.5614.7619.2310.6821.0237.4970.09
SGPA [6]implicit36.3414.444.786.8410.1315.038.4917.7325.5759.18
DualPoseNet [20]hybrid58.8425.498.289.3017.2619.059.3819.1873.8296.37
RBP-Pose [46]hybrid35.924.660.370.600.530.800.750.9639.7383.55
GPV-Pose [10]explicit15.280.260.100.700.140.962.252.965.3133.70
HS-Pose [47]explicit62.6523.024.264.8510.4911.614.9611.7580.9397.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Table 3: Category-level performance on unseen categories. Models are trained on Omni6Dtrain and tested on Omni6Dout. Categories in the test set never appear in the training set, validating the algorithms' ability to generalize across categories." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.385, + 0.785, + 0.472 + ], + "angle": 0, + "content": "
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit7.560.950.180.400.801.650.652.368.8840.59
SGPA [6]implicit7.050.600.070.280.190.820.531.693.8728.28
DualPoseNet [20]hybrid36.8512.063.243.378.048.513.398.6478.0098.60
RBP-Pose [46]hybrid26.181.950.010.020.020.030.020.0316.7443.06
GPV-Pose [10]explicit10.970.140.030.180.120.570.301.077.1441.30
HS-Pose [47]explicit36.758.921.541.664.675.161.755.3879.9598.27
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.639 + ], + "angle": 0, + "content": "that the increased complexity of the dataset could push model architectures to their theoretical limits, possibly revealing intrinsic characteristics otherwise obscured in less complex scenarios. For example, SPD, SGPA is particularly proficient in predicting rotation, and SPD achieves the highest score in \\( n^{\\circ}mcm \\). This could be due to its implicit network's propensity for generating more reliable rotational forecasts. Meanwhile, DualPoseNet and HS-Pose provide more accurate predictions for translation and score higher in IoU. This could be associated with the characteristic of models with explicit networks to produce better translations and size estimates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Our large-vocabulary dataset, encompassing a broad spectrum of shapes and appearances, enables a comprehensive evaluation of diverse category-level pose estimation methods. This serves not only as a robust test of an algorithm's generalizability but also as a valuable tool in understanding the advantages offered by different algorithmic structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Generalization Performance. We evaluate algorithms on \\(\\mathrm{Omni6D}_{out}\\) to assess their inter-category generalization capabilities. The outcomes are presented in Tab. 3. Notably, DualPoseNet and HS-Pose emerged as superior performers, outclassing others across all metrics, thereby demonstrating excellent generalization abilities. Contrastingly, implicit methods including SPD and SPGA exhibited marked limitations. Qualitative results are shown in Fig. S11 in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Drawing parallels with the observations from Tab. 2, we found that metrics such as translation and IoU were relatively easier to excel in, suggesting superior" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.77, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.265, + 0.788, + 0.362 + ], + "angle": 0, + "content": "Fig. 5: Category-Wise Performance on Omni6D Dataset. The x-axis, moving from left to right, sequentially represents: the number of objects within a category (Semantic Category), the number of objects within a cluster clustered based on shape priors (Shape Category) and the diversity of instances within a category. The y-axis depicts category or clustered group results for IoU75 and 5°2 cm metrics. Each plotted point illustrates the algorithm's result for a specific category or cluster, while the line showcases the trend of the linear fit for the scattered points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.785, + 0.466 + ], + "angle": 0, + "content": "generalization abilities in translation and size prediction. Conversely, the generalization of rotation emerges as a considerable challenge in category-level 6D object pose estimation, especially within large-vocabulary scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.66 + ], + "angle": 0, + "content": "Category-wise Analysis. Based on the \\(\\mathrm{IoU}_{75}\\) and \\(5^{\\circ}2\\) cm metrics, we conducted a detailed category-wise analysis of the results from Tab. 2. Left columns in Fig. 5 illustrate the correlation between category-level 6D pose estimation performance and the number of instances within each category in Omni6Dtrain. Middle columns in Fig. 5 analyze the correlation between cluster-level average performance and cluster size based on the clustering results described in Fig. 3e. We found that the performance of pose estimation for each category is more strongly correlated with the number of instances within clusters than with semantic categories, showing a positive correlation. This suggests that shape categories have a greater impact on training than semantic categories do. Notably, algorithms like SPD, SGPA, and RBP-Pose that utilize shape prior structures are particularly sensitive to this influence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Right columns in Fig. 5 reveal the correlation of pose estimation performance relative to instance diversity within each category in the training set. We measured instance diversity by calculating the mean chamfer distance [1] among all pairs of instances in each category. The results show that as diversity within a category increases, pose estimation performance tends to improve. This observation aligns with the assertion made by [23]: The key to the success of prior-based methods lies in the deformation modules, which learns to synthesize world-space target objects and explicitly builds the correspondence between camera and world-space. As the number of instances increases and the diversity within a shape category expands, the model's capacity to learn deformation from priors to actual instance shapes is strengthened, leading to improved results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.445, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.287, + 0.425, + 0.3 + ], + "angle": 0, + "content": "(a) Category inventory of clsn" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.153, + 0.747, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.287, + 0.686, + 0.3 + ], + "angle": 0, + "content": "(b) Our finetune strategy" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig.6: Our finetune strategy. (a) Category inventory of c1sn within Omni6D dataset. The angle of each sector in the chart reflects the relative size of the instance count within that category. (b) In each fine-tuning step, we double the category count, copying trained global features and old category parameters into the new network while initializing the new category parameters. An observable deepening of color is indicative of the escalating count of training iterations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.424, + 0.571, + 0.44 + ], + "angle": 0, + "content": "4.4 Fine-Tuning from Limited Categories" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.522 + ], + "angle": 0, + "content": "We propose a finetuning strategy that helps extend methods from a limited set of categories to large-vocabulary. We take SPD [34], DualPoseNet [20], and HS-Pose [47] as examples which belong to three different network architectures and show good performance on Omni6D\\(_{test}\\). We respectively take their best models on CAMERA as our pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Initiating the fine-tuning process, we utilize three categories: bottle, bowl, and cup, which are concurrently present in both Omni6D and CAMERA datasets, aligning with the cls3 category. By facilitating the training on Omni6D-cls3, we enable a transfer of the model from CAMERA to Omni6D. Following the method illustrated in Fig. 6b, we engage in an iterative fine-tuning process on a progressively expanded category dataset until it reaches our desired number. In our experiments, we set this target number to be 48 categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.704 + ], + "angle": 0, + "content": "In parallel, we conduct training from scratch separately on cls3, cls6, ..., and cls48 as a comparison, employing the same number of training iterations. As shown in Fig. 7, even with an exponential increase in the number of categories, pre-trained models remain pivotal in our fine-tuning strategy. The performance of fine-tuning consistently outperforms that of training from scratch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "However, regardless of whether the training approach is finetuning or training from scratch, a decline in performance is observed as the number of categories increases. The decline rates for SPD and DualPoseNet are slower, coupled with an initial augmentation in performance due to increased training data and iterations. In contrast, HS-Pose experiences a more rapid decline, with fine-tuned \\(5^{\\circ}2\\) cm results dropping from initial \\(62.52\\%\\) to \\(14.42\\%\\). Models that excel in tasks involving a limited number of categories may not necessarily maintain their superiority in large-vocabulary tasks, they might be surpassed by models that are more robust and easier to train." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.147, + 0.756, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Fig. 7: Finetuned results. Each figure's x-axis represents the number of categories in the training and test set, while the y-axis displays the outcomes of \\(5^{\\circ}2\\) cm, \\(5^{\\circ}\\) and \\(2cm\\) metrics. Each row, from top to bottom, sequentially employs three methods: SPD [34], DualPoseNet [20], and HS-Pose [47]. The figures depict the outcomes derived from two training strategies as the number of training categories increases, accompanied by the gradual expansion of corresponding test sets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.386, + 0.47 + ], + "angle": 0, + "content": "4.5 Visual Realism" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.784, + 0.588 + ], + "angle": 0, + "content": "Due to the complexity of collecting and annotating real-world data, contemporary datasets like NOCS [40] are composed of a large amount of synthetic data and a small portion of real-world data. While collecting real data is relatively straightforward when the number of categories is limited, gathering well-annotated real-world data for pose estimation tasks involving large vocabulary categories becomes a monumental task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.785, + 0.652 + ], + "angle": 0, + "content": "Our Omni6D dataset, which includes large vocabulary objects, is also derived from rendering. However, the incorporation of real-scanned objects significantly enhances the realism of the rendered images. As depicted in Fig. 8, Omni6D receives a score of \\(2.69 \\pm 0.39\\), surpassing the results obtained by CAMERA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.792 + ], + "angle": 0, + "content": "Given these significant advantages, our dataset excels not only in large-vocabulary scenarios but also in real-world scenes. As depicted in Tab. 4, We use DualPoseNet [20] to train on the common categories in REAL [40] and Omni6D, namely bottles, bowls, and mugs. We train separately on the two datasets and their mix. The results show that Omni6D models perform well on REAL275, and training on the mixed dataset outperforms using REAL or Omni6D datasets alone. This demonstrates that our dataset enables the direct transfer of models to real-world scenes. Moreover, it seamlessly supplements the existing real-world dataset, enabling joint training of models on our dataset and the real-world data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To further validate the sim2real capability of models trained with Omni6D, we constructed a real-world dataset, Omni6D-Real, comprising 30 scenes, 39 categories, 73 instances, and 1k images. We captured RGBD images with Azure" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.144, + 0.39, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.231, + 0.348, + 0.252 + ], + "angle": 0, + "content": "Omni6D \\(2.69\\pm 0.39\\)" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.144, + 0.532, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.23, + 0.491, + 0.252 + ], + "angle": 0, + "content": "CAMERA \\(1.55\\pm 0.08\\)" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.144, + 0.676, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.23, + 0.634, + 0.252 + ], + "angle": 0, + "content": "REAL \\(3.53\\pm 0.28\\)" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.144, + 0.757, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.23, + 0.746, + 0.252 + ], + "angle": 0, + "content": "WILD6D \\(4.38\\pm 0.23\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 8: Comparison of Visual Realism. We evaluated the visual realism of Omni6D in comparison to other datasets through a survey involving 70 human subjects. We randomly selected 10 images from each dataset and introduced noise by blending in 5 images from COCO [21], which included captured photos, and \\(\\mathrm{SKETCH}^4\\), which comprised rendered images. Subjects were asked to rate the realism of sampled images on a scale from 1 (least realistic) to 5 (most realistic). We report the mean and standard deviation and include a sampled image from the study." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Table 4: Performance on REAL275 with Different Training Sets. It compares how different training sets influence DualPoseNet's performance on REAL275 [40], providing insights into the model's ability to generalize in real-world tasks using Omni6D." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.436, + 0.758, + 0.487 + ], + "angle": 0, + "content": "
Train dataRealismIoU50IoU755°2cm5°5cm10°2cm10°5cm5°2cm
Omni6DReal-Scanned78.7632.696.558.8015.0021.3811.2049.54
REAL [40]Real84.5143.438.7610.4021.2425.3913.0169.46
REAL+Omni6DMixed85.2858.5914.1017.8330.1038.9720.9671.00
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.785, + 0.545 + ], + "angle": 0, + "content": "Kinect \\(\\mathrm{DK}^3\\) and preprocessed them using SAM [18] for object masks and ICP [2] for point cloud registration. Details are provided in Appendix Section D." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.565, + 0.36, + 0.58 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.716 + ], + "angle": 0, + "content": "In conclusion, this paper introduces Omni6D, a novel 6D pose estimation dataset with large-vocabulary categories and intricate scenes. We evaluate existing category-level 6D object pose estimation methods on this benchmark, analyze its challenges, and propose a fine-tuning strategy for large-vocabulary scenarios. Limitations. Our dataset, though more complex, doesn't fully encompass all real-world challenges. Additionally, our fine-tuning strategy effectively extends methods from a small set to a larger one, but its efficacy may decrease with growing category diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Future Work. Our study paves the way for diverse research avenues. An immediate next step is expanding the Omni6D dataset with more object types and scenes for comprehensive coverage. Additionally, annotating videos for scanned objects will validate algorithms' large-vocab pose estimation in real-world scenarios. Designing new training strategies for coping with increasing category diversity presents an intriguing challenge." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.536, + 0.826 + ], + "angle": 0, + "content": "3 https://learn.microsoft.com/azure/kinect-dk/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.826, + 0.39, + 0.84 + ], + "angle": 0, + "content": "4 https://sketchfab.com/" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.81, + 0.536, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.27 + ], + "angle": 0, + "content": "This research is supported by Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.326, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.367 + ], + "angle": 0, + "content": "1. Barrow, H.G., Tenenbaum, J.M., Bolles, R.C., Wolf, H.C.: Parametric correspondence and chamfer matching: Two new techniques for image matching. In: IJCAI. pp. 659-663. William Kaufmann (1977)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.369, + 0.785, + 0.394 + ], + "angle": 0, + "content": "2. Besl, P.J., McKay, N.D.: A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell. 14(2), 239-256 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.436 + ], + "angle": 0, + "content": "3. Brachmann, E., Michel, F., Krull, A., Yang, M.Y., Gumhold, S., Rother, C.: Uncertainty-driven 6d pose estimation of objects and scenes from a single RGB image. In: CVPR. pp. 3364-3372 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3d: A large benchmark and model for 3d object detection in the wild. In: CVPR. pp. 13154-13164 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.52 + ], + "angle": 0, + "content": "5. Chen, D., Li, J., Wang, Z., Xu, K.: Learning canonical shape space for category-level 6d object pose and size estimation. In: CVPR. pp. 11970-11979. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.548 + ], + "angle": 0, + "content": "6. Chen, K., Dou, Q.: SGPA: structure-guided prior adaptation for category-level 6d object pose estimation. In: ICCV. pp. 2753-2762 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.59 + ], + "angle": 0, + "content": "7. Chen, W., Jia, X., Chang, H.J., Duan, J., Shen, L., Leonardis, A.: Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In: CVPR. pp. 1581-1590 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.618 + ], + "angle": 0, + "content": "8. Chen, X., Dong, Z., Song, J., Geiger, A., Hilliges, O.: Category level object pose estimation via neural analysis-by-synthesis. In: ECCV (26). pp. 139-156 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.659 + ], + "angle": 0, + "content": "9. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. J. Open Source Softw. 8(83), 4901 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.701 + ], + "angle": 0, + "content": "0. Di, Y., Zhang, R., Lou, Z., Manhardt, F., Ji, X., Navab, N., Tombari, F.: Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In: CVPR. pp. 6771-6781 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "1. Du, G., Wang, K., Lian, S., Zhao, K.: Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artif. Intell. Rev. 54(3), 1677-1734 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.784 + ], + "angle": 0, + "content": "2. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "3. Gao, X., Zhang, T.: Introduction to Visual SLAM - From Theory to Practice. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Huang, S., Qi, S., Xiao, Y., Zhu, Y., Wu, Y.N., Zhu, S.: Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. In: NeurIPS. pp. 206-217 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "16. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: ICRA. pp. 10632-10640. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.272 + ], + "angle": 0, + "content": "17. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape, appearance, and pose optimization. In: ECCV (2). Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.785, + 0.339 + ], + "angle": 0, + "content": "19. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis. 128(3), 657-678 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.339, + 0.785, + 0.38 + ], + "angle": 0, + "content": "20. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: ICCV. pp. 3540-3549 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.435 + ], + "angle": 0, + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.462 + ], + "angle": 0, + "content": "22. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS. pp. 1323-1330 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.488 + ], + "angle": 0, + "content": "23. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. CoRR abs/2303.13479 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.515 + ], + "angle": 0, + "content": "24. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: ECCV (2). pp. 499-516 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.556 + ], + "angle": 0, + "content": "25. Lunayach, M., Zakharov, S., Chen, D., Ambrus, R., Kira, Z., Irshad, M.Z.: FSD: fast self-supervised single RGB-D to categorical 3d objects. CoRR abs/2310.12974 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.583 + ], + "angle": 0, + "content": "26. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph. 22(12), 2633-2651 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "27. Murtagh, F., Legendre, P.: Ward's hierarchical agglomerative clustering method: Which algorithms implement ward's criterion? J. Classif. 31(3), 274-295 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.609, + 0.785, + 0.651 + ], + "angle": 0, + "content": "28. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: CVPR. pp. 52-61 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.65, + 0.785, + 0.691 + ], + "angle": 0, + "content": "29. Peng, W., Yan, J., Wen, H., Sun, Y.: Self-supervised category-level 6d object pose estimation with deep implicit shape representation. In: AAAI. pp. 2082-2090. AAAI Press (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "30. Rad, M., Lepetit, V.: BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: ICCV. pp. 3848-3856 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "31. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.W.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR. pp. 2930-2937 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "32. Song, C., Song, J., Huang, Q.: Hybridpose: 6d object pose estimation under hybrid representations. In: CVPR. pp. 428-437 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Su, Y., Rambach, J.R., Minaskan, N., Lesur, P., Pagani, A., Stricker, D.: Deep multi-state object pose estimation for augmented reality assembly. In: ISMAR Adjunct. pp. 222-227. IEEE (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Omni6D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "34. Tian, M., Ang, M.H., Lee, G.H.: Shape prior deformation for categorical 6d object pose and size estimation. In: ECCV (21). pp. 530-546 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "35. Tremblay, J., To, T., Sundaralingam, B., Xiang, Y., Fox, D., Birchfield, S.: Deep object pose estimation for semantic robotic grasping of household objects. In: CoRL. pp. 306-316 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "36. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4), 376-380 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "37. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: ICRA. pp. 10059-10066 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "38. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. CoRR abs/2203.10339 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "39. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: CVPR. pp. 16611-16621 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.785, + 0.396 + ], + "angle": 0, + "content": "40. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR. pp. 2642-2651 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "41. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR. pp. 803-814 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.44, + 0.785, + 0.479 + ], + "angle": 0, + "content": "42. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. In: Robotics: Science and Systems (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "43. Zakharov, S., Shugurov, I., Ilic, S.: DPOD: dense 6d pose object detector in RGB images. CoRR abs/1902.11020 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "44. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semisupervised learning approach and A new dataset. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "45. Zhang, K., Fu, Y., Borse, S., Cai, H., Porikli, F., Wang, X.: Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In: ICLR. OpenReview.net (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "46. Zhang, R., Di, Y., Lou, Z., Manhardt, F., Tombari, F., Ji, X.: Rbp-posed: Residual bounding box projection for category-level pose estimation. In: ECCV (1). pp. 655-672 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "47. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: CVPR. pp. 17163-17173 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.66 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_origin.pdf b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d7f485796402500d295206eec1a929e8e1426b07 --- /dev/null +++ b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/bc855166-1f26-4e55-94e7-58dd5c2e8746_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8baf66e85553cc3248590819a98d3ac6798bbece195716f89e307e26f42d2bca +size 7709889 diff --git a/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/full.md b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d6cfbe938efa0d54158f6a69c54e34b4b9e7ba3a --- /dev/null +++ b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/full.md @@ -0,0 +1,297 @@ +# Omni6D: Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation + +Mengchen Zhang $^{1,2}$ , Tong Wu $^{3}$ , Tai Wang $^{2}$ , Tengfei Wang $^{2}$ , Ziwei Liu $^{4}$ , and Dahua Lin $^{2,3}$ + +1 Zhejiang University, Zhejiang, China + +$^{2}$ Shanghai Artificial Intelligence Laboratory, Shanghai, China + +3 The Chinese University of Hong Kong, Hong Kong SAR + +$^{4}$ Nanyang Technological University, Singapore + +{zhangmengchen, wangtai, wangtengfei}@pjlab.org.cn, + +{wt020,dhlin}@ie.cuhk.edu.hk,ziwei.liu@ntu.edu.sg + +![](images/55b5a6610150eb955d988781bccb266b91dce88136f8d16961fa3019f16d46c2.jpg) +Fig.1: Omni6D is a dataset for 6D object pose and size estimation with large vocabulary categories and rich annotations. (a) showcases ground truth of RGB image, depth map and NOCS map. (b) presents shape priors derived from a variational autoencoder [5] with adjusted canonical poses. (c) provides examples of the rotational symmetry of objects we have annotated, indicating the multiples of angles by which the shape remains unchanged when rotated around the xyz axes. + +Abstract. 6D object pose estimation aims at determining an object's translation, rotation, and scale, typically from a single RGBD image. Recent advancements have expanded this estimation from instance-level to category-level, allowing models to generalize across unseen instances within the same category. However, this generalization is limited by the narrow range of categories covered by existing datasets, such as NOCS, which also tend to overlook common real-world challenges like occlusion. To tackle these challenges, we introduce Omni6D, a comprehensive RGBD dataset featuring a wide range of categories and varied backgrounds, elevating the task to a more realistic context. 1) The dataset comprises an extensive spectrum of 166 categories, 4688 instances adjusted to the canonical pose, and over 0.8 million captures, significantly broadening the scope for evaluation. 2) We introduce a symmetry-aware + +metric and conduct systematic benchmarks of existing algorithms on Omni6D, offering a thorough exploration of new challenges and insights. 3) Additionally, we propose an effective fine-tuning approach that adapts models from previous datasets to our extensive vocabulary setting. We believe this initiative will pave the way for new insights and substantial progress in both the industrial and academic fields, pushing forward the boundaries of general 6D pose estimation. + +Keywords: 6DoF Pose Estimation $\cdot$ Large Vocabulary Dataset $\cdot$ Metrics and Benchmarks + +# 1 Introduction + +6D pose estimation aims at predicting the position, orientation, and size of objects in a 3D space using RGB (D) images, enabling various applications such as augmented/virtual reality [26,33], robot manipulation [11,35], and scene understanding [15,28]. + +Early instance-level pose estimation approaches [32, 38, 39, 42, 43] typically involve providing instance CAD models and predicting poses of instances that were seen during training, restricting the generalization to unseen objects. In contrast, recent research has shifted towards category-level 6D object pose estimation [6-8, 10, 16, 17, 20, 24, 25, 29, 34, 37, 40, 44-47], which learns category prior from a large number of instances within a category, allowing for pose estimation of new instances within the same category without the need for CAD models. By learning on a diverse range of categories, category-level approaches could be a more versatile solution for 6D pose estimation in real-world scenarios. + +However, most existing datasets [22, 40, 44] are limited to a small number of object categories, typically less than 10, as shown in Tab. 1, hindering their practical applicability to complex scenes. + +To overcome the limitations in previous category-level 6D pose estimation datasets, such as limited category numbers, lack of instance diversity within categories, and overly simplistic scenes, this paper presents a novel category-level dataset dubbed Omni6D for 6D pose estimation. Omni6D significantly extends the number of object categories to 166, and includes 4,688 real-scanned and well-annotated instance objects with a diverse range of shapes, sizes, and textures. The constructed benchmark includes 0.8M images featuring complex scenes with various occlusions, changing lighting conditions, complex backgrounds, and varying viewpoints. For each scene, we provide the rendered image, depth map, NOCS map, and instance mask. Also, considering the widespread rotational symmetry in objects, we examine three types of rotational invariance where an object maintains its original shape under following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we introduce a symmetry-aware metric to specifically address rotational invariance. Every object in Omni6D is adjusted to the canonical pose and annotated with rotational symmetry around three axes. + +Table 1: Comparisons between Omni6D and existing datasets. Omni6D significantly extends the range of everyday object categories and instances. + +
DatasetsModeRealism# Categories # Instances# Images
ShapeNet-SRNCars [22]RGBSynthetic13514
Sim2Real Cars [22]RGBReal110
CAMERA [40]RGBDSynthetic61085
REAL [40]RGBDReal642
Wild6D [44]RGBDReal51722
Omni6DRGBDReal-Scanned1664,688
+ +Including a broader range of categories, our dataset offers a more comprehensive and challenging evaluation benchmark for category-level 6D object pose estimation. Utilizing Omni6D, we train and analyze existing algorithms, initiating a profound exploration of the challenges and vital elements involved in category-level estimation within large-vocabulary categories. Additionally, we assess these algorithms' capability to generalize across categories, and carry out a category-wise analysis. Experiments show that our dataset presents a more challenging benchmark for 6D pose estimation, highlighting the need for more robust and generalized pose estimation approaches. As an initial attempt, we present a finetuning strategy that assists in broadening the scope of existing approaches from a limited range of categories to a broader vocabulary. Moreover, we conduct an analysis of the domain gap between our dataset and real-world dataset, emphasizing the benefits of their combined use. + +Our dataset will be publicly available to the research community, which will foster future research on more practical and robust 6D pose estimation algorithms and pave the way for broader applications. + +# 2 Related Work + +Existing work on category-level 6D object pose estimation can be generally divided into two types. After extracting features from images or point clouds, they compute Rotation, Translation, and Size (RTS) either through implicit point correspondence or explicit regression. + +Existing Datasets. The most commonly used dataset for category-level 6D object pose estimation is NOCS [40], comprising both the synthetic CAMERA dataset and the real-world REAL dataset. CAMERA includes 300k RGBD images of 31 indoor scenes with 1,085 object instances across 6 categories, while REAL mirrors the categories in CAMERA and includes 8k RGBD images capturing 42 instances in 18 real scenes. Wild6D [44] consists of 5,166 videos with 1.1 million images over 1,722 object instances in 5 categories. ShapeNet-SRN Cars dataset and Sim2Real Cars dataset proposed in iNerf [22] both exclusively include a single car category. The former includes 3,514 instances derived from ShapeNet cars, while the latter is extracted from videos capturing 10 distinct unseen car models. These datasets are limited by their narrow range of categories, hindering their ability to generalize broadly. Additionally, most training images + +are synthetic and lack realism, and their scenes are overly simplified, failing to account for common real-world challenges like occlusions. + +Implicit Methods. Implicit methods are based on point correspondence [6, 20, 24, 34, 37, 40, 44, 46]. NOCS [40], one of the pioneering works in this area, introduced the concept of Normalized Object Coordinate Space (NOCS). The final pose and size of the object are obtained by matching the predicted NOCS map with the observed depth input using the Umeyama algorithm [36] and RANSAC algorithm [12]. + +Subsequent algorithms such as DualPoseNet, RBP-Net and RePoNet [20,44, 46] have continued to develop along the vein of NOCS, implicitly solving for pose after predicting the NOCS map. SPD [34] proposed a category-level shape prior, subsequently deforming this shape prior (i.e., average shape) to fit observed point cloud. SGPA, RePoNet, and CATRE [6,24,44] continue to develop along SPD's category-level shape prior approach. Algorithms like 6-PACK and SGPA [6,37] extract low-rank structure points, i.e., keypoints, from dense observed point clouds. 6-PACK [37] predicts interframe motion of target instances through keypoint matching, while SGPA [6] employs keypoints for more effective incorporation of sparse structural information during prior adaptation. These methods rely heavily on the RANSAC process to eliminate outliers, making them non-differentiable and time-consuming. + +Explicit Methods. Explicit methods are based on direct pose regression [7, 10, 20, 24, 46, 47]. DualPoseNet and RBP-Net [20, 46] conduct both explicit and implicit training, where one parallel pose decoder explicitly regresses the pose. CATRE [24], recognizes the inherent difference between estimations of rotation and translation/size, explicitly regressing their residuals and carrying out an iterative pose estimation process. FS-Net [7] designs an autoencoder with 3D Graphic Convolution for latent feature extraction and separates the predictions for rotation and translation/size into two distinct networks: one estimates translation/size through two residuals, while the other handles rotation prediction by estimating deflections on two orthogonal axes. GPV-Pose and HS-Pose [10, 47] utilize the same foundational mechanism introduced by FS-Net [7]. GPV-Pose [10] proposes a decoupled confidence-driven rotation representation that facilitates geometrically-aware recovery of correlated rotation matrices and introduces a new geometry-guided point-by-point voting paradigm for robust retrieval of 3D object bounding boxes. Meanwhile, HS-Pose [47] extends 3D-GC to extract mixed-range latent features from point cloud data through a simple network structure known as the HS layer. + +# 3 Omni6D Dataset + +# 3.1 Construction + +Dataset Collection. As shown in Tab. 1, Omni6D comprises 4,688 instances across an impressive span of 166 categories. Each instance is a high-resolution + +textured mesh, obtained using Shining 3D scanner $^{1}$ and Artec Eva 3D scanner $^{2}$ , collected from OmniObject3D [41]. We normalize object models to fit within a $(-1,1)^{3}(m^{3})$ three-dimensional space, and align objects within each category to a consistent canonical pose. In the latest dataset, Omni6D-xl builds upon and extends Omni6D, comprising 13,161 instances across an impressive span of 339 categories. For more details, please refer to Appendix Section C. + +**Rendering.** We employ stratified sampling to split instances within each category, subsequently dividing them into training, validation, and test sets in a 7:2:1 ratio. In the construction of our dataset, we utilize 9 room models from the Replica dataset as backdrops. For each scenery setup, we randomly select a room model to act as the background, along with 6 - 8 object instance models, which are allowed to perform free-fall motion within the room model, resulting in random scattering in a specific section of the room. Each object model is scaled by a random factor ranging from 0.8 to 1.2 as part of our data augmentation strategy. Considering the attention center of the combined instance models as the origin point, the camera randomly selects ten positions within a radius of $8 - 9m$ and an elevation angle range between $30 - 90^{\circ}$ . The camera then performs rendering at these selected positions while facing towards the attention center. + +Setting. We utilize BlenderProc 2.5.0 [9] to implement the aforementioned rendering process. The intrinsic parameters of the camera are set to [577.5, 577.5, 319.5, 239.5], with an image size specified as $640 \times 480$ . Our approach ensures the diversity and breadth of the dataset, making it suitable for rigorous testing and yielding accurate results. + +# 3.2 Data Annotations + +Rich Annotations. Each rendered output includes the ground truth class label, instance mask, NOCS mapping [40], depth map, as well as 6D pose and size. Fig. 1 exhibits a selection of rendered outputs. To reduce the storage size of the dataset, we encode high-precision depth maps into RGB images by multiplying depth by 10,000, rounding to nearest integer, and converting to base 256. The resulting three digits represent RGB channels. + +Rotational Invariance. Rotational invariance implies that a symmetric object can retain its original shape after rotation by certain angles. Many common objects have this property. As shown in Fig. 6, we define the coordinate system as a right-handed system with the x-axis pointing outwards and the y-axis oriented upwards. We contemplate three cases of rotational invariance where an object maintains its original shape after following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we denote the case of no rotational invariance around the axis as Sym-0. According to these definitions, all objects in Omni6D are annotated for their rotational symmetry around the xyz-axes. It's worth noting that symmetry attributes may differ among instances within the same category, requiring instance-level rather + +![](images/21a06c09d06bd4bceb684939b8b7eea0d18a22f25c84e14e1f783f15922c5b0f.jpg) +Fig. 2: Symmetry statistics. The figure demonstrates different symmetry cases using object instances and provides a quantitative representation of the occurrence frequency for various combinations of distinct symmetry cases across the xyz-axes. + +than category-level annotations. Fig. 6 illustrates all kinds of symmetry cases using object instances and quantifies their occurrence frequency. Fig. 1 selects several examples to provide a more visual explanation of rotational invariance. These considerations are then integrated into our evaluation protocols in Sec. 4.2. + +# 3.3 Dataset Statistics + +Spatial Statistics. Omni6D aims to overcome challenges in estimating poses for occluded object instances. Fig. 3a and Fig. 3b shows the spatial distribution of point clouds and objects by projecting their centroids on the XY-plane (top) and XZ-plane (bottom) [4]. Fig. 3c depicts the relative object size distribution, defined as the square root of the object-to-image area ratio. We observe that the spatial distribution of Omni6D is similar to that of CAMERA and REAL, with a greater resemblance to CAMERA despite having a closer depth range. However, a more pronounced discrepancy between the spatial distribution of point clouds and objects is evident in Omni6D compared to CAMERA and REAL. This observation suggests a higher occurrence of occlusion scenes in Omni6D, highlighting the intricate challenges it presents to 6D object pose estimation. Nonetheless, as depicted in Fig. 4a, algorithms trained on Omni6D demonstrate their robustness in tackling these complexities. + +Angular Deviation. Omni6D enables accurate pose estimation using only the lower half or bottom appearance of objects. Fig. 3d depicts the density of angular deviations from the upward direction, i.e. y-axis. Our dataset displays a more uniform distribution of object angles relative to the upward axis and exhibits greater deviation from the canonical pose angles. Unlike NOCS, which primarily + +![](images/8fee4c399f65e261ce78ad239b6c70672ec07711d2cd00d69f74712914c2c832.jpg) + +![](images/39c29b7165ee116496b5b46d2b01809f3ea22361cecf798569bb9851a1b529a0.jpg) + +![](images/7d37c58de0e61aa6e6d242ce52c589c3410f4c955e4e848f5f7bba27e754f4a4.jpg) +(c) Relative 2D object size + +![](images/3d2a6dc09651eda420e55261b4d686340073f719d92e6cbff552b2ab33ce5d30.jpg) +(d) Angular deviation + +![](images/c18e3c527805bd2f8dd6ab3bc6734b5033be4bee52a798a040a219a804cd5ad0.jpg) +(e) Clustering results +Fig. 3: Omni6D analysis. (a) distribution of point cloud centroids, (b) distribution of object centroids on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (c) density of relative 2D object size, (d) density of angular deviation from the upward direction, (e) Omni6D dataset clustering results. The angle of each sector in the chart reflects the relative size of the instance count within that category. + +uses upright object placement, Omni6D utilizes physical simulations for freefall object positioning [9]. As a result, it presents more challenging and diverse pose estimation scenes. Training on Omni6D enhances algorithms' robustness to object rotation angles, as evidenced by the image in Fig. 4b. + +Shape Priors. We obtain the mean latent embedding and shape prior for each category from the variational autoencoder [5]. Fig. 1 showcases categorical shape priors, each displaying unique characteristics, facilitating an intuitive association between point cloud shapes and corresponding real-world entities. Meanwhile, Fig. 3e explains clustering results based on categorical latent embeddings, where we employ agglomerative clustering [27] to group categories into 20 clusters. It highlights the geometric coherence among semantically identical objects (especially man-made ones) in Omni6D dataset and further confirms that these categorical shape priors can effectively leverage the wealth of shape information from numerous similar objects to elucidate category features. These insights provide a theoretical basis for applications of category-level 6D object pose estimation using our Omni6D dataset. + +# 4 Evaluation and Analysis + +# 4.1 Experimental Setup + +Datasets. Our experimentation utilized two datasets, namely Omni6D and Omni6Dout. Omni6D are partitioned into training, validation, and test sets in a 7:2:1 ratio, denoted as Omni6Dtrain, Omni6Dval and Omni6Dtest respectively. These sets are further subdivided into subsets with increasing category sizes of 3, 6, 12, 24, and 48. We denote the subset containing $n$ categories as c1sn. Each subset includes all classes present in the previous subset with additional + +![](images/71ab0bbfbe8f9e114c0dfa45c2482f0cda0dc4ae1eca47cd529806e901147a56.jpg) +(a) Challenges from occluded object + +![](images/a884f1b76d3e22b47a9e9ae129f19ed947e2be829c487c81fbde988623259513.jpg) + +![](images/8f6e3fcb3935ac8383fdd002bbc432fd4255be8def1274eebab5495446190d70.jpg) +Fig. 4: Challenges from Omni6D. (a) Algorithms trained on Omni6D can overcome challenges in estimating poses for occluded object instances. The left shows an occluded object instance at the edge of the image, while the right image shows an object instance obstructed by other objects. (b) Algorithm trained on Omni6D can accurately estimate poses with only the lower half or bottom appearance of an object. The green and red colors respectively denote the ground truth and predicted 3D bounding boxes. The blue and orange lines on the boxes separately highlight the intersecting lines of the frontal face and the top face of the two 3D bounding boxes, while the darker lines indicate the bottom of the bounding boxes. + +![](images/586f974ef75a54088c7f57fd9a14b7105d4364da27f6377bab913ac916f10a06.jpg) +(b) Challenges from bottom views + +classes included to meet the desired total. Fig. 6a presents the specific categories included in c1sn and their respective sizes relative to each other. Omni6Dout is utilized as an additional test set to measure our algorithm's inter-category generalization. This dataset, constructed similarly to Omni6D, encompasses 52 models spanning 17 categories unseen in Omni6D, along with 4762 images. For additional details on datasets, please refer to the appendix. + +Details. All experiments are carried out on a server equipped with an Intel(R) Xeon(R) Gold 6248R CPU @ 3.00GHz and an NVIDIA A100-SXM4-80GB GPU. We maintain consistency in parameters and strategies throughout training, ensuring uniformity in our experiment environment. Given the challenges of semantic classification with a large vocabulary, we use ground truth masks to mitigate the impact of low-quality classification on pose estimation results. + +# 4.2 Symmetry-Aware Evaluation + +Basic Evaluation Metrics. We utilize the average accuracy of Intersection over 3D Union (IoU) [14] in object detection, and $n^{\circ}m$ cm in pose estimation. We further decompose $n^{\circ}m$ cm [19,31] to individually evaluate the model's predictive error $n^{\circ}$ for pose and $m$ cm for translation. For these three types of errors, the thresholds considered are $\{50\%, 75\%\}$ , $\{5^{\circ}, 10^{\circ}\}$ and $\{2cm,5cm\}$ [3, 30, 42]. Additionally, we set a detection threshold for objects requiring at least a $10\%$ overlap between predicted and ground-truth bounding boxes. + +Our Symmetry-Aware Metrics. Due to NOCS's limited categories, traditional algorithms mainly handle basic symmetry cases, such as rotational symmetry around the y-axis. However, Omni6D has a wider range of objects with different rotational invariances across multiple axes. Fig. 6 provides symmetry statistics for Omni6D objects. To alleviate this issue, we propose a symmetry- + +Algorithm 1 Compute Our Symmetry-Aware Metric $L_{s}$ +1: procedure SYMMETRIC_METRIC(L, R, $n_x$ , $n_y$ , $n_z$ ) +2: $\Theta_0 = \{0^\circ\}$ +3: $\Theta_2 = \{0^\circ, 90^\circ, 180^\circ, 270^\circ\}$ +4: $\Theta_3 = \{0^\circ, 180^\circ\}$ // Rotations around Sym-1 axis need not be considered. +5: $c =$ count(1 occurrences in $\{n_x, n_y, n_z\}$ ) +6: if $c \geq 2$ then // The object is a sphere. +7: $L_s = L(R^*, R)$ +8: else if $c == 1$ then // Rotations around Sym-1 axis can be disregarded. +9: Without loss of generality, assume $n_x == 1$ . +10: $L_s = \min_{\theta_y \in \Theta_{ny}, \theta_z \in \Theta_{nz}} L(R_{\theta_y, \theta_z}^*, R)$ +11: else if $c == 0$ then // Simply enumerate all cases. +12: $L_s = \min_{\theta_x \in \Theta_{nx}, \theta_y \in \Theta_{ny}, \theta_z \in \Theta_{nz}} L(R_{\theta_x, \theta_y, \theta_z}^*, R)$ +13: end if +14: return $L_s$ +15: end procedure + +aware metric. Unlike prior works focusing solely on the y-axis, our method considers rotation symmetry around all three axes. + +We define the relevant variables as follows: $L_{s}$ denotes our symmetry-aware metric, $L$ denotes the original metric. $R$ stands for the ground truth rotation matrix, while $R^{*}$ represents the predicted rotation matrix. $R_{\theta_x,\theta_y,\theta_z}^*$ corresponds to the predicted rotation matrix after sequentially rotating by $\theta_{x},\theta_{y}$ , and $\theta_{z}$ degrees around the xyz axes. The rotational invariance cases around the x, y, and z axes are denoted as Sym- $n_{x}$ , Sym- $n_{y}$ , and Sym- $n_{z}$ , where $n_x,n_y$ , and $n_z$ are the respective rotation parameters. Objects that align with Sym- $n$ around an axis maintain their original shape when rotated by an angle from $\Theta_{n}$ . + +Since the Euler angles are compact [13], the most straightforward approach is to determine the category of rotational invariance for each axis $\{\mathbf{x},\mathbf{y},\mathbf{z}\}$ sequentially, as mentioned in 3.2. To simplify computations, we set $\Theta_0 = \{0^\circ \}$ , $\Theta_{1} = \{0^{\circ},1^{\circ},\dots,359^{\circ}\}$ , $\Theta_{2} = \{0^{\circ},90^{\circ},180^{\circ},270^{\circ}\}$ , $\Theta_{3} = \{0^{\circ},180^{\circ}\}$ . We can define $L_{s}$ as $L_{s} = \min_{\theta_{x}\in \Theta_{n_{x}},\theta_{y}\in \Theta_{n_{y}},\theta_{z}\in \Theta_{n_{z}}}L(R_{\theta_{x},\theta_{y},\theta_{z}}^{*},R)$ . + +However, due to the singularity of Euler angles [13], we can simplify the above rotation transformation. The pseudo-code implementation of our Symmetry-Aware Evaluation is provided in Algorithm 1. It allows us to simplify what was originally at most $360^{3}$ computations to a maximum of only $4^{3}$ computations. + +# 4.3 Large-Vocabulary 6D Pose and Size Estimation + +Performance on Omni6D. We present results of algorithms [6,10,34,46,47] trained on Omni6Dtrain and tested on Omni6Dtest. We compare their quantitative results in Tab. 2 and their qualitative results in Fig. S10 in Appendix. The performance disparity among algorithms for category-level 6D object pose estimation becomes markedly pronounced when applied to large-vocabulary datasets, in contrast to the more consistent performance previously observed on the Real and CAMERA datasets [40]. This highlights the inherent strengths and weaknesses across various model structures. + +This observation suggests the potential importance of our large-vocabulary dataset in uncovering the relative performance of different models. It appears + +Table 2: Category-level performance on Omni6D dataset. Models are trained on Omni6Dtrain and tested on Omni6Dtest. Instances within each category in the test set are unseen during training, substantiating the algorithms' capacity to generalize within individual categories under large-vocabulary settings. Bold and underlined results indicate the best and second-best performers. + +
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit44.5620.377.559.5614.7619.2310.6821.0237.4970.09
SGPA [6]implicit36.3414.444.786.8410.1315.038.4917.7325.5759.18
DualPoseNet [20]hybrid58.8425.498.289.3017.2619.059.3819.1873.8296.37
RBP-Pose [46]hybrid35.924.660.370.600.530.800.750.9639.7383.55
GPV-Pose [10]explicit15.280.260.100.700.140.962.252.965.3133.70
HS-Pose [47]explicit62.6523.024.264.8510.4911.614.9611.7580.9397.78
+ +Table 3: Category-level performance on unseen categories. Models are trained on Omni6Dtrain and tested on Omni6Dout. Categories in the test set never appear in the training set, validating the algorithms' ability to generalize across categories. + +
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit7.560.950.180.400.801.650.652.368.8840.59
SGPA [6]implicit7.050.600.070.280.190.820.531.693.8728.28
DualPoseNet [20]hybrid36.8512.063.243.378.048.513.398.6478.0098.60
RBP-Pose [46]hybrid26.181.950.010.020.020.030.020.0316.7443.06
GPV-Pose [10]explicit10.970.140.030.180.120.570.301.077.1441.30
HS-Pose [47]explicit36.758.921.541.664.675.161.755.3879.9598.27
+ +that the increased complexity of the dataset could push model architectures to their theoretical limits, possibly revealing intrinsic characteristics otherwise obscured in less complex scenarios. For example, SPD, SGPA is particularly proficient in predicting rotation, and SPD achieves the highest score in $n^{\circ}mcm$ . This could be due to its implicit network's propensity for generating more reliable rotational forecasts. Meanwhile, DualPoseNet and HS-Pose provide more accurate predictions for translation and score higher in IoU. This could be associated with the characteristic of models with explicit networks to produce better translations and size estimates. + +Our large-vocabulary dataset, encompassing a broad spectrum of shapes and appearances, enables a comprehensive evaluation of diverse category-level pose estimation methods. This serves not only as a robust test of an algorithm's generalizability but also as a valuable tool in understanding the advantages offered by different algorithmic structures. + +Generalization Performance. We evaluate algorithms on $\mathrm{Omni6D}_{out}$ to assess their inter-category generalization capabilities. The outcomes are presented in Tab. 3. Notably, DualPoseNet and HS-Pose emerged as superior performers, outclassing others across all metrics, thereby demonstrating excellent generalization abilities. Contrastingly, implicit methods including SPD and SPGA exhibited marked limitations. Qualitative results are shown in Fig. S11 in Appendix. + +Drawing parallels with the observations from Tab. 2, we found that metrics such as translation and IoU were relatively easier to excel in, suggesting superior + +![](images/4570cd25a0b82c7cbb1192ae3e87459036b14737d54672c088e5deb7b36383b5.jpg) +Fig. 5: Category-Wise Performance on Omni6D Dataset. The x-axis, moving from left to right, sequentially represents: the number of objects within a category (Semantic Category), the number of objects within a cluster clustered based on shape priors (Shape Category) and the diversity of instances within a category. The y-axis depicts category or clustered group results for IoU75 and 5°2 cm metrics. Each plotted point illustrates the algorithm's result for a specific category or cluster, while the line showcases the trend of the linear fit for the scattered points. + +generalization abilities in translation and size prediction. Conversely, the generalization of rotation emerges as a considerable challenge in category-level 6D object pose estimation, especially within large-vocabulary scenes. + +Category-wise Analysis. Based on the $\mathrm{IoU}_{75}$ and $5^{\circ}2$ cm metrics, we conducted a detailed category-wise analysis of the results from Tab. 2. Left columns in Fig. 5 illustrate the correlation between category-level 6D pose estimation performance and the number of instances within each category in Omni6Dtrain. Middle columns in Fig. 5 analyze the correlation between cluster-level average performance and cluster size based on the clustering results described in Fig. 3e. We found that the performance of pose estimation for each category is more strongly correlated with the number of instances within clusters than with semantic categories, showing a positive correlation. This suggests that shape categories have a greater impact on training than semantic categories do. Notably, algorithms like SPD, SGPA, and RBP-Pose that utilize shape prior structures are particularly sensitive to this influence. + +Right columns in Fig. 5 reveal the correlation of pose estimation performance relative to instance diversity within each category in the training set. We measured instance diversity by calculating the mean chamfer distance [1] among all pairs of instances in each category. The results show that as diversity within a category increases, pose estimation performance tends to improve. This observation aligns with the assertion made by [23]: The key to the success of prior-based methods lies in the deformation modules, which learns to synthesize world-space target objects and explicitly builds the correspondence between camera and world-space. As the number of instances increases and the diversity within a shape category expands, the model's capacity to learn deformation from priors to actual instance shapes is strengthened, leading to improved results. + +![](images/c85a544023dec18ffe03f680575a37364787a74cbda9b5523875a42459d5637a.jpg) +(a) Category inventory of clsn + +![](images/e3ef99cf02c9d254bf358461b49c61e658fc66d4a7ae50b861cceaee9b6b7f9d.jpg) +(b) Our finetune strategy +Fig.6: Our finetune strategy. (a) Category inventory of c1sn within Omni6D dataset. The angle of each sector in the chart reflects the relative size of the instance count within that category. (b) In each fine-tuning step, we double the category count, copying trained global features and old category parameters into the new network while initializing the new category parameters. An observable deepening of color is indicative of the escalating count of training iterations. + +# 4.4 Fine-Tuning from Limited Categories + +We propose a finetuning strategy that helps extend methods from a limited set of categories to large-vocabulary. We take SPD [34], DualPoseNet [20], and HS-Pose [47] as examples which belong to three different network architectures and show good performance on Omni6D $_{test}$ . We respectively take their best models on CAMERA as our pre-trained models. + +Initiating the fine-tuning process, we utilize three categories: bottle, bowl, and cup, which are concurrently present in both Omni6D and CAMERA datasets, aligning with the cls3 category. By facilitating the training on Omni6D-cls3, we enable a transfer of the model from CAMERA to Omni6D. Following the method illustrated in Fig. 6b, we engage in an iterative fine-tuning process on a progressively expanded category dataset until it reaches our desired number. In our experiments, we set this target number to be 48 categories. + +In parallel, we conduct training from scratch separately on cls3, cls6, ..., and cls48 as a comparison, employing the same number of training iterations. As shown in Fig. 7, even with an exponential increase in the number of categories, pre-trained models remain pivotal in our fine-tuning strategy. The performance of fine-tuning consistently outperforms that of training from scratch. + +However, regardless of whether the training approach is finetuning or training from scratch, a decline in performance is observed as the number of categories increases. The decline rates for SPD and DualPoseNet are slower, coupled with an initial augmentation in performance due to increased training data and iterations. In contrast, HS-Pose experiences a more rapid decline, with fine-tuned $5^{\circ}2$ cm results dropping from initial $62.52\%$ to $14.42\%$ . Models that excel in tasks involving a limited number of categories may not necessarily maintain their superiority in large-vocabulary tasks, they might be surpassed by models that are more robust and easier to train. + +![](images/ba9ac324d8671bccc8c28c4b9fcfa9dcde22967901e28bfec40073491d65830a.jpg) +Fig. 7: Finetuned results. Each figure's x-axis represents the number of categories in the training and test set, while the y-axis displays the outcomes of $5^{\circ}2$ cm, $5^{\circ}$ and $2cm$ metrics. Each row, from top to bottom, sequentially employs three methods: SPD [34], DualPoseNet [20], and HS-Pose [47]. The figures depict the outcomes derived from two training strategies as the number of training categories increases, accompanied by the gradual expansion of corresponding test sets. + +# 4.5 Visual Realism + +Due to the complexity of collecting and annotating real-world data, contemporary datasets like NOCS [40] are composed of a large amount of synthetic data and a small portion of real-world data. While collecting real data is relatively straightforward when the number of categories is limited, gathering well-annotated real-world data for pose estimation tasks involving large vocabulary categories becomes a monumental task. + +Our Omni6D dataset, which includes large vocabulary objects, is also derived from rendering. However, the incorporation of real-scanned objects significantly enhances the realism of the rendered images. As depicted in Fig. 8, Omni6D receives a score of $2.69 \pm 0.39$ , surpassing the results obtained by CAMERA. + +Given these significant advantages, our dataset excels not only in large-vocabulary scenarios but also in real-world scenes. As depicted in Tab. 4, We use DualPoseNet [20] to train on the common categories in REAL [40] and Omni6D, namely bottles, bowls, and mugs. We train separately on the two datasets and their mix. The results show that Omni6D models perform well on REAL275, and training on the mixed dataset outperforms using REAL or Omni6D datasets alone. This demonstrates that our dataset enables the direct transfer of models to real-world scenes. Moreover, it seamlessly supplements the existing real-world dataset, enabling joint training of models on our dataset and the real-world data. + +To further validate the sim2real capability of models trained with Omni6D, we constructed a real-world dataset, Omni6D-Real, comprising 30 scenes, 39 categories, 73 instances, and 1k images. We captured RGBD images with Azure + +![](images/38cafe713d1ef714db0a8789a5bbedf739e0033b567d6dbaa80f835060781305.jpg) +Omni6D $2.69\pm 0.39$ + +![](images/55df81ccc0ce4ee128d7ec79f7423764222c93e70de8753013ec6fbeee0fe8d8.jpg) +CAMERA $1.55\pm 0.08$ +Fig. 8: Comparison of Visual Realism. We evaluated the visual realism of Omni6D in comparison to other datasets through a survey involving 70 human subjects. We randomly selected 10 images from each dataset and introduced noise by blending in 5 images from COCO [21], which included captured photos, and $\mathrm{SKETCH}^4$ , which comprised rendered images. Subjects were asked to rate the realism of sampled images on a scale from 1 (least realistic) to 5 (most realistic). We report the mean and standard deviation and include a sampled image from the study. + +![](images/7b9cffb67d1dbd3750ea2af15078b90bead68e931b3c33ee7a8e542584b0f8ab.jpg) +REAL $3.53\pm 0.28$ + +![](images/f37437a1c1f431216714574d7ee1872c7966b598f3e50fea098885e97b7140f2.jpg) +WILD6D $4.38\pm 0.23$ + +Table 4: Performance on REAL275 with Different Training Sets. It compares how different training sets influence DualPoseNet's performance on REAL275 [40], providing insights into the model's ability to generalize in real-world tasks using Omni6D. + +
Train dataRealismIoU50IoU755°2cm5°5cm10°2cm10°5cm5°2cm
Omni6DReal-Scanned78.7632.696.558.8015.0021.3811.2049.54
REAL [40]Real84.5143.438.7610.4021.2425.3913.0169.46
REAL+Omni6DMixed85.2858.5914.1017.8330.1038.9720.9671.00
+ +Kinect $\mathrm{DK}^3$ and preprocessed them using SAM [18] for object masks and ICP [2] for point cloud registration. Details are provided in Appendix Section D. + +# 5 Conclusion + +In conclusion, this paper introduces Omni6D, a novel 6D pose estimation dataset with large-vocabulary categories and intricate scenes. We evaluate existing category-level 6D object pose estimation methods on this benchmark, analyze its challenges, and propose a fine-tuning strategy for large-vocabulary scenarios. Limitations. Our dataset, though more complex, doesn't fully encompass all real-world challenges. Additionally, our fine-tuning strategy effectively extends methods from a small set to a larger one, but its efficacy may decrease with growing category diversity. + +Future Work. Our study paves the way for diverse research avenues. An immediate next step is expanding the Omni6D dataset with more object types and scenes for comprehensive coverage. Additionally, annotating videos for scanned objects will validate algorithms' large-vocab pose estimation in real-world scenarios. Designing new training strategies for coping with increasing category diversity presents an intriguing challenge. + +# Acknowledgements + +This research is supported by Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative. + +# References + +1. Barrow, H.G., Tenenbaum, J.M., Bolles, R.C., Wolf, H.C.: Parametric correspondence and chamfer matching: Two new techniques for image matching. In: IJCAI. pp. 659-663. William Kaufmann (1977) +2. Besl, P.J., McKay, N.D.: A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell. 14(2), 239-256 (1992) +3. Brachmann, E., Michel, F., Krull, A., Yang, M.Y., Gumhold, S., Rother, C.: Uncertainty-driven 6d pose estimation of objects and scenes from a single RGB image. In: CVPR. pp. 3364-3372 (2016) +4. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3d: A large benchmark and model for 3d object detection in the wild. In: CVPR. pp. 13154-13164 (2023) +5. Chen, D., Li, J., Wang, Z., Xu, K.: Learning canonical shape space for category-level 6d object pose and size estimation. In: CVPR. pp. 11970-11979. Computer Vision Foundation / IEEE (2020) +6. Chen, K., Dou, Q.: SGPA: structure-guided prior adaptation for category-level 6d object pose estimation. In: ICCV. pp. 2753-2762 (2021) +7. Chen, W., Jia, X., Chang, H.J., Duan, J., Shen, L., Leonardis, A.: Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In: CVPR. pp. 1581-1590 (2021) +8. Chen, X., Dong, Z., Song, J., Geiger, A., Hilliges, O.: Category level object pose estimation via neural analysis-by-synthesis. In: ECCV (26). pp. 139-156 (2020) +9. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. J. Open Source Softw. 8(83), 4901 (2023) +0. Di, Y., Zhang, R., Lou, Z., Manhardt, F., Ji, X., Navab, N., Tombari, F.: Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In: CVPR. pp. 6771-6781 (2022) +1. Du, G., Wang, K., Lian, S., Zhao, K.: Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artif. Intell. Rev. 54(3), 1677-1734 (2021) +2. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981) +3. Gao, X., Zhang, T.: Introduction to Visual SLAM - From Theory to Practice. Springer (2021) +4. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013) + +15. Huang, S., Qi, S., Xiao, Y., Zhu, Y., Wu, Y.N., Zhu, S.: Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. In: NeurIPS. pp. 206-217 (2018) +16. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: ICRA. pp. 10632-10640. IEEE (2022) +17. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape, appearance, and pose optimization. In: ECCV (2). Springer (2022) +18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +19. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis. 128(3), 657-678 (2020) +20. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: ICCV. pp. 3540-3549 (2021) +21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755 (2014) +22. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS. pp. 1323-1330 (2021) +23. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. CoRR abs/2303.13479 (2023) +24. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: ECCV (2). pp. 499-516 (2022) +25. Lunayach, M., Zakharov, S., Chen, D., Ambrus, R., Kira, Z., Irshad, M.Z.: FSD: fast self-supervised single RGB-D to categorical 3d objects. CoRR abs/2310.12974 (2023) +26. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph. 22(12), 2633-2651 (2016) +27. Murtagh, F., Legendre, P.: Ward's hierarchical agglomerative clustering method: Which algorithms implement ward's criterion? J. Classif. 31(3), 274-295 (2014) +28. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: CVPR. pp. 52-61 (2020) +29. Peng, W., Yan, J., Wen, H., Sun, Y.: Self-supervised category-level 6d object pose estimation with deep implicit shape representation. In: AAAI. pp. 2082-2090. AAAI Press (2022) +30. Rad, M., Lepetit, V.: BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: ICCV. pp. 3848-3856 (2017) +31. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.W.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR. pp. 2930-2937 (2013) +32. Song, C., Song, J., Huang, Q.: Hybridpose: 6d object pose estimation under hybrid representations. In: CVPR. pp. 428-437 (2020) +33. Su, Y., Rambach, J.R., Minaskan, N., Lesur, P., Pagani, A., Stricker, D.: Deep multi-state object pose estimation for augmented reality assembly. In: ISMAR Adjunct. pp. 222-227. IEEE (2019) + +34. Tian, M., Ang, M.H., Lee, G.H.: Shape prior deformation for categorical 6d object pose and size estimation. In: ECCV (21). pp. 530-546 (2020) +35. Tremblay, J., To, T., Sundaralingam, B., Xiang, Y., Fox, D., Birchfield, S.: Deep object pose estimation for semantic robotic grasping of household objects. In: CoRL. pp. 306-316 (2018) +36. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4), 376-380 (1991) +37. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: ICRA. pp. 10059-10066 (2020) +38. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. CoRR abs/2203.10339 (2022) +39. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: CVPR. pp. 16611-16621 (2021) +40. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR. pp. 2642-2651 (2019) +41. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR. pp. 803-814 (2023) +42. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. In: Robotics: Science and Systems (2018) +43. Zakharov, S., Shugurov, I., Ilic, S.: DPOD: dense 6d pose object detector in RGB images. CoRR abs/1902.11020 (2019) +44. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semisupervised learning approach and A new dataset. In: NeurIPS (2022) +45. Zhang, K., Fu, Y., Borse, S., Cai, H., Porikli, F., Wang, X.: Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In: ICLR. OpenReview.net (2023) +46. Zhang, R., Di, Y., Lou, Z., Manhardt, F., Tombari, F., Ji, X.: Rbp-posed: Residual bounding box projection for category-level pose estimation. In: ECCV (1). pp. 655-672 (2022) +47. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: CVPR. pp. 17163-17173 (2023) \ No newline at end of file diff --git a/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/images.zip b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..052b5f20e444d70667a1aa17156daef7bcdfa01d --- /dev/null +++ b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22f68cf868169ba0b597bdcb213c048dd4657985c71c8d48ed0796e644aae691 +size 462795 diff --git a/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/layout.json b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c2590f7fbba8eb63b60d9068ad870f372a39e6e5 --- /dev/null +++ b/2024/Omni6D_ Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation/layout.json @@ -0,0 +1,9047 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 112, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 112, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 112, + 473, + 148 + ], + "type": "text", + "content": "Omni6D: Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "spans": [ + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "text", + "content": "Mengchen Zhang" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "text", + "content": ", Tong Wu" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "text", + "content": ", Tai Wang" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "text", + "content": ", Tengfei Wang" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "text", + "content": ", Ziwei Liu" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "text", + "content": ", and Dahua Lin" + }, + { + "bbox": [ + 139, + 167, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 226, + 201, + 386, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 201, + 386, + 213 + ], + "spans": [ + { + "bbox": [ + 226, + 201, + 386, + 213 + ], + "type": "text", + "content": "1 Zhejiang University, Zhejiang, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 213, + 434, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 213, + 434, + 224 + ], + "spans": [ + { + "bbox": [ + 179, + 213, + 434, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 213, + 434, + 224 + ], + "type": "text", + "content": " Shanghai Artificial Intelligence Laboratory, Shanghai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 223, + 425, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 223, + 425, + 236 + ], + "spans": [ + { + "bbox": [ + 187, + 223, + 425, + 236 + ], + "type": "text", + "content": "3 The Chinese University of Hong Kong, Hong Kong SAR" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 235, + 403, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 235, + 403, + 247 + ], + "spans": [ + { + "bbox": [ + 209, + 235, + 403, + 247 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 209, + 235, + 403, + 247 + ], + "type": "text", + "content": " Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 189, + 247, + 422, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 247, + 422, + 257 + ], + "spans": [ + { + "bbox": [ + 189, + 247, + 422, + 257 + ], + "type": "text", + "content": "{zhangmengchen, wangtai, wangtengfei}@pjlab.org.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 258, + 425, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 258, + 425, + 269 + ], + "spans": [ + { + "bbox": [ + 188, + 258, + 425, + 269 + ], + "type": "text", + "content": "{wt020,dhlin}@ie.cuhk.edu.hk,ziwei.liu@ntu.edu.sg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 137, + 289, + 479, + 426 + ], + "blocks": [ + { + "bbox": [ + 137, + 289, + 479, + 426 + ], + "lines": [ + { + "bbox": [ + 137, + 289, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 479, + 426 + ], + "type": "image", + "image_path": "55b5a6610150eb955d988781bccb266b91dce88136f8d16961fa3019f16d46c2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 437, + 482, + 504 + ], + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 504 + ], + "type": "text", + "content": "Fig.1: Omni6D is a dataset for 6D object pose and size estimation with large vocabulary categories and rich annotations. (a) showcases ground truth of RGB image, depth map and NOCS map. (b) presents shape priors derived from a variational autoencoder [5] with adjusted canonical poses. (c) provides examples of the rotational symmetry of objects we have annotated, indicating the multiples of angles by which the shape remains unchanged when rotated around the xyz axes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 522, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 522, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 522, + 455, + 666 + ], + "type": "text", + "content": "Abstract. 6D object pose estimation aims at determining an object's translation, rotation, and scale, typically from a single RGBD image. Recent advancements have expanded this estimation from instance-level to category-level, allowing models to generalize across unseen instances within the same category. However, this generalization is limited by the narrow range of categories covered by existing datasets, such as NOCS, which also tend to overlook common real-world challenges like occlusion. To tackle these challenges, we introduce Omni6D, a comprehensive RGBD dataset featuring a wide range of categories and varied backgrounds, elevating the task to a more realistic context. 1) The dataset comprises an extensive spectrum of 166 categories, 4688 instances adjusted to the canonical pose, and over 0.8 million captures, significantly broadening the scope for evaluation. 2) We introduce a symmetry-aware" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 196 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 196 + ], + "type": "text", + "content": "metric and conduct systematic benchmarks of existing algorithms on Omni6D, offering a thorough exploration of new challenges and insights. 3) Additionally, we propose an effective fine-tuning approach that adapts models from previous datasets to our extensive vocabulary setting. We believe this initiative will pave the way for new insights and substantial progress in both the industrial and academic fields, pushing forward the boundaries of general 6D pose estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 205, + 455, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 205, + 455, + 228 + ], + "spans": [ + { + "bbox": [ + 160, + 205, + 455, + 228 + ], + "type": "text", + "content": "Keywords: 6DoF Pose Estimation " + }, + { + "bbox": [ + 160, + 205, + 455, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 205, + 455, + 228 + ], + "type": "text", + "content": " Large Vocabulary Dataset " + }, + { + "bbox": [ + 160, + 205, + 455, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 205, + 455, + 228 + ], + "type": "text", + "content": " Metrics and Benchmarks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 251, + 230, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 230, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 230, + 264 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 280, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 327 + ], + "type": "text", + "content": "6D pose estimation aims at predicting the position, orientation, and size of objects in a 3D space using RGB (D) images, enabling various applications such as augmented/virtual reality [26,33], robot manipulation [11,35], and scene understanding [15,28]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "type": "text", + "content": "Early instance-level pose estimation approaches [32, 38, 39, 42, 43] typically involve providing instance CAD models and predicting poses of instances that were seen during training, restricting the generalization to unseen objects. In contrast, recent research has shifted towards category-level 6D object pose estimation [6-8, 10, 16, 17, 20, 24, 25, 29, 34, 37, 40, 44-47], which learns category prior from a large number of instances within a category, allowing for pose estimation of new instances within the same category without the need for CAD models. By learning on a diverse range of categories, category-level approaches could be a more versatile solution for 6D pose estimation in real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 473 + ], + "type": "text", + "content": "However, most existing datasets [22, 40, 44] are limited to a small number of object categories, typically less than 10, as shown in Tab. 1, hindering their practical applicability to complex scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 475, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 666 + ], + "type": "text", + "content": "To overcome the limitations in previous category-level 6D pose estimation datasets, such as limited category numbers, lack of instance diversity within categories, and overly simplistic scenes, this paper presents a novel category-level dataset dubbed Omni6D for 6D pose estimation. Omni6D significantly extends the number of object categories to 166, and includes 4,688 real-scanned and well-annotated instance objects with a diverse range of shapes, sizes, and textures. The constructed benchmark includes 0.8M images featuring complex scenes with various occlusions, changing lighting conditions, complex backgrounds, and varying viewpoints. For each scene, we provide the rendered image, depth map, NOCS map, and instance mask. Also, considering the widespread rotational symmetry in objects, we examine three types of rotational invariance where an object maintains its original shape under following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we introduce a symmetry-aware metric to specifically address rotational invariance. Every object in Omni6D is adjusted to the canonical pose and annotated with rotational symmetry around three axes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 146, + 454, + 221 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Comparisons between Omni6D and existing datasets. Omni6D significantly extends the range of everyday object categories and instances." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 146, + 454, + 221 + ], + "lines": [ + { + "bbox": [ + 159, + 146, + 454, + 221 + ], + "spans": [ + { + "bbox": [ + 159, + 146, + 454, + 221 + ], + "type": "table", + "html": "
DatasetsModeRealism# Categories # Instances# Images
ShapeNet-SRNCars [22]RGBSynthetic13514
Sim2Real Cars [22]RGBReal110
CAMERA [40]RGBDSynthetic61085
REAL [40]RGBDReal642
Wild6D [44]RGBDReal51722
Omni6DRGBDReal-Scanned1664,688
", + "image_path": "c956b8a1bf6ec64f44009d2c61c9b4f21b46619d7aea3acb8888e6b17a4644e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 242, + 481, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 242, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 242, + 481, + 396 + ], + "type": "text", + "content": "Including a broader range of categories, our dataset offers a more comprehensive and challenging evaluation benchmark for category-level 6D object pose estimation. Utilizing Omni6D, we train and analyze existing algorithms, initiating a profound exploration of the challenges and vital elements involved in category-level estimation within large-vocabulary categories. Additionally, we assess these algorithms' capability to generalize across categories, and carry out a category-wise analysis. Experiments show that our dataset presents a more challenging benchmark for 6D pose estimation, highlighting the need for more robust and generalized pose estimation approaches. As an initial attempt, we present a finetuning strategy that assists in broadening the scope of existing approaches from a limited range of categories to a broader vocabulary. Moreover, we conduct an analysis of the domain gap between our dataset and real-world dataset, emphasizing the benefits of their combined use." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 397, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 433 + ], + "type": "text", + "content": "Our dataset will be publicly available to the research community, which will foster future research on more practical and robust 6D pose estimation algorithms and pave the way for broader applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 449, + 236, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 236, + 462 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 236, + 462 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 474, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 480, + 521 + ], + "type": "text", + "content": "Existing work on category-level 6D object pose estimation can be generally divided into two types. After extracting features from images or point clouds, they compute Rotation, Translation, and Size (RTS) either through implicit point correspondence or explicit regression." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "type": "text", + "content": "Existing Datasets. The most commonly used dataset for category-level 6D object pose estimation is NOCS [40], comprising both the synthetic CAMERA dataset and the real-world REAL dataset. CAMERA includes 300k RGBD images of 31 indoor scenes with 1,085 object instances across 6 categories, while REAL mirrors the categories in CAMERA and includes 8k RGBD images capturing 42 instances in 18 real scenes. Wild6D [44] consists of 5,166 videos with 1.1 million images over 1,722 object instances in 5 categories. ShapeNet-SRN Cars dataset and Sim2Real Cars dataset proposed in iNerf [22] both exclusively include a single car category. The former includes 3,514 instances derived from ShapeNet cars, while the latter is extracted from videos capturing 10 distinct unseen car models. These datasets are limited by their narrow range of categories, hindering their ability to generalize broadly. Additionally, most training images" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "are synthetic and lack realism, and their scenes are overly simplified, failing to account for common real-world challenges like occlusions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 481, + 213 + ], + "type": "text", + "content": "Implicit Methods. Implicit methods are based on point correspondence [6, 20, 24, 34, 37, 40, 44, 46]. NOCS [40], one of the pioneering works in this area, introduced the concept of Normalized Object Coordinate Space (NOCS). The final pose and size of the object are obtained by matching the predicted NOCS map with the observed depth input using the Umeyama algorithm [36] and RANSAC algorithm [12]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 214, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 357 + ], + "type": "text", + "content": "Subsequent algorithms such as DualPoseNet, RBP-Net and RePoNet [20,44, 46] have continued to develop along the vein of NOCS, implicitly solving for pose after predicting the NOCS map. SPD [34] proposed a category-level shape prior, subsequently deforming this shape prior (i.e., average shape) to fit observed point cloud. SGPA, RePoNet, and CATRE [6,24,44] continue to develop along SPD's category-level shape prior approach. Algorithms like 6-PACK and SGPA [6,37] extract low-rank structure points, i.e., keypoints, from dense observed point clouds. 6-PACK [37] predicts interframe motion of target instances through keypoint matching, while SGPA [6] employs keypoints for more effective incorporation of sparse structural information during prior adaptation. These methods rely heavily on the RANSAC process to eliminate outliers, making them non-differentiable and time-consuming." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 562 + ], + "type": "text", + "content": "Explicit Methods. Explicit methods are based on direct pose regression [7, 10, 20, 24, 46, 47]. DualPoseNet and RBP-Net [20, 46] conduct both explicit and implicit training, where one parallel pose decoder explicitly regresses the pose. CATRE [24], recognizes the inherent difference between estimations of rotation and translation/size, explicitly regressing their residuals and carrying out an iterative pose estimation process. FS-Net [7] designs an autoencoder with 3D Graphic Convolution for latent feature extraction and separates the predictions for rotation and translation/size into two distinct networks: one estimates translation/size through two residuals, while the other handles rotation prediction by estimating deflections on two orthogonal axes. GPV-Pose and HS-Pose [10, 47] utilize the same foundational mechanism introduced by FS-Net [7]. GPV-Pose [10] proposes a decoupled confidence-driven rotation representation that facilitates geometrically-aware recovery of correlated rotation matrices and introduces a new geometry-guided point-by-point voting paradigm for robust retrieval of 3D object bounding boxes. Meanwhile, HS-Pose [47] extends 3D-GC to extract mixed-range latent features from point cloud data through a simple network structure known as the HS layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 586, + 254, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 254, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 254, + 598 + ], + "type": "text", + "content": "3 Omni6D Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 616, + 227, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 227, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 227, + 628 + ], + "type": "text", + "content": "3.1 Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Dataset Collection. As shown in Tab. 1, Omni6D comprises 4,688 instances across an impressive span of 166 categories. Each instance is a high-resolution" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "text", + "content": "textured mesh, obtained using Shining 3D scanner" + }, + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "text", + "content": " and Artec Eva 3D scanner" + }, + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "text", + "content": ", collected from OmniObject3D [41]. We normalize object models to fit within a " + }, + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "inline_equation", + "content": "(-1,1)^{3}(m^{3})" + }, + { + "bbox": [ + 130, + 115, + 479, + 186 + ], + "type": "text", + "content": " three-dimensional space, and align objects within each category to a consistent canonical pose. In the latest dataset, Omni6D-xl builds upon and extends Omni6D, comprising 13,161 instances across an impressive span of 339 categories. For more details, please refer to Appendix Section C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 331 + ], + "type": "text", + "content": "**Rendering.** We employ stratified sampling to split instances within each category, subsequently dividing them into training, validation, and test sets in a 7:2:1 ratio. In the construction of our dataset, we utilize 9 room models from the Replica dataset as backdrops. For each scenery setup, we randomly select a room model to act as the background, along with 6 - 8 object instance models, which are allowed to perform free-fall motion within the room model, resulting in random scattering in a specific section of the room. Each object model is scaled by a random factor ranging from 0.8 to 1.2 as part of our data augmentation strategy. Considering the attention center of the combined instance models as the origin point, the camera randomly selects ten positions within a radius of " + }, + { + "bbox": [ + 130, + 188, + 480, + 331 + ], + "type": "inline_equation", + "content": "8 - 9m" + }, + { + "bbox": [ + 130, + 188, + 480, + 331 + ], + "type": "text", + "content": " and an elevation angle range between " + }, + { + "bbox": [ + 130, + 188, + 480, + 331 + ], + "type": "inline_equation", + "content": "30 - 90^{\\circ}" + }, + { + "bbox": [ + 130, + 188, + 480, + 331 + ], + "type": "text", + "content": ". The camera then performs rendering at these selected positions while facing towards the attention center." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "text", + "content": "Setting. We utilize BlenderProc 2.5.0 [9] to implement the aforementioned rendering process. The intrinsic parameters of the camera are set to [577.5, 577.5, 319.5, 239.5], with an image size specified as " + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "text", + "content": ". Our approach ensures the diversity and breadth of the dataset, making it suitable for rigorous testing and yielding accurate results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 411, + 250, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 250, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 250, + 422 + ], + "type": "text", + "content": "3.2 Data Annotations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 431, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 479, + 502 + ], + "type": "text", + "content": "Rich Annotations. Each rendered output includes the ground truth class label, instance mask, NOCS mapping [40], depth map, as well as 6D pose and size. Fig. 1 exhibits a selection of rendered outputs. To reduce the storage size of the dataset, we encode high-precision depth maps into RGB images by multiplying depth by 10,000, rounding to nearest integer, and converting to base 256. The resulting three digits represent RGB channels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 503, + 480, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 480, + 634 + ], + "type": "text", + "content": "Rotational Invariance. Rotational invariance implies that a symmetric object can retain its original shape after rotation by certain angles. Many common objects have this property. As shown in Fig. 6, we define the coordinate system as a right-handed system with the x-axis pointing outwards and the y-axis oriented upwards. We contemplate three cases of rotational invariance where an object maintains its original shape after following rotations: any degrees (Sym-1), multiples of 90 degrees (Sym-2) and 180 degrees (Sym-3). Additionally, we denote the case of no rotational invariance around the axis as Sym-0. According to these definitions, all objects in Omni6D are annotated for their rotational symmetry around the xyz-axes. It's worth noting that symmetry attributes may differ among instances within the same category, requiring instance-level rather" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 252, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 252, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 252, + 654 + ], + "type": "text", + "content": "1 https://www.einscan.com/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 654, + 246, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 246, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 246, + 665 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 654, + 246, + 665 + ], + "type": "text", + "content": " https://www.artec3d.cn/" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 304 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 304 + ], + "type": "image", + "image_path": "21a06c09d06bd4bceb684939b8b7eea0d18a22f25c84e14e1f783f15922c5b0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 346 + ], + "type": "text", + "content": "Fig. 2: Symmetry statistics. The figure demonstrates different symmetry cases using object instances and provides a quantitative representation of the occurrence frequency for various combinations of distinct symmetry cases across the xyz-axes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 367, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 416 + ], + "type": "text", + "content": "than category-level annotations. Fig. 6 illustrates all kinds of symmetry cases using object instances and quantifies their occurrence frequency. Fig. 1 selects several examples to provide a more visual explanation of rotational invariance. These considerations are then integrated into our evaluation protocols in Sec. 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 431, + 250, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 250, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 250, + 443 + ], + "type": "text", + "content": "3.3 Dataset Statistics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 605 + ], + "type": "text", + "content": "Spatial Statistics. Omni6D aims to overcome challenges in estimating poses for occluded object instances. Fig. 3a and Fig. 3b shows the spatial distribution of point clouds and objects by projecting their centroids on the XY-plane (top) and XZ-plane (bottom) [4]. Fig. 3c depicts the relative object size distribution, defined as the square root of the object-to-image area ratio. We observe that the spatial distribution of Omni6D is similar to that of CAMERA and REAL, with a greater resemblance to CAMERA despite having a closer depth range. However, a more pronounced discrepancy between the spatial distribution of point clouds and objects is evident in Omni6D compared to CAMERA and REAL. This observation suggests a higher occurrence of occlusion scenes in Omni6D, highlighting the intricate challenges it presents to 6D object pose estimation. Nonetheless, as depicted in Fig. 4a, algorithms trained on Omni6D demonstrate their robustness in tackling these complexities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Angular Deviation. Omni6D enables accurate pose estimation using only the lower half or bottom appearance of objects. Fig. 3d depicts the density of angular deviations from the upward direction, i.e. y-axis. Our dataset displays a more uniform distribution of object angles relative to the upward axis and exhibits greater deviation from the canonical pose angles. Unlike NOCS, which primarily" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 223, + 172 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 223, + 172 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 223, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 223, + 172 + ], + "type": "image", + "image_path": "8fee4c399f65e261ce78ad239b6c70672ec07711d2cd00d69f74712914c2c832.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 116, + 317, + 172 + ], + "blocks": [ + { + "bbox": [ + 233, + 116, + 317, + 172 + ], + "lines": [ + { + "bbox": [ + 233, + 116, + 317, + 172 + ], + "spans": [ + { + "bbox": [ + 233, + 116, + 317, + 172 + ], + "type": "image", + "image_path": "39c29b7165ee116496b5b46d2b01809f3ea22361cecf798569bb9851a1b529a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 175, + 227, + 243 + ], + "blocks": [ + { + "bbox": [ + 137, + 175, + 227, + 243 + ], + "lines": [ + { + "bbox": [ + 137, + 175, + 227, + 243 + ], + "spans": [ + { + "bbox": [ + 137, + 175, + 227, + 243 + ], + "type": "image", + "image_path": "7d37c58de0e61aa6e6d242ce52c589c3410f4c955e4e848f5f7bba27e754f4a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 244, + 228, + 253 + ], + "lines": [ + { + "bbox": [ + 136, + 244, + 228, + 253 + ], + "spans": [ + { + "bbox": [ + 136, + 244, + 228, + 253 + ], + "type": "text", + "content": "(c) Relative 2D object size" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 175, + 319, + 243 + ], + "blocks": [ + { + "bbox": [ + 239, + 175, + 319, + 243 + ], + "lines": [ + { + "bbox": [ + 239, + 175, + 319, + 243 + ], + "spans": [ + { + "bbox": [ + 239, + 175, + 319, + 243 + ], + "type": "image", + "image_path": "3d2a6dc09651eda420e55261b4d686340073f719d92e6cbff552b2ab33ce5d30.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 244, + 316, + 253 + ], + "lines": [ + { + "bbox": [ + 237, + 244, + 316, + 253 + ], + "spans": [ + { + "bbox": [ + 237, + 244, + 316, + 253 + ], + "type": "text", + "content": "(d) Angular deviation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 327, + 116, + 451, + 241 + ], + "blocks": [ + { + "bbox": [ + 327, + 116, + 451, + 241 + ], + "lines": [ + { + "bbox": [ + 327, + 116, + 451, + 241 + ], + "spans": [ + { + "bbox": [ + 327, + 116, + 451, + 241 + ], + "type": "image", + "image_path": "c18e3c527805bd2f8dd6ab3bc6734b5033be4bee52a798a040a219a804cd5ad0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 243, + 428, + 252 + ], + "lines": [ + { + "bbox": [ + 351, + 243, + 428, + 252 + ], + "spans": [ + { + "bbox": [ + 351, + 243, + 428, + 252 + ], + "type": "text", + "content": "(e) Clustering results" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 262, + 480, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 318 + ], + "type": "text", + "content": "Fig. 3: Omni6D analysis. (a) distribution of point cloud centroids, (b) distribution of object centroids on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (c) density of relative 2D object size, (d) density of angular deviation from the upward direction, (e) Omni6D dataset clustering results. The angle of each sector in the chart reflects the relative size of the instance count within that category." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 341, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 480, + 389 + ], + "type": "text", + "content": "uses upright object placement, Omni6D utilizes physical simulations for freefall object positioning [9]. As a result, it presents more challenging and diverse pose estimation scenes. Training on Omni6D enhances algorithms' robustness to object rotation angles, as evidenced by the image in Fig. 4b." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 389, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 533 + ], + "type": "text", + "content": "Shape Priors. We obtain the mean latent embedding and shape prior for each category from the variational autoencoder [5]. Fig. 1 showcases categorical shape priors, each displaying unique characteristics, facilitating an intuitive association between point cloud shapes and corresponding real-world entities. Meanwhile, Fig. 3e explains clustering results based on categorical latent embeddings, where we employ agglomerative clustering [27] to group categories into 20 clusters. It highlights the geometric coherence among semantically identical objects (especially man-made ones) in Omni6D dataset and further confirms that these categorical shape priors can effectively leverage the wealth of shape information from numerous similar objects to elucidate category features. These insights provide a theoretical basis for applications of category-level 6D object pose estimation using our Omni6D dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 550, + 297, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 297, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 297, + 564 + ], + "type": "text", + "content": "4 Evaluation and Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 574, + 261, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 261, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 261, + 586 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Datasets. Our experimentation utilized two datasets, namely Omni6D and Omni6Dout. Omni6D are partitioned into training, validation, and test sets in a 7:2:1 ratio, denoted as Omni6Dtrain, Omni6Dval and Omni6Dtest respectively. These sets are further subdivided into subsets with increasing category sizes of 3, 6, 12, 24, and 48. We denote the subset containing " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " categories as c1sn. Each subset includes all classes present in the previous subset with additional" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 227, + 192 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 227, + 192 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 227, + 192 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 227, + 192 + ], + "type": "image", + "image_path": "71ab0bbfbe8f9e114c0dfa45c2482f0cda0dc4ae1eca47cd529806e901147a56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 193, + 292, + 202 + ], + "lines": [ + { + "bbox": [ + 163, + 193, + 292, + 202 + ], + "spans": [ + { + "bbox": [ + 163, + 193, + 292, + 202 + ], + "type": "text", + "content": "(a) Challenges from occluded object" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 306, + 192 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 306, + 192 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 306, + 192 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 306, + 192 + ], + "type": "image", + "image_path": "a884f1b76d3e22b47a9e9ae129f19ed947e2be829c487c81fbde988623259513.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 386, + 192 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 386, + 192 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 386, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 386, + 192 + ], + "type": "image", + "image_path": "8f6e3fcb3935ac8383fdd002bbc432fd4255be8def1274eebab5495446190d70.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 212, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 310 + ], + "type": "text", + "content": "Fig. 4: Challenges from Omni6D. (a) Algorithms trained on Omni6D can overcome challenges in estimating poses for occluded object instances. The left shows an occluded object instance at the edge of the image, while the right image shows an object instance obstructed by other objects. (b) Algorithm trained on Omni6D can accurately estimate poses with only the lower half or bottom appearance of an object. The green and red colors respectively denote the ground truth and predicted 3D bounding boxes. The blue and orange lines on the boxes separately highlight the intersecting lines of the frontal face and the top face of the two 3D bounding boxes, while the darker lines indicate the bottom of the bounding boxes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 386, + 114, + 465, + 192 + ], + "blocks": [ + { + "bbox": [ + 386, + 114, + 465, + 192 + ], + "lines": [ + { + "bbox": [ + 386, + 114, + 465, + 192 + ], + "spans": [ + { + "bbox": [ + 386, + 114, + 465, + 192 + ], + "type": "image", + "image_path": "586f974ef75a54088c7f57fd9a14b7105d4364da27f6377bab913ac916f10a06.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 193, + 447, + 203 + ], + "lines": [ + { + "bbox": [ + 325, + 193, + 447, + 203 + ], + "spans": [ + { + "bbox": [ + 325, + 193, + 447, + 203 + ], + "type": "text", + "content": "(b) Challenges from bottom views" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 335, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 407 + ], + "type": "text", + "content": "classes included to meet the desired total. Fig. 6a presents the specific categories included in c1sn and their respective sizes relative to each other. Omni6Dout is utilized as an additional test set to measure our algorithm's inter-category generalization. This dataset, constructed similarly to Omni6D, encompasses 52 models spanning 17 categories unseen in Omni6D, along with 4762 images. For additional details on datasets, please refer to the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 479 + ], + "type": "text", + "content": "Details. All experiments are carried out on a server equipped with an Intel(R) Xeon(R) Gold 6248R CPU @ 3.00GHz and an NVIDIA A100-SXM4-80GB GPU. We maintain consistency in parameters and strategies throughout training, ensuring uniformity in our experiment environment. Given the challenges of semantic classification with a large vocabulary, we use ground truth masks to mitigate the impact of low-quality classification on pose estimation results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 499, + 306, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 306, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 306, + 511 + ], + "type": "text", + "content": "4.2 Symmetry-Aware Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Basic Evaluation Metrics. We utilize the average accuracy of Intersection over 3D Union (IoU) [14] in object detection, and " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "n^{\\circ}m" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " cm in pose estimation. We further decompose " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "n^{\\circ}m" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " cm [19,31] to individually evaluate the model's predictive error " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "n^{\\circ}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " for pose and " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " cm for translation. For these three types of errors, the thresholds considered are " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\{50\\%, 75\\%\\}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\{5^{\\circ}, 10^{\\circ}\\}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\{2cm,5cm\\}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " [3, 30, 42]. Additionally, we set a detection threshold for objects requiring at least a " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " overlap between predicted and ground-truth bounding boxes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Our Symmetry-Aware Metrics. Due to NOCS's limited categories, traditional algorithms mainly handle basic symmetry cases, such as rotational symmetry around the y-axis. However, Omni6D has a wider range of objects with different rotational invariances across multiple axes. Fig. 6 provides symmetry statistics for Omni6D objects. To alleviate this issue, we propose a symmetry-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 130, + 482, + 262 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 384, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 384, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 384, + 128 + ], + "type": "text", + "content": "Algorithm 1 Compute Our Symmetry-Aware Metric " + }, + { + "bbox": [ + 133, + 116, + 384, + 128 + ], + "type": "inline_equation", + "content": "L_{s}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": "1: procedure SYMMETRIC_METRIC(L, R, " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "n_x" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "n_y" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "n_z" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": ") \n2: " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\Theta_0 = \\{0^\\circ\\}" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\Theta_2 = \\{0^\\circ, 90^\\circ, 180^\\circ, 270^\\circ\\}" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\Theta_3 = \\{0^\\circ, 180^\\circ\\}" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " // Rotations around Sym-1 axis need not be considered. \n5: " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "c =" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " count(1 occurrences in " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\{n_x, n_y, n_z\\}" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": ") \n6: if " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "c \\geq 2" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " then // The object is a sphere. \n7: " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "L_s = L(R^*, R)" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " \n8: else if " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "c == 1" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " then // Rotations around Sym-1 axis can be disregarded. \n9: Without loss of generality, assume " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "n_x == 1" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": ". \n10: " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "L_s = \\min_{\\theta_y \\in \\Theta_{ny}, \\theta_z \\in \\Theta_{nz}} L(R_{\\theta_y, \\theta_z}^*, R)" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " \n11: else if " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "c == 0" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " then // Simply enumerate all cases. \n12: " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "L_s = \\min_{\\theta_x \\in \\Theta_{nx}, \\theta_y \\in \\Theta_{ny}, \\theta_z \\in \\Theta_{nz}} L(R_{\\theta_x, \\theta_y, \\theta_z}^*, R)" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " \n13: end if \n14: return " + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 134, + 130, + 482, + 262 + ], + "type": "text", + "content": " \n15: end procedure" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 283, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 479, + 306 + ], + "type": "text", + "content": "aware metric. Unlike prior works focusing solely on the y-axis, our method considers rotation symmetry around all three axes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": "We define the relevant variables as follows: " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " denotes our symmetry-aware metric, " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " denotes the original metric. " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " stands for the ground truth rotation matrix, while " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " represents the predicted rotation matrix. " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "R_{\\theta_x,\\theta_y,\\theta_z}^*" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " corresponds to the predicted rotation matrix after sequentially rotating by " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "\\theta_{x},\\theta_{y}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "\\theta_{z}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " degrees around the xyz axes. The rotational invariance cases around the x, y, and z axes are denoted as Sym- " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "n_{x}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": ", Sym- " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "n_{y}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": ", and Sym- " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "n_{z}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "n_x,n_y" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "n_z" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " are the respective rotation parameters. Objects that align with Sym- " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": " around an axis maintain their original shape when rotated by an angle from " + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "inline_equation", + "content": "\\Theta_{n}" + }, + { + "bbox": [ + 130, + 307, + 480, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": "Since the Euler angles are compact [13], the most straightforward approach is to determine the category of rotational invariance for each axis " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x},\\mathbf{y},\\mathbf{z}\\}" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": " sequentially, as mentioned in 3.2. To simplify computations, we set " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "\\Theta_0 = \\{0^\\circ \\}" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "\\Theta_{1} = \\{0^{\\circ},1^{\\circ},\\dots,359^{\\circ}\\}" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "\\Theta_{2} = \\{0^{\\circ},90^{\\circ},180^{\\circ},270^{\\circ}\\}" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "\\Theta_{3} = \\{0^{\\circ},180^{\\circ}\\}" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": ". We can define " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "inline_equation", + "content": "L_{s} = \\min_{\\theta_{x}\\in \\Theta_{n_{x}},\\theta_{y}\\in \\Theta_{n_{y}},\\theta_{z}\\in \\Theta_{n_{z}}}L(R_{\\theta_{x},\\theta_{y},\\theta_{z}}^{*},R)" + }, + { + "bbox": [ + 130, + 403, + 480, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 480, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 480, + 511 + ], + "type": "text", + "content": "However, due to the singularity of Euler angles [13], we can simplify the above rotation transformation. The pseudo-code implementation of our Symmetry-Aware Evaluation is provided in Algorithm 1. It allows us to simplify what was originally at most " + }, + { + "bbox": [ + 130, + 464, + 480, + 511 + ], + "type": "inline_equation", + "content": "360^{3}" + }, + { + "bbox": [ + 130, + 464, + 480, + 511 + ], + "type": "text", + "content": " computations to a maximum of only " + }, + { + "bbox": [ + 130, + 464, + 480, + 511 + ], + "type": "inline_equation", + "content": "4^{3}" + }, + { + "bbox": [ + 130, + 464, + 480, + 511 + ], + "type": "text", + "content": " computations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 527, + 399, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 399, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 399, + 540 + ], + "type": "text", + "content": "4.3 Large-Vocabulary 6D Pose and Size Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 640 + ], + "type": "text", + "content": "Performance on Omni6D. We present results of algorithms [6,10,34,46,47] trained on Omni6Dtrain and tested on Omni6Dtest. We compare their quantitative results in Tab. 2 and their qualitative results in Fig. S10 in Appendix. The performance disparity among algorithms for category-level 6D object pose estimation becomes markedly pronounced when applied to large-vocabulary datasets, in contrast to the more consistent performance previously observed on the Real and CAMERA datasets [40]. This highlights the inherent strengths and weaknesses across various model structures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "This observation suggests the potential importance of our large-vocabulary dataset in uncovering the relative performance of different models. It appears" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 179, + 480, + 249 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Category-level performance on Omni6D dataset. Models are trained on Omni6Dtrain and tested on Omni6Dtest. Instances within each category in the test set are unseen during training, substantiating the algorithms' capacity to generalize within individual categories under large-vocabulary settings. Bold and underlined results indicate the best and second-best performers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 179, + 480, + 249 + ], + "lines": [ + { + "bbox": [ + 133, + 179, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 179, + 480, + 249 + ], + "type": "table", + "html": "
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit44.5620.377.559.5614.7619.2310.6821.0237.4970.09
SGPA [6]implicit36.3414.444.786.8410.1315.038.4917.7325.5759.18
DualPoseNet [20]hybrid58.8425.498.289.3017.2619.059.3819.1873.8296.37
RBP-Pose [46]hybrid35.924.660.370.600.530.800.750.9639.7383.55
GPV-Pose [10]explicit15.280.260.100.700.140.962.252.965.3133.70
HS-Pose [47]explicit62.6523.024.264.8510.4911.614.9611.7580.9397.78
", + "image_path": "e6f45eb75e29e5ebdaca5cad65b230e7f225bfcab51f337ae31cac986691cce6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 304, + 480, + 373 + ], + "blocks": [ + { + "bbox": [ + 130, + 261, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 295 + ], + "type": "text", + "content": "Table 3: Category-level performance on unseen categories. Models are trained on Omni6Dtrain and tested on Omni6Dout. Categories in the test set never appear in the training set, validating the algorithms' ability to generalize across categories." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 304, + 480, + 373 + ], + "lines": [ + { + "bbox": [ + 133, + 304, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 480, + 373 + ], + "type": "table", + "html": "
MethodsNetworkIoU50IoU755°2cm5°5cm10°2cm10°5cm5°10°2cm5cm
SPD [34]implicit7.560.950.180.400.801.650.652.368.8840.59
SGPA [6]implicit7.050.600.070.280.190.820.531.693.8728.28
DualPoseNet [20]hybrid36.8512.063.243.378.048.513.398.6478.0098.60
RBP-Pose [46]hybrid26.181.950.010.020.020.030.020.0316.7443.06
GPV-Pose [10]explicit10.970.140.030.180.120.570.301.077.1441.30
HS-Pose [47]explicit36.758.921.541.664.675.161.755.3879.9598.27
", + "image_path": "585bedaa562e7b9201e8a889ed9f401c61069c0c42d843d52194ac43db91ff16.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "text", + "content": "that the increased complexity of the dataset could push model architectures to their theoretical limits, possibly revealing intrinsic characteristics otherwise obscured in less complex scenarios. For example, SPD, SGPA is particularly proficient in predicting rotation, and SPD achieves the highest score in " + }, + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "inline_equation", + "content": "n^{\\circ}mcm" + }, + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "text", + "content": ". This could be due to its implicit network's propensity for generating more reliable rotational forecasts. Meanwhile, DualPoseNet and HS-Pose provide more accurate predictions for translation and score higher in IoU. This could be associated with the characteristic of models with explicit networks to produce better translations and size estimates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "text", + "content": "Our large-vocabulary dataset, encompassing a broad spectrum of shapes and appearances, enables a comprehensive evaluation of diverse category-level pose estimation methods. This serves not only as a robust test of an algorithm's generalizability but also as a valuable tool in understanding the advantages offered by different algorithmic structures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": "Generalization Performance. We evaluate algorithms on " + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{Omni6D}_{out}" + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": " to assess their inter-category generalization capabilities. The outcomes are presented in Tab. 3. Notably, DualPoseNet and HS-Pose emerged as superior performers, outclassing others across all metrics, thereby demonstrating excellent generalization abilities. Contrastingly, implicit methods including SPD and SPGA exhibited marked limitations. Qualitative results are shown in Fig. S11 in Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Drawing parallels with the observations from Tab. 2, we found that metrics such as translation and IoU were relatively easier to excel in, suggesting superior" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 471, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 471, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 471, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 471, + 198 + ], + "type": "image", + "image_path": "4570cd25a0b82c7cbb1192ae3e87459036b14737d54672c088e5deb7b36383b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 209, + 482, + 286 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 286 + ], + "type": "text", + "content": "Fig. 5: Category-Wise Performance on Omni6D Dataset. The x-axis, moving from left to right, sequentially represents: the number of objects within a category (Semantic Category), the number of objects within a cluster clustered based on shape priors (Shape Category) and the diversity of instances within a category. The y-axis depicts category or clustered group results for IoU75 and 5°2 cm metrics. Each plotted point illustrates the algorithm's result for a specific category or cluster, while the line showcases the trend of the linear fit for the scattered points." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 332, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 480, + 369 + ], + "type": "text", + "content": "generalization abilities in translation and size prediction. Conversely, the generalization of rotation emerges as a considerable challenge in category-level 6D object pose estimation, especially within large-vocabulary scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 522 + ], + "type": "text", + "content": "Category-wise Analysis. Based on the " + }, + { + "bbox": [ + 130, + 379, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 130, + 379, + 482, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 379, + 482, + 522 + ], + "type": "inline_equation", + "content": "5^{\\circ}2" + }, + { + "bbox": [ + 130, + 379, + 482, + 522 + ], + "type": "text", + "content": " cm metrics, we conducted a detailed category-wise analysis of the results from Tab. 2. Left columns in Fig. 5 illustrate the correlation between category-level 6D pose estimation performance and the number of instances within each category in Omni6Dtrain. Middle columns in Fig. 5 analyze the correlation between cluster-level average performance and cluster size based on the clustering results described in Fig. 3e. We found that the performance of pose estimation for each category is more strongly correlated with the number of instances within clusters than with semantic categories, showing a positive correlation. This suggests that shape categories have a greater impact on training than semantic categories do. Notably, algorithms like SPD, SGPA, and RBP-Pose that utilize shape prior structures are particularly sensitive to this influence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Right columns in Fig. 5 reveal the correlation of pose estimation performance relative to instance diversity within each category in the training set. We measured instance diversity by calculating the mean chamfer distance [1] among all pairs of instances in each category. The results show that as diversity within a category increases, pose estimation performance tends to improve. This observation aligns with the assertion made by [23]: The key to the success of prior-based methods lies in the deformation modules, which learns to synthesize world-space target objects and explicitly builds the correspondence between camera and world-space. As the number of instances increases and the diversity within a shape category expands, the model's capacity to learn deformation from priors to actual instance shapes is strengthened, leading to improved results." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 272, + 224 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 272, + 224 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 272, + 224 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 272, + 224 + ], + "type": "image", + "image_path": "c85a544023dec18ffe03f680575a37364787a74cbda9b5523875a42459d5637a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 227, + 260, + 237 + ], + "lines": [ + { + "bbox": [ + 150, + 227, + 260, + 237 + ], + "spans": [ + { + "bbox": [ + 150, + 227, + 260, + 237 + ], + "type": "text", + "content": "(a) Category inventory of clsn" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 280, + 121, + 457, + 224 + ], + "blocks": [ + { + "bbox": [ + 280, + 121, + 457, + 224 + ], + "lines": [ + { + "bbox": [ + 280, + 121, + 457, + 224 + ], + "spans": [ + { + "bbox": [ + 280, + 121, + 457, + 224 + ], + "type": "image", + "image_path": "e3ef99cf02c9d254bf358461b49c61e658fc66d4a7ae50b861cceaee9b6b7f9d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 227, + 419, + 237 + ], + "lines": [ + { + "bbox": [ + 328, + 227, + 419, + 237 + ], + "spans": [ + { + "bbox": [ + 328, + 227, + 419, + 237 + ], + "type": "text", + "content": "(b) Our finetune strategy" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 246, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 312 + ], + "type": "text", + "content": "Fig.6: Our finetune strategy. (a) Category inventory of c1sn within Omni6D dataset. The angle of each sector in the chart reflects the relative size of the instance count within that category. (b) In each fine-tuning step, we double the category count, copying trained global features and old category parameters into the new network while initializing the new category parameters. An observable deepening of color is indicative of the escalating count of training iterations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 335, + 349, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 349, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 349, + 348 + ], + "type": "text", + "content": "4.4 Fine-Tuning from Limited Categories" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": "We propose a finetuning strategy that helps extend methods from a limited set of categories to large-vocabulary. We take SPD [34], DualPoseNet [20], and HS-Pose [47] as examples which belong to three different network architectures and show good performance on Omni6D" + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "inline_equation", + "content": "_{test}" + }, + { + "bbox": [ + 130, + 354, + 482, + 413 + ], + "type": "text", + "content": ". We respectively take their best models on CAMERA as our pre-trained models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "content": "Initiating the fine-tuning process, we utilize three categories: bottle, bowl, and cup, which are concurrently present in both Omni6D and CAMERA datasets, aligning with the cls3 category. By facilitating the training on Omni6D-cls3, we enable a transfer of the model from CAMERA to Omni6D. Following the method illustrated in Fig. 6b, we engage in an iterative fine-tuning process on a progressively expanded category dataset until it reaches our desired number. In our experiments, we set this target number to be 48 categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 557 + ], + "type": "text", + "content": "In parallel, we conduct training from scratch separately on cls3, cls6, ..., and cls48 as a comparison, employing the same number of training iterations. As shown in Fig. 7, even with an exponential increase in the number of categories, pre-trained models remain pivotal in our fine-tuning strategy. The performance of fine-tuning consistently outperforms that of training from scratch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "However, regardless of whether the training approach is finetuning or training from scratch, a decline in performance is observed as the number of categories increases. The decline rates for SPD and DualPoseNet are slower, coupled with an initial augmentation in performance due to increased training data and iterations. In contrast, HS-Pose experiences a more rapid decline, with fine-tuned " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "5^{\\circ}2" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " cm results dropping from initial " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "62.52\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "14.42\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ". Models that excel in tasks involving a limited number of categories may not necessarily maintain their superiority in large-vocabulary tasks, they might be surpassed by models that are more robust and easier to train." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 116, + 462, + 254 + ], + "blocks": [ + { + "bbox": [ + 151, + 116, + 462, + 254 + ], + "lines": [ + { + "bbox": [ + 151, + 116, + 462, + 254 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 462, + 254 + ], + "type": "image", + "image_path": "ba9ac324d8671bccc8c28c4b9fcfa9dcde22967901e28bfec40073491d65830a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "text", + "content": "Fig. 7: Finetuned results. Each figure's x-axis represents the number of categories in the training and test set, while the y-axis displays the outcomes of " + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "inline_equation", + "content": "5^{\\circ}2" + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "text", + "content": " cm, " + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "inline_equation", + "content": "2cm" + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "text", + "content": " metrics. Each row, from top to bottom, sequentially employs three methods: SPD [34], DualPoseNet [20], and HS-Pose [47]. The figures depict the outcomes derived from two training strategies as the number of training categories increases, accompanied by the gradual expansion of corresponding test sets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 361, + 236, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 236, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 236, + 372 + ], + "type": "text", + "content": "4.5 Visual Realism" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 394, + 479, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 479, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 479, + 465 + ], + "type": "text", + "content": "Due to the complexity of collecting and annotating real-world data, contemporary datasets like NOCS [40] are composed of a large amount of synthetic data and a small portion of real-world data. While collecting real data is relatively straightforward when the number of categories is limited, gathering well-annotated real-world data for pose estimation tasks involving large vocabulary categories becomes a monumental task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 480, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 516 + ], + "type": "text", + "content": "Our Omni6D dataset, which includes large vocabulary objects, is also derived from rendering. However, the incorporation of real-scanned objects significantly enhances the realism of the rendered images. As depicted in Fig. 8, Omni6D receives a score of " + }, + { + "bbox": [ + 130, + 468, + 480, + 516 + ], + "type": "inline_equation", + "content": "2.69 \\pm 0.39" + }, + { + "bbox": [ + 130, + 468, + 480, + 516 + ], + "type": "text", + "content": ", surpassing the results obtained by CAMERA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 519, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 627 + ], + "type": "text", + "content": "Given these significant advantages, our dataset excels not only in large-vocabulary scenarios but also in real-world scenes. As depicted in Tab. 4, We use DualPoseNet [20] to train on the common categories in REAL [40] and Omni6D, namely bottles, bowls, and mugs. We train separately on the two datasets and their mix. The results show that Omni6D models perform well on REAL275, and training on the mixed dataset outperforms using REAL or Omni6D datasets alone. This demonstrates that our dataset enables the direct transfer of models to real-world scenes. Moreover, it seamlessly supplements the existing real-world dataset, enabling joint training of models on our dataset and the real-world data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "To further validate the sim2real capability of models trained with Omni6D, we constructed a real-world dataset, Omni6D-Real, comprising 30 scenes, 39 categories, 73 instances, and 1k images. We captured RGBD images with Azure" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 114, + 238, + 180 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 238, + 180 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 238, + 180 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 238, + 180 + ], + "type": "image", + "image_path": "38cafe713d1ef714db0a8789a5bbedf739e0033b567d6dbaa80f835060781305.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 182, + 212, + 199 + ], + "lines": [ + { + "bbox": [ + 176, + 182, + 212, + 199 + ], + "spans": [ + { + "bbox": [ + 176, + 182, + 212, + 199 + ], + "type": "text", + "content": "Omni6D " + }, + { + "bbox": [ + 176, + 182, + 212, + 199 + ], + "type": "inline_equation", + "content": "2.69\\pm 0.39" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 114, + 325, + 180 + ], + "blocks": [ + { + "bbox": [ + 238, + 114, + 325, + 180 + ], + "lines": [ + { + "bbox": [ + 238, + 114, + 325, + 180 + ], + "spans": [ + { + "bbox": [ + 238, + 114, + 325, + 180 + ], + "type": "image", + "image_path": "55df81ccc0ce4ee128d7ec79f7423764222c93e70de8753013ec6fbeee0fe8d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 182, + 300, + 199 + ], + "lines": [ + { + "bbox": [ + 264, + 182, + 300, + 199 + ], + "spans": [ + { + "bbox": [ + 264, + 182, + 300, + 199 + ], + "type": "text", + "content": "CAMERA " + }, + { + "bbox": [ + 264, + 182, + 300, + 199 + ], + "type": "inline_equation", + "content": "1.55\\pm 0.08" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "text", + "content": "Fig. 8: Comparison of Visual Realism. We evaluated the visual realism of Omni6D in comparison to other datasets through a survey involving 70 human subjects. We randomly selected 10 images from each dataset and introduced noise by blending in 5 images from COCO [21], which included captured photos, and " + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{SKETCH}^4" + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "text", + "content": ", which comprised rendered images. Subjects were asked to rate the realism of sampled images on a scale from 1 (least realistic) to 5 (most realistic). We report the mean and standard deviation and include a sampled image from the study." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 325, + 114, + 413, + 180 + ], + "blocks": [ + { + "bbox": [ + 325, + 114, + 413, + 180 + ], + "lines": [ + { + "bbox": [ + 325, + 114, + 413, + 180 + ], + "spans": [ + { + "bbox": [ + 325, + 114, + 413, + 180 + ], + "type": "image", + "image_path": "7b9cffb67d1dbd3750ea2af15078b90bead68e931b3c33ee7a8e542584b0f8ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 182, + 388, + 199 + ], + "lines": [ + { + "bbox": [ + 351, + 182, + 388, + 199 + ], + "spans": [ + { + "bbox": [ + 351, + 182, + 388, + 199 + ], + "type": "text", + "content": "REAL " + }, + { + "bbox": [ + 351, + 182, + 388, + 199 + ], + "type": "inline_equation", + "content": "3.53\\pm 0.28" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 413, + 114, + 463, + 180 + ], + "blocks": [ + { + "bbox": [ + 413, + 114, + 463, + 180 + ], + "lines": [ + { + "bbox": [ + 413, + 114, + 463, + 180 + ], + "spans": [ + { + "bbox": [ + 413, + 114, + 463, + 180 + ], + "type": "image", + "image_path": "f37437a1c1f431216714574d7ee1872c7966b598f3e50fea098885e97b7140f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 182, + 456, + 199 + ], + "lines": [ + { + "bbox": [ + 419, + 182, + 456, + 199 + ], + "spans": [ + { + "bbox": [ + 419, + 182, + 456, + 199 + ], + "type": "text", + "content": "WILD6D " + }, + { + "bbox": [ + 419, + 182, + 456, + 199 + ], + "type": "inline_equation", + "content": "4.38\\pm 0.23" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 148, + 345, + 463, + 385 + ], + "blocks": [ + { + "bbox": [ + 130, + 301, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 335 + ], + "type": "text", + "content": "Table 4: Performance on REAL275 with Different Training Sets. It compares how different training sets influence DualPoseNet's performance on REAL275 [40], providing insights into the model's ability to generalize in real-world tasks using Omni6D." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 345, + 463, + 385 + ], + "lines": [ + { + "bbox": [ + 148, + 345, + 463, + 385 + ], + "spans": [ + { + "bbox": [ + 148, + 345, + 463, + 385 + ], + "type": "table", + "html": "
Train dataRealismIoU50IoU755°2cm5°5cm10°2cm10°5cm5°2cm
Omni6DReal-Scanned78.7632.696.558.8015.0021.3811.2049.54
REAL [40]Real84.5143.438.7610.4021.2425.3913.0169.46
REAL+Omni6DMixed85.2858.5914.1017.8330.1038.9720.9671.00
", + "image_path": "c8adf5872a9f5b977cc30b55b749aaba40ae431fc9353a762329178fda12eee1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 405, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 480, + 431 + ], + "type": "text", + "content": "Kinect " + }, + { + "bbox": [ + 130, + 405, + 480, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{DK}^3" + }, + { + "bbox": [ + 130, + 405, + 480, + 431 + ], + "type": "text", + "content": " and preprocessed them using SAM [18] for object masks and ICP [2] for point cloud registration. Details are provided in Appendix Section D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 447, + 220, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 220, + 459 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 220, + 459 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": "In conclusion, this paper introduces Omni6D, a novel 6D pose estimation dataset with large-vocabulary categories and intricate scenes. We evaluate existing category-level 6D object pose estimation methods on this benchmark, analyze its challenges, and propose a fine-tuning strategy for large-vocabulary scenarios. Limitations. Our dataset, though more complex, doesn't fully encompass all real-world challenges. Additionally, our fine-tuning strategy effectively extends methods from a small set to a larger one, but its efficacy may decrease with growing category diversity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": "Future Work. Our study paves the way for diverse research avenues. An immediate next step is expanding the Omni6D dataset with more object types and scenes for comprehensive coverage. Additionally, annotating videos for scanned objects will validate algorithms' large-vocab pose estimation in real-world scenarios. Designing new training strategies for coping with increasing category diversity presents an intriguing challenge." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 328, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 328, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 328, + 654 + ], + "type": "text", + "content": "3 https://learn.microsoft.com/azure/kinect-dk/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 654, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 238, + 665 + ], + "type": "text", + "content": "4 https://sketchfab.com/" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "This research is supported by Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 199, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 199, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 199, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 257, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 290 + ], + "type": "text", + "content": "1. Barrow, H.G., Tenenbaum, J.M., Bolles, R.C., Wolf, H.C.: Parametric correspondence and chamfer matching: Two new techniques for image matching. In: IJCAI. pp. 659-663. William Kaufmann (1977)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 292, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 480, + 312 + ], + "type": "text", + "content": "2. Besl, P.J., McKay, N.D.: A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell. 14(2), 239-256 (1992)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "type": "text", + "content": "3. Brachmann, E., Michel, F., Krull, A., Yang, M.Y., Gumhold, S., Rother, C.: Uncertainty-driven 6d pose estimation of objects and scenes from a single RGB image. In: CVPR. pp. 3364-3372 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "4. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3d: A large benchmark and model for 3d object detection in the wild. In: CVPR. pp. 13154-13164 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "text", + "content": "5. Chen, D., Li, J., Wang, Z., Xu, K.: Learning canonical shape space for category-level 6d object pose and size estimation. In: CVPR. pp. 11970-11979. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 434 + ], + "type": "text", + "content": "6. Chen, K., Dou, Q.: SGPA: structure-guided prior adaptation for category-level 6d object pose estimation. In: ICCV. pp. 2753-2762 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 467 + ], + "type": "text", + "content": "7. Chen, W., Jia, X., Chang, H.J., Duan, J., Shen, L., Leonardis, A.: Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In: CVPR. pp. 1581-1590 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "text", + "content": "8. Chen, X., Dong, Z., Song, J., Geiger, A., Hilliges, O.: Category level object pose estimation via neural analysis-by-synthesis. In: ECCV (26). pp. 139-156 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 521 + ], + "type": "text", + "content": "9. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. J. Open Source Softw. 8(83), 4901 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "0. Di, Y., Zhang, R., Lou, Z., Manhardt, F., Ji, X., Navab, N., Tombari, F.: Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In: CVPR. pp. 6771-6781 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "1. Du, G., Wang, K., Lian, S., Zhao, K.: Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artif. Intell. Rev. 54(3), 1677-1734 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "text", + "content": "2. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "3. Gao, X., Zhang, T.: Introduction to Visual SLAM - From Theory to Practice. Springer (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "4. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Huang, S., Qi, S., Xiao, Y., Zhu, Y., Wu, Y.N., Zhu, S.: Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. In: NeurIPS. pp. 206-217 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "16. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: ICRA. pp. 10632-10640. IEEE (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 215 + ], + "type": "text", + "content": "17. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape, appearance, and pose optimization. In: ECCV (2). Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "text", + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 480, + 268 + ], + "type": "text", + "content": "19. Li, Y., Wang, G., Ji, X., Xiang, Y., Fox, D.: Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis. 128(3), 657-678 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 480, + 300 + ], + "type": "text", + "content": "20. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: ICCV. pp. 3540-3549 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 344 + ], + "type": "text", + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "text", + "content": "22. Lin, Y., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.: inerf: Inverting neural radiance fields for pose estimation. In: IROS. pp. 1323-1330 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "text", + "content": "23. Liu, J., Chen, Y., Ye, X., Qi, X.: Prior-free category-level pose estimation with implicit space transformation. CoRR abs/2303.13479 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "24. Liu, X., Wang, G., Li, Y., Ji, X.: CATRE: iterative point clouds alignment for category-level object pose refinement. In: ECCV (2). pp. 499-516 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "text", + "content": "25. Lunayach, M., Zakharov, S., Chen, D., Ambrus, R., Kira, Z., Irshad, M.Z.: FSD: fast self-supervised single RGB-D to categorical 3d objects. CoRR abs/2310.12974 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "text", + "content": "26. Marchand, E., Uchiyama, H., Spindler, F.: Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph. 22(12), 2633-2651 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "27. Murtagh, F., Legendre, P.: Ward's hierarchical agglomerative clustering method: Which algorithms implement ward's criterion? J. Classif. 31(3), 274-295 (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 482, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 480, + 515 + ], + "type": "text", + "content": "28. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: CVPR. pp. 52-61 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 514, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 480, + 547 + ], + "type": "text", + "content": "29. Peng, W., Yan, J., Wen, H., Sun, Y.: Self-supervised category-level 6d object pose estimation with deep implicit shape representation. In: AAAI. pp. 2082-2090. AAAI Press (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "type": "text", + "content": "30. Rad, M., Lepetit, V.: BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: ICCV. pp. 3848-3856 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 480, + 611 + ], + "type": "text", + "content": "31. Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.W.: Scene coordinate regression forests for camera relocalization in RGB-D images. In: CVPR. pp. 2930-2937 (2013)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "32. Song, C., Song, J., Huang, Q.: Hybridpose: 6d object pose estimation under hybrid representations. In: CVPR. pp. 428-437 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "33. Su, Y., Rambach, J.R., Minaskan, N., Lesur, P., Pagani, A., Stricker, D.: Deep multi-state object pose estimation for augmented reality assembly. In: ISMAR Adjunct. pp. 222-227. IEEE (2019)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 522 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "34. Tian, M., Ang, M.H., Lee, G.H.: Shape prior deformation for categorical 6d object pose and size estimation. In: ECCV (21). pp. 530-546 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "35. Tremblay, J., To, T., Sundaralingam, B., Xiang, Y., Fox, D., Birchfield, S.: Deep object pose estimation for semantic robotic grasping of household objects. In: CoRL. pp. 306-316 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "36. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4), 376-380 (1991)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "37. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: ICRA. pp. 10059-10066 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 480, + 248 + ], + "type": "text", + "content": "38. Wang, G., Manhardt, F., Liu, X., Ji, X., Tombari, F.: Occlusion-aware self-supervised monocular 6d object pose estimation. CoRR abs/2203.10339 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "text", + "content": "39. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: CVPR. pp. 16611-16621 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 282, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 480, + 313 + ], + "type": "text", + "content": "40. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: CVPR. pp. 2642-2651 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "text", + "content": "41. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR. pp. 803-814 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 348, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 348, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 133, + 348, + 480, + 379 + ], + "type": "text", + "content": "42. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. In: Robotics: Science and Systems (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 480, + 402 + ], + "type": "text", + "content": "43. Zakharov, S., Shugurov, I., Ilic, S.: DPOD: dense 6d pose object detector in RGB images. CoRR abs/1902.11020 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 480, + 424 + ], + "type": "text", + "content": "44. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semisupervised learning approach and A new dataset. In: NeurIPS (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 425, + 480, + 456 + ], + "type": "text", + "content": "45. Zhang, K., Fu, Y., Borse, S., Cai, H., Porikli, F., Wang, X.: Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In: ICLR. OpenReview.net (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "type": "text", + "content": "46. Zhang, R., Di, Y., Lou, Z., Manhardt, F., Tombari, F., Ji, X.: Rbp-posed: Residual bounding box projection for category-level pose estimation. In: ECCV (1). pp. 655-672 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 480, + 522 + ], + "type": "text", + "content": "47. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-posed: Hybrid scope feature extraction for category-level object pose estimation. In: CVPR. pp. 17163-17173 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Omni6D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_content_list.json b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c836ab91a1e2b96fa768b8a52bbab03e0b3687a3 --- /dev/null +++ b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_content_list.json @@ -0,0 +1,1878 @@ +[ + { + "type": "text", + "text": "OmniACT: A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web", + "text_level": 1, + "bbox": [ + 225, + 141, + 776, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Raghav Kapoor1*, Yash Parag Butala1*, Melisa Russak2, Jing Yu Koh1, Kiran Kamble2, Waseem AlShikh2, and Ruslan Salakhutdinov1", + "bbox": [ + 215, + 234, + 784, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Carnegie Mellon University", + "bbox": [ + 401, + 276, + 599, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Writer.com", + "bbox": [ + 455, + 290, + 545, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{raghavka,ypb}@cs.cmu.edu", + "bbox": [ + 401, + 306, + 602, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://huggingface.co/datasets/Writer/omniact", + "bbox": [ + 321, + 320, + 678, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb400ca699f9d9bd89b67da7f139d9ad5d2f0a68d5306cf940f4420398338896.jpg", + "image_caption": [ + "Fig. 1: OmniACT dataset and benchmark for enabling autonomous human-computer interaction agents. The left shows an image paired with a natural language task description as the input, the right shows the resulting action script to be executed on the screen. Examples are presented from Stocks, Apartments, and Weather application." + ], + "image_footnote": [], + "bbox": [ + 246, + 356, + 392, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task Description", + "text_level": 1, + "bbox": [ + 403, + 364, + 473, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Check the rate of change in Google's stock price over last one month", + "bbox": [ + 400, + 396, + 473, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Find Rental Properties in Seattle, Washington", + "bbox": [ + 411, + 468, + 460, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Using the popup opened, scroll over to find weather in Chicago on 18th September", + "bbox": [ + 403, + 537, + 475, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Action Script", + "text_level": 1, + "bbox": [ + 524, + 366, + 581, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "import pyautogui", + "bbox": [ + 501, + 395, + 568, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pyautogui.moveTo(210, 179) \npyautoguidragTo(243, 179, button=\"left\")", + "bbox": [ + 501, + 407, + 586, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "import pyautogui", + "bbox": [ + 501, + 464, + 566, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pyautogui.click(398,167) pyautogui.write(\"Seattle,WA\") pyautogui.write(\"enter\")", + "bbox": [ + 501, + 478, + 602, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "import pyautogui", + "bbox": [ + 501, + 541, + 566, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pyautogui.goiveFo(410, 207) \npyautogui.hscroll(10) \npyautogui.clclick(410, 207)", + "bbox": [ + 501, + 554, + 602, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Illustration", + "text_level": 1, + "bbox": [ + 660, + 366, + 704, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fbff3ff17c5112751dfce0ff6d8f3cb2e02d617fb6b14f2b8f659187ae31e250.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 380, + 750, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a67c4c9abf4c04345eae133c5acfb623d982b7ab80e57194b856869ab0cff89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 452, + 750, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61ca544866e92d7d2bf7f122d060d7b52459617b33e9c37ec3b129c83b8a31c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 492, + 745, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e3c7cb3e88bffc5744b780c3183e94f4fc7ca0c664132376252b5207140f516.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 522, + 750, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. For decades, human-computer interaction has fundamentally been manual. Even today, almost all productive work done on the computer necessitates human input at every step. Autonomous virtual agents represent an exciting step in automating many of these menial tasks. Virtual agents would empower users with limited technical proficiency to harness the full possibilities of computer systems. They could also enable the efficient streamlining of numerous computer tasks, ranging from calendar management to complex travel bookings, with minimal human intervention. In this paper, we introduce OmniACT, the first-of-a-kind", + "bbox": [ + 259, + 691, + 740, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally. The order is determined by dice rolling.", + "bbox": [ + 230, + 825, + 743, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dataset and benchmark for assessing an agent's capability to generate executable programs to accomplish computer tasks. Our scope extends beyond traditional web automation, covering a diverse range of desktop applications. The dataset consists of fundamental tasks such as \"Play the next song\", as well as longer horizon tasks such as \"Send an email to John Doe mentioning the time and place to meet\". Specifically, given a pair of screen image and a visually-grounded natural language task, the goal is to generate a script capable of fully executing the task. We run several strong baseline language model agents on our benchmark. The strongest baseline, GPT-4, performs the best on our benchmark. However, its performance level still reaches only $15\\%$ of the human proficiency in generating executable scripts capable of completing the task, demonstrating the challenge of our task for conventional web agents. Our benchmark provides a platform to measure and evaluate the progress of language model agents in automating computer tasks and motivates future work towards building multimodal models that bridge large language models and the visual grounding of computer screens.", + "bbox": [ + 261, + 147, + 738, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Generalist Agent $\\cdot$ Multimodal Machine Learning $\\cdot$ Dataset and Benchmark $\\cdot$ Vision Language Understanding $\\cdot$ UI grounding $\\cdot$ Human-computer interaction", + "bbox": [ + 261, + 396, + 746, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 462, + 374, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Performing computer tasks based on natural language instructions has been a long-standing goal of artificial intelligence [49]. One concrete objective in the line of research is to develop generalist agents that can assist humans in doing computer tasks [21], such as \"Order a pizza from Domino's\" or \"Write a message to John.\" The agent should be able to open the application and perform the task. Executing these actions on a personal computer involves a sequence of interactions with a mouse and keyboard. For example, the simple task of writing an email involves hovering over the application icon, clicking it, clicking the 'New Email' button, writing the content of the email, and clicking send. Successfully sending an email requires accurately predicting the correct action at each step and accurately executing it, which is a herculean task even for the best agents today [14].", + "bbox": [ + 217, + 492, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A generalist agent for computer tasks must understand natural language instructions, process visual screenshots, and produce the correct sequence of actions to be performed to achieve the intended task. Several existing approaches focus on building agents based on the HTML model [9, 40, 62]. However, this approach introduces several challenges and constraints. These agents are limited to web applications and often struggle with complex or long-context HTML code. They cannot interact with native desktop applications or perform tasks that span multiple applications, like drafting an email using text from a code editor, without significant alterations. Furthermore, HTML-based agents, which are inherently powered by text-only language models, typically underperform in tasks requiring visual cues, such as identifying and clicking a blue button", + "bbox": [ + 217, + 674, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on a desktop's top-right corner. In contrast, humans can easily understand UI elements like dropdown menus, typable areas, redirections, and options with just a glance.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Towards the goal of developing a generalist autonomous agent with robust visual and user interface (UI) understanding capabilities, we introduce a new task and dataset, OmniACT, containing over 9.8K pairs of images and instructions (Figure 1) across different operating systems and the web. This dataset includes screenshots of various UI screens and corresponding natural language instructions. The objective of these instructions is to generate executable commands using the PyAutoGUI Python library [1]. PyAutoGUI enables the automation of the mouse and keyboard operations, which helps to facilitate interactions with various native applications across macOS, Windows, and Linux. This simplifies completing specified tasks across different web domains and native desktop applications.", + "bbox": [ + 212, + 191, + 784, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate several language model-based agent baselines on this dataset, including LLaMA [47], Vicuna [7], Palmyra-X (43B) [2], InstructPalmyra-30B [45], GPT 3.5, and GPT-4 [32]. We experiment with fine-tuning Vicuna-13B and LLaMA-13B models using QLoRA [10]. We also benchmark multimodal baseline LLaVa-v1.5-7B, LLaVa-v1.5-13B [47], Gemini-Pro [44] and GPT-4-vision-preview [55] for the task. Our findings highlight the necessity for a multimodal model capable of executing these tasks, and our analysis provides insights into promising future work in the space. Our key contributions are outlined as follows:", + "bbox": [ + 212, + 357, + 785, + 479 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We release a novel dataset of desktop and website applications consisting of over 9.8K natural language tasks, UI screens, and corresponding code snippets collected through human annotation. We introduce custom performance metrics tailored for computer tasks.", + "2. We propose DetACT, a module for creating textual representations of the screen using signals from OCR, color, and icon-template matching.", + "3. We conduct a comprehensive benchmark and analysis of state-of-the-art LLMs and multimodal models on our benchmark. Our results show that OmniACT is a challenging task for even the best LLM agents today, and existing models are far below human performance." + ], + "bbox": [ + 220, + 488, + 785, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 662, + 387, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 UI Understanding", + "text_level": 1, + "bbox": [ + 214, + 693, + 413, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "User interface (UI) understanding has garnered interest from researchers in the machine learning and human-computer interaction communities, evolving with various models focusing on understanding the semantics of mobile and web user interfaces. UIBert [3], PixelBERT [16], ActionBert [15], VUT [25], Screen2Words [48], WidgetCaptioning [24] and Pix2Act [39] are notable models in this area. They propose approaches for learning the user-interface semantics of the mobile screen using the image and view hierarchy. These models have demonstrated effectiveness in tasks like capability prediction, screen segmentation and", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/8d48eb443c512036321444e8686fc6fc5c06aad51b713709dd42daba7b1694ea.jpg", + "table_caption": [ + "Table 1: Comparison of OmniACT with other related benchmarks." + ], + "table_footnote": [], + "table_body": "
DatasetsSizeEnv TypeTask HeterogeneityReal-World PortrayalExecutional CorrectnessSupports Continuous Scale Desktop AppsAdaptive EvaluationTask
VisualWebArena [20]910WebYesYesYesNoNoWeb Navigation
WebArena [62]812WebYesYesYesNoNoWeb Navigation
Mind2Web [9]2350WebYesYesNoNoNoWeb Navigation
WebShop [56]12000 ProductsWebNoNoYesNoNoWeb Navigation
RUSS [53]80WebYesYesNoNoNoWeb Navigation
WebSRC [6]2735WebYesYes-NoNoQA
MiniWoB++ [17]100Mobile WebsitesNoNoYesNoNoWeb Navigation
PixelHelp [23]187MobileYesYesNoNoNoUI Grounding
MetaGUI [42]1125MobileYesYesYesNoNoMobile Navigation
MoTIF [5]756MobileYesYesYesNoNoMobile Navigation
AITW [35]715142Mobile and WebYesYesYesNoNoMobile/Web Navigation
OmniACT (Ours)9802Desktop and WebYesYesYesYesYesCode Generation
", + "bbox": [ + 215, + 171, + 785, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "understanding, and screen caption generation. Lexi [4] and Spotlight [22] propose models that use vision-only inputs to minimize the reliance on metadata such as view hierarchy. Furata et al. [11] demonstrates the use of fine-tuning for multimodal web navigation. The majority of machine learning models trained for UI understanding leverage the Rico dataset [8] and its extensions, which contain 64,462 unique Android screens and metadata. In addition, [4] released the UICaptions dataset, which consists of diverse image-captions pairs across a wide range of applications. PixelHelp [23] also released a corpus to train models that can interpret natural language instructions and map them to mobile UI actions.", + "bbox": [ + 212, + 344, + 784, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Autonomous Computer Agents", + "text_level": 1, + "bbox": [ + 215, + 500, + 517, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The advent of large language models (LLMs) has been pivotal in the rapid advancement of agents that operate on web pages. Recent research such as ViperGPT [43] Chameleon [29], RCI Agent [18], VisProg [12], and [31] employ LLMs for planning or action prediction in developing autonomous agents. Benchmark datasets, such as MiniWoB [40], WebShop [56], Macaw-LLM [30], ASH-Prompting [41] Mind2Web [9], WebArena [62], AgentBench [28] and VisualWebArena [20] have also been proposed to measure the ability of LLM-based agents to automate web tasks. These methods mainly involve agents that operate on a text-based Document Object Model (DOM) of HTML scripts. This limits their understanding of screen context, which is crucial for the model's decision-making and action-taking processes. To address this limitation, [35] released Android in the Wild, a dataset comprising screens, natural language instructions, and corresponding actions. Following this, [59] proposed a multimodal model, AutoUI, which is designed to build an agent on the Android in the Wild dataset confined to the Android ecosystem. WebAgent [13] utilized Flan-U-PaLM, for grounded code generation, and HTML-T5 and showed improvement on real-world websites.", + "bbox": [ + 212, + 522, + 785, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current benchmarks for autonomous agents focus mainly on the Web or Android environments, posing challenges for tasks involving desktop applications or spanning multiple applications beyond the web domain. The absence of established benchmarks and datasets in this area, coupled with basic methods", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 455, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for extracting user interface (UI) elements, underscores the need for significant progress in developing more versatile autonomous agents capable of handling diverse tasks beyond the current scope. To highlight the unique features that OmniACT introduces in the assessment of capable autonomous agents, we provide a comparison between the existing benchmarks and our proposed benchmark, OmniACT, in Table 1.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3637b64b82edc539bc8230362a244c2b1b77bf9d9c6277baaa4ab34339e3319e.jpg", + "image_caption": [ + "Fig. 2: Data Collection Pipeline. (1) We select over 60 applications and websites to ensure diversity, (2) segment the screen through human-annotated bounding boxes, (3) label the bounding boxes based on functionality, (4) ask student volunteers to come up with tasks, given a screen image, and (5) reverse map the textual labels to coordinates and filter the scripts based on execution and syntax." + ], + "image_footnote": [], + "bbox": [ + 245, + 262, + 759, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 OmniACT", + "text_level": 1, + "bbox": [ + 215, + 599, + 354, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a novel dataset and benchmark, OmniACT, which measures the performance of autonomous agents on both web and desktop applications. Compared to previous benchmarks which focus on text-based reasoning [9, 17, 40, 56, 62], our benchmark aims to measure multimodal agents that bridge large language model planners and UI understanding vision models. OmniACT can be accomplished as a standalone task as it is not under a mock environment.", + "bbox": [ + 212, + 628, + 787, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All actions that a human can execute on the computer can be encoded in the PyAutoGUI [1] Python framework. This framework allows a user to execute keyboard and mouse operations by running Python code. The PyAutoGUI code to execute these tasks is shown in the third column of Figure 1. For other computer tasks, the PyAutoGUI library provides functions such as 'press', 'write', and 'scroll' which can be used to execute the task. Our dataset consists of parallel data of natural language tasks, UI screenshots, and ground truth PyAutoGUI scripts that achieve successful execution.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Task Formulation", + "text_level": 1, + "bbox": [ + 215, + 146, + 405, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an input state of a computer defined by the screen $S$ and the task description $T$ in natural language, the goal of the task is to output a sequence of actions $A$ that can successfully accomplish the task $T$ within a screenshot $S \\in \\{ \\text{Linux, Windows, MacOS, Webpage} \\}$ . Formally, the task can be defined as learning the transition function $f: T \\times S \\to A$ . During dataset collection, we ensure that all task descriptions $T$ are feasible and can be accomplished in the current screenshot $S$ . To reduce ambiguity and facilitate better evaluation, we ensure that task descriptions are detailed and unambiguous. Tasks can also be visually grounded (e.g., 'Click the red button to start recording') or natural language based (e.g., 'Click the My Account button'). We define the action space using the functionalities in the PyAutoGUI library: $A \\in \\{\\text{'click', 'dragTo', 'scroll', 'write', \\ldots}\\}$ . The exhaustive list of actions is provided in Table 2. Our action space is much larger than other benchmarks [9,40,62] that resort to two or three interaction options. Mouse actions such as 'moveTo', 'click', 'rightClick', 'doubleClick', and 'dragTo', additionally require screen coordinates as arguments, which indicate the pixel location of the action.", + "bbox": [ + 212, + 169, + 787, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 1 illustrates sample tasks and corresponding outputs for three applications within OmniACT: (1) Stocks (MacOS), (2) Apartments.com (web page), and (3) Weather (MacOS). The first column depicts the input image, and the second column shows the natural language task that is to be executed on the current screen. To execute these tasks, a user must accurately perform a series of operations using the mouse and keyboard. Eg: to check the rate of change in Google's stock price over the last month, the mouse has to be moved to the last month and dragged while holding the left-click button to the current month.", + "bbox": [ + 212, + 412, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Dataset Preparation", + "text_level": 1, + "bbox": [ + 215, + 553, + 429, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To prepare our dataset, we followed a pipelined approach, as summarized in Figure 2. We first selected a variety of applications and websites. For each application or website, we created bounding boxes around key UI elements and labeled them according to their functionality, which is crucial for assisting human annotators in writing accurate PyAutoGUI scripts. After each script is written, we converted the labels back into numeric coordinates, allowing us to align the scripts precisely with the locations of the UI elements. Finally, we thoroughly reviewed each script, focusing on its executability and adherence to syntax standards. This ensured the high quality and functionality of our dataset, making it a valuable resource for training and evaluating autonomous agents.", + "bbox": [ + 212, + 577, + 787, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Application/Website Selection To test the computer agents' generalization ability across different tasks, we collect tasks across multiple domains on both desktop and web applications. In total, we collect and annotate 9802 data points (Table 3), with the split between desktop and web applications approximately 3:1. The emphasis on desktop applications, which do not contain Document Object Model (DOM) hierarchies unlike HTML-based web pages, presents a more", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "complex multimodal challenge where visual cues are crucial. We collect tasks from applications within the three most popular operating systems. We select 22 native applications from MacOS, and 8 each from Linux and Windows. We annotate roughly 3 to 4 screens for every application. The full list of applications is provided in the Appendix.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Many common computer tasks today are still performed through web applications, so we also collect 3-4 screenshots from 27 different web applications. To ensure diversity in task intents, we categorize these tasks into one of the following 6 categories: (1) Shopping, (2) Entertainment, (3) Service, (4) Government, (5) Travel, (6) Health. Inspired by the methodology of [9], these categories were selected to cover a wide range of user intents and functionalities.", + "bbox": [ + 212, + 222, + 514, + 402 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/efd3839d384fbecee73b13876012435c71bbd9d96d15ee0c14640fba843ab165.jpg", + "table_caption": [ + "Table 2: Action types supported by OmniACT and the number of instances for each action in the dataset." + ], + "table_footnote": [], + "table_body": "
TypeAction%
MouseClick63.73
Double Click0.58
Right Click0.77
Move/Hover1.85
Drag0.29
Scroll1.68
Horizontal Scroll0.17
KeyboardPress16.28
Hotkey3.00
Write11.65
", + "bbox": [ + 583, + 309, + 728, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UI Screen Segmentation To collect gold-standard data, we first annotate and segment the screen by identifying the", + "bbox": [ + 212, + 425, + 514, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bounding boxes present on the screen. We employ slightly different techniques for web and desktop applications to create the bounding boxes:", + "bbox": [ + 212, + 470, + 782, + 501 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Desktop Applications: We build a custom annotation interface based on PyQt5 $^3$ to create bounding boxes manually over a screen image using a simple drag-and-click mechanism. This custom interface expedites the process and allows us to get highly accurate gold-label data for desktop images.", + "2. Websites: For webpages, we write JavaScript code to extract all interactable (click, type, etc.) regions from HTML source code. We also extract banners, dropowns, submit, and radio buttons from the screen. We filter the elements to retain only those that are visible and interactable within the screen." + ], + "bbox": [ + 220, + 513, + 782, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Functionality Tagging To map each bounding box to its correct functional description, we leverage Amazon MTurk workers (see details in Appendix), who are given an image with a bounding box and are required to write the correct description or label of the bounding box's function. For example, given an image of an Amazon webpage with a search bar, the annotator labels it as \"find-product-search-bar\". The logical descriptions are used to create tasks in a structured manner without the need to identify individual bounding box coordinates.", + "bbox": [ + 212, + 657, + 782, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task Creation Our approach for each screen involves utilizing all human-annotated bounding boxes and their labels to create tasks that can be executed", + "bbox": [ + 212, + 785, + 782, + 816 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3 https://pypi.org/project/PyQt5/", + "bbox": [ + 217, + 823, + 472, + 839 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91bc6bf190024aeb93b022e0af17e277beeacd4e3c1c9b2f438622bc2f6c41f0.jpg", + "image_caption": [ + "Fig. 3: DetACT Module. Given an initial image and a natural language task description, we use a pipelined approach to run OCR and SAM on the screen. The outputs from SAM are then used by icon and color-matching modules to obtain an exhaustive set of useful UI elements. The list of elements is passed through LLM based filter to select only the elements related to the given task." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 782, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "within the confines of a single screen. These tasks are designed to be visually grounded in order to measure the capabilities of multimodal agents. We plan to release the bounding box and their corresponding labels as the metadata for evaluation purposes.", + "bbox": [ + 212, + 436, + 784, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For dataset compilation, college students with basic Python programming skills served as annotators, accessing API references for PyAutoGUI and examples of potential tasks. Each student generated multiple tasks, each accompanied by three alternative natural language reformulations. For instance, \"What is $3 + 2$ ?\" might be reformulated as \"Calculate the sum of 2 and $3$ \" or \"Add two to three\". To avoid train-test leakage, rephrased tasks were consistently placed in the same dataset split. Further details on the annotation process are available in the Appendix.", + "bbox": [ + 212, + 498, + 787, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reverse Mapping and Filtering To ensure high-quality data, we incorporate an additional step into the data collection pipeline. We build scripts to map the text-based labels of each bounding box back to their numeric coordinates, and then match the syntax and verify if the task will be executed on the screen. Using this filter, we remove all the non-working or syntactically incorrect data points and finally manually review the set of tasks.", + "bbox": [ + 212, + 642, + 517, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After filtering, we obtain 9802 human-annotated, gold-label data points across more than 200 desktop and web screens", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3ff6e279b9f9ca02e4e290ee118f367f93872ab2d707191e6b54439e215e34e7.jpg", + "table_caption": [ + "Table 3: Dataset distribution across splits and platforms." + ], + "table_footnote": [], + "table_body": "
DomainTrain ValidationTestTotal
DesktopMac OS3028444786
Linux761126247
Windows1573216458
Web-1427206530
Total67899922,021
", + "bbox": [ + 537, + 715, + 771, + 789 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 455, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(Table 3), split into train, validation, and test sets in a 7:1:2 ratio. All collected data will be publicly released to encourage future work on multimodal agents.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 199, + 434, + 215 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we detail various evaluation metrics for benchmarking model performance on the OmniACT dataset. UI screens have additional constraints such as spatial relevance which are not factored in most conventional similarity-based metrics such as BLEU [34], CodeBLEU [36], BERTScore [58] and CodeBERTScore [61]. For example, a valid click action is usually not constrained to a single coordinate but can be any coordinate within a specified region. In the event of invalid coordinate predictions, an agent that predicts coordinates further away from the valid region should invoke a higher penalty compared to an agent that predicted coordinates close to the region. We propose two new metrics adapted: Sequence Score (Section 4.1) and Action Score (Section 4.2) aimed at utilizing UI information.", + "bbox": [ + 212, + 231, + 782, + 396 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Sequence Score", + "text_level": 1, + "bbox": [ + 215, + 419, + 388, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The sequence score measures whether the predicted action sequence (e.g., 'click', 'write', 'press') exactly matches the gold sequence. Since predicting the first action in the sequence is relatively straightforward and later actions are more difficult, we define sequence score as follows:", + "bbox": [ + 212, + 444, + 782, + 503 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S e q S c o r e} _ {i} = \\left\\{ \\begin{array}{l l} \\beta_ {1} + \\beta_ {2} * (s - 1) & \\text {i f a l l a c t i o n s m a t c h} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 507, + 784, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $s$ is the action sequence length, $\\beta_{1}$ is set to 0.1 and $\\beta_{2}$ is set to 1.", + "bbox": [ + 215, + 547, + 732, + 563 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Action Score", + "text_level": 1, + "bbox": [ + 215, + 585, + 369, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The action score measures how well a code snippet containing the correct action sequence can perform the task. Specifically, for a script with a correct action sequence, we introduce penalties for inaccurate behavior. The penalties are described below:", + "bbox": [ + 212, + 609, + 782, + 670 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "1. Click penalty $(M)$ : For actions 'click', 'rightClick', 'doubleClick', 'moveTo', and 'dragTo', we penalize code snippets where predicted coordinates lie outside of the bounding box of the UI element. The click penalty for the $j^{th}$ action of the $i^{th}$ example is defined as:", + "bbox": [ + 222, + 681, + 782, + 742 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 1 - \\frac {\\mu}{\\mu + L _ {2}} & \\text {i f S e q S c o r e} _ {i} > 0 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 744, + 784, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here $L_{2}$ corresponds to the smallest Euclidean distance between the predicted coordinate and bounding box. $L_{2}$ is zero when the predicted coordinate lies within the target bounding box. $\\mu$ is the Dirichlet smoothing", + "bbox": [ + 241, + 794, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "coefficient which we dynamically set to the inverse of the length of the diagonal of the bounding box. This ensures that the penalty for points outside the bounding box varies based on the size of the bounding box. For two predicted points with the same $L_{2}$ , the metric penalizes more heavily if the box is larger. This is sound with the intuition that the chances of clicking on a larger box are higher and should be penalized more in case of a mistake.", + "bbox": [ + 240, + 146, + 787, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "2. Key penalty $(K)$ : For actions 'press' and 'hotkey', we check whether the set of keys in the target code (represented as $GK_{i}^{j}$ ) and predicted code (represented as $PK_{i}^{j}$ ) are the same. It is formally defined as:", + "bbox": [ + 222, + 237, + 787, + 284 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nK _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 0 & \\text {i f} G K _ {i} ^ {j} = P K _ {i} ^ {j} \\text {a n d S e q S c o r e} _ {i} > 0 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 287, + 785, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3. Write penalty $(W_{p})$ : For action type 'write', we penalize the output for the sentence to be typed. Specifically, we the employ BLEU score [34], and compute:", + "bbox": [ + 222, + 329, + 785, + 375 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 1 - B L E U \\left(G S _ {i} ^ {j}, P S _ {i} ^ {j}\\right) & \\text {i f S e q S c o r e} _ {i} > 1 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 377, + 785, + 416 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Here, $GS_{i}^{j}$ represents the actual sentence to be typed, and $PS_{i}^{j}$ represents the sentence predicted by the model in the $j^{th}$ action of example $i$ .", + "bbox": [ + 241, + 420, + 785, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the above equations, $(\\alpha_{i})$ is the weighting factor:", + "bbox": [ + 215, + 463, + 589, + 478 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\operatorname {S e q S c o r e} _ {i} / \\text {l e n g t h o f s e q u e n c e} i \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 481, + 785, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This ensures that the action score $\\in [0,1]$ . The mean action score is calculated as follows:", + "bbox": [ + 212, + 501, + 785, + 529 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\text {A c t i o n S c o r e} = \\frac {\\sum_ {i} \\max \\left(\\operatorname {S e q S c o r e} _ {i} - \\sum_ {j} \\left(M _ {i} ^ {j} + K _ {i} ^ {j} + W _ {i} ^ {j}\\right) , 0\\right)}{\\sum_ {i} \\operatorname {S e q S c o r e} _ {i}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 531, + 785, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 DetACT: DETecting ACTions from UI", + "text_level": 1, + "bbox": [ + 215, + 592, + 637, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Understanding UI screens is crucial for multimodal computer tasks. Web-based agents typically use language-only inputs from the HTML DOM. This is insufficient for comprehending the full extent of an application UI, as many components may not be easily described with HTML code. To address this, we propose DetACT, which allows us to convert images of UI layouts into structured code and text outputs for a downstream LLM. DetACT is a system comprised of three distinct modules: the text module, the icon module, and the color module.", + "bbox": [ + 212, + 623, + 787, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "1. Text Extraction: We use the EasyOCR model4 to parse over the UI screens and collect all text-based elements. Along with the text, we also note the locations of each of these elements. This is depicted in Figure 3, along with a list of text elements found on the screen using the OCR Module. We segment and classify the different regions within the screenshot using the", + "bbox": [ + 222, + 739, + 787, + 818 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 455, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/JaidedAI/EasyOCR", + "bbox": [ + 218, + 824, + 506, + 840 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/877231d43634aee2879a1c206ac0fe7070408c08015ff5c3ce1a8c4b850a6958.jpg", + "image_caption": [ + "Fig. 4: Baseline Model Architecture. Image and task descriptions are sent to DetACT module, which gives a filtered list of UI elements relevant to feed into the prompt along with the task. We also show the prompt structure used for action script generation. This structure is passed through the LLM (along with the image for multimodal LLM) to generate the automation script." + ], + "image_footnote": [], + "bbox": [ + 334, + 145, + 669, + 406 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Segment Anything Model (SAM) [19]. From the outputs, we filter out the non-textual segments for our icon and color detection.", + "bbox": [ + 240, + 522, + 784, + 551 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Icon Module: For matching with the appropriate icon, we use a pack of 1600 icons5 as templates. Each of these icons is labeled with their appropriate functionality and is matched with the filtered outputs SAM [19]. For the similarity of the two images, we resize the reference icons and segmented region of interest (ROI) to the same size, and convert both images to grayscale. After this, we use the Structural Similarity Index (SSIM) [52], to find the closest match of the ROI to the icons in our set, and select the ones above the SSIM threshold of 0.95. As seen in Figure 3, a few icons matched on the screen are Globe icon, Calendar icon, Person icon, and Location icon; each depicting a different use case.", + "3. Color Module: Finally, to place all segments of interest into appropriate buckets of colors, we average the RGB pixel values over the ROI and, based on that value, bucket them into different color categories. We categorize colors differently based on the human perspective of the ranges of each color. To avoid ambiguity, we consider eleven major colors, namely yellow, blue, green, red, pink, violet, white, black, orange, brown, and grey. We record the center of the element along with the color." + ], + "bbox": [ + 222, + 554, + 787, + 811 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "5 https://icomoon.io/", + "bbox": [ + 220, + 824, + 382, + 839 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Once all the elements of each category are extracted with their coordinates, we then filter these UI elements by prompting GPT-4 [32]. We ensure that the elements selected are suited only for our task, for which we also provide the task description in our prompts along with the list of elements. Full details of the prompt are provided in the appendix section of the paper. As we observe in Figure 3, given an image from the Expedia application, and a task (\"Click on the Black Location icon and enter the destination as Paris\"), the LLM filters out the elements to retain only \"Going To\", \"Location Icon\", and the Black colored elements from the screen. This is passed as input to the LLM or VLM backbone.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6Baselines", + "text_level": 1, + "bbox": [ + 215, + 304, + 341, + 320 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To evaluate the performance of existing language model-based agents, we conduct experiments with both language-based and multimodal baselines. The DetACT module takes in image and text descriptions of the task and outputs the color, icon, and text-based signals. This is concatenated to the prompt for the LLM prompt-based baselines (see Figure 4). Every prompt starts with a role assignment [60], followed by the detailed API reference of the PyAutoGUI function set, along with a textual description of their function. We then add five in-context examples from the training set that most closely match the task (based on the cosine similarity of the MiniLM [50] embeddings of the reference task and the train examples). We add a list of UI elements filtered by the DetACT module to the prompt. Finally, we provide the rules with the task description. For multimodal baselines, we also pass the image pixels to the vision encoder. We choose coordinate-based UI elements in the prompt as recent techniques like the Set-of-Mark (SOM) [54] prompting does not work for desktop settings since it is difficult to obtain interactive elements from the desktop screen images. We report the results of several baselines:", + "bbox": [ + 212, + 334, + 787, + 575 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Few-shot Generative LLM: We experiment with models from LLaMA-2 [47], Vicuna-1.5 [7], CodeLLaMA-34B [37], Palmyra [46], and GPT [32] series. We use the prompts structure as shown in Figure 4 to prompt the model. For LLaMA and CodeLLaMa, we reduce the prompt length to 2000 tokens by removing outputs from the DetACT module with lower confidence, as we observed poor performance on longer prompts. For the other models, we allow prompts with up to 4000 token sizes.", + "- Finetuned Generative LLM: We fine-tuned the LLaMA-13B model and Vicuna-13B using QLoRa [10] with rank 64 and scaling factor 16 for 300 steps to generate the code given screen description from the DetACT module and the instruction.", + "- Few-shot Generative Multimodal Models: As OmniACT is predominantly multimodal, with a majority of tasks being visually grounded, we conduct experiments with large multimodal models. Given the limited research in this domain [51, 57], there is a scarcity of available multimodal models with significant size adept for this task. Here, we experiment with [26, 27], providing a similar prompt as well as the screen image." + ], + "bbox": [ + 225, + 584, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "7 Results and Analysis", + "text_level": 1, + "bbox": [ + 215, + 143, + 454, + 162 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Table 4, we experiment with three different categories of models, namely Prompt-based LLMs, Fine-tuned LLMs, and Prompt-based Multimodal Models. GPT-4 is the best-performing approach, scoring higher on the sequence score and invoking lower penalties on coordinate predicting and text input. For prompt-only LLMs, the GPT-3.5-turbo and GPT-4 models outperform the other LLM baselines, including the LLaMA [47] and Vicuna [7] models. We observe that CodeLLaMA-34B [38], which is trained for code generation, also achieves a higher performance than other models of the same size at predicting the action sequences.", + "bbox": [ + 212, + 175, + 514, + 430 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fine-tuned models also perform much better than their few-shot prompt-only counterparts. Fine-tuning substantially improves LLaMA-13B's sequence score (4.80 to 8.92) and action score (1.62 to 2.14), as well as the other metrics. Despite this, we observed that both, prompt-based LLMs and finetuned LLMs face severe mouse penalties, especially on click coordinates. This is because they rely solely on text-based signals.", + "bbox": [ + 212, + 431, + 514, + 598 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0eda75a4d05cbc94fa4e0a4e8b0ed997ce4e639ad39c5f84ffd4f568682a7b24.jpg", + "table_caption": [ + "Table 4: Baseline Performance. (A) Prompt-only LLMs, (B) Fine Tuned LLMs, (C) Prompt-only Multimodal Models. The table represents the Sequence score (SS), click penalty $(M_p)$ , Key penalty $(K_p)$ , Write Penalty $(W_p)$ , and Action Score (AS). The best results for the (SS) and (AS) are highlighted." + ], + "table_footnote": [], + "table_body": "
ModelSS(↑)MpKpWpAS(↑)
Prompt based LLMs
LLaMA-7B [47]4.121.241.830.570.48
Vicuna-7B [7]3.881.171.510.430.77
LLaMA-13B [47]4.801.320.930.931.62
Vicuna-13B [7]5.441.650.941.061.78
Palmyra-Instruct-30B [45]7.515.680.120.401.31
CodeLLaMA-34B [38]10.092.992.710.663.72
Palmyra-X 43B [2]11.203.123.022.122.94
GPT-3.5-turbo-0613 [33]22.858.134.512.317.89
GPT-4 [32]32.7510.276.993.8911.60
Finetuned LLMs
LLaMA-13B FT8.924.611.430.742.14
Vicuna-13B FT8.784.121.310.632.72
Multimodal LLMs
LLaVA-v1.5-7B [27]13.234.731.241.445.82
LLaVA-v1.5-13B [26]20.566.073.442.858.19
Gemini-Pro [44]30.989.056.813.6611.46
GPT-4V [26]38.7210.537.144.0317.02
Human Performance82.230.120.361.6180.14
", + "bbox": [ + 526, + 332, + 785, + 549 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To address this, we experiment with multimodal language models (Table 4). We observe that the coordinate prediction improves significantly when we provide the entire image as input to the multimodal LLM, as this enables it to fully utilize the screen representation. In addition to open sourced models, we also experiment with the GPT-4-vision API [55] which shows that GPT-4 Vision [55] outperforms GPT-4 significantly on the Action Score along with improving the sequence score, which we attribute to the strong reasoning abilities of GPT-4 coupled with the improved visual understanding capabilities of the GPT-4-vision model [55]. These findings pave the way towards exciting new research directions on building multimodal models for long-horizon planning and code generation.", + "bbox": [ + 212, + 598, + 787, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Human performance over the task: OmniACT consists of visually complicated tasks, and tests various types of computer skills. In order to get a gauge of how well humans perform, we collect evaluation data from human evaluators. We split the test set uniformly amongst 10 human evaluators, and provided them with the screenshot and task instruction. We record the actions taken by the an-", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "notators, and measure their performance on our predefined metrics (Table 4). We find that users generally exhibit a high level of proficiency when attempting most tasks for the first time. However, there are instances where users face difficulties in successfully completing certain tasks. These are due to factors including the user's inability to fully comprehend the task, difficulties in grounding the task to the provided screenshot, or a lack of familiarity with the UI.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 214, + 257, + 532, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Autonomous virtual agents offer the potential to automate routine tasks, benefiting users with limited technical expertise. To solve this task, we introduce OmniACT, a unique dataset of 9.8K human-labeled data points. OmniACT benchmarks autonomous agents across a range of tasks on web and desktop applications. LLM-based agents, like GPT-4, achieve a respectable action score of 11.6 on our dataset. However, OmniACT presents a challenge for the current state-of-the-art language and multimodal models. It provides a direction for future research on foundational multimodal models that seamlessly integrate language and visual understanding of computer screens and stands poised to drive the next wave of advancements in generalist autonomous agents offering omnipotent assistance to humans.", + "bbox": [ + 212, + 287, + 787, + 454 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "9 Limitations", + "text_level": 1, + "bbox": [ + 214, + 473, + 362, + 489 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work introduces a valuable dataset, yet we recognize a few limitations that exist. State-of-the-art models like GPT-4, may exhibit susceptibility to hallucinations and bias towards specific data types, hindering broad applicability. Reliance on closed models like GPT-4V poses integration challenges due to high costs and time constraints. Despite efforts for equal representation and data collection without personal information, biases may be introduced as the dataset is exclusively in English, and human-curated content may have temporal biases.", + "bbox": [ + 212, + 503, + 787, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "10 Ethics Statement", + "text_level": 1, + "bbox": [ + 214, + 628, + 429, + 646 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As a part of the dataset creation process, we carefully review the pipeline at every stage, ensuring there is no personally identifiable information or offensive content, either during data collection or through the use of LLMs. For all purposes, we create dummy accounts that mimic real-like user content. To get the gold labels scripts we seek help from well-qualified student workers, approved through the institution, and get the bounding box data annotated through MTurk workers, both of whom are paid $25 per hour, which is greater than the minimum wage rate (We detail this process in the supplementary material). Human studies are also done with the help of student workers approved by the institution at the above-mentioned payscale. We also ensure that all groups have equitable representation and that no personal opinions are reflected in the dataset, avoiding bias during the collection as well as the annotation process.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We extend our heartfelt gratitude to Writer.com for their generous and unwavering support throughout this project. Their dedicated team's expertise and collaboration were invaluable in achieving our goals.", + "bbox": [ + 215, + 178, + 784, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 248, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Pyautogui: A cross-platform gui automation python module for human beings. https://github.com/asweigart/pyautogui (2023)", + "2. AlShikh, W., Daaboul, M., Goddard, K., Imel, B., Kamble, K., Kulkarni, P., Russak, M.: Becoming self-instruct: introducing early stopping criteria for minimal instruct tuning (2023)", + "3. Bai, C., Zang, X., Xu, Y., Sunkara, S., Rastogi, A., Chen, J., y Arcas, B.A.: Uibert: Learning generic multimodal representations for ui understanding (2021)", + "4. Banerjee, P., Mahajan, S., Arora, K., Baral, C., Riva, O.: Lexi: Self-supervised learning of the ui language (2023)", + "5. Burns, A., Arsan, D., Agrawal, S., Kumar, R., Saenko, K., Plummer, B.A.: Mobile app tasks with iterative feedback (motif): Addressing task feasibility in interactive visual environments", + "6. Chen, X., Zhao, Z., Chen, L., Ji, J., Zhang, D., Luo, A., Xiong, Y., Yu, K.: Websrc: A dataset for web-based structural reading comprehension. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 4173-4185 (2021)", + "7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/", + "8. Deka, B., Huang, Z., Franzen, C., Hibschman, J., Afergan, D., Li, Y., Nichols, J., Kumar, R.: Rico: A mobile app dataset for building data-driven design applications. In: Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology. p. 845-854. UIST '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3126594.3126651, https://doi.org/10.1145/3126594.3126651", + "9. Deng, X., Gu, Y., Zheng, B., Chen, S., Stevens, S., Wang, B., Sun, H., Su, Y.: Mind2web: Towards a generalist agent for the web. arXiv preprint arXiv:2306.06070 (2023)", + "10. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms (2023)", + "11. Furuta, H., Nachum, O., Lee, K.H., Matsuo, Y., Gu, S.S., Gur, I.: Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854 (2023)", + "12. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023)", + "13. Gur, I., Furuta, H., Huang, A., Safdari, M., Matsuo, Y., Eck, D., Faust, A.: A real-world webagent with planning, long context understanding, and program synthesis (2024)" + ], + "bbox": [ + 218, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Gur, I., Rueckert, U., Faust, A., Hakkani-Tur, D.: Learning to navigate the web. In: International Conference on Learning Representations (2018)", + "15. He, Z., Sunkara, S., Zang, X., Xu, Y., Liu, L., Wichers, N., Schubiner, G., Lee, R., Chen, J., y Arcas, B.A.: Actionbert: Leveraging user actions for semantic understanding of user interfaces (2021)", + "16. Huang, Z., Zeng, Z., Liu, B., Fu, D., Fu, J.: Pixel-bert: Aligning image pixels with text by deep multi-modal transformers (2020)", + "17. Humphreys, P.C., Raposo, D., Pohlen, T., Thornton, G., Chhaparia, R., Muldal, A., Abramson, J., Georgiev, P., Santoro, A., Lillicrap, T.: A data-driven approach for learning to control computers. In: International Conference on Machine Learning. pp. 9466-9482. PMLR (2022)", + "18. Kim, G., Baldi, P., McAleer, S.: Language models can solve computer tasks. arXiv preprint arXiv:2303.17491 (2023)", + "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything (2023)", + "20. Koh, J.Y., Lo, R., Jang, L., Duvvur, V., Lim, M.C., Huang, P.Y., Neubig, G., Zhou, S., Salakhutdinov, R., Fried, D.: Visualwebarena: Evaluating multimodal agents on realistic visual web tasks (2024)", + "21. LeCun, Y.: A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27 (2022)", + "22. Li, G., Li, Y.: Spotlight: Mobile ui understanding using vision-language models with a focus (2023)", + "23. Li, Y., He, J., Zhou, X., Zhang, Y., Baldridge, J.: Mapping natural language instructions to mobile ui action sequences (2020)", + "24. Li, Y., Li, G., He, L., Zheng, J., Li, H., Guan, Z.: Widget captioning: Generating natural language description for mobile user interface elements (2020)", + "25. Li, Y., Li, G., Zhou, X., Dehghani, M., Gritsenko, A.: Vut: Versatile ui transformer for multi-modal multi-task user interface modeling (2021)", + "26. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)", + "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)", + "28. Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., Zhang, S., Deng, X., Zeng, A., Du, Z., Zhang, C., Shen, S., Zhang, T., Su, Y., Sun, H., Huang, M., Dong, Y., Tang, J.: Agentbench: Evaluating llms as agents (2023)", + "29. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models (2023)", + "30. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023)", + "31. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback", + "32. OpenAI: Gpt-4 technical report (2023)", + "33. OpenAI: Introducing chatgpt (2023), https://openai.com/blog/chatgpt", + "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 454, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Rawles, C., Li, A., Rodriguez, D., Riva, O., Lillicrap, T.: Android in the wild: A large-scale dataset for android device control (2023)", + "36. Ren, S., Guo, D., Lu, S., Zhou, L., Liu, S., Tang, D., Sundaresan, N., Zhou, M., Blanco, A., Ma, S.: Codebleu: a method for automatic evaluation of code synthesis. arXiv preprint arXiv:2009.10297 (2020)", + "37. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023)", + "38. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023)", + "39. Shaw, P., Joshi, M., Cohan, J., Berant, J., Pasupat, P., Hu, H., Khandelwal, U., Lee, K., Toutanova, K.: From pixels to ui actions: Learning to follow instructions via graphical user interfaces. arXiv preprint arXiv:2306.00245 (2023)", + "40. Shi, T., Karpathy, A., Fan, L., Hernandez, J., Liang, P.: World of bits: An open-domain platform for web-based agents. In: International Conference on Machine Learning, pp. 3135-3144. PMLR (2017)", + "41. Sridhar, A., Lo, R., Xu, F.F., Zhu, H., Zhou, S.: Hierarchical prompting assists large language model on web navigation. arXiv preprint arXiv:2305.14257 (2023)", + "42. Sun, L., Chen, X., Chen, L., Dai, T., Zhu, Z., Yu, K.: Meta-gui: Towards multimodal conversational agents on mobile gui. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 6699-6712 (2022)", + "43. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning (2023)", + "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "45. team, W.E.: InstructPalmyra-30b : Instruct tuned Palmyra-Large model. https://dev.writer.com (2023)", + "46. team, W.E.: Palmyra-base Parameter Autoregressive Language Model. https://dev.writer.com (2023)", + "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)", + "48. Wang, B., Li, G., Zhou, X., Chen, Z., Grossman, T., Li, Y.: Screen2words: Automatic mobile ui summarization with multimodal learning. In: The 34th Annual ACM Symposium on User Interface Software and Technology. pp. 498-510 (2021)", + "49. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al.: A survey on large language model based autonomous agents. arXiv preprint arXiv:2308.11432 (2023)", + "50. Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems 33, 5776-5788 (2020)", + "51. Wang, X., Chen, G., Qian, G., Gao, P., Wei, X.Y., Wang, Y., Tian, Y., Gao, W.: Large-scale multi-modal pre-trained models: A comprehensive survey (2023)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OmniACT", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "53. Xu, N., Masling, S., Du, M., Campagna, G., Heck, L., Landay, J., Lam, M.: Grounding open-domain instructions to automate web support tasks. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 1022-1032 (2021)", + "54. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v (2023)", + "55. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)", + "56. Yao, S., Chen, H., Yang, J., Narasimhan, K.: Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems 35, 20744-20757 (2022)", + "57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models (2023)", + "58. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert (2020)", + "59. Zhang, Z., Zhang, A.: You only look at screens: Multimodal chain-of-action agents (2023)", + "60. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)", + "61. Zhou, S., Alon, U., Agarwal, S., Neubig, G.: Codebertscore: Evaluating code generation with pretrained models of code. arXiv preprint arXiv:2302.05527 (2023)", + "62. Zhou, S., Xu, F.F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Bisk, Y., Fried, D., Alon, U., et al.: Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 549 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Kapoor, Y. Butala et al.", + "bbox": [ + 271, + 114, + 454, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_model.json b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c11da46031d668891e383ce2197db287ffbdce4 --- /dev/null +++ b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_model.json @@ -0,0 +1,2634 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.226, + 0.142, + 0.777, + 0.207 + ], + "angle": 0, + "content": "OmniACT: A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.235, + 0.785, + 0.266 + ], + "angle": 0, + "content": "Raghav Kapoor1*, Yash Parag Butala1*, Melisa Russak2, Jing Yu Koh1, Kiran Kamble2, Waseem AlShikh2, and Ruslan Salakhutdinov1" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.277, + 0.601, + 0.291 + ], + "angle": 0, + "content": "1 Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.456, + 0.291, + 0.547, + 0.305 + ], + "angle": 0, + "content": "\\(^{2}\\) Writer.com" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.307, + 0.603, + 0.32 + ], + "angle": 0, + "content": "{raghavka,ypb}@cs.cmu.edu" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.321, + 0.679, + 0.334 + ], + "angle": 0, + "content": "https://huggingface.co/datasets/Writer/omniact" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.357, + 0.393, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.405, + 0.366, + 0.475, + 0.377 + ], + "angle": 0, + "content": "Task Description" + }, + { + "type": "text", + "bbox": [ + 0.401, + 0.397, + 0.474, + 0.432 + ], + "angle": 0, + "content": "Check the rate of change in Google's stock price over last one month" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.469, + 0.462, + 0.503 + ], + "angle": 0, + "content": "Find Rental Properties in Seattle, Washington" + }, + { + "type": "text", + "bbox": [ + 0.404, + 0.538, + 0.476, + 0.581 + ], + "angle": 0, + "content": "Using the popup opened, scroll over to find weather in Chicago on 18th September" + }, + { + "type": "title", + "bbox": [ + 0.526, + 0.367, + 0.582, + 0.376 + ], + "angle": 0, + "content": "Action Script" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.396, + 0.569, + 0.403 + ], + "angle": 0, + "content": "import pyautogui" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.409, + 0.588, + 0.433 + ], + "angle": 0, + "content": "pyautogui.moveTo(210, 179) \npyautoguidragTo(243, 179, button=\"left\")" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.465, + 0.568, + 0.472 + ], + "angle": 0, + "content": "import pyautogui" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.479, + 0.603, + 0.509 + ], + "angle": 0, + "content": "pyautogui.click(398,167) pyautogui.write(\"Seattle,WA\") pyautogui.write(\"enter\")" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.542, + 0.567, + 0.548 + ], + "angle": 0, + "content": "import pyautogui" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.555, + 0.603, + 0.58 + ], + "angle": 0, + "content": "pyautogui.goiveFo(410, 207) \npyautogui.hscroll(10) \npyautogui.clclick(410, 207)" + }, + { + "type": "title", + "bbox": [ + 0.661, + 0.367, + 0.705, + 0.375 + ], + "angle": 0, + "content": "Illustration" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.381, + 0.751, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.453, + 0.751, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.493, + 0.746, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.523, + 0.751, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.613, + 0.785, + 0.669 + ], + "angle": 0, + "content": "Fig. 1: OmniACT dataset and benchmark for enabling autonomous human-computer interaction agents. The left shows an image paired with a natural language task description as the input, the right shows the resulting action script to be executed on the screen. Examples are presented from Stocks, Apartments, and Weather application." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.692, + 0.741, + 0.817 + ], + "angle": 0, + "content": "Abstract. For decades, human-computer interaction has fundamentally been manual. Even today, almost all productive work done on the computer necessitates human input at every step. Autonomous virtual agents represent an exciting step in automating many of these menial tasks. Virtual agents would empower users with limited technical proficiency to harness the full possibilities of computer systems. They could also enable the efficient streamlining of numerous computer tasks, ranging from calendar management to complex travel bookings, with minimal human intervention. In this paper, we introduce OmniACT, the first-of-a-kind" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.744, + 0.84 + ], + "angle": 0, + "content": "* These authors contributed equally. The order is determined by dice rolling." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.148, + 0.74, + 0.384 + ], + "angle": 0, + "content": "dataset and benchmark for assessing an agent's capability to generate executable programs to accomplish computer tasks. Our scope extends beyond traditional web automation, covering a diverse range of desktop applications. The dataset consists of fundamental tasks such as \"Play the next song\", as well as longer horizon tasks such as \"Send an email to John Doe mentioning the time and place to meet\". Specifically, given a pair of screen image and a visually-grounded natural language task, the goal is to generate a script capable of fully executing the task. We run several strong baseline language model agents on our benchmark. The strongest baseline, GPT-4, performs the best on our benchmark. However, its performance level still reaches only \\(15\\%\\) of the human proficiency in generating executable scripts capable of completing the task, demonstrating the challenge of our task for conventional web agents. Our benchmark provides a platform to measure and evaluate the progress of language model agents in automating computer tasks and motivates future work towards building multimodal models that bridge large language models and the visual grounding of computer screens." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.397, + 0.747, + 0.438 + ], + "angle": 0, + "content": "Keywords: Generalist Agent \\(\\cdot\\) Multimodal Machine Learning \\(\\cdot\\) Dataset and Benchmark \\(\\cdot\\) Vision Language Understanding \\(\\cdot\\) UI grounding \\(\\cdot\\) Human-computer interaction" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.463, + 0.375, + 0.479 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.493, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Performing computer tasks based on natural language instructions has been a long-standing goal of artificial intelligence [49]. One concrete objective in the line of research is to develop generalist agents that can assist humans in doing computer tasks [21], such as \"Order a pizza from Domino's\" or \"Write a message to John.\" The agent should be able to open the application and perform the task. Executing these actions on a personal computer involves a sequence of interactions with a mouse and keyboard. For example, the simple task of writing an email involves hovering over the application icon, clicking it, clicking the 'New Email' button, writing the content of the email, and clicking send. Successfully sending an email requires accurately predicting the correct action at each step and accurately executing it, which is a herculean task even for the best agents today [14]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.675, + 0.787, + 0.84 + ], + "angle": 0, + "content": "A generalist agent for computer tasks must understand natural language instructions, process visual screenshots, and produce the correct sequence of actions to be performed to achieve the intended task. Several existing approaches focus on building agents based on the HTML model [9, 40, 62]. However, this approach introduces several challenges and constraints. These agents are limited to web applications and often struggle with complex or long-context HTML code. They cannot interact with native desktop applications or perform tasks that span multiple applications, like drafting an email using text from a code editor, without significant alterations. Furthermore, HTML-based agents, which are inherently powered by text-only language models, typically underperform in tasks requiring visual cues, such as identifying and clicking a blue button" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "on a desktop's top-right corner. In contrast, humans can easily understand UI elements like dropdown menus, typable areas, redirections, and options with just a glance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Towards the goal of developing a generalist autonomous agent with robust visual and user interface (UI) understanding capabilities, we introduce a new task and dataset, OmniACT, containing over 9.8K pairs of images and instructions (Figure 1) across different operating systems and the web. This dataset includes screenshots of various UI screens and corresponding natural language instructions. The objective of these instructions is to generate executable commands using the PyAutoGUI Python library [1]. PyAutoGUI enables the automation of the mouse and keyboard operations, which helps to facilitate interactions with various native applications across macOS, Windows, and Linux. This simplifies completing specified tasks across different web domains and native desktop applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.48 + ], + "angle": 0, + "content": "We evaluate several language model-based agent baselines on this dataset, including LLaMA [47], Vicuna [7], Palmyra-X (43B) [2], InstructPalmyra-30B [45], GPT 3.5, and GPT-4 [32]. We experiment with fine-tuning Vicuna-13B and LLaMA-13B models using QLoRA [10]. We also benchmark multimodal baseline LLaVa-v1.5-7B, LLaVa-v1.5-13B [47], Gemini-Pro [44] and GPT-4-vision-preview [55] for the task. Our findings highlight the necessity for a multimodal model capable of executing these tasks, and our analysis provides insights into promising future work in the space. Our key contributions are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.489, + 0.787, + 0.548 + ], + "angle": 0, + "content": "1. We release a novel dataset of desktop and website applications consisting of over 9.8K natural language tasks, UI screens, and corresponding code snippets collected through human annotation. We introduce custom performance metrics tailored for computer tasks." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.549, + 0.785, + 0.579 + ], + "angle": 0, + "content": "2. We propose DetACT, a module for creating textual representations of the screen using signals from OCR, color, and icon-template matching." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.58, + 0.785, + 0.639 + ], + "angle": 0, + "content": "3. We conduct a comprehensive benchmark and analysis of state-of-the-art LLMs and multimodal models on our benchmark. Our results show that OmniACT is a challenging task for even the best LLM agents today, and existing models are far below human performance." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.489, + 0.787, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.663, + 0.388, + 0.679 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.414, + 0.71 + ], + "angle": 0, + "content": "2.1 UI Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "User interface (UI) understanding has garnered interest from researchers in the machine learning and human-computer interaction communities, evolving with various models focusing on understanding the semantics of mobile and web user interfaces. UIBert [3], PixelBERT [16], ActionBert [15], VUT [25], Screen2Words [48], WidgetCaptioning [24] and Pix2Act [39] are notable models in this area. They propose approaches for learning the user-interface semantics of the mobile screen using the image and view hierarchy. These models have demonstrated effectiveness in tasks like capability prediction, screen segmentation and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.456, + 0.129 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.145, + 0.726, + 0.159 + ], + "angle": 0, + "content": "Table 1: Comparison of OmniACT with other related benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.172, + 0.787, + 0.32 + ], + "angle": 0, + "content": "
DatasetsSizeEnv TypeTask HeterogeneityReal-World PortrayalExecutional CorrectnessSupports Continuous Scale Desktop AppsAdaptive EvaluationTask
VisualWebArena [20]910WebYesYesYesNoNoWeb Navigation
WebArena [62]812WebYesYesYesNoNoWeb Navigation
Mind2Web [9]2350WebYesYesNoNoNoWeb Navigation
WebShop [56]12000 ProductsWebNoNoYesNoNoWeb Navigation
RUSS [53]80WebYesYesNoNoNoWeb Navigation
WebSRC [6]2735WebYesYes-NoNoQA
MiniWoB++ [17]100Mobile WebsitesNoNoYesNoNoWeb Navigation
PixelHelp [23]187MobileYesYesNoNoNoUI Grounding
MetaGUI [42]1125MobileYesYesYesNoNoMobile Navigation
MoTIF [5]756MobileYesYesYesNoNoMobile Navigation
AITW [35]715142Mobile and WebYesYesYesNoNoMobile/Web Navigation
OmniACT (Ours)9802Desktop and WebYesYesYesYesYesCode Generation
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.785, + 0.481 + ], + "angle": 0, + "content": "understanding, and screen caption generation. Lexi [4] and Spotlight [22] propose models that use vision-only inputs to minimize the reliance on metadata such as view hierarchy. Furata et al. [11] demonstrates the use of fine-tuning for multimodal web navigation. The majority of machine learning models trained for UI understanding leverage the Rico dataset [8] and its extensions, which contain 64,462 unique Android screens and metadata. In addition, [4] released the UICaptions dataset, which consists of diverse image-captions pairs across a wide range of applications. PixelHelp [23] also released a corpus to train models that can interpret natural language instructions and map them to mobile UI actions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.519, + 0.516 + ], + "angle": 0, + "content": "2.2 Autonomous Computer Agents" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.779 + ], + "angle": 0, + "content": "The advent of large language models (LLMs) has been pivotal in the rapid advancement of agents that operate on web pages. Recent research such as ViperGPT [43] Chameleon [29], RCI Agent [18], VisProg [12], and [31] employ LLMs for planning or action prediction in developing autonomous agents. Benchmark datasets, such as MiniWoB [40], WebShop [56], Macaw-LLM [30], ASH-Prompting [41] Mind2Web [9], WebArena [62], AgentBench [28] and VisualWebArena [20] have also been proposed to measure the ability of LLM-based agents to automate web tasks. These methods mainly involve agents that operate on a text-based Document Object Model (DOM) of HTML scripts. This limits their understanding of screen context, which is crucial for the model's decision-making and action-taking processes. To address this limitation, [35] released Android in the Wild, a dataset comprising screens, natural language instructions, and corresponding actions. Following this, [59] proposed a multimodal model, AutoUI, which is designed to build an agent on the Android in the Wild dataset confined to the Android ecosystem. WebAgent [13] utilized Flan-U-PaLM, for grounded code generation, and HTML-T5 and showed improvement on real-world websites." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Current benchmarks for autonomous agents focus mainly on the Web or Android environments, posing challenges for tasks involving desktop applications or spanning multiple applications beyond the web domain. The absence of established benchmarks and datasets in this area, coupled with basic methods" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "for extracting user interface (UI) elements, underscores the need for significant progress in developing more versatile autonomous agents capable of handling diverse tasks beyond the current scope. To highlight the unique features that OmniACT introduces in the assessment of capable autonomous agents, we provide a comparison between the existing benchmarks and our proposed benchmark, OmniACT, in Table 1." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.263, + 0.761, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.484, + 0.788, + 0.554 + ], + "angle": 0, + "content": "Fig. 2: Data Collection Pipeline. (1) We select over 60 applications and websites to ensure diversity, (2) segment the screen through human-annotated bounding boxes, (3) label the bounding boxes based on functionality, (4) ask student volunteers to come up with tasks, given a screen image, and (5) reverse map the textual labels to coordinates and filter the scripts based on execution and syntax." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.601, + 0.355, + 0.617 + ], + "angle": 0, + "content": "3 OmniACT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.719 + ], + "angle": 0, + "content": "We introduce a novel dataset and benchmark, OmniACT, which measures the performance of autonomous agents on both web and desktop applications. Compared to previous benchmarks which focus on text-based reasoning [9, 17, 40, 56, 62], our benchmark aims to measure multimodal agents that bridge large language model planners and UI understanding vision models. OmniACT can be accomplished as a standalone task as it is not under a mock environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "All actions that a human can execute on the computer can be encoded in the PyAutoGUI [1] Python framework. This framework allows a user to execute keyboard and mouse operations by running Python code. The PyAutoGUI code to execute these tasks is shown in the third column of Figure 1. For other computer tasks, the PyAutoGUI library provides functions such as 'press', 'write', and 'scroll' which can be used to execute the task. Our dataset consists of parallel data of natural language tasks, UI screenshots, and ground truth PyAutoGUI scripts that achieve successful execution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.406, + 0.16 + ], + "angle": 0, + "content": "3.1 Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Given an input state of a computer defined by the screen \\( S \\) and the task description \\( T \\) in natural language, the goal of the task is to output a sequence of actions \\( A \\) that can successfully accomplish the task \\( T \\) within a screenshot \\( S \\in \\{ \\text{Linux, Windows, MacOS, Webpage} \\} \\). Formally, the task can be defined as learning the transition function \\( f: T \\times S \\to A \\). During dataset collection, we ensure that all task descriptions \\( T \\) are feasible and can be accomplished in the current screenshot \\( S \\). To reduce ambiguity and facilitate better evaluation, we ensure that task descriptions are detailed and unambiguous. Tasks can also be visually grounded (e.g., 'Click the red button to start recording') or natural language based (e.g., 'Click the My Account button'). We define the action space using the functionalities in the PyAutoGUI library: \\( A \\in \\{\\text{'click', 'dragTo', 'scroll', 'write', \\ldots}\\} \\). The exhaustive list of actions is provided in Table 2. Our action space is much larger than other benchmarks [9,40,62] that resort to two or three interaction options. Mouse actions such as 'moveTo', 'click', 'rightClick', 'doubleClick', and 'dragTo', additionally require screen coordinates as arguments, which indicate the pixel location of the action." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Figure 1 illustrates sample tasks and corresponding outputs for three applications within OmniACT: (1) Stocks (MacOS), (2) Apartments.com (web page), and (3) Weather (MacOS). The first column depicts the input image, and the second column shows the natural language task that is to be executed on the current screen. To execute these tasks, a user must accurately perform a series of operations using the mouse and keyboard. Eg: to check the rate of change in Google's stock price over the last month, the mouse has to be moved to the last month and dragged while holding the left-click button to the current month." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.431, + 0.569 + ], + "angle": 0, + "content": "3.2 Dataset Preparation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.73 + ], + "angle": 0, + "content": "To prepare our dataset, we followed a pipelined approach, as summarized in Figure 2. We first selected a variety of applications and websites. For each application or website, we created bounding boxes around key UI elements and labeled them according to their functionality, which is crucial for assisting human annotators in writing accurate PyAutoGUI scripts. After each script is written, we converted the labels back into numeric coordinates, allowing us to align the scripts precisely with the locations of the UI elements. Finally, we thoroughly reviewed each script, focusing on its executability and adherence to syntax standards. This ensured the high quality and functionality of our dataset, making it a valuable resource for training and evaluating autonomous agents." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Application/Website Selection To test the computer agents' generalization ability across different tasks, we collect tasks across multiple domains on both desktop and web applications. In total, we collect and annotate 9802 data points (Table 3), with the split between desktop and web applications approximately 3:1. The emphasis on desktop applications, which do not contain Document Object Model (DOM) hierarchies unlike HTML-based web pages, presents a more" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.221 + ], + "angle": 0, + "content": "complex multimodal challenge where visual cues are crucial. We collect tasks from applications within the three most popular operating systems. We select 22 native applications from MacOS, and 8 each from Linux and Windows. We annotate roughly 3 to 4 screens for every application. The full list of applications is provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.516, + 0.403 + ], + "angle": 0, + "content": "Many common computer tasks today are still performed through web applications, so we also collect 3-4 screenshots from 27 different web applications. To ensure diversity in task intents, we categorize these tasks into one of the following 6 categories: (1) Shopping, (2) Entertainment, (3) Service, (4) Government, (5) Travel, (6) Health. Inspired by the methodology of [9], these categories were selected to cover a wide range of user intents and functionalities." + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.267, + 0.787, + 0.307 + ], + "angle": 0, + "content": "Table 2: Action types supported by OmniACT and the number of instances for each action in the dataset." + }, + { + "type": "table", + "bbox": [ + 0.584, + 0.31, + 0.73, + 0.415 + ], + "angle": 0, + "content": "
TypeAction%
MouseClick63.73
Double Click0.58
Right Click0.77
Move/Hover1.85
Drag0.29
Scroll1.68
Horizontal Scroll0.17
KeyboardPress16.28
Hotkey3.00
Write11.65
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.516, + 0.471 + ], + "angle": 0, + "content": "UI Screen Segmentation To collect gold-standard data, we first annotate and segment the screen by identifying the" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.784, + 0.502 + ], + "angle": 0, + "content": "bounding boxes present on the screen. We employ slightly different techniques for web and desktop applications to create the bounding boxes:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.514, + 0.784, + 0.574 + ], + "angle": 0, + "content": "1. Desktop Applications: We build a custom annotation interface based on PyQt5\\(^3\\) to create bounding boxes manually over a screen image using a simple drag-and-click mechanism. This custom interface expedites the process and allows us to get highly accurate gold-label data for desktop images." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.575, + 0.784, + 0.635 + ], + "angle": 0, + "content": "2. Websites: For webpages, we write JavaScript code to extract all interactable (click, type, etc.) regions from HTML source code. We also extract banners, dropowns, submit, and radio buttons from the screen. We filter the elements to retain only those that are visible and interactable within the screen." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.514, + 0.784, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.784, + 0.763 + ], + "angle": 0, + "content": "Functionality Tagging To map each bounding box to its correct functional description, we leverage Amazon MTurk workers (see details in Appendix), who are given an image with a bounding box and are required to write the correct description or label of the bounding box's function. For example, given an image of an Amazon webpage with a search bar, the annotator labels it as \"find-product-search-bar\". The logical descriptions are used to create tasks in a structured manner without the need to identify individual bounding box coordinates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.784, + 0.817 + ], + "angle": 0, + "content": "Task Creation Our approach for each screen involves utilizing all human-annotated bounding boxes and their labels to create tasks that can be executed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.473, + 0.84 + ], + "angle": 0, + "content": "3 https://pypi.org/project/PyQt5/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.456, + 0.129 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 3: DetACT Module. Given an initial image and a natural language task description, we use a pipelined approach to run OCR and SAM on the screen. The outputs from SAM are then used by icon and color-matching modules to obtain an exhaustive set of useful UI elements. The list of elements is passed through LLM based filter to select only the elements related to the given task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.498 + ], + "angle": 0, + "content": "within the confines of a single screen. These tasks are designed to be visually grounded in order to measure the capabilities of multimodal agents. We plan to release the bounding box and their corresponding labels as the metadata for evaluation purposes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.621 + ], + "angle": 0, + "content": "For dataset compilation, college students with basic Python programming skills served as annotators, accessing API references for PyAutoGUI and examples of potential tasks. Each student generated multiple tasks, each accompanied by three alternative natural language reformulations. For instance, \"What is \\(3 + 2\\)?\" might be reformulated as \"Calculate the sum of 2 and \\(3\\)\" or \"Add two to three\". To avoid train-test leakage, rephrased tasks were consistently placed in the same dataset split. Further details on the annotation process are available in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.518, + 0.809 + ], + "angle": 0, + "content": "Reverse Mapping and Filtering To ensure high-quality data, we incorporate an additional step into the data collection pipeline. We build scripts to map the text-based labels of each bounding box back to their numeric coordinates, and then match the syntax and verify if the task will be executed on the screen. Using this filter, we remove all the non-working or syntactically incorrect data points and finally manually review the set of tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "After filtering, we obtain 9802 human-annotated, gold-label data points across more than 200 desktop and web screens" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.688, + 0.788, + 0.715 + ], + "angle": 0, + "content": "Table 3: Dataset distribution across splits and platforms." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.717, + 0.772, + 0.79 + ], + "angle": 0, + "content": "
DomainTrain ValidationTestTotal
DesktopMac OS3028444786
Linux761126247
Windows1573216458
Web-1427206530
Total67899922,021
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "(Table 3), split into train, validation, and test sets in a 7:1:2 ratio. All collected data will be publicly released to encourage future work on multimodal agents." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.2, + 0.436, + 0.216 + ], + "angle": 0, + "content": "4 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.784, + 0.397 + ], + "angle": 0, + "content": "In this section, we detail various evaluation metrics for benchmarking model performance on the OmniACT dataset. UI screens have additional constraints such as spatial relevance which are not factored in most conventional similarity-based metrics such as BLEU [34], CodeBLEU [36], BERTScore [58] and CodeBERTScore [61]. For example, a valid click action is usually not constrained to a single coordinate but can be any coordinate within a specified region. In the event of invalid coordinate predictions, an agent that predicts coordinates further away from the valid region should invoke a higher penalty compared to an agent that predicted coordinates close to the region. We propose two new metrics adapted: Sequence Score (Section 4.1) and Action Score (Section 4.2) aimed at utilizing UI information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.42, + 0.389, + 0.435 + ], + "angle": 0, + "content": "4.1 Sequence Score" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.784, + 0.505 + ], + "angle": 0, + "content": "The sequence score measures whether the predicted action sequence (e.g., 'click', 'write', 'press') exactly matches the gold sequence. Since predicting the first action in the sequence is relatively straightforward and later actions are more difficult, we define sequence score as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.508, + 0.785, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S e q S c o r e} _ {i} = \\left\\{ \\begin{array}{l l} \\beta_ {1} + \\beta_ {2} * (s - 1) & \\text {i f a l l a c t i o n s m a t c h} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.549, + 0.733, + 0.564 + ], + "angle": 0, + "content": "where \\(s\\) is the action sequence length, \\(\\beta_{1}\\) is set to 0.1 and \\(\\beta_{2}\\) is set to 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.37, + 0.601 + ], + "angle": 0, + "content": "4.2 Action Score" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.784, + 0.671 + ], + "angle": 0, + "content": "The action score measures how well a code snippet containing the correct action sequence can perform the task. Specifically, for a script with a correct action sequence, we introduce penalties for inaccurate behavior. The penalties are described below:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.683, + 0.784, + 0.743 + ], + "angle": 0, + "content": "1. Click penalty \\((M)\\): For actions 'click', 'rightClick', 'doubleClick', 'moveTo', and 'dragTo', we penalize code snippets where predicted coordinates lie outside of the bounding box of the UI element. The click penalty for the \\(j^{th}\\) action of the \\(i^{th}\\) example is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.746, + 0.785, + 0.793 + ], + "angle": 0, + "content": "\\[\nM _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 1 - \\frac {\\mu}{\\mu + L _ {2}} & \\text {i f S e q S c o r e} _ {i} > 0 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.795, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Here \\( L_{2} \\) corresponds to the smallest Euclidean distance between the predicted coordinate and bounding box. \\( L_{2} \\) is zero when the predicted coordinate lies within the target bounding box. \\( \\mu \\) is the Dirichlet smoothing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.456, + 0.129 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "coefficient which we dynamically set to the inverse of the length of the diagonal of the bounding box. This ensures that the penalty for points outside the bounding box varies based on the size of the bounding box. For two predicted points with the same \\( L_{2} \\), the metric penalizes more heavily if the box is larger. This is sound with the intuition that the chances of clicking on a larger box are higher and should be penalized more in case of a mistake." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.238, + 0.788, + 0.285 + ], + "angle": 0, + "content": "2. Key penalty \\((K)\\): For actions 'press' and 'hotkey', we check whether the set of keys in the target code (represented as \\( GK_{i}^{j} \\)) and predicted code (represented as \\( PK_{i}^{j} \\)) are the same. It is formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.288, + 0.786, + 0.327 + ], + "angle": 0, + "content": "\\[\nK _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 0 & \\text {i f} G K _ {i} ^ {j} = P K _ {i} ^ {j} \\text {a n d S e q S c o r e} _ {i} > 0 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.33, + 0.787, + 0.375 + ], + "angle": 0, + "content": "3. Write penalty \\((W_{p})\\): For action type 'write', we penalize the output for the sentence to be typed. Specifically, we the employ BLEU score [34], and compute:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.378, + 0.786, + 0.417 + ], + "angle": 0, + "content": "\\[\nW _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 1 - B L E U \\left(G S _ {i} ^ {j}, P S _ {i} ^ {j}\\right) & \\text {i f S e q S c o r e} _ {i} > 1 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.421, + 0.787, + 0.453 + ], + "angle": 0, + "content": "Here, \\( GS_{i}^{j} \\) represents the actual sentence to be typed, and \\( PS_{i}^{j} \\) represents the sentence predicted by the model in the \\( j^{th} \\) action of example \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.464, + 0.591, + 0.479 + ], + "angle": 0, + "content": "In the above equations, \\((\\alpha_{i})\\) is the weighting factor:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.482, + 0.786, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\operatorname {S e q S c o r e} _ {i} / \\text {l e n g t h o f s e q u e n c e} i \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.786, + 0.53 + ], + "angle": 0, + "content": "This ensures that the action score \\(\\in [0,1]\\). The mean action score is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.532, + 0.786, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\text {A c t i o n S c o r e} = \\frac {\\sum_ {i} \\max \\left(\\operatorname {S e q S c o r e} _ {i} - \\sum_ {j} \\left(M _ {i} ^ {j} + K _ {i} ^ {j} + W _ {i} ^ {j}\\right) , 0\\right)}{\\sum_ {i} \\operatorname {S e q S c o r e} _ {i}} \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.593, + 0.638, + 0.611 + ], + "angle": 0, + "content": "5 DetACT: DETecting ACTions from UI" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Understanding UI screens is crucial for multimodal computer tasks. Web-based agents typically use language-only inputs from the HTML DOM. This is insufficient for comprehending the full extent of an application UI, as many components may not be easily described with HTML code. To address this, we propose DetACT, which allows us to convert images of UI layouts into structured code and text outputs for a downstream LLM. DetACT is a system comprised of three distinct modules: the text module, the icon module, and the color module." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.741, + 0.788, + 0.819 + ], + "angle": 0, + "content": "1. Text Extraction: We use the EasyOCR model4 to parse over the UI screens and collect all text-based elements. Along with the text, we also note the locations of each of these elements. This is depicted in Figure 3, along with a list of text elements found on the screen using the OCR Module. We segment and classify the different regions within the screenshot using the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.507, + 0.841 + ], + "angle": 0, + "content": "4 https://github.com/JaidedAI/EasyOCR" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.146, + 0.67, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.421, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 4: Baseline Model Architecture. Image and task descriptions are sent to DetACT module, which gives a filtered list of UI elements relevant to feed into the prompt along with the task. We also show the prompt structure used for action script generation. This structure is passed through the LLM (along with the image for multimodal LLM) to generate the automation script." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.523, + 0.785, + 0.553 + ], + "angle": 0, + "content": "Segment Anything Model (SAM) [19]. From the outputs, we filter out the non-textual segments for our icon and color detection." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.555, + 0.788, + 0.706 + ], + "angle": 0, + "content": "2. Icon Module: For matching with the appropriate icon, we use a pack of 1600 icons5 as templates. Each of these icons is labeled with their appropriate functionality and is matched with the filtered outputs SAM [19]. For the similarity of the two images, we resize the reference icons and segmented region of interest (ROI) to the same size, and convert both images to grayscale. After this, we use the Structural Similarity Index (SSIM) [52], to find the closest match of the ROI to the icons in our set, and select the ones above the SSIM threshold of 0.95. As seen in Figure 3, a few icons matched on the screen are Globe icon, Calendar icon, Person icon, and Location icon; each depicting a different use case." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.707, + 0.788, + 0.813 + ], + "angle": 0, + "content": "3. Color Module: Finally, to place all segments of interest into appropriate buckets of colors, we average the RGB pixel values over the ROI and, based on that value, bucket them into different color categories. We categorize colors differently based on the human perspective of the ranges of each color. To avoid ambiguity, we consider eleven major colors, namely yellow, blue, green, red, pink, violet, white, black, orange, brown, and grey. We record the center of the element along with the color." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.555, + 0.788, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "5 https://icomoon.io/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Once all the elements of each category are extracted with their coordinates, we then filter these UI elements by prompting GPT-4 [32]. We ensure that the elements selected are suited only for our task, for which we also provide the task description in our prompts along with the list of elements. Full details of the prompt are provided in the appendix section of the paper. As we observe in Figure 3, given an image from the Expedia application, and a task (\"Click on the Black Location icon and enter the destination as Paris\"), the LLM filters out the elements to retain only \"Going To\", \"Location Icon\", and the Black colored elements from the screen. This is passed as input to the LLM or VLM backbone." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.305, + 0.342, + 0.321 + ], + "angle": 0, + "content": "6Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.789, + 0.577 + ], + "angle": 0, + "content": "To evaluate the performance of existing language model-based agents, we conduct experiments with both language-based and multimodal baselines. The DetACT module takes in image and text descriptions of the task and outputs the color, icon, and text-based signals. This is concatenated to the prompt for the LLM prompt-based baselines (see Figure 4). Every prompt starts with a role assignment [60], followed by the detailed API reference of the PyAutoGUI function set, along with a textual description of their function. We then add five in-context examples from the training set that most closely match the task (based on the cosine similarity of the MiniLM [50] embeddings of the reference task and the train examples). We add a list of UI elements filtered by the DetACT module to the prompt. Finally, we provide the rules with the task description. For multimodal baselines, we also pass the image pixels to the vision encoder. We choose coordinate-based UI elements in the prompt as recent techniques like the Set-of-Mark (SOM) [54] prompting does not work for desktop settings since it is difficult to obtain interactive elements from the desktop screen images. We report the results of several baselines:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.585, + 0.788, + 0.689 + ], + "angle": 0, + "content": "- Few-shot Generative LLM: We experiment with models from LLaMA-2 [47], Vicuna-1.5 [7], CodeLLaMA-34B [37], Palmyra [46], and GPT [32] series. We use the prompts structure as shown in Figure 4 to prompt the model. For LLaMA and CodeLLaMa, we reduce the prompt length to 2000 tokens by removing outputs from the DetACT module with lower confidence, as we observed poor performance on longer prompts. For the other models, we allow prompts with up to 4000 token sizes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.69, + 0.788, + 0.748 + ], + "angle": 0, + "content": "- Finetuned Generative LLM: We fine-tuned the LLaMA-13B model and Vicuna-13B using QLoRa [10] with rank 64 and scaling factor 16 for 300 steps to generate the code given screen description from the DetACT module and the instruction." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "- Few-shot Generative Multimodal Models: As OmniACT is predominantly multimodal, with a majority of tasks being visually grounded, we conduct experiments with large multimodal models. Given the limited research in this domain [51, 57], there is a scarcity of available multimodal models with significant size adept for this task. Here, we experiment with [26, 27], providing a similar prompt as well as the screen image." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.585, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.455, + 0.163 + ], + "angle": 0, + "content": "7 Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.516, + 0.431 + ], + "angle": 0, + "content": "As shown in Table 4, we experiment with three different categories of models, namely Prompt-based LLMs, Fine-tuned LLMs, and Prompt-based Multimodal Models. GPT-4 is the best-performing approach, scoring higher on the sequence score and invoking lower penalties on coordinate predicting and text input. For prompt-only LLMs, the GPT-3.5-turbo and GPT-4 models outperform the other LLM baselines, including the LLaMA [47] and Vicuna [7] models. We observe that CodeLLaMA-34B [38], which is trained for code generation, also achieves a higher performance than other models of the same size at predicting the action sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.516, + 0.599 + ], + "angle": 0, + "content": "Fine-tuned models also perform much better than their few-shot prompt-only counterparts. Fine-tuning substantially improves LLaMA-13B's sequence score (4.80 to 8.92) and action score (1.62 to 2.14), as well as the other metrics. Despite this, we observed that both, prompt-based LLMs and finetuned LLMs face severe mouse penalties, especially on click coordinates. This is because they rely solely on text-based signals." + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.22, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Table 4: Baseline Performance. (A) Prompt-only LLMs, (B) Fine Tuned LLMs, (C) Prompt-only Multimodal Models. The table represents the Sequence score (SS), click penalty \\((M_p)\\), Key penalty \\((K_p)\\), Write Penalty \\((W_p)\\), and Action Score (AS). The best results for the (SS) and (AS) are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.333, + 0.787, + 0.55 + ], + "angle": 0, + "content": "
ModelSS(↑)MpKpWpAS(↑)
Prompt based LLMs
LLaMA-7B [47]4.121.241.830.570.48
Vicuna-7B [7]3.881.171.510.430.77
LLaMA-13B [47]4.801.320.930.931.62
Vicuna-13B [7]5.441.650.941.061.78
Palmyra-Instruct-30B [45]7.515.680.120.401.31
CodeLLaMA-34B [38]10.092.992.710.663.72
Palmyra-X 43B [2]11.203.123.022.122.94
GPT-3.5-turbo-0613 [33]22.858.134.512.317.89
GPT-4 [32]32.7510.276.993.8911.60
Finetuned LLMs
LLaMA-13B FT8.924.611.430.742.14
Vicuna-13B FT8.784.121.310.632.72
Multimodal LLMs
LLaVA-v1.5-7B [27]13.234.731.241.445.82
LLaVA-v1.5-13B [26]20.566.073.442.858.19
Gemini-Pro [44]30.989.056.813.6611.46
GPT-4V [26]38.7210.537.144.0317.02
Human Performance82.230.120.361.6180.14
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.751 + ], + "angle": 0, + "content": "To address this, we experiment with multimodal language models (Table 4). We observe that the coordinate prediction improves significantly when we provide the entire image as input to the multimodal LLM, as this enables it to fully utilize the screen representation. In addition to open sourced models, we also experiment with the GPT-4-vision API [55] which shows that GPT-4 Vision [55] outperforms GPT-4 significantly on the Action Score along with improving the sequence score, which we attribute to the strong reasoning abilities of GPT-4 coupled with the improved visual understanding capabilities of the GPT-4-vision model [55]. These findings pave the way towards exciting new research directions on building multimodal models for long-horizon planning and code generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Human performance over the task: OmniACT consists of visually complicated tasks, and tests various types of computer skills. In order to get a gauge of how well humans perform, we collect evaluation data from human evaluators. We split the test set uniformly amongst 10 human evaluators, and provided them with the screenshot and task instruction. We record the actions taken by the an-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "notators, and measure their performance on our predefined metrics (Table 4). We find that users generally exhibit a high level of proficiency when attempting most tasks for the first time. However, there are instances where users face difficulties in successfully completing certain tasks. These are due to factors including the user's inability to fully comprehend the task, difficulties in grounding the task to the provided screenshot, or a lack of familiarity with the UI." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.258, + 0.533, + 0.274 + ], + "angle": 0, + "content": "8 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.455 + ], + "angle": 0, + "content": "Autonomous virtual agents offer the potential to automate routine tasks, benefiting users with limited technical expertise. To solve this task, we introduce OmniACT, a unique dataset of 9.8K human-labeled data points. OmniACT benchmarks autonomous agents across a range of tasks on web and desktop applications. LLM-based agents, like GPT-4, achieve a respectable action score of 11.6 on our dataset. However, OmniACT presents a challenge for the current state-of-the-art language and multimodal models. It provides a direction for future research on foundational multimodal models that seamlessly integrate language and visual understanding of computer screens and stands poised to drive the next wave of advancements in generalist autonomous agents offering omnipotent assistance to humans." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.474, + 0.363, + 0.49 + ], + "angle": 0, + "content": "9 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.61 + ], + "angle": 0, + "content": "This work introduces a valuable dataset, yet we recognize a few limitations that exist. State-of-the-art models like GPT-4, may exhibit susceptibility to hallucinations and bias towards specific data types, hindering broad applicability. Reliance on closed models like GPT-4V poses integration challenges due to high costs and time constraints. Despite efforts for equal representation and data collection without personal information, biases may be introduced as the dataset is exclusively in English, and human-curated content may have temporal biases." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.431, + 0.647 + ], + "angle": 0, + "content": "10 Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As a part of the dataset creation process, we carefully review the pipeline at every stage, ensuring there is no personally identifiable information or offensive content, either during data collection or through the use of LLMs. For all purposes, we create dummy accounts that mimic real-like user content. To get the gold labels scripts we seek help from well-qualified student workers, approved through the institution, and get the bounding box data annotated through MTurk workers, both of whom are paid $25 per hour, which is greater than the minimum wage rate (We detail this process in the supplementary material). Human studies are also done with the help of student workers approved by the institution at the above-mentioned payscale. We also ensure that all groups have equitable representation and that no personal opinions are reflected in the dataset, avoiding bias during the collection as well as the annotation process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.224 + ], + "angle": 0, + "content": "We extend our heartfelt gratitude to Writer.com for their generous and unwavering support throughout this project. Their dedicated team's expertise and collaboration were invaluable in achieving our goals." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.249, + 0.323, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.784, + 0.31 + ], + "angle": 0, + "content": "1. Pyautogui: A cross-platform gui automation python module for human beings. https://github.com/asweigart/pyautogui (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. AlShikh, W., Daaboul, M., Goddard, K., Imel, B., Kamble, K., Kulkarni, P., Russak, M.: Becoming self-instruct: introducing early stopping criteria for minimal instruct tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.785, + 0.381 + ], + "angle": 0, + "content": "3. Bai, C., Zang, X., Xu, Y., Sunkara, S., Rastogi, A., Chen, J., y Arcas, B.A.: Uibert: Learning generic multimodal representations for ui understanding (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.408 + ], + "angle": 0, + "content": "4. Banerjee, P., Mahajan, S., Arora, K., Baral, C., Riva, O.: Lexi: Self-supervised learning of the ui language (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.785, + 0.448 + ], + "angle": 0, + "content": "5. Burns, A., Arsan, D., Agrawal, S., Kumar, R., Saenko, K., Plummer, B.A.: Mobile app tasks with iterative feedback (motif): Addressing task feasibility in interactive visual environments" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.785, + 0.505 + ], + "angle": 0, + "content": "6. Chen, X., Zhao, Z., Chen, L., Ji, J., Zhang, D., Luo, A., Xiong, Y., Yu, K.: Websrc: A dataset for web-based structural reading comprehension. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 4173-4185 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.56 + ], + "angle": 0, + "content": "7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.785, + 0.644 + ], + "angle": 0, + "content": "8. Deka, B., Huang, Z., Franzen, C., Hibschman, J., Afergan, D., Li, Y., Nichols, J., Kumar, R.: Rico: A mobile app dataset for building data-driven design applications. In: Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology. p. 845-854. UIST '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3126594.3126651, https://doi.org/10.1145/3126594.3126651" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.686 + ], + "angle": 0, + "content": "9. Deng, X., Gu, Y., Zheng, B., Chen, S., Stevens, S., Wang, B., Sun, H., Su, Y.: Mind2web: Towards a generalist agent for the web. arXiv preprint arXiv:2306.06070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.688, + 0.785, + 0.714 + ], + "angle": 0, + "content": "10. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.785, + 0.756 + ], + "angle": 0, + "content": "11. Furuta, H., Nachum, O., Lee, K.H., Matsuo, Y., Gu, S.S., Gur, I.: Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "12. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Gur, I., Furuta, H., Huang, A., Safdari, M., Matsuo, Y., Eck, D., Faust, A.: A real-world webagent with planning, long context understanding, and program synthesis (2024)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.128 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Gur, I., Rueckert, U., Faust, A., Hakkani-Tur, D.: Learning to navigate the web. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "15. He, Z., Sunkara, S., Zang, X., Xu, Y., Liu, L., Wichers, N., Schubiner, G., Lee, R., Chen, J., y Arcas, B.A.: Actionbert: Leveraging user actions for semantic understanding of user interfaces (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "16. Huang, Z., Zeng, Z., Liu, B., Fu, D., Fu, J.: Pixel-bert: Aligning image pixels with text by deep multi-modal transformers (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.299 + ], + "angle": 0, + "content": "17. Humphreys, P.C., Raposo, D., Pohlen, T., Thornton, G., Chhaparia, R., Muldal, A., Abramson, J., Georgiev, P., Santoro, A., Lillicrap, T.: A data-driven approach for learning to control computers. In: International Conference on Machine Learning. pp. 9466-9482. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "18. Kim, G., Baldi, P., McAleer, S.: Language models can solve computer tasks. arXiv preprint arXiv:2303.17491 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.329, + 0.785, + 0.369 + ], + "angle": 0, + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.41 + ], + "angle": 0, + "content": "20. Koh, J.Y., Lo, R., Jang, L., Duvvur, V., Lim, M.C., Huang, P.Y., Neubig, G., Zhou, S., Salakhutdinov, R., Fried, D.: Visualwebarena: Evaluating multimodal agents on realistic visual web tasks (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "21. LeCun, Y.: A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "22. Li, G., Li, Y.: Spotlight: Mobile ui understanding using vision-language models with a focus (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.493 + ], + "angle": 0, + "content": "23. Li, Y., He, J., Zhou, X., Zhang, Y., Baldridge, J.: Mapping natural language instructions to mobile ui action sequences (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "24. Li, Y., Li, G., He, L., Zheng, J., Li, H., Guan, Z.: Widget captioning: Generating natural language description for mobile user interface elements (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.524, + 0.785, + 0.549 + ], + "angle": 0, + "content": "25. Li, Y., Li, G., Zhou, X., Dehghani, M., Gritsenko, A.: Vut: Versatile ui transformer for multi-modal multi-task user interface modeling (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.576 + ], + "angle": 0, + "content": "26. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.689, + 0.59 + ], + "angle": 0, + "content": "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.592, + 0.785, + 0.646 + ], + "angle": 0, + "content": "28. Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., Zhang, S., Deng, X., Zeng, A., Du, Z., Zhang, C., Shen, S., Zhang, T., Su, Y., Sun, H., Huang, M., Dong, Y., Tang, J.: Agentbench: Evaluating llms as agents (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.648, + 0.785, + 0.687 + ], + "angle": 0, + "content": "29. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "30. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.769 + ], + "angle": 0, + "content": "31. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.503, + 0.784 + ], + "angle": 0, + "content": "32. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.735, + 0.799 + ], + "angle": 0, + "content": "33. OpenAI: Introducing chatgpt (2023), https://openai.com/blog/chatgpt" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.801, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniACT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "35. Rawles, C., Li, A., Rodriguez, D., Riva, O., Lillicrap, T.: Android in the wild: A large-scale dataset for android device control (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "36. Ren, S., Guo, D., Lu, S., Zhou, L., Liu, S., Tang, D., Sundaresan, N., Zhou, M., Blanco, A., Ma, S.: Codebleu: a method for automatic evaluation of code synthesis. arXiv preprint arXiv:2009.10297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.288 + ], + "angle": 0, + "content": "37. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.785, + 0.358 + ], + "angle": 0, + "content": "38. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.785, + 0.401 + ], + "angle": 0, + "content": "39. Shaw, P., Joshi, M., Cohan, J., Berant, J., Pasupat, P., Hu, H., Khandelwal, U., Lee, K., Toutanova, K.: From pixels to ui actions: Learning to follow instructions via graphical user interfaces. arXiv preprint arXiv:2306.00245 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.785, + 0.442 + ], + "angle": 0, + "content": "40. Shi, T., Karpathy, A., Fan, L., Hernandez, J., Liang, P.: World of bits: An open-domain platform for web-based agents. In: International Conference on Machine Learning, pp. 3135-3144. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.444, + 0.785, + 0.472 + ], + "angle": 0, + "content": "41. Sridhar, A., Lo, R., Xu, F.F., Zhu, H., Zhou, S.: Hierarchical prompting assists large language model on web navigation. arXiv preprint arXiv:2305.14257 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "42. Sun, L., Chen, X., Chen, L., Dai, T., Zhu, Z., Yu, K.: Meta-gui: Towards multimodal conversational agents on mobile gui. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 6699-6712 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.543 + ], + "angle": 0, + "content": "43. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.585 + ], + "angle": 0, + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.613 + ], + "angle": 0, + "content": "45. team, W.E.: InstructPalmyra-30b : Instruct tuned Palmyra-Large model. https://dev.writer.com (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.642 + ], + "angle": 0, + "content": "46. team, W.E.: Palmyra-base Parameter Autoregressive Language Model. https://dev.writer.com (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "48. Wang, B., Li, G., Zhou, X., Chen, Z., Grossman, T., Li, Y.: Screen2words: Automatic mobile ui summarization with multimodal learning. In: The 34th Annual ACM Symposium on User Interface Software and Technology. pp. 498-510 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "49. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al.: A survey on large language model based autonomous agents. arXiv preprint arXiv:2308.11432 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.812 + ], + "angle": 0, + "content": "50. Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems 33, 5776-5788 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "51. Wang, X., Chen, G., Qian, G., Gao, P., Wei, X.Y., Wang, Y., Tian, Y., Gao, W.: Large-scale multi-modal pre-trained models: A comprehensive survey (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.128 + ], + "angle": 0, + "content": "R. Kapoor, Y. Butala et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.245 + ], + "angle": 0, + "content": "53. Xu, N., Masling, S., Du, M., Campagna, G., Heck, L., Landay, J., Lam, M.: Grounding open-domain instructions to automate web support tasks. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 1022-1032 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "54. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "55. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.355 + ], + "angle": 0, + "content": "56. Yao, S., Chen, H., Yang, J., Narasimhan, K.: Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems 35, 20744-20757 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "58. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "59. Zhang, Z., Zhang, A.: You only look at screens: Multimodal chain-of-action agents (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.48 + ], + "angle": 0, + "content": "60. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "61. Zhou, S., Alon, U., Agarwal, S., Neubig, G.: Codebertscore: Evaluating code generation with pretrained models of code. arXiv preprint arXiv:2302.05527 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "62. Zhou, S., Xu, F.F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Bisk, Y., Fried, D., Alon, U., et al.: Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.55 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_origin.pdf b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7d40b0b94ed5a3f3eec00ba7397f451c2b06dd6e --- /dev/null +++ b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/4d3b02b6-e05d-4c83-bcd1-dea853d30a58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e13bc9fa0525e4614dd03a15300cd73acdbfeb7758bcb7f465fa43f14e1dc954 +size 6907657 diff --git a/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/full.md b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a008928153bac1449953c6ba95818fcaf64aea8f --- /dev/null +++ b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/full.md @@ -0,0 +1,326 @@ +# OmniACT: A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web + +Raghav Kapoor1*, Yash Parag Butala1*, Melisa Russak2, Jing Yu Koh1, Kiran Kamble2, Waseem AlShikh2, and Ruslan Salakhutdinov1 + +1 Carnegie Mellon University + +$^{2}$ Writer.com + +{raghavka,ypb}@cs.cmu.edu + +https://huggingface.co/datasets/Writer/omniact + +![](images/fb400ca699f9d9bd89b67da7f139d9ad5d2f0a68d5306cf940f4420398338896.jpg) +Fig. 1: OmniACT dataset and benchmark for enabling autonomous human-computer interaction agents. The left shows an image paired with a natural language task description as the input, the right shows the resulting action script to be executed on the screen. Examples are presented from Stocks, Apartments, and Weather application. + +# Task Description + +Check the rate of change in Google's stock price over last one month + +Find Rental Properties in Seattle, Washington + +Using the popup opened, scroll over to find weather in Chicago on 18th September + +# Action Script + +import pyautogui + +pyautogui.moveTo(210, 179) +pyautoguidragTo(243, 179, button="left") + +import pyautogui + +pyautogui.click(398,167) pyautogui.write("Seattle,WA") pyautogui.write("enter") + +import pyautogui + +pyautogui.goiveFo(410, 207) +pyautogui.hscroll(10) +pyautogui.clclick(410, 207) + +# Illustration + +![](images/fbff3ff17c5112751dfce0ff6d8f3cb2e02d617fb6b14f2b8f659187ae31e250.jpg) + +![](images/3a67c4c9abf4c04345eae133c5acfb623d982b7ab80e57194b856869ab0cff89.jpg) + +![](images/61ca544866e92d7d2bf7f122d060d7b52459617b33e9c37ec3b129c83b8a31c6.jpg) + +![](images/6e3c7cb3e88bffc5744b780c3183e94f4fc7ca0c664132376252b5207140f516.jpg) + +Abstract. For decades, human-computer interaction has fundamentally been manual. Even today, almost all productive work done on the computer necessitates human input at every step. Autonomous virtual agents represent an exciting step in automating many of these menial tasks. Virtual agents would empower users with limited technical proficiency to harness the full possibilities of computer systems. They could also enable the efficient streamlining of numerous computer tasks, ranging from calendar management to complex travel bookings, with minimal human intervention. In this paper, we introduce OmniACT, the first-of-a-kind + +dataset and benchmark for assessing an agent's capability to generate executable programs to accomplish computer tasks. Our scope extends beyond traditional web automation, covering a diverse range of desktop applications. The dataset consists of fundamental tasks such as "Play the next song", as well as longer horizon tasks such as "Send an email to John Doe mentioning the time and place to meet". Specifically, given a pair of screen image and a visually-grounded natural language task, the goal is to generate a script capable of fully executing the task. We run several strong baseline language model agents on our benchmark. The strongest baseline, GPT-4, performs the best on our benchmark. However, its performance level still reaches only $15\%$ of the human proficiency in generating executable scripts capable of completing the task, demonstrating the challenge of our task for conventional web agents. Our benchmark provides a platform to measure and evaluate the progress of language model agents in automating computer tasks and motivates future work towards building multimodal models that bridge large language models and the visual grounding of computer screens. + +Keywords: Generalist Agent $\cdot$ Multimodal Machine Learning $\cdot$ Dataset and Benchmark $\cdot$ Vision Language Understanding $\cdot$ UI grounding $\cdot$ Human-computer interaction + +# 1 Introduction + +Performing computer tasks based on natural language instructions has been a long-standing goal of artificial intelligence [49]. One concrete objective in the line of research is to develop generalist agents that can assist humans in doing computer tasks [21], such as "Order a pizza from Domino's" or "Write a message to John." The agent should be able to open the application and perform the task. Executing these actions on a personal computer involves a sequence of interactions with a mouse and keyboard. For example, the simple task of writing an email involves hovering over the application icon, clicking it, clicking the 'New Email' button, writing the content of the email, and clicking send. Successfully sending an email requires accurately predicting the correct action at each step and accurately executing it, which is a herculean task even for the best agents today [14]. + +A generalist agent for computer tasks must understand natural language instructions, process visual screenshots, and produce the correct sequence of actions to be performed to achieve the intended task. Several existing approaches focus on building agents based on the HTML model [9, 40, 62]. However, this approach introduces several challenges and constraints. These agents are limited to web applications and often struggle with complex or long-context HTML code. They cannot interact with native desktop applications or perform tasks that span multiple applications, like drafting an email using text from a code editor, without significant alterations. Furthermore, HTML-based agents, which are inherently powered by text-only language models, typically underperform in tasks requiring visual cues, such as identifying and clicking a blue button + +on a desktop's top-right corner. In contrast, humans can easily understand UI elements like dropdown menus, typable areas, redirections, and options with just a glance. + +Towards the goal of developing a generalist autonomous agent with robust visual and user interface (UI) understanding capabilities, we introduce a new task and dataset, OmniACT, containing over 9.8K pairs of images and instructions (Figure 1) across different operating systems and the web. This dataset includes screenshots of various UI screens and corresponding natural language instructions. The objective of these instructions is to generate executable commands using the PyAutoGUI Python library [1]. PyAutoGUI enables the automation of the mouse and keyboard operations, which helps to facilitate interactions with various native applications across macOS, Windows, and Linux. This simplifies completing specified tasks across different web domains and native desktop applications. + +We evaluate several language model-based agent baselines on this dataset, including LLaMA [47], Vicuna [7], Palmyra-X (43B) [2], InstructPalmyra-30B [45], GPT 3.5, and GPT-4 [32]. We experiment with fine-tuning Vicuna-13B and LLaMA-13B models using QLoRA [10]. We also benchmark multimodal baseline LLaVa-v1.5-7B, LLaVa-v1.5-13B [47], Gemini-Pro [44] and GPT-4-vision-preview [55] for the task. Our findings highlight the necessity for a multimodal model capable of executing these tasks, and our analysis provides insights into promising future work in the space. Our key contributions are outlined as follows: + +1. We release a novel dataset of desktop and website applications consisting of over 9.8K natural language tasks, UI screens, and corresponding code snippets collected through human annotation. We introduce custom performance metrics tailored for computer tasks. +2. We propose DetACT, a module for creating textual representations of the screen using signals from OCR, color, and icon-template matching. +3. We conduct a comprehensive benchmark and analysis of state-of-the-art LLMs and multimodal models on our benchmark. Our results show that OmniACT is a challenging task for even the best LLM agents today, and existing models are far below human performance. + +# 2 Related Work + +# 2.1 UI Understanding + +User interface (UI) understanding has garnered interest from researchers in the machine learning and human-computer interaction communities, evolving with various models focusing on understanding the semantics of mobile and web user interfaces. UIBert [3], PixelBERT [16], ActionBert [15], VUT [25], Screen2Words [48], WidgetCaptioning [24] and Pix2Act [39] are notable models in this area. They propose approaches for learning the user-interface semantics of the mobile screen using the image and view hierarchy. These models have demonstrated effectiveness in tasks like capability prediction, screen segmentation and + +Table 1: Comparison of OmniACT with other related benchmarks. + +
DatasetsSizeEnv TypeTask HeterogeneityReal-World PortrayalExecutional CorrectnessSupports Continuous Scale Desktop AppsAdaptive EvaluationTask
VisualWebArena [20]910WebYesYesYesNoNoWeb Navigation
WebArena [62]812WebYesYesYesNoNoWeb Navigation
Mind2Web [9]2350WebYesYesNoNoNoWeb Navigation
WebShop [56]12000 ProductsWebNoNoYesNoNoWeb Navigation
RUSS [53]80WebYesYesNoNoNoWeb Navigation
WebSRC [6]2735WebYesYes-NoNoQA
MiniWoB++ [17]100Mobile WebsitesNoNoYesNoNoWeb Navigation
PixelHelp [23]187MobileYesYesNoNoNoUI Grounding
MetaGUI [42]1125MobileYesYesYesNoNoMobile Navigation
MoTIF [5]756MobileYesYesYesNoNoMobile Navigation
AITW [35]715142Mobile and WebYesYesYesNoNoMobile/Web Navigation
OmniACT (Ours)9802Desktop and WebYesYesYesYesYesCode Generation
+ +understanding, and screen caption generation. Lexi [4] and Spotlight [22] propose models that use vision-only inputs to minimize the reliance on metadata such as view hierarchy. Furata et al. [11] demonstrates the use of fine-tuning for multimodal web navigation. The majority of machine learning models trained for UI understanding leverage the Rico dataset [8] and its extensions, which contain 64,462 unique Android screens and metadata. In addition, [4] released the UICaptions dataset, which consists of diverse image-captions pairs across a wide range of applications. PixelHelp [23] also released a corpus to train models that can interpret natural language instructions and map them to mobile UI actions. + +# 2.2 Autonomous Computer Agents + +The advent of large language models (LLMs) has been pivotal in the rapid advancement of agents that operate on web pages. Recent research such as ViperGPT [43] Chameleon [29], RCI Agent [18], VisProg [12], and [31] employ LLMs for planning or action prediction in developing autonomous agents. Benchmark datasets, such as MiniWoB [40], WebShop [56], Macaw-LLM [30], ASH-Prompting [41] Mind2Web [9], WebArena [62], AgentBench [28] and VisualWebArena [20] have also been proposed to measure the ability of LLM-based agents to automate web tasks. These methods mainly involve agents that operate on a text-based Document Object Model (DOM) of HTML scripts. This limits their understanding of screen context, which is crucial for the model's decision-making and action-taking processes. To address this limitation, [35] released Android in the Wild, a dataset comprising screens, natural language instructions, and corresponding actions. Following this, [59] proposed a multimodal model, AutoUI, which is designed to build an agent on the Android in the Wild dataset confined to the Android ecosystem. WebAgent [13] utilized Flan-U-PaLM, for grounded code generation, and HTML-T5 and showed improvement on real-world websites. + +Current benchmarks for autonomous agents focus mainly on the Web or Android environments, posing challenges for tasks involving desktop applications or spanning multiple applications beyond the web domain. The absence of established benchmarks and datasets in this area, coupled with basic methods + +for extracting user interface (UI) elements, underscores the need for significant progress in developing more versatile autonomous agents capable of handling diverse tasks beyond the current scope. To highlight the unique features that OmniACT introduces in the assessment of capable autonomous agents, we provide a comparison between the existing benchmarks and our proposed benchmark, OmniACT, in Table 1. + +![](images/3637b64b82edc539bc8230362a244c2b1b77bf9d9c6277baaa4ab34339e3319e.jpg) +Fig. 2: Data Collection Pipeline. (1) We select over 60 applications and websites to ensure diversity, (2) segment the screen through human-annotated bounding boxes, (3) label the bounding boxes based on functionality, (4) ask student volunteers to come up with tasks, given a screen image, and (5) reverse map the textual labels to coordinates and filter the scripts based on execution and syntax. + +# 3 OmniACT + +We introduce a novel dataset and benchmark, OmniACT, which measures the performance of autonomous agents on both web and desktop applications. Compared to previous benchmarks which focus on text-based reasoning [9, 17, 40, 56, 62], our benchmark aims to measure multimodal agents that bridge large language model planners and UI understanding vision models. OmniACT can be accomplished as a standalone task as it is not under a mock environment. + +All actions that a human can execute on the computer can be encoded in the PyAutoGUI [1] Python framework. This framework allows a user to execute keyboard and mouse operations by running Python code. The PyAutoGUI code to execute these tasks is shown in the third column of Figure 1. For other computer tasks, the PyAutoGUI library provides functions such as 'press', 'write', and 'scroll' which can be used to execute the task. Our dataset consists of parallel data of natural language tasks, UI screenshots, and ground truth PyAutoGUI scripts that achieve successful execution. + +# 3.1 Task Formulation + +Given an input state of a computer defined by the screen $S$ and the task description $T$ in natural language, the goal of the task is to output a sequence of actions $A$ that can successfully accomplish the task $T$ within a screenshot $S \in \{ \text{Linux, Windows, MacOS, Webpage} \}$ . Formally, the task can be defined as learning the transition function $f: T \times S \to A$ . During dataset collection, we ensure that all task descriptions $T$ are feasible and can be accomplished in the current screenshot $S$ . To reduce ambiguity and facilitate better evaluation, we ensure that task descriptions are detailed and unambiguous. Tasks can also be visually grounded (e.g., 'Click the red button to start recording') or natural language based (e.g., 'Click the My Account button'). We define the action space using the functionalities in the PyAutoGUI library: $A \in \{\text{'click', 'dragTo', 'scroll', 'write', \ldots}\}$ . The exhaustive list of actions is provided in Table 2. Our action space is much larger than other benchmarks [9,40,62] that resort to two or three interaction options. Mouse actions such as 'moveTo', 'click', 'rightClick', 'doubleClick', and 'dragTo', additionally require screen coordinates as arguments, which indicate the pixel location of the action. + +Figure 1 illustrates sample tasks and corresponding outputs for three applications within OmniACT: (1) Stocks (MacOS), (2) Apartments.com (web page), and (3) Weather (MacOS). The first column depicts the input image, and the second column shows the natural language task that is to be executed on the current screen. To execute these tasks, a user must accurately perform a series of operations using the mouse and keyboard. Eg: to check the rate of change in Google's stock price over the last month, the mouse has to be moved to the last month and dragged while holding the left-click button to the current month. + +# 3.2 Dataset Preparation + +To prepare our dataset, we followed a pipelined approach, as summarized in Figure 2. We first selected a variety of applications and websites. For each application or website, we created bounding boxes around key UI elements and labeled them according to their functionality, which is crucial for assisting human annotators in writing accurate PyAutoGUI scripts. After each script is written, we converted the labels back into numeric coordinates, allowing us to align the scripts precisely with the locations of the UI elements. Finally, we thoroughly reviewed each script, focusing on its executability and adherence to syntax standards. This ensured the high quality and functionality of our dataset, making it a valuable resource for training and evaluating autonomous agents. + +Application/Website Selection To test the computer agents' generalization ability across different tasks, we collect tasks across multiple domains on both desktop and web applications. In total, we collect and annotate 9802 data points (Table 3), with the split between desktop and web applications approximately 3:1. The emphasis on desktop applications, which do not contain Document Object Model (DOM) hierarchies unlike HTML-based web pages, presents a more + +complex multimodal challenge where visual cues are crucial. We collect tasks from applications within the three most popular operating systems. We select 22 native applications from MacOS, and 8 each from Linux and Windows. We annotate roughly 3 to 4 screens for every application. The full list of applications is provided in the Appendix. + +Many common computer tasks today are still performed through web applications, so we also collect 3-4 screenshots from 27 different web applications. To ensure diversity in task intents, we categorize these tasks into one of the following 6 categories: (1) Shopping, (2) Entertainment, (3) Service, (4) Government, (5) Travel, (6) Health. Inspired by the methodology of [9], these categories were selected to cover a wide range of user intents and functionalities. + +Table 2: Action types supported by OmniACT and the number of instances for each action in the dataset. + +
TypeAction%
MouseClick63.73
Double Click0.58
Right Click0.77
Move/Hover1.85
Drag0.29
Scroll1.68
Horizontal Scroll0.17
KeyboardPress16.28
Hotkey3.00
Write11.65
+ +UI Screen Segmentation To collect gold-standard data, we first annotate and segment the screen by identifying the + +bounding boxes present on the screen. We employ slightly different techniques for web and desktop applications to create the bounding boxes: + +1. Desktop Applications: We build a custom annotation interface based on PyQt5 $^3$ to create bounding boxes manually over a screen image using a simple drag-and-click mechanism. This custom interface expedites the process and allows us to get highly accurate gold-label data for desktop images. +2. Websites: For webpages, we write JavaScript code to extract all interactable (click, type, etc.) regions from HTML source code. We also extract banners, dropowns, submit, and radio buttons from the screen. We filter the elements to retain only those that are visible and interactable within the screen. + +Functionality Tagging To map each bounding box to its correct functional description, we leverage Amazon MTurk workers (see details in Appendix), who are given an image with a bounding box and are required to write the correct description or label of the bounding box's function. For example, given an image of an Amazon webpage with a search bar, the annotator labels it as "find-product-search-bar". The logical descriptions are used to create tasks in a structured manner without the need to identify individual bounding box coordinates. + +Task Creation Our approach for each screen involves utilizing all human-annotated bounding boxes and their labels to create tasks that can be executed + +![](images/91bc6bf190024aeb93b022e0af17e277beeacd4e3c1c9b2f438622bc2f6c41f0.jpg) +Fig. 3: DetACT Module. Given an initial image and a natural language task description, we use a pipelined approach to run OCR and SAM on the screen. The outputs from SAM are then used by icon and color-matching modules to obtain an exhaustive set of useful UI elements. The list of elements is passed through LLM based filter to select only the elements related to the given task. + +within the confines of a single screen. These tasks are designed to be visually grounded in order to measure the capabilities of multimodal agents. We plan to release the bounding box and their corresponding labels as the metadata for evaluation purposes. + +For dataset compilation, college students with basic Python programming skills served as annotators, accessing API references for PyAutoGUI and examples of potential tasks. Each student generated multiple tasks, each accompanied by three alternative natural language reformulations. For instance, "What is $3 + 2$ ?" might be reformulated as "Calculate the sum of 2 and $3$ " or "Add two to three". To avoid train-test leakage, rephrased tasks were consistently placed in the same dataset split. Further details on the annotation process are available in the Appendix. + +Reverse Mapping and Filtering To ensure high-quality data, we incorporate an additional step into the data collection pipeline. We build scripts to map the text-based labels of each bounding box back to their numeric coordinates, and then match the syntax and verify if the task will be executed on the screen. Using this filter, we remove all the non-working or syntactically incorrect data points and finally manually review the set of tasks. + +After filtering, we obtain 9802 human-annotated, gold-label data points across more than 200 desktop and web screens + +Table 3: Dataset distribution across splits and platforms. + +
DomainTrain ValidationTestTotal
DesktopMac OS3028444786
Linux761126247
Windows1573216458
Web-1427206530
Total67899922,021
+ +(Table 3), split into train, validation, and test sets in a 7:1:2 ratio. All collected data will be publicly released to encourage future work on multimodal agents. + +# 4 Evaluation Metrics + +In this section, we detail various evaluation metrics for benchmarking model performance on the OmniACT dataset. UI screens have additional constraints such as spatial relevance which are not factored in most conventional similarity-based metrics such as BLEU [34], CodeBLEU [36], BERTScore [58] and CodeBERTScore [61]. For example, a valid click action is usually not constrained to a single coordinate but can be any coordinate within a specified region. In the event of invalid coordinate predictions, an agent that predicts coordinates further away from the valid region should invoke a higher penalty compared to an agent that predicted coordinates close to the region. We propose two new metrics adapted: Sequence Score (Section 4.1) and Action Score (Section 4.2) aimed at utilizing UI information. + +# 4.1 Sequence Score + +The sequence score measures whether the predicted action sequence (e.g., 'click', 'write', 'press') exactly matches the gold sequence. Since predicting the first action in the sequence is relatively straightforward and later actions are more difficult, we define sequence score as follows: + +$$ +\operatorname {S e q S c o r e} _ {i} = \left\{ \begin{array}{l l} \beta_ {1} + \beta_ {2} * (s - 1) & \text {i f a l l a c t i o n s m a t c h} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {1} +$$ + +where $s$ is the action sequence length, $\beta_{1}$ is set to 0.1 and $\beta_{2}$ is set to 1. + +# 4.2 Action Score + +The action score measures how well a code snippet containing the correct action sequence can perform the task. Specifically, for a script with a correct action sequence, we introduce penalties for inaccurate behavior. The penalties are described below: + +1. Click penalty $(M)$ : For actions 'click', 'rightClick', 'doubleClick', 'moveTo', and 'dragTo', we penalize code snippets where predicted coordinates lie outside of the bounding box of the UI element. The click penalty for the $j^{th}$ action of the $i^{th}$ example is defined as: + +$$ +M _ {i} ^ {j} = \alpha_ {i} \times \left\{ \begin{array}{l l} 1 - \frac {\mu}{\mu + L _ {2}} & \text {i f S e q S c o r e} _ {i} > 0 \\ 1 & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +Here $L_{2}$ corresponds to the smallest Euclidean distance between the predicted coordinate and bounding box. $L_{2}$ is zero when the predicted coordinate lies within the target bounding box. $\mu$ is the Dirichlet smoothing + +coefficient which we dynamically set to the inverse of the length of the diagonal of the bounding box. This ensures that the penalty for points outside the bounding box varies based on the size of the bounding box. For two predicted points with the same $L_{2}$ , the metric penalizes more heavily if the box is larger. This is sound with the intuition that the chances of clicking on a larger box are higher and should be penalized more in case of a mistake. + +2. Key penalty $(K)$ : For actions 'press' and 'hotkey', we check whether the set of keys in the target code (represented as $GK_{i}^{j}$ ) and predicted code (represented as $PK_{i}^{j}$ ) are the same. It is formally defined as: + +$$ +K _ {i} ^ {j} = \alpha_ {i} \times \left\{ \begin{array}{l l} 0 & \text {i f} G K _ {i} ^ {j} = P K _ {i} ^ {j} \text {a n d S e q S c o r e} _ {i} > 0 \\ 1 & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +3. Write penalty $(W_{p})$ : For action type 'write', we penalize the output for the sentence to be typed. Specifically, we the employ BLEU score [34], and compute: + +$$ +W _ {i} ^ {j} = \alpha_ {i} \times \left\{ \begin{array}{l l} 1 - B L E U \left(G S _ {i} ^ {j}, P S _ {i} ^ {j}\right) & \text {i f S e q S c o r e} _ {i} > 1 \\ 1 & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +Here, $GS_{i}^{j}$ represents the actual sentence to be typed, and $PS_{i}^{j}$ represents the sentence predicted by the model in the $j^{th}$ action of example $i$ . + +In the above equations, $(\alpha_{i})$ is the weighting factor: + +$$ +\alpha_ {i} = \operatorname {S e q S c o r e} _ {i} / \text {l e n g t h o f s e q u e n c e} i \tag {5} +$$ + +This ensures that the action score $\in [0,1]$ . The mean action score is calculated as follows: + +$$ +\text {A c t i o n S c o r e} = \frac {\sum_ {i} \max \left(\operatorname {S e q S c o r e} _ {i} - \sum_ {j} \left(M _ {i} ^ {j} + K _ {i} ^ {j} + W _ {i} ^ {j}\right) , 0\right)}{\sum_ {i} \operatorname {S e q S c o r e} _ {i}} \tag {6} +$$ + +# 5 DetACT: DETecting ACTions from UI + +Understanding UI screens is crucial for multimodal computer tasks. Web-based agents typically use language-only inputs from the HTML DOM. This is insufficient for comprehending the full extent of an application UI, as many components may not be easily described with HTML code. To address this, we propose DetACT, which allows us to convert images of UI layouts into structured code and text outputs for a downstream LLM. DetACT is a system comprised of three distinct modules: the text module, the icon module, and the color module. + +1. Text Extraction: We use the EasyOCR model4 to parse over the UI screens and collect all text-based elements. Along with the text, we also note the locations of each of these elements. This is depicted in Figure 3, along with a list of text elements found on the screen using the OCR Module. We segment and classify the different regions within the screenshot using the + +![](images/877231d43634aee2879a1c206ac0fe7070408c08015ff5c3ce1a8c4b850a6958.jpg) +Fig. 4: Baseline Model Architecture. Image and task descriptions are sent to DetACT module, which gives a filtered list of UI elements relevant to feed into the prompt along with the task. We also show the prompt structure used for action script generation. This structure is passed through the LLM (along with the image for multimodal LLM) to generate the automation script. + +Segment Anything Model (SAM) [19]. From the outputs, we filter out the non-textual segments for our icon and color detection. + +2. Icon Module: For matching with the appropriate icon, we use a pack of 1600 icons5 as templates. Each of these icons is labeled with their appropriate functionality and is matched with the filtered outputs SAM [19]. For the similarity of the two images, we resize the reference icons and segmented region of interest (ROI) to the same size, and convert both images to grayscale. After this, we use the Structural Similarity Index (SSIM) [52], to find the closest match of the ROI to the icons in our set, and select the ones above the SSIM threshold of 0.95. As seen in Figure 3, a few icons matched on the screen are Globe icon, Calendar icon, Person icon, and Location icon; each depicting a different use case. +3. Color Module: Finally, to place all segments of interest into appropriate buckets of colors, we average the RGB pixel values over the ROI and, based on that value, bucket them into different color categories. We categorize colors differently based on the human perspective of the ranges of each color. To avoid ambiguity, we consider eleven major colors, namely yellow, blue, green, red, pink, violet, white, black, orange, brown, and grey. We record the center of the element along with the color. + +Once all the elements of each category are extracted with their coordinates, we then filter these UI elements by prompting GPT-4 [32]. We ensure that the elements selected are suited only for our task, for which we also provide the task description in our prompts along with the list of elements. Full details of the prompt are provided in the appendix section of the paper. As we observe in Figure 3, given an image from the Expedia application, and a task ("Click on the Black Location icon and enter the destination as Paris"), the LLM filters out the elements to retain only "Going To", "Location Icon", and the Black colored elements from the screen. This is passed as input to the LLM or VLM backbone. + +# 6Baselines + +To evaluate the performance of existing language model-based agents, we conduct experiments with both language-based and multimodal baselines. The DetACT module takes in image and text descriptions of the task and outputs the color, icon, and text-based signals. This is concatenated to the prompt for the LLM prompt-based baselines (see Figure 4). Every prompt starts with a role assignment [60], followed by the detailed API reference of the PyAutoGUI function set, along with a textual description of their function. We then add five in-context examples from the training set that most closely match the task (based on the cosine similarity of the MiniLM [50] embeddings of the reference task and the train examples). We add a list of UI elements filtered by the DetACT module to the prompt. Finally, we provide the rules with the task description. For multimodal baselines, we also pass the image pixels to the vision encoder. We choose coordinate-based UI elements in the prompt as recent techniques like the Set-of-Mark (SOM) [54] prompting does not work for desktop settings since it is difficult to obtain interactive elements from the desktop screen images. We report the results of several baselines: + +- Few-shot Generative LLM: We experiment with models from LLaMA-2 [47], Vicuna-1.5 [7], CodeLLaMA-34B [37], Palmyra [46], and GPT [32] series. We use the prompts structure as shown in Figure 4 to prompt the model. For LLaMA and CodeLLaMa, we reduce the prompt length to 2000 tokens by removing outputs from the DetACT module with lower confidence, as we observed poor performance on longer prompts. For the other models, we allow prompts with up to 4000 token sizes. +- Finetuned Generative LLM: We fine-tuned the LLaMA-13B model and Vicuna-13B using QLoRa [10] with rank 64 and scaling factor 16 for 300 steps to generate the code given screen description from the DetACT module and the instruction. +- Few-shot Generative Multimodal Models: As OmniACT is predominantly multimodal, with a majority of tasks being visually grounded, we conduct experiments with large multimodal models. Given the limited research in this domain [51, 57], there is a scarcity of available multimodal models with significant size adept for this task. Here, we experiment with [26, 27], providing a similar prompt as well as the screen image. + +# 7 Results and Analysis + +As shown in Table 4, we experiment with three different categories of models, namely Prompt-based LLMs, Fine-tuned LLMs, and Prompt-based Multimodal Models. GPT-4 is the best-performing approach, scoring higher on the sequence score and invoking lower penalties on coordinate predicting and text input. For prompt-only LLMs, the GPT-3.5-turbo and GPT-4 models outperform the other LLM baselines, including the LLaMA [47] and Vicuna [7] models. We observe that CodeLLaMA-34B [38], which is trained for code generation, also achieves a higher performance than other models of the same size at predicting the action sequences. + +Fine-tuned models also perform much better than their few-shot prompt-only counterparts. Fine-tuning substantially improves LLaMA-13B's sequence score (4.80 to 8.92) and action score (1.62 to 2.14), as well as the other metrics. Despite this, we observed that both, prompt-based LLMs and finetuned LLMs face severe mouse penalties, especially on click coordinates. This is because they rely solely on text-based signals. + +Table 4: Baseline Performance. (A) Prompt-only LLMs, (B) Fine Tuned LLMs, (C) Prompt-only Multimodal Models. The table represents the Sequence score (SS), click penalty $(M_p)$ , Key penalty $(K_p)$ , Write Penalty $(W_p)$ , and Action Score (AS). The best results for the (SS) and (AS) are highlighted. + +
ModelSS(↑)MpKpWpAS(↑)
Prompt based LLMs
LLaMA-7B [47]4.121.241.830.570.48
Vicuna-7B [7]3.881.171.510.430.77
LLaMA-13B [47]4.801.320.930.931.62
Vicuna-13B [7]5.441.650.941.061.78
Palmyra-Instruct-30B [45]7.515.680.120.401.31
CodeLLaMA-34B [38]10.092.992.710.663.72
Palmyra-X 43B [2]11.203.123.022.122.94
GPT-3.5-turbo-0613 [33]22.858.134.512.317.89
GPT-4 [32]32.7510.276.993.8911.60
Finetuned LLMs
LLaMA-13B FT8.924.611.430.742.14
Vicuna-13B FT8.784.121.310.632.72
Multimodal LLMs
LLaVA-v1.5-7B [27]13.234.731.241.445.82
LLaVA-v1.5-13B [26]20.566.073.442.858.19
Gemini-Pro [44]30.989.056.813.6611.46
GPT-4V [26]38.7210.537.144.0317.02
Human Performance82.230.120.361.6180.14
+ +To address this, we experiment with multimodal language models (Table 4). We observe that the coordinate prediction improves significantly when we provide the entire image as input to the multimodal LLM, as this enables it to fully utilize the screen representation. In addition to open sourced models, we also experiment with the GPT-4-vision API [55] which shows that GPT-4 Vision [55] outperforms GPT-4 significantly on the Action Score along with improving the sequence score, which we attribute to the strong reasoning abilities of GPT-4 coupled with the improved visual understanding capabilities of the GPT-4-vision model [55]. These findings pave the way towards exciting new research directions on building multimodal models for long-horizon planning and code generation. + +Human performance over the task: OmniACT consists of visually complicated tasks, and tests various types of computer skills. In order to get a gauge of how well humans perform, we collect evaluation data from human evaluators. We split the test set uniformly amongst 10 human evaluators, and provided them with the screenshot and task instruction. We record the actions taken by the an- + +notators, and measure their performance on our predefined metrics (Table 4). We find that users generally exhibit a high level of proficiency when attempting most tasks for the first time. However, there are instances where users face difficulties in successfully completing certain tasks. These are due to factors including the user's inability to fully comprehend the task, difficulties in grounding the task to the provided screenshot, or a lack of familiarity with the UI. + +# 8 Conclusion and Future Work + +Autonomous virtual agents offer the potential to automate routine tasks, benefiting users with limited technical expertise. To solve this task, we introduce OmniACT, a unique dataset of 9.8K human-labeled data points. OmniACT benchmarks autonomous agents across a range of tasks on web and desktop applications. LLM-based agents, like GPT-4, achieve a respectable action score of 11.6 on our dataset. However, OmniACT presents a challenge for the current state-of-the-art language and multimodal models. It provides a direction for future research on foundational multimodal models that seamlessly integrate language and visual understanding of computer screens and stands poised to drive the next wave of advancements in generalist autonomous agents offering omnipotent assistance to humans. + +# 9 Limitations + +This work introduces a valuable dataset, yet we recognize a few limitations that exist. State-of-the-art models like GPT-4, may exhibit susceptibility to hallucinations and bias towards specific data types, hindering broad applicability. Reliance on closed models like GPT-4V poses integration challenges due to high costs and time constraints. Despite efforts for equal representation and data collection without personal information, biases may be introduced as the dataset is exclusively in English, and human-curated content may have temporal biases. + +# 10 Ethics Statement + +As a part of the dataset creation process, we carefully review the pipeline at every stage, ensuring there is no personally identifiable information or offensive content, either during data collection or through the use of LLMs. For all purposes, we create dummy accounts that mimic real-like user content. To get the gold labels scripts we seek help from well-qualified student workers, approved through the institution, and get the bounding box data annotated through MTurk workers, both of whom are paid $25 per hour, which is greater than the minimum wage rate (We detail this process in the supplementary material). Human studies are also done with the help of student workers approved by the institution at the above-mentioned payscale. We also ensure that all groups have equitable representation and that no personal opinions are reflected in the dataset, avoiding bias during the collection as well as the annotation process. + +# Acknowledgements + +We extend our heartfelt gratitude to Writer.com for their generous and unwavering support throughout this project. Their dedicated team's expertise and collaboration were invaluable in achieving our goals. + +# References + +1. Pyautogui: A cross-platform gui automation python module for human beings. https://github.com/asweigart/pyautogui (2023) +2. AlShikh, W., Daaboul, M., Goddard, K., Imel, B., Kamble, K., Kulkarni, P., Russak, M.: Becoming self-instruct: introducing early stopping criteria for minimal instruct tuning (2023) +3. Bai, C., Zang, X., Xu, Y., Sunkara, S., Rastogi, A., Chen, J., y Arcas, B.A.: Uibert: Learning generic multimodal representations for ui understanding (2021) +4. Banerjee, P., Mahajan, S., Arora, K., Baral, C., Riva, O.: Lexi: Self-supervised learning of the ui language (2023) +5. Burns, A., Arsan, D., Agrawal, S., Kumar, R., Saenko, K., Plummer, B.A.: Mobile app tasks with iterative feedback (motif): Addressing task feasibility in interactive visual environments +6. Chen, X., Zhao, Z., Chen, L., Ji, J., Zhang, D., Luo, A., Xiong, Y., Yu, K.: Websrc: A dataset for web-based structural reading comprehension. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 4173-4185 (2021) +7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/ +8. Deka, B., Huang, Z., Franzen, C., Hibschman, J., Afergan, D., Li, Y., Nichols, J., Kumar, R.: Rico: A mobile app dataset for building data-driven design applications. In: Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology. p. 845-854. UIST '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3126594.3126651, https://doi.org/10.1145/3126594.3126651 +9. Deng, X., Gu, Y., Zheng, B., Chen, S., Stevens, S., Wang, B., Sun, H., Su, Y.: Mind2web: Towards a generalist agent for the web. arXiv preprint arXiv:2306.06070 (2023) +10. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms (2023) +11. Furuta, H., Nachum, O., Lee, K.H., Matsuo, Y., Gu, S.S., Gur, I.: Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854 (2023) +12. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023) +13. Gur, I., Furuta, H., Huang, A., Safdari, M., Matsuo, Y., Eck, D., Faust, A.: A real-world webagent with planning, long context understanding, and program synthesis (2024) + +14. Gur, I., Rueckert, U., Faust, A., Hakkani-Tur, D.: Learning to navigate the web. In: International Conference on Learning Representations (2018) +15. He, Z., Sunkara, S., Zang, X., Xu, Y., Liu, L., Wichers, N., Schubiner, G., Lee, R., Chen, J., y Arcas, B.A.: Actionbert: Leveraging user actions for semantic understanding of user interfaces (2021) +16. Huang, Z., Zeng, Z., Liu, B., Fu, D., Fu, J.: Pixel-bert: Aligning image pixels with text by deep multi-modal transformers (2020) +17. Humphreys, P.C., Raposo, D., Pohlen, T., Thornton, G., Chhaparia, R., Muldal, A., Abramson, J., Georgiev, P., Santoro, A., Lillicrap, T.: A data-driven approach for learning to control computers. In: International Conference on Machine Learning. pp. 9466-9482. PMLR (2022) +18. Kim, G., Baldi, P., McAleer, S.: Language models can solve computer tasks. arXiv preprint arXiv:2303.17491 (2023) +19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything (2023) +20. Koh, J.Y., Lo, R., Jang, L., Duvvur, V., Lim, M.C., Huang, P.Y., Neubig, G., Zhou, S., Salakhutdinov, R., Fried, D.: Visualwebarena: Evaluating multimodal agents on realistic visual web tasks (2024) +21. LeCun, Y.: A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27 (2022) +22. Li, G., Li, Y.: Spotlight: Mobile ui understanding using vision-language models with a focus (2023) +23. Li, Y., He, J., Zhou, X., Zhang, Y., Baldridge, J.: Mapping natural language instructions to mobile ui action sequences (2020) +24. Li, Y., Li, G., He, L., Zheng, J., Li, H., Guan, Z.: Widget captioning: Generating natural language description for mobile user interface elements (2020) +25. Li, Y., Li, G., Zhou, X., Dehghani, M., Gritsenko, A.: Vut: Versatile ui transformer for multi-modal multi-task user interface modeling (2021) +26. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023) +27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023) +28. Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., Zhang, S., Deng, X., Zeng, A., Du, Z., Zhang, C., Shen, S., Zhang, T., Su, Y., Sun, H., Huang, M., Dong, Y., Tang, J.: Agentbench: Evaluating llms as agents (2023) +29. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models (2023) +30. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023) +31. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback +32. OpenAI: Gpt-4 technical report (2023) +33. OpenAI: Introducing chatgpt (2023), https://openai.com/blog/chatgpt +34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002) + +35. Rawles, C., Li, A., Rodriguez, D., Riva, O., Lillicrap, T.: Android in the wild: A large-scale dataset for android device control (2023) +36. Ren, S., Guo, D., Lu, S., Zhou, L., Liu, S., Tang, D., Sundaresan, N., Zhou, M., Blanco, A., Ma, S.: Codebleu: a method for automatic evaluation of code synthesis. arXiv preprint arXiv:2009.10297 (2020) +37. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023) +38. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023) +39. Shaw, P., Joshi, M., Cohan, J., Berant, J., Pasupat, P., Hu, H., Khandelwal, U., Lee, K., Toutanova, K.: From pixels to ui actions: Learning to follow instructions via graphical user interfaces. arXiv preprint arXiv:2306.00245 (2023) +40. Shi, T., Karpathy, A., Fan, L., Hernandez, J., Liang, P.: World of bits: An open-domain platform for web-based agents. In: International Conference on Machine Learning, pp. 3135-3144. PMLR (2017) +41. Sridhar, A., Lo, R., Xu, F.F., Zhu, H., Zhou, S.: Hierarchical prompting assists large language model on web navigation. arXiv preprint arXiv:2305.14257 (2023) +42. Sun, L., Chen, X., Chen, L., Dai, T., Zhu, Z., Yu, K.: Meta-gui: Towards multimodal conversational agents on mobile gui. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 6699-6712 (2022) +43. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning (2023) +44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +45. team, W.E.: InstructPalmyra-30b : Instruct tuned Palmyra-Large model. https://dev.writer.com (2023) +46. team, W.E.: Palmyra-base Parameter Autoregressive Language Model. https://dev.writer.com (2023) +47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023) +48. Wang, B., Li, G., Zhou, X., Chen, Z., Grossman, T., Li, Y.: Screen2words: Automatic mobile ui summarization with multimodal learning. In: The 34th Annual ACM Symposium on User Interface Software and Technology. pp. 498-510 (2021) +49. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al.: A survey on large language model based autonomous agents. arXiv preprint arXiv:2308.11432 (2023) +50. Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems 33, 5776-5788 (2020) +51. Wang, X., Chen, G., Qian, G., Gao, P., Wei, X.Y., Wang, Y., Tian, Y., Gao, W.: Large-scale multi-modal pre-trained models: A comprehensive survey (2023) + +52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +53. Xu, N., Masling, S., Du, M., Campagna, G., Heck, L., Landay, J., Lam, M.: Grounding open-domain instructions to automate web support tasks. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 1022-1032 (2021) +54. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v (2023) +55. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023) +56. Yao, S., Chen, H., Yang, J., Narasimhan, K.: Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems 35, 20744-20757 (2022) +57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models (2023) +58. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert (2020) +59. Zhang, Z., Zhang, A.: You only look at screens: Multimodal chain-of-action agents (2023) +60. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023) +61. Zhou, S., Alon, U., Agarwal, S., Neubig, G.: Codebertscore: Evaluating code generation with pretrained models of code. arXiv preprint arXiv:2302.05527 (2023) +62. Zhou, S., Xu, F.F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Bisk, Y., Fried, D., Alon, U., et al.: Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854 (2023) \ No newline at end of file diff --git a/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/images.zip b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba8c4b14235cc6a34be48b0a29fc8c75c37f92e6 --- /dev/null +++ b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc61c2a8486f25e7bc57991751661aa0c12e0a466a04661450712d4e97dbb307 +size 371315 diff --git a/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/layout.json b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a97ae734cb409896111085ffcd166ba885bcda0 --- /dev/null +++ b/2024/OmniACT_ A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web/layout.json @@ -0,0 +1,8844 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 112, + 475, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 112, + 475, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 475, + 163 + ], + "type": "text", + "content": "OmniACT: A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 186, + 480, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 186, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 480, + 210 + ], + "type": "text", + "content": "Raghav Kapoor1*, Yash Parag Butala1*, Melisa Russak2, Jing Yu Koh1, Kiran Kamble2, Waseem AlShikh2, and Ruslan Salakhutdinov1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 246, + 219, + 367, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 219, + 367, + 230 + ], + "spans": [ + { + "bbox": [ + 246, + 219, + 367, + 230 + ], + "type": "text", + "content": "1 Carnegie Mellon University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 279, + 230, + 334, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 230, + 334, + 241 + ], + "spans": [ + { + "bbox": [ + 279, + 230, + 334, + 241 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 279, + 230, + 334, + 241 + ], + "type": "text", + "content": " Writer.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 243, + 369, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 243, + 369, + 253 + ], + "spans": [ + { + "bbox": [ + 246, + 243, + 369, + 253 + ], + "type": "text", + "content": "{raghavka,ypb}@cs.cmu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 254, + 415, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 254, + 415, + 264 + ], + "spans": [ + { + "bbox": [ + 197, + 254, + 415, + 264 + ], + "type": "text", + "content": "https://huggingface.co/datasets/Writer/omniact" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 282, + 240, + 473 + ], + "blocks": [ + { + "bbox": [ + 151, + 282, + 240, + 473 + ], + "lines": [ + { + "bbox": [ + 151, + 282, + 240, + 473 + ], + "spans": [ + { + "bbox": [ + 151, + 282, + 240, + 473 + ], + "type": "image", + "image_path": "fb400ca699f9d9bd89b67da7f139d9ad5d2f0a68d5306cf940f4420398338896.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 485, + 480, + 529 + ], + "lines": [ + { + "bbox": [ + 131, + 485, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 485, + 480, + 529 + ], + "type": "text", + "content": "Fig. 1: OmniACT dataset and benchmark for enabling autonomous human-computer interaction agents. The left shows an image paired with a natural language task description as the input, the right shows the resulting action script to be executed on the screen. Examples are presented from Stocks, Apartments, and Weather application." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 289, + 290, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 289, + 290, + 298 + ], + "spans": [ + { + "bbox": [ + 247, + 289, + 290, + 298 + ], + "type": "text", + "content": "Task Description" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 245, + 314, + 290, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 314, + 290, + 342 + ], + "spans": [ + { + "bbox": [ + 245, + 314, + 290, + 342 + ], + "type": "text", + "content": "Check the rate of change in Google's stock price over last one month" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 252, + 371, + 282, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 371, + 282, + 398 + ], + "spans": [ + { + "bbox": [ + 252, + 371, + 282, + 398 + ], + "type": "text", + "content": "Find Rental Properties in Seattle, Washington" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 247, + 426, + 291, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 426, + 291, + 460 + ], + "spans": [ + { + "bbox": [ + 247, + 426, + 291, + 460 + ], + "type": "text", + "content": "Using the popup opened, scroll over to find weather in Chicago on 18th September" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 290, + 356, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 290, + 356, + 297 + ], + "spans": [ + { + "bbox": [ + 321, + 290, + 356, + 297 + ], + "type": "text", + "content": "Action Script" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 313, + 348, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 313, + 348, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 313, + 348, + 319 + ], + "type": "text", + "content": "import pyautogui" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 323, + 359, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 323, + 359, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 359, + 342 + ], + "type": "text", + "content": "pyautogui.moveTo(210, 179) \npyautoguidragTo(243, 179, button=\"left\")" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 368, + 347, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 368, + 347, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 347, + 373 + ], + "type": "text", + "content": "import pyautogui" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 379, + 369, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 379, + 369, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 379, + 369, + 403 + ], + "type": "text", + "content": "pyautogui.click(398,167) pyautogui.write(\"Seattle,WA\") pyautogui.write(\"enter\")" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 429, + 347, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 347, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 347, + 434 + ], + "type": "text", + "content": "import pyautogui" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 439, + 369, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 369, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 369, + 459 + ], + "type": "text", + "content": "pyautogui.goiveFo(410, 207) \npyautogui.hscroll(10) \npyautogui.clclick(410, 207)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 404, + 290, + 431, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 290, + 431, + 297 + ], + "spans": [ + { + "bbox": [ + 404, + 290, + 431, + 297 + ], + "type": "text", + "content": "Illustration" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 376, + 301, + 459, + 354 + ], + "blocks": [ + { + "bbox": [ + 376, + 301, + 459, + 354 + ], + "lines": [ + { + "bbox": [ + 376, + 301, + 459, + 354 + ], + "spans": [ + { + "bbox": [ + 376, + 301, + 459, + 354 + ], + "type": "image", + "image_path": "fbff3ff17c5112751dfce0ff6d8f3cb2e02d617fb6b14f2b8f659187ae31e250.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 377, + 358, + 459, + 388 + ], + "blocks": [ + { + "bbox": [ + 377, + 358, + 459, + 388 + ], + "lines": [ + { + "bbox": [ + 377, + 358, + 459, + 388 + ], + "spans": [ + { + "bbox": [ + 377, + 358, + 459, + 388 + ], + "type": "image", + "image_path": "3a67c4c9abf4c04345eae133c5acfb623d982b7ab80e57194b856869ab0cff89.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 380, + 390, + 456, + 407 + ], + "blocks": [ + { + "bbox": [ + 380, + 390, + 456, + 407 + ], + "lines": [ + { + "bbox": [ + 380, + 390, + 456, + 407 + ], + "spans": [ + { + "bbox": [ + 380, + 390, + 456, + 407 + ], + "type": "image", + "image_path": "61ca544866e92d7d2bf7f122d060d7b52459617b33e9c37ec3b129c83b8a31c6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 377, + 414, + 459, + 468 + ], + "blocks": [ + { + "bbox": [ + 377, + 414, + 459, + 468 + ], + "lines": [ + { + "bbox": [ + 377, + 414, + 459, + 468 + ], + "spans": [ + { + "bbox": [ + 377, + 414, + 459, + 468 + ], + "type": "image", + "image_path": "6e3c7cb3e88bffc5744b780c3183e94f4fc7ca0c664132376252b5207140f516.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 159, + 548, + 453, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 548, + 453, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 548, + 453, + 647 + ], + "type": "text", + "content": "Abstract. For decades, human-computer interaction has fundamentally been manual. Even today, almost all productive work done on the computer necessitates human input at every step. Autonomous virtual agents represent an exciting step in automating many of these menial tasks. Virtual agents would empower users with limited technical proficiency to harness the full possibilities of computer systems. They could also enable the efficient streamlining of numerous computer tasks, ranging from calendar management to complex travel bookings, with minimal human intervention. In this paper, we introduce OmniACT, the first-of-a-kind" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 455, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 455, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 455, + 665 + ], + "type": "text", + "content": "* These authors contributed equally. The order is determined by dice rolling." + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 452, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 452, + 304 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 452, + 304 + ], + "type": "text", + "content": "dataset and benchmark for assessing an agent's capability to generate executable programs to accomplish computer tasks. Our scope extends beyond traditional web automation, covering a diverse range of desktop applications. The dataset consists of fundamental tasks such as \"Play the next song\", as well as longer horizon tasks such as \"Send an email to John Doe mentioning the time and place to meet\". Specifically, given a pair of screen image and a visually-grounded natural language task, the goal is to generate a script capable of fully executing the task. We run several strong baseline language model agents on our benchmark. The strongest baseline, GPT-4, performs the best on our benchmark. However, its performance level still reaches only " + }, + { + "bbox": [ + 160, + 117, + 452, + 304 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 160, + 117, + 452, + 304 + ], + "type": "text", + "content": " of the human proficiency in generating executable scripts capable of completing the task, demonstrating the challenge of our task for conventional web agents. Our benchmark provides a platform to measure and evaluate the progress of language model agents in automating computer tasks and motivates future work towards building multimodal models that bridge large language models and the visual grounding of computer screens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "spans": [ + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "text", + "content": "Keywords: Generalist Agent " + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "text", + "content": " Multimodal Machine Learning " + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "text", + "content": " Dataset and Benchmark " + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "text", + "content": " Vision Language Understanding " + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "text", + "content": " UI grounding " + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 314, + 457, + 346 + ], + "type": "text", + "content": " Human-computer interaction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 366, + 229, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 366, + 229, + 379 + ], + "spans": [ + { + "bbox": [ + 133, + 366, + 229, + 379 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 390, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 390, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 390, + 481, + 533 + ], + "type": "text", + "content": "Performing computer tasks based on natural language instructions has been a long-standing goal of artificial intelligence [49]. One concrete objective in the line of research is to develop generalist agents that can assist humans in doing computer tasks [21], such as \"Order a pizza from Domino's\" or \"Write a message to John.\" The agent should be able to open the application and perform the task. Executing these actions on a personal computer involves a sequence of interactions with a mouse and keyboard. For example, the simple task of writing an email involves hovering over the application icon, clicking it, clicking the 'New Email' button, writing the content of the email, and clicking send. Successfully sending an email requires accurately predicting the correct action at each step and accurately executing it, which is a herculean task even for the best agents today [14]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 534, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 481, + 665 + ], + "type": "text", + "content": "A generalist agent for computer tasks must understand natural language instructions, process visual screenshots, and produce the correct sequence of actions to be performed to achieve the intended task. Several existing approaches focus on building agents based on the HTML model [9, 40, 62]. However, this approach introduces several challenges and constraints. These agents are limited to web applications and often struggle with complex or long-context HTML code. They cannot interact with native desktop applications or perform tasks that span multiple applications, like drafting an email using text from a code editor, without significant alterations. Furthermore, HTML-based agents, which are inherently powered by text-only language models, typically underperform in tasks requiring visual cues, such as identifying and clicking a blue button" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "on a desktop's top-right corner. In contrast, humans can easily understand UI elements like dropdown menus, typable areas, redirections, and options with just a glance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 282 + ], + "type": "text", + "content": "Towards the goal of developing a generalist autonomous agent with robust visual and user interface (UI) understanding capabilities, we introduce a new task and dataset, OmniACT, containing over 9.8K pairs of images and instructions (Figure 1) across different operating systems and the web. This dataset includes screenshots of various UI screens and corresponding natural language instructions. The objective of these instructions is to generate executable commands using the PyAutoGUI Python library [1]. PyAutoGUI enables the automation of the mouse and keyboard operations, which helps to facilitate interactions with various native applications across macOS, Windows, and Linux. This simplifies completing specified tasks across different web domains and native desktop applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 380 + ], + "type": "text", + "content": "We evaluate several language model-based agent baselines on this dataset, including LLaMA [47], Vicuna [7], Palmyra-X (43B) [2], InstructPalmyra-30B [45], GPT 3.5, and GPT-4 [32]. We experiment with fine-tuning Vicuna-13B and LLaMA-13B models using QLoRA [10]. We also benchmark multimodal baseline LLaVa-v1.5-7B, LLaVa-v1.5-13B [47], Gemini-Pro [44] and GPT-4-vision-preview [55] for the task. Our findings highlight the necessity for a multimodal model capable of executing these tasks, and our analysis provides insights into promising future work in the space. Our key contributions are outlined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 387, + 481, + 506 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 387, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 387, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 136, + 387, + 481, + 434 + ], + "type": "text", + "content": "1. We release a novel dataset of desktop and website applications consisting of over 9.8K natural language tasks, UI screens, and corresponding code snippets collected through human annotation. We introduce custom performance metrics tailored for computer tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 434, + 480, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 434, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 135, + 434, + 480, + 458 + ], + "type": "text", + "content": "2. We propose DetACT, a module for creating textual representations of the screen using signals from OCR, color, and icon-template matching." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 459, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 459, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 135, + 459, + 480, + 506 + ], + "type": "text", + "content": "3. We conduct a comprehensive benchmark and analysis of state-of-the-art LLMs and multimodal models on our benchmark. Our results show that OmniACT is a challenging task for even the best LLM agents today, and existing models are far below human performance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 525, + 237, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 237, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 237, + 537 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 549, + 253, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 253, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 253, + 562 + ], + "type": "text", + "content": "2.1 UI Understanding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "User interface (UI) understanding has garnered interest from researchers in the machine learning and human-computer interaction communities, evolving with various models focusing on understanding the semantics of mobile and web user interfaces. UIBert [3], PixelBERT [16], ActionBert [15], VUT [25], Screen2Words [48], WidgetCaptioning [24] and Pix2Act [39] are notable models in this area. They propose approaches for learning the user-interface semantics of the mobile screen using the image and view hierarchy. These models have demonstrated effectiveness in tasks like capability prediction, screen segmentation and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 136, + 481, + 253 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "type": "text", + "content": "Table 1: Comparison of OmniACT with other related benchmarks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 136, + 481, + 253 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 481, + 253 + ], + "type": "table", + "html": "
DatasetsSizeEnv TypeTask HeterogeneityReal-World PortrayalExecutional CorrectnessSupports Continuous Scale Desktop AppsAdaptive EvaluationTask
VisualWebArena [20]910WebYesYesYesNoNoWeb Navigation
WebArena [62]812WebYesYesYesNoNoWeb Navigation
Mind2Web [9]2350WebYesYesNoNoNoWeb Navigation
WebShop [56]12000 ProductsWebNoNoYesNoNoWeb Navigation
RUSS [53]80WebYesYesNoNoNoWeb Navigation
WebSRC [6]2735WebYesYes-NoNoQA
MiniWoB++ [17]100Mobile WebsitesNoNoYesNoNoWeb Navigation
PixelHelp [23]187MobileYesYesNoNoNoUI Grounding
MetaGUI [42]1125MobileYesYesYesNoNoMobile Navigation
MoTIF [5]756MobileYesYesYesNoNoMobile Navigation
AITW [35]715142Mobile and WebYesYesYesNoNoMobile/Web Navigation
OmniACT (Ours)9802Desktop and WebYesYesYesYesYesCode Generation
", + "image_path": "8d48eb443c512036321444e8686fc6fc5c06aad51b713709dd42daba7b1694ea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 273, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 480, + 380 + ], + "type": "text", + "content": "understanding, and screen caption generation. Lexi [4] and Spotlight [22] propose models that use vision-only inputs to minimize the reliance on metadata such as view hierarchy. Furata et al. [11] demonstrates the use of fine-tuning for multimodal web navigation. The majority of machine learning models trained for UI understanding leverage the Rico dataset [8] and its extensions, which contain 64,462 unique Android screens and metadata. In addition, [4] released the UICaptions dataset, which consists of diverse image-captions pairs across a wide range of applications. PixelHelp [23] also released a corpus to train models that can interpret natural language instructions and map them to mobile UI actions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 396, + 317, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 317, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 317, + 408 + ], + "type": "text", + "content": "2.2 Autonomous Computer Agents" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 616 + ], + "type": "text", + "content": "The advent of large language models (LLMs) has been pivotal in the rapid advancement of agents that operate on web pages. Recent research such as ViperGPT [43] Chameleon [29], RCI Agent [18], VisProg [12], and [31] employ LLMs for planning or action prediction in developing autonomous agents. Benchmark datasets, such as MiniWoB [40], WebShop [56], Macaw-LLM [30], ASH-Prompting [41] Mind2Web [9], WebArena [62], AgentBench [28] and VisualWebArena [20] have also been proposed to measure the ability of LLM-based agents to automate web tasks. These methods mainly involve agents that operate on a text-based Document Object Model (DOM) of HTML scripts. This limits their understanding of screen context, which is crucial for the model's decision-making and action-taking processes. To address this limitation, [35] released Android in the Wild, a dataset comprising screens, natural language instructions, and corresponding actions. Following this, [59] proposed a multimodal model, AutoUI, which is designed to build an agent on the Android in the Wild dataset confined to the Android ecosystem. WebAgent [13] utilized Flan-U-PaLM, for grounded code generation, and HTML-T5 and showed improvement on real-world websites." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Current benchmarks for autonomous agents focus mainly on the Web or Android environments, posing challenges for tasks involving desktop applications or spanning multiple applications beyond the web domain. The absence of established benchmarks and datasets in this area, coupled with basic methods" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "for extracting user interface (UI) elements, underscores the need for significant progress in developing more versatile autonomous agents capable of handling diverse tasks beyond the current scope. To highlight the unique features that OmniACT introduces in the assessment of capable autonomous agents, we provide a comparison between the existing benchmarks and our proposed benchmark, OmniACT, in Table 1." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 150, + 208, + 465, + 371 + ], + "blocks": [ + { + "bbox": [ + 150, + 208, + 465, + 371 + ], + "lines": [ + { + "bbox": [ + 150, + 208, + 465, + 371 + ], + "spans": [ + { + "bbox": [ + 150, + 208, + 465, + 371 + ], + "type": "image", + "image_path": "3637b64b82edc539bc8230362a244c2b1b77bf9d9c6277baaa4ab34339e3319e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 383, + 482, + 438 + ], + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 438 + ], + "type": "text", + "content": "Fig. 2: Data Collection Pipeline. (1) We select over 60 applications and websites to ensure diversity, (2) segment the screen through human-annotated bounding boxes, (3) label the bounding boxes based on functionality, (4) ask student volunteers to come up with tasks, given a screen image, and (5) reverse map the textual labels to coordinates and filter the scripts based on execution and syntax." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 475, + 217, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 217, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 217, + 488 + ], + "type": "text", + "content": "3 OmniACT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "content": "We introduce a novel dataset and benchmark, OmniACT, which measures the performance of autonomous agents on both web and desktop applications. Compared to previous benchmarks which focus on text-based reasoning [9, 17, 40, 56, 62], our benchmark aims to measure multimodal agents that bridge large language model planners and UI understanding vision models. OmniACT can be accomplished as a standalone task as it is not under a mock environment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "All actions that a human can execute on the computer can be encoded in the PyAutoGUI [1] Python framework. This framework allows a user to execute keyboard and mouse operations by running Python code. The PyAutoGUI code to execute these tasks is shown in the third column of Figure 1. For other computer tasks, the PyAutoGUI library provides functions such as 'press', 'write', and 'scroll' which can be used to execute the task. Our dataset consists of parallel data of natural language tasks, UI screenshots, and ground truth PyAutoGUI scripts that achieve successful execution." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "type": "text", + "content": "3.1 Task Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": "Given an input state of a computer defined by the screen " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " and the task description " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " in natural language, the goal of the task is to output a sequence of actions " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " that can successfully accomplish the task " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " within a screenshot " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "S \\in \\{ \\text{Linux, Windows, MacOS, Webpage} \\}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ". Formally, the task can be defined as learning the transition function " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "f: T \\times S \\to A" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ". During dataset collection, we ensure that all task descriptions " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " are feasible and can be accomplished in the current screenshot " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ". To reduce ambiguity and facilitate better evaluation, we ensure that task descriptions are detailed and unambiguous. Tasks can also be visually grounded (e.g., 'Click the red button to start recording') or natural language based (e.g., 'Click the My Account button'). We define the action space using the functionalities in the PyAutoGUI library: " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "A \\in \\{\\text{'click', 'dragTo', 'scroll', 'write', \\ldots}\\}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ". The exhaustive list of actions is provided in Table 2. Our action space is much larger than other benchmarks [9,40,62] that resort to two or three interaction options. Mouse actions such as 'moveTo', 'click', 'rightClick', 'doubleClick', and 'dragTo', additionally require screen coordinates as arguments, which indicate the pixel location of the action." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 422 + ], + "type": "text", + "content": "Figure 1 illustrates sample tasks and corresponding outputs for three applications within OmniACT: (1) Stocks (MacOS), (2) Apartments.com (web page), and (3) Weather (MacOS). The first column depicts the input image, and the second column shows the natural language task that is to be executed on the current screen. To execute these tasks, a user must accurately perform a series of operations using the mouse and keyboard. Eg: to check the rate of change in Google's stock price over the last month, the mouse has to be moved to the last month and dragged while holding the left-click button to the current month." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 438, + 263, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 263, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 263, + 450 + ], + "type": "text", + "content": "3.2 Dataset Preparation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 457, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 578 + ], + "type": "text", + "content": "To prepare our dataset, we followed a pipelined approach, as summarized in Figure 2. We first selected a variety of applications and websites. For each application or website, we created bounding boxes around key UI elements and labeled them according to their functionality, which is crucial for assisting human annotators in writing accurate PyAutoGUI scripts. After each script is written, we converted the labels back into numeric coordinates, allowing us to align the scripts precisely with the locations of the UI elements. Finally, we thoroughly reviewed each script, focusing on its executability and adherence to syntax standards. This ensured the high quality and functionality of our dataset, making it a valuable resource for training and evaluating autonomous agents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Application/Website Selection To test the computer agents' generalization ability across different tasks, we collect tasks across multiple domains on both desktop and web applications. In total, we collect and annotate 9802 data points (Table 3), with the split between desktop and web applications approximately 3:1. The emphasis on desktop applications, which do not contain Document Object Model (DOM) hierarchies unlike HTML-based web pages, presents a more" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "complex multimodal challenge where visual cues are crucial. We collect tasks from applications within the three most popular operating systems. We select 22 native applications from MacOS, and 8 each from Linux and Windows. We annotate roughly 3 to 4 screens for every application. The full list of applications is provided in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 315, + 319 + ], + "type": "text", + "content": "Many common computer tasks today are still performed through web applications, so we also collect 3-4 screenshots from 27 different web applications. To ensure diversity in task intents, we categorize these tasks into one of the following 6 categories: (1) Shopping, (2) Entertainment, (3) Service, (4) Government, (5) Travel, (6) Health. Inspired by the methodology of [9], these categories were selected to cover a wide range of user intents and functionalities." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 357, + 245, + 446, + 328 + ], + "blocks": [ + { + "bbox": [ + 321, + 211, + 481, + 243 + ], + "lines": [ + { + "bbox": [ + 321, + 211, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 321, + 211, + 481, + 243 + ], + "type": "text", + "content": "Table 2: Action types supported by OmniACT and the number of instances for each action in the dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 357, + 245, + 446, + 328 + ], + "lines": [ + { + "bbox": [ + 357, + 245, + 446, + 328 + ], + "spans": [ + { + "bbox": [ + 357, + 245, + 446, + 328 + ], + "type": "table", + "html": "
TypeAction%
MouseClick63.73
Double Click0.58
Right Click0.77
Move/Hover1.85
Drag0.29
Scroll1.68
Horizontal Scroll0.17
KeyboardPress16.28
Hotkey3.00
Write11.65
", + "image_path": "efd3839d384fbecee73b13876012435c71bbd9d96d15ee0c14640fba843ab165.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 337, + 315, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 315, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 315, + 373 + ], + "type": "text", + "content": "UI Screen Segmentation To collect gold-standard data, we first annotate and segment the screen by identifying the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 479, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 479, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 479, + 397 + ], + "type": "text", + "content": "bounding boxes present on the screen. We employ slightly different techniques for web and desktop applications to create the bounding boxes:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 407, + 479, + 502 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 407, + 479, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 407, + 479, + 454 + ], + "spans": [ + { + "bbox": [ + 136, + 407, + 479, + 454 + ], + "type": "text", + "content": "1. Desktop Applications: We build a custom annotation interface based on PyQt5" + }, + { + "bbox": [ + 136, + 407, + 479, + 454 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 136, + 407, + 479, + 454 + ], + "type": "text", + "content": " to create bounding boxes manually over a screen image using a simple drag-and-click mechanism. This custom interface expedites the process and allows us to get highly accurate gold-label data for desktop images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 455, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 455, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 135, + 455, + 479, + 502 + ], + "type": "text", + "content": "2. Websites: For webpages, we write JavaScript code to extract all interactable (click, type, etc.) regions from HTML source code. We also extract banners, dropowns, submit, and radio buttons from the screen. We filter the elements to retain only those that are visible and interactable within the screen." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 521, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 604 + ], + "type": "text", + "content": "Functionality Tagging To map each bounding box to its correct functional description, we leverage Amazon MTurk workers (see details in Appendix), who are given an image with a bounding box and are required to write the correct description or label of the bounding box's function. For example, given an image of an Amazon webpage with a search bar, the annotator labels it as \"find-product-search-bar\". The logical descriptions are used to create tasks in a structured manner without the need to identify individual bounding box coordinates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "type": "text", + "content": "Task Creation Our approach for each screen involves utilizing all human-annotated bounding boxes and their labels to create tasks that can be executed" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 289, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 289, + 665 + ], + "type": "text", + "content": "3 https://pypi.org/project/PyQt5/" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 479, + 257 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 257 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 257 + ], + "type": "image", + "image_path": "91bc6bf190024aeb93b022e0af17e277beeacd4e3c1c9b2f438622bc2f6c41f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 323 + ], + "type": "text", + "content": "Fig. 3: DetACT Module. Given an initial image and a natural language task description, we use a pipelined approach to run OCR and SAM on the screen. The outputs from SAM are then used by icon and color-matching modules to obtain an exhaustive set of useful UI elements. The list of elements is passed through LLM based filter to select only the elements related to the given task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 346, + 480, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 394 + ], + "type": "text", + "content": "within the confines of a single screen. These tasks are designed to be visually grounded in order to measure the capabilities of multimodal agents. We plan to release the bounding box and their corresponding labels as the metadata for evaluation purposes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "content": "For dataset compilation, college students with basic Python programming skills served as annotators, accessing API references for PyAutoGUI and examples of potential tasks. Each student generated multiple tasks, each accompanied by three alternative natural language reformulations. For instance, \"What is " + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "inline_equation", + "content": "3 + 2" + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "content": "?\" might be reformulated as \"Calculate the sum of 2 and " + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "inline_equation", + "content": "3" + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "content": "\" or \"Add two to three\". To avoid train-test leakage, rephrased tasks were consistently placed in the same dataset split. Further details on the annotation process are available in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 317, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 317, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 317, + 640 + ], + "type": "text", + "content": "Reverse Mapping and Filtering To ensure high-quality data, we incorporate an additional step into the data collection pipeline. We build scripts to map the text-based labels of each bounding box back to their numeric coordinates, and then match the syntax and verify if the task will be executed on the screen. Using this filter, we remove all the non-working or syntactically incorrect data points and finally manually review the set of tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "After filtering, we obtain 9802 human-annotated, gold-label data points across more than 200 desktop and web screens" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 329, + 567, + 472, + 625 + ], + "blocks": [ + { + "bbox": [ + 321, + 544, + 482, + 566 + ], + "lines": [ + { + "bbox": [ + 321, + 544, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 321, + 544, + 482, + 566 + ], + "type": "text", + "content": "Table 3: Dataset distribution across splits and platforms." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 567, + 472, + 625 + ], + "lines": [ + { + "bbox": [ + 329, + 567, + 472, + 625 + ], + "spans": [ + { + "bbox": [ + 329, + 567, + 472, + 625 + ], + "type": "table", + "html": "
DomainTrain ValidationTestTotal
DesktopMac OS3028444786
Linux761126247
Windows1573216458
Web-1427206530
Total67899922,021
", + "image_path": "3ff6e279b9f9ca02e4e290ee118f367f93872ab2d707191e6b54439e215e34e7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "(Table 3), split into train, validation, and test sets in a 7:1:2 ratio. All collected data will be publicly released to encourage future work on multimodal agents." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 266, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 266, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 266, + 171 + ], + "type": "text", + "content": "4 Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 479, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 479, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 479, + 314 + ], + "type": "text", + "content": "In this section, we detail various evaluation metrics for benchmarking model performance on the OmniACT dataset. UI screens have additional constraints such as spatial relevance which are not factored in most conventional similarity-based metrics such as BLEU [34], CodeBLEU [36], BERTScore [58] and CodeBERTScore [61]. For example, a valid click action is usually not constrained to a single coordinate but can be any coordinate within a specified region. In the event of invalid coordinate predictions, an agent that predicts coordinates further away from the valid region should invoke a higher penalty compared to an agent that predicted coordinates close to the region. We propose two new metrics adapted: Sequence Score (Section 4.1) and Action Score (Section 4.2) aimed at utilizing UI information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 332, + 238, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 238, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 238, + 344 + ], + "type": "text", + "content": "4.1 Sequence Score" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 479, + 399 + ], + "type": "text", + "content": "The sequence score measures whether the predicted action sequence (e.g., 'click', 'write', 'press') exactly matches the gold sequence. Since predicting the first action in the sequence is relatively straightforward and later actions are more difficult, we define sequence score as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 190, + 402, + 480, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 402, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 190, + 402, + 480, + 434 + ], + "type": "interline_equation", + "content": "\\operatorname {S e q S c o r e} _ {i} = \\left\\{ \\begin{array}{l l} \\beta_ {1} + \\beta_ {2} * (s - 1) & \\text {i f a l l a c t i o n s m a t c h} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}", + "image_path": "88fd76273e0ceb30fa4e4329438eb667d3157eee0d7b9acac621a4d4621c1c3c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "text", + "content": " is the action sequence length, " + }, + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "text", + "content": " is set to 0.1 and " + }, + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 132, + 434, + 448, + 446 + ], + "type": "text", + "content": " is set to 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 464, + 226, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 226, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 226, + 475 + ], + "type": "text", + "content": "4.2 Action Score" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 483, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 479, + 531 + ], + "type": "text", + "content": "The action score measures how well a code snippet containing the correct action sequence can perform the task. Specifically, for a script with a correct action sequence, we introduce penalties for inaccurate behavior. The penalties are described below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "text", + "content": "1. Click penalty " + }, + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "text", + "content": ": For actions 'click', 'rightClick', 'doubleClick', 'moveTo', and 'dragTo', we penalize code snippets where predicted coordinates lie outside of the bounding box of the UI element. The click penalty for the " + }, + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "text", + "content": " action of the " + }, + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 136, + 540, + 479, + 588 + ], + "type": "text", + "content": " example is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 219, + 590, + 480, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 590, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 219, + 590, + 480, + 628 + ], + "type": "interline_equation", + "content": "M _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 1 - \\frac {\\mu}{\\mu + L _ {2}} & \\text {i f S e q S c o r e} _ {i} > 0 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "95a5132b5172244df62b7fc67d6c98fdfd114fb8af90c8c2fd74606617bd2dd8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "text", + "content": " corresponds to the smallest Euclidean distance between the predicted coordinate and bounding box. " + }, + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "text", + "content": " is zero when the predicted coordinate lies within the target bounding box. " + }, + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 148, + 629, + 480, + 666 + ], + "type": "text", + "content": " is the Dirichlet smoothing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 482, + 187 + ], + "type": "text", + "content": "coefficient which we dynamically set to the inverse of the length of the diagonal of the bounding box. This ensures that the penalty for points outside the bounding box varies based on the size of the bounding box. For two predicted points with the same " + }, + { + "bbox": [ + 147, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 147, + 116, + 482, + 187 + ], + "type": "text", + "content": ", the metric penalizes more heavily if the box is larger. This is sound with the intuition that the chances of clicking on a larger box are higher and should be penalized more in case of a mistake." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "text", + "content": "2. Key penalty " + }, + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "text", + "content": ": For actions 'press' and 'hotkey', we check whether the set of keys in the target code (represented as " + }, + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "inline_equation", + "content": "GK_{i}^{j}" + }, + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "text", + "content": ") and predicted code (represented as " + }, + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "inline_equation", + "content": "PK_{i}^{j}" + }, + { + "bbox": [ + 136, + 188, + 482, + 225 + ], + "type": "text", + "content": ") are the same. It is formally defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 202, + 228, + 481, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 202, + 228, + 481, + 258 + ], + "type": "interline_equation", + "content": "K _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 0 & \\text {i f} G K _ {i} ^ {j} = P K _ {i} ^ {j} \\text {a n d S e q S c o r e} _ {i} > 0 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "67f4719e24d1dff320c02ffbdbcb051f850c4158c27025f6fd27f9c89a0b8428.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 261, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 261, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 136, + 261, + 481, + 297 + ], + "type": "text", + "content": "3. Write penalty " + }, + { + "bbox": [ + 136, + 261, + 481, + 297 + ], + "type": "inline_equation", + "content": "(W_{p})" + }, + { + "bbox": [ + 136, + 261, + 481, + 297 + ], + "type": "text", + "content": ": For action type 'write', we penalize the output for the sentence to be typed. Specifically, we the employ BLEU score [34], and compute:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 299, + 481, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 299, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 194, + 299, + 481, + 330 + ], + "type": "interline_equation", + "content": "W _ {i} ^ {j} = \\alpha_ {i} \\times \\left\\{ \\begin{array}{l l} 1 - B L E U \\left(G S _ {i} ^ {j}, P S _ {i} ^ {j}\\right) & \\text {i f S e q S c o r e} _ {i} > 1 \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "1982ca79760031657a94ca16e8a8ab08b90acfe6ab6005308e10cf99cdcd162d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "inline_equation", + "content": "GS_{i}^{j}" + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "text", + "content": " represents the actual sentence to be typed, and " + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "inline_equation", + "content": "PS_{i}^{j}" + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "text", + "content": " represents the sentence predicted by the model in the " + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "text", + "content": " action of example " + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 148, + 333, + 481, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 367, + 361, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 361, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 361, + 379 + ], + "type": "text", + "content": "In the above equations, " + }, + { + "bbox": [ + 132, + 367, + 361, + 379 + ], + "type": "inline_equation", + "content": "(\\alpha_{i})" + }, + { + "bbox": [ + 132, + 367, + 361, + 379 + ], + "type": "text", + "content": " is the weighting factor:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 381, + 481, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 381, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 225, + 381, + 481, + 395 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\operatorname {S e q S c o r e} _ {i} / \\text {l e n g t h o f s e q u e n c e} i \\tag {5}", + "image_path": "542b1bb3cd6c3807221fa2a8f82f237f051c913c15f36b3a2008c8d4012f1423.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 397, + 481, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 481, + 419 + ], + "type": "text", + "content": "This ensures that the action score " + }, + { + "bbox": [ + 130, + 397, + 481, + 419 + ], + "type": "inline_equation", + "content": "\\in [0,1]" + }, + { + "bbox": [ + 130, + 397, + 481, + 419 + ], + "type": "text", + "content": ". The mean action score is calculated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 162, + 421, + 481, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 421, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 162, + 421, + 481, + 455 + ], + "type": "interline_equation", + "content": "\\text {A c t i o n S c o r e} = \\frac {\\sum_ {i} \\max \\left(\\operatorname {S e q S c o r e} _ {i} - \\sum_ {j} \\left(M _ {i} ^ {j} + K _ {i} ^ {j} + W _ {i} ^ {j}\\right) , 0\\right)}{\\sum_ {i} \\operatorname {S e q S c o r e} _ {i}} \\tag {6}", + "image_path": "a04c70fec5bca8adbc778299fe0219aa49c8fd7e6365666cea1705d027ae5e53.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 469, + 390, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 390, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 390, + 483 + ], + "type": "text", + "content": "5 DetACT: DETecting ACTions from UI" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "type": "text", + "content": "Understanding UI screens is crucial for multimodal computer tasks. Web-based agents typically use language-only inputs from the HTML DOM. This is insufficient for comprehending the full extent of an application UI, as many components may not be easily described with HTML code. To address this, we propose DetACT, which allows us to convert images of UI layouts into structured code and text outputs for a downstream LLM. DetACT is a system comprised of three distinct modules: the text module, the icon module, and the color module." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 136, + 586, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 586, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 136, + 586, + 482, + 648 + ], + "type": "text", + "content": "1. Text Extraction: We use the EasyOCR model4 to parse over the UI screens and collect all text-based elements. Along with the text, we also note the locations of each of these elements. This is depicted in Figure 3, along with a list of text elements found on the screen using the OCR Module. We segment and classify the different regions within the screenshot using the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 653, + 310, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 310, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 310, + 666 + ], + "type": "text", + "content": "4 https://github.com/JaidedAI/EasyOCR" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 205, + 115, + 410, + 322 + ], + "blocks": [ + { + "bbox": [ + 205, + 115, + 410, + 322 + ], + "lines": [ + { + "bbox": [ + 205, + 115, + 410, + 322 + ], + "spans": [ + { + "bbox": [ + 205, + 115, + 410, + 322 + ], + "type": "image", + "image_path": "877231d43634aee2879a1c206ac0fe7070408c08015ff5c3ce1a8c4b850a6958.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 333, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 131, + 333, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 333, + 482, + 389 + ], + "type": "text", + "content": "Fig. 4: Baseline Model Architecture. Image and task descriptions are sent to DetACT module, which gives a filtered list of UI elements relevant to feed into the prompt along with the task. We also show the prompt structure used for action script generation. This structure is passed through the LLM (along with the image for multimodal LLM) to generate the automation script." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 414, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 414, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 147, + 414, + 480, + 437 + ], + "type": "text", + "content": "Segment Anything Model (SAM) [19]. From the outputs, we filter out the non-textual segments for our icon and color detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 439, + 482, + 643 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 439, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 439, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 136, + 439, + 482, + 559 + ], + "type": "text", + "content": "2. Icon Module: For matching with the appropriate icon, we use a pack of 1600 icons5 as templates. Each of these icons is labeled with their appropriate functionality and is matched with the filtered outputs SAM [19]. For the similarity of the two images, we resize the reference icons and segmented region of interest (ROI) to the same size, and convert both images to grayscale. After this, we use the Structural Similarity Index (SSIM) [52], to find the closest match of the ROI to the icons in our set, and select the ones above the SSIM threshold of 0.95. As seen in Figure 3, a few icons matched on the screen are Globe icon, Calendar icon, Person icon, and Location icon; each depicting a different use case." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 559, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 559, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 136, + 559, + 482, + 643 + ], + "type": "text", + "content": "3. Color Module: Finally, to place all segments of interest into appropriate buckets of colors, we average the RGB pixel values over the ROI and, based on that value, bucket them into different color categories. We categorize colors differently based on the human perspective of the ranges of each color. To avoid ambiguity, we consider eleven major colors, namely yellow, blue, green, red, pink, violet, white, black, orange, brown, and grey. We record the center of the element along with the color." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 234, + 665 + ], + "type": "text", + "content": "5 https://icomoon.io/" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Once all the elements of each category are extracted with their coordinates, we then filter these UI elements by prompting GPT-4 [32]. We ensure that the elements selected are suited only for our task, for which we also provide the task description in our prompts along with the list of elements. Full details of the prompt are provided in the appendix section of the paper. As we observe in Figure 3, given an image from the Expedia application, and a task (\"Click on the Black Location icon and enter the destination as Paris\"), the LLM filters out the elements to retain only \"Going To\", \"Location Icon\", and the Black colored elements from the screen. This is passed as input to the LLM or VLM backbone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 241, + 209, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 209, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 209, + 254 + ], + "type": "text", + "content": "6Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 456 + ], + "type": "text", + "content": "To evaluate the performance of existing language model-based agents, we conduct experiments with both language-based and multimodal baselines. The DetACT module takes in image and text descriptions of the task and outputs the color, icon, and text-based signals. This is concatenated to the prompt for the LLM prompt-based baselines (see Figure 4). Every prompt starts with a role assignment [60], followed by the detailed API reference of the PyAutoGUI function set, along with a textual description of their function. We then add five in-context examples from the training set that most closely match the task (based on the cosine similarity of the MiniLM [50] embeddings of the reference task and the train examples). We add a list of UI elements filtered by the DetACT module to the prompt. Finally, we provide the rules with the task description. For multimodal baselines, we also pass the image pixels to the vision encoder. We choose coordinate-based UI elements in the prompt as recent techniques like the Set-of-Mark (SOM) [54] prompting does not work for desktop settings since it is difficult to obtain interactive elements from the desktop screen images. We report the results of several baselines:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 463, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 463, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 463, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 463, + 482, + 545 + ], + "type": "text", + "content": "- Few-shot Generative LLM: We experiment with models from LLaMA-2 [47], Vicuna-1.5 [7], CodeLLaMA-34B [37], Palmyra [46], and GPT [32] series. We use the prompts structure as shown in Figure 4 to prompt the model. For LLaMA and CodeLLaMa, we reduce the prompt length to 2000 tokens by removing outputs from the DetACT module with lower confidence, as we observed poor performance on longer prompts. For the other models, we allow prompts with up to 4000 token sizes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 546, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 482, + 592 + ], + "type": "text", + "content": "- Finetuned Generative LLM: We fine-tuned the LLaMA-13B model and Vicuna-13B using QLoRa [10] with rank 64 and scaling factor 16 for 300 steps to generate the code given screen description from the DetACT module and the instruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 482, + 665 + ], + "type": "text", + "content": "- Few-shot Generative Multimodal Models: As OmniACT is predominantly multimodal, with a majority of tasks being visually grounded, we conduct experiments with large multimodal models. Given the limited research in this domain [51, 57], there is a scarcity of available multimodal models with significant size adept for this task. Here, we experiment with [26, 27], providing a similar prompt as well as the screen image." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 278, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 278, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 278, + 129 + ], + "type": "text", + "content": "7 Results and Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 315, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 315, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 315, + 341 + ], + "type": "text", + "content": "As shown in Table 4, we experiment with three different categories of models, namely Prompt-based LLMs, Fine-tuned LLMs, and Prompt-based Multimodal Models. GPT-4 is the best-performing approach, scoring higher on the sequence score and invoking lower penalties on coordinate predicting and text input. For prompt-only LLMs, the GPT-3.5-turbo and GPT-4 models outperform the other LLM baselines, including the LLaMA [47] and Vicuna [7] models. We observe that CodeLLaMA-34B [38], which is trained for code generation, also achieves a higher performance than other models of the same size at predicting the action sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 315, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 315, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 315, + 474 + ], + "type": "text", + "content": "Fine-tuned models also perform much better than their few-shot prompt-only counterparts. Fine-tuning substantially improves LLaMA-13B's sequence score (4.80 to 8.92) and action score (1.62 to 2.14), as well as the other metrics. Despite this, we observed that both, prompt-based LLMs and finetuned LLMs face severe mouse penalties, especially on click coordinates. This is because they rely solely on text-based signals." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 322, + 263, + 481, + 435 + ], + "blocks": [ + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "type": "text", + "content": "Table 4: Baseline Performance. (A) Prompt-only LLMs, (B) Fine Tuned LLMs, (C) Prompt-only Multimodal Models. The table represents the Sequence score (SS), click penalty " + }, + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "type": "inline_equation", + "content": "(M_p)" + }, + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "type": "text", + "content": ", Key penalty " + }, + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "type": "inline_equation", + "content": "(K_p)" + }, + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "type": "text", + "content": ", Write Penalty " + }, + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "type": "inline_equation", + "content": "(W_p)" + }, + { + "bbox": [ + 321, + 174, + 482, + 262 + ], + "type": "text", + "content": ", and Action Score (AS). The best results for the (SS) and (AS) are highlighted." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 263, + 481, + 435 + ], + "lines": [ + { + "bbox": [ + 322, + 263, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 322, + 263, + 481, + 435 + ], + "type": "table", + "html": "
ModelSS(↑)MpKpWpAS(↑)
Prompt based LLMs
LLaMA-7B [47]4.121.241.830.570.48
Vicuna-7B [7]3.881.171.510.430.77
LLaMA-13B [47]4.801.320.930.931.62
Vicuna-13B [7]5.441.650.941.061.78
Palmyra-Instruct-30B [45]7.515.680.120.401.31
CodeLLaMA-34B [38]10.092.992.710.663.72
Palmyra-X 43B [2]11.203.123.022.122.94
GPT-3.5-turbo-0613 [33]22.858.134.512.317.89
GPT-4 [32]32.7510.276.993.8911.60
Finetuned LLMs
LLaMA-13B FT8.924.611.430.742.14
Vicuna-13B FT8.784.121.310.632.72
Multimodal LLMs
LLaVA-v1.5-7B [27]13.234.731.241.445.82
LLaVA-v1.5-13B [26]20.566.073.442.858.19
Gemini-Pro [44]30.989.056.813.6611.46
GPT-4V [26]38.7210.537.144.0317.02
Human Performance82.230.120.361.6180.14
", + "image_path": "0eda75a4d05cbc94fa4e0a4e8b0ed997ce4e639ad39c5f84ffd4f568682a7b24.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "type": "text", + "content": "To address this, we experiment with multimodal language models (Table 4). We observe that the coordinate prediction improves significantly when we provide the entire image as input to the multimodal LLM, as this enables it to fully utilize the screen representation. In addition to open sourced models, we also experiment with the GPT-4-vision API [55] which shows that GPT-4 Vision [55] outperforms GPT-4 significantly on the Action Score along with improving the sequence score, which we attribute to the strong reasoning abilities of GPT-4 coupled with the improved visual understanding capabilities of the GPT-4-vision model [55]. These findings pave the way towards exciting new research directions on building multimodal models for long-horizon planning and code generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Human performance over the task: OmniACT consists of visually complicated tasks, and tests various types of computer skills. In order to get a gauge of how well humans perform, we collect evaluation data from human evaluators. We split the test set uniformly amongst 10 human evaluators, and provided them with the screenshot and task instruction. We record the actions taken by the an-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "notators, and measure their performance on our predefined metrics (Table 4). We find that users generally exhibit a high level of proficiency when attempting most tasks for the first time. However, there are instances where users face difficulties in successfully completing certain tasks. These are due to factors including the user's inability to fully comprehend the task, difficulties in grounding the task to the provided screenshot, or a lack of familiarity with the UI." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 204, + 326, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 204, + 326, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 204, + 326, + 217 + ], + "type": "text", + "content": "8 Conclusion and Future Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 360 + ], + "type": "text", + "content": "Autonomous virtual agents offer the potential to automate routine tasks, benefiting users with limited technical expertise. To solve this task, we introduce OmniACT, a unique dataset of 9.8K human-labeled data points. OmniACT benchmarks autonomous agents across a range of tasks on web and desktop applications. LLM-based agents, like GPT-4, achieve a respectable action score of 11.6 on our dataset. However, OmniACT presents a challenge for the current state-of-the-art language and multimodal models. It provides a direction for future research on foundational multimodal models that seamlessly integrate language and visual understanding of computer screens and stands poised to drive the next wave of advancements in generalist autonomous agents offering omnipotent assistance to humans." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 375, + 222, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 222, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 222, + 388 + ], + "type": "text", + "content": "9 Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 483 + ], + "type": "text", + "content": "This work introduces a valuable dataset, yet we recognize a few limitations that exist. State-of-the-art models like GPT-4, may exhibit susceptibility to hallucinations and bias towards specific data types, hindering broad applicability. Reliance on closed models like GPT-4V poses integration challenges due to high costs and time constraints. Despite efforts for equal representation and data collection without personal information, biases may be introduced as the dataset is exclusively in English, and human-curated content may have temporal biases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 263, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 263, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 263, + 512 + ], + "type": "text", + "content": "10 Ethics Statement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "As a part of the dataset creation process, we carefully review the pipeline at every stage, ensuring there is no personally identifiable information or offensive content, either during data collection or through the use of LLMs. For all purposes, we create dummy accounts that mimic real-like user content. To get the gold labels scripts we seek help from well-qualified student workers, approved through the institution, and get the bounding box data annotated through MTurk workers, both of whom are paid $25 per hour, which is greater than the minimum wage rate (We detail this process in the supplementary material). Human studies are also done with the help of student workers approved by the institution at the above-mentioned payscale. We also ensure that all groups have equitable representation and that no personal opinions are reflected in the dataset, avoiding bias during the collection as well as the annotation process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 177 + ], + "type": "text", + "content": "We extend our heartfelt gratitude to Writer.com for their generous and unwavering support throughout this project. Their dedicated team's expertise and collaboration were invaluable in achieving our goals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 223, + 479, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 479, + 245 + ], + "type": "text", + "content": "1. Pyautogui: A cross-platform gui automation python module for human beings. https://github.com/asweigart/pyautogui (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "text", + "content": "2. AlShikh, W., Daaboul, M., Goddard, K., Imel, B., Kamble, K., Kulkarni, P., Russak, M.: Becoming self-instruct: introducing early stopping criteria for minimal instruct tuning (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 480, + 301 + ], + "type": "text", + "content": "3. Bai, C., Zang, X., Xu, Y., Sunkara, S., Rastogi, A., Chen, J., y Arcas, B.A.: Uibert: Learning generic multimodal representations for ui understanding (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "text", + "content": "4. Banerjee, P., Mahajan, S., Arora, K., Baral, C., Riva, O.: Lexi: Self-supervised learning of the ui language (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 354 + ], + "type": "text", + "content": "5. Burns, A., Arsan, D., Agrawal, S., Kumar, R., Saenko, K., Plummer, B.A.: Mobile app tasks with iterative feedback (motif): Addressing task feasibility in interactive visual environments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 356, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 480, + 399 + ], + "type": "text", + "content": "6. Chen, X., Zhao, Z., Chen, L., Ji, J., Zhang, D., Luo, A., Xiong, Y., Yu, K.: Websrc: A dataset for web-based structural reading comprehension. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 4173-4185 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 443 + ], + "type": "text", + "content": "7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 401, + 480, + 443 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 401, + 480, + 443 + ], + "type": "text", + "content": " chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 445, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 510 + ], + "type": "text", + "content": "8. Deka, B., Huang, Z., Franzen, C., Hibschman, J., Afergan, D., Li, Y., Nichols, J., Kumar, R.: Rico: A mobile app dataset for building data-driven design applications. In: Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology. p. 845-854. UIST '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3126594.3126651, https://doi.org/10.1145/3126594.3126651" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 511, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 543 + ], + "type": "text", + "content": "9. Deng, X., Gu, Y., Zheng, B., Chen, S., Stevens, S., Wang, B., Sun, H., Su, Y.: Mind2web: Towards a generalist agent for the web. arXiv preprint arXiv:2306.06070 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 480, + 565 + ], + "type": "text", + "content": "10. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 598 + ], + "type": "text", + "content": "11. Furuta, H., Nachum, O., Lee, K.H., Matsuo, Y., Gu, S.S., Gur, I.: Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "text", + "content": "12. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Gur, I., Furuta, H., Huang, A., Safdari, M., Matsuo, Y., Eck, D., Faust, A.: A real-world webagent with planning, long context understanding, and program synthesis (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Gur, I., Rueckert, U., Faust, A., Hakkani-Tur, D.: Learning to navigate the web. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "15. He, Z., Sunkara, S., Zang, X., Xu, Y., Liu, L., Wichers, N., Schubiner, G., Lee, R., Chen, J., y Arcas, B.A.: Actionbert: Leveraging user actions for semantic understanding of user interfaces (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "16. Huang, Z., Zeng, Z., Liu, B., Fu, D., Fu, J.: Pixel-bert: Aligning image pixels with text by deep multi-modal transformers (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "type": "text", + "content": "17. Humphreys, P.C., Raposo, D., Pohlen, T., Thornton, G., Chhaparia, R., Muldal, A., Abramson, J., Georgiev, P., Santoro, A., Lillicrap, T.: A data-driven approach for learning to control computers. In: International Conference on Machine Learning. pp. 9466-9482. PMLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "text", + "content": "18. Kim, G., Baldi, P., McAleer, S.: Language models can solve computer tasks. arXiv preprint arXiv:2303.17491 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 260, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 260, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 480, + 292 + ], + "type": "text", + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "20. Koh, J.Y., Lo, R., Jang, L., Duvvur, V., Lim, M.C., Huang, P.Y., Neubig, G., Zhou, S., Salakhutdinov, R., Fried, D.: Visualwebarena: Evaluating multimodal agents on realistic visual web tasks (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "21. LeCun, Y.: A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "22. Li, G., Li, Y.: Spotlight: Mobile ui understanding using vision-language models with a focus (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 370, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 390 + ], + "type": "text", + "content": "23. Li, Y., He, J., Zhou, X., Zhang, Y., Baldridge, J.: Mapping natural language instructions to mobile ui action sequences (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "24. Li, Y., Li, G., He, L., Zheng, J., Li, H., Guan, Z.: Widget captioning: Generating natural language description for mobile user interface elements (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 415, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 480, + 434 + ], + "type": "text", + "content": "25. Li, Y., Li, G., Zhou, X., Dehghani, M., Gritsenko, A.: Vut: Versatile ui transformer for multi-modal multi-task user interface modeling (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "text", + "content": "26. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 421, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 421, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 421, + 467 + ], + "type": "text", + "content": "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 480, + 511 + ], + "type": "text", + "content": "28. Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., Zhang, S., Deng, X., Zeng, A., Du, Z., Zhang, C., Shen, S., Zhang, T., Su, Y., Sun, H., Huang, M., Dong, Y., Tang, J.: Agentbench: Evaluating llms as agents (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "type": "text", + "content": "29. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "30. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 609 + ], + "type": "text", + "content": "31. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 307, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 307, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 307, + 620 + ], + "type": "text", + "content": "32. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 449, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 449, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 449, + 632 + ], + "type": "text", + "content": "33. OpenAI: Introducing chatgpt (2023), https://openai.com/blog/chatgpt" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 634, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 480, + 665 + ], + "type": "text", + "content": "34. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "35. Rawles, C., Li, A., Rodriguez, D., Riva, O., Lillicrap, T.: Android in the wild: A large-scale dataset for android device control (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "36. Ren, S., Guo, D., Lu, S., Zhou, L., Liu, S., Tang, D., Sundaresan, N., Zhou, M., Blanco, A., Ma, S.: Codebleu: a method for automatic evaluation of code synthesis. arXiv preprint arXiv:2009.10297 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 228 + ], + "type": "text", + "content": "37. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 283 + ], + "type": "text", + "content": "38. Rozière, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X.E., Adi, Y., Liu, J., Remez, T., Rapin, J., Kozhevnikov, A., Evtimov, I., Bitton, J., Bhatt, M., Ferrer, C.C., Grattafori, A., Xiong, W., Defossez, A., Copet, J., Azhar, F., Touvron, H., Martin, L., Usunier, N., Scialom, T., Synnaeve, G.: Code llama: Open foundation models for code (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "text", + "content": "39. Shaw, P., Joshi, M., Cohan, J., Berant, J., Pasupat, P., Hu, H., Khandelwal, U., Lee, K., Toutanova, K.: From pixels to ui actions: Learning to follow instructions via graphical user interfaces. arXiv preprint arXiv:2306.00245 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 318, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 480, + 350 + ], + "type": "text", + "content": "40. Shi, T., Karpathy, A., Fan, L., Hernandez, J., Liang, P.: World of bits: An open-domain platform for web-based agents. In: International Conference on Machine Learning, pp. 3135-3144. PMLR (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 351, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 480, + 373 + ], + "type": "text", + "content": "41. Sridhar, A., Lo, R., Xu, F.F., Zhu, H., Zhou, S.: Hierarchical prompting assists large language model on web navigation. arXiv preprint arXiv:2305.14257 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "42. Sun, L., Chen, X., Chen, L., Dai, T., Zhu, Z., Yu, K.: Meta-gui: Towards multimodal conversational agents on mobile gui. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 6699-6712 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "type": "text", + "content": "43. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 463 + ], + "type": "text", + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "text", + "content": "45. team, W.E.: InstructPalmyra-30b : Instruct tuned Palmyra-Large model. https://dev.writer.com (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "text", + "content": "46. team, W.E.: Palmyra-base Parameter Autoregressive Language Model. https://dev.writer.com (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "text", + "content": "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "text", + "content": "48. Wang, B., Li, G., Zhou, X., Chen, Z., Grossman, T., Li, Y.: Screen2words: Automatic mobile ui summarization with multimodal learning. In: The 34th Annual ACM Symposium on User Interface Software and Technology. pp. 498-510 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "text", + "content": "49. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al.: A survey on large language model based autonomous agents. arXiv preprint arXiv:2308.11432 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 643 + ], + "type": "text", + "content": "50. Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems 33, 5776-5788 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "51. Wang, X., Chen, G., Qian, G., Gao, P., Wei, X.Y., Wang, Y., Tian, Y., Gao, W.: Large-scale multi-modal pre-trained models: A comprehensive survey (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniACT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 435 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 194 + ], + "type": "text", + "content": "53. Xu, N., Masling, S., Du, M., Campagna, G., Heck, L., Landay, J., Lam, M.: Grounding open-domain instructions to automate web support tasks. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 1022-1032 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "54. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "55. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "type": "text", + "content": "56. Yao, S., Chen, H., Yang, J., Narasimhan, K.: Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems 35, 20744-20757 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "58. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "59. Zhang, Z., Zhang, A.: You only look at screens: Multimodal chain-of-action agents (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "60. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "61. Zhou, S., Alon, U., Agarwal, S., Neubig, G.: Codebertscore: Evaluating code generation with pretrained models of code. arXiv preprint arXiv:2302.05527 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "text", + "content": "62. Zhou, S., Xu, F.F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Bisk, Y., Fried, D., Alon, U., et al.: Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854 (2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "text", + "content": "R. Kapoor, Y. Butala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_content_list.json b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..745ccc5517283955698de4404b669855f44f4ea9 --- /dev/null +++ b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_content_list.json @@ -0,0 +1,1957 @@ +[ + { + "type": "text", + "text": "OmniNOCS: A unified NOCS dataset and model for 3D lifting of 2D objects", + "text_level": 1, + "bbox": [ + 220, + 141, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akshay Krishnan $^{1,2}$ , Abhijit Kundu $^{1}$ , Kevis-Kokitsi Maninis $^{1}$ , James Hays $^{2}$ , and Matthew Brown $^{1}$", + "bbox": [ + 215, + 210, + 787, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Google Research†", + "bbox": [ + 437, + 253, + 563, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Georgia Institute of Technology", + "bbox": [ + 398, + 268, + 602, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6cdaaff917a5eba3946dee9b9ca5ebdc840d3a5a95ab822c2bb341eca8803cc.jpg", + "image_caption": [ + "Fig. 1: We introduce OmniNOCS, a large-scale dataset with Normalized Object Coordinates (NOCS), instance masks, and 3D box annotations for objects across several classes, domains, and cameras. We also propose NOCSformer, a model trained on OmniNOCS that lifts each 2D object bounding box in an image to its corresponding 3D oriented box (pose) and 3D pointcloud (shape)." + ], + "image_footnote": [], + "bbox": [ + 222, + 291, + 393, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fd026d4e6665924bd49b1a033471c917353d1a4d0c01170f5b7c1310616d6f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 292, + 586, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a0d92c0efe8f14ba295dcc2483152ac8b33510ea8aef85261200f64e80b54f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 295, + 776, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose OmniNOCS, a large-scale monocular dataset with 3D Normalized Object Coordinate Space (NOCS) maps, object masks, and 3D bounding box annotations for indoor and outdoor scenes. OmniNOCS has 20 times more object classes and 200 times more instances than existing NOCS datasets (NOCS-Real275, Wild6D). We use OmniNOCS to train a novel, transformer-based monocular NOCS prediction model (NOCsformer) that can predict accurate NOCS, instance masks and poses from 2D object detections across diverse classes. It is the first NOCS model that can generalize to a broad range of classes when prompted with 2D boxes. We evaluate our model on the task of 3D oriented bounding box prediction, where it achieves comparable results to state-of-the-art 3D detection methods such as Cube R-CNN. Unlike other 3D detection methods, our model also provides detailed and accurate 3D object shape and segmentation. We propose a novel benchmark for the task of NOCS prediction based on OmniNOCS, which we hope will serve as a useful baseline for future work in this area. Our dataset and code will be at the project website: https://omninocs.github.io.", + "bbox": [ + 259, + 536, + 743, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 763, + 356, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicting the 6 Degree-of-Freedom (6DoF) pose and shape of objects from images is a crucial problem in 3D scene understanding. Robots need to understand the location,", + "bbox": [ + 212, + 787, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Now at Google DeepMind", + "bbox": [ + 215, + 825, + 364, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shape, and orientation of various objects to grasp and interact with them. Self-driving vehicles need to understand the location and heading of vehicles, pedestrians, and other objects on the road. In particular, the 3D orientation of these objects is crucial to predict their behavior. It is also important in AR/VR applications, as it allows users to interact with objects in meaningful ways. Predicting object shape and pose from monocular images is also a prerequisite to initialize 3D panoptic neural scene graphs [27, 39] and to track objects in videos [42]. These applications require approaches that generalize across a wide range of object classes, environments, and camera models.", + "bbox": [ + 217, + 146, + 787, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The problem of localizing 3D objects has been extensively studied through the lens of monocular 3D object detection or 6DoF pose estimation. The most commonly used approach is to represent objects as 3D cuboids, and train a model to regress the cuboid parameters from a 2D RoI of the object [7, 8, 28, 41, 51, 54]. An alternative method involves predicting corresponding 3D points in an object coordinate space followed by pose estimation using 3D-2D alignment [11, 12, 20, 33, 49]. However, all these methods are limited to narrow datasets collected on a single camera and context with typically very few object classes. Existing models are also trained separately for every dataset, preventing them from being used more widely.", + "bbox": [ + 217, + 270, + 787, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work [7] takes a step towards scalability by introducing the Omni3D dataset, which aggregates 3D detection datasets from different domains, and trains a Cube R-CNN model to regress 3D bounding boxes for 50 classes. However, Omni3D notably lacks consistently oriented, object-centric boxes, as the ground truth orientations are not canonical for object classes. Cube R-CNN [7] instead uses a Chamfer distance loss which is invariant to the predicted 3D cuboid orientation. This causes it to predict inconsistent object orientations, for example, often flipping the orientations of cars by $180^{\\circ}$ . Further, Omni3D is a detection dataset that does not provide object shape information.", + "bbox": [ + 217, + 409, + 787, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work aims to overcome the above shortcomings, by providing a new large-scale dataset with consistent object-centric ground truth boxes along with detailed shape. We argue for the use of Normalized Object Coordinate Space (NOCS) as proposed in [49] as a 3D object shape representation. NOCS represents both the canonical orientation and the shape of the visible surface of the object, properties that are essential for real-world applications such as self-driving and robotics. It can also be used to estimate the 3D bounding box of the object. However, all existing work on NOCS [24, 25, 32, 48, 49] only train small models on small datasets with fewer than 10 classes. We address the lack of diverse NOCS datasets by proposing a large-scale monocular NOCS dataset, OmniNOCS that has NOCS annotations, instance segmentation, and canonically oriented bounding boxes for 97 object classes, containing $380\\mathrm{k}$ images from 10 different data sources, making it the largest and most diverse NOCS dataset currently available. OmniNOCS includes all of the data from Omni3D (KITTI [19], nuScenes [10], ARKitScenes [5], Objectron [1], Hypersim [44], SUN-RGBD [45]), with the addition of Cityscapes [15], virtual KITTI [9], NOCS-Real275 [49], and the Waymo Open Dataset [46].", + "bbox": [ + 217, + 534, + 787, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work also introduces a new model, which we term \"NOCSformer\", that predicts NOCS coordinates and oriented 3D bounding boxes from monocular images and 2D detections for all the classes in OmniNOCS. NOCSformer leverages self-supervised pretrained ViTs [17, 40] and does not use any class-specific model parameters. This enables it scale to large vocabularies, and share information across semantically similar", + "bbox": [ + 217, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "object classes. In contrast, existing NOCS models use small NOCS heads with class-specific parameters that significantly limit their performance on large datasets like OmniNOCS. NOCSformer also predicts the 3D size of the object along with its 3D orientation using a learned PnP head. This allows NOCSformer to predict 3D oriented bounding boxes and object point cloud in metric scale for diverse object categories. Our experiments evaluate the quality of NOCS and bounding boxes predicted by NOCSformer in comparison to existing NOCS prediction or 3D detection models. We find that training on OmniNOCS allows NOCSformer to generalize to unseen datasets, even outperforming baselines trained on the target dataset in NOCS prediction accuracy.", + "bbox": [ + 218, + 145, + 787, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 240, + 282, + 468, + 297 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The OmniNOCS dataset: A new dataset containing Normalised Object Coordinates for 380k images in 97 categories, an order of magnitude larger on both counts than existing NOCS datasets.", + "- NOCSformer: A novel transformer-based architecture for predicting object NOCS, mask, and size from input 2D boxes, utilizing pre-trained self-supervised ViT backbones. NOCSformer is the first NOCS model to generalize to vocabularies with $90+$ classes and to unseen datasets, including images from internet collections.", + "- OmniNOCS benchmark: an evaluation framework with metrics for directly comparing different NOCS prediction algorithms on OmniNOCS, with baselines established via NOCSFormer." + ], + "bbox": [ + 225, + 300, + 784, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 217, + 474, + 361, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The task of predicting 3D object pose has been studied both in the context of camera/object pose estimation (6DoF) and 3D bounding box estimation (6DoF pose + 3DoF size). A further distinction can be made between methods that regress directly to object pose / bounding box parameters, and methods that compute per pixel coordinates or depth as an initial step. We review each of these paradigms in the sections below.", + "bbox": [ + 217, + 505, + 787, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D localization by regressing bounding boxes", + "text_level": 1, + "bbox": [ + 217, + 603, + 570, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Direct approaches to 3D object detection and pose estimation involve networks that output rotation, translation and scale parameters directly. Several works have explored different parameterizations in this setting, e.g., PoseCNN [54] uses regression of translation via centre direction + distance maps (which enable detection even under occlusion), and quaternion representation of rotation. BB8 [41] uses segmentation followed by a CNN to regress to 2D projections of the $8 \\times 3\\mathrm{D}$ bounding box corners. Multi-view monocular approaches have also been proposed, e.g., DETR3D [52] which uses DETR-style attention to reason about object interactions. Several techniques also make use of existing 2D bounding box predictors, either as an input to a 3D lifting approach [34], or as a constraint on 3D box predictions [38].", + "bbox": [ + 217, + 628, + 787, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another group of works focuses on predicting the alignment of 3D CAD models within various modalities: images [21, 29], videos [31, 35], or 3D scans [2, 3]. These approaches typically determine the object's 3D translation, rotation, and size (9DoF) and find a CAD model with a similar visual appearance.", + "bbox": [ + 217, + 779, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 3D localization from model-to-image alignment", + "text_level": 1, + "bbox": [ + 215, + 146, + 584, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An alternative approach to 3D localization is to first predict correspondence, either between views or to a normalized space, and then reason over redundant correspondences to establish pose. This has been done with both sparse and dense correspondences, e.g., AutoShape [33] makes use of sparse 2D to 3D correspondences, with a learned shape model and sparse 2D keypoints. [20] uses pairwise semantic correspondence from ViT, to find pose between a reference image and a sequence of targets.", + "bbox": [ + 212, + 167, + 787, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Several approaches use dense correspondences and the idea of Normalised Object Coordinate Spaces, or NOCS. In the original NOCS work, Wang et al. [49] use RGB-D data and 3D-3D correspondences for pose estimation. Follow-up work [11] estimates pose using PnP variants on 2D-3D correspondences. While NOCS [49] used the average spatial dimensions of the object category to define the object frame, [53] showed that an instance-specific NOCS coordinate frame can be used alongside a predicted instance size for the same purpose. Other methods combine coordinate regression and direct approaches, e.g., [48], which uses direct regression based on dense correspondences.", + "bbox": [ + 212, + 257, + 787, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to our work, MonoRUun [11] lifts 2D detections to 3D object coordinates without explicit class supervision. They also use a novel self-supervised coordinate regression training loss which obviates the need for detailed 3D ground truth. However, their representation does not include masks and therefore does not provide explicit object shape information. Their evaluation is also limited to 3 classes (car, pedestrian, cyclist) on the KITTI-Object test set. M3D-RPN [8] also has a single 3D head for multiple classes, jointly generating 2D and 3D bounding box predictions, though without shape information, and similarly limited to car, pedestrian and cyclist classes on KITTI.", + "bbox": [ + 212, + 378, + 787, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Monocular object localization / pose estimation datasets", + "text_level": 1, + "bbox": [ + 215, + 517, + 643, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For 3D object detection from monocular RGB images, most existing works use a small number of classes on a single dataset. Cube R-CNN/Omni3D [7] contribute towards creating a general purpose monocular 3D object detector. Cube R-CNN performs well over six 3D datasets: KITTI, SUN RGB-D, ARKitScenes, Objectron, nuScenes and Hypersim. However, the method is class specific, with Cube R-CNN trained only on 50 classes, and unable to work in the open class setting.", + "bbox": [ + 212, + 537, + 787, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cube R-CNN also uses a Chamfer loss on the predicted 3D box corners to deal with the inconsistent coordinate frame annotations in the underlying Omni3D datasets. The lack of direct orientation supervision results in inconsistent orientation predictions (for example, the positive x axis may point to the front or the back of the car for different instances). This orientation inconsistency is problematic for methods that seek to build detailed 3D models of object categories as a downstream task [27, 39].", + "bbox": [ + 212, + 628, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Other works are not limited to 3D boxes, and directly provide object shapes. These shapes come in the form of annotated point clouds (e.g., ScanNet [16]), or aligned CAD models (e.g., Scan2CAD [2], Pix3D [47], CAD-Estate [36]). Object point clouds are created by labeling points on scanned 3D scenes, a process that doesn't scale well to large datasets. Semi-automatically aligning CAD models is scale-able to a certain extent, but since it relies on retrieving existing models, the resulting shapes are rarely accurate, and the alignments are sensitive to deformations and movable parts. In contrast to these works, we propose a large-scale dataset of many different categories.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 379, + 127 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/9b6ca27155eb41b8344abde494fd93de510b97ba9fb9e956ca22813c78cf0d75.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNOCS GTReal#Images#Classes#Instances
CAMERA25 [49]✓×300k6184
NOCS-Real275 [49]✓✓7k624
Wild6D [18]×✓1M51.8K
OmniNOCS✓✓380k97>450k
", + "bbox": [ + 303, + 143, + 697, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Comparison to other NOCS datasets: Other existing NOCS datasets are limited in number of classes and instances. Note that Wild6D has 1M images, but these are used for self-supervised training, as it does not include NOCS ground truth (GT).", + "bbox": [ + 215, + 217, + 787, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/438dbb5de78e5233772616e4da6306952f437af2f155c8c65d4cfe25df4e33e0.jpg", + "image_caption": [ + "Fig. 2: Preparation of the OmniNOCS dataset: First, we ensure that object orientations are canonical across different all subsets of OmniNOCS. Next, we re-compute depth for datasets where depth is not available or is too noisy. Finally, we annotate objects with (pseudo) instance mask labels where ground truth masks are not available." + ], + "image_footnote": [], + "bbox": [ + 217, + 271, + 785, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 OmniNOCS dataset", + "text_level": 1, + "bbox": [ + 215, + 431, + 416, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Data statistics", + "text_level": 1, + "bbox": [ + 215, + 464, + 356, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We create OmniNOCS, a first-of-its-kind large and diverse NOCS dataset comprising data from several classes and domains. OmniNOCS uses data from self-driving scenes [9, 15, 19, 46], indoor scenes [5, 44, 45, 49], and object-centric videos [1]. Each of these sources use cameras with different parameters, ranging from phone cameras to wide FOV cameras mounted on self-driving cars. The number of object instances per image also varies widely, from single-object images [1] to images with hundreds of objects [44, 46].", + "bbox": [ + 212, + 489, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "OmniNOCS has 97 object categories across 380k images. This by far exceeds the diversity of existing NOCS datasets (with less than 6 categories as shown in Table 1). The number of instances exceeds that of other NOCS datasets by more than 2 orders of magnitude. It also contains more images with NOCS annotations (380k). Note that while Wild6D has 1M images, these do not have ground truth NOCS or pose annotations, and includes every frame of object-centric videos. OmniNOCS is also more diverse than the most diverse 3D detection dataset (Omni3D [7] with 98 categories). 307k images in OmniNOCS are from real world scenes and 73k are synthetic. 107k of the images in OmniNOCS are from outdoor scenes, 102k images are from complex indoor environments, and 140k are from object-centric videos.", + "bbox": [ + 212, + 580, + 789, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Data preparation", + "text_level": 1, + "bbox": [ + 215, + 753, + 377, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computing the NOCS for objects in an image requires canonically oriented 3D object bounding boxes, depth, and instance segmentations. Since the datasets we use were originally only proposed for monocular 3D detection, many of them lack dense depth or instance segmentation annotations (highlighted entries in Table 2). Additionally, NOCS", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8857892129c5c11d3852279137f8cbc2f147c75cdc8c922918ad788fa4998aef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
3D DatasetDepth sourceInstance mask source#Images#Classes
KITTI [19]LiDARHuman-labelled [23]7.4k7
SUN-RGBD [45]Depth CameraSegmentation model [26]10k72
Objectron [1]NeRFNeRF + Human label132k9
nuScenes [10]LiDARSegmentation Model [14]30k9
Hypersim [44]SyntheticSynthetic64k31
ARKitScenes [5]Depth CameraSegmentation model [26]60k15
Cityscapes 3D [15]StereoHuman-labelled3.4k8
Virtual KITTI [9]SyntheticSynthetic4.1k3
NOCS-Real275 [49]Depth cameraHuman-labelled7k6
Waymo OD [46]LiDARHuman-labelled62k7
OmniNOCS380k97
", + "bbox": [ + 245, + 142, + 753, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Constituent 3D detection datasets used in OmniNOCS: We augment many of these datasets with depth and masks, as they are missing in the original data (highlighted entries). In addition, we canonicalize the orientations of bounding boxes across all datasets. #Classes lists the number of classes we use from each dataset. Those above the dashed line are part of Omni3D [7].", + "bbox": [ + 212, + 325, + 787, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "also requires all instances within a category to have consistent orientation annotations. While this is certainly not true across different 3D detection datasets, it is also not true within some large synthetic datasets such as Hypersim. OmniNOCS aggregates several datasets, with the addition of 1) new depth estimated via Mip-NeRF reconstructions, 2) additional instance masks via segmentation models and human labels, 3) manual labelling of coordinate axes to give consistent object-centric coordinate frames. This enables NOCS to be computed for every image in the dataset. This multi-stage process for OmniNOCS creation is illustrated in Figure 2 and explained below. A summary of the resulting OmniNOCS dataset is provided in Table 2.", + "bbox": [ + 212, + 398, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Orientation canonicalization: Although the datasets we use contain oriented 3D bounding boxes, they vary in their level of canonicalization. Some datasets have their class-canonical orientations i.e., all instances of a particular class are oriented consistently within the dataset (for example, all cars in [19] have X axis pointing forward, and Z upwards). In such cases, we ensure that this canonicalization is consistent with all of OmniNOCS by applying a fixed class-specific offset orientation for the dataset. Datasets like Hypersim [44] have no class canonicalization at all, ie, although objects have tight bounding boxes, their XYZ axes directions are different for each instance. We manually label each object in such datasets to select the canonical orientation out of six possible orientations for the bounding box. More details on the labelling process are provided in the supplementary material. Some classes may have more than one choice for canonical orientations (due to symmetry), in which case an orientation is selected arbitrarily.", + "bbox": [ + 212, + 535, + 789, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Depth augmentation: For outdoor datasets, we use sparse depth from LiDAR if available. We recomputed the depth on Cityscapes using a state-of-the-art stereo depth model [55], as we found the depth from the original dataset to be noisy. Since Objectron [1] does not contain dense depth, we train Neural Radiance Fields [4, 37] for each video sequence in the dataset to obtain dense depth.", + "bbox": [ + 212, + 717, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance segmentation: Since many datasets we use were intended for 3D detection, they do not camera instance segmentations. We annotate ARKitScenes and SUNRGBD objects using instance masks from Segment-Anything (SAM) [26], prompting it with the", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "projected 3D bounding box. Although SUN-RGBD provides segmentation masks, we find that the SAM annotations are of superior quality. For Objectron, we create accurate masks by efficiently annotating in 3D space. Specifically, we apply the pipeline of [4] to each of the Objectron videos. We fuse the resulting NeRF depth-maps to create a mesh. We post-process the 3D mesh, and get rid of the redundant vertices while keeping the vertices of the object. Finally, we create the masks for multiple frames by measuring the distance between each pixel on the depth map and the object mesh: pixels whose depth is far from the surface of the mesh are discarded, a process which also handles occlusions. NOCS computation: As proposed in [49], the NOCS coordinate $x_{nocs}$ for a point on the object surface is defined as:", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {n o c s} = \\frac {1}{s} ^ {o b j} T _ {c a m} x _ {c a m} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 319, + 785, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where ${}^{obj}T_{cam} = \\left[{}^{obj}R_{cam}\\right|^{obj}t_{cam}$ is the 6DoF transformation from the camera to the object frame, $x_{cam}$ the 3D location of the point in the camera frame, and $s$ is a scalar, the size of the diagonal of the object's tight bounding box. It can be interpreted as the object shape scaled to a box with a unit diagonal. For each image, starting from a 3D pointcloud $(x_{cam})$ obtained from the depth, we collect points on each object $(x_{cam}^i)$ using its 3D bounding box and instance mask, which can be transformed to the NOCS coordinate $(x_{nocs}^i)$ using (1). We store NOCS as a 3-channel 2D map (i.e. $x_{nocs}^i$ at its corresponding 2D location obtained by projecting $x_{cam}^i$ ). Our final result is the OmniNOCS dataset which contains instance segmentations, NOCS maps, and 3D bounding boxes for objects across 97 classes.", + "bbox": [ + 212, + 359, + 787, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 NOCSformer model", + "text_level": 1, + "bbox": [ + 214, + 534, + 419, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose a novel architecture for monocular NOCS prediction termed \"NOCsformer\" trained on OmniNOCS. NOCsformer primarily uses self-attention layers, and is the largest trained monocular NOCS prediction model to date. NOCsformer also contains a 3D size head and a learned PnP head that are used to estimate a canonically oriented 3D object bounding box from the NOCS.", + "bbox": [ + 212, + 566, + 787, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Model architecture", + "text_level": 1, + "bbox": [ + 214, + 664, + 393, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 3, the NOCSformer architecture comprises an image backbone, a NOCS head, a size head, and a learned PnP head.", + "bbox": [ + 212, + 689, + 785, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Backbone:** Our backbone is a frozen Vision Transformer (ViT) [17], that uses DINOv2 weights [40] from self-supervised pretraining. This choice is motivated by recent findings on using frozen DINOv2 backbones for understanding depth, multi-view correspondences, or relative pose [20, 56]. Additionally, we use a Dense Prediction Transformer (DPT) [43] architecture to fuse low-resolution DINOv2 representations from multiple intermediate layers and upsample them by a factor of 8. This higher feature resolution is desirable to improve predictions on smaller objects. We train the DPT layers while keeping the ViT layers frozen. We use 2D input boxes to sample the DPT features using", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54ac7709470a5519de97dce65bee88326729adc990e8400865551bc77a2d6051.jpg", + "image_caption": [ + "Fig. 3: Architecture for NOCSformer: We use a transformer backbone to extract features from the input image, pool them using the 2D box RoIs, and feed the per-RoI features to the NOCS and size heads. Our novel NOCS head jointly predicts the NOCS and instance mask for the RoI. Our learned PnP head for pose estimation uses the predicted NOCS and instance mask to predict the projected 3D coordinate and 3D rotation of the object." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RoIAlign [22] into a 28x28 fixed resolution grid that is fed to the NOCS and size heads. While DPT has been previously leveraged for tasks that provide dense supervision, such as semantic segmentation and depth prediction, we use it in an instance-level prediction context, training it using supervision from object RoIs.", + "bbox": [ + 212, + 345, + 787, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NOCS head: All existing NOCS prediction models [32, 48, 49] use MaskRCNN heads [22] with convolutional layers, in some cases with a separate head for each NOCS coordinate [49]. While this works on small datasets, we find that it significantly limits performance on OmniNOCS. NOCSformer therefore uses a simple (but large) transformer NOCS head with 10 self-attention blocks and one final upsampling conv layer to jointly predict both the instance mask and the NOCS map in the same output head. Similar to [49], for each NOCS coordinate, we predict logits over non-overlapping bins, and use the softmax to obtain the final NOCS value. Our ablations show that this choice works best on our challenging OmniNOCS dataset.", + "bbox": [ + 212, + 406, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Size and learned PnP heads: We use a size head to predict the 3D size to scale the NOCS predictions to metric object coordinates. Our size head also uses self-attention blocks. Contrary to previous work [7,28,49], it does not use any class-specific layers or per-class average size statistics, as these limit scalability to larger and diverse vocabularies. Our learned PnP head is inspired from [32,48] and uses convolutional layers to predict the 3D orientation and a 2D projection of the 3D centroid. The orientation is predicted using the 6D partial rotation matrix representation [57] in allocentric space [7,28].", + "bbox": [ + 212, + 541, + 789, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Localization from NOCS predictions", + "text_level": 1, + "bbox": [ + 212, + 666, + 513, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To localize the object in 3D, we first scale the predicted NOCS using the predicted size to obtain unnormalized metric object coordinates. NOCSformer's learned PnP head predicts the object orientation in allocentric space, which is converted to the egocentric frame using the predicted projection of the 3D centroid. Using the unnormalized 3D object coordinates, object orientation, and the 3D centroid ray, the object depth is estimated from the corresponding 2D points.", + "bbox": [ + 212, + 688, + 787, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternatively, the 6DoF pose for the object can be solved directly (without using the learned PnP head) by solving a PnP problem using efficient solvers [30]. However, the orientations so obtained are known to be more sensitive to errors in NOCS coordinates, and less robust when compared to learned methods [32, 48].", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 379, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Losses", + "text_level": 1, + "bbox": [ + 215, + 146, + 303, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When training NOCSformer, we supervise the NOCS, mask, 3D size, orientation and centroid predictions using appropriate losses. We supervise predicted masks using a simple L2 loss with respect to the ground truth instance mask. Since our NOCS head uses binned prediction where the final NOCS value is a softmax over non-overlapping bin logits, we use a combination of cross-entropy and regression losses.", + "bbox": [ + 212, + 171, + 787, + 248 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a s k}} = \\left\\| \\operatorname {m a s k} _ {\\text {p r e d}} - \\operatorname {m a s k} _ {\\text {g t}} \\right\\| _ {2} \\\\ \\mathcal {L} _ {N O C S} = \\operatorname {s o f t m a x} \\mathrm {C E} (\\hat {\\mathbf {n}}, \\hat {\\mathbf {n}} _ {g t}) + | | \\mathbf {n} - \\mathbf {n} _ {g t} | | _ {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 262, + 658, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{n}}$ are the predicted logits over the discretized bins and $\\mathbf{n}$ the continuous NOCS coordinate prediction.", + "bbox": [ + 212, + 308, + 785, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In addition to the supervised loss above, we also use a variant of the self-supervised NOCS loss $\\mathcal{L}_{ss}$ from [11] that minimizes the reprojection error of the predicted NOCS using the ground truth pose and predicted mask. More details about $\\mathcal{L}_{ss}$ are in the supplementary material. The NOCS and mask losses are applied to the fixed resolution grid predictions. Since our NOCS ground truth can be sparse, it is applied at only those locations on the grid that have valid NOCS ground truth.", + "bbox": [ + 212, + 339, + 785, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since our 3D size head also uses binned prediction, we use a combination of softmax cross-entropy and L1 loss for supervision. Note that our size loss is normalized by the ground truth size, in order to penalize errors on smaller objects equally. For the learned PnP head, we also supervise the centroid and rotation predictions using L1 losses in their output space.", + "bbox": [ + 212, + 430, + 785, + 503 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i z e}} = \\operatorname {s o f t m a x C E} (\\hat {\\mathbf {s}}, \\hat {\\mathbf {s}} _ {g t}) + | \\mathbf {s} - \\mathbf {s} _ {\\mathbf {g} t} | / \\mathbf {s} _ {\\mathbf {g} t} \\\\ \\mathcal {L} _ {r o t} = \\left\\| ^ {c} \\mathbf {R} _ {o _ {p r e d}} - ^ {c} \\mathbf {R} _ {o _ {g t}} \\right\\| _ {1, 1} \\\\ \\mathcal {L} _ {\\text {c e n t r o i d}} = \\left\\| \\mathbf {c} _ {o} - \\mathbf {c} _ {o g t} \\right\\| _ {2} \\\\ \\mathcal {L} _ {P n P} = \\mathcal {L} _ {r o t} + \\mathcal {L} _ {c e n t r o i d} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 521, + 651, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The total loss for training NOCSformer is a weighted sum:", + "bbox": [ + 238, + 604, + 627, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = w _ {\\text {s i z e}} \\mathcal {L} _ {\\text {s i z e}} + w _ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + w _ {\\text {N O C S}} \\mathcal {L} _ {\\text {N O C S}} + w _ {s s} \\mathcal {L} _ {s s} + w _ {P n P} \\mathcal {L} _ {P n P}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 632, + 771, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 670, + 356, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although previous works predict NOCS accurately on a few categories [32, 48, 49], they only evaluate on 3D detection (localization) or pose estimation tasks, without quantifying the accuracy of the predicted NOCS. In section 5.1, we propose metrics and establish a benchmark to evaluate NOCS on the OmniNOCS dataset. We compare the localization accuracy of NOCSformer against existing benchmarks, by evaluating its localization accuracy on nuScenes [10] in section 5.2. We also evaluate the unique ability of NOCSformer to transfer to unseen datasets and domains. Finally, we provide ablations on critical design choices made in our model in section 5.4.", + "bbox": [ + 212, + 702, + 787, + 824 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2aa0ca2f97ae637fa30545efd1f1ae8588bb1f4cb61de62055b8d308fd639897.jpg", + "image_caption": [ + "Input: Image + 2D box" + ], + "image_footnote": [], + "bbox": [ + 223, + 172, + 406, + 267 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7b301e035cb9a8d792ee47cfe0d9599f6706dbcc92612ef7f286c08862e669ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 268, + 405, + 361 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/374cb9ec2555543a715f6dc158b5c9f71b8bf0a1361435068508944cbd3e1f21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 362, + 405, + 436 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c90f8e2d618abcd867d7ddadf1c909e5f57fd75cc105dc447ac07813344f5228.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 438, + 405, + 527 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/96c2e128ca476ddc8a22a180ed59a9c443c6f8d4ea3609bb8c624b657e935d26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 529, + 405, + 619 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a75054222dc8a2216e3c6c7b26086911937082e67b44a8ebffcdc7467c9d1cd2.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 223, + 635, + 310, + 720 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0ef75d46db01758059058fdc5084c36ae8f7da4761f8b7e7b386cd2c7d3f1770.jpg", + "image_caption": [ + "Output" + ], + "image_footnote": [], + "bbox": [ + 310, + 635, + 395, + 720 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7c4c4a5913f8a345de088583bbb8d42af25a19035a8a360d0213197d5bb428f5.jpg", + "image_caption": [ + "Output: NOCS + 3D box" + ], + "image_footnote": [], + "bbox": [ + 419, + 172, + 601, + 267 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7fe99a872f668d3e02d09b6a612fa8fa521921e1e8846569e8b5053137341184.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 268, + 601, + 361 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a7d3b1bc3efc9aa5aff2ea352c5a4e219af6004624e4599109ad87551ed346fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 362, + 601, + 436 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d1e6c396d4789631aed34fd84b72253a13bbc7c4726a8b8a8d149548ebd6d902.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 438, + 601, + 527 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c93c50b71f3a0679cbee0d98cd5ddb9b7d40baa192daae55e87c6277130e90f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 529, + 601, + 619 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/edaef087919be95f65a5f8c5ffbe06550f0d08f4af2413dd3ce1bfa071c9615f.jpg", + "image_caption": [ + "3D shape + pose" + ], + "image_footnote": [], + "bbox": [ + 401, + 636, + 516, + 712 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/135fba02399742b52f7bf8d47d37cdc60ac04fdf5c906eeef00a8a3e17f2c426.jpg", + "image_caption": [ + "Input", + "Fig. 4: Example results of our single unified NOCSformer model across various datasets and object classes. The left column shows input images and query 2D bounding boxes. The center column shows the NOCS+instance maps predicted by NOCSformer along with the estimated 3D pose overlaid on the input image. The NOCS can be used with the 3D boxes and object size to lift the objects to a 3D pointcloud, which is shown in the right column. Last row contains two examples." + ], + "image_footnote": [], + "bbox": [ + 534, + 635, + 607, + 720 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3a6dd6f3997c50cc674624ac3f3d733e29435e73516117f81e5c2b75dd73fed8.jpg", + "image_caption": [ + "Output: 3D shapes and poses" + ], + "image_footnote": [], + "bbox": [ + 607, + 172, + 781, + 361 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/097c300d06565cea7c36582b922499faea658fd528ff385a152f90fca82cbed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 371, + 777, + 440 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/28e41a371836d1d79a9b4750352f22b011c92bc830eb73c3691e3ae75ee379db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 440, + 779, + 522 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1cbaa09de46cdf437354ebc4676ed9502833e92d77b062bdf21cac896a6c73b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 531, + 761, + 614 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/22cc7db8e1952ec5bf580d17ccce4c8628cc0343d2d6594597231b55fd5e1863.jpg", + "image_caption": [ + "Output", + "3D shape + pose" + ], + "image_footnote": [], + "bbox": [ + 609, + 635, + 683, + 720 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7e24f158250b514172e978841e99570edd2f3e7d1ccdce649cdf5224a056b9d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 640, + 777, + 708 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 218, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets: For our experiments, we train NOCSformer on the OmniNOCS dataset containing 97 classes, holding out images from the NOCS-Real275 dataset. We use NOCS-Real275 to evaluate NOCSformer's cross-dataset generalization capabilities by not training on it. NOCS-Real275 has 6 classes that overlap with the rest of OmniNOCS, although they are not the top 20 classes. It also differs from the rest of OmniNOCS in terms of camera parameters and the context of objects in the scene.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 NOCS prediction accuracy", + "text_level": 1, + "bbox": [ + 214, + 258, + 444, + 273 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We propose the use of NOCS mAE and NOCS mPSNR to evaluate the accuracy of predicted NOCS. For each object the mean Absolute Error (mAE) and mPSNR are computed for all points in the intersection of ground truth and predicted masks, and reported as a mean of per-category means. However, the NOCS mAE/mPSNR do not penalize undersegmentations / sparse predictions. Since we also would like the NOCS predictions to span the full visible instance, we also evaluate the 2D mask mIoU.", + "bbox": [ + 212, + 285, + 787, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate NOCSformer on a subset of OmniNOCS that contains 75 classes with accurate ground truth NOCS and masks, see Table 3. It is able to predict NOCS with errors less than $9\\%$ (0.089 mAE on OmniNOCS). We also evaluate on the held-out NOCS-Real275 dataset, which was not used to train NOCSformer. We find that the zero-shot NOCSformer outperforms the NOCS baseline from [49] that was fully trained on this dataset, a strong indication of the generalization capabilities of NOCSformer.", + "bbox": [ + 212, + 375, + 787, + 467 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 3D localization accuracy", + "text_level": 1, + "bbox": [ + 214, + 488, + 428, + 503 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "NOCS provides dense 3D-2D correspondences that can be used to estimate the 3D object oriented bounding box. This is done by using the predicted object size and solving a PnP problem, as explained in Sec. 4.2. We evaluate the accuracy of our estimated bounding boxes in different settings:", + "bbox": [ + 212, + 513, + 785, + 573 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Outdoor scenes: We compare NOCSformer's 3D localization accuracy to that of other 3D detection models using the nuScenes true positive localization metrics on the challenging nuScenes dataset (included in OmniNOCS). We compare to Cube R-CNN, since it is the only other model that generalizes across diverse datasets. However, Cube R-CNN has two notable differences to NOCSformer: 1) being a detection model, it also jointly detects 2D object regions, and 2) it uses Chamfer loss causing it to have high orientation errors. As a more comparable baseline, we use a variant of our model (termed \"Cubeformer\"), by replacing our NOCS head with the cube head from Cube R-CNN", + "bbox": [ + 212, + 574, + 787, + 696 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b77edfab2528231a68f4af343c465b429d28c40490ef491c9b373dfe85d7054a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNOCS-Real275OmniNOCS
NOCS MAE↓NOCS PSNR↑Mask IoU↑NOCS MAE↓NOCS PSNR↑Mask IoU↑
NOCS baseline [49]0.12116.34586.10---
NOCSformer0.10718.52789.030.09420.24578.50
", + "bbox": [ + 218, + 720, + 782, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: NOCS and mask prediction evaluation: We report metrics for NOCSformer on 75 classes of OmniNOCS: it is the first model that is capable on predicting NOCS on such diverse data. Additionally, NOCSformer outperforms [49] which trained on NOCS-Real275, without training on NOCS-Real275.", + "bbox": [ + 212, + 781, + 787, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/36b4c61456202fa8216ed3b716aa91001338343ab65fd800c04acd774d438344.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMulti-datasetmATE (m)↓mAOE (rad)↓mASE (%)↓mIoU↑
FCOS3D [51]X0.7770.4000.231-
PGD [50]X0.6750.3990.236-
EProPNP [12]X0.6760.3630.263-
EProPNP + TTA [12]X0.6530.3190.255-
Cube R-CNN [7]✓0.6501.3050.2830.349
Cubeformer✓0.7900.5730.3010.280
NOCSformer✓0.8870.5580.2910.377
", + "bbox": [ + 245, + 142, + 753, + 263 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "without a Chamfer loss. It uses the input 2D boxes. More details about Cubeformer are in the supplementary material. From Table 4, we find that NOCSformer's box orientations are canonical and more accurate than both Cube R-CNN and Cubeformer. Moreover, the boxes estimated from NOCSformer's NOCS predictions are comparable in translation and scale errors to those of Cube R-CNN and Cubeformer, even though it does not directly regress depth. While all 3 methods generalize across several classes and domains, baselines trained solely on nuScenes localization (top half) significantly outperform them on this task.", + "bbox": [ + 212, + 363, + 787, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cross-dataset generalization (indoor): Here, we hold out NOCS-Real275 from OmniNOCS when training NOCSformer, and use it to evaluate NOCSformer's ability to generalize to unseen domains. NOCS-Real275 features a tabletop setting with multiple objects, which is semantically different from our other indoor datasets. We compare against Cube R-CNN [7] and the NOCS [49] model that was trained from scratch on this dataset alone. [49] also uses class-specific model parameters and additional losses for symmetric objects that do not scale to larger vocabularies.", + "bbox": [ + 212, + 486, + 787, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results are shown in Table 5. We find that NOCSformer is more accurate at transferring to this unseen dataset compared to Cube R-CNN, indicating the generalizability of NOCS-based localization methods over that of box-regression methods. Note that the mAP metrics used in the NOCS-Real275 datasets may also be affected by the false positives/negatives from Cube R-CNN. However, both models are worse than the supervised NOCS baseline.", + "bbox": [ + 212, + 593, + 787, + 684 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b486b9e61039fc93e7a94311ff611443e26d0ab2c993cb545b7e318b1b450033.jpg", + "table_caption": [ + "Table 4: Outdoor localization: Comparison of 3D localization accuracy on the nuScenes subset of OmniNOCS. Top half shows methods that are only trained on nuScenes: while these perform better on nuScenes itself, they do not generalize to other datasets and classes. OmniNOCS is competitive with Cube R-CNN on mIoU while also predicting canonical orientations and object coordinates. Note that mIoU and orientation error only uses the predicted yaw orientation component." + ], + "table_footnote": [], + "table_body": "
MethodTransferred?Depth input?mAP
3D IoU @ 253D IoU @ 50
NOCS model [49]X✓79.672.4
CubeRCNN [7]✓X14.94.1
NOCSformer✓X43.510.6
", + "bbox": [ + 315, + 712, + 683, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5: Cross-dataset generalization: Comparison of localization accuracy on NOCS-Real275 for all classes, using mAP at different thresholds. This dataset has been held out when training CubeRCNN and NOCSformer. NOCSformer is able to generalize to NOCS-Real275 without any additional training.", + "bbox": [ + 212, + 781, + 787, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 379, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7c23736e5a82817d3573db53649829c2d26c417528908f20903bc18b98a89633.jpg", + "image_caption": [ + "Fig. 5: Generalization across datasets: NOCSformer can generalize to new datasets that present new camera models and object domains. We show this zero-shot capability of NOCSformer (bottom row) on the NOCS-Real275 test set without training on NOCS-Real275 dataset. The predictions from a NOCS baseline [49] trained explicitly on the NOCS-Real275 dataset are shown in the top row." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 357 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 3D orientation accuracy", + "text_level": 1, + "bbox": [ + 215, + 473, + 424, + 489 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A key challenge for 3D localization models trained on multiple datasets [7] is predicting canonical object orientations, as these datasets typically differ in their object pose conventions. OmniNOCS ensures that object poses in all constituent datasets are consistent within a category, enabling our model to predict canonical orientations. We evaluate this by computing the accuracy at different orientation error threshold on KITTI-val [13] in Table 6. We compare the accuracy of our orientation predictions to those of Cube R-CNN [7], which are trained using Chamfer loss, and therefore suffers at predicting canonical orientations. We also compare against our Cubeformer baseline that directly supervises the orientations. We find that orientations estimated using NOCSformer are more consistent and accurate than both our baselines.", + "bbox": [ + 212, + 505, + 787, + 656 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b3d10bba7163610c13afa5a32194c8dbcabda78678a0e4c05ff2081feb80e99c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGravity axisHeading axis (X)
Accuracy @ 1 degAccuracy @ 5 degAccuracy @ 90 degAccuracy 5 degAccuracy @ 10 degAccuracy @ 90 deg
CubeRCNN5.7518.2951.7419.4321.5534.17
Cubeformer80.7998.52100.041.9955.8877.26
NOCSformer84.2499.95100.049.1659.6581.99
", + "bbox": [ + 246, + 683, + 751, + 772 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: 3D orientation accuracy for models trained on multiple datasets: Models that use Chamfer distance (CubeRCNN) for supervising orientation heads end up being inconsistent in their orientation predictions. The results are averaged over 5 classes in the KITTI-val subset [13].", + "bbox": [ + 212, + 781, + 787, + 824 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a21901c120e8bd14572d38ec79463acc51f5a7bf2b1ec3f1686f8859056f93f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureNOCS PSNR↑Mask IoU↑
NOCSformer20.43181.69
Hourglass head-3.064-3.16
MaskRCNN head-5.23-21.27
w/o discretized prediction-4.11-
", + "bbox": [ + 338, + 143, + 666, + 213 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 7: Comparison of different architecture choices for the NOCS and mask prediction head: We experiment with transformer, hourglass and MaskRCNN architectures. We compare whether discretized prediction of NOCS is better than continuous regression for transformer heads. We quantify the difference in performance when using separate heads for mask and NOCS prediction as opposed to a single head.", + "bbox": [ + 212, + 232, + 787, + 303 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Ablations", + "text_level": 1, + "bbox": [ + 215, + 347, + 323, + 361 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "NOCS head architecture: While all existing architectures use a few CNN layers for their NOCS heads [49], we find that this significantly limits the NOCS prediction performance when scaling to more classes. Using our transformer NOCS head provides a 5.23dB improvement in NOCS PSNR and a $21.27\\%$ improvement in mask mIoU. We also experiment with a larger convolutional head: the Hourglass model from [6]. We find that the NOCSformer NOCS heads are even better than Hourglass, with an improvement of 3.06dB on NOCS PSNR and $3.16\\%$ on mask IoU (as shown in Table 7).", + "bbox": [ + 212, + 378, + 787, + 482 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Discretized versus continuous prediction of NOCS: NOCSformer models the final NOCS prediction as a classification layer with 50 non-overlapping bins. This was observed to be better than a linear regression layer with a MaskRCNN head in [49]. We find that it is also significantly better when using a transformer head, improving the NOCS prediction PSNR by 4.11 dB, as in Table 7.", + "bbox": [ + 212, + 484, + 787, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Future work", + "text_level": 1, + "bbox": [ + 214, + 589, + 491, + 606 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper has introduced a new large scale dataset of Normalized Object Coordinates (NOCS) for a wide variety of object classes in indoor and outdoor scenes. It has also proposed a single transformer-based model NOCSformer that can predict object 3D pose, size, and NOCS for any of these objects given 2D bounding box inputs. These allows our model to obtain 3D shape and oriented bounding boxes of objects in metric scale from a single input image. This represents the first attempt to generalize NOCS estimation beyond small datasets of narrow domain, increasing the number of object categories available by an order of magnitude. We hope this provides a means for others to explore large-scale monocular 3D object pose and shape estimation.", + "bbox": [ + 212, + 627, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Some limitations of our current method include the handling of symmetric objects, where the coordinate system has multiple possible solutions, and reflected geometry, such as left / right shoes. Future work could address these issues, for example, minimizing over multiple coordinate frames in the loss for symmetric objects, and potential estimation of left / right coordinate frames for reflected geometry.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2021)", + "2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: CVPR (2019)", + "3. Avetisyan, A., Dai, A., Nießner, M.: End-to-end cad model retrieval and 9dof alignment in 3d scans. In: ICCV (2019)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)", + "5. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1) (2021), https://openreview.net/forum?id=tjZjv_qh_CE", + "6. Birodkar, V., Lu, Z., Li, S., Rathod, V., Huang, J.: The surprising impact of mask-head architecture on novel class segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7015-7025 (2021)", + "7. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023)", + "8. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019)", + "9. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2. arXiv preprint arXiv:2001.10773 (2020)", + "10. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "11. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "12. Chen, H., Wang, P., Wang, F., Tian, W., Xiong, L., Li, H.: Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2781-2790 (2022)", + "13. Chen, X., Kundu, K., Zhu, Y., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals using stereo imagery for accurate object class detection. IEEE transactions on pattern analysis and machine intelligence 40(5), 1259-1272 (2017)", + "14. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: CVPR (2020)", + "15. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)", + "16. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + ], + "bbox": [ + 220, + 179, + 785, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "18. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022)", + "19. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)", + "20. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: European Conference on Computer Vision. pp. 516-532. Springer (2022)", + "21. Gumeli, C., Dai, A., Nießner, M.: Roca: Robust cad model retrieval and alignment from a single image. In: CVPR (2022)", + "22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)", + "23. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 923-934 (2021)", + "24. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: IEEE International Conference on Robotics and Automation (ICRA) (2022), https://arxiv.org/abs/2203.01929", + "25. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape appearance and pose optimization. In: European Conference on Computer Vision (ECCV) (2022), https://arxiv.org/abs/2207.13691", + "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "27. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation. In: CVPR (2022)", + "28. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375", + "29. Kuo, W., Angelova, A., Lin, T.Y., Dai, A.: Mask2cad: 3d shape prediction by learning to segment and retrieve. In: ECCV (2020)", + "30. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision 81, 155-166 (2009). https://doi.org/ https://doi.org/10.1007/s11263-008-0152-6, http://infoscience. epfl.ch/record/160138", + "31. Li, K., DeTone, D., Chen, Y.F.S., Vo, M., Reid, I., Rezatofighi, H., Sweeney, C., Straub, J., Newcombe, R.: Odam: Object detection, association, and mapping using posed rgb video. In: ICCV (2021)", + "32. Li, Z., Wang, G., Ji, X.: Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7678-7687 (2019)", + "33. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15641-15650 (2021)", + "34. Manhardt, F., Kehl, W., Gaidon, A.: Roi-10d: Monocular lifting of 2d detection to 6d pose and metric shape. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2069-2078 (2019)" + ], + "bbox": [ + 215, + 147, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Vid2cad: Cad model alignment using multi-view constraints from videos. TPAMI (2022)", + "36. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Cad-estate: Large-scale cad model annotation in rgb videos. In: ICCV (2023)", + "37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "38. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017)", + "39. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020)", + "40. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "41. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: Proceedings of the IEEE international conference on computer vision. pp. 3828-3836 (2017)", + "42. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3D appearance, location & pose. In: CVPR (2022)", + "43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)", + "44. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: International Conference on Computer Vision (ICCV) 2021 (2021)", + "45. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 567-576 (2015). https://doi.org/10.1109/CVPR.2015.7298655", + "46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "47. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: CVPR (2018)", + "48. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16611-16621 (2021)", + "49. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "50. Wang, T., Xinge, Z., Pang, J., Lin, D.: Probabilistic and geometric depth: Detecting objects in perspective. In: Conference on Robot Learning. pp. 1475-1485. PMLR (2022)", + "51. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)", + "52. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022)" + ], + "bbox": [ + 212, + 146, + 789, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OmniNOCS", + "bbox": [ + 655, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Wen, B., Lian, W., Bekris, K., Schaal, S.: Catgrasp: Learning category-level task-relevant grasping in clutter from simulation. ICRA 2022 (2022)", + "54. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017)", + "55. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21919-21928 (2023)", + "56. Zhang, J., Herrmann, C., Hur, J., Cabrera, L.P., Jampani, V., Sun, D., Yang, M.H.: A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence arXiv preprint arxiv:2305.15347 (2023)", + "57. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + ], + "bbox": [ + 215, + 147, + 785, + 327 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Krishnan et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_model.json b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2d68fd737866af4276681eb563439776a06ed2c3 --- /dev/null +++ b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_model.json @@ -0,0 +1,2623 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.142, + 0.783, + 0.187 + ], + "angle": 0, + "content": "OmniNOCS: A unified NOCS dataset and model for 3D lifting of 2D objects" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.788, + 0.245 + ], + "angle": 0, + "content": "Akshay Krishnan\\(^{1,2}\\), Abhijit Kundu\\(^{1}\\), Kevis-Kokitsi Maninis\\(^{1}\\), James Hays\\(^{2}\\), and Matthew Brown\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.254, + 0.564, + 0.269 + ], + "angle": 0, + "content": "1 Google Research†" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.27, + 0.604, + 0.284 + ], + "angle": 0, + "content": "2 Georgia Institute of Technology" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.292, + 0.395, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.293, + 0.588, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.296, + 0.777, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.465, + 0.789, + 0.522 + ], + "angle": 0, + "content": "Fig. 1: We introduce OmniNOCS, a large-scale dataset with Normalized Object Coordinates (NOCS), instance masks, and 3D box annotations for objects across several classes, domains, and cameras. We also propose NOCSformer, a model trained on OmniNOCS that lifts each 2D object bounding box in an image to its corresponding 3D oriented box (pose) and 3D pointcloud (shape)." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.537, + 0.744, + 0.745 + ], + "angle": 0, + "content": "Abstract. We propose OmniNOCS, a large-scale monocular dataset with 3D Normalized Object Coordinate Space (NOCS) maps, object masks, and 3D bounding box annotations for indoor and outdoor scenes. OmniNOCS has 20 times more object classes and 200 times more instances than existing NOCS datasets (NOCS-Real275, Wild6D). We use OmniNOCS to train a novel, transformer-based monocular NOCS prediction model (NOCsformer) that can predict accurate NOCS, instance masks and poses from 2D object detections across diverse classes. It is the first NOCS model that can generalize to a broad range of classes when prompted with 2D boxes. We evaluate our model on the task of 3D oriented bounding box prediction, where it achieves comparable results to state-of-the-art 3D detection methods such as Cube R-CNN. Unlike other 3D detection methods, our model also provides detailed and accurate 3D object shape and segmentation. We propose a novel benchmark for the task of NOCS prediction based on OmniNOCS, which we hope will serve as a useful baseline for future work in this area. Our dataset and code will be at the project website: https://omninocs.github.io." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.765, + 0.357, + 0.781 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.789, + 0.789, + 0.821 + ], + "angle": 0, + "content": "Predicting the 6 Degree-of-Freedom (6DoF) pose and shape of objects from images is a crucial problem in 3D scene understanding. Robots need to understand the location," + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.826, + 0.365, + 0.84 + ], + "angle": 0, + "content": "†Now at Google DeepMind" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.378, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "shape, and orientation of various objects to grasp and interact with them. Self-driving vehicles need to understand the location and heading of vehicles, pedestrians, and other objects on the road. In particular, the 3D orientation of these objects is crucial to predict their behavior. It is also important in AR/VR applications, as it allows users to interact with objects in meaningful ways. Predicting object shape and pose from monocular images is also a prerequisite to initialize 3D panoptic neural scene graphs [27, 39] and to track objects in videos [42]. These applications require approaches that generalize across a wide range of object classes, environments, and camera models." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.271, + 0.788, + 0.407 + ], + "angle": 0, + "content": "The problem of localizing 3D objects has been extensively studied through the lens of monocular 3D object detection or 6DoF pose estimation. The most commonly used approach is to represent objects as 3D cuboids, and train a model to regress the cuboid parameters from a 2D RoI of the object [7, 8, 28, 41, 51, 54]. An alternative method involves predicting corresponding 3D points in an object coordinate space followed by pose estimation using 3D-2D alignment [11, 12, 20, 33, 49]. However, all these methods are limited to narrow datasets collected on a single camera and context with typically very few object classes. Existing models are also trained separately for every dataset, preventing them from being used more widely." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.41, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Recent work [7] takes a step towards scalability by introducing the Omni3D dataset, which aggregates 3D detection datasets from different domains, and trains a Cube R-CNN model to regress 3D bounding boxes for 50 classes. However, Omni3D notably lacks consistently oriented, object-centric boxes, as the ground truth orientations are not canonical for object classes. Cube R-CNN [7] instead uses a Chamfer distance loss which is invariant to the predicted 3D cuboid orientation. This causes it to predict inconsistent object orientations, for example, often flipping the orientations of cars by \\(180^{\\circ}\\). Further, Omni3D is a detection dataset that does not provide object shape information." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.535, + 0.788, + 0.762 + ], + "angle": 0, + "content": "Our work aims to overcome the above shortcomings, by providing a new large-scale dataset with consistent object-centric ground truth boxes along with detailed shape. We argue for the use of Normalized Object Coordinate Space (NOCS) as proposed in [49] as a 3D object shape representation. NOCS represents both the canonical orientation and the shape of the visible surface of the object, properties that are essential for real-world applications such as self-driving and robotics. It can also be used to estimate the 3D bounding box of the object. However, all existing work on NOCS [24, 25, 32, 48, 49] only train small models on small datasets with fewer than 10 classes. We address the lack of diverse NOCS datasets by proposing a large-scale monocular NOCS dataset, OmniNOCS that has NOCS annotations, instance segmentation, and canonically oriented bounding boxes for 97 object classes, containing \\(380\\mathrm{k}\\) images from 10 different data sources, making it the largest and most diverse NOCS dataset currently available. OmniNOCS includes all of the data from Omni3D (KITTI [19], nuScenes [10], ARKitScenes [5], Objectron [1], Hypersim [44], SUN-RGBD [45]), with the addition of Cityscapes [15], virtual KITTI [9], NOCS-Real275 [49], and the Waymo Open Dataset [46]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our work also introduces a new model, which we term \"NOCSformer\", that predicts NOCS coordinates and oriented 3D bounding boxes from monocular images and 2D detections for all the classes in OmniNOCS. NOCSformer leverages self-supervised pretrained ViTs [17, 40] and does not use any class-specific model parameters. This enables it scale to large vocabularies, and share information across semantically similar" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.788, + 0.282 + ], + "angle": 0, + "content": "object classes. In contrast, existing NOCS models use small NOCS heads with class-specific parameters that significantly limit their performance on large datasets like OmniNOCS. NOCSformer also predicts the 3D size of the object along with its 3D orientation using a learned PnP head. This allows NOCSformer to predict 3D oriented bounding boxes and object point cloud in metric scale for diverse object categories. Our experiments evaluate the quality of NOCS and bounding boxes predicted by NOCSformer in comparison to existing NOCS prediction or 3D detection models. We find that training on OmniNOCS allows NOCSformer to generalize to unseen datasets, even outperforming baselines trained on the target dataset in NOCS prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.284, + 0.47, + 0.298 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.301, + 0.784, + 0.342 + ], + "angle": 0, + "content": "- The OmniNOCS dataset: A new dataset containing Normalised Object Coordinates for 380k images in 97 categories, an order of magnitude larger on both counts than existing NOCS datasets." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.346, + 0.785, + 0.404 + ], + "angle": 0, + "content": "- NOCSformer: A novel transformer-based architecture for predicting object NOCS, mask, and size from input 2D boxes, utilizing pre-trained self-supervised ViT backbones. NOCSformer is the first NOCS model to generalize to vocabularies with \\(90+\\) classes and to unseen datasets, including images from internet collections." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.407, + 0.785, + 0.448 + ], + "angle": 0, + "content": "- OmniNOCS benchmark: an evaluation framework with metrics for directly comparing different NOCS prediction algorithms on OmniNOCS, with baselines established via NOCSFormer." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.301, + 0.785, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.475, + 0.362, + 0.491 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.506, + 0.788, + 0.582 + ], + "angle": 0, + "content": "The task of predicting 3D object pose has been studied both in the context of camera/object pose estimation (6DoF) and 3D bounding box estimation (6DoF pose + 3DoF size). A further distinction can be made between methods that regress directly to object pose / bounding box parameters, and methods that compute per pixel coordinates or depth as an initial step. We review each of these paradigms in the sections below." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.604, + 0.571, + 0.62 + ], + "angle": 0, + "content": "2.1 3D localization by regressing bounding boxes" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.629, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Direct approaches to 3D object detection and pose estimation involve networks that output rotation, translation and scale parameters directly. Several works have explored different parameterizations in this setting, e.g., PoseCNN [54] uses regression of translation via centre direction + distance maps (which enable detection even under occlusion), and quaternion representation of rotation. BB8 [41] uses segmentation followed by a CNN to regress to 2D projections of the \\(8 \\times 3\\mathrm{D}\\) bounding box corners. Multi-view monocular approaches have also been proposed, e.g., DETR3D [52] which uses DETR-style attention to reason about object interactions. Several techniques also make use of existing 2D bounding box predictors, either as an input to a 3D lifting approach [34], or as a constraint on 3D box predictions [38]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Another group of works focuses on predicting the alignment of 3D CAD models within various modalities: images [21, 29], videos [31, 35], or 3D scans [2, 3]. These approaches typically determine the object's 3D translation, rotation, and size (9DoF) and find a CAD model with a similar visual appearance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.38, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.586, + 0.162 + ], + "angle": 0, + "content": "2.2 3D localization from model-to-image alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.258 + ], + "angle": 0, + "content": "An alternative approach to 3D localization is to first predict correspondence, either between views or to a normalized space, and then reason over redundant correspondences to establish pose. This has been done with both sparse and dense correspondences, e.g., AutoShape [33] makes use of sparse 2D to 3D correspondences, with a learned shape model and sparse 2D keypoints. [20] uses pairwise semantic correspondence from ViT, to find pose between a reference image and a sequence of targets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Several approaches use dense correspondences and the idea of Normalised Object Coordinate Spaces, or NOCS. In the original NOCS work, Wang et al. [49] use RGB-D data and 3D-3D correspondences for pose estimation. Follow-up work [11] estimates pose using PnP variants on 2D-3D correspondences. While NOCS [49] used the average spatial dimensions of the object category to define the object frame, [53] showed that an instance-specific NOCS coordinate frame can be used alongside a predicted instance size for the same purpose. Other methods combine coordinate regression and direct approaches, e.g., [48], which uses direct regression based on dense correspondences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Similar to our work, MonoRUun [11] lifts 2D detections to 3D object coordinates without explicit class supervision. They also use a novel self-supervised coordinate regression training loss which obviates the need for detailed 3D ground truth. However, their representation does not include masks and therefore does not provide explicit object shape information. Their evaluation is also limited to 3 classes (car, pedestrian, cyclist) on the KITTI-Object test set. M3D-RPN [8] also has a single 3D head for multiple classes, jointly generating 2D and 3D bounding box predictions, though without shape information, and similarly limited to car, pedestrian and cyclist classes on KITTI." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.518, + 0.645, + 0.534 + ], + "angle": 0, + "content": "2.3 Monocular object localization / pose estimation datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.629 + ], + "angle": 0, + "content": "For 3D object detection from monocular RGB images, most existing works use a small number of classes on a single dataset. Cube R-CNN/Omni3D [7] contribute towards creating a general purpose monocular 3D object detector. Cube R-CNN performs well over six 3D datasets: KITTI, SUN RGB-D, ARKitScenes, Objectron, nuScenes and Hypersim. However, the method is class specific, with Cube R-CNN trained only on 50 classes, and unable to work in the open class setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Cube R-CNN also uses a Chamfer loss on the predicted 3D box corners to deal with the inconsistent coordinate frame annotations in the underlying Omni3D datasets. The lack of direct orientation supervision results in inconsistent orientation predictions (for example, the positive x axis may point to the front or the back of the car for different instances). This orientation inconsistency is problematic for methods that seek to build detailed 3D models of object categories as a downstream task [27, 39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Other works are not limited to 3D boxes, and directly provide object shapes. These shapes come in the form of annotated point clouds (e.g., ScanNet [16]), or aligned CAD models (e.g., Scan2CAD [2], Pix3D [47], CAD-Estate [36]). Object point clouds are created by labeling points on scanned 3D scenes, a process that doesn't scale well to large datasets. Semi-automatically aligning CAD models is scale-able to a certain extent, but since it relies on retrieving existing models, the resulting shapes are rarely accurate, and the alignments are sensitive to deformations and movable parts. In contrast to these works, we propose a large-scale dataset of many different categories." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.144, + 0.699, + 0.215 + ], + "angle": 0, + "content": "
DatasetNOCS GTReal#Images#Classes#Instances
CAMERA25 [49]✓×300k6184
NOCS-Real275 [49]✓✓7k624
Wild6D [18]×✓1M51.8K
OmniNOCS✓✓380k97>450k
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "Table 1: Comparison to other NOCS datasets: Other existing NOCS datasets are limited in number of classes and instances. Note that Wild6D has 1M images, but these are used for self-supervised training, as it does not include NOCS ground truth (GT)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.272, + 0.787, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.348, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: Preparation of the OmniNOCS dataset: First, we ensure that object orientations are canonical across different all subsets of OmniNOCS. Next, we re-compute depth for datasets where depth is not available or is too noisy. Finally, we annotate objects with (pseudo) instance mask labels where ground truth masks are not available." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.418, + 0.449 + ], + "angle": 0, + "content": "3 OmniNOCS dataset" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.465, + 0.357, + 0.479 + ], + "angle": 0, + "content": "3.1 Data statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.581 + ], + "angle": 0, + "content": "We create OmniNOCS, a first-of-its-kind large and diverse NOCS dataset comprising data from several classes and domains. OmniNOCS uses data from self-driving scenes [9, 15, 19, 46], indoor scenes [5, 44, 45, 49], and object-centric videos [1]. Each of these sources use cameras with different parameters, ranging from phone cameras to wide FOV cameras mounted on self-driving cars. The number of object instances per image also varies widely, from single-object images [1] to images with hundreds of objects [44, 46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.79, + 0.732 + ], + "angle": 0, + "content": "OmniNOCS has 97 object categories across 380k images. This by far exceeds the diversity of existing NOCS datasets (with less than 6 categories as shown in Table 1). The number of instances exceeds that of other NOCS datasets by more than 2 orders of magnitude. It also contains more images with NOCS annotations (380k). Note that while Wild6D has 1M images, these do not have ground truth NOCS or pose annotations, and includes every frame of object-centric videos. OmniNOCS is also more diverse than the most diverse 3D detection dataset (Omni3D [7] with 98 categories). 307k images in OmniNOCS are from real world scenes and 73k are synthetic. 107k of the images in OmniNOCS are from outdoor scenes, 102k images are from complex indoor environments, and 140k are from object-centric videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.379, + 0.77 + ], + "angle": 0, + "content": "3.2 Data preparation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Computing the NOCS for objects in an image requires canonically oriented 3D object bounding boxes, depth, and instance segmentations. Since the datasets we use were originally only proposed for monocular 3D detection, many of them lack dense depth or instance segmentation annotations (highlighted entries in Table 2). Additionally, NOCS" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.379, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.143, + 0.754, + 0.322 + ], + "angle": 0, + "content": "
3D DatasetDepth sourceInstance mask source#Images#Classes
KITTI [19]LiDARHuman-labelled [23]7.4k7
SUN-RGBD [45]Depth CameraSegmentation model [26]10k72
Objectron [1]NeRFNeRF + Human label132k9
nuScenes [10]LiDARSegmentation Model [14]30k9
Hypersim [44]SyntheticSynthetic64k31
ARKitScenes [5]Depth CameraSegmentation model [26]60k15
Cityscapes 3D [15]StereoHuman-labelled3.4k8
Virtual KITTI [9]SyntheticSynthetic4.1k3
NOCS-Real275 [49]Depth cameraHuman-labelled7k6
Waymo OD [46]LiDARHuman-labelled62k7
OmniNOCS380k97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Table 2: Constituent 3D detection datasets used in OmniNOCS: We augment many of these datasets with depth and masks, as they are missing in the original data (highlighted entries). In addition, we canonicalize the orientations of bounding boxes across all datasets. #Classes lists the number of classes we use from each dataset. Those above the dashed line are part of Omni3D [7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.534 + ], + "angle": 0, + "content": "also requires all instances within a category to have consistent orientation annotations. While this is certainly not true across different 3D detection datasets, it is also not true within some large synthetic datasets such as Hypersim. OmniNOCS aggregates several datasets, with the addition of 1) new depth estimated via Mip-NeRF reconstructions, 2) additional instance masks via segmentation models and human labels, 3) manual labelling of coordinate axes to give consistent object-centric coordinate frames. This enables NOCS to be computed for every image in the dataset. This multi-stage process for OmniNOCS creation is illustrated in Figure 2 and explained below. A summary of the resulting OmniNOCS dataset is provided in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.79, + 0.717 + ], + "angle": 0, + "content": "Orientation canonicalization: Although the datasets we use contain oriented 3D bounding boxes, they vary in their level of canonicalization. Some datasets have their class-canonical orientations i.e., all instances of a particular class are oriented consistently within the dataset (for example, all cars in [19] have X axis pointing forward, and Z upwards). In such cases, we ensure that this canonicalization is consistent with all of OmniNOCS by applying a fixed class-specific offset orientation for the dataset. Datasets like Hypersim [44] have no class canonicalization at all, ie, although objects have tight bounding boxes, their XYZ axes directions are different for each instance. We manually label each object in such datasets to select the canonical orientation out of six possible orientations for the bounding box. More details on the labelling process are provided in the supplementary material. Some classes may have more than one choice for canonical orientations (due to symmetry), in which case an orientation is selected arbitrarily." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Depth augmentation: For outdoor datasets, we use sparse depth from LiDAR if available. We recomputed the depth on Cityscapes using a state-of-the-art stereo depth model [55], as we found the depth from the original dataset to be noisy. Since Objectron [1] does not contain dense depth, we train Neural Radiance Fields [4, 37] for each video sequence in the dataset to obtain dense depth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Instance segmentation: Since many datasets we use were intended for 3D detection, they do not camera instance segmentations. We annotate ARKitScenes and SUNRGBD objects using instance masks from Segment-Anything (SAM) [26], prompting it with the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "projected 3D bounding box. Although SUN-RGBD provides segmentation masks, we find that the SAM annotations are of superior quality. For Objectron, we create accurate masks by efficiently annotating in 3D space. Specifically, we apply the pipeline of [4] to each of the Objectron videos. We fuse the resulting NeRF depth-maps to create a mesh. We post-process the 3D mesh, and get rid of the redundant vertices while keeping the vertices of the object. Finally, we create the masks for multiple frames by measuring the distance between each pixel on the depth map and the object mesh: pixels whose depth is far from the surface of the mesh are discarded, a process which also handles occlusions. NOCS computation: As proposed in [49], the NOCS coordinate \\( x_{nocs} \\) for a point on the object surface is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.32, + 0.786, + 0.35 + ], + "angle": 0, + "content": "\\[\nx _ {n o c s} = \\frac {1}{s} ^ {o b j} T _ {c a m} x _ {c a m} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.512 + ], + "angle": 0, + "content": "where \\( {}^{obj}T_{cam} = \\left[{}^{obj}R_{cam}\\right|^{obj}t_{cam} \\) is the 6DoF transformation from the camera to the object frame, \\( x_{cam} \\) the 3D location of the point in the camera frame, and \\( s \\) is a scalar, the size of the diagonal of the object's tight bounding box. It can be interpreted as the object shape scaled to a box with a unit diagonal. For each image, starting from a 3D pointcloud \\( (x_{cam}) \\) obtained from the depth, we collect points on each object \\( (x_{cam}^i) \\) using its 3D bounding box and instance mask, which can be transformed to the NOCS coordinate \\( (x_{nocs}^i) \\) using (1). We store NOCS as a 3-channel 2D map (i.e. \\( x_{nocs}^i \\) at its corresponding 2D location obtained by projecting \\( x_{cam}^i \\)). Our final result is the OmniNOCS dataset which contains instance segmentations, NOCS maps, and 3D bounding boxes for objects across 97 classes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.535, + 0.421, + 0.551 + ], + "angle": 0, + "content": "4 NOCSformer model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.643 + ], + "angle": 0, + "content": "We propose a novel architecture for monocular NOCS prediction termed \"NOCsformer\" trained on OmniNOCS. NOCsformer primarily uses self-attention layers, and is the largest trained monocular NOCS prediction model to date. NOCsformer also contains a 3D size head and a learned PnP head that are used to estimate a canonically oriented 3D object bounding box from the NOCS." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.394, + 0.679 + ], + "angle": 0, + "content": "4.1 Model architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.719 + ], + "angle": 0, + "content": "As shown in Figure 3, the NOCSformer architecture comprises an image backbone, a NOCS head, a size head, and a learned PnP head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "**Backbone:** Our backbone is a frozen Vision Transformer (ViT) [17], that uses DINOv2 weights [40] from self-supervised pretraining. This choice is motivated by recent findings on using frozen DINOv2 backbones for understanding depth, multi-view correspondences, or relative pose [20, 56]. Additionally, we use a Dense Prediction Transformer (DPT) [43] architecture to fuse low-resolution DINOv2 representations from multiple intermediate layers and upsample them by a factor of 8. This higher feature resolution is desirable to improve predictions on smaller objects. We train the DPT layers while keeping the ViT layers frozen. We use 2D input boxes to sample the DPT features using" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.38, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.789, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.79, + 0.334 + ], + "angle": 0, + "content": "Fig. 3: Architecture for NOCSformer: We use a transformer backbone to extract features from the input image, pool them using the 2D box RoIs, and feed the per-RoI features to the NOCS and size heads. Our novel NOCS head jointly predicts the NOCS and instance mask for the RoI. Our learned PnP head for pose estimation uses the predicted NOCS and instance mask to predict the projected 3D coordinate and 3D rotation of the object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.789, + 0.406 + ], + "angle": 0, + "content": "RoIAlign [22] into a 28x28 fixed resolution grid that is fed to the NOCS and size heads. While DPT has been previously leveraged for tasks that provide dense supervision, such as semantic segmentation and depth prediction, we use it in an instance-level prediction context, training it using supervision from object RoIs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.542 + ], + "angle": 0, + "content": "NOCS head: All existing NOCS prediction models [32, 48, 49] use MaskRCNN heads [22] with convolutional layers, in some cases with a separate head for each NOCS coordinate [49]. While this works on small datasets, we find that it significantly limits performance on OmniNOCS. NOCSformer therefore uses a simple (but large) transformer NOCS head with 10 self-attention blocks and one final upsampling conv layer to jointly predict both the instance mask and the NOCS map in the same output head. Similar to [49], for each NOCS coordinate, we predict logits over non-overlapping bins, and use the softmax to obtain the final NOCS value. Our ablations show that this choice works best on our challenging OmniNOCS dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.79, + 0.648 + ], + "angle": 0, + "content": "Size and learned PnP heads: We use a size head to predict the 3D size to scale the NOCS predictions to metric object coordinates. Our size head also uses self-attention blocks. Contrary to previous work [7,28,49], it does not use any class-specific layers or per-class average size statistics, as these limit scalability to larger and diverse vocabularies. Our learned PnP head is inspired from [32,48] and uses convolutional layers to predict the 3D orientation and a 2D projection of the 3D centroid. The orientation is predicted using the 6D partial rotation matrix representation [57] in allocentric space [7,28]." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.667, + 0.514, + 0.683 + ], + "angle": 0, + "content": "4.2 Localization from NOCS predictions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.78 + ], + "angle": 0, + "content": "To localize the object in 3D, we first scale the predicted NOCS using the predicted size to obtain unnormalized metric object coordinates. NOCSformer's learned PnP head predicts the object orientation in allocentric space, which is converted to the egocentric frame using the predicted projection of the 3D centroid. Using the unnormalized 3D object coordinates, object orientation, and the 3D centroid ray, the object depth is estimated from the corresponding 2D points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Alternatively, the 6DoF pose for the object can be solved directly (without using the learned PnP head) by solving a PnP problem using efficient solvers [30]. However, the orientations so obtained are known to be more sensitive to errors in NOCS coordinates, and less robust when compared to learned methods [32, 48]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.304, + 0.16 + ], + "angle": 0, + "content": "4.3 Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.249 + ], + "angle": 0, + "content": "When training NOCSformer, we supervise the NOCS, mask, 3D size, orientation and centroid predictions using appropriate losses. We supervise predicted masks using a simple L2 loss with respect to the ground truth instance mask. Since our NOCS head uses binned prediction where the final NOCS value is a softmax over non-overlapping bin logits, we use a combination of cross-entropy and regression losses." + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.263, + 0.659, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a s k}} = \\left\\| \\operatorname {m a s k} _ {\\text {p r e d}} - \\operatorname {m a s k} _ {\\text {g t}} \\right\\| _ {2} \\\\ \\mathcal {L} _ {N O C S} = \\operatorname {s o f t m a x} \\mathrm {C E} (\\hat {\\mathbf {n}}, \\hat {\\mathbf {n}} _ {g t}) + | | \\mathbf {n} - \\mathbf {n} _ {g t} | | _ {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.786, + 0.339 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{n}}\\) are the predicted logits over the discretized bins and \\(\\mathbf{n}\\) the continuous NOCS coordinate prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.787, + 0.43 + ], + "angle": 0, + "content": "In addition to the supervised loss above, we also use a variant of the self-supervised NOCS loss \\(\\mathcal{L}_{ss}\\) from [11] that minimizes the reprojection error of the predicted NOCS using the ground truth pose and predicted mask. More details about \\(\\mathcal{L}_{ss}\\) are in the supplementary material. The NOCS and mask losses are applied to the fixed resolution grid predictions. Since our NOCS ground truth can be sparse, it is applied at only those locations on the grid that have valid NOCS ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.787, + 0.505 + ], + "angle": 0, + "content": "Since our 3D size head also uses binned prediction, we use a combination of softmax cross-entropy and L1 loss for supervision. Note that our size loss is normalized by the ground truth size, in order to penalize errors on smaller objects equally. For the learned PnP head, we also supervise the centroid and rotation predictions using L1 losses in their output space." + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.522, + 0.652, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i z e}} = \\operatorname {s o f t m a x C E} (\\hat {\\mathbf {s}}, \\hat {\\mathbf {s}} _ {g t}) + | \\mathbf {s} - \\mathbf {s} _ {\\mathbf {g} t} | / \\mathbf {s} _ {\\mathbf {g} t} \\\\ \\mathcal {L} _ {r o t} = \\left\\| ^ {c} \\mathbf {R} _ {o _ {p r e d}} - ^ {c} \\mathbf {R} _ {o _ {g t}} \\right\\| _ {1, 1} \\\\ \\mathcal {L} _ {\\text {c e n t r o i d}} = \\left\\| \\mathbf {c} _ {o} - \\mathbf {c} _ {o g t} \\right\\| _ {2} \\\\ \\mathcal {L} _ {P n P} = \\mathcal {L} _ {r o t} + \\mathcal {L} _ {c e n t r o i d} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.605, + 0.629, + 0.621 + ], + "angle": 0, + "content": "The total loss for training NOCSformer is a weighted sum:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.633, + 0.772, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = w _ {\\text {s i z e}} \\mathcal {L} _ {\\text {s i z e}} + w _ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + w _ {\\text {N O C S}} \\mathcal {L} _ {\\text {N O C S}} + w _ {s s} \\mathcal {L} _ {s s} + w _ {P n P} \\mathcal {L} _ {P n P}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.357, + 0.688 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Although previous works predict NOCS accurately on a few categories [32, 48, 49], they only evaluate on 3D detection (localization) or pose estimation tasks, without quantifying the accuracy of the predicted NOCS. In section 5.1, we propose metrics and establish a benchmark to evaluate NOCS on the OmniNOCS dataset. We compare the localization accuracy of NOCSformer against existing benchmarks, by evaluating its localization accuracy on nuScenes [10] in section 5.2. We also evaluate the unique ability of NOCSformer to transfer to unseen datasets and domains. Finally, we provide ablations on critical design choices made in our model in section 5.4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.219, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.378, + 0.127 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.156, + 0.381, + 0.168 + ], + "angle": 0, + "content": "Input: Image + 2D box" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.174, + 0.407, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.269, + 0.406, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.363, + 0.406, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.439, + 0.406, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.53, + 0.406, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.625, + 0.283, + 0.635 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.636, + 0.311, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.626, + 0.373, + 0.635 + ], + "angle": 0, + "content": "Output" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.636, + 0.396, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.156, + 0.569, + 0.168 + ], + "angle": 0, + "content": "Output: NOCS + 3D box" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.174, + 0.602, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.269, + 0.602, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.363, + 0.602, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.439, + 0.602, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.53, + 0.602, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.625, + 0.5, + 0.636 + ], + "angle": 0, + "content": "3D shape + pose" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.637, + 0.517, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.625, + 0.586, + 0.635 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.636, + 0.608, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.156, + 0.781, + 0.169 + ], + "angle": 0, + "content": "Output: 3D shapes and poses" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.174, + 0.782, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.372, + 0.779, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.441, + 0.78, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.532, + 0.762, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.625, + 0.67, + 0.635 + ], + "angle": 0, + "content": "Output" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.636, + 0.684, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.625, + 0.782, + 0.635 + ], + "angle": 0, + "content": "3D shape + pose" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.641, + 0.779, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Fig. 4: Example results of our single unified NOCSformer model across various datasets and object classes. The left column shows input images and query 2D bounding boxes. The center column shows the NOCS+instance maps predicted by NOCSformer along with the estimated 3D pose overlaid on the input image. The NOCS can be used with the 3D boxes and object size to lift the objects to a 3D pointcloud, which is shown in the right column. Last row contains two examples." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Datasets: For our experiments, we train NOCSformer on the OmniNOCS dataset containing 97 classes, holding out images from the NOCS-Real275 dataset. We use NOCS-Real275 to evaluate NOCSformer's cross-dataset generalization capabilities by not training on it. NOCS-Real275 has 6 classes that overlap with the rest of OmniNOCS, although they are not the top 20 classes. It also differs from the rest of OmniNOCS in terms of camera parameters and the context of objects in the scene." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.26, + 0.445, + 0.275 + ], + "angle": 0, + "content": "5.1 NOCS prediction accuracy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.376 + ], + "angle": 0, + "content": "We propose the use of NOCS mAE and NOCS mPSNR to evaluate the accuracy of predicted NOCS. For each object the mean Absolute Error (mAE) and mPSNR are computed for all points in the intersection of ground truth and predicted masks, and reported as a mean of per-category means. However, the NOCS mAE/mPSNR do not penalize undersegmentations / sparse predictions. Since we also would like the NOCS predictions to span the full visible instance, we also evaluate the 2D mask mIoU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.468 + ], + "angle": 0, + "content": "We evaluate NOCSformer on a subset of OmniNOCS that contains 75 classes with accurate ground truth NOCS and masks, see Table 3. It is able to predict NOCS with errors less than \\(9\\%\\) (0.089 mAE on OmniNOCS). We also evaluate on the held-out NOCS-Real275 dataset, which was not used to train NOCSformer. We find that the zero-shot NOCSformer outperforms the NOCS baseline from [49] that was fully trained on this dataset, a strong indication of the generalization capabilities of NOCSformer." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.489, + 0.429, + 0.505 + ], + "angle": 0, + "content": "5.2 3D localization accuracy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.787, + 0.574 + ], + "angle": 0, + "content": "NOCS provides dense 3D-2D correspondences that can be used to estimate the 3D object oriented bounding box. This is done by using the predicted object size and solving a PnP problem, as explained in Sec. 4.2. We evaluate the accuracy of our estimated bounding boxes in different settings:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Outdoor scenes: We compare NOCSformer's 3D localization accuracy to that of other 3D detection models using the nuScenes true positive localization metrics on the challenging nuScenes dataset (included in OmniNOCS). We compare to Cube R-CNN, since it is the only other model that generalizes across diverse datasets. However, Cube R-CNN has two notable differences to NOCSformer: 1) being a detection model, it also jointly detects 2D object regions, and 2) it uses Chamfer loss causing it to have high orientation errors. As a more comparable baseline, we use a variant of our model (termed \"Cubeformer\"), by replacing our NOCS head with the cube head from Cube R-CNN" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.721, + 0.784, + 0.772 + ], + "angle": 0, + "content": "
MethodNOCS-Real275OmniNOCS
NOCS MAE↓NOCS PSNR↑Mask IoU↑NOCS MAE↓NOCS PSNR↑Mask IoU↑
NOCS baseline [49]0.12116.34586.10---
NOCSformer0.10718.52789.030.09420.24578.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.782, + 0.788, + 0.838 + ], + "angle": 0, + "content": "Table 3: NOCS and mask prediction evaluation: We report metrics for NOCSformer on 75 classes of OmniNOCS: it is the first model that is capable on predicting NOCS on such diverse data. Additionally, NOCSformer outperforms [49] which trained on NOCS-Real275, without training on NOCS-Real275." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.38, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.143, + 0.754, + 0.264 + ], + "angle": 0, + "content": "
MethodMulti-datasetmATE (m)↓mAOE (rad)↓mASE (%)↓mIoU↑
FCOS3D [51]X0.7770.4000.231-
PGD [50]X0.6750.3990.236-
EProPNP [12]X0.6760.3630.263-
EProPNP + TTA [12]X0.6530.3190.255-
Cube R-CNN [7]✓0.6501.3050.2830.349
Cubeformer✓0.7900.5730.3010.280
NOCSformer✓0.8870.5580.2910.377
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.272, + 0.789, + 0.335 + ], + "angle": 0, + "content": "Table 4: Outdoor localization: Comparison of 3D localization accuracy on the nuScenes subset of OmniNOCS. Top half shows methods that are only trained on nuScenes: while these perform better on nuScenes itself, they do not generalize to other datasets and classes. OmniNOCS is competitive with Cube R-CNN on mIoU while also predicting canonical orientations and object coordinates. Note that mIoU and orientation error only uses the predicted yaw orientation component." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.789, + 0.485 + ], + "angle": 0, + "content": "without a Chamfer loss. It uses the input 2D boxes. More details about Cubeformer are in the supplementary material. From Table 4, we find that NOCSformer's box orientations are canonical and more accurate than both Cube R-CNN and Cubeformer. Moreover, the boxes estimated from NOCSformer's NOCS predictions are comparable in translation and scale errors to those of Cube R-CNN and Cubeformer, even though it does not directly regress depth. While all 3 methods generalize across several classes and domains, baselines trained solely on nuScenes localization (top half) significantly outperform them on this task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.593 + ], + "angle": 0, + "content": "Cross-dataset generalization (indoor): Here, we hold out NOCS-Real275 from OmniNOCS when training NOCSformer, and use it to evaluate NOCSformer's ability to generalize to unseen domains. NOCS-Real275 features a tabletop setting with multiple objects, which is semantically different from our other indoor datasets. We compare against Cube R-CNN [7] and the NOCS [49] model that was trained from scratch on this dataset alone. [49] also uses class-specific model parameters and additional losses for symmetric objects that do not scale to larger vocabularies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.789, + 0.685 + ], + "angle": 0, + "content": "The results are shown in Table 5. We find that NOCSformer is more accurate at transferring to this unseen dataset compared to Cube R-CNN, indicating the generalizability of NOCS-based localization methods over that of box-regression methods. Note that the mAP metrics used in the NOCS-Real275 datasets may also be affected by the false positives/negatives from Cube R-CNN. However, both models are worse than the supervised NOCS baseline." + }, + { + "type": "table", + "bbox": [ + 0.316, + 0.713, + 0.684, + 0.778 + ], + "angle": 0, + "content": "
MethodTransferred?Depth input?mAP
3D IoU @ 253D IoU @ 50
NOCS model [49]X✓79.672.4
CubeRCNN [7]✓X14.94.1
NOCSformer✓X43.510.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.782, + 0.789, + 0.839 + ], + "angle": 0, + "content": "Table 5: Cross-dataset generalization: Comparison of localization accuracy on NOCS-Real275 for all classes, using mAP at different thresholds. This dataset has been held out when training CubeRCNN and NOCSformer. NOCSformer is able to generalize to NOCS-Real275 without any additional training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.788, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Fig. 5: Generalization across datasets: NOCSformer can generalize to new datasets that present new camera models and object domains. We show this zero-shot capability of NOCSformer (bottom row) on the NOCS-Real275 test set without training on NOCS-Real275 dataset. The predictions from a NOCS baseline [49] trained explicitly on the NOCS-Real275 dataset are shown in the top row." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.474, + 0.426, + 0.49 + ], + "angle": 0, + "content": "5.3 3D orientation accuracy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.657 + ], + "angle": 0, + "content": "A key challenge for 3D localization models trained on multiple datasets [7] is predicting canonical object orientations, as these datasets typically differ in their object pose conventions. OmniNOCS ensures that object poses in all constituent datasets are consistent within a category, enabling our model to predict canonical orientations. We evaluate this by computing the accuracy at different orientation error threshold on KITTI-val [13] in Table 6. We compare the accuracy of our orientation predictions to those of Cube R-CNN [7], which are trained using Chamfer loss, and therefore suffers at predicting canonical orientations. We also compare against our Cubeformer baseline that directly supervises the orientations. We find that orientations estimated using NOCSformer are more consistent and accurate than both our baselines." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.684, + 0.753, + 0.773 + ], + "angle": 0, + "content": "
MethodGravity axisHeading axis (X)
Accuracy @ 1 degAccuracy @ 5 degAccuracy @ 90 degAccuracy 5 degAccuracy @ 10 degAccuracy @ 90 deg
CubeRCNN5.7518.2951.7419.4321.5534.17
Cubeformer80.7998.52100.041.9955.8877.26
NOCSformer84.2499.95100.049.1659.6581.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.782, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Table 6: 3D orientation accuracy for models trained on multiple datasets: Models that use Chamfer distance (CubeRCNN) for supervising orientation heads end up being inconsistent in their orientation predictions. The results are averaged over 5 classes in the KITTI-val subset [13]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.379, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.144, + 0.668, + 0.214 + ], + "angle": 0, + "content": "
ArchitectureNOCS PSNR↑Mask IoU↑
NOCSformer20.43181.69
Hourglass head-3.064-3.16
MaskRCNN head-5.23-21.27
w/o discretized prediction-4.11-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Table 7: Comparison of different architecture choices for the NOCS and mask prediction head: We experiment with transformer, hourglass and MaskRCNN architectures. We compare whether discretized prediction of NOCS is better than continuous regression for transformer heads. We quantify the difference in performance when using separate heads for mask and NOCS prediction as opposed to a single head." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.348, + 0.325, + 0.362 + ], + "angle": 0, + "content": "5.4 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.483 + ], + "angle": 0, + "content": "NOCS head architecture: While all existing architectures use a few CNN layers for their NOCS heads [49], we find that this significantly limits the NOCS prediction performance when scaling to more classes. Using our transformer NOCS head provides a 5.23dB improvement in NOCS PSNR and a \\(21.27\\%\\) improvement in mask mIoU. We also experiment with a larger convolutional head: the Hourglass model from [6]. We find that the NOCSformer NOCS heads are even better than Hourglass, with an improvement of 3.06dB on NOCS PSNR and \\(3.16\\%\\) on mask IoU (as shown in Table 7)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Discretized versus continuous prediction of NOCS: NOCSformer models the final NOCS prediction as a classification layer with 50 non-overlapping bins. This was observed to be better than a linear regression layer with a MaskRCNN head in [49]. We find that it is also significantly better when using a transformer head, improving the NOCS prediction PSNR by 4.11 dB, as in Table 7." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.492, + 0.607 + ], + "angle": 0, + "content": "6 Conclusion and Future work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.764 + ], + "angle": 0, + "content": "This paper has introduced a new large scale dataset of Normalized Object Coordinates (NOCS) for a wide variety of object classes in indoor and outdoor scenes. It has also proposed a single transformer-based model NOCSformer that can predict object 3D pose, size, and NOCS for any of these objects given 2D bounding box inputs. These allows our model to obtain 3D shape and oriented bounding boxes of objects in metric scale from a single input image. This represents the first attempt to generalize NOCS estimation beyond small datasets of narrow domain, increasing the number of object categories available by an order of magnitude. We hope this provides a means for others to explore large-scale monocular 3D object pose and shape estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Some limitations of our current method include the handling of symmetric objects, where the coordinate system has multiple possible solutions, and reflected geometry, such as left / right shoes. Future work could address these issues, for example, minimizing over multiple coordinate frames in the loss for symmetric objects, and potential estimation of left / right coordinate frames for reflected geometry." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.786, + 0.221 + ], + "angle": 0, + "content": "1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.786, + 0.249 + ], + "angle": 0, + "content": "2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.786, + 0.278 + ], + "angle": 0, + "content": "3. Avetisyan, A., Dai, A., Nießner, M.: End-to-end cad model retrieval and 9dof alignment in 3d scans. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.28, + 0.787, + 0.306 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.787, + 0.376 + ], + "angle": 0, + "content": "5. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1) (2021), https://openreview.net/forum?id=tjZjv_qh_CE" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.786, + 0.418 + ], + "angle": 0, + "content": "6. Birodkar, V., Lu, Z., Li, S., Rathod, V., Huang, J.: The surprising impact of mask-head architecture on novel class segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7015-7025 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.786, + 0.46 + ], + "angle": 0, + "content": "7. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.786, + 0.502 + ], + "angle": 0, + "content": "8. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.786, + 0.53 + ], + "angle": 0, + "content": "9. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2. arXiv preprint arXiv:2001.10773 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.532, + 0.787, + 0.587 + ], + "angle": 0, + "content": "10. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.589, + 0.787, + 0.629 + ], + "angle": 0, + "content": "11. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.631, + 0.787, + 0.685 + ], + "angle": 0, + "content": "12. Chen, H., Wang, P., Wang, F., Tian, W., Xiong, L., Li, H.: Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2781-2790 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.687, + 0.787, + 0.728 + ], + "angle": 0, + "content": "13. Chen, X., Kundu, K., Zhu, Y., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals using stereo imagery for accurate object class detection. IEEE transactions on pattern analysis and machine intelligence 40(5), 1259-1272 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.729, + 0.787, + 0.769 + ], + "angle": 0, + "content": "14. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "15. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.813, + 0.787, + 0.839 + ], + "angle": 0, + "content": "16. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.18, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.788, + 0.189 + ], + "angle": 0, + "content": "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "18. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.257 + ], + "angle": 0, + "content": "19. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.284 + ], + "angle": 0, + "content": "20. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: European Conference on Computer Vision. pp. 516-532. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.31 + ], + "angle": 0, + "content": "21. Gumeli, C., Dai, A., Nießner, M.: Roca: Robust cad model retrieval and alignment from a single image. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.337 + ], + "angle": 0, + "content": "22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.337, + 0.787, + 0.392 + ], + "angle": 0, + "content": "23. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 923-934 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.445 + ], + "angle": 0, + "content": "24. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: IEEE International Conference on Robotics and Automation (ICRA) (2022), https://arxiv.org/abs/2203.01929" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.488 + ], + "angle": 0, + "content": "25. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape appearance and pose optimization. In: European Conference on Computer Vision (ECCV) (2022), https://arxiv.org/abs/2207.13691" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.554 + ], + "angle": 0, + "content": "27. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.787, + 0.595 + ], + "angle": 0, + "content": "28. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.622 + ], + "angle": 0, + "content": "29. Kuo, W., Angelova, A., Lin, T.Y., Dai, A.: Mask2cad: 3d shape prediction by learning to segment and retrieve. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.787, + 0.676 + ], + "angle": 0, + "content": "30. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision 81, 155-166 (2009). https://doi.org/ https://doi.org/10.1007/s11263-008-0152-6, http://infoscience. epfl.ch/record/160138" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.717 + ], + "angle": 0, + "content": "31. Li, K., DeTone, D., Chen, Y.F.S., Vo, M., Reid, I., Rezatofighi, H., Sweeney, C., Straub, J., Newcombe, R.: Odam: Object detection, association, and mapping using posed rgb video. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "32. Li, Z., Wang, G., Ji, X.: Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7678-7687 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "33. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15641-15650 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Manhardt, F., Kehl, W., Gaidon, A.: Roi-10d: Monocular lifting of 2d detection to 6d pose and metric shape. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2069-2078 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.656, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniNOCS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "35. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Vid2cad: Cad model alignment using multi-view constraints from videos. TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.79, + 0.203 + ], + "angle": 0, + "content": "36. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Cad-estate: Large-scale cad model annotation in rgb videos. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.79, + 0.231 + ], + "angle": 0, + "content": "37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "38. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "39. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.355 + ], + "angle": 0, + "content": "40. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "41. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: Proceedings of the IEEE international conference on computer vision. pp. 3828-3836 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.424 + ], + "angle": 0, + "content": "42. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3D appearance, location & pose. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.465 + ], + "angle": 0, + "content": "43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.508 + ], + "angle": 0, + "content": "44. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: International Conference on Computer Vision (ICCV) 2021 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.549 + ], + "angle": 0, + "content": "45. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 567-576 (2015). https://doi.org/10.1109/CVPR.2015.7298655" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.619 + ], + "angle": 0, + "content": "46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.646 + ], + "angle": 0, + "content": "47. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.688 + ], + "angle": 0, + "content": "48. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16611-16621 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.729 + ], + "angle": 0, + "content": "49. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.729, + 0.788, + 0.757 + ], + "angle": 0, + "content": "50. Wang, T., Xinge, Z., Pang, J., Lin, D.: Probabilistic and geometric depth: Detecting objects in perspective. In: Conference on Robot Learning. pp. 1475-1485. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.798 + ], + "angle": 0, + "content": "51. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.798, + 0.788, + 0.84 + ], + "angle": 0, + "content": "52. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.79, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.378, + 0.128 + ], + "angle": 0, + "content": "A. Krishnan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "53. Wen, B., Lian, W., Bekris, K., Schaal, S.: Catgrasp: Learning category-level task-relevant grasping in clutter from simulation. ICRA 2022 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "54. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "55. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21919-21928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.285 + ], + "angle": 0, + "content": "56. Zhang, J., Herrmann, C., Hur, J., Cabrera, L.P., Jampani, V., Sun, D., Yang, M.H.: A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence arXiv preprint arxiv:2305.15347 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "57. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_origin.pdf b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9b40d03b26861532367f6e240b87b6727017ef1e --- /dev/null +++ b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/98a3347b-c5d5-499d-bd79-bc9ee828386d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24747185eef53a0ef66fbe128b908f1d3d92d3e867fd2095649d7e479bd83e7c +size 2001914 diff --git a/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/full.md b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9279c3f8fe1e7b67ee381a0f4e120952280de400 --- /dev/null +++ b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/full.md @@ -0,0 +1,328 @@ +# OmniNOCS: A unified NOCS dataset and model for 3D lifting of 2D objects + +Akshay Krishnan $^{1,2}$ , Abhijit Kundu $^{1}$ , Kevis-Kokitsi Maninis $^{1}$ , James Hays $^{2}$ , and Matthew Brown $^{1}$ + +1 Google Research† + +2 Georgia Institute of Technology + +![](images/c6cdaaff917a5eba3946dee9b9ca5ebdc840d3a5a95ab822c2bb341eca8803cc.jpg) +Fig. 1: We introduce OmniNOCS, a large-scale dataset with Normalized Object Coordinates (NOCS), instance masks, and 3D box annotations for objects across several classes, domains, and cameras. We also propose NOCSformer, a model trained on OmniNOCS that lifts each 2D object bounding box in an image to its corresponding 3D oriented box (pose) and 3D pointcloud (shape). + +![](images/6fd026d4e6665924bd49b1a033471c917353d1a4d0c01170f5b7c1310616d6f7.jpg) + +![](images/6a0d92c0efe8f14ba295dcc2483152ac8b33510ea8aef85261200f64e80b54f2.jpg) + +Abstract. We propose OmniNOCS, a large-scale monocular dataset with 3D Normalized Object Coordinate Space (NOCS) maps, object masks, and 3D bounding box annotations for indoor and outdoor scenes. OmniNOCS has 20 times more object classes and 200 times more instances than existing NOCS datasets (NOCS-Real275, Wild6D). We use OmniNOCS to train a novel, transformer-based monocular NOCS prediction model (NOCsformer) that can predict accurate NOCS, instance masks and poses from 2D object detections across diverse classes. It is the first NOCS model that can generalize to a broad range of classes when prompted with 2D boxes. We evaluate our model on the task of 3D oriented bounding box prediction, where it achieves comparable results to state-of-the-art 3D detection methods such as Cube R-CNN. Unlike other 3D detection methods, our model also provides detailed and accurate 3D object shape and segmentation. We propose a novel benchmark for the task of NOCS prediction based on OmniNOCS, which we hope will serve as a useful baseline for future work in this area. Our dataset and code will be at the project website: https://omninocs.github.io. + +# 1 Introduction + +Predicting the 6 Degree-of-Freedom (6DoF) pose and shape of objects from images is a crucial problem in 3D scene understanding. Robots need to understand the location, + +shape, and orientation of various objects to grasp and interact with them. Self-driving vehicles need to understand the location and heading of vehicles, pedestrians, and other objects on the road. In particular, the 3D orientation of these objects is crucial to predict their behavior. It is also important in AR/VR applications, as it allows users to interact with objects in meaningful ways. Predicting object shape and pose from monocular images is also a prerequisite to initialize 3D panoptic neural scene graphs [27, 39] and to track objects in videos [42]. These applications require approaches that generalize across a wide range of object classes, environments, and camera models. + +The problem of localizing 3D objects has been extensively studied through the lens of monocular 3D object detection or 6DoF pose estimation. The most commonly used approach is to represent objects as 3D cuboids, and train a model to regress the cuboid parameters from a 2D RoI of the object [7, 8, 28, 41, 51, 54]. An alternative method involves predicting corresponding 3D points in an object coordinate space followed by pose estimation using 3D-2D alignment [11, 12, 20, 33, 49]. However, all these methods are limited to narrow datasets collected on a single camera and context with typically very few object classes. Existing models are also trained separately for every dataset, preventing them from being used more widely. + +Recent work [7] takes a step towards scalability by introducing the Omni3D dataset, which aggregates 3D detection datasets from different domains, and trains a Cube R-CNN model to regress 3D bounding boxes for 50 classes. However, Omni3D notably lacks consistently oriented, object-centric boxes, as the ground truth orientations are not canonical for object classes. Cube R-CNN [7] instead uses a Chamfer distance loss which is invariant to the predicted 3D cuboid orientation. This causes it to predict inconsistent object orientations, for example, often flipping the orientations of cars by $180^{\circ}$ . Further, Omni3D is a detection dataset that does not provide object shape information. + +Our work aims to overcome the above shortcomings, by providing a new large-scale dataset with consistent object-centric ground truth boxes along with detailed shape. We argue for the use of Normalized Object Coordinate Space (NOCS) as proposed in [49] as a 3D object shape representation. NOCS represents both the canonical orientation and the shape of the visible surface of the object, properties that are essential for real-world applications such as self-driving and robotics. It can also be used to estimate the 3D bounding box of the object. However, all existing work on NOCS [24, 25, 32, 48, 49] only train small models on small datasets with fewer than 10 classes. We address the lack of diverse NOCS datasets by proposing a large-scale monocular NOCS dataset, OmniNOCS that has NOCS annotations, instance segmentation, and canonically oriented bounding boxes for 97 object classes, containing $380\mathrm{k}$ images from 10 different data sources, making it the largest and most diverse NOCS dataset currently available. OmniNOCS includes all of the data from Omni3D (KITTI [19], nuScenes [10], ARKitScenes [5], Objectron [1], Hypersim [44], SUN-RGBD [45]), with the addition of Cityscapes [15], virtual KITTI [9], NOCS-Real275 [49], and the Waymo Open Dataset [46]. + +Our work also introduces a new model, which we term "NOCSformer", that predicts NOCS coordinates and oriented 3D bounding boxes from monocular images and 2D detections for all the classes in OmniNOCS. NOCSformer leverages self-supervised pretrained ViTs [17, 40] and does not use any class-specific model parameters. This enables it scale to large vocabularies, and share information across semantically similar + +object classes. In contrast, existing NOCS models use small NOCS heads with class-specific parameters that significantly limit their performance on large datasets like OmniNOCS. NOCSformer also predicts the 3D size of the object along with its 3D orientation using a learned PnP head. This allows NOCSformer to predict 3D oriented bounding boxes and object point cloud in metric scale for diverse object categories. Our experiments evaluate the quality of NOCS and bounding boxes predicted by NOCSformer in comparison to existing NOCS prediction or 3D detection models. We find that training on OmniNOCS allows NOCSformer to generalize to unseen datasets, even outperforming baselines trained on the target dataset in NOCS prediction accuracy. + +In summary, our contributions are: + +- The OmniNOCS dataset: A new dataset containing Normalised Object Coordinates for 380k images in 97 categories, an order of magnitude larger on both counts than existing NOCS datasets. +- NOCSformer: A novel transformer-based architecture for predicting object NOCS, mask, and size from input 2D boxes, utilizing pre-trained self-supervised ViT backbones. NOCSformer is the first NOCS model to generalize to vocabularies with $90+$ classes and to unseen datasets, including images from internet collections. +- OmniNOCS benchmark: an evaluation framework with metrics for directly comparing different NOCS prediction algorithms on OmniNOCS, with baselines established via NOCSFormer. + +# 2 Related work + +The task of predicting 3D object pose has been studied both in the context of camera/object pose estimation (6DoF) and 3D bounding box estimation (6DoF pose + 3DoF size). A further distinction can be made between methods that regress directly to object pose / bounding box parameters, and methods that compute per pixel coordinates or depth as an initial step. We review each of these paradigms in the sections below. + +# 2.1 3D localization by regressing bounding boxes + +Direct approaches to 3D object detection and pose estimation involve networks that output rotation, translation and scale parameters directly. Several works have explored different parameterizations in this setting, e.g., PoseCNN [54] uses regression of translation via centre direction + distance maps (which enable detection even under occlusion), and quaternion representation of rotation. BB8 [41] uses segmentation followed by a CNN to regress to 2D projections of the $8 \times 3\mathrm{D}$ bounding box corners. Multi-view monocular approaches have also been proposed, e.g., DETR3D [52] which uses DETR-style attention to reason about object interactions. Several techniques also make use of existing 2D bounding box predictors, either as an input to a 3D lifting approach [34], or as a constraint on 3D box predictions [38]. + +Another group of works focuses on predicting the alignment of 3D CAD models within various modalities: images [21, 29], videos [31, 35], or 3D scans [2, 3]. These approaches typically determine the object's 3D translation, rotation, and size (9DoF) and find a CAD model with a similar visual appearance. + +# 2.2 3D localization from model-to-image alignment + +An alternative approach to 3D localization is to first predict correspondence, either between views or to a normalized space, and then reason over redundant correspondences to establish pose. This has been done with both sparse and dense correspondences, e.g., AutoShape [33] makes use of sparse 2D to 3D correspondences, with a learned shape model and sparse 2D keypoints. [20] uses pairwise semantic correspondence from ViT, to find pose between a reference image and a sequence of targets. + +Several approaches use dense correspondences and the idea of Normalised Object Coordinate Spaces, or NOCS. In the original NOCS work, Wang et al. [49] use RGB-D data and 3D-3D correspondences for pose estimation. Follow-up work [11] estimates pose using PnP variants on 2D-3D correspondences. While NOCS [49] used the average spatial dimensions of the object category to define the object frame, [53] showed that an instance-specific NOCS coordinate frame can be used alongside a predicted instance size for the same purpose. Other methods combine coordinate regression and direct approaches, e.g., [48], which uses direct regression based on dense correspondences. + +Similar to our work, MonoRUun [11] lifts 2D detections to 3D object coordinates without explicit class supervision. They also use a novel self-supervised coordinate regression training loss which obviates the need for detailed 3D ground truth. However, their representation does not include masks and therefore does not provide explicit object shape information. Their evaluation is also limited to 3 classes (car, pedestrian, cyclist) on the KITTI-Object test set. M3D-RPN [8] also has a single 3D head for multiple classes, jointly generating 2D and 3D bounding box predictions, though without shape information, and similarly limited to car, pedestrian and cyclist classes on KITTI. + +# 2.3 Monocular object localization / pose estimation datasets + +For 3D object detection from monocular RGB images, most existing works use a small number of classes on a single dataset. Cube R-CNN/Omni3D [7] contribute towards creating a general purpose monocular 3D object detector. Cube R-CNN performs well over six 3D datasets: KITTI, SUN RGB-D, ARKitScenes, Objectron, nuScenes and Hypersim. However, the method is class specific, with Cube R-CNN trained only on 50 classes, and unable to work in the open class setting. + +Cube R-CNN also uses a Chamfer loss on the predicted 3D box corners to deal with the inconsistent coordinate frame annotations in the underlying Omni3D datasets. The lack of direct orientation supervision results in inconsistent orientation predictions (for example, the positive x axis may point to the front or the back of the car for different instances). This orientation inconsistency is problematic for methods that seek to build detailed 3D models of object categories as a downstream task [27, 39]. + +Other works are not limited to 3D boxes, and directly provide object shapes. These shapes come in the form of annotated point clouds (e.g., ScanNet [16]), or aligned CAD models (e.g., Scan2CAD [2], Pix3D [47], CAD-Estate [36]). Object point clouds are created by labeling points on scanned 3D scenes, a process that doesn't scale well to large datasets. Semi-automatically aligning CAD models is scale-able to a certain extent, but since it relies on retrieving existing models, the resulting shapes are rarely accurate, and the alignments are sensitive to deformations and movable parts. In contrast to these works, we propose a large-scale dataset of many different categories. + +
DatasetNOCS GTReal#Images#Classes#Instances
CAMERA25 [49]✓×300k6184
NOCS-Real275 [49]✓✓7k624
Wild6D [18]×✓1M51.8K
OmniNOCS✓✓380k97>450k
+ +Table 1: Comparison to other NOCS datasets: Other existing NOCS datasets are limited in number of classes and instances. Note that Wild6D has 1M images, but these are used for self-supervised training, as it does not include NOCS ground truth (GT). + +![](images/438dbb5de78e5233772616e4da6306952f437af2f155c8c65d4cfe25df4e33e0.jpg) +Fig. 2: Preparation of the OmniNOCS dataset: First, we ensure that object orientations are canonical across different all subsets of OmniNOCS. Next, we re-compute depth for datasets where depth is not available or is too noisy. Finally, we annotate objects with (pseudo) instance mask labels where ground truth masks are not available. + +# 3 OmniNOCS dataset + +# 3.1 Data statistics + +We create OmniNOCS, a first-of-its-kind large and diverse NOCS dataset comprising data from several classes and domains. OmniNOCS uses data from self-driving scenes [9, 15, 19, 46], indoor scenes [5, 44, 45, 49], and object-centric videos [1]. Each of these sources use cameras with different parameters, ranging from phone cameras to wide FOV cameras mounted on self-driving cars. The number of object instances per image also varies widely, from single-object images [1] to images with hundreds of objects [44, 46]. + +OmniNOCS has 97 object categories across 380k images. This by far exceeds the diversity of existing NOCS datasets (with less than 6 categories as shown in Table 1). The number of instances exceeds that of other NOCS datasets by more than 2 orders of magnitude. It also contains more images with NOCS annotations (380k). Note that while Wild6D has 1M images, these do not have ground truth NOCS or pose annotations, and includes every frame of object-centric videos. OmniNOCS is also more diverse than the most diverse 3D detection dataset (Omni3D [7] with 98 categories). 307k images in OmniNOCS are from real world scenes and 73k are synthetic. 107k of the images in OmniNOCS are from outdoor scenes, 102k images are from complex indoor environments, and 140k are from object-centric videos. + +# 3.2 Data preparation + +Computing the NOCS for objects in an image requires canonically oriented 3D object bounding boxes, depth, and instance segmentations. Since the datasets we use were originally only proposed for monocular 3D detection, many of them lack dense depth or instance segmentation annotations (highlighted entries in Table 2). Additionally, NOCS + +
3D DatasetDepth sourceInstance mask source#Images#Classes
KITTI [19]LiDARHuman-labelled [23]7.4k7
SUN-RGBD [45]Depth CameraSegmentation model [26]10k72
Objectron [1]NeRFNeRF + Human label132k9
nuScenes [10]LiDARSegmentation Model [14]30k9
Hypersim [44]SyntheticSynthetic64k31
ARKitScenes [5]Depth CameraSegmentation model [26]60k15
Cityscapes 3D [15]StereoHuman-labelled3.4k8
Virtual KITTI [9]SyntheticSynthetic4.1k3
NOCS-Real275 [49]Depth cameraHuman-labelled7k6
Waymo OD [46]LiDARHuman-labelled62k7
OmniNOCS380k97
+ +Table 2: Constituent 3D detection datasets used in OmniNOCS: We augment many of these datasets with depth and masks, as they are missing in the original data (highlighted entries). In addition, we canonicalize the orientations of bounding boxes across all datasets. #Classes lists the number of classes we use from each dataset. Those above the dashed line are part of Omni3D [7]. + +also requires all instances within a category to have consistent orientation annotations. While this is certainly not true across different 3D detection datasets, it is also not true within some large synthetic datasets such as Hypersim. OmniNOCS aggregates several datasets, with the addition of 1) new depth estimated via Mip-NeRF reconstructions, 2) additional instance masks via segmentation models and human labels, 3) manual labelling of coordinate axes to give consistent object-centric coordinate frames. This enables NOCS to be computed for every image in the dataset. This multi-stage process for OmniNOCS creation is illustrated in Figure 2 and explained below. A summary of the resulting OmniNOCS dataset is provided in Table 2. + +Orientation canonicalization: Although the datasets we use contain oriented 3D bounding boxes, they vary in their level of canonicalization. Some datasets have their class-canonical orientations i.e., all instances of a particular class are oriented consistently within the dataset (for example, all cars in [19] have X axis pointing forward, and Z upwards). In such cases, we ensure that this canonicalization is consistent with all of OmniNOCS by applying a fixed class-specific offset orientation for the dataset. Datasets like Hypersim [44] have no class canonicalization at all, ie, although objects have tight bounding boxes, their XYZ axes directions are different for each instance. We manually label each object in such datasets to select the canonical orientation out of six possible orientations for the bounding box. More details on the labelling process are provided in the supplementary material. Some classes may have more than one choice for canonical orientations (due to symmetry), in which case an orientation is selected arbitrarily. + +Depth augmentation: For outdoor datasets, we use sparse depth from LiDAR if available. We recomputed the depth on Cityscapes using a state-of-the-art stereo depth model [55], as we found the depth from the original dataset to be noisy. Since Objectron [1] does not contain dense depth, we train Neural Radiance Fields [4, 37] for each video sequence in the dataset to obtain dense depth. + +Instance segmentation: Since many datasets we use were intended for 3D detection, they do not camera instance segmentations. We annotate ARKitScenes and SUNRGBD objects using instance masks from Segment-Anything (SAM) [26], prompting it with the + +projected 3D bounding box. Although SUN-RGBD provides segmentation masks, we find that the SAM annotations are of superior quality. For Objectron, we create accurate masks by efficiently annotating in 3D space. Specifically, we apply the pipeline of [4] to each of the Objectron videos. We fuse the resulting NeRF depth-maps to create a mesh. We post-process the 3D mesh, and get rid of the redundant vertices while keeping the vertices of the object. Finally, we create the masks for multiple frames by measuring the distance between each pixel on the depth map and the object mesh: pixels whose depth is far from the surface of the mesh are discarded, a process which also handles occlusions. NOCS computation: As proposed in [49], the NOCS coordinate $x_{nocs}$ for a point on the object surface is defined as: + +$$ +x _ {n o c s} = \frac {1}{s} ^ {o b j} T _ {c a m} x _ {c a m} \tag {1} +$$ + +where ${}^{obj}T_{cam} = \left[{}^{obj}R_{cam}\right|^{obj}t_{cam}$ is the 6DoF transformation from the camera to the object frame, $x_{cam}$ the 3D location of the point in the camera frame, and $s$ is a scalar, the size of the diagonal of the object's tight bounding box. It can be interpreted as the object shape scaled to a box with a unit diagonal. For each image, starting from a 3D pointcloud $(x_{cam})$ obtained from the depth, we collect points on each object $(x_{cam}^i)$ using its 3D bounding box and instance mask, which can be transformed to the NOCS coordinate $(x_{nocs}^i)$ using (1). We store NOCS as a 3-channel 2D map (i.e. $x_{nocs}^i$ at its corresponding 2D location obtained by projecting $x_{cam}^i$ ). Our final result is the OmniNOCS dataset which contains instance segmentations, NOCS maps, and 3D bounding boxes for objects across 97 classes. + +# 4 NOCSformer model + +We propose a novel architecture for monocular NOCS prediction termed "NOCsformer" trained on OmniNOCS. NOCsformer primarily uses self-attention layers, and is the largest trained monocular NOCS prediction model to date. NOCsformer also contains a 3D size head and a learned PnP head that are used to estimate a canonically oriented 3D object bounding box from the NOCS. + +# 4.1 Model architecture + +As shown in Figure 3, the NOCSformer architecture comprises an image backbone, a NOCS head, a size head, and a learned PnP head. + +**Backbone:** Our backbone is a frozen Vision Transformer (ViT) [17], that uses DINOv2 weights [40] from self-supervised pretraining. This choice is motivated by recent findings on using frozen DINOv2 backbones for understanding depth, multi-view correspondences, or relative pose [20, 56]. Additionally, we use a Dense Prediction Transformer (DPT) [43] architecture to fuse low-resolution DINOv2 representations from multiple intermediate layers and upsample them by a factor of 8. This higher feature resolution is desirable to improve predictions on smaller objects. We train the DPT layers while keeping the ViT layers frozen. We use 2D input boxes to sample the DPT features using + +![](images/54ac7709470a5519de97dce65bee88326729adc990e8400865551bc77a2d6051.jpg) +Fig. 3: Architecture for NOCSformer: We use a transformer backbone to extract features from the input image, pool them using the 2D box RoIs, and feed the per-RoI features to the NOCS and size heads. Our novel NOCS head jointly predicts the NOCS and instance mask for the RoI. Our learned PnP head for pose estimation uses the predicted NOCS and instance mask to predict the projected 3D coordinate and 3D rotation of the object. + +RoIAlign [22] into a 28x28 fixed resolution grid that is fed to the NOCS and size heads. While DPT has been previously leveraged for tasks that provide dense supervision, such as semantic segmentation and depth prediction, we use it in an instance-level prediction context, training it using supervision from object RoIs. + +NOCS head: All existing NOCS prediction models [32, 48, 49] use MaskRCNN heads [22] with convolutional layers, in some cases with a separate head for each NOCS coordinate [49]. While this works on small datasets, we find that it significantly limits performance on OmniNOCS. NOCSformer therefore uses a simple (but large) transformer NOCS head with 10 self-attention blocks and one final upsampling conv layer to jointly predict both the instance mask and the NOCS map in the same output head. Similar to [49], for each NOCS coordinate, we predict logits over non-overlapping bins, and use the softmax to obtain the final NOCS value. Our ablations show that this choice works best on our challenging OmniNOCS dataset. + +Size and learned PnP heads: We use a size head to predict the 3D size to scale the NOCS predictions to metric object coordinates. Our size head also uses self-attention blocks. Contrary to previous work [7,28,49], it does not use any class-specific layers or per-class average size statistics, as these limit scalability to larger and diverse vocabularies. Our learned PnP head is inspired from [32,48] and uses convolutional layers to predict the 3D orientation and a 2D projection of the 3D centroid. The orientation is predicted using the 6D partial rotation matrix representation [57] in allocentric space [7,28]. + +# 4.2 Localization from NOCS predictions + +To localize the object in 3D, we first scale the predicted NOCS using the predicted size to obtain unnormalized metric object coordinates. NOCSformer's learned PnP head predicts the object orientation in allocentric space, which is converted to the egocentric frame using the predicted projection of the 3D centroid. Using the unnormalized 3D object coordinates, object orientation, and the 3D centroid ray, the object depth is estimated from the corresponding 2D points. + +Alternatively, the 6DoF pose for the object can be solved directly (without using the learned PnP head) by solving a PnP problem using efficient solvers [30]. However, the orientations so obtained are known to be more sensitive to errors in NOCS coordinates, and less robust when compared to learned methods [32, 48]. + +# 4.3 Losses + +When training NOCSformer, we supervise the NOCS, mask, 3D size, orientation and centroid predictions using appropriate losses. We supervise predicted masks using a simple L2 loss with respect to the ground truth instance mask. Since our NOCS head uses binned prediction where the final NOCS value is a softmax over non-overlapping bin logits, we use a combination of cross-entropy and regression losses. + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m a s k}} = \left\| \operatorname {m a s k} _ {\text {p r e d}} - \operatorname {m a s k} _ {\text {g t}} \right\| _ {2} \\ \mathcal {L} _ {N O C S} = \operatorname {s o f t m a x} \mathrm {C E} (\hat {\mathbf {n}}, \hat {\mathbf {n}} _ {g t}) + | | \mathbf {n} - \mathbf {n} _ {g t} | | _ {1} \\ \end{array} +$$ + +where $\hat{\mathbf{n}}$ are the predicted logits over the discretized bins and $\mathbf{n}$ the continuous NOCS coordinate prediction. + +In addition to the supervised loss above, we also use a variant of the self-supervised NOCS loss $\mathcal{L}_{ss}$ from [11] that minimizes the reprojection error of the predicted NOCS using the ground truth pose and predicted mask. More details about $\mathcal{L}_{ss}$ are in the supplementary material. The NOCS and mask losses are applied to the fixed resolution grid predictions. Since our NOCS ground truth can be sparse, it is applied at only those locations on the grid that have valid NOCS ground truth. + +Since our 3D size head also uses binned prediction, we use a combination of softmax cross-entropy and L1 loss for supervision. Note that our size loss is normalized by the ground truth size, in order to penalize errors on smaller objects equally. For the learned PnP head, we also supervise the centroid and rotation predictions using L1 losses in their output space. + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s i z e}} = \operatorname {s o f t m a x C E} (\hat {\mathbf {s}}, \hat {\mathbf {s}} _ {g t}) + | \mathbf {s} - \mathbf {s} _ {\mathbf {g} t} | / \mathbf {s} _ {\mathbf {g} t} \\ \mathcal {L} _ {r o t} = \left\| ^ {c} \mathbf {R} _ {o _ {p r e d}} - ^ {c} \mathbf {R} _ {o _ {g t}} \right\| _ {1, 1} \\ \mathcal {L} _ {\text {c e n t r o i d}} = \left\| \mathbf {c} _ {o} - \mathbf {c} _ {o g t} \right\| _ {2} \\ \mathcal {L} _ {P n P} = \mathcal {L} _ {r o t} + \mathcal {L} _ {c e n t r o i d} \\ \end{array} +$$ + +The total loss for training NOCSformer is a weighted sum: + +$$ +\mathcal {L} _ {\text {t o t a l}} = w _ {\text {s i z e}} \mathcal {L} _ {\text {s i z e}} + w _ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + w _ {\text {N O C S}} \mathcal {L} _ {\text {N O C S}} + w _ {s s} \mathcal {L} _ {s s} + w _ {P n P} \mathcal {L} _ {P n P} +$$ + +# 5 Experiments + +Although previous works predict NOCS accurately on a few categories [32, 48, 49], they only evaluate on 3D detection (localization) or pose estimation tasks, without quantifying the accuracy of the predicted NOCS. In section 5.1, we propose metrics and establish a benchmark to evaluate NOCS on the OmniNOCS dataset. We compare the localization accuracy of NOCSformer against existing benchmarks, by evaluating its localization accuracy on nuScenes [10] in section 5.2. We also evaluate the unique ability of NOCSformer to transfer to unseen datasets and domains. Finally, we provide ablations on critical design choices made in our model in section 5.4. + +![](images/2aa0ca2f97ae637fa30545efd1f1ae8588bb1f4cb61de62055b8d308fd639897.jpg) +Input: Image + 2D box + +![](images/7b301e035cb9a8d792ee47cfe0d9599f6706dbcc92612ef7f286c08862e669ac.jpg) + +![](images/374cb9ec2555543a715f6dc158b5c9f71b8bf0a1361435068508944cbd3e1f21.jpg) + +![](images/c90f8e2d618abcd867d7ddadf1c909e5f57fd75cc105dc447ac07813344f5228.jpg) + +![](images/96c2e128ca476ddc8a22a180ed59a9c443c6f8d4ea3609bb8c624b657e935d26.jpg) + +![](images/a75054222dc8a2216e3c6c7b26086911937082e67b44a8ebffcdc7467c9d1cd2.jpg) +Input + +![](images/0ef75d46db01758059058fdc5084c36ae8f7da4761f8b7e7b386cd2c7d3f1770.jpg) +Output + +![](images/7c4c4a5913f8a345de088583bbb8d42af25a19035a8a360d0213197d5bb428f5.jpg) +Output: NOCS + 3D box + +![](images/7fe99a872f668d3e02d09b6a612fa8fa521921e1e8846569e8b5053137341184.jpg) + +![](images/a7d3b1bc3efc9aa5aff2ea352c5a4e219af6004624e4599109ad87551ed346fe.jpg) + +![](images/d1e6c396d4789631aed34fd84b72253a13bbc7c4726a8b8a8d149548ebd6d902.jpg) + +![](images/c93c50b71f3a0679cbee0d98cd5ddb9b7d40baa192daae55e87c6277130e90f3.jpg) + +![](images/edaef087919be95f65a5f8c5ffbe06550f0d08f4af2413dd3ce1bfa071c9615f.jpg) +3D shape + pose + +![](images/135fba02399742b52f7bf8d47d37cdc60ac04fdf5c906eeef00a8a3e17f2c426.jpg) +Input +Fig. 4: Example results of our single unified NOCSformer model across various datasets and object classes. The left column shows input images and query 2D bounding boxes. The center column shows the NOCS+instance maps predicted by NOCSformer along with the estimated 3D pose overlaid on the input image. The NOCS can be used with the 3D boxes and object size to lift the objects to a 3D pointcloud, which is shown in the right column. Last row contains two examples. + +![](images/3a6dd6f3997c50cc674624ac3f3d733e29435e73516117f81e5c2b75dd73fed8.jpg) +Output: 3D shapes and poses + +![](images/097c300d06565cea7c36582b922499faea658fd528ff385a152f90fca82cbed5.jpg) + +![](images/28e41a371836d1d79a9b4750352f22b011c92bc830eb73c3691e3ae75ee379db.jpg) + +![](images/1cbaa09de46cdf437354ebc4676ed9502833e92d77b062bdf21cac896a6c73b9.jpg) + +![](images/22cc7db8e1952ec5bf580d17ccce4c8628cc0343d2d6594597231b55fd5e1863.jpg) +Output +3D shape + pose + +![](images/7e24f158250b514172e978841e99570edd2f3e7d1ccdce649cdf5224a056b9d3.jpg) + +Datasets: For our experiments, we train NOCSformer on the OmniNOCS dataset containing 97 classes, holding out images from the NOCS-Real275 dataset. We use NOCS-Real275 to evaluate NOCSformer's cross-dataset generalization capabilities by not training on it. NOCS-Real275 has 6 classes that overlap with the rest of OmniNOCS, although they are not the top 20 classes. It also differs from the rest of OmniNOCS in terms of camera parameters and the context of objects in the scene. + +# 5.1 NOCS prediction accuracy + +We propose the use of NOCS mAE and NOCS mPSNR to evaluate the accuracy of predicted NOCS. For each object the mean Absolute Error (mAE) and mPSNR are computed for all points in the intersection of ground truth and predicted masks, and reported as a mean of per-category means. However, the NOCS mAE/mPSNR do not penalize undersegmentations / sparse predictions. Since we also would like the NOCS predictions to span the full visible instance, we also evaluate the 2D mask mIoU. + +We evaluate NOCSformer on a subset of OmniNOCS that contains 75 classes with accurate ground truth NOCS and masks, see Table 3. It is able to predict NOCS with errors less than $9\%$ (0.089 mAE on OmniNOCS). We also evaluate on the held-out NOCS-Real275 dataset, which was not used to train NOCSformer. We find that the zero-shot NOCSformer outperforms the NOCS baseline from [49] that was fully trained on this dataset, a strong indication of the generalization capabilities of NOCSformer. + +# 5.2 3D localization accuracy + +NOCS provides dense 3D-2D correspondences that can be used to estimate the 3D object oriented bounding box. This is done by using the predicted object size and solving a PnP problem, as explained in Sec. 4.2. We evaluate the accuracy of our estimated bounding boxes in different settings: + +Outdoor scenes: We compare NOCSformer's 3D localization accuracy to that of other 3D detection models using the nuScenes true positive localization metrics on the challenging nuScenes dataset (included in OmniNOCS). We compare to Cube R-CNN, since it is the only other model that generalizes across diverse datasets. However, Cube R-CNN has two notable differences to NOCSformer: 1) being a detection model, it also jointly detects 2D object regions, and 2) it uses Chamfer loss causing it to have high orientation errors. As a more comparable baseline, we use a variant of our model (termed "Cubeformer"), by replacing our NOCS head with the cube head from Cube R-CNN + +
MethodNOCS-Real275OmniNOCS
NOCS MAE↓NOCS PSNR↑Mask IoU↑NOCS MAE↓NOCS PSNR↑Mask IoU↑
NOCS baseline [49]0.12116.34586.10---
NOCSformer0.10718.52789.030.09420.24578.50
+ +Table 3: NOCS and mask prediction evaluation: We report metrics for NOCSformer on 75 classes of OmniNOCS: it is the first model that is capable on predicting NOCS on such diverse data. Additionally, NOCSformer outperforms [49] which trained on NOCS-Real275, without training on NOCS-Real275. + +
MethodMulti-datasetmATE (m)↓mAOE (rad)↓mASE (%)↓mIoU↑
FCOS3D [51]X0.7770.4000.231-
PGD [50]X0.6750.3990.236-
EProPNP [12]X0.6760.3630.263-
EProPNP + TTA [12]X0.6530.3190.255-
Cube R-CNN [7]✓0.6501.3050.2830.349
Cubeformer✓0.7900.5730.3010.280
NOCSformer✓0.8870.5580.2910.377
+ +without a Chamfer loss. It uses the input 2D boxes. More details about Cubeformer are in the supplementary material. From Table 4, we find that NOCSformer's box orientations are canonical and more accurate than both Cube R-CNN and Cubeformer. Moreover, the boxes estimated from NOCSformer's NOCS predictions are comparable in translation and scale errors to those of Cube R-CNN and Cubeformer, even though it does not directly regress depth. While all 3 methods generalize across several classes and domains, baselines trained solely on nuScenes localization (top half) significantly outperform them on this task. + +Cross-dataset generalization (indoor): Here, we hold out NOCS-Real275 from OmniNOCS when training NOCSformer, and use it to evaluate NOCSformer's ability to generalize to unseen domains. NOCS-Real275 features a tabletop setting with multiple objects, which is semantically different from our other indoor datasets. We compare against Cube R-CNN [7] and the NOCS [49] model that was trained from scratch on this dataset alone. [49] also uses class-specific model parameters and additional losses for symmetric objects that do not scale to larger vocabularies. + +The results are shown in Table 5. We find that NOCSformer is more accurate at transferring to this unseen dataset compared to Cube R-CNN, indicating the generalizability of NOCS-based localization methods over that of box-regression methods. Note that the mAP metrics used in the NOCS-Real275 datasets may also be affected by the false positives/negatives from Cube R-CNN. However, both models are worse than the supervised NOCS baseline. + +Table 4: Outdoor localization: Comparison of 3D localization accuracy on the nuScenes subset of OmniNOCS. Top half shows methods that are only trained on nuScenes: while these perform better on nuScenes itself, they do not generalize to other datasets and classes. OmniNOCS is competitive with Cube R-CNN on mIoU while also predicting canonical orientations and object coordinates. Note that mIoU and orientation error only uses the predicted yaw orientation component. + +
MethodTransferred?Depth input?mAP
3D IoU @ 253D IoU @ 50
NOCS model [49]X✓79.672.4
CubeRCNN [7]✓X14.94.1
NOCSformer✓X43.510.6
+ +Table 5: Cross-dataset generalization: Comparison of localization accuracy on NOCS-Real275 for all classes, using mAP at different thresholds. This dataset has been held out when training CubeRCNN and NOCSformer. NOCSformer is able to generalize to NOCS-Real275 without any additional training. + +![](images/7c23736e5a82817d3573db53649829c2d26c417528908f20903bc18b98a89633.jpg) +Fig. 5: Generalization across datasets: NOCSformer can generalize to new datasets that present new camera models and object domains. We show this zero-shot capability of NOCSformer (bottom row) on the NOCS-Real275 test set without training on NOCS-Real275 dataset. The predictions from a NOCS baseline [49] trained explicitly on the NOCS-Real275 dataset are shown in the top row. + +# 5.3 3D orientation accuracy + +A key challenge for 3D localization models trained on multiple datasets [7] is predicting canonical object orientations, as these datasets typically differ in their object pose conventions. OmniNOCS ensures that object poses in all constituent datasets are consistent within a category, enabling our model to predict canonical orientations. We evaluate this by computing the accuracy at different orientation error threshold on KITTI-val [13] in Table 6. We compare the accuracy of our orientation predictions to those of Cube R-CNN [7], which are trained using Chamfer loss, and therefore suffers at predicting canonical orientations. We also compare against our Cubeformer baseline that directly supervises the orientations. We find that orientations estimated using NOCSformer are more consistent and accurate than both our baselines. + +
MethodGravity axisHeading axis (X)
Accuracy @ 1 degAccuracy @ 5 degAccuracy @ 90 degAccuracy 5 degAccuracy @ 10 degAccuracy @ 90 deg
CubeRCNN5.7518.2951.7419.4321.5534.17
Cubeformer80.7998.52100.041.9955.8877.26
NOCSformer84.2499.95100.049.1659.6581.99
+ +Table 6: 3D orientation accuracy for models trained on multiple datasets: Models that use Chamfer distance (CubeRCNN) for supervising orientation heads end up being inconsistent in their orientation predictions. The results are averaged over 5 classes in the KITTI-val subset [13]. + +
ArchitectureNOCS PSNR↑Mask IoU↑
NOCSformer20.43181.69
Hourglass head-3.064-3.16
MaskRCNN head-5.23-21.27
w/o discretized prediction-4.11-
+ +Table 7: Comparison of different architecture choices for the NOCS and mask prediction head: We experiment with transformer, hourglass and MaskRCNN architectures. We compare whether discretized prediction of NOCS is better than continuous regression for transformer heads. We quantify the difference in performance when using separate heads for mask and NOCS prediction as opposed to a single head. + +# 5.4 Ablations + +NOCS head architecture: While all existing architectures use a few CNN layers for their NOCS heads [49], we find that this significantly limits the NOCS prediction performance when scaling to more classes. Using our transformer NOCS head provides a 5.23dB improvement in NOCS PSNR and a $21.27\%$ improvement in mask mIoU. We also experiment with a larger convolutional head: the Hourglass model from [6]. We find that the NOCSformer NOCS heads are even better than Hourglass, with an improvement of 3.06dB on NOCS PSNR and $3.16\%$ on mask IoU (as shown in Table 7). + +Discretized versus continuous prediction of NOCS: NOCSformer models the final NOCS prediction as a classification layer with 50 non-overlapping bins. This was observed to be better than a linear regression layer with a MaskRCNN head in [49]. We find that it is also significantly better when using a transformer head, improving the NOCS prediction PSNR by 4.11 dB, as in Table 7. + +# 6 Conclusion and Future work + +This paper has introduced a new large scale dataset of Normalized Object Coordinates (NOCS) for a wide variety of object classes in indoor and outdoor scenes. It has also proposed a single transformer-based model NOCSformer that can predict object 3D pose, size, and NOCS for any of these objects given 2D bounding box inputs. These allows our model to obtain 3D shape and oriented bounding boxes of objects in metric scale from a single input image. This represents the first attempt to generalize NOCS estimation beyond small datasets of narrow domain, increasing the number of object categories available by an order of magnitude. We hope this provides a means for others to explore large-scale monocular 3D object pose and shape estimation. + +Some limitations of our current method include the handling of symmetric objects, where the coordinate system has multiple possible solutions, and reflected geometry, such as left / right shoes. Future work could address these issues, for example, minimizing over multiple coordinate frames in the loss for symmetric objects, and potential estimation of left / right coordinate frames for reflected geometry. + +# References + +1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2021) +2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: CVPR (2019) +3. Avetisyan, A., Dai, A., Nießner, M.: End-to-end cad model retrieval and 9dof alignment in 3d scans. In: ICCV (2019) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022) +5. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1) (2021), https://openreview.net/forum?id=tjZjv_qh_CE +6. Birodkar, V., Lu, Z., Li, S., Rathod, V., Huang, J.: The surprising impact of mask-head architecture on novel class segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7015-7025 (2021) +7. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023) +8. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019) +9. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2. arXiv preprint arXiv:2001.10773 (2020) +10. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +11. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +12. Chen, H., Wang, P., Wang, F., Tian, W., Xiong, L., Li, H.: Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2781-2790 (2022) +13. Chen, X., Kundu, K., Zhu, Y., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals using stereo imagery for accurate object class detection. IEEE transactions on pattern analysis and machine intelligence 40(5), 1259-1272 (2017) +14. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: CVPR (2020) +15. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016) +16. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) + +17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +18. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022) +19. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012) +20. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: European Conference on Computer Vision. pp. 516-532. Springer (2022) +21. Gumeli, C., Dai, A., Nießner, M.: Roca: Robust cad model retrieval and alignment from a single image. In: CVPR (2022) +22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) +23. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 923-934 (2021) +24. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: IEEE International Conference on Robotics and Automation (ICRA) (2022), https://arxiv.org/abs/2203.01929 +25. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape appearance and pose optimization. In: European Conference on Computer Vision (ECCV) (2022), https://arxiv.org/abs/2207.13691 +26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +27. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation. In: CVPR (2022) +28. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375 +29. Kuo, W., Angelova, A., Lin, T.Y., Dai, A.: Mask2cad: 3d shape prediction by learning to segment and retrieve. In: ECCV (2020) +30. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision 81, 155-166 (2009). https://doi.org/ https://doi.org/10.1007/s11263-008-0152-6, http://infoscience. epfl.ch/record/160138 +31. Li, K., DeTone, D., Chen, Y.F.S., Vo, M., Reid, I., Rezatofighi, H., Sweeney, C., Straub, J., Newcombe, R.: Odam: Object detection, association, and mapping using posed rgb video. In: ICCV (2021) +32. Li, Z., Wang, G., Ji, X.: Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7678-7687 (2019) +33. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15641-15650 (2021) +34. Manhardt, F., Kehl, W., Gaidon, A.: Roi-10d: Monocular lifting of 2d detection to 6d pose and metric shape. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2069-2078 (2019) + +35. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Vid2cad: Cad model alignment using multi-view constraints from videos. TPAMI (2022) +36. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Cad-estate: Large-scale cad model annotation in rgb videos. In: ICCV (2023) +37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +38. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017) +39. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020) +40. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +41. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: Proceedings of the IEEE international conference on computer vision. pp. 3828-3836 (2017) +42. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3D appearance, location & pose. In: CVPR (2022) +43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) +44. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: International Conference on Computer Vision (ICCV) 2021 (2021) +45. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 567-576 (2015). https://doi.org/10.1109/CVPR.2015.7298655 +46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +47. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: CVPR (2018) +48. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16611-16621 (2021) +49. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +50. Wang, T., Xinge, Z., Pang, J., Lin, D.: Probabilistic and geometric depth: Detecting objects in perspective. In: Conference on Robot Learning. pp. 1475-1485. PMLR (2022) +51. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021) +52. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022) + +53. Wen, B., Lian, W., Bekris, K., Schaal, S.: Catgrasp: Learning category-level task-relevant grasping in clutter from simulation. ICRA 2022 (2022) +54. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017) +55. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21919-21928 (2023) +56. Zhang, J., Herrmann, C., Hur, J., Cabrera, L.P., Jampani, V., Sun, D., Yang, M.H.: A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence arXiv preprint arxiv:2305.15347 (2023) +57. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019) \ No newline at end of file diff --git a/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/images.zip b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2074a8326ebe6245f2bc344af2b92d8935b9c792 --- /dev/null +++ b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:398162b5a41070c4d2466819de2584c6b2432fae8241d3d11864a37703eff4c5 +size 624722 diff --git a/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/layout.json b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..14107d17db46f679f641a14377906359f64dc111 --- /dev/null +++ b/2024/OmniNOCS_ A unified NOCS dataset and model for 3D lifting of 2D objects/layout.json @@ -0,0 +1,8761 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 112, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 112, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 112, + 479, + 148 + ], + "type": "text", + "content": "OmniNOCS: A unified NOCS dataset and model for 3D lifting of 2D objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": "Akshay Krishnan" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", Abhijit Kundu" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", Kevis-Kokitsi Maninis" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", James Hays" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", and Matthew Brown" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 268, + 201, + 345, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 201, + 345, + 213 + ], + "spans": [ + { + "bbox": [ + 268, + 201, + 345, + 213 + ], + "type": "text", + "content": "1 Google Research†" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 213, + 369, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 213, + 369, + 224 + ], + "spans": [ + { + "bbox": [ + 244, + 213, + 369, + 224 + ], + "type": "text", + "content": "2 Georgia Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 231, + 241, + 364 + ], + "blocks": [ + { + "bbox": [ + 136, + 231, + 241, + 364 + ], + "lines": [ + { + "bbox": [ + 136, + 231, + 241, + 364 + ], + "spans": [ + { + "bbox": [ + 136, + 231, + 241, + 364 + ], + "type": "image", + "image_path": "c6cdaaff917a5eba3946dee9b9ca5ebdc840d3a5a95ab822c2bb341eca8803cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 368, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 413 + ], + "type": "text", + "content": "Fig. 1: We introduce OmniNOCS, a large-scale dataset with Normalized Object Coordinates (NOCS), instance masks, and 3D box annotations for objects across several classes, domains, and cameras. We also propose NOCSformer, a model trained on OmniNOCS that lifts each 2D object bounding box in an image to its corresponding 3D oriented box (pose) and 3D pointcloud (shape)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 243, + 232, + 359, + 363 + ], + "blocks": [ + { + "bbox": [ + 243, + 232, + 359, + 363 + ], + "lines": [ + { + "bbox": [ + 243, + 232, + 359, + 363 + ], + "spans": [ + { + "bbox": [ + 243, + 232, + 359, + 363 + ], + "type": "image", + "image_path": "6fd026d4e6665924bd49b1a033471c917353d1a4d0c01170f5b7c1310616d6f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 360, + 234, + 475, + 364 + ], + "blocks": [ + { + "bbox": [ + 360, + 234, + 475, + 364 + ], + "lines": [ + { + "bbox": [ + 360, + 234, + 475, + 364 + ], + "spans": [ + { + "bbox": [ + 360, + 234, + 475, + 364 + ], + "type": "image", + "image_path": "6a0d92c0efe8f14ba295dcc2483152ac8b33510ea8aef85261200f64e80b54f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 425, + 455, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 425, + 455, + 590 + ], + "spans": [ + { + "bbox": [ + 159, + 425, + 455, + 590 + ], + "type": "text", + "content": "Abstract. We propose OmniNOCS, a large-scale monocular dataset with 3D Normalized Object Coordinate Space (NOCS) maps, object masks, and 3D bounding box annotations for indoor and outdoor scenes. OmniNOCS has 20 times more object classes and 200 times more instances than existing NOCS datasets (NOCS-Real275, Wild6D). We use OmniNOCS to train a novel, transformer-based monocular NOCS prediction model (NOCsformer) that can predict accurate NOCS, instance masks and poses from 2D object detections across diverse classes. It is the first NOCS model that can generalize to a broad range of classes when prompted with 2D boxes. We evaluate our model on the task of 3D oriented bounding box prediction, where it achieves comparable results to state-of-the-art 3D detection methods such as Cube R-CNN. Unlike other 3D detection methods, our model also provides detailed and accurate 3D object shape and segmentation. We propose a novel benchmark for the task of NOCS prediction based on OmniNOCS, which we hope will serve as a useful baseline for future work in this area. Our dataset and code will be at the project website: https://omninocs.github.io." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 605, + 218, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 218, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 218, + 618 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "type": "text", + "content": "Predicting the 6 Degree-of-Freedom (6DoF) pose and shape of objects from images is a crucial problem in 3D scene understanding. Robots need to understand the location," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 654, + 223, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 223, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 223, + 665 + ], + "type": "text", + "content": "†Now at Google DeepMind" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 212 + ], + "type": "text", + "content": "shape, and orientation of various objects to grasp and interact with them. Self-driving vehicles need to understand the location and heading of vehicles, pedestrians, and other objects on the road. In particular, the 3D orientation of these objects is crucial to predict their behavior. It is also important in AR/VR applications, as it allows users to interact with objects in meaningful ways. Predicting object shape and pose from monocular images is also a prerequisite to initialize 3D panoptic neural scene graphs [27, 39] and to track objects in videos [42]. These applications require approaches that generalize across a wide range of object classes, environments, and camera models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 214, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 482, + 322 + ], + "type": "text", + "content": "The problem of localizing 3D objects has been extensively studied through the lens of monocular 3D object detection or 6DoF pose estimation. The most commonly used approach is to represent objects as 3D cuboids, and train a model to regress the cuboid parameters from a 2D RoI of the object [7, 8, 28, 41, 51, 54]. An alternative method involves predicting corresponding 3D points in an object coordinate space followed by pose estimation using 3D-2D alignment [11, 12, 20, 33, 49]. However, all these methods are limited to narrow datasets collected on a single camera and context with typically very few object classes. Existing models are also trained separately for every dataset, preventing them from being used more widely." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 324, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 324, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 133, + 324, + 482, + 420 + ], + "type": "text", + "content": "Recent work [7] takes a step towards scalability by introducing the Omni3D dataset, which aggregates 3D detection datasets from different domains, and trains a Cube R-CNN model to regress 3D bounding boxes for 50 classes. However, Omni3D notably lacks consistently oriented, object-centric boxes, as the ground truth orientations are not canonical for object classes. Cube R-CNN [7] instead uses a Chamfer distance loss which is invariant to the predicted 3D cuboid orientation. This causes it to predict inconsistent object orientations, for example, often flipping the orientations of cars by " + }, + { + "bbox": [ + 133, + 324, + 482, + 420 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 133, + 324, + 482, + 420 + ], + "type": "text", + "content": ". Further, Omni3D is a detection dataset that does not provide object shape information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 423, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 423, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 133, + 423, + 482, + 603 + ], + "type": "text", + "content": "Our work aims to overcome the above shortcomings, by providing a new large-scale dataset with consistent object-centric ground truth boxes along with detailed shape. We argue for the use of Normalized Object Coordinate Space (NOCS) as proposed in [49] as a 3D object shape representation. NOCS represents both the canonical orientation and the shape of the visible surface of the object, properties that are essential for real-world applications such as self-driving and robotics. It can also be used to estimate the 3D bounding box of the object. However, all existing work on NOCS [24, 25, 32, 48, 49] only train small models on small datasets with fewer than 10 classes. We address the lack of diverse NOCS datasets by proposing a large-scale monocular NOCS dataset, OmniNOCS that has NOCS annotations, instance segmentation, and canonically oriented bounding boxes for 97 object classes, containing " + }, + { + "bbox": [ + 133, + 423, + 482, + 603 + ], + "type": "inline_equation", + "content": "380\\mathrm{k}" + }, + { + "bbox": [ + 133, + 423, + 482, + 603 + ], + "type": "text", + "content": " images from 10 different data sources, making it the largest and most diverse NOCS dataset currently available. OmniNOCS includes all of the data from Omni3D (KITTI [19], nuScenes [10], ARKitScenes [5], Objectron [1], Hypersim [44], SUN-RGBD [45]), with the addition of Cityscapes [15], virtual KITTI [9], NOCS-Real275 [49], and the Waymo Open Dataset [46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 605, + 482, + 665 + ], + "type": "text", + "content": "Our work also introduces a new model, which we term \"NOCSformer\", that predicts NOCS coordinates and oriented 3D bounding boxes from monocular images and 2D detections for all the classes in OmniNOCS. NOCSformer leverages self-supervised pretrained ViTs [17, 40] and does not use any class-specific model parameters. This enables it scale to large vocabularies, and share information across semantically similar" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 223 + ], + "type": "text", + "content": "object classes. In contrast, existing NOCS models use small NOCS heads with class-specific parameters that significantly limit their performance on large datasets like OmniNOCS. NOCSformer also predicts the 3D size of the object along with its 3D orientation using a learned PnP head. This allows NOCSformer to predict 3D oriented bounding boxes and object point cloud in metric scale for diverse object categories. Our experiments evaluate the quality of NOCS and bounding boxes predicted by NOCSformer in comparison to existing NOCS prediction or 3D detection models. We find that training on OmniNOCS allows NOCSformer to generalize to unseen datasets, even outperforming baselines trained on the target dataset in NOCS prediction accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 224, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 224, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 147, + 224, + 287, + 236 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 238, + 480, + 354 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 238, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 479, + 270 + ], + "type": "text", + "content": "- The OmniNOCS dataset: A new dataset containing Normalised Object Coordinates for 380k images in 97 categories, an order of magnitude larger on both counts than existing NOCS datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 274, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 319 + ], + "type": "text", + "content": "- NOCSformer: A novel transformer-based architecture for predicting object NOCS, mask, and size from input 2D boxes, utilizing pre-trained self-supervised ViT backbones. NOCSformer is the first NOCS model to generalize to vocabularies with " + }, + { + "bbox": [ + 138, + 274, + 480, + 319 + ], + "type": "inline_equation", + "content": "90+" + }, + { + "bbox": [ + 138, + 274, + 480, + 319 + ], + "type": "text", + "content": " classes and to unseen datasets, including images from internet collections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 322, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 354 + ], + "type": "text", + "content": "- OmniNOCS benchmark: an evaluation framework with metrics for directly comparing different NOCS prediction algorithms on OmniNOCS, with baselines established via NOCSFormer." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 376, + 221, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 376, + 221, + 388 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 221, + 388 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 400, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 400, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 400, + 482, + 460 + ], + "type": "text", + "content": "The task of predicting 3D object pose has been studied both in the context of camera/object pose estimation (6DoF) and 3D bounding box estimation (6DoF pose + 3DoF size). A further distinction can be made between methods that regress directly to object pose / bounding box parameters, and methods that compute per pixel coordinates or depth as an initial step. We review each of these paradigms in the sections below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 478, + 349, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 478, + 349, + 491 + ], + "spans": [ + { + "bbox": [ + 133, + 478, + 349, + 491 + ], + "type": "text", + "content": "2.1 3D localization by regressing bounding boxes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 498, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 482, + 617 + ], + "type": "text", + "content": "Direct approaches to 3D object detection and pose estimation involve networks that output rotation, translation and scale parameters directly. Several works have explored different parameterizations in this setting, e.g., PoseCNN [54] uses regression of translation via centre direction + distance maps (which enable detection even under occlusion), and quaternion representation of rotation. BB8 [41] uses segmentation followed by a CNN to regress to 2D projections of the " + }, + { + "bbox": [ + 133, + 498, + 482, + 617 + ], + "type": "inline_equation", + "content": "8 \\times 3\\mathrm{D}" + }, + { + "bbox": [ + 133, + 498, + 482, + 617 + ], + "type": "text", + "content": " bounding box corners. Multi-view monocular approaches have also been proposed, e.g., DETR3D [52] which uses DETR-style attention to reason about object interactions. Several techniques also make use of existing 2D bounding box predictors, either as an input to a 3D lifting approach [34], or as a constraint on 3D box predictions [38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 617, + 482, + 665 + ], + "type": "text", + "content": "Another group of works focuses on predicting the alignment of 3D CAD models within various modalities: images [21, 29], videos [31, 35], or 3D scans [2, 3]. These approaches typically determine the object's 3D translation, rotation, and size (9DoF) and find a CAD model with a similar visual appearance." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "text", + "content": "2.2 3D localization from model-to-image alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 204 + ], + "type": "text", + "content": "An alternative approach to 3D localization is to first predict correspondence, either between views or to a normalized space, and then reason over redundant correspondences to establish pose. This has been done with both sparse and dense correspondences, e.g., AutoShape [33] makes use of sparse 2D to 3D correspondences, with a learned shape model and sparse 2D keypoints. [20] uses pairwise semantic correspondence from ViT, to find pose between a reference image and a sequence of targets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 300 + ], + "type": "text", + "content": "Several approaches use dense correspondences and the idea of Normalised Object Coordinate Spaces, or NOCS. In the original NOCS work, Wang et al. [49] use RGB-D data and 3D-3D correspondences for pose estimation. Follow-up work [11] estimates pose using PnP variants on 2D-3D correspondences. While NOCS [49] used the average spatial dimensions of the object category to define the object frame, [53] showed that an instance-specific NOCS coordinate frame can be used alongside a predicted instance size for the same purpose. Other methods combine coordinate regression and direct approaches, e.g., [48], which uses direct regression based on dense correspondences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "type": "text", + "content": "Similar to our work, MonoRUun [11] lifts 2D detections to 3D object coordinates without explicit class supervision. They also use a novel self-supervised coordinate regression training loss which obviates the need for detailed 3D ground truth. However, their representation does not include masks and therefore does not provide explicit object shape information. Their evaluation is also limited to 3 classes (car, pedestrian, cyclist) on the KITTI-Object test set. M3D-RPN [8] also has a single 3D head for multiple classes, jointly generating 2D and 3D bounding box predictions, though without shape information, and similarly limited to car, pedestrian and cyclist classes on KITTI." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 410, + 394, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 394, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 394, + 422 + ], + "type": "text", + "content": "2.3 Monocular object localization / pose estimation datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 498 + ], + "type": "text", + "content": "For 3D object detection from monocular RGB images, most existing works use a small number of classes on a single dataset. Cube R-CNN/Omni3D [7] contribute towards creating a general purpose monocular 3D object detector. Cube R-CNN performs well over six 3D datasets: KITTI, SUN RGB-D, ARKitScenes, Objectron, nuScenes and Hypersim. However, the method is class specific, with Cube R-CNN trained only on 50 classes, and unable to work in the open class setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "content": "Cube R-CNN also uses a Chamfer loss on the predicted 3D box corners to deal with the inconsistent coordinate frame annotations in the underlying Omni3D datasets. The lack of direct orientation supervision results in inconsistent orientation predictions (for example, the positive x axis may point to the front or the back of the car for different instances). This orientation inconsistency is problematic for methods that seek to build detailed 3D models of object categories as a downstream task [27, 39]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Other works are not limited to 3D boxes, and directly provide object shapes. These shapes come in the form of annotated point clouds (e.g., ScanNet [16]), or aligned CAD models (e.g., Scan2CAD [2], Pix3D [47], CAD-Estate [36]). Object point clouds are created by labeling points on scanned 3D scenes, a process that doesn't scale well to large datasets. Semi-automatically aligning CAD models is scale-able to a certain extent, but since it relies on retrieving existing models, the resulting shapes are rarely accurate, and the alignments are sensitive to deformations and movable parts. In contrast to these works, we propose a large-scale dataset of many different categories." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 114, + 427, + 170 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 427, + 170 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 427, + 170 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 427, + 170 + ], + "type": "table", + "html": "
DatasetNOCS GTReal#Images#Classes#Instances
CAMERA25 [49]✓×300k6184
NOCS-Real275 [49]✓✓7k624
Wild6D [18]×✓1M51.8K
OmniNOCS✓✓380k97>450k
", + "image_path": "9b6ca27155eb41b8344abde494fd93de510b97ba9fb9e956ca22813c78cf0d75.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "text", + "content": "Table 1: Comparison to other NOCS datasets: Other existing NOCS datasets are limited in number of classes and instances. Note that Wild6D has 1M images, but these are used for self-supervised training, as it does not include NOCS ground truth (GT)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 215, + 481, + 267 + ], + "blocks": [ + { + "bbox": [ + 133, + 215, + 481, + 267 + ], + "lines": [ + { + "bbox": [ + 133, + 215, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 215, + 481, + 267 + ], + "type": "image", + "image_path": "438dbb5de78e5233772616e4da6306952f437af2f155c8c65d4cfe25df4e33e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 275, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 275, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: Preparation of the OmniNOCS dataset: First, we ensure that object orientations are canonical across different all subsets of OmniNOCS. Next, we re-compute depth for datasets where depth is not available or is too noisy. Finally, we annotate objects with (pseudo) instance mask labels where ground truth masks are not available." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 255, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 255, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 255, + 355 + ], + "type": "text", + "content": "3 OmniNOCS dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 368, + 218, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 218, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 218, + 379 + ], + "type": "text", + "content": "3.1 Data statistics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": "We create OmniNOCS, a first-of-its-kind large and diverse NOCS dataset comprising data from several classes and domains. OmniNOCS uses data from self-driving scenes [9, 15, 19, 46], indoor scenes [5, 44, 45, 49], and object-centric videos [1]. Each of these sources use cameras with different parameters, ranging from phone cameras to wide FOV cameras mounted on self-driving cars. The number of object instances per image also varies widely, from single-object images [1] to images with hundreds of objects [44, 46]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 460, + 483, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 483, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 483, + 579 + ], + "type": "text", + "content": "OmniNOCS has 97 object categories across 380k images. This by far exceeds the diversity of existing NOCS datasets (with less than 6 categories as shown in Table 1). The number of instances exceeds that of other NOCS datasets by more than 2 orders of magnitude. It also contains more images with NOCS annotations (380k). Note that while Wild6D has 1M images, these do not have ground truth NOCS or pose annotations, and includes every frame of object-centric videos. OmniNOCS is also more diverse than the most diverse 3D detection dataset (Omni3D [7] with 98 categories). 307k images in OmniNOCS are from real world scenes and 73k are synthetic. 107k of the images in OmniNOCS are from outdoor scenes, 102k images are from complex indoor environments, and 140k are from object-centric videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 597, + 231, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 231, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 231, + 609 + ], + "type": "text", + "content": "3.2 Data preparation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Computing the NOCS for objects in an image requires canonically oriented 3D object bounding boxes, depth, and instance segmentations. Since the datasets we use were originally only proposed for monocular 3D detection, many of them lack dense depth or instance segmentation annotations (highlighted entries in Table 2). Additionally, NOCS" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 113, + 461, + 255 + ], + "blocks": [ + { + "bbox": [ + 150, + 113, + 461, + 255 + ], + "lines": [ + { + "bbox": [ + 150, + 113, + 461, + 255 + ], + "spans": [ + { + "bbox": [ + 150, + 113, + 461, + 255 + ], + "type": "table", + "html": "
3D DatasetDepth sourceInstance mask source#Images#Classes
KITTI [19]LiDARHuman-labelled [23]7.4k7
SUN-RGBD [45]Depth CameraSegmentation model [26]10k72
Objectron [1]NeRFNeRF + Human label132k9
nuScenes [10]LiDARSegmentation Model [14]30k9
Hypersim [44]SyntheticSynthetic64k31
ARKitScenes [5]Depth CameraSegmentation model [26]60k15
Cityscapes 3D [15]StereoHuman-labelled3.4k8
Virtual KITTI [9]SyntheticSynthetic4.1k3
NOCS-Real275 [49]Depth cameraHuman-labelled7k6
Waymo OD [46]LiDARHuman-labelled62k7
OmniNOCS380k97
", + "image_path": "8857892129c5c11d3852279137f8cbc2f147c75cdc8c922918ad788fa4998aef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 258, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 304 + ], + "type": "text", + "content": "Table 2: Constituent 3D detection datasets used in OmniNOCS: We augment many of these datasets with depth and masks, as they are missing in the original data (highlighted entries). In addition, we canonicalize the orientations of bounding boxes across all datasets. #Classes lists the number of classes we use from each dataset. Those above the dashed line are part of Omni3D [7]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 316, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 422 + ], + "type": "text", + "content": "also requires all instances within a category to have consistent orientation annotations. While this is certainly not true across different 3D detection datasets, it is also not true within some large synthetic datasets such as Hypersim. OmniNOCS aggregates several datasets, with the addition of 1) new depth estimated via Mip-NeRF reconstructions, 2) additional instance masks via segmentation models and human labels, 3) manual labelling of coordinate axes to give consistent object-centric coordinate frames. This enables NOCS to be computed for every image in the dataset. This multi-stage process for OmniNOCS creation is illustrated in Figure 2 and explained below. A summary of the resulting OmniNOCS dataset is provided in Table 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 424, + 483, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 483, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 483, + 567 + ], + "type": "text", + "content": "Orientation canonicalization: Although the datasets we use contain oriented 3D bounding boxes, they vary in their level of canonicalization. Some datasets have their class-canonical orientations i.e., all instances of a particular class are oriented consistently within the dataset (for example, all cars in [19] have X axis pointing forward, and Z upwards). In such cases, we ensure that this canonicalization is consistent with all of OmniNOCS by applying a fixed class-specific offset orientation for the dataset. Datasets like Hypersim [44] have no class canonicalization at all, ie, although objects have tight bounding boxes, their XYZ axes directions are different for each instance. We manually label each object in such datasets to select the canonical orientation out of six possible orientations for the bounding box. More details on the labelling process are provided in the supplementary material. Some classes may have more than one choice for canonical orientations (due to symmetry), in which case an orientation is selected arbitrarily." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "content": "Depth augmentation: For outdoor datasets, we use sparse depth from LiDAR if available. We recomputed the depth on Cityscapes using a state-of-the-art stereo depth model [55], as we found the depth from the original dataset to be noisy. Since Objectron [1] does not contain dense depth, we train Neural Radiance Fields [4, 37] for each video sequence in the dataset to obtain dense depth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Instance segmentation: Since many datasets we use were intended for 3D detection, they do not camera instance segmentations. We annotate ARKitScenes and SUNRGBD objects using instance masks from Segment-Anything (SAM) [26], prompting it with the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "projected 3D bounding box. Although SUN-RGBD provides segmentation masks, we find that the SAM annotations are of superior quality. For Objectron, we create accurate masks by efficiently annotating in 3D space. Specifically, we apply the pipeline of [4] to each of the Objectron videos. We fuse the resulting NeRF depth-maps to create a mesh. We post-process the 3D mesh, and get rid of the redundant vertices while keeping the vertices of the object. Finally, we create the masks for multiple frames by measuring the distance between each pixel on the depth map and the object mesh: pixels whose depth is far from the surface of the mesh are discarded, a process which also handles occlusions. NOCS computation: As proposed in [49], the NOCS coordinate " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "x_{nocs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " for a point on the object surface is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 252, + 253, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 253, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 252, + 253, + 481, + 277 + ], + "type": "interline_equation", + "content": "x _ {n o c s} = \\frac {1}{s} ^ {o b j} T _ {c a m} x _ {c a m} \\tag {1}", + "image_path": "c34e8cdcb8243ed2744f707d6ec6e2c838c6e224f93d227136e372f613aff9c1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "{}^{obj}T_{cam} = \\left[{}^{obj}R_{cam}\\right|^{obj}t_{cam}" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": " is the 6DoF transformation from the camera to the object frame, " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "x_{cam}" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": " the 3D location of the point in the camera frame, and " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": " is a scalar, the size of the diagonal of the object's tight bounding box. It can be interpreted as the object shape scaled to a box with a unit diagonal. For each image, starting from a 3D pointcloud " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "(x_{cam})" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": " obtained from the depth, we collect points on each object " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "(x_{cam}^i)" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": " using its 3D bounding box and instance mask, which can be transformed to the NOCS coordinate " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "(x_{nocs}^i)" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": " using (1). We store NOCS as a 3-channel 2D map (i.e. " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "x_{nocs}^i" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": " at its corresponding 2D location obtained by projecting " + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "inline_equation", + "content": "x_{cam}^i" + }, + { + "bbox": [ + 130, + 285, + 482, + 405 + ], + "type": "text", + "content": "). Our final result is the OmniNOCS dataset which contains instance segmentations, NOCS maps, and 3D bounding boxes for objects across 97 classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 423, + 257, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 257, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 257, + 436 + ], + "type": "text", + "content": "4 NOCSformer model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 509 + ], + "type": "text", + "content": "We propose a novel architecture for monocular NOCS prediction termed \"NOCsformer\" trained on OmniNOCS. NOCsformer primarily uses self-attention layers, and is the largest trained monocular NOCS prediction model to date. NOCsformer also contains a 3D size head and a learned PnP head that are used to estimate a canonically oriented 3D object bounding box from the NOCS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 526, + 241, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 241, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 241, + 537 + ], + "type": "text", + "content": "4.1 Model architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 569 + ], + "type": "text", + "content": "As shown in Figure 3, the NOCSformer architecture comprises an image backbone, a NOCS head, a size head, and a learned PnP head." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "**Backbone:** Our backbone is a frozen Vision Transformer (ViT) [17], that uses DINOv2 weights [40] from self-supervised pretraining. This choice is motivated by recent findings on using frozen DINOv2 backbones for understanding depth, multi-view correspondences, or relative pose [20, 56]. Additionally, we use a Dense Prediction Transformer (DPT) [43] architecture to fuse low-resolution DINOv2 representations from multiple intermediate layers and upsample them by a factor of 8. This higher feature resolution is desirable to improve predictions on smaller objects. We train the DPT layers while keeping the ViT layers frozen. We use 2D input boxes to sample the DPT features using" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 196 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 196 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 196 + ], + "type": "image", + "image_path": "54ac7709470a5519de97dce65bee88326729adc990e8400865551bc77a2d6051.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 483, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 483, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 483, + 264 + ], + "type": "text", + "content": "Fig. 3: Architecture for NOCSformer: We use a transformer backbone to extract features from the input image, pool them using the 2D box RoIs, and feed the per-RoI features to the NOCS and size heads. Our novel NOCS head jointly predicts the NOCS and instance mask for the RoI. Our learned PnP head for pose estimation uses the predicted NOCS and instance mask to predict the projected 3D coordinate and 3D rotation of the object." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 274, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 321 + ], + "type": "text", + "content": "RoIAlign [22] into a 28x28 fixed resolution grid that is fed to the NOCS and size heads. While DPT has been previously leveraged for tasks that provide dense supervision, such as semantic segmentation and depth prediction, we use it in an instance-level prediction context, training it using supervision from object RoIs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 429 + ], + "type": "text", + "content": "NOCS head: All existing NOCS prediction models [32, 48, 49] use MaskRCNN heads [22] with convolutional layers, in some cases with a separate head for each NOCS coordinate [49]. While this works on small datasets, we find that it significantly limits performance on OmniNOCS. NOCSformer therefore uses a simple (but large) transformer NOCS head with 10 self-attention blocks and one final upsampling conv layer to jointly predict both the instance mask and the NOCS map in the same output head. Similar to [49], for each NOCS coordinate, we predict logits over non-overlapping bins, and use the softmax to obtain the final NOCS value. Our ablations show that this choice works best on our challenging OmniNOCS dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 429, + 483, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 483, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 483, + 513 + ], + "type": "text", + "content": "Size and learned PnP heads: We use a size head to predict the 3D size to scale the NOCS predictions to metric object coordinates. Our size head also uses self-attention blocks. Contrary to previous work [7,28,49], it does not use any class-specific layers or per-class average size statistics, as these limit scalability to larger and diverse vocabularies. Our learned PnP head is inspired from [32,48] and uses convolutional layers to predict the 3D orientation and a 2D projection of the 3D centroid. The orientation is predicted using the 6D partial rotation matrix representation [57] in allocentric space [7,28]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 314, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 314, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 314, + 540 + ], + "type": "text", + "content": "4.2 Localization from NOCS predictions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "content": "To localize the object in 3D, we first scale the predicted NOCS using the predicted size to obtain unnormalized metric object coordinates. NOCSformer's learned PnP head predicts the object orientation in allocentric space, which is converted to the egocentric frame using the predicted projection of the 3D centroid. Using the unnormalized 3D object coordinates, object orientation, and the 3D centroid ray, the object depth is estimated from the corresponding 2D points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "Alternatively, the 6DoF pose for the object can be solved directly (without using the learned PnP head) by solving a PnP problem using efficient solvers [30]. However, the orientations so obtained are known to be more sensitive to errors in NOCS coordinates, and less robust when compared to learned methods [32, 48]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 186, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 186, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 186, + 126 + ], + "type": "text", + "content": "4.3 Losses" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 197 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 197 + ], + "type": "text", + "content": "When training NOCSformer, we supervise the NOCS, mask, 3D size, orientation and centroid predictions using appropriate losses. We supervise predicted masks using a simple L2 loss with respect to the ground truth instance mask. Since our NOCS head uses binned prediction where the final NOCS value is a softmax over non-overlapping bin logits, we use a combination of cross-entropy and regression losses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 210, + 208, + 403, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 208, + 403, + 236 + ], + "spans": [ + { + "bbox": [ + 210, + 208, + 403, + 236 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m a s k}} = \\left\\| \\operatorname {m a s k} _ {\\text {p r e d}} - \\operatorname {m a s k} _ {\\text {g t}} \\right\\| _ {2} \\\\ \\mathcal {L} _ {N O C S} = \\operatorname {s o f t m a x} \\mathrm {C E} (\\hat {\\mathbf {n}}, \\hat {\\mathbf {n}} _ {g t}) + | | \\mathbf {n} - \\mathbf {n} _ {g t} | | _ {1} \\\\ \\end{array}", + "image_path": "d0ff83d67f12ed49ff5e1c98be5cd9283c805e52e290e37a32ebc871803d249b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 244, + 481, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 481, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 244, + 481, + 268 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 130, + 244, + 481, + 268 + ], + "type": "text", + "content": " are the predicted logits over the discretized bins and " + }, + { + "bbox": [ + 130, + 244, + 481, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 244, + 481, + 268 + ], + "type": "text", + "content": " the continuous NOCS coordinate prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 269, + 481, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 481, + 340 + ], + "type": "text", + "content": "In addition to the supervised loss above, we also use a variant of the self-supervised NOCS loss " + }, + { + "bbox": [ + 130, + 269, + 481, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ss}" + }, + { + "bbox": [ + 130, + 269, + 481, + 340 + ], + "type": "text", + "content": " from [11] that minimizes the reprojection error of the predicted NOCS using the ground truth pose and predicted mask. More details about " + }, + { + "bbox": [ + 130, + 269, + 481, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ss}" + }, + { + "bbox": [ + 130, + 269, + 481, + 340 + ], + "type": "text", + "content": " are in the supplementary material. The NOCS and mask losses are applied to the fixed resolution grid predictions. Since our NOCS ground truth can be sparse, it is applied at only those locations on the grid that have valid NOCS ground truth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 341, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 481, + 399 + ], + "type": "text", + "content": "Since our 3D size head also uses binned prediction, we use a combination of softmax cross-entropy and L1 loss for supervision. Note that our size loss is normalized by the ground truth size, in order to penalize errors on smaller objects equally. For the learned PnP head, we also supervise the centroid and rotation predictions using L1 losses in their output space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 413, + 399, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 413, + 399, + 468 + ], + "spans": [ + { + "bbox": [ + 216, + 413, + 399, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s i z e}} = \\operatorname {s o f t m a x C E} (\\hat {\\mathbf {s}}, \\hat {\\mathbf {s}} _ {g t}) + | \\mathbf {s} - \\mathbf {s} _ {\\mathbf {g} t} | / \\mathbf {s} _ {\\mathbf {g} t} \\\\ \\mathcal {L} _ {r o t} = \\left\\| ^ {c} \\mathbf {R} _ {o _ {p r e d}} - ^ {c} \\mathbf {R} _ {o _ {g t}} \\right\\| _ {1, 1} \\\\ \\mathcal {L} _ {\\text {c e n t r o i d}} = \\left\\| \\mathbf {c} _ {o} - \\mathbf {c} _ {o g t} \\right\\| _ {2} \\\\ \\mathcal {L} _ {P n P} = \\mathcal {L} _ {r o t} + \\mathcal {L} _ {c e n t r o i d} \\\\ \\end{array}", + "image_path": "b1e61f4f37be9908ff5eaad735a7c193473bc9eba3fe592f39f7187ac6649132.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 479, + 384, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 479, + 384, + 491 + ], + "spans": [ + { + "bbox": [ + 146, + 479, + 384, + 491 + ], + "type": "text", + "content": "The total loss for training NOCSformer is a weighted sum:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 501, + 472, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 501, + 472, + 514 + ], + "spans": [ + { + "bbox": [ + 140, + 501, + 472, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = w _ {\\text {s i z e}} \\mathcal {L} _ {\\text {s i z e}} + w _ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + w _ {\\text {N O C S}} \\mathcal {L} _ {\\text {N O C S}} + w _ {s s} \\mathcal {L} _ {s s} + w _ {P n P} \\mathcal {L} _ {P n P}", + "image_path": "743c1135b03627a9b00cf4419d05e46525e617e2d9fe068c88b61045903c71d4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 531, + 218, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 218, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 218, + 544 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 556, + 482, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 653 + ], + "type": "text", + "content": "Although previous works predict NOCS accurately on a few categories [32, 48, 49], they only evaluate on 3D detection (localization) or pose estimation tasks, without quantifying the accuracy of the predicted NOCS. In section 5.1, we propose metrics and establish a benchmark to evaluate NOCS on the OmniNOCS dataset. We compare the localization accuracy of NOCSformer against existing benchmarks, by evaluating its localization accuracy on nuScenes [10] in section 5.2. We also evaluate the unique ability of NOCSformer to transfer to unseen datasets and domains. Finally, we provide ablations on critical design choices made in our model in section 5.4." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 137, + 249, + 212 + ], + "blocks": [ + { + "bbox": [ + 156, + 123, + 233, + 133 + ], + "lines": [ + { + "bbox": [ + 156, + 123, + 233, + 133 + ], + "spans": [ + { + "bbox": [ + 156, + 123, + 233, + 133 + ], + "type": "text", + "content": "Input: Image + 2D box" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 137, + 249, + 212 + ], + "lines": [ + { + "bbox": [ + 137, + 137, + 249, + 212 + ], + "spans": [ + { + "bbox": [ + 137, + 137, + 249, + 212 + ], + "type": "image", + "image_path": "2aa0ca2f97ae637fa30545efd1f1ae8588bb1f4cb61de62055b8d308fd639897.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 213, + 248, + 286 + ], + "blocks": [ + { + "bbox": [ + 137, + 213, + 248, + 286 + ], + "lines": [ + { + "bbox": [ + 137, + 213, + 248, + 286 + ], + "spans": [ + { + "bbox": [ + 137, + 213, + 248, + 286 + ], + "type": "image", + "image_path": "7b301e035cb9a8d792ee47cfe0d9599f6706dbcc92612ef7f286c08862e669ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 287, + 248, + 346 + ], + "blocks": [ + { + "bbox": [ + 137, + 287, + 248, + 346 + ], + "lines": [ + { + "bbox": [ + 137, + 287, + 248, + 346 + ], + "spans": [ + { + "bbox": [ + 137, + 287, + 248, + 346 + ], + "type": "image", + "image_path": "374cb9ec2555543a715f6dc158b5c9f71b8bf0a1361435068508944cbd3e1f21.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 347, + 248, + 418 + ], + "blocks": [ + { + "bbox": [ + 137, + 347, + 248, + 418 + ], + "lines": [ + { + "bbox": [ + 137, + 347, + 248, + 418 + ], + "spans": [ + { + "bbox": [ + 137, + 347, + 248, + 418 + ], + "type": "image", + "image_path": "c90f8e2d618abcd867d7ddadf1c909e5f57fd75cc105dc447ac07813344f5228.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 419, + 248, + 491 + ], + "blocks": [ + { + "bbox": [ + 137, + 419, + 248, + 491 + ], + "lines": [ + { + "bbox": [ + 137, + 419, + 248, + 491 + ], + "spans": [ + { + "bbox": [ + 137, + 419, + 248, + 491 + ], + "type": "image", + "image_path": "96c2e128ca476ddc8a22a180ed59a9c443c6f8d4ea3609bb8c624b657e935d26.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 137, + 503, + 190, + 571 + ], + "blocks": [ + { + "bbox": [ + 154, + 495, + 173, + 502 + ], + "lines": [ + { + "bbox": [ + 154, + 495, + 173, + 502 + ], + "spans": [ + { + "bbox": [ + 154, + 495, + 173, + 502 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 503, + 190, + 571 + ], + "lines": [ + { + "bbox": [ + 137, + 503, + 190, + 571 + ], + "spans": [ + { + "bbox": [ + 137, + 503, + 190, + 571 + ], + "type": "image", + "image_path": "a75054222dc8a2216e3c6c7b26086911937082e67b44a8ebffcdc7467c9d1cd2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 190, + 503, + 242, + 571 + ], + "blocks": [ + { + "bbox": [ + 204, + 495, + 228, + 502 + ], + "lines": [ + { + "bbox": [ + 204, + 495, + 228, + 502 + ], + "spans": [ + { + "bbox": [ + 204, + 495, + 228, + 502 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 503, + 242, + 571 + ], + "lines": [ + { + "bbox": [ + 190, + 503, + 242, + 571 + ], + "spans": [ + { + "bbox": [ + 190, + 503, + 242, + 571 + ], + "type": "image", + "image_path": "0ef75d46db01758059058fdc5084c36ae8f7da4761f8b7e7b386cd2c7d3f1770.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 257, + 137, + 368, + 212 + ], + "blocks": [ + { + "bbox": [ + 264, + 123, + 348, + 133 + ], + "lines": [ + { + "bbox": [ + 264, + 123, + 348, + 133 + ], + "spans": [ + { + "bbox": [ + 264, + 123, + 348, + 133 + ], + "type": "text", + "content": "Output: NOCS + 3D box" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 137, + 368, + 212 + ], + "lines": [ + { + "bbox": [ + 257, + 137, + 368, + 212 + ], + "spans": [ + { + "bbox": [ + 257, + 137, + 368, + 212 + ], + "type": "image", + "image_path": "7c4c4a5913f8a345de088583bbb8d42af25a19035a8a360d0213197d5bb428f5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 257, + 213, + 368, + 286 + ], + "blocks": [ + { + "bbox": [ + 257, + 213, + 368, + 286 + ], + "lines": [ + { + "bbox": [ + 257, + 213, + 368, + 286 + ], + "spans": [ + { + "bbox": [ + 257, + 213, + 368, + 286 + ], + "type": "image", + "image_path": "7fe99a872f668d3e02d09b6a612fa8fa521921e1e8846569e8b5053137341184.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 257, + 287, + 368, + 346 + ], + "blocks": [ + { + "bbox": [ + 257, + 287, + 368, + 346 + ], + "lines": [ + { + "bbox": [ + 257, + 287, + 368, + 346 + ], + "spans": [ + { + "bbox": [ + 257, + 287, + 368, + 346 + ], + "type": "image", + "image_path": "a7d3b1bc3efc9aa5aff2ea352c5a4e219af6004624e4599109ad87551ed346fe.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 257, + 347, + 368, + 418 + ], + "blocks": [ + { + "bbox": [ + 257, + 347, + 368, + 418 + ], + "lines": [ + { + "bbox": [ + 257, + 347, + 368, + 418 + ], + "spans": [ + { + "bbox": [ + 257, + 347, + 368, + 418 + ], + "type": "image", + "image_path": "d1e6c396d4789631aed34fd84b72253a13bbc7c4726a8b8a8d149548ebd6d902.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 257, + 419, + 368, + 491 + ], + "blocks": [ + { + "bbox": [ + 257, + 419, + 368, + 491 + ], + "lines": [ + { + "bbox": [ + 257, + 419, + 368, + 491 + ], + "spans": [ + { + "bbox": [ + 257, + 419, + 368, + 491 + ], + "type": "image", + "image_path": "c93c50b71f3a0679cbee0d98cd5ddb9b7d40baa192daae55e87c6277130e90f3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 246, + 504, + 316, + 564 + ], + "blocks": [ + { + "bbox": [ + 247, + 495, + 306, + 503 + ], + "lines": [ + { + "bbox": [ + 247, + 495, + 306, + 503 + ], + "spans": [ + { + "bbox": [ + 247, + 495, + 306, + 503 + ], + "type": "text", + "content": "3D shape + pose" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 504, + 316, + 564 + ], + "lines": [ + { + "bbox": [ + 246, + 504, + 316, + 564 + ], + "spans": [ + { + "bbox": [ + 246, + 504, + 316, + 564 + ], + "type": "image", + "image_path": "edaef087919be95f65a5f8c5ffbe06550f0d08f4af2413dd3ce1bfa071c9615f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 327, + 503, + 372, + 571 + ], + "blocks": [ + { + "bbox": [ + 340, + 495, + 358, + 502 + ], + "lines": [ + { + "bbox": [ + 340, + 495, + 358, + 502 + ], + "spans": [ + { + "bbox": [ + 340, + 495, + 358, + 502 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 327, + 503, + 372, + 571 + ], + "lines": [ + { + "bbox": [ + 327, + 503, + 372, + 571 + ], + "spans": [ + { + "bbox": [ + 327, + 503, + 372, + 571 + ], + "type": "image", + "image_path": "135fba02399742b52f7bf8d47d37cdc60ac04fdf5c906eeef00a8a3e17f2c426.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 586, + 482, + 653 + ], + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 653 + ], + "type": "text", + "content": "Fig. 4: Example results of our single unified NOCSformer model across various datasets and object classes. The left column shows input images and query 2D bounding boxes. The center column shows the NOCS+instance maps predicted by NOCSformer along with the estimated 3D pose overlaid on the input image. The NOCS can be used with the 3D boxes and object size to lift the objects to a 3D pointcloud, which is shown in the right column. Last row contains two examples." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 372, + 137, + 478, + 286 + ], + "blocks": [ + { + "bbox": [ + 375, + 123, + 477, + 133 + ], + "lines": [ + { + "bbox": [ + 375, + 123, + 477, + 133 + ], + "spans": [ + { + "bbox": [ + 375, + 123, + 477, + 133 + ], + "type": "text", + "content": "Output: 3D shapes and poses" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 137, + 478, + 286 + ], + "lines": [ + { + "bbox": [ + 372, + 137, + 478, + 286 + ], + "spans": [ + { + "bbox": [ + 372, + 137, + 478, + 286 + ], + "type": "image", + "image_path": "3a6dd6f3997c50cc674624ac3f3d733e29435e73516117f81e5c2b75dd73fed8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 376, + 294, + 476, + 349 + ], + "blocks": [ + { + "bbox": [ + 376, + 294, + 476, + 349 + ], + "lines": [ + { + "bbox": [ + 376, + 294, + 476, + 349 + ], + "spans": [ + { + "bbox": [ + 376, + 294, + 476, + 349 + ], + "type": "image", + "image_path": "097c300d06565cea7c36582b922499faea658fd528ff385a152f90fca82cbed5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 379, + 349, + 477, + 414 + ], + "blocks": [ + { + "bbox": [ + 379, + 349, + 477, + 414 + ], + "lines": [ + { + "bbox": [ + 379, + 349, + 477, + 414 + ], + "spans": [ + { + "bbox": [ + 379, + 349, + 477, + 414 + ], + "type": "image", + "image_path": "28e41a371836d1d79a9b4750352f22b011c92bc830eb73c3691e3ae75ee379db.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 384, + 421, + 466, + 487 + ], + "blocks": [ + { + "bbox": [ + 384, + 421, + 466, + 487 + ], + "lines": [ + { + "bbox": [ + 384, + 421, + 466, + 487 + ], + "spans": [ + { + "bbox": [ + 384, + 421, + 466, + 487 + ], + "type": "image", + "image_path": "1cbaa09de46cdf437354ebc4676ed9502833e92d77b062bdf21cac896a6c73b9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 373, + 503, + 418, + 571 + ], + "blocks": [ + { + "bbox": [ + 386, + 495, + 410, + 502 + ], + "lines": [ + { + "bbox": [ + 386, + 495, + 410, + 502 + ], + "spans": [ + { + "bbox": [ + 386, + 495, + 410, + 502 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 373, + 503, + 418, + 571 + ], + "lines": [ + { + "bbox": [ + 373, + 503, + 418, + 571 + ], + "spans": [ + { + "bbox": [ + 373, + 503, + 418, + 571 + ], + "type": "image", + "image_path": "22cc7db8e1952ec5bf580d17ccce4c8628cc0343d2d6594597231b55fd5e1863.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 495, + 478, + 502 + ], + "lines": [ + { + "bbox": [ + 419, + 495, + 478, + 502 + ], + "spans": [ + { + "bbox": [ + 419, + 495, + 478, + 502 + ], + "type": "text", + "content": "3D shape + pose" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 421, + 507, + 476, + 561 + ], + "blocks": [ + { + "bbox": [ + 421, + 507, + 476, + 561 + ], + "lines": [ + { + "bbox": [ + 421, + 507, + 476, + 561 + ], + "spans": [ + { + "bbox": [ + 421, + 507, + 476, + 561 + ], + "type": "image", + "image_path": "7e24f158250b514172e978841e99570edd2f3e7d1ccdce649cdf5224a056b9d3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Datasets: For our experiments, we train NOCSformer on the OmniNOCS dataset containing 97 classes, holding out images from the NOCS-Real275 dataset. We use NOCS-Real275 to evaluate NOCSformer's cross-dataset generalization capabilities by not training on it. NOCS-Real275 has 6 classes that overlap with the rest of OmniNOCS, although they are not the top 20 classes. It also differs from the rest of OmniNOCS in terms of camera parameters and the context of objects in the scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 272, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 272, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 272, + 217 + ], + "type": "text", + "content": "5.1 NOCS prediction accuracy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 297 + ], + "type": "text", + "content": "We propose the use of NOCS mAE and NOCS mPSNR to evaluate the accuracy of predicted NOCS. For each object the mean Absolute Error (mAE) and mPSNR are computed for all points in the intersection of ground truth and predicted masks, and reported as a mean of per-category means. However, the NOCS mAE/mPSNR do not penalize undersegmentations / sparse predictions. Since we also would like the NOCS predictions to span the full visible instance, we also evaluate the 2D mask mIoU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 370 + ], + "type": "text", + "content": "We evaluate NOCSformer on a subset of OmniNOCS that contains 75 classes with accurate ground truth NOCS and masks, see Table 3. It is able to predict NOCS with errors less than " + }, + { + "bbox": [ + 130, + 297, + 482, + 370 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 370 + ], + "type": "text", + "content": " (0.089 mAE on OmniNOCS). We also evaluate on the held-out NOCS-Real275 dataset, which was not used to train NOCSformer. We find that the zero-shot NOCSformer outperforms the NOCS baseline from [49] that was fully trained on this dataset, a strong indication of the generalization capabilities of NOCSformer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 387, + 262, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 262, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 262, + 399 + ], + "type": "text", + "content": "5.2 3D localization accuracy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 481, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 454 + ], + "type": "text", + "content": "NOCS provides dense 3D-2D correspondences that can be used to estimate the 3D object oriented bounding box. This is done by using the predicted object size and solving a PnP problem, as explained in Sec. 4.2. We evaluate the accuracy of our estimated bounding boxes in different settings:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "content": "Outdoor scenes: We compare NOCSformer's 3D localization accuracy to that of other 3D detection models using the nuScenes true positive localization metrics on the challenging nuScenes dataset (included in OmniNOCS). We compare to Cube R-CNN, since it is the only other model that generalizes across diverse datasets. However, Cube R-CNN has two notable differences to NOCSformer: 1) being a detection model, it also jointly detects 2D object regions, and 2) it uses Chamfer loss causing it to have high orientation errors. As a more comparable baseline, we use a variant of our model (termed \"Cubeformer\"), by replacing our NOCS head with the cube head from Cube R-CNN" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 134, + 571, + 479, + 611 + ], + "blocks": [ + { + "bbox": [ + 134, + 571, + 479, + 611 + ], + "lines": [ + { + "bbox": [ + 134, + 571, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 571, + 479, + 611 + ], + "type": "table", + "html": "
MethodNOCS-Real275OmniNOCS
NOCS MAE↓NOCS PSNR↑Mask IoU↑NOCS MAE↓NOCS PSNR↑Mask IoU↑
NOCS baseline [49]0.12116.34586.10---
NOCSformer0.10718.52789.030.09420.24578.50
", + "image_path": "b77edfab2528231a68f4af343c465b429d28c40490ef491c9b373dfe85d7054a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 619, + 482, + 663 + ], + "lines": [ + { + "bbox": [ + 130, + 619, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 482, + 663 + ], + "type": "text", + "content": "Table 3: NOCS and mask prediction evaluation: We report metrics for NOCSformer on 75 classes of OmniNOCS: it is the first model that is capable on predicting NOCS on such diverse data. Additionally, NOCSformer outperforms [49] which trained on NOCS-Real275, without training on NOCS-Real275." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 113, + 461, + 209 + ], + "blocks": [ + { + "bbox": [ + 150, + 113, + 461, + 209 + ], + "lines": [ + { + "bbox": [ + 150, + 113, + 461, + 209 + ], + "spans": [ + { + "bbox": [ + 150, + 113, + 461, + 209 + ], + "type": "table", + "html": "
MethodMulti-datasetmATE (m)↓mAOE (rad)↓mASE (%)↓mIoU↑
FCOS3D [51]X0.7770.4000.231-
PGD [50]X0.6750.3990.236-
EProPNP [12]X0.6760.3630.263-
EProPNP + TTA [12]X0.6530.3190.255-
Cube R-CNN [7]✓0.6501.3050.2830.349
Cubeformer✓0.7900.5730.3010.280
NOCSformer✓0.8870.5580.2910.377
", + "image_path": "36b4c61456202fa8216ed3b716aa91001338343ab65fd800c04acd774d438344.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": "without a Chamfer loss. It uses the input 2D boxes. More details about Cubeformer are in the supplementary material. From Table 4, we find that NOCSformer's box orientations are canonical and more accurate than both Cube R-CNN and Cubeformer. Moreover, the boxes estimated from NOCSformer's NOCS predictions are comparable in translation and scale errors to those of Cube R-CNN and Cubeformer, even though it does not directly regress depth. While all 3 methods generalize across several classes and domains, baselines trained solely on nuScenes localization (top half) significantly outperform them on this task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 469 + ], + "type": "text", + "content": "Cross-dataset generalization (indoor): Here, we hold out NOCS-Real275 from OmniNOCS when training NOCSformer, and use it to evaluate NOCSformer's ability to generalize to unseen domains. NOCS-Real275 features a tabletop setting with multiple objects, which is semantically different from our other indoor datasets. We compare against Cube R-CNN [7] and the NOCS [49] model that was trained from scratch on this dataset alone. [49] also uses class-specific model parameters and additional losses for symmetric objects that do not scale to larger vocabularies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 542 + ], + "type": "text", + "content": "The results are shown in Table 5. We find that NOCSformer is more accurate at transferring to this unseen dataset compared to Cube R-CNN, indicating the generalizability of NOCS-based localization methods over that of box-regression methods. Note that the mAP metrics used in the NOCS-Real275 datasets may also be affected by the false positives/negatives from Cube R-CNN. However, both models are worse than the supervised NOCS baseline." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 193, + 564, + 418, + 616 + ], + "blocks": [ + { + "bbox": [ + 130, + 215, + 482, + 265 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 265 + ], + "type": "text", + "content": "Table 4: Outdoor localization: Comparison of 3D localization accuracy on the nuScenes subset of OmniNOCS. Top half shows methods that are only trained on nuScenes: while these perform better on nuScenes itself, they do not generalize to other datasets and classes. OmniNOCS is competitive with Cube R-CNN on mIoU while also predicting canonical orientations and object coordinates. Note that mIoU and orientation error only uses the predicted yaw orientation component." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 193, + 564, + 418, + 616 + ], + "lines": [ + { + "bbox": [ + 193, + 564, + 418, + 616 + ], + "spans": [ + { + "bbox": [ + 193, + 564, + 418, + 616 + ], + "type": "table", + "html": "
MethodTransferred?Depth input?mAP
3D IoU @ 253D IoU @ 50
NOCS model [49]X✓79.672.4
CubeRCNN [7]✓X14.94.1
NOCSformer✓X43.510.6
", + "image_path": "b486b9e61039fc93e7a94311ff611443e26d0ab2c993cb545b7e318b1b450033.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 619, + 482, + 664 + ], + "lines": [ + { + "bbox": [ + 130, + 619, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 482, + 664 + ], + "type": "text", + "content": "Table 5: Cross-dataset generalization: Comparison of localization accuracy on NOCS-Real275 for all classes, using mAP at different thresholds. This dataset has been held out when training CubeRCNN and NOCSformer. NOCSformer is able to generalize to NOCS-Real275 without any additional training." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 283 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 283 + ], + "type": "image", + "image_path": "7c23736e5a82817d3573db53649829c2d26c417528908f20903bc18b98a89633.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 350 + ], + "type": "text", + "content": "Fig. 5: Generalization across datasets: NOCSformer can generalize to new datasets that present new camera models and object domains. We show this zero-shot capability of NOCSformer (bottom row) on the NOCS-Real275 test set without training on NOCS-Real275 dataset. The predictions from a NOCS baseline [49] trained explicitly on the NOCS-Real275 dataset are shown in the top row." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 375, + 260, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 260, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 260, + 388 + ], + "type": "text", + "content": "5.3 3D orientation accuracy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 520 + ], + "type": "text", + "content": "A key challenge for 3D localization models trained on multiple datasets [7] is predicting canonical object orientations, as these datasets typically differ in their object pose conventions. OmniNOCS ensures that object poses in all constituent datasets are consistent within a category, enabling our model to predict canonical orientations. We evaluate this by computing the accuracy at different orientation error threshold on KITTI-val [13] in Table 6. We compare the accuracy of our orientation predictions to those of Cube R-CNN [7], which are trained using Chamfer loss, and therefore suffers at predicting canonical orientations. We also compare against our Cubeformer baseline that directly supervises the orientations. We find that orientations estimated using NOCSformer are more consistent and accurate than both our baselines." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 151, + 541, + 460, + 612 + ], + "blocks": [ + { + "bbox": [ + 151, + 541, + 460, + 612 + ], + "lines": [ + { + "bbox": [ + 151, + 541, + 460, + 612 + ], + "spans": [ + { + "bbox": [ + 151, + 541, + 460, + 612 + ], + "type": "table", + "html": "
MethodGravity axisHeading axis (X)
Accuracy @ 1 degAccuracy @ 5 degAccuracy @ 90 degAccuracy 5 degAccuracy @ 10 degAccuracy @ 90 deg
CubeRCNN5.7518.2951.7419.4321.5534.17
Cubeformer80.7998.52100.041.9955.8877.26
NOCSformer84.2499.95100.049.1659.6581.99
", + "image_path": "b3d10bba7163610c13afa5a32194c8dbcabda78678a0e4c05ff2081feb80e99c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 619, + 482, + 653 + ], + "lines": [ + { + "bbox": [ + 130, + 619, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 482, + 653 + ], + "type": "text", + "content": "Table 6: 3D orientation accuracy for models trained on multiple datasets: Models that use Chamfer distance (CubeRCNN) for supervising orientation heads end up being inconsistent in their orientation predictions. The results are averaged over 5 classes in the KITTI-val subset [13]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 207, + 114, + 408, + 169 + ], + "blocks": [ + { + "bbox": [ + 207, + 114, + 408, + 169 + ], + "lines": [ + { + "bbox": [ + 207, + 114, + 408, + 169 + ], + "spans": [ + { + "bbox": [ + 207, + 114, + 408, + 169 + ], + "type": "table", + "html": "
ArchitectureNOCS PSNR↑Mask IoU↑
NOCSformer20.43181.69
Hourglass head-3.064-3.16
MaskRCNN head-5.23-21.27
w/o discretized prediction-4.11-
", + "image_path": "a21901c120e8bd14572d38ec79463acc51f5a7bf2b1ec3f1686f8859056f93f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 184, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 240 + ], + "type": "text", + "content": "Table 7: Comparison of different architecture choices for the NOCS and mask prediction head: We experiment with transformer, hourglass and MaskRCNN architectures. We compare whether discretized prediction of NOCS is better than continuous regression for transformer heads. We quantify the difference in performance when using separate heads for mask and NOCS prediction as opposed to a single head." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 275, + 198, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 198, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 198, + 286 + ], + "type": "text", + "content": "5.4 Ablations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 382 + ], + "type": "text", + "content": "NOCS head architecture: While all existing architectures use a few CNN layers for their NOCS heads [49], we find that this significantly limits the NOCS prediction performance when scaling to more classes. Using our transformer NOCS head provides a 5.23dB improvement in NOCS PSNR and a " + }, + { + "bbox": [ + 130, + 300, + 482, + 382 + ], + "type": "inline_equation", + "content": "21.27\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 382 + ], + "type": "text", + "content": " improvement in mask mIoU. We also experiment with a larger convolutional head: the Hourglass model from [6]. We find that the NOCSformer NOCS heads are even better than Hourglass, with an improvement of 3.06dB on NOCS PSNR and " + }, + { + "bbox": [ + 130, + 300, + 482, + 382 + ], + "type": "inline_equation", + "content": "3.16\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 382 + ], + "type": "text", + "content": " on mask IoU (as shown in Table 7)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 445 + ], + "type": "text", + "content": "Discretized versus continuous prediction of NOCS: NOCSformer models the final NOCS prediction as a classification layer with 50 non-overlapping bins. This was observed to be better than a linear regression layer with a MaskRCNN head in [49]. We find that it is also significantly better when using a transformer head, improving the NOCS prediction PSNR by 4.11 dB, as in Table 7." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 467, + 301, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 301, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 301, + 480 + ], + "type": "text", + "content": "6 Conclusion and Future work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 605 + ], + "type": "text", + "content": "This paper has introduced a new large scale dataset of Normalized Object Coordinates (NOCS) for a wide variety of object classes in indoor and outdoor scenes. It has also proposed a single transformer-based model NOCSformer that can predict object 3D pose, size, and NOCS for any of these objects given 2D bounding box inputs. These allows our model to obtain 3D shape and oriented bounding boxes of objects in metric scale from a single input image. This represents the first attempt to generalize NOCS estimation beyond small datasets of narrow domain, increasing the number of object categories available by an order of magnitude. We hope this provides a means for others to explore large-scale monocular 3D object pose and shape estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Some limitations of our current method include the handling of symmetric objects, where the coordinate system has multiple possible solutions, and reflected geometry, such as left / right shoes. Future work could address these issues, for example, minimizing over multiple coordinate frames in the loss for symmetric objects, and potential estimation of left / right coordinate frames for reflected geometry." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 142, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 142, + 481, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 481, + 175 + ], + "type": "text", + "content": "1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 481, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 481, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 481, + 197 + ], + "type": "text", + "content": "2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: CVPR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 481, + 220 + ], + "type": "text", + "content": "3. Avetisyan, A., Dai, A., Nießner, M.: End-to-end cad model retrieval and 9dof alignment in 3d scans. In: ICCV (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 221, + 481, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 221, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 221, + 481, + 242 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 297 + ], + "type": "text", + "content": "5. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Joffe, B., Kurz, D., Schwartz, A., Shulman, E.: ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1) (2021), https://openreview.net/forum?id=tjZjv_qh_CE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 331 + ], + "type": "text", + "content": "6. Birodkar, V., Lu, Z., Li, S., Rathod, V., Huang, J.: The surprising impact of mask-head architecture on novel class segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7015-7025 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "text", + "content": "7. Brazil, G., Kumar, A., Straub, J., Ravi, N., Johnson, J., Gkioxari, G.: Omni3D: A large benchmark and model for 3D object detection in the wild. In: CVPR. IEEE, Vancouver, Canada (June 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "text", + "content": "8. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9287-9296 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 419 + ], + "type": "text", + "content": "9. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2. arXiv preprint arXiv:2001.10773 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 421, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 421, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 135, + 421, + 481, + 464 + ], + "type": "text", + "content": "10. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 466, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 466, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 135, + 466, + 481, + 498 + ], + "type": "text", + "content": "11. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: Monorun: Monocular 3d object detection by reconstruction and uncertainty propagation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 499, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 499, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 135, + 499, + 481, + 542 + ], + "type": "text", + "content": "12. Chen, H., Wang, P., Wang, F., Tian, W., Xiong, L., Li, H.: Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2781-2790 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 544, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 135, + 544, + 481, + 576 + ], + "type": "text", + "content": "13. Chen, X., Kundu, K., Zhu, Y., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals using stereo imagery for accurate object class detection. IEEE transactions on pattern analysis and machine intelligence 40(5), 1259-1272 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 135, + 577, + 481, + 609 + ], + "type": "text", + "content": "14. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: CVPR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 135, + 610, + 481, + 643 + ], + "type": "text", + "content": "15. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 643, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 481, + 664 + ], + "type": "text", + "content": "16. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 482, + 149 + ], + "type": "text", + "content": "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 482, + 181 + ], + "type": "text", + "content": "18. Fu, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "text", + "content": "19. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "text", + "content": "20. Goodwin, W., Vaze, S., Havoutis, I., Posner, I.: Zero-shot category-level object pose estimation. In: European Conference on Computer Vision. pp. 516-532. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 245 + ], + "type": "text", + "content": "21. Gumeli, C., Dai, A., Nießner, M.: Roca: Robust cad model retrieval and alignment from a single image. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 266 + ], + "type": "text", + "content": "22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 266, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 481, + 310 + ], + "type": "text", + "content": "23. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: Monocinis: Camera independent monocular 3d object detection using instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 923-934 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 310, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 352 + ], + "type": "text", + "content": "24. Irshad, M.Z., Kollar, T., Laskey, M., Stone, K., Kira, Z.: Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In: IEEE International Conference on Robotics and Automation (ICRA) (2022), https://arxiv.org/abs/2203.01929" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "text", + "content": "25. Irshad, M.Z., Zakharov, S., Ambrus, R., Kollar, T., Kira, Z., Gaidon, A.: Shapo: Implicit representations for multi-object shape appearance and pose optimization. In: European Conference on Computer Vision (ECCV) (2022), https://arxiv.org/abs/2207.13691" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 438 + ], + "type": "text", + "content": "27. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 438, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 481, + 471 + ], + "type": "text", + "content": "28. Kundu, A., Li, Y., Rehg, J.M.: 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In: CVPR (2018). https://doi.org/10.1109/CVPR.2018.00375, https://doi.org/10.1109/CVPR.2018.00375" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "type": "text", + "content": "29. Kuo, W., Angelova, A., Lin, T.Y., Dai, A.: Mask2cad: 3d shape prediction by learning to segment and retrieve. In: ECCV (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 492, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 481, + 535 + ], + "type": "text", + "content": "30. Lepetit, V., Moreno-Noguer, F., Fua, P.: Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision 81, 155-166 (2009). https://doi.org/ https://doi.org/10.1007/s11263-008-0152-6, http://infoscience. epfl.ch/record/160138" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "text", + "content": "31. Li, K., DeTone, D., Chen, Y.F.S., Vo, M., Reid, I., Rezatofighi, H., Sweeney, C., Straub, J., Newcombe, R.: Odam: Object detection, association, and mapping using posed rgb video. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "32. Li, Z., Wang, G., Ji, X.: Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7678-7687 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "33. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15641-15650 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "34. Manhardt, F., Kehl, W., Gaidon, A.: Roi-10d: Monocular lifting of 2d detection to 6d pose and metric shape. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2069-2078 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 483, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "35. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Vid2cad: Cad model alignment using multi-view constraints from videos. TPAMI (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 483, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 483, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 483, + 160 + ], + "type": "text", + "content": "36. Maninis, K.K., Popov, S., Nießner, M., Ferrari, V.: Cad-estate: Large-scale cad model annotation in rgb videos. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 483, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 483, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 483, + 182 + ], + "type": "text", + "content": "37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "38. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "39. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "40. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "41. Rad, M., Lepetit, V.: Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In: Proceedings of the IEEE international conference on computer vision. pp. 3828-3836 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "text", + "content": "42. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3D appearance, location & pose. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 335, + 482, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 368 + ], + "type": "text", + "content": "43. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "44. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: International Conference on Computer Vision (ICCV) 2021 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 482, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 434 + ], + "type": "text", + "content": "45. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 567-576 (2015). https://doi.org/10.1109/CVPR.2015.7298655" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 434, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 490 + ], + "type": "text", + "content": "46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 490, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 511 + ], + "type": "text", + "content": "47. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: CVPR (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 544 + ], + "type": "text", + "content": "48. Wang, G., Manhardt, F., Tombari, F., Ji, X.: Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16611-16621 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 544, + 482, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 577 + ], + "type": "text", + "content": "49. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 577, + 482, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 482, + 599 + ], + "type": "text", + "content": "50. Wang, T., Xinge, Z., Pang, J., Lin, D.: Probabilistic and geometric depth: Detecting objects in perspective. In: Conference on Robot Learning. pp. 1475-1485. PMLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "text", + "content": "51. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "52. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180-191. PMLR (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 401, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniNOCS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 259 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "53. Wen, B., Lian, W., Bekris, K., Schaal, S.: Catgrasp: Learning category-level task-relevant grasping in clutter from simulation. ICRA 2022 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "54. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "55. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21919-21928 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "56. Zhang, J., Herrmann, C., Hur, J., Cabrera, L.P., Jampani, V., Sun, D., Yang, M.H.: A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence arXiv preprint arxiv:2305.15347 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "57. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "A. Krishnan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_content_list.json b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7bcc985197a831c843d2c8bfff0539f645ece9b8 --- /dev/null +++ b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_content_list.json @@ -0,0 +1,2377 @@ +[ + { + "type": "text", + "text": "OmniSSR: Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model", + "text_level": 1, + "bbox": [ + 233, + 140, + 767, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runyi Li†, Xuhan Sheng†, Weiqi Li, and Jian Zhang", + "bbox": [ + 305, + 210, + 699, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Electronic and Computer Engineering, Peking University, China {lirunyi, shengxuhan, liweiqi}@stu.pku.edu.cn {zhangjian.sz}@pku.edu.cn", + "bbox": [ + 256, + 239, + 746, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Omnidirectional images (ODIs) are commonly used in real-world visual tasks, and high-resolution ODIs help improve the performance of related visual tasks. Most existing super-resolution methods for ODIs use end-to-end learning strategies, resulting in inferior realness of generated images and a lack of effective out-of-domain generalization capabilities in training methods. Image generation methods represented by diffusion model provide strong priors for visual tasks and have been proven to be effectively applied to image restoration tasks. Leveraging the image priors of the Stable Diffusion (SD) model, we achieve omnidirectional image Super Resolution with both fidelity and realness, dubbed as OmniSSR. Firstly, we transform the equirectangular projection (ERP) images into tangent projection (TP) images, whose distribution approximates the planar image domain. Then, we use SD to iteratively sample initial high-resolution results. At each denoising iteration, we further correct and update the initial results using the proposed Octadecaplex Tangent Information Interaction (OTII) and Gradient Decomposition (GD) technique to ensure better consistency. Finally, the TP images are transformed back to obtain the final high-resolution results. Our method is zero-shot, requiring no training or fine-tuning. Experiments of our method on two benchmark datasets demonstrate the effectiveness of our proposed method.", + "bbox": [ + 264, + 321, + 738, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Omnidirectional Imaging $\\cdot$ Super-Resolution $\\cdot$ Latent Diffusion Model", + "bbox": [ + 264, + 627, + 736, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 218, + 722, + 374, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Omnidirectional images (ODIs) capture the entire scene in all directions, exceeding the narrow field of view (FOV) offered by planar images. Super-Resolution", + "bbox": [ + 217, + 755, + 784, + 786 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contributor $\\boxtimes$ Corresponding author This work was supported in part by National Science Foundation of China (No. 62372016) and Shenzhen Research Project Grant (No. JSGGZD20220822095800001).", + "bbox": [ + 235, + 796, + 784, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(SR) techniques enhance the visual quality of ODIs by increasing their resolution, thereby revealing finer details and enabling more accurate scene analysis and interpretation. This becomes particularly crucial in applications like virtual reality and surveillance, where high-resolution ODIs are essential for precise perception and decision-making.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current research in omnidirectional image super-resolution (ODISR) explores various methodologies to enhance the resolution of ODIs [14,36]. SphereSR [59] addresses non-uniformity in different projections by learning upsampling processes and ensuring information consistency using LIIF [5]. OSRT [60] designs a distortion-aware Transformer to modulate equirectangular projection (ERP) distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods remarkably. However, existing ODISR methods face the following challenges: (1) The majority are end-to-end models that can only produce a deterministic output, always better data fidelity but worse visual perception quality [17]. It's promising to develop a generation-based model, but requiring high data demands, yet high-resolution ODIs are high cost to collect [54,55]. (2) Most methods directly perform SR on ERP format ODIs, while users usually watch ODIs in a narrow FOV using tangent projection (TP). So another promising direction is to use off-the-shelf planar models on TP images. Recent times have witnessed the introduction and widespread application of diffusion models [23,43], especially Stable Diffusion (SD) [38], which have provided a robust backbone for visual tasks [21,24,56,61], including SR [30,40,47,51,52,62]. However, if TP images are trivially one-by-one processed using diffusion-based SR models, they will exhibit discrepancies in the overlapping region when reprojected onto the ERP image. As a result, the global continuity is compromised.", + "bbox": [ + 212, + 223, + 787, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging the strong image prior provided by SD, we propose the first diffusion-based zero-shot method for ODISR, named OmniSSR. Specifically, we propose Octadecaplex Tangent Information Interaction (OTII). OTII entails iterative conversion of intermediate SR results between ERP and TP, bridging the domain gap between ODIs and planar images. Building upon OTII, we further employ an approximate analytical solution of gradient descent, namely as Gradient Decomposition, to guide high-fidelity, high-quality omnidirectional image SR. By capitalizing on SD's effective image prior, our approach strikes a balance between fidelity and realness, ensuring that the restored ODIs exhibit both fidelity to the input data and realistic visual details. This method shows potential for advancing the current state of ODISR, providing improved resolution and visual quality across various applications. Fig. 1 showcases results fully demonstrating the superiority and performance of our proposed methods.", + "bbox": [ + 212, + 529, + 787, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 238, + 727, + 607, + 741 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose OmniSSR, the first zero-shot ODISR method, using an off-the-shelf diffusion-based model, requiring no training or fine-tuning, leveraging existing image generation model priors to solve ODISR task.", + "- To bridge the domain gap between ODIs and planar images, we introduce Octadecaplex Tangent Information Interaction by repeatedly transforming" + ], + "bbox": [ + 225, + 761, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6ac71c55133d5a4b08665715de6b12acbba8f34c2673ad66ddd81cf56ad040db.jpg", + "image_caption": [ + "(a) Input low-resolution ERP image" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 393, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/190d9e5c51da8136830c6232a827c9aa1efaba08f16c4afe093ba374594cf240.jpg", + "image_caption": [ + "Degraded tangent image" + ], + "image_footnote": [], + "bbox": [ + 401, + 145, + 488, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/879565a0658a71bfcbda8b4f2d3b2bc3e444063d1c1fa0ab43f0587f20de7526.jpg", + "image_caption": [ + "Restored by OSRT" + ], + "image_footnote": [], + "bbox": [ + 500, + 145, + 586, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aa214b3782fdb33ef8a9ee5f1a9ea71fae5259094da3e7cdfd81f20ffd36e2d5.jpg", + "image_caption": [ + "Restored by OmniSSR" + ], + "image_footnote": [], + "bbox": [ + 596, + 145, + 684, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92878b1636165f5382bb60c5bd106c1975f4a9828b3623e262c61e3c2b49d295.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 696, + 145, + 784, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aaa341522c4a9f9e98e480a239146ccd5e106e49683f721c8356fdf458d422fa.jpg", + "image_caption": [ + "(b) Input low-resolution ERP image" + ], + "image_footnote": [], + "bbox": [ + 217, + 233, + 392, + 300 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0dd37afa508f60ab7420e295835f7c76128acc59f5e704052d4493526b2ceb94.jpg", + "image_caption": [ + "Degraded tangent image" + ], + "image_footnote": [], + "bbox": [ + 401, + 233, + 488, + 300 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9f30918bea3777a60968a73c3de20b972eed6387a7cfbd63b062475e81a346fe.jpg", + "image_caption": [ + "Restored by OSRT", + "Fig. 1: We address omnidirectional image super-resolution in a zero-shot manner via OmniSSR. Presented above are select outcomes that sketch the essence of OmniSSR compared with current state-of-the-art approach OSRT [60]. Part (a) and (b) illustrate that OmniSSR upholds fidelity and visual realness at the same time, providing vivid and realistic details, while OSRT outputs over-smoothed and distorted results. Zoom in for more details." + ], + "image_footnote": [], + "bbox": [ + 500, + 233, + 586, + 300 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1658a722c4f092baf6dec82180441937b062474937d82f77bc9b31dbb8ed3ecf.jpg", + "image_caption": [ + "Restored by OmniSSR" + ], + "image_footnote": [], + "bbox": [ + 596, + 233, + 684, + 300 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d010a53b7a099e9c3a72d109fd332f0e962bc3679bd46ce293cb2f67d46c5cc.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 696, + 233, + 784, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ODIs between ERP format and TP format, enabling ODISR task with pretrained diffusion models on planar images.", + "bbox": [ + 240, + 436, + 784, + 465 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- By iteratively updating images using the developed Gradient Decomposition technique, we introduce consistency constraints into the sampling process of the latent diffusion model, ensuring a trade-off between fidelity and realness in the reconstructed results.", + "- Extensive experiments are conducted on the benchmark datasets, demonstrating the superior performance of our method over existing state-of-the-art approaches, which validate the effectiveness of OmniSSR." + ], + "bbox": [ + 225, + 467, + 785, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 598, + 385, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Single Image Super-Resolution (SISR)", + "text_level": 1, + "bbox": [ + 214, + 631, + 576, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image super-resolution methods based on deep learning have undergone significant development over an extended period. Currently, they can be broadly classified into two categories of solutions. The first category involves end-to-end network training methods, which utilize image pairs consisting of low-resolution degraded images and high-resolution ground truth images for network training [6-8, 27, 31, 32, 63, 66]. The network architectures employed in this category include CNN [16], Transformers [46], etc. The second category employs image generation models as priors, such as GAN [20], diffusion models [23, 43], etc., where low-resolution images are used as conditions to generate high-resolution images. We will mainly introduce the methods using generative prior.", + "bbox": [ + 212, + 657, + 784, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single Image SR using GAN prior In SR works utilizing GAN pri- ors [3,12,33,37,58], including real-world scenarios [8,49,50,65], pre-trained GAN", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "networks are employed to transform image features into latent space, where the corresponding latent code for the high-resolution image is searched, ultimately yielding the reconstructed high-resolution result.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single Image SR using diffusion prior The diffusion model provides a powerful image prior, and the diffusion sampling process can generate highly realistic images. This strong prior distribution can be applied to various image restoration tasks, including super-resolution [9, 10, 19, 40, 42, 51]. Image-domain diffusion models directly provide prior distributions of image-domain data. DDNM [51] based on the mathematical method of Range-Null space Decomposition, iteratively refines content on the zero space, combining image prior content in the value domain to achieve image restoration. DDRM [25] uses SVD decomposition to obtain restoration results, which is similar to DDNM. DPS [9] transforms the image super-resolution problem into an optimization problem with consistency constraints, using gradient descent algorithms to guide the generation of image-domain diffusion models. GDP [19] further uses such gradient to update the degradation operator to tackle blind image inverse problems. Other methods including MCG [10], DDS [9] and unified control of diffusion generation [19, 42] use same strategy for image restoration, especially image super-resolution.", + "bbox": [ + 212, + 191, + 787, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Omnidirectional Image Super-Resolution", + "text_level": 1, + "bbox": [ + 214, + 452, + 599, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Omnidirectional image super-resolution (ODISR) aims to enhance the resolution of omnidirectional or 360-degree images, which are commonly captured by cameras with a wide field of view. This field has garnered increasing attention due to its applications in virtual reality, omnidirectional video streaming, and surveillance. Several approaches have been proposed to address the unique challenges of ODISR [1, 2, 35, 44]. For instance, Kämäräinen et al. [18] propose a deep learning-based approach for omnidirectional super-resolution, leveraging convolutional neural networks to effectively upscale low-resolution omnidirectional images while preserving spatial details. Similarly, Smolic et al. [36] introduce a novel omnidirectional super-resolution algorithm utilizing generative adversarial networks (GANs) to enhance the visual quality of omnidirectional images by hallucinating high-frequency details.", + "bbox": [ + 212, + 474, + 787, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For evaluation purposes, researchers commonly utilize datasets such as the ODI-SR dataset from LAU-Net [13], and SUN 360 Panorama dataset [53]. These datasets enable the quantitative assessment of ODISR algorithms across various scenarios and facilitate fair comparisons between different methods.", + "bbox": [ + 212, + 655, + 787, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 734, + 330, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first briefly introduce the preliminary background of our method (Sec. 3.1), and give an overall view of our proposed OmniSSR (Sec. 3.2). Then, we discuss the designs of Octadecaplex Tangent Information Interaction, which transform ODIs between ERP and TP formats with pre-upsampling strategy (Sec 3.3), and the Gradient Decomposition correction (Sec. 3.4).", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 146, + 372, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathrm{ERP} \\leftrightarrow \\mathrm{TP}$ Transformation The essence of projection transformations between ERP and TP lie in determining the positions of target image pixels within the source image and computing their corresponding pixel values using interpolation algorithms, as digital images are always stored discretely [28]. Therefore, the $\\mathrm{ERP} \\rightarrow \\mathrm{TP}$ transformation involves locating the TP image pixels on the ERP imaging plane, and vice versa. Gnomonic projection [11] provides the correspondence between ERP image pixels and TP image pixels.", + "bbox": [ + 212, + 169, + 782, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a pixel $P_{e}(x_{e},y_{e})$ within the ERP image, we first find its corresponding pixel $P_{s}(\\theta ,\\phi)$ on the unit sphere using Eq. 1:", + "bbox": [ + 212, + 276, + 784, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = 2 \\pi x _ {e} / W, \\phi = \\pi y _ {e} / H, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 316, + 784, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ and $W$ are the height and width of the ERP image. The Cartesian coordinates of the ERP image and the angular coordinates on the unit sphere exhibit a straightforward one-to-one linear relationship, suggesting a conceptual equivalence between these two projection formats.", + "bbox": [ + 212, + 340, + 782, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the spherical coordinates of the tangent plane center $(\\theta_c, \\phi_c)$ , The transformation from $P_s(\\theta, \\phi)$ to $P_t(x_t, y_t)$ , i.e. $\\mathrm{ERP} \\to \\mathrm{TP}$ , is defined as:", + "bbox": [ + 212, + 402, + 782, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t} = \\left(\\cos (\\phi) \\sin \\left(\\theta - \\theta_ {c}\\right)\\right) / \\zeta , \\tag {2} \\\\ y _ {t} = \\left(\\cos \\left(\\phi_ {c}\\right) \\sin (\\phi) - \\sin \\left(\\phi_ {c}\\right) \\cos (\\phi) \\cos \\left(\\theta - \\theta_ {c}\\right)\\right) / \\zeta , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 441, + 784, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\zeta = \\sin (\\phi_c)\\sin (\\phi) + \\cos (\\phi_c)\\cos (\\phi)\\cos (\\theta -\\theta_c).$", + "bbox": [ + 214, + 488, + 591, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The corresponding inverse transformation, i.e. $\\mathrm{TP} \\rightarrow \\mathrm{ERP}$ , is:", + "bbox": [ + 238, + 503, + 676, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta = \\theta_ {c} + \\arctan \\left(\\left(x _ {t} \\sin (c)\\right) / \\left(\\rho \\cos \\left(\\phi_ {1}\\right) \\cos (c) - y _ {t} \\sin \\left(\\phi_ {c}\\right) \\sin (c)\\right)\\right), \\\\ \\phi = \\arcsin \\left(\\cos (c) \\sin \\left(\\phi_ {c}\\right) + y _ {t} \\sin (c) \\cos \\left(\\phi_ {c}\\right) / \\rho\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 527, + 784, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho = \\sqrt{x_t^2 + y_t^2}$ and $c = \\arctan (\\rho)$ .", + "bbox": [ + 214, + 575, + 501, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With Eq. 2 and Eq. 3, we can build one-to-one forward and inverse mapping functions between pixels on the ERP image and pixels on the TP images. An illustration of the $\\mathrm{ERP} \\to \\mathrm{TP}$ transformation is shown in Fig. 2(a).", + "bbox": [ + 212, + 593, + 782, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Iterative Denoising for Super-Resolution Utilizing the rich image priors provided by SD, we can super-resolve planar images. During initialization, the images are passed through the encoder $\\mathcal{E}$ of SD to obtain latent codes, which are then added to pure noise to generate initial noise $\\mathbf{z}_T$ . Following the approach proposed by StableSR [47], we pass the images through a time-aware adapter $\\mathcal{T}$ . This adapter network structure is similar to the down-sampling part in denoising UNet, taking the image and the time step $t$ of diffusion sampling as inputs to obtain the latent code feature for step $t$ . This feature, along with the latent code $\\mathbf{z}_t$ for each step and the time step $t$ , is then passed through denoising UNet to calculate the denoised result $\\mathbf{z}_{0|t}$ and the latent code $\\mathbf{z}_{t - 1}$ for the next sampling step. By iterating this process $T$ times, we can obtain the final super-resolution result via decoder $\\mathcal{D}$ of SD, yielding high-resolution images.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4fd2290aad0f8599fc02ad333da083137fca0b2a1c9da880711a869ad92e8a6.jpg", + "image_caption": [ + "(a) Equirectangular Projection $\\rightarrow$ Tangent Projection" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 544, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54e8b5067afbd7f4152bc0a06645d53a3b3ac2783204893c1eb94d54f11121e7.jpg", + "image_caption": [ + "(b) Pre-Upsampling Strategy", + "Fig. 2: Details about gnomonic transformations. (a) conversion from ERP to TP. (b) pre-upsampling proposed in Octadecaplex Tangent Information Interaction (Sec. 3.3) mitigating loss during transformation." + ], + "image_footnote": [], + "bbox": [ + 550, + 143, + 784, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Overview", + "text_level": 1, + "bbox": [ + 215, + 367, + 339, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/786d8ad01e6419d6a5b472aa6e445112d61640237c5fccdbbc4090fa2f2ffabb.jpg", + "image_caption": [ + "(b) Detailed illustration of GD Correction", + "Fig. 3: Overview of our proposed OmniSSR. Input low-resolution omnidirectional image $\\mathbf{E}_{init}$ in ERP format is first projected onto Tangent Projection (TP) images $\\mathbf{x}_{init}^{(1)},\\mathbf{x}_{init}^{(2)},\\dots,\\mathbf{x}_{init}^{(m)}$ , then iteratively refined via Stable Diffusion (SD) with a time-aware adapter and controllable feature wrapping (CFW) module. In each step of diffusion sampling, we adopt the Gradient Decomposition (GD) correction technique to introduce consistency constraints for the restored intermediate results. After $T$ steps of sampling, we obtain the final result $\\tilde{\\mathbf{E}}_0$ with high resolution and better visual quality." + ], + "image_footnote": [], + "bbox": [ + 217, + 410, + 787, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach can be divided into three parts. The first part is pre-processing, where we initially up-sample the low-resolution ERP images $\\mathbf{E}_{init}$ with bicubic interpolation to target high-resolution size, then project them onto tangent planes to obtain a series of TP images. These TP images are transformed to the latent space by the SD encoder, iteratively processed through denoising UNet and time-aware adapter network, and then decoded to obtain high-resolution TP images. During each denoising step, these TP images are transformed back via inverse transformation to ERP images, employing the Gradient Decomposition correction to ensure consistency constraints in diffusion sampling. After $T$", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ecadd64d7687b53a39fb41e296ffa42a6ef7114d6b0e3f893506ebd26b6a97fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithm 1: OmniSSR PipelineAlgorithm 2: Iterative Denois-ing with GD Correction
Input: Einit, F, F-1, A, A†, E, D, TInput: Einit, F, F-1, A, A†, E, D, T, Output: Latent code {z(1), z(2), ..., z(m)}
Output: SR result E0for t = T to 1 do
1{x(1), x(2), ..., x(m)} = F(Einit)for i = 1 to m do
2for i = 1 to m doεt = εθ(z(t), T(z(i), t), t)
3z(i)init = Ε(x(i))z(0|t) = 1/√αt(z(t)i - εt√1-αt)
4ε(i) ∼ N(0, I)x(0|i) = D(z(i))
5zT(i) = √αTz(i) + √1-αTε(i)end
6endE0|t = F-1{x(1), x(2), ..., x(m)}
7Get{z(0), z(2), ..., z(m)} from Algo. 2E0|t = E0|t + γeA†(Einit - AE0|t)
8for i = 1 to m do{x(1), x(2), ..., x(m)} = F(E0|t)
9x(0)i = D(z(0))for i = 1 to m do
10endz(0|i) = (1-γl)z(0|i) + γlE(x(i))
11E0 = F-1{x(0), x(2), ..., x(m)}z(t-1) ∼ p(z(t-1)|z(t,i), x(0|i))
12E0 = E0 + γpA†(Einit - AE0)end
13return E0
14end
15return {z(0), z(2), ..., z(m)}
", + "bbox": [ + 214, + 146, + 769, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iterations, the final super-resolution result is obtained. A formulaic description for OmniSSR is shown in Algo. 1. Fig. 3 shows the overview of our pipeline.", + "bbox": [ + 212, + 438, + 782, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Octadecaplex Tangent Information Interaction (OTII)", + "text_level": 1, + "bbox": [ + 214, + 489, + 705, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation To apply SD for ODISR, a straightforward way is to perform the $\\mathrm{ERP} \\rightarrow \\mathrm{TP}$ transformation on the input ERP image. Then, each obtained TP image is fed into the SD-based model for SR. Finally, the $\\mathrm{TP} \\rightarrow \\mathrm{ERP}$ transformation yields the ultimate super-resolved ERP image. OmniFusion [28] employs a similar approach for depth estimation. However, this simplistic strategy fractures the inherent global coherence of ODIs, leading to pixel-level discontinuities in the fused ERP images. Moreover, interpolation algorithms cause significant information loss in the original projection transformations, resulting in more blurred images. If applied multiple times, this exacerbates the information loss even further. To mitigate this, a trivial solution is to increase the pixel count (resolution) of the intermediate projection imaging plane. However, excessively high resolutions in TP images can introduce unnecessary computational overhead during the denoising stage and potentially compromise the denoising performance. (see Supplementary Materials for details)", + "bbox": [ + 212, + 515, + 787, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Information Interaction and Pre-upsampling Based on the analysis of the Motivation in 3.3, we propose OTII by alternating the intermediate results between ERP and TP formats at each denoising step, where a single ERP image is represented by 18 TP images. From Sec. 3.1, we can achieve the $\\mathrm{ERP}\\rightarrow \\mathrm{TP}$ transformation (denoted as $\\mathcal{F}(\\cdot)$ ) and the $\\mathrm{TP}\\rightarrow \\mathrm{ERP}$ transformation (denoted as $\\mathcal{F}^{-1}(\\cdot)$ ). Through the $\\mathrm{ERP}\\rightarrow \\mathrm{TP}$ transformation, we can convert distorted ERP", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images into TP images with content distributions that approximate those of planar images. This enables the use of the original SD super-resolution method for planar images. Conversely, through the $\\mathrm{TP} \\rightarrow \\mathrm{ERP}$ transformation, we can fuse information between different TP images holistically, while providing ERP-format input for the subsequent GD Correction in Sec. 3.4. To handle information loss during projection transformation, we further propose to pre-upsample the source image before projection transformations, as shown in Fig. 2(b). Our experiments in Sec. 4.4 demonstrate that this pre-upsampling strategy can significantly mitigate the information loss caused by projection transformations.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Gradient Decomposition (GD) Correction for Fidelity", + "text_level": 1, + "bbox": [ + 214, + 305, + 705, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SD-based methods, as introduced in Sec. 3.1, can perform SR on sliced TP images. However, relying solely on the SR results from SD may lack consistency and fail to accurately preserve the original semantic information and details of the low-resolution image. To enhance the consistency of the SR results from SD, we opt to use convex optimization methods to iteratively refine them. Modeling the SR task as an image inverse problem, the following equation is formulated:", + "bbox": [ + 212, + 330, + 787, + 422 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathbf {A x} + \\mathbf {n}, \\quad \\mathbf {n} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 435, + 785, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ represents the original image, $\\mathbf{y}$ denotes the degraded result, $\\mathbf{A}$ is the degradation operator (e.g., bicubic downsampling for super-resolution), and $\\mathbf{n}$ is random noise. The objective we aim to solve can be expressed as the following convex optimization problem:", + "bbox": [ + 212, + 462, + 787, + 523 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} | | \\mathbf {y} - \\mathbf {A x} | | _ {2} ^ {2} + \\lambda \\mathcal {R} (\\mathbf {x}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 535, + 785, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the first term is the data-fidelity term, ensuring the consistency of image reconstruction, and the second term is the regulation term, ensuring the sparsity of the reconstruction result, thus making the image more realistic. The regularization term can be the 1-norm, Total Variation, etc. The aforementioned convex optimization problem can be solved using a series of algorithms, such as gradient descent, ADMM [4], etc. Considering the trade-off between time and performance, we turn to find a solution based on gradient descent, and provide an approximate analytical solution composed of a fidelity term and a realness term, named \"Gradient Decomposition (GD)\":", + "bbox": [ + 212, + 569, + 787, + 707 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {E}} _ {0 | t} = \\mathbf {E} _ {0 | t} + \\alpha \\nabla_ {\\mathbf {E} _ {0 | t}} \\left\\| \\mathbf {E} _ {i n i t} - \\mathbf {A} \\mathbf {E} _ {0 | t} \\right\\| _ {F} = \\mathbf {E} _ {0 | t} + \\alpha \\times 2 \\left(\\mathbf {A} ^ {\\dagger} \\mathbf {E} _ {i n i t} - \\mathbf {A} ^ {\\dagger} \\mathbf {A} \\mathbf {E} _ {0 | t}\\right) \\\\ = \\mathbf {E} _ {0 | t} + \\gamma \\mathbf {A} ^ {\\dagger} \\left(\\mathbf {E} _ {i n i t} - \\mathbf {A} \\mathbf {E} _ {0 | t}\\right) = \\gamma \\mathbf {A} ^ {\\dagger} \\mathbf {E} _ {i n i t} + (\\mathbf {I} - \\gamma \\mathbf {A} ^ {\\dagger} \\mathbf {A}) \\mathbf {E} _ {0 | t} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 718, + 785, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{A}^{\\dagger}$ denotes pseudo-inverse of degradation operator $\\mathbf{A}$ , $\\mathbf{E}_{init}$ denotes initial low-resolution ERP input, $\\mathbf{E}_{0|t}$ denotes restored result by SD, $\\tilde{\\mathbf{E}}_{0|t}$ denotes corrected result by GD, $\\alpha$ denotes the learning rate of gradient descent, and $\\gamma$", + "bbox": [ + 212, + 771, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "This claim will be further illustrated in subsequent experiments.", + "bbox": [ + 217, + 824, + 661, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "denotes the simplified hyper-parameter which is further tuned using grid search. The final setting of $\\gamma$ on different stages is shown in Sec. 4.1, and the ablation studies of parameter choice are in Sec. 4.4.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This technique could be seen as a step of gradient descent optimization, and the optimized result could be decomposed of (1) $\\gamma \\mathbf{A}^{\\dagger}\\mathbf{E}_{init}$ , which ensures the consistency of the generated result, and (2) $(\\mathbf{I} - \\gamma \\mathbf{A}^{\\dagger}\\mathbf{A})\\mathbf{E}_{0|t}$ , which serves as the iteratively updated generated result to improve its realness; $\\gamma$ is a hyperparameter balancing the data fidelity and visual quality. For a better diversity and generality of the SR process, we expand this solution to latent space, and obtain the denoising result from both denoising UNet and corrected TP images (Algo. 2 line 11). A more detailed understanding of the iterative denoising process and application of GD correction could be referred to Algo. 2.", + "bbox": [ + 212, + 191, + 787, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 352, + 375, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 385, + 455, + 400 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets and Pretrained Models We choose the test set of ODI-SR dataset from LAU-Net [13] and SUN 360 Panorama dataset [53], comprising 97 and 100 omnidirectional images respectively, for experimental evaluation. The ground truth images are of size $1024 \\times 2048$ pixels. In SR methods such as GDP [19] and PSLD [39] for planar images, we partitioned the images into several $256 \\times 256$ patches and performed super-resolution on each patch individually.", + "bbox": [ + 212, + 410, + 785, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For pre-trained models, we adopt from StableSR [47], which provided a SR network for planar images based on SD. This network architecture includes a time-aware adapter, a controllable feature wrapping (CFW) module, and the original SD structure from HuggingFace. All of them are kept untrained in our proposed OmniSSR.", + "bbox": [ + 212, + 501, + 785, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Settings We set diffusion sampling steps to 200, which is the same as StableSR. The steps for other diffusion-based methods are set the same as their default settings (e.g. 1000 steps for PSLD). The degradation for low-resolution ERP images is bicubic down-sampling, and the implementation of its pseudo-inverse can be referred from code of DDRM $[25]^2$ . For choices of hyper-parameter $\\gamma$ in GD correction, we set $\\gamma_p = 1.0$ , $\\gamma_e = 1.0$ , $\\gamma_l = 0.5$ . Our code is developed via PyTorch on NVIDIA 3090Ti GPU.", + "bbox": [ + 212, + 599, + 787, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparison of OmniSSR with diffusion-based methods", + "text_level": 1, + "bbox": [ + 214, + 729, + 715, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the performance of proposed OmniSSR, we compare our method with recent state-of-the-art zero-shot methods for single image SR task: DPS [9], DDRM [25], GDP [19] which are based on the image-domain diffusion model,", + "bbox": [ + 212, + 755, + 785, + 801 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "2 https://github.com/bahjat-kawar/ddrm", + "bbox": [ + 217, + 810, + 500, + 825 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{3}$ Code is at https://github.com/LiRunyi2001/OmniSSR.", + "bbox": [ + 220, + 825, + 602, + 839 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e398a17989a1ff47394648f61d2ffea67cbca161612a8f8259a432c62b9805c2.jpg", + "table_caption": [ + "Table 1: SR results under bicubic downsampling on ODI-SR and SUN 360 Panorama datasets. For tasks not implemented in those papers, we mark N/A in the corresponding results. The best results are shown in bold, and the second best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodScaleODI-SRSUN 360 Panorama
WS-PSNR↑WS-SSIM↑FID↓LPIPS↓WS-PSNR↑WS-SSIM↑FID↓LPIPS↓
Bicubic28.140.834324.000.216428.670.853729.250.1933
DDRM [25]27.900.831712.280.166129.550.867013.100.1426
DPS [9]20.990.6194148.300.524921.440.6598148.830.5175
GDP [19]27.890.815726.560.272428.600.837628.020.2445
PSLD [39]N/AN/AN/AN/AN/AN/AN/AN/A
DiffIR [52]23.770.658357.230.468723.540.677558.060.4658
StableSR [47]22.700.645844.870.303923.300.690743.490.2858
OmniSSR28.570.854013.010.157529.690.878112.990.1459
Bicubic25.430.705950.840.375525.490.722955.990.3656
DDRM [25]25.430.736732.690.320625.830.744332.930.3304
DPS [9]24.750.6594120.740.491121.090.6119175.21430.5541
GDP [19]23.160.669277.430.426023.750.656990.230.4240
PSLD [39]21.720.5498107.990.532921.750.5828141.490.5461
DiffIR [52]24.010.677054.140.436723.900.701450.370.4235
StableSR [47]23.330.657749.950.313523.990.699846.030.3023
OmniSSR25.770.727930.970.297726.010.748134.580.2963
", + "bbox": [ + 220, + 199, + 782, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and PSLD [39], which is based on latent diffusion model. We also choose supervised diffusion-based super-resolution approaches including StableSR [47] and DiffIR [52] for comparison. We conduct experiments on $\\times 2$ and $\\times 4$ SR with ERP bicubic downsampling, on ODI-SR test set and SUN test set. We choose WS-PSNR [45], WS-SSIM [67], FID [22], and LPIPS [64] as the main metrics.", + "bbox": [ + 212, + 416, + 784, + 491 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative results are presented in Tab. 1. With proposed OTII and GD correction, OmniSSR out-performs previous methods in terms of both Fidelity (from WS-PSNR and WS-SSIM) and Realness (from FID, LPIPS), which shows superior performance to existing diffusion-based methods for ODISR tasks on different scales.", + "bbox": [ + 212, + 491, + 785, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative results are shown in Fig. 4 and Fig. 5, which illustrates the visualization of SR results on SUN test set and ODI-SR test set with $\\times 2$ and $\\times 4$ scales, by different methods.", + "bbox": [ + 212, + 566, + 785, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The visual results indicate that our OmniSSR exhibits superior capability for detail recovery compared to other methods, particularly evident in textual elements (e.g., the text \"flapping\" in upper part of Fig. 4), complex objects (e.g., the black desk with a screen in lower part of Fig. 4, patterns above the white door in lower part of Fig. 5), and small-scale objects (e.g., the person and clock behind the desk in upper part of Fig. 5). OmniSSR demonstrates the ability to recover highly detailed and realistic visual effects from TP images.", + "bbox": [ + 212, + 612, + 785, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison with end-to-end supervised methods", + "text_level": 1, + "bbox": [ + 215, + 739, + 666, + 755 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The experiments of comparison in Sec. 4.2 are mainly focused on zero-shot image super-resolution methods, and supervised single image super-resolution methods, where the approaches are not trained or fine-tuned on omnidirectional images. In this part, we will compare OmniSSR to supervised end-to-end methods with end-to-end training on ODI datasets, including SwinIR and OSRT. Besides the", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e4c3ff6c545218f159267cef5882c1603f2c8bf24709a01975de6d37f76fc2e5.jpg", + "image_caption": [ + "$\\therefore m - 1 \\neq 0$ ;" + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 379, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/77002c1d00337de418c645adf178eaa35e56c1a3f5ee83e53bfeb849f3eaed87.jpg", + "image_caption": [ + "HR PSNR/SSIM" + ], + "image_footnote": [], + "bbox": [ + 380, + 143, + 470, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/05650ecbcf953f8c27820e324f67b077efd87da717d44393573f8cf043065397.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 143, + 562, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/118344129e4a2eb4ef49ca66cfce650818a4e189833fe66125991db2b6b1a1e0.jpg", + "image_caption": [ + "DPS [9] 24.02dB/0.5849" + ], + "image_footnote": [], + "bbox": [ + 566, + 143, + 655, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d982945947d1b2335bb0a61a5486f3b5c04193c4644c94c0a4f33c6ff494dbe5.jpg", + "image_caption": [ + "DDRM [25] 30.04dB/0.8855" + ], + "image_footnote": [], + "bbox": [ + 665, + 143, + 753, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/82b8f5727c3feb46d1fab616286acba8ecf4e4d130640730b693e03e94627b25.jpg", + "image_caption": [ + "SUN 360 $(\\times 2)$ : 001" + ], + "image_footnote": [], + "bbox": [ + 228, + 231, + 379, + 297 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0c3a37a737516d7755f59d9cc28cf2ea82c2e766d12fbad042740b8a943120a8.jpg", + "image_caption": [ + "GDP [19] 0.07dB/0.8802" + ], + "image_footnote": [], + "bbox": [ + 380, + 231, + 470, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/51e080788bc5266f80850c4fd2c7d67f12d19c5ba9a9dd08ba45a05525c3fbc1.jpg", + "image_caption": [ + "Bicubic 28.67dB/0.8317", + "DiffIR [52] 24.47dB/0.6421" + ], + "image_footnote": [], + "bbox": [ + 473, + 231, + 562, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d7f3e583f56f782601858dfe22fecbd03b12d177240296fb18e580a05c6757d9.jpg", + "image_caption": [ + "StableSR [47] 23.81dB/0.7384" + ], + "image_footnote": [], + "bbox": [ + 566, + 231, + 655, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/026ddc985ebdab03ea056438df9d09337ff31b83aeb64778b6cc93f49748ddad.jpg", + "image_caption": [ + "OmniSSR(ours) 30.15dB/0.8859" + ], + "image_footnote": [], + "bbox": [ + 663, + 231, + 754, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a22c76f1a9c57cf9ec0cd05b2cf7728d59df730cc8b188461f2ea001da8ec4f1.jpg", + "image_caption": [ + "SUN 360 $(\\times 4)$ : 009", + "GDP [19] 23.02dB/0.7525", + "Fig. 4: Visualized comparison of $\\times 2$ and $\\times 4$ SR results on SUN 360 test set. 001 and 009 are the ID numbers in the test set filenames. We also calculate the PSNR and SSIM to HR ground truth of each SR result and downsampled image." + ], + "image_footnote": [], + "bbox": [ + 228, + 321, + 379, + 474 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ad6731bd33a9a430eda461540342ab99da7f72d821a6de2efa230ce8f2f0bc06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 321, + 470, + 474 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fb91862be9975ce4a1bb80cd177b676301418460af1b0b6581f80b2790e37779.jpg", + "image_caption": [ + "DiffIR [52] 23.48dB/0.7799" + ], + "image_footnote": [], + "bbox": [ + 473, + 321, + 562, + 474 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d8c4ff0b5f831768bbd1ad40dfa23441b9d5b0b105e7250788095ac4b9050705.jpg", + "image_caption": [ + "StableSR [47] 24.62dB/0.7981" + ], + "image_footnote": [], + "bbox": [ + 566, + 321, + 655, + 474 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/90267770031892416bc9530e51251ba52189be4ad3bb981dd57eed42669e6495.jpg", + "image_caption": [ + "OmniSSR(ours) 26.53dB/0.8265" + ], + "image_footnote": [], + "bbox": [ + 663, + 321, + 753, + 474 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "main metrics in Sec. 4.2, we also use NIQE [34] and DISTS [15] to evaluate the visual perception of SR outputs. Results are presented in Tab. 2, which shows that although our OmniSSR exhibits inferior fidelity metrics compared to end-to-end supervised methods trained directly on ODI datasets, it demonstrates notable improvements in the visual quality and authenticity of super-resolved images. Notably, end-to-end methods often produce smoothed reconstructions with distortions, whereas our approach preserves finer details and adheres more closely to the realistic distribution. Considering that our method has never been trained or tuned on ODI datasets, nor having omnidirectional images prior, this result is acceptable.", + "bbox": [ + 212, + 578, + 787, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 753, + 401, + 767 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We first sequentially validate the performance improvement of the proposed strategy in OmniSSR including input image type, OTII and GD correction, on the ODI-SR test set with $\\times 2$ SR task, thereby demonstrating the significance of these strategies. The details are demonstrated as follows:", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0bdc65fb2176dac12f5298b3ca6d8fcb1ec664fb7773e237fbc775869930a5f4.jpg", + "image_caption": [ + "ODI-SR $(\\times 2)$ :067" + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 379, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0752235e31564fb4f752b63de28f5a0ac4c30935b0e53ef6caf8179ebf4bf697.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 380, + 143, + 470, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cf16280fa802fd9d07b9ff4c5314099f7089002bddd8c755718783f03a4ed9e0.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 473, + 143, + 562, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/543e56ef5b919493be7a8272c2ca28f09f569e5d38f7712965492320a2909fd0.jpg", + "image_caption": [ + "DPS [9]" + ], + "image_footnote": [], + "bbox": [ + 566, + 143, + 655, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/527ed727580df771e6de1ebb3b5bef03127e6a1ae23209ef51281ef67497fac8.jpg", + "image_caption": [ + "PSNR/S", + "GDP [19]" + ], + "image_footnote": [], + "bbox": [ + 380, + 232, + 470, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c3b5c7e640014d61d969b992d50e17ef486601a8713b4ea9cf721f5a150659aa.jpg", + "image_caption": [ + "27.67dB/0.8095", + "DiffIR [52]" + ], + "image_footnote": [], + "bbox": [ + 473, + 232, + 562, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fbb72815f1c1de0e2f751e61bdb32e69dd9832622603ed3647c242b650b3f5c7.jpg", + "image_caption": [ + "22.93dB/0.5653", + "StableSR [47]" + ], + "image_footnote": [], + "bbox": [ + 566, + 232, + 655, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/359bdce6a7e8122ad244e0f897c40642adfd7cd00819502c5a6608ac55d62143.jpg", + "image_caption": [ + "DDRM [25]" + ], + "image_footnote": [], + "bbox": [ + 665, + 143, + 753, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a39408904009596fae7d49e6a19b743fb7fff6944807c0196b6c4431195a9f4a.jpg", + "image_caption": [ + "29.91dB/0.8809", + "OmniSSR(ours)" + ], + "image_footnote": [], + "bbox": [ + 663, + 232, + 753, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d9dd5fa49be977880334bf244e60346700d2164088881783a7cfe8c0758cb98c.jpg", + "image_caption": [ + "ODI-SR $(\\times 4)$ :049", + "Fig. 5: Visualized comparison of $\\times 2$ and $\\times 4$ SR results on ODI-SR test set. 067 and 049 are the ID numbers in test set filenames. We also calculate the PSNR and SSIM between ground truth and each SR result as well as the downsampled image." + ], + "image_footnote": [], + "bbox": [ + 228, + 321, + 379, + 474 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/93c9af5a620013e7780e5afa1e28077d1cec55b360e465d761e40a0ee6c62dfb.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 380, + 321, + 470, + 388 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1b960db6d669d53cef5bcf03dc771877d1cca6a1dcbdf7989e6b9516c3c4e692.jpg", + "image_caption": [ + "22.65dB/0.6248", + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 473, + 321, + 562, + 387 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/665b7a80abefad187ae780411eb88428615dd5b47d299a62e7eef6793e7a5976.jpg", + "image_caption": [ + "21.80dB/0.5892", + "DPS [9]" + ], + "image_footnote": [], + "bbox": [ + 566, + 321, + 655, + 387 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "26.21dB/0.7574", + "bbox": [ + 566, + 398, + 656, + 407 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ceaf816b663668d5f27c284215943761f50a52998a3e3458c9f928edb319c082.jpg", + "image_caption": [ + "PSNR/SSIM", + "GDP [19] 24.15dB/0.7179" + ], + "image_footnote": [], + "bbox": [ + 380, + 398, + 470, + 474 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bfa6cc4eb8ac106c9f8c23112c1587b60c9c8ae66410ef9371f606367ada7017.jpg", + "image_caption": [ + "25.44dB/0.7536", + "DiffIR [52]", + ".25dB/0.7594" + ], + "image_footnote": [], + "bbox": [ + 473, + 398, + 562, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "26.21dB/0.7574", + "bbox": [ + 566, + 398, + 656, + 407 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "26.21dB/0.7574", + "bbox": [ + 566, + 398, + 656, + 407 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2b9a5aceeddd7d72e94d00bc0cc4f0520d244efa9dd886d51156ed52d0128903.jpg", + "image_caption": [ + "29.99dB/0.8798", + "DDRM [25]" + ], + "image_footnote": [], + "bbox": [ + 663, + 321, + 753, + 387 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/87f0ff13584b3ff238d34f0281a2ee9ae79b7b540089a9bb21222902908299ec.jpg", + "image_caption": [ + "27.12dB/0.8129", + "OmniSSR(ours)", + "27.20dB/0.8168" + ], + "image_footnote": [], + "bbox": [ + 663, + 407, + 753, + 474 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) we do not use any proposed strategy in the SR task, which is equivalent to the vanilla StableSR baseline;", + "2) we transform the degraded ERP image to TP images and feed them separately into StableSR pipeline, instead of directly inputting ERP images;", + "3) based on 2), we add OTII strategy during the denoising process of SD (Algo. 2 line 7);", + "4) based on 2), we add GD correction at the post-processing stage (Algo. 1 line 12) of the overall pipeline;", + "5) based on 3) and 4), we add GD correction at every step and post-processing stage of sampling, to improve the consistency of the restored result." + ], + "bbox": [ + 212, + 579, + 785, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Note that the execution of GD correction requires the execution of OTII in the denoising process simultaneously, there is no scenario where only GD correction is executed without the execution of OTII in the denoising process.", + "bbox": [ + 212, + 734, + 784, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative results of ablation studies are shown in Tab. 3. From the result shown below, we could come to the claim that the OTII helps improve the performance on the domain level, and the transformation between ERP and TP images provides information fusion among adjacent TP images. Our proposal of", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b43309cf1b7992571640fe32db857034d4d050d51a1962a710fdc2b8eb059731.jpg", + "table_caption": [ + "Table 2: Comparison on $\\times 4$ SR task with supervised methods trained on ODI-SR dataset, including SwinIR and OSRT. The best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
MethodDatasetWS-PSNR↑WS-SSIM↑FID↓LPIPS↓NIQE↓DISTS↓
SwinIR [29]26.760.762027.940.33215.39610.1710
OSRT [60]ODI-SR26.890.764627.390.32585.43640.1695
OmniSSR25.770.727930.970.29775.28910.1541
SwinIR [29]26.020.769239.900.34195.24400.1325
OSRT [60]SUN 36026.330.776639.220.33645.29840.1312
OmniSSR26.010.748134.580.29635.13290.1299
", + "bbox": [ + 272, + 184, + 725, + 272 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/566a66533920dfe81b91ceecd2f59b400f8b11554e6bd4af3b7de3c7fbd087cc.jpg", + "table_caption": [ + "Table 3: Ablation studies of OmniSSR on input type, OTII, and GD correction, on the test set of the ODI-SR dataset. Best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
Input typeOTIIGD CorrectionWS-PSNR↑WS-SSIM↑FID↓LPIPS↓
ERP××22.690.645844.870.3039
TP××23.530.684943.910.3113
TP✓×23.740.684765.350.3748
TP×✓ (in post-process only)26.770.819215.410.1691
TP✓✓28.580.854013.010.1575
", + "bbox": [ + 264, + 325, + 738, + 400 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Gradient Decomposition corrects such restoration result, improving fidelity and realness significantly at the same time, and it would be better if it is applied at each step of the overall denoising pipeline. Tab. 4 shows the effect of mitigating information loss via proposed pre-upsampling strategy.", + "bbox": [ + 212, + 426, + 784, + 488 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4da33366ec8ad894a06bf0a9b601184e2fc449d8ab20e8fa199af3a5e6d11cf6.jpg", + "image_caption": [ + "Fig.6: Ablation of choices on $\\gamma_{p}$ , $\\gamma_{e}$ and $\\gamma_{l}$ . For better readability, WS-PSNR and LPIPS are chosen as evaluation metrics for fidelity and visual quality, respectively, to demonstrate the performance under different choices of the gamma parameter. We illustrate the results of (a) $\\gamma_{p}$ and $\\gamma_{e}$ fixed, while adjusting $\\gamma_{l}$ ; (b) $\\gamma_{e}$ and $\\gamma_{l}$ fixed, while adjusting $\\gamma_{p}$ ; (c) $\\gamma_{p}$ and $\\gamma_{l}$ fixed, while adjusting $\\gamma_{e}$ . It can be observed that when $\\gamma_{p} = 1$ , $\\gamma_{e} = 1$ , and $\\gamma_{l} = 0.5$ , OmniSSR achieves the relatively best performance." + ], + "image_footnote": [], + "bbox": [ + 248, + 515, + 415, + 609 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5ad62970d5261ee1e556964b7162f20ac9220f05687d019e45e9cd5af8a3b221.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 515, + 584, + 609 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/79cf981a76a17afe2ec979ca890f842edaa11cb91a5c60efe71ba9c4ef1a517e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 515, + 754, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For $\\gamma$ in the GD correction technique, we use grid search to obtain better results on ODI-SR dataset and $\\times 4$ SR task. Fig. 6 shows performance on different choices of $\\gamma_{p}$ in Algo. 1 line 12, $\\gamma_{e}$ in Algo. 2 line 8, and $\\gamma_{l}$ in Algo. 2 line 11. The entire ablation of $\\gamma_{p},\\gamma_{e}$ and $\\gamma_{l}$ , with WS-PSNR, WS-SSIM, FID and LPIPS score all calculated and compared, will be provided in Supplementary Materials.", + "bbox": [ + 212, + 734, + 784, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To evaluate the generalizability of our proposed modules, including Pre-Upsampling, OTII, and GD correction, we further conducted ablation studies", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/932fd3dbfab607561b114af73fb44bc830ec27dfcf6f450faf4c6e4a975885f3.jpg", + "table_caption": [ + "Table 4: Results of pre-upsampling strategy on different scales, where $(x,y)$ denotes bicubic-based upsampling at $x\\times$ scale to ERP before $\\mathrm{ERP}\\rightarrow \\mathrm{TP}$ , and $y\\times$ scale to TP before $\\mathrm{TP}\\rightarrow \\mathrm{ERP}$ transformation. Best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
\\( \\mathrm{{ERP}} \\rightarrow \\mathrm{{TP}} \\rightarrow \\mathrm{{ERP}} \\)(1,1)(1,4)(4,1)(4,2)(2,4)(4,4)
WS-PSNR↑28.9838.1128.9933.9138.0538.18
WS-SSIM↑0.88590.98380.88620.96260.98370.9841
", + "bbox": [ + 321, + 199, + 679, + 239 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "on two super-resolution backbones, StableSR and SwinIR. The results underscore substantial performance enhancements facilitated by our modules across both backbones, which is provided in Supplementary Materials.", + "bbox": [ + 212, + 267, + 782, + 313 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitation and Discussion", + "text_level": 1, + "bbox": [ + 215, + 334, + 503, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Although OmniSSR bridges the gap between omnidirectional and planar images, achieving competitive performance and better visual results in ODISR, it still exhibits the following limitations: (1) The inference of the diffusion model requires a considerable amount of time, approximately 14 minutes per ERP-formatted omnidirectional image to be super-resolved into size $1024 \\times 2048$ , making real-time super-resolution challenging; (2) Multiple conversions between ERP and TP are required in the pipeline, leading to improved performance but consuming additional inference time; (3) Further exploration of the convex optimization properties of GD correction is warranted, such as designing gradient term coefficients adaptive to reconstruction results and degradation types.", + "bbox": [ + 212, + 364, + 784, + 515 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This study explores the application of image generation models to ODISR tasks. In future work, the framework behind OmniSSR can be extended beyond the confines of image super-resolution in a single scenario and venture into more complex ODI-based real-world scenarios. These include ODI editing, ODI inpainting, enhancing the quality of 3D Gaussian Splatting scenes [26,41] obtained after super-resolving ERP images, as well as enhancing the quality of omnidirectional videos [48], and other possible diffusion-based applications [57,68].", + "bbox": [ + 212, + 516, + 785, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 643, + 359, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper leverages the image prior of Stable Diffusion (SD) and employs the Octadecaplex Tangent Information Interaction (OTII) to achieve zero-shot omnidirectional image super-resolution. Additionally, we propose the Gradient Decomposition (GD) correction based on convex optimization algorithms to refine the initial super-resolution results, enhancing the fidelity and realness of the restored images. The superior performance of our proposed method, OmniSSR, is demonstrated on benchmark datasets. By bridging the gap between omnidirectional and planar images, we establish a training-free approach, mitigating the data demand and over-fitting associated with end-to-end training. The application scope of our method can be further extended to various applications, presenting potential value across multiple visual tasks.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. An, H., Zhang, X.: Perception-oriented omnidirectional image super-resolution based on transformer network. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2023)", + "2. Arican, Z., Frossard, P.: Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing (TIP) (2011)", + "3. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for image super-resolution and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)", + "4. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging (2016)", + "5. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "6. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision (ICCV) (2023)", + "7. Cheng, M., Ma, H., Ma, Q., Sun, X., Li, W., Zhang, Z., Sheng, X., Zhao, S., Li, J., Zhang, L.: Hybrid transformer and cnn attention network for stereo image superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "8. Chong, M., Yanze, W., Xintao, W., Chao, D., Jian, Z., Ying, S.: Metric learning based interactive modulation for real-world super-resolution. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)", + "9. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022)", + "0. Chung, H., Sim, B., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "1. Coxeter, H.S.M.: Introduction to geometry. John Wiley & Sons, Inc. (1961)", + "2. Daras, G., Dean, J., Jalal, A., Dimakis, A.: Intermediate layer optimization for inverse problems using deep generative models. In: Proceedings of the International Conference on Machine Learning (ICML) (2021)", + "3. Deng, X., Wang, H., Xu, M., Guo, Y., Song, Y., Yang, L.: Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "4. Deng, X., Wang, H., Xu, M., Li, L., Wang, Z.: Omnidirectional image superresolution via latitude adaptive network. IEEE Transactions on Multimedia (TMM) (2022)", + "5. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)", + "6. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2015)", + "7. Duan, H., Zhai, G., Min, X., Zhu, Y., Fang, Y., Yang, X.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the IEEE International Symposium on Circuits and Systems (ISCAS) (2018)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Fakour-Sevom, V., Guldogan, E., Kämäräinen, J.K.: 360 panorama superresolution using deep convolutional networks. In: Proceedings of the Int. Conf. on Computer Vision Theory and Applications (VISAPP) (2018)", + "19. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2014)", + "21. Guo, L., Tao, T., Cai, X., Zhu, Z., Huang, J., Zhu, L., Gu, Z., Tang, H., Zhou, R., Han, S., et al.: Cas-diffcom: Cascaded diffusion model for infant longitudinal super-resolution 3d medical image completion. arXiv preprint arXiv:2402.13776 (2024)", + "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017)", + "23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "24. Jiang, Y., Chan, K.C., Wang, X., Loy, C.C., Liu, Z.: Reference-based image and video super-resolution via $c^2$ -matching. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)", + "25. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of the ICLR Workshop on Deep Generative Models for Highly Structured Data (ICLRW) (2022)", + "26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) (2023)", + "27. Li, W., Chen, B., Zhang, J.: D3c2-net: Dual-domain deep convolutional coding network for compressive sensing. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) (2024)", + "28. Li, Y., Guo, Y., Yan, Z., Huang, X., Duan, Y., Ren, L.: Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "29. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (2021)", + "30. Liu, J., Wang, Q., Fan, H., Wang, Y., Tang, Y., Qu, L.: Residual denoising diffusion models. arXiv preprint arXiv:2308.13712 (2023)", + "31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2022)", + "32. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2020)", + "33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "34. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters (SPL) (2013)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Nishiyama, A., Ikehata, S., Aizawa, K.: $360^{\\circ}$ single image super resolution via distortion-aware network and distorted perspective images. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2021)", + "36. Ozcinar, C., Rana, A., Smolic, A.: Super-resolution of omnidirectional images using adversarial learning. In: Proceedings of the IEEE International Workshop on Multimedia Signal Processing (MMSPW) (2019)", + "37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)", + "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "39. Rout, L., Raoof, N., Daras, G., Caramanis, C., Dimakis, A., Shakkottai, S.: Solving linear inverse problems provably via posterior sampling with latent diffusion models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "40. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)", + "41. Schonbein, M., Geiger, A.: Omnidirectional 3d reconstruction in augmented manhattan worlds. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014)", + "42. Song, J., Zhang, Q., Yin, H., Mardani, M., Liu, M.Y., Kautz, J., Chen, Y., Vahdat, A.: Loss-guided diffusion models for plug-and-play controllable generation. In: Proceedings of the International Conference on Machine Learning (ICML) (2023)", + "43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Proceedings of the International Conference on Learning Representations (ICLR) (2020)", + "44. Sun, X., Li, W., Zhang, Z., Ma, Q., Sheng, X., Cheng, M., Ma, H., Zhao, S., Zhang, J., Li, J., et al.: Opdn: Omnidirectional position-aware deformable network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2023)", + "45. Sun, Y., Lu, A., Yu, L.: Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters (SPL) (2017)", + "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017)", + "47. Wang, J., Yue, Z., Zhou, S., Chan, K., Loy, C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)", + "48. Wang, Q., Li, W., Mou, C., Cheng, X., Zhang, J.: 360dvd: Controllable panorama video generation with 360-degree video diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "49. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "50. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European Conference on Computer Vision Workshops (ECCVW) (2018)", + "51. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "53. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2012)", + "54. Yagi, Y.: Omnidirectional sensing and its applications. IEICE Transactions on Information and Systems (TOIS) (1999)", + "55. Yamazawa, K., Yagi, Y., Yachida, M.: Omnidirectional imaging with hyperboloidal projection. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (1993)", + "56. Yang, S., Zhou, Y., Liu, Z., Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: Proceedings of the SIGGRAPH Asia 2023 Conference Papers (SIGGRAPH Asia) (2023)", + "57. Yang, S., Mou, C., Yu, J., Wang, Y., Meng, X., Zhang, J.: Neural video fields editing. arXiv preprint arXiv:2312.08882 (2023)", + "58. Yinhuai, W., Yujie, H., Jiwen, Y., Jian, Z.: Gan prior based null-space learning for consistent super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023)", + "59. Yoon, Y., Chung, I., Wang, L., Yoon, K.J.: Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "60. Yu, F., Wang, X., Cao, M., Li, G., Shan, Y., Dong, C.: Osrt: Omnidirectional image super-resolution with distortion-aware transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "61. Yu, J., Zhang, X., Xu, Y., Zhang, J.: Cross: Diffusion model makes controllable, robust and secure image steganography. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "62. Yue, Z., Wang, J., Loy, C.C.: Resshift: Efficient diffusion model for image superresolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "63. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "64. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "65. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "66. Zhang, X., Zhang, Y., Xiong, R., Sun, Q., Zhang, J.: Herosnet: Hyperspectral explicable reconstruction and optimal sampling deep network for snapshot compressive imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "67. Zhou, Y., Yu, M., Ma, H., Shao, H., Jiang, G.: Weighted-to-spherically-uniform sism objective quality evaluation for panoramic video. In: Proceedings of the IEEE International Conference on Signal Processing (ICSP) (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + }, + { + "type": "ref_text", + "text": "68. Zongying, L., Hao, L., Liuzhenghao, L., Bin, L., Junwu, Z., Yu-Chian, C.C., Li, Y., Yonghong, T.: Taxdiff: Taxonomic-guided diffusion model for protein sequence generation. arXiv preprint arXiv:2402.17156 (2024)", + "bbox": [ + 217, + 146, + 787, + 189 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "OmniSSR", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_model.json b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..203aa2f992331669f8e8704d5ba50b4c42ed3394 --- /dev/null +++ b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_model.json @@ -0,0 +1,3659 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.768, + 0.186 + ], + "angle": 0, + "content": "OmniSSR: Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.212, + 0.7, + 0.229 + ], + "angle": 0, + "content": "Runyi Li†, Xuhan Sheng†, Weiqi Li, and Jian Zhang" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.24, + 0.747, + 0.282 + ], + "angle": 0, + "content": "School of Electronic and Computer Engineering, Peking University, China {lirunyi, shengxuhan, liweiqi}@stu.pku.edu.cn {zhangjian.sz}@pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.322, + 0.74, + 0.612 + ], + "angle": 0, + "content": "Abstract. Omnidirectional images (ODIs) are commonly used in real-world visual tasks, and high-resolution ODIs help improve the performance of related visual tasks. Most existing super-resolution methods for ODIs use end-to-end learning strategies, resulting in inferior realness of generated images and a lack of effective out-of-domain generalization capabilities in training methods. Image generation methods represented by diffusion model provide strong priors for visual tasks and have been proven to be effectively applied to image restoration tasks. Leveraging the image priors of the Stable Diffusion (SD) model, we achieve omnidirectional image Super Resolution with both fidelity and realness, dubbed as OmniSSR. Firstly, we transform the equirectangular projection (ERP) images into tangent projection (TP) images, whose distribution approximates the planar image domain. Then, we use SD to iteratively sample initial high-resolution results. At each denoising iteration, we further correct and update the initial results using the proposed Octadecaplex Tangent Information Interaction (OTII) and Gradient Decomposition (GD) technique to ensure better consistency. Finally, the TP images are transformed back to obtain the final high-resolution results. Our method is zero-shot, requiring no training or fine-tuning. Experiments of our method on two benchmark datasets demonstrate the effectiveness of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.628, + 0.738, + 0.654 + ], + "angle": 0, + "content": "Keywords: Omnidirectional Imaging \\(\\cdot\\) Super-Resolution \\(\\cdot\\) Latent Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.723, + 0.375, + 0.739 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.756, + 0.785, + 0.787 + ], + "angle": 0, + "content": "Omnidirectional images (ODIs) capture the entire scene in all directions, exceeding the narrow field of view (FOV) offered by planar images. Super-Resolution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.236, + 0.797, + 0.785, + 0.84 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contributor \\(\\boxtimes\\) Corresponding author This work was supported in part by National Science Foundation of China (No. 62372016) and Shenzhen Research Project Grant (No. JSGGZD20220822095800001)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "(SR) techniques enhance the visual quality of ODIs by increasing their resolution, thereby revealing finer details and enabling more accurate scene analysis and interpretation. This becomes particularly crucial in applications like virtual reality and surveillance, where high-resolution ODIs are essential for precise perception and decision-making." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.788, + 0.527 + ], + "angle": 0, + "content": "Current research in omnidirectional image super-resolution (ODISR) explores various methodologies to enhance the resolution of ODIs [14,36]. SphereSR [59] addresses non-uniformity in different projections by learning upsampling processes and ensuring information consistency using LIIF [5]. OSRT [60] designs a distortion-aware Transformer to modulate equirectangular projection (ERP) distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods remarkably. However, existing ODISR methods face the following challenges: (1) The majority are end-to-end models that can only produce a deterministic output, always better data fidelity but worse visual perception quality [17]. It's promising to develop a generation-based model, but requiring high data demands, yet high-resolution ODIs are high cost to collect [54,55]. (2) Most methods directly perform SR on ERP format ODIs, while users usually watch ODIs in a narrow FOV using tangent projection (TP). So another promising direction is to use off-the-shelf planar models on TP images. Recent times have witnessed the introduction and widespread application of diffusion models [23,43], especially Stable Diffusion (SD) [38], which have provided a robust backbone for visual tasks [21,24,56,61], including SR [30,40,47,51,52,62]. However, if TP images are trivially one-by-one processed using diffusion-based SR models, they will exhibit discrepancies in the overlapping region when reprojected onto the ERP image. As a result, the global continuity is compromised." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Leveraging the strong image prior provided by SD, we propose the first diffusion-based zero-shot method for ODISR, named OmniSSR. Specifically, we propose Octadecaplex Tangent Information Interaction (OTII). OTII entails iterative conversion of intermediate SR results between ERP and TP, bridging the domain gap between ODIs and planar images. Building upon OTII, we further employ an approximate analytical solution of gradient descent, namely as Gradient Decomposition, to guide high-fidelity, high-quality omnidirectional image SR. By capitalizing on SD's effective image prior, our approach strikes a balance between fidelity and realness, ensuring that the restored ODIs exhibit both fidelity to the input data and realistic visual details. This method shows potential for advancing the current state of ODISR, providing improved resolution and visual quality across various applications. Fig. 1 showcases results fully demonstrating the superiority and performance of our proposed methods." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.728, + 0.608, + 0.742 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.762, + 0.787, + 0.808 + ], + "angle": 0, + "content": "- We propose OmniSSR, the first zero-shot ODISR method, using an off-the-shelf diffusion-based model, requiring no training or fine-tuning, leveraging existing image generation model priors to solve ODISR task." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- To bridge the domain gap between ODIs and planar images, we introduce Octadecaplex Tangent Information Interaction by repeatedly transforming" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.762, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.394, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.213, + 0.37, + 0.222 + ], + "angle": 0, + "content": "(a) Input low-resolution ERP image" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.146, + 0.49, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.214, + 0.496, + 0.223 + ], + "angle": 0, + "content": "Degraded tangent image" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.146, + 0.588, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.214, + 0.581, + 0.223 + ], + "angle": 0, + "content": "Restored by OSRT" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.146, + 0.686, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.214, + 0.689, + 0.222 + ], + "angle": 0, + "content": "Restored by OmniSSR" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.146, + 0.785, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.214, + 0.767, + 0.222 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.234, + 0.393, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.302, + 0.37, + 0.312 + ], + "angle": 0, + "content": "(b) Input low-resolution ERP image" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.234, + 0.49, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.302, + 0.495, + 0.312 + ], + "angle": 0, + "content": "Degraded tangent image" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.234, + 0.588, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.302, + 0.583, + 0.312 + ], + "angle": 0, + "content": "Restored by OSRT" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.234, + 0.686, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.302, + 0.688, + 0.311 + ], + "angle": 0, + "content": "Restored by OmniSSR" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.234, + 0.785, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.302, + 0.767, + 0.311 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.405 + ], + "angle": 0, + "content": "Fig. 1: We address omnidirectional image super-resolution in a zero-shot manner via OmniSSR. Presented above are select outcomes that sketch the essence of OmniSSR compared with current state-of-the-art approach OSRT [60]. Part (a) and (b) illustrate that OmniSSR upholds fidelity and visual realness at the same time, providing vivid and realistic details, while OSRT outputs over-smoothed and distorted results. Zoom in for more details." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.437, + 0.785, + 0.467 + ], + "angle": 0, + "content": "ODIs between ERP format and TP format, enabling ODISR task with pretrained diffusion models on planar images." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.468, + 0.787, + 0.527 + ], + "angle": 0, + "content": "- By iteratively updating images using the developed Gradient Decomposition technique, we introduce consistency constraints into the sampling process of the latent diffusion model, ensuring a trade-off between fidelity and realness in the reconstructed results." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.529, + 0.785, + 0.573 + ], + "angle": 0, + "content": "- Extensive experiments are conducted on the benchmark datasets, demonstrating the superior performance of our method over existing state-of-the-art approaches, which validate the effectiveness of OmniSSR." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.468, + 0.787, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.387, + 0.615 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.632, + 0.578, + 0.648 + ], + "angle": 0, + "content": "2.1 Single Image Super-Resolution (SISR)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Image super-resolution methods based on deep learning have undergone significant development over an extended period. Currently, they can be broadly classified into two categories of solutions. The first category involves end-to-end network training methods, which utilize image pairs consisting of low-resolution degraded images and high-resolution ground truth images for network training [6-8, 27, 31, 32, 63, 66]. The network architectures employed in this category include CNN [16], Transformers [46], etc. The second category employs image generation models as priors, such as GAN [20], diffusion models [23, 43], etc., where low-resolution images are used as conditions to generate high-resolution images. We will mainly introduce the methods using generative prior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Single Image SR using GAN prior In SR works utilizing GAN pri- ors [3,12,33,37,58], including real-world scenarios [8,49,50,65], pre-trained GAN" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "networks are employed to transform image features into latent space, where the corresponding latent code for the high-resolution image is searched, ultimately yielding the reconstructed high-resolution result." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Single Image SR using diffusion prior The diffusion model provides a powerful image prior, and the diffusion sampling process can generate highly realistic images. This strong prior distribution can be applied to various image restoration tasks, including super-resolution [9, 10, 19, 40, 42, 51]. Image-domain diffusion models directly provide prior distributions of image-domain data. DDNM [51] based on the mathematical method of Range-Null space Decomposition, iteratively refines content on the zero space, combining image prior content in the value domain to achieve image restoration. DDRM [25] uses SVD decomposition to obtain restoration results, which is similar to DDNM. DPS [9] transforms the image super-resolution problem into an optimization problem with consistency constraints, using gradient descent algorithms to guide the generation of image-domain diffusion models. GDP [19] further uses such gradient to update the degradation operator to tackle blind image inverse problems. Other methods including MCG [10], DDS [9] and unified control of diffusion generation [19, 42] use same strategy for image restoration, especially image super-resolution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.453, + 0.6, + 0.469 + ], + "angle": 0, + "content": "2.2 Omnidirectional Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.656 + ], + "angle": 0, + "content": "Omnidirectional image super-resolution (ODISR) aims to enhance the resolution of omnidirectional or 360-degree images, which are commonly captured by cameras with a wide field of view. This field has garnered increasing attention due to its applications in virtual reality, omnidirectional video streaming, and surveillance. Several approaches have been proposed to address the unique challenges of ODISR [1, 2, 35, 44]. For instance, Kämäräinen et al. [18] propose a deep learning-based approach for omnidirectional super-resolution, leveraging convolutional neural networks to effectively upscale low-resolution omnidirectional images while preserving spatial details. Similarly, Smolic et al. [36] introduce a novel omnidirectional super-resolution algorithm utilizing generative adversarial networks (GANs) to enhance the visual quality of omnidirectional images by hallucinating high-frequency details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.717 + ], + "angle": 0, + "content": "For evaluation purposes, researchers commonly utilize datasets such as the ODI-SR dataset from LAU-Net [13], and SUN 360 Panorama dataset [53]. These datasets enable the quantitative assessment of ODISR algorithms across various scenarios and facilitate fair comparisons between different methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.736, + 0.331, + 0.753 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this section, we first briefly introduce the preliminary background of our method (Sec. 3.1), and give an overall view of our proposed OmniSSR (Sec. 3.2). Then, we discuss the designs of Octadecaplex Tangent Information Interaction, which transform ODIs between ERP and TP formats with pre-upsampling strategy (Sec 3.3), and the Gradient Decomposition correction (Sec. 3.4)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.373, + 0.16 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.784, + 0.275 + ], + "angle": 0, + "content": "\\(\\mathrm{ERP} \\leftrightarrow \\mathrm{TP}\\) Transformation The essence of projection transformations between ERP and TP lie in determining the positions of target image pixels within the source image and computing their corresponding pixel values using interpolation algorithms, as digital images are always stored discretely [28]. Therefore, the \\(\\mathrm{ERP} \\rightarrow \\mathrm{TP}\\) transformation involves locating the TP image pixels on the ERP imaging plane, and vice versa. Gnomonic projection [11] provides the correspondence between ERP image pixels and TP image pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.785, + 0.307 + ], + "angle": 0, + "content": "For a pixel \\( P_{e}(x_{e},y_{e}) \\) within the ERP image, we first find its corresponding pixel \\( P_{s}(\\theta ,\\phi) \\) on the unit sphere using Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.318, + 0.785, + 0.333 + ], + "angle": 0, + "content": "\\[\n\\theta = 2 \\pi x _ {e} / W, \\phi = \\pi y _ {e} / H, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.784, + 0.401 + ], + "angle": 0, + "content": "where \\( H \\) and \\( W \\) are the height and width of the ERP image. The Cartesian coordinates of the ERP image and the angular coordinates on the unit sphere exhibit a straightforward one-to-one linear relationship, suggesting a conceptual equivalence between these two projection formats." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.784, + 0.432 + ], + "angle": 0, + "content": "Given the spherical coordinates of the tangent plane center \\((\\theta_c, \\phi_c)\\), The transformation from \\(P_s(\\theta, \\phi)\\) to \\(P_t(x_t, y_t)\\), i.e. \\(\\mathrm{ERP} \\to \\mathrm{TP}\\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.443, + 0.785, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t} = \\left(\\cos (\\phi) \\sin \\left(\\theta - \\theta_ {c}\\right)\\right) / \\zeta , \\tag {2} \\\\ y _ {t} = \\left(\\cos \\left(\\phi_ {c}\\right) \\sin (\\phi) - \\sin \\left(\\phi_ {c}\\right) \\cos (\\phi) \\cos \\left(\\theta - \\theta_ {c}\\right)\\right) / \\zeta , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.489, + 0.593, + 0.504 + ], + "angle": 0, + "content": "where \\(\\zeta = \\sin (\\phi_c)\\sin (\\phi) + \\cos (\\phi_c)\\cos (\\phi)\\cos (\\theta -\\theta_c).\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.505, + 0.678, + 0.52 + ], + "angle": 0, + "content": "The corresponding inverse transformation, i.e. \\(\\mathrm{TP} \\rightarrow \\mathrm{ERP}\\), is:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.529, + 0.785, + 0.566 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta = \\theta_ {c} + \\arctan \\left(\\left(x _ {t} \\sin (c)\\right) / \\left(\\rho \\cos \\left(\\phi_ {1}\\right) \\cos (c) - y _ {t} \\sin \\left(\\phi_ {c}\\right) \\sin (c)\\right)\\right), \\\\ \\phi = \\arcsin \\left(\\cos (c) \\sin \\left(\\phi_ {c}\\right) + y _ {t} \\sin (c) \\cos \\left(\\phi_ {c}\\right) / \\rho\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.577, + 0.503, + 0.593 + ], + "angle": 0, + "content": "where \\(\\rho = \\sqrt{x_t^2 + y_t^2}\\) and \\(c = \\arctan (\\rho)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.784, + 0.638 + ], + "angle": 0, + "content": "With Eq. 2 and Eq. 3, we can build one-to-one forward and inverse mapping functions between pixels on the ERP image and pixels on the TP images. An illustration of the \\(\\mathrm{ERP} \\to \\mathrm{TP}\\) transformation is shown in Fig. 2(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Iterative Denoising for Super-Resolution Utilizing the rich image priors provided by SD, we can super-resolve planar images. During initialization, the images are passed through the encoder \\(\\mathcal{E}\\) of SD to obtain latent codes, which are then added to pure noise to generate initial noise \\(\\mathbf{z}_T\\). Following the approach proposed by StableSR [47], we pass the images through a time-aware adapter \\(\\mathcal{T}\\). This adapter network structure is similar to the down-sampling part in denoising UNet, taking the image and the time step \\(t\\) of diffusion sampling as inputs to obtain the latent code feature for step \\(t\\). This feature, along with the latent code \\(\\mathbf{z}_t\\) for each step and the time step \\(t\\), is then passed through denoising UNet to calculate the denoised result \\(\\mathbf{z}_{0|t}\\) and the latent code \\(\\mathbf{z}_{t - 1}\\) for the next sampling step. By iterating this process \\(T\\) times, we can obtain the final super-resolution result via decoder \\(\\mathcal{D}\\) of SD, yielding high-resolution images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.545, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.277, + 0.482, + 0.287 + ], + "angle": 0, + "content": "(a) Equirectangular Projection \\(\\rightarrow\\) Tangent Projection" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.144, + 0.785, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.277, + 0.725, + 0.288 + ], + "angle": 0, + "content": "(b) Pre-Upsampling Strategy" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.34 + ], + "angle": 0, + "content": "Fig. 2: Details about gnomonic transformations. (a) conversion from ERP to TP. (b) pre-upsampling proposed in Octadecaplex Tangent Information Interaction (Sec. 3.3) mitigating loss during transformation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.368, + 0.341, + 0.382 + ], + "angle": 0, + "content": "3.2 Overview" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.411, + 0.789, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.562, + 0.564, + 0.57 + ], + "angle": 0, + "content": "(b) Detailed illustration of GD Correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.682 + ], + "angle": 0, + "content": "Fig. 3: Overview of our proposed OmniSSR. Input low-resolution omnidirectional image \\(\\mathbf{E}_{init}\\) in ERP format is first projected onto Tangent Projection (TP) images \\(\\mathbf{x}_{init}^{(1)},\\mathbf{x}_{init}^{(2)},\\dots,\\mathbf{x}_{init}^{(m)}\\), then iteratively refined via Stable Diffusion (SD) with a time-aware adapter and controllable feature wrapping (CFW) module. In each step of diffusion sampling, we adopt the Gradient Decomposition (GD) correction technique to introduce consistency constraints for the restored intermediate results. After \\(T\\) steps of sampling, we obtain the final result \\(\\tilde{\\mathbf{E}}_0\\) with high resolution and better visual quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our approach can be divided into three parts. The first part is pre-processing, where we initially up-sample the low-resolution ERP images \\(\\mathbf{E}_{init}\\) with bicubic interpolation to target high-resolution size, then project them onto tangent planes to obtain a series of TP images. These TP images are transformed to the latent space by the SD encoder, iteratively processed through denoising UNet and time-aware adapter network, and then decoded to obtain high-resolution TP images. During each denoising step, these TP images are transformed back via inverse transformation to ERP images, employing the Gradient Decomposition correction to ensure consistency constraints in diffusion sampling. After \\(T\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table", + "bbox": [ + 0.215, + 0.147, + 0.77, + 0.413 + ], + "angle": 0, + "content": "
Algorithm 1: OmniSSR PipelineAlgorithm 2: Iterative Denois-ing with GD Correction
Input: Einit, F, F-1, A, A†, E, D, TInput: Einit, F, F-1, A, A†, E, D, T, Output: Latent code {z(1), z(2), ..., z(m)}
Output: SR result E0for t = T to 1 do
1{x(1), x(2), ..., x(m)} = F(Einit)for i = 1 to m do
2for i = 1 to m doεt = εθ(z(t), T(z(i), t), t)
3z(i)init = Ε(x(i))z(0|t) = 1/√αt(z(t)i - εt√1-αt)
4ε(i) ∼ N(0, I)x(0|i) = D(z(i))
5zT(i) = √αTz(i) + √1-αTε(i)end
6endE0|t = F-1{x(1), x(2), ..., x(m)}
7Get{z(0), z(2), ..., z(m)} from Algo. 2E0|t = E0|t + γeA†(Einit - AE0|t)
8for i = 1 to m do{x(1), x(2), ..., x(m)} = F(E0|t)
9x(0)i = D(z(0))for i = 1 to m do
10endz(0|i) = (1-γl)z(0|i) + γlE(x(i))
11E0 = F-1{x(0), x(2), ..., x(m)}z(t-1) ∼ p(z(t-1)|z(t,i), x(0|i))
12E0 = E0 + γpA†(Einit - AE0)end
13return E0
14end
15return {z(0), z(2), ..., z(m)}
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.784, + 0.47 + ], + "angle": 0, + "content": "iterations, the final super-resolution result is obtained. A formulaic description for OmniSSR is shown in Algo. 1. Fig. 3 shows the overview of our pipeline." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.491, + 0.706, + 0.508 + ], + "angle": 0, + "content": "3.3 Octadecaplex Tangent Information Interaction (OTII)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.729 + ], + "angle": 0, + "content": "Motivation To apply SD for ODISR, a straightforward way is to perform the \\(\\mathrm{ERP} \\rightarrow \\mathrm{TP}\\) transformation on the input ERP image. Then, each obtained TP image is fed into the SD-based model for SR. Finally, the \\(\\mathrm{TP} \\rightarrow \\mathrm{ERP}\\) transformation yields the ultimate super-resolved ERP image. OmniFusion [28] employs a similar approach for depth estimation. However, this simplistic strategy fractures the inherent global coherence of ODIs, leading to pixel-level discontinuities in the fused ERP images. Moreover, interpolation algorithms cause significant information loss in the original projection transformations, resulting in more blurred images. If applied multiple times, this exacerbates the information loss even further. To mitigate this, a trivial solution is to increase the pixel count (resolution) of the intermediate projection imaging plane. However, excessively high resolutions in TP images can introduce unnecessary computational overhead during the denoising stage and potentially compromise the denoising performance. (see Supplementary Materials for details)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Information Interaction and Pre-upsampling Based on the analysis of the Motivation in 3.3, we propose OTII by alternating the intermediate results between ERP and TP formats at each denoising step, where a single ERP image is represented by 18 TP images. From Sec. 3.1, we can achieve the \\(\\mathrm{ERP}\\rightarrow \\mathrm{TP}\\) transformation (denoted as \\(\\mathcal{F}(\\cdot)\\)) and the \\(\\mathrm{TP}\\rightarrow \\mathrm{ERP}\\) transformation (denoted as \\(\\mathcal{F}^{-1}(\\cdot)\\)). Through the \\(\\mathrm{ERP}\\rightarrow \\mathrm{TP}\\) transformation, we can convert distorted ERP" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "images into TP images with content distributions that approximate those of planar images. This enables the use of the original SD super-resolution method for planar images. Conversely, through the \\(\\mathrm{TP} \\rightarrow \\mathrm{ERP}\\) transformation, we can fuse information between different TP images holistically, while providing ERP-format input for the subsequent GD Correction in Sec. 3.4. To handle information loss during projection transformation, we further propose to pre-upsample the source image before projection transformations, as shown in Fig. 2(b). Our experiments in Sec. 4.4 demonstrate that this pre-upsampling strategy can significantly mitigate the information loss caused by projection transformations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.306, + 0.707, + 0.323 + ], + "angle": 0, + "content": "3.4 Gradient Decomposition (GD) Correction for Fidelity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.424 + ], + "angle": 0, + "content": "SD-based methods, as introduced in Sec. 3.1, can perform SR on sliced TP images. However, relying solely on the SR results from SD may lack consistency and fail to accurately preserve the original semantic information and details of the low-resolution image. To enhance the consistency of the SR results from SD, we opt to use convex optimization methods to iteratively refine them. Modeling the SR task as an image inverse problem, the following equation is formulated:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.436, + 0.786, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathbf {A x} + \\mathbf {n}, \\quad \\mathbf {n} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.524 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) represents the original image, \\(\\mathbf{y}\\) denotes the degraded result, \\(\\mathbf{A}\\) is the degradation operator (e.g., bicubic downsampling for super-resolution), and \\(\\mathbf{n}\\) is random noise. The objective we aim to solve can be expressed as the following convex optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.536, + 0.786, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} | | \\mathbf {y} - \\mathbf {A x} | | _ {2} ^ {2} + \\lambda \\mathcal {R} (\\mathbf {x}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.708 + ], + "angle": 0, + "content": "where the first term is the data-fidelity term, ensuring the consistency of image reconstruction, and the second term is the regulation term, ensuring the sparsity of the reconstruction result, thus making the image more realistic. The regularization term can be the 1-norm, Total Variation, etc. The aforementioned convex optimization problem can be solved using a series of algorithms, such as gradient descent, ADMM [4], etc. Considering the trade-off between time and performance, we turn to find a solution based on gradient descent, and provide an approximate analytical solution composed of a fidelity term and a realness term, named \"Gradient Decomposition (GD)\":" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.719, + 0.786, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mathbf {E}} _ {0 | t} = \\mathbf {E} _ {0 | t} + \\alpha \\nabla_ {\\mathbf {E} _ {0 | t}} \\left\\| \\mathbf {E} _ {i n i t} - \\mathbf {A} \\mathbf {E} _ {0 | t} \\right\\| _ {F} = \\mathbf {E} _ {0 | t} + \\alpha \\times 2 \\left(\\mathbf {A} ^ {\\dagger} \\mathbf {E} _ {i n i t} - \\mathbf {A} ^ {\\dagger} \\mathbf {A} \\mathbf {E} _ {0 | t}\\right) \\\\ = \\mathbf {E} _ {0 | t} + \\gamma \\mathbf {A} ^ {\\dagger} \\left(\\mathbf {E} _ {i n i t} - \\mathbf {A} \\mathbf {E} _ {0 | t}\\right) = \\gamma \\mathbf {A} ^ {\\dagger} \\mathbf {E} _ {i n i t} + (\\mathbf {I} - \\gamma \\mathbf {A} ^ {\\dagger} \\mathbf {A}) \\mathbf {E} _ {0 | t} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}^{\\dagger}\\) denotes pseudo-inverse of degradation operator \\(\\mathbf{A}\\), \\(\\mathbf{E}_{init}\\) denotes initial low-resolution ERP input, \\(\\mathbf{E}_{0|t}\\) denotes restored result by SD, \\(\\tilde{\\mathbf{E}}_{0|t}\\) denotes corrected result by GD, \\(\\alpha\\) denotes the learning rate of gradient descent, and \\(\\gamma\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.663, + 0.841 + ], + "angle": 0, + "content": "This claim will be further illustrated in subsequent experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "denotes the simplified hyper-parameter which is further tuned using grid search. The final setting of \\(\\gamma\\) on different stages is shown in Sec. 4.1, and the ablation studies of parameter choice are in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.329 + ], + "angle": 0, + "content": "This technique could be seen as a step of gradient descent optimization, and the optimized result could be decomposed of (1) \\(\\gamma \\mathbf{A}^{\\dagger}\\mathbf{E}_{init}\\), which ensures the consistency of the generated result, and (2) \\((\\mathbf{I} - \\gamma \\mathbf{A}^{\\dagger}\\mathbf{A})\\mathbf{E}_{0|t}\\), which serves as the iteratively updated generated result to improve its realness; \\(\\gamma\\) is a hyperparameter balancing the data fidelity and visual quality. For a better diversity and generality of the SR process, we expand this solution to latent space, and obtain the denoising result from both denoising UNet and corrected TP images (Algo. 2 line 11). A more detailed understanding of the iterative denoising process and application of GD correction could be referred to Algo. 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.353, + 0.376, + 0.371 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.386, + 0.457, + 0.401 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.501 + ], + "angle": 0, + "content": "Datasets and Pretrained Models We choose the test set of ODI-SR dataset from LAU-Net [13] and SUN 360 Panorama dataset [53], comprising 97 and 100 omnidirectional images respectively, for experimental evaluation. The ground truth images are of size \\(1024 \\times 2048\\) pixels. In SR methods such as GDP [19] and PSLD [39] for planar images, we partitioned the images into several \\(256 \\times 256\\) patches and performed super-resolution on each patch individually." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.787, + 0.579 + ], + "angle": 0, + "content": "For pre-trained models, we adopt from StableSR [47], which provided a SR network for planar images based on SD. This network architecture includes a time-aware adapter, a controllable feature wrapping (CFW) module, and the original SD structure from HuggingFace. All of them are kept untrained in our proposed OmniSSR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Settings We set diffusion sampling steps to 200, which is the same as StableSR. The steps for other diffusion-based methods are set the same as their default settings (e.g. 1000 steps for PSLD). The degradation for low-resolution ERP images is bicubic down-sampling, and the implementation of its pseudo-inverse can be referred from code of DDRM \\([25]^2\\). For choices of hyper-parameter \\(\\gamma\\) in GD correction, we set \\(\\gamma_p = 1.0\\), \\(\\gamma_e = 1.0\\), \\(\\gamma_l = 0.5\\). Our code is developed via PyTorch on NVIDIA 3090Ti GPU." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.731, + 0.717, + 0.746 + ], + "angle": 0, + "content": "4.2 Comparison of OmniSSR with diffusion-based methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.787, + 0.803 + ], + "angle": 0, + "content": "To evaluate the performance of proposed OmniSSR, we compare our method with recent state-of-the-art zero-shot methods for single image SR task: DPS [9], DDRM [25], GDP [19] which are based on the image-domain diffusion model," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.5, + 0.826 + ], + "angle": 0, + "content": "2 https://github.com/bahjat-kawar/ddrm" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.826, + 0.603, + 0.84 + ], + "angle": 0, + "content": "\\(^{3}\\) Code is at https://github.com/LiRunyi2001/OmniSSR." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.603, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: SR results under bicubic downsampling on ODI-SR and SUN 360 Panorama datasets. For tasks not implemented in those papers, we mark N/A in the corresponding results. The best results are shown in bold, and the second best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.2, + 0.784, + 0.389 + ], + "angle": 0, + "content": "
MethodScaleODI-SRSUN 360 Panorama
WS-PSNR↑WS-SSIM↑FID↓LPIPS↓WS-PSNR↑WS-SSIM↑FID↓LPIPS↓
Bicubic28.140.834324.000.216428.670.853729.250.1933
DDRM [25]27.900.831712.280.166129.550.867013.100.1426
DPS [9]20.990.6194148.300.524921.440.6598148.830.5175
GDP [19]27.890.815726.560.272428.600.837628.020.2445
PSLD [39]N/AN/AN/AN/AN/AN/AN/AN/A
DiffIR [52]23.770.658357.230.468723.540.677558.060.4658
StableSR [47]22.700.645844.870.303923.300.690743.490.2858
OmniSSR28.570.854013.010.157529.690.878112.990.1459
Bicubic25.430.705950.840.375525.490.722955.990.3656
DDRM [25]25.430.736732.690.320625.830.744332.930.3304
DPS [9]24.750.6594120.740.491121.090.6119175.21430.5541
GDP [19]23.160.669277.430.426023.750.656990.230.4240
PSLD [39]21.720.5498107.990.532921.750.5828141.490.5461
DiffIR [52]24.010.677054.140.436723.900.701450.370.4235
StableSR [47]23.330.657749.950.313523.990.699846.030.3023
OmniSSR25.770.727930.970.297726.010.748134.580.2963
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.492 + ], + "angle": 0, + "content": "and PSLD [39], which is based on latent diffusion model. We also choose supervised diffusion-based super-resolution approaches including StableSR [47] and DiffIR [52] for comparison. We conduct experiments on \\(\\times 2\\) and \\(\\times 4\\) SR with ERP bicubic downsampling, on ODI-SR test set and SUN test set. We choose WS-PSNR [45], WS-SSIM [67], FID [22], and LPIPS [64] as the main metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.567 + ], + "angle": 0, + "content": "Quantitative results are presented in Tab. 1. With proposed OTII and GD correction, OmniSSR out-performs previous methods in terms of both Fidelity (from WS-PSNR and WS-SSIM) and Realness (from FID, LPIPS), which shows superior performance to existing diffusion-based methods for ODISR tasks on different scales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.612 + ], + "angle": 0, + "content": "Qualitative results are shown in Fig. 4 and Fig. 5, which illustrates the visualization of SR results on SUN test set and ODI-SR test set with \\(\\times 2\\) and \\(\\times 4\\) scales, by different methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.719 + ], + "angle": 0, + "content": "The visual results indicate that our OmniSSR exhibits superior capability for detail recovery compared to other methods, particularly evident in textual elements (e.g., the text \"flapping\" in upper part of Fig. 4), complex objects (e.g., the black desk with a screen in lower part of Fig. 4, patterns above the white door in lower part of Fig. 5), and small-scale objects (e.g., the person and clock behind the desk in upper part of Fig. 5). OmniSSR demonstrates the ability to recover highly detailed and realistic visual effects from TP images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.667, + 0.756 + ], + "angle": 0, + "content": "4.3 Comparison with end-to-end supervised methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The experiments of comparison in Sec. 4.2 are mainly focused on zero-shot image super-resolution methods, and supervised single image super-resolution methods, where the approaches are not trained or fine-tuned on omnidirectional images. In this part, we will compare OmniSSR to supervised end-to-end methods with end-to-end training on ODI datasets, including SwinIR and OSRT. Besides the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.144, + 0.38, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.212, + 0.36, + 0.232 + ], + "angle": 0, + "content": "\\( \\therefore m - 1 \\neq 0 \\) ;" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.144, + 0.471, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.212, + 0.464, + 0.232 + ], + "angle": 0, + "content": "HR PSNR/SSIM" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.144, + 0.563, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.213, + 0.564, + 0.233 + ], + "angle": 0, + "content": "Bicubic 28.67dB/0.8317" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.144, + 0.656, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.212, + 0.657, + 0.233 + ], + "angle": 0, + "content": "DPS [9] 24.02dB/0.5849" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.144, + 0.754, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.212, + 0.755, + 0.233 + ], + "angle": 0, + "content": "DDRM [25] 30.04dB/0.8855" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.232, + 0.38, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.3, + 0.36, + 0.311 + ], + "angle": 0, + "content": "SUN 360 \\((\\times 2)\\) : 001" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.232, + 0.471, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.301, + 0.469, + 0.321 + ], + "angle": 0, + "content": "GDP [19] 0.07dB/0.8802" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.232, + 0.563, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.301, + 0.563, + 0.321 + ], + "angle": 0, + "content": "DiffIR [52] 24.47dB/0.6421" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.232, + 0.656, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.301, + 0.657, + 0.321 + ], + "angle": 0, + "content": "StableSR [47] 23.81dB/0.7384" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.232, + 0.755, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.301, + 0.758, + 0.321 + ], + "angle": 0, + "content": "OmniSSR(ours) 30.15dB/0.8859" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.322, + 0.38, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.476, + 0.36, + 0.487 + ], + "angle": 0, + "content": "SUN 360 \\((\\times 4)\\) : 009" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.322, + 0.471, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.477, + 0.471, + 0.498 + ], + "angle": 0, + "content": "GDP [19] 23.02dB/0.7525" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.322, + 0.563, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.477, + 0.563, + 0.498 + ], + "angle": 0, + "content": "DiffIR [52] 23.48dB/0.7799" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.322, + 0.656, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.477, + 0.657, + 0.498 + ], + "angle": 0, + "content": "StableSR [47] 24.62dB/0.7981" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.322, + 0.754, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.477, + 0.758, + 0.498 + ], + "angle": 0, + "content": "OmniSSR(ours) 26.53dB/0.8265" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.51, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Fig. 4: Visualized comparison of \\(\\times 2\\) and \\(\\times 4\\) SR results on SUN 360 test set. 001 and 009 are the ID numbers in the test set filenames. We also calculate the PSNR and SSIM to HR ground truth of each SR result and downsampled image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.789, + 0.732 + ], + "angle": 0, + "content": "main metrics in Sec. 4.2, we also use NIQE [34] and DISTS [15] to evaluate the visual perception of SR outputs. Results are presented in Tab. 2, which shows that although our OmniSSR exhibits inferior fidelity metrics compared to end-to-end supervised methods trained directly on ODI datasets, it demonstrates notable improvements in the visual quality and authenticity of super-resolved images. Notably, end-to-end methods often produce smoothed reconstructions with distortions, whereas our approach preserves finer details and adheres more closely to the realistic distribution. Considering that our method has never been trained or tuned on ODI datasets, nor having omnidirectional images prior, this result is acceptable." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.403, + 0.768 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We first sequentially validate the performance improvement of the proposed strategy in OmniSSR including input image type, OTII and GD correction, on the ODI-SR test set with \\(\\times 2\\) SR task, thereby demonstrating the significance of these strategies. The details are demonstrated as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.144, + 0.38, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.3, + 0.358, + 0.311 + ], + "angle": 0, + "content": "ODI-SR \\((\\times 2)\\) :067" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.144, + 0.471, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.213, + 0.437, + 0.222 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.223, + 0.436, + 0.232 + ], + "angle": 0, + "content": "PSNR/S" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.144, + 0.563, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.213, + 0.541, + 0.223 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.223, + 0.564, + 0.232 + ], + "angle": 0, + "content": "27.67dB/0.8095" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.144, + 0.656, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.213, + 0.634, + 0.223 + ], + "angle": 0, + "content": "DPS [9]" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.223, + 0.657, + 0.232 + ], + "angle": 0, + "content": "22.93dB/0.5653" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.301, + 0.453, + 0.312 + ], + "angle": 0, + "content": "GDP [19]" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.233, + 0.563, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.301, + 0.549, + 0.312 + ], + "angle": 0, + "content": "DiffIR [52]" + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.312, + 0.563, + 0.321 + ], + "angle": 0, + "content": "22.65dB/0.6248" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.233, + 0.656, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.301, + 0.652, + 0.312 + ], + "angle": 0, + "content": "StableSR [47]" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.312, + 0.657, + 0.321 + ], + "angle": 0, + "content": "21.80dB/0.5892" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.144, + 0.754, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.213, + 0.742, + 0.223 + ], + "angle": 0, + "content": "DDRM [25]" + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.223, + 0.755, + 0.233 + ], + "angle": 0, + "content": "29.91dB/0.8809" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.233, + 0.754, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.301, + 0.758, + 0.311 + ], + "angle": 0, + "content": "OmniSSR(ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.312, + 0.754, + 0.321 + ], + "angle": 0, + "content": "29.99dB/0.8798" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.322, + 0.38, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.476, + 0.358, + 0.487 + ], + "angle": 0, + "content": "ODI-SR \\((\\times 4)\\) :049" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.322, + 0.471, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.39, + 0.436, + 0.399 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.4, + 0.465, + 0.41 + ], + "angle": 0, + "content": "PSNR/SSIM" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.322, + 0.563, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.389, + 0.542, + 0.398 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.4, + 0.564, + 0.409 + ], + "angle": 0, + "content": "25.44dB/0.7536" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.322, + 0.656, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.389, + 0.634, + 0.399 + ], + "angle": 0, + "content": "DPS [9]" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.399, + 0.657, + 0.409 + ], + "angle": 0, + "content": "26.21dB/0.7574" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.4, + 0.471, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.477, + 0.474, + 0.498 + ], + "angle": 0, + "content": "GDP [19] 24.15dB/0.7179" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.4, + 0.563, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.477, + 0.549, + 0.488 + ], + "angle": 0, + "content": "DiffIR [52]" + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.488, + 0.563, + 0.497 + ], + "angle": 0, + "content": ".25dB/0.7594" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.4, + 0.657, + 0.409 + ], + "angle": 0, + "content": "26.21dB/0.7574" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.4, + 0.657, + 0.409 + ], + "angle": 0, + "content": "26.21dB/0.7574" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.322, + 0.754, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.389, + 0.742, + 0.399 + ], + "angle": 0, + "content": "DDRM [25]" + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.399, + 0.755, + 0.409 + ], + "angle": 0, + "content": "27.12dB/0.8129" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.409, + 0.754, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.477, + 0.758, + 0.488 + ], + "angle": 0, + "content": "OmniSSR(ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.488, + 0.755, + 0.498 + ], + "angle": 0, + "content": "27.20dB/0.8168" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.51, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Fig. 5: Visualized comparison of \\(\\times 2\\) and \\(\\times 4\\) SR results on ODI-SR test set. 067 and 049 are the ID numbers in test set filenames. We also calculate the PSNR and SSIM between ground truth and each SR result as well as the downsampled image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.785, + 0.61 + ], + "angle": 0, + "content": "1) we do not use any proposed strategy in the SR task, which is equivalent to the vanilla StableSR baseline;" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.641 + ], + "angle": 0, + "content": "2) we transform the degraded ERP image to TP images and feed them separately into StableSR pipeline, instead of directly inputting ERP images;" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.785, + 0.671 + ], + "angle": 0, + "content": "3) based on 2), we add OTII strategy during the denoising process of SD (Algo. 2 line 7);" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.785, + 0.703 + ], + "angle": 0, + "content": "4) based on 2), we add GD correction at the post-processing stage (Algo. 1 line 12) of the overall pipeline;" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.787, + 0.734 + ], + "angle": 0, + "content": "5) based on 3) and 4), we add GD correction at every step and post-processing stage of sampling, to improve the consistency of the restored result." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Note that the execution of GD correction requires the execution of OTII in the denoising process simultaneously, there is no scenario where only GD correction is executed without the execution of OTII in the denoising process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Quantitative results of ablation studies are shown in Tab. 3. From the result shown below, we could come to the claim that the OTII helps improve the performance on the domain level, and the transformation between ERP and TP images provides information fusion among adjacent TP images. Our proposal of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 2: Comparison on \\( \\times 4 \\) SR task with supervised methods trained on ODI-SR dataset, including SwinIR and OSRT. The best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.185, + 0.726, + 0.273 + ], + "angle": 0, + "content": "
MethodDatasetWS-PSNR↑WS-SSIM↑FID↓LPIPS↓NIQE↓DISTS↓
SwinIR [29]26.760.762027.940.33215.39610.1710
OSRT [60]ODI-SR26.890.764627.390.32585.43640.1695
OmniSSR25.770.727930.970.29775.28910.1541
SwinIR [29]26.020.769239.900.34195.24400.1325
OSRT [60]SUN 36026.330.776639.220.33645.29840.1312
OmniSSR26.010.748134.580.29635.13290.1299
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.286, + 0.785, + 0.313 + ], + "angle": 0, + "content": "Table 3: Ablation studies of OmniSSR on input type, OTII, and GD correction, on the test set of the ODI-SR dataset. Best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.327, + 0.739, + 0.401 + ], + "angle": 0, + "content": "
Input typeOTIIGD CorrectionWS-PSNR↑WS-SSIM↑FID↓LPIPS↓
ERP××22.690.645844.870.3039
TP××23.530.684943.910.3113
TP✓×23.740.684765.350.3748
TP×✓ (in post-process only)26.770.819215.410.1691
TP✓✓28.580.854013.010.1575
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.785, + 0.489 + ], + "angle": 0, + "content": "Gradient Decomposition corrects such restoration result, improving fidelity and realness significantly at the same time, and it would be better if it is applied at each step of the overall denoising pipeline. Tab. 4 shows the effect of mitigating information loss via proposed pre-upsampling strategy." + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.516, + 0.416, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.516, + 0.585, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.516, + 0.755, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.625, + 0.785, + 0.709 + ], + "angle": 0, + "content": "Fig.6: Ablation of choices on \\(\\gamma_{p}\\), \\(\\gamma_{e}\\) and \\(\\gamma_{l}\\). For better readability, WS-PSNR and LPIPS are chosen as evaluation metrics for fidelity and visual quality, respectively, to demonstrate the performance under different choices of the gamma parameter. We illustrate the results of (a) \\(\\gamma_{p}\\) and \\(\\gamma_{e}\\) fixed, while adjusting \\(\\gamma_{l}\\); (b) \\(\\gamma_{e}\\) and \\(\\gamma_{l}\\) fixed, while adjusting \\(\\gamma_{p}\\); (c) \\(\\gamma_{p}\\) and \\(\\gamma_{l}\\) fixed, while adjusting \\(\\gamma_{e}\\). It can be observed that when \\(\\gamma_{p} = 1\\), \\(\\gamma_{e} = 1\\), and \\(\\gamma_{l} = 0.5\\), OmniSSR achieves the relatively best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.81 + ], + "angle": 0, + "content": "For \\(\\gamma\\) in the GD correction technique, we use grid search to obtain better results on ODI-SR dataset and \\(\\times 4\\) SR task. Fig. 6 shows performance on different choices of \\(\\gamma_{p}\\) in Algo. 1 line 12, \\(\\gamma_{e}\\) in Algo. 2 line 8, and \\(\\gamma_{l}\\) in Algo. 2 line 11. The entire ablation of \\(\\gamma_{p},\\gamma_{e}\\) and \\(\\gamma_{l}\\), with WS-PSNR, WS-SSIM, FID and LPIPS score all calculated and compared, will be provided in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To evaluate the generalizability of our proposed modules, including Pre-Upsampling, OTII, and GD correction, we further conducted ablation studies" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Results of pre-upsampling strategy on different scales, where \\((x,y)\\) denotes bicubic-based upsampling at \\(x\\times\\) scale to ERP before \\(\\mathrm{ERP}\\rightarrow \\mathrm{TP}\\), and \\(y\\times\\) scale to TP before \\(\\mathrm{TP}\\rightarrow \\mathrm{ERP}\\) transformation. Best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.323, + 0.2, + 0.681, + 0.24 + ], + "angle": 0, + "content": "
\\( \\mathrm{{ERP}} \\rightarrow \\mathrm{{TP}} \\rightarrow \\mathrm{{ERP}} \\)(1,1)(1,4)(4,1)(4,2)(2,4)(4,4)
WS-PSNR↑28.9838.1128.9933.9138.0538.18
WS-SSIM↑0.88590.98380.88620.96260.98370.9841
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.784, + 0.314 + ], + "angle": 0, + "content": "on two super-resolution backbones, StableSR and SwinIR. The results underscore substantial performance enhancements facilitated by our modules across both backbones, which is provided in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.335, + 0.504, + 0.351 + ], + "angle": 0, + "content": "5 Limitation and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.785, + 0.516 + ], + "angle": 0, + "content": "Although OmniSSR bridges the gap between omnidirectional and planar images, achieving competitive performance and better visual results in ODISR, it still exhibits the following limitations: (1) The inference of the diffusion model requires a considerable amount of time, approximately 14 minutes per ERP-formatted omnidirectional image to be super-resolved into size \\(1024 \\times 2048\\), making real-time super-resolution challenging; (2) Multiple conversions between ERP and TP are required in the pipeline, leading to improved performance but consuming additional inference time; (3) Further exploration of the convex optimization properties of GD correction is warranted, such as designing gradient term coefficients adaptive to reconstruction results and degradation types." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.787, + 0.624 + ], + "angle": 0, + "content": "This study explores the application of image generation models to ODISR tasks. In future work, the framework behind OmniSSR can be extended beyond the confines of image super-resolution in a single scenario and venture into more complex ODI-based real-world scenarios. These include ODI editing, ODI inpainting, enhancing the quality of 3D Gaussian Splatting scenes [26,41] obtained after super-resolving ERP images, as well as enhancing the quality of omnidirectional videos [48], and other possible diffusion-based applications [57,68]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.36, + 0.66 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper leverages the image prior of Stable Diffusion (SD) and employs the Octadecaplex Tangent Information Interaction (OTII) to achieve zero-shot omnidirectional image super-resolution. Additionally, we propose the Gradient Decomposition (GD) correction based on convex optimization algorithms to refine the initial super-resolution results, enhancing the fidelity and realness of the restored images. The superior performance of our proposed method, OmniSSR, is demonstrated on benchmark datasets. By bridging the gap between omnidirectional and planar images, we establish a training-free approach, mitigating the data demand and over-fitting associated with end-to-end training. The application scope of our method can be further extended to various applications, presenting potential value across multiple visual tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. An, H., Zhang, X.: Perception-oriented omnidirectional image super-resolution based on transformer network. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Arican, Z., Frossard, P.: Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing (TIP) (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.282 + ], + "angle": 0, + "content": "3. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for image super-resolution and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.323 + ], + "angle": 0, + "content": "4. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.365 + ], + "angle": 0, + "content": "5. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "6. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.459 + ], + "angle": 0, + "content": "7. Cheng, M., Ma, H., Ma, Q., Sun, X., Li, W., Zhang, Z., Sheng, X., Zhao, S., Li, J., Zhang, L.: Hybrid transformer and cnn attention network for stereo image superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "8. Chong, M., Yanze, W., Xintao, W., Chao, D., Jian, Z., Ying, S.: Metric learning based interactive modulation for real-world super-resolution. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "9. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "0. Chung, H., Sim, B., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.744, + 0.581 + ], + "angle": 0, + "content": "1. Coxeter, H.S.M.: Introduction to geometry. John Wiley & Sons, Inc. (1961)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "2. Daras, G., Dean, J., Jalal, A., Dimakis, A.: Intermediate layer optimization for inverse problems using deep generative models. In: Proceedings of the International Conference on Machine Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.676 + ], + "angle": 0, + "content": "3. Deng, X., Wang, H., Xu, M., Guo, Y., Song, Y., Yang, L.: Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "4. Deng, X., Wang, H., Xu, M., Li, L., Wang, Z.: Omnidirectional image superresolution via latitude adaptive network. IEEE Transactions on Multimedia (TMM) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "5. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Duan, H., Zhai, G., Min, X., Zhu, Y., Fang, Y., Yang, X.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the IEEE International Symposium on Circuits and Systems (ISCAS) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Fakour-Sevom, V., Guldogan, E., Kämäräinen, J.K.: 360 panorama superresolution using deep convolutional networks. In: Proceedings of the Int. Conf. on Computer Vision Theory and Applications (VISAPP) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "19. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.342 + ], + "angle": 0, + "content": "21. Guo, L., Tao, T., Cai, X., Zhu, Z., Huang, J., Zhu, L., Gu, Z., Tang, H., Zhou, R., Han, S., et al.: Cas-diffcom: Cascaded diffusion model for infant longitudinal super-resolution 3d medical image completion. arXiv preprint arXiv:2402.13776 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.384 + ], + "angle": 0, + "content": "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.412 + ], + "angle": 0, + "content": "23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.453 + ], + "angle": 0, + "content": "24. Jiang, Y., Chan, K.C., Wang, X., Loy, C.C., Liu, Z.: Reference-based image and video super-resolution via \\( c^2 \\)-matching. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.495 + ], + "angle": 0, + "content": "25. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of the ICLR Workshop on Deep Generative Models for Highly Structured Data (ICLRW) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.523 + ], + "angle": 0, + "content": "26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.786, + 0.563 + ], + "angle": 0, + "content": "27. Li, W., Chen, B., Zhang, J.: D3c2-net: Dual-domain deep convolutional coding network for compressive sensing. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.605 + ], + "angle": 0, + "content": "28. Li, Y., Guo, Y., Yan, Z., Huang, X., Duan, Y., Ren, L.: Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.647 + ], + "angle": 0, + "content": "29. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "30. Liu, J., Wang, Q., Fan, H., Wang, Y., Tang, Y., Qu, L.: Residual denoising diffusion models. arXiv preprint arXiv:2308.13712 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.716 + ], + "angle": 0, + "content": "31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.786, + 0.758 + ], + "angle": 0, + "content": "32. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.813 + ], + "angle": 0, + "content": "33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.841 + ], + "angle": 0, + "content": "34. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters (SPL) (2013)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Nishiyama, A., Ikehata, S., Aizawa, K.: \\(360^{\\circ}\\) single image super resolution via distortion-aware network and distorted perspective images. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "36. Ozcinar, C., Rana, A., Smolic, A.: Super-resolution of omnidirectional images using adversarial learning. In: Proceedings of the IEEE International Workshop on Multimedia Signal Processing (MMSPW) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.271 + ], + "angle": 0, + "content": "37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.311 + ], + "angle": 0, + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.366 + ], + "angle": 0, + "content": "39. Rout, L., Raoof, N., Daras, G., Caramanis, C., Dimakis, A., Shakkottai, S.: Solving linear inverse problems provably via posterior sampling with latent diffusion models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.407 + ], + "angle": 0, + "content": "40. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.447 + ], + "angle": 0, + "content": "41. Schonbein, M., Geiger, A.: Omnidirectional 3d reconstruction in augmented manhattan worlds. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.489 + ], + "angle": 0, + "content": "42. Song, J., Zhang, Q., Yin, H., Mardani, M., Liu, M.Y., Kautz, J., Chen, Y., Vahdat, A.: Loss-guided diffusion models for plug-and-play controllable generation. In: Proceedings of the International Conference on Machine Learning (ICML) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.529 + ], + "angle": 0, + "content": "43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Proceedings of the International Conference on Learning Representations (ICLR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.583 + ], + "angle": 0, + "content": "44. Sun, X., Li, W., Zhang, Z., Ma, Q., Sheng, X., Cheng, M., Ma, H., Zhao, S., Zhang, J., Li, J., et al.: Opdn: Omnidirectional position-aware deformable network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.61 + ], + "angle": 0, + "content": "45. Sun, Y., Lu, A., Yu, L.: Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters (SPL) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.651 + ], + "angle": 0, + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.678 + ], + "angle": 0, + "content": "47. Wang, J., Yue, Z., Zhou, S., Chan, K., Loy, C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.732 + ], + "angle": 0, + "content": "48. Wang, Q., Li, W., Mou, C., Cheng, X., Zhang, J.: 360dvd: Controllable panorama video generation with 360-degree video diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.773 + ], + "angle": 0, + "content": "49. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.813 + ], + "angle": 0, + "content": "50. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European Conference on Computer Vision Workshops (ECCVW) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "51. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "52. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.233 + ], + "angle": 0, + "content": "53. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.261 + ], + "angle": 0, + "content": "54. Yagi, Y.: Omnidirectional sensing and its applications. IEICE Transactions on Information and Systems (TOIS) (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.786, + 0.303 + ], + "angle": 0, + "content": "55. Yamazawa, K., Yagi, Y., Yachida, M.: Omnidirectional imaging with hyperboloidal projection. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.786, + 0.346 + ], + "angle": 0, + "content": "56. Yang, S., Zhou, Y., Liu, Z., Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: Proceedings of the SIGGRAPH Asia 2023 Conference Papers (SIGGRAPH Asia) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.786, + 0.374 + ], + "angle": 0, + "content": "57. Yang, S., Mou, C., Yu, J., Wang, Y., Meng, X., Zhang, J.: Neural video fields editing. arXiv preprint arXiv:2312.08882 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.786, + 0.417 + ], + "angle": 0, + "content": "58. Yinhuai, W., Yujie, H., Jiwen, Y., Jian, Z.: Gan prior based null-space learning for consistent super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.786, + 0.473 + ], + "angle": 0, + "content": "59. Yoon, Y., Chung, I., Wang, L., Yoon, K.J.: Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.786, + 0.53 + ], + "angle": 0, + "content": "60. Yu, F., Wang, X., Cao, M., Li, G., Shan, Y., Dong, C.: Osrt: Omnidirectional image super-resolution with distortion-aware transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.786, + 0.572 + ], + "angle": 0, + "content": "61. Yu, J., Zhang, X., Xu, Y., Zhang, J.: Cross: Diffusion model makes controllable, robust and secure image steganography. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.786, + 0.614 + ], + "angle": 0, + "content": "62. Yue, Z., Wang, J., Loy, C.C.: Resshift: Efficient diffusion model for image superresolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.786, + 0.657 + ], + "angle": 0, + "content": "63. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.786, + 0.699 + ], + "angle": 0, + "content": "64. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.786, + 0.742 + ], + "angle": 0, + "content": "65. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.786, + 0.798 + ], + "angle": 0, + "content": "66. Zhang, X., Zhang, Y., Xiong, R., Sun, Q., Zhang, J.: Herosnet: Hyperspectral explicable reconstruction and optimal sampling deep network for snapshot compressive imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "67. Zhou, Y., Yu, M., Ma, H., Shao, H., Jiang, G.: Weighted-to-spherically-uniform sism objective quality evaluation for panoramic video. In: Proceedings of the IEEE International Conference on Signal Processing (ICSP) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OmniSSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "68. Zongying, L., Hao, L., Liuzhenghao, L., Bin, L., Junwu, Z., Yu-Chian, C.C., Li, Y., Yonghong, T.: Taxdiff: Taxonomic-guided diffusion model for protein sequence generation. arXiv preprint arXiv:2402.17156 (2024)" + } + ] +] \ No newline at end of file diff --git a/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_origin.pdf b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c078cd381e442db044d193486d22eacab36a88f1 --- /dev/null +++ b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/7452c846-9323-4781-952d-a6062713a9d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e6f867b5dab93d19fd37f9f1065e7bcbde79efd23453b7dc3d89dcab46a7290 +size 19378180 diff --git a/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/full.md b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d67690cc3d9959f9bf5d012585ccce31a73aa5b9 --- /dev/null +++ b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/full.md @@ -0,0 +1,441 @@ +# OmniSSR: Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model + +Runyi Li†, Xuhan Sheng†, Weiqi Li, and Jian Zhang + +School of Electronic and Computer Engineering, Peking University, China {lirunyi, shengxuhan, liweiqi}@stu.pku.edu.cn {zhangjian.sz}@pku.edu.cn + +Abstract. Omnidirectional images (ODIs) are commonly used in real-world visual tasks, and high-resolution ODIs help improve the performance of related visual tasks. Most existing super-resolution methods for ODIs use end-to-end learning strategies, resulting in inferior realness of generated images and a lack of effective out-of-domain generalization capabilities in training methods. Image generation methods represented by diffusion model provide strong priors for visual tasks and have been proven to be effectively applied to image restoration tasks. Leveraging the image priors of the Stable Diffusion (SD) model, we achieve omnidirectional image Super Resolution with both fidelity and realness, dubbed as OmniSSR. Firstly, we transform the equirectangular projection (ERP) images into tangent projection (TP) images, whose distribution approximates the planar image domain. Then, we use SD to iteratively sample initial high-resolution results. At each denoising iteration, we further correct and update the initial results using the proposed Octadecaplex Tangent Information Interaction (OTII) and Gradient Decomposition (GD) technique to ensure better consistency. Finally, the TP images are transformed back to obtain the final high-resolution results. Our method is zero-shot, requiring no training or fine-tuning. Experiments of our method on two benchmark datasets demonstrate the effectiveness of our proposed method. + +Keywords: Omnidirectional Imaging $\cdot$ Super-Resolution $\cdot$ Latent Diffusion Model + +# 1 Introduction + +Omnidirectional images (ODIs) capture the entire scene in all directions, exceeding the narrow field of view (FOV) offered by planar images. Super-Resolution + +(SR) techniques enhance the visual quality of ODIs by increasing their resolution, thereby revealing finer details and enabling more accurate scene analysis and interpretation. This becomes particularly crucial in applications like virtual reality and surveillance, where high-resolution ODIs are essential for precise perception and decision-making. + +Current research in omnidirectional image super-resolution (ODISR) explores various methodologies to enhance the resolution of ODIs [14,36]. SphereSR [59] addresses non-uniformity in different projections by learning upsampling processes and ensuring information consistency using LIIF [5]. OSRT [60] designs a distortion-aware Transformer to modulate equirectangular projection (ERP) distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods remarkably. However, existing ODISR methods face the following challenges: (1) The majority are end-to-end models that can only produce a deterministic output, always better data fidelity but worse visual perception quality [17]. It's promising to develop a generation-based model, but requiring high data demands, yet high-resolution ODIs are high cost to collect [54,55]. (2) Most methods directly perform SR on ERP format ODIs, while users usually watch ODIs in a narrow FOV using tangent projection (TP). So another promising direction is to use off-the-shelf planar models on TP images. Recent times have witnessed the introduction and widespread application of diffusion models [23,43], especially Stable Diffusion (SD) [38], which have provided a robust backbone for visual tasks [21,24,56,61], including SR [30,40,47,51,52,62]. However, if TP images are trivially one-by-one processed using diffusion-based SR models, they will exhibit discrepancies in the overlapping region when reprojected onto the ERP image. As a result, the global continuity is compromised. + +Leveraging the strong image prior provided by SD, we propose the first diffusion-based zero-shot method for ODISR, named OmniSSR. Specifically, we propose Octadecaplex Tangent Information Interaction (OTII). OTII entails iterative conversion of intermediate SR results between ERP and TP, bridging the domain gap between ODIs and planar images. Building upon OTII, we further employ an approximate analytical solution of gradient descent, namely as Gradient Decomposition, to guide high-fidelity, high-quality omnidirectional image SR. By capitalizing on SD's effective image prior, our approach strikes a balance between fidelity and realness, ensuring that the restored ODIs exhibit both fidelity to the input data and realistic visual details. This method shows potential for advancing the current state of ODISR, providing improved resolution and visual quality across various applications. Fig. 1 showcases results fully demonstrating the superiority and performance of our proposed methods. + +Our main contributions are summarized as follows: + +- We propose OmniSSR, the first zero-shot ODISR method, using an off-the-shelf diffusion-based model, requiring no training or fine-tuning, leveraging existing image generation model priors to solve ODISR task. +- To bridge the domain gap between ODIs and planar images, we introduce Octadecaplex Tangent Information Interaction by repeatedly transforming + +![](images/6ac71c55133d5a4b08665715de6b12acbba8f34c2673ad66ddd81cf56ad040db.jpg) +(a) Input low-resolution ERP image + +![](images/190d9e5c51da8136830c6232a827c9aa1efaba08f16c4afe093ba374594cf240.jpg) +Degraded tangent image + +![](images/879565a0658a71bfcbda8b4f2d3b2bc3e444063d1c1fa0ab43f0587f20de7526.jpg) +Restored by OSRT + +![](images/aa214b3782fdb33ef8a9ee5f1a9ea71fae5259094da3e7cdfd81f20ffd36e2d5.jpg) +Restored by OmniSSR + +![](images/92878b1636165f5382bb60c5bd106c1975f4a9828b3623e262c61e3c2b49d295.jpg) +Ground truth + +![](images/aaa341522c4a9f9e98e480a239146ccd5e106e49683f721c8356fdf458d422fa.jpg) +(b) Input low-resolution ERP image + +![](images/0dd37afa508f60ab7420e295835f7c76128acc59f5e704052d4493526b2ceb94.jpg) +Degraded tangent image + +![](images/9f30918bea3777a60968a73c3de20b972eed6387a7cfbd63b062475e81a346fe.jpg) +Restored by OSRT +Fig. 1: We address omnidirectional image super-resolution in a zero-shot manner via OmniSSR. Presented above are select outcomes that sketch the essence of OmniSSR compared with current state-of-the-art approach OSRT [60]. Part (a) and (b) illustrate that OmniSSR upholds fidelity and visual realness at the same time, providing vivid and realistic details, while OSRT outputs over-smoothed and distorted results. Zoom in for more details. + +![](images/1658a722c4f092baf6dec82180441937b062474937d82f77bc9b31dbb8ed3ecf.jpg) +Restored by OmniSSR + +![](images/4d010a53b7a099e9c3a72d109fd332f0e962bc3679bd46ce293cb2f67d46c5cc.jpg) +Ground truth + +ODIs between ERP format and TP format, enabling ODISR task with pretrained diffusion models on planar images. + +- By iteratively updating images using the developed Gradient Decomposition technique, we introduce consistency constraints into the sampling process of the latent diffusion model, ensuring a trade-off between fidelity and realness in the reconstructed results. +- Extensive experiments are conducted on the benchmark datasets, demonstrating the superior performance of our method over existing state-of-the-art approaches, which validate the effectiveness of OmniSSR. + +# 2 Related Work + +# 2.1 Single Image Super-Resolution (SISR) + +Image super-resolution methods based on deep learning have undergone significant development over an extended period. Currently, they can be broadly classified into two categories of solutions. The first category involves end-to-end network training methods, which utilize image pairs consisting of low-resolution degraded images and high-resolution ground truth images for network training [6-8, 27, 31, 32, 63, 66]. The network architectures employed in this category include CNN [16], Transformers [46], etc. The second category employs image generation models as priors, such as GAN [20], diffusion models [23, 43], etc., where low-resolution images are used as conditions to generate high-resolution images. We will mainly introduce the methods using generative prior. + +Single Image SR using GAN prior In SR works utilizing GAN pri- ors [3,12,33,37,58], including real-world scenarios [8,49,50,65], pre-trained GAN + +networks are employed to transform image features into latent space, where the corresponding latent code for the high-resolution image is searched, ultimately yielding the reconstructed high-resolution result. + +Single Image SR using diffusion prior The diffusion model provides a powerful image prior, and the diffusion sampling process can generate highly realistic images. This strong prior distribution can be applied to various image restoration tasks, including super-resolution [9, 10, 19, 40, 42, 51]. Image-domain diffusion models directly provide prior distributions of image-domain data. DDNM [51] based on the mathematical method of Range-Null space Decomposition, iteratively refines content on the zero space, combining image prior content in the value domain to achieve image restoration. DDRM [25] uses SVD decomposition to obtain restoration results, which is similar to DDNM. DPS [9] transforms the image super-resolution problem into an optimization problem with consistency constraints, using gradient descent algorithms to guide the generation of image-domain diffusion models. GDP [19] further uses such gradient to update the degradation operator to tackle blind image inverse problems. Other methods including MCG [10], DDS [9] and unified control of diffusion generation [19, 42] use same strategy for image restoration, especially image super-resolution. + +# 2.2 Omnidirectional Image Super-Resolution + +Omnidirectional image super-resolution (ODISR) aims to enhance the resolution of omnidirectional or 360-degree images, which are commonly captured by cameras with a wide field of view. This field has garnered increasing attention due to its applications in virtual reality, omnidirectional video streaming, and surveillance. Several approaches have been proposed to address the unique challenges of ODISR [1, 2, 35, 44]. For instance, Kämäräinen et al. [18] propose a deep learning-based approach for omnidirectional super-resolution, leveraging convolutional neural networks to effectively upscale low-resolution omnidirectional images while preserving spatial details. Similarly, Smolic et al. [36] introduce a novel omnidirectional super-resolution algorithm utilizing generative adversarial networks (GANs) to enhance the visual quality of omnidirectional images by hallucinating high-frequency details. + +For evaluation purposes, researchers commonly utilize datasets such as the ODI-SR dataset from LAU-Net [13], and SUN 360 Panorama dataset [53]. These datasets enable the quantitative assessment of ODISR algorithms across various scenarios and facilitate fair comparisons between different methods. + +# 3 Method + +In this section, we first briefly introduce the preliminary background of our method (Sec. 3.1), and give an overall view of our proposed OmniSSR (Sec. 3.2). Then, we discuss the designs of Octadecaplex Tangent Information Interaction, which transform ODIs between ERP and TP formats with pre-upsampling strategy (Sec 3.3), and the Gradient Decomposition correction (Sec. 3.4). + +# 3.1 Preliminaries + +$\mathrm{ERP} \leftrightarrow \mathrm{TP}$ Transformation The essence of projection transformations between ERP and TP lie in determining the positions of target image pixels within the source image and computing their corresponding pixel values using interpolation algorithms, as digital images are always stored discretely [28]. Therefore, the $\mathrm{ERP} \rightarrow \mathrm{TP}$ transformation involves locating the TP image pixels on the ERP imaging plane, and vice versa. Gnomonic projection [11] provides the correspondence between ERP image pixels and TP image pixels. + +For a pixel $P_{e}(x_{e},y_{e})$ within the ERP image, we first find its corresponding pixel $P_{s}(\theta ,\phi)$ on the unit sphere using Eq. 1: + +$$ +\theta = 2 \pi x _ {e} / W, \phi = \pi y _ {e} / H, \tag {1} +$$ + +where $H$ and $W$ are the height and width of the ERP image. The Cartesian coordinates of the ERP image and the angular coordinates on the unit sphere exhibit a straightforward one-to-one linear relationship, suggesting a conceptual equivalence between these two projection formats. + +Given the spherical coordinates of the tangent plane center $(\theta_c, \phi_c)$ , The transformation from $P_s(\theta, \phi)$ to $P_t(x_t, y_t)$ , i.e. $\mathrm{ERP} \to \mathrm{TP}$ , is defined as: + +$$ +\begin{array}{l} x _ {t} = \left(\cos (\phi) \sin \left(\theta - \theta_ {c}\right)\right) / \zeta , \tag {2} \\ y _ {t} = \left(\cos \left(\phi_ {c}\right) \sin (\phi) - \sin \left(\phi_ {c}\right) \cos (\phi) \cos \left(\theta - \theta_ {c}\right)\right) / \zeta , \\ \end{array} +$$ + +where $\zeta = \sin (\phi_c)\sin (\phi) + \cos (\phi_c)\cos (\phi)\cos (\theta -\theta_c).$ + +The corresponding inverse transformation, i.e. $\mathrm{TP} \rightarrow \mathrm{ERP}$ , is: + +$$ +\begin{array}{l} \theta = \theta_ {c} + \arctan \left(\left(x _ {t} \sin (c)\right) / \left(\rho \cos \left(\phi_ {1}\right) \cos (c) - y _ {t} \sin \left(\phi_ {c}\right) \sin (c)\right)\right), \\ \phi = \arcsin \left(\cos (c) \sin \left(\phi_ {c}\right) + y _ {t} \sin (c) \cos \left(\phi_ {c}\right) / \rho\right), \tag {3} \\ \end{array} +$$ + +where $\rho = \sqrt{x_t^2 + y_t^2}$ and $c = \arctan (\rho)$ . + +With Eq. 2 and Eq. 3, we can build one-to-one forward and inverse mapping functions between pixels on the ERP image and pixels on the TP images. An illustration of the $\mathrm{ERP} \to \mathrm{TP}$ transformation is shown in Fig. 2(a). + +Iterative Denoising for Super-Resolution Utilizing the rich image priors provided by SD, we can super-resolve planar images. During initialization, the images are passed through the encoder $\mathcal{E}$ of SD to obtain latent codes, which are then added to pure noise to generate initial noise $\mathbf{z}_T$ . Following the approach proposed by StableSR [47], we pass the images through a time-aware adapter $\mathcal{T}$ . This adapter network structure is similar to the down-sampling part in denoising UNet, taking the image and the time step $t$ of diffusion sampling as inputs to obtain the latent code feature for step $t$ . This feature, along with the latent code $\mathbf{z}_t$ for each step and the time step $t$ , is then passed through denoising UNet to calculate the denoised result $\mathbf{z}_{0|t}$ and the latent code $\mathbf{z}_{t - 1}$ for the next sampling step. By iterating this process $T$ times, we can obtain the final super-resolution result via decoder $\mathcal{D}$ of SD, yielding high-resolution images. + +![](images/a4fd2290aad0f8599fc02ad333da083137fca0b2a1c9da880711a869ad92e8a6.jpg) +(a) Equirectangular Projection $\rightarrow$ Tangent Projection + +![](images/54e8b5067afbd7f4152bc0a06645d53a3b3ac2783204893c1eb94d54f11121e7.jpg) +(b) Pre-Upsampling Strategy +Fig. 2: Details about gnomonic transformations. (a) conversion from ERP to TP. (b) pre-upsampling proposed in Octadecaplex Tangent Information Interaction (Sec. 3.3) mitigating loss during transformation. + +# 3.2 Overview + +![](images/786d8ad01e6419d6a5b472aa6e445112d61640237c5fccdbbc4090fa2f2ffabb.jpg) +(b) Detailed illustration of GD Correction +Fig. 3: Overview of our proposed OmniSSR. Input low-resolution omnidirectional image $\mathbf{E}_{init}$ in ERP format is first projected onto Tangent Projection (TP) images $\mathbf{x}_{init}^{(1)},\mathbf{x}_{init}^{(2)},\dots,\mathbf{x}_{init}^{(m)}$ , then iteratively refined via Stable Diffusion (SD) with a time-aware adapter and controllable feature wrapping (CFW) module. In each step of diffusion sampling, we adopt the Gradient Decomposition (GD) correction technique to introduce consistency constraints for the restored intermediate results. After $T$ steps of sampling, we obtain the final result $\tilde{\mathbf{E}}_0$ with high resolution and better visual quality. + +Our approach can be divided into three parts. The first part is pre-processing, where we initially up-sample the low-resolution ERP images $\mathbf{E}_{init}$ with bicubic interpolation to target high-resolution size, then project them onto tangent planes to obtain a series of TP images. These TP images are transformed to the latent space by the SD encoder, iteratively processed through denoising UNet and time-aware adapter network, and then decoded to obtain high-resolution TP images. During each denoising step, these TP images are transformed back via inverse transformation to ERP images, employing the Gradient Decomposition correction to ensure consistency constraints in diffusion sampling. After $T$ + +
Algorithm 1: OmniSSR PipelineAlgorithm 2: Iterative Denois-ing with GD Correction
Input: Einit, F, F-1, A, A†, E, D, TInput: Einit, F, F-1, A, A†, E, D, T, Output: Latent code {z(1), z(2), ..., z(m)}
Output: SR result E0for t = T to 1 do
1{x(1), x(2), ..., x(m)} = F(Einit)for i = 1 to m do
2for i = 1 to m doεt = εθ(z(t), T(z(i), t), t)
3z(i)init = Ε(x(i))z(0|t) = 1/√αt(z(t)i - εt√1-αt)
4ε(i) ∼ N(0, I)x(0|i) = D(z(i))
5zT(i) = √αTz(i) + √1-αTε(i)end
6endE0|t = F-1{x(1), x(2), ..., x(m)}
7Get{z(0), z(2), ..., z(m)} from Algo. 2E0|t = E0|t + γeA†(Einit - AE0|t)
8for i = 1 to m do{x(1), x(2), ..., x(m)} = F(E0|t)
9x(0)i = D(z(0))for i = 1 to m do
10endz(0|i) = (1-γl)z(0|i) + γlE(x(i))
11E0 = F-1{x(0), x(2), ..., x(m)}z(t-1) ∼ p(z(t-1)|z(t,i), x(0|i))
12E0 = E0 + γpA†(Einit - AE0)end
13return E0
14end
15return {z(0), z(2), ..., z(m)}
+ +iterations, the final super-resolution result is obtained. A formulaic description for OmniSSR is shown in Algo. 1. Fig. 3 shows the overview of our pipeline. + +# 3.3 Octadecaplex Tangent Information Interaction (OTII) + +Motivation To apply SD for ODISR, a straightforward way is to perform the $\mathrm{ERP} \rightarrow \mathrm{TP}$ transformation on the input ERP image. Then, each obtained TP image is fed into the SD-based model for SR. Finally, the $\mathrm{TP} \rightarrow \mathrm{ERP}$ transformation yields the ultimate super-resolved ERP image. OmniFusion [28] employs a similar approach for depth estimation. However, this simplistic strategy fractures the inherent global coherence of ODIs, leading to pixel-level discontinuities in the fused ERP images. Moreover, interpolation algorithms cause significant information loss in the original projection transformations, resulting in more blurred images. If applied multiple times, this exacerbates the information loss even further. To mitigate this, a trivial solution is to increase the pixel count (resolution) of the intermediate projection imaging plane. However, excessively high resolutions in TP images can introduce unnecessary computational overhead during the denoising stage and potentially compromise the denoising performance. (see Supplementary Materials for details) + +Information Interaction and Pre-upsampling Based on the analysis of the Motivation in 3.3, we propose OTII by alternating the intermediate results between ERP and TP formats at each denoising step, where a single ERP image is represented by 18 TP images. From Sec. 3.1, we can achieve the $\mathrm{ERP}\rightarrow \mathrm{TP}$ transformation (denoted as $\mathcal{F}(\cdot)$ ) and the $\mathrm{TP}\rightarrow \mathrm{ERP}$ transformation (denoted as $\mathcal{F}^{-1}(\cdot)$ ). Through the $\mathrm{ERP}\rightarrow \mathrm{TP}$ transformation, we can convert distorted ERP + +images into TP images with content distributions that approximate those of planar images. This enables the use of the original SD super-resolution method for planar images. Conversely, through the $\mathrm{TP} \rightarrow \mathrm{ERP}$ transformation, we can fuse information between different TP images holistically, while providing ERP-format input for the subsequent GD Correction in Sec. 3.4. To handle information loss during projection transformation, we further propose to pre-upsample the source image before projection transformations, as shown in Fig. 2(b). Our experiments in Sec. 4.4 demonstrate that this pre-upsampling strategy can significantly mitigate the information loss caused by projection transformations. + +# 3.4 Gradient Decomposition (GD) Correction for Fidelity + +SD-based methods, as introduced in Sec. 3.1, can perform SR on sliced TP images. However, relying solely on the SR results from SD may lack consistency and fail to accurately preserve the original semantic information and details of the low-resolution image. To enhance the consistency of the SR results from SD, we opt to use convex optimization methods to iteratively refine them. Modeling the SR task as an image inverse problem, the following equation is formulated: + +$$ +\mathbf {y} = \mathbf {A x} + \mathbf {n}, \quad \mathbf {n} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), \tag {4} +$$ + +where $\mathbf{x}$ represents the original image, $\mathbf{y}$ denotes the degraded result, $\mathbf{A}$ is the degradation operator (e.g., bicubic downsampling for super-resolution), and $\mathbf{n}$ is random noise. The objective we aim to solve can be expressed as the following convex optimization problem: + +$$ +\underset {\mathbf {x}} {\operatorname {a r g m i n}} | | \mathbf {y} - \mathbf {A x} | | _ {2} ^ {2} + \lambda \mathcal {R} (\mathbf {x}), \tag {5} +$$ + +where the first term is the data-fidelity term, ensuring the consistency of image reconstruction, and the second term is the regulation term, ensuring the sparsity of the reconstruction result, thus making the image more realistic. The regularization term can be the 1-norm, Total Variation, etc. The aforementioned convex optimization problem can be solved using a series of algorithms, such as gradient descent, ADMM [4], etc. Considering the trade-off between time and performance, we turn to find a solution based on gradient descent, and provide an approximate analytical solution composed of a fidelity term and a realness term, named "Gradient Decomposition (GD)": + +$$ +\begin{array}{l} \tilde {\mathbf {E}} _ {0 | t} = \mathbf {E} _ {0 | t} + \alpha \nabla_ {\mathbf {E} _ {0 | t}} \left\| \mathbf {E} _ {i n i t} - \mathbf {A} \mathbf {E} _ {0 | t} \right\| _ {F} = \mathbf {E} _ {0 | t} + \alpha \times 2 \left(\mathbf {A} ^ {\dagger} \mathbf {E} _ {i n i t} - \mathbf {A} ^ {\dagger} \mathbf {A} \mathbf {E} _ {0 | t}\right) \\ = \mathbf {E} _ {0 | t} + \gamma \mathbf {A} ^ {\dagger} \left(\mathbf {E} _ {i n i t} - \mathbf {A} \mathbf {E} _ {0 | t}\right) = \gamma \mathbf {A} ^ {\dagger} \mathbf {E} _ {i n i t} + (\mathbf {I} - \gamma \mathbf {A} ^ {\dagger} \mathbf {A}) \mathbf {E} _ {0 | t} \tag {6} \\ \end{array} +$$ + +where $\mathbf{A}^{\dagger}$ denotes pseudo-inverse of degradation operator $\mathbf{A}$ , $\mathbf{E}_{init}$ denotes initial low-resolution ERP input, $\mathbf{E}_{0|t}$ denotes restored result by SD, $\tilde{\mathbf{E}}_{0|t}$ denotes corrected result by GD, $\alpha$ denotes the learning rate of gradient descent, and $\gamma$ + +denotes the simplified hyper-parameter which is further tuned using grid search. The final setting of $\gamma$ on different stages is shown in Sec. 4.1, and the ablation studies of parameter choice are in Sec. 4.4. + +This technique could be seen as a step of gradient descent optimization, and the optimized result could be decomposed of (1) $\gamma \mathbf{A}^{\dagger}\mathbf{E}_{init}$ , which ensures the consistency of the generated result, and (2) $(\mathbf{I} - \gamma \mathbf{A}^{\dagger}\mathbf{A})\mathbf{E}_{0|t}$ , which serves as the iteratively updated generated result to improve its realness; $\gamma$ is a hyperparameter balancing the data fidelity and visual quality. For a better diversity and generality of the SR process, we expand this solution to latent space, and obtain the denoising result from both denoising UNet and corrected TP images (Algo. 2 line 11). A more detailed understanding of the iterative denoising process and application of GD correction could be referred to Algo. 2. + +# 4 Experiments + +# 4.1 Implementation Details + +Datasets and Pretrained Models We choose the test set of ODI-SR dataset from LAU-Net [13] and SUN 360 Panorama dataset [53], comprising 97 and 100 omnidirectional images respectively, for experimental evaluation. The ground truth images are of size $1024 \times 2048$ pixels. In SR methods such as GDP [19] and PSLD [39] for planar images, we partitioned the images into several $256 \times 256$ patches and performed super-resolution on each patch individually. + +For pre-trained models, we adopt from StableSR [47], which provided a SR network for planar images based on SD. This network architecture includes a time-aware adapter, a controllable feature wrapping (CFW) module, and the original SD structure from HuggingFace. All of them are kept untrained in our proposed OmniSSR. + +Settings We set diffusion sampling steps to 200, which is the same as StableSR. The steps for other diffusion-based methods are set the same as their default settings (e.g. 1000 steps for PSLD). The degradation for low-resolution ERP images is bicubic down-sampling, and the implementation of its pseudo-inverse can be referred from code of DDRM $[25]^2$ . For choices of hyper-parameter $\gamma$ in GD correction, we set $\gamma_p = 1.0$ , $\gamma_e = 1.0$ , $\gamma_l = 0.5$ . Our code is developed via PyTorch on NVIDIA 3090Ti GPU. + +# 4.2 Comparison of OmniSSR with diffusion-based methods + +To evaluate the performance of proposed OmniSSR, we compare our method with recent state-of-the-art zero-shot methods for single image SR task: DPS [9], DDRM [25], GDP [19] which are based on the image-domain diffusion model, + +Table 1: SR results under bicubic downsampling on ODI-SR and SUN 360 Panorama datasets. For tasks not implemented in those papers, we mark N/A in the corresponding results. The best results are shown in bold, and the second best results are underlined. + +
MethodScaleODI-SRSUN 360 Panorama
WS-PSNR↑WS-SSIM↑FID↓LPIPS↓WS-PSNR↑WS-SSIM↑FID↓LPIPS↓
Bicubic28.140.834324.000.216428.670.853729.250.1933
DDRM [25]27.900.831712.280.166129.550.867013.100.1426
DPS [9]20.990.6194148.300.524921.440.6598148.830.5175
GDP [19]27.890.815726.560.272428.600.837628.020.2445
PSLD [39]N/AN/AN/AN/AN/AN/AN/AN/A
DiffIR [52]23.770.658357.230.468723.540.677558.060.4658
StableSR [47]22.700.645844.870.303923.300.690743.490.2858
OmniSSR28.570.854013.010.157529.690.878112.990.1459
Bicubic25.430.705950.840.375525.490.722955.990.3656
DDRM [25]25.430.736732.690.320625.830.744332.930.3304
DPS [9]24.750.6594120.740.491121.090.6119175.21430.5541
GDP [19]23.160.669277.430.426023.750.656990.230.4240
PSLD [39]21.720.5498107.990.532921.750.5828141.490.5461
DiffIR [52]24.010.677054.140.436723.900.701450.370.4235
StableSR [47]23.330.657749.950.313523.990.699846.030.3023
OmniSSR25.770.727930.970.297726.010.748134.580.2963
+ +and PSLD [39], which is based on latent diffusion model. We also choose supervised diffusion-based super-resolution approaches including StableSR [47] and DiffIR [52] for comparison. We conduct experiments on $\times 2$ and $\times 4$ SR with ERP bicubic downsampling, on ODI-SR test set and SUN test set. We choose WS-PSNR [45], WS-SSIM [67], FID [22], and LPIPS [64] as the main metrics. + +Quantitative results are presented in Tab. 1. With proposed OTII and GD correction, OmniSSR out-performs previous methods in terms of both Fidelity (from WS-PSNR and WS-SSIM) and Realness (from FID, LPIPS), which shows superior performance to existing diffusion-based methods for ODISR tasks on different scales. + +Qualitative results are shown in Fig. 4 and Fig. 5, which illustrates the visualization of SR results on SUN test set and ODI-SR test set with $\times 2$ and $\times 4$ scales, by different methods. + +The visual results indicate that our OmniSSR exhibits superior capability for detail recovery compared to other methods, particularly evident in textual elements (e.g., the text "flapping" in upper part of Fig. 4), complex objects (e.g., the black desk with a screen in lower part of Fig. 4, patterns above the white door in lower part of Fig. 5), and small-scale objects (e.g., the person and clock behind the desk in upper part of Fig. 5). OmniSSR demonstrates the ability to recover highly detailed and realistic visual effects from TP images. + +# 4.3 Comparison with end-to-end supervised methods + +The experiments of comparison in Sec. 4.2 are mainly focused on zero-shot image super-resolution methods, and supervised single image super-resolution methods, where the approaches are not trained or fine-tuned on omnidirectional images. In this part, we will compare OmniSSR to supervised end-to-end methods with end-to-end training on ODI datasets, including SwinIR and OSRT. Besides the + +![](images/e4c3ff6c545218f159267cef5882c1603f2c8bf24709a01975de6d37f76fc2e5.jpg) +$\therefore m - 1 \neq 0$ ; + +![](images/77002c1d00337de418c645adf178eaa35e56c1a3f5ee83e53bfeb849f3eaed87.jpg) +HR PSNR/SSIM + +![](images/05650ecbcf953f8c27820e324f67b077efd87da717d44393573f8cf043065397.jpg) + +![](images/118344129e4a2eb4ef49ca66cfce650818a4e189833fe66125991db2b6b1a1e0.jpg) +DPS [9] 24.02dB/0.5849 + +![](images/d982945947d1b2335bb0a61a5486f3b5c04193c4644c94c0a4f33c6ff494dbe5.jpg) +DDRM [25] 30.04dB/0.8855 + +![](images/82b8f5727c3feb46d1fab616286acba8ecf4e4d130640730b693e03e94627b25.jpg) +SUN 360 $(\times 2)$ : 001 + +![](images/0c3a37a737516d7755f59d9cc28cf2ea82c2e766d12fbad042740b8a943120a8.jpg) +GDP [19] 0.07dB/0.8802 + +![](images/51e080788bc5266f80850c4fd2c7d67f12d19c5ba9a9dd08ba45a05525c3fbc1.jpg) +Bicubic 28.67dB/0.8317 +DiffIR [52] 24.47dB/0.6421 + +![](images/d7f3e583f56f782601858dfe22fecbd03b12d177240296fb18e580a05c6757d9.jpg) +StableSR [47] 23.81dB/0.7384 + +![](images/026ddc985ebdab03ea056438df9d09337ff31b83aeb64778b6cc93f49748ddad.jpg) +OmniSSR(ours) 30.15dB/0.8859 + +![](images/a22c76f1a9c57cf9ec0cd05b2cf7728d59df730cc8b188461f2ea001da8ec4f1.jpg) +SUN 360 $(\times 4)$ : 009 +GDP [19] 23.02dB/0.7525 +Fig. 4: Visualized comparison of $\times 2$ and $\times 4$ SR results on SUN 360 test set. 001 and 009 are the ID numbers in the test set filenames. We also calculate the PSNR and SSIM to HR ground truth of each SR result and downsampled image. + +![](images/ad6731bd33a9a430eda461540342ab99da7f72d821a6de2efa230ce8f2f0bc06.jpg) + +![](images/fb91862be9975ce4a1bb80cd177b676301418460af1b0b6581f80b2790e37779.jpg) +DiffIR [52] 23.48dB/0.7799 + +![](images/d8c4ff0b5f831768bbd1ad40dfa23441b9d5b0b105e7250788095ac4b9050705.jpg) +StableSR [47] 24.62dB/0.7981 + +![](images/90267770031892416bc9530e51251ba52189be4ad3bb981dd57eed42669e6495.jpg) +OmniSSR(ours) 26.53dB/0.8265 + +main metrics in Sec. 4.2, we also use NIQE [34] and DISTS [15] to evaluate the visual perception of SR outputs. Results are presented in Tab. 2, which shows that although our OmniSSR exhibits inferior fidelity metrics compared to end-to-end supervised methods trained directly on ODI datasets, it demonstrates notable improvements in the visual quality and authenticity of super-resolved images. Notably, end-to-end methods often produce smoothed reconstructions with distortions, whereas our approach preserves finer details and adheres more closely to the realistic distribution. Considering that our method has never been trained or tuned on ODI datasets, nor having omnidirectional images prior, this result is acceptable. + +# 4.4 Ablation Studies + +We first sequentially validate the performance improvement of the proposed strategy in OmniSSR including input image type, OTII and GD correction, on the ODI-SR test set with $\times 2$ SR task, thereby demonstrating the significance of these strategies. The details are demonstrated as follows: + +![](images/0bdc65fb2176dac12f5298b3ca6d8fcb1ec664fb7773e237fbc775869930a5f4.jpg) +ODI-SR $(\times 2)$ :067 + +![](images/0752235e31564fb4f752b63de28f5a0ac4c30935b0e53ef6caf8179ebf4bf697.jpg) +HR + +![](images/cf16280fa802fd9d07b9ff4c5314099f7089002bddd8c755718783f03a4ed9e0.jpg) +Bicubic + +![](images/543e56ef5b919493be7a8272c2ca28f09f569e5d38f7712965492320a2909fd0.jpg) +DPS [9] + +![](images/527ed727580df771e6de1ebb3b5bef03127e6a1ae23209ef51281ef67497fac8.jpg) +PSNR/S +GDP [19] + +![](images/c3b5c7e640014d61d969b992d50e17ef486601a8713b4ea9cf721f5a150659aa.jpg) +27.67dB/0.8095 +DiffIR [52] + +![](images/fbb72815f1c1de0e2f751e61bdb32e69dd9832622603ed3647c242b650b3f5c7.jpg) +22.93dB/0.5653 +StableSR [47] + +![](images/359bdce6a7e8122ad244e0f897c40642adfd7cd00819502c5a6608ac55d62143.jpg) +DDRM [25] + +![](images/a39408904009596fae7d49e6a19b743fb7fff6944807c0196b6c4431195a9f4a.jpg) +29.91dB/0.8809 +OmniSSR(ours) + +![](images/d9dd5fa49be977880334bf244e60346700d2164088881783a7cfe8c0758cb98c.jpg) +ODI-SR $(\times 4)$ :049 +Fig. 5: Visualized comparison of $\times 2$ and $\times 4$ SR results on ODI-SR test set. 067 and 049 are the ID numbers in test set filenames. We also calculate the PSNR and SSIM between ground truth and each SR result as well as the downsampled image. + +![](images/93c9af5a620013e7780e5afa1e28077d1cec55b360e465d761e40a0ee6c62dfb.jpg) +HR + +![](images/1b960db6d669d53cef5bcf03dc771877d1cca6a1dcbdf7989e6b9516c3c4e692.jpg) +22.65dB/0.6248 +Bicubic + +![](images/665b7a80abefad187ae780411eb88428615dd5b47d299a62e7eef6793e7a5976.jpg) +21.80dB/0.5892 +DPS [9] + +26.21dB/0.7574 + +![](images/ceaf816b663668d5f27c284215943761f50a52998a3e3458c9f928edb319c082.jpg) +PSNR/SSIM +GDP [19] 24.15dB/0.7179 + +![](images/bfa6cc4eb8ac106c9f8c23112c1587b60c9c8ae66410ef9371f606367ada7017.jpg) +25.44dB/0.7536 +DiffIR [52] +.25dB/0.7594 + +26.21dB/0.7574 + +26.21dB/0.7574 + +![](images/2b9a5aceeddd7d72e94d00bc0cc4f0520d244efa9dd886d51156ed52d0128903.jpg) +29.99dB/0.8798 +DDRM [25] + +![](images/87f0ff13584b3ff238d34f0281a2ee9ae79b7b540089a9bb21222902908299ec.jpg) +27.12dB/0.8129 +OmniSSR(ours) +27.20dB/0.8168 + +1) we do not use any proposed strategy in the SR task, which is equivalent to the vanilla StableSR baseline; +2) we transform the degraded ERP image to TP images and feed them separately into StableSR pipeline, instead of directly inputting ERP images; +3) based on 2), we add OTII strategy during the denoising process of SD (Algo. 2 line 7); +4) based on 2), we add GD correction at the post-processing stage (Algo. 1 line 12) of the overall pipeline; +5) based on 3) and 4), we add GD correction at every step and post-processing stage of sampling, to improve the consistency of the restored result. + +Note that the execution of GD correction requires the execution of OTII in the denoising process simultaneously, there is no scenario where only GD correction is executed without the execution of OTII in the denoising process. + +Quantitative results of ablation studies are shown in Tab. 3. From the result shown below, we could come to the claim that the OTII helps improve the performance on the domain level, and the transformation between ERP and TP images provides information fusion among adjacent TP images. Our proposal of + +Table 2: Comparison on $\times 4$ SR task with supervised methods trained on ODI-SR dataset, including SwinIR and OSRT. The best results are shown in bold. + +
MethodDatasetWS-PSNR↑WS-SSIM↑FID↓LPIPS↓NIQE↓DISTS↓
SwinIR [29]26.760.762027.940.33215.39610.1710
OSRT [60]ODI-SR26.890.764627.390.32585.43640.1695
OmniSSR25.770.727930.970.29775.28910.1541
SwinIR [29]26.020.769239.900.34195.24400.1325
OSRT [60]SUN 36026.330.776639.220.33645.29840.1312
OmniSSR26.010.748134.580.29635.13290.1299
+ +Table 3: Ablation studies of OmniSSR on input type, OTII, and GD correction, on the test set of the ODI-SR dataset. Best results are shown in bold. + +
Input typeOTIIGD CorrectionWS-PSNR↑WS-SSIM↑FID↓LPIPS↓
ERP××22.690.645844.870.3039
TP××23.530.684943.910.3113
TP✓×23.740.684765.350.3748
TP×✓ (in post-process only)26.770.819215.410.1691
TP✓✓28.580.854013.010.1575
+ +Gradient Decomposition corrects such restoration result, improving fidelity and realness significantly at the same time, and it would be better if it is applied at each step of the overall denoising pipeline. Tab. 4 shows the effect of mitigating information loss via proposed pre-upsampling strategy. + +![](images/4da33366ec8ad894a06bf0a9b601184e2fc449d8ab20e8fa199af3a5e6d11cf6.jpg) +Fig.6: Ablation of choices on $\gamma_{p}$ , $\gamma_{e}$ and $\gamma_{l}$ . For better readability, WS-PSNR and LPIPS are chosen as evaluation metrics for fidelity and visual quality, respectively, to demonstrate the performance under different choices of the gamma parameter. We illustrate the results of (a) $\gamma_{p}$ and $\gamma_{e}$ fixed, while adjusting $\gamma_{l}$ ; (b) $\gamma_{e}$ and $\gamma_{l}$ fixed, while adjusting $\gamma_{p}$ ; (c) $\gamma_{p}$ and $\gamma_{l}$ fixed, while adjusting $\gamma_{e}$ . It can be observed that when $\gamma_{p} = 1$ , $\gamma_{e} = 1$ , and $\gamma_{l} = 0.5$ , OmniSSR achieves the relatively best performance. + +![](images/5ad62970d5261ee1e556964b7162f20ac9220f05687d019e45e9cd5af8a3b221.jpg) + +![](images/79cf981a76a17afe2ec979ca890f842edaa11cb91a5c60efe71ba9c4ef1a517e.jpg) + +For $\gamma$ in the GD correction technique, we use grid search to obtain better results on ODI-SR dataset and $\times 4$ SR task. Fig. 6 shows performance on different choices of $\gamma_{p}$ in Algo. 1 line 12, $\gamma_{e}$ in Algo. 2 line 8, and $\gamma_{l}$ in Algo. 2 line 11. The entire ablation of $\gamma_{p},\gamma_{e}$ and $\gamma_{l}$ , with WS-PSNR, WS-SSIM, FID and LPIPS score all calculated and compared, will be provided in Supplementary Materials. + +To evaluate the generalizability of our proposed modules, including Pre-Upsampling, OTII, and GD correction, we further conducted ablation studies + +Table 4: Results of pre-upsampling strategy on different scales, where $(x,y)$ denotes bicubic-based upsampling at $x\times$ scale to ERP before $\mathrm{ERP}\rightarrow \mathrm{TP}$ , and $y\times$ scale to TP before $\mathrm{TP}\rightarrow \mathrm{ERP}$ transformation. Best results are shown in bold. + +
\( \mathrm{{ERP}} \rightarrow \mathrm{{TP}} \rightarrow \mathrm{{ERP}} \)(1,1)(1,4)(4,1)(4,2)(2,4)(4,4)
WS-PSNR↑28.9838.1128.9933.9138.0538.18
WS-SSIM↑0.88590.98380.88620.96260.98370.9841
+ +on two super-resolution backbones, StableSR and SwinIR. The results underscore substantial performance enhancements facilitated by our modules across both backbones, which is provided in Supplementary Materials. + +# 5 Limitation and Discussion + +Although OmniSSR bridges the gap between omnidirectional and planar images, achieving competitive performance and better visual results in ODISR, it still exhibits the following limitations: (1) The inference of the diffusion model requires a considerable amount of time, approximately 14 minutes per ERP-formatted omnidirectional image to be super-resolved into size $1024 \times 2048$ , making real-time super-resolution challenging; (2) Multiple conversions between ERP and TP are required in the pipeline, leading to improved performance but consuming additional inference time; (3) Further exploration of the convex optimization properties of GD correction is warranted, such as designing gradient term coefficients adaptive to reconstruction results and degradation types. + +This study explores the application of image generation models to ODISR tasks. In future work, the framework behind OmniSSR can be extended beyond the confines of image super-resolution in a single scenario and venture into more complex ODI-based real-world scenarios. These include ODI editing, ODI inpainting, enhancing the quality of 3D Gaussian Splatting scenes [26,41] obtained after super-resolving ERP images, as well as enhancing the quality of omnidirectional videos [48], and other possible diffusion-based applications [57,68]. + +# 6 Conclusion + +This paper leverages the image prior of Stable Diffusion (SD) and employs the Octadecaplex Tangent Information Interaction (OTII) to achieve zero-shot omnidirectional image super-resolution. Additionally, we propose the Gradient Decomposition (GD) correction based on convex optimization algorithms to refine the initial super-resolution results, enhancing the fidelity and realness of the restored images. The superior performance of our proposed method, OmniSSR, is demonstrated on benchmark datasets. By bridging the gap between omnidirectional and planar images, we establish a training-free approach, mitigating the data demand and over-fitting associated with end-to-end training. The application scope of our method can be further extended to various applications, presenting potential value across multiple visual tasks. + +# References + +1. An, H., Zhang, X.: Perception-oriented omnidirectional image super-resolution based on transformer network. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2023) +2. Arican, Z., Frossard, P.: Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing (TIP) (2011) +3. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for image super-resolution and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022) +4. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging (2016) +5. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +6. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision (ICCV) (2023) +7. Cheng, M., Ma, H., Ma, Q., Sun, X., Li, W., Zhang, Z., Sheng, X., Zhao, S., Li, J., Zhang, L.: Hybrid transformer and cnn attention network for stereo image superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +8. Chong, M., Yanze, W., Xintao, W., Chao, D., Jian, Z., Ying, S.: Metric learning based interactive modulation for real-world super-resolution. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) +9. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022) +0. Chung, H., Sim, B., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2022) +1. Coxeter, H.S.M.: Introduction to geometry. John Wiley & Sons, Inc. (1961) +2. Daras, G., Dean, J., Jalal, A., Dimakis, A.: Intermediate layer optimization for inverse problems using deep generative models. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) +3. Deng, X., Wang, H., Xu, M., Guo, Y., Song, Y., Yang, L.: Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +4. Deng, X., Wang, H., Xu, M., Li, L., Wang, Z.: Omnidirectional image superresolution via latitude adaptive network. IEEE Transactions on Multimedia (TMM) (2022) +5. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020) +6. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2015) +7. Duan, H., Zhai, G., Min, X., Zhu, Y., Fang, Y., Yang, X.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the IEEE International Symposium on Circuits and Systems (ISCAS) (2018) + +18. Fakour-Sevom, V., Guldogan, E., Kämäräinen, J.K.: 360 panorama superresolution using deep convolutional networks. In: Proceedings of the Int. Conf. on Computer Vision Theory and Applications (VISAPP) (2018) +19. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2014) +21. Guo, L., Tao, T., Cai, X., Zhu, Z., Huang, J., Zhu, L., Gu, Z., Tang, H., Zhou, R., Han, S., et al.: Cas-diffcom: Cascaded diffusion model for infant longitudinal super-resolution 3d medical image completion. arXiv preprint arXiv:2402.13776 (2024) +22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017) +23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2020) +24. Jiang, Y., Chan, K.C., Wang, X., Loy, C.C., Liu, Z.: Reference-based image and video super-resolution via $c^2$ -matching. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022) +25. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of the ICLR Workshop on Deep Generative Models for Highly Structured Data (ICLRW) (2022) +26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) (2023) +27. Li, W., Chen, B., Zhang, J.: D3c2-net: Dual-domain deep convolutional coding network for compressive sensing. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) (2024) +28. Li, Y., Guo, Y., Yan, Z., Huang, X., Duan, Y., Ren, L.: Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +29. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (2021) +30. Liu, J., Wang, Q., Fan, H., Wang, Y., Tang, Y., Qu, L.: Residual denoising diffusion models. arXiv preprint arXiv:2308.13712 (2023) +31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2022) +32. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2020) +33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +34. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters (SPL) (2013) + +35. Nishiyama, A., Ikehata, S., Aizawa, K.: $360^{\circ}$ single image super resolution via distortion-aware network and distorted perspective images. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2021) +36. Ozcinar, C., Rana, A., Smolic, A.: Super-resolution of omnidirectional images using adversarial learning. In: Proceedings of the IEEE International Workshop on Multimedia Signal Processing (MMSPW) (2019) +37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021) +38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +39. Rout, L., Raoof, N., Daras, G., Caramanis, C., Dimakis, A., Shakkottai, S.: Solving linear inverse problems provably via posterior sampling with latent diffusion models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023) +40. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022) +41. Schonbein, M., Geiger, A.: Omnidirectional 3d reconstruction in augmented manhattan worlds. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014) +42. Song, J., Zhang, Q., Yin, H., Mardani, M., Liu, M.Y., Kautz, J., Chen, Y., Vahdat, A.: Loss-guided diffusion models for plug-and-play controllable generation. In: Proceedings of the International Conference on Machine Learning (ICML) (2023) +43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Proceedings of the International Conference on Learning Representations (ICLR) (2020) +44. Sun, X., Li, W., Zhang, Z., Ma, Q., Sheng, X., Cheng, M., Ma, H., Zhao, S., Zhang, J., Li, J., et al.: Opdn: Omnidirectional position-aware deformable network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2023) +45. Sun, Y., Lu, A., Yu, L.: Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters (SPL) (2017) +46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017) +47. Wang, J., Yue, Z., Zhou, S., Chan, K., Loy, C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023) +48. Wang, Q., Li, W., Mou, C., Cheng, X., Zhang, J.: 360dvd: Controllable panorama video generation with 360-degree video diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +49. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +50. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European Conference on Computer Vision Workshops (ECCVW) (2018) +51. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022) + +52. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +53. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2012) +54. Yagi, Y.: Omnidirectional sensing and its applications. IEICE Transactions on Information and Systems (TOIS) (1999) +55. Yamazawa, K., Yagi, Y., Yachida, M.: Omnidirectional imaging with hyperboloidal projection. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (1993) +56. Yang, S., Zhou, Y., Liu, Z., Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: Proceedings of the SIGGRAPH Asia 2023 Conference Papers (SIGGRAPH Asia) (2023) +57. Yang, S., Mou, C., Yu, J., Wang, Y., Meng, X., Zhang, J.: Neural video fields editing. arXiv preprint arXiv:2312.08882 (2023) +58. Yinhuai, W., Yujie, H., Jiwen, Y., Jian, Z.: Gan prior based null-space learning for consistent super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023) +59. Yoon, Y., Chung, I., Wang, L., Yoon, K.J.: Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +60. Yu, F., Wang, X., Cao, M., Li, G., Shan, Y., Dong, C.: Osrt: Omnidirectional image super-resolution with distortion-aware transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +61. Yu, J., Zhang, X., Xu, Y., Zhang, J.: Cross: Diffusion model makes controllable, robust and secure image steganography. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023) +62. Yue, Z., Wang, J., Loy, C.C.: Resshift: Efficient diffusion model for image superresolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) +63. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +64. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +65. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) +66. Zhang, X., Zhang, Y., Xiong, R., Sun, Q., Zhang, J.: Herosnet: Hyperspectral explicable reconstruction and optimal sampling deep network for snapshot compressive imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +67. Zhou, Y., Yu, M., Ma, H., Shao, H., Jiang, G.: Weighted-to-spherically-uniform sism objective quality evaluation for panoramic video. In: Proceedings of the IEEE International Conference on Signal Processing (ICSP) (2018) + +68. Zongying, L., Hao, L., Liuzhenghao, L., Bin, L., Junwu, Z., Yu-Chian, C.C., Li, Y., Yonghong, T.: Taxdiff: Taxonomic-guided diffusion model for protein sequence generation. arXiv preprint arXiv:2402.17156 (2024) \ No newline at end of file diff --git a/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/images.zip b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e712314d2d00fea67268966e9a9ffdc318af0f08 --- /dev/null +++ b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d4783c7ca056febbf45fbc7a9c3d3358f55b09578498d56256af3928a384a4d +size 729790 diff --git a/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/layout.json b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ef2278ea9ad155b0ff26aa6ca376102f15509c8b --- /dev/null +++ b/2024/OmniSSR_ Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model/layout.json @@ -0,0 +1,13743 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 470, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 470, + 147 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 470, + 147 + ], + "type": "text", + "content": "OmniSSR: Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 187, + 167, + 428, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 167, + 428, + 181 + ], + "spans": [ + { + "bbox": [ + 187, + 167, + 428, + 181 + ], + "type": "text", + "content": "Runyi Li†, Xuhan Sheng†, Weiqi Li, and Jian Zhang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 190, + 457, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 190, + 457, + 223 + ], + "spans": [ + { + "bbox": [ + 157, + 190, + 457, + 223 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Peking University, China {lirunyi, shengxuhan, liweiqi}@stu.pku.edu.cn {zhangjian.sz}@pku.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 255, + 452, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 255, + 452, + 484 + ], + "spans": [ + { + "bbox": [ + 162, + 255, + 452, + 484 + ], + "type": "text", + "content": "Abstract. Omnidirectional images (ODIs) are commonly used in real-world visual tasks, and high-resolution ODIs help improve the performance of related visual tasks. Most existing super-resolution methods for ODIs use end-to-end learning strategies, resulting in inferior realness of generated images and a lack of effective out-of-domain generalization capabilities in training methods. Image generation methods represented by diffusion model provide strong priors for visual tasks and have been proven to be effectively applied to image restoration tasks. Leveraging the image priors of the Stable Diffusion (SD) model, we achieve omnidirectional image Super Resolution with both fidelity and realness, dubbed as OmniSSR. Firstly, we transform the equirectangular projection (ERP) images into tangent projection (TP) images, whose distribution approximates the planar image domain. Then, we use SD to iteratively sample initial high-resolution results. At each denoising iteration, we further correct and update the initial results using the proposed Octadecaplex Tangent Information Interaction (OTII) and Gradient Decomposition (GD) technique to ensure better consistency. Finally, the TP images are transformed back to obtain the final high-resolution results. Our method is zero-shot, requiring no training or fine-tuning. Experiments of our method on two benchmark datasets demonstrate the effectiveness of our proposed method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 497, + 451, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 497, + 451, + 517 + ], + "spans": [ + { + "bbox": [ + 162, + 497, + 451, + 517 + ], + "type": "text", + "content": "Keywords: Omnidirectional Imaging " + }, + { + "bbox": [ + 162, + 497, + 451, + 517 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 162, + 497, + 451, + 517 + ], + "type": "text", + "content": " Super-Resolution " + }, + { + "bbox": [ + 162, + 497, + 451, + 517 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 162, + 497, + 451, + 517 + ], + "type": "text", + "content": " Latent Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 572, + 229, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 572, + 229, + 585 + ], + "spans": [ + { + "bbox": [ + 134, + 572, + 229, + 585 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 598, + 480, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 598, + 480, + 623 + ], + "spans": [ + { + "bbox": [ + 133, + 598, + 480, + 623 + ], + "type": "text", + "content": "Omnidirectional images (ODIs) capture the entire scene in all directions, exceeding the narrow field of view (FOV) offered by planar images. Super-Resolution" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 144, + 631, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 631, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 144, + 631, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 144, + 631, + 480, + 665 + ], + "type": "text", + "content": " Equal contributor " + }, + { + "bbox": [ + 144, + 631, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 144, + 631, + 480, + 665 + ], + "type": "text", + "content": " Corresponding author This work was supported in part by National Science Foundation of China (No. 62372016) and Shenzhen Research Project Grant (No. JSGGZD20220822095800001)." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "(SR) techniques enhance the visual quality of ODIs by increasing their resolution, thereby revealing finer details and enabling more accurate scene analysis and interpretation. This becomes particularly crucial in applications like virtual reality and surveillance, where high-resolution ODIs are essential for precise perception and decision-making." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 482, + 417 + ], + "type": "text", + "content": "Current research in omnidirectional image super-resolution (ODISR) explores various methodologies to enhance the resolution of ODIs [14,36]. SphereSR [59] addresses non-uniformity in different projections by learning upsampling processes and ensuring information consistency using LIIF [5]. OSRT [60] designs a distortion-aware Transformer to modulate equirectangular projection (ERP) distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods remarkably. However, existing ODISR methods face the following challenges: (1) The majority are end-to-end models that can only produce a deterministic output, always better data fidelity but worse visual perception quality [17]. It's promising to develop a generation-based model, but requiring high data demands, yet high-resolution ODIs are high cost to collect [54,55]. (2) Most methods directly perform SR on ERP format ODIs, while users usually watch ODIs in a narrow FOV using tangent projection (TP). So another promising direction is to use off-the-shelf planar models on TP images. Recent times have witnessed the introduction and widespread application of diffusion models [23,43], especially Stable Diffusion (SD) [38], which have provided a robust backbone for visual tasks [21,24,56,61], including SR [30,40,47,51,52,62]. However, if TP images are trivially one-by-one processed using diffusion-based SR models, they will exhibit discrepancies in the overlapping region when reprojected onto the ERP image. As a result, the global continuity is compromised." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 419, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 575 + ], + "type": "text", + "content": "Leveraging the strong image prior provided by SD, we propose the first diffusion-based zero-shot method for ODISR, named OmniSSR. Specifically, we propose Octadecaplex Tangent Information Interaction (OTII). OTII entails iterative conversion of intermediate SR results between ERP and TP, bridging the domain gap between ODIs and planar images. Building upon OTII, we further employ an approximate analytical solution of gradient descent, namely as Gradient Decomposition, to guide high-fidelity, high-quality omnidirectional image SR. By capitalizing on SD's effective image prior, our approach strikes a balance between fidelity and realness, ensuring that the restored ODIs exhibit both fidelity to the input data and realistic visual details. This method shows potential for advancing the current state of ODISR, providing improved resolution and visual quality across various applications. Fig. 1 showcases results fully demonstrating the superiority and performance of our proposed methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 576, + 372, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 576, + 372, + 587 + ], + "spans": [ + { + "bbox": [ + 146, + 576, + 372, + 587 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 603, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 603, + 481, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 603, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 138, + 603, + 481, + 639 + ], + "type": "text", + "content": "- We propose OmniSSR, the first zero-shot ODISR method, using an off-the-shelf diffusion-based model, requiring no training or fine-tuning, leveraging existing image generation model priors to solve ODISR task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "content": "- To bridge the domain gap between ODIs and planar images, we introduce Octadecaplex Tangent Information Interaction by repeatedly transforming" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 241, + 167 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 241, + 167 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 241, + 167 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 241, + 167 + ], + "type": "image", + "image_path": "6ac71c55133d5a4b08665715de6b12acbba8f34c2673ad66ddd81cf56ad040db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 168, + 226, + 175 + ], + "lines": [ + { + "bbox": [ + 136, + 168, + 226, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 168, + 226, + 175 + ], + "type": "text", + "content": "(a) Input low-resolution ERP image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 115, + 299, + 168 + ], + "blocks": [ + { + "bbox": [ + 246, + 115, + 299, + 168 + ], + "lines": [ + { + "bbox": [ + 246, + 115, + 299, + 168 + ], + "spans": [ + { + "bbox": [ + 246, + 115, + 299, + 168 + ], + "type": "image", + "image_path": "190d9e5c51da8136830c6232a827c9aa1efaba08f16c4afe093ba374594cf240.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 169, + 303, + 176 + ], + "lines": [ + { + "bbox": [ + 244, + 169, + 303, + 176 + ], + "spans": [ + { + "bbox": [ + 244, + 169, + 303, + 176 + ], + "type": "text", + "content": "Degraded tangent image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 115, + 359, + 168 + ], + "blocks": [ + { + "bbox": [ + 306, + 115, + 359, + 168 + ], + "lines": [ + { + "bbox": [ + 306, + 115, + 359, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 359, + 168 + ], + "type": "image", + "image_path": "879565a0658a71bfcbda8b4f2d3b2bc3e444063d1c1fa0ab43f0587f20de7526.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 169, + 355, + 176 + ], + "lines": [ + { + "bbox": [ + 310, + 169, + 355, + 176 + ], + "spans": [ + { + "bbox": [ + 310, + 169, + 355, + 176 + ], + "type": "text", + "content": "Restored by OSRT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 365, + 115, + 419, + 168 + ], + "blocks": [ + { + "bbox": [ + 365, + 115, + 419, + 168 + ], + "lines": [ + { + "bbox": [ + 365, + 115, + 419, + 168 + ], + "spans": [ + { + "bbox": [ + 365, + 115, + 419, + 168 + ], + "type": "image", + "image_path": "aa214b3782fdb33ef8a9ee5f1a9ea71fae5259094da3e7cdfd81f20ffd36e2d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 169, + 421, + 175 + ], + "lines": [ + { + "bbox": [ + 366, + 169, + 421, + 175 + ], + "spans": [ + { + "bbox": [ + 366, + 169, + 421, + 175 + ], + "type": "text", + "content": "Restored by OmniSSR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 115, + 480, + 168 + ], + "blocks": [ + { + "bbox": [ + 426, + 115, + 480, + 168 + ], + "lines": [ + { + "bbox": [ + 426, + 115, + 480, + 168 + ], + "spans": [ + { + "bbox": [ + 426, + 115, + 480, + 168 + ], + "type": "image", + "image_path": "92878b1636165f5382bb60c5bd106c1975f4a9828b3623e262c61e3c2b49d295.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 169, + 469, + 175 + ], + "lines": [ + { + "bbox": [ + 437, + 169, + 469, + 175 + ], + "spans": [ + { + "bbox": [ + 437, + 169, + 469, + 175 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 185, + 240, + 238 + ], + "blocks": [ + { + "bbox": [ + 133, + 185, + 240, + 238 + ], + "lines": [ + { + "bbox": [ + 133, + 185, + 240, + 238 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 240, + 238 + ], + "type": "image", + "image_path": "aaa341522c4a9f9e98e480a239146ccd5e106e49683f721c8356fdf458d422fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 239, + 226, + 247 + ], + "lines": [ + { + "bbox": [ + 136, + 239, + 226, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 239, + 226, + 247 + ], + "type": "text", + "content": "(b) Input low-resolution ERP image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 246, + 185, + 299, + 238 + ], + "blocks": [ + { + "bbox": [ + 246, + 185, + 299, + 238 + ], + "lines": [ + { + "bbox": [ + 246, + 185, + 299, + 238 + ], + "spans": [ + { + "bbox": [ + 246, + 185, + 299, + 238 + ], + "type": "image", + "image_path": "0dd37afa508f60ab7420e295835f7c76128acc59f5e704052d4493526b2ceb94.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 239, + 302, + 247 + ], + "lines": [ + { + "bbox": [ + 244, + 239, + 302, + 247 + ], + "spans": [ + { + "bbox": [ + 244, + 239, + 302, + 247 + ], + "type": "text", + "content": "Degraded tangent image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 306, + 185, + 359, + 238 + ], + "blocks": [ + { + "bbox": [ + 306, + 185, + 359, + 238 + ], + "lines": [ + { + "bbox": [ + 306, + 185, + 359, + 238 + ], + "spans": [ + { + "bbox": [ + 306, + 185, + 359, + 238 + ], + "type": "image", + "image_path": "9f30918bea3777a60968a73c3de20b972eed6387a7cfbd63b062475e81a346fe.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 239, + 356, + 247 + ], + "lines": [ + { + "bbox": [ + 310, + 239, + 356, + 247 + ], + "spans": [ + { + "bbox": [ + 310, + 239, + 356, + 247 + ], + "type": "text", + "content": "Restored by OSRT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 255, + 480, + 320 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 320 + ], + "type": "text", + "content": "Fig. 1: We address omnidirectional image super-resolution in a zero-shot manner via OmniSSR. Presented above are select outcomes that sketch the essence of OmniSSR compared with current state-of-the-art approach OSRT [60]. Part (a) and (b) illustrate that OmniSSR upholds fidelity and visual realness at the same time, providing vivid and realistic details, while OSRT outputs over-smoothed and distorted results. Zoom in for more details." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 365, + 185, + 419, + 238 + ], + "blocks": [ + { + "bbox": [ + 365, + 185, + 419, + 238 + ], + "lines": [ + { + "bbox": [ + 365, + 185, + 419, + 238 + ], + "spans": [ + { + "bbox": [ + 365, + 185, + 419, + 238 + ], + "type": "image", + "image_path": "1658a722c4f092baf6dec82180441937b062474937d82f77bc9b31dbb8ed3ecf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 239, + 421, + 246 + ], + "lines": [ + { + "bbox": [ + 366, + 239, + 421, + 246 + ], + "spans": [ + { + "bbox": [ + 366, + 239, + 421, + 246 + ], + "type": "text", + "content": "Restored by OmniSSR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 426, + 185, + 480, + 238 + ], + "blocks": [ + { + "bbox": [ + 426, + 185, + 480, + 238 + ], + "lines": [ + { + "bbox": [ + 426, + 185, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 426, + 185, + 480, + 238 + ], + "type": "image", + "image_path": "4d010a53b7a099e9c3a72d109fd332f0e962bc3679bd46ce293cb2f67d46c5cc.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 239, + 469, + 246 + ], + "lines": [ + { + "bbox": [ + 437, + 239, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 437, + 239, + 469, + 246 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 147, + 346, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 346, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 147, + 346, + 480, + 369 + ], + "type": "text", + "content": "ODIs between ERP format and TP format, enabling ODISR task with pretrained diffusion models on planar images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 370, + 481, + 453 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 138, + 370, + 481, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 417 + ], + "type": "text", + "content": "- By iteratively updating images using the developed Gradient Decomposition technique, we introduce consistency constraints into the sampling process of the latent diffusion model, ensuring a trade-off between fidelity and realness in the reconstructed results." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 138, + 418, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 453 + ], + "type": "text", + "content": "- Extensive experiments are conducted on the benchmark datasets, demonstrating the superior performance of our method over existing state-of-the-art approaches, which validate the effectiveness of OmniSSR." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 474, + 236, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 236, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 236, + 487 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 131, + 500, + 353, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 353, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 353, + 513 + ], + "type": "text", + "content": "2.1 Single Image Super-Resolution (SISR)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 521, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 640 + ], + "type": "text", + "content": "Image super-resolution methods based on deep learning have undergone significant development over an extended period. Currently, they can be broadly classified into two categories of solutions. The first category involves end-to-end network training methods, which utilize image pairs consisting of low-resolution degraded images and high-resolution ground truth images for network training [6-8, 27, 31, 32, 63, 66]. The network architectures employed in this category include CNN [16], Transformers [46], etc. The second category employs image generation models as priors, such as GAN [20], diffusion models [23, 43], etc., where low-resolution images are used as conditions to generate high-resolution images. We will mainly introduce the methods using generative prior." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Single Image SR using GAN prior In SR works utilizing GAN pri- ors [3,12,33,37,58], including real-world scenarios [8,49,50,65], pre-trained GAN" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "networks are employed to transform image features into latent space, where the corresponding latent code for the high-resolution image is searched, ultimately yielding the reconstructed high-resolution result." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 344 + ], + "type": "text", + "content": "Single Image SR using diffusion prior The diffusion model provides a powerful image prior, and the diffusion sampling process can generate highly realistic images. This strong prior distribution can be applied to various image restoration tasks, including super-resolution [9, 10, 19, 40, 42, 51]. Image-domain diffusion models directly provide prior distributions of image-domain data. DDNM [51] based on the mathematical method of Range-Null space Decomposition, iteratively refines content on the zero space, combining image prior content in the value domain to achieve image restoration. DDRM [25] uses SVD decomposition to obtain restoration results, which is similar to DDNM. DPS [9] transforms the image super-resolution problem into an optimization problem with consistency constraints, using gradient descent algorithms to guide the generation of image-domain diffusion models. GDP [19] further uses such gradient to update the degradation operator to tackle blind image inverse problems. Other methods including MCG [10], DDS [9] and unified control of diffusion generation [19, 42] use same strategy for image restoration, especially image super-resolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 358, + 367, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 367, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 367, + 371 + ], + "type": "text", + "content": "2.2 Omnidirectional Image Super-Resolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "type": "text", + "content": "Omnidirectional image super-resolution (ODISR) aims to enhance the resolution of omnidirectional or 360-degree images, which are commonly captured by cameras with a wide field of view. This field has garnered increasing attention due to its applications in virtual reality, omnidirectional video streaming, and surveillance. Several approaches have been proposed to address the unique challenges of ODISR [1, 2, 35, 44]. For instance, Kämäräinen et al. [18] propose a deep learning-based approach for omnidirectional super-resolution, leveraging convolutional neural networks to effectively upscale low-resolution omnidirectional images while preserving spatial details. Similarly, Smolic et al. [36] introduce a novel omnidirectional super-resolution algorithm utilizing generative adversarial networks (GANs) to enhance the visual quality of omnidirectional images by hallucinating high-frequency details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "text", + "content": "For evaluation purposes, researchers commonly utilize datasets such as the ODI-SR dataset from LAU-Net [13], and SUN 360 Panorama dataset [53]. These datasets enable the quantitative assessment of ODISR algorithms across various scenarios and facilitate fair comparisons between different methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 582, + 202, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 202, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 202, + 596 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "In this section, we first briefly introduce the preliminary background of our method (Sec. 3.1), and give an overall view of our proposed OmniSSR (Sec. 3.2). Then, we discuss the designs of Octadecaplex Tangent Information Interaction, which transform ODIs between ERP and TP formats with pre-upsampling strategy (Sec 3.3), and the Gradient Decomposition correction (Sec. 3.4)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 228, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 228, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 228, + 126 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP} \\leftrightarrow \\mathrm{TP}" + }, + { + "bbox": [ + 130, + 134, + 479, + 217 + ], + "type": "text", + "content": " Transformation The essence of projection transformations between ERP and TP lie in determining the positions of target image pixels within the source image and computing their corresponding pixel values using interpolation algorithms, as digital images are always stored discretely [28]. Therefore, the " + }, + { + "bbox": [ + 130, + 134, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP} \\rightarrow \\mathrm{TP}" + }, + { + "bbox": [ + 130, + 134, + 479, + 217 + ], + "type": "text", + "content": " transformation involves locating the TP image pixels on the ERP imaging plane, and vice versa. Gnomonic projection [11] provides the correspondence between ERP image pixels and TP image pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 219, + 480, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 480, + 243 + ], + "type": "text", + "content": "For a pixel " + }, + { + "bbox": [ + 130, + 219, + 480, + 243 + ], + "type": "inline_equation", + "content": "P_{e}(x_{e},y_{e})" + }, + { + "bbox": [ + 130, + 219, + 480, + 243 + ], + "type": "text", + "content": " within the ERP image, we first find its corresponding pixel " + }, + { + "bbox": [ + 130, + 219, + 480, + 243 + ], + "type": "inline_equation", + "content": "P_{s}(\\theta ,\\phi)" + }, + { + "bbox": [ + 130, + 219, + 480, + 243 + ], + "type": "text", + "content": " on the unit sphere using Eq. 1:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 251, + 480, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 251, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 251, + 251, + 480, + 263 + ], + "type": "interline_equation", + "content": "\\theta = 2 \\pi x _ {e} / W, \\phi = \\pi y _ {e} / H, \\tag {1}", + "image_path": "521ea8765e484454ffae7d65c1a6da755b80871c60f4d1f3038fe34e0e0d05be.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 270, + 479, + 317 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 270, + 479, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 479, + 317 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 270, + 479, + 317 + ], + "type": "text", + "content": " are the height and width of the ERP image. The Cartesian coordinates of the ERP image and the angular coordinates on the unit sphere exhibit a straightforward one-to-one linear relationship, suggesting a conceptual equivalence between these two projection formats." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "text", + "content": "Given the spherical coordinates of the tangent plane center " + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "inline_equation", + "content": "(\\theta_c, \\phi_c)" + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "text", + "content": ", The transformation from " + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "inline_equation", + "content": "P_s(\\theta, \\phi)" + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "inline_equation", + "content": "P_t(x_t, y_t)" + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP} \\to \\mathrm{TP}" + }, + { + "bbox": [ + 130, + 319, + 479, + 342 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 194, + 350, + 480, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 350, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 194, + 350, + 480, + 380 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t} = \\left(\\cos (\\phi) \\sin \\left(\\theta - \\theta_ {c}\\right)\\right) / \\zeta , \\tag {2} \\\\ y _ {t} = \\left(\\cos \\left(\\phi_ {c}\\right) \\sin (\\phi) - \\sin \\left(\\phi_ {c}\\right) \\cos (\\phi) \\cos \\left(\\theta - \\theta_ {c}\\right)\\right) / \\zeta , \\\\ \\end{array}", + "image_path": "32323d3301edd6f7d10cd02c6653386a55b57abcf0fadc25883805cf4e84e86a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 387, + 362, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 362, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 362, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 387, + 362, + 399 + ], + "type": "inline_equation", + "content": "\\zeta = \\sin (\\phi_c)\\sin (\\phi) + \\cos (\\phi_c)\\cos (\\phi)\\cos (\\theta -\\theta_c)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 399, + 414, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 399, + 414, + 411 + ], + "spans": [ + { + "bbox": [ + 146, + 399, + 414, + 411 + ], + "type": "text", + "content": "The corresponding inverse transformation, i.e. " + }, + { + "bbox": [ + 146, + 399, + 414, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{TP} \\rightarrow \\mathrm{ERP}" + }, + { + "bbox": [ + 146, + 399, + 414, + 411 + ], + "type": "text", + "content": ", is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 167, + 418, + 480, + 448 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 418, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 167, + 418, + 480, + 448 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta = \\theta_ {c} + \\arctan \\left(\\left(x _ {t} \\sin (c)\\right) / \\left(\\rho \\cos \\left(\\phi_ {1}\\right) \\cos (c) - y _ {t} \\sin \\left(\\phi_ {c}\\right) \\sin (c)\\right)\\right), \\\\ \\phi = \\arcsin \\left(\\cos (c) \\sin \\left(\\phi_ {c}\\right) + y _ {t} \\sin (c) \\cos \\left(\\phi_ {c}\\right) / \\rho\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "ed3a31b92e3e7f282ab04d6f5538006aa324da78b15b8f2f6c9465a2fe2f87ac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 456, + 307, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 307, + 469 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 307, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 456, + 307, + 469 + ], + "type": "inline_equation", + "content": "\\rho = \\sqrt{x_t^2 + y_t^2}" + }, + { + "bbox": [ + 131, + 456, + 307, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 456, + 307, + 469 + ], + "type": "inline_equation", + "content": "c = \\arctan (\\rho)" + }, + { + "bbox": [ + 131, + 456, + 307, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 470, + 479, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 479, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 479, + 505 + ], + "type": "text", + "content": "With Eq. 2 and Eq. 3, we can build one-to-one forward and inverse mapping functions between pixels on the ERP image and pixels on the TP images. An illustration of the " + }, + { + "bbox": [ + 130, + 470, + 479, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP} \\to \\mathrm{TP}" + }, + { + "bbox": [ + 130, + 470, + 479, + 505 + ], + "type": "text", + "content": " transformation is shown in Fig. 2(a)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Iterative Denoising for Super-Resolution Utilizing the rich image priors provided by SD, we can super-resolve planar images. During initialization, the images are passed through the encoder " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of SD to obtain latent codes, which are then added to pure noise to generate initial noise " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". Following the approach proposed by StableSR [47], we pass the images through a time-aware adapter " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". This adapter network structure is similar to the down-sampling part in denoising UNet, taking the image and the time step " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of diffusion sampling as inputs to obtain the latent code feature for step " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". This feature, along with the latent code " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " for each step and the time step " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", is then passed through denoising UNet to calculate the denoised result " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0|t}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and the latent code " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t - 1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " for the next sampling step. By iterating this process " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " times, we can obtain the final super-resolution result via decoder " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of SD, yielding high-resolution images." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 333, + 217 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 333, + 217 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 333, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 333, + 217 + ], + "type": "image", + "image_path": "a4fd2290aad0f8599fc02ad333da083137fca0b2a1c9da880711a869ad92e8a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 219, + 294, + 227 + ], + "lines": [ + { + "bbox": [ + 171, + 219, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 171, + 219, + 294, + 227 + ], + "type": "text", + "content": "(a) Equirectangular Projection " + }, + { + "bbox": [ + 171, + 219, + 294, + 227 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 171, + 219, + 294, + 227 + ], + "type": "text", + "content": " Tangent Projection" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 337, + 114, + 480, + 217 + ], + "blocks": [ + { + "bbox": [ + 337, + 114, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 337, + 114, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 337, + 114, + 480, + 217 + ], + "type": "image", + "image_path": "54e8b5067afbd7f4152bc0a06645d53a3b3ac2783204893c1eb94d54f11121e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 219, + 443, + 228 + ], + "lines": [ + { + "bbox": [ + 375, + 219, + 443, + 228 + ], + "spans": [ + { + "bbox": [ + 375, + 219, + 443, + 228 + ], + "type": "text", + "content": "(b) Pre-Upsampling Strategy" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 236, + 479, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 269 + ], + "type": "text", + "content": "Fig. 2: Details about gnomonic transformations. (a) conversion from ERP to TP. (b) pre-upsampling proposed in Octadecaplex Tangent Information Interaction (Sec. 3.3) mitigating loss during transformation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 291, + 208, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 208, + 302 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 208, + 302 + ], + "type": "text", + "content": "3.2 Overview" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 325, + 482, + 445 + ], + "blocks": [ + { + "bbox": [ + 133, + 325, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 133, + 325, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 325, + 482, + 445 + ], + "type": "image", + "image_path": "786d8ad01e6419d6a5b472aa6e445112d61640237c5fccdbbc4090fa2f2ffabb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 445, + 345, + 451 + ], + "lines": [ + { + "bbox": [ + 249, + 445, + 345, + 451 + ], + "spans": [ + { + "bbox": [ + 249, + 445, + 345, + 451 + ], + "type": "text", + "content": "(b) Detailed illustration of GD Correction" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "text", + "content": "Fig. 3: Overview of our proposed OmniSSR. Input low-resolution omnidirectional image " + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{init}" + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "text", + "content": " in ERP format is first projected onto Tangent Projection (TP) images " + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{init}^{(1)},\\mathbf{x}_{init}^{(2)},\\dots,\\mathbf{x}_{init}^{(m)}" + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "text", + "content": ", then iteratively refined via Stable Diffusion (SD) with a time-aware adapter and controllable feature wrapping (CFW) module. In each step of diffusion sampling, we adopt the Gradient Decomposition (GD) correction technique to introduce consistency constraints for the restored intermediate results. After " + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "text", + "content": " steps of sampling, we obtain the final result " + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_0" + }, + { + "bbox": [ + 130, + 460, + 482, + 540 + ], + "type": "text", + "content": " with high resolution and better visual quality." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Our approach can be divided into three parts. The first part is pre-processing, where we initially up-sample the low-resolution ERP images " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{init}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " with bicubic interpolation to target high-resolution size, then project them onto tangent planes to obtain a series of TP images. These TP images are transformed to the latent space by the SD encoder, iteratively processed through denoising UNet and time-aware adapter network, and then decoded to obtain high-resolution TP images. During each denoising step, these TP images are transformed back via inverse transformation to ERP images, employing the Gradient Decomposition correction to ensure consistency constraints in diffusion sampling. After " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 131, + 116, + 471, + 327 + ], + "blocks": [ + { + "bbox": [ + 131, + 116, + 471, + 327 + ], + "lines": [ + { + "bbox": [ + 131, + 116, + 471, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 471, + 327 + ], + "type": "table", + "html": "
Algorithm 1: OmniSSR PipelineAlgorithm 2: Iterative Denois-ing with GD Correction
Input: Einit, F, F-1, A, A†, E, D, TInput: Einit, F, F-1, A, A†, E, D, T, Output: Latent code {z(1), z(2), ..., z(m)}
Output: SR result E0for t = T to 1 do
1{x(1), x(2), ..., x(m)} = F(Einit)for i = 1 to m do
2for i = 1 to m doεt = εθ(z(t), T(z(i), t), t)
3z(i)init = Ε(x(i))z(0|t) = 1/√αt(z(t)i - εt√1-αt)
4ε(i) ∼ N(0, I)x(0|i) = D(z(i))
5zT(i) = √αTz(i) + √1-αTε(i)end
6endE0|t = F-1{x(1), x(2), ..., x(m)}
7Get{z(0), z(2), ..., z(m)} from Algo. 2E0|t = E0|t + γeA†(Einit - AE0|t)
8for i = 1 to m do{x(1), x(2), ..., x(m)} = F(E0|t)
9x(0)i = D(z(0))for i = 1 to m do
10endz(0|i) = (1-γl)z(0|i) + γlE(x(i))
11E0 = F-1{x(0), x(2), ..., x(m)}z(t-1) ∼ p(z(t-1)|z(t,i), x(0|i))
12E0 = E0 + γpA†(Einit - AE0)end
13return E0
14end
15return {z(0), z(2), ..., z(m)}
", + "image_path": "ecadd64d7687b53a39fb41e296ffa42a6ef7114d6b0e3f893506ebd26b6a97fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 479, + 372 + ], + "type": "text", + "content": "iterations, the final super-resolution result is obtained. A formulaic description for OmniSSR is shown in Algo. 1. Fig. 3 shows the overview of our pipeline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 388, + 432, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 432, + 402 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 432, + 402 + ], + "type": "text", + "content": "3.3 Octadecaplex Tangent Information Interaction (OTII)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 577 + ], + "type": "text", + "content": "Motivation To apply SD for ODISR, a straightforward way is to perform the " + }, + { + "bbox": [ + 130, + 408, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP} \\rightarrow \\mathrm{TP}" + }, + { + "bbox": [ + 130, + 408, + 482, + 577 + ], + "type": "text", + "content": " transformation on the input ERP image. Then, each obtained TP image is fed into the SD-based model for SR. Finally, the " + }, + { + "bbox": [ + 130, + 408, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\mathrm{TP} \\rightarrow \\mathrm{ERP}" + }, + { + "bbox": [ + 130, + 408, + 482, + 577 + ], + "type": "text", + "content": " transformation yields the ultimate super-resolved ERP image. OmniFusion [28] employs a similar approach for depth estimation. However, this simplistic strategy fractures the inherent global coherence of ODIs, leading to pixel-level discontinuities in the fused ERP images. Moreover, interpolation algorithms cause significant information loss in the original projection transformations, resulting in more blurred images. If applied multiple times, this exacerbates the information loss even further. To mitigate this, a trivial solution is to increase the pixel count (resolution) of the intermediate projection imaging plane. However, excessively high resolutions in TP images can introduce unnecessary computational overhead during the denoising stage and potentially compromise the denoising performance. (see Supplementary Materials for details)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Information Interaction and Pre-upsampling Based on the analysis of the Motivation in 3.3, we propose OTII by alternating the intermediate results between ERP and TP formats at each denoising step, where a single ERP image is represented by 18 TP images. From Sec. 3.1, we can achieve the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP}\\rightarrow \\mathrm{TP}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " transformation (denoted as " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ") and the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{TP}\\rightarrow \\mathrm{ERP}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " transformation (denoted as " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\cdot)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "). Through the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP}\\rightarrow \\mathrm{TP}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " transformation, we can convert distorted ERP" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "images into TP images with content distributions that approximate those of planar images. This enables the use of the original SD super-resolution method for planar images. Conversely, through the " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{TP} \\rightarrow \\mathrm{ERP}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " transformation, we can fuse information between different TP images holistically, while providing ERP-format input for the subsequent GD Correction in Sec. 3.4. To handle information loss during projection transformation, we further propose to pre-upsample the source image before projection transformations, as shown in Fig. 2(b). Our experiments in Sec. 4.4 demonstrate that this pre-upsampling strategy can significantly mitigate the information loss caused by projection transformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 242, + 432, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 432, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 432, + 255 + ], + "type": "text", + "content": "3.4 Gradient Decomposition (GD) Correction for Fidelity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 335 + ], + "type": "text", + "content": "SD-based methods, as introduced in Sec. 3.1, can perform SR on sliced TP images. However, relying solely on the SR results from SD may lack consistency and fail to accurately preserve the original semantic information and details of the low-resolution image. To enhance the consistency of the SR results from SD, we opt to use convex optimization methods to iteratively refine them. Modeling the SR task as an image inverse problem, the following equation is formulated:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 245, + 345, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 345, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 245, + 345, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathbf {A x} + \\mathbf {n}, \\quad \\mathbf {n} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {4}", + "image_path": "77544c469312bd8c1ba573b93b695c89d4ebde06bd876ecc45b42bf4c42ff69a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "text", + "content": " represents the original image, " + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "text", + "content": " denotes the degraded result, " + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "text", + "content": " is the degradation operator (e.g., bicubic downsampling for super-resolution), and " + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 366, + 482, + 415 + ], + "type": "text", + "content": " is random noise. The objective we aim to solve can be expressed as the following convex optimization problem:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 424, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 424, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 246, + 424, + 481, + 444 + ], + "type": "interline_equation", + "content": "\\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} | | \\mathbf {y} - \\mathbf {A x} | | _ {2} ^ {2} + \\lambda \\mathcal {R} (\\mathbf {x}), \\tag {5}", + "image_path": "9ef7ebe0e4f57e2c507f942635ed380e2151c730e340d357accb68d1a0f01035.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 482, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 560 + ], + "type": "text", + "content": "where the first term is the data-fidelity term, ensuring the consistency of image reconstruction, and the second term is the regulation term, ensuring the sparsity of the reconstruction result, thus making the image more realistic. The regularization term can be the 1-norm, Total Variation, etc. The aforementioned convex optimization problem can be solved using a series of algorithms, such as gradient descent, ADMM [4], etc. Considering the trade-off between time and performance, we turn to find a solution based on gradient descent, and provide an approximate analytical solution composed of a fidelity term and a realness term, named \"Gradient Decomposition (GD)\":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 569, + 481, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 569, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 143, + 569, + 481, + 610 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mathbf {E}} _ {0 | t} = \\mathbf {E} _ {0 | t} + \\alpha \\nabla_ {\\mathbf {E} _ {0 | t}} \\left\\| \\mathbf {E} _ {i n i t} - \\mathbf {A} \\mathbf {E} _ {0 | t} \\right\\| _ {F} = \\mathbf {E} _ {0 | t} + \\alpha \\times 2 \\left(\\mathbf {A} ^ {\\dagger} \\mathbf {E} _ {i n i t} - \\mathbf {A} ^ {\\dagger} \\mathbf {A} \\mathbf {E} _ {0 | t}\\right) \\\\ = \\mathbf {E} _ {0 | t} + \\gamma \\mathbf {A} ^ {\\dagger} \\left(\\mathbf {E} _ {i n i t} - \\mathbf {A} \\mathbf {E} _ {0 | t}\\right) = \\gamma \\mathbf {A} ^ {\\dagger} \\mathbf {E} _ {i n i t} + (\\mathbf {I} - \\gamma \\mathbf {A} ^ {\\dagger} \\mathbf {A}) \\mathbf {E} _ {0 | t} \\tag {6} \\\\ \\end{array}", + "image_path": "4f203f65ab6f3b327a96d2d35351576b3028226f52aae83c5cf536f0ba44cca2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{\\dagger}" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " denotes pseudo-inverse of degradation operator " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{init}" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " denotes initial low-resolution ERP input, " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{0|t}" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " denotes restored result by SD, " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{0|t}" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " denotes corrected result by GD, " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " denotes the learning rate of gradient descent, and " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 405, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 405, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 405, + 666 + ], + "type": "text", + "content": "This claim will be further illustrated in subsequent experiments." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "denotes the simplified hyper-parameter which is further tuned using grid search. The final setting of " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " on different stages is shown in Sec. 4.1, and the ablation studies of parameter choice are in Sec. 4.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": "This technique could be seen as a step of gradient descent optimization, and the optimized result could be decomposed of (1) " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\gamma \\mathbf{A}^{\\dagger}\\mathbf{E}_{init}" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": ", which ensures the consistency of the generated result, and (2) " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "(\\mathbf{I} - \\gamma \\mathbf{A}^{\\dagger}\\mathbf{A})\\mathbf{E}_{0|t}" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": ", which serves as the iteratively updated generated result to improve its realness; " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " is a hyperparameter balancing the data fidelity and visual quality. For a better diversity and generality of the SR process, we expand this solution to latent space, and obtain the denoising result from both denoising UNet and corrected TP images (Algo. 2 line 11). A more detailed understanding of the iterative denoising process and application of GD correction could be referred to Algo. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 279, + 230, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 279, + 230, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 230, + 293 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 305, + 279, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 305, + 279, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 279, + 317 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "content": "Datasets and Pretrained Models We choose the test set of ODI-SR dataset from LAU-Net [13] and SUN 360 Panorama dataset [53], comprising 97 and 100 omnidirectional images respectively, for experimental evaluation. The ground truth images are of size " + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "inline_equation", + "content": "1024 \\times 2048" + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "content": " pixels. In SR methods such as GDP [19] and PSLD [39] for planar images, we partitioned the images into several " + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 325, + 481, + 396 + ], + "type": "text", + "content": " patches and performed super-resolution on each patch individually." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 481, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 481, + 458 + ], + "type": "text", + "content": "For pre-trained models, we adopt from StableSR [47], which provided a SR network for planar images based on SD. This network architecture includes a time-aware adapter, a controllable feature wrapping (CFW) module, and the original SD structure from HuggingFace. All of them are kept untrained in our proposed OmniSSR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": "Settings We set diffusion sampling steps to 200, which is the same as StableSR. The steps for other diffusion-based methods are set the same as their default settings (e.g. 1000 steps for PSLD). The degradation for low-resolution ERP images is bicubic down-sampling, and the implementation of its pseudo-inverse can be referred from code of DDRM " + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "[25]^2" + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": ". For choices of hyper-parameter " + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": " in GD correction, we set " + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\gamma_p = 1.0" + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\gamma_e = 1.0" + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\gamma_l = 0.5" + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": ". Our code is developed via PyTorch on NVIDIA 3090Ti GPU." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 578, + 438, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 578, + 438, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 578, + 438, + 590 + ], + "type": "text", + "content": "4.2 Comparison of OmniSSR with diffusion-based methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 598, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 481, + 635 + ], + "type": "text", + "content": "To evaluate the performance of proposed OmniSSR, we compare our method with recent state-of-the-art zero-shot methods for single image SR task: DPS [9], DDRM [25], GDP [19] which are based on the image-domain diffusion model," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 306, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 306, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 306, + 654 + ], + "type": "text", + "content": "2 https://github.com/bahjat-kawar/ddrm" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 654, + 369, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 369, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 369, + 665 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 654, + 369, + 665 + ], + "type": "text", + "content": " Code is at https://github.com/LiRunyi2001/OmniSSR." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 158, + 479, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: SR results under bicubic downsampling on ODI-SR and SUN 360 Panorama datasets. For tasks not implemented in those papers, we mark N/A in the corresponding results. The best results are shown in bold, and the second best results are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 158, + 479, + 308 + ], + "lines": [ + { + "bbox": [ + 135, + 158, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 135, + 158, + 479, + 308 + ], + "type": "table", + "html": "
MethodScaleODI-SRSUN 360 Panorama
WS-PSNR↑WS-SSIM↑FID↓LPIPS↓WS-PSNR↑WS-SSIM↑FID↓LPIPS↓
Bicubic28.140.834324.000.216428.670.853729.250.1933
DDRM [25]27.900.831712.280.166129.550.867013.100.1426
DPS [9]20.990.6194148.300.524921.440.6598148.830.5175
GDP [19]27.890.815726.560.272428.600.837628.020.2445
PSLD [39]N/AN/AN/AN/AN/AN/AN/AN/A
DiffIR [52]23.770.658357.230.468723.540.677558.060.4658
StableSR [47]22.700.645844.870.303923.300.690743.490.2858
OmniSSR28.570.854013.010.157529.690.878112.990.1459
Bicubic25.430.705950.840.375525.490.722955.990.3656
DDRM [25]25.430.736732.690.320625.830.744332.930.3304
DPS [9]24.750.6594120.740.491121.090.6119175.21430.5541
GDP [19]23.160.669277.430.426023.750.656990.230.4240
PSLD [39]21.720.5498107.990.532921.750.5828141.490.5461
DiffIR [52]24.010.677054.140.436723.900.701450.370.4235
StableSR [47]23.330.657749.950.313523.990.699846.030.3023
OmniSSR25.770.727930.970.297726.010.748134.580.2963
", + "image_path": "e398a17989a1ff47394648f61d2ffea67cbca161612a8f8259a432c62b9805c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 330, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 389 + ], + "type": "text", + "content": "and PSLD [39], which is based on latent diffusion model. We also choose supervised diffusion-based super-resolution approaches including StableSR [47] and DiffIR [52] for comparison. We conduct experiments on " + }, + { + "bbox": [ + 130, + 330, + 480, + 389 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 330, + 480, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 330, + 480, + 389 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 330, + 480, + 389 + ], + "type": "text", + "content": " SR with ERP bicubic downsampling, on ODI-SR test set and SUN test set. We choose WS-PSNR [45], WS-SSIM [67], FID [22], and LPIPS [64] as the main metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 449 + ], + "type": "text", + "content": "Quantitative results are presented in Tab. 1. With proposed OTII and GD correction, OmniSSR out-performs previous methods in terms of both Fidelity (from WS-PSNR and WS-SSIM) and Realness (from FID, LPIPS), which shows superior performance to existing diffusion-based methods for ODISR tasks on different scales." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": "Qualitative results are shown in Fig. 4 and Fig. 5, which illustrates the visualization of SR results on SUN test set and ODI-SR test set with " + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": " scales, by different methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 569 + ], + "type": "text", + "content": "The visual results indicate that our OmniSSR exhibits superior capability for detail recovery compared to other methods, particularly evident in textual elements (e.g., the text \"flapping\" in upper part of Fig. 4), complex objects (e.g., the black desk with a screen in lower part of Fig. 4, patterns above the white door in lower part of Fig. 5), and small-scale objects (e.g., the person and clock behind the desk in upper part of Fig. 5). OmniSSR demonstrates the ability to recover highly detailed and realistic visual effects from TP images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 586, + 408, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 408, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 408, + 598 + ], + "type": "text", + "content": "4.3 Comparison with end-to-end supervised methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "The experiments of comparison in Sec. 4.2 are mainly focused on zero-shot image super-resolution methods, and supervised single image super-resolution methods, where the approaches are not trained or fine-tuned on omnidirectional images. In this part, we will compare OmniSSR to supervised end-to-end methods with end-to-end training on ODI datasets, including SwinIR and OSRT. Besides the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 232, + 167 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 232, + 167 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 232, + 167 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 232, + 167 + ], + "type": "image", + "image_path": "e4c3ff6c545218f159267cef5882c1603f2c8bf24709a01975de6d37f76fc2e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 167, + 220, + 183 + ], + "lines": [ + { + "bbox": [ + 153, + 167, + 220, + 183 + ], + "spans": [ + { + "bbox": [ + 153, + 167, + 220, + 183 + ], + "type": "inline_equation", + "content": "\\therefore m - 1 \\neq 0" + }, + { + "bbox": [ + 153, + 167, + 220, + 183 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 114, + 288, + 167 + ], + "blocks": [ + { + "bbox": [ + 233, + 114, + 288, + 167 + ], + "lines": [ + { + "bbox": [ + 233, + 114, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 233, + 114, + 288, + 167 + ], + "type": "image", + "image_path": "77002c1d00337de418c645adf178eaa35e56c1a3f5ee83e53bfeb849f3eaed87.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 167, + 283, + 183 + ], + "lines": [ + { + "bbox": [ + 238, + 167, + 283, + 183 + ], + "spans": [ + { + "bbox": [ + 238, + 167, + 283, + 183 + ], + "type": "text", + "content": "HR PSNR/SSIM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 290, + 114, + 344, + 167 + ], + "blocks": [ + { + "bbox": [ + 290, + 114, + 344, + 167 + ], + "lines": [ + { + "bbox": [ + 290, + 114, + 344, + 167 + ], + "spans": [ + { + "bbox": [ + 290, + 114, + 344, + 167 + ], + "type": "image", + "image_path": "05650ecbcf953f8c27820e324f67b077efd87da717d44393573f8cf043065397.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 347, + 114, + 401, + 167 + ], + "blocks": [ + { + "bbox": [ + 347, + 114, + 401, + 167 + ], + "lines": [ + { + "bbox": [ + 347, + 114, + 401, + 167 + ], + "spans": [ + { + "bbox": [ + 347, + 114, + 401, + 167 + ], + "type": "image", + "image_path": "118344129e4a2eb4ef49ca66cfce650818a4e189833fe66125991db2b6b1a1e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 167, + 402, + 184 + ], + "lines": [ + { + "bbox": [ + 346, + 167, + 402, + 184 + ], + "spans": [ + { + "bbox": [ + 346, + 167, + 402, + 184 + ], + "type": "text", + "content": "DPS [9] 24.02dB/0.5849" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 407, + 114, + 461, + 167 + ], + "blocks": [ + { + "bbox": [ + 407, + 114, + 461, + 167 + ], + "lines": [ + { + "bbox": [ + 407, + 114, + 461, + 167 + ], + "spans": [ + { + "bbox": [ + 407, + 114, + 461, + 167 + ], + "type": "image", + "image_path": "d982945947d1b2335bb0a61a5486f3b5c04193c4644c94c0a4f33c6ff494dbe5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 167, + 462, + 184 + ], + "lines": [ + { + "bbox": [ + 408, + 167, + 462, + 184 + ], + "spans": [ + { + "bbox": [ + 408, + 167, + 462, + 184 + ], + "type": "text", + "content": "DDRM [25] 30.04dB/0.8855" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 140, + 183, + 232, + 236 + ], + "blocks": [ + { + "bbox": [ + 140, + 183, + 232, + 236 + ], + "lines": [ + { + "bbox": [ + 140, + 183, + 232, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 183, + 232, + 236 + ], + "type": "image", + "image_path": "82b8f5727c3feb46d1fab616286acba8ecf4e4d130640730b693e03e94627b25.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 237, + 220, + 246 + ], + "lines": [ + { + "bbox": [ + 152, + 237, + 220, + 246 + ], + "spans": [ + { + "bbox": [ + 152, + 237, + 220, + 246 + ], + "type": "text", + "content": "SUN 360 " + }, + { + "bbox": [ + 152, + 237, + 220, + 246 + ], + "type": "inline_equation", + "content": "(\\times 2)" + }, + { + "bbox": [ + 152, + 237, + 220, + 246 + ], + "type": "text", + "content": " : 001" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 233, + 183, + 288, + 237 + ], + "blocks": [ + { + "bbox": [ + 233, + 183, + 288, + 237 + ], + "lines": [ + { + "bbox": [ + 233, + 183, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 233, + 183, + 288, + 237 + ], + "type": "image", + "image_path": "0c3a37a737516d7755f59d9cc28cf2ea82c2e766d12fbad042740b8a943120a8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 238, + 287, + 254 + ], + "lines": [ + { + "bbox": [ + 237, + 238, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 237, + 238, + 287, + 254 + ], + "type": "text", + "content": "GDP [19] 0.07dB/0.8802" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 290, + 183, + 344, + 237 + ], + "blocks": [ + { + "bbox": [ + 290, + 168, + 345, + 184 + ], + "lines": [ + { + "bbox": [ + 290, + 168, + 345, + 184 + ], + "spans": [ + { + "bbox": [ + 290, + 168, + 345, + 184 + ], + "type": "text", + "content": "Bicubic 28.67dB/0.8317" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 183, + 344, + 237 + ], + "lines": [ + { + "bbox": [ + 290, + 183, + 344, + 237 + ], + "spans": [ + { + "bbox": [ + 290, + 183, + 344, + 237 + ], + "type": "image", + "image_path": "51e080788bc5266f80850c4fd2c7d67f12d19c5ba9a9dd08ba45a05525c3fbc1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 238, + 344, + 254 + ], + "lines": [ + { + "bbox": [ + 290, + 238, + 344, + 254 + ], + "spans": [ + { + "bbox": [ + 290, + 238, + 344, + 254 + ], + "type": "text", + "content": "DiffIR [52] 24.47dB/0.6421" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 347, + 183, + 401, + 237 + ], + "blocks": [ + { + "bbox": [ + 347, + 183, + 401, + 237 + ], + "lines": [ + { + "bbox": [ + 347, + 183, + 401, + 237 + ], + "spans": [ + { + "bbox": [ + 347, + 183, + 401, + 237 + ], + "type": "image", + "image_path": "d7f3e583f56f782601858dfe22fecbd03b12d177240296fb18e580a05c6757d9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 238, + 402, + 254 + ], + "lines": [ + { + "bbox": [ + 347, + 238, + 402, + 254 + ], + "spans": [ + { + "bbox": [ + 347, + 238, + 402, + 254 + ], + "type": "text", + "content": "StableSR [47] 23.81dB/0.7384" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 406, + 183, + 462, + 237 + ], + "blocks": [ + { + "bbox": [ + 406, + 183, + 462, + 237 + ], + "lines": [ + { + "bbox": [ + 406, + 183, + 462, + 237 + ], + "spans": [ + { + "bbox": [ + 406, + 183, + 462, + 237 + ], + "type": "image", + "image_path": "026ddc985ebdab03ea056438df9d09337ff31b83aeb64778b6cc93f49748ddad.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 238, + 463, + 254 + ], + "lines": [ + { + "bbox": [ + 404, + 238, + 463, + 254 + ], + "spans": [ + { + "bbox": [ + 404, + 238, + 463, + 254 + ], + "type": "text", + "content": "OmniSSR(ours) 30.15dB/0.8859" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 140, + 255, + 232, + 376 + ], + "blocks": [ + { + "bbox": [ + 140, + 255, + 232, + 376 + ], + "lines": [ + { + "bbox": [ + 140, + 255, + 232, + 376 + ], + "spans": [ + { + "bbox": [ + 140, + 255, + 232, + 376 + ], + "type": "image", + "image_path": "a22c76f1a9c57cf9ec0cd05b2cf7728d59df730cc8b188461f2ea001da8ec4f1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 376, + 220, + 385 + ], + "lines": [ + { + "bbox": [ + 152, + 376, + 220, + 385 + ], + "spans": [ + { + "bbox": [ + 152, + 376, + 220, + 385 + ], + "type": "text", + "content": "SUN 360 " + }, + { + "bbox": [ + 152, + 376, + 220, + 385 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 152, + 376, + 220, + 385 + ], + "type": "text", + "content": " : 009" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 377, + 288, + 394 + ], + "lines": [ + { + "bbox": [ + 232, + 377, + 288, + 394 + ], + "spans": [ + { + "bbox": [ + 232, + 377, + 288, + 394 + ], + "type": "text", + "content": "GDP [19] 23.02dB/0.7525" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 403, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 131, + 403, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 482, + 439 + ], + "type": "text", + "content": "Fig. 4: Visualized comparison of " + }, + { + "bbox": [ + 131, + 403, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 131, + 403, + 482, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 403, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 403, + 482, + 439 + ], + "type": "text", + "content": " SR results on SUN 360 test set. 001 and 009 are the ID numbers in the test set filenames. We also calculate the PSNR and SSIM to HR ground truth of each SR result and downsampled image." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 233, + 255, + 288, + 376 + ], + "blocks": [ + { + "bbox": [ + 233, + 255, + 288, + 376 + ], + "lines": [ + { + "bbox": [ + 233, + 255, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 233, + 255, + 288, + 376 + ], + "type": "image", + "image_path": "ad6731bd33a9a430eda461540342ab99da7f72d821a6de2efa230ce8f2f0bc06.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 290, + 255, + 344, + 376 + ], + "blocks": [ + { + "bbox": [ + 290, + 255, + 344, + 376 + ], + "lines": [ + { + "bbox": [ + 290, + 255, + 344, + 376 + ], + "spans": [ + { + "bbox": [ + 290, + 255, + 344, + 376 + ], + "type": "image", + "image_path": "fb91862be9975ce4a1bb80cd177b676301418460af1b0b6581f80b2790e37779.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 377, + 344, + 394 + ], + "lines": [ + { + "bbox": [ + 290, + 377, + 344, + 394 + ], + "spans": [ + { + "bbox": [ + 290, + 377, + 344, + 394 + ], + "type": "text", + "content": "DiffIR [52] 23.48dB/0.7799" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 347, + 255, + 401, + 376 + ], + "blocks": [ + { + "bbox": [ + 347, + 255, + 401, + 376 + ], + "lines": [ + { + "bbox": [ + 347, + 255, + 401, + 376 + ], + "spans": [ + { + "bbox": [ + 347, + 255, + 401, + 376 + ], + "type": "image", + "image_path": "d8c4ff0b5f831768bbd1ad40dfa23441b9d5b0b105e7250788095ac4b9050705.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 377, + 402, + 394 + ], + "lines": [ + { + "bbox": [ + 348, + 377, + 402, + 394 + ], + "spans": [ + { + "bbox": [ + 348, + 377, + 402, + 394 + ], + "type": "text", + "content": "StableSR [47] 24.62dB/0.7981" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 406, + 255, + 461, + 376 + ], + "blocks": [ + { + "bbox": [ + 406, + 255, + 461, + 376 + ], + "lines": [ + { + "bbox": [ + 406, + 255, + 461, + 376 + ], + "spans": [ + { + "bbox": [ + 406, + 255, + 461, + 376 + ], + "type": "image", + "image_path": "90267770031892416bc9530e51251ba52189be4ad3bb981dd57eed42669e6495.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 377, + 463, + 394 + ], + "lines": [ + { + "bbox": [ + 404, + 377, + 463, + 394 + ], + "spans": [ + { + "bbox": [ + 404, + 377, + 463, + 394 + ], + "type": "text", + "content": "OmniSSR(ours) 26.53dB/0.8265" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 458, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 579 + ], + "type": "text", + "content": "main metrics in Sec. 4.2, we also use NIQE [34] and DISTS [15] to evaluate the visual perception of SR outputs. Results are presented in Tab. 2, which shows that although our OmniSSR exhibits inferior fidelity metrics compared to end-to-end supervised methods trained directly on ODI datasets, it demonstrates notable improvements in the visual quality and authenticity of super-resolved images. Notably, end-to-end methods often produce smoothed reconstructions with distortions, whereas our approach preserves finer details and adheres more closely to the realistic distribution. Considering that our method has never been trained or tuned on ODI datasets, nor having omnidirectional images prior, this result is acceptable." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 597, + 246, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 246, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 246, + 608 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We first sequentially validate the performance improvement of the proposed strategy in OmniSSR including input image type, OTII and GD correction, on the ODI-SR test set with " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " SR task, thereby demonstrating the significance of these strategies. The details are demonstrated as follows:" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 232, + 236 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 232, + 236 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 232, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 232, + 236 + ], + "type": "image", + "image_path": "0bdc65fb2176dac12f5298b3ca6d8fcb1ec664fb7773e237fbc775869930a5f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 237, + 219, + 246 + ], + "lines": [ + { + "bbox": [ + 153, + 237, + 219, + 246 + ], + "spans": [ + { + "bbox": [ + 153, + 237, + 219, + 246 + ], + "type": "text", + "content": "ODI-SR " + }, + { + "bbox": [ + 153, + 237, + 219, + 246 + ], + "type": "inline_equation", + "content": "(\\times 2)" + }, + { + "bbox": [ + 153, + 237, + 219, + 246 + ], + "type": "text", + "content": " :067" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 114, + 288, + 167 + ], + "blocks": [ + { + "bbox": [ + 233, + 114, + 288, + 167 + ], + "lines": [ + { + "bbox": [ + 233, + 114, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 233, + 114, + 288, + 167 + ], + "type": "image", + "image_path": "0752235e31564fb4f752b63de28f5a0ac4c30935b0e53ef6caf8179ebf4bf697.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 168, + 267, + 175 + ], + "lines": [ + { + "bbox": [ + 253, + 168, + 267, + 175 + ], + "spans": [ + { + "bbox": [ + 253, + 168, + 267, + 175 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 290, + 114, + 344, + 167 + ], + "blocks": [ + { + "bbox": [ + 290, + 114, + 344, + 167 + ], + "lines": [ + { + "bbox": [ + 290, + 114, + 344, + 167 + ], + "spans": [ + { + "bbox": [ + 290, + 114, + 344, + 167 + ], + "type": "image", + "image_path": "cf16280fa802fd9d07b9ff4c5314099f7089002bddd8c755718783f03a4ed9e0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 168, + 331, + 176 + ], + "lines": [ + { + "bbox": [ + 304, + 168, + 331, + 176 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 331, + 176 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 347, + 114, + 401, + 167 + ], + "blocks": [ + { + "bbox": [ + 347, + 114, + 401, + 167 + ], + "lines": [ + { + "bbox": [ + 347, + 114, + 401, + 167 + ], + "spans": [ + { + "bbox": [ + 347, + 114, + 401, + 167 + ], + "type": "image", + "image_path": "543e56ef5b919493be7a8272c2ca28f09f569e5d38f7712965492320a2909fd0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 168, + 388, + 176 + ], + "lines": [ + { + "bbox": [ + 360, + 168, + 388, + 176 + ], + "spans": [ + { + "bbox": [ + 360, + 168, + 388, + 176 + ], + "type": "text", + "content": "DPS [9]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 233, + 184, + 288, + 237 + ], + "blocks": [ + { + "bbox": [ + 238, + 176, + 266, + 183 + ], + "lines": [ + { + "bbox": [ + 238, + 176, + 266, + 183 + ], + "spans": [ + { + "bbox": [ + 238, + 176, + 266, + 183 + ], + "type": "text", + "content": "PSNR/S" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 184, + 288, + 237 + ], + "lines": [ + { + "bbox": [ + 233, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 233, + 184, + 288, + 237 + ], + "type": "image", + "image_path": "527ed727580df771e6de1ebb3b5bef03127e6a1ae23209ef51281ef67497fac8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 238, + 277, + 247 + ], + "lines": [ + { + "bbox": [ + 244, + 238, + 277, + 247 + ], + "spans": [ + { + "bbox": [ + 244, + 238, + 277, + 247 + ], + "type": "text", + "content": "GDP [19]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 290, + 184, + 344, + 237 + ], + "blocks": [ + { + "bbox": [ + 290, + 176, + 345, + 183 + ], + "lines": [ + { + "bbox": [ + 290, + 176, + 345, + 183 + ], + "spans": [ + { + "bbox": [ + 290, + 176, + 345, + 183 + ], + "type": "text", + "content": "27.67dB/0.8095" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 184, + 344, + 237 + ], + "lines": [ + { + "bbox": [ + 290, + 184, + 344, + 237 + ], + "spans": [ + { + "bbox": [ + 290, + 184, + 344, + 237 + ], + "type": "image", + "image_path": "c3b5c7e640014d61d969b992d50e17ef486601a8713b4ea9cf721f5a150659aa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 238, + 335, + 247 + ], + "lines": [ + { + "bbox": [ + 298, + 238, + 335, + 247 + ], + "spans": [ + { + "bbox": [ + 298, + 238, + 335, + 247 + ], + "type": "text", + "content": "DiffIR [52]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 347, + 184, + 401, + 237 + ], + "blocks": [ + { + "bbox": [ + 347, + 176, + 402, + 183 + ], + "lines": [ + { + "bbox": [ + 347, + 176, + 402, + 183 + ], + "spans": [ + { + "bbox": [ + 347, + 176, + 402, + 183 + ], + "type": "text", + "content": "22.93dB/0.5653" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 184, + 401, + 237 + ], + "lines": [ + { + "bbox": [ + 347, + 184, + 401, + 237 + ], + "spans": [ + { + "bbox": [ + 347, + 184, + 401, + 237 + ], + "type": "image", + "image_path": "fbb72815f1c1de0e2f751e61bdb32e69dd9832622603ed3647c242b650b3f5c7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 238, + 399, + 247 + ], + "lines": [ + { + "bbox": [ + 350, + 238, + 399, + 247 + ], + "spans": [ + { + "bbox": [ + 350, + 238, + 399, + 247 + ], + "type": "text", + "content": "StableSR [47]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 407, + 114, + 461, + 167 + ], + "blocks": [ + { + "bbox": [ + 407, + 114, + 461, + 167 + ], + "lines": [ + { + "bbox": [ + 407, + 114, + 461, + 167 + ], + "spans": [ + { + "bbox": [ + 407, + 114, + 461, + 167 + ], + "type": "image", + "image_path": "359bdce6a7e8122ad244e0f897c40642adfd7cd00819502c5a6608ac55d62143.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 168, + 454, + 176 + ], + "lines": [ + { + "bbox": [ + 411, + 168, + 454, + 176 + ], + "spans": [ + { + "bbox": [ + 411, + 168, + 454, + 176 + ], + "type": "text", + "content": "DDRM [25]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 406, + 184, + 461, + 237 + ], + "blocks": [ + { + "bbox": [ + 406, + 176, + 462, + 184 + ], + "lines": [ + { + "bbox": [ + 406, + 176, + 462, + 184 + ], + "spans": [ + { + "bbox": [ + 406, + 176, + 462, + 184 + ], + "type": "text", + "content": "29.91dB/0.8809" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 184, + 461, + 237 + ], + "lines": [ + { + "bbox": [ + 406, + 184, + 461, + 237 + ], + "spans": [ + { + "bbox": [ + 406, + 184, + 461, + 237 + ], + "type": "image", + "image_path": "a39408904009596fae7d49e6a19b743fb7fff6944807c0196b6c4431195a9f4a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 238, + 463, + 246 + ], + "lines": [ + { + "bbox": [ + 404, + 238, + 463, + 246 + ], + "spans": [ + { + "bbox": [ + 404, + 238, + 463, + 246 + ], + "type": "text", + "content": "OmniSSR(ours)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 140, + 255, + 232, + 376 + ], + "blocks": [ + { + "bbox": [ + 140, + 255, + 232, + 376 + ], + "lines": [ + { + "bbox": [ + 140, + 255, + 232, + 376 + ], + "spans": [ + { + "bbox": [ + 140, + 255, + 232, + 376 + ], + "type": "image", + "image_path": "d9dd5fa49be977880334bf244e60346700d2164088881783a7cfe8c0758cb98c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 376, + 219, + 385 + ], + "lines": [ + { + "bbox": [ + 153, + 376, + 219, + 385 + ], + "spans": [ + { + "bbox": [ + 153, + 376, + 219, + 385 + ], + "type": "text", + "content": "ODI-SR " + }, + { + "bbox": [ + 153, + 376, + 219, + 385 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 153, + 376, + 219, + 385 + ], + "type": "text", + "content": " :049" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "type": "text", + "content": "Fig. 5: Visualized comparison of " + }, + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "type": "text", + "content": " SR results on ODI-SR test set. 067 and 049 are the ID numbers in test set filenames. We also calculate the PSNR and SSIM between ground truth and each SR result as well as the downsampled image." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 233, + 255, + 288, + 308 + ], + "blocks": [ + { + "bbox": [ + 233, + 255, + 288, + 308 + ], + "lines": [ + { + "bbox": [ + 233, + 255, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 233, + 255, + 288, + 308 + ], + "type": "image", + "image_path": "93c9af5a620013e7780e5afa1e28077d1cec55b360e465d761e40a0ee6c62dfb.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 308, + 266, + 316 + ], + "lines": [ + { + "bbox": [ + 253, + 308, + 266, + 316 + ], + "spans": [ + { + "bbox": [ + 253, + 308, + 266, + 316 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 290, + 255, + 344, + 307 + ], + "blocks": [ + { + "bbox": [ + 290, + 247, + 344, + 254 + ], + "lines": [ + { + "bbox": [ + 290, + 247, + 344, + 254 + ], + "spans": [ + { + "bbox": [ + 290, + 247, + 344, + 254 + ], + "type": "text", + "content": "22.65dB/0.6248" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 255, + 344, + 307 + ], + "lines": [ + { + "bbox": [ + 290, + 255, + 344, + 307 + ], + "spans": [ + { + "bbox": [ + 290, + 255, + 344, + 307 + ], + "type": "image", + "image_path": "1b960db6d669d53cef5bcf03dc771877d1cca6a1dcbdf7989e6b9516c3c4e692.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 308, + 331, + 315 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 331, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 331, + 315 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 347, + 255, + 401, + 307 + ], + "blocks": [ + { + "bbox": [ + 347, + 247, + 402, + 254 + ], + "lines": [ + { + "bbox": [ + 347, + 247, + 402, + 254 + ], + "spans": [ + { + "bbox": [ + 347, + 247, + 402, + 254 + ], + "type": "text", + "content": "21.80dB/0.5892" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 255, + 401, + 307 + ], + "lines": [ + { + "bbox": [ + 347, + 255, + 401, + 307 + ], + "spans": [ + { + "bbox": [ + 347, + 255, + 401, + 307 + ], + "type": "image", + "image_path": "665b7a80abefad187ae780411eb88428615dd5b47d299a62e7eef6793e7a5976.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 308, + 388, + 316 + ], + "lines": [ + { + "bbox": [ + 360, + 308, + 388, + 316 + ], + "spans": [ + { + "bbox": [ + 360, + 308, + 388, + 316 + ], + "type": "text", + "content": "DPS [9]" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "spans": [ + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "type": "text", + "content": "26.21dB/0.7574" + } + ] + } + ], + "index": 37, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 233, + 316, + 288, + 376 + ], + "blocks": [ + { + "bbox": [ + 238, + 316, + 284, + 324 + ], + "lines": [ + { + "bbox": [ + 238, + 316, + 284, + 324 + ], + "spans": [ + { + "bbox": [ + 238, + 316, + 284, + 324 + ], + "type": "text", + "content": "PSNR/SSIM" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 316, + 288, + 376 + ], + "lines": [ + { + "bbox": [ + 233, + 316, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 233, + 316, + 288, + 376 + ], + "type": "image", + "image_path": "ceaf816b663668d5f27c284215943761f50a52998a3e3458c9f928edb319c082.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 377, + 290, + 394 + ], + "lines": [ + { + "bbox": [ + 233, + 377, + 290, + 394 + ], + "spans": [ + { + "bbox": [ + 233, + 377, + 290, + 394 + ], + "type": "text", + "content": "GDP [19] 24.15dB/0.7179" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 290, + 316, + 344, + 376 + ], + "blocks": [ + { + "bbox": [ + 290, + 316, + 345, + 323 + ], + "lines": [ + { + "bbox": [ + 290, + 316, + 345, + 323 + ], + "spans": [ + { + "bbox": [ + 290, + 316, + 345, + 323 + ], + "type": "text", + "content": "25.44dB/0.7536" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 316, + 344, + 376 + ], + "lines": [ + { + "bbox": [ + 290, + 316, + 344, + 376 + ], + "spans": [ + { + "bbox": [ + 290, + 316, + 344, + 376 + ], + "type": "image", + "image_path": "bfa6cc4eb8ac106c9f8c23112c1587b60c9c8ae66410ef9371f606367ada7017.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 377, + 335, + 386 + ], + "lines": [ + { + "bbox": [ + 298, + 377, + 335, + 386 + ], + "spans": [ + { + "bbox": [ + 298, + 377, + 335, + 386 + ], + "type": "text", + "content": "DiffIR [52]" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 386, + 344, + 393 + ], + "lines": [ + { + "bbox": [ + 298, + 386, + 344, + 393 + ], + "spans": [ + { + "bbox": [ + 298, + 386, + 344, + 393 + ], + "type": "text", + "content": ".25dB/0.7594" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "spans": [ + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "type": "text", + "content": "26.21dB/0.7574" + } + ] + } + ], + "index": 43, + "type": "text" + }, + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "spans": [ + { + "bbox": [ + 347, + 316, + 402, + 323 + ], + "type": "text", + "content": "26.21dB/0.7574" + } + ] + } + ], + "index": 44, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 406, + 255, + 461, + 307 + ], + "blocks": [ + { + "bbox": [ + 406, + 247, + 461, + 254 + ], + "lines": [ + { + "bbox": [ + 406, + 247, + 461, + 254 + ], + "spans": [ + { + "bbox": [ + 406, + 247, + 461, + 254 + ], + "type": "text", + "content": "29.99dB/0.8798" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 255, + 461, + 307 + ], + "lines": [ + { + "bbox": [ + 406, + 255, + 461, + 307 + ], + "spans": [ + { + "bbox": [ + 406, + 255, + 461, + 307 + ], + "type": "image", + "image_path": "2b9a5aceeddd7d72e94d00bc0cc4f0520d244efa9dd886d51156ed52d0128903.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 308, + 454, + 316 + ], + "lines": [ + { + "bbox": [ + 413, + 308, + 454, + 316 + ], + "spans": [ + { + "bbox": [ + 413, + 308, + 454, + 316 + ], + "type": "text", + "content": "DDRM [25]" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 406, + 323, + 461, + 376 + ], + "blocks": [ + { + "bbox": [ + 406, + 316, + 462, + 323 + ], + "lines": [ + { + "bbox": [ + 406, + 316, + 462, + 323 + ], + "spans": [ + { + "bbox": [ + 406, + 316, + 462, + 323 + ], + "type": "text", + "content": "27.12dB/0.8129" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 323, + 461, + 376 + ], + "lines": [ + { + "bbox": [ + 406, + 323, + 461, + 376 + ], + "spans": [ + { + "bbox": [ + 406, + 323, + 461, + 376 + ], + "type": "image", + "image_path": "87f0ff13584b3ff238d34f0281a2ee9ae79b7b540089a9bb21222902908299ec.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 377, + 463, + 386 + ], + "lines": [ + { + "bbox": [ + 403, + 377, + 463, + 386 + ], + "spans": [ + { + "bbox": [ + 403, + 377, + 463, + 386 + ], + "type": "text", + "content": "OmniSSR(ours)" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 404, + 386, + 462, + 394 + ], + "lines": [ + { + "bbox": [ + 404, + 386, + 462, + 394 + ], + "spans": [ + { + "bbox": [ + 404, + 386, + 462, + 394 + ], + "type": "text", + "content": "27.20dB/0.8168" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "bbox": [ + 130, + 459, + 481, + 581 + ], + "type": "list", + "angle": 0, + "index": 57, + "blocks": [ + { + "bbox": [ + 130, + 459, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 480, + 483 + ], + "type": "text", + "content": "1) we do not use any proposed strategy in the SR task, which is equivalent to the vanilla StableSR baseline;" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 507 + ], + "type": "text", + "content": "2) we transform the degraded ERP image to TP images and feed them separately into StableSR pipeline, instead of directly inputting ERP images;" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 130, + 508, + 480, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 480, + 531 + ], + "type": "text", + "content": "3) based on 2), we add OTII strategy during the denoising process of SD (Algo. 2 line 7);" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 130, + 532, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 480, + 556 + ], + "type": "text", + "content": "4) based on 2), we add GD correction at the post-processing stage (Algo. 1 line 12) of the overall pipeline;" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 130, + 556, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 481, + 581 + ], + "type": "text", + "content": "5) based on 3) and 4), we add GD correction at every step and post-processing stage of sampling, to improve the consistency of the restored result." + } + ] + } + ], + "index": 56 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 582, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 616 + ], + "type": "text", + "content": "Note that the execution of GD correction requires the execution of OTII in the denoising process simultaneously, there is no scenario where only GD correction is executed without the execution of OTII in the denoising process." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Quantitative results of ablation studies are shown in Tab. 3. From the result shown below, we could come to the claim that the OTII helps improve the performance on the domain level, and the transformation between ERP and TP images provides information fusion among adjacent TP images. Our proposal of" + } + ] + } + ], + "index": 59 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 146, + 444, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Comparison on " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": " SR task with supervised methods trained on ODI-SR dataset, including SwinIR and OSRT. The best results are shown in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 146, + 444, + 216 + ], + "lines": [ + { + "bbox": [ + 167, + 146, + 444, + 216 + ], + "spans": [ + { + "bbox": [ + 167, + 146, + 444, + 216 + ], + "type": "table", + "html": "
MethodDatasetWS-PSNR↑WS-SSIM↑FID↓LPIPS↓NIQE↓DISTS↓
SwinIR [29]26.760.762027.940.33215.39610.1710
OSRT [60]ODI-SR26.890.764627.390.32585.43640.1695
OmniSSR25.770.727930.970.29775.28910.1541
SwinIR [29]26.020.769239.900.34195.24400.1325
OSRT [60]SUN 36026.330.776639.220.33645.29840.1312
OmniSSR26.010.748134.580.29635.13290.1299
", + "image_path": "b43309cf1b7992571640fe32db857034d4d050d51a1962a710fdc2b8eb059731.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 162, + 258, + 452, + 317 + ], + "blocks": [ + { + "bbox": [ + 132, + 226, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 226, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 480, + 247 + ], + "type": "text", + "content": "Table 3: Ablation studies of OmniSSR on input type, OTII, and GD correction, on the test set of the ODI-SR dataset. Best results are shown in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 258, + 452, + 317 + ], + "lines": [ + { + "bbox": [ + 162, + 258, + 452, + 317 + ], + "spans": [ + { + "bbox": [ + 162, + 258, + 452, + 317 + ], + "type": "table", + "html": "
Input typeOTIIGD CorrectionWS-PSNR↑WS-SSIM↑FID↓LPIPS↓
ERP××22.690.645844.870.3039
TP××23.530.684943.910.3113
TP✓×23.740.684765.350.3748
TP×✓ (in post-process only)26.770.819215.410.1691
TP✓✓28.580.854013.010.1575
", + "image_path": "566a66533920dfe81b91ceecd2f59b400f8b11554e6bd4af3b7de3c7fbd087cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 480, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 387 + ], + "type": "text", + "content": "Gradient Decomposition corrects such restoration result, improving fidelity and realness significantly at the same time, and it would be better if it is applied at each step of the overall denoising pipeline. Tab. 4 shows the effect of mitigating information loss via proposed pre-upsampling strategy." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 152, + 408, + 254, + 483 + ], + "blocks": [ + { + "bbox": [ + 152, + 408, + 254, + 483 + ], + "lines": [ + { + "bbox": [ + 152, + 408, + 254, + 483 + ], + "spans": [ + { + "bbox": [ + 152, + 408, + 254, + 483 + ], + "type": "image", + "image_path": "4da33366ec8ad894a06bf0a9b601184e2fc449d8ab20e8fa199af3a5e6d11cf6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "lines": [ + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": "Fig.6: Ablation of choices on " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{p}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{e}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{l}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": ". For better readability, WS-PSNR and LPIPS are chosen as evaluation metrics for fidelity and visual quality, respectively, to demonstrate the performance under different choices of the gamma parameter. We illustrate the results of (a) " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{p}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{e}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": " fixed, while adjusting " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{l}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": "; (b) " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{e}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{l}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": " fixed, while adjusting " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{p}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": "; (c) " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{p}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{l}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": " fixed, while adjusting " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{e}" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": ". It can be observed that when " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{p} = 1" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{e} = 1" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "inline_equation", + "content": "\\gamma_{l} = 0.5" + }, + { + "bbox": [ + 130, + 495, + 480, + 561 + ], + "type": "text", + "content": ", OmniSSR achieves the relatively best performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 256, + 408, + 358, + 483 + ], + "blocks": [ + { + "bbox": [ + 256, + 408, + 358, + 483 + ], + "lines": [ + { + "bbox": [ + 256, + 408, + 358, + 483 + ], + "spans": [ + { + "bbox": [ + 256, + 408, + 358, + 483 + ], + "type": "image", + "image_path": "5ad62970d5261ee1e556964b7162f20ac9220f05687d019e45e9cd5af8a3b221.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 408, + 462, + 483 + ], + "blocks": [ + { + "bbox": [ + 359, + 408, + 462, + 483 + ], + "lines": [ + { + "bbox": [ + 359, + 408, + 462, + 483 + ], + "spans": [ + { + "bbox": [ + 359, + 408, + 462, + 483 + ], + "type": "image", + "image_path": "79cf981a76a17afe2ec979ca890f842edaa11cb91a5c60efe71ba9c4ef1a517e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": " in the GD correction technique, we use grid search to obtain better results on ODI-SR dataset and " + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": " SR task. Fig. 6 shows performance on different choices of " + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\gamma_{p}" + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": " in Algo. 1 line 12, " + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\gamma_{e}" + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": " in Algo. 2 line 8, and " + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\gamma_{l}" + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": " in Algo. 2 line 11. The entire ablation of " + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\gamma_{p},\\gamma_{e}" + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\gamma_{l}" + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": ", with WS-PSNR, WS-SSIM, FID and LPIPS score all calculated and compared, will be provided in Supplementary Materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "To evaluate the generalizability of our proposed modules, including Pre-Upsampling, OTII, and GD correction, we further conducted ablation studies" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 197, + 158, + 416, + 190 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Results of pre-upsampling strategy on different scales, where " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " denotes bicubic-based upsampling at " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "x\\times" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " scale to ERP before " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{ERP}\\rightarrow \\mathrm{TP}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "y\\times" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " scale to TP before " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{TP}\\rightarrow \\mathrm{ERP}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " transformation. Best results are shown in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 197, + 158, + 416, + 190 + ], + "lines": [ + { + "bbox": [ + 197, + 158, + 416, + 190 + ], + "spans": [ + { + "bbox": [ + 197, + 158, + 416, + 190 + ], + "type": "table", + "html": "
\\( \\mathrm{{ERP}} \\rightarrow \\mathrm{{TP}} \\rightarrow \\mathrm{{ERP}} \\)(1,1)(1,4)(4,1)(4,2)(2,4)(4,4)
WS-PSNR↑28.9838.1128.9933.9138.0538.18
WS-SSIM↑0.88590.98380.88620.96260.98370.9841
", + "image_path": "932fd3dbfab607561b114af73fb44bc830ec27dfcf6f450faf4c6e4a975885f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "text", + "content": "on two super-resolution backbones, StableSR and SwinIR. The results underscore substantial performance enhancements facilitated by our modules across both backbones, which is provided in Supplementary Materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 265, + 308, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 265, + 308, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 265, + 308, + 277 + ], + "type": "text", + "content": "5 Limitation and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 289, + 480, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 408 + ], + "type": "text", + "content": "Although OmniSSR bridges the gap between omnidirectional and planar images, achieving competitive performance and better visual results in ODISR, it still exhibits the following limitations: (1) The inference of the diffusion model requires a considerable amount of time, approximately 14 minutes per ERP-formatted omnidirectional image to be super-resolved into size " + }, + { + "bbox": [ + 130, + 289, + 480, + 408 + ], + "type": "inline_equation", + "content": "1024 \\times 2048" + }, + { + "bbox": [ + 130, + 289, + 480, + 408 + ], + "type": "text", + "content": ", making real-time super-resolution challenging; (2) Multiple conversions between ERP and TP are required in the pipeline, leading to improved performance but consuming additional inference time; (3) Further exploration of the convex optimization properties of GD correction is warranted, such as designing gradient term coefficients adaptive to reconstruction results and degradation types." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 409, + 481, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 481, + 494 + ], + "type": "text", + "content": "This study explores the application of image generation models to ODISR tasks. In future work, the framework behind OmniSSR can be extended beyond the confines of image super-resolution in a single scenario and venture into more complex ODI-based real-world scenarios. These include ODI editing, ODI inpainting, enhancing the quality of 3D Gaussian Splatting scenes [26,41] obtained after super-resolving ERP images, as well as enhancing the quality of omnidirectional videos [48], and other possible diffusion-based applications [57,68]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "This paper leverages the image prior of Stable Diffusion (SD) and employs the Octadecaplex Tangent Information Interaction (OTII) to achieve zero-shot omnidirectional image super-resolution. Additionally, we propose the Gradient Decomposition (GD) correction based on convex optimization algorithms to refine the initial super-resolution results, enhancing the fidelity and realness of the restored images. The superior performance of our proposed method, OmniSSR, is demonstrated on benchmark datasets. By bridging the gap between omnidirectional and planar images, we establish a training-free approach, mitigating the data demand and over-fitting associated with end-to-end training. The application scope of our method can be further extended to various applications, presenting potential value across multiple visual tasks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. An, H., Zhang, X.: Perception-oriented omnidirectional image super-resolution based on transformer network. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "type": "text", + "content": "2. Arican, Z., Frossard, P.: Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing (TIP) (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "type": "text", + "content": "3. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for image super-resolution and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "4. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "type": "text", + "content": "5. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "text", + "content": "6. Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X., Yu, F.: Dual aggregation transformer for image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision (ICCV) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 363 + ], + "type": "text", + "content": "7. Cheng, M., Ma, H., Ma, Q., Sun, X., Li, W., Zhang, Z., Sheng, X., Zhao, S., Li, J., Zhang, L.: Hybrid transformer and cnn attention network for stereo image superresolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "text", + "content": "8. Chong, M., Yanze, W., Xintao, W., Chao, D., Jian, Z., Ying, S.: Metric learning based interactive modulation for real-world super-resolution. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 418 + ], + "type": "text", + "content": "9. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "text", + "content": "0. Chung, H., Sim, B., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 449, + 455, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 455, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 455, + 460 + ], + "type": "text", + "content": "1. Coxeter, H.S.M.: Introduction to geometry. John Wiley & Sons, Inc. (1961)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "text", + "content": "2. Daras, G., Dean, J., Jalal, A., Dimakis, A.: Intermediate layer optimization for inverse problems using deep generative models. In: Proceedings of the International Conference on Machine Learning (ICML) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 492, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 535 + ], + "type": "text", + "content": "3. Deng, X., Wang, H., Xu, M., Guo, Y., Song, Y., Yang, L.: Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "4. Deng, X., Wang, H., Xu, M., Li, L., Wang, Z.: Omnidirectional image superresolution via latitude adaptive network. IEEE Transactions on Multimedia (TMM) (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "5. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "6. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Duan, H., Zhai, G., Min, X., Zhu, Y., Fang, Y., Yang, X.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the IEEE International Symposium on Circuits and Systems (ISCAS) (2018)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Fakour-Sevom, V., Guldogan, E., Kämäräinen, J.K.: 360 panorama superresolution using deep convolutional networks. In: Proceedings of the Int. Conf. on Computer Vision Theory and Applications (VISAPP) (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "19. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "21. Guo, L., Tao, T., Cai, X., Zhu, Z., Huang, J., Zhu, L., Gu, Z., Tang, H., Zhou, R., Han, S., et al.: Cas-diffcom: Cascaded diffusion model for infant longitudinal super-resolution 3d medical image completion. arXiv preprint arXiv:2402.13776 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "text", + "content": "23. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "24. Jiang, Y., Chan, K.C., Wang, X., Loy, C.C., Liu, Z.: Reference-based image and video super-resolution via " + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "inline_equation", + "content": "c^2" + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "-matching. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "25. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Proceedings of the ICLR Workshop on Deep Generative Models for Highly Structured Data (ICLRW) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 414 + ], + "type": "text", + "content": "26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "27. Li, W., Chen, B., Zhang, J.: D3c2-net: Dual-domain deep convolutional coding network for compressive sensing. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "28. Li, Y., Guo, Y., Yan, Z., Huang, X., Duan, Y., Ren, L.: Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "29. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "30. Liu, J., Wang, Q., Fan, H., Wang, Y., Tang, Y., Qu, L.: Residual denoising diffusion models. arXiv preprint arXiv:2308.13712 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "31. Lu, Z., Li, J., Liu, H., Huang, C., Zhang, L., Zeng, T.: Transformer for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "32. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "text", + "content": "33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "34. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters (SPL) (2013)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Nishiyama, A., Ikehata, S., Aizawa, K.: " + }, + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": " single image super resolution via distortion-aware network and distorted perspective images. In: Proceedings of the IEEE International Conference on Image Processing (ICIP) (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "36. Ozcinar, C., Rana, A., Smolic, A.: Super-resolution of omnidirectional images using adversarial learning. In: Proceedings of the IEEE International Workshop on Multimedia Signal Processing (MMSPW) (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "text", + "content": "37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "text", + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 289 + ], + "type": "text", + "content": "39. Rout, L., Raoof, N., Daras, G., Caramanis, C., Dimakis, A., Shakkottai, S.: Solving linear inverse problems provably via posterior sampling with latent diffusion models. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "type": "text", + "content": "40. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "type": "text", + "content": "41. Schonbein, M., Geiger, A.: Omnidirectional 3d reconstruction in augmented manhattan worlds. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 387 + ], + "type": "text", + "content": "42. Song, J., Zhang, Q., Yin, H., Mardani, M., Liu, M.Y., Kautz, J., Chen, Y., Vahdat, A.: Loss-guided diffusion models for plug-and-play controllable generation. In: Proceedings of the International Conference on Machine Learning (ICML) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "text", + "content": "43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Proceedings of the International Conference on Learning Representations (ICLR) (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 418, + 482, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 461 + ], + "type": "text", + "content": "44. Sun, X., Li, W., Zhang, Z., Ma, Q., Sheng, X., Cheng, M., Ma, H., Zhao, S., Zhang, J., Li, J., et al.: Opdn: Omnidirectional position-aware deformable network for omnidirectional image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 461, + 482, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 483 + ], + "type": "text", + "content": "45. Sun, Y., Lu, A., Yu, L.: Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters (SPL) (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "type": "text", + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "text", + "content": "47. Wang, J., Yue, Z., Zhou, S., Chan, K., Loy, C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 536, + 482, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 579 + ], + "type": "text", + "content": "48. Wang, Q., Li, W., Mou, C., Cheng, X., Zhang, J.: 360dvd: Controllable panorama video generation with 360-degree video diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 579, + 482, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 612 + ], + "type": "text", + "content": "49. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 612, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 643 + ], + "type": "text", + "content": "50. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European Conference on Computer Vision Workshops (ECCVW) (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "51. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "text", + "content": "52. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "53. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using panoramic place representation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "text", + "content": "54. Yagi, Y.: Omnidirectional sensing and its applications. IEICE Transactions on Information and Systems (TOIS) (1999)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "text", + "content": "55. Yamazawa, K., Yagi, Y., Yachida, M.: Omnidirectional imaging with hyperboloidal projection. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (1993)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "56. Yang, S., Zhou, Y., Liu, Z., Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: Proceedings of the SIGGRAPH Asia 2023 Conference Papers (SIGGRAPH Asia) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "type": "text", + "content": "57. Yang, S., Mou, C., Yu, J., Wang, Y., Meng, X., Zhang, J.: Neural video fields editing. arXiv preprint arXiv:2312.08882 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 296, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 330 + ], + "type": "text", + "content": "58. Yinhuai, W., Yujie, H., Jiwen, Y., Jian, Z.: Gan prior based null-space learning for consistent super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "type": "text", + "content": "59. Yoon, Y., Chung, I., Wang, L., Yoon, K.J.: Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 419 + ], + "type": "text", + "content": "60. Yu, F., Wang, X., Cao, M., Li, G., Shan, Y., Dong, C.: Osrt: Omnidirectional image super-resolution with distortion-aware transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "61. Yu, J., Zhang, X., Xu, Y., Zhang, J.: Cross: Diffusion model makes controllable, robust and secure image steganography. In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "62. Yue, Z., Wang, J., Loy, C.C.: Resshift: Efficient diffusion model for image superresolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "text", + "content": "63. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "64. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 587 + ], + "type": "text", + "content": "65. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "66. Zhang, X., Zhang, Y., Xiong, R., Sun, Q., Zhang, J.: Herosnet: Hyperspectral explicable reconstruction and optimal sampling deep network for snapshot compressive imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "67. Zhou, Y., Yu, M., Ma, H., Shao, H., Jiang, G.: Weighted-to-spherically-uniform sism objective quality evaluation for panoramic video. In: Proceedings of the IEEE International Conference on Signal Processing (ICSP) (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "text", + "content": "68. Zongying, L., Hao, L., Liuzhenghao, L., Bin, L., Junwu, Z., Yu-Chian, C.C., Li, Y., Yonghong, T.: Taxdiff: Taxonomic-guided diffusion model for protein sequence generation. arXiv preprint arXiv:2402.17156 (2024)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OmniSSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_content_list.json b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5f5c0b7b38bb22e04559df53a12242f2fb3d440d --- /dev/null +++ b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_content_list.json @@ -0,0 +1,2039 @@ +[ + { + "type": "text", + "text": "OmniSat: Self-Supervised Modality Fusion for Earth Observation", + "text_level": 1, + "bbox": [ + 284, + 140, + 718, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guillaume Astruc $^{1,3,4}$ , Nicolas Gonthier $^{1,2}$ , Clement Mallet $^{1}$ , and Loic Landrieu $^{1,3}$", + "bbox": [ + 344, + 210, + 656, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France \n2 IGN, France \n3 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France \n4 CNES, France", + "bbox": [ + 251, + 253, + 750, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The diversity and complementarity of sensors available for Earth Observations (EO) calls for developing bespoke self-supervised multimodal learning approaches. However, current multimodal EO datasets and models typically focus on a single data type, either mono-date images or time series, which limits their impact. To address this issue, we introduce OmniSat, a novel architecture able to merge diverse EO modalities into expressive features without labels by exploiting their alignment. To demonstrate the advantages of our approach, we create two new multimodal datasets by augmenting existing ones with new modalities. As demonstrated for three downstream tasks—forestry, land cover classification, and crop mapping—OmniSat can learn rich representations without supervision, leading to state-of-the-art performances in semi- and fully supervised settings. Furthermore, our multimodal pretraining scheme improves performance even when only one modality is available for inference. The code and dataset are available at https://github.com/gastruc/OmniSat.", + "bbox": [ + 261, + 323, + 743, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Earth observation $\\cdot$ Multi-modality $\\cdot$ Self-supervised learning", + "bbox": [ + 261, + 530, + 699, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 570, + 356, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised multimodal learning has recently gathered significant interest within computer vision [38, 82, 102]. Earth Observation (EO) is particularly well-suited for developing and evaluating such approaches [29, 51], thanks to the large amount of open-access data captured by sensing technologies with complementary capabilities [36, 79]. Combining different sources of EO observations is crucial for several high-impact applications, including environmental [21, 80, 83] and climate monitoring [56, 97], as well as improving food security [66]. Moreover, learning with few or no labels is essential for developing regions with limited data annotation capabilities [5, 55, 62].", + "bbox": [ + 212, + 601, + 787, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite this potential, most multimodal EO datasets and models focus on a single data type, either mono-date images or time series. This limitation prevents them from simultaneously leveraging the spatial resolution of aerial images [57, 64], the temporal and spectral resolutions of optical satellite time series [26], and the resilience of radar to weather effects [4, 65]. Additionally, existing approaches are often limited for a given set of sensors, limiting their applicability.", + "bbox": [ + 212, + 720, + 787, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we introduce OmniSat, a novel architecture designed for the self-supervised fusion of diverse EO data. Existing multimodal approaches often map", + "bbox": [ + 215, + 810, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multiple unrelated observations from different modalities to one pivot modality [38, 82] or a shared latent space [39, 84]. In contrast, OmniSat merges multiple views of the same area from different modalities into a single representation combining the specific information of each modality [14, 41, 73].", + "bbox": [ + 212, + 146, + 785, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In computer vision, obtaining finely aligned multimodal observations generally requires specialized sensors [54, 59, 67] or the computation of complex mappings between modalities [23, 75]. On the other hand, EO data can be naturally aligned with georeferencing. To leverage this property, we adapt multimodal contrastive learning [49, 72] and cross-modal masked auto-encoding techniques [43] to learn rich multimodal EO representations with a generalist fusion scheme and without annotations.", + "bbox": [ + 212, + 205, + 785, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the scarcity of EO datasets with a diverse range of heterogeneous modalities (see Tab. 1), we enrich the TreeSatAI [2] and PASTIS-R [33, 34] datasets with new aligned modalities. This allows us to evaluate OmniSat's ability to handle an arbitrary number of inputs with varying natures and resolutions. Our contributions can be summarized as follows:", + "bbox": [ + 212, + 294, + 785, + 368 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce OmniSat, a new model that learns to combine varied sources of EO observations in a self-supervised manner, resulting in richer joint representations that capture the unique characteristics of each modality.", + "- We augment two EO benchmarks to create the first datasets with three modalities of different natures (very high resolution images, optical and SAR time series).", + "- We demonstrate that OmniSat can leverage diverse modalities to learn rich representations, establishing new states-of-the-art for tree species, crop type, and land cover classification. Furthermore, our cross modal self-supervised training scheme improves performance even when only one modality is available during inference." + ], + "bbox": [ + 225, + 369, + 784, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 523, + 366, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section provides an overview of self-supervised and multimodal learning, emphasizing the specificities of their usage for Earth observation. Lastly, we highlight the scarcity of multimodal EO datasets with diverse data types.", + "bbox": [ + 212, + 549, + 785, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Learning. This technique consists in learning expressive data representations without labels by using a pretext task. This approach has been particularly successful for natural language [52] and image [70] analysis. Initially focused on discriminative tasks [37, 68, 100], recent self-supervised approaches for images can be categorized as contrastive or generative.", + "bbox": [ + 212, + 601, + 785, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive methods minimize the distance between representations of paired samples, often the same image under different transformations, and maximize the distance with other samples [15, 17, 46]. More efficient methods only consider positive samples and avoid mode collapse by introducing various asymmetries [18, 42] or normalization [16]. Such approaches have been successfully adapted to EO, for which samples are paired according to their location [89] or time of acquisition [7, 63].", + "bbox": [ + 212, + 676, + 785, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative methods reason at the level of individual token—a small portion of the input, typically a patch for images [25]. The objective is to reconstruct the masked tokens of an input image in pixel [10,45,95] or feature space [6]. This principle has been successfully adapted to EO analysis [20, 30, 99], and was further extended to handle multiple spatial scales [74], multimodality [29, 51], or hyperspectral observations [50, 60].", + "bbox": [ + 212, + 765, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e2dc96a91de8034a8210f663d9facd29e2169cbbfd9fa04b941da5ac5a6fa868.jpg", + "table_caption": [ + "Table 1: Publicly Available Multimodal EO Datasets. We provide in parenthesis the spatial resolutions of the single-date images and labels, and the temporal resolutions of time series. S1/S2 denotes Sentinel-1 and 2. $\\star$ : modalities added in this work." + ], + "table_footnote": [], + "table_body": "
DatasetModalitiesLabels
images (single date)time series
SpaceNet6 [81]SAR+optical (0.5m-2m)Xbuilding footprint (<1m)
TreeSatAI [2]aerial + S1/S2 (0.2-10m)Xforestry (60m)
BigEarthNet [86]S1/S2 (10m)Xland cover (100m)
DFC20 [76]S1/S2 (10m)Xland cover (500m)
MDAS [48]S1/S2 + hyperspectral (2.2-10m)Xland cover (0.25m)
DOFA [96]NAIP + Gaofen + S1/S2 + EnMAP (1-30m)XX
PASTIS-R [33, 34]XS1/S2 (30-140 / year)agriculture (10m)
SSL4EO-S12 [92]XS1/S2 (4 / year)X
DFC21-DSE [61]XS1/S2 + LS8 (3-9/year)human activity (500m)
MapInWild [28]XS1/S2 (4 / years)protected areas (10m)
SEN12MS-CR-TS [27]XS1/S2 (30 / years)cloud cover (10m)
MultiSenGE [93]XS1/S2 (30-140 / years)land cover (10m)
FLAIR [31]aerial (0.2m)S2 (20-114 / year)land cover (0.2m)
Satlas [12]NAIP (0.5 -2m)S2 (8-12 / year)various
PASTIS-HD★ SPOT 6-7 (1.5m)S1/S2 (30-140 / year)agriculture (10m)
TreeSatAI-TSaerial (0.2m)★ S1/S2 (10-70 / year)forestry (60m)
", + "bbox": [ + 220, + 195, + 781, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several hybrid approaches combine the discriminative power of contrastive methods and the scalability of generative objectives for natural images [70, 101] and EO data [29]. Our proposed OmniSat model also implements both mechanisms. A key feature is that we leverage the precise alignment between different sources of EO data to contrastively match small patches of different modalities rather than entire images or time series.", + "bbox": [ + 212, + 547, + 787, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Multimodal Learning. Multimodal computer vision has received a lot of interest [13], notably due to the success of cross-modal pre-training [72]. Recent models align the embeddings of heterogeneous modalities such as video and sound [49], depth and images [44], text and image [3, 9], or multiple combinations of these modalities [38, 39, 82, 84].", + "bbox": [ + 212, + 628, + 787, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal learning also has a long history in EO [58, 71, 98] due to the large variety and complementarity of sensors [36, 79]. However, recent transformer-based architectures [90] for EO are often limited to one type of modality, be it a single image [20, 74] or time-series [34, 87]. For example, CROMA [29] and PRESTO [88] are specifically designed for paired optical and radar observations, but cannot handle Very High Resolution (VHR) data. USat [51] considers images with different resolutions, but only takes a single date within a time series. UT&T [31] can natively take single and multi-date observations of different modalities, but cannot be easily pre-trained in a self-supervised manner since it relies on convolutions and an ad-hoc late fusion scheme.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/adaafb273b142fd66f620e4b304cd2da63772604f023462796b53e942b0fafe0.jpg", + "image_caption": [ + "Sentinel-2 time series" + ], + "image_footnote": [], + "bbox": [ + 264, + 143, + 383, + 243 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b99d8f55146e578a2c79a563fcf0efaa3b12e81db16e33da5a4225569c1e9431.jpg", + "image_caption": [ + "(a) FLAIR" + ], + "image_footnote": [], + "bbox": [ + 279, + 256, + 369, + 323 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e7dbf1494e3f9941f7ef4e45d4678844922fdf9001530804a691d96a677f73f.jpg", + "image_caption": [ + "$\\star$ Sentinel-2 time series" + ], + "image_footnote": [], + "bbox": [ + 413, + 143, + 581, + 243 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27a40a7871f836a6b6bcc40bf98238db9b90ba694fdfaa2296a4ca4f295c76c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 255, + 581, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49b1ea627ffa945a279abb165733b2c50fb82deed2f86384c6eba40fc2666e62.jpg", + "image_caption": [ + "$\\star$ Sentinel-1 time series", + "(b) TreeSatAI-TS", + "Fig. 1: Datasets. We represent three tiles from the considered multilabel classification datasets: FLAIR (a), TreeSatAI-TS (b) and PASTIS-HD (c). TreeSatAI-TS is a new dataset built by replacing the single-date Sentinel-1 and 2 images of TreeSatAI [2] by year-long time series. PASTIS-HD (c) adds VHR satellite images to PASTIS-R [34]. $\\star$ : modalities added in this work." + ], + "image_footnote": [], + "bbox": [ + 416, + 292, + 580, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5fc81623380a7f0c94f69f6c4145dede360af252f84a0be6a46016463dfbbbc7.jpg", + "image_caption": [ + "VHR satellite $1.5\\mathrm{m}$" + ], + "image_footnote": [], + "bbox": [ + 614, + 154, + 730, + 243 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0a84d509b9e4e6c42eefd7bc8c1cf3ae9129dad99db837392e258848538730ab.jpg", + "image_caption": [ + "PASTIS: Sentinel-2 time series" + ], + "image_footnote": [], + "bbox": [ + 596, + 253, + 743, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/95ac3c5edf81174e746a0aa5ea123991c7d2d45a2b0049f594b7786afda33ee9.jpg", + "image_caption": [ + "PASTIS-R: Sentinel-1 time series", + "(c) PASTIS-HD" + ], + "image_footnote": [], + "bbox": [ + 599, + 292, + 745, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal EO Datasets. As reported in Table 1, many multimodal EO datasets use Sentinel-1 [11] and 2 [26] data for applications ranging from land cover to forestry analysis and fire detection. We also note that most multimodal datasets only contain data of one type: mono-date image or time series. Several datasets (BigEarthNet [86], DFC20 [76], MDAS [48]) select a single date from time series. However, single Sentinel-1 and 2 acquisitions can be significantly affected by rain and cloud cover, respectively. Furthermore, capturing the temporal dynamics is crucial to characterize the phenology of vegetation [91],", + "bbox": [ + 212, + 430, + 787, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FLAIR [31] is the first multimodal EO dataset to propose both very high spatial resolution $(\\leq 2\\mathrm{m})$ and high temporal resolution $(>4$ images/year). Satlas [12] combines Sentinel-2 time series and for $5\\%$ to tiles (continental US), very high definition NAIP images. The functional map of the World [19] integrates observations from various sensors, but most areas are only observed with one sensor. Two other datasets contain time series and single images from multiple sources, but were not available at the time of writing this article: IARPA-SMART [40] and DOFA [96].", + "bbox": [ + 212, + 550, + 787, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To showcase how OmniSat can consume an arbitrary number of modalities with different spatial, spectral, and temporal resolutions, we selected two commonly used EO benchmarks, TreeSatAI [2] and PASTIS-R [34], whose focus on crop type mapping and forestry differs from the land cover analysis of FLAIR. We added new modalities to these datasets to reach three distinct data types: VHR aerial images, optical time series, and SAR time series. See Fig. 1 for an illustration, and Sec. 4.1 for more details on how we extended these datasets.", + "bbox": [ + 212, + 654, + 787, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 780, + 316, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider a tile $x$ observed through a set $\\mathbf{M}$ of $M$ distinct sensors or modalities. The goal of the OmniSat model is to learn in a self-supervised fashion to combine", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G.Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a4ab131949202b7eea26d3f6ad97cfa1a4fd862e9f1aef86144ee95446301553.jpg", + "image_caption": [ + "Fig. 2: OmniSat Architecture. We illustrate OmniSat for $M = 3$ modalities, and a tile split into $P = 4$ patches. The $M \\times P$ input tokens $x_{\\mathbf{P}}^{\\mathbf{M}}$ are encoded by $M$ modality-specific encoders $\\mathcal{E}^{\\mathrm{M}}$ , yielding the token representations $f_{\\mathbf{P}}^{\\mathbf{M}}$ . The module $\\mathcal{C}$ combines them into multimodal patch representations $f_{\\mathbf{P}}^{\\star}$ . The token embeddings $f_{\\mathbf{P}}^{\\mathbf{M}}$ are supervised by a contrastive cross-modal objective. We also use a reconstruction objective: the masked multimodal representations $f_{\\mathbf{P}}^{\\star}$ are decoded by modality-specific networks $\\mathcal{D}^{\\mathrm{M}}$ to reconstruct their corresponding inputs in $x_{\\mathbf{P}}^{\\mathbf{M}}$ ." + ], + "image_footnote": [], + "bbox": [ + 240, + 146, + 759, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "all modalities $\\mathbf{M}$ into a multimodal representation $f^{\\star}$ . We first provide details about OmniSat's architecture in Sec. 3.1. We then explain our our training scheme, which consists of a cross-modal contrastive objective (Sec. 3.2) and a multimodal masked encoding task (Sec. 3.3). Finally, we present the implementation details in Sec. 3.4. The overall method is represented in Fig. 2.", + "bbox": [ + 212, + 510, + 787, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Architecture", + "text_level": 1, + "bbox": [ + 215, + 618, + 346, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents the tokenization process, the structures of the encoder and decoder for each modality, and the architecture of the modality combiner network.", + "bbox": [ + 212, + 652, + 785, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Tokenization. All available modalities are spatially aligned through georeferencing. This allows us to divide the tile $x$ into a set $\\mathbf{P}$ of $P$ non-overlapping patches consistently across all modalities: $x_{p}^{\\mathbf{M}} = \\{x_{p}^{m}\\}_{m\\in \\mathbf{M}}$ corresponds to $M$ distinct views of the same patch $p$ with different modalities. Each modality $m$ takes its values in a space $\\varOmega^m$ such that $x_{p}^{m}\\in \\varOmega^{m}$ . We index tokens with pairs $(m,p)$ , defined for each modality $m$ and patch $p$ , for a total of $M\\times P$ tokens.", + "bbox": [ + 212, + 690, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Time series from Sentinel satellites may experience registration errors spanning several meters, complicating their precise alignment with high-resolution imagery. However, using temporal sequences of satellite data mitigates these errors as aggregation over time tends to balance out misalignments.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d67a4b45ff08d54139e398482e768f27ddaa24c13bd2614d0ef3f752fdd9cea.jpg", + "image_caption": [ + "(a) Image Encoder and Decoder." + ], + "image_footnote": [], + "bbox": [ + 233, + 148, + 380, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aeca5451ce9232f54390e78aabf62bcc3a8cbfe918721b1c3e96a2952cef638d.jpg", + "image_caption": [ + "(b) Temporal Encoder and Decoder." + ], + "image_footnote": [], + "bbox": [ + 405, + 147, + 547, + 309 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38032c22751dbd4962ef7b38027372d7c4e3f01048f4ac3ab6fcb3aa51de3663.jpg", + "image_caption": [ + "(c) Modality Combining Network.", + "Fig. 3: OmniSat Architecture. OmniSat is composed of dedicated patch encoders for image (a) and time series b, here represented for a length of $L = 4$ time stamps. The modality combining module $\\mathcal{C}$ is depicted in (c) with $P = 2$ and $M = 3$ . Elements colored in orange are learned networks or parameters." + ], + "image_footnote": [], + "bbox": [ + 568, + 145, + 779, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoder-Decoder for Images. We split image tiles split into small square patches: $\\varOmega^{\\mathrm{img}}=\\mathbb{R}^{C\\times W\\times W}$ with $W$ the size of the patches in pixels and $C$ the number of channels. As shown in Fig. 3a, we encode these inputs with a sequence of convolutions and max-pool layers until the spatial dimension is fully collapsed. Decoding involves a symmetric sequence of convolutions and un-pooling layers. Contrary to existing masked auto-encoders, we pass the pooling indices from the encoder's max-pooling to the decoder's un-pooling in the manner of SegNet [8]. This dispenses the encoder from learning the intra-patch spatial configuration. This allows the image encoder to focus on the radiometric information, which may be more relevant depending on the application.", + "bbox": [ + 212, + 419, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoder-Decoder for Time Series. Each temporal patch is represented by $L$ sequential observations with $C$ channels: $\\Omega^{\\mathrm{TS}} = \\mathbb{R}^{C\\times L}$ , each associated with a time stamp. We encode the temporal patches using a Lightweight Temporal Attention Encoder (LTAE) model [32], an efficient network for geospatial time series processing. We decode vector representations into time series by repeating the vector $L$ times across the temporal dimension, adding a temporal encoding for each time step, and using an MLP to map the results to size $C$ . See Fig. 3b for an illustration.", + "bbox": [ + 212, + 561, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optical time series are notoriously affected by clouds [85]. This may affect the validity of the reconstruction task: the decoder cannot know which observations are cloudy, making the reconstruction objective unpredictable. To circumvent this issue, we use the temporal attention maps of the encoder's LTAE to select dates to reconstruct: cloudless observations are more informative and should have a higher attention score [78]. We only consider in the reconstruction loss $\\mathcal{L}_{\\mathrm{reconstr}}$ the top $25\\%$ dates in terms of the LTAE's attention maps.", + "bbox": [ + 212, + 667, + 789, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Modality Combining Network. The modality combining network $\\mathcal{C}$ , represented in Fig. 3c, takes the $M \\times P$ token embeddings $f_{\\mathbf{P}}^{\\mathbf{M}}$ , some of whom can potentially be masked. We equip each token with a Euclidean relative positional encoding [94], calculated based on their patch's position $\\{r(p,q) \\mid (p,q) \\in \\mathbf{P}^2\\}$ , allowing each token to", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G.Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "selectively consider its spatial surroundings. As most EO data are captured from above (satellite or aerial), their distribution is invariant by horizontal translation, making this choice of encoding preferable to an absolute position encoding.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The modality combining module $\\mathcal{C}$ starts with a series of $B$ residual self-attention blocks connecting all tokens across modality. We then perform cross-attention between the resulting token embeddings $g_{\\mathbf{P}}^{\\mathbf{M}} \\in \\mathbb{R}^{d \\times M \\times P}$ and $P$ copies $f_{\\mathbf{P}}^{\\mathrm{comb}}$ of a modality combining token $f^{\\mathrm{comb}} \\in \\mathbb{R}^{d}$ learned as a free parameter. Each copy of $f_{p}^{\\mathrm{comb}}$ is spatially located at the patch $p$ for the relative positional encoding $r$ . The module $\\mathcal{C}$ outputs $P$ multimodal encodings $f_{\\mathbf{P}}^{\\star}$ combining all available modalities for each patch:", + "bbox": [ + 212, + 190, + 784, + 281 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\mathbf {P}} ^ {\\mathbf {M}} = \\text {s e l f - a t t e n t i o n} \\left(f _ {\\mathbf {P}} ^ {\\mathbf {M}}; r\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 290, + 787, + 308 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathbf {P}} ^ {f \\star} = \\text {c r o s s - a t t e n t i o n} \\left(f _ {\\mathbf {P}} ^ {\\text {c o m b}}, g _ {\\mathbf {P}} ^ {\\mathbf {M}}; r\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 311, + 785, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Contrastive Objective", + "text_level": 1, + "bbox": [ + 214, + 348, + 408, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We denote by $f_{p}^{m}$ the $d$ -dimensional encodings of the input patch $x_{p}^{m}$ given by their dedicated encoders. We propose to supervise the embeddings $f_{p}^{m}$ with a contrastive objective encouraging spatial consistency across modalities. Indeed, while each modality captures distinct characteristics of $p$ , all encodings $f_{p}^{m}$ share the same latent variable: the semantic content of the patch.", + "bbox": [ + 212, + 371, + 784, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In practice, we want $f_{p}^{m}$ to be closer to $f_{p}^{n}$ for $n \\neq m$ , than to $f_{q}^{n}$ for other patches $q \\neq p$ . We define $\\mathbf{B}$ as the set of patches within the current batch of observations. We adapt the classic InfoNCE loss [69] to our setting with two main differences, illustrated in Fig. 4. (i) Each token $(m,p)$ has $M - 1$ positive matches: the tokens corresponding to the same patch $p$ but viewed in another modality $n \\neq m$ ; and (ii) as EO observations are generally spatially regular, nearby patches may be visually indistinguishable. Therefore, we exclude from the negative matches of $(m,p)$ all tokens in modality $m$ and which are too close to $p$ . To this end, we remove the set $T(m,p)$ of tokens with modality $m$ and whose patches are in the same tile as $p$ . Our loss function $\\mathcal{L}_{\\mathrm{contrast}}$ is defined as such:", + "bbox": [ + 212, + 445, + 785, + 580 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a s t}} = \\frac {1}{M | \\mathbf {B} |} \\sum_ {(m, p) \\in \\mathbf {M} \\times \\mathbf {B}} \\log \\left(\\frac {\\sum_ {n \\neq m} \\exp \\left(\\left\\langle f _ {p} ^ {m} , f _ {p} ^ {n} \\right\\rangle / \\gamma\\right)}{\\sum_ {(n , q) \\in \\mathbf {M} \\times \\mathbf {B} \\backslash T (m , p)} \\exp \\left(\\left\\langle f _ {p} ^ {m} , f _ {q} ^ {n} \\right\\rangle / \\gamma\\right)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 590, + 785, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\gamma$ a temperature parameter, and $\\langle \\cdot ,\\cdot \\rangle$ the scalar product in $\\mathbb{R}^d$ . This function, specifically designed for geospatial data, allows us to contrast individual patches across modalities, which is not typically feasible for natural images. However, as the contrastive objective aligns multimodal representations, the patch encoders may be encouraged to overlook the distinct attributes of their respective modality. Instead, they may focus only on features shared by all modalities, i.e., their common denominator. To ensure that encoders also capture modality-specific information, we incorporate a reconstruction objective, detailed in Sec. 3.3.", + "bbox": [ + 212, + 645, + 787, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Multimodal Reconstruction Objective", + "text_level": 1, + "bbox": [ + 214, + 786, + 519, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, we mask a fraction of tokens $\\mathbf{K} \\subset \\mathbf{M} \\times \\mathbf{P}$ and replace their embeddings with a learned vector $f^{\\mathrm{mask}} \\in \\mathbb{R}^{d}$ . Note that the masking can differ across modalities,", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/509f9339c2eb39b7a94dc06dc20da0ac9a4c337a553e964b614eddf521e8846d.jpg", + "image_caption": [ + "Fig.4:Contrastive Loss. We represent the token matching matrix for two tiles $\\mathrm{Tile}_1$ and $\\mathrm{Tile}_2$ viewed across 3 modalities $m_{1}$ $m_2$ ,and $m_3$ Tile is composed of the patches $p_1$ and $p_2$ ,while $\\mathrm{Tile}_2$ comprises $q_{1}$ and $q_{2}$ .In contrast to classic approaches which ignore the diagonal and assign each sample with a single positive match, our loss defines operates at the patch level, considers multiple positives per token, and excludes tokens in a block-diagonal fashion." + ], + "image_footnote": [ + "+ positive match\n- negative match\no ignored" + ], + "bbox": [ + 218, + 143, + 475, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and some patches may be entirely masked. All tokens are then processed by the modality combining network $\\mathcal{C}$ , which outputs $P$ multimodal embeddings $f_{\\mathbf{P}}^{\\star}$ :", + "bbox": [ + 214, + 364, + 787, + 396 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathbf {P}} ^ {\\star} = \\mathcal {C} \\left(\\left\\{f _ {p} ^ {m} \\right\\} _ {(m, p) \\notin \\mathbf {K}} \\cup \\left\\{f ^ {\\text {m a s k}} \\right\\} _ {(m, p) \\in \\mathbf {K}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 405, + 787, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To encourage the patch embeddings $f_{\\mathbf{P}}^{*}$ to capture information from all modalities, we build a multimodal reconstruction objective. We denote by $\\mathcal{D}^m: \\mathbb{R}^d \\mapsto \\Omega^m$ the dedicated decoder of each modality $m$ and write the reconstruction loss as:", + "bbox": [ + 214, + 435, + 789, + 481 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n s t r}} = \\frac {1}{| \\mathbf {K} |} \\sum_ {(m, p) \\in \\mathbf {K}} \\frac {1}{\\dim (\\Omega^ {m})} \\| \\mathcal {D} ^ {m} \\left(f _ {p} ^ {\\star}\\right) - x _ {p} ^ {m} \\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 489, + 787, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\dim (\\varOmega^m)$ the dimension of $\\varOmega^m$ . The total loss is the sum of $\\mathcal{L}_{\\mathrm{reconstr}}$ and $\\mathcal{L}_{\\mathrm{contrast}}$ .", + "bbox": [ + 214, + 539, + 781, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 575, + 421, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We detail here the specific parameters chosen in all our experiments.", + "bbox": [ + 214, + 599, + 666, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tokenization. We split each tile along a regular spatial grid to produce a set of non-overlapping patches $\\mathbf{P}$ consistent across all modalities. For TreeSat and FLAIR, we use a $10 \\times 10\\mathrm{m}$ grid, meaning that the VHR input tokens are small image patches of size $50 \\times 50$ with $0.2\\mathrm{m}$ per pixel. The patches of Sentinel observations with a resolution of $10\\mathrm{m}$ are single-pixel temporal sequences of spectral measurements. For PASTIS-HD, we use a $40 \\times 40\\mathrm{m}$ grid, meaning that the VHR patches are of size $40 \\times 40$ with $1.0\\mathrm{m}$ per pixel. The patches of Sentinel observations [26] are $4 \\times 4$ image time series which we spatially flatten before encoding.", + "bbox": [ + 214, + 621, + 787, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameters. To show the versatility of OmniSat, we use the same configuration throughout all experiments. The embedding size is $d = 256$ , resulting in image encoders and decoders with 3.6M and 1.8M parameters, 403k and 96k for optical time series, and 402k and 95k for radar time series. The modality combiner module is composed of $B = 6$ residual self-attention blocks and a single cross-attention block, for a total of 3.6M parameters. We train our model on 3 A6000 GPUs with a batch size of 128", + "bbox": [ + 214, + 750, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G.Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "multimodal tiles per GPU and set the contrastive temperature $\\gamma$ to 0.1. We train our model with the ADAM optimizer [53], with a learning rate of $10^{-4}$ for pretraining and $2\\times 10^{-5}$ for fine-tuning, and a ReduceLROnPlateau scheduler [1] with a patience of 10 epochs and a decay rate of 0.1. When re-implementing competing methods, we use the hyperparameters of their open-source repository.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 244, + 357, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate OmniSat's performance across three multimodal datasets, including two new datasets introduced in this work, and presented in Sec. 4.1. We outline our experimental protocol and our adaptation of competing methods in Sec. 4.2. We then present our quantitative results and analysis in Sec. 4.3 and conduct an ablation study in Sec. 4.4.", + "bbox": [ + 212, + 275, + 787, + 335 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 358, + 316, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We consider three multimodal datasets: FLAIR [31], and the augmented TreeSatAI-TS [2] and PASTIS-HD [33, 34]. See Fig. 1 for an illustration of these two last datasets.", + "bbox": [ + 212, + 382, + 785, + 412 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "TreeSatAI-TS: TreeSatAI [2] is a multimodal dataset for tree species identification, containing 50,381 tiles of $60 \\times 60\\mathrm{m}$ with multi-label annotations for 20 classes and all taken in Germany. Each tile is associated with a very high resolution RGB and near-infrared (NIR) image ( $0.2\\mathrm{m}$ pixel resolution), a single Sentinel-2 multi-spectral image ( $10\\mathrm{m}$ per pixel resolution, 10 bands), and a single Sentinel-1 radar image ( $10\\mathrm{m}$ per pixel resolution, 3 bands: two polarization channels and their ratio).", + "bbox": [ + 212, + 419, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Motivated by the fact that fine-grained vegetation discrimination relies heavily on temporal dynamics information [91], we introduce TreeSatAI- $\\mathrm{TS}^1$ . This extended version uses open-source data to add Sentinel-1 and Sentinel-2 time series to each tile, spanning the closest available year to the VHR observation for Sentinel-2. Note that due to the weather patterns and position of the area of interest with respect to Sentinel-2's orbit, the optical time series is particularly irregular and occluded, with up to $50\\%$ of acquisitions being non-exploitable. Despite this challenge, we included the raw observations without pre-processing, whereas TreeSatAI's single-date images have been manually selected.", + "bbox": [ + 212, + 508, + 787, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PASTIS-HD: The PASTIS dataset [33], is designed for semantic and panoptic segmentation of agricultural parcels using Sentinel-2 time series and covers 18 crop types across 2,433 image time series with dimensions of $1280 \\times 1280\\mathrm{m}$ . Each series contains between 38 and 61 observations with 10 spectral bands. PASTIS-R [34] adds the corresponding Sentinel-1 radar time series. We only used the ascendant time series of Sentinel-1 for our training and evaluation, for a total of 169,587 radar images with three bands.", + "bbox": [ + 212, + 636, + 787, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To enhance the spatial resolution and utility of PASTIS, we introduce PASTISHD², which integrates contemporary VHR satellite images (SPOT 6-7 [24]). We apply orthorectification and pansharpening, resample the resulting images to a 1m resolution, and finally convert them to 8 bits. We follow the protocol of Irvin et al. [51] to use the", + "bbox": [ + 212, + 739, + 787, + 801 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1The dataset is available at https://huggingface.co/datasets/IGNF/TreeSatAI-Time-Series.", + "bbox": [ + 214, + 810, + 743, + 824 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "2The dataset is available at https://huggingface.co/datasets/IGNF/PASTIS-HD.", + "bbox": [ + 214, + 824, + 679, + 838 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "dense annotations for a multi-label classification task: each patch is associated with the labels of all of its pixels. This conversion allows us to evaluate all methods in the same setting and configuration as TreeSatAI.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "FLAIR. The FLAIR dataset [31] combines VHR aerial images with time series data. It comprises 77,762 aerial tiles ( $512 \\times 512$ pixels, $0.2\\mathrm{m}$ resolution) with five channels (RGB, near-infrared, and a normalized digital surface model) taken in France, alongside corresponding Sentinel-2 time series ( $10\\mathrm{m}$ resolution, 10 spectral bands, 20 to 114 observations per year). We apply the same processing as PASTIS to use the dense annotation for a multi-label classification task.", + "bbox": [ + 212, + 199, + 789, + 289 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental Setting", + "text_level": 1, + "bbox": [ + 214, + 314, + 405, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This section details our experimental protocol and our adaption of competing algorithms.", + "bbox": [ + 212, + 340, + 787, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol. All experiments follow a similar setting:", + "bbox": [ + 214, + 363, + 630, + 377 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pre-training (optional). Methods that support self-supervised pre-training (OmniSat, SatMAE [20], ScaleMAE [74], CROMA [29]) are pre-trained for up to 250 epochs on the entire training set without access to labels.", + "- Fine-Tuning. We propose two settings for fine-tuning:" + ], + "bbox": [ + 225, + 378, + 785, + 436 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Fully Supervised Fine-Tuning. We train the resulting models using all the labels in the training set.", + "- Semi-Supervised Fine-Tuning. We use a portion of $10\\%$ or $20\\%$ of the training set, stratified by the distribution of classes, to fine-tune the models. For models without pre-training, this corresponds to supervision in the low-data regime." + ], + "bbox": [ + 253, + 439, + 784, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- Unimodal and Multimodal Evaluation. We evaluate all methods using each available modality independently and combining all supported modalities.", + "bbox": [ + 225, + 516, + 785, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Adapting Competing Approaches. We report the performance of several methods taken from the literature on our considered datasets: LightGBM [2], PRESTO [88], and MOSAIKS [77]. However, few existing methods can operate on single- and multi-date data at the same time. To ensure a fair evaluation of competing approaches, we modify various state-of-the-art models to handle a broader combination of modalities. We provide details on these changes in the appendix.", + "bbox": [ + 212, + 555, + 787, + 646 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Numerical Experiments and Analysis", + "text_level": 1, + "bbox": [ + 214, + 670, + 516, + 684 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we report our model's performance and efficiency compared to other approaches across the considered datasets and propose our analysis.", + "bbox": [ + 212, + 696, + 785, + 727 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "TreeSatAI-TS. Tab. 2 presents the performance of different models on TreeSatAI and TreeSatAI-TS. We report several key observations:", + "bbox": [ + 212, + 734, + 785, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- Benefit of Time Series. For the original TreeSatAI dataset with single-date Sentinel1/2 observations, none of the pre-training schemes significantly improve performance beyond simple baselines such as ResNet, PSE, or MLP, even in a semi-supervised setting. In particular, single-date S1 observations yield low performance for all methods (below 20 F1-score), emphasizing the need to use the entire time series.", + "bbox": [ + 225, + 765, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G.Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5d2858b392ceab9018227921915a045d9153167f7f0643228cc7d1b1b53b66e6.jpg", + "table_caption": [ + "Table 2: Performance on TreeSatAI-TS. We report the weighted F1 for multi-label tree species classification on TreeSatAI (TSAI) and our extended TreeSatAI-TS (TSAI-TS) dataset when fine-tuning with $10\\%$ and $100\\%$ of training labels. The first line of the table is the modality used for evaluation. We distinguish methods that are best for one modality within a dataset, best in a dataset across all modalities, and the best overall performance. $\\star$ : late feature fusion with a ResNet pre-trained on ImageNet. $\\blacksquare$ : Foundation model trained on extensive external data. $\\dagger$ : model evaluated on this dataset for the first time." + ], + "table_footnote": [], + "table_body": "
Modelpre-trainingAllSentinel-1Sentinel-2VHR Image
10%100%10%100%10%100%10%100%
Evaluated on TreeSatAI: single date for Sentinel-1 and Sentinel-2
† PSE [35]None47.2*68.1*11.514.648.558.3--
† ViT [25]None42.757.18.717.539.857.336.751.7
MLP [2]None42.6*71.5*3.410.122.152.0--
ResNet [2]ImageNet------58.870.1
LightGBM [2]ImageNet-54.3*-11.9-48.2-44.0
PRESTO [88]---19.8-46.3--
† DOFA [96]59.571.611.619.348.257.051.667.5
MOSAIKS [22,77]TSAI-----56.0--
† CROMA [29]TSAI49.661.010.112.747.855.7--
† SatMAE [20]TSAI46.161.5--40.349.744.161.4
† ScaleMAE [74]TSAI47.662.5--46.755.246.963.6
OmniSat (ours)TSAI56.270.45.36.448.457.152.868.9
Evaluated on TreeSatAI-TS: Sentinel-1 and Sentinel-2 time series spanning one year
† PSE+LTAE [35]None59.4*71.2*42.652.444.057.2--
† UT&T [31]ImageNet43.856.742.355.241.557.044.355.9
† DOFA [96]41.871.30.00.025.039.451.667.5
† Scale-MAE [74]TSAI-TS44.160.4--11.031.546.963.6
OmniSat (ours)None52.273.331.655.933.949.751.471.0
OmniSat (ours)TSAI-TS61.174.248.256.751.462.958.370.5
", + "bbox": [ + 218, + 250, + 784, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "OmniSat exhibits significantly improved results on TreeSatAI-TS, with or without pretraining. Image models struggles to extract meaningful features temporally aggregated temporal observations, while OmniSat learn rich dynamic features.", + "bbox": [ + 240, + 657, + 787, + 704 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The foundation model DOFA [96], with 111M parameters and a large closed-source training set, outperforms all models when evaluated on single-date modalities. However, OmniSat reaches higher performances on TreeSatAI-TS with only 10 million parameters, which we attribute to its ability to leverage temporal modalities.", + "bbox": [ + 240, + 704, + 787, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "- Benefits of Multimodality. When using all modalities, OmniSat outperforms all competing methods by a margin of $3\\%$ F1-score. The multimodal performance of OmniSat and CROMA, which learn to combine data sources, is strictly superior to the F1-score of their best modality by $3.7\\%$ and $5.3\\%$ points, respectively. Conversely, the performance of methods that rely on late-fusion (SatMAE, ScaleMAE,", + "bbox": [ + 225, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d13c290ff9ebebdf8e6fb1180059ddc0f169baf761486f08a97fc810cd4a2473.jpg", + "table_caption": [ + "Table 3: Performance on PASTIS-HD. We report the macro-averaged F1-score for crop-type multi-class classification on the PASTIS-HD dataset. We distinguish methods that are best for one modality, best in a dataset across all modalities. $\\star$ : late feature fusion with a ResNet. $\\dagger$ : model evaluated on this dataset for the first time." + ], + "table_footnote": [], + "table_body": "
Modelpre-trainedAllSentinel-1Sentinel-2VHR image
20%100%20%100%20%100%20%100%
† UTAE [33, 34]None36.8*46.9*20.140.732.737.6--
† ResNet50 [47]ImageNet------57.659.3
† UT&T [31]ImageNet54.253.558.862.854.961.351.149.8
† DOFA [96]53.755.736.741.550.853.447.954.8
† Scale-MAE [74]PASTIS-HD42.042.2--41.246.148.851.9
† CROMA [29]PASTIS-HD57.560.155.356.153.056.7--
OmniSat (ours)No42.059.158.260.251.760.147.352.8
OmniSat (ours)PASTIS-HD62.669.960.869.061.870.854.659.3
", + "bbox": [ + 218, + 209, + 782, + 385 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/676a6993c1abb48d833a1d362184f58e62d58de409d0443d1e0b48066416521a.jpg", + "table_caption": [ + "Table 4: Performance on FLAIR. We report the macro-averaged F1-score for land cover multiclass classification on the FLAIR dataset. We distinguish methods that are best for one modality and best in a dataset. $\\dagger$ : model evaluated on this dataset for the first time." + ], + "table_footnote": [], + "table_body": "
Modelpre-trainedAllSentinel-2VHR Image
10%100%10%100%10%100%
† UT&T [31]ImageNet44.248.857.462.058.965.5
† DOFA [96]70.674.957.061.066.872.1
† ScaleMAE [74]FLAIR63.170.052.561.061.267.3
OmniSat (ours)No62.570.056.165.464.771.5
OmniSat (ours)FLAIR60.673.456.865.465.271.6
", + "bbox": [ + 269, + 454, + 725, + 588 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ViT) is comparable to their best modality. This demonstrates the value of learning to combine information from different sources end-to-end.", + "bbox": [ + 240, + 619, + 785, + 648 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "- Benefits of Cross-Modal Pre-Training. With access to all modalities, our self-supervised pre-training improves by $0.9\\%$ point the F1-score of the model fine-tuned on $100\\%$ of labels, compared to not pre-training, and $8.9\\%$ when using only $10\\%$ of labels. This shows that our pre-training leads to more expressive multimodal features. Interestingly, when performing inference with Sentinel-2 time series alone, the performance increase linked to the pre-training becomes $13.2\\%$ with $100\\%$ labels and $17.5\\%$ with $10\\%$ . This illustrates that our self-supervised pre-training scheme improves the features learned by each encoder despite not relying on annotated data.", + "bbox": [ + 225, + 650, + 787, + 772 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experiments on PASTIS-HD. The analysis of the performance of various models on PASTIS-HD is reported in Tab. 3, and is consistent with the ones of TreeSatAI-TS. First, by learning to combine all modalities despite their different resolutions, OmniSat achieves state-of-the-art results on this benchmark. Second, our cross-modal pretraining", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G.Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/de0d8468d42cad0b4ccc124f4aeb30b473d15b7ff6c72aeacecfa27a202211ca.jpg", + "image_caption": [ + "Fig. 5: Efficiency. We report the best performance of different models between TreeSatAI and TreeSatAI-TS, with pre-training and fine-tuning using $100\\%$ of labels. The area of the markers is proportional to the training time, broken down in pre-training and fine-tuning when applicable." + ], + "image_footnote": [], + "bbox": [ + 238, + 143, + 781, + 325 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "significantly improves OmniSat's performance in the multimodal $(+10.8\\mathrm{pF1}$ -score with $100\\%$ of training label) and all single-modality settings (8.8 points for Sentinel-1, 10.7 for Sentinel-2, and 6.5 for the VHR images).", + "bbox": [ + 212, + 401, + 787, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Experiments on FLAIR. We report in Tab. 4 the results on the bimodal FLAIR dataset for multilabel classification. OmniSat outperforms the much larger ScaleMAE [74] and UT&T [31] models with $100\\%$ of labels and both modalities by $3.4\\%$ . Our pre-training scheme had a smaller impact than for the TreeSatAI-TS experiment. We attribute this to the fact that only two modalities are available, which decreases the supervisory power of our cross-modal contrastive objective and our multimodal reconstruction loss. This highlights a limitation of OmniSat: the model needs to be pre-trained on a modality-rich dataset to achieve its best performance.", + "bbox": [ + 212, + 454, + 787, + 575 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Efficiency Evaluation. We plot in Fig. 5 the best performance between TreeSatAI and TreeSatAI-TS for different models according to their size and training time. OmniSat is more compact, faster to train, and performs better than all evaluated models, including the DOFA foundation model. The highly-specialized combination of PSE, LTAE, and ResNet is a strong contender, outperforming significantly larger models with generic encoding-decoding schemes.", + "bbox": [ + 212, + 582, + 787, + 674 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 690, + 364, + 705 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we report the results of several experiments evaluating the impact and validity of our main design choices, see Tab. 5.", + "bbox": [ + 212, + 710, + 785, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "a) Encoder/Decoder Architecture. We propose several improvements to the standard image encoder-decoder scheme used in computer vision to accommodate the specificities of EO data. In particular, passing the max-pool indices from the image patch encoder to its decoder allows the learned representation to focus on characterizing the spectral signature instead of fine-grained spatial information, and leads to a performance increase of $0.7\\%$ in the full supervision setting.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/98fd15b3a5a51ab06fb7ae8978d16f472716ba6f17232aa23201a20afb861fba.jpg", + "table_caption": [ + "Table 5: Ablation Study. We present the impact of several design choices on the TreeSatAI-TS dataset, measured in terms of macro-averaged F1-score." + ], + "table_footnote": [], + "table_body": "
Experiment10%100%Experiment10%100%
OmniSat61.174.2b) no contrastive loss55.673.4
a) no index bypass57.573.5b) naive contrastive loss57.872.2
a) no date filtering58.271.6b) no reconstruction loss59.072.2
", + "bbox": [ + 282, + 181, + 715, + 253 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As clouds frequently obstruct optical time series, we use a unsupervised date-filtering scheme to reconstruct only meaningful acquisitions. This approach leads to a significant improvement of $3.6\\%$ , showcasing the benefit of developing modality-aware approaches for EO.", + "bbox": [ + 212, + 284, + 785, + 344 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "b) Role of Loss Functions. When training without contrastive loss, we observe a decrease in performance of $0.8\\%$ in the fully supervised regime, and a more pronounced drop of $5.5\\%$ in the semi-supervised regime. This demonstrates how learning consistent encoding across encoders facilitates their subsequent fusion. Interestingly, when implementing a naive contrastive loss that considers all negative examples from the batch, the decrease is greater than simply removing this loss ( $2\\%$ in full supervision). This strategy may introduce indistinguishable negative examples and perturb the learning process.", + "bbox": [ + 212, + 352, + 785, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also remove the reconstruction loss, meaning that only the encoders are learned contrastively during pre-training. This results in a drop of $2\\%$ F1-score point, illustrating the importance of pre-training the transformer $\\mathcal{C}$ alongside its encoders.", + "bbox": [ + 212, + 455, + 785, + 501 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. All datasets used in our experiments are based in Europe, primarily due to the availability of open-access annotations. This regional focus prevents us from evaluating our model's performance in tropical and developing countries, which present unique challenges in terms of label provision, heterogeneity, and complex classes.", + "bbox": [ + 212, + 508, + 785, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A limitation of our pre-training scheme is its dependence on a sufficient number of aligned modalities, as illustrated by its moderate impact on the bimodal FLAIR dataset.", + "bbox": [ + 212, + 568, + 785, + 597 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 625, + 344, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced OmniSat, a new architecture for the self-supervised modality fusion of Earth Observation (EO) data from multiple sources. To facilitate its evaluation, we augmented two existing datasets with new modalities of different natures and resolutions. We experimentally showed that leveraging diverse modalities with a flexible model improves the model's performance in both fully and semi-supervised settings. Moreover, our training scheme can exploit the spatial alignment of multiple modalities to improve our model's unimodal performance. Finally, we proposed several improvements to leverage the unique structure of EO data in the architecture of our model, such as automatic date filtering for reconstructing time series. We hope that our promising results and new datasets will encourage the computer vision community to consider EO data as a playing field for evaluating and developing novel self-supervised multimodal algorithms.", + "bbox": [ + 212, + 657, + 787, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G.Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by ANR project READY3D ANR-19-CE23-0007, and was granted access to the HPC resources of IDRIS under the allocations AD011014719 and AD011014286R1 made by GENCI. We thank Anatol Garioud and Sébastien Giordano for their help on the creation of TreeSatAI-TS and PASTIS-HD datasets. The SPOT images are opendata thanks to the Dataterra Dinamis initiative in the case of the \"Couverture France DINAMIS\" program. We thank Jordi Inglada for inspiring discussions and valuable feedback.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 305, + 310, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. PyTorch: ReduceLROOnPlateau.org/docs/stable/generated/torch.optim.lr_scheduler.ReduceLROOnPlateau.html#torch.optim.lr_scheduler.ReduceLROOnPlateau, accessed: 2024-02-29 9", + "2. Ahlswede, S., Schulz, C., Gava, C., Helber, P., Bischke, B., Förster, M., Arias, F., Hees, J., Demir, B., Kleinschmit, B.: TreeSatAI Benchmark Archive: A multi-sensor, multi-label dataset for tree species classification in remote sensing. Earth System Science Data Discussions (2022) 2, 3, 4, 9, 10, 11", + "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: A visual language model for few-shot learning. In: NeurIPS (2022) 3", + "4. Amitrano, D., Di Martino, G., Guida, R., Iervolino, P., Iodice, A., Papa, M.N., Riccio, D., Ruello, G.: Earth environmental monitoring using multi-temporal synthetic aperture radar: A critical review of selected applications. Remote Sensing (2021) 1", + "5. Anderson, K., Ryan, B., Sonntag, W., Kavvada, A., Friedl, L.: Earth observation in service of the 2030 agenda for sustainable development. Geo-spatial Information Science (2017) 1", + "6. Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding predictive architecture. In: CVPR (2023) 2", + "7. Ayush, K., Uzkent, B., Meng, C., Tanmay, K., Burke, M., Lobell, D., Ermon, S.: Geography-aware self-supervised learning. In: ICCV (2021) 2", + "8. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE TPAMI (2017) 6", + "9. Baevski, A., Hsu, W.N., Xu, Q., Babu, A., Gu, J., Auli, M.: Data2vec: A general framework for self-supervised learning in speech, vision and language. In: ICML (2022) 3", + "10. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2021) 2", + "11. Bao, X., Zhang, R., Lv, J., Wu, R., Zhang, H., Chen, J., Zhang, B., Ouyang, X., Liu, G.: Vegetation descriptors from Sentinel-1 SAR data for crop growth monitoring. ISPRS Journal of Photogrammetry and Remote Sensing (2023) 4", + "12. Bastani, F., Wolters, P., Gupta, R., Ferdinando, J., Kembhavi, A.: SatlasPretrain: A large-scale dataset for remote sensing image understanding. In: ICCV (2023) 3, 4", + "13. Bayoudh, K., Knani, R., Hamdaoui, F., Mtibaa, A.: A survey on deep multimodal learning for computer vision: Advances, trends, applications, and datasets. The Visual Computer (2022) 3" + ], + "bbox": [ + 225, + 335, + 789, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Benedetti, P., Ienco, D., Gaetano, R., Ose, K., Pensa, R.G., Dupuy, S.: $\\mathbf{M}^3$ -fusion: A deep learning architecture for multiscale multimodal multitemporal satellite data fusion. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2018) 2", + "15. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020) 2", + "16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 2", + "17. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020) 2", + "18. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021) 2", + "19. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018) 4", + "20. Cong, Y., Khanna, S., Meng, C., Liu, P., Rozi, E., He, Y., Burke, M., Lobell, D., Ermon, S.: SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In: NeurIPS (2022) 2, 3, 10, 11", + "21. Coppin, P., Lambin, E., Jonckheere, I., Muys, B.: Digital change detection methods in natural ecosystem monitoring: A review. Analysis of multi-temporal remote sensing images (2002) 1", + "22. Corley, I., Robinson, C., Dodhia, R., Ferres, J.M.L., Najafirad, P.: Revisiting pre-trained remote sensing model benchmarks: Resizing and normalization matters. arXiv preprint arXiv:2305.13456 (2023) 11", + "23. Dai, A., Nießner, M.: 3DMV: Joint 3D-multi-view prediction for 3D semantic scene segmentation. In: ECCV (2018) 2", + "24. DataTerra Dinamis: Diffusion OpenData Dinamis, https://dinamis.data-terra.org/opendata/, accessed: 2023-12-15 9", + "25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 11", + "26. Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B., Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F., Bargellini, P.: Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment (2012) 1, 4, 8", + "27. Ebel, P., Xu, Y., Schmitt, M., Zhu, X.X.: SEN12MS-CR-TS: A remote-sensing data set for multimodal multitemporal cloud removal. IEEE TGRS (2022) 3", + "28. Ekim, B., Stomberg, T.T., Roscher, R., Schmitt, M.: MapInWild: A remote sensing dataset to address the question of what makes nature wild. IEEE Geoscience and Remote Sensing Magazine (2023) 3", + "29. Fuller, A., Millard, K., Green, J.R.: CROMA: Remote sensing representations with contrastive radar-optical masked autoencoders. In: NeurIPS (2023) 1, 2, 3, 10, 11, 12", + "30. Gao, Y., Sun, X., Liu, C.: A general self-supervised framework for remote sensing image classification. Remote Sensing (2022) 2", + "31. Garioud, A., Gonthier, N., Landrieu, L., De Wit, A., Valette, M., Poupee, M., Giordano, S., Wattrelos, B.: FLAIR: A country-scale land cover semantic segmentation dataset from multi-source optical imagery. In: NeurIPS Dataset and Benchmark (2023) 3, 4, 9, 10, 11, 12, 13", + "32. Garnot, V.S.F., Landrieu, L.: Lightweight temporal self-attention for classifying satellite images time series. In: Advanced Analytics and Learning on Temporal Data: ECML PKDD Workshop (2020) 6", + "33. Garnot, V.S.F., Landrieu, L.: Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In: ICCV (2021) 2, 3, 9, 12" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 218, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Garnot, V.S.F., Landrieu, L., Chehata, N.: Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing (2022) 2, 3, 4, 9, 12", + "35. Garnot, V.S.F., Landrieu, L., Giordano, S., Chehata, N.: Satellite image time series classification with pixel-set encoders and temporal self-attention. In: CVPR (2020) 11", + "36. Ghamisi, P., Rasti, B., Yokoya, N., Wang, Q., Hofle, B., Bruzzone, L., Bovolo, F., Chi, M., Anders, K., Gloaguen, R., et al.: Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine (2019) 1, 3", + "37. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018) 2", + "38. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023) 1, 2, 3", + "39. Girdhar, R., Singh, M., Ravi, N., van der Maaten, L., Joulin, A., Misra, I.: Omnivore: A single model for many visual modalities. In: CVPR (2022) 2, 3", + "40. Goldberg, H.R., Ratto, C.R., Banerjee, A., Kelbaugh, M.T., Giglio, M., Vermote, E.F.: Automated global-scale detection and characterization of anthropogenic activity using multi-source satellite-based remote sensing imagery. In: Geospatial Informatics XIII. SPIE (2023) 4", + "41. Greenwell, C., Crall, J., Purri, M., Dana, K., Jacobs, N., Hadzic, A., Workman, S., Leotta, M.: WATCH: Wide-area terrestrial change hypercube. In: WACV (2024) 2", + "42. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. In: NeurIPS (2020) 2", + "43. Hackstein, J., Sumbul, G., Clasen, K.N., Demir, B.: Exploring masked autoencoders for sensor-agnostic image retrieval in remote sensing. arXiv preprint arXiv:2401.07782 (2024) 2", + "44. Hazirbas, C., Ma, L., Domokos, C., Cremers, D.: FuseNet: Incorporating depth into semantic segmentation via fusion-based CNN architecture. In: ACCV (2017) 3", + "45. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2", + "46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 2", + "47. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 12", + "48. Hu, J., Liu, R., Hong, D., Camero, A., Yao, J., Schneider, M., Kurz, F., Segl, K., Zhu, X.X.: MDAS: A new multimodal benchmark dataset for remote sensing. Earth System Science Data Discussions (2022) 3, 4", + "49. Huang, P.Y., Sharma, V., Xu, H., Ryali, C., Fan, H., Li, Y., Li, S.W., Ghosh, G., Malik, J., Feichtenhofer, C.: MAViL: Masked audio-video learners. In: NeurIPS (2023) 2, 3", + "50. Ibanez, D., Fernandez-Beltran, R., Pla, F., Yokoya, N.: Masked auto-encoding spectral-spatial transformer for hyperspectral image classification. IEEE TGRS (2022) 2", + "51. Irvin, J., Tao, L., Zhou, J., Ma, Y., Nashold, L., Liu, B., Ng, A.Y.: USat: A unified self-supervised encoder for multi-sensor satellite imagery. arXiv preprint arXiv:2312.02199 (2023) 1, 2, 3, 9", + "52. Kenton, J.D.M.W.C., Toutanova, L.K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 2", + "53. Kingma, D.P., Ba, J.: ADAM: A method for stochastic optimization. ICLR (2015) 9", + "54. Krispel, G., Opitz, M., Waltner, G., Possegger, H., Bischof, H.: Fuseseg: LiDAR point cloud segmentation fusing multi-modal data. In: WACV (2020) 2" + ], + "bbox": [ + 223, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Kuffer, M., Thomson, D.R., Boo, G., Mahabir, R., Grippa, T., Vanhuysse, S., Engstrom, R., Ndugwa, R., Makau, J., Darin, E., et al.: The role of Earth observation in an integrated deprived area mapping \"system\" for low-to-middle income countries. Remote sensing (2020) 1", + "56. Lacoste, A., Sherwin, E.D., Kerner, H., Alemohammad, H., Lütjens, B., Irvin, J., Dao, D., Chang, A., Gunturkun, M., Drouin, A., et al.: Toward foundation models for Earth monitoring: Proposal for a climate change benchmark. arXiv preprint arXiv:2112.00570 (2021) 1", + "57. Li, D., Tong, Q., Li, R., Gong, J., Zhang, L.: Current issues in high-resolution Earth observation technology. Science China Earth sciences (2012) 1", + "58. Li, J., Hong, D., Gao, L., Yao, J., Zheng, K., Zhang, B., Chanussot, J.: Deep learning in multimodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation 112, 102926 (2022) 3", + "59. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE TPAMI (2022) 2", + "60. Liu, Y., Li, X., Hua, Z., Xia, C., Zhao, L.: A band selection method with masked convolutional autoencoder for hyperspectral image. IEEE Geoscience and Remote Sensing Letters (2022) 2", + "61. Ma, Y., Li, Y., Feng, K., Xia, Y., Huang, Q., Zhang, H., Prieur, C., Licciardi, G., Malha, H., Chanussot, J., et al.: The outcome of the 2021 IEEE GRSS data fusion contest-Track DSE: Detection of settlements without electricity. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3", + "62. Mai, G., Huang, W., Sun, J., Song, S., Mishra, D., Liu, N., Gao, S., Liu, T., Cong, G., Hu, Y., et al.: On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798 (2023) 1", + "63. Manas, O., Lacoste, A., Giró-i Nieto, X., Vazquez, D., Rodriguez, P.: Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In: ICCV (2021) 2", + "64. Manfreda, S., McCabe, M.F., Miller, P.E., Lucas, R., Pajuelo Madrigal, V., Mallinis, G., Ben Dor, E., Helman, D., Estes, L., Ciraolo, G., et al.: On the use of unmanned aerial systems for environmental monitoring. Remote sensing p. 641 (2018) 1", + "65. Moreira, A., Prats-Iraola, P., Younis, M., Krieger, G., Hajnsek, I., Papathanassiou, K.P.: A tutorial on synthetic aperture radar. IEEE Geoscience and Remote Sensing Magazine (2013) 1", + "66. Nakalembe, C.: Urgent and critical need for sub-saharan african countries to invest in Earth observation-based agricultural early warning and monitoring systems. Environmental Research Letters (2020) 1", + "67. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from RGBD images. In: ECCV (2012) 2", + "68. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016) 2", + "69. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 7", + "70. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. TLMR (2023) 2, 3", + "71. Pohl, C., Van Genderen, J.L.: Multisensor image fusion in remote sensing: Concepts, methods and applications. International journal of remote sensing (1998) 3", + "72. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3" + ], + "bbox": [ + 223, + 146, + 787, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 218, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G.Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "73. Recasens, A., Lin, J., Carreira, J., Jaegle, D., Wang, L., Alayrac, J.b., Luc, P., Miech, A., Smaira, L., Hemsley, R., et al.: Zorro: The masked multimodal transformer. arXiv preprint arXiv:2301.09595 (2023) 2", + "74. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: ICCV (2023) 2, 3, 10, 11, 12, 13", + "75. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3D semantic segmentation. In: CVPR (2022) 2", + "76. Robinson, C., Malkin, K., Jojic, N., Chen, H., Qin, R., Xiao, C., Schmitt, M., Ghamisi, P., Hänsch, R., Yokoya, N.: Global land-cover mapping with weak supervision: Outcome of the 2020 IEEE GRSS data fusion contest. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3, 4", + "77. Rolf, E., Proctor, J., Carleton, T., Bolliger, I., Shankar, V., Ishihara, M., Recht, B., Hsiang, S.: A generalizable and accessible approach to machine learning with global satellite imagery. Nature communications (2021) 10, 11", + "78. Rußwurm, M., Körner, M.: Self-attention for raw optical satellite time series classification. ISPRS Journal of Photogrammetry and Remote Sensing (2020) 6", + "79. Schmitt, M., Zhu, X.X.: Data fusion and remote sensing: An ever-growing relationship. IEEE Geoscience and Remote Sensing Magazine (2016) 1, 3", + "80. Secades, C., O'Connor, B., Brown, C., Walpole, M., et al.: Earth observation for biodiversity monitoring: A review of current approaches and future opportunities for tracking progress towards the aichi biodiversity targets. CBD technical series (2014) 1", + "81. Shermeyer, J., Hogan, D., Brown, J., Van Etten, A., Weir, N., Pacifici, F., Hansch, R., Bastidas, A., Soenen, S., Bacastow, T., et al.: Spacenet 6: Multi-sensor all weather mapping dataset. In: CVPR Workshop EarthVision (2020) 3", + "82. Shukor, M., Dancette, C., Rame, A., Cord, M.: UnIVAL: Unified model for image, video, audio and language tasks. TMLR (2023) 1, 2, 3", + "83. Skidmore, A.K., Coops, N.C., Neinavaz, E., Ali, A., Schaepman, M.E., Paganini, M., Kissling, W.D., Vihervaara, P., Darvishzadeh, R., Feilhauer, H., et al.: Priority list of biodiversity metrics to observe from space. Nature Ecology & Evolution (2021) 1", + "84. Srivastava, S., Sharma, G.: OmniVec: Learning robust representations with cross modal sharing. In: WACV (2024) 2, 3", + "85. Sudmanns, M., Tiede, D., Augustin, H., Lang, S.: Assessing global Sentinel-2 coverage dynamics and data availability for operational Earth observation (EO) applications using the EO-Compass. International Journal of Digital Earth (2019) 6", + "86. Sumbul, G., De Wall, A., Kreuziger, T., Marcelino, F., Costa, H., Benevides, P., Caetano, M., Demir, B., Markl, V.: BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine (2021) 3, 4", + "87. Tarasiou, M., Chavez, E., Zafeiriou, S.: ViTs for SITS: Vision transformers for satellite image time series. In: CVPR (2023) 3", + "88. Tseng, G., Zvonkov, I., Purohit, M., Rolnick, D., Kerner, H.: Lightweight, pre-trained transformers for remote sensing timeseries. arXiv preprint arXiv:2304.14065 (2023) 3, 10, 11", + "89. Tseng, W.H., Lé, H.A., Boulch, A., Lefèvre, S., Tiede, D.: CROCO: Cross-modal contrastive learning for localization of Earth observation data. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 2", + "90. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3" + ], + "bbox": [ + 225, + 147, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Cross-Modal Token Alignment", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 769, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "91. Vrieling, A., Meroni, M., Darvishzadeh, R., Skidmore, A.K., Wang, T., Zurita-Milla, R., Oosterbeek, K., O'Connor, B., Paganini, M.: Vegetation phenology from Sentinel-2 and field cameras for a Dutch barrier island. Remote sensing of environment (2018) 4, 9", + "92. Wang, Y., Graham, N.A.A., Xiong, Z., Liu, C., Albrecht, C.M., Zhu, X.X.: SSL4EO-S12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in Earth observation. IEEE Geoscience and Remote Sensing Magazine (2023) 3", + "93. Wenger, R., Puisant, A., Weber, J., Idoumghar, L., Forestier, G.: MultiSenGE: A multimodal and multitemporal benchmark dataset for land use/land cover remote sensing applications. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 3", + "94. Wu, K., Peng, H., Chen, M., Fu, J., Chao, H.: Rethinking and improving relative position encoding for vision transformer. In: ICCV (2021) 6", + "95. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMim: A simple framework for masked image modeling. In: CVPR (2022) 2", + "96. Xiong, Z., Wang, Y., Zhang, F., Stewart, A.J., Hanna, J., Borth, D., Papoutsis, I., Saux, B.L., Camps-Valls, G., Zhu, X.X.: Neural plasticity-inspired foundation model for observing the Earth crossing modalities. arXiv preprint arXiv:2403.15356 (2024) 3, 4, 11, 12", + "97. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change (2013) 1", + "98. Yang, M.Y., Landrieu, L., Tuia, D., Toth, C.: Multi-modal learning in photogrammetry and remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing (2021) 3", + "99. Yuan, Y., Lin, L., Liu, Q., Hang, R., Zhou, Z.G.: SITS-Former: A pre-trained spatio-spectral-temporal representation model for sentinel-2 time series classification. International Journal of Applied Earth Observation and Geoinformation (2022) 2", + "100. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016) 2", + "101. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: Image BERT pre-training with online tokenizer. In: ICLR (2022) 3", + "102. Zong, Y., Mac Aodha, O., Hospedales, T.: Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008 (2023) 1" + ], + "bbox": [ + 220, + 146, + 787, + 549 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "G. Astruc et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_model.json b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..75d7866d2fef5567182f2dab787d4bf13b804af0 --- /dev/null +++ b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_model.json @@ -0,0 +1,3309 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.285, + 0.141, + 0.72, + 0.184 + ], + "angle": 0, + "content": "OmniSat: Self-Supervised Modality Fusion for Earth Observation" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.212, + 0.658, + 0.243 + ], + "angle": 0, + "content": "Guillaume Astruc \\(^{1,3,4}\\), Nicolas Gonthier \\(^{1,2}\\), Clement Mallet \\(^{1}\\), and Loic Landrieu \\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.255, + 0.75, + 0.283 + ], + "angle": 0, + "content": "1 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France \n2 IGN, France \n3 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France \n4 CNES, France" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.324, + 0.744, + 0.518 + ], + "angle": 0, + "content": "Abstract. The diversity and complementarity of sensors available for Earth Observations (EO) calls for developing bespoke self-supervised multimodal learning approaches. However, current multimodal EO datasets and models typically focus on a single data type, either mono-date images or time series, which limits their impact. To address this issue, we introduce OmniSat, a novel architecture able to merge diverse EO modalities into expressive features without labels by exploiting their alignment. To demonstrate the advantages of our approach, we create two new multimodal datasets by augmenting existing ones with new modalities. As demonstrated for three downstream tasks—forestry, land cover classification, and crop mapping—OmniSat can learn rich representations without supervision, leading to state-of-the-art performances in semi- and fully supervised settings. Furthermore, our multimodal pretraining scheme improves performance even when only one modality is available for inference. The code and dataset are available at https://github.com/gastruc/OmniSat." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.531, + 0.7, + 0.546 + ], + "angle": 0, + "content": "Keywords: Earth observation \\(\\cdot\\) Multi-modality \\(\\cdot\\) Self-supervised learning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.571, + 0.357, + 0.587 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.722 + ], + "angle": 0, + "content": "Self-supervised multimodal learning has recently gathered significant interest within computer vision [38, 82, 102]. Earth Observation (EO) is particularly well-suited for developing and evaluating such approaches [29, 51], thanks to the large amount of open-access data captured by sensing technologies with complementary capabilities [36, 79]. Combining different sources of EO observations is crucial for several high-impact applications, including environmental [21, 80, 83] and climate monitoring [56, 97], as well as improving food security [66]. Moreover, learning with few or no labels is essential for developing regions with limited data annotation capabilities [5, 55, 62]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.788, + 0.811 + ], + "angle": 0, + "content": "Despite this potential, most multimodal EO datasets and models focus on a single data type, either mono-date images or time series. This limitation prevents them from simultaneously leveraging the spatial resolution of aerial images [57, 64], the temporal and spectral resolutions of optical satellite time series [26], and the resilience of radar to weather effects [4, 65]. Additionally, existing approaches are often limited for a given set of sensors, limiting their applicability." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address these challenges, we introduce OmniSat, a novel architecture designed for the self-supervised fusion of diverse EO data. Existing multimodal approaches often map" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.128 + ], + "angle": 0, + "content": "G. Astruc et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "multiple unrelated observations from different modalities to one pivot modality [38, 82] or a shared latent space [39, 84]. In contrast, OmniSat merges multiple views of the same area from different modalities into a single representation combining the specific information of each modality [14, 41, 73]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.787, + 0.295 + ], + "angle": 0, + "content": "In computer vision, obtaining finely aligned multimodal observations generally requires specialized sensors [54, 59, 67] or the computation of complex mappings between modalities [23, 75]. On the other hand, EO data can be naturally aligned with georeferencing. To leverage this property, we adapt multimodal contrastive learning [49, 72] and cross-modal masked auto-encoding techniques [43] to learn rich multimodal EO representations with a generalist fusion scheme and without annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.787, + 0.369 + ], + "angle": 0, + "content": "To address the scarcity of EO datasets with a diverse range of heterogeneous modalities (see Tab. 1), we enrich the TreeSatAI [2] and PASTIS-R [33, 34] datasets with new aligned modalities. This allows us to evaluate OmniSat's ability to handle an arbitrary number of inputs with varying natures and resolutions. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.37, + 0.784, + 0.414 + ], + "angle": 0, + "content": "- We introduce OmniSat, a new model that learns to combine varied sources of EO observations in a self-supervised manner, resulting in richer joint representations that capture the unique characteristics of each modality." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.416, + 0.785, + 0.444 + ], + "angle": 0, + "content": "- We augment two EO benchmarks to create the first datasets with three modalities of different natures (very high resolution images, optical and SAR time series)." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.446, + 0.785, + 0.505 + ], + "angle": 0, + "content": "- We demonstrate that OmniSat can leverage diverse modalities to learn rich representations, establishing new states-of-the-art for tree species, crop type, and land cover classification. Furthermore, our cross modal self-supervised training scheme improves performance even when only one modality is available during inference." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.37, + 0.785, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.367, + 0.54 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.595 + ], + "angle": 0, + "content": "This section provides an overview of self-supervised and multimodal learning, emphasizing the specificities of their usage for Earth observation. Lastly, we highlight the scarcity of multimodal EO datasets with diverse data types." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.787, + 0.677 + ], + "angle": 0, + "content": "Self-Supervised Learning. This technique consists in learning expressive data representations without labels by using a pretext task. This approach has been particularly successful for natural language [52] and image [70] analysis. Initially focused on discriminative tasks [37, 68, 100], recent self-supervised approaches for images can be categorized as contrastive or generative." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.787, + 0.766 + ], + "angle": 0, + "content": "Contrastive methods minimize the distance between representations of paired samples, often the same image under different transformations, and maximize the distance with other samples [15, 17, 46]. More efficient methods only consider positive samples and avoid mode collapse by introducing various asymmetries [18, 42] or normalization [16]. Such approaches have been successfully adapted to EO, for which samples are paired according to their location [89] or time of acquisition [7, 63]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Generative methods reason at the level of individual token—a small portion of the input, typically a patch for images [25]. The objective is to reconstruct the masked tokens of an input image in pixel [10,45,95] or feature space [6]. This principle has been successfully adapted to EO analysis [20, 30, 99], and was further extended to handle multiple spatial scales [74], multimodality [29, 51], or hyperspectral observations [50, 60]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Publicly Available Multimodal EO Datasets. We provide in parenthesis the spatial resolutions of the single-date images and labels, and the temporal resolutions of time series. S1/S2 denotes Sentinel-1 and 2. \\(\\star\\): modalities added in this work." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.196, + 0.782, + 0.498 + ], + "angle": 0, + "content": "
DatasetModalitiesLabels
images (single date)time series
SpaceNet6 [81]SAR+optical (0.5m-2m)Xbuilding footprint (<1m)
TreeSatAI [2]aerial + S1/S2 (0.2-10m)Xforestry (60m)
BigEarthNet [86]S1/S2 (10m)Xland cover (100m)
DFC20 [76]S1/S2 (10m)Xland cover (500m)
MDAS [48]S1/S2 + hyperspectral (2.2-10m)Xland cover (0.25m)
DOFA [96]NAIP + Gaofen + S1/S2 + EnMAP (1-30m)XX
PASTIS-R [33, 34]XS1/S2 (30-140 / year)agriculture (10m)
SSL4EO-S12 [92]XS1/S2 (4 / year)X
DFC21-DSE [61]XS1/S2 + LS8 (3-9/year)human activity (500m)
MapInWild [28]XS1/S2 (4 / years)protected areas (10m)
SEN12MS-CR-TS [27]XS1/S2 (30 / years)cloud cover (10m)
MultiSenGE [93]XS1/S2 (30-140 / years)land cover (10m)
FLAIR [31]aerial (0.2m)S2 (20-114 / year)land cover (0.2m)
Satlas [12]NAIP (0.5 -2m)S2 (8-12 / year)various
PASTIS-HD★ SPOT 6-7 (1.5m)S1/S2 (30-140 / year)agriculture (10m)
TreeSatAI-TSaerial (0.2m)★ S1/S2 (10-70 / year)forestry (60m)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Several hybrid approaches combine the discriminative power of contrastive methods and the scalability of generative objectives for natural images [70, 101] and EO data [29]. Our proposed OmniSat model also implements both mechanisms. A key feature is that we leverage the precise alignment between different sources of EO data to contrastively match small patches of different modalities rather than entire images or time series." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Self-Supervised Multimodal Learning. Multimodal computer vision has received a lot of interest [13], notably due to the success of cross-modal pre-training [72]. Recent models align the embeddings of heterogeneous modalities such as video and sound [49], depth and images [44], text and image [3, 9], or multiple combinations of these modalities [38, 39, 82, 84]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Multimodal learning also has a long history in EO [58, 71, 98] due to the large variety and complementarity of sensors [36, 79]. However, recent transformer-based architectures [90] for EO are often limited to one type of modality, be it a single image [20, 74] or time-series [34, 87]. For example, CROMA [29] and PRESTO [88] are specifically designed for paired optical and radar observations, but cannot handle Very High Resolution (VHR) data. USat [51] considers images with different resolutions, but only takes a single date within a time series. UT&T [31] can natively take single and multi-date observations of different modalities, but cannot be easily pre-trained in a self-supervised manner since it relies on convolutions and an ad-hoc late fusion scheme." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.127 + ], + "angle": 0, + "content": "G.Astruc et al." + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.144, + 0.384, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.246, + 0.376, + 0.255 + ], + "angle": 0, + "content": "Sentinel-2 time series" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.257, + 0.37, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.328, + 0.351, + 0.338 + ], + "angle": 0, + "content": "(a) FLAIR" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.144, + 0.583, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.246, + 0.558, + 0.255 + ], + "angle": 0, + "content": "\\(\\star\\) Sentinel-2 time series" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.256, + 0.582, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.285, + 0.558, + 0.294 + ], + "angle": 0, + "content": "\\(\\star\\) Sentinel-1 time series" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.294, + 0.581, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.327, + 0.543, + 0.338 + ], + "angle": 0, + "content": "(b) TreeSatAI-TS" + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.144, + 0.728, + 0.154 + ], + "angle": 0, + "content": "VHR satellite \\(1.5\\mathrm{m}\\)" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.155, + 0.731, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.245, + 0.746, + 0.255 + ], + "angle": 0, + "content": "PASTIS: Sentinel-2 time series" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.255, + 0.745, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.285, + 0.751, + 0.294 + ], + "angle": 0, + "content": "PASTIS-R: Sentinel-1 time series" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.294, + 0.746, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.327, + 0.712, + 0.338 + ], + "angle": 0, + "content": "(c) PASTIS-HD" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.346, + 0.789, + 0.402 + ], + "angle": 0, + "content": "Fig. 1: Datasets. We represent three tiles from the considered multilabel classification datasets: FLAIR (a), TreeSatAI-TS (b) and PASTIS-HD (c). TreeSatAI-TS is a new dataset built by replacing the single-date Sentinel-1 and 2 images of TreeSatAI [2] by year-long time series. PASTIS-HD (c) adds VHR satellite images to PASTIS-R [34]. \\(\\star\\): modalities added in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.789, + 0.55 + ], + "angle": 0, + "content": "Multimodal EO Datasets. As reported in Table 1, many multimodal EO datasets use Sentinel-1 [11] and 2 [26] data for applications ranging from land cover to forestry analysis and fire detection. We also note that most multimodal datasets only contain data of one type: mono-date image or time series. Several datasets (BigEarthNet [86], DFC20 [76], MDAS [48]) select a single date from time series. However, single Sentinel-1 and 2 acquisitions can be significantly affected by rain and cloud cover, respectively. Furthermore, capturing the temporal dynamics is crucial to characterize the phenology of vegetation [91]," + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.789, + 0.655 + ], + "angle": 0, + "content": "FLAIR [31] is the first multimodal EO dataset to propose both very high spatial resolution \\((\\leq 2\\mathrm{m})\\) and high temporal resolution \\((>4\\) images/year). Satlas [12] combines Sentinel-2 time series and for \\(5\\%\\) to tiles (continental US), very high definition NAIP images. The functional map of the World [19] integrates observations from various sensors, but most areas are only observed with one sensor. Two other datasets contain time series and single images from multiple sources, but were not available at the time of writing this article: IARPA-SMART [40] and DOFA [96]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.789, + 0.759 + ], + "angle": 0, + "content": "To showcase how OmniSat can consume an arbitrary number of modalities with different spatial, spectral, and temporal resolutions, we selected two commonly used EO benchmarks, TreeSatAI [2] and PASTIS-R [34], whose focus on crop type mapping and forestry differs from the land cover analysis of FLAIR. We added new modalities to these datasets to reach three distinct data types: VHR aerial images, optical time series, and SAR time series. See Fig. 1 for an illustration, and Sec. 4.1 for more details on how we extended these datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.781, + 0.317, + 0.796 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We consider a tile \\( x \\) observed through a set \\( \\mathbf{M} \\) of \\( M \\) distinct sensors or modalities. The goal of the OmniSat model is to learn in a self-supervised fashion to combine" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.147, + 0.761, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.392, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Fig. 2: OmniSat Architecture. We illustrate OmniSat for \\( M = 3 \\) modalities, and a tile split into \\( P = 4 \\) patches. The \\( M \\times P \\) input tokens \\( x_{\\mathbf{P}}^{\\mathbf{M}} \\) are encoded by \\( M \\) modality-specific encoders \\( \\mathcal{E}^{\\mathrm{M}} \\), yielding the token representations \\( f_{\\mathbf{P}}^{\\mathbf{M}} \\). The module \\( \\mathcal{C} \\) combines them into multimodal patch representations \\( f_{\\mathbf{P}}^{\\star} \\). The token embeddings \\( f_{\\mathbf{P}}^{\\mathbf{M}} \\) are supervised by a contrastive cross-modal objective. We also use a reconstruction objective: the masked multimodal representations \\( f_{\\mathbf{P}}^{\\star} \\) are decoded by modality-specific networks \\( \\mathcal{D}^{\\mathrm{M}} \\) to reconstruct their corresponding inputs in \\( x_{\\mathbf{P}}^{\\mathbf{M}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.587 + ], + "angle": 0, + "content": "all modalities \\(\\mathbf{M}\\) into a multimodal representation \\(f^{\\star}\\). We first provide details about OmniSat's architecture in Sec. 3.1. We then explain our our training scheme, which consists of a cross-modal contrastive objective (Sec. 3.2) and a multimodal masked encoding task (Sec. 3.3). Finally, we present the implementation details in Sec. 3.4. The overall method is represented in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.347, + 0.633 + ], + "angle": 0, + "content": "3.1 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.787, + 0.684 + ], + "angle": 0, + "content": "This section presents the tokenization process, the structures of the encoder and decoder for each modality, and the architecture of the modality combiner network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Multimodal Tokenization. All available modalities are spatially aligned through georeferencing. This allows us to divide the tile \\( x \\) into a set \\( \\mathbf{P} \\) of \\( P \\) non-overlapping patches consistently across all modalities: \\( x_{p}^{\\mathbf{M}} = \\{x_{p}^{m}\\}_{m\\in \\mathbf{M}} \\) corresponds to \\( M \\) distinct views of the same patch \\( p \\) with different modalities. Each modality \\( m \\) takes its values in a space \\( \\varOmega^m \\) such that \\( x_{p}^{m}\\in \\varOmega^{m} \\). We index tokens with pairs \\( (m,p) \\), defined for each modality \\( m \\) and patch \\( p \\), for a total of \\( M\\times P \\) tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Time series from Sentinel satellites may experience registration errors spanning several meters, complicating their precise alignment with high-resolution imagery. However, using temporal sequences of satellite data mitigates these errors as aggregation over time tends to balance out misalignments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.127 + ], + "angle": 0, + "content": "G.Astruc et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.149, + 0.382, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.31, + 0.387, + 0.322 + ], + "angle": 0, + "content": "(a) Image Encoder and Decoder." + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.148, + 0.548, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.311, + 0.562, + 0.322 + ], + "angle": 0, + "content": "(b) Temporal Encoder and Decoder." + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.146, + 0.78, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.31, + 0.758, + 0.322 + ], + "angle": 0, + "content": "(c) Modality Combining Network." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 3: OmniSat Architecture. OmniSat is composed of dedicated patch encoders for image (a) and time series b, here represented for a length of \\( L = 4 \\) time stamps. The modality combining module \\( \\mathcal{C} \\) is depicted in (c) with \\( P = 2 \\) and \\( M = 3 \\). Elements colored in orange are learned networks or parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.789, + 0.557 + ], + "angle": 0, + "content": "Encoder-Decoder for Images. We split image tiles split into small square patches: \\(\\varOmega^{\\mathrm{img}}=\\mathbb{R}^{C\\times W\\times W}\\) with \\(W\\) the size of the patches in pixels and \\(C\\) the number of channels. As shown in Fig. 3a, we encode these inputs with a sequence of convolutions and max-pool layers until the spatial dimension is fully collapsed. Decoding involves a symmetric sequence of convolutions and un-pooling layers. Contrary to existing masked auto-encoders, we pass the pooling indices from the encoder's max-pooling to the decoder's un-pooling in the manner of SegNet [8]. This dispenses the encoder from learning the intra-patch spatial configuration. This allows the image encoder to focus on the radiometric information, which may be more relevant depending on the application." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.789, + 0.668 + ], + "angle": 0, + "content": "Encoder-Decoder for Time Series. Each temporal patch is represented by \\(L\\) sequential observations with \\(C\\) channels: \\(\\Omega^{\\mathrm{TS}} = \\mathbb{R}^{C\\times L}\\), each associated with a time stamp. We encode the temporal patches using a Lightweight Temporal Attention Encoder (LTAE) model [32], an efficient network for geospatial time series processing. We decode vector representations into time series by repeating the vector \\(L\\) times across the temporal dimension, adding a temporal encoding for each time step, and using an MLP to map the results to size \\(C\\). See Fig. 3b for an illustration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.79, + 0.773 + ], + "angle": 0, + "content": "Optical time series are notoriously affected by clouds [85]. This may affect the validity of the reconstruction task: the decoder cannot know which observations are cloudy, making the reconstruction objective unpredictable. To circumvent this issue, we use the temporal attention maps of the encoder's LTAE to select dates to reconstruct: cloudless observations are more informative and should have a higher attention score [78]. We only consider in the reconstruction loss \\(\\mathcal{L}_{\\mathrm{reconstr}}\\) the top \\(25\\%\\) dates in terms of the LTAE's attention maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Modality Combining Network. The modality combining network \\(\\mathcal{C}\\), represented in Fig. 3c, takes the \\(M \\times P\\) token embeddings \\(f_{\\mathbf{P}}^{\\mathbf{M}}\\), some of whom can potentially be masked. We equip each token with a Euclidean relative positional encoding [94], calculated based on their patch's position \\(\\{r(p,q) \\mid (p,q) \\in \\mathbf{P}^2\\}\\), allowing each token to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "selectively consider its spatial surroundings. As most EO data are captured from above (satellite or aerial), their distribution is invariant by horizontal translation, making this choice of encoding preferable to an absolute position encoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.191, + 0.785, + 0.282 + ], + "angle": 0, + "content": "The modality combining module \\(\\mathcal{C}\\) starts with a series of \\(B\\) residual self-attention blocks connecting all tokens across modality. We then perform cross-attention between the resulting token embeddings \\(g_{\\mathbf{P}}^{\\mathbf{M}} \\in \\mathbb{R}^{d \\times M \\times P}\\) and \\(P\\) copies \\(f_{\\mathbf{P}}^{\\mathrm{comb}}\\) of a modality combining token \\(f^{\\mathrm{comb}} \\in \\mathbb{R}^{d}\\) learned as a free parameter. Each copy of \\(f_{p}^{\\mathrm{comb}}\\) is spatially located at the patch \\(p\\) for the relative positional encoding \\(r\\). The module \\(\\mathcal{C}\\) outputs \\(P\\) multimodal encodings \\(f_{\\mathbf{P}}^{\\star}\\) combining all available modalities for each patch:" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.291, + 0.788, + 0.309 + ], + "angle": 0, + "content": "\\[\ng _ {\\mathbf {P}} ^ {\\mathbf {M}} = \\text {s e l f - a t t e n t i o n} \\left(f _ {\\mathbf {P}} ^ {\\mathbf {M}}; r\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.312, + 0.787, + 0.331 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathbf {P}} ^ {f \\star} = \\text {c r o s s - a t t e n t i o n} \\left(f _ {\\mathbf {P}} ^ {\\text {c o m b}}, g _ {\\mathbf {P}} ^ {\\mathbf {M}}; r\\right). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.349, + 0.41, + 0.364 + ], + "angle": 0, + "content": "3.2 Contrastive Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.446 + ], + "angle": 0, + "content": "We denote by \\( f_{p}^{m} \\) the \\( d \\)-dimensional encodings of the input patch \\( x_{p}^{m} \\) given by their dedicated encoders. We propose to supervise the embeddings \\( f_{p}^{m} \\) with a contrastive objective encouraging spatial consistency across modalities. Indeed, while each modality captures distinct characteristics of \\( p \\), all encodings \\( f_{p}^{m} \\) share the same latent variable: the semantic content of the patch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.786, + 0.581 + ], + "angle": 0, + "content": "In practice, we want \\( f_{p}^{m} \\) to be closer to \\( f_{p}^{n} \\) for \\( n \\neq m \\), than to \\( f_{q}^{n} \\) for other patches \\( q \\neq p \\). We define \\( \\mathbf{B} \\) as the set of patches within the current batch of observations. We adapt the classic InfoNCE loss [69] to our setting with two main differences, illustrated in Fig. 4. (i) Each token \\( (m,p) \\) has \\( M - 1 \\) positive matches: the tokens corresponding to the same patch \\( p \\) but viewed in another modality \\( n \\neq m \\); and (ii) as EO observations are generally spatially regular, nearby patches may be visually indistinguishable. Therefore, we exclude from the negative matches of \\( (m,p) \\) all tokens in modality \\( m \\) and which are too close to \\( p \\). To this end, we remove the set \\( T(m,p) \\) of tokens with modality \\( m \\) and whose patches are in the same tile as \\( p \\). Our loss function \\( \\mathcal{L}_{\\mathrm{contrast}} \\) is defined as such:" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.591, + 0.787, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t r a s t}} = \\frac {1}{M | \\mathbf {B} |} \\sum_ {(m, p) \\in \\mathbf {M} \\times \\mathbf {B}} \\log \\left(\\frac {\\sum_ {n \\neq m} \\exp \\left(\\left\\langle f _ {p} ^ {m} , f _ {p} ^ {n} \\right\\rangle / \\gamma\\right)}{\\sum_ {(n , q) \\in \\mathbf {M} \\times \\mathbf {B} \\backslash T (m , p)} \\exp \\left(\\left\\langle f _ {p} ^ {m} , f _ {q} ^ {n} \\right\\rangle / \\gamma\\right)}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.768 + ], + "angle": 0, + "content": "with \\(\\gamma\\) a temperature parameter, and \\(\\langle \\cdot ,\\cdot \\rangle\\) the scalar product in \\(\\mathbb{R}^d\\). This function, specifically designed for geospatial data, allows us to contrast individual patches across modalities, which is not typically feasible for natural images. However, as the contrastive objective aligns multimodal representations, the patch encoders may be encouraged to overlook the distinct attributes of their respective modality. Instead, they may focus only on features shared by all modalities, i.e., their common denominator. To ensure that encoders also capture modality-specific information, we incorporate a reconstruction objective, detailed in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.521, + 0.803 + ], + "angle": 0, + "content": "3.3 Multimodal Reconstruction Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "During training, we mask a fraction of tokens \\(\\mathbf{K} \\subset \\mathbf{M} \\times \\mathbf{P}\\) and replace their embeddings with a learned vector \\(f^{\\mathrm{mask}} \\in \\mathbb{R}^{d}\\). Note that the masking can differ across modalities," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.127 + ], + "angle": 0, + "content": "G.Astruc et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.476, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.147, + 0.788, + 0.285 + ], + "angle": 0, + "content": "Fig.4:Contrastive Loss. We represent the token matching matrix for two tiles \\(\\mathrm{Tile}_1\\) and \\(\\mathrm{Tile}_2\\) viewed across 3 modalities \\(m_{1}\\) \\(m_2\\) ,and \\(m_3\\) Tile is composed of the patches \\(p_1\\) and \\(p_2\\) ,while \\(\\mathrm{Tile}_2\\) comprises \\(q_{1}\\) and \\(q_{2}\\) .In contrast to classic approaches which ignore the diagonal and assign each sample with a single positive match, our loss defines operates at the patch level, considers multiple positives per token, and excludes tokens in a block-diagonal fashion." + }, + { + "type": "image_footnote", + "bbox": [ + 0.581, + 0.29, + 0.702, + 0.335 + ], + "angle": 0, + "content": "+ positive match\n- negative match\no ignored" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.788, + 0.397 + ], + "angle": 0, + "content": "and some patches may be entirely masked. All tokens are then processed by the modality combining network \\(\\mathcal{C}\\), which outputs \\(P\\) multimodal embeddings \\(f_{\\mathbf{P}}^{\\star}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.406, + 0.788, + 0.427 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathbf {P}} ^ {\\star} = \\mathcal {C} \\left(\\left\\{f _ {p} ^ {m} \\right\\} _ {(m, p) \\notin \\mathbf {K}} \\cup \\left\\{f ^ {\\text {m a s k}} \\right\\} _ {(m, p) \\in \\mathbf {K}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.436, + 0.79, + 0.482 + ], + "angle": 0, + "content": "To encourage the patch embeddings \\( f_{\\mathbf{P}}^{*} \\) to capture information from all modalities, we build a multimodal reconstruction objective. We denote by \\( \\mathcal{D}^m: \\mathbb{R}^d \\mapsto \\Omega^m \\) the dedicated decoder of each modality \\( m \\) and write the reconstruction loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.49, + 0.788, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n s t r}} = \\frac {1}{| \\mathbf {K} |} \\sum_ {(m, p) \\in \\mathbf {K}} \\frac {1}{\\dim (\\Omega^ {m})} \\| \\mathcal {D} ^ {m} \\left(f _ {p} ^ {\\star}\\right) - x _ {p} ^ {m} \\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.54, + 0.782, + 0.558 + ], + "angle": 0, + "content": "with \\(\\dim (\\varOmega^m)\\) the dimension of \\(\\varOmega^m\\). The total loss is the sum of \\(\\mathcal{L}_{\\mathrm{reconstr}}\\) and \\(\\mathcal{L}_{\\mathrm{contrast}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.576, + 0.422, + 0.592 + ], + "angle": 0, + "content": "3.4 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.6, + 0.667, + 0.615 + ], + "angle": 0, + "content": "We detail here the specific parameters chosen in all our experiments." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.622, + 0.789, + 0.744 + ], + "angle": 0, + "content": "Tokenization. We split each tile along a regular spatial grid to produce a set of non-overlapping patches \\(\\mathbf{P}\\) consistent across all modalities. For TreeSat and FLAIR, we use a \\(10 \\times 10\\mathrm{m}\\) grid, meaning that the VHR input tokens are small image patches of size \\(50 \\times 50\\) with \\(0.2\\mathrm{m}\\) per pixel. The patches of Sentinel observations with a resolution of \\(10\\mathrm{m}\\) are single-pixel temporal sequences of spectral measurements. For PASTIS-HD, we use a \\(40 \\times 40\\mathrm{m}\\) grid, meaning that the VHR patches are of size \\(40 \\times 40\\) with \\(1.0\\mathrm{m}\\) per pixel. The patches of Sentinel observations [26] are \\(4 \\times 4\\) image time series which we spatially flatten before encoding." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Hyperparameters. To show the versatility of OmniSat, we use the same configuration throughout all experiments. The embedding size is \\( d = 256 \\), resulting in image encoders and decoders with 3.6M and 1.8M parameters, 403k and 96k for optical time series, and 402k and 95k for radar time series. The modality combiner module is composed of \\( B = 6 \\) residual self-attention blocks and a single cross-attention block, for a total of 3.6M parameters. We train our model on 3 A6000 GPUs with a batch size of 128" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "multimodal tiles per GPU and set the contrastive temperature \\(\\gamma\\) to 0.1. We train our model with the ADAM optimizer [53], with a learning rate of \\(10^{-4}\\) for pretraining and \\(2\\times 10^{-5}\\) for fine-tuning, and a ReduceLROnPlateau scheduler [1] with a patience of 10 epochs and a decay rate of 0.1. When re-implementing competing methods, we use the hyperparameters of their open-source repository." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.245, + 0.358, + 0.263 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.337 + ], + "angle": 0, + "content": "We evaluate OmniSat's performance across three multimodal datasets, including two new datasets introduced in this work, and presented in Sec. 4.1. We outline our experimental protocol and our adaptation of competing methods in Sec. 4.2. We then present our quantitative results and analysis in Sec. 4.3 and conduct an ablation study in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.359, + 0.318, + 0.372 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.787, + 0.413 + ], + "angle": 0, + "content": "We consider three multimodal datasets: FLAIR [31], and the augmented TreeSatAI-TS [2] and PASTIS-HD [33, 34]. See Fig. 1 for an illustration of these two last datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.51 + ], + "angle": 0, + "content": "TreeSatAI-TS: TreeSatAI [2] is a multimodal dataset for tree species identification, containing 50,381 tiles of \\(60 \\times 60\\mathrm{m}\\) with multi-label annotations for 20 classes and all taken in Germany. Each tile is associated with a very high resolution RGB and near-infrared (NIR) image (\\(0.2\\mathrm{m}\\) pixel resolution), a single Sentinel-2 multi-spectral image (\\(10\\mathrm{m}\\) per pixel resolution, 10 bands), and a single Sentinel-1 radar image (\\(10\\mathrm{m}\\) per pixel resolution, 3 bands: two polarization channels and their ratio)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Motivated by the fact that fine-grained vegetation discrimination relies heavily on temporal dynamics information [91], we introduce TreeSatAI- \\(\\mathrm{TS}^1\\). This extended version uses open-source data to add Sentinel-1 and Sentinel-2 time series to each tile, spanning the closest available year to the VHR observation for Sentinel-2. Note that due to the weather patterns and position of the area of interest with respect to Sentinel-2's orbit, the optical time series is particularly irregular and occluded, with up to \\(50\\%\\) of acquisitions being non-exploitable. Despite this challenge, we included the raw observations without pre-processing, whereas TreeSatAI's single-date images have been manually selected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.741 + ], + "angle": 0, + "content": "PASTIS-HD: The PASTIS dataset [33], is designed for semantic and panoptic segmentation of agricultural parcels using Sentinel-2 time series and covers 18 crop types across 2,433 image time series with dimensions of \\(1280 \\times 1280\\mathrm{m}\\). Each series contains between 38 and 61 observations with 10 spectral bands. PASTIS-R [34] adds the corresponding Sentinel-1 radar time series. We only used the ascendant time series of Sentinel-1 for our training and evaluation, for a total of 169,587 radar images with three bands." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.803 + ], + "angle": 0, + "content": "To enhance the spatial resolution and utility of PASTIS, we introduce PASTISHD², which integrates contemporary VHR satellite images (SPOT 6-7 [24]). We apply orthorectification and pansharpening, resample the resulting images to a 1m resolution, and finally convert them to 8 bits. We follow the protocol of Irvin et al. [51] to use the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.811, + 0.744, + 0.825 + ], + "angle": 0, + "content": "1The dataset is available at https://huggingface.co/datasets/IGNF/TreeSatAI-Time-Series." + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.825, + 0.681, + 0.839 + ], + "angle": 0, + "content": "2The dataset is available at https://huggingface.co/datasets/IGNF/PASTIS-HD." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.811, + 0.744, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "G.Astruc et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "dense annotations for a multi-label classification task: each patch is associated with the labels of all of its pixels. This conversion allows us to evaluate all methods in the same setting and configuration as TreeSatAI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.2, + 0.79, + 0.29 + ], + "angle": 0, + "content": "FLAIR. The FLAIR dataset [31] combines VHR aerial images with time series data. It comprises 77,762 aerial tiles (\\(512 \\times 512\\) pixels, \\(0.2\\mathrm{m}\\) resolution) with five channels (RGB, near-infrared, and a normalized digital surface model) taken in France, alongside corresponding Sentinel-2 time series (\\(10\\mathrm{m}\\) resolution, 10 spectral bands, 20 to 114 observations per year). We apply the same processing as PASTIS to use the dense annotation for a multi-label classification task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.315, + 0.406, + 0.33 + ], + "angle": 0, + "content": "4.2 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.789, + 0.357 + ], + "angle": 0, + "content": "This section details our experimental protocol and our adaption of competing algorithms." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.364, + 0.631, + 0.378 + ], + "angle": 0, + "content": "Evaluation Protocol. All experiments follow a similar setting:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.379, + 0.787, + 0.422 + ], + "angle": 0, + "content": "- Pre-training (optional). Methods that support self-supervised pre-training (OmniSat, SatMAE [20], ScaleMAE [74], CROMA [29]) are pre-trained for up to 250 epochs on the entire training set without access to labels." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.424, + 0.604, + 0.437 + ], + "angle": 0, + "content": "- Fine-Tuning. We propose two settings for fine-tuning:" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.379, + 0.787, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.44, + 0.785, + 0.469 + ], + "angle": 0, + "content": "- Fully Supervised Fine-Tuning. We train the resulting models using all the labels in the training set." + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.471, + 0.785, + 0.514 + ], + "angle": 0, + "content": "- Semi-Supervised Fine-Tuning. We use a portion of \\(10\\%\\) or \\(20\\%\\) of the training set, stratified by the distribution of classes, to fine-tune the models. For models without pre-training, this corresponds to supervision in the low-data regime." + }, + { + "type": "list", + "bbox": [ + 0.254, + 0.44, + 0.785, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.517, + 0.787, + 0.547 + ], + "angle": 0, + "content": "- Unimodal and Multimodal Evaluation. We evaluate all methods using each available modality independently and combining all supported modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.789, + 0.647 + ], + "angle": 0, + "content": "Adapting Competing Approaches. We report the performance of several methods taken from the literature on our considered datasets: LightGBM [2], PRESTO [88], and MOSAIKS [77]. However, few existing methods can operate on single- and multi-date data at the same time. To ensure a fair evaluation of competing approaches, we modify various state-of-the-art models to handle a broader combination of modalities. We provide details on these changes in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.671, + 0.517, + 0.685 + ], + "angle": 0, + "content": "4.3 Numerical Experiments and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.787, + 0.728 + ], + "angle": 0, + "content": "In this section, we report our model's performance and efficiency compared to other approaches across the considered datasets and propose our analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.786, + 0.763 + ], + "angle": 0, + "content": "TreeSatAI-TS. Tab. 2 presents the performance of different models on TreeSatAI and TreeSatAI-TS. We report several key observations:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Benefit of Time Series. For the original TreeSatAI dataset with single-date Sentinel1/2 observations, none of the pre-training schemes significantly improve performance beyond simple baselines such as ResNet, PSE, or MLP, even in a semi-supervised setting. In particular, single-date S1 observations yield low performance for all methods (below 20 F1-score), emphasizing the need to use the entire time series." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.242 + ], + "angle": 0, + "content": "Table 2: Performance on TreeSatAI-TS. We report the weighted F1 for multi-label tree species classification on TreeSatAI (TSAI) and our extended TreeSatAI-TS (TSAI-TS) dataset when fine-tuning with \\(10\\%\\) and \\(100\\%\\) of training labels. The first line of the table is the modality used for evaluation. We distinguish methods that are best for one modality within a dataset, best in a dataset across all modalities, and the best overall performance. \\(\\star\\): late feature fusion with a ResNet pre-trained on ImageNet. \\(\\blacksquare\\): Foundation model trained on extensive external data. \\(\\dagger\\): model evaluated on this dataset for the first time." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.251, + 0.785, + 0.626 + ], + "angle": 0, + "content": "
Modelpre-trainingAllSentinel-1Sentinel-2VHR Image
10%100%10%100%10%100%10%100%
Evaluated on TreeSatAI: single date for Sentinel-1 and Sentinel-2
† PSE [35]None47.2*68.1*11.514.648.558.3--
† ViT [25]None42.757.18.717.539.857.336.751.7
MLP [2]None42.6*71.5*3.410.122.152.0--
ResNet [2]ImageNet------58.870.1
LightGBM [2]ImageNet-54.3*-11.9-48.2-44.0
PRESTO [88]---19.8-46.3--
† DOFA [96]59.571.611.619.348.257.051.667.5
MOSAIKS [22,77]TSAI-----56.0--
† CROMA [29]TSAI49.661.010.112.747.855.7--
† SatMAE [20]TSAI46.161.5--40.349.744.161.4
† ScaleMAE [74]TSAI47.662.5--46.755.246.963.6
OmniSat (ours)TSAI56.270.45.36.448.457.152.868.9
Evaluated on TreeSatAI-TS: Sentinel-1 and Sentinel-2 time series spanning one year
† PSE+LTAE [35]None59.4*71.2*42.652.444.057.2--
† UT&T [31]ImageNet43.856.742.355.241.557.044.355.9
† DOFA [96]41.871.30.00.025.039.451.667.5
† Scale-MAE [74]TSAI-TS44.160.4--11.031.546.963.6
OmniSat (ours)None52.273.331.655.933.949.751.471.0
OmniSat (ours)TSAI-TS61.174.248.256.751.462.958.370.5
" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.659, + 0.788, + 0.705 + ], + "angle": 0, + "content": "OmniSat exhibits significantly improved results on TreeSatAI-TS, with or without pretraining. Image models struggles to extract meaningful features temporally aggregated temporal observations, while OmniSat learn rich dynamic features." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.705, + 0.788, + 0.763 + ], + "angle": 0, + "content": "The foundation model DOFA [96], with 111M parameters and a large closed-source training set, outperforms all models when evaluated on single-date modalities. However, OmniSat reaches higher performances on TreeSatAI-TS with only 10 million parameters, which we attribute to its ability to leverage temporal modalities." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "- Benefits of Multimodality. When using all modalities, OmniSat outperforms all competing methods by a margin of \\(3\\%\\) F1-score. The multimodal performance of OmniSat and CROMA, which learn to combine data sources, is strictly superior to the F1-score of their best modality by \\(3.7\\%\\) and \\(5.3\\%\\) points, respectively. Conversely, the performance of methods that rely on late-fusion (SatMAE, ScaleMAE," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "G.Astruc et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 3: Performance on PASTIS-HD. We report the macro-averaged F1-score for crop-type multi-class classification on the PASTIS-HD dataset. We distinguish methods that are best for one modality, best in a dataset across all modalities. \\(\\star\\): late feature fusion with a ResNet. \\(\\dagger\\): model evaluated on this dataset for the first time." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.21, + 0.784, + 0.386 + ], + "angle": 0, + "content": "
Modelpre-trainedAllSentinel-1Sentinel-2VHR image
20%100%20%100%20%100%20%100%
† UTAE [33, 34]None36.8*46.9*20.140.732.737.6--
† ResNet50 [47]ImageNet------57.659.3
† UT&T [31]ImageNet54.253.558.862.854.961.351.149.8
† DOFA [96]53.755.736.741.550.853.447.954.8
† Scale-MAE [74]PASTIS-HD42.042.2--41.246.148.851.9
† CROMA [29]PASTIS-HD57.560.155.356.153.056.7--
OmniSat (ours)No42.059.158.260.251.760.147.352.8
OmniSat (ours)PASTIS-HD62.669.960.869.061.870.854.659.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.404, + 0.79, + 0.446 + ], + "angle": 0, + "content": "Table 4: Performance on FLAIR. We report the macro-averaged F1-score for land cover multiclass classification on the FLAIR dataset. We distinguish methods that are best for one modality and best in a dataset. \\(\\dagger\\): model evaluated on this dataset for the first time." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.455, + 0.726, + 0.589 + ], + "angle": 0, + "content": "
Modelpre-trainedAllSentinel-2VHR Image
10%100%10%100%10%100%
† UT&T [31]ImageNet44.248.857.462.058.965.5
† DOFA [96]70.674.957.061.066.872.1
† ScaleMAE [74]FLAIR63.170.052.561.061.267.3
OmniSat (ours)No62.570.056.165.464.771.5
OmniSat (ours)FLAIR60.673.456.865.465.271.6
" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.621, + 0.787, + 0.65 + ], + "angle": 0, + "content": "ViT) is comparable to their best modality. This demonstrates the value of learning to combine information from different sources end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.651, + 0.789, + 0.773 + ], + "angle": 0, + "content": "- Benefits of Cross-Modal Pre-Training. With access to all modalities, our self-supervised pre-training improves by \\(0.9\\%\\) point the F1-score of the model fine-tuned on \\(100\\%\\) of labels, compared to not pre-training, and \\(8.9\\%\\) when using only \\(10\\%\\) of labels. This shows that our pre-training leads to more expressive multimodal features. Interestingly, when performing inference with Sentinel-2 time series alone, the performance increase linked to the pre-training becomes \\(13.2\\%\\) with \\(100\\%\\) labels and \\(17.5\\%\\) with \\(10\\%\\). This illustrates that our self-supervised pre-training scheme improves the features learned by each encoder despite not relying on annotated data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Experiments on PASTIS-HD. The analysis of the performance of various models on PASTIS-HD is reported in Tab. 3, and is consistent with the ones of TreeSatAI-TS. First, by learning to combine all modalities despite their different resolutions, OmniSat achieves state-of-the-art results on this benchmark. Second, our cross-modal pretraining" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.145, + 0.782, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Fig. 5: Efficiency. We report the best performance of different models between TreeSatAI and TreeSatAI-TS, with pre-training and fine-tuning using \\(100\\%\\) of labels. The area of the markers is proportional to the training time, broken down in pre-training and fine-tuning when applicable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.448 + ], + "angle": 0, + "content": "significantly improves OmniSat's performance in the multimodal \\((+10.8\\mathrm{pF1}\\) -score with \\(100\\%\\) of training label) and all single-modality settings (8.8 points for Sentinel-1, 10.7 for Sentinel-2, and 6.5 for the VHR images)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.789, + 0.577 + ], + "angle": 0, + "content": "Experiments on FLAIR. We report in Tab. 4 the results on the bimodal FLAIR dataset for multilabel classification. OmniSat outperforms the much larger ScaleMAE [74] and UT&T [31] models with \\(100\\%\\) of labels and both modalities by \\(3.4\\%\\). Our pre-training scheme had a smaller impact than for the TreeSatAI-TS experiment. We attribute this to the fact that only two modalities are available, which decreases the supervisory power of our cross-modal contrastive objective and our multimodal reconstruction loss. This highlights a limitation of OmniSat: the model needs to be pre-trained on a modality-rich dataset to achieve its best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.675 + ], + "angle": 0, + "content": "Efficiency Evaluation. We plot in Fig. 5 the best performance between TreeSatAI and TreeSatAI-TS for different models according to their size and training time. OmniSat is more compact, faster to train, and performs better than all evaluated models, including the DOFA foundation model. The highly-specialized combination of PSE, LTAE, and ResNet is a strong contender, outperforming significantly larger models with generic encoding-decoding schemes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.365, + 0.707 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.787, + 0.743 + ], + "angle": 0, + "content": "In this section, we report the results of several experiments evaluating the impact and validity of our main design choices, see Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "a) Encoder/Decoder Architecture. We propose several improvements to the standard image encoder-decoder scheme used in computer vision to accommodate the specificities of EO data. In particular, passing the max-pool indices from the image patch encoder to its decoder allows the learned representation to focus on characterizing the spectral signature instead of fine-grained spatial information, and leads to a performance increase of \\(0.7\\%\\) in the full supervision setting." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "G.Astruc et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 5: Ablation Study. We present the impact of several design choices on the TreeSatAI-TS dataset, measured in terms of macro-averaged F1-score." + }, + { + "type": "table", + "bbox": [ + 0.284, + 0.182, + 0.716, + 0.255 + ], + "angle": 0, + "content": "
Experiment10%100%Experiment10%100%
OmniSat61.174.2b) no contrastive loss55.673.4
a) no index bypass57.573.5b) naive contrastive loss57.872.2
a) no date filtering58.271.6b) no reconstruction loss59.072.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.787, + 0.345 + ], + "angle": 0, + "content": "As clouds frequently obstruct optical time series, we use a unsupervised date-filtering scheme to reconstruct only meaningful acquisitions. This approach leads to a significant improvement of \\(3.6\\%\\), showcasing the benefit of developing modality-aware approaches for EO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.787, + 0.457 + ], + "angle": 0, + "content": "b) Role of Loss Functions. When training without contrastive loss, we observe a decrease in performance of \\(0.8\\%\\) in the fully supervised regime, and a more pronounced drop of \\(5.5\\%\\) in the semi-supervised regime. This demonstrates how learning consistent encoding across encoders facilitates their subsequent fusion. Interestingly, when implementing a naive contrastive loss that considers all negative examples from the batch, the decrease is greater than simply removing this loss (\\(2\\%\\) in full supervision). This strategy may introduce indistinguishable negative examples and perturb the learning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.502 + ], + "angle": 0, + "content": "We also remove the reconstruction loss, meaning that only the encoders are learned contrastively during pre-training. This results in a drop of \\(2\\%\\) F1-score point, illustrating the importance of pre-training the transformer \\(\\mathcal{C}\\) alongside its encoders." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Limitations. All datasets used in our experiments are based in Europe, primarily due to the availability of open-access annotations. This regional focus prevents us from evaluating our model's performance in tropical and developing countries, which present unique challenges in terms of label provision, heterogeneity, and complex classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.598 + ], + "angle": 0, + "content": "A limitation of our pre-training scheme is its dependence on a sufficient number of aligned modalities, as illustrated by its moderate impact on the bimodal FLAIR dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.345, + 0.64 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.826 + ], + "angle": 0, + "content": "We introduced OmniSat, a new architecture for the self-supervised modality fusion of Earth Observation (EO) data from multiple sources. To facilitate its evaluation, we augmented two existing datasets with new modalities of different natures and resolutions. We experimentally showed that leveraging diverse modalities with a flexible model improves the model's performance in both fully and semi-supervised settings. Moreover, our training scheme can exploit the spatial alignment of multiple modalities to improve our model's unimodal performance. Finally, we proposed several improvements to leverage the unique structure of EO data in the architecture of our model, such as automatic date filtering for reconstructing time series. We hope that our promising results and new datasets will encourage the computer vision community to consider EO data as a playing field for evaluating and developing novel self-supervised multimodal algorithms." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This work was supported by ANR project READY3D ANR-19-CE23-0007, and was granted access to the HPC resources of IDRIS under the allocations AD011014719 and AD011014286R1 made by GENCI. We thank Anatol Garioud and Sébastien Giordano for their help on the creation of TreeSatAI-TS and PASTIS-HD datasets. The SPOT images are opendata thanks to the Dataterra Dinamis initiative in the case of the \"Couverture France DINAMIS\" program. We thank Jordi Inglada for inspiring discussions and valuable feedback." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.306, + 0.312, + 0.322 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.337, + 0.79, + 0.378 + ], + "angle": 0, + "content": "1. PyTorch: ReduceLROOnPlateau.org/docs/stable/generated/torch.optim.lr_scheduler.ReduceLROOnPlateau.html#torch.optim.lr_scheduler.ReduceLROOnPlateau, accessed: 2024-02-29 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.379, + 0.79, + 0.435 + ], + "angle": 0, + "content": "2. Ahlswede, S., Schulz, C., Gava, C., Helber, P., Bischke, B., Förster, M., Arias, F., Hees, J., Demir, B., Kleinschmit, B.: TreeSatAI Benchmark Archive: A multi-sensor, multi-label dataset for tree species classification in remote sensing. Earth System Science Data Discussions (2022) 2, 3, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.435, + 0.788, + 0.505 + ], + "angle": 0, + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: A visual language model for few-shot learning. In: NeurIPS (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.505, + 0.788, + 0.547 + ], + "angle": 0, + "content": "4. Amitrano, D., Di Martino, G., Guida, R., Iervolino, P., Iodice, A., Papa, M.N., Riccio, D., Ruello, G.: Earth environmental monitoring using multi-temporal synthetic aperture radar: A critical review of selected applications. Remote Sensing (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.547, + 0.788, + 0.575 + ], + "angle": 0, + "content": "5. Anderson, K., Ryan, B., Sonntag, W., Kavvada, A., Friedl, L.: Earth observation in service of the 2030 agenda for sustainable development. Geo-spatial Information Science (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.575, + 0.788, + 0.616 + ], + "angle": 0, + "content": "6. Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding predictive architecture. In: CVPR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.616, + 0.788, + 0.644 + ], + "angle": 0, + "content": "7. Ayush, K., Uzkent, B., Meng, C., Tanmay, K., Burke, M., Lobell, D., Ermon, S.: Geography-aware self-supervised learning. In: ICCV (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.644, + 0.788, + 0.672 + ], + "angle": 0, + "content": "8. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE TPAMI (2017) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.231, + 0.672, + 0.788, + 0.701 + ], + "angle": 0, + "content": "9. Baevski, A., Hsu, W.N., Xu, Q., Babu, A., Gu, J., Auli, M.: Data2vec: A general framework for self-supervised learning in speech, vision and language. In: ICML (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.788, + 0.728 + ], + "angle": 0, + "content": "10. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.788, + 0.771 + ], + "angle": 0, + "content": "11. Bao, X., Zhang, R., Lv, J., Wu, R., Zhang, H., Chen, J., Zhang, B., Ouyang, X., Liu, G.: Vegetation descriptors from Sentinel-1 SAR data for crop growth monitoring. ISPRS Journal of Photogrammetry and Remote Sensing (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.788, + 0.799 + ], + "angle": 0, + "content": "12. Bastani, F., Wolters, P., Gupta, R., Ferdinando, J., Kembhavi, A.: SatlasPretrain: A large-scale dataset for remote sensing image understanding. In: ICCV (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "13. Bayoudh, K., Knani, R., Hamdaoui, F., Mtibaa, A.: A survey on deep multimodal learning for computer vision: Advances, trends, applications, and datasets. The Visual Computer (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.337, + 0.79, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.219, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.128 + ], + "angle": 0, + "content": "G. Astruc et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Benedetti, P., Ienco, D., Gaetano, R., Ose, K., Pensa, R.G., Dupuy, S.: \\(\\mathbf{M}^3\\)-fusion: A deep learning architecture for multiscale multimodal multitemporal satellite data fusion. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "15. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.245, + 0.787, + 0.272 + ], + "angle": 0, + "content": "17. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.772, + 0.286 + ], + "angle": 0, + "content": "18. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.787, + 0.313 + ], + "angle": 0, + "content": "19. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "20. Cong, Y., Khanna, S., Meng, C., Liu, P., Rozi, E., He, Y., Burke, M., Lobell, D., Ermon, S.: SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In: NeurIPS (2022) 2, 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.787, + 0.395 + ], + "angle": 0, + "content": "21. Coppin, P., Lambin, E., Jonckheere, I., Muys, B.: Digital change detection methods in natural ecosystem monitoring: A review. Analysis of multi-temporal remote sensing images (2002) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "22. Corley, I., Robinson, C., Dodhia, R., Ferres, J.M.L., Najafirad, P.: Revisiting pre-trained remote sensing model benchmarks: Resizing and normalization matters. arXiv preprint arXiv:2305.13456 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "23. Dai, A., Nießner, M.: 3DMV: Joint 3D-multi-view prediction for 3D semantic scene segmentation. In: ECCV (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.787, + 0.493 + ], + "angle": 0, + "content": "24. DataTerra Dinamis: Diffusion OpenData Dinamis, https://dinamis.data-terra.org/opendata/, accessed: 2023-12-15 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.787, + 0.59 + ], + "angle": 0, + "content": "26. Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B., Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F., Bargellini, P.: Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment (2012) 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.787, + 0.618 + ], + "angle": 0, + "content": "27. Ebel, P., Xu, Y., Schmitt, M., Zhu, X.X.: SEN12MS-CR-TS: A remote-sensing data set for multimodal multitemporal cloud removal. IEEE TGRS (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "28. Ekim, B., Stomberg, T.T., Roscher, R., Schmitt, M.: MapInWild: A remote sensing dataset to address the question of what makes nature wild. IEEE Geoscience and Remote Sensing Magazine (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.687 + ], + "angle": 0, + "content": "29. Fuller, A., Millard, K., Green, J.R.: CROMA: Remote sensing representations with contrastive radar-optical masked autoencoders. In: NeurIPS (2023) 1, 2, 3, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.715 + ], + "angle": 0, + "content": "30. Gao, Y., Sun, X., Liu, C.: A general self-supervised framework for remote sensing image classification. Remote Sensing (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.787, + 0.769 + ], + "angle": 0, + "content": "31. Garioud, A., Gonthier, N., Landrieu, L., De Wit, A., Valette, M., Poupee, M., Giordano, S., Wattrelos, B.: FLAIR: A country-scale land cover semantic segmentation dataset from multi-source optical imagery. In: NeurIPS Dataset and Benchmark (2023) 3, 4, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "32. Garnot, V.S.F., Landrieu, L.: Lightweight temporal self-attention for classifying satellite images time series. In: Advanced Analytics and Learning on Temporal Data: ECML PKDD Workshop (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Garnot, V.S.F., Landrieu, L.: Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In: ICCV (2021) 2, 3, 9, 12" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.786, + 0.188 + ], + "angle": 0, + "content": "34. Garnot, V.S.F., Landrieu, L., Chehata, N.: Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing (2022) 2, 3, 4, 9, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "35. Garnot, V.S.F., Landrieu, L., Giordano, S., Chehata, N.: Satellite image time series classification with pixel-set encoders and temporal self-attention. In: CVPR (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "36. Ghamisi, P., Rasti, B., Yokoya, N., Wang, Q., Hofle, B., Bruzzone, L., Bovolo, F., Chi, M., Anders, K., Gloaguen, R., et al.: Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine (2019) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "37. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.3, + 0.787, + 0.327 + ], + "angle": 0, + "content": "38. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023) 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.328, + 0.787, + 0.355 + ], + "angle": 0, + "content": "39. Girdhar, R., Singh, M., Ravi, N., van der Maaten, L., Joulin, A., Misra, I.: Omnivore: A single model for many visual modalities. In: CVPR (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.356, + 0.787, + 0.409 + ], + "angle": 0, + "content": "40. Goldberg, H.R., Ratto, C.R., Banerjee, A., Kelbaugh, M.T., Giglio, M., Vermote, E.F.: Automated global-scale detection and characterization of anthropogenic activity using multi-source satellite-based remote sensing imagery. In: Geospatial Informatics XIII. SPIE (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.787, + 0.438 + ], + "angle": 0, + "content": "41. Greenwell, C., Crall, J., Purri, M., Dana, K., Jacobs, N., Hadzic, A., Workman, S., Leotta, M.: WATCH: Wide-area terrestrial change hypercube. In: WACV (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "42. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. In: NeurIPS (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.787, + 0.519 + ], + "angle": 0, + "content": "43. Hackstein, J., Sumbul, G., Clasen, K.N., Demir, B.: Exploring masked autoencoders for sensor-agnostic image retrieval in remote sensing. arXiv preprint arXiv:2401.07782 (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.787, + 0.55 + ], + "angle": 0, + "content": "44. Hazirbas, C., Ma, L., Domokos, C., Cremers, D.: FuseNet: Incorporating depth into semantic segmentation via fusion-based CNN architecture. In: ACCV (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "45. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.604 + ], + "angle": 0, + "content": "46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.787, + 0.632 + ], + "angle": 0, + "content": "47. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.787, + 0.673 + ], + "angle": 0, + "content": "48. Hu, J., Liu, R., Hong, D., Camero, A., Yao, J., Schneider, M., Kurz, F., Segl, K., Zhu, X.X.: MDAS: A new multimodal benchmark dataset for remote sensing. Earth System Science Data Discussions (2022) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.787, + 0.702 + ], + "angle": 0, + "content": "49. Huang, P.Y., Sharma, V., Xu, H., Ryali, C., Fan, H., Li, Y., Li, S.W., Ghosh, G., Malik, J., Feichtenhofer, C.: MAViL: Masked audio-video learners. In: NeurIPS (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "50. Ibanez, D., Fernandez-Beltran, R., Pla, F., Yokoya, N.: Masked auto-encoding spectral-spatial transformer for hyperspectral image classification. IEEE TGRS (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.787, + 0.77 + ], + "angle": 0, + "content": "51. Irvin, J., Tao, L., Zhou, J., Ma, Y., Nashold, L., Liu, B., Ng, A.Y.: USat: A unified self-supervised encoder for multi-sensor satellite imagery. arXiv preprint arXiv:2312.02199 (2023) 1, 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "52. Kenton, J.D.M.W.C., Toutanova, L.K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.749, + 0.812 + ], + "angle": 0, + "content": "53. Kingma, D.P., Ba, J.: ADAM: A method for stochastic optimization. ICLR (2015) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "54. Krispel, G., Opitz, M., Waltner, G., Possegger, H., Bischof, H.: Fuseseg: LiDAR point cloud segmentation fusing multi-modal data. In: WACV (2020) 2" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.219, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.128 + ], + "angle": 0, + "content": "G.Astruc et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "55. Kuffer, M., Thomson, D.R., Boo, G., Mahabir, R., Grippa, T., Vanhuysse, S., Engstrom, R., Ndugwa, R., Makau, J., Darin, E., et al.: The role of Earth observation in an integrated deprived area mapping \"system\" for low-to-middle income countries. Remote sensing (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.204, + 0.788, + 0.258 + ], + "angle": 0, + "content": "56. Lacoste, A., Sherwin, E.D., Kerner, H., Alemohammad, H., Lütjens, B., Irvin, J., Dao, D., Chang, A., Gunturkun, M., Drouin, A., et al.: Toward foundation models for Earth monitoring: Proposal for a climate change benchmark. arXiv preprint arXiv:2112.00570 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.259, + 0.788, + 0.285 + ], + "angle": 0, + "content": "57. Li, D., Tong, Q., Li, R., Gong, J., Zhang, L.: Current issues in high-resolution Earth observation technology. Science China Earth sciences (2012) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.287, + 0.788, + 0.327 + ], + "angle": 0, + "content": "58. Li, J., Hong, D., Gao, L., Yao, J., Zheng, K., Zhang, B., Chanussot, J.: Deep learning in multimodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation 112, 102926 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.328, + 0.788, + 0.355 + ], + "angle": 0, + "content": "59. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE TPAMI (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.356, + 0.788, + 0.395 + ], + "angle": 0, + "content": "60. Liu, Y., Li, X., Hua, Z., Xia, C., Zhao, L.: A band selection method with masked convolutional autoencoder for hyperspectral image. IEEE Geoscience and Remote Sensing Letters (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.398, + 0.788, + 0.452 + ], + "angle": 0, + "content": "61. Ma, Y., Li, Y., Feng, K., Xia, Y., Huang, Q., Zhang, H., Prieur, C., Licciardi, G., Malha, H., Chanussot, J., et al.: The outcome of the 2021 IEEE GRSS data fusion contest-Track DSE: Detection of settlements without electricity. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.453, + 0.788, + 0.493 + ], + "angle": 0, + "content": "62. Mai, G., Huang, W., Sun, J., Song, S., Mishra, D., Liu, N., Gao, S., Liu, T., Cong, G., Hu, Y., et al.: On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.494, + 0.788, + 0.522 + ], + "angle": 0, + "content": "63. Manas, O., Lacoste, A., Giró-i Nieto, X., Vazquez, D., Rodriguez, P.: Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In: ICCV (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "64. Manfreda, S., McCabe, M.F., Miller, P.E., Lucas, R., Pajuelo Madrigal, V., Mallinis, G., Ben Dor, E., Helman, D., Estes, L., Ciraolo, G., et al.: On the use of unmanned aerial systems for environmental monitoring. Remote sensing p. 641 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.564, + 0.788, + 0.603 + ], + "angle": 0, + "content": "65. Moreira, A., Prats-Iraola, P., Younis, M., Krieger, G., Hajnsek, I., Papathanassiou, K.P.: A tutorial on synthetic aperture radar. IEEE Geoscience and Remote Sensing Magazine (2013) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.605, + 0.788, + 0.645 + ], + "angle": 0, + "content": "66. Nakalembe, C.: Urgent and critical need for sub-saharan african countries to invest in Earth observation-based agricultural early warning and monitoring systems. Environmental Research Letters (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.647, + 0.788, + 0.673 + ], + "angle": 0, + "content": "67. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from RGBD images. In: ECCV (2012) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.675, + 0.788, + 0.702 + ], + "angle": 0, + "content": "68. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.702, + 0.788, + 0.729 + ], + "angle": 0, + "content": "69. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.73, + 0.788, + 0.77 + ], + "angle": 0, + "content": "70. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. TLMR (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.772, + 0.788, + 0.799 + ], + "angle": 0, + "content": "71. Pohl, C., Van Genderen, J.L.: Multisensor image fusion in remote sensing: Concepts, methods and applications. International journal of remote sensing (1998) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "72. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Token Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.148, + 0.788, + 0.189 + ], + "angle": 0, + "content": "73. Recasens, A., Lin, J., Carreira, J., Jaegle, D., Wang, L., Alayrac, J.b., Luc, P., Miech, A., Smaira, L., Hemsley, R., et al.: Zorro: The masked multimodal transformer. arXiv preprint arXiv:2301.09595 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.788, + 0.232 + ], + "angle": 0, + "content": "74. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: ICCV (2023) 2, 3, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.788, + 0.26 + ], + "angle": 0, + "content": "75. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3D semantic segmentation. In: CVPR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.788, + 0.316 + ], + "angle": 0, + "content": "76. Robinson, C., Malkin, K., Jojic, N., Chen, H., Qin, R., Xiao, C., Schmitt, M., Ghamisi, P., Hänsch, R., Yokoya, N.: Global land-cover mapping with weak supervision: Outcome of the 2020 IEEE GRSS data fusion contest. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.788, + 0.358 + ], + "angle": 0, + "content": "77. Rolf, E., Proctor, J., Carleton, T., Bolliger, I., Shankar, V., Ishihara, M., Recht, B., Hsiang, S.: A generalizable and accessible approach to machine learning with global satellite imagery. Nature communications (2021) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.788, + 0.387 + ], + "angle": 0, + "content": "78. Rußwurm, M., Körner, M.: Self-attention for raw optical satellite time series classification. ISPRS Journal of Photogrammetry and Remote Sensing (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.788, + 0.416 + ], + "angle": 0, + "content": "79. Schmitt, M., Zhu, X.X.: Data fusion and remote sensing: An ever-growing relationship. IEEE Geoscience and Remote Sensing Magazine (2016) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.788, + 0.458 + ], + "angle": 0, + "content": "80. Secades, C., O'Connor, B., Brown, C., Walpole, M., et al.: Earth observation for biodiversity monitoring: A review of current approaches and future opportunities for tracking progress towards the aichi biodiversity targets. CBD technical series (2014) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.788, + 0.5 + ], + "angle": 0, + "content": "81. Shermeyer, J., Hogan, D., Brown, J., Van Etten, A., Weir, N., Pacifici, F., Hansch, R., Bastidas, A., Soenen, S., Bacastow, T., et al.: Spacenet 6: Multi-sensor all weather mapping dataset. In: CVPR Workshop EarthVision (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.788, + 0.529 + ], + "angle": 0, + "content": "82. Shukor, M., Dancette, C., Rame, A., Cord, M.: UnIVAL: Unified model for image, video, audio and language tasks. TMLR (2023) 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.788, + 0.572 + ], + "angle": 0, + "content": "83. Skidmore, A.K., Coops, N.C., Neinavaz, E., Ali, A., Schaepman, M.E., Paganini, M., Kissling, W.D., Vihervaara, P., Darvishzadeh, R., Feilhauer, H., et al.: Priority list of biodiversity metrics to observe from space. Nature Ecology & Evolution (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.573, + 0.788, + 0.599 + ], + "angle": 0, + "content": "84. Srivastava, S., Sharma, G.: OmniVec: Learning robust representations with cross modal sharing. In: WACV (2024) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.788, + 0.642 + ], + "angle": 0, + "content": "85. Sudmanns, M., Tiede, D., Augustin, H., Lang, S.: Assessing global Sentinel-2 coverage dynamics and data availability for operational Earth observation (EO) applications using the EO-Compass. International Journal of Digital Earth (2019) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.788, + 0.698 + ], + "angle": 0, + "content": "86. Sumbul, G., De Wall, A., Kreuziger, T., Marcelino, F., Costa, H., Benevides, P., Caetano, M., Demir, B., Markl, V.: BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.788, + 0.726 + ], + "angle": 0, + "content": "87. Tarasiou, M., Chavez, E., Zafeiriou, S.: ViTs for SITS: Vision transformers for satellite image time series. In: CVPR (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.788, + 0.768 + ], + "angle": 0, + "content": "88. Tseng, G., Zvonkov, I., Purohit, M., Rolnick, D., Kerner, H.: Lightweight, pre-trained transformers for remote sensing timeseries. arXiv preprint arXiv:2304.14065 (2023) 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.77, + 0.788, + 0.811 + ], + "angle": 0, + "content": "89. Tseng, W.H., Lé, H.A., Boulch, A., Lefèvre, S., Tiede, D.: CROCO: Cross-modal contrastive learning for localization of Earth observation data. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "90. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.148, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.128 + ], + "angle": 0, + "content": "G. Astruc et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "91. Vrieling, A., Meroni, M., Darvishzadeh, R., Skidmore, A.K., Wang, T., Zurita-Milla, R., Oosterbeek, K., O'Connor, B., Paganini, M.: Vegetation phenology from Sentinel-2 and field cameras for a Dutch barrier island. Remote sensing of environment (2018) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "92. Wang, Y., Graham, N.A.A., Xiong, Z., Liu, C., Albrecht, C.M., Zhu, X.X.: SSL4EO-S12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in Earth observation. IEEE Geoscience and Remote Sensing Magazine (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.232, + 0.788, + 0.285 + ], + "angle": 0, + "content": "93. Wenger, R., Puisant, A., Weber, J., Idoumghar, L., Forestier, G.: MultiSenGE: A multimodal and multitemporal benchmark dataset for land use/land cover remote sensing applications. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "94. Wu, K., Peng, H., Chen, M., Fu, J., Chao, H.: Rethinking and improving relative position encoding for vision transformer. In: ICCV (2021) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "95. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMim: A simple framework for masked image modeling. In: CVPR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "96. Xiong, Z., Wang, Y., Zhang, F., Stewart, A.J., Hanna, J., Borth, D., Papoutsis, I., Saux, B.L., Camps-Valls, G., Zhu, X.X.: Neural plasticity-inspired foundation model for observing the Earth crossing modalities. arXiv preprint arXiv:2403.15356 (2024) 3, 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "97. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change (2013) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "98. Yang, M.Y., Landrieu, L., Tuia, D., Toth, C.: Multi-modal learning in photogrammetry and remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.787, + 0.481 + ], + "angle": 0, + "content": "99. Yuan, Y., Lin, L., Liu, Q., Hang, R., Zhou, Z.G.: SITS-Former: A pre-trained spatio-spectral-temporal representation model for sentinel-2 time series classification. International Journal of Applied Earth Observation and Geoinformation (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.482, + 0.733, + 0.494 + ], + "angle": 0, + "content": "100. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "101. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: Image BERT pre-training with online tokenizer. In: ICLR (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "102. Zong, Y., Mac Aodha, O., Hospedales, T.: Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008 (2023) 1" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.147, + 0.788, + 0.55 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_origin.pdf b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46267b0c0d90a167341f7dcbcf6c5215d133edf5 --- /dev/null +++ b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/cd5afd38-8b9c-4bb7-8670-4848d17215ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42aae080ecb0f4529d3733d0b2eed5bde8f349ac7291e2d4d21ee9235e188eba +size 2275571 diff --git a/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/full.md b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7578ab59ab2972ce353c2dd682bc2f99d168726a --- /dev/null +++ b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/full.md @@ -0,0 +1,390 @@ +# OmniSat: Self-Supervised Modality Fusion for Earth Observation + +Guillaume Astruc $^{1,3,4}$ , Nicolas Gonthier $^{1,2}$ , Clement Mallet $^{1}$ , and Loic Landrieu $^{1,3}$ + +1 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France +2 IGN, France +3 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France +4 CNES, France + +Abstract. The diversity and complementarity of sensors available for Earth Observations (EO) calls for developing bespoke self-supervised multimodal learning approaches. However, current multimodal EO datasets and models typically focus on a single data type, either mono-date images or time series, which limits their impact. To address this issue, we introduce OmniSat, a novel architecture able to merge diverse EO modalities into expressive features without labels by exploiting their alignment. To demonstrate the advantages of our approach, we create two new multimodal datasets by augmenting existing ones with new modalities. As demonstrated for three downstream tasks—forestry, land cover classification, and crop mapping—OmniSat can learn rich representations without supervision, leading to state-of-the-art performances in semi- and fully supervised settings. Furthermore, our multimodal pretraining scheme improves performance even when only one modality is available for inference. The code and dataset are available at https://github.com/gastruc/OmniSat. + +Keywords: Earth observation $\cdot$ Multi-modality $\cdot$ Self-supervised learning + +# 1 Introduction + +Self-supervised multimodal learning has recently gathered significant interest within computer vision [38, 82, 102]. Earth Observation (EO) is particularly well-suited for developing and evaluating such approaches [29, 51], thanks to the large amount of open-access data captured by sensing technologies with complementary capabilities [36, 79]. Combining different sources of EO observations is crucial for several high-impact applications, including environmental [21, 80, 83] and climate monitoring [56, 97], as well as improving food security [66]. Moreover, learning with few or no labels is essential for developing regions with limited data annotation capabilities [5, 55, 62]. + +Despite this potential, most multimodal EO datasets and models focus on a single data type, either mono-date images or time series. This limitation prevents them from simultaneously leveraging the spatial resolution of aerial images [57, 64], the temporal and spectral resolutions of optical satellite time series [26], and the resilience of radar to weather effects [4, 65]. Additionally, existing approaches are often limited for a given set of sensors, limiting their applicability. + +To address these challenges, we introduce OmniSat, a novel architecture designed for the self-supervised fusion of diverse EO data. Existing multimodal approaches often map + +multiple unrelated observations from different modalities to one pivot modality [38, 82] or a shared latent space [39, 84]. In contrast, OmniSat merges multiple views of the same area from different modalities into a single representation combining the specific information of each modality [14, 41, 73]. + +In computer vision, obtaining finely aligned multimodal observations generally requires specialized sensors [54, 59, 67] or the computation of complex mappings between modalities [23, 75]. On the other hand, EO data can be naturally aligned with georeferencing. To leverage this property, we adapt multimodal contrastive learning [49, 72] and cross-modal masked auto-encoding techniques [43] to learn rich multimodal EO representations with a generalist fusion scheme and without annotations. + +To address the scarcity of EO datasets with a diverse range of heterogeneous modalities (see Tab. 1), we enrich the TreeSatAI [2] and PASTIS-R [33, 34] datasets with new aligned modalities. This allows us to evaluate OmniSat's ability to handle an arbitrary number of inputs with varying natures and resolutions. Our contributions can be summarized as follows: + +- We introduce OmniSat, a new model that learns to combine varied sources of EO observations in a self-supervised manner, resulting in richer joint representations that capture the unique characteristics of each modality. +- We augment two EO benchmarks to create the first datasets with three modalities of different natures (very high resolution images, optical and SAR time series). +- We demonstrate that OmniSat can leverage diverse modalities to learn rich representations, establishing new states-of-the-art for tree species, crop type, and land cover classification. Furthermore, our cross modal self-supervised training scheme improves performance even when only one modality is available during inference. + +# 2 Related Work + +This section provides an overview of self-supervised and multimodal learning, emphasizing the specificities of their usage for Earth observation. Lastly, we highlight the scarcity of multimodal EO datasets with diverse data types. + +Self-Supervised Learning. This technique consists in learning expressive data representations without labels by using a pretext task. This approach has been particularly successful for natural language [52] and image [70] analysis. Initially focused on discriminative tasks [37, 68, 100], recent self-supervised approaches for images can be categorized as contrastive or generative. + +Contrastive methods minimize the distance between representations of paired samples, often the same image under different transformations, and maximize the distance with other samples [15, 17, 46]. More efficient methods only consider positive samples and avoid mode collapse by introducing various asymmetries [18, 42] or normalization [16]. Such approaches have been successfully adapted to EO, for which samples are paired according to their location [89] or time of acquisition [7, 63]. + +Generative methods reason at the level of individual token—a small portion of the input, typically a patch for images [25]. The objective is to reconstruct the masked tokens of an input image in pixel [10,45,95] or feature space [6]. This principle has been successfully adapted to EO analysis [20, 30, 99], and was further extended to handle multiple spatial scales [74], multimodality [29, 51], or hyperspectral observations [50, 60]. + +Table 1: Publicly Available Multimodal EO Datasets. We provide in parenthesis the spatial resolutions of the single-date images and labels, and the temporal resolutions of time series. S1/S2 denotes Sentinel-1 and 2. $\star$ : modalities added in this work. + +
DatasetModalitiesLabels
images (single date)time series
SpaceNet6 [81]SAR+optical (0.5m-2m)Xbuilding footprint (<1m)
TreeSatAI [2]aerial + S1/S2 (0.2-10m)Xforestry (60m)
BigEarthNet [86]S1/S2 (10m)Xland cover (100m)
DFC20 [76]S1/S2 (10m)Xland cover (500m)
MDAS [48]S1/S2 + hyperspectral (2.2-10m)Xland cover (0.25m)
DOFA [96]NAIP + Gaofen + S1/S2 + EnMAP (1-30m)XX
PASTIS-R [33, 34]XS1/S2 (30-140 / year)agriculture (10m)
SSL4EO-S12 [92]XS1/S2 (4 / year)X
DFC21-DSE [61]XS1/S2 + LS8 (3-9/year)human activity (500m)
MapInWild [28]XS1/S2 (4 / years)protected areas (10m)
SEN12MS-CR-TS [27]XS1/S2 (30 / years)cloud cover (10m)
MultiSenGE [93]XS1/S2 (30-140 / years)land cover (10m)
FLAIR [31]aerial (0.2m)S2 (20-114 / year)land cover (0.2m)
Satlas [12]NAIP (0.5 -2m)S2 (8-12 / year)various
PASTIS-HD★ SPOT 6-7 (1.5m)S1/S2 (30-140 / year)agriculture (10m)
TreeSatAI-TSaerial (0.2m)★ S1/S2 (10-70 / year)forestry (60m)
+ +Several hybrid approaches combine the discriminative power of contrastive methods and the scalability of generative objectives for natural images [70, 101] and EO data [29]. Our proposed OmniSat model also implements both mechanisms. A key feature is that we leverage the precise alignment between different sources of EO data to contrastively match small patches of different modalities rather than entire images or time series. + +Self-Supervised Multimodal Learning. Multimodal computer vision has received a lot of interest [13], notably due to the success of cross-modal pre-training [72]. Recent models align the embeddings of heterogeneous modalities such as video and sound [49], depth and images [44], text and image [3, 9], or multiple combinations of these modalities [38, 39, 82, 84]. + +Multimodal learning also has a long history in EO [58, 71, 98] due to the large variety and complementarity of sensors [36, 79]. However, recent transformer-based architectures [90] for EO are often limited to one type of modality, be it a single image [20, 74] or time-series [34, 87]. For example, CROMA [29] and PRESTO [88] are specifically designed for paired optical and radar observations, but cannot handle Very High Resolution (VHR) data. USat [51] considers images with different resolutions, but only takes a single date within a time series. UT&T [31] can natively take single and multi-date observations of different modalities, but cannot be easily pre-trained in a self-supervised manner since it relies on convolutions and an ad-hoc late fusion scheme. + +![](images/adaafb273b142fd66f620e4b304cd2da63772604f023462796b53e942b0fafe0.jpg) +Sentinel-2 time series + +![](images/b99d8f55146e578a2c79a563fcf0efaa3b12e81db16e33da5a4225569c1e9431.jpg) +(a) FLAIR + +![](images/9e7dbf1494e3f9941f7ef4e45d4678844922fdf9001530804a691d96a677f73f.jpg) +$\star$ Sentinel-2 time series + +![](images/27a40a7871f836a6b6bcc40bf98238db9b90ba694fdfaa2296a4ca4f295c76c7.jpg) + +![](images/49b1ea627ffa945a279abb165733b2c50fb82deed2f86384c6eba40fc2666e62.jpg) +$\star$ Sentinel-1 time series +(b) TreeSatAI-TS +Fig. 1: Datasets. We represent three tiles from the considered multilabel classification datasets: FLAIR (a), TreeSatAI-TS (b) and PASTIS-HD (c). TreeSatAI-TS is a new dataset built by replacing the single-date Sentinel-1 and 2 images of TreeSatAI [2] by year-long time series. PASTIS-HD (c) adds VHR satellite images to PASTIS-R [34]. $\star$ : modalities added in this work. + +![](images/5fc81623380a7f0c94f69f6c4145dede360af252f84a0be6a46016463dfbbbc7.jpg) +VHR satellite $1.5\mathrm{m}$ + +![](images/0a84d509b9e4e6c42eefd7bc8c1cf3ae9129dad99db837392e258848538730ab.jpg) +PASTIS: Sentinel-2 time series + +![](images/95ac3c5edf81174e746a0aa5ea123991c7d2d45a2b0049f594b7786afda33ee9.jpg) +PASTIS-R: Sentinel-1 time series +(c) PASTIS-HD + +Multimodal EO Datasets. As reported in Table 1, many multimodal EO datasets use Sentinel-1 [11] and 2 [26] data for applications ranging from land cover to forestry analysis and fire detection. We also note that most multimodal datasets only contain data of one type: mono-date image or time series. Several datasets (BigEarthNet [86], DFC20 [76], MDAS [48]) select a single date from time series. However, single Sentinel-1 and 2 acquisitions can be significantly affected by rain and cloud cover, respectively. Furthermore, capturing the temporal dynamics is crucial to characterize the phenology of vegetation [91], + +FLAIR [31] is the first multimodal EO dataset to propose both very high spatial resolution $(\leq 2\mathrm{m})$ and high temporal resolution $(>4$ images/year). Satlas [12] combines Sentinel-2 time series and for $5\%$ to tiles (continental US), very high definition NAIP images. The functional map of the World [19] integrates observations from various sensors, but most areas are only observed with one sensor. Two other datasets contain time series and single images from multiple sources, but were not available at the time of writing this article: IARPA-SMART [40] and DOFA [96]. + +To showcase how OmniSat can consume an arbitrary number of modalities with different spatial, spectral, and temporal resolutions, we selected two commonly used EO benchmarks, TreeSatAI [2] and PASTIS-R [34], whose focus on crop type mapping and forestry differs from the land cover analysis of FLAIR. We added new modalities to these datasets to reach three distinct data types: VHR aerial images, optical time series, and SAR time series. See Fig. 1 for an illustration, and Sec. 4.1 for more details on how we extended these datasets. + +# 3 Method + +We consider a tile $x$ observed through a set $\mathbf{M}$ of $M$ distinct sensors or modalities. The goal of the OmniSat model is to learn in a self-supervised fashion to combine + +![](images/a4ab131949202b7eea26d3f6ad97cfa1a4fd862e9f1aef86144ee95446301553.jpg) +Fig. 2: OmniSat Architecture. We illustrate OmniSat for $M = 3$ modalities, and a tile split into $P = 4$ patches. The $M \times P$ input tokens $x_{\mathbf{P}}^{\mathbf{M}}$ are encoded by $M$ modality-specific encoders $\mathcal{E}^{\mathrm{M}}$ , yielding the token representations $f_{\mathbf{P}}^{\mathbf{M}}$ . The module $\mathcal{C}$ combines them into multimodal patch representations $f_{\mathbf{P}}^{\star}$ . The token embeddings $f_{\mathbf{P}}^{\mathbf{M}}$ are supervised by a contrastive cross-modal objective. We also use a reconstruction objective: the masked multimodal representations $f_{\mathbf{P}}^{\star}$ are decoded by modality-specific networks $\mathcal{D}^{\mathrm{M}}$ to reconstruct their corresponding inputs in $x_{\mathbf{P}}^{\mathbf{M}}$ . + +all modalities $\mathbf{M}$ into a multimodal representation $f^{\star}$ . We first provide details about OmniSat's architecture in Sec. 3.1. We then explain our our training scheme, which consists of a cross-modal contrastive objective (Sec. 3.2) and a multimodal masked encoding task (Sec. 3.3). Finally, we present the implementation details in Sec. 3.4. The overall method is represented in Fig. 2. + +# 3.1 Architecture + +This section presents the tokenization process, the structures of the encoder and decoder for each modality, and the architecture of the modality combiner network. + +Multimodal Tokenization. All available modalities are spatially aligned through georeferencing. This allows us to divide the tile $x$ into a set $\mathbf{P}$ of $P$ non-overlapping patches consistently across all modalities: $x_{p}^{\mathbf{M}} = \{x_{p}^{m}\}_{m\in \mathbf{M}}$ corresponds to $M$ distinct views of the same patch $p$ with different modalities. Each modality $m$ takes its values in a space $\varOmega^m$ such that $x_{p}^{m}\in \varOmega^{m}$ . We index tokens with pairs $(m,p)$ , defined for each modality $m$ and patch $p$ , for a total of $M\times P$ tokens. + +Time series from Sentinel satellites may experience registration errors spanning several meters, complicating their precise alignment with high-resolution imagery. However, using temporal sequences of satellite data mitigates these errors as aggregation over time tends to balance out misalignments. + +![](images/6d67a4b45ff08d54139e398482e768f27ddaa24c13bd2614d0ef3f752fdd9cea.jpg) +(a) Image Encoder and Decoder. + +![](images/aeca5451ce9232f54390e78aabf62bcc3a8cbfe918721b1c3e96a2952cef638d.jpg) +(b) Temporal Encoder and Decoder. + +![](images/38032c22751dbd4962ef7b38027372d7c4e3f01048f4ac3ab6fcb3aa51de3663.jpg) +(c) Modality Combining Network. +Fig. 3: OmniSat Architecture. OmniSat is composed of dedicated patch encoders for image (a) and time series b, here represented for a length of $L = 4$ time stamps. The modality combining module $\mathcal{C}$ is depicted in (c) with $P = 2$ and $M = 3$ . Elements colored in orange are learned networks or parameters. + +Encoder-Decoder for Images. We split image tiles split into small square patches: $\varOmega^{\mathrm{img}}=\mathbb{R}^{C\times W\times W}$ with $W$ the size of the patches in pixels and $C$ the number of channels. As shown in Fig. 3a, we encode these inputs with a sequence of convolutions and max-pool layers until the spatial dimension is fully collapsed. Decoding involves a symmetric sequence of convolutions and un-pooling layers. Contrary to existing masked auto-encoders, we pass the pooling indices from the encoder's max-pooling to the decoder's un-pooling in the manner of SegNet [8]. This dispenses the encoder from learning the intra-patch spatial configuration. This allows the image encoder to focus on the radiometric information, which may be more relevant depending on the application. + +Encoder-Decoder for Time Series. Each temporal patch is represented by $L$ sequential observations with $C$ channels: $\Omega^{\mathrm{TS}} = \mathbb{R}^{C\times L}$ , each associated with a time stamp. We encode the temporal patches using a Lightweight Temporal Attention Encoder (LTAE) model [32], an efficient network for geospatial time series processing. We decode vector representations into time series by repeating the vector $L$ times across the temporal dimension, adding a temporal encoding for each time step, and using an MLP to map the results to size $C$ . See Fig. 3b for an illustration. + +Optical time series are notoriously affected by clouds [85]. This may affect the validity of the reconstruction task: the decoder cannot know which observations are cloudy, making the reconstruction objective unpredictable. To circumvent this issue, we use the temporal attention maps of the encoder's LTAE to select dates to reconstruct: cloudless observations are more informative and should have a higher attention score [78]. We only consider in the reconstruction loss $\mathcal{L}_{\mathrm{reconstr}}$ the top $25\%$ dates in terms of the LTAE's attention maps. + +Modality Combining Network. The modality combining network $\mathcal{C}$ , represented in Fig. 3c, takes the $M \times P$ token embeddings $f_{\mathbf{P}}^{\mathbf{M}}$ , some of whom can potentially be masked. We equip each token with a Euclidean relative positional encoding [94], calculated based on their patch's position $\{r(p,q) \mid (p,q) \in \mathbf{P}^2\}$ , allowing each token to + +selectively consider its spatial surroundings. As most EO data are captured from above (satellite or aerial), their distribution is invariant by horizontal translation, making this choice of encoding preferable to an absolute position encoding. + +The modality combining module $\mathcal{C}$ starts with a series of $B$ residual self-attention blocks connecting all tokens across modality. We then perform cross-attention between the resulting token embeddings $g_{\mathbf{P}}^{\mathbf{M}} \in \mathbb{R}^{d \times M \times P}$ and $P$ copies $f_{\mathbf{P}}^{\mathrm{comb}}$ of a modality combining token $f^{\mathrm{comb}} \in \mathbb{R}^{d}$ learned as a free parameter. Each copy of $f_{p}^{\mathrm{comb}}$ is spatially located at the patch $p$ for the relative positional encoding $r$ . The module $\mathcal{C}$ outputs $P$ multimodal encodings $f_{\mathbf{P}}^{\star}$ combining all available modalities for each patch: + +$$ +g _ {\mathbf {P}} ^ {\mathbf {M}} = \text {s e l f - a t t e n t i o n} \left(f _ {\mathbf {P}} ^ {\mathbf {M}}; r\right) \tag {1} +$$ + +$$ +f _ {\mathbf {P}} ^ {f \star} = \text {c r o s s - a t t e n t i o n} \left(f _ {\mathbf {P}} ^ {\text {c o m b}}, g _ {\mathbf {P}} ^ {\mathbf {M}}; r\right). \tag {2} +$$ + +# 3.2 Contrastive Objective + +We denote by $f_{p}^{m}$ the $d$ -dimensional encodings of the input patch $x_{p}^{m}$ given by their dedicated encoders. We propose to supervise the embeddings $f_{p}^{m}$ with a contrastive objective encouraging spatial consistency across modalities. Indeed, while each modality captures distinct characteristics of $p$ , all encodings $f_{p}^{m}$ share the same latent variable: the semantic content of the patch. + +In practice, we want $f_{p}^{m}$ to be closer to $f_{p}^{n}$ for $n \neq m$ , than to $f_{q}^{n}$ for other patches $q \neq p$ . We define $\mathbf{B}$ as the set of patches within the current batch of observations. We adapt the classic InfoNCE loss [69] to our setting with two main differences, illustrated in Fig. 4. (i) Each token $(m,p)$ has $M - 1$ positive matches: the tokens corresponding to the same patch $p$ but viewed in another modality $n \neq m$ ; and (ii) as EO observations are generally spatially regular, nearby patches may be visually indistinguishable. Therefore, we exclude from the negative matches of $(m,p)$ all tokens in modality $m$ and which are too close to $p$ . To this end, we remove the set $T(m,p)$ of tokens with modality $m$ and whose patches are in the same tile as $p$ . Our loss function $\mathcal{L}_{\mathrm{contrast}}$ is defined as such: + +$$ +\mathcal {L} _ {\text {c o n t r a s t}} = \frac {1}{M | \mathbf {B} |} \sum_ {(m, p) \in \mathbf {M} \times \mathbf {B}} \log \left(\frac {\sum_ {n \neq m} \exp \left(\left\langle f _ {p} ^ {m} , f _ {p} ^ {n} \right\rangle / \gamma\right)}{\sum_ {(n , q) \in \mathbf {M} \times \mathbf {B} \backslash T (m , p)} \exp \left(\left\langle f _ {p} ^ {m} , f _ {q} ^ {n} \right\rangle / \gamma\right)}\right), \tag {3} +$$ + +with $\gamma$ a temperature parameter, and $\langle \cdot ,\cdot \rangle$ the scalar product in $\mathbb{R}^d$ . This function, specifically designed for geospatial data, allows us to contrast individual patches across modalities, which is not typically feasible for natural images. However, as the contrastive objective aligns multimodal representations, the patch encoders may be encouraged to overlook the distinct attributes of their respective modality. Instead, they may focus only on features shared by all modalities, i.e., their common denominator. To ensure that encoders also capture modality-specific information, we incorporate a reconstruction objective, detailed in Sec. 3.3. + +# 3.3 Multimodal Reconstruction Objective + +During training, we mask a fraction of tokens $\mathbf{K} \subset \mathbf{M} \times \mathbf{P}$ and replace their embeddings with a learned vector $f^{\mathrm{mask}} \in \mathbb{R}^{d}$ . Note that the masking can differ across modalities, + +Fig.4:Contrastive Loss. We represent the token matching matrix for two tiles $\mathrm{Tile}_1$ and $\mathrm{Tile}_2$ viewed across 3 modalities $m_{1}$ $m_2$ ,and $m_3$ Tile is composed of the patches $p_1$ and $p_2$ ,while $\mathrm{Tile}_2$ comprises $q_{1}$ and $q_{2}$ .In contrast to classic approaches which ignore the diagonal and assign each sample with a single positive match, our loss defines operates at the patch level, considers multiple positives per token, and excludes tokens in a block-diagonal fashion. +![](images/509f9339c2eb39b7a94dc06dc20da0ac9a4c337a553e964b614eddf521e8846d.jpg) ++ positive match +- negative match +o ignored + +and some patches may be entirely masked. All tokens are then processed by the modality combining network $\mathcal{C}$ , which outputs $P$ multimodal embeddings $f_{\mathbf{P}}^{\star}$ : + +$$ +f _ {\mathbf {P}} ^ {\star} = \mathcal {C} \left(\left\{f _ {p} ^ {m} \right\} _ {(m, p) \notin \mathbf {K}} \cup \left\{f ^ {\text {m a s k}} \right\} _ {(m, p) \in \mathbf {K}}\right). \tag {4} +$$ + +To encourage the patch embeddings $f_{\mathbf{P}}^{*}$ to capture information from all modalities, we build a multimodal reconstruction objective. We denote by $\mathcal{D}^m: \mathbb{R}^d \mapsto \Omega^m$ the dedicated decoder of each modality $m$ and write the reconstruction loss as: + +$$ +\mathcal {L} _ {\text {r e c o n s t r}} = \frac {1}{| \mathbf {K} |} \sum_ {(m, p) \in \mathbf {K}} \frac {1}{\dim (\Omega^ {m})} \| \mathcal {D} ^ {m} \left(f _ {p} ^ {\star}\right) - x _ {p} ^ {m} \| ^ {2}, \tag {5} +$$ + +with $\dim (\varOmega^m)$ the dimension of $\varOmega^m$ . The total loss is the sum of $\mathcal{L}_{\mathrm{reconstr}}$ and $\mathcal{L}_{\mathrm{contrast}}$ . + +# 3.4 Implementation Details + +We detail here the specific parameters chosen in all our experiments. + +Tokenization. We split each tile along a regular spatial grid to produce a set of non-overlapping patches $\mathbf{P}$ consistent across all modalities. For TreeSat and FLAIR, we use a $10 \times 10\mathrm{m}$ grid, meaning that the VHR input tokens are small image patches of size $50 \times 50$ with $0.2\mathrm{m}$ per pixel. The patches of Sentinel observations with a resolution of $10\mathrm{m}$ are single-pixel temporal sequences of spectral measurements. For PASTIS-HD, we use a $40 \times 40\mathrm{m}$ grid, meaning that the VHR patches are of size $40 \times 40$ with $1.0\mathrm{m}$ per pixel. The patches of Sentinel observations [26] are $4 \times 4$ image time series which we spatially flatten before encoding. + +Hyperparameters. To show the versatility of OmniSat, we use the same configuration throughout all experiments. The embedding size is $d = 256$ , resulting in image encoders and decoders with 3.6M and 1.8M parameters, 403k and 96k for optical time series, and 402k and 95k for radar time series. The modality combiner module is composed of $B = 6$ residual self-attention blocks and a single cross-attention block, for a total of 3.6M parameters. We train our model on 3 A6000 GPUs with a batch size of 128 + +multimodal tiles per GPU and set the contrastive temperature $\gamma$ to 0.1. We train our model with the ADAM optimizer [53], with a learning rate of $10^{-4}$ for pretraining and $2\times 10^{-5}$ for fine-tuning, and a ReduceLROnPlateau scheduler [1] with a patience of 10 epochs and a decay rate of 0.1. When re-implementing competing methods, we use the hyperparameters of their open-source repository. + +# 4 Experiments + +We evaluate OmniSat's performance across three multimodal datasets, including two new datasets introduced in this work, and presented in Sec. 4.1. We outline our experimental protocol and our adaptation of competing methods in Sec. 4.2. We then present our quantitative results and analysis in Sec. 4.3 and conduct an ablation study in Sec. 4.4. + +# 4.1 Datasets + +We consider three multimodal datasets: FLAIR [31], and the augmented TreeSatAI-TS [2] and PASTIS-HD [33, 34]. See Fig. 1 for an illustration of these two last datasets. + +TreeSatAI-TS: TreeSatAI [2] is a multimodal dataset for tree species identification, containing 50,381 tiles of $60 \times 60\mathrm{m}$ with multi-label annotations for 20 classes and all taken in Germany. Each tile is associated with a very high resolution RGB and near-infrared (NIR) image ( $0.2\mathrm{m}$ pixel resolution), a single Sentinel-2 multi-spectral image ( $10\mathrm{m}$ per pixel resolution, 10 bands), and a single Sentinel-1 radar image ( $10\mathrm{m}$ per pixel resolution, 3 bands: two polarization channels and their ratio). + +Motivated by the fact that fine-grained vegetation discrimination relies heavily on temporal dynamics information [91], we introduce TreeSatAI- $\mathrm{TS}^1$ . This extended version uses open-source data to add Sentinel-1 and Sentinel-2 time series to each tile, spanning the closest available year to the VHR observation for Sentinel-2. Note that due to the weather patterns and position of the area of interest with respect to Sentinel-2's orbit, the optical time series is particularly irregular and occluded, with up to $50\%$ of acquisitions being non-exploitable. Despite this challenge, we included the raw observations without pre-processing, whereas TreeSatAI's single-date images have been manually selected. + +PASTIS-HD: The PASTIS dataset [33], is designed for semantic and panoptic segmentation of agricultural parcels using Sentinel-2 time series and covers 18 crop types across 2,433 image time series with dimensions of $1280 \times 1280\mathrm{m}$ . Each series contains between 38 and 61 observations with 10 spectral bands. PASTIS-R [34] adds the corresponding Sentinel-1 radar time series. We only used the ascendant time series of Sentinel-1 for our training and evaluation, for a total of 169,587 radar images with three bands. + +To enhance the spatial resolution and utility of PASTIS, we introduce PASTISHD², which integrates contemporary VHR satellite images (SPOT 6-7 [24]). We apply orthorectification and pansharpening, resample the resulting images to a 1m resolution, and finally convert them to 8 bits. We follow the protocol of Irvin et al. [51] to use the + +dense annotations for a multi-label classification task: each patch is associated with the labels of all of its pixels. This conversion allows us to evaluate all methods in the same setting and configuration as TreeSatAI. + +FLAIR. The FLAIR dataset [31] combines VHR aerial images with time series data. It comprises 77,762 aerial tiles ( $512 \times 512$ pixels, $0.2\mathrm{m}$ resolution) with five channels (RGB, near-infrared, and a normalized digital surface model) taken in France, alongside corresponding Sentinel-2 time series ( $10\mathrm{m}$ resolution, 10 spectral bands, 20 to 114 observations per year). We apply the same processing as PASTIS to use the dense annotation for a multi-label classification task. + +# 4.2 Experimental Setting + +This section details our experimental protocol and our adaption of competing algorithms. + +Evaluation Protocol. All experiments follow a similar setting: + +- Pre-training (optional). Methods that support self-supervised pre-training (OmniSat, SatMAE [20], ScaleMAE [74], CROMA [29]) are pre-trained for up to 250 epochs on the entire training set without access to labels. +- Fine-Tuning. We propose two settings for fine-tuning: + +- Fully Supervised Fine-Tuning. We train the resulting models using all the labels in the training set. +- Semi-Supervised Fine-Tuning. We use a portion of $10\%$ or $20\%$ of the training set, stratified by the distribution of classes, to fine-tune the models. For models without pre-training, this corresponds to supervision in the low-data regime. + +- Unimodal and Multimodal Evaluation. We evaluate all methods using each available modality independently and combining all supported modalities. + +Adapting Competing Approaches. We report the performance of several methods taken from the literature on our considered datasets: LightGBM [2], PRESTO [88], and MOSAIKS [77]. However, few existing methods can operate on single- and multi-date data at the same time. To ensure a fair evaluation of competing approaches, we modify various state-of-the-art models to handle a broader combination of modalities. We provide details on these changes in the appendix. + +# 4.3 Numerical Experiments and Analysis + +In this section, we report our model's performance and efficiency compared to other approaches across the considered datasets and propose our analysis. + +TreeSatAI-TS. Tab. 2 presents the performance of different models on TreeSatAI and TreeSatAI-TS. We report several key observations: + +- Benefit of Time Series. For the original TreeSatAI dataset with single-date Sentinel1/2 observations, none of the pre-training schemes significantly improve performance beyond simple baselines such as ResNet, PSE, or MLP, even in a semi-supervised setting. In particular, single-date S1 observations yield low performance for all methods (below 20 F1-score), emphasizing the need to use the entire time series. + +Table 2: Performance on TreeSatAI-TS. We report the weighted F1 for multi-label tree species classification on TreeSatAI (TSAI) and our extended TreeSatAI-TS (TSAI-TS) dataset when fine-tuning with $10\%$ and $100\%$ of training labels. The first line of the table is the modality used for evaluation. We distinguish methods that are best for one modality within a dataset, best in a dataset across all modalities, and the best overall performance. $\star$ : late feature fusion with a ResNet pre-trained on ImageNet. $\blacksquare$ : Foundation model trained on extensive external data. $\dagger$ : model evaluated on this dataset for the first time. + +
Modelpre-trainingAllSentinel-1Sentinel-2VHR Image
10%100%10%100%10%100%10%100%
Evaluated on TreeSatAI: single date for Sentinel-1 and Sentinel-2
† PSE [35]None47.2*68.1*11.514.648.558.3--
† ViT [25]None42.757.18.717.539.857.336.751.7
MLP [2]None42.6*71.5*3.410.122.152.0--
ResNet [2]ImageNet------58.870.1
LightGBM [2]ImageNet-54.3*-11.9-48.2-44.0
PRESTO [88]---19.8-46.3--
† DOFA [96]59.571.611.619.348.257.051.667.5
MOSAIKS [22,77]TSAI-----56.0--
† CROMA [29]TSAI49.661.010.112.747.855.7--
† SatMAE [20]TSAI46.161.5--40.349.744.161.4
† ScaleMAE [74]TSAI47.662.5--46.755.246.963.6
OmniSat (ours)TSAI56.270.45.36.448.457.152.868.9
Evaluated on TreeSatAI-TS: Sentinel-1 and Sentinel-2 time series spanning one year
† PSE+LTAE [35]None59.4*71.2*42.652.444.057.2--
† UT&T [31]ImageNet43.856.742.355.241.557.044.355.9
† DOFA [96]41.871.30.00.025.039.451.667.5
† Scale-MAE [74]TSAI-TS44.160.4--11.031.546.963.6
OmniSat (ours)None52.273.331.655.933.949.751.471.0
OmniSat (ours)TSAI-TS61.174.248.256.751.462.958.370.5
+ +OmniSat exhibits significantly improved results on TreeSatAI-TS, with or without pretraining. Image models struggles to extract meaningful features temporally aggregated temporal observations, while OmniSat learn rich dynamic features. + +The foundation model DOFA [96], with 111M parameters and a large closed-source training set, outperforms all models when evaluated on single-date modalities. However, OmniSat reaches higher performances on TreeSatAI-TS with only 10 million parameters, which we attribute to its ability to leverage temporal modalities. + +- Benefits of Multimodality. When using all modalities, OmniSat outperforms all competing methods by a margin of $3\%$ F1-score. The multimodal performance of OmniSat and CROMA, which learn to combine data sources, is strictly superior to the F1-score of their best modality by $3.7\%$ and $5.3\%$ points, respectively. Conversely, the performance of methods that rely on late-fusion (SatMAE, ScaleMAE, + +Table 3: Performance on PASTIS-HD. We report the macro-averaged F1-score for crop-type multi-class classification on the PASTIS-HD dataset. We distinguish methods that are best for one modality, best in a dataset across all modalities. $\star$ : late feature fusion with a ResNet. $\dagger$ : model evaluated on this dataset for the first time. + +
Modelpre-trainedAllSentinel-1Sentinel-2VHR image
20%100%20%100%20%100%20%100%
† UTAE [33, 34]None36.8*46.9*20.140.732.737.6--
† ResNet50 [47]ImageNet------57.659.3
† UT&T [31]ImageNet54.253.558.862.854.961.351.149.8
† DOFA [96]53.755.736.741.550.853.447.954.8
† Scale-MAE [74]PASTIS-HD42.042.2--41.246.148.851.9
† CROMA [29]PASTIS-HD57.560.155.356.153.056.7--
OmniSat (ours)No42.059.158.260.251.760.147.352.8
OmniSat (ours)PASTIS-HD62.669.960.869.061.870.854.659.3
+ +Table 4: Performance on FLAIR. We report the macro-averaged F1-score for land cover multiclass classification on the FLAIR dataset. We distinguish methods that are best for one modality and best in a dataset. $\dagger$ : model evaluated on this dataset for the first time. + +
Modelpre-trainedAllSentinel-2VHR Image
10%100%10%100%10%100%
† UT&T [31]ImageNet44.248.857.462.058.965.5
† DOFA [96]70.674.957.061.066.872.1
† ScaleMAE [74]FLAIR63.170.052.561.061.267.3
OmniSat (ours)No62.570.056.165.464.771.5
OmniSat (ours)FLAIR60.673.456.865.465.271.6
+ +ViT) is comparable to their best modality. This demonstrates the value of learning to combine information from different sources end-to-end. + +- Benefits of Cross-Modal Pre-Training. With access to all modalities, our self-supervised pre-training improves by $0.9\%$ point the F1-score of the model fine-tuned on $100\%$ of labels, compared to not pre-training, and $8.9\%$ when using only $10\%$ of labels. This shows that our pre-training leads to more expressive multimodal features. Interestingly, when performing inference with Sentinel-2 time series alone, the performance increase linked to the pre-training becomes $13.2\%$ with $100\%$ labels and $17.5\%$ with $10\%$ . This illustrates that our self-supervised pre-training scheme improves the features learned by each encoder despite not relying on annotated data. + +Experiments on PASTIS-HD. The analysis of the performance of various models on PASTIS-HD is reported in Tab. 3, and is consistent with the ones of TreeSatAI-TS. First, by learning to combine all modalities despite their different resolutions, OmniSat achieves state-of-the-art results on this benchmark. Second, our cross-modal pretraining + +![](images/de0d8468d42cad0b4ccc124f4aeb30b473d15b7ff6c72aeacecfa27a202211ca.jpg) +Fig. 5: Efficiency. We report the best performance of different models between TreeSatAI and TreeSatAI-TS, with pre-training and fine-tuning using $100\%$ of labels. The area of the markers is proportional to the training time, broken down in pre-training and fine-tuning when applicable. + +significantly improves OmniSat's performance in the multimodal $(+10.8\mathrm{pF1}$ -score with $100\%$ of training label) and all single-modality settings (8.8 points for Sentinel-1, 10.7 for Sentinel-2, and 6.5 for the VHR images). + +Experiments on FLAIR. We report in Tab. 4 the results on the bimodal FLAIR dataset for multilabel classification. OmniSat outperforms the much larger ScaleMAE [74] and UT&T [31] models with $100\%$ of labels and both modalities by $3.4\%$ . Our pre-training scheme had a smaller impact than for the TreeSatAI-TS experiment. We attribute this to the fact that only two modalities are available, which decreases the supervisory power of our cross-modal contrastive objective and our multimodal reconstruction loss. This highlights a limitation of OmniSat: the model needs to be pre-trained on a modality-rich dataset to achieve its best performance. + +Efficiency Evaluation. We plot in Fig. 5 the best performance between TreeSatAI and TreeSatAI-TS for different models according to their size and training time. OmniSat is more compact, faster to train, and performs better than all evaluated models, including the DOFA foundation model. The highly-specialized combination of PSE, LTAE, and ResNet is a strong contender, outperforming significantly larger models with generic encoding-decoding schemes. + +# 4.4 Ablation Study + +In this section, we report the results of several experiments evaluating the impact and validity of our main design choices, see Tab. 5. + +a) Encoder/Decoder Architecture. We propose several improvements to the standard image encoder-decoder scheme used in computer vision to accommodate the specificities of EO data. In particular, passing the max-pool indices from the image patch encoder to its decoder allows the learned representation to focus on characterizing the spectral signature instead of fine-grained spatial information, and leads to a performance increase of $0.7\%$ in the full supervision setting. + +Table 5: Ablation Study. We present the impact of several design choices on the TreeSatAI-TS dataset, measured in terms of macro-averaged F1-score. + +
Experiment10%100%Experiment10%100%
OmniSat61.174.2b) no contrastive loss55.673.4
a) no index bypass57.573.5b) naive contrastive loss57.872.2
a) no date filtering58.271.6b) no reconstruction loss59.072.2
+ +As clouds frequently obstruct optical time series, we use a unsupervised date-filtering scheme to reconstruct only meaningful acquisitions. This approach leads to a significant improvement of $3.6\%$ , showcasing the benefit of developing modality-aware approaches for EO. + +b) Role of Loss Functions. When training without contrastive loss, we observe a decrease in performance of $0.8\%$ in the fully supervised regime, and a more pronounced drop of $5.5\%$ in the semi-supervised regime. This demonstrates how learning consistent encoding across encoders facilitates their subsequent fusion. Interestingly, when implementing a naive contrastive loss that considers all negative examples from the batch, the decrease is greater than simply removing this loss ( $2\%$ in full supervision). This strategy may introduce indistinguishable negative examples and perturb the learning process. + +We also remove the reconstruction loss, meaning that only the encoders are learned contrastively during pre-training. This results in a drop of $2\%$ F1-score point, illustrating the importance of pre-training the transformer $\mathcal{C}$ alongside its encoders. + +Limitations. All datasets used in our experiments are based in Europe, primarily due to the availability of open-access annotations. This regional focus prevents us from evaluating our model's performance in tropical and developing countries, which present unique challenges in terms of label provision, heterogeneity, and complex classes. + +A limitation of our pre-training scheme is its dependence on a sufficient number of aligned modalities, as illustrated by its moderate impact on the bimodal FLAIR dataset. + +# 5 Conclusion + +We introduced OmniSat, a new architecture for the self-supervised modality fusion of Earth Observation (EO) data from multiple sources. To facilitate its evaluation, we augmented two existing datasets with new modalities of different natures and resolutions. We experimentally showed that leveraging diverse modalities with a flexible model improves the model's performance in both fully and semi-supervised settings. Moreover, our training scheme can exploit the spatial alignment of multiple modalities to improve our model's unimodal performance. Finally, we proposed several improvements to leverage the unique structure of EO data in the architecture of our model, such as automatic date filtering for reconstructing time series. We hope that our promising results and new datasets will encourage the computer vision community to consider EO data as a playing field for evaluating and developing novel self-supervised multimodal algorithms. + +# Acknowledgements + +This work was supported by ANR project READY3D ANR-19-CE23-0007, and was granted access to the HPC resources of IDRIS under the allocations AD011014719 and AD011014286R1 made by GENCI. We thank Anatol Garioud and Sébastien Giordano for their help on the creation of TreeSatAI-TS and PASTIS-HD datasets. The SPOT images are opendata thanks to the Dataterra Dinamis initiative in the case of the "Couverture France DINAMIS" program. We thank Jordi Inglada for inspiring discussions and valuable feedback. + +# References + +1. PyTorch: ReduceLROOnPlateau.org/docs/stable/generated/torch.optim.lr_scheduler.ReduceLROOnPlateau.html#torch.optim.lr_scheduler.ReduceLROOnPlateau, accessed: 2024-02-29 9 +2. Ahlswede, S., Schulz, C., Gava, C., Helber, P., Bischke, B., Förster, M., Arias, F., Hees, J., Demir, B., Kleinschmit, B.: TreeSatAI Benchmark Archive: A multi-sensor, multi-label dataset for tree species classification in remote sensing. Earth System Science Data Discussions (2022) 2, 3, 4, 9, 10, 11 +3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: A visual language model for few-shot learning. In: NeurIPS (2022) 3 +4. Amitrano, D., Di Martino, G., Guida, R., Iervolino, P., Iodice, A., Papa, M.N., Riccio, D., Ruello, G.: Earth environmental monitoring using multi-temporal synthetic aperture radar: A critical review of selected applications. Remote Sensing (2021) 1 +5. Anderson, K., Ryan, B., Sonntag, W., Kavvada, A., Friedl, L.: Earth observation in service of the 2030 agenda for sustainable development. Geo-spatial Information Science (2017) 1 +6. Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding predictive architecture. In: CVPR (2023) 2 +7. Ayush, K., Uzkent, B., Meng, C., Tanmay, K., Burke, M., Lobell, D., Ermon, S.: Geography-aware self-supervised learning. In: ICCV (2021) 2 +8. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE TPAMI (2017) 6 +9. Baevski, A., Hsu, W.N., Xu, Q., Babu, A., Gu, J., Auli, M.: Data2vec: A general framework for self-supervised learning in speech, vision and language. In: ICML (2022) 3 +10. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2021) 2 +11. Bao, X., Zhang, R., Lv, J., Wu, R., Zhang, H., Chen, J., Zhang, B., Ouyang, X., Liu, G.: Vegetation descriptors from Sentinel-1 SAR data for crop growth monitoring. ISPRS Journal of Photogrammetry and Remote Sensing (2023) 4 +12. Bastani, F., Wolters, P., Gupta, R., Ferdinando, J., Kembhavi, A.: SatlasPretrain: A large-scale dataset for remote sensing image understanding. In: ICCV (2023) 3, 4 +13. Bayoudh, K., Knani, R., Hamdaoui, F., Mtibaa, A.: A survey on deep multimodal learning for computer vision: Advances, trends, applications, and datasets. The Visual Computer (2022) 3 + +14. Benedetti, P., Ienco, D., Gaetano, R., Ose, K., Pensa, R.G., Dupuy, S.: $\mathbf{M}^3$ -fusion: A deep learning architecture for multiscale multimodal multitemporal satellite data fusion. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2018) 2 +15. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020) 2 +16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 2 +17. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020) 2 +18. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021) 2 +19. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018) 4 +20. Cong, Y., Khanna, S., Meng, C., Liu, P., Rozi, E., He, Y., Burke, M., Lobell, D., Ermon, S.: SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In: NeurIPS (2022) 2, 3, 10, 11 +21. Coppin, P., Lambin, E., Jonckheere, I., Muys, B.: Digital change detection methods in natural ecosystem monitoring: A review. Analysis of multi-temporal remote sensing images (2002) 1 +22. Corley, I., Robinson, C., Dodhia, R., Ferres, J.M.L., Najafirad, P.: Revisiting pre-trained remote sensing model benchmarks: Resizing and normalization matters. arXiv preprint arXiv:2305.13456 (2023) 11 +23. Dai, A., Nießner, M.: 3DMV: Joint 3D-multi-view prediction for 3D semantic scene segmentation. In: ECCV (2018) 2 +24. DataTerra Dinamis: Diffusion OpenData Dinamis, https://dinamis.data-terra.org/opendata/, accessed: 2023-12-15 9 +25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 11 +26. Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B., Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F., Bargellini, P.: Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment (2012) 1, 4, 8 +27. Ebel, P., Xu, Y., Schmitt, M., Zhu, X.X.: SEN12MS-CR-TS: A remote-sensing data set for multimodal multitemporal cloud removal. IEEE TGRS (2022) 3 +28. Ekim, B., Stomberg, T.T., Roscher, R., Schmitt, M.: MapInWild: A remote sensing dataset to address the question of what makes nature wild. IEEE Geoscience and Remote Sensing Magazine (2023) 3 +29. Fuller, A., Millard, K., Green, J.R.: CROMA: Remote sensing representations with contrastive radar-optical masked autoencoders. In: NeurIPS (2023) 1, 2, 3, 10, 11, 12 +30. Gao, Y., Sun, X., Liu, C.: A general self-supervised framework for remote sensing image classification. Remote Sensing (2022) 2 +31. Garioud, A., Gonthier, N., Landrieu, L., De Wit, A., Valette, M., Poupee, M., Giordano, S., Wattrelos, B.: FLAIR: A country-scale land cover semantic segmentation dataset from multi-source optical imagery. In: NeurIPS Dataset and Benchmark (2023) 3, 4, 9, 10, 11, 12, 13 +32. Garnot, V.S.F., Landrieu, L.: Lightweight temporal self-attention for classifying satellite images time series. In: Advanced Analytics and Learning on Temporal Data: ECML PKDD Workshop (2020) 6 +33. Garnot, V.S.F., Landrieu, L.: Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In: ICCV (2021) 2, 3, 9, 12 + +34. Garnot, V.S.F., Landrieu, L., Chehata, N.: Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing (2022) 2, 3, 4, 9, 12 +35. Garnot, V.S.F., Landrieu, L., Giordano, S., Chehata, N.: Satellite image time series classification with pixel-set encoders and temporal self-attention. In: CVPR (2020) 11 +36. Ghamisi, P., Rasti, B., Yokoya, N., Wang, Q., Hofle, B., Bruzzone, L., Bovolo, F., Chi, M., Anders, K., Gloaguen, R., et al.: Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine (2019) 1, 3 +37. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018) 2 +38. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023) 1, 2, 3 +39. Girdhar, R., Singh, M., Ravi, N., van der Maaten, L., Joulin, A., Misra, I.: Omnivore: A single model for many visual modalities. In: CVPR (2022) 2, 3 +40. Goldberg, H.R., Ratto, C.R., Banerjee, A., Kelbaugh, M.T., Giglio, M., Vermote, E.F.: Automated global-scale detection and characterization of anthropogenic activity using multi-source satellite-based remote sensing imagery. In: Geospatial Informatics XIII. SPIE (2023) 4 +41. Greenwell, C., Crall, J., Purri, M., Dana, K., Jacobs, N., Hadzic, A., Workman, S., Leotta, M.: WATCH: Wide-area terrestrial change hypercube. In: WACV (2024) 2 +42. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. In: NeurIPS (2020) 2 +43. Hackstein, J., Sumbul, G., Clasen, K.N., Demir, B.: Exploring masked autoencoders for sensor-agnostic image retrieval in remote sensing. arXiv preprint arXiv:2401.07782 (2024) 2 +44. Hazirbas, C., Ma, L., Domokos, C., Cremers, D.: FuseNet: Incorporating depth into semantic segmentation via fusion-based CNN architecture. In: ACCV (2017) 3 +45. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2 +46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 2 +47. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 12 +48. Hu, J., Liu, R., Hong, D., Camero, A., Yao, J., Schneider, M., Kurz, F., Segl, K., Zhu, X.X.: MDAS: A new multimodal benchmark dataset for remote sensing. Earth System Science Data Discussions (2022) 3, 4 +49. Huang, P.Y., Sharma, V., Xu, H., Ryali, C., Fan, H., Li, Y., Li, S.W., Ghosh, G., Malik, J., Feichtenhofer, C.: MAViL: Masked audio-video learners. In: NeurIPS (2023) 2, 3 +50. Ibanez, D., Fernandez-Beltran, R., Pla, F., Yokoya, N.: Masked auto-encoding spectral-spatial transformer for hyperspectral image classification. IEEE TGRS (2022) 2 +51. Irvin, J., Tao, L., Zhou, J., Ma, Y., Nashold, L., Liu, B., Ng, A.Y.: USat: A unified self-supervised encoder for multi-sensor satellite imagery. arXiv preprint arXiv:2312.02199 (2023) 1, 2, 3, 9 +52. Kenton, J.D.M.W.C., Toutanova, L.K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 2 +53. Kingma, D.P., Ba, J.: ADAM: A method for stochastic optimization. ICLR (2015) 9 +54. Krispel, G., Opitz, M., Waltner, G., Possegger, H., Bischof, H.: Fuseseg: LiDAR point cloud segmentation fusing multi-modal data. In: WACV (2020) 2 + +55. Kuffer, M., Thomson, D.R., Boo, G., Mahabir, R., Grippa, T., Vanhuysse, S., Engstrom, R., Ndugwa, R., Makau, J., Darin, E., et al.: The role of Earth observation in an integrated deprived area mapping "system" for low-to-middle income countries. Remote sensing (2020) 1 +56. Lacoste, A., Sherwin, E.D., Kerner, H., Alemohammad, H., Lütjens, B., Irvin, J., Dao, D., Chang, A., Gunturkun, M., Drouin, A., et al.: Toward foundation models for Earth monitoring: Proposal for a climate change benchmark. arXiv preprint arXiv:2112.00570 (2021) 1 +57. Li, D., Tong, Q., Li, R., Gong, J., Zhang, L.: Current issues in high-resolution Earth observation technology. Science China Earth sciences (2012) 1 +58. Li, J., Hong, D., Gao, L., Yao, J., Zheng, K., Zhang, B., Chanussot, J.: Deep learning in multimodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation 112, 102926 (2022) 3 +59. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE TPAMI (2022) 2 +60. Liu, Y., Li, X., Hua, Z., Xia, C., Zhao, L.: A band selection method with masked convolutional autoencoder for hyperspectral image. IEEE Geoscience and Remote Sensing Letters (2022) 2 +61. Ma, Y., Li, Y., Feng, K., Xia, Y., Huang, Q., Zhang, H., Prieur, C., Licciardi, G., Malha, H., Chanussot, J., et al.: The outcome of the 2021 IEEE GRSS data fusion contest-Track DSE: Detection of settlements without electricity. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3 +62. Mai, G., Huang, W., Sun, J., Song, S., Mishra, D., Liu, N., Gao, S., Liu, T., Cong, G., Hu, Y., et al.: On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798 (2023) 1 +63. Manas, O., Lacoste, A., Giró-i Nieto, X., Vazquez, D., Rodriguez, P.: Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In: ICCV (2021) 2 +64. Manfreda, S., McCabe, M.F., Miller, P.E., Lucas, R., Pajuelo Madrigal, V., Mallinis, G., Ben Dor, E., Helman, D., Estes, L., Ciraolo, G., et al.: On the use of unmanned aerial systems for environmental monitoring. Remote sensing p. 641 (2018) 1 +65. Moreira, A., Prats-Iraola, P., Younis, M., Krieger, G., Hajnsek, I., Papathanassiou, K.P.: A tutorial on synthetic aperture radar. IEEE Geoscience and Remote Sensing Magazine (2013) 1 +66. Nakalembe, C.: Urgent and critical need for sub-saharan african countries to invest in Earth observation-based agricultural early warning and monitoring systems. Environmental Research Letters (2020) 1 +67. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from RGBD images. In: ECCV (2012) 2 +68. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016) 2 +69. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 7 +70. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. TLMR (2023) 2, 3 +71. Pohl, C., Van Genderen, J.L.: Multisensor image fusion in remote sensing: Concepts, methods and applications. International journal of remote sensing (1998) 3 +72. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3 + +73. Recasens, A., Lin, J., Carreira, J., Jaegle, D., Wang, L., Alayrac, J.b., Luc, P., Miech, A., Smaira, L., Hemsley, R., et al.: Zorro: The masked multimodal transformer. arXiv preprint arXiv:2301.09595 (2023) 2 +74. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: ICCV (2023) 2, 3, 10, 11, 12, 13 +75. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3D semantic segmentation. In: CVPR (2022) 2 +76. Robinson, C., Malkin, K., Jojic, N., Chen, H., Qin, R., Xiao, C., Schmitt, M., Ghamisi, P., Hänsch, R., Yokoya, N.: Global land-cover mapping with weak supervision: Outcome of the 2020 IEEE GRSS data fusion contest. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3, 4 +77. Rolf, E., Proctor, J., Carleton, T., Bolliger, I., Shankar, V., Ishihara, M., Recht, B., Hsiang, S.: A generalizable and accessible approach to machine learning with global satellite imagery. Nature communications (2021) 10, 11 +78. Rußwurm, M., Körner, M.: Self-attention for raw optical satellite time series classification. ISPRS Journal of Photogrammetry and Remote Sensing (2020) 6 +79. Schmitt, M., Zhu, X.X.: Data fusion and remote sensing: An ever-growing relationship. IEEE Geoscience and Remote Sensing Magazine (2016) 1, 3 +80. Secades, C., O'Connor, B., Brown, C., Walpole, M., et al.: Earth observation for biodiversity monitoring: A review of current approaches and future opportunities for tracking progress towards the aichi biodiversity targets. CBD technical series (2014) 1 +81. Shermeyer, J., Hogan, D., Brown, J., Van Etten, A., Weir, N., Pacifici, F., Hansch, R., Bastidas, A., Soenen, S., Bacastow, T., et al.: Spacenet 6: Multi-sensor all weather mapping dataset. In: CVPR Workshop EarthVision (2020) 3 +82. Shukor, M., Dancette, C., Rame, A., Cord, M.: UnIVAL: Unified model for image, video, audio and language tasks. TMLR (2023) 1, 2, 3 +83. Skidmore, A.K., Coops, N.C., Neinavaz, E., Ali, A., Schaepman, M.E., Paganini, M., Kissling, W.D., Vihervaara, P., Darvishzadeh, R., Feilhauer, H., et al.: Priority list of biodiversity metrics to observe from space. Nature Ecology & Evolution (2021) 1 +84. Srivastava, S., Sharma, G.: OmniVec: Learning robust representations with cross modal sharing. In: WACV (2024) 2, 3 +85. Sudmanns, M., Tiede, D., Augustin, H., Lang, S.: Assessing global Sentinel-2 coverage dynamics and data availability for operational Earth observation (EO) applications using the EO-Compass. International Journal of Digital Earth (2019) 6 +86. Sumbul, G., De Wall, A., Kreuziger, T., Marcelino, F., Costa, H., Benevides, P., Caetano, M., Demir, B., Markl, V.: BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine (2021) 3, 4 +87. Tarasiou, M., Chavez, E., Zafeiriou, S.: ViTs for SITS: Vision transformers for satellite image time series. In: CVPR (2023) 3 +88. Tseng, G., Zvonkov, I., Purohit, M., Rolnick, D., Kerner, H.: Lightweight, pre-trained transformers for remote sensing timeseries. arXiv preprint arXiv:2304.14065 (2023) 3, 10, 11 +89. Tseng, W.H., Lé, H.A., Boulch, A., Lefèvre, S., Tiede, D.: CROCO: Cross-modal contrastive learning for localization of Earth observation data. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 2 +90. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3 + +91. Vrieling, A., Meroni, M., Darvishzadeh, R., Skidmore, A.K., Wang, T., Zurita-Milla, R., Oosterbeek, K., O'Connor, B., Paganini, M.: Vegetation phenology from Sentinel-2 and field cameras for a Dutch barrier island. Remote sensing of environment (2018) 4, 9 +92. Wang, Y., Graham, N.A.A., Xiong, Z., Liu, C., Albrecht, C.M., Zhu, X.X.: SSL4EO-S12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in Earth observation. IEEE Geoscience and Remote Sensing Magazine (2023) 3 +93. Wenger, R., Puisant, A., Weber, J., Idoumghar, L., Forestier, G.: MultiSenGE: A multimodal and multitemporal benchmark dataset for land use/land cover remote sensing applications. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 3 +94. Wu, K., Peng, H., Chen, M., Fu, J., Chao, H.: Rethinking and improving relative position encoding for vision transformer. In: ICCV (2021) 6 +95. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMim: A simple framework for masked image modeling. In: CVPR (2022) 2 +96. Xiong, Z., Wang, Y., Zhang, F., Stewart, A.J., Hanna, J., Borth, D., Papoutsis, I., Saux, B.L., Camps-Valls, G., Zhu, X.X.: Neural plasticity-inspired foundation model for observing the Earth crossing modalities. arXiv preprint arXiv:2403.15356 (2024) 3, 4, 11, 12 +97. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change (2013) 1 +98. Yang, M.Y., Landrieu, L., Tuia, D., Toth, C.: Multi-modal learning in photogrammetry and remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing (2021) 3 +99. Yuan, Y., Lin, L., Liu, Q., Hang, R., Zhou, Z.G.: SITS-Former: A pre-trained spatio-spectral-temporal representation model for sentinel-2 time series classification. International Journal of Applied Earth Observation and Geoinformation (2022) 2 +100. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016) 2 +101. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: Image BERT pre-training with online tokenizer. In: ICLR (2022) 3 +102. Zong, Y., Mac Aodha, O., Hospedales, T.: Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008 (2023) 1 \ No newline at end of file diff --git a/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/images.zip b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba806df3ac420b9f5d3976da4704fdc3136f930d --- /dev/null +++ b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:221cb01d366ce720f4398255c9bb711b12ba85bf2503589a946c1400084c236f +size 551820 diff --git a/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/layout.json b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6526e82e56fb6296632fbd60eaba880184467ef8 --- /dev/null +++ b/2024/OmniSat_ Self-Supervised Modality Fusion for Earth Observation/layout.json @@ -0,0 +1,13630 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 174, + 111, + 440, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 111, + 440, + 145 + ], + "spans": [ + { + "bbox": [ + 174, + 111, + 440, + 145 + ], + "type": "text", + "content": "OmniSat: Self-Supervised Modality Fusion for Earth Observation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "spans": [ + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "text", + "content": "Guillaume Astruc " + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "inline_equation", + "content": "^{1,3,4}" + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "text", + "content": ", Nicolas Gonthier " + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "text", + "content": ", Clement Mallet " + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "text", + "content": ", and Loic Landrieu " + }, + { + "bbox": [ + 211, + 167, + 402, + 192 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 154, + 201, + 459, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 201, + 459, + 224 + ], + "spans": [ + { + "bbox": [ + 154, + 201, + 459, + 224 + ], + "type": "text", + "content": "1 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France \n2 IGN, France \n3 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France \n4 CNES, France" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 256, + 455, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 455, + 410 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 455, + 410 + ], + "type": "text", + "content": "Abstract. The diversity and complementarity of sensors available for Earth Observations (EO) calls for developing bespoke self-supervised multimodal learning approaches. However, current multimodal EO datasets and models typically focus on a single data type, either mono-date images or time series, which limits their impact. To address this issue, we introduce OmniSat, a novel architecture able to merge diverse EO modalities into expressive features without labels by exploiting their alignment. To demonstrate the advantages of our approach, we create two new multimodal datasets by augmenting existing ones with new modalities. As demonstrated for three downstream tasks—forestry, land cover classification, and crop mapping—OmniSat can learn rich representations without supervision, leading to state-of-the-art performances in semi- and fully supervised settings. Furthermore, our multimodal pretraining scheme improves performance even when only one modality is available for inference. The code and dataset are available at https://github.com/gastruc/OmniSat." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 420, + 428, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 420, + 428, + 432 + ], + "spans": [ + { + "bbox": [ + 160, + 420, + 428, + 432 + ], + "type": "text", + "content": "Keywords: Earth observation " + }, + { + "bbox": [ + 160, + 420, + 428, + 432 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 420, + 428, + 432 + ], + "type": "text", + "content": " Multi-modality " + }, + { + "bbox": [ + 160, + 420, + 428, + 432 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 420, + 428, + 432 + ], + "type": "text", + "content": " Self-supervised learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 452, + 218, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 452, + 218, + 464 + ], + "spans": [ + { + "bbox": [ + 133, + 452, + 218, + 464 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 476, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 571 + ], + "type": "text", + "content": "Self-supervised multimodal learning has recently gathered significant interest within computer vision [38, 82, 102]. Earth Observation (EO) is particularly well-suited for developing and evaluating such approaches [29, 51], thanks to the large amount of open-access data captured by sensing technologies with complementary capabilities [36, 79]. Combining different sources of EO observations is crucial for several high-impact applications, including environmental [21, 80, 83] and climate monitoring [56, 97], as well as improving food security [66]. Moreover, learning with few or no labels is essential for developing regions with limited data annotation capabilities [5, 55, 62]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 571, + 482, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 642 + ], + "type": "text", + "content": "Despite this potential, most multimodal EO datasets and models focus on a single data type, either mono-date images or time series. This limitation prevents them from simultaneously leveraging the spatial resolution of aerial images [57, 64], the temporal and spectral resolutions of optical satellite time series [26], and the resilience of radar to weather effects [4, 65]. Additionally, existing approaches are often limited for a given set of sensors, limiting their applicability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 642, + 481, + 665 + ], + "type": "text", + "content": "To address these challenges, we introduce OmniSat, a novel architecture designed for the self-supervised fusion of diverse EO data. Existing multimodal approaches often map" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 163 + ], + "type": "text", + "content": "multiple unrelated observations from different modalities to one pivot modality [38, 82] or a shared latent space [39, 84]. In contrast, OmniSat merges multiple views of the same area from different modalities into a single representation combining the specific information of each modality [14, 41, 73]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 481, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 481, + 233 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 481, + 233 + ], + "type": "text", + "content": "In computer vision, obtaining finely aligned multimodal observations generally requires specialized sensors [54, 59, 67] or the computation of complex mappings between modalities [23, 75]. On the other hand, EO data can be naturally aligned with georeferencing. To leverage this property, we adapt multimodal contrastive learning [49, 72] and cross-modal masked auto-encoding techniques [43] to learn rich multimodal EO representations with a generalist fusion scheme and without annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 233, + 481, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 481, + 292 + ], + "type": "text", + "content": "To address the scarcity of EO datasets with a diverse range of heterogeneous modalities (see Tab. 1), we enrich the TreeSatAI [2] and PASTIS-R [33, 34] datasets with new aligned modalities. This allows us to evaluate OmniSat's ability to handle an arbitrary number of inputs with varying natures and resolutions. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 293, + 480, + 399 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 293, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 479, + 327 + ], + "type": "text", + "content": "- We introduce OmniSat, a new model that learns to combine varied sources of EO observations in a self-supervised manner, resulting in richer joint representations that capture the unique characteristics of each modality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 329, + 480, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 480, + 351 + ], + "type": "text", + "content": "- We augment two EO benchmarks to create the first datasets with three modalities of different natures (very high resolution images, optical and SAR time series)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 353, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 399 + ], + "type": "text", + "content": "- We demonstrate that OmniSat can leverage diverse modalities to learn rich representations, establishing new states-of-the-art for tree species, crop type, and land cover classification. Furthermore, our cross modal self-supervised training scheme improves performance even when only one modality is available during inference." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 415, + 224, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 224, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 224, + 427 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 435, + 481, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 471 + ], + "type": "text", + "content": "This section provides an overview of self-supervised and multimodal learning, emphasizing the specificities of their usage for Earth observation. Lastly, we highlight the scarcity of multimodal EO datasets with diverse data types." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 476, + 481, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 481, + 536 + ], + "type": "text", + "content": "Self-Supervised Learning. This technique consists in learning expressive data representations without labels by using a pretext task. This approach has been particularly successful for natural language [52] and image [70] analysis. Initially focused on discriminative tasks [37, 68, 100], recent self-supervised approaches for images can be categorized as contrastive or generative." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 536, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 606 + ], + "type": "text", + "content": "Contrastive methods minimize the distance between representations of paired samples, often the same image under different transformations, and maximize the distance with other samples [15, 17, 46]. More efficient methods only consider positive samples and avoid mode collapse by introducing various asymmetries [18, 42] or normalization [16]. Such approaches have been successfully adapted to EO, for which samples are paired according to their location [89] or time of acquisition [7, 63]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "content": "Generative methods reason at the level of individual token—a small portion of the input, typically a patch for images [25]. The objective is to reconstruct the masked tokens of an input image in pixel [10,45,95] or feature space [6]. This principle has been successfully adapted to EO analysis [20, 30, 99], and was further extended to handle multiple spatial scales [74], multimodality [29, 51], or hyperspectral observations [50, 60]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "G. Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 155, + 478, + 394 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Publicly Available Multimodal EO Datasets. We provide in parenthesis the spatial resolutions of the single-date images and labels, and the temporal resolutions of time series. S1/S2 denotes Sentinel-1 and 2. " + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": ": modalities added in this work." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 155, + 478, + 394 + ], + "lines": [ + { + "bbox": [ + 135, + 155, + 478, + 394 + ], + "spans": [ + { + "bbox": [ + 135, + 155, + 478, + 394 + ], + "type": "table", + "html": "
DatasetModalitiesLabels
images (single date)time series
SpaceNet6 [81]SAR+optical (0.5m-2m)Xbuilding footprint (<1m)
TreeSatAI [2]aerial + S1/S2 (0.2-10m)Xforestry (60m)
BigEarthNet [86]S1/S2 (10m)Xland cover (100m)
DFC20 [76]S1/S2 (10m)Xland cover (500m)
MDAS [48]S1/S2 + hyperspectral (2.2-10m)Xland cover (0.25m)
DOFA [96]NAIP + Gaofen + S1/S2 + EnMAP (1-30m)XX
PASTIS-R [33, 34]XS1/S2 (30-140 / year)agriculture (10m)
SSL4EO-S12 [92]XS1/S2 (4 / year)X
DFC21-DSE [61]XS1/S2 + LS8 (3-9/year)human activity (500m)
MapInWild [28]XS1/S2 (4 / years)protected areas (10m)
SEN12MS-CR-TS [27]XS1/S2 (30 / years)cloud cover (10m)
MultiSenGE [93]XS1/S2 (30-140 / years)land cover (10m)
FLAIR [31]aerial (0.2m)S2 (20-114 / year)land cover (0.2m)
Satlas [12]NAIP (0.5 -2m)S2 (8-12 / year)various
PASTIS-HD★ SPOT 6-7 (1.5m)S1/S2 (30-140 / year)agriculture (10m)
TreeSatAI-TSaerial (0.2m)★ S1/S2 (10-70 / year)forestry (60m)
", + "image_path": "e2dc96a91de8034a8210f663d9facd29e2169cbbfd9fa04b941da5ac5a6fa868.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "content": "Several hybrid approaches combine the discriminative power of contrastive methods and the scalability of generative objectives for natural images [70, 101] and EO data [29]. Our proposed OmniSat model also implements both mechanisms. A key feature is that we leverage the precise alignment between different sources of EO data to contrastively match small patches of different modalities rather than entire images or time series." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "content": "Self-Supervised Multimodal Learning. Multimodal computer vision has received a lot of interest [13], notably due to the success of cross-modal pre-training [72]. Recent models align the embeddings of heterogeneous modalities such as video and sound [49], depth and images [44], text and image [3, 9], or multiple combinations of these modalities [38, 39, 82, 84]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Multimodal learning also has a long history in EO [58, 71, 98] due to the large variety and complementarity of sensors [36, 79]. However, recent transformer-based architectures [90] for EO are often limited to one type of modality, be it a single image [20, 74] or time-series [34, 87]. For example, CROMA [29] and PRESTO [88] are specifically designed for paired optical and radar observations, but cannot handle Very High Resolution (VHR) data. USat [51] considers images with different resolutions, but only takes a single date within a time series. UT&T [31] can natively take single and multi-date observations of different modalities, but cannot be easily pre-trained in a self-supervised manner since it relies on convolutions and an ad-hoc late fusion scheme." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 114, + 235, + 193 + ], + "blocks": [ + { + "bbox": [ + 162, + 114, + 235, + 193 + ], + "lines": [ + { + "bbox": [ + 162, + 114, + 235, + 193 + ], + "spans": [ + { + "bbox": [ + 162, + 114, + 235, + 193 + ], + "type": "image", + "image_path": "adaafb273b142fd66f620e4b304cd2da63772604f023462796b53e942b0fafe0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 194, + 230, + 201 + ], + "lines": [ + { + "bbox": [ + 167, + 194, + 230, + 201 + ], + "spans": [ + { + "bbox": [ + 167, + 194, + 230, + 201 + ], + "type": "text", + "content": "Sentinel-2 time series" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 203, + 226, + 256 + ], + "blocks": [ + { + "bbox": [ + 171, + 203, + 226, + 256 + ], + "lines": [ + { + "bbox": [ + 171, + 203, + 226, + 256 + ], + "spans": [ + { + "bbox": [ + 171, + 203, + 226, + 256 + ], + "type": "image", + "image_path": "b99d8f55146e578a2c79a563fcf0efaa3b12e81db16e33da5a4225569c1e9431.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 259, + 214, + 267 + ], + "lines": [ + { + "bbox": [ + 182, + 259, + 214, + 267 + ], + "spans": [ + { + "bbox": [ + 182, + 259, + 214, + 267 + ], + "type": "text", + "content": "(a) FLAIR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 114, + 356, + 193 + ], + "blocks": [ + { + "bbox": [ + 253, + 114, + 356, + 193 + ], + "lines": [ + { + "bbox": [ + 253, + 114, + 356, + 193 + ], + "spans": [ + { + "bbox": [ + 253, + 114, + 356, + 193 + ], + "type": "image", + "image_path": "9e7dbf1494e3f9941f7ef4e45d4678844922fdf9001530804a691d96a677f73f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 194, + 341, + 201 + ], + "lines": [ + { + "bbox": [ + 270, + 194, + 341, + 201 + ], + "spans": [ + { + "bbox": [ + 270, + 194, + 341, + 201 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 270, + 194, + 341, + 201 + ], + "type": "text", + "content": " Sentinel-2 time series" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 255, + 202, + 356, + 224 + ], + "blocks": [ + { + "bbox": [ + 255, + 202, + 356, + 224 + ], + "lines": [ + { + "bbox": [ + 255, + 202, + 356, + 224 + ], + "spans": [ + { + "bbox": [ + 255, + 202, + 356, + 224 + ], + "type": "image", + "image_path": "27a40a7871f836a6b6bcc40bf98238db9b90ba694fdfaa2296a4ca4f295c76c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 255, + 232, + 355, + 255 + ], + "blocks": [ + { + "bbox": [ + 270, + 225, + 341, + 232 + ], + "lines": [ + { + "bbox": [ + 270, + 225, + 341, + 232 + ], + "spans": [ + { + "bbox": [ + 270, + 225, + 341, + 232 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 270, + 225, + 341, + 232 + ], + "type": "text", + "content": " Sentinel-1 time series" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 232, + 355, + 255 + ], + "lines": [ + { + "bbox": [ + 255, + 232, + 355, + 255 + ], + "spans": [ + { + "bbox": [ + 255, + 232, + 355, + 255 + ], + "type": "image", + "image_path": "49b1ea627ffa945a279abb165733b2c50fb82deed2f86384c6eba40fc2666e62.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 258, + 332, + 267 + ], + "lines": [ + { + "bbox": [ + 278, + 258, + 332, + 267 + ], + "spans": [ + { + "bbox": [ + 278, + 258, + 332, + 267 + ], + "type": "text", + "content": "(b) TreeSatAI-TS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 274, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 131, + 274, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 482, + 318 + ], + "type": "text", + "content": "Fig. 1: Datasets. We represent three tiles from the considered multilabel classification datasets: FLAIR (a), TreeSatAI-TS (b) and PASTIS-HD (c). TreeSatAI-TS is a new dataset built by replacing the single-date Sentinel-1 and 2 images of TreeSatAI [2] by year-long time series. PASTIS-HD (c) adds VHR satellite images to PASTIS-R [34]. " + }, + { + "bbox": [ + 131, + 274, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 131, + 274, + 482, + 318 + ], + "type": "text", + "content": ": modalities added in this work." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 376, + 122, + 447, + 193 + ], + "blocks": [ + { + "bbox": [ + 378, + 114, + 445, + 121 + ], + "lines": [ + { + "bbox": [ + 378, + 114, + 445, + 121 + ], + "spans": [ + { + "bbox": [ + 378, + 114, + 445, + 121 + ], + "type": "text", + "content": "VHR satellite " + }, + { + "bbox": [ + 378, + 114, + 445, + 121 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 122, + 447, + 193 + ], + "lines": [ + { + "bbox": [ + 376, + 122, + 447, + 193 + ], + "spans": [ + { + "bbox": [ + 376, + 122, + 447, + 193 + ], + "type": "image", + "image_path": "5fc81623380a7f0c94f69f6c4145dede360af252f84a0be6a46016463dfbbbc7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 365, + 201, + 455, + 224 + ], + "blocks": [ + { + "bbox": [ + 367, + 194, + 456, + 201 + ], + "lines": [ + { + "bbox": [ + 367, + 194, + 456, + 201 + ], + "spans": [ + { + "bbox": [ + 367, + 194, + 456, + 201 + ], + "type": "text", + "content": "PASTIS: Sentinel-2 time series" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 201, + 455, + 224 + ], + "lines": [ + { + "bbox": [ + 365, + 201, + 455, + 224 + ], + "spans": [ + { + "bbox": [ + 365, + 201, + 455, + 224 + ], + "type": "image", + "image_path": "0a84d509b9e4e6c42eefd7bc8c1cf3ae9129dad99db837392e258848538730ab.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 367, + 232, + 456, + 255 + ], + "blocks": [ + { + "bbox": [ + 364, + 225, + 459, + 232 + ], + "lines": [ + { + "bbox": [ + 364, + 225, + 459, + 232 + ], + "spans": [ + { + "bbox": [ + 364, + 225, + 459, + 232 + ], + "type": "text", + "content": "PASTIS-R: Sentinel-1 time series" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 232, + 456, + 255 + ], + "lines": [ + { + "bbox": [ + 367, + 232, + 456, + 255 + ], + "spans": [ + { + "bbox": [ + 367, + 232, + 456, + 255 + ], + "type": "image", + "image_path": "95ac3c5edf81174e746a0aa5ea123991c7d2d45a2b0049f594b7786afda33ee9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 258, + 435, + 267 + ], + "lines": [ + { + "bbox": [ + 388, + 258, + 435, + 267 + ], + "spans": [ + { + "bbox": [ + 388, + 258, + 435, + 267 + ], + "type": "text", + "content": "(c) PASTIS-HD" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "type": "text", + "content": "Multimodal EO Datasets. As reported in Table 1, many multimodal EO datasets use Sentinel-1 [11] and 2 [26] data for applications ranging from land cover to forestry analysis and fire detection. We also note that most multimodal datasets only contain data of one type: mono-date image or time series. Several datasets (BigEarthNet [86], DFC20 [76], MDAS [48]) select a single date from time series. However, single Sentinel-1 and 2 acquisitions can be significantly affected by rain and cloud cover, respectively. Furthermore, capturing the temporal dynamics is crucial to characterize the phenology of vegetation [91]," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "text", + "content": "FLAIR [31] is the first multimodal EO dataset to propose both very high spatial resolution " + }, + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "inline_equation", + "content": "(\\leq 2\\mathrm{m})" + }, + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "text", + "content": " and high temporal resolution " + }, + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "inline_equation", + "content": "(>4" + }, + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "text", + "content": " images/year). Satlas [12] combines Sentinel-2 time series and for " + }, + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 436, + 482, + 518 + ], + "type": "text", + "content": " to tiles (continental US), very high definition NAIP images. The functional map of the World [19] integrates observations from various sensors, but most areas are only observed with one sensor. Two other datasets contain time series and single images from multiple sources, but were not available at the time of writing this article: IARPA-SMART [40] and DOFA [96]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 518, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 601 + ], + "type": "text", + "content": "To showcase how OmniSat can consume an arbitrary number of modalities with different spatial, spectral, and temporal resolutions, we selected two commonly used EO benchmarks, TreeSatAI [2] and PASTIS-R [34], whose focus on crop type mapping and forestry differs from the land cover analysis of FLAIR. We added new modalities to these datasets to reach three distinct data types: VHR aerial images, optical time series, and SAR time series. See Fig. 1 for an illustration, and Sec. 4.1 for more details on how we extended these datasets." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 131, + 618, + 194, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 618, + 194, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 618, + 194, + 630 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "We consider a tile " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " observed through a set " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " distinct sensors or modalities. The goal of the OmniSat model is to learn in a self-supervised fashion to combine" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "G.Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 116, + 465, + 305 + ], + "blocks": [ + { + "bbox": [ + 147, + 116, + 465, + 305 + ], + "lines": [ + { + "bbox": [ + 147, + 116, + 465, + 305 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 465, + 305 + ], + "type": "image", + "image_path": "a4ab131949202b7eea26d3f6ad97cfa1a4fd862e9f1aef86144ee95446301553.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": "Fig. 2: OmniSat Architecture. We illustrate OmniSat for " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " modalities, and a tile split into " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "P = 4" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " patches. The " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "M \\times P" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " input tokens " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "x_{\\mathbf{P}}^{\\mathbf{M}}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " are encoded by " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " modality-specific encoders " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{\\mathrm{M}}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": ", yielding the token representations " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\mathbf{M}}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": ". The module " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " combines them into multimodal patch representations " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\star}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": ". The token embeddings " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\mathbf{M}}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " are supervised by a contrastive cross-modal objective. We also use a reconstruction objective: the masked multimodal representations " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\star}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " are decoded by modality-specific networks " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{M}}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": " to reconstruct their corresponding inputs in " + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "inline_equation", + "content": "x_{\\mathbf{P}}^{\\mathbf{M}}" + }, + { + "bbox": [ + 131, + 310, + 482, + 377 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "text", + "content": "all modalities " + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "text", + "content": " into a multimodal representation " + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "inline_equation", + "content": "f^{\\star}" + }, + { + "bbox": [ + 130, + 404, + 482, + 464 + ], + "type": "text", + "content": ". We first provide details about OmniSat's architecture in Sec. 3.1. We then explain our our training scheme, which consists of a cross-modal contrastive objective (Sec. 3.2) and a multimodal masked encoding task (Sec. 3.3). Finally, we present the implementation details in Sec. 3.4. The overall method is represented in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 490, + 212, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 212, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 212, + 501 + ], + "type": "text", + "content": "3.1 Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 517, + 481, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 481, + 541 + ], + "type": "text", + "content": "This section presents the tokenization process, the structures of the encoder and decoder for each modality, and the architecture of the modality combiner network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": "Multimodal Tokenization. All available modalities are spatially aligned through georeferencing. This allows us to divide the tile " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " into a set " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " non-overlapping patches consistently across all modalities: " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "x_{p}^{\\mathbf{M}} = \\{x_{p}^{m}\\}_{m\\in \\mathbf{M}}" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " corresponds to " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " distinct views of the same patch " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " with different modalities. Each modality " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " takes its values in a space " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\varOmega^m" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "x_{p}^{m}\\in \\varOmega^{m}" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": ". We index tokens with pairs " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "(m,p)" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": ", defined for each modality " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " and patch " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": ", for a total of " + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "inline_equation", + "content": "M\\times P" + }, + { + "bbox": [ + 130, + 547, + 482, + 617 + ], + "type": "text", + "content": " tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Time series from Sentinel satellites may experience registration errors spanning several meters, complicating their precise alignment with high-resolution imagery. However, using temporal sequences of satellite data mitigates these errors as aggregation over time tends to balance out misalignments." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 448, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 118, + 233, + 243 + ], + "blocks": [ + { + "bbox": [ + 143, + 118, + 233, + 243 + ], + "lines": [ + { + "bbox": [ + 143, + 118, + 233, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 118, + 233, + 243 + ], + "type": "image", + "image_path": "6d67a4b45ff08d54139e398482e768f27ddaa24c13bd2614d0ef3f752fdd9cea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 245, + 236, + 255 + ], + "lines": [ + { + "bbox": [ + 137, + 245, + 236, + 255 + ], + "spans": [ + { + "bbox": [ + 137, + 245, + 236, + 255 + ], + "type": "text", + "content": "(a) Image Encoder and Decoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 117, + 335, + 245 + ], + "blocks": [ + { + "bbox": [ + 248, + 117, + 335, + 245 + ], + "lines": [ + { + "bbox": [ + 248, + 117, + 335, + 245 + ], + "spans": [ + { + "bbox": [ + 248, + 117, + 335, + 245 + ], + "type": "image", + "image_path": "aeca5451ce9232f54390e78aabf62bcc3a8cbfe918721b1c3e96a2952cef638d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 246, + 343, + 255 + ], + "lines": [ + { + "bbox": [ + 239, + 246, + 343, + 255 + ], + "spans": [ + { + "bbox": [ + 239, + 246, + 343, + 255 + ], + "type": "text", + "content": "(b) Temporal Encoder and Decoder." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 348, + 115, + 477, + 243 + ], + "blocks": [ + { + "bbox": [ + 348, + 115, + 477, + 243 + ], + "lines": [ + { + "bbox": [ + 348, + 115, + 477, + 243 + ], + "spans": [ + { + "bbox": [ + 348, + 115, + 477, + 243 + ], + "type": "image", + "image_path": "38032c22751dbd4962ef7b38027372d7c4e3f01048f4ac3ab6fcb3aa51de3663.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 245, + 463, + 255 + ], + "lines": [ + { + "bbox": [ + 361, + 245, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 361, + 245, + 463, + 255 + ], + "type": "text", + "content": "(c) Modality Combining Network." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "text", + "content": "Fig. 3: OmniSat Architecture. OmniSat is composed of dedicated patch encoders for image (a) and time series b, here represented for a length of " + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "text", + "content": " time stamps. The modality combining module " + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "text", + "content": " is depicted in (c) with " + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "inline_equation", + "content": "P = 2" + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 130, + 258, + 482, + 302 + ], + "type": "text", + "content": ". Elements colored in orange are learned networks or parameters." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "text", + "content": "Encoder-Decoder for Images. We split image tiles split into small square patches: " + }, + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\varOmega^{\\mathrm{img}}=\\mathbb{R}^{C\\times W\\times W}" + }, + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "text", + "content": " the size of the patches in pixels and " + }, + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 332, + 482, + 441 + ], + "type": "text", + "content": " the number of channels. As shown in Fig. 3a, we encode these inputs with a sequence of convolutions and max-pool layers until the spatial dimension is fully collapsed. Decoding involves a symmetric sequence of convolutions and un-pooling layers. Contrary to existing masked auto-encoders, we pass the pooling indices from the encoder's max-pooling to the decoder's un-pooling in the manner of SegNet [8]. This dispenses the encoder from learning the intra-patch spatial configuration. This allows the image encoder to focus on the radiometric information, which may be more relevant depending on the application." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "content": "Encoder-Decoder for Time Series. Each temporal patch is represented by " + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "content": " sequential observations with " + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "content": " channels: " + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\Omega^{\\mathrm{TS}} = \\mathbb{R}^{C\\times L}" + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "content": ", each associated with a time stamp. We encode the temporal patches using a Lightweight Temporal Attention Encoder (LTAE) model [32], an efficient network for geospatial time series processing. We decode vector representations into time series by repeating the vector " + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "content": " times across the temporal dimension, adding a temporal encoding for each time step, and using an MLP to map the results to size " + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 445, + 482, + 529 + ], + "type": "text", + "content": ". See Fig. 3b for an illustration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 529, + 483, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 483, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 483, + 612 + ], + "type": "text", + "content": "Optical time series are notoriously affected by clouds [85]. This may affect the validity of the reconstruction task: the decoder cannot know which observations are cloudy, making the reconstruction objective unpredictable. To circumvent this issue, we use the temporal attention maps of the encoder's LTAE to select dates to reconstruct: cloudless observations are more informative and should have a higher attention score [78]. We only consider in the reconstruction loss " + }, + { + "bbox": [ + 130, + 529, + 483, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reconstr}}" + }, + { + "bbox": [ + 130, + 529, + 483, + 612 + ], + "type": "text", + "content": " the top " + }, + { + "bbox": [ + 130, + 529, + 483, + 612 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 529, + 483, + 612 + ], + "type": "text", + "content": " dates in terms of the LTAE's attention maps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Modality Combining Network. The modality combining network " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ", represented in Fig. 3c, takes the " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "M \\times P" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " token embeddings " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\mathbf{M}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ", some of whom can potentially be masked. We equip each token with a Euclidean relative positional encoding [94], calculated based on their patch's position " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{r(p,q) \\mid (p,q) \\in \\mathbf{P}^2\\}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ", allowing each token to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "G.Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "selectively consider its spatial surroundings. As most EO data are captured from above (satellite or aerial), their distribution is invariant by horizontal translation, making this choice of encoding preferable to an absolute position encoding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": "The modality combining module " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " starts with a series of " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " residual self-attention blocks connecting all tokens across modality. We then perform cross-attention between the resulting token embeddings " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "g_{\\mathbf{P}}^{\\mathbf{M}} \\in \\mathbb{R}^{d \\times M \\times P}" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " copies " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\mathrm{comb}}" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " of a modality combining token " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{comb}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " learned as a free parameter. Each copy of " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "f_{p}^{\\mathrm{comb}}" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " is spatially located at the patch " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " for the relative positional encoding " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": ". The module " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " outputs " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " multimodal encodings " + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\star}" + }, + { + "bbox": [ + 130, + 151, + 480, + 223 + ], + "type": "text", + "content": " combining all available modalities for each patch:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 230, + 482, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 230, + 482, + 244 + ], + "spans": [ + { + "bbox": [ + 227, + 230, + 482, + 244 + ], + "type": "interline_equation", + "content": "g _ {\\mathbf {P}} ^ {\\mathbf {M}} = \\text {s e l f - a t t e n t i o n} \\left(f _ {\\mathbf {P}} ^ {\\mathbf {M}}; r\\right) \\tag {1}", + "image_path": "6de589fe4cb6c0fa71707fb2eb8ee9f27b666764a17de169bd10a2d0e0924f31.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 247, + 481, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 247, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 231, + 247, + 481, + 262 + ], + "type": "interline_equation", + "content": "f _ {\\mathbf {P}} ^ {f \\star} = \\text {c r o s s - a t t e n t i o n} \\left(f _ {\\mathbf {P}} ^ {\\text {c o m b}}, g _ {\\mathbf {P}} ^ {\\mathbf {M}}; r\\right). \\tag {2}", + "image_path": "486f51104b2eb31737694ea9586cff5648ccb808658953751f46681314aaa241.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 276, + 250, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 276, + 250, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 250, + 288 + ], + "type": "text", + "content": "3.2 Contrastive Objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "content": "We denote by " + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "inline_equation", + "content": "f_{p}^{m}" + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "content": "-dimensional encodings of the input patch " + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "inline_equation", + "content": "x_{p}^{m}" + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "content": " given by their dedicated encoders. We propose to supervise the embeddings " + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "inline_equation", + "content": "f_{p}^{m}" + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "content": " with a contrastive objective encouraging spatial consistency across modalities. Indeed, while each modality captures distinct characteristics of " + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "content": ", all encodings " + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "inline_equation", + "content": "f_{p}^{m}" + }, + { + "bbox": [ + 130, + 294, + 480, + 353 + ], + "type": "text", + "content": " share the same latent variable: the semantic content of the patch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": "In practice, we want " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "f_{p}^{m}" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " to be closer to " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "f_{p}^{n}" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "n \\neq m" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": ", than to " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "f_{q}^{n}" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " for other patches " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "q \\neq p" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " as the set of patches within the current batch of observations. We adapt the classic InfoNCE loss [69] to our setting with two main differences, illustrated in Fig. 4. (i) Each token " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "(m,p)" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " positive matches: the tokens corresponding to the same patch " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " but viewed in another modality " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "n \\neq m" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": "; and (ii) as EO observations are generally spatially regular, nearby patches may be visually indistinguishable. Therefore, we exclude from the negative matches of " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "(m,p)" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " all tokens in modality " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " and which are too close to " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": ". To this end, we remove the set " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "T(m,p)" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " of tokens with modality " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " and whose patches are in the same tile as " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": ". Our loss function " + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrast}}" + }, + { + "bbox": [ + 130, + 353, + 481, + 460 + ], + "type": "text", + "content": " is defined as such:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 468, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 468, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 141, + 468, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t r a s t}} = \\frac {1}{M | \\mathbf {B} |} \\sum_ {(m, p) \\in \\mathbf {M} \\times \\mathbf {B}} \\log \\left(\\frac {\\sum_ {n \\neq m} \\exp \\left(\\left\\langle f _ {p} ^ {m} , f _ {p} ^ {n} \\right\\rangle / \\gamma\\right)}{\\sum_ {(n , q) \\in \\mathbf {M} \\times \\mathbf {B} \\backslash T (m , p)} \\exp \\left(\\left\\langle f _ {p} ^ {m} , f _ {q} ^ {n} \\right\\rangle / \\gamma\\right)}\\right), \\tag {3}", + "image_path": "269f535f851fb1443908e5da368176f1cbff2c47fe4fe8ac3700f813d51bbd5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " a temperature parameter, and " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " the scalar product in " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": ". This function, specifically designed for geospatial data, allows us to contrast individual patches across modalities, which is not typically feasible for natural images. However, as the contrastive objective aligns multimodal representations, the patch encoders may be encouraged to overlook the distinct attributes of their respective modality. Instead, they may focus only on features shared by all modalities, i.e., their common denominator. To ensure that encoders also capture modality-specific information, we incorporate a reconstruction objective, detailed in Sec. 3.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 623, + 318, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 318, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 318, + 635 + ], + "type": "text", + "content": "3.3 Multimodal Reconstruction Objective" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "During training, we mask a fraction of tokens " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\subset \\mathbf{M} \\times \\mathbf{P}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " and replace their embeddings with a learned vector " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{mask}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ". Note that the masking can differ across modalities," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 291, + 268 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 291, + 268 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 291, + 268 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 291, + 268 + ], + "type": "image", + "image_path": "509f9339c2eb39b7a94dc06dc20da0ac9a4c337a553e964b614eddf521e8846d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "lines": [ + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": "Fig.4:Contrastive Loss. We represent the token matching matrix for two tiles " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "\\mathrm{Tile}_1" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "\\mathrm{Tile}_2" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " viewed across 3 modalities " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "m_{1}" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "m_2" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "m_3" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " Tile is composed of the patches " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " ,while " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "\\mathrm{Tile}_2" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "q_{1}" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "inline_equation", + "content": "q_{2}" + }, + { + "bbox": [ + 302, + 116, + 482, + 225 + ], + "type": "text", + "content": " .In contrast to classic approaches which ignore the diagonal and assign each sample with a single positive match, our loss defines operates at the patch level, considers multiple positives per token, and excludes tokens in a block-diagonal fashion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 229, + 429, + 265 + ], + "lines": [ + { + "bbox": [ + 355, + 229, + 429, + 265 + ], + "spans": [ + { + "bbox": [ + 355, + 229, + 429, + 265 + ], + "type": "text", + "content": "+ positive match\n- negative match\no ignored" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "text", + "content": "and some patches may be entirely masked. All tokens are then processed by the modality combining network " + }, + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "text", + "content": ", which outputs " + }, + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "text", + "content": " multimodal embeddings " + }, + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{\\star}" + }, + { + "bbox": [ + 131, + 289, + 482, + 314 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 321, + 482, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 321, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 214, + 321, + 482, + 338 + ], + "type": "interline_equation", + "content": "f _ {\\mathbf {P}} ^ {\\star} = \\mathcal {C} \\left(\\left\\{f _ {p} ^ {m} \\right\\} _ {(m, p) \\notin \\mathbf {K}} \\cup \\left\\{f ^ {\\text {m a s k}} \\right\\} _ {(m, p) \\in \\mathbf {K}}\\right). \\tag {4}", + "image_path": "e24db91b5a072a6ba64512b6e642073251c28e47e716466f0fbed4146ad1c726.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "spans": [ + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "text", + "content": "To encourage the patch embeddings " + }, + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{P}}^{*}" + }, + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "text", + "content": " to capture information from all modalities, we build a multimodal reconstruction objective. We denote by " + }, + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^m: \\mathbb{R}^d \\mapsto \\Omega^m" + }, + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "text", + "content": " the dedicated decoder of each modality " + }, + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 345, + 483, + 381 + ], + "type": "text", + "content": " and write the reconstruction loss as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 194, + 388, + 482, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 388, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 194, + 388, + 482, + 419 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n s t r}} = \\frac {1}{| \\mathbf {K} |} \\sum_ {(m, p) \\in \\mathbf {K}} \\frac {1}{\\dim (\\Omega^ {m})} \\| \\mathcal {D} ^ {m} \\left(f _ {p} ^ {\\star}\\right) - x _ {p} ^ {m} \\| ^ {2}, \\tag {5}", + "image_path": "4a7afc5465fda6a4471f6fb309107bb81481973e99753f9309d842e8de3f5231.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "inline_equation", + "content": "\\dim (\\varOmega^m)" + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "text", + "content": " the dimension of " + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "inline_equation", + "content": "\\varOmega^m" + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "text", + "content": ". The total loss is the sum of " + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reconstr}}" + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrast}}" + }, + { + "bbox": [ + 131, + 427, + 478, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 456, + 258, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 258, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 258, + 468 + ], + "type": "text", + "content": "3.4 Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 475, + 408, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 408, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 408, + 487 + ], + "type": "text", + "content": "We detail here the specific parameters chosen in all our experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": "Tokenization. We split each tile along a regular spatial grid to produce a set of non-overlapping patches " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " consistent across all modalities. For TreeSat and FLAIR, we use a " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "10 \\times 10\\mathrm{m}" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " grid, meaning that the VHR input tokens are small image patches of size " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "50 \\times 50" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " per pixel. The patches of Sentinel observations with a resolution of " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " are single-pixel temporal sequences of spectral measurements. For PASTIS-HD, we use a " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "40 \\times 40\\mathrm{m}" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " grid, meaning that the VHR patches are of size " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "40 \\times 40" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{m}" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " per pixel. The patches of Sentinel observations [26] are " + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 131, + 492, + 482, + 589 + ], + "type": "text", + "content": " image time series which we spatially flatten before encoding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 594, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 483, + 666 + ], + "type": "text", + "content": "Hyperparameters. To show the versatility of OmniSat, we use the same configuration throughout all experiments. The embedding size is " + }, + { + "bbox": [ + 131, + 594, + 483, + 666 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 131, + 594, + 483, + 666 + ], + "type": "text", + "content": ", resulting in image encoders and decoders with 3.6M and 1.8M parameters, 403k and 96k for optical time series, and 402k and 95k for radar time series. The modality combiner module is composed of " + }, + { + "bbox": [ + 131, + 594, + 483, + 666 + ], + "type": "inline_equation", + "content": "B = 6" + }, + { + "bbox": [ + 131, + 594, + 483, + 666 + ], + "type": "text", + "content": " residual self-attention blocks and a single cross-attention block, for a total of 3.6M parameters. We train our model on 3 A6000 GPUs with a batch size of 128" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "G.Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "multimodal tiles per GPU and set the contrastive temperature " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " to 0.1. We train our model with the ADAM optimizer [53], with a learning rate of " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " for pretraining and " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "2\\times 10^{-5}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " for fine-tuning, and a ReduceLROnPlateau scheduler [1] with a patience of 10 epochs and a decay rate of 0.1. When re-implementing competing methods, we use the hyperparameters of their open-source repository." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 194, + 219, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 219, + 208 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 219, + 208 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 218, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 266 + ], + "type": "text", + "content": "We evaluate OmniSat's performance across three multimodal datasets, including two new datasets introduced in this work, and presented in Sec. 4.1. We outline our experimental protocol and our adaptation of competing methods in Sec. 4.2. We then present our quantitative results and analysis in Sec. 4.3 and conduct an ablation study in Sec. 4.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 284, + 194, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 284, + 194, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 194, + 294 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 303, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 481, + 327 + ], + "type": "text", + "content": "We consider three multimodal datasets: FLAIR [31], and the augmented TreeSatAI-TS [2] and PASTIS-HD [33, 34]. See Fig. 1 for an illustration of these two last datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "text", + "content": "TreeSatAI-TS: TreeSatAI [2] is a multimodal dataset for tree species identification, containing 50,381 tiles of " + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "inline_equation", + "content": "60 \\times 60\\mathrm{m}" + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "text", + "content": " with multi-label annotations for 20 classes and all taken in Germany. Each tile is associated with a very high resolution RGB and near-infrared (NIR) image (" + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "text", + "content": " pixel resolution), a single Sentinel-2 multi-spectral image (" + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "text", + "content": " per pixel resolution, 10 bands), and a single Sentinel-1 radar image (" + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 130, + 332, + 482, + 403 + ], + "type": "text", + "content": " per pixel resolution, 3 bands: two polarization channels and their ratio)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "Motivated by the fact that fine-grained vegetation discrimination relies heavily on temporal dynamics information [91], we introduce TreeSatAI- " + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{TS}^1" + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": ". This extended version uses open-source data to add Sentinel-1 and Sentinel-2 time series to each tile, spanning the closest available year to the VHR observation for Sentinel-2. Note that due to the weather patterns and position of the area of interest with respect to Sentinel-2's orbit, the optical time series is particularly irregular and occluded, with up to " + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": " of acquisitions being non-exploitable. Despite this challenge, we included the raw observations without pre-processing, whereas TreeSatAI's single-date images have been manually selected." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 504, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 586 + ], + "type": "text", + "content": "PASTIS-HD: The PASTIS dataset [33], is designed for semantic and panoptic segmentation of agricultural parcels using Sentinel-2 time series and covers 18 crop types across 2,433 image time series with dimensions of " + }, + { + "bbox": [ + 130, + 504, + 482, + 586 + ], + "type": "inline_equation", + "content": "1280 \\times 1280\\mathrm{m}" + }, + { + "bbox": [ + 130, + 504, + 482, + 586 + ], + "type": "text", + "content": ". Each series contains between 38 and 61 observations with 10 spectral bands. PASTIS-R [34] adds the corresponding Sentinel-1 radar time series. We only used the ascendant time series of Sentinel-1 for our training and evaluation, for a total of 169,587 radar images with three bands." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 586, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 635 + ], + "type": "text", + "content": "To enhance the spatial resolution and utility of PASTIS, we introduce PASTISHD², which integrates contemporary VHR satellite images (SPOT 6-7 [24]). We apply orthorectification and pansharpening, resample the resulting images to a 1m resolution, and finally convert them to 8 bits. We follow the protocol of Irvin et al. [51] to use the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 642, + 455, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 455, + 653 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 455, + 653 + ], + "type": "text", + "content": "1The dataset is available at https://huggingface.co/datasets/IGNF/TreeSatAI-Time-Series." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 653, + 416, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 416, + 664 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 416, + 664 + ], + "type": "text", + "content": "2The dataset is available at https://huggingface.co/datasets/IGNF/PASTIS-HD." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "dense annotations for a multi-label classification task: each patch is associated with the labels of all of its pixels. This conversion allows us to evaluate all methods in the same setting and configuration as TreeSatAI." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "text", + "content": "FLAIR. The FLAIR dataset [31] combines VHR aerial images with time series data. It comprises 77,762 aerial tiles (" + }, + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "text", + "content": " pixels, " + }, + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "text", + "content": " resolution) with five channels (RGB, near-infrared, and a normalized digital surface model) taken in France, alongside corresponding Sentinel-2 time series (" + }, + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 130, + 158, + 483, + 229 + ], + "type": "text", + "content": " resolution, 10 spectral bands, 20 to 114 observations per year). We apply the same processing as PASTIS to use the dense annotation for a multi-label classification task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 249, + 248, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 249, + 248, + 261 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 248, + 261 + ], + "type": "text", + "content": "4.2 Experimental Setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 282 + ], + "type": "text", + "content": "This section details our experimental protocol and our adaption of competing algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 288, + 386, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 386, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 386, + 299 + ], + "type": "text", + "content": "Evaluation Protocol. All experiments follow a similar setting:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 481, + 346 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 300, + 481, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 334 + ], + "type": "text", + "content": "- Pre-training (optional). Methods that support self-supervised pre-training (OmniSat, SatMAE [20], ScaleMAE [74], CROMA [29]) are pre-trained for up to 250 epochs on the entire training set without access to labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 369, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 369, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 369, + 346 + ], + "type": "text", + "content": "- Fine-Tuning. We propose two settings for fine-tuning:" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 155, + 348, + 480, + 407 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 155, + 348, + 480, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 348, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 155, + 348, + 480, + 371 + ], + "type": "text", + "content": "- Fully Supervised Fine-Tuning. We train the resulting models using all the labels in the training set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 156, + 373, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 373, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 156, + 373, + 480, + 407 + ], + "type": "text", + "content": "- Semi-Supervised Fine-Tuning. We use a portion of " + }, + { + "bbox": [ + 156, + 373, + 480, + 407 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 156, + 373, + 480, + 407 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 156, + 373, + 480, + 407 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 156, + 373, + 480, + 407 + ], + "type": "text", + "content": " of the training set, stratified by the distribution of classes, to fine-tune the models. For models without pre-training, this corresponds to supervision in the low-data regime." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 138, + 409, + 481, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 481, + 433 + ], + "type": "text", + "content": "- Unimodal and Multimodal Evaluation. We evaluate all methods using each available modality independently and combining all supported modalities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 440, + 482, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 512 + ], + "type": "text", + "content": "Adapting Competing Approaches. We report the performance of several methods taken from the literature on our considered datasets: LightGBM [2], PRESTO [88], and MOSAIKS [77]. However, few existing methods can operate on single- and multi-date data at the same time. To ensure a fair evaluation of competing approaches, we modify various state-of-the-art models to handle a broader combination of modalities. We provide details on these changes in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 531, + 316, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 316, + 542 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 316, + 542 + ], + "type": "text", + "content": "4.3 Numerical Experiments and Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 552, + 481, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 481, + 576 + ], + "type": "text", + "content": "In this section, we report our model's performance and efficiency compared to other approaches across the considered datasets and propose our analysis." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 582, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 604 + ], + "type": "text", + "content": "TreeSatAI-TS. Tab. 2 presents the performance of different models on TreeSatAI and TreeSatAI-TS. We report several key observations:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 606, + 481, + 665 + ], + "type": "text", + "content": "- Benefit of Time Series. For the original TreeSatAI dataset with single-date Sentinel1/2 observations, none of the pre-training schemes significantly improve performance beyond simple baselines such as ResNet, PSE, or MLP, even in a semi-supervised setting. In particular, single-date S1 observations yield low performance for all methods (below 20 F1-score), emphasizing the need to use the entire time series." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "G.Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 198, + 480, + 495 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "text", + "content": "Table 2: Performance on TreeSatAI-TS. We report the weighted F1 for multi-label tree species classification on TreeSatAI (TSAI) and our extended TreeSatAI-TS (TSAI-TS) dataset when fine-tuning with " + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "text", + "content": " of training labels. The first line of the table is the modality used for evaluation. We distinguish methods that are best for one modality within a dataset, best in a dataset across all modalities, and the best overall performance. " + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "text", + "content": ": late feature fusion with a ResNet pre-trained on ImageNet. " + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "inline_equation", + "content": "\\blacksquare" + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "text", + "content": ": Foundation model trained on extensive external data. " + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 191 + ], + "type": "text", + "content": ": model evaluated on this dataset for the first time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 198, + 480, + 495 + ], + "lines": [ + { + "bbox": [ + 134, + 198, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 134, + 198, + 480, + 495 + ], + "type": "table", + "html": "
Modelpre-trainingAllSentinel-1Sentinel-2VHR Image
10%100%10%100%10%100%10%100%
Evaluated on TreeSatAI: single date for Sentinel-1 and Sentinel-2
† PSE [35]None47.2*68.1*11.514.648.558.3--
† ViT [25]None42.757.18.717.539.857.336.751.7
MLP [2]None42.6*71.5*3.410.122.152.0--
ResNet [2]ImageNet------58.870.1
LightGBM [2]ImageNet-54.3*-11.9-48.2-44.0
PRESTO [88]---19.8-46.3--
† DOFA [96]59.571.611.619.348.257.051.667.5
MOSAIKS [22,77]TSAI-----56.0--
† CROMA [29]TSAI49.661.010.112.747.855.7--
† SatMAE [20]TSAI46.161.5--40.349.744.161.4
† ScaleMAE [74]TSAI47.662.5--46.755.246.963.6
OmniSat (ours)TSAI56.270.45.36.448.457.152.868.9
Evaluated on TreeSatAI-TS: Sentinel-1 and Sentinel-2 time series spanning one year
† PSE+LTAE [35]None59.4*71.2*42.652.444.057.2--
† UT&T [31]ImageNet43.856.742.355.241.557.044.355.9
† DOFA [96]41.871.30.00.025.039.451.667.5
† Scale-MAE [74]TSAI-TS44.160.4--11.031.546.963.6
OmniSat (ours)None52.273.331.655.933.949.751.471.0
OmniSat (ours)TSAI-TS61.174.248.256.751.462.958.370.5
", + "image_path": "5d2858b392ceab9018227921915a045d9153167f7f0643228cc7d1b1b53b66e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 521, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 521, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 147, + 521, + 482, + 558 + ], + "type": "text", + "content": "OmniSat exhibits significantly improved results on TreeSatAI-TS, with or without pretraining. Image models struggles to extract meaningful features temporally aggregated temporal observations, while OmniSat learn rich dynamic features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 558, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 558, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 147, + 558, + 482, + 604 + ], + "type": "text", + "content": "The foundation model DOFA [96], with 111M parameters and a large closed-source training set, outperforms all models when evaluated on single-date modalities. However, OmniSat reaches higher performances on TreeSatAI-TS with only 10 million parameters, which we attribute to its ability to leverage temporal modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "text", + "content": "- Benefits of Multimodality. When using all modalities, OmniSat outperforms all competing methods by a margin of " + }, + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "text", + "content": " F1-score. The multimodal performance of OmniSat and CROMA, which learn to combine data sources, is strictly superior to the F1-score of their best modality by " + }, + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 138, + 605, + 482, + 665 + ], + "type": "text", + "content": " points, respectively. Conversely, the performance of methods that rely on late-fusion (SatMAE, ScaleMAE," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 166, + 479, + 305 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 3: Performance on PASTIS-HD. We report the macro-averaged F1-score for crop-type multi-class classification on the PASTIS-HD dataset. We distinguish methods that are best for one modality, best in a dataset across all modalities. " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": ": late feature fusion with a ResNet. " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": ": model evaluated on this dataset for the first time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 166, + 479, + 305 + ], + "lines": [ + { + "bbox": [ + 134, + 166, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 134, + 166, + 479, + 305 + ], + "type": "table", + "html": "
Modelpre-trainedAllSentinel-1Sentinel-2VHR image
20%100%20%100%20%100%20%100%
† UTAE [33, 34]None36.8*46.9*20.140.732.737.6--
† ResNet50 [47]ImageNet------57.659.3
† UT&T [31]ImageNet54.253.558.862.854.961.351.149.8
† DOFA [96]53.755.736.741.550.853.447.954.8
† Scale-MAE [74]PASTIS-HD42.042.2--41.246.148.851.9
† CROMA [29]PASTIS-HD57.560.155.356.153.056.7--
OmniSat (ours)No42.059.158.260.251.760.147.352.8
OmniSat (ours)PASTIS-HD62.669.960.869.061.870.854.659.3
", + "image_path": "d13c290ff9ebebdf8e6fb1180059ddc0f169baf761486f08a97fc810cd4a2473.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 165, + 360, + 444, + 466 + ], + "blocks": [ + { + "bbox": [ + 130, + 319, + 483, + 353 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 483, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 483, + 353 + ], + "type": "text", + "content": "Table 4: Performance on FLAIR. We report the macro-averaged F1-score for land cover multiclass classification on the FLAIR dataset. We distinguish methods that are best for one modality and best in a dataset. " + }, + { + "bbox": [ + 130, + 319, + 483, + 353 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 319, + 483, + 353 + ], + "type": "text", + "content": ": model evaluated on this dataset for the first time." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 360, + 444, + 466 + ], + "lines": [ + { + "bbox": [ + 165, + 360, + 444, + 466 + ], + "spans": [ + { + "bbox": [ + 165, + 360, + 444, + 466 + ], + "type": "table", + "html": "
Modelpre-trainedAllSentinel-2VHR Image
10%100%10%100%10%100%
† UT&T [31]ImageNet44.248.857.462.058.965.5
† DOFA [96]70.674.957.061.066.872.1
† ScaleMAE [74]FLAIR63.170.052.561.061.267.3
OmniSat (ours)No62.570.056.165.464.771.5
OmniSat (ours)FLAIR60.673.456.865.465.271.6
", + "image_path": "676a6993c1abb48d833a1d362184f58e62d58de409d0443d1e0b48066416521a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 491, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 491, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 147, + 491, + 481, + 514 + ], + "type": "text", + "content": "ViT) is comparable to their best modality. This demonstrates the value of learning to combine information from different sources end-to-end." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": "- Benefits of Cross-Modal Pre-Training. With access to all modalities, our self-supervised pre-training improves by " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": " point the F1-score of the model fine-tuned on " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": " of labels, compared to not pre-training, and " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": " when using only " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": " of labels. This shows that our pre-training leads to more expressive multimodal features. Interestingly, when performing inference with Sentinel-2 time series alone, the performance increase linked to the pre-training becomes " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "13.2\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": " labels and " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "17.5\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 138, + 515, + 482, + 612 + ], + "type": "text", + "content": ". This illustrates that our self-supervised pre-training scheme improves the features learned by each encoder despite not relying on annotated data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Experiments on PASTIS-HD. The analysis of the performance of various models on PASTIS-HD is reported in Tab. 3, and is consistent with the ones of TreeSatAI-TS. First, by learning to combine all modalities despite their different resolutions, OmniSat achieves state-of-the-art results on this benchmark. Second, our cross-modal pretraining" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "G.Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 114, + 478, + 258 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 478, + 258 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 478, + 258 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 478, + 258 + ], + "type": "image", + "image_path": "de0d8468d42cad0b4ccc124f4aeb30b473d15b7ff6c72aeacecfa27a202211ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "type": "text", + "content": "Fig. 5: Efficiency. We report the best performance of different models between TreeSatAI and TreeSatAI-TS, with pre-training and fine-tuning using " + }, + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 262, + 482, + 298 + ], + "type": "text", + "content": " of labels. The area of the markers is proportional to the training time, broken down in pre-training and fine-tuning when applicable." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "text", + "content": "significantly improves OmniSat's performance in the multimodal " + }, + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "inline_equation", + "content": "(+10.8\\mathrm{pF1}" + }, + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "text", + "content": " -score with " + }, + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 318, + 482, + 354 + ], + "type": "text", + "content": " of training label) and all single-modality settings (8.8 points for Sentinel-1, 10.7 for Sentinel-2, and 6.5 for the VHR images)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 360, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 456 + ], + "type": "text", + "content": "Experiments on FLAIR. We report in Tab. 4 the results on the bimodal FLAIR dataset for multilabel classification. OmniSat outperforms the much larger ScaleMAE [74] and UT&T [31] models with " + }, + { + "bbox": [ + 130, + 360, + 482, + 456 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 360, + 482, + 456 + ], + "type": "text", + "content": " of labels and both modalities by " + }, + { + "bbox": [ + 130, + 360, + 482, + 456 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 360, + 482, + 456 + ], + "type": "text", + "content": ". Our pre-training scheme had a smaller impact than for the TreeSatAI-TS experiment. We attribute this to the fact that only two modalities are available, which decreases the supervisory power of our cross-modal contrastive objective and our multimodal reconstruction loss. This highlights a limitation of OmniSat: the model needs to be pre-trained on a modality-rich dataset to achieve its best performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "type": "text", + "content": "Efficiency Evaluation. We plot in Fig. 5 the best performance between TreeSatAI and TreeSatAI-TS for different models according to their size and training time. OmniSat is more compact, faster to train, and performs better than all evaluated models, including the DOFA foundation model. The highly-specialized combination of PSE, LTAE, and ResNet is a strong contender, outperforming significantly larger models with generic encoding-decoding schemes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 547, + 223, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 223, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 223, + 559 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 481, + 588 + ], + "type": "text", + "content": "In this section, we report the results of several experiments evaluating the impact and validity of our main design choices, see Tab. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "a) Encoder/Decoder Architecture. We propose several improvements to the standard image encoder-decoder scheme used in computer vision to accommodate the specificities of EO data. In particular, passing the max-pool indices from the image patch encoder to its decoder allows the learned representation to focus on characterizing the spectral signature instead of fine-grained spatial information, and leads to a performance increase of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " in the full supervision setting." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 144, + 438, + 201 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 5: Ablation Study. We present the impact of several design choices on the TreeSatAI-TS dataset, measured in terms of macro-averaged F1-score." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 144, + 438, + 201 + ], + "lines": [ + { + "bbox": [ + 173, + 144, + 438, + 201 + ], + "spans": [ + { + "bbox": [ + 173, + 144, + 438, + 201 + ], + "type": "table", + "html": "
Experiment10%100%Experiment10%100%
OmniSat61.174.2b) no contrastive loss55.673.4
a) no index bypass57.573.5b) naive contrastive loss57.872.2
a) no date filtering58.271.6b) no reconstruction loss59.072.2
", + "image_path": "98fd15b3a5a51ab06fb7ae8978d16f472716ba6f17232aa23201a20afb861fba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 481, + 273 + ], + "type": "text", + "content": "As clouds frequently obstruct optical time series, we use a unsupervised date-filtering scheme to reconstruct only meaningful acquisitions. This approach leads to a significant improvement of " + }, + { + "bbox": [ + 130, + 225, + 481, + 273 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 130, + 225, + 481, + 273 + ], + "type": "text", + "content": ", showcasing the benefit of developing modality-aware approaches for EO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "text", + "content": "b) Role of Loss Functions. When training without contrastive loss, we observe a decrease in performance of " + }, + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "text", + "content": " in the fully supervised regime, and a more pronounced drop of " + }, + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "text", + "content": " in the semi-supervised regime. This demonstrates how learning consistent encoding across encoders facilitates their subsequent fusion. Interestingly, when implementing a naive contrastive loss that considers all negative examples from the batch, the decrease is greater than simply removing this loss (" + }, + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 279, + 481, + 361 + ], + "type": "text", + "content": " in full supervision). This strategy may introduce indistinguishable negative examples and perturb the learning process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "content": "We also remove the reconstruction loss, meaning that only the encoders are learned contrastively during pre-training. This results in a drop of " + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "content": " F1-score point, illustrating the importance of pre-training the transformer " + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "content": " alongside its encoders." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 450 + ], + "type": "text", + "content": "Limitations. All datasets used in our experiments are based in Europe, primarily due to the availability of open-access annotations. This regional focus prevents us from evaluating our model's performance in tropical and developing countries, which present unique challenges in terms of label provision, heterogeneity, and complex classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 481, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 473 + ], + "type": "text", + "content": "A limitation of our pre-training scheme is its dependence on a sufficient number of aligned modalities, as illustrated by its moderate impact on the bimodal FLAIR dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 495, + 211, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 211, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 211, + 506 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "type": "text", + "content": "We introduced OmniSat, a new architecture for the self-supervised modality fusion of Earth Observation (EO) data from multiple sources. To facilitate its evaluation, we augmented two existing datasets with new modalities of different natures and resolutions. We experimentally showed that leveraging diverse modalities with a flexible model improves the model's performance in both fully and semi-supervised settings. Moreover, our training scheme can exploit the spatial alignment of multiple modalities to improve our model's unimodal performance. Finally, we proposed several improvements to leverage the unique structure of EO data in the architecture of our model, such as automatic date filtering for reconstructing time series. We hope that our promising results and new datasets will encourage the computer vision community to consider EO data as a playing field for evaluating and developing novel self-supervised multimodal algorithms." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "G.Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "This work was supported by ANR project READY3D ANR-19-CE23-0007, and was granted access to the HPC resources of IDRIS under the allocations AD011014719 and AD011014286R1 made by GENCI. We thank Anatol Garioud and Sébastien Giordano for their help on the creation of TreeSatAI-TS and PASTIS-HD datasets. The SPOT images are opendata thanks to the Dataterra Dinamis initiative in the case of the \"Couverture France DINAMIS\" program. We thank Jordi Inglada for inspiring discussions and valuable feedback." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 242, + 190, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 190, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 190, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 266, + 483, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 141, + 266, + 483, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 266, + 483, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 483, + 299 + ], + "type": "text", + "content": "1. PyTorch: ReduceLROOnPlateau.org/docs/stable/generated/torch.optim.lr_scheduler.ReduceLROOnPlateau.html#torch.optim.lr_scheduler.ReduceLROOnPlateau, accessed: 2024-02-29 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 300, + 483, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 300, + 483, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 483, + 344 + ], + "type": "text", + "content": "2. Ahlswede, S., Schulz, C., Gava, C., Helber, P., Bischke, B., Förster, M., Arias, F., Hees, J., Demir, B., Kleinschmit, B.: TreeSatAI Benchmark Archive: A multi-sensor, multi-label dataset for tree species classification in remote sensing. Earth System Science Data Discussions (2022) 2, 3, 4, 9, 10, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 344, + 482, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 344, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 482, + 399 + ], + "type": "text", + "content": "3. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: A visual language model for few-shot learning. In: NeurIPS (2022) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 399, + 482, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 399, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 482, + 433 + ], + "type": "text", + "content": "4. Amitrano, D., Di Martino, G., Guida, R., Iervolino, P., Iodice, A., Papa, M.N., Riccio, D., Ruello, G.: Earth environmental monitoring using multi-temporal synthetic aperture radar: A critical review of selected applications. Remote Sensing (2021) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 433, + 482, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 433, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 141, + 433, + 482, + 455 + ], + "type": "text", + "content": "5. Anderson, K., Ryan, B., Sonntag, W., Kavvada, A., Friedl, L.: Earth observation in service of the 2030 agenda for sustainable development. Geo-spatial Information Science (2017) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 455, + 482, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 455, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 141, + 455, + 482, + 487 + ], + "type": "text", + "content": "6. Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding predictive architecture. In: CVPR (2023) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 487, + 482, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 487, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 141, + 487, + 482, + 510 + ], + "type": "text", + "content": "7. Ayush, K., Uzkent, B., Meng, C., Tanmay, K., Burke, M., Lobell, D., Ermon, S.: Geography-aware self-supervised learning. In: ICCV (2021) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 510, + 482, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 510, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 141, + 510, + 482, + 532 + ], + "type": "text", + "content": "8. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE TPAMI (2017) 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 532, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 532, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 141, + 532, + 482, + 555 + ], + "type": "text", + "content": "9. Baevski, A., Hsu, W.N., Xu, Q., Babu, A., Gu, J., Auli, M.: Data2vec: A general framework for self-supervised learning in speech, vision and language. In: ICML (2022) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 482, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 482, + 576 + ], + "type": "text", + "content": "10. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2021) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 576, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 482, + 610 + ], + "type": "text", + "content": "11. Bao, X., Zhang, R., Lv, J., Wu, R., Zhang, H., Chen, J., Zhang, B., Ouyang, X., Liu, G.: Vegetation descriptors from Sentinel-1 SAR data for crop growth monitoring. ISPRS Journal of Photogrammetry and Remote Sensing (2023) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 610, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 482, + 632 + ], + "type": "text", + "content": "12. Bastani, F., Wolters, P., Gupta, R., Ferdinando, J., Kembhavi, A.: SatlasPretrain: A large-scale dataset for remote sensing image understanding. In: ICCV (2023) 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 482, + 665 + ], + "type": "text", + "content": "13. Bayoudh, K., Knani, R., Hamdaoui, F., Mtibaa, A.: A survey on deep multimodal learning for computer vision: Advances, trends, applications, and datasets. The Visual Computer (2022) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Benedetti, P., Ienco, D., Gaetano, R., Ose, K., Pensa, R.G., Dupuy, S.: " + }, + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3" + }, + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "-fusion: A deep learning architecture for multiscale multimodal multitemporal satellite data fusion. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2018) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 481, + 171 + ], + "type": "text", + "content": "15. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 481, + 193 + ], + "type": "text", + "content": "16. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 481, + 215 + ], + "type": "text", + "content": "17. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 216, + 472, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 472, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 472, + 226 + ], + "type": "text", + "content": "18. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 227, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 481, + 247 + ], + "type": "text", + "content": "19. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: CVPR (2018) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "type": "text", + "content": "20. Cong, Y., Khanna, S., Meng, C., Liu, P., Rozi, E., He, Y., Burke, M., Lobell, D., Ermon, S.: SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In: NeurIPS (2022) 2, 3, 10, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 282, + 481, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 312 + ], + "type": "text", + "content": "21. Coppin, P., Lambin, E., Jonckheere, I., Muys, B.: Digital change detection methods in natural ecosystem monitoring: A review. Analysis of multi-temporal remote sensing images (2002) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 481, + 346 + ], + "type": "text", + "content": "22. Corley, I., Robinson, C., Dodhia, R., Ferres, J.M.L., Najafirad, P.: Revisiting pre-trained remote sensing model benchmarks: Resizing and normalization matters. arXiv preprint arXiv:2305.13456 (2023) 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 481, + 369 + ], + "type": "text", + "content": "23. Dai, A., Nießner, M.: 3DMV: Joint 3D-multi-view prediction for 3D semantic scene segmentation. In: ECCV (2018) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 369, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 481, + 390 + ], + "type": "text", + "content": "24. DataTerra Dinamis: Diffusion OpenData Dinamis, https://dinamis.data-terra.org/opendata/, accessed: 2023-12-15 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 423 + ], + "type": "text", + "content": "25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 425, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 481, + 467 + ], + "type": "text", + "content": "26. Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B., Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F., Bargellini, P.: Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment (2012) 1, 4, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 468, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 489 + ], + "type": "text", + "content": "27. Ebel, P., Xu, Y., Schmitt, M., Zhu, X.X.: SEN12MS-CR-TS: A remote-sensing data set for multimodal multitemporal cloud removal. IEEE TGRS (2022) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "text", + "content": "28. Ekim, B., Stomberg, T.T., Roscher, R., Schmitt, M.: MapInWild: A remote sensing dataset to address the question of what makes nature wild. IEEE Geoscience and Remote Sensing Magazine (2023) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "text", + "content": "29. Fuller, A., Millard, K., Green, J.R.: CROMA: Remote sensing representations with contrastive radar-optical masked autoencoders. In: NeurIPS (2023) 1, 2, 3, 10, 11, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "type": "text", + "content": "30. Gao, Y., Sun, X., Liu, C.: A general self-supervised framework for remote sensing image classification. Remote Sensing (2022) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 609 + ], + "type": "text", + "content": "31. Garioud, A., Gonthier, N., Landrieu, L., De Wit, A., Valette, M., Poupee, M., Giordano, S., Wattrelos, B.: FLAIR: A country-scale land cover semantic segmentation dataset from multi-source optical imagery. In: NeurIPS Dataset and Benchmark (2023) 3, 4, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "32. Garnot, V.S.F., Landrieu, L.: Lightweight temporal self-attention for classifying satellite images time series. In: Advanced Analytics and Learning on Temporal Data: ECML PKDD Workshop (2020) 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "33. Garnot, V.S.F., Landrieu, L.: Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In: ICCV (2021) 2, 3, 9, 12" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "G. Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 148 + ], + "type": "text", + "content": "34. Garnot, V.S.F., Landrieu, L., Chehata, N.: Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing (2022) 2, 3, 4, 9, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 481, + 171 + ], + "type": "text", + "content": "35. Garnot, V.S.F., Landrieu, L., Giordano, S., Chehata, N.: Satellite image time series classification with pixel-set encoders and temporal self-attention. In: CVPR (2020) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 481, + 215 + ], + "type": "text", + "content": "36. Ghamisi, P., Rasti, B., Yokoya, N., Wang, Q., Hofle, B., Bruzzone, L., Bovolo, F., Chi, M., Anders, K., Gloaguen, R., et al.: Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine (2019) 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 481, + 237 + ], + "type": "text", + "content": "37. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 237, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 237, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 237, + 481, + 258 + ], + "type": "text", + "content": "38. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023) 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 481, + 281 + ], + "type": "text", + "content": "39. Girdhar, R., Singh, M., Ravi, N., van der Maaten, L., Joulin, A., Misra, I.: Omnivore: A single model for many visual modalities. In: CVPR (2022) 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 281, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 481, + 323 + ], + "type": "text", + "content": "40. Goldberg, H.R., Ratto, C.R., Banerjee, A., Kelbaugh, M.T., Giglio, M., Vermote, E.F.: Automated global-scale detection and characterization of anthropogenic activity using multi-source satellite-based remote sensing imagery. In: Geospatial Informatics XIII. SPIE (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 325, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 481, + 346 + ], + "type": "text", + "content": "41. Greenwell, C., Crall, J., Purri, M., Dana, K., Jacobs, N., Hadzic, A., Workman, S., Leotta, M.: WATCH: Wide-area terrestrial change hypercube. In: WACV (2024) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "type": "text", + "content": "42. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. In: NeurIPS (2020) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 380, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 411 + ], + "type": "text", + "content": "43. Hackstein, J., Sumbul, G., Clasen, K.N., Demir, B.: Exploring masked autoencoders for sensor-agnostic image retrieval in remote sensing. arXiv preprint arXiv:2401.07782 (2024) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 412, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 435 + ], + "type": "text", + "content": "44. Hazirbas, C., Ma, L., Domokos, C., Cremers, D.: FuseNet: Incorporating depth into semantic segmentation via fusion-based CNN architecture. In: ACCV (2017) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 481, + 456 + ], + "type": "text", + "content": "45. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 457, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 478 + ], + "type": "text", + "content": "46. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 479, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 500 + ], + "type": "text", + "content": "47. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "text", + "content": "48. Hu, J., Liu, R., Hong, D., Camero, A., Yao, J., Schneider, M., Kurz, F., Segl, K., Zhu, X.X.: MDAS: A new multimodal benchmark dataset for remote sensing. Earth System Science Data Discussions (2022) 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "text", + "content": "49. Huang, P.Y., Sharma, V., Xu, H., Ryali, C., Fan, H., Li, Y., Li, S.W., Ghosh, G., Malik, J., Feichtenhofer, C.: MAViL: Masked audio-video learners. In: NeurIPS (2023) 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 578 + ], + "type": "text", + "content": "50. Ibanez, D., Fernandez-Beltran, R., Pla, F., Yokoya, N.: Masked auto-encoding spectral-spatial transformer for hyperspectral image classification. IEEE TGRS (2022) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "type": "text", + "content": "51. Irvin, J., Tao, L., Zhou, J., Ma, Y., Nashold, L., Liu, B., Ng, A.Y.: USat: A unified self-supervised encoder for multi-sensor satellite imagery. arXiv preprint arXiv:2312.02199 (2023) 1, 2, 3, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "type": "text", + "content": "52. Kenton, J.D.M.W.C., Toutanova, L.K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 632, + 458, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 458, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 458, + 643 + ], + "type": "text", + "content": "53. Kingma, D.P., Ba, J.: ADAM: A method for stochastic optimization. ICLR (2015) 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "54. Krispel, G., Opitz, M., Waltner, G., Possegger, H., Bischof, H.: Fuseseg: LiDAR point cloud segmentation fusing multi-modal data. In: WACV (2020) 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 482, + 160 + ], + "type": "text", + "content": "55. Kuffer, M., Thomson, D.R., Boo, G., Mahabir, R., Grippa, T., Vanhuysse, S., Engstrom, R., Ndugwa, R., Makau, J., Darin, E., et al.: The role of Earth observation in an integrated deprived area mapping \"system\" for low-to-middle income countries. Remote sensing (2020) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 161, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 161, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 137, + 161, + 482, + 204 + ], + "type": "text", + "content": "56. Lacoste, A., Sherwin, E.D., Kerner, H., Alemohammad, H., Lütjens, B., Irvin, J., Dao, D., Chang, A., Gunturkun, M., Drouin, A., et al.: Toward foundation models for Earth monitoring: Proposal for a climate change benchmark. arXiv preprint arXiv:2112.00570 (2021) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 205, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 205, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 482, + 225 + ], + "type": "text", + "content": "57. Li, D., Tong, Q., Li, R., Gong, J., Zhang, L.: Current issues in high-resolution Earth observation technology. Science China Earth sciences (2012) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 227, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 227, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 137, + 227, + 482, + 258 + ], + "type": "text", + "content": "58. Li, J., Hong, D., Gao, L., Yao, J., Zheng, K., Zhang, B., Chanussot, J.: Deep learning in multimodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation 112, 102926 (2022) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 259, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 259, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 137, + 259, + 482, + 281 + ], + "type": "text", + "content": "59. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE TPAMI (2022) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 281, + 482, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 281, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 137, + 281, + 482, + 312 + ], + "type": "text", + "content": "60. Liu, Y., Li, X., Hua, Z., Xia, C., Zhao, L.: A band selection method with masked convolutional autoencoder for hyperspectral image. IEEE Geoscience and Remote Sensing Letters (2022) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 315, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 315, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 137, + 315, + 482, + 357 + ], + "type": "text", + "content": "61. Ma, Y., Li, Y., Feng, K., Xia, Y., Huang, Q., Zhang, H., Prieur, C., Licciardi, G., Malha, H., Chanussot, J., et al.: The outcome of the 2021 IEEE GRSS data fusion contest-Track DSE: Detection of settlements without electricity. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 358, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 358, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 137, + 358, + 482, + 390 + ], + "type": "text", + "content": "62. Mai, G., Huang, W., Sun, J., Song, S., Mishra, D., Liu, N., Gao, S., Liu, T., Cong, G., Hu, Y., et al.: On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798 (2023) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 391, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 391, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 137, + 391, + 482, + 413 + ], + "type": "text", + "content": "63. Manas, O., Lacoste, A., Giró-i Nieto, X., Vazquez, D., Rodriguez, P.: Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In: ICCV (2021) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 137, + 413, + 482, + 445 + ], + "type": "text", + "content": "64. Manfreda, S., McCabe, M.F., Miller, P.E., Lucas, R., Pajuelo Madrigal, V., Mallinis, G., Ben Dor, E., Helman, D., Estes, L., Ciraolo, G., et al.: On the use of unmanned aerial systems for environmental monitoring. Remote sensing p. 641 (2018) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 446, + 482, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 446, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 137, + 446, + 482, + 477 + ], + "type": "text", + "content": "65. Moreira, A., Prats-Iraola, P., Younis, M., Krieger, G., Hajnsek, I., Papathanassiou, K.P.: A tutorial on synthetic aperture radar. IEEE Geoscience and Remote Sensing Magazine (2013) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 479, + 482, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 479, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 137, + 479, + 482, + 510 + ], + "type": "text", + "content": "66. Nakalembe, C.: Urgent and critical need for sub-saharan african countries to invest in Earth observation-based agricultural early warning and monitoring systems. Environmental Research Letters (2020) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 512, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 512, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 137, + 512, + 482, + 533 + ], + "type": "text", + "content": "67. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from RGBD images. In: ECCV (2012) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 534, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 534, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 137, + 534, + 482, + 555 + ], + "type": "text", + "content": "68. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 555, + 482, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 555, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 137, + 555, + 482, + 577 + ], + "type": "text", + "content": "69. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 578, + 482, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 578, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 137, + 578, + 482, + 609 + ], + "type": "text", + "content": "70. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. TLMR (2023) 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 137, + 611, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 611, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 137, + 611, + 482, + 632 + ], + "type": "text", + "content": "71. Pohl, C., Van Genderen, J.L.: Multisensor image fusion in remote sensing: Concepts, methods and applications. International journal of remote sensing (1998) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 137, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 632, + 482, + 665 + ], + "type": "text", + "content": "72. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "G.Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 117, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 117, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 117, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 117, + 482, + 149 + ], + "type": "text", + "content": "73. Recasens, A., Lin, J., Carreira, J., Jaegle, D., Wang, L., Alayrac, J.b., Luc, P., Miech, A., Smaira, L., Hemsley, R., et al.: Zorro: The masked multimodal transformer. arXiv preprint arXiv:2301.09595 (2023) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 482, + 183 + ], + "type": "text", + "content": "74. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: ICCV (2023) 2, 3, 10, 11, 12, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 482, + 205 + ], + "type": "text", + "content": "75. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3D semantic segmentation. In: CVPR (2022) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 482, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 482, + 250 + ], + "type": "text", + "content": "76. Robinson, C., Malkin, K., Jojic, N., Chen, H., Qin, R., Xiao, C., Schmitt, M., Ghamisi, P., Hänsch, R., Yokoya, N.: Global land-cover mapping with weak supervision: Outcome of the 2020 IEEE GRSS data fusion contest. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021) 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 482, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 482, + 283 + ], + "type": "text", + "content": "77. Rolf, E., Proctor, J., Carleton, T., Bolliger, I., Shankar, V., Ishihara, M., Recht, B., Hsiang, S.: A generalizable and accessible approach to machine learning with global satellite imagery. Nature communications (2021) 10, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 482, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 482, + 306 + ], + "type": "text", + "content": "78. Rußwurm, M., Körner, M.: Self-attention for raw optical satellite time series classification. ISPRS Journal of Photogrammetry and Remote Sensing (2020) 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 482, + 329 + ], + "type": "text", + "content": "79. Schmitt, M., Zhu, X.X.: Data fusion and remote sensing: An ever-growing relationship. IEEE Geoscience and Remote Sensing Magazine (2016) 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 482, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 482, + 362 + ], + "type": "text", + "content": "80. Secades, C., O'Connor, B., Brown, C., Walpole, M., et al.: Earth observation for biodiversity monitoring: A review of current approaches and future opportunities for tracking progress towards the aichi biodiversity targets. CBD technical series (2014) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 364, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 482, + 396 + ], + "type": "text", + "content": "81. Shermeyer, J., Hogan, D., Brown, J., Van Etten, A., Weir, N., Pacifici, F., Hansch, R., Bastidas, A., Soenen, S., Bacastow, T., et al.: Spacenet 6: Multi-sensor all weather mapping dataset. In: CVPR Workshop EarthVision (2020) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 397, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 482, + 418 + ], + "type": "text", + "content": "82. Shukor, M., Dancette, C., Rame, A., Cord, M.: UnIVAL: Unified model for image, video, audio and language tasks. TMLR (2023) 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 420, + 482, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 482, + 453 + ], + "type": "text", + "content": "83. Skidmore, A.K., Coops, N.C., Neinavaz, E., Ali, A., Schaepman, M.E., Paganini, M., Kissling, W.D., Vihervaara, P., Darvishzadeh, R., Feilhauer, H., et al.: Priority list of biodiversity metrics to observe from space. Nature Ecology & Evolution (2021) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 453, + 482, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 482, + 474 + ], + "type": "text", + "content": "84. Srivastava, S., Sharma, G.: OmniVec: Learning robust representations with cross modal sharing. In: WACV (2024) 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 475, + 482, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 482, + 508 + ], + "type": "text", + "content": "85. Sudmanns, M., Tiede, D., Augustin, H., Lang, S.: Assessing global Sentinel-2 coverage dynamics and data availability for operational Earth observation (EO) applications using the EO-Compass. International Journal of Digital Earth (2019) 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 509, + 482, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 482, + 552 + ], + "type": "text", + "content": "86. Sumbul, G., De Wall, A., Kreuziger, T., Marcelino, F., Costa, H., Benevides, P., Caetano, M., Demir, B., Markl, V.: BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine (2021) 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 554, + 482, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 482, + 574 + ], + "type": "text", + "content": "87. Tarasiou, M., Chavez, E., Zafeiriou, S.: ViTs for SITS: Vision transformers for satellite image time series. In: CVPR (2023) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 576, + 482, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 482, + 608 + ], + "type": "text", + "content": "88. Tseng, G., Zvonkov, I., Purohit, M., Rolnick, D., Kerner, H.: Lightweight, pre-trained transformers for remote sensing timeseries. arXiv preprint arXiv:2304.14065 (2023) 3, 10, 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 609, + 482, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 609, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 609, + 482, + 642 + ], + "type": "text", + "content": "89. Tseng, W.H., Lé, H.A., Boulch, A., Lefèvre, S., Tiede, D.: CROCO: Cross-modal contrastive learning for localization of Earth observation data. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 482, + 665 + ], + "type": "text", + "content": "90. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Token Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 482, + 435 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "text", + "content": "91. Vrieling, A., Meroni, M., Darvishzadeh, R., Skidmore, A.K., Wang, T., Zurita-Milla, R., Oosterbeek, K., O'Connor, B., Paganini, M.: Vegetation phenology from Sentinel-2 and field cameras for a Dutch barrier island. Remote sensing of environment (2018) 4, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 482, + 183 + ], + "type": "text", + "content": "92. Wang, Y., Graham, N.A.A., Xiong, Z., Liu, C., Albrecht, C.M., Zhu, X.X.: SSL4EO-S12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in Earth observation. IEEE Geoscience and Remote Sensing Magazine (2023) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 183, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 183, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 137, + 183, + 482, + 225 + ], + "type": "text", + "content": "93. Wenger, R., Puisant, A., Weber, J., Idoumghar, L., Forestier, G.: MultiSenGE: A multimodal and multitemporal benchmark dataset for land use/land cover remote sensing applications. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2022) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 481, + 248 + ], + "type": "text", + "content": "94. Wu, K., Peng, H., Chen, M., Fu, J., Chao, H.: Rethinking and improving relative position encoding for vision transformer. In: ICCV (2021) 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "type": "text", + "content": "95. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMim: A simple framework for masked image modeling. In: CVPR (2022) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 303 + ], + "type": "text", + "content": "96. Xiong, Z., Wang, Y., Zhang, F., Stewart, A.J., Hanna, J., Borth, D., Papoutsis, I., Saux, B.L., Camps-Valls, G., Zhu, X.X.: Neural plasticity-inspired foundation model for observing the Earth crossing modalities. arXiv preprint arXiv:2403.15356 (2024) 3, 4, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 325 + ], + "type": "text", + "content": "97. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change (2013) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "text", + "content": "98. Yang, M.Y., Landrieu, L., Tuia, D., Toth, C.: Multi-modal learning in photogrammetry and remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing (2021) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "text", + "content": "99. Yuan, Y., Lin, L., Liu, Q., Hang, R., Zhou, Z.G.: SITS-Former: A pre-trained spatio-spectral-temporal representation model for sentinel-2 time series classification. International Journal of Applied Earth Observation and Geoinformation (2022) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 381, + 448, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 381, + 448, + 391 + ], + "spans": [ + { + "bbox": [ + 135, + 381, + 448, + 391 + ], + "type": "text", + "content": "100. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 135, + 392, + 481, + 413 + ], + "type": "text", + "content": "101. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: Image BERT pre-training with online tokenizer. In: ICLR (2022) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 135, + 414, + 481, + 435 + ], + "type": "text", + "content": "102. Zong, Y., Mac Aodha, O., Hospedales, T.: Self-supervised multimodal learning: A survey. arXiv preprint arXiv:2304.01008 (2023) 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "G. Astruc et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_content_list.json b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5abeb32b6c502e3066380dce8fa4d9e3885b984 --- /dev/null +++ b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_content_list.json @@ -0,0 +1,1849 @@ +[ + { + "type": "text", + "text": "Omniview-Tuning: Boosting Viewpoint Invariance of Vision-Language Pre-training Models", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shouwei Ruan1, Yinpeng Dong2,5, Hanqing Liu3, Yao Huang1, Hang Su2,4, and Xingxing Wei1,3*", + "bbox": [ + 225, + 213, + 776, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Artificial Intelligence, Beihang University, Beijing 100191, China", + "bbox": [ + 236, + 255, + 763, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Dept. of Comp. Sci. & Tech., Institute for AI, BNrist Center, THBI Lab,", + "bbox": [ + 243, + 271, + 759, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, 100084, China", + "bbox": [ + 236, + 285, + 761, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Hangzhou Innovation Institute, Beihang University, Hangzhou 311228, China", + "bbox": [ + 236, + 297, + 763, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Zhongguancun Laboratory, Beijing, 100080, China", + "bbox": [ + 326, + 311, + 674, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 RealAI", + "bbox": [ + 468, + 325, + 532, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shouweiruan,hqliu,y_huang,xxwei}@buaa.edu.cn,{dongyinpeng,suhangss}@tsinghua.edu.cn", + "bbox": [ + 241, + 338, + 759, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision-Language Pre-training (VLP) models like CLIP have achieved remarkable success in computer vision and particularly demonstrated superior robustness to distribution shifts of 2D images. However, their robustness under 3D viewpoint variations is still limited, which can hinder the development for real-world applications. This paper successfully addresses this concern while keeping VLPs' original performance by breaking through two primary obstacles: 1) the scarcity of training data and 2) the suboptimal fine-tuning paradigms. To combat data scarcity, we build the Multi-View Caption (MVCap) dataset — a comprehensive collection of over four million multi-view image-text pairs across more than 100K objects, providing more potential for VLP models to develop generalizable viewpoint-invariant representations. To address the limitations of existing paradigms in performance trade-offs and training efficiency, we design a novel fine-tuning framework named Omniview-Tuning (OVT). Specifically, OVT introduces a Cross-Viewpoint Alignment objective through a minimax-like optimization strategy, which effectively aligns representations of identical objects from diverse viewpoints without causing overfitting. Additionally, OVT fine-tunes VLP models in a parameter-efficient manner, leading to minimal computational cost. Extensive experiments on various VLP models with different architectures validate that OVT significantly improves the models' resilience to viewpoint shifts and keeps the original performance, establishing a pioneering standard for boosting the viewpoint invariance of VLP models. The code and dataset are available via https://github.com/Heathcliff-saku/Omniview_Tuning", + "bbox": [ + 259, + 402, + 740, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision-Language Pre-training $\\cdot$ Viewpoint Invariance", + "bbox": [ + 261, + 763, + 702, + 777 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd8e9d5f691bd16af676e8d99a9563cf46c77023904d859ca71d05d7d7b9f1d5.jpg", + "image_caption": [ + "Fig. 1: The Challenge of Viewpoint Invariance in VLP. We selected benchmarks representing clean distributions (ImageNet-1K [14], CIFAR-100 [27]), common 2D-OOD (ImageNet-V2 [42], ImageNet-R(endition) [21], ImageNet-Sketch [55]), and viewpoint-OOD (ImageNet-V(微观)+ [46], OOD-CV(Pose) [61], MIRO [7]). We display samples from these data distributions (left) and report the Top-1 accuracy of the original CLIP (ViT-L/14) and our improved OVT-CLIP (ViT-L/14) (right)." + ], + "image_footnote": [], + "bbox": [ + 228, + 147, + 504, + 260 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8730a5b29bf7ea8c98929862ba32a92437859482f0ad87e48c984d656b4a10b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 147, + 772, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 369, + 375, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Pre-training (VLP) models, such as CLIP [40] and BLIP [29], have shown great promise in learning transferable representations across various tasks. By aligning images and texts in a joint embedding space with a large corpus of paired image-text data, VLP models exhibit exceptional representation and generalization capabilities that surpass traditional task-specific models. Owing to this, the VLP models serve as foundation models for numerous tasks, including visual recognition [40], visual question answering [1,33], and image generation [41,47]. Moreover, these models can effectively integrate real-world visual inputs with humankind instructions, leading to their increasing use in physical-world applications, like autonomous driving [62] and embodied robotics [28,57].", + "bbox": [ + 212, + 401, + 785, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides their expressive power, VLP models have also shown excellent robustness under out-of-distribution (OOD) data [17, 40, 54, 60], including common corruptions [6, 15, 22], stylistic changes [21, 55], and natural distribution shifts [21, 23, 42]. However, a recent study [45] identifies that although VLP models excel at handling 2D-OOD samples, they suffer significant performance degradation under 3D viewpoint changes, revealing a notable shortcoming of the existing VLP models. As demonstrated in Fig. 1, when dealing with the benchmarks concerned with 3D viewpoint shifts [16, 46, 61], CLIP's performance is obviously lower than that on 2D-OOD benchmarks. This large gap likely stems from limited coverage of diverse viewpoints in the training datasets [18, 49, 53], which is crucial for learning viewpoint-invariant representations. As VLP models are increasingly deployed in real-world environments where viewpoint shifts often occur, enhancing their resilience to such changes is urgent and essential.", + "bbox": [ + 212, + 553, + 785, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, this paper sets out to enhance the viewpoint invariance of VLP models while preserving the original performance as much as possible. However, achieving this goal meets the following challenges: (1) Data scarcity: acquiring VLP training data that covers a wide range of viewpoint variations is particularly challenging compared to conventional image-text pair data. Although some datasets introduced for task-specific models do include viewpoint", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0600b22fc4662a7f0843e15114af70f943e67ed0e86a56fd09cc3570769254a0.jpg", + "image_caption": [ + "Fig. 2: Method Overview. (A) We create the first multi-view image caption dataset by collecting multi-view samples from existing 3D object and video datasets, and generating category-guided descriptions using VLLMs. (B) The proposed Omniview-Tuning takes multi-view image caption data as input, employs the cross-view alignment objective to encourage the model to learn viewpoint-invariant representations, and achieves efficient fine-tuning by updating Viformer and LoRA parameters." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 743, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "variations [4, 23, 35, 61], they often lack the textual descriptions vital for VLP. Even the largest available multi-view datasets [10, 46, 59] fall short in terms of scale, category coverage, and viewpoint diversity, thereby limiting the potential for VLP models to develop generalizable viewpoint-invariant representations. (2) Inappropriate paradigms: traditional approaches, which often regard viewpoint changes as adversarial attacks and employ adversarial training paradigms for enhancing invariance [2, 45, 46], are not entirely suitable for VLP models. Such frameworks typically entail a trade-off between robustness and accuracy—a balance that requires more careful consideration for foundation VLP models, where our aim is not solely to improve viewpoint invariance but, more importantly, to bridge the gap between it and the original performance. Furthermore, these approaches necessitate extra 3D reconstruction and neural rendering to capture adversarial viewpoints, leading to prohibitive computational costs for large-scale VLP models. For instance, tuning ResNet-50 with VIAT [46] under a dataset of just 1K objects demands around 400 GPU hours. Therefore, it is important to make training more efficient and less resource-intensive.", + "bbox": [ + 212, + 597, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the above discussions, this paper conducts a pioneering exploration of the viewpoint invariance of VLP models. Specifically, we address the aforementioned challenges by making the following contributions:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Million-scale multi-view image-text training set. We introduce a large-scale Multi-View Caption (MVCap) dataset tailored for viewpoint invariance of VLP models, comprising over 4.6 million multi-view image-text pairs across more than 100K objects. To assemble a diverse collection of multi-view image-text pairs, we amalgamate various 3D assets with real-world multi-view data. This process involves an extensive selection and rendering of multi-view images from existing datasets. We then utilize a Vision Large Language Model (VLLM) for automated caption generation to obtain semantically rich textual descriptions without extensive manual efforts. To ensure category consistency across varying viewpoints in the generated captions, we implement a category-guided prompting strategy, which maintains accuracy in textual descriptions for different viewpoints of the same object or scene (details in Sec. 3).", + "bbox": [ + 212, + 196, + 787, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Effective framework for enhancing VLP's viewpoint invariance. We propose Omniview-Tuning (OVT), a novel framework designed to enhance the viewpoint invariance of prevalent VLP models. As illustrated in Fig. 2, OVT employs multi-view image-text pairs for training additional learnable components. To amplify the model's proficiency in learning viewpoint-invariant representations, we introduce a Cross-viewpoint Alignment objective, ensuring that representations of the same object from different viewpoints are close and unified in the high-dimensional feature space. To prevent performance trade-offs due to the concept drift from aggressive viewpoint alignment, we innovatively construct the optimization paradigm of OVT in a minimax-like form. The optimization process includes identifying extreme outlier viewpoints during the maximization step, while optimizing the model's invariant representation for these outlier samples in the minimization step. This strategy enables the model to focus more on the worst-case viewpoint samples, thereby maximally preserving the original embedding distribution and avoiding performance degradation while saving computational costs. Moreover, OVT is designed in a Parameter-Efficient Fine-Tuning manner to improve efficiency, and creatively incorporates two trainable parameter modules: an embedding transformation module named VIForner and the Low-Rank Adaptation (LoRA [25]) weights, to acquire additional viewpoint invariance capabilities efficiently.", + "bbox": [ + 212, + 382, + 787, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extensive experiments across various VLP architectures and tasks. We conduct extensive experiments to show the efficacy of the OVT framework in improving the viewpoint invariance for VLP models while maintaining performance on clean data and 2D-OOD samples. For example, by fine-tuning CLIP with OVT on different architectures (ViT-B/32, ViT-B/16, and ViT-L/14), the Top-1 accuracy on viewpoint-OOD benchmarks increased by an average of $9.6\\%$ , $10.2\\%$ , and $8.9\\%$ , respectively, with only a minimal sacrifice on 2D-OOD benchmarks by an average of $2.6\\%$ , $1.4\\%$ , and $0.2\\%$ . Furthermore, serving as the visual encoder in VLLMs (e.g., LLaVa [33]), OVT-CLIP also effectively improves viewpoint invariance in image captioning and visual question answering tasks.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.1 Viewpoint Invariance and Robustness", + "text_level": 1, + "bbox": [ + 215, + 180, + 571, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Viewpoint invariance is a key property of human vision [5] but is usually lacking in computer vision models [2, 16]. Addressing viewpoint invariance and robustness involves strategies like data augmentation and adversarial learning. Early efforts aim to enhance viewpoint robustness by incorporating datasets enriched with viewpoint variations [4, 23, 35, 61]. For example, Madan et al. encourage models to learn viewpoint-robust representations by incorporating object-pose combinations [35]. However, these methods often falter under malicious viewpoint perturbations due to their inability to capture the worst-case viewpoint samples. Recently, achieving viewpoint invariance within the adversarial training paradigm has shown promise [2, 16, 20, 46]. By treating viewpoint variations as an adversarial attack, Alcorn et al. employ a differentiable renderer to train models against adversarial viewpoints optimized from a limited 3D objects set [2]. Recent studies, such as Viewfool [16] and VIAT [45, 46], have introduced neural radiance field (NeRF) [38, 39], enabling the characterization of adversarial viewpoint distributions from 2D multi-view inputs. Besides, studies in the self-supervised domain employ a label-free paradigm to improve viewpoint invariance, but these methods often require the introduction of complex structures, such as large-scale graph structures [56] and viewpoint generators [12]. Distinct from previous studies, our work pioneers the improvement of viewpoint invariance representation within large-scale VLP models, which is facilitated through suitable training data and refined fine-tuning methodologies.", + "bbox": [ + 212, + 209, + 787, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Vision-Language Pre-training", + "text_level": 1, + "bbox": [ + 215, + 553, + 506, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the realm of VLP, significant strides have been made in understanding and bridging the semantic gap between visual and textual information. Despite the variety of existing VLP paradigms, such as single-stream encoder (e.g., VisualBERT [31] and UNITER [8], etc.) or dual-stream encoder equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by ALIGN [30] and OpenAI's CLIP [40] dominates the field. CLIP, in particular, has gained widespread attention for its ability to perform zero-shot classification tasks by adopting a vast corpus of internet-collected image-text pairs, demonstrating the power of large-scale contrastive pre-training. Thus, Our investigation primarily focuses on these VLP architectures. Building upon these foundational works, subsequent iterations like open-CLIP [26], EVA-CLIP [51,52], and Meta-CLIP [58] have introduced nuanced enhancements. These refinements, ranging from the incorporation of more expansive high-quality image-text datasets and improved training methodologies, have collectively contributed to performance uplifts. BLIP [29], meanwhile, introduces a bootstrapping mechanism by the proposed captioner and filter module that achieve significant performance improvements on various downstream tasks.", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/79a000a71b7b79afe28a4be0f938f31b588006053ab792cde997d9ac788a9328.jpg", + "table_caption": [ + "Table 1: Comparison of current large-scale multi-view datasets. \"Spherical\" indicates whether the viewpoints cover spherical space, \"Diversity\" assesses the diversity of viewpoints, and \"Caption\" indicates whether textual descriptions are provided." + ], + "table_footnote": [], + "table_body": "
DatasetYear#Obj.#Cat.#Avg. View.#SampleImage DomainSphericalDiversityCaption
OOWL [24]201950025240120KRealX★★X
CO3D [43]202118.6k50~801.5MRealX★X
ABO [10]20227.9k6330238KSyntheticX★X
IM3D [46]20231.0k100100100KSynthetic✓★★★X
MVImgNet [59]2023219.0k238~306.5MRealX★★X
MVCap (Ours)202494.6k1600100/~304.6MSynthetic+Real✓★★★✓
", + "bbox": [ + 217, + 198, + 782, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Multi-view Caption Dataset", + "text_level": 1, + "bbox": [ + 215, + 295, + 522, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We recognize that one of the key challenges in achieving viewpoint invariance for VLP is the scarcity of training data that offer comprehensive viewpoints. As shown in Tab. 1, existing large-scale multi-view datasets [10,24,43,46,59] typically lack in either sample diversity, category breadth, or textual descriptions, limiting their effectiveness for supporting VLP models to achieve viewpoint invariance. To address these limitations, we next introduce the MVCap dataset.", + "bbox": [ + 212, + 325, + 782, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Multi-View Image Collection", + "text_level": 1, + "bbox": [ + 215, + 438, + 501, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We commence by gathering a multi-view image collection $\\mathcal{D} = \\{I_{ij} \\mid i = 1,2,\\dots,N; j = 1,2,\\dots,M_i\\}$ , where $N$ and $M_i$ represent the counts of objects and their viewpoints, respectively. To cover various categories from virtual to real-world scenes, we integrate samples from Objaverse [13], IM3D [46], and MVImgNet [59]. Since the original 3D dataset includes a fair share of noisy and semantically indistinct objects, we leverage semantic embeddings provided by OpenShape [34] to conduct cosine similarity sorting based on the embeddings of customized labels. Finally, we filter 24,495 virtual 3D objects endowed with distinct semantic clarity and cover over 1,600 categories. For each chosen 3D object, we employ Blender to render 100 random viewpoint images from the upper hemisphere, ensuring a comprehensive and varied viewpoint representation in our collected samples. We also incorporate objects from MVImgNet with over 30 valid viewpoints (video frames), thereby acquiring a substantial number of real-world multi-view samples to enrich the dataset's content and quality further.", + "bbox": [ + 212, + 460, + 787, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Category-Guided Caption Generation", + "text_level": 1, + "bbox": [ + 215, + 694, + 571, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The granularity and precision of textual descriptions are pivotal in VLP training, as they influence the model's generalization capabilities and the variety of visual concepts learned. Relying solely on simple prompt engineering, such as \"a photo of [category],\" may introduce biases and limit the model's generalizability, whereas manual annotation is costly. To circumvent this, we utilize InstructBLIP-flant5xl [11], a leading VLLM, to create multi-view captions automatically. However, such VLLMs also grapple with viewpoint invariance, where the model's responses to different viewpoints can often be category-inconsistent,", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56f637aa9ccc719991a6d62e8b11923aab6c6fdc5e0f236ed0833ff97e337b0d.jpg", + "image_caption": [ + "Fig. 3: Generated multi-view captions with common and category-guided prompts." + ], + "image_footnote": [], + "bbox": [ + 227, + 150, + 429, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dea7630e8ef29e614e56232b11409b5ba7ca3294d6375173085d9033740b4f06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 148, + 555, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3021dcae3c64a46f248edda2cd7f61114ed0fabe1aa4b3a642d39c83ae826b9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 152, + 776, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as depicted in Fig. 3. This situation presents a \"chicken or egg\" dilemma: we hope to use a viewpoint-invariant model to supply data for viewpoint invariance training. We address this by the design of category-guided prompting. Specifically, we use prompts containing ground-truth category information to eliminate the hallucination of large VLLMs in response to viewpoint-shifted inputs, thereby generating category-consistent multi-view captions. Formally, the forward process for generating captions can be represented as follows:", + "bbox": [ + 212, + 265, + 784, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$T_{ij} = \\mathcal{G}[I_{ij},\\mathrm{Prompt}(c_i)]$ Prompt $(c_{i}) = \"$ Write a short description for the image, (1) noting that the main instance of the image is $a < c_{i} > .\"$", + "bbox": [ + 294, + 381, + 784, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c_{i} \\in C$ denotes the category label for the $i$ -th object, and $\\mathcal{G}$ denotes the forward process of InstructBLIP. This yields the multi-view image-text pairs $\\tilde{\\mathcal{D}} = \\{\\langle I_{ij}, T_{ij} \\rangle \\mid i = 1, 2, \\dots, N; j = 1, 2, \\dots, M_i\\}$ , which can be utilized for the viewpoint invariance fine-tuning of VLP models.", + "bbox": [ + 214, + 439, + 784, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 omniview-Tuning", + "text_level": 1, + "bbox": [ + 215, + 522, + 424, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Preliminaries: Contrastive Vision-Language Pre-training", + "text_level": 1, + "bbox": [ + 214, + 554, + 725, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Despite the variety of existing VLP paradigms, such as single-stream or dual-stream architectures equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by CLIP et al. [30, 40] dominates the field. Thus, Our investigation primarily focuses on this VLP architecture.", + "bbox": [ + 212, + 579, + 784, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without the loss of generality, these VLP models are composed of a visual encoder $E_{\\mathbf{W}_{\\mathbf{v}}}: I \\to z^{I} \\in \\mathbb{R}^{d}$ and a text encoder $E_{\\mathbf{W}_{\\mathbf{t}}} : T \\to z^{T} \\in \\mathbb{R}^{d}$ , which maps visual and textual inputs to a unified high-dimensional feature space $\\mathbb{R}^d$ , respectively, where $\\mathbf{W}_{\\mathbf{v}}$ and $\\mathbf{W}_{\\mathbf{t}}$ are weight matrices of two encoders. Given a large corpus of image-text pairs $\\{\\langle I_i, T_i \\rangle\\}_{i=1}^N$ , VLP models typically employ an image-text contrastive (ITC) loss as the training objective:", + "bbox": [ + 212, + 638, + 784, + 729 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I T C} = \\frac {1}{2} \\left(\\mathcal {L} _ {I \\rightarrow T} + \\mathcal {L} _ {T \\rightarrow I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 739, + 784, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is composed of an image-to-text and a text-to-image terms formulated as:", + "bbox": [ + 214, + 768, + 782, + 782 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {I \\rightarrow T} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(d \\left(z _ {i} ^ {I} , z _ {i} ^ {T}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(d \\left(z _ {i} ^ {I} , z _ {k} ^ {T}\\right) / \\tau\\right)}, \\tag {3} \\\\ \\mathcal {L} _ {T \\to I} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (d (z _ {i} ^ {T} , z _ {i} ^ {I}) / \\tau)}{\\sum_ {k = 1} ^ {N} \\exp (d (z _ {i} ^ {T} , z _ {k} ^ {I}) / \\tau)}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 794, + 784, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ represents a learnable temperature parameter, $z^I$ and $z^T$ denote the image and text embeddings, respectively. The $\\mathcal{L}_{ITC}$ maximizes the similarity between matched image-text pairs while minimizing the similarity for mismatched pairs, thus enabling the alignment of visual and textual information to the same feature space, bringing the embeddings of matched pairs closer. Following [29, 30, 40], the proposed Omniview-Tuning implements $\\mathcal{L}_{ITC}$ for aligning multi-view images with text modalities, which is explained in the next section.", + "bbox": [ + 212, + 145, + 787, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 273, + 439, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Viewpoint Invariance of Vision-Language Pre-training. In computer vision scenario, viewpoint invariance implies that model $f(\\cdot)$ can provide consistent predictions or representations given any different views of the identical object or scene [46]. Formally, given a collection of multi-view images $\\mathcal{D} = \\{I_{ij} \\mid i = 1,2,\\dots,N; j = 1,2,\\dots,M_i\\}$ , viewpoint invariance is required:", + "bbox": [ + 212, + 299, + 785, + 376 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(I _ {i j}\\right) = f \\left(I _ {i j ^ {\\prime}}\\right), \\quad \\forall i, j, j ^ {\\prime} \\text {w i t h} j \\neq j ^ {\\prime}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 387, + 785, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $i$ is the index of the object/scene, $j$ and $j'$ are indexes of two viewpoint samples. However, in the context of dual-stream VLP models, this concept requires a more refined interpretation. For VLP models, viewpoint invariance necessitates that the visual representations (i.e., the embeddings inferred from the visual encoder) from different viewpoints be sufficiently close in the feature space. Assuming $I_{ij}$ and $I_{ij'}$ are images from different viewpoints of the same object, this requirement can be formulated as follows:", + "bbox": [ + 212, + 414, + 787, + 521 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd \\left[ E _ {\\mathbf {W} _ {\\mathbf {v}}} \\left(I _ {i j}\\right), E _ {\\mathbf {W} _ {\\mathbf {v}}} \\left(I _ {i j ^ {\\prime}}\\right) \\right] \\leq \\epsilon , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 532, + 785, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $d(\\cdot)$ denotes a distance metric in the representation space, such as cosine distance, $\\epsilon$ represents the maximum variance allowed.", + "bbox": [ + 212, + 566, + 785, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization Objectives of Omniview-Tuning. Although images from different viewpoints often correspond to slightly varying textual descriptions, influenced by context, grammatical structure, and linguistic ambiguity, this variation could be significantly amplified in the high-dimensional representation space [9, 44]. Therefore, relying solely on image-text alignment may not suffice to adequately align embeddings from different viewpoints. Starting from the definition of viewpoint invariance, we introduce a cross-viewpoint alignment objective within the $\\mathcal{L}_{ITC}$ to directly encourage the model to learn invariant representations between different viewpoints, rather than relying on the indirect alignment through textual descriptions. This can be seen as a regularization that forces the model to obtain viewpoint invariance, even when such invariance is not explicitly articulated in the textual descriptions. With this consideration, given a multi-view training set $\\mathcal{D}$ , the optimization problem is defined as follows:", + "bbox": [ + 212, + 597, + 787, + 792 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {W} _ {\\mathbf {v}}, \\mathbf {W} _ {\\mathbf {t}}} \\left[ \\mathcal {L} _ {I T C} + \\lambda \\cdot \\underbrace {\\sum_ {i} \\sum_ {j \\neq j ^ {\\prime}} d \\left(z _ {i j} ^ {I} , z _ {i j ^ {\\prime}} ^ {I}\\right)} _ {\\mathcal {L} _ {V C}} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 806, + 785, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the first term represents the image-text alignment used in the pre-training process, while the second term signifies the cross-viewpoint alignment goal mentioned above, referred to as Viewpoint Consistency loss $(\\mathcal{L}_{VC})$ , which aims to minimize the cosine distance between embeddings from different viewpoints. $\\lambda$ is a hyperparameter that balances the importance of two loss terms.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Optimization Strategy", + "text_level": 1, + "bbox": [ + 215, + 241, + 447, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, the naive way to achieve viewpoint invariance is to calculate the loss terms in Eq. (6) based on the forward process of encoders, then update the encoders' weight using gradient descent. However, it has a relatively high time complexity to solve Eq. (6) because current $\\mathcal{L}_{VC}$ requires iterating over every possible combination of viewpoints. Therefore, we endeavor to provide a more effective implementation for the original optimization problem. Drawing from the advantages of adversarial training [36, 46], we frame the optimization of the $\\mathcal{L}_{VC}$ in a minimax format, rewriting the original problem Eq. (6) as:", + "bbox": [ + 212, + 263, + 787, + 387 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\mathbf {W} _ {\\mathbf {v}}, \\mathbf {W} _ {\\mathbf {t}}} \\left[ \\mathcal {L} _ {I T C} + \\lambda \\cdot \\underbrace {\\max _ {\\mathcal {O} = \\{O _ {i} \\} _ {i = 1} ^ {N} , | O _ {i} | = K} \\sum_ {i = 1} ^ {N} \\sum_ {j \\in \\mathcal {O}} l \\left(z _ {i j} ^ {I} , z _ {C _ {i}} ^ {I}\\right)} \\right], \\tag {7} \\\\ \\mathrm {w h e r e} l (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) = \\max \\left[ d (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) + m, 0 \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 393, + 785, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{O} = \\{O_i\\}_{i=1}^N$ is the outlier viewpoints set, $z_{C_i}^I$ are anchor viewpoint embeddings of each object, and $l(\\cdot)$ is the cosine distance with a margin $m$ . During the optimization, The maximization step first identifies the collection of top- $K$ outlier viewpoints $\\mathcal{O}$ , which are the viewpoint samples with the highest degree of representational deviation. Then, the minimization step encourages the outlier viewpoint embeddings to converge towards corresponding anchor viewpoint embeddings $z_{C_i}^I$ . We obtain $z_{C_i}^I$ by calculating the nearest-neighbor weighted embedding centroid of each object:", + "bbox": [ + 212, + 470, + 787, + 592 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} z _ {C _ {i}} ^ {I} = \\sum_ {j = 1} ^ {M _ {i}} \\tilde {\\omega} _ {i j} \\cdot z _ {i j} ^ {I}, \\\\ \\mathrm {w h e r e} \\tilde {\\omega} _ {i j} = \\omega_ {i j} / \\sum_ {j} \\omega_ {i j}, \\omega_ {i j} = 1 / \\sum_ {z _ {i h} ^ {I} \\in \\mathcal {Q} _ {i j}} d (z _ {i j} ^ {I}, z _ {i h} ^ {I}) \\end{array} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 598, + 785, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}_{ij} = \\{z_{ih}^I\\}_{h=1}^5$ is the top-5 nearest neighbours of each viewpoint embedding $z_{ij}^I$ . As for the outlier viewpoints set, we define them as the viewpoints with the top- $K$ farthest cosine distances from $z_{C_i}^I$ .", + "bbox": [ + 212, + 643, + 787, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The adoption of this strategy offers dual advantages: Firstly, it allows the model to focus solely on extreme outlier viewpoints, preventing concept drift and potential overfitting to the fine-tuning dataset that results from excessive alignment. Secondly, this approach reduces computational overhead and significantly enhances optimization efficiency.", + "bbox": [ + 212, + 691, + 787, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Parameter-Efficient Modules", + "text_level": 1, + "bbox": [ + 215, + 787, + 500, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To mitigate the impact of full parameters update on the original performance and enhance training efficiency, we achieve viewpoint invariance by efficiently", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/aa8940cfe67daeb44d280088a90f02feb57c89777a4f40bbcfde2385f8723805.jpg", + "image_caption": [ + "Fig. 4: Visualization results for zero-shot classification results." + ], + "image_footnote": [], + "bbox": [ + 235, + 146, + 769, + 277 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "fine-tuning the parameters of the visual encoder while keeping the text encoder frozen. Inspired by LoRA [25], we perform low-rank decomposition on the weights of the visual encoder $\\mathbf{W}_{\\mathbf{v}} \\in \\mathbb{R}^{n \\times m}$ to substitute full-parameter update:", + "bbox": [ + 212, + 316, + 784, + 363 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {W}} _ {\\mathbf {v}} = \\mathbf {W} _ {\\mathbf {v}} + \\Delta W = \\mathbf {W} _ {\\mathbf {v}} + \\mathbf {B} \\mathbf {A}, \\quad \\text {w h e r e} \\mathbf {B} \\in \\mathbb {R} ^ {m \\times r}, \\mathbf {A} \\in \\mathbb {R} ^ {r \\times n}, r \\ll \\min (n, m), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 373, + 784, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ and $\\mathbf{B}$ are two learnable low-rank parameter matrices, which we apply to the self-attention layers of the visual encoder and update them during fine-tuning while freezing the original pre-trained weights. Compared to directly adjusting the original network weights (i.e., full-parameter fine-tuning manner in TeCoA [37] and FARE [48]), this approach enables us to improve the model's viewpoint invariance representation capability while better preserving the original performance. Drawing inspiration from the success of CLIP-Adapter [19], which improves CLIP's performance in few-shot scenarios by introducing linear layers after the encoder, we propose a similar module called VIformer: $f_{\\pmb{\\theta}}:z^{I}\\in \\mathbb{R}^{d}\\to s^{I}\\in \\mathbb{R}^{d}$ , where $\\pmb{\\theta}$ is the weight. Unlike CLIP-Adapter, VIformer transforms the original embeddings $z^{I}$ by introducing self-attention layers in a learnable manner to extract and retain specific viewpoint-invariant key components $s^I$ . Combining LoRA and VIformer modules, the forward process of image encoding can be represented as follows:", + "bbox": [ + 212, + 405, + 787, + 617 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {z} ^ {I} = \\alpha \\cdot f _ {\\boldsymbol {\\theta}} (z ^ {I}) + (1 - \\alpha) \\cdot z ^ {I} \\\\ = \\alpha \\cdot f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {W} _ {\\mathbf {v}} \\cdot I + \\mathbf {B} \\mathbf {A} \\cdot I\\right) + (1 - \\alpha) \\cdot \\left(\\mathbf {W} _ {\\mathbf {v}} \\cdot I + \\mathbf {B} \\mathbf {A} \\cdot I\\right), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 627, + 784, + 664 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where the constant value $\\alpha$ denotes the residual ratio to balance achieving original performance and viewpoint invariance performance. Therefore, for Eq. (6), we now only need to update $\\mathbf{A}$ , $\\mathbf{B}$ , and $\\pmb{\\theta}$ , rather than the entire weights of VLP. To facilitate the understanding of the OVT training process, we provide the pseudocode for OVT as shown in Appendix D", + "bbox": [ + 212, + 676, + 787, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 776, + 375, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our evaluation of Omniview-Tuning spans several downstream tasks, including zero-shot classification, image captioning, and vision question answering.", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 269, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cd6a3ea38e44fda31d8d47dc867ff4abf99e196c7be31ff90b75ba0d609facef.jpg", + "table_caption": [ + "Table 2: Configurations of OVT and zero-shot Top-1 accuracy (\\%) on ImageNet-1K with ImageNet-V+. The number in parentheses shows the performance change relative to the pre-trained weights. Through OVT training, each model maintains the performance on ImageNet-1K (IN-1K) while significantly improving the performance on ImageNet-V+ (IN-V+.), narrowing the performance gap." + ], + "table_footnote": [], + "table_body": "
ModelPretrain WeightPretrain DataTotal #Param.Trainable #Param.Image SizeBatch Size#Iter.IN-1KIN-View+
OVT-OpenCLIP ViT-B/32OpenCLIP-ViT-B/32LAION (2B)151M6.6M22451235k67.8 (↑1.3)59.5 (↑22.4)
OVT-OpenCLIP ViT-B/16OpenCLIP-ViT-B/16LAION (2B)149M6.6M22451235k69.7 (↑2.1)61.7 (↑17.5)
OVT-OpenCLIP ViT-L/14OpenCLIP-ViT-L/14LAION (2B)428M11.8M22425620k77.3 (↑2.1)69.8 (↑16.6)
OVT-MetaCLIP ViT-B/32MetaCLIP-ViT-B/32Common crawl (2.5B)151M6.6M22451240k69.7 (↑2.1)54.8 (↑13.8)
OVT-MetaCLIP ViT-B/16MetaCLIP-ViT-B/16Common crawl (2.5B)149M6.6M22451240k73.8 (↑1.7)64.8 (↑15.2)
OVT-MetaCLIP ViT-L/14MetaCLIP-ViT-L/14Common crawl (2.5B)428M11.8M22425620k77.7 (↓1.4)75.4 (↑9.0)
OVT-BLIP ViT-B/16Salesforce-BLIP-ViT-B/16COCO et al. (129M)234M4.3M22425620k61.7 (↑8.8)54.8 (↑18.0)
", + "bbox": [ + 215, + 226, + 785, + 327 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For zero-shot classification (Sec. 5.1), we conduct evaluations for CLIP [40] and BLIP [29] architectures. For image captioning and vision question answering (Sec. 5.2), we replace the visual encoders in Vision Large Language Models (VLLMs) with our fine-tuned versions. We adopt LLaVA-1.5 [32,33], and OpenFlamingo [3], the most advanced open-source VLLMs available. Additionally, we present the ablation study and convergence analysis of our approach in Sec. 5.3.", + "bbox": [ + 212, + 354, + 787, + 448 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Evaluation of Zero-Shot Classification", + "text_level": 1, + "bbox": [ + 214, + 468, + 571, + 482 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We adopt the official CLIP (OpenAI CLIP [40]) and the community open-source version (OpenCLIP [26]) as baselines. Additionally, we include the current state-of-the-art Eva02-CLIP [51] and MetaCLIP [58] as another set of baselines to compare CLIP trained with improved techniques and extensive training data. For BLIP, we use the official implementation [29] as the baseline. Settings. We train two series of CLIP models using our OVT framework and MVCap dataset, each series comprising three different visual encoder architectures (ViT-B/32, ViT-B/16, and ViT-L/14). OVT-OpenCLIP are fine-tuned on the original weights of OpenCLIP, while OVT-MetaCLIP are based on MetaCLIP. The fine-tuning settings are detailed in Tab. 2. We standardized the $\\lambda = 1.0$ , $\\alpha = 0.1$ , and the outlier viewpoints number $K = 5$ and set the LoRA rank at 8. Datasets and Metrics. We employ a various set of benchmarks for evaluation, including clean data distributions [14, 27], common 2D-OOD [21, 23, 42, 55, 61], and most importantly, viewpoint-OOD [7, 16, 46, 61]. For each benchmark, we report Top-1 and Top-5 accuracy and average accuracy across all benchmarks. The evaluations follow the standard prompting engineering of CLIP [40].", + "bbox": [ + 212, + 492, + 785, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results and Discussions. Tab. 3 summarizes the performance of our OVT-trained VLP models against various VLP versions, including their accuracy across different benchmarks and average accuracy across clean, common-OOD, and Viewpoint-OOD domains. We can draw the following conclusions:", + "bbox": [ + 212, + 734, + 785, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(1) OVT significantly enhances the models' invariance to Viewpoint-OOD samples. Across different VLP architectures and visual encoders, OVT-trained models perform best on almost all viewpoint-OOD benchmarks. On the average", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b2e138e934d2a949bc324e3c70bcfc264ae7adf9ce2a413c9460f3665eff4b27.jpg", + "table_caption": [ + "Table 3: Top-1/Top-5 zero-shot accuracy (%) under different benchmarks" + ], + "table_footnote": [], + "table_body": "
ModelCleanCommon-ODViewpoint-OD
ImageNet-100 [66]ImageNet-IR [14]Cllr-100 [27]Avg. Acc.ImageNet-Vi [14]ImageNet-Sch. [53]ImageNet-ODD [25]ImageNet-Rem. [21]OOD-CV [61]Avg. Acc.ImageNet-View [68]ImageNet-View + [68]OOD-CV-Pose [63]MIBO [7]Avg. Top-1Total Avg. Acc.
A. Comparisons with VIT-B/32 baselines
OpenAI CLIP77.5/93.963.3/88.864.3/88.168.4/90.255.8/83.442.2/70.333.4/62.250.7/75.450.2/82.646.5/74.844.5/65.427.5/52.447.2/84.526.5/59.436.4/65.448.6/75.5
Open CLIP81.1/95.366.5/89.975.8/94.074.5/93.058.1/83.953.6/79.334.8/64.461.0/81.953.5/81.952.2/78.354.4/72.137.1/63.246.9/81.633.0/69.242.8/71.554.6/79.7
OVT-OpenCLIP80.9/95.667.8/90.865.0/89.371.2/91.958.0/84.245.8/73.442.8/75.050.3/71.451.7/79.549.7/76.761.9/81.259.5/85.652.8/82.535.4/80.152.4/82.456.0/82.4
MetaCLIP80.7/95.667.6/90.577.7/95.275.3/93.859.5/85.455.9/81.432.4/62.563.2/83.852.0/84.252.6/79.561.4/76.741.0/67.848.9/87.934.8/73.246.5/76.456.3/82.0
OVT-MetaCLIP80.7/95.669.7/92.071.8/93.074.0/93.560.6/85.847.8/75.843.5/73.849.0/70.850.1/80.150.2/77.264.0/79.254.8/80.455.1/84.835.6/77.052.4/80.356.9/82.3
B. Comparisons with VIT-B/16 baselines
OpenAI CLIP82.1/95.768.3/91.967.2/89.472.5/92.361.8/87.448.2/76.327.7/55.759.1/83.052.2/84.649.8/77.451.6/68.936.9/63.853.4/86.830.1/66.143.0/71.453.2/79.1
Open CLIP83.2/96.270.1/91.877.0/94.876.8/94.362.2/87.056.0/82.030.7/59.864.9/85.654.3/82.753.6/79.458.1/74.444.2/70.948.5/84.034.6/74.646.4/76.057.0/82.0
OVT-OpenCLIP83.9/97.071.9/93.169.0/90.774.9/93.664.0/88.650.5/77.936.8/68.957.0/77.256.3/84.552.9/79.465.4/80.761.7/85.856.9/87.442.4/84.956.6/84.759.6/84.7
EVA-CLIP85.3/96.574.6/94.287.5/98.082.5/96.367.0/89.857.6/82.321.3/47.369.6/87.553.1/83.153.7/78.061.8/76.644.3/69.453.9/87.432.9/73.248.2/76.659.1/82.1
MetaCLIP84.3/97.272.1/93.478.9/95.478.4/95.365.0/89.360.1/84.826.2/56.470.2/89.352.3/85.454.8/81.064.2/79.449.6/76.148.9/90.938.5/78.750.3/81.259.2/84.7
OVT-MetaCLIP83.4/97.473.8/94.173.9/93.677.0/95.065.9/89.453.6/81.036.2/68.859.0/79.651.6/83.853.2/80.169.7/84.064.8/87.355.2/87.839.2/82.957.2/85.560.5/85.6
C. Comparisons with VIT-L/14 baselines
OpenAI CLIP86.5/97.475.4/94.676.5/93.379.5/95.169.8/90.959.5/84.318.6/43.872.8/91.452.9/88.854.7/79.860.3/75.645.8/71.547.9/88.238.0/74.148.0/77.358.6/82.8
Open CLIP86.8/97.875.2/94.383.7/96.781.9/96.267.7/90.263.2/86.424.0/50.574.5/91.254.5/85.056.8/80.665.7/78.153.2/76.752.4/90.542.3/83.053.4/82.161.9/85.0
OVT-OpenCLIP89.0/97.877.3/95.379.2/95.381.8/96.169.6/91.561.9/86.027.5/55.471.3/88.756.4/87.057.3/81.772.2/86.669.8/89.757.3/94.150.0/89.362.3/89.965.1/88.1
EVA-CLIP88.5/97.979.6/96.090.6/98.686.3/97.572.8/92.768.0/89.116.3/40.082.8/95.754.7/87.458.9/81.071.5/82.361.1/81.754.4/94.539.6/86.156.6/86.165.0/86.8
MetaCLIP88.3/98.379.1/95.984.1/96.983.8/97.072.5/92.668.9/89.817.0/40.681.8/95.156.6/87.559.3/81.177.3/89.366.4/87.058.9/93.448.1/89.662.7/89.866.6/88.0
OVT-MetaCLIP88.8/97.577.7/95.984.0/96.983.5/96.870.8/92.264.4/87.920.8/47.077.0/92.756.3/89.357.8/81.879.3/90.675.4/93.057.0/94.446.4/93.864.5/92.966.5/89.3
D. Comparisons with BLIP VIT-L/B16 baselines
BLIP76.6/93.352.9/80.267.0/88.365.5/87.347.3/74.751.0/76.625.6/53.464.3/83.853.9/87.648.4/75.255.2/68.236.8/63.350.8/89.227.0/66.142.4/71.950.7/77.1
OVT-BLIP82.2/97.061.7/88.866.6/88.970.2/91.553.7/82.946.5/74.233.8/62.757.4/77.956.4/87.349.6/77.062.6/79.054.8/79.955.2/89.531.5/73.251.0/80.455.2/81.8
", + "bbox": [ + 215, + 171, + 785, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "accuracy of viewpoint-OOD datasets, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 shows improvements of $9.6\\%$ , $10.2\\%$ , and $8.9\\%$ over OpenCLIP, respectively. OVT-BLIP demonstrated an average improvement of $8.6\\%$ . (2) While enhancing viewpoint invariance, OVT maintains performance on clean samples and 2D-OOD without significant performance trade-offs. For 2D-OOD benchmarks, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 sacrifice only $2.6\\%$ , $1.4\\%$ , and $0.2\\%$ accuracy. (3) Compared to earlier CLIP baselines, the recently developed MetaCLIP exhibits better zero-shot performance. Based on this, OVT further enhances its performance under viewpoint-OOD samples. Visualization. We showcase OVT-OpenCLIP and the original OpenCLIP prediction on several viewpoint-OOD samples. As illustrated in Fig. 4, OVT-CLIP successfully predicts the categories of images from various unusual viewpoints in all cases, whereas the original CLIP is prone to make incorrect predictions.", + "bbox": [ + 212, + 472, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Performance on Other Tasks", + "text_level": 1, + "bbox": [ + 215, + 691, + 500, + 707 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Settings. As LLaVA and Openflamingo use the OpenAI CLIP (ViT-L/14) to encode vision inputs, we applied OVT to this model in this section and comparing with other OpenAI CLIP (ViT-L/14) versions in image captioning tasks. The training setup remains consistent with the OVT-OpenCLIP described in Tab. 2. Baselines. In addition to comparing with the original OpenAI CLIP version, we also select $TeCoA^4$ [37] and $FARE^4$ [48], robust CLIP models based on adversarial training, as baselines, which have been proven to possess good resistance to adversarial samples in image captioning task.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ec503e52c3f4437a948ad01c15cd1eb3fab20f684a30a729f9698efb2c8c010b.jpg", + "table_caption": [ + "Table 4: Image captioning performance under clean distribution samples and viewpoint-OOD samples from Real-world and Synthetic domains. We utilize the MP-Net [50] to calculate the similarity between generated descriptions and ground-truth labels, considering predictions successful if they exceed the similarity threshold $\\beta$ ." + ], + "table_footnote": [], + "table_body": "
ModelVisual EncoderReal-world DomainSynthetic Domain
OOD-CV (iid) [61]β@1.0β@0.5β@Adp.OOD-CV (Pose) [61]β@1.0β@0.5β@Adp.IM3D [46]β@1.0β@0.5β@Adp.ImageNet-V+ [46]β@1.0β@0.5β@Adp.
LLaVa-7bOpenAI CLIP(ViT-L/14)44.161.167.546.453.658.746.753.358.820.425.532.1
TeCoA4[37](ViT-L/14)41.958.965.536.141.649.226.330.142.68.711.622.6
FARE4[48](ViT-L/14)42.158.965.240.245.950.835.239.249.212.715.823.1
OVT-CLIP(ViT-L/14)43.559.565.946.553.659.149.454.061.826.431.941.0
LLaVa-13bOpenAI CLIP(ViT-L/14)45.468.070.648.658.660.848.756.760.821.228.432.5
TeCoA4[37](ViT-L/14)42.467.072.237.448.951.325.028.641.58.410.921.8
FARE4[48](ViT-L/14)43.966.771.141.952.154.836.141.448.612.115.920.8
OVT-CLIP(ViT-L/14)45.767.370.848.258.661.950.458.963.226.436.240.9
", + "bbox": [ + 217, + 212, + 785, + 339 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/966b2b7e73c38d4bdda01e628a71669dd308034e071e2454b3036d4c2bdc62e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 358, + 272, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d6e47f114e0cbdcf420d52e2e3c71b928caa97b4f83d166f578b68c359a6600f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 359, + 401, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/469b1c5701e4df162717f0773a2816158e876dbec0c8babe70dd526650b27211.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 359, + 459, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7e2383e2660a341ad04013475d3c73368e55a530b4d8866c0b193260ac4ed703.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 359, + 584, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c506a10688a0963e44f542dea37037a56ecc7232abe0974915d44eefebcbaf64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 396, + 272, + 430 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/85a4559c15e28c2b0274591aa8605fffbb3b51938ba907a5639068216059ff65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 396, + 401, + 430 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6a3500f5ad9bddd5d8d6004f5daa241989d16620e66758c9ef59d6a0a65621d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 396, + 459, + 430 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/50f184330a8a2503329c2ffb537eebe88295d6837189dbc57cb23663fa823fe2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 396, + 584, + 430 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/328b2d77f496f5b4420c9f5f53bb5b478edaef7b62477e9d772192f08ee72d5e.jpg", + "image_caption": [ + "Fig. 5: The image descriptions generated by LLaVa-13B using our OVT-CLIP and the original OpenAI CLIP as vision encoder, where red texts indicates incorrect category descriptions, and green texts represents correct." + ], + "image_footnote": [], + "bbox": [ + 220, + 431, + 274, + 468 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bb53e203ab2bfc77f45d51e3c11e732f2590994e2dc9e5ba9a5fe532b89b74a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 431, + 401, + 469 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/565e0550f42ea728db8158e630a725072a6e02952c402d5984a985e90d9eade4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 431, + 450, + 468 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1f9962b8863478dbe866f31889ade1c338cd7bc6bfe62ec33607f41560425b63.jpg", + "image_caption": [ + "Fig.6: The Top-1 accuracy of OVT-OpenCLIP (ViT-B/16) with the iterations increases." + ], + "image_footnote": [], + "bbox": [ + 462, + 431, + 584, + 469 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/02873cab1d52a52a340c73d57b93b0f413bce70449cb1e6828f51b84cff74dea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 358, + 781, + 468 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Datasets and Metrics. Given the absence of caption benchmarks that include viewpoint-changing OOD samples, we conduct evaluations using existing viewpoint-OOD datasets, including real-world datasets (using OOD-CV (iid) to represent clean distribution and OOD-CV (Pose) for viewpoint-OOD) and synthetic datasets (using IM3D [46] for clean distribution and ImageNet-V+ for viewpoint-OOD). We adopt word embedding distance to calculate the accuracy of the captioning task. By adopting MPNet [50], a state-of-the-art textual embedding model, we measure the similarity between keywords in the generated description and the ground-truth categories. Then assess the accuracy by counting the number of samples that exceed a specific similarity threshold $\\beta$ .", + "bbox": [ + 212, + 551, + 787, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results and Discussions. Tab. 4 shows the image captioning accuracy of CLIP models under different training strategies, considering $\\beta$ at 1.0 (indicating predictions involve ground-truth categories), 0.5, and $Adp$ . (meaning $\\beta$ is equal to the average similarity in the clean distribution). We found that OVT-CLIP improves the accuracy of descriptions generated by LLaVa for viewpoint-OOD samples while maintaining its performance on corresponding clean distributions. When used as the visual encoder for the LLaVa-7B model, OVT-CLIP achieved an $8.9\\%$ increase in accuracy compared to the original CLIP model weights under $\\beta = Adp$ . Besides, we find that although robust CLIP versions maintain perfor", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/42043bd602183900b6ab184896c5e6da8f315e70c06c7bb769fa2988c8b4fdb1.jpg", + "table_caption": [ + "Table 5: Average Top-1/Top-5 zero-shot accuracy (\\%) under different data distributions within various ablation settings." + ], + "table_footnote": [], + "table_body": "
LITCVIFormerLVCTotal Avg.Clean Avg.Common-OOD Avg.Viewpoint-OOD Avg.
XXX61.985.081.996.256.880.653.482.1
✓XX61.985.7 (↑0.7)79.4 (↓2.5)95.6 (↓0.6)56.2 (↓0.6)81.4 (↑0.8)56.0 (↑2.6)83.8 (↑1.7)
✓✓X62.2 (↑0.3)86.2 (↑1.2)79.9 (↓2.0)95.4 (↓0.8)55.4 (↓1.4)81.6 (↑1.0)57.5 (↑4.1)85.2 (↑3.1)
✓✓✓65.1 (↑3.2)88.1 (↑3.1)81.8 (↓0.1)96.1 (↓0.1)57.3 (↑0.5)81.7 (↑1.1)62.3 (↑8.9)89.9 (↑7.8)
", + "bbox": [ + 217, + 184, + 785, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "mance on clean distribution samples, they experience a significant performance decline when facing viewpoint-OOD samples. We select some examples with the generated description in Fig. 5. For the visual question-answering task, we used OpenFlamingo as the VLLMs. The results are reported in the Appendix A.", + "bbox": [ + 212, + 279, + 784, + 339 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.3 Ablation Studies and Additional Results", + "text_level": 1, + "bbox": [ + 215, + 363, + 594, + 378 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our ablation studies focus on the VIFormer and the $\\mathcal{L}_{VC}$ within the Omniview-Tuning framework. Tab. 5 shows the Top-1/Top-5 acc of OVT-OpenCLIP (ViT-L/14) across various data distributions and different ablation settings. Beyond the original OpenCLIP, we set a baseline that only uses $\\mathcal{L}_{ITC}$ for fine-tuning. Keeping other training settings fixed, reliance solely on $\\mathcal{L}_{ITC}$ led to a more significant performance decline in clean and 2D-OOD samples while achieving limited viewpoint OOD performance improvement (2.6%/1.7%). The integration of VI-Former led to further improvements in viewpoint OOD accuracy (4.1%/3.1%). With the further addition of $\\mathcal{L}_{VC}$ , the improvement in viewpoint OOD performance is most significant (8.9%/7.8%), and it also reduces performance sacrifices in other data distributions. More ablation analyses are available in Appendix B. Furthermore, we report OVT's training convergence, depicted in Fig. 6. We display the Top-1 accuracy evolution for OVT-OpenCLIP (ViT-B/16) across various training iterations. We observe that around 40K iterations, with a batch size of 512, are sufficient for effective convergence, thus achieving a balance in performance across different data distributions.", + "bbox": [ + 212, + 388, + 787, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 656, + 367, + 671 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To tackle the challenge of 3D viewpoint in VLP models, this paper first introduced the MVCap dataset, a million-scale collection of image-text pairs with diverse viewpoint variations. Building upon this groundwork, we then proposed the Omniview-Tuning framework, which incorporates a novel Cross-Viewpoint Alignment objective in a parameter-efficient manner, effectively enhancing the VLP models' ability to generate viewpoint-invariant representations. Moreover, through extensive experiments, we successfully verified that Omniview-Tuning could bring significant improvements in viewpoint invariance while preserving the original performance. These advancements provide valuable insights and a standard for future research on viewpoint invariance in foundation models.", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the Project of the National Natural Science Foundation of China (No. 62076018, 92370124, 62276149, 92248303), the Fundamental Research Funds for the Central Universities, and Tsinghua-Alibaba Joint Research Program. Y. Dong was also supported by the China National Postdoctoral Program for Innovative Talents and CCF-BaiChuan-Ebtech Foundation Model Fund.", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 292, + 321, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 23716-23736 (2022)", + "2. Alcorn, M.A., Li, Q., Gong, Z., Wang, C., Mai, L., Ku, W.S., Nguyen, A.: Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4845-4854 (2019)", + "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)", + "5. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987)", + "6. Calian, D.A., Stimberg, F., Wiles, O., Rebuffi, S.A., Gyorgy, A., Mann, T., Gowal, S.: Defending against image corruptions through adversarial augmentations. arXiv preprint arXiv:2104.01086 (2021)", + "7. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. European Conference on Computer Vision (ECCV) (2022)", + "8. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020)", + "9. Chuang, Y.S., Dangovski, R., Luo, H., Zhang, Y., Chang, S., Soljacic, M., Li, S.W., Yih, W.t., Kim, Y., Glass, J.: Diffcse: Difference-based contrastive learning for sentence embeddings. arXiv preprint arXiv:2204.10298 (2022)", + "10. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21126-21136 (2022)", + "11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + ], + "bbox": [ + 218, + 325, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Das, S., Ryoo, M.S.: ViewCLR: Learning self-supervised video representation for unseen viewpoints. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5573-5583 (2023)", + "13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)", + "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "15. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1022-1032 (2023)", + "16. Dong, Y., Ruan, S., Su, H., Kang, C., Wei, X., Zhu, J.: View fool: Evaluating the robustness of visual recognition to adversarial viewpoints. Advances in Neural Information Processing Systems 35, 36789-36803 (2022)", + "17. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: Proceedings of the 39th International Conference on Machine Learning. pp. 6216-6234 (2022)", + "18. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)", + "19. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)", + "20. Hamdi, A., Ghanem, B.: Towards analyzing semantic robustness of deep neural networks. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 22-38. Springer (2020)", + "21. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)", + "22. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)", + "23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)", + "24. Ho, C.H., Leung, B., Sandstrom, E., Chang, Y., Vasconcelos, N.: Catastrophic child's play: Easy to perform, hard to defend adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9229-9237 (2019)", + "25. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "26. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below." + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)", + "29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)", + "31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019)", + "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (2023)", + "34. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in Neural Information Processing Systems 36 (2024)", + "35. Madan, S., Henry, T., Dozier, J., Ho, H., Bhandari, N., Sasaki, T., Durand, F., Pfister, H., Boix, X.: When and how cnns generalize to out-of-distribution category-viewpoint combinations. arXiv preprint arXiv:2007.08032 (2020)", + "36. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (ICLR) (2018)", + "37. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016 (2022)", + "38. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Proceedings of the 38th International Conference on Machine Learning. pp. 8821-8831 (2021)", + "42. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)", + "43. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021)", + "44. Rong, X.: word2vec parameter learning explained. arXiv preprint arXiv:1411.2738 (2014)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528 (2023)", + "46. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Towards viewpoint-invariant visual recognition via adversarial training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4709-4719 (2023)", + "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Advances in Neural Information Processing Systems (2022)", + "48. Schlarmann, C., Singh, N.D., Croce, F., Hein, M.: Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336 (2024)", + "49. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "50. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020)", + "51. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)", + "52. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024)", + "53. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016)", + "54. Tu, W., Deng, W., Gedeon, T.: A closer look at the robustness of contrastive language-image pre-training (CLIP). In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "55. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019)", + "56. Wang, X., He, K., Gupta, A.: Transitive invariance for self-supervised visual representation learning. In: Proceedings of the IEEE international conference on computer vision. pp. 1329-1338 (2017)", + "57. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023)", + "58. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023)", + "59. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023)", + "60. Zhang, Y., Huang, Y., Sun, Y., Liu, C., Zhao, Z., Fang, Z., Wang, Y., Chen, H., Yang, X., Wei, X., et al.: Benchmarking trustworthiness of multimodal large language models: A comprehensive study. arXiv preprint arXiv:2406.07057 (2024)", + "61. Zhao, B., Yu, S., Ma, W., Yu, M., Mei, S., Wang, A., He, J., Yuille, A., Kortylewski, A.: Ood-cv: A benchmark for robustness to individual nuisances in real-world out-of-distribution shifts. In: ICML 2022 Shift Happens Workshop (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Ruan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 17 + }, + { + "type": "ref_text", + "text": "62. Zhou, X., Liu, M., Zagar, B.L., Yurtsever, E., Knoll, A.C.: Vision language models in autonomous driving and intelligent transportation systems. arXiv preprint arXiv:2310.14414 (2023)", + "bbox": [ + 217, + 146, + 784, + 189 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Omniview-Tuning", + "bbox": [ + 609, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_model.json b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fd85dff6260c5889de13107b5c61994bc02633c6 --- /dev/null +++ b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_model.json @@ -0,0 +1,2504 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Omniview-Tuning: Boosting Viewpoint Invariance of Vision-Language Pre-training Models" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.214, + 0.777, + 0.246 + ], + "angle": 0, + "content": "Shouwei Ruan1, Yinpeng Dong2,5, Hanqing Liu3, Yao Huang1, Hang Su2,4, and Xingxing Wei1,3*" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.256, + 0.764, + 0.271 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Artificial Intelligence, Beihang University, Beijing 100191, China" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.272, + 0.76, + 0.285 + ], + "angle": 0, + "content": "\\(^{2}\\) Dept. of Comp. Sci. & Tech., Institute for AI, BNrist Center, THBI Lab," + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.286, + 0.762, + 0.299 + ], + "angle": 0, + "content": "Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, 100084, China" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.299, + 0.764, + 0.312 + ], + "angle": 0, + "content": "3 Hangzhou Innovation Institute, Beihang University, Hangzhou 311228, China" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.313, + 0.676, + 0.326 + ], + "angle": 0, + "content": "4 Zhongguancun Laboratory, Beijing, 100080, China" + }, + { + "type": "text", + "bbox": [ + 0.47, + 0.326, + 0.534, + 0.338 + ], + "angle": 0, + "content": "5 RealAI" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.339, + 0.761, + 0.35 + ], + "angle": 0, + "content": "{shouweiruan,hqliu,y_huang,xxwei}@buaa.edu.cn,{dongyinpeng,suhangss}@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.403, + 0.741, + 0.751 + ], + "angle": 0, + "content": "Abstract. Vision-Language Pre-training (VLP) models like CLIP have achieved remarkable success in computer vision and particularly demonstrated superior robustness to distribution shifts of 2D images. However, their robustness under 3D viewpoint variations is still limited, which can hinder the development for real-world applications. This paper successfully addresses this concern while keeping VLPs' original performance by breaking through two primary obstacles: 1) the scarcity of training data and 2) the suboptimal fine-tuning paradigms. To combat data scarcity, we build the Multi-View Caption (MVCap) dataset — a comprehensive collection of over four million multi-view image-text pairs across more than 100K objects, providing more potential for VLP models to develop generalizable viewpoint-invariant representations. To address the limitations of existing paradigms in performance trade-offs and training efficiency, we design a novel fine-tuning framework named Omniview-Tuning (OVT). Specifically, OVT introduces a Cross-Viewpoint Alignment objective through a minimax-like optimization strategy, which effectively aligns representations of identical objects from diverse viewpoints without causing overfitting. Additionally, OVT fine-tunes VLP models in a parameter-efficient manner, leading to minimal computational cost. Extensive experiments on various VLP models with different architectures validate that OVT significantly improves the models' resilience to viewpoint shifts and keeps the original performance, establishing a pioneering standard for boosting the viewpoint invariance of VLP models. The code and dataset are available via https://github.com/Heathcliff-saku/Omniview_Tuning" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.765, + 0.704, + 0.779 + ], + "angle": 0, + "content": "Keywords: Vision-Language Pre-training \\(\\cdot\\) Viewpoint Invariance" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.385, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.148, + 0.506, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.148, + 0.774, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 1: The Challenge of Viewpoint Invariance in VLP. We selected benchmarks representing clean distributions (ImageNet-1K [14], CIFAR-100 [27]), common 2D-OOD (ImageNet-V2 [42], ImageNet-R(endition) [21], ImageNet-Sketch [55]), and viewpoint-OOD (ImageNet-V(微观)+ [46], OOD-CV(Pose) [61], MIRO [7]). We display samples from these data distributions (left) and report the Top-1 accuracy of the original CLIP (ViT-L/14) and our improved OVT-CLIP (ViT-L/14) (right)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.37, + 0.376, + 0.387 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.787, + 0.553 + ], + "angle": 0, + "content": "Vision-Language Pre-training (VLP) models, such as CLIP [40] and BLIP [29], have shown great promise in learning transferable representations across various tasks. By aligning images and texts in a joint embedding space with a large corpus of paired image-text data, VLP models exhibit exceptional representation and generalization capabilities that surpass traditional task-specific models. Owing to this, the VLP models serve as foundation models for numerous tasks, including visual recognition [40], visual question answering [1,33], and image generation [41,47]. Moreover, these models can effectively integrate real-world visual inputs with humankind instructions, leading to their increasing use in physical-world applications, like autonomous driving [62] and embodied robotics [28,57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Besides their expressive power, VLP models have also shown excellent robustness under out-of-distribution (OOD) data [17, 40, 54, 60], including common corruptions [6, 15, 22], stylistic changes [21, 55], and natural distribution shifts [21, 23, 42]. However, a recent study [45] identifies that although VLP models excel at handling 2D-OOD samples, they suffer significant performance degradation under 3D viewpoint changes, revealing a notable shortcoming of the existing VLP models. As demonstrated in Fig. 1, when dealing with the benchmarks concerned with 3D viewpoint shifts [16, 46, 61], CLIP's performance is obviously lower than that on 2D-OOD benchmarks. This large gap likely stems from limited coverage of diverse viewpoints in the training datasets [18, 49, 53], which is crucial for learning viewpoint-invariant representations. As VLP models are increasingly deployed in real-world environments where viewpoint shifts often occur, enhancing their resilience to such changes is urgent and essential." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To address this problem, this paper sets out to enhance the viewpoint invariance of VLP models while preserving the original performance as much as possible. However, achieving this goal meets the following challenges: (1) Data scarcity: acquiring VLP training data that covers a wide range of viewpoint variations is particularly challenging compared to conventional image-text pair data. Although some datasets introduced for task-specific models do include viewpoint" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.145, + 0.744, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.565 + ], + "angle": 0, + "content": "Fig. 2: Method Overview. (A) We create the first multi-view image caption dataset by collecting multi-view samples from existing 3D object and video datasets, and generating category-guided descriptions using VLLMs. (B) The proposed Omniview-Tuning takes multi-view image caption data as input, employs the cross-view alignment objective to encourage the model to learn viewpoint-invariant representations, and achieves efficient fine-tuning by updating Viformer and LoRA parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.841 + ], + "angle": 0, + "content": "variations [4, 23, 35, 61], they often lack the textual descriptions vital for VLP. Even the largest available multi-view datasets [10, 46, 59] fall short in terms of scale, category coverage, and viewpoint diversity, thereby limiting the potential for VLP models to develop generalizable viewpoint-invariant representations. (2) Inappropriate paradigms: traditional approaches, which often regard viewpoint changes as adversarial attacks and employ adversarial training paradigms for enhancing invariance [2, 45, 46], are not entirely suitable for VLP models. Such frameworks typically entail a trade-off between robustness and accuracy—a balance that requires more careful consideration for foundation VLP models, where our aim is not solely to improve viewpoint invariance but, more importantly, to bridge the gap between it and the original performance. Furthermore, these approaches necessitate extra 3D reconstruction and neural rendering to capture adversarial viewpoints, leading to prohibitive computational costs for large-scale VLP models. For instance, tuning ResNet-50 with VIAT [46] under a dataset of just 1K objects demands around 400 GPU hours. Therefore, it is important to make training more efficient and less resource-intensive." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "Based on the above discussions, this paper conducts a pioneering exploration of the viewpoint invariance of VLP models. Specifically, we address the aforementioned challenges by making the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Million-scale multi-view image-text training set. We introduce a large-scale Multi-View Caption (MVCap) dataset tailored for viewpoint invariance of VLP models, comprising over 4.6 million multi-view image-text pairs across more than 100K objects. To assemble a diverse collection of multi-view image-text pairs, we amalgamate various 3D assets with real-world multi-view data. This process involves an extensive selection and rendering of multi-view images from existing datasets. We then utilize a Vision Large Language Model (VLLM) for automated caption generation to obtain semantically rich textual descriptions without extensive manual efforts. To ensure category consistency across varying viewpoints in the generated captions, we implement a category-guided prompting strategy, which maintains accuracy in textual descriptions for different viewpoints of the same object or scene (details in Sec. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Effective framework for enhancing VLP's viewpoint invariance. We propose Omniview-Tuning (OVT), a novel framework designed to enhance the viewpoint invariance of prevalent VLP models. As illustrated in Fig. 2, OVT employs multi-view image-text pairs for training additional learnable components. To amplify the model's proficiency in learning viewpoint-invariant representations, we introduce a Cross-viewpoint Alignment objective, ensuring that representations of the same object from different viewpoints are close and unified in the high-dimensional feature space. To prevent performance trade-offs due to the concept drift from aggressive viewpoint alignment, we innovatively construct the optimization paradigm of OVT in a minimax-like form. The optimization process includes identifying extreme outlier viewpoints during the maximization step, while optimizing the model's invariant representation for these outlier samples in the minimization step. This strategy enables the model to focus more on the worst-case viewpoint samples, thereby maximally preserving the original embedding distribution and avoiding performance degradation while saving computational costs. Moreover, OVT is designed in a Parameter-Efficient Fine-Tuning manner to improve efficiency, and creatively incorporates two trainable parameter modules: an embedding transformation module named VIForner and the Low-Rank Adaptation (LoRA [25]) weights, to acquire additional viewpoint invariance capabilities efficiently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Extensive experiments across various VLP architectures and tasks. We conduct extensive experiments to show the efficacy of the OVT framework in improving the viewpoint invariance for VLP models while maintaining performance on clean data and 2D-OOD samples. For example, by fine-tuning CLIP with OVT on different architectures (ViT-B/32, ViT-B/16, and ViT-L/14), the Top-1 accuracy on viewpoint-OOD benchmarks increased by an average of \\(9.6\\%\\), \\(10.2\\%\\), and \\(8.9\\%\\), respectively, with only a minimal sacrifice on 2D-OOD benchmarks by an average of \\(2.6\\%\\), \\(1.4\\%\\), and \\(0.2\\%\\). Furthermore, serving as the visual encoder in VLLMs (e.g., LLaVa [33]), OVT-CLIP also effectively improves viewpoint invariance in image captioning and visual question answering tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.181, + 0.573, + 0.197 + ], + "angle": 0, + "content": "2.1 Viewpoint Invariance and Robustness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.789, + 0.529 + ], + "angle": 0, + "content": "Viewpoint invariance is a key property of human vision [5] but is usually lacking in computer vision models [2, 16]. Addressing viewpoint invariance and robustness involves strategies like data augmentation and adversarial learning. Early efforts aim to enhance viewpoint robustness by incorporating datasets enriched with viewpoint variations [4, 23, 35, 61]. For example, Madan et al. encourage models to learn viewpoint-robust representations by incorporating object-pose combinations [35]. However, these methods often falter under malicious viewpoint perturbations due to their inability to capture the worst-case viewpoint samples. Recently, achieving viewpoint invariance within the adversarial training paradigm has shown promise [2, 16, 20, 46]. By treating viewpoint variations as an adversarial attack, Alcorn et al. employ a differentiable renderer to train models against adversarial viewpoints optimized from a limited 3D objects set [2]. Recent studies, such as Viewfool [16] and VIAT [45, 46], have introduced neural radiance field (NeRF) [38, 39], enabling the characterization of adversarial viewpoint distributions from 2D multi-view inputs. Besides, studies in the self-supervised domain employ a label-free paradigm to improve viewpoint invariance, but these methods often require the introduction of complex structures, such as large-scale graph structures [56] and viewpoint generators [12]. Distinct from previous studies, our work pioneers the improvement of viewpoint invariance representation within large-scale VLP models, which is facilitated through suitable training data and refined fine-tuning methodologies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.507, + 0.57 + ], + "angle": 0, + "content": "2.2 Vision-Language Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In the realm of VLP, significant strides have been made in understanding and bridging the semantic gap between visual and textual information. Despite the variety of existing VLP paradigms, such as single-stream encoder (e.g., VisualBERT [31] and UNITER [8], etc.) or dual-stream encoder equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by ALIGN [30] and OpenAI's CLIP [40] dominates the field. CLIP, in particular, has gained widespread attention for its ability to perform zero-shot classification tasks by adopting a vast corpus of internet-collected image-text pairs, demonstrating the power of large-scale contrastive pre-training. Thus, Our investigation primarily focuses on these VLP architectures. Building upon these foundational works, subsequent iterations like open-CLIP [26], EVA-CLIP [51,52], and Meta-CLIP [58] have introduced nuanced enhancements. These refinements, ranging from the incorporation of more expansive high-quality image-text datasets and improved training methodologies, have collectively contributed to performance uplifts. BLIP [29], meanwhile, introduces a bootstrapping mechanism by the proposed captioner and filter module that achieve significant performance improvements on various downstream tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison of current large-scale multi-view datasets. \"Spherical\" indicates whether the viewpoints cover spherical space, \"Diversity\" assesses the diversity of viewpoints, and \"Caption\" indicates whether textual descriptions are provided." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.784, + 0.272 + ], + "angle": 0, + "content": "
DatasetYear#Obj.#Cat.#Avg. View.#SampleImage DomainSphericalDiversityCaption
OOWL [24]201950025240120KRealX★★X
CO3D [43]202118.6k50~801.5MRealX★X
ABO [10]20227.9k6330238KSyntheticX★X
IM3D [46]20231.0k100100100KSynthetic✓★★★X
MVImgNet [59]2023219.0k238~306.5MRealX★★X
MVCap (Ours)202494.6k1600100/~304.6MSynthetic+Real✓★★★✓
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.296, + 0.523, + 0.314 + ], + "angle": 0, + "content": "3 Multi-view Caption Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.418 + ], + "angle": 0, + "content": "We recognize that one of the key challenges in achieving viewpoint invariance for VLP is the scarcity of training data that offer comprehensive viewpoints. As shown in Tab. 1, existing large-scale multi-view datasets [10,24,43,46,59] typically lack in either sample diversity, category breadth, or textual descriptions, limiting their effectiveness for supporting VLP models to achieve viewpoint invariance. To address these limitations, we next introduce the MVCap dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.439, + 0.502, + 0.455 + ], + "angle": 0, + "content": "3.1 Multi-View Image Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.675 + ], + "angle": 0, + "content": "We commence by gathering a multi-view image collection \\(\\mathcal{D} = \\{I_{ij} \\mid i = 1,2,\\dots,N; j = 1,2,\\dots,M_i\\}\\), where \\(N\\) and \\(M_i\\) represent the counts of objects and their viewpoints, respectively. To cover various categories from virtual to real-world scenes, we integrate samples from Objaverse [13], IM3D [46], and MVImgNet [59]. Since the original 3D dataset includes a fair share of noisy and semantically indistinct objects, we leverage semantic embeddings provided by OpenShape [34] to conduct cosine similarity sorting based on the embeddings of customized labels. Finally, we filter 24,495 virtual 3D objects endowed with distinct semantic clarity and cover over 1,600 categories. For each chosen 3D object, we employ Blender to render 100 random viewpoint images from the upper hemisphere, ensuring a comprehensive and varied viewpoint representation in our collected samples. We also incorporate objects from MVImgNet with over 30 valid viewpoints (video frames), thereby acquiring a substantial number of real-world multi-view samples to enrich the dataset's content and quality further." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.573, + 0.712 + ], + "angle": 0, + "content": "3.2 Category-Guided Caption Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.786, + 0.841 + ], + "angle": 0, + "content": "The granularity and precision of textual descriptions are pivotal in VLP training, as they influence the model's generalization capabilities and the variety of visual concepts learned. Relying solely on simple prompt engineering, such as \"a photo of [category],\" may introduce biases and limit the model's generalizability, whereas manual annotation is costly. To circumvent this, we utilize InstructBLIP-flant5xl [11], a leading VLLM, to create multi-view captions automatically. However, such VLLMs also grapple with viewpoint invariance, where the model's responses to different viewpoints can often be category-inconsistent," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.151, + 0.431, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.149, + 0.557, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.153, + 0.777, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.24, + 0.777, + 0.255 + ], + "angle": 0, + "content": "Fig. 3: Generated multi-view captions with common and category-guided prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.785, + 0.372 + ], + "angle": 0, + "content": "as depicted in Fig. 3. This situation presents a \"chicken or egg\" dilemma: we hope to use a viewpoint-invariant model to supply data for viewpoint invariance training. We address this by the design of category-guided prompting. Specifically, we use prompts containing ground-truth category information to eliminate the hallucination of large VLLMs in response to viewpoint-shifted inputs, thereby generating category-consistent multi-view captions. Formally, the forward process for generating captions can be represented as follows:" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.382, + 0.785, + 0.429 + ], + "angle": 0, + "content": "\\(T_{ij} = \\mathcal{G}[I_{ij},\\mathrm{Prompt}(c_i)]\\) Prompt \\((c_{i}) = \"\\) Write a short description for the image, (1) noting that the main instance of the image is \\(a < c_{i} > .\"\\)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.44, + 0.785, + 0.5 + ], + "angle": 0, + "content": "where \\(c_{i} \\in C\\) denotes the category label for the \\(i\\)-th object, and \\(\\mathcal{G}\\) denotes the forward process of InstructBLIP. This yields the multi-view image-text pairs \\(\\tilde{\\mathcal{D}} = \\{\\langle I_{ij}, T_{ij} \\rangle \\mid i = 1, 2, \\dots, N; j = 1, 2, \\dots, M_i\\}\\), which can be utilized for the viewpoint invariance fine-tuning of VLP models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.426, + 0.54 + ], + "angle": 0, + "content": "4 omniview-Tuning" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.555, + 0.726, + 0.57 + ], + "angle": 0, + "content": "4.1 Preliminaries: Contrastive Vision-Language Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.785, + 0.639 + ], + "angle": 0, + "content": "Despite the variety of existing VLP paradigms, such as single-stream or dual-stream architectures equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by CLIP et al. [30, 40] dominates the field. Thus, Our investigation primarily focuses on this VLP architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.785, + 0.73 + ], + "angle": 0, + "content": "Without the loss of generality, these VLP models are composed of a visual encoder \\( E_{\\mathbf{W}_{\\mathbf{v}}}: I \\to z^{I} \\in \\mathbb{R}^{d} \\) and a text encoder \\( E_{\\mathbf{W}_{\\mathbf{t}}} : T \\to z^{T} \\in \\mathbb{R}^{d} \\), which maps visual and textual inputs to a unified high-dimensional feature space \\( \\mathbb{R}^d \\), respectively, where \\( \\mathbf{W}_{\\mathbf{v}} \\) and \\( \\mathbf{W}_{\\mathbf{t}} \\) are weight matrices of two encoders. Given a large corpus of image-text pairs \\( \\{\\langle I_i, T_i \\rangle\\}_{i=1}^N \\), VLP models typically employ an image-text contrastive (ITC) loss as the training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.741, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I T C} = \\frac {1}{2} \\left(\\mathcal {L} _ {I \\rightarrow T} + \\mathcal {L} _ {T \\rightarrow I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.77, + 0.784, + 0.784 + ], + "angle": 0, + "content": "which is composed of an image-to-text and a text-to-image terms formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.795, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {I \\rightarrow T} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(d \\left(z _ {i} ^ {I} , z _ {i} ^ {T}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(d \\left(z _ {i} ^ {I} , z _ {k} ^ {T}\\right) / \\tau\\right)}, \\tag {3} \\\\ \\mathcal {L} _ {T \\to I} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (d (z _ {i} ^ {T} , z _ {i} ^ {I}) / \\tau)}{\\sum_ {k = 1} ^ {N} \\exp (d (z _ {i} ^ {T} , z _ {k} ^ {I}) / \\tau)}, \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.254 + ], + "angle": 0, + "content": "where \\(\\tau\\) represents a learnable temperature parameter, \\(z^I\\) and \\(z^T\\) denote the image and text embeddings, respectively. The \\(\\mathcal{L}_{ITC}\\) maximizes the similarity between matched image-text pairs while minimizing the similarity for mismatched pairs, thus enabling the alignment of visual and textual information to the same feature space, bringing the embeddings of matched pairs closer. Following [29, 30, 40], the proposed Omniview-Tuning implements \\(\\mathcal{L}_{ITC}\\) for aligning multi-view images with text modalities, which is explained in the next section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.275, + 0.441, + 0.289 + ], + "angle": 0, + "content": "4.2 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.377 + ], + "angle": 0, + "content": "Viewpoint Invariance of Vision-Language Pre-training. In computer vision scenario, viewpoint invariance implies that model \\( f(\\cdot) \\) can provide consistent predictions or representations given any different views of the identical object or scene [46]. Formally, given a collection of multi-view images \\( \\mathcal{D} = \\{I_{ij} \\mid i = 1,2,\\dots,N; j = 1,2,\\dots,M_i\\} \\), viewpoint invariance is required:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.388, + 0.786, + 0.406 + ], + "angle": 0, + "content": "\\[\nf \\left(I _ {i j}\\right) = f \\left(I _ {i j ^ {\\prime}}\\right), \\quad \\forall i, j, j ^ {\\prime} \\text {w i t h} j \\neq j ^ {\\prime}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.522 + ], + "angle": 0, + "content": "where \\( i \\) is the index of the object/scene, \\( j \\) and \\( j' \\) are indexes of two viewpoint samples. However, in the context of dual-stream VLP models, this concept requires a more refined interpretation. For VLP models, viewpoint invariance necessitates that the visual representations (i.e., the embeddings inferred from the visual encoder) from different viewpoints be sufficiently close in the feature space. Assuming \\( I_{ij} \\) and \\( I_{ij'} \\) are images from different viewpoints of the same object, this requirement can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.533, + 0.787, + 0.558 + ], + "angle": 0, + "content": "\\[\nd \\left[ E _ {\\mathbf {W} _ {\\mathbf {v}}} \\left(I _ {i j}\\right), E _ {\\mathbf {W} _ {\\mathbf {v}}} \\left(I _ {i j ^ {\\prime}}\\right) \\right] \\leq \\epsilon , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.597 + ], + "angle": 0, + "content": "where \\( d(\\cdot) \\) denotes a distance metric in the representation space, such as cosine distance, \\( \\epsilon \\) represents the maximum variance allowed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Optimization Objectives of Omniview-Tuning. Although images from different viewpoints often correspond to slightly varying textual descriptions, influenced by context, grammatical structure, and linguistic ambiguity, this variation could be significantly amplified in the high-dimensional representation space [9, 44]. Therefore, relying solely on image-text alignment may not suffice to adequately align embeddings from different viewpoints. Starting from the definition of viewpoint invariance, we introduce a cross-viewpoint alignment objective within the \\(\\mathcal{L}_{ITC}\\) to directly encourage the model to learn invariant representations between different viewpoints, rather than relying on the indirect alignment through textual descriptions. This can be seen as a regularization that forces the model to obtain viewpoint invariance, even when such invariance is not explicitly articulated in the textual descriptions. With this consideration, given a multi-view training set \\(\\mathcal{D}\\), the optimization problem is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.808, + 0.787, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {W} _ {\\mathbf {v}}, \\mathbf {W} _ {\\mathbf {t}}} \\left[ \\mathcal {L} _ {I T C} + \\lambda \\cdot \\underbrace {\\sum_ {i} \\sum_ {j \\neq j ^ {\\prime}} d \\left(z _ {i j} ^ {I} , z _ {i j ^ {\\prime}} ^ {I}\\right)} _ {\\mathcal {L} _ {V C}} \\right], \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "where the first term represents the image-text alignment used in the pre-training process, while the second term signifies the cross-viewpoint alignment goal mentioned above, referred to as Viewpoint Consistency loss \\((\\mathcal{L}_{VC})\\), which aims to minimize the cosine distance between embeddings from different viewpoints. \\(\\lambda\\) is a hyperparameter that balances the importance of two loss terms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.242, + 0.449, + 0.259 + ], + "angle": 0, + "content": "4.3 Optimization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.789, + 0.388 + ], + "angle": 0, + "content": "In summary, the naive way to achieve viewpoint invariance is to calculate the loss terms in Eq. (6) based on the forward process of encoders, then update the encoders' weight using gradient descent. However, it has a relatively high time complexity to solve Eq. (6) because current \\(\\mathcal{L}_{VC}\\) requires iterating over every possible combination of viewpoints. Therefore, we endeavor to provide a more effective implementation for the original optimization problem. Drawing from the advantages of adversarial training [36, 46], we frame the optimization of the \\(\\mathcal{L}_{VC}\\) in a minimax format, rewriting the original problem Eq. (6) as:" + }, + { + "type": "equation", + "bbox": [ + 0.269, + 0.394, + 0.786, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\mathbf {W} _ {\\mathbf {v}}, \\mathbf {W} _ {\\mathbf {t}}} \\left[ \\mathcal {L} _ {I T C} + \\lambda \\cdot \\underbrace {\\max _ {\\mathcal {O} = \\{O _ {i} \\} _ {i = 1} ^ {N} , | O _ {i} | = K} \\sum_ {i = 1} ^ {N} \\sum_ {j \\in \\mathcal {O}} l \\left(z _ {i j} ^ {I} , z _ {C _ {i}} ^ {I}\\right)} \\right], \\tag {7} \\\\ \\mathrm {w h e r e} l (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) = \\max \\left[ d (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) + m, 0 \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.789, + 0.593 + ], + "angle": 0, + "content": "where \\(\\mathcal{O} = \\{O_i\\}_{i=1}^N\\) is the outlier viewpoints set, \\(z_{C_i}^I\\) are anchor viewpoint embeddings of each object, and \\(l(\\cdot)\\) is the cosine distance with a margin \\(m\\). During the optimization, The maximization step first identifies the collection of top-\\(K\\) outlier viewpoints \\(\\mathcal{O}\\), which are the viewpoint samples with the highest degree of representational deviation. Then, the minimization step encourages the outlier viewpoint embeddings to converge towards corresponding anchor viewpoint embeddings \\(z_{C_i}^I\\). We obtain \\(z_{C_i}^I\\) by calculating the nearest-neighbor weighted embedding centroid of each object:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.599, + 0.787, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} z _ {C _ {i}} ^ {I} = \\sum_ {j = 1} ^ {M _ {i}} \\tilde {\\omega} _ {i j} \\cdot z _ {i j} ^ {I}, \\\\ \\mathrm {w h e r e} \\tilde {\\omega} _ {i j} = \\omega_ {i j} / \\sum_ {j} \\omega_ {i j}, \\omega_ {i j} = 1 / \\sum_ {z _ {i h} ^ {I} \\in \\mathcal {Q} _ {i j}} d (z _ {i j} ^ {I}, z _ {i h} ^ {I}) \\end{array} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.693 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}_{ij} = \\{z_{ih}^I\\}_{h=1}^5\\) is the top-5 nearest neighbours of each viewpoint embedding \\(z_{ij}^I\\). As for the outlier viewpoints set, we define them as the viewpoints with the top-\\(K\\) farthest cosine distances from \\(z_{C_i}^I\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.77 + ], + "angle": 0, + "content": "The adoption of this strategy offers dual advantages: Firstly, it allows the model to focus solely on extreme outlier viewpoints, preventing concept drift and potential overfitting to the fine-tuning dataset that results from excessive alignment. Secondly, this approach reduces computational overhead and significantly enhances optimization efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.501, + 0.803 + ], + "angle": 0, + "content": "4.4 Parameter-Efficient Modules" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To mitigate the impact of full parameters update on the original performance and enhance training efficiency, we achieve viewpoint invariance by efficiently" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.147, + 0.77, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.29, + 0.71, + 0.305 + ], + "angle": 0, + "content": "Fig. 4: Visualization results for zero-shot classification results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.785, + 0.364 + ], + "angle": 0, + "content": "fine-tuning the parameters of the visual encoder while keeping the text encoder frozen. Inspired by LoRA [25], we perform low-rank decomposition on the weights of the visual encoder \\(\\mathbf{W}_{\\mathbf{v}} \\in \\mathbb{R}^{n \\times m}\\) to substitute full-parameter update:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.374, + 0.785, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {W}} _ {\\mathbf {v}} = \\mathbf {W} _ {\\mathbf {v}} + \\Delta W = \\mathbf {W} _ {\\mathbf {v}} + \\mathbf {B} \\mathbf {A}, \\quad \\text {w h e r e} \\mathbf {B} \\in \\mathbb {R} ^ {m \\times r}, \\mathbf {A} \\in \\mathbb {R} ^ {r \\times n}, r \\ll \\min (n, m), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.618 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) are two learnable low-rank parameter matrices, which we apply to the self-attention layers of the visual encoder and update them during fine-tuning while freezing the original pre-trained weights. Compared to directly adjusting the original network weights (i.e., full-parameter fine-tuning manner in TeCoA [37] and FARE [48]), this approach enables us to improve the model's viewpoint invariance representation capability while better preserving the original performance. Drawing inspiration from the success of CLIP-Adapter [19], which improves CLIP's performance in few-shot scenarios by introducing linear layers after the encoder, we propose a similar module called VIformer: \\(f_{\\pmb{\\theta}}:z^{I}\\in \\mathbb{R}^{d}\\to s^{I}\\in \\mathbb{R}^{d}\\), where \\(\\pmb{\\theta}\\) is the weight. Unlike CLIP-Adapter, VIformer transforms the original embeddings \\(z^{I}\\) by introducing self-attention layers in a learnable manner to extract and retain specific viewpoint-invariant key components \\(s^I\\). Combining LoRA and VIformer modules, the forward process of image encoding can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.628, + 0.785, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {z} ^ {I} = \\alpha \\cdot f _ {\\boldsymbol {\\theta}} (z ^ {I}) + (1 - \\alpha) \\cdot z ^ {I} \\\\ = \\alpha \\cdot f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {W} _ {\\mathbf {v}} \\cdot I + \\mathbf {B} \\mathbf {A} \\cdot I\\right) + (1 - \\alpha) \\cdot \\left(\\mathbf {W} _ {\\mathbf {v}} \\cdot I + \\mathbf {B} \\mathbf {A} \\cdot I\\right), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.753 + ], + "angle": 0, + "content": "where the constant value \\(\\alpha\\) denotes the residual ratio to balance achieving original performance and viewpoint invariance performance. Therefore, for Eq. (6), we now only need to update \\(\\mathbf{A}\\), \\(\\mathbf{B}\\), and \\(\\pmb{\\theta}\\), rather than the entire weights of VLP. To facilitate the understanding of the OVT training process, we provide the pseudocode for OVT as shown in Appendix D" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.777, + 0.376, + 0.794 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.842 + ], + "angle": 0, + "content": "Our evaluation of Omniview-Tuning spans several downstream tasks, including zero-shot classification, image captioning, and vision question answering." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Configurations of OVT and zero-shot Top-1 accuracy (\\%) on ImageNet-1K with ImageNet-V+. The number in parentheses shows the performance change relative to the pre-trained weights. Through OVT training, each model maintains the performance on ImageNet-1K (IN-1K) while significantly improving the performance on ImageNet-V+ (IN-V+.), narrowing the performance gap." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.227, + 0.787, + 0.328 + ], + "angle": 0, + "content": "
ModelPretrain WeightPretrain DataTotal #Param.Trainable #Param.Image SizeBatch Size#Iter.IN-1KIN-View+
OVT-OpenCLIP ViT-B/32OpenCLIP-ViT-B/32LAION (2B)151M6.6M22451235k67.8 (↑1.3)59.5 (↑22.4)
OVT-OpenCLIP ViT-B/16OpenCLIP-ViT-B/16LAION (2B)149M6.6M22451235k69.7 (↑2.1)61.7 (↑17.5)
OVT-OpenCLIP ViT-L/14OpenCLIP-ViT-L/14LAION (2B)428M11.8M22425620k77.3 (↑2.1)69.8 (↑16.6)
OVT-MetaCLIP ViT-B/32MetaCLIP-ViT-B/32Common crawl (2.5B)151M6.6M22451240k69.7 (↑2.1)54.8 (↑13.8)
OVT-MetaCLIP ViT-B/16MetaCLIP-ViT-B/16Common crawl (2.5B)149M6.6M22451240k73.8 (↑1.7)64.8 (↑15.2)
OVT-MetaCLIP ViT-L/14MetaCLIP-ViT-L/14Common crawl (2.5B)428M11.8M22425620k77.7 (↓1.4)75.4 (↑9.0)
OVT-BLIP ViT-B/16Salesforce-BLIP-ViT-B/16COCO et al. (129M)234M4.3M22425620k61.7 (↑8.8)54.8 (↑18.0)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.449 + ], + "angle": 0, + "content": "For zero-shot classification (Sec. 5.1), we conduct evaluations for CLIP [40] and BLIP [29] architectures. For image captioning and vision question answering (Sec. 5.2), we replace the visual encoders in Vision Large Language Models (VLLMs) with our fine-tuned versions. We adopt LLaVA-1.5 [32,33], and OpenFlamingo [3], the most advanced open-source VLLMs available. Additionally, we present the ablation study and convergence analysis of our approach in Sec. 5.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.469, + 0.573, + 0.483 + ], + "angle": 0, + "content": "5.1 Evaluation of Zero-Shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.735 + ], + "angle": 0, + "content": "Baselines. We adopt the official CLIP (OpenAI CLIP [40]) and the community open-source version (OpenCLIP [26]) as baselines. Additionally, we include the current state-of-the-art Eva02-CLIP [51] and MetaCLIP [58] as another set of baselines to compare CLIP trained with improved techniques and extensive training data. For BLIP, we use the official implementation [29] as the baseline. Settings. We train two series of CLIP models using our OVT framework and MVCap dataset, each series comprising three different visual encoder architectures (ViT-B/32, ViT-B/16, and ViT-L/14). OVT-OpenCLIP are fine-tuned on the original weights of OpenCLIP, while OVT-MetaCLIP are based on MetaCLIP. The fine-tuning settings are detailed in Tab. 2. We standardized the \\(\\lambda = 1.0\\), \\(\\alpha = 0.1\\), and the outlier viewpoints number \\(K = 5\\) and set the LoRA rank at 8. Datasets and Metrics. We employ a various set of benchmarks for evaluation, including clean data distributions [14, 27], common 2D-OOD [21, 23, 42, 55, 61], and most importantly, viewpoint-OOD [7, 16, 46, 61]. For each benchmark, we report Top-1 and Top-5 accuracy and average accuracy across all benchmarks. The evaluations follow the standard prompting engineering of CLIP [40]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Results and Discussions. Tab. 3 summarizes the performance of our OVT-trained VLP models against various VLP versions, including their accuracy across different benchmarks and average accuracy across clean, common-OOD, and Viewpoint-OOD domains. We can draw the following conclusions:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "(1) OVT significantly enhances the models' invariance to Viewpoint-OOD samples. Across different VLP architectures and visual encoders, OVT-trained models perform best on almost all viewpoint-OOD benchmarks. On the average" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.253, + 0.145, + 0.75, + 0.16 + ], + "angle": 0, + "content": "Table 3: Top-1/Top-5 zero-shot accuracy (%) under different benchmarks" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.172, + 0.787, + 0.443 + ], + "angle": 0, + "content": "
ModelCleanCommon-ODViewpoint-OD
ImageNet-100 [66]ImageNet-IR [14]Cllr-100 [27]Avg. Acc.ImageNet-Vi [14]ImageNet-Sch. [53]ImageNet-ODD [25]ImageNet-Rem. [21]OOD-CV [61]Avg. Acc.ImageNet-View [68]ImageNet-View + [68]OOD-CV-Pose [63]MIBO [7]Avg. Top-1Total Avg. Acc.
A. Comparisons with VIT-B/32 baselines
OpenAI CLIP77.5/93.963.3/88.864.3/88.168.4/90.255.8/83.442.2/70.333.4/62.250.7/75.450.2/82.646.5/74.844.5/65.427.5/52.447.2/84.526.5/59.436.4/65.448.6/75.5
Open CLIP81.1/95.366.5/89.975.8/94.074.5/93.058.1/83.953.6/79.334.8/64.461.0/81.953.5/81.952.2/78.354.4/72.137.1/63.246.9/81.633.0/69.242.8/71.554.6/79.7
OVT-OpenCLIP80.9/95.667.8/90.865.0/89.371.2/91.958.0/84.245.8/73.442.8/75.050.3/71.451.7/79.549.7/76.761.9/81.259.5/85.652.8/82.535.4/80.152.4/82.456.0/82.4
MetaCLIP80.7/95.667.6/90.577.7/95.275.3/93.859.5/85.455.9/81.432.4/62.563.2/83.852.0/84.252.6/79.561.4/76.741.0/67.848.9/87.934.8/73.246.5/76.456.3/82.0
OVT-MetaCLIP80.7/95.669.7/92.071.8/93.074.0/93.560.6/85.847.8/75.843.5/73.849.0/70.850.1/80.150.2/77.264.0/79.254.8/80.455.1/84.835.6/77.052.4/80.356.9/82.3
B. Comparisons with VIT-B/16 baselines
OpenAI CLIP82.1/95.768.3/91.967.2/89.472.5/92.361.8/87.448.2/76.327.7/55.759.1/83.052.2/84.649.8/77.451.6/68.936.9/63.853.4/86.830.1/66.143.0/71.453.2/79.1
Open CLIP83.2/96.270.1/91.877.0/94.876.8/94.362.2/87.056.0/82.030.7/59.864.9/85.654.3/82.753.6/79.458.1/74.444.2/70.948.5/84.034.6/74.646.4/76.057.0/82.0
OVT-OpenCLIP83.9/97.071.9/93.169.0/90.774.9/93.664.0/88.650.5/77.936.8/68.957.0/77.256.3/84.552.9/79.465.4/80.761.7/85.856.9/87.442.4/84.956.6/84.759.6/84.7
EVA-CLIP85.3/96.574.6/94.287.5/98.082.5/96.367.0/89.857.6/82.321.3/47.369.6/87.553.1/83.153.7/78.061.8/76.644.3/69.453.9/87.432.9/73.248.2/76.659.1/82.1
MetaCLIP84.3/97.272.1/93.478.9/95.478.4/95.365.0/89.360.1/84.826.2/56.470.2/89.352.3/85.454.8/81.064.2/79.449.6/76.148.9/90.938.5/78.750.3/81.259.2/84.7
OVT-MetaCLIP83.4/97.473.8/94.173.9/93.677.0/95.065.9/89.453.6/81.036.2/68.859.0/79.651.6/83.853.2/80.169.7/84.064.8/87.355.2/87.839.2/82.957.2/85.560.5/85.6
C. Comparisons with VIT-L/14 baselines
OpenAI CLIP86.5/97.475.4/94.676.5/93.379.5/95.169.8/90.959.5/84.318.6/43.872.8/91.452.9/88.854.7/79.860.3/75.645.8/71.547.9/88.238.0/74.148.0/77.358.6/82.8
Open CLIP86.8/97.875.2/94.383.7/96.781.9/96.267.7/90.263.2/86.424.0/50.574.5/91.254.5/85.056.8/80.665.7/78.153.2/76.752.4/90.542.3/83.053.4/82.161.9/85.0
OVT-OpenCLIP89.0/97.877.3/95.379.2/95.381.8/96.169.6/91.561.9/86.027.5/55.471.3/88.756.4/87.057.3/81.772.2/86.669.8/89.757.3/94.150.0/89.362.3/89.965.1/88.1
EVA-CLIP88.5/97.979.6/96.090.6/98.686.3/97.572.8/92.768.0/89.116.3/40.082.8/95.754.7/87.458.9/81.071.5/82.361.1/81.754.4/94.539.6/86.156.6/86.165.0/86.8
MetaCLIP88.3/98.379.1/95.984.1/96.983.8/97.072.5/92.668.9/89.817.0/40.681.8/95.156.6/87.559.3/81.177.3/89.366.4/87.058.9/93.448.1/89.662.7/89.866.6/88.0
OVT-MetaCLIP88.8/97.577.7/95.984.0/96.983.5/96.870.8/92.264.4/87.920.8/47.077.0/92.756.3/89.357.8/81.879.3/90.675.4/93.057.0/94.446.4/93.864.5/92.966.5/89.3
D. Comparisons with BLIP VIT-L/B16 baselines
BLIP76.6/93.352.9/80.267.0/88.365.5/87.347.3/74.751.0/76.625.6/53.464.3/83.853.9/87.648.4/75.255.2/68.236.8/63.350.8/89.227.0/66.142.4/71.950.7/77.1
OVT-BLIP82.2/97.061.7/88.866.6/88.970.2/91.553.7/82.946.5/74.233.8/62.757.4/77.956.4/87.349.6/77.062.6/79.054.8/79.955.2/89.531.5/73.251.0/80.455.2/81.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.673 + ], + "angle": 0, + "content": "accuracy of viewpoint-OOD datasets, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 shows improvements of \\(9.6\\%\\), \\(10.2\\%\\), and \\(8.9\\%\\) over OpenCLIP, respectively. OVT-BLIP demonstrated an average improvement of \\(8.6\\%\\). (2) While enhancing viewpoint invariance, OVT maintains performance on clean samples and 2D-OOD without significant performance trade-offs. For 2D-OOD benchmarks, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 sacrifice only \\(2.6\\%\\), \\(1.4\\%\\), and \\(0.2\\%\\) accuracy. (3) Compared to earlier CLIP baselines, the recently developed MetaCLIP exhibits better zero-shot performance. Based on this, OVT further enhances its performance under viewpoint-OOD samples. Visualization. We showcase OVT-OpenCLIP and the original OpenCLIP prediction on several viewpoint-OOD samples. As illustrated in Fig. 4, OVT-CLIP successfully predicts the categories of images from various unusual viewpoints in all cases, whereas the original CLIP is prone to make incorrect predictions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.5, + 0.708 + ], + "angle": 0, + "content": "5.2 Performance on Other Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Settings. As LLaVA and Openflamingo use the OpenAI CLIP (ViT-L/14) to encode vision inputs, we applied OVT to this model in this section and comparing with other OpenAI CLIP (ViT-L/14) versions in image captioning tasks. The training setup remains consistent with the OVT-OpenCLIP described in Tab. 2. Baselines. In addition to comparing with the original OpenAI CLIP version, we also select \\( TeCoA^4 \\) [37] and \\( FARE^4 \\) [48], robust CLIP models based on adversarial training, as baselines, which have been proven to possess good resistance to adversarial samples in image captioning task." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Image captioning performance under clean distribution samples and viewpoint-OOD samples from Real-world and Synthetic domains. We utilize the MP-Net [50] to calculate the similarity between generated descriptions and ground-truth labels, considering predictions successful if they exceed the similarity threshold \\(\\beta\\)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.787, + 0.34 + ], + "angle": 0, + "content": "
ModelVisual EncoderReal-world DomainSynthetic Domain
OOD-CV (iid) [61]β@1.0β@0.5β@Adp.OOD-CV (Pose) [61]β@1.0β@0.5β@Adp.IM3D [46]β@1.0β@0.5β@Adp.ImageNet-V+ [46]β@1.0β@0.5β@Adp.
LLaVa-7bOpenAI CLIP(ViT-L/14)44.161.167.546.453.658.746.753.358.820.425.532.1
TeCoA4[37](ViT-L/14)41.958.965.536.141.649.226.330.142.68.711.622.6
FARE4[48](ViT-L/14)42.158.965.240.245.950.835.239.249.212.715.823.1
OVT-CLIP(ViT-L/14)43.559.565.946.553.659.149.454.061.826.431.941.0
LLaVa-13bOpenAI CLIP(ViT-L/14)45.468.070.648.658.660.848.756.760.821.228.432.5
TeCoA4[37](ViT-L/14)42.467.072.237.448.951.325.028.641.58.410.921.8
FARE4[48](ViT-L/14)43.966.771.141.952.154.836.141.448.612.115.920.8
OVT-CLIP(ViT-L/14)45.767.370.848.258.661.950.458.963.226.436.240.9
" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.359, + 0.273, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.36, + 0.402, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.36, + 0.46, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.36, + 0.585, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.397, + 0.274, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.397, + 0.402, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.397, + 0.46, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.397, + 0.585, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.433, + 0.276, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.433, + 0.402, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.433, + 0.452, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.433, + 0.586, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.481, + 0.589, + 0.537 + ], + "angle": 0, + "content": "Fig. 5: The image descriptions generated by LLaVa-13B using our OVT-CLIP and the original OpenAI CLIP as vision encoder, where red texts indicates incorrect category descriptions, and green texts represents correct." + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.359, + 0.782, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.481, + 0.788, + 0.536 + ], + "angle": 0, + "content": "Fig.6: The Top-1 accuracy of OVT-OpenCLIP (ViT-B/16) with the iterations increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Datasets and Metrics. Given the absence of caption benchmarks that include viewpoint-changing OOD samples, we conduct evaluations using existing viewpoint-OOD datasets, including real-world datasets (using OOD-CV (iid) to represent clean distribution and OOD-CV (Pose) for viewpoint-OOD) and synthetic datasets (using IM3D [46] for clean distribution and ImageNet-V+ for viewpoint-OOD). We adopt word embedding distance to calculate the accuracy of the captioning task. By adopting MPNet [50], a state-of-the-art textual embedding model, we measure the similarity between keywords in the generated description and the ground-truth categories. Then assess the accuracy by counting the number of samples that exceed a specific similarity threshold \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Results and Discussions. Tab. 4 shows the image captioning accuracy of CLIP models under different training strategies, considering \\(\\beta\\) at 1.0 (indicating predictions involve ground-truth categories), 0.5, and \\(Adp\\). (meaning \\(\\beta\\) is equal to the average similarity in the clean distribution). We found that OVT-CLIP improves the accuracy of descriptions generated by LLaVa for viewpoint-OOD samples while maintaining its performance on corresponding clean distributions. When used as the visual encoder for the LLaVa-7B model, OVT-CLIP achieved an \\(8.9\\%\\) increase in accuracy compared to the original CLIP model weights under \\(\\beta = Adp\\). Besides, we find that although robust CLIP versions maintain perfor" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 5: Average Top-1/Top-5 zero-shot accuracy (\\%) under different data distributions within various ablation settings." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.787, + 0.25 + ], + "angle": 0, + "content": "
LITCVIFormerLVCTotal Avg.Clean Avg.Common-OOD Avg.Viewpoint-OOD Avg.
XXX61.985.081.996.256.880.653.482.1
✓XX61.985.7 (↑0.7)79.4 (↓2.5)95.6 (↓0.6)56.2 (↓0.6)81.4 (↑0.8)56.0 (↑2.6)83.8 (↑1.7)
✓✓X62.2 (↑0.3)86.2 (↑1.2)79.9 (↓2.0)95.4 (↓0.8)55.4 (↓1.4)81.6 (↑1.0)57.5 (↑4.1)85.2 (↑3.1)
✓✓✓65.1 (↑3.2)88.1 (↑3.1)81.8 (↓0.1)96.1 (↓0.1)57.3 (↑0.5)81.7 (↑1.1)62.3 (↑8.9)89.9 (↑7.8)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.785, + 0.34 + ], + "angle": 0, + "content": "mance on clean distribution samples, they experience a significant performance decline when facing viewpoint-OOD samples. We select some examples with the generated description in Fig. 5. For the visual question-answering task, we used OpenFlamingo as the VLLMs. The results are reported in the Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.364, + 0.596, + 0.379 + ], + "angle": 0, + "content": "5.3 Ablation Studies and Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Our ablation studies focus on the VIFormer and the \\(\\mathcal{L}_{VC}\\) within the Omniview-Tuning framework. Tab. 5 shows the Top-1/Top-5 acc of OVT-OpenCLIP (ViT-L/14) across various data distributions and different ablation settings. Beyond the original OpenCLIP, we set a baseline that only uses \\(\\mathcal{L}_{ITC}\\) for fine-tuning. Keeping other training settings fixed, reliance solely on \\(\\mathcal{L}_{ITC}\\) led to a more significant performance decline in clean and 2D-OOD samples while achieving limited viewpoint OOD performance improvement (2.6%/1.7%). The integration of VI-Former led to further improvements in viewpoint OOD accuracy (4.1%/3.1%). With the further addition of \\(\\mathcal{L}_{VC}\\), the improvement in viewpoint OOD performance is most significant (8.9%/7.8%), and it also reduces performance sacrifices in other data distributions. More ablation analyses are available in Appendix B. Furthermore, we report OVT's training convergence, depicted in Fig. 6. We display the Top-1 accuracy evolution for OVT-OpenCLIP (ViT-B/16) across various training iterations. We observe that around 40K iterations, with a batch size of 512, are sufficient for effective convergence, thus achieving a balance in performance across different data distributions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.368, + 0.672 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To tackle the challenge of 3D viewpoint in VLP models, this paper first introduced the MVCap dataset, a million-scale collection of image-text pairs with diverse viewpoint variations. Building upon this groundwork, we then proposed the Omniview-Tuning framework, which incorporates a novel Cross-Viewpoint Alignment objective in a parameter-efficient manner, effectively enhancing the VLP models' ability to generate viewpoint-invariant representations. Moreover, through extensive experiments, we successfully verified that Omniview-Tuning could bring significant improvements in viewpoint invariance while preserving the original performance. These advancements provide valuable insights and a standard for future research on viewpoint invariance in foundation models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.269 + ], + "angle": 0, + "content": "This work was supported by the Project of the National Natural Science Foundation of China (No. 62076018, 92370124, 62276149, 92248303), the Fundamental Research Funds for the Central Universities, and Tsinghua-Alibaba Joint Research Program. Y. Dong was also supported by the China National Postdoctoral Program for Innovative Talents and CCF-BaiChuan-Ebtech Foundation Model Fund." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.293, + 0.323, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.38 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.436 + ], + "angle": 0, + "content": "2. Alcorn, M.A., Li, Q., Gong, Z., Wang, C., Mai, L., Ku, W.S., Nguyen, A.: Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4845-4854 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.492 + ], + "angle": 0, + "content": "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.548 + ], + "angle": 0, + "content": "4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.576 + ], + "angle": 0, + "content": "5. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "6. Calian, D.A., Stimberg, F., Wiles, O., Rebuffi, S.A., Gyorgy, A., Mann, T., Gowal, S.: Defending against image corruptions through adversarial augmentations. arXiv preprint arXiv:2104.01086 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.659 + ], + "angle": 0, + "content": "7. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "8. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "9. Chuang, Y.S., Dangovski, R., Luo, H., Zhang, Y., Chang, S., Soljacic, M., Li, S.W., Yih, W.t., Kim, Y., Glass, J.: Diffcse: Difference-based contrastive learning for sentence embeddings. arXiv preprint arXiv:2204.10298 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.799 + ], + "angle": 0, + "content": "10. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21126-21136 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.326, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "12. Das, S., Ryoo, M.S.: ViewCLR: Learning self-supervised video representation for unseen viewpoints. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5573-5583 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "15. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1022-1032 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "16. Dong, Y., Ruan, S., Su, H., Kang, C., Wei, X., Zhu, J.: View fool: Evaluating the robustness of visual recognition to adversarial viewpoints. Advances in Neural Information Processing Systems 35, 36789-36803 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "17. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: Proceedings of the 39th International Conference on Machine Learning. pp. 6216-6234 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "18. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "19. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "20. Hamdi, A., Ghanem, B.: Towards analyzing semantic robustness of deep neural networks. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 22-38. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.619 + ], + "angle": 0, + "content": "21. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "22. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.743 + ], + "angle": 0, + "content": "24. Ho, C.H., Leung, B., Sandstrom, E., Chang, Y., Vasconcelos, N.: Catastrophic child's play: Easy to perform, hard to defend adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9229-9237 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "25. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "26. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.203 + ], + "angle": 0, + "content": "28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.245 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.327 + ], + "angle": 0, + "content": "31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.355 + ], + "angle": 0, + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.787, + 0.383 + ], + "angle": 0, + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.787, + 0.424 + ], + "angle": 0, + "content": "34. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.787, + 0.466 + ], + "angle": 0, + "content": "35. Madan, S., Henry, T., Dozier, J., Ho, H., Bhandari, N., Sasaki, T., Durand, F., Pfister, H., Boix, X.: When and how cnns generalize to out-of-distribution category-viewpoint combinations. arXiv preprint arXiv:2007.08032 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.787, + 0.508 + ], + "angle": 0, + "content": "36. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (ICLR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.536 + ], + "angle": 0, + "content": "37. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "38. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.787, + 0.618 + ], + "angle": 0, + "content": "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.674 + ], + "angle": 0, + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.715 + ], + "angle": 0, + "content": "41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Proceedings of the 38th International Conference on Machine Learning. pp. 8821-8831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.787, + 0.757 + ], + "angle": 0, + "content": "42. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.757, + 0.787, + 0.813 + ], + "angle": 0, + "content": "43. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "44. Rong, X.: word2vec parameter learning explained. arXiv preprint arXiv:1411.2738 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Ruan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "46. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Towards viewpoint-invariant visual recognition via adversarial training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4709-4719 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.326 + ], + "angle": 0, + "content": "48. Schlarmann, C., Singh, N.D., Croce, F., Hein, M.: Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.785, + 0.38 + ], + "angle": 0, + "content": "49. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.785, + 0.42 + ], + "angle": 0, + "content": "50. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.447 + ], + "angle": 0, + "content": "51. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "52. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "53. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "54. Tu, W., Deng, W., Gedeon, T.: A closer look at the robustness of contrastive language-image pre-training (CLIP). In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "55. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "56. Wang, X., He, K., Gupta, A.: Transitive invariance for self-supervised visual representation learning. In: Proceedings of the IEEE international conference on computer vision. pp. 1329-1338 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "57. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.704 + ], + "angle": 0, + "content": "58. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.759 + ], + "angle": 0, + "content": "59. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "60. Zhang, Y., Huang, Y., Sun, Y., Liu, C., Zhao, Z., Fang, Z., Wang, Y., Chen, H., Yang, X., Wei, X., et al.: Benchmarking trustworthiness of multimodal large language models: A comprehensive study. arXiv preprint arXiv:2406.07057 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "61. Zhao, B., Yu, S., Ma, W., Yu, M., Mei, S., Wang, A., He, J., Yuille, A., Kortylewski, A.: Ood-cv: A benchmark for robustness to individual nuisances in real-world out-of-distribution shifts. In: ICML 2022 Shift Happens Workshop (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Omniview-Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.19 + ], + "angle": 0, + "content": "62. Zhou, X., Liu, M., Zagar, B.L., Yurtsever, E., Knoll, A.C.: Vision language models in autonomous driving and intelligent transportation systems. arXiv preprint arXiv:2310.14414 (2023)" + } + ] +] \ No newline at end of file diff --git a/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_origin.pdf b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..04f90bcbfe8d5caa206894edf2256be767ee7152 --- /dev/null +++ b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/9e6d386a-8f34-4571-9fc7-5ebd88f69b97_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fce67bba14fc488852663e1f3b779a1a65c52cd9c365802b5d73dd77d0e841b +size 4197935 diff --git a/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/full.md b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e182e0ae1565507ada5a88845b210c50665526bc --- /dev/null +++ b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/full.md @@ -0,0 +1,320 @@ +# Omniview-Tuning: Boosting Viewpoint Invariance of Vision-Language Pre-training Models + +Shouwei Ruan1, Yinpeng Dong2,5, Hanqing Liu3, Yao Huang1, Hang Su2,4, and Xingxing Wei1,3* + +$^{1}$ Institute of Artificial Intelligence, Beihang University, Beijing 100191, China + +$^{2}$ Dept. of Comp. Sci. & Tech., Institute for AI, BNrist Center, THBI Lab, + +Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, 100084, China + +3 Hangzhou Innovation Institute, Beihang University, Hangzhou 311228, China + +4 Zhongguancun Laboratory, Beijing, 100080, China + +5 RealAI + +{shouweiruan,hqliu,y_huang,xxwei}@buaa.edu.cn,{dongyinpeng,suhangss}@tsinghua.edu.cn + +Abstract. Vision-Language Pre-training (VLP) models like CLIP have achieved remarkable success in computer vision and particularly demonstrated superior robustness to distribution shifts of 2D images. However, their robustness under 3D viewpoint variations is still limited, which can hinder the development for real-world applications. This paper successfully addresses this concern while keeping VLPs' original performance by breaking through two primary obstacles: 1) the scarcity of training data and 2) the suboptimal fine-tuning paradigms. To combat data scarcity, we build the Multi-View Caption (MVCap) dataset — a comprehensive collection of over four million multi-view image-text pairs across more than 100K objects, providing more potential for VLP models to develop generalizable viewpoint-invariant representations. To address the limitations of existing paradigms in performance trade-offs and training efficiency, we design a novel fine-tuning framework named Omniview-Tuning (OVT). Specifically, OVT introduces a Cross-Viewpoint Alignment objective through a minimax-like optimization strategy, which effectively aligns representations of identical objects from diverse viewpoints without causing overfitting. Additionally, OVT fine-tunes VLP models in a parameter-efficient manner, leading to minimal computational cost. Extensive experiments on various VLP models with different architectures validate that OVT significantly improves the models' resilience to viewpoint shifts and keeps the original performance, establishing a pioneering standard for boosting the viewpoint invariance of VLP models. The code and dataset are available via https://github.com/Heathcliff-saku/Omniview_Tuning + +Keywords: Vision-Language Pre-training $\cdot$ Viewpoint Invariance + +![](images/dd8e9d5f691bd16af676e8d99a9563cf46c77023904d859ca71d05d7d7b9f1d5.jpg) +Fig. 1: The Challenge of Viewpoint Invariance in VLP. We selected benchmarks representing clean distributions (ImageNet-1K [14], CIFAR-100 [27]), common 2D-OOD (ImageNet-V2 [42], ImageNet-R(endition) [21], ImageNet-Sketch [55]), and viewpoint-OOD (ImageNet-V(微观)+ [46], OOD-CV(Pose) [61], MIRO [7]). We display samples from these data distributions (left) and report the Top-1 accuracy of the original CLIP (ViT-L/14) and our improved OVT-CLIP (ViT-L/14) (right). + +![](images/8730a5b29bf7ea8c98929862ba32a92437859482f0ad87e48c984d656b4a10b8.jpg) + +# 1 Introduction + +Vision-Language Pre-training (VLP) models, such as CLIP [40] and BLIP [29], have shown great promise in learning transferable representations across various tasks. By aligning images and texts in a joint embedding space with a large corpus of paired image-text data, VLP models exhibit exceptional representation and generalization capabilities that surpass traditional task-specific models. Owing to this, the VLP models serve as foundation models for numerous tasks, including visual recognition [40], visual question answering [1,33], and image generation [41,47]. Moreover, these models can effectively integrate real-world visual inputs with humankind instructions, leading to their increasing use in physical-world applications, like autonomous driving [62] and embodied robotics [28,57]. + +Besides their expressive power, VLP models have also shown excellent robustness under out-of-distribution (OOD) data [17, 40, 54, 60], including common corruptions [6, 15, 22], stylistic changes [21, 55], and natural distribution shifts [21, 23, 42]. However, a recent study [45] identifies that although VLP models excel at handling 2D-OOD samples, they suffer significant performance degradation under 3D viewpoint changes, revealing a notable shortcoming of the existing VLP models. As demonstrated in Fig. 1, when dealing with the benchmarks concerned with 3D viewpoint shifts [16, 46, 61], CLIP's performance is obviously lower than that on 2D-OOD benchmarks. This large gap likely stems from limited coverage of diverse viewpoints in the training datasets [18, 49, 53], which is crucial for learning viewpoint-invariant representations. As VLP models are increasingly deployed in real-world environments where viewpoint shifts often occur, enhancing their resilience to such changes is urgent and essential. + +To address this problem, this paper sets out to enhance the viewpoint invariance of VLP models while preserving the original performance as much as possible. However, achieving this goal meets the following challenges: (1) Data scarcity: acquiring VLP training data that covers a wide range of viewpoint variations is particularly challenging compared to conventional image-text pair data. Although some datasets introduced for task-specific models do include viewpoint + +![](images/0600b22fc4662a7f0843e15114af70f943e67ed0e86a56fd09cc3570769254a0.jpg) +Fig. 2: Method Overview. (A) We create the first multi-view image caption dataset by collecting multi-view samples from existing 3D object and video datasets, and generating category-guided descriptions using VLLMs. (B) The proposed Omniview-Tuning takes multi-view image caption data as input, employs the cross-view alignment objective to encourage the model to learn viewpoint-invariant representations, and achieves efficient fine-tuning by updating Viformer and LoRA parameters. + +variations [4, 23, 35, 61], they often lack the textual descriptions vital for VLP. Even the largest available multi-view datasets [10, 46, 59] fall short in terms of scale, category coverage, and viewpoint diversity, thereby limiting the potential for VLP models to develop generalizable viewpoint-invariant representations. (2) Inappropriate paradigms: traditional approaches, which often regard viewpoint changes as adversarial attacks and employ adversarial training paradigms for enhancing invariance [2, 45, 46], are not entirely suitable for VLP models. Such frameworks typically entail a trade-off between robustness and accuracy—a balance that requires more careful consideration for foundation VLP models, where our aim is not solely to improve viewpoint invariance but, more importantly, to bridge the gap between it and the original performance. Furthermore, these approaches necessitate extra 3D reconstruction and neural rendering to capture adversarial viewpoints, leading to prohibitive computational costs for large-scale VLP models. For instance, tuning ResNet-50 with VIAT [46] under a dataset of just 1K objects demands around 400 GPU hours. Therefore, it is important to make training more efficient and less resource-intensive. + +Based on the above discussions, this paper conducts a pioneering exploration of the viewpoint invariance of VLP models. Specifically, we address the aforementioned challenges by making the following contributions: + +Million-scale multi-view image-text training set. We introduce a large-scale Multi-View Caption (MVCap) dataset tailored for viewpoint invariance of VLP models, comprising over 4.6 million multi-view image-text pairs across more than 100K objects. To assemble a diverse collection of multi-view image-text pairs, we amalgamate various 3D assets with real-world multi-view data. This process involves an extensive selection and rendering of multi-view images from existing datasets. We then utilize a Vision Large Language Model (VLLM) for automated caption generation to obtain semantically rich textual descriptions without extensive manual efforts. To ensure category consistency across varying viewpoints in the generated captions, we implement a category-guided prompting strategy, which maintains accuracy in textual descriptions for different viewpoints of the same object or scene (details in Sec. 3). + +Effective framework for enhancing VLP's viewpoint invariance. We propose Omniview-Tuning (OVT), a novel framework designed to enhance the viewpoint invariance of prevalent VLP models. As illustrated in Fig. 2, OVT employs multi-view image-text pairs for training additional learnable components. To amplify the model's proficiency in learning viewpoint-invariant representations, we introduce a Cross-viewpoint Alignment objective, ensuring that representations of the same object from different viewpoints are close and unified in the high-dimensional feature space. To prevent performance trade-offs due to the concept drift from aggressive viewpoint alignment, we innovatively construct the optimization paradigm of OVT in a minimax-like form. The optimization process includes identifying extreme outlier viewpoints during the maximization step, while optimizing the model's invariant representation for these outlier samples in the minimization step. This strategy enables the model to focus more on the worst-case viewpoint samples, thereby maximally preserving the original embedding distribution and avoiding performance degradation while saving computational costs. Moreover, OVT is designed in a Parameter-Efficient Fine-Tuning manner to improve efficiency, and creatively incorporates two trainable parameter modules: an embedding transformation module named VIForner and the Low-Rank Adaptation (LoRA [25]) weights, to acquire additional viewpoint invariance capabilities efficiently. + +Extensive experiments across various VLP architectures and tasks. We conduct extensive experiments to show the efficacy of the OVT framework in improving the viewpoint invariance for VLP models while maintaining performance on clean data and 2D-OOD samples. For example, by fine-tuning CLIP with OVT on different architectures (ViT-B/32, ViT-B/16, and ViT-L/14), the Top-1 accuracy on viewpoint-OOD benchmarks increased by an average of $9.6\%$ , $10.2\%$ , and $8.9\%$ , respectively, with only a minimal sacrifice on 2D-OOD benchmarks by an average of $2.6\%$ , $1.4\%$ , and $0.2\%$ . Furthermore, serving as the visual encoder in VLLMs (e.g., LLaVa [33]), OVT-CLIP also effectively improves viewpoint invariance in image captioning and visual question answering tasks. + +# 2 Related Work + +# 2.1 Viewpoint Invariance and Robustness + +Viewpoint invariance is a key property of human vision [5] but is usually lacking in computer vision models [2, 16]. Addressing viewpoint invariance and robustness involves strategies like data augmentation and adversarial learning. Early efforts aim to enhance viewpoint robustness by incorporating datasets enriched with viewpoint variations [4, 23, 35, 61]. For example, Madan et al. encourage models to learn viewpoint-robust representations by incorporating object-pose combinations [35]. However, these methods often falter under malicious viewpoint perturbations due to their inability to capture the worst-case viewpoint samples. Recently, achieving viewpoint invariance within the adversarial training paradigm has shown promise [2, 16, 20, 46]. By treating viewpoint variations as an adversarial attack, Alcorn et al. employ a differentiable renderer to train models against adversarial viewpoints optimized from a limited 3D objects set [2]. Recent studies, such as Viewfool [16] and VIAT [45, 46], have introduced neural radiance field (NeRF) [38, 39], enabling the characterization of adversarial viewpoint distributions from 2D multi-view inputs. Besides, studies in the self-supervised domain employ a label-free paradigm to improve viewpoint invariance, but these methods often require the introduction of complex structures, such as large-scale graph structures [56] and viewpoint generators [12]. Distinct from previous studies, our work pioneers the improvement of viewpoint invariance representation within large-scale VLP models, which is facilitated through suitable training data and refined fine-tuning methodologies. + +# 2.2 Vision-Language Pre-training + +In the realm of VLP, significant strides have been made in understanding and bridging the semantic gap between visual and textual information. Despite the variety of existing VLP paradigms, such as single-stream encoder (e.g., VisualBERT [31] and UNITER [8], etc.) or dual-stream encoder equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by ALIGN [30] and OpenAI's CLIP [40] dominates the field. CLIP, in particular, has gained widespread attention for its ability to perform zero-shot classification tasks by adopting a vast corpus of internet-collected image-text pairs, demonstrating the power of large-scale contrastive pre-training. Thus, Our investigation primarily focuses on these VLP architectures. Building upon these foundational works, subsequent iterations like open-CLIP [26], EVA-CLIP [51,52], and Meta-CLIP [58] have introduced nuanced enhancements. These refinements, ranging from the incorporation of more expansive high-quality image-text datasets and improved training methodologies, have collectively contributed to performance uplifts. BLIP [29], meanwhile, introduces a bootstrapping mechanism by the proposed captioner and filter module that achieve significant performance improvements on various downstream tasks. + +Table 1: Comparison of current large-scale multi-view datasets. "Spherical" indicates whether the viewpoints cover spherical space, "Diversity" assesses the diversity of viewpoints, and "Caption" indicates whether textual descriptions are provided. + +
DatasetYear#Obj.#Cat.#Avg. View.#SampleImage DomainSphericalDiversityCaption
OOWL [24]201950025240120KRealX★★X
CO3D [43]202118.6k50~801.5MRealX★X
ABO [10]20227.9k6330238KSyntheticX★X
IM3D [46]20231.0k100100100KSynthetic✓★★★X
MVImgNet [59]2023219.0k238~306.5MRealX★★X
MVCap (Ours)202494.6k1600100/~304.6MSynthetic+Real✓★★★✓
+ +# 3 Multi-view Caption Dataset + +We recognize that one of the key challenges in achieving viewpoint invariance for VLP is the scarcity of training data that offer comprehensive viewpoints. As shown in Tab. 1, existing large-scale multi-view datasets [10,24,43,46,59] typically lack in either sample diversity, category breadth, or textual descriptions, limiting their effectiveness for supporting VLP models to achieve viewpoint invariance. To address these limitations, we next introduce the MVCap dataset. + +# 3.1 Multi-View Image Collection + +We commence by gathering a multi-view image collection $\mathcal{D} = \{I_{ij} \mid i = 1,2,\dots,N; j = 1,2,\dots,M_i\}$ , where $N$ and $M_i$ represent the counts of objects and their viewpoints, respectively. To cover various categories from virtual to real-world scenes, we integrate samples from Objaverse [13], IM3D [46], and MVImgNet [59]. Since the original 3D dataset includes a fair share of noisy and semantically indistinct objects, we leverage semantic embeddings provided by OpenShape [34] to conduct cosine similarity sorting based on the embeddings of customized labels. Finally, we filter 24,495 virtual 3D objects endowed with distinct semantic clarity and cover over 1,600 categories. For each chosen 3D object, we employ Blender to render 100 random viewpoint images from the upper hemisphere, ensuring a comprehensive and varied viewpoint representation in our collected samples. We also incorporate objects from MVImgNet with over 30 valid viewpoints (video frames), thereby acquiring a substantial number of real-world multi-view samples to enrich the dataset's content and quality further. + +# 3.2 Category-Guided Caption Generation + +The granularity and precision of textual descriptions are pivotal in VLP training, as they influence the model's generalization capabilities and the variety of visual concepts learned. Relying solely on simple prompt engineering, such as "a photo of [category]," may introduce biases and limit the model's generalizability, whereas manual annotation is costly. To circumvent this, we utilize InstructBLIP-flant5xl [11], a leading VLLM, to create multi-view captions automatically. However, such VLLMs also grapple with viewpoint invariance, where the model's responses to different viewpoints can often be category-inconsistent, + +![](images/56f637aa9ccc719991a6d62e8b11923aab6c6fdc5e0f236ed0833ff97e337b0d.jpg) +Fig. 3: Generated multi-view captions with common and category-guided prompts. + +![](images/dea7630e8ef29e614e56232b11409b5ba7ca3294d6375173085d9033740b4f06.jpg) + +![](images/3021dcae3c64a46f248edda2cd7f61114ed0fabe1aa4b3a642d39c83ae826b9d.jpg) + +as depicted in Fig. 3. This situation presents a "chicken or egg" dilemma: we hope to use a viewpoint-invariant model to supply data for viewpoint invariance training. We address this by the design of category-guided prompting. Specifically, we use prompts containing ground-truth category information to eliminate the hallucination of large VLLMs in response to viewpoint-shifted inputs, thereby generating category-consistent multi-view captions. Formally, the forward process for generating captions can be represented as follows: + +$T_{ij} = \mathcal{G}[I_{ij},\mathrm{Prompt}(c_i)]$ Prompt $(c_{i}) = "$ Write a short description for the image, (1) noting that the main instance of the image is $a < c_{i} > ."$ + +where $c_{i} \in C$ denotes the category label for the $i$ -th object, and $\mathcal{G}$ denotes the forward process of InstructBLIP. This yields the multi-view image-text pairs $\tilde{\mathcal{D}} = \{\langle I_{ij}, T_{ij} \rangle \mid i = 1, 2, \dots, N; j = 1, 2, \dots, M_i\}$ , which can be utilized for the viewpoint invariance fine-tuning of VLP models. + +# 4 omniview-Tuning + +# 4.1 Preliminaries: Contrastive Vision-Language Pre-training + +Despite the variety of existing VLP paradigms, such as single-stream or dual-stream architectures equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by CLIP et al. [30, 40] dominates the field. Thus, Our investigation primarily focuses on this VLP architecture. + +Without the loss of generality, these VLP models are composed of a visual encoder $E_{\mathbf{W}_{\mathbf{v}}}: I \to z^{I} \in \mathbb{R}^{d}$ and a text encoder $E_{\mathbf{W}_{\mathbf{t}}} : T \to z^{T} \in \mathbb{R}^{d}$ , which maps visual and textual inputs to a unified high-dimensional feature space $\mathbb{R}^d$ , respectively, where $\mathbf{W}_{\mathbf{v}}$ and $\mathbf{W}_{\mathbf{t}}$ are weight matrices of two encoders. Given a large corpus of image-text pairs $\{\langle I_i, T_i \rangle\}_{i=1}^N$ , VLP models typically employ an image-text contrastive (ITC) loss as the training objective: + +$$ +\mathcal {L} _ {I T C} = \frac {1}{2} \left(\mathcal {L} _ {I \rightarrow T} + \mathcal {L} _ {T \rightarrow I}\right), \tag {2} +$$ + +which is composed of an image-to-text and a text-to-image terms formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {I \rightarrow T} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp \left(d \left(z _ {i} ^ {I} , z _ {i} ^ {T}\right) / \tau\right)}{\sum_ {k = 1} ^ {N} \exp \left(d \left(z _ {i} ^ {I} , z _ {k} ^ {T}\right) / \tau\right)}, \tag {3} \\ \mathcal {L} _ {T \to I} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp (d (z _ {i} ^ {T} , z _ {i} ^ {I}) / \tau)}{\sum_ {k = 1} ^ {N} \exp (d (z _ {i} ^ {T} , z _ {k} ^ {I}) / \tau)}, \\ \end{array} +$$ + +where $\tau$ represents a learnable temperature parameter, $z^I$ and $z^T$ denote the image and text embeddings, respectively. The $\mathcal{L}_{ITC}$ maximizes the similarity between matched image-text pairs while minimizing the similarity for mismatched pairs, thus enabling the alignment of visual and textual information to the same feature space, bringing the embeddings of matched pairs closer. Following [29, 30, 40], the proposed Omniview-Tuning implements $\mathcal{L}_{ITC}$ for aligning multi-view images with text modalities, which is explained in the next section. + +# 4.2 Problem Formulation + +Viewpoint Invariance of Vision-Language Pre-training. In computer vision scenario, viewpoint invariance implies that model $f(\cdot)$ can provide consistent predictions or representations given any different views of the identical object or scene [46]. Formally, given a collection of multi-view images $\mathcal{D} = \{I_{ij} \mid i = 1,2,\dots,N; j = 1,2,\dots,M_i\}$ , viewpoint invariance is required: + +$$ +f \left(I _ {i j}\right) = f \left(I _ {i j ^ {\prime}}\right), \quad \forall i, j, j ^ {\prime} \text {w i t h} j \neq j ^ {\prime}, \tag {4} +$$ + +where $i$ is the index of the object/scene, $j$ and $j'$ are indexes of two viewpoint samples. However, in the context of dual-stream VLP models, this concept requires a more refined interpretation. For VLP models, viewpoint invariance necessitates that the visual representations (i.e., the embeddings inferred from the visual encoder) from different viewpoints be sufficiently close in the feature space. Assuming $I_{ij}$ and $I_{ij'}$ are images from different viewpoints of the same object, this requirement can be formulated as follows: + +$$ +d \left[ E _ {\mathbf {W} _ {\mathbf {v}}} \left(I _ {i j}\right), E _ {\mathbf {W} _ {\mathbf {v}}} \left(I _ {i j ^ {\prime}}\right) \right] \leq \epsilon , \tag {5} +$$ + +where $d(\cdot)$ denotes a distance metric in the representation space, such as cosine distance, $\epsilon$ represents the maximum variance allowed. + +Optimization Objectives of Omniview-Tuning. Although images from different viewpoints often correspond to slightly varying textual descriptions, influenced by context, grammatical structure, and linguistic ambiguity, this variation could be significantly amplified in the high-dimensional representation space [9, 44]. Therefore, relying solely on image-text alignment may not suffice to adequately align embeddings from different viewpoints. Starting from the definition of viewpoint invariance, we introduce a cross-viewpoint alignment objective within the $\mathcal{L}_{ITC}$ to directly encourage the model to learn invariant representations between different viewpoints, rather than relying on the indirect alignment through textual descriptions. This can be seen as a regularization that forces the model to obtain viewpoint invariance, even when such invariance is not explicitly articulated in the textual descriptions. With this consideration, given a multi-view training set $\mathcal{D}$ , the optimization problem is defined as follows: + +$$ +\min _ {\mathbf {W} _ {\mathbf {v}}, \mathbf {W} _ {\mathbf {t}}} \left[ \mathcal {L} _ {I T C} + \lambda \cdot \underbrace {\sum_ {i} \sum_ {j \neq j ^ {\prime}} d \left(z _ {i j} ^ {I} , z _ {i j ^ {\prime}} ^ {I}\right)} _ {\mathcal {L} _ {V C}} \right], \tag {6} +$$ + +where the first term represents the image-text alignment used in the pre-training process, while the second term signifies the cross-viewpoint alignment goal mentioned above, referred to as Viewpoint Consistency loss $(\mathcal{L}_{VC})$ , which aims to minimize the cosine distance between embeddings from different viewpoints. $\lambda$ is a hyperparameter that balances the importance of two loss terms. + +# 4.3 Optimization Strategy + +In summary, the naive way to achieve viewpoint invariance is to calculate the loss terms in Eq. (6) based on the forward process of encoders, then update the encoders' weight using gradient descent. However, it has a relatively high time complexity to solve Eq. (6) because current $\mathcal{L}_{VC}$ requires iterating over every possible combination of viewpoints. Therefore, we endeavor to provide a more effective implementation for the original optimization problem. Drawing from the advantages of adversarial training [36, 46], we frame the optimization of the $\mathcal{L}_{VC}$ in a minimax format, rewriting the original problem Eq. (6) as: + +$$ +\begin{array}{l} \min _ {\mathbf {W} _ {\mathbf {v}}, \mathbf {W} _ {\mathbf {t}}} \left[ \mathcal {L} _ {I T C} + \lambda \cdot \underbrace {\max _ {\mathcal {O} = \{O _ {i} \} _ {i = 1} ^ {N} , | O _ {i} | = K} \sum_ {i = 1} ^ {N} \sum_ {j \in \mathcal {O}} l \left(z _ {i j} ^ {I} , z _ {C _ {i}} ^ {I}\right)} \right], \tag {7} \\ \mathrm {w h e r e} l (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) = \max \left[ d (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) + m, 0 \right], \\ \end{array} +$$ + +where $\mathcal{O} = \{O_i\}_{i=1}^N$ is the outlier viewpoints set, $z_{C_i}^I$ are anchor viewpoint embeddings of each object, and $l(\cdot)$ is the cosine distance with a margin $m$ . During the optimization, The maximization step first identifies the collection of top- $K$ outlier viewpoints $\mathcal{O}$ , which are the viewpoint samples with the highest degree of representational deviation. Then, the minimization step encourages the outlier viewpoint embeddings to converge towards corresponding anchor viewpoint embeddings $z_{C_i}^I$ . We obtain $z_{C_i}^I$ by calculating the nearest-neighbor weighted embedding centroid of each object: + +$$ +\begin{array}{r} z _ {C _ {i}} ^ {I} = \sum_ {j = 1} ^ {M _ {i}} \tilde {\omega} _ {i j} \cdot z _ {i j} ^ {I}, \\ \mathrm {w h e r e} \tilde {\omega} _ {i j} = \omega_ {i j} / \sum_ {j} \omega_ {i j}, \omega_ {i j} = 1 / \sum_ {z _ {i h} ^ {I} \in \mathcal {Q} _ {i j}} d (z _ {i j} ^ {I}, z _ {i h} ^ {I}) \end{array} \tag {8} +$$ + +where $\mathcal{Q}_{ij} = \{z_{ih}^I\}_{h=1}^5$ is the top-5 nearest neighbours of each viewpoint embedding $z_{ij}^I$ . As for the outlier viewpoints set, we define them as the viewpoints with the top- $K$ farthest cosine distances from $z_{C_i}^I$ . + +The adoption of this strategy offers dual advantages: Firstly, it allows the model to focus solely on extreme outlier viewpoints, preventing concept drift and potential overfitting to the fine-tuning dataset that results from excessive alignment. Secondly, this approach reduces computational overhead and significantly enhances optimization efficiency. + +# 4.4 Parameter-Efficient Modules + +To mitigate the impact of full parameters update on the original performance and enhance training efficiency, we achieve viewpoint invariance by efficiently + +![](images/aa8940cfe67daeb44d280088a90f02feb57c89777a4f40bbcfde2385f8723805.jpg) +Fig. 4: Visualization results for zero-shot classification results. + +fine-tuning the parameters of the visual encoder while keeping the text encoder frozen. Inspired by LoRA [25], we perform low-rank decomposition on the weights of the visual encoder $\mathbf{W}_{\mathbf{v}} \in \mathbb{R}^{n \times m}$ to substitute full-parameter update: + +$$ +\tilde {\mathbf {W}} _ {\mathbf {v}} = \mathbf {W} _ {\mathbf {v}} + \Delta W = \mathbf {W} _ {\mathbf {v}} + \mathbf {B} \mathbf {A}, \quad \text {w h e r e} \mathbf {B} \in \mathbb {R} ^ {m \times r}, \mathbf {A} \in \mathbb {R} ^ {r \times n}, r \ll \min (n, m), \tag {9} +$$ + +where $\mathbf{A}$ and $\mathbf{B}$ are two learnable low-rank parameter matrices, which we apply to the self-attention layers of the visual encoder and update them during fine-tuning while freezing the original pre-trained weights. Compared to directly adjusting the original network weights (i.e., full-parameter fine-tuning manner in TeCoA [37] and FARE [48]), this approach enables us to improve the model's viewpoint invariance representation capability while better preserving the original performance. Drawing inspiration from the success of CLIP-Adapter [19], which improves CLIP's performance in few-shot scenarios by introducing linear layers after the encoder, we propose a similar module called VIformer: $f_{\pmb{\theta}}:z^{I}\in \mathbb{R}^{d}\to s^{I}\in \mathbb{R}^{d}$ , where $\pmb{\theta}$ is the weight. Unlike CLIP-Adapter, VIformer transforms the original embeddings $z^{I}$ by introducing self-attention layers in a learnable manner to extract and retain specific viewpoint-invariant key components $s^I$ . Combining LoRA and VIformer modules, the forward process of image encoding can be represented as follows: + +$$ +\begin{array}{l} \tilde {z} ^ {I} = \alpha \cdot f _ {\boldsymbol {\theta}} (z ^ {I}) + (1 - \alpha) \cdot z ^ {I} \\ = \alpha \cdot f _ {\boldsymbol {\theta}} \left(\mathbf {W} _ {\mathbf {v}} \cdot I + \mathbf {B} \mathbf {A} \cdot I\right) + (1 - \alpha) \cdot \left(\mathbf {W} _ {\mathbf {v}} \cdot I + \mathbf {B} \mathbf {A} \cdot I\right), \tag {10} \\ \end{array} +$$ + +where the constant value $\alpha$ denotes the residual ratio to balance achieving original performance and viewpoint invariance performance. Therefore, for Eq. (6), we now only need to update $\mathbf{A}$ , $\mathbf{B}$ , and $\pmb{\theta}$ , rather than the entire weights of VLP. To facilitate the understanding of the OVT training process, we provide the pseudocode for OVT as shown in Appendix D + +# 5 Experiments + +Our evaluation of Omniview-Tuning spans several downstream tasks, including zero-shot classification, image captioning, and vision question answering. + +Table 2: Configurations of OVT and zero-shot Top-1 accuracy (\%) on ImageNet-1K with ImageNet-V+. The number in parentheses shows the performance change relative to the pre-trained weights. Through OVT training, each model maintains the performance on ImageNet-1K (IN-1K) while significantly improving the performance on ImageNet-V+ (IN-V+.), narrowing the performance gap. + +
ModelPretrain WeightPretrain DataTotal #Param.Trainable #Param.Image SizeBatch Size#Iter.IN-1KIN-View+
OVT-OpenCLIP ViT-B/32OpenCLIP-ViT-B/32LAION (2B)151M6.6M22451235k67.8 (↑1.3)59.5 (↑22.4)
OVT-OpenCLIP ViT-B/16OpenCLIP-ViT-B/16LAION (2B)149M6.6M22451235k69.7 (↑2.1)61.7 (↑17.5)
OVT-OpenCLIP ViT-L/14OpenCLIP-ViT-L/14LAION (2B)428M11.8M22425620k77.3 (↑2.1)69.8 (↑16.6)
OVT-MetaCLIP ViT-B/32MetaCLIP-ViT-B/32Common crawl (2.5B)151M6.6M22451240k69.7 (↑2.1)54.8 (↑13.8)
OVT-MetaCLIP ViT-B/16MetaCLIP-ViT-B/16Common crawl (2.5B)149M6.6M22451240k73.8 (↑1.7)64.8 (↑15.2)
OVT-MetaCLIP ViT-L/14MetaCLIP-ViT-L/14Common crawl (2.5B)428M11.8M22425620k77.7 (↓1.4)75.4 (↑9.0)
OVT-BLIP ViT-B/16Salesforce-BLIP-ViT-B/16COCO et al. (129M)234M4.3M22425620k61.7 (↑8.8)54.8 (↑18.0)
+ +For zero-shot classification (Sec. 5.1), we conduct evaluations for CLIP [40] and BLIP [29] architectures. For image captioning and vision question answering (Sec. 5.2), we replace the visual encoders in Vision Large Language Models (VLLMs) with our fine-tuned versions. We adopt LLaVA-1.5 [32,33], and OpenFlamingo [3], the most advanced open-source VLLMs available. Additionally, we present the ablation study and convergence analysis of our approach in Sec. 5.3. + +# 5.1 Evaluation of Zero-Shot Classification + +Baselines. We adopt the official CLIP (OpenAI CLIP [40]) and the community open-source version (OpenCLIP [26]) as baselines. Additionally, we include the current state-of-the-art Eva02-CLIP [51] and MetaCLIP [58] as another set of baselines to compare CLIP trained with improved techniques and extensive training data. For BLIP, we use the official implementation [29] as the baseline. Settings. We train two series of CLIP models using our OVT framework and MVCap dataset, each series comprising three different visual encoder architectures (ViT-B/32, ViT-B/16, and ViT-L/14). OVT-OpenCLIP are fine-tuned on the original weights of OpenCLIP, while OVT-MetaCLIP are based on MetaCLIP. The fine-tuning settings are detailed in Tab. 2. We standardized the $\lambda = 1.0$ , $\alpha = 0.1$ , and the outlier viewpoints number $K = 5$ and set the LoRA rank at 8. Datasets and Metrics. We employ a various set of benchmarks for evaluation, including clean data distributions [14, 27], common 2D-OOD [21, 23, 42, 55, 61], and most importantly, viewpoint-OOD [7, 16, 46, 61]. For each benchmark, we report Top-1 and Top-5 accuracy and average accuracy across all benchmarks. The evaluations follow the standard prompting engineering of CLIP [40]. + +Results and Discussions. Tab. 3 summarizes the performance of our OVT-trained VLP models against various VLP versions, including their accuracy across different benchmarks and average accuracy across clean, common-OOD, and Viewpoint-OOD domains. We can draw the following conclusions: + +(1) OVT significantly enhances the models' invariance to Viewpoint-OOD samples. Across different VLP architectures and visual encoders, OVT-trained models perform best on almost all viewpoint-OOD benchmarks. On the average + +Table 3: Top-1/Top-5 zero-shot accuracy (%) under different benchmarks + +
ModelCleanCommon-ODViewpoint-OD
ImageNet-100 [66]ImageNet-IR [14]Cllr-100 [27]Avg. Acc.ImageNet-Vi [14]ImageNet-Sch. [53]ImageNet-ODD [25]ImageNet-Rem. [21]OOD-CV [61]Avg. Acc.ImageNet-View [68]ImageNet-View + [68]OOD-CV-Pose [63]MIBO [7]Avg. Top-1Total Avg. Acc.
A. Comparisons with VIT-B/32 baselines
OpenAI CLIP77.5/93.963.3/88.864.3/88.168.4/90.255.8/83.442.2/70.333.4/62.250.7/75.450.2/82.646.5/74.844.5/65.427.5/52.447.2/84.526.5/59.436.4/65.448.6/75.5
Open CLIP81.1/95.366.5/89.975.8/94.074.5/93.058.1/83.953.6/79.334.8/64.461.0/81.953.5/81.952.2/78.354.4/72.137.1/63.246.9/81.633.0/69.242.8/71.554.6/79.7
OVT-OpenCLIP80.9/95.667.8/90.865.0/89.371.2/91.958.0/84.245.8/73.442.8/75.050.3/71.451.7/79.549.7/76.761.9/81.259.5/85.652.8/82.535.4/80.152.4/82.456.0/82.4
MetaCLIP80.7/95.667.6/90.577.7/95.275.3/93.859.5/85.455.9/81.432.4/62.563.2/83.852.0/84.252.6/79.561.4/76.741.0/67.848.9/87.934.8/73.246.5/76.456.3/82.0
OVT-MetaCLIP80.7/95.669.7/92.071.8/93.074.0/93.560.6/85.847.8/75.843.5/73.849.0/70.850.1/80.150.2/77.264.0/79.254.8/80.455.1/84.835.6/77.052.4/80.356.9/82.3
B. Comparisons with VIT-B/16 baselines
OpenAI CLIP82.1/95.768.3/91.967.2/89.472.5/92.361.8/87.448.2/76.327.7/55.759.1/83.052.2/84.649.8/77.451.6/68.936.9/63.853.4/86.830.1/66.143.0/71.453.2/79.1
Open CLIP83.2/96.270.1/91.877.0/94.876.8/94.362.2/87.056.0/82.030.7/59.864.9/85.654.3/82.753.6/79.458.1/74.444.2/70.948.5/84.034.6/74.646.4/76.057.0/82.0
OVT-OpenCLIP83.9/97.071.9/93.169.0/90.774.9/93.664.0/88.650.5/77.936.8/68.957.0/77.256.3/84.552.9/79.465.4/80.761.7/85.856.9/87.442.4/84.956.6/84.759.6/84.7
EVA-CLIP85.3/96.574.6/94.287.5/98.082.5/96.367.0/89.857.6/82.321.3/47.369.6/87.553.1/83.153.7/78.061.8/76.644.3/69.453.9/87.432.9/73.248.2/76.659.1/82.1
MetaCLIP84.3/97.272.1/93.478.9/95.478.4/95.365.0/89.360.1/84.826.2/56.470.2/89.352.3/85.454.8/81.064.2/79.449.6/76.148.9/90.938.5/78.750.3/81.259.2/84.7
OVT-MetaCLIP83.4/97.473.8/94.173.9/93.677.0/95.065.9/89.453.6/81.036.2/68.859.0/79.651.6/83.853.2/80.169.7/84.064.8/87.355.2/87.839.2/82.957.2/85.560.5/85.6
C. Comparisons with VIT-L/14 baselines
OpenAI CLIP86.5/97.475.4/94.676.5/93.379.5/95.169.8/90.959.5/84.318.6/43.872.8/91.452.9/88.854.7/79.860.3/75.645.8/71.547.9/88.238.0/74.148.0/77.358.6/82.8
Open CLIP86.8/97.875.2/94.383.7/96.781.9/96.267.7/90.263.2/86.424.0/50.574.5/91.254.5/85.056.8/80.665.7/78.153.2/76.752.4/90.542.3/83.053.4/82.161.9/85.0
OVT-OpenCLIP89.0/97.877.3/95.379.2/95.381.8/96.169.6/91.561.9/86.027.5/55.471.3/88.756.4/87.057.3/81.772.2/86.669.8/89.757.3/94.150.0/89.362.3/89.965.1/88.1
EVA-CLIP88.5/97.979.6/96.090.6/98.686.3/97.572.8/92.768.0/89.116.3/40.082.8/95.754.7/87.458.9/81.071.5/82.361.1/81.754.4/94.539.6/86.156.6/86.165.0/86.8
MetaCLIP88.3/98.379.1/95.984.1/96.983.8/97.072.5/92.668.9/89.817.0/40.681.8/95.156.6/87.559.3/81.177.3/89.366.4/87.058.9/93.448.1/89.662.7/89.866.6/88.0
OVT-MetaCLIP88.8/97.577.7/95.984.0/96.983.5/96.870.8/92.264.4/87.920.8/47.077.0/92.756.3/89.357.8/81.879.3/90.675.4/93.057.0/94.446.4/93.864.5/92.966.5/89.3
D. Comparisons with BLIP VIT-L/B16 baselines
BLIP76.6/93.352.9/80.267.0/88.365.5/87.347.3/74.751.0/76.625.6/53.464.3/83.853.9/87.648.4/75.255.2/68.236.8/63.350.8/89.227.0/66.142.4/71.950.7/77.1
OVT-BLIP82.2/97.061.7/88.866.6/88.970.2/91.553.7/82.946.5/74.233.8/62.757.4/77.956.4/87.349.6/77.062.6/79.054.8/79.955.2/89.531.5/73.251.0/80.455.2/81.8
+ +accuracy of viewpoint-OOD datasets, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 shows improvements of $9.6\%$ , $10.2\%$ , and $8.9\%$ over OpenCLIP, respectively. OVT-BLIP demonstrated an average improvement of $8.6\%$ . (2) While enhancing viewpoint invariance, OVT maintains performance on clean samples and 2D-OOD without significant performance trade-offs. For 2D-OOD benchmarks, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 sacrifice only $2.6\%$ , $1.4\%$ , and $0.2\%$ accuracy. (3) Compared to earlier CLIP baselines, the recently developed MetaCLIP exhibits better zero-shot performance. Based on this, OVT further enhances its performance under viewpoint-OOD samples. Visualization. We showcase OVT-OpenCLIP and the original OpenCLIP prediction on several viewpoint-OOD samples. As illustrated in Fig. 4, OVT-CLIP successfully predicts the categories of images from various unusual viewpoints in all cases, whereas the original CLIP is prone to make incorrect predictions. + +# 5.2 Performance on Other Tasks + +Settings. As LLaVA and Openflamingo use the OpenAI CLIP (ViT-L/14) to encode vision inputs, we applied OVT to this model in this section and comparing with other OpenAI CLIP (ViT-L/14) versions in image captioning tasks. The training setup remains consistent with the OVT-OpenCLIP described in Tab. 2. Baselines. In addition to comparing with the original OpenAI CLIP version, we also select $TeCoA^4$ [37] and $FARE^4$ [48], robust CLIP models based on adversarial training, as baselines, which have been proven to possess good resistance to adversarial samples in image captioning task. + +Table 4: Image captioning performance under clean distribution samples and viewpoint-OOD samples from Real-world and Synthetic domains. We utilize the MP-Net [50] to calculate the similarity between generated descriptions and ground-truth labels, considering predictions successful if they exceed the similarity threshold $\beta$ . + +
ModelVisual EncoderReal-world DomainSynthetic Domain
OOD-CV (iid) [61]β@1.0β@0.5β@Adp.OOD-CV (Pose) [61]β@1.0β@0.5β@Adp.IM3D [46]β@1.0β@0.5β@Adp.ImageNet-V+ [46]β@1.0β@0.5β@Adp.
LLaVa-7bOpenAI CLIP(ViT-L/14)44.161.167.546.453.658.746.753.358.820.425.532.1
TeCoA4[37](ViT-L/14)41.958.965.536.141.649.226.330.142.68.711.622.6
FARE4[48](ViT-L/14)42.158.965.240.245.950.835.239.249.212.715.823.1
OVT-CLIP(ViT-L/14)43.559.565.946.553.659.149.454.061.826.431.941.0
LLaVa-13bOpenAI CLIP(ViT-L/14)45.468.070.648.658.660.848.756.760.821.228.432.5
TeCoA4[37](ViT-L/14)42.467.072.237.448.951.325.028.641.58.410.921.8
FARE4[48](ViT-L/14)43.966.771.141.952.154.836.141.448.612.115.920.8
OVT-CLIP(ViT-L/14)45.767.370.848.258.661.950.458.963.226.436.240.9
+ +![](images/966b2b7e73c38d4bdda01e628a71669dd308034e071e2454b3036d4c2bdc62e9.jpg) + +![](images/d6e47f114e0cbdcf420d52e2e3c71b928caa97b4f83d166f578b68c359a6600f.jpg) + +![](images/469b1c5701e4df162717f0773a2816158e876dbec0c8babe70dd526650b27211.jpg) + +![](images/7e2383e2660a341ad04013475d3c73368e55a530b4d8866c0b193260ac4ed703.jpg) + +![](images/c506a10688a0963e44f542dea37037a56ecc7232abe0974915d44eefebcbaf64.jpg) + +![](images/85a4559c15e28c2b0274591aa8605fffbb3b51938ba907a5639068216059ff65.jpg) + +![](images/6a3500f5ad9bddd5d8d6004f5daa241989d16620e66758c9ef59d6a0a65621d1.jpg) + +![](images/50f184330a8a2503329c2ffb537eebe88295d6837189dbc57cb23663fa823fe2.jpg) + +![](images/328b2d77f496f5b4420c9f5f53bb5b478edaef7b62477e9d772192f08ee72d5e.jpg) +Fig. 5: The image descriptions generated by LLaVa-13B using our OVT-CLIP and the original OpenAI CLIP as vision encoder, where red texts indicates incorrect category descriptions, and green texts represents correct. + +![](images/bb53e203ab2bfc77f45d51e3c11e732f2590994e2dc9e5ba9a5fe532b89b74a2.jpg) + +![](images/565e0550f42ea728db8158e630a725072a6e02952c402d5984a985e90d9eade4.jpg) + +![](images/1f9962b8863478dbe866f31889ade1c338cd7bc6bfe62ec33607f41560425b63.jpg) +Fig.6: The Top-1 accuracy of OVT-OpenCLIP (ViT-B/16) with the iterations increases. + +![](images/02873cab1d52a52a340c73d57b93b0f413bce70449cb1e6828f51b84cff74dea.jpg) + +Datasets and Metrics. Given the absence of caption benchmarks that include viewpoint-changing OOD samples, we conduct evaluations using existing viewpoint-OOD datasets, including real-world datasets (using OOD-CV (iid) to represent clean distribution and OOD-CV (Pose) for viewpoint-OOD) and synthetic datasets (using IM3D [46] for clean distribution and ImageNet-V+ for viewpoint-OOD). We adopt word embedding distance to calculate the accuracy of the captioning task. By adopting MPNet [50], a state-of-the-art textual embedding model, we measure the similarity between keywords in the generated description and the ground-truth categories. Then assess the accuracy by counting the number of samples that exceed a specific similarity threshold $\beta$ . + +Results and Discussions. Tab. 4 shows the image captioning accuracy of CLIP models under different training strategies, considering $\beta$ at 1.0 (indicating predictions involve ground-truth categories), 0.5, and $Adp$ . (meaning $\beta$ is equal to the average similarity in the clean distribution). We found that OVT-CLIP improves the accuracy of descriptions generated by LLaVa for viewpoint-OOD samples while maintaining its performance on corresponding clean distributions. When used as the visual encoder for the LLaVa-7B model, OVT-CLIP achieved an $8.9\%$ increase in accuracy compared to the original CLIP model weights under $\beta = Adp$ . Besides, we find that although robust CLIP versions maintain perfor + +Table 5: Average Top-1/Top-5 zero-shot accuracy (\%) under different data distributions within various ablation settings. + +
LITCVIFormerLVCTotal Avg.Clean Avg.Common-OOD Avg.Viewpoint-OOD Avg.
XXX61.985.081.996.256.880.653.482.1
✓XX61.985.7 (↑0.7)79.4 (↓2.5)95.6 (↓0.6)56.2 (↓0.6)81.4 (↑0.8)56.0 (↑2.6)83.8 (↑1.7)
✓✓X62.2 (↑0.3)86.2 (↑1.2)79.9 (↓2.0)95.4 (↓0.8)55.4 (↓1.4)81.6 (↑1.0)57.5 (↑4.1)85.2 (↑3.1)
✓✓✓65.1 (↑3.2)88.1 (↑3.1)81.8 (↓0.1)96.1 (↓0.1)57.3 (↑0.5)81.7 (↑1.1)62.3 (↑8.9)89.9 (↑7.8)
+ +mance on clean distribution samples, they experience a significant performance decline when facing viewpoint-OOD samples. We select some examples with the generated description in Fig. 5. For the visual question-answering task, we used OpenFlamingo as the VLLMs. The results are reported in the Appendix A. + +# 5.3 Ablation Studies and Additional Results + +Our ablation studies focus on the VIFormer and the $\mathcal{L}_{VC}$ within the Omniview-Tuning framework. Tab. 5 shows the Top-1/Top-5 acc of OVT-OpenCLIP (ViT-L/14) across various data distributions and different ablation settings. Beyond the original OpenCLIP, we set a baseline that only uses $\mathcal{L}_{ITC}$ for fine-tuning. Keeping other training settings fixed, reliance solely on $\mathcal{L}_{ITC}$ led to a more significant performance decline in clean and 2D-OOD samples while achieving limited viewpoint OOD performance improvement (2.6%/1.7%). The integration of VI-Former led to further improvements in viewpoint OOD accuracy (4.1%/3.1%). With the further addition of $\mathcal{L}_{VC}$ , the improvement in viewpoint OOD performance is most significant (8.9%/7.8%), and it also reduces performance sacrifices in other data distributions. More ablation analyses are available in Appendix B. Furthermore, we report OVT's training convergence, depicted in Fig. 6. We display the Top-1 accuracy evolution for OVT-OpenCLIP (ViT-B/16) across various training iterations. We observe that around 40K iterations, with a batch size of 512, are sufficient for effective convergence, thus achieving a balance in performance across different data distributions. + +# 6 Conclusions + +To tackle the challenge of 3D viewpoint in VLP models, this paper first introduced the MVCap dataset, a million-scale collection of image-text pairs with diverse viewpoint variations. Building upon this groundwork, we then proposed the Omniview-Tuning framework, which incorporates a novel Cross-Viewpoint Alignment objective in a parameter-efficient manner, effectively enhancing the VLP models' ability to generate viewpoint-invariant representations. Moreover, through extensive experiments, we successfully verified that Omniview-Tuning could bring significant improvements in viewpoint invariance while preserving the original performance. These advancements provide valuable insights and a standard for future research on viewpoint invariance in foundation models. + +# Acknowledgments + +This work was supported by the Project of the National Natural Science Foundation of China (No. 62076018, 92370124, 62276149, 92248303), the Fundamental Research Funds for the Central Universities, and Tsinghua-Alibaba Joint Research Program. Y. Dong was also supported by the China National Postdoctoral Program for Innovative Talents and CCF-BaiChuan-Ebtech Foundation Model Fund. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 23716-23736 (2022) +2. Alcorn, M.A., Li, Q., Gong, Z., Wang, C., Mai, L., Ku, W.S., Nguyen, A.: Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4845-4854 (2019) +3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019) +5. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987) +6. Calian, D.A., Stimberg, F., Wiles, O., Rebuffi, S.A., Gyorgy, A., Mann, T., Gowal, S.: Defending against image corruptions through adversarial augmentations. arXiv preprint arXiv:2104.01086 (2021) +7. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. European Conference on Computer Vision (ECCV) (2022) +8. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020) +9. Chuang, Y.S., Dangovski, R., Luo, H., Zhang, Y., Chang, S., Soljacic, M., Li, S.W., Yih, W.t., Kim, Y., Glass, J.: Diffcse: Difference-based contrastive learning for sentence embeddings. arXiv preprint arXiv:2204.10298 (2022) +10. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21126-21136 (2022) +11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) + +12. Das, S., Ryoo, M.S.: ViewCLR: Learning self-supervised video representation for unseen viewpoints. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5573-5583 (2023) +13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023) +14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +15. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1022-1032 (2023) +16. Dong, Y., Ruan, S., Su, H., Kang, C., Wei, X., Zhu, J.: View fool: Evaluating the robustness of visual recognition to adversarial viewpoints. Advances in Neural Information Processing Systems 35, 36789-36803 (2022) +17. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: Proceedings of the 39th International Conference on Machine Learning. pp. 6216-6234 (2022) +18. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) +19. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023) +20. Hamdi, A., Ghanem, B.: Towards analyzing semantic robustness of deep neural networks. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 22-38. Springer (2020) +21. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) +22. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) +23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021) +24. Ho, C.H., Leung, B., Sandstrom, E., Chang, Y., Vasconcelos, N.: Catastrophic child's play: Easy to perform, hard to defend adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9229-9237 (2019) +25. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +26. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below. + +27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) +29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021) +31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019) +32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (2023) +34. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in Neural Information Processing Systems 36 (2024) +35. Madan, S., Henry, T., Dozier, J., Ho, H., Bhandari, N., Sasaki, T., Durand, F., Pfister, H., Boix, X.: When and how cnns generalize to out-of-distribution category-viewpoint combinations. arXiv preprint arXiv:2007.08032 (2020) +36. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (ICLR) (2018) +37. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016 (2022) +38. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Proceedings of the 38th International Conference on Machine Learning. pp. 8821-8831 (2021) +42. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019) +43. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021) +44. Rong, X.: word2vec parameter learning explained. arXiv preprint arXiv:1411.2738 (2014) + +45. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528 (2023) +46. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Towards viewpoint-invariant visual recognition via adversarial training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4709-4719 (2023) +47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Advances in Neural Information Processing Systems (2022) +48. Schlarmann, C., Singh, N.D., Croce, F., Hein, M.: Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336 (2024) +49. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +50. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020) +51. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023) +52. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024) +53. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016) +54. Tu, W., Deng, W., Gedeon, T.: A closer look at the robustness of contrastive language-image pre-training (CLIP). In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +55. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019) +56. Wang, X., He, K., Gupta, A.: Transitive invariance for self-supervised visual representation learning. In: Proceedings of the IEEE international conference on computer vision. pp. 1329-1338 (2017) +57. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023) +58. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023) +59. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023) +60. Zhang, Y., Huang, Y., Sun, Y., Liu, C., Zhao, Z., Fang, Z., Wang, Y., Chen, H., Yang, X., Wei, X., et al.: Benchmarking trustworthiness of multimodal large language models: A comprehensive study. arXiv preprint arXiv:2406.07057 (2024) +61. Zhao, B., Yu, S., Ma, W., Yu, M., Mei, S., Wang, A., He, J., Yuille, A., Kortylewski, A.: Ood-cv: A benchmark for robustness to individual nuisances in real-world out-of-distribution shifts. In: ICML 2022 Shift Happens Workshop (2022) + +62. Zhou, X., Liu, M., Zagar, B.L., Yurtsever, E., Knoll, A.C.: Vision language models in autonomous driving and intelligent transportation systems. arXiv preprint arXiv:2310.14414 (2023) \ No newline at end of file diff --git a/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/images.zip b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..798cdbeb08a100c4409ba761676a84cc78001ac9 --- /dev/null +++ b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db2827525dcdfa040e4e5f0949128b0df88227863e042c6ccce0447c56821b67 +size 682032 diff --git a/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/layout.json b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2acc61602cdee92f89237542f6334fca538beef1 --- /dev/null +++ b/2024/Omniview-Tuning_ Boosting Viewpoint Invariance of Vision-Language Pre-training Models/layout.json @@ -0,0 +1,9625 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Omniview-Tuning: Boosting Viewpoint Invariance of Vision-Language Pre-training Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 169, + 475, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 169, + 475, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 169, + 475, + 194 + ], + "type": "text", + "content": "Shouwei Ruan1, Yinpeng Dong2,5, Hanqing Liu3, Yao Huang1, Hang Su2,4, and Xingxing Wei1,3*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 202, + 467, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 202, + 467, + 214 + ], + "spans": [ + { + "bbox": [ + 145, + 202, + 467, + 214 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 202, + 467, + 214 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Beihang University, Beijing 100191, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 215, + 465, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 215, + 465, + 225 + ], + "spans": [ + { + "bbox": [ + 149, + 215, + 465, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 215, + 465, + 225 + ], + "type": "text", + "content": " Dept. of Comp. Sci. & Tech., Institute for AI, BNrist Center, THBI Lab," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 226, + 466, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 226, + 466, + 236 + ], + "spans": [ + { + "bbox": [ + 145, + 226, + 466, + 236 + ], + "type": "text", + "content": "Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, 100084, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 236, + 467, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 236, + 467, + 247 + ], + "spans": [ + { + "bbox": [ + 145, + 236, + 467, + 247 + ], + "type": "text", + "content": "3 Hangzhou Innovation Institute, Beihang University, Hangzhou 311228, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 247, + 413, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 247, + 413, + 258 + ], + "spans": [ + { + "bbox": [ + 200, + 247, + 413, + 258 + ], + "type": "text", + "content": "4 Zhongguancun Laboratory, Beijing, 100080, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 287, + 258, + 326, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 258, + 326, + 267 + ], + "spans": [ + { + "bbox": [ + 287, + 258, + 326, + 267 + ], + "type": "text", + "content": "5 RealAI" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 148, + 268, + 465, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 268, + 465, + 277 + ], + "spans": [ + { + "bbox": [ + 148, + 268, + 465, + 277 + ], + "type": "text", + "content": "{shouweiruan,hqliu,y_huang,xxwei}@buaa.edu.cn,{dongyinpeng,suhangss}@tsinghua.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 319, + 453, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 319, + 453, + 594 + ], + "spans": [ + { + "bbox": [ + 159, + 319, + 453, + 594 + ], + "type": "text", + "content": "Abstract. Vision-Language Pre-training (VLP) models like CLIP have achieved remarkable success in computer vision and particularly demonstrated superior robustness to distribution shifts of 2D images. However, their robustness under 3D viewpoint variations is still limited, which can hinder the development for real-world applications. This paper successfully addresses this concern while keeping VLPs' original performance by breaking through two primary obstacles: 1) the scarcity of training data and 2) the suboptimal fine-tuning paradigms. To combat data scarcity, we build the Multi-View Caption (MVCap) dataset — a comprehensive collection of over four million multi-view image-text pairs across more than 100K objects, providing more potential for VLP models to develop generalizable viewpoint-invariant representations. To address the limitations of existing paradigms in performance trade-offs and training efficiency, we design a novel fine-tuning framework named Omniview-Tuning (OVT). Specifically, OVT introduces a Cross-Viewpoint Alignment objective through a minimax-like optimization strategy, which effectively aligns representations of identical objects from diverse viewpoints without causing overfitting. Additionally, OVT fine-tunes VLP models in a parameter-efficient manner, leading to minimal computational cost. Extensive experiments on various VLP models with different architectures validate that OVT significantly improves the models' resilience to viewpoint shifts and keeps the original performance, establishing a pioneering standard for boosting the viewpoint invariance of VLP models. The code and dataset are available via https://github.com/Heathcliff-saku/Omniview_Tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 605, + 430, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 605, + 430, + 616 + ], + "spans": [ + { + "bbox": [ + 160, + 605, + 430, + 616 + ], + "type": "text", + "content": "Keywords: Vision-Language Pre-training " + }, + { + "bbox": [ + 160, + 605, + 430, + 616 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 605, + 430, + 616 + ], + "type": "text", + "content": " Viewpoint Invariance" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 235, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 117, + 309, + 206 + ], + "blocks": [ + { + "bbox": [ + 140, + 117, + 309, + 206 + ], + "lines": [ + { + "bbox": [ + 140, + 117, + 309, + 206 + ], + "spans": [ + { + "bbox": [ + 140, + 117, + 309, + 206 + ], + "type": "image", + "image_path": "dd8e9d5f691bd16af676e8d99a9563cf46c77023904d859ca71d05d7d7b9f1d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 285 + ], + "type": "text", + "content": "Fig. 1: The Challenge of Viewpoint Invariance in VLP. We selected benchmarks representing clean distributions (ImageNet-1K [14], CIFAR-100 [27]), common 2D-OOD (ImageNet-V2 [42], ImageNet-R(endition) [21], ImageNet-Sketch [55]), and viewpoint-OOD (ImageNet-V(微观)+ [46], OOD-CV(Pose) [61], MIRO [7]). We display samples from these data distributions (left) and report the Top-1 accuracy of the original CLIP (ViT-L/14) and our improved OVT-CLIP (ViT-L/14) (right)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 117, + 473, + 206 + ], + "blocks": [ + { + "bbox": [ + 314, + 117, + 473, + 206 + ], + "lines": [ + { + "bbox": [ + 314, + 117, + 473, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 117, + 473, + 206 + ], + "type": "image", + "image_path": "8730a5b29bf7ea8c98929862ba32a92437859482f0ad87e48c984d656b4a10b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 293, + 230, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 230, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 230, + 306 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 481, + 437 + ], + "type": "text", + "content": "Vision-Language Pre-training (VLP) models, such as CLIP [40] and BLIP [29], have shown great promise in learning transferable representations across various tasks. By aligning images and texts in a joint embedding space with a large corpus of paired image-text data, VLP models exhibit exceptional representation and generalization capabilities that surpass traditional task-specific models. Owing to this, the VLP models serve as foundation models for numerous tasks, including visual recognition [40], visual question answering [1,33], and image generation [41,47]. Moreover, these models can effectively integrate real-world visual inputs with humankind instructions, leading to their increasing use in physical-world applications, like autonomous driving [62] and embodied robotics [28,57]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 481, + 593 + ], + "type": "text", + "content": "Besides their expressive power, VLP models have also shown excellent robustness under out-of-distribution (OOD) data [17, 40, 54, 60], including common corruptions [6, 15, 22], stylistic changes [21, 55], and natural distribution shifts [21, 23, 42]. However, a recent study [45] identifies that although VLP models excel at handling 2D-OOD samples, they suffer significant performance degradation under 3D viewpoint changes, revealing a notable shortcoming of the existing VLP models. As demonstrated in Fig. 1, when dealing with the benchmarks concerned with 3D viewpoint shifts [16, 46, 61], CLIP's performance is obviously lower than that on 2D-OOD benchmarks. This large gap likely stems from limited coverage of diverse viewpoints in the training datasets [18, 49, 53], which is crucial for learning viewpoint-invariant representations. As VLP models are increasingly deployed in real-world environments where viewpoint shifts often occur, enhancing their resilience to such changes is urgent and essential." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "To address this problem, this paper sets out to enhance the viewpoint invariance of VLP models while preserving the original performance as much as possible. However, achieving this goal meets the following challenges: (1) Data scarcity: acquiring VLP training data that covers a wide range of viewpoint variations is particularly challenging compared to conventional image-text pair data. Although some datasets introduced for task-specific models do include viewpoint" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 455, + 371 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 455, + 371 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 455, + 371 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 455, + 371 + ], + "type": "image", + "image_path": "0600b22fc4662a7f0843e15114af70f943e67ed0e86a56fd09cc3570769254a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 379, + 482, + 447 + ], + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 447 + ], + "type": "text", + "content": "Fig. 2: Method Overview. (A) We create the first multi-view image caption dataset by collecting multi-view samples from existing 3D object and video datasets, and generating category-guided descriptions using VLLMs. (B) The proposed Omniview-Tuning takes multi-view image caption data as input, employs the cross-view alignment objective to encourage the model to learn viewpoint-invariant representations, and achieves efficient fine-tuning by updating Viformer and LoRA parameters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": "variations [4, 23, 35, 61], they often lack the textual descriptions vital for VLP. Even the largest available multi-view datasets [10, 46, 59] fall short in terms of scale, category coverage, and viewpoint diversity, thereby limiting the potential for VLP models to develop generalizable viewpoint-invariant representations. (2) Inappropriate paradigms: traditional approaches, which often regard viewpoint changes as adversarial attacks and employ adversarial training paradigms for enhancing invariance [2, 45, 46], are not entirely suitable for VLP models. Such frameworks typically entail a trade-off between robustness and accuracy—a balance that requires more careful consideration for foundation VLP models, where our aim is not solely to improve viewpoint invariance but, more importantly, to bridge the gap between it and the original performance. Furthermore, these approaches necessitate extra 3D reconstruction and neural rendering to capture adversarial viewpoints, leading to prohibitive computational costs for large-scale VLP models. For instance, tuning ResNet-50 with VIAT [46] under a dataset of just 1K objects demands around 400 GPU hours. Therefore, it is important to make training more efficient and less resource-intensive." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Based on the above discussions, this paper conducts a pioneering exploration of the viewpoint invariance of VLP models. Specifically, we address the aforementioned challenges by making the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 300 + ], + "type": "text", + "content": "Million-scale multi-view image-text training set. We introduce a large-scale Multi-View Caption (MVCap) dataset tailored for viewpoint invariance of VLP models, comprising over 4.6 million multi-view image-text pairs across more than 100K objects. To assemble a diverse collection of multi-view image-text pairs, we amalgamate various 3D assets with real-world multi-view data. This process involves an extensive selection and rendering of multi-view images from existing datasets. We then utilize a Vision Large Language Model (VLLM) for automated caption generation to obtain semantically rich textual descriptions without extensive manual efforts. To ensure category consistency across varying viewpoints in the generated captions, we implement a category-guided prompting strategy, which maintains accuracy in textual descriptions for different viewpoints of the same object or scene (details in Sec. 3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 542 + ], + "type": "text", + "content": "Effective framework for enhancing VLP's viewpoint invariance. We propose Omniview-Tuning (OVT), a novel framework designed to enhance the viewpoint invariance of prevalent VLP models. As illustrated in Fig. 2, OVT employs multi-view image-text pairs for training additional learnable components. To amplify the model's proficiency in learning viewpoint-invariant representations, we introduce a Cross-viewpoint Alignment objective, ensuring that representations of the same object from different viewpoints are close and unified in the high-dimensional feature space. To prevent performance trade-offs due to the concept drift from aggressive viewpoint alignment, we innovatively construct the optimization paradigm of OVT in a minimax-like form. The optimization process includes identifying extreme outlier viewpoints during the maximization step, while optimizing the model's invariant representation for these outlier samples in the minimization step. This strategy enables the model to focus more on the worst-case viewpoint samples, thereby maximally preserving the original embedding distribution and avoiding performance degradation while saving computational costs. Moreover, OVT is designed in a Parameter-Efficient Fine-Tuning manner to improve efficiency, and creatively incorporates two trainable parameter modules: an embedding transformation module named VIForner and the Low-Rank Adaptation (LoRA [25]) weights, to acquire additional viewpoint invariance capabilities efficiently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Extensive experiments across various VLP architectures and tasks. We conduct extensive experiments to show the efficacy of the OVT framework in improving the viewpoint invariance for VLP models while maintaining performance on clean data and 2D-OOD samples. For example, by fine-tuning CLIP with OVT on different architectures (ViT-B/32, ViT-B/16, and ViT-L/14), the Top-1 accuracy on viewpoint-OOD benchmarks increased by an average of " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", respectively, with only a minimal sacrifice on 2D-OOD benchmarks by an average of " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". Furthermore, serving as the visual encoder in VLLMs (e.g., LLaVa [33]), OVT-CLIP also effectively improves viewpoint invariance in image captioning and visual question answering tasks." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 350, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 350, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 350, + 156 + ], + "type": "text", + "content": "2.1 Viewpoint Invariance and Robustness" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 166, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 482, + 418 + ], + "type": "text", + "content": "Viewpoint invariance is a key property of human vision [5] but is usually lacking in computer vision models [2, 16]. Addressing viewpoint invariance and robustness involves strategies like data augmentation and adversarial learning. Early efforts aim to enhance viewpoint robustness by incorporating datasets enriched with viewpoint variations [4, 23, 35, 61]. For example, Madan et al. encourage models to learn viewpoint-robust representations by incorporating object-pose combinations [35]. However, these methods often falter under malicious viewpoint perturbations due to their inability to capture the worst-case viewpoint samples. Recently, achieving viewpoint invariance within the adversarial training paradigm has shown promise [2, 16, 20, 46]. By treating viewpoint variations as an adversarial attack, Alcorn et al. employ a differentiable renderer to train models against adversarial viewpoints optimized from a limited 3D objects set [2]. Recent studies, such as Viewfool [16] and VIAT [45, 46], have introduced neural radiance field (NeRF) [38, 39], enabling the characterization of adversarial viewpoint distributions from 2D multi-view inputs. Besides, studies in the self-supervised domain employ a label-free paradigm to improve viewpoint invariance, but these methods often require the introduction of complex structures, such as large-scale graph structures [56] and viewpoint generators [12]. Distinct from previous studies, our work pioneers the improvement of viewpoint invariance representation within large-scale VLP models, which is facilitated through suitable training data and refined fine-tuning methodologies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 438, + 310, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 310, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 310, + 451 + ], + "type": "text", + "content": "2.2 Vision-Language Pre-training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "In the realm of VLP, significant strides have been made in understanding and bridging the semantic gap between visual and textual information. Despite the variety of existing VLP paradigms, such as single-stream encoder (e.g., VisualBERT [31] and UNITER [8], etc.) or dual-stream encoder equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by ALIGN [30] and OpenAI's CLIP [40] dominates the field. CLIP, in particular, has gained widespread attention for its ability to perform zero-shot classification tasks by adopting a vast corpus of internet-collected image-text pairs, demonstrating the power of large-scale contrastive pre-training. Thus, Our investigation primarily focuses on these VLP architectures. Building upon these foundational works, subsequent iterations like open-CLIP [26], EVA-CLIP [51,52], and Meta-CLIP [58] have introduced nuanced enhancements. These refinements, ranging from the incorporation of more expansive high-quality image-text datasets and improved training methodologies, have collectively contributed to performance uplifts. BLIP [29], meanwhile, introduces a bootstrapping mechanism by the proposed captioner and filter module that achieve significant performance improvements on various downstream tasks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 479, + 215 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison of current large-scale multi-view datasets. \"Spherical\" indicates whether the viewpoints cover spherical space, \"Diversity\" assesses the diversity of viewpoints, and \"Caption\" indicates whether textual descriptions are provided." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 479, + 215 + ], + "type": "table", + "html": "
DatasetYear#Obj.#Cat.#Avg. View.#SampleImage DomainSphericalDiversityCaption
OOWL [24]201950025240120KRealX★★X
CO3D [43]202118.6k50~801.5MRealX★X
ABO [10]20227.9k6330238KSyntheticX★X
IM3D [46]20231.0k100100100KSynthetic✓★★★X
MVImgNet [59]2023219.0k238~306.5MRealX★★X
MVCap (Ours)202494.6k1600100/~304.6MSynthetic+Real✓★★★✓
", + "image_path": "79a000a71b7b79afe28a4be0f938f31b588006053ab792cde997d9ac788a9328.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 234, + 320, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 234, + 320, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 320, + 248 + ], + "type": "text", + "content": "3 Multi-view Caption Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 331 + ], + "type": "text", + "content": "We recognize that one of the key challenges in achieving viewpoint invariance for VLP is the scarcity of training data that offer comprehensive viewpoints. As shown in Tab. 1, existing large-scale multi-view datasets [10,24,43,46,59] typically lack in either sample diversity, category breadth, or textual descriptions, limiting their effectiveness for supporting VLP models to achieve viewpoint invariance. To address these limitations, we next introduce the MVCap dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 347, + 307, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 307, + 360 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 307, + 360 + ], + "type": "text", + "content": "3.1 Multi-View Image Collection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "content": "We commence by gathering a multi-view image collection " + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{I_{ij} \\mid i = 1,2,\\dots,N; j = 1,2,\\dots,M_i\\}" + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "content": " represent the counts of objects and their viewpoints, respectively. To cover various categories from virtual to real-world scenes, we integrate samples from Objaverse [13], IM3D [46], and MVImgNet [59]. Since the original 3D dataset includes a fair share of noisy and semantically indistinct objects, we leverage semantic embeddings provided by OpenShape [34] to conduct cosine similarity sorting based on the embeddings of customized labels. Finally, we filter 24,495 virtual 3D objects endowed with distinct semantic clarity and cover over 1,600 categories. For each chosen 3D object, we employ Blender to render 100 random viewpoint images from the upper hemisphere, ensuring a comprehensive and varied viewpoint representation in our collected samples. We also incorporate objects from MVImgNet with over 30 valid viewpoints (video frames), thereby acquiring a substantial number of real-world multi-view samples to enrich the dataset's content and quality further." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 550, + 350, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 350, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 350, + 563 + ], + "type": "text", + "content": "3.2 Category-Guided Caption Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "The granularity and precision of textual descriptions are pivotal in VLP training, as they influence the model's generalization capabilities and the variety of visual concepts learned. Relying solely on simple prompt engineering, such as \"a photo of [category],\" may introduce biases and limit the model's generalizability, whereas manual annotation is costly. To circumvent this, we utilize InstructBLIP-flant5xl [11], a leading VLLM, to create multi-view captions automatically. However, such VLLMs also grapple with viewpoint invariance, where the model's responses to different viewpoints can often be category-inconsistent," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 119, + 263, + 176 + ], + "blocks": [ + { + "bbox": [ + 139, + 119, + 263, + 176 + ], + "lines": [ + { + "bbox": [ + 139, + 119, + 263, + 176 + ], + "spans": [ + { + "bbox": [ + 139, + 119, + 263, + 176 + ], + "type": "image", + "image_path": "56f637aa9ccc719991a6d62e8b11923aab6c6fdc5e0f236ed0833ff97e337b0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 190, + 475, + 201 + ], + "lines": [ + { + "bbox": [ + 136, + 190, + 475, + 201 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 475, + 201 + ], + "type": "text", + "content": "Fig. 3: Generated multi-view captions with common and category-guided prompts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 118, + 340, + 178 + ], + "blocks": [ + { + "bbox": [ + 265, + 118, + 340, + 178 + ], + "lines": [ + { + "bbox": [ + 265, + 118, + 340, + 178 + ], + "spans": [ + { + "bbox": [ + 265, + 118, + 340, + 178 + ], + "type": "image", + "image_path": "dea7630e8ef29e614e56232b11409b5ba7ca3294d6375173085d9033740b4f06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 344, + 121, + 475, + 176 + ], + "blocks": [ + { + "bbox": [ + 344, + 121, + 475, + 176 + ], + "lines": [ + { + "bbox": [ + 344, + 121, + 475, + 176 + ], + "spans": [ + { + "bbox": [ + 344, + 121, + 475, + 176 + ], + "type": "image", + "image_path": "3021dcae3c64a46f248edda2cd7f61114ed0fabe1aa4b3a642d39c83ae826b9d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 210, + 480, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 480, + 294 + ], + "type": "text", + "content": "as depicted in Fig. 3. This situation presents a \"chicken or egg\" dilemma: we hope to use a viewpoint-invariant model to supply data for viewpoint invariance training. We address this by the design of category-guided prompting. Specifically, we use prompts containing ground-truth category information to eliminate the hallucination of large VLLMs in response to viewpoint-shifted inputs, thereby generating category-consistent multi-view captions. Formally, the forward process for generating captions can be represented as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 180, + 302, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 302, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 180, + 302, + 480, + 339 + ], + "type": "inline_equation", + "content": "T_{ij} = \\mathcal{G}[I_{ij},\\mathrm{Prompt}(c_i)]" + }, + { + "bbox": [ + 180, + 302, + 480, + 339 + ], + "type": "text", + "content": " Prompt " + }, + { + "bbox": [ + 180, + 302, + 480, + 339 + ], + "type": "inline_equation", + "content": "(c_{i}) = \"" + }, + { + "bbox": [ + 180, + 302, + 480, + 339 + ], + "type": "text", + "content": " Write a short description for the image, (1) noting that the main instance of the image is " + }, + { + "bbox": [ + 180, + 302, + 480, + 339 + ], + "type": "inline_equation", + "content": "a < c_{i} > .\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "inline_equation", + "content": "c_{i} \\in C" + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "text", + "content": " denotes the category label for the " + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "text", + "content": "-th object, and " + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "text", + "content": " denotes the forward process of InstructBLIP. This yields the multi-view image-text pairs " + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}} = \\{\\langle I_{ij}, T_{ij} \\rangle \\mid i = 1, 2, \\dots, N; j = 1, 2, \\dots, M_i\\}" + }, + { + "bbox": [ + 131, + 348, + 480, + 396 + ], + "type": "text", + "content": ", which can be utilized for the viewpoint invariance fine-tuning of VLP models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 414, + 260, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 260, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 260, + 427 + ], + "type": "text", + "content": "4 omniview-Tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 439, + 444, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 439, + 444, + 451 + ], + "spans": [ + { + "bbox": [ + 131, + 439, + 444, + 451 + ], + "type": "text", + "content": "4.1 Preliminaries: Contrastive Vision-Language Pre-training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 459, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 480, + 506 + ], + "type": "text", + "content": "Despite the variety of existing VLP paradigms, such as single-stream or dual-stream architectures equipped with diverse training objectives, the dual-stream contrastive learning architecture exemplified by CLIP et al. [30, 40] dominates the field. Thus, Our investigation primarily focuses on this VLP architecture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "content": "Without the loss of generality, these VLP models are composed of a visual encoder " + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "inline_equation", + "content": "E_{\\mathbf{W}_{\\mathbf{v}}}: I \\to z^{I} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "inline_equation", + "content": "E_{\\mathbf{W}_{\\mathbf{t}}} : T \\to z^{T} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "content": ", which maps visual and textual inputs to a unified high-dimensional feature space " + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{v}}" + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{t}}" + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "content": " are weight matrices of two encoders. Given a large corpus of image-text pairs " + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\{\\langle I_i, T_i \\rangle\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 506, + 480, + 578 + ], + "type": "text", + "content": ", VLP models typically employ an image-text contrastive (ITC) loss as the training objective:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 247, + 586, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 586, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 247, + 586, + 480, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I T C} = \\frac {1}{2} \\left(\\mathcal {L} _ {I \\rightarrow T} + \\mathcal {L} _ {T \\rightarrow I}\\right), \\tag {2}", + "image_path": "9bfe88beaebaba75aa1c03a41e5ef60949380ae1e5d356f8cd53196db17e4c8a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 609, + 479, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 479, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 479, + 620 + ], + "type": "text", + "content": "which is composed of an image-to-text and a text-to-image terms formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 212, + 629, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 629, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 212, + 629, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {I \\rightarrow T} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(d \\left(z _ {i} ^ {I} , z _ {i} ^ {T}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(d \\left(z _ {i} ^ {I} , z _ {k} ^ {T}\\right) / \\tau\\right)}, \\tag {3} \\\\ \\mathcal {L} _ {T \\to I} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (d (z _ {i} ^ {T} , z _ {i} ^ {I}) / \\tau)}{\\sum_ {k = 1} ^ {N} \\exp (d (z _ {i} ^ {T} , z _ {k} ^ {I}) / \\tau)}, \\\\ \\end{array}", + "image_path": "66dc69ae2a3153b0c6dd1e55905a9b7ffec96c7dd41b5fd5f1bf65588498e989.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "text", + "content": " represents a learnable temperature parameter, " + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "inline_equation", + "content": "z^I" + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "inline_equation", + "content": "z^T" + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "text", + "content": " denote the image and text embeddings, respectively. The " + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ITC}" + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "text", + "content": " maximizes the similarity between matched image-text pairs while minimizing the similarity for mismatched pairs, thus enabling the alignment of visual and textual information to the same feature space, bringing the embeddings of matched pairs closer. Following [29, 30, 40], the proposed Omniview-Tuning implements " + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ITC}" + }, + { + "bbox": [ + 130, + 115, + 482, + 201 + ], + "type": "text", + "content": " for aligning multi-view images with text modalities, which is explained in the next section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 269, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 269, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 269, + 228 + ], + "type": "text", + "content": "4.2 Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 481, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 298 + ], + "type": "text", + "content": "Viewpoint Invariance of Vision-Language Pre-training. In computer vision scenario, viewpoint invariance implies that model " + }, + { + "bbox": [ + 130, + 237, + 481, + 298 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 237, + 481, + 298 + ], + "type": "text", + "content": " can provide consistent predictions or representations given any different views of the identical object or scene [46]. Formally, given a collection of multi-view images " + }, + { + "bbox": [ + 130, + 237, + 481, + 298 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{I_{ij} \\mid i = 1,2,\\dots,N; j = 1,2,\\dots,M_i\\}" + }, + { + "bbox": [ + 130, + 237, + 481, + 298 + ], + "type": "text", + "content": ", viewpoint invariance is required:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 224, + 307, + 481, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 307, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 224, + 307, + 481, + 321 + ], + "type": "interline_equation", + "content": "f \\left(I _ {i j}\\right) = f \\left(I _ {i j ^ {\\prime}}\\right), \\quad \\forall i, j, j ^ {\\prime} \\text {w i t h} j \\neq j ^ {\\prime}, \\tag {4}", + "image_path": "03ef8246d8a28655c368703375fdecbd20c2b779c70e141d96512dcfc6f3b1e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " is the index of the object/scene, " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "j'" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " are indexes of two viewpoint samples. However, in the context of dual-stream VLP models, this concept requires a more refined interpretation. For VLP models, viewpoint invariance necessitates that the visual representations (i.e., the embeddings inferred from the visual encoder) from different viewpoints be sufficiently close in the feature space. Assuming " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "I_{ij}" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "I_{ij'}" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " are images from different viewpoints of the same object, this requirement can be formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 245, + 422, + 481, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 422, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 245, + 422, + 481, + 441 + ], + "type": "interline_equation", + "content": "d \\left[ E _ {\\mathbf {W} _ {\\mathbf {v}}} \\left(I _ {i j}\\right), E _ {\\mathbf {W} _ {\\mathbf {v}}} \\left(I _ {i j ^ {\\prime}}\\right) \\right] \\leq \\epsilon , \\tag {5}", + "image_path": "7d3bc1548d695c982e425d5f16e9e006827afc2ade6ec02ba8f5df6c9c555ea9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 449, + 481, + 472 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 130, + 449, + 481, + 472 + ], + "type": "text", + "content": " denotes a distance metric in the representation space, such as cosine distance, " + }, + { + "bbox": [ + 130, + 449, + 481, + 472 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 449, + 481, + 472 + ], + "type": "text", + "content": " represents the maximum variance allowed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": "Optimization Objectives of Omniview-Tuning. Although images from different viewpoints often correspond to slightly varying textual descriptions, influenced by context, grammatical structure, and linguistic ambiguity, this variation could be significantly amplified in the high-dimensional representation space [9, 44]. Therefore, relying solely on image-text alignment may not suffice to adequately align embeddings from different viewpoints. Starting from the definition of viewpoint invariance, we introduce a cross-viewpoint alignment objective within the " + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ITC}" + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": " to directly encourage the model to learn invariant representations between different viewpoints, rather than relying on the indirect alignment through textual descriptions. This can be seen as a regularization that forces the model to obtain viewpoint invariance, even when such invariance is not explicitly articulated in the textual descriptions. With this consideration, given a multi-view training set " + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": ", the optimization problem is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 639, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 639, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 217, + 639, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {W} _ {\\mathbf {v}}, \\mathbf {W} _ {\\mathbf {t}}} \\left[ \\mathcal {L} _ {I T C} + \\lambda \\cdot \\underbrace {\\sum_ {i} \\sum_ {j \\neq j ^ {\\prime}} d \\left(z _ {i j} ^ {I} , z _ {i j ^ {\\prime}} ^ {I}\\right)} _ {\\mathcal {L} _ {V C}} \\right], \\tag {6}", + "image_path": "a107f4a3af7aad0c141319a0be9c677a9b0b6099621ec6297a67b42dd0247cb6.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "where the first term represents the image-text alignment used in the pre-training process, while the second term signifies the cross-viewpoint alignment goal mentioned above, referred to as Viewpoint Consistency loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{VC})" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", which aims to minimize the cosine distance between embeddings from different viewpoints. " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is a hyperparameter that balances the importance of two loss terms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 191, + 274, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 191, + 274, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 274, + 205 + ], + "type": "text", + "content": "4.3 Optimization Strategy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 307 + ], + "type": "text", + "content": "In summary, the naive way to achieve viewpoint invariance is to calculate the loss terms in Eq. (6) based on the forward process of encoders, then update the encoders' weight using gradient descent. However, it has a relatively high time complexity to solve Eq. (6) because current " + }, + { + "bbox": [ + 130, + 209, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VC}" + }, + { + "bbox": [ + 130, + 209, + 482, + 307 + ], + "type": "text", + "content": " requires iterating over every possible combination of viewpoints. Therefore, we endeavor to provide a more effective implementation for the original optimization problem. Drawing from the advantages of adversarial training [36, 46], we frame the optimization of the " + }, + { + "bbox": [ + 130, + 209, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VC}" + }, + { + "bbox": [ + 130, + 209, + 482, + 307 + ], + "type": "text", + "content": " in a minimax format, rewriting the original problem Eq. (6) as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 312, + 481, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 312, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 164, + 312, + 481, + 367 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\mathbf {W} _ {\\mathbf {v}}, \\mathbf {W} _ {\\mathbf {t}}} \\left[ \\mathcal {L} _ {I T C} + \\lambda \\cdot \\underbrace {\\max _ {\\mathcal {O} = \\{O _ {i} \\} _ {i = 1} ^ {N} , | O _ {i} | = K} \\sum_ {i = 1} ^ {N} \\sum_ {j \\in \\mathcal {O}} l \\left(z _ {i j} ^ {I} , z _ {C _ {i}} ^ {I}\\right)} \\right], \\tag {7} \\\\ \\mathrm {w h e r e} l (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) = \\max \\left[ d (z _ {i j} ^ {I}, z _ {C _ {i}} ^ {I}) + m, 0 \\right], \\\\ \\end{array}", + "image_path": "dc8ed46a12477c45efc4725d5eb8610f8a3ffa4aa69d43e26b6d65c5c5a7e8bd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{O_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": " is the outlier viewpoints set, " + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "z_{C_i}^I" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": " are anchor viewpoint embeddings of each object, and " + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "l(\\cdot)" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": " is the cosine distance with a margin " + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": ". During the optimization, The maximization step first identifies the collection of top-" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": " outlier viewpoints " + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": ", which are the viewpoint samples with the highest degree of representational deviation. Then, the minimization step encourages the outlier viewpoint embeddings to converge towards corresponding anchor viewpoint embeddings " + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "z_{C_i}^I" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": ". We obtain " + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "z_{C_i}^I" + }, + { + "bbox": [ + 130, + 373, + 482, + 469 + ], + "type": "text", + "content": " by calculating the nearest-neighbor weighted embedding centroid of each object:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 185, + 474, + 481, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 474, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 185, + 474, + 481, + 506 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} z _ {C _ {i}} ^ {I} = \\sum_ {j = 1} ^ {M _ {i}} \\tilde {\\omega} _ {i j} \\cdot z _ {i j} ^ {I}, \\\\ \\mathrm {w h e r e} \\tilde {\\omega} _ {i j} = \\omega_ {i j} / \\sum_ {j} \\omega_ {i j}, \\omega_ {i j} = 1 / \\sum_ {z _ {i h} ^ {I} \\in \\mathcal {Q} _ {i j}} d (z _ {i j} ^ {I}, z _ {i h} ^ {I}) \\end{array} \\tag {8}", + "image_path": "8fb0ff405b140c3c608b501df241201ec09caeb8f98ffe42a116f1a393831db8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{ij} = \\{z_{ih}^I\\}_{h=1}^5" + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "text", + "content": " is the top-5 nearest neighbours of each viewpoint embedding " + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "inline_equation", + "content": "z_{ij}^I" + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "text", + "content": ". As for the outlier viewpoints set, we define them as the viewpoints with the top-" + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "text", + "content": " farthest cosine distances from " + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "inline_equation", + "content": "z_{C_i}^I" + }, + { + "bbox": [ + 130, + 510, + 482, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": "The adoption of this strategy offers dual advantages: Firstly, it allows the model to focus solely on extreme outlier viewpoints, preventing concept drift and potential overfitting to the fine-tuning dataset that results from excessive alignment. Secondly, this approach reduces computational overhead and significantly enhances optimization efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 306, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 306, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 306, + 635 + ], + "type": "text", + "content": "4.4 Parameter-Efficient Modules" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "To mitigate the impact of full parameters update on the original performance and enhance training efficiency, we achieve viewpoint invariance by efficiently" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 116, + 471, + 220 + ], + "blocks": [ + { + "bbox": [ + 144, + 116, + 471, + 220 + ], + "lines": [ + { + "bbox": [ + 144, + 116, + 471, + 220 + ], + "spans": [ + { + "bbox": [ + 144, + 116, + 471, + 220 + ], + "type": "image", + "image_path": "aa8940cfe67daeb44d280088a90f02feb57c89777a4f40bbcfde2385f8723805.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 229, + 434, + 241 + ], + "lines": [ + { + "bbox": [ + 178, + 229, + 434, + 241 + ], + "spans": [ + { + "bbox": [ + 178, + 229, + 434, + 241 + ], + "type": "text", + "content": "Fig. 4: Visualization results for zero-shot classification results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "type": "text", + "content": "fine-tuning the parameters of the visual encoder while keeping the text encoder frozen. Inspired by LoRA [25], we perform low-rank decomposition on the weights of the visual encoder " + }, + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{v}} \\in \\mathbb{R}^{n \\times m}" + }, + { + "bbox": [ + 130, + 251, + 480, + 288 + ], + "type": "text", + "content": " to substitute full-parameter update:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 296, + 480, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 296, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 133, + 296, + 480, + 321 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {W}} _ {\\mathbf {v}} = \\mathbf {W} _ {\\mathbf {v}} + \\Delta W = \\mathbf {W} _ {\\mathbf {v}} + \\mathbf {B} \\mathbf {A}, \\quad \\text {w h e r e} \\mathbf {B} \\in \\mathbb {R} ^ {m \\times r}, \\mathbf {A} \\in \\mathbb {R} ^ {r \\times n}, r \\ll \\min (n, m), \\tag {9}", + "image_path": "fd7de427ee0d24cf544773ba374f1d4f9de34844ba62c1467c84a98eb4827e1f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "content": " are two learnable low-rank parameter matrices, which we apply to the self-attention layers of the visual encoder and update them during fine-tuning while freezing the original pre-trained weights. Compared to directly adjusting the original network weights (i.e., full-parameter fine-tuning manner in TeCoA [37] and FARE [48]), this approach enables us to improve the model's viewpoint invariance representation capability while better preserving the original performance. Drawing inspiration from the success of CLIP-Adapter [19], which improves CLIP's performance in few-shot scenarios by introducing linear layers after the encoder, we propose a similar module called VIformer: " + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "inline_equation", + "content": "f_{\\pmb{\\theta}}:z^{I}\\in \\mathbb{R}^{d}\\to s^{I}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "content": " is the weight. Unlike CLIP-Adapter, VIformer transforms the original embeddings " + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "inline_equation", + "content": "z^{I}" + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "content": " by introducing self-attention layers in a learnable manner to extract and retain specific viewpoint-invariant key components " + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "inline_equation", + "content": "s^I" + }, + { + "bbox": [ + 130, + 321, + 482, + 489 + ], + "type": "text", + "content": ". Combining LoRA and VIformer modules, the forward process of image encoding can be represented as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 176, + 497, + 480, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 497, + 480, + 526 + ], + "spans": [ + { + "bbox": [ + 176, + 497, + 480, + 526 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {z} ^ {I} = \\alpha \\cdot f _ {\\boldsymbol {\\theta}} (z ^ {I}) + (1 - \\alpha) \\cdot z ^ {I} \\\\ = \\alpha \\cdot f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {W} _ {\\mathbf {v}} \\cdot I + \\mathbf {B} \\mathbf {A} \\cdot I\\right) + (1 - \\alpha) \\cdot \\left(\\mathbf {W} _ {\\mathbf {v}} \\cdot I + \\mathbf {B} \\mathbf {A} \\cdot I\\right), \\tag {10} \\\\ \\end{array}", + "image_path": "091bb8837b77808d66c4c71f81c19a82d4a9e404ebf336f46ab17cbd33462976.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "text", + "content": "where the constant value " + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "text", + "content": " denotes the residual ratio to balance achieving original performance and viewpoint invariance performance. Therefore, for Eq. (6), we now only need to update " + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 536, + 482, + 596 + ], + "type": "text", + "content": ", rather than the entire weights of VLP. To facilitate the understanding of the OVT training process, we provide the pseudocode for OVT as shown in Appendix D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "Our evaluation of Omniview-Tuning spans several downstream tasks, including zero-shot classification, image captioning, and vision question answering." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 179, + 481, + 259 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Configurations of OVT and zero-shot Top-1 accuracy (\\%) on ImageNet-1K with ImageNet-V+. The number in parentheses shows the performance change relative to the pre-trained weights. Through OVT training, each model maintains the performance on ImageNet-1K (IN-1K) while significantly improving the performance on ImageNet-V+ (IN-V+.), narrowing the performance gap." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 179, + 481, + 259 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 481, + 259 + ], + "type": "table", + "html": "
ModelPretrain WeightPretrain DataTotal #Param.Trainable #Param.Image SizeBatch Size#Iter.IN-1KIN-View+
OVT-OpenCLIP ViT-B/32OpenCLIP-ViT-B/32LAION (2B)151M6.6M22451235k67.8 (↑1.3)59.5 (↑22.4)
OVT-OpenCLIP ViT-B/16OpenCLIP-ViT-B/16LAION (2B)149M6.6M22451235k69.7 (↑2.1)61.7 (↑17.5)
OVT-OpenCLIP ViT-L/14OpenCLIP-ViT-L/14LAION (2B)428M11.8M22425620k77.3 (↑2.1)69.8 (↑16.6)
OVT-MetaCLIP ViT-B/32MetaCLIP-ViT-B/32Common crawl (2.5B)151M6.6M22451240k69.7 (↑2.1)54.8 (↑13.8)
OVT-MetaCLIP ViT-B/16MetaCLIP-ViT-B/16Common crawl (2.5B)149M6.6M22451240k73.8 (↑1.7)64.8 (↑15.2)
OVT-MetaCLIP ViT-L/14MetaCLIP-ViT-L/14Common crawl (2.5B)428M11.8M22425620k77.7 (↓1.4)75.4 (↑9.0)
OVT-BLIP ViT-B/16Salesforce-BLIP-ViT-B/16COCO et al. (129M)234M4.3M22425620k61.7 (↑8.8)54.8 (↑18.0)
", + "image_path": "cd6a3ea38e44fda31d8d47dc867ff4abf99e196c7be31ff90b75ba0d609facef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 355 + ], + "type": "text", + "content": "For zero-shot classification (Sec. 5.1), we conduct evaluations for CLIP [40] and BLIP [29] architectures. For image captioning and vision question answering (Sec. 5.2), we replace the visual encoders in Vision Large Language Models (VLLMs) with our fine-tuned versions. We adopt LLaVA-1.5 [32,33], and OpenFlamingo [3], the most advanced open-source VLLMs available. Additionally, we present the ablation study and convergence analysis of our approach in Sec. 5.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 371, + 350, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 371, + 350, + 382 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 350, + 382 + ], + "type": "text", + "content": "5.1 Evaluation of Zero-Shot Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "text", + "content": "Baselines. We adopt the official CLIP (OpenAI CLIP [40]) and the community open-source version (OpenCLIP [26]) as baselines. Additionally, we include the current state-of-the-art Eva02-CLIP [51] and MetaCLIP [58] as another set of baselines to compare CLIP trained with improved techniques and extensive training data. For BLIP, we use the official implementation [29] as the baseline. Settings. We train two series of CLIP models using our OVT framework and MVCap dataset, each series comprising three different visual encoder architectures (ViT-B/32, ViT-B/16, and ViT-L/14). OVT-OpenCLIP are fine-tuned on the original weights of OpenCLIP, while OVT-MetaCLIP are based on MetaCLIP. The fine-tuning settings are detailed in Tab. 2. We standardized the " + }, + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "text", + "content": ", and the outlier viewpoints number " + }, + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 130, + 390, + 481, + 582 + ], + "type": "text", + "content": " and set the LoRA rank at 8. Datasets and Metrics. We employ a various set of benchmarks for evaluation, including clean data distributions [14, 27], common 2D-OOD [21, 23, 42, 55, 61], and most importantly, viewpoint-OOD [7, 16, 46, 61]. For each benchmark, we report Top-1 and Top-5 accuracy and average accuracy across all benchmarks. The evaluations follow the standard prompting engineering of CLIP [40]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "type": "text", + "content": "Results and Discussions. Tab. 3 summarizes the performance of our OVT-trained VLP models against various VLP versions, including their accuracy across different benchmarks and average accuracy across clean, common-OOD, and Viewpoint-OOD domains. We can draw the following conclusions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "(1) OVT significantly enhances the models' invariance to Viewpoint-OOD samples. Across different VLP architectures and visual encoders, OVT-trained models perform best on almost all viewpoint-OOD benchmarks. On the average" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 136, + 481, + 350 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 459, + 126 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 459, + 126 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 459, + 126 + ], + "type": "text", + "content": "Table 3: Top-1/Top-5 zero-shot accuracy (%) under different benchmarks" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 136, + 481, + 350 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 481, + 350 + ], + "type": "table", + "html": "
ModelCleanCommon-ODViewpoint-OD
ImageNet-100 [66]ImageNet-IR [14]Cllr-100 [27]Avg. Acc.ImageNet-Vi [14]ImageNet-Sch. [53]ImageNet-ODD [25]ImageNet-Rem. [21]OOD-CV [61]Avg. Acc.ImageNet-View [68]ImageNet-View + [68]OOD-CV-Pose [63]MIBO [7]Avg. Top-1Total Avg. Acc.
A. Comparisons with VIT-B/32 baselines
OpenAI CLIP77.5/93.963.3/88.864.3/88.168.4/90.255.8/83.442.2/70.333.4/62.250.7/75.450.2/82.646.5/74.844.5/65.427.5/52.447.2/84.526.5/59.436.4/65.448.6/75.5
Open CLIP81.1/95.366.5/89.975.8/94.074.5/93.058.1/83.953.6/79.334.8/64.461.0/81.953.5/81.952.2/78.354.4/72.137.1/63.246.9/81.633.0/69.242.8/71.554.6/79.7
OVT-OpenCLIP80.9/95.667.8/90.865.0/89.371.2/91.958.0/84.245.8/73.442.8/75.050.3/71.451.7/79.549.7/76.761.9/81.259.5/85.652.8/82.535.4/80.152.4/82.456.0/82.4
MetaCLIP80.7/95.667.6/90.577.7/95.275.3/93.859.5/85.455.9/81.432.4/62.563.2/83.852.0/84.252.6/79.561.4/76.741.0/67.848.9/87.934.8/73.246.5/76.456.3/82.0
OVT-MetaCLIP80.7/95.669.7/92.071.8/93.074.0/93.560.6/85.847.8/75.843.5/73.849.0/70.850.1/80.150.2/77.264.0/79.254.8/80.455.1/84.835.6/77.052.4/80.356.9/82.3
B. Comparisons with VIT-B/16 baselines
OpenAI CLIP82.1/95.768.3/91.967.2/89.472.5/92.361.8/87.448.2/76.327.7/55.759.1/83.052.2/84.649.8/77.451.6/68.936.9/63.853.4/86.830.1/66.143.0/71.453.2/79.1
Open CLIP83.2/96.270.1/91.877.0/94.876.8/94.362.2/87.056.0/82.030.7/59.864.9/85.654.3/82.753.6/79.458.1/74.444.2/70.948.5/84.034.6/74.646.4/76.057.0/82.0
OVT-OpenCLIP83.9/97.071.9/93.169.0/90.774.9/93.664.0/88.650.5/77.936.8/68.957.0/77.256.3/84.552.9/79.465.4/80.761.7/85.856.9/87.442.4/84.956.6/84.759.6/84.7
EVA-CLIP85.3/96.574.6/94.287.5/98.082.5/96.367.0/89.857.6/82.321.3/47.369.6/87.553.1/83.153.7/78.061.8/76.644.3/69.453.9/87.432.9/73.248.2/76.659.1/82.1
MetaCLIP84.3/97.272.1/93.478.9/95.478.4/95.365.0/89.360.1/84.826.2/56.470.2/89.352.3/85.454.8/81.064.2/79.449.6/76.148.9/90.938.5/78.750.3/81.259.2/84.7
OVT-MetaCLIP83.4/97.473.8/94.173.9/93.677.0/95.065.9/89.453.6/81.036.2/68.859.0/79.651.6/83.853.2/80.169.7/84.064.8/87.355.2/87.839.2/82.957.2/85.560.5/85.6
C. Comparisons with VIT-L/14 baselines
OpenAI CLIP86.5/97.475.4/94.676.5/93.379.5/95.169.8/90.959.5/84.318.6/43.872.8/91.452.9/88.854.7/79.860.3/75.645.8/71.547.9/88.238.0/74.148.0/77.358.6/82.8
Open CLIP86.8/97.875.2/94.383.7/96.781.9/96.267.7/90.263.2/86.424.0/50.574.5/91.254.5/85.056.8/80.665.7/78.153.2/76.752.4/90.542.3/83.053.4/82.161.9/85.0
OVT-OpenCLIP89.0/97.877.3/95.379.2/95.381.8/96.169.6/91.561.9/86.027.5/55.471.3/88.756.4/87.057.3/81.772.2/86.669.8/89.757.3/94.150.0/89.362.3/89.965.1/88.1
EVA-CLIP88.5/97.979.6/96.090.6/98.686.3/97.572.8/92.768.0/89.116.3/40.082.8/95.754.7/87.458.9/81.071.5/82.361.1/81.754.4/94.539.6/86.156.6/86.165.0/86.8
MetaCLIP88.3/98.379.1/95.984.1/96.983.8/97.072.5/92.668.9/89.817.0/40.681.8/95.156.6/87.559.3/81.177.3/89.366.4/87.058.9/93.448.1/89.662.7/89.866.6/88.0
OVT-MetaCLIP88.8/97.577.7/95.984.0/96.983.5/96.870.8/92.264.4/87.920.8/47.077.0/92.756.3/89.357.8/81.879.3/90.675.4/93.057.0/94.446.4/93.864.5/92.966.5/89.3
D. Comparisons with BLIP VIT-L/B16 baselines
BLIP76.6/93.352.9/80.267.0/88.365.5/87.347.3/74.751.0/76.625.6/53.464.3/83.853.9/87.648.4/75.255.2/68.236.8/63.350.8/89.227.0/66.142.4/71.950.7/77.1
OVT-BLIP82.2/97.061.7/88.866.6/88.970.2/91.553.7/82.946.5/74.233.8/62.757.4/77.956.4/87.349.6/77.062.6/79.054.8/79.955.2/89.531.5/73.251.0/80.455.2/81.8
", + "image_path": "b2e138e934d2a949bc324e3c70bcfc264ae7adf9ce2a413c9460f3665eff4b27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": "accuracy of viewpoint-OOD datasets, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 shows improvements of " + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": " over OpenCLIP, respectively. OVT-BLIP demonstrated an average improvement of " + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": ". (2) While enhancing viewpoint invariance, OVT maintains performance on clean samples and 2D-OOD without significant performance trade-offs. For 2D-OOD benchmarks, OVT-OpenCLIP with ViT-B/32, ViT-B/16, and ViT-L/14 sacrifice only " + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 533 + ], + "type": "text", + "content": " accuracy. (3) Compared to earlier CLIP baselines, the recently developed MetaCLIP exhibits better zero-shot performance. Based on this, OVT further enhances its performance under viewpoint-OOD samples. Visualization. We showcase OVT-OpenCLIP and the original OpenCLIP prediction on several viewpoint-OOD samples. As illustrated in Fig. 4, OVT-CLIP successfully predicts the categories of images from various unusual viewpoints in all cases, whereas the original CLIP is prone to make incorrect predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 548, + 306, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 306, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 306, + 560 + ], + "type": "text", + "content": "5.2 Performance on Other Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Settings. As LLaVA and Openflamingo use the OpenAI CLIP (ViT-L/14) to encode vision inputs, we applied OVT to this model in this section and comparing with other OpenAI CLIP (ViT-L/14) versions in image captioning tasks. The training setup remains consistent with the OVT-OpenCLIP described in Tab. 2. Baselines. In addition to comparing with the original OpenAI CLIP version, we also select " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "TeCoA^4" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " [37] and " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "FARE^4" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " [48], robust CLIP models based on adversarial training, as baselines, which have been proven to possess good resistance to adversarial samples in image captioning task." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 481, + 269 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Image captioning performance under clean distribution samples and viewpoint-OOD samples from Real-world and Synthetic domains. We utilize the MP-Net [50] to calculate the similarity between generated descriptions and ground-truth labels, considering predictions successful if they exceed the similarity threshold " + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 481, + 269 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 481, + 269 + ], + "type": "table", + "html": "
ModelVisual EncoderReal-world DomainSynthetic Domain
OOD-CV (iid) [61]β@1.0β@0.5β@Adp.OOD-CV (Pose) [61]β@1.0β@0.5β@Adp.IM3D [46]β@1.0β@0.5β@Adp.ImageNet-V+ [46]β@1.0β@0.5β@Adp.
LLaVa-7bOpenAI CLIP(ViT-L/14)44.161.167.546.453.658.746.753.358.820.425.532.1
TeCoA4[37](ViT-L/14)41.958.965.536.141.649.226.330.142.68.711.622.6
FARE4[48](ViT-L/14)42.158.965.240.245.950.835.239.249.212.715.823.1
OVT-CLIP(ViT-L/14)43.559.565.946.553.659.149.454.061.826.431.941.0
LLaVa-13bOpenAI CLIP(ViT-L/14)45.468.070.648.658.660.848.756.760.821.228.432.5
TeCoA4[37](ViT-L/14)42.467.072.237.448.951.325.028.641.58.410.921.8
FARE4[48](ViT-L/14)43.966.771.141.952.154.836.141.448.612.115.920.8
OVT-CLIP(ViT-L/14)45.767.370.848.258.661.950.458.963.226.436.240.9
", + "image_path": "ec503e52c3f4437a948ad01c15cd1eb3fab20f684a30a729f9698efb2c8c010b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 284, + 167, + 312 + ], + "blocks": [ + { + "bbox": [ + 134, + 284, + 167, + 312 + ], + "lines": [ + { + "bbox": [ + 134, + 284, + 167, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 284, + 167, + 312 + ], + "type": "image", + "image_path": "966b2b7e73c38d4bdda01e628a71669dd308034e071e2454b3036d4c2bdc62e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 285, + 246, + 312 + ], + "blocks": [ + { + "bbox": [ + 168, + 285, + 246, + 312 + ], + "lines": [ + { + "bbox": [ + 168, + 285, + 246, + 312 + ], + "spans": [ + { + "bbox": [ + 168, + 285, + 246, + 312 + ], + "type": "image", + "image_path": "d6e47f114e0cbdcf420d52e2e3c71b928caa97b4f83d166f578b68c359a6600f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 247, + 285, + 281, + 312 + ], + "blocks": [ + { + "bbox": [ + 247, + 285, + 281, + 312 + ], + "lines": [ + { + "bbox": [ + 247, + 285, + 281, + 312 + ], + "spans": [ + { + "bbox": [ + 247, + 285, + 281, + 312 + ], + "type": "image", + "image_path": "469b1c5701e4df162717f0773a2816158e876dbec0c8babe70dd526650b27211.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 283, + 285, + 358, + 312 + ], + "blocks": [ + { + "bbox": [ + 283, + 285, + 358, + 312 + ], + "lines": [ + { + "bbox": [ + 283, + 285, + 358, + 312 + ], + "spans": [ + { + "bbox": [ + 283, + 285, + 358, + 312 + ], + "type": "image", + "image_path": "7e2383e2660a341ad04013475d3c73368e55a530b4d8866c0b193260ac4ed703.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 134, + 314, + 167, + 341 + ], + "blocks": [ + { + "bbox": [ + 134, + 314, + 167, + 341 + ], + "lines": [ + { + "bbox": [ + 134, + 314, + 167, + 341 + ], + "spans": [ + { + "bbox": [ + 134, + 314, + 167, + 341 + ], + "type": "image", + "image_path": "c506a10688a0963e44f542dea37037a56ecc7232abe0974915d44eefebcbaf64.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 168, + 314, + 246, + 341 + ], + "blocks": [ + { + "bbox": [ + 168, + 314, + 246, + 341 + ], + "lines": [ + { + "bbox": [ + 168, + 314, + 246, + 341 + ], + "spans": [ + { + "bbox": [ + 168, + 314, + 246, + 341 + ], + "type": "image", + "image_path": "85a4559c15e28c2b0274591aa8605fffbb3b51938ba907a5639068216059ff65.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 247, + 314, + 281, + 341 + ], + "blocks": [ + { + "bbox": [ + 247, + 314, + 281, + 341 + ], + "lines": [ + { + "bbox": [ + 247, + 314, + 281, + 341 + ], + "spans": [ + { + "bbox": [ + 247, + 314, + 281, + 341 + ], + "type": "image", + "image_path": "6a3500f5ad9bddd5d8d6004f5daa241989d16620e66758c9ef59d6a0a65621d1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 283, + 314, + 358, + 341 + ], + "blocks": [ + { + "bbox": [ + 283, + 314, + 358, + 341 + ], + "lines": [ + { + "bbox": [ + 283, + 314, + 358, + 341 + ], + "spans": [ + { + "bbox": [ + 283, + 314, + 358, + 341 + ], + "type": "image", + "image_path": "50f184330a8a2503329c2ffb537eebe88295d6837189dbc57cb23663fa823fe2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 342, + 168, + 371 + ], + "blocks": [ + { + "bbox": [ + 135, + 342, + 168, + 371 + ], + "lines": [ + { + "bbox": [ + 135, + 342, + 168, + 371 + ], + "spans": [ + { + "bbox": [ + 135, + 342, + 168, + 371 + ], + "type": "image", + "image_path": "328b2d77f496f5b4420c9f5f53bb5b478edaef7b62477e9d772192f08ee72d5e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 380, + 360, + 425 + ], + "lines": [ + { + "bbox": [ + 131, + 380, + 360, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 360, + 425 + ], + "type": "text", + "content": "Fig. 5: The image descriptions generated by LLaVa-13B using our OVT-CLIP and the original OpenAI CLIP as vision encoder, where red texts indicates incorrect category descriptions, and green texts represents correct." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 169, + 342, + 246, + 372 + ], + "blocks": [ + { + "bbox": [ + 169, + 342, + 246, + 372 + ], + "lines": [ + { + "bbox": [ + 169, + 342, + 246, + 372 + ], + "spans": [ + { + "bbox": [ + 169, + 342, + 246, + 372 + ], + "type": "image", + "image_path": "bb53e203ab2bfc77f45d51e3c11e732f2590994e2dc9e5ba9a5fe532b89b74a2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 253, + 342, + 276, + 371 + ], + "blocks": [ + { + "bbox": [ + 253, + 342, + 276, + 371 + ], + "lines": [ + { + "bbox": [ + 253, + 342, + 276, + 371 + ], + "spans": [ + { + "bbox": [ + 253, + 342, + 276, + 371 + ], + "type": "image", + "image_path": "565e0550f42ea728db8158e630a725072a6e02952c402d5984a985e90d9eade4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 283, + 342, + 358, + 372 + ], + "blocks": [ + { + "bbox": [ + 283, + 342, + 358, + 372 + ], + "lines": [ + { + "bbox": [ + 283, + 342, + 358, + 372 + ], + "spans": [ + { + "bbox": [ + 283, + 342, + 358, + 372 + ], + "type": "image", + "image_path": "1f9962b8863478dbe866f31889ade1c338cd7bc6bfe62ec33607f41560425b63.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 380, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 373, + 380, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 373, + 380, + 482, + 424 + ], + "type": "text", + "content": "Fig.6: The Top-1 accuracy of OVT-OpenCLIP (ViT-B/16) with the iterations increases." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 375, + 284, + 478, + 371 + ], + "blocks": [ + { + "bbox": [ + 375, + 284, + 478, + 371 + ], + "lines": [ + { + "bbox": [ + 375, + 284, + 478, + 371 + ], + "spans": [ + { + "bbox": [ + 375, + 284, + 478, + 371 + ], + "type": "image", + "image_path": "02873cab1d52a52a340c73d57b93b0f413bce70449cb1e6828f51b84cff74dea.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "content": "Datasets and Metrics. Given the absence of caption benchmarks that include viewpoint-changing OOD samples, we conduct evaluations using existing viewpoint-OOD datasets, including real-world datasets (using OOD-CV (iid) to represent clean distribution and OOD-CV (Pose) for viewpoint-OOD) and synthetic datasets (using IM3D [46] for clean distribution and ImageNet-V+ for viewpoint-OOD). We adopt word embedding distance to calculate the accuracy of the captioning task. By adopting MPNet [50], a state-of-the-art textual embedding model, we measure the similarity between keywords in the generated description and the ground-truth categories. Then assess the accuracy by counting the number of samples that exceed a specific similarity threshold " + }, + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Results and Discussions. Tab. 4 shows the image captioning accuracy of CLIP models under different training strategies, considering " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " at 1.0 (indicating predictions involve ground-truth categories), 0.5, and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "Adp" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". (meaning " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " is equal to the average similarity in the clean distribution). We found that OVT-CLIP improves the accuracy of descriptions generated by LLaVa for viewpoint-OOD samples while maintaining its performance on corresponding clean distributions. When used as the visual encoder for the LLaVa-7B model, OVT-CLIP achieved an " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " increase in accuracy compared to the original CLIP model weights under " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta = Adp" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". Besides, we find that although robust CLIP versions maintain perfor" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 481, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 5: Average Top-1/Top-5 zero-shot accuracy (\\%) under different data distributions within various ablation settings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 481, + 198 + ], + "type": "table", + "html": "
LITCVIFormerLVCTotal Avg.Clean Avg.Common-OOD Avg.Viewpoint-OOD Avg.
XXX61.985.081.996.256.880.653.482.1
✓XX61.985.7 (↑0.7)79.4 (↓2.5)95.6 (↓0.6)56.2 (↓0.6)81.4 (↑0.8)56.0 (↑2.6)83.8 (↑1.7)
✓✓X62.2 (↑0.3)86.2 (↑1.2)79.9 (↓2.0)95.4 (↓0.8)55.4 (↓1.4)81.6 (↑1.0)57.5 (↑4.1)85.2 (↑3.1)
✓✓✓65.1 (↑3.2)88.1 (↑3.1)81.8 (↓0.1)96.1 (↓0.1)57.3 (↑0.5)81.7 (↑1.1)62.3 (↑8.9)89.9 (↑7.8)
", + "image_path": "42043bd602183900b6ab184896c5e6da8f315e70c06c7bb769fa2988c8b4fdb1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 480, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 480, + 269 + ], + "type": "text", + "content": "mance on clean distribution samples, they experience a significant performance decline when facing viewpoint-OOD samples. We select some examples with the generated description in Fig. 5. For the visual question-answering task, we used OpenFlamingo as the VLLMs. The results are reported in the Appendix A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 288, + 364, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 288, + 364, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 364, + 300 + ], + "type": "text", + "content": "5.3 Ablation Studies and Additional Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": "Our ablation studies focus on the VIFormer and the " + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VC}" + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": " within the Omniview-Tuning framework. Tab. 5 shows the Top-1/Top-5 acc of OVT-OpenCLIP (ViT-L/14) across various data distributions and different ablation settings. Beyond the original OpenCLIP, we set a baseline that only uses " + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ITC}" + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": " for fine-tuning. Keeping other training settings fixed, reliance solely on " + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ITC}" + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": " led to a more significant performance decline in clean and 2D-OOD samples while achieving limited viewpoint OOD performance improvement (2.6%/1.7%). The integration of VI-Former led to further improvements in viewpoint OOD accuracy (4.1%/3.1%). With the further addition of " + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VC}" + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": ", the improvement in viewpoint OOD performance is most significant (8.9%/7.8%), and it also reduces performance sacrifices in other data distributions. More ablation analyses are available in Appendix B. Furthermore, we report OVT's training convergence, depicted in Fig. 6. We display the Top-1 accuracy evolution for OVT-OpenCLIP (ViT-B/16) across various training iterations. We observe that around 40K iterations, with a batch size of 512, are sufficient for effective convergence, thus achieving a balance in performance across different data distributions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 520, + 225, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 225, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 225, + 532 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "To tackle the challenge of 3D viewpoint in VLP models, this paper first introduced the MVCap dataset, a million-scale collection of image-text pairs with diverse viewpoint variations. Building upon this groundwork, we then proposed the Omniview-Tuning framework, which incorporates a novel Cross-Viewpoint Alignment objective in a parameter-efficient manner, effectively enhancing the VLP models' ability to generate viewpoint-invariant representations. Moreover, through extensive experiments, we successfully verified that Omniview-Tuning could bring significant improvements in viewpoint invariance while preserving the original performance. These advancements provide valuable insights and a standard for future research on viewpoint invariance in foundation models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "This work was supported by the Project of the National Natural Science Foundation of China (No. 62076018, 92370124, 62276149, 92248303), the Fundamental Research Funds for the Central Universities, and Tsinghua-Alibaba Joint Research Program. Y. Dong was also supported by the China National Postdoctoral Program for Innovative Talents and CCF-BaiChuan-Ebtech Foundation Model Fund." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 232, + 197, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 232, + 197, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 232, + 197, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 258, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 258, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 300 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 302, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 345 + ], + "type": "text", + "content": "2. Alcorn, M.A., Li, Q., Gong, Z., Wang, C., Mai, L., Ku, W.S., Nguyen, A.: Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4845-4854 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 346, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 389 + ], + "type": "text", + "content": "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 391, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 434 + ], + "type": "text", + "content": "4. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "type": "text", + "content": "5. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "text", + "content": "6. Calian, D.A., Stimberg, F., Wiles, O., Rebuffi, S.A., Gyorgy, A., Mann, T., Gowal, S.: Defending against image corruptions through adversarial augmentations. arXiv preprint arXiv:2104.01086 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "text", + "content": "7. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "8. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "9. Chuang, Y.S., Dangovski, R., Luo, H., Zhang, Y., Chang, S., Soljacic, M., Li, S.W., Yih, W.t., Kim, Y., Glass, J.: Diffcse: Difference-based contrastive learning for sentence embeddings. arXiv preprint arXiv:2204.10298 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 632 + ], + "type": "text", + "content": "10. Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21126-21136 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "12. Das, S., Ryoo, M.S.: ViewCLR: Learning self-supervised video representation for unseen viewpoints. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5573-5583 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "15. Dong, Y., Kang, C., Zhang, J., Zhu, Z., Wang, Y., Yang, X., Su, H., Wei, X., Zhu, J.: Benchmarking robustness of 3d object detection to common corruptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1022-1032 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "text", + "content": "16. Dong, Y., Ruan, S., Su, H., Kang, C., Wei, X., Zhu, J.: View fool: Evaluating the robustness of visual recognition to adversarial viewpoints. Advances in Neural Information Processing Systems 35, 36789-36803 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "text", + "content": "17. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: Proceedings of the 39th International Conference on Machine Learning. pp. 6216-6234 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "18. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "19. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "20. Hamdi, A., Ghanem, B.: Towards analyzing semantic robustness of deep neural networks. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 22-38. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 490 + ], + "type": "text", + "content": "21. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "22. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "23. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15262-15271 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "type": "text", + "content": "24. Ho, C.H., Leung, B., Sandstrom, E., Chang, Y., Vasconcelos, N.: Catastrophic child's play: Easy to perform, hard to defend adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9229-9237 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "25. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "26. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "text", + "content": "28. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "type": "text", + "content": "29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 226 + ], + "type": "text", + "content": "30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 258 + ], + "type": "text", + "content": "31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "type": "text", + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "type": "text", + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 303, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 481, + 335 + ], + "type": "text", + "content": "34. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H.: Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 335, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 481, + 369 + ], + "type": "text", + "content": "35. Madan, S., Henry, T., Dozier, J., Ho, H., Bhandari, N., Sasaki, T., Durand, F., Pfister, H., Boix, X.: When and how cnns generalize to out-of-distribution category-viewpoint combinations. arXiv preprint arXiv:2007.08032 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 402 + ], + "type": "text", + "content": "36. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (ICLR) (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 424 + ], + "type": "text", + "content": "37. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 456 + ], + "type": "text", + "content": "38. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 456, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 481, + 489 + ], + "type": "text", + "content": "39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 490, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 533 + ], + "type": "text", + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "type": "text", + "content": "41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: Proceedings of the 38th International Conference on Machine Learning. pp. 8821-8831 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 566, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 481, + 599 + ], + "type": "text", + "content": "42. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 599, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 481, + 643 + ], + "type": "text", + "content": "43. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10901-10911 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "type": "text", + "content": "44. Rong, X.: word2vec parameter learning explained. arXiv preprint arXiv:1411.2738 (2014)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "45. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "46. Ruan, S., Dong, Y., Su, H., Peng, J., Chen, N., Wei, X.: Towards viewpoint-invariant visual recognition via adversarial training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4709-4719 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 225 + ], + "type": "text", + "content": "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "type": "text", + "content": "48. Schlarmann, C., Singh, N.D., Croce, F., Hein, M.: Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 258, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 300 + ], + "type": "text", + "content": "49. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "50. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "type": "text", + "content": "51. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "52. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "53. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "54. Tu, W., Deng, W., Gedeon, T.: A closer look at the robustness of contrastive language-image pre-training (CLIP). In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "text", + "content": "55. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "56. Wang, X., He, K., Gupta, A.: Transitive invariance for self-supervised visual representation learning. In: Proceedings of the IEEE international conference on computer vision. pp. 1329-1338 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "57. Wu, Z., Wang, Z., Xu, X., Lu, J., Yan, H.: Embodied task planning with large language models. arXiv preprint arXiv:2307.01848 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "58. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "text", + "content": "59. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9150-9161 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "60. Zhang, Y., Huang, Y., Sun, Y., Liu, C., Zhao, Z., Fang, Z., Wang, Y., Chen, H., Yang, X., Wei, X., et al.: Benchmarking trustworthiness of multimodal large language models: A comprehensive study. arXiv preprint arXiv:2406.07057 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "61. Zhao, B., Yu, S., Ma, W., Yu, M., Mei, S., Wang, A., He, J., Yuille, A., Kortylewski, A.: Ood-cv: A benchmark for robustness to individual nuisances in real-world out-of-distribution shifts. In: ICML 2022 Shift Happens Workshop (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Ruan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 150 + ], + "type": "text", + "content": "62. Zhou, X., Liu, M., Zagar, B.L., Yurtsever, E., Knoll, A.C.: Vision language models in autonomous driving and intelligent transportation systems. arXiv preprint arXiv:2310.14414 (2023)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 102 + ], + "type": "text", + "content": "Omniview-Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_content_list.json b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..48a7ff5f644dbd23dc094c51b3044cfd0a40e958 --- /dev/null +++ b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_content_list.json @@ -0,0 +1,1821 @@ +[ + { + "type": "text", + "text": "On Calibration of Object Detectors: Pitfalls, Evaluation and Baselines", + "text_level": 1, + "bbox": [ + 246, + 140, + 754, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Selim Kuzucu\\*, Kemal Oksuz\\*, Jonathan Sadeghi, and Puneet K. Dokania", + "bbox": [ + 215, + 212, + 785, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Five AI Ltd., United Kingdom {selim.kuzucu2, kemal.oksuz, jonathan.sadeghi, puneet.dokania}@five.ai", + "bbox": [ + 228, + 239, + 771, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Reliable usage of object detectors require them to be calibrated—a crucial problem that requires careful attention. Recent approaches towards this involve (1) designing new loss functions to obtain calibrated detectors by training them from scratch, and (2) post-hoc Temperature Scaling (TS) that learns to scale the likelihood of a trained detector to output calibrated predictions. These approaches are then evaluated based on a combination of Detection Expected Calibration Error (D-ECE) and Average Precision. In this work, via extensive analysis and insights, we highlight that these recent evaluation frameworks, evaluation metrics, and the use of TS have notable drawbacks leading to incorrect conclusions. As a step towards fixing these issues, we propose a principled evaluation framework to jointly measure calibration and accuracy of object detectors. We also tailor efficient and easy-to-use post-hoc calibration approaches such as Platt Scaling and Isotonic Regression specifically for object detection task. Contrary to the common notion, our experiments show that once designed and evaluated properly, post-hoc calibrators, which are extremely cheap to build and use, are much more powerful and effective than the recent train-time calibration methods. To illustrate, D-DETR with our post-hoc Isotonic Regression calibrator outperforms the recent train-time state-of-the-art calibration method Cal-DETR by more than 7 D-ECE on the COCO dataset. Additionally, we propose improved versions of the recently proposed Localization-aware ECE and show the efficacy of our method on these metrics. Code is available at: https://github.com/fiveai/detection_calibration.", + "bbox": [ + 261, + 306, + 766, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Calibration $\\cdot$ Object Detection $\\cdot$ Performance Evaluation", + "bbox": [ + 261, + 651, + 728, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 691, + 375, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detectors have been widely-used in a variety of safety-critical applications related to, but not limited to, autonomous driving [6, 9, 13, 14, 61, 65] and medical imaging [22, 28, 29, 64]. In addition to being accurate, their confidence estimates should also allow characterization of their error behaviour to make them reliable. This feature, known as calibration, can enable a model to provide valuable information to subsequent systems playing crucial role in making", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. SK contributed during his internship at Five AI Oxford team.", + "bbox": [ + 236, + 824, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7e2b008d9e7a264bfc5bb9abac53da3f2feb9a4a639484f78ea7abe627f7b2a.jpg", + "image_caption": [ + "(a) D-ECE" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 357, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a2687f6620eb68b4dc6e79b3f122c1689f977dd67b2b7414500ac4fa3b20133.jpg", + "image_caption": [ + "(b) AP", + "Fig. 1: The performance of different detectors over operating confidence thresholds on COCO minitest. Orange: Faster R-CNN, Green: RS R-CNN, Purple: ATSS, Red: PAA, Blue: D-DETR. All measures are lower better except AP. It is not trivial to identify an operating threshold and compare detectors, especially when the common evaluation [30, 41–43, 52], combining D-ECE for calibration and AP for accuracy, is used. Instead, we use LaECE0 and Localisation-Recall-Precision Error (LRP)." + ], + "image_footnote": [], + "bbox": [ + 361, + 145, + 500, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9ae363959673ad87211e181932b1e66023918fb521ed7df0166e5834d0b8c275.jpg", + "image_caption": [ + "(c) LaECEo" + ], + "image_footnote": [], + "bbox": [ + 501, + 145, + 640, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0cd98c320f1287d7ad646c9f8c773df7be549c494a0ef3ea458a0063e82dbe54.jpg", + "image_caption": [ + "(d) LRP" + ], + "image_footnote": [], + "bbox": [ + 643, + 148, + 781, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "safety-critical decisions [4, 24, 26, 37, 39]. Despite its importance, calibration of detectors is a relatively underexplored area in the literature and requires significant attention. Therefore, in this work, we focus on different aspects of the evaluation framework that is now being adopted by most recent works building calibrated detectors and discuss their pitfalls and propose fixes. Additionally, we tailor the well-known post-hoc calibration methods to improve the calibration of a given object detector (trained) with minimal effort.", + "bbox": [ + 212, + 364, + 787, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Naturally, practitioners prefer detectors that perform well in terms of both accuracy and calibration, which we will refer to as joint performance. However, unlike classification, choosing the best performing model is non-trivial for object detection. This is because different detectors commonly yield detection sets with varying cardinalities for the same image, and this difference in population size is shown to affect the joint performance evaluation [48]. Furthermore, when object detectors are used in practice, an operating threshold is normally chosen [4, 24, 26, 31, 36, 37, 39], and the choice of this threshold directly influences a detector's performance. Thus, comparing the performance of a detector in terms of calibration or accuracy over different operating thresholds as well as with different detectors is not straightforward, as illustrated in Fig. 1.", + "bbox": [ + 212, + 474, + 787, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We assert that a framework for joint evaluation should follow certain basic principles. Firstly, the detectors should be evaluated on a thresholded set of detections to align with their practical usage. While doing so, the evaluation framework will require a principled model-dependent threshold selection mechanism, as the confidence distribution of each detector can differ significantly [47]. Secondly, the calibration should enforce the confidence scores to provide unambiguous and fine-grained information about the detection quality. For example, if the confidence score represents the localisation quality of a detection, this provides more fine-grained information than representing whether the object is detected or not. Thirdly, the datasets should be properly-designed for evaluation. That is, the training, validation (val.) and in-distribution (ID) test splits should be sampled from the same underlying distribution, and additionally, the domain-shifted test splits — which are crucial for safety-critical applications — should", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/382af17139916493efbe02eec5bd7dfd57a4aa38934a5d6c73e1c92af4de55ec.jpg", + "table_caption": [ + "Table 1: Principles of joint performance evaluation of object detectors in terms of accuracy and calibration, and whether existing evaluation approaches violate them." + ], + "table_footnote": [], + "table_body": "
Principles of Joint EvaluationD-ECE-style [30, 41–43, 52]LaECE-style [48]CE-style [55]Ours
Model-dependent threshold selectionX✓X✓
Unambiguous & fine-grained confidence scoresXXX✓
Properly-designed datasetsXXX✓
Properly-trained detectors & calibratorsX✓X✓
", + "bbox": [ + 215, + 184, + 785, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be included. Finally, baseline detectors and calibration methods must be trained properly, as otherwise the evaluation might provide misleading conclusions.", + "bbox": [ + 212, + 277, + 785, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are three approaches for jointly evaluating accuracy and calibration:", + "bbox": [ + 238, + 308, + 777, + 321 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\circ$ D-CEC-style [30,41-43,52] thresholds the detections commonly from a confidence of 0.30 to compute Detection Expected Calibration Error (D-CEC) and use top-100 detections from each image for Average Precision (AP),", + "$\\circ$ LaECE-style [48] enforces the detectors to be thresholded properly, and combine Localisation-aware Expected Calibration Error (LaECE) with LRP [47],", + "- CE-style [55] thresholds the detections from a confidence score of 0.50 to obtain Calibration Error (CE) and AP." + ], + "bbox": [ + 225, + 324, + 784, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As summarized in Tab. 1, these evaluations do not adhere to the basic principles mentioned above. To exemplify, D-CEE-style evaluation — the most common evaluation approach [30, 41–43, 52] — uses different operating thresholds for calibration and accuracy, which does not align well with the practical usage of detectors. Also, using a fixed threshold (i.e., 0.30) for all detectors artificially promotes certain detectors. To illustrate, while D-CEE-style evaluation ranks the green detector as the worst in Fig. 1(a), the green one yields the best D-CEE at 0.70. Besides, as shown in Fig. 1(b), AP is maximized at the confidence of 0 (leading to too many detections with low confidence) for all the detectors, and thus AP cannot be used to obtain a proper operating threshold [47, 48]. In terms of conveying fine-grained information, D-CEE aims to align confidence with the precision only, which effectively ignores the localisation quality of the detections, a crucial performance aspect of object detection. Finally, this type of evaluation also has limitations in terms of dataset splits and the chosen baselines as we explore in Sec. 3.", + "bbox": [ + 212, + 431, + 787, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Having proper baseline calibration methods is also essential to monitor the progress in the field. Recently proposed train-time calibration methods commonly employ an auxiliary loss term to regularize the confidence scores during training [30,41-43,52]. Such methods are shown to be effective against the Temperature Scaling (TS) [15], which is used as the only post-hoc calibration baseline. Post-hoc calibrators are obtained on a held-out val. set, and hence can easily be applied to any off-the-shelf detector. Despite their potential advantages, unlike for classification [15,20,23,38,56,62,69], post-hoc calibration methods have not been explored for object detection sufficiently [30,48].", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce a joint evaluation framework which respects the aforementioned principles (Tab. 1), and thus address the critical drawbacks of existing evaluation approaches. That is, we first define $\\mathrm{LaECE}_0$ and $\\mathrm{LaACE}_0$ ,", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as novel calibration errors, each of which aims to align the detection confidence scores with their localisation qualities. Thus, the detectors respecting $\\mathrm{LaECE}_0$ and $\\mathrm{LaACE}_0$ provide informative confidence estimates about their behaviours. We measure accuracy using LRP [45,47], which requires a proper combination of false-positive (FP), false-negative (FN) and localisation errors. Thereby requiring the detectors to be properly-thresholded as shown by the bell-like curves in Fig. 1(d). Also, we design three datasets with different characteristics, and introduce Platt Scaling (PS) as well as Isotonic Regression (IR) as highly effective post-hoc calibrators tailored to object detection. Our main contributions are:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We identify various quirks and assumptions in state-of-the-art (SOTA) methods in quantifying miscalibration of object detectors and show that they, if not treated properly, can provide misleading conclusions.", + "bbox": [ + 225, + 282, + 787, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a framework for joint evaluation consisting of properly-designed datasets, evaluation measures tailored to practical usage of object detectors and baseline post-hoc calibration methods. We show that our framework addresses the drawbacks of existing approaches.", + "bbox": [ + 227, + 327, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to the literature, we show that, if designed properly, post-hoc calibrators can significantly outperform the SOTA training time calibration methods. To illustrate, on the common COCO benchmark, D-DETR with our IR calibrator outperforms the SOTA Cal-DETR [43] significantly: (i) by more than 7 points in terms of D-ECE and (ii) $\\sim 4$ points in terms of our challenging LaECE $_0$ .", + "bbox": [ + 227, + 388, + 787, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Background and Notation", + "text_level": 1, + "bbox": [ + 215, + 501, + 503, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Detectors and Evaluating their Accuracy Denoting the set of $M$ objects in an image $X$ by $\\{b_i, c_i\\}^M$ where $b_i \\in \\mathbb{R}^4$ is a bounding box and $c_i \\in \\{1, \\ldots, K\\}$ is its class; an object detector produces the bounding box $\\hat{b}_i$ , the class label $\\hat{c}_i$ and the confidence score $\\hat{p}_i$ for the objects in $X$ , i.e., $f(X) = \\{\\hat{c}_i, \\hat{b}_i, \\hat{p}_i\\}^N$ with $N$ being the number of predictions. During evaluation, each detection is first labelled as a true-positive (TP) or a FP using a matching function $\\psi(\\cdot)$ relying on an Intersection-over-Union (IoU) threshold $\\tau$ to validate TPs. We assume $\\psi(i)$ returns the index of the object that a TP $i$ matches to; else $i$ is a FP and $\\psi(i) = -1$ . Then, AP [11, 16, 33], the common accuracy measure, corresponds to the area under the Precision Recall (PR) curve. Though widely-used, AP has been criticized recently from different aspects [5, 45, 47, 50, 58]. To illustrate, AP is maximized when the number of detections increases [48] as shown in Fig. 1(b). Therefore, AP does not help choosing an operating threshold, which is critical for practical deployment. As an alternative, LRP [45, 47] combines the numbers of TP, FP, FN with the localisation error of the detections, which are denoted by $\\mathrm{N_{TP}}$ , $\\mathrm{N_{FP}}$ , $\\mathrm{N_{FN}}$ and $\\mathcal{E}_{loc}(i) \\in [0, 1]$ respectively:", + "bbox": [ + 212, + 532, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L R P} = \\frac {1}{\\mathrm {N} _ {\\mathrm {F P}} + \\mathrm {N} _ {\\mathrm {F N}} + \\mathrm {N} _ {\\mathrm {T P}}} \\left(\\mathrm {N} _ {\\mathrm {F P}} + \\mathrm {N} _ {\\mathrm {F N}} + \\sum_ {\\psi (i) > 0} \\mathcal {E} _ {l o c} (i)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 789, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike AP, LRP requires the detection set to be thresholded properly as both FPs and FNs are penalized in Eq. (1).", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluating the Calibration of Object Detectors The alignment of accuracy and confidence of a model, termed calibration, is extensively studied for classification [8, 15, 27, 40, 44, 63]. That is, a classifier is calibrated if its accuracy is $p$ for the predictions with confidence of $p$ for all $p \\in [0,1]$ . For object detection, [30] extends this definition to enforce that the confidence matches the precision of the detector, $\\mathbb{P}(\\hat{c}_i = c_i|\\hat{p}_i) = \\hat{p}_i, \\forall \\hat{p}_i \\in [0,1]$ , where $\\mathbb{P}(\\hat{c}_i = c_i|\\hat{p}_i)$ is the precision. Then, discretizing the confidence space into $J$ bins, D-ECE is", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} - \\mathrm {E C E} = \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} \\right|}{\\left| \\hat {\\mathcal {D}} \\right|} \\left| \\bar {p} _ {j} - \\operatorname {p r e c i s i o n} (j) \\right|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 295, + 785, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{D}$ and $\\hat{D}_j$ are the set of all detections and the detections in the $j$ -th bin, and $\\bar{p}_j$ and precision $(j)$ are the average confidence and the precision of the detections in the $j$ -th bin. Alternatively, considering that object detection is a joint task of classification and localisation, LaECE [48] aims to match the confidence with the product of precision and average IoU of TPs. Also, to prevent certain classes from dominating the error, LaECE is introduced as a class-wise measure. Using superscript $c$ to refer to each class and $\\mathrm{IoU}^c (j)$ as the average IoU of $\\hat{D}_j^c$ , LaECE is defined as:", + "bbox": [ + 212, + 349, + 787, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L a E C E} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} ^ {c} \\right|}{\\left| \\hat {\\mathcal {D}} ^ {c} \\right|} \\left| \\bar {p} _ {j} ^ {c} - \\operatorname {p r e c i s i o n} ^ {c} (j) \\times \\mathrm {I o U} ^ {c} (j) \\right|. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 486, + 785, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Calibration Methods in Object Detection The existing methods for calibrating object detectors can be split into two groups: (1) Training-time calibration approaches [41-43,52,55] regularize the model to yield calibrated confidence scores during training, which is generally achieved by an additive auxiliary loss. (2) Post-hoc calibration methods use a held-out val. set to fit a calibration function that maps the predicted confidence to the calibrated confidence. Specifically, TS [15] is the only method considered as a baseline for recent training time methods [41,42,52,55]. As an alternative, IR [1-3,66] is used within a limited scope for a specific task called Self-aware Object Detection [48]. Furthermore, its effectiveness neither on a wide range of detectors nor against existing training-time calibration approaches has yet been investigated.", + "bbox": [ + 212, + 539, + 787, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Analysis of the Common D-ECE-style Evaluation", + "text_level": 1, + "bbox": [ + 215, + 729, + 728, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "D-ECE-style evaluation is the most common evaluation approach adopted by several methods [30,41-43,52]. For that reason, here we provide a comprehensive analysis of this evaluation approach and analyse the LaECE-style and CE-style evaluations in App. B. Our analyses below are based on the principles outlined in Sec. 1 and Tab. 1, and show that all approaches have notable drawbacks.", + "bbox": [ + 212, + 761, + 787, + 838 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Model-dependent threshold selection. As AP is obtained using the top-100 detections and D-ECE is computed on detections thresholded above 0.30, D-ECE-style evaluation uses two different detection sets. This inconsistency is not reflective of how detectors are used in practice. Also, we observe that a fixed threshold of 0.30 for evaluating the calibration induces a bias for certain detectors. To illustrate, we compare the performance of different calibration methods over different thresholds in Fig. 2, where CalDETR [43] performs the best only for the threshold 0.30 and the post-hoc TS significantly outperforms it on all other thresholds. Therefore, this method of evaluation is sensitive to the choice of threshold, leading to ambiguity on the best performing method.", + "bbox": [ + 212, + 146, + 540, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55857e30c4cb45b7f556d6325a5db863f4533c87ecfbd07c8a0e872e2356e10b.jpg", + "image_caption": [ + "Fig.2: Comparison of calibration methods in terms of D-ECE on COCO mini-test using D-DETR [71]. Post-hoc TS and IR calibrators are obtained on a subset of Objects365 [60] as in D-ECE-style evaluation." + ], + "image_footnote": [], + "bbox": [ + 553, + 148, + 784, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. Fine-grained confidence scores. Manipulating Eq. (2), we show in App. B that D-ECE for the $j$ -th bin can be expressed as,", + "bbox": [ + 215, + 419, + 784, + 450 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\sum_ {\\hat {b} _ {i} \\in \\hat {\\mathcal {D}} _ {j}, \\psi (i) > 0} (\\hat {p} _ {i} - 1) + \\sum_ {\\hat {b} _ {i} \\in \\hat {\\mathcal {D}} _ {j}, \\psi (i) = - 1} \\hat {p} _ {i} \\right|. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 460, + 785, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Eq. (4) implies that D-ECE is minimized when the confidence scores $\\hat{p}_i$ of TPs are 1 and those of FPs are 0, which is also how the prediction-target pairs are usually constructed to train post-hoc TS [30, 41-43, 52]. Even if the detector is perfectly calibrated for these binary targets, the confidence scores do not provide information about localisation quality as illustrated by binary-valued $\\hat{p}_{\\mathrm{D - ECE}}$ for both detections in Fig. 3(b). Also, Popordanoska et al. [55] utilise D-ECE in a COCO-style manner, that is they average D-ECE over different TP validation IoU thresholds similar to COCO-style AP [33]. However, we observe that this way of using D-ECE can promote ambiguous confidence scores. As an example, given two IoU thresholds $\\tau_{1}$ and $\\tau_{2}$ , a detection $\\hat{b}_i$ with $\\tau_{1} \\leq \\mathrm{IoU}(\\hat{b}_{i}, b_{\\psi(i)}) < \\tau_{2}$ is a TP for $\\tau_{1}$ but a FP for $\\tau_{2}$ . Thus, given Eq. (4), it follows that $\\hat{b}_i$ has contradictory confidence targets for $\\tau_{1}$ and $\\tau_{2}$ . This is illustrated in Fig. 3(d) in which D-ECE $_{\\mathrm{C}}$ (red line) remains constant regardless of the confidence. Thus, using D-ECE (or another calibration measure) in this way should be avoided.", + "bbox": [ + 212, + 518, + 787, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Properly-designed datasets. In the literature, the val. set to obtain the post-hoc calibrator is typically taken from a different dataset than the ID dataset [41-43, 52]. Specifically, the post-hoc calibrators are obtained on a subset from Objects365 [60] and BDD100K [65] for the models trained with COCO [33] and Cityscapes [9] respectively. Hence, as expected, a different dataset inevitably induces domain shift, affecting the performance of the post-hoc calibrator [51]. To show that, following existing approaches, we obtain an IR calibrator [48] on", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73f300fb1e05cc44a585f3248d032376a9711b3637f925edf6cd8be00ca2550a.jpg", + "image_caption": [ + "(a) Uncalibrated detections" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 405, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8889571d3ed27f1a9cbbb1a4416fa51054e95fe999997570e6adc0ece4628325.jpg", + "image_caption": [ + "(b) Calibrated detections" + ], + "image_footnote": [], + "bbox": [ + 406, + 143, + 594, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3da4bf1e3647babc0688f30aa0ac6043dd45637f257b9ba2643a5fdec3929738.jpg", + "image_caption": [ + "(c) Calibrated detections (Ours)" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 782, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e47267df99eec3bccc9b443e6ca1768bfe5710836689ea57d2cd96169249157e.jpg", + "image_caption": [ + "(d) TP for $\\tau = 0.5$" + ], + "image_footnote": [], + "bbox": [ + 225, + 261, + 406, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9475da39d774a2523d85c0b4448391780aacb3c9b874bf94e453437a3dc8ed3c.jpg", + "image_caption": [ + "(e) FP for $\\tau = 0.5$ , TP for $\\tau = 0$" + ], + "image_footnote": [], + "bbox": [ + 410, + 261, + 591, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd1ccdfbff0d84bf901eec177dc03d33597e7a65b87a5fb835f788a5b9102f29.jpg", + "image_caption": [ + "(f) FP for $\\tau = 0$", + "Fig. 3: A pictorial comparison of the different calibration errors. (a) Uncalibrated detections of D-DETR on an image from [65]. The detections on the left and right have IoUs of 0.74 and 0.48 with the objects. (b) Calibrated detections in terms of D-ECE and LaECE using $\\tau = 0.50$ , and D-ECEc, COCO-style D-ECE as in [55]. $\\mathrm{D - ECE_C} = ?$ as calibration error does not have a global minimum as shown in (d). (c) Calibrated detections in terms of LaECEo and LaACEo in which confidence matches IoU. (d-f) Calibration errors for different types of detections, for which LaACEo behave the same as LaECEo, hence excluded for clarity. App. B presents the details." + ], + "image_footnote": [], + "bbox": [ + 596, + 261, + 777, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Objects365 and compare it with the one obtained on the ID val. set in terms of D-ECE-style evaluation. Tab. 2 shows that the latter IR now outperforms (i) the former one by $\\sim 11$ D-ECE and (ii) SOTA Cal-DETR [43] by 7.4 D-ECE, showing the importance of dataset design for proper evaluation.", + "bbox": [ + 212, + 537, + 784, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Properly-trained detectors and calibrators. Though Cityscapes is commonly used in the literature [41-43,52], the models trained on this dataset follow COCO-style training. Specifically, D-DETR [71] was trained on Cityscapes for only 50 epochs though the training set of Cityscapes is $\\sim 40\\times$ smaller than that of COCO (3K vs. $\\sim 118\\mathrm{K}$ ). We now tailor the training of D-DETR for Cityscapes by (i) $4\\times$ longer training considering the smaller training set and (ii) increasing the training image scale considering the original resolution following [7,9]. We kept all other hyperparameters as they are for both Cal-DETR and D-DETR, and App. B presents the details. Tab. 3 shows that, once trained in this setting, D-DETR is more accurate and calibrated than Cal-DETR.", + "bbox": [ + 212, + 599, + 787, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 A Framework for Joint Evaluation of Object Detectors", + "text_level": 1, + "bbox": [ + 214, + 782, + 781, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now present our evaluation approach that respects to the principles in Sec. 1.", + "bbox": [ + 214, + 824, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f41c453fe2cd3a0716343d95f357b615423e7baa55bc4e9125a89e6c79031c2c.jpg", + "table_caption": [ + "Table 2: Effect of using domain-shifted val. set on IR calibrator. Results are reported on COCO-minitest. Val. set is N/A for uncalibrated D-DETR and training time calibration method Cal-DETR." + ], + "table_footnote": [], + "table_body": "
MethodVal setD-ECEAP ↑
D-DETRN/A12.844.1
Cal-DETRN/A8.744.4
IRObjects36514.244.1
IRCOCO1.3 (+7.4)44.1
", + "bbox": [ + 225, + 227, + 478, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab99d42d8ff835f52878471f2243e4889531baded620cc2569fdb688fc9edd24.jpg", + "table_caption": [ + "Table 3: COCO training settings are commonly adopted while training DDETR on Cityscapes. When trained with larger images and longer, DETR performs slightly better than Cal-DETR." + ], + "table_footnote": [], + "table_body": "
MethodTraining StyleD-ECEAP↑
D-DETR [43]COCO13.826.8
Cal-DETR [43]COCO8.428.4
Cal-DETRCityscapes4.034.9
D-DETRCityscapes2.936.1
", + "bbox": [ + 509, + 226, + 785, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Towards Fine-grained Calibrated Detection Confidence Scores", + "text_level": 1, + "bbox": [ + 215, + 325, + 772, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Calibration refers to the alignment of accuracy and confidence of a model. Therefore, for an object detector to be calibrated, its confidence should respect both classification and localisation accuracy. We discussed in Sec. 3 that D-ECE, as the common calibration measure, only considers the precision of a detector, thereby ignoring its localisation performance (Eq. (2)). LaECE [48], defined in Eq. (3) as an alternative to D-ECE, enforces the confidence scores to represent the product of precision and average IoU of TPs. Thus, LaECE considers IoUs of only TPs, and effectively ignores the localisation qualities of detections if their IoU is less than the TP validation threshold $\\tau > 0$ . We assert that this selection mechanism based on IoU unnecessarily limits the information conveyed by the confidence score. We illustrate this on the right car in Fig. 3(b) for which LaECE requires a target confidence of 0 ( $\\hat{p}_{\\mathrm{LaECE}} = 0$ ) as its IoU is less than $\\tau = 0.50$ . However, instead of conveying a 0 confidence and implying no detection, representing its IoU by $\\hat{p}_i$ provides additional information. Hence, we propose using $\\tau = 0$ , in which case the calibration criterion of LaECE reduces to,", + "bbox": [ + 212, + 349, + 787, + 575 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\hat {b} _ {i} \\in B _ {i} (\\hat {p} _ {i})} [ \\mathrm {I o U} (\\hat {b} _ {i}, b _ {\\psi (i)}) ] = \\hat {p} _ {i}, \\forall \\hat {p} _ {i} \\in [ 0, 1 ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 587, + 785, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where we define $\\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)}) = 0$ for FPs when $\\tau = 0$ , $B_i(\\hat{p}_i)$ is the set of boxes with the confidence of $\\hat{p}_i$ and $b_{\\psi(i)}$ is the ground-truth box that $\\hat{b}_i$ matches with. To derive the calibration error for Eq. (5), we follow LaECE by using $J = 25$ equally-spaced bins and averaging over class-wise errors and define,", + "bbox": [ + 212, + 617, + 787, + 681 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L a E C E} _ {0} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} ^ {c} \\right|}{\\left| \\hat {\\mathcal {D}} ^ {c} \\right|} \\left| \\bar {p} _ {j} ^ {c} - \\operatorname {I o} \\mathrm {U} ^ {c} (j) \\right|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 691, + 785, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{D}}^c$ and $\\hat{\\mathcal{D}}_j^c$ denote the set of all detections and those in $j$ th bin respectively, $\\bar{p}_j^c$ is the average confidence score and $\\mathrm{IoU}^c (j)$ is the average IoU of detections in the $j$ -th bin for class $c$ , and the subscript 0 refers to the chosen $\\tau$ which is 0. Furthermore, similar to the classification literature [40, 44], we define Localisation-aware Adaptive Calibration Error (LaACE) using an adaptive binning approach in which the number of detections in each bin is equal. In order", + "bbox": [ + 212, + 746, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d4b33f40c53aef58411a483d128cab4c338fd618eb219e0bb684b60d5e0f27d.jpg", + "table_caption": [ + "Table 4: Datasets for evaluating object detection and instance segmentation methods." + ], + "table_footnote": [], + "table_body": "
TypeTrain setVal setID test setDomain-shifted test set
Common ObjectsCOCO trainCOCO minivalCOCO minitestCOCO minitest-C, Obj45K
Autonomous DrivingCS trainCS minivalCS minitestCS minitest-C, Foggy-CS
Long-tailed ObjectsLVIS trainLVIS minivalLVIS minitestLVIS minitest-C
", + "bbox": [ + 228, + 170, + 769, + 226 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to capture the model behaviour precisely, we adopt the extreme case in which each bin has only one detection, resulting in an easy-to-interpret measure which corresponds to the mean absolute error between the confidence and the IoU,", + "bbox": [ + 212, + 251, + 782, + 297 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L a A C E} _ {0} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {i = 1} ^ {| \\hat {\\mathcal {D}} ^ {c} |} \\frac {1}{| \\hat {\\mathcal {D}} ^ {c} |} \\left| \\hat {p} _ {i} - \\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right) \\right|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 308, + 784, + 349 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As we show in App. C, $\\mathrm{LaECE}_0$ and $\\mathrm{LaACE}_0$ are both minimized when $\\hat{p}_i = \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})$ for all detections, which is also a necessary condition for $\\mathrm{LaACE}_0$ . Hence, as illustrated on the right car in Fig. 3(c) and (e), $\\mathrm{LaECE}_0$ and $\\mathrm{LaACE}_0$ requires conveying more fine-grained information compared to other measures.", + "bbox": [ + 214, + 358, + 782, + 419 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Model-dependent Thresholding for Proper Joint Evaluation", + "text_level": 1, + "bbox": [ + 214, + 439, + 751, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In practice, object detectors employ an operating threshold to preferably output only TPs with high recall. However, AP as the common performance measure does not enable cross-validating such a threshold as it is maximized when the recall is maximized despite a drop in precision [47, 48]. This can be observed in Fig. 1(b) where AP consistently decreases as the confidence threshold increases. Alternatively, LRP (Eq. 1) prefers detectors with high precision, recall and low localisation error as illustrated by the bell-like curves in Fig. 1(d). This is because, unlike AP, LRP severely penalises detectors with low recall or precision, making it a perfect fit for our framework. As a result, we consider $\\mathrm{LaECE}_0$ and $\\mathrm{LaACE}_0$ with LRP and require each model to be thresholded properly.", + "bbox": [ + 212, + 463, + 787, + 614 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Properly-designed Datasets", + "text_level": 1, + "bbox": [ + 214, + 635, + 488, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We curate three datasets summarized in Tab. 4: (i) COCO [33] including common daily objects; (ii) Cityscapes [9] with autonomous driving scenes; and (iii) LVIS [16], a challenging dataset focusing on the calibration of long-tailed detection. For each dataset, we ensure that train, val. and ID test sets are sampled from the same distribution, and include domain-shifted test sets. As these datasets do not have public labels for test sets, we randomly split their val. sets into two as minival and minitest similar to [17,48,49]. In such a way, we provide ID val. sets to enable obtaining post-hoc calibrators and the operating thresholds properly. For domain-shifted test sets, we apply common corruptions [21] to the ID test sets, and include Obj45K [48,60] and Foggy Cityscapes [59] as more realistic shifts. Our datasets also have mask annotations and hence they can be used to evaluate instance segmentation methods. App. C includes further details.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Baseline Post-hoc Calibrators Tailored to Object Detection", + "text_level": 1, + "bbox": [ + 215, + 146, + 750, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "It is essential to develop post-hoc calibration methods tailored to object detection, which has certain differences from the classification task. Existing methods [41-43,52,55] use only TS as a baseline without considering the peculiarities of detection. Specifically, a single temperature parameter $T$ is learned to adjust the predictive distribution while the confidence $\\hat{p}_i$ is commonly assumed to be a Bernoulli random variable [30]. However, PS, fitting both a scale and a shift parameter, is the widely-accepted calibration approach when the underlying distribution is Bernoulli [15,54]. Also, how to construct a useful subset of the detections to train the post-hoc calibrators has not been explored. To address these shortcomings, we present (i) Platt Scaling in which the bias term makes a notable difference in the performance, and (ii) Isotonic Regression by modeling the calibration as a regression problem. Before introducing them, we now present an overview on how we determine the set of detections to train the calibrators.", + "bbox": [ + 212, + 167, + 787, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overview We obtain post-hoc calibrators on a held out val. set using the detections that are similar to those seen at inference to prevent low-scoring detections from dominating the training of the calibrator. To do so, we crossvalidate a calibration threshold $\\bar{u}^c$ for each class $c$ and train a class-specific calibrator $\\zeta^c: [0,1] \\to [0,1]$ using the detections with higher scores than $\\bar{u}^c$ . Still, as $\\zeta^c(\\cdot)$ changes the confidence scores, we need another threshold $\\bar{v}^c$ , as the operating threshold, to remove the redundant detections after calibration. Following the accuracy measure, we cross-validate $\\bar{u}^c$ and $\\bar{v}^c$ using LRP. As for inference time, for the $i$ -th detection $(\\hat{p}_i, \\hat{b}_i, \\hat{c}_i)$ , if $\\hat{p}_i \\geq \\bar{u}^{\\hat{c}_i}$ , it survives to the calibrator and then $\\hat{p}_i^{cal} = \\zeta^{\\hat{c}_i}(\\hat{p}_i)$ . Finally, if $\\hat{p}_i^{cal} \\geq \\bar{v}^{\\hat{c}_i}$ , the $i$ -th detection is an output of the detector. Alg. A.1 and A.2 provide the details. For $\\zeta^c(\\cdot)$ , we prefer monotonically increasing functions in order not to affect the ranking of the detections significantly and to keep their accuracy as we detail below.", + "bbox": [ + 212, + 364, + 787, + 560 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Distribution Calibration via Platt Scaling Assuming that $\\hat{p}_i$ is sampled from Bernoulli distribution $\\mathcal{B}(\\cdot)$ , we aim to minimize the Negative Log-Likelihood (NLL) of the predictions on the target distribution $\\mathcal{B}(\\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)}))$ using PS [54]. Accordingly, we recover the logits, and then shift and scale the logits to obtain the calibrated probabilities $\\hat{p}_i^{cal}$ ,", + "bbox": [ + 214, + 561, + 787, + 638 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {i} ^ {\\text {c a l}} = \\sigma \\left(a \\sigma^ {- 1} \\left(\\hat {p} _ {i}\\right) + b\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 646, + 784, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\sigma(\\cdot)$ is the sigmoid and $\\sigma^{-1}(\\cdot)$ is its inverse, as well as $a \\geq 0$ and $b$ are the learnable parameters. We derive the NLL for the $i$ th detection in App. C as", + "bbox": [ + 214, + 670, + 785, + 700 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n- \\left(\\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right) \\log \\left(\\hat {p} _ {i} ^ {c a l}\\right) + \\left(1 - \\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right)\\right) \\log \\left(1 - \\hat {p} _ {i} ^ {c a l}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 289, + 708, + 784, + 724 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Please note that Eq. (9), which is in fact the cross-entropy loss, is minimized if $\\hat{p}_i^{cal} = \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})$ when $\\mathrm{LaECE}_0$ and $\\mathrm{LaACE}_0$ are minimized. We optimize Eq. (9) via the second-order optimization strategy L-BFGS [35] following [30].", + "bbox": [ + 214, + 732, + 785, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Confidence Calibration via Isotonic Regression As an alternative perspective, $\\hat{p}_i$ can also be directly calibrated by modelling the calibration as a regression task. To do so, we construct the prediction-target pairs $(\\{\\hat{p}_i, \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})\\})$ on the held-out val. set and then fit an IR model using scikit-learn [53].", + "bbox": [ + 212, + 780, + 802, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a6caf504bb9c345e4131867507e1fc47f2f601eaad25d091169716240d3edc84.jpg", + "table_caption": [ + "Table 5: Comparison with SOTA methods in terms of other evaluation measures on COCO [33]. LRP is reported on LRP-optimal thresholds obtained on val. set. AP is reported on top-100 detections. $\\tau$ is taken as 0.50. All measures are lower-better, except AP. Bold: the best, underlined: second best. PS: Platt Scaling, IR: Isotonic Regression." + ], + "table_footnote": [], + "table_body": "
Cal.\nTypeMethodCalibration (thr. 0.30)Calibration (LRP thr.)Accuracy
D-ECELaECED-ECELaECELRPAP↑
Uncal.D-DETR [71]12.813.215.012.166.344.1
Training\nTimeMbLS [34]15.616.318.715.865.944.3
MDCA [19]12.213.514.312.666.443.8
TCD [42]12.413.114.412.366.644.0
BPC [41]9.813.111.412.366.843.6
Cal-DETR [43]8.712.99.711.866.044.4
Post-hoc\n(Ours)PS for D-ECE0.9(+7.8)16.32.4(+7.3)15.866.344.1
PS for LaECE11.011.5(+1.4)9.410.1(+1.7)66.344.1
IR for D-ECE1.3(+7.4)15.72.6(+7.1)15.366.244.1
IR for LaECE10.28.9(+4.0)9.38.2(+3.6)66.343.7
", + "bbox": [ + 228, + 212, + 779, + 359 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adapting Our Approach to Different Calibration Objectives Until now, we considered post-hoc calibrators for $\\mathrm{LaECE_0}$ and $\\mathrm{LaACE_0}$ though in practice different measures can be preferred. Our post-hoc calibrators can easily be adapted for such cases by considering the dataset design and optimisation criterion. To illustrate, for D-ECE-style evaluation, the calibration dataset is to be class-agnostic where the detections are thresholded from 0.30 with prediction-target pairs for IR as $(\\{\\hat{p}_i,0\\})$ and $(\\{\\hat{p}_i,1\\})$ for FPs and TPs respectively.", + "bbox": [ + 212, + 388, + 787, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experimental Evaluation", + "text_level": 1, + "bbox": [ + 215, + 518, + 493, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We now show that our post-hoc calibration approaches consistently outperform training time calibration methods by significant margins (Sec. 5.1) and that they generalize to any detector and can thus be used as a strong baseline (Sec. 5.2).", + "bbox": [ + 212, + 550, + 787, + 597 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Comparing Our Baselines with SOTA Calibration Methods", + "text_level": 1, + "bbox": [ + 212, + 618, + 750, + 633 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Here, we compare PS and IR with recent training-time calibration methods considering various evaluation approaches. As these training-time methods mostly rely on D-DETR, we also use D-DETR with ResNet-50 [18]. We obtain the detectors of training time approaches trained with COCO dataset from their official repositories, whereas we incorporate Cityscapes into their official repositories and train them using the recommended setting in Tab. 3.", + "bbox": [ + 212, + 643, + 787, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on Other Evaluation Approaches Before moving on to our evaluation approach, we first show that our PS and IR outperform all existing training time methods on existing evaluation approaches. For that, we consider D-ECE and the LaECE from $\\tau = 0.5$ by including two different evaluation settings for each: (i) the detection set is obtained from the fixed threshold of 0.30 following the convention [30,41-43], and (ii) the operating thresholds are cross-validated using LRP. Following their standard usage, we use 10 and 25", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/727b8852eb2b5186fd2e01914b37d1292e6934f15d9b7661729664bea9e6cf66.jpg", + "table_caption": [ + "Table 6: Comparison with SOTA calibration methods on Common Objects using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration." + ], + "table_footnote": [], + "table_body": "
Calibration TypeMethodCOCO minitest (ID)COCO-C (Domain Shift)Obj45K (Domain Shift)
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
UncalibratedD-DETR [71]12.727.157.314.628.771.516.435.872.0
Training-timeMbLS [34]16.530.356.816.831.171.817.337.171.6
MDCA [19]13.127.257.514.528.771.816.635.672.2
TCD [42]13.026.757.614.628.371.916.335.571.7
BPC [41]12.425.557.714.127.172.117.334.572.0
Cal-DETR [43]11.624.656.213.826.470.618.835.371.1
Post-hoc (Ours)Platt Scaling9.623.557.312.825.671.517.033.772.0
Isotonic Regression7.723.157.210.725.371.517.233.372.0
(+3.9)(+1.5)(+3.1)(+1.1)(-0.9)(+1.2)
", + "bbox": [ + 235, + 199, + 772, + 347 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f8fc842e44c50a16585600aefd8dd232053601a4c0842c9a45cdd371665d7fad.jpg", + "table_caption": [ + "Table 7: Comparison with SOTA on Autonomous Driving using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration." + ], + "table_footnote": [], + "table_body": "
Calibration TypeMethodCityscapes minitest (ID) LaECE0|LaACE0|LRPCityscapes-C (Domain Shift) LaECE0|LaACE0|LRPFoggy Cityscapes (Domain Shift) LaECE0|LaACE0|LRP
UncalibratedD-DETR [71]20.326.057.221.425.680.218.522.369.4
Training-timeTCD [42]16.831.759.223.232.481.624.433.871.6
BPC [41]23.831.864.928.133.383.724.730.973.8
Cal-DETR [43]21.325.356.923.026.480.820.023.271.0
Post-hoc (Ours)Platt Scaling9.623.357.217.726.280.211.321.669.4
Isotonic Regression9.0(+7.8)23.7(+1.6)56.816.4(+5.0)25.8(-0.2)80.510.0(+8.5)21.2(+1.1)69.5
", + "bbox": [ + 235, + 425, + 772, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "bins to compute D-ECE and LaECE respectively. We optimize PS and IR by considering the calibration objective as described in Sec. 4.4. Tab. 5 shows that PS and IR outperform SOTA Cal-DETR significantly by more than 7 D-ECE and up to 4 LaECE on COCO minitest. Please note that all previous approaches are optimized for $D$ -ECE thresholded from 0.30, in terms of which our PS yields only 0.9 $D$ -ECE improving the SOTA by 7.8. Tab. 5 also suggests that post-hoc calibrators perform the best when the calibration objective is aligned with the measure. App. D shows that our observations generalize to Cityscapes.", + "bbox": [ + 212, + 592, + 787, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Common Objects Setting We now evaluate detectors using our evaluation approach. Tab. 6 shows that IR and PS share the top-2 entries on almost all test subsets by preserving the accuracy (LRP) of D-DETR. Specifically, our gains on ID set and COCO-C are significant, where IR outperforms Cal-DETR by around $3 - 4\\mathrm{LaECE}_0$ and $1.0 - 1.5\\mathrm{LaACE}_0$ . As for Obj45K, the challenging test set with natural shift, IR and PS improve $\\mathrm{LaACE}_0$ but perform slightly worse in terms of $\\mathrm{LaECE}_0$ . This is an expected drawback of post-hoc approaches when the domain shift is large as they are trained only with ID val. set [51].", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3ab2636fc33bdeeae40195539f60483527ee5150b2d102a71e1d2f924109141d.jpg", + "table_caption": [ + "Table 8: Comparison with TS using D-DETR. Bold: the best, underlined: second best. X : domain-shifted val. set is used to obtain thresholds and calibrators, decreasing the accuracy (red font). Bias term only exists for PS ( $b$ in Eq. (8)), thus $\\mathrm{N}/\\mathrm{A}$ for IR." + ], + "table_footnote": [], + "table_body": "
MethodAblations on DatasetAblations on ModelCOCO minitestCityscapes minitest
ID Val. SetThresholdClass-wiseBias TermLaECE0LaACE0LRPLaECE0LaACE0LRP
Temperature Scaling (Current Baseline)✘12.320.861.521.025.960.3
✓12.523.157.320.926.357.2
Ablations on Temperature Scaling✓✓11.324.857.313.325.557.2
✓✓12.422.957.323.127.557.2
✓✓✓10.624.257.312.724.657.2
Platt Scaling (Ours) Isotonic Regression (Ours)✓✓✓✓9.623.557.39.623.357.2
✓✓✓N/A7.723.157.29.023.756.8
", + "bbox": [ + 223, + 199, + 782, + 310 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/38a715e43f9018cc99cd9903891029e0a88b8d3f4816091d92ef0f456f7486f5.jpg", + "table_caption": [ + "Table 9: Calibrating and evaluating different object detectors. We use Common Objects setting and report the results on COCO minitest. * denotes the detectors in Fig. 1. Bold: the best, underlined: second best for calibration." + ], + "table_footnote": [], + "table_body": "
TypeDetectorBackboneUncalibratedPlatt ScalingIsotonic RegressionAP ↑
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
One-StagePAA [25]*R5015.928.159.79.724.359.77.723.859.743.2
ATSS [70]*R5019.134.059.510.324.759.58.524.159.543.1
GFL [32]R5013.728.559.310.324.559.38.324.059.343.0
VFNet [68]R5013.925.857.710.725.157.78.324.657.744.8
Two-StageFaster R-CNN [57]*R5027.029.960.410.423.860.48.623.560.440.1
RS R-CNN [46]*R5019.728.958.710.223.558.78.123.058.842.4
DETR-likeD-DETR [71]*R5012.727.157.39.623.557.37.723.157.244.1
UP-DETR [10]R5034.435.255.810.022.655.88.222.255.942.9
DINO [67]R5013.626.953.610.623.553.68.922.853.650.4
OVODGLIP [31]Swin-T13.025.349.09.222.449.07.721.849.055.7
G. DINO [36]Swin-T13.827.546.98.921.946.97.721.347.058.3
SOTACo-DETR [72]Swin-L10.823.041.58.620.241.56.719.341.664.5
EVA [12]ViT(EVA)17.421.241.28.620.341.27.119.941.264.5
MoCaE [49]N/A10.621.440.78.920.440.77.319.940.765.0
", + "bbox": [ + 225, + 378, + 782, + 568 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Autonomous Driving Setting Tab. 7 shows that our approaches consistently outperform all training time calibration approaches on this setting as well. Specifically, our gains are very significant ranging between 5.0-8.5 $\\mathrm{LaECE}_0$ compared to the SOTA Cal-DETR, further presenting the efficacy of our approaches.", + "bbox": [ + 212, + 598, + 784, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with Existing Temperature Scaling Baseline and Ablations Tab. 8 compares TS for different design choices as well as with our PS and IR. Please note that $\\pmb{x}$ corresponds to the baseline setting used in the recent approaches [41-43,52] that employ Objects365 [60] and BDD100K [60] as domain-shifted val. sets for obtaining the calibrator. Due to this domain shift, the accuracy of TS degrades by up to 4 LRP, in red font, as the operating thresholds obtained on these val. sets do not generalize to the ID set; showing that it is crucial to use an ID val set. In ablations, thresholding the detections and class-wise calibrators generally improves the performance of TS and a more notable gain is observed once the bias term is used in PS. Our PS outperforms TS baseline obtained on ID val. set by $\\sim 3$ LaECE $_0$ on COCO and 11.4 LaECE $_0$ on Cityscapes. Finally, IR performs on par or better compared to PS.", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Calibrating and Evaluating Different Detection Methods", + "text_level": 1, + "bbox": [ + 215, + 146, + 727, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Another benefit of our post-hoc calibrators is that they generalize to any object detector, thus they can be reliably used as baselines. To show that, we calibrate 14 different detectors with a diverse set of architectures using PS and IR in Tab. 9. The results suggest that both IR and PS perform better than uncalibrated detectors. IR consistently outperforms PS as it fits multiple piece-wise linear functions while PS learns only two parameters (Eq. (8)). Specifically, IR", + "bbox": [ + 212, + 169, + 468, + 381 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1f243872d76195e0be22a46a33633ca48ce9818721b25c4d71b2626689a7532b.jpg", + "image_caption": [ + "(a) Uncalibrated", + "Fig. 4: The reliability diagrams of UP-DETR." + ], + "image_footnote": [], + "bbox": [ + 488, + 202, + 627, + 308 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ce0488e4f241f90660917f7120d337b9bdd7e2f4e0a03927de10cc76a0fa2a30.jpg", + "image_caption": [ + "(b) Calibrated by IR" + ], + "image_footnote": [], + "bbox": [ + 638, + 202, + 774, + 308 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "decreases the range of $\\mathrm{LaECE_0}$ from 10.6-34.4 to 6.7-8.9 on COCO minitest by preserving the accuracy, making it a solid baseline. For further insights, Fig. 4 provides the reliability diagrams of the overconfident UP-DETR, of which IR significantly improves its calibration. As for SOTA, MoCaE performs the best in terms of accuracy with 40.7 LRP while Co-DETR has the best calibration with $6.7\\mathrm{LaECE}_0$ and $19.3\\mathrm{LaACE}_0$ , which the future work should aim to surpass. App. D.3 includes our results on our Long-tailed Objects setting (including object detection and instance segmentation), showing the effectiveness of our post-hoc calibrators on this challenging setting as well.", + "bbox": [ + 212, + 381, + 785, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The performance measures we use in our evaluation framework are also easily interpretable. For accuracy, LRP is a weighted combination of its FP, FN and localisation error components (App. A), which, as an example, are 10.1, 22.7 and 18.8 respectively for Co-DETR [72] calibrated with IR. Also considering $\\mathrm{LaACE}_0 = 19.3$ , one can easily infer that: once deployed with the operating thresholds determined by our framework, Co-DETR finds $77.3\\%$ of the objects with $89.9\\%$ precision and $81.2\\%$ IoU where $19.3\\%$ is the mean absolute error of the confidence to represent IoU. We believe these intuitive measures will enable practitioners to make better decision when deploying object detectors.", + "bbox": [ + 212, + 517, + 787, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 674, + 369, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The progress in a field heavily relies on the evaluation tools and the baselines used. In this paper, we showed that existing evaluation tools for calibration as well as the baseline post-hoc calibrators for object detectors have significant drawbacks. We remedied that by introducing an evaluation framework including baseline post-hoc calibrators tailored to object detection. Our experiments suggested that, once evaluated and designed properly, the post-hoc calibrators significantly outperform all existing training-time calibrators. This implies the need for research to develop better calibration techniques for object detection, for which, we believe, our evaluation framework will be an essential pillar.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ayer, M., Brunk, H.D., Ewing, G.M., Reid, W.T., Silverman, E.: An empirical distribution function for sampling with incomplete information. The annals of mathematical statistics pp. 641-647 (1955)", + "2. Barlow, R.E., Brunk, H.D.: The isotonic regression problem and its dual. Journal of the American Statistical Association 67(337), 140-147 (1972)", + "3. Best, M.J., Chakravarti, N.: Active set algorithms for isotonic regression; a unifying framework. Mathematical Programming 47(1), 425-439 (1990)", + "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2016). https://doi.org/10.1109/icip.2016.7533003, http://dx.doi.org/10.1109/ICIP.2016.7533003", + "5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: The IEEE European Conference on Computer Vision (ECCV) (2020)", + "6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv 1906.07155 (2019)", + "8. Cheng, J., Vasconcelos, N.: Calibrating deep neural networks by pairwise constraints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "0. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514", + "1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) 88(2), 303-338 (2010)", + "2. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)", + "4. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (Nov 2019). https://doi.org/10.1002/rob.21918, http://dx.doi.org/10.1002/rob.21918" + ], + "bbox": [ + 225, + 179, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1321-1330. PMLR (2017)", + "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "17. Harakeh, A., Waslander, S.L.: Estimating and evaluating regression predictive uncertainty in deep object detectors. In: International Conference on Learning Representations (ICLR) (2021)", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "19. Hebbalaguppe, R., Prakash, J., Madan, N., Arora, C.: A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16081-16090 (June 2022)", + "20. Hekler, A., Brinker, T.J., Buettner, F.: Test time augmentation meets post-hoc calibration: Uncertainty quantification under real-world conditions. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14856-14864 (Jun 2023). https://doi.org/10.1609/aaai.v37i12.26735, https://ojs.aaai.org/index.php/AAAI/article/view/26735", + "21. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (ICLR) (2019)", + "22. Jin, C., Udupa, J.K., Zhao, L., Tong, Y., Odhner, D., Pednekar, G., Nag, S., Lewis, S., Poole, N., Mannikeri, S., Govindasamy, S., Singh, A., Camaratta, J., Owens, S., Torigian, D.A.: Object recognition in medical images via anatomy-guided deep learning. Medical Image Analysis 81, 102527 (2022). https://doi.org/https://doi.org/10.1016/j.media.2022.102527, https://www.sciencedirect.com/science/article/pii/S1361841522001748", + "23. Joy, T., Pinto, F., Lim, S.N., Torr, P.H., Dokania, P.K.: Sample-dependent adaptive temperature scaling for improved calibration. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14919-14926 (Jun 2023). https://doi.org/10. 1609/aaii.v37i12.26742, https://ojs.aaii.org/index.php/AAAI/article/view/26742", + "24. Karimi, D., Dou, H., Warfield, S.K., Gholipour, A.: Deep learning with noisy labels: Exploring techniques and remedies in medical image analysis. Medical Image Analysis 65, 101759 (2020). https://doi.org/https://doi.org/10.1016/j.media.2020.101759, https://www.sciencedirect.com/science/article/pii/S1361841520301237", + "25. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020)", + "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)", + "27. Kumar, A., Liang, P.S., Ma, T.: Verified uncertainty calibration. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Kumar, N., Verma, R., Anand, D., Zhou, Y., Onder, O.F., Tsougenis, E., Chen, H., Heng, P.A., Li, J., Hu, Z., Wang, Y., Koohbanani, N.A., Jahanifar, M., Tajeddin, N.Z., Gooya, A., Rajpoot, N., Ren, X., Zhou, S., Wang, Q., Shen, D., Yang, C.K., Weng, C.H., Yu, W.H., Yeh, C.Y., Yang, S., Xu, S., Yeung, P.H., Sun, P., Mahbod, A., Schaefer, G., Ellinger, I., Ecker, R., Smedby, O., Wang, C., Chidester, B., Ton, T.V., Tran, M.T., Ma, J., Do, M.N., Graham, S., Vu, Q.D., Kwak, J.T., Gunda, A., Chanduri, R., Hu, C., Zhou, X., Lotfi, D., Safdari, R., Kascenas, A., O'Neil, A., Eschweiler, D., Stegmaier, J., Cui, Y., Yin, B., Chen, K., Tian, X., Gruening, P., Barth, E., Arbel, E., Remer, I., Ben-Dor, A., Sirazitdinova, E., Kohl, M., Braunewell, S., Li, Y., Xie, X., Shen, L., Ma, J., Baksi, K.D., Khan, M.A., Choo, J., Colomer, A., Naranjo, V., Pei, L., Iftekharuddin, K.M., Roy, K., Bhattacharjee, D., Pedraza, A., Bueno, M.G., Devanathan, S., Radhakrishnan, S., Koduganty, P., Wu, Z., Cai, G., Liu, X., Wang, Y., Sethi, A.: A multi-organ nucleus segmentation challenge. IEEE Transactions on Medical Imaging 39(5), 1380–1391 (2020). https://doi.org/10.1109/TMI.2019.2947628", + "29. Kumar, N., Verma, R., Sharma, S., Bhargava, S., Vahadane, A., Sethi, A.: A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging 36(7), 1550-1560 (2017). https://doi.org/10.1109/TMI.2017.2677499", + "30. Kuppers, F., Kronenberger, J., Shantia, A., Haselhoff, A.: Multivariate confidence calibration for object detection. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020)", + "31. Li, L.H., Zhang, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "32. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014)", + "34. Liu, B., Ayed, I.B., Galdran, A., Dolz, J.: The devil is in the margin: Margin-based label smoothing for network calibration. In: CVPR (2022)", + "35. Liu, D.C., Nocedal, J.: On the limited memory bfgs method for large scale optimization. Math. Program. 45(1-3), 503-528 (1989), http://dblp.uni-trier.de/db/journals/mp/mp45.htm1#LiuN89", + "36. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "37. Lu, Y., Lu, C., Tang, C.K.: Online video object detection using association LSTM. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 2363-2371 (2017). https://doi.org/10.1109/ICCV.2017.257", + "38. Ma, X., Blaschko, M.B.: Meta-cal: Well-controlled post-hoc calibration by ranking. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 7235-7245. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/ma21a.html", + "39. Mehrtash, A., Wells, W.M., Tempany, C.M., Abolmaesumi, P., Kapur, T.: Confidence calibration and predictive uncertainty estimation for deep medical image" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation. IEEE Transactions on Medical Imaging 39(12), 3868-3878 (Dec 2020). https://doi.org/10.1109/tmi.2020.3006437, http://dx.doi.org/10.1109/TMI.2020.3006437", + "40. Mukhoti, J., Kulharia, V., Sanyal, A., Golodetz, S., Torr, P., Dokania, P.: Calibrating deep neural networks using focal loss. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 15288-15299. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/aeb7b30ef1d024a76f21a1d40e30c302-Paper.pdf", + "41. Munir, M.A., Khan, M.H., Khan, S., Khan, F.S.: Bridging precision and confidence: A train-time loss for calibrating object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11474-11483 (June 2023)", + "42. Munir, M.A., Khan, M.H., Sarfraz, M., Ali, M.: Towards improving calibration in object detection under domain shift. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 38706-38718. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/fcd812a51b8f8d05cfea22e3c9c4b369-Paper-Conference.pdf", + "43. Munir, M.A., Khan, S., Khan, M.H., Ali, M., Khan, F.: Cal-DETR: Calibrated detection transformer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=4SkPTD6XNP", + "44. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)", + "45. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Localization recall precision (LRP): A new performance metric for object detection. In: The European Conference on Computer Vision (ECCV) (2018)", + "46. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: The International Conference on Computer Vision (ICCV) (2021)", + "47. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: One metric to measure them all: Localisation recall precision (lrp) for evaluating visual detection tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-1 (2021)", + "48. Oksuz, K., Joy, T., Dokania, P.K.: Towards building self-aware object detectors via reliable uncertainty quantification and calibration. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "49. Oksuz, K., Kuzucu, S., Joy, T., Dokania, P.K.: Mocae: Mixture of calibrated experts significantly improves object detection. arXiv preprint arXiv:2309.14976 (2023)", + "50. Otani, M., Togashi, R., Nakashima, Y., Rahtu, E., Heikkilä, J., Satoh, S.: Optimal correction cost for object detection evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21107-21115 (2022)", + "51. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In: Advances in Neural Information Processing Systems. vol. 32 (2019)", + "52. Pathiraja, B., Gunawardhana, M., Khan, M.H.: Multiclass confidence and localization calibration for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)", + "54. Platt, J.: Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Adv. Large Margin Classif. 10 (06 2000)", + "55. Popordanoska, T., Tiulpin, A., Blaschko, M.B.: Beyond classification: Definition and density-based estimation of calibration in object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 585-594 (January 2024)", + "56. Rahimi, A., Mensink, T., Gupta, K., Ajanthan, T., Sminchisescu, C., Hartley, R.: Post-hoc calibration of neural networks by g-layers. arXiv preprint arXiv:2006.12807 (2020)", + "57. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017)", + "58. Rezatofighi, H., Nguyen, T.T.D., Vo, B., Vo, B., Savarese, S., Reid, I.D.: How trustworthy are the existing performance evaluations for basic vision tasks? arXiv e-prints:2008.03533 (2020)", + "59. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision 126(9), 973-992 (Sep 2018), https://doi.org/10.1007/s11263-018-1072-8", + "60. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2019)", + "61. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)", + "62. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11809-11820. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/61f3a6dbc9120ea78ef75544826c814e-Paper.pdf", + "63. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "64. Yan, K., Wang, X., Lu, L., Summers, R.M.: Deepleson: Automated deep mining, categorization and detection of significant radiology image findings using large-scale clinical lesion annotations. arXiv preprint arXiv:1710.01766 (2017)", + "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "66. Zadrozny, B., Elkan, C.: Transforming classifier scores into accurate multiclass probability estimates. In: Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 694-699 (2002)", + "67. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "On Calibration of Object Detectors", + "bbox": [ + 493, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "68. Zhang, H., Wang, Y., Dayoub, F., Sünderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "69. Zhang, J., Yao, W., Chen, X., Feng, L.: Transferable post-hoc calibration on pretrained transformers in noisy text classification. Proceedings of the AAAI Conference on Artificial Intelligence 37(11), 13940-13948 (Jun 2023). https://doi.org/10.1609/aaai.v37i11.26632, https://ojs.aai.org/index.php/AAAI/article/view/26632", + "70. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "71. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable $\\{\\mathrm{detr}\\}$ : Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR) (2021)", + "72. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 369 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Kuzucu et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_model.json b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f5950cbb32984d5b0da38db2d0e8e624ae15ef88 --- /dev/null +++ b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_model.json @@ -0,0 +1,2759 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.141, + 0.756, + 0.185 + ], + "angle": 0, + "content": "On Calibration of Object Detectors: Pitfalls, Evaluation and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.213, + 0.786, + 0.228 + ], + "angle": 0, + "content": "Selim Kuzucu\\*, Kemal Oksuz\\*, Jonathan Sadeghi, and Puneet K. Dokania" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.24, + 0.772, + 0.269 + ], + "angle": 0, + "content": "Five AI Ltd., United Kingdom {selim.kuzucu2, kemal.oksuz, jonathan.sadeghi, puneet.dokania}@five.ai" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.307, + 0.767, + 0.64 + ], + "angle": 0, + "content": "Abstract. Reliable usage of object detectors require them to be calibrated—a crucial problem that requires careful attention. Recent approaches towards this involve (1) designing new loss functions to obtain calibrated detectors by training them from scratch, and (2) post-hoc Temperature Scaling (TS) that learns to scale the likelihood of a trained detector to output calibrated predictions. These approaches are then evaluated based on a combination of Detection Expected Calibration Error (D-ECE) and Average Precision. In this work, via extensive analysis and insights, we highlight that these recent evaluation frameworks, evaluation metrics, and the use of TS have notable drawbacks leading to incorrect conclusions. As a step towards fixing these issues, we propose a principled evaluation framework to jointly measure calibration and accuracy of object detectors. We also tailor efficient and easy-to-use post-hoc calibration approaches such as Platt Scaling and Isotonic Regression specifically for object detection task. Contrary to the common notion, our experiments show that once designed and evaluated properly, post-hoc calibrators, which are extremely cheap to build and use, are much more powerful and effective than the recent train-time calibration methods. To illustrate, D-DETR with our post-hoc Isotonic Regression calibrator outperforms the recent train-time state-of-the-art calibration method Cal-DETR by more than 7 D-ECE on the COCO dataset. Additionally, we propose improved versions of the recently proposed Localization-aware ECE and show the efficacy of our method on these metrics. Code is available at: https://github.com/fiveai/detection_calibration." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.652, + 0.729, + 0.667 + ], + "angle": 0, + "content": "Keywords: Calibration \\(\\cdot\\) Object Detection \\(\\cdot\\) Performance Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.693, + 0.377, + 0.709 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Object detectors have been widely-used in a variety of safety-critical applications related to, but not limited to, autonomous driving [6, 9, 13, 14, 61, 65] and medical imaging [22, 28, 29, 64]. In addition to being accurate, their confidence estimates should also allow characterization of their error behaviour to make them reliable. This feature, known as calibration, can enable a model to provide valuable information to subsequent systems playing crucial role in making" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.786, + 0.84 + ], + "angle": 0, + "content": "*Equal contributions. SK contributed during his internship at Five AI Oxford team." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.359, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.223, + 0.326, + 0.235 + ], + "angle": 0, + "content": "(a) D-ECE" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.146, + 0.501, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.223, + 0.454, + 0.235 + ], + "angle": 0, + "content": "(b) AP" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.146, + 0.641, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.223, + 0.609, + 0.235 + ], + "angle": 0, + "content": "(c) LaECEo" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.149, + 0.782, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.223, + 0.739, + 0.235 + ], + "angle": 0, + "content": "(d) LRP" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.246, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Fig. 1: The performance of different detectors over operating confidence thresholds on COCO minitest. Orange: Faster R-CNN, Green: RS R-CNN, Purple: ATSS, Red: PAA, Blue: D-DETR. All measures are lower better except AP. It is not trivial to identify an operating threshold and compare detectors, especially when the common evaluation [30, 41–43, 52], combining D-ECE for calibration and AP for accuracy, is used. Instead, we use LaECE0 and Localisation-Recall-Precision Error (LRP)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.789, + 0.472 + ], + "angle": 0, + "content": "safety-critical decisions [4, 24, 26, 37, 39]. Despite its importance, calibration of detectors is a relatively underexplored area in the literature and requires significant attention. Therefore, in this work, we focus on different aspects of the evaluation framework that is now being adopted by most recent works building calibrated detectors and discuss their pitfalls and propose fixes. Additionally, we tailor the well-known post-hoc calibration methods to improve the calibration of a given object detector (trained) with minimal effort." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.642 + ], + "angle": 0, + "content": "Naturally, practitioners prefer detectors that perform well in terms of both accuracy and calibration, which we will refer to as joint performance. However, unlike classification, choosing the best performing model is non-trivial for object detection. This is because different detectors commonly yield detection sets with varying cardinalities for the same image, and this difference in population size is shown to affect the joint performance evaluation [48]. Furthermore, when object detectors are used in practice, an operating threshold is normally chosen [4, 24, 26, 31, 36, 37, 39], and the choice of this threshold directly influences a detector's performance. Thus, comparing the performance of a detector in terms of calibration or accuracy over different operating thresholds as well as with different detectors is not straightforward, as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We assert that a framework for joint evaluation should follow certain basic principles. Firstly, the detectors should be evaluated on a thresholded set of detections to align with their practical usage. While doing so, the evaluation framework will require a principled model-dependent threshold selection mechanism, as the confidence distribution of each detector can differ significantly [47]. Secondly, the calibration should enforce the confidence scores to provide unambiguous and fine-grained information about the detection quality. For example, if the confidence score represents the localisation quality of a detection, this provides more fine-grained information than representing whether the object is detected or not. Thirdly, the datasets should be properly-designed for evaluation. That is, the training, validation (val.) and in-distribution (ID) test splits should be sampled from the same underlying distribution, and additionally, the domain-shifted test splits — which are crucial for safety-critical applications — should" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Principles of joint performance evaluation of object detectors in terms of accuracy and calibration, and whether existing evaluation approaches violate them." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.787, + 0.247 + ], + "angle": 0, + "content": "
Principles of Joint EvaluationD-ECE-style [30, 41–43, 52]LaECE-style [48]CE-style [55]Ours
Model-dependent threshold selectionX✓X✓
Unambiguous & fine-grained confidence scoresXXX✓
Properly-designed datasetsXXX✓
Properly-trained detectors & calibratorsX✓X✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.786, + 0.308 + ], + "angle": 0, + "content": "be included. Finally, baseline detectors and calibration methods must be trained properly, as otherwise the evaluation might provide misleading conclusions." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.309, + 0.778, + 0.323 + ], + "angle": 0, + "content": "There are three approaches for jointly evaluating accuracy and calibration:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.325, + 0.784, + 0.369 + ], + "angle": 0, + "content": "\\(\\circ\\) D-CEC-style [30,41-43,52] thresholds the detections commonly from a confidence of 0.30 to compute Detection Expected Calibration Error (D-CEC) and use top-100 detections from each image for Average Precision (AP)," + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.37, + 0.785, + 0.4 + ], + "angle": 0, + "content": "\\(\\circ\\) LaECE-style [48] enforces the detectors to be thresholded properly, and combine Localisation-aware Expected Calibration Error (LaECE) with LRP [47]," + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.4, + 0.785, + 0.43 + ], + "angle": 0, + "content": "- CE-style [55] thresholds the detections from a confidence score of 0.50 to obtain Calibration Error (CE) and AP." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.325, + 0.785, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.656 + ], + "angle": 0, + "content": "As summarized in Tab. 1, these evaluations do not adhere to the basic principles mentioned above. To exemplify, D-CEE-style evaluation — the most common evaluation approach [30, 41–43, 52] — uses different operating thresholds for calibration and accuracy, which does not align well with the practical usage of detectors. Also, using a fixed threshold (i.e., 0.30) for all detectors artificially promotes certain detectors. To illustrate, while D-CEE-style evaluation ranks the green detector as the worst in Fig. 1(a), the green one yields the best D-CEE at 0.70. Besides, as shown in Fig. 1(b), AP is maximized at the confidence of 0 (leading to too many detections with low confidence) for all the detectors, and thus AP cannot be used to obtain a proper operating threshold [47, 48]. In terms of conveying fine-grained information, D-CEE aims to align confidence with the precision only, which effectively ignores the localisation quality of the detections, a crucial performance aspect of object detection. Finally, this type of evaluation also has limitations in terms of dataset splits and the chosen baselines as we explore in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Having proper baseline calibration methods is also essential to monitor the progress in the field. Recently proposed train-time calibration methods commonly employ an auxiliary loss term to regularize the confidence scores during training [30,41-43,52]. Such methods are shown to be effective against the Temperature Scaling (TS) [15], which is used as the only post-hoc calibration baseline. Post-hoc calibrators are obtained on a held-out val. set, and hence can easily be applied to any off-the-shelf detector. Despite their potential advantages, unlike for classification [15,20,23,38,56,62,69], post-hoc calibration methods have not been explored for object detection sufficiently [30,48]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce a joint evaluation framework which respects the aforementioned principles (Tab. 1), and thus address the critical drawbacks of existing evaluation approaches. That is, we first define \\(\\mathrm{LaECE}_0\\) and \\(\\mathrm{LaACE}_0\\)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "as novel calibration errors, each of which aims to align the detection confidence scores with their localisation qualities. Thus, the detectors respecting \\(\\mathrm{LaECE}_0\\) and \\(\\mathrm{LaACE}_0\\) provide informative confidence estimates about their behaviours. We measure accuracy using LRP [45,47], which requires a proper combination of false-positive (FP), false-negative (FN) and localisation errors. Thereby requiring the detectors to be properly-thresholded as shown by the bell-like curves in Fig. 1(d). Also, we design three datasets with different characteristics, and introduce Platt Scaling (PS) as well as Isotonic Regression (IR) as highly effective post-hoc calibrators tailored to object detection. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.283, + 0.788, + 0.328 + ], + "angle": 0, + "content": "We identify various quirks and assumptions in state-of-the-art (SOTA) methods in quantifying miscalibration of object detectors and show that they, if not treated properly, can provide misleading conclusions." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.328, + 0.788, + 0.389 + ], + "angle": 0, + "content": "We introduce a framework for joint evaluation consisting of properly-designed datasets, evaluation measures tailored to practical usage of object detectors and baseline post-hoc calibration methods. We show that our framework addresses the drawbacks of existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.389, + 0.788, + 0.48 + ], + "angle": 0, + "content": "In contrast to the literature, we show that, if designed properly, post-hoc calibrators can significantly outperform the SOTA training time calibration methods. To illustrate, on the common COCO benchmark, D-DETR with our IR calibrator outperforms the SOTA Cal-DETR [43] significantly: (i) by more than 7 points in terms of D-ECE and (ii) \\(\\sim 4\\) points in terms of our challenging LaECE\\(_0\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.502, + 0.504, + 0.52 + ], + "angle": 0, + "content": "2 Background and Notation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Object Detectors and Evaluating their Accuracy Denoting the set of \\(M\\) objects in an image \\(X\\) by \\(\\{b_i, c_i\\}^M\\) where \\(b_i \\in \\mathbb{R}^4\\) is a bounding box and \\(c_i \\in \\{1, \\ldots, K\\}\\) is its class; an object detector produces the bounding box \\(\\hat{b}_i\\), the class label \\(\\hat{c}_i\\) and the confidence score \\(\\hat{p}_i\\) for the objects in \\(X\\), i.e., \\(f(X) = \\{\\hat{c}_i, \\hat{b}_i, \\hat{p}_i\\}^N\\) with \\(N\\) being the number of predictions. During evaluation, each detection is first labelled as a true-positive (TP) or a FP using a matching function \\(\\psi(\\cdot)\\) relying on an Intersection-over-Union (IoU) threshold \\(\\tau\\) to validate TPs. We assume \\(\\psi(i)\\) returns the index of the object that a TP \\(i\\) matches to; else \\(i\\) is a FP and \\(\\psi(i) = -1\\). Then, AP [11, 16, 33], the common accuracy measure, corresponds to the area under the Precision Recall (PR) curve. Though widely-used, AP has been criticized recently from different aspects [5, 45, 47, 50, 58]. To illustrate, AP is maximized when the number of detections increases [48] as shown in Fig. 1(b). Therefore, AP does not help choosing an operating threshold, which is critical for practical deployment. As an alternative, LRP [45, 47] combines the numbers of TP, FP, FN with the localisation error of the detections, which are denoted by \\(\\mathrm{N_{TP}}\\), \\(\\mathrm{N_{FP}}\\), \\(\\mathrm{N_{FN}}\\) and \\(\\mathcal{E}_{loc}(i) \\in [0, 1]\\) respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.79, + 0.787, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L R P} = \\frac {1}{\\mathrm {N} _ {\\mathrm {F P}} + \\mathrm {N} _ {\\mathrm {F N}} + \\mathrm {N} _ {\\mathrm {T P}}} \\left(\\mathrm {N} _ {\\mathrm {F P}} + \\mathrm {N} _ {\\mathrm {F N}} + \\sum_ {\\psi (i) > 0} \\mathcal {E} _ {l o c} (i)\\right). \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Unlike AP, LRP requires the detection set to be thresholded properly as both FPs and FNs are penalized in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Evaluating the Calibration of Object Detectors The alignment of accuracy and confidence of a model, termed calibration, is extensively studied for classification [8, 15, 27, 40, 44, 63]. That is, a classifier is calibrated if its accuracy is \\( p \\) for the predictions with confidence of \\( p \\) for all \\( p \\in [0,1] \\). For object detection, [30] extends this definition to enforce that the confidence matches the precision of the detector, \\( \\mathbb{P}(\\hat{c}_i = c_i|\\hat{p}_i) = \\hat{p}_i, \\forall \\hat{p}_i \\in [0,1] \\), where \\( \\mathbb{P}(\\hat{c}_i = c_i|\\hat{p}_i) \\) is the precision. Then, discretizing the confidence space into \\( J \\) bins, D-ECE is" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.296, + 0.786, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} - \\mathrm {E C E} = \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} \\right|}{\\left| \\hat {\\mathcal {D}} \\right|} \\left| \\bar {p} _ {j} - \\operatorname {p r e c i s i o n} (j) \\right|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.475 + ], + "angle": 0, + "content": "where \\(\\hat{D}\\) and \\(\\hat{D}_j\\) are the set of all detections and the detections in the \\(j\\)-th bin, and \\(\\bar{p}_j\\) and precision \\((j)\\) are the average confidence and the precision of the detections in the \\(j\\)-th bin. Alternatively, considering that object detection is a joint task of classification and localisation, LaECE [48] aims to match the confidence with the product of precision and average IoU of TPs. Also, to prevent certain classes from dominating the error, LaECE is introduced as a class-wise measure. Using superscript \\(c\\) to refer to each class and \\(\\mathrm{IoU}^c (j)\\) as the average IoU of \\(\\hat{D}_j^c\\), LaECE is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.487, + 0.786, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L a E C E} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} ^ {c} \\right|}{\\left| \\hat {\\mathcal {D}} ^ {c} \\right|} \\left| \\bar {p} _ {j} ^ {c} - \\operatorname {p r e c i s i o n} ^ {c} (j) \\times \\mathrm {I o U} ^ {c} (j) \\right|. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.708 + ], + "angle": 0, + "content": "Calibration Methods in Object Detection The existing methods for calibrating object detectors can be split into two groups: (1) Training-time calibration approaches [41-43,52,55] regularize the model to yield calibrated confidence scores during training, which is generally achieved by an additive auxiliary loss. (2) Post-hoc calibration methods use a held-out val. set to fit a calibration function that maps the predicted confidence to the calibrated confidence. Specifically, TS [15] is the only method considered as a baseline for recent training time methods [41,42,52,55]. As an alternative, IR [1-3,66] is used within a limited scope for a specific task called Self-aware Object Detection [48]. Furthermore, its effectiveness neither on a wide range of detectors nor against existing training-time calibration approaches has yet been investigated." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.73, + 0.73, + 0.748 + ], + "angle": 0, + "content": "3 Analysis of the Common D-ECE-style Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.762, + 0.788, + 0.839 + ], + "angle": 0, + "content": "D-ECE-style evaluation is the most common evaluation approach adopted by several methods [30,41-43,52]. For that reason, here we provide a comprehensive analysis of this evaluation approach and analyse the LaECE-style and CE-style evaluations in App. B. Our analyses below are based on the principles outlined in Sec. 1 and Tab. 1, and show that all approaches have notable drawbacks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.541, + 0.419 + ], + "angle": 0, + "content": "1. Model-dependent threshold selection. As AP is obtained using the top-100 detections and D-ECE is computed on detections thresholded above 0.30, D-ECE-style evaluation uses two different detection sets. This inconsistency is not reflective of how detectors are used in practice. Also, we observe that a fixed threshold of 0.30 for evaluating the calibration induces a bias for certain detectors. To illustrate, we compare the performance of different calibration methods over different thresholds in Fig. 2, where CalDETR [43] performs the best only for the threshold 0.30 and the post-hoc TS significantly outperforms it on all other thresholds. Therefore, this method of evaluation is sensitive to the choice of threshold, leading to ambiguity on the best performing method." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.15, + 0.785, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.284, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig.2: Comparison of calibration methods in terms of D-ECE on COCO mini-test using D-DETR [71]. Post-hoc TS and IR calibrators are obtained on a subset of Objects365 [60] as in D-ECE-style evaluation." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.42, + 0.785, + 0.451 + ], + "angle": 0, + "content": "2. Fine-grained confidence scores. Manipulating Eq. (2), we show in App. B that D-ECE for the \\(j\\)-th bin can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.462, + 0.786, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\left| \\sum_ {\\hat {b} _ {i} \\in \\hat {\\mathcal {D}} _ {j}, \\psi (i) > 0} (\\hat {p} _ {i} - 1) + \\sum_ {\\hat {b} _ {i} \\in \\hat {\\mathcal {D}} _ {j}, \\psi (i) = - 1} \\hat {p} _ {i} \\right|. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Eq. (4) implies that D-ECE is minimized when the confidence scores \\(\\hat{p}_i\\) of TPs are 1 and those of FPs are 0, which is also how the prediction-target pairs are usually constructed to train post-hoc TS [30, 41-43, 52]. Even if the detector is perfectly calibrated for these binary targets, the confidence scores do not provide information about localisation quality as illustrated by binary-valued \\(\\hat{p}_{\\mathrm{D - ECE}}\\) for both detections in Fig. 3(b). Also, Popordanoska et al. [55] utilise D-ECE in a COCO-style manner, that is they average D-ECE over different TP validation IoU thresholds similar to COCO-style AP [33]. However, we observe that this way of using D-ECE can promote ambiguous confidence scores. As an example, given two IoU thresholds \\(\\tau_{1}\\) and \\(\\tau_{2}\\), a detection \\(\\hat{b}_i\\) with \\(\\tau_{1} \\leq \\mathrm{IoU}(\\hat{b}_{i}, b_{\\psi(i)}) < \\tau_{2}\\) is a TP for \\(\\tau_{1}\\) but a FP for \\(\\tau_{2}\\). Thus, given Eq. (4), it follows that \\(\\hat{b}_i\\) has contradictory confidence targets for \\(\\tau_{1}\\) and \\(\\tau_{2}\\). This is illustrated in Fig. 3(d) in which D-ECE\\(_{\\mathrm{C}}\\) (red line) remains constant regardless of the confidence. Thus, using D-ECE (or another calibration measure) in this way should be avoided." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "3. Properly-designed datasets. In the literature, the val. set to obtain the post-hoc calibrator is typically taken from a different dataset than the ID dataset [41-43, 52]. Specifically, the post-hoc calibrators are obtained on a subset from Objects365 [60] and BDD100K [65] for the models trained with COCO [33] and Cityscapes [9] respectively. Hence, as expected, a different dataset inevitably induces domain shift, affecting the performance of the post-hoc calibrator [51]. To show that, following existing approaches, we obtain an IR calibrator [48] on" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.406, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.248, + 0.395, + 0.26 + ], + "angle": 0, + "content": "(a) Uncalibrated detections" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.145, + 0.596, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.248, + 0.576, + 0.26 + ], + "angle": 0, + "content": "(b) Calibrated detections" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.145, + 0.784, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.248, + 0.783, + 0.26 + ], + "angle": 0, + "content": "(c) Calibrated detections (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.262, + 0.408, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.369, + 0.372, + 0.381 + ], + "angle": 0, + "content": "(d) TP for \\(\\tau = 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.262, + 0.593, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.369, + 0.593, + 0.382 + ], + "angle": 0, + "content": "(e) FP for \\(\\tau = 0.5\\), TP for \\(\\tau = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.262, + 0.779, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.369, + 0.737, + 0.381 + ], + "angle": 0, + "content": "(f) FP for \\(\\tau = 0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 3: A pictorial comparison of the different calibration errors. (a) Uncalibrated detections of D-DETR on an image from [65]. The detections on the left and right have IoUs of 0.74 and 0.48 with the objects. (b) Calibrated detections in terms of D-ECE and LaECE using \\(\\tau = 0.50\\), and D-ECEc, COCO-style D-ECE as in [55]. \\(\\mathrm{D - ECE_C} = ?\\) as calibration error does not have a global minimum as shown in (d). (c) Calibrated detections in terms of LaECEo and LaACEo in which confidence matches IoU. (d-f) Calibration errors for different types of detections, for which LaACEo behave the same as LaECEo, hence excluded for clarity. App. B presents the details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.599 + ], + "angle": 0, + "content": "Objects365 and compare it with the one obtained on the ID val. set in terms of D-ECE-style evaluation. Tab. 2 shows that the latter IR now outperforms (i) the former one by \\(\\sim 11\\) D-ECE and (ii) SOTA Cal-DETR [43] by 7.4 D-ECE, showing the importance of dataset design for proper evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.751 + ], + "angle": 0, + "content": "4. Properly-trained detectors and calibrators. Though Cityscapes is commonly used in the literature [41-43,52], the models trained on this dataset follow COCO-style training. Specifically, D-DETR [71] was trained on Cityscapes for only 50 epochs though the training set of Cityscapes is \\(\\sim 40\\times\\) smaller than that of COCO (3K vs. \\(\\sim 118\\mathrm{K}\\)). We now tailor the training of D-DETR for Cityscapes by (i) \\(4\\times\\) longer training considering the smaller training set and (ii) increasing the training image scale considering the original resolution following [7,9]. We kept all other hyperparameters as they are for both Cal-DETR and D-DETR, and App. B presents the details. Tab. 3 shows that, once trained in this setting, D-DETR is more accurate and calibrated than Cal-DETR." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.782, + 0.801 + ], + "angle": 0, + "content": "4 A Framework for Joint Evaluation of Object Detectors" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.785, + 0.841 + ], + "angle": 0, + "content": "We now present our evaluation approach that respects to the principles in Sec. 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.495, + 0.215 + ], + "angle": 0, + "content": "Table 2: Effect of using domain-shifted val. set on IR calibrator. Results are reported on COCO-minitest. Val. set is N/A for uncalibrated D-DETR and training time calibration method Cal-DETR." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.228, + 0.48, + 0.298 + ], + "angle": 0, + "content": "
MethodVal setD-ECEAP ↑
D-DETRN/A12.844.1
Cal-DETRN/A8.744.4
IRObjects36514.244.1
IRCOCO1.3 (+7.4)44.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: COCO training settings are commonly adopted while training DDETR on Cityscapes. When trained with larger images and longer, DETR performs slightly better than Cal-DETR." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.227, + 0.787, + 0.298 + ], + "angle": 0, + "content": "
MethodTraining StyleD-ECEAP↑
D-DETR [43]COCO13.826.8
Cal-DETR [43]COCO8.428.4
Cal-DETRCityscapes4.034.9
D-DETRCityscapes2.936.1
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.326, + 0.774, + 0.342 + ], + "angle": 0, + "content": "4.1 Towards Fine-grained Calibrated Detection Confidence Scores" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.789, + 0.577 + ], + "angle": 0, + "content": "Calibration refers to the alignment of accuracy and confidence of a model. Therefore, for an object detector to be calibrated, its confidence should respect both classification and localisation accuracy. We discussed in Sec. 3 that D-ECE, as the common calibration measure, only considers the precision of a detector, thereby ignoring its localisation performance (Eq. (2)). LaECE [48], defined in Eq. (3) as an alternative to D-ECE, enforces the confidence scores to represent the product of precision and average IoU of TPs. Thus, LaECE considers IoUs of only TPs, and effectively ignores the localisation qualities of detections if their IoU is less than the TP validation threshold \\(\\tau > 0\\). We assert that this selection mechanism based on IoU unnecessarily limits the information conveyed by the confidence score. We illustrate this on the right car in Fig. 3(b) for which LaECE requires a target confidence of 0 (\\(\\hat{p}_{\\mathrm{LaECE}} = 0\\)) as its IoU is less than \\(\\tau = 0.50\\). However, instead of conveying a 0 confidence and implying no detection, representing its IoU by \\(\\hat{p}_i\\) provides additional information. Hence, we propose using \\(\\tau = 0\\), in which case the calibration criterion of LaECE reduces to," + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.588, + 0.787, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\hat {b} _ {i} \\in B _ {i} (\\hat {p} _ {i})} [ \\mathrm {I o U} (\\hat {b} _ {i}, b _ {\\psi (i)}) ] = \\hat {p} _ {i}, \\forall \\hat {p} _ {i} \\in [ 0, 1 ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.683 + ], + "angle": 0, + "content": "where we define \\(\\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)}) = 0\\) for FPs when \\(\\tau = 0\\), \\(B_i(\\hat{p}_i)\\) is the set of boxes with the confidence of \\(\\hat{p}_i\\) and \\(b_{\\psi(i)}\\) is the ground-truth box that \\(\\hat{b}_i\\) matches with. To derive the calibration error for Eq. (5), we follow LaECE by using \\(J = 25\\) equally-spaced bins and averaging over class-wise errors and define," + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.693, + 0.787, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L a E C E} _ {0} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} ^ {c} \\right|}{\\left| \\hat {\\mathcal {D}} ^ {c} \\right|} \\left| \\bar {p} _ {j} ^ {c} - \\operatorname {I o} \\mathrm {U} ^ {c} (j) \\right|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{D}}^c\\) and \\(\\hat{\\mathcal{D}}_j^c\\) denote the set of all detections and those in \\(j\\)th bin respectively, \\(\\bar{p}_j^c\\) is the average confidence score and \\(\\mathrm{IoU}^c (j)\\) is the average IoU of detections in the \\(j\\)-th bin for class \\(c\\), and the subscript 0 refers to the chosen \\(\\tau\\) which is 0. Furthermore, similar to the classification literature [40, 44], we define Localisation-aware Adaptive Calibration Error (LaACE) using an adaptive binning approach in which the number of detections in each bin is equal. In order" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.159 + ], + "angle": 0, + "content": "Table 4: Datasets for evaluating object detection and instance segmentation methods." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.171, + 0.77, + 0.227 + ], + "angle": 0, + "content": "
TypeTrain setVal setID test setDomain-shifted test set
Common ObjectsCOCO trainCOCO minivalCOCO minitestCOCO minitest-C, Obj45K
Autonomous DrivingCS trainCS minivalCS minitestCS minitest-C, Foggy-CS
Long-tailed ObjectsLVIS trainLVIS minivalLVIS minitestLVIS minitest-C
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.252, + 0.784, + 0.298 + ], + "angle": 0, + "content": "to capture the model behaviour precisely, we adopt the extreme case in which each bin has only one detection, resulting in an easy-to-interpret measure which corresponds to the mean absolute error between the confidence and the IoU," + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.309, + 0.785, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L a A C E} _ {0} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {i = 1} ^ {| \\hat {\\mathcal {D}} ^ {c} |} \\frac {1}{| \\hat {\\mathcal {D}} ^ {c} |} \\left| \\hat {p} _ {i} - \\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right) \\right|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.359, + 0.784, + 0.42 + ], + "angle": 0, + "content": "As we show in App. C, \\(\\mathrm{LaECE}_0\\) and \\(\\mathrm{LaACE}_0\\) are both minimized when \\(\\hat{p}_i = \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})\\) for all detections, which is also a necessary condition for \\(\\mathrm{LaACE}_0\\). Hence, as illustrated on the right car in Fig. 3(c) and (e), \\(\\mathrm{LaECE}_0\\) and \\(\\mathrm{LaACE}_0\\) requires conveying more fine-grained information compared to other measures." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.44, + 0.753, + 0.456 + ], + "angle": 0, + "content": "4.2 Model-dependent Thresholding for Proper Joint Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.616 + ], + "angle": 0, + "content": "In practice, object detectors employ an operating threshold to preferably output only TPs with high recall. However, AP as the common performance measure does not enable cross-validating such a threshold as it is maximized when the recall is maximized despite a drop in precision [47, 48]. This can be observed in Fig. 1(b) where AP consistently decreases as the confidence threshold increases. Alternatively, LRP (Eq. 1) prefers detectors with high precision, recall and low localisation error as illustrated by the bell-like curves in Fig. 1(d). This is because, unlike AP, LRP severely penalises detectors with low recall or precision, making it a perfect fit for our framework. As a result, we consider \\(\\mathrm{LaECE}_0\\) and \\(\\mathrm{LaACE}_0\\) with LRP and require each model to be thresholded properly." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.636, + 0.49, + 0.651 + ], + "angle": 0, + "content": "4.3 Properly-designed Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We curate three datasets summarized in Tab. 4: (i) COCO [33] including common daily objects; (ii) Cityscapes [9] with autonomous driving scenes; and (iii) LVIS [16], a challenging dataset focusing on the calibration of long-tailed detection. For each dataset, we ensure that train, val. and ID test sets are sampled from the same distribution, and include domain-shifted test sets. As these datasets do not have public labels for test sets, we randomly split their val. sets into two as minival and minitest similar to [17,48,49]. In such a way, we provide ID val. sets to enable obtaining post-hoc calibrators and the operating thresholds properly. For domain-shifted test sets, we apply common corruptions [21] to the ID test sets, and include Obj45K [48,60] and Foggy Cityscapes [59] as more realistic shifts. Our datasets also have mask annotations and hence they can be used to evaluate instance segmentation methods. App. C includes further details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.75, + 0.162 + ], + "angle": 0, + "content": "4.4 Baseline Post-hoc Calibrators Tailored to Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.365 + ], + "angle": 0, + "content": "It is essential to develop post-hoc calibration methods tailored to object detection, which has certain differences from the classification task. Existing methods [41-43,52,55] use only TS as a baseline without considering the peculiarities of detection. Specifically, a single temperature parameter \\( T \\) is learned to adjust the predictive distribution while the confidence \\( \\hat{p}_i \\) is commonly assumed to be a Bernoulli random variable [30]. However, PS, fitting both a scale and a shift parameter, is the widely-accepted calibration approach when the underlying distribution is Bernoulli [15,54]. Also, how to construct a useful subset of the detections to train the post-hoc calibrators has not been explored. To address these shortcomings, we present (i) Platt Scaling in which the bias term makes a notable difference in the performance, and (ii) Isotonic Regression by modeling the calibration as a regression problem. Before introducing them, we now present an overview on how we determine the set of detections to train the calibrators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.561 + ], + "angle": 0, + "content": "Overview We obtain post-hoc calibrators on a held out val. set using the detections that are similar to those seen at inference to prevent low-scoring detections from dominating the training of the calibrator. To do so, we crossvalidate a calibration threshold \\(\\bar{u}^c\\) for each class \\(c\\) and train a class-specific calibrator \\(\\zeta^c: [0,1] \\to [0,1]\\) using the detections with higher scores than \\(\\bar{u}^c\\). Still, as \\(\\zeta^c(\\cdot)\\) changes the confidence scores, we need another threshold \\(\\bar{v}^c\\), as the operating threshold, to remove the redundant detections after calibration. Following the accuracy measure, we cross-validate \\(\\bar{u}^c\\) and \\(\\bar{v}^c\\) using LRP. As for inference time, for the \\(i\\)-th detection \\((\\hat{p}_i, \\hat{b}_i, \\hat{c}_i)\\), if \\(\\hat{p}_i \\geq \\bar{u}^{\\hat{c}_i}\\), it survives to the calibrator and then \\(\\hat{p}_i^{cal} = \\zeta^{\\hat{c}_i}(\\hat{p}_i)\\). Finally, if \\(\\hat{p}_i^{cal} \\geq \\bar{v}^{\\hat{c}_i}\\), the \\(i\\)-th detection is an output of the detector. Alg. A.1 and A.2 provide the details. For \\(\\zeta^c(\\cdot)\\), we prefer monotonically increasing functions in order not to affect the ranking of the detections significantly and to keep their accuracy as we detail below." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.562, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Distribution Calibration via Platt Scaling Assuming that \\(\\hat{p}_i\\) is sampled from Bernoulli distribution \\(\\mathcal{B}(\\cdot)\\), we aim to minimize the Negative Log-Likelihood (NLL) of the predictions on the target distribution \\(\\mathcal{B}(\\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)}))\\) using PS [54]. Accordingly, we recover the logits, and then shift and scale the logits to obtain the calibrated probabilities \\(\\hat{p}_i^{cal}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.647, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {i} ^ {\\text {c a l}} = \\sigma \\left(a \\sigma^ {- 1} \\left(\\hat {p} _ {i}\\right) + b\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.671, + 0.787, + 0.702 + ], + "angle": 0, + "content": "where \\(\\sigma(\\cdot)\\) is the sigmoid and \\(\\sigma^{-1}(\\cdot)\\) is its inverse, as well as \\(a \\geq 0\\) and \\(b\\) are the learnable parameters. We derive the NLL for the \\(i\\)th detection in App. C as" + }, + { + "type": "equation", + "bbox": [ + 0.29, + 0.709, + 0.785, + 0.726 + ], + "angle": 0, + "content": "\\[\n- \\left(\\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right) \\log \\left(\\hat {p} _ {i} ^ {c a l}\\right) + \\left(1 - \\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right)\\right) \\log \\left(1 - \\hat {p} _ {i} ^ {c a l}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.733, + 0.787, + 0.781 + ], + "angle": 0, + "content": "Please note that Eq. (9), which is in fact the cross-entropy loss, is minimized if \\(\\hat{p}_i^{cal} = \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})\\) when \\(\\mathrm{LaECE}_0\\) and \\(\\mathrm{LaACE}_0\\) are minimized. We optimize Eq. (9) via the second-order optimization strategy L-BFGS [35] following [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.803, + 0.842 + ], + "angle": 0, + "content": "Confidence Calibration via Isotonic Regression As an alternative perspective, \\(\\hat{p}_i\\) can also be directly calibrated by modelling the calibration as a regression task. To do so, we construct the prediction-target pairs \\((\\{\\hat{p}_i, \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})\\})\\) on the held-out val. set and then fit an IR model using scikit-learn [53]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Comparison with SOTA methods in terms of other evaluation measures on COCO [33]. LRP is reported on LRP-optimal thresholds obtained on val. set. AP is reported on top-100 detections. \\(\\tau\\) is taken as 0.50. All measures are lower-better, except AP. Bold: the best, underlined: second best. PS: Platt Scaling, IR: Isotonic Regression." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.213, + 0.78, + 0.361 + ], + "angle": 0, + "content": "
Cal.\nTypeMethodCalibration (thr. 0.30)Calibration (LRP thr.)Accuracy
D-ECELaECED-ECELaECELRPAP↑
Uncal.D-DETR [71]12.813.215.012.166.344.1
Training\nTimeMbLS [34]15.616.318.715.865.944.3
MDCA [19]12.213.514.312.666.443.8
TCD [42]12.413.114.412.366.644.0
BPC [41]9.813.111.412.366.843.6
Cal-DETR [43]8.712.99.711.866.044.4
Post-hoc\n(Ours)PS for D-ECE0.9(+7.8)16.32.4(+7.3)15.866.344.1
PS for LaECE11.011.5(+1.4)9.410.1(+1.7)66.344.1
IR for D-ECE1.3(+7.4)15.72.6(+7.1)15.366.244.1
IR for LaECE10.28.9(+4.0)9.38.2(+3.6)66.343.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.789, + 0.496 + ], + "angle": 0, + "content": "Adapting Our Approach to Different Calibration Objectives Until now, we considered post-hoc calibrators for \\(\\mathrm{LaECE_0}\\) and \\(\\mathrm{LaACE_0}\\) though in practice different measures can be preferred. Our post-hoc calibrators can easily be adapted for such cases by considering the dataset design and optimisation criterion. To illustrate, for D-ECE-style evaluation, the calibration dataset is to be class-agnostic where the detections are thresholded from 0.30 with prediction-target pairs for IR as \\((\\{\\hat{p}_i,0\\})\\) and \\((\\{\\hat{p}_i,1\\})\\) for FPs and TPs respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.519, + 0.494, + 0.538 + ], + "angle": 0, + "content": "5 Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.598 + ], + "angle": 0, + "content": "We now show that our post-hoc calibration approaches consistently outperform training time calibration methods by significant margins (Sec. 5.1) and that they generalize to any detector and can thus be used as a strong baseline (Sec. 5.2)." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.619, + 0.75, + 0.635 + ], + "angle": 0, + "content": "5.1 Comparing Our Baselines with SOTA Calibration Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Here, we compare PS and IR with recent training-time calibration methods considering various evaluation approaches. As these training-time methods mostly rely on D-DETR, we also use D-DETR with ResNet-50 [18]. We obtain the detectors of training time approaches trained with COCO dataset from their official repositories, whereas we incorporate Cityscapes into their official repositories and train them using the recommended setting in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Comparison on Other Evaluation Approaches Before moving on to our evaluation approach, we first show that our PS and IR outperform all existing training time methods on existing evaluation approaches. For that, we consider D-ECE and the LaECE from \\(\\tau = 0.5\\) by including two different evaluation settings for each: (i) the detection set is obtained from the fixed threshold of 0.30 following the convention [30,41-43], and (ii) the operating thresholds are cross-validated using LRP. Following their standard usage, we use 10 and 25" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 6: Comparison with SOTA calibration methods on Common Objects using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.2, + 0.773, + 0.348 + ], + "angle": 0, + "content": "
Calibration TypeMethodCOCO minitest (ID)COCO-C (Domain Shift)Obj45K (Domain Shift)
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
UncalibratedD-DETR [71]12.727.157.314.628.771.516.435.872.0
Training-timeMbLS [34]16.530.356.816.831.171.817.337.171.6
MDCA [19]13.127.257.514.528.771.816.635.672.2
TCD [42]13.026.757.614.628.371.916.335.571.7
BPC [41]12.425.557.714.127.172.117.334.572.0
Cal-DETR [43]11.624.656.213.826.470.618.835.371.1
Post-hoc (Ours)Platt Scaling9.623.557.312.825.671.517.033.772.0
Isotonic Regression7.723.157.210.725.371.517.233.372.0
(+3.9)(+1.5)(+3.1)(+1.1)(-0.9)(+1.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.371, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Table 7: Comparison with SOTA on Autonomous Driving using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.426, + 0.773, + 0.555 + ], + "angle": 0, + "content": "
Calibration TypeMethodCityscapes minitest (ID) LaECE0|LaACE0|LRPCityscapes-C (Domain Shift) LaECE0|LaACE0|LRPFoggy Cityscapes (Domain Shift) LaECE0|LaACE0|LRP
UncalibratedD-DETR [71]20.326.057.221.425.680.218.522.369.4
Training-timeTCD [42]16.831.759.223.232.481.624.433.871.6
BPC [41]23.831.864.928.133.383.724.730.973.8
Cal-DETR [43]21.325.356.923.026.480.820.023.271.0
Post-hoc (Ours)Platt Scaling9.623.357.217.726.280.211.321.669.4
Isotonic Regression9.0(+7.8)23.7(+1.6)56.816.4(+5.0)25.8(-0.2)80.510.0(+8.5)21.2(+1.1)69.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.716 + ], + "angle": 0, + "content": "bins to compute D-ECE and LaECE respectively. We optimize PS and IR by considering the calibration objective as described in Sec. 4.4. Tab. 5 shows that PS and IR outperform SOTA Cal-DETR significantly by more than 7 D-ECE and up to 4 LaECE on COCO minitest. Please note that all previous approaches are optimized for \\(D\\)-ECE thresholded from 0.30, in terms of which our PS yields only 0.9 \\(D\\)-ECE improving the SOTA by 7.8. Tab. 5 also suggests that post-hoc calibrators perform the best when the calibration objective is aligned with the measure. App. D shows that our observations generalize to Cityscapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Common Objects Setting We now evaluate detectors using our evaluation approach. Tab. 6 shows that IR and PS share the top-2 entries on almost all test subsets by preserving the accuracy (LRP) of D-DETR. Specifically, our gains on ID set and COCO-C are significant, where IR outperforms Cal-DETR by around \\(3 - 4\\mathrm{LaECE}_0\\) and \\(1.0 - 1.5\\mathrm{LaACE}_0\\). As for Obj45K, the challenging test set with natural shift, IR and PS improve \\(\\mathrm{LaACE}_0\\) but perform slightly worse in terms of \\(\\mathrm{LaECE}_0\\). This is an expected drawback of post-hoc approaches when the domain shift is large as they are trained only with ID val. set [51]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 8: Comparison with TS using D-DETR. Bold: the best, underlined: second best. X : domain-shifted val. set is used to obtain thresholds and calibrators, decreasing the accuracy (red font). Bias term only exists for PS ( \\( b \\) in Eq. (8)), thus \\( \\mathrm{N}/\\mathrm{A} \\) for IR." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.2, + 0.784, + 0.311 + ], + "angle": 0, + "content": "
MethodAblations on DatasetAblations on ModelCOCO minitestCityscapes minitest
ID Val. SetThresholdClass-wiseBias TermLaECE0LaACE0LRPLaECE0LaACE0LRP
Temperature Scaling (Current Baseline)✘12.320.861.521.025.960.3
✓12.523.157.320.926.357.2
Ablations on Temperature Scaling✓✓11.324.857.313.325.557.2
✓✓12.422.957.323.127.557.2
✓✓✓10.624.257.312.724.657.2
Platt Scaling (Ours) Isotonic Regression (Ours)✓✓✓✓9.623.557.39.623.357.2
✓✓✓N/A7.723.157.29.023.756.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.324, + 0.784, + 0.366 + ], + "angle": 0, + "content": "Table 9: Calibrating and evaluating different object detectors. We use Common Objects setting and report the results on COCO minitest. * denotes the detectors in Fig. 1. Bold: the best, underlined: second best for calibration." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.38, + 0.784, + 0.569 + ], + "angle": 0, + "content": "
TypeDetectorBackboneUncalibratedPlatt ScalingIsotonic RegressionAP ↑
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
One-StagePAA [25]*R5015.928.159.79.724.359.77.723.859.743.2
ATSS [70]*R5019.134.059.510.324.759.58.524.159.543.1
GFL [32]R5013.728.559.310.324.559.38.324.059.343.0
VFNet [68]R5013.925.857.710.725.157.78.324.657.744.8
Two-StageFaster R-CNN [57]*R5027.029.960.410.423.860.48.623.560.440.1
RS R-CNN [46]*R5019.728.958.710.223.558.78.123.058.842.4
DETR-likeD-DETR [71]*R5012.727.157.39.623.557.37.723.157.244.1
UP-DETR [10]R5034.435.255.810.022.655.88.222.255.942.9
DINO [67]R5013.626.953.610.623.553.68.922.853.650.4
OVODGLIP [31]Swin-T13.025.349.09.222.449.07.721.849.055.7
G. DINO [36]Swin-T13.827.546.98.921.946.97.721.347.058.3
SOTACo-DETR [72]Swin-L10.823.041.58.620.241.56.719.341.664.5
EVA [12]ViT(EVA)17.421.241.28.620.341.27.119.941.264.5
MoCaE [49]N/A10.621.440.78.920.440.77.319.940.765.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.659 + ], + "angle": 0, + "content": "Autonomous Driving Setting Tab. 7 shows that our approaches consistently outperform all training time calibration approaches on this setting as well. Specifically, our gains are very significant ranging between 5.0-8.5 \\(\\mathrm{LaECE}_0\\) compared to the SOTA Cal-DETR, further presenting the efficacy of our approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison with Existing Temperature Scaling Baseline and Ablations Tab. 8 compares TS for different design choices as well as with our PS and IR. Please note that \\(\\pmb{x}\\) corresponds to the baseline setting used in the recent approaches [41-43,52] that employ Objects365 [60] and BDD100K [60] as domain-shifted val. sets for obtaining the calibrator. Due to this domain shift, the accuracy of TS degrades by up to 4 LRP, in red font, as the operating thresholds obtained on these val. sets do not generalize to the ID set; showing that it is crucial to use an ID val set. In ablations, thresholding the detections and class-wise calibrators generally improves the performance of TS and a more notable gain is observed once the bias term is used in PS. Our PS outperforms TS baseline obtained on ID val. set by \\(\\sim 3\\) LaECE\\(_0\\) on COCO and 11.4 LaECE\\(_0\\) on Cityscapes. Finally, IR performs on par or better compared to PS." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.728, + 0.163 + ], + "angle": 0, + "content": "5.2 Calibrating and Evaluating Different Detection Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.47, + 0.382 + ], + "angle": 0, + "content": "Another benefit of our post-hoc calibrators is that they generalize to any object detector, thus they can be reliably used as baselines. To show that, we calibrate 14 different detectors with a diverse set of architectures using PS and IR in Tab. 9. The results suggest that both IR and PS perform better than uncalibrated detectors. IR consistently outperforms PS as it fits multiple piece-wise linear functions while PS learns only two parameters (Eq. (8)). Specifically, IR" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.203, + 0.629, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.314, + 0.611, + 0.326 + ], + "angle": 0, + "content": "(a) Uncalibrated" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.203, + 0.776, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.314, + 0.769, + 0.326 + ], + "angle": 0, + "content": "(b) Calibrated by IR" + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.337, + 0.785, + 0.351 + ], + "angle": 0, + "content": "Fig. 4: The reliability diagrams of UP-DETR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.787, + 0.517 + ], + "angle": 0, + "content": "decreases the range of \\(\\mathrm{LaECE_0}\\) from 10.6-34.4 to 6.7-8.9 on COCO minitest by preserving the accuracy, making it a solid baseline. For further insights, Fig. 4 provides the reliability diagrams of the overconfident UP-DETR, of which IR significantly improves its calibration. As for SOTA, MoCaE performs the best in terms of accuracy with 40.7 LRP while Co-DETR has the best calibration with \\(6.7\\mathrm{LaECE}_0\\) and \\(19.3\\mathrm{LaACE}_0\\), which the future work should aim to surpass. App. D.3 includes our results on our Long-tailed Objects setting (including object detection and instance segmentation), showing the effectiveness of our post-hoc calibrators on this challenging setting as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.789, + 0.653 + ], + "angle": 0, + "content": "The performance measures we use in our evaluation framework are also easily interpretable. For accuracy, LRP is a weighted combination of its FP, FN and localisation error components (App. A), which, as an example, are 10.1, 22.7 and 18.8 respectively for Co-DETR [72] calibrated with IR. Also considering \\(\\mathrm{LaACE}_0 = 19.3\\), one can easily infer that: once deployed with the operating thresholds determined by our framework, Co-DETR finds \\(77.3\\%\\) of the objects with \\(89.9\\%\\) precision and \\(81.2\\%\\) IoU where \\(19.3\\%\\) is the mean absolute error of the confidence to represent IoU. We believe these intuitive measures will enable practitioners to make better decision when deploying object detectors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.37, + 0.691 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The progress in a field heavily relies on the evaluation tools and the baselines used. In this paper, we showed that existing evaluation tools for calibration as well as the baseline post-hoc calibrators for object detectors have significant drawbacks. We remedied that by introducing an evaluation framework including baseline post-hoc calibrators tailored to object detection. Our experiments suggested that, once evaluated and designed properly, the post-hoc calibrators significantly outperform all existing training-time calibrators. This implies the need for research to develop better calibration techniques for object detection, for which, we believe, our evaluation framework will be an essential pillar." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.223 + ], + "angle": 0, + "content": "1. Ayer, M., Brunk, H.D., Ewing, G.M., Reid, W.T., Silverman, E.: An empirical distribution function for sampling with incomplete information. The annals of mathematical statistics pp. 641-647 (1955)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.251 + ], + "angle": 0, + "content": "2. Barlow, R.E., Brunk, H.D.: The isotonic regression problem and its dual. Journal of the American Statistical Association 67(337), 140-147 (1972)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.279 + ], + "angle": 0, + "content": "3. Best, M.J., Chakravarti, N.: Active set algorithms for isotonic regression; a unifying framework. Mathematical Programming 47(1), 425-439 (1990)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.28, + 0.785, + 0.335 + ], + "angle": 0, + "content": "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2016). https://doi.org/10.1109/icip.2016.7533003, http://dx.doi.org/10.1109/ICIP.2016.7533003" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.785, + 0.378 + ], + "angle": 0, + "content": "5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: The IEEE European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.434 + ], + "angle": 0, + "content": "6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.49 + ], + "angle": 0, + "content": "7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv 1906.07155 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.532 + ], + "angle": 0, + "content": "8. Cheng, J., Vasconcelos, N.: Calibrating deep neural networks by pairwise constraints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.588 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.643 + ], + "angle": 0, + "content": "0. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.785, + 0.686 + ], + "angle": 0, + "content": "1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) 88(2), 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.785, + 0.742 + ], + "angle": 0, + "content": "2. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.785 + ], + "angle": 0, + "content": "3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.839 + ], + "angle": 0, + "content": "4. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (Nov 2019). https://doi.org/10.1002/rob.21918, http://dx.doi.org/10.1002/rob.21918" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1321-1330. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.289 + ], + "angle": 0, + "content": "17. Harakeh, A., Waslander, S.L.: Estimating and evaluating regression predictive uncertainty in deep object detectors. In: International Conference on Learning Representations (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.389 + ], + "angle": 0, + "content": "19. Hebbalaguppe, R., Prakash, J., Madan, N., Arora, C.: A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16081-16090 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.459 + ], + "angle": 0, + "content": "20. Hekler, A., Brinker, T.J., Buettner, F.: Test time augmentation meets post-hoc calibration: Uncertainty quantification under real-world conditions. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14856-14864 (Jun 2023). https://doi.org/10.1609/aaai.v37i12.26735, https://ojs.aaai.org/index.php/AAAI/article/view/26735" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.501 + ], + "angle": 0, + "content": "21. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (ICLR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.585 + ], + "angle": 0, + "content": "22. Jin, C., Udupa, J.K., Zhao, L., Tong, Y., Odhner, D., Pednekar, G., Nag, S., Lewis, S., Poole, N., Mannikeri, S., Govindasamy, S., Singh, A., Camaratta, J., Owens, S., Torigian, D.A.: Object recognition in medical images via anatomy-guided deep learning. Medical Image Analysis 81, 102527 (2022). https://doi.org/https://doi.org/10.1016/j.media.2022.102527, https://www.sciencedirect.com/science/article/pii/S1361841522001748" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.655 + ], + "angle": 0, + "content": "23. Joy, T., Pinto, F., Lim, S.N., Torr, P.H., Dokania, P.K.: Sample-dependent adaptive temperature scaling for improved calibration. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14919-14926 (Jun 2023). https://doi.org/10. 1609/aaii.v37i12.26742, https://ojs.aaii.org/index.php/AAAI/article/view/26742" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.724 + ], + "angle": 0, + "content": "24. Karimi, D., Dou, H., Warfield, S.K., Gholipour, A.: Deep learning with noisy labels: Exploring techniques and remedies in medical image analysis. Medical Image Analysis 65, 101759 (2020). https://doi.org/https://doi.org/10.1016/j.media.2020.101759, https://www.sciencedirect.com/science/article/pii/S1361841520301237" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.727, + 0.787, + 0.755 + ], + "angle": 0, + "content": "25. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.811 + ], + "angle": 0, + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "27. Kumar, A., Liang, P.S., Ma, T.: Verified uncertainty calibration. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.355 + ], + "angle": 0, + "content": "28. Kumar, N., Verma, R., Anand, D., Zhou, Y., Onder, O.F., Tsougenis, E., Chen, H., Heng, P.A., Li, J., Hu, Z., Wang, Y., Koohbanani, N.A., Jahanifar, M., Tajeddin, N.Z., Gooya, A., Rajpoot, N., Ren, X., Zhou, S., Wang, Q., Shen, D., Yang, C.K., Weng, C.H., Yu, W.H., Yeh, C.Y., Yang, S., Xu, S., Yeung, P.H., Sun, P., Mahbod, A., Schaefer, G., Ellinger, I., Ecker, R., Smedby, O., Wang, C., Chidester, B., Ton, T.V., Tran, M.T., Ma, J., Do, M.N., Graham, S., Vu, Q.D., Kwak, J.T., Gunda, A., Chanduri, R., Hu, C., Zhou, X., Lotfi, D., Safdari, R., Kascenas, A., O'Neil, A., Eschweiler, D., Stegmaier, J., Cui, Y., Yin, B., Chen, K., Tian, X., Gruening, P., Barth, E., Arbel, E., Remer, I., Ben-Dor, A., Sirazitdinova, E., Kohl, M., Braunewell, S., Li, Y., Xie, X., Shen, L., Ma, J., Baksi, K.D., Khan, M.A., Choo, J., Colomer, A., Naranjo, V., Pei, L., Iftekharuddin, K.M., Roy, K., Bhattacharjee, D., Pedraza, A., Bueno, M.G., Devanathan, S., Radhakrishnan, S., Koduganty, P., Wu, Z., Cai, G., Liu, X., Wang, Y., Sethi, A.: A multi-organ nucleus segmentation challenge. IEEE Transactions on Medical Imaging 39(5), 1380–1391 (2020). https://doi.org/10.1109/TMI.2019.2947628" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.411 + ], + "angle": 0, + "content": "29. Kumar, N., Verma, R., Sharma, S., Bhargava, S., Vahadane, A., Sethi, A.: A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging 36(7), 1550-1560 (2017). https://doi.org/10.1109/TMI.2017.2677499" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.452 + ], + "angle": 0, + "content": "30. Kuppers, F., Kronenberger, J., Shantia, A., Haselhoff, A.: Multivariate confidence calibration for object detection. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.508 + ], + "angle": 0, + "content": "31. Li, L.H., Zhang, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.55 + ], + "angle": 0, + "content": "32. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.786, + 0.592 + ], + "angle": 0, + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.619 + ], + "angle": 0, + "content": "34. Liu, B., Ayed, I.B., Galdran, A., Dolz, J.: The devil is in the margin: Margin-based label smoothing for network calibration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.66 + ], + "angle": 0, + "content": "35. Liu, D.C., Nocedal, J.: On the limited memory bfgs method for large scale optimization. Math. Program. 45(1-3), 503-528 (1989), http://dblp.uni-trier.de/db/journals/mp/mp45.htm1#LiuN89" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "36. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.743 + ], + "angle": 0, + "content": "37. Lu, Y., Lu, C., Tang, C.K.: Online video object detection using association LSTM. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 2363-2371 (2017). https://doi.org/10.1109/ICCV.2017.257" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.811 + ], + "angle": 0, + "content": "38. Ma, X., Blaschko, M.B.: Meta-cal: Well-controlled post-hoc calibration by ranking. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 7235-7245. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/ma21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "39. Mehrtash, A., Wells, W.M., Tempany, C.M., Abolmaesumi, P., Kapur, T.: Confidence calibration and predictive uncertainty estimation for deep medical image" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.188 + ], + "angle": 0, + "content": "segmentation. IEEE Transactions on Medical Imaging 39(12), 3868-3878 (Dec 2020). https://doi.org/10.1109/tmi.2020.3006437, http://dx.doi.org/10.1109/TMI.2020.3006437" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.26 + ], + "angle": 0, + "content": "40. Mukhoti, J., Kulharia, V., Sanyal, A., Golodetz, S., Torr, P., Dokania, P.: Calibrating deep neural networks using focal loss. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 15288-15299. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/aeb7b30ef1d024a76f21a1d40e30c302-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.316 + ], + "angle": 0, + "content": "41. Munir, M.A., Khan, M.H., Khan, S., Khan, F.S.: Bridging precision and confidence: A train-time loss for calibrating object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11474-11483 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.317, + 0.787, + 0.4 + ], + "angle": 0, + "content": "42. Munir, M.A., Khan, M.H., Sarfraz, M., Ali, M.: Towards improving calibration in object detection under domain shift. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 38706-38718. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/fcd812a51b8f8d05cfea22e3c9c4b369-Paper-Conference.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.401, + 0.787, + 0.443 + ], + "angle": 0, + "content": "43. Munir, M.A., Khan, S., Khan, M.H., Ali, M., Khan, F.: Cal-DETR: Calibrated detection transformer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=4SkPTD6XNP" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.444, + 0.787, + 0.486 + ], + "angle": 0, + "content": "44. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.486, + 0.787, + 0.528 + ], + "angle": 0, + "content": "45. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Localization recall precision (LRP): A new performance metric for object detection. In: The European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.572 + ], + "angle": 0, + "content": "46. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: The International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.614 + ], + "angle": 0, + "content": "47. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: One metric to measure them all: Localisation recall precision (lrp) for evaluating visual detection tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-1 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.655 + ], + "angle": 0, + "content": "48. Oksuz, K., Joy, T., Dokania, P.K.: Towards building self-aware object detectors via reliable uncertainty quantification and calibration. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.685 + ], + "angle": 0, + "content": "49. Oksuz, K., Kuzucu, S., Joy, T., Dokania, P.K.: Mocae: Mixture of calibrated experts significantly improves object detection. arXiv preprint arXiv:2309.14976 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.741 + ], + "angle": 0, + "content": "50. Otani, M., Togashi, R., Nakashima, Y., Rahtu, E., Heikkilä, J., Satoh, S.: Optimal correction cost for object detection evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21107-21115 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.787, + 0.798 + ], + "angle": 0, + "content": "51. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In: Advances in Neural Information Processing Systems. vol. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "52. Pathiraja, B., Gunawardhana, M., Khan, M.H.: Multiclass confidence and localization calibration for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.494, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Calibration of Object Detectors" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.229 + ], + "angle": 0, + "content": "54. Platt, J.: Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Adv. Large Margin Classif. 10 (06 2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.787, + 0.284 + ], + "angle": 0, + "content": "55. Popordanoska, T., Tiulpin, A., Blaschko, M.B.: Beyond classification: Definition and density-based estimation of calibration in object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 585-594 (January 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.787, + 0.324 + ], + "angle": 0, + "content": "56. Rahimi, A., Mensink, T., Gupta, K., Ajanthan, T., Sminchisescu, C., Hartley, R.: Post-hoc calibration of neural networks by g-layers. arXiv preprint arXiv:2006.12807 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.787, + 0.365 + ], + "angle": 0, + "content": "57. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.787, + 0.405 + ], + "angle": 0, + "content": "58. Rezatofighi, H., Nguyen, T.T.D., Vo, B., Vo, B., Savarese, S., Reid, I.D.: How trustworthy are the existing performance evaluations for basic vision tasks? arXiv e-prints:2008.03533 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.787, + 0.445 + ], + "angle": 0, + "content": "59. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision 126(9), 973-992 (Sep 2018), https://doi.org/10.1007/s11263-018-1072-8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "60. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.486, + 0.787, + 0.541 + ], + "angle": 0, + "content": "61. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.541, + 0.787, + 0.623 + ], + "angle": 0, + "content": "62. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11809-11820. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/61f3a6dbc9120ea78ef75544826c814e-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.623, + 0.787, + 0.663 + ], + "angle": 0, + "content": "63. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.663, + 0.787, + 0.704 + ], + "angle": 0, + "content": "64. Yan, K., Wang, X., Lu, L., Summers, R.M.: Deepleson: Automated deep mining, categorization and detection of significant radiology image findings using large-scale clinical lesion annotations. arXiv preprint arXiv:1710.01766 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.759 + ], + "angle": 0, + "content": "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "66. Zadrozny, B., Elkan, C.: Transforming classifier scores into accurate multiclass probability estimates. In: Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 694-699 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "67. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "Kuzucu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "68. Zhang, H., Wang, Y., Dayoub, F., Sünderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.258 + ], + "angle": 0, + "content": "69. Zhang, J., Yao, W., Chen, X., Feng, L.: Transferable post-hoc calibration on pretrained transformers in noisy text classification. Proceedings of the AAAI Conference on Artificial Intelligence 37(11), 13940-13948 (Jun 2023). https://doi.org/10.1609/aaai.v37i11.26632, https://ojs.aai.org/index.php/AAAI/article/view/26632" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "70. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "71. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable \\(\\{\\mathrm{detr}\\}\\): Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "72. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_origin.pdf b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6aa2acaba0d0d9dea5153b033b25fa24ac682bab --- /dev/null +++ b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/be195471-6153-4286-905c-575a7ab7761b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e94a834d9edefc26aa877a32328813d85476b03a3f7b8db5231c9f437698302 +size 1757057 diff --git a/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/full.md b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db0aaa484ea8119d5bae4b3868db586ff961859e --- /dev/null +++ b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/full.md @@ -0,0 +1,337 @@ +# On Calibration of Object Detectors: Pitfalls, Evaluation and Baselines + +Selim Kuzucu\*, Kemal Oksuz\*, Jonathan Sadeghi, and Puneet K. Dokania + +Five AI Ltd., United Kingdom {selim.kuzucu2, kemal.oksuz, jonathan.sadeghi, puneet.dokania}@five.ai + +Abstract. Reliable usage of object detectors require them to be calibrated—a crucial problem that requires careful attention. Recent approaches towards this involve (1) designing new loss functions to obtain calibrated detectors by training them from scratch, and (2) post-hoc Temperature Scaling (TS) that learns to scale the likelihood of a trained detector to output calibrated predictions. These approaches are then evaluated based on a combination of Detection Expected Calibration Error (D-ECE) and Average Precision. In this work, via extensive analysis and insights, we highlight that these recent evaluation frameworks, evaluation metrics, and the use of TS have notable drawbacks leading to incorrect conclusions. As a step towards fixing these issues, we propose a principled evaluation framework to jointly measure calibration and accuracy of object detectors. We also tailor efficient and easy-to-use post-hoc calibration approaches such as Platt Scaling and Isotonic Regression specifically for object detection task. Contrary to the common notion, our experiments show that once designed and evaluated properly, post-hoc calibrators, which are extremely cheap to build and use, are much more powerful and effective than the recent train-time calibration methods. To illustrate, D-DETR with our post-hoc Isotonic Regression calibrator outperforms the recent train-time state-of-the-art calibration method Cal-DETR by more than 7 D-ECE on the COCO dataset. Additionally, we propose improved versions of the recently proposed Localization-aware ECE and show the efficacy of our method on these metrics. Code is available at: https://github.com/fiveai/detection_calibration. + +Keywords: Calibration $\cdot$ Object Detection $\cdot$ Performance Evaluation + +# 1 Introduction + +Object detectors have been widely-used in a variety of safety-critical applications related to, but not limited to, autonomous driving [6, 9, 13, 14, 61, 65] and medical imaging [22, 28, 29, 64]. In addition to being accurate, their confidence estimates should also allow characterization of their error behaviour to make them reliable. This feature, known as calibration, can enable a model to provide valuable information to subsequent systems playing crucial role in making + +![](images/d7e2b008d9e7a264bfc5bb9abac53da3f2feb9a4a639484f78ea7abe627f7b2a.jpg) +(a) D-ECE + +![](images/3a2687f6620eb68b4dc6e79b3f122c1689f977dd67b2b7414500ac4fa3b20133.jpg) +(b) AP +Fig. 1: The performance of different detectors over operating confidence thresholds on COCO minitest. Orange: Faster R-CNN, Green: RS R-CNN, Purple: ATSS, Red: PAA, Blue: D-DETR. All measures are lower better except AP. It is not trivial to identify an operating threshold and compare detectors, especially when the common evaluation [30, 41–43, 52], combining D-ECE for calibration and AP for accuracy, is used. Instead, we use LaECE0 and Localisation-Recall-Precision Error (LRP). + +![](images/9ae363959673ad87211e181932b1e66023918fb521ed7df0166e5834d0b8c275.jpg) +(c) LaECEo + +![](images/0cd98c320f1287d7ad646c9f8c773df7be549c494a0ef3ea458a0063e82dbe54.jpg) +(d) LRP + +safety-critical decisions [4, 24, 26, 37, 39]. Despite its importance, calibration of detectors is a relatively underexplored area in the literature and requires significant attention. Therefore, in this work, we focus on different aspects of the evaluation framework that is now being adopted by most recent works building calibrated detectors and discuss their pitfalls and propose fixes. Additionally, we tailor the well-known post-hoc calibration methods to improve the calibration of a given object detector (trained) with minimal effort. + +Naturally, practitioners prefer detectors that perform well in terms of both accuracy and calibration, which we will refer to as joint performance. However, unlike classification, choosing the best performing model is non-trivial for object detection. This is because different detectors commonly yield detection sets with varying cardinalities for the same image, and this difference in population size is shown to affect the joint performance evaluation [48]. Furthermore, when object detectors are used in practice, an operating threshold is normally chosen [4, 24, 26, 31, 36, 37, 39], and the choice of this threshold directly influences a detector's performance. Thus, comparing the performance of a detector in terms of calibration or accuracy over different operating thresholds as well as with different detectors is not straightforward, as illustrated in Fig. 1. + +We assert that a framework for joint evaluation should follow certain basic principles. Firstly, the detectors should be evaluated on a thresholded set of detections to align with their practical usage. While doing so, the evaluation framework will require a principled model-dependent threshold selection mechanism, as the confidence distribution of each detector can differ significantly [47]. Secondly, the calibration should enforce the confidence scores to provide unambiguous and fine-grained information about the detection quality. For example, if the confidence score represents the localisation quality of a detection, this provides more fine-grained information than representing whether the object is detected or not. Thirdly, the datasets should be properly-designed for evaluation. That is, the training, validation (val.) and in-distribution (ID) test splits should be sampled from the same underlying distribution, and additionally, the domain-shifted test splits — which are crucial for safety-critical applications — should + +Table 1: Principles of joint performance evaluation of object detectors in terms of accuracy and calibration, and whether existing evaluation approaches violate them. + +
Principles of Joint EvaluationD-ECE-style [30, 41–43, 52]LaECE-style [48]CE-style [55]Ours
Model-dependent threshold selectionX✓X✓
Unambiguous & fine-grained confidence scoresXXX✓
Properly-designed datasetsXXX✓
Properly-trained detectors & calibratorsX✓X✓
+ +be included. Finally, baseline detectors and calibration methods must be trained properly, as otherwise the evaluation might provide misleading conclusions. + +There are three approaches for jointly evaluating accuracy and calibration: + +$\circ$ D-CEC-style [30,41-43,52] thresholds the detections commonly from a confidence of 0.30 to compute Detection Expected Calibration Error (D-CEC) and use top-100 detections from each image for Average Precision (AP), +$\circ$ LaECE-style [48] enforces the detectors to be thresholded properly, and combine Localisation-aware Expected Calibration Error (LaECE) with LRP [47], +- CE-style [55] thresholds the detections from a confidence score of 0.50 to obtain Calibration Error (CE) and AP. + +As summarized in Tab. 1, these evaluations do not adhere to the basic principles mentioned above. To exemplify, D-CEE-style evaluation — the most common evaluation approach [30, 41–43, 52] — uses different operating thresholds for calibration and accuracy, which does not align well with the practical usage of detectors. Also, using a fixed threshold (i.e., 0.30) for all detectors artificially promotes certain detectors. To illustrate, while D-CEE-style evaluation ranks the green detector as the worst in Fig. 1(a), the green one yields the best D-CEE at 0.70. Besides, as shown in Fig. 1(b), AP is maximized at the confidence of 0 (leading to too many detections with low confidence) for all the detectors, and thus AP cannot be used to obtain a proper operating threshold [47, 48]. In terms of conveying fine-grained information, D-CEE aims to align confidence with the precision only, which effectively ignores the localisation quality of the detections, a crucial performance aspect of object detection. Finally, this type of evaluation also has limitations in terms of dataset splits and the chosen baselines as we explore in Sec. 3. + +Having proper baseline calibration methods is also essential to monitor the progress in the field. Recently proposed train-time calibration methods commonly employ an auxiliary loss term to regularize the confidence scores during training [30,41-43,52]. Such methods are shown to be effective against the Temperature Scaling (TS) [15], which is used as the only post-hoc calibration baseline. Post-hoc calibrators are obtained on a held-out val. set, and hence can easily be applied to any off-the-shelf detector. Despite their potential advantages, unlike for classification [15,20,23,38,56,62,69], post-hoc calibration methods have not been explored for object detection sufficiently [30,48]. + +In this paper, we introduce a joint evaluation framework which respects the aforementioned principles (Tab. 1), and thus address the critical drawbacks of existing evaluation approaches. That is, we first define $\mathrm{LaECE}_0$ and $\mathrm{LaACE}_0$ , + +as novel calibration errors, each of which aims to align the detection confidence scores with their localisation qualities. Thus, the detectors respecting $\mathrm{LaECE}_0$ and $\mathrm{LaACE}_0$ provide informative confidence estimates about their behaviours. We measure accuracy using LRP [45,47], which requires a proper combination of false-positive (FP), false-negative (FN) and localisation errors. Thereby requiring the detectors to be properly-thresholded as shown by the bell-like curves in Fig. 1(d). Also, we design three datasets with different characteristics, and introduce Platt Scaling (PS) as well as Isotonic Regression (IR) as highly effective post-hoc calibrators tailored to object detection. Our main contributions are: + +We identify various quirks and assumptions in state-of-the-art (SOTA) methods in quantifying miscalibration of object detectors and show that they, if not treated properly, can provide misleading conclusions. + +We introduce a framework for joint evaluation consisting of properly-designed datasets, evaluation measures tailored to practical usage of object detectors and baseline post-hoc calibration methods. We show that our framework addresses the drawbacks of existing approaches. + +In contrast to the literature, we show that, if designed properly, post-hoc calibrators can significantly outperform the SOTA training time calibration methods. To illustrate, on the common COCO benchmark, D-DETR with our IR calibrator outperforms the SOTA Cal-DETR [43] significantly: (i) by more than 7 points in terms of D-ECE and (ii) $\sim 4$ points in terms of our challenging LaECE $_0$ . + +# 2 Background and Notation + +Object Detectors and Evaluating their Accuracy Denoting the set of $M$ objects in an image $X$ by $\{b_i, c_i\}^M$ where $b_i \in \mathbb{R}^4$ is a bounding box and $c_i \in \{1, \ldots, K\}$ is its class; an object detector produces the bounding box $\hat{b}_i$ , the class label $\hat{c}_i$ and the confidence score $\hat{p}_i$ for the objects in $X$ , i.e., $f(X) = \{\hat{c}_i, \hat{b}_i, \hat{p}_i\}^N$ with $N$ being the number of predictions. During evaluation, each detection is first labelled as a true-positive (TP) or a FP using a matching function $\psi(\cdot)$ relying on an Intersection-over-Union (IoU) threshold $\tau$ to validate TPs. We assume $\psi(i)$ returns the index of the object that a TP $i$ matches to; else $i$ is a FP and $\psi(i) = -1$ . Then, AP [11, 16, 33], the common accuracy measure, corresponds to the area under the Precision Recall (PR) curve. Though widely-used, AP has been criticized recently from different aspects [5, 45, 47, 50, 58]. To illustrate, AP is maximized when the number of detections increases [48] as shown in Fig. 1(b). Therefore, AP does not help choosing an operating threshold, which is critical for practical deployment. As an alternative, LRP [45, 47] combines the numbers of TP, FP, FN with the localisation error of the detections, which are denoted by $\mathrm{N_{TP}}$ , $\mathrm{N_{FP}}$ , $\mathrm{N_{FN}}$ and $\mathcal{E}_{loc}(i) \in [0, 1]$ respectively: + +$$ +\mathrm {L R P} = \frac {1}{\mathrm {N} _ {\mathrm {F P}} + \mathrm {N} _ {\mathrm {F N}} + \mathrm {N} _ {\mathrm {T P}}} \left(\mathrm {N} _ {\mathrm {F P}} + \mathrm {N} _ {\mathrm {F N}} + \sum_ {\psi (i) > 0} \mathcal {E} _ {l o c} (i)\right). \tag {1} +$$ + +Unlike AP, LRP requires the detection set to be thresholded properly as both FPs and FNs are penalized in Eq. (1). + +Evaluating the Calibration of Object Detectors The alignment of accuracy and confidence of a model, termed calibration, is extensively studied for classification [8, 15, 27, 40, 44, 63]. That is, a classifier is calibrated if its accuracy is $p$ for the predictions with confidence of $p$ for all $p \in [0,1]$ . For object detection, [30] extends this definition to enforce that the confidence matches the precision of the detector, $\mathbb{P}(\hat{c}_i = c_i|\hat{p}_i) = \hat{p}_i, \forall \hat{p}_i \in [0,1]$ , where $\mathbb{P}(\hat{c}_i = c_i|\hat{p}_i)$ is the precision. Then, discretizing the confidence space into $J$ bins, D-ECE is + +$$ +\mathrm {D} - \mathrm {E C E} = \sum_ {j = 1} ^ {J} \frac {\left| \hat {\mathcal {D}} _ {j} \right|}{\left| \hat {\mathcal {D}} \right|} \left| \bar {p} _ {j} - \operatorname {p r e c i s i o n} (j) \right|, \tag {2} +$$ + +where $\hat{D}$ and $\hat{D}_j$ are the set of all detections and the detections in the $j$ -th bin, and $\bar{p}_j$ and precision $(j)$ are the average confidence and the precision of the detections in the $j$ -th bin. Alternatively, considering that object detection is a joint task of classification and localisation, LaECE [48] aims to match the confidence with the product of precision and average IoU of TPs. Also, to prevent certain classes from dominating the error, LaECE is introduced as a class-wise measure. Using superscript $c$ to refer to each class and $\mathrm{IoU}^c (j)$ as the average IoU of $\hat{D}_j^c$ , LaECE is defined as: + +$$ +\mathrm {L a E C E} = \frac {1}{K} \sum_ {c = 1} ^ {K} \sum_ {j = 1} ^ {J} \frac {\left| \hat {\mathcal {D}} _ {j} ^ {c} \right|}{\left| \hat {\mathcal {D}} ^ {c} \right|} \left| \bar {p} _ {j} ^ {c} - \operatorname {p r e c i s i o n} ^ {c} (j) \times \mathrm {I o U} ^ {c} (j) \right|. \tag {3} +$$ + +Calibration Methods in Object Detection The existing methods for calibrating object detectors can be split into two groups: (1) Training-time calibration approaches [41-43,52,55] regularize the model to yield calibrated confidence scores during training, which is generally achieved by an additive auxiliary loss. (2) Post-hoc calibration methods use a held-out val. set to fit a calibration function that maps the predicted confidence to the calibrated confidence. Specifically, TS [15] is the only method considered as a baseline for recent training time methods [41,42,52,55]. As an alternative, IR [1-3,66] is used within a limited scope for a specific task called Self-aware Object Detection [48]. Furthermore, its effectiveness neither on a wide range of detectors nor against existing training-time calibration approaches has yet been investigated. + +# 3 Analysis of the Common D-ECE-style Evaluation + +D-ECE-style evaluation is the most common evaluation approach adopted by several methods [30,41-43,52]. For that reason, here we provide a comprehensive analysis of this evaluation approach and analyse the LaECE-style and CE-style evaluations in App. B. Our analyses below are based on the principles outlined in Sec. 1 and Tab. 1, and show that all approaches have notable drawbacks. + +1. Model-dependent threshold selection. As AP is obtained using the top-100 detections and D-ECE is computed on detections thresholded above 0.30, D-ECE-style evaluation uses two different detection sets. This inconsistency is not reflective of how detectors are used in practice. Also, we observe that a fixed threshold of 0.30 for evaluating the calibration induces a bias for certain detectors. To illustrate, we compare the performance of different calibration methods over different thresholds in Fig. 2, where CalDETR [43] performs the best only for the threshold 0.30 and the post-hoc TS significantly outperforms it on all other thresholds. Therefore, this method of evaluation is sensitive to the choice of threshold, leading to ambiguity on the best performing method. + +![](images/55857e30c4cb45b7f556d6325a5db863f4533c87ecfbd07c8a0e872e2356e10b.jpg) +Fig.2: Comparison of calibration methods in terms of D-ECE on COCO mini-test using D-DETR [71]. Post-hoc TS and IR calibrators are obtained on a subset of Objects365 [60] as in D-ECE-style evaluation. + +2. Fine-grained confidence scores. Manipulating Eq. (2), we show in App. B that D-ECE for the $j$ -th bin can be expressed as, + +$$ +\left| \sum_ {\hat {b} _ {i} \in \hat {\mathcal {D}} _ {j}, \psi (i) > 0} (\hat {p} _ {i} - 1) + \sum_ {\hat {b} _ {i} \in \hat {\mathcal {D}} _ {j}, \psi (i) = - 1} \hat {p} _ {i} \right|. \tag {4} +$$ + +Eq. (4) implies that D-ECE is minimized when the confidence scores $\hat{p}_i$ of TPs are 1 and those of FPs are 0, which is also how the prediction-target pairs are usually constructed to train post-hoc TS [30, 41-43, 52]. Even if the detector is perfectly calibrated for these binary targets, the confidence scores do not provide information about localisation quality as illustrated by binary-valued $\hat{p}_{\mathrm{D - ECE}}$ for both detections in Fig. 3(b). Also, Popordanoska et al. [55] utilise D-ECE in a COCO-style manner, that is they average D-ECE over different TP validation IoU thresholds similar to COCO-style AP [33]. However, we observe that this way of using D-ECE can promote ambiguous confidence scores. As an example, given two IoU thresholds $\tau_{1}$ and $\tau_{2}$ , a detection $\hat{b}_i$ with $\tau_{1} \leq \mathrm{IoU}(\hat{b}_{i}, b_{\psi(i)}) < \tau_{2}$ is a TP for $\tau_{1}$ but a FP for $\tau_{2}$ . Thus, given Eq. (4), it follows that $\hat{b}_i$ has contradictory confidence targets for $\tau_{1}$ and $\tau_{2}$ . This is illustrated in Fig. 3(d) in which D-ECE $_{\mathrm{C}}$ (red line) remains constant regardless of the confidence. Thus, using D-ECE (or another calibration measure) in this way should be avoided. + +3. Properly-designed datasets. In the literature, the val. set to obtain the post-hoc calibrator is typically taken from a different dataset than the ID dataset [41-43, 52]. Specifically, the post-hoc calibrators are obtained on a subset from Objects365 [60] and BDD100K [65] for the models trained with COCO [33] and Cityscapes [9] respectively. Hence, as expected, a different dataset inevitably induces domain shift, affecting the performance of the post-hoc calibrator [51]. To show that, following existing approaches, we obtain an IR calibrator [48] on + +![](images/73f300fb1e05cc44a585f3248d032376a9711b3637f925edf6cd8be00ca2550a.jpg) +(a) Uncalibrated detections + +![](images/8889571d3ed27f1a9cbbb1a4416fa51054e95fe999997570e6adc0ece4628325.jpg) +(b) Calibrated detections + +![](images/3da4bf1e3647babc0688f30aa0ac6043dd45637f257b9ba2643a5fdec3929738.jpg) +(c) Calibrated detections (Ours) + +![](images/e47267df99eec3bccc9b443e6ca1768bfe5710836689ea57d2cd96169249157e.jpg) +(d) TP for $\tau = 0.5$ + +![](images/9475da39d774a2523d85c0b4448391780aacb3c9b874bf94e453437a3dc8ed3c.jpg) +(e) FP for $\tau = 0.5$ , TP for $\tau = 0$ + +![](images/bd1ccdfbff0d84bf901eec177dc03d33597e7a65b87a5fb835f788a5b9102f29.jpg) +(f) FP for $\tau = 0$ +Fig. 3: A pictorial comparison of the different calibration errors. (a) Uncalibrated detections of D-DETR on an image from [65]. The detections on the left and right have IoUs of 0.74 and 0.48 with the objects. (b) Calibrated detections in terms of D-ECE and LaECE using $\tau = 0.50$ , and D-ECEc, COCO-style D-ECE as in [55]. $\mathrm{D - ECE_C} = ?$ as calibration error does not have a global minimum as shown in (d). (c) Calibrated detections in terms of LaECEo and LaACEo in which confidence matches IoU. (d-f) Calibration errors for different types of detections, for which LaACEo behave the same as LaECEo, hence excluded for clarity. App. B presents the details. + +Objects365 and compare it with the one obtained on the ID val. set in terms of D-ECE-style evaluation. Tab. 2 shows that the latter IR now outperforms (i) the former one by $\sim 11$ D-ECE and (ii) SOTA Cal-DETR [43] by 7.4 D-ECE, showing the importance of dataset design for proper evaluation. + +4. Properly-trained detectors and calibrators. Though Cityscapes is commonly used in the literature [41-43,52], the models trained on this dataset follow COCO-style training. Specifically, D-DETR [71] was trained on Cityscapes for only 50 epochs though the training set of Cityscapes is $\sim 40\times$ smaller than that of COCO (3K vs. $\sim 118\mathrm{K}$ ). We now tailor the training of D-DETR for Cityscapes by (i) $4\times$ longer training considering the smaller training set and (ii) increasing the training image scale considering the original resolution following [7,9]. We kept all other hyperparameters as they are for both Cal-DETR and D-DETR, and App. B presents the details. Tab. 3 shows that, once trained in this setting, D-DETR is more accurate and calibrated than Cal-DETR. + +# 4 A Framework for Joint Evaluation of Object Detectors + +We now present our evaluation approach that respects to the principles in Sec. 1. + +Table 2: Effect of using domain-shifted val. set on IR calibrator. Results are reported on COCO-minitest. Val. set is N/A for uncalibrated D-DETR and training time calibration method Cal-DETR. + +
MethodVal setD-ECEAP ↑
D-DETRN/A12.844.1
Cal-DETRN/A8.744.4
IRObjects36514.244.1
IRCOCO1.3 (+7.4)44.1
+ +Table 3: COCO training settings are commonly adopted while training DDETR on Cityscapes. When trained with larger images and longer, DETR performs slightly better than Cal-DETR. + +
MethodTraining StyleD-ECEAP↑
D-DETR [43]COCO13.826.8
Cal-DETR [43]COCO8.428.4
Cal-DETRCityscapes4.034.9
D-DETRCityscapes2.936.1
+ +# 4.1 Towards Fine-grained Calibrated Detection Confidence Scores + +Calibration refers to the alignment of accuracy and confidence of a model. Therefore, for an object detector to be calibrated, its confidence should respect both classification and localisation accuracy. We discussed in Sec. 3 that D-ECE, as the common calibration measure, only considers the precision of a detector, thereby ignoring its localisation performance (Eq. (2)). LaECE [48], defined in Eq. (3) as an alternative to D-ECE, enforces the confidence scores to represent the product of precision and average IoU of TPs. Thus, LaECE considers IoUs of only TPs, and effectively ignores the localisation qualities of detections if their IoU is less than the TP validation threshold $\tau > 0$ . We assert that this selection mechanism based on IoU unnecessarily limits the information conveyed by the confidence score. We illustrate this on the right car in Fig. 3(b) for which LaECE requires a target confidence of 0 ( $\hat{p}_{\mathrm{LaECE}} = 0$ ) as its IoU is less than $\tau = 0.50$ . However, instead of conveying a 0 confidence and implying no detection, representing its IoU by $\hat{p}_i$ provides additional information. Hence, we propose using $\tau = 0$ , in which case the calibration criterion of LaECE reduces to, + +$$ +\mathbb {E} _ {\hat {b} _ {i} \in B _ {i} (\hat {p} _ {i})} [ \mathrm {I o U} (\hat {b} _ {i}, b _ {\psi (i)}) ] = \hat {p} _ {i}, \forall \hat {p} _ {i} \in [ 0, 1 ], \tag {5} +$$ + +where we define $\mathrm{IoU}(\hat{b}_i, b_{\psi(i)}) = 0$ for FPs when $\tau = 0$ , $B_i(\hat{p}_i)$ is the set of boxes with the confidence of $\hat{p}_i$ and $b_{\psi(i)}$ is the ground-truth box that $\hat{b}_i$ matches with. To derive the calibration error for Eq. (5), we follow LaECE by using $J = 25$ equally-spaced bins and averaging over class-wise errors and define, + +$$ +\mathrm {L a E C E} _ {0} = \frac {1}{K} \sum_ {c = 1} ^ {K} \sum_ {j = 1} ^ {J} \frac {\left| \hat {\mathcal {D}} _ {j} ^ {c} \right|}{\left| \hat {\mathcal {D}} ^ {c} \right|} \left| \bar {p} _ {j} ^ {c} - \operatorname {I o} \mathrm {U} ^ {c} (j) \right|, \tag {6} +$$ + +where $\hat{\mathcal{D}}^c$ and $\hat{\mathcal{D}}_j^c$ denote the set of all detections and those in $j$ th bin respectively, $\bar{p}_j^c$ is the average confidence score and $\mathrm{IoU}^c (j)$ is the average IoU of detections in the $j$ -th bin for class $c$ , and the subscript 0 refers to the chosen $\tau$ which is 0. Furthermore, similar to the classification literature [40, 44], we define Localisation-aware Adaptive Calibration Error (LaACE) using an adaptive binning approach in which the number of detections in each bin is equal. In order + +Table 4: Datasets for evaluating object detection and instance segmentation methods. + +
TypeTrain setVal setID test setDomain-shifted test set
Common ObjectsCOCO trainCOCO minivalCOCO minitestCOCO minitest-C, Obj45K
Autonomous DrivingCS trainCS minivalCS minitestCS minitest-C, Foggy-CS
Long-tailed ObjectsLVIS trainLVIS minivalLVIS minitestLVIS minitest-C
+ +to capture the model behaviour precisely, we adopt the extreme case in which each bin has only one detection, resulting in an easy-to-interpret measure which corresponds to the mean absolute error between the confidence and the IoU, + +$$ +\mathrm {L a A C E} _ {0} = \frac {1}{K} \sum_ {c = 1} ^ {K} \sum_ {i = 1} ^ {| \hat {\mathcal {D}} ^ {c} |} \frac {1}{| \hat {\mathcal {D}} ^ {c} |} \left| \hat {p} _ {i} - \operatorname {I o U} \left(\hat {b} _ {i}, b _ {\psi (i)}\right) \right|. \tag {7} +$$ + +As we show in App. C, $\mathrm{LaECE}_0$ and $\mathrm{LaACE}_0$ are both minimized when $\hat{p}_i = \mathrm{IoU}(\hat{b}_i, b_{\psi(i)})$ for all detections, which is also a necessary condition for $\mathrm{LaACE}_0$ . Hence, as illustrated on the right car in Fig. 3(c) and (e), $\mathrm{LaECE}_0$ and $\mathrm{LaACE}_0$ requires conveying more fine-grained information compared to other measures. + +# 4.2 Model-dependent Thresholding for Proper Joint Evaluation + +In practice, object detectors employ an operating threshold to preferably output only TPs with high recall. However, AP as the common performance measure does not enable cross-validating such a threshold as it is maximized when the recall is maximized despite a drop in precision [47, 48]. This can be observed in Fig. 1(b) where AP consistently decreases as the confidence threshold increases. Alternatively, LRP (Eq. 1) prefers detectors with high precision, recall and low localisation error as illustrated by the bell-like curves in Fig. 1(d). This is because, unlike AP, LRP severely penalises detectors with low recall or precision, making it a perfect fit for our framework. As a result, we consider $\mathrm{LaECE}_0$ and $\mathrm{LaACE}_0$ with LRP and require each model to be thresholded properly. + +# 4.3 Properly-designed Datasets + +We curate three datasets summarized in Tab. 4: (i) COCO [33] including common daily objects; (ii) Cityscapes [9] with autonomous driving scenes; and (iii) LVIS [16], a challenging dataset focusing on the calibration of long-tailed detection. For each dataset, we ensure that train, val. and ID test sets are sampled from the same distribution, and include domain-shifted test sets. As these datasets do not have public labels for test sets, we randomly split their val. sets into two as minival and minitest similar to [17,48,49]. In such a way, we provide ID val. sets to enable obtaining post-hoc calibrators and the operating thresholds properly. For domain-shifted test sets, we apply common corruptions [21] to the ID test sets, and include Obj45K [48,60] and Foggy Cityscapes [59] as more realistic shifts. Our datasets also have mask annotations and hence they can be used to evaluate instance segmentation methods. App. C includes further details. + +# 4.4 Baseline Post-hoc Calibrators Tailored to Object Detection + +It is essential to develop post-hoc calibration methods tailored to object detection, which has certain differences from the classification task. Existing methods [41-43,52,55] use only TS as a baseline without considering the peculiarities of detection. Specifically, a single temperature parameter $T$ is learned to adjust the predictive distribution while the confidence $\hat{p}_i$ is commonly assumed to be a Bernoulli random variable [30]. However, PS, fitting both a scale and a shift parameter, is the widely-accepted calibration approach when the underlying distribution is Bernoulli [15,54]. Also, how to construct a useful subset of the detections to train the post-hoc calibrators has not been explored. To address these shortcomings, we present (i) Platt Scaling in which the bias term makes a notable difference in the performance, and (ii) Isotonic Regression by modeling the calibration as a regression problem. Before introducing them, we now present an overview on how we determine the set of detections to train the calibrators. + +Overview We obtain post-hoc calibrators on a held out val. set using the detections that are similar to those seen at inference to prevent low-scoring detections from dominating the training of the calibrator. To do so, we crossvalidate a calibration threshold $\bar{u}^c$ for each class $c$ and train a class-specific calibrator $\zeta^c: [0,1] \to [0,1]$ using the detections with higher scores than $\bar{u}^c$ . Still, as $\zeta^c(\cdot)$ changes the confidence scores, we need another threshold $\bar{v}^c$ , as the operating threshold, to remove the redundant detections after calibration. Following the accuracy measure, we cross-validate $\bar{u}^c$ and $\bar{v}^c$ using LRP. As for inference time, for the $i$ -th detection $(\hat{p}_i, \hat{b}_i, \hat{c}_i)$ , if $\hat{p}_i \geq \bar{u}^{\hat{c}_i}$ , it survives to the calibrator and then $\hat{p}_i^{cal} = \zeta^{\hat{c}_i}(\hat{p}_i)$ . Finally, if $\hat{p}_i^{cal} \geq \bar{v}^{\hat{c}_i}$ , the $i$ -th detection is an output of the detector. Alg. A.1 and A.2 provide the details. For $\zeta^c(\cdot)$ , we prefer monotonically increasing functions in order not to affect the ranking of the detections significantly and to keep their accuracy as we detail below. + +Distribution Calibration via Platt Scaling Assuming that $\hat{p}_i$ is sampled from Bernoulli distribution $\mathcal{B}(\cdot)$ , we aim to minimize the Negative Log-Likelihood (NLL) of the predictions on the target distribution $\mathcal{B}(\mathrm{IoU}(\hat{b}_i, b_{\psi(i)}))$ using PS [54]. Accordingly, we recover the logits, and then shift and scale the logits to obtain the calibrated probabilities $\hat{p}_i^{cal}$ , + +$$ +\hat {p} _ {i} ^ {\text {c a l}} = \sigma \left(a \sigma^ {- 1} \left(\hat {p} _ {i}\right) + b\right), \tag {8} +$$ + +where $\sigma(\cdot)$ is the sigmoid and $\sigma^{-1}(\cdot)$ is its inverse, as well as $a \geq 0$ and $b$ are the learnable parameters. We derive the NLL for the $i$ th detection in App. C as + +$$ +- \left(\operatorname {I o U} \left(\hat {b} _ {i}, b _ {\psi (i)}\right) \log \left(\hat {p} _ {i} ^ {c a l}\right) + \left(1 - \operatorname {I o U} \left(\hat {b} _ {i}, b _ {\psi (i)}\right)\right) \log \left(1 - \hat {p} _ {i} ^ {c a l}\right)\right). \tag {9} +$$ + +Please note that Eq. (9), which is in fact the cross-entropy loss, is minimized if $\hat{p}_i^{cal} = \mathrm{IoU}(\hat{b}_i, b_{\psi(i)})$ when $\mathrm{LaECE}_0$ and $\mathrm{LaACE}_0$ are minimized. We optimize Eq. (9) via the second-order optimization strategy L-BFGS [35] following [30]. + +Confidence Calibration via Isotonic Regression As an alternative perspective, $\hat{p}_i$ can also be directly calibrated by modelling the calibration as a regression task. To do so, we construct the prediction-target pairs $(\{\hat{p}_i, \mathrm{IoU}(\hat{b}_i, b_{\psi(i)})\})$ on the held-out val. set and then fit an IR model using scikit-learn [53]. + +Table 5: Comparison with SOTA methods in terms of other evaluation measures on COCO [33]. LRP is reported on LRP-optimal thresholds obtained on val. set. AP is reported on top-100 detections. $\tau$ is taken as 0.50. All measures are lower-better, except AP. Bold: the best, underlined: second best. PS: Platt Scaling, IR: Isotonic Regression. + +
Cal. +TypeMethodCalibration (thr. 0.30)Calibration (LRP thr.)Accuracy
D-ECELaECED-ECELaECELRPAP↑
Uncal.D-DETR [71]12.813.215.012.166.344.1
Training +TimeMbLS [34]15.616.318.715.865.944.3
MDCA [19]12.213.514.312.666.443.8
TCD [42]12.413.114.412.366.644.0
BPC [41]9.813.111.412.366.843.6
Cal-DETR [43]8.712.99.711.866.044.4
Post-hoc +(Ours)PS for D-ECE0.9(+7.8)16.32.4(+7.3)15.866.344.1
PS for LaECE11.011.5(+1.4)9.410.1(+1.7)66.344.1
IR for D-ECE1.3(+7.4)15.72.6(+7.1)15.366.244.1
IR for LaECE10.28.9(+4.0)9.38.2(+3.6)66.343.7
+ +Adapting Our Approach to Different Calibration Objectives Until now, we considered post-hoc calibrators for $\mathrm{LaECE_0}$ and $\mathrm{LaACE_0}$ though in practice different measures can be preferred. Our post-hoc calibrators can easily be adapted for such cases by considering the dataset design and optimisation criterion. To illustrate, for D-ECE-style evaluation, the calibration dataset is to be class-agnostic where the detections are thresholded from 0.30 with prediction-target pairs for IR as $(\{\hat{p}_i,0\})$ and $(\{\hat{p}_i,1\})$ for FPs and TPs respectively. + +# 5 Experimental Evaluation + +We now show that our post-hoc calibration approaches consistently outperform training time calibration methods by significant margins (Sec. 5.1) and that they generalize to any detector and can thus be used as a strong baseline (Sec. 5.2). + +# 5.1 Comparing Our Baselines with SOTA Calibration Methods + +Here, we compare PS and IR with recent training-time calibration methods considering various evaluation approaches. As these training-time methods mostly rely on D-DETR, we also use D-DETR with ResNet-50 [18]. We obtain the detectors of training time approaches trained with COCO dataset from their official repositories, whereas we incorporate Cityscapes into their official repositories and train them using the recommended setting in Tab. 3. + +Comparison on Other Evaluation Approaches Before moving on to our evaluation approach, we first show that our PS and IR outperform all existing training time methods on existing evaluation approaches. For that, we consider D-ECE and the LaECE from $\tau = 0.5$ by including two different evaluation settings for each: (i) the detection set is obtained from the fixed threshold of 0.30 following the convention [30,41-43], and (ii) the operating thresholds are cross-validated using LRP. Following their standard usage, we use 10 and 25 + +Table 6: Comparison with SOTA calibration methods on Common Objects using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration. + +
Calibration TypeMethodCOCO minitest (ID)COCO-C (Domain Shift)Obj45K (Domain Shift)
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
UncalibratedD-DETR [71]12.727.157.314.628.771.516.435.872.0
Training-timeMbLS [34]16.530.356.816.831.171.817.337.171.6
MDCA [19]13.127.257.514.528.771.816.635.672.2
TCD [42]13.026.757.614.628.371.916.335.571.7
BPC [41]12.425.557.714.127.172.117.334.572.0
Cal-DETR [43]11.624.656.213.826.470.618.835.371.1
Post-hoc (Ours)Platt Scaling9.623.557.312.825.671.517.033.772.0
Isotonic Regression7.723.157.210.725.371.517.233.372.0
(+3.9)(+1.5)(+3.1)(+1.1)(-0.9)(+1.2)
+ +Table 7: Comparison with SOTA on Autonomous Driving using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration. + +
Calibration TypeMethodCityscapes minitest (ID) LaECE0|LaACE0|LRPCityscapes-C (Domain Shift) LaECE0|LaACE0|LRPFoggy Cityscapes (Domain Shift) LaECE0|LaACE0|LRP
UncalibratedD-DETR [71]20.326.057.221.425.680.218.522.369.4
Training-timeTCD [42]16.831.759.223.232.481.624.433.871.6
BPC [41]23.831.864.928.133.383.724.730.973.8
Cal-DETR [43]21.325.356.923.026.480.820.023.271.0
Post-hoc (Ours)Platt Scaling9.623.357.217.726.280.211.321.669.4
Isotonic Regression9.0(+7.8)23.7(+1.6)56.816.4(+5.0)25.8(-0.2)80.510.0(+8.5)21.2(+1.1)69.5
+ +bins to compute D-ECE and LaECE respectively. We optimize PS and IR by considering the calibration objective as described in Sec. 4.4. Tab. 5 shows that PS and IR outperform SOTA Cal-DETR significantly by more than 7 D-ECE and up to 4 LaECE on COCO minitest. Please note that all previous approaches are optimized for $D$ -ECE thresholded from 0.30, in terms of which our PS yields only 0.9 $D$ -ECE improving the SOTA by 7.8. Tab. 5 also suggests that post-hoc calibrators perform the best when the calibration objective is aligned with the measure. App. D shows that our observations generalize to Cityscapes. + +Common Objects Setting We now evaluate detectors using our evaluation approach. Tab. 6 shows that IR and PS share the top-2 entries on almost all test subsets by preserving the accuracy (LRP) of D-DETR. Specifically, our gains on ID set and COCO-C are significant, where IR outperforms Cal-DETR by around $3 - 4\mathrm{LaECE}_0$ and $1.0 - 1.5\mathrm{LaACE}_0$ . As for Obj45K, the challenging test set with natural shift, IR and PS improve $\mathrm{LaACE}_0$ but perform slightly worse in terms of $\mathrm{LaECE}_0$ . This is an expected drawback of post-hoc approaches when the domain shift is large as they are trained only with ID val. set [51]. + +Table 8: Comparison with TS using D-DETR. Bold: the best, underlined: second best. X : domain-shifted val. set is used to obtain thresholds and calibrators, decreasing the accuracy (red font). Bias term only exists for PS ( $b$ in Eq. (8)), thus $\mathrm{N}/\mathrm{A}$ for IR. + +
MethodAblations on DatasetAblations on ModelCOCO minitestCityscapes minitest
ID Val. SetThresholdClass-wiseBias TermLaECE0LaACE0LRPLaECE0LaACE0LRP
Temperature Scaling (Current Baseline)✘12.320.861.521.025.960.3
✓12.523.157.320.926.357.2
Ablations on Temperature Scaling✓✓11.324.857.313.325.557.2
✓✓12.422.957.323.127.557.2
✓✓✓10.624.257.312.724.657.2
Platt Scaling (Ours) Isotonic Regression (Ours)✓✓✓✓9.623.557.39.623.357.2
✓✓✓N/A7.723.157.29.023.756.8
+ +Table 9: Calibrating and evaluating different object detectors. We use Common Objects setting and report the results on COCO minitest. * denotes the detectors in Fig. 1. Bold: the best, underlined: second best for calibration. + +
TypeDetectorBackboneUncalibratedPlatt ScalingIsotonic RegressionAP ↑
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
One-StagePAA [25]*R5015.928.159.79.724.359.77.723.859.743.2
ATSS [70]*R5019.134.059.510.324.759.58.524.159.543.1
GFL [32]R5013.728.559.310.324.559.38.324.059.343.0
VFNet [68]R5013.925.857.710.725.157.78.324.657.744.8
Two-StageFaster R-CNN [57]*R5027.029.960.410.423.860.48.623.560.440.1
RS R-CNN [46]*R5019.728.958.710.223.558.78.123.058.842.4
DETR-likeD-DETR [71]*R5012.727.157.39.623.557.37.723.157.244.1
UP-DETR [10]R5034.435.255.810.022.655.88.222.255.942.9
DINO [67]R5013.626.953.610.623.553.68.922.853.650.4
OVODGLIP [31]Swin-T13.025.349.09.222.449.07.721.849.055.7
G. DINO [36]Swin-T13.827.546.98.921.946.97.721.347.058.3
SOTACo-DETR [72]Swin-L10.823.041.58.620.241.56.719.341.664.5
EVA [12]ViT(EVA)17.421.241.28.620.341.27.119.941.264.5
MoCaE [49]N/A10.621.440.78.920.440.77.319.940.765.0
+ +Autonomous Driving Setting Tab. 7 shows that our approaches consistently outperform all training time calibration approaches on this setting as well. Specifically, our gains are very significant ranging between 5.0-8.5 $\mathrm{LaECE}_0$ compared to the SOTA Cal-DETR, further presenting the efficacy of our approaches. + +Comparison with Existing Temperature Scaling Baseline and Ablations Tab. 8 compares TS for different design choices as well as with our PS and IR. Please note that $\pmb{x}$ corresponds to the baseline setting used in the recent approaches [41-43,52] that employ Objects365 [60] and BDD100K [60] as domain-shifted val. sets for obtaining the calibrator. Due to this domain shift, the accuracy of TS degrades by up to 4 LRP, in red font, as the operating thresholds obtained on these val. sets do not generalize to the ID set; showing that it is crucial to use an ID val set. In ablations, thresholding the detections and class-wise calibrators generally improves the performance of TS and a more notable gain is observed once the bias term is used in PS. Our PS outperforms TS baseline obtained on ID val. set by $\sim 3$ LaECE $_0$ on COCO and 11.4 LaECE $_0$ on Cityscapes. Finally, IR performs on par or better compared to PS. + +# 5.2 Calibrating and Evaluating Different Detection Methods + +Another benefit of our post-hoc calibrators is that they generalize to any object detector, thus they can be reliably used as baselines. To show that, we calibrate 14 different detectors with a diverse set of architectures using PS and IR in Tab. 9. The results suggest that both IR and PS perform better than uncalibrated detectors. IR consistently outperforms PS as it fits multiple piece-wise linear functions while PS learns only two parameters (Eq. (8)). Specifically, IR + +![](images/1f243872d76195e0be22a46a33633ca48ce9818721b25c4d71b2626689a7532b.jpg) +(a) Uncalibrated +Fig. 4: The reliability diagrams of UP-DETR. + +![](images/ce0488e4f241f90660917f7120d337b9bdd7e2f4e0a03927de10cc76a0fa2a30.jpg) +(b) Calibrated by IR + +decreases the range of $\mathrm{LaECE_0}$ from 10.6-34.4 to 6.7-8.9 on COCO minitest by preserving the accuracy, making it a solid baseline. For further insights, Fig. 4 provides the reliability diagrams of the overconfident UP-DETR, of which IR significantly improves its calibration. As for SOTA, MoCaE performs the best in terms of accuracy with 40.7 LRP while Co-DETR has the best calibration with $6.7\mathrm{LaECE}_0$ and $19.3\mathrm{LaACE}_0$ , which the future work should aim to surpass. App. D.3 includes our results on our Long-tailed Objects setting (including object detection and instance segmentation), showing the effectiveness of our post-hoc calibrators on this challenging setting as well. + +The performance measures we use in our evaluation framework are also easily interpretable. For accuracy, LRP is a weighted combination of its FP, FN and localisation error components (App. A), which, as an example, are 10.1, 22.7 and 18.8 respectively for Co-DETR [72] calibrated with IR. Also considering $\mathrm{LaACE}_0 = 19.3$ , one can easily infer that: once deployed with the operating thresholds determined by our framework, Co-DETR finds $77.3\%$ of the objects with $89.9\%$ precision and $81.2\%$ IoU where $19.3\%$ is the mean absolute error of the confidence to represent IoU. We believe these intuitive measures will enable practitioners to make better decision when deploying object detectors. + +# 6 Conclusions + +The progress in a field heavily relies on the evaluation tools and the baselines used. In this paper, we showed that existing evaluation tools for calibration as well as the baseline post-hoc calibrators for object detectors have significant drawbacks. We remedied that by introducing an evaluation framework including baseline post-hoc calibrators tailored to object detection. Our experiments suggested that, once evaluated and designed properly, the post-hoc calibrators significantly outperform all existing training-time calibrators. This implies the need for research to develop better calibration techniques for object detection, for which, we believe, our evaluation framework will be an essential pillar. + +# References + +1. Ayer, M., Brunk, H.D., Ewing, G.M., Reid, W.T., Silverman, E.: An empirical distribution function for sampling with incomplete information. The annals of mathematical statistics pp. 641-647 (1955) +2. Barlow, R.E., Brunk, H.D.: The isotonic regression problem and its dual. Journal of the American Statistical Association 67(337), 140-147 (1972) +3. Best, M.J., Chakravarti, N.: Active set algorithms for isotonic regression; a unifying framework. Mathematical Programming 47(1), 425-439 (1990) +4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2016). https://doi.org/10.1109/icip.2016.7533003, http://dx.doi.org/10.1109/ICIP.2016.7533003 +5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: The IEEE European Conference on Computer Vision (ECCV) (2020) +6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv 1906.07155 (2019) +8. Cheng, J., Vasconcelos, N.: Calibrating deep neural networks by pairwise constraints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +0. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514 +1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) 88(2), 303-338 (2010) +2. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012) +4. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (Nov 2019). https://doi.org/10.1002/rob.21918, http://dx.doi.org/10.1002/rob.21918 + +15. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1321-1330. PMLR (2017) +16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +17. Harakeh, A., Waslander, S.L.: Estimating and evaluating regression predictive uncertainty in deep object detectors. In: International Conference on Learning Representations (ICLR) (2021) +18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +19. Hebbalaguppe, R., Prakash, J., Madan, N., Arora, C.: A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16081-16090 (June 2022) +20. Hekler, A., Brinker, T.J., Buettner, F.: Test time augmentation meets post-hoc calibration: Uncertainty quantification under real-world conditions. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14856-14864 (Jun 2023). https://doi.org/10.1609/aaai.v37i12.26735, https://ojs.aaai.org/index.php/AAAI/article/view/26735 +21. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (ICLR) (2019) +22. Jin, C., Udupa, J.K., Zhao, L., Tong, Y., Odhner, D., Pednekar, G., Nag, S., Lewis, S., Poole, N., Mannikeri, S., Govindasamy, S., Singh, A., Camaratta, J., Owens, S., Torigian, D.A.: Object recognition in medical images via anatomy-guided deep learning. Medical Image Analysis 81, 102527 (2022). https://doi.org/https://doi.org/10.1016/j.media.2022.102527, https://www.sciencedirect.com/science/article/pii/S1361841522001748 +23. Joy, T., Pinto, F., Lim, S.N., Torr, P.H., Dokania, P.K.: Sample-dependent adaptive temperature scaling for improved calibration. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14919-14926 (Jun 2023). https://doi.org/10. 1609/aaii.v37i12.26742, https://ojs.aaii.org/index.php/AAAI/article/view/26742 +24. Karimi, D., Dou, H., Warfield, S.K., Gholipour, A.: Deep learning with noisy labels: Exploring techniques and remedies in medical image analysis. Medical Image Analysis 65, 101759 (2020). https://doi.org/https://doi.org/10.1016/j.media.2020.101759, https://www.sciencedirect.com/science/article/pii/S1361841520301237 +25. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020) +26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023) +27. Kumar, A., Liang, P.S., Ma, T.: Verified uncertainty calibration. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019) + +28. Kumar, N., Verma, R., Anand, D., Zhou, Y., Onder, O.F., Tsougenis, E., Chen, H., Heng, P.A., Li, J., Hu, Z., Wang, Y., Koohbanani, N.A., Jahanifar, M., Tajeddin, N.Z., Gooya, A., Rajpoot, N., Ren, X., Zhou, S., Wang, Q., Shen, D., Yang, C.K., Weng, C.H., Yu, W.H., Yeh, C.Y., Yang, S., Xu, S., Yeung, P.H., Sun, P., Mahbod, A., Schaefer, G., Ellinger, I., Ecker, R., Smedby, O., Wang, C., Chidester, B., Ton, T.V., Tran, M.T., Ma, J., Do, M.N., Graham, S., Vu, Q.D., Kwak, J.T., Gunda, A., Chanduri, R., Hu, C., Zhou, X., Lotfi, D., Safdari, R., Kascenas, A., O'Neil, A., Eschweiler, D., Stegmaier, J., Cui, Y., Yin, B., Chen, K., Tian, X., Gruening, P., Barth, E., Arbel, E., Remer, I., Ben-Dor, A., Sirazitdinova, E., Kohl, M., Braunewell, S., Li, Y., Xie, X., Shen, L., Ma, J., Baksi, K.D., Khan, M.A., Choo, J., Colomer, A., Naranjo, V., Pei, L., Iftekharuddin, K.M., Roy, K., Bhattacharjee, D., Pedraza, A., Bueno, M.G., Devanathan, S., Radhakrishnan, S., Koduganty, P., Wu, Z., Cai, G., Liu, X., Wang, Y., Sethi, A.: A multi-organ nucleus segmentation challenge. IEEE Transactions on Medical Imaging 39(5), 1380–1391 (2020). https://doi.org/10.1109/TMI.2019.2947628 +29. Kumar, N., Verma, R., Sharma, S., Bhargava, S., Vahadane, A., Sethi, A.: A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging 36(7), 1550-1560 (2017). https://doi.org/10.1109/TMI.2017.2677499 +30. Kuppers, F., Kronenberger, J., Shantia, A., Haselhoff, A.: Multivariate confidence calibration for object detection. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020) +31. Li, L.H., Zhang, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +32. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014) +34. Liu, B., Ayed, I.B., Galdran, A., Dolz, J.: The devil is in the margin: Margin-based label smoothing for network calibration. In: CVPR (2022) +35. Liu, D.C., Nocedal, J.: On the limited memory bfgs method for large scale optimization. Math. Program. 45(1-3), 503-528 (1989), http://dblp.uni-trier.de/db/journals/mp/mp45.htm1#LiuN89 +36. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +37. Lu, Y., Lu, C., Tang, C.K.: Online video object detection using association LSTM. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 2363-2371 (2017). https://doi.org/10.1109/ICCV.2017.257 +38. Ma, X., Blaschko, M.B.: Meta-cal: Well-controlled post-hoc calibration by ranking. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 7235-7245. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/ma21a.html +39. Mehrtash, A., Wells, W.M., Tempany, C.M., Abolmaesumi, P., Kapur, T.: Confidence calibration and predictive uncertainty estimation for deep medical image + +segmentation. IEEE Transactions on Medical Imaging 39(12), 3868-3878 (Dec 2020). https://doi.org/10.1109/tmi.2020.3006437, http://dx.doi.org/10.1109/TMI.2020.3006437 +40. Mukhoti, J., Kulharia, V., Sanyal, A., Golodetz, S., Torr, P., Dokania, P.: Calibrating deep neural networks using focal loss. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 15288-15299. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/aeb7b30ef1d024a76f21a1d40e30c302-Paper.pdf +41. Munir, M.A., Khan, M.H., Khan, S., Khan, F.S.: Bridging precision and confidence: A train-time loss for calibrating object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11474-11483 (June 2023) +42. Munir, M.A., Khan, M.H., Sarfraz, M., Ali, M.: Towards improving calibration in object detection under domain shift. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 38706-38718. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/fcd812a51b8f8d05cfea22e3c9c4b369-Paper-Conference.pdf +43. Munir, M.A., Khan, S., Khan, M.H., Ali, M., Khan, F.: Cal-DETR: Calibrated detection transformer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=4SkPTD6XNP +44. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019) +45. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Localization recall precision (LRP): A new performance metric for object detection. In: The European Conference on Computer Vision (ECCV) (2018) +46. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: The International Conference on Computer Vision (ICCV) (2021) +47. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: One metric to measure them all: Localisation recall precision (lrp) for evaluating visual detection tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-1 (2021) +48. Oksuz, K., Joy, T., Dokania, P.K.: Towards building self-aware object detectors via reliable uncertainty quantification and calibration. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +49. Oksuz, K., Kuzucu, S., Joy, T., Dokania, P.K.: Mocae: Mixture of calibrated experts significantly improves object detection. arXiv preprint arXiv:2309.14976 (2023) +50. Otani, M., Togashi, R., Nakashima, Y., Rahtu, E., Heikkilä, J., Satoh, S.: Optimal correction cost for object detection evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21107-21115 (2022) +51. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In: Advances in Neural Information Processing Systems. vol. 32 (2019) +52. Pathiraja, B., Gunawardhana, M., Khan, M.H.: Multiclass confidence and localization calibration for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) + +53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011) +54. Platt, J.: Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Adv. Large Margin Classif. 10 (06 2000) +55. Popordanoska, T., Tiulpin, A., Blaschko, M.B.: Beyond classification: Definition and density-based estimation of calibration in object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 585-594 (January 2024) +56. Rahimi, A., Mensink, T., Gupta, K., Ajanthan, T., Sminchisescu, C., Hartley, R.: Post-hoc calibration of neural networks by g-layers. arXiv preprint arXiv:2006.12807 (2020) +57. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017) +58. Rezatofighi, H., Nguyen, T.T.D., Vo, B., Vo, B., Savarese, S., Reid, I.D.: How trustworthy are the existing performance evaluations for basic vision tasks? arXiv e-prints:2008.03533 (2020) +59. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision 126(9), 973-992 (Sep 2018), https://doi.org/10.1007/s11263-018-1072-8 +60. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2019) +61. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) +62. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11809-11820. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/61f3a6dbc9120ea78ef75544826c814e-Paper.pdf +63. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +64. Yan, K., Wang, X., Lu, L., Summers, R.M.: Deepleson: Automated deep mining, categorization and detection of significant radiology image findings using large-scale clinical lesion annotations. arXiv preprint arXiv:1710.01766 (2017) +65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +66. Zadrozny, B., Elkan, C.: Transforming classifier scores into accurate multiclass probability estimates. In: Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 694-699 (2002) +67. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) + +68. Zhang, H., Wang, Y., Dayoub, F., Sünderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +69. Zhang, J., Yao, W., Chen, X., Feng, L.: Transferable post-hoc calibration on pretrained transformers in noisy text classification. Proceedings of the AAAI Conference on Artificial Intelligence 37(11), 13940-13948 (Jun 2023). https://doi.org/10.1609/aaai.v37i11.26632, https://ojs.aai.org/index.php/AAAI/article/view/26632 +70. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +71. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable $\{\mathrm{detr}\}$ : Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR) (2021) +72. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023) \ No newline at end of file diff --git a/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/images.zip b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0aba20c9fc5b26a2406390839cab06b7f959e93 --- /dev/null +++ b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c878991347e8b2f0c2cd547f36cf1084c01a4fb9e8a8958837a44d20be828a3 +size 632987 diff --git a/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/layout.json b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bbb779b99e0233bf3d5e10eaaecae541cc8b1c45 --- /dev/null +++ b/2024/On Calibration of Object Detectors_ Pitfalls, Evaluation and Baselines/layout.json @@ -0,0 +1,11596 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 111, + 462, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 111, + 462, + 146 + ], + "spans": [ + { + "bbox": [ + 151, + 111, + 462, + 146 + ], + "type": "text", + "content": "On Calibration of Object Detectors: Pitfalls, Evaluation and Baselines" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 168, + 481, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 481, + 180 + ], + "type": "text", + "content": "Selim Kuzucu\\*, Kemal Oksuz\\*, Jonathan Sadeghi, and Puneet K. Dokania" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 190, + 472, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 190, + 472, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 190, + 472, + 213 + ], + "type": "text", + "content": "Five AI Ltd., United Kingdom {selim.kuzucu2, kemal.oksuz, jonathan.sadeghi, puneet.dokania}@five.ai" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 469, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 469, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 469, + 506 + ], + "type": "text", + "content": "Abstract. Reliable usage of object detectors require them to be calibrated—a crucial problem that requires careful attention. Recent approaches towards this involve (1) designing new loss functions to obtain calibrated detectors by training them from scratch, and (2) post-hoc Temperature Scaling (TS) that learns to scale the likelihood of a trained detector to output calibrated predictions. These approaches are then evaluated based on a combination of Detection Expected Calibration Error (D-ECE) and Average Precision. In this work, via extensive analysis and insights, we highlight that these recent evaluation frameworks, evaluation metrics, and the use of TS have notable drawbacks leading to incorrect conclusions. As a step towards fixing these issues, we propose a principled evaluation framework to jointly measure calibration and accuracy of object detectors. We also tailor efficient and easy-to-use post-hoc calibration approaches such as Platt Scaling and Isotonic Regression specifically for object detection task. Contrary to the common notion, our experiments show that once designed and evaluated properly, post-hoc calibrators, which are extremely cheap to build and use, are much more powerful and effective than the recent train-time calibration methods. To illustrate, D-DETR with our post-hoc Isotonic Regression calibrator outperforms the recent train-time state-of-the-art calibration method Cal-DETR by more than 7 D-ECE on the COCO dataset. Additionally, we propose improved versions of the recently proposed Localization-aware ECE and show the efficacy of our method on these metrics. Code is available at: https://github.com/fiveai/detection_calibration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 516, + 446, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 516, + 446, + 528 + ], + "spans": [ + { + "bbox": [ + 160, + 516, + 446, + 528 + ], + "type": "text", + "content": "Keywords: Calibration " + }, + { + "bbox": [ + 160, + 516, + 446, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 516, + 446, + 528 + ], + "type": "text", + "content": " Object Detection " + }, + { + "bbox": [ + 160, + 516, + 446, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 516, + 446, + 528 + ], + "type": "text", + "content": " Performance Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 548, + 230, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 230, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 230, + 561 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Object detectors have been widely-used in a variety of safety-critical applications related to, but not limited to, autonomous driving [6, 9, 13, 14, 61, 65] and medical imaging [22, 28, 29, 64]. In addition to being accurate, their confidence estimates should also allow characterization of their error behaviour to make them reliable. This feature, known as calibration, can enable a model to provide valuable information to subsequent systems playing crucial role in making" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 653, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 481, + 665 + ], + "type": "text", + "content": "*Equal contributions. SK contributed during his internship at Five AI Oxford team." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 219, + 175 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 219, + 175 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 219, + 175 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 219, + 175 + ], + "type": "image", + "image_path": "d7e2b008d9e7a264bfc5bb9abac53da3f2feb9a4a639484f78ea7abe627f7b2a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 176, + 199, + 186 + ], + "lines": [ + { + "bbox": [ + 158, + 176, + 199, + 186 + ], + "spans": [ + { + "bbox": [ + 158, + 176, + 199, + 186 + ], + "type": "text", + "content": "(a) D-ECE" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 115, + 306, + 175 + ], + "blocks": [ + { + "bbox": [ + 221, + 115, + 306, + 175 + ], + "lines": [ + { + "bbox": [ + 221, + 115, + 306, + 175 + ], + "spans": [ + { + "bbox": [ + 221, + 115, + 306, + 175 + ], + "type": "image", + "image_path": "3a2687f6620eb68b4dc6e79b3f122c1689f977dd67b2b7414500ac4fa3b20133.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 176, + 277, + 186 + ], + "lines": [ + { + "bbox": [ + 251, + 176, + 277, + 186 + ], + "spans": [ + { + "bbox": [ + 251, + 176, + 277, + 186 + ], + "type": "text", + "content": "(b) AP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 194, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 194, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 482, + 262 + ], + "type": "text", + "content": "Fig. 1: The performance of different detectors over operating confidence thresholds on COCO minitest. Orange: Faster R-CNN, Green: RS R-CNN, Purple: ATSS, Red: PAA, Blue: D-DETR. All measures are lower better except AP. It is not trivial to identify an operating threshold and compare detectors, especially when the common evaluation [30, 41–43, 52], combining D-ECE for calibration and AP for accuracy, is used. Instead, we use LaECE0 and Localisation-Recall-Precision Error (LRP)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 115, + 392, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 115, + 392, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 115, + 392, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 392, + 175 + ], + "type": "image", + "image_path": "9ae363959673ad87211e181932b1e66023918fb521ed7df0166e5834d0b8c275.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 176, + 372, + 186 + ], + "lines": [ + { + "bbox": [ + 328, + 176, + 372, + 186 + ], + "spans": [ + { + "bbox": [ + 328, + 176, + 372, + 186 + ], + "type": "text", + "content": "(c) LaECEo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 118, + 478, + 175 + ], + "blocks": [ + { + "bbox": [ + 394, + 118, + 478, + 175 + ], + "lines": [ + { + "bbox": [ + 394, + 118, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 394, + 118, + 478, + 175 + ], + "type": "image", + "image_path": "0cd98c320f1287d7ad646c9f8c773df7be549c494a0ef3ea458a0063e82dbe54.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 176, + 452, + 186 + ], + "lines": [ + { + "bbox": [ + 420, + 176, + 452, + 186 + ], + "spans": [ + { + "bbox": [ + 420, + 176, + 452, + 186 + ], + "type": "text", + "content": "(d) LRP" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 373 + ], + "type": "text", + "content": "safety-critical decisions [4, 24, 26, 37, 39]. Despite its importance, calibration of detectors is a relatively underexplored area in the literature and requires significant attention. Therefore, in this work, we focus on different aspects of the evaluation framework that is now being adopted by most recent works building calibrated detectors and discuss their pitfalls and propose fixes. Additionally, we tailor the well-known post-hoc calibration methods to improve the calibration of a given object detector (trained) with minimal effort." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 376, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 508 + ], + "type": "text", + "content": "Naturally, practitioners prefer detectors that perform well in terms of both accuracy and calibration, which we will refer to as joint performance. However, unlike classification, choosing the best performing model is non-trivial for object detection. This is because different detectors commonly yield detection sets with varying cardinalities for the same image, and this difference in population size is shown to affect the joint performance evaluation [48]. Furthermore, when object detectors are used in practice, an operating threshold is normally chosen [4, 24, 26, 31, 36, 37, 39], and the choice of this threshold directly influences a detector's performance. Thus, comparing the performance of a detector in terms of calibration or accuracy over different operating thresholds as well as with different detectors is not straightforward, as illustrated in Fig. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We assert that a framework for joint evaluation should follow certain basic principles. Firstly, the detectors should be evaluated on a thresholded set of detections to align with their practical usage. While doing so, the evaluation framework will require a principled model-dependent threshold selection mechanism, as the confidence distribution of each detector can differ significantly [47]. Secondly, the calibration should enforce the confidence scores to provide unambiguous and fine-grained information about the detection quality. For example, if the confidence score represents the localisation quality of a detection, this provides more fine-grained information than representing whether the object is detected or not. Thirdly, the datasets should be properly-designed for evaluation. That is, the training, validation (val.) and in-distribution (ID) test splits should be sampled from the same underlying distribution, and additionally, the domain-shifted test splits — which are crucial for safety-critical applications — should" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 195 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Principles of joint performance evaluation of object detectors in terms of accuracy and calibration, and whether existing evaluation approaches violate them." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 195 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 195 + ], + "type": "table", + "html": "
Principles of Joint EvaluationD-ECE-style [30, 41–43, 52]LaECE-style [48]CE-style [55]Ours
Model-dependent threshold selectionX✓X✓
Unambiguous & fine-grained confidence scoresXXX✓
Properly-designed datasetsXXX✓
Properly-trained detectors & calibratorsX✓X✓
", + "image_path": "382af17139916493efbe02eec5bd7dfd57a4aa38934a5d6c73e1c92af4de55ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 481, + 243 + ], + "type": "text", + "content": "be included. Finally, baseline detectors and calibration methods must be trained properly, as otherwise the evaluation might provide misleading conclusions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 244, + 476, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 244, + 476, + 255 + ], + "spans": [ + { + "bbox": [ + 146, + 244, + 476, + 255 + ], + "type": "text", + "content": "There are three approaches for jointly evaluating accuracy and calibration:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 340 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 257, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 479, + 292 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 138, + 257, + 479, + 292 + ], + "type": "text", + "content": " D-CEC-style [30,41-43,52] thresholds the detections commonly from a confidence of 0.30 to compute Detection Expected Calibration Error (D-CEC) and use top-100 detections from each image for Average Precision (AP)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 138, + 293, + 480, + 316 + ], + "type": "text", + "content": " LaECE-style [48] enforces the detectors to be thresholded properly, and combine Localisation-aware Expected Calibration Error (LaECE) with LRP [47]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 316, + 480, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 316, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 139, + 316, + 480, + 340 + ], + "type": "text", + "content": "- CE-style [55] thresholds the detections from a confidence score of 0.50 to obtain Calibration Error (CE) and AP." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 342, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 519 + ], + "type": "text", + "content": "As summarized in Tab. 1, these evaluations do not adhere to the basic principles mentioned above. To exemplify, D-CEE-style evaluation — the most common evaluation approach [30, 41–43, 52] — uses different operating thresholds for calibration and accuracy, which does not align well with the practical usage of detectors. Also, using a fixed threshold (i.e., 0.30) for all detectors artificially promotes certain detectors. To illustrate, while D-CEE-style evaluation ranks the green detector as the worst in Fig. 1(a), the green one yields the best D-CEE at 0.70. Besides, as shown in Fig. 1(b), AP is maximized at the confidence of 0 (leading to too many detections with low confidence) for all the detectors, and thus AP cannot be used to obtain a proper operating threshold [47, 48]. In terms of conveying fine-grained information, D-CEE aims to align confidence with the precision only, which effectively ignores the localisation quality of the detections, a crucial performance aspect of object detection. Finally, this type of evaluation also has limitations in terms of dataset splits and the chosen baselines as we explore in Sec. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "Having proper baseline calibration methods is also essential to monitor the progress in the field. Recently proposed train-time calibration methods commonly employ an auxiliary loss term to regularize the confidence scores during training [30,41-43,52]. Such methods are shown to be effective against the Temperature Scaling (TS) [15], which is used as the only post-hoc calibration baseline. Post-hoc calibrators are obtained on a held-out val. set, and hence can easily be applied to any off-the-shelf detector. Despite their potential advantages, unlike for classification [15,20,23,38,56,62,69], post-hoc calibration methods have not been explored for object detection sufficiently [30,48]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce a joint evaluation framework which respects the aforementioned principles (Tab. 1), and thus address the critical drawbacks of existing evaluation approaches. That is, we first define " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "as novel calibration errors, each of which aims to align the detection confidence scores with their localisation qualities. Thus, the detectors respecting " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " provide informative confidence estimates about their behaviours. We measure accuracy using LRP [45,47], which requires a proper combination of false-positive (FP), false-negative (FN) and localisation errors. Thereby requiring the detectors to be properly-thresholded as shown by the bell-like curves in Fig. 1(d). Also, we design three datasets with different characteristics, and introduce Platt Scaling (PS) as well as Isotonic Regression (IR) as highly effective post-hoc calibrators tailored to object detection. Our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 224, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 482, + 259 + ], + "type": "text", + "content": "We identify various quirks and assumptions in state-of-the-art (SOTA) methods in quantifying miscalibration of object detectors and show that they, if not treated properly, can provide misleading conclusions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 259, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 259, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 139, + 259, + 482, + 308 + ], + "type": "text", + "content": "We introduce a framework for joint evaluation consisting of properly-designed datasets, evaluation measures tailored to practical usage of object detectors and baseline post-hoc calibration methods. We show that our framework addresses the drawbacks of existing approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 308, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 308, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 139, + 308, + 482, + 380 + ], + "type": "text", + "content": "In contrast to the literature, we show that, if designed properly, post-hoc calibrators can significantly outperform the SOTA training time calibration methods. To illustrate, on the common COCO benchmark, D-DETR with our IR calibrator outperforms the SOTA Cal-DETR [43] significantly: (i) by more than 7 points in terms of D-ECE and (ii) " + }, + { + "bbox": [ + 139, + 308, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 139, + 308, + 482, + 380 + ], + "type": "text", + "content": " points in terms of our challenging LaECE" + }, + { + "bbox": [ + 139, + 308, + 482, + 380 + ], + "type": "inline_equation", + "content": "_0" + }, + { + "bbox": [ + 139, + 308, + 482, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 397, + 308, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 308, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 308, + 411 + ], + "type": "text", + "content": "2 Background and Notation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": "Object Detectors and Evaluating their Accuracy Denoting the set of " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " objects in an image " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\{b_i, c_i\\}^M" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "b_i \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " is a bounding box and " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "c_i \\in \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " is its class; an object detector produces the bounding box " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": ", the class label " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " and the confidence score " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " for the objects in " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "f(X) = \\{\\hat{c}_i, \\hat{b}_i, \\hat{p}_i\\}^N" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " being the number of predictions. During evaluation, each detection is first labelled as a true-positive (TP) or a FP using a matching function " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " relying on an Intersection-over-Union (IoU) threshold " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " to validate TPs. We assume " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\psi(i)" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " returns the index of the object that a TP " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " matches to; else " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " is a FP and " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\psi(i) = -1" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": ". Then, AP [11, 16, 33], the common accuracy measure, corresponds to the area under the Precision Recall (PR) curve. Though widely-used, AP has been criticized recently from different aspects [5, 45, 47, 50, 58]. To illustrate, AP is maximized when the number of detections increases [48] as shown in Fig. 1(b). Therefore, AP does not help choosing an operating threshold, which is critical for practical deployment. As an alternative, LRP [45, 47] combines the numbers of TP, FP, FN with the localisation error of the detections, which are denoted by " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{N_{TP}}" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{N_{FP}}" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{N_{FN}}" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{loc}(i) \\in [0, 1]" + }, + { + "bbox": [ + 130, + 422, + 482, + 617 + ], + "type": "text", + "content": " respectively:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 178, + 625, + 481, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 625, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 178, + 625, + 481, + 665 + ], + "type": "interline_equation", + "content": "\\mathrm {L R P} = \\frac {1}{\\mathrm {N} _ {\\mathrm {F P}} + \\mathrm {N} _ {\\mathrm {F N}} + \\mathrm {N} _ {\\mathrm {T P}}} \\left(\\mathrm {N} _ {\\mathrm {F P}} + \\mathrm {N} _ {\\mathrm {F N}} + \\sum_ {\\psi (i) > 0} \\mathcal {E} _ {l o c} (i)\\right). \\tag {1}", + "image_path": "178b60f7640222c64f6304ca165e71993809771f56aaab432479f852292cf619.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "Unlike AP, LRP requires the detection set to be thresholded properly as both FPs and FNs are penalized in Eq. (1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "Evaluating the Calibration of Object Detectors The alignment of accuracy and confidence of a model, termed calibration, is extensively studied for classification [8, 15, 27, 40, 44, 63]. That is, a classifier is calibrated if its accuracy is " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " for the predictions with confidence of " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "p \\in [0,1]" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ". For object detection, [30] extends this definition to enforce that the confidence matches the precision of the detector, " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\hat{c}_i = c_i|\\hat{p}_i) = \\hat{p}_i, \\forall \\hat{p}_i \\in [0,1]" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\hat{c}_i = c_i|\\hat{p}_i)" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " is the precision. Then, discretizing the confidence space into " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " bins, D-ECE is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 224, + 234, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 234, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 224, + 234, + 481, + 266 + ], + "type": "interline_equation", + "content": "\\mathrm {D} - \\mathrm {E C E} = \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} \\right|}{\\left| \\hat {\\mathcal {D}} \\right|} \\left| \\bar {p} _ {j} - \\operatorname {p r e c i s i o n} (j) \\right|, \\tag {2}", + "image_path": "f6f1659769e3bd8673b95ae33b0c0f61edab977246188908a1e419e8336f2efd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{D}_j" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": " are the set of all detections and the detections in the " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": "-th bin, and " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\bar{p}_j" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": " and precision " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "(j)" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": " are the average confidence and the precision of the detections in the " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": "-th bin. Alternatively, considering that object detection is a joint task of classification and localisation, LaECE [48] aims to match the confidence with the product of precision and average IoU of TPs. Also, to prevent certain classes from dominating the error, LaECE is introduced as a class-wise measure. Using superscript " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": " to refer to each class and " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}^c (j)" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": " as the average IoU of " + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{D}_j^c" + }, + { + "bbox": [ + 130, + 277, + 482, + 376 + ], + "type": "text", + "content": ", LaECE is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 385, + 481, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 184, + 385, + 481, + 418 + ], + "type": "interline_equation", + "content": "\\mathrm {L a E C E} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} ^ {c} \\right|}{\\left| \\hat {\\mathcal {D}} ^ {c} \\right|} \\left| \\bar {p} _ {j} ^ {c} - \\operatorname {p r e c i s i o n} ^ {c} (j) \\times \\mathrm {I o U} ^ {c} (j) \\right|. \\tag {3}", + "image_path": "2158779433deb284d5caf250ff008c82b1f26926b9700b3d01d7930aa71b501b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 427, + 482, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 560 + ], + "type": "text", + "content": "Calibration Methods in Object Detection The existing methods for calibrating object detectors can be split into two groups: (1) Training-time calibration approaches [41-43,52,55] regularize the model to yield calibrated confidence scores during training, which is generally achieved by an additive auxiliary loss. (2) Post-hoc calibration methods use a held-out val. set to fit a calibration function that maps the predicted confidence to the calibrated confidence. Specifically, TS [15] is the only method considered as a baseline for recent training time methods [41,42,52,55]. As an alternative, IR [1-3,66] is used within a limited scope for a specific task called Self-aware Object Detection [48]. Furthermore, its effectiveness neither on a wide range of detectors nor against existing training-time calibration approaches has yet been investigated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 578, + 446, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 446, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 446, + 592 + ], + "type": "text", + "content": "3 Analysis of the Common D-ECE-style Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 603, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 603, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 603, + 482, + 664 + ], + "type": "text", + "content": "D-ECE-style evaluation is the most common evaluation approach adopted by several methods [30,41-43,52]. For that reason, here we provide a comprehensive analysis of this evaluation approach and analyse the LaECE-style and CE-style evaluations in App. B. Our analyses below are based on the principles outlined in Sec. 1 and Tab. 1, and show that all approaches have notable drawbacks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 448, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 331, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 331, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 331, + 331 + ], + "type": "text", + "content": "1. Model-dependent threshold selection. As AP is obtained using the top-100 detections and D-ECE is computed on detections thresholded above 0.30, D-ECE-style evaluation uses two different detection sets. This inconsistency is not reflective of how detectors are used in practice. Also, we observe that a fixed threshold of 0.30 for evaluating the calibration induces a bias for certain detectors. To illustrate, we compare the performance of different calibration methods over different thresholds in Fig. 2, where CalDETR [43] performs the best only for the threshold 0.30 and the post-hoc TS significantly outperforms it on all other thresholds. Therefore, this method of evaluation is sensitive to the choice of threshold, leading to ambiguity on the best performing method." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 339, + 118, + 480, + 216 + ], + "blocks": [ + { + "bbox": [ + 339, + 118, + 480, + 216 + ], + "lines": [ + { + "bbox": [ + 339, + 118, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 339, + 118, + 480, + 216 + ], + "type": "image", + "image_path": "55857e30c4cb45b7f556d6325a5db863f4533c87ecfbd07c8a0e872e2356e10b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 224, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 335, + 224, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 335, + 224, + 482, + 301 + ], + "type": "text", + "content": "Fig.2: Comparison of calibration methods in terms of D-ECE on COCO mini-test using D-DETR [71]. Post-hoc TS and IR calibrators are obtained on a subset of Objects365 [60] as in D-ECE-style evaluation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "type": "text", + "content": "2. Fine-grained confidence scores. Manipulating Eq. (2), we show in App. B that D-ECE for the " + }, + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "type": "text", + "content": "-th bin can be expressed as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 365, + 481, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 365, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 219, + 365, + 481, + 401 + ], + "type": "interline_equation", + "content": "\\left| \\sum_ {\\hat {b} _ {i} \\in \\hat {\\mathcal {D}} _ {j}, \\psi (i) > 0} (\\hat {p} _ {i} - 1) + \\sum_ {\\hat {b} _ {i} \\in \\hat {\\mathcal {D}} _ {j}, \\psi (i) = - 1} \\hat {p} _ {i} \\right|. \\tag {4}", + "image_path": "0fe3cbf7216f8707bb920459e13472abda06390bada8193c7c2029b7e579e719.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": "Eq. (4) implies that D-ECE is minimized when the confidence scores " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " of TPs are 1 and those of FPs are 0, which is also how the prediction-target pairs are usually constructed to train post-hoc TS [30, 41-43, 52]. Even if the detector is perfectly calibrated for these binary targets, the confidence scores do not provide information about localisation quality as illustrated by binary-valued " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\mathrm{D - ECE}}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " for both detections in Fig. 3(b). Also, Popordanoska et al. [55] utilise D-ECE in a COCO-style manner, that is they average D-ECE over different TP validation IoU thresholds similar to COCO-style AP [33]. However, we observe that this way of using D-ECE can promote ambiguous confidence scores. As an example, given two IoU thresholds " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": ", a detection " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{1} \\leq \\mathrm{IoU}(\\hat{b}_{i}, b_{\\psi(i)}) < \\tau_{2}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " is a TP for " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " but a FP for " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": ". Thus, given Eq. (4), it follows that " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " has contradictory confidence targets for " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": ". This is illustrated in Fig. 3(d) in which D-ECE" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "inline_equation", + "content": "_{\\mathrm{C}}" + }, + { + "bbox": [ + 130, + 411, + 482, + 582 + ], + "type": "text", + "content": " (red line) remains constant regardless of the confidence. Thus, using D-ECE (or another calibration measure) in this way should be avoided." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "3. Properly-designed datasets. In the literature, the val. set to obtain the post-hoc calibrator is typically taken from a different dataset than the ID dataset [41-43, 52]. Specifically, the post-hoc calibrators are obtained on a subset from Objects365 [60] and BDD100K [65] for the models trained with COCO [33] and Cityscapes [9] respectively. Hence, as expected, a different dataset inevitably induces domain shift, affecting the performance of the post-hoc calibrator [51]. To show that, following existing approaches, we obtain an IR calibrator [48] on" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 248, + 195 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 248, + 195 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 248, + 195 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 248, + 195 + ], + "type": "image", + "image_path": "73f300fb1e05cc44a585f3248d032376a9711b3637f925edf6cd8be00ca2550a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 196, + 241, + 205 + ], + "lines": [ + { + "bbox": [ + 143, + 196, + 241, + 205 + ], + "spans": [ + { + "bbox": [ + 143, + 196, + 241, + 205 + ], + "type": "text", + "content": "(a) Uncalibrated detections" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 364, + 195 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 364, + 195 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 364, + 195 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 364, + 195 + ], + "type": "image", + "image_path": "8889571d3ed27f1a9cbbb1a4416fa51054e95fe999997570e6adc0ece4628325.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 196, + 352, + 205 + ], + "lines": [ + { + "bbox": [ + 262, + 196, + 352, + 205 + ], + "spans": [ + { + "bbox": [ + 262, + 196, + 352, + 205 + ], + "type": "text", + "content": "(b) Calibrated detections" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 479, + 195 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 479, + 195 + ], + "type": "image", + "image_path": "3da4bf1e3647babc0688f30aa0ac6043dd45637f257b9ba2643a5fdec3929738.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 196, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 366, + 196, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 366, + 196, + 479, + 205 + ], + "type": "text", + "content": "(c) Calibrated detections (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 138, + 207, + 249, + 289 + ], + "blocks": [ + { + "bbox": [ + 138, + 207, + 249, + 289 + ], + "lines": [ + { + "bbox": [ + 138, + 207, + 249, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 249, + 289 + ], + "type": "image", + "image_path": "e47267df99eec3bccc9b443e6ca1768bfe5710836689ea57d2cd96169249157e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 292, + 227, + 301 + ], + "lines": [ + { + "bbox": [ + 159, + 292, + 227, + 301 + ], + "spans": [ + { + "bbox": [ + 159, + 292, + 227, + 301 + ], + "type": "text", + "content": "(d) TP for " + }, + { + "bbox": [ + 159, + 292, + 227, + 301 + ], + "type": "inline_equation", + "content": "\\tau = 0.5" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 251, + 207, + 362, + 289 + ], + "blocks": [ + { + "bbox": [ + 251, + 207, + 362, + 289 + ], + "lines": [ + { + "bbox": [ + 251, + 207, + 362, + 289 + ], + "spans": [ + { + "bbox": [ + 251, + 207, + 362, + 289 + ], + "type": "image", + "image_path": "9475da39d774a2523d85c0b4448391780aacb3c9b874bf94e453437a3dc8ed3c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 292, + 362, + 302 + ], + "lines": [ + { + "bbox": [ + 251, + 292, + 362, + 302 + ], + "spans": [ + { + "bbox": [ + 251, + 292, + 362, + 302 + ], + "type": "text", + "content": "(e) FP for " + }, + { + "bbox": [ + 251, + 292, + 362, + 302 + ], + "type": "inline_equation", + "content": "\\tau = 0.5" + }, + { + "bbox": [ + 251, + 292, + 362, + 302 + ], + "type": "text", + "content": ", TP for " + }, + { + "bbox": [ + 251, + 292, + 362, + 302 + ], + "type": "inline_equation", + "content": "\\tau = 0" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 207, + 476, + 289 + ], + "blocks": [ + { + "bbox": [ + 365, + 207, + 476, + 289 + ], + "lines": [ + { + "bbox": [ + 365, + 207, + 476, + 289 + ], + "spans": [ + { + "bbox": [ + 365, + 207, + 476, + 289 + ], + "type": "image", + "image_path": "bd1ccdfbff0d84bf901eec177dc03d33597e7a65b87a5fb835f788a5b9102f29.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 292, + 451, + 301 + ], + "lines": [ + { + "bbox": [ + 391, + 292, + 451, + 301 + ], + "spans": [ + { + "bbox": [ + 391, + 292, + 451, + 301 + ], + "type": "text", + "content": "(f) FP for " + }, + { + "bbox": [ + 391, + 292, + 451, + 301 + ], + "type": "inline_equation", + "content": "\\tau = 0" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 311, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 399 + ], + "type": "text", + "content": "Fig. 3: A pictorial comparison of the different calibration errors. (a) Uncalibrated detections of D-DETR on an image from [65]. The detections on the left and right have IoUs of 0.74 and 0.48 with the objects. (b) Calibrated detections in terms of D-ECE and LaECE using " + }, + { + "bbox": [ + 130, + 311, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\tau = 0.50" + }, + { + "bbox": [ + 130, + 311, + 482, + 399 + ], + "type": "text", + "content": ", and D-ECEc, COCO-style D-ECE as in [55]. " + }, + { + "bbox": [ + 130, + 311, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{D - ECE_C} = ?" + }, + { + "bbox": [ + 130, + 311, + 482, + 399 + ], + "type": "text", + "content": " as calibration error does not have a global minimum as shown in (d). (c) Calibrated detections in terms of LaECEo and LaACEo in which confidence matches IoU. (d-f) Calibration errors for different types of detections, for which LaACEo behave the same as LaECEo, hence excluded for clarity. App. B presents the details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 426, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 474 + ], + "type": "text", + "content": "Objects365 and compare it with the one obtained on the ID val. set in terms of D-ECE-style evaluation. Tab. 2 shows that the latter IR now outperforms (i) the former one by " + }, + { + "bbox": [ + 130, + 426, + 480, + 474 + ], + "type": "inline_equation", + "content": "\\sim 11" + }, + { + "bbox": [ + 130, + 426, + 480, + 474 + ], + "type": "text", + "content": " D-ECE and (ii) SOTA Cal-DETR [43] by 7.4 D-ECE, showing the importance of dataset design for proper evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "text", + "content": "4. Properly-trained detectors and calibrators. Though Cityscapes is commonly used in the literature [41-43,52], the models trained on this dataset follow COCO-style training. Specifically, D-DETR [71] was trained on Cityscapes for only 50 epochs though the training set of Cityscapes is " + }, + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\sim 40\\times" + }, + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "text", + "content": " smaller than that of COCO (3K vs. " + }, + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\sim 118\\mathrm{K}" + }, + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "text", + "content": "). We now tailor the training of D-DETR for Cityscapes by (i) " + }, + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 130, + 475, + 482, + 594 + ], + "type": "text", + "content": " longer training considering the smaller training set and (ii) increasing the training image scale considering the original resolution following [7,9]. We kept all other hyperparameters as they are for both Cal-DETR and D-DETR, and App. B presents the details. Tab. 3 shows that, once trained in this setting, D-DETR is more accurate and calibrated than Cal-DETR." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 620, + 478, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 478, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 478, + 634 + ], + "type": "text", + "content": "4 A Framework for Joint Evaluation of Object Detectors" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 653, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 480, + 666 + ], + "type": "text", + "content": "We now present our evaluation approach that respects to the principles in Sec. 1." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 448, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 180, + 293, + 236 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 302, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 302, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 302, + 170 + ], + "type": "text", + "content": "Table 2: Effect of using domain-shifted val. set on IR calibrator. Results are reported on COCO-minitest. Val. set is N/A for uncalibrated D-DETR and training time calibration method Cal-DETR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 180, + 293, + 236 + ], + "lines": [ + { + "bbox": [ + 138, + 180, + 293, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 293, + 236 + ], + "type": "table", + "html": "
MethodVal setD-ECEAP ↑
D-DETRN/A12.844.1
Cal-DETRN/A8.744.4
IRObjects36514.244.1
IRCOCO1.3 (+7.4)44.1
", + "image_path": "f41c453fe2cd3a0716343d95f357b615423e7baa55bc4e9125a89e6c79031c2c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 312, + 179, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 311, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 311, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: COCO training settings are commonly adopted while training DDETR on Cityscapes. When trained with larger images and longer, DETR performs slightly better than Cal-DETR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 179, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 312, + 179, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 312, + 179, + 481, + 236 + ], + "type": "table", + "html": "
MethodTraining StyleD-ECEAP↑
D-DETR [43]COCO13.826.8
Cal-DETR [43]COCO8.428.4
Cal-DETRCityscapes4.034.9
D-DETRCityscapes2.936.1
", + "image_path": "ab99d42d8ff835f52878471f2243e4889531baded620cc2569fdb688fc9edd24.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 258, + 473, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 473, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 473, + 270 + ], + "type": "text", + "content": "4.1 Towards Fine-grained Calibrated Detection Confidence Scores" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "text", + "content": "Calibration refers to the alignment of accuracy and confidence of a model. Therefore, for an object detector to be calibrated, its confidence should respect both classification and localisation accuracy. We discussed in Sec. 3 that D-ECE, as the common calibration measure, only considers the precision of a detector, thereby ignoring its localisation performance (Eq. (2)). LaECE [48], defined in Eq. (3) as an alternative to D-ECE, enforces the confidence scores to represent the product of precision and average IoU of TPs. Thus, LaECE considers IoUs of only TPs, and effectively ignores the localisation qualities of detections if their IoU is less than the TP validation threshold " + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\tau > 0" + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "text", + "content": ". We assert that this selection mechanism based on IoU unnecessarily limits the information conveyed by the confidence score. We illustrate this on the right car in Fig. 3(b) for which LaECE requires a target confidence of 0 (" + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\mathrm{LaECE}} = 0" + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "text", + "content": ") as its IoU is less than " + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\tau = 0.50" + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "text", + "content": ". However, instead of conveying a 0 confidence and implying no detection, representing its IoU by " + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "text", + "content": " provides additional information. Hence, we propose using " + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 130, + 277, + 482, + 456 + ], + "type": "text", + "content": ", in which case the calibration criterion of LaECE reduces to," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 218, + 465, + 481, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 465, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 218, + 465, + 481, + 480 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\hat {b} _ {i} \\in B _ {i} (\\hat {p} _ {i})} [ \\mathrm {I o U} (\\hat {b} _ {i}, b _ {\\psi (i)}) ] = \\hat {p} _ {i}, \\forall \\hat {p} _ {i} \\in [ 0, 1 ], \\tag {5}", + "image_path": "159cf7f55b3384211037c14a51b282cd84b70c4700c7c15b133265e352ed3fa5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": "where we define " + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)}) = 0" + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": " for FPs when " + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "inline_equation", + "content": "B_i(\\hat{p}_i)" + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": " is the set of boxes with the confidence of " + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "inline_equation", + "content": "b_{\\psi(i)}" + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": " is the ground-truth box that " + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": " matches with. To derive the calibration error for Eq. (5), we follow LaECE by using " + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "inline_equation", + "content": "J = 25" + }, + { + "bbox": [ + 130, + 489, + 482, + 540 + ], + "type": "text", + "content": " equally-spaced bins and averaging over class-wise errors and define," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 548, + 481, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 548, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 216, + 548, + 481, + 582 + ], + "type": "interline_equation", + "content": "\\mathrm {L a E C E} _ {0} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {j = 1} ^ {J} \\frac {\\left| \\hat {\\mathcal {D}} _ {j} ^ {c} \\right|}{\\left| \\hat {\\mathcal {D}} ^ {c} \\right|} \\left| \\bar {p} _ {j} ^ {c} - \\operatorname {I o} \\mathrm {U} ^ {c} (j) \\right|, \\tag {6}", + "image_path": "754439feb74d4055613b01ff02fcfc6976ce09ee31fcf2deda24a21b14cf6c29.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}^c" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}_j^c" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " denote the set of all detections and those in " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": "th bin respectively, " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{p}_j^c" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " is the average confidence score and " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}^c (j)" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " is the average IoU of detections in the " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": "-th bin for class " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": ", and the subscript 0 refers to the chosen " + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 591, + 482, + 666 + ], + "type": "text", + "content": " which is 0. Furthermore, similar to the classification literature [40, 44], we define Localisation-aware Adaptive Calibration Error (LaACE) using an adaptive binning approach in which the number of detections in each bin is equal. In order" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 135, + 471, + 179 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "type": "text", + "content": "Table 4: Datasets for evaluating object detection and instance segmentation methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 135, + 471, + 179 + ], + "lines": [ + { + "bbox": [ + 140, + 135, + 471, + 179 + ], + "spans": [ + { + "bbox": [ + 140, + 135, + 471, + 179 + ], + "type": "table", + "html": "
TypeTrain setVal setID test setDomain-shifted test set
Common ObjectsCOCO trainCOCO minivalCOCO minitestCOCO minitest-C, Obj45K
Autonomous DrivingCS trainCS minivalCS minitestCS minitest-C, Foggy-CS
Long-tailed ObjectsLVIS trainLVIS minivalLVIS minitestLVIS minitest-C
", + "image_path": "5d4b33f40c53aef58411a483d128cab4c338fd618eb219e0bb684b60d5e0f27d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 199, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 199, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 479, + 236 + ], + "type": "text", + "content": "to capture the model behaviour precisely, we adopt the extreme case in which each bin has only one detection, resulting in an easy-to-interpret measure which corresponds to the mean absolute error between the confidence and the IoU," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 244, + 480, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 244, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 205, + 244, + 480, + 277 + ], + "type": "interline_equation", + "content": "\\mathrm {L a A C E} _ {0} = \\frac {1}{K} \\sum_ {c = 1} ^ {K} \\sum_ {i = 1} ^ {| \\hat {\\mathcal {D}} ^ {c} |} \\frac {1}{| \\hat {\\mathcal {D}} ^ {c} |} \\left| \\hat {p} _ {i} - \\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right) \\right|. \\tag {7}", + "image_path": "34a337d1ad38d4cc7b33b2e9826b2fef720020534b83517fb2e870ed297bd553.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "content": "As we show in App. C, " + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "content": " are both minimized when " + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\hat{p}_i = \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})" + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "content": " for all detections, which is also a necessary condition for " + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "content": ". Hence, as illustrated on the right car in Fig. 3(c) and (e), " + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 131, + 284, + 479, + 332 + ], + "type": "text", + "content": " requires conveying more fine-grained information compared to other measures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 348, + 460, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 460, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 460, + 361 + ], + "type": "text", + "content": "4.2 Model-dependent Thresholding for Proper Joint Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": "In practice, object detectors employ an operating threshold to preferably output only TPs with high recall. However, AP as the common performance measure does not enable cross-validating such a threshold as it is maximized when the recall is maximized despite a drop in precision [47, 48]. This can be observed in Fig. 1(b) where AP consistently decreases as the confidence threshold increases. Alternatively, LRP (Eq. 1) prefers detectors with high precision, recall and low localisation error as illustrated by the bell-like curves in Fig. 1(d). This is because, unlike AP, LRP severely penalises detectors with low recall or precision, making it a perfect fit for our framework. As a result, we consider " + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": " with LRP and require each model to be thresholded properly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 503, + 299, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 299, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 299, + 515 + ], + "type": "text", + "content": "4.3 Properly-designed Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We curate three datasets summarized in Tab. 4: (i) COCO [33] including common daily objects; (ii) Cityscapes [9] with autonomous driving scenes; and (iii) LVIS [16], a challenging dataset focusing on the calibration of long-tailed detection. For each dataset, we ensure that train, val. and ID test sets are sampled from the same distribution, and include domain-shifted test sets. As these datasets do not have public labels for test sets, we randomly split their val. sets into two as minival and minitest similar to [17,48,49]. In such a way, we provide ID val. sets to enable obtaining post-hoc calibrators and the operating thresholds properly. For domain-shifted test sets, we apply common corruptions [21] to the ID test sets, and include Obj45K [48,60] and Foggy Cityscapes [59] as more realistic shifts. Our datasets also have mask annotations and hence they can be used to evaluate instance segmentation methods. App. C includes further details." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 459, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 459, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 459, + 128 + ], + "type": "text", + "content": "4.4 Baseline Post-hoc Calibrators Tailored to Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 289 + ], + "type": "text", + "content": "It is essential to develop post-hoc calibration methods tailored to object detection, which has certain differences from the classification task. Existing methods [41-43,52,55] use only TS as a baseline without considering the peculiarities of detection. Specifically, a single temperature parameter " + }, + { + "bbox": [ + 130, + 133, + 482, + 289 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 133, + 482, + 289 + ], + "type": "text", + "content": " is learned to adjust the predictive distribution while the confidence " + }, + { + "bbox": [ + 130, + 133, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 133, + 482, + 289 + ], + "type": "text", + "content": " is commonly assumed to be a Bernoulli random variable [30]. However, PS, fitting both a scale and a shift parameter, is the widely-accepted calibration approach when the underlying distribution is Bernoulli [15,54]. Also, how to construct a useful subset of the detections to train the post-hoc calibrators has not been explored. To address these shortcomings, we present (i) Platt Scaling in which the bias term makes a notable difference in the performance, and (ii) Isotonic Regression by modeling the calibration as a regression problem. Before introducing them, we now present an overview on how we determine the set of detections to train the calibrators." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": "Overview We obtain post-hoc calibrators on a held out val. set using the detections that are similar to those seen at inference to prevent low-scoring detections from dominating the training of the calibrator. To do so, we crossvalidate a calibration threshold " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\bar{u}^c" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": " and train a class-specific calibrator " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\zeta^c: [0,1] \\to [0,1]" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": " using the detections with higher scores than " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\bar{u}^c" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": ". Still, as " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\zeta^c(\\cdot)" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": " changes the confidence scores, we need another threshold " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\bar{v}^c" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": ", as the operating threshold, to remove the redundant detections after calibration. Following the accuracy measure, we cross-validate " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\bar{u}^c" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\bar{v}^c" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": " using LRP. As for inference time, for the " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": "-th detection " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "(\\hat{p}_i, \\hat{b}_i, \\hat{c}_i)" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\hat{p}_i \\geq \\bar{u}^{\\hat{c}_i}" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": ", it survives to the calibrator and then " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\hat{p}_i^{cal} = \\zeta^{\\hat{c}_i}(\\hat{p}_i)" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": ". Finally, if " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\hat{p}_i^{cal} \\geq \\bar{v}^{\\hat{c}_i}" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": "-th detection is an output of the detector. Alg. A.1 and A.2 provide the details. For " + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\zeta^c(\\cdot)" + }, + { + "bbox": [ + 130, + 289, + 482, + 444 + ], + "type": "text", + "content": ", we prefer monotonically increasing functions in order not to affect the ranking of the detections significantly and to keep their accuracy as we detail below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "text", + "content": "Distribution Calibration via Platt Scaling Assuming that " + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "text", + "content": " is sampled from Bernoulli distribution " + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\cdot)" + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "text", + "content": ", we aim to minimize the Negative Log-Likelihood (NLL) of the predictions on the target distribution " + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)}))" + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "text", + "content": " using PS [54]. Accordingly, we recover the logits, and then shift and scale the logits to obtain the calibrated probabilities " + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\hat{p}_i^{cal}" + }, + { + "bbox": [ + 131, + 445, + 482, + 506 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 256, + 512, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 512, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 256, + 512, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {i} ^ {\\text {c a l}} = \\sigma \\left(a \\sigma^ {- 1} \\left(\\hat {p} _ {i}\\right) + b\\right), \\tag {8}", + "image_path": "b84cf8fce4e87089ef6995d9b3443992623a57c01e1fbea59d85accf2570729a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "text", + "content": " is the sigmoid and " + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\sigma^{-1}(\\cdot)" + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "text", + "content": " is its inverse, as well as " + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "inline_equation", + "content": "a \\geq 0" + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "text", + "content": " are the learnable parameters. We derive the NLL for the " + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 531, + 481, + 555 + ], + "type": "text", + "content": "th detection in App. C as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 177, + 561, + 480, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 561, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 177, + 561, + 480, + 574 + ], + "type": "interline_equation", + "content": "- \\left(\\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right) \\log \\left(\\hat {p} _ {i} ^ {c a l}\\right) + \\left(1 - \\operatorname {I o U} \\left(\\hat {b} _ {i}, b _ {\\psi (i)}\\right)\\right) \\log \\left(1 - \\hat {p} _ {i} ^ {c a l}\\right)\\right). \\tag {9}", + "image_path": "2a9cf5eb32075ce3c51d969ce832eb0e622f80bd53a1ec84a3a0413b5ec05b6d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "text", + "content": "Please note that Eq. (9), which is in fact the cross-entropy loss, is minimized if " + }, + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "inline_equation", + "content": "\\hat{p}_i^{cal} = \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})" + }, + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 131, + 580, + 481, + 618 + ], + "type": "text", + "content": " are minimized. We optimize Eq. (9) via the second-order optimization strategy L-BFGS [35] following [30]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 618, + 491, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 491, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 491, + 666 + ], + "type": "text", + "content": "Confidence Calibration via Isotonic Regression As an alternative perspective, " + }, + { + "bbox": [ + 130, + 618, + 491, + 666 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 618, + 491, + 666 + ], + "type": "text", + "content": " can also be directly calibrated by modelling the calibration as a regression task. To do so, we construct the prediction-target pairs " + }, + { + "bbox": [ + 130, + 618, + 491, + 666 + ], + "type": "inline_equation", + "content": "(\\{\\hat{p}_i, \\mathrm{IoU}(\\hat{b}_i, b_{\\psi(i)})\\})" + }, + { + "bbox": [ + 130, + 618, + 491, + 666 + ], + "type": "text", + "content": " on the held-out val. set and then fit an IR model using scikit-learn [53]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 168, + 477, + 285 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Comparison with SOTA methods in terms of other evaluation measures on COCO [33]. LRP is reported on LRP-optimal thresholds obtained on val. set. AP is reported on top-100 detections. " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " is taken as 0.50. All measures are lower-better, except AP. Bold: the best, underlined: second best. PS: Platt Scaling, IR: Isotonic Regression." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 168, + 477, + 285 + ], + "lines": [ + { + "bbox": [ + 140, + 168, + 477, + 285 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 477, + 285 + ], + "type": "table", + "html": "
Cal.\nTypeMethodCalibration (thr. 0.30)Calibration (LRP thr.)Accuracy
D-ECELaECED-ECELaECELRPAP↑
Uncal.D-DETR [71]12.813.215.012.166.344.1
Training\nTimeMbLS [34]15.616.318.715.865.944.3
MDCA [19]12.213.514.312.666.443.8
TCD [42]12.413.114.412.366.644.0
BPC [41]9.813.111.412.366.843.6
Cal-DETR [43]8.712.99.711.866.044.4
Post-hoc\n(Ours)PS for D-ECE0.9(+7.8)16.32.4(+7.3)15.866.344.1
PS for LaECE11.011.5(+1.4)9.410.1(+1.7)66.344.1
IR for D-ECE1.3(+7.4)15.72.6(+7.1)15.366.244.1
IR for LaECE10.28.9(+4.0)9.38.2(+3.6)66.343.7
", + "image_path": "a6caf504bb9c345e4131867507e1fc47f2f601eaad25d091169716240d3edc84.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "content": "Adapting Our Approach to Different Calibration Objectives Until now, we considered post-hoc calibrators for " + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE_0}" + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE_0}" + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "content": " though in practice different measures can be preferred. Our post-hoc calibrators can easily be adapted for such cases by considering the dataset design and optimisation criterion. To illustrate, for D-ECE-style evaluation, the calibration dataset is to be class-agnostic where the detections are thresholded from 0.30 with prediction-target pairs for IR as " + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "inline_equation", + "content": "(\\{\\hat{p}_i,0\\})" + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "inline_equation", + "content": "(\\{\\hat{p}_i,1\\})" + }, + { + "bbox": [ + 130, + 308, + 482, + 392 + ], + "type": "text", + "content": " for FPs and TPs respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 411, + 302, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 302, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 302, + 426 + ], + "type": "text", + "content": "5 Experimental Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 473 + ], + "type": "text", + "content": "We now show that our post-hoc calibration approaches consistently outperform training time calibration methods by significant margins (Sec. 5.1) and that they generalize to any detector and can thus be used as a strong baseline (Sec. 5.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 490, + 459, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 459, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 459, + 502 + ], + "type": "text", + "content": "5.1 Comparing Our Baselines with SOTA Calibration Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": "Here, we compare PS and IR with recent training-time calibration methods considering various evaluation approaches. As these training-time methods mostly rely on D-DETR, we also use D-DETR with ResNet-50 [18]. We obtain the detectors of training time approaches trained with COCO dataset from their official repositories, whereas we incorporate Cityscapes into their official repositories and train them using the recommended setting in Tab. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Comparison on Other Evaluation Approaches Before moving on to our evaluation approach, we first show that our PS and IR outperform all existing training time methods on existing evaluation approaches. For that, we consider D-ECE and the LaECE from " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\tau = 0.5" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " by including two different evaluation settings for each: (i) the detection set is obtained from the fixed threshold of 0.30 following the convention [30,41-43], and (ii) the operating thresholds are cross-validated using LRP. Following their standard usage, we use 10 and 25" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 158, + 473, + 275 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 6: Comparison with SOTA calibration methods on Common Objects using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 158, + 473, + 275 + ], + "lines": [ + { + "bbox": [ + 144, + 158, + 473, + 275 + ], + "spans": [ + { + "bbox": [ + 144, + 158, + 473, + 275 + ], + "type": "table", + "html": "
Calibration TypeMethodCOCO minitest (ID)COCO-C (Domain Shift)Obj45K (Domain Shift)
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
UncalibratedD-DETR [71]12.727.157.314.628.771.516.435.872.0
Training-timeMbLS [34]16.530.356.816.831.171.817.337.171.6
MDCA [19]13.127.257.514.528.771.816.635.672.2
TCD [42]13.026.757.614.628.371.916.335.571.7
BPC [41]12.425.557.714.127.172.117.334.572.0
Cal-DETR [43]11.624.656.213.826.470.618.835.371.1
Post-hoc (Ours)Platt Scaling9.623.557.312.825.671.517.033.772.0
Isotonic Regression7.723.157.210.725.371.517.233.372.0
(+3.9)(+1.5)(+3.1)(+1.1)(-0.9)(+1.2)
", + "image_path": "727b8852eb2b5186fd2e01914b37d1292e6934f15d9b7661729664bea9e6cf66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 144, + 337, + 473, + 439 + ], + "blocks": [ + { + "bbox": [ + 132, + 293, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 327 + ], + "type": "text", + "content": "Table 7: Comparison with SOTA on Autonomous Driving using our proposed evaluation. Our gains (green/red) are reported for IR compared to the best existing approach. Bold: the best, underlined: second best in terms of calibration." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 337, + 473, + 439 + ], + "lines": [ + { + "bbox": [ + 144, + 337, + 473, + 439 + ], + "spans": [ + { + "bbox": [ + 144, + 337, + 473, + 439 + ], + "type": "table", + "html": "
Calibration TypeMethodCityscapes minitest (ID) LaECE0|LaACE0|LRPCityscapes-C (Domain Shift) LaECE0|LaACE0|LRPFoggy Cityscapes (Domain Shift) LaECE0|LaACE0|LRP
UncalibratedD-DETR [71]20.326.057.221.425.680.218.522.369.4
Training-timeTCD [42]16.831.759.223.232.481.624.433.871.6
BPC [41]23.831.864.928.133.383.724.730.973.8
Cal-DETR [43]21.325.356.923.026.480.820.023.271.0
Post-hoc (Ours)Platt Scaling9.623.357.217.726.280.211.321.669.4
Isotonic Regression9.0(+7.8)23.7(+1.6)56.816.4(+5.0)25.8(-0.2)80.510.0(+8.5)21.2(+1.1)69.5
", + "image_path": "f8fc842e44c50a16585600aefd8dd232053601a4c0842c9a45cdd371665d7fad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 469, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 567 + ], + "type": "text", + "content": "bins to compute D-ECE and LaECE respectively. We optimize PS and IR by considering the calibration objective as described in Sec. 4.4. Tab. 5 shows that PS and IR outperform SOTA Cal-DETR significantly by more than 7 D-ECE and up to 4 LaECE on COCO minitest. Please note that all previous approaches are optimized for " + }, + { + "bbox": [ + 130, + 469, + 482, + 567 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 469, + 482, + 567 + ], + "type": "text", + "content": "-ECE thresholded from 0.30, in terms of which our PS yields only 0.9 " + }, + { + "bbox": [ + 130, + 469, + 482, + 567 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 469, + 482, + 567 + ], + "type": "text", + "content": "-ECE improving the SOTA by 7.8. Tab. 5 also suggests that post-hoc calibrators perform the best when the calibration objective is aligned with the measure. App. D shows that our observations generalize to Cityscapes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Common Objects Setting We now evaluate detectors using our evaluation approach. Tab. 6 shows that IR and PS share the top-2 entries on almost all test subsets by preserving the accuracy (LRP) of D-DETR. Specifically, our gains on ID set and COCO-C are significant, where IR outperforms Cal-DETR by around " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 - 4\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.0 - 1.5\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". As for Obj45K, the challenging test set with natural shift, IR and PS improve " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " but perform slightly worse in terms of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". This is an expected drawback of post-hoc approaches when the domain shift is large as they are trained only with ID val. set [51]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 158, + 479, + 246 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 8: Comparison with TS using D-DETR. Bold: the best, underlined: second best. X : domain-shifted val. set is used to obtain thresholds and calibrators, decreasing the accuracy (red font). Bias term only exists for PS ( " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " in Eq. (8)), thus " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{N}/\\mathrm{A}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " for IR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 158, + 479, + 246 + ], + "lines": [ + { + "bbox": [ + 137, + 158, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 158, + 479, + 246 + ], + "type": "table", + "html": "
MethodAblations on DatasetAblations on ModelCOCO minitestCityscapes minitest
ID Val. SetThresholdClass-wiseBias TermLaECE0LaACE0LRPLaECE0LaACE0LRP
Temperature Scaling (Current Baseline)✘12.320.861.521.025.960.3
✓12.523.157.320.926.357.2
Ablations on Temperature Scaling✓✓11.324.857.313.325.557.2
✓✓12.422.957.323.127.557.2
✓✓✓10.624.257.312.724.657.2
Platt Scaling (Ours) Isotonic Regression (Ours)✓✓✓✓9.623.557.39.623.357.2
✓✓✓N/A7.723.157.29.023.756.8
", + "image_path": "3ab2636fc33bdeeae40195539f60483527ee5150b2d102a71e1d2f924109141d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 300, + 479, + 450 + ], + "blocks": [ + { + "bbox": [ + 132, + 256, + 479, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 256, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 479, + 289 + ], + "type": "text", + "content": "Table 9: Calibrating and evaluating different object detectors. We use Common Objects setting and report the results on COCO minitest. * denotes the detectors in Fig. 1. Bold: the best, underlined: second best for calibration." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 300, + 479, + 450 + ], + "lines": [ + { + "bbox": [ + 138, + 300, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 479, + 450 + ], + "type": "table", + "html": "
TypeDetectorBackboneUncalibratedPlatt ScalingIsotonic RegressionAP ↑
LaECE0LaACE0LRPLaECE0LaACE0LRPLaECE0LaACE0LRP
One-StagePAA [25]*R5015.928.159.79.724.359.77.723.859.743.2
ATSS [70]*R5019.134.059.510.324.759.58.524.159.543.1
GFL [32]R5013.728.559.310.324.559.38.324.059.343.0
VFNet [68]R5013.925.857.710.725.157.78.324.657.744.8
Two-StageFaster R-CNN [57]*R5027.029.960.410.423.860.48.623.560.440.1
RS R-CNN [46]*R5019.728.958.710.223.558.78.123.058.842.4
DETR-likeD-DETR [71]*R5012.727.157.39.623.557.37.723.157.244.1
UP-DETR [10]R5034.435.255.810.022.655.88.222.255.942.9
DINO [67]R5013.626.953.610.623.553.68.922.853.650.4
OVODGLIP [31]Swin-T13.025.349.09.222.449.07.721.849.055.7
G. DINO [36]Swin-T13.827.546.98.921.946.97.721.347.058.3
SOTACo-DETR [72]Swin-L10.823.041.58.620.241.56.719.341.664.5
EVA [12]ViT(EVA)17.421.241.28.620.341.27.119.941.264.5
MoCaE [49]N/A10.621.440.78.920.440.77.319.940.765.0
", + "image_path": "38a715e43f9018cc99cd9903891029e0a88b8d3f4816091d92ef0f456f7486f5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "text", + "content": "Autonomous Driving Setting Tab. 7 shows that our approaches consistently outperform all training time calibration approaches on this setting as well. Specifically, our gains are very significant ranging between 5.0-8.5 " + }, + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "text", + "content": " compared to the SOTA Cal-DETR, further presenting the efficacy of our approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Comparison with Existing Temperature Scaling Baseline and Ablations Tab. 8 compares TS for different design choices as well as with our PS and IR. Please note that " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " corresponds to the baseline setting used in the recent approaches [41-43,52] that employ Objects365 [60] and BDD100K [60] as domain-shifted val. sets for obtaining the calibrator. Due to this domain shift, the accuracy of TS degrades by up to 4 LRP, in red font, as the operating thresholds obtained on these val. sets do not generalize to the ID set; showing that it is crucial to use an ID val set. In ablations, thresholding the detections and class-wise calibrators generally improves the performance of TS and a more notable gain is observed once the bias term is used in PS. Our PS outperforms TS baseline obtained on ID val. set by " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 3" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " LaECE" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "_0" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " on COCO and 11.4 LaECE" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "_0" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " on Cityscapes. Finally, IR performs on par or better compared to PS." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 445, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 445, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 445, + 129 + ], + "type": "text", + "content": "5.2 Calibrating and Evaluating Different Detection Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 287, + 302 + ], + "type": "text", + "content": "Another benefit of our post-hoc calibrators is that they generalize to any object detector, thus they can be reliably used as baselines. To show that, we calibrate 14 different detectors with a diverse set of architectures using PS and IR in Tab. 9. The results suggest that both IR and PS perform better than uncalibrated detectors. IR consistently outperforms PS as it fits multiple piece-wise linear functions while PS learns only two parameters (Eq. (8)). Specifically, IR" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 299, + 160, + 384, + 244 + ], + "blocks": [ + { + "bbox": [ + 299, + 160, + 384, + 244 + ], + "lines": [ + { + "bbox": [ + 299, + 160, + 384, + 244 + ], + "spans": [ + { + "bbox": [ + 299, + 160, + 384, + 244 + ], + "type": "image", + "image_path": "1f243872d76195e0be22a46a33633ca48ce9818721b25c4d71b2626689a7532b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 248, + 373, + 258 + ], + "lines": [ + { + "bbox": [ + 312, + 248, + 373, + 258 + ], + "spans": [ + { + "bbox": [ + 312, + 248, + 373, + 258 + ], + "type": "text", + "content": "(a) Uncalibrated" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 266, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 294, + 266, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 294, + 266, + 480, + 277 + ], + "type": "text", + "content": "Fig. 4: The reliability diagrams of UP-DETR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 160, + 474, + 244 + ], + "blocks": [ + { + "bbox": [ + 391, + 160, + 474, + 244 + ], + "lines": [ + { + "bbox": [ + 391, + 160, + 474, + 244 + ], + "spans": [ + { + "bbox": [ + 391, + 160, + 474, + 244 + ], + "type": "image", + "image_path": "ce0488e4f241f90660917f7120d337b9bdd7e2f4e0a03927de10cc76a0fa2a30.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 248, + 470, + 258 + ], + "lines": [ + { + "bbox": [ + 395, + 248, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 395, + 248, + 470, + 258 + ], + "type": "text", + "content": "(b) Calibrated by IR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "text", + "content": "decreases the range of " + }, + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "inline_equation", + "content": "\\mathrm{LaECE_0}" + }, + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "text", + "content": " from 10.6-34.4 to 6.7-8.9 on COCO minitest by preserving the accuracy, making it a solid baseline. For further insights, Fig. 4 provides the reliability diagrams of the overconfident UP-DETR, of which IR significantly improves its calibration. As for SOTA, MoCaE performs the best in terms of accuracy with 40.7 LRP while Co-DETR has the best calibration with " + }, + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "inline_equation", + "content": "6.7\\mathrm{LaECE}_0" + }, + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "inline_equation", + "content": "19.3\\mathrm{LaACE}_0" + }, + { + "bbox": [ + 130, + 302, + 481, + 409 + ], + "type": "text", + "content": ", which the future work should aim to surpass. App. D.3 includes our results on our Long-tailed Objects setting (including object detection and instance segmentation), showing the effectiveness of our post-hoc calibrators on this challenging setting as well." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": "The performance measures we use in our evaluation framework are also easily interpretable. For accuracy, LRP is a weighted combination of its FP, FN and localisation error components (App. A), which, as an example, are 10.1, 22.7 and 18.8 respectively for Co-DETR [72] calibrated with IR. Also considering " + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathrm{LaACE}_0 = 19.3" + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": ", one can easily infer that: once deployed with the operating thresholds determined by our framework, Co-DETR finds " + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "inline_equation", + "content": "77.3\\%" + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": " of the objects with " + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "inline_equation", + "content": "89.9\\%" + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": " precision and " + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "inline_equation", + "content": "81.2\\%" + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": " IoU where " + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "inline_equation", + "content": "19.3\\%" + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": " is the mean absolute error of the confidence to represent IoU. We believe these intuitive measures will enable practitioners to make better decision when deploying object detectors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 534, + 226, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 226, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 226, + 547 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "The progress in a field heavily relies on the evaluation tools and the baselines used. In this paper, we showed that existing evaluation tools for calibration as well as the baseline post-hoc calibrators for object detectors have significant drawbacks. We remedied that by introducing an evaluation framework including baseline post-hoc calibrators tailored to object detection. Our experiments suggested that, once evaluated and designed properly, the post-hoc calibrators significantly outperform all existing training-time calibrators. This implies the need for research to develop better calibration techniques for object detection, for which, we believe, our evaluation framework will be an essential pillar." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 142, + 480, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 480, + 176 + ], + "type": "text", + "content": "1. Ayer, M., Brunk, H.D., Ewing, G.M., Reid, W.T., Silverman, E.: An empirical distribution function for sampling with incomplete information. The annals of mathematical statistics pp. 641-647 (1955)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "text", + "content": "2. Barlow, R.E., Brunk, H.D.: The isotonic regression problem and its dual. Journal of the American Statistical Association 67(337), 140-147 (1972)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "text", + "content": "3. Best, M.J., Chakravarti, N.: Active set algorithms for isotonic regression; a unifying framework. Mathematical Programming 47(1), 425-439 (1990)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 221, + 480, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 221, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 221, + 480, + 265 + ], + "type": "text", + "content": "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE International Conference on Image Processing (ICIP). IEEE (Sep 2016). https://doi.org/10.1109/icip.2016.7533003, http://dx.doi.org/10.1109/ICIP.2016.7533003" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 265, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 480, + 299 + ], + "type": "text", + "content": "5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: The IEEE European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 343 + ], + "type": "text", + "content": "6. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 388 + ], + "type": "text", + "content": "7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv 1906.07155 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "text", + "content": "8. Cheng, J., Vasconcelos, N.: Calibrating deep neural networks by pairwise constraints. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 421, + 480, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 465 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 466, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 509 + ], + "type": "text", + "content": "0. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "type": "text", + "content": "1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) 88(2), 303-338 (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 543, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 480, + 587 + ], + "type": "text", + "content": "2. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 587, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 621 + ], + "type": "text", + "content": "3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "type": "text", + "content": "4. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (Nov 2019). https://doi.org/10.1002/rob.21918, http://dx.doi.org/10.1002/rob.21918" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1321-1330. PMLR (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "text", + "content": "17. Harakeh, A., Waslander, S.L.: Estimating and evaluating regression predictive uncertainty in deep object detectors. In: International Conference on Learning Representations (ICLR) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 308 + ], + "type": "text", + "content": "19. Hebbalaguppe, R., Prakash, J., Madan, N., Arora, C.: A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16081-16090 (June 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 363 + ], + "type": "text", + "content": "20. Hekler, A., Brinker, T.J., Buettner, F.: Test time augmentation meets post-hoc calibration: Uncertainty quantification under real-world conditions. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14856-14864 (Jun 2023). https://doi.org/10.1609/aaai.v37i12.26735, https://ojs.aaai.org/index.php/AAAI/article/view/26735" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "21. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (ICLR) (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 397, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 463 + ], + "type": "text", + "content": "22. Jin, C., Udupa, J.K., Zhao, L., Tong, Y., Odhner, D., Pednekar, G., Nag, S., Lewis, S., Poole, N., Mannikeri, S., Govindasamy, S., Singh, A., Camaratta, J., Owens, S., Torigian, D.A.: Object recognition in medical images via anatomy-guided deep learning. Medical Image Analysis 81, 102527 (2022). https://doi.org/https://doi.org/10.1016/j.media.2022.102527, https://www.sciencedirect.com/science/article/pii/S1361841522001748" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 464, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 518 + ], + "type": "text", + "content": "23. Joy, T., Pinto, F., Lim, S.N., Torr, P.H., Dokania, P.K.: Sample-dependent adaptive temperature scaling for improved calibration. Proceedings of the AAAI Conference on Artificial Intelligence 37(12), 14919-14926 (Jun 2023). https://doi.org/10. 1609/aaii.v37i12.26742, https://ojs.aaii.org/index.php/AAAI/article/view/26742" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 520, + 481, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 573 + ], + "type": "text", + "content": "24. Karimi, D., Dou, H., Warfield, S.K., Gholipour, A.: Deep learning with noisy labels: Exploring techniques and remedies in medical image analysis. Medical Image Analysis 65, 101759 (2020). https://doi.org/https://doi.org/10.1016/j.media.2020.101759, https://www.sciencedirect.com/science/article/pii/S1361841520301237" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 575, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 481, + 597 + ], + "type": "text", + "content": "25. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 599, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 642 + ], + "type": "text", + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "27. Kumar, A., Liang, P.S., Ma, T.: Verified uncertainty calibration. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 281 + ], + "type": "text", + "content": "28. Kumar, N., Verma, R., Anand, D., Zhou, Y., Onder, O.F., Tsougenis, E., Chen, H., Heng, P.A., Li, J., Hu, Z., Wang, Y., Koohbanani, N.A., Jahanifar, M., Tajeddin, N.Z., Gooya, A., Rajpoot, N., Ren, X., Zhou, S., Wang, Q., Shen, D., Yang, C.K., Weng, C.H., Yu, W.H., Yeh, C.Y., Yang, S., Xu, S., Yeung, P.H., Sun, P., Mahbod, A., Schaefer, G., Ellinger, I., Ecker, R., Smedby, O., Wang, C., Chidester, B., Ton, T.V., Tran, M.T., Ma, J., Do, M.N., Graham, S., Vu, Q.D., Kwak, J.T., Gunda, A., Chanduri, R., Hu, C., Zhou, X., Lotfi, D., Safdari, R., Kascenas, A., O'Neil, A., Eschweiler, D., Stegmaier, J., Cui, Y., Yin, B., Chen, K., Tian, X., Gruening, P., Barth, E., Arbel, E., Remer, I., Ben-Dor, A., Sirazitdinova, E., Kohl, M., Braunewell, S., Li, Y., Xie, X., Shen, L., Ma, J., Baksi, K.D., Khan, M.A., Choo, J., Colomer, A., Naranjo, V., Pei, L., Iftekharuddin, K.M., Roy, K., Bhattacharjee, D., Pedraza, A., Bueno, M.G., Devanathan, S., Radhakrishnan, S., Koduganty, P., Wu, Z., Cai, G., Liu, X., Wang, Y., Sethi, A.: A multi-organ nucleus segmentation challenge. IEEE Transactions on Medical Imaging 39(5), 1380–1391 (2020). https://doi.org/10.1109/TMI.2019.2947628" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "text", + "content": "29. Kumar, N., Verma, R., Sharma, S., Bhargava, S., Vahadane, A., Sethi, A.: A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging 36(7), 1550-1560 (2017). https://doi.org/10.1109/TMI.2017.2677499" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "30. Kuppers, F., Kronenberger, J., Shantia, A., Haselhoff, A.: Multivariate confidence calibration for object detection. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "31. Li, L.H., Zhang, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "32. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "text", + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "34. Liu, B., Ayed, I.B., Galdran, A., Dolz, J.: The devil is in the margin: Margin-based label smoothing for network calibration. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "35. Liu, D.C., Nocedal, J.: On the limited memory bfgs method for large scale optimization. Math. Program. 45(1-3), 503-528 (1989), http://dblp.uni-trier.de/db/journals/mp/mp45.htm1#LiuN89" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "36. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "37. Lu, Y., Lu, C., Tang, C.K.: Online video object detection using association LSTM. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 2363-2371 (2017). https://doi.org/10.1109/ICCV.2017.257" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 589, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 642 + ], + "type": "text", + "content": "38. Ma, X., Blaschko, M.B.: Meta-cal: Well-controlled post-hoc calibration by ranking. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 7235-7245. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/ma21a.html" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "39. Mehrtash, A., Wells, W.M., Tempany, C.M., Abolmaesumi, P., Kapur, T.: Confidence calibration and predictive uncertainty estimation for deep medical image" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 148 + ], + "type": "text", + "content": "segmentation. IEEE Transactions on Medical Imaging 39(12), 3868-3878 (Dec 2020). https://doi.org/10.1109/tmi.2020.3006437, http://dx.doi.org/10.1109/TMI.2020.3006437" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "text", + "content": "40. Mukhoti, J., Kulharia, V., Sanyal, A., Golodetz, S., Torr, P., Dokania, P.: Calibrating deep neural networks using focal loss. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 15288-15299. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/aeb7b30ef1d024a76f21a1d40e30c302-Paper.pdf" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 250 + ], + "type": "text", + "content": "41. Munir, M.A., Khan, M.H., Khan, S., Khan, F.S.: Bridging precision and confidence: A train-time loss for calibrating object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11474-11483 (June 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 316 + ], + "type": "text", + "content": "42. Munir, M.A., Khan, M.H., Sarfraz, M., Ali, M.: Towards improving calibration in object detection under domain shift. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 38706-38718. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/fcd812a51b8f8d05cfea22e3c9c4b369-Paper-Conference.pdf" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 317, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 317, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 317, + 481, + 350 + ], + "type": "text", + "content": "43. Munir, M.A., Khan, S., Khan, M.H., Ali, M., Khan, F.: Cal-DETR: Calibrated detection transformer. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=4SkPTD6XNP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 351, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 481, + 384 + ], + "type": "text", + "content": "44. Nixon, J., Dusenberry, M.W., Zhang, L., Jerfel, G., Tran, D.: Measuring calibration in deep learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 384, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 481, + 418 + ], + "type": "text", + "content": "45. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Localization recall precision (LRP): A new performance metric for object detection. In: The European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 418, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 453 + ], + "type": "text", + "content": "46. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: The International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "47. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: One metric to measure them all: Localisation recall precision (lrp) for evaluating visual detection tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-1 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "text", + "content": "48. Oksuz, K., Joy, T., Dokania, P.K.: Towards building self-aware object detectors via reliable uncertainty quantification and calibration. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "type": "text", + "content": "49. Oksuz, K., Kuzucu, S., Joy, T., Dokania, P.K.: Mocae: Mixture of calibrated experts significantly improves object detection. arXiv preprint arXiv:2309.14976 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "text", + "content": "50. Otani, M., Togashi, R., Nakashima, Y., Rahtu, E., Heikkilä, J., Satoh, S.: Optimal correction cost for object detection evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21107-21115 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 586, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 632 + ], + "type": "text", + "content": "51. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In: Advances in Neural Information Processing Systems. vol. 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "52. Pathiraja, B., Gunawardhana, M., Khan, M.H.: Multiclass confidence and localization calibration for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 181 + ], + "type": "text", + "content": "54. Platt, J.: Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Adv. Large Margin Classif. 10 (06 2000)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 481, + 224 + ], + "type": "text", + "content": "55. Popordanoska, T., Tiulpin, A., Blaschko, M.B.: Beyond classification: Definition and density-based estimation of calibration in object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 585-594 (January 2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 256 + ], + "type": "text", + "content": "56. Rahimi, A., Mensink, T., Gupta, K., Ajanthan, T., Sminchisescu, C., Hartley, R.: Post-hoc calibration of neural networks by g-layers. arXiv preprint arXiv:2006.12807 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 481, + 289 + ], + "type": "text", + "content": "57. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 320 + ], + "type": "text", + "content": "58. Rezatofighi, H., Nguyen, T.T.D., Vo, B., Vo, B., Savarese, S., Reid, I.D.: How trustworthy are the existing performance evaluations for basic vision tasks? arXiv e-prints:2008.03533 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 481, + 352 + ], + "type": "text", + "content": "59. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision 126(9), 973-992 (Sep 2018), https://doi.org/10.1007/s11263-018-1072-8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 385 + ], + "type": "text", + "content": "60. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 384, + 481, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 481, + 428 + ], + "type": "text", + "content": "61. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 428, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 481, + 493 + ], + "type": "text", + "content": "62. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11809-11820. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/61f3a6dbc9120ea78ef75544826c814e-Paper.pdf" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 481, + 525 + ], + "type": "text", + "content": "63. Wang, D.B., Feng, L., Zhang, M.L.: Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 481, + 557 + ], + "type": "text", + "content": "64. Yan, K., Wang, X., Lu, L., Summers, R.M.: Deepleson: Automated deep mining, categorization and detection of significant radiology image findings using large-scale clinical lesion annotations. arXiv preprint arXiv:1710.01766 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 601 + ], + "type": "text", + "content": "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "text", + "content": "66. Zadrozny, B., Elkan, C.: Transforming classifier scores into accurate multiclass probability estimates. In: Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 694-699 (2002)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "text", + "content": "67. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Calibration of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 293 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "68. Zhang, H., Wang, Y., Dayoub, F., Sünderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 204 + ], + "type": "text", + "content": "69. Zhang, J., Yao, W., Chen, X., Feng, L.: Transferable post-hoc calibration on pretrained transformers in noisy text classification. Proceedings of the AAAI Conference on Artificial Intelligence 37(11), 13940-13948 (Jun 2023). https://doi.org/10.1609/aaai.v37i11.26632, https://ojs.aai.org/index.php/AAAI/article/view/26632" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "70. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "71. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable " + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{detr}\\}" + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": ": Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (ICLR) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "72. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "Kuzucu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_content_list.json b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..710af5a9fb1bf4b50843795a955a8de3332096c1 --- /dev/null +++ b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zihu Wang $^{1}$ , Lingqiao Liu $^{2}$ , Scott Ricardo Figueroa Weston $^{1}$ , Samuel Tian $^{3}$ , and Peng Li $^{1}$", + "bbox": [ + 238, + 234, + 761, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of California, Santa Barbara CA 93106, USA \n2 The University of Adelaide, Adelaide South Australia 5001, Australia \n3 Carnegie Mellon University, Pittsburgh PA 15213, USA \n{zihu_wang, scotricardo, lip}@ucsb.edu {lingqiao.liu}@adelaide.edu.au {samuel.tian7}@gmail.com", + "bbox": [ + 238, + 277, + 761, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Self-Supervised Learning (SSL) has become a prominent approach for acquiring visual representations across various tasks, yet its application in fine-grained visual recognition (FGVR) is challenged by the intricate task of distinguishing subtle differences between categories. To overcome this, we introduce an novel strategy that boosts SSL's ability to extract critical discriminative features vital for FGVR. This approach creates synthesized data pairs to guide the model to focus on discriminative features critical for FGVR during SSL. We start by identifying non-discriminative features using two main criteria: features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss. We then introduce perturbations to these non-discriminative features while preserving discriminative ones. A decoder is employed to reconstruct images from both perturbed and original feature vectors to create data pairs. An encoder is trained on such generated data pairs to become invariant to variations in non-discriminative dimensions while focusing on discriminative features, thereby improving the model's performance in FGVR tasks. We demonstrate the promising FGVR performance of the proposed approach through extensive evaluation on a wide variety of datasets.", + "bbox": [ + 261, + 392, + 743, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Self-Supervised representation learning $\\cdot$ fine-grained visual recognition $\\cdot$ learning from generated data", + "bbox": [ + 259, + 670, + 740, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 728, + 375, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In computer vision, Fine-grained Visual Recognition (FGVR) focuses on identifying subcategories of visual data, such as bird species [1, 38], aircraft variants [29], and vehicle models [25]. Different from studies on large-scaled general image datasets [26, 32, 37], FGVR tasks highlight the challenge of distinguishing subtle visual patterns.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-Supervised Learning (SSL) methods have recently largely advanced the domain of visual representation learning, circumventing the necessity for human-provided annotations. In SSL, many contrastive learning approaches achieve state-of-the-art performance by learning the similarities between data pairs derived from augmentations of identical source images [3,4,6,15,16,27,40]. These methods facilitate the transferability of the learned representations across a wide range of visual recognition problems [4,13,14,16]. Despite these advancements, it has been suggested that SSL may prioritize general visual similarities, instead of critical subtle features in FGVR tasks, which leads to SSL's 'coarse-label bias' [9]. Furthermore, recent studies [24,35,36] have highlighted a tendency among existing SSL methods to be distracted by task-irrelevant features, consequently failing to capture FGVR-relevant patterns.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose an innovative self-supervised learning strategy focusing on selectively extracting highly discriminative features while disregarding less informative, noisy ones. Our approach involves the generation of new contrastive data pairs from the latent feature space of the encoder, training the encoder to prioritize critical objects within these pairs. To facilitate this, a decoder is employed to generate data based on the latent feature space, reconstructing an image's feature vector and its perturbed counterpart to form each data pair. As the data pairs are generated to guide the encoder to learn key features and to be invariant to variations in non-discriminative features, in each feature vector, only those dimensions associated with non-discriminative patterns are perturbed.", + "bbox": [ + 212, + 332, + 787, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, at the core of our methods are two non-discriminative feature identification techniques. Firstly, Grad-CAM [33] induced from the SSL loss to the latent feature space highlights dimensions relevant to FGVR [35,36]. Thus, we introduce greater perturbation to those less highlighted dimensions. Despite the conventional view that dimensional collapse in SSL—characterized by some encoder dimensions producing constant outputs—is undesirable [6, 23, 28], recent literature [10, 45] suggests that inducing such collapse in task-irrelevant feature dimensions yields beneficial outcomes. As shown in Fig. 2, our empirical studies indicate that in encoders pre-trained by SSL methods, there are always dimensions with low variance across the dataset which cannot effectively separate data from different categories. We thus treat these low-variance dimensions as task-irrelevant and introduce perturbations to them. The two aforementioned perturbation components are then combined and applied to the latent feature vector of each image. Images are then reconstructed from the perturbed and original feature vectors to form contrastive pairs for a contrastive loss [4, 30]. Such a framework encourages the encoder to learn the key features highlighted by Grad-CAM and to reduce variance and induce collapse in those non-discriminative dimensions with low variance across the dataset in the latent space.", + "bbox": [ + 212, + 488, + 787, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed fine-grained feature learning method can be incorporated into various existing SSL methods. We use SimSiam [6] and MoCo v2 [16] as baseline methods and incorporate our proposed technique into these methods. Experiments across various fine-grained visual datasets show the effectiveness of our method. The proposed method provides a great improvement over baseline", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods. Our methods built on MoCo v2 outperforms existing state-of-the-art Self-Supervised fine-grained visual recognition methods in numerous downstream tasks.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 215, + 395, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Self-Supervised Contrastive Learning", + "text_level": 1, + "bbox": [ + 214, + 250, + 568, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning (SSL) facilitates the learning of visual representations without the need for labeled data. Among various SSL methodologies, contrastive learning has emerged as a promising technique. With the InfoNCE loss [30] and its variants [3, 4, 6, 7, 16] being introduced as the objectives for optimization, contrastive approaches treat different views of the same image as positive data pairs, while views from different images are considered negative data pairs. The goal for the encoder is to minimize the distance between positive pairs and maximize it between negative pairs within its representation space [2, 4, 16, 39]. Methods such as BYOL [15] and SimSiam [6] rely exclusively on positive pairs. Additionally, the issue of dimensional collapse, where some encoder dimensions output constant values, is discussed in [6, 23, 28], along with proposed solutions to mitigate this phenomenon. Nonetheless, recent studies [10, 45] have shown that the collapse of dimensions associated with task-irrelevant features can enhance the performance in downstream visual recognition tasks.", + "bbox": [ + 212, + 276, + 787, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Fine-Grained Visual Recognition in Self-Supervised Learning", + "text_level": 1, + "bbox": [ + 214, + 511, + 764, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While encoders pre-trained by Self-Supervised Learning (SSL) methods demonstrate transferability and generalizability in many tasks [4, 6, 16, 21, 41], studies [9, 24, 35] reveal SSL's limitations in capturing essential features for Fine-Grained Visual Recognition (FGVR). To enhance SSL's capability in identifying critical features, several works concentrate on refining data augmentations. Approaches such as SAGA [43], CAST [34], and ContrastiveCrop [31] adopt attention-guided heatmaps to locate and better crop key objects in images. DiLo [44] introduces a novel augmentation by merging key image objects with different backgrounds to generate additional views. Contrary to methods that modify images directly, our approach involves perturbing feature vectors and generating realistic images from the latent feature space to enhance the encoder's discriminative capacity. Another line of research employs auxiliary neural networks connected to the encoder's convolutional layers for improving encoder's attention on salient regions. LEWEL [20] trains an additional head to adaptively aggregate features. Techniques such as CVSA [12] and [42] train a network to fit segmentation annotations or outputs of pre-trained saliency detectors. Similarly, LCR [35] and SAM [36] train the network to align with Grad-CAM, treating Grad-CAM as the ground truth for the encoder's attention maps. Our method proposes training the encoder on generated data pairs to learn critical features. In addition to Grad-CAM, we use dimension variance as", + "bbox": [ + 212, + 537, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a criterion for identifying non-discriminative features. Low-variance dimensions where data points across the dataset are not well separated are treated less crucial. Besides, SimCore [24] pre-trains an encoder on a target dataset, then using it to select more relevant data from a large-scaled dataset to expand the training set, upon which a new encoder is retrained for downstream tasks.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27326784c35c0e697b96d6dffa584166d661b08c2002cee2e55ec4c899bfa391.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 247, + 640, + 426 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8208488b42b70d05eb31277d521366e59433fc37c5e0dcbea0c7aa3d97f13164.jpg", + "image_caption": [ + "(b)", + "Fig. 1: The overview of the proposed method. (a) Our method can be incorporated into various existing SSL methods. A decoder is utilized to generate images from both the original feature vector and its perturbed counterpart to form data pairs. The overall loss consists three terms: a conventional contrastive loss, a reconstruction loss (ensuring the decoder evolves with the encoder), and a proposed contrastive loss on the generated pairs. (b) We propose two techniques to identify and perturb non-discriminative features in a feature vector, i.e., features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss." + ], + "image_footnote": [], + "bbox": [ + 640, + 247, + 787, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 628, + 328, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Background", + "text_level": 1, + "bbox": [ + 214, + 662, + 362, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-Supervised Contrastive Learning Without need for labels, self-supervised contrastive learning learns to represent data $\\mathbf{x} \\in \\mathbb{R}^m$ in a lower dimensional space $\\mathbb{R}^n$ by learning the similarities among data samples. Typically, in contrastive learning, the model consists two components, an encoder $f_{\\theta_e}: \\mathbb{R}^m \\to \\mathbb{R}^n$ that maps data to a latent feature space $\\mathcal{V} \\subseteq \\mathbb{R}^n$ , and a projection head $g_{\\theta_p}: \\mathbb{R}^n \\to \\mathbb{R}^k$ which projects the latent feature vectors in $\\mathcal{V} \\subseteq \\mathbb{R}^n$ to a lower dimensional representation space $\\mathcal{Z} \\subseteq \\mathbb{R}^k$ where the contrastive loss is applied. Formally, given a batch $\\mathcal{B}$ of unlabeled data, every image $\\mathbf{x}$ in it is augmented by two random augmentation $\\mathcal{T}_1$ and $\\mathcal{T}_2$ to acquire two views, i.e., $\\mathbf{x}' = \\mathcal{T}_1(\\mathbf{x})$ , $\\mathbf{x}'' = \\mathcal{T}_2(\\mathbf{x})$ . Views augmented from the same image are considered a positive", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pair, while those acquired from different images form negative pairs. Two augmented views of all images form a new batch $\\mathcal{B}_a$ which doubles the size of $\\mathcal{B}$ . The encoder and projection head are then used to represent the views in $\\mathcal{Z}$ , i.e., $\\mathbf{z}' = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}'))$ , $\\mathbf{z}'' = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}''))$ . A contrastive loss $\\mathcal{L}_C$ is then defined in $\\mathcal{Z}$ space:", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} (\\mathbf {z} ^ {\\prime}, \\mathbf {z} ^ {\\prime \\prime}) = - \\log \\frac {\\exp (\\mathbf {z} ^ {\\prime} \\cdot \\mathbf {z} ^ {\\prime \\prime} / \\tau)}{\\sum_ {\\mathbf {z _ {i}} \\in \\mathcal {B} _ {a} , \\mathbf {z _ {i}} \\neq \\mathbf {z} ^ {\\prime} , \\mathbf {z _ {i}} \\neq \\mathbf {z} ^ {\\prime \\prime}} \\exp (\\mathbf {z} ^ {\\prime} \\cdot \\mathbf {z _ {i}} / \\tau)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 233, + 785, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature hyperparameter.", + "bbox": [ + 215, + 273, + 535, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although there are subtle differences between different contrastive methods, the contrastive loss are defined similarly. MoCo [5,16] introduces a large memory of negative representations. SimSiam [6] and BYOL [15] discard negative pairs and learn solely from positive pairs.", + "bbox": [ + 212, + 289, + 782, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overview", + "text_level": 1, + "bbox": [ + 215, + 372, + 339, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overview of our method is illustrated in Fig. 1. The essence of the proposed method is learning discriminative fine-grained visual features from synthesized data pairs which are reconstructed from latent feature vectors by a decoder $h_{\\theta_d}$ . During training, to ensure that the decoder follows the evolution of the encoder, we use Mean Square Error (MSE) as the loss function to optimize the decoder. Our empirical studies show that the decoder produces images of better quality when it is trained on non-augmented images. The following reconstruction loss $\\mathcal{L}_R$ is calculated for each image.", + "bbox": [ + 212, + 397, + 784, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\frac {1}{m} \\| \\mathbf {x} - \\hat {\\mathbf {x}} \\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 530, + 784, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x} \\in \\mathbb{R}^m$ is non-augmented image, and $\\hat{\\mathbf{x}} = h_{\\theta_d}(f_{\\theta_e}(\\mathbf{x}))$ is the reconstruction of it.", + "bbox": [ + 212, + 564, + 782, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to producing $\\hat{\\mathbf{x}}$ , we generate $\\hat{\\mathbf{x}}_p$ from $\\mathbf{v}_p \\in \\mathcal{V}$ , which is a perturbed version of $\\mathbf{v}$ where non-discriminative features are perturbed. A positive data pair is formed between $\\hat{\\mathbf{x}}$ and $\\hat{\\mathbf{x}}_p$ from which the encoder learns discriminative features while disregarding task-irrelevant ones. In Sec. 3.3 and Sec. 3.4, we introduce two methods of identifying crucial discriminative dimensions in the latent feature space $\\mathcal{V}$ .", + "bbox": [ + 212, + 594, + 784, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Identifying key dimensions via Grad-CAM", + "text_level": 1, + "bbox": [ + 215, + 708, + 612, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Grad-CAM [33], a widely used saliency detection technique, uses the gradient of the target loss with respect to intermediate features of the network to produce an attention map highlighting regions in the features that contribute to minimizing the loss. As in our proposed self-supervised method, labels are not available during training, we thus choose the contrastive loss as the target. To identify important features within an image's feature vector $\\mathbf{v} = f_{\\theta_e}(\\mathbf{x}) \\in \\mathbb{R}^n$ , we form positive pair between $\\mathbf{x}$ and an augmented view $\\mathbf{x}''$ to calculate a contrastive loss", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_C(\\mathbf{z}, \\mathbf{z}^{\\prime \\prime})$ , where $\\mathbf{z} = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}))$ , $\\mathbf{z}^{\\prime \\prime} = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}^{\\prime \\prime}))$ . As it is shown in Fig. 1, the Grad-CAM score vector $\\pmb{\\eta} = \\{\\eta_i\\}_{i=1}^n \\in \\mathbb{R}^n$ is calculated by gradient of the contrastive loss with respect to the feature vector $\\mathbf{v}$ .", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\eta_ {i} = \\operatorname {R e L U} \\left(\\frac {\\partial \\mathcal {L} _ {C} \\left(g _ {\\theta_ {p}} (\\mathbf {v}) , g _ {\\theta_ {p}} \\left(\\mathbf {v} ^ {\\prime \\prime}\\right)\\right)}{\\partial v _ {i}} \\cdot v _ {i}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 202, + 785, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mathbf{v} = f_{\\theta_e}(\\mathbf{x})$ , $\\mathbf{v}'' = f_{\\theta_e}(\\mathbf{x}'')$ . $v_i$ denotes the $i_{th}$ element of $\\mathbf{v}$ . The application of $\\mathrm{ReLU}(\\cdot)$ makes $\\eta_i > 0$ for all $i \\in \\{1,2,\\dots,n\\}$ . Note that the original Grad-CAM calculates gradient with respect to feature maps of the last convolutional layer. To measure the saliency of feature dimensions, we calculate gradient with respect to feature vectors. Higher Grad-CAM scores in $\\pmb{\\eta}$ represent corresponding dimension's higher contribution to data discrimination in contrastive learning.", + "bbox": [ + 212, + 239, + 782, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the Grad-CAM scores, random Gaussian noise is then introduced as perturbation to $\\mathbf{v}$ . We first scale all elements in $\\pmb{\\eta}$ to [0,1] by min-max normalization.", + "bbox": [ + 212, + 330, + 782, + 375 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\eta} _ {i} = \\frac {\\eta_ {i} - \\operatorname* {m i n} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\}}{\\operatorname* {m a x} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\} - \\operatorname* {m i n} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 388, + 785, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{\\eta}_i$ is the $i_{th}$ element of the normalized Grad-CAM score vector $\\bar{\\eta}$ . After the normalization, a random Gaussian noise perturbation vector $\\tilde{\\mathbf{v}}^g\\in \\mathbb{R}^n$ is calculated as follows.", + "bbox": [ + 214, + 422, + 784, + 468 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {v}} ^ {g} = \\left\\{\\tilde {v} _ {i} ^ {g}: \\tilde {v} _ {i} ^ {g} \\sim \\mathcal {N} \\left(0, \\epsilon_ {g} \\cdot \\left(1 - \\bar {\\eta} _ {i}\\right)\\right) \\right\\} _ {i = 1} ^ {n} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 484, + 785, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{g}$ is a hyperparameter that controls the standard deviation of Gaussian noise. In such a perturbation, all elements are sampled from i.i.d. zero-mean Gaussian distributions. Importantly, dimensions with lower normalized Grad-CAM scores $\\bar{\\eta}_i$ receive noise from a Gaussian distribution with a greater standard deviation, implying a higher likelihood of more significant noise affecting dimensions with lower Grad-CAM scores. And on average, key dimensions with higher Grad-CAM scores are affected less which helps preserve crucial features in the original images.", + "bbox": [ + 212, + 506, + 785, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Determining feature's task-relevance via dimension variance", + "text_level": 1, + "bbox": [ + 214, + 648, + 754, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the feature perturbation technique elaborated in Sec. 3.3, we propose another technique to determine and perturb task-irrelevant dimensions.", + "bbox": [ + 212, + 672, + 782, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In SSL, dimensional collapse is a phenomenon where some encoder dimensions output constant values [6,19,23,28]. As data points are not separated along such dimensions, these dimensions can not be used to perform downstream visual recognition. However, recent works [10, 45] suggest that collapse of dimensions which are related to downstream task-irrelevant features can be beneficial. In spite of the potential benefit, how to induce beneficial dimensional collapse is not illustrated by existing SSL studies.", + "bbox": [ + 212, + 704, + 784, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As it is illustrated in Fig. 2, our empirical studies show that, in the latent feature space of encoders pre-trained by SSL methods, typically, data points are", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fdb9f0018841dd77ec83ea145e7e04a265c1965e3fe1630682532587eeaf7364.jpg", + "image_caption": [ + "(a) CUB-200" + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 411, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f137a9b8ffd2e101b08f94d2b992892d0b047bdcd4a3338fcbdc7e25d9a63c86.jpg", + "image_caption": [ + "(b) Stanford Cars", + "Fig. 2: An illustration of data distribution in the feature space of encoders pre-trained by MoCo v2 [5]. Blue and red dots represent feature vectors of two categories' data from 3 fine-grained datasets, CUB-200 [38], Stanford Cars [25], and FGVC-Aircraft [29]. $v_{min}$ and $v_{max}$ are the dimensions in the feature space where data has the minimal and maximal variance across the dataset. Probability density curve fitting of each category along each dimension is attached to the corresponding axis. Different classes are separated much better along $v_{max}$ than $v_{min}$ ." + ], + "image_footnote": [], + "bbox": [ + 418, + 143, + 593, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f36ce91d6ec23470844e58c3271456c866b03d698c2c85e871a0626e69c0a64.jpg", + "image_caption": [ + "(c) FGVC-Aircraft" + ], + "image_footnote": [], + "bbox": [ + 599, + 143, + 776, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "not well separated along dimensions with low variance. Variance along these dimensions thus introduces noise to downstream classification. Therefore, we treat such dimensions as task-irrelevant and propose a technique to induce collapse in these dimensions to guide the encoder's to be invariant to variations of such features. This technique starts with estimating the dataset's variance along each feature dimension in a feature vector memory bank $\\mathbf{M} \\in \\mathbb{R}^{D \\times n}$ of size $D$ . During training, whenever the encoder is provided with a batch of data, its feature vectors will be stored in the memory to replace the oldest batch of feature vectors in it. Variance of each dimension across the dataset can be approximated in $\\mathbf{M}$ to get a variance vector $\\mathbf{s} = \\{s_i : \\sigma^2(\\bar{\\mathbf{w}}_i), i = 1, 2, \\dots, n\\} \\in \\mathbb{R}^n$ . Here $\\bar{\\mathbf{w}}_i \\in \\mathbb{R}^D$ is the $\\ell_2$ -normalized $i_{th}$ column vector of $\\mathbf{M}$ . A feature represented by dimension $i$ is considered less discriminative if its corresponding variance $s_i < \\kappa$ where $\\kappa$ is a threshold hyperparameter. To introduce random noise to those less discriminative dimensions, similar to Eq. (4), we first apply min-max normalization to $\\mathbf{s}$ to acquire $\\bar{\\mathbf{s}} = \\{\\bar{s}_i\\}_{i=1}^n$ . We then calculate the random noise vector $\\tilde{\\mathbf{v}}^{var} = \\{\\tilde{v}_i^{var}\\}_{i=1}^n$ to be applied.", + "bbox": [ + 212, + 407, + 787, + 648 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {v} _ {i} ^ {v a r} = \\left\\{ \\begin{array}{l l} u _ {i} \\sim \\mathcal {N} \\left(0, \\epsilon_ {v a r} \\cdot \\left(1 - \\bar {s} _ {i}\\right)\\right) & \\text {i f} s _ {i} < \\kappa \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 659, + 785, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $\\epsilon_{var}$ defines the standard deviations of the i.i.d. Gaussian distributions. Similar to $\\tilde{\\mathbf{v}}^g$ , $\\tilde{\\mathbf{v}}^{var}$ introduces greater noise to lower-variance dimensions.", + "bbox": [ + 214, + 704, + 782, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Learning from reconstructed data pairs", + "text_level": 1, + "bbox": [ + 214, + 755, + 586, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the two feature perturbation techniques proposed in Sec. 3.3 and Sec. 3.4, we can finally perturb the feature vector $\\mathbf{v}$ by adding the two random Gaussian noise to it to obtain its perturbed version $\\mathbf{v}_p$ , i.e., $\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g + \\tilde{\\mathbf{v}}^{var}$ . As it is shown in Fig. 1, $\\mathbf{v}$ and $\\mathbf{v}_p$ are reconstructed by the decoder to produce", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\hat{\\mathbf{x}}$ and $\\hat{\\mathbf{x}}_p$ , respectively. In $\\hat{\\mathbf{x}}_p$ , key patterns from $\\hat{\\mathbf{x}}$ are preserved, while those contribute less to Grad-CAM or deemed less discriminative across the dataset by the low-variance criterion are perturbed.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then form a positive pair between $\\hat{\\mathbf{x}}$ and $\\hat{\\mathbf{x}}_p$ , addressing extracting key features and disregarding non-discriminative ones. To this end, we propose to pass the representations $\\hat{\\mathbf{z}} = g_{\\theta_p}(f_{\\theta_e}(\\hat{\\mathbf{x}}))$ and $\\hat{\\mathbf{z}}_p = g_{\\theta_p}(f_{\\theta_e}(\\hat{\\mathbf{x}}_p))$ to a contrastive loss $\\mathcal{L}_{C_p}$ .", + "bbox": [ + 212, + 191, + 784, + 253 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C _ {p}} (\\hat {\\mathbf {z}}, \\hat {\\mathbf {z}} _ {p}) = - \\log \\frac {\\exp (\\hat {\\mathbf {z}} \\cdot \\hat {\\mathbf {z}} _ {p} / \\tau)}{\\sum_ {\\hat {\\mathbf {z}} _ {\\mathrm {i}} \\in \\mathcal {B} _ {r} , \\hat {\\mathbf {z}} _ {\\mathrm {i}} \\neq \\hat {\\mathbf {z}} , \\hat {\\mathbf {z}} _ {\\mathrm {i}} \\neq \\hat {\\mathbf {z}} _ {p}} \\exp (\\hat {\\mathbf {z}} \\cdot \\hat {\\mathbf {z}} _ {\\mathrm {i}} / \\tau)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 265, + 784, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{B}_r$ is the set of all reconstructed and perturbed images from the current batch $\\mathcal{B}$ . By forming a positive pair between $\\hat{\\mathbf{x}}$ and $\\hat{\\mathbf{x}}_p$ , Eq. (7) requires the encoder to be invariant to those perturbed less discriminative features.", + "bbox": [ + 212, + 305, + 782, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we write the training loss $\\mathcal{L}$ of our method as follows.", + "bbox": [ + 238, + 351, + 684, + 364 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {C} + \\alpha \\cdot \\mathcal {L} _ {R} + \\nu \\cdot \\mathcal {L} _ {C _ {p}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 383, + 784, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\alpha$ and $\\nu$ are hyperparameters that control the weight of $\\mathcal{L}_R$ and $\\mathcal{L}_{C_p}$ in training. The pseudocode of our method is provided in Appendices.", + "bbox": [ + 212, + 405, + 782, + 435 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ebbd489159a4b0ae2c05fe2c6ecd4600be24e7a6f77e4949041a2fdf3443845.jpg", + "table_caption": [ + "Table 1: Performance comparison on three datasets. Our method is compared with MoCo v2 [5] and ResNet50 supervised pre-trained on ImageNet-1k [11]. Top-1 classification accuracy (in%) is reported when model is evaluated on $100\\%$ , $50\\%$ , and $20\\%$ of all labels. Rank-1, rank-5, and mAP (in%) in image retrieval are reported." + ], + "table_footnote": [], + "table_body": "
DatasetMethodsClassificationImage Retrieval
100%50%20%rank-1rank-5mAP
CUB-200ResNet5063.0655.7142.2440.3968.9415.88
MoCo V263.9856.3542.6339.7267.1415.91
Ours66.1760.8449.6942.0669.5919.70
Stanford CarsResNet5061.4149.8533.5729.2854.667.01
MoCo V262.0251.0835.4430.5156.157.13
Ours65.6054.3640.2435.8161.9410.02
FGVC-AircraftResNet5049.7941.0233.6129.4851.3910.17
MoCo V251.1344.3436.4230.0252.8711.24
Ours55.2849.3741.1033.2756.8012.69
", + "bbox": [ + 251, + 530, + 759, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 734, + 375, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 214, + 768, + 431, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. Experiments are conducted across five fine-grained visual datasets. We adopt on three widely used fine-grained datasets. Caltech-UCSD Birds 200 -2011 (CUB-200) dataset [38] contains 5994 training data and 5794 testing data", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of 200 categories of birds. Stanford Cars (Cars) [25] has 196 classes of car models where 8144 data and 8041 data are in its training and testing split respectively. FGVC-Aircraft (Aircraft) [29] has 100 classes where 6667 images are for training and 3333 images are for testing. Additionally, we consider German Traffic Sign Recognition Benchmark (GTSRB) [18] which contains 43 classes of traffic signs. This dataset is usually used in autonomous driving and smart cities development. We take 4800 images for training and 3750 images for testing from GTSRB. We also evaluate the effectiveness of our method on ISIC2017 [8], a medical image dataset with 3 categories of skin lesion analysis where 2000 images are for training and 600 images are for testing.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Settings. All methods adopt ResNet-50 [17] as the encoder backbone where weights are initialized by loading ImageNet-1k [32] pre-trained model. Using two state-of-the-art SSL method, MoCo v2 [5] and SimSiam [6], as the baseline methods, we incorporate our proposed method in their framework. For the sake of fair comparison, encoders of all methods are pre-trained for 100 epochs. And the training batch size is set to 128 for all methods. More detailed encoder pre-training settings are provided in Appendices.", + "bbox": [ + 212, + 330, + 787, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Additionally, in our method, we use a feature vector memory bank of size $D = 5632$ . For the training loss in Eq. (8), we choose $\\alpha = 1$ and $\\nu = 0.5$ . When introducing noise described in Eq. (5) and Eq. (6), we choose $\\epsilon_g = 0.1$ , $\\epsilon_{var} = 0.05$ , and $\\kappa = 0.02$ . To ensure the quality of reconstructed images, before encoder training, we freeze the encoder parameters and pre-train the decoder on target datasets by a reconstruction loss. We provide decoder pre-training details in Appendices.", + "bbox": [ + 212, + 439, + 787, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Performance Evaluation Protocols. Linear evaluation is a widely adopted protocol for assessing the performance of learned representations in visual recognition. This approach freezes the parameters of the pre-trained encoder and attaches a linear classifier to it. The classifier is then trained to perform classification. Our linear evaluation setup follows [16], detailed further in the Appendices.", + "bbox": [ + 212, + 578, + 787, + 654 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The task of image retrieval [22, 35, 41] serves as another pivotal method for evaluating the performance of representation learning. Without adjusting any model parameters, it searches the nearest neighbors of a query image in the latent feature space for images that share the same label with the query image. The effectiveness of the evaluated model is quantified by recording the proportion of retrieved images that fall into the same categories as the query image. We present three commonly utilized metrics of retrieval performance: rank-1, rank-5, and mean Average Precision (mAP).", + "bbox": [ + 212, + 656, + 787, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, to illustrate the effect of the proposed method, attention maps generated by Grad-CAM on encoders trained by different methods are compared. Images from the training datasets, along with their corresponding reconstructed and perturbed versions, are also provided.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/12e8fe9d26da5ca54b89991606ebfea71352e106b4af5279b8235f15af5bde59.jpg", + "image_caption": [ + "Fig. 3: Grad-CAM attention visualized on images. Our proposed method is incorporated into MoCo v2 and SimSiam and compared with them." + ], + "image_footnote": [], + "bbox": [ + 287, + 148, + 718, + 438 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 510, + 372, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison between the proposed method and baseline methods As among all configurations of our method, the one based on MoCo v2 achieves the best overall performance, this configuration is thus called 'ours' in the experiments. We first compare our method with MoCo v2 in Tab. 1. 'ResNet50' in Tab. 1 represents the encoder supervised pre-trained on ImageNet-1k [11].", + "bbox": [ + 212, + 537, + 784, + 612 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In linear evaluation, we evaluate pre-trained encoders using different proportions of labels. Across all three datasets, our proposed method achieves an average top-1 accuracy improvement of $3.31\\%$ , $4.27\\%$ , and $5.51\\%$ over MoCo v2, with $100\\%$ , $50\\%$ , and $20\\%$ of all labels, respectively. This advancement highlights the efficacy of our data pairs generation technique, enabling great performance in classification tasks. Notably, the remarkable improvement in label-insufficient scenarios highlights our method's ability to learn more generalizable features from unlabeled data.", + "bbox": [ + 212, + 613, + 784, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We further evaluate the effectiveness of our method by image retrieval tasks. Performance in these tasks serves as a measure of semantic consistency in the learned latent feature space. Our method outperforms MoCo v2 on all datasets in rank-1, rank-5, and mAP. Invariant to irrelevant patterns, our approach ensures that images from the same category, which exhibit discriminative patterns, are more closely clustered in the feature space. This distribution enhances performance in image retrieval.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1f8df497a97bd69aa92724a393d9c83d4edda584b175113748bd414468cf5c02.jpg", + "table_caption": [ + "Table 2: Performance comparison on ISIC2017 [8] and GTSRB [18]. Top-1 classification accuracy (in %) and rank-1 (in %) of image retrieval is reported." + ], + "table_footnote": [], + "table_body": "
MethodsISIC2017GTSRB
ClassificationRetrievalClassificationRetrieval
MoCo V2 [5]64.3366.3488.2597.46
Ours66.9267.9091.6798.59
SimSiam [6]58.8364.4681.1092.81
SimSiam + Ours63.7865.5586.9494.42
", + "bbox": [ + 285, + 184, + 723, + 276 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Additionally, we visualize Grad-CAM attention in MoCo v2 and our method in Fig. 3. Unlike MoCo v2, which may concentrate on background regions irrelevant to visual recognition, our method exhibits enhanced precision in identifying and focusing on pivotal objects within the images, effectively minimizing the influence of background distractions.", + "bbox": [ + 212, + 311, + 782, + 385 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Furthermore, we also integrate the proposed technique into a state-of-the-art negative pair-free method SimSiam [6], denoted as 'SimSiam+ours' in our experiments. Our framework significantly outperforms the original SimSiam in both linear evaluation and image retrieval tasks, as demonstrated in Tab. 3. Attention comparison is also visualized by Grad-CAM and compared with SimSiam in Fig. 3.", + "bbox": [ + 212, + 390, + 784, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To comprehensively assess the effectiveness of our proposed method, we utilize two more fine-grained datasets: the traffic sign visual dataset GTSRB and the medical image dataset ISIC2017. The results, as shown in Tab. 2, indicate that our approach enhances the performance of Self-Supervised Learning (SSL), demonstrating its potential in real-world applications for FGVR tasks.", + "bbox": [ + 212, + 484, + 784, + 560 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0a46f1d4e2971cf37909159ed9bf0601b7c56af89377a9eb4156dbd38ae9f578.jpg", + "table_caption": [ + "Table 3: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported." + ], + "table_footnote": [], + "table_body": "
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
supervised77.4688.6085.93---
Dino [3]16.7414.3312.07---
SimSiam [6]46.7545.7238.5216.2412.4518.49
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
DiLo [44]62.97-----
CVSA [12]63.02-----
LEWEL [20]64.5962.9151.9039.9132.3631.09
ContrastiveCrop [31]64.2363.2952.0439.8432.7130.37
SAM-SSL-Bilinear [36]64.9462.8552.8340.0833.1930.52
LCR [35]65.2463.9653.2241.2634.7431.55
SimSiam+Ours57.8053.6347.5024.6719.7224.56
Ours66.1765.6055.2842.0635.8133.27
", + "bbox": [ + 235, + 647, + 774, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with state-of-the-art self-supervised FGVR methods In this section, we compare our proposed method against existing self-supervised learning (SSL) techniques renowned for their enhanced fine-grained visual recognition capabilities [12, 20, 31, 35, 36, 44], as detailed in Tab. 3. Methods like Dino [3], SimSiam [6], and MoCo v2 [16], which are not optimized for fine-grained feature extraction are also listed. Supervised training performance is included to provide a comprehensive comparison. Top-1 accuracy in linear evaluation and rank-1 in image retrieval tasks are reported.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Tab. 3, our proposed method achieves the best overall performance in both image retrieval and linear evaluation tasks. DiLo [44], CVSA [12], and ContrastiveCrop [31] innovate with novel data augmentation techniques which directly modify the original images. In contrary, our method generates more realistic images from the learned feature space, highlighting the learning of FGVR-related features. And unlike SAM [36] and LCR [35] which train an auxiliary network to directly fit the encoder's attention to Grad-CAM, our method learns from generated data to highlight dimensions with high Grad-CAM scores and introduce dimensional collapse to non-discriminative features.", + "bbox": [ + 212, + 267, + 787, + 402 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bfd8b76c7cf61f0a01b53550519c520603a962e92a7afbdb51e62a49ba3e59dc.jpg", + "image_caption": [ + "Fig. 4: Generated data pairs on CUB-200, Stanford Cars, and FGVC-Aircraft. The original images are also included." + ], + "image_footnote": [], + "bbox": [ + 346, + 433, + 655, + 797 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generated data pairs of the proposed method To better understand why the proposed technique enhances the fine-grained visual feature extraction capability, we show the generated data pairs from different datasets in Fig. 4. The perturbed images, as illustrated in Fig. 4, show that they largely retain the original data's key objects, with modifications primarily appearing as subtle changes in background or less important regions, e.g., changes of the tree's branch behind a bird, alterations in vehicle light's textures and adjustments in an airplane's exterior finish. These modifications do not affect the defining features of the subjects. Remarkably, some perturbations lead to entirely new objects that maintain the original's identity. For instance, transforming a side view of a car into a front view, or depicting an aircraft in flight from a grounded position. These images, obtained by modifying latent semantics of original images, are difficult to obtain through traditional data augmentation techniques defined in the original data space. They efficiently guide the encoder in identifying which features to prioritize and which to ignore, enhancing its learning performance in FGVR.", + "bbox": [ + 217, + 145, + 785, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/02e36873604bea656674784d92eae2dfb0af2a9af4ef0ca3abedf042a36aba68.jpg", + "image_caption": [ + "Fig. 5: Performance comparison of encoders trained by $\\mathcal{L}_C + \\alpha \\cdot \\mathcal{L}_R$ with respect to different $\\alpha$ value (green solid line). Top-1 classification accuracy on Stanford Cars is reported. MoCo v2 (red dashed line) and our method (blue dashed line) are included for comparison." + ], + "image_footnote": [], + "bbox": [ + 320, + 435, + 651, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Effectiveness of the reconstruction loss in contrastive learning", + "text_level": 1, + "bbox": [ + 217, + 736, + 769, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As described in Eq. (8), our model's overall training loss, $\\mathcal{L}$ , includes a reconstruction loss term, $\\mathcal{L}_R$ . To assess $\\mathcal{L}_R$ 's effect on self-supervised contrastive learning, we incorporate a decoder into MoCo v2 and train the encoder by a loss $\\mathcal{L}_C + \\alpha \\cdot \\mathcal{L}_R$ on the Stanford Cars dataset, varying $\\alpha$ in the loss function. The results, shown in Fig. 5, indicate that $\\alpha$ values of 0.5 and 1.0 enhance top-1", + "bbox": [ + 215, + 763, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "classification accuracy the most over MoCo v2. Generally, $\\mathcal{L}_R$ provides a modest improvement (less than $1\\%$ ) to Self-Supervised FGVC.", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Effectiveness of the two feature perturbation techniques", + "text_level": 1, + "bbox": [ + 214, + 196, + 723, + 214 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As detailed in Sec. 3.3 and Sec. 3.4, two non-discriminative feature perturbation techniques are proposed to introduce noise $\\tilde{\\mathbf{v}}^g$ and $\\tilde{\\mathbf{v}}^{var}$ , respectively. We conduct further experiments to assess the effectiveness of each technique in identifying and perturbing task-irrelevant features. In Tab. 4, we evaluate three different configurations of our method: (1) Ours: $\\mathbf{v}_p$ is obtained by adding both noise components to $\\mathbf{v}$ , i.e., $\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g +\\tilde{\\mathbf{v}}^{var}$ ; (2) Ours - Grad-CAM: $\\mathbf{v}_p$ is obtained by adding only the noise generated by the low Grad-CAM scores criterion, i.e., $\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g$ ; (3) Ours - low-var: $\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^{var}$ .", + "bbox": [ + 212, + 220, + 787, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As shown in Tab. 4, all configurations achieve competitive results comparing with existing state-of-the-art self-supervised FGVR methods. And when combining two noise components $\\tilde{\\mathbf{v}}^g$ and $\\tilde{\\mathbf{v}}^{var}$ , our methods achieve the best overall performance.", + "bbox": [ + 212, + 343, + 787, + 402 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/2f21a9e68d02676597c04fab460d9735b6204dd7b9c08cd750d722498c75dcad.jpg", + "table_caption": [ + "Table 4: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported." + ], + "table_footnote": [], + "table_body": "
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
Ours - Grad-CAM66.0464.1854.1941.0235.0832.14
Ours - low-var65.9164.3453.1141.6934.5531.23
Ours66.1765.6055.2842.0635.8133.27
", + "bbox": [ + 243, + 481, + 769, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 628, + 359, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To enhance the performance of Self-Supervised Learning (SSL) in Fine-grained Visual Recognition (FGVR) tasks, this paper introduces a novel approach where an encoder learns discriminative features from generated images. By introducing noise to features deemed non-discriminative by two proposed criteria, we generate synthetic data from both the original and perturbed feature vectors by a decoder, thus forming data pairs that emphasize learning key features for FGVR. Our approach outperforms existing methods across various datasets in many downstream tasks. While the proposed approach also offers a modest boost to SSL performance in non-fine-grained visual recognition tasks—as detailed in the Appendices—the gains are notably more substantial in FGVR contexts. The refinement of our methodology for application to large-scaled general datasets remains an avenue for future research works.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. This material is based upon work supported by the National Science Foundation under Grant No. 1956313.", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 199, + 321, + 215 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014)", + "2. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020)", + "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "6. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021)", + "7. Chuang, C.Y., Robinson, J., Lin, Y.C., Torralba, A., Jegelka, S.: Debiased contrastive learning. Advances in neural information processing systems 33, 8765-8775 (2020)", + "8. Codella, N.C., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). pp. 168-172. IEEE (2018)", + "9. Cole, E., Yang, X., Wilber, K., Mac Aodha, O., Belongie, S.: When does contrastive visual representation learning work? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14755-14764 (2022)", + "0. Cosentino, R., Sengupta, A., Avestimehr, S., Soltanolkotabi, M., Ortega, A., Willke, T., Tepper, M.: Toward a geometrical understanding of self-supervised contrastive learning. arXiv preprint arXiv:2205.06926 (2022)", + "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009)", + "2. Di Wu, S.L., Zang, Z., Wang, K., Shang, L., Sun, B., Li, H., Li, S.Z.: Align yourself: Self-supervised pre-training for fine-grained recognition via saliency alignment. arXiv preprint arXiv:2106.15788 2(7), 8 (2021)", + "3. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5414-5423 (2021)", + "4. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision. pp. 237-253. Springer (2022)" + ], + "bbox": [ + 225, + 231, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)", + "16. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "18. Houben, S., Stallkamp, J., Salmen, J., Schlipsing, M., Igel, C.: Detection of traffic signs in real-world images: The German Traffic Sign Detection Benchmark. In: International Joint Conference on Neural Networks. No. 1288 (2013)", + "19. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021)", + "20. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Learning where to learn in cross-view self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14451-14460 (2022)", + "21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8845-8855 (2021)", + "22. Jang, Y.K., Cho, N.I.: Self-supervised product quantization for deep unsupervised image retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12085-12094 (2021)", + "23. Jing, L., Vincent, P., LeCun, Y., Tian, Y.: Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348 (2021)", + "24. Kim, S., Bae, S., Yun, S.Y.: Coreset sampling from open-set for fine-grained self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7537-7547 (2023)", + "25. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)", + "26. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "27. Lee, H., Lee, K., Lee, K., Lee, H., Shin, J.: Improving transferability of representations via augmentation-aware self-supervision. Advances in Neural Information Processing Systems 34, 17710-17722 (2021)", + "28. Li, A.C., Efros, A.A., Pathak, D.: Understanding collapse in non-contrastive siamese representation learning. In: European Conference on Computer Vision. pp. 490-505. Springer (2022)", + "29. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)", + "30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "31. Peng, X., Wang, K., Zhu, Z., Wang, M., You, Y.: Crafting better contrastive views for siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16031-16040 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015). https://doi.org/10.1007/s11263-015-0816-y", + "33. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "34. Selvaraju, R.R., Desai, K., Johnson, J., Naik, N.: Casting your model: Learning to localize improves self-supervised representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11058-11067 (2021)", + "35. Shu, Y., van den Hengel, A., Liu, L.: Learning common rationale to improve self-supervised representation for fine-grained visual recognition problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11392-11401 (2023)", + "36. Shu, Y., Yu, B., Xu, H., Liu, L.: Improving fine-grained visual recognition in low data regimes via self-boosting attention mechanism. In: European Conference on Computer Vision. pp. 449-465. Springer (2022)", + "37. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016)", + "38. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "39. Wang, T., Isola, P.: Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In: International Conference on Machine Learning. pp. 9929-9939. PMLR (2020)", + "40. Wang, Z., Wang, Y., Hu, H., Li, P.: Contrastive learning with consistent representations. arXiv preprint arXiv:2302.01541 (2023)", + "41. Xiao, T., Wang, X., Efros, A.A., Darrell, T.: What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659 (2020)", + "42. Yao, Y., Ye, C., He, J., Elsayed, G.F.: Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23282-23291 (2023)", + "43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L.: Saga: Self-augmentation with guided attention for representation learning. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3463-3467. IEEE (2022)", + "44. Zhao, N., Wu, Z., Lau, R.W., Lin, S.: Distilling localization for self-supervised representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10990-10998 (2021)", + "45. Ziyin, L., Lubana, E.S., Ueda, M., Tanaka, H.: What shapes the loss landscape of self-supervised learning? arXiv preprint arXiv:2210.00638 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 743 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Self-Supervised Fine-Grained Visual Recognition", + "bbox": [ + 405, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_model.json b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..19f07f2a294e10a05c94e9ad702450c6931f8288 --- /dev/null +++ b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_model.json @@ -0,0 +1,2137 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.261, + 0.141, + 0.744, + 0.21 + ], + "angle": 0, + "content": "On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.235, + 0.763, + 0.267 + ], + "angle": 0, + "content": "Zihu Wang\\(^{1}\\), Lingqiao Liu\\(^{2}\\), Scott Ricardo Figueroa Weston\\(^{1}\\), Samuel Tian\\(^{3}\\), and Peng Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.278, + 0.763, + 0.349 + ], + "angle": 0, + "content": "1 University of California, Santa Barbara CA 93106, USA \n2 The University of Adelaide, Adelaide South Australia 5001, Australia \n3 Carnegie Mellon University, Pittsburgh PA 15213, USA \n{zihu_wang, scotricardo, lip}@ucsb.edu {lingqiao.liu}@adelaide.edu.au {samuel.tian7}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.393, + 0.744, + 0.657 + ], + "angle": 0, + "content": "Abstract. Self-Supervised Learning (SSL) has become a prominent approach for acquiring visual representations across various tasks, yet its application in fine-grained visual recognition (FGVR) is challenged by the intricate task of distinguishing subtle differences between categories. To overcome this, we introduce an novel strategy that boosts SSL's ability to extract critical discriminative features vital for FGVR. This approach creates synthesized data pairs to guide the model to focus on discriminative features critical for FGVR during SSL. We start by identifying non-discriminative features using two main criteria: features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss. We then introduce perturbations to these non-discriminative features while preserving discriminative ones. A decoder is employed to reconstruct images from both perturbed and original feature vectors to create data pairs. An encoder is trained on such generated data pairs to become invariant to variations in non-discriminative dimensions while focusing on discriminative features, thereby improving the model's performance in FGVR tasks. We demonstrate the promising FGVR performance of the proposed approach through extensive evaluation on a wide variety of datasets." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.671, + 0.741, + 0.7 + ], + "angle": 0, + "content": "Keywords: Self-Supervised representation learning \\(\\cdot\\) fine-grained visual recognition \\(\\cdot\\) learning from generated data" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.729, + 0.377, + 0.745 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In computer vision, Fine-grained Visual Recognition (FGVR) focuses on identifying subcategories of visual data, such as bird species [1, 38], aircraft variants [29], and vehicle models [25]. Different from studies on large-scaled general image datasets [26, 32, 37], FGVR tasks highlight the challenge of distinguishing subtle visual patterns." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Self-Supervised Learning (SSL) methods have recently largely advanced the domain of visual representation learning, circumventing the necessity for human-provided annotations. In SSL, many contrastive learning approaches achieve state-of-the-art performance by learning the similarities between data pairs derived from augmentations of identical source images [3,4,6,15,16,27,40]. These methods facilitate the transferability of the learned representations across a wide range of visual recognition problems [4,13,14,16]. Despite these advancements, it has been suggested that SSL may prioritize general visual similarities, instead of critical subtle features in FGVR tasks, which leads to SSL's 'coarse-label bias' [9]. Furthermore, recent studies [24,35,36] have highlighted a tendency among existing SSL methods to be distracted by task-irrelevant features, consequently failing to capture FGVR-relevant patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.485 + ], + "angle": 0, + "content": "To this end, we propose an innovative self-supervised learning strategy focusing on selectively extracting highly discriminative features while disregarding less informative, noisy ones. Our approach involves the generation of new contrastive data pairs from the latent feature space of the encoder, training the encoder to prioritize critical objects within these pairs. To facilitate this, a decoder is employed to generate data based on the latent feature space, reconstructing an image's feature vector and its perturbed counterpart to form each data pair. As the data pairs are generated to guide the encoder to learn key features and to be invariant to variations in non-discriminative features, in each feature vector, only those dimensions associated with non-discriminative patterns are perturbed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.761 + ], + "angle": 0, + "content": "Therefore, at the core of our methods are two non-discriminative feature identification techniques. Firstly, Grad-CAM [33] induced from the SSL loss to the latent feature space highlights dimensions relevant to FGVR [35,36]. Thus, we introduce greater perturbation to those less highlighted dimensions. Despite the conventional view that dimensional collapse in SSL—characterized by some encoder dimensions producing constant outputs—is undesirable [6, 23, 28], recent literature [10, 45] suggests that inducing such collapse in task-irrelevant feature dimensions yields beneficial outcomes. As shown in Fig. 2, our empirical studies indicate that in encoders pre-trained by SSL methods, there are always dimensions with low variance across the dataset which cannot effectively separate data from different categories. We thus treat these low-variance dimensions as task-irrelevant and introduce perturbations to them. The two aforementioned perturbation components are then combined and applied to the latent feature vector of each image. Images are then reconstructed from the perturbed and original feature vectors to form contrastive pairs for a contrastive loss [4, 30]. Such a framework encourages the encoder to learn the key features highlighted by Grad-CAM and to reduce variance and induce collapse in those non-discriminative dimensions with low variance across the dataset in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our proposed fine-grained feature learning method can be incorporated into various existing SSL methods. We use SimSiam [6] and MoCo v2 [16] as baseline methods and incorporate our proposed technique into these methods. Experiments across various fine-grained visual datasets show the effectiveness of our method. The proposed method provides a great improvement over baseline" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "methods. Our methods built on MoCo v2 outperforms existing state-of-the-art Self-Supervised fine-grained visual recognition methods in numerous downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.217, + 0.396, + 0.234 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.251, + 0.569, + 0.267 + ], + "angle": 0, + "content": "2.1 Self-Supervised Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Self-Supervised Learning (SSL) facilitates the learning of visual representations without the need for labeled data. Among various SSL methodologies, contrastive learning has emerged as a promising technique. With the InfoNCE loss [30] and its variants [3, 4, 6, 7, 16] being introduced as the objectives for optimization, contrastive approaches treat different views of the same image as positive data pairs, while views from different images are considered negative data pairs. The goal for the encoder is to minimize the distance between positive pairs and maximize it between negative pairs within its representation space [2, 4, 16, 39]. Methods such as BYOL [15] and SimSiam [6] rely exclusively on positive pairs. Additionally, the issue of dimensional collapse, where some encoder dimensions output constant values, is discussed in [6, 23, 28], along with proposed solutions to mitigate this phenomenon. Nonetheless, recent studies [10, 45] have shown that the collapse of dimensions associated with task-irrelevant features can enhance the performance in downstream visual recognition tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.512, + 0.766, + 0.529 + ], + "angle": 0, + "content": "2.2 Fine-Grained Visual Recognition in Self-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While encoders pre-trained by Self-Supervised Learning (SSL) methods demonstrate transferability and generalizability in many tasks [4, 6, 16, 21, 41], studies [9, 24, 35] reveal SSL's limitations in capturing essential features for Fine-Grained Visual Recognition (FGVR). To enhance SSL's capability in identifying critical features, several works concentrate on refining data augmentations. Approaches such as SAGA [43], CAST [34], and ContrastiveCrop [31] adopt attention-guided heatmaps to locate and better crop key objects in images. DiLo [44] introduces a novel augmentation by merging key image objects with different backgrounds to generate additional views. Contrary to methods that modify images directly, our approach involves perturbing feature vectors and generating realistic images from the latent feature space to enhance the encoder's discriminative capacity. Another line of research employs auxiliary neural networks connected to the encoder's convolutional layers for improving encoder's attention on salient regions. LEWEL [20] trains an additional head to adaptively aggregate features. Techniques such as CVSA [12] and [42] train a network to fit segmentation annotations or outputs of pre-trained saliency detectors. Similarly, LCR [35] and SAM [36] train the network to align with Grad-CAM, treating Grad-CAM as the ground truth for the encoder's attention maps. Our method proposes training the encoder on generated data pairs to learn critical features. In addition to Grad-CAM, we use dimension variance as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "a criterion for identifying non-discriminative features. Low-variance dimensions where data points across the dataset are not well separated are treated less crucial. Besides, SimCore [24] pre-trains an encoder on a target dataset, then using it to select more relevant data from a large-scaled dataset to expand the training set, upon which a new encoder is retrained for downstream tasks." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.248, + 0.641, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.43, + 0.443, + 0.444 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.248, + 0.788, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.43, + 0.726, + 0.443 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.574 + ], + "angle": 0, + "content": "Fig. 1: The overview of the proposed method. (a) Our method can be incorporated into various existing SSL methods. A decoder is utilized to generate images from both the original feature vector and its perturbed counterpart to form data pairs. The overall loss consists three terms: a conventional contrastive loss, a reconstruction loss (ensuring the decoder evolves with the encoder), and a proposed contrastive loss on the generated pairs. (b) We propose two techniques to identify and perturb non-discriminative features in a feature vector, i.e., features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.629, + 0.33, + 0.645 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.663, + 0.363, + 0.679 + ], + "angle": 0, + "content": "3.1 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Self-Supervised Contrastive Learning Without need for labels, self-supervised contrastive learning learns to represent data \\(\\mathbf{x} \\in \\mathbb{R}^m\\) in a lower dimensional space \\(\\mathbb{R}^n\\) by learning the similarities among data samples. Typically, in contrastive learning, the model consists two components, an encoder \\(f_{\\theta_e}: \\mathbb{R}^m \\to \\mathbb{R}^n\\) that maps data to a latent feature space \\(\\mathcal{V} \\subseteq \\mathbb{R}^n\\), and a projection head \\(g_{\\theta_p}: \\mathbb{R}^n \\to \\mathbb{R}^k\\) which projects the latent feature vectors in \\(\\mathcal{V} \\subseteq \\mathbb{R}^n\\) to a lower dimensional representation space \\(\\mathcal{Z} \\subseteq \\mathbb{R}^k\\) where the contrastive loss is applied. Formally, given a batch \\(\\mathcal{B}\\) of unlabeled data, every image \\(\\mathbf{x}\\) in it is augmented by two random augmentation \\(\\mathcal{T}_1\\) and \\(\\mathcal{T}_2\\) to acquire two views, i.e., \\(\\mathbf{x}' = \\mathcal{T}_1(\\mathbf{x})\\), \\(\\mathbf{x}'' = \\mathcal{T}_2(\\mathbf{x})\\). Views augmented from the same image are considered a positive" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "pair, while those acquired from different images form negative pairs. Two augmented views of all images form a new batch \\(\\mathcal{B}_a\\) which doubles the size of \\(\\mathcal{B}\\). The encoder and projection head are then used to represent the views in \\(\\mathcal{Z}\\), i.e., \\(\\mathbf{z}' = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}'))\\), \\(\\mathbf{z}'' = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}''))\\). A contrastive loss \\(\\mathcal{L}_C\\) is then defined in \\(\\mathcal{Z}\\) space:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.234, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} (\\mathbf {z} ^ {\\prime}, \\mathbf {z} ^ {\\prime \\prime}) = - \\log \\frac {\\exp (\\mathbf {z} ^ {\\prime} \\cdot \\mathbf {z} ^ {\\prime \\prime} / \\tau)}{\\sum_ {\\mathbf {z _ {i}} \\in \\mathcal {B} _ {a} , \\mathbf {z _ {i}} \\neq \\mathbf {z} ^ {\\prime} , \\mathbf {z _ {i}} \\neq \\mathbf {z} ^ {\\prime \\prime}} \\exp (\\mathbf {z} ^ {\\prime} \\cdot \\mathbf {z _ {i}} / \\tau)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.274, + 0.536, + 0.289 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.784, + 0.351 + ], + "angle": 0, + "content": "Although there are subtle differences between different contrastive methods, the contrastive loss are defined similarly. MoCo [5,16] introduces a large memory of negative representations. SimSiam [6] and BYOL [15] discard negative pairs and learn solely from positive pairs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.341, + 0.387 + ], + "angle": 0, + "content": "3.2 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.52 + ], + "angle": 0, + "content": "The overview of our method is illustrated in Fig. 1. The essence of the proposed method is learning discriminative fine-grained visual features from synthesized data pairs which are reconstructed from latent feature vectors by a decoder \\( h_{\\theta_d} \\). During training, to ensure that the decoder follows the evolution of the encoder, we use Mean Square Error (MSE) as the loss function to optimize the decoder. Our empirical studies show that the decoder produces images of better quality when it is trained on non-augmented images. The following reconstruction loss \\( \\mathcal{L}_R \\) is calculated for each image." + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.531, + 0.785, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\frac {1}{m} \\| \\mathbf {x} - \\hat {\\mathbf {x}} \\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.784, + 0.594 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} \\in \\mathbb{R}^m\\) is non-augmented image, and \\(\\hat{\\mathbf{x}} = h_{\\theta_d}(f_{\\theta_e}(\\mathbf{x}))\\) is the reconstruction of it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.785, + 0.685 + ], + "angle": 0, + "content": "In addition to producing \\(\\hat{\\mathbf{x}}\\), we generate \\(\\hat{\\mathbf{x}}_p\\) from \\(\\mathbf{v}_p \\in \\mathcal{V}\\), which is a perturbed version of \\(\\mathbf{v}\\) where non-discriminative features are perturbed. A positive data pair is formed between \\(\\hat{\\mathbf{x}}\\) and \\(\\hat{\\mathbf{x}}_p\\) from which the encoder learns discriminative features while disregarding task-irrelevant ones. In Sec. 3.3 and Sec. 3.4, we introduce two methods of identifying crucial discriminative dimensions in the latent feature space \\(\\mathcal{V}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.613, + 0.724 + ], + "angle": 0, + "content": "3.3 Identifying key dimensions via Grad-CAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Grad-CAM [33], a widely used saliency detection technique, uses the gradient of the target loss with respect to intermediate features of the network to produce an attention map highlighting regions in the features that contribute to minimizing the loss. As in our proposed self-supervised method, labels are not available during training, we thus choose the contrastive loss as the target. To identify important features within an image's feature vector \\(\\mathbf{v} = f_{\\theta_e}(\\mathbf{x}) \\in \\mathbb{R}^n\\), we form positive pair between \\(\\mathbf{x}\\) and an augmented view \\(\\mathbf{x}''\\) to calculate a contrastive loss" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_C(\\mathbf{z}, \\mathbf{z}^{\\prime \\prime})\\), where \\(\\mathbf{z} = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}))\\), \\(\\mathbf{z}^{\\prime \\prime} = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}^{\\prime \\prime}))\\). As it is shown in Fig. 1, the Grad-CAM score vector \\(\\pmb{\\eta} = \\{\\eta_i\\}_{i=1}^n \\in \\mathbb{R}^n\\) is calculated by gradient of the contrastive loss with respect to the feature vector \\(\\mathbf{v}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.203, + 0.786, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\eta_ {i} = \\operatorname {R e L U} \\left(\\frac {\\partial \\mathcal {L} _ {C} \\left(g _ {\\theta_ {p}} (\\mathbf {v}) , g _ {\\theta_ {p}} \\left(\\mathbf {v} ^ {\\prime \\prime}\\right)\\right)}{\\partial v _ {i}} \\cdot v _ {i}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.784, + 0.331 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{v} = f_{\\theta_e}(\\mathbf{x})\\), \\(\\mathbf{v}'' = f_{\\theta_e}(\\mathbf{x}'')\\). \\(v_i\\) denotes the \\(i_{th}\\) element of \\(\\mathbf{v}\\). The application of \\(\\mathrm{ReLU}(\\cdot)\\) makes \\(\\eta_i > 0\\) for all \\(i \\in \\{1,2,\\dots,n\\}\\). Note that the original Grad-CAM calculates gradient with respect to feature maps of the last convolutional layer. To measure the saliency of feature dimensions, we calculate gradient with respect to feature vectors. Higher Grad-CAM scores in \\(\\pmb{\\eta}\\) represent corresponding dimension's higher contribution to data discrimination in contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.376 + ], + "angle": 0, + "content": "With the Grad-CAM scores, random Gaussian noise is then introduced as perturbation to \\(\\mathbf{v}\\). We first scale all elements in \\(\\pmb{\\eta}\\) to [0,1] by min-max normalization." + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.389, + 0.786, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\bar {\\eta} _ {i} = \\frac {\\eta_ {i} - \\operatorname* {m i n} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\}}{\\operatorname* {m a x} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\} - \\operatorname* {m i n} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.424, + 0.785, + 0.469 + ], + "angle": 0, + "content": "where \\(\\bar{\\eta}_i\\) is the \\(i_{th}\\) element of the normalized Grad-CAM score vector \\(\\bar{\\eta}\\). After the normalization, a random Gaussian noise perturbation vector \\(\\tilde{\\mathbf{v}}^g\\in \\mathbb{R}^n\\) is calculated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.485, + 0.786, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {v}} ^ {g} = \\left\\{\\tilde {v} _ {i} ^ {g}: \\tilde {v} _ {i} ^ {g} \\sim \\mathcal {N} \\left(0, \\epsilon_ {g} \\cdot \\left(1 - \\bar {\\eta} _ {i}\\right)\\right) \\right\\} _ {i = 1} ^ {n} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.786, + 0.629 + ], + "angle": 0, + "content": "where \\(\\epsilon_{g}\\) is a hyperparameter that controls the standard deviation of Gaussian noise. In such a perturbation, all elements are sampled from i.i.d. zero-mean Gaussian distributions. Importantly, dimensions with lower normalized Grad-CAM scores \\(\\bar{\\eta}_i\\) receive noise from a Gaussian distribution with a greater standard deviation, implying a higher likelihood of more significant noise affecting dimensions with lower Grad-CAM scores. And on average, key dimensions with higher Grad-CAM scores are affected less which helps preserve crucial features in the original images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.65, + 0.756, + 0.667 + ], + "angle": 0, + "content": "3.4 Determining feature's task-relevance via dimension variance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.784, + 0.703 + ], + "angle": 0, + "content": "In addition to the feature perturbation technique elaborated in Sec. 3.3, we propose another technique to determine and perturb task-irrelevant dimensions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.809 + ], + "angle": 0, + "content": "In SSL, dimensional collapse is a phenomenon where some encoder dimensions output constant values [6,19,23,28]. As data points are not separated along such dimensions, these dimensions can not be used to perform downstream visual recognition. However, recent works [10, 45] suggest that collapse of dimensions which are related to downstream task-irrelevant features can be beneficial. In spite of the potential benefit, how to induce beneficial dimensional collapse is not illustrated by existing SSL studies." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "As it is illustrated in Fig. 2, our empirical studies show that, in the latent feature space of encoders pre-trained by SSL methods, typically, data points are" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.413, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.258, + 0.347, + 0.268 + ], + "angle": 0, + "content": "(a) CUB-200" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.144, + 0.594, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.258, + 0.535, + 0.267 + ], + "angle": 0, + "content": "(b) Stanford Cars" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.144, + 0.777, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.258, + 0.726, + 0.268 + ], + "angle": 0, + "content": "(c) FGVC-Aircraft" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.281, + 0.789, + 0.379 + ], + "angle": 0, + "content": "Fig. 2: An illustration of data distribution in the feature space of encoders pre-trained by MoCo v2 [5]. Blue and red dots represent feature vectors of two categories' data from 3 fine-grained datasets, CUB-200 [38], Stanford Cars [25], and FGVC-Aircraft [29]. \\(v_{min}\\) and \\(v_{max}\\) are the dimensions in the feature space where data has the minimal and maximal variance across the dataset. Probability density curve fitting of each category along each dimension is attached to the corresponding axis. Different classes are separated much better along \\(v_{max}\\) than \\(v_{min}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.65 + ], + "angle": 0, + "content": "not well separated along dimensions with low variance. Variance along these dimensions thus introduces noise to downstream classification. Therefore, we treat such dimensions as task-irrelevant and propose a technique to induce collapse in these dimensions to guide the encoder's to be invariant to variations of such features. This technique starts with estimating the dataset's variance along each feature dimension in a feature vector memory bank \\(\\mathbf{M} \\in \\mathbb{R}^{D \\times n}\\) of size \\(D\\). During training, whenever the encoder is provided with a batch of data, its feature vectors will be stored in the memory to replace the oldest batch of feature vectors in it. Variance of each dimension across the dataset can be approximated in \\(\\mathbf{M}\\) to get a variance vector \\(\\mathbf{s} = \\{s_i : \\sigma^2(\\bar{\\mathbf{w}}_i), i = 1, 2, \\dots, n\\} \\in \\mathbb{R}^n\\). Here \\(\\bar{\\mathbf{w}}_i \\in \\mathbb{R}^D\\) is the \\(\\ell_2\\)-normalized \\(i_{th}\\) column vector of \\(\\mathbf{M}\\). A feature represented by dimension \\(i\\) is considered less discriminative if its corresponding variance \\(s_i < \\kappa\\) where \\(\\kappa\\) is a threshold hyperparameter. To introduce random noise to those less discriminative dimensions, similar to Eq. (4), we first apply min-max normalization to \\(\\mathbf{s}\\) to acquire \\(\\bar{\\mathbf{s}} = \\{\\bar{s}_i\\}_{i=1}^n\\). We then calculate the random noise vector \\(\\tilde{\\mathbf{v}}^{var} = \\{\\tilde{v}_i^{var}\\}_{i=1}^n\\) to be applied." + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.66, + 0.787, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\tilde {v} _ {i} ^ {v a r} = \\left\\{ \\begin{array}{l l} u _ {i} \\sim \\mathcal {N} \\left(0, \\epsilon_ {v a r} \\cdot \\left(1 - \\bar {s} _ {i}\\right)\\right) & \\text {i f} s _ {i} < \\kappa \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.784, + 0.734 + ], + "angle": 0, + "content": "Here, \\(\\epsilon_{var}\\) defines the standard deviations of the i.i.d. Gaussian distributions. Similar to \\(\\tilde{\\mathbf{v}}^g\\), \\(\\tilde{\\mathbf{v}}^{var}\\) introduces greater noise to lower-variance dimensions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.588, + 0.772 + ], + "angle": 0, + "content": "3.5 Learning from reconstructed data pairs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "With the two feature perturbation techniques proposed in Sec. 3.3 and Sec. 3.4, we can finally perturb the feature vector \\(\\mathbf{v}\\) by adding the two random Gaussian noise to it to obtain its perturbed version \\(\\mathbf{v}_p\\), i.e., \\(\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g + \\tilde{\\mathbf{v}}^{var}\\). As it is shown in Fig. 1, \\(\\mathbf{v}\\) and \\(\\mathbf{v}_p\\) are reconstructed by the decoder to produce" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "\\(\\hat{\\mathbf{x}}\\) and \\(\\hat{\\mathbf{x}}_p\\), respectively. In \\(\\hat{\\mathbf{x}}_p\\), key patterns from \\(\\hat{\\mathbf{x}}\\) are preserved, while those contribute less to Grad-CAM or deemed less discriminative across the dataset by the low-variance criterion are perturbed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.255 + ], + "angle": 0, + "content": "We then form a positive pair between \\(\\hat{\\mathbf{x}}\\) and \\(\\hat{\\mathbf{x}}_p\\), addressing extracting key features and disregarding non-discriminative ones. To this end, we propose to pass the representations \\(\\hat{\\mathbf{z}} = g_{\\theta_p}(f_{\\theta_e}(\\hat{\\mathbf{x}}))\\) and \\(\\hat{\\mathbf{z}}_p = g_{\\theta_p}(f_{\\theta_e}(\\hat{\\mathbf{x}}_p))\\) to a contrastive loss \\(\\mathcal{L}_{C_p}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.266, + 0.785, + 0.301 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C _ {p}} (\\hat {\\mathbf {z}}, \\hat {\\mathbf {z}} _ {p}) = - \\log \\frac {\\exp (\\hat {\\mathbf {z}} \\cdot \\hat {\\mathbf {z}} _ {p} / \\tau)}{\\sum_ {\\hat {\\mathbf {z}} _ {\\mathrm {i}} \\in \\mathcal {B} _ {r} , \\hat {\\mathbf {z}} _ {\\mathrm {i}} \\neq \\hat {\\mathbf {z}} , \\hat {\\mathbf {z}} _ {\\mathrm {i}} \\neq \\hat {\\mathbf {z}} _ {p}} \\exp (\\hat {\\mathbf {z}} \\cdot \\hat {\\mathbf {z}} _ {\\mathrm {i}} / \\tau)} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.784, + 0.351 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}_r\\) is the set of all reconstructed and perturbed images from the current batch \\(\\mathcal{B}\\). By forming a positive pair between \\(\\hat{\\mathbf{x}}\\) and \\(\\hat{\\mathbf{x}}_p\\), Eq. (7) requires the encoder to be invariant to those perturbed less discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.352, + 0.685, + 0.366 + ], + "angle": 0, + "content": "Finally, we write the training loss \\(\\mathcal{L}\\) of our method as follows." + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.384, + 0.785, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {C} + \\alpha \\cdot \\mathcal {L} _ {R} + \\nu \\cdot \\mathcal {L} _ {C _ {p}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.436 + ], + "angle": 0, + "content": "\\(\\alpha\\) and \\(\\nu\\) are hyperparameters that control the weight of \\(\\mathcal{L}_R\\) and \\(\\mathcal{L}_{C_p}\\) in training. The pseudocode of our method is provided in Appendices." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Table 1: Performance comparison on three datasets. Our method is compared with MoCo v2 [5] and ResNet50 supervised pre-trained on ImageNet-1k [11]. Top-1 classification accuracy (in%) is reported when model is evaluated on \\(100\\%\\), \\(50\\%\\), and \\(20\\%\\) of all labels. Rank-1, rank-5, and mAP (in%) in image retrieval are reported." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.531, + 0.76, + 0.683 + ], + "angle": 0, + "content": "
DatasetMethodsClassificationImage Retrieval
100%50%20%rank-1rank-5mAP
CUB-200ResNet5063.0655.7142.2440.3968.9415.88
MoCo V263.9856.3542.6339.7267.1415.91
Ours66.1760.8449.6942.0669.5919.70
Stanford CarsResNet5061.4149.8533.5729.2854.667.01
MoCo V262.0251.0835.4430.5156.157.13
Ours65.6054.3640.2435.8161.9410.02
FGVC-AircraftResNet5049.7941.0233.6129.4851.3910.17
MoCo V251.1344.3436.4230.0252.8711.24
Ours55.2849.3741.1033.2756.8012.69
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.736, + 0.377, + 0.754 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.433, + 0.784 + ], + "angle": 0, + "content": "4.1 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Datasets. Experiments are conducted across five fine-grained visual datasets. We adopt on three widely used fine-grained datasets. Caltech-UCSD Birds 200 -2011 (CUB-200) dataset [38] contains 5994 training data and 5794 testing data" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "of 200 categories of birds. Stanford Cars (Cars) [25] has 196 classes of car models where 8144 data and 8041 data are in its training and testing split respectively. FGVC-Aircraft (Aircraft) [29] has 100 classes where 6667 images are for training and 3333 images are for testing. Additionally, we consider German Traffic Sign Recognition Benchmark (GTSRB) [18] which contains 43 classes of traffic signs. This dataset is usually used in autonomous driving and smart cities development. We take 4800 images for training and 3750 images for testing from GTSRB. We also evaluate the effectiveness of our method on ISIC2017 [8], a medical image dataset with 3 categories of skin lesion analysis where 2000 images are for training and 600 images are for testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Training Settings. All methods adopt ResNet-50 [17] as the encoder backbone where weights are initialized by loading ImageNet-1k [32] pre-trained model. Using two state-of-the-art SSL method, MoCo v2 [5] and SimSiam [6], as the baseline methods, we incorporate our proposed method in their framework. For the sake of fair comparison, encoders of all methods are pre-trained for 100 epochs. And the training batch size is set to 128 for all methods. More detailed encoder pre-training settings are provided in Appendices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Additionally, in our method, we use a feature vector memory bank of size \\( D = 5632 \\). For the training loss in Eq. (8), we choose \\( \\alpha = 1 \\) and \\( \\nu = 0.5 \\). When introducing noise described in Eq. (5) and Eq. (6), we choose \\( \\epsilon_g = 0.1 \\), \\( \\epsilon_{var} = 0.05 \\), and \\( \\kappa = 0.02 \\). To ensure the quality of reconstructed images, before encoder training, we freeze the encoder parameters and pre-train the decoder on target datasets by a reconstruction loss. We provide decoder pre-training details in Appendices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Performance Evaluation Protocols. Linear evaluation is a widely adopted protocol for assessing the performance of learned representations in visual recognition. This approach freezes the parameters of the pre-trained encoder and attaches a linear classifier to it. The classifier is then trained to perform classification. Our linear evaluation setup follows [16], detailed further in the Appendices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.778 + ], + "angle": 0, + "content": "The task of image retrieval [22, 35, 41] serves as another pivotal method for evaluating the performance of representation learning. Without adjusting any model parameters, it searches the nearest neighbors of a query image in the latent feature space for images that share the same label with the query image. The effectiveness of the evaluated model is quantified by recording the proportion of retrieved images that fall into the same categories as the query image. We present three commonly utilized metrics of retrieval performance: rank-1, rank-5, and mean Average Precision (mAP)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Furthermore, to illustrate the effect of the proposed method, attention maps generated by Grad-CAM on encoders trained by different methods are compared. Images from the training datasets, along with their corresponding reconstructed and perturbed versions, are also provided." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.149, + 0.72, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.451, + 0.785, + 0.481 + ], + "angle": 0, + "content": "Fig. 3: Grad-CAM attention visualized on images. Our proposed method is incorporated into MoCo v2 and SimSiam and compared with them." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.373, + 0.525 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.613 + ], + "angle": 0, + "content": "Comparison between the proposed method and baseline methods As among all configurations of our method, the one based on MoCo v2 achieves the best overall performance, this configuration is thus called 'ours' in the experiments. We first compare our method with MoCo v2 in Tab. 1. 'ResNet50' in Tab. 1 represents the encoder supervised pre-trained on ImageNet-1k [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.734 + ], + "angle": 0, + "content": "In linear evaluation, we evaluate pre-trained encoders using different proportions of labels. Across all three datasets, our proposed method achieves an average top-1 accuracy improvement of \\(3.31\\%\\), \\(4.27\\%\\), and \\(5.51\\%\\) over MoCo v2, with \\(100\\%\\), \\(50\\%\\), and \\(20\\%\\) of all labels, respectively. This advancement highlights the efficacy of our data pairs generation technique, enabling great performance in classification tasks. Notably, the remarkable improvement in label-insufficient scenarios highlights our method's ability to learn more generalizable features from unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We further evaluate the effectiveness of our method by image retrieval tasks. Performance in these tasks serves as a measure of semantic consistency in the learned latent feature space. Our method outperforms MoCo v2 on all datasets in rank-1, rank-5, and mAP. Invariant to irrelevant patterns, our approach ensures that images from the same category, which exhibit discriminative patterns, are more closely clustered in the feature space. This distribution enhances performance in image retrieval." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Performance comparison on ISIC2017 [8] and GTSRB [18]. Top-1 classification accuracy (in %) and rank-1 (in %) of image retrieval is reported." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.185, + 0.725, + 0.277 + ], + "angle": 0, + "content": "
MethodsISIC2017GTSRB
ClassificationRetrievalClassificationRetrieval
MoCo V2 [5]64.3366.3488.2597.46
Ours66.9267.9091.6798.59
SimSiam [6]58.8364.4681.1092.81
SimSiam + Ours63.7865.5586.9494.42
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.784, + 0.386 + ], + "angle": 0, + "content": "Additionally, we visualize Grad-CAM attention in MoCo v2 and our method in Fig. 3. Unlike MoCo v2, which may concentrate on background regions irrelevant to visual recognition, our method exhibits enhanced precision in identifying and focusing on pivotal objects within the images, effectively minimizing the influence of background distractions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.481 + ], + "angle": 0, + "content": "Furthermore, we also integrate the proposed technique into a state-of-the-art negative pair-free method SimSiam [6], denoted as 'SimSiam+ours' in our experiments. Our framework significantly outperforms the original SimSiam in both linear evaluation and image retrieval tasks, as demonstrated in Tab. 3. Attention comparison is also visualized by Grad-CAM and compared with SimSiam in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.561 + ], + "angle": 0, + "content": "To comprehensively assess the effectiveness of our proposed method, we utilize two more fine-grained datasets: the traffic sign visual dataset GTSRB and the medical image dataset ISIC2017. The results, as shown in Tab. 2, indicate that our approach enhances the performance of Self-Supervised Learning (SSL), demonstrating its potential in real-world applications for FGVR tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.595, + 0.784, + 0.636 + ], + "angle": 0, + "content": "Table 3: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.648, + 0.776, + 0.838 + ], + "angle": 0, + "content": "
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
supervised77.4688.6085.93---
Dino [3]16.7414.3312.07---
SimSiam [6]46.7545.7238.5216.2412.4518.49
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
DiLo [44]62.97-----
CVSA [12]63.02-----
LEWEL [20]64.5962.9151.9039.9132.3631.09
ContrastiveCrop [31]64.2363.2952.0439.8432.7130.37
SAM-SSL-Bilinear [36]64.9462.8552.8340.0833.1930.52
LCR [35]65.2463.9653.2241.2634.7431.55
SimSiam+Ours57.8053.6347.5024.6719.7224.56
Ours66.1765.6055.2842.0635.8133.27
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Comparison with state-of-the-art self-supervised FGVR methods In this section, we compare our proposed method against existing self-supervised learning (SSL) techniques renowned for their enhanced fine-grained visual recognition capabilities [12, 20, 31, 35, 36, 44], as detailed in Tab. 3. Methods like Dino [3], SimSiam [6], and MoCo v2 [16], which are not optimized for fine-grained feature extraction are also listed. Supervised training performance is included to provide a comprehensive comparison. Top-1 accuracy in linear evaluation and rank-1 in image retrieval tasks are reported." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.404 + ], + "angle": 0, + "content": "In Tab. 3, our proposed method achieves the best overall performance in both image retrieval and linear evaluation tasks. DiLo [44], CVSA [12], and ContrastiveCrop [31] innovate with novel data augmentation techniques which directly modify the original images. In contrary, our method generates more realistic images from the learned feature space, highlighting the learning of FGVR-related features. And unlike SAM [36] and LCR [35] which train an auxiliary network to directly fit the encoder's attention to Grad-CAM, our method learns from generated data to highlight dimensions with high Grad-CAM scores and introduce dimensional collapse to non-discriminative features." + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.434, + 0.656, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.808, + 0.788, + 0.838 + ], + "angle": 0, + "content": "Fig. 4: Generated data pairs on CUB-200, Stanford Cars, and FGVC-Aircraft. The original images are also included." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.388 + ], + "angle": 0, + "content": "Generated data pairs of the proposed method To better understand why the proposed technique enhances the fine-grained visual feature extraction capability, we show the generated data pairs from different datasets in Fig. 4. The perturbed images, as illustrated in Fig. 4, show that they largely retain the original data's key objects, with modifications primarily appearing as subtle changes in background or less important regions, e.g., changes of the tree's branch behind a bird, alterations in vehicle light's textures and adjustments in an airplane's exterior finish. These modifications do not affect the defining features of the subjects. Remarkably, some perturbations lead to entirely new objects that maintain the original's identity. For instance, transforming a side view of a car into a front view, or depicting an aircraft in flight from a grounded position. These images, obtained by modifying latent semantics of original images, are difficult to obtain through traditional data augmentation techniques defined in the original data space. They efficiently guide the encoder in identifying which features to prioritize and which to ignore, enhancing its learning performance in FGVR." + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.436, + 0.652, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.623, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Fig. 5: Performance comparison of encoders trained by \\(\\mathcal{L}_C + \\alpha \\cdot \\mathcal{L}_R\\) with respect to different \\(\\alpha\\) value (green solid line). Top-1 classification accuracy on Stanford Cars is reported. MoCo v2 (red dashed line) and our method (blue dashed line) are included for comparison." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.737, + 0.771, + 0.753 + ], + "angle": 0, + "content": "4.3 Effectiveness of the reconstruction loss in contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "As described in Eq. (8), our model's overall training loss, \\(\\mathcal{L}\\), includes a reconstruction loss term, \\(\\mathcal{L}_R\\). To assess \\(\\mathcal{L}_R\\)'s effect on self-supervised contrastive learning, we incorporate a decoder into MoCo v2 and train the encoder by a loss \\(\\mathcal{L}_C + \\alpha \\cdot \\mathcal{L}_R\\) on the Stanford Cars dataset, varying \\(\\alpha\\) in the loss function. The results, shown in Fig. 5, indicate that \\(\\alpha\\) values of 0.5 and 1.0 enhance top-1" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "classification accuracy the most over MoCo v2. Generally, \\(\\mathcal{L}_R\\) provides a modest improvement (less than \\(1\\%\\)) to Self-Supervised FGVC." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.198, + 0.725, + 0.215 + ], + "angle": 0, + "content": "4.4 Effectiveness of the two feature perturbation techniques" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.344 + ], + "angle": 0, + "content": "As detailed in Sec. 3.3 and Sec. 3.4, two non-discriminative feature perturbation techniques are proposed to introduce noise \\(\\tilde{\\mathbf{v}}^g\\) and \\(\\tilde{\\mathbf{v}}^{var}\\), respectively. We conduct further experiments to assess the effectiveness of each technique in identifying and perturbing task-irrelevant features. In Tab. 4, we evaluate three different configurations of our method: (1) Ours: \\(\\mathbf{v}_p\\) is obtained by adding both noise components to \\(\\mathbf{v}\\), i.e., \\(\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g +\\tilde{\\mathbf{v}}^{var}\\); (2) Ours - Grad-CAM: \\(\\mathbf{v}_p\\) is obtained by adding only the noise generated by the low Grad-CAM scores criterion, i.e., \\(\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g\\); (3) Ours - low-var: \\(\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^{var}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.404 + ], + "angle": 0, + "content": "As shown in Tab. 4, all configurations achieve competitive results comparing with existing state-of-the-art self-supervised FGVR methods. And when combining two noise components \\(\\tilde{\\mathbf{v}}^g\\) and \\(\\tilde{\\mathbf{v}}^{var}\\), our methods achieve the best overall performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.429, + 0.79, + 0.471 + ], + "angle": 0, + "content": "Table 4: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.482, + 0.77, + 0.581 + ], + "angle": 0, + "content": "
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
Ours - Grad-CAM66.0464.1854.1941.0235.0832.14
Ours - low-var65.9164.3453.1141.6934.5531.23
Ours66.1765.6055.2842.0635.8133.27
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.629, + 0.36, + 0.645 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To enhance the performance of Self-Supervised Learning (SSL) in Fine-grained Visual Recognition (FGVR) tasks, this paper introduces a novel approach where an encoder learns discriminative features from generated images. By introducing noise to features deemed non-discriminative by two proposed criteria, we generate synthetic data from both the original and perturbed feature vectors by a decoder, thus forming data pairs that emphasize learning key features for FGVR. Our approach outperforms existing methods across various datasets in many downstream tasks. While the proposed approach also offers a modest boost to SSL performance in non-fine-grained visual recognition tasks—as detailed in the Appendices—the gains are notably more substantial in FGVR contexts. The refinement of our methodology for application to large-scaled general datasets remains an avenue for future research works." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.177 + ], + "angle": 0, + "content": "Acknowledgment. This material is based upon work supported by the National Science Foundation under Grant No. 1956313." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.2, + 0.323, + 0.216 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.786, + 0.286 + ], + "angle": 0, + "content": "1. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.786, + 0.328 + ], + "angle": 0, + "content": "2. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.786, + 0.37 + ], + "angle": 0, + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.786, + 0.412 + ], + "angle": 0, + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.786, + 0.481 + ], + "angle": 0, + "content": "6. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "7. Chuang, C.Y., Robinson, J., Lin, Y.C., Torralba, A., Jegelka, S.: Debiased contrastive learning. Advances in neural information processing systems 33, 8765-8775 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.786, + 0.605 + ], + "angle": 0, + "content": "8. Codella, N.C., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). pp. 168-172. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.786, + 0.647 + ], + "angle": 0, + "content": "9. Cole, E., Yang, X., Wilber, K., Mac Aodha, O., Belongie, S.: When does contrastive visual representation learning work? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14755-14764 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "0. Cosentino, R., Sengupta, A., Avestimehr, S., Soltanolkotabi, M., Ortega, A., Willke, T., Tepper, M.: Toward a geometrical understanding of self-supervised contrastive learning. arXiv preprint arXiv:2205.06926 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.786, + 0.716 + ], + "angle": 0, + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.758 + ], + "angle": 0, + "content": "2. Di Wu, S.L., Zang, Z., Wang, K., Shang, L., Sun, B., Li, H., Li, S.Z.: Align yourself: Self-supervised pre-training for fine-grained recognition via saliency alignment. arXiv preprint arXiv:2106.15788 2(7), 8 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "3. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5414-5423 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "4. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision. pp. 237-253. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.232, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "16. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.288 + ], + "angle": 0, + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.331 + ], + "angle": 0, + "content": "18. Houben, S., Stallkamp, J., Salmen, J., Schlipsing, M., Igel, C.: Detection of traffic signs in real-world images: The German Traffic Sign Detection Benchmark. In: International Joint Conference on Neural Networks. No. 1288 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "19. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "20. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Learning where to learn in cross-view self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14451-14460 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.471 + ], + "angle": 0, + "content": "21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8845-8855 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "22. Jang, Y.K., Cho, N.I.: Self-supervised product quantization for deep unsupervised image retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12085-12094 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.543 + ], + "angle": 0, + "content": "23. Jing, L., Vincent, P., LeCun, Y., Tian, Y.: Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.585 + ], + "angle": 0, + "content": "24. Kim, S., Bae, S., Yun, S.Y.: Coreset sampling from open-set for fine-grained self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7537-7547 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "25. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.655 + ], + "angle": 0, + "content": "26. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "27. Lee, H., Lee, K., Lee, K., Lee, H., Shin, J.: Improving transferability of representations via augmentation-aware self-supervision. Advances in Neural Information Processing Systems 34, 17710-17722 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "28. Li, A.C., Efros, A.A., Pathak, D.: Understanding collapse in non-contrastive siamese representation learning. In: European Conference on Computer Vision. pp. 490-505. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.769 + ], + "angle": 0, + "content": "29. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Peng, X., Wang, K., Zhu, Z., Wang, M., You, Y.: Crafting better contrastive views for siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16031-16040 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Self-Supervised Fine-Grained Visual Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "32. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015). https://doi.org/10.1007/s11263-015-0816-y" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.788, + 0.259 + ], + "angle": 0, + "content": "33. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.301 + ], + "angle": 0, + "content": "34. Selvaraju, R.R., Desai, K., Johnson, J., Naik, N.: Casting your model: Learning to localize improves self-supervised representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11058-11067 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.357 + ], + "angle": 0, + "content": "35. Shu, Y., van den Hengel, A., Liu, L.: Learning common rationale to improve self-supervised representation for fine-grained visual recognition problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11392-11401 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.398 + ], + "angle": 0, + "content": "36. Shu, Y., Yu, B., Xu, H., Liu, L.: Improving fine-grained visual recognition in low data regimes via self-boosting attention mechanism. In: European Conference on Computer Vision. pp. 449-465. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.44 + ], + "angle": 0, + "content": "37. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.468 + ], + "angle": 0, + "content": "38. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.509 + ], + "angle": 0, + "content": "39. Wang, T., Isola, P.: Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In: International Conference on Machine Learning. pp. 9929-9939. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.536 + ], + "angle": 0, + "content": "40. Wang, Z., Wang, Y., Hu, H., Li, P.: Contrastive learning with consistent representations. arXiv preprint arXiv:2302.01541 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.564 + ], + "angle": 0, + "content": "41. Xiao, T., Wang, X., Efros, A.A., Darrell, T.: What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.619 + ], + "angle": 0, + "content": "42. Yao, Y., Ye, C., He, J., Elsayed, G.F.: Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23282-23291 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.675 + ], + "angle": 0, + "content": "43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L.: Saga: Self-augmentation with guided attention for representation learning. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3463-3467. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.717 + ], + "angle": 0, + "content": "44. Zhao, N., Wu, Z., Lau, R.W., Lin, S.: Distilling localization for self-supervised representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10990-10998 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.744 + ], + "angle": 0, + "content": "45. Ziyin, L., Lubana, E.S., Ueda, M., Tanaka, H.: What shapes the loss landscape of self-supervised learning? arXiv preprint arXiv:2210.00638 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.744 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_origin.pdf b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..752888f98211ce91e2767371f914a515094a2649 --- /dev/null +++ b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/6c9295fb-3a98-4cfe-ae7d-7698be6784b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f055117d122f53d1b9c322b57181793d31c601120c54a3803b42e94b317f898c +size 3081372 diff --git a/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/full.md b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28f4a63fcc8ae94e205cbb77201bc366caa8e508 --- /dev/null +++ b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/full.md @@ -0,0 +1,278 @@ +# On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition + +Zihu Wang $^{1}$ , Lingqiao Liu $^{2}$ , Scott Ricardo Figueroa Weston $^{1}$ , Samuel Tian $^{3}$ , and Peng Li $^{1}$ + +1 University of California, Santa Barbara CA 93106, USA +2 The University of Adelaide, Adelaide South Australia 5001, Australia +3 Carnegie Mellon University, Pittsburgh PA 15213, USA +{zihu_wang, scotricardo, lip}@ucsb.edu {lingqiao.liu}@adelaide.edu.au {samuel.tian7}@gmail.com + +Abstract. Self-Supervised Learning (SSL) has become a prominent approach for acquiring visual representations across various tasks, yet its application in fine-grained visual recognition (FGVR) is challenged by the intricate task of distinguishing subtle differences between categories. To overcome this, we introduce an novel strategy that boosts SSL's ability to extract critical discriminative features vital for FGVR. This approach creates synthesized data pairs to guide the model to focus on discriminative features critical for FGVR during SSL. We start by identifying non-discriminative features using two main criteria: features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss. We then introduce perturbations to these non-discriminative features while preserving discriminative ones. A decoder is employed to reconstruct images from both perturbed and original feature vectors to create data pairs. An encoder is trained on such generated data pairs to become invariant to variations in non-discriminative dimensions while focusing on discriminative features, thereby improving the model's performance in FGVR tasks. We demonstrate the promising FGVR performance of the proposed approach through extensive evaluation on a wide variety of datasets. + +Keywords: Self-Supervised representation learning $\cdot$ fine-grained visual recognition $\cdot$ learning from generated data + +# 1 Introduction + +In computer vision, Fine-grained Visual Recognition (FGVR) focuses on identifying subcategories of visual data, such as bird species [1, 38], aircraft variants [29], and vehicle models [25]. Different from studies on large-scaled general image datasets [26, 32, 37], FGVR tasks highlight the challenge of distinguishing subtle visual patterns. + +Self-Supervised Learning (SSL) methods have recently largely advanced the domain of visual representation learning, circumventing the necessity for human-provided annotations. In SSL, many contrastive learning approaches achieve state-of-the-art performance by learning the similarities between data pairs derived from augmentations of identical source images [3,4,6,15,16,27,40]. These methods facilitate the transferability of the learned representations across a wide range of visual recognition problems [4,13,14,16]. Despite these advancements, it has been suggested that SSL may prioritize general visual similarities, instead of critical subtle features in FGVR tasks, which leads to SSL's 'coarse-label bias' [9]. Furthermore, recent studies [24,35,36] have highlighted a tendency among existing SSL methods to be distracted by task-irrelevant features, consequently failing to capture FGVR-relevant patterns. + +To this end, we propose an innovative self-supervised learning strategy focusing on selectively extracting highly discriminative features while disregarding less informative, noisy ones. Our approach involves the generation of new contrastive data pairs from the latent feature space of the encoder, training the encoder to prioritize critical objects within these pairs. To facilitate this, a decoder is employed to generate data based on the latent feature space, reconstructing an image's feature vector and its perturbed counterpart to form each data pair. As the data pairs are generated to guide the encoder to learn key features and to be invariant to variations in non-discriminative features, in each feature vector, only those dimensions associated with non-discriminative patterns are perturbed. + +Therefore, at the core of our methods are two non-discriminative feature identification techniques. Firstly, Grad-CAM [33] induced from the SSL loss to the latent feature space highlights dimensions relevant to FGVR [35,36]. Thus, we introduce greater perturbation to those less highlighted dimensions. Despite the conventional view that dimensional collapse in SSL—characterized by some encoder dimensions producing constant outputs—is undesirable [6, 23, 28], recent literature [10, 45] suggests that inducing such collapse in task-irrelevant feature dimensions yields beneficial outcomes. As shown in Fig. 2, our empirical studies indicate that in encoders pre-trained by SSL methods, there are always dimensions with low variance across the dataset which cannot effectively separate data from different categories. We thus treat these low-variance dimensions as task-irrelevant and introduce perturbations to them. The two aforementioned perturbation components are then combined and applied to the latent feature vector of each image. Images are then reconstructed from the perturbed and original feature vectors to form contrastive pairs for a contrastive loss [4, 30]. Such a framework encourages the encoder to learn the key features highlighted by Grad-CAM and to reduce variance and induce collapse in those non-discriminative dimensions with low variance across the dataset in the latent space. + +Our proposed fine-grained feature learning method can be incorporated into various existing SSL methods. We use SimSiam [6] and MoCo v2 [16] as baseline methods and incorporate our proposed technique into these methods. Experiments across various fine-grained visual datasets show the effectiveness of our method. The proposed method provides a great improvement over baseline + +methods. Our methods built on MoCo v2 outperforms existing state-of-the-art Self-Supervised fine-grained visual recognition methods in numerous downstream tasks. + +# 2 Related Works + +# 2.1 Self-Supervised Contrastive Learning + +Self-Supervised Learning (SSL) facilitates the learning of visual representations without the need for labeled data. Among various SSL methodologies, contrastive learning has emerged as a promising technique. With the InfoNCE loss [30] and its variants [3, 4, 6, 7, 16] being introduced as the objectives for optimization, contrastive approaches treat different views of the same image as positive data pairs, while views from different images are considered negative data pairs. The goal for the encoder is to minimize the distance between positive pairs and maximize it between negative pairs within its representation space [2, 4, 16, 39]. Methods such as BYOL [15] and SimSiam [6] rely exclusively on positive pairs. Additionally, the issue of dimensional collapse, where some encoder dimensions output constant values, is discussed in [6, 23, 28], along with proposed solutions to mitigate this phenomenon. Nonetheless, recent studies [10, 45] have shown that the collapse of dimensions associated with task-irrelevant features can enhance the performance in downstream visual recognition tasks. + +# 2.2 Fine-Grained Visual Recognition in Self-Supervised Learning + +While encoders pre-trained by Self-Supervised Learning (SSL) methods demonstrate transferability and generalizability in many tasks [4, 6, 16, 21, 41], studies [9, 24, 35] reveal SSL's limitations in capturing essential features for Fine-Grained Visual Recognition (FGVR). To enhance SSL's capability in identifying critical features, several works concentrate on refining data augmentations. Approaches such as SAGA [43], CAST [34], and ContrastiveCrop [31] adopt attention-guided heatmaps to locate and better crop key objects in images. DiLo [44] introduces a novel augmentation by merging key image objects with different backgrounds to generate additional views. Contrary to methods that modify images directly, our approach involves perturbing feature vectors and generating realistic images from the latent feature space to enhance the encoder's discriminative capacity. Another line of research employs auxiliary neural networks connected to the encoder's convolutional layers for improving encoder's attention on salient regions. LEWEL [20] trains an additional head to adaptively aggregate features. Techniques such as CVSA [12] and [42] train a network to fit segmentation annotations or outputs of pre-trained saliency detectors. Similarly, LCR [35] and SAM [36] train the network to align with Grad-CAM, treating Grad-CAM as the ground truth for the encoder's attention maps. Our method proposes training the encoder on generated data pairs to learn critical features. In addition to Grad-CAM, we use dimension variance as + +a criterion for identifying non-discriminative features. Low-variance dimensions where data points across the dataset are not well separated are treated less crucial. Besides, SimCore [24] pre-trains an encoder on a target dataset, then using it to select more relevant data from a large-scaled dataset to expand the training set, upon which a new encoder is retrained for downstream tasks. + +![](images/27326784c35c0e697b96d6dffa584166d661b08c2002cee2e55ec4c899bfa391.jpg) +(a) + +![](images/8208488b42b70d05eb31277d521366e59433fc37c5e0dcbea0c7aa3d97f13164.jpg) +(b) +Fig. 1: The overview of the proposed method. (a) Our method can be incorporated into various existing SSL methods. A decoder is utilized to generate images from both the original feature vector and its perturbed counterpart to form data pairs. The overall loss consists three terms: a conventional contrastive loss, a reconstruction loss (ensuring the decoder evolves with the encoder), and a proposed contrastive loss on the generated pairs. (b) We propose two techniques to identify and perturb non-discriminative features in a feature vector, i.e., features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss. + +# 3 Method + +# 3.1 Background + +Self-Supervised Contrastive Learning Without need for labels, self-supervised contrastive learning learns to represent data $\mathbf{x} \in \mathbb{R}^m$ in a lower dimensional space $\mathbb{R}^n$ by learning the similarities among data samples. Typically, in contrastive learning, the model consists two components, an encoder $f_{\theta_e}: \mathbb{R}^m \to \mathbb{R}^n$ that maps data to a latent feature space $\mathcal{V} \subseteq \mathbb{R}^n$ , and a projection head $g_{\theta_p}: \mathbb{R}^n \to \mathbb{R}^k$ which projects the latent feature vectors in $\mathcal{V} \subseteq \mathbb{R}^n$ to a lower dimensional representation space $\mathcal{Z} \subseteq \mathbb{R}^k$ where the contrastive loss is applied. Formally, given a batch $\mathcal{B}$ of unlabeled data, every image $\mathbf{x}$ in it is augmented by two random augmentation $\mathcal{T}_1$ and $\mathcal{T}_2$ to acquire two views, i.e., $\mathbf{x}' = \mathcal{T}_1(\mathbf{x})$ , $\mathbf{x}'' = \mathcal{T}_2(\mathbf{x})$ . Views augmented from the same image are considered a positive + +pair, while those acquired from different images form negative pairs. Two augmented views of all images form a new batch $\mathcal{B}_a$ which doubles the size of $\mathcal{B}$ . The encoder and projection head are then used to represent the views in $\mathcal{Z}$ , i.e., $\mathbf{z}' = g_{\theta_p}(f_{\theta_e}(\mathbf{x}'))$ , $\mathbf{z}'' = g_{\theta_p}(f_{\theta_e}(\mathbf{x}''))$ . A contrastive loss $\mathcal{L}_C$ is then defined in $\mathcal{Z}$ space: + +$$ +\mathcal {L} _ {C} (\mathbf {z} ^ {\prime}, \mathbf {z} ^ {\prime \prime}) = - \log \frac {\exp (\mathbf {z} ^ {\prime} \cdot \mathbf {z} ^ {\prime \prime} / \tau)}{\sum_ {\mathbf {z _ {i}} \in \mathcal {B} _ {a} , \mathbf {z _ {i}} \neq \mathbf {z} ^ {\prime} , \mathbf {z _ {i}} \neq \mathbf {z} ^ {\prime \prime}} \exp (\mathbf {z} ^ {\prime} \cdot \mathbf {z _ {i}} / \tau)} \tag {1} +$$ + +where $\tau$ is the temperature hyperparameter. + +Although there are subtle differences between different contrastive methods, the contrastive loss are defined similarly. MoCo [5,16] introduces a large memory of negative representations. SimSiam [6] and BYOL [15] discard negative pairs and learn solely from positive pairs. + +# 3.2 Overview + +The overview of our method is illustrated in Fig. 1. The essence of the proposed method is learning discriminative fine-grained visual features from synthesized data pairs which are reconstructed from latent feature vectors by a decoder $h_{\theta_d}$ . During training, to ensure that the decoder follows the evolution of the encoder, we use Mean Square Error (MSE) as the loss function to optimize the decoder. Our empirical studies show that the decoder produces images of better quality when it is trained on non-augmented images. The following reconstruction loss $\mathcal{L}_R$ is calculated for each image. + +$$ +\mathcal {L} _ {R} = \frac {1}{m} \| \mathbf {x} - \hat {\mathbf {x}} \| _ {2} ^ {2} \tag {2} +$$ + +where $\mathbf{x} \in \mathbb{R}^m$ is non-augmented image, and $\hat{\mathbf{x}} = h_{\theta_d}(f_{\theta_e}(\mathbf{x}))$ is the reconstruction of it. + +In addition to producing $\hat{\mathbf{x}}$ , we generate $\hat{\mathbf{x}}_p$ from $\mathbf{v}_p \in \mathcal{V}$ , which is a perturbed version of $\mathbf{v}$ where non-discriminative features are perturbed. A positive data pair is formed between $\hat{\mathbf{x}}$ and $\hat{\mathbf{x}}_p$ from which the encoder learns discriminative features while disregarding task-irrelevant ones. In Sec. 3.3 and Sec. 3.4, we introduce two methods of identifying crucial discriminative dimensions in the latent feature space $\mathcal{V}$ . + +# 3.3 Identifying key dimensions via Grad-CAM + +Grad-CAM [33], a widely used saliency detection technique, uses the gradient of the target loss with respect to intermediate features of the network to produce an attention map highlighting regions in the features that contribute to minimizing the loss. As in our proposed self-supervised method, labels are not available during training, we thus choose the contrastive loss as the target. To identify important features within an image's feature vector $\mathbf{v} = f_{\theta_e}(\mathbf{x}) \in \mathbb{R}^n$ , we form positive pair between $\mathbf{x}$ and an augmented view $\mathbf{x}''$ to calculate a contrastive loss + +$\mathcal{L}_C(\mathbf{z}, \mathbf{z}^{\prime \prime})$ , where $\mathbf{z} = g_{\theta_p}(f_{\theta_e}(\mathbf{x}))$ , $\mathbf{z}^{\prime \prime} = g_{\theta_p}(f_{\theta_e}(\mathbf{x}^{\prime \prime}))$ . As it is shown in Fig. 1, the Grad-CAM score vector $\pmb{\eta} = \{\eta_i\}_{i=1}^n \in \mathbb{R}^n$ is calculated by gradient of the contrastive loss with respect to the feature vector $\mathbf{v}$ . + +$$ +\eta_ {i} = \operatorname {R e L U} \left(\frac {\partial \mathcal {L} _ {C} \left(g _ {\theta_ {p}} (\mathbf {v}) , g _ {\theta_ {p}} \left(\mathbf {v} ^ {\prime \prime}\right)\right)}{\partial v _ {i}} \cdot v _ {i}\right) \tag {3} +$$ + +Here, $\mathbf{v} = f_{\theta_e}(\mathbf{x})$ , $\mathbf{v}'' = f_{\theta_e}(\mathbf{x}'')$ . $v_i$ denotes the $i_{th}$ element of $\mathbf{v}$ . The application of $\mathrm{ReLU}(\cdot)$ makes $\eta_i > 0$ for all $i \in \{1,2,\dots,n\}$ . Note that the original Grad-CAM calculates gradient with respect to feature maps of the last convolutional layer. To measure the saliency of feature dimensions, we calculate gradient with respect to feature vectors. Higher Grad-CAM scores in $\pmb{\eta}$ represent corresponding dimension's higher contribution to data discrimination in contrastive learning. + +With the Grad-CAM scores, random Gaussian noise is then introduced as perturbation to $\mathbf{v}$ . We first scale all elements in $\pmb{\eta}$ to [0,1] by min-max normalization. + +$$ +\bar {\eta} _ {i} = \frac {\eta_ {i} - \operatorname* {m i n} \left\{\eta_ {j} , j = 1 , 2 , \dots , n \right\}}{\operatorname* {m a x} \left\{\eta_ {j} , j = 1 , 2 , \dots , n \right\} - \operatorname* {m i n} \left\{\eta_ {j} , j = 1 , 2 , \dots , n \right\}} \tag {4} +$$ + +where $\bar{\eta}_i$ is the $i_{th}$ element of the normalized Grad-CAM score vector $\bar{\eta}$ . After the normalization, a random Gaussian noise perturbation vector $\tilde{\mathbf{v}}^g\in \mathbb{R}^n$ is calculated as follows. + +$$ +\tilde {\mathbf {v}} ^ {g} = \left\{\tilde {v} _ {i} ^ {g}: \tilde {v} _ {i} ^ {g} \sim \mathcal {N} \left(0, \epsilon_ {g} \cdot \left(1 - \bar {\eta} _ {i}\right)\right) \right\} _ {i = 1} ^ {n} \tag {5} +$$ + +where $\epsilon_{g}$ is a hyperparameter that controls the standard deviation of Gaussian noise. In such a perturbation, all elements are sampled from i.i.d. zero-mean Gaussian distributions. Importantly, dimensions with lower normalized Grad-CAM scores $\bar{\eta}_i$ receive noise from a Gaussian distribution with a greater standard deviation, implying a higher likelihood of more significant noise affecting dimensions with lower Grad-CAM scores. And on average, key dimensions with higher Grad-CAM scores are affected less which helps preserve crucial features in the original images. + +# 3.4 Determining feature's task-relevance via dimension variance + +In addition to the feature perturbation technique elaborated in Sec. 3.3, we propose another technique to determine and perturb task-irrelevant dimensions. + +In SSL, dimensional collapse is a phenomenon where some encoder dimensions output constant values [6,19,23,28]. As data points are not separated along such dimensions, these dimensions can not be used to perform downstream visual recognition. However, recent works [10, 45] suggest that collapse of dimensions which are related to downstream task-irrelevant features can be beneficial. In spite of the potential benefit, how to induce beneficial dimensional collapse is not illustrated by existing SSL studies. + +As it is illustrated in Fig. 2, our empirical studies show that, in the latent feature space of encoders pre-trained by SSL methods, typically, data points are + +![](images/fdb9f0018841dd77ec83ea145e7e04a265c1965e3fe1630682532587eeaf7364.jpg) +(a) CUB-200 + +![](images/f137a9b8ffd2e101b08f94d2b992892d0b047bdcd4a3338fcbdc7e25d9a63c86.jpg) +(b) Stanford Cars +Fig. 2: An illustration of data distribution in the feature space of encoders pre-trained by MoCo v2 [5]. Blue and red dots represent feature vectors of two categories' data from 3 fine-grained datasets, CUB-200 [38], Stanford Cars [25], and FGVC-Aircraft [29]. $v_{min}$ and $v_{max}$ are the dimensions in the feature space where data has the minimal and maximal variance across the dataset. Probability density curve fitting of each category along each dimension is attached to the corresponding axis. Different classes are separated much better along $v_{max}$ than $v_{min}$ . + +![](images/5f36ce91d6ec23470844e58c3271456c866b03d698c2c85e871a0626e69c0a64.jpg) +(c) FGVC-Aircraft + +not well separated along dimensions with low variance. Variance along these dimensions thus introduces noise to downstream classification. Therefore, we treat such dimensions as task-irrelevant and propose a technique to induce collapse in these dimensions to guide the encoder's to be invariant to variations of such features. This technique starts with estimating the dataset's variance along each feature dimension in a feature vector memory bank $\mathbf{M} \in \mathbb{R}^{D \times n}$ of size $D$ . During training, whenever the encoder is provided with a batch of data, its feature vectors will be stored in the memory to replace the oldest batch of feature vectors in it. Variance of each dimension across the dataset can be approximated in $\mathbf{M}$ to get a variance vector $\mathbf{s} = \{s_i : \sigma^2(\bar{\mathbf{w}}_i), i = 1, 2, \dots, n\} \in \mathbb{R}^n$ . Here $\bar{\mathbf{w}}_i \in \mathbb{R}^D$ is the $\ell_2$ -normalized $i_{th}$ column vector of $\mathbf{M}$ . A feature represented by dimension $i$ is considered less discriminative if its corresponding variance $s_i < \kappa$ where $\kappa$ is a threshold hyperparameter. To introduce random noise to those less discriminative dimensions, similar to Eq. (4), we first apply min-max normalization to $\mathbf{s}$ to acquire $\bar{\mathbf{s}} = \{\bar{s}_i\}_{i=1}^n$ . We then calculate the random noise vector $\tilde{\mathbf{v}}^{var} = \{\tilde{v}_i^{var}\}_{i=1}^n$ to be applied. + +$$ +\tilde {v} _ {i} ^ {v a r} = \left\{ \begin{array}{l l} u _ {i} \sim \mathcal {N} \left(0, \epsilon_ {v a r} \cdot \left(1 - \bar {s} _ {i}\right)\right) & \text {i f} s _ {i} < \kappa \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {6} +$$ + +Here, $\epsilon_{var}$ defines the standard deviations of the i.i.d. Gaussian distributions. Similar to $\tilde{\mathbf{v}}^g$ , $\tilde{\mathbf{v}}^{var}$ introduces greater noise to lower-variance dimensions. + +# 3.5 Learning from reconstructed data pairs + +With the two feature perturbation techniques proposed in Sec. 3.3 and Sec. 3.4, we can finally perturb the feature vector $\mathbf{v}$ by adding the two random Gaussian noise to it to obtain its perturbed version $\mathbf{v}_p$ , i.e., $\mathbf{v}_p = \mathbf{v} + \tilde{\mathbf{v}}^g + \tilde{\mathbf{v}}^{var}$ . As it is shown in Fig. 1, $\mathbf{v}$ and $\mathbf{v}_p$ are reconstructed by the decoder to produce + +$\hat{\mathbf{x}}$ and $\hat{\mathbf{x}}_p$ , respectively. In $\hat{\mathbf{x}}_p$ , key patterns from $\hat{\mathbf{x}}$ are preserved, while those contribute less to Grad-CAM or deemed less discriminative across the dataset by the low-variance criterion are perturbed. + +We then form a positive pair between $\hat{\mathbf{x}}$ and $\hat{\mathbf{x}}_p$ , addressing extracting key features and disregarding non-discriminative ones. To this end, we propose to pass the representations $\hat{\mathbf{z}} = g_{\theta_p}(f_{\theta_e}(\hat{\mathbf{x}}))$ and $\hat{\mathbf{z}}_p = g_{\theta_p}(f_{\theta_e}(\hat{\mathbf{x}}_p))$ to a contrastive loss $\mathcal{L}_{C_p}$ . + +$$ +\mathcal {L} _ {C _ {p}} (\hat {\mathbf {z}}, \hat {\mathbf {z}} _ {p}) = - \log \frac {\exp (\hat {\mathbf {z}} \cdot \hat {\mathbf {z}} _ {p} / \tau)}{\sum_ {\hat {\mathbf {z}} _ {\mathrm {i}} \in \mathcal {B} _ {r} , \hat {\mathbf {z}} _ {\mathrm {i}} \neq \hat {\mathbf {z}} , \hat {\mathbf {z}} _ {\mathrm {i}} \neq \hat {\mathbf {z}} _ {p}} \exp (\hat {\mathbf {z}} \cdot \hat {\mathbf {z}} _ {\mathrm {i}} / \tau)} \tag {7} +$$ + +where $\mathcal{B}_r$ is the set of all reconstructed and perturbed images from the current batch $\mathcal{B}$ . By forming a positive pair between $\hat{\mathbf{x}}$ and $\hat{\mathbf{x}}_p$ , Eq. (7) requires the encoder to be invariant to those perturbed less discriminative features. + +Finally, we write the training loss $\mathcal{L}$ of our method as follows. + +$$ +\mathcal {L} = \mathcal {L} _ {C} + \alpha \cdot \mathcal {L} _ {R} + \nu \cdot \mathcal {L} _ {C _ {p}} \tag {8} +$$ + +$\alpha$ and $\nu$ are hyperparameters that control the weight of $\mathcal{L}_R$ and $\mathcal{L}_{C_p}$ in training. The pseudocode of our method is provided in Appendices. + +Table 1: Performance comparison on three datasets. Our method is compared with MoCo v2 [5] and ResNet50 supervised pre-trained on ImageNet-1k [11]. Top-1 classification accuracy (in%) is reported when model is evaluated on $100\%$ , $50\%$ , and $20\%$ of all labels. Rank-1, rank-5, and mAP (in%) in image retrieval are reported. + +
DatasetMethodsClassificationImage Retrieval
100%50%20%rank-1rank-5mAP
CUB-200ResNet5063.0655.7142.2440.3968.9415.88
MoCo V263.9856.3542.6339.7267.1415.91
Ours66.1760.8449.6942.0669.5919.70
Stanford CarsResNet5061.4149.8533.5729.2854.667.01
MoCo V262.0251.0835.4430.5156.157.13
Ours65.6054.3640.2435.8161.9410.02
FGVC-AircraftResNet5049.7941.0233.6129.4851.3910.17
MoCo V251.1344.3436.4230.0252.8711.24
Ours55.2849.3741.1033.2756.8012.69
+ +# 4 Experiments + +# 4.1 Experiment Settings + +Datasets. Experiments are conducted across five fine-grained visual datasets. We adopt on three widely used fine-grained datasets. Caltech-UCSD Birds 200 -2011 (CUB-200) dataset [38] contains 5994 training data and 5794 testing data + +of 200 categories of birds. Stanford Cars (Cars) [25] has 196 classes of car models where 8144 data and 8041 data are in its training and testing split respectively. FGVC-Aircraft (Aircraft) [29] has 100 classes where 6667 images are for training and 3333 images are for testing. Additionally, we consider German Traffic Sign Recognition Benchmark (GTSRB) [18] which contains 43 classes of traffic signs. This dataset is usually used in autonomous driving and smart cities development. We take 4800 images for training and 3750 images for testing from GTSRB. We also evaluate the effectiveness of our method on ISIC2017 [8], a medical image dataset with 3 categories of skin lesion analysis where 2000 images are for training and 600 images are for testing. + +Training Settings. All methods adopt ResNet-50 [17] as the encoder backbone where weights are initialized by loading ImageNet-1k [32] pre-trained model. Using two state-of-the-art SSL method, MoCo v2 [5] and SimSiam [6], as the baseline methods, we incorporate our proposed method in their framework. For the sake of fair comparison, encoders of all methods are pre-trained for 100 epochs. And the training batch size is set to 128 for all methods. More detailed encoder pre-training settings are provided in Appendices. + +Additionally, in our method, we use a feature vector memory bank of size $D = 5632$ . For the training loss in Eq. (8), we choose $\alpha = 1$ and $\nu = 0.5$ . When introducing noise described in Eq. (5) and Eq. (6), we choose $\epsilon_g = 0.1$ , $\epsilon_{var} = 0.05$ , and $\kappa = 0.02$ . To ensure the quality of reconstructed images, before encoder training, we freeze the encoder parameters and pre-train the decoder on target datasets by a reconstruction loss. We provide decoder pre-training details in Appendices. + +Performance Evaluation Protocols. Linear evaluation is a widely adopted protocol for assessing the performance of learned representations in visual recognition. This approach freezes the parameters of the pre-trained encoder and attaches a linear classifier to it. The classifier is then trained to perform classification. Our linear evaluation setup follows [16], detailed further in the Appendices. + +The task of image retrieval [22, 35, 41] serves as another pivotal method for evaluating the performance of representation learning. Without adjusting any model parameters, it searches the nearest neighbors of a query image in the latent feature space for images that share the same label with the query image. The effectiveness of the evaluated model is quantified by recording the proportion of retrieved images that fall into the same categories as the query image. We present three commonly utilized metrics of retrieval performance: rank-1, rank-5, and mean Average Precision (mAP). + +Furthermore, to illustrate the effect of the proposed method, attention maps generated by Grad-CAM on encoders trained by different methods are compared. Images from the training datasets, along with their corresponding reconstructed and perturbed versions, are also provided. + +![](images/12e8fe9d26da5ca54b89991606ebfea71352e106b4af5279b8235f15af5bde59.jpg) +Fig. 3: Grad-CAM attention visualized on images. Our proposed method is incorporated into MoCo v2 and SimSiam and compared with them. + +# 4.2 Main Results + +Comparison between the proposed method and baseline methods As among all configurations of our method, the one based on MoCo v2 achieves the best overall performance, this configuration is thus called 'ours' in the experiments. We first compare our method with MoCo v2 in Tab. 1. 'ResNet50' in Tab. 1 represents the encoder supervised pre-trained on ImageNet-1k [11]. + +In linear evaluation, we evaluate pre-trained encoders using different proportions of labels. Across all three datasets, our proposed method achieves an average top-1 accuracy improvement of $3.31\%$ , $4.27\%$ , and $5.51\%$ over MoCo v2, with $100\%$ , $50\%$ , and $20\%$ of all labels, respectively. This advancement highlights the efficacy of our data pairs generation technique, enabling great performance in classification tasks. Notably, the remarkable improvement in label-insufficient scenarios highlights our method's ability to learn more generalizable features from unlabeled data. + +We further evaluate the effectiveness of our method by image retrieval tasks. Performance in these tasks serves as a measure of semantic consistency in the learned latent feature space. Our method outperforms MoCo v2 on all datasets in rank-1, rank-5, and mAP. Invariant to irrelevant patterns, our approach ensures that images from the same category, which exhibit discriminative patterns, are more closely clustered in the feature space. This distribution enhances performance in image retrieval. + +Table 2: Performance comparison on ISIC2017 [8] and GTSRB [18]. Top-1 classification accuracy (in %) and rank-1 (in %) of image retrieval is reported. + +
MethodsISIC2017GTSRB
ClassificationRetrievalClassificationRetrieval
MoCo V2 [5]64.3366.3488.2597.46
Ours66.9267.9091.6798.59
SimSiam [6]58.8364.4681.1092.81
SimSiam + Ours63.7865.5586.9494.42
+ +Additionally, we visualize Grad-CAM attention in MoCo v2 and our method in Fig. 3. Unlike MoCo v2, which may concentrate on background regions irrelevant to visual recognition, our method exhibits enhanced precision in identifying and focusing on pivotal objects within the images, effectively minimizing the influence of background distractions. + +Furthermore, we also integrate the proposed technique into a state-of-the-art negative pair-free method SimSiam [6], denoted as 'SimSiam+ours' in our experiments. Our framework significantly outperforms the original SimSiam in both linear evaluation and image retrieval tasks, as demonstrated in Tab. 3. Attention comparison is also visualized by Grad-CAM and compared with SimSiam in Fig. 3. + +To comprehensively assess the effectiveness of our proposed method, we utilize two more fine-grained datasets: the traffic sign visual dataset GTSRB and the medical image dataset ISIC2017. The results, as shown in Tab. 2, indicate that our approach enhances the performance of Self-Supervised Learning (SSL), demonstrating its potential in real-world applications for FGVR tasks. + +Table 3: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported. + +
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
supervised77.4688.6085.93---
Dino [3]16.7414.3312.07---
SimSiam [6]46.7545.7238.5216.2412.4518.49
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
DiLo [44]62.97-----
CVSA [12]63.02-----
LEWEL [20]64.5962.9151.9039.9132.3631.09
ContrastiveCrop [31]64.2363.2952.0439.8432.7130.37
SAM-SSL-Bilinear [36]64.9462.8552.8340.0833.1930.52
LCR [35]65.2463.9653.2241.2634.7431.55
SimSiam+Ours57.8053.6347.5024.6719.7224.56
Ours66.1765.6055.2842.0635.8133.27
+ +Comparison with state-of-the-art self-supervised FGVR methods In this section, we compare our proposed method against existing self-supervised learning (SSL) techniques renowned for their enhanced fine-grained visual recognition capabilities [12, 20, 31, 35, 36, 44], as detailed in Tab. 3. Methods like Dino [3], SimSiam [6], and MoCo v2 [16], which are not optimized for fine-grained feature extraction are also listed. Supervised training performance is included to provide a comprehensive comparison. Top-1 accuracy in linear evaluation and rank-1 in image retrieval tasks are reported. + +In Tab. 3, our proposed method achieves the best overall performance in both image retrieval and linear evaluation tasks. DiLo [44], CVSA [12], and ContrastiveCrop [31] innovate with novel data augmentation techniques which directly modify the original images. In contrary, our method generates more realistic images from the learned feature space, highlighting the learning of FGVR-related features. And unlike SAM [36] and LCR [35] which train an auxiliary network to directly fit the encoder's attention to Grad-CAM, our method learns from generated data to highlight dimensions with high Grad-CAM scores and introduce dimensional collapse to non-discriminative features. + +![](images/bfd8b76c7cf61f0a01b53550519c520603a962e92a7afbdb51e62a49ba3e59dc.jpg) +Fig. 4: Generated data pairs on CUB-200, Stanford Cars, and FGVC-Aircraft. The original images are also included. + +Generated data pairs of the proposed method To better understand why the proposed technique enhances the fine-grained visual feature extraction capability, we show the generated data pairs from different datasets in Fig. 4. The perturbed images, as illustrated in Fig. 4, show that they largely retain the original data's key objects, with modifications primarily appearing as subtle changes in background or less important regions, e.g., changes of the tree's branch behind a bird, alterations in vehicle light's textures and adjustments in an airplane's exterior finish. These modifications do not affect the defining features of the subjects. Remarkably, some perturbations lead to entirely new objects that maintain the original's identity. For instance, transforming a side view of a car into a front view, or depicting an aircraft in flight from a grounded position. These images, obtained by modifying latent semantics of original images, are difficult to obtain through traditional data augmentation techniques defined in the original data space. They efficiently guide the encoder in identifying which features to prioritize and which to ignore, enhancing its learning performance in FGVR. + +![](images/02e36873604bea656674784d92eae2dfb0af2a9af4ef0ca3abedf042a36aba68.jpg) +Fig. 5: Performance comparison of encoders trained by $\mathcal{L}_C + \alpha \cdot \mathcal{L}_R$ with respect to different $\alpha$ value (green solid line). Top-1 classification accuracy on Stanford Cars is reported. MoCo v2 (red dashed line) and our method (blue dashed line) are included for comparison. + +# 4.3 Effectiveness of the reconstruction loss in contrastive learning + +As described in Eq. (8), our model's overall training loss, $\mathcal{L}$ , includes a reconstruction loss term, $\mathcal{L}_R$ . To assess $\mathcal{L}_R$ 's effect on self-supervised contrastive learning, we incorporate a decoder into MoCo v2 and train the encoder by a loss $\mathcal{L}_C + \alpha \cdot \mathcal{L}_R$ on the Stanford Cars dataset, varying $\alpha$ in the loss function. The results, shown in Fig. 5, indicate that $\alpha$ values of 0.5 and 1.0 enhance top-1 + +classification accuracy the most over MoCo v2. Generally, $\mathcal{L}_R$ provides a modest improvement (less than $1\%$ ) to Self-Supervised FGVC. + +# 4.4 Effectiveness of the two feature perturbation techniques + +As detailed in Sec. 3.3 and Sec. 3.4, two non-discriminative feature perturbation techniques are proposed to introduce noise $\tilde{\mathbf{v}}^g$ and $\tilde{\mathbf{v}}^{var}$ , respectively. We conduct further experiments to assess the effectiveness of each technique in identifying and perturbing task-irrelevant features. In Tab. 4, we evaluate three different configurations of our method: (1) Ours: $\mathbf{v}_p$ is obtained by adding both noise components to $\mathbf{v}$ , i.e., $\mathbf{v}_p = \mathbf{v} + \tilde{\mathbf{v}}^g +\tilde{\mathbf{v}}^{var}$ ; (2) Ours - Grad-CAM: $\mathbf{v}_p$ is obtained by adding only the noise generated by the low Grad-CAM scores criterion, i.e., $\mathbf{v}_p = \mathbf{v} + \tilde{\mathbf{v}}^g$ ; (3) Ours - low-var: $\mathbf{v}_p = \mathbf{v} + \tilde{\mathbf{v}}^{var}$ . + +As shown in Tab. 4, all configurations achieve competitive results comparing with existing state-of-the-art self-supervised FGVR methods. And when combining two noise components $\tilde{\mathbf{v}}^g$ and $\tilde{\mathbf{v}}^{var}$ , our methods achieve the best overall performance. + +Table 4: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported. + +
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
Ours - Grad-CAM66.0464.1854.1941.0235.0832.14
Ours - low-var65.9164.3453.1141.6934.5531.23
Ours66.1765.6055.2842.0635.8133.27
+ +# 5 Conclusion + +To enhance the performance of Self-Supervised Learning (SSL) in Fine-grained Visual Recognition (FGVR) tasks, this paper introduces a novel approach where an encoder learns discriminative features from generated images. By introducing noise to features deemed non-discriminative by two proposed criteria, we generate synthetic data from both the original and perturbed feature vectors by a decoder, thus forming data pairs that emphasize learning key features for FGVR. Our approach outperforms existing methods across various datasets in many downstream tasks. While the proposed approach also offers a modest boost to SSL performance in non-fine-grained visual recognition tasks—as detailed in the Appendices—the gains are notably more substantial in FGVR contexts. The refinement of our methodology for application to large-scaled general datasets remains an avenue for future research works. + +Acknowledgment. This material is based upon work supported by the National Science Foundation under Grant No. 1956313. + +# References + +1. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014) +2. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020) +3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +6. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021) +7. Chuang, C.Y., Robinson, J., Lin, Y.C., Torralba, A., Jegelka, S.: Debiased contrastive learning. Advances in neural information processing systems 33, 8765-8775 (2020) +8. Codella, N.C., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). pp. 168-172. IEEE (2018) +9. Cole, E., Yang, X., Wilber, K., Mac Aodha, O., Belongie, S.: When does contrastive visual representation learning work? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14755-14764 (2022) +0. Cosentino, R., Sengupta, A., Avestimehr, S., Soltanolkotabi, M., Ortega, A., Willke, T., Tepper, M.: Toward a geometrical understanding of self-supervised contrastive learning. arXiv preprint arXiv:2205.06926 (2022) +1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009) +2. Di Wu, S.L., Zang, Z., Wang, K., Shang, L., Sun, B., Li, H., Li, S.Z.: Align yourself: Self-supervised pre-training for fine-grained recognition via saliency alignment. arXiv preprint arXiv:2106.15788 2(7), 8 (2021) +3. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5414-5423 (2021) +4. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision. pp. 237-253. Springer (2022) + +15. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) +16. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +18. Houben, S., Stallkamp, J., Salmen, J., Schlipsing, M., Igel, C.: Detection of traffic signs in real-world images: The German Traffic Sign Detection Benchmark. In: International Joint Conference on Neural Networks. No. 1288 (2013) +19. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021) +20. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Learning where to learn in cross-view self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14451-14460 (2022) +21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8845-8855 (2021) +22. Jang, Y.K., Cho, N.I.: Self-supervised product quantization for deep unsupervised image retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12085-12094 (2021) +23. Jing, L., Vincent, P., LeCun, Y., Tian, Y.: Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348 (2021) +24. Kim, S., Bae, S., Yun, S.Y.: Coreset sampling from open-set for fine-grained self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7537-7547 (2023) +25. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013) +26. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +27. Lee, H., Lee, K., Lee, K., Lee, H., Shin, J.: Improving transferability of representations via augmentation-aware self-supervision. Advances in Neural Information Processing Systems 34, 17710-17722 (2021) +28. Li, A.C., Efros, A.A., Pathak, D.: Understanding collapse in non-contrastive siamese representation learning. In: European Conference on Computer Vision. pp. 490-505. Springer (2022) +29. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) +30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +31. Peng, X., Wang, K., Zhu, Z., Wang, M., You, Y.: Crafting better contrastive views for siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16031-16040 (2022) + +32. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015). https://doi.org/10.1007/s11263-015-0816-y +33. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +34. Selvaraju, R.R., Desai, K., Johnson, J., Naik, N.: Casting your model: Learning to localize improves self-supervised representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11058-11067 (2021) +35. Shu, Y., van den Hengel, A., Liu, L.: Learning common rationale to improve self-supervised representation for fine-grained visual recognition problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11392-11401 (2023) +36. Shu, Y., Yu, B., Xu, H., Liu, L.: Improving fine-grained visual recognition in low data regimes via self-boosting attention mechanism. In: European Conference on Computer Vision. pp. 449-465. Springer (2022) +37. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016) +38. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +39. Wang, T., Isola, P.: Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In: International Conference on Machine Learning. pp. 9929-9939. PMLR (2020) +40. Wang, Z., Wang, Y., Hu, H., Li, P.: Contrastive learning with consistent representations. arXiv preprint arXiv:2302.01541 (2023) +41. Xiao, T., Wang, X., Efros, A.A., Darrell, T.: What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659 (2020) +42. Yao, Y., Ye, C., He, J., Elsayed, G.F.: Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23282-23291 (2023) +43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L.: Saga: Self-augmentation with guided attention for representation learning. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3463-3467. IEEE (2022) +44. Zhao, N., Wu, Z., Lau, R.W., Lin, S.: Distilling localization for self-supervised representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10990-10998 (2021) +45. Ziyin, L., Lubana, E.S., Ueda, M., Tanaka, H.: What shapes the loss landscape of self-supervised learning? arXiv preprint arXiv:2210.00638 (2022) \ No newline at end of file diff --git a/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/images.zip b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0e000dc79199bc200533511af160931adfcc4e73 --- /dev/null +++ b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98400cd06b200add4724d05aa151e98b3d94546b93d9cd03842a52905b400353 +size 501595 diff --git a/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/layout.json b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..405568397c59b7cb74d39f7b129ca0bd45bbf610 --- /dev/null +++ b/2024/On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition/layout.json @@ -0,0 +1,9414 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "text", + "content": "On Learning Discriminative Features from Synthesized Data for Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "spans": [ + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "text", + "content": "Zihu Wang" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "text", + "content": ", Lingqiao Liu" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "text", + "content": ", Scott Ricardo Figueroa Weston" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "text", + "content": ", Samuel Tian" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "text", + "content": ", and Peng Li" + }, + { + "bbox": [ + 146, + 186, + 466, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 220, + 466, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 220, + 466, + 276 + ], + "spans": [ + { + "bbox": [ + 146, + 220, + 466, + 276 + ], + "type": "text", + "content": "1 University of California, Santa Barbara CA 93106, USA \n2 The University of Adelaide, Adelaide South Australia 5001, Australia \n3 Carnegie Mellon University, Pittsburgh PA 15213, USA \n{zihu_wang, scotricardo, lip}@ucsb.edu {lingqiao.liu}@adelaide.edu.au {samuel.tian7}@gmail.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 311, + 455, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 311, + 455, + 520 + ], + "spans": [ + { + "bbox": [ + 160, + 311, + 455, + 520 + ], + "type": "text", + "content": "Abstract. Self-Supervised Learning (SSL) has become a prominent approach for acquiring visual representations across various tasks, yet its application in fine-grained visual recognition (FGVR) is challenged by the intricate task of distinguishing subtle differences between categories. To overcome this, we introduce an novel strategy that boosts SSL's ability to extract critical discriminative features vital for FGVR. This approach creates synthesized data pairs to guide the model to focus on discriminative features critical for FGVR during SSL. We start by identifying non-discriminative features using two main criteria: features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss. We then introduce perturbations to these non-discriminative features while preserving discriminative ones. A decoder is employed to reconstruct images from both perturbed and original feature vectors to create data pairs. An encoder is trained on such generated data pairs to become invariant to variations in non-discriminative dimensions while focusing on discriminative features, thereby improving the model's performance in FGVR tasks. We demonstrate the promising FGVR performance of the proposed approach through extensive evaluation on a wide variety of datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 531, + 453, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 531, + 453, + 554 + ], + "spans": [ + { + "bbox": [ + 159, + 531, + 453, + 554 + ], + "type": "text", + "content": "Keywords: Self-Supervised representation learning " + }, + { + "bbox": [ + 159, + 531, + 453, + 554 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 531, + 453, + 554 + ], + "type": "text", + "content": " fine-grained visual recognition " + }, + { + "bbox": [ + 159, + 531, + 453, + 554 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 531, + 453, + 554 + ], + "type": "text", + "content": " learning from generated data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 577, + 230, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 230, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 230, + 590 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In computer vision, Fine-grained Visual Recognition (FGVR) focuses on identifying subcategories of visual data, such as bird species [1, 38], aircraft variants [29], and vehicle models [25]. Different from studies on large-scaled general image datasets [26, 32, 37], FGVR tasks highlight the challenge of distinguishing subtle visual patterns." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "Self-Supervised Learning (SSL) methods have recently largely advanced the domain of visual representation learning, circumventing the necessity for human-provided annotations. In SSL, many contrastive learning approaches achieve state-of-the-art performance by learning the similarities between data pairs derived from augmentations of identical source images [3,4,6,15,16,27,40]. These methods facilitate the transferability of the learned representations across a wide range of visual recognition problems [4,13,14,16]. Despite these advancements, it has been suggested that SSL may prioritize general visual similarities, instead of critical subtle features in FGVR tasks, which leads to SSL's 'coarse-label bias' [9]. Furthermore, recent studies [24,35,36] have highlighted a tendency among existing SSL methods to be distracted by task-irrelevant features, consequently failing to capture FGVR-relevant patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": "To this end, we propose an innovative self-supervised learning strategy focusing on selectively extracting highly discriminative features while disregarding less informative, noisy ones. Our approach involves the generation of new contrastive data pairs from the latent feature space of the encoder, training the encoder to prioritize critical objects within these pairs. To facilitate this, a decoder is employed to generate data based on the latent feature space, reconstructing an image's feature vector and its perturbed counterpart to form each data pair. As the data pairs are generated to guide the encoder to learn key features and to be invariant to variations in non-discriminative features, in each feature vector, only those dimensions associated with non-discriminative patterns are perturbed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 387, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 602 + ], + "type": "text", + "content": "Therefore, at the core of our methods are two non-discriminative feature identification techniques. Firstly, Grad-CAM [33] induced from the SSL loss to the latent feature space highlights dimensions relevant to FGVR [35,36]. Thus, we introduce greater perturbation to those less highlighted dimensions. Despite the conventional view that dimensional collapse in SSL—characterized by some encoder dimensions producing constant outputs—is undesirable [6, 23, 28], recent literature [10, 45] suggests that inducing such collapse in task-irrelevant feature dimensions yields beneficial outcomes. As shown in Fig. 2, our empirical studies indicate that in encoders pre-trained by SSL methods, there are always dimensions with low variance across the dataset which cannot effectively separate data from different categories. We thus treat these low-variance dimensions as task-irrelevant and introduce perturbations to them. The two aforementioned perturbation components are then combined and applied to the latent feature vector of each image. Images are then reconstructed from the perturbed and original feature vectors to form contrastive pairs for a contrastive loss [4, 30]. Such a framework encourages the encoder to learn the key features highlighted by Grad-CAM and to reduce variance and induce collapse in those non-discriminative dimensions with low variance across the dataset in the latent space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Our proposed fine-grained feature learning method can be incorporated into various existing SSL methods. We use SimSiam [6] and MoCo v2 [16] as baseline methods and incorporate our proposed technique into these methods. Experiments across various fine-grained visual datasets show the effectiveness of our method. The proposed method provides a great improvement over baseline" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "methods. Our methods built on MoCo v2 outperforms existing state-of-the-art Self-Supervised fine-grained visual recognition methods in numerous downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 171, + 242, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 242, + 185 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 242, + 185 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 198, + 348, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 348, + 211 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 348, + 211 + ], + "type": "text", + "content": "2.1 Self-Supervised Contrastive Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 219, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 387 + ], + "type": "text", + "content": "Self-Supervised Learning (SSL) facilitates the learning of visual representations without the need for labeled data. Among various SSL methodologies, contrastive learning has emerged as a promising technique. With the InfoNCE loss [30] and its variants [3, 4, 6, 7, 16] being introduced as the objectives for optimization, contrastive approaches treat different views of the same image as positive data pairs, while views from different images are considered negative data pairs. The goal for the encoder is to minimize the distance between positive pairs and maximize it between negative pairs within its representation space [2, 4, 16, 39]. Methods such as BYOL [15] and SimSiam [6] rely exclusively on positive pairs. Additionally, the issue of dimensional collapse, where some encoder dimensions output constant values, is discussed in [6, 23, 28], along with proposed solutions to mitigate this phenomenon. Nonetheless, recent studies [10, 45] have shown that the collapse of dimensions associated with task-irrelevant features can enhance the performance in downstream visual recognition tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 405, + 468, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 405, + 468, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 468, + 418 + ], + "type": "text", + "content": "2.2 Fine-Grained Visual Recognition in Self-Supervised Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "While encoders pre-trained by Self-Supervised Learning (SSL) methods demonstrate transferability and generalizability in many tasks [4, 6, 16, 21, 41], studies [9, 24, 35] reveal SSL's limitations in capturing essential features for Fine-Grained Visual Recognition (FGVR). To enhance SSL's capability in identifying critical features, several works concentrate on refining data augmentations. Approaches such as SAGA [43], CAST [34], and ContrastiveCrop [31] adopt attention-guided heatmaps to locate and better crop key objects in images. DiLo [44] introduces a novel augmentation by merging key image objects with different backgrounds to generate additional views. Contrary to methods that modify images directly, our approach involves perturbing feature vectors and generating realistic images from the latent feature space to enhance the encoder's discriminative capacity. Another line of research employs auxiliary neural networks connected to the encoder's convolutional layers for improving encoder's attention on salient regions. LEWEL [20] trains an additional head to adaptively aggregate features. Techniques such as CVSA [12] and [42] train a network to fit segmentation annotations or outputs of pre-trained saliency detectors. Similarly, LCR [35] and SAM [36] train the network to align with Grad-CAM, treating Grad-CAM as the ground truth for the encoder's attention maps. Our method proposes training the encoder on generated data pairs to learn critical features. In addition to Grad-CAM, we use dimension variance as" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "a criterion for identifying non-discriminative features. Low-variance dimensions where data points across the dataset are not well separated are treated less crucial. Besides, SimCore [24] pre-trains an encoder on a target dataset, then using it to select more relevant data from a large-scaled dataset to expand the training set, upon which a new encoder is retrained for downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 196, + 392, + 338 + ], + "blocks": [ + { + "bbox": [ + 133, + 196, + 392, + 338 + ], + "lines": [ + { + "bbox": [ + 133, + 196, + 392, + 338 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 392, + 338 + ], + "type": "image", + "image_path": "27326784c35c0e697b96d6dffa584166d661b08c2002cee2e55ec4c899bfa391.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 340, + 271, + 351 + ], + "lines": [ + { + "bbox": [ + 258, + 340, + 271, + 351 + ], + "spans": [ + { + "bbox": [ + 258, + 340, + 271, + 351 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 392, + 196, + 482, + 338 + ], + "blocks": [ + { + "bbox": [ + 392, + 196, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 392, + 196, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 392, + 196, + 482, + 338 + ], + "type": "image", + "image_path": "8208488b42b70d05eb31277d521366e59433fc37c5e0dcbea0c7aa3d97f13164.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 340, + 444, + 350 + ], + "lines": [ + { + "bbox": [ + 430, + 340, + 444, + 350 + ], + "spans": [ + { + "bbox": [ + 430, + 340, + 444, + 350 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 366, + 482, + 454 + ], + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 454 + ], + "type": "text", + "content": "Fig. 1: The overview of the proposed method. (a) Our method can be incorporated into various existing SSL methods. A decoder is utilized to generate images from both the original feature vector and its perturbed counterpart to form data pairs. The overall loss consists three terms: a conventional contrastive loss, a reconstruction loss (ensuring the decoder evolves with the encoder), and a proposed contrastive loss on the generated pairs. (b) We propose two techniques to identify and perturb non-discriminative features in a feature vector, i.e., features with low variance that fail to effectively separate data and those deemed less important by Grad-CAM induced from the SSL loss." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 498, + 201, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 201, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 201, + 510 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 525, + 222, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 222, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 222, + 537 + ], + "type": "text", + "content": "3.1 Background" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "Self-Supervised Contrastive Learning Without need for labels, self-supervised contrastive learning learns to represent data " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " in a lower dimensional space " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " by learning the similarities among data samples. Typically, in contrastive learning, the model consists two components, an encoder " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{\\theta_e}: \\mathbb{R}^m \\to \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " that maps data to a latent feature space " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\subseteq \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ", and a projection head " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "g_{\\theta_p}: \\mathbb{R}^n \\to \\mathbb{R}^k" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " which projects the latent feature vectors in " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\subseteq \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " to a lower dimensional representation space " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\subseteq \\mathbb{R}^k" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " where the contrastive loss is applied. Formally, given a batch " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " of unlabeled data, every image " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " in it is augmented by two random augmentation " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " to acquire two views, i.e., " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' = \\mathcal{T}_1(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'' = \\mathcal{T}_2(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ". Views augmented from the same image are considered a positive" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "pair, while those acquired from different images form negative pairs. Two augmented views of all images form a new batch " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_a" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " which doubles the size of " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". The encoder and projection head are then used to represent the views in " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{z}' = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}'))" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{z}'' = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}''))" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". A contrastive loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " is then defined in " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " space:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 185, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 185, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 198, + 185, + 481, + 213 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} (\\mathbf {z} ^ {\\prime}, \\mathbf {z} ^ {\\prime \\prime}) = - \\log \\frac {\\exp (\\mathbf {z} ^ {\\prime} \\cdot \\mathbf {z} ^ {\\prime \\prime} / \\tau)}{\\sum_ {\\mathbf {z _ {i}} \\in \\mathcal {B} _ {a} , \\mathbf {z _ {i}} \\neq \\mathbf {z} ^ {\\prime} , \\mathbf {z _ {i}} \\neq \\mathbf {z} ^ {\\prime \\prime}} \\exp (\\mathbf {z} ^ {\\prime} \\cdot \\mathbf {z _ {i}} / \\tau)} \\tag {1}", + "image_path": "ecdda6735af12c6bbbea00ecf3e525c17d5e61f064eb962888ccd328b8bb3793.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 328, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 328, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 328, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 217, + 328, + 228 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 217, + 328, + 228 + ], + "type": "text", + "content": " is the temperature hyperparameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "text", + "content": "Although there are subtle differences between different contrastive methods, the contrastive loss are defined similarly. MoCo [5,16] introduces a large memory of negative representations. SimSiam [6] and BYOL [15] discard negative pairs and learn solely from positive pairs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 295, + 208, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 208, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 208, + 306 + ], + "type": "text", + "content": "3.2 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "text", + "content": "The overview of our method is illustrated in Fig. 1. The essence of the proposed method is learning discriminative fine-grained visual features from synthesized data pairs which are reconstructed from latent feature vectors by a decoder " + }, + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "inline_equation", + "content": "h_{\\theta_d}" + }, + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "text", + "content": ". During training, to ensure that the decoder follows the evolution of the encoder, we use Mean Square Error (MSE) as the loss function to optimize the decoder. Our empirical studies show that the decoder produces images of better quality when it is trained on non-augmented images. The following reconstruction loss " + }, + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "text", + "content": " is calculated for each image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 269, + 420, + 480, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 420, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 269, + 420, + 480, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\frac {1}{m} \\| \\mathbf {x} - \\hat {\\mathbf {x}} \\| _ {2} ^ {2} \\tag {2}", + "image_path": "dfd7c41db42d9718dc4bafedc0aab4ccb7027b162448fc31876044334ec0c5a4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 447, + 479, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 479, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 479, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 447, + 479, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 447, + 479, + 470 + ], + "type": "text", + "content": " is non-augmented image, and " + }, + { + "bbox": [ + 130, + 447, + 479, + 470 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = h_{\\theta_d}(f_{\\theta_e}(\\mathbf{x}))" + }, + { + "bbox": [ + 130, + 447, + 479, + 470 + ], + "type": "text", + "content": " is the reconstruction of it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": "In addition to producing " + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_p" + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": ", which is a perturbed version of " + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": " where non-discriminative features are perturbed. A positive data pair is formed between " + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_p" + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": " from which the encoder learns discriminative features while disregarding task-irrelevant ones. In Sec. 3.3 and Sec. 3.4, we introduce two methods of identifying crucial discriminative dimensions in the latent feature space " + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 471, + 480, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 561, + 375, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 375, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 375, + 573 + ], + "type": "text", + "content": "3.3 Identifying key dimensions via Grad-CAM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "Grad-CAM [33], a widely used saliency detection technique, uses the gradient of the target loss with respect to intermediate features of the network to produce an attention map highlighting regions in the features that contribute to minimizing the loss. As in our proposed self-supervised method, labels are not available during training, we thus choose the contrastive loss as the target. To identify important features within an image's feature vector " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = f_{\\theta_e}(\\mathbf{x}) \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ", we form positive pair between " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": " and an augmented view " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}''" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": " to calculate a contrastive loss" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 448, + 103 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C(\\mathbf{z}, \\mathbf{z}^{\\prime \\prime})" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}))" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\prime \\prime} = g_{\\theta_p}(f_{\\theta_e}(\\mathbf{x}^{\\prime \\prime}))" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": ". As it is shown in Fig. 1, the Grad-CAM score vector " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta} = \\{\\eta_i\\}_{i=1}^n \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": " is calculated by gradient of the contrastive loss with respect to the feature vector " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 160, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 160, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 225, + 160, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\eta_ {i} = \\operatorname {R e L U} \\left(\\frac {\\partial \\mathcal {L} _ {C} \\left(g _ {\\theta_ {p}} (\\mathbf {v}) , g _ {\\theta_ {p}} \\left(\\mathbf {v} ^ {\\prime \\prime}\\right)\\right)}{\\partial v _ {i}} \\cdot v _ {i}\\right) \\tag {3}", + "image_path": "d048bd085f56580e6971ba5ed798f74b4d8380eff6d1892b4c9358500f99cf1f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = f_{\\theta_e}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{v}'' = f_{\\theta_e}(\\mathbf{x}'')" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": " element of " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": ". The application of " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{ReLU}(\\cdot)" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": " makes " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "\\eta_i > 0" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,n\\}" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": ". Note that the original Grad-CAM calculates gradient with respect to feature maps of the last convolutional layer. To measure the saliency of feature dimensions, we calculate gradient with respect to feature vectors. Higher Grad-CAM scores in " + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta}" + }, + { + "bbox": [ + 130, + 190, + 479, + 262 + ], + "type": "text", + "content": " represent corresponding dimension's higher contribution to data discrimination in contrastive learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "text", + "content": "With the Grad-CAM scores, random Gaussian noise is then introduced as perturbation to " + }, + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "text", + "content": ". We first scale all elements in " + }, + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta}" + }, + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "text", + "content": " to [0,1] by min-max normalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 308, + 481, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 308, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 187, + 308, + 481, + 333 + ], + "type": "interline_equation", + "content": "\\bar {\\eta} _ {i} = \\frac {\\eta_ {i} - \\operatorname* {m i n} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\}}{\\operatorname* {m a x} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\} - \\operatorname* {m i n} \\left\\{\\eta_ {j} , j = 1 , 2 , \\dots , n \\right\\}} \\tag {4}", + "image_path": "3b2d9c242e1a3ac238253295d9dcc4bf47f28d50089a52692c88b82c16881ed4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\bar{\\eta}_i" + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "text", + "content": " element of the normalized Grad-CAM score vector " + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\bar{\\eta}" + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "text", + "content": ". After the normalization, a random Gaussian noise perturbation vector " + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^g\\in \\mathbb{R}^n" + }, + { + "bbox": [ + 131, + 335, + 480, + 371 + ], + "type": "text", + "content": " is calculated as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 224, + 384, + 481, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 384, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 224, + 384, + 481, + 397 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {v}} ^ {g} = \\left\\{\\tilde {v} _ {i} ^ {g}: \\tilde {v} _ {i} ^ {g} \\sim \\mathcal {N} \\left(0, \\epsilon_ {g} \\cdot \\left(1 - \\bar {\\eta} _ {i}\\right)\\right) \\right\\} _ {i = 1} ^ {n} \\tag {5}", + "image_path": "2a5df228bef5d1b00600f1305d184b2652568d7c04f9c79df78ad4fff8d343c0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 401, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 481, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 401, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_{g}" + }, + { + "bbox": [ + 130, + 401, + 481, + 498 + ], + "type": "text", + "content": " is a hyperparameter that controls the standard deviation of Gaussian noise. In such a perturbation, all elements are sampled from i.i.d. zero-mean Gaussian distributions. Importantly, dimensions with lower normalized Grad-CAM scores " + }, + { + "bbox": [ + 130, + 401, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\bar{\\eta}_i" + }, + { + "bbox": [ + 130, + 401, + 481, + 498 + ], + "type": "text", + "content": " receive noise from a Gaussian distribution with a greater standard deviation, implying a higher likelihood of more significant noise affecting dimensions with lower Grad-CAM scores. And on average, key dimensions with higher Grad-CAM scores are affected less which helps preserve crucial features in the original images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 514, + 462, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 462, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 462, + 528 + ], + "type": "text", + "content": "3.4 Determining feature's task-relevance via dimension variance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 479, + 556 + ], + "type": "text", + "content": "In addition to the feature perturbation technique elaborated in Sec. 3.3, we propose another technique to determine and perturb task-irrelevant dimensions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": "In SSL, dimensional collapse is a phenomenon where some encoder dimensions output constant values [6,19,23,28]. As data points are not separated along such dimensions, these dimensions can not be used to perform downstream visual recognition. However, recent works [10, 45] suggest that collapse of dimensions which are related to downstream task-irrelevant features can be beneficial. In spite of the potential benefit, how to induce beneficial dimensional collapse is not illustrated by existing SSL studies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "As it is illustrated in Fig. 2, our empirical studies show that, in the latent feature space of encoders pre-trained by SSL methods, typically, data points are" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 252, + 200 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 252, + 200 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 252, + 200 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 252, + 200 + ], + "type": "image", + "image_path": "fdb9f0018841dd77ec83ea145e7e04a265c1965e3fe1630682532587eeaf7364.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 204, + 212, + 212 + ], + "lines": [ + { + "bbox": [ + 181, + 204, + 212, + 212 + ], + "spans": [ + { + "bbox": [ + 181, + 204, + 212, + 212 + ], + "type": "text", + "content": "(a) CUB-200" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 114, + 363, + 202 + ], + "blocks": [ + { + "bbox": [ + 256, + 114, + 363, + 202 + ], + "lines": [ + { + "bbox": [ + 256, + 114, + 363, + 202 + ], + "spans": [ + { + "bbox": [ + 256, + 114, + 363, + 202 + ], + "type": "image", + "image_path": "f137a9b8ffd2e101b08f94d2b992892d0b047bdcd4a3338fcbdc7e25d9a63c86.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 204, + 327, + 211 + ], + "lines": [ + { + "bbox": [ + 287, + 204, + 327, + 211 + ], + "spans": [ + { + "bbox": [ + 287, + 204, + 327, + 211 + ], + "type": "text", + "content": "(b) Stanford Cars" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "text", + "content": "Fig. 2: An illustration of data distribution in the feature space of encoders pre-trained by MoCo v2 [5]. Blue and red dots represent feature vectors of two categories' data from 3 fine-grained datasets, CUB-200 [38], Stanford Cars [25], and FGVC-Aircraft [29]. " + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "inline_equation", + "content": "v_{min}" + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "inline_equation", + "content": "v_{max}" + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "text", + "content": " are the dimensions in the feature space where data has the minimal and maximal variance across the dataset. Probability density curve fitting of each category along each dimension is attached to the corresponding axis. Different classes are separated much better along " + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "inline_equation", + "content": "v_{max}" + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "inline_equation", + "content": "v_{min}" + }, + { + "bbox": [ + 131, + 222, + 482, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 367, + 114, + 475, + 198 + ], + "blocks": [ + { + "bbox": [ + 367, + 114, + 475, + 198 + ], + "lines": [ + { + "bbox": [ + 367, + 114, + 475, + 198 + ], + "spans": [ + { + "bbox": [ + 367, + 114, + 475, + 198 + ], + "type": "image", + "image_path": "5f36ce91d6ec23470844e58c3271456c866b03d698c2c85e871a0626e69c0a64.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 204, + 444, + 212 + ], + "lines": [ + { + "bbox": [ + 400, + 204, + 444, + 212 + ], + "spans": [ + { + "bbox": [ + 400, + 204, + 444, + 212 + ], + "type": "text", + "content": "(c) FGVC-Aircraft" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": "not well separated along dimensions with low variance. Variance along these dimensions thus introduces noise to downstream classification. Therefore, we treat such dimensions as task-irrelevant and propose a technique to induce collapse in these dimensions to guide the encoder's to be invariant to variations of such features. This technique starts with estimating the dataset's variance along each feature dimension in a feature vector memory bank " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{D \\times n}" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": ". During training, whenever the encoder is provided with a batch of data, its feature vectors will be stored in the memory to replace the oldest batch of feature vectors in it. Variance of each dimension across the dataset can be approximated in " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " to get a variance vector " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = \\{s_i : \\sigma^2(\\bar{\\mathbf{w}}_i), i = 1, 2, \\dots, n\\} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{w}}_i \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": "-normalized " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " column vector of " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": ". A feature represented by dimension " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " is considered less discriminative if its corresponding variance " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "s_i < \\kappa" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " is a threshold hyperparameter. To introduce random noise to those less discriminative dimensions, similar to Eq. (4), we first apply min-max normalization to " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " to acquire " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{s}} = \\{\\bar{s}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": ". We then calculate the random noise vector " + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^{var} = \\{\\tilde{v}_i^{var}\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 323, + 482, + 514 + ], + "type": "text", + "content": " to be applied." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 522, + 481, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 205, + 522, + 481, + 555 + ], + "type": "interline_equation", + "content": "\\tilde {v} _ {i} ^ {v a r} = \\left\\{ \\begin{array}{l l} u _ {i} \\sim \\mathcal {N} \\left(0, \\epsilon_ {v a r} \\cdot \\left(1 - \\bar {s} _ {i}\\right)\\right) & \\text {i f} s _ {i} < \\kappa \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}", + "image_path": "8cbfad1ad4f1fcb12ea129ae7a2641c53c73f7e002637205a898c189b469f378.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\epsilon_{var}" + }, + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "text", + "content": " defines the standard deviations of the i.i.d. Gaussian distributions. Similar to " + }, + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^g" + }, + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^{var}" + }, + { + "bbox": [ + 131, + 558, + 479, + 581 + ], + "type": "text", + "content": " introduces greater noise to lower-variance dimensions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 598, + 359, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 359, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 359, + 611 + ], + "type": "text", + "content": "3.5 Learning from reconstructed data pairs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "With the two feature perturbation techniques proposed in Sec. 3.3 and Sec. 3.4, we can finally perturb the feature vector " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " by adding the two random Gaussian noise to it to obtain its perturbed version " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g + \\tilde{\\mathbf{v}}^{var}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". As it is shown in Fig. 1, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " are reconstructed by the decoder to produce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", respectively. In " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", key patterns from " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " are preserved, while those contribute less to Grad-CAM or deemed less discriminative across the dataset by the low-variance criterion are perturbed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "text", + "content": "We then form a positive pair between " + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_p" + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "text", + "content": ", addressing extracting key features and disregarding non-discriminative ones. To this end, we propose to pass the representations " + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}} = g_{\\theta_p}(f_{\\theta_e}(\\hat{\\mathbf{x}}))" + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_p = g_{\\theta_p}(f_{\\theta_e}(\\hat{\\mathbf{x}}_p))" + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "text", + "content": " to a contrastive loss " + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{C_p}" + }, + { + "bbox": [ + 130, + 152, + 480, + 201 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 210, + 480, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 210, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 201, + 210, + 480, + 238 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C _ {p}} (\\hat {\\mathbf {z}}, \\hat {\\mathbf {z}} _ {p}) = - \\log \\frac {\\exp (\\hat {\\mathbf {z}} \\cdot \\hat {\\mathbf {z}} _ {p} / \\tau)}{\\sum_ {\\hat {\\mathbf {z}} _ {\\mathrm {i}} \\in \\mathcal {B} _ {r} , \\hat {\\mathbf {z}} _ {\\mathrm {i}} \\neq \\hat {\\mathbf {z}} , \\hat {\\mathbf {z}} _ {\\mathrm {i}} \\neq \\hat {\\mathbf {z}} _ {p}} \\exp (\\hat {\\mathbf {z}} \\cdot \\hat {\\mathbf {z}} _ {\\mathrm {i}} / \\tau)} \\tag {7}", + "image_path": "6bd2038f19b204d1a9e852bdf7c11443f98983690dc20d87cd74e03ca5490b12.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_r" + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "content": " is the set of all reconstructed and perturbed images from the current batch " + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "content": ". By forming a positive pair between " + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_p" + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "content": ", Eq. (7) requires the encoder to be invariant to those perturbed less discriminative features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 278, + 419, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 278, + 419, + 289 + ], + "spans": [ + { + "bbox": [ + 146, + 278, + 419, + 289 + ], + "type": "text", + "content": "Finally, we write the training loss " + }, + { + "bbox": [ + 146, + 278, + 419, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 146, + 278, + 419, + 289 + ], + "type": "text", + "content": " of our method as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 304, + 480, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 304, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 249, + 304, + 480, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {C} + \\alpha \\cdot \\mathcal {L} _ {R} + \\nu \\cdot \\mathcal {L} _ {C _ {p}} \\tag {8}", + "image_path": "f59145b7577c3856d7205b64095fda8f9be72304a57bb10c75e2b216a4f2b0e0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "text", + "content": " are hyperparameters that control the weight of " + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{C_p}" + }, + { + "bbox": [ + 130, + 321, + 479, + 345 + ], + "type": "text", + "content": " in training. The pseudocode of our method is provided in Appendices." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 154, + 420, + 465, + 540 + ], + "blocks": [ + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "type": "text", + "content": "Table 1: Performance comparison on three datasets. Our method is compared with MoCo v2 [5] and ResNet50 supervised pre-trained on ImageNet-1k [11]. Top-1 classification accuracy (in%) is reported when model is evaluated on " + }, + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 366, + 480, + 411 + ], + "type": "text", + "content": " of all labels. Rank-1, rank-5, and mAP (in%) in image retrieval are reported." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 420, + 465, + 540 + ], + "lines": [ + { + "bbox": [ + 154, + 420, + 465, + 540 + ], + "spans": [ + { + "bbox": [ + 154, + 420, + 465, + 540 + ], + "type": "table", + "html": "
DatasetMethodsClassificationImage Retrieval
100%50%20%rank-1rank-5mAP
CUB-200ResNet5063.0655.7142.2440.3968.9415.88
MoCo V263.9856.3542.6339.7267.1415.91
Ours66.1760.8449.6942.0669.5919.70
Stanford CarsResNet5061.4149.8533.5729.2854.667.01
MoCo V262.0251.0835.4430.5156.157.13
Ours65.6054.3640.2435.8161.9410.02
FGVC-AircraftResNet5049.7941.0233.6129.4851.3910.17
MoCo V251.1344.3436.4230.0252.8711.24
Ours55.2849.3741.1033.2756.8012.69
", + "image_path": "5ebbd489159a4b0ae2c05fe2c6ecd4600be24e7a6f77e4949041a2fdf3443845.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 582, + 230, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 230, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 230, + 597 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 609, + 264, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 264, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 264, + 620 + ], + "type": "text", + "content": "4.1 Experiment Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Datasets. Experiments are conducted across five fine-grained visual datasets. We adopt on three widely used fine-grained datasets. Caltech-UCSD Birds 200 -2011 (CUB-200) dataset [38] contains 5994 training data and 5794 testing data" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "of 200 categories of birds. Stanford Cars (Cars) [25] has 196 classes of car models where 8144 data and 8041 data are in its training and testing split respectively. FGVC-Aircraft (Aircraft) [29] has 100 classes where 6667 images are for training and 3333 images are for testing. Additionally, we consider German Traffic Sign Recognition Benchmark (GTSRB) [18] which contains 43 classes of traffic signs. This dataset is usually used in autonomous driving and smart cities development. We take 4800 images for training and 3750 images for testing from GTSRB. We also evaluate the effectiveness of our method on ISIC2017 [8], a medical image dataset with 3 categories of skin lesion analysis where 2000 images are for training and 600 images are for testing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": "Training Settings. All methods adopt ResNet-50 [17] as the encoder backbone where weights are initialized by loading ImageNet-1k [32] pre-trained model. Using two state-of-the-art SSL method, MoCo v2 [5] and SimSiam [6], as the baseline methods, we incorporate our proposed method in their framework. For the sake of fair comparison, encoders of all methods are pre-trained for 100 epochs. And the training batch size is set to 128 for all methods. More detailed encoder pre-training settings are provided in Appendices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "content": "Additionally, in our method, we use a feature vector memory bank of size " + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "inline_equation", + "content": "D = 5632" + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "content": ". For the training loss in Eq. (8), we choose " + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\nu = 0.5" + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "content": ". When introducing noise described in Eq. (5) and Eq. (6), we choose " + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\epsilon_g = 0.1" + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\epsilon_{var} = 0.05" + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\kappa = 0.02" + }, + { + "bbox": [ + 130, + 348, + 482, + 433 + ], + "type": "text", + "content": ". To ensure the quality of reconstructed images, before encoder training, we freeze the encoder parameters and pre-train the decoder on target datasets by a reconstruction loss. We provide decoder pre-training details in Appendices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 458, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 518 + ], + "type": "text", + "content": "Performance Evaluation Protocols. Linear evaluation is a widely adopted protocol for assessing the performance of learned representations in visual recognition. This approach freezes the parameters of the pre-trained encoder and attaches a linear classifier to it. The classifier is then trained to perform classification. Our linear evaluation setup follows [16], detailed further in the Appendices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 520, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 616 + ], + "type": "text", + "content": "The task of image retrieval [22, 35, 41] serves as another pivotal method for evaluating the performance of representation learning. Without adjusting any model parameters, it searches the nearest neighbors of a query image in the latent feature space for images that share the same label with the query image. The effectiveness of the evaluated model is quantified by recording the proportion of retrieved images that fall into the same categories as the query image. We present three commonly utilized metrics of retrieval performance: rank-1, rank-5, and mean Average Precision (mAP)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, to illustrate the effect of the proposed method, attention maps generated by Grad-CAM on encoders trained by different methods are compared. Images from the training datasets, along with their corresponding reconstructed and perturbed versions, are also provided." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 118, + 440, + 347 + ], + "blocks": [ + { + "bbox": [ + 176, + 118, + 440, + 347 + ], + "lines": [ + { + "bbox": [ + 176, + 118, + 440, + 347 + ], + "spans": [ + { + "bbox": [ + 176, + 118, + 440, + 347 + ], + "type": "image", + "image_path": "12e8fe9d26da5ca54b89991606ebfea71352e106b4af5279b8235f15af5bde59.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 357, + 480, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 357, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 480, + 380 + ], + "type": "text", + "content": "Fig. 3: Grad-CAM attention visualized on images. Our proposed method is incorporated into MoCo v2 and SimSiam and compared with them." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 404, + 228, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 228, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 228, + 415 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 485 + ], + "type": "text", + "content": "Comparison between the proposed method and baseline methods As among all configurations of our method, the one based on MoCo v2 achieves the best overall performance, this configuration is thus called 'ours' in the experiments. We first compare our method with MoCo v2 in Tab. 1. 'ResNet50' in Tab. 1 represents the encoder supervised pre-trained on ImageNet-1k [11]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": "In linear evaluation, we evaluate pre-trained encoders using different proportions of labels. Across all three datasets, our proposed method achieves an average top-1 accuracy improvement of " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "3.31\\%" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "4.27\\%" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "5.51\\%" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": " over MoCo v2, with " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": " of all labels, respectively. This advancement highlights the efficacy of our data pairs generation technique, enabling great performance in classification tasks. Notably, the remarkable improvement in label-insufficient scenarios highlights our method's ability to learn more generalizable features from unlabeled data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "We further evaluate the effectiveness of our method by image retrieval tasks. Performance in these tasks serves as a measure of semantic consistency in the learned latent feature space. Our method outperforms MoCo v2 on all datasets in rank-1, rank-5, and mAP. Invariant to irrelevant patterns, our approach ensures that images from the same category, which exhibit discriminative patterns, are more closely clustered in the feature space. This distribution enhances performance in image retrieval." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 175, + 146, + 443, + 219 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Performance comparison on ISIC2017 [8] and GTSRB [18]. Top-1 classification accuracy (in %) and rank-1 (in %) of image retrieval is reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 146, + 443, + 219 + ], + "lines": [ + { + "bbox": [ + 175, + 146, + 443, + 219 + ], + "spans": [ + { + "bbox": [ + 175, + 146, + 443, + 219 + ], + "type": "table", + "html": "
MethodsISIC2017GTSRB
ClassificationRetrievalClassificationRetrieval
MoCo V2 [5]64.3366.3488.2597.46
Ours66.9267.9091.6798.59
SimSiam [6]58.8364.4681.1092.81
SimSiam + Ours63.7865.5586.9494.42
", + "image_path": "1f8df497a97bd69aa92724a393d9c83d4edda584b175113748bd414468cf5c02.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 479, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 305 + ], + "type": "text", + "content": "Additionally, we visualize Grad-CAM attention in MoCo v2 and our method in Fig. 3. Unlike MoCo v2, which may concentrate on background regions irrelevant to visual recognition, our method exhibits enhanced precision in identifying and focusing on pivotal objects within the images, effectively minimizing the influence of background distractions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 380 + ], + "type": "text", + "content": "Furthermore, we also integrate the proposed technique into a state-of-the-art negative pair-free method SimSiam [6], denoted as 'SimSiam+ours' in our experiments. Our framework significantly outperforms the original SimSiam in both linear evaluation and image retrieval tasks, as demonstrated in Tab. 3. Attention comparison is also visualized by Grad-CAM and compared with SimSiam in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 444 + ], + "type": "text", + "content": "To comprehensively assess the effectiveness of our proposed method, we utilize two more fine-grained datasets: the traffic sign visual dataset GTSRB and the medical image dataset ISIC2017. The results, as shown in Tab. 2, indicate that our approach enhances the performance of Self-Supervised Learning (SSL), demonstrating its potential in real-world applications for FGVR tasks." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 144, + 513, + 474, + 663 + ], + "blocks": [ + { + "bbox": [ + 132, + 471, + 479, + 503 + ], + "lines": [ + { + "bbox": [ + 132, + 471, + 479, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 479, + 503 + ], + "type": "text", + "content": "Table 3: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 513, + 474, + 663 + ], + "lines": [ + { + "bbox": [ + 144, + 513, + 474, + 663 + ], + "spans": [ + { + "bbox": [ + 144, + 513, + 474, + 663 + ], + "type": "table", + "html": "
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
supervised77.4688.6085.93---
Dino [3]16.7414.3312.07---
SimSiam [6]46.7545.7238.5216.2412.4518.49
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
DiLo [44]62.97-----
CVSA [12]63.02-----
LEWEL [20]64.5962.9151.9039.9132.3631.09
ContrastiveCrop [31]64.2363.2952.0439.8432.7130.37
SAM-SSL-Bilinear [36]64.9462.8552.8340.0833.1930.52
LCR [35]65.2463.9653.2241.2634.7431.55
SimSiam+Ours57.8053.6347.5024.6719.7224.56
Ours66.1765.6055.2842.0635.8133.27
", + "image_path": "0a46f1d4e2971cf37909159ed9bf0601b7c56af89377a9eb4156dbd38ae9f578.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "Comparison with state-of-the-art self-supervised FGVR methods In this section, we compare our proposed method against existing self-supervised learning (SSL) techniques renowned for their enhanced fine-grained visual recognition capabilities [12, 20, 31, 35, 36, 44], as detailed in Tab. 3. Methods like Dino [3], SimSiam [6], and MoCo v2 [16], which are not optimized for fine-grained feature extraction are also listed. Supervised training performance is included to provide a comprehensive comparison. Top-1 accuracy in linear evaluation and rank-1 in image retrieval tasks are reported." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": "In Tab. 3, our proposed method achieves the best overall performance in both image retrieval and linear evaluation tasks. DiLo [44], CVSA [12], and ContrastiveCrop [31] innovate with novel data augmentation techniques which directly modify the original images. In contrary, our method generates more realistic images from the learned feature space, highlighting the learning of FGVR-related features. And unlike SAM [36] and LCR [35] which train an auxiliary network to directly fit the encoder's attention to Grad-CAM, our method learns from generated data to highlight dimensions with high Grad-CAM scores and introduce dimensional collapse to non-discriminative features." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 212, + 343, + 401, + 632 + ], + "blocks": [ + { + "bbox": [ + 212, + 343, + 401, + 632 + ], + "lines": [ + { + "bbox": [ + 212, + 343, + 401, + 632 + ], + "spans": [ + { + "bbox": [ + 212, + 343, + 401, + 632 + ], + "type": "image", + "image_path": "bfd8b76c7cf61f0a01b53550519c520603a962e92a7afbdb51e62a49ba3e59dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 639, + 482, + 663 + ], + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 663 + ], + "type": "text", + "content": "Fig. 4: Generated data pairs on CUB-200, Stanford Cars, and FGVC-Aircraft. The original images are also included." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 307 + ], + "type": "text", + "content": "Generated data pairs of the proposed method To better understand why the proposed technique enhances the fine-grained visual feature extraction capability, we show the generated data pairs from different datasets in Fig. 4. The perturbed images, as illustrated in Fig. 4, show that they largely retain the original data's key objects, with modifications primarily appearing as subtle changes in background or less important regions, e.g., changes of the tree's branch behind a bird, alterations in vehicle light's textures and adjustments in an airplane's exterior finish. These modifications do not affect the defining features of the subjects. Remarkably, some perturbations lead to entirely new objects that maintain the original's identity. For instance, transforming a side view of a car into a front view, or depicting an aircraft in flight from a grounded position. These images, obtained by modifying latent semantics of original images, are difficult to obtain through traditional data augmentation techniques defined in the original data space. They efficiently guide the encoder in identifying which features to prioritize and which to ignore, enhancing its learning performance in FGVR." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 196, + 345, + 399, + 483 + ], + "blocks": [ + { + "bbox": [ + 196, + 345, + 399, + 483 + ], + "lines": [ + { + "bbox": [ + 196, + 345, + 399, + 483 + ], + "spans": [ + { + "bbox": [ + 196, + 345, + 399, + 483 + ], + "type": "image", + "image_path": "02e36873604bea656674784d92eae2dfb0af2a9af4ef0ca3abedf042a36aba68.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 493, + 482, + 537 + ], + "lines": [ + { + "bbox": [ + 132, + 493, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 482, + 537 + ], + "type": "text", + "content": "Fig. 5: Performance comparison of encoders trained by " + }, + { + "bbox": [ + 132, + 493, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C + \\alpha \\cdot \\mathcal{L}_R" + }, + { + "bbox": [ + 132, + 493, + 482, + 537 + ], + "type": "text", + "content": " with respect to different " + }, + { + "bbox": [ + 132, + 493, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 493, + 482, + 537 + ], + "type": "text", + "content": " value (green solid line). Top-1 classification accuracy on Stanford Cars is reported. MoCo v2 (red dashed line) and our method (blue dashed line) are included for comparison." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 583, + 471, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 583, + 471, + 596 + ], + "spans": [ + { + "bbox": [ + 133, + 583, + 471, + 596 + ], + "type": "text", + "content": "4.3 Effectiveness of the reconstruction loss in contrastive learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": "As described in Eq. (8), our model's overall training loss, " + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": ", includes a reconstruction loss term, " + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": ". To assess " + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": "'s effect on self-supervised contrastive learning, we incorporate a decoder into MoCo v2 and train the encoder by a loss " + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C + \\alpha \\cdot \\mathcal{L}_R" + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": " on the Stanford Cars dataset, varying " + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": " in the loss function. The results, shown in Fig. 5, indicate that " + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": " values of 0.5 and 1.0 enhance top-1" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "classification accuracy the most over MoCo v2. Generally, " + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": " provides a modest improvement (less than " + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": ") to Self-Supervised FGVC." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 156, + 443, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 156, + 443, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 156, + 443, + 170 + ], + "type": "text", + "content": "4.4 Effectiveness of the two feature perturbation techniques" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": "As detailed in Sec. 3.3 and Sec. 3.4, two non-discriminative feature perturbation techniques are proposed to introduce noise " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^g" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^{var}" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": ", respectively. We conduct further experiments to assess the effectiveness of each technique in identifying and perturbing task-irrelevant features. In Tab. 4, we evaluate three different configurations of our method: (1) Ours: " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " is obtained by adding both noise components to " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g +\\tilde{\\mathbf{v}}^{var}" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": "; (2) Ours - Grad-CAM: " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " is obtained by adding only the noise generated by the low Grad-CAM scores criterion, i.e., " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^g" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": "; (3) Ours - low-var: " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p = \\mathbf{v} + \\tilde{\\mathbf{v}}^{var}" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "text", + "content": "As shown in Tab. 4, all configurations achieve competitive results comparing with existing state-of-the-art self-supervised FGVR methods. And when combining two noise components " + }, + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^g" + }, + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}^{var}" + }, + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "text", + "content": ", our methods achieve the best overall performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 149, + 381, + 471, + 460 + ], + "blocks": [ + { + "bbox": [ + 131, + 339, + 483, + 373 + ], + "lines": [ + { + "bbox": [ + 131, + 339, + 483, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 483, + 373 + ], + "type": "text", + "content": "Table 4: Comparison with state-of-the-art self-supervised FGVC methods. Supervised training is also included. Top-1 classification accuracy (in %) and rank-1 image retrieval (in %) are reported." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 381, + 471, + 460 + ], + "lines": [ + { + "bbox": [ + 149, + 381, + 471, + 460 + ], + "spans": [ + { + "bbox": [ + 149, + 381, + 471, + 460 + ], + "type": "table", + "html": "
MethodClassificationImage Retrieval
CUB-200CarsAircraftCUB-200CarsAircraft
MoCo V2 [5]63.9862.0251.1339.7230.5130.02
Ours - Grad-CAM66.0464.1854.1941.0235.0832.14
Ours - low-var65.9164.3453.1141.6934.5531.23
Ours66.1765.6055.2842.0635.8133.27
", + "image_path": "2f21a9e68d02676597c04fab460d9735b6204dd7b9c08cd750d722498c75dcad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 498, + 220, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 220, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 220, + 510 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "To enhance the performance of Self-Supervised Learning (SSL) in Fine-grained Visual Recognition (FGVR) tasks, this paper introduces a novel approach where an encoder learns discriminative features from generated images. By introducing noise to features deemed non-discriminative by two proposed criteria, we generate synthetic data from both the original and perturbed feature vectors by a decoder, thus forming data pairs that emphasize learning key features for FGVR. Our approach outperforms existing methods across various datasets in many downstream tasks. While the proposed approach also offers a modest boost to SSL performance in non-fine-grained visual recognition tasks—as detailed in the Appendices—the gains are notably more substantial in FGVR contexts. The refinement of our methodology for application to large-scaled general datasets remains an avenue for future research works." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Acknowledgment. This material is based upon work supported by the National Science Foundation under Grant No. 1956313." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 197, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 197, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 197, + 171 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 481, + 226 + ], + "type": "text", + "content": "1. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "text", + "content": "2. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 260, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 481, + 293 + ], + "type": "text", + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "text", + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "text", + "content": "5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "text", + "content": "6. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "type": "text", + "content": "7. Chuang, C.Y., Robinson, J., Lin, Y.C., Torralba, A., Jegelka, S.: Debiased contrastive learning. Advances in neural information processing systems 33, 8765-8775 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 413, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 481, + 479 + ], + "type": "text", + "content": "8. Codella, N.C., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). pp. 168-172. IEEE (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "type": "text", + "content": "9. Cole, E., Yang, X., Wilber, K., Mac Aodha, O., Belongie, S.: When does contrastive visual representation learning work? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14755-14764 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "text", + "content": "0. Cosentino, R., Sengupta, A., Avestimehr, S., Soltanolkotabi, M., Ortega, A., Willke, T., Tepper, M.: Toward a geometrical understanding of self-supervised contrastive learning. arXiv preprint arXiv:2205.06926 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "text", + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "text", + "content": "2. Di Wu, S.L., Zang, Z., Wang, K., Shang, L., Sun, B., Li, H., Li, S.Z.: Align yourself: Self-supervised pre-training for fine-grained recognition via saliency alignment. arXiv preprint arXiv:2106.15788 2(7), 8 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "3. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5414-5423 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "4. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision. pp. 237-253. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "16. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "text", + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "text", + "content": "18. Houben, S., Stallkamp, J., Salmen, J., Schlipsing, M., Igel, C.: Detection of traffic signs in real-world images: The German Traffic Sign Detection Benchmark. In: International Joint Conference on Neural Networks. No. 1288 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "19. Hua, T., Wang, W., Xue, Z., Ren, S., Wang, Y., Zhao, H.: On feature decorrelation in self-supervised learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9598-9608 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "20. Huang, L., You, S., Zheng, M., Wang, F., Qian, C., Yamasaki, T.: Learning where to learn in cross-view self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14451-14460 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "text", + "content": "21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8845-8855 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "22. Jang, Y.K., Cho, N.I.: Self-supervised product quantization for deep unsupervised image retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12085-12094 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "type": "text", + "content": "23. Jing, L., Vincent, P., LeCun, Y., Tian, Y.: Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "text", + "content": "24. Kim, S., Bae, S., Yun, S.Y.: Coreset sampling from open-set for fine-grained self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7537-7547 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "25. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 518 + ], + "type": "text", + "content": "26. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "27. Lee, H., Lee, K., Lee, K., Lee, H., Shin, J.: Improving transferability of representations via augmentation-aware self-supervision. Advances in Neural Information Processing Systems 34, 17710-17722 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "text", + "content": "28. Li, A.C., Efros, A.A., Pathak, D.: Understanding collapse in non-contrastive siamese representation learning. In: European Conference on Computer Vision. pp. 490-505. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "29. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "text", + "content": "30. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Peng, X., Wang, K., Zhu, Z., Wang, M., You, Y.: Crafting better contrastive views for siamese representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16031-16040 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 589 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "text", + "content": "32. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015). https://doi.org/10.1007/s11263-015-0816-y" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 482, + 205 + ], + "type": "text", + "content": "33. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 238 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 238 + ], + "type": "text", + "content": "34. Selvaraju, R.R., Desai, K., Johnson, J., Naik, N.: Casting your model: Learning to localize improves self-supervised representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11058-11067 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 238, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 282 + ], + "type": "text", + "content": "35. Shu, Y., van den Hengel, A., Liu, L.: Learning common rationale to improve self-supervised representation for fine-grained visual recognition problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11392-11401 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "text", + "content": "36. Shu, Y., Yu, B., Xu, H., Liu, L.: Improving fine-grained visual recognition in low data regimes via self-boosting attention mechanism. In: European Conference on Computer Vision. pp. 449-465. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 482, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 348 + ], + "type": "text", + "content": "37. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2), 64-73 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 348, + 482, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 370 + ], + "type": "text", + "content": "38. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "type": "text", + "content": "39. Wang, T., Isola, P.: Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In: International Conference on Machine Learning. pp. 9929-9939. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 424 + ], + "type": "text", + "content": "40. Wang, Z., Wang, Y., Hu, H., Li, P.: Contrastive learning with consistent representations. arXiv preprint arXiv:2302.01541 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "type": "text", + "content": "41. Xiao, T., Wang, X., Efros, A.A., Darrell, T.: What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 446, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 490 + ], + "type": "text", + "content": "42. Yao, Y., Ye, C., He, J., Elsayed, G.F.: Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23282-23291 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 490, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 534 + ], + "type": "text", + "content": "43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L.: Saga: Self-augmentation with guided attention for representation learning. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3463-3467. IEEE (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 567 + ], + "type": "text", + "content": "44. Zhao, N., Wu, Z., Lau, R.W., Lin, S.: Distilling localization for self-supervised representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10990-10998 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "text", + "content": "45. Ziyin, L., Lubana, E.S., Ueda, M., Tanaka, H.: What shapes the loss landscape of self-supervised learning? arXiv preprint arXiv:2210.00638 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 447, + 102 + ], + "type": "text", + "content": "Self-Supervised Fine-Grained Visual Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_content_list.json b/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ae4323b77782fd0f4f35d45225129d6c068e872e --- /dev/null +++ b/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_content_list.json @@ -0,0 +1,1492 @@ +[ + { + "type": "text", + "text": "On Pretraining Data Diversity for Self-Supervised Learning", + "text_level": 1, + "bbox": [ + 326, + 140, + 674, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hasan Abed Al Kader Hammoud $^{1,2\\star\\diamond}$ , Tuhin Das $^{2\\star}$ , Fabio Pizzati $^{2\\star}$ Philip H.S. Torr $^{2}$ , and Adel Bibi $^{2\\dagger}$ , Bernard Ghanem $^{1\\dagger}$", + "bbox": [ + 256, + 210, + 743, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 KAUST", + "bbox": [ + 462, + 253, + 535, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Oxford", + "bbox": [ + 424, + 268, + 576, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We explore the impact of training with more diverse datasets, characterized by the number of unique samples, on the performance of self-supervised learning (SSL) under a fixed computational budget. Our findings demonstrate that increasing pretraining data diversity enhances SSL performance, albeit only when the distribution distance to the downstream data is minimal. Notably, even with an exceptionally large pretraining data diversity achieved through methods like web crawling or diffusion-generated data, among other ways, the distribution shift remains a challenge. Our experiments are comprehensive with seven SSL methods using large-scale datasets such as ImageNet and YFCC100M amounting to over 200 GPU days. The code and trained models will be available at https://github.com/hammoudhasan/DiversitySSL.", + "bbox": [ + 261, + 316, + 738, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: self-supervised learning $\\cdot$ distribution shift $\\cdot$ data diversity", + "bbox": [ + 261, + 497, + 736, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 558, + 374, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) has recently emerged as a new paradigm to pretrain large vision models at scale [26, 27, 47]. Leveraging the ability to learn from unlabelled data, pretraining on millions—or even billions [27, 53]—of images turned from an unachievable goal to a common practice. This exposure to extremely diverse datasets, i.e., composed of a remarkable number of unique samples, granted impressive performance and unprecedented generalization capabilities to a growing number of vision models [47]. Large-scale datasets, in conjunction with substantial computational resources, have been the key driving forces for the success of SSL-based approaches. For instance, SEER [26], pretrained for approximately 11 GPU years on a billion images, exemplifies the massive computation and data resources employed for these models. It has become the implicit norm that increasing computation and data is beneficial, without any detailed analysis of how they separately impact SSL effectiveness. In particular, it is not clear to what extent large datasets are responsible for the impressive generalization capabilities of SSL models. Indeed, consider the", + "bbox": [ + 212, + 589, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Equal supervision Ⓞ Work done during a research visit at Oxford", + "bbox": [ + 217, + 824, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9efeea3bf096c200ff57ab6b0029609544d4d0d2f34fea7b857437a9d115c3d2.jpg", + "image_caption": [ + "Fig. 1: Impact of Pretraining Diversity: Self-supervised learning (SSL) can be used to pretrain vision models on small datasets closely aligned to the downstream task, e.g., pets classification, hence with a small distribution shift (top, wild animals pretraining). Conversely, we could pretrain on an extensively varied dataset, with wide distribution differences (outdoor scenes, bottom). We study the role of pretraining diversity in SSL under a fixed budget, and highlight its effects in relationship to the distribution shift." + ], + "image_footnote": [], + "bbox": [ + 264, + 145, + 733, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "example in Figure 1. Assuming a fixed monetary budget, in the form of computational expenses, for pretraining a vision model under the SSL paradigm, does iterating over a large set of images work best given a downstream task, or it is better to train repeatedly on a smaller set of samples visually close to the ones of the downstream? In this context, we face the ever-lasting problem of distribution shift on vision models [58]. Without a proper understanding of the impact of data, there could be a wrong allocation of efforts in increasing data or computation, leading to inefficiencies in the deployment process of SSL models.", + "bbox": [ + 212, + 522, + 787, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we study the role of the diversity of the SSL pretraining data, under a fixed computational budget, when the pretraining data matches the distribution of the downstream data, as well as when they differ. Our experiments span various SSL methods [4,8-10, 14, 29, 65] and multiple datasets [5,18,38-40, 44,48,55,57,69], and are tested under several computational budgets amounting to a total of 200 GPU days. We summarize our contributions below:", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We show that SSL pretraining strategies are currently data-inefficient in compensating for distribution shifts. Under normalized computational costs, we verify that pretraining on large datasets with high diversity cannot outperform the pretraining on datasets with limited diversity, but with a lower distribution shift with respect to the downstream tasks.", + "2. We conclude that there is a wide margin for improvement in the performance of current SSL methods on out-of-distribution classes, and we propose in" + ], + "bbox": [ + 220, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sights for a fair evaluation. This shows the need for future research on new SSL techniques leveraging data diversity better, to improve generalization capabilities beyond the training data distribution.", + "bbox": [ + 240, + 146, + 782, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. We propose a novel strategy for a computationally-normalized evaluation of SSL and a carefully designed set of experiments focusing on pretraining data diversity, enabling us to draw our novel conclusions.", + "bbox": [ + 220, + 191, + 782, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ultimately, our work provides a comprehensive analysis, from experiments worth 200 GPU days on NVIDIA A100, on the interaction between data diversity and computational resources, and their impact on the performance of SSL models.", + "bbox": [ + 212, + 237, + 782, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 306, + 387, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning Early work in self-supervised learning used simple pretext tasks, such as relative patch prediction [19, 20], image colorization [67], image rotation prediction [25], or solving jigsaw puzzles [45] to train feature extractors in absence of annotations. Representations learned with those methods are limitedly effective, hence more recent literature has moved towards more sophisticated approaches, such as using image augmentations to generate correlated views of a training sample, and learning to extract augmentation-invariant representations for these correlated pairs. Among these multi-view methods, many exploit contrastive losses [1,8,10-12,32,35,36,42,43,46], enforcing similarity between views of the same image (positives) and dissimilarity from other images (negatives). Due to the need of many negative samples, contrastive methods often require large batch sizes to work effectively [10,14]. Cluster-based methods such as SwAV [8], DINO [9] and DeepCluster v2 [6] learn generalizable representations by grouping samples into cluster prototypes. Others exploit prediction of features with siamese networks [13], learn features in a teacher-student fashion [29], or use redundancy reduction techniques [4,65]. Finally, masked image modeling [3,31,63] emerged as a scalable alternative for Vision Transformers [21], learning representations by predicting masked image patches as SSL task.", + "bbox": [ + 212, + 340, + 787, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pretraining at Scale SSL pretraining is most effective at scale, benefiting from large datasets and great computational resources. Initial attempts at large-scale pretraining were made by combining contrastive learning and clustering [7, 56]. SEER [26] was trained on 1 billion internet images with SwAV [8] on a RegNet [49] backbone. The importance of model scaling to leverage large datasets is also shown in [27, 28, 64], as well as the need for longer training times [64]. Additional strategies are necessary to achieve best performance [17, 24, 66]. These works focus on reaching the best representations, without considerations about training costs, thus encouraging unfair comparisons across trainings. Preliminary papers [15, 22] found that increasing data may lead to a modest boost in downstream performance when the downstream distribution changes. While they give insights on the pretraining diversity, its relation with the distribution shift under normalized computation is still not investigated.", + "bbox": [ + 212, + 628, + 787, + 824 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distribution Shift in SSL Some studies have investigated how the pretraining data affects downstream performance on other domains. In general, the distribution shift increases when factors such as visual content and data acquisition procedure differ in two data ensembles [58]. For various pretraining datasets, preliminary works [15, 28, 37, 41] observed that the best performing representations were learned on datasets that were similar to the downstream test task. Additionally, combining datasets before pretraining or combining self-supervised features learned from various datasets did not lead to significant improvements in [15]. In [52, 68], they showcase higher generalization capabilities of SSL models compared to their supervised counterparts, for several downstream tasks in the presence of a distribution shift. In [59], they pretrained on several datasets, observing different generalization depending on the appearance of the pretraining dataset. Furthermore, initial considerations on the impact of external data on downstream tasks under distribution shift have been proposed in [60]. Some compensate the distribution shift with the scale of training datasets [30]. However, there is still a lack of a fair, computation-normalized evaluation that allows to study the effects of the distribution shift in a controlled environment.", + "bbox": [ + 212, + 145, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 426, + 382, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining We first outline the general pretraining procedure common to state-of-the-art self-supervised learning methods. Specific differences among these methods are detailed in the supplementary material. The overall pretraining pipeline, common across many SSL approaches [4, 8-10, 29, 65], goes as follows: (1) Data Sampling: from a large-scale, unlabeled upstream pretraining dataset, denoted as $\\mathbb{D}_{\\mathrm{SSL}}$ , an image $\\mathbf{x}$ is randomly sampled; (2) View Generation: two correlated views, $\\tilde{\\mathbf{x}}_A$ and $\\tilde{\\mathbf{x}}_B$ , are generated from $\\mathbf{x}$ using two image augmentation functions sampled at random. These random augmentations include random resized cropping, horizontal flipping, blurring, and color adjustments [2], among others; (3) Feature Extraction and Projection: the correlated views undergo feature extraction through a network, $f_{\\theta_f}$ , parameterized by $\\theta_f$ , such as a ResNet [33] or a ViT [21], leading to representations $\\mathbf{h}_A = f_{\\theta_f}(\\tilde{\\mathbf{x}}_A)$ and $\\mathbf{h}_B = f_{\\theta_f}(\\tilde{\\mathbf{x}}_B)$ . A linear projection head, $g_{\\theta_g}$ , parameterized by $\\theta_g$ , then maps these representations to a latent space, resulting in $\\mathbf{z}_A = g_{\\theta_g}(\\mathbf{h}_A)$ and $\\mathbf{z}_B = g_{\\theta_g}(\\mathbf{h}_B)$ ; (4) Joint Optimization: The feature extractor $f_{\\theta_f}$ and the projection head $g_{\\theta_g}$ are jointly optimized according to the following objective:", + "bbox": [ + 212, + 458, + 787, + 703 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {f} ^ {*}, \\theta_ {g} ^ {*} = \\underset {\\theta_ {f}, \\theta_ {g}} {\\arg \\min } \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {D} _ {\\mathrm {S S L}}} \\mathcal {L} _ {\\mathrm {S S L}} \\left(\\mathbf {z} _ {A}, \\mathbf {z} _ {B}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 714, + 785, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{SSL}}$ is a loss function specific to the SSL pretraining method. After pretraining, the feature extractor $f_{\\theta_f^*}$ can be deployed for various downstream tasks such as image classification, object detection, or segmentation. This is typically achieved by training a task-specific head. Alternatively, the feature extractor $f_{\\theta_f^*}$ can be fine-tuned or used with a $k$ -nearest neighbors (kNN) classifier.", + "bbox": [ + 212, + 747, + 787, + 825 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Linear Probing There are several ways to evaluate the performance of a self-supervised learning method such as linear probing [10, 14, 29], kNN [8, 9, 62, 71], and few-shot evaluation [23, 26]. Consistently with the general protocol established in the literature [10, 15, 32, 67], we use linear probing to measure the quality of the features extracted for classification tasks. The procedure is: (1) a labeled downstream dataset, $\\mathbb{D}_{\\mathrm{task}}$ , consisting of image-class pairs $(\\mathbf{x}, \\mathbf{y}) \\sim \\mathbb{D}_{\\mathrm{task}}$ , is selected for evaluation. (2) For each image $\\mathbf{x}$ , its representation is extracted using the pretrained feature extractor $f_{\\theta_f^*}$ , after which the linear classification head $t_{\\theta_t}$ , parameterized by $\\theta_t$ , is then applied to obtain $t_{\\theta_t}(f_{\\theta_f^*}(\\mathbf{x}))$ . (3) The linear head $t_{\\theta_t}$ is optimized as follows:", + "bbox": [ + 212, + 146, + 787, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} ^ {*} = \\underset {\\theta_ {t}} {\\arg \\min } \\mathbb {E} _ {(\\mathbf {x}, \\mathbf {y}) \\sim \\mathbb {D} _ {\\mathrm {t a s k}}} \\left[ \\mathcal {L} _ {\\mathrm {t a s k}} (t _ {\\theta_ {t}} (f _ {\\theta_ {f} ^ {*}} (\\mathbf {x})), \\mathbf {y}) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 309, + 785, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that only the parameters of the head $\\theta_t$ are optimized, while the parameters $\\theta_f^*$ of the feature extractor are frozen. The quality of features extracted by $f_{\\theta_f^*}$ is directly inferred from the classification accuracy achieved on the test set of $\\mathbb{D}_{\\mathrm{task}}$ , which serves as the primary indicator of the quality of the extracted features.", + "bbox": [ + 214, + 348, + 787, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Normalized Evaluation", + "text_level": 1, + "bbox": [ + 214, + 431, + 473, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We stress how for a correct understanding of the impact of data diversity we need to analyze its effects isolating them from the impact of increased computation. To enable this, we introduce (1) a computational budget used to normalize computation across experiments, and (2) a quantification of the data diversity seen by models during pretraining.", + "bbox": [ + 212, + 460, + 787, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computational Budget Current progress in SSL pretraining simultaneously scales computational budget and dataset size to achieve the best performance [26]. This makes it difficult to assess the reasons behind the success of many SSL algorithms. Do these methods benefit mainly from the large amounts of computation, i.e., running SSL pretraining for large numbers of epochs, or do they benefit from data diversity in larger datasets containing a vast variety of visual features? To perform a disentangled evaluation of these two factors, we first introduce $\\mathcal{C}$ as a measure of computational budget, which quantifies the total number of images an SSL method is allowed to process during pretraining. This is calculated as $\\mathcal{C} = N\\cdot \\mathcal{E}$ , where $N$ is the number of unique samples in the pretraining dataset $\\mathbb{D}_{\\mathrm{SSL}}$ , hence the data diversity of the dataset, and $\\mathcal{E}$ is the number of pretraining epochs. Constraining multiple models pretrained with SSL to a fixed computational budget $\\mathcal{C}$ allows for meaningful comparison, as it is guaranteed that all SSL methods will have processed the same number of pretraining images.", + "bbox": [ + 212, + 551, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantifying Pretraining Diversity Even under normalized computation $\\mathcal{C}$ , various SSL approaches may be exposed to different data diversity as they are trained with different datasets. For instance, a model trained for $\\mathcal{E} = 1000$ epochs on a dataset of size $N = 1000$ will see less diversity than a model pretrained for", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f9334e48c8e74747622928a0d93e597863a98139da79078362a45d2722b6b237.jpg", + "table_caption": [ + "Table 1: Impact of Data Diversity on CIFAR-100 and Tiny ImageNet SSL Pretraining Performance: We study the effects of diversity on CIFAR-100 a and Tiny ImageNet b across seven different methods and three data diversity settings for a ResNet-18 pretraining, where for all, $\\mathbb{D}_{\\mathrm{task}} = \\mathbb{D}_{\\mathrm{SSL}}$ . This comparison includes analyzing classification accuracy through linear probing on the training set and evaluation on the test set of the respective datasets. Although performance fluctuates among different methods, a consistent trend is observed: higher data diversity typically leads to the generation of higher quality representations." + ], + "table_footnote": [ + "$\\mathcal{C} = 50\\times 10^{6}$", + "(a) CIFAR-100" + ], + "table_body": "
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
50.138.044.341.114.144.112.515.7
250.550.154.351.045.450.451.635.9
501.058.658.458.356.455.855.640.9
", + "bbox": [ + 217, + 267, + 496, + 333 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/44088994b57550bc6a6946212e8c6c9eb3533885ba6a90ecb43f469d970516eb.jpg", + "table_caption": [], + "table_footnote": [ + "$\\mathcal{C} = 50\\times 10^{6}$", + "(b) Tiny ImageNet" + ], + "table_body": "
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
100.236.941.035.634.237.636.634.5
501.048.852.048.143.648.846.544.4
1002.049.855.650.347.554.148.647.3
", + "bbox": [ + 503, + 268, + 782, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{E} = 1$ epoch on a dataset of size $N = 10^6$ , despite that they are both pretrained under normalized computation, processing the same number of images. Hence, to capture this notion of exposure to diversity while pretraining under normalized computation $\\mathcal{C}$ , we define a pretraining diversity $\\mathcal{D}$ , which captures the number of unique samples encountered during training given a fixed $\\mathcal{C}$ as $\\mathcal{D} = N / \\mathcal{C} = 1 / \\varepsilon$ . A model pretrained with larger $\\mathcal{D}$ indicates that the model is presented with a larger number of unique images during training with fixed $\\mathcal{C}$ , and hence is exposed to more pretraining data diversity. In the next section, we explore the effects of variations in $\\mathcal{D}$ on SSL performance under a distribution shift, while keeping $\\mathcal{C}$ constant.", + "bbox": [ + 212, + 405, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 Fixed Budget SSL: In & Out-of-Distribution", + "text_level": 1, + "bbox": [ + 214, + 579, + 683, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Configuration We evaluate seven SSL methods: SimCLR [10], MoCoV3 [14], VICReg [4], DINO [9], BYOL [29], Barlow Twins [65], and SwAV [8]. We use different datasets for both pretraining $\\mathbb{D}_{\\mathrm{SSL}}$ and linear probing $\\mathbb{D}_{\\mathrm{task}}$ for different sections. We use solo-learn [16] as a codebase. For each method, we use the default parameters provided when available, otherwise, we conduct a hyperparameter search to ensure proper convergence.", + "bbox": [ + 212, + 611, + 787, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 Performance on the Same Distribution", + "text_level": 1, + "bbox": [ + 214, + 724, + 581, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We aim to answer a first question: does collecting more samples from the same distribution help SSL pretraining with a fixed budget? We conduct experiments to capture how the pretraining diversity $\\mathcal{D}$ influences SSL pretraining within a normalized $\\mathcal{C}$ , focusing on the simplest setting where the upstream and downstream datasets belong to the same distribution such that $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}$ . This serves as a fundamental ground for our further experiments.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup For pretraining, we use CIFAR-100 [39] and Tiny ImageNet [40] as $\\mathbb{D}_{\\mathrm{SSL}}$ which contain $50\\times 10^{3}$ and $100\\times 10^{3}$ images, respectively. We set $\\mathcal{C} = 50\\times 10^6$ chosen such that it allows all methods to converge during pretraining. In Section 6, we study the effect of varying the budget $\\mathcal{C}$ . We pretrain on subsets of $\\mathbb{D}_{\\mathrm{SSL}}$ with different sizes $N$ (10%, 50%, and 100% of $\\mathbb{D}_{\\mathrm{SSL}}$ ), enabling us to observe the effects of pretraining diversity $\\mathcal{D}$ on training outcomes where $\\mathcal{E}$ is adjusted accordingly to match the budget $\\mathcal{C}$ . For example, using 100% of CIFAR-100 involves 1000 epochs of training, while 10% and 50% of the dataset lead to 10000 and 2000 epochs, respectively. These subsets of $\\mathbb{D}_{\\mathrm{SSL}}$ are created by shuffling the dataset and then selecting the first 10%, 50%, or 100% split. All models in this section use a ResNet-18 [33] backbone pretrained from scratch. For evaluation, we use in-distribution linear probing, i.e., $\\mathbb{D}_{\\mathrm{task}} = \\mathbb{D}_{\\mathrm{SSL}}$ , where performance is measured by classification accuracy on the test set.", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results The results of linear probing are presented in Tables 1a and 1b for CIFAR-100 and Tiny ImageNet, respectively. While different SSL methods show varying levels of performance, we observe that, for all methods, an increase in the diversity $\\mathcal{D}$ consistently leads to higher linear probing accuracy, suggesting that including more in-distribution samples in $\\mathbb{D}_{\\mathrm{SSL}}$ helps under a normalized $\\mathcal{C}$ . For example, SimCLR achieves an accuracy of 37.95 when $10\\%$ of CIFAR-100 is provided, whereas this performance improve by around $12\\%$ when only $50\\%$ of the dataset is provided and by another $8\\%$ after pretraining on $100\\%$ of unique samples. This suggests that SSL methods substantially benefit from having a more diverse pretraining set in computationally budgeted in-distribution evaluation, a fundamental verification that allows us to proceed in our analysis.", + "bbox": [ + 212, + 358, + 787, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Insight (1)", + "text_level": 1, + "bbox": [ + 240, + 536, + 333, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When the distributions of the upstream and downstream tasks are the same, i.e., $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}$ , in a normalized computation setup, increasing pretraining diversity $\\mathcal{D}$ proves to be an effective strategy for enhancing SSL pretraining performance.", + "bbox": [ + 238, + 564, + 761, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 Increasing Data Diversity", + "text_level": 1, + "bbox": [ + 215, + 661, + 473, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As observed in Section 5.1, if $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}$ having a more diverse pretraining set benefits SSL methods, under a normalized computation assumption. However, to increase diversity, sampling from the same distribution $\\mathbb{D}_{\\mathrm{task}}$ to extend $\\mathbb{D}_{\\mathrm{SSL}}$ is not always attainable. In fact, pretraining data is often sampled from a distribution different than $\\mathbb{D}_{\\mathrm{task}}$ . The added samples will then introduce a distribution shift between $\\mathbb{D}_{\\mathrm{SSL}}$ and $\\mathbb{D}_{\\mathrm{task}}$ . Then, we raise the following question: is increasing pretraining diversity still beneficial when there is a distribution shift between $\\mathbb{D}_{\\mathrm{SSL}}$ and $\\mathbb{D}_{\\mathrm{task}}$ , i.e., $\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}$ ? We explore strategies for acquiring new data for increasing $\\mathcal{D}$ , namely, including existing samples, crawled internet data, and data synthetically generated. To evaluate the effects of the distribution", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37aedb84907c1bd3433b5fd0bd22e76fa4499276e05d60c6494e4a009d20084e.jpg", + "image_caption": [ + "Fig. 2: Data Collection Strategies: We analyze strategies for collecting additional data (A), i.e., collecting more source data, crawling the web or using synthetic images. Using a class prior (top row) simulates In-distribution trainings. We also collect images without class prior (bottom row) to analyze the interactions between diversity and Out-of-distribution classes." + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 702, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shift in a controlled scenario, we analyze distributions closer to $\\mathbb{D}_{\\mathrm{task}}$ by using a class prior (in-distribution) and without a class prior (out-of-distribution).", + "bbox": [ + 212, + 426, + 787, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Setup We use ImageNet-100 [57] as $\\mathbb{D}_{\\mathrm{task}}$ , and we construct multiple $\\mathbb{D}_{\\mathrm{SSL}}$ to evaluate the effects of different data collection strategies and the distribution shift. We first introduce a set $\\mathbb{B}$ composed of $65\\times 10^{3}$ images from ImageNet-100 (50% of the dataset), which we use as a baseline for $\\mathbb{D}_{\\mathrm{SSL}}$ with minimum diversity. We denote the 100 classes in ImageNet-100 as $\\mathcal{T}_{100}$ , and sample $\\mathbb{B}$ uniformly including images from all classes. Next, we compare with pretraining on more diverse datasets as $\\mathbb{D}_{\\mathrm{SSL}}$ , imposing $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{B} \\cup \\mathbb{A}$ where $\\mathbb{A}$ includes $65\\times 10^{3}$ images sampled with one of three strategies. To highlight the effects of the distribution shift, we include in $\\mathbb{A}$ either images from $In$ -distribution classes, i.e., selecting images from classes overlapping with $\\mathcal{T}_{100}$ , or images from $Out$ -of-distribution classes, which do not overlap with $\\mathcal{T}_{100}$ . This is to study the effects of the distribution shift, since we do not assume access to downstream classes in real scenarios. For the $Out$ -of-distribution samples, we define $\\mathbb{A}$ as (1) random images sampled from the full ImageNet [18] dataset; (2) images crawled from Flickr, Bing, and DuckDuckGo; or (3) images generated with Stable Diffusion V2.1 [51]. We respectively call these sets $\\mathbb{A}_{\\mathrm{Source}}^{Out}$ , $\\mathbb{A}_{\\mathrm{Web}}^{Out}$ , and $\\mathbb{A}_{\\mathrm{Synthetic}}^{Out}$ . We also define their $In$ -distribution counterparts as $\\mathbb{A}_{\\mathrm{Source}}^{In}$ , $\\mathbb{A}_{\\mathrm{Web}}^{In}$ , and $\\mathbb{A}_{\\mathrm{Synthetic}}^{In}$ , respectively. Note how $\\mathbb{B} \\cup \\mathbb{A}_{\\mathrm{Source}}^{In}$ is the full ImageNet-100, coherently with Section 5.1. Figure 2 shows each collection strategy, for which we provide implementation details in supplementary material. Although many factors (such as the images appearance) impact the distribution shift, using a class prior with any strategy would result in closer distribution with respect to the same strategy without class priors [15]. We pretrain a ResNet-18 from scratch, with the same settings of Section 5.1, and $\\mathcal{C} = 50\\times 10^6$ . Note that this results in $\\mathcal{D} = 1.25\\times 10^{-3}$ for pretraining on $\\mathbb{B}$ , and $\\mathcal{D} = 2.5\\times 10^{-3}$ for pretraining on $\\mathbb{B} \\cup \\mathbb{A}$ , introducing a $\\mathcal{D}$ difference of a factor of 2.", + "bbox": [ + 212, + 458, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/499aef39458c4cf65338b219cc5dccb382b983eb764f0832660f835fe8136d0d.jpg", + "image_caption": [ + "Fig. 3: Effect of Various Data Sources on SSL Pretraining: We use a baseline set $\\mathbb{B}$ (black dashed line), comprising $65 \\times 10^{3}$ images from ImageNet-100, for pretraining a ResNet-18 with $\\mathcal{C} = 50 \\times 10^{6}$ . Augmenting $\\mathbb{B}$ with In-distribution images enhances performance (above black line), while Out-of-distribution augmentations reduce it (below black line)." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results We report the linear probing accuracies for pretraining on each $\\mathbb{B} \\cup \\mathbb{A}$ as bars in Figure 3, showing the $\\mathcal{C}$ -normalized training on $\\mathbb{B}$ as a dashed line. Surprisingly, without class priors $(\\mathbf{Out})$ , including $\\mathbb{A}_{\\mathrm{Source}}^{Out}$ , $\\mathbb{A}_{\\mathrm{Web}}^{Out}$ , and $\\mathbb{A}_{\\mathrm{Synthetic}}^{Out}$ underperforms compared to pretraining on $\\mathbb{B}$ only. For instance, while SimCLR on $\\mathbb{B}$ scores $72.3\\%$ accuracy, increasing diversity reduces the accuracy by $1\\%$ in all cases. This might appear to conflict with our findings in Section 5.1, however, the inclusion of $\\mathbf{Out}$ samples leads to $\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}$ , since we sample only classes not included in $\\mathbb{D}_{\\mathrm{task}}$ . We infer that, with normalized $\\mathcal{C}$ , increasing $\\mathcal{D}$ without distribution priors may negatively impact performance. Conversely, when class priors are available $(\\mathbf{In})$ , increasing pretraining diversity $\\mathcal{D}$ improves performance compared to $\\mathbb{B}$ pretraining. For instance, pretraining on $\\mathbb{A}_{\\mathrm{Web}}^{In}$ performs comparably to augmenting with additional ImageNet samples $(\\mathbb{A}_{\\mathrm{Source}}^{In})$ , as in the case of SimCLR where the inclusion of $\\mathbb{A}_{\\mathrm{Web}}^{In}$ scores only $0.2\\%$ lower than $\\mathbb{A}_{\\mathrm{Source}}^{In}$ . Including $\\mathbb{A}_{\\mathrm{Synthetic}}^{In}$ data also helps, although more limitedly due to the visual discrepancies between generated and real images. Ultimately, the effectiveness of diversity is linked to the distribution shift. These findings highlight the impact of the distribution shift on computationally-normalized SSL pretraining and help define evaluation practices for this task (see Section 7).", + "bbox": [ + 212, + 359, + 787, + 632 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Insight (2)", + "text_level": 1, + "bbox": [ + 240, + 642, + 333, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When the distributions of the upstream and downstream tasks differ, $\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}$ , and in a normalized computation setup, increasing pretraining diversity $\\mathcal{D}$ may harm SSL pretraining performance, reflecting the adverse effects of distribution shift.", + "bbox": [ + 238, + 670, + 761, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3 Scaling Pretraining Diversity", + "text_level": 1, + "bbox": [ + 215, + 768, + 501, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We showed that diversity improves pretraining performance when the training set and downstream task share the same data distribution $(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}})$ , as discussed in Section 5.1. This may change when a distribution shift is introduced,", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/edd5a49ed0b7a3a5341403c828aabb977afce7e29ddd7536d0ec6cc8eaa674a9.jpg", + "table_caption": [ + "Table 2: Performance of ImageNet and YFCC100M SSL Pretraining on Various Downstream Tasks: We train ResNet-50 and ViT-B/16 under normalized computation $(\\mathcal{C} = 98\\times 10^{6})$ using SimCLR (left) and MoCoV3 (right) on ImageNet (IN) and YFCC100M (YFCC) with multiple $\\mathcal{D}$ . Despite the significantly larger $\\mathcal{D}$ when trained on YFCC100M, these models cannot offset the effects of the distribution shift and are outperformed by models pretrained on ImageNet in most downstreams." + ], + "table_footnote": [ + "$\\mathcal{C} = 98\\times 10^{6}$" + ], + "table_body": "
SimCLR
\\( \\mathbb{D}_{\\text{SSL}} \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3156.943.082.373.945.459.9
IN0.2562.6161.145.584.076.047.064.4
IN0.6406.5463.746.884.278.848.367.2
IN1.28113.064.546.484.679.548.868.0
YFCC98.17100057.337.276.658.950.162.1
ViT-B/16IN0.1281.3154.237.281.870.344.664.3
IN0.2562.6161.339.583.877.747.269.4
IN0.6406.5465.539.984.680.449.173.1
IN1.28113.066.739.583.781.750.073.6
YFCC98.17100054.525.072.559.749.565.4
", + "bbox": [ + 215, + 241, + 496, + 393 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3ef61d7f3f67f77f54c546e172a6b5bc5134fa65c1663ff8bd541c16645ae8dd.jpg", + "table_caption": [], + "table_footnote": [ + "$\\mathcal{C} = 98\\times 10^{6}$" + ], + "table_body": "
MoCoV3
\\( \\mathbb{D}_{\\text{SSL}} \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3158.140.681.876.0445.1
IN0.2562.6162.945.385.079.847.6
IN0.6406.5465.448.486.181.949.2
IN1.28113.065.948.886.682.649.5
YFCC98.17100060.442.282.666.350.767.3
ViT-B/16IN0.1281.3157.933.482.878.046.7
IN0.2562.6163.735.085.382.948.4
IN0.6406.5467.239.585.085.849.7
IN1.28113.068.841.986.586.550.3
YFCC98.17100057.225.270.343.850.364.0
", + "bbox": [ + 501, + 241, + 784, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "as explored in Section 5.2. However, it is still unclear from Section 5.2, whether including a larger number of samples, and thus increasing considerably the pretraining diversity, can compensate for the negative effects of the distribution shift. To address this, the following section presents larger-scale experiments, employing significantly varied $\\mathcal{D}$ values, aiming to explore the interplay between pretraining diversity and different distributions using millions of samples.", + "bbox": [ + 212, + 441, + 787, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Setup For our large-scale pretraining experiments, we set $\\mathbb{D}_{\\mathrm{SSL}}$ to be two datasets of significantly different sizes: ImageNet [18] and YFCC100M [55], comprising 1.28 million and 98 million images, respectively. Following Section 5.1, we explore multiple $\\mathcal{D}$ values for pretraining. We set $\\mathcal{C} = 98 \\times 10^6$ , which corresponds to one epoch on YFCC100M, iterating once through each of its 98 million images to maximize diversity $(\\mathcal{D} = 1)$ . Normalizing $\\mathcal{C}$ (see Section 5.1), we pretrain on ImageNet for approximately 77 epochs, cumulatively utilizing 98 million samples. Due to the extensive cost of these experiments, we focus on SimCLR and MoCoV3 only. We also employ larger architectures, namely ResNet-50 [33] and ViT-B/16 [21], to leverage the extensive scale of the pretraining datasets. We also use multiple $\\mathbb{D}_{\\mathrm{task}}$ including ImageNet [18], Stanford Cars [38], Flowers-102 [44], Oxford-IIIT Pets [48], Places365 [69], and Food-101 [5].", + "bbox": [ + 212, + 550, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results The results of our large-scale experiments are detailed in Table 2. Consistently with findings in Section 5.1, increasing $\\mathcal{D}$ leads to better pretraining efficacy when $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}$ . This is evident when ImageNet is used for both pretraining and downstream classification, reinforcing that our observations hold even at a larger scale. Instead, models pretrained on YFCC100M show substantially lower performance compared to those pretrained on ImageNet, although", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "having much higher $\\mathcal{D}$ . This highlights the inefficiency of collecting data indiscriminately without considering distribution shifts. To stress this, note how the model pretrained on YFCC100M ( $D = 1$ ) often performs similarly to those pretrained with drastically lower $\\mathcal{D}$ on ImageNet ( $D = 1.31 \\times 10^{-3}$ ). This aligns with our observations in Section 5.2, emphasizing that distribution differences remain a significant factor even when training with large datasets. However, the YFCC100M-pretrained model outperforms the ImageNet-pretrained model in Places365, suggesting a closer distribution relationship between YFCC100M and Places365. We explore this hypothesis further in Section 6, where we analyze distribution distances with existing metrics. Ultimately, our analysis highlights that scaling the data is not an effective solution for compensating the distribution shift, when computational resources are normalized.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Insight (3)", + "text_level": 1, + "bbox": [ + 241, + 339, + 331, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Even an extremely large data diversity cannot mitigate the distribution shift under normalized computation. This emphasizes the importance of further research in how effectively SSL pretraining algorithms generalize.", + "bbox": [ + 238, + 367, + 761, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Additional Analysis", + "text_level": 1, + "bbox": [ + 215, + 435, + 442, + 452 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Previously, we introduced a normalized evaluation to study the relation of $\\mathcal{D}$ and data distributions. We showed that, although diversity helps when $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}$ , vision models are unable to compensate for the distributional differences with normalized costs. We now analyze additional elements supporting our findings.", + "bbox": [ + 212, + 460, + 787, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Distribution Distances Using FID & VisualDNA In Section 5.3, we showed that pretraining on ImageNet typically outperforms YFCC100M on a variety of downstream tasks, with Places365 being the exception. We speculate that the distribution of ImageNet is more aligned with those of the downstream tasks compared to YFCC100M. To verify this, we evaluate the similarity between the datasets using FID [34] and VisualDNA [50] and report results in Table 3. With both FID or VisualDNA, the distribution of ImageNet is always closer to the downstream tasks, except for Places365 where YFCC100M is closer. This aligns with the lower performance of ImageNet on this dataset only (Table 2), further suggesting that the performance drop is caused by the distribution shift.", + "bbox": [ + 212, + 537, + 787, + 690 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Importance of Normalized Computation We now study the impact of normalizing computation on the performance of SSL methods. We aim to understand if $\\mathcal{C}$ is not normalized across methods, this will lead to misleading conclusions, and thus motivating our normalized computation. In Table 4, we compare the performance of ResNet-50 and ViT-B/16 models pretrained using MoCoV3 with $\\mathcal{C} = 98 \\times 10^6$ (as used in Section 5.3) against a tripled budget of $\\mathcal{C} = 294 \\times 10^6$ . We show that inconsistencies arise when pretraining data diversity $\\mathcal{D}$ and computation $\\mathcal{C}$ are not fairly compared. For instance, in scenarios highlighted in red, we notice that a lower pretraining diversity coupled with a higher", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Comparing VisualDNA and FID Scores Across Datasets: We assess the relationship between VisualDNA [50] and FID Score [34] for various large-scale $\\mathbb{D}_{\\mathrm{SSL}}$ and several downstream tasks. For VisualDNA, activations are extracted as suggested [50] with MUGS [70] with ViT-B/16 (MUGS) or DINO v2 [47] with ViT-B/16 (DINO v2), while FID activations are obtained via an Inception network [54]. Consistently, across all metrics, the ranking of dataset distances between $\\mathbb{D}_{\\mathrm{SSL}}$ and various $\\mathbb{D}_{\\mathrm{task}}$ aligns with the accuracy ranking in Table 2. Models exhibiting lower VisualDNA/FID scores benefit more from the diversity in pretraining data for SSL.", + "bbox": [ + 215, + 143, + 491, + 380 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b98970d2b868e58612791ef9532b45d3bb331f34d40e0adb672974134e1ee861.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DSSLBackboneVisualDNA↓
INCarsFlow.PetsPlacesFood
INMUGS011.4912.466.097.199.08
YFCCMUGS3.7211.5712.717.936.209.72
INDINO v207.056.954.433.527.26
YFCCDINO v22.377.127.465.752.747.90
DSSLBackboneFID↓
INCarsFlow.PetsPlacesFood
INInception0143.4192.988.964.9114.9
YFCCInception48.1174.3214.8145.838.9154.5
", + "bbox": [ + 218, + 392, + 491, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Importance of Normalization. We report the accuracy of MoCoV3 trained on ImageNet (IN) with different data diversity and variable $\\mathcal{C}$ , for the in-distribution assumption $(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}})$ . For a given network, cells in red highlight inconsistencies where although the model trained with $\\mathcal{C} = 294 \\times 10^6$ has seen less samples, it outperforms the best model trained with one third of the computational budget $(\\mathcal{C} = 90 \\times 10^6)$ , showing the importance of normalization for understanding how models exploit data.", + "bbox": [ + 501, + 147, + 782, + 329 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b2522af5faa0c747ba7fe09f340dc11c9cdcb516a517d5d8c19d1ab2a8b5e221.jpg", + "table_caption": [ + "MoCoV3" + ], + "table_footnote": [], + "table_body": "
Network\\( \\mathbb{D}_{\\text{SSL}} \\)\\( \\mathcal{C} \\times 10^6 \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy† IN
ResNet-50IN980.1281.3158.1
IN980.6406.5465.4
IN981.28113.065.9
IN2940.1280.4357.4
IN2940.6402.1769.8
IN2941.2814.3571.4
ViT-B/16IN980.1281.3157.9
IN980.6406.5467.2
IN981.28113.068.8
IN2940.1280.4356.9
IN2940.6402.1771.9
IN2941.2814.3574.9
", + "bbox": [ + 509, + 352, + 777, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "computational budget can yield better results. We observe that under pretraining diversity of only $\\mathcal{D} = 13\\times 10^{-3}$ and a computational budget of $\\mathcal{C} = 98\\times 10^6$ ResNet-50 only enjoys a $65.9\\%$ accuracy compared against with $69.8\\%$ with a smaller pretraining diversity of $2.17\\times 10^{-3}$ but with a larger computational budget of $294\\times 10^{6}$ . This shows that without normalized computation, it could be incorrectly concluded that pretraining diversity does not play a significant role.", + "bbox": [ + 212, + 547, + 787, + 640 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Model Saturation We aim to study the pretraining trend on YFCC100M, and if adding more samples could compensate the distribution shift with ImageNet. Hence, we extend the YFCC100M experiments from Section 5.3, examining various subsets—specifically $0.1\\%$ , $1\\%$ , $10\\%$ , and $100\\%$ of the dataset. Again, we normalize the computational budget to $\\mathcal{C} = 98 \\times 10^6$ , equivalent to one epoch on YFCC100M. The results of linear probing models pretrained using SimCLR and MoCoV3 with ResNet-50 and ViT-B/16 on ImageNet are shown in Figure 4. Interestingly, we obtain a performance plateau. This observation points to a saturation point in performance, showing that simply increasing the dataset further would not bridge the gap between the $\\mathbb{D}_{\\mathrm{SSL}}$ and $\\mathbb{D}_{\\mathrm{task}}$ . Hence, arbitrarily increasing pretraining diversity $\\mathcal{D}$ is not sufficient to bridge the distribution shift problem. Here $\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}$ , hence still aligning with our findings in Section 5.1.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a0b7c13e724ad3bd7d335cdc10821de1486178e39607abbcd2524a10e6098ece.jpg", + "image_caption": [ + "Fig. 4: Data Diversity Impact on YFCC100M Pretraining Performance: Pretraining $(\\mathcal{C} = 98\\times 10^{6})$ of networks with $\\mathbb{D}_{\\mathrm{SSL}} = \\mathrm{YFCC100M}$ and $\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}$ for several dataset subsets. In the presence of a distribution shift, performance tends to saturate and does not benefit from additional data diversity." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 243 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/97e66e7c08440b54ffa4f3f1dc840b86b6b1a7e28771e1d9c7242bbcaa6a9526.jpg", + "table_caption": [ + "Table 5: Evaluating Network Accuracy With Varied Label Quantity: We evaluate the accuracy of networks trained on ImageNet (IN) and YFCC100M (YFCC) with different labeling percentages of $\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}$ . The increased diversity still does not compensate for the distribution shift. However, for in-distribution data, one can get away with fewer labels with more diverse pretraining data." + ], + "table_footnote": [], + "table_body": "
SimCLR
Network\\( \\mathbb{D}_{\\text{SSL}} \\)N×106D×10-3Accuracy†
5%10%50%
ResNet-50IN1.281.3150.354.261.7
YFCC98.2100039.444.354.0
ViT-B/16IN1.281.3155.259.465.6
YFCC98.2100040.745.753.3
C=98×10^6
", + "bbox": [ + 220, + 429, + 496, + 508 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a186ff28a4385e0e76328b728e3b193a41f53469e4ff6a1fd264f4693f91f6fd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NetworkMoCoV3
\\( \\mathbb{D}_{\\text{SSL}} \\)N×106D×10-3Accuracy↑
5%10%50%
ResNet-50IN1.281.3152.156.263.8
YFCC98.2100042.848.257.7
ViT-B/16IN1.281.3159.363.068.3
YFCC98.2100043.949.056.0
C=98×10^6
", + "bbox": [ + 504, + 429, + 785, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Label Quantity A question that arises in our setting is whether having a higher pretraining diversity leads to requiring less labeled samples during linear probing. In this section, we focus on understanding the impact of increased $\\mathcal{D}$ on the number of labeled samples required for effective linear probing in downstream tasks. We use the same trained models from Section 5.3, i.e., SimCLR and MoCoV3, with ResNet-50 and ViT-B/16 architectures. In this setup, we set $\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}$ , with two upstream dataset scenarios: ImageNet (where $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}$ ) and YFCC100M (where $\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}$ ). Our experiments are summarized in Table 5. We note that with ViT-B/16, if ImageNet is used for pretraining, linear probing with just $5\\%$ labeled samples can surpass the performance achieved using $100\\%$ labeled data when YFCC100M serves as $\\mathbb{D}_{\\mathrm{SSL}}$ . This also implies that when $\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}$ , a higher quantity of labeled data is necessary to attain competitive performance, and increasing $\\mathcal{D}$ does not bring considerable benefits on label quantity under distribution shift. This implies that our findings in Section 5.3 hold regardless of the linear probing labeled set. We note that in scenarios where $\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}$ , using only $50\\%$ of the labeled data can achieve similar performance as utilizing the full $100\\%$ of labeled samples, implying that increasing $\\mathcal{D}$ leads to reduced label requirement efforts for downstream tasks.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 Discussion", + "text_level": 1, + "bbox": [ + 215, + 143, + 354, + 160 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Main Conclusions Our set of experiments leads to Insight (3) in Section 5.3, revealing that with normalized computational costs, SSL pretrainings with large diversity cannot compensate for the distribution shift. This is surprising, since the variety of information that SSL algorithms could benefit from during training is much higher in large generic datasets than in small ones. Since our evaluation is cost-normalized, (3) also shows that SSL strategies are not efficiently exploiting the pretraining diversity on large datasets for representation extraction. This inefficiency implies a wide margin for improvement of generalization of SSL models, making better use of the computational power used for training. The role of existing models must also be discussed in this context. Following Insights (1) and (2), respectively in Sections 5.1 and 5.2, we have studied how in-distribution and out-of-distribution data impact performance in a controlled scenario. We argue that this should be taken into account in the evaluation of the performance of SSL models. Indeed, training at scale may enlarge the in-distribution space, including classes of the downstream tasks in the training set. In recent literature, this is a design choice to maximize performance on popular benchmarks [47]. While this allows for achieving impressive results in many tasks, we stress that this does not permit a fair evaluation. Now, we summarize practical takeaways.", + "bbox": [ + 212, + 185, + 787, + 458 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Training Takeaways Coherently with our findings in Section 5.1 and similarly to prior art [15,28,37], we find that aligned distributions benefit performance, in particular increasing $\\mathbb{D}_{\\mathrm{SSL}}$ diversity helps, as long as the distribution of the new samples match those of the downstream task data. Differently from the state-of-the-art, we demonstrate that this holds also in a computationally-normalized setting, implying that collecting large scale in-distribution data matching the downstream task is be an effective approach for improving SSL. Hence, for practical applications distribution priors should be used, if available. On the contrary, for a fair evaluation of models, this should not be the case, as specified below.", + "bbox": [ + 212, + 458, + 787, + 594 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Evaluation Takeaways Our analysis reveals that to permit a fair evaluation of SSL methods, computationally normalized tests are necessary to avoid inconsistencies, as shown in Section 6. Moreover, it is crucial to identify out-of-distribution downstream tasks for a correct evaluation of generalization performance. By evaluating only on $\\mathbb{D}_{\\mathrm{task}}$ with a low distribution shift, there is a risk of reporting inflated metrics, not representative of a real gain in generalization. This is important, since new SSL approaches may be reporting higher downstream performance when pretrained on a different dataset. We relate this to Sections 5 and 6, where we show that increasing the computation and the in-distribution data, respectively, can improve performance. Ultimately, wrong practices may result in incorrectly concluding that an SSL algorithm is performing better.", + "bbox": [ + 212, + 595, + 787, + 762 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Differences With Language Models In Section 5.3 we showed that even very diverse datasets (YFCC100M), fall short in satisfactory generalization performance. In addition to further exploration into generalization for SSL pretraining, this open doors to investigating why language models enjoy enhanced generalization when trained on highly diverse data compared to vision models [61].", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by SDAIA-KAUST Center of Excellence in Data Science, Artificial Intelligence (SDAIA-KAUST AI). Fabio Pizzati is financed by KAUST (Grant DFR07910). Philip H.S. Torr thanks the Royal Academy of Engineering for their support. This work is supported by a UKRI grant Turing AI Fellowship (EP/W002981/1).", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bachman, P., Hjelm, R.D., Buchwalter, W.: Learning representations by maximizing mutual information across views. In: NeurIPS (2019)", + "2. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023)", + "3. Bao, H., Dong, L., Piao, S., Wei, F.: BEit: BERT pre-training of image transformers. In: ICML (2022)", + "4. Bardes, A., Ponce, J., LeCun, Y.: VICReg: Variance-invariance-covariance regularization for self-supervised learning. In: ICLR (2022)", + "5. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-Mining discriminative components with random forests. In: ECCV (2014)", + "6. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: ECCV (2018)", + "7. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: ICCV (2019)", + "8. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020)", + "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)", + "0. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)", + "1. Chen, T., Kornblith, S., Swersky, K., Norouzi, M., Hinton, G.: Big self-supervised models are strong semi-supervised learners. In: NeurIPS (2020)", + "2. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "3. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021)", + "4. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV (2021)", + "5. Cole, E., Yang, X., Wilber, K., Aodha, O.M., Belongie, S.: When does contrastive visual representation learning work? In: CVPR (2022)", + "6. Costa, V.G.T.d., Fini, E., Nabi, M., Sebe, N., Ricci, E.: solo-learn: A library of self-supervised methods for visual representation learning. In: JMLR (2022)", + "7. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: ICML (2023)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + ], + "bbox": [ + 225, + 310, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: ICCV (2015)", + "20. Doersch, C., Zisserman, A.: Multi-task self-supervised visual learning. In: ICCV (2017)", + "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "22. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv (2021)", + "23. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: CVPR (2021)", + "24. Fini, E., Astolfi, P., Romero-Soriano, A., Verbeek, J., Drozdzal, M.: Improved baselines for vision-language pre-training. TMLR (2023)", + "25. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018)", + "26. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021)", + "27. Goyal, P., Duval, Q., Seessel, I., Caron, M., Misra, I., Sagun, L., Joulin, A., Bojanowski, P.: Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360 (2022)", + "28. Goyal, P., Mahajan, D., Gupta, A., Misra, I.: Scaling and benchmarking self-supervised visual representation learning. In: ICCV (2019)", + "29. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent: A new approach to self-supervised learning. In: NeurIPS. vol. 33 (2020)", + "30. Hammoud, H.A.A.K., Itani, H., Pizzati, F., Torr, P., Bibi, A., Ghanem, B.: Synthclip: Are we ready for a fully synthetic clip training? arXiv (2024)", + "31. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)", + "32. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)", + "33. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017)", + "35. Hjelm, R.D., Fedorov, A., Lavoie-Marchildon, S., Grewal, K., Bachman, P., Trischler, A., Bengio, Y.: Learning deep representations by mutual information estimation and maximization. In: ICLR (2018)", + "36. Henaff, O.J., Srinivas, A., De Fauw, J., Razavi, A., Doersch, C., Eslami, S.M.A., Oord, A.v.d.: Data-efficient image recognition with contrastive predictive coding. In: ICML (2019)", + "37. Kotar, K., Ilharco, G., Schmidt, L., Ehsani, K., Mottaghi, R.: Contrasting contrastive self-supervised representation learning pipelines. In: ICCV (2021)", + "38. Krause, J., Deng, J., Stark, M., Fei-Fei, L.: Collecting a large-scale dataset of fine-grained cars. In: FGVC (2013)", + "39. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N (2015)", + "41. Li, A.C., Brown, E.L., Efros, A.A., Pathak, D.: Internet explorer: Targeted representation learning on the open web. In: ICML (2023)", + "42. Li, J., Zhou, P., Xiong, C., Hoi, S.C.H.: Prototypical contrastive learning of unsupervised representations. In: ICLR (2021)", + "43. Misra, I., Maaten, L.v.d.: Self-supervised learning of pretext-invariant representations. In: CVPR (2020)", + "44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP (2008)", + "45. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016)", + "46. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "47. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINoV2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR (2012)", + "49. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: CVPR (2020)", + "50. Ramtoula, B., Gadd, M., Newman, P., De Martini, D.: Visual DNA: Representing and comparing images using distributions of neuron activations. In: CVPR (2023)", + "51. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "52. Shi, Y., Daunhawer, I., Vogt, J.E., Torr, P., Sanyal, A.: How robust is unsupervised representation learning to distribution shift? In: ICLR (2022)", + "53. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: ICCV (2017)", + "54. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: CVPR (2016)", + "55. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: YFCC100M: The new data in multimedia research. In: Commun. ACM (2016)", + "56. Tian, Y., Henaff, O.J., Van Den Oord, A.: Divide and contrast: Self-supervised learning from uncurated data. In: ICCV (2021)", + "57. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: ECCV (2020)", + "58. Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: CVPR (2011)", + "59. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Gool, L.V.: Revisiting contrastive methods for unsupervised learning of visual representations. In: NeurIPS (2021)", + "60. Van Horn, G., Cole, E., Beery, S., Wilber, K., Belongie, S., Mac Aodha, O.: Benchmarking representation learning for natural world image collections. In: CVPR (2021)", + "61. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. TMLR (2022)", + "62. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via non-parametric instance discrimination. In: CVPR (2018)", + "63. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMIM: A simple framework for masked image modeling. In: CVPR (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On Pretraining Data Diversity for SSL", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Wei, Y., Dai, Q., Hu, H.: On data scaling in masked image modeling. In: CVPR (2023)", + "65. Zbontar, J., Jing, L., Misra, I., LeCun, Y., Deny, S.: Barlow Twins: Self-supervised learning via redundancy reduction. In: ICML (2021)", + "66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR (2022)", + "67. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016)", + "68. Zhao, N., Wu, Z., Lau, R.W.H., Lin, S.: What makes instance discrimination good for transfer learning? In: ICML (2021)", + "69. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. T-PAMI (2017)", + "70. Zhou, P., Zhou, Y., Si, C., Yu, W., Ng, T.K., Yan, S.: Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415 (2022)", + "71. Zhuang, C., Zhai, A.L., Yamins, D.: Local aggregation for unsupervised learning of visual embeddings. In: ICCV (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 354 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Hammoud et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_model.json b/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e715a333932eab329248cbab119ee8617f2ad3fa --- /dev/null +++ b/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_model.json @@ -0,0 +1,2348 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.328, + 0.141, + 0.675, + 0.186 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for Self-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.212, + 0.744, + 0.244 + ], + "angle": 0, + "content": "Hasan Abed Al Kader Hammoud\\(^{1,2\\star\\diamond}\\), Tuhin Das\\(^{2\\star}\\), Fabio Pizzati\\(^{2\\star}\\) Philip H.S. Torr\\(^{2}\\), and Adel Bibi\\(^{2\\dagger}\\), Bernard Ghanem\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.464, + 0.254, + 0.536, + 0.268 + ], + "angle": 0, + "content": "1 KAUST" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.269, + 0.578, + 0.283 + ], + "angle": 0, + "content": "2 University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.318, + 0.74, + 0.484 + ], + "angle": 0, + "content": "Abstract. We explore the impact of training with more diverse datasets, characterized by the number of unique samples, on the performance of self-supervised learning (SSL) under a fixed computational budget. Our findings demonstrate that increasing pretraining data diversity enhances SSL performance, albeit only when the distribution distance to the downstream data is minimal. Notably, even with an exceptionally large pretraining data diversity achieved through methods like web crawling or diffusion-generated data, among other ways, the distribution shift remains a challenge. Our experiments are comprehensive with seven SSL methods using large-scale datasets such as ImageNet and YFCC100M amounting to over 200 GPU days. The code and trained models will be available at https://github.com/hammoudhasan/DiversitySSL." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.498, + 0.737, + 0.513 + ], + "angle": 0, + "content": "Keywords: self-supervised learning \\(\\cdot\\) distribution shift \\(\\cdot\\) data diversity" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.559, + 0.375, + 0.575 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) has recently emerged as a new paradigm to pretrain large vision models at scale [26, 27, 47]. Leveraging the ability to learn from unlabelled data, pretraining on millions—or even billions [27, 53]—of images turned from an unachievable goal to a common practice. This exposure to extremely diverse datasets, i.e., composed of a remarkable number of unique samples, granted impressive performance and unprecedented generalization capabilities to a growing number of vision models [47]. Large-scale datasets, in conjunction with substantial computational resources, have been the key driving forces for the success of SSL-based approaches. For instance, SEER [26], pretrained for approximately 11 GPU years on a billion images, exemplifies the massive computation and data resources employed for these models. It has become the implicit norm that increasing computation and data is beneficial, without any detailed analysis of how they separately impact SSL effectiveness. In particular, it is not clear to what extent large datasets are responsible for the impressive generalization capabilities of SSL models. Indeed, consider the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.787, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution † Equal supervision Ⓞ Work done during a research visit at Oxford" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.146, + 0.734, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 1: Impact of Pretraining Diversity: Self-supervised learning (SSL) can be used to pretrain vision models on small datasets closely aligned to the downstream task, e.g., pets classification, hence with a small distribution shift (top, wild animals pretraining). Conversely, we could pretrain on an extensively varied dataset, with wide distribution differences (outdoor scenes, bottom). We study the role of pretraining diversity in SSL under a fixed budget, and highlight its effects in relationship to the distribution shift." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.644 + ], + "angle": 0, + "content": "example in Figure 1. Assuming a fixed monetary budget, in the form of computational expenses, for pretraining a vision model under the SSL paradigm, does iterating over a large set of images work best given a downstream task, or it is better to train repeatedly on a smaller set of samples visually close to the ones of the downstream? In this context, we face the ever-lasting problem of distribution shift on vision models [58]. Without a proper understanding of the impact of data, there could be a wrong allocation of efforts in increasing data or computation, leading to inefficiencies in the deployment process of SSL models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.735 + ], + "angle": 0, + "content": "In this paper, we study the role of the diversity of the SSL pretraining data, under a fixed computational budget, when the pretraining data matches the distribution of the downstream data, as well as when they differ. Our experiments span various SSL methods [4,8-10, 14, 29, 65] and multiple datasets [5,18,38-40, 44,48,55,57,69], and are tested under several computational budgets amounting to a total of 200 GPU days. We summarize our contributions below:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.735, + 0.788, + 0.809 + ], + "angle": 0, + "content": "1. We show that SSL pretraining strategies are currently data-inefficient in compensating for distribution shifts. Under normalized computational costs, we verify that pretraining on large datasets with high diversity cannot outperform the pretraining on datasets with limited diversity, but with a lower distribution shift with respect to the downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "2. We conclude that there is a wide margin for improvement in the performance of current SSL methods on out-of-distribution classes, and we propose in" + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "sights for a fair evaluation. This shows the need for future research on new SSL techniques leveraging data diversity better, to improve generalization capabilities beyond the training data distribution." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.192, + 0.784, + 0.237 + ], + "angle": 0, + "content": "3. We propose a novel strategy for a computationally-normalized evaluation of SSL and a carefully designed set of experiments focusing on pretraining data diversity, enabling us to draw our novel conclusions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.283 + ], + "angle": 0, + "content": "Ultimately, our work provides a comprehensive analysis, from experiments worth 200 GPU days on NVIDIA A100, on the interaction between data diversity and computational resources, and their impact on the performance of SSL models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.388, + 0.324 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Self-Supervised Learning Early work in self-supervised learning used simple pretext tasks, such as relative patch prediction [19, 20], image colorization [67], image rotation prediction [25], or solving jigsaw puzzles [45] to train feature extractors in absence of annotations. Representations learned with those methods are limitedly effective, hence more recent literature has moved towards more sophisticated approaches, such as using image augmentations to generate correlated views of a training sample, and learning to extract augmentation-invariant representations for these correlated pairs. Among these multi-view methods, many exploit contrastive losses [1,8,10-12,32,35,36,42,43,46], enforcing similarity between views of the same image (positives) and dissimilarity from other images (negatives). Due to the need of many negative samples, contrastive methods often require large batch sizes to work effectively [10,14]. Cluster-based methods such as SwAV [8], DINO [9] and DeepCluster v2 [6] learn generalizable representations by grouping samples into cluster prototypes. Others exploit prediction of features with siamese networks [13], learn features in a teacher-student fashion [29], or use redundancy reduction techniques [4,65]. Finally, masked image modeling [3,31,63] emerged as a scalable alternative for Vision Transformers [21], learning representations by predicting masked image patches as SSL task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Pretraining at Scale SSL pretraining is most effective at scale, benefiting from large datasets and great computational resources. Initial attempts at large-scale pretraining were made by combining contrastive learning and clustering [7, 56]. SEER [26] was trained on 1 billion internet images with SwAV [8] on a RegNet [49] backbone. The importance of model scaling to leverage large datasets is also shown in [27, 28, 64], as well as the need for longer training times [64]. Additional strategies are necessary to achieve best performance [17, 24, 66]. These works focus on reaching the best representations, without considerations about training costs, thus encouraging unfair comparisons across trainings. Preliminary papers [15, 22] found that increasing data may lead to a modest boost in downstream performance when the downstream distribution changes. While they give insights on the pretraining diversity, its relation with the distribution shift under normalized computation is still not investigated." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Distribution Shift in SSL Some studies have investigated how the pretraining data affects downstream performance on other domains. In general, the distribution shift increases when factors such as visual content and data acquisition procedure differ in two data ensembles [58]. For various pretraining datasets, preliminary works [15, 28, 37, 41] observed that the best performing representations were learned on datasets that were similar to the downstream test task. Additionally, combining datasets before pretraining or combining self-supervised features learned from various datasets did not lead to significant improvements in [15]. In [52, 68], they showcase higher generalization capabilities of SSL models compared to their supervised counterparts, for several downstream tasks in the presence of a distribution shift. In [59], they pretrained on several datasets, observing different generalization depending on the appearance of the pretraining dataset. Furthermore, initial considerations on the impact of external data on downstream tasks under distribution shift have been proposed in [60]. Some compensate the distribution shift with the scale of training datasets [30]. However, there is still a lack of a fair, computation-normalized evaluation that allows to study the effects of the distribution shift in a controlled environment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.383, + 0.443 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.789, + 0.704 + ], + "angle": 0, + "content": "Pretraining We first outline the general pretraining procedure common to state-of-the-art self-supervised learning methods. Specific differences among these methods are detailed in the supplementary material. The overall pretraining pipeline, common across many SSL approaches [4, 8-10, 29, 65], goes as follows: (1) Data Sampling: from a large-scale, unlabeled upstream pretraining dataset, denoted as \\(\\mathbb{D}_{\\mathrm{SSL}}\\), an image \\(\\mathbf{x}\\) is randomly sampled; (2) View Generation: two correlated views, \\(\\tilde{\\mathbf{x}}_A\\) and \\(\\tilde{\\mathbf{x}}_B\\), are generated from \\(\\mathbf{x}\\) using two image augmentation functions sampled at random. These random augmentations include random resized cropping, horizontal flipping, blurring, and color adjustments [2], among others; (3) Feature Extraction and Projection: the correlated views undergo feature extraction through a network, \\(f_{\\theta_f}\\), parameterized by \\(\\theta_f\\), such as a ResNet [33] or a ViT [21], leading to representations \\(\\mathbf{h}_A = f_{\\theta_f}(\\tilde{\\mathbf{x}}_A)\\) and \\(\\mathbf{h}_B = f_{\\theta_f}(\\tilde{\\mathbf{x}}_B)\\). A linear projection head, \\(g_{\\theta_g}\\), parameterized by \\(\\theta_g\\), then maps these representations to a latent space, resulting in \\(\\mathbf{z}_A = g_{\\theta_g}(\\mathbf{h}_A)\\) and \\(\\mathbf{z}_B = g_{\\theta_g}(\\mathbf{h}_B)\\); (4) Joint Optimization: The feature extractor \\(f_{\\theta_f}\\) and the projection head \\(g_{\\theta_g}\\) are jointly optimized according to the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.715, + 0.787, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\theta_ {f} ^ {*}, \\theta_ {g} ^ {*} = \\underset {\\theta_ {f}, \\theta_ {g}} {\\arg \\min } \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {D} _ {\\mathrm {S S L}}} \\mathcal {L} _ {\\mathrm {S S L}} \\left(\\mathbf {z} _ {A}, \\mathbf {z} _ {B}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.788, + 0.826 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{SSL}}\\) is a loss function specific to the SSL pretraining method. After pretraining, the feature extractor \\(f_{\\theta_f^*}\\) can be deployed for various downstream tasks such as image classification, object detection, or segmentation. This is typically achieved by training a task-specific head. Alternatively, the feature extractor \\(f_{\\theta_f^*}\\) can be fine-tuned or used with a \\(k\\)-nearest neighbors (kNN) classifier." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Linear Probing There are several ways to evaluate the performance of a self-supervised learning method such as linear probing [10, 14, 29], kNN [8, 9, 62, 71], and few-shot evaluation [23, 26]. Consistently with the general protocol established in the literature [10, 15, 32, 67], we use linear probing to measure the quality of the features extracted for classification tasks. The procedure is: (1) a labeled downstream dataset, \\(\\mathbb{D}_{\\mathrm{task}}\\), consisting of image-class pairs \\((\\mathbf{x}, \\mathbf{y}) \\sim \\mathbb{D}_{\\mathrm{task}}\\), is selected for evaluation. (2) For each image \\(\\mathbf{x}\\), its representation is extracted using the pretrained feature extractor \\(f_{\\theta_f^*}\\), after which the linear classification head \\(t_{\\theta_t}\\), parameterized by \\(\\theta_t\\), is then applied to obtain \\(t_{\\theta_t}(f_{\\theta_f^*}(\\mathbf{x}))\\). (3) The linear head \\(t_{\\theta_t}\\) is optimized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.31, + 0.786, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} ^ {*} = \\underset {\\theta_ {t}} {\\arg \\min } \\mathbb {E} _ {(\\mathbf {x}, \\mathbf {y}) \\sim \\mathbb {D} _ {\\mathrm {t a s k}}} \\left[ \\mathcal {L} _ {\\mathrm {t a s k}} (t _ {\\theta_ {t}} (f _ {\\theta_ {f} ^ {*}} (\\mathbf {x})), \\mathbf {y}) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.349, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Note that only the parameters of the head \\(\\theta_t\\) are optimized, while the parameters \\(\\theta_f^*\\) of the feature extractor are frozen. The quality of features extracted by \\(f_{\\theta_f^*}\\) is directly inferred from the classification accuracy achieved on the test set of \\(\\mathbb{D}_{\\mathrm{task}}\\), which serves as the primary indicator of the quality of the extracted features." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.432, + 0.475, + 0.449 + ], + "angle": 0, + "content": "4 Normalized Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.539 + ], + "angle": 0, + "content": "We stress how for a correct understanding of the impact of data diversity we need to analyze its effects isolating them from the impact of increased computation. To enable this, we introduce (1) a computational budget used to normalize computation across experiments, and (2) a quantification of the data diversity seen by models during pretraining." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Computational Budget Current progress in SSL pretraining simultaneously scales computational budget and dataset size to achieve the best performance [26]. This makes it difficult to assess the reasons behind the success of many SSL algorithms. Do these methods benefit mainly from the large amounts of computation, i.e., running SSL pretraining for large numbers of epochs, or do they benefit from data diversity in larger datasets containing a vast variety of visual features? To perform a disentangled evaluation of these two factors, we first introduce \\(\\mathcal{C}\\) as a measure of computational budget, which quantifies the total number of images an SSL method is allowed to process during pretraining. This is calculated as \\(\\mathcal{C} = N\\cdot \\mathcal{E}\\), where \\(N\\) is the number of unique samples in the pretraining dataset \\(\\mathbb{D}_{\\mathrm{SSL}}\\), hence the data diversity of the dataset, and \\(\\mathcal{E}\\) is the number of pretraining epochs. Constraining multiple models pretrained with SSL to a fixed computational budget \\(\\mathcal{C}\\) allows for meaningful comparison, as it is guaranteed that all SSL methods will have processed the same number of pretraining images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Quantifying Pretraining Diversity Even under normalized computation \\(\\mathcal{C}\\), various SSL approaches may be exposed to different data diversity as they are trained with different datasets. For instance, a model trained for \\(\\mathcal{E} = 1000\\) epochs on a dataset of size \\(N = 1000\\) will see less diversity than a model pretrained for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.256 + ], + "angle": 0, + "content": "Table 1: Impact of Data Diversity on CIFAR-100 and Tiny ImageNet SSL Pretraining Performance: We study the effects of diversity on CIFAR-100 a and Tiny ImageNet b across seven different methods and three data diversity settings for a ResNet-18 pretraining, where for all, \\(\\mathbb{D}_{\\mathrm{task}} = \\mathbb{D}_{\\mathrm{SSL}}\\). This comparison includes analyzing classification accuracy through linear probing on the training set and evaluation on the test set of the respective datasets. Although performance fluctuates among different methods, a consistent trend is observed: higher data diversity typically leads to the generation of higher quality representations." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.268, + 0.498, + 0.334 + ], + "angle": 0, + "content": "
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
50.138.044.341.114.144.112.515.7
250.550.154.351.045.450.451.635.9
501.058.658.458.356.455.855.640.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.321, + 0.338, + 0.393, + 0.349 + ], + "angle": 0, + "content": "\\(\\mathcal{C} = 50\\times 10^{6}\\)" + }, + { + "type": "table_footnote", + "bbox": [ + 0.316, + 0.352, + 0.404, + 0.364 + ], + "angle": 0, + "content": "(a) CIFAR-100" + }, + { + "type": "list", + "bbox": [ + 0.316, + 0.338, + 0.404, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.269, + 0.783, + 0.334 + ], + "angle": 0, + "content": "
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
100.236.941.035.634.237.636.634.5
501.048.852.048.143.648.846.544.4
1002.049.855.650.347.554.148.647.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.607, + 0.338, + 0.678, + 0.349 + ], + "angle": 0, + "content": "\\(\\mathcal{C} = 50\\times 10^{6}\\)" + }, + { + "type": "table_footnote", + "bbox": [ + 0.589, + 0.352, + 0.7, + 0.364 + ], + "angle": 0, + "content": "(b) Tiny ImageNet" + }, + { + "type": "list", + "bbox": [ + 0.589, + 0.338, + 0.7, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.557 + ], + "angle": 0, + "content": "\\(\\mathcal{E} = 1\\) epoch on a dataset of size \\(N = 10^6\\), despite that they are both pretrained under normalized computation, processing the same number of images. Hence, to capture this notion of exposure to diversity while pretraining under normalized computation \\(\\mathcal{C}\\), we define a pretraining diversity \\(\\mathcal{D}\\), which captures the number of unique samples encountered during training given a fixed \\(\\mathcal{C}\\) as \\(\\mathcal{D} = N / \\mathcal{C} = 1 / \\varepsilon\\). A model pretrained with larger \\(\\mathcal{D}\\) indicates that the model is presented with a larger number of unique images during training with fixed \\(\\mathcal{C}\\), and hence is exposed to more pretraining data diversity. In the next section, we explore the effects of variations in \\(\\mathcal{D}\\) on SSL performance under a distribution shift, while keeping \\(\\mathcal{C}\\) constant." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.58, + 0.684, + 0.598 + ], + "angle": 0, + "content": "5 Fixed Budget SSL: In & Out-of-Distribution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Training Configuration We evaluate seven SSL methods: SimCLR [10], MoCoV3 [14], VICReg [4], DINO [9], BYOL [29], Barlow Twins [65], and SwAV [8]. We use different datasets for both pretraining \\(\\mathbb{D}_{\\mathrm{SSL}}\\) and linear probing \\(\\mathbb{D}_{\\mathrm{task}}\\) for different sections. We use solo-learn [16] as a codebase. For each method, we use the default parameters provided when available, otherwise, we conduct a hyperparameter search to ensure proper convergence." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.583, + 0.74 + ], + "angle": 0, + "content": "5.1 Performance on the Same Distribution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We aim to answer a first question: does collecting more samples from the same distribution help SSL pretraining with a fixed budget? We conduct experiments to capture how the pretraining diversity \\(\\mathcal{D}\\) influences SSL pretraining within a normalized \\(\\mathcal{C}\\), focusing on the simplest setting where the upstream and downstream datasets belong to the same distribution such that \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}\\). This serves as a fundamental ground for our further experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Setup For pretraining, we use CIFAR-100 [39] and Tiny ImageNet [40] as \\(\\mathbb{D}_{\\mathrm{SSL}}\\) which contain \\(50\\times 10^{3}\\) and \\(100\\times 10^{3}\\) images, respectively. We set \\(\\mathcal{C} = 50\\times 10^6\\) chosen such that it allows all methods to converge during pretraining. In Section 6, we study the effect of varying the budget \\(\\mathcal{C}\\). We pretrain on subsets of \\(\\mathbb{D}_{\\mathrm{SSL}}\\) with different sizes \\(N\\) (10%, 50%, and 100% of \\(\\mathbb{D}_{\\mathrm{SSL}}\\)), enabling us to observe the effects of pretraining diversity \\(\\mathcal{D}\\) on training outcomes where \\(\\mathcal{E}\\) is adjusted accordingly to match the budget \\(\\mathcal{C}\\). For example, using 100% of CIFAR-100 involves 1000 epochs of training, while 10% and 50% of the dataset lead to 10000 and 2000 epochs, respectively. These subsets of \\(\\mathbb{D}_{\\mathrm{SSL}}\\) are created by shuffling the dataset and then selecting the first 10%, 50%, or 100% split. All models in this section use a ResNet-18 [33] backbone pretrained from scratch. For evaluation, we use in-distribution linear probing, i.e., \\(\\mathbb{D}_{\\mathrm{task}} = \\mathbb{D}_{\\mathrm{SSL}}\\), where performance is measured by classification accuracy on the test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Results The results of linear probing are presented in Tables 1a and 1b for CIFAR-100 and Tiny ImageNet, respectively. While different SSL methods show varying levels of performance, we observe that, for all methods, an increase in the diversity \\(\\mathcal{D}\\) consistently leads to higher linear probing accuracy, suggesting that including more in-distribution samples in \\(\\mathbb{D}_{\\mathrm{SSL}}\\) helps under a normalized \\(\\mathcal{C}\\). For example, SimCLR achieves an accuracy of 37.95 when \\(10\\%\\) of CIFAR-100 is provided, whereas this performance improve by around \\(12\\%\\) when only \\(50\\%\\) of the dataset is provided and by another \\(8\\%\\) after pretraining on \\(100\\%\\) of unique samples. This suggests that SSL methods substantially benefit from having a more diverse pretraining set in computationally budgeted in-distribution evaluation, a fundamental verification that allows us to proceed in our analysis." + }, + { + "type": "title", + "bbox": [ + 0.241, + 0.537, + 0.334, + 0.554 + ], + "angle": 0, + "content": "Insight (1)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.565, + 0.763, + 0.626 + ], + "angle": 0, + "content": "When the distributions of the upstream and downstream tasks are the same, i.e., \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}\\), in a normalized computation setup, increasing pretraining diversity \\(\\mathcal{D}\\) proves to be an effective strategy for enhancing SSL pretraining performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.475, + 0.677 + ], + "angle": 0, + "content": "5.2 Increasing Data Diversity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As observed in Section 5.1, if \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}\\) having a more diverse pretraining set benefits SSL methods, under a normalized computation assumption. However, to increase diversity, sampling from the same distribution \\(\\mathbb{D}_{\\mathrm{task}}\\) to extend \\(\\mathbb{D}_{\\mathrm{SSL}}\\) is not always attainable. In fact, pretraining data is often sampled from a distribution different than \\(\\mathbb{D}_{\\mathrm{task}}\\). The added samples will then introduce a distribution shift between \\(\\mathbb{D}_{\\mathrm{SSL}}\\) and \\(\\mathbb{D}_{\\mathrm{task}}\\). Then, we raise the following question: is increasing pretraining diversity still beneficial when there is a distribution shift between \\(\\mathbb{D}_{\\mathrm{SSL}}\\) and \\(\\mathbb{D}_{\\mathrm{task}}\\), i.e., \\(\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}\\)? We explore strategies for acquiring new data for increasing \\(\\mathcal{D}\\), namely, including existing samples, crawled internet data, and data synthetically generated. To evaluate the effects of the distribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.127 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.145, + 0.703, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 2: Data Collection Strategies: We analyze strategies for collecting additional data (A), i.e., collecting more source data, crawling the web or using synthetic images. Using a class prior (top row) simulates In-distribution trainings. We also collect images without class prior (bottom row) to analyze the interactions between diversity and Out-of-distribution classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.458 + ], + "angle": 0, + "content": "shift in a controlled scenario, we analyze distributions closer to \\(\\mathbb{D}_{\\mathrm{task}}\\) by using a class prior (in-distribution) and without a class prior (out-of-distribution)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Setup We use ImageNet-100 [57] as \\(\\mathbb{D}_{\\mathrm{task}}\\), and we construct multiple \\(\\mathbb{D}_{\\mathrm{SSL}}\\) to evaluate the effects of different data collection strategies and the distribution shift. We first introduce a set \\(\\mathbb{B}\\) composed of \\(65\\times 10^{3}\\) images from ImageNet-100 (50% of the dataset), which we use as a baseline for \\(\\mathbb{D}_{\\mathrm{SSL}}\\) with minimum diversity. We denote the 100 classes in ImageNet-100 as \\(\\mathcal{T}_{100}\\), and sample \\(\\mathbb{B}\\) uniformly including images from all classes. Next, we compare with pretraining on more diverse datasets as \\(\\mathbb{D}_{\\mathrm{SSL}}\\), imposing \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{B} \\cup \\mathbb{A}\\) where \\(\\mathbb{A}\\) includes \\(65\\times 10^{3}\\) images sampled with one of three strategies. To highlight the effects of the distribution shift, we include in \\(\\mathbb{A}\\) either images from \\(In\\)-distribution classes, i.e., selecting images from classes overlapping with \\(\\mathcal{T}_{100}\\), or images from \\(Out\\)-of-distribution classes, which do not overlap with \\(\\mathcal{T}_{100}\\). This is to study the effects of the distribution shift, since we do not assume access to downstream classes in real scenarios. For the \\(Out\\)-of-distribution samples, we define \\(\\mathbb{A}\\) as (1) random images sampled from the full ImageNet [18] dataset; (2) images crawled from Flickr, Bing, and DuckDuckGo; or (3) images generated with Stable Diffusion V2.1 [51]. We respectively call these sets \\(\\mathbb{A}_{\\mathrm{Source}}^{Out}\\), \\(\\mathbb{A}_{\\mathrm{Web}}^{Out}\\), and \\(\\mathbb{A}_{\\mathrm{Synthetic}}^{Out}\\). We also define their \\(In\\)-distribution counterparts as \\(\\mathbb{A}_{\\mathrm{Source}}^{In}\\), \\(\\mathbb{A}_{\\mathrm{Web}}^{In}\\), and \\(\\mathbb{A}_{\\mathrm{Synthetic}}^{In}\\), respectively. Note how \\(\\mathbb{B} \\cup \\mathbb{A}_{\\mathrm{Source}}^{In}\\) is the full ImageNet-100, coherently with Section 5.1. Figure 2 shows each collection strategy, for which we provide implementation details in supplementary material. Although many factors (such as the images appearance) impact the distribution shift, using a class prior with any strategy would result in closer distribution with respect to the same strategy without class priors [15]. We pretrain a ResNet-18 from scratch, with the same settings of Section 5.1, and \\(\\mathcal{C} = 50\\times 10^6\\). Note that this results in \\(\\mathcal{D} = 1.25\\times 10^{-3}\\) for pretraining on \\(\\mathbb{B}\\), and \\(\\mathcal{D} = 2.5\\times 10^{-3}\\) for pretraining on \\(\\mathbb{B} \\cup \\mathbb{A}\\), introducing a \\(\\mathcal{D}\\) difference of a factor of 2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Fig. 3: Effect of Various Data Sources on SSL Pretraining: We use a baseline set \\(\\mathbb{B}\\) (black dashed line), comprising \\(65 \\times 10^{3}\\) images from ImageNet-100, for pretraining a ResNet-18 with \\(\\mathcal{C} = 50 \\times 10^{6}\\). Augmenting \\(\\mathbb{B}\\) with In-distribution images enhances performance (above black line), while Out-of-distribution augmentations reduce it (below black line)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.789, + 0.633 + ], + "angle": 0, + "content": "Results We report the linear probing accuracies for pretraining on each \\(\\mathbb{B} \\cup \\mathbb{A}\\) as bars in Figure 3, showing the \\(\\mathcal{C}\\)-normalized training on \\(\\mathbb{B}\\) as a dashed line. Surprisingly, without class priors \\((\\mathbf{Out})\\), including \\(\\mathbb{A}_{\\mathrm{Source}}^{Out}\\), \\(\\mathbb{A}_{\\mathrm{Web}}^{Out}\\), and \\(\\mathbb{A}_{\\mathrm{Synthetic}}^{Out}\\) underperforms compared to pretraining on \\(\\mathbb{B}\\) only. For instance, while SimCLR on \\(\\mathbb{B}\\) scores \\(72.3\\%\\) accuracy, increasing diversity reduces the accuracy by \\(1\\%\\) in all cases. This might appear to conflict with our findings in Section 5.1, however, the inclusion of \\(\\mathbf{Out}\\) samples leads to \\(\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}\\), since we sample only classes not included in \\(\\mathbb{D}_{\\mathrm{task}}\\). We infer that, with normalized \\(\\mathcal{C}\\), increasing \\(\\mathcal{D}\\) without distribution priors may negatively impact performance. Conversely, when class priors are available \\((\\mathbf{In})\\), increasing pretraining diversity \\(\\mathcal{D}\\) improves performance compared to \\(\\mathbb{B}\\) pretraining. For instance, pretraining on \\(\\mathbb{A}_{\\mathrm{Web}}^{In}\\) performs comparably to augmenting with additional ImageNet samples \\((\\mathbb{A}_{\\mathrm{Source}}^{In})\\), as in the case of SimCLR where the inclusion of \\(\\mathbb{A}_{\\mathrm{Web}}^{In}\\) scores only \\(0.2\\%\\) lower than \\(\\mathbb{A}_{\\mathrm{Source}}^{In}\\). Including \\(\\mathbb{A}_{\\mathrm{Synthetic}}^{In}\\) data also helps, although more limitedly due to the visual discrepancies between generated and real images. Ultimately, the effectiveness of diversity is linked to the distribution shift. These findings highlight the impact of the distribution shift on computationally-normalized SSL pretraining and help define evaluation practices for this task (see Section 7)." + }, + { + "type": "title", + "bbox": [ + 0.241, + 0.643, + 0.334, + 0.66 + ], + "angle": 0, + "content": "Insight (2)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.671, + 0.763, + 0.732 + ], + "angle": 0, + "content": "When the distributions of the upstream and downstream tasks differ, \\(\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}\\), and in a normalized computation setup, increasing pretraining diversity \\(\\mathcal{D}\\) may harm SSL pretraining performance, reflecting the adverse effects of distribution shift." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.502, + 0.785 + ], + "angle": 0, + "content": "5.3 Scaling Pretraining Diversity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We showed that diversity improves pretraining performance when the training set and downstream task share the same data distribution \\((\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}})\\), as discussed in Section 5.1. This may change when a distribution shift is introduced," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: Performance of ImageNet and YFCC100M SSL Pretraining on Various Downstream Tasks: We train ResNet-50 and ViT-B/16 under normalized computation \\((\\mathcal{C} = 98\\times 10^{6})\\) using SimCLR (left) and MoCoV3 (right) on ImageNet (IN) and YFCC100M (YFCC) with multiple \\(\\mathcal{D}\\). Despite the significantly larger \\(\\mathcal{D}\\) when trained on YFCC100M, these models cannot offset the effects of the distribution shift and are outperformed by models pretrained on ImageNet in most downstreams." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.242, + 0.498, + 0.395 + ], + "angle": 0, + "content": "
SimCLR
\\( \\mathbb{D}_{\\text{SSL}} \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3156.943.082.373.945.459.9
IN0.2562.6161.145.584.076.047.064.4
IN0.6406.5463.746.884.278.848.367.2
IN1.28113.064.546.484.679.548.868.0
YFCC98.17100057.337.276.658.950.162.1
ViT-B/16IN0.1281.3154.237.281.870.344.664.3
IN0.2562.6161.339.583.877.747.269.4
IN0.6406.5465.539.984.680.449.173.1
IN1.28113.066.739.583.781.750.073.6
YFCC98.17100054.525.072.559.749.565.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.326, + 0.399, + 0.389, + 0.409 + ], + "angle": 0, + "content": "\\(\\mathcal{C} = 98\\times 10^{6}\\)" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.242, + 0.785, + 0.395 + ], + "angle": 0, + "content": "
MoCoV3
\\( \\mathbb{D}_{\\text{SSL}} \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3158.140.681.876.0445.1
IN0.2562.6162.945.385.079.847.6
IN0.6406.5465.448.486.181.949.2
IN1.28113.065.948.886.682.649.5
YFCC98.17100060.442.282.666.350.767.3
ViT-B/16IN0.1281.3157.933.482.878.046.7
IN0.2562.6163.735.085.382.948.4
IN0.6406.5467.239.585.085.849.7
IN1.28113.068.841.986.586.550.3
YFCC98.17100057.225.270.343.850.364.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.612, + 0.398, + 0.673, + 0.408 + ], + "angle": 0, + "content": "\\(\\mathcal{C} = 98\\times 10^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.537 + ], + "angle": 0, + "content": "as explored in Section 5.2. However, it is still unclear from Section 5.2, whether including a larger number of samples, and thus increasing considerably the pretraining diversity, can compensate for the negative effects of the distribution shift. To address this, the following section presents larger-scale experiments, employing significantly varied \\(\\mathcal{D}\\) values, aiming to explore the interplay between pretraining diversity and different distributions using millions of samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.789, + 0.734 + ], + "angle": 0, + "content": "Setup For our large-scale pretraining experiments, we set \\(\\mathbb{D}_{\\mathrm{SSL}}\\) to be two datasets of significantly different sizes: ImageNet [18] and YFCC100M [55], comprising 1.28 million and 98 million images, respectively. Following Section 5.1, we explore multiple \\(\\mathcal{D}\\) values for pretraining. We set \\(\\mathcal{C} = 98 \\times 10^6\\), which corresponds to one epoch on YFCC100M, iterating once through each of its 98 million images to maximize diversity \\((\\mathcal{D} = 1)\\). Normalizing \\(\\mathcal{C}\\) (see Section 5.1), we pretrain on ImageNet for approximately 77 epochs, cumulatively utilizing 98 million samples. Due to the extensive cost of these experiments, we focus on SimCLR and MoCoV3 only. We also employ larger architectures, namely ResNet-50 [33] and ViT-B/16 [21], to leverage the extensive scale of the pretraining datasets. We also use multiple \\(\\mathbb{D}_{\\mathrm{task}}\\) including ImageNet [18], Stanford Cars [38], Flowers-102 [44], Oxford-IIIT Pets [48], Places365 [69], and Food-101 [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Results The results of our large-scale experiments are detailed in Table 2. Consistently with findings in Section 5.1, increasing \\(\\mathcal{D}\\) leads to better pretraining efficacy when \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}\\). This is evident when ImageNet is used for both pretraining and downstream classification, reinforcing that our observations hold even at a larger scale. Instead, models pretrained on YFCC100M show substantially lower performance compared to those pretrained on ImageNet, although" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "having much higher \\(\\mathcal{D}\\). This highlights the inefficiency of collecting data indiscriminately without considering distribution shifts. To stress this, note how the model pretrained on YFCC100M (\\(D = 1\\)) often performs similarly to those pretrained with drastically lower \\(\\mathcal{D}\\) on ImageNet (\\(D = 1.31 \\times 10^{-3}\\)). This aligns with our observations in Section 5.2, emphasizing that distribution differences remain a significant factor even when training with large datasets. However, the YFCC100M-pretrained model outperforms the ImageNet-pretrained model in Places365, suggesting a closer distribution relationship between YFCC100M and Places365. We explore this hypothesis further in Section 6, where we analyze distribution distances with existing metrics. Ultimately, our analysis highlights that scaling the data is not an effective solution for compensating the distribution shift, when computational resources are normalized." + }, + { + "type": "title", + "bbox": [ + 0.242, + 0.34, + 0.333, + 0.356 + ], + "angle": 0, + "content": "Insight (3)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.368, + 0.763, + 0.414 + ], + "angle": 0, + "content": "Even an extremely large data diversity cannot mitigate the distribution shift under normalized computation. This emphasizes the importance of further research in how effectively SSL pretraining algorithms generalize." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.444, + 0.453 + ], + "angle": 0, + "content": "6 Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Previously, we introduced a normalized evaluation to study the relation of \\(\\mathcal{D}\\) and data distributions. We showed that, although diversity helps when \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}\\), vision models are unable to compensate for the distributional differences with normalized costs. We now analyze additional elements supporting our findings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Distribution Distances Using FID & VisualDNA In Section 5.3, we showed that pretraining on ImageNet typically outperforms YFCC100M on a variety of downstream tasks, with Places365 being the exception. We speculate that the distribution of ImageNet is more aligned with those of the downstream tasks compared to YFCC100M. To verify this, we evaluate the similarity between the datasets using FID [34] and VisualDNA [50] and report results in Table 3. With both FID or VisualDNA, the distribution of ImageNet is always closer to the downstream tasks, except for Places365 where YFCC100M is closer. This aligns with the lower performance of ImageNet on this dataset only (Table 2), further suggesting that the performance drop is caused by the distribution shift." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Importance of Normalized Computation We now study the impact of normalizing computation on the performance of SSL methods. We aim to understand if \\(\\mathcal{C}\\) is not normalized across methods, this will lead to misleading conclusions, and thus motivating our normalized computation. In Table 4, we compare the performance of ResNet-50 and ViT-B/16 models pretrained using MoCoV3 with \\(\\mathcal{C} = 98 \\times 10^6\\) (as used in Section 5.3) against a tripled budget of \\(\\mathcal{C} = 294 \\times 10^6\\). We show that inconsistencies arise when pretraining data diversity \\(\\mathcal{D}\\) and computation \\(\\mathcal{C}\\) are not fairly compared. For instance, in scenarios highlighted in red, we notice that a lower pretraining diversity coupled with a higher" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.145, + 0.493, + 0.381 + ], + "angle": 0, + "content": "Table 3: Comparing VisualDNA and FID Scores Across Datasets: We assess the relationship between VisualDNA [50] and FID Score [34] for various large-scale \\(\\mathbb{D}_{\\mathrm{SSL}}\\) and several downstream tasks. For VisualDNA, activations are extracted as suggested [50] with MUGS [70] with ViT-B/16 (MUGS) or DINO v2 [47] with ViT-B/16 (DINO v2), while FID activations are obtained via an Inception network [54]. Consistently, across all metrics, the ranking of dataset distances between \\(\\mathbb{D}_{\\mathrm{SSL}}\\) and various \\(\\mathbb{D}_{\\mathrm{task}}\\) aligns with the accuracy ranking in Table 2. Models exhibiting lower VisualDNA/FID scores benefit more from the diversity in pretraining data for SSL." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.393, + 0.493, + 0.524 + ], + "angle": 0, + "content": "
DSSLBackboneVisualDNA↓
INCarsFlow.PetsPlacesFood
INMUGS011.4912.466.097.199.08
YFCCMUGS3.7211.5712.717.936.209.72
INDINO v207.056.954.433.527.26
YFCCDINO v22.377.127.465.752.747.90
DSSLBackboneFID↓
INCarsFlow.PetsPlacesFood
INInception0143.4192.988.964.9114.9
YFCCInception48.1174.3214.8145.838.9154.5
" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.148, + 0.784, + 0.33 + ], + "angle": 0, + "content": "Table 4: Importance of Normalization. We report the accuracy of MoCoV3 trained on ImageNet (IN) with different data diversity and variable \\(\\mathcal{C}\\), for the in-distribution assumption \\((\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}})\\). For a given network, cells in red highlight inconsistencies where although the model trained with \\(\\mathcal{C} = 294 \\times 10^6\\) has seen less samples, it outperforms the best model trained with one third of the computational budget \\((\\mathcal{C} = 90 \\times 10^6)\\), showing the importance of normalization for understanding how models exploit data." + }, + { + "type": "table_caption", + "bbox": [ + 0.61, + 0.341, + 0.678, + 0.352 + ], + "angle": 0, + "content": "MoCoV3" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.353, + 0.779, + 0.525 + ], + "angle": 0, + "content": "
Network\\( \\mathbb{D}_{\\text{SSL}} \\)\\( \\mathcal{C} \\times 10^6 \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy† IN
ResNet-50IN980.1281.3158.1
IN980.6406.5465.4
IN981.28113.065.9
IN2940.1280.4357.4
IN2940.6402.1769.8
IN2941.2814.3571.4
ViT-B/16IN980.1281.3157.9
IN980.6406.5467.2
IN981.28113.068.8
IN2940.1280.4356.9
IN2940.6402.1771.9
IN2941.2814.3574.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.641 + ], + "angle": 0, + "content": "computational budget can yield better results. We observe that under pretraining diversity of only \\(\\mathcal{D} = 13\\times 10^{-3}\\) and a computational budget of \\(\\mathcal{C} = 98\\times 10^6\\) ResNet-50 only enjoys a \\(65.9\\%\\) accuracy compared against with \\(69.8\\%\\) with a smaller pretraining diversity of \\(2.17\\times 10^{-3}\\) but with a larger computational budget of \\(294\\times 10^{6}\\). This shows that without normalized computation, it could be incorrectly concluded that pretraining diversity does not play a significant role." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Model Saturation We aim to study the pretraining trend on YFCC100M, and if adding more samples could compensate the distribution shift with ImageNet. Hence, we extend the YFCC100M experiments from Section 5.3, examining various subsets—specifically \\(0.1\\%\\), \\(1\\%\\), \\(10\\%\\), and \\(100\\%\\) of the dataset. Again, we normalize the computational budget to \\(\\mathcal{C} = 98 \\times 10^6\\), equivalent to one epoch on YFCC100M. The results of linear probing models pretrained using SimCLR and MoCoV3 with ResNet-50 and ViT-B/16 on ImageNet are shown in Figure 4. Interestingly, we obtain a performance plateau. This observation points to a saturation point in performance, showing that simply increasing the dataset further would not bridge the gap between the \\(\\mathbb{D}_{\\mathrm{SSL}}\\) and \\(\\mathbb{D}_{\\mathrm{task}}\\). Hence, arbitrarily increasing pretraining diversity \\(\\mathcal{D}\\) is not sufficient to bridge the distribution shift problem. Here \\(\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}\\), hence still aligning with our findings in Section 5.1." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Fig. 4: Data Diversity Impact on YFCC100M Pretraining Performance: Pretraining \\((\\mathcal{C} = 98\\times 10^{6})\\) of networks with \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathrm{YFCC100M}\\) and \\(\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}\\) for several dataset subsets. In the presence of a distribution shift, performance tends to saturate and does not benefit from additional data diversity." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Table 5: Evaluating Network Accuracy With Varied Label Quantity: We evaluate the accuracy of networks trained on ImageNet (IN) and YFCC100M (YFCC) with different labeling percentages of \\(\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}\\). The increased diversity still does not compensate for the distribution shift. However, for in-distribution data, one can get away with fewer labels with more diverse pretraining data." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.43, + 0.498, + 0.509 + ], + "angle": 0, + "content": "
SimCLR
Network\\( \\mathbb{D}_{\\text{SSL}} \\)N×106D×10-3Accuracy†
5%10%50%
ResNet-50IN1.281.3150.354.261.7
YFCC98.2100039.444.354.0
ViT-B/16IN1.281.3155.259.465.6
YFCC98.2100040.745.753.3
C=98×10^6
" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.43, + 0.787, + 0.509 + ], + "angle": 0, + "content": "
NetworkMoCoV3
\\( \\mathbb{D}_{\\text{SSL}} \\)N×106D×10-3Accuracy↑
5%10%50%
ResNet-50IN1.281.3152.156.263.8
YFCC98.2100042.848.257.7
ViT-B/16IN1.281.3159.363.068.3
YFCC98.2100043.949.056.0
C=98×10^6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Label Quantity A question that arises in our setting is whether having a higher pretraining diversity leads to requiring less labeled samples during linear probing. In this section, we focus on understanding the impact of increased \\(\\mathcal{D}\\) on the number of labeled samples required for effective linear probing in downstream tasks. We use the same trained models from Section 5.3, i.e., SimCLR and MoCoV3, with ResNet-50 and ViT-B/16 architectures. In this setup, we set \\(\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}\\), with two upstream dataset scenarios: ImageNet (where \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}\\)) and YFCC100M (where \\(\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}\\)). Our experiments are summarized in Table 5. We note that with ViT-B/16, if ImageNet is used for pretraining, linear probing with just \\(5\\%\\) labeled samples can surpass the performance achieved using \\(100\\%\\) labeled data when YFCC100M serves as \\(\\mathbb{D}_{\\mathrm{SSL}}\\). This also implies that when \\(\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}\\), a higher quantity of labeled data is necessary to attain competitive performance, and increasing \\(\\mathcal{D}\\) does not bring considerable benefits on label quantity under distribution shift. This implies that our findings in Section 5.3 hold regardless of the linear probing labeled set. We note that in scenarios where \\(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}\\), using only \\(50\\%\\) of the labeled data can achieve similar performance as utilizing the full \\(100\\%\\) of labeled samples, implying that increasing \\(\\mathcal{D}\\) leads to reduced label requirement efforts for downstream tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.355, + 0.161 + ], + "angle": 0, + "content": "7 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.186, + 0.788, + 0.459 + ], + "angle": 0, + "content": "Main Conclusions Our set of experiments leads to Insight (3) in Section 5.3, revealing that with normalized computational costs, SSL pretrainings with large diversity cannot compensate for the distribution shift. This is surprising, since the variety of information that SSL algorithms could benefit from during training is much higher in large generic datasets than in small ones. Since our evaluation is cost-normalized, (3) also shows that SSL strategies are not efficiently exploiting the pretraining diversity on large datasets for representation extraction. This inefficiency implies a wide margin for improvement of generalization of SSL models, making better use of the computational power used for training. The role of existing models must also be discussed in this context. Following Insights (1) and (2), respectively in Sections 5.1 and 5.2, we have studied how in-distribution and out-of-distribution data impact performance in a controlled scenario. We argue that this should be taken into account in the evaluation of the performance of SSL models. Indeed, training at scale may enlarge the in-distribution space, including classes of the downstream tasks in the training set. In recent literature, this is a design choice to maximize performance on popular benchmarks [47]. While this allows for achieving impressive results in many tasks, we stress that this does not permit a fair evaluation. Now, we summarize practical takeaways." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Training Takeaways Coherently with our findings in Section 5.1 and similarly to prior art [15,28,37], we find that aligned distributions benefit performance, in particular increasing \\(\\mathbb{D}_{\\mathrm{SSL}}\\) diversity helps, as long as the distribution of the new samples match those of the downstream task data. Differently from the state-of-the-art, we demonstrate that this holds also in a computationally-normalized setting, implying that collecting large scale in-distribution data matching the downstream task is be an effective approach for improving SSL. Hence, for practical applications distribution priors should be used, if available. On the contrary, for a fair evaluation of models, this should not be the case, as specified below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Evaluation Takeaways Our analysis reveals that to permit a fair evaluation of SSL methods, computationally normalized tests are necessary to avoid inconsistencies, as shown in Section 6. Moreover, it is crucial to identify out-of-distribution downstream tasks for a correct evaluation of generalization performance. By evaluating only on \\(\\mathbb{D}_{\\mathrm{task}}\\) with a low distribution shift, there is a risk of reporting inflated metrics, not representative of a real gain in generalization. This is important, since new SSL approaches may be reporting higher downstream performance when pretrained on a different dataset. We relate this to Sections 5 and 6, where we show that increasing the computation and the in-distribution data, respectively, can improve performance. Ultimately, wrong practices may result in incorrectly concluding that an SSL algorithm is performing better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Differences With Language Models In Section 5.3 we showed that even very diverse datasets (YFCC100M), fall short in satisfactory generalization performance. In addition to further exploration into generalization for SSL pretraining, this open doors to investigating why language models enjoy enhanced generalization when trained on highly diverse data compared to vision models [61]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.254 + ], + "angle": 0, + "content": "This work was supported by SDAIA-KAUST Center of Excellence in Data Science, Artificial Intelligence (SDAIA-KAUST AI). Fabio Pizzati is financed by KAUST (Grant DFR07910). Philip H.S. Torr thanks the Royal Academy of Engineering for their support. This work is supported by a UKRI grant Turing AI Fellowship (EP/W002981/1)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.278, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Bachman, P., Hjelm, R.D., Buchwalter, W.: Learning representations by maximizing mutual information across views. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.38 + ], + "angle": 0, + "content": "2. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.407 + ], + "angle": 0, + "content": "3. Bao, H., Dong, L., Piao, S., Wei, F.: BEit: BERT pre-training of image transformers. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.785, + 0.435 + ], + "angle": 0, + "content": "4. Bardes, A., Ponce, J., LeCun, Y.: VICReg: Variance-invariance-covariance regularization for self-supervised learning. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.464 + ], + "angle": 0, + "content": "5. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-Mining discriminative components with random forests. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.466, + 0.785, + 0.492 + ], + "angle": 0, + "content": "6. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.519 + ], + "angle": 0, + "content": "7. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.547 + ], + "angle": 0, + "content": "8. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.575 + ], + "angle": 0, + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.603 + ], + "angle": 0, + "content": "0. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.631 + ], + "angle": 0, + "content": "1. Chen, T., Kornblith, S., Swersky, K., Norouzi, M., Hinton, G.: Big self-supervised models are strong semi-supervised learners. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.659 + ], + "angle": 0, + "content": "2. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.686 + ], + "angle": 0, + "content": "3. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.715 + ], + "angle": 0, + "content": "4. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "5. Cole, E., Yang, X., Wilber, K., Aodha, O.M., Belongie, S.: When does contrastive visual representation learning work? In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.77 + ], + "angle": 0, + "content": "6. Costa, V.G.T.d., Fini, E., Nabi, M., Sebe, N., Ricci, E.: solo-learn: A library of self-supervised methods for visual representation learning. In: JMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "7. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. Doersch, C., Zisserman, A.: Multi-task self-supervised visual learning. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.258 + ], + "angle": 0, + "content": "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "22. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "23. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "24. Fini, E., Astolfi, P., Romero-Soriano, A., Verbeek, J., Drozdzal, M.: Improved baselines for vision-language pre-training. TMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "25. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "26. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "27. Goyal, P., Duval, Q., Seessel, I., Caron, M., Misra, I., Sagun, L., Joulin, A., Bojanowski, P.: Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "28. Goyal, P., Mahajan, D., Gupta, A., Misra, I.: Scaling and benchmarking self-supervised visual representation learning. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "29. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent: A new approach to self-supervised learning. In: NeurIPS. vol. 33 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "30. Hammoud, H.A.A.K., Itani, H., Pizzati, F., Torr, P., Bibi, A., Ghanem, B.: Synthclip: Are we ready for a fully synthetic clip training? arXiv (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "31. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "32. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "33. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "35. Hjelm, R.D., Fedorov, A., Lavoie-Marchildon, S., Grewal, K., Bachman, P., Trischler, A., Bengio, Y.: Learning deep representations by mutual information estimation and maximization. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "36. Henaff, O.J., Srinivas, A., De Fauw, J., Razavi, A., Doersch, C., Eslami, S.M.A., Oord, A.v.d.: Data-efficient image recognition with contrastive predictive coding. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "37. Kotar, K., Ilharco, G., Schmidt, L., Ehsani, K., Mottaghi, R.: Contrasting contrastive self-supervised representation learning pipelines. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "38. Krause, J., Deng, J., Stark, M., Fei-Fei, L.: Collecting a large-scale dataset of fine-grained cars. In: FGVC (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Pretraining Data Diversity for SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.762, + 0.162 + ], + "angle": 0, + "content": "40. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.785, + 0.189 + ], + "angle": 0, + "content": "41. Li, A.C., Brown, E.L., Efros, A.A., Pathak, D.: Internet explorer: Targeted representation learning on the open web. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.218 + ], + "angle": 0, + "content": "42. Li, J., Zhou, P., Xiong, C., Hoi, S.C.H.: Prototypical contrastive learning of unsupervised representations. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "43. Misra, I., Maaten, L.v.d.: Self-supervised learning of pretext-invariant representations. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.273 + ], + "angle": 0, + "content": "44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "45. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "46. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "47. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINoV2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.425 + ], + "angle": 0, + "content": "49. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.453 + ], + "angle": 0, + "content": "50. Ramtoula, B., Gadd, M., Newman, P., De Martini, D.: Visual DNA: Representing and comparing images using distributions of neuron activations. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.48 + ], + "angle": 0, + "content": "51. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "52. Shi, Y., Daunhawer, I., Vogt, J.E., Torr, P., Sanyal, A.: How robust is unsupervised representation learning to distribution shift? In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "53. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "54. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "55. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: YFCC100M: The new data in multimedia research. In: Commun. ACM (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.633 + ], + "angle": 0, + "content": "56. Tian, Y., Henaff, O.J., Van Den Oord, A.: Divide and contrast: Self-supervised learning from uncurated data. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.647 + ], + "angle": 0, + "content": "57. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.74, + 0.661 + ], + "angle": 0, + "content": "58. Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: CVPR (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "59. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Gool, L.V.: Revisiting contrastive methods for unsupervised learning of visual representations. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "60. Van Horn, G., Cole, E., Beery, S., Wilber, K., Belongie, S., Mac Aodha, O.: Benchmarking representation learning for natural world image collections. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "61. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. TMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "62. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via non-parametric instance discrimination. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "63. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMIM: A simple framework for masked image modeling. In: CVPR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Hammoud et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Wei, Y., Dai, Q., Hu, H.: On data scaling in masked image modeling. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "65. Zbontar, J., Jing, L., Misra, I., LeCun, Y., Deny, S.: Barlow Twins: Self-supervised learning via redundancy reduction. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.23 + ], + "angle": 0, + "content": "66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.767, + 0.245 + ], + "angle": 0, + "content": "67. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "68. Zhao, N., Wu, Z., Lau, R.W.H., Lin, S.: What makes instance discrimination good for transfer learning? In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.785, + 0.3 + ], + "angle": 0, + "content": "69. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. T-PAMI (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "70. Zhou, P., Zhou, Y., Si, C., Yu, W., Ng, T.K., Yan, S.: Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "71. Zhuang, C., Zhai, A.L., Yamins, D.: Local aggregation for unsupervised learning of visual embeddings. In: ICCV (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_origin.pdf b/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4ef5153578669015b9de6ad5feadf70b84e3c0ba --- /dev/null +++ b/2024/On Pretraining Data Diversity for Self-Supervised Learning/2c80ad0d-b5b2-4cfa-82ca-d7334be6fefa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33e276703f60baf7ef92cbbf95e36a1fee0bfb9e0231a6d6b6dae6c006bbbd85 +size 5232060 diff --git a/2024/On Pretraining Data Diversity for Self-Supervised Learning/full.md b/2024/On Pretraining Data Diversity for Self-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc1c2c24ec58255195dae682402e46f5fe6e049a --- /dev/null +++ b/2024/On Pretraining Data Diversity for Self-Supervised Learning/full.md @@ -0,0 +1,265 @@ +# On Pretraining Data Diversity for Self-Supervised Learning + +Hasan Abed Al Kader Hammoud $^{1,2\star\diamond}$ , Tuhin Das $^{2\star}$ , Fabio Pizzati $^{2\star}$ Philip H.S. Torr $^{2}$ , and Adel Bibi $^{2\dagger}$ , Bernard Ghanem $^{1\dagger}$ + +1 KAUST + +2 University of Oxford + +Abstract. We explore the impact of training with more diverse datasets, characterized by the number of unique samples, on the performance of self-supervised learning (SSL) under a fixed computational budget. Our findings demonstrate that increasing pretraining data diversity enhances SSL performance, albeit only when the distribution distance to the downstream data is minimal. Notably, even with an exceptionally large pretraining data diversity achieved through methods like web crawling or diffusion-generated data, among other ways, the distribution shift remains a challenge. Our experiments are comprehensive with seven SSL methods using large-scale datasets such as ImageNet and YFCC100M amounting to over 200 GPU days. The code and trained models will be available at https://github.com/hammoudhasan/DiversitySSL. + +Keywords: self-supervised learning $\cdot$ distribution shift $\cdot$ data diversity + +# 1 Introduction + +Self-supervised learning (SSL) has recently emerged as a new paradigm to pretrain large vision models at scale [26, 27, 47]. Leveraging the ability to learn from unlabelled data, pretraining on millions—or even billions [27, 53]—of images turned from an unachievable goal to a common practice. This exposure to extremely diverse datasets, i.e., composed of a remarkable number of unique samples, granted impressive performance and unprecedented generalization capabilities to a growing number of vision models [47]. Large-scale datasets, in conjunction with substantial computational resources, have been the key driving forces for the success of SSL-based approaches. For instance, SEER [26], pretrained for approximately 11 GPU years on a billion images, exemplifies the massive computation and data resources employed for these models. It has become the implicit norm that increasing computation and data is beneficial, without any detailed analysis of how they separately impact SSL effectiveness. In particular, it is not clear to what extent large datasets are responsible for the impressive generalization capabilities of SSL models. Indeed, consider the + +![](images/9efeea3bf096c200ff57ab6b0029609544d4d0d2f34fea7b857437a9d115c3d2.jpg) +Fig. 1: Impact of Pretraining Diversity: Self-supervised learning (SSL) can be used to pretrain vision models on small datasets closely aligned to the downstream task, e.g., pets classification, hence with a small distribution shift (top, wild animals pretraining). Conversely, we could pretrain on an extensively varied dataset, with wide distribution differences (outdoor scenes, bottom). We study the role of pretraining diversity in SSL under a fixed budget, and highlight its effects in relationship to the distribution shift. + +example in Figure 1. Assuming a fixed monetary budget, in the form of computational expenses, for pretraining a vision model under the SSL paradigm, does iterating over a large set of images work best given a downstream task, or it is better to train repeatedly on a smaller set of samples visually close to the ones of the downstream? In this context, we face the ever-lasting problem of distribution shift on vision models [58]. Without a proper understanding of the impact of data, there could be a wrong allocation of efforts in increasing data or computation, leading to inefficiencies in the deployment process of SSL models. + +In this paper, we study the role of the diversity of the SSL pretraining data, under a fixed computational budget, when the pretraining data matches the distribution of the downstream data, as well as when they differ. Our experiments span various SSL methods [4,8-10, 14, 29, 65] and multiple datasets [5,18,38-40, 44,48,55,57,69], and are tested under several computational budgets amounting to a total of 200 GPU days. We summarize our contributions below: + +1. We show that SSL pretraining strategies are currently data-inefficient in compensating for distribution shifts. Under normalized computational costs, we verify that pretraining on large datasets with high diversity cannot outperform the pretraining on datasets with limited diversity, but with a lower distribution shift with respect to the downstream tasks. +2. We conclude that there is a wide margin for improvement in the performance of current SSL methods on out-of-distribution classes, and we propose in + +sights for a fair evaluation. This shows the need for future research on new SSL techniques leveraging data diversity better, to improve generalization capabilities beyond the training data distribution. + +3. We propose a novel strategy for a computationally-normalized evaluation of SSL and a carefully designed set of experiments focusing on pretraining data diversity, enabling us to draw our novel conclusions. + +Ultimately, our work provides a comprehensive analysis, from experiments worth 200 GPU days on NVIDIA A100, on the interaction between data diversity and computational resources, and their impact on the performance of SSL models. + +# 2 Related Work + +Self-Supervised Learning Early work in self-supervised learning used simple pretext tasks, such as relative patch prediction [19, 20], image colorization [67], image rotation prediction [25], or solving jigsaw puzzles [45] to train feature extractors in absence of annotations. Representations learned with those methods are limitedly effective, hence more recent literature has moved towards more sophisticated approaches, such as using image augmentations to generate correlated views of a training sample, and learning to extract augmentation-invariant representations for these correlated pairs. Among these multi-view methods, many exploit contrastive losses [1,8,10-12,32,35,36,42,43,46], enforcing similarity between views of the same image (positives) and dissimilarity from other images (negatives). Due to the need of many negative samples, contrastive methods often require large batch sizes to work effectively [10,14]. Cluster-based methods such as SwAV [8], DINO [9] and DeepCluster v2 [6] learn generalizable representations by grouping samples into cluster prototypes. Others exploit prediction of features with siamese networks [13], learn features in a teacher-student fashion [29], or use redundancy reduction techniques [4,65]. Finally, masked image modeling [3,31,63] emerged as a scalable alternative for Vision Transformers [21], learning representations by predicting masked image patches as SSL task. + +Pretraining at Scale SSL pretraining is most effective at scale, benefiting from large datasets and great computational resources. Initial attempts at large-scale pretraining were made by combining contrastive learning and clustering [7, 56]. SEER [26] was trained on 1 billion internet images with SwAV [8] on a RegNet [49] backbone. The importance of model scaling to leverage large datasets is also shown in [27, 28, 64], as well as the need for longer training times [64]. Additional strategies are necessary to achieve best performance [17, 24, 66]. These works focus on reaching the best representations, without considerations about training costs, thus encouraging unfair comparisons across trainings. Preliminary papers [15, 22] found that increasing data may lead to a modest boost in downstream performance when the downstream distribution changes. While they give insights on the pretraining diversity, its relation with the distribution shift under normalized computation is still not investigated. + +Distribution Shift in SSL Some studies have investigated how the pretraining data affects downstream performance on other domains. In general, the distribution shift increases when factors such as visual content and data acquisition procedure differ in two data ensembles [58]. For various pretraining datasets, preliminary works [15, 28, 37, 41] observed that the best performing representations were learned on datasets that were similar to the downstream test task. Additionally, combining datasets before pretraining or combining self-supervised features learned from various datasets did not lead to significant improvements in [15]. In [52, 68], they showcase higher generalization capabilities of SSL models compared to their supervised counterparts, for several downstream tasks in the presence of a distribution shift. In [59], they pretrained on several datasets, observing different generalization depending on the appearance of the pretraining dataset. Furthermore, initial considerations on the impact of external data on downstream tasks under distribution shift have been proposed in [60]. Some compensate the distribution shift with the scale of training datasets [30]. However, there is still a lack of a fair, computation-normalized evaluation that allows to study the effects of the distribution shift in a controlled environment. + +# 3 Preliminaries + +Pretraining We first outline the general pretraining procedure common to state-of-the-art self-supervised learning methods. Specific differences among these methods are detailed in the supplementary material. The overall pretraining pipeline, common across many SSL approaches [4, 8-10, 29, 65], goes as follows: (1) Data Sampling: from a large-scale, unlabeled upstream pretraining dataset, denoted as $\mathbb{D}_{\mathrm{SSL}}$ , an image $\mathbf{x}$ is randomly sampled; (2) View Generation: two correlated views, $\tilde{\mathbf{x}}_A$ and $\tilde{\mathbf{x}}_B$ , are generated from $\mathbf{x}$ using two image augmentation functions sampled at random. These random augmentations include random resized cropping, horizontal flipping, blurring, and color adjustments [2], among others; (3) Feature Extraction and Projection: the correlated views undergo feature extraction through a network, $f_{\theta_f}$ , parameterized by $\theta_f$ , such as a ResNet [33] or a ViT [21], leading to representations $\mathbf{h}_A = f_{\theta_f}(\tilde{\mathbf{x}}_A)$ and $\mathbf{h}_B = f_{\theta_f}(\tilde{\mathbf{x}}_B)$ . A linear projection head, $g_{\theta_g}$ , parameterized by $\theta_g$ , then maps these representations to a latent space, resulting in $\mathbf{z}_A = g_{\theta_g}(\mathbf{h}_A)$ and $\mathbf{z}_B = g_{\theta_g}(\mathbf{h}_B)$ ; (4) Joint Optimization: The feature extractor $f_{\theta_f}$ and the projection head $g_{\theta_g}$ are jointly optimized according to the following objective: + +$$ +\theta_ {f} ^ {*}, \theta_ {g} ^ {*} = \underset {\theta_ {f}, \theta_ {g}} {\arg \min } \mathbb {E} _ {\mathbf {x} \sim \mathbb {D} _ {\mathrm {S S L}}} \mathcal {L} _ {\mathrm {S S L}} \left(\mathbf {z} _ {A}, \mathbf {z} _ {B}\right), \tag {1} +$$ + +where $\mathcal{L}_{\mathrm{SSL}}$ is a loss function specific to the SSL pretraining method. After pretraining, the feature extractor $f_{\theta_f^*}$ can be deployed for various downstream tasks such as image classification, object detection, or segmentation. This is typically achieved by training a task-specific head. Alternatively, the feature extractor $f_{\theta_f^*}$ can be fine-tuned or used with a $k$ -nearest neighbors (kNN) classifier. + +Linear Probing There are several ways to evaluate the performance of a self-supervised learning method such as linear probing [10, 14, 29], kNN [8, 9, 62, 71], and few-shot evaluation [23, 26]. Consistently with the general protocol established in the literature [10, 15, 32, 67], we use linear probing to measure the quality of the features extracted for classification tasks. The procedure is: (1) a labeled downstream dataset, $\mathbb{D}_{\mathrm{task}}$ , consisting of image-class pairs $(\mathbf{x}, \mathbf{y}) \sim \mathbb{D}_{\mathrm{task}}$ , is selected for evaluation. (2) For each image $\mathbf{x}$ , its representation is extracted using the pretrained feature extractor $f_{\theta_f^*}$ , after which the linear classification head $t_{\theta_t}$ , parameterized by $\theta_t$ , is then applied to obtain $t_{\theta_t}(f_{\theta_f^*}(\mathbf{x}))$ . (3) The linear head $t_{\theta_t}$ is optimized as follows: + +$$ +\theta_ {t} ^ {*} = \underset {\theta_ {t}} {\arg \min } \mathbb {E} _ {(\mathbf {x}, \mathbf {y}) \sim \mathbb {D} _ {\mathrm {t a s k}}} \left[ \mathcal {L} _ {\mathrm {t a s k}} (t _ {\theta_ {t}} (f _ {\theta_ {f} ^ {*}} (\mathbf {x})), \mathbf {y}) \right], \tag {2} +$$ + +Note that only the parameters of the head $\theta_t$ are optimized, while the parameters $\theta_f^*$ of the feature extractor are frozen. The quality of features extracted by $f_{\theta_f^*}$ is directly inferred from the classification accuracy achieved on the test set of $\mathbb{D}_{\mathrm{task}}$ , which serves as the primary indicator of the quality of the extracted features. + +# 4 Normalized Evaluation + +We stress how for a correct understanding of the impact of data diversity we need to analyze its effects isolating them from the impact of increased computation. To enable this, we introduce (1) a computational budget used to normalize computation across experiments, and (2) a quantification of the data diversity seen by models during pretraining. + +Computational Budget Current progress in SSL pretraining simultaneously scales computational budget and dataset size to achieve the best performance [26]. This makes it difficult to assess the reasons behind the success of many SSL algorithms. Do these methods benefit mainly from the large amounts of computation, i.e., running SSL pretraining for large numbers of epochs, or do they benefit from data diversity in larger datasets containing a vast variety of visual features? To perform a disentangled evaluation of these two factors, we first introduce $\mathcal{C}$ as a measure of computational budget, which quantifies the total number of images an SSL method is allowed to process during pretraining. This is calculated as $\mathcal{C} = N\cdot \mathcal{E}$ , where $N$ is the number of unique samples in the pretraining dataset $\mathbb{D}_{\mathrm{SSL}}$ , hence the data diversity of the dataset, and $\mathcal{E}$ is the number of pretraining epochs. Constraining multiple models pretrained with SSL to a fixed computational budget $\mathcal{C}$ allows for meaningful comparison, as it is guaranteed that all SSL methods will have processed the same number of pretraining images. + +Quantifying Pretraining Diversity Even under normalized computation $\mathcal{C}$ , various SSL approaches may be exposed to different data diversity as they are trained with different datasets. For instance, a model trained for $\mathcal{E} = 1000$ epochs on a dataset of size $N = 1000$ will see less diversity than a model pretrained for + +Table 1: Impact of Data Diversity on CIFAR-100 and Tiny ImageNet SSL Pretraining Performance: We study the effects of diversity on CIFAR-100 a and Tiny ImageNet b across seven different methods and three data diversity settings for a ResNet-18 pretraining, where for all, $\mathbb{D}_{\mathrm{task}} = \mathbb{D}_{\mathrm{SSL}}$ . This comparison includes analyzing classification accuracy through linear probing on the training set and evaluation on the test set of the respective datasets. Although performance fluctuates among different methods, a consistent trend is observed: higher data diversity typically leads to the generation of higher quality representations. + +
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
50.138.044.341.114.144.112.515.7
250.550.154.351.045.450.451.635.9
501.058.658.458.356.455.855.640.9
+ +$\mathcal{C} = 50\times 10^{6}$ +(a) CIFAR-100 + +
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
100.236.941.035.634.237.636.634.5
501.048.852.048.143.648.846.544.4
1002.049.855.650.347.554.148.647.3
+ +$\mathcal{C} = 50\times 10^{6}$ +(b) Tiny ImageNet + +$\mathcal{E} = 1$ epoch on a dataset of size $N = 10^6$ , despite that they are both pretrained under normalized computation, processing the same number of images. Hence, to capture this notion of exposure to diversity while pretraining under normalized computation $\mathcal{C}$ , we define a pretraining diversity $\mathcal{D}$ , which captures the number of unique samples encountered during training given a fixed $\mathcal{C}$ as $\mathcal{D} = N / \mathcal{C} = 1 / \varepsilon$ . A model pretrained with larger $\mathcal{D}$ indicates that the model is presented with a larger number of unique images during training with fixed $\mathcal{C}$ , and hence is exposed to more pretraining data diversity. In the next section, we explore the effects of variations in $\mathcal{D}$ on SSL performance under a distribution shift, while keeping $\mathcal{C}$ constant. + +# 5 Fixed Budget SSL: In & Out-of-Distribution + +Training Configuration We evaluate seven SSL methods: SimCLR [10], MoCoV3 [14], VICReg [4], DINO [9], BYOL [29], Barlow Twins [65], and SwAV [8]. We use different datasets for both pretraining $\mathbb{D}_{\mathrm{SSL}}$ and linear probing $\mathbb{D}_{\mathrm{task}}$ for different sections. We use solo-learn [16] as a codebase. For each method, we use the default parameters provided when available, otherwise, we conduct a hyperparameter search to ensure proper convergence. + +# 5.1 Performance on the Same Distribution + +We aim to answer a first question: does collecting more samples from the same distribution help SSL pretraining with a fixed budget? We conduct experiments to capture how the pretraining diversity $\mathcal{D}$ influences SSL pretraining within a normalized $\mathcal{C}$ , focusing on the simplest setting where the upstream and downstream datasets belong to the same distribution such that $\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}}$ . This serves as a fundamental ground for our further experiments. + +Setup For pretraining, we use CIFAR-100 [39] and Tiny ImageNet [40] as $\mathbb{D}_{\mathrm{SSL}}$ which contain $50\times 10^{3}$ and $100\times 10^{3}$ images, respectively. We set $\mathcal{C} = 50\times 10^6$ chosen such that it allows all methods to converge during pretraining. In Section 6, we study the effect of varying the budget $\mathcal{C}$ . We pretrain on subsets of $\mathbb{D}_{\mathrm{SSL}}$ with different sizes $N$ (10%, 50%, and 100% of $\mathbb{D}_{\mathrm{SSL}}$ ), enabling us to observe the effects of pretraining diversity $\mathcal{D}$ on training outcomes where $\mathcal{E}$ is adjusted accordingly to match the budget $\mathcal{C}$ . For example, using 100% of CIFAR-100 involves 1000 epochs of training, while 10% and 50% of the dataset lead to 10000 and 2000 epochs, respectively. These subsets of $\mathbb{D}_{\mathrm{SSL}}$ are created by shuffling the dataset and then selecting the first 10%, 50%, or 100% split. All models in this section use a ResNet-18 [33] backbone pretrained from scratch. For evaluation, we use in-distribution linear probing, i.e., $\mathbb{D}_{\mathrm{task}} = \mathbb{D}_{\mathrm{SSL}}$ , where performance is measured by classification accuracy on the test set. + +Results The results of linear probing are presented in Tables 1a and 1b for CIFAR-100 and Tiny ImageNet, respectively. While different SSL methods show varying levels of performance, we observe that, for all methods, an increase in the diversity $\mathcal{D}$ consistently leads to higher linear probing accuracy, suggesting that including more in-distribution samples in $\mathbb{D}_{\mathrm{SSL}}$ helps under a normalized $\mathcal{C}$ . For example, SimCLR achieves an accuracy of 37.95 when $10\%$ of CIFAR-100 is provided, whereas this performance improve by around $12\%$ when only $50\%$ of the dataset is provided and by another $8\%$ after pretraining on $100\%$ of unique samples. This suggests that SSL methods substantially benefit from having a more diverse pretraining set in computationally budgeted in-distribution evaluation, a fundamental verification that allows us to proceed in our analysis. + +# Insight (1) + +When the distributions of the upstream and downstream tasks are the same, i.e., $\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}}$ , in a normalized computation setup, increasing pretraining diversity $\mathcal{D}$ proves to be an effective strategy for enhancing SSL pretraining performance. + +# 5.2 Increasing Data Diversity + +As observed in Section 5.1, if $\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}}$ having a more diverse pretraining set benefits SSL methods, under a normalized computation assumption. However, to increase diversity, sampling from the same distribution $\mathbb{D}_{\mathrm{task}}$ to extend $\mathbb{D}_{\mathrm{SSL}}$ is not always attainable. In fact, pretraining data is often sampled from a distribution different than $\mathbb{D}_{\mathrm{task}}$ . The added samples will then introduce a distribution shift between $\mathbb{D}_{\mathrm{SSL}}$ and $\mathbb{D}_{\mathrm{task}}$ . Then, we raise the following question: is increasing pretraining diversity still beneficial when there is a distribution shift between $\mathbb{D}_{\mathrm{SSL}}$ and $\mathbb{D}_{\mathrm{task}}$ , i.e., $\mathbb{D}_{\mathrm{SSL}} \neq \mathbb{D}_{\mathrm{task}}$ ? We explore strategies for acquiring new data for increasing $\mathcal{D}$ , namely, including existing samples, crawled internet data, and data synthetically generated. To evaluate the effects of the distribution + +![](images/37aedb84907c1bd3433b5fd0bd22e76fa4499276e05d60c6494e4a009d20084e.jpg) +Fig. 2: Data Collection Strategies: We analyze strategies for collecting additional data (A), i.e., collecting more source data, crawling the web or using synthetic images. Using a class prior (top row) simulates In-distribution trainings. We also collect images without class prior (bottom row) to analyze the interactions between diversity and Out-of-distribution classes. + +shift in a controlled scenario, we analyze distributions closer to $\mathbb{D}_{\mathrm{task}}$ by using a class prior (in-distribution) and without a class prior (out-of-distribution). + +Setup We use ImageNet-100 [57] as $\mathbb{D}_{\mathrm{task}}$ , and we construct multiple $\mathbb{D}_{\mathrm{SSL}}$ to evaluate the effects of different data collection strategies and the distribution shift. We first introduce a set $\mathbb{B}$ composed of $65\times 10^{3}$ images from ImageNet-100 (50% of the dataset), which we use as a baseline for $\mathbb{D}_{\mathrm{SSL}}$ with minimum diversity. We denote the 100 classes in ImageNet-100 as $\mathcal{T}_{100}$ , and sample $\mathbb{B}$ uniformly including images from all classes. Next, we compare with pretraining on more diverse datasets as $\mathbb{D}_{\mathrm{SSL}}$ , imposing $\mathbb{D}_{\mathrm{SSL}} = \mathbb{B} \cup \mathbb{A}$ where $\mathbb{A}$ includes $65\times 10^{3}$ images sampled with one of three strategies. To highlight the effects of the distribution shift, we include in $\mathbb{A}$ either images from $In$ -distribution classes, i.e., selecting images from classes overlapping with $\mathcal{T}_{100}$ , or images from $Out$ -of-distribution classes, which do not overlap with $\mathcal{T}_{100}$ . This is to study the effects of the distribution shift, since we do not assume access to downstream classes in real scenarios. For the $Out$ -of-distribution samples, we define $\mathbb{A}$ as (1) random images sampled from the full ImageNet [18] dataset; (2) images crawled from Flickr, Bing, and DuckDuckGo; or (3) images generated with Stable Diffusion V2.1 [51]. We respectively call these sets $\mathbb{A}_{\mathrm{Source}}^{Out}$ , $\mathbb{A}_{\mathrm{Web}}^{Out}$ , and $\mathbb{A}_{\mathrm{Synthetic}}^{Out}$ . We also define their $In$ -distribution counterparts as $\mathbb{A}_{\mathrm{Source}}^{In}$ , $\mathbb{A}_{\mathrm{Web}}^{In}$ , and $\mathbb{A}_{\mathrm{Synthetic}}^{In}$ , respectively. Note how $\mathbb{B} \cup \mathbb{A}_{\mathrm{Source}}^{In}$ is the full ImageNet-100, coherently with Section 5.1. Figure 2 shows each collection strategy, for which we provide implementation details in supplementary material. Although many factors (such as the images appearance) impact the distribution shift, using a class prior with any strategy would result in closer distribution with respect to the same strategy without class priors [15]. We pretrain a ResNet-18 from scratch, with the same settings of Section 5.1, and $\mathcal{C} = 50\times 10^6$ . Note that this results in $\mathcal{D} = 1.25\times 10^{-3}$ for pretraining on $\mathbb{B}$ , and $\mathcal{D} = 2.5\times 10^{-3}$ for pretraining on $\mathbb{B} \cup \mathbb{A}$ , introducing a $\mathcal{D}$ difference of a factor of 2. + +![](images/499aef39458c4cf65338b219cc5dccb382b983eb764f0832660f835fe8136d0d.jpg) +Fig. 3: Effect of Various Data Sources on SSL Pretraining: We use a baseline set $\mathbb{B}$ (black dashed line), comprising $65 \times 10^{3}$ images from ImageNet-100, for pretraining a ResNet-18 with $\mathcal{C} = 50 \times 10^{6}$ . Augmenting $\mathbb{B}$ with In-distribution images enhances performance (above black line), while Out-of-distribution augmentations reduce it (below black line). + +Results We report the linear probing accuracies for pretraining on each $\mathbb{B} \cup \mathbb{A}$ as bars in Figure 3, showing the $\mathcal{C}$ -normalized training on $\mathbb{B}$ as a dashed line. Surprisingly, without class priors $(\mathbf{Out})$ , including $\mathbb{A}_{\mathrm{Source}}^{Out}$ , $\mathbb{A}_{\mathrm{Web}}^{Out}$ , and $\mathbb{A}_{\mathrm{Synthetic}}^{Out}$ underperforms compared to pretraining on $\mathbb{B}$ only. For instance, while SimCLR on $\mathbb{B}$ scores $72.3\%$ accuracy, increasing diversity reduces the accuracy by $1\%$ in all cases. This might appear to conflict with our findings in Section 5.1, however, the inclusion of $\mathbf{Out}$ samples leads to $\mathbb{D}_{\mathrm{SSL}} \neq \mathbb{D}_{\mathrm{task}}$ , since we sample only classes not included in $\mathbb{D}_{\mathrm{task}}$ . We infer that, with normalized $\mathcal{C}$ , increasing $\mathcal{D}$ without distribution priors may negatively impact performance. Conversely, when class priors are available $(\mathbf{In})$ , increasing pretraining diversity $\mathcal{D}$ improves performance compared to $\mathbb{B}$ pretraining. For instance, pretraining on $\mathbb{A}_{\mathrm{Web}}^{In}$ performs comparably to augmenting with additional ImageNet samples $(\mathbb{A}_{\mathrm{Source}}^{In})$ , as in the case of SimCLR where the inclusion of $\mathbb{A}_{\mathrm{Web}}^{In}$ scores only $0.2\%$ lower than $\mathbb{A}_{\mathrm{Source}}^{In}$ . Including $\mathbb{A}_{\mathrm{Synthetic}}^{In}$ data also helps, although more limitedly due to the visual discrepancies between generated and real images. Ultimately, the effectiveness of diversity is linked to the distribution shift. These findings highlight the impact of the distribution shift on computationally-normalized SSL pretraining and help define evaluation practices for this task (see Section 7). + +# Insight (2) + +When the distributions of the upstream and downstream tasks differ, $\mathbb{D}_{\mathrm{SSL}} \neq \mathbb{D}_{\mathrm{task}}$ , and in a normalized computation setup, increasing pretraining diversity $\mathcal{D}$ may harm SSL pretraining performance, reflecting the adverse effects of distribution shift. + +# 5.3 Scaling Pretraining Diversity + +We showed that diversity improves pretraining performance when the training set and downstream task share the same data distribution $(\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}})$ , as discussed in Section 5.1. This may change when a distribution shift is introduced, + +Table 2: Performance of ImageNet and YFCC100M SSL Pretraining on Various Downstream Tasks: We train ResNet-50 and ViT-B/16 under normalized computation $(\mathcal{C} = 98\times 10^{6})$ using SimCLR (left) and MoCoV3 (right) on ImageNet (IN) and YFCC100M (YFCC) with multiple $\mathcal{D}$ . Despite the significantly larger $\mathcal{D}$ when trained on YFCC100M, these models cannot offset the effects of the distribution shift and are outperformed by models pretrained on ImageNet in most downstreams. + +
SimCLR
\( \mathbb{D}_{\text{SSL}} \)\( N \times 10^6 \)\( \mathcal{D} \times 10^{-3} \)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3156.943.082.373.945.459.9
IN0.2562.6161.145.584.076.047.064.4
IN0.6406.5463.746.884.278.848.367.2
IN1.28113.064.546.484.679.548.868.0
YFCC98.17100057.337.276.658.950.162.1
ViT-B/16IN0.1281.3154.237.281.870.344.664.3
IN0.2562.6161.339.583.877.747.269.4
IN0.6406.5465.539.984.680.449.173.1
IN1.28113.066.739.583.781.750.073.6
YFCC98.17100054.525.072.559.749.565.4
+ +$\mathcal{C} = 98\times 10^{6}$ + +
MoCoV3
\( \mathbb{D}_{\text{SSL}} \)\( N \times 10^6 \)\( \mathcal{D} \times 10^{-3} \)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3158.140.681.876.0445.1
IN0.2562.6162.945.385.079.847.6
IN0.6406.5465.448.486.181.949.2
IN1.28113.065.948.886.682.649.5
YFCC98.17100060.442.282.666.350.767.3
ViT-B/16IN0.1281.3157.933.482.878.046.7
IN0.2562.6163.735.085.382.948.4
IN0.6406.5467.239.585.085.849.7
IN1.28113.068.841.986.586.550.3
YFCC98.17100057.225.270.343.850.364.0
+ +$\mathcal{C} = 98\times 10^{6}$ + +as explored in Section 5.2. However, it is still unclear from Section 5.2, whether including a larger number of samples, and thus increasing considerably the pretraining diversity, can compensate for the negative effects of the distribution shift. To address this, the following section presents larger-scale experiments, employing significantly varied $\mathcal{D}$ values, aiming to explore the interplay between pretraining diversity and different distributions using millions of samples. + +Setup For our large-scale pretraining experiments, we set $\mathbb{D}_{\mathrm{SSL}}$ to be two datasets of significantly different sizes: ImageNet [18] and YFCC100M [55], comprising 1.28 million and 98 million images, respectively. Following Section 5.1, we explore multiple $\mathcal{D}$ values for pretraining. We set $\mathcal{C} = 98 \times 10^6$ , which corresponds to one epoch on YFCC100M, iterating once through each of its 98 million images to maximize diversity $(\mathcal{D} = 1)$ . Normalizing $\mathcal{C}$ (see Section 5.1), we pretrain on ImageNet for approximately 77 epochs, cumulatively utilizing 98 million samples. Due to the extensive cost of these experiments, we focus on SimCLR and MoCoV3 only. We also employ larger architectures, namely ResNet-50 [33] and ViT-B/16 [21], to leverage the extensive scale of the pretraining datasets. We also use multiple $\mathbb{D}_{\mathrm{task}}$ including ImageNet [18], Stanford Cars [38], Flowers-102 [44], Oxford-IIIT Pets [48], Places365 [69], and Food-101 [5]. + +Results The results of our large-scale experiments are detailed in Table 2. Consistently with findings in Section 5.1, increasing $\mathcal{D}$ leads to better pretraining efficacy when $\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}}$ . This is evident when ImageNet is used for both pretraining and downstream classification, reinforcing that our observations hold even at a larger scale. Instead, models pretrained on YFCC100M show substantially lower performance compared to those pretrained on ImageNet, although + +having much higher $\mathcal{D}$ . This highlights the inefficiency of collecting data indiscriminately without considering distribution shifts. To stress this, note how the model pretrained on YFCC100M ( $D = 1$ ) often performs similarly to those pretrained with drastically lower $\mathcal{D}$ on ImageNet ( $D = 1.31 \times 10^{-3}$ ). This aligns with our observations in Section 5.2, emphasizing that distribution differences remain a significant factor even when training with large datasets. However, the YFCC100M-pretrained model outperforms the ImageNet-pretrained model in Places365, suggesting a closer distribution relationship between YFCC100M and Places365. We explore this hypothesis further in Section 6, where we analyze distribution distances with existing metrics. Ultimately, our analysis highlights that scaling the data is not an effective solution for compensating the distribution shift, when computational resources are normalized. + +# Insight (3) + +Even an extremely large data diversity cannot mitigate the distribution shift under normalized computation. This emphasizes the importance of further research in how effectively SSL pretraining algorithms generalize. + +# 6 Additional Analysis + +Previously, we introduced a normalized evaluation to study the relation of $\mathcal{D}$ and data distributions. We showed that, although diversity helps when $\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}}$ , vision models are unable to compensate for the distributional differences with normalized costs. We now analyze additional elements supporting our findings. + +Distribution Distances Using FID & VisualDNA In Section 5.3, we showed that pretraining on ImageNet typically outperforms YFCC100M on a variety of downstream tasks, with Places365 being the exception. We speculate that the distribution of ImageNet is more aligned with those of the downstream tasks compared to YFCC100M. To verify this, we evaluate the similarity between the datasets using FID [34] and VisualDNA [50] and report results in Table 3. With both FID or VisualDNA, the distribution of ImageNet is always closer to the downstream tasks, except for Places365 where YFCC100M is closer. This aligns with the lower performance of ImageNet on this dataset only (Table 2), further suggesting that the performance drop is caused by the distribution shift. + +Importance of Normalized Computation We now study the impact of normalizing computation on the performance of SSL methods. We aim to understand if $\mathcal{C}$ is not normalized across methods, this will lead to misleading conclusions, and thus motivating our normalized computation. In Table 4, we compare the performance of ResNet-50 and ViT-B/16 models pretrained using MoCoV3 with $\mathcal{C} = 98 \times 10^6$ (as used in Section 5.3) against a tripled budget of $\mathcal{C} = 294 \times 10^6$ . We show that inconsistencies arise when pretraining data diversity $\mathcal{D}$ and computation $\mathcal{C}$ are not fairly compared. For instance, in scenarios highlighted in red, we notice that a lower pretraining diversity coupled with a higher + +Table 3: Comparing VisualDNA and FID Scores Across Datasets: We assess the relationship between VisualDNA [50] and FID Score [34] for various large-scale $\mathbb{D}_{\mathrm{SSL}}$ and several downstream tasks. For VisualDNA, activations are extracted as suggested [50] with MUGS [70] with ViT-B/16 (MUGS) or DINO v2 [47] with ViT-B/16 (DINO v2), while FID activations are obtained via an Inception network [54]. Consistently, across all metrics, the ranking of dataset distances between $\mathbb{D}_{\mathrm{SSL}}$ and various $\mathbb{D}_{\mathrm{task}}$ aligns with the accuracy ranking in Table 2. Models exhibiting lower VisualDNA/FID scores benefit more from the diversity in pretraining data for SSL. + +
DSSLBackboneVisualDNA↓
INCarsFlow.PetsPlacesFood
INMUGS011.4912.466.097.199.08
YFCCMUGS3.7211.5712.717.936.209.72
INDINO v207.056.954.433.527.26
YFCCDINO v22.377.127.465.752.747.90
DSSLBackboneFID↓
INCarsFlow.PetsPlacesFood
INInception0143.4192.988.964.9114.9
YFCCInception48.1174.3214.8145.838.9154.5
+ +Table 4: Importance of Normalization. We report the accuracy of MoCoV3 trained on ImageNet (IN) with different data diversity and variable $\mathcal{C}$ , for the in-distribution assumption $(\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}})$ . For a given network, cells in red highlight inconsistencies where although the model trained with $\mathcal{C} = 294 \times 10^6$ has seen less samples, it outperforms the best model trained with one third of the computational budget $(\mathcal{C} = 90 \times 10^6)$ , showing the importance of normalization for understanding how models exploit data. + +MoCoV3 + +
Network\( \mathbb{D}_{\text{SSL}} \)\( \mathcal{C} \times 10^6 \)\( N \times 10^6 \)\( \mathcal{D} \times 10^{-3} \)Accuracy† IN
ResNet-50IN980.1281.3158.1
IN980.6406.5465.4
IN981.28113.065.9
IN2940.1280.4357.4
IN2940.6402.1769.8
IN2941.2814.3571.4
ViT-B/16IN980.1281.3157.9
IN980.6406.5467.2
IN981.28113.068.8
IN2940.1280.4356.9
IN2940.6402.1771.9
IN2941.2814.3574.9
+ +computational budget can yield better results. We observe that under pretraining diversity of only $\mathcal{D} = 13\times 10^{-3}$ and a computational budget of $\mathcal{C} = 98\times 10^6$ ResNet-50 only enjoys a $65.9\%$ accuracy compared against with $69.8\%$ with a smaller pretraining diversity of $2.17\times 10^{-3}$ but with a larger computational budget of $294\times 10^{6}$ . This shows that without normalized computation, it could be incorrectly concluded that pretraining diversity does not play a significant role. + +Model Saturation We aim to study the pretraining trend on YFCC100M, and if adding more samples could compensate the distribution shift with ImageNet. Hence, we extend the YFCC100M experiments from Section 5.3, examining various subsets—specifically $0.1\%$ , $1\%$ , $10\%$ , and $100\%$ of the dataset. Again, we normalize the computational budget to $\mathcal{C} = 98 \times 10^6$ , equivalent to one epoch on YFCC100M. The results of linear probing models pretrained using SimCLR and MoCoV3 with ResNet-50 and ViT-B/16 on ImageNet are shown in Figure 4. Interestingly, we obtain a performance plateau. This observation points to a saturation point in performance, showing that simply increasing the dataset further would not bridge the gap between the $\mathbb{D}_{\mathrm{SSL}}$ and $\mathbb{D}_{\mathrm{task}}$ . Hence, arbitrarily increasing pretraining diversity $\mathcal{D}$ is not sufficient to bridge the distribution shift problem. Here $\mathbb{D}_{\mathrm{SSL}} \neq \mathbb{D}_{\mathrm{task}}$ , hence still aligning with our findings in Section 5.1. + +![](images/a0b7c13e724ad3bd7d335cdc10821de1486178e39607abbcd2524a10e6098ece.jpg) +Fig. 4: Data Diversity Impact on YFCC100M Pretraining Performance: Pretraining $(\mathcal{C} = 98\times 10^{6})$ of networks with $\mathbb{D}_{\mathrm{SSL}} = \mathrm{YFCC100M}$ and $\mathbb{D}_{\mathrm{task}} = \mathrm{ImageNet}$ for several dataset subsets. In the presence of a distribution shift, performance tends to saturate and does not benefit from additional data diversity. + +Table 5: Evaluating Network Accuracy With Varied Label Quantity: We evaluate the accuracy of networks trained on ImageNet (IN) and YFCC100M (YFCC) with different labeling percentages of $\mathbb{D}_{\mathrm{task}} = \mathrm{ImageNet}$ . The increased diversity still does not compensate for the distribution shift. However, for in-distribution data, one can get away with fewer labels with more diverse pretraining data. + +
SimCLR
Network\( \mathbb{D}_{\text{SSL}} \)N×106D×10-3Accuracy†
5%10%50%
ResNet-50IN1.281.3150.354.261.7
YFCC98.2100039.444.354.0
ViT-B/16IN1.281.3155.259.465.6
YFCC98.2100040.745.753.3
C=98×10^6
+ +
NetworkMoCoV3
\( \mathbb{D}_{\text{SSL}} \)N×106D×10-3Accuracy↑
5%10%50%
ResNet-50IN1.281.3152.156.263.8
YFCC98.2100042.848.257.7
ViT-B/16IN1.281.3159.363.068.3
YFCC98.2100043.949.056.0
C=98×10^6
+ +Label Quantity A question that arises in our setting is whether having a higher pretraining diversity leads to requiring less labeled samples during linear probing. In this section, we focus on understanding the impact of increased $\mathcal{D}$ on the number of labeled samples required for effective linear probing in downstream tasks. We use the same trained models from Section 5.3, i.e., SimCLR and MoCoV3, with ResNet-50 and ViT-B/16 architectures. In this setup, we set $\mathbb{D}_{\mathrm{task}} = \mathrm{ImageNet}$ , with two upstream dataset scenarios: ImageNet (where $\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}}$ ) and YFCC100M (where $\mathbb{D}_{\mathrm{SSL}} \neq \mathbb{D}_{\mathrm{task}}$ ). Our experiments are summarized in Table 5. We note that with ViT-B/16, if ImageNet is used for pretraining, linear probing with just $5\%$ labeled samples can surpass the performance achieved using $100\%$ labeled data when YFCC100M serves as $\mathbb{D}_{\mathrm{SSL}}$ . This also implies that when $\mathbb{D}_{\mathrm{SSL}} \neq \mathbb{D}_{\mathrm{task}}$ , a higher quantity of labeled data is necessary to attain competitive performance, and increasing $\mathcal{D}$ does not bring considerable benefits on label quantity under distribution shift. This implies that our findings in Section 5.3 hold regardless of the linear probing labeled set. We note that in scenarios where $\mathbb{D}_{\mathrm{SSL}} = \mathbb{D}_{\mathrm{task}}$ , using only $50\%$ of the labeled data can achieve similar performance as utilizing the full $100\%$ of labeled samples, implying that increasing $\mathcal{D}$ leads to reduced label requirement efforts for downstream tasks. + +# 7 Discussion + +Main Conclusions Our set of experiments leads to Insight (3) in Section 5.3, revealing that with normalized computational costs, SSL pretrainings with large diversity cannot compensate for the distribution shift. This is surprising, since the variety of information that SSL algorithms could benefit from during training is much higher in large generic datasets than in small ones. Since our evaluation is cost-normalized, (3) also shows that SSL strategies are not efficiently exploiting the pretraining diversity on large datasets for representation extraction. This inefficiency implies a wide margin for improvement of generalization of SSL models, making better use of the computational power used for training. The role of existing models must also be discussed in this context. Following Insights (1) and (2), respectively in Sections 5.1 and 5.2, we have studied how in-distribution and out-of-distribution data impact performance in a controlled scenario. We argue that this should be taken into account in the evaluation of the performance of SSL models. Indeed, training at scale may enlarge the in-distribution space, including classes of the downstream tasks in the training set. In recent literature, this is a design choice to maximize performance on popular benchmarks [47]. While this allows for achieving impressive results in many tasks, we stress that this does not permit a fair evaluation. Now, we summarize practical takeaways. + +Training Takeaways Coherently with our findings in Section 5.1 and similarly to prior art [15,28,37], we find that aligned distributions benefit performance, in particular increasing $\mathbb{D}_{\mathrm{SSL}}$ diversity helps, as long as the distribution of the new samples match those of the downstream task data. Differently from the state-of-the-art, we demonstrate that this holds also in a computationally-normalized setting, implying that collecting large scale in-distribution data matching the downstream task is be an effective approach for improving SSL. Hence, for practical applications distribution priors should be used, if available. On the contrary, for a fair evaluation of models, this should not be the case, as specified below. + +Evaluation Takeaways Our analysis reveals that to permit a fair evaluation of SSL methods, computationally normalized tests are necessary to avoid inconsistencies, as shown in Section 6. Moreover, it is crucial to identify out-of-distribution downstream tasks for a correct evaluation of generalization performance. By evaluating only on $\mathbb{D}_{\mathrm{task}}$ with a low distribution shift, there is a risk of reporting inflated metrics, not representative of a real gain in generalization. This is important, since new SSL approaches may be reporting higher downstream performance when pretrained on a different dataset. We relate this to Sections 5 and 6, where we show that increasing the computation and the in-distribution data, respectively, can improve performance. Ultimately, wrong practices may result in incorrectly concluding that an SSL algorithm is performing better. + +Differences With Language Models In Section 5.3 we showed that even very diverse datasets (YFCC100M), fall short in satisfactory generalization performance. In addition to further exploration into generalization for SSL pretraining, this open doors to investigating why language models enjoy enhanced generalization when trained on highly diverse data compared to vision models [61]. + +# Acknowledgements + +This work was supported by SDAIA-KAUST Center of Excellence in Data Science, Artificial Intelligence (SDAIA-KAUST AI). Fabio Pizzati is financed by KAUST (Grant DFR07910). Philip H.S. Torr thanks the Royal Academy of Engineering for their support. This work is supported by a UKRI grant Turing AI Fellowship (EP/W002981/1). + +# References + +1. Bachman, P., Hjelm, R.D., Buchwalter, W.: Learning representations by maximizing mutual information across views. In: NeurIPS (2019) +2. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023) +3. Bao, H., Dong, L., Piao, S., Wei, F.: BEit: BERT pre-training of image transformers. In: ICML (2022) +4. Bardes, A., Ponce, J., LeCun, Y.: VICReg: Variance-invariance-covariance regularization for self-supervised learning. In: ICLR (2022) +5. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-Mining discriminative components with random forests. In: ECCV (2014) +6. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: ECCV (2018) +7. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: ICCV (2019) +8. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020) +9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) +0. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020) +1. Chen, T., Kornblith, S., Swersky, K., Norouzi, M., Hinton, G.: Big self-supervised models are strong semi-supervised learners. In: NeurIPS (2020) +2. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +3. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021) +4. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV (2021) +5. Cole, E., Yang, X., Wilber, K., Aodha, O.M., Belongie, S.: When does contrastive visual representation learning work? In: CVPR (2022) +6. Costa, V.G.T.d., Fini, E., Nabi, M., Sebe, N., Ricci, E.: solo-learn: A library of self-supervised methods for visual representation learning. In: JMLR (2022) +7. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: ICML (2023) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) + +19. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: ICCV (2015) +20. Doersch, C., Zisserman, A.: Multi-task self-supervised visual learning. In: ICCV (2017) +21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +22. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv (2021) +23. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: CVPR (2021) +24. Fini, E., Astolfi, P., Romero-Soriano, A., Verbeek, J., Drozdzal, M.: Improved baselines for vision-language pre-training. TMLR (2023) +25. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018) +26. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021) +27. Goyal, P., Duval, Q., Seessel, I., Caron, M., Misra, I., Sagun, L., Joulin, A., Bojanowski, P.: Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360 (2022) +28. Goyal, P., Mahajan, D., Gupta, A., Misra, I.: Scaling and benchmarking self-supervised visual representation learning. In: ICCV (2019) +29. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent: A new approach to self-supervised learning. In: NeurIPS. vol. 33 (2020) +30. Hammoud, H.A.A.K., Itani, H., Pizzati, F., Torr, P., Bibi, A., Ghanem, B.: Synthclip: Are we ready for a fully synthetic clip training? arXiv (2024) +31. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) +32. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) +33. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017) +35. Hjelm, R.D., Fedorov, A., Lavoie-Marchildon, S., Grewal, K., Bachman, P., Trischler, A., Bengio, Y.: Learning deep representations by mutual information estimation and maximization. In: ICLR (2018) +36. Henaff, O.J., Srinivas, A., De Fauw, J., Razavi, A., Doersch, C., Eslami, S.M.A., Oord, A.v.d.: Data-efficient image recognition with contrastive predictive coding. In: ICML (2019) +37. Kotar, K., Ilharco, G., Schmidt, L., Ehsani, K., Mottaghi, R.: Contrasting contrastive self-supervised representation learning pipelines. In: ICCV (2021) +38. Krause, J., Deng, J., Stark, M., Fei-Fei, L.: Collecting a large-scale dataset of fine-grained cars. In: FGVC (2013) +39. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) + +40. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N (2015) +41. Li, A.C., Brown, E.L., Efros, A.A., Pathak, D.: Internet explorer: Targeted representation learning on the open web. In: ICML (2023) +42. Li, J., Zhou, P., Xiong, C., Hoi, S.C.H.: Prototypical contrastive learning of unsupervised representations. In: ICLR (2021) +43. Misra, I., Maaten, L.v.d.: Self-supervised learning of pretext-invariant representations. In: CVPR (2020) +44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP (2008) +45. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016) +46. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +47. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINoV2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR (2012) +49. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: CVPR (2020) +50. Ramtoula, B., Gadd, M., Newman, P., De Martini, D.: Visual DNA: Representing and comparing images using distributions of neuron activations. In: CVPR (2023) +51. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +52. Shi, Y., Daunhawer, I., Vogt, J.E., Torr, P., Sanyal, A.: How robust is unsupervised representation learning to distribution shift? In: ICLR (2022) +53. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: ICCV (2017) +54. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: CVPR (2016) +55. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: YFCC100M: The new data in multimedia research. In: Commun. ACM (2016) +56. Tian, Y., Henaff, O.J., Van Den Oord, A.: Divide and contrast: Self-supervised learning from uncurated data. In: ICCV (2021) +57. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: ECCV (2020) +58. Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: CVPR (2011) +59. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Gool, L.V.: Revisiting contrastive methods for unsupervised learning of visual representations. In: NeurIPS (2021) +60. Van Horn, G., Cole, E., Beery, S., Wilber, K., Belongie, S., Mac Aodha, O.: Benchmarking representation learning for natural world image collections. In: CVPR (2021) +61. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. TMLR (2022) +62. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via non-parametric instance discrimination. In: CVPR (2018) +63. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMIM: A simple framework for masked image modeling. In: CVPR (2022) + +64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Wei, Y., Dai, Q., Hu, H.: On data scaling in masked image modeling. In: CVPR (2023) +65. Zbontar, J., Jing, L., Misra, I., LeCun, Y., Deny, S.: Barlow Twins: Self-supervised learning via redundancy reduction. In: ICML (2021) +66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR (2022) +67. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016) +68. Zhao, N., Wu, Z., Lau, R.W.H., Lin, S.: What makes instance discrimination good for transfer learning? In: ICML (2021) +69. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. T-PAMI (2017) +70. Zhou, P., Zhou, Y., Si, C., Yu, W., Ng, T.K., Yan, S.: Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415 (2022) +71. Zhuang, C., Zhai, A.L., Yamins, D.: Local aggregation for unsupervised learning of visual embeddings. In: ICCV (2019) \ No newline at end of file diff --git a/2024/On Pretraining Data Diversity for Self-Supervised Learning/images.zip b/2024/On Pretraining Data Diversity for Self-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c411a0c61df5d70a834fc4d33010325e95c77357 --- /dev/null +++ b/2024/On Pretraining Data Diversity for Self-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b3afd26014a01055fe077a74e86d0b6a981f8fe65f98127a55f50db0fce82a9 +size 402021 diff --git a/2024/On Pretraining Data Diversity for Self-Supervised Learning/layout.json b/2024/On Pretraining Data Diversity for Self-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8c229c54cf8f31bf358e53a0b17e518f7da55f1f --- /dev/null +++ b/2024/On Pretraining Data Diversity for Self-Supervised Learning/layout.json @@ -0,0 +1,11627 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 200, + 111, + 413, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 111, + 413, + 147 + ], + "spans": [ + { + "bbox": [ + 200, + 111, + 413, + 147 + ], + "type": "text", + "content": "On Pretraining Data Diversity for Self-Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "text", + "content": "Hasan Abed Al Kader Hammoud" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{1,2\\star\\diamond}" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "text", + "content": ", Tuhin Das" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "text", + "content": ", Fabio Pizzati" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "text", + "content": " Philip H.S. Torr" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "text", + "content": ", and Adel Bibi" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "text", + "content": ", Bernard Ghanem" + }, + { + "bbox": [ + 157, + 167, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 283, + 201, + 328, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 201, + 328, + 212 + ], + "spans": [ + { + "bbox": [ + 283, + 201, + 328, + 212 + ], + "type": "text", + "content": "1 KAUST" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 260, + 213, + 353, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 213, + 353, + 224 + ], + "spans": [ + { + "bbox": [ + 260, + 213, + 353, + 224 + ], + "type": "text", + "content": "2 University of Oxford" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 251, + 452, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 452, + 383 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 452, + 383 + ], + "type": "text", + "content": "Abstract. We explore the impact of training with more diverse datasets, characterized by the number of unique samples, on the performance of self-supervised learning (SSL) under a fixed computational budget. Our findings demonstrate that increasing pretraining data diversity enhances SSL performance, albeit only when the distribution distance to the downstream data is minimal. Notably, even with an exceptionally large pretraining data diversity achieved through methods like web crawling or diffusion-generated data, among other ways, the distribution shift remains a challenge. Our experiments are comprehensive with seven SSL methods using large-scale datasets such as ImageNet and YFCC100M amounting to over 200 GPU days. The code and trained models will be available at https://github.com/hammoudhasan/DiversitySSL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 394, + 451, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 394, + 451, + 406 + ], + "spans": [ + { + "bbox": [ + 160, + 394, + 451, + 406 + ], + "type": "text", + "content": "Keywords: self-supervised learning " + }, + { + "bbox": [ + 160, + 394, + 451, + 406 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 394, + 451, + 406 + ], + "type": "text", + "content": " distribution shift " + }, + { + "bbox": [ + 160, + 394, + 451, + 406 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 394, + 451, + 406 + ], + "type": "text", + "content": " data diversity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 442, + 229, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 229, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 229, + 455 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 467, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 647 + ], + "type": "text", + "content": "Self-supervised learning (SSL) has recently emerged as a new paradigm to pretrain large vision models at scale [26, 27, 47]. Leveraging the ability to learn from unlabelled data, pretraining on millions—or even billions [27, 53]—of images turned from an unachievable goal to a common practice. This exposure to extremely diverse datasets, i.e., composed of a remarkable number of unique samples, granted impressive performance and unprecedented generalization capabilities to a growing number of vision models [47]. Large-scale datasets, in conjunction with substantial computational resources, have been the key driving forces for the success of SSL-based approaches. For instance, SEER [26], pretrained for approximately 11 GPU years on a billion images, exemplifies the massive computation and data resources employed for these models. It has become the implicit norm that increasing computation and data is beneficial, without any detailed analysis of how they separately impact SSL effectiveness. In particular, it is not clear to what extent large datasets are responsible for the impressive generalization capabilities of SSL models. Indeed, consider the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 481, + 665 + ], + "type": "text", + "content": "* Equal contribution † Equal supervision Ⓞ Work done during a research visit at Oxford" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 115, + 449, + 316 + ], + "blocks": [ + { + "bbox": [ + 162, + 115, + 449, + 316 + ], + "lines": [ + { + "bbox": [ + 162, + 115, + 449, + 316 + ], + "spans": [ + { + "bbox": [ + 162, + 115, + 449, + 316 + ], + "type": "image", + "image_path": "9efeea3bf096c200ff57ab6b0029609544d4d0d2f34fea7b857437a9d115c3d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "text", + "content": "Fig. 1: Impact of Pretraining Diversity: Self-supervised learning (SSL) can be used to pretrain vision models on small datasets closely aligned to the downstream task, e.g., pets classification, hence with a small distribution shift (top, wild animals pretraining). Conversely, we could pretrain on an extensively varied dataset, with wide distribution differences (outdoor scenes, bottom). We study the role of pretraining diversity in SSL under a fixed budget, and highlight its effects in relationship to the distribution shift." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 414, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 510 + ], + "type": "text", + "content": "example in Figure 1. Assuming a fixed monetary budget, in the form of computational expenses, for pretraining a vision model under the SSL paradigm, does iterating over a large set of images work best given a downstream task, or it is better to train repeatedly on a smaller set of samples visually close to the ones of the downstream? In this context, we face the ever-lasting problem of distribution shift on vision models [58]. Without a proper understanding of the impact of data, there could be a wrong allocation of efforts in increasing data or computation, leading to inefficiencies in the deployment process of SSL models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "In this paper, we study the role of the diversity of the SSL pretraining data, under a fixed computational budget, when the pretraining data matches the distribution of the downstream data, as well as when they differ. Our experiments span various SSL methods [4,8-10, 14, 29, 65] and multiple datasets [5,18,38-40, 44,48,55,57,69], and are tested under several computational budgets amounting to a total of 200 GPU days. We summarize our contributions below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 582, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 135, + 582, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 582, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 135, + 582, + 482, + 640 + ], + "type": "text", + "content": "1. We show that SSL pretraining strategies are currently data-inefficient in compensating for distribution shifts. Under normalized computational costs, we verify that pretraining on large datasets with high diversity cannot outperform the pretraining on datasets with limited diversity, but with a lower distribution shift with respect to the downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 135, + 641, + 482, + 666 + ], + "type": "text", + "content": "2. We conclude that there is a wide margin for improvement in the performance of current SSL methods on out-of-distribution classes, and we propose in" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 479, + 151 + ], + "type": "text", + "content": "sights for a fair evaluation. This shows the need for future research on new SSL techniques leveraging data diversity better, to improve generalization capabilities beyond the training data distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 152, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 152, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 135, + 152, + 479, + 187 + ], + "type": "text", + "content": "3. We propose a novel strategy for a computationally-normalized evaluation of SSL and a carefully designed set of experiments focusing on pretraining data diversity, enabling us to draw our novel conclusions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 224 + ], + "type": "text", + "content": "Ultimately, our work provides a comprehensive analysis, from experiments worth 200 GPU days on NVIDIA A100, on the interaction between data diversity and computational resources, and their impact on the performance of SSL models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 243, + 237, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 237, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 237, + 256 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 486 + ], + "type": "text", + "content": "Self-Supervised Learning Early work in self-supervised learning used simple pretext tasks, such as relative patch prediction [19, 20], image colorization [67], image rotation prediction [25], or solving jigsaw puzzles [45] to train feature extractors in absence of annotations. Representations learned with those methods are limitedly effective, hence more recent literature has moved towards more sophisticated approaches, such as using image augmentations to generate correlated views of a training sample, and learning to extract augmentation-invariant representations for these correlated pairs. Among these multi-view methods, many exploit contrastive losses [1,8,10-12,32,35,36,42,43,46], enforcing similarity between views of the same image (positives) and dissimilarity from other images (negatives). Due to the need of many negative samples, contrastive methods often require large batch sizes to work effectively [10,14]. Cluster-based methods such as SwAV [8], DINO [9] and DeepCluster v2 [6] learn generalizable representations by grouping samples into cluster prototypes. Others exploit prediction of features with siamese networks [13], learn features in a teacher-student fashion [29], or use redundancy reduction techniques [4,65]. Finally, masked image modeling [3,31,63] emerged as a scalable alternative for Vision Transformers [21], learning representations by predicting masked image patches as SSL task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 653 + ], + "type": "text", + "content": "Pretraining at Scale SSL pretraining is most effective at scale, benefiting from large datasets and great computational resources. Initial attempts at large-scale pretraining were made by combining contrastive learning and clustering [7, 56]. SEER [26] was trained on 1 billion internet images with SwAV [8] on a RegNet [49] backbone. The importance of model scaling to leverage large datasets is also shown in [27, 28, 64], as well as the need for longer training times [64]. Additional strategies are necessary to achieve best performance [17, 24, 66]. These works focus on reaching the best representations, without considerations about training costs, thus encouraging unfair comparisons across trainings. Preliminary papers [15, 22] found that increasing data may lead to a modest boost in downstream performance when the downstream distribution changes. While they give insights on the pretraining diversity, its relation with the distribution shift under normalized computation is still not investigated." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": "Distribution Shift in SSL Some studies have investigated how the pretraining data affects downstream performance on other domains. In general, the distribution shift increases when factors such as visual content and data acquisition procedure differ in two data ensembles [58]. For various pretraining datasets, preliminary works [15, 28, 37, 41] observed that the best performing representations were learned on datasets that were similar to the downstream test task. Additionally, combining datasets before pretraining or combining self-supervised features learned from various datasets did not lead to significant improvements in [15]. In [52, 68], they showcase higher generalization capabilities of SSL models compared to their supervised counterparts, for several downstream tasks in the presence of a distribution shift. In [59], they pretrained on several datasets, observing different generalization depending on the appearance of the pretraining dataset. Furthermore, initial considerations on the impact of external data on downstream tasks under distribution shift have been proposed in [60]. Some compensate the distribution shift with the scale of training datasets [30]. However, there is still a lack of a fair, computation-normalized evaluation that allows to study the effects of the distribution shift in a controlled environment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 338, + 234, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 234, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 234, + 350 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": "Pretraining We first outline the general pretraining procedure common to state-of-the-art self-supervised learning methods. Specific differences among these methods are detailed in the supplementary material. The overall pretraining pipeline, common across many SSL approaches [4, 8-10, 29, 65], goes as follows: (1) Data Sampling: from a large-scale, unlabeled upstream pretraining dataset, denoted as " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": ", an image " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": " is randomly sampled; (2) View Generation: two correlated views, " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_A" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_B" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": ", are generated from " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": " using two image augmentation functions sampled at random. These random augmentations include random resized cropping, horizontal flipping, blurring, and color adjustments [2], among others; (3) Feature Extraction and Projection: the correlated views undergo feature extraction through a network, " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "f_{\\theta_f}" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\theta_f" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": ", such as a ResNet [33] or a ViT [21], leading to representations " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_A = f_{\\theta_f}(\\tilde{\\mathbf{x}}_A)" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_B = f_{\\theta_f}(\\tilde{\\mathbf{x}}_B)" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": ". A linear projection head, " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "g_{\\theta_g}" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\theta_g" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": ", then maps these representations to a latent space, resulting in " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_A = g_{\\theta_g}(\\mathbf{h}_A)" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_B = g_{\\theta_g}(\\mathbf{h}_B)" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": "; (4) Joint Optimization: The feature extractor " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "f_{\\theta_f}" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": " and the projection head " + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "inline_equation", + "content": "g_{\\theta_g}" + }, + { + "bbox": [ + 130, + 363, + 482, + 557 + ], + "type": "text", + "content": " are jointly optimized according to the following objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 566, + 481, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 566, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 219, + 566, + 481, + 587 + ], + "type": "interline_equation", + "content": "\\theta_ {f} ^ {*}, \\theta_ {g} ^ {*} = \\underset {\\theta_ {f}, \\theta_ {g}} {\\arg \\min } \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {D} _ {\\mathrm {S S L}}} \\mathcal {L} _ {\\mathrm {S S L}} \\left(\\mathbf {z} _ {A}, \\mathbf {z} _ {B}\\right), \\tag {1}", + "image_path": "0b6937c57116638e943ee2fe84c208c5f2eca7bc169d013cbb3ec3295e332ed7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "text", + "content": " is a loss function specific to the SSL pretraining method. After pretraining, the feature extractor " + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "inline_equation", + "content": "f_{\\theta_f^*}" + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "text", + "content": " can be deployed for various downstream tasks such as image classification, object detection, or segmentation. This is typically achieved by training a task-specific head. Alternatively, the feature extractor " + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "inline_equation", + "content": "f_{\\theta_f^*}" + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "text", + "content": " can be fine-tuned or used with a " + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 592, + 482, + 654 + ], + "type": "text", + "content": "-nearest neighbors (kNN) classifier." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": "Linear Probing There are several ways to evaluate the performance of a self-supervised learning method such as linear probing [10, 14, 29], kNN [8, 9, 62, 71], and few-shot evaluation [23, 26]. Consistently with the general protocol established in the literature [10, 15, 32, 67], we use linear probing to measure the quality of the features extracted for classification tasks. The procedure is: (1) a labeled downstream dataset, " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ", consisting of image-class pairs " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, \\mathbf{y}) \\sim \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ", is selected for evaluation. (2) For each image " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ", its representation is extracted using the pretrained feature extractor " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "f_{\\theta_f^*}" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ", after which the linear classification head " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "t_{\\theta_t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ", is then applied to obtain " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "t_{\\theta_t}(f_{\\theta_f^*}(\\mathbf{x}))" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ". (3) The linear head " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "t_{\\theta_t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": " is optimized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 245, + 481, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 245, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 202, + 245, + 481, + 269 + ], + "type": "interline_equation", + "content": "\\theta_ {t} ^ {*} = \\underset {\\theta_ {t}} {\\arg \\min } \\mathbb {E} _ {(\\mathbf {x}, \\mathbf {y}) \\sim \\mathbb {D} _ {\\mathrm {t a s k}}} \\left[ \\mathcal {L} _ {\\mathrm {t a s k}} (t _ {\\theta_ {t}} (f _ {\\theta_ {f} ^ {*}} (\\mathbf {x})), \\mathbf {y}) \\right], \\tag {2}", + "image_path": "407d1a030402ecc87371c8afc2011c68ae57ea754805217061a1d0e4c71e4401.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "text", + "content": "Note that only the parameters of the head " + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "text", + "content": " are optimized, while the parameters " + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\theta_f^*" + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "text", + "content": " of the feature extractor are frozen. The quality of features extracted by " + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "inline_equation", + "content": "f_{\\theta_f^*}" + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "text", + "content": " is directly inferred from the classification accuracy achieved on the test set of " + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 131, + 276, + 482, + 324 + ], + "type": "text", + "content": ", which serves as the primary indicator of the quality of the extracted features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 342, + 290, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 290, + 355 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 290, + 355 + ], + "type": "text", + "content": "4 Normalized Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 426 + ], + "type": "text", + "content": "We stress how for a correct understanding of the impact of data diversity we need to analyze its effects isolating them from the impact of increased computation. To enable this, we introduce (1) a computational budget used to normalize computation across experiments, and (2) a quantification of the data diversity seen by models during pretraining." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": "Computational Budget Current progress in SSL pretraining simultaneously scales computational budget and dataset size to achieve the best performance [26]. This makes it difficult to assess the reasons behind the success of many SSL algorithms. Do these methods benefit mainly from the large amounts of computation, i.e., running SSL pretraining for large numbers of epochs, or do they benefit from data diversity in larger datasets containing a vast variety of visual features? To perform a disentangled evaluation of these two factors, we first introduce " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " as a measure of computational budget, which quantifies the total number of images an SSL method is allowed to process during pretraining. This is calculated as " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = N\\cdot \\mathcal{E}" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " is the number of unique samples in the pretraining dataset " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": ", hence the data diversity of the dataset, and " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " is the number of pretraining epochs. Constraining multiple models pretrained with SSL to a fixed computational budget " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " allows for meaningful comparison, as it is guaranteed that all SSL methods will have processed the same number of pretraining images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Quantifying Pretraining Diversity Even under normalized computation " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", various SSL approaches may be exposed to different data diversity as they are trained with different datasets. For instance, a model trained for " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 1000" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " epochs on a dataset of size " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "N = 1000" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " will see less diversity than a model pretrained for" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 212, + 304, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "text", + "content": "Table 1: Impact of Data Diversity on CIFAR-100 and Tiny ImageNet SSL Pretraining Performance: We study the effects of diversity on CIFAR-100 a and Tiny ImageNet b across seven different methods and three data diversity settings for a ResNet-18 pretraining, where for all, " + }, + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}} = \\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 202 + ], + "type": "text", + "content": ". This comparison includes analyzing classification accuracy through linear probing on the training set and evaluation on the test set of the respective datasets. Although performance fluctuates among different methods, a consistent trend is observed: higher data diversity typically leads to the generation of higher quality representations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 212, + 304, + 264 + ], + "lines": [ + { + "bbox": [ + 133, + 212, + 304, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 212, + 304, + 264 + ], + "type": "table", + "html": "
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
50.138.044.341.114.144.112.515.7
250.550.154.351.045.450.451.635.9
501.058.658.458.356.455.855.640.9
", + "image_path": "f9334e48c8e74747622928a0d93e597863a98139da79078362a45d2722b6b237.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 196, + 267, + 240, + 276 + ], + "lines": [ + { + "bbox": [ + 196, + 267, + 240, + 276 + ], + "spans": [ + { + "bbox": [ + 196, + 267, + 240, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 50\\times 10^{6}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 193, + 278, + 247, + 288 + ], + "lines": [ + { + "bbox": [ + 193, + 278, + 247, + 288 + ], + "spans": [ + { + "bbox": [ + 193, + 278, + 247, + 288 + ], + "type": "text", + "content": "(a) CIFAR-100" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 308, + 213, + 479, + 264 + ], + "blocks": [ + { + "bbox": [ + 308, + 213, + 479, + 264 + ], + "lines": [ + { + "bbox": [ + 308, + 213, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 308, + 213, + 479, + 264 + ], + "type": "table", + "html": "
N×103D×10-3Accuracy↑
SimCLR B.T.BYOLSwAVVICRMoCo3DINO
100.236.941.035.634.237.636.634.5
501.048.852.048.143.648.846.544.4
1002.049.855.650.347.554.148.647.3
", + "image_path": "44088994b57550bc6a6946212e8c6c9eb3533885ba6a90ecb43f469d970516eb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 371, + 267, + 414, + 276 + ], + "lines": [ + { + "bbox": [ + 371, + 267, + 414, + 276 + ], + "spans": [ + { + "bbox": [ + 371, + 267, + 414, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 50\\times 10^{6}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 360, + 278, + 428, + 288 + ], + "lines": [ + { + "bbox": [ + 360, + 278, + 428, + 288 + ], + "spans": [ + { + "bbox": [ + 360, + 278, + 428, + 288 + ], + "type": "text", + "content": "(b) Tiny ImageNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 1" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": " epoch on a dataset of size " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "N = 10^6" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": ", despite that they are both pretrained under normalized computation, processing the same number of images. Hence, to capture this notion of exposure to diversity while pretraining under normalized computation " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": ", we define a pretraining diversity " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": ", which captures the number of unique samples encountered during training given a fixed " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = N / \\mathcal{C} = 1 / \\varepsilon" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": ". A model pretrained with larger " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": " indicates that the model is presented with a larger number of unique images during training with fixed " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": ", and hence is exposed to more pretraining data diversity. In the next section, we explore the effects of variations in " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": " on SSL performance under a distribution shift, while keeping " + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": " constant." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 459, + 418, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 418, + 473 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 418, + 473 + ], + "type": "text", + "content": "5 Fixed Budget SSL: In & Out-of-Distribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 484, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 557 + ], + "type": "text", + "content": "Training Configuration We evaluate seven SSL methods: SimCLR [10], MoCoV3 [14], VICReg [4], DINO [9], BYOL [29], Barlow Twins [65], and SwAV [8]. We use different datasets for both pretraining " + }, + { + "bbox": [ + 130, + 484, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 484, + 482, + 557 + ], + "type": "text", + "content": " and linear probing " + }, + { + "bbox": [ + 130, + 484, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 484, + 482, + 557 + ], + "type": "text", + "content": " for different sections. We use solo-learn [16] as a codebase. For each method, we use the default parameters provided when available, otherwise, we conduct a hyperparameter search to ensure proper convergence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 574, + 356, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 356, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 356, + 586 + ], + "type": "text", + "content": "5.1 Performance on the Same Distribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We aim to answer a first question: does collecting more samples from the same distribution help SSL pretraining with a fixed budget? We conduct experiments to capture how the pretraining diversity " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " influences SSL pretraining within a normalized " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", focusing on the simplest setting where the upstream and downstream datasets belong to the same distribution such that " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". This serves as a fundamental ground for our further experiments." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "Setup For pretraining, we use CIFAR-100 [39] and Tiny ImageNet [40] as " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " which contain " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "50\\times 10^{3}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "100\\times 10^{3}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " images, respectively. We set " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 50\\times 10^6" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " chosen such that it allows all methods to converge during pretraining. In Section 6, we study the effect of varying the budget " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ". We pretrain on subsets of " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " with different sizes " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " (10%, 50%, and 100% of " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "), enabling us to observe the effects of pretraining diversity " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " on training outcomes where " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " is adjusted accordingly to match the budget " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ". For example, using 100% of CIFAR-100 involves 1000 epochs of training, while 10% and 50% of the dataset lead to 10000 and 2000 epochs, respectively. These subsets of " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " are created by shuffling the dataset and then selecting the first 10%, 50%, or 100% split. All models in this section use a ResNet-18 [33] backbone pretrained from scratch. For evaluation, we use in-distribution linear probing, i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}} = \\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ", where performance is measured by classification accuracy on the test set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": "Results The results of linear probing are presented in Tables 1a and 1b for CIFAR-100 and Tiny ImageNet, respectively. While different SSL methods show varying levels of performance, we observe that, for all methods, an increase in the diversity " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " consistently leads to higher linear probing accuracy, suggesting that including more in-distribution samples in " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " helps under a normalized " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": ". For example, SimCLR achieves an accuracy of 37.95 when " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " of CIFAR-100 is provided, whereas this performance improve by around " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " when only " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " of the dataset is provided and by another " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " after pretraining on " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " of unique samples. This suggests that SSL methods substantially benefit from having a more diverse pretraining set in computationally budgeted in-distribution evaluation, a fundamental verification that allows us to proceed in our analysis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 425, + 204, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 425, + 204, + 438 + ], + "spans": [ + { + "bbox": [ + 147, + 425, + 204, + 438 + ], + "type": "text", + "content": "Insight (1)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 447, + 466, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 447, + 466, + 495 + ], + "spans": [ + { + "bbox": [ + 146, + 447, + 466, + 495 + ], + "type": "text", + "content": "When the distributions of the upstream and downstream tasks are the same, i.e., " + }, + { + "bbox": [ + 146, + 447, + 466, + 495 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 146, + 447, + 466, + 495 + ], + "type": "text", + "content": ", in a normalized computation setup, increasing pretraining diversity " + }, + { + "bbox": [ + 146, + 447, + 466, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 146, + 447, + 466, + 495 + ], + "type": "text", + "content": " proves to be an effective strategy for enhancing SSL pretraining performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 524, + 290, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 290, + 536 + ], + "type": "text", + "content": "5.2 Increasing Data Diversity" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "As observed in Section 5.1, if " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " having a more diverse pretraining set benefits SSL methods, under a normalized computation assumption. However, to increase diversity, sampling from the same distribution " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " to extend " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " is not always attainable. In fact, pretraining data is often sampled from a distribution different than " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". The added samples will then introduce a distribution shift between " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". Then, we raise the following question: is increasing pretraining diversity still beneficial when there is a distribution shift between " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "? We explore strategies for acquiring new data for increasing " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ", namely, including existing samples, crawled internet data, and data synthetically generated. To evaluate the effects of the distribution" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 430, + 244 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 430, + 244 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 430, + 244 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 430, + 244 + ], + "type": "image", + "image_path": "37aedb84907c1bd3433b5fd0bd22e76fa4499276e05d60c6494e4a009d20084e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 310 + ], + "type": "text", + "content": "Fig. 2: Data Collection Strategies: We analyze strategies for collecting additional data (A), i.e., collecting more source data, crawling the web or using synthetic images. Using a class prior (top row) simulates In-distribution trainings. We also collect images without class prior (bottom row) to analyze the interactions between diversity and Out-of-distribution classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 362 + ], + "type": "text", + "content": "shift in a controlled scenario, we analyze distributions closer to " + }, + { + "bbox": [ + 130, + 338, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 338, + 482, + 362 + ], + "type": "text", + "content": " by using a class prior (in-distribution) and without a class prior (out-of-distribution)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": "Setup We use ImageNet-100 [57] as " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", and we construct multiple " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " to evaluate the effects of different data collection strategies and the distribution shift. We first introduce a set " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " composed of " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "65\\times 10^{3}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " images from ImageNet-100 (50% of the dataset), which we use as a baseline for " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " with minimum diversity. We denote the 100 classes in ImageNet-100 as " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{100}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", and sample " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " uniformly including images from all classes. Next, we compare with pretraining on more diverse datasets as " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", imposing " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{B} \\cup \\mathbb{A}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " includes " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "65\\times 10^{3}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " images sampled with one of three strategies. To highlight the effects of the distribution shift, we include in " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " either images from " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "In" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": "-distribution classes, i.e., selecting images from classes overlapping with " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{100}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", or images from " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "Out" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": "-of-distribution classes, which do not overlap with " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{100}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ". This is to study the effects of the distribution shift, since we do not assume access to downstream classes in real scenarios. For the " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "Out" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": "-of-distribution samples, we define " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " as (1) random images sampled from the full ImageNet [18] dataset; (2) images crawled from Flickr, Bing, and DuckDuckGo; or (3) images generated with Stable Diffusion V2.1 [51]. We respectively call these sets " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Source}}^{Out}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Web}}^{Out}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Synthetic}}^{Out}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ". We also define their " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "In" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": "-distribution counterparts as " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Source}}^{In}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Web}}^{In}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Synthetic}}^{In}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Note how " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{B} \\cup \\mathbb{A}_{\\mathrm{Source}}^{In}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " is the full ImageNet-100, coherently with Section 5.1. Figure 2 shows each collection strategy, for which we provide implementation details in supplementary material. Although many factors (such as the images appearance) impact the distribution shift, using a class prior with any strategy would result in closer distribution with respect to the same strategy without class priors [15]. We pretrain a ResNet-18 from scratch, with the same settings of Section 5.1, and " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 50\\times 10^6" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ". Note that this results in " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = 1.25\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " for pretraining on " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = 2.5\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " for pretraining on " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{B} \\cup \\mathbb{A}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": ", introducing a " + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 363, + 482, + 666 + ], + "type": "text", + "content": " difference of a factor of 2." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 194 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 194 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 194 + ], + "type": "image", + "image_path": "499aef39458c4cf65338b219cc5dccb382b983eb764f0832660f835fe8136d0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": "Fig. 3: Effect of Various Data Sources on SSL Pretraining: We use a baseline set " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": " (black dashed line), comprising " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "65 \\times 10^{3}" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": " images from ImageNet-100, for pretraining a ResNet-18 with " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 50 \\times 10^{6}" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": ". Augmenting " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": " with In-distribution images enhances performance (above black line), while Out-of-distribution augmentations reduce it (below black line)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": "Results We report the linear probing accuracies for pretraining on each " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{B} \\cup \\mathbb{A}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " as bars in Figure 3, showing the " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": "-normalized training on " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " as a dashed line. Surprisingly, without class priors " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "(\\mathbf{Out})" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ", including " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Source}}^{Out}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Web}}^{Out}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Synthetic}}^{Out}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " underperforms compared to pretraining on " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " only. For instance, while SimCLR on " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " scores " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "72.3\\%" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " accuracy, increasing diversity reduces the accuracy by " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " in all cases. This might appear to conflict with our findings in Section 5.1, however, the inclusion of " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{Out}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " samples leads to " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ", since we sample only classes not included in " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ". We infer that, with normalized " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ", increasing " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " without distribution priors may negatively impact performance. Conversely, when class priors are available " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "(\\mathbf{In})" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ", increasing pretraining diversity " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " improves performance compared to " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " pretraining. For instance, pretraining on " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Web}}^{In}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " performs comparably to augmenting with additional ImageNet samples " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "(\\mathbb{A}_{\\mathrm{Source}}^{In})" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ", as in the case of SimCLR where the inclusion of " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Web}}^{In}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " scores only " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " lower than " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Source}}^{In}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": ". Including " + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{\\mathrm{Synthetic}}^{In}" + }, + { + "bbox": [ + 130, + 285, + 482, + 501 + ], + "type": "text", + "content": " data also helps, although more limitedly due to the visual discrepancies between generated and real images. Ultimately, the effectiveness of diversity is linked to the distribution shift. These findings highlight the impact of the distribution shift on computationally-normalized SSL pretraining and help define evaluation practices for this task (see Section 7)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 509, + 204, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 509, + 204, + 522 + ], + "spans": [ + { + "bbox": [ + 147, + 509, + 204, + 522 + ], + "type": "text", + "content": "Insight (2)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 531, + 466, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 531, + 466, + 579 + ], + "spans": [ + { + "bbox": [ + 146, + 531, + 466, + 579 + ], + "type": "text", + "content": "When the distributions of the upstream and downstream tasks differ, " + }, + { + "bbox": [ + 146, + 531, + 466, + 579 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 146, + 531, + 466, + 579 + ], + "type": "text", + "content": ", and in a normalized computation setup, increasing pretraining diversity " + }, + { + "bbox": [ + 146, + 531, + 466, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 146, + 531, + 466, + 579 + ], + "type": "text", + "content": " may harm SSL pretraining performance, reflecting the adverse effects of distribution shift." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 609, + 307, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 307, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 307, + 621 + ], + "type": "text", + "content": "5.3 Scaling Pretraining Diversity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We showed that diversity improves pretraining performance when the training set and downstream task share the same data distribution " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}})" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": ", as discussed in Section 5.1. This may change when a distribution shift is introduced," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 191, + 304, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: Performance of ImageNet and YFCC100M SSL Pretraining on Various Downstream Tasks: We train ResNet-50 and ViT-B/16 under normalized computation " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "(\\mathcal{C} = 98\\times 10^{6})" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": " using SimCLR (left) and MoCoV3 (right) on ImageNet (IN) and YFCC100M (YFCC) with multiple " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": ". Despite the significantly larger " + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 114, + 482, + 180 + ], + "type": "text", + "content": " when trained on YFCC100M, these models cannot offset the effects of the distribution shift and are outperformed by models pretrained on ImageNet in most downstreams." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 191, + 304, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 191, + 304, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 304, + 312 + ], + "type": "table", + "html": "
SimCLR
\\( \\mathbb{D}_{\\text{SSL}} \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3156.943.082.373.945.459.9
IN0.2562.6161.145.584.076.047.064.4
IN0.6406.5463.746.884.278.848.367.2
IN1.28113.064.546.484.679.548.868.0
YFCC98.17100057.337.276.658.950.162.1
ViT-B/16IN0.1281.3154.237.281.870.344.664.3
IN0.2562.6161.339.583.877.747.269.4
IN0.6406.5465.539.984.680.449.173.1
IN1.28113.066.739.583.781.750.073.6
YFCC98.17100054.525.072.559.749.565.4
", + "image_path": "edd5a49ed0b7a3a5341403c828aabb977afce7e29ddd7536d0ec6cc8eaa674a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 199, + 316, + 238, + 323 + ], + "lines": [ + { + "bbox": [ + 199, + 316, + 238, + 323 + ], + "spans": [ + { + "bbox": [ + 199, + 316, + 238, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 98\\times 10^{6}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 191, + 480, + 312 + ], + "blocks": [ + { + "bbox": [ + 307, + 191, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 191, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 480, + 312 + ], + "type": "table", + "html": "
MoCoV3
\\( \\mathbb{D}_{\\text{SSL}} \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy↑
INCarsFlow.PetsPlacesFood
ResNet-50IN0.1281.3158.140.681.876.0445.1
IN0.2562.6162.945.385.079.847.6
IN0.6406.5465.448.486.181.949.2
IN1.28113.065.948.886.682.649.5
YFCC98.17100060.442.282.666.350.767.3
ViT-B/16IN0.1281.3157.933.482.878.046.7
IN0.2562.6163.735.085.382.948.4
IN0.6406.5467.239.585.085.849.7
IN1.28113.068.841.986.586.550.3
YFCC98.17100057.225.270.343.850.364.0
", + "image_path": "3ef61d7f3f67f77f54c546e172a6b5bc5134fa65c1663ff8bd541c16645ae8dd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 374, + 315, + 411, + 323 + ], + "lines": [ + { + "bbox": [ + 374, + 315, + 411, + 323 + ], + "spans": [ + { + "bbox": [ + 374, + 315, + 411, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 98\\times 10^{6}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 425 + ], + "type": "text", + "content": "as explored in Section 5.2. However, it is still unclear from Section 5.2, whether including a larger number of samples, and thus increasing considerably the pretraining diversity, can compensate for the negative effects of the distribution shift. To address this, the following section presents larger-scale experiments, employing significantly varied " + }, + { + "bbox": [ + 130, + 350, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 350, + 482, + 425 + ], + "type": "text", + "content": " values, aiming to explore the interplay between pretraining diversity and different distributions using millions of samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": "Setup For our large-scale pretraining experiments, we set " + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": " to be two datasets of significantly different sizes: ImageNet [18] and YFCC100M [55], comprising 1.28 million and 98 million images, respectively. Following Section 5.1, we explore multiple " + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": " values for pretraining. We set " + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 98 \\times 10^6" + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": ", which corresponds to one epoch on YFCC100M, iterating once through each of its 98 million images to maximize diversity " + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "inline_equation", + "content": "(\\mathcal{D} = 1)" + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": ". Normalizing " + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": " (see Section 5.1), we pretrain on ImageNet for approximately 77 epochs, cumulatively utilizing 98 million samples. Due to the extensive cost of these experiments, we focus on SimCLR and MoCoV3 only. We also employ larger architectures, namely ResNet-50 [33] and ViT-B/16 [21], to leverage the extensive scale of the pretraining datasets. We also use multiple " + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": " including ImageNet [18], Stanford Cars [38], Flowers-102 [44], Oxford-IIIT Pets [48], Places365 [69], and Food-101 [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "Results The results of our large-scale experiments are detailed in Table 2. Consistently with findings in Section 5.1, increasing " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " leads to better pretraining efficacy when " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ". This is evident when ImageNet is used for both pretraining and downstream classification, reinforcing that our observations hold even at a larger scale. Instead, models pretrained on YFCC100M show substantially lower performance compared to those pretrained on ImageNet, although" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "having much higher " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ". This highlights the inefficiency of collecting data indiscriminately without considering distribution shifts. To stress this, note how the model pretrained on YFCC100M (" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "D = 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ") often performs similarly to those pretrained with drastically lower " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " on ImageNet (" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "D = 1.31 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "). This aligns with our observations in Section 5.2, emphasizing that distribution differences remain a significant factor even when training with large datasets. However, the YFCC100M-pretrained model outperforms the ImageNet-pretrained model in Places365, suggesting a closer distribution relationship between YFCC100M and Places365. We explore this hypothesis further in Section 6, where we analyze distribution distances with existing metrics. Ultimately, our analysis highlights that scaling the data is not an effective solution for compensating the distribution shift, when computational resources are normalized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 269, + 203, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 269, + 203, + 281 + ], + "spans": [ + { + "bbox": [ + 148, + 269, + 203, + 281 + ], + "type": "text", + "content": "Insight (3)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 291, + 466, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 291, + 466, + 327 + ], + "spans": [ + { + "bbox": [ + 146, + 291, + 466, + 327 + ], + "type": "text", + "content": "Even an extremely large data diversity cannot mitigate the distribution shift under normalized computation. This emphasizes the importance of further research in how effectively SSL pretraining algorithms generalize." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 345, + 271, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 271, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 271, + 358 + ], + "type": "text", + "content": "6 Additional Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 415 + ], + "type": "text", + "content": "Previously, we introduced a normalized evaluation to study the relation of " + }, + { + "bbox": [ + 130, + 365, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 365, + 482, + 415 + ], + "type": "text", + "content": " and data distributions. We showed that, although diversity helps when " + }, + { + "bbox": [ + 130, + 365, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 415 + ], + "type": "text", + "content": ", vision models are unable to compensate for the distributional differences with normalized costs. We now analyze additional elements supporting our findings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": "Distribution Distances Using FID & VisualDNA In Section 5.3, we showed that pretraining on ImageNet typically outperforms YFCC100M on a variety of downstream tasks, with Places365 being the exception. We speculate that the distribution of ImageNet is more aligned with those of the downstream tasks compared to YFCC100M. To verify this, we evaluate the similarity between the datasets using FID [34] and VisualDNA [50] and report results in Table 3. With both FID or VisualDNA, the distribution of ImageNet is always closer to the downstream tasks, except for Places365 where YFCC100M is closer. This aligns with the lower performance of ImageNet on this dataset only (Table 2), further suggesting that the performance drop is caused by the distribution shift." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Importance of Normalized Computation We now study the impact of normalizing computation on the performance of SSL methods. We aim to understand if " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " is not normalized across methods, this will lead to misleading conclusions, and thus motivating our normalized computation. In Table 4, we compare the performance of ResNet-50 and ViT-B/16 models pretrained using MoCoV3 with " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 98 \\times 10^6" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " (as used in Section 5.3) against a tripled budget of " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 294 \\times 10^6" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". We show that inconsistencies arise when pretraining data diversity " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " and computation " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " are not fairly compared. For instance, in scenarios highlighted in red, we notice that a lower pretraining diversity coupled with a higher" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "text", + "content": "Table 3: Comparing VisualDNA and FID Scores Across Datasets: We assess the relationship between VisualDNA [50] and FID Score [34] for various large-scale " + }, + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "text", + "content": " and several downstream tasks. For VisualDNA, activations are extracted as suggested [50] with MUGS [70] with ViT-B/16 (MUGS) or DINO v2 [47] with ViT-B/16 (DINO v2), while FID activations are obtained via an Inception network [54]. Consistently, across all metrics, the ranking of dataset distances between " + }, + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "text", + "content": " and various " + }, + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 132, + 114, + 301, + 301 + ], + "type": "text", + "content": " aligns with the accuracy ranking in Table 2. Models exhibiting lower VisualDNA/FID scores benefit more from the diversity in pretraining data for SSL." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 311, + 301, + 415 + ], + "blocks": [ + { + "bbox": [ + 134, + 311, + 301, + 415 + ], + "lines": [ + { + "bbox": [ + 134, + 311, + 301, + 415 + ], + "spans": [ + { + "bbox": [ + 134, + 311, + 301, + 415 + ], + "type": "table", + "html": "
DSSLBackboneVisualDNA↓
INCarsFlow.PetsPlacesFood
INMUGS011.4912.466.097.199.08
YFCCMUGS3.7211.5712.717.936.209.72
INDINO v207.056.954.433.527.26
YFCCDINO v22.377.127.465.752.747.90
DSSLBackboneFID↓
INCarsFlow.PetsPlacesFood
INInception0143.4192.988.964.9114.9
YFCCInception48.1174.3214.8145.838.9154.5
", + "image_path": "b98970d2b868e58612791ef9532b45d3bb331f34d40e0adb672974134e1ee861.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "text", + "content": "Table 4: Importance of Normalization. We report the accuracy of MoCoV3 trained on ImageNet (IN) with different data diversity and variable " + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "text", + "content": ", for the in-distribution assumption " + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "inline_equation", + "content": "(\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}})" + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "text", + "content": ". For a given network, cells in red highlight inconsistencies where although the model trained with " + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 294 \\times 10^6" + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "text", + "content": " has seen less samples, it outperforms the best model trained with one third of the computational budget " + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "inline_equation", + "content": "(\\mathcal{C} = 90 \\times 10^6)" + }, + { + "bbox": [ + 307, + 117, + 479, + 261 + ], + "type": "text", + "content": ", showing the importance of normalization for understanding how models exploit data." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 312, + 279, + 476, + 415 + ], + "blocks": [ + { + "bbox": [ + 373, + 270, + 414, + 278 + ], + "lines": [ + { + "bbox": [ + 373, + 270, + 414, + 278 + ], + "spans": [ + { + "bbox": [ + 373, + 270, + 414, + 278 + ], + "type": "text", + "content": "MoCoV3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 279, + 476, + 415 + ], + "lines": [ + { + "bbox": [ + 312, + 279, + 476, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 279, + 476, + 415 + ], + "type": "table", + "html": "
Network\\( \\mathbb{D}_{\\text{SSL}} \\)\\( \\mathcal{C} \\times 10^6 \\)\\( N \\times 10^6 \\)\\( \\mathcal{D} \\times 10^{-3} \\)Accuracy† IN
ResNet-50IN980.1281.3158.1
IN980.6406.5465.4
IN981.28113.065.9
IN2940.1280.4357.4
IN2940.6402.1769.8
IN2941.2814.3571.4
ViT-B/16IN980.1281.3157.9
IN980.6406.5467.2
IN981.28113.068.8
IN2940.1280.4356.9
IN2940.6402.1771.9
IN2941.2814.3574.9
", + "image_path": "b2522af5faa0c747ba7fe09f340dc11c9cdcb516a517d5d8c19d1ab2a8b5e221.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "content": "computational budget can yield better results. We observe that under pretraining diversity of only " + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = 13\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "content": " and a computational budget of " + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 98\\times 10^6" + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "content": " ResNet-50 only enjoys a " + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "inline_equation", + "content": "65.9\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "content": " accuracy compared against with " + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "inline_equation", + "content": "69.8\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "content": " with a smaller pretraining diversity of " + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "inline_equation", + "content": "2.17\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "content": " but with a larger computational budget of " + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "inline_equation", + "content": "294\\times 10^{6}" + }, + { + "bbox": [ + 130, + 434, + 482, + 507 + ], + "type": "text", + "content": ". This shows that without normalized computation, it could be incorrectly concluded that pretraining diversity does not play a significant role." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Model Saturation We aim to study the pretraining trend on YFCC100M, and if adding more samples could compensate the distribution shift with ImageNet. Hence, we extend the YFCC100M experiments from Section 5.3, examining various subsets—specifically " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of the dataset. Again, we normalize the computational budget to " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 98 \\times 10^6" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", equivalent to one epoch on YFCC100M. The results of linear probing models pretrained using SimCLR and MoCoV3 with ResNet-50 and ViT-B/16 on ImageNet are shown in Figure 4. Interestingly, we obtain a performance plateau. This observation points to a saturation point in performance, showing that simply increasing the dataset further would not bridge the gap between the " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". Hence, arbitrarily increasing pretraining diversity " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is not sufficient to bridge the distribution shift problem. Here " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", hence still aligning with our findings in Section 5.1." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 193 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 193 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 193 + ], + "type": "image", + "image_path": "a0b7c13e724ad3bd7d335cdc10821de1486178e39607abbcd2524a10e6098ece.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "text", + "content": "Fig. 4: Data Diversity Impact on YFCC100M Pretraining Performance: Pretraining " + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "inline_equation", + "content": "(\\mathcal{C} = 98\\times 10^{6})" + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "text", + "content": " of networks with " + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathrm{YFCC100M}" + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}" + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "text", + "content": " for several dataset subsets. In the presence of a distribution shift, performance tends to saturate and does not benefit from additional data diversity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 340, + 304, + 403 + ], + "blocks": [ + { + "bbox": [ + 130, + 274, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 329 + ], + "type": "text", + "content": "Table 5: Evaluating Network Accuracy With Varied Label Quantity: We evaluate the accuracy of networks trained on ImageNet (IN) and YFCC100M (YFCC) with different labeling percentages of " + }, + { + "bbox": [ + 130, + 274, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}" + }, + { + "bbox": [ + 130, + 274, + 482, + 329 + ], + "type": "text", + "content": ". The increased diversity still does not compensate for the distribution shift. However, for in-distribution data, one can get away with fewer labels with more diverse pretraining data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 340, + 304, + 403 + ], + "lines": [ + { + "bbox": [ + 135, + 340, + 304, + 403 + ], + "spans": [ + { + "bbox": [ + 135, + 340, + 304, + 403 + ], + "type": "table", + "html": "
SimCLR
Network\\( \\mathbb{D}_{\\text{SSL}} \\)N×106D×10-3Accuracy†
5%10%50%
ResNet-50IN1.281.3150.354.261.7
YFCC98.2100039.444.354.0
ViT-B/16IN1.281.3155.259.465.6
YFCC98.2100040.745.753.3
C=98×10^6
", + "image_path": "97e66e7c08440b54ffa4f3f1dc840b86b6b1a7e28771e1d9c7242bbcaa6a9526.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 340, + 481, + 403 + ], + "blocks": [ + { + "bbox": [ + 309, + 340, + 481, + 403 + ], + "lines": [ + { + "bbox": [ + 309, + 340, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 309, + 340, + 481, + 403 + ], + "type": "table", + "html": "
NetworkMoCoV3
\\( \\mathbb{D}_{\\text{SSL}} \\)N×106D×10-3Accuracy↑
5%10%50%
ResNet-50IN1.281.3152.156.263.8
YFCC98.2100042.848.257.7
ViT-B/16IN1.281.3159.363.068.3
YFCC98.2100043.949.056.0
C=98×10^6
", + "image_path": "a186ff28a4385e0e76328b728e3b193a41f53469e4ff6a1fd264f4693f91f6fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Label Quantity A question that arises in our setting is whether having a higher pretraining diversity leads to requiring less labeled samples during linear probing. In this section, we focus on understanding the impact of increased " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " on the number of labeled samples required for effective linear probing in downstream tasks. We use the same trained models from Section 5.3, i.e., SimCLR and MoCoV3, with ResNet-50 and ViT-B/16 architectures. In this setup, we set " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}} = \\mathrm{ImageNet}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", with two upstream dataset scenarios: ImageNet (where " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ") and YFCC100M (where " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "). Our experiments are summarized in Table 5. We note that with ViT-B/16, if ImageNet is used for pretraining, linear probing with just " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " labeled samples can surpass the performance achieved using " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " labeled data when YFCC100M serves as " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ". This also implies that when " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} \\neq \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", a higher quantity of labeled data is necessary to attain competitive performance, and increasing " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " does not bring considerable benefits on label quantity under distribution shift. This implies that our findings in Section 5.3 hold regardless of the linear probing labeled set. We note that in scenarios where " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}} = \\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", using only " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " of the labeled data can achieve similar performance as utilizing the full " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " of labeled samples, implying that increasing " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " leads to reduced label requirement efforts for downstream tasks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 217, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 217, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 217, + 127 + ], + "type": "text", + "content": "7 Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 147, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 147, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 147, + 482, + 363 + ], + "type": "text", + "content": "Main Conclusions Our set of experiments leads to Insight (3) in Section 5.3, revealing that with normalized computational costs, SSL pretrainings with large diversity cannot compensate for the distribution shift. This is surprising, since the variety of information that SSL algorithms could benefit from during training is much higher in large generic datasets than in small ones. Since our evaluation is cost-normalized, (3) also shows that SSL strategies are not efficiently exploiting the pretraining diversity on large datasets for representation extraction. This inefficiency implies a wide margin for improvement of generalization of SSL models, making better use of the computational power used for training. The role of existing models must also be discussed in this context. Following Insights (1) and (2), respectively in Sections 5.1 and 5.2, we have studied how in-distribution and out-of-distribution data impact performance in a controlled scenario. We argue that this should be taken into account in the evaluation of the performance of SSL models. Indeed, training at scale may enlarge the in-distribution space, including classes of the downstream tasks in the training set. In recent literature, this is a design choice to maximize performance on popular benchmarks [47]. While this allows for achieving impressive results in many tasks, we stress that this does not permit a fair evaluation. Now, we summarize practical takeaways." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 363, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 471 + ], + "type": "text", + "content": "Training Takeaways Coherently with our findings in Section 5.1 and similarly to prior art [15,28,37], we find that aligned distributions benefit performance, in particular increasing " + }, + { + "bbox": [ + 130, + 363, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 471 + ], + "type": "text", + "content": " diversity helps, as long as the distribution of the new samples match those of the downstream task data. Differently from the state-of-the-art, we demonstrate that this holds also in a computationally-normalized setting, implying that collecting large scale in-distribution data matching the downstream task is be an effective approach for improving SSL. Hence, for practical applications distribution priors should be used, if available. On the contrary, for a fair evaluation of models, this should not be the case, as specified below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": "Evaluation Takeaways Our analysis reveals that to permit a fair evaluation of SSL methods, computationally normalized tests are necessary to avoid inconsistencies, as shown in Section 6. Moreover, it is crucial to identify out-of-distribution downstream tasks for a correct evaluation of generalization performance. By evaluating only on " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{task}}" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": " with a low distribution shift, there is a risk of reporting inflated metrics, not representative of a real gain in generalization. This is important, since new SSL approaches may be reporting higher downstream performance when pretrained on a different dataset. We relate this to Sections 5 and 6, where we show that increasing the computation and the in-distribution data, respectively, can improve performance. Ultimately, wrong practices may result in incorrectly concluding that an SSL algorithm is performing better." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Differences With Language Models In Section 5.3 we showed that even very diverse datasets (YFCC100M), fall short in satisfactory generalization performance. In addition to further exploration into generalization for SSL pretraining, this open doors to investigating why language models enjoy enhanced generalization when trained on highly diverse data compared to vision models [61]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "This work was supported by SDAIA-KAUST Center of Excellence in Data Science, Artificial Intelligence (SDAIA-KAUST AI). Fabio Pizzati is financed by KAUST (Grant DFR07910). Philip H.S. Torr thanks the Royal Academy of Engineering for their support. This work is supported by a UKRI grant Turing AI Fellowship (EP/W002981/1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 246, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "text", + "content": "1. Bachman, P., Hjelm, R.D., Buchwalter, W.: Learning representations by maximizing mutual information across views. In: NeurIPS (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "text", + "content": "2. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 302, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 322 + ], + "type": "text", + "content": "3. Bao, H., Dong, L., Piao, S., Wei, F.: BEit: BERT pre-training of image transformers. In: ICML (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "text", + "content": "4. Bardes, A., Ponce, J., LeCun, Y.: VICReg: Variance-invariance-covariance regularization for self-supervised learning. In: ICLR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 367 + ], + "type": "text", + "content": "5. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-Mining discriminative components with random forests. In: ECCV (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 369, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 389 + ], + "type": "text", + "content": "6. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: ECCV (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 391, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 411 + ], + "type": "text", + "content": "7. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: ICCV (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 433 + ], + "type": "text", + "content": "8. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "text", + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 477 + ], + "type": "text", + "content": "0. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 479, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 499 + ], + "type": "text", + "content": "1. Chen, T., Kornblith, S., Swersky, K., Norouzi, M., Hinton, G.: Big self-supervised models are strong semi-supervised learners. In: NeurIPS (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "text", + "content": "2. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 523, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 543 + ], + "type": "text", + "content": "3. Chen, X., He, K.: Exploring simple siamese representation learning. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "text", + "content": "4. Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "5. Cole, E., Yang, X., Wilber, K., Aodha, O.M., Belongie, S.: When does contrastive visual representation learning work? In: CVPR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 609 + ], + "type": "text", + "content": "6. Costa, V.G.T.d., Fini, E., Nabi, M., Sebe, N., Ricci, E.: solo-learn: A library of self-supervised methods for visual representation learning. In: JMLR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "7. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: ICML (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: ICCV (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "20. Doersch, C., Zisserman, A.: Multi-task self-supervised visual learning. In: ICCV (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 204 + ], + "type": "text", + "content": "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "22. El-Nouby, A., Izacard, G., Touvron, H., Laptev, I., Jegou, H., Grave, E.: Are large-scale datasets necessary for self-supervised pre-training? arXiv (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "23. Ericsson, L., Gouk, H., Hospedales, T.M.: How well do self-supervised models transfer? In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "24. Fini, E., Astolfi, P., Romero-Soriano, A., Verbeek, J., Drozdzal, M.: Improved baselines for vision-language pre-training. TMLR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "25. Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by predicting image rotations. In: ICLR (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "26. Goyal, P., Caron, M., Lefaudeauux, B., Xu, M., Wang, P., Pai, V., Singh, M., Liptchinsky, V., Misra, I., Joulin, A., et al.: Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "27. Goyal, P., Duval, Q., Seessel, I., Caron, M., Misra, I., Sagun, L., Joulin, A., Bojanowski, P.: Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "28. Goyal, P., Mahajan, D., Gupta, A., Misra, I.: Scaling and benchmarking self-supervised visual representation learning. In: ICCV (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "29. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent: A new approach to self-supervised learning. In: NeurIPS. vol. 33 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "text", + "content": "30. Hammoud, H.A.A.K., Itani, H., Pizzati, F., Torr, P., Bibi, A., Ghanem, B.: Synthclip: Are we ready for a fully synthetic clip training? arXiv (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "31. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "32. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "33. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "35. Hjelm, R.D., Fedorov, A., Lavoie-Marchildon, S., Grewal, K., Bachman, P., Trischler, A., Bengio, Y.: Learning deep representations by mutual information estimation and maximization. In: ICLR (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "36. Henaff, O.J., Srinivas, A., De Fauw, J., Razavi, A., Doersch, C., Eslami, S.M.A., Oord, A.v.d.: Data-efficient image recognition with contrastive predictive coding. In: ICML (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "37. Kotar, K., Ilharco, G., Schmidt, L., Ehsani, K., Mottaghi, R.: Contrasting contrastive self-supervised representation learning pipelines. In: ICCV (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "38. Krause, J., Deng, J., Stark, M., Fei-Fei, L.: Collecting a large-scale dataset of fine-grained cars. In: FGVC (2013)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "39. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 466, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 466, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 466, + 128 + ], + "type": "text", + "content": "40. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 480, + 149 + ], + "type": "text", + "content": "41. Li, A.C., Brown, E.L., Efros, A.A., Pathak, D.: Internet explorer: Targeted representation learning on the open web. In: ICML (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 149, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 172 + ], + "type": "text", + "content": "42. Li, J., Zhou, P., Xiong, C., Hoi, S.C.H.: Prototypical contrastive learning of unsupervised representations. In: ICLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "text", + "content": "43. Misra, I., Maaten, L.v.d.: Self-supervised learning of pretext-invariant representations. In: CVPR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "44. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP (2008)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "45. Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: ECCV (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "46. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "47. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: DINoV2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "48. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 314, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 336 + ], + "type": "text", + "content": "49. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: CVPR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "type": "text", + "content": "50. Ramtoula, B., Gadd, M., Newman, P., De Martini, D.: Visual DNA: Representing and comparing images using distributions of neuron activations. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "text", + "content": "51. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "52. Shi, Y., Daunhawer, I., Vogt, J.E., Torr, P., Sanyal, A.: How robust is unsupervised representation learning to distribution shift? In: ICLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "53. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: ICCV (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "54. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: CVPR (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "55. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: YFCC100M: The new data in multimedia research. In: Commun. ACM (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "56. Tian, Y., Henaff, O.J., Van Den Oord, A.: Divide and contrast: Self-supervised learning from uncurated data. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "type": "text", + "content": "57. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 512, + 452, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 452, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 452, + 523 + ], + "type": "text", + "content": "58. Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: CVPR (2011)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "59. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Gool, L.V.: Revisiting contrastive methods for unsupervised learning of visual representations. In: NeurIPS (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "60. Van Horn, G., Cole, E., Beery, S., Wilber, K., Belongie, S., Mac Aodha, O.: Benchmarking representation learning for natural world image collections. In: CVPR (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "61. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. TMLR (2022)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "62. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via non-parametric instance discrimination. In: CVPR (2018)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "63. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMIM: A simple framework for masked image modeling. In: CVPR (2022)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Pretraining Data Diversity for SSL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 281 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "64. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Wei, Y., Dai, Q., Hu, H.: On data scaling in masked image modeling. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "65. Zbontar, J., Jing, L., Misra, I., LeCun, Y., Deny, S.: Barlow Twins: Self-supervised learning via redundancy reduction. In: ICML (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 469, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 469, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 469, + 194 + ], + "type": "text", + "content": "67. Zhang, R., Isola, P., Efros, A.A.: Colorful image colorization. In: ECCV (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "68. Zhao, N., Wu, Z., Lau, R.W.H., Lin, S.: What makes instance discrimination good for transfer learning? In: ICML (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "text", + "content": "69. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. T-PAMI (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "70. Zhou, P., Zhou, Y., Si, C., Yu, W., Ng, T.K., Yan, S.: Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "71. Zhuang, C., Zhai, A.L., Yamins, D.: Local aggregation for unsupervised learning of visual embeddings. In: ICCV (2019)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Hammoud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_content_list.json b/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..20968e88987d06afaab6012617742107915c77f8 --- /dev/null +++ b/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_content_list.json @@ -0,0 +1,2015 @@ +[ + { + "type": "text", + "text": "On Spectral Properties of Gradient-based Explanation Methods", + "text_level": 1, + "bbox": [ + 264, + 140, + 738, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amir Mehrpanah1, Erik Englesson1, and Hossein Azizpour1", + "bbox": [ + 269, + 210, + 732, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, Royal Institute of Technology, Stockholm, Sweden {amirme,engless,azizpour}@kth.se", + "bbox": [ + 218, + 239, + 781, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Understanding the behavior of deep networks is crucial to increase our confidence in their results. Despite an extensive body of work for explaining their predictions, researchers have faced reliability issues, which can be attributed to insufficient formalism. In our research, we adopt novel probabilistic and spectral perspectives to formally analyze explanation methods. Our study reveals a pervasive spectral bias stemming from the use of gradient, and sheds light on some common design choices that have been discovered experimentally, in particular, the use of squared gradient and input perturbation. We further characterize how the choice of perturbation hyperparameters in explanation methods, such as SmoothGrad, can lead to inconsistent explanations and introduce two remedies based on our proposed formalism: (i) a mechanism to determine a standard perturbation scale, and (ii) an aggregation method which we call SpectralLens. Finally, we substantiate our theoretical results through quantitative evaluations.", + "bbox": [ + 261, + 297, + 738, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Probabilistic Machine Learning $\\cdot$ Gradient-based Explanation Methods $\\cdot$ Probabilistic Pixel Attribution Techniques $\\cdot$ Explainability $\\cdot$ Deep Neural Networks $\\cdot$ Spectral Analysis", + "bbox": [ + 261, + 518, + 738, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 584, + 375, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ensuring the interpretability of machine learning models is crucial for user's trust and understanding [2, 6, 45]. While considerable efforts have been dedicated to advancing explainability in the field [37, 47, 49, 51, 57, 62], there is a notable gap in systematically formalizing and consolidating existing knowledge [5, 9, 16, 19]. Insufficient formal analysis leaves too much room for different interpretations of the explanations, making them unreliable [27] which is detrimental to the very use case of explanations, i.e., human-critical applications. Therefore, more attention is essential to retrospectively analyze and structure the wealth of work done, creating a cohesive foundation for explainability [39, 61].", + "bbox": [ + 212, + 613, + 787, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One main reason for the unreliability of explanation methods is their susceptibility to hyperparameters, leading to inconsistent explanations [4, 13, 33, 36, 48], see Fig. 1. Among the several works that formally study explanation methods [19, 37, 57], to our knowledge, none shed light on such inconsistencies. Therefore, in this work, we adopt a novel probabilistic and spectral perspective to characterize the observed inconsistencies and propose preliminary solutions.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e68eef4b31cc9eeeeb66fd83a96ef769983ba8c769329e421842f6b592caf38.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 357, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6790d550b909fe52e3755320774161cf06857cfd3ca86a9329e5809e880bef88.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 364, + 145, + 498, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03c52b4bd53b511552856ccfd5f167b88f3d2a9db724060736f777498edc1087.jpg", + "image_caption": [ + "(c)", + "Fig. 1: The figure illustrates sample explanations generated by varying hyperparameters of SmoothGrad-Squared [30, 55] for ResNet-50 on ImageNet (see Fig. 16 in the Appendix for more examples). It is evident from the figure that using different hyperparameters results in diverse explanations. Such variations have been demonstrated in recent studies for other explanation techniques as well [4, 13, 33, 48]." + ], + "image_footnote": [], + "bbox": [ + 506, + 145, + 638, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/43820e97b3834c99683d7a63a41c4cd1303e12cb49f6ea9acd56f4234bf3ec06.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 645, + 145, + 779, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Contribution. We iterate our contributions in three parts:", + "bbox": [ + 215, + 377, + 686, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Analysis: We introduce a probabilistic representation for post-hoc gradient-based explanation methods in Sec. 2.1, and perform a spectral analysis of them in Sec. 3.1.", + "- Identified Issues: Our examination reveals a shared spectral bias stemming from the fundamental assumption of utilizing gradients to measure feature contributions. We show that the combined spectral effects of gradient and perturbation lead to band-pass filters modulating the explanations, which could lead to inconsistencies in the explanations, discussed in Sec. 3.2. Moreover, we conduct a spectral analysis of a few specific perturbation distributions that have been proposed in the literature in Sec. 3.3.", + "- Solutions and Evaluations: To address inconsistencies, we investigate two straightforward solutions that are finding an optimal perturbation scale based on the cosine similarity of the perturbation and the classifier in the frequency domain in Sec. 4.1, and an aggregation mechanism named SpectralLens in Sec. 4.2. Finally, we evaluate the proposed methods using the pixel removal strategy [48], acknowledging their effectiveness and limitations under Sec. 5." + ], + "bbox": [ + 225, + 407, + 785, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 215, + 707, + 369, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notation. We consider a classification setting, that is, modeling $p(y|\\pmb{x})$ for labels $y \\in \\{1,\\dots,k\\}$ and $\\pmb{x} \\in \\mathbb{R}^n$ the set of input images. We denote the $\\log p(y|\\pmb{x})$ by $f$ . We denote standard Fourier bases by $\\{\\psi(\\pmb{\\omega},\\pmb{x})\\}_{\\pmb{\\omega} \\in \\mathbb{R}^n}$ , according to which, under some regularity conditions [43], $f$ can be decomposed as $f(\\pmb{x}) = \\int \\widehat{f}(\\pmb{\\omega}) \\psi(\\pmb{\\omega},\\pmb{x}) d\\pmb{\\omega}$ . In this representation, $\\widehat{f}$ denotes the Fourier transform (FT) of $f$ . We denote the power spectral density (PSD) of $f$ by $S_f(\\pmb{\\omega}) = |\\widehat{f}(\\pmb{\\omega})|^2$ .", + "bbox": [ + 212, + 744, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 A Probabilistic Representation of Explanations", + "text_level": 1, + "bbox": [ + 214, + 146, + 651, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this view, each explanation method consists of three components: (1) a distribution for perturbation of the input, (2) an explainer, and (3) a summary statistic. We use $p(\\tilde{\\pmb{x}} | \\pmb{x}, \\sigma)$ to denote the distribution of perturbations of the input sample parameterized by $\\sigma$ , referred to as the perturbation scale. A definition of \"explanation\" is recovering the contribution of features to a prediction, which emanates from human's natural counterfactual reasoning [19]. Therefore, the explainer is usually the gradient with respect to input $\\nabla_{\\pmb{x}}$ , or approximations of it, e.g. the finite difference, given by $\\Delta_{\\pmb{x}} f(\\pmb{x}) = \\frac{f(\\tilde{\\pmb{x}}) - f(\\pmb{x})}{\\tilde{\\pmb{x}} - \\pmb{x}}$ (see Appendix D). Finally, we assume that the summary statistic is the expectation of the outputs, or the squared outputs, of the explainer, e.g. $(\\nabla f(\\tilde{\\pmb{x}}))^2$ , over the chosen perturbation distribution, $p(\\tilde{\\pmb{x}} | \\pmb{x}, \\sigma)$ , denoted as $\\mathbb{E}_{p(\\tilde{\\pmb{x}})}\\left[(\\nabla f(\\tilde{\\pmb{x}}))^2\\right]$ , where the operations are carried out pixelwise. This is referred to as the explanation of $f$ at $\\pmb{x}$ with a hyperparameter $\\sigma$ . This probabilistic view can, in fact, unify different explanation methods into a single form (see Appendix C). We provide theoretical justifications for why using squared gradients is preferable to gradients, explaining the empirical findings [30], and will discuss both cases in Sec. 3.1.", + "bbox": [ + 212, + 169, + 787, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Spectral Analysis of Explanations", + "text_level": 1, + "bbox": [ + 214, + 434, + 578, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section demonstrates the spectral property in explanations when using the gradient explainer and its interaction with perturbations, where a significant difference in the frequency of learned features leads to a selective attribution of features. In the following sections, we draw on insights from spectral analysis of signals and systems theory, utilizing this viewpoint to gain a deeper understanding of explanation methods. We demonstrate a spectral bias related to gradient-based explanation approaches, identify its source, and interpret various methods in the literature as efforts to alleviate it through the incorporation of low-pass filters, either explicitly [55, 57] or implicitly [42, 62]. We illustrate that the inconsistencies in explanation techniques stem from different strategies for mitigating the spectral bias. Finally, we suggest two straightforward approaches to tackle these inconsistencies: determining an optimal perturbation scale in Sec. 4.1 and aggregating the explanations in Sec. 4.2.", + "bbox": [ + 212, + 464, + 787, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Writing a spectral representation for prediction, as shown in Appendix A.1, the model output with a perturbation distribution can be written as a convolution between the function and the perturbation distribution $f * p$ , therefore we may use \"perturbation kernel\" and \"perturbation distribution\" interchangeably.", + "bbox": [ + 212, + 660, + 785, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Spectral Representation of Explainers", + "text_level": 1, + "bbox": [ + 214, + 741, + 576, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, by drawing a connection between explainability techniques and the literature on spectral signal processing [12, 38, 56], we take a novel avenue for our analysis. This connection enables us to have a spectral representation of the explanations, as discussed in the next sections (see Appendix A for the derivations). It is important to note that the Fourier bases in our work is 1D and", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "per pixel, and should not be confused with the traditional Fourier bases used in image processing literature that expands along spatial dimensions.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gradients in Spectral Domain. For explanation methods that use gradients and a perturbation kernel, $p(\\tilde{\\pmb{x}})$ we have the following (see Appendix A.2).", + "bbox": [ + 215, + 207, + 782, + 238 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} [ \\nabla f (\\tilde {\\boldsymbol {x}}) ] = \\int i 2 \\pi \\omega \\widehat {f} (\\omega) \\widehat {p} (\\omega) d \\omega (1) \\\\ = - 2 \\pi \\left[ \\int \\omega \\left(\\widehat {f} _ {E} (\\omega) \\widehat {p} _ {O} (\\omega) + \\widehat {f} _ {O} (\\omega) \\widehat {p} _ {E} (\\omega)\\right) d \\omega \\right] (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 253, + 784, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\widehat{p}$ denotes the FT of the perturbation kernel. We used $\\widehat{f}_E$ and $\\widehat{f}_O$ to denote a function $\\widehat{f}$ is decomposed into real even and imaginary odd components $\\widehat{f} = i\\widehat{f}_O + \\widehat{f}_E^1$ . The negative sign suggests that in case of using the gradients, we need to negate the actual values to obtain rankings. We will elaborate more on the negation of the attributions in Sec. 5.", + "bbox": [ + 215, + 334, + 782, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 1. Note the symmetry in the form of the spectral representation of the gradients. This form implies that it is inevitable for any perturbation kernel to extract asymmetric information about even and odd parts of the classifier $f$ . Therefore, the gradient explainer with a kernel symmetric about the sample being explained, only extracts information about the odd parts of the classifier. This is problematic because the attribution is expected to reflect the total contribution to the prediction.", + "bbox": [ + 215, + 431, + 782, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will use this interesting symmetry later in Sec. 3.3 for analysis of a specific perturbation kernel.", + "bbox": [ + 215, + 554, + 782, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Squared Gradients in Spectral Domain. On the other hand, for explanation methods which use squared gradients, we have the following (see Appendix A.3).", + "bbox": [ + 215, + 616, + 782, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] = 4 \\pi^ {2} \\int \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) \\mathrm {S} _ {\\sqrt {p}} (\\boldsymbol {\\omega}) d \\boldsymbol {\\omega} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 662, + 782, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{S}_f$ and $\\mathrm{S}_{\\sqrt{p}}$ denote the PSD of the model and square root of the perturbation kernel respectively.", + "bbox": [ + 215, + 703, + 782, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We discuss the benefits of using squared gradients as in Eq. (3) over gradients as in Eq. (2) later in Sec. 3.3. However, it can be noticed here that the squared gradients, unlike gradients, are always non-negative, extract information about the PSD of $f$ , and do not have a complex symmetric structure.", + "bbox": [ + 215, + 734, + 782, + 795 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 Also note that as we are using real valued functions, other components i.e. real odd and imaginary even, are zero.", + "bbox": [ + 217, + 810, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Gradient Explainers with Perturbations Are Band-Pass Filters", + "text_level": 1, + "bbox": [ + 215, + 146, + 777, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our probabilistic framework, various explanation methods can be decomposed into three components (1) an explainer, (2) a perturbation kernel and (3) a summary statistic. In this section, we analyze the spectral effects of the two core components, namely gradient and perturbation. Importantly, we show that gradient acts as a high-pass filter, and perturbation acts as a low-pass filter. Before that, we need a more formal definition for a high-pass and low-pass filter that suits our study. Assuming that $\\widehat{m} (\\omega)$ denotes the FT of a perturbation kernel, we define low- and high-pass as follows.", + "bbox": [ + 212, + 178, + 782, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 1. A filter $\\widehat{m}(\\omega)$ is low pass, if for a given $M > 0$ there exists a threshold $\\omega^{*}$ such that for any $\\|\\omega\\| > \\|\\omega^{*}\\|$ , we have $\\|\\widehat{m}(\\omega)\\| < M$ , and it is high pass if such a condition holds for $\\|\\omega\\| < \\|\\omega^{*}\\|$ .", + "bbox": [ + 214, + 313, + 784, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For better readability, the proofs for the theorems and propositions are outlined in Appendix B, and we provide intuition as we progress.", + "bbox": [ + 214, + 372, + 782, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. Any perturbation kernel $p(\\tilde{\\pmb{x}})$ mitigates the attribution of high-frequency features; simply put, any perturbation is a low-pass filter.", + "bbox": [ + 214, + 420, + 782, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The intuition is that any modification to a signal, would only remove information, and, as the intensity of perturbation increases, the removal goes from high-frequency components toward lower frequencies.", + "bbox": [ + 214, + 467, + 782, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. The gradient operator amplifies the attribution of high-frequency features; simply put, the gradient operator is a high-pass filter.", + "bbox": [ + 214, + 531, + 782, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the stated theorems, one can see that a gradient-based explanation with perturbation creates a band-pass filter, whose shape is controlled by the perturbation hyperparameters (see Figs. 2 and 3).", + "bbox": [ + 214, + 579, + 782, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 2. This amplification-mitigation trade-off can be seen in several explanation methods, where perturbation kernels (low-pass filters) are utilized to mitigate the amplification effect (the high-pass filter) caused by the gradient. Therefore, the combined effect of the two filters usually forms a band-pass filter (see Appendix C for a non-comprehensive list).", + "bbox": [ + 214, + 642, + 782, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Spectral Analysis of Specific Perturbation Kernels", + "text_level": 1, + "bbox": [ + 214, + 747, + 674, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we analyze the properties of special kernels used in the literature with remarks or propositions that justify, and thus shed light on, previous work's experimental observations. In most cases, we expand upon the squared gradients, Eq. (3), due to its analytical simplicity.", + "bbox": [ + 214, + 779, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VanillaGrad via a Dirac Kernel. The very naive explanation method can be obtained with a Dirac kernel $p(\\tilde{\\pmb{x}}) = \\delta_{\\pmb{x}}(\\tilde{\\pmb{x}})$ known as VanillaGrad (VG). Analyzing the squared version of VG and following Eq. (3), we know that the FT of the Dirac kernel is $S_{\\sqrt{p}}(\\omega) = 1(\\omega)$ , where $\\mathbf{1}(\\omega)$ is the constant function mapping all values to 1. Thus, it can be seen that we recover the scaled PSD of $f$ , i.e. $\\| \\boldsymbol{\\omega} \\|^2 \\mathrm{S}_f(\\boldsymbol{\\omega})$ .", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 3. Scaling $\\widehat{f}$ by frequency can be interpreted as the bias of the gradient explainer towards exaggeration of the attribution of high-frequency features [31]. Beside other justifications for gradient sparsity [8], spectral representation of $\\mathrm{VG}^2$ and domination of high-frequency features give an orthogonal justification for why VanillaGrad explanations being usually scattered and noisy (see Fig. 12).", + "bbox": [ + 212, + 250, + 787, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SmoothGrad via a Gaussian Kernel. Letting $p(\\tilde{\\pmb{x}}) = \\mathcal{N}(\\pmb{x},\\sigma^2)$ , i.e. a Gaussian with variance $\\sigma^2$ , centered on the sample being explained, we recover variants of an explanation method known as SmoothGrad (SG) or the squared gradient version SmoothGrad-Squared (SG²) [30, 55], we can rewrite Eq. (3) as follows:", + "bbox": [ + 212, + 349, + 787, + 424 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] \\propto \\int \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) e ^ {- 8 \\pi^ {2} \\sigma^ {2} \\| \\boldsymbol {\\omega} \\| ^ {2}} d \\boldsymbol {\\omega} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 425, + 785, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This shows that the contributions $\\mathrm{S}_f(\\omega)$ are modulated by the combined spectral effect of the Gaussian perturbation kernel and gradient operator, which is a band pass filter of the form $\\| \\boldsymbol{\\omega} \\|^2 e^{-8\\pi^2 \\sigma^2} \\| \\boldsymbol{\\omega} \\|^2$ , whose shape is controlled by the kernel variance (see Fig. 2). The explanations could be inconsistent, but generally become smoother as $\\sigma$ increases, which is a result of the shrinkage of the attribution of high-frequency features that can dominate the explanations with a small $\\sigma$ (see Fig. 1).", + "bbox": [ + 212, + 460, + 787, + 568 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b31ae8acbda21c62db3680ca167d5e8643f9577041981134d95a585adb9f706a.jpg", + "image_caption": [ + "Fig.2: A visualization of frequency bands, where $\\mathrm{S}_{\\sqrt{p}}(\\omega)$ is defined according to Eq. (4), show that change in the noise scale $\\sigma$ leads to band-pass filters with different modes, hence, representing information of different frequencies." + ], + "image_footnote": [], + "bbox": [ + 382, + 601, + 624, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When generating explanations, one is tempted to look for hyperparameters that make the explanation reflect what they think it should [16]. However, monitoring explanations from only one frequency band is potentially misleading and is likely to contradict itself simply by changing hyperparameters (see Fig. 5) [27,33].", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67c2e0419ad1f1d36380c85575144747ffdd4dd7dc6e82bf1f4f862896522cfb.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 385, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53bc8bee8d3cd47038cd74474421e58ef512951a596f609f1b8f0f684b07c2b0.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 393, + 148, + 552, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad1af17a5be4a7a8c28d04ec78998923adf17d112e888a31077027331948287b.jpg", + "image_caption": [ + "(c)", + "Fig. 3: A 1D visualization of amplification (b) and mitigation (d) effects caused by gradient and perturbation, respectively. For simplicity, we have shown a rectangular pulse (c) convolved with $\\nabla f$ , but the kernel shape is usually more complex in practice, an example shown in Fig. 2. Moreover, functions in blue are the Fourier representation of the black functions. Due to the gradient operator, small high-frequency jitters in the decision boundary (a) amplify proportionally to their frequency (b). To remove the effect of high-frequency components of the decision boundary that usually dominate the explanations, researchers have proposed different perturbation kernels, visualized in (d). Theorem 2 corresponds to $(\\mathrm{a}\\rightarrow \\mathrm{b})$ and Theorem 1 corresponds to $(\\mathrm{b}\\rightarrow \\mathrm{d})$ ." + ], + "image_footnote": [], + "bbox": [ + 565, + 148, + 772, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We formalize this statement, also known as the Rashomon effect (see Appendix B for the proof).", + "bbox": [ + 212, + 449, + 784, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proposition 1. The SmoothGrad-Squared pixel attribution can contradict itself by changing the hyperparameters. Simply, the Rashomon effect can occur in the SmoothGrad-Squared explanations.", + "bbox": [ + 214, + 492, + 787, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuition. This proposition suggests a reason, i.e. removing different frequency bands, for inconsistencies we encounter in $\\mathrm{SG}^2$ explanations, which has been unknown so far [48]. Note that this occurs with different kernels, but we have chosen the one that leads to a simpler proof.", + "bbox": [ + 214, + 554, + 787, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "IntegratedGrad via a Rect Kernel. Setting the perturbation kernel to $p(\\tilde{\\boldsymbol{x}}) = \\frac{1}{\\sigma|\\boldsymbol{s} - \\boldsymbol{x}|}\\mathrm{Rect}(\\frac{\\tilde{\\boldsymbol{x}} - \\boldsymbol{x}}{\\sigma(\\boldsymbol{s} - \\boldsymbol{x})} -\\frac{1}{2})^2$ , where $s$ is a chosen baseline image, we recover the perturbation kernel of a well known method called IntegratedGrad (IG) [57]. We use the spectral representation of gradients in Eq. (2), for a proposition about the gradient signs.", + "bbox": [ + 214, + 638, + 787, + 715 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} [ \\nabla f (\\tilde {\\boldsymbol {x}}) ] \\propto \\int i \\widehat {f} (\\boldsymbol {\\omega}) \\sin (\\pi \\sigma (\\boldsymbol {s} - \\boldsymbol {x}) \\boldsymbol {\\omega}) e ^ {i \\pi \\boldsymbol {\\omega} \\boldsymbol {r}} d \\boldsymbol {\\omega} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 726, + 785, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $\\boldsymbol{r} = (1 + 2\\frac{\\boldsymbol{x}}{\\sigma(\\boldsymbol{s} - \\boldsymbol{x})})$ , appears because the kernel is not centered on the sample being explained. Note that the Fourier operator linearly expands along each dimension, but we use vector forms for notational simplicity. Furthermore, to", + "bbox": [ + 214, + 768, + 787, + 816 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^2$ Rect function is defined as $\\mathrm{Rect}(\\tilde{\\pmb{x}}) = \\frac{1}{2}\\mathrm{sign}(\\frac{1}{2} -|\\tilde{\\pmb{x}} |) + \\frac{1}{2}$", + "bbox": [ + 217, + 823, + 612, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "study the effect of elementwise multiplication on the explanation quantitatively, we use IG to denote integrated gradients without the elementwise multiplication with input and XIG to denote the version with elementwise multiplication.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In explanation visualizations, the negative gradient sign has been used to show a negative contribution to the prediction [7, 53], to our knowledge it has never been formally justified. In the next proposition, we state that the negative gradient sign can be the direct result of the chosen perturbation kernel, formally shown in Appendix B.", + "bbox": [ + 212, + 191, + 784, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 2. The gradient sign in $IG$ is affected by the perturbation level.", + "bbox": [ + 214, + 277, + 771, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This proposition entails that the negative sign in the gradient explainer does not necessarily reflect a negative contribution to the prediction, nor the positive sign necessarily reflects a positive contribution (see Figs. 13 and 14).", + "bbox": [ + 212, + 301, + 782, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark 4. Since we assume continuity wrt $\\sigma$ , in our analysis, by Bolzano's theorem [50], there are hyperparameters that set the attribution to zero. Therefore, the gradient explainer may not produce a useful ranking compared to that of the gradient squared.", + "bbox": [ + 214, + 356, + 784, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark 5. Proposition 2 also shows that adding explanations with different hyperparameters may cause destructive interference when the different values of $\\omega \\widehat{p} (\\omega)$ are out of phase. Hence, squared gradients is a more appropriate design choice, unless the use of phase information is planned as part of the method. This has been discovered experimentally [30] and confirmed in Sec. 5.", + "bbox": [ + 214, + 426, + 785, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Proposed Solutions Based on the Spectral Analysis", + "text_level": 1, + "bbox": [ + 214, + 523, + 746, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Finding an Optimal Perturbation Scale", + "text_level": 1, + "bbox": [ + 214, + 555, + 584, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that the squared gradients in spectral domain (Eq. (3)) can be formulated as an inner product between the scaled PSD of the perturbation distribution and the PSD of the classifier i.e. $\\mathbb{E}_{p(\\tilde{\\boldsymbol{x}})}\\left[(\\nabla f(\\tilde{\\boldsymbol{x}}))^2\\right] = 4\\pi^2\\langle \\mathrm{S}_f,\\| \\boldsymbol {\\omega}\\| ^2\\mathrm{S}_{\\sqrt{p}}\\rangle$ Assuming unimodality, a nonzero contribution, and that the norm $\\left\\| \\left\\| \\boldsymbol {\\omega}\\right\\| ^2\\mathrm{S}_{\\sqrt{p}}\\right\\|$ exists, then the cosine similarity is maximized at a certain perturbation scale that has the highest similarity with $\\mathrm{S}_f$ . A higher cosine similarity guarantees recovering more information about the classifier.", + "bbox": [ + 212, + 579, + 784, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Definition 2. The cosine similarity between the PSD of the classifier and that of the perturbation kernel given by", + "bbox": [ + 214, + 694, + 785, + 724 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S i m i l a r i t y} \\left(\\mathrm {S} _ {f}, \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}}\\right) = 4 \\pi^ {2} \\frac {\\left\\langle \\mathrm {S} _ {f} , \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\right\\rangle}{\\| \\mathrm {S} _ {f} \\| \\left\\| \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\right\\|} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 734, + 784, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This quantity can be used as a way to determine the optimal perturbation scale per pixel to maximize the information extracted with the squared gradients. Note that, the word \"optimal\" refers to extracting the most information by the inner product, whether deemed relevant by a human or not..", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For one of the perturbation kernels we have presented earlier in Sec. 3.3, we have optimized this quantity on multiple levels, such as dataset level Fig. 4, image level Tab. 1, and, pixel level Fig. 5d. Finally, we have experimentally verified the effectiveness of this approach in Sec. 5.", + "bbox": [ + 214, + 146, + 787, + 208 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Computing the Similarity of Classifier and Gaussian Kernel. To compute the similarity, we first need to evaluate the kernel norm as follows", + "bbox": [ + 214, + 224, + 785, + 258 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2} e ^ {- 8 \\pi^ {2} \\sigma^ {2} \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2}} \\right\\| = \\left(\\int \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {4} e ^ {- 1 6 \\pi^ {2} \\sigma^ {2} \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2}} d \\boldsymbol {\\omega}\\right) ^ {\\frac {1}{2}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 262, + 785, + 299 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\propto \\sigma^ {\\frac {- 5}{2}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 464, + 301, + 785, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Therefore, the cosine similarity for the Gaussian perturbation kernel can be computed as follows:", + "bbox": [ + 214, + 328, + 785, + 358 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S i m i l a r i t y} (\\mathrm {S} _ {\\sqrt {p}}, \\mathrm {S} _ {f}) \\propto \\sigma^ {\\frac {5}{2}} \\langle \\mathrm {S} _ {f}, \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\rangle\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 364, + 645, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Using the L1 norm for the squared gradients of flattened explanations. We have observed experimentally that $\\mathrm{S}_f$ depends on the input scale, hence we use a convex interpolation with noise to support the assumption that $\\mathrm{S}_f$ remains constant (see Appendix E.2 and Fig. 8). The results are shown in Fig. 4.", + "bbox": [ + 214, + 388, + 785, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Remark 6. The optimal perturbation scale, determined by maximizing cosine similarity, offers a standard method for selecting the perturbation scale in $\\mathrm{SG^2}$ . Given that user discretion in the selection of perturbation scale contributes to inconsistencies, as illustrated in Fig. 1, opting for the optimal perturbation scale mitigates these inconsistencies in the results. Hence, it represents a more suitable design choice. In Sec. 5 we demonstrate that this approach not only enhances $\\mathrm{SG^2}$ 's robustness against inconsistencies but also its adaptability to diverse datasets, which in turn can improve the quantitative performance.", + "bbox": [ + 214, + 455, + 787, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Aggregation of Information With Explanation Prior", + "text_level": 1, + "bbox": [ + 214, + 602, + 689, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have shown that explanations are influenced by specific frequency bands, rather than representing the importance of individual pixels for prediction. Instead, they indicate the contributions within a specific frequency range. This observation motivates us to investigate a method for aggregation of information across different bands, which closely resembles an ensemble of IG and SG [55,57] which we call SpectralLens.", + "bbox": [ + 214, + 623, + 785, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Definition 3. SpectralLens is an ensemble explanation method, that aggregates information across different frequency bands as follows", + "bbox": [ + 214, + 720, + 785, + 752 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {p (\\sigma)} \\left[ \\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] \\right] \\propto \\iint \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) \\| \\boldsymbol {\\omega} \\| ^ {2} e ^ {- 8 \\pi^ {2} \\sigma \\| \\boldsymbol {\\omega} \\| ^ {2}} d \\boldsymbol {\\omega} p (\\sigma) d \\sigma \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 757, + 785, + 789 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Where we refer to this method by $SL^2$ as it is using the squared gradients and we call $p(\\sigma)$ as an explanation prior, gauging the frequency bands that we choose to represent in an explanation.", + "bbox": [ + 214, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/58a578b8a285dba16384c1190133078cef1105bae86180afc2741924d8b891df.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 496, + 243 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ad50c71b764988b8360a2f65904a4624eee08e61389feaeff4d15c2c8584a998.jpg", + "image_caption": [ + "(b)", + "Fig. 4: Cosine similarity between kernel and classifier, defined in Eq. (6), is measured for the Gaussian perturbation kernel and ResNet50 averaged on the 10K images of the ImageNet validation set (a) and Food101 (b). This metric assesses the similarity between the PSD of the perturbation kernel and that of the classifier, indicating the perturbation level's effectiveness for kernels with finite norm in the Fourier domain. When using a Dirac explanation prior in Eq. (9), where a perturbation level is assigned per image or dataset, the optimal choice for the perturbation level is where the similarity is maximized across the image or dataset respectively. Note that this level of noise refers to a band-pass filter, shown in Fig. 2, which suggests the consideration of a latent distribution on the frequency of labeling, reflecting the workforce's mindset during data labeling." + ], + "image_footnote": [], + "bbox": [ + 508, + 148, + 777, + 243 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Remark 7. $\\mathrm{SL}^2$ exhibits robustness against inconsistencies in $\\mathrm{SG}^2$ due to its aggregation of information from various bands. Therefore, it is a preferable design choice not to employ a Dirac explanation prior with $SG^2$ , as it eliminates inconsistencies, depicted in Fig. 1. Nevertheless, as demonstrated in the Sec. 5, this approach may enhance the quantitative performance of explanations.", + "bbox": [ + 212, + 455, + 787, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Intuition. Considering Eq. (4), we realize that the combination of perturbation and gradient explainer creates overlapping band-pass carrier waves that modulate the amplitude of the spectral density of the classifier. Therefore, a new summary statistic is required that can reflect a wide range of frequency bands. We explicitly identified the spectral properties of the explanations, recognizing that not all frequencies are equally desired. Explanations with very high frequencies primarily attribute predictions to pixels. Therefore, we utilize an explanation prior $p(\\sigma)$ for a nuanced control of the spectral properties of explanations. Although there are many possible alternatives for selecting an explanation prior, we used $p(\\sigma) = \\mathrm{Unif}(0,1)$ for simplicity.", + "bbox": [ + 212, + 547, + 787, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Definition 4. We define $\\mathbf{ArgLens}(AL)$ as the point at which the $SG^2$ is maximized for pixel $x_i$ . Formally defined as", + "bbox": [ + 212, + 709, + 785, + 741 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A r g L e n s} \\left(x _ {i}\\right) = \\arg \\max _ {\\sigma} \\left\\{\\mathrm {S G} ^ {2} \\left(x _ {i}\\right) \\right\\} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 750, + 785, + 773 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\sigma$ is the hyperparameter of the perturbation kernel.", + "bbox": [ + 215, + 782, + 627, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AL shows the frequency band at which the pixel's attribution is maximized. Alternatively, it can be interpreted as the robustness of a pixel to noise. We", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "have visualized the AL for the Gaussian kernel in Fig. 5 and discussed the way it helps the interpretation of explanations in Sec. 4.2.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The aggregation of explanations from different frequencies is visualized in Fig. 5. Moreover, the basic explanation methods are obtainable with a Dirac explanation prior $p(\\sigma) = \\delta(\\sigma)$ . Therefore, it can be used as a unified formulation for gradient-based explanation methods in spectral domain.", + "bbox": [ + 212, + 176, + 787, + 238 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Interesting Properties of SpectralLens and ArgLens. In the absence of a robust mathematical framework, the interpretation of explanations may be subjective. Thus, explanation methods must be grounded in solid theoretical principles. $\\mathrm{SL}^2$ addresses potential inconsistencies by aggregating attributions across frequency bands, ensuring fair representation of all pixels' contributions.", + "bbox": [ + 212, + 258, + 787, + 333 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As $\\mathrm{SG^2}$ results in bands-pass filters (see Sec. 3.3), its attributions reflect the changes in the decision boundary within a certain frequency band. Furthermore, AL reveals the frequency band at which the attribution is maximized. Hence, lower AL indicates that the attributions are likely governed by high-frequency components of the decision boundary, more sensitive to input noise, consequently less robust. An ideal classifier with smooth decision boundaries, i.e. robust features, leads to high AL everywhere, with a high intensity in $\\mathrm{SL^2}$ on relevant pixels and low intensity on irrelevant pixels.", + "bbox": [ + 212, + 334, + 787, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Given the different purposes of pixels within an image, the frequency of activation may vary and can be visualized via AL. Since $\\mathrm{SL}^2$ and AL offer complementary information, they should be interpreted together. This approach offers a deeper understanding compared to a simplistic perspective, as it has been shown by Rahaman et al. [43] that high-frequency regions emerge later in training, are more challenging to fit, and are prone to change. On the contrary, low-frequency components of the decision boundary manifest earlier in training, are easier to fit, and exhibit greater stability with respect to noise.", + "bbox": [ + 212, + 455, + 787, + 574 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "It is important to clarify that \"high-frequency\" refers to features considered high-frequency by the model, and not necessarily frequency across spatial dimensions (e.g. width, height). Nonetheless, empirical findings demonstrate that high-frequency features in spatial domain are more likely to be identified as high frequency by the model (see Appendix I).", + "bbox": [ + 212, + 575, + 787, + 650 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 672, + 356, + 690 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Given the experimental discoveries that squared gradients lead to better results [30], we evaluate both scenarios for different hyperparameters and compare them with SL in Tabs. 1 and 2, refer to Appendix E for implementation details and reproducibility. For the purposes of this study, we use pixel removal strategy, despite its problems [30]. The spectral analysis of pixel removal strategy and possible solutions are deferred to future research.", + "bbox": [ + 212, + 704, + 787, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We report single-pixel most-important-first (MoRF) insertion and deletion scores of the methods to assess the design choices previously proposed by our theoretical framework: (1) utilizing squared gradients instead of gradients in", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/748a8d9134f66ebad795cb6170cdd6324a4fdf8226acf661c91c6173a4a42d83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 145, + 357, + 250 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1bf4a60f123842bd0e33e7d4229d36537730d83575dba59526babc51101c6d06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 145, + 498, + 250 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/772a0ff07b8a7e44411c06fca752f52abb2b569a465fee1bb7ca4c577914c191.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 638, + 250 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f61bc8e42c32c6a0ce10e697499b84b7284c279125c24493aa109e49a85ac190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 146, + 779, + 250 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cc40f30369167a69d73bdea5a27af10b38294dbbbd1c1455353f9d1b9ae6bd87.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 222, + 251, + 357, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d707b0fbd240bcc7fda17dcfbe7b08a4d9531c03ec08fe624b9e2d35a2aaeb4d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 364, + 252, + 498, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1a932a9c1aeb327b28a9292702767225f034d638bae2be886111fe4f8591b6a8.jpg", + "image_caption": [ + "(c)", + "Fig. 5: This figure shows example visualizations of explanations generated by $\\mathrm{SG^2}$ implementing design decisions suggested in our work in Remarks 6 and 7. (b) shows the $\\mathrm{SG^2}$ visualized at the maximum similarity according to Fig. 4, for this image $\\sigma = 0.5$ . (c) shows the SL attributions. Lastly, (d) shows the frequencies obtained from AL (defined in Eq. (10)), which correspond to the perturbation level at which $\\mathrm{SG^2}$ is maximized. Since we are interested in frequencies rather than the actual perturbation levels, we visualize 2 stds above and below the mean of frequencies obtained by $\\omega = 1 / (1 + \\mathrm{AL})$ (see Appendix I for more examples). The heatmaps in (b) and (c) show how much the model is relying on the pixel, while (d) shows the frequency at which the feature's contribution maximizes. In the provided examples, we can see that the model is relying on the nose (top) and eye (bottom) while the same region is of a lower frequency compared to the head. This can be interpreted as the model finding the nose and eye more stable features compared to the head." + ], + "image_footnote": [], + "bbox": [ + 504, + 252, + 638, + 354 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/961e15fc803ece74636792d3c11d15aa26594b32db6c08462fd39ef4337e4f08.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 643, + 252, + 779, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Remark 5, (2) finding a standard perturbation scale in Remark 6, and (3) employing a uniform explanation prior instead of Dirac, stated in Remark 7 (see Fig. 7 for an ablation study of kernel variance and Appendix G for an ablation study of patch removal instead of single pixel removal). We evaluated the suggested design choices on ImageNet [20] and Food101 [1] in Tab. 1. However, smaller datasets, such as CIFAR, are omitted because their resolution restricts the potential gap between high- and low-frequency features.", + "bbox": [ + 212, + 590, + 787, + 699 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Limitations of Our Framework. The choice of independent Gaussian perturbation is limited due to its inability to incorporate prior knowledge about the input space or data manifold characteristics, and other data modalities such as graphs and time series. However, preferring simpler analytical forms within our framework limits the alternative choices. Efficiency is another challenge compared to basic methods, as performing back-props at various noise scales is costly. Moreover, our attribution of frequencies in visualizations, such as SL, may lack the desired generalizability across data points, as high-frequency re", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/785c663824415506b622e5327b8b9af3375350e1363b63664bc95e201293ddff.jpg", + "table_caption": [ + "Table 1: This table presents the assessment of the proposed design choices in Remarks 5 to 7 across different methods, highlighting statistically significant improvements with bold fonts. We conduct t-tests comparing gradient methods to their squared-gradient counterparts to evaluate the design choice in Remark 5. Furthermore, we conduct t-tests between $\\mathrm{SG}^2$ with the original heuristic and those ensuring consistency to evaluate design decisions in Remarks 6 and 7. Overlines and underlines denote highest or lowest values within columns, respectively. Although there are no squared versions of Occlusion Sensitivity (OS) [5] and RISE [42], they are included as low-frequency baselines. While the dimensions of boolean masks utilized in OS and RISE can implicitly regulate the sampling variance (see Appendix C), the exponential increase in the required number of samples in high-frequency explanations renders them less practical. SG scores follow the heuristic of the SmoothGrad paper [55] i.e. $\\sigma = (\\max(x) - \\min(x))$ 0.1. XIG indicates the input multiplication in the original implementation of IntegratedGrad [57], while IG does not apply such transformation. Rankings for methods using gradients are obtained from negative gradients, including RISE and Occlusion Sensitivity, according to Eq. (2). Furthermore, $\\mathrm{SG}_{\\mathrm{Opt.}}^2$ uses an optimized kernel per image and $\\mathrm{SL}^2$ uses a uniform explanation prior. Comparing SG² and $\\mathrm{SG}_{\\mathrm{Opt.}}^2$ , we can see that $\\mathrm{SG}_{\\mathrm{Opt.}}^2$ can adapt to the statistics of the Food101. Therefore, $\\mathrm{SG}_{\\mathrm{Opt.}}^2$ , while not subject to inconsistencies (shown in Fig. 1), can improve the performance due to its adaptability. Hence, we can ensure the consistency of SG² without compromising its performance. Also, note that the input multiplication enhances visual quality, but degrades the performance." + ], + "table_footnote": [], + "table_body": "
ImageNetFood101
Insertion(↑)Deletion(↓)Insertion(↑)Deletion(↓)
GradientsOS0.384 ± 0.0080.237 ± 0.0100.378 ± 0.0110.209 ± 0.009
RISE0.494 ± 0.0140.174 ± 0.0070.410 ± 0.0130.198 ± 0.007
SG0.212 ± 0.0100.179 ± 0.0090.179 ± 0.0080.137 ± 0.007
XIG0.208 ± 0.0090.204 ± 0.0080.101 ± 0.0070.177 ± 0.007
IG0.201 ± 0.0100.182 ± 0.0090.145 ± 0.0060.131 ± 0.005
Gradients2SG20.465 ± 0.0130.157 ± 0.0070.371 ± 0.0120.162 ± 0.006
XIG20.334 ± 0.0140.187 ± 0.0080.204 ± 0.0100.263 ± 0.007
IG20.340 ± 0.0120.168 ± 0.0070.320 ± 0.0090.135 ± 0.005
Gradients2Opt.0.491 ± 0.0150.187 ± 0.0070.394 ± 0.0130.167 ± 0.006
SL20.493 ± 0.0160.169 ± 0.0070.395 ± 0.0140.157 ± 0.007
", + "bbox": [ + 269, + 460, + 733, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "gions in one image may correspond to low-frequency regions in another image (see Appendix I). Identifying generalizable attributions across images remains a subject for future investigation.", + "bbox": [ + 212, + 691, + 787, + 738 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Related Works", + "text_level": 1, + "bbox": [ + 214, + 761, + 397, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Probabilistic and Spectral View of Explanations. Despite sporadic attempts to analyze explanations from a probabilistic view [17, 59], this view is more common in the literature around the uncertainty quantification of expla", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "nations [29,40,54]. Our work is closely related to Covert et al. [19], if their work is seen from a probabilistic perspective. Aside from the works around the spectral view of deep networks [10,11,15,43,58], as of our current knowledge, there is no work adopting a spectral view of explanations along pixels' value. Nonetheless, there are works that build on a spectral view along spatial dimensions [34,35], extending an optimization-based definition of explanations [22].", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Unifying Views to Explanations. Covert et al. [19] propose an information theoretic framework that unifies several explanation methods, focusing on three design choices: (1) feature removal method, (2) which model behavior to analyze, and (3) summary statistics. Our probabilistic framework, being similar to Covert's, offers a distinct viewpoint relevant to our spectral analysis. It can be used as a unifying view of explanations (see Appendix C), inheriting from Covert's framework. Notably, our work aligns with the broader context of research under finding unifying representations of explanations [3,26,37].", + "bbox": [ + 212, + 258, + 789, + 381 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Input-space or Function-space Perturbations. Methods such as NoiseGrad [14], Grad-CAM [51], RectifiedGrad [32], Layer-wise Relevance Propagation [7], DeepLift [53], and even Sharpness-Aware Minimization (SAM) [24] employ function space perturbations. Utilizing such distributions offers data modality independence, whereas input space perturbation remains agnostic about the model architecture. While acknowledging the duality between function and input space distributions, our focus on explicit input space distributions excludes these methods from our study. Although we recognize the utility of integrating function space perturbation into our mathematical model, we defer this aspect to future investigations. It can be noted, however, that extending this work to function space perturbation will connect it to the literature on double descent and smooth interpolation of samples [25, 41].", + "bbox": [ + 212, + 402, + 789, + 587 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 607, + 359, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we deepen understanding of pixel attribution methodologies by rigorously analyzing their theoretical underpinnings, revealing a common spectral bias. Our analysis supports the experimental finding of recent works regarding the efficacy of using squared gradients. Moreover, we demonstrate how spectral properties can lead to inconsistent explanations and propose two design choices informed by our theoretical framework to address these inconsistencies: standardizing hyperparameters based on cosine similarity between the perturbation kernel and classifier, and aggregating explanations with a given prior. Despite the effectiveness of our approach, we acknowledge limitations and suggest avenues for further research. Lastly, we assess the suggested design choices using deletion and insertion scores.", + "bbox": [ + 212, + 638, + 787, + 806 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank Giovanni Luca Marchetti for an early review of our work and his kind feedback. This project is partially supported by Region Stockholm through MedTechLabs, and Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Scientific computation was enabled by the supercomputing resource Berzelius, provided by the National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg foundation.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Food-101 - Mining Discriminative Components with Random Forests, https://data.vision.ee.ethz.ch/cvl/datasets-extra/food-101/", + "2. Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., Kim, B.: Sanity Checks for Saliency Maps (Nov 2020). https://doi.org/10.48550/arXiv.1810.03292, http://arxiv.org/abs/1810.03292, arXiv:1810.03292 [cs, stat]", + "3. Agarwal, S., Jabbari, S., Agarwal, C., Upadhyay, S., Wu, Z.S., Lakkaraju, H.: Towards the Unification and Robustness of Perturbation and Gradient Based Explanations (Jul 2021). https://doi.org/10.48550/arXiv.2102.10618, http://arxiv.org/abs/2102.10618, arXiv:2102.10618 [cs]", + "4. Alvarez-Melis, D., Jaakkola, T.S.: On the Robustness of Interpretability Methods (Jun 2018). https://doi.org/10.48550/arXiv.1806.08049, http://arxiv.org/abs/1806.08049, arXiv:1806.08049 [cs, stat]", + "5. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for Deep Neural Networks (Mar 2018). https://doi.org/10.48550/arXiv.1711.06104, http://arxiv.org/abs/1711.06104, arXiv:1711.06104 [cs, stat]", + "6. Arrieta, A.B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., Chatila, R., Herrera, F.: Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI (Dec 2019). https://doi.org/10.48550/arXiv.1910.10045, http://arxiv.org/abs/1910.10045, arXiv:1910.10045 [cs]", + "7. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS ONE 10(7), e0130140 (Jul 2015). https://doi.org/10.1371/journal.pone.0130140, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0130140", + "8. Balduzzi, D., Frean, M., Leary, L., Lewis, J.P., Ma, K.W.D., McWilliams, B.: The Shattered Gradients Problem: If resnets are the answer, then what is the question? (Jun 2018). https://doi.org/10.48550/arXiv.1702.08591, http://arxiv.org/abs/1702.08591, arXiv:1702.08591 [cs, stat]", + "9. Bansal, N., Agarwal, C., Nguyen, A.: SAM: The Sensitivity of Attribution Methods to Hyperparameters (Apr 2020). https://doi.org/10.48550/arXiv.2003.08754, http://arxiv.org/abs/2003.08754, arXiv:2003.08754 [cs]", + "10. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency Bias in Neural Networks for Input of Non-Uniform Density (Mar 2020), https://api-semanticscholar.org/CorpusID:212644664" + ], + "bbox": [ + 218, + 339, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Benbarka, N., Hofer, T., Ul-Moqeet Riaz, H., Zell, A.: Seeing Implicit Neural Representations as Fourier Series. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 2283-2292 (Jan 2022). https://doi.org/10.1109/WACV51458.2022.00234, https://ieeexplore.ieee.org/document/9706795/", + "12. Blackman, R.B., Tukey, J.W.: The measurement of power spectra from the point of view of communications engineering — Part I. The Bell System Technical Journal 37(1), 185–282 (Jan 1958). https://doi.org/10.1002/j.1538-7305.1958.tb03874.x, http://ieeexplore.ieee.org/document/6768513/", + "13. Brughmans, D., Melis, L., Martens, D.: Disagreement amongst counterfactual explanations: How transparency can be deceptive (Apr 2023). https://doi.org/10.48550/arXiv.2304.12667, http://arxiv.org/abs/2304.12667 [cs]", + "14. Bykov, K., Hedström, A., Nakajima, S., Höhne, M.M.C.: NoiseGrad: Enhancing Explanations by Introducing Stochasticity to Model Weights (May 2022). https://doi.org/10.48550/arXiv.2106.10185, http://arxiv.org/abs/2106.10185, arXiv:2106.10185 [cs]", + "15. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards Understanding the Spectral Bias of Deep Learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. pp. 2205-2211. International Joint Conferences on Artificial Intelligence Organization, Montreal, Canada (Aug 2021). https://doi.org/10.24963/ijcai.2021/304, https://www.ijcai.org/proceedings/2021/304", + "16. Chen, H., Janizek, J.D., Lundberg, S., Lee, S.I.: True to the Model or True to the Data? (Jun 2020). https://doi.org/10.48550/arXiv.2006.16234, http://arxiv.org/abs/2006.16234, arXiv:2006.16234 [cs, stat]", + "17. Chen, J., Song, L., Wainwright, M.J., Jordan, M.I.: Learning to Explain: An Information-Theoretic Perspective on Model Interpretation (Jun 2018). https://doi.org/10.48550/arXiv.1802.07814, http://arxiv.org/abs/1802.07814, arXiv:1802.07814 [cs, stat]", + "18. Cohen, L.: The generalization of the Wiener-Khinchin theorem. In: Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181). vol. 3, pp. 1577-1580. IEEE (1998), https://ieeexplore.ieee.org/abstract/document/681753/", + "19. Covert, I., Lundberg, S., Lee, S.I.: Explaining by Removing: A Unified Framework for Model Explanation (May 2022). https://doi.org/10.48550/arXiv.2011.14878, http://arxiv.org/abs/2011.14878, arXiv:2011.14878 [cs, stat]", + "20. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (Jun 2009). https://doi.org/10.1109/CVPR.2009.5206848, https://ieeexplore.ieee.org/document/5206848, iSSN: 1063-6919", + "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Jun 2021). https://doi.org/10.48550/arXiv.2010.11929, http://arxiv.org/abs/2010.11929, arXiv:2010.11929 [cs]", + "22. Fong, R., Patrick, M., Vedaldi, A.: Understanding Deep Networks via Extremal Perturbations and Smooth Masks (Oct 2019). https://doi.org/10.48550/arXiv.1910.08485, http://arxiv.org/abs/1910.08485, arXiv:1910.08485 [cs, stat]" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Fong, R., Vedaldi, A.: Interpretable Explanations of Black Boxes by Meaningful Perturbation. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 3449-3457 (Oct 2017). https://doi.org/10.1109/ICCV.2017.371, http://arxiv.org/abs/1704.03296, arXiv:1704.03296 [cs, stat]", + "24. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-Aware Minimization for Efficiently Improving Generalization (Apr 2021). https://doi.org/10.48550/arXiv.2010.01412, http://arxiv.org/abs/2010.01412, arXiv:2010.01412 [cs, stat]", + "25. Gamba, M., Azizpour, H., Björkman, M.: On the Lipschitz Constant of Deep Networks and Double Descent (Nov 2023). https://doi.org/10.48550/arXiv.2301.12309, http://arxiv.org/abs/2301.12309, arXiv:2301.12309 [cs]", + "26. Han, T., Srinivas, S., Lakkaraju, H.: Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations (Dec 2022). https://doi.org/10.48550/arXiv.2206.01254, http://arxiv.org/abs/2206.01254, arXiv:2206.01254 [cs]", + "27. Harel, N., Gilad-Bachrach, R., Obolski, U.: Inherent Inconsistencies of Feature Importance (Jun 2022). https://doi.org/10.48550/arXiv.2206.08204, http://arxiv.org/abs/2206.08204, arXiv:2206.08204 [cs]", + "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition (Dec 2015). https://doi.org/10.48550/arXiv.1512.03385, http://arxiv.org/abs/1512.03385, arXiv:1512.03385 [cs]", + "29. Hill, D., Masoomi, A., Ghimire, S., Torop, M., Dy, J.: Explanation Uncertainty with Decision Boundary Awareness (Nov 2022). https://doi.org/10.48550/arXiv.2210.02419, http://arxiv.org/abs/2210.02419, arXiv:2210.02419 [cs]", + "30. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A Benchmark for Interpretability Methods in Deep Neural Networks (Nov 2019). https://doi.org/10.48550/arXiv.1806.10758, http://arxiv.org/abs/1806.10758, arXiv:1806.10758 [cs, stat]", + "31. Hou, X., Zhang, L.: Saliency Detection: A Spectral Residual Approach. In: 2007 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8 (Jun 2007). https://doi.org/10.1109/CVPR.2007.383267, https://ieeexplore.ieee.org/document/4270292, iSSN: 1063-6919", + "32. Kim, B., Seo, J., Jeon, S., Koo, J., Choe, J., Jeon, T.: Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps (Sep 2019). https://doi.org/10.48550/arXiv.1902.04893, http://arxiv.org/abs/1902.04893, arXiv:1902.04893 [cs, stat]", + "33. Kindermans, P.J., Hooker, S., Adebayo, J., Alber, M., Schütt, K.T., Dähne, S., Erhan, D., Kim, B.: The (Un)reliability of saliency methods (Nov 2017). https://doi.org/10.48550/arXiv.1711.00867, http://arxiv.org/abs/1711.00867, arXiv:1711.00867 [cs, stat]", + "34. Kolek, S., Nguyen, D.A., Levie, R., Bruna, J., Kutyniok, G.: Cartoon Explanations of Image Classifiers (Oct 2022). https://doi.org/10.48550/arXiv.2110.03485, http://arxiv.org/abs/2110.03485, arXiv:2110.03485 [cs]", + "35. Kolek, S., Windsesheim, R., Loarca, H.A., Kutyniok, G., Levie, R.: Explaining Image Classifiers with Multiscale Directional Image Representation (Apr 2023). https://doi.org/10.48550/arXiv.2211.12857, http://arxiv.org/abs/2211.12857, arXiv:2211.12857 [cs]", + "36. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The Disagreement Problem in Explainable Machine Learning: A Practitioner's Perspective (Feb 2022). https://doi.org/10.48550/arXiv.2202.01602, http://arxiv.org/abs/2202.01602, arXiv:2202.01602 [cs]" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)", + "38. Marmarelis, P.Z., Marmarelis, V.Z.: The White-Noise Method in System Identification. In: Marmarelis, P.Z., Marmarelis, V.Z. (eds.) Analysis of Physiological Systems: The White-Noise Approach, pp. 131-180. Computers in Biology and Medicine, Springer US, Boston, MA (1978). https://doi.org/10.1007/978-1-4613-3970-0_4, https://doi.org/10.1007/978-1-4613-3970-0_4", + "39. Marques-Silva, J., Ignatiev, A.: Delivering Trustworthy AI through Formal XAI. Proceedings of the AAAI Conference on Artificial Intelligence 36(11), 12342-12350 (Jun 2022). https://doi.org/10.1609/aaai.v36i11.21499, https://ojs.aaai.org/index.php/AAAI/article/view/21499", + "40. Marx, C., Park, Y., Hasson, H., Wang, Y., Ermon, S., Huan, L.: But Are You Sure? An Uncertainty-Aware Perspective on Explainable AI. In: Proceedings of The 26th International Conference on Artificial Intelligence and Statistics. pp. 7375-7391. PMLR (Apr 2023), https://proceedings.mlr.press/v206/marx23a.htm1", + "41. Nakkiran, P., Kaplun, G., Bansal, Y., Yang, T., Barak, B., Sutskever, I.: Deep Double Descent: Where Bigger Models and More Data Hurt (Dec 2019). https://doi.org/10.48550/arXiv.1912.02292, http://arxiv.org/abs/1912.02292, arXiv:1912.02292 [cs, stat]", + "42. Petsiuk, V., Das, A., Saenko, K.: RISE: Randomized Input Sampling for Explanation of Black-box Models (Sep 2018). https://doi.org/10.48550/arXiv.1806.07421, http://arxiv.org/abs/1806.07421, arXiv:1806.07421 [cs]", + "43. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F.A., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks (May 2019). https://doi.org/10.48550/arXiv.1806.08734, http://arxiv.org/abs/1806.08734, arXiv:1806.08734 [cs, stat]", + "44. Rahman, M.: Applications of Fourier Transforms to Generalized Functions. WIT Press (2011), google-Books-ID: k_rdcKaUdr4C", + "45. Ras, G., van Gerven, M., Haselager, P.: Explanation Methods in Deep Learning: Users, Values, Concerns and Challenges. In: Escalante, H.J., Escalera, S., Guyon, I., Baró, X., Güçlütürk, Y., Güçlü, U., van Gerven, M. (eds.) Explainable and Interpretable Models in Computer Vision and Machine Learning, pp. 19-36. The Springer Series on Challenges in Machine Learning, Springer International Publishing, Cham (2018). https://doi.org/10.1007/978-3-319-98131-4_2, https://doi.org/10.1007/978-3-319-98131-4_2", + "46. Reiter, H., Stegeman, J.D.: Classical Harmonic Analysis and Locally Compact Groups. Oxford University Press (Nov 2000). https://doi.org/10.1093/oso/9780198511892.001.0001, https://academic.oup.com/book/54460", + "47. Ribeiro, M.T., Singh, S., Guestrin, C.: \"Why Should I Trust You?\": Explaining the Predictions of Any Classifier (Aug 2016). https://doi.org/10.48550/arXiv.1602.04938, http://arxiv.org/abs/1602.04938, arXiv:1602.04938 [cs, stat]", + "48. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A Consistent and Efficient Evaluation Strategy for Attribution Methods (Jun 2022). https://doi.org/10.48550/arXiv.2202.00449, http://arxiv.org/abs/2202.00449, arXiv:2202.00449 [cs]", + "49. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence 1(5), 206-215 (May 2019). https://doi.org/10.1038/s42256-019-0048-x, https://www.nature.com/articles/s42256-019-0048-x", + "50. Rudin, W.: Principles of Mathematical Analysis. McGraw-Hill (1976), google-Books-ID: kwqzPAAACAAJ" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Mehrpanah et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision 128(2), 336-359 (Feb 2020). https://doi.org/10.1007/s11263-019-01228-7, http://arxiv.org/abs/1610.02391, arXiv:1610.02391 [cs]", + "52. Serov, V.: The Riemann-Lebesgue Lemma. In: Serov, V. (ed.) Fourier Series, Fourier Transform and Their Applications to Mathematical Physics, pp. 33-35. Applied Mathematical Sciences, Springer International Publishing, Cham (2017). https://doi.org/10.1007/978-3-319-65262-7_6, https://doi.org/10.1007/978-3-319-65262-7_6", + "53. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not Just a Black Box: Learning Important Features Through Propagating Activation Differences (Apr 2017). https://doi.org/10.48550/arXiv.1605.01713, http://arxiv.org/abs/1605.01713, arXiv:1605.01713 [cs]", + "54. Slack, D., Hilgard, S., Singh, S., Lakkaraju, H.: Reliable Post hoc Explanations: Modeling Uncertainty in Explainability (Nov 2021). https://doi.org/10.48550/arXiv.2008.05030, http://arxiv.org/abs/2008.05030, arXiv:2008.05030 [cs, stat]", + "55. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: SmoothGrad: removing noise by adding noise (Jun 2017). https://doi.org/10.48550/arXiv.1706.03825, http://arxiv.org/abs/1706.03825, arXiv:1706.03825 [cs, stat]", + "56. Stoica, P., Moses, R.L.: Spectral analysis of signals, vol. 452. Pearson Prentice Hall Upper Saddle River, NJ (2005), http://user.it.uu.se/~ps/SAS-new.pdf", + "57. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic Attribution for Deep Networks (Jun 2017). https://doi.org/10.48550/arXiv.1703.01365, http://arxiv.org/abs/1703.01365, arXiv:1703.01365 [cs]", + "58. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. ArXiv (Jun 2020), https://api_semanticscholar.org/CorpusID:219791950", + "59. Wang, E., Khosravi, P., Broeck, G.V.d.: Probabilistic Sufficient Explanations (May 2021). https://doi.org/10.48550/arXiv.2105.10118, http://arxiv.org/abs/2105.10118, arXiv:2105.10118 [cs]", + "60. Wang, Y., Wang, X.: \"Why Not Other Classes?\": Towards Class-Contrastive Backpropagation Explanations. Advances in Neural Information Processing Systems 35, 9085-9097 (Dec 2022), https://proceedings.neurips.cc/paper_files/paper/2022/bitical/3b7a66b2d1258e892c89f485b8f896e0-AAbstract-Conference.html", + "61. Wilming, R., Kieslich, L., Clark, B., Haufe, S.: Theoretical Behavior of XAI Methods in the Presence of Suppressor Variables (Jun 2023). https://doi.org/10.48550/arXiv.2306.01464, http://arxiv.org/abs/2306.01464 [cs, stat]", + "62. Zeiler, M.D., Fergus, R.: Visualizing and Understanding Convolutional Networks (Nov 2013). https://doi.org/10.48550/arXiv.1311.2901, http://arxiv.org/abs/1311.2901, arXiv:1311.2901 [cs]" + ], + "bbox": [ + 212, + 146, + 787, + 757 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Spectral Properties of Gradient-based Explanation Methods", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_model.json b/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a09cc02185328e249e75b93a6687422d56e18de2 --- /dev/null +++ b/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_model.json @@ -0,0 +1,2801 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.266, + 0.141, + 0.74, + 0.186 + ], + "angle": 0, + "content": "On Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.212, + 0.733, + 0.228 + ], + "angle": 0, + "content": "Amir Mehrpanah1, Erik Englesson1, and Hossein Azizpour1" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.24, + 0.782, + 0.269 + ], + "angle": 0, + "content": "Department of Computer Science, Royal Institute of Technology, Stockholm, Sweden {amirme,engless,azizpour}@kth.se" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.299, + 0.74, + 0.506 + ], + "angle": 0, + "content": "Abstract. Understanding the behavior of deep networks is crucial to increase our confidence in their results. Despite an extensive body of work for explaining their predictions, researchers have faced reliability issues, which can be attributed to insufficient formalism. In our research, we adopt novel probabilistic and spectral perspectives to formally analyze explanation methods. Our study reveals a pervasive spectral bias stemming from the use of gradient, and sheds light on some common design choices that have been discovered experimentally, in particular, the use of squared gradient and input perturbation. We further characterize how the choice of perturbation hyperparameters in explanation methods, such as SmoothGrad, can lead to inconsistent explanations and introduce two remedies based on our proposed formalism: (i) a mechanism to determine a standard perturbation scale, and (ii) an aggregation method which we call SpectralLens. Finally, we substantiate our theoretical results through quantitative evaluations." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.52, + 0.74, + 0.562 + ], + "angle": 0, + "content": "Keywords: Probabilistic Machine Learning \\(\\cdot\\) Gradient-based Explanation Methods \\(\\cdot\\) Probabilistic Pixel Attribution Techniques \\(\\cdot\\) Explainability \\(\\cdot\\) Deep Neural Networks \\(\\cdot\\) Spectral Analysis" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.585, + 0.377, + 0.601 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Ensuring the interpretability of machine learning models is crucial for user's trust and understanding [2, 6, 45]. While considerable efforts have been dedicated to advancing explainability in the field [37, 47, 49, 51, 57, 62], there is a notable gap in systematically formalizing and consolidating existing knowledge [5, 9, 16, 19]. Insufficient formal analysis leaves too much room for different interpretations of the explanations, making them unreliable [27] which is detrimental to the very use case of explanations, i.e., human-critical applications. Therefore, more attention is essential to retrospectively analyze and structure the wealth of work done, creating a cohesive foundation for explainability [39, 61]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "One main reason for the unreliability of explanation methods is their susceptibility to hyperparameters, leading to inconsistent explanations [4, 13, 33, 36, 48], see Fig. 1. Among the several works that formally study explanation methods [19, 37, 57], to our knowledge, none shed light on such inconsistencies. Therefore, in this work, we adopt a novel probabilistic and spectral perspective to characterize the observed inconsistencies and propose preliminary solutions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.359, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.253, + 0.302, + 0.264 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.146, + 0.499, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.253, + 0.442, + 0.264 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.146, + 0.639, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.253, + 0.582, + 0.265 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.146, + 0.78, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.253, + 0.724, + 0.265 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 1: The figure illustrates sample explanations generated by varying hyperparameters of SmoothGrad-Squared [30, 55] for ResNet-50 on ImageNet (see Fig. 16 in the Appendix for more examples). It is evident from the figure that using different hyperparameters results in diverse explanations. Such variations have been demonstrated in recent studies for other explanation techniques as well [4, 13, 33, 48]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.378, + 0.687, + 0.394 + ], + "angle": 0, + "content": "Our Contribution. We iterate our contributions in three parts:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.408, + 0.787, + 0.451 + ], + "angle": 0, + "content": "- Analysis: We introduce a probabilistic representation for post-hoc gradient-based explanation methods in Sec. 2.1, and perform a spectral analysis of them in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.455, + 0.787, + 0.56 + ], + "angle": 0, + "content": "- Identified Issues: Our examination reveals a shared spectral bias stemming from the fundamental assumption of utilizing gradients to measure feature contributions. We show that the combined spectral effects of gradient and perturbation lead to band-pass filters modulating the explanations, which could lead to inconsistencies in the explanations, discussed in Sec. 3.2. Moreover, we conduct a spectral analysis of a few specific perturbation distributions that have been proposed in the literature in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.561, + 0.787, + 0.666 + ], + "angle": 0, + "content": "- Solutions and Evaluations: To address inconsistencies, we investigate two straightforward solutions that are finding an optimal perturbation scale based on the cosine similarity of the perturbation and the classifier in the frequency domain in Sec. 4.1, and an aggregation mechanism named SpectralLens in Sec. 4.2. Finally, we evaluate the proposed methods using the pixel removal strategy [48], acknowledging their effectiveness and limitations under Sec. 5." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.408, + 0.787, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.37, + 0.726 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Notation. We consider a classification setting, that is, modeling \\( p(y|\\pmb{x}) \\) for labels \\( y \\in \\{1,\\dots,k\\} \\) and \\( \\pmb{x} \\in \\mathbb{R}^n \\) the set of input images. We denote the \\( \\log p(y|\\pmb{x}) \\) by \\( f \\). We denote standard Fourier bases by \\( \\{\\psi(\\pmb{\\omega},\\pmb{x})\\}_{\\pmb{\\omega} \\in \\mathbb{R}^n} \\), according to which, under some regularity conditions [43], \\( f \\) can be decomposed as \\( f(\\pmb{x}) = \\int \\widehat{f}(\\pmb{\\omega}) \\psi(\\pmb{\\omega},\\pmb{x}) d\\pmb{\\omega} \\). In this representation, \\( \\widehat{f} \\) denotes the Fourier transform (FT) of \\( f \\). We denote the power spectral density (PSD) of \\( f \\) by \\( S_f(\\pmb{\\omega}) = |\\widehat{f}(\\pmb{\\omega})|^2 \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.147, + 0.652, + 0.162 + ], + "angle": 0, + "content": "2.1 A Probabilistic Representation of Explanations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.414 + ], + "angle": 0, + "content": "In this view, each explanation method consists of three components: (1) a distribution for perturbation of the input, (2) an explainer, and (3) a summary statistic. We use \\( p(\\tilde{\\pmb{x}} | \\pmb{x}, \\sigma) \\) to denote the distribution of perturbations of the input sample parameterized by \\( \\sigma \\), referred to as the perturbation scale. A definition of \"explanation\" is recovering the contribution of features to a prediction, which emanates from human's natural counterfactual reasoning [19]. Therefore, the explainer is usually the gradient with respect to input \\( \\nabla_{\\pmb{x}} \\), or approximations of it, e.g. the finite difference, given by \\( \\Delta_{\\pmb{x}} f(\\pmb{x}) = \\frac{f(\\tilde{\\pmb{x}}) - f(\\pmb{x})}{\\tilde{\\pmb{x}} - \\pmb{x}} \\) (see Appendix D). Finally, we assume that the summary statistic is the expectation of the outputs, or the squared outputs, of the explainer, e.g. \\( (\\nabla f(\\tilde{\\pmb{x}}))^2 \\), over the chosen perturbation distribution, \\( p(\\tilde{\\pmb{x}} | \\pmb{x}, \\sigma) \\), denoted as \\( \\mathbb{E}_{p(\\tilde{\\pmb{x}})}\\left[(\\nabla f(\\tilde{\\pmb{x}}))^2\\right] \\), where the operations are carried out pixelwise. This is referred to as the explanation of \\( f \\) at \\( \\pmb{x} \\) with a hyperparameter \\( \\sigma \\). This probabilistic view can, in fact, unify different explanation methods into a single form (see Appendix C). We provide theoretical justifications for why using squared gradients is preferable to gradients, explaining the empirical findings [30], and will discuss both cases in Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.435, + 0.579, + 0.453 + ], + "angle": 0, + "content": "3 Spectral Analysis of Explanations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.66 + ], + "angle": 0, + "content": "This section demonstrates the spectral property in explanations when using the gradient explainer and its interaction with perturbations, where a significant difference in the frequency of learned features leads to a selective attribution of features. In the following sections, we draw on insights from spectral analysis of signals and systems theory, utilizing this viewpoint to gain a deeper understanding of explanation methods. We demonstrate a spectral bias related to gradient-based explanation approaches, identify its source, and interpret various methods in the literature as efforts to alleviate it through the incorporation of low-pass filters, either explicitly [55, 57] or implicitly [42, 62]. We illustrate that the inconsistencies in explanation techniques stem from different strategies for mitigating the spectral bias. Finally, we suggest two straightforward approaches to tackle these inconsistencies: determining an optimal perturbation scale in Sec. 4.1 and aggregating the explanations in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.787, + 0.722 + ], + "angle": 0, + "content": "Writing a spectral representation for prediction, as shown in Appendix A.1, the model output with a perturbation distribution can be written as a convolution between the function and the perturbation distribution \\( f * p \\), therefore we may use \"perturbation kernel\" and \"perturbation distribution\" interchangeably." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.742, + 0.577, + 0.758 + ], + "angle": 0, + "content": "3.1 Spectral Representation of Explainers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Here, by drawing a connection between explainability techniques and the literature on spectral signal processing [12, 38, 56], we take a novel avenue for our analysis. This connection enables us to have a spectral representation of the explanations, as discussed in the next sections (see Appendix A for the derivations). It is important to note that the Fourier bases in our work is 1D and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "per pixel, and should not be confused with the traditional Fourier bases used in image processing literature that expands along spatial dimensions." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.208, + 0.784, + 0.239 + ], + "angle": 0, + "content": "Gradients in Spectral Domain. For explanation methods that use gradients and a perturbation kernel, \\( p(\\tilde{\\pmb{x}}) \\) we have the following (see Appendix A.2)." + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.254, + 0.785, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} [ \\nabla f (\\tilde {\\boldsymbol {x}}) ] = \\int i 2 \\pi \\omega \\widehat {f} (\\omega) \\widehat {p} (\\omega) d \\omega (1) \\\\ = - 2 \\pi \\left[ \\int \\omega \\left(\\widehat {f} _ {E} (\\omega) \\widehat {p} _ {O} (\\omega) + \\widehat {f} _ {O} (\\omega) \\widehat {p} _ {E} (\\omega)\\right) d \\omega \\right] (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.335, + 0.784, + 0.412 + ], + "angle": 0, + "content": "where \\(\\widehat{p}\\) denotes the FT of the perturbation kernel. We used \\(\\widehat{f}_E\\) and \\(\\widehat{f}_O\\) to denote a function \\(\\widehat{f}\\) is decomposed into real even and imaginary odd components \\(\\widehat{f} = i\\widehat{f}_O + \\widehat{f}_E^1\\). The negative sign suggests that in case of using the gradients, we need to negate the actual values to obtain rankings. We will elaborate more on the negation of the attributions in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.432, + 0.784, + 0.537 + ], + "angle": 0, + "content": "Remark 1. Note the symmetry in the form of the spectral representation of the gradients. This form implies that it is inevitable for any perturbation kernel to extract asymmetric information about even and odd parts of the classifier \\( f \\). Therefore, the gradient explainer with a kernel symmetric about the sample being explained, only extracts information about the odd parts of the classifier. This is problematic because the attribution is expected to reflect the total contribution to the prediction." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.555, + 0.784, + 0.586 + ], + "angle": 0, + "content": "We will use this interesting symmetry later in Sec. 3.3 for analysis of a specific perturbation kernel." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.617, + 0.784, + 0.647 + ], + "angle": 0, + "content": "Squared Gradients in Spectral Domain. On the other hand, for explanation methods which use squared gradients, we have the following (see Appendix A.3)." + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.664, + 0.784, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] = 4 \\pi^ {2} \\int \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) \\mathrm {S} _ {\\sqrt {p}} (\\boldsymbol {\\omega}) d \\boldsymbol {\\omega} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.704, + 0.784, + 0.734 + ], + "angle": 0, + "content": "where \\( \\mathrm{S}_f \\) and \\( \\mathrm{S}_{\\sqrt{p}} \\) denote the PSD of the model and square root of the perturbation kernel respectively." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.736, + 0.784, + 0.796 + ], + "angle": 0, + "content": "We discuss the benefits of using squared gradients as in Eq. (3) over gradients as in Eq. (2) later in Sec. 3.3. However, it can be noticed here that the squared gradients, unlike gradients, are always non-negative, extract information about the PSD of \\( f \\), and do not have a complex symmetric structure." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.784, + 0.84 + ], + "angle": 0, + "content": "1 Also note that as we are using real valued functions, other components i.e. real odd and imaginary even, are zero." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.779, + 0.162 + ], + "angle": 0, + "content": "3.2 Gradient Explainers with Perturbations Are Band-Pass Filters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.784, + 0.3 + ], + "angle": 0, + "content": "In our probabilistic framework, various explanation methods can be decomposed into three components (1) an explainer, (2) a perturbation kernel and (3) a summary statistic. In this section, we analyze the spectral effects of the two core components, namely gradient and perturbation. Importantly, we show that gradient acts as a high-pass filter, and perturbation acts as a low-pass filter. Before that, we need a more formal definition for a high-pass and low-pass filter that suits our study. Assuming that \\(\\widehat{m} (\\omega)\\) denotes the FT of a perturbation kernel, we define low- and high-pass as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.314, + 0.785, + 0.36 + ], + "angle": 0, + "content": "Definition 1. A filter \\(\\widehat{m}(\\omega)\\) is low pass, if for a given \\(M > 0\\) there exists a threshold \\(\\omega^{*}\\) such that for any \\(\\|\\omega\\| > \\|\\omega^{*}\\|\\), we have \\(\\|\\widehat{m}(\\omega)\\| < M\\), and it is high pass if such a condition holds for \\(\\|\\omega\\| < \\|\\omega^{*}\\|\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.373, + 0.784, + 0.403 + ], + "angle": 0, + "content": "For better readability, the proofs for the theorems and propositions are outlined in Appendix B, and we provide intuition as we progress." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.421, + 0.784, + 0.452 + ], + "angle": 0, + "content": "Theorem 1. Any perturbation kernel \\( p(\\tilde{\\pmb{x}}) \\) mitigates the attribution of high-frequency features; simply put, any perturbation is a low-pass filter." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.468, + 0.784, + 0.514 + ], + "angle": 0, + "content": "The intuition is that any modification to a signal, would only remove information, and, as the intensity of perturbation increases, the removal goes from high-frequency components toward lower frequencies." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.532, + 0.784, + 0.563 + ], + "angle": 0, + "content": "Theorem 2. The gradient operator amplifies the attribution of high-frequency features; simply put, the gradient operator is a high-pass filter." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.58, + 0.784, + 0.624 + ], + "angle": 0, + "content": "From the stated theorems, one can see that a gradient-based explanation with perturbation creates a band-pass filter, whose shape is controlled by the perturbation hyperparameters (see Figs. 2 and 3)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.643, + 0.784, + 0.718 + ], + "angle": 0, + "content": "Remark 2. This amplification-mitigation trade-off can be seen in several explanation methods, where perturbation kernels (low-pass filters) are utilized to mitigate the amplification effect (the high-pass filter) caused by the gradient. Therefore, the combined effect of the two filters usually forms a band-pass filter (see Appendix C for a non-comprehensive list)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.748, + 0.675, + 0.763 + ], + "angle": 0, + "content": "3.3 Spectral Analysis of Specific Perturbation Kernels" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this section, we analyze the properties of special kernels used in the literature with remarks or propositions that justify, and thus shed light on, previous work's experimental observations. In most cases, we expand upon the squared gradients, Eq. (3), due to its analytical simplicity." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "VanillaGrad via a Dirac Kernel. The very naive explanation method can be obtained with a Dirac kernel \\( p(\\tilde{\\pmb{x}}) = \\delta_{\\pmb{x}}(\\tilde{\\pmb{x}}) \\) known as VanillaGrad (VG). Analyzing the squared version of VG and following Eq. (3), we know that the FT of the Dirac kernel is \\( S_{\\sqrt{p}}(\\omega) = 1(\\omega) \\), where \\( \\mathbf{1}(\\omega) \\) is the constant function mapping all values to 1. Thus, it can be seen that we recover the scaled PSD of \\( f \\), i.e. \\( \\| \\boldsymbol{\\omega} \\|^2 \\mathrm{S}_f(\\boldsymbol{\\omega}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Remark 3. Scaling \\(\\widehat{f}\\) by frequency can be interpreted as the bias of the gradient explainer towards exaggeration of the attribution of high-frequency features [31]. Beside other justifications for gradient sparsity [8], spectral representation of \\(\\mathrm{VG}^2\\) and domination of high-frequency features give an orthogonal justification for why VanillaGrad explanations being usually scattered and noisy (see Fig. 12)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.425 + ], + "angle": 0, + "content": "SmoothGrad via a Gaussian Kernel. Letting \\( p(\\tilde{\\pmb{x}}) = \\mathcal{N}(\\pmb{x},\\sigma^2) \\), i.e. a Gaussian with variance \\( \\sigma^2 \\), centered on the sample being explained, we recover variants of an explanation method known as SmoothGrad (SG) or the squared gradient version SmoothGrad-Squared (SG²) [30, 55], we can rewrite Eq. (3) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.426, + 0.786, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] \\propto \\int \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) e ^ {- 8 \\pi^ {2} \\sigma^ {2} \\| \\boldsymbol {\\omega} \\| ^ {2}} d \\boldsymbol {\\omega} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.569 + ], + "angle": 0, + "content": "This shows that the contributions \\( \\mathrm{S}_f(\\omega) \\) are modulated by the combined spectral effect of the Gaussian perturbation kernel and gradient operator, which is a band pass filter of the form \\( \\| \\boldsymbol{\\omega} \\|^2 e^{-8\\pi^2 \\sigma^2} \\| \\boldsymbol{\\omega} \\|^2 \\), whose shape is controlled by the kernel variance (see Fig. 2). The explanations could be inconsistent, but generally become smoother as \\( \\sigma \\) increases, which is a result of the shrinkage of the attribution of high-frequency features that can dominate the explanations with a small \\( \\sigma \\) (see Fig. 1)." + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.602, + 0.625, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Fig.2: A visualization of frequency bands, where \\( \\mathrm{S}_{\\sqrt{p}}(\\omega) \\) is defined according to Eq. (4), show that change in the noise scale \\( \\sigma \\) leads to band-pass filters with different modes, hence, representing information of different frequencies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "When generating explanations, one is tempted to look for hyperparameters that make the explanation reflect what they think it should [16]. However, monitoring explanations from only one frequency band is potentially misleading and is likely to contradict itself simply by changing hyperparameters (see Fig. 5) [27,33]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.149, + 0.387, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.271, + 0.319, + 0.283 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.394, + 0.149, + 0.553, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.271, + 0.482, + 0.283 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.149, + 0.774, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.271, + 0.593, + 0.283 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.294, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Fig. 3: A 1D visualization of amplification (b) and mitigation (d) effects caused by gradient and perturbation, respectively. For simplicity, we have shown a rectangular pulse (c) convolved with \\(\\nabla f\\), but the kernel shape is usually more complex in practice, an example shown in Fig. 2. Moreover, functions in blue are the Fourier representation of the black functions. Due to the gradient operator, small high-frequency jitters in the decision boundary (a) amplify proportionally to their frequency (b). To remove the effect of high-frequency components of the decision boundary that usually dominate the explanations, researchers have proposed different perturbation kernels, visualized in (d). Theorem 2 corresponds to \\((\\mathrm{a}\\rightarrow \\mathrm{b})\\) and Theorem 1 corresponds to \\((\\mathrm{b}\\rightarrow \\mathrm{d})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.785, + 0.481 + ], + "angle": 0, + "content": "We formalize this statement, also known as the Rashomon effect (see Appendix B for the proof)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.493, + 0.789, + 0.54 + ], + "angle": 0, + "content": "Proposition 1. The SmoothGrad-Squared pixel attribution can contradict itself by changing the hyperparameters. Simply, the Rashomon effect can occur in the SmoothGrad-Squared explanations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.555, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Intuition. This proposition suggests a reason, i.e. removing different frequency bands, for inconsistencies we encounter in \\(\\mathrm{SG}^2\\) explanations, which has been unknown so far [48]. Note that this occurs with different kernels, but we have chosen the one that leads to a simpler proof." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.639, + 0.788, + 0.716 + ], + "angle": 0, + "content": "IntegratedGrad via a Rect Kernel. Setting the perturbation kernel to \\( p(\\tilde{\\boldsymbol{x}}) = \\frac{1}{\\sigma|\\boldsymbol{s} - \\boldsymbol{x}|}\\mathrm{Rect}(\\frac{\\tilde{\\boldsymbol{x}} - \\boldsymbol{x}}{\\sigma(\\boldsymbol{s} - \\boldsymbol{x})} -\\frac{1}{2})^2 \\), where \\( s \\) is a chosen baseline image, we recover the perturbation kernel of a well known method called IntegratedGrad (IG) [57]. We use the spectral representation of gradients in Eq. (2), for a proposition about the gradient signs." + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.727, + 0.787, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} [ \\nabla f (\\tilde {\\boldsymbol {x}}) ] \\propto \\int i \\widehat {f} (\\boldsymbol {\\omega}) \\sin (\\pi \\sigma (\\boldsymbol {s} - \\boldsymbol {x}) \\boldsymbol {\\omega}) e ^ {i \\pi \\boldsymbol {\\omega} \\boldsymbol {r}} d \\boldsymbol {\\omega} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.769, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Where \\(\\boldsymbol{r} = (1 + 2\\frac{\\boldsymbol{x}}{\\sigma(\\boldsymbol{s} - \\boldsymbol{x})})\\), appears because the kernel is not centered on the sample being explained. Note that the Fourier operator linearly expands along each dimension, but we use vector forms for notational simplicity. Furthermore, to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.614, + 0.843 + ], + "angle": 0, + "content": "\\(^2\\) Rect function is defined as \\(\\mathrm{Rect}(\\tilde{\\pmb{x}}) = \\frac{1}{2}\\mathrm{sign}(\\frac{1}{2} -|\\tilde{\\pmb{x}} |) + \\frac{1}{2}\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "study the effect of elementwise multiplication on the explanation quantitatively, we use IG to denote integrated gradients without the elementwise multiplication with input and XIG to denote the version with elementwise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.268 + ], + "angle": 0, + "content": "In explanation visualizations, the negative gradient sign has been used to show a negative contribution to the prediction [7, 53], to our knowledge it has never been formally justified. In the next proposition, we state that the negative gradient sign can be the direct result of the chosen perturbation kernel, formally shown in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.772, + 0.294 + ], + "angle": 0, + "content": "Proposition 2. The gradient sign in \\(IG\\) is affected by the perturbation level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.784, + 0.349 + ], + "angle": 0, + "content": "This proposition entails that the negative sign in the gradient explainer does not necessarily reflect a negative contribution to the prediction, nor the positive sign necessarily reflects a positive contribution (see Figs. 13 and 14)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.357, + 0.785, + 0.418 + ], + "angle": 0, + "content": "Remark 4. Since we assume continuity wrt \\(\\sigma\\), in our analysis, by Bolzano's theorem [50], there are hyperparameters that set the attribution to zero. Therefore, the gradient explainer may not produce a useful ranking compared to that of the gradient squared." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.428, + 0.787, + 0.504 + ], + "angle": 0, + "content": "Remark 5. Proposition 2 also shows that adding explanations with different hyperparameters may cause destructive interference when the different values of \\(\\omega \\widehat{p} (\\omega)\\) are out of phase. Hence, squared gradients is a more appropriate design choice, unless the use of phase information is planned as part of the method. This has been discovered experimentally [30] and confirmed in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.525, + 0.747, + 0.543 + ], + "angle": 0, + "content": "4 Proposed Solutions Based on the Spectral Analysis" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.556, + 0.586, + 0.571 + ], + "angle": 0, + "content": "4.1 Finding an Optimal Perturbation Scale" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.785, + 0.687 + ], + "angle": 0, + "content": "Note that the squared gradients in spectral domain (Eq. (3)) can be formulated as an inner product between the scaled PSD of the perturbation distribution and the PSD of the classifier i.e. \\(\\mathbb{E}_{p(\\tilde{\\boldsymbol{x}})}\\left[(\\nabla f(\\tilde{\\boldsymbol{x}}))^2\\right] = 4\\pi^2\\langle \\mathrm{S}_f,\\| \\boldsymbol {\\omega}\\| ^2\\mathrm{S}_{\\sqrt{p}}\\rangle\\) Assuming unimodality, a nonzero contribution, and that the norm \\(\\left\\| \\left\\| \\boldsymbol {\\omega}\\right\\| ^2\\mathrm{S}_{\\sqrt{p}}\\right\\|\\) exists, then the cosine similarity is maximized at a certain perturbation scale that has the highest similarity with \\(\\mathrm{S}_f\\) . A higher cosine similarity guarantees recovering more information about the classifier." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.695, + 0.787, + 0.726 + ], + "angle": 0, + "content": "Definition 2. The cosine similarity between the PSD of the classifier and that of the perturbation kernel given by" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.735, + 0.785, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\text {S i m i l a r i t y} \\left(\\mathrm {S} _ {f}, \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}}\\right) = 4 \\pi^ {2} \\frac {\\left\\langle \\mathrm {S} _ {f} , \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\right\\rangle}{\\| \\mathrm {S} _ {f} \\| \\left\\| \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\right\\|} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This quantity can be used as a way to determine the optimal perturbation scale per pixel to maximize the information extracted with the squared gradients. Note that, the word \"optimal\" refers to extracting the most information by the inner product, whether deemed relevant by a human or not.." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "For one of the perturbation kernels we have presented earlier in Sec. 3.3, we have optimized this quantity on multiple levels, such as dataset level Fig. 4, image level Tab. 1, and, pixel level Fig. 5d. Finally, we have experimentally verified the effectiveness of this approach in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.226, + 0.787, + 0.259 + ], + "angle": 0, + "content": "Computing the Similarity of Classifier and Gaussian Kernel. To compute the similarity, we first need to evaluate the kernel norm as follows" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.263, + 0.786, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2} e ^ {- 8 \\pi^ {2} \\sigma^ {2} \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2}} \\right\\| = \\left(\\int \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {4} e ^ {- 1 6 \\pi^ {2} \\sigma^ {2} \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2}} d \\boldsymbol {\\omega}\\right) ^ {\\frac {1}{2}} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.465, + 0.303, + 0.786, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\propto \\sigma^ {\\frac {- 5}{2}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.329, + 0.787, + 0.359 + ], + "angle": 0, + "content": "Therefore, the cosine similarity for the Gaussian perturbation kernel can be computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.365, + 0.646, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S i m i l a r i t y} (\\mathrm {S} _ {\\sqrt {p}}, \\mathrm {S} _ {f}) \\propto \\sigma^ {\\frac {5}{2}} \\langle \\mathrm {S} _ {f}, \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\rangle\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.389, + 0.787, + 0.452 + ], + "angle": 0, + "content": "Using the L1 norm for the squared gradients of flattened explanations. We have observed experimentally that \\( \\mathrm{S}_f \\) depends on the input scale, hence we use a convex interpolation with noise to support the assumption that \\( \\mathrm{S}_f \\) remains constant (see Appendix E.2 and Fig. 8). The results are shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.457, + 0.788, + 0.579 + ], + "angle": 0, + "content": "Remark 6. The optimal perturbation scale, determined by maximizing cosine similarity, offers a standard method for selecting the perturbation scale in \\(\\mathrm{SG^2}\\). Given that user discretion in the selection of perturbation scale contributes to inconsistencies, as illustrated in Fig. 1, opting for the optimal perturbation scale mitigates these inconsistencies in the results. Hence, it represents a more suitable design choice. In Sec. 5 we demonstrate that this approach not only enhances \\(\\mathrm{SG^2}\\)'s robustness against inconsistencies but also its adaptability to diverse datasets, which in turn can improve the quantitative performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.603, + 0.691, + 0.619 + ], + "angle": 0, + "content": "4.2 Aggregation of Information With Explanation Prior" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.787, + 0.716 + ], + "angle": 0, + "content": "We have shown that explanations are influenced by specific frequency bands, rather than representing the importance of individual pixels for prediction. Instead, they indicate the contributions within a specific frequency range. This observation motivates us to investigate a method for aggregation of information across different bands, which closely resembles an ensemble of IG and SG [55,57] which we call SpectralLens." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.721, + 0.787, + 0.753 + ], + "angle": 0, + "content": "Definition 3. SpectralLens is an ensemble explanation method, that aggregates information across different frequency bands as follows" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.758, + 0.787, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {p (\\sigma)} \\left[ \\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] \\right] \\propto \\iint \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) \\| \\boldsymbol {\\omega} \\| ^ {2} e ^ {- 8 \\pi^ {2} \\sigma \\| \\boldsymbol {\\omega} \\| ^ {2}} d \\boldsymbol {\\omega} p (\\sigma) d \\sigma \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Where we refer to this method by \\( SL^2 \\) as it is using the squared gradients and we call \\( p(\\sigma) \\) as an explanation prior, gauging the frequency bands that we choose to represent in an explanation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.497, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.251, + 0.371, + 0.264 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.149, + 0.779, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.252, + 0.653, + 0.264 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.789, + 0.43 + ], + "angle": 0, + "content": "Fig. 4: Cosine similarity between kernel and classifier, defined in Eq. (6), is measured for the Gaussian perturbation kernel and ResNet50 averaged on the 10K images of the ImageNet validation set (a) and Food101 (b). This metric assesses the similarity between the PSD of the perturbation kernel and that of the classifier, indicating the perturbation level's effectiveness for kernels with finite norm in the Fourier domain. When using a Dirac explanation prior in Eq. (9), where a perturbation level is assigned per image or dataset, the optimal choice for the perturbation level is where the similarity is maximized across the image or dataset respectively. Note that this level of noise refers to a band-pass filter, shown in Fig. 2, which suggests the consideration of a latent distribution on the frequency of labeling, reflecting the workforce's mindset during data labeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Remark 7. \\(\\mathrm{SL}^2\\) exhibits robustness against inconsistencies in \\(\\mathrm{SG}^2\\) due to its aggregation of information from various bands. Therefore, it is a preferable design choice not to employ a Dirac explanation prior with \\(SG^2\\), as it eliminates inconsistencies, depicted in Fig. 1. Nevertheless, as demonstrated in the Sec. 5, this approach may enhance the quantitative performance of explanations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Intuition. Considering Eq. (4), we realize that the combination of perturbation and gradient explainer creates overlapping band-pass carrier waves that modulate the amplitude of the spectral density of the classifier. Therefore, a new summary statistic is required that can reflect a wide range of frequency bands. We explicitly identified the spectral properties of the explanations, recognizing that not all frequencies are equally desired. Explanations with very high frequencies primarily attribute predictions to pixels. Therefore, we utilize an explanation prior \\( p(\\sigma) \\) for a nuanced control of the spectral properties of explanations. Although there are many possible alternatives for selecting an explanation prior, we used \\( p(\\sigma) = \\mathrm{Unif}(0,1) \\) for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.787, + 0.742 + ], + "angle": 0, + "content": "Definition 4. We define \\(\\mathbf{ArgLens}(AL)\\) as the point at which the \\(SG^2\\) is maximized for pixel \\(x_i\\). Formally defined as" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.751, + 0.786, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A r g L e n s} \\left(x _ {i}\\right) = \\arg \\max _ {\\sigma} \\left\\{\\mathrm {S G} ^ {2} \\left(x _ {i}\\right) \\right\\} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.784, + 0.628, + 0.8 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the hyperparameter of the perturbation kernel." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "AL shows the frequency band at which the pixel's attribution is maximized. Alternatively, it can be interpreted as the robustness of a pixel to noise. We" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "have visualized the AL for the Gaussian kernel in Fig. 5 and discussed the way it helps the interpretation of explanations in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.239 + ], + "angle": 0, + "content": "The aggregation of explanations from different frequencies is visualized in Fig. 5. Moreover, the basic explanation methods are obtainable with a Dirac explanation prior \\( p(\\sigma) = \\delta(\\sigma) \\). Therefore, it can be used as a unified formulation for gradient-based explanation methods in spectral domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Interesting Properties of SpectralLens and ArgLens. In the absence of a robust mathematical framework, the interpretation of explanations may be subjective. Thus, explanation methods must be grounded in solid theoretical principles. \\(\\mathrm{SL}^2\\) addresses potential inconsistencies by aggregating attributions across frequency bands, ensuring fair representation of all pixels' contributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.456 + ], + "angle": 0, + "content": "As \\(\\mathrm{SG^2}\\) results in bands-pass filters (see Sec. 3.3), its attributions reflect the changes in the decision boundary within a certain frequency band. Furthermore, AL reveals the frequency band at which the attribution is maximized. Hence, lower AL indicates that the attributions are likely governed by high-frequency components of the decision boundary, more sensitive to input noise, consequently less robust. An ideal classifier with smooth decision boundaries, i.e. robust features, leads to high AL everywhere, with a high intensity in \\(\\mathrm{SL^2}\\) on relevant pixels and low intensity on irrelevant pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Given the different purposes of pixels within an image, the frequency of activation may vary and can be visualized via AL. Since \\(\\mathrm{SL}^2\\) and AL offer complementary information, they should be interpreted together. This approach offers a deeper understanding compared to a simplistic perspective, as it has been shown by Rahaman et al. [43] that high-frequency regions emerge later in training, are more challenging to fit, and are prone to change. On the contrary, low-frequency components of the decision boundary manifest earlier in training, are easier to fit, and exhibit greater stability with respect to noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.651 + ], + "angle": 0, + "content": "It is important to clarify that \"high-frequency\" refers to features considered high-frequency by the model, and not necessarily frequency across spatial dimensions (e.g. width, height). Nonetheless, empirical findings demonstrate that high-frequency features in spatial domain are more likely to be identified as high frequency by the model (see Appendix I)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.357, + 0.691 + ], + "angle": 0, + "content": "5 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Given the experimental discoveries that squared gradients lead to better results [30], we evaluate both scenarios for different hyperparameters and compare them with SL in Tabs. 1 and 2, refer to Appendix E for implementation details and reproducibility. For the purposes of this study, we use pixel removal strategy, despite its problems [30]. The spectral analysis of pixel removal strategy and possible solutions are deferred to future research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We report single-pixel most-important-first (MoRF) insertion and deletion scores of the methods to assess the design choices previously proposed by our theoretical framework: (1) utilizing squared gradients instead of gradients in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.359, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.146, + 0.499, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.639, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.147, + 0.781, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.252, + 0.359, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.359, + 0.302, + 0.371 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.253, + 0.499, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.359, + 0.443, + 0.371 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.253, + 0.639, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.359, + 0.583, + 0.371 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.253, + 0.781, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.359, + 0.724, + 0.371 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Fig. 5: This figure shows example visualizations of explanations generated by \\(\\mathrm{SG^2}\\) implementing design decisions suggested in our work in Remarks 6 and 7. (b) shows the \\(\\mathrm{SG^2}\\) visualized at the maximum similarity according to Fig. 4, for this image \\(\\sigma = 0.5\\). (c) shows the SL attributions. Lastly, (d) shows the frequencies obtained from AL (defined in Eq. (10)), which correspond to the perturbation level at which \\(\\mathrm{SG^2}\\) is maximized. Since we are interested in frequencies rather than the actual perturbation levels, we visualize 2 stds above and below the mean of frequencies obtained by \\(\\omega = 1 / (1 + \\mathrm{AL})\\) (see Appendix I for more examples). The heatmaps in (b) and (c) show how much the model is relying on the pixel, while (d) shows the frequency at which the feature's contribution maximizes. In the provided examples, we can see that the model is relying on the nose (top) and eye (bottom) while the same region is of a lower frequency compared to the head. This can be interpreted as the model finding the nose and eye more stable features compared to the head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Remark 5, (2) finding a standard perturbation scale in Remark 6, and (3) employing a uniform explanation prior instead of Dirac, stated in Remark 7 (see Fig. 7 for an ablation study of kernel variance and Appendix G for an ablation study of patch removal instead of single pixel removal). We evaluated the suggested design choices on ImageNet [20] and Food101 [1] in Tab. 1. However, smaller datasets, such as CIFAR, are omitted because their resolution restricts the potential gap between high- and low-frequency features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Limitations of Our Framework. The choice of independent Gaussian perturbation is limited due to its inability to incorporate prior knowledge about the input space or data manifold characteristics, and other data modalities such as graphs and time series. However, preferring simpler analytical forms within our framework limits the alternative choices. Efficiency is another challenge compared to basic methods, as performing back-props at various noise scales is costly. Moreover, our attribution of frequencies in visualizations, such as SL, may lack the desired generalizability across data points, as high-frequency re" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Table 1: This table presents the assessment of the proposed design choices in Remarks 5 to 7 across different methods, highlighting statistically significant improvements with bold fonts. We conduct t-tests comparing gradient methods to their squared-gradient counterparts to evaluate the design choice in Remark 5. Furthermore, we conduct t-tests between \\(\\mathrm{SG}^2\\) with the original heuristic and those ensuring consistency to evaluate design decisions in Remarks 6 and 7. Overlines and underlines denote highest or lowest values within columns, respectively. Although there are no squared versions of Occlusion Sensitivity (OS) [5] and RISE [42], they are included as low-frequency baselines. While the dimensions of boolean masks utilized in OS and RISE can implicitly regulate the sampling variance (see Appendix C), the exponential increase in the required number of samples in high-frequency explanations renders them less practical. SG scores follow the heuristic of the SmoothGrad paper [55] i.e. \\(\\sigma = (\\max(x) - \\min(x))\\) 0.1. XIG indicates the input multiplication in the original implementation of IntegratedGrad [57], while IG does not apply such transformation. Rankings for methods using gradients are obtained from negative gradients, including RISE and Occlusion Sensitivity, according to Eq. (2). Furthermore, \\(\\mathrm{SG}_{\\mathrm{Opt.}}^2\\) uses an optimized kernel per image and \\(\\mathrm{SL}^2\\) uses a uniform explanation prior. Comparing SG² and \\(\\mathrm{SG}_{\\mathrm{Opt.}}^2\\), we can see that \\(\\mathrm{SG}_{\\mathrm{Opt.}}^2\\) can adapt to the statistics of the Food101. Therefore, \\(\\mathrm{SG}_{\\mathrm{Opt.}}^2\\), while not subject to inconsistencies (shown in Fig. 1), can improve the performance due to its adaptability. Hence, we can ensure the consistency of SG² without compromising its performance. Also, note that the input multiplication enhances visual quality, but degrades the performance." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.462, + 0.734, + 0.664 + ], + "angle": 0, + "content": "
ImageNetFood101
Insertion(↑)Deletion(↓)Insertion(↑)Deletion(↓)
GradientsOS0.384 ± 0.0080.237 ± 0.0100.378 ± 0.0110.209 ± 0.009
RISE0.494 ± 0.0140.174 ± 0.0070.410 ± 0.0130.198 ± 0.007
SG0.212 ± 0.0100.179 ± 0.0090.179 ± 0.0080.137 ± 0.007
XIG0.208 ± 0.0090.204 ± 0.0080.101 ± 0.0070.177 ± 0.007
IG0.201 ± 0.0100.182 ± 0.0090.145 ± 0.0060.131 ± 0.005
Gradients2SG20.465 ± 0.0130.157 ± 0.0070.371 ± 0.0120.162 ± 0.006
XIG20.334 ± 0.0140.187 ± 0.0080.204 ± 0.0100.263 ± 0.007
IG20.340 ± 0.0120.168 ± 0.0070.320 ± 0.0090.135 ± 0.005
Gradients2Opt.0.491 ± 0.0150.187 ± 0.0070.394 ± 0.0130.167 ± 0.006
SL20.493 ± 0.0160.169 ± 0.0070.395 ± 0.0140.157 ± 0.007
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.739 + ], + "angle": 0, + "content": "gions in one image may correspond to low-frequency regions in another image (see Appendix I). Identifying generalizable attributions across images remains a subject for future investigation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.762, + 0.398, + 0.779 + ], + "angle": 0, + "content": "6 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Probabilistic and Spectral View of Explanations. Despite sporadic attempts to analyze explanations from a probabilistic view [17, 59], this view is more common in the literature around the uncertainty quantification of expla" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "nations [29,40,54]. Our work is closely related to Covert et al. [19], if their work is seen from a probabilistic perspective. Aside from the works around the spectral view of deep networks [10,11,15,43,58], as of our current knowledge, there is no work adopting a spectral view of explanations along pixels' value. Nonetheless, there are works that build on a spectral view along spatial dimensions [34,35], extending an optimization-based definition of explanations [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.79, + 0.382 + ], + "angle": 0, + "content": "Unifying Views to Explanations. Covert et al. [19] propose an information theoretic framework that unifies several explanation methods, focusing on three design choices: (1) feature removal method, (2) which model behavior to analyze, and (3) summary statistics. Our probabilistic framework, being similar to Covert's, offers a distinct viewpoint relevant to our spectral analysis. It can be used as a unifying view of explanations (see Appendix C), inheriting from Covert's framework. Notably, our work aligns with the broader context of research under finding unifying representations of explanations [3,26,37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.79, + 0.588 + ], + "angle": 0, + "content": "Input-space or Function-space Perturbations. Methods such as NoiseGrad [14], Grad-CAM [51], RectifiedGrad [32], Layer-wise Relevance Propagation [7], DeepLift [53], and even Sharpness-Aware Minimization (SAM) [24] employ function space perturbations. Utilizing such distributions offers data modality independence, whereas input space perturbation remains agnostic about the model architecture. While acknowledging the duality between function and input space distributions, our focus on explicit input space distributions excludes these methods from our study. Although we recognize the utility of integrating function space perturbation into our mathematical model, we defer this aspect to future investigations. It can be noted, however, that extending this work to function space perturbation will connect it to the literature on double descent and smooth interpolation of samples [25, 41]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.608, + 0.36, + 0.625 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.789, + 0.807 + ], + "angle": 0, + "content": "In this study, we deepen understanding of pixel attribution methodologies by rigorously analyzing their theoretical underpinnings, revealing a common spectral bias. Our analysis supports the experimental finding of recent works regarding the efficacy of using squared gradients. Moreover, we demonstrate how spectral properties can lead to inconsistent explanations and propose two design choices informed by our theoretical framework to address these inconsistencies: standardizing hyperparameters based on cosine similarity between the perturbation kernel and classifier, and aggregating explanations with a given prior. Despite the effectiveness of our approach, we acknowledge limitations and suggest avenues for further research. Lastly, we assess the suggested design choices using deletion and insertion scores." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "We would like to thank Giovanni Luca Marchetti for an early review of our work and his kind feedback. This project is partially supported by Region Stockholm through MedTechLabs, and Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Scientific computation was enabled by the supercomputing resource Berzelius, provided by the National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg foundation." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.786, + 0.368 + ], + "angle": 0, + "content": "1. Food-101 - Mining Discriminative Components with Random Forests, https://data.vision.ee.ethz.ch/cvl/datasets-extra/food-101/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.369, + 0.786, + 0.41 + ], + "angle": 0, + "content": "2. Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., Kim, B.: Sanity Checks for Saliency Maps (Nov 2020). https://doi.org/10.48550/arXiv.1810.03292, http://arxiv.org/abs/1810.03292, arXiv:1810.03292 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.786, + 0.465 + ], + "angle": 0, + "content": "3. Agarwal, S., Jabbari, S., Agarwal, C., Upadhyay, S., Wu, Z.S., Lakkaraju, H.: Towards the Unification and Robustness of Perturbation and Gradient Based Explanations (Jul 2021). https://doi.org/10.48550/arXiv.2102.10618, http://arxiv.org/abs/2102.10618, arXiv:2102.10618 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.466, + 0.786, + 0.508 + ], + "angle": 0, + "content": "4. Alvarez-Melis, D., Jaakkola, T.S.: On the Robustness of Interpretability Methods (Jun 2018). https://doi.org/10.48550/arXiv.1806.08049, http://arxiv.org/abs/1806.08049, arXiv:1806.08049 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.786, + 0.564 + ], + "angle": 0, + "content": "5. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for Deep Neural Networks (Mar 2018). https://doi.org/10.48550/arXiv.1711.06104, http://arxiv.org/abs/1711.06104, arXiv:1711.06104 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.786, + 0.632 + ], + "angle": 0, + "content": "6. Arrieta, A.B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., Chatila, R., Herrera, F.: Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI (Dec 2019). https://doi.org/10.48550/arXiv.1910.10045, http://arxiv.org/abs/1910.10045, arXiv:1910.10045 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.786, + 0.701 + ], + "angle": 0, + "content": "7. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS ONE 10(7), e0130140 (Jul 2015). https://doi.org/10.1371/journal.pone.0130140, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0130140" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.758 + ], + "angle": 0, + "content": "8. Balduzzi, D., Frean, M., Leary, L., Lewis, J.P., Ma, K.W.D., McWilliams, B.: The Shattered Gradients Problem: If resnets are the answer, then what is the question? (Jun 2018). https://doi.org/10.48550/arXiv.1702.08591, http://arxiv.org/abs/1702.08591, arXiv:1702.08591 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "9. Bansal, N., Agarwal, C., Nguyen, A.: SAM: The Sensitivity of Attribution Methods to Hyperparameters (Apr 2020). https://doi.org/10.48550/arXiv.2003.08754, http://arxiv.org/abs/2003.08754, arXiv:2003.08754 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "10. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency Bias in Neural Networks for Input of Non-Uniform Density (Mar 2020), https://api-semanticscholar.org/CorpusID:212644664" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.34, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "11. Benbarka, N., Hofer, T., Ul-Moqeet Riaz, H., Zell, A.: Seeing Implicit Neural Representations as Fourier Series. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 2283-2292 (Jan 2022). https://doi.org/10.1109/WACV51458.2022.00234, https://ieeexplore.ieee.org/document/9706795/" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.26 + ], + "angle": 0, + "content": "12. Blackman, R.B., Tukey, J.W.: The measurement of power spectra from the point of view of communications engineering — Part I. The Bell System Technical Journal 37(1), 185–282 (Jan 1958). https://doi.org/10.1002/j.1538-7305.1958.tb03874.x, http://ieeexplore.ieee.org/document/6768513/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.317 + ], + "angle": 0, + "content": "13. Brughmans, D., Melis, L., Martens, D.: Disagreement amongst counterfactual explanations: How transparency can be deceptive (Apr 2023). https://doi.org/10.48550/arXiv.2304.12667, http://arxiv.org/abs/2304.12667 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.374 + ], + "angle": 0, + "content": "14. Bykov, K., Hedström, A., Nakajima, S., Höhne, M.M.C.: NoiseGrad: Enhancing Explanations by Introducing Stochasticity to Model Weights (May 2022). https://doi.org/10.48550/arXiv.2106.10185, http://arxiv.org/abs/2106.10185, arXiv:2106.10185 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.457 + ], + "angle": 0, + "content": "15. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards Understanding the Spectral Bias of Deep Learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. pp. 2205-2211. International Joint Conferences on Artificial Intelligence Organization, Montreal, Canada (Aug 2021). https://doi.org/10.24963/ijcai.2021/304, https://www.ijcai.org/proceedings/2021/304" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "16. Chen, H., Janizek, J.D., Lundberg, S., Lee, S.I.: True to the Model or True to the Data? (Jun 2020). https://doi.org/10.48550/arXiv.2006.16234, http://arxiv.org/abs/2006.16234, arXiv:2006.16234 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.558 + ], + "angle": 0, + "content": "17. Chen, J., Song, L., Wainwright, M.J., Jordan, M.I.: Learning to Explain: An Information-Theoretic Perspective on Model Interpretation (Jun 2018). https://doi.org/10.48550/arXiv.1802.07814, http://arxiv.org/abs/1802.07814, arXiv:1802.07814 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.613 + ], + "angle": 0, + "content": "18. Cohen, L.: The generalization of the Wiener-Khinchin theorem. In: Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181). vol. 3, pp. 1577-1580. IEEE (1998), https://ieeexplore.ieee.org/abstract/document/681753/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.657 + ], + "angle": 0, + "content": "19. Covert, I., Lundberg, S., Lee, S.I.: Explaining by Removing: A Unified Framework for Model Explanation (May 2022). https://doi.org/10.48550/arXiv.2011.14878, http://arxiv.org/abs/2011.14878, arXiv:2011.14878 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.726 + ], + "angle": 0, + "content": "20. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (Jun 2009). https://doi.org/10.1109/CVPR.2009.5206848, https://ieeexplore.ieee.org/document/5206848, iSSN: 1063-6919" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.798 + ], + "angle": 0, + "content": "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Jun 2021). https://doi.org/10.48550/arXiv.2010.11929, http://arxiv.org/abs/2010.11929, arXiv:2010.11929 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "22. Fong, R., Patrick, M., Vedaldi, A.: Understanding Deep Networks via Extremal Perturbations and Smooth Masks (Oct 2019). https://doi.org/10.48550/arXiv.1910.08485, http://arxiv.org/abs/1910.08485, arXiv:1910.08485 [cs, stat]" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "23. Fong, R., Vedaldi, A.: Interpretable Explanations of Black Boxes by Meaningful Perturbation. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 3449-3457 (Oct 2017). https://doi.org/10.1109/ICCV.2017.371, http://arxiv.org/abs/1704.03296, arXiv:1704.03296 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.258 + ], + "angle": 0, + "content": "24. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-Aware Minimization for Efficiently Improving Generalization (Apr 2021). https://doi.org/10.48550/arXiv.2010.01412, http://arxiv.org/abs/2010.01412, arXiv:2010.01412 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "25. Gamba, M., Azizpour, H., Björkman, M.: On the Lipschitz Constant of Deep Networks and Double Descent (Nov 2023). https://doi.org/10.48550/arXiv.2301.12309, http://arxiv.org/abs/2301.12309, arXiv:2301.12309 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.787, + 0.354 + ], + "angle": 0, + "content": "26. Han, T., Srinivas, S., Lakkaraju, H.: Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations (Dec 2022). https://doi.org/10.48550/arXiv.2206.01254, http://arxiv.org/abs/2206.01254, arXiv:2206.01254 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "27. Harel, N., Gilad-Bachrach, R., Obolski, U.: Inherent Inconsistencies of Feature Importance (Jun 2022). https://doi.org/10.48550/arXiv.2206.08204, http://arxiv.org/abs/2206.08204, arXiv:2206.08204 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.434 + ], + "angle": 0, + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition (Dec 2015). https://doi.org/10.48550/arXiv.1512.03385, http://arxiv.org/abs/1512.03385, arXiv:1512.03385 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.475 + ], + "angle": 0, + "content": "29. Hill, D., Masoomi, A., Ghimire, S., Torop, M., Dy, J.: Explanation Uncertainty with Decision Boundary Awareness (Nov 2022). https://doi.org/10.48550/arXiv.2210.02419, http://arxiv.org/abs/2210.02419, arXiv:2210.02419 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.528 + ], + "angle": 0, + "content": "30. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A Benchmark for Interpretability Methods in Deep Neural Networks (Nov 2019). https://doi.org/10.48550/arXiv.1806.10758, http://arxiv.org/abs/1806.10758, arXiv:1806.10758 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.527, + 0.787, + 0.582 + ], + "angle": 0, + "content": "31. Hou, X., Zhang, L.: Saliency Detection: A Spectral Residual Approach. In: 2007 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8 (Jun 2007). https://doi.org/10.1109/CVPR.2007.383267, https://ieeexplore.ieee.org/document/4270292, iSSN: 1063-6919" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.581, + 0.787, + 0.637 + ], + "angle": 0, + "content": "32. Kim, B., Seo, J., Jeon, S., Koo, J., Choe, J., Jeon, T.: Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps (Sep 2019). https://doi.org/10.48550/arXiv.1902.04893, http://arxiv.org/abs/1902.04893, arXiv:1902.04893 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.787, + 0.692 + ], + "angle": 0, + "content": "33. Kindermans, P.J., Hooker, S., Adebayo, J., Alber, M., Schütt, K.T., Dähne, S., Erhan, D., Kim, B.: The (Un)reliability of saliency methods (Nov 2017). https://doi.org/10.48550/arXiv.1711.00867, http://arxiv.org/abs/1711.00867, arXiv:1711.00867 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "34. Kolek, S., Nguyen, D.A., Levie, R., Bruna, J., Kutyniok, G.: Cartoon Explanations of Image Classifiers (Oct 2022). https://doi.org/10.48550/arXiv.2110.03485, http://arxiv.org/abs/2110.03485, arXiv:2110.03485 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.786 + ], + "angle": 0, + "content": "35. Kolek, S., Windsesheim, R., Loarca, H.A., Kutyniok, G., Levie, R.: Explaining Image Classifiers with Multiscale Directional Image Representation (Apr 2023). https://doi.org/10.48550/arXiv.2211.12857, http://arxiv.org/abs/2211.12857, arXiv:2211.12857 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.842 + ], + "angle": 0, + "content": "36. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The Disagreement Problem in Explainable Machine Learning: A Practitioner's Perspective (Feb 2022). https://doi.org/10.48550/arXiv.2202.01602, http://arxiv.org/abs/2202.01602, arXiv:2202.01602 [cs]" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Mehrpanah et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "37. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.787, + 0.243 + ], + "angle": 0, + "content": "38. Marmarelis, P.Z., Marmarelis, V.Z.: The White-Noise Method in System Identification. In: Marmarelis, P.Z., Marmarelis, V.Z. (eds.) Analysis of Physiological Systems: The White-Noise Approach, pp. 131-180. Computers in Biology and Medicine, Springer US, Boston, MA (1978). https://doi.org/10.1007/978-1-4613-3970-0_4, https://doi.org/10.1007/978-1-4613-3970-0_4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.786, + 0.297 + ], + "angle": 0, + "content": "39. Marques-Silva, J., Ignatiev, A.: Delivering Trustworthy AI through Formal XAI. Proceedings of the AAAI Conference on Artificial Intelligence 36(11), 12342-12350 (Jun 2022). https://doi.org/10.1609/aaai.v36i11.21499, https://ojs.aaai.org/index.php/AAAI/article/view/21499" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.786, + 0.352 + ], + "angle": 0, + "content": "40. Marx, C., Park, Y., Hasson, H., Wang, Y., Ermon, S., Huan, L.: But Are You Sure? An Uncertainty-Aware Perspective on Explainable AI. In: Proceedings of The 26th International Conference on Artificial Intelligence and Statistics. pp. 7375-7391. PMLR (Apr 2023), https://proceedings.mlr.press/v206/marx23a.htm1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.786, + 0.407 + ], + "angle": 0, + "content": "41. Nakkiran, P., Kaplun, G., Bansal, Y., Yang, T., Barak, B., Sutskever, I.: Deep Double Descent: Where Bigger Models and More Data Hurt (Dec 2019). https://doi.org/10.48550/arXiv.1912.02292, http://arxiv.org/abs/1912.02292, arXiv:1912.02292 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.786, + 0.447 + ], + "angle": 0, + "content": "42. Petsiuk, V., Das, A., Saenko, K.: RISE: Randomized Input Sampling for Explanation of Black-box Models (Sep 2018). https://doi.org/10.48550/arXiv.1806.07421, http://arxiv.org/abs/1806.07421, arXiv:1806.07421 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.786, + 0.502 + ], + "angle": 0, + "content": "43. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F.A., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks (May 2019). https://doi.org/10.48550/arXiv.1806.08734, http://arxiv.org/abs/1806.08734, arXiv:1806.08734 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.786, + 0.528 + ], + "angle": 0, + "content": "44. Rahman, M.: Applications of Fourier Transforms to Generalized Functions. WIT Press (2011), google-Books-ID: k_rdcKaUdr4C" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.786, + 0.623 + ], + "angle": 0, + "content": "45. Ras, G., van Gerven, M., Haselager, P.: Explanation Methods in Deep Learning: Users, Values, Concerns and Challenges. In: Escalante, H.J., Escalera, S., Guyon, I., Baró, X., Güçlütürk, Y., Güçlü, U., van Gerven, M. (eds.) Explainable and Interpretable Models in Computer Vision and Machine Learning, pp. 19-36. The Springer Series on Challenges in Machine Learning, Springer International Publishing, Cham (2018). https://doi.org/10.1007/978-3-319-98131-4_2, https://doi.org/10.1007/978-3-319-98131-4_2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.786, + 0.664 + ], + "angle": 0, + "content": "46. Reiter, H., Stegeman, J.D.: Classical Harmonic Analysis and Locally Compact Groups. Oxford University Press (Nov 2000). https://doi.org/10.1093/oso/9780198511892.001.0001, https://academic.oup.com/book/54460" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.786, + 0.705 + ], + "angle": 0, + "content": "47. Ribeiro, M.T., Singh, S., Guestrin, C.: \"Why Should I Trust You?\": Explaining the Predictions of Any Classifier (Aug 2016). https://doi.org/10.48550/arXiv.1602.04938, http://arxiv.org/abs/1602.04938, arXiv:1602.04938 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.786, + 0.759 + ], + "angle": 0, + "content": "48. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A Consistent and Efficient Evaluation Strategy for Attribution Methods (Jun 2022). https://doi.org/10.48550/arXiv.2202.00449, http://arxiv.org/abs/2202.00449, arXiv:2202.00449 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.813 + ], + "angle": 0, + "content": "49. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence 1(5), 206-215 (May 2019). https://doi.org/10.1038/s42256-019-0048-x, https://www.nature.com/articles/s42256-019-0048-x" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "50. Rudin, W.: Principles of Mathematical Analysis. McGraw-Hill (1976), google-Books-ID: kwqzPAAACAAJ" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spectral Properties of Gradient-based Explanation Methods" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.218 + ], + "angle": 0, + "content": "51. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision 128(2), 336-359 (Feb 2020). https://doi.org/10.1007/s11263-019-01228-7, http://arxiv.org/abs/1610.02391, arXiv:1610.02391 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.285 + ], + "angle": 0, + "content": "52. Serov, V.: The Riemann-Lebesgue Lemma. In: Serov, V. (ed.) Fourier Series, Fourier Transform and Their Applications to Mathematical Physics, pp. 33-35. Applied Mathematical Sciences, Springer International Publishing, Cham (2017). https://doi.org/10.1007/978-3-319-65262-7_6, https://doi.org/10.1007/978-3-319-65262-7_6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.342 + ], + "angle": 0, + "content": "53. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not Just a Black Box: Learning Important Features Through Propagating Activation Differences (Apr 2017). https://doi.org/10.48550/arXiv.1605.01713, http://arxiv.org/abs/1605.01713, arXiv:1605.01713 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.396 + ], + "angle": 0, + "content": "54. Slack, D., Hilgard, S., Singh, S., Lakkaraju, H.: Reliable Post hoc Explanations: Modeling Uncertainty in Explainability (Nov 2021). https://doi.org/10.48550/arXiv.2008.05030, http://arxiv.org/abs/2008.05030, arXiv:2008.05030 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.439 + ], + "angle": 0, + "content": "55. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: SmoothGrad: removing noise by adding noise (Jun 2017). https://doi.org/10.48550/arXiv.1706.03825, http://arxiv.org/abs/1706.03825, arXiv:1706.03825 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.468 + ], + "angle": 0, + "content": "56. Stoica, P., Moses, R.L.: Spectral analysis of signals, vol. 452. Pearson Prentice Hall Upper Saddle River, NJ (2005), http://user.it.uu.se/~ps/SAS-new.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.509 + ], + "angle": 0, + "content": "57. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic Attribution for Deep Networks (Jun 2017). https://doi.org/10.48550/arXiv.1703.01365, http://arxiv.org/abs/1703.01365, arXiv:1703.01365 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.563 + ], + "angle": 0, + "content": "58. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. ArXiv (Jun 2020), https://api_semanticscholar.org/CorpusID:219791950" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "59. Wang, E., Khosravi, P., Broeck, G.V.d.: Probabilistic Sufficient Explanations (May 2021). https://doi.org/10.48550/arXiv.2105.10118, http://arxiv.org/abs/2105.10118, arXiv:2105.10118 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.659 + ], + "angle": 0, + "content": "60. Wang, Y., Wang, X.: \"Why Not Other Classes?\": Towards Class-Contrastive Backpropagation Explanations. Advances in Neural Information Processing Systems 35, 9085-9097 (Dec 2022), https://proceedings.neurips.cc/paper_files/paper/2022/bitical/3b7a66b2d1258e892c89f485b8f896e0-AAbstract-Conference.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.716 + ], + "angle": 0, + "content": "61. Wilming, R., Kieslich, L., Clark, B., Haufe, S.: Theoretical Behavior of XAI Methods in the Presence of Suppressor Variables (Jun 2023). https://doi.org/10.48550/arXiv.2306.01464, http://arxiv.org/abs/2306.01464 [cs, stat]" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.758 + ], + "angle": 0, + "content": "62. Zeiler, M.D., Fergus, R.: Visualizing and Understanding Convolutional Networks (Nov 2013). https://doi.org/10.48550/arXiv.1311.2901, http://arxiv.org/abs/1311.2901, arXiv:1311.2901 [cs]" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.758 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_origin.pdf b/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..838fa212dedaa6db7e500ab316936a1144b3d803 --- /dev/null +++ b/2024/On Spectral Properties of Gradient-based Explanation Methods/b6f0f6e4-cee6-4eb8-8867-cce658641afc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1417fe092e1f1c8a589b613e9399d9cb13187eda864139c3d98e7c5fd58c344 +size 5108392 diff --git a/2024/On Spectral Properties of Gradient-based Explanation Methods/full.md b/2024/On Spectral Properties of Gradient-based Explanation Methods/full.md new file mode 100644 index 0000000000000000000000000000000000000000..98e3019b7c6865e6b08f5147e6e2ebdac3875eee --- /dev/null +++ b/2024/On Spectral Properties of Gradient-based Explanation Methods/full.md @@ -0,0 +1,356 @@ +# On Spectral Properties of Gradient-based Explanation Methods + +Amir Mehrpanah1, Erik Englesson1, and Hossein Azizpour1 + +Department of Computer Science, Royal Institute of Technology, Stockholm, Sweden {amirme,engless,azizpour}@kth.se + +Abstract. Understanding the behavior of deep networks is crucial to increase our confidence in their results. Despite an extensive body of work for explaining their predictions, researchers have faced reliability issues, which can be attributed to insufficient formalism. In our research, we adopt novel probabilistic and spectral perspectives to formally analyze explanation methods. Our study reveals a pervasive spectral bias stemming from the use of gradient, and sheds light on some common design choices that have been discovered experimentally, in particular, the use of squared gradient and input perturbation. We further characterize how the choice of perturbation hyperparameters in explanation methods, such as SmoothGrad, can lead to inconsistent explanations and introduce two remedies based on our proposed formalism: (i) a mechanism to determine a standard perturbation scale, and (ii) an aggregation method which we call SpectralLens. Finally, we substantiate our theoretical results through quantitative evaluations. + +Keywords: Probabilistic Machine Learning $\cdot$ Gradient-based Explanation Methods $\cdot$ Probabilistic Pixel Attribution Techniques $\cdot$ Explainability $\cdot$ Deep Neural Networks $\cdot$ Spectral Analysis + +# 1 Introduction + +Ensuring the interpretability of machine learning models is crucial for user's trust and understanding [2, 6, 45]. While considerable efforts have been dedicated to advancing explainability in the field [37, 47, 49, 51, 57, 62], there is a notable gap in systematically formalizing and consolidating existing knowledge [5, 9, 16, 19]. Insufficient formal analysis leaves too much room for different interpretations of the explanations, making them unreliable [27] which is detrimental to the very use case of explanations, i.e., human-critical applications. Therefore, more attention is essential to retrospectively analyze and structure the wealth of work done, creating a cohesive foundation for explainability [39, 61]. + +One main reason for the unreliability of explanation methods is their susceptibility to hyperparameters, leading to inconsistent explanations [4, 13, 33, 36, 48], see Fig. 1. Among the several works that formally study explanation methods [19, 37, 57], to our knowledge, none shed light on such inconsistencies. Therefore, in this work, we adopt a novel probabilistic and spectral perspective to characterize the observed inconsistencies and propose preliminary solutions. + +![](images/0e68eef4b31cc9eeeeb66fd83a96ef769983ba8c769329e421842f6b592caf38.jpg) +(a) + +![](images/6790d550b909fe52e3755320774161cf06857cfd3ca86a9329e5809e880bef88.jpg) +(b) + +![](images/03c52b4bd53b511552856ccfd5f167b88f3d2a9db724060736f777498edc1087.jpg) +(c) +Fig. 1: The figure illustrates sample explanations generated by varying hyperparameters of SmoothGrad-Squared [30, 55] for ResNet-50 on ImageNet (see Fig. 16 in the Appendix for more examples). It is evident from the figure that using different hyperparameters results in diverse explanations. Such variations have been demonstrated in recent studies for other explanation techniques as well [4, 13, 33, 48]. + +![](images/43820e97b3834c99683d7a63a41c4cd1303e12cb49f6ea9acd56f4234bf3ec06.jpg) +(d) + +Our Contribution. We iterate our contributions in three parts: + +- Analysis: We introduce a probabilistic representation for post-hoc gradient-based explanation methods in Sec. 2.1, and perform a spectral analysis of them in Sec. 3.1. +- Identified Issues: Our examination reveals a shared spectral bias stemming from the fundamental assumption of utilizing gradients to measure feature contributions. We show that the combined spectral effects of gradient and perturbation lead to band-pass filters modulating the explanations, which could lead to inconsistencies in the explanations, discussed in Sec. 3.2. Moreover, we conduct a spectral analysis of a few specific perturbation distributions that have been proposed in the literature in Sec. 3.3. +- Solutions and Evaluations: To address inconsistencies, we investigate two straightforward solutions that are finding an optimal perturbation scale based on the cosine similarity of the perturbation and the classifier in the frequency domain in Sec. 4.1, and an aggregation mechanism named SpectralLens in Sec. 4.2. Finally, we evaluate the proposed methods using the pixel removal strategy [48], acknowledging their effectiveness and limitations under Sec. 5. + +# 2 Background + +Notation. We consider a classification setting, that is, modeling $p(y|\pmb{x})$ for labels $y \in \{1,\dots,k\}$ and $\pmb{x} \in \mathbb{R}^n$ the set of input images. We denote the $\log p(y|\pmb{x})$ by $f$ . We denote standard Fourier bases by $\{\psi(\pmb{\omega},\pmb{x})\}_{\pmb{\omega} \in \mathbb{R}^n}$ , according to which, under some regularity conditions [43], $f$ can be decomposed as $f(\pmb{x}) = \int \widehat{f}(\pmb{\omega}) \psi(\pmb{\omega},\pmb{x}) d\pmb{\omega}$ . In this representation, $\widehat{f}$ denotes the Fourier transform (FT) of $f$ . We denote the power spectral density (PSD) of $f$ by $S_f(\pmb{\omega}) = |\widehat{f}(\pmb{\omega})|^2$ . + +# 2.1 A Probabilistic Representation of Explanations + +In this view, each explanation method consists of three components: (1) a distribution for perturbation of the input, (2) an explainer, and (3) a summary statistic. We use $p(\tilde{\pmb{x}} | \pmb{x}, \sigma)$ to denote the distribution of perturbations of the input sample parameterized by $\sigma$ , referred to as the perturbation scale. A definition of "explanation" is recovering the contribution of features to a prediction, which emanates from human's natural counterfactual reasoning [19]. Therefore, the explainer is usually the gradient with respect to input $\nabla_{\pmb{x}}$ , or approximations of it, e.g. the finite difference, given by $\Delta_{\pmb{x}} f(\pmb{x}) = \frac{f(\tilde{\pmb{x}}) - f(\pmb{x})}{\tilde{\pmb{x}} - \pmb{x}}$ (see Appendix D). Finally, we assume that the summary statistic is the expectation of the outputs, or the squared outputs, of the explainer, e.g. $(\nabla f(\tilde{\pmb{x}}))^2$ , over the chosen perturbation distribution, $p(\tilde{\pmb{x}} | \pmb{x}, \sigma)$ , denoted as $\mathbb{E}_{p(\tilde{\pmb{x}})}\left[(\nabla f(\tilde{\pmb{x}}))^2\right]$ , where the operations are carried out pixelwise. This is referred to as the explanation of $f$ at $\pmb{x}$ with a hyperparameter $\sigma$ . This probabilistic view can, in fact, unify different explanation methods into a single form (see Appendix C). We provide theoretical justifications for why using squared gradients is preferable to gradients, explaining the empirical findings [30], and will discuss both cases in Sec. 3.1. + +# 3 Spectral Analysis of Explanations + +This section demonstrates the spectral property in explanations when using the gradient explainer and its interaction with perturbations, where a significant difference in the frequency of learned features leads to a selective attribution of features. In the following sections, we draw on insights from spectral analysis of signals and systems theory, utilizing this viewpoint to gain a deeper understanding of explanation methods. We demonstrate a spectral bias related to gradient-based explanation approaches, identify its source, and interpret various methods in the literature as efforts to alleviate it through the incorporation of low-pass filters, either explicitly [55, 57] or implicitly [42, 62]. We illustrate that the inconsistencies in explanation techniques stem from different strategies for mitigating the spectral bias. Finally, we suggest two straightforward approaches to tackle these inconsistencies: determining an optimal perturbation scale in Sec. 4.1 and aggregating the explanations in Sec. 4.2. + +Writing a spectral representation for prediction, as shown in Appendix A.1, the model output with a perturbation distribution can be written as a convolution between the function and the perturbation distribution $f * p$ , therefore we may use "perturbation kernel" and "perturbation distribution" interchangeably. + +# 3.1 Spectral Representation of Explainers + +Here, by drawing a connection between explainability techniques and the literature on spectral signal processing [12, 38, 56], we take a novel avenue for our analysis. This connection enables us to have a spectral representation of the explanations, as discussed in the next sections (see Appendix A for the derivations). It is important to note that the Fourier bases in our work is 1D and + +per pixel, and should not be confused with the traditional Fourier bases used in image processing literature that expands along spatial dimensions. + +Gradients in Spectral Domain. For explanation methods that use gradients and a perturbation kernel, $p(\tilde{\pmb{x}})$ we have the following (see Appendix A.2). + +$$ +\begin{array}{l} \mathbb {E} _ {p (\tilde {\boldsymbol {x}})} [ \nabla f (\tilde {\boldsymbol {x}}) ] = \int i 2 \pi \omega \widehat {f} (\omega) \widehat {p} (\omega) d \omega (1) \\ = - 2 \pi \left[ \int \omega \left(\widehat {f} _ {E} (\omega) \widehat {p} _ {O} (\omega) + \widehat {f} _ {O} (\omega) \widehat {p} _ {E} (\omega)\right) d \omega \right] (2) \\ \end{array} +$$ + +where $\widehat{p}$ denotes the FT of the perturbation kernel. We used $\widehat{f}_E$ and $\widehat{f}_O$ to denote a function $\widehat{f}$ is decomposed into real even and imaginary odd components $\widehat{f} = i\widehat{f}_O + \widehat{f}_E^1$ . The negative sign suggests that in case of using the gradients, we need to negate the actual values to obtain rankings. We will elaborate more on the negation of the attributions in Sec. 5. + +Remark 1. Note the symmetry in the form of the spectral representation of the gradients. This form implies that it is inevitable for any perturbation kernel to extract asymmetric information about even and odd parts of the classifier $f$ . Therefore, the gradient explainer with a kernel symmetric about the sample being explained, only extracts information about the odd parts of the classifier. This is problematic because the attribution is expected to reflect the total contribution to the prediction. + +We will use this interesting symmetry later in Sec. 3.3 for analysis of a specific perturbation kernel. + +Squared Gradients in Spectral Domain. On the other hand, for explanation methods which use squared gradients, we have the following (see Appendix A.3). + +$$ +\mathbb {E} _ {p (\tilde {\boldsymbol {x}})} \left[ (\nabla f (\tilde {\boldsymbol {x}})) ^ {2} \right] = 4 \pi^ {2} \int \| \boldsymbol {\omega} \| ^ {2} \mathrm {S} _ {f} (\boldsymbol {\omega}) \mathrm {S} _ {\sqrt {p}} (\boldsymbol {\omega}) d \boldsymbol {\omega} \tag {3} +$$ + +where $\mathrm{S}_f$ and $\mathrm{S}_{\sqrt{p}}$ denote the PSD of the model and square root of the perturbation kernel respectively. + +We discuss the benefits of using squared gradients as in Eq. (3) over gradients as in Eq. (2) later in Sec. 3.3. However, it can be noticed here that the squared gradients, unlike gradients, are always non-negative, extract information about the PSD of $f$ , and do not have a complex symmetric structure. + +# 3.2 Gradient Explainers with Perturbations Are Band-Pass Filters + +In our probabilistic framework, various explanation methods can be decomposed into three components (1) an explainer, (2) a perturbation kernel and (3) a summary statistic. In this section, we analyze the spectral effects of the two core components, namely gradient and perturbation. Importantly, we show that gradient acts as a high-pass filter, and perturbation acts as a low-pass filter. Before that, we need a more formal definition for a high-pass and low-pass filter that suits our study. Assuming that $\widehat{m} (\omega)$ denotes the FT of a perturbation kernel, we define low- and high-pass as follows. + +Definition 1. A filter $\widehat{m}(\omega)$ is low pass, if for a given $M > 0$ there exists a threshold $\omega^{*}$ such that for any $\|\omega\| > \|\omega^{*}\|$ , we have $\|\widehat{m}(\omega)\| < M$ , and it is high pass if such a condition holds for $\|\omega\| < \|\omega^{*}\|$ . + +For better readability, the proofs for the theorems and propositions are outlined in Appendix B, and we provide intuition as we progress. + +Theorem 1. Any perturbation kernel $p(\tilde{\pmb{x}})$ mitigates the attribution of high-frequency features; simply put, any perturbation is a low-pass filter. + +The intuition is that any modification to a signal, would only remove information, and, as the intensity of perturbation increases, the removal goes from high-frequency components toward lower frequencies. + +Theorem 2. The gradient operator amplifies the attribution of high-frequency features; simply put, the gradient operator is a high-pass filter. + +From the stated theorems, one can see that a gradient-based explanation with perturbation creates a band-pass filter, whose shape is controlled by the perturbation hyperparameters (see Figs. 2 and 3). + +Remark 2. This amplification-mitigation trade-off can be seen in several explanation methods, where perturbation kernels (low-pass filters) are utilized to mitigate the amplification effect (the high-pass filter) caused by the gradient. Therefore, the combined effect of the two filters usually forms a band-pass filter (see Appendix C for a non-comprehensive list). + +# 3.3 Spectral Analysis of Specific Perturbation Kernels + +In this section, we analyze the properties of special kernels used in the literature with remarks or propositions that justify, and thus shed light on, previous work's experimental observations. In most cases, we expand upon the squared gradients, Eq. (3), due to its analytical simplicity. + +VanillaGrad via a Dirac Kernel. The very naive explanation method can be obtained with a Dirac kernel $p(\tilde{\pmb{x}}) = \delta_{\pmb{x}}(\tilde{\pmb{x}})$ known as VanillaGrad (VG). Analyzing the squared version of VG and following Eq. (3), we know that the FT of the Dirac kernel is $S_{\sqrt{p}}(\omega) = 1(\omega)$ , where $\mathbf{1}(\omega)$ is the constant function mapping all values to 1. Thus, it can be seen that we recover the scaled PSD of $f$ , i.e. $\| \boldsymbol{\omega} \|^2 \mathrm{S}_f(\boldsymbol{\omega})$ . + +Remark 3. Scaling $\widehat{f}$ by frequency can be interpreted as the bias of the gradient explainer towards exaggeration of the attribution of high-frequency features [31]. Beside other justifications for gradient sparsity [8], spectral representation of $\mathrm{VG}^2$ and domination of high-frequency features give an orthogonal justification for why VanillaGrad explanations being usually scattered and noisy (see Fig. 12). + +SmoothGrad via a Gaussian Kernel. Letting $p(\tilde{\pmb{x}}) = \mathcal{N}(\pmb{x},\sigma^2)$ , i.e. a Gaussian with variance $\sigma^2$ , centered on the sample being explained, we recover variants of an explanation method known as SmoothGrad (SG) or the squared gradient version SmoothGrad-Squared (SG²) [30, 55], we can rewrite Eq. (3) as follows: + +$$ +\mathbb {E} _ {p (\tilde {\boldsymbol {x}})} \left[ (\nabla f (\tilde {\boldsymbol {x}})) ^ {2} \right] \propto \int \| \boldsymbol {\omega} \| ^ {2} \mathrm {S} _ {f} (\boldsymbol {\omega}) e ^ {- 8 \pi^ {2} \sigma^ {2} \| \boldsymbol {\omega} \| ^ {2}} d \boldsymbol {\omega} \tag {4} +$$ + +This shows that the contributions $\mathrm{S}_f(\omega)$ are modulated by the combined spectral effect of the Gaussian perturbation kernel and gradient operator, which is a band pass filter of the form $\| \boldsymbol{\omega} \|^2 e^{-8\pi^2 \sigma^2} \| \boldsymbol{\omega} \|^2$ , whose shape is controlled by the kernel variance (see Fig. 2). The explanations could be inconsistent, but generally become smoother as $\sigma$ increases, which is a result of the shrinkage of the attribution of high-frequency features that can dominate the explanations with a small $\sigma$ (see Fig. 1). + +![](images/b31ae8acbda21c62db3680ca167d5e8643f9577041981134d95a585adb9f706a.jpg) +Fig.2: A visualization of frequency bands, where $\mathrm{S}_{\sqrt{p}}(\omega)$ is defined according to Eq. (4), show that change in the noise scale $\sigma$ leads to band-pass filters with different modes, hence, representing information of different frequencies. + +When generating explanations, one is tempted to look for hyperparameters that make the explanation reflect what they think it should [16]. However, monitoring explanations from only one frequency band is potentially misleading and is likely to contradict itself simply by changing hyperparameters (see Fig. 5) [27,33]. + +![](images/67c2e0419ad1f1d36380c85575144747ffdd4dd7dc6e82bf1f4f862896522cfb.jpg) +(a) + +![](images/53bc8bee8d3cd47038cd74474421e58ef512951a596f609f1b8f0f684b07c2b0.jpg) +(b) + +![](images/ad1af17a5be4a7a8c28d04ec78998923adf17d112e888a31077027331948287b.jpg) +(c) +Fig. 3: A 1D visualization of amplification (b) and mitigation (d) effects caused by gradient and perturbation, respectively. For simplicity, we have shown a rectangular pulse (c) convolved with $\nabla f$ , but the kernel shape is usually more complex in practice, an example shown in Fig. 2. Moreover, functions in blue are the Fourier representation of the black functions. Due to the gradient operator, small high-frequency jitters in the decision boundary (a) amplify proportionally to their frequency (b). To remove the effect of high-frequency components of the decision boundary that usually dominate the explanations, researchers have proposed different perturbation kernels, visualized in (d). Theorem 2 corresponds to $(\mathrm{a}\rightarrow \mathrm{b})$ and Theorem 1 corresponds to $(\mathrm{b}\rightarrow \mathrm{d})$ . + +We formalize this statement, also known as the Rashomon effect (see Appendix B for the proof). + +Proposition 1. The SmoothGrad-Squared pixel attribution can contradict itself by changing the hyperparameters. Simply, the Rashomon effect can occur in the SmoothGrad-Squared explanations. + +Intuition. This proposition suggests a reason, i.e. removing different frequency bands, for inconsistencies we encounter in $\mathrm{SG}^2$ explanations, which has been unknown so far [48]. Note that this occurs with different kernels, but we have chosen the one that leads to a simpler proof. + +IntegratedGrad via a Rect Kernel. Setting the perturbation kernel to $p(\tilde{\boldsymbol{x}}) = \frac{1}{\sigma|\boldsymbol{s} - \boldsymbol{x}|}\mathrm{Rect}(\frac{\tilde{\boldsymbol{x}} - \boldsymbol{x}}{\sigma(\boldsymbol{s} - \boldsymbol{x})} -\frac{1}{2})^2$ , where $s$ is a chosen baseline image, we recover the perturbation kernel of a well known method called IntegratedGrad (IG) [57]. We use the spectral representation of gradients in Eq. (2), for a proposition about the gradient signs. + +$$ +\mathbb {E} _ {p (\tilde {\boldsymbol {x}})} [ \nabla f (\tilde {\boldsymbol {x}}) ] \propto \int i \widehat {f} (\boldsymbol {\omega}) \sin (\pi \sigma (\boldsymbol {s} - \boldsymbol {x}) \boldsymbol {\omega}) e ^ {i \pi \boldsymbol {\omega} \boldsymbol {r}} d \boldsymbol {\omega} \tag {5} +$$ + +Where $\boldsymbol{r} = (1 + 2\frac{\boldsymbol{x}}{\sigma(\boldsymbol{s} - \boldsymbol{x})})$ , appears because the kernel is not centered on the sample being explained. Note that the Fourier operator linearly expands along each dimension, but we use vector forms for notational simplicity. Furthermore, to + +study the effect of elementwise multiplication on the explanation quantitatively, we use IG to denote integrated gradients without the elementwise multiplication with input and XIG to denote the version with elementwise multiplication. + +In explanation visualizations, the negative gradient sign has been used to show a negative contribution to the prediction [7, 53], to our knowledge it has never been formally justified. In the next proposition, we state that the negative gradient sign can be the direct result of the chosen perturbation kernel, formally shown in Appendix B. + +Proposition 2. The gradient sign in $IG$ is affected by the perturbation level. + +This proposition entails that the negative sign in the gradient explainer does not necessarily reflect a negative contribution to the prediction, nor the positive sign necessarily reflects a positive contribution (see Figs. 13 and 14). + +Remark 4. Since we assume continuity wrt $\sigma$ , in our analysis, by Bolzano's theorem [50], there are hyperparameters that set the attribution to zero. Therefore, the gradient explainer may not produce a useful ranking compared to that of the gradient squared. + +Remark 5. Proposition 2 also shows that adding explanations with different hyperparameters may cause destructive interference when the different values of $\omega \widehat{p} (\omega)$ are out of phase. Hence, squared gradients is a more appropriate design choice, unless the use of phase information is planned as part of the method. This has been discovered experimentally [30] and confirmed in Sec. 5. + +# 4 Proposed Solutions Based on the Spectral Analysis + +# 4.1 Finding an Optimal Perturbation Scale + +Note that the squared gradients in spectral domain (Eq. (3)) can be formulated as an inner product between the scaled PSD of the perturbation distribution and the PSD of the classifier i.e. $\mathbb{E}_{p(\tilde{\boldsymbol{x}})}\left[(\nabla f(\tilde{\boldsymbol{x}}))^2\right] = 4\pi^2\langle \mathrm{S}_f,\| \boldsymbol {\omega}\| ^2\mathrm{S}_{\sqrt{p}}\rangle$ Assuming unimodality, a nonzero contribution, and that the norm $\left\| \left\| \boldsymbol {\omega}\right\| ^2\mathrm{S}_{\sqrt{p}}\right\|$ exists, then the cosine similarity is maximized at a certain perturbation scale that has the highest similarity with $\mathrm{S}_f$ . A higher cosine similarity guarantees recovering more information about the classifier. + +Definition 2. The cosine similarity between the PSD of the classifier and that of the perturbation kernel given by + +$$ +\text {S i m i l a r i t y} \left(\mathrm {S} _ {f}, \| \boldsymbol {\omega} \| ^ {2} \mathrm {S} _ {\sqrt {p}}\right) = 4 \pi^ {2} \frac {\left\langle \mathrm {S} _ {f} , \| \boldsymbol {\omega} \| ^ {2} \mathrm {S} _ {\sqrt {p}} \right\rangle}{\| \mathrm {S} _ {f} \| \left\| \| \boldsymbol {\omega} \| ^ {2} \mathrm {S} _ {\sqrt {p}} \right\|} \tag {6} +$$ + +This quantity can be used as a way to determine the optimal perturbation scale per pixel to maximize the information extracted with the squared gradients. Note that, the word "optimal" refers to extracting the most information by the inner product, whether deemed relevant by a human or not.. + +For one of the perturbation kernels we have presented earlier in Sec. 3.3, we have optimized this quantity on multiple levels, such as dataset level Fig. 4, image level Tab. 1, and, pixel level Fig. 5d. Finally, we have experimentally verified the effectiveness of this approach in Sec. 5. + +Computing the Similarity of Classifier and Gaussian Kernel. To compute the similarity, we first need to evaluate the kernel norm as follows + +$$ +\left\| \left\| \boldsymbol {\omega} \right\| ^ {2} e ^ {- 8 \pi^ {2} \sigma^ {2} \left\| \boldsymbol {\omega} \right\| ^ {2}} \right\| = \left(\int \left\| \boldsymbol {\omega} \right\| ^ {4} e ^ {- 1 6 \pi^ {2} \sigma^ {2} \left\| \boldsymbol {\omega} \right\| ^ {2}} d \boldsymbol {\omega}\right) ^ {\frac {1}{2}} \tag {7} +$$ + +$$ +\propto \sigma^ {\frac {- 5}{2}} \tag {8} +$$ + +Therefore, the cosine similarity for the Gaussian perturbation kernel can be computed as follows: + +$$ +\mathrm {S i m i l a r i t y} (\mathrm {S} _ {\sqrt {p}}, \mathrm {S} _ {f}) \propto \sigma^ {\frac {5}{2}} \langle \mathrm {S} _ {f}, \| \boldsymbol {\omega} \| ^ {2} \mathrm {S} _ {\sqrt {p}} \rangle +$$ + +Using the L1 norm for the squared gradients of flattened explanations. We have observed experimentally that $\mathrm{S}_f$ depends on the input scale, hence we use a convex interpolation with noise to support the assumption that $\mathrm{S}_f$ remains constant (see Appendix E.2 and Fig. 8). The results are shown in Fig. 4. + +Remark 6. The optimal perturbation scale, determined by maximizing cosine similarity, offers a standard method for selecting the perturbation scale in $\mathrm{SG^2}$ . Given that user discretion in the selection of perturbation scale contributes to inconsistencies, as illustrated in Fig. 1, opting for the optimal perturbation scale mitigates these inconsistencies in the results. Hence, it represents a more suitable design choice. In Sec. 5 we demonstrate that this approach not only enhances $\mathrm{SG^2}$ 's robustness against inconsistencies but also its adaptability to diverse datasets, which in turn can improve the quantitative performance. + +# 4.2 Aggregation of Information With Explanation Prior + +We have shown that explanations are influenced by specific frequency bands, rather than representing the importance of individual pixels for prediction. Instead, they indicate the contributions within a specific frequency range. This observation motivates us to investigate a method for aggregation of information across different bands, which closely resembles an ensemble of IG and SG [55,57] which we call SpectralLens. + +Definition 3. SpectralLens is an ensemble explanation method, that aggregates information across different frequency bands as follows + +$$ +\mathbb {E} _ {p (\sigma)} \left[ \mathbb {E} _ {p (\tilde {\boldsymbol {x}})} \left[ (\nabla f (\tilde {\boldsymbol {x}})) ^ {2} \right] \right] \propto \iint \mathrm {S} _ {f} (\boldsymbol {\omega}) \| \boldsymbol {\omega} \| ^ {2} e ^ {- 8 \pi^ {2} \sigma \| \boldsymbol {\omega} \| ^ {2}} d \boldsymbol {\omega} p (\sigma) d \sigma \tag {9} +$$ + +Where we refer to this method by $SL^2$ as it is using the squared gradients and we call $p(\sigma)$ as an explanation prior, gauging the frequency bands that we choose to represent in an explanation. + +![](images/58a578b8a285dba16384c1190133078cef1105bae86180afc2741924d8b891df.jpg) +(a) + +![](images/ad50c71b764988b8360a2f65904a4624eee08e61389feaeff4d15c2c8584a998.jpg) +(b) +Fig. 4: Cosine similarity between kernel and classifier, defined in Eq. (6), is measured for the Gaussian perturbation kernel and ResNet50 averaged on the 10K images of the ImageNet validation set (a) and Food101 (b). This metric assesses the similarity between the PSD of the perturbation kernel and that of the classifier, indicating the perturbation level's effectiveness for kernels with finite norm in the Fourier domain. When using a Dirac explanation prior in Eq. (9), where a perturbation level is assigned per image or dataset, the optimal choice for the perturbation level is where the similarity is maximized across the image or dataset respectively. Note that this level of noise refers to a band-pass filter, shown in Fig. 2, which suggests the consideration of a latent distribution on the frequency of labeling, reflecting the workforce's mindset during data labeling. + +Remark 7. $\mathrm{SL}^2$ exhibits robustness against inconsistencies in $\mathrm{SG}^2$ due to its aggregation of information from various bands. Therefore, it is a preferable design choice not to employ a Dirac explanation prior with $SG^2$ , as it eliminates inconsistencies, depicted in Fig. 1. Nevertheless, as demonstrated in the Sec. 5, this approach may enhance the quantitative performance of explanations. + +Intuition. Considering Eq. (4), we realize that the combination of perturbation and gradient explainer creates overlapping band-pass carrier waves that modulate the amplitude of the spectral density of the classifier. Therefore, a new summary statistic is required that can reflect a wide range of frequency bands. We explicitly identified the spectral properties of the explanations, recognizing that not all frequencies are equally desired. Explanations with very high frequencies primarily attribute predictions to pixels. Therefore, we utilize an explanation prior $p(\sigma)$ for a nuanced control of the spectral properties of explanations. Although there are many possible alternatives for selecting an explanation prior, we used $p(\sigma) = \mathrm{Unif}(0,1)$ for simplicity. + +Definition 4. We define $\mathbf{ArgLens}(AL)$ as the point at which the $SG^2$ is maximized for pixel $x_i$ . Formally defined as + +$$ +\operatorname {A r g L e n s} \left(x _ {i}\right) = \arg \max _ {\sigma} \left\{\mathrm {S G} ^ {2} \left(x _ {i}\right) \right\} \tag {10} +$$ + +where $\sigma$ is the hyperparameter of the perturbation kernel. + +AL shows the frequency band at which the pixel's attribution is maximized. Alternatively, it can be interpreted as the robustness of a pixel to noise. We + +have visualized the AL for the Gaussian kernel in Fig. 5 and discussed the way it helps the interpretation of explanations in Sec. 4.2. + +The aggregation of explanations from different frequencies is visualized in Fig. 5. Moreover, the basic explanation methods are obtainable with a Dirac explanation prior $p(\sigma) = \delta(\sigma)$ . Therefore, it can be used as a unified formulation for gradient-based explanation methods in spectral domain. + +Interesting Properties of SpectralLens and ArgLens. In the absence of a robust mathematical framework, the interpretation of explanations may be subjective. Thus, explanation methods must be grounded in solid theoretical principles. $\mathrm{SL}^2$ addresses potential inconsistencies by aggregating attributions across frequency bands, ensuring fair representation of all pixels' contributions. + +As $\mathrm{SG^2}$ results in bands-pass filters (see Sec. 3.3), its attributions reflect the changes in the decision boundary within a certain frequency band. Furthermore, AL reveals the frequency band at which the attribution is maximized. Hence, lower AL indicates that the attributions are likely governed by high-frequency components of the decision boundary, more sensitive to input noise, consequently less robust. An ideal classifier with smooth decision boundaries, i.e. robust features, leads to high AL everywhere, with a high intensity in $\mathrm{SL^2}$ on relevant pixels and low intensity on irrelevant pixels. + +Given the different purposes of pixels within an image, the frequency of activation may vary and can be visualized via AL. Since $\mathrm{SL}^2$ and AL offer complementary information, they should be interpreted together. This approach offers a deeper understanding compared to a simplistic perspective, as it has been shown by Rahaman et al. [43] that high-frequency regions emerge later in training, are more challenging to fit, and are prone to change. On the contrary, low-frequency components of the decision boundary manifest earlier in training, are easier to fit, and exhibit greater stability with respect to noise. + +It is important to clarify that "high-frequency" refers to features considered high-frequency by the model, and not necessarily frequency across spatial dimensions (e.g. width, height). Nonetheless, empirical findings demonstrate that high-frequency features in spatial domain are more likely to be identified as high frequency by the model (see Appendix I). + +# 5 Evaluation + +Given the experimental discoveries that squared gradients lead to better results [30], we evaluate both scenarios for different hyperparameters and compare them with SL in Tabs. 1 and 2, refer to Appendix E for implementation details and reproducibility. For the purposes of this study, we use pixel removal strategy, despite its problems [30]. The spectral analysis of pixel removal strategy and possible solutions are deferred to future research. + +We report single-pixel most-important-first (MoRF) insertion and deletion scores of the methods to assess the design choices previously proposed by our theoretical framework: (1) utilizing squared gradients instead of gradients in + +![](images/748a8d9134f66ebad795cb6170cdd6324a4fdf8226acf661c91c6173a4a42d83.jpg) + +![](images/1bf4a60f123842bd0e33e7d4229d36537730d83575dba59526babc51101c6d06.jpg) + +![](images/772a0ff07b8a7e44411c06fca752f52abb2b569a465fee1bb7ca4c577914c191.jpg) + +![](images/f61bc8e42c32c6a0ce10e697499b84b7284c279125c24493aa109e49a85ac190.jpg) + +![](images/cc40f30369167a69d73bdea5a27af10b38294dbbbd1c1455353f9d1b9ae6bd87.jpg) +(a) + +![](images/d707b0fbd240bcc7fda17dcfbe7b08a4d9531c03ec08fe624b9e2d35a2aaeb4d.jpg) +(b) + +![](images/1a932a9c1aeb327b28a9292702767225f034d638bae2be886111fe4f8591b6a8.jpg) +(c) +Fig. 5: This figure shows example visualizations of explanations generated by $\mathrm{SG^2}$ implementing design decisions suggested in our work in Remarks 6 and 7. (b) shows the $\mathrm{SG^2}$ visualized at the maximum similarity according to Fig. 4, for this image $\sigma = 0.5$ . (c) shows the SL attributions. Lastly, (d) shows the frequencies obtained from AL (defined in Eq. (10)), which correspond to the perturbation level at which $\mathrm{SG^2}$ is maximized. Since we are interested in frequencies rather than the actual perturbation levels, we visualize 2 stds above and below the mean of frequencies obtained by $\omega = 1 / (1 + \mathrm{AL})$ (see Appendix I for more examples). The heatmaps in (b) and (c) show how much the model is relying on the pixel, while (d) shows the frequency at which the feature's contribution maximizes. In the provided examples, we can see that the model is relying on the nose (top) and eye (bottom) while the same region is of a lower frequency compared to the head. This can be interpreted as the model finding the nose and eye more stable features compared to the head. + +![](images/961e15fc803ece74636792d3c11d15aa26594b32db6c08462fd39ef4337e4f08.jpg) +(d) + +Remark 5, (2) finding a standard perturbation scale in Remark 6, and (3) employing a uniform explanation prior instead of Dirac, stated in Remark 7 (see Fig. 7 for an ablation study of kernel variance and Appendix G for an ablation study of patch removal instead of single pixel removal). We evaluated the suggested design choices on ImageNet [20] and Food101 [1] in Tab. 1. However, smaller datasets, such as CIFAR, are omitted because their resolution restricts the potential gap between high- and low-frequency features. + +Limitations of Our Framework. The choice of independent Gaussian perturbation is limited due to its inability to incorporate prior knowledge about the input space or data manifold characteristics, and other data modalities such as graphs and time series. However, preferring simpler analytical forms within our framework limits the alternative choices. Efficiency is another challenge compared to basic methods, as performing back-props at various noise scales is costly. Moreover, our attribution of frequencies in visualizations, such as SL, may lack the desired generalizability across data points, as high-frequency re + +Table 1: This table presents the assessment of the proposed design choices in Remarks 5 to 7 across different methods, highlighting statistically significant improvements with bold fonts. We conduct t-tests comparing gradient methods to their squared-gradient counterparts to evaluate the design choice in Remark 5. Furthermore, we conduct t-tests between $\mathrm{SG}^2$ with the original heuristic and those ensuring consistency to evaluate design decisions in Remarks 6 and 7. Overlines and underlines denote highest or lowest values within columns, respectively. Although there are no squared versions of Occlusion Sensitivity (OS) [5] and RISE [42], they are included as low-frequency baselines. While the dimensions of boolean masks utilized in OS and RISE can implicitly regulate the sampling variance (see Appendix C), the exponential increase in the required number of samples in high-frequency explanations renders them less practical. SG scores follow the heuristic of the SmoothGrad paper [55] i.e. $\sigma = (\max(x) - \min(x))$ 0.1. XIG indicates the input multiplication in the original implementation of IntegratedGrad [57], while IG does not apply such transformation. Rankings for methods using gradients are obtained from negative gradients, including RISE and Occlusion Sensitivity, according to Eq. (2). Furthermore, $\mathrm{SG}_{\mathrm{Opt.}}^2$ uses an optimized kernel per image and $\mathrm{SL}^2$ uses a uniform explanation prior. Comparing SG² and $\mathrm{SG}_{\mathrm{Opt.}}^2$ , we can see that $\mathrm{SG}_{\mathrm{Opt.}}^2$ can adapt to the statistics of the Food101. Therefore, $\mathrm{SG}_{\mathrm{Opt.}}^2$ , while not subject to inconsistencies (shown in Fig. 1), can improve the performance due to its adaptability. Hence, we can ensure the consistency of SG² without compromising its performance. Also, note that the input multiplication enhances visual quality, but degrades the performance. + +
ImageNetFood101
Insertion(↑)Deletion(↓)Insertion(↑)Deletion(↓)
GradientsOS0.384 ± 0.0080.237 ± 0.0100.378 ± 0.0110.209 ± 0.009
RISE0.494 ± 0.0140.174 ± 0.0070.410 ± 0.0130.198 ± 0.007
SG0.212 ± 0.0100.179 ± 0.0090.179 ± 0.0080.137 ± 0.007
XIG0.208 ± 0.0090.204 ± 0.0080.101 ± 0.0070.177 ± 0.007
IG0.201 ± 0.0100.182 ± 0.0090.145 ± 0.0060.131 ± 0.005
Gradients2SG20.465 ± 0.0130.157 ± 0.0070.371 ± 0.0120.162 ± 0.006
XIG20.334 ± 0.0140.187 ± 0.0080.204 ± 0.0100.263 ± 0.007
IG20.340 ± 0.0120.168 ± 0.0070.320 ± 0.0090.135 ± 0.005
Gradients2Opt.0.491 ± 0.0150.187 ± 0.0070.394 ± 0.0130.167 ± 0.006
SL20.493 ± 0.0160.169 ± 0.0070.395 ± 0.0140.157 ± 0.007
+ +gions in one image may correspond to low-frequency regions in another image (see Appendix I). Identifying generalizable attributions across images remains a subject for future investigation. + +# 6 Related Works + +Probabilistic and Spectral View of Explanations. Despite sporadic attempts to analyze explanations from a probabilistic view [17, 59], this view is more common in the literature around the uncertainty quantification of expla + +nations [29,40,54]. Our work is closely related to Covert et al. [19], if their work is seen from a probabilistic perspective. Aside from the works around the spectral view of deep networks [10,11,15,43,58], as of our current knowledge, there is no work adopting a spectral view of explanations along pixels' value. Nonetheless, there are works that build on a spectral view along spatial dimensions [34,35], extending an optimization-based definition of explanations [22]. + +Unifying Views to Explanations. Covert et al. [19] propose an information theoretic framework that unifies several explanation methods, focusing on three design choices: (1) feature removal method, (2) which model behavior to analyze, and (3) summary statistics. Our probabilistic framework, being similar to Covert's, offers a distinct viewpoint relevant to our spectral analysis. It can be used as a unifying view of explanations (see Appendix C), inheriting from Covert's framework. Notably, our work aligns with the broader context of research under finding unifying representations of explanations [3,26,37]. + +Input-space or Function-space Perturbations. Methods such as NoiseGrad [14], Grad-CAM [51], RectifiedGrad [32], Layer-wise Relevance Propagation [7], DeepLift [53], and even Sharpness-Aware Minimization (SAM) [24] employ function space perturbations. Utilizing such distributions offers data modality independence, whereas input space perturbation remains agnostic about the model architecture. While acknowledging the duality between function and input space distributions, our focus on explicit input space distributions excludes these methods from our study. Although we recognize the utility of integrating function space perturbation into our mathematical model, we defer this aspect to future investigations. It can be noted, however, that extending this work to function space perturbation will connect it to the literature on double descent and smooth interpolation of samples [25, 41]. + +# 7 Conclusion + +In this study, we deepen understanding of pixel attribution methodologies by rigorously analyzing their theoretical underpinnings, revealing a common spectral bias. Our analysis supports the experimental finding of recent works regarding the efficacy of using squared gradients. Moreover, we demonstrate how spectral properties can lead to inconsistent explanations and propose two design choices informed by our theoretical framework to address these inconsistencies: standardizing hyperparameters based on cosine similarity between the perturbation kernel and classifier, and aggregating explanations with a given prior. Despite the effectiveness of our approach, we acknowledge limitations and suggest avenues for further research. Lastly, we assess the suggested design choices using deletion and insertion scores. + +# Acknowledgements + +We would like to thank Giovanni Luca Marchetti for an early review of our work and his kind feedback. This project is partially supported by Region Stockholm through MedTechLabs, and Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Scientific computation was enabled by the supercomputing resource Berzelius, provided by the National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg foundation. + +# References + +1. Food-101 - Mining Discriminative Components with Random Forests, https://data.vision.ee.ethz.ch/cvl/datasets-extra/food-101/ +2. Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., Kim, B.: Sanity Checks for Saliency Maps (Nov 2020). https://doi.org/10.48550/arXiv.1810.03292, http://arxiv.org/abs/1810.03292, arXiv:1810.03292 [cs, stat] +3. Agarwal, S., Jabbari, S., Agarwal, C., Upadhyay, S., Wu, Z.S., Lakkaraju, H.: Towards the Unification and Robustness of Perturbation and Gradient Based Explanations (Jul 2021). https://doi.org/10.48550/arXiv.2102.10618, http://arxiv.org/abs/2102.10618, arXiv:2102.10618 [cs] +4. Alvarez-Melis, D., Jaakkola, T.S.: On the Robustness of Interpretability Methods (Jun 2018). https://doi.org/10.48550/arXiv.1806.08049, http://arxiv.org/abs/1806.08049, arXiv:1806.08049 [cs, stat] +5. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for Deep Neural Networks (Mar 2018). https://doi.org/10.48550/arXiv.1711.06104, http://arxiv.org/abs/1711.06104, arXiv:1711.06104 [cs, stat] +6. Arrieta, A.B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., Chatila, R., Herrera, F.: Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI (Dec 2019). https://doi.org/10.48550/arXiv.1910.10045, http://arxiv.org/abs/1910.10045, arXiv:1910.10045 [cs] +7. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS ONE 10(7), e0130140 (Jul 2015). https://doi.org/10.1371/journal.pone.0130140, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0130140 +8. Balduzzi, D., Frean, M., Leary, L., Lewis, J.P., Ma, K.W.D., McWilliams, B.: The Shattered Gradients Problem: If resnets are the answer, then what is the question? (Jun 2018). https://doi.org/10.48550/arXiv.1702.08591, http://arxiv.org/abs/1702.08591, arXiv:1702.08591 [cs, stat] +9. Bansal, N., Agarwal, C., Nguyen, A.: SAM: The Sensitivity of Attribution Methods to Hyperparameters (Apr 2020). https://doi.org/10.48550/arXiv.2003.08754, http://arxiv.org/abs/2003.08754, arXiv:2003.08754 [cs] +10. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency Bias in Neural Networks for Input of Non-Uniform Density (Mar 2020), https://api-semanticscholar.org/CorpusID:212644664 + +11. Benbarka, N., Hofer, T., Ul-Moqeet Riaz, H., Zell, A.: Seeing Implicit Neural Representations as Fourier Series. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 2283-2292 (Jan 2022). https://doi.org/10.1109/WACV51458.2022.00234, https://ieeexplore.ieee.org/document/9706795/ +12. Blackman, R.B., Tukey, J.W.: The measurement of power spectra from the point of view of communications engineering — Part I. The Bell System Technical Journal 37(1), 185–282 (Jan 1958). https://doi.org/10.1002/j.1538-7305.1958.tb03874.x, http://ieeexplore.ieee.org/document/6768513/ +13. Brughmans, D., Melis, L., Martens, D.: Disagreement amongst counterfactual explanations: How transparency can be deceptive (Apr 2023). https://doi.org/10.48550/arXiv.2304.12667, http://arxiv.org/abs/2304.12667 [cs] +14. Bykov, K., Hedström, A., Nakajima, S., Höhne, M.M.C.: NoiseGrad: Enhancing Explanations by Introducing Stochasticity to Model Weights (May 2022). https://doi.org/10.48550/arXiv.2106.10185, http://arxiv.org/abs/2106.10185, arXiv:2106.10185 [cs] +15. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards Understanding the Spectral Bias of Deep Learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. pp. 2205-2211. International Joint Conferences on Artificial Intelligence Organization, Montreal, Canada (Aug 2021). https://doi.org/10.24963/ijcai.2021/304, https://www.ijcai.org/proceedings/2021/304 +16. Chen, H., Janizek, J.D., Lundberg, S., Lee, S.I.: True to the Model or True to the Data? (Jun 2020). https://doi.org/10.48550/arXiv.2006.16234, http://arxiv.org/abs/2006.16234, arXiv:2006.16234 [cs, stat] +17. Chen, J., Song, L., Wainwright, M.J., Jordan, M.I.: Learning to Explain: An Information-Theoretic Perspective on Model Interpretation (Jun 2018). https://doi.org/10.48550/arXiv.1802.07814, http://arxiv.org/abs/1802.07814, arXiv:1802.07814 [cs, stat] +18. Cohen, L.: The generalization of the Wiener-Khinchin theorem. In: Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181). vol. 3, pp. 1577-1580. IEEE (1998), https://ieeexplore.ieee.org/abstract/document/681753/ +19. Covert, I., Lundberg, S., Lee, S.I.: Explaining by Removing: A Unified Framework for Model Explanation (May 2022). https://doi.org/10.48550/arXiv.2011.14878, http://arxiv.org/abs/2011.14878, arXiv:2011.14878 [cs, stat] +20. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (Jun 2009). https://doi.org/10.1109/CVPR.2009.5206848, https://ieeexplore.ieee.org/document/5206848, iSSN: 1063-6919 +21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Jun 2021). https://doi.org/10.48550/arXiv.2010.11929, http://arxiv.org/abs/2010.11929, arXiv:2010.11929 [cs] +22. Fong, R., Patrick, M., Vedaldi, A.: Understanding Deep Networks via Extremal Perturbations and Smooth Masks (Oct 2019). https://doi.org/10.48550/arXiv.1910.08485, http://arxiv.org/abs/1910.08485, arXiv:1910.08485 [cs, stat] + +23. Fong, R., Vedaldi, A.: Interpretable Explanations of Black Boxes by Meaningful Perturbation. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 3449-3457 (Oct 2017). https://doi.org/10.1109/ICCV.2017.371, http://arxiv.org/abs/1704.03296, arXiv:1704.03296 [cs, stat] +24. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-Aware Minimization for Efficiently Improving Generalization (Apr 2021). https://doi.org/10.48550/arXiv.2010.01412, http://arxiv.org/abs/2010.01412, arXiv:2010.01412 [cs, stat] +25. Gamba, M., Azizpour, H., Björkman, M.: On the Lipschitz Constant of Deep Networks and Double Descent (Nov 2023). https://doi.org/10.48550/arXiv.2301.12309, http://arxiv.org/abs/2301.12309, arXiv:2301.12309 [cs] +26. Han, T., Srinivas, S., Lakkaraju, H.: Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations (Dec 2022). https://doi.org/10.48550/arXiv.2206.01254, http://arxiv.org/abs/2206.01254, arXiv:2206.01254 [cs] +27. Harel, N., Gilad-Bachrach, R., Obolski, U.: Inherent Inconsistencies of Feature Importance (Jun 2022). https://doi.org/10.48550/arXiv.2206.08204, http://arxiv.org/abs/2206.08204, arXiv:2206.08204 [cs] +28. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition (Dec 2015). https://doi.org/10.48550/arXiv.1512.03385, http://arxiv.org/abs/1512.03385, arXiv:1512.03385 [cs] +29. Hill, D., Masoomi, A., Ghimire, S., Torop, M., Dy, J.: Explanation Uncertainty with Decision Boundary Awareness (Nov 2022). https://doi.org/10.48550/arXiv.2210.02419, http://arxiv.org/abs/2210.02419, arXiv:2210.02419 [cs] +30. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A Benchmark for Interpretability Methods in Deep Neural Networks (Nov 2019). https://doi.org/10.48550/arXiv.1806.10758, http://arxiv.org/abs/1806.10758, arXiv:1806.10758 [cs, stat] +31. Hou, X., Zhang, L.: Saliency Detection: A Spectral Residual Approach. In: 2007 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8 (Jun 2007). https://doi.org/10.1109/CVPR.2007.383267, https://ieeexplore.ieee.org/document/4270292, iSSN: 1063-6919 +32. Kim, B., Seo, J., Jeon, S., Koo, J., Choe, J., Jeon, T.: Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps (Sep 2019). https://doi.org/10.48550/arXiv.1902.04893, http://arxiv.org/abs/1902.04893, arXiv:1902.04893 [cs, stat] +33. Kindermans, P.J., Hooker, S., Adebayo, J., Alber, M., Schütt, K.T., Dähne, S., Erhan, D., Kim, B.: The (Un)reliability of saliency methods (Nov 2017). https://doi.org/10.48550/arXiv.1711.00867, http://arxiv.org/abs/1711.00867, arXiv:1711.00867 [cs, stat] +34. Kolek, S., Nguyen, D.A., Levie, R., Bruna, J., Kutyniok, G.: Cartoon Explanations of Image Classifiers (Oct 2022). https://doi.org/10.48550/arXiv.2110.03485, http://arxiv.org/abs/2110.03485, arXiv:2110.03485 [cs] +35. Kolek, S., Windsesheim, R., Loarca, H.A., Kutyniok, G., Levie, R.: Explaining Image Classifiers with Multiscale Directional Image Representation (Apr 2023). https://doi.org/10.48550/arXiv.2211.12857, http://arxiv.org/abs/2211.12857, arXiv:2211.12857 [cs] +36. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The Disagreement Problem in Explainable Machine Learning: A Practitioner's Perspective (Feb 2022). https://doi.org/10.48550/arXiv.2202.01602, http://arxiv.org/abs/2202.01602, arXiv:2202.01602 [cs] + +37. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017) +38. Marmarelis, P.Z., Marmarelis, V.Z.: The White-Noise Method in System Identification. In: Marmarelis, P.Z., Marmarelis, V.Z. (eds.) Analysis of Physiological Systems: The White-Noise Approach, pp. 131-180. Computers in Biology and Medicine, Springer US, Boston, MA (1978). https://doi.org/10.1007/978-1-4613-3970-0_4, https://doi.org/10.1007/978-1-4613-3970-0_4 +39. Marques-Silva, J., Ignatiev, A.: Delivering Trustworthy AI through Formal XAI. Proceedings of the AAAI Conference on Artificial Intelligence 36(11), 12342-12350 (Jun 2022). https://doi.org/10.1609/aaai.v36i11.21499, https://ojs.aaai.org/index.php/AAAI/article/view/21499 +40. Marx, C., Park, Y., Hasson, H., Wang, Y., Ermon, S., Huan, L.: But Are You Sure? An Uncertainty-Aware Perspective on Explainable AI. In: Proceedings of The 26th International Conference on Artificial Intelligence and Statistics. pp. 7375-7391. PMLR (Apr 2023), https://proceedings.mlr.press/v206/marx23a.htm1 +41. Nakkiran, P., Kaplun, G., Bansal, Y., Yang, T., Barak, B., Sutskever, I.: Deep Double Descent: Where Bigger Models and More Data Hurt (Dec 2019). https://doi.org/10.48550/arXiv.1912.02292, http://arxiv.org/abs/1912.02292, arXiv:1912.02292 [cs, stat] +42. Petsiuk, V., Das, A., Saenko, K.: RISE: Randomized Input Sampling for Explanation of Black-box Models (Sep 2018). https://doi.org/10.48550/arXiv.1806.07421, http://arxiv.org/abs/1806.07421, arXiv:1806.07421 [cs] +43. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F.A., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks (May 2019). https://doi.org/10.48550/arXiv.1806.08734, http://arxiv.org/abs/1806.08734, arXiv:1806.08734 [cs, stat] +44. Rahman, M.: Applications of Fourier Transforms to Generalized Functions. WIT Press (2011), google-Books-ID: k_rdcKaUdr4C +45. Ras, G., van Gerven, M., Haselager, P.: Explanation Methods in Deep Learning: Users, Values, Concerns and Challenges. In: Escalante, H.J., Escalera, S., Guyon, I., Baró, X., Güçlütürk, Y., Güçlü, U., van Gerven, M. (eds.) Explainable and Interpretable Models in Computer Vision and Machine Learning, pp. 19-36. The Springer Series on Challenges in Machine Learning, Springer International Publishing, Cham (2018). https://doi.org/10.1007/978-3-319-98131-4_2, https://doi.org/10.1007/978-3-319-98131-4_2 +46. Reiter, H., Stegeman, J.D.: Classical Harmonic Analysis and Locally Compact Groups. Oxford University Press (Nov 2000). https://doi.org/10.1093/oso/9780198511892.001.0001, https://academic.oup.com/book/54460 +47. Ribeiro, M.T., Singh, S., Guestrin, C.: "Why Should I Trust You?": Explaining the Predictions of Any Classifier (Aug 2016). https://doi.org/10.48550/arXiv.1602.04938, http://arxiv.org/abs/1602.04938, arXiv:1602.04938 [cs, stat] +48. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A Consistent and Efficient Evaluation Strategy for Attribution Methods (Jun 2022). https://doi.org/10.48550/arXiv.2202.00449, http://arxiv.org/abs/2202.00449, arXiv:2202.00449 [cs] +49. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence 1(5), 206-215 (May 2019). https://doi.org/10.1038/s42256-019-0048-x, https://www.nature.com/articles/s42256-019-0048-x +50. Rudin, W.: Principles of Mathematical Analysis. McGraw-Hill (1976), google-Books-ID: kwqzPAAACAAJ + +51. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision 128(2), 336-359 (Feb 2020). https://doi.org/10.1007/s11263-019-01228-7, http://arxiv.org/abs/1610.02391, arXiv:1610.02391 [cs] +52. Serov, V.: The Riemann-Lebesgue Lemma. In: Serov, V. (ed.) Fourier Series, Fourier Transform and Their Applications to Mathematical Physics, pp. 33-35. Applied Mathematical Sciences, Springer International Publishing, Cham (2017). https://doi.org/10.1007/978-3-319-65262-7_6, https://doi.org/10.1007/978-3-319-65262-7_6 +53. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not Just a Black Box: Learning Important Features Through Propagating Activation Differences (Apr 2017). https://doi.org/10.48550/arXiv.1605.01713, http://arxiv.org/abs/1605.01713, arXiv:1605.01713 [cs] +54. Slack, D., Hilgard, S., Singh, S., Lakkaraju, H.: Reliable Post hoc Explanations: Modeling Uncertainty in Explainability (Nov 2021). https://doi.org/10.48550/arXiv.2008.05030, http://arxiv.org/abs/2008.05030, arXiv:2008.05030 [cs, stat] +55. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: SmoothGrad: removing noise by adding noise (Jun 2017). https://doi.org/10.48550/arXiv.1706.03825, http://arxiv.org/abs/1706.03825, arXiv:1706.03825 [cs, stat] +56. Stoica, P., Moses, R.L.: Spectral analysis of signals, vol. 452. Pearson Prentice Hall Upper Saddle River, NJ (2005), http://user.it.uu.se/~ps/SAS-new.pdf +57. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic Attribution for Deep Networks (Jun 2017). https://doi.org/10.48550/arXiv.1703.01365, http://arxiv.org/abs/1703.01365, arXiv:1703.01365 [cs] +58. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. ArXiv (Jun 2020), https://api_semanticscholar.org/CorpusID:219791950 +59. Wang, E., Khosravi, P., Broeck, G.V.d.: Probabilistic Sufficient Explanations (May 2021). https://doi.org/10.48550/arXiv.2105.10118, http://arxiv.org/abs/2105.10118, arXiv:2105.10118 [cs] +60. Wang, Y., Wang, X.: "Why Not Other Classes?": Towards Class-Contrastive Backpropagation Explanations. Advances in Neural Information Processing Systems 35, 9085-9097 (Dec 2022), https://proceedings.neurips.cc/paper_files/paper/2022/bitical/3b7a66b2d1258e892c89f485b8f896e0-AAbstract-Conference.html +61. Wilming, R., Kieslich, L., Clark, B., Haufe, S.: Theoretical Behavior of XAI Methods in the Presence of Suppressor Variables (Jun 2023). https://doi.org/10.48550/arXiv.2306.01464, http://arxiv.org/abs/2306.01464 [cs, stat] +62. Zeiler, M.D., Fergus, R.: Visualizing and Understanding Convolutional Networks (Nov 2013). https://doi.org/10.48550/arXiv.1311.2901, http://arxiv.org/abs/1311.2901, arXiv:1311.2901 [cs] \ No newline at end of file diff --git a/2024/On Spectral Properties of Gradient-based Explanation Methods/images.zip b/2024/On Spectral Properties of Gradient-based Explanation Methods/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ff52ae7ce000542fe57fc0f0e5e55d2b5ac5f6b3 --- /dev/null +++ b/2024/On Spectral Properties of Gradient-based Explanation Methods/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf883b6baef84ba51a2cba11f5564c01a1a218d64acc417175e48ab80bf3052d +size 310857 diff --git a/2024/On Spectral Properties of Gradient-based Explanation Methods/layout.json b/2024/On Spectral Properties of Gradient-based Explanation Methods/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9eb46b214f56f3f47670b4cd934ca966ba33739e --- /dev/null +++ b/2024/On Spectral Properties of Gradient-based Explanation Methods/layout.json @@ -0,0 +1,10762 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 111, + 452, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 111, + 452, + 147 + ], + "spans": [ + { + "bbox": [ + 162, + 111, + 452, + 147 + ], + "type": "text", + "content": "On Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "spans": [ + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "content": "Amir Mehrpanah1, Erik Englesson1, and Hossein Azizpour1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 190, + 478, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 190, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 190, + 478, + 213 + ], + "type": "text", + "content": "Department of Computer Science, Royal Institute of Technology, Stockholm, Sweden {amirme,engless,azizpour}@kth.se" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 236, + 452, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 452, + 400 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 452, + 400 + ], + "type": "text", + "content": "Abstract. Understanding the behavior of deep networks is crucial to increase our confidence in their results. Despite an extensive body of work for explaining their predictions, researchers have faced reliability issues, which can be attributed to insufficient formalism. In our research, we adopt novel probabilistic and spectral perspectives to formally analyze explanation methods. Our study reveals a pervasive spectral bias stemming from the use of gradient, and sheds light on some common design choices that have been discovered experimentally, in particular, the use of squared gradient and input perturbation. We further characterize how the choice of perturbation hyperparameters in explanation methods, such as SmoothGrad, can lead to inconsistent explanations and introduce two remedies based on our proposed formalism: (i) a mechanism to determine a standard perturbation scale, and (ii) an aggregation method which we call SpectralLens. Finally, we substantiate our theoretical results through quantitative evaluations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "spans": [ + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "text", + "content": "Keywords: Probabilistic Machine Learning " + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "text", + "content": " Gradient-based Explanation Methods " + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "text", + "content": " Probabilistic Pixel Attribution Techniques " + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "text", + "content": " Explainability " + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "text", + "content": " Deep Neural Networks " + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 411, + 452, + 445 + ], + "type": "text", + "content": " Spectral Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "content": "Ensuring the interpretability of machine learning models is crucial for user's trust and understanding [2, 6, 45]. While considerable efforts have been dedicated to advancing explainability in the field [37, 47, 49, 51, 57, 62], there is a notable gap in systematically formalizing and consolidating existing knowledge [5, 9, 16, 19]. Insufficient formal analysis leaves too much room for different interpretations of the explanations, making them unreliable [27] which is detrimental to the very use case of explanations, i.e., human-critical applications. Therefore, more attention is essential to retrospectively analyze and structure the wealth of work done, creating a cohesive foundation for explainability [39, 61]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "One main reason for the unreliability of explanation methods is their susceptibility to hyperparameters, leading to inconsistent explanations [4, 13, 33, 36, 48], see Fig. 1. Among the several works that formally study explanation methods [19, 37, 57], to our knowledge, none shed light on such inconsistencies. Therefore, in this work, we adopt a novel probabilistic and spectral perspective to characterize the observed inconsistencies and propose preliminary solutions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 219, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 219, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 219, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 219, + 198 + ], + "type": "image", + "image_path": "0e68eef4b31cc9eeeeb66fd83a96ef769983ba8c769329e421842f6b592caf38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 200, + 184, + 209 + ], + "lines": [ + { + "bbox": [ + 172, + 200, + 184, + 209 + ], + "spans": [ + { + "bbox": [ + 172, + 200, + 184, + 209 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 115, + 305, + 198 + ], + "blocks": [ + { + "bbox": [ + 223, + 115, + 305, + 198 + ], + "lines": [ + { + "bbox": [ + 223, + 115, + 305, + 198 + ], + "spans": [ + { + "bbox": [ + 223, + 115, + 305, + 198 + ], + "type": "image", + "image_path": "6790d550b909fe52e3755320774161cf06857cfd3ca86a9329e5809e880bef88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 200, + 270, + 209 + ], + "lines": [ + { + "bbox": [ + 258, + 200, + 270, + 209 + ], + "spans": [ + { + "bbox": [ + 258, + 200, + 270, + 209 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 115, + 391, + 198 + ], + "blocks": [ + { + "bbox": [ + 310, + 115, + 391, + 198 + ], + "lines": [ + { + "bbox": [ + 310, + 115, + 391, + 198 + ], + "spans": [ + { + "bbox": [ + 310, + 115, + 391, + 198 + ], + "type": "image", + "image_path": "03c52b4bd53b511552856ccfd5f167b88f3d2a9db724060736f777498edc1087.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 200, + 356, + 209 + ], + "lines": [ + { + "bbox": [ + 345, + 200, + 356, + 209 + ], + "spans": [ + { + "bbox": [ + 345, + 200, + 356, + 209 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "type": "text", + "content": "Fig. 1: The figure illustrates sample explanations generated by varying hyperparameters of SmoothGrad-Squared [30, 55] for ResNet-50 on ImageNet (see Fig. 16 in the Appendix for more examples). It is evident from the figure that using different hyperparameters results in diverse explanations. Such variations have been demonstrated in recent studies for other explanation techniques as well [4, 13, 33, 48]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 115, + 477, + 198 + ], + "blocks": [ + { + "bbox": [ + 395, + 115, + 477, + 198 + ], + "lines": [ + { + "bbox": [ + 395, + 115, + 477, + 198 + ], + "spans": [ + { + "bbox": [ + 395, + 115, + 477, + 198 + ], + "type": "image", + "image_path": "43820e97b3834c99683d7a63a41c4cd1303e12cb49f6ea9acd56f4234bf3ec06.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 200, + 443, + 209 + ], + "lines": [ + { + "bbox": [ + 430, + 200, + 443, + 209 + ], + "spans": [ + { + "bbox": [ + 430, + 200, + 443, + 209 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 299, + 420, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 420, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 420, + 312 + ], + "type": "text", + "content": "Our Contribution. We iterate our contributions in three parts:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 323, + 481, + 527 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 323, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 357 + ], + "type": "text", + "content": "- Analysis: We introduce a probabilistic representation for post-hoc gradient-based explanation methods in Sec. 2.1, and perform a spectral analysis of them in Sec. 3.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 360, + 481, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 481, + 443 + ], + "type": "text", + "content": "- Identified Issues: Our examination reveals a shared spectral bias stemming from the fundamental assumption of utilizing gradients to measure feature contributions. We show that the combined spectral effects of gradient and perturbation lead to band-pass filters modulating the explanations, which could lead to inconsistencies in the explanations, discussed in Sec. 3.2. Moreover, we conduct a spectral analysis of a few specific perturbation distributions that have been proposed in the literature in Sec. 3.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 444, + 481, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 481, + 527 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 481, + 527 + ], + "type": "text", + "content": "- Solutions and Evaluations: To address inconsistencies, we investigate two straightforward solutions that are finding an optimal perturbation scale based on the cosine similarity of the perturbation and the classifier in the frequency domain in Sec. 4.1, and an aggregation mechanism named SpectralLens in Sec. 4.2. Finally, we evaluate the proposed methods using the pixel removal strategy [48], acknowledging their effectiveness and limitations under Sec. 5." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 560, + 226, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 226, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 226, + 574 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": "Notation. We consider a classification setting, that is, modeling " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "p(y|\\pmb{x})" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": " for labels " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "y \\in \\{1,\\dots,k\\}" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": " the set of input images. We denote the " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\log p(y|\\pmb{x})" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": ". We denote standard Fourier bases by " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{\\psi(\\pmb{\\omega},\\pmb{x})\\}_{\\pmb{\\omega} \\in \\mathbb{R}^n}" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": ", according to which, under some regularity conditions [43], " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": " can be decomposed as " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}) = \\int \\widehat{f}(\\pmb{\\omega}) \\psi(\\pmb{\\omega},\\pmb{x}) d\\pmb{\\omega}" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": ". In this representation, " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\widehat{f}" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": " denotes the Fourier transform (FT) of " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": ". We denote the power spectral density (PSD) of " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "inline_equation", + "content": "S_f(\\pmb{\\omega}) = |\\widehat{f}(\\pmb{\\omega})|^2" + }, + { + "bbox": [ + 130, + 590, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 399, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 399, + 128 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 399, + 128 + ], + "type": "text", + "content": "2.1 A Probabilistic Representation of Explanations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": "In this view, each explanation method consists of three components: (1) a distribution for perturbation of the input, (2) an explainer, and (3) a summary statistic. We use " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\pmb{x}} | \\pmb{x}, \\sigma)" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": " to denote the distribution of perturbations of the input sample parameterized by " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": ", referred to as the perturbation scale. A definition of \"explanation\" is recovering the contribution of features to a prediction, which emanates from human's natural counterfactual reasoning [19]. Therefore, the explainer is usually the gradient with respect to input " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}}" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": ", or approximations of it, e.g. the finite difference, given by " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\Delta_{\\pmb{x}} f(\\pmb{x}) = \\frac{f(\\tilde{\\pmb{x}}) - f(\\pmb{x})}{\\tilde{\\pmb{x}} - \\pmb{x}}" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": " (see Appendix D). Finally, we assume that the summary statistic is the expectation of the outputs, or the squared outputs, of the explainer, e.g. " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "(\\nabla f(\\tilde{\\pmb{x}}))^2" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": ", over the chosen perturbation distribution, " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\pmb{x}} | \\pmb{x}, \\sigma)" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{p(\\tilde{\\pmb{x}})}\\left[(\\nabla f(\\tilde{\\pmb{x}}))^2\\right]" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": ", where the operations are carried out pixelwise. This is referred to as the explanation of " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": " with a hyperparameter " + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 134, + 482, + 327 + ], + "type": "text", + "content": ". This probabilistic view can, in fact, unify different explanation methods into a single form (see Appendix C). We provide theoretical justifications for why using squared gradients is preferable to gradients, explaining the empirical findings [30], and will discuss both cases in Sec. 3.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 344, + 354, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 344, + 354, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 354, + 358 + ], + "type": "text", + "content": "3 Spectral Analysis of Explanations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 522 + ], + "type": "text", + "content": "This section demonstrates the spectral property in explanations when using the gradient explainer and its interaction with perturbations, where a significant difference in the frequency of learned features leads to a selective attribution of features. In the following sections, we draw on insights from spectral analysis of signals and systems theory, utilizing this viewpoint to gain a deeper understanding of explanation methods. We demonstrate a spectral bias related to gradient-based explanation approaches, identify its source, and interpret various methods in the literature as efforts to alleviate it through the incorporation of low-pass filters, either explicitly [55, 57] or implicitly [42, 62]. We illustrate that the inconsistencies in explanation techniques stem from different strategies for mitigating the spectral bias. Finally, we suggest two straightforward approaches to tackle these inconsistencies: determining an optimal perturbation scale in Sec. 4.1 and aggregating the explanations in Sec. 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "text", + "content": "Writing a spectral representation for prediction, as shown in Appendix A.1, the model output with a perturbation distribution can be written as a convolution between the function and the perturbation distribution " + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "inline_equation", + "content": "f * p" + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "text", + "content": ", therefore we may use \"perturbation kernel\" and \"perturbation distribution\" interchangeably." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 587, + 353, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 353, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 353, + 600 + ], + "type": "text", + "content": "3.1 Spectral Representation of Explainers" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Here, by drawing a connection between explainability techniques and the literature on spectral signal processing [12, 38, 56], we take a novel avenue for our analysis. This connection enables us to have a spectral representation of the explanations, as discussed in the next sections (see Appendix A for the derivations). It is important to note that the Fourier bases in our work is 1D and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "per pixel, and should not be confused with the traditional Fourier bases used in image processing literature that expands along spatial dimensions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 164, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 164, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 479, + 189 + ], + "type": "text", + "content": "Gradients in Spectral Domain. For explanation methods that use gradients and a perturbation kernel, " + }, + { + "bbox": [ + 132, + 164, + 479, + 189 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\pmb{x}})" + }, + { + "bbox": [ + 132, + 164, + 479, + 189 + ], + "type": "text", + "content": " we have the following (see Appendix A.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 201, + 480, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 201, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 176, + 201, + 480, + 253 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} [ \\nabla f (\\tilde {\\boldsymbol {x}}) ] = \\int i 2 \\pi \\omega \\widehat {f} (\\omega) \\widehat {p} (\\omega) d \\omega (1) \\\\ = - 2 \\pi \\left[ \\int \\omega \\left(\\widehat {f} _ {E} (\\omega) \\widehat {p} _ {O} (\\omega) + \\widehat {f} _ {O} (\\omega) \\widehat {p} _ {E} (\\omega)\\right) d \\omega \\right] (2) \\\\ \\end{array}", + "image_path": "54b162b5314908b5c3ffd2c29d87c43343ebf18924cd802cceb8f58735243bb3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\widehat{p}" + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "text", + "content": " denotes the FT of the perturbation kernel. We used " + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\widehat{f}_E" + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\widehat{f}_O" + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "text", + "content": " to denote a function " + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\widehat{f}" + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "text", + "content": " is decomposed into real even and imaginary odd components " + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\widehat{f} = i\\widehat{f}_O + \\widehat{f}_E^1" + }, + { + "bbox": [ + 132, + 265, + 479, + 326 + ], + "type": "text", + "content": ". The negative sign suggests that in case of using the gradients, we need to negate the actual values to obtain rankings. We will elaborate more on the negation of the attributions in Sec. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 342, + 479, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 479, + 425 + ], + "type": "text", + "content": "Remark 1. Note the symmetry in the form of the spectral representation of the gradients. This form implies that it is inevitable for any perturbation kernel to extract asymmetric information about even and odd parts of the classifier " + }, + { + "bbox": [ + 132, + 342, + 479, + 425 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 342, + 479, + 425 + ], + "type": "text", + "content": ". Therefore, the gradient explainer with a kernel symmetric about the sample being explained, only extracts information about the odd parts of the classifier. This is problematic because the attribution is expected to reflect the total contribution to the prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 439, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 479, + 464 + ], + "type": "text", + "content": "We will use this interesting symmetry later in Sec. 3.3 for analysis of a specific perturbation kernel." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 488, + 479, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 488, + 479, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 479, + 512 + ], + "type": "text", + "content": "Squared Gradients in Spectral Domain. On the other hand, for explanation methods which use squared gradients, we have the following (see Appendix A.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 525, + 479, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 525, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 201, + 525, + 479, + 548 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] = 4 \\pi^ {2} \\int \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) \\mathrm {S} _ {\\sqrt {p}} (\\boldsymbol {\\omega}) d \\boldsymbol {\\omega} \\tag {3}", + "image_path": "6f79bffe17a06a0ed4c52d7db227ebb13e9a8058dd0943b34150c4adc5f503be.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 557, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 479, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 557, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_f" + }, + { + "bbox": [ + 132, + 557, + 479, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 557, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_{\\sqrt{p}}" + }, + { + "bbox": [ + 132, + 557, + 479, + 581 + ], + "type": "text", + "content": " denote the PSD of the model and square root of the perturbation kernel respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 582, + 479, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 479, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 479, + 630 + ], + "type": "text", + "content": "We discuss the benefits of using squared gradients as in Eq. (3) over gradients as in Eq. (2) later in Sec. 3.3. However, it can be noticed here that the squared gradients, unlike gradients, are always non-negative, extract information about the PSD of " + }, + { + "bbox": [ + 132, + 582, + 479, + 630 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 582, + 479, + 630 + ], + "type": "text", + "content": ", and do not have a complex symmetric structure." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "type": "text", + "content": "1 Also note that as we are using real valued functions, other components i.e. real odd and imaginary even, are zero." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 476, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 476, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 476, + 128 + ], + "type": "text", + "content": "3.2 Gradient Explainers with Perturbations Are Band-Pass Filters" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 479, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 479, + 237 + ], + "type": "text", + "content": "In our probabilistic framework, various explanation methods can be decomposed into three components (1) an explainer, (2) a perturbation kernel and (3) a summary statistic. In this section, we analyze the spectral effects of the two core components, namely gradient and perturbation. Importantly, we show that gradient acts as a high-pass filter, and perturbation acts as a low-pass filter. Before that, we need a more formal definition for a high-pass and low-pass filter that suits our study. Assuming that " + }, + { + "bbox": [ + 130, + 141, + 479, + 237 + ], + "type": "inline_equation", + "content": "\\widehat{m} (\\omega)" + }, + { + "bbox": [ + 130, + 141, + 479, + 237 + ], + "type": "text", + "content": " denotes the FT of a perturbation kernel, we define low- and high-pass as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "content": "Definition 1. A filter " + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\widehat{m}(\\omega)" + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "content": " is low pass, if for a given " + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "inline_equation", + "content": "M > 0" + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "content": " there exists a threshold " + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\omega^{*}" + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "content": " such that for any " + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\|\\omega\\| > \\|\\omega^{*}\\|" + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\|\\widehat{m}(\\omega)\\| < M" + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "content": ", and it is high pass if such a condition holds for " + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "inline_equation", + "content": "\\|\\omega\\| < \\|\\omega^{*}\\|" + }, + { + "bbox": [ + 131, + 248, + 480, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 295, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 295, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 479, + 319 + ], + "type": "text", + "content": "For better readability, the proofs for the theorems and propositions are outlined in Appendix B, and we provide intuition as we progress." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 333, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 333, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 333, + 479, + 357 + ], + "type": "text", + "content": "Theorem 1. Any perturbation kernel " + }, + { + "bbox": [ + 131, + 333, + 479, + 357 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\pmb{x}})" + }, + { + "bbox": [ + 131, + 333, + 479, + 357 + ], + "type": "text", + "content": " mitigates the attribution of high-frequency features; simply put, any perturbation is a low-pass filter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "text", + "content": "The intuition is that any modification to a signal, would only remove information, and, as the intensity of perturbation increases, the removal goes from high-frequency components toward lower frequencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 421, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 421, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 421, + 479, + 445 + ], + "type": "text", + "content": "Theorem 2. The gradient operator amplifies the attribution of high-frequency features; simply put, the gradient operator is a high-pass filter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 459, + 479, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 479, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 479, + 494 + ], + "type": "text", + "content": "From the stated theorems, one can see that a gradient-based explanation with perturbation creates a band-pass filter, whose shape is controlled by the perturbation hyperparameters (see Figs. 2 and 3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 509, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 479, + 568 + ], + "type": "text", + "content": "Remark 2. This amplification-mitigation trade-off can be seen in several explanation methods, where perturbation kernels (low-pass filters) are utilized to mitigate the amplification effect (the high-pass filter) caused by the gradient. Therefore, the combined effect of the two filters usually forms a band-pass filter (see Appendix C for a non-comprehensive list)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 592, + 413, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 592, + 413, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 592, + 413, + 604 + ], + "type": "text", + "content": "3.3 Spectral Analysis of Specific Perturbation Kernels" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": "In this section, we analyze the properties of special kernels used in the literature with remarks or propositions that justify, and thus shed light on, previous work's experimental observations. In most cases, we expand upon the squared gradients, Eq. (3), due to its analytical simplicity." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "VanillaGrad via a Dirac Kernel. The very naive explanation method can be obtained with a Dirac kernel " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\pmb{x}}) = \\delta_{\\pmb{x}}(\\tilde{\\pmb{x}})" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " known as VanillaGrad (VG). Analyzing the squared version of VG and following Eq. (3), we know that the FT of the Dirac kernel is " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "S_{\\sqrt{p}}(\\omega) = 1(\\omega)" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(\\omega)" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is the constant function mapping all values to 1. Thus, it can be seen that we recover the scaled PSD of " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\| \\boldsymbol{\\omega} \\|^2 \\mathrm{S}_f(\\boldsymbol{\\omega})" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "text", + "content": "Remark 3. Scaling " + }, + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\widehat{f}" + }, + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "text", + "content": " by frequency can be interpreted as the bias of the gradient explainer towards exaggeration of the attribution of high-frequency features [31]. Beside other justifications for gradient sparsity [8], spectral representation of " + }, + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{VG}^2" + }, + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "text", + "content": " and domination of high-frequency features give an orthogonal justification for why VanillaGrad explanations being usually scattered and noisy (see Fig. 12)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 336 + ], + "type": "text", + "content": "SmoothGrad via a Gaussian Kernel. Letting " + }, + { + "bbox": [ + 130, + 277, + 482, + 336 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\pmb{x}}) = \\mathcal{N}(\\pmb{x},\\sigma^2)" + }, + { + "bbox": [ + 130, + 277, + 482, + 336 + ], + "type": "text", + "content": ", i.e. a Gaussian with variance " + }, + { + "bbox": [ + 130, + 277, + 482, + 336 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 130, + 277, + 482, + 336 + ], + "type": "text", + "content": ", centered on the sample being explained, we recover variants of an explanation method known as SmoothGrad (SG) or the squared gradient version SmoothGrad-Squared (SG²) [30, 55], we can rewrite Eq. (3) as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 337, + 481, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 337, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 200, + 337, + 481, + 360 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] \\propto \\int \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) e ^ {- 8 \\pi^ {2} \\sigma^ {2} \\| \\boldsymbol {\\omega} \\| ^ {2}} d \\boldsymbol {\\omega} \\tag {4}", + "image_path": "958d6e5e5c0e1bd86893ef9fd3b90358311f4890f12fb8a7697c39135ee74dad.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": "This shows that the contributions " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_f(\\omega)" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " are modulated by the combined spectral effect of the Gaussian perturbation kernel and gradient operator, which is a band pass filter of the form " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\| \\boldsymbol{\\omega} \\|^2 e^{-8\\pi^2 \\sigma^2} \\| \\boldsymbol{\\omega} \\|^2" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": ", whose shape is controlled by the kernel variance (see Fig. 2). The explanations could be inconsistent, but generally become smoother as " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " increases, which is a result of the shrinkage of the attribution of high-frequency features that can dominate the explanations with a small " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " (see Fig. 1)." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 234, + 476, + 382, + 544 + ], + "blocks": [ + { + "bbox": [ + 234, + 476, + 382, + 544 + ], + "lines": [ + { + "bbox": [ + 234, + 476, + 382, + 544 + ], + "spans": [ + { + "bbox": [ + 234, + 476, + 382, + 544 + ], + "type": "image", + "image_path": "b31ae8acbda21c62db3680ca167d5e8643f9577041981134d95a585adb9f706a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 559, + 482, + 594 + ], + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 594 + ], + "type": "text", + "content": "Fig.2: A visualization of frequency bands, where " + }, + { + "bbox": [ + 130, + 559, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_{\\sqrt{p}}(\\omega)" + }, + { + "bbox": [ + 130, + 559, + 482, + 594 + ], + "type": "text", + "content": " is defined according to Eq. (4), show that change in the noise scale " + }, + { + "bbox": [ + 130, + 559, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 559, + 482, + 594 + ], + "type": "text", + "content": " leads to band-pass filters with different modes, hence, representing information of different frequencies." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "When generating explanations, one is tempted to look for hyperparameters that make the explanation reflect what they think it should [16]. However, monitoring explanations from only one frequency band is potentially misleading and is likely to contradict itself simply by changing hyperparameters (see Fig. 5) [27,33]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 236, + 208 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 236, + 208 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 236, + 208 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 236, + 208 + ], + "type": "image", + "image_path": "67c2e0419ad1f1d36380c85575144747ffdd4dd7dc6e82bf1f4f862896522cfb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 214, + 195, + 224 + ], + "lines": [ + { + "bbox": [ + 182, + 214, + 195, + 224 + ], + "spans": [ + { + "bbox": [ + 182, + 214, + 195, + 224 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 241, + 118, + 338, + 209 + ], + "blocks": [ + { + "bbox": [ + 241, + 118, + 338, + 209 + ], + "lines": [ + { + "bbox": [ + 241, + 118, + 338, + 209 + ], + "spans": [ + { + "bbox": [ + 241, + 118, + 338, + 209 + ], + "type": "image", + "image_path": "53bc8bee8d3cd47038cd74474421e58ef512951a596f609f1b8f0f684b07c2b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 214, + 294, + 224 + ], + "lines": [ + { + "bbox": [ + 282, + 214, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 282, + 214, + 294, + 224 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 346, + 118, + 473, + 211 + ], + "blocks": [ + { + "bbox": [ + 346, + 118, + 473, + 211 + ], + "lines": [ + { + "bbox": [ + 346, + 118, + 473, + 211 + ], + "spans": [ + { + "bbox": [ + 346, + 118, + 473, + 211 + ], + "type": "image", + "image_path": "ad1af17a5be4a7a8c28d04ec78998923adf17d112e888a31077027331948287b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 214, + 362, + 224 + ], + "lines": [ + { + "bbox": [ + 351, + 214, + 362, + 224 + ], + "spans": [ + { + "bbox": [ + 351, + 214, + 362, + 224 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "type": "text", + "content": "Fig. 3: A 1D visualization of amplification (b) and mitigation (d) effects caused by gradient and perturbation, respectively. For simplicity, we have shown a rectangular pulse (c) convolved with " + }, + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\nabla f" + }, + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "type": "text", + "content": ", but the kernel shape is usually more complex in practice, an example shown in Fig. 2. Moreover, functions in blue are the Fourier representation of the black functions. Due to the gradient operator, small high-frequency jitters in the decision boundary (a) amplify proportionally to their frequency (b). To remove the effect of high-frequency components of the decision boundary that usually dominate the explanations, researchers have proposed different perturbation kernels, visualized in (d). Theorem 2 corresponds to " + }, + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "type": "inline_equation", + "content": "(\\mathrm{a}\\rightarrow \\mathrm{b})" + }, + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "type": "text", + "content": " and Theorem 1 corresponds to " + }, + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "type": "inline_equation", + "content": "(\\mathrm{b}\\rightarrow \\mathrm{d})" + }, + { + "bbox": [ + 130, + 232, + 482, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 356, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 480, + 380 + ], + "type": "text", + "content": "We formalize this statement, also known as the Rashomon effect (see Appendix B for the proof)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 390, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 482, + 427 + ], + "type": "text", + "content": "Proposition 1. The SmoothGrad-Squared pixel attribution can contradict itself by changing the hyperparameters. Simply, the Rashomon effect can occur in the SmoothGrad-Squared explanations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 439, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 439, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 439, + 482, + 487 + ], + "type": "text", + "content": "Intuition. This proposition suggests a reason, i.e. removing different frequency bands, for inconsistencies we encounter in " + }, + { + "bbox": [ + 131, + 439, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathrm{SG}^2" + }, + { + "bbox": [ + 131, + 439, + 482, + 487 + ], + "type": "text", + "content": " explanations, which has been unknown so far [48]. Note that this occurs with different kernels, but we have chosen the one that leads to a simpler proof." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 506, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 482, + 567 + ], + "type": "text", + "content": "IntegratedGrad via a Rect Kernel. Setting the perturbation kernel to " + }, + { + "bbox": [ + 131, + 506, + 482, + 567 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\boldsymbol{x}}) = \\frac{1}{\\sigma|\\boldsymbol{s} - \\boldsymbol{x}|}\\mathrm{Rect}(\\frac{\\tilde{\\boldsymbol{x}} - \\boldsymbol{x}}{\\sigma(\\boldsymbol{s} - \\boldsymbol{x})} -\\frac{1}{2})^2" + }, + { + "bbox": [ + 131, + 506, + 482, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 506, + 482, + 567 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 506, + 482, + 567 + ], + "type": "text", + "content": " is a chosen baseline image, we recover the perturbation kernel of a well known method called IntegratedGrad (IG) [57]. We use the spectral representation of gradients in Eq. (2), for a proposition about the gradient signs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 196, + 575, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 575, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 196, + 575, + 481, + 600 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} [ \\nabla f (\\tilde {\\boldsymbol {x}}) ] \\propto \\int i \\widehat {f} (\\boldsymbol {\\omega}) \\sin (\\pi \\sigma (\\boldsymbol {s} - \\boldsymbol {x}) \\boldsymbol {\\omega}) e ^ {i \\pi \\boldsymbol {\\omega} \\boldsymbol {r}} d \\boldsymbol {\\omega} \\tag {5}", + "image_path": "515c0dfaa69eb13b95413990dc243c979bf2f32192b6df701b026dbcec468d1a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 609, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 482, + 647 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 609, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r} = (1 + 2\\frac{\\boldsymbol{x}}{\\sigma(\\boldsymbol{s} - \\boldsymbol{x})})" + }, + { + "bbox": [ + 131, + 609, + 482, + 647 + ], + "type": "text", + "content": ", appears because the kernel is not centered on the sample being explained. Note that the Fourier operator linearly expands along each dimension, but we use vector forms for notational simplicity. Furthermore, to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "type": "text", + "content": " Rect function is defined as " + }, + { + "bbox": [ + 133, + 652, + 375, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{Rect}(\\tilde{\\pmb{x}}) = \\frac{1}{2}\\mathrm{sign}(\\frac{1}{2} -|\\tilde{\\pmb{x}} |) + \\frac{1}{2}" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "study the effect of elementwise multiplication on the explanation quantitatively, we use IG to denote integrated gradients without the elementwise multiplication with input and XIG to denote the version with elementwise multiplication." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": "In explanation visualizations, the negative gradient sign has been used to show a negative contribution to the prediction [7, 53], to our knowledge it has never been formally justified. In the next proposition, we state that the negative gradient sign can be the direct result of the chosen perturbation kernel, formally shown in Appendix B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 472, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 472, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 472, + 232 + ], + "type": "text", + "content": "Proposition 2. The gradient sign in " + }, + { + "bbox": [ + 131, + 220, + 472, + 232 + ], + "type": "inline_equation", + "content": "IG" + }, + { + "bbox": [ + 131, + 220, + 472, + 232 + ], + "type": "text", + "content": " is affected by the perturbation level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 479, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 276 + ], + "type": "text", + "content": "This proposition entails that the negative sign in the gradient explainer does not necessarily reflect a negative contribution to the prediction, nor the positive sign necessarily reflects a positive contribution (see Figs. 13 and 14)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 282, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 282, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 480, + 331 + ], + "type": "text", + "content": "Remark 4. Since we assume continuity wrt " + }, + { + "bbox": [ + 131, + 282, + 480, + 331 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 282, + 480, + 331 + ], + "type": "text", + "content": ", in our analysis, by Bolzano's theorem [50], there are hyperparameters that set the attribution to zero. Therefore, the gradient explainer may not produce a useful ranking compared to that of the gradient squared." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 338, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 481, + 399 + ], + "type": "text", + "content": "Remark 5. Proposition 2 also shows that adding explanations with different hyperparameters may cause destructive interference when the different values of " + }, + { + "bbox": [ + 131, + 338, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\omega \\widehat{p} (\\omega)" + }, + { + "bbox": [ + 131, + 338, + 481, + 399 + ], + "type": "text", + "content": " are out of phase. Hence, squared gradients is a more appropriate design choice, unless the use of phase information is planned as part of the method. This has been discovered experimentally [30] and confirmed in Sec. 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 415, + 457, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 457, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 457, + 430 + ], + "type": "text", + "content": "4 Proposed Solutions Based on the Spectral Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 440, + 358, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 440, + 358, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 440, + 358, + 452 + ], + "type": "text", + "content": "4.1 Finding an Optimal Perturbation Scale" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "text", + "content": "Note that the squared gradients in spectral domain (Eq. (3)) can be formulated as an inner product between the scaled PSD of the perturbation distribution and the PSD of the classifier i.e. " + }, + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{p(\\tilde{\\boldsymbol{x}})}\\left[(\\nabla f(\\tilde{\\boldsymbol{x}}))^2\\right] = 4\\pi^2\\langle \\mathrm{S}_f,\\| \\boldsymbol {\\omega}\\| ^2\\mathrm{S}_{\\sqrt{p}}\\rangle" + }, + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "text", + "content": " Assuming unimodality, a nonzero contribution, and that the norm " + }, + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\left\\| \\left\\| \\boldsymbol {\\omega}\\right\\| ^2\\mathrm{S}_{\\sqrt{p}}\\right\\|" + }, + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "text", + "content": " exists, then the cosine similarity is maximized at a certain perturbation scale that has the highest similarity with " + }, + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_f" + }, + { + "bbox": [ + 130, + 459, + 480, + 544 + ], + "type": "text", + "content": " . A higher cosine similarity guarantees recovering more information about the classifier." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 550, + 481, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 481, + 574 + ], + "type": "text", + "content": "Definition 2. The cosine similarity between the PSD of the classifier and that of the perturbation kernel given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 202, + 582, + 480, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 582, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 202, + 582, + 480, + 611 + ], + "type": "interline_equation", + "content": "\\text {S i m i l a r i t y} \\left(\\mathrm {S} _ {f}, \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}}\\right) = 4 \\pi^ {2} \\frac {\\left\\langle \\mathrm {S} _ {f} , \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\right\\rangle}{\\| \\mathrm {S} _ {f} \\| \\left\\| \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\right\\|} \\tag {6}", + "image_path": "5a905148ac631a06090d728be42e7f5fda51b0e31b63635a3a3d630b0bcff0cc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "This quantity can be used as a way to determine the optimal perturbation scale per pixel to maximize the information extracted with the squared gradients. Note that, the word \"optimal\" refers to extracting the most information by the inner product, whether deemed relevant by a human or not.." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 165 + ], + "type": "text", + "content": "For one of the perturbation kernels we have presented earlier in Sec. 3.3, we have optimized this quantity on multiple levels, such as dataset level Fig. 4, image level Tab. 1, and, pixel level Fig. 5d. Finally, we have experimentally verified the effectiveness of this approach in Sec. 5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 178, + 481, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 178, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 178, + 481, + 205 + ], + "type": "text", + "content": "Computing the Similarity of Classifier and Gaussian Kernel. To compute the similarity, we first need to evaluate the kernel norm as follows" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 208, + 481, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 208, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 198, + 208, + 481, + 237 + ], + "type": "interline_equation", + "content": "\\left\\| \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2} e ^ {- 8 \\pi^ {2} \\sigma^ {2} \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2}} \\right\\| = \\left(\\int \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {4} e ^ {- 1 6 \\pi^ {2} \\sigma^ {2} \\left\\| \\boldsymbol {\\omega} \\right\\| ^ {2}} d \\boldsymbol {\\omega}\\right) ^ {\\frac {1}{2}} \\tag {7}", + "image_path": "110374cb9818fc59a968a1a7e0ddfbf3bd09ce3346551b4809b179cdd8e12a9f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 284, + 239, + 481, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 239, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 284, + 239, + 481, + 255 + ], + "type": "interline_equation", + "content": "\\propto \\sigma^ {\\frac {- 5}{2}} \\tag {8}", + "image_path": "0cf1957c12251c26e8c33604ef0142fd08ba39ed86c048d7e4bfd3609cd1094d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 260, + 481, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 260, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 260, + 481, + 284 + ], + "type": "text", + "content": "Therefore, the cosine similarity for the Gaussian perturbation kernel can be computed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 220, + 289, + 395, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 289, + 395, + 304 + ], + "spans": [ + { + "bbox": [ + 220, + 289, + 395, + 304 + ], + "type": "interline_equation", + "content": "\\mathrm {S i m i l a r i t y} (\\mathrm {S} _ {\\sqrt {p}}, \\mathrm {S} _ {f}) \\propto \\sigma^ {\\frac {5}{2}} \\langle \\mathrm {S} _ {f}, \\| \\boldsymbol {\\omega} \\| ^ {2} \\mathrm {S} _ {\\sqrt {p}} \\rangle", + "image_path": "2ee231c49428102d632eef200a06ab6aae3e92b355ea6d99e9b85d413e8113e1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 308, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 481, + 357 + ], + "type": "text", + "content": "Using the L1 norm for the squared gradients of flattened explanations. We have observed experimentally that " + }, + { + "bbox": [ + 131, + 308, + 481, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_f" + }, + { + "bbox": [ + 131, + 308, + 481, + 357 + ], + "type": "text", + "content": " depends on the input scale, hence we use a convex interpolation with noise to support the assumption that " + }, + { + "bbox": [ + 131, + 308, + 481, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_f" + }, + { + "bbox": [ + 131, + 308, + 481, + 357 + ], + "type": "text", + "content": " remains constant (see Appendix E.2 and Fig. 8). The results are shown in Fig. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 361, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 482, + 458 + ], + "type": "text", + "content": "Remark 6. The optimal perturbation scale, determined by maximizing cosine similarity, offers a standard method for selecting the perturbation scale in " + }, + { + "bbox": [ + 131, + 361, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathrm{SG^2}" + }, + { + "bbox": [ + 131, + 361, + 482, + 458 + ], + "type": "text", + "content": ". Given that user discretion in the selection of perturbation scale contributes to inconsistencies, as illustrated in Fig. 1, opting for the optimal perturbation scale mitigates these inconsistencies in the results. Hence, it represents a more suitable design choice. In Sec. 5 we demonstrate that this approach not only enhances " + }, + { + "bbox": [ + 131, + 361, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathrm{SG^2}" + }, + { + "bbox": [ + 131, + 361, + 482, + 458 + ], + "type": "text", + "content": "'s robustness against inconsistencies but also its adaptability to diverse datasets, which in turn can improve the quantitative performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 477, + 422, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 422, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 422, + 490 + ], + "type": "text", + "content": "4.2 Aggregation of Information With Explanation Prior" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 494, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 481, + 567 + ], + "type": "text", + "content": "We have shown that explanations are influenced by specific frequency bands, rather than representing the importance of individual pixels for prediction. Instead, they indicate the contributions within a specific frequency range. This observation motivates us to investigate a method for aggregation of information across different bands, which closely resembles an ensemble of IG and SG [55,57] which we call SpectralLens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "text", + "content": "Definition 3. SpectralLens is an ensemble explanation method, that aggregates information across different frequency bands as follows" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 169, + 600, + 481, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 600, + 481, + 625 + ], + "spans": [ + { + "bbox": [ + 169, + 600, + 481, + 625 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {p (\\sigma)} \\left[ \\mathbb {E} _ {p (\\tilde {\\boldsymbol {x}})} \\left[ (\\nabla f (\\tilde {\\boldsymbol {x}})) ^ {2} \\right] \\right] \\propto \\iint \\mathrm {S} _ {f} (\\boldsymbol {\\omega}) \\| \\boldsymbol {\\omega} \\| ^ {2} e ^ {- 8 \\pi^ {2} \\sigma \\| \\boldsymbol {\\omega} \\| ^ {2}} d \\boldsymbol {\\omega} p (\\sigma) d \\sigma \\tag {9}", + "image_path": "14786f4e18e4070acf25c5f724f6afb33da55b36d4972bd5bfea584943919dad.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": "Where we refer to this method by " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "SL^2" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": " as it is using the squared gradients and we call " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "p(\\sigma)" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": " as an explanation prior, gauging the frequency bands that we choose to represent in an explanation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 103 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 304, + 193 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 304, + 193 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 304, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 304, + 193 + ], + "type": "image", + "image_path": "58a578b8a285dba16384c1190133078cef1105bae86180afc2741924d8b891df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 198, + 227, + 209 + ], + "lines": [ + { + "bbox": [ + 214, + 198, + 227, + 209 + ], + "spans": [ + { + "bbox": [ + 214, + 198, + 227, + 209 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 311, + 118, + 476, + 193 + ], + "blocks": [ + { + "bbox": [ + 311, + 118, + 476, + 193 + ], + "lines": [ + { + "bbox": [ + 311, + 118, + 476, + 193 + ], + "spans": [ + { + "bbox": [ + 311, + 118, + 476, + 193 + ], + "type": "image", + "image_path": "ad50c71b764988b8360a2f65904a4624eee08e61389feaeff4d15c2c8584a998.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 199, + 399, + 209 + ], + "lines": [ + { + "bbox": [ + 387, + 199, + 399, + 209 + ], + "spans": [ + { + "bbox": [ + 387, + 199, + 399, + 209 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 217, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 340 + ], + "type": "text", + "content": "Fig. 4: Cosine similarity between kernel and classifier, defined in Eq. (6), is measured for the Gaussian perturbation kernel and ResNet50 averaged on the 10K images of the ImageNet validation set (a) and Food101 (b). This metric assesses the similarity between the PSD of the perturbation kernel and that of the classifier, indicating the perturbation level's effectiveness for kernels with finite norm in the Fourier domain. When using a Dirac explanation prior in Eq. (9), where a perturbation level is assigned per image or dataset, the optimal choice for the perturbation level is where the similarity is maximized across the image or dataset respectively. Note that this level of noise refers to a band-pass filter, shown in Fig. 2, which suggests the consideration of a latent distribution on the frequency of labeling, reflecting the workforce's mindset during data labeling." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "content": "Remark 7. " + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{SL}^2" + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "content": " exhibits robustness against inconsistencies in " + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{SG}^2" + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "content": " due to its aggregation of information from various bands. Therefore, it is a preferable design choice not to employ a Dirac explanation prior with " + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "inline_equation", + "content": "SG^2" + }, + { + "bbox": [ + 130, + 361, + 482, + 422 + ], + "type": "text", + "content": ", as it eliminates inconsistencies, depicted in Fig. 1. Nevertheless, as demonstrated in the Sec. 5, this approach may enhance the quantitative performance of explanations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "content": "Intuition. Considering Eq. (4), we realize that the combination of perturbation and gradient explainer creates overlapping band-pass carrier waves that modulate the amplitude of the spectral density of the classifier. Therefore, a new summary statistic is required that can reflect a wide range of frequency bands. We explicitly identified the spectral properties of the explanations, recognizing that not all frequencies are equally desired. Explanations with very high frequencies primarily attribute predictions to pixels. Therefore, we utilize an explanation prior " + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "inline_equation", + "content": "p(\\sigma)" + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "content": " for a nuanced control of the spectral properties of explanations. Although there are many possible alternatives for selecting an explanation prior, we used " + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "inline_equation", + "content": "p(\\sigma) = \\mathrm{Unif}(0,1)" + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "content": " for simplicity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "content": "Definition 4. We define " + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{ArgLens}(AL)" + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "content": " as the point at which the " + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "inline_equation", + "content": "SG^2" + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "content": " is maximized for pixel " + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "content": ". Formally defined as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 230, + 594, + 481, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 594, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 230, + 594, + 481, + 613 + ], + "type": "interline_equation", + "content": "\\operatorname {A r g L e n s} \\left(x _ {i}\\right) = \\arg \\max _ {\\sigma} \\left\\{\\mathrm {S G} ^ {2} \\left(x _ {i}\\right) \\right\\} \\tag {10}", + "image_path": "324048816424fd2db3c97e8a564aff5114a1d84e7ec47ff95ac8ad04e62d54da.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 620, + 384, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 384, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 384, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 620, + 384, + 633 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 132, + 620, + 384, + 633 + ], + "type": "text", + "content": " is the hyperparameter of the perturbation kernel." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "AL shows the frequency band at which the pixel's attribution is maximized. Alternatively, it can be interpreted as the robustness of a pixel to noise. We" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "have visualized the AL for the Gaussian kernel in Fig. 5 and discussed the way it helps the interpretation of explanations in Sec. 4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "content": "The aggregation of explanations from different frequencies is visualized in Fig. 5. Moreover, the basic explanation methods are obtainable with a Dirac explanation prior " + }, + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "inline_equation", + "content": "p(\\sigma) = \\delta(\\sigma)" + }, + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "content": ". Therefore, it can be used as a unified formulation for gradient-based explanation methods in spectral domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 264 + ], + "type": "text", + "content": "Interesting Properties of SpectralLens and ArgLens. In the absence of a robust mathematical framework, the interpretation of explanations may be subjective. Thus, explanation methods must be grounded in solid theoretical principles. " + }, + { + "bbox": [ + 130, + 205, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{SL}^2" + }, + { + "bbox": [ + 130, + 205, + 482, + 264 + ], + "type": "text", + "content": " addresses potential inconsistencies by aggregating attributions across frequency bands, ensuring fair representation of all pixels' contributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 361 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 130, + 265, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{SG^2}" + }, + { + "bbox": [ + 130, + 265, + 482, + 361 + ], + "type": "text", + "content": " results in bands-pass filters (see Sec. 3.3), its attributions reflect the changes in the decision boundary within a certain frequency band. Furthermore, AL reveals the frequency band at which the attribution is maximized. Hence, lower AL indicates that the attributions are likely governed by high-frequency components of the decision boundary, more sensitive to input noise, consequently less robust. An ideal classifier with smooth decision boundaries, i.e. robust features, leads to high AL everywhere, with a high intensity in " + }, + { + "bbox": [ + 130, + 265, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{SL^2}" + }, + { + "bbox": [ + 130, + 265, + 482, + 361 + ], + "type": "text", + "content": " on relevant pixels and low intensity on irrelevant pixels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 455 + ], + "type": "text", + "content": "Given the different purposes of pixels within an image, the frequency of activation may vary and can be visualized via AL. Since " + }, + { + "bbox": [ + 130, + 361, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathrm{SL}^2" + }, + { + "bbox": [ + 130, + 361, + 482, + 455 + ], + "type": "text", + "content": " and AL offer complementary information, they should be interpreted together. This approach offers a deeper understanding compared to a simplistic perspective, as it has been shown by Rahaman et al. [43] that high-frequency regions emerge later in training, are more challenging to fit, and are prone to change. On the contrary, low-frequency components of the decision boundary manifest earlier in training, are easier to fit, and exhibit greater stability with respect to noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": "It is important to clarify that \"high-frequency\" refers to features considered high-frequency by the model, and not necessarily frequency across spatial dimensions (e.g. width, height). Nonetheless, empirical findings demonstrate that high-frequency features in spatial domain are more likely to be identified as high frequency by the model (see Appendix I)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 533, + 218, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 218, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 218, + 547 + ], + "type": "text", + "content": "5 Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "type": "text", + "content": "Given the experimental discoveries that squared gradients lead to better results [30], we evaluate both scenarios for different hyperparameters and compare them with SL in Tabs. 1 and 2, refer to Appendix E for implementation details and reproducibility. For the purposes of this study, we use pixel removal strategy, despite its problems [30]. The spectral analysis of pixel removal strategy and possible solutions are deferred to future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "We report single-pixel most-important-first (MoRF) insertion and deletion scores of the methods to assess the design choices previously proposed by our theoretical framework: (1) utilizing squared gradients instead of gradients in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 90, + 447, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 219, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 219, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 219, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 219, + 198 + ], + "type": "image", + "image_path": "748a8d9134f66ebad795cb6170cdd6324a4fdf8226acf661c91c6173a4a42d83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 115, + 305, + 198 + ], + "blocks": [ + { + "bbox": [ + 223, + 115, + 305, + 198 + ], + "lines": [ + { + "bbox": [ + 223, + 115, + 305, + 198 + ], + "spans": [ + { + "bbox": [ + 223, + 115, + 305, + 198 + ], + "type": "image", + "image_path": "1bf4a60f123842bd0e33e7d4229d36537730d83575dba59526babc51101c6d06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 391, + 198 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 391, + 198 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 391, + 198 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 391, + 198 + ], + "type": "image", + "image_path": "772a0ff07b8a7e44411c06fca752f52abb2b569a465fee1bb7ca4c577914c191.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 116, + 477, + 198 + ], + "blocks": [ + { + "bbox": [ + 394, + 116, + 477, + 198 + ], + "lines": [ + { + "bbox": [ + 394, + 116, + 477, + 198 + ], + "spans": [ + { + "bbox": [ + 394, + 116, + 477, + 198 + ], + "type": "image", + "image_path": "f61bc8e42c32c6a0ce10e697499b84b7284c279125c24493aa109e49a85ac190.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 136, + 199, + 219, + 281 + ], + "blocks": [ + { + "bbox": [ + 136, + 199, + 219, + 281 + ], + "lines": [ + { + "bbox": [ + 136, + 199, + 219, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 199, + 219, + 281 + ], + "type": "image", + "image_path": "cc40f30369167a69d73bdea5a27af10b38294dbbbd1c1455353f9d1b9ae6bd87.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 284, + 184, + 293 + ], + "lines": [ + { + "bbox": [ + 171, + 284, + 184, + 293 + ], + "spans": [ + { + "bbox": [ + 171, + 284, + 184, + 293 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 200, + 305, + 281 + ], + "blocks": [ + { + "bbox": [ + 223, + 200, + 305, + 281 + ], + "lines": [ + { + "bbox": [ + 223, + 200, + 305, + 281 + ], + "spans": [ + { + "bbox": [ + 223, + 200, + 305, + 281 + ], + "type": "image", + "image_path": "d707b0fbd240bcc7fda17dcfbe7b08a4d9531c03ec08fe624b9e2d35a2aaeb4d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 284, + 271, + 293 + ], + "lines": [ + { + "bbox": [ + 258, + 284, + 271, + 293 + ], + "spans": [ + { + "bbox": [ + 258, + 284, + 271, + 293 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 200, + 391, + 281 + ], + "blocks": [ + { + "bbox": [ + 309, + 200, + 391, + 281 + ], + "lines": [ + { + "bbox": [ + 309, + 200, + 391, + 281 + ], + "spans": [ + { + "bbox": [ + 309, + 200, + 391, + 281 + ], + "type": "image", + "image_path": "1a932a9c1aeb327b28a9292702767225f034d638bae2be886111fe4f8591b6a8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 284, + 356, + 293 + ], + "lines": [ + { + "bbox": [ + 345, + 284, + 356, + 293 + ], + "spans": [ + { + "bbox": [ + 345, + 284, + 356, + 293 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "content": "Fig. 5: This figure shows example visualizations of explanations generated by " + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{SG^2}" + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "content": " implementing design decisions suggested in our work in Remarks 6 and 7. (b) shows the " + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{SG^2}" + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "content": " visualized at the maximum similarity according to Fig. 4, for this image " + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\sigma = 0.5" + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "content": ". (c) shows the SL attributions. Lastly, (d) shows the frequencies obtained from AL (defined in Eq. (10)), which correspond to the perturbation level at which " + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{SG^2}" + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "content": " is maximized. Since we are interested in frequencies rather than the actual perturbation levels, we visualize 2 stds above and below the mean of frequencies obtained by " + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\omega = 1 / (1 + \\mathrm{AL})" + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "content": " (see Appendix I for more examples). The heatmaps in (b) and (c) show how much the model is relying on the pixel, while (d) shows the frequency at which the feature's contribution maximizes. In the provided examples, we can see that the model is relying on the nose (top) and eye (bottom) while the same region is of a lower frequency compared to the head. This can be interpreted as the model finding the nose and eye more stable features compared to the head." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 394, + 200, + 477, + 281 + ], + "blocks": [ + { + "bbox": [ + 394, + 200, + 477, + 281 + ], + "lines": [ + { + "bbox": [ + 394, + 200, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 394, + 200, + 477, + 281 + ], + "type": "image", + "image_path": "961e15fc803ece74636792d3c11d15aa26594b32db6c08462fd39ef4337e4f08.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 284, + 443, + 293 + ], + "lines": [ + { + "bbox": [ + 430, + 284, + 443, + 293 + ], + "spans": [ + { + "bbox": [ + 430, + 284, + 443, + 293 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 554 + ], + "type": "text", + "content": "Remark 5, (2) finding a standard perturbation scale in Remark 6, and (3) employing a uniform explanation prior instead of Dirac, stated in Remark 7 (see Fig. 7 for an ablation study of kernel variance and Appendix G for an ablation study of patch removal instead of single pixel removal). We evaluated the suggested design choices on ImageNet [20] and Food101 [1] in Tab. 1. However, smaller datasets, such as CIFAR, are omitted because their resolution restricts the potential gap between high- and low-frequency features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Limitations of Our Framework. The choice of independent Gaussian perturbation is limited due to its inability to incorporate prior knowledge about the input space or data manifold characteristics, and other data modalities such as graphs and time series. However, preferring simpler analytical forms within our framework limits the alternative choices. Efficiency is another challenge compared to basic methods, as performing back-props at various noise scales is costly. Moreover, our attribution of frequencies in visualizations, such as SL, may lack the desired generalizability across data points, as high-frequency re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 365, + 449, + 525 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": "Table 1: This table presents the assessment of the proposed design choices in Remarks 5 to 7 across different methods, highlighting statistically significant improvements with bold fonts. We conduct t-tests comparing gradient methods to their squared-gradient counterparts to evaluate the design choice in Remark 5. Furthermore, we conduct t-tests between " + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{SG}^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": " with the original heuristic and those ensuring consistency to evaluate design decisions in Remarks 6 and 7. Overlines and underlines denote highest or lowest values within columns, respectively. Although there are no squared versions of Occlusion Sensitivity (OS) [5] and RISE [42], they are included as low-frequency baselines. While the dimensions of boolean masks utilized in OS and RISE can implicitly regulate the sampling variance (see Appendix C), the exponential increase in the required number of samples in high-frequency explanations renders them less practical. SG scores follow the heuristic of the SmoothGrad paper [55] i.e. " + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\sigma = (\\max(x) - \\min(x))" + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": " 0.1. XIG indicates the input multiplication in the original implementation of IntegratedGrad [57], while IG does not apply such transformation. Rankings for methods using gradients are obtained from negative gradients, including RISE and Occlusion Sensitivity, according to Eq. (2). Furthermore, " + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{SG}_{\\mathrm{Opt.}}^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": " uses an optimized kernel per image and " + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{SL}^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": " uses a uniform explanation prior. Comparing SG² and " + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{SG}_{\\mathrm{Opt.}}^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": ", we can see that " + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{SG}_{\\mathrm{Opt.}}^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": " can adapt to the statistics of the Food101. Therefore, " + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{SG}_{\\mathrm{Opt.}}^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 357 + ], + "type": "text", + "content": ", while not subject to inconsistencies (shown in Fig. 1), can improve the performance due to its adaptability. Hence, we can ensure the consistency of SG² without compromising its performance. Also, note that the input multiplication enhances visual quality, but degrades the performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 365, + 449, + 525 + ], + "lines": [ + { + "bbox": [ + 165, + 365, + 449, + 525 + ], + "spans": [ + { + "bbox": [ + 165, + 365, + 449, + 525 + ], + "type": "table", + "html": "
ImageNetFood101
Insertion(↑)Deletion(↓)Insertion(↑)Deletion(↓)
GradientsOS0.384 ± 0.0080.237 ± 0.0100.378 ± 0.0110.209 ± 0.009
RISE0.494 ± 0.0140.174 ± 0.0070.410 ± 0.0130.198 ± 0.007
SG0.212 ± 0.0100.179 ± 0.0090.179 ± 0.0080.137 ± 0.007
XIG0.208 ± 0.0090.204 ± 0.0080.101 ± 0.0070.177 ± 0.007
IG0.201 ± 0.0100.182 ± 0.0090.145 ± 0.0060.131 ± 0.005
Gradients2SG20.465 ± 0.0130.157 ± 0.0070.371 ± 0.0120.162 ± 0.006
XIG20.334 ± 0.0140.187 ± 0.0080.204 ± 0.0100.263 ± 0.007
IG20.340 ± 0.0120.168 ± 0.0070.320 ± 0.0090.135 ± 0.005
Gradients2Opt.0.491 ± 0.0150.187 ± 0.0070.394 ± 0.0130.167 ± 0.006
SL20.493 ± 0.0160.169 ± 0.0070.395 ± 0.0140.157 ± 0.007
", + "image_path": "785c663824415506b622e5327b8b9af3375350e1363b63664bc95e201293ddff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "content": "gions in one image may correspond to low-frequency regions in another image (see Appendix I). Identifying generalizable attributions across images remains a subject for future investigation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "type": "text", + "content": "6 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Probabilistic and Spectral View of Explanations. Despite sporadic attempts to analyze explanations from a probabilistic view [17, 59], this view is more common in the literature around the uncertainty quantification of expla" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "nations [29,40,54]. Our work is closely related to Covert et al. [19], if their work is seen from a probabilistic perspective. Aside from the works around the spectral view of deep networks [10,11,15,43,58], as of our current knowledge, there is no work adopting a spectral view of explanations along pixels' value. Nonetheless, there are works that build on a spectral view along spatial dimensions [34,35], extending an optimization-based definition of explanations [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 205, + 483, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 483, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 483, + 302 + ], + "type": "text", + "content": "Unifying Views to Explanations. Covert et al. [19] propose an information theoretic framework that unifies several explanation methods, focusing on three design choices: (1) feature removal method, (2) which model behavior to analyze, and (3) summary statistics. Our probabilistic framework, being similar to Covert's, offers a distinct viewpoint relevant to our spectral analysis. It can be used as a unifying view of explanations (see Appendix C), inheriting from Covert's framework. Notably, our work aligns with the broader context of research under finding unifying representations of explanations [3,26,37]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 483, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 483, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 483, + 465 + ], + "type": "text", + "content": "Input-space or Function-space Perturbations. Methods such as NoiseGrad [14], Grad-CAM [51], RectifiedGrad [32], Layer-wise Relevance Propagation [7], DeepLift [53], and even Sharpness-Aware Minimization (SAM) [24] employ function space perturbations. Utilizing such distributions offers data modality independence, whereas input space perturbation remains agnostic about the model architecture. While acknowledging the duality between function and input space distributions, our focus on explicit input space distributions excludes these methods from our study. Although we recognize the utility of integrating function space perturbation into our mathematical model, we defer this aspect to future investigations. It can be noted, however, that extending this work to function space perturbation will connect it to the literature on double descent and smooth interpolation of samples [25, 41]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 481, + 220, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 220, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 220, + 495 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 506, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 639 + ], + "type": "text", + "content": "In this study, we deepen understanding of pixel attribution methodologies by rigorously analyzing their theoretical underpinnings, revealing a common spectral bias. Our analysis supports the experimental finding of recent works regarding the efficacy of using squared gradients. Moreover, we demonstrate how spectral properties can lead to inconsistent explanations and propose two design choices informed by our theoretical framework to address these inconsistencies: standardizing hyperparameters based on cosine similarity between the perturbation kernel and classifier, and aggregating explanations with a given prior. Despite the effectiveness of our approach, we acknowledge limitations and suggest avenues for further research. Lastly, we assess the suggested design choices using deletion and insertion scores." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "We would like to thank Giovanni Luca Marchetti for an early review of our work and his kind feedback. This project is partially supported by Region Stockholm through MedTechLabs, and Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Scientific computation was enabled by the supercomputing resource Berzelius, provided by the National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg foundation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 269, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 269, + 481, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 291 + ], + "type": "text", + "content": "1. Food-101 - Mining Discriminative Components with Random Forests, https://data.vision.ee.ethz.ch/cvl/datasets-extra/food-101/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 292, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 481, + 324 + ], + "type": "text", + "content": "2. Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., Kim, B.: Sanity Checks for Saliency Maps (Nov 2020). https://doi.org/10.48550/arXiv.1810.03292, http://arxiv.org/abs/1810.03292, arXiv:1810.03292 [cs, stat]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 325, + 481, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 481, + 368 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 481, + 368 + ], + "type": "text", + "content": "3. Agarwal, S., Jabbari, S., Agarwal, C., Upadhyay, S., Wu, Z.S., Lakkaraju, H.: Towards the Unification and Robustness of Perturbation and Gradient Based Explanations (Jul 2021). https://doi.org/10.48550/arXiv.2102.10618, http://arxiv.org/abs/2102.10618, arXiv:2102.10618 [cs]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 481, + 402 + ], + "type": "text", + "content": "4. Alvarez-Melis, D., Jaakkola, T.S.: On the Robustness of Interpretability Methods (Jun 2018). https://doi.org/10.48550/arXiv.1806.08049, http://arxiv.org/abs/1806.08049, arXiv:1806.08049 [cs, stat]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 402, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 481, + 446 + ], + "type": "text", + "content": "5. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for Deep Neural Networks (Mar 2018). https://doi.org/10.48550/arXiv.1711.06104, http://arxiv.org/abs/1711.06104, arXiv:1711.06104 [cs, stat]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "type": "text", + "content": "6. Arrieta, A.B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., Chatila, R., Herrera, F.: Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI (Dec 2019). https://doi.org/10.48550/arXiv.1910.10045, http://arxiv.org/abs/1910.10045, arXiv:1910.10045 [cs]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 501, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 555 + ], + "type": "text", + "content": "7. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS ONE 10(7), e0130140 (Jul 2015). https://doi.org/10.1371/journal.pone.0130140, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0130140" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 555, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 600 + ], + "type": "text", + "content": "8. Balduzzi, D., Frean, M., Leary, L., Lewis, J.P., Ma, K.W.D., McWilliams, B.: The Shattered Gradients Problem: If resnets are the answer, then what is the question? (Jun 2018). https://doi.org/10.48550/arXiv.1702.08591, http://arxiv.org/abs/1702.08591, arXiv:1702.08591 [cs, stat]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "9. Bansal, N., Agarwal, C., Nguyen, A.: SAM: The Sensitivity of Attribution Methods to Hyperparameters (Apr 2020). https://doi.org/10.48550/arXiv.2003.08754, http://arxiv.org/abs/2003.08754, arXiv:2003.08754 [cs]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "10. Basri, R., Galun, M., Geifman, A., Jacobs, D., Kasten, Y., Kritchman, S.: Frequency Bias in Neural Networks for Input of Non-Uniform Density (Mar 2020), https://api-semanticscholar.org/CorpusID:212644664" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "11. Benbarka, N., Hofer, T., Ul-Moqeet Riaz, H., Zell, A.: Seeing Implicit Neural Representations as Fourier Series. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 2283-2292 (Jan 2022). https://doi.org/10.1109/WACV51458.2022.00234, https://ieeexplore.ieee.org/document/9706795/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "text", + "content": "12. Blackman, R.B., Tukey, J.W.: The measurement of power spectra from the point of view of communications engineering — Part I. The Bell System Technical Journal 37(1), 185–282 (Jan 1958). https://doi.org/10.1002/j.1538-7305.1958.tb03874.x, http://ieeexplore.ieee.org/document/6768513/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "type": "text", + "content": "13. Brughmans, D., Melis, L., Martens, D.: Disagreement amongst counterfactual explanations: How transparency can be deceptive (Apr 2023). https://doi.org/10.48550/arXiv.2304.12667, http://arxiv.org/abs/2304.12667 [cs]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "text", + "content": "14. Bykov, K., Hedström, A., Nakajima, S., Höhne, M.M.C.: NoiseGrad: Enhancing Explanations by Introducing Stochasticity to Model Weights (May 2022). https://doi.org/10.48550/arXiv.2106.10185, http://arxiv.org/abs/2106.10185, arXiv:2106.10185 [cs]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 297, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 361 + ], + "type": "text", + "content": "15. Cao, Y., Fang, Z., Wu, Y., Zhou, D.X., Gu, Q.: Towards Understanding the Spectral Bias of Deep Learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. pp. 2205-2211. International Joint Conferences on Artificial Intelligence Organization, Montreal, Canada (Aug 2021). https://doi.org/10.24963/ijcai.2021/304, https://www.ijcai.org/proceedings/2021/304" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "16. Chen, H., Janizek, J.D., Lundberg, S., Lee, S.I.: True to the Model or True to the Data? (Jun 2020). https://doi.org/10.48550/arXiv.2006.16234, http://arxiv.org/abs/2006.16234, arXiv:2006.16234 [cs, stat]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 396, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 441 + ], + "type": "text", + "content": "17. Chen, J., Song, L., Wainwright, M.J., Jordan, M.I.: Learning to Explain: An Information-Theoretic Perspective on Model Interpretation (Jun 2018). https://doi.org/10.48550/arXiv.1802.07814, http://arxiv.org/abs/1802.07814, arXiv:1802.07814 [cs, stat]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "text", + "content": "18. Cohen, L.: The generalization of the Wiener-Khinchin theorem. In: Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181). vol. 3, pp. 1577-1580. IEEE (1998), https://ieeexplore.ieee.org/abstract/document/681753/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "text", + "content": "19. Covert, I., Lundberg, S., Lee, S.I.: Explaining by Removing: A Unified Framework for Model Explanation (May 2022). https://doi.org/10.48550/arXiv.2011.14878, http://arxiv.org/abs/2011.14878, arXiv:2011.14878 [cs, stat]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 521, + 481, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 574 + ], + "type": "text", + "content": "20. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (Jun 2009). https://doi.org/10.1109/CVPR.2009.5206848, https://ieeexplore.ieee.org/document/5206848, iSSN: 1063-6919" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 576, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 632 + ], + "type": "text", + "content": "21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Jun 2021). https://doi.org/10.48550/arXiv.2010.11929, http://arxiv.org/abs/2010.11929, arXiv:2010.11929 [cs]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "22. Fong, R., Patrick, M., Vedaldi, A.: Understanding Deep Networks via Extremal Perturbations and Smooth Masks (Oct 2019). https://doi.org/10.48550/arXiv.1910.08485, http://arxiv.org/abs/1910.08485, arXiv:1910.08485 [cs, stat]" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "text", + "content": "23. Fong, R., Vedaldi, A.: Interpretable Explanations of Black Boxes by Meaningful Perturbation. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 3449-3457 (Oct 2017). https://doi.org/10.1109/ICCV.2017.371, http://arxiv.org/abs/1704.03296, arXiv:1704.03296 [cs, stat]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 204 + ], + "type": "text", + "content": "24. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-Aware Minimization for Efficiently Improving Generalization (Apr 2021). https://doi.org/10.48550/arXiv.2010.01412, http://arxiv.org/abs/2010.01412, arXiv:2010.01412 [cs, stat]" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "25. Gamba, M., Azizpour, H., Björkman, M.: On the Lipschitz Constant of Deep Networks and Double Descent (Nov 2023). https://doi.org/10.48550/arXiv.2301.12309, http://arxiv.org/abs/2301.12309, arXiv:2301.12309 [cs]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 235, + 481, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 280 + ], + "type": "text", + "content": "26. Han, T., Srinivas, S., Lakkaraju, H.: Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations (Dec 2022). https://doi.org/10.48550/arXiv.2206.01254, http://arxiv.org/abs/2206.01254, arXiv:2206.01254 [cs]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "27. Harel, N., Gilad-Bachrach, R., Obolski, U.: Inherent Inconsistencies of Feature Importance (Jun 2022). https://doi.org/10.48550/arXiv.2206.08204, http://arxiv.org/abs/2206.08204, arXiv:2206.08204 [cs]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "type": "text", + "content": "28. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition (Dec 2015). https://doi.org/10.48550/arXiv.1512.03385, http://arxiv.org/abs/1512.03385, arXiv:1512.03385 [cs]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 342, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 376 + ], + "type": "text", + "content": "29. Hill, D., Masoomi, A., Ghimire, S., Torop, M., Dy, J.: Explanation Uncertainty with Decision Boundary Awareness (Nov 2022). https://doi.org/10.48550/arXiv.2210.02419, http://arxiv.org/abs/2210.02419, arXiv:2210.02419 [cs]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "text", + "content": "30. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A Benchmark for Interpretability Methods in Deep Neural Networks (Nov 2019). https://doi.org/10.48550/arXiv.1806.10758, http://arxiv.org/abs/1806.10758, arXiv:1806.10758 [cs, stat]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "type": "text", + "content": "31. Hou, X., Zhang, L.: Saliency Detection: A Spectral Residual Approach. In: 2007 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1-8 (Jun 2007). https://doi.org/10.1109/CVPR.2007.383267, https://ieeexplore.ieee.org/document/4270292, iSSN: 1063-6919" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 460, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 504 + ], + "type": "text", + "content": "32. Kim, B., Seo, J., Jeon, S., Koo, J., Choe, J., Jeon, T.: Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps (Sep 2019). https://doi.org/10.48550/arXiv.1902.04893, http://arxiv.org/abs/1902.04893, arXiv:1902.04893 [cs, stat]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 503, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 481, + 548 + ], + "type": "text", + "content": "33. Kindermans, P.J., Hooker, S., Adebayo, J., Alber, M., Schütt, K.T., Dähne, S., Erhan, D., Kim, B.: The (Un)reliability of saliency methods (Nov 2017). https://doi.org/10.48550/arXiv.1711.00867, http://arxiv.org/abs/1711.00867, arXiv:1711.00867 [cs, stat]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "text", + "content": "34. Kolek, S., Nguyen, D.A., Levie, R., Bruna, J., Kutyniok, G.: Cartoon Explanations of Image Classifiers (Oct 2022). https://doi.org/10.48550/arXiv.2110.03485, http://arxiv.org/abs/2110.03485, arXiv:2110.03485 [cs]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 578, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 622 + ], + "type": "text", + "content": "35. Kolek, S., Windsesheim, R., Loarca, H.A., Kutyniok, G., Levie, R.: Explaining Image Classifiers with Multiscale Directional Image Representation (Apr 2023). https://doi.org/10.48550/arXiv.2211.12857, http://arxiv.org/abs/2211.12857, arXiv:2211.12857 [cs]" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "text", + "content": "36. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The Disagreement Problem in Explainable Machine Learning: A Practitioner's Perspective (Feb 2022). https://doi.org/10.48550/arXiv.2202.01602, http://arxiv.org/abs/2202.01602, arXiv:2202.01602 [cs]" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "37. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 481, + 192 + ], + "type": "text", + "content": "38. Marmarelis, P.Z., Marmarelis, V.Z.: The White-Noise Method in System Identification. In: Marmarelis, P.Z., Marmarelis, V.Z. (eds.) Analysis of Physiological Systems: The White-Noise Approach, pp. 131-180. Computers in Biology and Medicine, Springer US, Boston, MA (1978). https://doi.org/10.1007/978-1-4613-3970-0_4, https://doi.org/10.1007/978-1-4613-3970-0_4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 235 + ], + "type": "text", + "content": "39. Marques-Silva, J., Ignatiev, A.: Delivering Trustworthy AI through Formal XAI. Proceedings of the AAAI Conference on Artificial Intelligence 36(11), 12342-12350 (Jun 2022). https://doi.org/10.1609/aaai.v36i11.21499, https://ojs.aaai.org/index.php/AAAI/article/view/21499" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 235, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 278 + ], + "type": "text", + "content": "40. Marx, C., Park, Y., Hasson, H., Wang, Y., Ermon, S., Huan, L.: But Are You Sure? An Uncertainty-Aware Perspective on Explainable AI. In: Proceedings of The 26th International Conference on Artificial Intelligence and Statistics. pp. 7375-7391. PMLR (Apr 2023), https://proceedings.mlr.press/v206/marx23a.htm1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "type": "text", + "content": "41. Nakkiran, P., Kaplun, G., Bansal, Y., Yang, T., Barak, B., Sutskever, I.: Deep Double Descent: Where Bigger Models and More Data Hurt (Dec 2019). https://doi.org/10.48550/arXiv.1912.02292, http://arxiv.org/abs/1912.02292, arXiv:1912.02292 [cs, stat]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "text", + "content": "42. Petsiuk, V., Das, A., Saenko, K.: RISE: Randomized Input Sampling for Explanation of Black-box Models (Sep 2018). https://doi.org/10.48550/arXiv.1806.07421, http://arxiv.org/abs/1806.07421, arXiv:1806.07421 [cs]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 354, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 397 + ], + "type": "text", + "content": "43. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F.A., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks (May 2019). https://doi.org/10.48550/arXiv.1806.08734, http://arxiv.org/abs/1806.08734, arXiv:1806.08734 [cs, stat]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "44. Rahman, M.: Applications of Fourier Transforms to Generalized Functions. WIT Press (2011), google-Books-ID: k_rdcKaUdr4C" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 418, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 493 + ], + "type": "text", + "content": "45. Ras, G., van Gerven, M., Haselager, P.: Explanation Methods in Deep Learning: Users, Values, Concerns and Challenges. In: Escalante, H.J., Escalera, S., Guyon, I., Baró, X., Güçlütürk, Y., Güçlü, U., van Gerven, M. (eds.) Explainable and Interpretable Models in Computer Vision and Machine Learning, pp. 19-36. The Springer Series on Challenges in Machine Learning, Springer International Publishing, Cham (2018). https://doi.org/10.1007/978-3-319-98131-4_2, https://doi.org/10.1007/978-3-319-98131-4_2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "46. Reiter, H., Stegeman, J.D.: Classical Harmonic Analysis and Locally Compact Groups. Oxford University Press (Nov 2000). https://doi.org/10.1093/oso/9780198511892.001.0001, https://academic.oup.com/book/54460" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 525, + 481, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 558 + ], + "type": "text", + "content": "47. Ribeiro, M.T., Singh, S., Guestrin, C.: \"Why Should I Trust You?\": Explaining the Predictions of Any Classifier (Aug 2016). https://doi.org/10.48550/arXiv.1602.04938, http://arxiv.org/abs/1602.04938, arXiv:1602.04938 [cs, stat]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 558, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 601 + ], + "type": "text", + "content": "48. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A Consistent and Efficient Evaluation Strategy for Attribution Methods (Jun 2022). https://doi.org/10.48550/arXiv.2202.00449, http://arxiv.org/abs/2202.00449, arXiv:2202.00449 [cs]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "text", + "content": "49. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence 1(5), 206-215 (May 2019). https://doi.org/10.1038/s42256-019-0048-x, https://www.nature.com/articles/s42256-019-0048-x" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "50. Rudin, W.: Principles of Mathematical Analysis. McGraw-Hill (1976), google-Books-ID: kwqzPAAACAAJ" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Mehrpanah et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 600 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "text", + "content": "51. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision 128(2), 336-359 (Feb 2020). https://doi.org/10.1007/s11263-019-01228-7, http://arxiv.org/abs/1610.02391, arXiv:1610.02391 [cs]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 172, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 225 + ], + "type": "text", + "content": "52. Serov, V.: The Riemann-Lebesgue Lemma. In: Serov, V. (ed.) Fourier Series, Fourier Transform and Their Applications to Mathematical Physics, pp. 33-35. Applied Mathematical Sciences, Springer International Publishing, Cham (2017). https://doi.org/10.1007/978-3-319-65262-7_6, https://doi.org/10.1007/978-3-319-65262-7_6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "type": "text", + "content": "53. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not Just a Black Box: Learning Important Features Through Propagating Activation Differences (Apr 2017). https://doi.org/10.48550/arXiv.1605.01713, http://arxiv.org/abs/1605.01713, arXiv:1605.01713 [cs]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 313 + ], + "type": "text", + "content": "54. Slack, D., Hilgard, S., Singh, S., Lakkaraju, H.: Reliable Post hoc Explanations: Modeling Uncertainty in Explainability (Nov 2021). https://doi.org/10.48550/arXiv.2008.05030, http://arxiv.org/abs/2008.05030, arXiv:2008.05030 [cs, stat]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "text", + "content": "55. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: SmoothGrad: removing noise by adding noise (Jun 2017). https://doi.org/10.48550/arXiv.1706.03825, http://arxiv.org/abs/1706.03825, arXiv:1706.03825 [cs, stat]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 347, + 482, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 370 + ], + "type": "text", + "content": "56. Stoica, P., Moses, R.L.: Spectral analysis of signals, vol. 452. Pearson Prentice Hall Upper Saddle River, NJ (2005), http://user.it.uu.se/~ps/SAS-new.pdf" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "type": "text", + "content": "57. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic Attribution for Deep Networks (Jun 2017). https://doi.org/10.48550/arXiv.1703.01365, http://arxiv.org/abs/1703.01365, arXiv:1703.01365 [cs]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "type": "text", + "content": "58. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. ArXiv (Jun 2020), https://api_semanticscholar.org/CorpusID:219791950" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "59. Wang, E., Khosravi, P., Broeck, G.V.d.: Probabilistic Sufficient Explanations (May 2021). https://doi.org/10.48550/arXiv.2105.10118, http://arxiv.org/abs/2105.10118, arXiv:2105.10118 [cs]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 482, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 521 + ], + "type": "text", + "content": "60. Wang, Y., Wang, X.: \"Why Not Other Classes?\": Towards Class-Contrastive Backpropagation Explanations. Advances in Neural Information Processing Systems 35, 9085-9097 (Dec 2022), https://proceedings.neurips.cc/paper_files/paper/2022/bitical/3b7a66b2d1258e892c89f485b8f896e0-AAbstract-Conference.html" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 522, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 567 + ], + "type": "text", + "content": "61. Wilming, R., Kieslich, L., Clark, B., Haufe, S.: Theoretical Behavior of XAI Methods in the Presence of Suppressor Variables (Jun 2023). https://doi.org/10.48550/arXiv.2306.01464, http://arxiv.org/abs/2306.01464 [cs, stat]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "text", + "content": "62. Zeiler, M.D., Fergus, R.: Visualizing and Understanding Convolutional Networks (Nov 2013). https://doi.org/10.48550/arXiv.1311.2901, http://arxiv.org/abs/1311.2901, arXiv:1311.2901 [cs]" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spectral Properties of Gradient-based Explanation Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_content_list.json b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..77795baf805fde4eae1f53245b05fc2cd855a0d7 --- /dev/null +++ b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_content_list.json @@ -0,0 +1,1970 @@ +[ + { + "type": "text", + "text": "On the Approximation Risk of Few-Shot Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 269, + 140, + 733, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuan Wang $^{1,2}$ , Zhong Ji $^{1,2}$ , Xiyao Liu $^{3}$ , Yanwei Pang $^{1,2}$ , and Jungong Han $^{4}$", + "bbox": [ + 220, + 210, + 781, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tianjin University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ State Key Laboratory of Robotics, SIA, CAS $^{4}$ University of Sheffield", + "bbox": [ + 253, + 253, + 746, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Few-Shot Class-Incremental Learning (FSCIL) aims to learn new concepts with few training samples while preserving previously acquired knowledge. Although promising performance has been achieved, there remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Therefore, we thoroughly explore the approximation risk of FSCIL, encompassing both transfer and consistency risks. By tightening the upper bounds of these risks, we derive practical guidelines for designing and training FSCIL models. These guidelines include (1) expanding training datasets for base classes, (2) preventing excessive focus on specific features, (3) optimizing classification margin discrepancy, and (4) ensuring unbiased classification across both base and novel classes. Leveraging these insights, we conduct comprehensive experiments to validate our principles, achieving state-of-the-art performance on three FSCIL benchmark datasets. Code is available at https://github.com/xwangrs/Approximation_FSCIL-ECCV2024.git.", + "bbox": [ + 259, + 316, + 743, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Few-Shot Class-Incremental Learning $\\cdot$ Few-Shot Learning $\\cdot$ Class Incremental Learning $\\cdot$ Approximation Risk", + "bbox": [ + 261, + 537, + 741, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 590, + 375, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-Shot Class-Incremental Learning (FSCIL) [4, 23, 34], an emergent learning paradigm, aims to learn novel classes with limited training data, while preserving the knowledge acquired from base classes. This paradigm faces two primary issues: (1) the catastrophic forgetting of established base class knowledge when integrating new information, and (2) the overfitting to few-shot novel classes.", + "bbox": [ + 212, + 619, + 787, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To effectively address these issues, researchers have proposed various approaches, broadly categorized into three groups [34]: data-based [13, 17, 38], structure-based [23, 32, 36], and optimization-based [2, 4, 35] approaches. Despite their proven effectiveness, there still remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Central to this gap is the question: What are the key factors in solving FSCIL problems? To answer this question, our research delves into the basic statistical principles of FSCIL, aiming to guide the design and training of FSCIL models.", + "bbox": [ + 212, + 696, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: Zhong Ji, Email: jizhong@tju.edu.cn", + "bbox": [ + 217, + 824, + 632, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff4f8134f9ea14acd57e3a7063a331d7a92653bd8d7e1fad08e3312ce5f0bdcc.jpg", + "image_caption": [ + "(a) Transfer Risk" + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 498, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3446c78168d56bbc9b5a2334ec2bebf819807655c980ad745d92ac29386943db.jpg", + "image_caption": [ + "(b) Consistency Risk", + "Fig. 1: The upper bounds of the transfer and consistency risks. In Fig. 1 (a), the horizontal axis represents the number of training samples for base classes, whereas the vertical axis quantifies the magnitude of the transfer risk. Similarly, in Fig. 1 (b), the horizontal axis reflects the classification margin discrepancy, while the vertical axis quantifies the magnitude of the consistency risk. The parameters $\\eta > 0$ and $\\kappa > 0$ serve as dynamic scaling factors, regulating the transfer and consistency risks." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 743, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin with fundamental assumptions: Homogeneity, Regularity conditions, and Realizability. These assumptions, grounded in practical scenarios, form the foundation of our theory.", + "bbox": [ + 214, + 407, + 787, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these assumptions, we formulate the gap between empirically obtained and theoretically optimal FSCIL models as an optimization problem, whose objective function is defined as the Approximation Risk of FSCIL. This risk includes two components: transfer risk and consistency risk. The transfer risk [24] evaluates the efficacy of the shared representations in knowledge transfer from base to novel classes. A low transfer risk indicates effective utilization of base classes knowledge to compensate for the scarcity of novel class samples, thereby preventing overfitting in these novel classes. Meanwhile, the consistency risk measures performance drop when incorporating novel classes, with a low risk signifying robust and stable performance throughout class-incremental tasks, thereby preventing the catastrophic forgetting of base classes.", + "bbox": [ + 212, + 452, + 787, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging statistical learning theory [6, 14, 21, 26], we explore the upper bounds of transfer and consistency risks, then identifying their determinants2:", + "bbox": [ + 214, + 619, + 787, + 648 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\eta \\cdot \\frac {1}{\\sqrt {N}}} _ {\\text {T r a n s f e r R i s k}} + \\underbrace {\\log \\left(1 + \\frac {\\kappa}{\\exp (\\nu)}\\right)} _ {\\text {C o n s i s t e n c y R i s k}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 655, + 785, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $N$ denotes the number of training samples for the base classes; $\\nu$ indicates the classification margin discrepancy, defined as the infimum of classifier logits for correct classifications minus the supremum of those for incorrect classifications; the parameters $\\eta$ and $\\kappa$ are the dynamic scaling factors for transfer and consistency risks, respectively. We will further discuss and elaborate on methods for minimizing these risks by tightening their upper bounds.", + "bbox": [ + 214, + 712, + 787, + 804 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Proofs in Section 4.", + "bbox": [ + 217, + 810, + 367, + 824 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2 Proofs in Section 5.", + "bbox": [ + 218, + 824, + 364, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Fig. 1 (a), it is evident that increasing the number of training samples for base classes leads to a significant reduction in the transfer risk. Theoretically, this risk could potentially be minimized to a negligible level with a sufficiently large and diverse dataset. However, in practical scenarios, gathering sufficient data for base classes remains a challenge, thereby prompting reliance on foundation models [19,31]. Crucially, for the Homogeneity assumption to hold, the marginal distribution of the pre-trained data should align with that of the base class data. Failing to meet this condition could lead to an increase in the transfer risk.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In exploring how $N$ affects the transfer risk, it becomes imperative to fully comprehend the role of the dynamic scaling factor $\\eta$ . From a theoretical perspective, $\\eta$ represents a multivariate polynomial that is related to the maximum activation values across each layer of the neural networks. This could be interpreted as the model's selective attention to specific features. An increase in $\\eta$ values indicates a marked emphasis on certain features, often at the expense of neglecting others, thereby increasing the transfer risk. This phenomenon is known as the \"supervision collapse\" [3, 9]. To mitigate this issue, we propose to maintain a balanced distribution of activation values across the network layers to prevent extreme impulse values. Furthermore, our theoretical analysis indicates that Vision Transformers (ViTs) [5], owing to their self-attention mechanism [25], exhibit a lower transfer risk compared to Convolutional Neural Networks (CNNs) [8]. Thus, in addressing the FSCIL problem, our preference is towards the deployment of ViT models.", + "bbox": [ + 212, + 267, + 784, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evidenced in Fig. 1 (b), a clear inverse relationship exists between the classification margin discrepancy and the consistency risk, indicating that enhancing the inter-class discrimination of representations can theoretically diminish the consistency risk to minimal levels. This requires models to maximize the classification margin discrepancy on base classes, ensuring comprehensive training of these models. Moreover, our theory also indicates that the scaling factor $\\kappa$ reaches its lower bound when the model exhibits unbiased classification across both base and novel classes. Achieving this balance involves training classifiers for each task independently before combining them into a unified classifier.", + "bbox": [ + 212, + 479, + 784, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 240, + 614, + 553, + 628 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Theoretical Framework: We develop a new optimization theory that provides a novel perspective for understanding and addressing challenges in FSCIL.", + "- Risk Analysis: We decompose the approximation risk of FSCIL into transfer and consistency risks, defining their upper boundaries and and key determinants, thereby offering guidance for model design and training.", + "- Practical Strategies: We propose practical guidelines to address the challenges in FSCIL. These include expanding training datasets for base classes, preventing excessive focus on specific features, optimizing classification margin discrepancy, and ensuring unbiased classification across both base and novel classes." + ], + "bbox": [ + 225, + 643, + 784, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Informed by these insights, we achieve state-of-the-art performance on three benchmark datasets, demonstrating the effectiveness of our theory.", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, FSCIL [23,28] has emerged as a paradigm to incrementally learn novel classes with a limited number of samples, while preserving previously learned knowledge. According to the survey [34], existing approaches can be broadly categorized into three groups: data-based approaches [13,17,38], structure-based approaches [23,32,36], and optimization-based approaches [2,4,35].", + "bbox": [ + 212, + 176, + 785, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The data-based approaches address FSCIL from the perspective of training data, which involves data replay-based methods [13] and pseudo scenarios-based methods [17, 38]. For example, Kukleva et al. [13] propose a three-stage framework where the base and the novel class samples are efficiently replayed to calibrate the classification bias. Zhou et al. [38] propose ForwArd Compatible Training (FACT) for FSCIL, which preassigns multiple virtual prototypes in the embedding space by creatively mixuping the known class samples, and Peng et al. [17] have similar ideas about utilizing the Mixup [33] to expand the datasets.", + "bbox": [ + 212, + 253, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The structure-based approaches address FSCIL from the perspective of network architecture design, which involves dynamic structure-based methods [23, 32] and attention-based methods [36]. For example, Zhang et al. [32] propose Continually Evolved Classifier (CEC) that employs a graph model to propagate context information between classifiers for adaptation. Tao et al. [23] propose to use a neural \"gas\" network to learn and preserve the topology of the feature manifold formed by the base and the novel classes in FSCIL. Zhao et al. [36] propose an attention-based aggregation module that selectively merges predictions from the base branch and the novel branch.", + "bbox": [ + 212, + 375, + 787, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The optimization-based approaches address FSCIL from the perspective of algorithm optimization, which involves representation learning-based methods [35] and knowledge distillation methods [2, 4]. Zhao et al. [35] concentrate on the dilemma between the slow forgetting of old knowledge and the fast adaptation to novel knowledge. Then a multi-grained \"slow vs fast\" learning strategy is proposed to cope with this dilemma from both intra-space and the inter-space. Dong et al. [4] propose a exemplar relation distillation incremental learning framework to balance the tasks of old-knowledge preserving and new-knowledge adaptation. Cui et al. [2] propose to distill reliable knowledge from the reference model, then implement an uncertainty-aware distillation module that combines uncertainty-guided knowledge refinement with adaptive distillation.", + "bbox": [ + 212, + 510, + 787, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Beyond the above approaches, we explore the basic statistical principles underlying FSCIL, then conclude with practical guidelines for the design and training of FSCIL models.", + "bbox": [ + 212, + 676, + 785, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 746, + 370, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Definition. FSCIL involves a base task followed by a sequence of incremental tasks, each identified by a unique task identifier $t$ . Each task is constituted by a distinct training set, $\\mathcal{X}_{train}^{t}$ , and a testing set, $\\mathcal{X}_{test}^{t}$ , which are associated with their respective distinct label spaces $\\mathcal{Y}^t$ , such that $\\forall i,j$ with", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$i \\neq j$ , $\\mathcal{Y}^i \\cap \\mathcal{Y}^j = \\emptyset$ . For a particular task $t$ , training is confined to $\\mathcal{X}_{train}^t$ , while testing is performed across all encountered tasks. Notably, $\\mathcal{X}_{train}^0$ encompasses a substantial number of samples for the base task, while $\\mathcal{X}_{train}^{t > 0}$ are limited to $M$ samples per task.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To tackle FSCIL, popular work [17,30-32] assumes a shared feature representation, enabling knowledge transfer from the base to incremental tasks. Building upon this premise, we introduce the following:", + "bbox": [ + 212, + 207, + 782, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assumption 1 (Homogeneity) In FSCIL, the data pairs $\\{(\\mathbf{x}_{t,i},y_{t,i})\\}$ for each task $t$ , are i.i.d. drawn from a distribution $\\mathbb{P}_t$ over $\\mathcal{X}^t\\times \\mathcal{Y}^t$ :", + "bbox": [ + 214, + 258, + 787, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} _ {t} (\\mathbf {x}, y) = \\mathbb {P} _ {\\mathbf {x}} (\\mathbf {x}) \\cdot \\mathbb {P} _ {y | \\mathbf {x}} \\left(y \\mid f _ {t} ^ {\\star} \\circ \\mathbf {h} ^ {\\star} (\\mathbf {x})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 300, + 784, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{h}:\\mathbb{R}^d\\to \\mathbb{R}^r$ denotes a shared feature representation, and $f_{t}:\\mathbb{R}^{r}\\rightarrow \\mathbb{R}^{c}$ represents a task-specific mapping. Crucially, this assumption entails that all tasks are subject to an identical marginal distribution $\\mathbb{P}_{\\mathbf{x}}$ , indicative of a unified dataset shared across the tasks.", + "bbox": [ + 214, + 325, + 787, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Approximation Risk of FSCIL Models", + "text_level": 1, + "bbox": [ + 214, + 407, + 620, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As explicated in existing work [23,32], the training procedure for FSCIL models is roughly partitioned into two distinct phases: the base and incremental phases.", + "bbox": [ + 212, + 436, + 782, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base Phase. During the base phase, the primary objective is the learning of an optimal shared representation $\\mathbf{h}$ in conjunction with the base classifier $f_{0}$ . This is accomplished by minimizing the empirical risk $\\hat{R}_{\\mathrm{base}}$ :", + "bbox": [ + 212, + 469, + 782, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {R} _ {\\text {b a s e}} \\left(f _ {0}, \\mathbf {h}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell \\left(f _ {0} \\circ \\mathbf {h} \\left(\\mathbf {x} _ {0, i}\\right), y _ {0, i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 523, + 784, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell(\\cdot, \\cdot)$ symbolizes the loss function; $\\mathbf{x}_{0,i} \\in \\mathcal{X}_{\\mathrm{train}}^0$ and $y_{0,i} \\in \\mathcal{Y}^0$ . The composition operator $\\circ$ denotes the consecutive application of functions such that $f_0 \\circ \\mathbf{h}(\\mathbf{x}) = f_0(\\mathbf{h}(\\mathbf{x}))$ . If Assumption 1 holds and the dataset size $N$ is sufficiently large ( $N \\gg 1$ ), the estimator $(\\hat{f}_0, \\hat{\\mathbf{h}})$ , derived from the minimization of $\\hat{R}_{\\mathrm{base}}$ , serves as an effective proxy for the optimal function pair $(f_0^\\star, h^\\star)$ . This base phase is crucial in forming a robust baseline for the learning agent, enabling a seamless transition and effective integration with the novel information presented in subsequent incremental learning tasks.", + "bbox": [ + 212, + 571, + 787, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Incremental Phase. The incremental training process continues with a sequence of incremental tasks, where the learning agent learns the task-specific classifier $f_{t}$ by minimizing empirical risk $\\hat{R}_{\\mathrm{novel}}$ over $M$ new training samples3:", + "bbox": [ + 212, + 695, + 787, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {R} _ {\\text {n o v e l}} \\left(\\left[ f _ {0}, f _ {t} \\right], \\mathbf {h}\\right) = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\ell \\left(\\left[ f _ {0}, f _ {t} \\right] \\circ \\mathbf {h} \\left(\\mathbf {x} _ {t, m}\\right), y _ {t, m}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 750, + 784, + 790 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 475, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3 For brevity in our theoretical exposition, we confine this discussion to a single incremental task; nevertheless, the principles delineated here are also applicable to sequential tasks.", + "bbox": [ + 217, + 796, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[f_0, f_t]$ signifies the concatenation of the base and incremental classifiers; $\\mathbf{x}_{t,m} \\in \\mathcal{X}_{\\mathrm{train}}^t$ and $y_{t,m} \\in \\mathcal{Y}^t$ . The challenge herein lies in the fact that for $M \\ll N$ , the estimator $([\\hat{f}_0, \\hat{f}_t], \\hat{\\mathbf{h}})$ , obtained from minimizing $\\mathbf{h} \\in \\mathcal{H}$ fails to estimate their underlying counterparts $([f_0^\\star, f_t^\\star], h^\\star)$ . To address this limitation, we assess the efficacy of the estimator $([\\hat{f}_0, \\hat{f}_t], \\hat{\\mathbf{h}})$ by its excess risk on the incremental task $t$ , defined as the approximation risk of FSCIL:", + "bbox": [ + 215, + 146, + 787, + 239 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\mathrm {A R}} = R _ {\\text {n o v e l}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(\\left[ f _ {0} ^ {\\star}, f _ {t} ^ {\\star} \\right], \\mathbf {h} ^ {\\star}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 251, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Approximation Risk. To clarify the approximation risk, we deconstruct $\\varDelta_{\\mathrm{AR}}=\\varDelta_{\\mathrm{transfer}}+\\varDelta R_{\\mathrm{novel}}^{4}$ . These components are defined as:", + "bbox": [ + 215, + 280, + 785, + 311 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\text {t r a n s f e r}} = R _ {\\text {n o v e l}} \\left(\\hat {f} _ {t}, \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(f _ {t} ^ {\\star}, \\mathbf {h} ^ {\\star}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 321, + 784, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta R _ {\\text {n o v e l}} = R _ {\\text {n o v e l}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(\\hat {f} _ {t}, \\hat {\\mathbf {h}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 343, + 784, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Within this framework, $\\Delta_{\\mathrm{AR}}$ is deconstructed into two principal components: $\\Delta_{\\mathrm{transfer}}$ and $\\Delta R_{\\mathrm{novel}}$ . The former signifies the transfer risk, assessing the expected prediction risk when transferring knowledge to novel tasks. The latter encapsulates the adaptation risk, which particularly measures the divergence between the composite classifier $[f_0, f_t]$ and the task-specific classifier $f_t$ .", + "bbox": [ + 215, + 372, + 785, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Beyond $\\varDelta_{\\mathrm{AR}}$ , FSCIL additionally necessitates the consistent performance across incremental tasks. This is formally represented by a constraint encapsulating the adaptation risk pertaining to the base classifier:", + "bbox": [ + 215, + 448, + 785, + 492 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta R _ {\\text {b a s e}} = R _ {\\text {b a s e}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {b a s e}} \\left(\\hat {f} _ {0}, \\hat {\\mathbf {h}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 503, + 784, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In synthesis, the final version of the approximation risk is represented by Eq. (5), incorporating the constraint given by Eq. (8), and is achieved by applying the penalty method:", + "bbox": [ + 215, + 532, + 785, + 578 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {f _ {0}, f _ {t}, \\mathbf {h}} \\underbrace {\\Delta_ {\\text {t r a n s f e r}}} _ {\\text {T r a n s f e r R i s k}} + \\underbrace {\\Delta R _ {\\text {n o v e l}} + \\gamma \\Delta R _ {\\text {b a s e}}} _ {\\text {C o n s i s t e n c y R i s k}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 592, + 784, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ represents the penalty coefficient.", + "bbox": [ + 215, + 635, + 519, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 Theoretical Results And Applications", + "text_level": 1, + "bbox": [ + 215, + 672, + 616, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, the basic statistical principles underlying FSCIL are presented. Initially, we make the following standard, mild regularity assumptions on the loss function $\\ell(\\cdot, \\cdot)$ , the function class of tasks $\\mathcal{F}$ , and the function class of shared representations $\\mathcal{H}$ .", + "bbox": [ + 212, + 704, + 785, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assumption 2 (Regularity conditions) The following conditions hold:", + "bbox": [ + 215, + 775, + 754, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- The loss function $\\ell(\\cdot, \\cdot)$ is $B$ -bounded, and $\\ell(\\cdot, y)$ is $\\Gamma$ -Lipschitz for all $y \\in \\mathcal{Y}$ .", + "bbox": [ + 222, + 799, + 785, + 816 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "Proof In Appendix 9.4", + "bbox": [ + 218, + 824, + 385, + 839 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34a1ef1ed210b78c27434966106ce0b4cd6c4b8b6ccaae157a959e4cd320b25c.jpg", + "image_caption": [ + "Fig.2: Architecture-Agnostic Modifications to FSCIL Models. Specific adjustments are indicated in red." + ], + "image_footnote": [], + "bbox": [ + 333, + 143, + 671, + 250 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The function $f$ is $\\Gamma(\\mathcal{F})$ -Lipschitz regarding the $\\ell_2$ distance for any $f \\in \\mathcal{F}$ .", + "- The composed function $f \\circ \\mathbf{h}$ is bounded: $\\sup_{\\mathbf{x} \\in \\mathcal{X}} |f \\circ \\mathbf{h}(\\mathbf{x})| \\leq D_{\\mathcal{X}}$ , for any $f \\in \\mathcal{F}, \\mathbf{h} \\in \\mathcal{H}$ ." + ], + "bbox": [ + 223, + 320, + 784, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, the realizability assumption is adopted, positing that the true task functions and the true representation are contained in $\\mathcal{F}$ and $\\mathcal{H}$ , respectively.", + "bbox": [ + 212, + 376, + 782, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Assumption 3 (Realizability) The true representation $\\mathbf{h}^{\\star}$ is encompassed by $\\mathcal{H}$ . Concurrently, for all tasks considered—both base and incremental—the true task-specific functions $f_{t}^{\\star}$ are also contained in $\\mathcal{F}$ .", + "bbox": [ + 214, + 419, + 784, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on Regularity and Realizability assumptions, we analyze transfer learning risk $\\varDelta_{\\mathrm{transfer}}$ and consistency risk $\\varDelta R_{\\mathrm{novel}}+\\gamma\\varDelta R_{\\mathrm{base}}$ in Eq. (9) separately.", + "bbox": [ + 212, + 477, + 782, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 Transfer Risk", + "text_level": 1, + "bbox": [ + 214, + 531, + 374, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 1. (Proof in Appendix 9.2) If Assumptions 2 and 3 hold, and the base task is $(\\rho,0)$ diverse, with probability $1 - 2\\delta$ , the transfer risk $\\Delta_{\\text{transfer}}$ in Eq. (9) is bounded by:", + "bbox": [ + 214, + 555, + 782, + 599 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {O} \\Big (\\frac {\\Gamma}{\\rho} \\cdot \\Gamma (\\mathcal {F}) \\cdot 4 D \\prod_ {l = 1} ^ {L} Z (l) \\sqrt {\\frac {\\log 2 J}{N}} + \\frac {\\Gamma D _ {\\mathcal {X} ^ {0}}}{\\rho N ^ {2}} + B \\Big [ \\frac {1}{\\rho} \\cdot \\sqrt {\\frac {\\log (2 / \\delta)}{N}} + \\sqrt {\\frac {\\log (2 / \\delta)}{M}} \\Big ] \\Big),\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 608, + 769, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N$ and $M$ quantify the number of training samples of the base and incremental task $t$ respectively; $\\| \\mathbf{x} \\| \\leq D$ ; $L$ is the depth of the neural network; $Z(l) = \\frac{\\| h^{(l)} \\|_{\\infty}}{\\| h^{(l-1)} \\|_{\\infty}}$ , where $h^{(l)}$ denotes the feature representations of the $l$ -th layer.", + "bbox": [ + 214, + 654, + 782, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Application. To bridge the basic statistical principles and the practical guidelines, we undertake the subsequent analysis:", + "bbox": [ + 212, + 717, + 782, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(1) From Theorem 1, it becomes apparent that the quantity $N$ , representing the number of training samples for the base task, primarily influences the upper bound of the transfer risk. As $N$ increases, the upper bound tightens at a rate proportional to $O\\left(\\frac{1}{\\sqrt{N}}\\right)$ . This relation shows that expanding training datasets for base classes is the most straightforward and potent strategy for addressing the FSCIL challenge, especially when the initial dataset is insufficiently large.", + "bbox": [ + 212, + 748, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 475, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In scenarios where expanding the data pool is not feasible, the adoption of a foundation model [19,27] becomes a strategic alternative, which is substantiated by our analysis of the transfer risk.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(2) Theoretically, the upper bound of transfer risk is proportional to $\\prod_{l=1}^{L} Z(l)$ , where $Z(l)$ is the ratio of the infinity norm of features at layer $l$ to that at layer $l-1$ . While reducing $\\prod_{l=1}^{L} Z(l)$ could lower transfer risk, it might impair the feature extraction, so we do not adopt this modification. In original ResNets and ViTs, each layer typically includes Layer or Batch Normalization, standardizing feature means to 0 and variances to 1. However, this normalization does not limit the infinity norm of features, potentially causing excessively large values in some dimensions and increasing $Z(l)$ , then raising transfer risk. To mitigate this, we constrain each layer's features to $[-1,1]$ , setting $Z(l) = 1$ . Since normalization centers most feature values around zero, we introduce the tanh function to cap them. The tanh function bounds values within $[-1,1]$ and has a slope of 1 at zero, effectively constraining extreme values without significantly impacting original models. Similarly, we also use norm + tanh to constrain the representation $\\mathbf{h}$ .", + "bbox": [ + 212, + 191, + 789, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, $D_{\\mathcal{X}^0}$ represents the highest probability value of the predicted outcome, exhibiting a positive correlation with the upper bound. The essential strategy to reduce this bound involves calibrating the model to ensure its predictions are tempered, thereby preventing overconfidence. In conclusion, the analysis of both $\\prod_{l=1}^{L} Z(l)$ and $D_{\\mathcal{X}^0}$ shows that preventing excessive focus on specific features is an effective strategy for addressing FSCIL challenge. As shown in Fig. 2, we propose three architecture-agnostic modifications:", + "bbox": [ + 212, + 390, + 784, + 494 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Incorporation of the tanh activation function subsequent to each normalization layer in the original models;", + "- Integration of the layer normalization and tanh activation function prior to the generation of shared representations;", + "- Application of the label smooth in the training process." + ], + "bbox": [ + 225, + 501, + 782, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark. The term $Z(l) = \\frac{\\|h^{(l)}\\|_{\\infty}}{\\|h^{(l - 1)}\\|_{\\infty}}$ quantifies the ratio of maximal activation value in the $l$ -th layer compared to that in the $(l - 1)$ -th layer. It serves as an indicator of the proportional activation dynamics spanning all dimensions between consecutive layers. This metric emphasizes the need for a balanced distribution of activation values throughout the network layers to inhibit the incidence of impulse activation. Such regulation ensures that the model maintains an equitable emphasis on all features, thereby supporting the underlying expectation of $D_{\\chi^0}$ for the model to prevent overconfidence in its predictions.", + "bbox": [ + 212, + 580, + 784, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Corollary 1 (Proof in Appendix 9.3). Under conditions from Theorem 1, Vision Transformers (ViTs) exhibit lower transfer risks than Convolutional Neural Networks (CNNs).", + "bbox": [ + 214, + 713, + 785, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark. This finding substantiates the efficacy of ViTs in leveraging shared representations for novel tasks, underscoring their robustness in transfer learning scenarios. Furthermore, the reduced transfer risk with ViTs as compared to CNNs suggests a pivotal architectural advantage that may inform future developments in neural network design.", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Consistency Risk", + "text_level": 1, + "bbox": [ + 215, + 146, + 403, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Theorem 2 (Proof in Appendix 9.5). If Assumptions 2 and 3 hold, and given $K_{0}$ and $K_{t}$ classes in the base and incremental tasks respectively, the consistency risk $\\Delta R_{\\text{novel}} + \\gamma \\Delta R_{\\text{base}}$ in Eq. (9) is upper-bounded by:", + "bbox": [ + 215, + 171, + 787, + 217 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\log \\left(\\left(1 + \\frac {K _ {0}}{K _ {t}} v _ {t}\\right) \\left(1 + \\frac {K _ {t}}{K _ {0}} v _ {0}\\right) ^ {\\gamma}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 228, + 784, + 260 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $v_{t}$ and $v_{0}$ involve supremum and infimum of logits:", + "bbox": [ + 215, + 268, + 629, + 285 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} = \\exp \\left(\\sup \\left(\\phi_ {t \\rightarrow 0}\\right) - \\inf \\left(\\phi_ {t \\rightarrow t}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 297, + 784, + 311 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nv _ {0} = \\exp \\left(\\sup \\left(\\phi_ {0 \\rightarrow t}\\right) - \\inf \\left(\\phi_ {0 \\rightarrow 0}\\right)\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 316, + 784, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The classifier logits $\\phi_{t\\to 0}$ , $\\phi_{t\\to t}$ , $\\phi_{0\\to 0}$ , and $\\phi_{0\\to t}$ respectively quantify how samples from task $t$ and the task 0 (base task) are classified, either correctly or in-correctly.", + "bbox": [ + 215, + 342, + 787, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Application. Building on the above analysis about the transfer risk, we proceed with the following detailed analysis about the consistency risk:", + "bbox": [ + 212, + 398, + 785, + 428 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(3) From Theorem 2, we observe that the consistency risk is inherently tied to the classification margin discrepancy of the model, which is quantified by the difference between the infimum of the classifier logits for correct classification and the supremum for incorrect classification. It infers that the consistency risk approaches nullity as the predictive accuracy of the model approaches perfection.", + "(4) The upper bound of the consistency risk represents a balance between $v_{t}$ and $v_{0}$ . Introducing a bias in categorization leads to inverse variations in $v_{t}$ and $v_{0}$ , indicating that ensuring unbiased classification across both base and novel classes can completely nullify the consistency risk. This implies strategically designed biases might reach the minimal risk bound but are not optimally efficient. Key lies in improving classification margin discrepancy while maintaining a balanced classification bias." + ], + "bbox": [ + 212, + 429, + 785, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.1 Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 664, + 568, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Our study employs three diverse datasets: mini-ImageNet [20], CIFAR100 [12], and CUB200 [29]. The mini-ImageNet, a subset of ImageNet [20], includes 600 images from 100 classes, split into 60 base and 40 novel classes. These novel classes are divided into 8 incremental tasks, each a 5-way 5-shot task. CIFAR100 comprises 100 classes with 600 images each, partitioned into 60 base and 40 novel classes, the latter organized into 8 incremental 5-way 5-shot tasks. CUB200, containing 11,788 images of 200 bird species, is split into 100 base and 100 novel classes, with the latter arranged into 10 incremental 10-way 5-shot tasks. In this work, when the model is trained from scratch, the image sizes for mini-ImageNet, CIFAR100, and CUB200 are set at $84 \\times 84$ , $32 \\times 32$ ,", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f9642e2ab951a4ac4525b9bf6ca88216180644ed3a91123186dc9781524fe032.jpg", + "table_caption": [ + "Table 1: Classification Accuracy (\\%) on CUB200 in the 10-Way 5-Shot FS-CIL Setting. The notation ViT-B/16 specifically denotes the ViT-Base model [5] with a designated patch size of 16. The symbols $\\diamond$ and $\\dagger$ signify models derived from the CLIP [19] and those pre-trained on the ImageNet-1K dataset respectively. The symbol $\\downarrow$ is employed to indicate that lower values are more desirable." + ], + "table_footnote": [], + "table_body": "
MethodBackboneAccuracy in each session(%)DR↓(%)
012345678910
\\( CEC^† \\) [32]ResNet1875.8571.9468.5063.562.4358.2757.7355.8154.8353.5252.2831.07
\\( MCNet^† \\) [10]ResNet1877.5773.9670.4765.8166.1663.8162.0961.8260.4160.0959.0823.84
\\( FACT^† \\) [38]ResNet1875.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9424.98
\\( ALICE^† \\) [17]ResNet1877.472.770.667.265.963.462.961.960.560.660.122.35
\\( CABD^† \\) [37]ResNet1879.1275.6373.2169.9368.3266.3065.1564.9664.2064.0363.8119.35
\\( NC-FSCIL^† \\) [30]ResNet1880.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4426.12
\\( SAVC^† \\) [22]ResNet1881.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5023.64
\\( LDC^† \\) [15]ResNet1877.8976.9374.6470.0668.8867.1564.8364.1663.0362.3961.5820.94
\\( CoCoOp^° \\) [39]ViT-B/1680.377.175.873.972.468.265.263.961.160.357.228.76
\\( Prompt^° \\) [31]ViT-B/1684.383.280.979.277.772.472.169.968.867.466.820.76
\\(ours^{\\dagger}\\)ResNet1879.8777.3674.9270.6270.6468.5568.1467.9366.7366.6765.6917.75
\\(ours^{\\circ}\\)ViT-B/1686.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
\\(ours^{\\dagger}\\)ViT-B/1686.6985.3884.3482.4383.0881.0781.3881.4281.2181.1181.306.22
", + "bbox": [ + 246, + 226, + 759, + 378 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and $224 \\times 224$ respectively. In scenarios utilizing a foundation model, the images from all three datasets are resized to $224 \\times 224$ .", + "bbox": [ + 212, + 405, + 785, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Following previous work [23, 34], and guided by our theory, we select ViTs as our primary models across all datasets. Specifically, we utilize two ViT-B/16 variants: one from CLIP [19] and another pre-trained on ImageNet-1K. During the base phase, we train the ViT models on the base classes for 50 epochs using the AdamW optimizer. The learning rate is initially set to 0, then linearly rises to 0.00001 during the first 20 epochs, followed by a linear decrease back to 0 for the rest of the epochs. For a fair comparison, ResNet18 serves as an auxiliary model across datasets, trained on base classes for 500 epochs using AdamW. The learning rate for ResNet18 follows a similar pattern: beginning at 0, increasing linearly to 0.1 within the first 50 epochs, and subsequently diminishing back to 0. In the incremental phase, we freeze the backbone and update only the task-specific classifiers.", + "bbox": [ + 212, + 435, + 787, + 616 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following precedents [16, 23], our evaluation employs Top-1 classification accuracy across cumulative testing sets after each learning session. While traditional models often employ the Performance Drop (PD) metric, defined as $\\mathrm{PD} = (\\mathcal{A}_0 - \\mathcal{A}_N)$ , where $\\mathcal{A}_0$ and $\\mathcal{A}_N$ represent the accuracies at the initial and final sessions, we propose a refined metric. Recognizing that model forgetting encompasses both the volume of initially acquired knowledge and the extent of subsequent knowledge loss, we introduce the Performance Drop Rate (DR), calculated as $\\mathrm{DR} = (\\mathcal{A}_0 - \\mathcal{A}_N) / \\mathcal{A}_0$ . This metric more accurately captures the effect of knowledge loss, factoring in the initial knowledge base.", + "bbox": [ + 212, + 617, + 787, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.2 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 212, + 771, + 620, + 787 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We benchmark our model against state-of-the-art performances on three datasets: CUB200, CIFAR100, and mini-ImageNet, with results in Fig. 3 and detailed CUB200 data in Table 1 (see supplementary materials for other datasets).", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/86d59389063e912ebcfd5ed177ed4836ec583f921c92379c5e16dbb05206f5d4.jpg", + "image_caption": [ + "Fig. 3: Comparison with State-of-the-Art on Three Datasets: CUB200, mini-ImageNet, and CIFAR100. For detailed performance metrics, please see Table 1 and refer to our supplementary material." + ], + "image_footnote": [ + "--- CEC --- FACT --- CABD --- SAVC --- CoCoOp : ours (RosNet18) : ours (ImageNet) \n--- MCNet --- ALICE --- NC-FSCIL --- LDC --- Prompt : ours (CLIP) -- C-FSCL" + ], + "bbox": [ + 233, + 145, + 411, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2ccefc01fe4ebb75c0711cafaee7ba4741ba3177f5cea85e0dbfdb0341e3ac06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 146, + 589, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b926463328533480f934a14a73b98a04ada6a4075dfa43ee69656dbea10e72bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 146, + 769, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1 shows that our ViT model (CLIP) achieves an improvement of $8.42\\%$ in the final session compared to the sub-optimal method, and the DR metric decreases by $7.76\\%$ . When ImageNet-1K is used for pre-training, our ViT model records a boost of $17.49\\%$ in the last session compared to the second best method, with a DR reduction of $13.13\\%$ . As for ResNet18, our ResNet18 model exhibits a $1.88\\%$ improvement in the final session over the sub-optimal ResNet18 model, with a $1.6\\%$ reduction in the DR metric. These results indicate a decrease in catastrophic forgetting and overfitting, attributed to the application of our guiding principles for reducing the approximation risk.", + "bbox": [ + 212, + 342, + 787, + 478 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "From Fig. 3, similar trends are observed on the mini-ImageNet and CIFAR100, though the improvements of ViTs are less pronounced. This observation can be attributed to mini-ImageNet being a subset of ImageNet, which violates the standard FSCIL setting, and to the unique image characteristics of CIFAR100, such as smaller image size and lack of background, which differ from the typical natural image distributions used in pre-training.", + "bbox": [ + 212, + 479, + 787, + 569 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Remarkably, even though CLIP has a broader pre-training dataset compared to ImageNet-1K, performance with CLIP remains consistently lower. This phenomenon might be explained by the significant difference between the marginal distributions of the three benchmark datasets and the data used for pre-training of CLIP. Conversely, ImageNet-1K aligns more closely with these distributions, better fulfilling the Homogeneity assumption in our study.", + "bbox": [ + 212, + 569, + 787, + 660 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 680, + 390, + 696 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we assess the impact of four key components of our ViT model on the CUB200 dataset (see supplementary materials for ResNet18). Detailed in Table 2, these components show varied performance effects. The FT strategy exhibits the lowest Top-1 accuracy (38.01% in the final session) and highest DR metric (40.99%), indicating a pronounced bias towards novel classes and substantial consistency risk. Conversely, FR strategy results in improved Top-1 accuracy (52.41%) and lower DR metric (36.63%), yet struggles with catastrophic forgetting. The DS strategy, separating training of base and novel classifiers, is the most effective in reducing bias and consistency risk.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 475, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d6c213b29fa5a0c12d1b2f93b8aeda6613b8f13525e52b3cb2d5ad61e6f0c686.jpg", + "table_caption": [ + "Table 2: Ablation Study on CUB200 in the 10-Way 5-Shot FSCIL Setting. With ViT-B/16 as the backbone, our study encompasses: \"Stabilize\", enhancing output stability via tanh function after normalization; \"Constrain\", limiting representation values using layer normalization and tanh function; \"LS\" (Label Smoothing) to reduce model overconfidence; and \"BC\" (Base Classifier), employing Fine-Tuning (FT), Freezing (FR), and Discarding (DS) in the incremental phase." + ], + "table_footnote": [], + "table_body": "
StabilizeConstrainLSBCAccuracy in each session(%)DR↓(%)
012345678910
FT64.4259.3255.0450.6248.9145.2243.9742.1340.5939.5038.0140.99
FR82.7176.3671.3366.9564.8861.0358.9457.4855.7554.1252.4136.63
DS87.0583.5081.0077.6676.5674.1273.0073.1772.3271.8770.9718.47
✓DS87.1584.4181.6578.1177.4675.1874.4473.8773.1772.9772.0717.30
✓DS86.7784.0281.5678.3177.6375.4474.3974.4873.4073.1172.2916.69
✓DS87.4984.0982.1978.5677.9375.4975.5374.6773.4673.4272.7716.82
✓✓DS87.5184.0381.9278.4378.2275.7475.9175.1973.5973.8373.3416.19
✓✓DS87.6384.1381.7778.0578.0075.8075.4074.7974.4374.2573.7115.88
✓✓DS87.4684.5382.7779.3278.9676.6276.8575.8774.5474.4773.9815.41
✓✓✓DS86.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
", + "bbox": [ + 245, + 241, + 756, + 363 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d8ec9f2a782e1195ba9107f18dbbe8e0aa0cc9f7a2514f7b533dac72e6608229.jpg", + "image_caption": [ + "Fig. 4: Impact of Dataset Size in Transfer Risk with ResNet18 and ViT-S/6 Backbones Trained from Scratch on mini-ImageNet. In both subfigures, the horizontal axis represents the dataset size per base class, while the vertical axis lists training techniques. These include Data Augmentation [1,7,11] and Mixup [18,33], essential for expanding the training dataset." + ], + "image_footnote": [], + "bbox": [ + 254, + 376, + 750, + 506 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Integration of the tanh function for layer stabilization improves the DR metric by about $1\\%$ , mitigating overfitting. Similarly, constrained feature representations through layer normalization and tanh function slightly decrease base class accuracy but enhance the DR metric by $1\\%$ , indicating reduced overfitting. Implementing label smoothing technique raises both base classification accuracy and DR metric by $1\\%$ , attributed to lowering transfer risk.", + "bbox": [ + 212, + 614, + 787, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Combining these components yields superior performance, highlighting their synergistic effect in FSCIL.", + "bbox": [ + 212, + 705, + 787, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.4 Dataset Size in Transfer Risk", + "text_level": 1, + "bbox": [ + 214, + 755, + 504, + 770 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we empirically investigate the correlation between dataset size and transfer risk in neural networks, focusing on ResNet18 and ViT-S/6 architectures. Fig. 4 (a) and (b), show a clear correlation between increased dataset size per class and improved final accuracy for both models. This supports the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/82d4361764532fca01f019eeacd3aa538b826867a95c02a5b0305909f46722a9.jpg", + "image_caption": [ + "Fig. 5: Visualization of ResNet18 and ViT-S/6 on mini-ImageNet. Images are divided by a red dotted line: left for base classes and right for novel classes. Beneath each image, the predicted probability is numerically displayed. The star symbol denotes models in our implementation. In ViT visualizations, image brightness reflects attention level. Darker images signify attention concentrated on specific patches, while brighter ones indicate a more global attention on the whole image." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 784, + 364 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "idea that increasing the dataset size $(N)$ effectively tightens the upper bound of transfer risk, thereby enhancing models' capacity for incremental learning.", + "bbox": [ + 212, + 489, + 787, + 520 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "When direct sample collection is limited, Data Augmentation and Mixup techniques present a viable alternative. These methods create additional samples, enriching the training dataset. This is evident in Figures 4 (a) and (b), where both ResNet18 and ViT-S models show significant performance gains with these techniques. This reinforces our theory: indirect sample generation via Data Augmentation and Mixup can enhance model transferability, mirroring the advantages of direct sample collection.", + "bbox": [ + 212, + 520, + 787, + 626 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.5 Analysis of the Model Overconfidence", + "text_level": 1, + "bbox": [ + 214, + 648, + 576, + 664 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we provide a visual results to illustrate the influence of prediction confidence on transfer risk. The first and third rows of Fig. 5 indicate that models (CNNs or ViTs) naturally focus on small, highly discriminative regions within base class samples, which leads to overconfidence in these predictions. This propensity leads to misclassification during the transfer of models to novel classes, a phenomenon commonly referred to as supervision collapse.", + "bbox": [ + 212, + 672, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our approach aims to temper the model's over-reliance on base class predictions. As evidenced in the second and fourth rows of Fig. 5, lowering the probability for base class predictions enables the model to cover broader regions. This expanded scope assists in focusing on distinctive features of novel classes, thus enhancing prediction accuracy for these classes.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 477, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0bcff4c875352c7c80404e9e1894a3291af13784ebe30b895cf89986a6b73c0a.jpg", + "image_caption": [ + "(a) -30.84" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 362, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9eaf6be963d74c27aae2da552a86958a8f7bf1ff25fbf47e372a286d263cf395.jpg", + "image_caption": [ + "(b) -15.78", + "Fig. 6: The t-SNE visualizations of the logits for test samples on mini-ImageNet, under varying degrees of Classification Margin Discrepancy, are depicted with distinct colors representing different classes. The architectures for Fig. (a), (b), and (c) are based on ViT-S/6 models, each trained from scratch. Conversely, Fig. (d) employs a ViT-B/16 backbone, which is pre-trained on the ImageNet dataset." + ], + "image_footnote": [], + "bbox": [ + 364, + 143, + 500, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ac3bf5461c70696b846bc870d8af86972fd1b34d4d7e8b391a6bf796d6eaa651.jpg", + "image_caption": [ + "(c)-10.18" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 643, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/aff127abaeaf51443acab0cbe32b056a586180222167a6b33995d30ea6c13f40.jpg", + "image_caption": [ + "(d)-3.5" + ], + "image_footnote": [], + "bbox": [ + 645, + 143, + 784, + 215 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.6 Classification Margin Discrepancy", + "text_level": 1, + "bbox": [ + 214, + 343, + 542, + 358 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As demonstrated in Fig. 6, we observe a progressive enhancement in the separability of logits for test samples, correlating with an incremental increase in the Classification Margin Discrepancy. This trend not only signifies a reduction in the consistency risk but also marks a significant step towards achieving more reliable and robust classification performance in FSCIL settings. These findings empirically validate our theoretical framework, suggesting that optimizing for classification margin discrepancy directly contributes to mitigating consistency risks and enhancing model adaptability across incremental learning tasks.", + "bbox": [ + 212, + 368, + 787, + 489 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 511, + 359, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, we have conducted a thorough exploration on the approximation risk of FSCIL. By focusing on mitigating transfer and consistency risks and tightening their upper bounds, we have developed practical guidelines for the design and training of FSCIL models. These guidelines have expanded training datasets for base classes, optimized classification margin discrepancy, prevented excessive focus on specific features, and ensured unbiased classification across both base and novel classes. Our extensive experiments have validated these principles, culminating in state-of-the-art performance on three benchmark datasets.", + "bbox": [ + 212, + 542, + 787, + 664 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Limitation", + "text_level": 1, + "bbox": [ + 214, + 686, + 356, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, Homogeneity and Regularity conditions are foundational assumptions. Homogeneity, based on the i.i.d assumption, indicates samples in a dataset share similar characteristics like hues and lighting. Regularity involves boundedness for model convergence and the Lipschitz condition, crucial for representation learning. Practically, comparative experiments show that even when the pretraining and FSCIL benchmark datasets have different marginal distributions, our methods still achieve better results, albeit with smaller gains. This indicates that our theory generally holds but comes at the cost of some practicality.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160403, by the National Natural Science Foundation of China (NSFC) under Grants 62176178, 62106152, and 62303447, by the Postdoctoral Innovation Talents Support Program under No. BX20230399, and by the China Postdoctoral Science Foundation under No. 2023M743702.", + "bbox": [ + 212, + 180, + 787, + 270 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 297, + 321, + 313 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Balestriero, R., Bottou, L., LeCun, Y.: The effects of regularization and data augmentation are class dependent. In: Advances in Neural Information Processing Systems. vol. 35, pp. 37878-37891 (2022)", + "2. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023), early Access", + "3. Doersch, C., Gupta, A., Zisserman, A.: Crosstransformers: spatially-aware few-shot transfer. In: Advances in Neural Information Processing Systems. pp. 21981-21993 (2020)", + "4. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021)", + "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)", + "6. Golowich, N., Rakhlin, A., Shamir, O.: Size-independent sample complexity of neural networks. In: Conference On Learning Theory. pp. 297-299. PMLR (2018)", + "7. Hao, X., Zhu, Y., Appalaraju, S., Zhang, A., Zhang, W., Li, B., Li, M.: Mixgen: A new multi-modal data augmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 379-389 (2023)", + "8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)", + "9. Hiller, M., Ma, R., Harandi, M., Drummond, T.: Rethinking generalization in few-shot classification. In: Advances in Neural Information Processing Systems. pp. 1-13 (2022)", + "10. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing 32, 937-948 (2023)", + "11. Ji, Z., Jiao, Z., Wang, Q., Pang, Y., Han, J.: Imbalance mitigation for continual learning via knowledge decoupling and dual enhanced contrastive learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-14 (2024), early Access", + "12. Krizhevsky, A., Nair, V., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009)" + ], + "bbox": [ + 217, + 332, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Kukleva, A., Kuehne, H., Schiele, B.: Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9029 (2021)", + "14. Ledoux, M., Talagrand, M.: Probability in Banach Spaces: isoperimetry and processes, vol. 23. Springer Science & Business Media (1991)", + "15. Liu, B., Yang, B., Xie, L., Wang, R., Tian, Q., Ye, Q.: Learnable distribution calibration for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-8 (2023)", + "16. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2337-2345 (2021)", + "17. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022)", + "18. Pinto, F., Yang, H., Lim, S.N., Torr, P., Dokania, P.: Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In: Advances in Neural Information Processing Systems. vol. 35, pp. 14608-14622 (2022)", + "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "20. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015)", + "21. Shalev-Shwartz, S., Ben-David, S.: Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press (2014)", + "22. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24183-24192 (June 2023)", + "23. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020)", + "24. Tripuraneni, N., Jordan, M., Jin, C.: On the theory of transfer learning: The importance of task diversity. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7852-7862 (2020)", + "25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30, pp. 1-11 (2017)", + "26. Wainwright, M.J.: High-dimensional statistics: A non-asymptotic viewpoint, vol. 48. Cambridge university press (2019)", + "27. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023)", + "28. Wang, X., Ji, Z., Pang, Y., Yu, Y.: A cognition-driven framework for few-shot class-incremental learning. Neurocomputing 600, 128118 (2024)", + "29. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations. pp. 1-13 (2023)", + "31. Yoon, I.U., Choi, T.M., Lee, S.K., Kim, Y.M., Kim, J.H.: Image-object-specific prompt learning for few-shot class-incremental learning. arXiv preprint arXiv:2309.02833 (2023)", + "32. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12455-12464 (2021)", + "33. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)", + "34. Zhang, J., Liu, L., Silven, O., Pietikainen, M., Hu, D.: Few-shot class-incremental learning: A survey. arXiv preprint arXiv:2308.06764 (2023)", + "35. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-11 (2021)", + "36. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (2023)", + "37. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (June 2023)", + "38. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022)", + "39. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 563 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On the Approximation Risk of FSCIL", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_model.json b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_model.json new file mode 100644 index 0000000000000000000000000000000000000000..afc719321c8fd03ed67a7379f963535722202bc6 --- /dev/null +++ b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_model.json @@ -0,0 +1,2566 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.271, + 0.141, + 0.734, + 0.187 + ], + "angle": 0, + "content": "On the Approximation Risk of Few-Shot Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.212, + 0.782, + 0.243 + ], + "angle": 0, + "content": "Xuan Wang\\(^{1,2}\\), Zhong Ji\\(^{1,2}\\), Xiyao Liu\\(^{3}\\), Yanwei Pang\\(^{1,2}\\), and Jungong Han\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.255, + 0.748, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\) Tianjin University \\(^{2}\\) Shanghai Artificial Intelligence Laboratory \\(^{3}\\) State Key Laboratory of Robotics, SIA, CAS \\(^{4}\\) University of Sheffield" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.317, + 0.744, + 0.526 + ], + "angle": 0, + "content": "Abstract. Few-Shot Class-Incremental Learning (FSCIL) aims to learn new concepts with few training samples while preserving previously acquired knowledge. Although promising performance has been achieved, there remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Therefore, we thoroughly explore the approximation risk of FSCIL, encompassing both transfer and consistency risks. By tightening the upper bounds of these risks, we derive practical guidelines for designing and training FSCIL models. These guidelines include (1) expanding training datasets for base classes, (2) preventing excessive focus on specific features, (3) optimizing classification margin discrepancy, and (4) ensuring unbiased classification across both base and novel classes. Leveraging these insights, we conduct comprehensive experiments to validate our principles, achieving state-of-the-art performance on three FSCIL benchmark datasets. Code is available at https://github.com/xwangrs/Approximation_FSCIL-ECCV2024.git." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.538, + 0.743, + 0.568 + ], + "angle": 0, + "content": "Keywords: Few-Shot Class-Incremental Learning \\(\\cdot\\) Few-Shot Learning \\(\\cdot\\) Class Incremental Learning \\(\\cdot\\) Approximation Risk" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.377, + 0.607 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Few-Shot Class-Incremental Learning (FSCIL) [4, 23, 34], an emergent learning paradigm, aims to learn novel classes with limited training data, while preserving the knowledge acquired from base classes. This paradigm faces two primary issues: (1) the catastrophic forgetting of established base class knowledge when integrating new information, and (2) the overfitting to few-shot novel classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.818 + ], + "angle": 0, + "content": "To effectively address these issues, researchers have proposed various approaches, broadly categorized into three groups [34]: data-based [13, 17, 38], structure-based [23, 32, 36], and optimization-based [2, 4, 35] approaches. Despite their proven effectiveness, there still remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Central to this gap is the question: What are the key factors in solving FSCIL problems? To answer this question, our research delves into the basic statistical principles of FSCIL, aiming to guide the design and training of FSCIL models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.633, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author: Zhong Ji, Email: jizhong@tju.edu.cn" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.145, + 0.499, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.273, + 0.434, + 0.285 + ], + "angle": 0, + "content": "(a) Transfer Risk" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.144, + 0.744, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.273, + 0.686, + 0.285 + ], + "angle": 0, + "content": "(b) Consistency Risk" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.297, + 0.789, + 0.382 + ], + "angle": 0, + "content": "Fig. 1: The upper bounds of the transfer and consistency risks. In Fig. 1 (a), the horizontal axis represents the number of training samples for base classes, whereas the vertical axis quantifies the magnitude of the transfer risk. Similarly, in Fig. 1 (b), the horizontal axis reflects the classification margin discrepancy, while the vertical axis quantifies the magnitude of the consistency risk. The parameters \\(\\eta > 0\\) and \\(\\kappa > 0\\) serve as dynamic scaling factors, regulating the transfer and consistency risks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.789, + 0.453 + ], + "angle": 0, + "content": "We begin with fundamental assumptions: Homogeneity, Regularity conditions, and Realizability. These assumptions, grounded in practical scenarios, form the foundation of our theory." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.789, + 0.619 + ], + "angle": 0, + "content": "Based on these assumptions, we formulate the gap between empirically obtained and theoretically optimal FSCIL models as an optimization problem, whose objective function is defined as the Approximation Risk of FSCIL. This risk includes two components: transfer risk and consistency risk. The transfer risk [24] evaluates the efficacy of the shared representations in knowledge transfer from base to novel classes. A low transfer risk indicates effective utilization of base classes knowledge to compensate for the scarcity of novel class samples, thereby preventing overfitting in these novel classes. Meanwhile, the consistency risk measures performance drop when incorporating novel classes, with a low risk signifying robust and stable performance throughout class-incremental tasks, thereby preventing the catastrophic forgetting of base classes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.62, + 0.789, + 0.65 + ], + "angle": 0, + "content": "Leveraging statistical learning theory [6, 14, 21, 26], we explore the upper bounds of transfer and consistency risks, then identifying their determinants2:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.656, + 0.787, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\eta \\cdot \\frac {1}{\\sqrt {N}}} _ {\\text {T r a n s f e r R i s k}} + \\underbrace {\\log \\left(1 + \\frac {\\kappa}{\\exp (\\nu)}\\right)} _ {\\text {C o n s i s t e n c y R i s k}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.713, + 0.789, + 0.805 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of training samples for the base classes; \\(\\nu\\) indicates the classification margin discrepancy, defined as the infimum of classifier logits for correct classifications minus the supremum of those for incorrect classifications; the parameters \\(\\eta\\) and \\(\\kappa\\) are the dynamic scaling factors for transfer and consistency risks, respectively. We will further discuss and elaborate on methods for minimizing these risks by tightening their upper bounds." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.368, + 0.825 + ], + "angle": 0, + "content": "1 Proofs in Section 4." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.366, + 0.839 + ], + "angle": 0, + "content": "2 Proofs in Section 5." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.368, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "In Fig. 1 (a), it is evident that increasing the number of training samples for base classes leads to a significant reduction in the transfer risk. Theoretically, this risk could potentially be minimized to a negligible level with a sufficiently large and diverse dataset. However, in practical scenarios, gathering sufficient data for base classes remains a challenge, thereby prompting reliance on foundation models [19,31]. Crucially, for the Homogeneity assumption to hold, the marginal distribution of the pre-trained data should align with that of the base class data. Failing to meet this condition could lead to an increase in the transfer risk." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.478 + ], + "angle": 0, + "content": "In exploring how \\( N \\) affects the transfer risk, it becomes imperative to fully comprehend the role of the dynamic scaling factor \\( \\eta \\). From a theoretical perspective, \\( \\eta \\) represents a multivariate polynomial that is related to the maximum activation values across each layer of the neural networks. This could be interpreted as the model's selective attention to specific features. An increase in \\( \\eta \\) values indicates a marked emphasis on certain features, often at the expense of neglecting others, thereby increasing the transfer risk. This phenomenon is known as the \"supervision collapse\" [3, 9]. To mitigate this issue, we propose to maintain a balanced distribution of activation values across the network layers to prevent extreme impulse values. Furthermore, our theoretical analysis indicates that Vision Transformers (ViTs) [5], owing to their self-attention mechanism [25], exhibit a lower transfer risk compared to Convolutional Neural Networks (CNNs) [8]. Thus, in addressing the FSCIL problem, our preference is towards the deployment of ViT models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.785, + 0.614 + ], + "angle": 0, + "content": "Evidenced in Fig. 1 (b), a clear inverse relationship exists between the classification margin discrepancy and the consistency risk, indicating that enhancing the inter-class discrimination of representations can theoretically diminish the consistency risk to minimal levels. This requires models to maximize the classification margin discrepancy on base classes, ensuring comprehensive training of these models. Moreover, our theory also indicates that the scaling factor \\(\\kappa\\) reaches its lower bound when the model exhibits unbiased classification across both base and novel classes. Achieving this balance involves training classifiers for each task independently before combining them into a unified classifier." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.616, + 0.555, + 0.63 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.645, + 0.784, + 0.674 + ], + "angle": 0, + "content": "- Theoretical Framework: We develop a new optimization theory that provides a novel perspective for understanding and addressing challenges in FSCIL." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.72 + ], + "angle": 0, + "content": "- Risk Analysis: We decompose the approximation risk of FSCIL into transfer and consistency risks, defining their upper boundaries and and key determinants, thereby offering guidance for model design and training." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.721, + 0.784, + 0.794 + ], + "angle": 0, + "content": "- Practical Strategies: We propose practical guidelines to address the challenges in FSCIL. These include expanding training datasets for base classes, preventing excessive focus on specific features, optimizing classification margin discrepancy, and ensuring unbiased classification across both base and novel classes." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Informed by these insights, we achieve state-of-the-art performance on three benchmark datasets, demonstrating the effectiveness of our theory." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.39, + 0.163 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.254 + ], + "angle": 0, + "content": "Recently, FSCIL [23,28] has emerged as a paradigm to incrementally learn novel classes with a limited number of samples, while preserving previously learned knowledge. According to the survey [34], existing approaches can be broadly categorized into three groups: data-based approaches [13,17,38], structure-based approaches [23,32,36], and optimization-based approaches [2,4,35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.789, + 0.375 + ], + "angle": 0, + "content": "The data-based approaches address FSCIL from the perspective of training data, which involves data replay-based methods [13] and pseudo scenarios-based methods [17, 38]. For example, Kukleva et al. [13] propose a three-stage framework where the base and the novel class samples are efficiently replayed to calibrate the classification bias. Zhou et al. [38] propose ForwArd Compatible Training (FACT) for FSCIL, which preassigns multiple virtual prototypes in the embedding space by creatively mixuping the known class samples, and Peng et al. [17] have similar ideas about utilizing the Mixup [33] to expand the datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.789, + 0.51 + ], + "angle": 0, + "content": "The structure-based approaches address FSCIL from the perspective of network architecture design, which involves dynamic structure-based methods [23, 32] and attention-based methods [36]. For example, Zhang et al. [32] propose Continually Evolved Classifier (CEC) that employs a graph model to propagate context information between classifiers for adaptation. Tao et al. [23] propose to use a neural \"gas\" network to learn and preserve the topology of the feature manifold formed by the base and the novel classes in FSCIL. Zhao et al. [36] propose an attention-based aggregation module that selectively merges predictions from the base branch and the novel branch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.789, + 0.677 + ], + "angle": 0, + "content": "The optimization-based approaches address FSCIL from the perspective of algorithm optimization, which involves representation learning-based methods [35] and knowledge distillation methods [2, 4]. Zhao et al. [35] concentrate on the dilemma between the slow forgetting of old knowledge and the fast adaptation to novel knowledge. Then a multi-grained \"slow vs fast\" learning strategy is proposed to cope with this dilemma from both intra-space and the inter-space. Dong et al. [4] propose a exemplar relation distillation incremental learning framework to balance the tasks of old-knowledge preserving and new-knowledge adaptation. Cui et al. [2] propose to distill reliable knowledge from the reference model, then implement an uncertainty-aware distillation module that combines uncertainty-guided knowledge refinement with adaptive distillation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.787, + 0.723 + ], + "angle": 0, + "content": "Beyond the above approaches, we explore the basic statistical principles underlying FSCIL, then conclude with practical guidelines for the design and training of FSCIL models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.371, + 0.765 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Problem Definition. FSCIL involves a base task followed by a sequence of incremental tasks, each identified by a unique task identifier \\( t \\). Each task is constituted by a distinct training set, \\( \\mathcal{X}_{train}^{t} \\), and a testing set, \\( \\mathcal{X}_{test}^{t} \\), which are associated with their respective distinct label spaces \\( \\mathcal{Y}^t \\), such that \\( \\forall i,j \\) with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "\\(i \\neq j\\), \\(\\mathcal{Y}^i \\cap \\mathcal{Y}^j = \\emptyset\\). For a particular task \\(t\\), training is confined to \\(\\mathcal{X}_{train}^t\\), while testing is performed across all encountered tasks. Notably, \\(\\mathcal{X}_{train}^0\\) encompasses a substantial number of samples for the base task, while \\(\\mathcal{X}_{train}^{t > 0}\\) are limited to \\(M\\) samples per task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.784, + 0.254 + ], + "angle": 0, + "content": "To tackle FSCIL, popular work [17,30-32] assumes a shared feature representation, enabling knowledge transfer from the base to incremental tasks. Building upon this premise, we introduce the following:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.26, + 0.788, + 0.292 + ], + "angle": 0, + "content": "Assumption 1 (Homogeneity) In FSCIL, the data pairs \\(\\{(\\mathbf{x}_{t,i},y_{t,i})\\}\\) for each task \\(t\\), are i.i.d. drawn from a distribution \\(\\mathbb{P}_t\\) over \\(\\mathcal{X}^t\\times \\mathcal{Y}^t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.301, + 0.785, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} _ {t} (\\mathbf {x}, y) = \\mathbb {P} _ {\\mathbf {x}} (\\mathbf {x}) \\cdot \\mathbb {P} _ {y | \\mathbf {x}} \\left(y \\mid f _ {t} ^ {\\star} \\circ \\mathbf {h} ^ {\\star} (\\mathbf {x})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.326, + 0.788, + 0.388 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}:\\mathbb{R}^d\\to \\mathbb{R}^r\\) denotes a shared feature representation, and \\(f_{t}:\\mathbb{R}^{r}\\rightarrow \\mathbb{R}^{c}\\) represents a task-specific mapping. Crucially, this assumption entails that all tasks are subject to an identical marginal distribution \\(\\mathbb{P}_{\\mathbf{x}}\\), indicative of a unified dataset shared across the tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.409, + 0.621, + 0.427 + ], + "angle": 0, + "content": "4 Approximation Risk of FSCIL Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.784, + 0.469 + ], + "angle": 0, + "content": "As explicated in existing work [23,32], the training procedure for FSCIL models is roughly partitioned into two distinct phases: the base and incremental phases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.784, + 0.515 + ], + "angle": 0, + "content": "Base Phase. During the base phase, the primary objective is the learning of an optimal shared representation \\(\\mathbf{h}\\) in conjunction with the base classifier \\(f_{0}\\). This is accomplished by minimizing the empirical risk \\(\\hat{R}_{\\mathrm{base}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.524, + 0.785, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\hat {R} _ {\\text {b a s e}} \\left(f _ {0}, \\mathbf {h}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell \\left(f _ {0} \\circ \\mathbf {h} \\left(\\mathbf {x} _ {0, i}\\right), y _ {0, i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.695 + ], + "angle": 0, + "content": "where \\(\\ell(\\cdot, \\cdot)\\) symbolizes the loss function; \\(\\mathbf{x}_{0,i} \\in \\mathcal{X}_{\\mathrm{train}}^0\\) and \\(y_{0,i} \\in \\mathcal{Y}^0\\). The composition operator \\(\\circ\\) denotes the consecutive application of functions such that \\(f_0 \\circ \\mathbf{h}(\\mathbf{x}) = f_0(\\mathbf{h}(\\mathbf{x}))\\). If Assumption 1 holds and the dataset size \\(N\\) is sufficiently large (\\(N \\gg 1\\)), the estimator \\((\\hat{f}_0, \\hat{\\mathbf{h}})\\), derived from the minimization of \\(\\hat{R}_{\\mathrm{base}}\\), serves as an effective proxy for the optimal function pair \\((f_0^\\star, h^\\star)\\). This base phase is crucial in forming a robust baseline for the learning agent, enabling a seamless transition and effective integration with the novel information presented in subsequent incremental learning tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Incremental Phase. The incremental training process continues with a sequence of incremental tasks, where the learning agent learns the task-specific classifier \\( f_{t} \\) by minimizing empirical risk \\( \\hat{R}_{\\mathrm{novel}} \\) over \\( M \\) new training samples3:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.751, + 0.785, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\hat {R} _ {\\text {n o v e l}} \\left(\\left[ f _ {0}, f _ {t} \\right], \\mathbf {h}\\right) = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\ell \\left(\\left[ f _ {0}, f _ {t} \\right] \\circ \\mathbf {h} \\left(\\mathbf {x} _ {t, m}\\right), y _ {t, m}\\right), \\tag {4}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.842 + ], + "angle": 0, + "content": "3 For brevity in our theoretical exposition, we confine this discussion to a single incremental task; nevertheless, the principles delineated here are also applicable to sequential tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.241 + ], + "angle": 0, + "content": "where \\([f_0, f_t]\\) signifies the concatenation of the base and incremental classifiers; \\(\\mathbf{x}_{t,m} \\in \\mathcal{X}_{\\mathrm{train}}^t\\) and \\(y_{t,m} \\in \\mathcal{Y}^t\\). The challenge herein lies in the fact that for \\(M \\ll N\\), the estimator \\(([\\hat{f}_0, \\hat{f}_t], \\hat{\\mathbf{h}})\\), obtained from minimizing \\(\\mathbf{h} \\in \\mathcal{H}\\) fails to estimate their underlying counterparts \\(([f_0^\\star, f_t^\\star], h^\\star)\\). To address this limitation, we assess the efficacy of the estimator \\(([\\hat{f}_0, \\hat{f}_t], \\hat{\\mathbf{h}})\\) by its excess risk on the incremental task \\(t\\), defined as the approximation risk of FSCIL:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.252, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\mathrm {A R}} = R _ {\\text {n o v e l}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(\\left[ f _ {0} ^ {\\star}, f _ {t} ^ {\\star} \\right], \\mathbf {h} ^ {\\star}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.281, + 0.787, + 0.312 + ], + "angle": 0, + "content": "Approximation Risk. To clarify the approximation risk, we deconstruct \\(\\varDelta_{\\mathrm{AR}}=\\varDelta_{\\mathrm{transfer}}+\\varDelta R_{\\mathrm{novel}}^{4}\\). These components are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.323, + 0.785, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\text {t r a n s f e r}} = R _ {\\text {n o v e l}} \\left(\\hat {f} _ {t}, \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(f _ {t} ^ {\\star}, \\mathbf {h} ^ {\\star}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.344, + 0.785, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\Delta R _ {\\text {n o v e l}} = R _ {\\text {n o v e l}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(\\hat {f} _ {t}, \\hat {\\mathbf {h}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.373, + 0.787, + 0.448 + ], + "angle": 0, + "content": "Within this framework, \\(\\Delta_{\\mathrm{AR}}\\) is deconstructed into two principal components: \\(\\Delta_{\\mathrm{transfer}}\\) and \\(\\Delta R_{\\mathrm{novel}}\\). The former signifies the transfer risk, assessing the expected prediction risk when transferring knowledge to novel tasks. The latter encapsulates the adaptation risk, which particularly measures the divergence between the composite classifier \\([f_0, f_t]\\) and the task-specific classifier \\(f_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.449, + 0.787, + 0.493 + ], + "angle": 0, + "content": "Beyond \\(\\varDelta_{\\mathrm{AR}}\\), FSCIL additionally necessitates the consistent performance across incremental tasks. This is formally represented by a constraint encapsulating the adaptation risk pertaining to the base classifier:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.505, + 0.785, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\Delta R _ {\\text {b a s e}} = R _ {\\text {b a s e}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {b a s e}} \\left(\\hat {f} _ {0}, \\hat {\\mathbf {h}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.534, + 0.787, + 0.579 + ], + "angle": 0, + "content": "In synthesis, the final version of the approximation risk is represented by Eq. (5), incorporating the constraint given by Eq. (8), and is achieved by applying the penalty method:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.593, + 0.785, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\min _ {f _ {0}, f _ {t}, \\mathbf {h}} \\underbrace {\\Delta_ {\\text {t r a n s f e r}}} _ {\\text {T r a n s f e r R i s k}} + \\underbrace {\\Delta R _ {\\text {n o v e l}} + \\gamma \\Delta R _ {\\text {b a s e}}} _ {\\text {C o n s i s t e n c y R i s k}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.636, + 0.52, + 0.651 + ], + "angle": 0, + "content": "where \\(\\gamma\\) represents the penalty coefficient." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.617, + 0.691 + ], + "angle": 0, + "content": "5 Theoretical Results And Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.766 + ], + "angle": 0, + "content": "In this section, the basic statistical principles underlying FSCIL are presented. Initially, we make the following standard, mild regularity assumptions on the loss function \\(\\ell(\\cdot, \\cdot)\\), the function class of tasks \\(\\mathcal{F}\\), and the function class of shared representations \\(\\mathcal{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.776, + 0.755, + 0.792 + ], + "angle": 0, + "content": "Assumption 2 (Regularity conditions) The following conditions hold:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.8, + 0.787, + 0.817 + ], + "angle": 0, + "content": "- The loss function \\(\\ell(\\cdot, \\cdot)\\) is \\(B\\)-bounded, and \\(\\ell(\\cdot, y)\\) is \\(\\Gamma\\)-Lipschitz for all \\(y \\in \\mathcal{Y}\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "Proof In Appendix 9.4" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.145, + 0.672, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.785, + 0.29 + ], + "angle": 0, + "content": "Fig.2: Architecture-Agnostic Modifications to FSCIL Models. Specific adjustments are indicated in red." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.321, + 0.777, + 0.336 + ], + "angle": 0, + "content": "- The function \\( f \\) is \\( \\Gamma(\\mathcal{F}) \\)-Lipschitz regarding the \\( \\ell_2 \\) distance for any \\( f \\in \\mathcal{F} \\)." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.336, + 0.785, + 0.366 + ], + "angle": 0, + "content": "- The composed function \\( f \\circ \\mathbf{h} \\) is bounded: \\( \\sup_{\\mathbf{x} \\in \\mathcal{X}} |f \\circ \\mathbf{h}(\\mathbf{x})| \\leq D_{\\mathcal{X}} \\), for any \\( f \\in \\mathcal{F}, \\mathbf{h} \\in \\mathcal{H} \\)." + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.321, + 0.785, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.407 + ], + "angle": 0, + "content": "Furthermore, the realizability assumption is adopted, positing that the true task functions and the true representation are contained in \\(\\mathcal{F}\\) and \\(\\mathcal{H}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.42, + 0.785, + 0.466 + ], + "angle": 0, + "content": "Assumption 3 (Realizability) The true representation \\(\\mathbf{h}^{\\star}\\) is encompassed by \\(\\mathcal{H}\\). Concurrently, for all tasks considered—both base and incremental—the true task-specific functions \\(f_{t}^{\\star}\\) are also contained in \\(\\mathcal{F}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.784, + 0.509 + ], + "angle": 0, + "content": "Based on Regularity and Realizability assumptions, we analyze transfer learning risk \\(\\varDelta_{\\mathrm{transfer}}\\) and consistency risk \\(\\varDelta R_{\\mathrm{novel}}+\\gamma\\varDelta R_{\\mathrm{base}}\\) in Eq. (9) separately." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.532, + 0.375, + 0.545 + ], + "angle": 0, + "content": "5.1 Transfer Risk" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.556, + 0.784, + 0.6 + ], + "angle": 0, + "content": "Theorem 1. (Proof in Appendix 9.2) If Assumptions 2 and 3 hold, and the base task is \\((\\rho,0)\\) diverse, with probability \\(1 - 2\\delta\\), the transfer risk \\(\\Delta_{\\text{transfer}}\\) in Eq. (9) is bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.609, + 0.77, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\tilde {O} \\Big (\\frac {\\Gamma}{\\rho} \\cdot \\Gamma (\\mathcal {F}) \\cdot 4 D \\prod_ {l = 1} ^ {L} Z (l) \\sqrt {\\frac {\\log 2 J}{N}} + \\frac {\\Gamma D _ {\\mathcal {X} ^ {0}}}{\\rho N ^ {2}} + B \\Big [ \\frac {1}{\\rho} \\cdot \\sqrt {\\frac {\\log (2 / \\delta)}{N}} + \\sqrt {\\frac {\\log (2 / \\delta)}{M}} \\Big ] \\Big),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.655, + 0.784, + 0.709 + ], + "angle": 0, + "content": "where \\( N \\) and \\( M \\) quantify the number of training samples of the base and incremental task \\( t \\) respectively; \\( \\| \\mathbf{x} \\| \\leq D \\); \\( L \\) is the depth of the neural network; \\( Z(l) = \\frac{\\| h^{(l)} \\|_{\\infty}}{\\| h^{(l-1)} \\|_{\\infty}} \\), where \\( h^{(l)} \\) denotes the feature representations of the \\( l \\)-th layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.784, + 0.748 + ], + "angle": 0, + "content": "Application. To bridge the basic statistical principles and the practical guidelines, we undertake the subsequent analysis:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.785, + 0.84 + ], + "angle": 0, + "content": "(1) From Theorem 1, it becomes apparent that the quantity \\( N \\), representing the number of training samples for the base task, primarily influences the upper bound of the transfer risk. As \\( N \\) increases, the upper bound tightens at a rate proportional to \\( O\\left(\\frac{1}{\\sqrt{N}}\\right) \\). This relation shows that expanding training datasets for base classes is the most straightforward and potent strategy for addressing the FSCIL challenge, especially when the initial dataset is insufficiently large." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "In scenarios where expanding the data pool is not feasible, the adoption of a foundation model [19,27] becomes a strategic alternative, which is substantiated by our analysis of the transfer risk." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.79, + 0.389 + ], + "angle": 0, + "content": "(2) Theoretically, the upper bound of transfer risk is proportional to \\(\\prod_{l=1}^{L} Z(l)\\), where \\(Z(l)\\) is the ratio of the infinity norm of features at layer \\(l\\) to that at layer \\(l-1\\). While reducing \\(\\prod_{l=1}^{L} Z(l)\\) could lower transfer risk, it might impair the feature extraction, so we do not adopt this modification. In original ResNets and ViTs, each layer typically includes Layer or Batch Normalization, standardizing feature means to 0 and variances to 1. However, this normalization does not limit the infinity norm of features, potentially causing excessively large values in some dimensions and increasing \\(Z(l)\\), then raising transfer risk. To mitigate this, we constrain each layer's features to \\([-1,1]\\), setting \\(Z(l) = 1\\). Since normalization centers most feature values around zero, we introduce the tanh function to cap them. The tanh function bounds values within \\([-1,1]\\) and has a slope of 1 at zero, effectively constraining extreme values without significantly impacting original models. Similarly, we also use norm + tanh to constrain the representation \\(\\mathbf{h}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.496 + ], + "angle": 0, + "content": "Besides, \\( D_{\\mathcal{X}^0} \\) represents the highest probability value of the predicted outcome, exhibiting a positive correlation with the upper bound. The essential strategy to reduce this bound involves calibrating the model to ensure its predictions are tempered, thereby preventing overconfidence. In conclusion, the analysis of both \\( \\prod_{l=1}^{L} Z(l) \\) and \\( D_{\\mathcal{X}^0} \\) shows that preventing excessive focus on specific features is an effective strategy for addressing FSCIL challenge. As shown in Fig. 2, we propose three architecture-agnostic modifications:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.784, + 0.53 + ], + "angle": 0, + "content": "- Incorporation of the tanh activation function subsequent to each normalization layer in the original models;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.531, + 0.784, + 0.56 + ], + "angle": 0, + "content": "- Integration of the layer normalization and tanh activation function prior to the generation of shared representations;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.561, + 0.642, + 0.575 + ], + "angle": 0, + "content": "- Application of the label smooth in the training process." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.502, + 0.784, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.785, + 0.709 + ], + "angle": 0, + "content": "Remark. The term \\( Z(l) = \\frac{\\|h^{(l)}\\|_{\\infty}}{\\|h^{(l - 1)}\\|_{\\infty}} \\) quantifies the ratio of maximal activation value in the \\( l \\)-th layer compared to that in the \\( (l - 1) \\)-th layer. It serves as an indicator of the proportional activation dynamics spanning all dimensions between consecutive layers. This metric emphasizes the need for a balanced distribution of activation values throughout the network layers to inhibit the incidence of impulse activation. Such regulation ensures that the model maintains an equitable emphasis on all features, thereby supporting the underlying expectation of \\( D_{\\chi^0} \\) for the model to prevent overconfidence in its predictions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.714, + 0.787, + 0.759 + ], + "angle": 0, + "content": "Corollary 1 (Proof in Appendix 9.3). Under conditions from Theorem 1, Vision Transformers (ViTs) exhibit lower transfer risks than Convolutional Neural Networks (CNNs)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Remark. This finding substantiates the efficacy of ViTs in leveraging shared representations for novel tasks, underscoring their robustness in transfer learning scenarios. Furthermore, the reduced transfer risk with ViTs as compared to CNNs suggests a pivotal architectural advantage that may inform future developments in neural network design." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.404, + 0.162 + ], + "angle": 0, + "content": "5.2 Consistency Risk" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.172, + 0.788, + 0.218 + ], + "angle": 0, + "content": "Theorem 2 (Proof in Appendix 9.5). If Assumptions 2 and 3 hold, and given \\( K_{0} \\) and \\( K_{t} \\) classes in the base and incremental tasks respectively, the consistency risk \\( \\Delta R_{\\text{novel}} + \\gamma \\Delta R_{\\text{base}} \\) in Eq. (9) is upper-bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.229, + 0.785, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\log \\left(\\left(1 + \\frac {K _ {0}}{K _ {t}} v _ {t}\\right) \\left(1 + \\frac {K _ {t}}{K _ {0}} v _ {0}\\right) ^ {\\gamma}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.27, + 0.63, + 0.286 + ], + "angle": 0, + "content": "where \\( v_{t} \\) and \\( v_{0} \\) involve supremum and infimum of logits:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.298, + 0.785, + 0.313 + ], + "angle": 0, + "content": "\\[\nv _ {t} = \\exp \\left(\\sup \\left(\\phi_ {t \\rightarrow 0}\\right) - \\inf \\left(\\phi_ {t \\rightarrow t}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.317, + 0.785, + 0.332 + ], + "angle": 0, + "content": "\\[\nv _ {0} = \\exp \\left(\\sup \\left(\\phi_ {0 \\rightarrow t}\\right) - \\inf \\left(\\phi_ {0 \\rightarrow 0}\\right)\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.343, + 0.788, + 0.389 + ], + "angle": 0, + "content": "The classifier logits \\(\\phi_{t\\to 0}\\), \\(\\phi_{t\\to t}\\), \\(\\phi_{0\\to 0}\\), and \\(\\phi_{0\\to t}\\) respectively quantify how samples from task \\(t\\) and the task 0 (base task) are classified, either correctly or in-correctly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.786, + 0.429 + ], + "angle": 0, + "content": "Application. Building on the above analysis about the transfer risk, we proceed with the following detailed analysis about the consistency risk:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.786, + 0.504 + ], + "angle": 0, + "content": "(3) From Theorem 2, we observe that the consistency risk is inherently tied to the classification margin discrepancy of the model, which is quantified by the difference between the infimum of the classifier logits for correct classification and the supremum for incorrect classification. It infers that the consistency risk approaches nullity as the predictive accuracy of the model approaches perfection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.61 + ], + "angle": 0, + "content": "(4) The upper bound of the consistency risk represents a balance between \\( v_{t} \\) and \\( v_{0} \\). Introducing a bias in categorization leads to inverse variations in \\( v_{t} \\) and \\( v_{0} \\), indicating that ensuring unbiased classification across both base and novel classes can completely nullify the consistency risk. This implies strategically designed biases might reach the minimal risk bound but are not optimally efficient. Key lies in improving classification margin discrepancy while maintaining a balanced classification bias." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.43, + 0.787, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.376, + 0.65 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.57, + 0.68 + ], + "angle": 0, + "content": "6.1 Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. Our study employs three diverse datasets: mini-ImageNet [20], CIFAR100 [12], and CUB200 [29]. The mini-ImageNet, a subset of ImageNet [20], includes 600 images from 100 classes, split into 60 base and 40 novel classes. These novel classes are divided into 8 incremental tasks, each a 5-way 5-shot task. CIFAR100 comprises 100 classes with 600 images each, partitioned into 60 base and 40 novel classes, the latter organized into 8 incremental 5-way 5-shot tasks. CUB200, containing 11,788 images of 200 bird species, is split into 100 base and 100 novel classes, with the latter arranged into 10 incremental 10-way 5-shot tasks. In this work, when the model is trained from scratch, the image sizes for mini-ImageNet, CIFAR100, and CUB200 are set at \\(84 \\times 84\\), \\(32 \\times 32\\)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Classification Accuracy (\\%) on CUB200 in the 10-Way 5-Shot FS-CIL Setting. The notation ViT-B/16 specifically denotes the ViT-Base model [5] with a designated patch size of 16. The symbols \\(\\diamond\\) and \\(\\dagger\\) signify models derived from the CLIP [19] and those pre-trained on the ImageNet-1K dataset respectively. The symbol \\(\\downarrow\\) is employed to indicate that lower values are more desirable." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.227, + 0.761, + 0.38 + ], + "angle": 0, + "content": "
MethodBackboneAccuracy in each session(%)DR↓(%)
012345678910
\\( CEC^† \\) [32]ResNet1875.8571.9468.5063.562.4358.2757.7355.8154.8353.5252.2831.07
\\( MCNet^† \\) [10]ResNet1877.5773.9670.4765.8166.1663.8162.0961.8260.4160.0959.0823.84
\\( FACT^† \\) [38]ResNet1875.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9424.98
\\( ALICE^† \\) [17]ResNet1877.472.770.667.265.963.462.961.960.560.660.122.35
\\( CABD^† \\) [37]ResNet1879.1275.6373.2169.9368.3266.3065.1564.9664.2064.0363.8119.35
\\( NC-FSCIL^† \\) [30]ResNet1880.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4426.12
\\( SAVC^† \\) [22]ResNet1881.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5023.64
\\( LDC^† \\) [15]ResNet1877.8976.9374.6470.0668.8867.1564.8364.1663.0362.3961.5820.94
\\( CoCoOp^° \\) [39]ViT-B/1680.377.175.873.972.468.265.263.961.160.357.228.76
\\( Prompt^° \\) [31]ViT-B/1684.383.280.979.277.772.472.169.968.867.466.820.76
\\(ours^{\\dagger}\\)ResNet1879.8777.3674.9270.6270.6468.5568.1467.9366.7366.6765.6917.75
\\(ours^{\\circ}\\)ViT-B/1686.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
\\(ours^{\\dagger}\\)ViT-B/1686.6985.3884.3482.4383.0881.0781.3881.4281.2181.1181.306.22
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.786, + 0.433 + ], + "angle": 0, + "content": "and \\(224 \\times 224\\) respectively. In scenarios utilizing a foundation model, the images from all three datasets are resized to \\(224 \\times 224\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Implementation Details. Following previous work [23, 34], and guided by our theory, we select ViTs as our primary models across all datasets. Specifically, we utilize two ViT-B/16 variants: one from CLIP [19] and another pre-trained on ImageNet-1K. During the base phase, we train the ViT models on the base classes for 50 epochs using the AdamW optimizer. The learning rate is initially set to 0, then linearly rises to 0.00001 during the first 20 epochs, followed by a linear decrease back to 0 for the rest of the epochs. For a fair comparison, ResNet18 serves as an auxiliary model across datasets, trained on base classes for 500 epochs using AdamW. The learning rate for ResNet18 follows a similar pattern: beginning at 0, increasing linearly to 0.1 within the first 50 epochs, and subsequently diminishing back to 0. In the incremental phase, we freeze the backbone and update only the task-specific classifiers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Evaluation Metrics. Following precedents [16, 23], our evaluation employs Top-1 classification accuracy across cumulative testing sets after each learning session. While traditional models often employ the Performance Drop (PD) metric, defined as \\(\\mathrm{PD} = (\\mathcal{A}_0 - \\mathcal{A}_N)\\), where \\(\\mathcal{A}_0\\) and \\(\\mathcal{A}_N\\) represent the accuracies at the initial and final sessions, we propose a refined metric. Recognizing that model forgetting encompasses both the volume of initially acquired knowledge and the extent of subsequent knowledge loss, we introduce the Performance Drop Rate (DR), calculated as \\(\\mathrm{DR} = (\\mathcal{A}_0 - \\mathcal{A}_N) / \\mathcal{A}_0\\). This metric more accurately captures the effect of knowledge loss, factoring in the initial knowledge base." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.772, + 0.622, + 0.788 + ], + "angle": 0, + "content": "6.2 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We benchmark our model against state-of-the-art performances on three datasets: CUB200, CIFAR100, and mini-ImageNet, with results in Fig. 3 and detailed CUB200 data in Table 1 (see supplementary materials for other datasets)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.146, + 0.413, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.147, + 0.59, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.147, + 0.77, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.294, + 0.242, + 0.71, + 0.259 + ], + "angle": 0, + "content": "--- CEC --- FACT --- CABD --- SAVC --- CoCoOp : ours (RosNet18) : ours (ImageNet) \n--- MCNet --- ALICE --- NC-FSCIL --- LDC --- Prompt : ours (CLIP) -- C-FSCL" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.273, + 0.787, + 0.316 + ], + "angle": 0, + "content": "Fig. 3: Comparison with State-of-the-Art on Three Datasets: CUB200, mini-ImageNet, and CIFAR100. For detailed performance metrics, please see Table 1 and refer to our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.479 + ], + "angle": 0, + "content": "Table 1 shows that our ViT model (CLIP) achieves an improvement of \\(8.42\\%\\) in the final session compared to the sub-optimal method, and the DR metric decreases by \\(7.76\\%\\). When ImageNet-1K is used for pre-training, our ViT model records a boost of \\(17.49\\%\\) in the last session compared to the second best method, with a DR reduction of \\(13.13\\%\\). As for ResNet18, our ResNet18 model exhibits a \\(1.88\\%\\) improvement in the final session over the sub-optimal ResNet18 model, with a \\(1.6\\%\\) reduction in the DR metric. These results indicate a decrease in catastrophic forgetting and overfitting, attributed to the application of our guiding principles for reducing the approximation risk." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.57 + ], + "angle": 0, + "content": "From Fig. 3, similar trends are observed on the mini-ImageNet and CIFAR100, though the improvements of ViTs are less pronounced. This observation can be attributed to mini-ImageNet being a subset of ImageNet, which violates the standard FSCIL setting, and to the unique image characteristics of CIFAR100, such as smaller image size and lack of background, which differ from the typical natural image distributions used in pre-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Remarkably, even though CLIP has a broader pre-training dataset compared to ImageNet-1K, performance with CLIP remains consistently lower. This phenomenon might be explained by the significant difference between the marginal distributions of the three benchmark datasets and the data used for pre-training of CLIP. Conversely, ImageNet-1K aligns more closely with these distributions, better fulfilling the Homogeneity assumption in our study." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.391, + 0.697 + ], + "angle": 0, + "content": "6.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this section, we assess the impact of four key components of our ViT model on the CUB200 dataset (see supplementary materials for ResNet18). Detailed in Table 2, these components show varied performance effects. The FT strategy exhibits the lowest Top-1 accuracy (38.01% in the final session) and highest DR metric (40.99%), indicating a pronounced bias towards novel classes and substantial consistency risk. Conversely, FR strategy results in improved Top-1 accuracy (52.41%) and lower DR metric (36.63%), yet struggles with catastrophic forgetting. The DS strategy, separating training of base and novel classifiers, is the most effective in reducing bias and consistency risk." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 2: Ablation Study on CUB200 in the 10-Way 5-Shot FSCIL Setting. With ViT-B/16 as the backbone, our study encompasses: \"Stabilize\", enhancing output stability via tanh function after normalization; \"Constrain\", limiting representation values using layer normalization and tanh function; \"LS\" (Label Smoothing) to reduce model overconfidence; and \"BC\" (Base Classifier), employing Fine-Tuning (FT), Freezing (FR), and Discarding (DS) in the incremental phase." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.242, + 0.758, + 0.364 + ], + "angle": 0, + "content": "
StabilizeConstrainLSBCAccuracy in each session(%)DR↓(%)
012345678910
FT64.4259.3255.0450.6248.9145.2243.9742.1340.5939.5038.0140.99
FR82.7176.3671.3366.9564.8861.0358.9457.4855.7554.1252.4136.63
DS87.0583.5081.0077.6676.5674.1273.0073.1772.3271.8770.9718.47
✓DS87.1584.4181.6578.1177.4675.1874.4473.8773.1772.9772.0717.30
✓DS86.7784.0281.5678.3177.6375.4474.3974.4873.4073.1172.2916.69
✓DS87.4984.0982.1978.5677.9375.4975.5374.6773.4673.4272.7716.82
✓✓DS87.5184.0381.9278.4378.2275.7475.9175.1973.5973.8373.3416.19
✓✓DS87.6384.1381.7778.0578.0075.8075.4074.7974.4374.2573.7115.88
✓✓DS87.4684.5382.7779.3278.9676.6276.8575.8774.5474.4773.9815.41
✓✓✓DS86.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.377, + 0.751, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Fig. 4: Impact of Dataset Size in Transfer Risk with ResNet18 and ViT-S/6 Backbones Trained from Scratch on mini-ImageNet. In both subfigures, the horizontal axis represents the dataset size per base class, while the vertical axis lists training techniques. These include Data Augmentation [1,7,11] and Mixup [18,33], essential for expanding the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Integration of the tanh function for layer stabilization improves the DR metric by about \\(1\\%\\), mitigating overfitting. Similarly, constrained feature representations through layer normalization and tanh function slightly decrease base class accuracy but enhance the DR metric by \\(1\\%\\), indicating reduced overfitting. Implementing label smoothing technique raises both base classification accuracy and DR metric by \\(1\\%\\), attributed to lowering transfer risk." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Combining these components yields superior performance, highlighting their synergistic effect in FSCIL." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.506, + 0.771 + ], + "angle": 0, + "content": "6.4 Dataset Size in Transfer Risk" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this section, we empirically investigate the correlation between dataset size and transfer risk in neural networks, focusing on ResNet18 and ViT-S/6 architectures. Fig. 4 (a) and (b), show a clear correlation between increased dataset size per class and improved final accuracy for both models. This supports the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.478, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.785, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig. 5: Visualization of ResNet18 and ViT-S/6 on mini-ImageNet. Images are divided by a red dotted line: left for base classes and right for novel classes. Beneath each image, the predicted probability is numerically displayed. The star symbol denotes models in our implementation. In ViT visualizations, image brightness reflects attention level. Darker images signify attention concentrated on specific patches, while brighter ones indicate a more global attention on the whole image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.521 + ], + "angle": 0, + "content": "idea that increasing the dataset size \\((N)\\) effectively tightens the upper bound of transfer risk, thereby enhancing models' capacity for incremental learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.627 + ], + "angle": 0, + "content": "When direct sample collection is limited, Data Augmentation and Mixup techniques present a viable alternative. These methods create additional samples, enriching the training dataset. This is evident in Figures 4 (a) and (b), where both ResNet18 and ViT-S models show significant performance gains with these techniques. This reinforces our theory: indirect sample generation via Data Augmentation and Mixup can enhance model transferability, mirroring the advantages of direct sample collection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.649, + 0.577, + 0.665 + ], + "angle": 0, + "content": "6.5 Analysis of the Model Overconfidence" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.765 + ], + "angle": 0, + "content": "In this section, we provide a visual results to illustrate the influence of prediction confidence on transfer risk. The first and third rows of Fig. 5 indicate that models (CNNs or ViTs) naturally focus on small, highly discriminative regions within base class samples, which leads to overconfidence in these predictions. This propensity leads to misclassification during the transfer of models to novel classes, a phenomenon commonly referred to as supervision collapse." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our approach aims to temper the model's over-reliance on base class predictions. As evidenced in the second and fourth rows of Fig. 5, lowering the probability for base class predictions enables the model to cover broader regions. This expanded scope assists in focusing on distinctive features of novel classes, thus enhancing prediction accuracy for these classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.363, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.22, + 0.321, + 0.232 + ], + "angle": 0, + "content": "(a) -30.84" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.144, + 0.501, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.22, + 0.463, + 0.233 + ], + "angle": 0, + "content": "(b) -15.78" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.645, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.22, + 0.603, + 0.233 + ], + "angle": 0, + "content": "(c)-10.18" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.144, + 0.785, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.22, + 0.738, + 0.233 + ], + "angle": 0, + "content": "(d)-3.5" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 6: The t-SNE visualizations of the logits for test samples on mini-ImageNet, under varying degrees of Classification Margin Discrepancy, are depicted with distinct colors representing different classes. The architectures for Fig. (a), (b), and (c) are based on ViT-S/6 models, each trained from scratch. Conversely, Fig. (d) employs a ViT-B/16 backbone, which is pre-trained on the ImageNet dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.344, + 0.544, + 0.359 + ], + "angle": 0, + "content": "6.6 Classification Margin Discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.49 + ], + "angle": 0, + "content": "As demonstrated in Fig. 6, we observe a progressive enhancement in the separability of logits for test samples, correlating with an incremental increase in the Classification Margin Discrepancy. This trend not only signifies a reduction in the consistency risk but also marks a significant step towards achieving more reliable and robust classification performance in FSCIL settings. These findings empirically validate our theoretical framework, suggesting that optimizing for classification margin discrepancy directly contributes to mitigating consistency risks and enhancing model adaptability across incremental learning tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.512, + 0.36, + 0.528 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.665 + ], + "angle": 0, + "content": "In conclusion, we have conducted a thorough exploration on the approximation risk of FSCIL. By focusing on mitigating transfer and consistency risks and tightening their upper bounds, we have developed practical guidelines for the design and training of FSCIL models. These guidelines have expanded training datasets for base classes, optimized classification margin discrepancy, prevented excessive focus on specific features, and ensured unbiased classification across both base and novel classes. Our extensive experiments have validated these principles, culminating in state-of-the-art performance on three benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.688, + 0.357, + 0.704 + ], + "angle": 0, + "content": "8 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, Homogeneity and Regularity conditions are foundational assumptions. Homogeneity, based on the i.i.d assumption, indicates samples in a dataset share similar characteristics like hues and lighting. Regularity involves boundedness for model convergence and the Lipschitz condition, crucial for representation learning. Practically, comparative experiments show that even when the pretraining and FSCIL benchmark datasets have different marginal distributions, our methods still achieve better results, albeit with smaller gains. This indicates that our theory generally holds but comes at the cost of some practicality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.271 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160403, by the National Natural Science Foundation of China (NSFC) under Grants 62176178, 62106152, and 62303447, by the Postdoctoral Innovation Talents Support Program under No. BX20230399, and by the China Postdoctoral Science Foundation under No. 2023M743702." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.298, + 0.323, + 0.314 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.787, + 0.375 + ], + "angle": 0, + "content": "1. Balestriero, R., Bottou, L., LeCun, Y.: The effects of regularization and data augmentation are class dependent. In: Advances in Neural Information Processing Systems. vol. 35, pp. 37878-37891 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.787, + 0.417 + ], + "angle": 0, + "content": "2. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023), early Access" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.787, + 0.46 + ], + "angle": 0, + "content": "3. Doersch, C., Gupta, A., Zisserman, A.: Crosstransformers: spatially-aware few-shot transfer. In: Advances in Neural Information Processing Systems. pp. 21981-21993 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.787, + 0.502 + ], + "angle": 0, + "content": "4. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.558 + ], + "angle": 0, + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.787, + 0.586 + ], + "angle": 0, + "content": "6. Golowich, N., Rakhlin, A., Shamir, O.: Size-independent sample complexity of neural networks. In: Conference On Learning Theory. pp. 297-299. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.787, + 0.629 + ], + "angle": 0, + "content": "7. Hao, X., Zhu, Y., Appalaraju, S., Zhang, A., Zhang, W., Li, B., Li, M.: Mixgen: A new multi-modal data augmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 379-389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.787, + 0.671 + ], + "angle": 0, + "content": "8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.713 + ], + "angle": 0, + "content": "9. Hiller, M., Ma, R., Harandi, M., Drummond, T.: Rethinking generalization in few-shot classification. In: Advances in Neural Information Processing Systems. pp. 1-13 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.715, + 0.787, + 0.755 + ], + "angle": 0, + "content": "10. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing 32, 937-948 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.787, + 0.81 + ], + "angle": 0, + "content": "11. Ji, Z., Jiao, Z., Wang, Q., Pang, Y., Han, J.: Imbalance mitigation for continual learning via knowledge decoupling and dual enhanced contrastive learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-14 (2024), early Access" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "12. Krizhevsky, A., Nair, V., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.333, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "X. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Kukleva, A., Kuehne, H., Schiele, B.: Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9029 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.785, + 0.218 + ], + "angle": 0, + "content": "14. Ledoux, M., Talagrand, M.: Probability in Banach Spaces: isoperimetry and processes, vol. 23. Springer Science & Business Media (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.261 + ], + "angle": 0, + "content": "15. Liu, B., Yang, B., Xie, L., Wang, R., Tian, Q., Ye, Q.: Learnable distribution calibration for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-8 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.289 + ], + "angle": 0, + "content": "16. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2337-2345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.291, + 0.785, + 0.331 + ], + "angle": 0, + "content": "17. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.333, + 0.785, + 0.374 + ], + "angle": 0, + "content": "18. Pinto, F., Yang, H., Lim, S.N., Torr, P., Dokania, P.: Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In: Advances in Neural Information Processing Systems. vol. 35, pp. 14608-14622 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.376, + 0.785, + 0.431 + ], + "angle": 0, + "content": "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.473 + ], + "angle": 0, + "content": "20. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.785, + 0.5 + ], + "angle": 0, + "content": "21. Shalev-Shwartz, S., Ben-David, S.: Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.503, + 0.785, + 0.557 + ], + "angle": 0, + "content": "22. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24183-24192 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.559, + 0.785, + 0.599 + ], + "angle": 0, + "content": "23. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "24. Tripuraneni, N., Jordan, M., Jin, C.: On the theory of transfer learning: The importance of task diversity. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7852-7862 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.644, + 0.785, + 0.685 + ], + "angle": 0, + "content": "25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30, pp. 1-11 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.785, + 0.713 + ], + "angle": 0, + "content": "26. Wainwright, M.J.: High-dimensional statistics: A non-asymptotic viewpoint, vol. 48. Cambridge university press (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.785, + 0.769 + ], + "angle": 0, + "content": "27. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "28. Wang, X., Ji, Z., Pang, Y., Yu, Y.: A cognition-driven framework for few-shot class-incremental learning. Neurocomputing 600, 128118 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Approximation Risk of FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations. pp. 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "31. Yoon, I.U., Choi, T.M., Lee, S.K., Kim, Y.M., Kim, J.H.: Image-object-specific prompt learning for few-shot class-incremental learning. arXiv preprint arXiv:2309.02833 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "32. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12455-12464 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.3 + ], + "angle": 0, + "content": "33. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "34. Zhang, J., Liu, L., Silven, O., Pietikainen, M., Hu, D.: Few-shot class-incremental learning: A survey. arXiv preprint arXiv:2308.06764 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "35. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-11 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.424 + ], + "angle": 0, + "content": "36. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.48 + ], + "angle": 0, + "content": "37. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.522 + ], + "angle": 0, + "content": "38. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.564 + ], + "angle": 0, + "content": "39. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.564 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_origin.pdf b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f0ab8a17b8d0857d281903d9aa061168a47fc842 --- /dev/null +++ b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/72ad6bc5-15cc-4de3-b1bf-33c0f70fa389_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed6e17445be41a4f7b5144a9bee2d0d415c5c88aa69b0c7d4f1ea353268e6c98 +size 7202403 diff --git a/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/full.md b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db5cfbce4cd8abd1e8a829758ec4a0bdbc986f88 --- /dev/null +++ b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/full.md @@ -0,0 +1,341 @@ +# On the Approximation Risk of Few-Shot Class-Incremental Learning + +Xuan Wang $^{1,2}$ , Zhong Ji $^{1,2}$ , Xiyao Liu $^{3}$ , Yanwei Pang $^{1,2}$ , and Jungong Han $^{4}$ + +$^{1}$ Tianjin University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ State Key Laboratory of Robotics, SIA, CAS $^{4}$ University of Sheffield + +Abstract. Few-Shot Class-Incremental Learning (FSCIL) aims to learn new concepts with few training samples while preserving previously acquired knowledge. Although promising performance has been achieved, there remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Therefore, we thoroughly explore the approximation risk of FSCIL, encompassing both transfer and consistency risks. By tightening the upper bounds of these risks, we derive practical guidelines for designing and training FSCIL models. These guidelines include (1) expanding training datasets for base classes, (2) preventing excessive focus on specific features, (3) optimizing classification margin discrepancy, and (4) ensuring unbiased classification across both base and novel classes. Leveraging these insights, we conduct comprehensive experiments to validate our principles, achieving state-of-the-art performance on three FSCIL benchmark datasets. Code is available at https://github.com/xwangrs/Approximation_FSCIL-ECCV2024.git. + +Keywords: Few-Shot Class-Incremental Learning $\cdot$ Few-Shot Learning $\cdot$ Class Incremental Learning $\cdot$ Approximation Risk + +# 1 Introduction + +Few-Shot Class-Incremental Learning (FSCIL) [4, 23, 34], an emergent learning paradigm, aims to learn novel classes with limited training data, while preserving the knowledge acquired from base classes. This paradigm faces two primary issues: (1) the catastrophic forgetting of established base class knowledge when integrating new information, and (2) the overfitting to few-shot novel classes. + +To effectively address these issues, researchers have proposed various approaches, broadly categorized into three groups [34]: data-based [13, 17, 38], structure-based [23, 32, 36], and optimization-based [2, 4, 35] approaches. Despite their proven effectiveness, there still remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Central to this gap is the question: What are the key factors in solving FSCIL problems? To answer this question, our research delves into the basic statistical principles of FSCIL, aiming to guide the design and training of FSCIL models. + +![](images/ff4f8134f9ea14acd57e3a7063a331d7a92653bd8d7e1fad08e3312ce5f0bdcc.jpg) +(a) Transfer Risk + +![](images/3446c78168d56bbc9b5a2334ec2bebf819807655c980ad745d92ac29386943db.jpg) +(b) Consistency Risk +Fig. 1: The upper bounds of the transfer and consistency risks. In Fig. 1 (a), the horizontal axis represents the number of training samples for base classes, whereas the vertical axis quantifies the magnitude of the transfer risk. Similarly, in Fig. 1 (b), the horizontal axis reflects the classification margin discrepancy, while the vertical axis quantifies the magnitude of the consistency risk. The parameters $\eta > 0$ and $\kappa > 0$ serve as dynamic scaling factors, regulating the transfer and consistency risks. + +We begin with fundamental assumptions: Homogeneity, Regularity conditions, and Realizability. These assumptions, grounded in practical scenarios, form the foundation of our theory. + +Based on these assumptions, we formulate the gap between empirically obtained and theoretically optimal FSCIL models as an optimization problem, whose objective function is defined as the Approximation Risk of FSCIL. This risk includes two components: transfer risk and consistency risk. The transfer risk [24] evaluates the efficacy of the shared representations in knowledge transfer from base to novel classes. A low transfer risk indicates effective utilization of base classes knowledge to compensate for the scarcity of novel class samples, thereby preventing overfitting in these novel classes. Meanwhile, the consistency risk measures performance drop when incorporating novel classes, with a low risk signifying robust and stable performance throughout class-incremental tasks, thereby preventing the catastrophic forgetting of base classes. + +Leveraging statistical learning theory [6, 14, 21, 26], we explore the upper bounds of transfer and consistency risks, then identifying their determinants2: + +$$ +\underbrace {\eta \cdot \frac {1}{\sqrt {N}}} _ {\text {T r a n s f e r R i s k}} + \underbrace {\log \left(1 + \frac {\kappa}{\exp (\nu)}\right)} _ {\text {C o n s i s t e n c y R i s k}}, \tag {1} +$$ + +where $N$ denotes the number of training samples for the base classes; $\nu$ indicates the classification margin discrepancy, defined as the infimum of classifier logits for correct classifications minus the supremum of those for incorrect classifications; the parameters $\eta$ and $\kappa$ are the dynamic scaling factors for transfer and consistency risks, respectively. We will further discuss and elaborate on methods for minimizing these risks by tightening their upper bounds. + +In Fig. 1 (a), it is evident that increasing the number of training samples for base classes leads to a significant reduction in the transfer risk. Theoretically, this risk could potentially be minimized to a negligible level with a sufficiently large and diverse dataset. However, in practical scenarios, gathering sufficient data for base classes remains a challenge, thereby prompting reliance on foundation models [19,31]. Crucially, for the Homogeneity assumption to hold, the marginal distribution of the pre-trained data should align with that of the base class data. Failing to meet this condition could lead to an increase in the transfer risk. + +In exploring how $N$ affects the transfer risk, it becomes imperative to fully comprehend the role of the dynamic scaling factor $\eta$ . From a theoretical perspective, $\eta$ represents a multivariate polynomial that is related to the maximum activation values across each layer of the neural networks. This could be interpreted as the model's selective attention to specific features. An increase in $\eta$ values indicates a marked emphasis on certain features, often at the expense of neglecting others, thereby increasing the transfer risk. This phenomenon is known as the "supervision collapse" [3, 9]. To mitigate this issue, we propose to maintain a balanced distribution of activation values across the network layers to prevent extreme impulse values. Furthermore, our theoretical analysis indicates that Vision Transformers (ViTs) [5], owing to their self-attention mechanism [25], exhibit a lower transfer risk compared to Convolutional Neural Networks (CNNs) [8]. Thus, in addressing the FSCIL problem, our preference is towards the deployment of ViT models. + +Evidenced in Fig. 1 (b), a clear inverse relationship exists between the classification margin discrepancy and the consistency risk, indicating that enhancing the inter-class discrimination of representations can theoretically diminish the consistency risk to minimal levels. This requires models to maximize the classification margin discrepancy on base classes, ensuring comprehensive training of these models. Moreover, our theory also indicates that the scaling factor $\kappa$ reaches its lower bound when the model exhibits unbiased classification across both base and novel classes. Achieving this balance involves training classifiers for each task independently before combining them into a unified classifier. + +We summarize our contributions as follows: + +- Theoretical Framework: We develop a new optimization theory that provides a novel perspective for understanding and addressing challenges in FSCIL. +- Risk Analysis: We decompose the approximation risk of FSCIL into transfer and consistency risks, defining their upper boundaries and and key determinants, thereby offering guidance for model design and training. +- Practical Strategies: We propose practical guidelines to address the challenges in FSCIL. These include expanding training datasets for base classes, preventing excessive focus on specific features, optimizing classification margin discrepancy, and ensuring unbiased classification across both base and novel classes. + +Informed by these insights, we achieve state-of-the-art performance on three benchmark datasets, demonstrating the effectiveness of our theory. + +# 2 Related Work + +Recently, FSCIL [23,28] has emerged as a paradigm to incrementally learn novel classes with a limited number of samples, while preserving previously learned knowledge. According to the survey [34], existing approaches can be broadly categorized into three groups: data-based approaches [13,17,38], structure-based approaches [23,32,36], and optimization-based approaches [2,4,35]. + +The data-based approaches address FSCIL from the perspective of training data, which involves data replay-based methods [13] and pseudo scenarios-based methods [17, 38]. For example, Kukleva et al. [13] propose a three-stage framework where the base and the novel class samples are efficiently replayed to calibrate the classification bias. Zhou et al. [38] propose ForwArd Compatible Training (FACT) for FSCIL, which preassigns multiple virtual prototypes in the embedding space by creatively mixuping the known class samples, and Peng et al. [17] have similar ideas about utilizing the Mixup [33] to expand the datasets. + +The structure-based approaches address FSCIL from the perspective of network architecture design, which involves dynamic structure-based methods [23, 32] and attention-based methods [36]. For example, Zhang et al. [32] propose Continually Evolved Classifier (CEC) that employs a graph model to propagate context information between classifiers for adaptation. Tao et al. [23] propose to use a neural "gas" network to learn and preserve the topology of the feature manifold formed by the base and the novel classes in FSCIL. Zhao et al. [36] propose an attention-based aggregation module that selectively merges predictions from the base branch and the novel branch. + +The optimization-based approaches address FSCIL from the perspective of algorithm optimization, which involves representation learning-based methods [35] and knowledge distillation methods [2, 4]. Zhao et al. [35] concentrate on the dilemma between the slow forgetting of old knowledge and the fast adaptation to novel knowledge. Then a multi-grained "slow vs fast" learning strategy is proposed to cope with this dilemma from both intra-space and the inter-space. Dong et al. [4] propose a exemplar relation distillation incremental learning framework to balance the tasks of old-knowledge preserving and new-knowledge adaptation. Cui et al. [2] propose to distill reliable knowledge from the reference model, then implement an uncertainty-aware distillation module that combines uncertainty-guided knowledge refinement with adaptive distillation. + +Beyond the above approaches, we explore the basic statistical principles underlying FSCIL, then conclude with practical guidelines for the design and training of FSCIL models. + +# 3 Preliminary + +Problem Definition. FSCIL involves a base task followed by a sequence of incremental tasks, each identified by a unique task identifier $t$ . Each task is constituted by a distinct training set, $\mathcal{X}_{train}^{t}$ , and a testing set, $\mathcal{X}_{test}^{t}$ , which are associated with their respective distinct label spaces $\mathcal{Y}^t$ , such that $\forall i,j$ with + +$i \neq j$ , $\mathcal{Y}^i \cap \mathcal{Y}^j = \emptyset$ . For a particular task $t$ , training is confined to $\mathcal{X}_{train}^t$ , while testing is performed across all encountered tasks. Notably, $\mathcal{X}_{train}^0$ encompasses a substantial number of samples for the base task, while $\mathcal{X}_{train}^{t > 0}$ are limited to $M$ samples per task. + +To tackle FSCIL, popular work [17,30-32] assumes a shared feature representation, enabling knowledge transfer from the base to incremental tasks. Building upon this premise, we introduce the following: + +Assumption 1 (Homogeneity) In FSCIL, the data pairs $\{(\mathbf{x}_{t,i},y_{t,i})\}$ for each task $t$ , are i.i.d. drawn from a distribution $\mathbb{P}_t$ over $\mathcal{X}^t\times \mathcal{Y}^t$ : + +$$ +\mathbb {P} _ {t} (\mathbf {x}, y) = \mathbb {P} _ {\mathbf {x}} (\mathbf {x}) \cdot \mathbb {P} _ {y | \mathbf {x}} \left(y \mid f _ {t} ^ {\star} \circ \mathbf {h} ^ {\star} (\mathbf {x})\right), \tag {2} +$$ + +where $\mathbf{h}:\mathbb{R}^d\to \mathbb{R}^r$ denotes a shared feature representation, and $f_{t}:\mathbb{R}^{r}\rightarrow \mathbb{R}^{c}$ represents a task-specific mapping. Crucially, this assumption entails that all tasks are subject to an identical marginal distribution $\mathbb{P}_{\mathbf{x}}$ , indicative of a unified dataset shared across the tasks. + +# 4 Approximation Risk of FSCIL Models + +As explicated in existing work [23,32], the training procedure for FSCIL models is roughly partitioned into two distinct phases: the base and incremental phases. + +Base Phase. During the base phase, the primary objective is the learning of an optimal shared representation $\mathbf{h}$ in conjunction with the base classifier $f_{0}$ . This is accomplished by minimizing the empirical risk $\hat{R}_{\mathrm{base}}$ : + +$$ +\hat {R} _ {\text {b a s e}} \left(f _ {0}, \mathbf {h}\right) = \frac {1}{N} \sum_ {i = 1} ^ {N} \ell \left(f _ {0} \circ \mathbf {h} \left(\mathbf {x} _ {0, i}\right), y _ {0, i}\right), \tag {3} +$$ + +where $\ell(\cdot, \cdot)$ symbolizes the loss function; $\mathbf{x}_{0,i} \in \mathcal{X}_{\mathrm{train}}^0$ and $y_{0,i} \in \mathcal{Y}^0$ . The composition operator $\circ$ denotes the consecutive application of functions such that $f_0 \circ \mathbf{h}(\mathbf{x}) = f_0(\mathbf{h}(\mathbf{x}))$ . If Assumption 1 holds and the dataset size $N$ is sufficiently large ( $N \gg 1$ ), the estimator $(\hat{f}_0, \hat{\mathbf{h}})$ , derived from the minimization of $\hat{R}_{\mathrm{base}}$ , serves as an effective proxy for the optimal function pair $(f_0^\star, h^\star)$ . This base phase is crucial in forming a robust baseline for the learning agent, enabling a seamless transition and effective integration with the novel information presented in subsequent incremental learning tasks. + +Incremental Phase. The incremental training process continues with a sequence of incremental tasks, where the learning agent learns the task-specific classifier $f_{t}$ by minimizing empirical risk $\hat{R}_{\mathrm{novel}}$ over $M$ new training samples3: + +$$ +\hat {R} _ {\text {n o v e l}} \left(\left[ f _ {0}, f _ {t} \right], \mathbf {h}\right) = \frac {1}{M} \sum_ {m = 1} ^ {M} \ell \left(\left[ f _ {0}, f _ {t} \right] \circ \mathbf {h} \left(\mathbf {x} _ {t, m}\right), y _ {t, m}\right), \tag {4} +$$ + +where $[f_0, f_t]$ signifies the concatenation of the base and incremental classifiers; $\mathbf{x}_{t,m} \in \mathcal{X}_{\mathrm{train}}^t$ and $y_{t,m} \in \mathcal{Y}^t$ . The challenge herein lies in the fact that for $M \ll N$ , the estimator $([\hat{f}_0, \hat{f}_t], \hat{\mathbf{h}})$ , obtained from minimizing $\mathbf{h} \in \mathcal{H}$ fails to estimate their underlying counterparts $([f_0^\star, f_t^\star], h^\star)$ . To address this limitation, we assess the efficacy of the estimator $([\hat{f}_0, \hat{f}_t], \hat{\mathbf{h}})$ by its excess risk on the incremental task $t$ , defined as the approximation risk of FSCIL: + +$$ +\Delta_ {\mathrm {A R}} = R _ {\text {n o v e l}} \left(\left[ \hat {f} _ {0}, \hat {f} _ {t} \right], \hat {\mathbf {h}}\right) - R _ {\text {n o v e l}} \left(\left[ f _ {0} ^ {\star}, f _ {t} ^ {\star} \right], \mathbf {h} ^ {\star}\right). \tag {5} +$$ + +Approximation Risk. To clarify the approximation risk, we deconstruct $\varDelta_{\mathrm{AR}}=\varDelta_{\mathrm{transfer}}+\varDelta R_{\mathrm{novel}}^{4}$ . These components are defined as: + +$$ +\Delta_ {\text {t r a n s f e r}} = R _ {\text {n o v e l}} \left(\hat {f} _ {t}, \hat {\mathbf {h}}\right) - R _ {\text {n o v e l}} \left(f _ {t} ^ {\star}, \mathbf {h} ^ {\star}\right), \tag {6} +$$ + +$$ +\Delta R _ {\text {n o v e l}} = R _ {\text {n o v e l}} \left(\left[ \hat {f} _ {0}, \hat {f} _ {t} \right], \hat {\mathbf {h}}\right) - R _ {\text {n o v e l}} \left(\hat {f} _ {t}, \hat {\mathbf {h}}\right). \tag {7} +$$ + +Within this framework, $\Delta_{\mathrm{AR}}$ is deconstructed into two principal components: $\Delta_{\mathrm{transfer}}$ and $\Delta R_{\mathrm{novel}}$ . The former signifies the transfer risk, assessing the expected prediction risk when transferring knowledge to novel tasks. The latter encapsulates the adaptation risk, which particularly measures the divergence between the composite classifier $[f_0, f_t]$ and the task-specific classifier $f_t$ . + +Beyond $\varDelta_{\mathrm{AR}}$ , FSCIL additionally necessitates the consistent performance across incremental tasks. This is formally represented by a constraint encapsulating the adaptation risk pertaining to the base classifier: + +$$ +\Delta R _ {\text {b a s e}} = R _ {\text {b a s e}} \left(\left[ \hat {f} _ {0}, \hat {f} _ {t} \right], \hat {\mathbf {h}}\right) - R _ {\text {b a s e}} \left(\hat {f} _ {0}, \hat {\mathbf {h}}\right). \tag {8} +$$ + +In synthesis, the final version of the approximation risk is represented by Eq. (5), incorporating the constraint given by Eq. (8), and is achieved by applying the penalty method: + +$$ +\min _ {f _ {0}, f _ {t}, \mathbf {h}} \underbrace {\Delta_ {\text {t r a n s f e r}}} _ {\text {T r a n s f e r R i s k}} + \underbrace {\Delta R _ {\text {n o v e l}} + \gamma \Delta R _ {\text {b a s e}}} _ {\text {C o n s i s t e n c y R i s k}}, \tag {9} +$$ + +where $\gamma$ represents the penalty coefficient. + +# 5 Theoretical Results And Applications + +In this section, the basic statistical principles underlying FSCIL are presented. Initially, we make the following standard, mild regularity assumptions on the loss function $\ell(\cdot, \cdot)$ , the function class of tasks $\mathcal{F}$ , and the function class of shared representations $\mathcal{H}$ . + +Assumption 2 (Regularity conditions) The following conditions hold: + +- The loss function $\ell(\cdot, \cdot)$ is $B$ -bounded, and $\ell(\cdot, y)$ is $\Gamma$ -Lipschitz for all $y \in \mathcal{Y}$ . + +![](images/34a1ef1ed210b78c27434966106ce0b4cd6c4b8b6ccaae157a959e4cd320b25c.jpg) +Fig.2: Architecture-Agnostic Modifications to FSCIL Models. Specific adjustments are indicated in red. + +- The function $f$ is $\Gamma(\mathcal{F})$ -Lipschitz regarding the $\ell_2$ distance for any $f \in \mathcal{F}$ . +- The composed function $f \circ \mathbf{h}$ is bounded: $\sup_{\mathbf{x} \in \mathcal{X}} |f \circ \mathbf{h}(\mathbf{x})| \leq D_{\mathcal{X}}$ , for any $f \in \mathcal{F}, \mathbf{h} \in \mathcal{H}$ . + +Furthermore, the realizability assumption is adopted, positing that the true task functions and the true representation are contained in $\mathcal{F}$ and $\mathcal{H}$ , respectively. + +Assumption 3 (Realizability) The true representation $\mathbf{h}^{\star}$ is encompassed by $\mathcal{H}$ . Concurrently, for all tasks considered—both base and incremental—the true task-specific functions $f_{t}^{\star}$ are also contained in $\mathcal{F}$ . + +Based on Regularity and Realizability assumptions, we analyze transfer learning risk $\varDelta_{\mathrm{transfer}}$ and consistency risk $\varDelta R_{\mathrm{novel}}+\gamma\varDelta R_{\mathrm{base}}$ in Eq. (9) separately. + +# 5.1 Transfer Risk + +Theorem 1. (Proof in Appendix 9.2) If Assumptions 2 and 3 hold, and the base task is $(\rho,0)$ diverse, with probability $1 - 2\delta$ , the transfer risk $\Delta_{\text{transfer}}$ in Eq. (9) is bounded by: + +$$ +\tilde {O} \Big (\frac {\Gamma}{\rho} \cdot \Gamma (\mathcal {F}) \cdot 4 D \prod_ {l = 1} ^ {L} Z (l) \sqrt {\frac {\log 2 J}{N}} + \frac {\Gamma D _ {\mathcal {X} ^ {0}}}{\rho N ^ {2}} + B \Big [ \frac {1}{\rho} \cdot \sqrt {\frac {\log (2 / \delta)}{N}} + \sqrt {\frac {\log (2 / \delta)}{M}} \Big ] \Big), +$$ + +where $N$ and $M$ quantify the number of training samples of the base and incremental task $t$ respectively; $\| \mathbf{x} \| \leq D$ ; $L$ is the depth of the neural network; $Z(l) = \frac{\| h^{(l)} \|_{\infty}}{\| h^{(l-1)} \|_{\infty}}$ , where $h^{(l)}$ denotes the feature representations of the $l$ -th layer. + +Application. To bridge the basic statistical principles and the practical guidelines, we undertake the subsequent analysis: + +(1) From Theorem 1, it becomes apparent that the quantity $N$ , representing the number of training samples for the base task, primarily influences the upper bound of the transfer risk. As $N$ increases, the upper bound tightens at a rate proportional to $O\left(\frac{1}{\sqrt{N}}\right)$ . This relation shows that expanding training datasets for base classes is the most straightforward and potent strategy for addressing the FSCIL challenge, especially when the initial dataset is insufficiently large. + +In scenarios where expanding the data pool is not feasible, the adoption of a foundation model [19,27] becomes a strategic alternative, which is substantiated by our analysis of the transfer risk. + +(2) Theoretically, the upper bound of transfer risk is proportional to $\prod_{l=1}^{L} Z(l)$ , where $Z(l)$ is the ratio of the infinity norm of features at layer $l$ to that at layer $l-1$ . While reducing $\prod_{l=1}^{L} Z(l)$ could lower transfer risk, it might impair the feature extraction, so we do not adopt this modification. In original ResNets and ViTs, each layer typically includes Layer or Batch Normalization, standardizing feature means to 0 and variances to 1. However, this normalization does not limit the infinity norm of features, potentially causing excessively large values in some dimensions and increasing $Z(l)$ , then raising transfer risk. To mitigate this, we constrain each layer's features to $[-1,1]$ , setting $Z(l) = 1$ . Since normalization centers most feature values around zero, we introduce the tanh function to cap them. The tanh function bounds values within $[-1,1]$ and has a slope of 1 at zero, effectively constraining extreme values without significantly impacting original models. Similarly, we also use norm + tanh to constrain the representation $\mathbf{h}$ . + +Besides, $D_{\mathcal{X}^0}$ represents the highest probability value of the predicted outcome, exhibiting a positive correlation with the upper bound. The essential strategy to reduce this bound involves calibrating the model to ensure its predictions are tempered, thereby preventing overconfidence. In conclusion, the analysis of both $\prod_{l=1}^{L} Z(l)$ and $D_{\mathcal{X}^0}$ shows that preventing excessive focus on specific features is an effective strategy for addressing FSCIL challenge. As shown in Fig. 2, we propose three architecture-agnostic modifications: + +- Incorporation of the tanh activation function subsequent to each normalization layer in the original models; +- Integration of the layer normalization and tanh activation function prior to the generation of shared representations; +- Application of the label smooth in the training process. + +Remark. The term $Z(l) = \frac{\|h^{(l)}\|_{\infty}}{\|h^{(l - 1)}\|_{\infty}}$ quantifies the ratio of maximal activation value in the $l$ -th layer compared to that in the $(l - 1)$ -th layer. It serves as an indicator of the proportional activation dynamics spanning all dimensions between consecutive layers. This metric emphasizes the need for a balanced distribution of activation values throughout the network layers to inhibit the incidence of impulse activation. Such regulation ensures that the model maintains an equitable emphasis on all features, thereby supporting the underlying expectation of $D_{\chi^0}$ for the model to prevent overconfidence in its predictions. + +Corollary 1 (Proof in Appendix 9.3). Under conditions from Theorem 1, Vision Transformers (ViTs) exhibit lower transfer risks than Convolutional Neural Networks (CNNs). + +Remark. This finding substantiates the efficacy of ViTs in leveraging shared representations for novel tasks, underscoring their robustness in transfer learning scenarios. Furthermore, the reduced transfer risk with ViTs as compared to CNNs suggests a pivotal architectural advantage that may inform future developments in neural network design. + +# 5.2 Consistency Risk + +Theorem 2 (Proof in Appendix 9.5). If Assumptions 2 and 3 hold, and given $K_{0}$ and $K_{t}$ classes in the base and incremental tasks respectively, the consistency risk $\Delta R_{\text{novel}} + \gamma \Delta R_{\text{base}}$ in Eq. (9) is upper-bounded by: + +$$ +\log \left(\left(1 + \frac {K _ {0}}{K _ {t}} v _ {t}\right) \left(1 + \frac {K _ {t}}{K _ {0}} v _ {0}\right) ^ {\gamma}\right), \tag {10} +$$ + +where $v_{t}$ and $v_{0}$ involve supremum and infimum of logits: + +$$ +v _ {t} = \exp \left(\sup \left(\phi_ {t \rightarrow 0}\right) - \inf \left(\phi_ {t \rightarrow t}\right)\right), \tag {11} +$$ + +$$ +v _ {0} = \exp \left(\sup \left(\phi_ {0 \rightarrow t}\right) - \inf \left(\phi_ {0 \rightarrow 0}\right)\right). \tag {12} +$$ + +The classifier logits $\phi_{t\to 0}$ , $\phi_{t\to t}$ , $\phi_{0\to 0}$ , and $\phi_{0\to t}$ respectively quantify how samples from task $t$ and the task 0 (base task) are classified, either correctly or in-correctly. + +Application. Building on the above analysis about the transfer risk, we proceed with the following detailed analysis about the consistency risk: + +(3) From Theorem 2, we observe that the consistency risk is inherently tied to the classification margin discrepancy of the model, which is quantified by the difference between the infimum of the classifier logits for correct classification and the supremum for incorrect classification. It infers that the consistency risk approaches nullity as the predictive accuracy of the model approaches perfection. +(4) The upper bound of the consistency risk represents a balance between $v_{t}$ and $v_{0}$ . Introducing a bias in categorization leads to inverse variations in $v_{t}$ and $v_{0}$ , indicating that ensuring unbiased classification across both base and novel classes can completely nullify the consistency risk. This implies strategically designed biases might reach the minimal risk bound but are not optimally efficient. Key lies in improving classification margin discrepancy while maintaining a balanced classification bias. + +# 6 Experiments + +# 6.1 Datasets and Implementation Details + +Datasets. Our study employs three diverse datasets: mini-ImageNet [20], CIFAR100 [12], and CUB200 [29]. The mini-ImageNet, a subset of ImageNet [20], includes 600 images from 100 classes, split into 60 base and 40 novel classes. These novel classes are divided into 8 incremental tasks, each a 5-way 5-shot task. CIFAR100 comprises 100 classes with 600 images each, partitioned into 60 base and 40 novel classes, the latter organized into 8 incremental 5-way 5-shot tasks. CUB200, containing 11,788 images of 200 bird species, is split into 100 base and 100 novel classes, with the latter arranged into 10 incremental 10-way 5-shot tasks. In this work, when the model is trained from scratch, the image sizes for mini-ImageNet, CIFAR100, and CUB200 are set at $84 \times 84$ , $32 \times 32$ , + +Table 1: Classification Accuracy (\%) on CUB200 in the 10-Way 5-Shot FS-CIL Setting. The notation ViT-B/16 specifically denotes the ViT-Base model [5] with a designated patch size of 16. The symbols $\diamond$ and $\dagger$ signify models derived from the CLIP [19] and those pre-trained on the ImageNet-1K dataset respectively. The symbol $\downarrow$ is employed to indicate that lower values are more desirable. + +
MethodBackboneAccuracy in each session(%)DR↓(%)
012345678910
\( CEC^† \) [32]ResNet1875.8571.9468.5063.562.4358.2757.7355.8154.8353.5252.2831.07
\( MCNet^† \) [10]ResNet1877.5773.9670.4765.8166.1663.8162.0961.8260.4160.0959.0823.84
\( FACT^† \) [38]ResNet1875.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9424.98
\( ALICE^† \) [17]ResNet1877.472.770.667.265.963.462.961.960.560.660.122.35
\( CABD^† \) [37]ResNet1879.1275.6373.2169.9368.3266.3065.1564.9664.2064.0363.8119.35
\( NC-FSCIL^† \) [30]ResNet1880.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4426.12
\( SAVC^† \) [22]ResNet1881.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5023.64
\( LDC^† \) [15]ResNet1877.8976.9374.6470.0668.8867.1564.8364.1663.0362.3961.5820.94
\( CoCoOp^° \) [39]ViT-B/1680.377.175.873.972.468.265.263.961.160.357.228.76
\( Prompt^° \) [31]ViT-B/1684.383.280.979.277.772.472.169.968.867.466.820.76
\(ours^{\dagger}\)ResNet1879.8777.3674.9270.6270.6468.5568.1467.9366.7366.6765.6917.75
\(ours^{\circ}\)ViT-B/1686.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
\(ours^{\dagger}\)ViT-B/1686.6985.3884.3482.4383.0881.0781.3881.4281.2181.1181.306.22
+ +and $224 \times 224$ respectively. In scenarios utilizing a foundation model, the images from all three datasets are resized to $224 \times 224$ . + +Implementation Details. Following previous work [23, 34], and guided by our theory, we select ViTs as our primary models across all datasets. Specifically, we utilize two ViT-B/16 variants: one from CLIP [19] and another pre-trained on ImageNet-1K. During the base phase, we train the ViT models on the base classes for 50 epochs using the AdamW optimizer. The learning rate is initially set to 0, then linearly rises to 0.00001 during the first 20 epochs, followed by a linear decrease back to 0 for the rest of the epochs. For a fair comparison, ResNet18 serves as an auxiliary model across datasets, trained on base classes for 500 epochs using AdamW. The learning rate for ResNet18 follows a similar pattern: beginning at 0, increasing linearly to 0.1 within the first 50 epochs, and subsequently diminishing back to 0. In the incremental phase, we freeze the backbone and update only the task-specific classifiers. + +Evaluation Metrics. Following precedents [16, 23], our evaluation employs Top-1 classification accuracy across cumulative testing sets after each learning session. While traditional models often employ the Performance Drop (PD) metric, defined as $\mathrm{PD} = (\mathcal{A}_0 - \mathcal{A}_N)$ , where $\mathcal{A}_0$ and $\mathcal{A}_N$ represent the accuracies at the initial and final sessions, we propose a refined metric. Recognizing that model forgetting encompasses both the volume of initially acquired knowledge and the extent of subsequent knowledge loss, we introduce the Performance Drop Rate (DR), calculated as $\mathrm{DR} = (\mathcal{A}_0 - \mathcal{A}_N) / \mathcal{A}_0$ . This metric more accurately captures the effect of knowledge loss, factoring in the initial knowledge base. + +# 6.2 Comparison with State-of-the-Art Methods + +We benchmark our model against state-of-the-art performances on three datasets: CUB200, CIFAR100, and mini-ImageNet, with results in Fig. 3 and detailed CUB200 data in Table 1 (see supplementary materials for other datasets). + +Fig. 3: Comparison with State-of-the-Art on Three Datasets: CUB200, mini-ImageNet, and CIFAR100. For detailed performance metrics, please see Table 1 and refer to our supplementary material. +![](images/86d59389063e912ebcfd5ed177ed4836ec583f921c92379c5e16dbb05206f5d4.jpg) +--- CEC --- FACT --- CABD --- SAVC --- CoCoOp : ours (RosNet18) : ours (ImageNet) +--- MCNet --- ALICE --- NC-FSCIL --- LDC --- Prompt : ours (CLIP) -- C-FSCL + +![](images/2ccefc01fe4ebb75c0711cafaee7ba4741ba3177f5cea85e0dbfdb0341e3ac06.jpg) + +![](images/b926463328533480f934a14a73b98a04ada6a4075dfa43ee69656dbea10e72bb.jpg) + +Table 1 shows that our ViT model (CLIP) achieves an improvement of $8.42\%$ in the final session compared to the sub-optimal method, and the DR metric decreases by $7.76\%$ . When ImageNet-1K is used for pre-training, our ViT model records a boost of $17.49\%$ in the last session compared to the second best method, with a DR reduction of $13.13\%$ . As for ResNet18, our ResNet18 model exhibits a $1.88\%$ improvement in the final session over the sub-optimal ResNet18 model, with a $1.6\%$ reduction in the DR metric. These results indicate a decrease in catastrophic forgetting and overfitting, attributed to the application of our guiding principles for reducing the approximation risk. + +From Fig. 3, similar trends are observed on the mini-ImageNet and CIFAR100, though the improvements of ViTs are less pronounced. This observation can be attributed to mini-ImageNet being a subset of ImageNet, which violates the standard FSCIL setting, and to the unique image characteristics of CIFAR100, such as smaller image size and lack of background, which differ from the typical natural image distributions used in pre-training. + +Remarkably, even though CLIP has a broader pre-training dataset compared to ImageNet-1K, performance with CLIP remains consistently lower. This phenomenon might be explained by the significant difference between the marginal distributions of the three benchmark datasets and the data used for pre-training of CLIP. Conversely, ImageNet-1K aligns more closely with these distributions, better fulfilling the Homogeneity assumption in our study. + +# 6.3 Ablation Study + +In this section, we assess the impact of four key components of our ViT model on the CUB200 dataset (see supplementary materials for ResNet18). Detailed in Table 2, these components show varied performance effects. The FT strategy exhibits the lowest Top-1 accuracy (38.01% in the final session) and highest DR metric (40.99%), indicating a pronounced bias towards novel classes and substantial consistency risk. Conversely, FR strategy results in improved Top-1 accuracy (52.41%) and lower DR metric (36.63%), yet struggles with catastrophic forgetting. The DS strategy, separating training of base and novel classifiers, is the most effective in reducing bias and consistency risk. + +Table 2: Ablation Study on CUB200 in the 10-Way 5-Shot FSCIL Setting. With ViT-B/16 as the backbone, our study encompasses: "Stabilize", enhancing output stability via tanh function after normalization; "Constrain", limiting representation values using layer normalization and tanh function; "LS" (Label Smoothing) to reduce model overconfidence; and "BC" (Base Classifier), employing Fine-Tuning (FT), Freezing (FR), and Discarding (DS) in the incremental phase. + +
StabilizeConstrainLSBCAccuracy in each session(%)DR↓(%)
012345678910
FT64.4259.3255.0450.6248.9145.2243.9742.1340.5939.5038.0140.99
FR82.7176.3671.3366.9564.8861.0358.9457.4855.7554.1252.4136.63
DS87.0583.5081.0077.6676.5674.1273.0073.1772.3271.8770.9718.47
✓DS87.1584.4181.6578.1177.4675.1874.4473.8773.1772.9772.0717.30
✓DS86.7784.0281.5678.3177.6375.4474.3974.4873.4073.1172.2916.69
✓DS87.4984.0982.1978.5677.9375.4975.5374.6773.4673.4272.7716.82
✓✓DS87.5184.0381.9278.4378.2275.7475.9175.1973.5973.8373.3416.19
✓✓DS87.6384.1381.7778.0578.0075.8075.4074.7974.4374.2573.7115.88
✓✓DS87.4684.5382.7779.3278.9676.6276.8575.8774.5474.4773.9815.41
✓✓✓DS86.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
+ +![](images/d8ec9f2a782e1195ba9107f18dbbe8e0aa0cc9f7a2514f7b533dac72e6608229.jpg) +Fig. 4: Impact of Dataset Size in Transfer Risk with ResNet18 and ViT-S/6 Backbones Trained from Scratch on mini-ImageNet. In both subfigures, the horizontal axis represents the dataset size per base class, while the vertical axis lists training techniques. These include Data Augmentation [1,7,11] and Mixup [18,33], essential for expanding the training dataset. + +Integration of the tanh function for layer stabilization improves the DR metric by about $1\%$ , mitigating overfitting. Similarly, constrained feature representations through layer normalization and tanh function slightly decrease base class accuracy but enhance the DR metric by $1\%$ , indicating reduced overfitting. Implementing label smoothing technique raises both base classification accuracy and DR metric by $1\%$ , attributed to lowering transfer risk. + +Combining these components yields superior performance, highlighting their synergistic effect in FSCIL. + +# 6.4 Dataset Size in Transfer Risk + +In this section, we empirically investigate the correlation between dataset size and transfer risk in neural networks, focusing on ResNet18 and ViT-S/6 architectures. Fig. 4 (a) and (b), show a clear correlation between increased dataset size per class and improved final accuracy for both models. This supports the + +![](images/82d4361764532fca01f019eeacd3aa538b826867a95c02a5b0305909f46722a9.jpg) +Fig. 5: Visualization of ResNet18 and ViT-S/6 on mini-ImageNet. Images are divided by a red dotted line: left for base classes and right for novel classes. Beneath each image, the predicted probability is numerically displayed. The star symbol denotes models in our implementation. In ViT visualizations, image brightness reflects attention level. Darker images signify attention concentrated on specific patches, while brighter ones indicate a more global attention on the whole image. + +idea that increasing the dataset size $(N)$ effectively tightens the upper bound of transfer risk, thereby enhancing models' capacity for incremental learning. + +When direct sample collection is limited, Data Augmentation and Mixup techniques present a viable alternative. These methods create additional samples, enriching the training dataset. This is evident in Figures 4 (a) and (b), where both ResNet18 and ViT-S models show significant performance gains with these techniques. This reinforces our theory: indirect sample generation via Data Augmentation and Mixup can enhance model transferability, mirroring the advantages of direct sample collection. + +# 6.5 Analysis of the Model Overconfidence + +In this section, we provide a visual results to illustrate the influence of prediction confidence on transfer risk. The first and third rows of Fig. 5 indicate that models (CNNs or ViTs) naturally focus on small, highly discriminative regions within base class samples, which leads to overconfidence in these predictions. This propensity leads to misclassification during the transfer of models to novel classes, a phenomenon commonly referred to as supervision collapse. + +Our approach aims to temper the model's over-reliance on base class predictions. As evidenced in the second and fourth rows of Fig. 5, lowering the probability for base class predictions enables the model to cover broader regions. This expanded scope assists in focusing on distinctive features of novel classes, thus enhancing prediction accuracy for these classes. + +![](images/0bcff4c875352c7c80404e9e1894a3291af13784ebe30b895cf89986a6b73c0a.jpg) +(a) -30.84 + +![](images/9eaf6be963d74c27aae2da552a86958a8f7bf1ff25fbf47e372a286d263cf395.jpg) +(b) -15.78 +Fig. 6: The t-SNE visualizations of the logits for test samples on mini-ImageNet, under varying degrees of Classification Margin Discrepancy, are depicted with distinct colors representing different classes. The architectures for Fig. (a), (b), and (c) are based on ViT-S/6 models, each trained from scratch. Conversely, Fig. (d) employs a ViT-B/16 backbone, which is pre-trained on the ImageNet dataset. + +![](images/ac3bf5461c70696b846bc870d8af86972fd1b34d4d7e8b391a6bf796d6eaa651.jpg) +(c)-10.18 + +![](images/aff127abaeaf51443acab0cbe32b056a586180222167a6b33995d30ea6c13f40.jpg) +(d)-3.5 + +# 6.6 Classification Margin Discrepancy + +As demonstrated in Fig. 6, we observe a progressive enhancement in the separability of logits for test samples, correlating with an incremental increase in the Classification Margin Discrepancy. This trend not only signifies a reduction in the consistency risk but also marks a significant step towards achieving more reliable and robust classification performance in FSCIL settings. These findings empirically validate our theoretical framework, suggesting that optimizing for classification margin discrepancy directly contributes to mitigating consistency risks and enhancing model adaptability across incremental learning tasks. + +# 7 Conclusion + +In conclusion, we have conducted a thorough exploration on the approximation risk of FSCIL. By focusing on mitigating transfer and consistency risks and tightening their upper bounds, we have developed practical guidelines for the design and training of FSCIL models. These guidelines have expanded training datasets for base classes, optimized classification margin discrepancy, prevented excessive focus on specific features, and ensured unbiased classification across both base and novel classes. Our extensive experiments have validated these principles, culminating in state-of-the-art performance on three benchmark datasets. + +# 8 Limitation + +In this paper, Homogeneity and Regularity conditions are foundational assumptions. Homogeneity, based on the i.i.d assumption, indicates samples in a dataset share similar characteristics like hues and lighting. Regularity involves boundedness for model convergence and the Lipschitz condition, crucial for representation learning. Practically, comparative experiments show that even when the pretraining and FSCIL benchmark datasets have different marginal distributions, our methods still achieve better results, albeit with smaller gains. This indicates that our theory generally holds but comes at the cost of some practicality. + +# Acknowledgement + +This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160403, by the National Natural Science Foundation of China (NSFC) under Grants 62176178, 62106152, and 62303447, by the Postdoctoral Innovation Talents Support Program under No. BX20230399, and by the China Postdoctoral Science Foundation under No. 2023M743702. + +# References + +1. Balestriero, R., Bottou, L., LeCun, Y.: The effects of regularization and data augmentation are class dependent. In: Advances in Neural Information Processing Systems. vol. 35, pp. 37878-37891 (2022) +2. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023), early Access +3. Doersch, C., Gupta, A., Zisserman, A.: Crosstransformers: spatially-aware few-shot transfer. In: Advances in Neural Information Processing Systems. pp. 21981-21993 (2020) +4. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021) +5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021) +6. Golowich, N., Rakhlin, A., Shamir, O.: Size-independent sample complexity of neural networks. In: Conference On Learning Theory. pp. 297-299. PMLR (2018) +7. Hao, X., Zhu, Y., Appalaraju, S., Zhang, A., Zhang, W., Li, B., Li, M.: Mixgen: A new multi-modal data augmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 379-389 (2023) +8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) +9. Hiller, M., Ma, R., Harandi, M., Drummond, T.: Rethinking generalization in few-shot classification. In: Advances in Neural Information Processing Systems. pp. 1-13 (2022) +10. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing 32, 937-948 (2023) +11. Ji, Z., Jiao, Z., Wang, Q., Pang, Y., Han, J.: Imbalance mitigation for continual learning via knowledge decoupling and dual enhanced contrastive learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-14 (2024), early Access +12. Krizhevsky, A., Nair, V., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009) + +13. Kukleva, A., Kuehne, H., Schiele, B.: Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9029 (2021) +14. Ledoux, M., Talagrand, M.: Probability in Banach Spaces: isoperimetry and processes, vol. 23. Springer Science & Business Media (1991) +15. Liu, B., Yang, B., Xie, L., Wang, R., Tian, Q., Ye, Q.: Learnable distribution calibration for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-8 (2023) +16. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2337-2345 (2021) +17. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022) +18. Pinto, F., Yang, H., Lim, S.N., Torr, P., Dokania, P.: Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In: Advances in Neural Information Processing Systems. vol. 35, pp. 14608-14622 (2022) +19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +20. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015) +21. Shalev-Shwartz, S., Ben-David, S.: Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press (2014) +22. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24183-24192 (June 2023) +23. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020) +24. Tripuraneni, N., Jordan, M., Jin, C.: On the theory of transfer learning: The importance of task diversity. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7852-7862 (2020) +25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30, pp. 1-11 (2017) +26. Wainwright, M.J.: High-dimensional statistics: A non-asymptotic viewpoint, vol. 48. Cambridge university press (2019) +27. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023) +28. Wang, X., Ji, Z., Pang, Y., Yu, Y.: A cognition-driven framework for few-shot class-incremental learning. Neurocomputing 600, 128118 (2024) +29. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011) + +30. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations. pp. 1-13 (2023) +31. Yoon, I.U., Choi, T.M., Lee, S.K., Kim, Y.M., Kim, J.H.: Image-object-specific prompt learning for few-shot class-incremental learning. arXiv preprint arXiv:2309.02833 (2023) +32. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12455-12464 (2021) +33. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) +34. Zhang, J., Liu, L., Silven, O., Pietikainen, M., Hu, D.: Few-shot class-incremental learning: A survey. arXiv preprint arXiv:2308.06764 (2023) +35. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-11 (2021) +36. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (2023) +37. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (June 2023) +38. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022) +39. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) \ No newline at end of file diff --git a/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/images.zip b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80cd2deab7132a6c275aa9b7285af727efcbfc4f --- /dev/null +++ b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15bd5ec8b5143956ae7d71ca427d194eb5eeff5f73d305f16829a2b1ac5c6841 +size 471439 diff --git a/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/layout.json b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f61c800b5c56e4b8af13082eb51e2018a5b9c49b --- /dev/null +++ b/2024/On the Approximation Risk of Few-Shot Class-Incremental Learning/layout.json @@ -0,0 +1,11050 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "text", + "content": "On the Approximation Risk of Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "spans": [ + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "text", + "content": "Xuan Wang" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "text", + "content": ", Zhong Ji" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "text", + "content": ", Xiyao Liu" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "text", + "content": ", Yanwei Pang" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "text", + "content": ", and Jungong Han" + }, + { + "bbox": [ + 135, + 167, + 478, + 192 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "text", + "content": " Tianjin University " + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "text", + "content": " Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "text", + "content": " State Key Laboratory of Robotics, SIA, CAS " + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 155, + 201, + 457, + 225 + ], + "type": "text", + "content": " University of Sheffield" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 251, + 455, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 251, + 455, + 416 + ], + "spans": [ + { + "bbox": [ + 159, + 251, + 455, + 416 + ], + "type": "text", + "content": "Abstract. Few-Shot Class-Incremental Learning (FSCIL) aims to learn new concepts with few training samples while preserving previously acquired knowledge. Although promising performance has been achieved, there remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Therefore, we thoroughly explore the approximation risk of FSCIL, encompassing both transfer and consistency risks. By tightening the upper bounds of these risks, we derive practical guidelines for designing and training FSCIL models. These guidelines include (1) expanding training datasets for base classes, (2) preventing excessive focus on specific features, (3) optimizing classification margin discrepancy, and (4) ensuring unbiased classification across both base and novel classes. Leveraging these insights, we conduct comprehensive experiments to validate our principles, achieving state-of-the-art performance on three FSCIL benchmark datasets. Code is available at https://github.com/xwangrs/Approximation_FSCIL-ECCV2024.git." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "spans": [ + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "text", + "content": "Keywords: Few-Shot Class-Incremental Learning " + }, + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "text", + "content": " Few-Shot Learning " + }, + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "text", + "content": " Class Incremental Learning " + }, + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 426, + 454, + 449 + ], + "type": "text", + "content": " Approximation Risk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 468, + 230, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 230, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 230, + 480 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": "Few-Shot Class-Incremental Learning (FSCIL) [4, 23, 34], an emergent learning paradigm, aims to learn novel classes with limited training data, while preserving the knowledge acquired from base classes. This paradigm faces two primary issues: (1) the catastrophic forgetting of established base class knowledge when integrating new information, and (2) the overfitting to few-shot novel classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "type": "text", + "content": "To effectively address these issues, researchers have proposed various approaches, broadly categorized into three groups [34]: data-based [13, 17, 38], structure-based [23, 32, 36], and optimization-based [2, 4, 35] approaches. Despite their proven effectiveness, there still remains an underexplored aspect regarding the basic statistical principles underlying FSCIL. Central to this gap is the question: What are the key factors in solving FSCIL problems? To answer this question, our research delves into the basic statistical principles of FSCIL, aiming to guide the design and training of FSCIL models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "text", + "content": "* Corresponding author: Zhong Ji, Email: jizhong@tju.edu.cn" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 305, + 213 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 305, + 213 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 305, + 213 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 305, + 213 + ], + "type": "image", + "image_path": "ff4f8134f9ea14acd57e3a7063a331d7a92653bd8d7e1fad08e3312ce5f0bdcc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 216, + 265, + 225 + ], + "lines": [ + { + "bbox": [ + 201, + 216, + 265, + 225 + ], + "spans": [ + { + "bbox": [ + 201, + 216, + 265, + 225 + ], + "type": "text", + "content": "(a) Transfer Risk" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 455, + 213 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 455, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 455, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 455, + 213 + ], + "type": "image", + "image_path": "3446c78168d56bbc9b5a2334ec2bebf819807655c980ad745d92ac29386943db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 216, + 419, + 225 + ], + "lines": [ + { + "bbox": [ + 343, + 216, + 419, + 225 + ], + "spans": [ + { + "bbox": [ + 343, + 216, + 419, + 225 + ], + "type": "text", + "content": "(b) Consistency Risk" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 235, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 131, + 235, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 482, + 302 + ], + "type": "text", + "content": "Fig. 1: The upper bounds of the transfer and consistency risks. In Fig. 1 (a), the horizontal axis represents the number of training samples for base classes, whereas the vertical axis quantifies the magnitude of the transfer risk. Similarly, in Fig. 1 (b), the horizontal axis reflects the classification margin discrepancy, while the vertical axis quantifies the magnitude of the consistency risk. The parameters " + }, + { + "bbox": [ + 131, + 235, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\eta > 0" + }, + { + "bbox": [ + 131, + 235, + 482, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 235, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\kappa > 0" + }, + { + "bbox": [ + 131, + 235, + 482, + 302 + ], + "type": "text", + "content": " serve as dynamic scaling factors, regulating the transfer and consistency risks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 323, + 482, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 482, + 358 + ], + "type": "text", + "content": "We begin with fundamental assumptions: Homogeneity, Regularity conditions, and Realizability. These assumptions, grounded in practical scenarios, form the foundation of our theory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 358, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 490 + ], + "type": "text", + "content": "Based on these assumptions, we formulate the gap between empirically obtained and theoretically optimal FSCIL models as an optimization problem, whose objective function is defined as the Approximation Risk of FSCIL. This risk includes two components: transfer risk and consistency risk. The transfer risk [24] evaluates the efficacy of the shared representations in knowledge transfer from base to novel classes. A low transfer risk indicates effective utilization of base classes knowledge to compensate for the scarcity of novel class samples, thereby preventing overfitting in these novel classes. Meanwhile, the consistency risk measures performance drop when incorporating novel classes, with a low risk signifying robust and stable performance throughout class-incremental tasks, thereby preventing the catastrophic forgetting of base classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 491, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 482, + 514 + ], + "type": "text", + "content": "Leveraging statistical learning theory [6, 14, 21, 26], we explore the upper bounds of transfer and consistency risks, then identifying their determinants2:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 239, + 519, + 481, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 519, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 239, + 519, + 481, + 559 + ], + "type": "interline_equation", + "content": "\\underbrace {\\eta \\cdot \\frac {1}{\\sqrt {N}}} _ {\\text {T r a n s f e r R i s k}} + \\underbrace {\\log \\left(1 + \\frac {\\kappa}{\\exp (\\nu)}\\right)} _ {\\text {C o n s i s t e n c y R i s k}}, \\tag {1}", + "image_path": "e20a05248f688698bc562a8aa8449bec7c3fae3396d6a530880b08c2d4278116.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "text", + "content": " denotes the number of training samples for the base classes; " + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "text", + "content": " indicates the classification margin discrepancy, defined as the infimum of classifier logits for correct classifications minus the supremum of those for incorrect classifications; the parameters " + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 131, + 564, + 482, + 637 + ], + "type": "text", + "content": " are the dynamic scaling factors for transfer and consistency risks, respectively. We will further discuss and elaborate on methods for minimizing these risks by tightening their upper bounds." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 225, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 225, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 225, + 653 + ], + "type": "text", + "content": "1 Proofs in Section 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 653, + 223, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 223, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 223, + 664 + ], + "type": "text", + "content": "2 Proofs in Section 5." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "In Fig. 1 (a), it is evident that increasing the number of training samples for base classes leads to a significant reduction in the transfer risk. Theoretically, this risk could potentially be minimized to a negligible level with a sufficiently large and diverse dataset. However, in practical scenarios, gathering sufficient data for base classes remains a challenge, thereby prompting reliance on foundation models [19,31]. Crucially, for the Homogeneity assumption to hold, the marginal distribution of the pre-trained data should align with that of the base class data. Failing to meet this condition could lead to an increase in the transfer risk." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "text", + "content": "In exploring how " + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "text", + "content": " affects the transfer risk, it becomes imperative to fully comprehend the role of the dynamic scaling factor " + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "text", + "content": ". From a theoretical perspective, " + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "text", + "content": " represents a multivariate polynomial that is related to the maximum activation values across each layer of the neural networks. This could be interpreted as the model's selective attention to specific features. An increase in " + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 212, + 480, + 378 + ], + "type": "text", + "content": " values indicates a marked emphasis on certain features, often at the expense of neglecting others, thereby increasing the transfer risk. This phenomenon is known as the \"supervision collapse\" [3, 9]. To mitigate this issue, we propose to maintain a balanced distribution of activation values across the network layers to prevent extreme impulse values. Furthermore, our theoretical analysis indicates that Vision Transformers (ViTs) [5], owing to their self-attention mechanism [25], exhibit a lower transfer risk compared to Convolutional Neural Networks (CNNs) [8]. Thus, in addressing the FSCIL problem, our preference is towards the deployment of ViT models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 480, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 486 + ], + "type": "text", + "content": "Evidenced in Fig. 1 (b), a clear inverse relationship exists between the classification margin discrepancy and the consistency risk, indicating that enhancing the inter-class discrimination of representations can theoretically diminish the consistency risk to minimal levels. This requires models to maximize the classification margin discrepancy on base classes, ensuring comprehensive training of these models. Moreover, our theory also indicates that the scaling factor " + }, + { + "bbox": [ + 130, + 380, + 480, + 486 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 380, + 480, + 486 + ], + "type": "text", + "content": " reaches its lower bound when the model exhibits unbiased classification across both base and novel classes. Achieving this balance involves training classifiers for each task independently before combining them into a unified classifier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 487, + 339, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 487, + 339, + 498 + ], + "spans": [ + { + "bbox": [ + 147, + 487, + 339, + 498 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 510, + 480, + 628 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 510, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 479, + 533 + ], + "type": "text", + "content": "- Theoretical Framework: We develop a new optimization theory that provides a novel perspective for understanding and addressing challenges in FSCIL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 535, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 570 + ], + "type": "text", + "content": "- Risk Analysis: We decompose the approximation risk of FSCIL into transfer and consistency risks, defining their upper boundaries and and key determinants, thereby offering guidance for model design and training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 571, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 571, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 571, + 479, + 628 + ], + "type": "text", + "content": "- Practical Strategies: We propose practical guidelines to address the challenges in FSCIL. These include expanding training datasets for base classes, preventing excessive focus on specific features, optimizing classification margin discrepancy, and ensuring unbiased classification across both base and novel classes." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Informed by these insights, we achieve state-of-the-art performance on three benchmark datasets, demonstrating the effectiveness of our theory." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 129 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 201 + ], + "type": "text", + "content": "Recently, FSCIL [23,28] has emerged as a paradigm to incrementally learn novel classes with a limited number of samples, while preserving previously learned knowledge. According to the survey [34], existing approaches can be broadly categorized into three groups: data-based approaches [13,17,38], structure-based approaches [23,32,36], and optimization-based approaches [2,4,35]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": "The data-based approaches address FSCIL from the perspective of training data, which involves data replay-based methods [13] and pseudo scenarios-based methods [17, 38]. For example, Kukleva et al. [13] propose a three-stage framework where the base and the novel class samples are efficiently replayed to calibrate the classification bias. Zhou et al. [38] propose ForwArd Compatible Training (FACT) for FSCIL, which preassigns multiple virtual prototypes in the embedding space by creatively mixuping the known class samples, and Peng et al. [17] have similar ideas about utilizing the Mixup [33] to expand the datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 403 + ], + "type": "text", + "content": "The structure-based approaches address FSCIL from the perspective of network architecture design, which involves dynamic structure-based methods [23, 32] and attention-based methods [36]. For example, Zhang et al. [32] propose Continually Evolved Classifier (CEC) that employs a graph model to propagate context information between classifiers for adaptation. Tao et al. [23] propose to use a neural \"gas\" network to learn and preserve the topology of the feature manifold formed by the base and the novel classes in FSCIL. Zhao et al. [36] propose an attention-based aggregation module that selectively merges predictions from the base branch and the novel branch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 404, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 536 + ], + "type": "text", + "content": "The optimization-based approaches address FSCIL from the perspective of algorithm optimization, which involves representation learning-based methods [35] and knowledge distillation methods [2, 4]. Zhao et al. [35] concentrate on the dilemma between the slow forgetting of old knowledge and the fast adaptation to novel knowledge. Then a multi-grained \"slow vs fast\" learning strategy is proposed to cope with this dilemma from both intra-space and the inter-space. Dong et al. [4] propose a exemplar relation distillation incremental learning framework to balance the tasks of old-knowledge preserving and new-knowledge adaptation. Cui et al. [2] propose to distill reliable knowledge from the reference model, then implement an uncertainty-aware distillation module that combines uncertainty-guided knowledge refinement with adaptive distillation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 536, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 572 + ], + "type": "text", + "content": "Beyond the above approaches, we explore the basic statistical principles underlying FSCIL, then conclude with practical guidelines for the design and training of FSCIL models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 591, + 227, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 227, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 227, + 605 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Problem Definition. FSCIL involves a base task followed by a sequence of incremental tasks, each identified by a unique task identifier " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Each task is constituted by a distinct training set, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{train}^{t}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", and a testing set, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{test}^{t}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", which are associated with their respective distinct label spaces " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\forall i,j" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^i \\cap \\mathcal{Y}^j = \\emptyset" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ". For a particular task " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ", training is confined to " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{train}^t" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ", while testing is performed across all encountered tasks. Notably, " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{train}^0" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " encompasses a substantial number of samples for the base task, while " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{train}^{t > 0}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " are limited to " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " samples per task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 479, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 479, + 201 + ], + "type": "text", + "content": "To tackle FSCIL, popular work [17,30-32] assumes a shared feature representation, enabling knowledge transfer from the base to incremental tasks. Building upon this premise, we introduce the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "text", + "content": "Assumption 1 (Homogeneity) In FSCIL, the data pairs " + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_{t,i},y_{t,i})\\}" + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "text", + "content": " for each task " + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "text", + "content": ", are i.i.d. drawn from a distribution " + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_t" + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^t\\times \\mathcal{Y}^t" + }, + { + "bbox": [ + 131, + 205, + 482, + 231 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 227, + 238, + 480, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 238, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 227, + 238, + 480, + 251 + ], + "type": "interline_equation", + "content": "\\mathbb {P} _ {t} (\\mathbf {x}, y) = \\mathbb {P} _ {\\mathbf {x}} (\\mathbf {x}) \\cdot \\mathbb {P} _ {y | \\mathbf {x}} \\left(y \\mid f _ {t} ^ {\\star} \\circ \\mathbf {h} ^ {\\star} (\\mathbf {x})\\right), \\tag {2}", + "image_path": "52141a4e1b24b3bc28af719daa904d1a269700269518a382856f4256c43e9f70.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{h}:\\mathbb{R}^d\\to \\mathbb{R}^r" + }, + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "text", + "content": " denotes a shared feature representation, and " + }, + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "inline_equation", + "content": "f_{t}:\\mathbb{R}^{r}\\rightarrow \\mathbb{R}^{c}" + }, + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "text", + "content": " represents a task-specific mapping. Crucially, this assumption entails that all tasks are subject to an identical marginal distribution " + }, + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{\\mathbf{x}}" + }, + { + "bbox": [ + 131, + 258, + 482, + 307 + ], + "type": "text", + "content": ", indicative of a unified dataset shared across the tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 323, + 380, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 380, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 380, + 338 + ], + "type": "text", + "content": "4 Approximation Risk of FSCIL Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "content": "As explicated in existing work [23,32], the training procedure for FSCIL models is roughly partitioned into two distinct phases: the base and incremental phases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "text", + "content": "Base Phase. During the base phase, the primary objective is the learning of an optimal shared representation " + }, + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "text", + "content": " in conjunction with the base classifier " + }, + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "text", + "content": ". This is accomplished by minimizing the empirical risk " + }, + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "inline_equation", + "content": "\\hat{R}_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 372, + 479, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 219, + 415, + 480, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 219, + 415, + 480, + 446 + ], + "type": "interline_equation", + "content": "\\hat {R} _ {\\text {b a s e}} \\left(f _ {0}, \\mathbf {h}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell \\left(f _ {0} \\circ \\mathbf {h} \\left(\\mathbf {x} _ {0, i}\\right), y _ {0, i}\\right), \\tag {3}", + "image_path": "3f25a5d3532979fe61a17f7d3b7ff5998cd5529de87738c89d5fd4cac44d2c2d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": " symbolizes the loss function; " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,i} \\in \\mathcal{X}_{\\mathrm{train}}^0" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "y_{0,i} \\in \\mathcal{Y}^0" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": ". The composition operator " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": " denotes the consecutive application of functions such that " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "f_0 \\circ \\mathbf{h}(\\mathbf{x}) = f_0(\\mathbf{h}(\\mathbf{x}))" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": ". If Assumption 1 holds and the dataset size " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": " is sufficiently large (" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "N \\gg 1" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": "), the estimator " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "(\\hat{f}_0, \\hat{\\mathbf{h}})" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": ", derived from the minimization of " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\hat{R}_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": ", serves as an effective proxy for the optimal function pair " + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "inline_equation", + "content": "(f_0^\\star, h^\\star)" + }, + { + "bbox": [ + 130, + 453, + 482, + 550 + ], + "type": "text", + "content": ". This base phase is crucial in forming a robust baseline for the learning agent, enabling a seamless transition and effective integration with the novel information presented in subsequent incremental learning tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "content": "Incremental Phase. The incremental training process continues with a sequence of incremental tasks, where the learning agent learns the task-specific classifier " + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "content": " by minimizing empirical risk " + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "inline_equation", + "content": "\\hat{R}_{\\mathrm{novel}}" + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 551, + 482, + 588 + ], + "type": "text", + "content": " new training samples3:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 194, + 594, + 480, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 594, + 480, + 626 + ], + "spans": [ + { + "bbox": [ + 194, + 594, + 480, + 626 + ], + "type": "interline_equation", + "content": "\\hat {R} _ {\\text {n o v e l}} \\left(\\left[ f _ {0}, f _ {t} \\right], \\mathbf {h}\\right) = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\ell \\left(\\left[ f _ {0}, f _ {t} \\right] \\circ \\mathbf {h} \\left(\\mathbf {x} _ {t, m}\\right), y _ {t, m}\\right), \\tag {4}", + "image_path": "76a1442637960a9c9d10a71b79d922204c41d9b3b1044389f64be74218b2fb1b.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": "3 For brevity in our theoretical exposition, we confine this discussion to a single incremental task; nevertheless, the principles delineated here are also applicable to sequential tasks." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "[f_0, f_t]" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " signifies the concatenation of the base and incremental classifiers; " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t,m} \\in \\mathcal{X}_{\\mathrm{train}}^t" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "y_{t,m} \\in \\mathcal{Y}^t" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ". The challenge herein lies in the fact that for " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "M \\ll N" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ", the estimator " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "([\\hat{f}_0, \\hat{f}_t], \\hat{\\mathbf{h}})" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ", obtained from minimizing " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathcal{H}" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " fails to estimate their underlying counterparts " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "([f_0^\\star, f_t^\\star], h^\\star)" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ". To address this limitation, we assess the efficacy of the estimator " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "([\\hat{f}_0, \\hat{f}_t], \\hat{\\mathbf{h}})" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " by its excess risk on the incremental task " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ", defined as the approximation risk of FSCIL:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 208, + 199, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 199, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 208, + 199, + 481, + 213 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\mathrm {A R}} = R _ {\\text {n o v e l}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(\\left[ f _ {0} ^ {\\star}, f _ {t} ^ {\\star} \\right], \\mathbf {h} ^ {\\star}\\right). \\tag {5}", + "image_path": "24ebf5a2672e3a824dfb6aebcfab40f6b0a3c1a3efe7b9932a58f885d13d0249.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 222, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 481, + 247 + ], + "type": "text", + "content": "Approximation Risk. To clarify the approximation risk, we deconstruct " + }, + { + "bbox": [ + 132, + 222, + 481, + 247 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{AR}}=\\varDelta_{\\mathrm{transfer}}+\\varDelta R_{\\mathrm{novel}}^{4}" + }, + { + "bbox": [ + 132, + 222, + 481, + 247 + ], + "type": "text", + "content": ". These components are defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 213, + 255, + 480, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 255, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 213, + 255, + 480, + 269 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\text {t r a n s f e r}} = R _ {\\text {n o v e l}} \\left(\\hat {f} _ {t}, \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(f _ {t} ^ {\\star}, \\mathbf {h} ^ {\\star}\\right), \\tag {6}", + "image_path": "3471ff999a01b694ad9191e88c92ff1f5063f7034239c8ff93fe483750175a64.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 272, + 480, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 272, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 213, + 272, + 480, + 285 + ], + "type": "interline_equation", + "content": "\\Delta R _ {\\text {n o v e l}} = R _ {\\text {n o v e l}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {n o v e l}} \\left(\\hat {f} _ {t}, \\hat {\\mathbf {h}}\\right). \\tag {7}", + "image_path": "287a9db728757e45381e9fca9aa28831252473862c24a01fbe578287022d007a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "text", + "content": "Within this framework, " + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{AR}}" + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "text", + "content": " is deconstructed into two principal components: " + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{transfer}}" + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "inline_equation", + "content": "\\Delta R_{\\mathrm{novel}}" + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "text", + "content": ". The former signifies the transfer risk, assessing the expected prediction risk when transferring knowledge to novel tasks. The latter encapsulates the adaptation risk, which particularly measures the divergence between the composite classifier " + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "inline_equation", + "content": "[f_0, f_t]" + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "text", + "content": " and the task-specific classifier " + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 132, + 295, + 481, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 355, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 481, + 390 + ], + "type": "text", + "content": "Beyond " + }, + { + "bbox": [ + 132, + 355, + 481, + 390 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{AR}}" + }, + { + "bbox": [ + 132, + 355, + 481, + 390 + ], + "type": "text", + "content": ", FSCIL additionally necessitates the consistent performance across incremental tasks. This is formally represented by a constraint encapsulating the adaptation risk pertaining to the base classifier:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 217, + 399, + 480, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 399, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 217, + 399, + 480, + 413 + ], + "type": "interline_equation", + "content": "\\Delta R _ {\\text {b a s e}} = R _ {\\text {b a s e}} \\left(\\left[ \\hat {f} _ {0}, \\hat {f} _ {t} \\right], \\hat {\\mathbf {h}}\\right) - R _ {\\text {b a s e}} \\left(\\hat {f} _ {0}, \\hat {\\mathbf {h}}\\right). \\tag {8}", + "image_path": "72004f3c477a8c5f6546b32df23f9e6fda67e49cd3940a3280bf9eb5e5919ad9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 422, + 481, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 481, + 458 + ], + "type": "text", + "content": "In synthesis, the final version of the approximation risk is represented by Eq. (5), incorporating the constraint given by Eq. (8), and is achieved by applying the penalty method:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 222, + 469, + 480, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 469, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 222, + 469, + 480, + 494 + ], + "type": "interline_equation", + "content": "\\min _ {f _ {0}, f _ {t}, \\mathbf {h}} \\underbrace {\\Delta_ {\\text {t r a n s f e r}}} _ {\\text {T r a n s f e r R i s k}} + \\underbrace {\\Delta R _ {\\text {n o v e l}} + \\gamma \\Delta R _ {\\text {b a s e}}} _ {\\text {C o n s i s t e n c y R i s k}}, \\tag {9}", + "image_path": "bd6be9f543757239a86dfcd8d4834837a82b98a2d54b739516b7102950787c48.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 503, + 318, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 318, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 318, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 503, + 318, + 515 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 503, + 318, + 515 + ], + "type": "text", + "content": " represents the penalty coefficient." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 533, + 377, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 377, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 377, + 547 + ], + "type": "text", + "content": "5 Theoretical Results And Applications" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "text", + "content": "In this section, the basic statistical principles underlying FSCIL are presented. Initially, we make the following standard, mild regularity assumptions on the loss function " + }, + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "text", + "content": ", the function class of tasks " + }, + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "text", + "content": ", and the function class of shared representations " + }, + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 558, + 481, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 614, + 462, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 462, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 462, + 627 + ], + "type": "text", + "content": "Assumption 2 (Regularity conditions) The following conditions hold:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "text", + "content": "- The loss function " + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot, \\cdot)" + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "text", + "content": "-bounded, and " + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot, y)" + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "text", + "content": "-Lipschitz for all " + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 136, + 633, + 481, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "type": "text", + "content": "Proof In Appendix 9.4" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 204, + 114, + 411, + 198 + ], + "blocks": [ + { + "bbox": [ + 204, + 114, + 411, + 198 + ], + "lines": [ + { + "bbox": [ + 204, + 114, + 411, + 198 + ], + "spans": [ + { + "bbox": [ + 204, + 114, + 411, + 198 + ], + "type": "image", + "image_path": "34a1ef1ed210b78c27434966106ce0b4cd6c4b8b6ccaae157a959e4cd320b25c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 480, + 229 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 480, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 480, + 229 + ], + "type": "text", + "content": "Fig.2: Architecture-Agnostic Modifications to FSCIL Models. Specific adjustments are indicated in red." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 254, + 480, + 289 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "spans": [ + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "text", + "content": "- The function " + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "inline_equation", + "content": "\\Gamma(\\mathcal{F})" + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "text", + "content": "-Lipschitz regarding the " + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "text", + "content": " distance for any " + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}" + }, + { + "bbox": [ + 137, + 254, + 475, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "text", + "content": "- The composed function " + }, + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "inline_equation", + "content": "f \\circ \\mathbf{h}" + }, + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "text", + "content": " is bounded: " + }, + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "inline_equation", + "content": "\\sup_{\\mathbf{x} \\in \\mathcal{X}} |f \\circ \\mathbf{h}(\\mathbf{x})| \\leq D_{\\mathcal{X}}" + }, + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "text", + "content": ", for any " + }, + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}, \\mathbf{h} \\in \\mathcal{H}" + }, + { + "bbox": [ + 137, + 266, + 480, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 298, + 479, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 322 + ], + "type": "text", + "content": "Furthermore, the realizability assumption is adopted, positing that the true task functions and the true representation are contained in " + }, + { + "bbox": [ + 130, + 298, + 479, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 298, + 479, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 298, + 479, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 298, + 479, + 322 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "text", + "content": "Assumption 3 (Realizability) The true representation " + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{\\star}" + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "text", + "content": " is encompassed by " + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "text", + "content": ". Concurrently, for all tasks considered—both base and incremental—the true task-specific functions " + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "inline_equation", + "content": "f_{t}^{\\star}" + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "text", + "content": " are also contained in " + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 131, + 332, + 480, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 378, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 479, + 403 + ], + "type": "text", + "content": "Based on Regularity and Realizability assumptions, we analyze transfer learning risk " + }, + { + "bbox": [ + 130, + 378, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{transfer}}" + }, + { + "bbox": [ + 130, + 378, + 479, + 403 + ], + "type": "text", + "content": " and consistency risk " + }, + { + "bbox": [ + 130, + 378, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\varDelta R_{\\mathrm{novel}}+\\gamma\\varDelta R_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 378, + 479, + 403 + ], + "type": "text", + "content": " in Eq. (9) separately." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 421, + 229, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 421, + 229, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 421, + 229, + 431 + ], + "type": "text", + "content": "5.1 Transfer Risk" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "text", + "content": "Theorem 1. (Proof in Appendix 9.2) If Assumptions 2 and 3 hold, and the base task is " + }, + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "inline_equation", + "content": "(\\rho,0)" + }, + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "text", + "content": " diverse, with probability " + }, + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "inline_equation", + "content": "1 - 2\\delta" + }, + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "text", + "content": ", the transfer risk " + }, + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\Delta_{\\text{transfer}}" + }, + { + "bbox": [ + 131, + 440, + 479, + 475 + ], + "type": "text", + "content": " in Eq. (9) is bounded by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 482, + 471, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 482, + 471, + 512 + ], + "spans": [ + { + "bbox": [ + 142, + 482, + 471, + 512 + ], + "type": "interline_equation", + "content": "\\tilde {O} \\Big (\\frac {\\Gamma}{\\rho} \\cdot \\Gamma (\\mathcal {F}) \\cdot 4 D \\prod_ {l = 1} ^ {L} Z (l) \\sqrt {\\frac {\\log 2 J}{N}} + \\frac {\\Gamma D _ {\\mathcal {X} ^ {0}}}{\\rho N ^ {2}} + B \\Big [ \\frac {1}{\\rho} \\cdot \\sqrt {\\frac {\\log (2 / \\delta)}{N}} + \\sqrt {\\frac {\\log (2 / \\delta)}{M}} \\Big ] \\Big),", + "image_path": "d44dfd9251f33d6e44bcb28ff6256fc6e8d55d4a7f96eb6e84aaa95cc026eb5a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": " quantify the number of training samples of the base and incremental task " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": " respectively; " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{x} \\| \\leq D" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": " is the depth of the neural network; " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "Z(l) = \\frac{\\| h^{(l)} \\|_{\\infty}}{\\| h^{(l-1)} \\|_{\\infty}}" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "h^{(l)}" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": " denotes the feature representations of the " + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 518, + 479, + 561 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 568, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 479, + 592 + ], + "type": "text", + "content": "Application. To bridge the basic statistical principles and the practical guidelines, we undertake the subsequent analysis:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": "(1) From Theorem 1, it becomes apparent that the quantity " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": ", representing the number of training samples for the base task, primarily influences the upper bound of the transfer risk. As " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": " increases, the upper bound tightens at a rate proportional to " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "O\\left(\\frac{1}{\\sqrt{N}}\\right)" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": ". This relation shows that expanding training datasets for base classes is the most straightforward and potent strategy for addressing the FSCIL challenge, especially when the initial dataset is insufficiently large." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "In scenarios where expanding the data pool is not feasible, the adoption of a foundation model [19,27] becomes a strategic alternative, which is substantiated by our analysis of the transfer risk." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": "(2) Theoretically, the upper bound of transfer risk is proportional to " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "\\prod_{l=1}^{L} Z(l)" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "Z(l)" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": " is the ratio of the infinity norm of features at layer " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": " to that at layer " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "l-1" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": ". While reducing " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "\\prod_{l=1}^{L} Z(l)" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": " could lower transfer risk, it might impair the feature extraction, so we do not adopt this modification. In original ResNets and ViTs, each layer typically includes Layer or Batch Normalization, standardizing feature means to 0 and variances to 1. However, this normalization does not limit the infinity norm of features, potentially causing excessively large values in some dimensions and increasing " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "Z(l)" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": ", then raising transfer risk. To mitigate this, we constrain each layer's features to " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": ", setting " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "Z(l) = 1" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": ". Since normalization centers most feature values around zero, we introduce the tanh function to cap them. The tanh function bounds values within " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": " and has a slope of 1 at zero, effectively constraining extreme values without significantly impacting original models. Similarly, we also use norm + tanh to constrain the representation " + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 130, + 152, + 483, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "text", + "content": "Besides, " + }, + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "inline_equation", + "content": "D_{\\mathcal{X}^0}" + }, + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "text", + "content": " represents the highest probability value of the predicted outcome, exhibiting a positive correlation with the upper bound. The essential strategy to reduce this bound involves calibrating the model to ensure its predictions are tempered, thereby preventing overconfidence. In conclusion, the analysis of both " + }, + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\prod_{l=1}^{L} Z(l)" + }, + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "inline_equation", + "content": "D_{\\mathcal{X}^0}" + }, + { + "bbox": [ + 130, + 309, + 480, + 392 + ], + "type": "text", + "content": " shows that preventing excessive focus on specific features is an effective strategy for addressing FSCIL challenge. As shown in Fig. 2, we propose three architecture-agnostic modifications:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 397, + 479, + 455 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 397, + 479, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 479, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 479, + 419 + ], + "type": "text", + "content": "- Incorporation of the tanh activation function subsequent to each normalization layer in the original models;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 420, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 479, + 443 + ], + "type": "text", + "content": "- Integration of the layer normalization and tanh activation function prior to the generation of shared representations;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 444, + 392, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 392, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 392, + 455 + ], + "type": "text", + "content": "- Application of the label smooth in the training process." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "text", + "content": "Remark. The term " + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "inline_equation", + "content": "Z(l) = \\frac{\\|h^{(l)}\\|_{\\infty}}{\\|h^{(l - 1)}\\|_{\\infty}}" + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "text", + "content": " quantifies the ratio of maximal activation value in the " + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "text", + "content": "-th layer compared to that in the " + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "inline_equation", + "content": "(l - 1)" + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "text", + "content": "-th layer. It serves as an indicator of the proportional activation dynamics spanning all dimensions between consecutive layers. This metric emphasizes the need for a balanced distribution of activation values throughout the network layers to inhibit the incidence of impulse activation. Such regulation ensures that the model maintains an equitable emphasis on all features, thereby supporting the underlying expectation of " + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "inline_equation", + "content": "D_{\\chi^0}" + }, + { + "bbox": [ + 130, + 460, + 480, + 561 + ], + "type": "text", + "content": " for the model to prevent overconfidence in its predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 565, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 565, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 565, + 481, + 601 + ], + "type": "text", + "content": "Corollary 1 (Proof in Appendix 9.3). Under conditions from Theorem 1, Vision Transformers (ViTs) exhibit lower transfer risks than Convolutional Neural Networks (CNNs)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Remark. This finding substantiates the efficacy of ViTs in leveraging shared representations for novel tasks, underscoring their robustness in transfer learning scenarios. Furthermore, the reduced transfer risk with ViTs as compared to CNNs suggests a pivotal architectural advantage that may inform future developments in neural network design." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 247, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 247, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 247, + 128 + ], + "type": "text", + "content": "5.2 Consistency Risk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "text", + "content": "Theorem 2 (Proof in Appendix 9.5). If Assumptions 2 and 3 hold, and given " + }, + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "inline_equation", + "content": "K_{0}" + }, + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "text", + "content": " classes in the base and incremental tasks respectively, the consistency risk " + }, + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "inline_equation", + "content": "\\Delta R_{\\text{novel}} + \\gamma \\Delta R_{\\text{base}}" + }, + { + "bbox": [ + 132, + 136, + 482, + 172 + ], + "type": "text", + "content": " in Eq. (9) is upper-bounded by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 181, + 480, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 181, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 238, + 181, + 480, + 206 + ], + "type": "interline_equation", + "content": "\\log \\left(\\left(1 + \\frac {K _ {0}}{K _ {t}} v _ {t}\\right) \\left(1 + \\frac {K _ {t}}{K _ {0}} v _ {0}\\right) ^ {\\gamma}\\right), \\tag {10}", + "image_path": "f64ecec6bfce770e063cddf7f9c91b522d17789b17863a6041be3fbd8f2774a1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 213, + 385, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 385, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 385, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 213, + 385, + 226 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 132, + 213, + 385, + 226 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 213, + 385, + 226 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 132, + 213, + 385, + 226 + ], + "type": "text", + "content": " involve supremum and infimum of logits:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 234, + 236, + 480, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 236, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 234, + 236, + 480, + 247 + ], + "type": "interline_equation", + "content": "v _ {t} = \\exp \\left(\\sup \\left(\\phi_ {t \\rightarrow 0}\\right) - \\inf \\left(\\phi_ {t \\rightarrow t}\\right)\\right), \\tag {11}", + "image_path": "bc8df2b3c7d0a3d6e46ec24d96a03f9bc15a0f5f925159cfb8e036b269625624.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 251, + 480, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 251, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 231, + 251, + 480, + 262 + ], + "type": "interline_equation", + "content": "v _ {0} = \\exp \\left(\\sup \\left(\\phi_ {0 \\rightarrow t}\\right) - \\inf \\left(\\phi_ {0 \\rightarrow 0}\\right)\\right). \\tag {12}", + "image_path": "9e30362d1cfc4e41af2595bc71ee23920384192d35ea9ed6eea23d18bb7957e3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "text", + "content": "The classifier logits " + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\phi_{t\\to 0}" + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\phi_{t\\to t}" + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\phi_{0\\to 0}" + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\phi_{0\\to t}" + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "text", + "content": " respectively quantify how samples from task " + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 271, + 482, + 308 + ], + "type": "text", + "content": " and the task 0 (base task) are classified, either correctly or in-correctly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 316, + 481, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 481, + 339 + ], + "type": "text", + "content": "Application. Building on the above analysis about the transfer risk, we proceed with the following detailed analysis about the consistency risk:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 340, + 481, + 483 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 340, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 399 + ], + "type": "text", + "content": "(3) From Theorem 2, we observe that the consistency risk is inherently tied to the classification margin discrepancy of the model, which is quantified by the difference between the infimum of the classifier logits for correct classification and the supremum for incorrect classification. It infers that the consistency risk approaches nullity as the predictive accuracy of the model approaches perfection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "text", + "content": "(4) The upper bound of the consistency risk represents a balance between " + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "text", + "content": ". Introducing a bias in categorization leads to inverse variations in " + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 130, + 399, + 481, + 483 + ], + "type": "text", + "content": ", indicating that ensuring unbiased classification across both base and novel classes can completely nullify the consistency risk. This implies strategically designed biases might reach the minimal risk bound but are not optimally efficient. Key lies in improving classification margin discrepancy while maintaining a balanced classification bias." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 526, + 348, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 348, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 348, + 538 + ], + "type": "text", + "content": "6.1 Datasets and Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Datasets. Our study employs three diverse datasets: mini-ImageNet [20], CIFAR100 [12], and CUB200 [29]. The mini-ImageNet, a subset of ImageNet [20], includes 600 images from 100 classes, split into 60 base and 40 novel classes. These novel classes are divided into 8 incremental tasks, each a 5-way 5-shot task. CIFAR100 comprises 100 classes with 600 images each, partitioned into 60 base and 40 novel classes, the latter organized into 8 incremental 5-way 5-shot tasks. CUB200, containing 11,788 images of 200 bird species, is split into 100 base and 100 novel classes, with the latter arranged into 10 incremental 10-way 5-shot tasks. In this work, when the model is trained from scratch, the image sizes for mini-ImageNet, CIFAR100, and CUB200 are set at " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "84 \\times 84" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 179, + 465, + 300 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Classification Accuracy (\\%) on CUB200 in the 10-Way 5-Shot FS-CIL Setting. The notation ViT-B/16 specifically denotes the ViT-Base model [5] with a designated patch size of 16. The symbols " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " signify models derived from the CLIP [19] and those pre-trained on the ImageNet-1K dataset respectively. The symbol " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " is employed to indicate that lower values are more desirable." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 179, + 465, + 300 + ], + "lines": [ + { + "bbox": [ + 151, + 179, + 465, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 179, + 465, + 300 + ], + "type": "table", + "html": "
MethodBackboneAccuracy in each session(%)DR↓(%)
012345678910
\\( CEC^† \\) [32]ResNet1875.8571.9468.5063.562.4358.2757.7355.8154.8353.5252.2831.07
\\( MCNet^† \\) [10]ResNet1877.5773.9670.4765.8166.1663.8162.0961.8260.4160.0959.0823.84
\\( FACT^† \\) [38]ResNet1875.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9424.98
\\( ALICE^† \\) [17]ResNet1877.472.770.667.265.963.462.961.960.560.660.122.35
\\( CABD^† \\) [37]ResNet1879.1275.6373.2169.9368.3266.3065.1564.9664.2064.0363.8119.35
\\( NC-FSCIL^† \\) [30]ResNet1880.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4426.12
\\( SAVC^† \\) [22]ResNet1881.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5023.64
\\( LDC^† \\) [15]ResNet1877.8976.9374.6470.0668.8867.1564.8364.1663.0362.3961.5820.94
\\( CoCoOp^° \\) [39]ViT-B/1680.377.175.873.972.468.265.263.961.160.357.228.76
\\( Prompt^° \\) [31]ViT-B/1684.383.280.979.277.772.472.169.968.867.466.820.76
\\(ours^{\\dagger}\\)ResNet1879.8777.3674.9270.6270.6468.5568.1467.9366.7366.6765.6917.75
\\(ours^{\\circ}\\)ViT-B/1686.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
\\(ours^{\\dagger}\\)ViT-B/1686.6985.3884.3482.4383.0881.0781.3881.4281.2181.1181.306.22
", + "image_path": "f9642e2ab951a4ac4525b9bf6ca88216180644ed3a91123186dc9781524fe032.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 321, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 481, + 342 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 130, + 321, + 481, + 342 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 321, + 481, + 342 + ], + "type": "text", + "content": " respectively. In scenarios utilizing a foundation model, the images from all three datasets are resized to " + }, + { + "bbox": [ + 130, + 321, + 481, + 342 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 321, + 481, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 488 + ], + "type": "text", + "content": "Implementation Details. Following previous work [23, 34], and guided by our theory, we select ViTs as our primary models across all datasets. Specifically, we utilize two ViT-B/16 variants: one from CLIP [19] and another pre-trained on ImageNet-1K. During the base phase, we train the ViT models on the base classes for 50 epochs using the AdamW optimizer. The learning rate is initially set to 0, then linearly rises to 0.00001 during the first 20 epochs, followed by a linear decrease back to 0 for the rest of the epochs. For a fair comparison, ResNet18 serves as an auxiliary model across datasets, trained on base classes for 500 epochs using AdamW. The learning rate for ResNet18 follows a similar pattern: beginning at 0, increasing linearly to 0.1 within the first 50 epochs, and subsequently diminishing back to 0. In the incremental phase, we freeze the backbone and update only the task-specific classifiers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "text", + "content": "Evaluation Metrics. Following precedents [16, 23], our evaluation employs Top-1 classification accuracy across cumulative testing sets after each learning session. While traditional models often employ the Performance Drop (PD) metric, defined as " + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathrm{PD} = (\\mathcal{A}_0 - \\mathcal{A}_N)" + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_0" + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_N" + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "text", + "content": " represent the accuracies at the initial and final sessions, we propose a refined metric. Recognizing that model forgetting encompasses both the volume of initially acquired knowledge and the extent of subsequent knowledge loss, we introduce the Performance Drop Rate (DR), calculated as " + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathrm{DR} = (\\mathcal{A}_0 - \\mathcal{A}_N) / \\mathcal{A}_0" + }, + { + "bbox": [ + 130, + 489, + 482, + 596 + ], + "type": "text", + "content": ". This metric more accurately captures the effect of knowledge loss, factoring in the initial knowledge base." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 611, + 380, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 380, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 380, + 624 + ], + "type": "text", + "content": "6.2 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "We benchmark our model against state-of-the-art performances on three datasets: CUB200, CIFAR100, and mini-ImageNet, with results in Fig. 3 and detailed CUB200 data in Table 1 (see supplementary materials for other datasets)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 252, + 186 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 252, + 186 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 252, + 186 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 252, + 186 + ], + "type": "image", + "image_path": "86d59389063e912ebcfd5ed177ed4836ec583f921c92379c5e16dbb05206f5d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 191, + 434, + 205 + ], + "lines": [ + { + "bbox": [ + 179, + 191, + 434, + 205 + ], + "spans": [ + { + "bbox": [ + 179, + 191, + 434, + 205 + ], + "type": "text", + "content": "--- CEC --- FACT --- CABD --- SAVC --- CoCoOp : ours (RosNet18) : ours (ImageNet) \n--- MCNet --- ALICE --- NC-FSCIL --- LDC --- Prompt : ours (CLIP) -- C-FSCL" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 216, + 481, + 250 + ], + "lines": [ + { + "bbox": [ + 131, + 216, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 481, + 250 + ], + "type": "text", + "content": "Fig. 3: Comparison with State-of-the-Art on Three Datasets: CUB200, mini-ImageNet, and CIFAR100. For detailed performance metrics, please see Table 1 and refer to our supplementary material." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 116, + 361, + 186 + ], + "blocks": [ + { + "bbox": [ + 253, + 116, + 361, + 186 + ], + "lines": [ + { + "bbox": [ + 253, + 116, + 361, + 186 + ], + "spans": [ + { + "bbox": [ + 253, + 116, + 361, + 186 + ], + "type": "image", + "image_path": "2ccefc01fe4ebb75c0711cafaee7ba4741ba3177f5cea85e0dbfdb0341e3ac06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 361, + 116, + 471, + 186 + ], + "blocks": [ + { + "bbox": [ + 361, + 116, + 471, + 186 + ], + "lines": [ + { + "bbox": [ + 361, + 116, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 361, + 116, + 471, + 186 + ], + "type": "image", + "image_path": "b926463328533480f934a14a73b98a04ada6a4075dfa43ee69656dbea10e72bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "content": "Table 1 shows that our ViT model (CLIP) achieves an improvement of " + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "inline_equation", + "content": "8.42\\%" + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "content": " in the final session compared to the sub-optimal method, and the DR metric decreases by " + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "inline_equation", + "content": "7.76\\%" + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "content": ". When ImageNet-1K is used for pre-training, our ViT model records a boost of " + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "inline_equation", + "content": "17.49\\%" + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "content": " in the last session compared to the second best method, with a DR reduction of " + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "inline_equation", + "content": "13.13\\%" + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "content": ". As for ResNet18, our ResNet18 model exhibits a " + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "inline_equation", + "content": "1.88\\%" + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "content": " improvement in the final session over the sub-optimal ResNet18 model, with a " + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 271, + 482, + 379 + ], + "type": "text", + "content": " reduction in the DR metric. These results indicate a decrease in catastrophic forgetting and overfitting, attributed to the application of our guiding principles for reducing the approximation risk." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": "From Fig. 3, similar trends are observed on the mini-ImageNet and CIFAR100, though the improvements of ViTs are less pronounced. This observation can be attributed to mini-ImageNet being a subset of ImageNet, which violates the standard FSCIL setting, and to the unique image characteristics of CIFAR100, such as smaller image size and lack of background, which differ from the typical natural image distributions used in pre-training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 451, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 523 + ], + "type": "text", + "content": "Remarkably, even though CLIP has a broader pre-training dataset compared to ImageNet-1K, performance with CLIP remains consistently lower. This phenomenon might be explained by the significant difference between the marginal distributions of the three benchmark datasets and the data used for pre-training of CLIP. Conversely, ImageNet-1K aligns more closely with these distributions, better fulfilling the Homogeneity assumption in our study." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 539, + 239, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 239, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 239, + 552 + ], + "type": "text", + "content": "6.3 Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In this section, we assess the impact of four key components of our ViT model on the CUB200 dataset (see supplementary materials for ResNet18). Detailed in Table 2, these components show varied performance effects. The FT strategy exhibits the lowest Top-1 accuracy (38.01% in the final session) and highest DR metric (40.99%), indicating a pronounced bias towards novel classes and substantial consistency risk. Conversely, FR strategy results in improved Top-1 accuracy (52.41%) and lower DR metric (36.63%), yet struggles with catastrophic forgetting. The DS strategy, separating training of base and novel classifiers, is the most effective in reducing bias and consistency risk." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 191, + 463, + 288 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 2: Ablation Study on CUB200 in the 10-Way 5-Shot FSCIL Setting. With ViT-B/16 as the backbone, our study encompasses: \"Stabilize\", enhancing output stability via tanh function after normalization; \"Constrain\", limiting representation values using layer normalization and tanh function; \"LS\" (Label Smoothing) to reduce model overconfidence; and \"BC\" (Base Classifier), employing Fine-Tuning (FT), Freezing (FR), and Discarding (DS) in the incremental phase." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 191, + 463, + 288 + ], + "lines": [ + { + "bbox": [ + 150, + 191, + 463, + 288 + ], + "spans": [ + { + "bbox": [ + 150, + 191, + 463, + 288 + ], + "type": "table", + "html": "
StabilizeConstrainLSBCAccuracy in each session(%)DR↓(%)
012345678910
FT64.4259.3255.0450.6248.9145.2243.9742.1340.5939.5038.0140.99
FR82.7176.3671.3366.9564.8861.0358.9457.4855.7554.1252.4136.63
DS87.0583.5081.0077.6676.5674.1273.0073.1772.3271.8770.9718.47
✓DS87.1584.4181.6578.1177.4675.1874.4473.8773.1772.9772.0717.30
✓DS86.7784.0281.5678.3177.6375.4474.3974.4873.4073.1172.2916.69
✓DS87.4984.0982.1978.5677.9375.4975.5374.6773.4673.4272.7716.82
✓✓DS87.5184.0381.9278.4378.2275.7475.9175.1973.5973.8373.3416.19
✓✓DS87.6384.1381.7778.0578.0075.8075.4074.7974.4374.2573.7115.88
✓✓DS87.4684.5382.7779.3278.9676.6276.8575.8774.5474.4773.9815.41
✓✓✓DS86.4684.2182.6779.7979.5277.3276.7177.1675.7675.5675.2213.00
", + "image_path": "d6c213b29fa5a0c12d1b2f93b8aeda6613b8f13525e52b3cb2d5ad61e6f0c686.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 156, + 298, + 459, + 401 + ], + "blocks": [ + { + "bbox": [ + 156, + 298, + 459, + 401 + ], + "lines": [ + { + "bbox": [ + 156, + 298, + 459, + 401 + ], + "spans": [ + { + "bbox": [ + 156, + 298, + 459, + 401 + ], + "type": "image", + "image_path": "d8ec9f2a782e1195ba9107f18dbbe8e0aa0cc9f7a2514f7b533dac72e6608229.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 409, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 464 + ], + "type": "text", + "content": "Fig. 4: Impact of Dataset Size in Transfer Risk with ResNet18 and ViT-S/6 Backbones Trained from Scratch on mini-ImageNet. In both subfigures, the horizontal axis represents the dataset size per base class, while the vertical axis lists training techniques. These include Data Augmentation [1,7,11] and Mixup [18,33], essential for expanding the training dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "text", + "content": "Integration of the tanh function for layer stabilization improves the DR metric by about " + }, + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "text", + "content": ", mitigating overfitting. Similarly, constrained feature representations through layer normalization and tanh function slightly decrease base class accuracy but enhance the DR metric by " + }, + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "text", + "content": ", indicating reduced overfitting. Implementing label smoothing technique raises both base classification accuracy and DR metric by " + }, + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 487, + 482, + 558 + ], + "type": "text", + "content": ", attributed to lowering transfer risk." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 559, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 583 + ], + "type": "text", + "content": "Combining these components yields superior performance, highlighting their synergistic effect in FSCIL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 598, + 309, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 309, + 610 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 309, + 610 + ], + "type": "text", + "content": "6.4 Dataset Size in Transfer Risk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this section, we empirically investigate the correlation between dataset size and transfer risk in neural networks, focusing on ResNet18 and ViT-S/6 architectures. Fig. 4 (a) and (b), show a clear correlation between increased dataset size per class and improved final accuracy for both models. This supports the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 480, + 289 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 480, + 289 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 480, + 289 + ], + "type": "image", + "image_path": "82d4361764532fca01f019eeacd3aa538b826867a95c02a5b0305909f46722a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "type": "text", + "content": "Fig. 5: Visualization of ResNet18 and ViT-S/6 on mini-ImageNet. Images are divided by a red dotted line: left for base classes and right for novel classes. Beneath each image, the predicted probability is numerically displayed. The star symbol denotes models in our implementation. In ViT visualizations, image brightness reflects attention level. Darker images signify attention concentrated on specific patches, while brighter ones indicate a more global attention on the whole image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 388, + 482, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 412 + ], + "type": "text", + "content": "idea that increasing the dataset size " + }, + { + "bbox": [ + 130, + 388, + 482, + 412 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 130, + 388, + 482, + 412 + ], + "type": "text", + "content": " effectively tightens the upper bound of transfer risk, thereby enhancing models' capacity for incremental learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": "When direct sample collection is limited, Data Augmentation and Mixup techniques present a viable alternative. These methods create additional samples, enriching the training dataset. This is evident in Figures 4 (a) and (b), where both ResNet18 and ViT-S models show significant performance gains with these techniques. This reinforces our theory: indirect sample generation via Data Augmentation and Mixup can enhance model transferability, mirroring the advantages of direct sample collection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 514, + 353, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 353, + 526 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 353, + 526 + ], + "type": "text", + "content": "6.5 Analysis of the Model Overconfidence" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": "In this section, we provide a visual results to illustrate the influence of prediction confidence on transfer risk. The first and third rows of Fig. 5 indicate that models (CNNs or ViTs) naturally focus on small, highly discriminative regions within base class samples, which leads to overconfidence in these predictions. This propensity leads to misclassification during the transfer of models to novel classes, a phenomenon commonly referred to as supervision collapse." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Our approach aims to temper the model's over-reliance on base class predictions. As evidenced in the second and fourth rows of Fig. 5, lowering the probability for base class predictions enables the model to cover broader regions. This expanded scope assists in focusing on distinctive features of novel classes, thus enhancing prediction accuracy for these classes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 292, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 222, + 171 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 222, + 171 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 222, + 171 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 222, + 171 + ], + "type": "image", + "image_path": "0bcff4c875352c7c80404e9e1894a3291af13784ebe30b895cf89986a6b73c0a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 174, + 196, + 183 + ], + "lines": [ + { + "bbox": [ + 159, + 174, + 196, + 183 + ], + "spans": [ + { + "bbox": [ + 159, + 174, + 196, + 183 + ], + "type": "text", + "content": "(a) -30.84" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 114, + 306, + 171 + ], + "blocks": [ + { + "bbox": [ + 223, + 114, + 306, + 171 + ], + "lines": [ + { + "bbox": [ + 223, + 114, + 306, + 171 + ], + "spans": [ + { + "bbox": [ + 223, + 114, + 306, + 171 + ], + "type": "image", + "image_path": "9eaf6be963d74c27aae2da552a86958a8f7bf1ff25fbf47e372a286d263cf395.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 245, + 174, + 283, + 184 + ], + "lines": [ + { + "bbox": [ + 245, + 174, + 283, + 184 + ], + "spans": [ + { + "bbox": [ + 245, + 174, + 283, + 184 + ], + "type": "text", + "content": "(b) -15.78" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 248 + ], + "type": "text", + "content": "Fig. 6: The t-SNE visualizations of the logits for test samples on mini-ImageNet, under varying degrees of Classification Margin Discrepancy, are depicted with distinct colors representing different classes. The architectures for Fig. (a), (b), and (c) are based on ViT-S/6 models, each trained from scratch. Conversely, Fig. (d) employs a ViT-B/16 backbone, which is pre-trained on the ImageNet dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 394, + 171 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 394, + 171 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 394, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 394, + 171 + ], + "type": "image", + "image_path": "ac3bf5461c70696b846bc870d8af86972fd1b34d4d7e8b391a6bf796d6eaa651.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 174, + 369, + 184 + ], + "lines": [ + { + "bbox": [ + 332, + 174, + 369, + 184 + ], + "spans": [ + { + "bbox": [ + 332, + 174, + 369, + 184 + ], + "type": "text", + "content": "(c)-10.18" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 480, + 171 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 480, + 171 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 480, + 171 + ], + "type": "image", + "image_path": "aff127abaeaf51443acab0cbe32b056a586180222167a6b33995d30ea6c13f40.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 174, + 451, + 184 + ], + "lines": [ + { + "bbox": [ + 422, + 174, + 451, + 184 + ], + "spans": [ + { + "bbox": [ + 422, + 174, + 451, + 184 + ], + "type": "text", + "content": "(d)-3.5" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 272, + 332, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 272, + 332, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 332, + 284 + ], + "type": "text", + "content": "6.6 Classification Margin Discrepancy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 388 + ], + "type": "text", + "content": "As demonstrated in Fig. 6, we observe a progressive enhancement in the separability of logits for test samples, correlating with an incremental increase in the Classification Margin Discrepancy. This trend not only signifies a reduction in the consistency risk but also marks a significant step towards achieving more reliable and robust classification performance in FSCIL settings. These findings empirically validate our theoretical framework, suggesting that optimizing for classification margin discrepancy directly contributes to mitigating consistency risks and enhancing model adaptability across incremental learning tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 405, + 220, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 220, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 220, + 418 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 526 + ], + "type": "text", + "content": "In conclusion, we have conducted a thorough exploration on the approximation risk of FSCIL. By focusing on mitigating transfer and consistency risks and tightening their upper bounds, we have developed practical guidelines for the design and training of FSCIL models. These guidelines have expanded training datasets for base classes, optimized classification margin discrepancy, prevented excessive focus on specific features, and ensured unbiased classification across both base and novel classes. Our extensive experiments have validated these principles, culminating in state-of-the-art performance on three benchmark datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 544, + 218, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 218, + 557 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 218, + 557 + ], + "type": "text", + "content": "8 Limitation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this paper, Homogeneity and Regularity conditions are foundational assumptions. Homogeneity, based on the i.i.d assumption, indicates samples in a dataset share similar characteristics like hues and lighting. Regularity involves boundedness for model convergence and the Lipschitz condition, crucial for representation learning. Practically, comparative experiments show that even when the pretraining and FSCIL benchmark datasets have different marginal distributions, our methods still achieve better results, albeit with smaller gains. This indicates that our theory generally holds but comes at the cost of some practicality." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160403, by the National Natural Science Foundation of China (NSFC) under Grants 62176178, 62106152, and 62303447, by the Postdoctoral Innovation Talents Support Program under No. BX20230399, and by the China Postdoctoral Science Foundation under No. 2023M743702." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 263, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 263, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 481, + 297 + ], + "type": "text", + "content": "1. Balestriero, R., Bottou, L., LeCun, Y.: The effects of regularization and data augmentation are class dependent. In: Advances in Neural Information Processing Systems. vol. 35, pp. 37878-37891 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 298, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 481, + 330 + ], + "type": "text", + "content": "2. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023), early Access" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 331, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 364 + ], + "type": "text", + "content": "3. Doersch, C., Gupta, A., Zisserman, A.: Crosstransformers: spatially-aware few-shot transfer. In: Advances in Neural Information Processing Systems. pp. 21981-21993 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "text", + "content": "4. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 399, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 441 + ], + "type": "text", + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 443, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 481, + 464 + ], + "type": "text", + "content": "6. Golowich, N., Rakhlin, A., Shamir, O.: Size-independent sample complexity of neural networks. In: Conference On Learning Theory. pp. 297-299. PMLR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "text", + "content": "7. Hao, X., Zhu, Y., Appalaraju, S., Zhang, A., Zhang, W., Li, B., Li, M.: Mixgen: A new multi-modal data augmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 379-389 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 498, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 531 + ], + "type": "text", + "content": "8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "type": "text", + "content": "9. Hiller, M., Ma, R., Harandi, M., Drummond, T.: Rethinking generalization in few-shot classification. In: Advances in Neural Information Processing Systems. pp. 1-13 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 566, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 566, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 566, + 481, + 597 + ], + "type": "text", + "content": "10. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing 32, 937-948 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 599, + 481, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 481, + 641 + ], + "type": "text", + "content": "11. Ji, Z., Jiao, Z., Wang, Q., Pang, Y., Han, J.: Imbalance mitigation for continual learning via knowledge decoupling and dual enhanced contrastive learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-14 (2024), early Access" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "content": "12. Krizhevsky, A., Nair, V., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "13. Kukleva, A., Kuehne, H., Schiele, B.: Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9029 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "text", + "content": "14. Ledoux, M., Talagrand, M.: Probability in Banach Spaces: isoperimetry and processes, vol. 23. Springer Science & Business Media (1991)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "text", + "content": "15. Liu, B., Yang, B., Xie, L., Wang, R., Tian, Q., Ye, Q.: Learnable distribution calibration for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-8 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "text", + "content": "16. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2337-2345 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 230, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 230, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 230, + 480, + 262 + ], + "type": "text", + "content": "17. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 263, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 263, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 480, + 296 + ], + "type": "text", + "content": "18. Pinto, F., Yang, H., Lim, S.N., Torr, P., Dokania, P.: Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In: Advances in Neural Information Processing Systems. vol. 35, pp. 14608-14622 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 297, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 297, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 133, + 297, + 480, + 341 + ], + "type": "text", + "content": "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "20. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 396 + ], + "type": "text", + "content": "21. Shalev-Shwartz, S., Ben-David, S.: Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 398, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 398, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 398, + 480, + 441 + ], + "type": "text", + "content": "22. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24183-24192 (June 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 442, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 442, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 133, + 442, + 480, + 474 + ], + "type": "text", + "content": "23. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 480, + 508 + ], + "type": "text", + "content": "24. Tripuraneni, N., Jordan, M., Jin, C.: On the theory of transfer learning: The importance of task diversity. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7852-7862 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "text", + "content": "25. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30, pp. 1-11 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 564 + ], + "type": "text", + "content": "26. Wainwright, M.J.: High-dimensional statistics: A non-asymptotic viewpoint, vol. 48. Cambridge university press (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 566, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 480, + 609 + ], + "type": "text", + "content": "27. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "28. Wang, X., Ji, Z., Pang, Y., Yu, Y.: A cognition-driven framework for few-shot class-incremental learning. Neurocomputing 600, 128118 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "29. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "X. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 446 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations. pp. 1-13 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "31. Yoon, I.U., Choi, T.M., Lee, S.K., Kim, Y.M., Kim, J.H.: Image-object-specific prompt learning for few-shot class-incremental learning. arXiv preprint arXiv:2309.02833 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "32. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12455-12464 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "text", + "content": "33. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "34. Zhang, J., Liu, L., Silven, O., Pietikainen, M., Hu, D.: Few-shot class-incremental learning: A survey. arXiv preprint arXiv:2308.06764 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "35. Zhao, H., Fu, Y., Kang, M., Tian, Q., Wu, F., Li, X.: Mgmt: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1-11 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "text", + "content": "36. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "type": "text", + "content": "37. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11838-11847 (June 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "38. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9046-9056 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 446 + ], + "type": "text", + "content": "39. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Approximation Risk of FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_content_list.json b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee9fe266e30c48e204b7198224d65c4fff5a8d35 --- /dev/null +++ b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_content_list.json @@ -0,0 +1,1975 @@ +[ + { + "type": "text", + "text": "On the Error Analysis of 3D Gaussian Splitting and an Optimal Projection Strategy", + "text_level": 1, + "bbox": [ + 227, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Letian Huang $^{\\text{®}}$ , Jiayang Bai $^{\\text{®}}$ , Jie Guo $^{\\star \\text{®}}$ , Yuanqi Li $^{\\text{®}}$ , and Yanwen Guo $^{\\text{®}}$", + "bbox": [ + 230, + 212, + 772, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nanjing University, Nanjing, China \n{1thuang, jybai, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn \nhttps://letianhuang.github.io/op43dgs/", + "bbox": [ + 225, + 239, + 777, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D Gaussian Splatting has garnered extensive attention and application in real-time neural rendering. Concurrently, concerns have been raised about the limitations of this technology in aspects such as point cloud storage, performance, and robustness in sparse viewpoints, leading to various improvements. However, there has been a notable lack of attention to the fundamental problem of projection errors introduced by the local affine approximation inherent in the splatting itself, and the consequential impact of these errors on the quality of photo-realistic rendering. This paper addresses the projection error function of 3D Gaussian Splatting, commencing with the residual error from the first-order Taylor expansion of the projection function. The analysis establishes a correlation between the error and the Gaussian mean position. Subsequently, leveraging function optimization theory, this paper analyzes the function's minima to provide an optimal projection strategy for Gaussian Splatting referred to Optimal Gaussian Splatting, which can accommodate a variety of camera models. Experimental validation further confirms that this projection methodology reduces artifacts, resulting in a more convincingly realistic rendering.", + "bbox": [ + 261, + 320, + 743, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D deep learning $\\cdot$ view synthesis $\\cdot$ radiance fields $\\cdot$ 3D gaussians $\\cdot$ real-time rendering $\\cdot$ error analysis", + "bbox": [ + 261, + 583, + 743, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 375, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 3D scenes from 2D images and synthesizing novel views has been a critical task in computer vision and graphics. Recently, the trend of novel view synthesis is spearheaded by Neural Radiance Fields (NeRFs) [25] and its variants [1, 2], which achieves photo-realistic rendering quality. NeRF represents scenes with Multi-Layer Perceptron (MLP) and renders novel views using volumetric ray-marching. However, these MLP-based methods contain a relatively large number of layers, leading to lengthy training and rendering times, which still fall significantly short of real-time rendering. To accelerate rendering,", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77cd6b753a96e723145c87409706006a12e725f4bc39dbe589905d0347c0d3ee.jpg", + "image_caption": [ + "Fig. 1: By minimizing the projection error through error analysis, we have achieved an improvement in the rendering image quality compared to the original 3D-GS [16]. These images are cropped from the complete images of the TRAIN scene, in order to be highlighted." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 405, + 289 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/297487479d90a6b06985f07bd69cf50747d4c68ffcc67fc9eaa3b04d6a8b7098.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 143, + 596, + 289 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c66b8cd20ba5787cd72de9697c734ef0fd46c516a26b98b8a718c4f146f6237.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 143, + 784, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "many methods resort to auxiliary data structures, such as hash grids [26], points or voxel [8,36,37,39]. However, due to the dense sampling of points along rays, they still face challenges in achieving real-time performance.", + "bbox": [ + 212, + 386, + 782, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With a substantial demand for high-speed and photo-realistic rendering effects, 3D Gaussian Splatting (3D-GS) [16] has emerged as an efficient representation that can be rendered at high speed on the GPU. It has departed from the implicit scene representation using MLPs and instead opted for an explicit representation using Gaussian functions. As a result, 3D-GS avoids the expensive cost of densely sampling points in volume rendering, thus achieving real-time performance. Given a set of images with the camera parameters calibrated by SfM [31], 3D-GS aims to optimize a set of 3D Gaussians as graphical primitives to explicitly represent the scene. To render 2D images, these 3D Gaussians are projected onto the image plane ( $z = 1$ plane) via a projection function denoted as $\\phi$ for differentiable rasterization.", + "bbox": [ + 212, + 431, + 787, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, Gaussian functions are closed under affine transformations but not under projective transformations. Therefore, 3D-GS adopts a local affine approximation [44], specifically approximating the projection function with the first two terms of its Taylor expansion. Nevertheless, approximations introduce errors, and the local affine approximation similarly contributes to these errors which may lead to artifacts in the rendered images. This paper exploits the relationship between the error of 3D-GS and the Gaussian mean through the analysis of the Taylor remainder term. Additionally, it identifies the circumstances under which the error is minimized by determining the extremum of the error function.", + "bbox": [ + 212, + 598, + 787, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, based on the extremum analysis of the error function, we propose a novel optimal projection method to minimize the projection error in 3D-GS. This method requires only minor code modifications, does not affect real-time rendering performance, and yet achieves a significant improvement in rendering quality. Specifically, we project along the direction from the Gaussian mean to the camera center, where the projection plane is tangent to the line connecting the Gaussian mean and the camera center. Compared with the straightforward", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "approach of projecting 3D Gaussians onto the image plane of the camera, the proposed projection reduces expected error while maintaining real-time and realistic rendering performance. Furthermore, we find that the projection errors in 3D-GS tend to escalate with the continuous expansion of the field of view and the corresponding reduction in focal length. This increasing error leads to more and larger elongated Gaussians, as well as cloud-like Gaussian artifacts, consequently degrading the overall image quality. In contrast, our method exhibits greater realism and robustness. As illustrated in Figure 1, our method significantly reduces blurriness and artifacts caused by the projection error. Through experiments on various datasets and scenes, our method consistently outperforms the original projection method, particularly in settings with short focal lengths. And the method is easily implemented and adaptable to various camera models. To summarize, we provide the following contributions:", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide thorough error analysis for 3D-GS that may lead to artifacts and consequently degrade rendering quality. We have identified the correlation between this error and the Gaussian position.", + "- We derive the mathematical expectation of this error function and analyze when this function takes extrema through methods of function optimization.", + "- We propose a novel optimal projection capable of adapting to diverse camera models that adopts different tangent plane projections based on the directions from each Gaussian mean to the camera center instead of naively projecting all Gaussians onto the same plane" + ], + "bbox": [ + 225, + 347, + 784, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 500, + 387, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Novel view synthesis is a longstanding challenge in computer vision and graphics. From traditional techniques [3,5,10,11,19,31,33], to neural network-based scene representations [9, 15, 18, 24, 28, 29, 32], various approaches have struggled to address the problem of synthesizing a new view from captured images.", + "bbox": [ + 212, + 527, + 787, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 215, + 607, + 455, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Neural Radiance Field (NeRF) [25] stands out as a successful neural rendering method based on MLPs, primarily owing to its encoding of position and direction. This encoding allows for effective reconstruction of high-frequency information in scenes. Notable improvements on this encoding have been made by MipNeRF [1], NeRF-W [22], FreeNeRF [41], and Instant NGP [26]. These enhancements enable the handling of multi-resolution image inputs, multi-illumination with occlusion image inputs, sparse-view inputs, and achieve nearly real-time rendering capabilities, respectively. Barron et al. introduced MipNeRF360 [2] as an extension of MipNeRF [1] to address the issue of generating low-quality renderings for unbounded scenes in NeRF. Importantly, these methods thoroughly exploit the intrinsic capabilities of NeRF as an implicit scene representation without introducing additional model priors.", + "bbox": [ + 212, + 628, + 787, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, due to the implicit representation of scenes and the dense sampling of points along rays, they still face challenges in achieving real-time performance.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 215, + 146, + 449, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3D-GS) [16] has demonstrated notable advancements in rendering performance by departing from MLPs and ray sampling, opting instead for Gaussian functions and Gaussian Splatting. This paradigm shift has attracted considerable attention within the industry, leading to various studies building upon this methodology [6].", + "bbox": [ + 212, + 169, + 787, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D-GS finds application in SLAM systems [23], scene editing [7] and segmentation [4]. Efforts [20] have also been made to integrate 3D-GS with popular diffusion models [14, 34, 35] and dynamic scenes [21]. Recent endeavors aim to enhance the robustness in sparse-view scenarios [38, 40], performance, storage efficiency [27] and mesh reconstruction [12] of 3D-GS. However, these improvements do not specifically address errors associated with Gaussian projection. Potential methods for enhancing Gaussian functions include [30, 42]. Nevertheless, they introduce a Gaussian filter and a novel Gaussian depth sorting algorithm, respectively, without investigating the projection function and associated errors. These approaches still rely on projecting onto the $z = 1$ plane.", + "bbox": [ + 212, + 244, + 787, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper aims to explore the potential of 3D Gaussian Splatting by analyzing errors that may arise during the projection process. The analysis will delve into the factors contributing to these errors and propose methods to mitigate and reduce them.", + "bbox": [ + 212, + 396, + 787, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 478, + 382, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to NeRF [25], the input to 3D-GS consists of a set of images, together with the corresponding cameras calibrated by SFM [31]. However, in contrast to NeRF, 3D-GS takes the sparse point cloud generated during the SFM calibration process as input. From these points, it constructs a set of 3D Gaussians as graphical primitives explicitly representing the scene. Each 3D Gaussian is characterized by its position (mean) $\\pmb{\\mu}$ , covariance matrix $\\pmb{\\Sigma}$ , and opacity $\\alpha$ , and carries anisotropic spherical harmonics (SH) representing directional appearance component (color) of the radiance field, following standard practice [8,26]. Subsequently, these 3D Gaussians are projected onto the image plane ( $z = 1$ plane) via a projection function for differentiable rasterization. Next, we will provide a detailed explanation of the Gaussian projection process in 3D-GS.", + "bbox": [ + 212, + 507, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the world space, we define a 3D Gaussian function by $G$ , characterized by its mean $\\mu$ and covariance matrix $\\pmb{\\Sigma}$ . To project the Gaussian onto the image plane, the initial step in 3D-GS involves an affine transformation of this Gaussian function from the world space to the camera space via a viewing transformation matrix $\\mathbf{W}$ , yielding a transformed Gaussian function:", + "bbox": [ + 212, + 674, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (\\mathbf {x}) = \\exp \\left\\{- 1 / 2 (\\mathbf {W x} - \\mathbf {W} \\boldsymbol {\\mu}) ^ {\\top} (\\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top}) ^ {- 1} (\\mathbf {W x} - \\mathbf {W} \\boldsymbol {\\mu}) \\right\\}. \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 758, + 785, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote $G'$ as the 3D Gaussian in the camera space, with $\\mathbf{x}' = \\mathbf{W}\\mathbf{x}$ , $\\boldsymbol{\\mu}' = \\mathbf{W}\\boldsymbol{\\mu}$ and $\\boldsymbol{\\Sigma}' = \\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^\\top$ as the point, mean and covariance matrix in the camera space, respectively.", + "bbox": [ + 212, + 791, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, traditional 3D-GS requires projecting the Gaussian function in the camera space onto the $z = 1$ plane, with the projection function $\\varphi$ defined as:", + "bbox": [ + 212, + 146, + 782, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x _ {0}}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 189, + 785, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x_0} = [0,0,1]^\\top$ represents the projection of the camera space's origin onto this plane. Expanding this projection function to the first order using Taylor series yields:", + "bbox": [ + 212, + 224, + 782, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) = \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) + \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) \\left(\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}\\right) + R _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) \\tag {3} \\\\ \\approx \\varphi (\\boldsymbol {\\mu} ^ {\\prime}) + \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} (\\boldsymbol {\\mu} ^ {\\prime}) (\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 286, + 784, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\frac{\\partial\\varphi}{\\partial\\mathbf{x}'}\\left(\\boldsymbol{\\mu}'\\right)$ is the Jacobian of the affine approximation of the projective transformation, denoted as $\\mathbf{J}$ . Applying this local affine approximation by neglecting the Taylor remainder term allows to derive the 2D Gaussian function $G_{2\\mathrm{D}}$ projected onto the $z = 1$ plane as:", + "bbox": [ + 212, + 358, + 782, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} G _ {2 D} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\exp \\left\\{- 1 / 2 \\left(\\mathbf {J} \\mathbf {x} ^ {\\prime} - \\mathbf {J} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {\\top} \\left(\\mathbf {J} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} ^ {\\top}\\right) ^ {- 1} \\left(\\mathbf {J} \\mathbf {x} ^ {\\prime} - \\mathbf {J} \\boldsymbol {\\mu} ^ {\\prime}\\right) \\right\\} \\\\ \\approx \\exp \\left\\{- 1 / 2 \\left(\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) ^ {\\top} \\left(\\mathbf {J} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} ^ {\\top}\\right) ^ {- 1} \\left(\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) \\right\\}. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 438, + 784, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the rank of the matrix $\\mathbf{J}$ is 2, the inverse of $\\mathbf{J} \\pmb{\\Sigma}' \\mathbf{J}^\\top$ is, in fact, the inverse of the covariance matrix of the 2D Gaussian obtained by skipping the third row and column [44]. We denote $\\mathbf{x}_{2\\mathrm{D}} = \\varphi(\\mathbf{x}')$ , $\\pmb{\\mu}_{2\\mathrm{D}} = \\varphi(\\pmb{\\mu}')$ and $\\pmb{\\Sigma}_{2\\mathrm{D}} = \\mathbf{J} \\pmb{\\Sigma}' \\mathbf{J}^\\top$ as the point, mean and $2 \\times 2$ covariance matrix of $G_{2\\mathrm{D}}$ in the image space, respectively. For a pixel $(u, v)$ , we obtain the value of the 2D Gaussian function by querying $\\mathbf{x}_{2\\mathrm{D}} = [u, v]^\\top$ , and then perform alpha blending to derive the color of the pixel. The blue box in Figure 3 illustrates the projection process.", + "bbox": [ + 212, + 536, + 782, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Local Affine Approximation Error", + "text_level": 1, + "bbox": [ + 214, + 678, + 578, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From Equation 3 and Equation 4 we observe that 3D Gaussian Splatting introduces an approximation during the projection transformation, i.e. the local affine approximation. Consequently, 3D-GS actually adopts a 2D Gaussian function that is not the true projection of the initial 3D Gaussian. The error introduced by this approximation can be characterized by the Taylor remainder term in Equation 3:", + "bbox": [ + 212, + 712, + 782, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) - \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) \\left(\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 816, + 784, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 146, + 264, + 159 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\mathbb {I} \\otimes \\left(\\mathbf {x} _ {0} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} - \\boldsymbol {\\mu} ^ {\\prime} \\left(\\mathbf {x} _ {0} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\mathbf {x} _ {0}\\right) ^ {- 1} \\mathbf {x} _ {0} ^ {\\top}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 172, + 785, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbb{I}$ represents the identity matrix and $\\otimes$ denotes the multiplication of a matrix and a scalar. Clearly, this Taylor remainder term is a three-dimensional vector related to the random variable $\\mathbf{x}^{\\prime}$ and the mean $\\pmb{\\mu}^{\\prime}$ of the Gaussian function. Therefore, by computing the square of the Frobenius norm of this vector $\\| R_1(\\mathbf{x}')\\| _F^2$ and taking the mathematical expectation of this norm function with respect to the random variable $\\mathbf{x}^{\\prime}$ , we obtain an error function that depends solely on the Gaussian mean $\\pmb{\\mu}^{\\prime}$ :", + "bbox": [ + 215, + 210, + 787, + 323 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\int_ {\\mathbf {x} ^ {\\prime} \\in \\mathcal {X} ^ {\\prime}} \\| R _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) \\| _ {F} ^ {2} \\mathrm {d} \\mathbf {x} ^ {\\prime}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 335, + 785, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The simplified error function $\\epsilon$ involves three unit vectors: $\\mathbf{x}^{\\prime}$ , $\\mathbf{x_0}$ and $\\pmb{\\mu}^{\\prime}$ . Their spherical coordinates are given as follows:", + "bbox": [ + 215, + 392, + 785, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathbf {0}} = \\left[ \\begin{array}{c} \\sin (\\phi_ {0}) \\cos (\\theta_ {0}) \\\\ - \\sin (\\theta_ {0}) \\\\ \\cos (\\phi_ {0}) \\cos (\\theta_ {0}) \\end{array} \\right], \\mathbf {x} ^ {\\prime} = \\left[ \\begin{array}{c} \\sin (\\phi) \\cos (\\theta) \\\\ - \\sin (\\theta) \\\\ \\cos (\\phi) \\cos (\\theta) \\end{array} \\right], \\boldsymbol {\\mu} ^ {\\prime} = \\left[ \\begin{array}{c} \\sin (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu}) \\\\ - \\sin (\\theta_ {\\mu}) \\\\ \\cos (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu}) \\end{array} \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 435, + 785, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi_0 = 0$ and $\\theta_0 = 0$ . Please refer to the detailed proofs in the supplementary materials.", + "bbox": [ + 215, + 494, + 785, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The size of the integration region is related to the covariance of the Gaussian function. Since the current discussion specifically focuses on the impact of the mean on the error rather than the variance, let's assume the integration region to be $\\left\\{\\mathbf{x}' \\mid \\theta \\in [-\\pi/4 + \\theta_{\\mu}, \\pi/4 + \\theta_{\\mu}] \\land \\phi \\in [-\\pi/4 + \\phi_{\\mu}, \\pi/4 + \\phi_{\\mu}]\\right\\}$ . Then, based on Equation 5-8, we obtain:", + "bbox": [ + 215, + 525, + 785, + 606 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon (\\theta_ {\\mu}, \\phi_ {\\mu}) = \\int_ {- \\pi / 4 + \\theta_ {\\mu}} ^ {\\pi / 4 + \\theta_ {\\mu}} \\int_ {- \\pi / 4 + \\phi_ {\\mu}} ^ {\\pi / 4 + \\phi_ {\\mu}} \\left(- \\frac {\\sin (\\phi - \\phi_ {\\mu}) \\cos (\\theta)}{\\cos^ {2} (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu})} + \\tan (\\phi) - \\tan (\\phi_ {\\mu})\\right) ^ {2} + \\\\ \\left(\\frac {\\sin (\\theta)}{\\cos (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu})} - \\frac {\\sin (\\theta_ {\\mu}) \\cos (\\phi) \\cos (\\theta)}{\\cos^ {2} (\\phi_ {\\mu}) \\cos^ {2} (\\theta_ {\\mu})} + \\frac {\\tan (\\theta_ {\\mu})}{\\cos (\\phi_ {\\mu})} - \\frac {\\tan (\\theta)}{\\cos (\\phi)}\\right) ^ {2} d \\theta d \\phi . \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 619, + 800, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fortunately, this integral has a close-form expression, but its form is very complicated. We report the close-form expression in the supplemental materials and visually depict it in Figure 2.", + "bbox": [ + 215, + 722, + 785, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The error function is partially differentiated with respect to $\\theta_{\\mu}$ and $\\phi_{\\mu}$ . It is easy to check that at $\\theta_{\\mu} = \\theta_0 = 0$ and $\\phi_{\\mu} = \\phi_0 = 0$ , the following holds:", + "bbox": [ + 215, + 768, + 785, + 799 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\epsilon}{\\partial \\theta_ {\\mu}} (0, 0) = 0, \\frac {\\partial \\epsilon}{\\partial \\phi_ {\\mu}} (0, 0) = 0. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 811, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b6c085bd2f8652cdb99fcf863d9d73ed435b3c42003556c3f27351eb39cf9b3.jpg", + "image_caption": [ + "(a) $\\lambda = 0.095$" + ], + "image_footnote": [], + "bbox": [ + 289, + 148, + 480, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/361d2ffa91e732c015aff3b5487d1082d3f5f19b2e2fb35714ee61dc1642f9f4.jpg", + "image_caption": [ + "(b) $\\lambda = 0.95$", + "Fig. 2: Visualization of the 3D Gaussian Splitting error function $\\epsilon (\\theta_{\\mu},\\phi_{\\mu})$ under two distinct domains. The domain of the function is $\\{(\\theta_{\\mu},\\phi_{\\mu})\\mid \\theta_{\\mu}\\in [-\\lambda \\pi /4,\\lambda \\pi /4]\\land \\phi_{\\mu}\\in [-\\lambda \\pi /4,\\lambda \\pi /4]\\}$ where $(\\theta_{\\mu},\\phi_{\\mu})$ represents the polar coordinate of the 3D Gaussian mean and $\\lambda$ represents the scaling factor of the domain." + ], + "image_footnote": [], + "bbox": [ + 537, + 150, + 728, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, this point is actually an extremum point of the function. We also confirm that this point is the minimum of the function. Indeed, from Figure 2a, it is also apparent that the function is concave, with its minimum value occurring at $(0,0)$ . Additionally, it can be observed that the minimum value of the error function is greater than zero, due to the fact that Gaussian functions are not closed under projection transformations. Figure 2b shows that in the majority of the region close to the origin, the function values vary very gently. However, as the independent variable of the function approaches the boundaries, the function values increase rapidly, resulting in a substantial difference between the maximum and minimum values.", + "bbox": [ + 212, + 422, + 787, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above analysis indicates that the error is small and does not significantly affect the quality of rendered images. Therefore, traditional 3D-GS [16] utilizes local affine approximation but still successfully reconstructs the scene to obtain a high-quality novel view image. Nevertheless, the naive projection of all Gaussians onto the same plane $z = 1$ in 3D-GS may lead to larger projection errors for Gaussians farther from the plane center (the projection point of the camera center to the plane), causing artifacts. These artifacts are severe when using wide-angle lenses, significantly degrading the quality of the renderings.", + "bbox": [ + 212, + 574, + 787, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Optimal Gaussian Splatting", + "text_level": 1, + "bbox": [ + 214, + 717, + 522, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the analysis of the error function, we introduce Optimal Gaussian Splitting shown in Figure 3. This helps us to achieve smaller projection errors, resulting in higher-quality rendering. The first step, similar to the original 3D-GS [16], involves transforming world coordinates into the camera coordinate system. Instead of projecting Gaussians onto the $z = 1$ plane, each Gaussian is radially projected based on its mean along the line connecting it to the camera", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9916bebfb6385cf294cad8405418c9dfc2eb84de32f010cd889c9549cf88c2c2.jpg", + "image_caption": [ + "Fig. 3: Illustration of the rendering pipeline for our Optimal Gaussian Splitting and the projection of 3D-GS [16]. The blue box depicts the projection process of the original 3D-GS, which straightforwardly projects all Gaussians onto the same projection plane. In contrast, the red box illustrates our approach, where we project individual Gaussians onto corresponding tangent planes." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 393, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ab75e9dba7ebfc00a5f7e92503697fbeb0b2589e5f28bcd52833218da8eb61f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 146, + 785, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "center, projecting onto a plane tangent to the unit sphere and perpendicular to this line. After projection, colors of points on the unit sphere are computed by alpha blending the 2D Gaussians on the tangent plane. Finally, for each pixel in the image, rays are cast onto the unit sphere to retrieve the color of that pixel, resulting in the rendered image.", + "bbox": [ + 212, + 383, + 787, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Optimal Projection", + "text_level": 1, + "bbox": [ + 214, + 479, + 423, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We discover that the error function attains its minimum value when the projection of the Gaussian mean on the plane coincides with the projection from the camera center to the plane. Therefore, in Optimal Gaussian Splatting, we employ an optimal projection method. Specifically, instead of naively projecting different Gaussians onto the same plane $z = 1$ , we adopt distinct projection planes for each Gaussian. These projection planes are determined by the tangent planes formed by the Gaussian mean and the line connecting it to the camera center. The specific projection plane is formulated as:", + "bbox": [ + 212, + 501, + 787, + 622 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\cdot \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x} _ {\\mathbf {p}}\\right) = 0 \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 628, + 785, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 214, + 660, + 264, + 672 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathbf {p}} = \\varpi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\boldsymbol {\\mu} ^ {\\prime} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1 / 2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 671, + 785, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "represents the projection of the camera space's origin onto this plane and $\\varpi$ function represents the transformation that projects points onto the unit sphere. According to Equation 11, the optimal projection function $\\varphi_{\\mathrm{p}}$ is obtained as:", + "bbox": [ + 212, + 704, + 787, + 750 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {\\mathrm {p}} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} _ {\\mathbf {p}}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 755, + 785, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The corresponding local affine approximation Jacobian matrix $\\mathbf{J}_{\\mathbf{p}}$ is:", + "bbox": [ + 214, + 790, + 712, + 806 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} _ {\\mathbf {p}} = \\frac {\\partial \\varphi_ {\\mathrm {p}}}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\mathbb {I} \\otimes \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} - \\boldsymbol {\\mu} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\mathbf {x} _ {\\mathbf {p}}\\right) ^ {- 1} \\mathbf {x} _ {\\mathbf {p}} ^ {\\top}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 813, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When $\\pmb{\\mu}^{\\prime} = [\\mu_{x},\\mu_{y},\\mu_{z}]^{\\top}$ , the specific form of this Jacobian matrix is:", + "bbox": [ + 215, + 143, + 717, + 165 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} _ {\\mathbf {p}} = \\left[ \\begin{array}{l l l} \\frac {\\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {x} \\mu_ {y}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {x} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {y}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & \\frac {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {y} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {y} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\cfrac {3}{2}}} & \\frac {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\cfrac {3}{2}}} \\end{array} \\right]. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 176, + 785, + 266 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "It's very easy to implement modifications to the Jacobian matrix in the forward process.", + "bbox": [ + 214, + 275, + 782, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Unit Sphere Based Rasterizer", + "text_level": 1, + "bbox": [ + 215, + 328, + 506, + 343 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the proposed Optimal Projection, we have proposed a Unit Sphere Based Rasterizer to generate images. Through Optimal Projection, we obtain the projection of the three-dimensional Gaussians on the tangent plane of the unit sphere instead of obtaining the Gaussians in the image space. Therefore, we need to rasterize based on this unit sphere.", + "bbox": [ + 212, + 353, + 782, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, for a pixel $(u,v)$ on the image, similar to NeRF, we cast a ray. However, unlike NeRF, we do not involve the extensive sampling of points along the ray, which would significantly degrade performance. Instead, our focus is solely on determining which tangent-plane Gaussians the ray intersects on the unit sphere:", + "bbox": [ + 214, + 430, + 782, + 503 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {2 D}} = \\varphi_ {\\mathrm {p}} \\left(\\left[ \\begin{array}{c} (u - c _ {x}) / f _ {x} \\\\ (v - c _ {y}) / f _ {y} \\\\ 1 \\end{array} \\right]\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 505, + 784, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $c_x, c_y, f_x, f_y$ denote the intrinsic parameters of the pinhole camera. We then query the function values of these Gaussians for alpha blending to obtain the color", + "bbox": [ + 214, + 556, + 782, + 601 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nG _ {2 \\mathrm {D}} \\left(\\mathbf {x} _ {2 \\mathrm {D}}\\right) = \\exp \\left\\{- 1 / 2 \\left(\\mathbf {x} _ {2 \\mathrm {D}} - \\varphi_ {\\mathrm {p}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) ^ {\\top} \\left(\\mathbf {J} _ {\\mathbf {p}} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} _ {\\mathbf {p}} ^ {\\top}\\right) ^ {- 1} \\left(\\mathbf {x} _ {2 \\mathrm {D}} - \\varphi_ {\\mathrm {p}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) \\right\\}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 611, + 784, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that both $\\mathbf{J}_{\\mathbf{p}}$ in Equation 14 and $\\mathbf{J}$ in Equation 6 are matrices with a rank of 2. Consequently, the inverse of the covariance matrix of the 2D Gaussian is also obtained by skipping the third row and column. However, the third row and third column of $\\mathbf{J}_{\\mathbf{p}}\\boldsymbol{\\Sigma}'\\mathbf{J}_{\\mathbf{p}}^\\top$ are not entirely zeros, as opposed to $\\mathbf{J}\\boldsymbol{\\Sigma}'\\mathbf{J}^\\top$ . To ensure the equation still holds, an invertible matrix $\\mathbf{Q}$ needs to be used for left multiplication with $\\mathbf{x}_{2\\mathrm{D}}$ , $\\varphi_{\\mathrm{p}}(\\boldsymbol{\\mu}')$ and $\\mathbf{J}_{\\mathbf{p}}$ :", + "bbox": [ + 214, + 657, + 782, + 758 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\left[ \\begin{array}{c c c} \\frac {\\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}} & 0 & - \\frac {\\mu_ {x}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {y}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}} \\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} - \\frac {\\mu_ {y} \\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}} \\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} \\\\ \\frac {\\mu_ {x}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\mu_ {y}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} \\end{array} \\right]. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 771, + 784, + 844 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since our projection does not rely on the perspective image plane $z = 1$ as in the original 3D-GS [16], we can adapt to different camera models such as fisheye cameras and panoramic images by simply modifying the transformation from image space to camera space, as described in Equation 16.", + "bbox": [ + 215, + 146, + 785, + 207 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 215, + 229, + 375, + 247 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To validate the effectiveness of the Optimal Gaussian Splatting derived from theoretical error analysis, we conduct a series of experiments, comparing it with the original 3D-GS [16] and some current state-of-the-art methods.", + "bbox": [ + 215, + 260, + 785, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Implementation", + "text_level": 1, + "bbox": [ + 215, + 328, + 393, + 345 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We implemented Optimal Gaussian Splitting based on the PyTorch framework in 3D-GS [16] and wrote custom CUDA kernels for rasterization. We used the default parameters of 3D-GS to maintain consistency with the original 3D-GS and prevent other factors from introducing interference into the results.", + "bbox": [ + 215, + 352, + 785, + 414 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.2 Datasets", + "text_level": 1, + "bbox": [ + 215, + 435, + 334, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We test our algorithm on a total of 13 real scenes which are the same as those used in the original 3D-GS [16]. In particular, we evaluate our approach on the complete set of scenes featured in Mip-NeRF360 [2], which currently represents the state-of-the-art in NeRF rendering quality. Additionally, we test our method on two scenes from the Tanks & Temples dataset [17] and two scenes provided by Hedman et al. [13]. The selected scenes exhibit diverse capture styles, encompassing both confined indoor environments and expansive, unbounded outdoor settings.", + "bbox": [ + 215, + 460, + 785, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.3 Results", + "text_level": 1, + "bbox": [ + 215, + 603, + 323, + 618 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our algorithm with 3D-GS [16] and three state-of-the-art NeRF-based approaches: Mip-NeRF360 [2], InstantNGP [26] and Plenoxels [8]. All methods are configured using the same settings as outlined in the 3D-GS [16] to control variables, preventing the introduction of variables other than our Optimal Splatting method.", + "bbox": [ + 215, + 628, + 785, + 704 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative comparisons We adopt a train/test split for datasets following the methodology proposed by Mip-NeRF360 [2]. Specifically, every 8th photo is reserved for testing, ensuring a consistent and meaningful basis for comparisons to generate error metrics. Standard metrics such as PSNR, LPIPS [43], and SSIM are employed for evaluation. We report quantitative results in Table 1. While the 3D-GS significantly outperforms MipNeRF360 in rendering performance, its rendering quality slightly falls behind MipNeRF360, with a lower average PSNR across the 13 scenes. This might be attributed to MipNeRF360 not employing", + "bbox": [ + 215, + 718, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dbb36083d07d2bfd893a35d4f8c527c1ae011c21d1d0df91bb10822c1bfcbf08.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of our method compared to previous work across three datasets. The Mip-NeRF360 [2] dataset includes both indoor and outdoor scenarios. Tanks. and Deep. indicates the average results over the Tanks&Templates [17] and Deep Blending [13] datasets. Average represents the mean value across all scenarios in these datasets." + ], + "table_footnote": [], + "table_body": "
Dataset Method|MetricAverageMip-NeRF360Tanks. and Deep.
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Plenoxels [8]22.770.6660.45723.080.6250.46322.070.7570.444
INGP-Base [26]24.490.6990.37325.300.6710.37122.670.7600.377
INGP-Big [26]24.930.7240.33625.590.6990.33123.440.7810.347
M-NeRF360 [2]27.110.8030.24127.690.7920.23725.810.8300.251
3D-GS [16]26.920.8320.21427.210.8150.21426.270.8720.213
Ours27.170.8360.21027.480.8210.20926.440.8720.214
", + "bbox": [ + 232, + 226, + 769, + 333 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d8afd8b9cbf9090cd121a3e0a2c14d26a85ea766d67fa45af51a6b7a8b85ef67.jpg", + "image_caption": [ + "Fig. 4: We show comparisons of our method to previous methods and the corresponding ground truth images from held-out test views. The scenes are, from the top down: TRUCK from Tanks&Temples [17]; PLAYROOM from the Deep Blending dataset [13] and BONSAI, COUNTER from Mip-NeRF360 dataset [2]. Differences in quality highlighted by arrows/insets." + ], + "image_footnote": [], + "bbox": [ + 217, + 345, + 785, + 640 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "significant approximations throughout its entire pipeline. However, by replacing the original projection method of 3D-GS with our optimal projection, which incurs smaller errors, we not only maintain rendering performance far surpassing MipNeRF360 but also achieve a quality improvement.", + "bbox": [ + 212, + 734, + 787, + 796 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative comparisons For our qualitative experiments, we select only two methods, 3D-GS and MipNeRF360, as our baselines. It can be observed that", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3a9051b9664d848d67240f7def0e83431f2673daa8e62b3d249915406c2fd0f5.jpg", + "image_caption": [ + "Fig. 5: We derive the projection error function with respect to image coordinates $(u, v)$ and focal length through the transformation between image and polar coordinates $(\\theta_{\\mu}, \\phi_{\\mu})$ . We visualize this function, showing 3D-GS rendered images with long and short focal lengths, followed by the corresponding error functions." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 215 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "our method is capable of generating more realistic details, with fewer artifacts compared to 3D-GS [16]. And this is precisely the superiority brought about by our projection method, which results in smaller errors. Furthermore, We see that in some cases Mip-NeRF360 has remaining artifacts that our method avoids (e.g., black lines - in PLAYROOM).", + "bbox": [ + 212, + 313, + 787, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.4 Discussions", + "text_level": 1, + "bbox": [ + 214, + 415, + 356, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Impacts of Decreasing Focal Length Based on Equation 9, we further derive the projection error function with respect to image coordinates $(u,v)$ and focal length $f_{x},f_{y}$ through the coordinate transformation between image coordinates and polar coordinates $(\\theta_{\\mu},\\phi_{\\mu})$ . As illustrated in Figure 5, the peak of projection error mainly occurs at the image edges under a long focal length, resulting in artifacts at the image edges. As the focal length decreases, the field of view expands, leading to more Gaussians deviating from the projection center and an overall increase in projection error. In such cases, 3D-GS exhibit more artifacts, such as needle-like structures or cloud-like large Gaussians, which obscure parts that perform well under long focal length, thereby significantly degrading the overall image quality. Since our method employs a central radial projection for the projection plane, theoretically, we should not encounter such defects. To validate this perspective, we conduct experiments at various focal length reduction ratios (reduction ratios of $\\times 0.2$ and $\\times 0.3$ ).", + "bbox": [ + 212, + 445, + 787, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For quantitative experiments, considering that there is no ground truth for large field-of-view in the 13 scenes from the 3D-GS [16], we apply a focal length mask. This mask selects central parts of the rendered images with corresponding ground truth based on the focal length scaling factor for metric calculation. The results are presented in the Table 2. Whether it is at $\\times 0.2$ or $\\times 0.3$ reduction ratios, our method outperforms the original 3D-GS significantly across various scenes in terms of all metrics. As the focal length decreases from $\\times 0.3$ to $\\times 0.2$ , 3D-GS [16] experiences a significant degradation due to increased projection errors, while our method demonstrates greater robustness.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For quantitative experiments, we randomly select three scenes, each with two different focal lengths for demonstration. Additional experimental results can be found in the supplementary materials. It is evident from the Figure 6", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/17e301929eeb6052252eb7808976d57580f3465a6395899e58a383de9154b15e.jpg", + "table_caption": [ + "Table 2: We quantitatively compare our method with the original 3D-GS [16] using focal lengths of 0.2 and 0.3 times the original. Six scenes were selected from three datasets: Mip-NeRF360 [2], Tank & Templates [17], and Deep Blending [13]. The scenes include TRAIN and TRUCK from Tanks & Templates, DRJOHNSON and PLAYROOM from Deep Blending, and ROOM and TREEHILL from Mip-NeRF360." + ], + "table_footnote": [], + "table_body": "
MetricFocalMethodTrainTruckDrJohnsonPlayroomRoomTreehillAverage
PSNR↑×0.23D-GS [16]10.1212.3919.2416.3820.0215.3615.58
Ours16.6016.7023.5423.2923.6718.9820.46
×0.33D-GS [16]12.8514.8923.3622.5125.0018.3519.49
Ours17.5218.0724.8926.4025.8719.8722.10
SSIM↑×0.23D-GS [16]0.3250.3800.6160.6190.6800.3710.499
Ours0.5650.5370.7010.7520.7770.4350.628
×0.33D-GS [16]0.4710.5290.7400.7910.8120.4370.630
Ours0.5880.6050.7620.8440.8340.4720.684
LPIPS↓×0.23D-GS [16]0.5690.4790.2960.3550.2070.4740.397
Ours0.2770.2500.2160.2030.1780.3850.251
×0.33D-GS [16]0.3310.2760.2070.2130.1540.3970.263
Ours0.2430.2170.2010.1820.1580.3850.231
", + "bbox": [ + 228, + 226, + 776, + 402 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1268bf6aebcbd50d49bf512bc15f8bfaefad5b4482ba126462629acc764d7e79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 421, + 331, + 470 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/806d986099681f6ef4726210ae9e4b9c5b8f115a8eac887c480f830699b21553.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 421, + 446, + 470 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a570e1470766f89c58851a3a7d14140da7e461252c3df58833133d4be2898dc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 421, + 558, + 470 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/82604c3b8841aab712b2d98ac7fe985456453319cead0b3b56f272a9d7d3caec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 421, + 671, + 470 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6e8aaebb3966e9a3c5863fb4f05f0aa1e7c136011190b7bc6bae970af8aeebc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 421, + 784, + 470 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1f224a3d72de997304df760b78a94a2e419c07838d97bbc0f03dc4a6a018f810.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 472, + 331, + 530 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a663eea222528cd1f488ef9d04e4ad9b4b4a97d33efccbc9d731945a22ba54e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 472, + 446, + 530 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3c71e304e7fbcfb6e8e0394be300935b2c64df261c1cc52571831c561dcb34df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 472, + 558, + 530 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/53297f4f353236a57052531f0f138eb92184c2fae7608bc9d3d459f636185bd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 472, + 669, + 530 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/12ddc4272ca0c2cb88a52e98376d3d02b2ac9a5919e06c5b2738a9e379d44702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 472, + 784, + 530 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f044836e2f37e1982f5475183f90501bd25da44c4e7fccbe899100087fb8db94.jpg", + "image_caption": [ + "Fig. 6: We compare our method to the original 3D-GS [16] with different focal length reductions. Two focal lengths, 0.2 and 0.3 times the original, are tested. Ground truth $\\times 1$ in the left images indicates no focal length reduction. The red-boxed areas in the right images correspond to the same regions in the left images." + ], + "image_footnote": [], + "bbox": [ + 217, + 530, + 331, + 588 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b1286fd98708d78e2d6a03ba5bffd24b1279568d4a5344f63d6782de1b2b3b43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 530, + 446, + 588 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/19ff1f611e4ac50d9b297c394b68d04779d83f7718904e0acedccf2288b926d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 530, + 558, + 588 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a5252c499c4ad2f162536b50871bcbfe809e42822428f03066a981ab68b4c684.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 530, + 669, + 588 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e322cc1fa21d192bc0afe8cd143d1844a7593c9317e9b522a10e12e2b598ee0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 530, + 784, + 588 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "that our method exhibits greater robustness across various focal length settings, mitigating the artifacts introduced by projection in the original 3D-GS.", + "bbox": [ + 212, + 691, + 785, + 722 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Adaptability to Various Camera Models By straightforwardly modifying the transformation from image space to camera space, we have achieved adaptation to various camera models, with the caveat that the original 3D-GS's projection struggles to support these camera models. Figure 7 illustrates that our projection, in addition to conforming to the original pinhole camera according to Equation 16, can also accommodate fisheye cameras and generate panoramas.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1021eeabfb4dcb03869c0a00a9f5eb44804a6cb59c16fbaac84f6936b59e9391.jpg", + "image_caption": [ + "Fig. 7: Renderings of three scenes using our projection under three camera models. From left to right: the ground truth of the original perspective, wide-angle rendering with a pinhole camera, fisheye camera rendering, and the panorama." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 366 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation** Due to our method's independence from the image plane, it exhibits adaptability to various camera models. However, the slight increase in training time is a consequence of the subsequent transformation to the image plane after projection onto the tangent plane, rather than a direct projection onto the image plane. Fortunately, this method avoids the need for dense points sampling, ensuring the algorithm's time complexity remains unchanged and rendering performance stays unaffected. Moreover, it may be possible to mitigate the impact on training time, particularly concerning the corresponding backward propagation, through the implementation of optimized CUDA code.", + "bbox": [ + 212, + 449, + 787, + 587 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 609, + 359, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Through theoretical analysis, we establish the correlation between the projection error and the Gaussian mean in 3D-GS. By optimizing the error function, we discover that the projection error is minimized when projecting on the plane perpendicular to the line connecting the Gaussian mean and the camera center. Leveraging this insight, we propose a novel projection that yields high-quality images without compromising rendering performance and can be easily adapted to various camera models through simple modifications.", + "bbox": [ + 212, + 642, + 787, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, our primary focus is on analyzing the relationship between the projection error and the mean of the 3D Gaussian under the assumption of a constant covariance. The influence of Gaussian covariance on projection requires further discussion in future work. It may be possible to further explore the potential of Gaussian Splatting as an explicit representation technique for scenes.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5835-5844. IEEE (2021)", + "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "3. Buehler, C., Bosse, M., McMillan, L., Gortler, S.J., Cohen, M.F.: Unstructured lumigraph rendering. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2001, Los Angeles, California, USA, August 12-17, 2001. pp. 425-432. ACM (2001). https://doi.org/10.1145/383259.383309, https://doi.org/10.1145/383259.383309", + "4. Cen, J., Fang, J., Yang, C., Xie, L., Zhang, X., Shen, W., Tian, Q.: Segment any 3d gaussians. arXiv preprint arXiv:2312.00860 (2023)", + "5. Chaurasia, G., Duchêne, S., Sorkine-Hornung, O., Drettakis, G.: Depth synthesis and local warps for plausible image-based navigation. ACM Trans. Graph. 32(3), 30:1-30:12 (2013). https://doi.org/10.1145/2487228.2487238, https://doi.org/10.1145/2487228.2487238", + "6. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)", + "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21476-21485 (2024)", + "8. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "9. Genova, K., Cole, F., Sud, A., Sarna, A., Funkhouser, T.: Local deep implicit functions for 3d shape. In: CVPR (2020)", + "0. Goesele, M., Snavely, N., Curless, B., Hoppe, H., Seitz, S.M.: Multi-view stereo for community photo collections. In: IEEE 11th International Conference on Computer Vision, ICCV 2007, Rio de Janeiro, Brazil, October 14-20, 2007. pp. 1-8. IEEE Computer Society (2007). https://doi.org/10.1109/ICCV.2007.4408933, https://doi.org/10.1109/ICCV.2007.4408933", + "1. Gortler, S.J., Grzeszczuk, R., Szeliski, R., Cohen, M.F.: The lumigraph. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 43-54. ACM (1996). https://doi.org/10.1145/237170.237200, https://doi.org/10.1145/237170.237200", + "2. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2024)", + "3. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018)" + ], + "bbox": [ + 225, + 179, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual (2020), https://proceedings.neurips.cc/paper/2020/bit/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html", + "15. Jiang, C., Sud, A., Makadia, A., Huang, J., Nießner, M., Funkhouser, T.: Local implicit grid representations for 3d scenes. In: CVPR (2020)", + "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)", + "18. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Trans. Graph. 41(6), 201:1-201:15 (2022). https://doi.org/10.1145/3550454.3555497, https://doi.org/10.1145/3550454.3555497", + "19. Levoy, M., Hanrahan, P.: Light field rendering. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 31-42. ACM (1996). https://doi.org/10.1145/237170.237199, https://doi.org/10.1145/237170.237199", + "20. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023)", + "21. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 2024 International Conference on 3D Vision (3DV). pp. 800-809. IEEE (2024)", + "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00713", + "23. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)", + "24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR (2019)", + "25. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision (2020)", + "26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatting for accelerated novel view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10349-10358 (2024)", + "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In: CVPR (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR (2019)", + "30. Radl, L., Steiner, M., Parger, M., Weinrauch, A., Kerbl, B., Steinberger, M.: StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics 4(43) (2024)", + "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)", + "32. Sitzmann, V., Zollhoefer, M., Wetzstein, G.: Scene representation networks: Continuous 3D-structure-aware neural scene representations. In: NeurIPS (2019)", + "33. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. ACM Trans. Graph. 25(3), 835-846 (2006). https://doi.org/10.1145/1141911.1141964, https://doi.org/10.1145/1141911.1141964", + "34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 11895-11907 (2019), https://proceedings.neurips.cc/paper/2019/bit/3001ef257407d5a371a96cdcd947c7d93-Abstraction.html", + "35. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=PxTIG12RRHS", + "36. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5449-5459. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00538, https://doi.org/10.1109/CVPR52688.2022.00538", + "37. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C.T., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11358-11367. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01120", + "38. Xiong, H., Muttukuru, S., Upadhyay, R., Chari, P., Kadambi, A.: Sparsegs: Real-time $360^{\\circ}$ sparse view synthesis using gaussian splatting. arXiv e-prints pp. arXiv-2312 (2023)", + "39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5428-5438. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00536, https://doi.org/10.1109/CVPR52688.2022.00536", + "40. Yang, C., Li, S., Fang, J., Liang, R., Xie, L., Zhang, X., Shen, W., Tian, Q.: Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv preprint arXiv:2402.10259 (2024)", + "41. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8254-8263 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "43. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "44. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE Transactions on Visualization and Computer Graphics 8(3), 223-238 (2002)" + ], + "bbox": [ + 215, + 146, + 787, + 244 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "L. Huang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_model.json b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5e8e7198cede0221f8214da87f11720d707d5850 --- /dev/null +++ b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_model.json @@ -0,0 +1,2480 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "On the Error Analysis of 3D Gaussian Splitting and an Optimal Projection Strategy" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.213, + 0.773, + 0.229 + ], + "angle": 0, + "content": "Letian Huang\\(^{\\text{®}}\\), Jiayang Bai\\(^{\\text{®}}\\), Jie Guo\\(^{\\star \\text{®}}\\), Yuanqi Li\\(^{\\text{®}}\\), and Yanwen Guo\\(^{\\text{®}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.24, + 0.778, + 0.283 + ], + "angle": 0, + "content": "Nanjing University, Nanjing, China \n{1thuang, jybai, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn \nhttps://letianhuang.github.io/op43dgs/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.321, + 0.744, + 0.571 + ], + "angle": 0, + "content": "Abstract. 3D Gaussian Splatting has garnered extensive attention and application in real-time neural rendering. Concurrently, concerns have been raised about the limitations of this technology in aspects such as point cloud storage, performance, and robustness in sparse viewpoints, leading to various improvements. However, there has been a notable lack of attention to the fundamental problem of projection errors introduced by the local affine approximation inherent in the splatting itself, and the consequential impact of these errors on the quality of photo-realistic rendering. This paper addresses the projection error function of 3D Gaussian Splatting, commencing with the residual error from the first-order Taylor expansion of the projection function. The analysis establishes a correlation between the error and the Gaussian mean position. Subsequently, leveraging function optimization theory, this paper analyzes the function's minima to provide an optimal projection strategy for Gaussian Splatting referred to Optimal Gaussian Splatting, which can accommodate a variety of camera models. Experimental validation further confirms that this projection methodology reduces artifacts, resulting in a more convincingly realistic rendering." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.584, + 0.744, + 0.613 + ], + "angle": 0, + "content": "Keywords: 3D deep learning \\(\\cdot\\) view synthesis \\(\\cdot\\) radiance fields \\(\\cdot\\) 3D gaussians \\(\\cdot\\) real-time rendering \\(\\cdot\\) error analysis" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.663, + 0.377, + 0.679 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Reconstructing 3D scenes from 2D images and synthesizing novel views has been a critical task in computer vision and graphics. Recently, the trend of novel view synthesis is spearheaded by Neural Radiance Fields (NeRFs) [25] and its variants [1, 2], which achieves photo-realistic rendering quality. NeRF represents scenes with Multi-Layer Perceptron (MLP) and renders novel views using volumetric ray-marching. However, these MLP-based methods contain a relatively large number of layers, leading to lengthy training and rendering times, which still fall significantly short of real-time rendering. To accelerate rendering," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.387, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.406, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.144, + 0.597, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.144, + 0.785, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Fig. 1: By minimizing the projection error through error analysis, we have achieved an improvement in the rendering image quality compared to the original 3D-GS [16]. These images are cropped from the complete images of the TRAIN scene, in order to be highlighted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.784, + 0.432 + ], + "angle": 0, + "content": "many methods resort to auxiliary data structures, such as hash grids [26], points or voxel [8,36,37,39]. However, due to the dense sampling of points along rays, they still face challenges in achieving real-time performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.789, + 0.598 + ], + "angle": 0, + "content": "With a substantial demand for high-speed and photo-realistic rendering effects, 3D Gaussian Splatting (3D-GS) [16] has emerged as an efficient representation that can be rendered at high speed on the GPU. It has departed from the implicit scene representation using MLPs and instead opted for an explicit representation using Gaussian functions. As a result, 3D-GS avoids the expensive cost of densely sampling points in volume rendering, thus achieving real-time performance. Given a set of images with the camera parameters calibrated by SfM [31], 3D-GS aims to optimize a set of 3D Gaussians as graphical primitives to explicitly represent the scene. To render 2D images, these 3D Gaussians are projected onto the image plane (\\(z = 1\\) plane) via a projection function denoted as \\(\\phi\\) for differentiable rasterization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.735 + ], + "angle": 0, + "content": "Unfortunately, Gaussian functions are closed under affine transformations but not under projective transformations. Therefore, 3D-GS adopts a local affine approximation [44], specifically approximating the projection function with the first two terms of its Taylor expansion. Nevertheless, approximations introduce errors, and the local affine approximation similarly contributes to these errors which may lead to artifacts in the rendered images. This paper exploits the relationship between the error of 3D-GS and the Gaussian mean through the analysis of the Taylor remainder term. Additionally, it identifies the circumstances under which the error is minimized by determining the extremum of the error function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Finally, based on the extremum analysis of the error function, we propose a novel optimal projection method to minimize the projection error in 3D-GS. This method requires only minor code modifications, does not affect real-time rendering performance, and yet achieves a significant improvement in rendering quality. Specifically, we project along the direction from the Gaussian mean to the camera center, where the projection plane is tangent to the line connecting the Gaussian mean and the camera center. Compared with the straightforward" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "approach of projecting 3D Gaussians onto the image plane of the camera, the proposed projection reduces expected error while maintaining real-time and realistic rendering performance. Furthermore, we find that the projection errors in 3D-GS tend to escalate with the continuous expansion of the field of view and the corresponding reduction in focal length. This increasing error leads to more and larger elongated Gaussians, as well as cloud-like Gaussian artifacts, consequently degrading the overall image quality. In contrast, our method exhibits greater realism and robustness. As illustrated in Figure 1, our method significantly reduces blurriness and artifacts caused by the projection error. Through experiments on various datasets and scenes, our method consistently outperforms the original projection method, particularly in settings with short focal lengths. And the method is easily implemented and adaptable to various camera models. To summarize, we provide the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.348, + 0.785, + 0.392 + ], + "angle": 0, + "content": "- We provide thorough error analysis for 3D-GS that may lead to artifacts and consequently degrade rendering quality. We have identified the correlation between this error and the Gaussian position." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.422 + ], + "angle": 0, + "content": "- We derive the mathematical expectation of this error function and analyze when this function takes extrema through methods of function optimization." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.422, + 0.785, + 0.482 + ], + "angle": 0, + "content": "- We propose a novel optimal projection capable of adapting to diverse camera models that adopts different tangent plane projections based on the directions from each Gaussian mean to the camera center instead of naively projecting all Gaussians onto the same plane" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.348, + 0.785, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.388, + 0.518 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Novel view synthesis is a longstanding challenge in computer vision and graphics. From traditional techniques [3,5,10,11,19,31,33], to neural network-based scene representations [9, 15, 18, 24, 28, 29, 32], various approaches have struggled to address the problem of synthesizing a new view from captured images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.608, + 0.456, + 0.623 + ], + "angle": 0, + "content": "2.1 Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.81 + ], + "angle": 0, + "content": "The Neural Radiance Field (NeRF) [25] stands out as a successful neural rendering method based on MLPs, primarily owing to its encoding of position and direction. This encoding allows for effective reconstruction of high-frequency information in scenes. Notable improvements on this encoding have been made by MipNeRF [1], NeRF-W [22], FreeNeRF [41], and Instant NGP [26]. These enhancements enable the handling of multi-resolution image inputs, multi-illumination with occlusion image inputs, sparse-view inputs, and achieve nearly real-time rendering capabilities, respectively. Barron et al. introduced MipNeRF360 [2] as an extension of MipNeRF [1] to address the issue of generating low-quality renderings for unbounded scenes in NeRF. Importantly, these methods thoroughly exploit the intrinsic capabilities of NeRF as an implicit scene representation without introducing additional model priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "However, due to the implicit representation of scenes and the dense sampling of points along rays, they still face challenges in achieving real-time performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.45, + 0.163 + ], + "angle": 0, + "content": "2.2 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.246 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3D-GS) [16] has demonstrated notable advancements in rendering performance by departing from MLPs and ray sampling, opting instead for Gaussian functions and Gaussian Splatting. This paradigm shift has attracted considerable attention within the industry, leading to various studies building upon this methodology [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.397 + ], + "angle": 0, + "content": "3D-GS finds application in SLAM systems [23], scene editing [7] and segmentation [4]. Efforts [20] have also been made to integrate 3D-GS with popular diffusion models [14, 34, 35] and dynamic scenes [21]. Recent endeavors aim to enhance the robustness in sparse-view scenarios [38, 40], performance, storage efficiency [27] and mesh reconstruction [12] of 3D-GS. However, these improvements do not specifically address errors associated with Gaussian projection. Potential methods for enhancing Gaussian functions include [30, 42]. Nevertheless, they introduce a Gaussian filter and a novel Gaussian depth sorting algorithm, respectively, without investigating the projection function and associated errors. These approaches still rely on projecting onto the \\(z = 1\\) plane." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.456 + ], + "angle": 0, + "content": "This paper aims to explore the potential of 3D Gaussian Splatting by analyzing errors that may arise during the projection process. The analysis will delve into the factors contributing to these errors and propose methods to mitigate and reduce them." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.479, + 0.383, + 0.495 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Similar to NeRF [25], the input to 3D-GS consists of a set of images, together with the corresponding cameras calibrated by SFM [31]. However, in contrast to NeRF, 3D-GS takes the sparse point cloud generated during the SFM calibration process as input. From these points, it constructs a set of 3D Gaussians as graphical primitives explicitly representing the scene. Each 3D Gaussian is characterized by its position (mean) \\(\\pmb{\\mu}\\), covariance matrix \\(\\pmb{\\Sigma}\\), and opacity \\(\\alpha\\), and carries anisotropic spherical harmonics (SH) representing directional appearance component (color) of the radiance field, following standard practice [8,26]. Subsequently, these 3D Gaussians are projected onto the image plane (\\(z = 1\\) plane) via a projection function for differentiable rasterization. Next, we will provide a detailed explanation of the Gaussian projection process in 3D-GS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.75 + ], + "angle": 0, + "content": "In the world space, we define a 3D Gaussian function by \\( G \\), characterized by its mean \\( \\mu \\) and covariance matrix \\( \\pmb{\\Sigma} \\). To project the Gaussian onto the image plane, the initial step in 3D-GS involves an affine transformation of this Gaussian function from the world space to the camera space via a viewing transformation matrix \\( \\mathbf{W} \\), yielding a transformed Gaussian function:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.759, + 0.787, + 0.784 + ], + "angle": 0, + "content": "\\[\nG (\\mathbf {x}) = \\exp \\left\\{- 1 / 2 (\\mathbf {W x} - \\mathbf {W} \\boldsymbol {\\mu}) ^ {\\top} (\\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top}) ^ {- 1} (\\mathbf {W x} - \\mathbf {W} \\boldsymbol {\\mu}) \\right\\}. \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.792, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We denote \\(G'\\) as the 3D Gaussian in the camera space, with \\(\\mathbf{x}' = \\mathbf{W}\\mathbf{x}\\), \\(\\boldsymbol{\\mu}' = \\mathbf{W}\\boldsymbol{\\mu}\\) and \\(\\boldsymbol{\\Sigma}' = \\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^\\top\\) as the point, mean and covariance matrix in the camera space, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.731, + 0.131 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.189 + ], + "angle": 0, + "content": "Subsequently, traditional 3D-GS requires projecting the Gaussian function in the camera space onto the \\(z = 1\\) plane, with the projection function \\(\\varphi\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x _ {0}}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.784, + 0.275 + ], + "angle": 0, + "content": "where \\(\\mathbf{x_0} = [0,0,1]^\\top\\) represents the projection of the camera space's origin onto this plane. Expanding this projection function to the first order using Taylor series yields:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.287, + 0.785, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) = \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) + \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) \\left(\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}\\right) + R _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) \\tag {3} \\\\ \\approx \\varphi (\\boldsymbol {\\mu} ^ {\\prime}) + \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} (\\boldsymbol {\\mu} ^ {\\prime}) (\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.784, + 0.427 + ], + "angle": 0, + "content": "where \\(\\frac{\\partial\\varphi}{\\partial\\mathbf{x}'}\\left(\\boldsymbol{\\mu}'\\right)\\) is the Jacobian of the affine approximation of the projective transformation, denoted as \\(\\mathbf{J}\\). Applying this local affine approximation by neglecting the Taylor remainder term allows to derive the 2D Gaussian function \\(G_{2\\mathrm{D}}\\) projected onto the \\(z = 1\\) plane as:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.439, + 0.785, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} G _ {2 D} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\exp \\left\\{- 1 / 2 \\left(\\mathbf {J} \\mathbf {x} ^ {\\prime} - \\mathbf {J} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {\\top} \\left(\\mathbf {J} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} ^ {\\top}\\right) ^ {- 1} \\left(\\mathbf {J} \\mathbf {x} ^ {\\prime} - \\mathbf {J} \\boldsymbol {\\mu} ^ {\\prime}\\right) \\right\\} \\\\ \\approx \\exp \\left\\{- 1 / 2 \\left(\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) ^ {\\top} \\left(\\mathbf {J} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} ^ {\\top}\\right) ^ {- 1} \\left(\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) \\right\\}. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.784, + 0.653 + ], + "angle": 0, + "content": "Since the rank of the matrix \\(\\mathbf{J}\\) is 2, the inverse of \\(\\mathbf{J} \\pmb{\\Sigma}' \\mathbf{J}^\\top\\) is, in fact, the inverse of the covariance matrix of the 2D Gaussian obtained by skipping the third row and column [44]. We denote \\(\\mathbf{x}_{2\\mathrm{D}} = \\varphi(\\mathbf{x}')\\), \\(\\pmb{\\mu}_{2\\mathrm{D}} = \\varphi(\\pmb{\\mu}')\\) and \\(\\pmb{\\Sigma}_{2\\mathrm{D}} = \\mathbf{J} \\pmb{\\Sigma}' \\mathbf{J}^\\top\\) as the point, mean and \\(2 \\times 2\\) covariance matrix of \\(G_{2\\mathrm{D}}\\) in the image space, respectively. For a pixel \\((u, v)\\), we obtain the value of the 2D Gaussian function by querying \\(\\mathbf{x}_{2\\mathrm{D}} = [u, v]^\\top\\), and then perform alpha blending to derive the color of the pixel. The blue box in Figure 3 illustrates the projection process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.679, + 0.58, + 0.696 + ], + "angle": 0, + "content": "4 Local Affine Approximation Error" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.784, + 0.803 + ], + "angle": 0, + "content": "From Equation 3 and Equation 4 we observe that 3D Gaussian Splatting introduces an approximation during the projection transformation, i.e. the local affine approximation. Consequently, 3D-GS actually adopts a 2D Gaussian function that is not the true projection of the initial 3D Gaussian. The error introduced by this approximation can be characterized by the Taylor remainder term in Equation 3:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.817, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\nR _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) - \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) \\left(\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}\\right) \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.265, + 0.16 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.173, + 0.786, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\mathbb {I} \\otimes \\left(\\mathbf {x} _ {0} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} - \\boldsymbol {\\mu} ^ {\\prime} \\left(\\mathbf {x} _ {0} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\mathbf {x} _ {0}\\right) ^ {- 1} \\mathbf {x} _ {0} ^ {\\top}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.789, + 0.324 + ], + "angle": 0, + "content": "\\(\\mathbb{I}\\) represents the identity matrix and \\(\\otimes\\) denotes the multiplication of a matrix and a scalar. Clearly, this Taylor remainder term is a three-dimensional vector related to the random variable \\(\\mathbf{x}^{\\prime}\\) and the mean \\(\\pmb{\\mu}^{\\prime}\\) of the Gaussian function. Therefore, by computing the square of the Frobenius norm of this vector \\(\\| R_1(\\mathbf{x}')\\| _F^2\\) and taking the mathematical expectation of this norm function with respect to the random variable \\(\\mathbf{x}^{\\prime}\\), we obtain an error function that depends solely on the Gaussian mean \\(\\pmb{\\mu}^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.337, + 0.786, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\epsilon \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\int_ {\\mathbf {x} ^ {\\prime} \\in \\mathcal {X} ^ {\\prime}} \\| R _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) \\| _ {F} ^ {2} \\mathrm {d} \\mathbf {x} ^ {\\prime}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.393, + 0.787, + 0.424 + ], + "angle": 0, + "content": "The simplified error function \\(\\epsilon\\) involves three unit vectors: \\(\\mathbf{x}^{\\prime}\\), \\(\\mathbf{x_0}\\) and \\(\\pmb{\\mu}^{\\prime}\\). Their spherical coordinates are given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.436, + 0.786, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathbf {0}} = \\left[ \\begin{array}{c} \\sin (\\phi_ {0}) \\cos (\\theta_ {0}) \\\\ - \\sin (\\theta_ {0}) \\\\ \\cos (\\phi_ {0}) \\cos (\\theta_ {0}) \\end{array} \\right], \\mathbf {x} ^ {\\prime} = \\left[ \\begin{array}{c} \\sin (\\phi) \\cos (\\theta) \\\\ - \\sin (\\theta) \\\\ \\cos (\\phi) \\cos (\\theta) \\end{array} \\right], \\boldsymbol {\\mu} ^ {\\prime} = \\left[ \\begin{array}{c} \\sin (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu}) \\\\ - \\sin (\\theta_ {\\mu}) \\\\ \\cos (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu}) \\end{array} \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.495, + 0.786, + 0.524 + ], + "angle": 0, + "content": "where \\(\\phi_0 = 0\\) and \\(\\theta_0 = 0\\). Please refer to the detailed proofs in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.526, + 0.787, + 0.607 + ], + "angle": 0, + "content": "The size of the integration region is related to the covariance of the Gaussian function. Since the current discussion specifically focuses on the impact of the mean on the error rather than the variance, let's assume the integration region to be \\(\\left\\{\\mathbf{x}' \\mid \\theta \\in [-\\pi/4 + \\theta_{\\mu}, \\pi/4 + \\theta_{\\mu}] \\land \\phi \\in [-\\pi/4 + \\phi_{\\mu}, \\pi/4 + \\phi_{\\mu}]\\right\\}\\). Then, based on Equation 5-8, we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.62, + 0.802, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon (\\theta_ {\\mu}, \\phi_ {\\mu}) = \\int_ {- \\pi / 4 + \\theta_ {\\mu}} ^ {\\pi / 4 + \\theta_ {\\mu}} \\int_ {- \\pi / 4 + \\phi_ {\\mu}} ^ {\\pi / 4 + \\phi_ {\\mu}} \\left(- \\frac {\\sin (\\phi - \\phi_ {\\mu}) \\cos (\\theta)}{\\cos^ {2} (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu})} + \\tan (\\phi) - \\tan (\\phi_ {\\mu})\\right) ^ {2} + \\\\ \\left(\\frac {\\sin (\\theta)}{\\cos (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu})} - \\frac {\\sin (\\theta_ {\\mu}) \\cos (\\phi) \\cos (\\theta)}{\\cos^ {2} (\\phi_ {\\mu}) \\cos^ {2} (\\theta_ {\\mu})} + \\frac {\\tan (\\theta_ {\\mu})}{\\cos (\\phi_ {\\mu})} - \\frac {\\tan (\\theta)}{\\cos (\\phi)}\\right) ^ {2} d \\theta d \\phi . \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.723, + 0.786, + 0.768 + ], + "angle": 0, + "content": "Fortunately, this integral has a close-form expression, but its form is very complicated. We report the close-form expression in the supplemental materials and visually depict it in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.769, + 0.787, + 0.8 + ], + "angle": 0, + "content": "The error function is partially differentiated with respect to \\(\\theta_{\\mu}\\) and \\(\\phi_{\\mu}\\). It is easy to check that at \\(\\theta_{\\mu} = \\theta_0 = 0\\) and \\(\\phi_{\\mu} = \\phi_0 = 0\\), the following holds:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.813, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\epsilon}{\\partial \\theta_ {\\mu}} (0, 0) = 0, \\frac {\\partial \\epsilon}{\\partial \\phi_ {\\mu}} (0, 0) = 0. \\tag {10}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.731, + 0.131 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.15, + 0.481, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.3, + 0.416, + 0.313 + ], + "angle": 0, + "content": "(a) \\(\\lambda = 0.095\\)" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.151, + 0.729, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.301, + 0.668, + 0.313 + ], + "angle": 0, + "content": "(b) \\(\\lambda = 0.95\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.789, + 0.392 + ], + "angle": 0, + "content": "Fig. 2: Visualization of the 3D Gaussian Splitting error function \\(\\epsilon (\\theta_{\\mu},\\phi_{\\mu})\\) under two distinct domains. The domain of the function is \\(\\{(\\theta_{\\mu},\\phi_{\\mu})\\mid \\theta_{\\mu}\\in [-\\lambda \\pi /4,\\lambda \\pi /4]\\land \\phi_{\\mu}\\in [-\\lambda \\pi /4,\\lambda \\pi /4]\\}\\) where \\((\\theta_{\\mu},\\phi_{\\mu})\\) represents the polar coordinate of the 3D Gaussian mean and \\(\\lambda\\) represents the scaling factor of the domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.789, + 0.573 + ], + "angle": 0, + "content": "Therefore, this point is actually an extremum point of the function. We also confirm that this point is the minimum of the function. Indeed, from Figure 2a, it is also apparent that the function is concave, with its minimum value occurring at \\((0,0)\\). Additionally, it can be observed that the minimum value of the error function is greater than zero, due to the fact that Gaussian functions are not closed under projection transformations. Figure 2b shows that in the majority of the region close to the origin, the function values vary very gently. However, as the independent variable of the function approaches the boundaries, the function values increase rapidly, resulting in a substantial difference between the maximum and minimum values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.789, + 0.697 + ], + "angle": 0, + "content": "The above analysis indicates that the error is small and does not significantly affect the quality of rendered images. Therefore, traditional 3D-GS [16] utilizes local affine approximation but still successfully reconstructs the scene to obtain a high-quality novel view image. Nevertheless, the naive projection of all Gaussians onto the same plane \\( z = 1 \\) in 3D-GS may lead to larger projection errors for Gaussians farther from the plane center (the projection point of the camera center to the plane), causing artifacts. These artifacts are severe when using wide-angle lenses, significantly degrading the quality of the renderings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.718, + 0.524, + 0.738 + ], + "angle": 0, + "content": "5 Optimal Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Based on the analysis of the error function, we introduce Optimal Gaussian Splitting shown in Figure 3. This helps us to achieve smaller projection errors, resulting in higher-quality rendering. The first step, similar to the original 3D-GS [16], involves transforming world coordinates into the camera coordinate system. Instead of projecting Gaussians onto the \\(z = 1\\) plane, each Gaussian is radially projected based on its mean along the line connecting it to the camera" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.395, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.147, + 0.787, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.359 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the rendering pipeline for our Optimal Gaussian Splitting and the projection of 3D-GS [16]. The blue box depicts the projection process of the original 3D-GS, which straightforwardly projects all Gaussians onto the same projection plane. In contrast, the red box illustrates our approach, where we project individual Gaussians onto corresponding tangent planes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.462 + ], + "angle": 0, + "content": "center, projecting onto a plane tangent to the unit sphere and perpendicular to this line. After projection, colors of points on the unit sphere are computed by alpha blending the 2D Gaussians on the tangent plane. Finally, for each pixel in the image, rays are cast onto the unit sphere to retrieve the color of that pixel, resulting in the rendered image." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.48, + 0.424, + 0.496 + ], + "angle": 0, + "content": "5.1 Optimal Projection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.623 + ], + "angle": 0, + "content": "We discover that the error function attains its minimum value when the projection of the Gaussian mean on the plane coincides with the projection from the camera center to the plane. Therefore, in Optimal Gaussian Splatting, we employ an optimal projection method. Specifically, instead of naively projecting different Gaussians onto the same plane \\( z = 1 \\), we adopt distinct projection planes for each Gaussian. These projection planes are determined by the tangent planes formed by the Gaussian mean and the line connecting it to the camera center. The specific projection plane is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.629, + 0.787, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\cdot \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x} _ {\\mathbf {p}}\\right) = 0 \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.661, + 0.265, + 0.674 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.672, + 0.787, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathbf {p}} = \\varpi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\boldsymbol {\\mu} ^ {\\prime} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1 / 2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.751 + ], + "angle": 0, + "content": "represents the projection of the camera space's origin onto this plane and \\(\\varpi\\) function represents the transformation that projects points onto the unit sphere. According to Equation 11, the optimal projection function \\(\\varphi_{\\mathrm{p}}\\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.756, + 0.787, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {\\mathrm {p}} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} _ {\\mathbf {p}}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.791, + 0.714, + 0.808 + ], + "angle": 0, + "content": "The corresponding local affine approximation Jacobian matrix \\(\\mathbf{J}_{\\mathbf{p}}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.814, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} _ {\\mathbf {p}} = \\frac {\\partial \\varphi_ {\\mathrm {p}}}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\mathbb {I} \\otimes \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} - \\boldsymbol {\\mu} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\mathbf {x} _ {\\mathbf {p}}\\right) ^ {- 1} \\mathbf {x} _ {\\mathbf {p}} ^ {\\top}. \\tag {14}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.145, + 0.718, + 0.166 + ], + "angle": 0, + "content": "When \\(\\pmb{\\mu}^{\\prime} = [\\mu_{x},\\mu_{y},\\mu_{z}]^{\\top}\\), the specific form of this Jacobian matrix is:" + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.177, + 0.786, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} _ {\\mathbf {p}} = \\left[ \\begin{array}{l l l} \\frac {\\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {x} \\mu_ {y}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {x} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {y}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & \\frac {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {y} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {y} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\cfrac {3}{2}}} & \\frac {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\cfrac {3}{2}}} \\end{array} \\right]. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.276, + 0.784, + 0.306 + ], + "angle": 0, + "content": "It's very easy to implement modifications to the Jacobian matrix in the forward process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.329, + 0.507, + 0.344 + ], + "angle": 0, + "content": "5.2 Unit Sphere Based Rasterizer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.784, + 0.429 + ], + "angle": 0, + "content": "For the proposed Optimal Projection, we have proposed a Unit Sphere Based Rasterizer to generate images. Through Optimal Projection, we obtain the projection of the three-dimensional Gaussians on the tangent plane of the unit sphere instead of obtaining the Gaussians in the image space. Therefore, we need to rasterize based on this unit sphere." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.431, + 0.784, + 0.505 + ], + "angle": 0, + "content": "Specifically, for a pixel \\((u,v)\\) on the image, similar to NeRF, we cast a ray. However, unlike NeRF, we do not involve the extensive sampling of points along the ray, which would significantly degrade performance. Instead, our focus is solely on determining which tangent-plane Gaussians the ray intersects on the unit sphere:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.506, + 0.785, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {2 D}} = \\varphi_ {\\mathrm {p}} \\left(\\left[ \\begin{array}{c} (u - c _ {x}) / f _ {x} \\\\ (v - c _ {y}) / f _ {y} \\\\ 1 \\end{array} \\right]\\right) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.558, + 0.784, + 0.602 + ], + "angle": 0, + "content": "where \\( c_x, c_y, f_x, f_y \\) denote the intrinsic parameters of the pinhole camera. We then query the function values of these Gaussians for alpha blending to obtain the color" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.612, + 0.785, + 0.658 + ], + "angle": 0, + "content": "\\[\nG _ {2 \\mathrm {D}} \\left(\\mathbf {x} _ {2 \\mathrm {D}}\\right) = \\exp \\left\\{- 1 / 2 \\left(\\mathbf {x} _ {2 \\mathrm {D}} - \\varphi_ {\\mathrm {p}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) ^ {\\top} \\left(\\mathbf {J} _ {\\mathbf {p}} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} _ {\\mathbf {p}} ^ {\\top}\\right) ^ {- 1} \\left(\\mathbf {x} _ {2 \\mathrm {D}} - \\varphi_ {\\mathrm {p}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) \\right\\}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.784, + 0.759 + ], + "angle": 0, + "content": "Note that both \\(\\mathbf{J}_{\\mathbf{p}}\\) in Equation 14 and \\(\\mathbf{J}\\) in Equation 6 are matrices with a rank of 2. Consequently, the inverse of the covariance matrix of the 2D Gaussian is also obtained by skipping the third row and column. However, the third row and third column of \\(\\mathbf{J}_{\\mathbf{p}}\\boldsymbol{\\Sigma}'\\mathbf{J}_{\\mathbf{p}}^\\top\\) are not entirely zeros, as opposed to \\(\\mathbf{J}\\boldsymbol{\\Sigma}'\\mathbf{J}^\\top\\). To ensure the equation still holds, an invertible matrix \\(\\mathbf{Q}\\) needs to be used for left multiplication with \\(\\mathbf{x}_{2\\mathrm{D}}\\), \\(\\varphi_{\\mathrm{p}}(\\boldsymbol{\\mu}')\\) and \\(\\mathbf{J}_{\\mathbf{p}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.772, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\left[ \\begin{array}{c c c} \\frac {\\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}} & 0 & - \\frac {\\mu_ {x}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {y}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}} \\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} - \\frac {\\mu_ {y} \\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}} \\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} \\\\ \\frac {\\mu_ {x}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\mu_ {y}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} \\end{array} \\right]. \\tag {18}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.208 + ], + "angle": 0, + "content": "Since our projection does not rely on the perspective image plane \\( z = 1 \\) as in the original 3D-GS [16], we can adapt to different camera models such as fisheye cameras and panoramic images by simply modifying the transformation from image space to camera space, as described in Equation 16." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.23, + 0.377, + 0.248 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.308 + ], + "angle": 0, + "content": "To validate the effectiveness of the Optimal Gaussian Splatting derived from theoretical error analysis, we conduct a series of experiments, comparing it with the original 3D-GS [16] and some current state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.329, + 0.395, + 0.346 + ], + "angle": 0, + "content": "6.1 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.353, + 0.787, + 0.415 + ], + "angle": 0, + "content": "We implemented Optimal Gaussian Splitting based on the PyTorch framework in 3D-GS [16] and wrote custom CUDA kernels for rasterization. We used the default parameters of 3D-GS to maintain consistency with the original 3D-GS and prevent other factors from introducing interference into the results." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.436, + 0.335, + 0.451 + ], + "angle": 0, + "content": "6.2 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.582 + ], + "angle": 0, + "content": "We test our algorithm on a total of 13 real scenes which are the same as those used in the original 3D-GS [16]. In particular, we evaluate our approach on the complete set of scenes featured in Mip-NeRF360 [2], which currently represents the state-of-the-art in NeRF rendering quality. Additionally, we test our method on two scenes from the Tanks & Temples dataset [17] and two scenes provided by Hedman et al. [13]. The selected scenes exhibit diverse capture styles, encompassing both confined indoor environments and expansive, unbounded outdoor settings." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.604, + 0.324, + 0.619 + ], + "angle": 0, + "content": "6.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.705 + ], + "angle": 0, + "content": "We compare our algorithm with 3D-GS [16] and three state-of-the-art NeRF-based approaches: Mip-NeRF360 [2], InstantNGP [26] and Plenoxels [8]. All methods are configured using the same settings as outlined in the 3D-GS [16] to control variables, preventing the introduction of variables other than our Optimal Splatting method." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.719, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Quantitative comparisons We adopt a train/test split for datasets following the methodology proposed by Mip-NeRF360 [2]. Specifically, every 8th photo is reserved for testing, ensuring a consistent and meaningful basis for comparisons to generate error metrics. Standard metrics such as PSNR, LPIPS [43], and SSIM are employed for evaluation. We report quantitative results in Table 1. While the 3D-GS significantly outperforms MipNeRF360 in rendering performance, its rendering quality slightly falls behind MipNeRF360, with a lower average PSNR across the 13 scenes. This might be attributed to MipNeRF360 not employing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of our method compared to previous work across three datasets. The Mip-NeRF360 [2] dataset includes both indoor and outdoor scenarios. Tanks. and Deep. indicates the average results over the Tanks&Templates [17] and Deep Blending [13] datasets. Average represents the mean value across all scenarios in these datasets." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.227, + 0.77, + 0.334 + ], + "angle": 0, + "content": "
Dataset Method|MetricAverageMip-NeRF360Tanks. and Deep.
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Plenoxels [8]22.770.6660.45723.080.6250.46322.070.7570.444
INGP-Base [26]24.490.6990.37325.300.6710.37122.670.7600.377
INGP-Big [26]24.930.7240.33625.590.6990.33123.440.7810.347
M-NeRF360 [2]27.110.8030.24127.690.7920.23725.810.8300.251
3D-GS [16]26.920.8320.21427.210.8150.21426.270.8720.213
Ours27.170.8360.21027.480.8210.20926.440.8720.214
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.346, + 0.787, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Fig. 4: We show comparisons of our method to previous methods and the corresponding ground truth images from held-out test views. The scenes are, from the top down: TRUCK from Tanks&Temples [17]; PLAYROOM from the Deep Blending dataset [13] and BONSAI, COUNTER from Mip-NeRF360 dataset [2]. Differences in quality highlighted by arrows/insets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.736, + 0.788, + 0.797 + ], + "angle": 0, + "content": "significant approximations throughout its entire pipeline. However, by replacing the original projection method of 3D-GS with our optimal projection, which incurs smaller errors, we not only maintain rendering performance far surpassing MipNeRF360 but also achieve a quality improvement." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative comparisons For our qualitative experiments, we select only two methods, 3D-GS and MipNeRF360, as our baselines. It can be observed that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.226, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Fig. 5: We derive the projection error function with respect to image coordinates \\((u, v)\\) and focal length through the transformation between image and polar coordinates \\((\\theta_{\\mu}, \\phi_{\\mu})\\). We visualize this function, showing 3D-GS rendered images with long and short focal lengths, followed by the corresponding error functions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.39 + ], + "angle": 0, + "content": "our method is capable of generating more realistic details, with fewer artifacts compared to 3D-GS [16]. And this is precisely the superiority brought about by our projection method, which results in smaller errors. Furthermore, We see that in some cases Mip-NeRF360 has remaining artifacts that our method avoids (e.g., black lines - in PLAYROOM)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.416, + 0.357, + 0.43 + ], + "angle": 0, + "content": "6.4 Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Impacts of Decreasing Focal Length Based on Equation 9, we further derive the projection error function with respect to image coordinates \\((u,v)\\) and focal length \\(f_{x},f_{y}\\) through the coordinate transformation between image coordinates and polar coordinates \\((\\theta_{\\mu},\\phi_{\\mu})\\). As illustrated in Figure 5, the peak of projection error mainly occurs at the image edges under a long focal length, resulting in artifacts at the image edges. As the focal length decreases, the field of view expands, leading to more Gaussians deviating from the projection center and an overall increase in projection error. In such cases, 3D-GS exhibit more artifacts, such as needle-like structures or cloud-like large Gaussians, which obscure parts that perform well under long focal length, thereby significantly degrading the overall image quality. Since our method employs a central radial projection for the projection plane, theoretically, we should not encounter such defects. To validate this perspective, we conduct experiments at various focal length reduction ratios (reduction ratios of \\(\\times 0.2\\) and \\(\\times 0.3\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "For quantitative experiments, considering that there is no ground truth for large field-of-view in the 13 scenes from the 3D-GS [16], we apply a focal length mask. This mask selects central parts of the rendered images with corresponding ground truth based on the focal length scaling factor for metric calculation. The results are presented in the Table 2. Whether it is at \\(\\times 0.2\\) or \\(\\times 0.3\\) reduction ratios, our method outperforms the original 3D-GS significantly across various scenes in terms of all metrics. As the focal length decreases from \\(\\times 0.3\\) to \\(\\times 0.2\\), 3D-GS [16] experiences a significant degradation due to increased projection errors, while our method demonstrates greater robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For quantitative experiments, we randomly select three scenes, each with two different focal lengths for demonstration. Additional experimental results can be found in the supplementary materials. It is evident from the Figure 6" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: We quantitatively compare our method with the original 3D-GS [16] using focal lengths of 0.2 and 0.3 times the original. Six scenes were selected from three datasets: Mip-NeRF360 [2], Tank & Templates [17], and Deep Blending [13]. The scenes include TRAIN and TRUCK from Tanks & Templates, DRJOHNSON and PLAYROOM from Deep Blending, and ROOM and TREEHILL from Mip-NeRF360." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.227, + 0.777, + 0.404 + ], + "angle": 0, + "content": "
MetricFocalMethodTrainTruckDrJohnsonPlayroomRoomTreehillAverage
PSNR↑×0.23D-GS [16]10.1212.3919.2416.3820.0215.3615.58
Ours16.6016.7023.5423.2923.6718.9820.46
×0.33D-GS [16]12.8514.8923.3622.5125.0018.3519.49
Ours17.5218.0724.8926.4025.8719.8722.10
SSIM↑×0.23D-GS [16]0.3250.3800.6160.6190.6800.3710.499
Ours0.5650.5370.7010.7520.7770.4350.628
×0.33D-GS [16]0.4710.5290.7400.7910.8120.4370.630
Ours0.5880.6050.7620.8440.8340.4720.684
LPIPS↓×0.23D-GS [16]0.5690.4790.2960.3550.2070.4740.397
Ours0.2770.2500.2160.2030.1780.3850.251
×0.33D-GS [16]0.3310.2760.2070.2130.1540.3970.263
Ours0.2430.2170.2010.1820.1580.3850.231
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.422, + 0.332, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.422, + 0.447, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.422, + 0.559, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.422, + 0.672, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.422, + 0.785, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.473, + 0.332, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.473, + 0.447, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.473, + 0.559, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.473, + 0.671, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.473, + 0.785, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.531, + 0.332, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.531, + 0.447, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.531, + 0.559, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.531, + 0.671, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.531, + 0.785, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Fig. 6: We compare our method to the original 3D-GS [16] with different focal length reductions. Two focal lengths, 0.2 and 0.3 times the original, are tested. Ground truth \\(\\times 1\\) in the left images indicates no focal length reduction. The red-boxed areas in the right images correspond to the same regions in the left images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.787, + 0.723 + ], + "angle": 0, + "content": "that our method exhibits greater robustness across various focal length settings, mitigating the artifacts introduced by projection in the original 3D-GS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Adaptability to Various Camera Models By straightforwardly modifying the transformation from image space to camera space, we have achieved adaptation to various camera models, with the caveat that the original 3D-GS's projection struggles to support these camera models. Figure 7 illustrates that our projection, in addition to conforming to the original pinhole camera according to Equation 16, can also accommodate fisheye cameras and generate panoramas." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig. 7: Renderings of three scenes using our projection under three camera models. From left to right: the ground truth of the original perspective, wide-angle rendering with a pinhole camera, fisheye camera rendering, and the panorama." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.588 + ], + "angle": 0, + "content": "**Limitation** Due to our method's independence from the image plane, it exhibits adaptability to various camera models. However, the slight increase in training time is a consequence of the subsequent transformation to the image plane after projection onto the tangent plane, rather than a direct projection onto the image plane. Fortunately, this method avoids the need for dense points sampling, ensuring the algorithm's time complexity remains unchanged and rendering performance stays unaffected. Moreover, it may be possible to mitigate the impact on training time, particularly concerning the corresponding backward propagation, through the implementation of optimized CUDA code." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.36, + 0.627 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Through theoretical analysis, we establish the correlation between the projection error and the Gaussian mean in 3D-GS. By optimizing the error function, we discover that the projection error is minimized when projecting on the plane perpendicular to the line connecting the Gaussian mean and the camera center. Leveraging this insight, we propose a novel projection that yields high-quality images without compromising rendering performance and can be easily adapted to various camera models through simple modifications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this work, our primary focus is on analyzing the relationship between the projection error and the mean of the 3D Gaussian under the assumption of a constant covariance. The influence of Gaussian covariance on projection requires further discussion in future work. It may be possible to further explore the potential of Gaussian Splatting as an explicit representation technique for scenes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.786, + 0.236 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5835-5844. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.787, + 0.292 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.787, + 0.363 + ], + "angle": 0, + "content": "3. Buehler, C., Bosse, M., McMillan, L., Gortler, S.J., Cohen, M.F.: Unstructured lumigraph rendering. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2001, Los Angeles, California, USA, August 12-17, 2001. pp. 425-432. ACM (2001). https://doi.org/10.1145/383259.383309, https://doi.org/10.1145/383259.383309" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.39 + ], + "angle": 0, + "content": "4. Cen, J., Fang, J., Yang, C., Xie, L., Zhang, X., Shen, W., Tian, Q.: Segment any 3d gaussians. arXiv preprint arXiv:2312.00860 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.787, + 0.446 + ], + "angle": 0, + "content": "5. Chaurasia, G., Duchêne, S., Sorkine-Hornung, O., Drettakis, G.: Depth synthesis and local warps for plausible image-based navigation. ACM Trans. Graph. 32(3), 30:1-30:12 (2013). https://doi.org/10.1145/2487228.2487238, https://doi.org/10.1145/2487228.2487238" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.787, + 0.474 + ], + "angle": 0, + "content": "6. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.532 + ], + "angle": 0, + "content": "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21476-21485 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.787, + 0.574 + ], + "angle": 0, + "content": "8. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.787, + 0.602 + ], + "angle": 0, + "content": "9. Genova, K., Cole, F., Sud, A., Sarna, A., Funkhouser, T.: Local deep implicit functions for 3d shape. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.787, + 0.671 + ], + "angle": 0, + "content": "0. Goesele, M., Snavely, N., Curless, B., Hoppe, H., Seitz, S.M.: Multi-view stereo for community photo collections. In: IEEE 11th International Conference on Computer Vision, ICCV 2007, Rio de Janeiro, Brazil, October 14-20, 2007. pp. 1-8. IEEE Computer Society (2007). https://doi.org/10.1109/ICCV.2007.4408933, https://doi.org/10.1109/ICCV.2007.4408933" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.741 + ], + "angle": 0, + "content": "1. Gortler, S.J., Grzeszczuk, R., Szeliski, R., Cohen, M.F.: The lumigraph. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 43-54. ACM (1996). https://doi.org/10.1145/237170.237200, https://doi.org/10.1145/237170.237200" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.787, + 0.798 + ], + "angle": 0, + "content": "2. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "3. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.23 + ], + "angle": 0, + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual (2020), https://proceedings.neurips.cc/paper/2020/bit/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.786, + 0.259 + ], + "angle": 0, + "content": "15. Jiang, C., Sud, A., Makadia, A., Huang, J., Nießner, M., Funkhouser, T.: Local implicit grid representations for 3d scenes. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.286 + ], + "angle": 0, + "content": "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "18. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Trans. Graph. 41(6), 201:1-201:15 (2022). https://doi.org/10.1145/3550454.3555497, https://doi.org/10.1145/3550454.3555497" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.451 + ], + "angle": 0, + "content": "19. Levoy, M., Hanrahan, P.: Light field rendering. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 31-42. ACM (1996). https://doi.org/10.1145/237170.237199, https://doi.org/10.1145/237170.237199" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "20. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "21. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 2024 International Conference on 3D Vision (3DV). pp. 800-809. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.605 + ], + "angle": 0, + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00713" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "23. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "25. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatting for accelerated novel view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10349-10358 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.217 + ], + "angle": 0, + "content": "30. Radl, L., Steiner, M., Parger, M., Weinrauch, A., Kerbl, B., Steinberger, M.: StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics 4(43) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.258 + ], + "angle": 0, + "content": "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.259, + 0.787, + 0.286 + ], + "angle": 0, + "content": "32. Sitzmann, V., Zollhoefer, M., Wetzstein, G.: Scene representation networks: Continuous 3D-structure-aware neural scene representations. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "33. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. ACM Trans. Graph. 25(3), 835-846 (2006). https://doi.org/10.1145/1141911.1141964, https://doi.org/10.1145/1141911.1141964" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.424 + ], + "angle": 0, + "content": "34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 11895-11907 (2019), https://proceedings.neurips.cc/paper/2019/bit/3001ef257407d5a371a96cdcd947c7d93-Abstraction.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.493 + ], + "angle": 0, + "content": "35. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=PxTIG12RRHS" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.494, + 0.787, + 0.563 + ], + "angle": 0, + "content": "36. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5449-5459. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00538, https://doi.org/10.1109/CVPR52688.2022.00538" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.787, + 0.645 + ], + "angle": 0, + "content": "37. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C.T., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11358-11367. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01120" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.687 + ], + "angle": 0, + "content": "38. Xiong, H., Muttukuru, S., Upadhyay, R., Chari, P., Kadambi, A.: Sparsegs: Real-time \\(360^{\\circ}\\) sparse view synthesis using gaussian splatting. arXiv e-prints pp. arXiv-2312 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.688, + 0.787, + 0.757 + ], + "angle": 0, + "content": "39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5428-5438. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00536, https://doi.org/10.1109/CVPR52688.2022.00536" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "40. Yang, C., Li, S., Fang, J., Liang, R., Xie, L., Zhang, X., Shen, W., Tian, Q.: Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv preprint arXiv:2402.10259 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "41. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8254-8263 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "L. Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "42. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "43. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.246 + ], + "angle": 0, + "content": "44. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE Transactions on Visualization and Computer Graphics 8(3), 223-238 (2002)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.246 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_origin.pdf b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dfdebbdf7a22293e05627f1719e343ad705bc09b --- /dev/null +++ b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/b186b2c8-8ed4-4348-a524-63a72c671518_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cbe6c1cbfd85a8168e0c9fb42cadd48242efeab78a9a92fdb6898cc5a878bfe +size 9536544 diff --git a/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/full.md b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cbf5b28fd55c4266bfb2295f2968cca8482ff521 --- /dev/null +++ b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/full.md @@ -0,0 +1,350 @@ +# On the Error Analysis of 3D Gaussian Splitting and an Optimal Projection Strategy + +Letian Huang $^{\text{®}}$ , Jiayang Bai $^{\text{®}}$ , Jie Guo $^{\star \text{®}}$ , Yuanqi Li $^{\text{®}}$ , and Yanwen Guo $^{\text{®}}$ + +Nanjing University, Nanjing, China +{1thuang, jybai, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn +https://letianhuang.github.io/op43dgs/ + +Abstract. 3D Gaussian Splatting has garnered extensive attention and application in real-time neural rendering. Concurrently, concerns have been raised about the limitations of this technology in aspects such as point cloud storage, performance, and robustness in sparse viewpoints, leading to various improvements. However, there has been a notable lack of attention to the fundamental problem of projection errors introduced by the local affine approximation inherent in the splatting itself, and the consequential impact of these errors on the quality of photo-realistic rendering. This paper addresses the projection error function of 3D Gaussian Splatting, commencing with the residual error from the first-order Taylor expansion of the projection function. The analysis establishes a correlation between the error and the Gaussian mean position. Subsequently, leveraging function optimization theory, this paper analyzes the function's minima to provide an optimal projection strategy for Gaussian Splatting referred to Optimal Gaussian Splatting, which can accommodate a variety of camera models. Experimental validation further confirms that this projection methodology reduces artifacts, resulting in a more convincingly realistic rendering. + +Keywords: 3D deep learning $\cdot$ view synthesis $\cdot$ radiance fields $\cdot$ 3D gaussians $\cdot$ real-time rendering $\cdot$ error analysis + +# 1 Introduction + +Reconstructing 3D scenes from 2D images and synthesizing novel views has been a critical task in computer vision and graphics. Recently, the trend of novel view synthesis is spearheaded by Neural Radiance Fields (NeRFs) [25] and its variants [1, 2], which achieves photo-realistic rendering quality. NeRF represents scenes with Multi-Layer Perceptron (MLP) and renders novel views using volumetric ray-marching. However, these MLP-based methods contain a relatively large number of layers, leading to lengthy training and rendering times, which still fall significantly short of real-time rendering. To accelerate rendering, + +![](images/77cd6b753a96e723145c87409706006a12e725f4bc39dbe589905d0347c0d3ee.jpg) +Fig. 1: By minimizing the projection error through error analysis, we have achieved an improvement in the rendering image quality compared to the original 3D-GS [16]. These images are cropped from the complete images of the TRAIN scene, in order to be highlighted. + +![](images/297487479d90a6b06985f07bd69cf50747d4c68ffcc67fc9eaa3b04d6a8b7098.jpg) + +![](images/1c66b8cd20ba5787cd72de9697c734ef0fd46c516a26b98b8a718c4f146f6237.jpg) + +many methods resort to auxiliary data structures, such as hash grids [26], points or voxel [8,36,37,39]. However, due to the dense sampling of points along rays, they still face challenges in achieving real-time performance. + +With a substantial demand for high-speed and photo-realistic rendering effects, 3D Gaussian Splatting (3D-GS) [16] has emerged as an efficient representation that can be rendered at high speed on the GPU. It has departed from the implicit scene representation using MLPs and instead opted for an explicit representation using Gaussian functions. As a result, 3D-GS avoids the expensive cost of densely sampling points in volume rendering, thus achieving real-time performance. Given a set of images with the camera parameters calibrated by SfM [31], 3D-GS aims to optimize a set of 3D Gaussians as graphical primitives to explicitly represent the scene. To render 2D images, these 3D Gaussians are projected onto the image plane ( $z = 1$ plane) via a projection function denoted as $\phi$ for differentiable rasterization. + +Unfortunately, Gaussian functions are closed under affine transformations but not under projective transformations. Therefore, 3D-GS adopts a local affine approximation [44], specifically approximating the projection function with the first two terms of its Taylor expansion. Nevertheless, approximations introduce errors, and the local affine approximation similarly contributes to these errors which may lead to artifacts in the rendered images. This paper exploits the relationship between the error of 3D-GS and the Gaussian mean through the analysis of the Taylor remainder term. Additionally, it identifies the circumstances under which the error is minimized by determining the extremum of the error function. + +Finally, based on the extremum analysis of the error function, we propose a novel optimal projection method to minimize the projection error in 3D-GS. This method requires only minor code modifications, does not affect real-time rendering performance, and yet achieves a significant improvement in rendering quality. Specifically, we project along the direction from the Gaussian mean to the camera center, where the projection plane is tangent to the line connecting the Gaussian mean and the camera center. Compared with the straightforward + +approach of projecting 3D Gaussians onto the image plane of the camera, the proposed projection reduces expected error while maintaining real-time and realistic rendering performance. Furthermore, we find that the projection errors in 3D-GS tend to escalate with the continuous expansion of the field of view and the corresponding reduction in focal length. This increasing error leads to more and larger elongated Gaussians, as well as cloud-like Gaussian artifacts, consequently degrading the overall image quality. In contrast, our method exhibits greater realism and robustness. As illustrated in Figure 1, our method significantly reduces blurriness and artifacts caused by the projection error. Through experiments on various datasets and scenes, our method consistently outperforms the original projection method, particularly in settings with short focal lengths. And the method is easily implemented and adaptable to various camera models. To summarize, we provide the following contributions: + +- We provide thorough error analysis for 3D-GS that may lead to artifacts and consequently degrade rendering quality. We have identified the correlation between this error and the Gaussian position. +- We derive the mathematical expectation of this error function and analyze when this function takes extrema through methods of function optimization. +- We propose a novel optimal projection capable of adapting to diverse camera models that adopts different tangent plane projections based on the directions from each Gaussian mean to the camera center instead of naively projecting all Gaussians onto the same plane + +# 2 Related Work + +Novel view synthesis is a longstanding challenge in computer vision and graphics. From traditional techniques [3,5,10,11,19,31,33], to neural network-based scene representations [9, 15, 18, 24, 28, 29, 32], various approaches have struggled to address the problem of synthesizing a new view from captured images. + +# 2.1 Neural Radiance Fields + +The Neural Radiance Field (NeRF) [25] stands out as a successful neural rendering method based on MLPs, primarily owing to its encoding of position and direction. This encoding allows for effective reconstruction of high-frequency information in scenes. Notable improvements on this encoding have been made by MipNeRF [1], NeRF-W [22], FreeNeRF [41], and Instant NGP [26]. These enhancements enable the handling of multi-resolution image inputs, multi-illumination with occlusion image inputs, sparse-view inputs, and achieve nearly real-time rendering capabilities, respectively. Barron et al. introduced MipNeRF360 [2] as an extension of MipNeRF [1] to address the issue of generating low-quality renderings for unbounded scenes in NeRF. Importantly, these methods thoroughly exploit the intrinsic capabilities of NeRF as an implicit scene representation without introducing additional model priors. + +However, due to the implicit representation of scenes and the dense sampling of points along rays, they still face challenges in achieving real-time performance. + +# 2.2 3D Gaussian Splitting + +3D Gaussian Splatting (3D-GS) [16] has demonstrated notable advancements in rendering performance by departing from MLPs and ray sampling, opting instead for Gaussian functions and Gaussian Splatting. This paradigm shift has attracted considerable attention within the industry, leading to various studies building upon this methodology [6]. + +3D-GS finds application in SLAM systems [23], scene editing [7] and segmentation [4]. Efforts [20] have also been made to integrate 3D-GS with popular diffusion models [14, 34, 35] and dynamic scenes [21]. Recent endeavors aim to enhance the robustness in sparse-view scenarios [38, 40], performance, storage efficiency [27] and mesh reconstruction [12] of 3D-GS. However, these improvements do not specifically address errors associated with Gaussian projection. Potential methods for enhancing Gaussian functions include [30, 42]. Nevertheless, they introduce a Gaussian filter and a novel Gaussian depth sorting algorithm, respectively, without investigating the projection function and associated errors. These approaches still rely on projecting onto the $z = 1$ plane. + +This paper aims to explore the potential of 3D Gaussian Splatting by analyzing errors that may arise during the projection process. The analysis will delve into the factors contributing to these errors and propose methods to mitigate and reduce them. + +# 3 Preliminaries + +Similar to NeRF [25], the input to 3D-GS consists of a set of images, together with the corresponding cameras calibrated by SFM [31]. However, in contrast to NeRF, 3D-GS takes the sparse point cloud generated during the SFM calibration process as input. From these points, it constructs a set of 3D Gaussians as graphical primitives explicitly representing the scene. Each 3D Gaussian is characterized by its position (mean) $\pmb{\mu}$ , covariance matrix $\pmb{\Sigma}$ , and opacity $\alpha$ , and carries anisotropic spherical harmonics (SH) representing directional appearance component (color) of the radiance field, following standard practice [8,26]. Subsequently, these 3D Gaussians are projected onto the image plane ( $z = 1$ plane) via a projection function for differentiable rasterization. Next, we will provide a detailed explanation of the Gaussian projection process in 3D-GS. + +In the world space, we define a 3D Gaussian function by $G$ , characterized by its mean $\mu$ and covariance matrix $\pmb{\Sigma}$ . To project the Gaussian onto the image plane, the initial step in 3D-GS involves an affine transformation of this Gaussian function from the world space to the camera space via a viewing transformation matrix $\mathbf{W}$ , yielding a transformed Gaussian function: + +$$ +G (\mathbf {x}) = \exp \left\{- 1 / 2 (\mathbf {W x} - \mathbf {W} \boldsymbol {\mu}) ^ {\top} (\mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {\top}) ^ {- 1} (\mathbf {W x} - \mathbf {W} \boldsymbol {\mu}) \right\}. \qquad (1) +$$ + +We denote $G'$ as the 3D Gaussian in the camera space, with $\mathbf{x}' = \mathbf{W}\mathbf{x}$ , $\boldsymbol{\mu}' = \mathbf{W}\boldsymbol{\mu}$ and $\boldsymbol{\Sigma}' = \mathbf{W}\boldsymbol{\Sigma}\mathbf{W}^\top$ as the point, mean and covariance matrix in the camera space, respectively. + +Subsequently, traditional 3D-GS requires projecting the Gaussian function in the camera space onto the $z = 1$ plane, with the projection function $\varphi$ defined as: + +$$ +\varphi \left(\mathbf {x} ^ {\prime}\right) = \mathbf {x} ^ {\prime} \left(\mathbf {x _ {0}} ^ {\top} \mathbf {x} ^ {\prime}\right) ^ {- 1} \left(\mathbf {x _ {0}} ^ {\top} \mathbf {x _ {0}}\right) = \mathbf {x} ^ {\prime} \left(\mathbf {x _ {0}} ^ {\top} \mathbf {x} ^ {\prime}\right) ^ {- 1} \tag {2} +$$ + +where $\mathbf{x_0} = [0,0,1]^\top$ represents the projection of the camera space's origin onto this plane. Expanding this projection function to the first order using Taylor series yields: + +$$ +\begin{array}{l} \varphi \left(\mathbf {x} ^ {\prime}\right) = \varphi \left(\boldsymbol {\mu} ^ {\prime}\right) + \frac {\partial \varphi}{\partial \mathbf {x} ^ {\prime}} \left(\boldsymbol {\mu} ^ {\prime}\right) \left(\mathbf {x} ^ {\prime} - \boldsymbol {\mu} ^ {\prime}\right) + R _ {1} \left(\mathbf {x} ^ {\prime}\right) \tag {3} \\ \approx \varphi (\boldsymbol {\mu} ^ {\prime}) + \frac {\partial \varphi}{\partial \mathbf {x} ^ {\prime}} (\boldsymbol {\mu} ^ {\prime}) (\mathbf {x} ^ {\prime} - \boldsymbol {\mu} ^ {\prime}) \\ \end{array} +$$ + +where $\frac{\partial\varphi}{\partial\mathbf{x}'}\left(\boldsymbol{\mu}'\right)$ is the Jacobian of the affine approximation of the projective transformation, denoted as $\mathbf{J}$ . Applying this local affine approximation by neglecting the Taylor remainder term allows to derive the 2D Gaussian function $G_{2\mathrm{D}}$ projected onto the $z = 1$ plane as: + +$$ +\begin{array}{l} G _ {2 D} \left(\mathbf {x} ^ {\prime}\right) = \exp \left\{- 1 / 2 \left(\mathbf {J} \mathbf {x} ^ {\prime} - \mathbf {J} \boldsymbol {\mu} ^ {\prime}\right) ^ {\top} \left(\mathbf {J} \boldsymbol {\Sigma} ^ {\prime} \mathbf {J} ^ {\top}\right) ^ {- 1} \left(\mathbf {J} \mathbf {x} ^ {\prime} - \mathbf {J} \boldsymbol {\mu} ^ {\prime}\right) \right\} \\ \approx \exp \left\{- 1 / 2 \left(\varphi \left(\mathbf {x} ^ {\prime}\right) - \varphi \left(\boldsymbol {\mu} ^ {\prime}\right)\right) ^ {\top} \left(\mathbf {J} \boldsymbol {\Sigma} ^ {\prime} \mathbf {J} ^ {\top}\right) ^ {- 1} \left(\varphi \left(\mathbf {x} ^ {\prime}\right) - \varphi \left(\boldsymbol {\mu} ^ {\prime}\right)\right) \right\}. \tag {4} \\ \end{array} +$$ + +Since the rank of the matrix $\mathbf{J}$ is 2, the inverse of $\mathbf{J} \pmb{\Sigma}' \mathbf{J}^\top$ is, in fact, the inverse of the covariance matrix of the 2D Gaussian obtained by skipping the third row and column [44]. We denote $\mathbf{x}_{2\mathrm{D}} = \varphi(\mathbf{x}')$ , $\pmb{\mu}_{2\mathrm{D}} = \varphi(\pmb{\mu}')$ and $\pmb{\Sigma}_{2\mathrm{D}} = \mathbf{J} \pmb{\Sigma}' \mathbf{J}^\top$ as the point, mean and $2 \times 2$ covariance matrix of $G_{2\mathrm{D}}$ in the image space, respectively. For a pixel $(u, v)$ , we obtain the value of the 2D Gaussian function by querying $\mathbf{x}_{2\mathrm{D}} = [u, v]^\top$ , and then perform alpha blending to derive the color of the pixel. The blue box in Figure 3 illustrates the projection process. + +# 4 Local Affine Approximation Error + +From Equation 3 and Equation 4 we observe that 3D Gaussian Splatting introduces an approximation during the projection transformation, i.e. the local affine approximation. Consequently, 3D-GS actually adopts a 2D Gaussian function that is not the true projection of the initial 3D Gaussian. The error introduced by this approximation can be characterized by the Taylor remainder term in Equation 3: + +$$ +R _ {1} \left(\mathbf {x} ^ {\prime}\right) = \varphi \left(\mathbf {x} ^ {\prime}\right) - \varphi \left(\boldsymbol {\mu} ^ {\prime}\right) - \frac {\partial \varphi}{\partial \mathbf {x} ^ {\prime}} \left(\boldsymbol {\mu} ^ {\prime}\right) \left(\mathbf {x} ^ {\prime} - \boldsymbol {\mu} ^ {\prime}\right) \tag {5} +$$ + +where + +$$ +\frac {\partial \varphi}{\partial \mathbf {x} ^ {\prime}} \left(\boldsymbol {\mu} ^ {\prime}\right) = \mathbb {I} \otimes \left(\mathbf {x} _ {0} ^ {\top} \boldsymbol {\mu} ^ {\prime}\right) ^ {- 1} - \boldsymbol {\mu} ^ {\prime} \left(\mathbf {x} _ {0} ^ {\top} \boldsymbol {\mu} ^ {\prime}\right) ^ {- 1} \left(\boldsymbol {\mu} ^ {\prime \top} \mathbf {x} _ {0}\right) ^ {- 1} \mathbf {x} _ {0} ^ {\top}. \tag {6} +$$ + +$\mathbb{I}$ represents the identity matrix and $\otimes$ denotes the multiplication of a matrix and a scalar. Clearly, this Taylor remainder term is a three-dimensional vector related to the random variable $\mathbf{x}^{\prime}$ and the mean $\pmb{\mu}^{\prime}$ of the Gaussian function. Therefore, by computing the square of the Frobenius norm of this vector $\| R_1(\mathbf{x}')\| _F^2$ and taking the mathematical expectation of this norm function with respect to the random variable $\mathbf{x}^{\prime}$ , we obtain an error function that depends solely on the Gaussian mean $\pmb{\mu}^{\prime}$ : + +$$ +\epsilon \left(\boldsymbol {\mu} ^ {\prime}\right) = \int_ {\mathbf {x} ^ {\prime} \in \mathcal {X} ^ {\prime}} \| R _ {1} \left(\mathbf {x} ^ {\prime}\right) \| _ {F} ^ {2} \mathrm {d} \mathbf {x} ^ {\prime}. \tag {7} +$$ + +The simplified error function $\epsilon$ involves three unit vectors: $\mathbf{x}^{\prime}$ , $\mathbf{x_0}$ and $\pmb{\mu}^{\prime}$ . Their spherical coordinates are given as follows: + +$$ +\mathbf {x} _ {\mathbf {0}} = \left[ \begin{array}{c} \sin (\phi_ {0}) \cos (\theta_ {0}) \\ - \sin (\theta_ {0}) \\ \cos (\phi_ {0}) \cos (\theta_ {0}) \end{array} \right], \mathbf {x} ^ {\prime} = \left[ \begin{array}{c} \sin (\phi) \cos (\theta) \\ - \sin (\theta) \\ \cos (\phi) \cos (\theta) \end{array} \right], \boldsymbol {\mu} ^ {\prime} = \left[ \begin{array}{c} \sin (\phi_ {\mu}) \cos (\theta_ {\mu}) \\ - \sin (\theta_ {\mu}) \\ \cos (\phi_ {\mu}) \cos (\theta_ {\mu}) \end{array} \right] \tag {8} +$$ + +where $\phi_0 = 0$ and $\theta_0 = 0$ . Please refer to the detailed proofs in the supplementary materials. + +The size of the integration region is related to the covariance of the Gaussian function. Since the current discussion specifically focuses on the impact of the mean on the error rather than the variance, let's assume the integration region to be $\left\{\mathbf{x}' \mid \theta \in [-\pi/4 + \theta_{\mu}, \pi/4 + \theta_{\mu}] \land \phi \in [-\pi/4 + \phi_{\mu}, \pi/4 + \phi_{\mu}]\right\}$ . Then, based on Equation 5-8, we obtain: + +$$ +\begin{array}{l} \epsilon (\theta_ {\mu}, \phi_ {\mu}) = \int_ {- \pi / 4 + \theta_ {\mu}} ^ {\pi / 4 + \theta_ {\mu}} \int_ {- \pi / 4 + \phi_ {\mu}} ^ {\pi / 4 + \phi_ {\mu}} \left(- \frac {\sin (\phi - \phi_ {\mu}) \cos (\theta)}{\cos^ {2} (\phi_ {\mu}) \cos (\theta_ {\mu})} + \tan (\phi) - \tan (\phi_ {\mu})\right) ^ {2} + \\ \left(\frac {\sin (\theta)}{\cos (\phi_ {\mu}) \cos (\theta_ {\mu})} - \frac {\sin (\theta_ {\mu}) \cos (\phi) \cos (\theta)}{\cos^ {2} (\phi_ {\mu}) \cos^ {2} (\theta_ {\mu})} + \frac {\tan (\theta_ {\mu})}{\cos (\phi_ {\mu})} - \frac {\tan (\theta)}{\cos (\phi)}\right) ^ {2} d \theta d \phi . \tag {9} \\ \end{array} +$$ + +Fortunately, this integral has a close-form expression, but its form is very complicated. We report the close-form expression in the supplemental materials and visually depict it in Figure 2. + +The error function is partially differentiated with respect to $\theta_{\mu}$ and $\phi_{\mu}$ . It is easy to check that at $\theta_{\mu} = \theta_0 = 0$ and $\phi_{\mu} = \phi_0 = 0$ , the following holds: + +$$ +\frac {\partial \epsilon}{\partial \theta_ {\mu}} (0, 0) = 0, \frac {\partial \epsilon}{\partial \phi_ {\mu}} (0, 0) = 0. \tag {10} +$$ + +![](images/5b6c085bd2f8652cdb99fcf863d9d73ed435b3c42003556c3f27351eb39cf9b3.jpg) +(a) $\lambda = 0.095$ + +![](images/361d2ffa91e732c015aff3b5487d1082d3f5f19b2e2fb35714ee61dc1642f9f4.jpg) +(b) $\lambda = 0.95$ +Fig. 2: Visualization of the 3D Gaussian Splitting error function $\epsilon (\theta_{\mu},\phi_{\mu})$ under two distinct domains. The domain of the function is $\{(\theta_{\mu},\phi_{\mu})\mid \theta_{\mu}\in [-\lambda \pi /4,\lambda \pi /4]\land \phi_{\mu}\in [-\lambda \pi /4,\lambda \pi /4]\}$ where $(\theta_{\mu},\phi_{\mu})$ represents the polar coordinate of the 3D Gaussian mean and $\lambda$ represents the scaling factor of the domain. + +Therefore, this point is actually an extremum point of the function. We also confirm that this point is the minimum of the function. Indeed, from Figure 2a, it is also apparent that the function is concave, with its minimum value occurring at $(0,0)$ . Additionally, it can be observed that the minimum value of the error function is greater than zero, due to the fact that Gaussian functions are not closed under projection transformations. Figure 2b shows that in the majority of the region close to the origin, the function values vary very gently. However, as the independent variable of the function approaches the boundaries, the function values increase rapidly, resulting in a substantial difference between the maximum and minimum values. + +The above analysis indicates that the error is small and does not significantly affect the quality of rendered images. Therefore, traditional 3D-GS [16] utilizes local affine approximation but still successfully reconstructs the scene to obtain a high-quality novel view image. Nevertheless, the naive projection of all Gaussians onto the same plane $z = 1$ in 3D-GS may lead to larger projection errors for Gaussians farther from the plane center (the projection point of the camera center to the plane), causing artifacts. These artifacts are severe when using wide-angle lenses, significantly degrading the quality of the renderings. + +# 5 Optimal Gaussian Splatting + +Based on the analysis of the error function, we introduce Optimal Gaussian Splitting shown in Figure 3. This helps us to achieve smaller projection errors, resulting in higher-quality rendering. The first step, similar to the original 3D-GS [16], involves transforming world coordinates into the camera coordinate system. Instead of projecting Gaussians onto the $z = 1$ plane, each Gaussian is radially projected based on its mean along the line connecting it to the camera + +![](images/9916bebfb6385cf294cad8405418c9dfc2eb84de32f010cd889c9549cf88c2c2.jpg) +Fig. 3: Illustration of the rendering pipeline for our Optimal Gaussian Splitting and the projection of 3D-GS [16]. The blue box depicts the projection process of the original 3D-GS, which straightforwardly projects all Gaussians onto the same projection plane. In contrast, the red box illustrates our approach, where we project individual Gaussians onto corresponding tangent planes. + +![](images/2ab75e9dba7ebfc00a5f7e92503697fbeb0b2589e5f28bcd52833218da8eb61f.jpg) + +center, projecting onto a plane tangent to the unit sphere and perpendicular to this line. After projection, colors of points on the unit sphere are computed by alpha blending the 2D Gaussians on the tangent plane. Finally, for each pixel in the image, rays are cast onto the unit sphere to retrieve the color of that pixel, resulting in the rendered image. + +# 5.1 Optimal Projection + +We discover that the error function attains its minimum value when the projection of the Gaussian mean on the plane coincides with the projection from the camera center to the plane. Therefore, in Optimal Gaussian Splatting, we employ an optimal projection method. Specifically, instead of naively projecting different Gaussians onto the same plane $z = 1$ , we adopt distinct projection planes for each Gaussian. These projection planes are determined by the tangent planes formed by the Gaussian mean and the line connecting it to the camera center. The specific projection plane is formulated as: + +$$ +\mathbf {x} _ {\mathbf {p}} ^ {\top} \cdot \left(\mathbf {x} ^ {\prime} - \mathbf {x} _ {\mathbf {p}}\right) = 0 \tag {11} +$$ + +where + +$$ +\mathbf {x} _ {\mathbf {p}} = \varpi \left(\boldsymbol {\mu} ^ {\prime}\right) = \boldsymbol {\mu} ^ {\prime} \left(\boldsymbol {\mu} ^ {\prime \top} \boldsymbol {\mu} ^ {\prime}\right) ^ {- 1 / 2} \tag {12} +$$ + +represents the projection of the camera space's origin onto this plane and $\varpi$ function represents the transformation that projects points onto the unit sphere. According to Equation 11, the optimal projection function $\varphi_{\mathrm{p}}$ is obtained as: + +$$ +\varphi_ {\mathrm {p}} \left(\mathbf {x} ^ {\prime}\right) = \mathbf {x} ^ {\prime} \left(\mathbf {x} _ {\mathbf {p}} ^ {\top} \mathbf {x} ^ {\prime}\right) ^ {- 1} \left(\mathbf {x} _ {\mathbf {p}} ^ {\top} \mathbf {x} _ {\mathbf {p}}\right) = \mathbf {x} ^ {\prime} \left(\mathbf {x} _ {\mathbf {p}} ^ {\top} \mathbf {x} ^ {\prime}\right) ^ {- 1}. \tag {13} +$$ + +The corresponding local affine approximation Jacobian matrix $\mathbf{J}_{\mathbf{p}}$ is: + +$$ +\mathbf {J} _ {\mathbf {p}} = \frac {\partial \varphi_ {\mathrm {p}}}{\partial \mathbf {x} ^ {\prime}} \left(\boldsymbol {\mu} ^ {\prime}\right) = \mathbb {I} \otimes \left(\mathbf {x} _ {\mathbf {p}} ^ {\top} \boldsymbol {\mu} ^ {\prime}\right) ^ {- 1} - \boldsymbol {\mu} ^ {\prime} \left(\mathbf {x} _ {\mathbf {p}} ^ {\top} \boldsymbol {\mu} ^ {\prime}\right) ^ {- 1} \left(\boldsymbol {\mu} ^ {\prime \top} \mathbf {x} _ {\mathbf {p}}\right) ^ {- 1} \mathbf {x} _ {\mathbf {p}} ^ {\top}. \tag {14} +$$ + +When $\pmb{\mu}^{\prime} = [\mu_{x},\mu_{y},\mu_{z}]^{\top}$ , the specific form of this Jacobian matrix is: + +$$ +\mathbf {J} _ {\mathbf {p}} = \left[ \begin{array}{l l l} \frac {\mu_ {y} ^ {2} + \mu_ {z} ^ {2}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\frac {3}{2}}} & - \frac {\mu_ {x} \mu_ {y}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\frac {3}{2}}} & - \frac {\mu_ {x} \mu_ {z}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\frac {3}{2}}} \\ - \frac {\mu_ {x} \mu_ {y}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\frac {3}{2}}} & \frac {\mu_ {x} ^ {2} + \mu_ {z} ^ {2}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\frac {3}{2}}} & - \frac {\mu_ {y} \mu_ {z}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\frac {3}{2}}} \\ - \frac {\mu_ {x} \mu_ {z}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\frac {3}{2}}} & - \frac {\mu_ {y} \mu_ {z}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\cfrac {3}{2}}} & \frac {\mu_ {x} ^ {2} + \mu_ {y} ^ {2}}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}\right) ^ {\cfrac {3}{2}}} \end{array} \right]. \tag {15} +$$ + +It's very easy to implement modifications to the Jacobian matrix in the forward process. + +# 5.2 Unit Sphere Based Rasterizer + +For the proposed Optimal Projection, we have proposed a Unit Sphere Based Rasterizer to generate images. Through Optimal Projection, we obtain the projection of the three-dimensional Gaussians on the tangent plane of the unit sphere instead of obtaining the Gaussians in the image space. Therefore, we need to rasterize based on this unit sphere. + +Specifically, for a pixel $(u,v)$ on the image, similar to NeRF, we cast a ray. However, unlike NeRF, we do not involve the extensive sampling of points along the ray, which would significantly degrade performance. Instead, our focus is solely on determining which tangent-plane Gaussians the ray intersects on the unit sphere: + +$$ +\mathbf {x} _ {\mathrm {2 D}} = \varphi_ {\mathrm {p}} \left(\left[ \begin{array}{c} (u - c _ {x}) / f _ {x} \\ (v - c _ {y}) / f _ {y} \\ 1 \end{array} \right]\right) \tag {16} +$$ + +where $c_x, c_y, f_x, f_y$ denote the intrinsic parameters of the pinhole camera. We then query the function values of these Gaussians for alpha blending to obtain the color + +$$ +G _ {2 \mathrm {D}} \left(\mathbf {x} _ {2 \mathrm {D}}\right) = \exp \left\{- 1 / 2 \left(\mathbf {x} _ {2 \mathrm {D}} - \varphi_ {\mathrm {p}} \left(\boldsymbol {\mu} ^ {\prime}\right)\right) ^ {\top} \left(\mathbf {J} _ {\mathbf {p}} \boldsymbol {\Sigma} ^ {\prime} \mathbf {J} _ {\mathbf {p}} ^ {\top}\right) ^ {- 1} \left(\mathbf {x} _ {2 \mathrm {D}} - \varphi_ {\mathrm {p}} \left(\boldsymbol {\mu} ^ {\prime}\right)\right) \right\}. \tag {17} +$$ + +Note that both $\mathbf{J}_{\mathbf{p}}$ in Equation 14 and $\mathbf{J}$ in Equation 6 are matrices with a rank of 2. Consequently, the inverse of the covariance matrix of the 2D Gaussian is also obtained by skipping the third row and column. However, the third row and third column of $\mathbf{J}_{\mathbf{p}}\boldsymbol{\Sigma}'\mathbf{J}_{\mathbf{p}}^\top$ are not entirely zeros, as opposed to $\mathbf{J}\boldsymbol{\Sigma}'\mathbf{J}^\top$ . To ensure the equation still holds, an invertible matrix $\mathbf{Q}$ needs to be used for left multiplication with $\mathbf{x}_{2\mathrm{D}}$ , $\varphi_{\mathrm{p}}(\boldsymbol{\mu}')$ and $\mathbf{J}_{\mathbf{p}}$ : + +$$ +\mathbf {Q} = \left[ \begin{array}{c c c} \frac {\mu_ {z}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {z} ^ {2}}} & 0 & - \frac {\mu_ {x}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {z} ^ {2}}} \\ - \frac {\mu_ {x} \mu_ {y}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {z} ^ {2}} \sqrt {\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}}} & \frac {\sqrt {\mu_ {x} ^ {2} + \mu_ {z} ^ {2}}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}}} - \frac {\mu_ {y} \mu_ {z}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {z} ^ {2}} \sqrt {\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}}} \\ \frac {\mu_ {x}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}}} & \frac {\mu_ {y}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}}} & \frac {\mu_ {z}}{\sqrt {\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + \mu_ {z} ^ {2}}} \end{array} \right]. \tag {18} +$$ + +Since our projection does not rely on the perspective image plane $z = 1$ as in the original 3D-GS [16], we can adapt to different camera models such as fisheye cameras and panoramic images by simply modifying the transformation from image space to camera space, as described in Equation 16. + +# 6 Experiments + +To validate the effectiveness of the Optimal Gaussian Splatting derived from theoretical error analysis, we conduct a series of experiments, comparing it with the original 3D-GS [16] and some current state-of-the-art methods. + +# 6.1 Implementation + +We implemented Optimal Gaussian Splitting based on the PyTorch framework in 3D-GS [16] and wrote custom CUDA kernels for rasterization. We used the default parameters of 3D-GS to maintain consistency with the original 3D-GS and prevent other factors from introducing interference into the results. + +# 6.2 Datasets + +We test our algorithm on a total of 13 real scenes which are the same as those used in the original 3D-GS [16]. In particular, we evaluate our approach on the complete set of scenes featured in Mip-NeRF360 [2], which currently represents the state-of-the-art in NeRF rendering quality. Additionally, we test our method on two scenes from the Tanks & Temples dataset [17] and two scenes provided by Hedman et al. [13]. The selected scenes exhibit diverse capture styles, encompassing both confined indoor environments and expansive, unbounded outdoor settings. + +# 6.3 Results + +We compare our algorithm with 3D-GS [16] and three state-of-the-art NeRF-based approaches: Mip-NeRF360 [2], InstantNGP [26] and Plenoxels [8]. All methods are configured using the same settings as outlined in the 3D-GS [16] to control variables, preventing the introduction of variables other than our Optimal Splatting method. + +Quantitative comparisons We adopt a train/test split for datasets following the methodology proposed by Mip-NeRF360 [2]. Specifically, every 8th photo is reserved for testing, ensuring a consistent and meaningful basis for comparisons to generate error metrics. Standard metrics such as PSNR, LPIPS [43], and SSIM are employed for evaluation. We report quantitative results in Table 1. While the 3D-GS significantly outperforms MipNeRF360 in rendering performance, its rendering quality slightly falls behind MipNeRF360, with a lower average PSNR across the 13 scenes. This might be attributed to MipNeRF360 not employing + +Table 1: Quantitative evaluation of our method compared to previous work across three datasets. The Mip-NeRF360 [2] dataset includes both indoor and outdoor scenarios. Tanks. and Deep. indicates the average results over the Tanks&Templates [17] and Deep Blending [13] datasets. Average represents the mean value across all scenarios in these datasets. + +
Dataset Method|MetricAverageMip-NeRF360Tanks. and Deep.
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Plenoxels [8]22.770.6660.45723.080.6250.46322.070.7570.444
INGP-Base [26]24.490.6990.37325.300.6710.37122.670.7600.377
INGP-Big [26]24.930.7240.33625.590.6990.33123.440.7810.347
M-NeRF360 [2]27.110.8030.24127.690.7920.23725.810.8300.251
3D-GS [16]26.920.8320.21427.210.8150.21426.270.8720.213
Ours27.170.8360.21027.480.8210.20926.440.8720.214
+ +![](images/d8afd8b9cbf9090cd121a3e0a2c14d26a85ea766d67fa45af51a6b7a8b85ef67.jpg) +Fig. 4: We show comparisons of our method to previous methods and the corresponding ground truth images from held-out test views. The scenes are, from the top down: TRUCK from Tanks&Temples [17]; PLAYROOM from the Deep Blending dataset [13] and BONSAI, COUNTER from Mip-NeRF360 dataset [2]. Differences in quality highlighted by arrows/insets. + +significant approximations throughout its entire pipeline. However, by replacing the original projection method of 3D-GS with our optimal projection, which incurs smaller errors, we not only maintain rendering performance far surpassing MipNeRF360 but also achieve a quality improvement. + +Qualitative comparisons For our qualitative experiments, we select only two methods, 3D-GS and MipNeRF360, as our baselines. It can be observed that + +![](images/3a9051b9664d848d67240f7def0e83431f2673daa8e62b3d249915406c2fd0f5.jpg) +Fig. 5: We derive the projection error function with respect to image coordinates $(u, v)$ and focal length through the transformation between image and polar coordinates $(\theta_{\mu}, \phi_{\mu})$ . We visualize this function, showing 3D-GS rendered images with long and short focal lengths, followed by the corresponding error functions. + +our method is capable of generating more realistic details, with fewer artifacts compared to 3D-GS [16]. And this is precisely the superiority brought about by our projection method, which results in smaller errors. Furthermore, We see that in some cases Mip-NeRF360 has remaining artifacts that our method avoids (e.g., black lines - in PLAYROOM). + +# 6.4 Discussions + +Impacts of Decreasing Focal Length Based on Equation 9, we further derive the projection error function with respect to image coordinates $(u,v)$ and focal length $f_{x},f_{y}$ through the coordinate transformation between image coordinates and polar coordinates $(\theta_{\mu},\phi_{\mu})$ . As illustrated in Figure 5, the peak of projection error mainly occurs at the image edges under a long focal length, resulting in artifacts at the image edges. As the focal length decreases, the field of view expands, leading to more Gaussians deviating from the projection center and an overall increase in projection error. In such cases, 3D-GS exhibit more artifacts, such as needle-like structures or cloud-like large Gaussians, which obscure parts that perform well under long focal length, thereby significantly degrading the overall image quality. Since our method employs a central radial projection for the projection plane, theoretically, we should not encounter such defects. To validate this perspective, we conduct experiments at various focal length reduction ratios (reduction ratios of $\times 0.2$ and $\times 0.3$ ). + +For quantitative experiments, considering that there is no ground truth for large field-of-view in the 13 scenes from the 3D-GS [16], we apply a focal length mask. This mask selects central parts of the rendered images with corresponding ground truth based on the focal length scaling factor for metric calculation. The results are presented in the Table 2. Whether it is at $\times 0.2$ or $\times 0.3$ reduction ratios, our method outperforms the original 3D-GS significantly across various scenes in terms of all metrics. As the focal length decreases from $\times 0.3$ to $\times 0.2$ , 3D-GS [16] experiences a significant degradation due to increased projection errors, while our method demonstrates greater robustness. + +For quantitative experiments, we randomly select three scenes, each with two different focal lengths for demonstration. Additional experimental results can be found in the supplementary materials. It is evident from the Figure 6 + +Table 2: We quantitatively compare our method with the original 3D-GS [16] using focal lengths of 0.2 and 0.3 times the original. Six scenes were selected from three datasets: Mip-NeRF360 [2], Tank & Templates [17], and Deep Blending [13]. The scenes include TRAIN and TRUCK from Tanks & Templates, DRJOHNSON and PLAYROOM from Deep Blending, and ROOM and TREEHILL from Mip-NeRF360. + +
MetricFocalMethodTrainTruckDrJohnsonPlayroomRoomTreehillAverage
PSNR↑×0.23D-GS [16]10.1212.3919.2416.3820.0215.3615.58
Ours16.6016.7023.5423.2923.6718.9820.46
×0.33D-GS [16]12.8514.8923.3622.5125.0018.3519.49
Ours17.5218.0724.8926.4025.8719.8722.10
SSIM↑×0.23D-GS [16]0.3250.3800.6160.6190.6800.3710.499
Ours0.5650.5370.7010.7520.7770.4350.628
×0.33D-GS [16]0.4710.5290.7400.7910.8120.4370.630
Ours0.5880.6050.7620.8440.8340.4720.684
LPIPS↓×0.23D-GS [16]0.5690.4790.2960.3550.2070.4740.397
Ours0.2770.2500.2160.2030.1780.3850.251
×0.33D-GS [16]0.3310.2760.2070.2130.1540.3970.263
Ours0.2430.2170.2010.1820.1580.3850.231
+ +![](images/1268bf6aebcbd50d49bf512bc15f8bfaefad5b4482ba126462629acc764d7e79.jpg) + +![](images/806d986099681f6ef4726210ae9e4b9c5b8f115a8eac887c480f830699b21553.jpg) + +![](images/a570e1470766f89c58851a3a7d14140da7e461252c3df58833133d4be2898dc1.jpg) + +![](images/82604c3b8841aab712b2d98ac7fe985456453319cead0b3b56f272a9d7d3caec.jpg) + +![](images/6e8aaebb3966e9a3c5863fb4f05f0aa1e7c136011190b7bc6bae970af8aeebc4.jpg) + +![](images/1f224a3d72de997304df760b78a94a2e419c07838d97bbc0f03dc4a6a018f810.jpg) + +![](images/a663eea222528cd1f488ef9d04e4ad9b4b4a97d33efccbc9d731945a22ba54e8.jpg) + +![](images/3c71e304e7fbcfb6e8e0394be300935b2c64df261c1cc52571831c561dcb34df.jpg) + +![](images/53297f4f353236a57052531f0f138eb92184c2fae7608bc9d3d459f636185bd5.jpg) + +![](images/12ddc4272ca0c2cb88a52e98376d3d02b2ac9a5919e06c5b2738a9e379d44702.jpg) + +![](images/f044836e2f37e1982f5475183f90501bd25da44c4e7fccbe899100087fb8db94.jpg) +Fig. 6: We compare our method to the original 3D-GS [16] with different focal length reductions. Two focal lengths, 0.2 and 0.3 times the original, are tested. Ground truth $\times 1$ in the left images indicates no focal length reduction. The red-boxed areas in the right images correspond to the same regions in the left images. + +![](images/b1286fd98708d78e2d6a03ba5bffd24b1279568d4a5344f63d6782de1b2b3b43.jpg) + +![](images/19ff1f611e4ac50d9b297c394b68d04779d83f7718904e0acedccf2288b926d7.jpg) + +![](images/a5252c499c4ad2f162536b50871bcbfe809e42822428f03066a981ab68b4c684.jpg) + +![](images/e322cc1fa21d192bc0afe8cd143d1844a7593c9317e9b522a10e12e2b598ee0d.jpg) + +that our method exhibits greater robustness across various focal length settings, mitigating the artifacts introduced by projection in the original 3D-GS. + +Adaptability to Various Camera Models By straightforwardly modifying the transformation from image space to camera space, we have achieved adaptation to various camera models, with the caveat that the original 3D-GS's projection struggles to support these camera models. Figure 7 illustrates that our projection, in addition to conforming to the original pinhole camera according to Equation 16, can also accommodate fisheye cameras and generate panoramas. + +![](images/1021eeabfb4dcb03869c0a00a9f5eb44804a6cb59c16fbaac84f6936b59e9391.jpg) +Fig. 7: Renderings of three scenes using our projection under three camera models. From left to right: the ground truth of the original perspective, wide-angle rendering with a pinhole camera, fisheye camera rendering, and the panorama. + +**Limitation** Due to our method's independence from the image plane, it exhibits adaptability to various camera models. However, the slight increase in training time is a consequence of the subsequent transformation to the image plane after projection onto the tangent plane, rather than a direct projection onto the image plane. Fortunately, this method avoids the need for dense points sampling, ensuring the algorithm's time complexity remains unchanged and rendering performance stays unaffected. Moreover, it may be possible to mitigate the impact on training time, particularly concerning the corresponding backward propagation, through the implementation of optimized CUDA code. + +# 7 Conclusion + +Through theoretical analysis, we establish the correlation between the projection error and the Gaussian mean in 3D-GS. By optimizing the error function, we discover that the projection error is minimized when projecting on the plane perpendicular to the line connecting the Gaussian mean and the camera center. Leveraging this insight, we propose a novel projection that yields high-quality images without compromising rendering performance and can be easily adapted to various camera models through simple modifications. + +In this work, our primary focus is on analyzing the relationship between the projection error and the mean of the 3D Gaussian under the assumption of a constant covariance. The influence of Gaussian covariance on projection requires further discussion in future work. It may be possible to further explore the potential of Gaussian Splatting as an explicit representation technique for scenes. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5835-5844. IEEE (2021) +2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +3. Buehler, C., Bosse, M., McMillan, L., Gortler, S.J., Cohen, M.F.: Unstructured lumigraph rendering. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2001, Los Angeles, California, USA, August 12-17, 2001. pp. 425-432. ACM (2001). https://doi.org/10.1145/383259.383309, https://doi.org/10.1145/383259.383309 +4. Cen, J., Fang, J., Yang, C., Xie, L., Zhang, X., Shen, W., Tian, Q.: Segment any 3d gaussians. arXiv preprint arXiv:2312.00860 (2023) +5. Chaurasia, G., Duchêne, S., Sorkine-Hornung, O., Drettakis, G.: Depth synthesis and local warps for plausible image-based navigation. ACM Trans. Graph. 32(3), 30:1-30:12 (2013). https://doi.org/10.1145/2487228.2487238, https://doi.org/10.1145/2487228.2487238 +6. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024) +7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21476-21485 (2024) +8. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +9. Genova, K., Cole, F., Sud, A., Sarna, A., Funkhouser, T.: Local deep implicit functions for 3d shape. In: CVPR (2020) +0. Goesele, M., Snavely, N., Curless, B., Hoppe, H., Seitz, S.M.: Multi-view stereo for community photo collections. In: IEEE 11th International Conference on Computer Vision, ICCV 2007, Rio de Janeiro, Brazil, October 14-20, 2007. pp. 1-8. IEEE Computer Society (2007). https://doi.org/10.1109/ICCV.2007.4408933, https://doi.org/10.1109/ICCV.2007.4408933 +1. Gortler, S.J., Grzeszczuk, R., Szeliski, R., Cohen, M.F.: The lumigraph. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 43-54. ACM (1996). https://doi.org/10.1145/237170.237200, https://doi.org/10.1145/237170.237200 +2. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2024) +3. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018) + +14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual (2020), https://proceedings.neurips.cc/paper/2020/bit/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html +15. Jiang, C., Sud, A., Makadia, A., Huang, J., Nießner, M., Funkhouser, T.: Local implicit grid representations for 3d scenes. In: CVPR (2020) +16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017) +18. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Trans. Graph. 41(6), 201:1-201:15 (2022). https://doi.org/10.1145/3550454.3555497, https://doi.org/10.1145/3550454.3555497 +19. Levoy, M., Hanrahan, P.: Light field rendering. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 31-42. ACM (1996). https://doi.org/10.1145/237170.237199, https://doi.org/10.1145/237170.237199 +20. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023) +21. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 2024 International Conference on 3D Vision (3DV). pp. 800-809. IEEE (2024) +22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00713 +23. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024) +24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR (2019) +25. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision (2020) +26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatting for accelerated novel view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10349-10358 (2024) +28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In: CVPR (2019) + +29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR (2019) +30. Radl, L., Steiner, M., Parger, M., Weinrauch, A., Kerbl, B., Steinberger, M.: StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics 4(43) (2024) +31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) +32. Sitzmann, V., Zollhoefer, M., Wetzstein, G.: Scene representation networks: Continuous 3D-structure-aware neural scene representations. In: NeurIPS (2019) +33. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. ACM Trans. Graph. 25(3), 835-846 (2006). https://doi.org/10.1145/1141911.1141964, https://doi.org/10.1145/1141911.1141964 +34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 11895-11907 (2019), https://proceedings.neurips.cc/paper/2019/bit/3001ef257407d5a371a96cdcd947c7d93-Abstraction.html +35. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=PxTIG12RRHS +36. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5449-5459. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00538, https://doi.org/10.1109/CVPR52688.2022.00538 +37. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C.T., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11358-11367. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01120 +38. Xiong, H., Muttukuru, S., Upadhyay, R., Chari, P., Kadambi, A.: Sparsegs: Real-time $360^{\circ}$ sparse view synthesis using gaussian splatting. arXiv e-prints pp. arXiv-2312 (2023) +39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5428-5438. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00536, https://doi.org/10.1109/CVPR52688.2022.00536 +40. Yang, C., Li, S., Fang, J., Liang, R., Xie, L., Zhang, X., Shen, W., Tian, Q.: Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv preprint arXiv:2402.10259 (2024) +41. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8254-8263 (2023) + +42. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +43. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +44. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE Transactions on Visualization and Computer Graphics 8(3), 223-238 (2002) \ No newline at end of file diff --git a/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/images.zip b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16ef14538665cec3c4fa75f1331635cabdfbd8a8 --- /dev/null +++ b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c910e3afac57736879919dc85d3e3b53c6f72b874779fa3fd97944475a87b90e +size 779065 diff --git a/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/layout.json b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eed259a949d8aaf2bd2e26bcec272e5f66f66847 --- /dev/null +++ b/2024/On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy/layout.json @@ -0,0 +1,9679 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 139, + 111, + 477, + 148 + ], + "type": "text", + "content": "On the Error Analysis of 3D Gaussian Splitting and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "spans": [ + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "text", + "content": "Letian Huang" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "inline_equation", + "content": "^{\\text{®}}" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "text", + "content": ", Jiayang Bai" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "inline_equation", + "content": "^{\\text{®}}" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "text", + "content": ", Jie Guo" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "inline_equation", + "content": "^{\\star \\text{®}}" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "text", + "content": ", Yuanqi Li" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "inline_equation", + "content": "^{\\text{®}}" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "text", + "content": ", and Yanwen Guo" + }, + { + "bbox": [ + 141, + 168, + 473, + 181 + ], + "type": "inline_equation", + "content": "^{\\text{®}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 190, + 476, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 476, + 224 + ], + "type": "text", + "content": "Nanjing University, Nanjing, China \n{1thuang, jybai, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn \nhttps://letianhuang.github.io/op43dgs/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 254, + 455, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 254, + 455, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 254, + 455, + 452 + ], + "type": "text", + "content": "Abstract. 3D Gaussian Splatting has garnered extensive attention and application in real-time neural rendering. Concurrently, concerns have been raised about the limitations of this technology in aspects such as point cloud storage, performance, and robustness in sparse viewpoints, leading to various improvements. However, there has been a notable lack of attention to the fundamental problem of projection errors introduced by the local affine approximation inherent in the splatting itself, and the consequential impact of these errors on the quality of photo-realistic rendering. This paper addresses the projection error function of 3D Gaussian Splatting, commencing with the residual error from the first-order Taylor expansion of the projection function. The analysis establishes a correlation between the error and the Gaussian mean position. Subsequently, leveraging function optimization theory, this paper analyzes the function's minima to provide an optimal projection strategy for Gaussian Splatting referred to Optimal Gaussian Splatting, which can accommodate a variety of camera models. Experimental validation further confirms that this projection methodology reduces artifacts, resulting in a more convincingly realistic rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "spans": [ + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "text", + "content": "Keywords: 3D deep learning " + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "text", + "content": " view synthesis " + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "text", + "content": " radiance fields " + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "text", + "content": " 3D gaussians " + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "text", + "content": " real-time rendering " + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 455, + 485 + ], + "type": "text", + "content": " error analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 525, + 230, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 230, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 230, + 537 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "Reconstructing 3D scenes from 2D images and synthesizing novel views has been a critical task in computer vision and graphics. Recently, the trend of novel view synthesis is spearheaded by Neural Radiance Fields (NeRFs) [25] and its variants [1, 2], which achieves photo-realistic rendering quality. NeRF represents scenes with Multi-Layer Perceptron (MLP) and renders novel views using volumetric ray-marching. However, these MLP-based methods contain a relatively large number of layers, leading to lengthy training and rendering times, which still fall significantly short of real-time rendering. To accelerate rendering," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 248, + 229 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 248, + 229 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 248, + 229 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 248, + 229 + ], + "type": "image", + "image_path": "77cd6b753a96e723145c87409706006a12e725f4bc39dbe589905d0347c0d3ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 282 + ], + "type": "text", + "content": "Fig. 1: By minimizing the projection error through error analysis, we have achieved an improvement in the rendering image quality compared to the original 3D-GS [16]. These images are cropped from the complete images of the TRAIN scene, in order to be highlighted." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 365, + 229 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 365, + 229 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 365, + 229 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 365, + 229 + ], + "type": "image", + "image_path": "297487479d90a6b06985f07bd69cf50747d4c68ffcc67fc9eaa3b04d6a8b7098.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 480, + 229 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 480, + 229 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 480, + 229 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 480, + 229 + ], + "type": "image", + "image_path": "1c66b8cd20ba5787cd72de9697c734ef0fd46c516a26b98b8a718c4f146f6237.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 479, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 479, + 342 + ], + "type": "text", + "content": "many methods resort to auxiliary data structures, such as hash grids [26], points or voxel [8,36,37,39]. However, due to the dense sampling of points along rays, they still face challenges in achieving real-time performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "text", + "content": "With a substantial demand for high-speed and photo-realistic rendering effects, 3D Gaussian Splatting (3D-GS) [16] has emerged as an efficient representation that can be rendered at high speed on the GPU. It has departed from the implicit scene representation using MLPs and instead opted for an explicit representation using Gaussian functions. As a result, 3D-GS avoids the expensive cost of densely sampling points in volume rendering, thus achieving real-time performance. Given a set of images with the camera parameters calibrated by SfM [31], 3D-GS aims to optimize a set of 3D Gaussians as graphical primitives to explicitly represent the scene. To render 2D images, these 3D Gaussians are projected onto the image plane (" + }, + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "text", + "content": " plane) via a projection function denoted as " + }, + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "text", + "content": " for differentiable rasterization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 582 + ], + "type": "text", + "content": "Unfortunately, Gaussian functions are closed under affine transformations but not under projective transformations. Therefore, 3D-GS adopts a local affine approximation [44], specifically approximating the projection function with the first two terms of its Taylor expansion. Nevertheless, approximations introduce errors, and the local affine approximation similarly contributes to these errors which may lead to artifacts in the rendered images. This paper exploits the relationship between the error of 3D-GS and the Gaussian mean through the analysis of the Taylor remainder term. Additionally, it identifies the circumstances under which the error is minimized by determining the extremum of the error function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Finally, based on the extremum analysis of the error function, we propose a novel optimal projection method to minimize the projection error in 3D-GS. This method requires only minor code modifications, does not affect real-time rendering performance, and yet achieves a significant improvement in rendering quality. Specifically, we project along the direction from the Gaussian mean to the camera center, where the projection plane is tangent to the line connecting the Gaussian mean and the camera center. Compared with the straightforward" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "approach of projecting 3D Gaussians onto the image plane of the camera, the proposed projection reduces expected error while maintaining real-time and realistic rendering performance. Furthermore, we find that the projection errors in 3D-GS tend to escalate with the continuous expansion of the field of view and the corresponding reduction in focal length. This increasing error leads to more and larger elongated Gaussians, as well as cloud-like Gaussian artifacts, consequently degrading the overall image quality. In contrast, our method exhibits greater realism and robustness. As illustrated in Figure 1, our method significantly reduces blurriness and artifacts caused by the projection error. Through experiments on various datasets and scenes, our method consistently outperforms the original projection method, particularly in settings with short focal lengths. And the method is easily implemented and adaptable to various camera models. To summarize, we provide the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 275, + 480, + 381 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 275, + 480, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 275, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 275, + 480, + 310 + ], + "type": "text", + "content": "- We provide thorough error analysis for 3D-GS that may lead to artifacts and consequently degrade rendering quality. We have identified the correlation between this error and the Gaussian position." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 310, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 334 + ], + "type": "text", + "content": "- We derive the mathematical expectation of this error function and analyze when this function takes extrema through methods of function optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 334, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 480, + 381 + ], + "type": "text", + "content": "- We propose a novel optimal projection capable of adapting to diverse camera models that adopts different tangent plane projections based on the directions from each Gaussian mean to the camera center instead of naively projecting all Gaussians onto the same plane" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 396, + 237, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 237, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 237, + 410 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 418, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 468 + ], + "type": "text", + "content": "Novel view synthesis is a longstanding challenge in computer vision and graphics. From traditional techniques [3,5,10,11,19,31,33], to neural network-based scene representations [9, 15, 18, 24, 28, 29, 32], various approaches have struggled to address the problem of synthesizing a new view from captured images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 481, + 279, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 279, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 279, + 493 + ], + "type": "text", + "content": "2.1 Neural Radiance Fields" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": "The Neural Radiance Field (NeRF) [25] stands out as a successful neural rendering method based on MLPs, primarily owing to its encoding of position and direction. This encoding allows for effective reconstruction of high-frequency information in scenes. Notable improvements on this encoding have been made by MipNeRF [1], NeRF-W [22], FreeNeRF [41], and Instant NGP [26]. These enhancements enable the handling of multi-resolution image inputs, multi-illumination with occlusion image inputs, sparse-view inputs, and achieve nearly real-time rendering capabilities, respectively. Barron et al. introduced MipNeRF360 [2] as an extension of MipNeRF [1] to address the issue of generating low-quality renderings for unbounded scenes in NeRF. Importantly, these methods thoroughly exploit the intrinsic capabilities of NeRF as an implicit scene representation without introducing additional model priors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "However, due to the implicit representation of scenes and the dense sampling of points along rays, they still face challenges in achieving real-time performance." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 275, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 275, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 275, + 129 + ], + "type": "text", + "content": "2.2 3D Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 194 + ], + "type": "text", + "content": "3D Gaussian Splatting (3D-GS) [16] has demonstrated notable advancements in rendering performance by departing from MLPs and ray sampling, opting instead for Gaussian functions and Gaussian Splatting. This paradigm shift has attracted considerable attention within the industry, leading to various studies building upon this methodology [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 314 + ], + "type": "text", + "content": "3D-GS finds application in SLAM systems [23], scene editing [7] and segmentation [4]. Efforts [20] have also been made to integrate 3D-GS with popular diffusion models [14, 34, 35] and dynamic scenes [21]. Recent endeavors aim to enhance the robustness in sparse-view scenarios [38, 40], performance, storage efficiency [27] and mesh reconstruction [12] of 3D-GS. However, these improvements do not specifically address errors associated with Gaussian projection. Potential methods for enhancing Gaussian functions include [30, 42]. Nevertheless, they introduce a Gaussian filter and a novel Gaussian depth sorting algorithm, respectively, without investigating the projection function and associated errors. These approaches still rely on projecting onto the " + }, + { + "bbox": [ + 130, + 194, + 482, + 314 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 194, + 482, + 314 + ], + "type": "text", + "content": " plane." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 361 + ], + "type": "text", + "content": "This paper aims to explore the potential of 3D Gaussian Splatting by analyzing errors that may arise during the projection process. The analysis will delve into the factors contributing to these errors and propose methods to mitigate and reduce them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 379, + 234, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 234, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 234, + 392 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "text", + "content": "Similar to NeRF [25], the input to 3D-GS consists of a set of images, together with the corresponding cameras calibrated by SFM [31]. However, in contrast to NeRF, 3D-GS takes the sparse point cloud generated during the SFM calibration process as input. From these points, it constructs a set of 3D Gaussians as graphical primitives explicitly representing the scene. Each 3D Gaussian is characterized by its position (mean) " + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "text", + "content": ", covariance matrix " + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "text", + "content": ", and opacity " + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "text", + "content": ", and carries anisotropic spherical harmonics (SH) representing directional appearance component (color) of the radiance field, following standard practice [8,26]. Subsequently, these 3D Gaussians are projected onto the image plane (" + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 402, + 482, + 533 + ], + "type": "text", + "content": " plane) via a projection function for differentiable rasterization. Next, we will provide a detailed explanation of the Gaussian projection process in 3D-GS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": "In the world space, we define a 3D Gaussian function by " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": ", characterized by its mean " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": ". To project the Gaussian onto the image plane, the initial step in 3D-GS involves an affine transformation of this Gaussian function from the world space to the camera space via a viewing transformation matrix " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": ", yielding a transformed Gaussian function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 167, + 601, + 481, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 601, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 167, + 601, + 481, + 620 + ], + "type": "interline_equation", + "content": "G (\\mathbf {x}) = \\exp \\left\\{- 1 / 2 (\\mathbf {W x} - \\mathbf {W} \\boldsymbol {\\mu}) ^ {\\top} (\\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top}) ^ {- 1} (\\mathbf {W x} - \\mathbf {W} \\boldsymbol {\\mu}) \\right\\}. \\qquad (1)", + "image_path": "57224c27298a6de2e09fa8d3c182a60afeb4b6b728e39c217f1b3071c3724e6c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": "We denote " + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": " as the 3D Gaussian in the camera space, with " + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' = \\mathbf{W}\\mathbf{x}" + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}' = \\mathbf{W}\\boldsymbol{\\mu}" + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}' = \\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^\\top" + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": " as the point, mean and covariance matrix in the camera space, respectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 149 + ], + "type": "text", + "content": "Subsequently, traditional 3D-GS requires projecting the Gaussian function in the camera space onto the " + }, + { + "bbox": [ + 130, + 116, + 479, + 149 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 149 + ], + "type": "text", + "content": " plane, with the projection function " + }, + { + "bbox": [ + 130, + 116, + 479, + 149 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 116, + 479, + 149 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 150, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 198, + 150, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x _ {0}}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x _ {0}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\tag {2}", + "image_path": "bbe6b69d6237750a1aa4d42785debe437cae230390ab289acd98271823994a31.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 479, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 178, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{x_0} = [0,0,1]^\\top" + }, + { + "bbox": [ + 130, + 178, + 479, + 217 + ], + "type": "text", + "content": " represents the projection of the camera space's origin onto this plane. Expanding this projection function to the first order using Taylor series yields:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 227, + 480, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 227, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 197, + 227, + 480, + 274 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) = \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) + \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) \\left(\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}\\right) + R _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) \\tag {3} \\\\ \\approx \\varphi (\\boldsymbol {\\mu} ^ {\\prime}) + \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} (\\boldsymbol {\\mu} ^ {\\prime}) (\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}) \\\\ \\end{array}", + "image_path": "4f9c36d05152c8bab076a135bd2c452df1f6fbede457628b16f3ab5d78ba7050.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\varphi}{\\partial\\mathbf{x}'}\\left(\\boldsymbol{\\mu}'\\right)" + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective transformation, denoted as " + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "text", + "content": ". Applying this local affine approximation by neglecting the Taylor remainder term allows to derive the 2D Gaussian function " + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "inline_equation", + "content": "G_{2\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "text", + "content": " projected onto the " + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 284, + 479, + 338 + ], + "type": "text", + "content": " plane as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 347, + 480, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 347, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 134, + 347, + 480, + 412 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} G _ {2 D} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\exp \\left\\{- 1 / 2 \\left(\\mathbf {J} \\mathbf {x} ^ {\\prime} - \\mathbf {J} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {\\top} \\left(\\mathbf {J} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} ^ {\\top}\\right) ^ {- 1} \\left(\\mathbf {J} \\mathbf {x} ^ {\\prime} - \\mathbf {J} \\boldsymbol {\\mu} ^ {\\prime}\\right) \\right\\} \\\\ \\approx \\exp \\left\\{- 1 / 2 \\left(\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) ^ {\\top} \\left(\\mathbf {J} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} ^ {\\top}\\right) ^ {- 1} \\left(\\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) \\right\\}. \\tag {4} \\\\ \\end{array}", + "image_path": "1979bae998f16c0c46ba3080b9e68f530cbf90019e75015e85a0e433f5e437d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": "Since the rank of the matrix " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": " is 2, the inverse of " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\pmb{\\Sigma}' \\mathbf{J}^\\top" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": " is, in fact, the inverse of the covariance matrix of the 2D Gaussian obtained by skipping the third row and column [44]. We denote " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{2\\mathrm{D}} = \\varphi(\\mathbf{x}')" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{2\\mathrm{D}} = \\varphi(\\pmb{\\mu}')" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{2\\mathrm{D}} = \\mathbf{J} \\pmb{\\Sigma}' \\mathbf{J}^\\top" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": " as the point, mean and " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": " covariance matrix of " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "G_{2\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": " in the image space, respectively. For a pixel " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": ", we obtain the value of the 2D Gaussian function by querying " + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{2\\mathrm{D}} = [u, v]^\\top" + }, + { + "bbox": [ + 130, + 425, + 479, + 517 + ], + "type": "text", + "content": ", and then perform alpha blending to derive the color of the pixel. The blue box in Figure 3 illustrates the projection process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 537, + 354, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 537, + 354, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 537, + 354, + 551 + ], + "type": "text", + "content": "4 Local Affine Approximation Error" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 564, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 479, + 635 + ], + "type": "text", + "content": "From Equation 3 and Equation 4 we observe that 3D Gaussian Splatting introduces an approximation during the projection transformation, i.e. the local affine approximation. Consequently, 3D-GS actually adopts a 2D Gaussian function that is not the true projection of the initial 3D Gaussian. The error introduced by this approximation can be characterized by the Taylor remainder term in Equation 3:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 198, + 647, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 647, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 198, + 647, + 480, + 668 + ], + "type": "interline_equation", + "content": "R _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\varphi \\left(\\mathbf {x} ^ {\\prime}\\right) - \\varphi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) - \\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) \\left(\\mathbf {x} ^ {\\prime} - \\boldsymbol {\\mu} ^ {\\prime}\\right) \\tag {5}", + "image_path": "a89f7bf111524b5c5fa04e54741a5b6992383126c5c913c059c0c85a2be44aeb.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 137, + 481, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 137, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 172, + 137, + 481, + 159 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\varphi}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\mathbb {I} \\otimes \\left(\\mathbf {x} _ {0} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} - \\boldsymbol {\\mu} ^ {\\prime} \\left(\\mathbf {x} _ {0} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\mathbf {x} _ {0}\\right) ^ {- 1} \\mathbf {x} _ {0} ^ {\\top}. \\tag {6}", + "image_path": "ce1e42dc07e27f2b56b8d313a51fe5ed8e31e5cdbde6547445396fea4d154863.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "content": " represents the identity matrix and " + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "content": " denotes the multiplication of a matrix and a scalar. Clearly, this Taylor remainder term is a three-dimensional vector related to the random variable " + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\prime}" + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "content": " and the mean " + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{\\prime}" + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "content": " of the Gaussian function. Therefore, by computing the square of the Frobenius norm of this vector " + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\| R_1(\\mathbf{x}')\\| _F^2" + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "content": " and taking the mathematical expectation of this norm function with respect to the random variable " + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\prime}" + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "content": ", we obtain an error function that depends solely on the Gaussian mean " + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{\\prime}" + }, + { + "bbox": [ + 132, + 167, + 482, + 256 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 266, + 481, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 266, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 238, + 266, + 481, + 299 + ], + "type": "interline_equation", + "content": "\\epsilon \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\int_ {\\mathbf {x} ^ {\\prime} \\in \\mathcal {X} ^ {\\prime}} \\| R _ {1} \\left(\\mathbf {x} ^ {\\prime}\\right) \\| _ {F} ^ {2} \\mathrm {d} \\mathbf {x} ^ {\\prime}. \\tag {7}", + "image_path": "c41a86ee6f49ff66a5c44d27be94657cda0ad80a69f09ad5131632a999d3b6ce.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "text", + "content": "The simplified error function " + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "text", + "content": " involves three unit vectors: " + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\prime}" + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{x_0}" + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{\\prime}" + }, + { + "bbox": [ + 132, + 311, + 481, + 335 + ], + "type": "text", + "content": ". Their spherical coordinates are given as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 345, + 481, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 345, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 481, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathbf {0}} = \\left[ \\begin{array}{c} \\sin (\\phi_ {0}) \\cos (\\theta_ {0}) \\\\ - \\sin (\\theta_ {0}) \\\\ \\cos (\\phi_ {0}) \\cos (\\theta_ {0}) \\end{array} \\right], \\mathbf {x} ^ {\\prime} = \\left[ \\begin{array}{c} \\sin (\\phi) \\cos (\\theta) \\\\ - \\sin (\\theta) \\\\ \\cos (\\phi) \\cos (\\theta) \\end{array} \\right], \\boldsymbol {\\mu} ^ {\\prime} = \\left[ \\begin{array}{c} \\sin (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu}) \\\\ - \\sin (\\theta_ {\\mu}) \\\\ \\cos (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu}) \\end{array} \\right] \\tag {8}", + "image_path": "1b59c63bdc03cc6243f2f5ab62ade521f1528a70014dae2abd52b82111519802.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 392, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 392, + 481, + 415 + ], + "type": "inline_equation", + "content": "\\phi_0 = 0" + }, + { + "bbox": [ + 132, + 392, + 481, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 392, + 481, + 415 + ], + "type": "inline_equation", + "content": "\\theta_0 = 0" + }, + { + "bbox": [ + 132, + 392, + 481, + 415 + ], + "type": "text", + "content": ". Please refer to the detailed proofs in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 416, + 481, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 481, + 480 + ], + "type": "text", + "content": "The size of the integration region is related to the covariance of the Gaussian function. Since the current discussion specifically focuses on the impact of the mean on the error rather than the variance, let's assume the integration region to be " + }, + { + "bbox": [ + 132, + 416, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathbf{x}' \\mid \\theta \\in [-\\pi/4 + \\theta_{\\mu}, \\pi/4 + \\theta_{\\mu}] \\land \\phi \\in [-\\pi/4 + \\phi_{\\mu}, \\pi/4 + \\phi_{\\mu}]\\right\\}" + }, + { + "bbox": [ + 132, + 416, + 481, + 480 + ], + "type": "text", + "content": ". Then, based on Equation 5-8, we obtain:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 491, + 490, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 490, + 563 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 490, + 563 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon (\\theta_ {\\mu}, \\phi_ {\\mu}) = \\int_ {- \\pi / 4 + \\theta_ {\\mu}} ^ {\\pi / 4 + \\theta_ {\\mu}} \\int_ {- \\pi / 4 + \\phi_ {\\mu}} ^ {\\pi / 4 + \\phi_ {\\mu}} \\left(- \\frac {\\sin (\\phi - \\phi_ {\\mu}) \\cos (\\theta)}{\\cos^ {2} (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu})} + \\tan (\\phi) - \\tan (\\phi_ {\\mu})\\right) ^ {2} + \\\\ \\left(\\frac {\\sin (\\theta)}{\\cos (\\phi_ {\\mu}) \\cos (\\theta_ {\\mu})} - \\frac {\\sin (\\theta_ {\\mu}) \\cos (\\phi) \\cos (\\theta)}{\\cos^ {2} (\\phi_ {\\mu}) \\cos^ {2} (\\theta_ {\\mu})} + \\frac {\\tan (\\theta_ {\\mu})}{\\cos (\\phi_ {\\mu})} - \\frac {\\tan (\\theta)}{\\cos (\\phi)}\\right) ^ {2} d \\theta d \\phi . \\tag {9} \\\\ \\end{array}", + "image_path": "5f8570a227af726f1a02727ef8e7aebc04cec0093df1fa919fa10e38fcf952cc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 572, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 481, + 608 + ], + "type": "text", + "content": "Fortunately, this integral has a close-form expression, but its form is very complicated. We report the close-form expression in the supplemental materials and visually depict it in Figure 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "text", + "content": "The error function is partially differentiated with respect to " + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "inline_equation", + "content": "\\theta_{\\mu}" + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "inline_equation", + "content": "\\phi_{\\mu}" + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "text", + "content": ". It is easy to check that at " + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "inline_equation", + "content": "\\theta_{\\mu} = \\theta_0 = 0" + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "inline_equation", + "content": "\\phi_{\\mu} = \\phi_0 = 0" + }, + { + "bbox": [ + 132, + 609, + 481, + 633 + ], + "type": "text", + "content": ", the following holds:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 243, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 243, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\epsilon}{\\partial \\theta_ {\\mu}} (0, 0) = 0, \\frac {\\partial \\epsilon}{\\partial \\phi_ {\\mu}} (0, 0) = 0. \\tag {10}", + "image_path": "cbe34e8408abb81b4f4bc97f9497999881d5390a2c886f01941891399866e6b8.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 118, + 294, + 232 + ], + "blocks": [ + { + "bbox": [ + 177, + 118, + 294, + 232 + ], + "lines": [ + { + "bbox": [ + 177, + 118, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 177, + 118, + 294, + 232 + ], + "type": "image", + "image_path": "5b6c085bd2f8652cdb99fcf863d9d73ed435b3c42003556c3f27351eb39cf9b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 237, + 254, + 247 + ], + "lines": [ + { + "bbox": [ + 204, + 237, + 254, + 247 + ], + "spans": [ + { + "bbox": [ + 204, + 237, + 254, + 247 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 204, + 237, + 254, + 247 + ], + "type": "inline_equation", + "content": "\\lambda = 0.095" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 329, + 119, + 446, + 231 + ], + "blocks": [ + { + "bbox": [ + 329, + 119, + 446, + 231 + ], + "lines": [ + { + "bbox": [ + 329, + 119, + 446, + 231 + ], + "spans": [ + { + "bbox": [ + 329, + 119, + 446, + 231 + ], + "type": "image", + "image_path": "361d2ffa91e732c015aff3b5487d1082d3f5f19b2e2fb35714ee61dc1642f9f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 238, + 408, + 247 + ], + "lines": [ + { + "bbox": [ + 362, + 238, + 408, + 247 + ], + "spans": [ + { + "bbox": [ + 362, + 238, + 408, + 247 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 362, + 238, + 408, + 247 + ], + "type": "inline_equation", + "content": "\\lambda = 0.95" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "text", + "content": "Fig. 2: Visualization of the 3D Gaussian Splitting error function " + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\epsilon (\\theta_{\\mu},\\phi_{\\mu})" + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "text", + "content": " under two distinct domains. The domain of the function is " + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\{(\\theta_{\\mu},\\phi_{\\mu})\\mid \\theta_{\\mu}\\in [-\\lambda \\pi /4,\\lambda \\pi /4]\\land \\phi_{\\mu}\\in [-\\lambda \\pi /4,\\lambda \\pi /4]\\}" + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "inline_equation", + "content": "(\\theta_{\\mu},\\phi_{\\mu})" + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "text", + "content": " represents the polar coordinate of the 3D Gaussian mean and " + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 256, + 482, + 310 + ], + "type": "text", + "content": " represents the scaling factor of the domain." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 453 + ], + "type": "text", + "content": "Therefore, this point is actually an extremum point of the function. We also confirm that this point is the minimum of the function. Indeed, from Figure 2a, it is also apparent that the function is concave, with its minimum value occurring at " + }, + { + "bbox": [ + 130, + 335, + 482, + 453 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 130, + 335, + 482, + 453 + ], + "type": "text", + "content": ". Additionally, it can be observed that the minimum value of the error function is greater than zero, due to the fact that Gaussian functions are not closed under projection transformations. Figure 2b shows that in the majority of the region close to the origin, the function values vary very gently. However, as the independent variable of the function approaches the boundaries, the function values increase rapidly, resulting in a substantial difference between the maximum and minimum values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "content": "The above analysis indicates that the error is small and does not significantly affect the quality of rendered images. Therefore, traditional 3D-GS [16] utilizes local affine approximation but still successfully reconstructs the scene to obtain a high-quality novel view image. Nevertheless, the naive projection of all Gaussians onto the same plane " + }, + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "content": " in 3D-GS may lead to larger projection errors for Gaussians farther from the plane center (the projection point of the camera center to the plane), causing artifacts. These artifacts are severe when using wide-angle lenses, significantly degrading the quality of the renderings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 568, + 320, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 320, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 320, + 584 + ], + "type": "text", + "content": "5 Optimal Gaussian Splatting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Based on the analysis of the error function, we introduce Optimal Gaussian Splitting shown in Figure 3. This helps us to achieve smaller projection errors, resulting in higher-quality rendering. The first step, similar to the original 3D-GS [16], involves transforming world coordinates into the camera coordinate system. Instead of projecting Gaussians onto the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " plane, each Gaussian is radially projected based on its mean along the line connecting it to the camera" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 241, + 219 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 241, + 219 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 241, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 241, + 219 + ], + "type": "image", + "image_path": "9916bebfb6385cf294cad8405418c9dfc2eb84de32f010cd889c9549cf88c2c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": "Fig. 3: Illustration of the rendering pipeline for our Optimal Gaussian Splitting and the projection of 3D-GS [16]. The blue box depicts the projection process of the original 3D-GS, which straightforwardly projects all Gaussians onto the same projection plane. In contrast, the red box illustrates our approach, where we project individual Gaussians onto corresponding tangent planes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 116, + 481, + 219 + ], + "blocks": [ + { + "bbox": [ + 242, + 116, + 481, + 219 + ], + "lines": [ + { + "bbox": [ + 242, + 116, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 242, + 116, + 481, + 219 + ], + "type": "image", + "image_path": "2ab75e9dba7ebfc00a5f7e92503697fbeb0b2589e5f28bcd52833218da8eb61f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 365 + ], + "type": "text", + "content": "center, projecting onto a plane tangent to the unit sphere and perpendicular to this line. After projection, colors of points on the unit sphere are computed by alpha blending the 2D Gaussians on the tangent plane. Finally, for each pixel in the image, rays are cast onto the unit sphere to retrieve the color of that pixel, resulting in the rendered image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 380, + 259, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 259, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 259, + 392 + ], + "type": "text", + "content": "5.1 Optimal Projection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 493 + ], + "type": "text", + "content": "We discover that the error function attains its minimum value when the projection of the Gaussian mean on the plane coincides with the projection from the camera center to the plane. Therefore, in Optimal Gaussian Splatting, we employ an optimal projection method. Specifically, instead of naively projecting different Gaussians onto the same plane " + }, + { + "bbox": [ + 130, + 397, + 482, + 493 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 130, + 397, + 482, + 493 + ], + "type": "text", + "content": ", we adopt distinct projection planes for each Gaussian. These projection planes are determined by the tangent planes formed by the Gaussian mean and the line connecting it to the camera center. The specific projection plane is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 259, + 498, + 481, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 498, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 259, + 498, + 481, + 519 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\cdot \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x} _ {\\mathbf {p}}\\right) = 0 \\tag {11}", + "image_path": "b29743d1fbbb0421b1c8524d64cc332829008344d20428f783964b722be2ff29.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 523, + 162, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 162, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 162, + 533 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 532, + 481, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 532, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 233, + 532, + 481, + 555 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathbf {p}} = \\varpi \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\boldsymbol {\\mu} ^ {\\prime} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1 / 2} \\tag {12}", + "image_path": "fe4f353d44fe1492203842984fa1b2ac4d018867cca607e86fca7c6cb2f330af.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 594 + ], + "type": "text", + "content": "represents the projection of the camera space's origin onto this plane and " + }, + { + "bbox": [ + 130, + 558, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\varpi" + }, + { + "bbox": [ + 130, + 558, + 482, + 594 + ], + "type": "text", + "content": " function represents the transformation that projects points onto the unit sphere. According to Equation 11, the optimal projection function " + }, + { + "bbox": [ + 130, + 558, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 594 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 190, + 598, + 481, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 598, + 481, + 623 + ], + "spans": [ + { + "bbox": [ + 190, + 598, + 481, + 623 + ], + "type": "interline_equation", + "content": "\\varphi_ {\\mathrm {p}} \\left(\\mathbf {x} ^ {\\prime}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} _ {\\mathbf {p}}\\right) = \\mathbf {x} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\mathbf {x} ^ {\\prime}\\right) ^ {- 1}. \\tag {13}", + "image_path": "8258c384553935e70092c476c087c4fe087d0eabd8106ed2764579f12a44b710.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 626, + 436, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 626, + 436, + 639 + ], + "spans": [ + { + "bbox": [ + 131, + 626, + 436, + 639 + ], + "type": "text", + "content": "The corresponding local affine approximation Jacobian matrix " + }, + { + "bbox": [ + 131, + 626, + 436, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathbf{p}}" + }, + { + "bbox": [ + 131, + 626, + 436, + 639 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 145, + 644, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 644, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 145, + 644, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {J} _ {\\mathbf {p}} = \\frac {\\partial \\varphi_ {\\mathrm {p}}}{\\partial \\mathbf {x} ^ {\\prime}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right) = \\mathbb {I} \\otimes \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} - \\boldsymbol {\\mu} ^ {\\prime} \\left(\\mathbf {x} _ {\\mathbf {p}} ^ {\\top} \\boldsymbol {\\mu} ^ {\\prime}\\right) ^ {- 1} \\left(\\boldsymbol {\\mu} ^ {\\prime \\top} \\mathbf {x} _ {\\mathbf {p}}\\right) ^ {- 1} \\mathbf {x} _ {\\mathbf {p}} ^ {\\top}. \\tag {14}", + "image_path": "bb900232032818a5b229d93c4e40ae2e1c7c1c1d1eed92c3108d504c4eece1ce.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 439, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 439, + 131 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 439, + 131 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 132, + 114, + 439, + 131 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{\\prime} = [\\mu_{x},\\mu_{y},\\mu_{z}]^{\\top}" + }, + { + "bbox": [ + 132, + 114, + 439, + 131 + ], + "type": "text", + "content": ", the specific form of this Jacobian matrix is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 140, + 481, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 140, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 186, + 140, + 481, + 211 + ], + "type": "interline_equation", + "content": "\\mathbf {J} _ {\\mathbf {p}} = \\left[ \\begin{array}{l l l} \\frac {\\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {x} \\mu_ {y}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {x} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {y}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & \\frac {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {y} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\frac {3}{2}}} & - \\frac {\\mu_ {y} \\mu_ {z}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\cfrac {3}{2}}} & \\frac {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2}}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}\\right) ^ {\\cfrac {3}{2}}} \\end{array} \\right]. \\tag {15}", + "image_path": "01f60ff29493dfa257fc672f0a73ed34c48e64187fe459ae40fa800bc3bf6d24.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 218, + 479, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 218, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 479, + 242 + ], + "type": "text", + "content": "It's very easy to implement modifications to the Jacobian matrix in the forward process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 260, + 310, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 310, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 310, + 272 + ], + "type": "text", + "content": "5.2 Unit Sphere Based Rasterizer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 280, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 479, + 339 + ], + "type": "text", + "content": "For the proposed Optimal Projection, we have proposed a Unit Sphere Based Rasterizer to generate images. Through Optimal Projection, we obtain the projection of the three-dimensional Gaussians on the tangent plane of the unit sphere instead of obtaining the Gaussians in the image space. Therefore, we need to rasterize based on this unit sphere." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 341, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 479, + 399 + ], + "type": "text", + "content": "Specifically, for a pixel " + }, + { + "bbox": [ + 131, + 341, + 479, + 399 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 131, + 341, + 479, + 399 + ], + "type": "text", + "content": " on the image, similar to NeRF, we cast a ray. However, unlike NeRF, we do not involve the extensive sampling of points along the ray, which would significantly degrade performance. Instead, our focus is solely on determining which tangent-plane Gaussians the ray intersects on the unit sphere:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 244, + 400, + 480, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 400, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 244, + 400, + 480, + 437 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {2 D}} = \\varphi_ {\\mathrm {p}} \\left(\\left[ \\begin{array}{c} (u - c _ {x}) / f _ {x} \\\\ (v - c _ {y}) / f _ {y} \\\\ 1 \\end{array} \\right]\\right) \\tag {16}", + "image_path": "d7064279b148b8fe42cbcb4751d4e0216700d9510c8dd63946efe56ae681ba7d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "inline_equation", + "content": "c_x, c_y, f_x, f_y" + }, + { + "bbox": [ + 131, + 441, + 479, + 476 + ], + "type": "text", + "content": " denote the intrinsic parameters of the pinhole camera. We then query the function values of these Gaussians for alpha blending to obtain the color" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 484, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 484, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 134, + 484, + 480, + 521 + ], + "type": "interline_equation", + "content": "G _ {2 \\mathrm {D}} \\left(\\mathbf {x} _ {2 \\mathrm {D}}\\right) = \\exp \\left\\{- 1 / 2 \\left(\\mathbf {x} _ {2 \\mathrm {D}} - \\varphi_ {\\mathrm {p}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) ^ {\\top} \\left(\\mathbf {J} _ {\\mathbf {p}} \\boldsymbol {\\Sigma} ^ {\\prime} \\mathbf {J} _ {\\mathbf {p}} ^ {\\top}\\right) ^ {- 1} \\left(\\mathbf {x} _ {2 \\mathrm {D}} - \\varphi_ {\\mathrm {p}} \\left(\\boldsymbol {\\mu} ^ {\\prime}\\right)\\right) \\right\\}. \\tag {17}", + "image_path": "6a81b8dbb288b1f898f3b79b0f0eadb877913a7fac6c5342cd54d68a6e253250.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": "Note that both " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathbf{p}}" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": " in Equation 14 and " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": " in Equation 6 are matrices with a rank of 2. Consequently, the inverse of the covariance matrix of the 2D Gaussian is also obtained by skipping the third row and column. However, the third row and third column of " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathbf{p}}\\boldsymbol{\\Sigma}'\\mathbf{J}_{\\mathbf{p}}^\\top" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": " are not entirely zeros, as opposed to " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{J}\\boldsymbol{\\Sigma}'\\mathbf{J}^\\top" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": ". To ensure the equation still holds, an invertible matrix " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": " needs to be used for left multiplication with " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{p}}(\\boldsymbol{\\mu}')" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathbf{p}}" + }, + { + "bbox": [ + 131, + 521, + 479, + 601 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 611, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 611, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 159, + 611, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\left[ \\begin{array}{c c c} \\frac {\\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}} & 0 & - \\frac {\\mu_ {x}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}} \\\\ - \\frac {\\mu_ {x} \\mu_ {y}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}} \\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} - \\frac {\\mu_ {y} \\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {z} ^ {2}} \\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} \\\\ \\frac {\\mu_ {x}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\mu_ {y}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} & \\frac {\\mu_ {z}}{\\sqrt {\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + \\mu_ {z} ^ {2}}} \\end{array} \\right]. \\tag {18}", + "image_path": "a5769beccb359dcab7714908e9742471a9f130739d161ed410429a9c10e13b31.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 164 + ], + "type": "text", + "content": "Since our projection does not rely on the perspective image plane " + }, + { + "bbox": [ + 132, + 116, + 481, + 164 + ], + "type": "inline_equation", + "content": "z = 1" + }, + { + "bbox": [ + 132, + 116, + 481, + 164 + ], + "type": "text", + "content": " as in the original 3D-GS [16], we can adapt to different camera models such as fisheye cameras and panoramic images by simply modifying the transformation from image space to camera space, as described in Equation 16." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 230, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 230, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 230, + 196 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 243 + ], + "type": "text", + "content": "To validate the effectiveness of the Optimal Gaussian Splatting derived from theoretical error analysis, we conduct a series of experiments, comparing it with the original 3D-GS [16] and some current state-of-the-art methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 260, + 241, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 241, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 241, + 274 + ], + "type": "text", + "content": "6.1 Implementation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 279, + 481, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 481, + 328 + ], + "type": "text", + "content": "We implemented Optimal Gaussian Splitting based on the PyTorch framework in 3D-GS [16] and wrote custom CUDA kernels for rasterization. We used the default parameters of 3D-GS to maintain consistency with the original 3D-GS and prevent other factors from introducing interference into the results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 345, + 205, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 205, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 205, + 357 + ], + "type": "text", + "content": "6.2 Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 365, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 460 + ], + "type": "text", + "content": "We test our algorithm on a total of 13 real scenes which are the same as those used in the original 3D-GS [16]. In particular, we evaluate our approach on the complete set of scenes featured in Mip-NeRF360 [2], which currently represents the state-of-the-art in NeRF rendering quality. Additionally, we test our method on two scenes from the Tanks & Temples dataset [17] and two scenes provided by Hedman et al. [13]. The selected scenes exhibit diverse capture styles, encompassing both confined indoor environments and expansive, unbounded outdoor settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 478, + 198, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 198, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 198, + 490 + ], + "type": "text", + "content": "6.3 Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 498, + 481, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 558 + ], + "type": "text", + "content": "We compare our algorithm with 3D-GS [16] and three state-of-the-art NeRF-based approaches: Mip-NeRF360 [2], InstantNGP [26] and Plenoxels [8]. All methods are configured using the same settings as outlined in the 3D-GS [16] to control variables, preventing the introduction of variables other than our Optimal Splatting method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "text", + "content": "Quantitative comparisons We adopt a train/test split for datasets following the methodology proposed by Mip-NeRF360 [2]. Specifically, every 8th photo is reserved for testing, ensuring a consistent and meaningful basis for comparisons to generate error metrics. Standard metrics such as PSNR, LPIPS [43], and SSIM are employed for evaluation. We report quantitative results in Table 1. While the 3D-GS significantly outperforms MipNeRF360 in rendering performance, its rendering quality slightly falls behind MipNeRF360, with a lower average PSNR across the 13 scenes. This might be attributed to MipNeRF360 not employing" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 179, + 471, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of our method compared to previous work across three datasets. The Mip-NeRF360 [2] dataset includes both indoor and outdoor scenarios. Tanks. and Deep. indicates the average results over the Tanks&Templates [17] and Deep Blending [13] datasets. Average represents the mean value across all scenarios in these datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 179, + 471, + 264 + ], + "lines": [ + { + "bbox": [ + 142, + 179, + 471, + 264 + ], + "spans": [ + { + "bbox": [ + 142, + 179, + 471, + 264 + ], + "type": "table", + "html": "
Dataset Method|MetricAverageMip-NeRF360Tanks. and Deep.
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Plenoxels [8]22.770.6660.45723.080.6250.46322.070.7570.444
INGP-Base [26]24.490.6990.37325.300.6710.37122.670.7600.377
INGP-Big [26]24.930.7240.33625.590.6990.33123.440.7810.347
M-NeRF360 [2]27.110.8030.24127.690.7920.23725.810.8300.251
3D-GS [16]26.920.8320.21427.210.8150.21426.270.8720.213
Ours27.170.8360.21027.480.8210.20926.440.8720.214
", + "image_path": "dbb36083d07d2bfd893a35d4f8c527c1ae011c21d1d0df91bb10822c1bfcbf08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 274, + 481, + 507 + ], + "blocks": [ + { + "bbox": [ + 133, + 274, + 481, + 507 + ], + "lines": [ + { + "bbox": [ + 133, + 274, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 481, + 507 + ], + "type": "image", + "image_path": "d8afd8b9cbf9090cd121a3e0a2c14d26a85ea766d67fa45af51a6b7a8b85ef67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 515, + 482, + 571 + ], + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 571 + ], + "type": "text", + "content": "Fig. 4: We show comparisons of our method to previous methods and the corresponding ground truth images from held-out test views. The scenes are, from the top down: TRUCK from Tanks&Temples [17]; PLAYROOM from the Deep Blending dataset [13] and BONSAI, COUNTER from Mip-NeRF360 dataset [2]. Differences in quality highlighted by arrows/insets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": "significant approximations throughout its entire pipeline. However, by replacing the original projection method of 3D-GS with our optimal projection, which incurs smaller errors, we not only maintain rendering performance far surpassing MipNeRF360 but also achieve a quality improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Qualitative comparisons For our qualitative experiments, we select only two methods, 3D-GS and MipNeRF360, as our baselines. It can be observed that" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 171 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 171 + ], + "type": "image", + "image_path": "3a9051b9664d848d67240f7def0e83431f2673daa8e62b3d249915406c2fd0f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "type": "text", + "content": "Fig. 5: We derive the projection error function with respect to image coordinates " + }, + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "type": "text", + "content": " and focal length through the transformation between image and polar coordinates " + }, + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\theta_{\\mu}, \\phi_{\\mu})" + }, + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "type": "text", + "content": ". We visualize this function, showing 3D-GS rendered images with long and short focal lengths, followed by the corresponding error functions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 308 + ], + "type": "text", + "content": "our method is capable of generating more realistic details, with fewer artifacts compared to 3D-GS [16]. And this is precisely the superiority brought about by our projection method, which results in smaller errors. Furthermore, We see that in some cases Mip-NeRF360 has remaining artifacts that our method avoids (e.g., black lines - in PLAYROOM)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 329, + 218, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 329, + 218, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 329, + 218, + 340 + ], + "type": "text", + "content": "6.4 Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "text", + "content": "Impacts of Decreasing Focal Length Based on Equation 9, we further derive the projection error function with respect to image coordinates " + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "text", + "content": " and focal length " + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "inline_equation", + "content": "f_{x},f_{y}" + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "text", + "content": " through the coordinate transformation between image coordinates and polar coordinates " + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "inline_equation", + "content": "(\\theta_{\\mu},\\phi_{\\mu})" + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "text", + "content": ". As illustrated in Figure 5, the peak of projection error mainly occurs at the image edges under a long focal length, resulting in artifacts at the image edges. As the focal length decreases, the field of view expands, leading to more Gaussians deviating from the projection center and an overall increase in projection error. In such cases, 3D-GS exhibit more artifacts, such as needle-like structures or cloud-like large Gaussians, which obscure parts that perform well under long focal length, thereby significantly degrading the overall image quality. Since our method employs a central radial projection for the projection plane, theoretically, we should not encounter such defects. To validate this perspective, we conduct experiments at various focal length reduction ratios (reduction ratios of " + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\times 0.2" + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\times 0.3" + }, + { + "bbox": [ + 130, + 353, + 482, + 520 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "For quantitative experiments, considering that there is no ground truth for large field-of-view in the 13 scenes from the 3D-GS [16], we apply a focal length mask. This mask selects central parts of the rendered images with corresponding ground truth based on the focal length scaling factor for metric calculation. The results are presented in the Table 2. Whether it is at " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\times 0.2" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\times 0.3" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " reduction ratios, our method outperforms the original 3D-GS significantly across various scenes in terms of all metrics. As the focal length decreases from " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\times 0.3" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\times 0.2" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": ", 3D-GS [16] experiences a significant degradation due to increased projection errors, while our method demonstrates greater robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "For quantitative experiments, we randomly select three scenes, each with two different focal lengths for demonstration. Additional experimental results can be found in the supplementary materials. It is evident from the Figure 6" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 179, + 475, + 319 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: We quantitatively compare our method with the original 3D-GS [16] using focal lengths of 0.2 and 0.3 times the original. Six scenes were selected from three datasets: Mip-NeRF360 [2], Tank & Templates [17], and Deep Blending [13]. The scenes include TRAIN and TRUCK from Tanks & Templates, DRJOHNSON and PLAYROOM from Deep Blending, and ROOM and TREEHILL from Mip-NeRF360." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 179, + 475, + 319 + ], + "lines": [ + { + "bbox": [ + 140, + 179, + 475, + 319 + ], + "spans": [ + { + "bbox": [ + 140, + 179, + 475, + 319 + ], + "type": "table", + "html": "
MetricFocalMethodTrainTruckDrJohnsonPlayroomRoomTreehillAverage
PSNR↑×0.23D-GS [16]10.1212.3919.2416.3820.0215.3615.58
Ours16.6016.7023.5423.2923.6718.9820.46
×0.33D-GS [16]12.8514.8923.3622.5125.0018.3519.49
Ours17.5218.0724.8926.4025.8719.8722.10
SSIM↑×0.23D-GS [16]0.3250.3800.6160.6190.6800.3710.499
Ours0.5650.5370.7010.7520.7770.4350.628
×0.33D-GS [16]0.4710.5290.7400.7910.8120.4370.630
Ours0.5880.6050.7620.8440.8340.4720.684
LPIPS↓×0.23D-GS [16]0.5690.4790.2960.3550.2070.4740.397
Ours0.2770.2500.2160.2030.1780.3850.251
×0.33D-GS [16]0.3310.2760.2070.2130.1540.3970.263
Ours0.2430.2170.2010.1820.1580.3850.231
", + "image_path": "17e301929eeb6052252eb7808976d57580f3465a6395899e58a383de9154b15e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 334, + 203, + 373 + ], + "blocks": [ + { + "bbox": [ + 133, + 334, + 203, + 373 + ], + "lines": [ + { + "bbox": [ + 133, + 334, + 203, + 373 + ], + "spans": [ + { + "bbox": [ + 133, + 334, + 203, + 373 + ], + "type": "image", + "image_path": "1268bf6aebcbd50d49bf512bc15f8bfaefad5b4482ba126462629acc764d7e79.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 205, + 334, + 273, + 373 + ], + "blocks": [ + { + "bbox": [ + 205, + 334, + 273, + 373 + ], + "lines": [ + { + "bbox": [ + 205, + 334, + 273, + 373 + ], + "spans": [ + { + "bbox": [ + 205, + 334, + 273, + 373 + ], + "type": "image", + "image_path": "806d986099681f6ef4726210ae9e4b9c5b8f115a8eac887c480f830699b21553.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 273, + 334, + 342, + 373 + ], + "blocks": [ + { + "bbox": [ + 273, + 334, + 342, + 373 + ], + "lines": [ + { + "bbox": [ + 273, + 334, + 342, + 373 + ], + "spans": [ + { + "bbox": [ + 273, + 334, + 342, + 373 + ], + "type": "image", + "image_path": "a570e1470766f89c58851a3a7d14140da7e461252c3df58833133d4be2898dc1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 334, + 411, + 373 + ], + "blocks": [ + { + "bbox": [ + 342, + 334, + 411, + 373 + ], + "lines": [ + { + "bbox": [ + 342, + 334, + 411, + 373 + ], + "spans": [ + { + "bbox": [ + 342, + 334, + 411, + 373 + ], + "type": "image", + "image_path": "82604c3b8841aab712b2d98ac7fe985456453319cead0b3b56f272a9d7d3caec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 411, + 334, + 480, + 373 + ], + "blocks": [ + { + "bbox": [ + 411, + 334, + 480, + 373 + ], + "lines": [ + { + "bbox": [ + 411, + 334, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 411, + 334, + 480, + 373 + ], + "type": "image", + "image_path": "6e8aaebb3966e9a3c5863fb4f05f0aa1e7c136011190b7bc6bae970af8aeebc4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 374, + 203, + 420 + ], + "blocks": [ + { + "bbox": [ + 133, + 374, + 203, + 420 + ], + "lines": [ + { + "bbox": [ + 133, + 374, + 203, + 420 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 203, + 420 + ], + "type": "image", + "image_path": "1f224a3d72de997304df760b78a94a2e419c07838d97bbc0f03dc4a6a018f810.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 205, + 374, + 273, + 420 + ], + "blocks": [ + { + "bbox": [ + 205, + 374, + 273, + 420 + ], + "lines": [ + { + "bbox": [ + 205, + 374, + 273, + 420 + ], + "spans": [ + { + "bbox": [ + 205, + 374, + 273, + 420 + ], + "type": "image", + "image_path": "a663eea222528cd1f488ef9d04e4ad9b4b4a97d33efccbc9d731945a22ba54e8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 273, + 374, + 342, + 420 + ], + "blocks": [ + { + "bbox": [ + 273, + 374, + 342, + 420 + ], + "lines": [ + { + "bbox": [ + 273, + 374, + 342, + 420 + ], + "spans": [ + { + "bbox": [ + 273, + 374, + 342, + 420 + ], + "type": "image", + "image_path": "3c71e304e7fbcfb6e8e0394be300935b2c64df261c1cc52571831c561dcb34df.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 342, + 374, + 410, + 420 + ], + "blocks": [ + { + "bbox": [ + 342, + 374, + 410, + 420 + ], + "lines": [ + { + "bbox": [ + 342, + 374, + 410, + 420 + ], + "spans": [ + { + "bbox": [ + 342, + 374, + 410, + 420 + ], + "type": "image", + "image_path": "53297f4f353236a57052531f0f138eb92184c2fae7608bc9d3d459f636185bd5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 374, + 480, + 420 + ], + "blocks": [ + { + "bbox": [ + 410, + 374, + 480, + 420 + ], + "lines": [ + { + "bbox": [ + 410, + 374, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 410, + 374, + 480, + 420 + ], + "type": "image", + "image_path": "12ddc4272ca0c2cb88a52e98376d3d02b2ac9a5919e06c5b2738a9e379d44702.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 133, + 420, + 203, + 466 + ], + "blocks": [ + { + "bbox": [ + 133, + 420, + 203, + 466 + ], + "lines": [ + { + "bbox": [ + 133, + 420, + 203, + 466 + ], + "spans": [ + { + "bbox": [ + 133, + 420, + 203, + 466 + ], + "type": "image", + "image_path": "f044836e2f37e1982f5475183f90501bd25da44c4e7fccbe899100087fb8db94.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 475, + 482, + 521 + ], + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 521 + ], + "type": "text", + "content": "Fig. 6: We compare our method to the original 3D-GS [16] with different focal length reductions. Two focal lengths, 0.2 and 0.3 times the original, are tested. Ground truth " + }, + { + "bbox": [ + 130, + 475, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\times 1" + }, + { + "bbox": [ + 130, + 475, + 482, + 521 + ], + "type": "text", + "content": " in the left images indicates no focal length reduction. The red-boxed areas in the right images correspond to the same regions in the left images." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 205, + 420, + 273, + 466 + ], + "blocks": [ + { + "bbox": [ + 205, + 420, + 273, + 466 + ], + "lines": [ + { + "bbox": [ + 205, + 420, + 273, + 466 + ], + "spans": [ + { + "bbox": [ + 205, + 420, + 273, + 466 + ], + "type": "image", + "image_path": "b1286fd98708d78e2d6a03ba5bffd24b1279568d4a5344f63d6782de1b2b3b43.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 273, + 420, + 342, + 466 + ], + "blocks": [ + { + "bbox": [ + 273, + 420, + 342, + 466 + ], + "lines": [ + { + "bbox": [ + 273, + 420, + 342, + 466 + ], + "spans": [ + { + "bbox": [ + 273, + 420, + 342, + 466 + ], + "type": "image", + "image_path": "19ff1f611e4ac50d9b297c394b68d04779d83f7718904e0acedccf2288b926d7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 342, + 420, + 410, + 466 + ], + "blocks": [ + { + "bbox": [ + 342, + 420, + 410, + 466 + ], + "lines": [ + { + "bbox": [ + 342, + 420, + 410, + 466 + ], + "spans": [ + { + "bbox": [ + 342, + 420, + 410, + 466 + ], + "type": "image", + "image_path": "a5252c499c4ad2f162536b50871bcbfe809e42822428f03066a981ab68b4c684.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 410, + 420, + 480, + 466 + ], + "blocks": [ + { + "bbox": [ + 410, + 420, + 480, + 466 + ], + "lines": [ + { + "bbox": [ + 410, + 420, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 410, + 420, + 480, + 466 + ], + "type": "image", + "image_path": "e322cc1fa21d192bc0afe8cd143d1844a7593c9317e9b522a10e12e2b598ee0d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 548, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 481, + 572 + ], + "type": "text", + "content": "that our method exhibits greater robustness across various focal length settings, mitigating the artifacts introduced by projection in the original 3D-GS." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Adaptability to Various Camera Models By straightforwardly modifying the transformation from image space to camera space, we have achieved adaptation to various camera models, with the caveat that the original 3D-GS's projection struggles to support these camera models. Figure 7 illustrates that our projection, in addition to conforming to the original pinhole camera according to Equation 16, can also accommodate fisheye cameras and generate panoramas." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 290 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 290 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 290 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 290 + ], + "type": "image", + "image_path": "1021eeabfb4dcb03869c0a00a9f5eb44804a6cb59c16fbaac84f6936b59e9391.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 299, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 333 + ], + "type": "text", + "content": "Fig. 7: Renderings of three scenes using our projection under three camera models. From left to right: the ground truth of the original perspective, wide-angle rendering with a pinhole camera, fisheye camera rendering, and the panorama." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 465 + ], + "type": "text", + "content": "**Limitation** Due to our method's independence from the image plane, it exhibits adaptability to various camera models. However, the slight increase in training time is a consequence of the subsequent transformation to the image plane after projection onto the tangent plane, rather than a direct projection onto the image plane. Fortunately, this method avoids the need for dense points sampling, ensuring the algorithm's time complexity remains unchanged and rendering performance stays unaffected. Moreover, it may be possible to mitigate the impact on training time, particularly concerning the corresponding backward propagation, through the implementation of optimized CUDA code." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 220, + 496 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "type": "text", + "content": "Through theoretical analysis, we establish the correlation between the projection error and the Gaussian mean in 3D-GS. By optimizing the error function, we discover that the projection error is minimized when projecting on the plane perpendicular to the line connecting the Gaussian mean and the camera center. Leveraging this insight, we propose a novel projection that yields high-quality images without compromising rendering performance and can be easily adapted to various camera models through simple modifications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "In this work, our primary focus is on analyzing the relationship between the projection error and the mean of the 3D Gaussian under the assumption of a constant covariance. The influence of Gaussian covariance on projection requires further discussion in future work. It may be possible to further explore the potential of Gaussian Splatting as an explicit representation technique for scenes." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5835-5844. IEEE (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 481, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 481, + 231 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 232, + 481, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 287 + ], + "type": "text", + "content": "3. Buehler, C., Bosse, M., McMillan, L., Gortler, S.J., Cohen, M.F.: Unstructured lumigraph rendering. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2001, Los Angeles, California, USA, August 12-17, 2001. pp. 425-432. ACM (2001). https://doi.org/10.1145/383259.383309, https://doi.org/10.1145/383259.383309" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 288, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 308 + ], + "type": "text", + "content": "4. Cen, J., Fang, J., Yang, C., Xie, L., Zhang, X., Shen, W., Tian, Q.: Segment any 3d gaussians. arXiv preprint arXiv:2312.00860 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 310, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 353 + ], + "type": "text", + "content": "5. Chaurasia, G., Duchêne, S., Sorkine-Hornung, O., Drettakis, G.: Depth synthesis and local warps for plausible image-based navigation. ACM Trans. Graph. 32(3), 30:1-30:12 (2013). https://doi.org/10.1145/2487228.2487238, https://doi.org/10.1145/2487228.2487238" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "type": "text", + "content": "6. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 376, + 481, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 421 + ], + "type": "text", + "content": "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21476-21485 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 422, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 481, + 454 + ], + "type": "text", + "content": "8. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 455, + 481, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 481, + 476 + ], + "type": "text", + "content": "9. Genova, K., Cole, F., Sud, A., Sarna, A., Funkhouser, T.: Local deep implicit functions for 3d shape. In: CVPR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 477, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 481, + 531 + ], + "type": "text", + "content": "0. Goesele, M., Snavely, N., Curless, B., Hoppe, H., Seitz, S.M.: Multi-view stereo for community photo collections. In: IEEE 11th International Conference on Computer Vision, ICCV 2007, Rio de Janeiro, Brazil, October 14-20, 2007. pp. 1-8. IEEE Computer Society (2007). https://doi.org/10.1109/ICCV.2007.4408933, https://doi.org/10.1109/ICCV.2007.4408933" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 533, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 586 + ], + "type": "text", + "content": "1. Gortler, S.J., Grzeszczuk, R., Szeliski, R., Cohen, M.F.: The lumigraph. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 43-54. ACM (1996). https://doi.org/10.1145/237170.237200, https://doi.org/10.1145/237170.237200" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "type": "text", + "content": "2. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354-5363 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "3. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 182 + ], + "type": "text", + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual (2020), https://proceedings.neurips.cc/paper/2020/bit/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 205 + ], + "type": "text", + "content": "15. Jiang, C., Sud, A., Makadia, A., Huang, J., Nießner, M., Funkhouser, T.: Local implicit grid representations for 3d scenes. In: CVPR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "18. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Trans. Graph. 41(6), 201:1-201:15 (2022). https://doi.org/10.1145/3550454.3555497, https://doi.org/10.1145/3550454.3555497" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 357 + ], + "type": "text", + "content": "19. Levoy, M., Hanrahan, P.: Light field rendering. In: Fujii, J. (ed.) Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996. pp. 31-42. ACM (1996). https://doi.org/10.1145/237170.237199, https://doi.org/10.1145/237170.237199" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "20. Li, X., Wang, H., Tseng, K.K.: Gaussiandiffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "21. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 2024 International Conference on 3D Vision (3DV). pp. 800-809. IEEE (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 479 + ], + "type": "text", + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7210-7219. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00713" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "23. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "25. Mildenhall, B., Srinivasan, P., Tancik, M., Barron, J., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "27. Niedermayr, S., Stumpfegger, J., Westermann, R.: Compressed 3d gaussian splatting for accelerated novel view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10349-10358 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "text", + "content": "30. Radl, L., Steiner, M., Parger, M., Weinrauch, A., Kerbl, B., Steinberger, M.: StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics 4(43) (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "text", + "content": "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "32. Sitzmann, V., Zollhoefer, M., Wetzstein, G.: Scene representation networks: Continuous 3D-structure-aware neural scene representations. In: NeurIPS (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "33. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. ACM Trans. Graph. 25(3), 835-846 (2006). https://doi.org/10.1145/1141911.1141964, https://doi.org/10.1145/1141911.1141964" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 335 + ], + "type": "text", + "content": "34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 11895-11907 (2019), https://proceedings.neurips.cc/paper/2019/bit/3001ef257407d5a371a96cdcd947c7d93-Abstraction.html" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 390 + ], + "type": "text", + "content": "35. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=PxTIG12RRHS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 445 + ], + "type": "text", + "content": "36. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5449-5459. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00538, https://doi.org/10.1109/CVPR52688.2022.00538" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 446, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 510 + ], + "type": "text", + "content": "37. Takikawa, T., Litalien, J., Yin, K., Kreis, K., Loop, C.T., Nowrouzezahrai, D., Jacobson, A., McGuire, M., Fidler, S.: Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11358-11367. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01120" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "38. Xiong, H., Muttukuru, S., Upadhyay, R., Chari, P., Kadambi, A.: Sparsegs: Real-time " + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": " sparse view synthesis using gaussian splatting. arXiv e-prints pp. arXiv-2312 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 544, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 599 + ], + "type": "text", + "content": "39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 5428-5438. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00536, https://doi.org/10.1109/CVPR52688.2022.00536" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "40. Yang, C., Li, S., Fang, J., Liang, R., Xie, L., Zhang, X., Shen, W., Tian, Q.: Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv preprint arXiv:2402.10259 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "41. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8254-8263 (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Error Analysis of 3D-GS and an Optimal Projection Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 194 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "42. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "43. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 194 + ], + "type": "text", + "content": "44. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE Transactions on Visualization and Computer Graphics 8(3), 223-238 (2002)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "L. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_content_list.json b/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..002869f6a71b60b2a481110f9cc5e836be763a4f --- /dev/null +++ b/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_content_list.json @@ -0,0 +1,1844 @@ +[ + { + "type": "text", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "text_level": 1, + "bbox": [ + 308, + 141, + 694, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiarui Duan $^{1,3}$ , Haoling Li $^{2,3}$ , Haofei Zhang $^{1,3}$ , Hao Jiang $^{4}$ , Mengqi Xue $^{5}$ , Li Sun $^{6}$ , Mingli Song $^{1,3}$ , and Jie Song $^{2,3\\dagger}$", + "bbox": [ + 236, + 210, + 767, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 College of Computer Science and Technology, Zhejiang University", + "bbox": [ + 272, + 253, + 727, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Software Technology, Zhejiang University", + "bbox": [ + 323, + 268, + 678, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Key Laboratory of Visual Perception (Zhejiang University), Ministry of Education and Microsoft", + "bbox": [ + 217, + 282, + 781, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Alibaba Group", + "bbox": [ + 444, + 309, + 558, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Hangzhou City University", + "bbox": [ + 406, + 323, + 596, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Ningbo Innovation Center, Zhejiang University", + "bbox": [ + 336, + 337, + 663, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jerryduan,hollylee,haofeizhang,lsun,brooksong,sjie}@zju.edu.cn;", + "bbox": [ + 256, + 352, + 746, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aoshu.jh@alibaba-inc.com; mqxue@hzcu.edu.cn", + "bbox": [ + 336, + 366, + 663, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Attribution-based explanations are garnering increasing attention recently and have emerged as the predominant approach towards eXplanable Artificial Intelligence (XAI). However, the absence of consistent configurations and systematic investigations in prior literature impedes comprehensive evaluations of existing methodologies. In this work, we introduce Meta-Rank, an open platform for benchmarking attribution methods in the image domain. Presently, Meta-Rank assesses eight exemplary attribution methods using six renowned model architectures on four diverse datasets, employing both the Most Relevant First (MoRF) and Least Relevant First (LeRF) evaluation protocols. Through extensive experimentation, our benchmark reveals three insights in attribution evaluation endeavors: 1) evaluating attribution methods under disparate settings can yield divergent performance rankings; 2) although inconsistent across numerous cases, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts at consistent evaluation fare no better than baselines when extended to more heterogeneous models and datasets. Our findings underscore the necessity for future research in this domain to conduct rigorous evaluations encompassing a broader range of models and datasets, and to reassess the assumptions underlying the empirical success of different attribution methods. Our code is publicly available at https://github.com/TreeThree-R/Meta-Rank.", + "bbox": [ + 259, + 407, + 740, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Attribution evaluation $\\cdot$ Disparate settings $\\cdot$ Consistency", + "bbox": [ + 261, + 724, + 727, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 760, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Explainable Artificial Intelligence (XAI) [1] has emerged as a prominent research field within computer vision, with a multitude of approaches proposed for ex-", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 218, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plaining and interpreting the internal mechanisms of diverse neural network architectures across various application domains. Among these approaches, attribution methods have gained significant traction in the realm of deep learning. By generating an explanation map in the form of a heatmap (also referred to as a relevance map or a saliency map), these methods attribute the prediction to specific regions within the input image, thereby facilitating interpretation. Over the years, researchers have devised a large number of attribution methods from diverse perspectives, including gradient-based methods [5, 43, 44, 47, 48], perturbation-based methods [34, 37, 39, 53], and CAM-based methods [12, 42].", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, given the absence of ground truth regarding the inner workings of deep neural networks, evaluating attributions directly poses a formidable challenge. Presently, various studies have endeavored to assess the reliability of attribution methods, which can be roughly categorized into two schools: expert-grounded methods [6, 15, 16, 32, 35, 54], and functional-grounded methods [4, 17, 26, 37, 40, 41, 51]. Expert-grounded evaluations capture how well the explanations imitate the human-annotated importance of the image regions. Such methods not only require human effort, but also suffer from the faithfulness1 issue, as deep neural networks could rely on non-intuitive, spurious cues that differ from human perception [20]. In contrast, functional-based approaches evaluate attribution methods based on some fundamental axioms. For instance, feature ablation [37], as a functional-grounded evaluation, systematically removes features in a predetermined order and observes the resulting changes in predictions. This approach enables more faithful assessments of the attribution explanations by comparing the model behavior with and without specific individual features [19, 21, 26, 37, 39, 58]. Unfortunately, the current literature often evaluates attributions under divergent settings, with different datasets, models and evaluation protocols [26, 40]. The efficacy of attribution methods in practice is often called into question due to restricted and inconsistent experimental conditions.", + "bbox": [ + 212, + 284, + 787, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of this work is to gain a thorough understanding of the current state of attribution methods while setting the stage for critical problems to be worked on. To this end, we present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines. Our benchmark carefully examines eight widely-used attribution methods (Saliency [45], Input $\\odot$ Gradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5] and Guided Grad-CAM [42]), with a wide range of models (ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]), on various domain datasets (NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]), using two evaluation protocols (Most Relevant First and Least Relevant First). Specifically, we place a strong emphasis on comprehensive and consistent experimental settings that have been largely overlooked in previous works. Through extensive experiments on the exhaustive combinations of methods, datasets, models and evaluation protocols,", + "bbox": [ + 212, + 571, + 787, + 797 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Faithfulness refers to the degree of alignment between the results interpreted by an attribution method and the decision of the model.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our benchmark reveals three main findings regarding previous attribution evaluation endeavors: 1) evaluating attribution methods under disparate experimental settings, such as different model architectures, datasets and evaluation protocols, can yield highly divergent performance rankings; 2) although highly inconsistent across datasets, models and evaluation protocols, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts towards more consistent evaluation, like Remove and Debias (ROAD) [40], fare no better than simple baselines when it is extended to more heterogeneous models and datasets.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. Meta-Rank aggregates ranking results from a diverse array of settings, thereby mitigating biases associated with specific configurations. By employing this metric, we conduct systematic comparisons among existing attribution methods, and provide extensive experiment results and discussions on the status quo. Beyond extensive experiments, our benchmark is designed as a flexible framework that standardizes experimental settings and simplifies the integration of novel algorithmic implementations. We hope our work will not only facilitate reliable evaluations of attribution methods across a broader range of models and datasets, but also inspire the research community to delve further into the development of more consistent attribution methods.", + "bbox": [ + 212, + 282, + 787, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we make the following contributions:", + "bbox": [ + 238, + 464, + 599, + 478 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines.", + "- Through extensive experiments, we identify three pivotal insights from previous attribution evaluation efforts, which provide directions for future advancements in attribution research.", + "- We propose an evaluation metric to facilitate a more comprehensive benchmark comparison across multiple settings, with which extensive experiments and discussions are provided on the status quo." + ], + "bbox": [ + 227, + 486, + 782, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 628, + 387, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The development of attribution methods has prompted researchers to seek the best approach to assist models in downstream tasks. In this section, we delineate the progress of attribution evaluation chronologically.", + "bbox": [ + 212, + 657, + 787, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expert-grounded evaluations [6, 16, 35, 54, 55] rely on human judgment to evaluate attribution methods. For example, Zhang et al. [54] proposed the pointing-game method as a means of evaluating attention maps based on detecting whether the highest value point in the map hit the target category. However, such methods are inherently subjective and fail to guarantee the reliability of evaluation results.", + "bbox": [ + 212, + 703, + 795, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Functional-grounded evaluations [2,7,26,37,41] provide more objective evaluations of the model by comparing how it behaves when certain individual features are present or absent. Samek et al. [41] proposed two evaluation protocols,", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "namely Most Relevant First (MoRF) and Least Relevant First (LeRF), to observe the impact on the model's output. In MoRF, effective attribution methods should produce a greater reduction in model accuracy when removing features with higher attribution values, and vice versa. Similarly, [37] suggested evaluating attribution methods through the deletion and insertion of features by using their attribution values. These methods violate the assumption of a consistent distribution between training and test data, making it hard to determine if the decline in model performance solely results from feature deletion. Hooker et al. [26] further demonstrated that the performance degradation from feature deletion partly stems from such a distribution shift. To address this, they developed RemOve And Retrain (ROAR), which resolves inconsistent distributions by synchronously deleting the features of the training set and test set based on the attributions of each attribution method, and then retrained the model on the generated datasets. The effectiveness of the attribution method is measured by the retrained model's accuracy decline. Better methods identify more important features, whose removal causes sharper performance drops. ROAR has been widely adopted in evaluation studies [7, 22, 28]. However, due to its exorbitant expense, this type of evaluation method is unfeasible for scaling up. Rong et al. [40] discovered inconsistent MoRF and LeRF rankings from ROAR, which they attributed to information leakage based on information theory analysis. Thus, they proposed a noisy linear imputation method, named RemOve And Debias (ROAD), to effectively generate consistent results. However, as a pixel filling method, ROAD still suffers from several inherent defects, e.g., the features are replaced with substantial values instead of entirely dropped, which can hardly guarantee their neutral effects on the predictions. Deng et al. [17] noted the lack of unified theoretical frameworks for evaluating attribution methods. To resolve this limitation, they proposed a Taylor interaction system to unify fourteen attribution methods under a common mathematical formulation and suggested three principles for evaluation. However, applying this framework to new methods requires manual theoretical analysis, hampering both scalability and efficiency. Recently, attribution evaluation has been explored from more diverse perspectives [30, 38, 57], including comprehensive viewpoints [24, 25] as well as the consistency lens [8, 11, 26, 40]. Unfortunately, the absence of standardized experimental settings and systematic evaluations in previous literatures poses challenges for comparing existing methods universally. To overcome this limitation, we propose Meta-Rank, a novel evaluation benchmark, to provide a comprehensive evaluation of attribution methods across multiple settings.", + "bbox": [ + 212, + 142, + 787, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 729, + 382, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attribution methods explain deep models by generating saliency maps that highlight pixels critical for predictions. In our work, we benchmark eight prevalent attribution methods: Saliency [45], InputGradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5], and Guided Grad-CAM [42]. More details regarding these methods are recorded", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Appendix B. In the following, we formalize the key concepts in attribution evaluation, deriving a general assessment paradigm in XAI.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To begin with, let $\\mathcal{D} = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^n$ be an image classification dataset with $n$ samples and $C$ categories, where $y_{i}$ is the label of RGB image $\\mathbf{x}_i\\in \\mathbb{R}^{3\\times H\\times W}$ . Considering a deep network $\\hat{y} = f_{\\Theta}(\\mathbf{x})$ that predicts $\\hat{y}$ for input $\\mathbf{x}$ , an attribution method aims to generate a heatmap $H\\in \\mathbb{R}^{3\\times H\\times W}$ so that each input pixel is assigned to a saliency score to the network's prediction. Such a procedure can be formulated as the following form:", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH = \\mathsf {A t t r} _ {\\Gamma} (f _ {\\Theta} | \\mathbf {x}, \\hat {y}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 273, + 785, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathsf{Attr}_{\\varGamma}$ denotes a certain attribution method parameterized by $\\varGamma$", + "bbox": [ + 214, + 297, + 732, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the faithfulness of different attribution methods, we adopt two most common strategies: with progressively removing the most or least salient pixels of each image in the query dataset, we measure the performance degradation of the target model. The two strategies are usually referred to as the Most Relevant First (MoRF) and the Least Relevant First (LeRF) evaluation protocols. Specifically, we construct a top- $t$ removed query dataset $\\tilde{\\mathcal{D}}^{(t)}$ by removing the top- $t$ most or least salient pixels of the original image $\\mathbf{x}_i$ and derive the modified image $\\tilde{\\mathbf{x}}_i^{(t)}$ as follows:", + "bbox": [ + 212, + 311, + 785, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {i} ^ {(t)} = \\operatorname {R e m o v e} \\left(\\mathbf {x} _ {i}, H _ {i} | t\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 441, + 785, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the operation Remove specifies how features are eliminated, representing feature ablation in this paper.", + "bbox": [ + 212, + 467, + 782, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effective attribution methods can accurately pinpoint significant pixels, and the elimination of these pixels would result in a swift deterioration in the model's performance. Thus, the evaluation of attribution methods can be facilitated by quantifying the reduction in model performance, which is defined by:", + "bbox": [ + 212, + 498, + 784, + 558 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\varDelta = \\frac {1}{n} \\left(\\sum_ {i = 1} ^ {n} p \\left(\\hat {y} _ {i} = y _ {i} | \\mathbf {x} _ {i}\\right) - \\sum_ {i = 1} ^ {n} p \\left(\\tilde {y} _ {i} ^ {(t)} = y _ {i} | \\tilde {\\mathbf {x}} _ {i} ^ {(t)}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 564, + 785, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we focus on faithfulness to acquire a broadly applicable standard among multiple experimental settings. Ultimately, attribution methods are ranked according to this metric to generate a ranking list. The schematic diagram of the attribution and evaluation is depicted in Figure 1 (b).", + "bbox": [ + 212, + 595, + 785, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Meta-Rank Settings and Benchmark", + "text_level": 1, + "bbox": [ + 214, + 676, + 606, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To acquire a comprehensive comprehension of the experimental configurations employed in previous investigations, we present a summary of the datasets and models utilized in these works, as depicted in Table 1. Notably, despite the limited choice of datasets and models, there remains considerable variation in the evaluations conducted thus far. The issue of evaluation consistency has largely been overlooked in prior research endeavors. To bridge this gap, in this section, we first establish standardized settings and subsequently provide a comprehensive outline of the Meta-Rank evaluation benchmark. The workflow of MetaRank is illustrated in Figure 1.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0fc47b8c8201488df65994f3c8b750e081316ad81f48d2f6a6d21463ebf989ac.jpg", + "table_caption": [ + "Table 1: Comparison of experimental configurations employed in previous studies. The lack of consistency in these settings may lead to disparate observations. The term \"Cons.\" refers to whether the study addresses the issue of consistency. Only a limited number of works have explored the issue, but their scope is limited to specific datasets, models, or both." + ], + "table_footnote": [], + "table_body": "
MethodsVenueDatasetsModelsCons.
[45]ICLR 2014ImageNetConvNetX
[44]ICML 2016Tiny ImageNet, DNA sequenceVGG-16, 1 CNNX
[48]ICML 2017ImageNetGoogleNetX
[47]ICLR Workshop 2015ImageNet, CIFAR-10/1003 CNNsX
[43]ICML 2017MNIST, DNA sequence2 CNNsX
[53]ECCV 2014ImageNet, Caltech, PASCAL VOC2 CNNsX
[5]PLOS ONE 2015ImageNet, PASCAL VOC, MINIST1 DNN, 1 CNNX
[42]ICCV 2017ImageNet, PASCAL VOCVGG-16, AlexNetX
[26]NeurIPS 2019ImageNet, Birdsnap, Food-101ResNet-50X
[11]ECCV Workshop 2020ImageNetResNet-50✓
[3]FUZZ-IEEE 2022ImageNet, Dogs vs. Cats, MAME, MIT67AlexNet, VGG-16, ResNet-18X
[40]ICML 2022CIFAR-10, Food-101ResNet-18, ResNet-50✓
[10]ICLR Tiny Paper 2023ImageNetResNet-50X
[8]CVPR 2023ImageNetVGG-16, ResNet-34✓
[24]JMLR 2023ImageNetResNet-18X
", + "bbox": [ + 220, + 226, + 779, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Standardized Settings", + "text_level": 1, + "bbox": [ + 215, + 392, + 442, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The main idea of the proposed Meta-Rank is evaluating attribution explanations on a diverse array of datasets, models, and evaluation protocols.", + "bbox": [ + 212, + 417, + 782, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To streamline standardized evaluations of attribution methods, we meticulously select four commonly used datasets for benchmarking: NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]. NWPU-RESISC45 is a medium-scale dataset for remote sensing image classification, containing 45 categories of remote sensing images covering various scenes such as streets, airports, factories, forests, grasslands, and deserts. Food-101 is a medium-scale dataset for fine-grained food image classification, compassing a total of 101 different food categories. ImageNet is a large-scale dataset employed for image classification, consisting of 1,000 different categories. Places-365 is a large-scale dataset designed for the purposes of scene recognition, covering a wide range of 365 diverse scene categories. These datasets allow us to examine each attribution method under various domains, providing us with a more thorough view of the current literature. Further details are provided in Appendix C.", + "bbox": [ + 212, + 446, + 785, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. The selection of models presents a prominent source of inconsistency in previous research. In order to comprehensively validate the effectiveness of the attribution techniques across various models, we adopt six widely employed convolutional neural networks (CNNs) to evaluate existing attribution methods. These CNNs include ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]. It should be noted that most attribution methods are specifically tailored for CNNs and cannot be directly applied to vision transformers (ViTs) [13,33,52]. As such, for maintaining an equitable evaluation across attribution methods with a consistent set of evaluation settings, our benchmark does not take ViTs into consideration. Additional details regarding these used models can be found in Appendix C.", + "bbox": [ + 212, + 643, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocols. Recent advances have discovered that different evaluation protocols, MoRF and LeRF [41], may produce conflicting rankings of", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3416dfb6a9bd15b2155427cc2f6a3397ad242512cb4542cdb4677058c420d238.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 220, + 154, + 782, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99b684f16776c970dd34ed4c862f68e0ffbfb9a1bed356ef8ad69e3ffdd7a191.jpg", + "image_caption": [ + "(b)", + "Fig. 1: Meta-Rank benchmark. It is mainly divided into two stages: Test Case Generation and Meta-Rank (a). Test Case Generation: multiple factors (i.e., datasets, models and evaluation protocols) are combined to generate $\\tau$ different cases. MetaRank: (1) Case Execution. All $m$ competitors (i.e., attribution methods) are applied to these cases, resulting in a collection of rankings. The details of attribution evaluation on an individual case are provided in (b). (2) Ranking Fusion. All rankings $\\{R_1, R_2, \\ldots, R_\\tau\\}$ are subsequently fed into this module. The comparison of two competitors is transformed into the differences in their rankings across all cases, then integrated and converted into the discrepancy in Meta-Ranks. (3) Leaderboard. Ultimately, a unified leaderboard is obtained based on the Meta-Ranks of the competitors. “ $T$ ” is the test case, “ $C$ ” is the competitor, and “ $\\kappa$ ” is the Meta-Rank value." + ], + "image_footnote": [], + "bbox": [ + 243, + 297, + 782, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "attribute methods [40]. To provide a comprehensive view of the attribution methods, the proposed Meta-Rank considers both protocols for evaluation.", + "bbox": [ + 212, + 611, + 784, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Test case settings. To gain a generalized evaluation for attributions, our analytical experiments evaluate existing attribution methods on a wide array of experiment settings, necessitating the initial preparation of baselines. We first train ResNet-18, Inception-v4 and VGG-19 on NWPU-RESISC45 and Food101. Subsequently, we exclusively train EfficientNet_b2 and DenseNet-121 on the NWPU-RESISC45 dataset. These baselines are implemented with the PyTorch [36] framework. Adam [31] is used to optimize the models, with the learning rate of $1 \\times 10^{-4}$ and the weight decay rate of $5 \\times 10^{-4}$ . The cosine annealing is adopted as the learning rate decay schedule. Furthermore, we also incorporate pre-trained ResNet-18, ResNet-50 sourced from Places-365, and pre-trained ResNet-18, Inception-v4, VGG-19 from ImageNet-1k. The aforementioned baselines, combined with two evaluation protocols, are served as the test cases for the benchmark. Detailed information for each test case can be found in Appendix C.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 The Proposed Benchmark Metric", + "text_level": 1, + "bbox": [ + 215, + 146, + 539, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. The main idea of the Meta-Rank metric is viewing each specific configuration of the model, the dataset and the evaluation protocol as one test case. Meta-Rank aggregates the ranking results (i.e., rank the rankings) from a diverse array of test cases, thereby mitigating biases associated with specific configurations.", + "bbox": [ + 212, + 166, + 787, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Formally, let the triplet $\\langle D_i,M_j,P_k\\rangle$ be one test case for evaluating the attribution algorithms, where $D$ , $M$ and $P$ denote the dataset, model and evaluation protocol respectively. The subscript $i$ , $j$ and $k$ denote one specific configuration from the candidate datasets, models and protocols as described in Section 4.1. Let $\\mathcal{C} = \\{C_1,C_2,\\dots,C_m\\}$ be the set of the competitors (i.e., attribution methods). In the proposed benchmark, all the competitors are evaluated on every generated test case, thus producing a ranking among the competitors. For the test case $\\langle D_i,M_j,P_k\\rangle$ , we use $R_{(i,j,k)} = (r_1^{(i,j,k)},r_2^{(i,j,k)},\\dots,r_m^{(i,j,k)})$ to denote the rankings of all the competitors, where $r_m^{(i,j,k)}$ denotes the rank of the $m$ -th attribution method. After evaluating the attribution methods on all the test cases, the probability that $C_p$ outperforms $C_q$ is formally defined as follows:", + "bbox": [ + 212, + 257, + 787, + 431 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} _ {q \\prec p} = \\frac {1}{| \\Omega |} \\sum_ {(i, j, k)} ^ {\\Omega} \\mathbf {1} \\left(r _ {q} ^ {(i, j, k)} \\prec r _ {p} ^ {(i, j, k)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 434, + 785, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ denotes the indicator function, and $r_q \\prec r_p$ represents the ranking $r_q$ is outperformed by $r_p$ . $\\Omega$ is the joint space of datasets, models and evaluation protocols. The probability $\\mathsf{P}_{q \\prec p}$ is then converted into the logistic odds, obtaining a more sensitive metric $\\mathsf{Logit}_{q \\prec p}$ that can better capture slight differences between $C_p$ and $C_q$ :", + "bbox": [ + 212, + 470, + 787, + 547 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L o g i t} _ {q \\prec p} = \\log \\operatorname {o d d s} \\left(C _ {p}, C _ {q}\\right) = \\log \\left(\\mathrm {P} _ {q \\prec p} / \\left(1 - \\mathrm {P} _ {q \\prec p}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 551, + 785, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the above derivation process, we can obtain a set of standardized results $\\{\\mathsf{Logit}_{q\\prec p}\\}$ that quantify the performance differences between any two competitors on all test cases. Then, we transform the performance discrepancy into the difference on the Meta-Rank metric:", + "bbox": [ + 212, + 574, + 787, + 633 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\kappa_ {1} - \\kappa_ {2} = \\operatorname {L o g i t} _ {2 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 2 \\\\ \\kappa_ {1} - \\kappa_ {3} = \\operatorname {L o g i t} _ {3 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 3 \\\\ \\kappa_ {1} - \\kappa_ {4} = \\operatorname {L o g i t} _ {4 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 4 \\\\ \\dots & \\dots \\\\ \\kappa_ {2} - \\kappa_ {3} = \\operatorname {L o g i t} _ {3 \\prec 2}, & \\text {C o m p e t i t o r} \\# 2 v s. \\# 3 \\\\ \\kappa_ {2} - \\kappa_ {4} = \\operatorname {L o g i t} _ {4 \\prec 2}, & \\text {C o m p e t i t o r} \\# 2 v s. \\# 4 \\\\ \\dots & \\dots \\\\ \\kappa_ {m - 1} - \\kappa_ {m} = \\operatorname {L o g i t} _ {m \\prec m - 1}. & \\text {C o m p e t i t o r} \\# \\mathrm {m} - 1 v s. \\# \\mathrm {m} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 636, + 785, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\kappa_{m}$ represents the Meta-Rank value of $C_m$ . After solving the system of equations in Equation 6, we can derive a score $\\kappa$ for each competitor, serving as an identifier for its comprehensive performance. Through the arrangement of competitors in descending order of their $\\kappa$ values, a unified leaderboard can finally be constructed.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/590109a90a800eadccecc5eab54d2de43cdb215560376459c2a5d0785e63ea10.jpg", + "image_caption": [ + "(a) Food-ResNet18(M)" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 357, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/68874e528fe429141a2c9c7d61a04d0e54095d1c8a4bfbdf4d86aaaa0d3f0f9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 143, + 500, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a85a1efca9bcc4a1b86fc44273fd5ea8031d65eaece98a1b31e4e8727eda294f.jpg", + "image_caption": [ + "(c) IN-ResNet18(M)" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 640, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/214d3f34eb9f25c66f95e7cb360c87041b05e175b3f752a62413d9a9839a81d5.jpg", + "image_caption": [ + "(d) IN-ResNet18(L)" + ], + "image_footnote": [], + "bbox": [ + 645, + 143, + 789, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4a9275ddce357e16c3457891d944f9c20cb26f0a843591270f51633f2dd0b19d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 242, + 362, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3a29a1d4cf263775abb7900491272e2c4a772ee39711b09367989e63e93d5825.jpg", + "image_caption": [ + "(b) Food-ResNet18(L)" + ], + "image_footnote": [], + "bbox": [ + 364, + 242, + 504, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/028712ff39f0b7f997da024ec668f6d80a9120b28b308d61a32a8b715c86096a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 242, + 645, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ac8eb546ddf9941173d9921dbcfc4f714a5438e200a3476e040b5f2b357a870a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 242, + 789, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/19a1bd99e420012140627505a9084fd65189e0c941ae00c8a5707bdba2da8e95.jpg", + "image_caption": [ + "(e) Food-Inceptionv4(M)", + "(i) Food-VGG19(M)", + "Fig.2: Evaluation results on 12 test cases. These cases consist of two datasets: Food-101 (a)(b)(e)(f)(i)(j), ImageNet-1k (c)(d)(g)(h)(k)(l), three models: ResNet-18 (a)(b)(c)(d), Inception-v4 (e)(f)(g)(h), VGG-19 (i)(j)(k)(l), and two protocols: MoRF(a)(c)(e)(g)(i)(k), LeRF(b)(d)(f)(h)(j)(l). \"Baseline\" represents the accuracy of the model when no features are ablated." + ], + "image_footnote": [ + "Saliency GB -XG DeepLift -Deconv LRP -GuidedGradCam IG" + ], + "bbox": [ + 217, + 340, + 359, + 424 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f5a9a17f674cfaa427b019217c77464c00fac5988a1c3f652a534868db0efecb.jpg", + "image_caption": [ + "(f) Food-Inceptionv4(L)", + "(j) Food-VGG19(L)" + ], + "image_footnote": [], + "bbox": [ + 361, + 340, + 500, + 424 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e512d465e38b2b422dd7de213c84ae366157974d93239303ab2a48f7a3d7c5f3.jpg", + "image_caption": [ + "(g) IN-Inceptionv4(M)", + "(k) IN-VGG19(M)" + ], + "image_footnote": [], + "bbox": [ + 500, + 340, + 640, + 424 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/738ae351af61a1fb10f336a95de5f07bc5ad8345eeb7e78ee4d2f45f1eb6a829.jpg", + "image_caption": [ + "(h) IN-Inceptionv4(L)", + "(1) IN-VGG19(L)" + ], + "image_footnote": [], + "bbox": [ + 647, + 340, + 789, + 424 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Main Results", + "text_level": 1, + "bbox": [ + 215, + 566, + 382, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This section provides a detailed elaboration of extensive experiments on Meta-Rank, with a particular emphasis on the following pivotal questions.", + "bbox": [ + 212, + 597, + 785, + 628 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- RQ1: Are attribution methods consistently effective across diverse settings (i.e., datasets, models, evaluation protocols) and distinct checkpoints? (Section 5.1)", + "- RQ2: Does the state-of-the-art evaluation method, ROAD, yield consistent rankings across different test cases? (Section 5.2)", + "- RQ3: According to our proposed Meta-Rank, which attribution methods exhibit superior performance across a wide range of cases? (Section 5.3)", + "- RQ4: Can Meta-Rank conduct the evaluation of substantial methods within a feasible timeframe? (Section 5.4)" + ], + "bbox": [ + 225, + 636, + 805, + 756 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 RQ1: Consistency Investigation", + "text_level": 1, + "bbox": [ + 214, + 785, + 521, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we investigate the consistency of performance rankings for existing attribution methods across diverse evaluation settings. Specifically, we focus", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ac78e231fdee73c10604645bc441a62aa2af36108ef3d171d2427efdb13aa783.jpg", + "image_caption": [ + "(a) MoRF" + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 398, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/102d25f1a7c33b50dd58b5071d0f98620881792a79adb775e065fa8b37880800.jpg", + "image_caption": [ + "(b) LeRF", + "Fig. 3: Spearman correlation among nine test cases (NWPU, Food, and ImageNet datasets paired with ResNet-18, Inception-v4, and VGG-19 models) in MoRF (a) and LeRF (b), and between MoRF and LeRF on the same nine cases (c). The labels along the horizontal arrow are equivalent to those along the vertical arrow. Here, we define a correlation score higher than 0.8 as strongly correlated (■), between 0.6 and 0.8 as moderately correlated (▲), and lower than 0.6 as weakly correlated (▼)." + ], + "image_footnote": [], + "bbox": [ + 416, + 145, + 591, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/73436c5a4764c326cac077aa5ac60d5787c36f3afe5c6883b0d54b97f6b1df25.jpg", + "image_caption": [ + "(c) MoRF v.s. LeRF" + ], + "image_footnote": [], + "bbox": [ + 607, + 156, + 776, + 257 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "on the model-, dataset-, protocol- and checkpoint-level ranking consistencies by varying these factors in the evaluation. To this end, we assess the aforementioned eight attribution methods on all the generated test cases, i.e., with six models, four datasets and two ablation protocols (MoRF and LeRF). This yields a total of 26 groups of results, each corresponding to one test case. Some experimental results are presented in Figure 2. For more details of the experimental results, please refer to Appendix D.", + "bbox": [ + 212, + 401, + 787, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Model-level Consistency. It is evident that the attribution methods exhibit varying performance across different models. To elaborate: (1) LRP has the worst performance on the \"ImageNet-Inceptionv4(M)\" case but significantly better results on the \"ImageNet-ResNet18(M)\" and \"ImageNet-VGG19(M)\" cases. (2) Integrated Gradients and Guided Grad-CAM exhibit comparable performance on the \"Food-ResNet18(M)\" case, but show great disparity on the \"Food-VGG19(M)\" case. For a more clear measurement of the inconsistency, we calculate the Spearman Rank correlations among several cases and illustrate them in Figure 3. Here, we extract area under the curve (AUC) [4] of all attribution methods (detailed in Appendix D) to generate a ranking list for each case. In the MoRF protocol, removing the most critical pixels identified by the attribution method causes a substantial drop in model performance. Thus, a lower AUC value indicates a superior method. And the LeRF protocol follows the inverse principle. The results suggest a pronounced weakness in the consistency at the model level, with a strong correlation found in only 2 cases (i.e., the correlation between ResNet-18 and Inception-v4 on the Food-101 dataset in both MoRF and LeRF protocols) out of 18 cases.", + "bbox": [ + 212, + 507, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset-level Consistency. Analogously, attribution methods have shown noticeable inconsistency in performance on different datasets. For instance, Guided Backpropagation demonstrates superior performance on \"ImageNet-VGG19(M)\" yet ranks third-worst on \"Food-VGG19(M)\". In addition, according to the results of Spearman analysis, only $\\frac{5}{18}$ cases show strong correlations at the dataset level.", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/80d36368f64414bdf589ee815780c21538c749e6ad56fd44ecf6d1f23d401917.jpg", + "table_caption": [ + "Table 2: AUC rankings of the eight methods at different checkpoints along the same trajectory. Here, ResNet-18 trained on NWPU-RESISC45 is taken as an illustrative example." + ], + "table_footnote": [], + "table_body": "
Epoch-50Epoch-100Epoch-150Epoch-200
MoRF
LeRF
SaliencyIXGDeconvGuidedGradCam
GBDeepLiftLRPIG
", + "bbox": [ + 220, + 227, + 488, + 354 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e2f5b8da111f2816471a99754fc94f8271d5d85d2ab588518051c728bd2c68f9.jpg", + "table_caption": [ + "Table 3: Comparison of Spearman results among four cases between ROAD and feature ablation. $a \\mid b$ represents a pair of Spearman results, where $a$ denotes using feature ablation (in Section 5.1) and $b$ denotes using noisy linear imputation. The settings of color blocks in this table are the same as those in Figure 3." + ], + "table_footnote": [], + "table_body": "
NWPU-ResNet18Food-Inceptionv4
MoRF (#1)LeRF (#2)MoRF (#3)LeRF (#4)
#1| || || || |||
#2| || || || |||
#3| || || || |||
#4| || || || |||
", + "bbox": [ + 514, + 267, + 781, + 357 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Protocol-level Consistency. The evaluation results indicate that significant inconsistencies between the MoRF and LeRF protocols still exist. A detailed comparison between \"Food-Inceptionv4(M)\" and \"Food-Inceptionv4(L)\" reveals marked discrepancies in Saliency method under MoRF versus LeRF. Besides, the Spearman correlation between MoRF and LeRF is computed for nine test cases (in Figure 3 (c)). It can be easily seen that the correlation of all cases at the protocol level is either moderately correlated or weakly correlated, with no strong correlation observed.", + "bbox": [ + 212, + 383, + 787, + 503 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Checkpoint-level Consistency. We also consider the consistency of attribution methods at different training stages of the model. To illustrate, using ResNet-18 trained on NWPU-RESISC45 as a case study, we extract distinct checkpoints from the same training trajectory and employ two evaluation protocols for attribution analysis. The results of the AUC rankings are recorded in Table 2. It can be easily observed that the rankings of attribution methods demonstrate high consistency across different checkpoints. In MoRF, the outcomes of Epoch-200 differ by only one distinction from those at both Epoch-100 and Epoch-150, while exhibiting slightly greater disparities when compared to those at Epoch-50. Similarly for LeRF, the rankings at Epoch-200 also exhibit a single deviation from those at Epoch-100 and Epoch-150, respectively, and show two distinctions in comparison to those at Epoch-50. These findings suggest a notable consistency in the ranking of attribution methods across different checkpoints. Moreover, the rankings tend to stabilize rapidly as the model converges. More detailed results under other settings can be found in Appendix D.", + "bbox": [ + 212, + 505, + 787, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 RQ2: Necessity of Meta-Rank", + "text_level": 1, + "bbox": [ + 215, + 753, + 509, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To further substantiate the indispensability of the Meta-Rank benchmark, we employ the state-of-the-art attribution evaluation method known as ROAD for assessing the attribution methods across various cases. Starting from information theoretic analyses, ROAD identifies the class information leakage problem", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/09b333279a5ca05efe63f46641be21266a0be9e893c3868406db4103accdaab2.jpg", + "table_caption": [ + "Table 4: Leaderboard of the eight attribution methods. The larger Meta-Rank indicates the stronger ability of attribution method. The top three methods exhibit more positive performance, while the remaining five methods perform negatively." + ], + "table_footnote": [], + "table_body": "
MethodsMeta-Rank Score#Rank
InputGradient2.32001
Integrated Gradients2.07362
DeepLift1.07803
Saliency-0.04834
LRP-0.22345
Guided Backpropagation-0.25856
Guided Grad-CAM-0.96797
Deconvolution-3.97348
", + "bbox": [ + 217, + 252, + 496, + 337 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1e282c6aecdfde2b2ad839227b11212ca6466ad488dbbedb9e8978ce71fb8b28.jpg", + "image_caption": [ + "Fig. 4: Time consumption of the \"Ranking Fusion\" module in Meta-Rank. 2, 4, 8, 16, 32, and 64 are the number of evaluated attribution methods. The gray bar represents the relative error." + ], + "image_footnote": [], + "bbox": [ + 529, + 143, + 767, + 266 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "dictated by the ablation mask and finally derives the noisy linear imputation for minimally revealing imputation, which is empirically demonstrated to produce consistent rankings of attribution methods between evaluations with MoRF and LeRF. However, it overlooks consistency at both the model and dataset levels. Thus, we randomly select four cases, namely NWPU-ResNet18(M), NWPU-ResNet18(L), Food-Inceptionv4(M), and Food-Inceptionv4(L), to assess the eight aforementioned attribution methods using ROAD. The evaluation metric used in this study aligns with those in RQ1. Table 3 presents the Spearman correlations among the evaluation results of the four cases. Further information regarding the accuracy decay curves is available in Appendix E.", + "bbox": [ + 212, + 378, + 784, + 530 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "It can be easily seen that, even when ROAD is employed as a substitute for feature ablation, the correlation between the rankings of attribution methods on different cases remains inconsistent. Specifically, all correlations demonstrate either moderate or weak, with no instances of strong correlation. This indicates that relying solely on ROAD for evaluating attribution methods across distinct cases is inadequate. Consequently, there is a critical need for a cross-case benchmark to ensure robust ranking results.", + "bbox": [ + 212, + 531, + 785, + 636 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 RQ3: Attribution Evaluation with Meta-Rank", + "text_level": 1, + "bbox": [ + 215, + 661, + 638, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Following the evaluation scheme outlined in the Meta-Rank benchmark (described in Section 4), we have carried out a unified measurement of the eight attribution methods. The consistency validation experiments in Section 5.1 generate evaluation results for attribution methods across distinct configurations, encompassing four datasets, six models, and two protocols, resulting in 26 sets of data. We then transform these results into AUC scores, which serve as the performance metric for the attribution methods.", + "bbox": [ + 212, + 688, + 784, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Meta-Rank generates a comprehensive value for each attribution method. These uniquely determined scores enable us to establish a unified ranking list for all the methods. The finalized leaderboard for the eight attribution methods", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "is presented in Table 4. The results reveal InputGradient as the top-performing approach, while Deconvolution exhibits the poorest performance. (1) In particular, InputGradient achieves the highest score among the eight methods. By element-wise multiplying the gradients with the inputs, InputGradient produces more stable attributions in the face of perturbations, which contributes to its superior performance across multiple test cases. (2) Furthermore, the second-ranked method, Integrated Gradients, also shows great potential, with its score closely approaching that of the top-ranked method. (3) In contrast, Deconvolution obtains the lowest score, likely attributable to the hierarchical structure of the network accumulating noise during backpropagation. It is worth noting that a Meta-Rank score can assume both positive and negative values. Positive values signify comparatively superior attribution performance, whereas negative values indicate inferior performance.", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The experimental findings demonstrate that Meta-Rank furnishes a comprehensive benchmark for attribution evaluation across diverse test cases, while effectively ameliorating the issue of inconsistent performance across heterogeneous cases. It computes the discrepancies between methods and converts them into the differences in capacity to derive the Meta-Rank. This result is unique, consistent, and aligned with practical application needs.", + "bbox": [ + 212, + 342, + 787, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 RQ4: Efficiency of Meta-Rank", + "text_level": 1, + "bbox": [ + 214, + 452, + 511, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To measure the time consumption, we conduct experiments on the CPU device (Intel(R) Xeon(R) Platinum 8260L CPU @ 2.40GHz) using different numbers of methods. Here, we focus on the additional \"Ranking Fusion\" module, as the \"Case Execution\" module has been generically applied in all evaluations. In particular, we compare the time expenses of Meta-Rank during the calculation of the 2, 4, 8, 16, 32, and 64 methods. It is important to highlight that the first three test groups (i.e., 2, 4, and 8 methods) rely on real data gathered from previous experiments, whereas the last three test groups (i.e., 16, 32, and 64 methods) employ simulated data randomly generated. The data here refers to a set of accuracies (i.e., $\\{acc_i\\}$ in Figure 1) obtained on each test case according to every attribution method.", + "bbox": [ + 212, + 474, + 787, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results in Figure 4 indicate that as the number of attribution methods increases, the computation time also increases. However, since the number of attribution methods is finite, the computational time remains within an acceptable range. For example, when 64 methods are involved in the Meta-Rank evaluation, the time required to generate the leaderboard is approximately 0.12s, which completely meets expectations. Additionally, using GPUs for Meta-Rank calculations can further enhance the computation speed. This efficiency is beneficial for the promotion and application of Meta-Rank.", + "bbox": [ + 212, + 640, + 787, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 214, + 781, + 354, + 796 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Choice of faithfulness metric. Although there are many dimensions for attribution evaluation in previous work [24], faithfulness remains the most solid", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and general criteria for benchmarking attribution methods. In the context of multiple experimental settings, as such, we opt to use it as the base metric $\\Delta$ in our framework to facilitate subsequent Meta-Rank assessment and then enable broader evaluation on novel tasks. Meanwhile, Meta-Rank can also accommodate evaluations on other dimensions by enabling the derivation of a ranking list along each respective evaluation metric.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Missingness bias in feature ablation. Previous work has discussed the missingness bias problem caused by the removal of pixels through masking values in CNNs [29]. Hooker et al. proposed ROAR to address this issue, but it inadvertently introduced the problem of information leakage. Subsequently, Rong et al. developed ROAD to mitigate this leakage, but it failed to fully remove the pixels (explained in Appendix E). We recognize that completely eliminating this bias poses substantial difficulties, since the root of missingness bias lies in the operation of removing pixels within CNNs. Furthermore, the impact of bias exhibits variability across individual cases, also posing additional challenges in quantification and measurement. However, we believe that with a more diverse array of test cases, the missingness bias is more neutralized. Future research will further explore solutions for eliminating this bias.", + "bbox": [ + 212, + 237, + 787, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Scalability of Meta-Rank. When introducing new attribution methods to the ranking, we initially conduct tests on several existing cases. The time complexity of evaluating a method on an individual case is linear, which has been demonstrated in previous studies [37, 40]. Subsequently, the obtained results, in conjunction with historical data, are fed into the \"Ranking Fusion\" module to generate an updated ranking list. The time overhead associated with this process is also optimistic, as illustrated in Section 5.4. This flexibility allows for the inclusion of any attribution method into the leaderboard at any time. Future endeavors will focus on expanding the corpus of test cases and pursuing further computational efficiency by leveraging parallel computing paradigms.", + "bbox": [ + 212, + 419, + 787, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 595, + 359, + 611 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To address the lack of systematic research in feature attribution, we propose Meta-Rank, a novel standardized benchmark for robust and unified evaluation of methods. Meta-Rank aggregates relative comparisons between all method pairs across diverse settings, mitigating inconsistencies. Through extensive experiments, we reveal three critical insights: diverse evaluation settings yield divergent performance rankings, assessment results remain consistent across distinct checkpoints along the same training trajectory, and previous attempts at consistent evaluation exhibit limitations on heterogeneous models and datasets. Meta-Rank provides a leaderboard that highlights the broader applicability of InputGradient and Integrated Gradients. Remarkably, Meta-Rank exhibits exceptional computational efficiency, enabling ranking computations to be completed within milliseconds even for a substantial number of methods. In the future, we hope the proposed benchmark will facilitate the development of faithful and consistent explanations, fostering transparency and accountability.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (62106220), Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies, Ningbo 2025 Science and Technology Innovation Major Project (No. 2022Z072), and the Fundamental Research Funds for the Central Universities (226-2024-00058).", + "bbox": [ + 212, + 174, + 810, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adadi, A., Berrada, M.: Peeking inside the black-box: a survey on explainable artificial intelligence (xai). IEEE access 6, 52138-52160 (2018)", + "2. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104 (2017)", + "3. Arias-Duart, A., Parés, F., Garcia-Gasulla, D., Giménez-Ábalos, V.: Focus! rating xai methods and finding biases. In: 2022 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). pp. 1-8. IEEE (2022)", + "4. Atanasova, P., Simonsen, J.G., Lioma, C., Augenstein, I.: A diagnostic study of explainability techniques for text classification. arXiv preprint arXiv:2009.13295 (2020)", + "5. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015)", + "6. Bastings, J., Aziz, W., Titov, I.: Interpretable neural predictions with differentiable binary variables. arXiv preprint arXiv:1905.08160 (2019)", + "7. Bhatt, U., Weller, A., Moura, J.M.: Evaluating and aggregating feature-based model explanations. arXiv preprint arXiv:2005.00631 (2020)", + "8. Binder, A., Weber, L., Lapuschkin, S., Montavon, G., Müller, K.R., Samek, W.: Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16143-16152 (2023)", + "9. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014)", + "10. Brocki, L., Chung, N.C.: Evaluation of interpretability methods and perturbation artifacts in deep neural networks. arXiv preprint arXiv:2203.02928 (2022)", + "1. Brunke, L., Agrawal, P., George, N.: Evaluating input perturbation methods for interpreting cnns and saliency map comparison. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 120-134. Springer (2020)", + "2. Chattopadhy, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Gradcam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018)", + "3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 782-791 (2021)" + ], + "bbox": [ + 225, + 301, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)", + "15. Colin, J., Fel, T., Cadène, R., Serre, T.: What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems 35, 2832-2845 (2022)", + "16. Covert, I., Lundberg, S.M., Lee, S.I.: Understanding global feature contributions with additive importance measures. Advances in Neural Information Processing Systems 33, 17212-17223 (2020)", + "17. Deng, H., Zou, N., Du, M., Chen, W., Feng, G., Yang, Z., Li, Z., Zhang, Q.: Understanding and unifying fourteen attribution methods with taylor interactions. arXiv preprint arXiv:2303.01506 (2023)", + "18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "19. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017)", + "20. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)", + "21. Goyal, Y., Wu, Z., Ernst, J., Batra, D., Parikh, D., Lee, S.: Counterfactual visual explanations. In: International Conference on Machine Learning. pp. 2376-2384. PMLR (2019)", + "22. Hartley, T., Sidorov, K., Willis, C., Marshall, D.: Explaining failure: Investigation of surprise and expectation in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 12-13 (2020)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "24. Hedström, A., Weber, L., Krakowczyk, D., Bareeva, D., Motzkus, F., Samek, W., Lapuschkin, S., Höhne, M.M.C.: Quantus: An explainable ai toolkit for responsible evaluation of neural network explanations and beyond. Journal of Machine Learning Research 24(34), 1-11 (2023)", + "25. Hesse, R., Schaub-Meyer, S., Roth, S.: Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3981-3991 (2023)", + "26. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems 32 (2019)", + "27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)", + "28. Ismail, A.A., Gunady, M., Corrada Bravo, H., Feizi, S.: Benchmarking deep learning interpretability in time series predictions. Advances in neural information processing systems 33, 6441-6452 (2020)", + "29. Jain, S., Salman, H., Wong, E., Zhang, P., Vineet, V., Vermprala, S., Madry, A.: Missingness bias in model debugging. arXiv preprint arXiv:2204.08945 (2022)", + "30. Khakzar, A., Khorsandi, P., Nobahari, R., Navab, N.: Do explanations explain? model knows best. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10244-10253 (2022)" + ], + "bbox": [ + 217, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "32. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602 (2022)", + "33. Leem, S., Seo, H.: Attention guided cam: Visual explanations of vision transformer guided by self-attention. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2956-2964 (2024)", + "34. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)", + "35. Nguyen, G., Kim, D., Nguyen, A.: The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems 34, 26422-26436 (2021)", + "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32, 8026-8037 (2019)", + "37. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)", + "38. Rao, S., Böhle, M., Schiele, B.: Better understanding differences in attribution methods via systematic evaluations. arXiv preprint arXiv:2303.11884 (2023)", + "39. Ribeiro, M.T., Singh, S., Guestrin, C.: \"why should i trust you?\" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016)", + "40. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A consistent and efficient evaluation strategy for attribution methods. In: International Conference on Machine Learning. pp. 18770-18795. PMLR (2022)", + "41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.R.: Evaluating the visualization of what a deep neural network has learned. IEEE transactions on neural networks and learning systems 28(11), 2660-2673 (2016)", + "42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "43. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017)", + "44. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not just a black box: Learning important features through propagating activation differences. arXiv preprint arXiv:1605.01713 (2016)", + "45. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: visualising image classification models and saliency maps. In: Proceedings of the International Conference on Learning Representations (ICLR). ICLR (2014)", + "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "47. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014)", + "48. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On the Evaluation Consistency of Attribution-based Explanations", + "bbox": [ + 290, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)", + "50. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)", + "51. Wang, Y., Wang, X.: A unified study of machine learning explanation evaluation metrics. arXiv preprint arXiv:2203.14265 (2022)", + "52. Xue, M., Huang, Q., Zhang, H., Cheng, L., Song, J., Wu, M., Song, M.: Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In: 33rd International Joint Conference on Artificial Intelligence, IJCAI-24. International Joint Conferences on Artificial Intelligence Organization (2024)", + "53. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 818-833. Springer (2014)", + "54. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-down neural attention by excitation backprop. International Journal of Computer Vision 126(10), 1084-1102 (2018)", + "55. Zhang, Y., Khakzar, A., Li, Y., Farshad, A., Kim, S.T., Navab, N.: Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems 34, 20040-20051 (2021)", + "56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)", + "57. Zhou, Y., Booth, S., Ribeiro, M.T., Shah, J.: Do feature attribution methods correctly attribute features? In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9623-9633 (2022)", + "58. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. arXiv preprint arXiv:1702.04595 (2017)" + ], + "bbox": [ + 215, + 147, + 784, + 563 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Duan et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_model.json b/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..15355d6ebcdce5a06efecaaf5e2bfaaa06d49740 --- /dev/null +++ b/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_model.json @@ -0,0 +1,2689 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.31, + 0.142, + 0.696, + 0.187 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.212, + 0.768, + 0.245 + ], + "angle": 0, + "content": "Jiarui Duan\\(^{1,3}\\), Haoling Li\\(^{2,3}\\), Haofei Zhang\\(^{1,3}\\), Hao Jiang\\(^{4}\\), Mengqi Xue\\(^{5}\\), Li Sun\\(^{6}\\), Mingli Song\\(^{1,3}\\), and Jie Song\\(^{2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.255, + 0.728, + 0.269 + ], + "angle": 0, + "content": "1 College of Computer Science and Technology, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.269, + 0.679, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Software Technology, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.283, + 0.782, + 0.309 + ], + "angle": 0, + "content": "3 Key Laboratory of Visual Perception (Zhejiang University), Ministry of Education and Microsoft" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.31, + 0.56, + 0.324 + ], + "angle": 0, + "content": "4 Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.324, + 0.597, + 0.338 + ], + "angle": 0, + "content": "5 Hangzhou City University" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.338, + 0.665, + 0.353 + ], + "angle": 0, + "content": "\\(^{6}\\) Ningbo Innovation Center, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.353, + 0.748, + 0.367 + ], + "angle": 0, + "content": "{jerryduan,hollylee,haofeizhang,lsun,brooksong,sjie}@zju.edu.cn;" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.367, + 0.665, + 0.38 + ], + "angle": 0, + "content": "aoshu.jh@alibaba-inc.com; mqxue@hzcu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.408, + 0.741, + 0.713 + ], + "angle": 0, + "content": "Abstract. Attribution-based explanations are garnering increasing attention recently and have emerged as the predominant approach towards eXplanable Artificial Intelligence (XAI). However, the absence of consistent configurations and systematic investigations in prior literature impedes comprehensive evaluations of existing methodologies. In this work, we introduce Meta-Rank, an open platform for benchmarking attribution methods in the image domain. Presently, Meta-Rank assesses eight exemplary attribution methods using six renowned model architectures on four diverse datasets, employing both the Most Relevant First (MoRF) and Least Relevant First (LeRF) evaluation protocols. Through extensive experimentation, our benchmark reveals three insights in attribution evaluation endeavors: 1) evaluating attribution methods under disparate settings can yield divergent performance rankings; 2) although inconsistent across numerous cases, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts at consistent evaluation fare no better than baselines when extended to more heterogeneous models and datasets. Our findings underscore the necessity for future research in this domain to conduct rigorous evaluations encompassing a broader range of models and datasets, and to reassess the assumptions underlying the empirical success of different attribution methods. Our code is publicly available at https://github.com/TreeThree-R/Meta-Rank." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.725, + 0.728, + 0.739 + ], + "angle": 0, + "content": "Keywords: Attribution evaluation \\(\\cdot\\) Disparate settings \\(\\cdot\\) Consistency" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.761, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Explainable Artificial Intelligence (XAI) [1] has emerged as a prominent research field within computer vision, with a multitude of approaches proposed for ex-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "plaining and interpreting the internal mechanisms of diverse neural network architectures across various application domains. Among these approaches, attribution methods have gained significant traction in the realm of deep learning. By generating an explanation map in the form of a heatmap (also referred to as a relevance map or a saliency map), these methods attribute the prediction to specific regions within the input image, thereby facilitating interpretation. Over the years, researchers have devised a large number of attribution methods from diverse perspectives, including gradient-based methods [5, 43, 44, 47, 48], perturbation-based methods [34, 37, 39, 53], and CAM-based methods [12, 42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Nevertheless, given the absence of ground truth regarding the inner workings of deep neural networks, evaluating attributions directly poses a formidable challenge. Presently, various studies have endeavored to assess the reliability of attribution methods, which can be roughly categorized into two schools: expert-grounded methods [6, 15, 16, 32, 35, 54], and functional-grounded methods [4, 17, 26, 37, 40, 41, 51]. Expert-grounded evaluations capture how well the explanations imitate the human-annotated importance of the image regions. Such methods not only require human effort, but also suffer from the faithfulness1 issue, as deep neural networks could rely on non-intuitive, spurious cues that differ from human perception [20]. In contrast, functional-based approaches evaluate attribution methods based on some fundamental axioms. For instance, feature ablation [37], as a functional-grounded evaluation, systematically removes features in a predetermined order and observes the resulting changes in predictions. This approach enables more faithful assessments of the attribution explanations by comparing the model behavior with and without specific individual features [19, 21, 26, 37, 39, 58]. Unfortunately, the current literature often evaluates attributions under divergent settings, with different datasets, models and evaluation protocols [26, 40]. The efficacy of attribution methods in practice is often called into question due to restricted and inconsistent experimental conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.798 + ], + "angle": 0, + "content": "The goal of this work is to gain a thorough understanding of the current state of attribution methods while setting the stage for critical problems to be worked on. To this end, we present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines. Our benchmark carefully examines eight widely-used attribution methods (Saliency [45], Input\\(\\odot\\)Gradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5] and Guided Grad-CAM [42]), with a wide range of models (ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]), on various domain datasets (NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]), using two evaluation protocols (Most Relevant First and Least Relevant First). Specifically, we place a strong emphasis on comprehensive and consistent experimental settings that have been largely overlooked in previous works. Through extensive experiments on the exhaustive combinations of methods, datasets, models and evaluation protocols," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "1 Faithfulness refers to the degree of alignment between the results interpreted by an attribution method and the decision of the model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "our benchmark reveals three main findings regarding previous attribution evaluation endeavors: 1) evaluating attribution methods under disparate experimental settings, such as different model architectures, datasets and evaluation protocols, can yield highly divergent performance rankings; 2) although highly inconsistent across datasets, models and evaluation protocols, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts towards more consistent evaluation, like Remove and Debias (ROAD) [40], fare no better than simple baselines when it is extended to more heterogeneous models and datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.463 + ], + "angle": 0, + "content": "We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. Meta-Rank aggregates ranking results from a diverse array of settings, thereby mitigating biases associated with specific configurations. By employing this metric, we conduct systematic comparisons among existing attribution methods, and provide extensive experiment results and discussions on the status quo. Beyond extensive experiments, our benchmark is designed as a flexible framework that standardizes experimental settings and simplifies the integration of novel algorithmic implementations. We hope our work will not only facilitate reliable evaluations of attribution methods across a broader range of models and datasets, but also inspire the research community to delve further into the development of more consistent attribution methods." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.465, + 0.601, + 0.479 + ], + "angle": 0, + "content": "In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.487, + 0.784, + 0.514 + ], + "angle": 0, + "content": "- We present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.517, + 0.784, + 0.559 + ], + "angle": 0, + "content": "- Through extensive experiments, we identify three pivotal insights from previous attribution evaluation efforts, which provide directions for future advancements in attribution research." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.562, + 0.784, + 0.605 + ], + "angle": 0, + "content": "- We propose an evaluation metric to facilitate a more comprehensive benchmark comparison across multiple settings, with which extensive experiments and discussions are provided on the status quo." + }, + { + "type": "list", + "bbox": [ + 0.228, + 0.487, + 0.784, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.388, + 0.645 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.703 + ], + "angle": 0, + "content": "The development of attribution methods has prompted researchers to seek the best approach to assist models in downstream tasks. In this section, we delineate the progress of attribution evaluation chronologically." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.797, + 0.793 + ], + "angle": 0, + "content": "Expert-grounded evaluations [6, 16, 35, 54, 55] rely on human judgment to evaluate attribution methods. For example, Zhang et al. [54] proposed the pointing-game method as a means of evaluating attention maps based on detecting whether the highest value point in the map hit the target category. However, such methods are inherently subjective and fail to guarantee the reliability of evaluation results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Functional-grounded evaluations [2,7,26,37,41] provide more objective evaluations of the model by comparing how it behaves when certain individual features are present or absent. Samek et al. [41] proposed two evaluation protocols," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.143, + 0.788, + 0.709 + ], + "angle": 0, + "content": "namely Most Relevant First (MoRF) and Least Relevant First (LeRF), to observe the impact on the model's output. In MoRF, effective attribution methods should produce a greater reduction in model accuracy when removing features with higher attribution values, and vice versa. Similarly, [37] suggested evaluating attribution methods through the deletion and insertion of features by using their attribution values. These methods violate the assumption of a consistent distribution between training and test data, making it hard to determine if the decline in model performance solely results from feature deletion. Hooker et al. [26] further demonstrated that the performance degradation from feature deletion partly stems from such a distribution shift. To address this, they developed RemOve And Retrain (ROAR), which resolves inconsistent distributions by synchronously deleting the features of the training set and test set based on the attributions of each attribution method, and then retrained the model on the generated datasets. The effectiveness of the attribution method is measured by the retrained model's accuracy decline. Better methods identify more important features, whose removal causes sharper performance drops. ROAR has been widely adopted in evaluation studies [7, 22, 28]. However, due to its exorbitant expense, this type of evaluation method is unfeasible for scaling up. Rong et al. [40] discovered inconsistent MoRF and LeRF rankings from ROAR, which they attributed to information leakage based on information theory analysis. Thus, they proposed a noisy linear imputation method, named RemOve And Debias (ROAD), to effectively generate consistent results. However, as a pixel filling method, ROAD still suffers from several inherent defects, e.g., the features are replaced with substantial values instead of entirely dropped, which can hardly guarantee their neutral effects on the predictions. Deng et al. [17] noted the lack of unified theoretical frameworks for evaluating attribution methods. To resolve this limitation, they proposed a Taylor interaction system to unify fourteen attribution methods under a common mathematical formulation and suggested three principles for evaluation. However, applying this framework to new methods requires manual theoretical analysis, hampering both scalability and efficiency. Recently, attribution evaluation has been explored from more diverse perspectives [30, 38, 57], including comprehensive viewpoints [24, 25] as well as the consistency lens [8, 11, 26, 40]. Unfortunately, the absence of standardized experimental settings and systematic evaluations in previous literatures poses challenges for comparing existing methods universally. To overcome this limitation, we propose Meta-Rank, a novel evaluation benchmark, to provide a comprehensive evaluation of attribution methods across multiple settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.383, + 0.748 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Attribution methods explain deep models by generating saliency maps that highlight pixels critical for predictions. In our work, we benchmark eight prevalent attribution methods: Saliency [45], InputGradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5], and Guided Grad-CAM [42]. More details regarding these methods are recorded" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "in Appendix B. In the following, we formalize the key concepts in attribution evaluation, deriving a general assessment paradigm in XAI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "To begin with, let \\(\\mathcal{D} = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^n\\) be an image classification dataset with \\(n\\) samples and \\(C\\) categories, where \\(y_{i}\\) is the label of RGB image \\(\\mathbf{x}_i\\in \\mathbb{R}^{3\\times H\\times W}\\). Considering a deep network \\(\\hat{y} = f_{\\Theta}(\\mathbf{x})\\) that predicts \\(\\hat{y}\\) for input \\(\\mathbf{x}\\), an attribution method aims to generate a heatmap \\(H\\in \\mathbb{R}^{3\\times H\\times W}\\) so that each input pixel is assigned to a saliency score to the network's prediction. Such a procedure can be formulated as the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.274, + 0.786, + 0.293 + ], + "angle": 0, + "content": "\\[\nH = \\mathsf {A t t r} _ {\\Gamma} (f _ {\\Theta} | \\mathbf {x}, \\hat {y}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.298, + 0.733, + 0.312 + ], + "angle": 0, + "content": "where \\(\\mathsf{Attr}_{\\varGamma}\\) denotes a certain attribution method parameterized by \\(\\varGamma\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.787, + 0.435 + ], + "angle": 0, + "content": "To evaluate the faithfulness of different attribution methods, we adopt two most common strategies: with progressively removing the most or least salient pixels of each image in the query dataset, we measure the performance degradation of the target model. The two strategies are usually referred to as the Most Relevant First (MoRF) and the Least Relevant First (LeRF) evaluation protocols. Specifically, we construct a top-\\(t\\) removed query dataset \\(\\tilde{\\mathcal{D}}^{(t)}\\) by removing the top-\\(t\\) most or least salient pixels of the original image \\(\\mathbf{x}_i\\) and derive the modified image \\(\\tilde{\\mathbf{x}}_i^{(t)}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.442, + 0.786, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {i} ^ {(t)} = \\operatorname {R e m o v e} \\left(\\mathbf {x} _ {i}, H _ {i} | t\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.784, + 0.498 + ], + "angle": 0, + "content": "Here, the operation Remove specifies how features are eliminated, representing feature ablation in this paper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.785, + 0.559 + ], + "angle": 0, + "content": "Effective attribution methods can accurately pinpoint significant pixels, and the elimination of these pixels would result in a swift deterioration in the model's performance. Thus, the evaluation of attribution methods can be facilitated by quantifying the reduction in model performance, which is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.565, + 0.786, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\varDelta = \\frac {1}{n} \\left(\\sum_ {i = 1} ^ {n} p \\left(\\hat {y} _ {i} = y _ {i} | \\mathbf {x} _ {i}\\right) - \\sum_ {i = 1} ^ {n} p \\left(\\tilde {y} _ {i} ^ {(t)} = y _ {i} | \\tilde {\\mathbf {x}} _ {i} ^ {(t)}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.786, + 0.658 + ], + "angle": 0, + "content": "In this work, we focus on faithfulness to acquire a broadly applicable standard among multiple experimental settings. Ultimately, attribution methods are ranked according to this metric to generate a ranking list. The schematic diagram of the attribution and evaluation is depicted in Figure 1 (b)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.677, + 0.607, + 0.695 + ], + "angle": 0, + "content": "4 Meta-Rank Settings and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To acquire a comprehensive comprehension of the experimental configurations employed in previous investigations, we present a summary of the datasets and models utilized in these works, as depicted in Table 1. Notably, despite the limited choice of datasets and models, there remains considerable variation in the evaluations conducted thus far. The issue of evaluation consistency has largely been overlooked in prior research endeavors. To bridge this gap, in this section, we first establish standardized settings and subsequently provide a comprehensive outline of the Meta-Rank evaluation benchmark. The workflow of MetaRank is illustrated in Figure 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Comparison of experimental configurations employed in previous studies. The lack of consistency in these settings may lead to disparate observations. The term \"Cons.\" refers to whether the study addresses the issue of consistency. Only a limited number of works have explored the issue, but their scope is limited to specific datasets, models, or both." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.227, + 0.78, + 0.365 + ], + "angle": 0, + "content": "
MethodsVenueDatasetsModelsCons.
[45]ICLR 2014ImageNetConvNetX
[44]ICML 2016Tiny ImageNet, DNA sequenceVGG-16, 1 CNNX
[48]ICML 2017ImageNetGoogleNetX
[47]ICLR Workshop 2015ImageNet, CIFAR-10/1003 CNNsX
[43]ICML 2017MNIST, DNA sequence2 CNNsX
[53]ECCV 2014ImageNet, Caltech, PASCAL VOC2 CNNsX
[5]PLOS ONE 2015ImageNet, PASCAL VOC, MINIST1 DNN, 1 CNNX
[42]ICCV 2017ImageNet, PASCAL VOCVGG-16, AlexNetX
[26]NeurIPS 2019ImageNet, Birdsnap, Food-101ResNet-50X
[11]ECCV Workshop 2020ImageNetResNet-50✓
[3]FUZZ-IEEE 2022ImageNet, Dogs vs. Cats, MAME, MIT67AlexNet, VGG-16, ResNet-18X
[40]ICML 2022CIFAR-10, Food-101ResNet-18, ResNet-50✓
[10]ICLR Tiny Paper 2023ImageNetResNet-50X
[8]CVPR 2023ImageNetVGG-16, ResNet-34✓
[24]JMLR 2023ImageNetResNet-18X
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.393, + 0.444, + 0.409 + ], + "angle": 0, + "content": "4.1 Standardized Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.784, + 0.447 + ], + "angle": 0, + "content": "The main idea of the proposed Meta-Rank is evaluating attribution explanations on a diverse array of datasets, models, and evaluation protocols." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.787, + 0.644 + ], + "angle": 0, + "content": "Datasets. To streamline standardized evaluations of attribution methods, we meticulously select four commonly used datasets for benchmarking: NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]. NWPU-RESISC45 is a medium-scale dataset for remote sensing image classification, containing 45 categories of remote sensing images covering various scenes such as streets, airports, factories, forests, grasslands, and deserts. Food-101 is a medium-scale dataset for fine-grained food image classification, compassing a total of 101 different food categories. ImageNet is a large-scale dataset employed for image classification, consisting of 1,000 different categories. Places-365 is a large-scale dataset designed for the purposes of scene recognition, covering a wide range of 365 diverse scene categories. These datasets allow us to examine each attribution method under various domains, providing us with a more thorough view of the current literature. Further details are provided in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Models. The selection of models presents a prominent source of inconsistency in previous research. In order to comprehensively validate the effectiveness of the attribution techniques across various models, we adopt six widely employed convolutional neural networks (CNNs) to evaluate existing attribution methods. These CNNs include ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]. It should be noted that most attribution methods are specifically tailored for CNNs and cannot be directly applied to vision transformers (ViTs) [13,33,52]. As such, for maintaining an equitable evaluation across attribution methods with a consistent set of evaluation settings, our benchmark does not take ViTs into consideration. Additional details regarding these used models can be found in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluation protocols. Recent advances have discovered that different evaluation protocols, MoRF and LeRF [41], may produce conflicting rankings of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.145, + 0.238, + 0.153 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.155, + 0.784, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.296, + 0.238, + 0.309 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.299, + 0.784, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.58 + ], + "angle": 0, + "content": "Fig. 1: Meta-Rank benchmark. It is mainly divided into two stages: Test Case Generation and Meta-Rank (a). Test Case Generation: multiple factors (i.e., datasets, models and evaluation protocols) are combined to generate \\(\\tau\\) different cases. MetaRank: (1) Case Execution. All \\(m\\) competitors (i.e., attribution methods) are applied to these cases, resulting in a collection of rankings. The details of attribution evaluation on an individual case are provided in (b). (2) Ranking Fusion. All rankings \\(\\{R_1, R_2, \\ldots, R_\\tau\\}\\) are subsequently fed into this module. The comparison of two competitors is transformed into the differences in their rankings across all cases, then integrated and converted into the discrepancy in Meta-Ranks. (3) Leaderboard. Ultimately, a unified leaderboard is obtained based on the Meta-Ranks of the competitors. “\\(T\\)” is the test case, “\\(C\\)” is the competitor, and “\\(\\kappa\\)” is the Meta-Rank value." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.643 + ], + "angle": 0, + "content": "attribute methods [40]. To provide a comprehensive view of the attribution methods, the proposed Meta-Rank considers both protocols for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Test case settings. To gain a generalized evaluation for attributions, our analytical experiments evaluate existing attribution methods on a wide array of experiment settings, necessitating the initial preparation of baselines. We first train ResNet-18, Inception-v4 and VGG-19 on NWPU-RESISC45 and Food101. Subsequently, we exclusively train EfficientNet_b2 and DenseNet-121 on the NWPU-RESISC45 dataset. These baselines are implemented with the PyTorch [36] framework. Adam [31] is used to optimize the models, with the learning rate of \\( 1 \\times 10^{-4} \\) and the weight decay rate of \\( 5 \\times 10^{-4} \\). The cosine annealing is adopted as the learning rate decay schedule. Furthermore, we also incorporate pre-trained ResNet-18, ResNet-50 sourced from Places-365, and pre-trained ResNet-18, Inception-v4, VGG-19 from ImageNet-1k. The aforementioned baselines, combined with two evaluation protocols, are served as the test cases for the benchmark. Detailed information for each test case can be found in Appendix C." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.54, + 0.162 + ], + "angle": 0, + "content": "4.2 The Proposed Benchmark Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.167, + 0.788, + 0.257 + ], + "angle": 0, + "content": "We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. The main idea of the Meta-Rank metric is viewing each specific configuration of the model, the dataset and the evaluation protocol as one test case. Meta-Rank aggregates the ranking results (i.e., rank the rankings) from a diverse array of test cases, thereby mitigating biases associated with specific configurations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.789, + 0.432 + ], + "angle": 0, + "content": "Formally, let the triplet \\(\\langle D_i,M_j,P_k\\rangle\\) be one test case for evaluating the attribution algorithms, where \\(D\\), \\(M\\) and \\(P\\) denote the dataset, model and evaluation protocol respectively. The subscript \\(i\\), \\(j\\) and \\(k\\) denote one specific configuration from the candidate datasets, models and protocols as described in Section 4.1. Let \\(\\mathcal{C} = \\{C_1,C_2,\\dots,C_m\\}\\) be the set of the competitors (i.e., attribution methods). In the proposed benchmark, all the competitors are evaluated on every generated test case, thus producing a ranking among the competitors. For the test case \\(\\langle D_i,M_j,P_k\\rangle\\), we use \\(R_{(i,j,k)} = (r_1^{(i,j,k)},r_2^{(i,j,k)},\\dots,r_m^{(i,j,k)})\\) to denote the rankings of all the competitors, where \\(r_m^{(i,j,k)}\\) denotes the rank of the \\(m\\)-th attribution method. After evaluating the attribution methods on all the test cases, the probability that \\(C_p\\) outperforms \\(C_q\\) is formally defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.435, + 0.787, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} _ {q \\prec p} = \\frac {1}{| \\Omega |} \\sum_ {(i, j, k)} ^ {\\Omega} \\mathbf {1} \\left(r _ {q} ^ {(i, j, k)} \\prec r _ {p} ^ {(i, j, k)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.548 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) denotes the indicator function, and \\(r_q \\prec r_p\\) represents the ranking \\(r_q\\) is outperformed by \\(r_p\\). \\(\\Omega\\) is the joint space of datasets, models and evaluation protocols. The probability \\(\\mathsf{P}_{q \\prec p}\\) is then converted into the logistic odds, obtaining a more sensitive metric \\(\\mathsf{Logit}_{q \\prec p}\\) that can better capture slight differences between \\(C_p\\) and \\(C_q\\):" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.552, + 0.786, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L o g i t} _ {q \\prec p} = \\log \\operatorname {o d d s} \\left(C _ {p}, C _ {q}\\right) = \\log \\left(\\mathrm {P} _ {q \\prec p} / \\left(1 - \\mathrm {P} _ {q \\prec p}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Based on the above derivation process, we can obtain a set of standardized results \\(\\{\\mathsf{Logit}_{q\\prec p}\\}\\) that quantify the performance differences between any two competitors on all test cases. Then, we transform the performance discrepancy into the difference on the Meta-Rank metric:" + }, + { + "type": "equation", + "bbox": [ + 0.287, + 0.637, + 0.787, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\kappa_ {1} - \\kappa_ {2} = \\operatorname {L o g i t} _ {2 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 2 \\\\ \\kappa_ {1} - \\kappa_ {3} = \\operatorname {L o g i t} _ {3 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 3 \\\\ \\kappa_ {1} - \\kappa_ {4} = \\operatorname {L o g i t} _ {4 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 4 \\\\ \\dots & \\dots \\\\ \\kappa_ {2} - \\kappa_ {3} = \\operatorname {L o g i t} _ {3 \\prec 2}, & \\text {C o m p e t i t o r} \\# 2 v s. \\# 3 \\\\ \\kappa_ {2} - \\kappa_ {4} = \\operatorname {L o g i t} _ {4 \\prec 2}, & \\text {C o m p e t i t o r} \\# 2 v s. \\# 4 \\\\ \\dots & \\dots \\\\ \\kappa_ {m - 1} - \\kappa_ {m} = \\operatorname {L o g i t} _ {m \\prec m - 1}. & \\text {C o m p e t i t o r} \\# \\mathrm {m} - 1 v s. \\# \\mathrm {m} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Here, \\(\\kappa_{m}\\) represents the Meta-Rank value of \\(C_m\\). After solving the system of equations in Equation 6, we can derive a score \\(\\kappa\\) for each competitor, serving as an identifier for its comprehensive performance. Through the arrangement of competitors in descending order of their \\(\\kappa\\) values, a unified leaderboard can finally be constructed." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.359, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.232, + 0.356, + 0.242 + ], + "angle": 0, + "content": "(a) Food-ResNet18(M)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.501, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.232, + 0.496, + 0.243 + ], + "angle": 0, + "content": "(b) Food-ResNet18(L)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.642, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.231, + 0.632, + 0.243 + ], + "angle": 0, + "content": "(c) IN-ResNet18(M)" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.144, + 0.79, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.231, + 0.774, + 0.243 + ], + "angle": 0, + "content": "(d) IN-ResNet18(L)" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.243, + 0.364, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.329, + 0.364, + 0.34 + ], + "angle": 0, + "content": "(e) Food-Inceptionv4(M)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.243, + 0.505, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.329, + 0.505, + 0.34 + ], + "angle": 0, + "content": "(f) Food-Inceptionv4(L)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.243, + 0.647, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.33, + 0.642, + 0.341 + ], + "angle": 0, + "content": "(g) IN-Inceptionv4(M)" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.243, + 0.79, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.33, + 0.783, + 0.341 + ], + "angle": 0, + "content": "(h) IN-Inceptionv4(L)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.341, + 0.36, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.428, + 0.35, + 0.44 + ], + "angle": 0, + "content": "(i) Food-VGG19(M)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.341, + 0.501, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.428, + 0.489, + 0.44 + ], + "angle": 0, + "content": "(j) Food-VGG19(L)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.341, + 0.642, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.428, + 0.626, + 0.44 + ], + "angle": 0, + "content": "(k) IN-VGG19(M)" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.341, + 0.79, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.428, + 0.767, + 0.44 + ], + "angle": 0, + "content": "(1) IN-VGG19(L)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.229, + 0.444, + 0.773, + 0.454 + ], + "angle": 0, + "content": "Saliency GB -XG DeepLift -Deconv LRP -GuidedGradCam IG" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.471, + 0.789, + 0.541 + ], + "angle": 0, + "content": "Fig.2: Evaluation results on 12 test cases. These cases consist of two datasets: Food-101 (a)(b)(e)(f)(i)(j), ImageNet-1k (c)(d)(g)(h)(k)(l), three models: ResNet-18 (a)(b)(c)(d), Inception-v4 (e)(f)(g)(h), VGG-19 (i)(j)(k)(l), and two protocols: MoRF(a)(c)(e)(g)(i)(k), LeRF(b)(d)(f)(h)(j)(l). \"Baseline\" represents the accuracy of the model when no features are ablated." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.383, + 0.584 + ], + "angle": 0, + "content": "5 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.63 + ], + "angle": 0, + "content": "This section provides a detailed elaboration of extensive experiments on Meta-Rank, with a particular emphasis on the following pivotal questions." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.637, + 0.807, + 0.667 + ], + "angle": 0, + "content": "- RQ1: Are attribution methods consistently effective across diverse settings (i.e., datasets, models, evaluation protocols) and distinct checkpoints? (Section 5.1)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.667, + 0.785, + 0.696 + ], + "angle": 0, + "content": "- RQ2: Does the state-of-the-art evaluation method, ROAD, yield consistent rankings across different test cases? (Section 5.2)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.697, + 0.785, + 0.726 + ], + "angle": 0, + "content": "- RQ3: According to our proposed Meta-Rank, which attribution methods exhibit superior performance across a wide range of cases? (Section 5.3)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.727, + 0.785, + 0.757 + ], + "angle": 0, + "content": "- RQ4: Can Meta-Rank conduct the evaluation of substantial methods within a feasible timeframe? (Section 5.4)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.637, + 0.807, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.522, + 0.802 + ], + "angle": 0, + "content": "5.1 RQ1: Consistency Investigation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we investigate the consistency of performance rankings for existing attribution methods across diverse evaluation settings. Specifically, we focus" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.146, + 0.4, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.264, + 0.352, + 0.276 + ], + "angle": 0, + "content": "(a) MoRF" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.146, + 0.592, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.264, + 0.543, + 0.276 + ], + "angle": 0, + "content": "(b) LeRF" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.157, + 0.777, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.264, + 0.754, + 0.276 + ], + "angle": 0, + "content": "(c) MoRF v.s. LeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 3: Spearman correlation among nine test cases (NWPU, Food, and ImageNet datasets paired with ResNet-18, Inception-v4, and VGG-19 models) in MoRF (a) and LeRF (b), and between MoRF and LeRF on the same nine cases (c). The labels along the horizontal arrow are equivalent to those along the vertical arrow. Here, we define a correlation score higher than 0.8 as strongly correlated (■), between 0.6 and 0.8 as moderately correlated (▲), and lower than 0.6 as weakly correlated (▼)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.508 + ], + "angle": 0, + "content": "on the model-, dataset-, protocol- and checkpoint-level ranking consistencies by varying these factors in the evaluation. To this end, we assess the aforementioned eight attribution methods on all the generated test cases, i.e., with six models, four datasets and two ablation protocols (MoRF and LeRF). This yields a total of 26 groups of results, each corresponding to one test case. Some experimental results are presented in Figure 2. For more details of the experimental results, please refer to Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Model-level Consistency. It is evident that the attribution methods exhibit varying performance across different models. To elaborate: (1) LRP has the worst performance on the \"ImageNet-Inceptionv4(M)\" case but significantly better results on the \"ImageNet-ResNet18(M)\" and \"ImageNet-VGG19(M)\" cases. (2) Integrated Gradients and Guided Grad-CAM exhibit comparable performance on the \"Food-ResNet18(M)\" case, but show great disparity on the \"Food-VGG19(M)\" case. For a more clear measurement of the inconsistency, we calculate the Spearman Rank correlations among several cases and illustrate them in Figure 3. Here, we extract area under the curve (AUC) [4] of all attribution methods (detailed in Appendix D) to generate a ranking list for each case. In the MoRF protocol, removing the most critical pixels identified by the attribution method causes a substantial drop in model performance. Thus, a lower AUC value indicates a superior method. And the LeRF protocol follows the inverse principle. The results suggest a pronounced weakness in the consistency at the model level, with a strong correlation found in only 2 cases (i.e., the correlation between ResNet-18 and Inception-v4 on the Food-101 dataset in both MoRF and LeRF protocols) out of 18 cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Dataset-level Consistency. Analogously, attribution methods have shown noticeable inconsistency in performance on different datasets. For instance, Guided Backpropagation demonstrates superior performance on \"ImageNet-VGG19(M)\" yet ranks third-worst on \"Food-VGG19(M)\". In addition, according to the results of Spearman analysis, only \\(\\frac{5}{18}\\) cases show strong correlations at the dataset level." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.146, + 0.495, + 0.219 + ], + "angle": 0, + "content": "Table 2: AUC rankings of the eight methods at different checkpoints along the same trajectory. Here, ResNet-18 trained on NWPU-RESISC45 is taken as an illustrative example." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.228, + 0.49, + 0.356 + ], + "angle": 0, + "content": "
Epoch-50Epoch-100Epoch-150Epoch-200
MoRF
LeRF
SaliencyIXGDeconvGuidedGradCam
GBDeepLiftLRPIG
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.145, + 0.788, + 0.257 + ], + "angle": 0, + "content": "Table 3: Comparison of Spearman results among four cases between ROAD and feature ablation. \\( a \\mid b \\) represents a pair of Spearman results, where \\( a \\) denotes using feature ablation (in Section 5.1) and \\( b \\) denotes using noisy linear imputation. The settings of color blocks in this table are the same as those in Figure 3." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.268, + 0.782, + 0.358 + ], + "angle": 0, + "content": "
NWPU-ResNet18Food-Inceptionv4
MoRF (#1)LeRF (#2)MoRF (#3)LeRF (#4)
#1| || || || |||
#2| || || || |||
#3| || || || |||
#4| || || || |||
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.505 + ], + "angle": 0, + "content": "Protocol-level Consistency. The evaluation results indicate that significant inconsistencies between the MoRF and LeRF protocols still exist. A detailed comparison between \"Food-Inceptionv4(M)\" and \"Food-Inceptionv4(L)\" reveals marked discrepancies in Saliency method under MoRF versus LeRF. Besides, the Spearman correlation between MoRF and LeRF is computed for nine test cases (in Figure 3 (c)). It can be easily seen that the correlation of all cases at the protocol level is either moderately correlated or weakly correlated, with no strong correlation observed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.734 + ], + "angle": 0, + "content": "Checkpoint-level Consistency. We also consider the consistency of attribution methods at different training stages of the model. To illustrate, using ResNet-18 trained on NWPU-RESISC45 as a case study, we extract distinct checkpoints from the same training trajectory and employ two evaluation protocols for attribution analysis. The results of the AUC rankings are recorded in Table 2. It can be easily observed that the rankings of attribution methods demonstrate high consistency across different checkpoints. In MoRF, the outcomes of Epoch-200 differ by only one distinction from those at both Epoch-100 and Epoch-150, while exhibiting slightly greater disparities when compared to those at Epoch-50. Similarly for LeRF, the rankings at Epoch-200 also exhibit a single deviation from those at Epoch-100 and Epoch-150, respectively, and show two distinctions in comparison to those at Epoch-50. These findings suggest a notable consistency in the ranking of attribution methods across different checkpoints. Moreover, the rankings tend to stabilize rapidly as the model converges. More detailed results under other settings can be found in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.51, + 0.772 + ], + "angle": 0, + "content": "5.2 RQ2: Necessity of Meta-Rank" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To further substantiate the indispensability of the Meta-Rank benchmark, we employ the state-of-the-art attribution evaluation method known as ROAD for assessing the attribution methods across various cases. Starting from information theoretic analyses, ROAD identifies the class information leakage problem" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.156, + 0.498, + 0.24 + ], + "angle": 0, + "content": "Table 4: Leaderboard of the eight attribution methods. The larger Meta-Rank indicates the stronger ability of attribution method. The top three methods exhibit more positive performance, while the remaining five methods perform negatively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.253, + 0.498, + 0.338 + ], + "angle": 0, + "content": "
MethodsMeta-Rank Score#Rank
InputGradient2.32001
Integrated Gradients2.07362
DeepLift1.07803
Saliency-0.04834
LRP-0.22345
Guided Backpropagation-0.25856
Guided Grad-CAM-0.96797
Deconvolution-3.97348
" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.144, + 0.768, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.268, + 0.786, + 0.336 + ], + "angle": 0, + "content": "Fig. 4: Time consumption of the \"Ranking Fusion\" module in Meta-Rank. 2, 4, 8, 16, 32, and 64 are the number of evaluated attribution methods. The gray bar represents the relative error." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.531 + ], + "angle": 0, + "content": "dictated by the ablation mask and finally derives the noisy linear imputation for minimally revealing imputation, which is empirically demonstrated to produce consistent rankings of attribution methods between evaluations with MoRF and LeRF. However, it overlooks consistency at both the model and dataset levels. Thus, we randomly select four cases, namely NWPU-ResNet18(M), NWPU-ResNet18(L), Food-Inceptionv4(M), and Food-Inceptionv4(L), to assess the eight aforementioned attribution methods using ROAD. The evaluation metric used in this study aligns with those in RQ1. Table 3 presents the Spearman correlations among the evaluation results of the four cases. Further information regarding the accuracy decay curves is available in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.787, + 0.637 + ], + "angle": 0, + "content": "It can be easily seen that, even when ROAD is employed as a substitute for feature ablation, the correlation between the rankings of attribution methods on different cases remains inconsistent. Specifically, all correlations demonstrate either moderate or weak, with no instances of strong correlation. This indicates that relying solely on ROAD for evaluating attribution methods across distinct cases is inadequate. Consequently, there is a critical need for a cross-case benchmark to ensure robust ranking results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.64, + 0.677 + ], + "angle": 0, + "content": "5.3 RQ3: Attribution Evaluation with Meta-Rank" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Following the evaluation scheme outlined in the Meta-Rank benchmark (described in Section 4), we have carried out a unified measurement of the eight attribution methods. The consistency validation experiments in Section 5.1 generate evaluation results for attribution methods across distinct configurations, encompassing four datasets, six models, and two protocols, resulting in 26 sets of data. We then transform these results into AUC scores, which serve as the performance metric for the attribution methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Meta-Rank generates a comprehensive value for each attribution method. These uniquely determined scores enable us to establish a unified ranking list for all the methods. The finalized leaderboard for the eight attribution methods" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "is presented in Table 4. The results reveal InputGradient as the top-performing approach, while Deconvolution exhibits the poorest performance. (1) In particular, InputGradient achieves the highest score among the eight methods. By element-wise multiplying the gradients with the inputs, InputGradient produces more stable attributions in the face of perturbations, which contributes to its superior performance across multiple test cases. (2) Furthermore, the second-ranked method, Integrated Gradients, also shows great potential, with its score closely approaching that of the top-ranked method. (3) In contrast, Deconvolution obtains the lowest score, likely attributable to the hierarchical structure of the network accumulating noise during backpropagation. It is worth noting that a Meta-Rank score can assume both positive and negative values. Positive values signify comparatively superior attribution performance, whereas negative values indicate inferior performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.435 + ], + "angle": 0, + "content": "The experimental findings demonstrate that Meta-Rank furnishes a comprehensive benchmark for attribution evaluation across diverse test cases, while effectively ameliorating the issue of inconsistent performance across heterogeneous cases. It computes the discrepancies between methods and converts them into the differences in capacity to derive the Meta-Rank. This result is unique, consistent, and aligned with practical application needs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.453, + 0.513, + 0.468 + ], + "angle": 0, + "content": "5.4 RQ4: Efficiency of Meta-Rank" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.64 + ], + "angle": 0, + "content": "To measure the time consumption, we conduct experiments on the CPU device (Intel(R) Xeon(R) Platinum 8260L CPU @ 2.40GHz) using different numbers of methods. Here, we focus on the additional \"Ranking Fusion\" module, as the \"Case Execution\" module has been generically applied in all evaluations. In particular, we compare the time expenses of Meta-Rank during the calculation of the 2, 4, 8, 16, 32, and 64 methods. It is important to highlight that the first three test groups (i.e., 2, 4, and 8 methods) rely on real data gathered from previous experiments, whereas the last three test groups (i.e., 16, 32, and 64 methods) employ simulated data randomly generated. The data here refers to a set of accuracies (i.e., \\(\\{acc_i\\}\\) in Figure 1) obtained on each test case according to every attribution method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.763 + ], + "angle": 0, + "content": "The results in Figure 4 indicate that as the number of attribution methods increases, the computation time also increases. However, since the number of attribution methods is finite, the computational time remains within an acceptable range. For example, when 64 methods are involved in the Meta-Rank evaluation, the time required to generate the leaderboard is approximately 0.12s, which completely meets expectations. Additionally, using GPUs for Meta-Rank calculations can further enhance the computation speed. This efficiency is beneficial for the promotion and application of Meta-Rank." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.782, + 0.355, + 0.797 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Choice of faithfulness metric. Although there are many dimensions for attribution evaluation in previous work [24], faithfulness remains the most solid" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "and general criteria for benchmarking attribution methods. In the context of multiple experimental settings, as such, we opt to use it as the base metric \\(\\Delta\\) in our framework to facilitate subsequent Meta-Rank assessment and then enable broader evaluation on novel tasks. Meanwhile, Meta-Rank can also accommodate evaluations on other dimensions by enabling the derivation of a ranking list along each respective evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Missingness bias in feature ablation. Previous work has discussed the missingness bias problem caused by the removal of pixels through masking values in CNNs [29]. Hooker et al. proposed ROAR to address this issue, but it inadvertently introduced the problem of information leakage. Subsequently, Rong et al. developed ROAD to mitigate this leakage, but it failed to fully remove the pixels (explained in Appendix E). We recognize that completely eliminating this bias poses substantial difficulties, since the root of missingness bias lies in the operation of removing pixels within CNNs. Furthermore, the impact of bias exhibits variability across individual cases, also posing additional challenges in quantification and measurement. However, we believe that with a more diverse array of test cases, the missingness bias is more neutralized. Future research will further explore solutions for eliminating this bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Scalability of Meta-Rank. When introducing new attribution methods to the ranking, we initially conduct tests on several existing cases. The time complexity of evaluating a method on an individual case is linear, which has been demonstrated in previous studies [37, 40]. Subsequently, the obtained results, in conjunction with historical data, are fed into the \"Ranking Fusion\" module to generate an updated ranking list. The time overhead associated with this process is also optimistic, as illustrated in Section 5.4. This flexibility allows for the inclusion of any attribution method into the leaderboard at any time. Future endeavors will focus on expanding the corpus of test cases and pursuing further computational efficiency by leveraging parallel computing paradigms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.36, + 0.612 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address the lack of systematic research in feature attribution, we propose Meta-Rank, a novel standardized benchmark for robust and unified evaluation of methods. Meta-Rank aggregates relative comparisons between all method pairs across diverse settings, mitigating inconsistencies. Through extensive experiments, we reveal three critical insights: diverse evaluation settings yield divergent performance rankings, assessment results remain consistent across distinct checkpoints along the same training trajectory, and previous attempts at consistent evaluation exhibit limitations on heterogeneous models and datasets. Meta-Rank provides a leaderboard that highlights the broader applicability of InputGradient and Integrated Gradients. Remarkably, Meta-Rank exhibits exceptional computational efficiency, enabling ranking computations to be completed within milliseconds even for a substantial number of methods. In the future, we hope the proposed benchmark will facilitate the development of faithful and consistent explanations, fostering transparency and accountability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.812, + 0.252 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (62106220), Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies, Ningbo 2025 Science and Technology Innovation Major Project (No. 2022Z072), and the Fundamental Research Funds for the Central Universities (226-2024-00058)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.787, + 0.331 + ], + "angle": 0, + "content": "1. Adadi, A., Berrada, M.: Peeking inside the black-box: a survey on explainable artificial intelligence (xai). IEEE access 6, 52138-52160 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.786, + 0.373 + ], + "angle": 0, + "content": "2. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.787, + 0.413 + ], + "angle": 0, + "content": "3. Arias-Duart, A., Parés, F., Garcia-Gasulla, D., Giménez-Ábalos, V.: Focus! rating xai methods and finding biases. In: 2022 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). pp. 1-8. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.414, + 0.787, + 0.455 + ], + "angle": 0, + "content": "4. Atanasova, P., Simonsen, J.G., Lioma, C., Augenstein, I.: A diagnostic study of explainability techniques for text classification. arXiv preprint arXiv:2009.13295 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.787, + 0.497 + ], + "angle": 0, + "content": "5. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.787, + 0.524 + ], + "angle": 0, + "content": "6. Bastings, J., Aziz, W., Titov, I.: Interpretable neural predictions with differentiable binary variables. arXiv preprint arXiv:1905.08160 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.787, + 0.551 + ], + "angle": 0, + "content": "7. Bhatt, U., Weller, A., Moura, J.M.: Evaluating and aggregating feature-based model explanations. arXiv preprint arXiv:2005.00631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.787, + 0.607 + ], + "angle": 0, + "content": "8. Binder, A., Weber, L., Lapuschkin, S., Montavon, G., Müller, K.R., Samek, W.: Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16143-16152 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.787, + 0.661 + ], + "angle": 0, + "content": "9. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.787, + 0.689 + ], + "angle": 0, + "content": "10. Brocki, L., Chung, N.C.: Evaluation of interpretability methods and perturbation artifacts in deep neural networks. arXiv preprint arXiv:2203.02928 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.744 + ], + "angle": 0, + "content": "1. Brunke, L., Agrawal, P., George, N.: Evaluating input perturbation methods for interpreting cnns and saliency map comparison. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 120-134. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "2. Chattopadhy, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Gradcam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 782-791 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.303, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "15. Colin, J., Fel, T., Cadène, R., Serre, T.: What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems 35, 2832-2845 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "16. Covert, I., Lundberg, S.M., Lee, S.I.: Understanding global feature contributions with additive importance measures. Advances in Neural Information Processing Systems 33, 17212-17223 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "17. Deng, H., Zou, N., Du, M., Chen, W., Feng, G., Yang, Z., Li, Z., Zhang, Q.: Understanding and unifying fourteen attribution methods with taylor interactions. arXiv preprint arXiv:2303.01506 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "19. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.383, + 0.785, + 0.424 + ], + "angle": 0, + "content": "20. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.424, + 0.785, + 0.466 + ], + "angle": 0, + "content": "21. Goyal, Y., Wu, Z., Ernst, J., Batra, D., Parikh, D., Lee, S.: Counterfactual visual explanations. In: International Conference on Machine Learning. pp. 2376-2384. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.466, + 0.785, + 0.508 + ], + "angle": 0, + "content": "22. Hartley, T., Sidorov, K., Willis, C., Marshall, D.: Explaining failure: Investigation of surprise and expectation in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 12-13 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.785, + 0.605 + ], + "angle": 0, + "content": "24. Hedström, A., Weber, L., Krakowczyk, D., Bareeva, D., Motzkus, F., Samek, W., Lapuschkin, S., Höhne, M.M.C.: Quantus: An explainable ai toolkit for responsible evaluation of neural network explanations and beyond. Journal of Machine Learning Research 24(34), 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.605, + 0.785, + 0.646 + ], + "angle": 0, + "content": "25. Hesse, R., Schaub-Meyer, S., Roth, S.: Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3981-3991 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.646, + 0.785, + 0.688 + ], + "angle": 0, + "content": "26. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.688, + 0.785, + 0.729 + ], + "angle": 0, + "content": "27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.729, + 0.785, + 0.771 + ], + "angle": 0, + "content": "28. Ismail, A.A., Gunady, M., Corrada Bravo, H., Feizi, S.: Benchmarking deep learning interpretability in time series predictions. Advances in neural information processing systems 33, 6441-6452 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "29. Jain, S., Salman, H., Wong, E., Zhang, P., Vineet, V., Vermprala, S., Madry, A.: Missingness bias in model debugging. arXiv preprint arXiv:2204.08945 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "30. Khakzar, A., Khorsandi, P., Nobahari, R., Navab, N.: Do explanations explain? model knows best. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10244-10253 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Evaluation Consistency of Attribution-based Explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "31. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.218 + ], + "angle": 0, + "content": "32. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.219, + 0.787, + 0.26 + ], + "angle": 0, + "content": "33. Leem, S., Seo, H.: Attention guided cam: Visual explanations of vision transformer guided by self-attention. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2956-2964 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.261, + 0.786, + 0.289 + ], + "angle": 0, + "content": "34. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "35. Nguyen, G., Kim, D., Nguyen, A.: The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems 34, 26422-26436 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.332, + 0.787, + 0.387 + ], + "angle": 0, + "content": "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32, 8026-8037 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.388, + 0.787, + 0.416 + ], + "angle": 0, + "content": "37. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.417, + 0.787, + 0.445 + ], + "angle": 0, + "content": "38. Rao, S., Böhle, M., Schiele, B.: Better understanding differences in attribution methods via systematic evaluations. arXiv preprint arXiv:2303.11884 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "39. Ribeiro, M.T., Singh, S., Guestrin, C.: \"why should i trust you?\" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "40. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A consistent and efficient evaluation strategy for attribution methods. In: International Conference on Machine Learning. pp. 18770-18795. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.R.: Evaluating the visualization of what a deep neural network has learned. IEEE transactions on neural networks and learning systems 28(11), 2660-2673 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.573, + 0.787, + 0.627 + ], + "angle": 0, + "content": "42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.629, + 0.787, + 0.67 + ], + "angle": 0, + "content": "43. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.671, + 0.787, + 0.712 + ], + "angle": 0, + "content": "44. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not just a black box: Learning important features through propagating activation differences. arXiv preprint arXiv:1605.01713 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.713, + 0.787, + 0.755 + ], + "angle": 0, + "content": "45. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: visualising image classification models and saliency maps. In: Proceedings of the International Conference on Learning Representations (ICLR). ICLR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.756, + 0.787, + 0.783 + ], + "angle": 0, + "content": "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.784, + 0.787, + 0.812 + ], + "angle": 0, + "content": "47. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "48. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Duan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "50. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "51. Wang, Y., Wang, X.: A unified study of machine learning explanation evaluation metrics. arXiv preprint arXiv:2203.14265 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.327 + ], + "angle": 0, + "content": "52. Xue, M., Huang, Q., Zhang, H., Cheng, L., Song, J., Wu, M., Song, M.: Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In: 33rd International Joint Conference on Artificial Intelligence, IJCAI-24. International Joint Conferences on Artificial Intelligence Organization (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "53. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 818-833. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "54. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-down neural attention by excitation backprop. International Journal of Computer Vision 126(10), 1084-1102 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.453 + ], + "angle": 0, + "content": "55. Zhang, Y., Khakzar, A., Li, Y., Farshad, A., Kim, S.T., Navab, N.: Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems 34, 20040-20051 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "57. Zhou, Y., Booth, S., Ribeiro, M.T., Shah, J.: Do feature attribution methods correctly attribute features? In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9623-9633 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.564 + ], + "angle": 0, + "content": "58. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. arXiv preprint arXiv:1702.04595 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.564 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_origin.pdf b/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0009ae073c29d81076d5f0d57d2dba599ac6cbc8 --- /dev/null +++ b/2024/On the Evaluation Consistency of Attribution-based Explanations/84f419f4-dac9-4b0b-a91c-aee1899c20cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85d9a702a2849ccf44241c34875f34faa36fe27864ae7eba356e5492195543cb +size 3167606 diff --git a/2024/On the Evaluation Consistency of Attribution-based Explanations/full.md b/2024/On the Evaluation Consistency of Attribution-based Explanations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2bbc56ef34f1176bdf561033451f79d38f5ec2a4 --- /dev/null +++ b/2024/On the Evaluation Consistency of Attribution-based Explanations/full.md @@ -0,0 +1,328 @@ +# On the Evaluation Consistency of Attribution-based Explanations + +Jiarui Duan $^{1,3}$ , Haoling Li $^{2,3}$ , Haofei Zhang $^{1,3}$ , Hao Jiang $^{4}$ , Mengqi Xue $^{5}$ , Li Sun $^{6}$ , Mingli Song $^{1,3}$ , and Jie Song $^{2,3\dagger}$ + +1 College of Computer Science and Technology, Zhejiang University + +$^{2}$ School of Software Technology, Zhejiang University + +3 Key Laboratory of Visual Perception (Zhejiang University), Ministry of Education and Microsoft + +4 Alibaba Group + +5 Hangzhou City University + +$^{6}$ Ningbo Innovation Center, Zhejiang University + +{jerryduan,hollylee,haofeizhang,lsun,brooksong,sjie}@zju.edu.cn; + +aoshu.jh@alibaba-inc.com; mqxue@hzcu.edu.cn + +Abstract. Attribution-based explanations are garnering increasing attention recently and have emerged as the predominant approach towards eXplanable Artificial Intelligence (XAI). However, the absence of consistent configurations and systematic investigations in prior literature impedes comprehensive evaluations of existing methodologies. In this work, we introduce Meta-Rank, an open platform for benchmarking attribution methods in the image domain. Presently, Meta-Rank assesses eight exemplary attribution methods using six renowned model architectures on four diverse datasets, employing both the Most Relevant First (MoRF) and Least Relevant First (LeRF) evaluation protocols. Through extensive experimentation, our benchmark reveals three insights in attribution evaluation endeavors: 1) evaluating attribution methods under disparate settings can yield divergent performance rankings; 2) although inconsistent across numerous cases, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts at consistent evaluation fare no better than baselines when extended to more heterogeneous models and datasets. Our findings underscore the necessity for future research in this domain to conduct rigorous evaluations encompassing a broader range of models and datasets, and to reassess the assumptions underlying the empirical success of different attribution methods. Our code is publicly available at https://github.com/TreeThree-R/Meta-Rank. + +Keywords: Attribution evaluation $\cdot$ Disparate settings $\cdot$ Consistency + +# 1 Introduction + +Explainable Artificial Intelligence (XAI) [1] has emerged as a prominent research field within computer vision, with a multitude of approaches proposed for ex- + +plaining and interpreting the internal mechanisms of diverse neural network architectures across various application domains. Among these approaches, attribution methods have gained significant traction in the realm of deep learning. By generating an explanation map in the form of a heatmap (also referred to as a relevance map or a saliency map), these methods attribute the prediction to specific regions within the input image, thereby facilitating interpretation. Over the years, researchers have devised a large number of attribution methods from diverse perspectives, including gradient-based methods [5, 43, 44, 47, 48], perturbation-based methods [34, 37, 39, 53], and CAM-based methods [12, 42]. + +Nevertheless, given the absence of ground truth regarding the inner workings of deep neural networks, evaluating attributions directly poses a formidable challenge. Presently, various studies have endeavored to assess the reliability of attribution methods, which can be roughly categorized into two schools: expert-grounded methods [6, 15, 16, 32, 35, 54], and functional-grounded methods [4, 17, 26, 37, 40, 41, 51]. Expert-grounded evaluations capture how well the explanations imitate the human-annotated importance of the image regions. Such methods not only require human effort, but also suffer from the faithfulness1 issue, as deep neural networks could rely on non-intuitive, spurious cues that differ from human perception [20]. In contrast, functional-based approaches evaluate attribution methods based on some fundamental axioms. For instance, feature ablation [37], as a functional-grounded evaluation, systematically removes features in a predetermined order and observes the resulting changes in predictions. This approach enables more faithful assessments of the attribution explanations by comparing the model behavior with and without specific individual features [19, 21, 26, 37, 39, 58]. Unfortunately, the current literature often evaluates attributions under divergent settings, with different datasets, models and evaluation protocols [26, 40]. The efficacy of attribution methods in practice is often called into question due to restricted and inconsistent experimental conditions. + +The goal of this work is to gain a thorough understanding of the current state of attribution methods while setting the stage for critical problems to be worked on. To this end, we present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines. Our benchmark carefully examines eight widely-used attribution methods (Saliency [45], Input $\odot$ Gradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5] and Guided Grad-CAM [42]), with a wide range of models (ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]), on various domain datasets (NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]), using two evaluation protocols (Most Relevant First and Least Relevant First). Specifically, we place a strong emphasis on comprehensive and consistent experimental settings that have been largely overlooked in previous works. Through extensive experiments on the exhaustive combinations of methods, datasets, models and evaluation protocols, + +our benchmark reveals three main findings regarding previous attribution evaluation endeavors: 1) evaluating attribution methods under disparate experimental settings, such as different model architectures, datasets and evaluation protocols, can yield highly divergent performance rankings; 2) although highly inconsistent across datasets, models and evaluation protocols, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts towards more consistent evaluation, like Remove and Debias (ROAD) [40], fare no better than simple baselines when it is extended to more heterogeneous models and datasets. + +We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. Meta-Rank aggregates ranking results from a diverse array of settings, thereby mitigating biases associated with specific configurations. By employing this metric, we conduct systematic comparisons among existing attribution methods, and provide extensive experiment results and discussions on the status quo. Beyond extensive experiments, our benchmark is designed as a flexible framework that standardizes experimental settings and simplifies the integration of novel algorithmic implementations. We hope our work will not only facilitate reliable evaluations of attribution methods across a broader range of models and datasets, but also inspire the research community to delve further into the development of more consistent attribution methods. + +In summary, we make the following contributions: + +- We present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines. +- Through extensive experiments, we identify three pivotal insights from previous attribution evaluation efforts, which provide directions for future advancements in attribution research. +- We propose an evaluation metric to facilitate a more comprehensive benchmark comparison across multiple settings, with which extensive experiments and discussions are provided on the status quo. + +# 2 Related Work + +The development of attribution methods has prompted researchers to seek the best approach to assist models in downstream tasks. In this section, we delineate the progress of attribution evaluation chronologically. + +Expert-grounded evaluations [6, 16, 35, 54, 55] rely on human judgment to evaluate attribution methods. For example, Zhang et al. [54] proposed the pointing-game method as a means of evaluating attention maps based on detecting whether the highest value point in the map hit the target category. However, such methods are inherently subjective and fail to guarantee the reliability of evaluation results. + +Functional-grounded evaluations [2,7,26,37,41] provide more objective evaluations of the model by comparing how it behaves when certain individual features are present or absent. Samek et al. [41] proposed two evaluation protocols, + +namely Most Relevant First (MoRF) and Least Relevant First (LeRF), to observe the impact on the model's output. In MoRF, effective attribution methods should produce a greater reduction in model accuracy when removing features with higher attribution values, and vice versa. Similarly, [37] suggested evaluating attribution methods through the deletion and insertion of features by using their attribution values. These methods violate the assumption of a consistent distribution between training and test data, making it hard to determine if the decline in model performance solely results from feature deletion. Hooker et al. [26] further demonstrated that the performance degradation from feature deletion partly stems from such a distribution shift. To address this, they developed RemOve And Retrain (ROAR), which resolves inconsistent distributions by synchronously deleting the features of the training set and test set based on the attributions of each attribution method, and then retrained the model on the generated datasets. The effectiveness of the attribution method is measured by the retrained model's accuracy decline. Better methods identify more important features, whose removal causes sharper performance drops. ROAR has been widely adopted in evaluation studies [7, 22, 28]. However, due to its exorbitant expense, this type of evaluation method is unfeasible for scaling up. Rong et al. [40] discovered inconsistent MoRF and LeRF rankings from ROAR, which they attributed to information leakage based on information theory analysis. Thus, they proposed a noisy linear imputation method, named RemOve And Debias (ROAD), to effectively generate consistent results. However, as a pixel filling method, ROAD still suffers from several inherent defects, e.g., the features are replaced with substantial values instead of entirely dropped, which can hardly guarantee their neutral effects on the predictions. Deng et al. [17] noted the lack of unified theoretical frameworks for evaluating attribution methods. To resolve this limitation, they proposed a Taylor interaction system to unify fourteen attribution methods under a common mathematical formulation and suggested three principles for evaluation. However, applying this framework to new methods requires manual theoretical analysis, hampering both scalability and efficiency. Recently, attribution evaluation has been explored from more diverse perspectives [30, 38, 57], including comprehensive viewpoints [24, 25] as well as the consistency lens [8, 11, 26, 40]. Unfortunately, the absence of standardized experimental settings and systematic evaluations in previous literatures poses challenges for comparing existing methods universally. To overcome this limitation, we propose Meta-Rank, a novel evaluation benchmark, to provide a comprehensive evaluation of attribution methods across multiple settings. + +# 3 Preliminaries + +Attribution methods explain deep models by generating saliency maps that highlight pixels critical for predictions. In our work, we benchmark eight prevalent attribution methods: Saliency [45], InputGradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5], and Guided Grad-CAM [42]. More details regarding these methods are recorded + +in Appendix B. In the following, we formalize the key concepts in attribution evaluation, deriving a general assessment paradigm in XAI. + +To begin with, let $\mathcal{D} = \{(\mathbf{x}_i,y_i)\}_{i = 1}^n$ be an image classification dataset with $n$ samples and $C$ categories, where $y_{i}$ is the label of RGB image $\mathbf{x}_i\in \mathbb{R}^{3\times H\times W}$ . Considering a deep network $\hat{y} = f_{\Theta}(\mathbf{x})$ that predicts $\hat{y}$ for input $\mathbf{x}$ , an attribution method aims to generate a heatmap $H\in \mathbb{R}^{3\times H\times W}$ so that each input pixel is assigned to a saliency score to the network's prediction. Such a procedure can be formulated as the following form: + +$$ +H = \mathsf {A t t r} _ {\Gamma} (f _ {\Theta} | \mathbf {x}, \hat {y}), \tag {1} +$$ + +where $\mathsf{Attr}_{\varGamma}$ denotes a certain attribution method parameterized by $\varGamma$ + +To evaluate the faithfulness of different attribution methods, we adopt two most common strategies: with progressively removing the most or least salient pixels of each image in the query dataset, we measure the performance degradation of the target model. The two strategies are usually referred to as the Most Relevant First (MoRF) and the Least Relevant First (LeRF) evaluation protocols. Specifically, we construct a top- $t$ removed query dataset $\tilde{\mathcal{D}}^{(t)}$ by removing the top- $t$ most or least salient pixels of the original image $\mathbf{x}_i$ and derive the modified image $\tilde{\mathbf{x}}_i^{(t)}$ as follows: + +$$ +\tilde {\mathbf {x}} _ {i} ^ {(t)} = \operatorname {R e m o v e} \left(\mathbf {x} _ {i}, H _ {i} | t\right). \tag {2} +$$ + +Here, the operation Remove specifies how features are eliminated, representing feature ablation in this paper. + +Effective attribution methods can accurately pinpoint significant pixels, and the elimination of these pixels would result in a swift deterioration in the model's performance. Thus, the evaluation of attribution methods can be facilitated by quantifying the reduction in model performance, which is defined by: + +$$ +\varDelta = \frac {1}{n} \left(\sum_ {i = 1} ^ {n} p \left(\hat {y} _ {i} = y _ {i} | \mathbf {x} _ {i}\right) - \sum_ {i = 1} ^ {n} p \left(\tilde {y} _ {i} ^ {(t)} = y _ {i} | \tilde {\mathbf {x}} _ {i} ^ {(t)}\right)\right). \tag {3} +$$ + +In this work, we focus on faithfulness to acquire a broadly applicable standard among multiple experimental settings. Ultimately, attribution methods are ranked according to this metric to generate a ranking list. The schematic diagram of the attribution and evaluation is depicted in Figure 1 (b). + +# 4 Meta-Rank Settings and Benchmark + +To acquire a comprehensive comprehension of the experimental configurations employed in previous investigations, we present a summary of the datasets and models utilized in these works, as depicted in Table 1. Notably, despite the limited choice of datasets and models, there remains considerable variation in the evaluations conducted thus far. The issue of evaluation consistency has largely been overlooked in prior research endeavors. To bridge this gap, in this section, we first establish standardized settings and subsequently provide a comprehensive outline of the Meta-Rank evaluation benchmark. The workflow of MetaRank is illustrated in Figure 1. + +Table 1: Comparison of experimental configurations employed in previous studies. The lack of consistency in these settings may lead to disparate observations. The term "Cons." refers to whether the study addresses the issue of consistency. Only a limited number of works have explored the issue, but their scope is limited to specific datasets, models, or both. + +
MethodsVenueDatasetsModelsCons.
[45]ICLR 2014ImageNetConvNetX
[44]ICML 2016Tiny ImageNet, DNA sequenceVGG-16, 1 CNNX
[48]ICML 2017ImageNetGoogleNetX
[47]ICLR Workshop 2015ImageNet, CIFAR-10/1003 CNNsX
[43]ICML 2017MNIST, DNA sequence2 CNNsX
[53]ECCV 2014ImageNet, Caltech, PASCAL VOC2 CNNsX
[5]PLOS ONE 2015ImageNet, PASCAL VOC, MINIST1 DNN, 1 CNNX
[42]ICCV 2017ImageNet, PASCAL VOCVGG-16, AlexNetX
[26]NeurIPS 2019ImageNet, Birdsnap, Food-101ResNet-50X
[11]ECCV Workshop 2020ImageNetResNet-50✓
[3]FUZZ-IEEE 2022ImageNet, Dogs vs. Cats, MAME, MIT67AlexNet, VGG-16, ResNet-18X
[40]ICML 2022CIFAR-10, Food-101ResNet-18, ResNet-50✓
[10]ICLR Tiny Paper 2023ImageNetResNet-50X
[8]CVPR 2023ImageNetVGG-16, ResNet-34✓
[24]JMLR 2023ImageNetResNet-18X
+ +# 4.1 Standardized Settings + +The main idea of the proposed Meta-Rank is evaluating attribution explanations on a diverse array of datasets, models, and evaluation protocols. + +Datasets. To streamline standardized evaluations of attribution methods, we meticulously select four commonly used datasets for benchmarking: NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]. NWPU-RESISC45 is a medium-scale dataset for remote sensing image classification, containing 45 categories of remote sensing images covering various scenes such as streets, airports, factories, forests, grasslands, and deserts. Food-101 is a medium-scale dataset for fine-grained food image classification, compassing a total of 101 different food categories. ImageNet is a large-scale dataset employed for image classification, consisting of 1,000 different categories. Places-365 is a large-scale dataset designed for the purposes of scene recognition, covering a wide range of 365 diverse scene categories. These datasets allow us to examine each attribution method under various domains, providing us with a more thorough view of the current literature. Further details are provided in Appendix C. + +Models. The selection of models presents a prominent source of inconsistency in previous research. In order to comprehensively validate the effectiveness of the attribution techniques across various models, we adopt six widely employed convolutional neural networks (CNNs) to evaluate existing attribution methods. These CNNs include ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]. It should be noted that most attribution methods are specifically tailored for CNNs and cannot be directly applied to vision transformers (ViTs) [13,33,52]. As such, for maintaining an equitable evaluation across attribution methods with a consistent set of evaluation settings, our benchmark does not take ViTs into consideration. Additional details regarding these used models can be found in Appendix C. + +Evaluation protocols. Recent advances have discovered that different evaluation protocols, MoRF and LeRF [41], may produce conflicting rankings of + +![](images/3416dfb6a9bd15b2155427cc2f6a3397ad242512cb4542cdb4677058c420d238.jpg) +(a) + +![](images/99b684f16776c970dd34ed4c862f68e0ffbfb9a1bed356ef8ad69e3ffdd7a191.jpg) +(b) +Fig. 1: Meta-Rank benchmark. It is mainly divided into two stages: Test Case Generation and Meta-Rank (a). Test Case Generation: multiple factors (i.e., datasets, models and evaluation protocols) are combined to generate $\tau$ different cases. MetaRank: (1) Case Execution. All $m$ competitors (i.e., attribution methods) are applied to these cases, resulting in a collection of rankings. The details of attribution evaluation on an individual case are provided in (b). (2) Ranking Fusion. All rankings $\{R_1, R_2, \ldots, R_\tau\}$ are subsequently fed into this module. The comparison of two competitors is transformed into the differences in their rankings across all cases, then integrated and converted into the discrepancy in Meta-Ranks. (3) Leaderboard. Ultimately, a unified leaderboard is obtained based on the Meta-Ranks of the competitors. “ $T$ ” is the test case, “ $C$ ” is the competitor, and “ $\kappa$ ” is the Meta-Rank value. + +attribute methods [40]. To provide a comprehensive view of the attribution methods, the proposed Meta-Rank considers both protocols for evaluation. + +Test case settings. To gain a generalized evaluation for attributions, our analytical experiments evaluate existing attribution methods on a wide array of experiment settings, necessitating the initial preparation of baselines. We first train ResNet-18, Inception-v4 and VGG-19 on NWPU-RESISC45 and Food101. Subsequently, we exclusively train EfficientNet_b2 and DenseNet-121 on the NWPU-RESISC45 dataset. These baselines are implemented with the PyTorch [36] framework. Adam [31] is used to optimize the models, with the learning rate of $1 \times 10^{-4}$ and the weight decay rate of $5 \times 10^{-4}$ . The cosine annealing is adopted as the learning rate decay schedule. Furthermore, we also incorporate pre-trained ResNet-18, ResNet-50 sourced from Places-365, and pre-trained ResNet-18, Inception-v4, VGG-19 from ImageNet-1k. The aforementioned baselines, combined with two evaluation protocols, are served as the test cases for the benchmark. Detailed information for each test case can be found in Appendix C. + +# 4.2 The Proposed Benchmark Metric + +We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. The main idea of the Meta-Rank metric is viewing each specific configuration of the model, the dataset and the evaluation protocol as one test case. Meta-Rank aggregates the ranking results (i.e., rank the rankings) from a diverse array of test cases, thereby mitigating biases associated with specific configurations. + +Formally, let the triplet $\langle D_i,M_j,P_k\rangle$ be one test case for evaluating the attribution algorithms, where $D$ , $M$ and $P$ denote the dataset, model and evaluation protocol respectively. The subscript $i$ , $j$ and $k$ denote one specific configuration from the candidate datasets, models and protocols as described in Section 4.1. Let $\mathcal{C} = \{C_1,C_2,\dots,C_m\}$ be the set of the competitors (i.e., attribution methods). In the proposed benchmark, all the competitors are evaluated on every generated test case, thus producing a ranking among the competitors. For the test case $\langle D_i,M_j,P_k\rangle$ , we use $R_{(i,j,k)} = (r_1^{(i,j,k)},r_2^{(i,j,k)},\dots,r_m^{(i,j,k)})$ to denote the rankings of all the competitors, where $r_m^{(i,j,k)}$ denotes the rank of the $m$ -th attribution method. After evaluating the attribution methods on all the test cases, the probability that $C_p$ outperforms $C_q$ is formally defined as follows: + +$$ +\mathrm {P} _ {q \prec p} = \frac {1}{| \Omega |} \sum_ {(i, j, k)} ^ {\Omega} \mathbf {1} \left(r _ {q} ^ {(i, j, k)} \prec r _ {p} ^ {(i, j, k)}\right), \tag {4} +$$ + +where $\mathbf{1}$ denotes the indicator function, and $r_q \prec r_p$ represents the ranking $r_q$ is outperformed by $r_p$ . $\Omega$ is the joint space of datasets, models and evaluation protocols. The probability $\mathsf{P}_{q \prec p}$ is then converted into the logistic odds, obtaining a more sensitive metric $\mathsf{Logit}_{q \prec p}$ that can better capture slight differences between $C_p$ and $C_q$ : + +$$ +\operatorname {L o g i t} _ {q \prec p} = \log \operatorname {o d d s} \left(C _ {p}, C _ {q}\right) = \log \left(\mathrm {P} _ {q \prec p} / \left(1 - \mathrm {P} _ {q \prec p}\right)\right). \tag {5} +$$ + +Based on the above derivation process, we can obtain a set of standardized results $\{\mathsf{Logit}_{q\prec p}\}$ that quantify the performance differences between any two competitors on all test cases. Then, we transform the performance discrepancy into the difference on the Meta-Rank metric: + +$$ +\left\{ \begin{array}{l l} \kappa_ {1} - \kappa_ {2} = \operatorname {L o g i t} _ {2 \prec 1}, & \text {C o m p e t i t o r} \# 1 v s. \# 2 \\ \kappa_ {1} - \kappa_ {3} = \operatorname {L o g i t} _ {3 \prec 1}, & \text {C o m p e t i t o r} \# 1 v s. \# 3 \\ \kappa_ {1} - \kappa_ {4} = \operatorname {L o g i t} _ {4 \prec 1}, & \text {C o m p e t i t o r} \# 1 v s. \# 4 \\ \dots & \dots \\ \kappa_ {2} - \kappa_ {3} = \operatorname {L o g i t} _ {3 \prec 2}, & \text {C o m p e t i t o r} \# 2 v s. \# 3 \\ \kappa_ {2} - \kappa_ {4} = \operatorname {L o g i t} _ {4 \prec 2}, & \text {C o m p e t i t o r} \# 2 v s. \# 4 \\ \dots & \dots \\ \kappa_ {m - 1} - \kappa_ {m} = \operatorname {L o g i t} _ {m \prec m - 1}. & \text {C o m p e t i t o r} \# \mathrm {m} - 1 v s. \# \mathrm {m} \end{array} \right. \tag {6} +$$ + +Here, $\kappa_{m}$ represents the Meta-Rank value of $C_m$ . After solving the system of equations in Equation 6, we can derive a score $\kappa$ for each competitor, serving as an identifier for its comprehensive performance. Through the arrangement of competitors in descending order of their $\kappa$ values, a unified leaderboard can finally be constructed. + +![](images/590109a90a800eadccecc5eab54d2de43cdb215560376459c2a5d0785e63ea10.jpg) +(a) Food-ResNet18(M) + +![](images/68874e528fe429141a2c9c7d61a04d0e54095d1c8a4bfbdf4d86aaaa0d3f0f9b.jpg) + +![](images/a85a1efca9bcc4a1b86fc44273fd5ea8031d65eaece98a1b31e4e8727eda294f.jpg) +(c) IN-ResNet18(M) + +![](images/214d3f34eb9f25c66f95e7cb360c87041b05e175b3f752a62413d9a9839a81d5.jpg) +(d) IN-ResNet18(L) + +![](images/4a9275ddce357e16c3457891d944f9c20cb26f0a843591270f51633f2dd0b19d.jpg) + +![](images/3a29a1d4cf263775abb7900491272e2c4a772ee39711b09367989e63e93d5825.jpg) +(b) Food-ResNet18(L) + +![](images/028712ff39f0b7f997da024ec668f6d80a9120b28b308d61a32a8b715c86096a.jpg) + +![](images/ac8eb546ddf9941173d9921dbcfc4f714a5438e200a3476e040b5f2b357a870a.jpg) + +(e) Food-Inceptionv4(M) +(i) Food-VGG19(M) +Fig.2: Evaluation results on 12 test cases. These cases consist of two datasets: Food-101 (a)(b)(e)(f)(i)(j), ImageNet-1k (c)(d)(g)(h)(k)(l), three models: ResNet-18 (a)(b)(c)(d), Inception-v4 (e)(f)(g)(h), VGG-19 (i)(j)(k)(l), and two protocols: MoRF(a)(c)(e)(g)(i)(k), LeRF(b)(d)(f)(h)(j)(l). "Baseline" represents the accuracy of the model when no features are ablated. +![](images/19a1bd99e420012140627505a9084fd65189e0c941ae00c8a5707bdba2da8e95.jpg) +Saliency GB -XG DeepLift -Deconv LRP -GuidedGradCam IG + +![](images/f5a9a17f674cfaa427b019217c77464c00fac5988a1c3f652a534868db0efecb.jpg) +(f) Food-Inceptionv4(L) +(j) Food-VGG19(L) + +![](images/e512d465e38b2b422dd7de213c84ae366157974d93239303ab2a48f7a3d7c5f3.jpg) +(g) IN-Inceptionv4(M) +(k) IN-VGG19(M) + +![](images/738ae351af61a1fb10f336a95de5f07bc5ad8345eeb7e78ee4d2f45f1eb6a829.jpg) +(h) IN-Inceptionv4(L) +(1) IN-VGG19(L) + +# 5 Main Results + +This section provides a detailed elaboration of extensive experiments on Meta-Rank, with a particular emphasis on the following pivotal questions. + +- RQ1: Are attribution methods consistently effective across diverse settings (i.e., datasets, models, evaluation protocols) and distinct checkpoints? (Section 5.1) +- RQ2: Does the state-of-the-art evaluation method, ROAD, yield consistent rankings across different test cases? (Section 5.2) +- RQ3: According to our proposed Meta-Rank, which attribution methods exhibit superior performance across a wide range of cases? (Section 5.3) +- RQ4: Can Meta-Rank conduct the evaluation of substantial methods within a feasible timeframe? (Section 5.4) + +# 5.1 RQ1: Consistency Investigation + +In this section, we investigate the consistency of performance rankings for existing attribution methods across diverse evaluation settings. Specifically, we focus + +![](images/ac78e231fdee73c10604645bc441a62aa2af36108ef3d171d2427efdb13aa783.jpg) +(a) MoRF + +![](images/102d25f1a7c33b50dd58b5071d0f98620881792a79adb775e065fa8b37880800.jpg) +(b) LeRF +Fig. 3: Spearman correlation among nine test cases (NWPU, Food, and ImageNet datasets paired with ResNet-18, Inception-v4, and VGG-19 models) in MoRF (a) and LeRF (b), and between MoRF and LeRF on the same nine cases (c). The labels along the horizontal arrow are equivalent to those along the vertical arrow. Here, we define a correlation score higher than 0.8 as strongly correlated (■), between 0.6 and 0.8 as moderately correlated (▲), and lower than 0.6 as weakly correlated (▼). + +![](images/73436c5a4764c326cac077aa5ac60d5787c36f3afe5c6883b0d54b97f6b1df25.jpg) +(c) MoRF v.s. LeRF + +on the model-, dataset-, protocol- and checkpoint-level ranking consistencies by varying these factors in the evaluation. To this end, we assess the aforementioned eight attribution methods on all the generated test cases, i.e., with six models, four datasets and two ablation protocols (MoRF and LeRF). This yields a total of 26 groups of results, each corresponding to one test case. Some experimental results are presented in Figure 2. For more details of the experimental results, please refer to Appendix D. + +Model-level Consistency. It is evident that the attribution methods exhibit varying performance across different models. To elaborate: (1) LRP has the worst performance on the "ImageNet-Inceptionv4(M)" case but significantly better results on the "ImageNet-ResNet18(M)" and "ImageNet-VGG19(M)" cases. (2) Integrated Gradients and Guided Grad-CAM exhibit comparable performance on the "Food-ResNet18(M)" case, but show great disparity on the "Food-VGG19(M)" case. For a more clear measurement of the inconsistency, we calculate the Spearman Rank correlations among several cases and illustrate them in Figure 3. Here, we extract area under the curve (AUC) [4] of all attribution methods (detailed in Appendix D) to generate a ranking list for each case. In the MoRF protocol, removing the most critical pixels identified by the attribution method causes a substantial drop in model performance. Thus, a lower AUC value indicates a superior method. And the LeRF protocol follows the inverse principle. The results suggest a pronounced weakness in the consistency at the model level, with a strong correlation found in only 2 cases (i.e., the correlation between ResNet-18 and Inception-v4 on the Food-101 dataset in both MoRF and LeRF protocols) out of 18 cases. + +Dataset-level Consistency. Analogously, attribution methods have shown noticeable inconsistency in performance on different datasets. For instance, Guided Backpropagation demonstrates superior performance on "ImageNet-VGG19(M)" yet ranks third-worst on "Food-VGG19(M)". In addition, according to the results of Spearman analysis, only $\frac{5}{18}$ cases show strong correlations at the dataset level. + +Table 2: AUC rankings of the eight methods at different checkpoints along the same trajectory. Here, ResNet-18 trained on NWPU-RESISC45 is taken as an illustrative example. + +
Epoch-50Epoch-100Epoch-150Epoch-200
MoRF
LeRF
SaliencyIXGDeconvGuidedGradCam
GBDeepLiftLRPIG
+ +Table 3: Comparison of Spearman results among four cases between ROAD and feature ablation. $a \mid b$ represents a pair of Spearman results, where $a$ denotes using feature ablation (in Section 5.1) and $b$ denotes using noisy linear imputation. The settings of color blocks in this table are the same as those in Figure 3. + +
NWPU-ResNet18Food-Inceptionv4
MoRF (#1)LeRF (#2)MoRF (#3)LeRF (#4)
#1| || || || |||
#2| || || || |||
#3| || || || |||
#4| || || || |||
+ +Protocol-level Consistency. The evaluation results indicate that significant inconsistencies between the MoRF and LeRF protocols still exist. A detailed comparison between "Food-Inceptionv4(M)" and "Food-Inceptionv4(L)" reveals marked discrepancies in Saliency method under MoRF versus LeRF. Besides, the Spearman correlation between MoRF and LeRF is computed for nine test cases (in Figure 3 (c)). It can be easily seen that the correlation of all cases at the protocol level is either moderately correlated or weakly correlated, with no strong correlation observed. + +Checkpoint-level Consistency. We also consider the consistency of attribution methods at different training stages of the model. To illustrate, using ResNet-18 trained on NWPU-RESISC45 as a case study, we extract distinct checkpoints from the same training trajectory and employ two evaluation protocols for attribution analysis. The results of the AUC rankings are recorded in Table 2. It can be easily observed that the rankings of attribution methods demonstrate high consistency across different checkpoints. In MoRF, the outcomes of Epoch-200 differ by only one distinction from those at both Epoch-100 and Epoch-150, while exhibiting slightly greater disparities when compared to those at Epoch-50. Similarly for LeRF, the rankings at Epoch-200 also exhibit a single deviation from those at Epoch-100 and Epoch-150, respectively, and show two distinctions in comparison to those at Epoch-50. These findings suggest a notable consistency in the ranking of attribution methods across different checkpoints. Moreover, the rankings tend to stabilize rapidly as the model converges. More detailed results under other settings can be found in Appendix D. + +# 5.2 RQ2: Necessity of Meta-Rank + +To further substantiate the indispensability of the Meta-Rank benchmark, we employ the state-of-the-art attribution evaluation method known as ROAD for assessing the attribution methods across various cases. Starting from information theoretic analyses, ROAD identifies the class information leakage problem + +Table 4: Leaderboard of the eight attribution methods. The larger Meta-Rank indicates the stronger ability of attribution method. The top three methods exhibit more positive performance, while the remaining five methods perform negatively. + +
MethodsMeta-Rank Score#Rank
InputGradient2.32001
Integrated Gradients2.07362
DeepLift1.07803
Saliency-0.04834
LRP-0.22345
Guided Backpropagation-0.25856
Guided Grad-CAM-0.96797
Deconvolution-3.97348
+ +![](images/1e282c6aecdfde2b2ad839227b11212ca6466ad488dbbedb9e8978ce71fb8b28.jpg) +Fig. 4: Time consumption of the "Ranking Fusion" module in Meta-Rank. 2, 4, 8, 16, 32, and 64 are the number of evaluated attribution methods. The gray bar represents the relative error. + +dictated by the ablation mask and finally derives the noisy linear imputation for minimally revealing imputation, which is empirically demonstrated to produce consistent rankings of attribution methods between evaluations with MoRF and LeRF. However, it overlooks consistency at both the model and dataset levels. Thus, we randomly select four cases, namely NWPU-ResNet18(M), NWPU-ResNet18(L), Food-Inceptionv4(M), and Food-Inceptionv4(L), to assess the eight aforementioned attribution methods using ROAD. The evaluation metric used in this study aligns with those in RQ1. Table 3 presents the Spearman correlations among the evaluation results of the four cases. Further information regarding the accuracy decay curves is available in Appendix E. + +It can be easily seen that, even when ROAD is employed as a substitute for feature ablation, the correlation between the rankings of attribution methods on different cases remains inconsistent. Specifically, all correlations demonstrate either moderate or weak, with no instances of strong correlation. This indicates that relying solely on ROAD for evaluating attribution methods across distinct cases is inadequate. Consequently, there is a critical need for a cross-case benchmark to ensure robust ranking results. + +# 5.3 RQ3: Attribution Evaluation with Meta-Rank + +Following the evaluation scheme outlined in the Meta-Rank benchmark (described in Section 4), we have carried out a unified measurement of the eight attribution methods. The consistency validation experiments in Section 5.1 generate evaluation results for attribution methods across distinct configurations, encompassing four datasets, six models, and two protocols, resulting in 26 sets of data. We then transform these results into AUC scores, which serve as the performance metric for the attribution methods. + +Meta-Rank generates a comprehensive value for each attribution method. These uniquely determined scores enable us to establish a unified ranking list for all the methods. The finalized leaderboard for the eight attribution methods + +is presented in Table 4. The results reveal InputGradient as the top-performing approach, while Deconvolution exhibits the poorest performance. (1) In particular, InputGradient achieves the highest score among the eight methods. By element-wise multiplying the gradients with the inputs, InputGradient produces more stable attributions in the face of perturbations, which contributes to its superior performance across multiple test cases. (2) Furthermore, the second-ranked method, Integrated Gradients, also shows great potential, with its score closely approaching that of the top-ranked method. (3) In contrast, Deconvolution obtains the lowest score, likely attributable to the hierarchical structure of the network accumulating noise during backpropagation. It is worth noting that a Meta-Rank score can assume both positive and negative values. Positive values signify comparatively superior attribution performance, whereas negative values indicate inferior performance. + +The experimental findings demonstrate that Meta-Rank furnishes a comprehensive benchmark for attribution evaluation across diverse test cases, while effectively ameliorating the issue of inconsistent performance across heterogeneous cases. It computes the discrepancies between methods and converts them into the differences in capacity to derive the Meta-Rank. This result is unique, consistent, and aligned with practical application needs. + +# 5.4 RQ4: Efficiency of Meta-Rank + +To measure the time consumption, we conduct experiments on the CPU device (Intel(R) Xeon(R) Platinum 8260L CPU @ 2.40GHz) using different numbers of methods. Here, we focus on the additional "Ranking Fusion" module, as the "Case Execution" module has been generically applied in all evaluations. In particular, we compare the time expenses of Meta-Rank during the calculation of the 2, 4, 8, 16, 32, and 64 methods. It is important to highlight that the first three test groups (i.e., 2, 4, and 8 methods) rely on real data gathered from previous experiments, whereas the last three test groups (i.e., 16, 32, and 64 methods) employ simulated data randomly generated. The data here refers to a set of accuracies (i.e., $\{acc_i\}$ in Figure 1) obtained on each test case according to every attribution method. + +The results in Figure 4 indicate that as the number of attribution methods increases, the computation time also increases. However, since the number of attribution methods is finite, the computational time remains within an acceptable range. For example, when 64 methods are involved in the Meta-Rank evaluation, the time required to generate the leaderboard is approximately 0.12s, which completely meets expectations. Additionally, using GPUs for Meta-Rank calculations can further enhance the computation speed. This efficiency is beneficial for the promotion and application of Meta-Rank. + +# 6 Discussion + +Choice of faithfulness metric. Although there are many dimensions for attribution evaluation in previous work [24], faithfulness remains the most solid + +and general criteria for benchmarking attribution methods. In the context of multiple experimental settings, as such, we opt to use it as the base metric $\Delta$ in our framework to facilitate subsequent Meta-Rank assessment and then enable broader evaluation on novel tasks. Meanwhile, Meta-Rank can also accommodate evaluations on other dimensions by enabling the derivation of a ranking list along each respective evaluation metric. + +Missingness bias in feature ablation. Previous work has discussed the missingness bias problem caused by the removal of pixels through masking values in CNNs [29]. Hooker et al. proposed ROAR to address this issue, but it inadvertently introduced the problem of information leakage. Subsequently, Rong et al. developed ROAD to mitigate this leakage, but it failed to fully remove the pixels (explained in Appendix E). We recognize that completely eliminating this bias poses substantial difficulties, since the root of missingness bias lies in the operation of removing pixels within CNNs. Furthermore, the impact of bias exhibits variability across individual cases, also posing additional challenges in quantification and measurement. However, we believe that with a more diverse array of test cases, the missingness bias is more neutralized. Future research will further explore solutions for eliminating this bias. + +Scalability of Meta-Rank. When introducing new attribution methods to the ranking, we initially conduct tests on several existing cases. The time complexity of evaluating a method on an individual case is linear, which has been demonstrated in previous studies [37, 40]. Subsequently, the obtained results, in conjunction with historical data, are fed into the "Ranking Fusion" module to generate an updated ranking list. The time overhead associated with this process is also optimistic, as illustrated in Section 5.4. This flexibility allows for the inclusion of any attribution method into the leaderboard at any time. Future endeavors will focus on expanding the corpus of test cases and pursuing further computational efficiency by leveraging parallel computing paradigms. + +# 7 Conclusion + +To address the lack of systematic research in feature attribution, we propose Meta-Rank, a novel standardized benchmark for robust and unified evaluation of methods. Meta-Rank aggregates relative comparisons between all method pairs across diverse settings, mitigating inconsistencies. Through extensive experiments, we reveal three critical insights: diverse evaluation settings yield divergent performance rankings, assessment results remain consistent across distinct checkpoints along the same training trajectory, and previous attempts at consistent evaluation exhibit limitations on heterogeneous models and datasets. Meta-Rank provides a leaderboard that highlights the broader applicability of InputGradient and Integrated Gradients. Remarkably, Meta-Rank exhibits exceptional computational efficiency, enabling ranking computations to be completed within milliseconds even for a substantial number of methods. In the future, we hope the proposed benchmark will facilitate the development of faithful and consistent explanations, fostering transparency and accountability. + +# Acknowledgements + +This work is supported by National Natural Science Foundation of China (62106220), Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies, Ningbo 2025 Science and Technology Innovation Major Project (No. 2022Z072), and the Fundamental Research Funds for the Central Universities (226-2024-00058). + +# References + +1. Adadi, A., Berrada, M.: Peeking inside the black-box: a survey on explainable artificial intelligence (xai). IEEE access 6, 52138-52160 (2018) +2. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104 (2017) +3. Arias-Duart, A., Parés, F., Garcia-Gasulla, D., Giménez-Ábalos, V.: Focus! rating xai methods and finding biases. In: 2022 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). pp. 1-8. IEEE (2022) +4. Atanasova, P., Simonsen, J.G., Lioma, C., Augenstein, I.: A diagnostic study of explainability techniques for text classification. arXiv preprint arXiv:2009.13295 (2020) +5. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015) +6. Bastings, J., Aziz, W., Titov, I.: Interpretable neural predictions with differentiable binary variables. arXiv preprint arXiv:1905.08160 (2019) +7. Bhatt, U., Weller, A., Moura, J.M.: Evaluating and aggregating feature-based model explanations. arXiv preprint arXiv:2005.00631 (2020) +8. Binder, A., Weber, L., Lapuschkin, S., Montavon, G., Müller, K.R., Samek, W.: Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16143-16152 (2023) +9. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014) +10. Brocki, L., Chung, N.C.: Evaluation of interpretability methods and perturbation artifacts in deep neural networks. arXiv preprint arXiv:2203.02928 (2022) +1. Brunke, L., Agrawal, P., George, N.: Evaluating input perturbation methods for interpreting cnns and saliency map comparison. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 120-134. Springer (2020) +2. Chattopadhy, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Gradcam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018) +3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 782-791 (2021) + +14. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017) +15. Colin, J., Fel, T., Cadène, R., Serre, T.: What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems 35, 2832-2845 (2022) +16. Covert, I., Lundberg, S.M., Lee, S.I.: Understanding global feature contributions with additive importance measures. Advances in Neural Information Processing Systems 33, 17212-17223 (2020) +17. Deng, H., Zou, N., Du, M., Chen, W., Feng, G., Yang, Z., Li, Z., Zhang, Q.: Understanding and unifying fourteen attribution methods with taylor interactions. arXiv preprint arXiv:2303.01506 (2023) +18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +19. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017) +20. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020) +21. Goyal, Y., Wu, Z., Ernst, J., Batra, D., Parikh, D., Lee, S.: Counterfactual visual explanations. In: International Conference on Machine Learning. pp. 2376-2384. PMLR (2019) +22. Hartley, T., Sidorov, K., Willis, C., Marshall, D.: Explaining failure: Investigation of surprise and expectation in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 12-13 (2020) +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +24. Hedström, A., Weber, L., Krakowczyk, D., Bareeva, D., Motzkus, F., Samek, W., Lapuschkin, S., Höhne, M.M.C.: Quantus: An explainable ai toolkit for responsible evaluation of neural network explanations and beyond. Journal of Machine Learning Research 24(34), 1-11 (2023) +25. Hesse, R., Schaub-Meyer, S., Roth, S.: Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3981-3991 (2023) +26. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems 32 (2019) +27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017) +28. Ismail, A.A., Gunady, M., Corrada Bravo, H., Feizi, S.: Benchmarking deep learning interpretability in time series predictions. Advances in neural information processing systems 33, 6441-6452 (2020) +29. Jain, S., Salman, H., Wong, E., Zhang, P., Vineet, V., Vermprala, S., Madry, A.: Missingness bias in model debugging. arXiv preprint arXiv:2204.08945 (2022) +30. Khakzar, A., Khorsandi, P., Nobahari, R., Navab, N.: Do explanations explain? model knows best. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10244-10253 (2022) + +31. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +32. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602 (2022) +33. Leem, S., Seo, H.: Attention guided cam: Visual explanations of vision transformer guided by self-attention. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2956-2964 (2024) +34. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017) +35. Nguyen, G., Kim, D., Nguyen, A.: The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems 34, 26422-26436 (2021) +36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32, 8026-8037 (2019) +37. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018) +38. Rao, S., Böhle, M., Schiele, B.: Better understanding differences in attribution methods via systematic evaluations. arXiv preprint arXiv:2303.11884 (2023) +39. Ribeiro, M.T., Singh, S., Guestrin, C.: "why should i trust you?" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016) +40. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A consistent and efficient evaluation strategy for attribution methods. In: International Conference on Machine Learning. pp. 18770-18795. PMLR (2022) +41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.R.: Evaluating the visualization of what a deep neural network has learned. IEEE transactions on neural networks and learning systems 28(11), 2660-2673 (2016) +42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +43. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017) +44. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not just a black box: Learning important features through propagating activation differences. arXiv preprint arXiv:1605.01713 (2016) +45. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: visualising image classification models and saliency maps. In: Proceedings of the International Conference on Learning Representations (ICLR). ICLR (2014) +46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +47. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014) +48. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017) + +49. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017) +50. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019) +51. Wang, Y., Wang, X.: A unified study of machine learning explanation evaluation metrics. arXiv preprint arXiv:2203.14265 (2022) +52. Xue, M., Huang, Q., Zhang, H., Cheng, L., Song, J., Wu, M., Song, M.: Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In: 33rd International Joint Conference on Artificial Intelligence, IJCAI-24. International Joint Conferences on Artificial Intelligence Organization (2024) +53. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 818-833. Springer (2014) +54. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-down neural attention by excitation backprop. International Journal of Computer Vision 126(10), 1084-1102 (2018) +55. Zhang, Y., Khakzar, A., Li, Y., Farshad, A., Kim, S.T., Navab, N.: Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems 34, 20040-20051 (2021) +56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017) +57. Zhou, Y., Booth, S., Ribeiro, M.T., Shah, J.: Do feature attribution methods correctly attribute features? In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9623-9633 (2022) +58. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. arXiv preprint arXiv:1702.04595 (2017) \ No newline at end of file diff --git a/2024/On the Evaluation Consistency of Attribution-based Explanations/images.zip b/2024/On the Evaluation Consistency of Attribution-based Explanations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..92bf2e69be28e02477996f01ff43bb4e004e799a --- /dev/null +++ b/2024/On the Evaluation Consistency of Attribution-based Explanations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c98dd737d66fe3f7d2e4fc7d2346017203cd99680a233f9d6ae28d1430785b06 +size 385806 diff --git a/2024/On the Evaluation Consistency of Attribution-based Explanations/layout.json b/2024/On the Evaluation Consistency of Attribution-based Explanations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17d29ffac377fda1bd6bc564ee1f244c67829b75 --- /dev/null +++ b/2024/On the Evaluation Consistency of Attribution-based Explanations/layout.json @@ -0,0 +1,9697 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 189, + 112, + 425, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 112, + 425, + 148 + ], + "spans": [ + { + "bbox": [ + 189, + 112, + 425, + 148 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": "Jiarui Duan" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Haoling Li" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Haofei Zhang" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Hao Jiang" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Mengqi Xue" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Li Sun" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Mingli Song" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": ", and Jie Song" + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{2,3\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 167, + 201, + 445, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 201, + 445, + 213 + ], + "spans": [ + { + "bbox": [ + 167, + 201, + 445, + 213 + ], + "type": "text", + "content": "1 College of Computer Science and Technology, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 213, + 415, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 213, + 415, + 224 + ], + "spans": [ + { + "bbox": [ + 198, + 213, + 415, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 213, + 415, + 224 + ], + "type": "text", + "content": " School of Software Technology, Zhejiang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 224, + 478, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 478, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 478, + 244 + ], + "type": "text", + "content": "3 Key Laboratory of Visual Perception (Zhejiang University), Ministry of Education and Microsoft" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 272, + 245, + 342, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 245, + 342, + 256 + ], + "spans": [ + { + "bbox": [ + 272, + 245, + 342, + 256 + ], + "type": "text", + "content": "4 Alibaba Group" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 256, + 365, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 256, + 365, + 267 + ], + "spans": [ + { + "bbox": [ + 249, + 256, + 365, + 267 + ], + "type": "text", + "content": "5 Hangzhou City University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 206, + 267, + 406, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 267, + 406, + 279 + ], + "spans": [ + { + "bbox": [ + 206, + 267, + 406, + 279 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 206, + 267, + 406, + 279 + ], + "type": "text", + "content": " Ningbo Innovation Center, Zhejiang University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 157, + 279, + 457, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 279, + 457, + 290 + ], + "spans": [ + { + "bbox": [ + 157, + 279, + 457, + 290 + ], + "type": "text", + "content": "{jerryduan,hollylee,haofeizhang,lsun,brooksong,sjie}@zju.edu.cn;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 206, + 290, + 406, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 290, + 406, + 300 + ], + "spans": [ + { + "bbox": [ + 206, + 290, + 406, + 300 + ], + "type": "text", + "content": "aoshu.jh@alibaba-inc.com; mqxue@hzcu.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 323, + 453, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 323, + 453, + 564 + ], + "spans": [ + { + "bbox": [ + 159, + 323, + 453, + 564 + ], + "type": "text", + "content": "Abstract. Attribution-based explanations are garnering increasing attention recently and have emerged as the predominant approach towards eXplanable Artificial Intelligence (XAI). However, the absence of consistent configurations and systematic investigations in prior literature impedes comprehensive evaluations of existing methodologies. In this work, we introduce Meta-Rank, an open platform for benchmarking attribution methods in the image domain. Presently, Meta-Rank assesses eight exemplary attribution methods using six renowned model architectures on four diverse datasets, employing both the Most Relevant First (MoRF) and Least Relevant First (LeRF) evaluation protocols. Through extensive experimentation, our benchmark reveals three insights in attribution evaluation endeavors: 1) evaluating attribution methods under disparate settings can yield divergent performance rankings; 2) although inconsistent across numerous cases, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts at consistent evaluation fare no better than baselines when extended to more heterogeneous models and datasets. Our findings underscore the necessity for future research in this domain to conduct rigorous evaluations encompassing a broader range of models and datasets, and to reassess the assumptions underlying the empirical success of different attribution methods. Our code is publicly available at https://github.com/TreeThree-R/Meta-Rank." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 574, + 445, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 574, + 445, + 585 + ], + "spans": [ + { + "bbox": [ + 160, + 574, + 445, + 585 + ], + "type": "text", + "content": "Keywords: Attribution evaluation " + }, + { + "bbox": [ + 160, + 574, + 445, + 585 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 574, + 445, + 585 + ], + "type": "text", + "content": " Disparate settings " + }, + { + "bbox": [ + 160, + 574, + 445, + 585 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 574, + 445, + 585 + ], + "type": "text", + "content": " Consistency" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "Explainable Artificial Intelligence (XAI) [1] has emerged as a prominent research field within computer vision, with a multitude of approaches proposed for ex-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 236, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "plaining and interpreting the internal mechanisms of diverse neural network architectures across various application domains. Among these approaches, attribution methods have gained significant traction in the realm of deep learning. By generating an explanation map in the form of a heatmap (also referred to as a relevance map or a saliency map), these methods attribute the prediction to specific regions within the input image, thereby facilitating interpretation. Over the years, researchers have devised a large number of attribution methods from diverse perspectives, including gradient-based methods [5, 43, 44, 47, 48], perturbation-based methods [34, 37, 39, 53], and CAM-based methods [12, 42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 225, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 451 + ], + "type": "text", + "content": "Nevertheless, given the absence of ground truth regarding the inner workings of deep neural networks, evaluating attributions directly poses a formidable challenge. Presently, various studies have endeavored to assess the reliability of attribution methods, which can be roughly categorized into two schools: expert-grounded methods [6, 15, 16, 32, 35, 54], and functional-grounded methods [4, 17, 26, 37, 40, 41, 51]. Expert-grounded evaluations capture how well the explanations imitate the human-annotated importance of the image regions. Such methods not only require human effort, but also suffer from the faithfulness1 issue, as deep neural networks could rely on non-intuitive, spurious cues that differ from human perception [20]. In contrast, functional-based approaches evaluate attribution methods based on some fundamental axioms. For instance, feature ablation [37], as a functional-grounded evaluation, systematically removes features in a predetermined order and observes the resulting changes in predictions. This approach enables more faithful assessments of the attribution explanations by comparing the model behavior with and without specific individual features [19, 21, 26, 37, 39, 58]. Unfortunately, the current literature often evaluates attributions under divergent settings, with different datasets, models and evaluation protocols [26, 40]. The efficacy of attribution methods in practice is often called into question due to restricted and inconsistent experimental conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 453, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 632 + ], + "type": "text", + "content": "The goal of this work is to gain a thorough understanding of the current state of attribution methods while setting the stage for critical problems to be worked on. To this end, we present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines. Our benchmark carefully examines eight widely-used attribution methods (Saliency [45], Input" + }, + { + "bbox": [ + 130, + 453, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 453, + 482, + 632 + ], + "type": "text", + "content": "Gradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5] and Guided Grad-CAM [42]), with a wide range of models (ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]), on various domain datasets (NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]), using two evaluation protocols (Most Relevant First and Least Relevant First). Specifically, we place a strong emphasis on comprehensive and consistent experimental settings that have been largely overlooked in previous works. Through extensive experiments on the exhaustive combinations of methods, datasets, models and evaluation protocols," + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "1 Faithfulness refers to the degree of alignment between the results interpreted by an attribution method and the decision of the model." + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "our benchmark reveals three main findings regarding previous attribution evaluation endeavors: 1) evaluating attribution methods under disparate experimental settings, such as different model architectures, datasets and evaluation protocols, can yield highly divergent performance rankings; 2) although highly inconsistent across datasets, models and evaluation protocols, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts towards more consistent evaluation, like Remove and Debias (ROAD) [40], fare no better than simple baselines when it is extended to more heterogeneous models and datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 366 + ], + "type": "text", + "content": "We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. Meta-Rank aggregates ranking results from a diverse array of settings, thereby mitigating biases associated with specific configurations. By employing this metric, we conduct systematic comparisons among existing attribution methods, and provide extensive experiment results and discussions on the status quo. Beyond extensive experiments, our benchmark is designed as a flexible framework that standardizes experimental settings and simplifies the integration of novel algorithmic implementations. We hope our work will not only facilitate reliable evaluations of attribution methods across a broader range of models and datasets, but also inspire the research community to delve further into the development of more consistent attribution methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 368, + 367, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 368, + 367, + 379 + ], + "spans": [ + { + "bbox": [ + 146, + 368, + 367, + 379 + ], + "type": "text", + "content": "In summary, we make the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 385, + 479, + 479 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 139, + 385, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 385, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 139, + 385, + 479, + 407 + ], + "type": "text", + "content": "- We present Meta-Rank, an open-sourced attribution benchmark featuring rigorous evaluations, comprehensive analyses as well as extensive baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 409, + 479, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 409, + 479, + 442 + ], + "spans": [ + { + "bbox": [ + 139, + 409, + 479, + 442 + ], + "type": "text", + "content": "- Through extensive experiments, we identify three pivotal insights from previous attribution evaluation efforts, which provide directions for future advancements in attribution research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 445, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 445, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 139, + 445, + 479, + 479 + ], + "type": "text", + "content": "- We propose an evaluation metric to facilitate a more comprehensive benchmark comparison across multiple settings, with which extensive experiments and discussions are provided on the status quo." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 498, + 237, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 237, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 237, + 510 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 556 + ], + "type": "text", + "content": "The development of attribution methods has prompted researchers to seek the best approach to assist models in downstream tasks. In this section, we delineate the progress of attribution evaluation chronologically." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 487, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 487, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 487, + 628 + ], + "type": "text", + "content": "Expert-grounded evaluations [6, 16, 35, 54, 55] rely on human judgment to evaluate attribution methods. For example, Zhang et al. [54] proposed the pointing-game method as a means of evaluating attention maps based on detecting whether the highest value point in the map hit the target category. However, such methods are inherently subjective and fail to guarantee the reliability of evaluation results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Functional-grounded evaluations [2,7,26,37,41] provide more objective evaluations of the model by comparing how it behaves when certain individual features are present or absent. Samek et al. [41] proposed two evaluation protocols," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 113, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 113, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 113, + 482, + 561 + ], + "type": "text", + "content": "namely Most Relevant First (MoRF) and Least Relevant First (LeRF), to observe the impact on the model's output. In MoRF, effective attribution methods should produce a greater reduction in model accuracy when removing features with higher attribution values, and vice versa. Similarly, [37] suggested evaluating attribution methods through the deletion and insertion of features by using their attribution values. These methods violate the assumption of a consistent distribution between training and test data, making it hard to determine if the decline in model performance solely results from feature deletion. Hooker et al. [26] further demonstrated that the performance degradation from feature deletion partly stems from such a distribution shift. To address this, they developed RemOve And Retrain (ROAR), which resolves inconsistent distributions by synchronously deleting the features of the training set and test set based on the attributions of each attribution method, and then retrained the model on the generated datasets. The effectiveness of the attribution method is measured by the retrained model's accuracy decline. Better methods identify more important features, whose removal causes sharper performance drops. ROAR has been widely adopted in evaluation studies [7, 22, 28]. However, due to its exorbitant expense, this type of evaluation method is unfeasible for scaling up. Rong et al. [40] discovered inconsistent MoRF and LeRF rankings from ROAR, which they attributed to information leakage based on information theory analysis. Thus, they proposed a noisy linear imputation method, named RemOve And Debias (ROAD), to effectively generate consistent results. However, as a pixel filling method, ROAD still suffers from several inherent defects, e.g., the features are replaced with substantial values instead of entirely dropped, which can hardly guarantee their neutral effects on the predictions. Deng et al. [17] noted the lack of unified theoretical frameworks for evaluating attribution methods. To resolve this limitation, they proposed a Taylor interaction system to unify fourteen attribution methods under a common mathematical formulation and suggested three principles for evaluation. However, applying this framework to new methods requires manual theoretical analysis, hampering both scalability and efficiency. Recently, attribution evaluation has been explored from more diverse perspectives [30, 38, 57], including comprehensive viewpoints [24, 25] as well as the consistency lens [8, 11, 26, 40]. Unfortunately, the absence of standardized experimental settings and systematic evaluations in previous literatures poses challenges for comparing existing methods universally. To overcome this limitation, we propose Meta-Rank, a novel evaluation benchmark, to provide a comprehensive evaluation of attribution methods across multiple settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 578, + 234, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 234, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 234, + 592 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Attribution methods explain deep models by generating saliency maps that highlight pixels critical for predictions. In our work, we benchmark eight prevalent attribution methods: Saliency [45], InputGradient [44], Integrated Gradients [48], Guided Backpropagation [47], DeepLift [43], Deconvolution [53], LRP [5], and Guided Grad-CAM [42]. More details regarding these methods are recorded" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "in Appendix B. In the following, we formalize the key concepts in attribution evaluation, deriving a general assessment paradigm in XAI." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "To begin with, let " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^n" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " be an image classification dataset with " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " samples and " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " categories, where " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " is the label of RGB image " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ". Considering a deep network " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\hat{y} = f_{\\Theta}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " that predicts " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " for input " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", an attribution method aims to generate a heatmap " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "H\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " so that each input pixel is assigned to a saliency score to the network's prediction. Such a procedure can be formulated as the following form:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 217, + 481, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 217, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 261, + 217, + 481, + 232 + ], + "type": "interline_equation", + "content": "H = \\mathsf {A t t r} _ {\\Gamma} (f _ {\\Theta} | \\mathbf {x}, \\hat {y}), \\tag {1}", + "image_path": "1881d1a9913ff5ca8bb7aaad537d1c2a12802f57322fc9ca96b70659096e54a9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 236, + 448, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 448, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 448, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 236, + 448, + 247 + ], + "type": "inline_equation", + "content": "\\mathsf{Attr}_{\\varGamma}" + }, + { + "bbox": [ + 131, + 236, + 448, + 247 + ], + "type": "text", + "content": " denotes a certain attribution method parameterized by " + }, + { + "bbox": [ + 131, + 236, + 448, + 247 + ], + "type": "inline_equation", + "content": "\\varGamma" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "text", + "content": "To evaluate the faithfulness of different attribution methods, we adopt two most common strategies: with progressively removing the most or least salient pixels of each image in the query dataset, we measure the performance degradation of the target model. The two strategies are usually referred to as the Most Relevant First (MoRF) and the Least Relevant First (LeRF) evaluation protocols. Specifically, we construct a top-" + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "text", + "content": " removed query dataset " + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}^{(t)}" + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "text", + "content": " by removing the top-" + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "text", + "content": " most or least salient pixels of the original image " + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "text", + "content": " and derive the modified image " + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i^{(t)}" + }, + { + "bbox": [ + 130, + 247, + 481, + 344 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 350, + 481, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 350, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 255, + 350, + 481, + 365 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {i} ^ {(t)} = \\operatorname {R e m o v e} \\left(\\mathbf {x} _ {i}, H _ {i} | t\\right). \\tag {2}", + "image_path": "6f2d1ba106246b584c0af1ffebf5d987ef297d0f744542378f434acc8f1e1b98.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 370, + 479, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 479, + 394 + ], + "type": "text", + "content": "Here, the operation Remove specifies how features are eliminated, representing feature ablation in this paper." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 395, + 480, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 480, + 442 + ], + "type": "text", + "content": "Effective attribution methods can accurately pinpoint significant pixels, and the elimination of these pixels would result in a swift deterioration in the model's performance. Thus, the evaluation of attribution methods can be facilitated by quantifying the reduction in model performance, which is defined by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 189, + 447, + 481, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 447, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 189, + 447, + 481, + 469 + ], + "type": "interline_equation", + "content": "\\varDelta = \\frac {1}{n} \\left(\\sum_ {i = 1} ^ {n} p \\left(\\hat {y} _ {i} = y _ {i} | \\mathbf {x} _ {i}\\right) - \\sum_ {i = 1} ^ {n} p \\left(\\tilde {y} _ {i} ^ {(t)} = y _ {i} | \\tilde {\\mathbf {x}} _ {i} ^ {(t)}\\right)\\right). \\tag {3}", + "image_path": "aa10695a25f16e20ebd1c5b18a851646854a19c459762061b640e5378d1d4a71.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 472, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 521 + ], + "type": "text", + "content": "In this work, we focus on faithfulness to acquire a broadly applicable standard among multiple experimental settings. Ultimately, attribution methods are ranked according to this metric to generate a ranking list. The schematic diagram of the attribution and evaluation is depicted in Figure 1 (b)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 536, + 371, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 371, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 371, + 550 + ], + "type": "text", + "content": "4 Meta-Rank Settings and Benchmark" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "To acquire a comprehensive comprehension of the experimental configurations employed in previous investigations, we present a summary of the datasets and models utilized in these works, as depicted in Table 1. Notably, despite the limited choice of datasets and models, there remains considerable variation in the evaluations conducted thus far. The issue of evaluation consistency has largely been overlooked in prior research endeavors. To bridge this gap, in this section, we first establish standardized settings and subsequently provide a comprehensive outline of the Meta-Rank evaluation benchmark. The workflow of MetaRank is illustrated in Figure 1." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 448, + 103 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 179, + 477, + 289 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Comparison of experimental configurations employed in previous studies. The lack of consistency in these settings may lead to disparate observations. The term \"Cons.\" refers to whether the study addresses the issue of consistency. Only a limited number of works have explored the issue, but their scope is limited to specific datasets, models, or both." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 179, + 477, + 289 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 477, + 289 + ], + "type": "table", + "html": "
MethodsVenueDatasetsModelsCons.
[45]ICLR 2014ImageNetConvNetX
[44]ICML 2016Tiny ImageNet, DNA sequenceVGG-16, 1 CNNX
[48]ICML 2017ImageNetGoogleNetX
[47]ICLR Workshop 2015ImageNet, CIFAR-10/1003 CNNsX
[43]ICML 2017MNIST, DNA sequence2 CNNsX
[53]ECCV 2014ImageNet, Caltech, PASCAL VOC2 CNNsX
[5]PLOS ONE 2015ImageNet, PASCAL VOC, MINIST1 DNN, 1 CNNX
[42]ICCV 2017ImageNet, PASCAL VOCVGG-16, AlexNetX
[26]NeurIPS 2019ImageNet, Birdsnap, Food-101ResNet-50X
[11]ECCV Workshop 2020ImageNetResNet-50✓
[3]FUZZ-IEEE 2022ImageNet, Dogs vs. Cats, MAME, MIT67AlexNet, VGG-16, ResNet-18X
[40]ICML 2022CIFAR-10, Food-101ResNet-18, ResNet-50✓
[10]ICLR Tiny Paper 2023ImageNetResNet-50X
[8]CVPR 2023ImageNetVGG-16, ResNet-34✓
[24]JMLR 2023ImageNetResNet-18X
", + "image_path": "0fc47b8c8201488df65994f3c8b750e081316ad81f48d2f6a6d21463ebf989ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 311, + 271, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 271, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 271, + 323 + ], + "type": "text", + "content": "4.1 Standardized Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 354 + ], + "type": "text", + "content": "The main idea of the proposed Meta-Rank is evaluating attribution explanations on a diverse array of datasets, models, and evaluation protocols." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 510 + ], + "type": "text", + "content": "Datasets. To streamline standardized evaluations of attribution methods, we meticulously select four commonly used datasets for benchmarking: NWPU-RESISC45 [14], Food-101 [9], ImageNet-1k [18], and Places-365 [56]. NWPU-RESISC45 is a medium-scale dataset for remote sensing image classification, containing 45 categories of remote sensing images covering various scenes such as streets, airports, factories, forests, grasslands, and deserts. Food-101 is a medium-scale dataset for fine-grained food image classification, compassing a total of 101 different food categories. ImageNet is a large-scale dataset employed for image classification, consisting of 1,000 different categories. Places-365 is a large-scale dataset designed for the purposes of scene recognition, covering a wide range of 365 diverse scene categories. These datasets allow us to examine each attribution method under various domains, providing us with a more thorough view of the current literature. Further details are provided in Appendix C." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": "Models. The selection of models presents a prominent source of inconsistency in previous research. In order to comprehensively validate the effectiveness of the attribution techniques across various models, we adopt six widely employed convolutional neural networks (CNNs) to evaluate existing attribution methods. These CNNs include ResNet-18 [23], ResNet-50 [23], Inception-v4 [49], VGG-19 [46], EfficientNet_b2 [50], and DenseNet-121 [27]. It should be noted that most attribution methods are specifically tailored for CNNs and cannot be directly applied to vision transformers (ViTs) [13,33,52]. As such, for maintaining an equitable evaluation across attribution methods with a consistent set of evaluation settings, our benchmark does not take ViTs into consideration. Additional details regarding these used models can be found in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Evaluation protocols. Recent advances have discovered that different evaluation protocols, MoRF and LeRF [41], may produce conflicting rankings of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 122, + 479, + 231 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 145, + 121 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 145, + 121 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 145, + 121 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 122, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 135, + 122, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 135, + 122, + 479, + 231 + ], + "type": "image", + "image_path": "3416dfb6a9bd15b2155427cc2f6a3397ad242512cb4542cdb4677058c420d238.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 236, + 479, + 325 + ], + "blocks": [ + { + "bbox": [ + 135, + 234, + 145, + 244 + ], + "lines": [ + { + "bbox": [ + 135, + 234, + 145, + 244 + ], + "spans": [ + { + "bbox": [ + 135, + 234, + 145, + 244 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 236, + 479, + 325 + ], + "lines": [ + { + "bbox": [ + 149, + 236, + 479, + 325 + ], + "spans": [ + { + "bbox": [ + 149, + 236, + 479, + 325 + ], + "type": "image", + "image_path": "99b684f16776c970dd34ed4c862f68e0ffbfb9a1bed356ef8ad69e3ffdd7a191.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "text", + "content": "Fig. 1: Meta-Rank benchmark. It is mainly divided into two stages: Test Case Generation and Meta-Rank (a). Test Case Generation: multiple factors (i.e., datasets, models and evaluation protocols) are combined to generate " + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "text", + "content": " different cases. MetaRank: (1) Case Execution. All " + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "text", + "content": " competitors (i.e., attribution methods) are applied to these cases, resulting in a collection of rankings. The details of attribution evaluation on an individual case are provided in (b). (2) Ranking Fusion. All rankings " + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\{R_1, R_2, \\ldots, R_\\tau\\}" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "text", + "content": " are subsequently fed into this module. The comparison of two competitors is transformed into the differences in their rankings across all cases, then integrated and converted into the discrepancy in Meta-Ranks. (3) Leaderboard. Ultimately, a unified leaderboard is obtained based on the Meta-Ranks of the competitors. “" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "text", + "content": "” is the test case, “" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "text", + "content": "” is the competitor, and “" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 337, + 482, + 459 + ], + "type": "text", + "content": "” is the Meta-Rank value." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 484, + 480, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 509 + ], + "type": "text", + "content": "attribute methods [40]. To provide a comprehensive view of the attribution methods, the proposed Meta-Rank considers both protocols for evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Test case settings. To gain a generalized evaluation for attributions, our analytical experiments evaluate existing attribution methods on a wide array of experiment settings, necessitating the initial preparation of baselines. We first train ResNet-18, Inception-v4 and VGG-19 on NWPU-RESISC45 and Food101. Subsequently, we exclusively train EfficientNet_b2 and DenseNet-121 on the NWPU-RESISC45 dataset. These baselines are implemented with the PyTorch [36] framework. Adam [31] is used to optimize the models, with the learning rate of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and the weight decay rate of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". The cosine annealing is adopted as the learning rate decay schedule. Furthermore, we also incorporate pre-trained ResNet-18, ResNet-50 sourced from Places-365, and pre-trained ResNet-18, Inception-v4, VGG-19 from ImageNet-1k. The aforementioned baselines, combined with two evaluation protocols, are served as the test cases for the benchmark. Detailed information for each test case can be found in Appendix C." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 330, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 330, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 330, + 128 + ], + "type": "text", + "content": "4.2 The Proposed Benchmark Metric" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 132, + 482, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 132, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 132, + 482, + 203 + ], + "type": "text", + "content": "We further propose a benchmark metric, named Meta-Rank, to facilitate a more comprehensive benchmark comparison between attribution methods. The main idea of the Meta-Rank metric is viewing each specific configuration of the model, the dataset and the evaluation protocol as one test case. Meta-Rank aggregates the ranking results (i.e., rank the rankings) from a diverse array of test cases, thereby mitigating biases associated with specific configurations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": "Formally, let the triplet " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\langle D_i,M_j,P_k\\rangle" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " be one test case for evaluating the attribution algorithms, where " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " denote the dataset, model and evaluation protocol respectively. The subscript " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " denote one specific configuration from the candidate datasets, models and protocols as described in Section 4.1. Let " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{C_1,C_2,\\dots,C_m\\}" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " be the set of the competitors (i.e., attribution methods). In the proposed benchmark, all the competitors are evaluated on every generated test case, thus producing a ranking among the competitors. For the test case " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\langle D_i,M_j,P_k\\rangle" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "R_{(i,j,k)} = (r_1^{(i,j,k)},r_2^{(i,j,k)},\\dots,r_m^{(i,j,k)})" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " to denote the rankings of all the competitors, where " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "r_m^{(i,j,k)}" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " denotes the rank of the " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": "-th attribution method. After evaluating the attribution methods on all the test cases, the probability that " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "C_p" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "inline_equation", + "content": "C_q" + }, + { + "bbox": [ + 130, + 204, + 482, + 342 + ], + "type": "text", + "content": " is formally defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 344, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 344, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 219, + 344, + 481, + 369 + ], + "type": "interline_equation", + "content": "\\mathrm {P} _ {q \\prec p} = \\frac {1}{| \\Omega |} \\sum_ {(i, j, k)} ^ {\\Omega} \\mathbf {1} \\left(r _ {q} ^ {(i, j, k)} \\prec r _ {p} ^ {(i, j, k)}\\right), \\tag {4}", + "image_path": "ab2792bda81b4e8522d93f0cfc8ff9183d0558e6c98f7ef8276470c41427b0a4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": " denotes the indicator function, and " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "r_q \\prec r_p" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": " represents the ranking " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "r_q" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": " is outperformed by " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "r_p" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": " is the joint space of datasets, models and evaluation protocols. The probability " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathsf{P}_{q \\prec p}" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": " is then converted into the logistic odds, obtaining a more sensitive metric " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathsf{Logit}_{q \\prec p}" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": " that can better capture slight differences between " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "C_p" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "inline_equation", + "content": "C_q" + }, + { + "bbox": [ + 130, + 373, + 482, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 437, + 481, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 437, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 188, + 437, + 481, + 452 + ], + "type": "interline_equation", + "content": "\\operatorname {L o g i t} _ {q \\prec p} = \\log \\operatorname {o d d s} \\left(C _ {p}, C _ {q}\\right) = \\log \\left(\\mathrm {P} _ {q \\prec p} / \\left(1 - \\mathrm {P} _ {q \\prec p}\\right)\\right). \\tag {5}", + "image_path": "8693b1d4f11dcafec16a4eff475fc214e611ccbd5b42636be7a382df287d2a76.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 502 + ], + "type": "text", + "content": "Based on the above derivation process, we can obtain a set of standardized results " + }, + { + "bbox": [ + 130, + 455, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{Logit}_{q\\prec p}\\}" + }, + { + "bbox": [ + 130, + 455, + 482, + 502 + ], + "type": "text", + "content": " that quantify the performance differences between any two competitors on all test cases. Then, we transform the performance discrepancy into the difference on the Meta-Rank metric:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 175, + 504, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 504, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 175, + 504, + 481, + 602 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\kappa_ {1} - \\kappa_ {2} = \\operatorname {L o g i t} _ {2 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 2 \\\\ \\kappa_ {1} - \\kappa_ {3} = \\operatorname {L o g i t} _ {3 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 3 \\\\ \\kappa_ {1} - \\kappa_ {4} = \\operatorname {L o g i t} _ {4 \\prec 1}, & \\text {C o m p e t i t o r} \\# 1 v s. \\# 4 \\\\ \\dots & \\dots \\\\ \\kappa_ {2} - \\kappa_ {3} = \\operatorname {L o g i t} _ {3 \\prec 2}, & \\text {C o m p e t i t o r} \\# 2 v s. \\# 3 \\\\ \\kappa_ {2} - \\kappa_ {4} = \\operatorname {L o g i t} _ {4 \\prec 2}, & \\text {C o m p e t i t o r} \\# 2 v s. \\# 4 \\\\ \\dots & \\dots \\\\ \\kappa_ {m - 1} - \\kappa_ {m} = \\operatorname {L o g i t} _ {m \\prec m - 1}. & \\text {C o m p e t i t o r} \\# \\mathrm {m} - 1 v s. \\# \\mathrm {m} \\end{array} \\right. \\tag {6}", + "image_path": "34a3ec4dddb9a3237c2e978edd3c47df77a83fe791efc099e01116a269e5dfd2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\kappa_{m}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " represents the Meta-Rank value of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_m" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". After solving the system of equations in Equation 6, we can derive a score " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " for each competitor, serving as an identifier for its comprehensive performance. Through the arrangement of competitors in descending order of their " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " values, a unified leaderboard can finally be constructed." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 219, + 182 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 219, + 182 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 219, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 219, + 182 + ], + "type": "image", + "image_path": "590109a90a800eadccecc5eab54d2de43cdb215560376459c2a5d0785e63ea10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 183, + 217, + 191 + ], + "lines": [ + { + "bbox": [ + 136, + 183, + 217, + 191 + ], + "spans": [ + { + "bbox": [ + 136, + 183, + 217, + 191 + ], + "type": "text", + "content": "(a) Food-ResNet18(M)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 306, + 182 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 306, + 182 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 306, + 182 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 306, + 182 + ], + "type": "image", + "image_path": "68874e528fe429141a2c9c7d61a04d0e54095d1c8a4bfbdf4d86aaaa0d3f0f9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 392, + 182 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 392, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 392, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 392, + 182 + ], + "type": "image", + "image_path": "a85a1efca9bcc4a1b86fc44273fd5ea8031d65eaece98a1b31e4e8727eda294f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 386, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 386, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 386, + 192 + ], + "type": "text", + "content": "(c) IN-ResNet18(M)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 483, + 182 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 483, + 182 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 483, + 182 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 483, + 182 + ], + "type": "image", + "image_path": "214d3f34eb9f25c66f95e7cb360c87041b05e175b3f752a62413d9a9839a81d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 182, + 473, + 192 + ], + "lines": [ + { + "bbox": [ + 403, + 182, + 473, + 192 + ], + "spans": [ + { + "bbox": [ + 403, + 182, + 473, + 192 + ], + "type": "text", + "content": "(d) IN-ResNet18(L)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 192, + 222, + 258 + ], + "blocks": [ + { + "bbox": [ + 135, + 192, + 222, + 258 + ], + "lines": [ + { + "bbox": [ + 135, + 192, + 222, + 258 + ], + "spans": [ + { + "bbox": [ + 135, + 192, + 222, + 258 + ], + "type": "image", + "image_path": "4a9275ddce357e16c3457891d944f9c20cb26f0a843591270f51633f2dd0b19d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 223, + 192, + 309, + 258 + ], + "blocks": [ + { + "bbox": [ + 223, + 183, + 303, + 192 + ], + "lines": [ + { + "bbox": [ + 223, + 183, + 303, + 192 + ], + "spans": [ + { + "bbox": [ + 223, + 183, + 303, + 192 + ], + "type": "text", + "content": "(b) Food-ResNet18(L)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 192, + 309, + 258 + ], + "lines": [ + { + "bbox": [ + 223, + 192, + 309, + 258 + ], + "spans": [ + { + "bbox": [ + 223, + 192, + 309, + 258 + ], + "type": "image", + "image_path": "3a29a1d4cf263775abb7900491272e2c4a772ee39711b09367989e63e93d5825.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 192, + 395, + 258 + ], + "blocks": [ + { + "bbox": [ + 309, + 192, + 395, + 258 + ], + "lines": [ + { + "bbox": [ + 309, + 192, + 395, + 258 + ], + "spans": [ + { + "bbox": [ + 309, + 192, + 395, + 258 + ], + "type": "image", + "image_path": "028712ff39f0b7f997da024ec668f6d80a9120b28b308d61a32a8b715c86096a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 396, + 192, + 483, + 258 + ], + "blocks": [ + { + "bbox": [ + 396, + 192, + 483, + 258 + ], + "lines": [ + { + "bbox": [ + 396, + 192, + 483, + 258 + ], + "spans": [ + { + "bbox": [ + 396, + 192, + 483, + 258 + ], + "type": "image", + "image_path": "ac8eb546ddf9941173d9921dbcfc4f714a5438e200a3476e040b5f2b357a870a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 133, + 270, + 220, + 336 + ], + "blocks": [ + { + "bbox": [ + 134, + 260, + 222, + 269 + ], + "lines": [ + { + "bbox": [ + 134, + 260, + 222, + 269 + ], + "spans": [ + { + "bbox": [ + 134, + 260, + 222, + 269 + ], + "type": "text", + "content": "(e) Food-Inceptionv4(M)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 270, + 220, + 336 + ], + "lines": [ + { + "bbox": [ + 133, + 270, + 220, + 336 + ], + "spans": [ + { + "bbox": [ + 133, + 270, + 220, + 336 + ], + "type": "image", + "image_path": "19a1bd99e420012140627505a9084fd65189e0c941ae00c8a5707bdba2da8e95.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 338, + 214, + 348 + ], + "lines": [ + { + "bbox": [ + 140, + 338, + 214, + 348 + ], + "spans": [ + { + "bbox": [ + 140, + 338, + 214, + 348 + ], + "type": "text", + "content": "(i) Food-VGG19(M)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 351, + 473, + 359 + ], + "lines": [ + { + "bbox": [ + 140, + 351, + 473, + 359 + ], + "spans": [ + { + "bbox": [ + 140, + 351, + 473, + 359 + ], + "type": "text", + "content": "Saliency GB -XG DeepLift -Deconv LRP -GuidedGradCam IG" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 373, + 482, + 428 + ], + "lines": [ + { + "bbox": [ + 131, + 373, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 482, + 428 + ], + "type": "text", + "content": "Fig.2: Evaluation results on 12 test cases. These cases consist of two datasets: Food-101 (a)(b)(e)(f)(i)(j), ImageNet-1k (c)(d)(g)(h)(k)(l), three models: ResNet-18 (a)(b)(c)(d), Inception-v4 (e)(f)(g)(h), VGG-19 (i)(j)(k)(l), and two protocols: MoRF(a)(c)(e)(g)(i)(k), LeRF(b)(d)(f)(h)(j)(l). \"Baseline\" represents the accuracy of the model when no features are ablated." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 221, + 270, + 306, + 336 + ], + "blocks": [ + { + "bbox": [ + 223, + 260, + 309, + 269 + ], + "lines": [ + { + "bbox": [ + 223, + 260, + 309, + 269 + ], + "spans": [ + { + "bbox": [ + 223, + 260, + 309, + 269 + ], + "type": "text", + "content": "(f) Food-Inceptionv4(L)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 270, + 306, + 336 + ], + "lines": [ + { + "bbox": [ + 221, + 270, + 306, + 336 + ], + "spans": [ + { + "bbox": [ + 221, + 270, + 306, + 336 + ], + "type": "image", + "image_path": "f5a9a17f674cfaa427b019217c77464c00fac5988a1c3f652a534868db0efecb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 338, + 299, + 348 + ], + "lines": [ + { + "bbox": [ + 229, + 338, + 299, + 348 + ], + "spans": [ + { + "bbox": [ + 229, + 338, + 299, + 348 + ], + "type": "text", + "content": "(j) Food-VGG19(L)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 306, + 270, + 392, + 336 + ], + "blocks": [ + { + "bbox": [ + 313, + 261, + 392, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 392, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 392, + 270 + ], + "type": "text", + "content": "(g) IN-Inceptionv4(M)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 270, + 392, + 336 + ], + "lines": [ + { + "bbox": [ + 306, + 270, + 392, + 336 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 392, + 336 + ], + "type": "image", + "image_path": "e512d465e38b2b422dd7de213c84ae366157974d93239303ab2a48f7a3d7c5f3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 338, + 383, + 348 + ], + "lines": [ + { + "bbox": [ + 317, + 338, + 383, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 383, + 348 + ], + "type": "text", + "content": "(k) IN-VGG19(M)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 396, + 270, + 483, + 336 + ], + "blocks": [ + { + "bbox": [ + 400, + 261, + 479, + 270 + ], + "lines": [ + { + "bbox": [ + 400, + 261, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 400, + 261, + 479, + 270 + ], + "type": "text", + "content": "(h) IN-Inceptionv4(L)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 270, + 483, + 336 + ], + "lines": [ + { + "bbox": [ + 396, + 270, + 483, + 336 + ], + "spans": [ + { + "bbox": [ + 396, + 270, + 483, + 336 + ], + "type": "image", + "image_path": "738ae351af61a1fb10f336a95de5f07bc5ad8345eeb7e78ee4d2f45f1eb6a829.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 338, + 469, + 348 + ], + "lines": [ + { + "bbox": [ + 407, + 338, + 469, + 348 + ], + "spans": [ + { + "bbox": [ + 407, + 338, + 469, + 348 + ], + "type": "text", + "content": "(1) IN-VGG19(L)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 449, + 234, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 234, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 234, + 462 + ], + "type": "text", + "content": "5 Main Results" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 473, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 498 + ], + "type": "text", + "content": "This section provides a detailed elaboration of extensive experiments on Meta-Rank, with a particular emphasis on the following pivotal questions." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 138, + 504, + 493, + 599 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 138, + 504, + 493, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 493, + 528 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 493, + 528 + ], + "type": "text", + "content": "- RQ1: Are attribution methods consistently effective across diverse settings (i.e., datasets, models, evaluation protocols) and distinct checkpoints? (Section 5.1)" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 139, + 528, + 480, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 528, + 480, + 551 + ], + "spans": [ + { + "bbox": [ + 139, + 528, + 480, + 551 + ], + "type": "text", + "content": "- RQ2: Does the state-of-the-art evaluation method, ROAD, yield consistent rankings across different test cases? (Section 5.2)" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 139, + 552, + 480, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 552, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 139, + 552, + 480, + 574 + ], + "type": "text", + "content": "- RQ3: According to our proposed Meta-Rank, which attribution methods exhibit superior performance across a wide range of cases? (Section 5.3)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 139, + 575, + 480, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 575, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 139, + 575, + 480, + 599 + ], + "type": "text", + "content": "- RQ4: Can Meta-Rank conduct the evaluation of substantial methods within a feasible timeframe? (Section 5.4)" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 622, + 319, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 319, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 319, + 635 + ], + "type": "text", + "content": "5.1 RQ1: Consistency Investigation" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we investigate the consistency of performance rankings for existing attribution methods across diverse evaluation settings. Specifically, we focus" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 244, + 205 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 244, + 205 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 244, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 244, + 205 + ], + "type": "image", + "image_path": "ac78e231fdee73c10604645bc441a62aa2af36108ef3d171d2427efdb13aa783.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 209, + 215, + 218 + ], + "lines": [ + { + "bbox": [ + 178, + 209, + 215, + 218 + ], + "spans": [ + { + "bbox": [ + 178, + 209, + 215, + 218 + ], + "type": "text", + "content": "(a) MoRF" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 115, + 362, + 205 + ], + "blocks": [ + { + "bbox": [ + 255, + 115, + 362, + 205 + ], + "lines": [ + { + "bbox": [ + 255, + 115, + 362, + 205 + ], + "spans": [ + { + "bbox": [ + 255, + 115, + 362, + 205 + ], + "type": "image", + "image_path": "102d25f1a7c33b50dd58b5071d0f98620881792a79adb775e065fa8b37880800.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 209, + 332, + 218 + ], + "lines": [ + { + "bbox": [ + 296, + 209, + 332, + 218 + ], + "spans": [ + { + "bbox": [ + 296, + 209, + 332, + 218 + ], + "type": "text", + "content": "(b) LeRF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 227, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 294 + ], + "type": "text", + "content": "Fig. 3: Spearman correlation among nine test cases (NWPU, Food, and ImageNet datasets paired with ResNet-18, Inception-v4, and VGG-19 models) in MoRF (a) and LeRF (b), and between MoRF and LeRF on the same nine cases (c). The labels along the horizontal arrow are equivalent to those along the vertical arrow. Here, we define a correlation score higher than 0.8 as strongly correlated (■), between 0.6 and 0.8 as moderately correlated (▲), and lower than 0.6 as weakly correlated (▼)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 372, + 124, + 475, + 204 + ], + "blocks": [ + { + "bbox": [ + 372, + 124, + 475, + 204 + ], + "lines": [ + { + "bbox": [ + 372, + 124, + 475, + 204 + ], + "spans": [ + { + "bbox": [ + 372, + 124, + 475, + 204 + ], + "type": "image", + "image_path": "73436c5a4764c326cac077aa5ac60d5787c36f3afe5c6883b0d54b97f6b1df25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 209, + 461, + 218 + ], + "lines": [ + { + "bbox": [ + 388, + 209, + 461, + 218 + ], + "spans": [ + { + "bbox": [ + 388, + 209, + 461, + 218 + ], + "type": "text", + "content": "(c) MoRF v.s. LeRF" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 318, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 402 + ], + "type": "text", + "content": "on the model-, dataset-, protocol- and checkpoint-level ranking consistencies by varying these factors in the evaluation. To this end, we assess the aforementioned eight attribution methods on all the generated test cases, i.e., with six models, four datasets and two ablation protocols (MoRF and LeRF). This yields a total of 26 groups of results, each corresponding to one test case. Some experimental results are presented in Figure 2. For more details of the experimental results, please refer to Appendix D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "text", + "content": "Model-level Consistency. It is evident that the attribution methods exhibit varying performance across different models. To elaborate: (1) LRP has the worst performance on the \"ImageNet-Inceptionv4(M)\" case but significantly better results on the \"ImageNet-ResNet18(M)\" and \"ImageNet-VGG19(M)\" cases. (2) Integrated Gradients and Guided Grad-CAM exhibit comparable performance on the \"Food-ResNet18(M)\" case, but show great disparity on the \"Food-VGG19(M)\" case. For a more clear measurement of the inconsistency, we calculate the Spearman Rank correlations among several cases and illustrate them in Figure 3. Here, we extract area under the curve (AUC) [4] of all attribution methods (detailed in Appendix D) to generate a ranking list for each case. In the MoRF protocol, removing the most critical pixels identified by the attribution method causes a substantial drop in model performance. Thus, a lower AUC value indicates a superior method. And the LeRF protocol follows the inverse principle. The results suggest a pronounced weakness in the consistency at the model level, with a strong correlation found in only 2 cases (i.e., the correlation between ResNet-18 and Inception-v4 on the Food-101 dataset in both MoRF and LeRF protocols) out of 18 cases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Dataset-level Consistency. Analogously, attribution methods have shown noticeable inconsistency in performance on different datasets. For instance, Guided Backpropagation demonstrates superior performance on \"ImageNet-VGG19(M)\" yet ranks third-worst on \"Food-VGG19(M)\". In addition, according to the results of Spearman analysis, only " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\frac{5}{18}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " cases show strong correlations at the dataset level." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 180, + 299, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 302, + 173 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 302, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 302, + 173 + ], + "type": "text", + "content": "Table 2: AUC rankings of the eight methods at different checkpoints along the same trajectory. Here, ResNet-18 trained on NWPU-RESISC45 is taken as an illustrative example." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 180, + 299, + 281 + ], + "lines": [ + { + "bbox": [ + 135, + 180, + 299, + 281 + ], + "spans": [ + { + "bbox": [ + 135, + 180, + 299, + 281 + ], + "type": "table", + "html": "
Epoch-50Epoch-100Epoch-150Epoch-200
MoRF
LeRF
SaliencyIXGDeconvGuidedGradCam
GBDeepLiftLRPIG
", + "image_path": "80d36368f64414bdf589ee815780c21538c749e6ad56fd44ecf6d1f23d401917.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 212, + 478, + 283 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "type": "text", + "content": "Table 3: Comparison of Spearman results among four cases between ROAD and feature ablation. " + }, + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "type": "inline_equation", + "content": "a \\mid b" + }, + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "type": "text", + "content": " represents a pair of Spearman results, where " + }, + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "type": "text", + "content": " denotes using feature ablation (in Section 5.1) and " + }, + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 309, + 114, + 482, + 203 + ], + "type": "text", + "content": " denotes using noisy linear imputation. The settings of color blocks in this table are the same as those in Figure 3." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 212, + 478, + 283 + ], + "lines": [ + { + "bbox": [ + 315, + 212, + 478, + 283 + ], + "spans": [ + { + "bbox": [ + 315, + 212, + 478, + 283 + ], + "type": "table", + "html": "
NWPU-ResNet18Food-Inceptionv4
MoRF (#1)LeRF (#2)MoRF (#3)LeRF (#4)
#1| || || || |||
#2| || || || |||
#3| || || || |||
#4| || || || |||
", + "image_path": "e2f5b8da111f2816471a99754fc94f8271d5d85d2ab588518051c728bd2c68f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 399 + ], + "type": "text", + "content": "Protocol-level Consistency. The evaluation results indicate that significant inconsistencies between the MoRF and LeRF protocols still exist. A detailed comparison between \"Food-Inceptionv4(M)\" and \"Food-Inceptionv4(L)\" reveals marked discrepancies in Saliency method under MoRF versus LeRF. Besides, the Spearman correlation between MoRF and LeRF is computed for nine test cases (in Figure 3 (c)). It can be easily seen that the correlation of all cases at the protocol level is either moderately correlated or weakly correlated, with no strong correlation observed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 400, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 581 + ], + "type": "text", + "content": "Checkpoint-level Consistency. We also consider the consistency of attribution methods at different training stages of the model. To illustrate, using ResNet-18 trained on NWPU-RESISC45 as a case study, we extract distinct checkpoints from the same training trajectory and employ two evaluation protocols for attribution analysis. The results of the AUC rankings are recorded in Table 2. It can be easily observed that the rankings of attribution methods demonstrate high consistency across different checkpoints. In MoRF, the outcomes of Epoch-200 differ by only one distinction from those at both Epoch-100 and Epoch-150, while exhibiting slightly greater disparities when compared to those at Epoch-50. Similarly for LeRF, the rankings at Epoch-200 also exhibit a single deviation from those at Epoch-100 and Epoch-150, respectively, and show two distinctions in comparison to those at Epoch-50. These findings suggest a notable consistency in the ranking of attribution methods across different checkpoints. Moreover, the rankings tend to stabilize rapidly as the model converges. More detailed results under other settings can be found in Appendix D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 312, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 312, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 312, + 611 + ], + "type": "text", + "content": "5.2 RQ2: Necessity of Meta-Rank" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To further substantiate the indispensability of the Meta-Rank benchmark, we employ the state-of-the-art attribution evaluation method known as ROAD for assessing the attribution methods across various cases. Starting from information theoretic analyses, ROAD identifies the class information leakage problem" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 200, + 304, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 123, + 304, + 190 + ], + "lines": [ + { + "bbox": [ + 132, + 123, + 304, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 123, + 304, + 190 + ], + "type": "text", + "content": "Table 4: Leaderboard of the eight attribution methods. The larger Meta-Rank indicates the stronger ability of attribution method. The top three methods exhibit more positive performance, while the remaining five methods perform negatively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 200, + 304, + 267 + ], + "lines": [ + { + "bbox": [ + 133, + 200, + 304, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 200, + 304, + 267 + ], + "type": "table", + "html": "
MethodsMeta-Rank Score#Rank
InputGradient2.32001
Integrated Gradients2.07362
DeepLift1.07803
Saliency-0.04834
LRP-0.22345
Guided Backpropagation-0.25856
Guided Grad-CAM-0.96797
Deconvolution-3.97348
", + "image_path": "09b333279a5ca05efe63f46641be21266a0be9e893c3868406db4103accdaab2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 324, + 114, + 470, + 211 + ], + "blocks": [ + { + "bbox": [ + 324, + 114, + 470, + 211 + ], + "lines": [ + { + "bbox": [ + 324, + 114, + 470, + 211 + ], + "spans": [ + { + "bbox": [ + 324, + 114, + 470, + 211 + ], + "type": "image", + "image_path": "1e282c6aecdfde2b2ad839227b11212ca6466ad488dbbedb9e8978ce71fb8b28.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 212, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 312, + 212, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 312, + 212, + 481, + 266 + ], + "type": "text", + "content": "Fig. 4: Time consumption of the \"Ranking Fusion\" module in Meta-Rank. 2, 4, 8, 16, 32, and 64 are the number of evaluated attribution methods. The gray bar represents the relative error." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 420 + ], + "type": "text", + "content": "dictated by the ablation mask and finally derives the noisy linear imputation for minimally revealing imputation, which is empirically demonstrated to produce consistent rankings of attribution methods between evaluations with MoRF and LeRF. However, it overlooks consistency at both the model and dataset levels. Thus, we randomly select four cases, namely NWPU-ResNet18(M), NWPU-ResNet18(L), Food-Inceptionv4(M), and Food-Inceptionv4(L), to assess the eight aforementioned attribution methods using ROAD. The evaluation metric used in this study aligns with those in RQ1. Table 3 presents the Spearman correlations among the evaluation results of the four cases. Further information regarding the accuracy decay curves is available in Appendix E." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 421, + 481, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 481, + 504 + ], + "type": "text", + "content": "It can be easily seen that, even when ROAD is employed as a substitute for feature ablation, the correlation between the rankings of attribution methods on different cases remains inconsistent. Specifically, all correlations demonstrate either moderate or weak, with no instances of strong correlation. This indicates that relying solely on ROAD for evaluating attribution methods across distinct cases is inadequate. Consequently, there is a critical need for a cross-case benchmark to ensure robust ranking results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 524, + 391, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 391, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 391, + 536 + ], + "type": "text", + "content": "5.3 RQ3: Attribution Evaluation with Meta-Rank" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 628 + ], + "type": "text", + "content": "Following the evaluation scheme outlined in the Meta-Rank benchmark (described in Section 4), we have carried out a unified measurement of the eight attribution methods. The consistency validation experiments in Section 5.1 generate evaluation results for attribution methods across distinct configurations, encompassing four datasets, six models, and two protocols, resulting in 26 sets of data. We then transform these results into AUC scores, which serve as the performance metric for the attribution methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Meta-Rank generates a comprehensive value for each attribution method. These uniquely determined scores enable us to establish a unified ranking list for all the methods. The finalized leaderboard for the eight attribution methods" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "is presented in Table 4. The results reveal InputGradient as the top-performing approach, while Deconvolution exhibits the poorest performance. (1) In particular, InputGradient achieves the highest score among the eight methods. By element-wise multiplying the gradients with the inputs, InputGradient produces more stable attributions in the face of perturbations, which contributes to its superior performance across multiple test cases. (2) Furthermore, the second-ranked method, Integrated Gradients, also shows great potential, with its score closely approaching that of the top-ranked method. (3) In contrast, Deconvolution obtains the lowest score, likely attributable to the hierarchical structure of the network accumulating noise during backpropagation. It is worth noting that a Meta-Rank score can assume both positive and negative values. Positive values signify comparatively superior attribution performance, whereas negative values indicate inferior performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 271, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 344 + ], + "type": "text", + "content": "The experimental findings demonstrate that Meta-Rank furnishes a comprehensive benchmark for attribution evaluation across diverse test cases, while effectively ameliorating the issue of inconsistent performance across heterogeneous cases. It computes the discrepancies between methods and converts them into the differences in capacity to derive the Meta-Rank. This result is unique, consistent, and aligned with practical application needs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 358, + 313, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 313, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 313, + 370 + ], + "type": "text", + "content": "5.4 RQ4: Efficiency of Meta-Rank" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "type": "text", + "content": "To measure the time consumption, we conduct experiments on the CPU device (Intel(R) Xeon(R) Platinum 8260L CPU @ 2.40GHz) using different numbers of methods. Here, we focus on the additional \"Ranking Fusion\" module, as the \"Case Execution\" module has been generically applied in all evaluations. In particular, we compare the time expenses of Meta-Rank during the calculation of the 2, 4, 8, 16, 32, and 64 methods. It is important to highlight that the first three test groups (i.e., 2, 4, and 8 methods) rely on real data gathered from previous experiments, whereas the last three test groups (i.e., 16, 32, and 64 methods) employ simulated data randomly generated. The data here refers to a set of accuracies (i.e., " + }, + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\{acc_i\\}" + }, + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "type": "text", + "content": " in Figure 1) obtained on each test case according to every attribution method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 604 + ], + "type": "text", + "content": "The results in Figure 4 indicate that as the number of attribution methods increases, the computation time also increases. However, since the number of attribution methods is finite, the computational time remains within an acceptable range. For example, when 64 methods are involved in the Meta-Rank evaluation, the time required to generate the leaderboard is approximately 0.12s, which completely meets expectations. Additionally, using GPUs for Meta-Rank calculations can further enhance the computation speed. This efficiency is beneficial for the promotion and application of Meta-Rank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 619, + 217, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 619, + 217, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 619, + 217, + 631 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Choice of faithfulness metric. Although there are many dimensions for attribution evaluation in previous work [24], faithfulness remains the most solid" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "and general criteria for benchmarking attribution methods. In the context of multiple experimental settings, as such, we opt to use it as the base metric " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " in our framework to facilitate subsequent Meta-Rank assessment and then enable broader evaluation on novel tasks. Meanwhile, Meta-Rank can also accommodate evaluations on other dimensions by enabling the derivation of a ranking list along each respective evaluation metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "content": "Missingness bias in feature ablation. Previous work has discussed the missingness bias problem caused by the removal of pixels through masking values in CNNs [29]. Hooker et al. proposed ROAR to address this issue, but it inadvertently introduced the problem of information leakage. Subsequently, Rong et al. developed ROAD to mitigate this leakage, but it failed to fully remove the pixels (explained in Appendix E). We recognize that completely eliminating this bias poses substantial difficulties, since the root of missingness bias lies in the operation of removing pixels within CNNs. Furthermore, the impact of bias exhibits variability across individual cases, also posing additional challenges in quantification and measurement. However, we believe that with a more diverse array of test cases, the missingness bias is more neutralized. Future research will further explore solutions for eliminating this bias." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "text", + "content": "Scalability of Meta-Rank. When introducing new attribution methods to the ranking, we initially conduct tests on several existing cases. The time complexity of evaluating a method on an individual case is linear, which has been demonstrated in previous studies [37, 40]. Subsequently, the obtained results, in conjunction with historical data, are fed into the \"Ranking Fusion\" module to generate an updated ranking list. The time overhead associated with this process is also optimistic, as illustrated in Section 5.4. This flexibility allows for the inclusion of any attribution method into the leaderboard at any time. Future endeavors will focus on expanding the corpus of test cases and pursuing further computational efficiency by leveraging parallel computing paradigms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 472, + 220, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 220, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 220, + 484 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "To address the lack of systematic research in feature attribution, we propose Meta-Rank, a novel standardized benchmark for robust and unified evaluation of methods. Meta-Rank aggregates relative comparisons between all method pairs across diverse settings, mitigating inconsistencies. Through extensive experiments, we reveal three critical insights: diverse evaluation settings yield divergent performance rankings, assessment results remain consistent across distinct checkpoints along the same training trajectory, and previous attempts at consistent evaluation exhibit limitations on heterogeneous models and datasets. Meta-Rank provides a leaderboard that highlights the broader applicability of InputGradient and Integrated Gradients. Remarkably, Meta-Rank exhibits exceptional computational efficiency, enabling ranking computations to be completed within milliseconds even for a substantial number of methods. In the future, we hope the proposed benchmark will facilitate the development of faithful and consistent explanations, fostering transparency and accountability." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 496, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 496, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 496, + 199 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (62106220), Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies, Ningbo 2025 Science and Technology Innovation Major Project (No. 2022Z072), and the Fundamental Research Funds for the Central Universities (226-2024-00058)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 239, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 262 + ], + "type": "text", + "content": "1. Adadi, A., Berrada, M.: Peeking inside the black-box: a survey on explainable artificial intelligence (xai). IEEE access 6, 52138-52160 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 295 + ], + "type": "text", + "content": "2. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. arXiv preprint arXiv:1711.06104 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 481, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 481, + 327 + ], + "type": "text", + "content": "3. Arias-Duart, A., Parés, F., Garcia-Gasulla, D., Giménez-Ábalos, V.: Focus! rating xai methods and finding biases. In: 2022 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). pp. 1-8. IEEE (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 481, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 481, + 360 + ], + "type": "text", + "content": "4. Atanasova, P., Simonsen, J.G., Lioma, C., Augenstein, I.: A diagnostic study of explainability techniques for text classification. arXiv preprint arXiv:2009.13295 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "type": "text", + "content": "5. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 393, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 481, + 415 + ], + "type": "text", + "content": "6. Bastings, J., Aziz, W., Titov, I.: Interpretable neural predictions with differentiable binary variables. arXiv preprint arXiv:1905.08160 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 481, + 436 + ], + "type": "text", + "content": "7. Bhatt, U., Weller, A., Moura, J.M.: Evaluating and aggregating feature-based model explanations. arXiv preprint arXiv:2005.00631 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "type": "text", + "content": "8. Binder, A., Weber, L., Lapuschkin, S., Montavon, G., Müller, K.R., Samek, W.: Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16143-16152 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 481, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 481, + 523 + ], + "type": "text", + "content": "9. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "text", + "content": "10. Brocki, L., Chung, N.C.: Evaluation of interpretability methods and perturbation artifacts in deep neural networks. arXiv preprint arXiv:2203.02928 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "type": "text", + "content": "1. Brunke, L., Agrawal, P., George, N.: Evaluating input perturbation methods for interpreting cnns and saliency map comparison. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 120-134. Springer (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "text", + "content": "2. Chattopadhy, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Gradcam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "3. Chefer, H., Gur, S., Wolf, L.: Transformer interpretability beyond attention visualization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 782-791 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "15. Colin, J., Fel, T., Cadène, R., Serre, T.: What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems 35, 2832-2845 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "16. Covert, I., Lundberg, S.M., Lee, S.I.: Understanding global feature contributions with additive importance measures. Advances in Neural Information Processing Systems 33, 17212-17223 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "17. Deng, H., Zou, N., Du, M., Chen, W., Feng, G., Yang, Z., Li, Z., Zhang, Q.: Understanding and unifying fourteen attribution methods with taylor interactions. arXiv preprint arXiv:2303.01506 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 237, + 480, + 270 + ], + "type": "text", + "content": "18. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 134, + 270, + 480, + 303 + ], + "type": "text", + "content": "19. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 303, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 303, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 303, + 480, + 335 + ], + "type": "text", + "content": "20. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 335, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 335, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 335, + 480, + 369 + ], + "type": "text", + "content": "21. Goyal, Y., Wu, Z., Ernst, J., Batra, D., Parikh, D., Lee, S.: Counterfactual visual explanations. In: International Conference on Machine Learning. pp. 2376-2384. PMLR (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 402 + ], + "type": "text", + "content": "22. Hartley, T., Sidorov, K., Willis, C., Marshall, D.: Explaining failure: Investigation of surprise and expectation in cnns. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 12-13 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 402, + 480, + 435 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 435, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 480, + 479 + ], + "type": "text", + "content": "24. Hedström, A., Weber, L., Krakowczyk, D., Bareeva, D., Motzkus, F., Samek, W., Lapuschkin, S., Höhne, M.M.C.: Quantus: An explainable ai toolkit for responsible evaluation of neural network explanations and beyond. Journal of Machine Learning Research 24(34), 1-11 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "text", + "content": "25. Hesse, R., Schaub-Meyer, S., Roth, S.: Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3981-3991 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 511, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 511, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 511, + 480, + 544 + ], + "type": "text", + "content": "26. Hooker, S., Erhan, D., Kindermans, P.J., Kim, B.: A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 544, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 480, + 577 + ], + "type": "text", + "content": "27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 577, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 577, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 577, + 480, + 610 + ], + "type": "text", + "content": "28. Ismail, A.A., Gunady, M., Corrada Bravo, H., Feizi, S.: Benchmarking deep learning interpretability in time series predictions. Advances in neural information processing systems 33, 6441-6452 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 610, + 480, + 632 + ], + "type": "text", + "content": "29. Jain, S., Salman, H., Wong, E., Zhang, P., Vineet, V., Vermprala, S., Madry, A.: Missingness bias in model debugging. arXiv preprint arXiv:2204.08945 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 666 + ], + "type": "text", + "content": "30. Khakzar, A., Khorsandi, P., Nobahari, R., Navab, N.: Do explanations explain? model knows best. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10244-10253 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "text", + "content": "31. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 172 + ], + "type": "text", + "content": "32. Krishna, S., Han, T., Gu, A., Pombra, J., Jabbari, S., Wu, S., Lakkaraju, H.: The disagreement problem in explainable machine learning: A practitioner's perspective. arXiv preprint arXiv:2202.01602 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 205 + ], + "type": "text", + "content": "33. Leem, S., Seo, H.: Attention guided cam: Visual explanations of vision transformer guided by self-attention. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2956-2964 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "34. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "35. Nguyen, G., Kim, D., Nguyen, A.: The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems 34, 26422-26436 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 306 + ], + "type": "text", + "content": "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32, 8026-8037 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 329 + ], + "type": "text", + "content": "37. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 352 + ], + "type": "text", + "content": "38. Rao, S., Böhle, M., Schiele, B.: Better understanding differences in attribution methods via systematic evaluations. arXiv preprint arXiv:2303.11884 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "39. Ribeiro, M.T., Singh, S., Guestrin, C.: \"why should i trust you?\" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "40. Rong, Y., Leemann, T., Borisov, V., Kasneci, G., Kasneci, E.: A consistent and efficient evaluation strategy for attribution methods. In: International Conference on Machine Learning. pp. 18770-18795. PMLR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.R.: Evaluating the visualization of what a deep neural network has learned. IEEE transactions on neural networks and learning systems 28(11), 2660-2673 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "type": "text", + "content": "42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "43. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "text", + "content": "44. Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not just a black box: Learning important features through propagating activation differences. arXiv preprint arXiv:1605.01713 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 564, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 597 + ], + "type": "text", + "content": "45. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: visualising image classification models and saliency maps. In: Proceedings of the International Conference on Learning Representations (ICLR). ICLR (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "type": "text", + "content": "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "text", + "content": "47. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "48. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 178, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Evaluation Consistency of Attribution-based Explanations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 446 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "49. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "50. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "51. Wang, Y., Wang, X.: A unified study of machine learning explanation evaluation metrics. arXiv preprint arXiv:2203.14265 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 258 + ], + "type": "text", + "content": "52. Xue, M., Huang, Q., Zhang, H., Cheng, L., Song, J., Wu, M., Song, M.: Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In: 33rd International Joint Conference on Artificial Intelligence, IJCAI-24. International Joint Conferences on Artificial Intelligence Organization (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "53. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 818-833. Springer (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "54. Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-down neural attention by excitation backprop. International Journal of Computer Vision 126(10), 1084-1102 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "text", + "content": "55. Zhang, Y., Khakzar, A., Li, Y., Farshad, A., Kim, S.T., Navab, N.: Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems 34, 20040-20051 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "57. Zhou, Y., Booth, S., Ribeiro, M.T., Shah, J.: Do feature attribution methods correctly attribute features? In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9623-9633 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "text", + "content": "58. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. arXiv preprint arXiv:1702.04595 (2017)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Duan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_content_list.json b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..896c813c01da57a1566941324686bd5e0ac6be71 --- /dev/null +++ b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_content_list.json @@ -0,0 +1,1670 @@ +[ + { + "type": "text", + "text": "On the Topology Awareness and Generalization Performance of Graph Neural Networks", + "text_level": 1, + "bbox": [ + 230, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junwei Su and Chuan Wu†", + "bbox": [ + 401, + 212, + 599, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, the University of Hong Kong {jwsu, cwu} @cs.hku.hk", + "bbox": [ + 289, + 239, + 712, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Many computer vision and machine learning problems are modelled as learning tasks on graphs, where graph neural networks (GNNs) have emerged as a dominant tool for learning representations of graph-structured data. A key feature of GNNs is their use of graph structures as input, enabling them to exploit the graphs' inherent topological properties—known as the topology awareness of GNNs. Despite the empirical successes of GNNs, the influence of topology awareness on generalization performance remains unexplored, particularly for node-level tasks that diverge from the assumption of data being independent and identically distributed (I.I.D.). The precise definition and characterization of the topology awareness of GNNs, especially concerning different topological features, are still unclear. This paper introduces a comprehensive framework to characterize the topology awareness of GNNs across any topological feature. Using this framework, we investigate the effects of topology awareness on GNN generalization performance. Contrary to the prevailing belief that enhancing the topology awareness of GNNs is always advantageous, our analysis reveals a critical insight: improving the topology awareness of GNNs may inadvertently lead to unfair generalization across structural groups, which might not be desired in some scenarios. Additionally, we conduct a case study using the intrinsic graph metric, the shortest-path distance, on various benchmark datasets. The empirical results of this case study confirm our theoretical insights. Moreover, we demonstrate the practical applicability of our framework by using it to tackle the cold start problem in graph active learning.", + "bbox": [ + 261, + 304, + 748, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 375, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many problems in computer vision and machine learning are modeled as learning tasks on graphs. For example, in semantic segmentation, graphs model the relationships between different image regions, enhancing accuracy and context-aware segmentation. Graph neural networks (GNNs) have emerged as a dominant class of machine learning models specifically designed for learning representations of graph-structured data. They have demonstrated considerable success in addressing a wide range of graph-related problems in various domains such as chemistry [10], biology [41], social networking [6,22,35], scene graph generation [49,54]", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†: corresponding authors", + "bbox": [ + 230, + 825, + 400, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and visual relationship detection [24,46,52]. A defining characteristic of GNNs is their use of a spatial approach through message passing on the graph structure for feature aggregation. This enables GNNs to preserve structural information or dependencies (referred to as topology awareness) from the underlying graph structure, allowing them to be highly effective in tasks such as node classification. Fig. 1 illustrates the overall learning process of GNNs.", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their practicality and potential, there remains a lack of theoretical understanding about GNNs, particularly in the semi-supervised node classification setting where the dependencies among the data differ significantly from other machine learning models [25]. In this setting, the goal is to leverage relations, as captured by the graph structure, among the data and a small set of labelled nodes to predict labels for the remaining nodes. Most of the existing theoretical studies of GNNs have focused on the connection between the message-passing mechanism of GNNs and the Weisfeiler-Lehman isomorphism test [19], aiming to understand the ability of GNNs to differentiate different graph structures in the learned representations, known as the expressive power of GNNs. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [28,32,51].", + "bbox": [ + 212, + 241, + 787, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as GNNs become more reliant on and sensitive (aware) of graph structure as input, they may exhibit different generalization performances towards certain structural subgroups (distinct data subsets grouped by structural similarity to the training set) within the data. The quantification of GNN generalization across distinct structural subgroups is termed structural subgroup generalization [25]. Such considerations are vital in GNN application and development. For instance, within protein-protein interaction networks, these structural subgroups could represent different molecular complexes, influencing the accuracy of interaction predictions. Similarly, understanding how the topology awareness of GNNs influences generalization is essential when devising sampling strategies for training. The extent to which the generalization performance of GNNs is influenced by specific structural features of graph data is critical in deciding the composition of training datasets. Despite its importance, an understanding of the relationship between the topology awareness of GNNs and its structural subgroup generalization is still lacking. Furthermore, characterizing the topology awareness of GNNs poses a challenge, especially considering that different domains and tasks may prioritize distinct structural aspects. Therefore, a versatile framework is needed to assess the topology awareness of GNNs in relation to various structures.", + "bbox": [ + 212, + 441, + 787, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this gap, in this paper, we propose a novel framework based on approximate metric embedding to study the relationship between structural subgroup generalization and topology awareness of GNNs in the context of semi-supervised node classification. The proposed framework allows for the investigation of the structural subgroup generalization of GNNs with respect to different structural subgroups. More concretely, the main contributions of this work are summarized as follows.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/70cc696bee85f371baf0fa14792b5a03e165bb9a7edfc83c0f5ffcc78b9fc29b.jpg", + "image_caption": [ + "Fig. 1. An illustration of the learning process in a 2-layer GNN. The message-passing mechanism leverages the graph structure to aggregate information, thereby generating the representation/embedding $h_a$ for the target vertex $a$ (highlighted in red)." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 781, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. We propose a novel, structure-agnostic framework using approximate metric embedding to examine the interplay between GNNs' structural subgroup generalization and topology awareness. This framework is versatile, accommodating various structural measures like shortest-path distance, and requires only the corresponding structural measure. Its simplicity in estimating key factors makes it applicable and generalizable to a wide range of scenarios.", + "bbox": [ + 222, + 356, + 782, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Through formal analysis within our framework, we establish a clear link between GNN topology awareness and their generalization performance (Theorem 1). We also demonstrate that while enhanced topology awareness boosts GNN expressiveness, it can result in uneven generalization performance, favouring subgroups more structurally similar to the training set (Theorem 2). Such structural property can be harmful (causing unfairness issues) or useful (informing design decisions) depending on the scenario. This challenges the prevailing belief that increased topology awareness universally benefits GNNs [28, 32, 51], emphasizing the importance of considering the relationship between topology awareness and generalization performance.", + "bbox": [ + 222, + 446, + 784, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. We validate our framework through a case study on shortest-path distance, highlighting its practicality and relevance. The results corroborate our theoretical findings, showing that GNNs with heightened awareness of shortest-path distances excel in classifying vertex groups closer to the training set. Moreover, we demonstrate how our findings can be applied to mitigate the cold start problem in graph active learning [11,15], highlighting the practical implications of our framework and results.", + "bbox": [ + 222, + 598, + 784, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 736, + 387, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Topology Awareness of GNNs", + "text_level": 1, + "bbox": [ + 212, + 768, + 508, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modern GNN architectures take inspiration from the Weisfeiler-Lehman isomorphism test [12, 19], which employs the graph structure to propagate information. Consequently, a significant portion of existing literature on GNNs concentrates", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on understanding their ability to differentiate various graph structures, known as the expressive power of GNNs [27-29, 47, 51, 53]. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties [37, 38] in the learned representation [28, 32, 51]. Nevertheless, the precise ramifications of heightened topology awareness on GNNs' generalization performance remain shrouded in ambiguity. The intricate variations within structures across different domains and tasks further compound this complexity, underscoring the necessity for a unified framework—akin to the one proposed in this paper—to shed light on these intricacies.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Generalization of GNNs", + "text_level": 1, + "bbox": [ + 215, + 321, + 460, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relatively few studies focus on understanding the generalization performance of GNNs in node-level tasks. [3,9,18,23,25,33,39,40,44,48] extend the conventional statistical learning or information-theoretic framework to GNNs, providing different generalization bounds based on concepts such as Vapnik-Chervonenkis dimension, mutual information, and algorithm stability. In light of this, despite empirical indications highlighting the critical role of GNNs' topology awareness in generalization performance [20,28,51], there remains limited understanding of the relationship between topology awareness of GNNs and their generalization. Our proposed framework addresses this gap by utilizing the distortion concept in the approximate metric embedding to connect the topology awareness of GNNs and their generalization performance.", + "bbox": [ + 212, + 349, + 789, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Active Learning in Graph", + "text_level": 1, + "bbox": [ + 215, + 541, + 473, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Active Learning (AL) seeks to enhance model performance within a limited labeling budget by identifying the most representative samples for annotation [31]. The AL methodology generally involves two phases: 1) selecting a small, preliminary subset of samples for labeling; 2) utilizing this labeled subset to iteratively discover the most informative sample for future labeling. The initial selection of an effective subset represents the cold-start problem in AL [11]. While AL has proven effective in settings with independent and identically distributed (i.i.d.) data, such as images and text, applying it to Graph Neural Networks (GNNs) poses unique challenges. These challenges stem from the necessity to account for node interdependencies and the computational complexity of iterative sample selection and retraining, especially when considering multi-hop neighbors. Recent AL methodologies for GNNs, such as AGE [5], FeatProp [45], and GraphPart [26], aim to overcome these obstacles by leveraging the graph structure and initial node features to boost efficiency and efficacy, although their success may significantly hinge on the quality of the initial features. In this work, we focus on the cold-start problem of graph active learning and introduce a novel approach (based on our analysis) that relies exclusively on graph structure, thus eliminating the reliance on feature quality at the onset of learning.", + "bbox": [ + 212, + 566, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Framework", + "text_level": 1, + "bbox": [ + 215, + 143, + 359, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed framework, which aims to investigate the intricate relationship between topology awareness and generalization performance in GNNs. The main idea of the framework is to use metric distortion to characterize the topology awareness of GNNs. With this characterization, we are able to show the relation between topology awareness of GNNs and their generalization performance.", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 GNN Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 290, + 421, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{G} = (\\mathcal{V},\\mathcal{E},\\mathbf{X})$ be an undirected graph where $\\mathcal{V} = \\{1,2,\\dots,n\\}$ is the set of $n$ nodes and $\\mathcal{E}\\subseteq \\mathcal{V}\\times \\mathcal{V}$ is the set of edges. Let $\\mathbf{X} = \\{x_{1},x_{2},\\ldots ,x_{n}\\}$ denote the set of node attribute vectors corresponding to all the nodes in $\\mathcal{V}$ . More specifically, $x_{v}\\in \\mathbf{X}$ is an m-dimensional vector which captures the attribute values of node $v\\in \\mathcal{V}$ . In addition, each vertex $v\\in \\mathcal{V}$ is associated with a node label $y_{v}\\in \\{1,\\dots,c\\}$ , where $c$ is the total number of classes.", + "bbox": [ + 212, + 316, + 787, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For our analysis, we focus on the transductive node classification setting [50], where the graph $\\mathcal{G}$ are observed prior to learning and labels are only available for a small subset of nodes $\\mathcal{V}_0\\subset \\mathcal{V}$ . Without loss of generality, we treat $\\mathcal{G}$ as fixed throughout our analysis and the randomness comes from the labels $y$ . Given a small set of the labelled nodes, $\\mathcal{V}_0\\subset \\mathcal{V}$ , the task of node-level semi-supervised learning is to learn a classifier $f$ , such that $y_{v} = f(v)$ , from a hypothesis space $\\mathcal{F}$ and perform it on the remaining unlabeled nodes. Let $\\mathcal{L}(.,.)\\mapsto \\mathbb{R}_{+}$ be a bounded continuous loss function. We denote by $R^{\\mathcal{L}}(f,\\mathcal{V})$ the generalization risk of the classifier $f$ with respect to $\\mathcal{L}$ and $\\mathcal{V}$ and it is defined as follows,", + "bbox": [ + 212, + 407, + 787, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {\\mathcal {L}} (f, \\mathcal {V}) = \\mathbb {E} _ {v \\sim \\mathcal {V}} [ \\mathcal {L} (f (v), y _ {v}) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 555, + 784, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our case, the hypothesis space is captured by a GNN and an additional read-out function. GNNs are commonly used as an embedding function and adopt a neighbourhood aggregation scheme, where the representation of a node is learnt iteratively by aggregating representations of its neighbours. After $k$ iterations of aggregation, a node's representation captures the information within its $k$ -hop neighbourhood in the graph. Formally, the $k$ -th layer of a GNN can be written as follows,", + "bbox": [ + 212, + 584, + 787, + 688 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {u} ^ {(k)} = \\operatorname {U p d a t e} ^ {(k)} \\left(\\operatorname {A g g r e g a t e} ^ {(k)} \\left(\\left\\{h _ {v} ^ {(k - 1)}, \\forall v \\in \\mathcal {N} (u) \\cup \\{u \\} \\right\\}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 700, + 784, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_u^{(k)}$ is the representation vector of node u at the k-th layer. $h_u^{(0)}$ is initialized to be $x_u$ , and $\\mathcal{N}(u)$ is the set of nodes directly adjacent to u. Different GNNs are compositions of aggregation and update layers of different designs. A GNN is commonly interpreted as an embedding function that combines the graph structure and node features to learn a representation vector $h_v$ of node v. Then the learnt representation is fed into a task-specific read-out function g (commonly a multilayer perceptron) for the prediction. $g$ maps the embedding", + "bbox": [ + 212, + 731, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vector $h_v$ to a label $y_v$ in the label set $\\mathbb{R}^c$ . The overall GNN prediction process can be viewed as a composition of $g \\circ \\mathcal{M}: \\mathcal{V} \\mapsto \\mathbb{R}^c$ . The overall GNN prediction process can be viewed as a composition of $g \\circ \\mathcal{M}: \\mathcal{V} \\mapsto \\mathbb{R}^c$ .", + "bbox": [ + 212, + 146, + 787, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Structural Subgroup", + "text_level": 1, + "bbox": [ + 215, + 215, + 433, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To study the generalization performance of GNNs on different structural subgroups, we assume the existence of a generalized distance (structural measure) function, $d_{s}(u,v)$ , that captures the similarity between two nodes $u,v$ with respect to structure $s$ . Furthermore, we assume that this structural measure function $d_{s}(u,v)$ and the vertex set $\\mathcal{V}$ form a metric space. A metric space is a fundamental mathematical object that allows for studying and measuring the relation among the elements in a given set, and it is formally defined as follows.", + "bbox": [ + 212, + 242, + 787, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 1 (metric space). A metric space is a pair $(\\mathcal{Q},d)$ , where $\\mathcal{Q}$ is a set and $d:\\mathcal{Q}\\times \\mathcal{Q}\\mapsto \\mathbb{R}$ is a metric satisfying the following axioms ( $x,y,z$ are arbitrary points of $\\mathcal{Q}$ ):", + "bbox": [ + 214, + 358, + 789, + 405 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} (M 1) d (x, y) \\geq 0, \\\\ (M 2) d (x, x) = 0, \\\\ (M 3) d (x, y) > 0, x \\neq y, \\\\ (M 4) d (x, y) = d (y, x), \\\\ (M 5) d (x, y) + d (y, z) \\geq d (x, z). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 417, + 614, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These properties ensure that the metric $d(.)$ can be used to quantify the distance or dissimilarity between any two elements in the set $\\mathcal{Q}$ . In the context of graph learning, the metric space $(\\mathcal{V},d_s)$ allows us to study the similarity between nodes in terms of the structural property $s$ and measure the structural distance between different subgroups of the test set. One natural and commonly used example in the graph is the metric space formed by the vertex set and the shortest-path distance.", + "bbox": [ + 212, + 521, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 2 (Structural Group Distance). Given a metric space $(\\mathcal{V}, d_s)$ with respect to structure $s$ and two distinct sets (groups) $\\mathcal{V}_1, \\mathcal{V}_2 \\subset \\mathcal{V}$ , we define the structural distance $D_s$ between an element $v \\in \\mathcal{V}$ to a group as", + "bbox": [ + 214, + 638, + 787, + 686 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {s} (v, \\mathcal {V} _ {1}) = \\min _ {u \\in \\mathcal {V} _ {1}} d _ {s} (v, u) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 699, + 784, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the structural distance between the two groups as", + "bbox": [ + 215, + 732, + 602, + 748 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {s} \\left(\\mathcal {V} _ {1}, \\mathcal {V} _ {2}\\right) = \\max _ {v \\in \\mathcal {V} _ {1}} \\min _ {u \\in \\mathcal {V} _ {2}} d _ {s} (v, u) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 761, + 784, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It should be noted that the distance function in Def. 2 is not symmetrical, and Eq. 4 captures the maximal structural difference of vertices in $\\mathcal{V}_1$ to $\\mathcal{V}_2$ . To provide further insight into the concept of a structural subgroup, let's", + "bbox": [ + 214, + 794, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 341, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "consider a scenario where $\\mathcal{V}_0\\subset \\mathcal{V}$ represents the training set and $d_{s}(.)$ represents the shortest-path distance. In this context, we can partition the remaining vertices $\\mathcal{V}\\backslash \\mathcal{V}_0$ into different structural subgroups based on their shortest-path distance to the nearest vertex in the training set. Specifically, we can define $\\mathcal{V}_k = \\{v\\in \\mathcal{V}\\backslash \\mathcal{V}_0|D_s(v,\\mathcal{V}_0) = k\\}$ . However, it is important to emphasize that our analysis does not assume a specific partition scheme, and the framework remains adaptable to various scenarios.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Topology Awareness", + "text_level": 1, + "bbox": [ + 214, + 271, + 429, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To characterize and formalize the topology awareness of GNNs, we introduce the concept of distortion, which measures the degree to which the graph structure is preserved in the embedding space. Mathematically, the distortion of a function between two general metric spaces is defined as,", + "bbox": [ + 212, + 292, + 785, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 3 (distortion). Given two metric spaces $(\\mathcal{Q},d)$ and $(\\mathcal{Q}',d')$ and a mapping $\\gamma :\\mathcal{Q}\\mapsto \\mathcal{Q}'$ $(\\mathcal{Q}(\\mathcal{Q}')$ is a set and $d(d^{\\prime})$ is a distance function)), $\\gamma$ is said to have a distortion $\\alpha \\geq 1$ , if there exists a constant $r > 0$ such that $\\forall u,v\\in \\mathcal{Q}$", + "bbox": [ + 214, + 359, + 789, + 421 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nr d (u, v) \\leq d ^ {\\prime} (\\gamma (u), \\gamma (v)) \\leq \\alpha r d (u, v).\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 421, + 633, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The definition above captures the maximum expansion and minimum contraction between the two metric spaces. In the extreme case when $\\alpha = 1$ , the mapping $\\gamma$ is the so-called isometric mapping that perfectly preserves the distance between any two points in the space. In other words, for any two points $x$ and $y$ in the domain of the function, their distance $d(x,y)$ in the original metric space is equal to the distance $d'(\\gamma(x), \\gamma(y))$ in the target metric space (up to a constant scaling factor). If the distortion rate of the mapping is close to 1, then this mapping is said to have a low distortion rate.", + "bbox": [ + 212, + 443, + 787, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the context of GNNs, evaluating distortion involves a straightforward comparison of embeddings in the embedding domain with the structural measures on the graph, as demonstrated in the case study. If the distortion rate between the graph domain and the embedding domain is close to 1, the resulting embedding would largely preserve the corresponding structural information from the graph space, thereby indicating high topology awareness. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. This concept of distortion is agnostic to specific structural configurations, enabling the study of diverse graph structures using appropriate distance functions. Moreover, its estimation can be readily achieved by comparing the embedding with the original structural measure, as illustrated later in the case study.", + "bbox": [ + 212, + 565, + 787, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Main Results", + "text_level": 1, + "bbox": [ + 214, + 781, + 382, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present the main results derived from our proposed framework for studying the generalization and fairness of GNNs with respect to different", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "structural groups in a graph. We start by demonstrating the relationship between the generalization performance of the model and the structural distance between the subgroups and the training set (Theorem 1). We then show that such a structural relation can induce disparities in the generalization performance across different subgroups (Theorem 2). Finally, we connect these theoretical results to the recent deep learning spline line theory [2] to provide additional insights. Detailed proof of the results presented in this paper can be found in the supplementary material.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Generalization Performance and Structural Distance", + "text_level": 1, + "bbox": [ + 215, + 292, + 694, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 1. Let $(\\mathcal{V}, d_s)$ be a metric space for structure $s$ . Let $\\mathcal{V}_0$ be a given labelled training set and $\\mathcal{V}_i \\subset \\mathcal{V} \\backslash \\mathcal{V}_0$ be an arbitrary test subgroup. Suppose $\\mathcal{M}$ is a GNN model with distortion $\\alpha$ with respect to the structural measure $d_s(:,.)$ , and $g$ and $\\mathcal{L}$ are Lipschitz, we have that", + "bbox": [ + 214, + 320, + 787, + 381 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {\\mathcal {L}} \\left(g \\circ \\mathcal {M}, \\mathcal {V} _ {i}\\right) \\leq R ^ {\\mathcal {L}} \\left(g \\circ \\mathcal {M}, \\mathcal {V} _ {0}\\right) + \\mathcal {O} \\left(\\alpha D _ {s} \\left(\\mathcal {V} _ {i}, \\mathcal {V} _ {0}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 393, + 785, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 1 states that the generalization risk of a classifier with respect to a structural subgroup depends on the empirical training loss of the classifier, the structural distance between the subgroup and the training set, and the distortion (topology awareness) of GNNs. Such a relation provides a new perspective on understanding the link between topology awareness and generalization performance in GNNs. Prior research in this area has mainly focused on the connection between GNNs and the Weisfeiler-Lehman algorithm, emphasizing the ability of GNNs to differentiate various graph structures. However, Theorem 1 highlights a direct link between topology awareness and generalization performance, and a potential issue of fairness in GNNs, as it shows that the model's performance on different structural subgroups is influenced by their distance to the training set. Next, we extend the result of Theorem 1 to analyze the relation between the generalization of different structural subgroups and their distance to the training set.", + "bbox": [ + 212, + 424, + 787, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Unfair Generalization Performance on Structural Subgroups", + "text_level": 1, + "bbox": [ + 215, + 661, + 759, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 2. Let $(\\mathcal{V}, d_s)$ be a metric space for structure $s$ . Suppose $\\mathcal{M}$ is a GNN model with small training loss on $\\mathcal{V}_0$ and distortion $\\alpha$ with respect to structural measure $d_s(:,.)$ , and $\\mathcal{L}$ and $g$ are smooth. Let $\\mathcal{V}_i, \\mathcal{V}_j \\subset \\mathcal{V} \\backslash \\mathcal{V}_0$ be two distinct subgroups in the local neighbourhood of the training set $\\mathcal{V}_0$ , then there exists constant $\\delta > 0$ (independent of the choice of subgroups) such that if", + "bbox": [ + 214, + 689, + 789, + 766 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nD _ {s} \\left(\\mathcal {V} _ {i}, \\mathcal {V} _ {0}\\right) \\geq \\alpha \\delta D _ {s} \\left(\\mathcal {V} _ {j}, \\mathcal {V} _ {0}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 779, + 599, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we have that,", + "bbox": [ + 215, + 808, + 315, + 821 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {\\mathcal {L}} (g \\circ \\mathcal {M}, \\mathcal {V} _ {i}) \\geq R ^ {\\mathcal {L}} (g \\circ \\mathcal {M}, \\mathcal {V} _ {j}).\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 823, + 617, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 341, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 2 states that when comparing two subgroups, if the difference in their structural distance to the training set is larger than the threshold $(\\delta \\alpha)$ , then GNNs exhibit unfair generalization performance with respect to these groups. Specifically, the theorem states that GNNs generalize better on subgroups that are structurally closer to the training set, which leads to accuracy disparities among different subgroups. Additionally, the theorem notes that the topology awareness of GNNs (measured by the distortion rate $\\alpha$ ) plays a key role in determining the strength of this accuracy disparity. This has important implications for the use of GNNs in critical applications, as it suggests that increased topology awareness in GNNs can lead to improved generalization performance but also a greater potential for unfairness in terms of accuracy disparities among different structural subgroups. As such, it is important to carefully consider the trade-offs between topology awareness and fairness when designing new GNNs and applying them to real-world problems.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Connection with Deep Learning Spline Theory", + "text_level": 1, + "bbox": [ + 215, + 393, + 645, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our theoretical findings reveal that when considering the performance of the model from the viewpoint of the embedding space, the training set creates a \"valley\" effect on generalization performance, with the training vertex at the bottom of the valley. Representations that are closer to the embedding of the training set perform better, while those farther away perform worse. This pattern has an interesting connection with deep learning spline theory [2]. This theory suggests that in DNNs, the learning process for classification divides the input/latent space into small circles (referred to as power diagrams), where representations within each circle are grouped, and representations closer to the centre of the circle are predicted with more confidence. Our findings suggest that in GNNs, the mapping of test data to these power diagrams is related to the structural similarity between test and training data in the graph. This connection provides further validation for our derived results.", + "bbox": [ + 212, + 431, + 488, + 839 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/81cad5d5b5b90ea693f888047329963422b316ff861f369dc6c95edcb6513a6f.jpg", + "image_caption": [ + "Fig.2. The loss landscape in embedding space induced by different training sets. When vertices $a$ and $d$ are selected as the training set instead of $b$ and $c$ , the test losses are smaller as the average distance between the training set and the other vertices is smaller (the neighbourhoods of $h_a$ and $h_d$ cover embeddings of more vertices in the embedding space)." + ], + "image_footnote": [], + "bbox": [ + 517, + 469, + 764, + 659 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 A Case Study on Shortest-Path Distance", + "text_level": 1, + "bbox": [ + 215, + 143, + 648, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we present a case study which utilizes the shortest path distance to validate the predictions and effectiveness of our proposed framework. We aim to demonstrate how our framework can be applied to a real-world scenario and how it can be used to understand the generalization performance of GNNs on different structural subgroups. The case study will provide insights and show the potential impact of our results on various critical applications of GNNs. In this case study, we aim to answer empirically the following questions: Given a GNN model that can preserve certain structure $s$ ( $\\alpha$ closed to 1), is the derived structural result observed on the generalization performance of the GNN model with respect to structural subgroups defined by $s$ ? Furthermore, how can our outcomes be harnessed to effectively tackle real-world challenges?", + "bbox": [ + 212, + 175, + 787, + 340 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overall, our case study demonstrates that our proposed framework can be applied to a real-world scenario and provides useful insights into the performance of GNNs in different structural subgroups. It also shows that the results of our framework can be generalized for different train/test pairs. Moreover, we showcase its practical utility by employing the findings to address the cold start problem within graph active learning.", + "bbox": [ + 212, + 342, + 787, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "General Setup. We use Deep Graph Library [43] for our implementation and experiment with four representative GNN models, GCN [17], Sage [13], GAT [42] and GCNII [7]. In all experiments, we follow closely the common GNN experimental setup, and multiple independent trials are conducted for each experiment to obtain average results. We adopt five popular and publicly available node classification datasets: Cora [34], Citeseer [21], CoraFull [4], CoauthorCS [36] and Ogbn-arxiv [16]. Due to space limitations, we only show part of the experimental results in the main body of the paper. Detailed descriptions of the datasets, hyperparameters and remaining experimental results can be found in the appendix.", + "bbox": [ + 212, + 431, + 787, + 582 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Structural Subgroups. We examine the accuracy disparity of subgroups defined by the shortest path distance (spd). In order to directly investigate the effect of structural distance on the generalization bound (Theorem 1), we first split the test nodes into subgroups by their shortest-path distance to the training set. For each test node $v$ , its distance to the training set $\\mathcal{V}_0$ is given by $d_v = D_{\\mathrm{spd}}(v, \\mathcal{V}_0)$ with the shortest path distance metric as defined in Def. 2. Then we group the test nodes according to this distance.", + "bbox": [ + 212, + 583, + 787, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Distance Awareness of the Selected GNNs. We first verify if the selected GNNs preserve (are aware of) graph distances (with a small distortion rate), by evaluating the relation between the graph distances of test vertices to the training set and the embedding distances (Euclidean distance) of vertex representations to the representations of the training set, i.e., the relation between $d_{\\mathrm{spd}}(v,\\mathcal{V}_0)$ and $d_{\\mathrm{Euclidean}}(\\mathcal{M}(v),\\mathcal{M}(\\mathcal{V}_0))$ . In this experiment, we train the GNN models on the default training set provided in the dataset. We extract vertices within the 5-hop neighbourhood of the training set and group the vertices based on their distances to the training set. 5-hop neighbourhood is chosen because the diameter of most real-life networks is around 5 and it gives a better visual", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "presentation. We then compute the average embedding distance of each group to the training set in the embedding space. Based on Def. 3, if an embedding has low distortion, we should expect to see linear relation with the slope being the scaling factor. As shown in Fig. 3, there exists a strong correlation between the graph distance and the embedding distance. The relation is near linear and the relative orders of distances are mostly preserved within the first few hops. This implies that the distortion $\\alpha$ is small with the selected GNNs, i.e., they are distance aware. It also indicates that the premise of the small distortion in our theoretical results can be commonly satisfied in practice.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Graph Distance and Generalization Performance", + "text_level": 1, + "bbox": [ + 214, + 303, + 665, + 319 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Next, we study the relation between graph distance and generalization performance following the same settings as in the previous experiment. We compute the average prediction accuracy in each group of vertices. In Fig. 4, '0 hop' indicates the training set. We observe that the selected GNNs (with low distortion) indeed tend to generalize better on the group of vertices that are closer to the training set, which validates Theorem 2.", + "bbox": [ + 212, + 327, + 787, + 416 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further validate if the distance between different pairs of the training sets and the test group is related to the generalization performance of GNNs in the test group. We randomly sample $k = 5$ vertices from each class as the training set $\\mathcal{V}_0$ . Then we evaluate the GNN model on the rest of the vertices $\\mathcal{V} \\backslash \\mathcal{V}_0$ . We make $k$ relatively small to enable a larger variance (among different trials) in the distance $D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)$ (as given in Def. 2) between the training set and the rest of the vertices. We conduct 30 independent trials each selecting a different training set and ranking all the trials in descending order of the mean graph distance between the training set and the rest of the vertices. In Table 1, Group 1 contains the 10 trials with the largest graph distance $D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)$ and Group 3 includes the 10 trials with the smallest graph distance $D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)$ . The average classification accuracy and the variance (value in the bracket) are computed for each group.", + "bbox": [ + 212, + 417, + 787, + 613 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Table 1, training GNNs on a training set with a smaller graph distance $D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)$ to the rest of the vertices, does generalize better, which validates our result and further enables us to use these insights for designing and developing new GNNs algorithms.", + "bbox": [ + 212, + 614, + 787, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Application in Cold Start Problem for Graph Active Learning", + "text_level": 1, + "bbox": [ + 214, + 695, + 774, + 710 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Active learning is a type of machine learning where the model is actively involved in the labelling process of the training data. In active learning, the model is initially trained on a small labelled dataset and then used to select the most informative examples from the unlabeled dataset to be labelled next. Unsatisfactory initial set selection can lead to slow learning progress and unfair performance, which is referred to as the cold start problem [11]. Effective resolution of this issue entails an initial set selection that adheres to the following criteria: 1) exclusively leveraging graph structure; 2) achieving robust generalization; and", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/946e531828fad46dc27183d8a11b403593b79f2805de9fbd550ec7b2cf3ac5a0.jpg", + "image_caption": [ + "(a) GCN, Cora" + ], + "image_footnote": [], + "bbox": [ + 220, + 159, + 343, + 239 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8738f1566eeb92e9f5718d510477bcea0c7251885d9d7280a45710018f9a40ca.jpg", + "image_caption": [ + "(b) Sage, Cora" + ], + "image_footnote": [], + "bbox": [ + 362, + 159, + 483, + 241 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a835381639d7517d35d2c14b2fa3da45149337efad5d434e0ca13a663ec4f0f7.jpg", + "image_caption": [ + "(c) GAT, Cora" + ], + "image_footnote": [], + "bbox": [ + 504, + 159, + 625, + 241 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4e9209739cca6e1a6356318383c88c548519cb90f423c12b3b4df02b07ded9cc.jpg", + "image_caption": [ + "(d) GCN2, Cora" + ], + "image_footnote": [], + "bbox": [ + 647, + 159, + 767, + 241 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6ce5e3918f3f657d548b573cf27a209d6ee92500da964e9b3006c2f6c6cafea7.jpg", + "image_caption": [ + "(e) GCN, Coauthor CS" + ], + "image_footnote": [], + "bbox": [ + 220, + 276, + 341, + 358 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/123f5f88dfabbbc933b0af2f1b8ab30efae89afa73bd4d3db9f1b12607ef3676.jpg", + "image_caption": [ + "(f) Sage, Coauthor CS" + ], + "image_footnote": [], + "bbox": [ + 362, + 277, + 483, + 358 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f426896bffeba2d89cdf89c9722fe1e4ed86383f4ebb82342c9acf88f00fef2a.jpg", + "image_caption": [ + "(g) GAT, Coauthor \nCS" + ], + "image_footnote": [], + "bbox": [ + 504, + 277, + 625, + 358 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/72b0fc676d728179f45ccd341544634676249e2a6b6695bb5a40ce957c28cd59.jpg", + "image_caption": [ + "(h) GCNII, Coauthor CS" + ], + "image_footnote": [], + "bbox": [ + 647, + 277, + 767, + 358 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1cd4281aedfa3b7694dcc00e692f29b10c2808b28d514f55e47017674cebcf2d.jpg", + "image_caption": [ + "Fig. 3. Graph distance vs. embedding distance. (a)-(d) are the relation between graph distance (shortest-path distance) and embedding distance (Euclidean) of the four representative GNNs on Cora. (e)-(h) are the relation between graph distance and embedding distance of the four representative GNNs on Coauthor CS. The highlighted regions represent the $95\\%$ confidence interval for that regression. (a)-(h) all demonstrate a rather linear relation between graph distance and embedding distance.", + "(a) Cora", + "Fig.4. Graph distance vs. accuracy. (a) is the relation between graph distance (shortest-path distance) and the accuracy of four representative GNNs on Cora. (b) is the relation between graph distance and the accuracy of four representative GNNs on Coauthor CS. Both (a) and (b) show that the graph distance is indeed a good indicator for generalization performance when GNNs are distance aware." + ], + "image_footnote": [], + "bbox": [ + 241, + 516, + 493, + 614 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fdf0a3bbf809c30410ce869893d47eb38ad60924425fa770cc533a82bcdcf6c2.jpg", + "image_caption": [ + "(b) Coauthor CS" + ], + "image_footnote": [], + "bbox": [ + 511, + 516, + 763, + 614 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3) ensuring fairness (with minimal discrepancy in generalization performance) across distinct subgroups. We illustrate how our derived findings can be instrumental in the selection of an effective initial labelled dataset.", + "bbox": [ + 212, + 750, + 784, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our empirical results show that the selected GNNs are highly distance-aware, meaning that their generalization performance is closely related to the coverage of the training set. To ensure that the model is fair and effective, the initial", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d6b2a12bdd7827705ee4621c6f13fc147917d7fa0d2ebc84d4eea1933aea046d.jpg", + "table_caption": [ + "Table 1. Accuracy of 30 independent trials with different train/test pairs grouped by $D_{s}(\\mathcal{V}\\backslash \\mathcal{V}_{0},\\mathcal{V}_{0})$ . Group 1 contains the 10 trials with the largest graph distance $D_{\\mathrm{spd}}(\\mathcal{V}\\backslash \\mathcal{V}_0,\\mathcal{V}_0)$ and Group 3 includes the 10 trials with the smallest graph distance $D_{\\mathrm{spd}}(\\mathcal{V}\\backslash \\mathcal{V}_0,\\mathcal{V}_0)$ . The average accuracy and variance of the 10 trials are recorded in the table. The results show that train/test pairs with closer structural similarity indeed have better generalization performance, as predicted by our theoretical results." + ], + "table_footnote": [], + "table_body": "
CoraCiteseer
GCNGATSageGCNIIGCNGATSageGCNII
Group 165.42 (5.03)68.03 (4.62)67.61 (5.02)61.42 (4.03)59.72 (5.41)56.78 (4.73)58.52 (4.85)55.46 (3.69)
Group 267.56 (4.23)70.94 (3.86)69.54 (4.84)66.42 (3.76)61.48 (5.15)60.33 (4.24)60.78 (3.60)58.42 (4.08)
Group 371.40 (3.59)73.52 (3.53)72.60 (4.23)72.18 (4.01)62.93 (4.14)64.63 (3.65)63.03 (4.36)63.23 (5.06)
", + "bbox": [ + 222, + 237, + 785, + 290 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7265c3fc318d258418fbc93b45e956af0d999ec318db87980d59fae4950fab48.jpg", + "table_caption": [ + "Table 2. Initial data labelling: performance comparison of different GNNs sampling strategies based on graph structure. Results are averaged among five trials. For each dataset, we select $0.5\\%$ of the total data as the initial set and evaluate on the rest of the vertices. Each entry corresponds to \"test accuracy (ACC)| maximum discrepancy (MD)\". $\\uparrow$ means the higher the better and $\\downarrow$ means the lower the better. Bold letter indicates the best sampling strategy for the given dataset and GNN model." + ], + "table_footnote": [], + "table_body": "
GCNSage
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)
R37.08|13.2336.35|10.238.13|3.2357.75|16.2337.69|12.0337.24|11.208.56|4.2049.72|15.52
Degree40.87|12.5236.27|9.2810.68|2.3251.50|18.2338.04|11.8735.26|12.0511.56|3.0242.78|13.02
Centrality42.75|10.4138.45|8.159.49|2.0853.48|16.1142.18|10.1540.81|9.0512.56|3.0845.65|10.11
PageRank47.87|9.5242.27|8.2512.48|2.1055.50|15.2350.15|9.1443.26|9.1513.15|2.9855.16|10.74
OURS49.03|8.2350.98|7.3215.25|1.2361.75|14.3258.57|8.2948.80|7.0314.76|2.0161.63|8.82
GATGCNII
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)|MD(↓)ACC(↑)|MD(↓)
R49.36|15.7843.53|16.279.55|3.0743.25|15.2031.53|14.0237.29|17.3210.98|3.0652.61|17.02
Degree54.08|13.5244.65|15.1710.50|2.7847.30|13.0432.19|13.8638.88|16.9810.40|2.7954.70|15.96
Centrality56.81|11.5248.15|13.0811.15|2.5248.50|13.1832.89|12.5739.98|16.0512.48|2.5256.88|15.18
PageRank60.91|11.0849.77|12.0712.94|2.1150.50|12.2334.04|12.0842.77|13.9514.56|2.1958.78|14.08
OURS62.76|10.5451.11|11.0713.39|1.8851.21|11.0235.61|11.0244.50|13.0415.19|1.9660.76|13.18
", + "bbox": [ + 222, + 400, + 785, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "labelled set should consist of vertices that are in close proximity to the rest of the input graph (in graph distance), as illustrated in Fig. 2. This can be formulated into the following optimization problem:", + "bbox": [ + 212, + 569, + 787, + 617 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\mathcal {V} _ {0} \\subset \\mathcal {V}} D _ {\\operatorname {s p d}} (\\mathcal {V} \\backslash \\mathcal {V} _ {0}, \\mathcal {V} _ {0}) \\\\ s. t. \\quad | \\mathcal {V} _ {0} | = k \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 631, + 575, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where $k$ is the given initial set size and $D_{s}(.)$ is given in Def. 2. Intuitively, Eq.6 aims to select a subset of vertices that exhibit the closest structural similarity (as defined by the graph distance in this case) to the remaining vertices. This problem corresponds to the well-known k-center problem in graph theory [14], which is NP-hard. While finding the optimal solution may be intractable, efficient heuristics exist for tackling Eq. 6. Exploring the best heuristic is beyond the scope of this paper. Instead, for our experiment, we adopt the classic farthest-first traversal algorithm to obtain a solution. We refer to this sample selection procedure as the coverage-based sampling algorithm and provide a detailed description of the method used in the experiment within the supplementary material.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 516, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We evaluate the initial set selected by our coverage-based sampling algorithm, comparing it to the existing baselines: (1) $\\mathbf{R}$ : uniform random sampling [30], and importance sampling with graph properties (2) Degree [15], (3) Centrality, (4) Density and (5) PageRank [5]. We apply each strategy to select an initial dataset of $0.5\\%$ of the total data for labelling (training). We evaluate the performance of the learnt model with the rest of the vertices on two metrics: 1) classification accuracy, which reflects the generalization performance, and 2) the maximum discrepancy among structural groups [1,8], which reflects the fairness. Let $R_{i}$ denote the mean classification accuracy of the vertex group that is $i$ -hop away from $\\mathcal{V}_0$ . Then, the maximum discrepancy is computed as $\\max \\{|R_i - R_j||i,j\\in 1,\\dots,r\\}$ , where $r$ is the maximum hop number. We use $r = 5$ to keep consistent with other experiments. The results in Table 2 show that coverage-based sampling achieves a significant improvement in model performance and fairness over the other strategies across different datasets, giving a better kick-start to the active learning process of GNNs.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 392, + 509, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we study and investigate GNNs which are fundamental to many computer vision and machine learning tasks. In particular, we explore the relationship between the topology awareness and generalization performance of GNNs. We introduce a novel framework that connects the structural awareness of GNNs with approximate metric embedding, offering a fresh perspective on their generalization capabilities in semi-supervised node classification tasks. This structure-agnostic framework facilitates a deeper, more intuitive comprehension of GNN generalizability across varied graph structures. Through a case study centered on graph distance, we demonstrate that our theoretical findings regarding structural results are reflected in the practical generalization performance of GNNs. Additionally, our findings shed light on the cold start problem in graph active learning and could influence various significant GNN applications.", + "bbox": [ + 212, + 421, + 787, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.1 Limitation and Future Works", + "text_level": 1, + "bbox": [ + 215, + 621, + 504, + 636 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our proposed framework introduces a novel angle on GNN generalization performance, yet it is not without limitations. The current approach interprets GNN embeddings in terms of metric mapping and describes the structure awareness of GNNs through distortion. This perspective, however, overlooks the specific dynamics that contribute to reduced distortion within GNNs. Investigating these underlying dynamics would significantly enrich this area of research. Moreover, our study primarily examines the transductive setting. Expanding this analysis to the inductive setting—where the model is trained on one graph but applied to another—would offer a more comprehensive view of GNN generalization capabilities.", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. We would like to thank the anonymous reviewers for their helpful comments. JS and CW are supported by grants from Hong Kong RGC under the contracts HKU 17207621, 17203522 and C7004-22G (CRF).", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ali, J., Babaei, M., Chakraborty, A., Mirzasoleiman, B., Gummadi, K.P., Singla, A.: On the fairness of time-critical influence maximization in social networks. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 1541-1542. IEEE (2022)", + "2. Balestriero, R., richard baraniuk: A spline theory of deep learning. In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 374-383. PMLR (10-15 Jul 2018), https://proceedings.mlr.press/v80/balestriero18b.htm1", + "3. Baranwal, A., Fountoulakis, K., Jagannath, A.: Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization. arXiv preprint arXiv:2102.06966 (2021)", + "4. Bojchevski, A., Gunnemann, S.: Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking (2018)", + "5. Cai, H., Zheng, V.W., Chang, K.C.C.: Active learning for graph embedding. arXiv preprint arXiv:1705.05085 (2017)", + "6. Chen, J., Zhu, J., Song, L.: Stochastic training of graph convolutional networks with variance reduction. arXiv preprint arXiv:1710.10568 (2017)", + "7. Chen, M., Wei, Z., Huang, Z., Ding, B., Li, Y.: Simple and deep graph convolutional networks (2020)", + "8. Dong, Y., Ma, J., Chen, C., Li, J.: Fairness in graph mining: A survey (2022). https://doi.org/10.48550/ARXIV.2204.09888, https://arxiv.org/abs/2204.09888", + "9. Esser, P., Chennuru Vankadara, L., Ghoshdastidar, D.: Learning theory can (sometimes) explain generalisation in graph neural networks. Advances in Neural Information Processing Systems 34, 27043-27056 (2021)", + "0. Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for quantum chemistry (2017)", + "11. Grimova, N., Macas, M., Gerla, V.: Addressing the cold start problem in active learning approach used for semi-automated sleep stages classification. In: 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). pp. 2249–2253. IEEE (2018)", + "12. Hamilton, W.L.: Graph representation learning. Synthesis Lectures on Artificial Intelligence and Machine Learning 14(3), 1-159 (2020)", + "13. Hamilton, W.L., Ying, R., Leskovec, J.: Inductive representation learning on large graphs (2018)", + "14. Hochbaum, D.S.: Approximating covering and packing problems: set cover, vertex cover, independent set, and related problems. In: Approximation algorithms for NP-hard problems, pp. 94-143 (1996)", + "15. Hu, S., Xiong, Z., Qu, M., Yuan, X., Côté, M.A., Liu, Z., Tang, J.: Graph policy network for transferable active learning on graphs. Advances in Neural Information Processing Systems 33, 10174-10185 (2020)", + "16. Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., Leskovec, J.: Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687 (2020)", + "17. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks (2017)" + ], + "bbox": [ + 218, + 181, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Lao, D., Yang, X., Wu, Q., Yan, J.: Variational inference for training graph neural networks in low-data regime through joint structure-label estimation. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 824-834 (2022)", + "19. Leman, A., Weisfeiler, B.: A reduction of a graph to a canonical form and an algebra arising during this reduction. Nauchno-Technicheskaya Informatiya 2(9), 12-16 (1968)", + "20. Li, P., Wang, Y., Wang, H., Leskovec, J.: Distance encoding: Design provably more powerful neural networks for graph representation learning (2020)", + "21. Li, Q., Han, Z., Wu, X.M.: Deeper insights into graph convolutional networks for semi-supervised learning. In: Thirty-Second AAAI conference on artificial intelligence (2018)", + "22. Li, Y., Yu, R., Shahabi, C., Liu, Y.: Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926 (2017)", + "23. Liao, R., Urtasun, R., Zemel, R.: A {pac}-bayesian approach to generalization bounds for graph neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=TR-Nj6nFx42", + "24. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors (2016)", + "25. Ma, J., Deng, J., Mei, Q.: Subgroup generalization and fairness of graph neural networks. arXiv preprint arXiv:2106.15535 (2021)", + "26. Ma, J., Ma, Z., Chai, J., Mei, Q.: Partition-based active learning for graph neural networks. arXiv preprint arXiv:2201.09391 (2022)", + "27. Morris, C., Ritzert, M., Fey, M., Hamilton, W.L., Lenssen, J.E., Rattan, G., Grohe, M.: Weisfeiler and leman go neural: Higher-order graph neural networks (2020)", + "28. Nishad, S., Agarwal, S., Bhattacharya, A., Ranu, S.: Graphreach: Position-aware graph neural network using reachability estimations (2021)", + "29. Oono, K., Suzuki, T.: Graph neural networks exponentially lose expressive power for node classification. arXiv preprint arXiv:1905.10947 (2019)", + "30. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Chen, X., Wang, X.: A survey of deep active learning (2020)", + "31. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021)", + "32. Sato, R.: A survey on the expressive power of graph neural networks. arXiv preprint arXiv:2003.04078 (2020)", + "33. Scarselli, F., Tsoi, A.C., Hagenbuchner, M.: The vapnik-chervonenkis dimension of graph and recursive neural networks. Neural Networks 108, 248-259 (2018)", + "34. Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., Eliassi-Rad, T.: Collective classification in network data. AI magazine 29(3), 93-93 (2008)", + "35. Sheng, G., Su, J., Huang, C., Wu, C.: Mspipe: Efficient temporal gnn training via staleness-aware pipeline. arXiv preprint arXiv:2402.15113 (2024)", + "36. Sinha, A., Shen, Z., Song, Y., Ma, H., Eide, D., Hsu, B.J., Wang, K.: An overview of microsoft academic service (mas) and applications. In: Proceedings of the 24th international conference on world wide web. pp. 243-246 (2015)", + "37. Su, J., Marbach, P.: Structure of core-periphery communities. In: International Conference on Complex Networks and Their Applications. pp. 151-161. Springer (2022)", + "38. Su, J., Marbach, P.: Structural properties of corea $\\mathbf{E}$ periphery communities. Advances in Complex Systems (ACS) 26(06), 1-43 (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J.Su et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Su, J., Zou, D., Wu, C.: Pres: Toward scalable memory-based dynamic graph neural networks (2024), https://arxiv.org/abs/2402.04284", + "40. Su, J., Zou, D., Zhang, Z., Wu, C.: Towards robust graph incremental learning on evolving graphs. In: International Conference on Machine Learning. pp. 32728-32748. PMLR (2023)", + "41. Tsubaki, M., Tomii, K., Sese, J.: Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics 35(2), 309-318 (2019)", + "42. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)", + "43. Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y., et al.: Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315 (2019)", + "44. Wu, Q., Zhang, H., Yan, J., Wipf, D.: Handling distribution shifts on graphs: An invariance perspective (2022)", + "45. Wu, Y., Xu, Y., Singh, A., Yang, Y., Dubrawski, A.: Active learning for graph neural networks via node feature propagation. arXiv preprint arXiv:1910.07567 (2019)", + "46. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing (2017)", + "47. Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 (2018)", + "48. Yang, C., Wu, Q., Wang, J., Yan, J.: Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps. arXiv preprint arXiv:2212.09034 (2022)", + "49. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation (2022)", + "50. Yang, Z., Cohen, W.W., Salakhutdinov, R.: Revisiting semi-supervised learning with graph embeddings (2016). https://doi.org/10.48550/ARXIV.1603.08861, https://arxiv.org/abs/1603.08861", + "51. You, J., Ying, R., Leskovec, J.: Position-aware graph neural networks (2019)", + "52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context (2018)", + "53. Zhou, K., Huang, X., Li, Y., Zha, D., Chen, R., Hu, X.: Towards deeper graph neural networks with differentiable group normalization (2020)", + "54. Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Feng, M., Zhao, X., Miao, Q., Shah, S.A.A., Bennamoun, M.: Scene graph generation: A comprehensive survey (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 672 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Topology Awareness Framework", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_model.json b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5bd44d5dce59df2a2a7327924d32eb8898019587 --- /dev/null +++ b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_model.json @@ -0,0 +1,2324 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.231, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "On the Topology Awareness and Generalization Performance of Graph Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.213, + 0.601, + 0.227 + ], + "angle": 0, + "content": "Junwei Su and Chuan Wu†" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.24, + 0.714, + 0.269 + ], + "angle": 0, + "content": "Department of Computer Science, the University of Hong Kong {jwsu, cwu} @cs.hku.hk" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.305, + 0.749, + 0.639 + ], + "angle": 0, + "content": "Abstract. Many computer vision and machine learning problems are modelled as learning tasks on graphs, where graph neural networks (GNNs) have emerged as a dominant tool for learning representations of graph-structured data. A key feature of GNNs is their use of graph structures as input, enabling them to exploit the graphs' inherent topological properties—known as the topology awareness of GNNs. Despite the empirical successes of GNNs, the influence of topology awareness on generalization performance remains unexplored, particularly for node-level tasks that diverge from the assumption of data being independent and identically distributed (I.I.D.). The precise definition and characterization of the topology awareness of GNNs, especially concerning different topological features, are still unclear. This paper introduces a comprehensive framework to characterize the topology awareness of GNNs across any topological feature. Using this framework, we investigate the effects of topology awareness on GNN generalization performance. Contrary to the prevailing belief that enhancing the topology awareness of GNNs is always advantageous, our analysis reveals a critical insight: improving the topology awareness of GNNs may inadvertently lead to unfair generalization across structural groups, which might not be desired in some scenarios. Additionally, we conduct a case study using the intrinsic graph metric, the shortest-path distance, on various benchmark datasets. The empirical results of this case study confirm our theoretical insights. Moreover, we demonstrate the practical applicability of our framework by using it to tackle the cold start problem in graph active learning." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.664, + 0.377, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Many problems in computer vision and machine learning are modeled as learning tasks on graphs. For example, in semantic segmentation, graphs model the relationships between different image regions, enhancing accuracy and context-aware segmentation. Graph neural networks (GNNs) have emerged as a dominant class of machine learning models specifically designed for learning representations of graph-structured data. They have demonstrated considerable success in addressing a wide range of graph-related problems in various domains such as chemistry [10], biology [41], social networking [6,22,35], scene graph generation [49,54]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.401, + 0.84 + ], + "angle": 0, + "content": "†: corresponding authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "and visual relationship detection [24,46,52]. A defining characteristic of GNNs is their use of a spatial approach through message passing on the graph structure for feature aggregation. This enables GNNs to preserve structural information or dependencies (referred to as topology awareness) from the underlying graph structure, allowing them to be highly effective in tasks such as node classification. Fig. 1 illustrates the overall learning process of GNNs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.44 + ], + "angle": 0, + "content": "Despite their practicality and potential, there remains a lack of theoretical understanding about GNNs, particularly in the semi-supervised node classification setting where the dependencies among the data differ significantly from other machine learning models [25]. In this setting, the goal is to leverage relations, as captured by the graph structure, among the data and a small set of labelled nodes to predict labels for the remaining nodes. Most of the existing theoretical studies of GNNs have focused on the connection between the message-passing mechanism of GNNs and the Weisfeiler-Lehman isomorphism test [19], aiming to understand the ability of GNNs to differentiate different graph structures in the learned representations, known as the expressive power of GNNs. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [28,32,51]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.73 + ], + "angle": 0, + "content": "However, as GNNs become more reliant on and sensitive (aware) of graph structure as input, they may exhibit different generalization performances towards certain structural subgroups (distinct data subsets grouped by structural similarity to the training set) within the data. The quantification of GNN generalization across distinct structural subgroups is termed structural subgroup generalization [25]. Such considerations are vital in GNN application and development. For instance, within protein-protein interaction networks, these structural subgroups could represent different molecular complexes, influencing the accuracy of interaction predictions. Similarly, understanding how the topology awareness of GNNs influences generalization is essential when devising sampling strategies for training. The extent to which the generalization performance of GNNs is influenced by specific structural features of graph data is critical in deciding the composition of training datasets. Despite its importance, an understanding of the relationship between the topology awareness of GNNs and its structural subgroup generalization is still lacking. Furthermore, characterizing the topology awareness of GNNs poses a challenge, especially considering that different domains and tasks may prioritize distinct structural aspects. Therefore, a versatile framework is needed to assess the topology awareness of GNNs in relation to various structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To address this gap, in this paper, we propose a novel framework based on approximate metric embedding to study the relationship between structural subgroup generalization and topology awareness of GNNs in the context of semi-supervised node classification. The proposed framework allows for the investigation of the structural subgroup generalization of GNNs with respect to different structural subgroups. More concretely, the main contributions of this work are summarized as follows." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.782, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Fig. 1. An illustration of the learning process in a 2-layer GNN. The message-passing mechanism leverages the graph structure to aggregate information, thereby generating the representation/embedding \\( h_a \\) for the target vertex \\( a \\) (highlighted in red)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.357, + 0.784, + 0.446 + ], + "angle": 0, + "content": "1. We propose a novel, structure-agnostic framework using approximate metric embedding to examine the interplay between GNNs' structural subgroup generalization and topology awareness. This framework is versatile, accommodating various structural measures like shortest-path distance, and requires only the corresponding structural measure. Its simplicity in estimating key factors makes it applicable and generalizable to a wide range of scenarios." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.448, + 0.785, + 0.598 + ], + "angle": 0, + "content": "2. Through formal analysis within our framework, we establish a clear link between GNN topology awareness and their generalization performance (Theorem 1). We also demonstrate that while enhanced topology awareness boosts GNN expressiveness, it can result in uneven generalization performance, favouring subgroups more structurally similar to the training set (Theorem 2). Such structural property can be harmful (causing unfairness issues) or useful (informing design decisions) depending on the scenario. This challenges the prevailing belief that increased topology awareness universally benefits GNNs [28, 32, 51], emphasizing the importance of considering the relationship between topology awareness and generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.599, + 0.785, + 0.704 + ], + "angle": 0, + "content": "3. We validate our framework through a case study on shortest-path distance, highlighting its practicality and relevance. The results corroborate our theoretical findings, showing that GNNs with heightened awareness of shortest-path distances excel in classifying vertex groups closer to the training set. Moreover, we demonstrate how our findings can be applied to mitigate the cold start problem in graph active learning [11,15], highlighting the practical implications of our framework and results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.737, + 0.388, + 0.753 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.769, + 0.509, + 0.785 + ], + "angle": 0, + "content": "2.1 Topology Awareness of GNNs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Modern GNN architectures take inspiration from the Weisfeiler-Lehman isomorphism test [12, 19], which employs the graph structure to propagate information. Consequently, a significant portion of existing literature on GNNs concentrates" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "on understanding their ability to differentiate various graph structures, known as the expressive power of GNNs [27-29, 47, 51, 53]. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties [37, 38] in the learned representation [28, 32, 51]. Nevertheless, the precise ramifications of heightened topology awareness on GNNs' generalization performance remain shrouded in ambiguity. The intricate variations within structures across different domains and tasks further compound this complexity, underscoring the necessity for a unified framework—akin to the one proposed in this paper—to shed light on these intricacies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.462, + 0.337 + ], + "angle": 0, + "content": "2.2 Generalization of GNNs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.79, + 0.518 + ], + "angle": 0, + "content": "Relatively few studies focus on understanding the generalization performance of GNNs in node-level tasks. [3,9,18,23,25,33,39,40,44,48] extend the conventional statistical learning or information-theoretic framework to GNNs, providing different generalization bounds based on concepts such as Vapnik-Chervonenkis dimension, mutual information, and algorithm stability. In light of this, despite empirical indications highlighting the critical role of GNNs' topology awareness in generalization performance [20,28,51], there remains limited understanding of the relationship between topology awareness of GNNs and their generalization. Our proposed framework addresses this gap by utilizing the distortion concept in the approximate metric embedding to connect the topology awareness of GNNs and their generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.475, + 0.558 + ], + "angle": 0, + "content": "2.3 Active Learning in Graph" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Active Learning (AL) seeks to enhance model performance within a limited labeling budget by identifying the most representative samples for annotation [31]. The AL methodology generally involves two phases: 1) selecting a small, preliminary subset of samples for labeling; 2) utilizing this labeled subset to iteratively discover the most informative sample for future labeling. The initial selection of an effective subset represents the cold-start problem in AL [11]. While AL has proven effective in settings with independent and identically distributed (i.i.d.) data, such as images and text, applying it to Graph Neural Networks (GNNs) poses unique challenges. These challenges stem from the necessity to account for node interdependencies and the computational complexity of iterative sample selection and retraining, especially when considering multi-hop neighbors. Recent AL methodologies for GNNs, such as AGE [5], FeatProp [45], and GraphPart [26], aim to overcome these obstacles by leveraging the graph structure and initial node features to boost efficiency and efficacy, although their success may significantly hinge on the quality of the initial features. In this work, we focus on the cold-start problem of graph active learning and introduce a novel approach (based on our analysis) that relies exclusively on graph structure, thus eliminating the reliance on feature quality at the onset of learning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.361, + 0.161 + ], + "angle": 0, + "content": "3 Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.269 + ], + "angle": 0, + "content": "In this section, we introduce our proposed framework, which aims to investigate the intricate relationship between topology awareness and generalization performance in GNNs. The main idea of the framework is to use metric distortion to characterize the topology awareness of GNNs. With this characterization, we are able to show the relation between topology awareness of GNNs and their generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.291, + 0.423, + 0.306 + ], + "angle": 0, + "content": "3.1 GNN Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.789, + 0.407 + ], + "angle": 0, + "content": "Let \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E},\\mathbf{X})\\) be an undirected graph where \\(\\mathcal{V} = \\{1,2,\\dots,n\\}\\) is the set of \\(n\\) nodes and \\(\\mathcal{E}\\subseteq \\mathcal{V}\\times \\mathcal{V}\\) is the set of edges. Let \\(\\mathbf{X} = \\{x_{1},x_{2},\\ldots ,x_{n}\\}\\) denote the set of node attribute vectors corresponding to all the nodes in \\(\\mathcal{V}\\). More specifically, \\(x_{v}\\in \\mathbf{X}\\) is an m-dimensional vector which captures the attribute values of node \\(v\\in \\mathcal{V}\\). In addition, each vertex \\(v\\in \\mathcal{V}\\) is associated with a node label \\(y_{v}\\in \\{1,\\dots,c\\}\\), where \\(c\\) is the total number of classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.543 + ], + "angle": 0, + "content": "For our analysis, we focus on the transductive node classification setting [50], where the graph \\(\\mathcal{G}\\) are observed prior to learning and labels are only available for a small subset of nodes \\(\\mathcal{V}_0\\subset \\mathcal{V}\\). Without loss of generality, we treat \\(\\mathcal{G}\\) as fixed throughout our analysis and the randomness comes from the labels \\(y\\). Given a small set of the labelled nodes, \\(\\mathcal{V}_0\\subset \\mathcal{V}\\), the task of node-level semi-supervised learning is to learn a classifier \\(f\\), such that \\(y_{v} = f(v)\\), from a hypothesis space \\(\\mathcal{F}\\) and perform it on the remaining unlabeled nodes. Let \\(\\mathcal{L}(.,.)\\mapsto \\mathbb{R}_{+}\\) be a bounded continuous loss function. We denote by \\(R^{\\mathcal{L}}(f,\\mathcal{V})\\) the generalization risk of the classifier \\(f\\) with respect to \\(\\mathcal{L}\\) and \\(\\mathcal{V}\\) and it is defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.556, + 0.785, + 0.573 + ], + "angle": 0, + "content": "\\[\nR ^ {\\mathcal {L}} (f, \\mathcal {V}) = \\mathbb {E} _ {v \\sim \\mathcal {V}} [ \\mathcal {L} (f (v), y _ {v}) ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.689 + ], + "angle": 0, + "content": "In our case, the hypothesis space is captured by a GNN and an additional read-out function. GNNs are commonly used as an embedding function and adopt a neighbourhood aggregation scheme, where the representation of a node is learnt iteratively by aggregating representations of its neighbours. After \\( k \\) iterations of aggregation, a node's representation captures the information within its \\( k \\)-hop neighbourhood in the graph. Formally, the \\( k \\)-th layer of a GNN can be written as follows," + }, + { + "type": "equation", + "bbox": [ + 0.287, + 0.701, + 0.785, + 0.72 + ], + "angle": 0, + "content": "\\[\nh _ {u} ^ {(k)} = \\operatorname {U p d a t e} ^ {(k)} \\left(\\operatorname {A g g r e g a t e} ^ {(k)} \\left(\\left\\{h _ {v} ^ {(k - 1)}, \\forall v \\in \\mathcal {N} (u) \\cup \\{u \\} \\right\\}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( h_u^{(k)} \\) is the representation vector of node u at the k-th layer. \\( h_u^{(0)} \\) is initialized to be \\( x_u \\), and \\( \\mathcal{N}(u) \\) is the set of nodes directly adjacent to u. Different GNNs are compositions of aggregation and update layers of different designs. A GNN is commonly interpreted as an embedding function that combines the graph structure and node features to learn a representation vector \\( h_v \\) of node v. Then the learnt representation is fed into a task-specific read-out function g (commonly a multilayer perceptron) for the prediction. \\( g \\) maps the embedding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.128 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "vector \\( h_v \\) to a label \\( y_v \\) in the label set \\( \\mathbb{R}^c \\). The overall GNN prediction process can be viewed as a composition of \\( g \\circ \\mathcal{M}: \\mathcal{V} \\mapsto \\mathbb{R}^c \\). The overall GNN prediction process can be viewed as a composition of \\( g \\circ \\mathcal{M}: \\mathcal{V} \\mapsto \\mathbb{R}^c \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.216, + 0.434, + 0.233 + ], + "angle": 0, + "content": "3.2 Structural Subgroup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.351 + ], + "angle": 0, + "content": "To study the generalization performance of GNNs on different structural subgroups, we assume the existence of a generalized distance (structural measure) function, \\( d_{s}(u,v) \\), that captures the similarity between two nodes \\( u,v \\) with respect to structure \\( s \\). Furthermore, we assume that this structural measure function \\( d_{s}(u,v) \\) and the vertex set \\( \\mathcal{V} \\) form a metric space. A metric space is a fundamental mathematical object that allows for studying and measuring the relation among the elements in a given set, and it is formally defined as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.359, + 0.79, + 0.406 + ], + "angle": 0, + "content": "Definition 1 (metric space). A metric space is a pair \\((\\mathcal{Q},d)\\), where \\(\\mathcal{Q}\\) is a set and \\(d:\\mathcal{Q}\\times \\mathcal{Q}\\mapsto \\mathbb{R}\\) is a metric satisfying the following axioms (\\(x,y,z\\) are arbitrary points of \\(\\mathcal{Q}\\)):" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.419, + 0.615, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} (M 1) d (x, y) \\geq 0, \\\\ (M 2) d (x, x) = 0, \\\\ (M 3) d (x, y) > 0, x \\neq y, \\\\ (M 4) d (x, y) = d (y, x), \\\\ (M 5) d (x, y) + d (y, z) \\geq d (x, z). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.629 + ], + "angle": 0, + "content": "These properties ensure that the metric \\( d(.) \\) can be used to quantify the distance or dissimilarity between any two elements in the set \\( \\mathcal{Q} \\). In the context of graph learning, the metric space \\( (\\mathcal{V},d_s) \\) allows us to study the similarity between nodes in terms of the structural property \\( s \\) and measure the structural distance between different subgroups of the test set. One natural and commonly used example in the graph is the metric space formed by the vertex set and the shortest-path distance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.64, + 0.789, + 0.687 + ], + "angle": 0, + "content": "Definition 2 (Structural Group Distance). Given a metric space \\((\\mathcal{V}, d_s)\\) with respect to structure \\(s\\) and two distinct sets (groups) \\(\\mathcal{V}_1, \\mathcal{V}_2 \\subset \\mathcal{V}\\), we define the structural distance \\(D_s\\) between an element \\(v \\in \\mathcal{V}\\) to a group as" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.7, + 0.785, + 0.721 + ], + "angle": 0, + "content": "\\[\nD _ {s} (v, \\mathcal {V} _ {1}) = \\min _ {u \\in \\mathcal {V} _ {1}} d _ {s} (v, u) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.733, + 0.604, + 0.749 + ], + "angle": 0, + "content": "and the structural distance between the two groups as" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.762, + 0.785, + 0.784 + ], + "angle": 0, + "content": "\\[\nD _ {s} \\left(\\mathcal {V} _ {1}, \\mathcal {V} _ {2}\\right) = \\max _ {v \\in \\mathcal {V} _ {1}} \\min _ {u \\in \\mathcal {V} _ {2}} d _ {s} (v, u) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "It should be noted that the distance function in Def. 2 is not symmetrical, and Eq. 4 captures the maximal structural difference of vertices in \\(\\mathcal{V}_1\\) to \\(\\mathcal{V}_2\\). To provide further insight into the concept of a structural subgroup, let's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "consider a scenario where \\(\\mathcal{V}_0\\subset \\mathcal{V}\\) represents the training set and \\(d_{s}(.)\\) represents the shortest-path distance. In this context, we can partition the remaining vertices \\(\\mathcal{V}\\backslash \\mathcal{V}_0\\) into different structural subgroups based on their shortest-path distance to the nearest vertex in the training set. Specifically, we can define \\(\\mathcal{V}_k = \\{v\\in \\mathcal{V}\\backslash \\mathcal{V}_0|D_s(v,\\mathcal{V}_0) = k\\}\\). However, it is important to emphasize that our analysis does not assume a specific partition scheme, and the framework remains adaptable to various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.272, + 0.431, + 0.288 + ], + "angle": 0, + "content": "3.3 Topology Awareness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.787, + 0.355 + ], + "angle": 0, + "content": "To characterize and formalize the topology awareness of GNNs, we introduce the concept of distortion, which measures the degree to which the graph structure is preserved in the embedding space. Mathematically, the distortion of a function between two general metric spaces is defined as," + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.361, + 0.79, + 0.422 + ], + "angle": 0, + "content": "Definition 3 (distortion). Given two metric spaces \\((\\mathcal{Q},d)\\) and \\((\\mathcal{Q}',d')\\) and a mapping \\(\\gamma :\\mathcal{Q}\\mapsto \\mathcal{Q}'\\) \\((\\mathcal{Q}(\\mathcal{Q}')\\) is a set and \\(d(d^{\\prime})\\) is a distance function)), \\(\\gamma\\) is said to have a distortion \\(\\alpha \\geq 1\\), if there exists a constant \\(r > 0\\) such that \\(\\forall u,v\\in \\mathcal{Q}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.422, + 0.635, + 0.438 + ], + "angle": 0, + "content": "\\[\nr d (u, v) \\leq d ^ {\\prime} (\\gamma (u), \\gamma (v)) \\leq \\alpha r d (u, v).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.564 + ], + "angle": 0, + "content": "The definition above captures the maximum expansion and minimum contraction between the two metric spaces. In the extreme case when \\(\\alpha = 1\\), the mapping \\(\\gamma\\) is the so-called isometric mapping that perfectly preserves the distance between any two points in the space. In other words, for any two points \\(x\\) and \\(y\\) in the domain of the function, their distance \\(d(x,y)\\) in the original metric space is equal to the distance \\(d'(\\gamma(x), \\gamma(y))\\) in the target metric space (up to a constant scaling factor). If the distortion rate of the mapping is close to 1, then this mapping is said to have a low distortion rate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.763 + ], + "angle": 0, + "content": "In the context of GNNs, evaluating distortion involves a straightforward comparison of embeddings in the embedding domain with the structural measures on the graph, as demonstrated in the case study. If the distortion rate between the graph domain and the embedding domain is close to 1, the resulting embedding would largely preserve the corresponding structural information from the graph space, thereby indicating high topology awareness. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. This concept of distortion is agnostic to specific structural configurations, enabling the study of diverse graph structures using appropriate distance functions. Moreover, its estimation can be readily achieved by comparing the embedding with the original structural measure, as illustrated later in the case study." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.782, + 0.383, + 0.797 + ], + "angle": 0, + "content": "4 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we present the main results derived from our proposed framework for studying the generalization and fairness of GNNs with respect to different" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.128 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "structural groups in a graph. We start by demonstrating the relationship between the generalization performance of the model and the structural distance between the subgroups and the training set (Theorem 1). We then show that such a structural relation can induce disparities in the generalization performance across different subgroups (Theorem 2). Finally, we connect these theoretical results to the recent deep learning spline line theory [2] to provide additional insights. Detailed proof of the results presented in this paper can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.696, + 0.308 + ], + "angle": 0, + "content": "4.1 Generalization Performance and Structural Distance" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.321, + 0.789, + 0.382 + ], + "angle": 0, + "content": "Theorem 1. Let \\((\\mathcal{V}, d_s)\\) be a metric space for structure \\(s\\). Let \\(\\mathcal{V}_0\\) be a given labelled training set and \\(\\mathcal{V}_i \\subset \\mathcal{V} \\backslash \\mathcal{V}_0\\) be an arbitrary test subgroup. Suppose \\(\\mathcal{M}\\) is a GNN model with distortion \\(\\alpha\\) with respect to the structural measure \\(d_s(:,.)\\), and \\(g\\) and \\(\\mathcal{L}\\) are Lipschitz, we have that" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.395, + 0.786, + 0.412 + ], + "angle": 0, + "content": "\\[\nR ^ {\\mathcal {L}} \\left(g \\circ \\mathcal {M}, \\mathcal {V} _ {i}\\right) \\leq R ^ {\\mathcal {L}} \\left(g \\circ \\mathcal {M}, \\mathcal {V} _ {0}\\right) + \\mathcal {O} \\left(\\alpha D _ {s} \\left(\\mathcal {V} _ {i}, \\mathcal {V} _ {0}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.789, + 0.637 + ], + "angle": 0, + "content": "Theorem 1 states that the generalization risk of a classifier with respect to a structural subgroup depends on the empirical training loss of the classifier, the structural distance between the subgroup and the training set, and the distortion (topology awareness) of GNNs. Such a relation provides a new perspective on understanding the link between topology awareness and generalization performance in GNNs. Prior research in this area has mainly focused on the connection between GNNs and the Weisfeiler-Lehman algorithm, emphasizing the ability of GNNs to differentiate various graph structures. However, Theorem 1 highlights a direct link between topology awareness and generalization performance, and a potential issue of fairness in GNNs, as it shows that the model's performance on different structural subgroups is influenced by their distance to the training set. Next, we extend the result of Theorem 1 to analyze the relation between the generalization of different structural subgroups and their distance to the training set." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.76, + 0.679 + ], + "angle": 0, + "content": "4.2 Unfair Generalization Performance on Structural Subgroups" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.69, + 0.79, + 0.767 + ], + "angle": 0, + "content": "Theorem 2. Let \\((\\mathcal{V}, d_s)\\) be a metric space for structure \\(s\\). Suppose \\(\\mathcal{M}\\) is a GNN model with small training loss on \\(\\mathcal{V}_0\\) and distortion \\(\\alpha\\) with respect to structural measure \\(d_s(:,.)\\), and \\(\\mathcal{L}\\) and \\(g\\) are smooth. Let \\(\\mathcal{V}_i, \\mathcal{V}_j \\subset \\mathcal{V} \\backslash \\mathcal{V}_0\\) be two distinct subgroups in the local neighbourhood of the training set \\(\\mathcal{V}_0\\), then there exists constant \\(\\delta > 0\\) (independent of the choice of subgroups) such that if" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.78, + 0.6, + 0.796 + ], + "angle": 0, + "content": "\\[\nD _ {s} \\left(\\mathcal {V} _ {i}, \\mathcal {V} _ {0}\\right) \\geq \\alpha \\delta D _ {s} \\left(\\mathcal {V} _ {j}, \\mathcal {V} _ {0}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.809, + 0.316, + 0.822 + ], + "angle": 0, + "content": "we have that," + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.824, + 0.618, + 0.842 + ], + "angle": 0, + "content": "\\[\nR ^ {\\mathcal {L}} (g \\circ \\mathcal {M}, \\mathcal {V} _ {i}) \\geq R ^ {\\mathcal {L}} (g \\circ \\mathcal {M}, \\mathcal {V} _ {j}).\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Theorem 2 states that when comparing two subgroups, if the difference in their structural distance to the training set is larger than the threshold \\((\\delta \\alpha)\\), then GNNs exhibit unfair generalization performance with respect to these groups. Specifically, the theorem states that GNNs generalize better on subgroups that are structurally closer to the training set, which leads to accuracy disparities among different subgroups. Additionally, the theorem notes that the topology awareness of GNNs (measured by the distortion rate \\(\\alpha\\)) plays a key role in determining the strength of this accuracy disparity. This has important implications for the use of GNNs in critical applications, as it suggests that increased topology awareness in GNNs can lead to improved generalization performance but also a greater potential for unfairness in terms of accuracy disparities among different structural subgroups. As such, it is important to carefully consider the trade-offs between topology awareness and fairness when designing new GNNs and applying them to real-world problems." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.646, + 0.41 + ], + "angle": 0, + "content": "4.3 Connection with Deep Learning Spline Theory" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.49, + 0.84 + ], + "angle": 0, + "content": "Our theoretical findings reveal that when considering the performance of the model from the viewpoint of the embedding space, the training set creates a \"valley\" effect on generalization performance, with the training vertex at the bottom of the valley. Representations that are closer to the embedding of the training set perform better, while those farther away perform worse. This pattern has an interesting connection with deep learning spline theory [2]. This theory suggests that in DNNs, the learning process for classification divides the input/latent space into small circles (referred to as power diagrams), where representations within each circle are grouped, and representations closer to the centre of the circle are predicted with more confidence. Our findings suggest that in GNNs, the mapping of test data to these power diagrams is related to the structural similarity between test and training data in the graph. This connection provides further validation for our derived results." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.47, + 0.765, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.691, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Fig.2. The loss landscape in embedding space induced by different training sets. When vertices \\(a\\) and \\(d\\) are selected as the training set instead of \\(b\\) and \\(c\\), the test losses are smaller as the average distance between the training set and the other vertices is smaller (the neighbourhoods of \\(h_a\\) and \\(h_d\\) cover embeddings of more vertices in the embedding space)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.649, + 0.163 + ], + "angle": 0, + "content": "5 A Case Study on Shortest-Path Distance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.341 + ], + "angle": 0, + "content": "In this section, we present a case study which utilizes the shortest path distance to validate the predictions and effectiveness of our proposed framework. We aim to demonstrate how our framework can be applied to a real-world scenario and how it can be used to understand the generalization performance of GNNs on different structural subgroups. The case study will provide insights and show the potential impact of our results on various critical applications of GNNs. In this case study, we aim to answer empirically the following questions: Given a GNN model that can preserve certain structure \\( s \\) (\\( \\alpha \\) closed to 1), is the derived structural result observed on the generalization performance of the GNN model with respect to structural subgroups defined by \\( s \\)? Furthermore, how can our outcomes be harnessed to effectively tackle real-world challenges?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Overall, our case study demonstrates that our proposed framework can be applied to a real-world scenario and provides useful insights into the performance of GNNs in different structural subgroups. It also shows that the results of our framework can be generalized for different train/test pairs. Moreover, we showcase its practical utility by employing the findings to address the cold start problem within graph active learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.583 + ], + "angle": 0, + "content": "General Setup. We use Deep Graph Library [43] for our implementation and experiment with four representative GNN models, GCN [17], Sage [13], GAT [42] and GCNII [7]. In all experiments, we follow closely the common GNN experimental setup, and multiple independent trials are conducted for each experiment to obtain average results. We adopt five popular and publicly available node classification datasets: Cora [34], Citeseer [21], CoraFull [4], CoauthorCS [36] and Ogbn-arxiv [16]. Due to space limitations, we only show part of the experimental results in the main body of the paper. Detailed descriptions of the datasets, hyperparameters and remaining experimental results can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Structural Subgroups. We examine the accuracy disparity of subgroups defined by the shortest path distance (spd). In order to directly investigate the effect of structural distance on the generalization bound (Theorem 1), we first split the test nodes into subgroups by their shortest-path distance to the training set. For each test node \\( v \\), its distance to the training set \\( \\mathcal{V}_0 \\) is given by \\( d_v = D_{\\mathrm{spd}}(v, \\mathcal{V}_0) \\) with the shortest path distance metric as defined in Def. 2. Then we group the test nodes according to this distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Distance Awareness of the Selected GNNs. We first verify if the selected GNNs preserve (are aware of) graph distances (with a small distortion rate), by evaluating the relation between the graph distances of test vertices to the training set and the embedding distances (Euclidean distance) of vertex representations to the representations of the training set, i.e., the relation between \\( d_{\\mathrm{spd}}(v,\\mathcal{V}_0) \\) and \\( d_{\\mathrm{Euclidean}}(\\mathcal{M}(v),\\mathcal{M}(\\mathcal{V}_0)) \\). In this experiment, we train the GNN models on the default training set provided in the dataset. We extract vertices within the 5-hop neighbourhood of the training set and group the vertices based on their distances to the training set. 5-hop neighbourhood is chosen because the diameter of most real-life networks is around 5 and it gives a better visual" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "presentation. We then compute the average embedding distance of each group to the training set in the embedding space. Based on Def. 3, if an embedding has low distortion, we should expect to see linear relation with the slope being the scaling factor. As shown in Fig. 3, there exists a strong correlation between the graph distance and the embedding distance. The relation is near linear and the relative orders of distances are mostly preserved within the first few hops. This implies that the distortion \\(\\alpha\\) is small with the selected GNNs, i.e., they are distance aware. It also indicates that the premise of the small distortion in our theoretical results can be commonly satisfied in practice." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.304, + 0.666, + 0.32 + ], + "angle": 0, + "content": "5.1 Graph Distance and Generalization Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Next, we study the relation between graph distance and generalization performance following the same settings as in the previous experiment. We compute the average prediction accuracy in each group of vertices. In Fig. 4, '0 hop' indicates the training set. We observe that the selected GNNs (with low distortion) indeed tend to generalize better on the group of vertices that are closer to the training set, which validates Theorem 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.614 + ], + "angle": 0, + "content": "We further validate if the distance between different pairs of the training sets and the test group is related to the generalization performance of GNNs in the test group. We randomly sample \\( k = 5 \\) vertices from each class as the training set \\( \\mathcal{V}_0 \\). Then we evaluate the GNN model on the rest of the vertices \\( \\mathcal{V} \\backslash \\mathcal{V}_0 \\). We make \\( k \\) relatively small to enable a larger variance (among different trials) in the distance \\( D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0) \\) (as given in Def. 2) between the training set and the rest of the vertices. We conduct 30 independent trials each selecting a different training set and ranking all the trials in descending order of the mean graph distance between the training set and the rest of the vertices. In Table 1, Group 1 contains the 10 trials with the largest graph distance \\( D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0) \\) and Group 3 includes the 10 trials with the smallest graph distance \\( D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0) \\). The average classification accuracy and the variance (value in the bracket) are computed for each group." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.675 + ], + "angle": 0, + "content": "As shown in Table 1, training GNNs on a training set with a smaller graph distance \\( D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0) \\) to the rest of the vertices, does generalize better, which validates our result and further enables us to use these insights for designing and developing new GNNs algorithms." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.776, + 0.712 + ], + "angle": 0, + "content": "5.2 Application in Cold Start Problem for Graph Active Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Active learning is a type of machine learning where the model is actively involved in the labelling process of the training data. In active learning, the model is initially trained on a small labelled dataset and then used to select the most informative examples from the unlabeled dataset to be labelled next. Unsatisfactory initial set selection can lead to slow learning progress and unfair performance, which is referred to as the cold start problem [11]. Effective resolution of this issue entails an initial set selection that adheres to the following criteria: 1) exclusively leveraging graph structure; 2) achieving robust generalization; and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.16, + 0.344, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.242, + 0.341, + 0.256 + ], + "angle": 0, + "content": "(a) GCN, Cora" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.16, + 0.484, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.243, + 0.481, + 0.257 + ], + "angle": 0, + "content": "(b) Sage, Cora" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.16, + 0.627, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.243, + 0.623, + 0.257 + ], + "angle": 0, + "content": "(c) GAT, Cora" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.16, + 0.769, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.243, + 0.771, + 0.257 + ], + "angle": 0, + "content": "(d) GCN2, Cora" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.277, + 0.342, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.361, + 0.359, + 0.387 + ], + "angle": 0, + "content": "(e) GCN, Coauthor CS" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.278, + 0.484, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.361, + 0.499, + 0.387 + ], + "angle": 0, + "content": "(f) Sage, Coauthor CS" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.278, + 0.627, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.361, + 0.642, + 0.387 + ], + "angle": 0, + "content": "(g) GAT, Coauthor \nCS" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.278, + 0.769, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.361, + 0.785, + 0.387 + ], + "angle": 0, + "content": "(h) GCNII, Coauthor CS" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 3. Graph distance vs. embedding distance. (a)-(d) are the relation between graph distance (shortest-path distance) and embedding distance (Euclidean) of the four representative GNNs on Cora. (e)-(h) are the relation between graph distance and embedding distance of the four representative GNNs on Coauthor CS. The highlighted regions represent the \\(95\\%\\) confidence interval for that regression. (a)-(h) all demonstrate a rather linear relation between graph distance and embedding distance." + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.517, + 0.495, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.619, + 0.395, + 0.633 + ], + "angle": 0, + "content": "(a) Cora" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.517, + 0.764, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.619, + 0.692, + 0.633 + ], + "angle": 0, + "content": "(b) Coauthor CS" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Fig.4. Graph distance vs. accuracy. (a) is the relation between graph distance (shortest-path distance) and the accuracy of four representative GNNs on Cora. (b) is the relation between graph distance and the accuracy of four representative GNNs on Coauthor CS. Both (a) and (b) show that the graph distance is indeed a good indicator for generalization performance when GNNs are distance aware." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.794 + ], + "angle": 0, + "content": "3) ensuring fairness (with minimal discrepancy in generalization performance) across distinct subgroups. We illustrate how our derived findings can be instrumental in the selection of an effective initial labelled dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our empirical results show that the selected GNNs are highly distance-aware, meaning that their generalization performance is closely related to the coverage of the training set. To ensure that the model is fair and effective, the initial" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1. Accuracy of 30 independent trials with different train/test pairs grouped by \\( D_{s}(\\mathcal{V}\\backslash \\mathcal{V}_{0},\\mathcal{V}_{0}) \\). Group 1 contains the 10 trials with the largest graph distance \\( D_{\\mathrm{spd}}(\\mathcal{V}\\backslash \\mathcal{V}_0,\\mathcal{V}_0) \\) and Group 3 includes the 10 trials with the smallest graph distance \\( D_{\\mathrm{spd}}(\\mathcal{V}\\backslash \\mathcal{V}_0,\\mathcal{V}_0) \\). The average accuracy and variance of the 10 trials are recorded in the table. The results show that train/test pairs with closer structural similarity indeed have better generalization performance, as predicted by our theoretical results." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.238, + 0.787, + 0.291 + ], + "angle": 0, + "content": "
CoraCiteseer
GCNGATSageGCNIIGCNGATSageGCNII
Group 165.42 (5.03)68.03 (4.62)67.61 (5.02)61.42 (4.03)59.72 (5.41)56.78 (4.73)58.52 (4.85)55.46 (3.69)
Group 267.56 (4.23)70.94 (3.86)69.54 (4.84)66.42 (3.76)61.48 (5.15)60.33 (4.24)60.78 (3.60)58.42 (4.08)
Group 371.40 (3.59)73.52 (3.53)72.60 (4.23)72.18 (4.01)62.93 (4.14)64.63 (3.65)63.03 (4.36)63.23 (5.06)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.306, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Table 2. Initial data labelling: performance comparison of different GNNs sampling strategies based on graph structure. Results are averaged among five trials. For each dataset, we select \\(0.5\\%\\) of the total data as the initial set and evaluate on the rest of the vertices. Each entry corresponds to \"test accuracy (ACC)| maximum discrepancy (MD)\". \\(\\uparrow\\) means the higher the better and \\(\\downarrow\\) means the lower the better. Bold letter indicates the best sampling strategy for the given dataset and GNN model." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.401, + 0.787, + 0.538 + ], + "angle": 0, + "content": "
GCNSage
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)
R37.08|13.2336.35|10.238.13|3.2357.75|16.2337.69|12.0337.24|11.208.56|4.2049.72|15.52
Degree40.87|12.5236.27|9.2810.68|2.3251.50|18.2338.04|11.8735.26|12.0511.56|3.0242.78|13.02
Centrality42.75|10.4138.45|8.159.49|2.0853.48|16.1142.18|10.1540.81|9.0512.56|3.0845.65|10.11
PageRank47.87|9.5242.27|8.2512.48|2.1055.50|15.2350.15|9.1443.26|9.1513.15|2.9855.16|10.74
OURS49.03|8.2350.98|7.3215.25|1.2361.75|14.3258.57|8.2948.80|7.0314.76|2.0161.63|8.82
GATGCNII
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)|MD(↓)ACC(↑)|MD(↓)
R49.36|15.7843.53|16.279.55|3.0743.25|15.2031.53|14.0237.29|17.3210.98|3.0652.61|17.02
Degree54.08|13.5244.65|15.1710.50|2.7847.30|13.0432.19|13.8638.88|16.9810.40|2.7954.70|15.96
Centrality56.81|11.5248.15|13.0811.15|2.5248.50|13.1832.89|12.5739.98|16.0512.48|2.5256.88|15.18
PageRank60.91|11.0849.77|12.0712.94|2.1150.50|12.2334.04|12.0842.77|13.9514.56|2.1958.78|14.08
OURS62.76|10.5451.11|11.0713.39|1.8851.21|11.0235.61|11.0244.50|13.0415.19|1.9660.76|13.18
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.618 + ], + "angle": 0, + "content": "labelled set should consist of vertices that are in close proximity to the rest of the input graph (in graph distance), as illustrated in Fig. 2. This can be formulated into the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.632, + 0.576, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\mathcal {V} _ {0} \\subset \\mathcal {V}} D _ {\\operatorname {s p d}} (\\mathcal {V} \\backslash \\mathcal {V} _ {0}, \\mathcal {V} _ {0}) \\\\ s. t. \\quad | \\mathcal {V} _ {0} | = k \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( k \\) is the given initial set size and \\( D_{s}(.) \\) is given in Def. 2. Intuitively, Eq.6 aims to select a subset of vertices that exhibit the closest structural similarity (as defined by the graph distance in this case) to the remaining vertices. This problem corresponds to the well-known k-center problem in graph theory [14], which is NP-hard. While finding the optimal solution may be intractable, efficient heuristics exist for tackling Eq. 6. Exploring the best heuristic is beyond the scope of this paper. Instead, for our experiment, we adopt the classic farthest-first traversal algorithm to obtain a solution. We refer to this sample selection procedure as the coverage-based sampling algorithm and provide a detailed description of the method used in the experiment within the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "We evaluate the initial set selected by our coverage-based sampling algorithm, comparing it to the existing baselines: (1) \\(\\mathbf{R}\\): uniform random sampling [30], and importance sampling with graph properties (2) Degree [15], (3) Centrality, (4) Density and (5) PageRank [5]. We apply each strategy to select an initial dataset of \\(0.5\\%\\) of the total data for labelling (training). We evaluate the performance of the learnt model with the rest of the vertices on two metrics: 1) classification accuracy, which reflects the generalization performance, and 2) the maximum discrepancy among structural groups [1,8], which reflects the fairness. Let \\(R_{i}\\) denote the mean classification accuracy of the vertex group that is \\(i\\)-hop away from \\(\\mathcal{V}_0\\). Then, the maximum discrepancy is computed as \\(\\max \\{|R_i - R_j||i,j\\in 1,\\dots,r\\}\\), where \\(r\\) is the maximum hop number. We use \\(r = 5\\) to keep consistent with other experiments. The results in Table 2 show that coverage-based sampling achieves a significant improvement in model performance and fairness over the other strategies across different datasets, giving a better kick-start to the active learning process of GNNs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.393, + 0.51, + 0.41 + ], + "angle": 0, + "content": "6 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.605 + ], + "angle": 0, + "content": "In this work, we study and investigate GNNs which are fundamental to many computer vision and machine learning tasks. In particular, we explore the relationship between the topology awareness and generalization performance of GNNs. We introduce a novel framework that connects the structural awareness of GNNs with approximate metric embedding, offering a fresh perspective on their generalization capabilities in semi-supervised node classification tasks. This structure-agnostic framework facilitates a deeper, more intuitive comprehension of GNN generalizability across varied graph structures. Through a case study centered on graph distance, we demonstrate that our theoretical findings regarding structural results are reflected in the practical generalization performance of GNNs. Additionally, our findings shed light on the cold start problem in graph active learning and could influence various significant GNN applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.506, + 0.637 + ], + "angle": 0, + "content": "6.1 Limitation and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Our proposed framework introduces a novel angle on GNN generalization performance, yet it is not without limitations. The current approach interprets GNN embeddings in terms of metric mapping and describes the structure awareness of GNNs through distortion. This perspective, however, overlooks the specific dynamics that contribute to reduced distortion within GNNs. Investigating these underlying dynamics would significantly enrich this area of research. Moreover, our study primarily examines the transductive setting. Expanding this analysis to the inductive setting—where the model is trained on one graph but applied to another—would offer a more comprehensive view of GNN generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Acknowledgement. We would like to thank the anonymous reviewers for their helpful comments. JS and CW are supported by grants from Hong Kong RGC under the contracts HKU 17207621, 17203522 and C7004-22G (CRF)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.518, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.183, + 0.786, + 0.239 + ], + "angle": 0, + "content": "1. Ali, J., Babaei, M., Chakraborty, A., Mirzasoleiman, B., Gummadi, K.P., Singla, A.: On the fairness of time-critical influence maximization in social networks. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 1541-1542. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.241, + 0.787, + 0.296 + ], + "angle": 0, + "content": "2. Balestriero, R., richard baraniuk: A spline theory of deep learning. In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 374-383. PMLR (10-15 Jul 2018), https://proceedings.mlr.press/v80/balestriero18b.htm1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.786, + 0.338 + ], + "angle": 0, + "content": "3. Baranwal, A., Fountoulakis, K., Jagannath, A.: Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization. arXiv preprint arXiv:2102.06966 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.786, + 0.368 + ], + "angle": 0, + "content": "4. Bojchevski, A., Gunnemann, S.: Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.369, + 0.786, + 0.396 + ], + "angle": 0, + "content": "5. Cai, H., Zheng, V.W., Chang, K.C.C.: Active learning for graph embedding. arXiv preprint arXiv:1705.05085 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.786, + 0.425 + ], + "angle": 0, + "content": "6. Chen, J., Zhu, J., Song, L.: Stochastic training of graph convolutional networks with variance reduction. arXiv preprint arXiv:1710.10568 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.786, + 0.453 + ], + "angle": 0, + "content": "7. Chen, M., Wei, Z., Huang, Z., Ding, B., Li, Y.: Simple and deep graph convolutional networks (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.786, + 0.495 + ], + "angle": 0, + "content": "8. Dong, Y., Ma, J., Chen, C., Li, J.: Fairness in graph mining: A survey (2022). https://doi.org/10.48550/ARXIV.2204.09888, https://arxiv.org/abs/2204.09888" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.786, + 0.539 + ], + "angle": 0, + "content": "9. Esser, P., Chennuru Vankadara, L., Ghoshdastidar, D.: Learning theory can (sometimes) explain generalisation in graph neural networks. Advances in Neural Information Processing Systems 34, 27043-27056 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.786, + 0.569 + ], + "angle": 0, + "content": "0. Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for quantum chemistry (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.571, + 0.786, + 0.624 + ], + "angle": 0, + "content": "11. Grimova, N., Macas, M., Gerla, V.: Addressing the cold start problem in active learning approach used for semi-automated sleep stages classification. In: 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). pp. 2249–2253. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.626, + 0.786, + 0.654 + ], + "angle": 0, + "content": "12. Hamilton, W.L.: Graph representation learning. Synthesis Lectures on Artificial Intelligence and Machine Learning 14(3), 1-159 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.655, + 0.786, + 0.683 + ], + "angle": 0, + "content": "13. Hamilton, W.L., Ying, R., Leskovec, J.: Inductive representation learning on large graphs (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.685, + 0.786, + 0.726 + ], + "angle": 0, + "content": "14. Hochbaum, D.S.: Approximating covering and packing problems: set cover, vertex cover, independent set, and related problems. In: Approximation algorithms for NP-hard problems, pp. 94-143 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.728, + 0.786, + 0.768 + ], + "angle": 0, + "content": "15. Hu, S., Xiong, Z., Qu, M., Yuan, X., Côté, M.A., Liu, Z., Tang, J.: Graph policy network for transferable active learning on graphs. Advances in Neural Information Processing Systems 33, 10174-10185 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.77, + 0.786, + 0.811 + ], + "angle": 0, + "content": "16. Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., Leskovec, J.: Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "17. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.183, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "J.Su et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.204 + ], + "angle": 0, + "content": "18. Lao, D., Yang, X., Wu, Q., Yan, J.: Variational inference for training graph neural networks in low-data regime through joint structure-label estimation. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 824-834 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "19. Leman, A., Weisfeiler, B.: A reduction of a graph to a canonical form and an algebra arising during this reduction. Nauchno-Technicheskaya Informatiya 2(9), 12-16 (1968)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "20. Li, P., Wang, Y., Wang, H., Leskovec, J.: Distance encoding: Design provably more powerful neural networks for graph representation learning (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.314 + ], + "angle": 0, + "content": "21. Li, Q., Han, Z., Wu, X.M.: Deeper insights into graph convolutional networks for semi-supervised learning. In: Thirty-Second AAAI conference on artificial intelligence (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "22. Li, Y., Yu, R., Shahabi, C., Liu, Y.: Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "23. Liao, R., Urtasun, R., Zemel, R.: A {pac}-bayesian approach to generalization bounds for graph neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=TR-Nj6nFx42" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "24. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "25. Ma, J., Deng, J., Mei, Q.: Subgroup generalization and fairness of graph neural networks. arXiv preprint arXiv:2106.15535 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.786, + 0.466 + ], + "angle": 0, + "content": "26. Ma, J., Ma, Z., Chai, J., Mei, Q.: Partition-based active learning for graph neural networks. arXiv preprint arXiv:2201.09391 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.494 + ], + "angle": 0, + "content": "27. Morris, C., Ritzert, M., Fey, M., Hamilton, W.L., Lenssen, J.E., Rattan, G., Grohe, M.: Weisfeiler and leman go neural: Higher-order graph neural networks (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.522 + ], + "angle": 0, + "content": "28. Nishad, S., Agarwal, S., Bhattacharya, A., Ranu, S.: Graphreach: Position-aware graph neural network using reachability estimations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.786, + 0.55 + ], + "angle": 0, + "content": "29. Oono, K., Suzuki, T.: Graph neural networks exponentially lose expressive power for node classification. arXiv preprint arXiv:1905.10947 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.577 + ], + "angle": 0, + "content": "30. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Chen, X., Wang, X.: A survey of deep active learning (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "31. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.646 + ], + "angle": 0, + "content": "32. Sato, R.: A survey on the expressive power of graph neural networks. arXiv preprint arXiv:2003.04078 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "33. Scarselli, F., Tsoi, A.C., Hagenbuchner, M.: The vapnik-chervonenkis dimension of graph and recursive neural networks. Neural Networks 108, 248-259 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.702 + ], + "angle": 0, + "content": "34. Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., Eliassi-Rad, T.: Collective classification in network data. AI magazine 29(3), 93-93 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.73 + ], + "angle": 0, + "content": "35. Sheng, G., Su, J., Huang, C., Wu, C.: Mspipe: Efficient temporal gnn training via staleness-aware pipeline. arXiv preprint arXiv:2402.15113 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.771 + ], + "angle": 0, + "content": "36. Sinha, A., Shen, Z., Song, Y., Ma, H., Eide, D., Hsu, B.J., Wang, K.: An overview of microsoft academic service (mas) and applications. In: Proceedings of the 24th international conference on world wide web. pp. 243-246 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.812 + ], + "angle": 0, + "content": "37. Su, J., Marbach, P.: Structure of core-periphery communities. In: International Conference on Complex Networks and Their Applications. pp. 151-161. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "38. Su, J., Marbach, P.: Structural properties of corea\\(\\mathbf{E}\\)periphery communities. Advances in Complex Systems (ACS) 26(06), 1-43 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.518, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Awareness Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "39. Su, J., Zou, D., Wu, C.: Pres: Toward scalable memory-based dynamic graph neural networks (2024), https://arxiv.org/abs/2402.04284" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "40. Su, J., Zou, D., Zhang, Z., Wu, C.: Towards robust graph incremental learning on evolving graphs. In: International Conference on Machine Learning. pp. 32728-32748. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "41. Tsubaki, M., Tomii, K., Sese, J.: Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics 35(2), 309-318 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "42. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "43. Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y., et al.: Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "44. Wu, Q., Zhang, H., Yan, J., Wipf, D.: Handling distribution shifts on graphs: An invariance perspective (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.396 + ], + "angle": 0, + "content": "45. Wu, Y., Xu, Y., Singh, A., Yang, Y., Dubrawski, A.: Active learning for graph neural networks via node feature propagation. arXiv preprint arXiv:1910.07567 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.424 + ], + "angle": 0, + "content": "46. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "47. Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "48. Yang, C., Wu, Q., Wang, J., Yan, J.: Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps. arXiv preprint arXiv:2212.09034 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.521 + ], + "angle": 0, + "content": "49. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.562 + ], + "angle": 0, + "content": "50. Yang, Z., Cohen, W.W., Salakhutdinov, R.: Revisiting semi-supervised learning with graph embeddings (2016). https://doi.org/10.48550/ARXIV.1603.08861, https://arxiv.org/abs/1603.08861" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.563, + 0.751, + 0.577 + ], + "angle": 0, + "content": "51. You, J., Ying, R., Leskovec, J.: Position-aware graph neural networks (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.632 + ], + "angle": 0, + "content": "53. Zhou, K., Huang, X., Li, Y., Zha, D., Chen, R., Hu, X.: Towards deeper graph neural networks with differentiable group normalization (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "54. Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Feng, M., Zhao, X., Miao, Q., Shah, S.A.A., Bennamoun, M.: Scene graph generation: A comprehensive survey (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_origin.pdf b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d939385ee911f8e525750df35b3bce1f30240d5a --- /dev/null +++ b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/a30a461c-3f3b-47ab-8426-e61337113303_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fdf7ab8bbd4f1b1631beda5d0fa8750c9b303b9ad8aa0cc6e3f7e87be37bceb +size 2472231 diff --git a/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/full.md b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..41d6af2e4ee75a9c8b274485307552aee2b6f2dd --- /dev/null +++ b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/full.md @@ -0,0 +1,299 @@ +# On the Topology Awareness and Generalization Performance of Graph Neural Networks + +Junwei Su and Chuan Wu† + +Department of Computer Science, the University of Hong Kong {jwsu, cwu} @cs.hku.hk + +Abstract. Many computer vision and machine learning problems are modelled as learning tasks on graphs, where graph neural networks (GNNs) have emerged as a dominant tool for learning representations of graph-structured data. A key feature of GNNs is their use of graph structures as input, enabling them to exploit the graphs' inherent topological properties—known as the topology awareness of GNNs. Despite the empirical successes of GNNs, the influence of topology awareness on generalization performance remains unexplored, particularly for node-level tasks that diverge from the assumption of data being independent and identically distributed (I.I.D.). The precise definition and characterization of the topology awareness of GNNs, especially concerning different topological features, are still unclear. This paper introduces a comprehensive framework to characterize the topology awareness of GNNs across any topological feature. Using this framework, we investigate the effects of topology awareness on GNN generalization performance. Contrary to the prevailing belief that enhancing the topology awareness of GNNs is always advantageous, our analysis reveals a critical insight: improving the topology awareness of GNNs may inadvertently lead to unfair generalization across structural groups, which might not be desired in some scenarios. Additionally, we conduct a case study using the intrinsic graph metric, the shortest-path distance, on various benchmark datasets. The empirical results of this case study confirm our theoretical insights. Moreover, we demonstrate the practical applicability of our framework by using it to tackle the cold start problem in graph active learning. + +# 1 Introduction + +Many problems in computer vision and machine learning are modeled as learning tasks on graphs. For example, in semantic segmentation, graphs model the relationships between different image regions, enhancing accuracy and context-aware segmentation. Graph neural networks (GNNs) have emerged as a dominant class of machine learning models specifically designed for learning representations of graph-structured data. They have demonstrated considerable success in addressing a wide range of graph-related problems in various domains such as chemistry [10], biology [41], social networking [6,22,35], scene graph generation [49,54] + +and visual relationship detection [24,46,52]. A defining characteristic of GNNs is their use of a spatial approach through message passing on the graph structure for feature aggregation. This enables GNNs to preserve structural information or dependencies (referred to as topology awareness) from the underlying graph structure, allowing them to be highly effective in tasks such as node classification. Fig. 1 illustrates the overall learning process of GNNs. + +Despite their practicality and potential, there remains a lack of theoretical understanding about GNNs, particularly in the semi-supervised node classification setting where the dependencies among the data differ significantly from other machine learning models [25]. In this setting, the goal is to leverage relations, as captured by the graph structure, among the data and a small set of labelled nodes to predict labels for the remaining nodes. Most of the existing theoretical studies of GNNs have focused on the connection between the message-passing mechanism of GNNs and the Weisfeiler-Lehman isomorphism test [19], aiming to understand the ability of GNNs to differentiate different graph structures in the learned representations, known as the expressive power of GNNs. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [28,32,51]. + +However, as GNNs become more reliant on and sensitive (aware) of graph structure as input, they may exhibit different generalization performances towards certain structural subgroups (distinct data subsets grouped by structural similarity to the training set) within the data. The quantification of GNN generalization across distinct structural subgroups is termed structural subgroup generalization [25]. Such considerations are vital in GNN application and development. For instance, within protein-protein interaction networks, these structural subgroups could represent different molecular complexes, influencing the accuracy of interaction predictions. Similarly, understanding how the topology awareness of GNNs influences generalization is essential when devising sampling strategies for training. The extent to which the generalization performance of GNNs is influenced by specific structural features of graph data is critical in deciding the composition of training datasets. Despite its importance, an understanding of the relationship between the topology awareness of GNNs and its structural subgroup generalization is still lacking. Furthermore, characterizing the topology awareness of GNNs poses a challenge, especially considering that different domains and tasks may prioritize distinct structural aspects. Therefore, a versatile framework is needed to assess the topology awareness of GNNs in relation to various structures. + +To address this gap, in this paper, we propose a novel framework based on approximate metric embedding to study the relationship between structural subgroup generalization and topology awareness of GNNs in the context of semi-supervised node classification. The proposed framework allows for the investigation of the structural subgroup generalization of GNNs with respect to different structural subgroups. More concretely, the main contributions of this work are summarized as follows. + +![](images/70cc696bee85f371baf0fa14792b5a03e165bb9a7edfc83c0f5ffcc78b9fc29b.jpg) +Fig. 1. An illustration of the learning process in a 2-layer GNN. The message-passing mechanism leverages the graph structure to aggregate information, thereby generating the representation/embedding $h_a$ for the target vertex $a$ (highlighted in red). + +1. We propose a novel, structure-agnostic framework using approximate metric embedding to examine the interplay between GNNs' structural subgroup generalization and topology awareness. This framework is versatile, accommodating various structural measures like shortest-path distance, and requires only the corresponding structural measure. Its simplicity in estimating key factors makes it applicable and generalizable to a wide range of scenarios. + +2. Through formal analysis within our framework, we establish a clear link between GNN topology awareness and their generalization performance (Theorem 1). We also demonstrate that while enhanced topology awareness boosts GNN expressiveness, it can result in uneven generalization performance, favouring subgroups more structurally similar to the training set (Theorem 2). Such structural property can be harmful (causing unfairness issues) or useful (informing design decisions) depending on the scenario. This challenges the prevailing belief that increased topology awareness universally benefits GNNs [28, 32, 51], emphasizing the importance of considering the relationship between topology awareness and generalization performance. + +3. We validate our framework through a case study on shortest-path distance, highlighting its practicality and relevance. The results corroborate our theoretical findings, showing that GNNs with heightened awareness of shortest-path distances excel in classifying vertex groups closer to the training set. Moreover, we demonstrate how our findings can be applied to mitigate the cold start problem in graph active learning [11,15], highlighting the practical implications of our framework and results. + +# 2 Related Work + +# 2.1 Topology Awareness of GNNs + +Modern GNN architectures take inspiration from the Weisfeiler-Lehman isomorphism test [12, 19], which employs the graph structure to propagate information. Consequently, a significant portion of existing literature on GNNs concentrates + +on understanding their ability to differentiate various graph structures, known as the expressive power of GNNs [27-29, 47, 51, 53]. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties [37, 38] in the learned representation [28, 32, 51]. Nevertheless, the precise ramifications of heightened topology awareness on GNNs' generalization performance remain shrouded in ambiguity. The intricate variations within structures across different domains and tasks further compound this complexity, underscoring the necessity for a unified framework—akin to the one proposed in this paper—to shed light on these intricacies. + +# 2.2 Generalization of GNNs + +Relatively few studies focus on understanding the generalization performance of GNNs in node-level tasks. [3,9,18,23,25,33,39,40,44,48] extend the conventional statistical learning or information-theoretic framework to GNNs, providing different generalization bounds based on concepts such as Vapnik-Chervonenkis dimension, mutual information, and algorithm stability. In light of this, despite empirical indications highlighting the critical role of GNNs' topology awareness in generalization performance [20,28,51], there remains limited understanding of the relationship between topology awareness of GNNs and their generalization. Our proposed framework addresses this gap by utilizing the distortion concept in the approximate metric embedding to connect the topology awareness of GNNs and their generalization performance. + +# 2.3 Active Learning in Graph + +Active Learning (AL) seeks to enhance model performance within a limited labeling budget by identifying the most representative samples for annotation [31]. The AL methodology generally involves two phases: 1) selecting a small, preliminary subset of samples for labeling; 2) utilizing this labeled subset to iteratively discover the most informative sample for future labeling. The initial selection of an effective subset represents the cold-start problem in AL [11]. While AL has proven effective in settings with independent and identically distributed (i.i.d.) data, such as images and text, applying it to Graph Neural Networks (GNNs) poses unique challenges. These challenges stem from the necessity to account for node interdependencies and the computational complexity of iterative sample selection and retraining, especially when considering multi-hop neighbors. Recent AL methodologies for GNNs, such as AGE [5], FeatProp [45], and GraphPart [26], aim to overcome these obstacles by leveraging the graph structure and initial node features to boost efficiency and efficacy, although their success may significantly hinge on the quality of the initial features. In this work, we focus on the cold-start problem of graph active learning and introduce a novel approach (based on our analysis) that relies exclusively on graph structure, thus eliminating the reliance on feature quality at the onset of learning. + +# 3 Framework + +In this section, we introduce our proposed framework, which aims to investigate the intricate relationship between topology awareness and generalization performance in GNNs. The main idea of the framework is to use metric distortion to characterize the topology awareness of GNNs. With this characterization, we are able to show the relation between topology awareness of GNNs and their generalization performance. + +# 3.1 GNN Preliminaries + +Let $\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{X})$ be an undirected graph where $\mathcal{V} = \{1,2,\dots,n\}$ is the set of $n$ nodes and $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$ is the set of edges. Let $\mathbf{X} = \{x_{1},x_{2},\ldots ,x_{n}\}$ denote the set of node attribute vectors corresponding to all the nodes in $\mathcal{V}$ . More specifically, $x_{v}\in \mathbf{X}$ is an m-dimensional vector which captures the attribute values of node $v\in \mathcal{V}$ . In addition, each vertex $v\in \mathcal{V}$ is associated with a node label $y_{v}\in \{1,\dots,c\}$ , where $c$ is the total number of classes. + +For our analysis, we focus on the transductive node classification setting [50], where the graph $\mathcal{G}$ are observed prior to learning and labels are only available for a small subset of nodes $\mathcal{V}_0\subset \mathcal{V}$ . Without loss of generality, we treat $\mathcal{G}$ as fixed throughout our analysis and the randomness comes from the labels $y$ . Given a small set of the labelled nodes, $\mathcal{V}_0\subset \mathcal{V}$ , the task of node-level semi-supervised learning is to learn a classifier $f$ , such that $y_{v} = f(v)$ , from a hypothesis space $\mathcal{F}$ and perform it on the remaining unlabeled nodes. Let $\mathcal{L}(.,.)\mapsto \mathbb{R}_{+}$ be a bounded continuous loss function. We denote by $R^{\mathcal{L}}(f,\mathcal{V})$ the generalization risk of the classifier $f$ with respect to $\mathcal{L}$ and $\mathcal{V}$ and it is defined as follows, + +$$ +R ^ {\mathcal {L}} (f, \mathcal {V}) = \mathbb {E} _ {v \sim \mathcal {V}} [ \mathcal {L} (f (v), y _ {v}) ] \tag {1} +$$ + +In our case, the hypothesis space is captured by a GNN and an additional read-out function. GNNs are commonly used as an embedding function and adopt a neighbourhood aggregation scheme, where the representation of a node is learnt iteratively by aggregating representations of its neighbours. After $k$ iterations of aggregation, a node's representation captures the information within its $k$ -hop neighbourhood in the graph. Formally, the $k$ -th layer of a GNN can be written as follows, + +$$ +h _ {u} ^ {(k)} = \operatorname {U p d a t e} ^ {(k)} \left(\operatorname {A g g r e g a t e} ^ {(k)} \left(\left\{h _ {v} ^ {(k - 1)}, \forall v \in \mathcal {N} (u) \cup \{u \} \right\}\right)\right), \tag {2} +$$ + +where $h_u^{(k)}$ is the representation vector of node u at the k-th layer. $h_u^{(0)}$ is initialized to be $x_u$ , and $\mathcal{N}(u)$ is the set of nodes directly adjacent to u. Different GNNs are compositions of aggregation and update layers of different designs. A GNN is commonly interpreted as an embedding function that combines the graph structure and node features to learn a representation vector $h_v$ of node v. Then the learnt representation is fed into a task-specific read-out function g (commonly a multilayer perceptron) for the prediction. $g$ maps the embedding + +vector $h_v$ to a label $y_v$ in the label set $\mathbb{R}^c$ . The overall GNN prediction process can be viewed as a composition of $g \circ \mathcal{M}: \mathcal{V} \mapsto \mathbb{R}^c$ . The overall GNN prediction process can be viewed as a composition of $g \circ \mathcal{M}: \mathcal{V} \mapsto \mathbb{R}^c$ . + +# 3.2 Structural Subgroup + +To study the generalization performance of GNNs on different structural subgroups, we assume the existence of a generalized distance (structural measure) function, $d_{s}(u,v)$ , that captures the similarity between two nodes $u,v$ with respect to structure $s$ . Furthermore, we assume that this structural measure function $d_{s}(u,v)$ and the vertex set $\mathcal{V}$ form a metric space. A metric space is a fundamental mathematical object that allows for studying and measuring the relation among the elements in a given set, and it is formally defined as follows. + +Definition 1 (metric space). A metric space is a pair $(\mathcal{Q},d)$ , where $\mathcal{Q}$ is a set and $d:\mathcal{Q}\times \mathcal{Q}\mapsto \mathbb{R}$ is a metric satisfying the following axioms ( $x,y,z$ are arbitrary points of $\mathcal{Q}$ ): + +$$ +\begin{array}{l} (M 1) d (x, y) \geq 0, \\ (M 2) d (x, x) = 0, \\ (M 3) d (x, y) > 0, x \neq y, \\ (M 4) d (x, y) = d (y, x), \\ (M 5) d (x, y) + d (y, z) \geq d (x, z). \\ \end{array} +$$ + +These properties ensure that the metric $d(.)$ can be used to quantify the distance or dissimilarity between any two elements in the set $\mathcal{Q}$ . In the context of graph learning, the metric space $(\mathcal{V},d_s)$ allows us to study the similarity between nodes in terms of the structural property $s$ and measure the structural distance between different subgroups of the test set. One natural and commonly used example in the graph is the metric space formed by the vertex set and the shortest-path distance. + +Definition 2 (Structural Group Distance). Given a metric space $(\mathcal{V}, d_s)$ with respect to structure $s$ and two distinct sets (groups) $\mathcal{V}_1, \mathcal{V}_2 \subset \mathcal{V}$ , we define the structural distance $D_s$ between an element $v \in \mathcal{V}$ to a group as + +$$ +D _ {s} (v, \mathcal {V} _ {1}) = \min _ {u \in \mathcal {V} _ {1}} d _ {s} (v, u) \tag {3} +$$ + +and the structural distance between the two groups as + +$$ +D _ {s} \left(\mathcal {V} _ {1}, \mathcal {V} _ {2}\right) = \max _ {v \in \mathcal {V} _ {1}} \min _ {u \in \mathcal {V} _ {2}} d _ {s} (v, u) \tag {4} +$$ + +It should be noted that the distance function in Def. 2 is not symmetrical, and Eq. 4 captures the maximal structural difference of vertices in $\mathcal{V}_1$ to $\mathcal{V}_2$ . To provide further insight into the concept of a structural subgroup, let's + +consider a scenario where $\mathcal{V}_0\subset \mathcal{V}$ represents the training set and $d_{s}(.)$ represents the shortest-path distance. In this context, we can partition the remaining vertices $\mathcal{V}\backslash \mathcal{V}_0$ into different structural subgroups based on their shortest-path distance to the nearest vertex in the training set. Specifically, we can define $\mathcal{V}_k = \{v\in \mathcal{V}\backslash \mathcal{V}_0|D_s(v,\mathcal{V}_0) = k\}$ . However, it is important to emphasize that our analysis does not assume a specific partition scheme, and the framework remains adaptable to various scenarios. + +# 3.3 Topology Awareness + +To characterize and formalize the topology awareness of GNNs, we introduce the concept of distortion, which measures the degree to which the graph structure is preserved in the embedding space. Mathematically, the distortion of a function between two general metric spaces is defined as, + +Definition 3 (distortion). Given two metric spaces $(\mathcal{Q},d)$ and $(\mathcal{Q}',d')$ and a mapping $\gamma :\mathcal{Q}\mapsto \mathcal{Q}'$ $(\mathcal{Q}(\mathcal{Q}')$ is a set and $d(d^{\prime})$ is a distance function)), $\gamma$ is said to have a distortion $\alpha \geq 1$ , if there exists a constant $r > 0$ such that $\forall u,v\in \mathcal{Q}$ + +$$ +r d (u, v) \leq d ^ {\prime} (\gamma (u), \gamma (v)) \leq \alpha r d (u, v). +$$ + +The definition above captures the maximum expansion and minimum contraction between the two metric spaces. In the extreme case when $\alpha = 1$ , the mapping $\gamma$ is the so-called isometric mapping that perfectly preserves the distance between any two points in the space. In other words, for any two points $x$ and $y$ in the domain of the function, their distance $d(x,y)$ in the original metric space is equal to the distance $d'(\gamma(x), \gamma(y))$ in the target metric space (up to a constant scaling factor). If the distortion rate of the mapping is close to 1, then this mapping is said to have a low distortion rate. + +In the context of GNNs, evaluating distortion involves a straightforward comparison of embeddings in the embedding domain with the structural measures on the graph, as demonstrated in the case study. If the distortion rate between the graph domain and the embedding domain is close to 1, the resulting embedding would largely preserve the corresponding structural information from the graph space, thereby indicating high topology awareness. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. This concept of distortion is agnostic to specific structural configurations, enabling the study of diverse graph structures using appropriate distance functions. Moreover, its estimation can be readily achieved by comparing the embedding with the original structural measure, as illustrated later in the case study. + +# 4 Main Results + +In this section, we present the main results derived from our proposed framework for studying the generalization and fairness of GNNs with respect to different + +structural groups in a graph. We start by demonstrating the relationship between the generalization performance of the model and the structural distance between the subgroups and the training set (Theorem 1). We then show that such a structural relation can induce disparities in the generalization performance across different subgroups (Theorem 2). Finally, we connect these theoretical results to the recent deep learning spline line theory [2] to provide additional insights. Detailed proof of the results presented in this paper can be found in the supplementary material. + +# 4.1 Generalization Performance and Structural Distance + +Theorem 1. Let $(\mathcal{V}, d_s)$ be a metric space for structure $s$ . Let $\mathcal{V}_0$ be a given labelled training set and $\mathcal{V}_i \subset \mathcal{V} \backslash \mathcal{V}_0$ be an arbitrary test subgroup. Suppose $\mathcal{M}$ is a GNN model with distortion $\alpha$ with respect to the structural measure $d_s(:,.)$ , and $g$ and $\mathcal{L}$ are Lipschitz, we have that + +$$ +R ^ {\mathcal {L}} \left(g \circ \mathcal {M}, \mathcal {V} _ {i}\right) \leq R ^ {\mathcal {L}} \left(g \circ \mathcal {M}, \mathcal {V} _ {0}\right) + \mathcal {O} \left(\alpha D _ {s} \left(\mathcal {V} _ {i}, \mathcal {V} _ {0}\right)\right). \tag {5} +$$ + +Theorem 1 states that the generalization risk of a classifier with respect to a structural subgroup depends on the empirical training loss of the classifier, the structural distance between the subgroup and the training set, and the distortion (topology awareness) of GNNs. Such a relation provides a new perspective on understanding the link between topology awareness and generalization performance in GNNs. Prior research in this area has mainly focused on the connection between GNNs and the Weisfeiler-Lehman algorithm, emphasizing the ability of GNNs to differentiate various graph structures. However, Theorem 1 highlights a direct link between topology awareness and generalization performance, and a potential issue of fairness in GNNs, as it shows that the model's performance on different structural subgroups is influenced by their distance to the training set. Next, we extend the result of Theorem 1 to analyze the relation between the generalization of different structural subgroups and their distance to the training set. + +# 4.2 Unfair Generalization Performance on Structural Subgroups + +Theorem 2. Let $(\mathcal{V}, d_s)$ be a metric space for structure $s$ . Suppose $\mathcal{M}$ is a GNN model with small training loss on $\mathcal{V}_0$ and distortion $\alpha$ with respect to structural measure $d_s(:,.)$ , and $\mathcal{L}$ and $g$ are smooth. Let $\mathcal{V}_i, \mathcal{V}_j \subset \mathcal{V} \backslash \mathcal{V}_0$ be two distinct subgroups in the local neighbourhood of the training set $\mathcal{V}_0$ , then there exists constant $\delta > 0$ (independent of the choice of subgroups) such that if + +$$ +D _ {s} \left(\mathcal {V} _ {i}, \mathcal {V} _ {0}\right) \geq \alpha \delta D _ {s} \left(\mathcal {V} _ {j}, \mathcal {V} _ {0}\right), +$$ + +we have that, + +$$ +R ^ {\mathcal {L}} (g \circ \mathcal {M}, \mathcal {V} _ {i}) \geq R ^ {\mathcal {L}} (g \circ \mathcal {M}, \mathcal {V} _ {j}). +$$ + +Theorem 2 states that when comparing two subgroups, if the difference in their structural distance to the training set is larger than the threshold $(\delta \alpha)$ , then GNNs exhibit unfair generalization performance with respect to these groups. Specifically, the theorem states that GNNs generalize better on subgroups that are structurally closer to the training set, which leads to accuracy disparities among different subgroups. Additionally, the theorem notes that the topology awareness of GNNs (measured by the distortion rate $\alpha$ ) plays a key role in determining the strength of this accuracy disparity. This has important implications for the use of GNNs in critical applications, as it suggests that increased topology awareness in GNNs can lead to improved generalization performance but also a greater potential for unfairness in terms of accuracy disparities among different structural subgroups. As such, it is important to carefully consider the trade-offs between topology awareness and fairness when designing new GNNs and applying them to real-world problems. + +# 4.3 Connection with Deep Learning Spline Theory + +Our theoretical findings reveal that when considering the performance of the model from the viewpoint of the embedding space, the training set creates a "valley" effect on generalization performance, with the training vertex at the bottom of the valley. Representations that are closer to the embedding of the training set perform better, while those farther away perform worse. This pattern has an interesting connection with deep learning spline theory [2]. This theory suggests that in DNNs, the learning process for classification divides the input/latent space into small circles (referred to as power diagrams), where representations within each circle are grouped, and representations closer to the centre of the circle are predicted with more confidence. Our findings suggest that in GNNs, the mapping of test data to these power diagrams is related to the structural similarity between test and training data in the graph. This connection provides further validation for our derived results. + +![](images/81cad5d5b5b90ea693f888047329963422b316ff861f369dc6c95edcb6513a6f.jpg) +Fig.2. The loss landscape in embedding space induced by different training sets. When vertices $a$ and $d$ are selected as the training set instead of $b$ and $c$ , the test losses are smaller as the average distance between the training set and the other vertices is smaller (the neighbourhoods of $h_a$ and $h_d$ cover embeddings of more vertices in the embedding space). + +# 5 A Case Study on Shortest-Path Distance + +In this section, we present a case study which utilizes the shortest path distance to validate the predictions and effectiveness of our proposed framework. We aim to demonstrate how our framework can be applied to a real-world scenario and how it can be used to understand the generalization performance of GNNs on different structural subgroups. The case study will provide insights and show the potential impact of our results on various critical applications of GNNs. In this case study, we aim to answer empirically the following questions: Given a GNN model that can preserve certain structure $s$ ( $\alpha$ closed to 1), is the derived structural result observed on the generalization performance of the GNN model with respect to structural subgroups defined by $s$ ? Furthermore, how can our outcomes be harnessed to effectively tackle real-world challenges? + +Overall, our case study demonstrates that our proposed framework can be applied to a real-world scenario and provides useful insights into the performance of GNNs in different structural subgroups. It also shows that the results of our framework can be generalized for different train/test pairs. Moreover, we showcase its practical utility by employing the findings to address the cold start problem within graph active learning. + +General Setup. We use Deep Graph Library [43] for our implementation and experiment with four representative GNN models, GCN [17], Sage [13], GAT [42] and GCNII [7]. In all experiments, we follow closely the common GNN experimental setup, and multiple independent trials are conducted for each experiment to obtain average results. We adopt five popular and publicly available node classification datasets: Cora [34], Citeseer [21], CoraFull [4], CoauthorCS [36] and Ogbn-arxiv [16]. Due to space limitations, we only show part of the experimental results in the main body of the paper. Detailed descriptions of the datasets, hyperparameters and remaining experimental results can be found in the appendix. + +Structural Subgroups. We examine the accuracy disparity of subgroups defined by the shortest path distance (spd). In order to directly investigate the effect of structural distance on the generalization bound (Theorem 1), we first split the test nodes into subgroups by their shortest-path distance to the training set. For each test node $v$ , its distance to the training set $\mathcal{V}_0$ is given by $d_v = D_{\mathrm{spd}}(v, \mathcal{V}_0)$ with the shortest path distance metric as defined in Def. 2. Then we group the test nodes according to this distance. + +Distance Awareness of the Selected GNNs. We first verify if the selected GNNs preserve (are aware of) graph distances (with a small distortion rate), by evaluating the relation between the graph distances of test vertices to the training set and the embedding distances (Euclidean distance) of vertex representations to the representations of the training set, i.e., the relation between $d_{\mathrm{spd}}(v,\mathcal{V}_0)$ and $d_{\mathrm{Euclidean}}(\mathcal{M}(v),\mathcal{M}(\mathcal{V}_0))$ . In this experiment, we train the GNN models on the default training set provided in the dataset. We extract vertices within the 5-hop neighbourhood of the training set and group the vertices based on their distances to the training set. 5-hop neighbourhood is chosen because the diameter of most real-life networks is around 5 and it gives a better visual + +presentation. We then compute the average embedding distance of each group to the training set in the embedding space. Based on Def. 3, if an embedding has low distortion, we should expect to see linear relation with the slope being the scaling factor. As shown in Fig. 3, there exists a strong correlation between the graph distance and the embedding distance. The relation is near linear and the relative orders of distances are mostly preserved within the first few hops. This implies that the distortion $\alpha$ is small with the selected GNNs, i.e., they are distance aware. It also indicates that the premise of the small distortion in our theoretical results can be commonly satisfied in practice. + +# 5.1 Graph Distance and Generalization Performance + +Next, we study the relation between graph distance and generalization performance following the same settings as in the previous experiment. We compute the average prediction accuracy in each group of vertices. In Fig. 4, '0 hop' indicates the training set. We observe that the selected GNNs (with low distortion) indeed tend to generalize better on the group of vertices that are closer to the training set, which validates Theorem 2. + +We further validate if the distance between different pairs of the training sets and the test group is related to the generalization performance of GNNs in the test group. We randomly sample $k = 5$ vertices from each class as the training set $\mathcal{V}_0$ . Then we evaluate the GNN model on the rest of the vertices $\mathcal{V} \backslash \mathcal{V}_0$ . We make $k$ relatively small to enable a larger variance (among different trials) in the distance $D_{\mathrm{spd}}(\mathcal{V} \backslash \mathcal{V}_0, \mathcal{V}_0)$ (as given in Def. 2) between the training set and the rest of the vertices. We conduct 30 independent trials each selecting a different training set and ranking all the trials in descending order of the mean graph distance between the training set and the rest of the vertices. In Table 1, Group 1 contains the 10 trials with the largest graph distance $D_{\mathrm{spd}}(\mathcal{V} \backslash \mathcal{V}_0, \mathcal{V}_0)$ and Group 3 includes the 10 trials with the smallest graph distance $D_{\mathrm{spd}}(\mathcal{V} \backslash \mathcal{V}_0, \mathcal{V}_0)$ . The average classification accuracy and the variance (value in the bracket) are computed for each group. + +As shown in Table 1, training GNNs on a training set with a smaller graph distance $D_{\mathrm{spd}}(\mathcal{V} \backslash \mathcal{V}_0, \mathcal{V}_0)$ to the rest of the vertices, does generalize better, which validates our result and further enables us to use these insights for designing and developing new GNNs algorithms. + +# 5.2 Application in Cold Start Problem for Graph Active Learning + +Active learning is a type of machine learning where the model is actively involved in the labelling process of the training data. In active learning, the model is initially trained on a small labelled dataset and then used to select the most informative examples from the unlabeled dataset to be labelled next. Unsatisfactory initial set selection can lead to slow learning progress and unfair performance, which is referred to as the cold start problem [11]. Effective resolution of this issue entails an initial set selection that adheres to the following criteria: 1) exclusively leveraging graph structure; 2) achieving robust generalization; and + +![](images/946e531828fad46dc27183d8a11b403593b79f2805de9fbd550ec7b2cf3ac5a0.jpg) +(a) GCN, Cora + +![](images/8738f1566eeb92e9f5718d510477bcea0c7251885d9d7280a45710018f9a40ca.jpg) +(b) Sage, Cora + +![](images/a835381639d7517d35d2c14b2fa3da45149337efad5d434e0ca13a663ec4f0f7.jpg) +(c) GAT, Cora + +![](images/4e9209739cca6e1a6356318383c88c548519cb90f423c12b3b4df02b07ded9cc.jpg) +(d) GCN2, Cora + +![](images/6ce5e3918f3f657d548b573cf27a209d6ee92500da964e9b3006c2f6c6cafea7.jpg) +(e) GCN, Coauthor CS + +![](images/123f5f88dfabbbc933b0af2f1b8ab30efae89afa73bd4d3db9f1b12607ef3676.jpg) +(f) Sage, Coauthor CS + +![](images/f426896bffeba2d89cdf89c9722fe1e4ed86383f4ebb82342c9acf88f00fef2a.jpg) +(g) GAT, Coauthor +CS + +![](images/72b0fc676d728179f45ccd341544634676249e2a6b6695bb5a40ce957c28cd59.jpg) +(h) GCNII, Coauthor CS + +![](images/1cd4281aedfa3b7694dcc00e692f29b10c2808b28d514f55e47017674cebcf2d.jpg) +Fig. 3. Graph distance vs. embedding distance. (a)-(d) are the relation between graph distance (shortest-path distance) and embedding distance (Euclidean) of the four representative GNNs on Cora. (e)-(h) are the relation between graph distance and embedding distance of the four representative GNNs on Coauthor CS. The highlighted regions represent the $95\%$ confidence interval for that regression. (a)-(h) all demonstrate a rather linear relation between graph distance and embedding distance. +(a) Cora +Fig.4. Graph distance vs. accuracy. (a) is the relation between graph distance (shortest-path distance) and the accuracy of four representative GNNs on Cora. (b) is the relation between graph distance and the accuracy of four representative GNNs on Coauthor CS. Both (a) and (b) show that the graph distance is indeed a good indicator for generalization performance when GNNs are distance aware. + +![](images/fdf0a3bbf809c30410ce869893d47eb38ad60924425fa770cc533a82bcdcf6c2.jpg) +(b) Coauthor CS + +3) ensuring fairness (with minimal discrepancy in generalization performance) across distinct subgroups. We illustrate how our derived findings can be instrumental in the selection of an effective initial labelled dataset. + +Our empirical results show that the selected GNNs are highly distance-aware, meaning that their generalization performance is closely related to the coverage of the training set. To ensure that the model is fair and effective, the initial + +Table 1. Accuracy of 30 independent trials with different train/test pairs grouped by $D_{s}(\mathcal{V}\backslash \mathcal{V}_{0},\mathcal{V}_{0})$ . Group 1 contains the 10 trials with the largest graph distance $D_{\mathrm{spd}}(\mathcal{V}\backslash \mathcal{V}_0,\mathcal{V}_0)$ and Group 3 includes the 10 trials with the smallest graph distance $D_{\mathrm{spd}}(\mathcal{V}\backslash \mathcal{V}_0,\mathcal{V}_0)$ . The average accuracy and variance of the 10 trials are recorded in the table. The results show that train/test pairs with closer structural similarity indeed have better generalization performance, as predicted by our theoretical results. + +
CoraCiteseer
GCNGATSageGCNIIGCNGATSageGCNII
Group 165.42 (5.03)68.03 (4.62)67.61 (5.02)61.42 (4.03)59.72 (5.41)56.78 (4.73)58.52 (4.85)55.46 (3.69)
Group 267.56 (4.23)70.94 (3.86)69.54 (4.84)66.42 (3.76)61.48 (5.15)60.33 (4.24)60.78 (3.60)58.42 (4.08)
Group 371.40 (3.59)73.52 (3.53)72.60 (4.23)72.18 (4.01)62.93 (4.14)64.63 (3.65)63.03 (4.36)63.23 (5.06)
+ +Table 2. Initial data labelling: performance comparison of different GNNs sampling strategies based on graph structure. Results are averaged among five trials. For each dataset, we select $0.5\%$ of the total data as the initial set and evaluate on the rest of the vertices. Each entry corresponds to "test accuracy (ACC)| maximum discrepancy (MD)". $\uparrow$ means the higher the better and $\downarrow$ means the lower the better. Bold letter indicates the best sampling strategy for the given dataset and GNN model. + +
GCNSage
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)
R37.08|13.2336.35|10.238.13|3.2357.75|16.2337.69|12.0337.24|11.208.56|4.2049.72|15.52
Degree40.87|12.5236.27|9.2810.68|2.3251.50|18.2338.04|11.8735.26|12.0511.56|3.0242.78|13.02
Centrality42.75|10.4138.45|8.159.49|2.0853.48|16.1142.18|10.1540.81|9.0512.56|3.0845.65|10.11
PageRank47.87|9.5242.27|8.2512.48|2.1055.50|15.2350.15|9.1443.26|9.1513.15|2.9855.16|10.74
OURS49.03|8.2350.98|7.3215.25|1.2361.75|14.3258.57|8.2948.80|7.0314.76|2.0161.63|8.82
GATGCNII
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)|MD(↓)ACC(↑)|MD(↓)
R49.36|15.7843.53|16.279.55|3.0743.25|15.2031.53|14.0237.29|17.3210.98|3.0652.61|17.02
Degree54.08|13.5244.65|15.1710.50|2.7847.30|13.0432.19|13.8638.88|16.9810.40|2.7954.70|15.96
Centrality56.81|11.5248.15|13.0811.15|2.5248.50|13.1832.89|12.5739.98|16.0512.48|2.5256.88|15.18
PageRank60.91|11.0849.77|12.0712.94|2.1150.50|12.2334.04|12.0842.77|13.9514.56|2.1958.78|14.08
OURS62.76|10.5451.11|11.0713.39|1.8851.21|11.0235.61|11.0244.50|13.0415.19|1.9660.76|13.18
+ +labelled set should consist of vertices that are in close proximity to the rest of the input graph (in graph distance), as illustrated in Fig. 2. This can be formulated into the following optimization problem: + +$$ +\begin{array}{l} \min _ {\mathcal {V} _ {0} \subset \mathcal {V}} D _ {\operatorname {s p d}} (\mathcal {V} \backslash \mathcal {V} _ {0}, \mathcal {V} _ {0}) \\ s. t. \quad | \mathcal {V} _ {0} | = k \\ \end{array} +$$ + +where $k$ is the given initial set size and $D_{s}(.)$ is given in Def. 2. Intuitively, Eq.6 aims to select a subset of vertices that exhibit the closest structural similarity (as defined by the graph distance in this case) to the remaining vertices. This problem corresponds to the well-known k-center problem in graph theory [14], which is NP-hard. While finding the optimal solution may be intractable, efficient heuristics exist for tackling Eq. 6. Exploring the best heuristic is beyond the scope of this paper. Instead, for our experiment, we adopt the classic farthest-first traversal algorithm to obtain a solution. We refer to this sample selection procedure as the coverage-based sampling algorithm and provide a detailed description of the method used in the experiment within the supplementary material. + +We evaluate the initial set selected by our coverage-based sampling algorithm, comparing it to the existing baselines: (1) $\mathbf{R}$ : uniform random sampling [30], and importance sampling with graph properties (2) Degree [15], (3) Centrality, (4) Density and (5) PageRank [5]. We apply each strategy to select an initial dataset of $0.5\%$ of the total data for labelling (training). We evaluate the performance of the learnt model with the rest of the vertices on two metrics: 1) classification accuracy, which reflects the generalization performance, and 2) the maximum discrepancy among structural groups [1,8], which reflects the fairness. Let $R_{i}$ denote the mean classification accuracy of the vertex group that is $i$ -hop away from $\mathcal{V}_0$ . Then, the maximum discrepancy is computed as $\max \{|R_i - R_j||i,j\in 1,\dots,r\}$ , where $r$ is the maximum hop number. We use $r = 5$ to keep consistent with other experiments. The results in Table 2 show that coverage-based sampling achieves a significant improvement in model performance and fairness over the other strategies across different datasets, giving a better kick-start to the active learning process of GNNs. + +# 6 Conclusion and Discussion + +In this work, we study and investigate GNNs which are fundamental to many computer vision and machine learning tasks. In particular, we explore the relationship between the topology awareness and generalization performance of GNNs. We introduce a novel framework that connects the structural awareness of GNNs with approximate metric embedding, offering a fresh perspective on their generalization capabilities in semi-supervised node classification tasks. This structure-agnostic framework facilitates a deeper, more intuitive comprehension of GNN generalizability across varied graph structures. Through a case study centered on graph distance, we demonstrate that our theoretical findings regarding structural results are reflected in the practical generalization performance of GNNs. Additionally, our findings shed light on the cold start problem in graph active learning and could influence various significant GNN applications. + +# 6.1 Limitation and Future Works + +Our proposed framework introduces a novel angle on GNN generalization performance, yet it is not without limitations. The current approach interprets GNN embeddings in terms of metric mapping and describes the structure awareness of GNNs through distortion. This perspective, however, overlooks the specific dynamics that contribute to reduced distortion within GNNs. Investigating these underlying dynamics would significantly enrich this area of research. Moreover, our study primarily examines the transductive setting. Expanding this analysis to the inductive setting—where the model is trained on one graph but applied to another—would offer a more comprehensive view of GNN generalization capabilities. + +Acknowledgement. We would like to thank the anonymous reviewers for their helpful comments. JS and CW are supported by grants from Hong Kong RGC under the contracts HKU 17207621, 17203522 and C7004-22G (CRF). + +# References + +1. Ali, J., Babaei, M., Chakraborty, A., Mirzasoleiman, B., Gummadi, K.P., Singla, A.: On the fairness of time-critical influence maximization in social networks. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 1541-1542. IEEE (2022) +2. Balestriero, R., richard baraniuk: A spline theory of deep learning. In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 374-383. PMLR (10-15 Jul 2018), https://proceedings.mlr.press/v80/balestriero18b.htm1 +3. Baranwal, A., Fountoulakis, K., Jagannath, A.: Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization. arXiv preprint arXiv:2102.06966 (2021) +4. Bojchevski, A., Gunnemann, S.: Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking (2018) +5. Cai, H., Zheng, V.W., Chang, K.C.C.: Active learning for graph embedding. arXiv preprint arXiv:1705.05085 (2017) +6. Chen, J., Zhu, J., Song, L.: Stochastic training of graph convolutional networks with variance reduction. arXiv preprint arXiv:1710.10568 (2017) +7. Chen, M., Wei, Z., Huang, Z., Ding, B., Li, Y.: Simple and deep graph convolutional networks (2020) +8. Dong, Y., Ma, J., Chen, C., Li, J.: Fairness in graph mining: A survey (2022). https://doi.org/10.48550/ARXIV.2204.09888, https://arxiv.org/abs/2204.09888 +9. Esser, P., Chennuru Vankadara, L., Ghoshdastidar, D.: Learning theory can (sometimes) explain generalisation in graph neural networks. Advances in Neural Information Processing Systems 34, 27043-27056 (2021) +0. Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for quantum chemistry (2017) +11. Grimova, N., Macas, M., Gerla, V.: Addressing the cold start problem in active learning approach used for semi-automated sleep stages classification. In: 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). pp. 2249–2253. IEEE (2018) +12. Hamilton, W.L.: Graph representation learning. Synthesis Lectures on Artificial Intelligence and Machine Learning 14(3), 1-159 (2020) +13. Hamilton, W.L., Ying, R., Leskovec, J.: Inductive representation learning on large graphs (2018) +14. Hochbaum, D.S.: Approximating covering and packing problems: set cover, vertex cover, independent set, and related problems. In: Approximation algorithms for NP-hard problems, pp. 94-143 (1996) +15. Hu, S., Xiong, Z., Qu, M., Yuan, X., Côté, M.A., Liu, Z., Tang, J.: Graph policy network for transferable active learning on graphs. Advances in Neural Information Processing Systems 33, 10174-10185 (2020) +16. Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., Leskovec, J.: Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687 (2020) +17. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks (2017) + +18. Lao, D., Yang, X., Wu, Q., Yan, J.: Variational inference for training graph neural networks in low-data regime through joint structure-label estimation. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 824-834 (2022) +19. Leman, A., Weisfeiler, B.: A reduction of a graph to a canonical form and an algebra arising during this reduction. Nauchno-Technicheskaya Informatiya 2(9), 12-16 (1968) +20. Li, P., Wang, Y., Wang, H., Leskovec, J.: Distance encoding: Design provably more powerful neural networks for graph representation learning (2020) +21. Li, Q., Han, Z., Wu, X.M.: Deeper insights into graph convolutional networks for semi-supervised learning. In: Thirty-Second AAAI conference on artificial intelligence (2018) +22. Li, Y., Yu, R., Shahabi, C., Liu, Y.: Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926 (2017) +23. Liao, R., Urtasun, R., Zemel, R.: A {pac}-bayesian approach to generalization bounds for graph neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=TR-Nj6nFx42 +24. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors (2016) +25. Ma, J., Deng, J., Mei, Q.: Subgroup generalization and fairness of graph neural networks. arXiv preprint arXiv:2106.15535 (2021) +26. Ma, J., Ma, Z., Chai, J., Mei, Q.: Partition-based active learning for graph neural networks. arXiv preprint arXiv:2201.09391 (2022) +27. Morris, C., Ritzert, M., Fey, M., Hamilton, W.L., Lenssen, J.E., Rattan, G., Grohe, M.: Weisfeiler and leman go neural: Higher-order graph neural networks (2020) +28. Nishad, S., Agarwal, S., Bhattacharya, A., Ranu, S.: Graphreach: Position-aware graph neural network using reachability estimations (2021) +29. Oono, K., Suzuki, T.: Graph neural networks exponentially lose expressive power for node classification. arXiv preprint arXiv:1905.10947 (2019) +30. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Chen, X., Wang, X.: A survey of deep active learning (2020) +31. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021) +32. Sato, R.: A survey on the expressive power of graph neural networks. arXiv preprint arXiv:2003.04078 (2020) +33. Scarselli, F., Tsoi, A.C., Hagenbuchner, M.: The vapnik-chervonenkis dimension of graph and recursive neural networks. Neural Networks 108, 248-259 (2018) +34. Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., Eliassi-Rad, T.: Collective classification in network data. AI magazine 29(3), 93-93 (2008) +35. Sheng, G., Su, J., Huang, C., Wu, C.: Mspipe: Efficient temporal gnn training via staleness-aware pipeline. arXiv preprint arXiv:2402.15113 (2024) +36. Sinha, A., Shen, Z., Song, Y., Ma, H., Eide, D., Hsu, B.J., Wang, K.: An overview of microsoft academic service (mas) and applications. In: Proceedings of the 24th international conference on world wide web. pp. 243-246 (2015) +37. Su, J., Marbach, P.: Structure of core-periphery communities. In: International Conference on Complex Networks and Their Applications. pp. 151-161. Springer (2022) +38. Su, J., Marbach, P.: Structural properties of corea $\mathbf{E}$ periphery communities. Advances in Complex Systems (ACS) 26(06), 1-43 (2023) + +39. Su, J., Zou, D., Wu, C.: Pres: Toward scalable memory-based dynamic graph neural networks (2024), https://arxiv.org/abs/2402.04284 +40. Su, J., Zou, D., Zhang, Z., Wu, C.: Towards robust graph incremental learning on evolving graphs. In: International Conference on Machine Learning. pp. 32728-32748. PMLR (2023) +41. Tsubaki, M., Tomii, K., Sese, J.: Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics 35(2), 309-318 (2019) +42. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017) +43. Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y., et al.: Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315 (2019) +44. Wu, Q., Zhang, H., Yan, J., Wipf, D.: Handling distribution shifts on graphs: An invariance perspective (2022) +45. Wu, Y., Xu, Y., Singh, A., Yang, Y., Dubrawski, A.: Active learning for graph neural networks via node feature propagation. arXiv preprint arXiv:1910.07567 (2019) +46. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing (2017) +47. Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 (2018) +48. Yang, C., Wu, Q., Wang, J., Yan, J.: Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps. arXiv preprint arXiv:2212.09034 (2022) +49. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation (2022) +50. Yang, Z., Cohen, W.W., Salakhutdinov, R.: Revisiting semi-supervised learning with graph embeddings (2016). https://doi.org/10.48550/ARXIV.1603.08861, https://arxiv.org/abs/1603.08861 +51. You, J., Ying, R., Leskovec, J.: Position-aware graph neural networks (2019) +52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context (2018) +53. Zhou, K., Huang, X., Li, Y., Zha, D., Chen, R., Hu, X.: Towards deeper graph neural networks with differentiable group normalization (2020) +54. Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Feng, M., Zhao, X., Miao, Q., Shah, S.A.A., Bennamoun, M.: Scene graph generation: A comprehensive survey (2022) \ No newline at end of file diff --git a/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/images.zip b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5000c15c8b494adbb7605e470f6eb45ffcdae968 --- /dev/null +++ b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f44df61ee6acce33a4785f1d9a28d208ce9e18169e07427ed70dba34dd037999 +size 273251 diff --git a/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/layout.json b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8969c971a8f420b7389bdbc6a4d09010733373ae --- /dev/null +++ b/2024/On the Topology Awareness and Generalization Performance of Graph Neural Networks/layout.json @@ -0,0 +1,10020 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "type": "text", + "content": "On the Topology Awareness and Generalization Performance of Graph Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 246, + 168, + 367, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 168, + 367, + 179 + ], + "spans": [ + { + "bbox": [ + 246, + 168, + 367, + 179 + ], + "type": "text", + "content": "Junwei Su and Chuan Wu†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 177, + 190, + 436, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 190, + 436, + 213 + ], + "spans": [ + { + "bbox": [ + 177, + 190, + 436, + 213 + ], + "type": "text", + "content": "Department of Computer Science, the University of Hong Kong {jwsu, cwu} @cs.hku.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 241, + 458, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 241, + 458, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 241, + 458, + 506 + ], + "type": "text", + "content": "Abstract. Many computer vision and machine learning problems are modelled as learning tasks on graphs, where graph neural networks (GNNs) have emerged as a dominant tool for learning representations of graph-structured data. A key feature of GNNs is their use of graph structures as input, enabling them to exploit the graphs' inherent topological properties—known as the topology awareness of GNNs. Despite the empirical successes of GNNs, the influence of topology awareness on generalization performance remains unexplored, particularly for node-level tasks that diverge from the assumption of data being independent and identically distributed (I.I.D.). The precise definition and characterization of the topology awareness of GNNs, especially concerning different topological features, are still unclear. This paper introduces a comprehensive framework to characterize the topology awareness of GNNs across any topological feature. Using this framework, we investigate the effects of topology awareness on GNN generalization performance. Contrary to the prevailing belief that enhancing the topology awareness of GNNs is always advantageous, our analysis reveals a critical insight: improving the topology awareness of GNNs may inadvertently lead to unfair generalization across structural groups, which might not be desired in some scenarios. Additionally, we conduct a case study using the intrinsic graph metric, the shortest-path distance, on various benchmark datasets. The empirical results of this case study confirm our theoretical insights. Moreover, we demonstrate the practical applicability of our framework by using it to tackle the cold start problem in graph active learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "Many problems in computer vision and machine learning are modeled as learning tasks on graphs. For example, in semantic segmentation, graphs model the relationships between different image regions, enhancing accuracy and context-aware segmentation. Graph neural networks (GNNs) have emerged as a dominant class of machine learning models specifically designed for learning representations of graph-structured data. They have demonstrated considerable success in addressing a wide range of graph-related problems in various domains such as chemistry [10], biology [41], social networking [6,22,35], scene graph generation [49,54]" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 245, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 245, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 245, + 665 + ], + "type": "text", + "content": "†: corresponding authors" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "and visual relationship detection [24,46,52]. A defining characteristic of GNNs is their use of a spatial approach through message passing on the graph structure for feature aggregation. This enables GNNs to preserve structural information or dependencies (referred to as topology awareness) from the underlying graph structure, allowing them to be highly effective in tasks such as node classification. Fig. 1 illustrates the overall learning process of GNNs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 348 + ], + "type": "text", + "content": "Despite their practicality and potential, there remains a lack of theoretical understanding about GNNs, particularly in the semi-supervised node classification setting where the dependencies among the data differ significantly from other machine learning models [25]. In this setting, the goal is to leverage relations, as captured by the graph structure, among the data and a small set of labelled nodes to predict labels for the remaining nodes. Most of the existing theoretical studies of GNNs have focused on the connection between the message-passing mechanism of GNNs and the Weisfeiler-Lehman isomorphism test [19], aiming to understand the ability of GNNs to differentiate different graph structures in the learned representations, known as the expressive power of GNNs. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [28,32,51]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 578 + ], + "type": "text", + "content": "However, as GNNs become more reliant on and sensitive (aware) of graph structure as input, they may exhibit different generalization performances towards certain structural subgroups (distinct data subsets grouped by structural similarity to the training set) within the data. The quantification of GNN generalization across distinct structural subgroups is termed structural subgroup generalization [25]. Such considerations are vital in GNN application and development. For instance, within protein-protein interaction networks, these structural subgroups could represent different molecular complexes, influencing the accuracy of interaction predictions. Similarly, understanding how the topology awareness of GNNs influences generalization is essential when devising sampling strategies for training. The extent to which the generalization performance of GNNs is influenced by specific structural features of graph data is critical in deciding the composition of training datasets. Despite its importance, an understanding of the relationship between the topology awareness of GNNs and its structural subgroup generalization is still lacking. Furthermore, characterizing the topology awareness of GNNs poses a challenge, especially considering that different domains and tasks may prioritize distinct structural aspects. Therefore, a versatile framework is needed to assess the topology awareness of GNNs in relation to various structures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "To address this gap, in this paper, we propose a novel framework based on approximate metric embedding to study the relationship between structural subgroup generalization and topology awareness of GNNs in the context of semi-supervised node classification. The proposed framework allows for the investigation of the structural subgroup generalization of GNNs with respect to different structural subgroups. More concretely, the main contributions of this work are summarized as follows." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 478, + 216 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 478, + 216 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 478, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 478, + 216 + ], + "type": "image", + "image_path": "70cc696bee85f371baf0fa14792b5a03e165bb9a7edfc83c0f5ffcc78b9fc29b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 259 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 259 + ], + "type": "text", + "content": "Fig. 1. An illustration of the learning process in a 2-layer GNN. The message-passing mechanism leverages the graph structure to aggregate information, thereby generating the representation/embedding " + }, + { + "bbox": [ + 130, + 225, + 482, + 259 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 130, + 225, + 482, + 259 + ], + "type": "text", + "content": " for the target vertex " + }, + { + "bbox": [ + 130, + 225, + 482, + 259 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 225, + 482, + 259 + ], + "type": "text", + "content": " (highlighted in red)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 282, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 282, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 136, + 282, + 479, + 353 + ], + "type": "text", + "content": "1. We propose a novel, structure-agnostic framework using approximate metric embedding to examine the interplay between GNNs' structural subgroup generalization and topology awareness. This framework is versatile, accommodating various structural measures like shortest-path distance, and requires only the corresponding structural measure. Its simplicity in estimating key factors makes it applicable and generalizable to a wide range of scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 354, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 354, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 136, + 354, + 480, + 473 + ], + "type": "text", + "content": "2. Through formal analysis within our framework, we establish a clear link between GNN topology awareness and their generalization performance (Theorem 1). We also demonstrate that while enhanced topology awareness boosts GNN expressiveness, it can result in uneven generalization performance, favouring subgroups more structurally similar to the training set (Theorem 2). Such structural property can be harmful (causing unfairness issues) or useful (informing design decisions) depending on the scenario. This challenges the prevailing belief that increased topology awareness universally benefits GNNs [28, 32, 51], emphasizing the importance of considering the relationship between topology awareness and generalization performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 474, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 474, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 136, + 474, + 480, + 557 + ], + "type": "text", + "content": "3. We validate our framework through a case study on shortest-path distance, highlighting its practicality and relevance. The results corroborate our theoretical findings, showing that GNNs with heightened awareness of shortest-path distances excel in classifying vertex groups closer to the training set. Moreover, we demonstrate how our findings can be applied to mitigate the cold start problem in graph active learning [11,15], highlighting the practical implications of our framework and results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 583, + 237, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 237, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 237, + 596 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 609, + 311, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 311, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 311, + 621 + ], + "type": "text", + "content": "2.1 Topology Awareness of GNNs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Modern GNN architectures take inspiration from the Weisfeiler-Lehman isomorphism test [12, 19], which employs the graph structure to propagate information. Consequently, a significant portion of existing literature on GNNs concentrates" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "on understanding their ability to differentiate various graph structures, known as the expressive power of GNNs [27-29, 47, 51, 53]. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties [37, 38] in the learned representation [28, 32, 51]. Nevertheless, the precise ramifications of heightened topology awareness on GNNs' generalization performance remain shrouded in ambiguity. The intricate variations within structures across different domains and tasks further compound this complexity, underscoring the necessity for a unified framework—akin to the one proposed in this paper—to shed light on these intricacies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 255, + 282, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 282, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 282, + 266 + ], + "type": "text", + "content": "2.2 Generalization of GNNs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 483, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 483, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 483, + 410 + ], + "type": "text", + "content": "Relatively few studies focus on understanding the generalization performance of GNNs in node-level tasks. [3,9,18,23,25,33,39,40,44,48] extend the conventional statistical learning or information-theoretic framework to GNNs, providing different generalization bounds based on concepts such as Vapnik-Chervonenkis dimension, mutual information, and algorithm stability. In light of this, despite empirical indications highlighting the critical role of GNNs' topology awareness in generalization performance [20,28,51], there remains limited understanding of the relationship between topology awareness of GNNs and their generalization. Our proposed framework addresses this gap by utilizing the distortion concept in the approximate metric embedding to connect the topology awareness of GNNs and their generalization performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 429, + 290, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 290, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 290, + 441 + ], + "type": "text", + "content": "2.3 Active Learning in Graph" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": "Active Learning (AL) seeks to enhance model performance within a limited labeling budget by identifying the most representative samples for annotation [31]. The AL methodology generally involves two phases: 1) selecting a small, preliminary subset of samples for labeling; 2) utilizing this labeled subset to iteratively discover the most informative sample for future labeling. The initial selection of an effective subset represents the cold-start problem in AL [11]. While AL has proven effective in settings with independent and identically distributed (i.i.d.) data, such as images and text, applying it to Graph Neural Networks (GNNs) poses unique challenges. These challenges stem from the necessity to account for node interdependencies and the computational complexity of iterative sample selection and retraining, especially when considering multi-hop neighbors. Recent AL methodologies for GNNs, such as AGE [5], FeatProp [45], and GraphPart [26], aim to overcome these obstacles by leveraging the graph structure and initial node features to boost efficiency and efficacy, although their success may significantly hinge on the quality of the initial features. In this work, we focus on the cold-start problem of graph active learning and introduce a novel approach (based on our analysis) that relies exclusively on graph structure, thus eliminating the reliance on feature quality at the onset of learning." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 220, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 220, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 220, + 127 + ], + "type": "text", + "content": "3 Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "In this section, we introduce our proposed framework, which aims to investigate the intricate relationship between topology awareness and generalization performance in GNNs. The main idea of the framework is to use metric distortion to characterize the topology awareness of GNNs. With this characterization, we are able to show the relation between topology awareness of GNNs and their generalization performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 230, + 258, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 258, + 242 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 258, + 242 + ], + "type": "text", + "content": "3.1 GNN Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E},\\mathbf{X})" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " be an undirected graph where " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{1,2,\\dots,n\\}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " is the set of " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " nodes and " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\subseteq \\mathcal{V}\\times \\mathcal{V}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " is the set of edges. Let " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{x_{1},x_{2},\\ldots ,x_{n}\\}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " denote the set of node attribute vectors corresponding to all the nodes in " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": ". More specifically, " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "x_{v}\\in \\mathbf{X}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " is an m-dimensional vector which captures the attribute values of node " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "v\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": ". In addition, each vertex " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "v\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " is associated with a node label " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "y_{v}\\in \\{1,\\dots,c\\}" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 251, + 482, + 322 + ], + "type": "text", + "content": " is the total number of classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": "For our analysis, we focus on the transductive node classification setting [50], where the graph " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " are observed prior to learning and labels are only available for a small subset of nodes " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0\\subset \\mathcal{V}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ". Without loss of generality, we treat " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " as fixed throughout our analysis and the randomness comes from the labels " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ". Given a small set of the labelled nodes, " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0\\subset \\mathcal{V}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ", the task of node-level semi-supervised learning is to learn a classifier " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "y_{v} = f(v)" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": ", from a hypothesis space " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " and perform it on the remaining unlabeled nodes. Let " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(.,.)\\mapsto \\mathbb{R}_{+}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " be a bounded continuous loss function. We denote by " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "R^{\\mathcal{L}}(f,\\mathcal{V})" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " the generalization risk of the classifier " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": " and it is defined as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 243, + 440, + 480, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 440, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 243, + 440, + 480, + 453 + ], + "type": "interline_equation", + "content": "R ^ {\\mathcal {L}} (f, \\mathcal {V}) = \\mathbb {E} _ {v \\sim \\mathcal {V}} [ \\mathcal {L} (f (v), y _ {v}) ] \\tag {1}", + "image_path": "ffcfebc096db579b7f59be430021841a4406c851f7358d5deda4fb0a11d1698e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "text", + "content": "In our case, the hypothesis space is captured by a GNN and an additional read-out function. GNNs are commonly used as an embedding function and adopt a neighbourhood aggregation scheme, where the representation of a node is learnt iteratively by aggregating representations of its neighbours. After " + }, + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "text", + "content": " iterations of aggregation, a node's representation captures the information within its " + }, + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "text", + "content": "-hop neighbourhood in the graph. Formally, the " + }, + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 463, + 482, + 545 + ], + "type": "text", + "content": "-th layer of a GNN can be written as follows," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 175, + 555, + 480, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 555, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 175, + 555, + 480, + 570 + ], + "type": "interline_equation", + "content": "h _ {u} ^ {(k)} = \\operatorname {U p d a t e} ^ {(k)} \\left(\\operatorname {A g g r e g a t e} ^ {(k)} \\left(\\left\\{h _ {v} ^ {(k - 1)}, \\forall v \\in \\mathcal {N} (u) \\cup \\{u \\} \\right\\}\\right)\\right), \\tag {2}", + "image_path": "cd37fd6fe328ded8082eb4025e41b3962b9f03c512699c7eb42d557dab810d46.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "h_u^{(k)}" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " is the representation vector of node u at the k-th layer. " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "h_u^{(0)}" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " is initialized to be " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "x_u" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(u)" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " is the set of nodes directly adjacent to u. Different GNNs are compositions of aggregation and update layers of different designs. A GNN is commonly interpreted as an embedding function that combines the graph structure and node features to learn a representation vector " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "h_v" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " of node v. Then the learnt representation is fed into a task-specific read-out function g (commonly a multilayer perceptron) for the prediction. " + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 579, + 482, + 667 + ], + "type": "text", + "content": " maps the embedding" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "vector " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "h_v" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": " to a label " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "y_v" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": " in the label set " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^c" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": ". The overall GNN prediction process can be viewed as a composition of " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "g \\circ \\mathcal{M}: \\mathcal{V} \\mapsto \\mathbb{R}^c" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": ". The overall GNN prediction process can be viewed as a composition of " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "g \\circ \\mathcal{M}: \\mathcal{V} \\mapsto \\mathbb{R}^c" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 265, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 265, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 265, + 184 + ], + "type": "text", + "content": "3.2 Structural Subgroup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "text", + "content": "To study the generalization performance of GNNs on different structural subgroups, we assume the existence of a generalized distance (structural measure) function, " + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "inline_equation", + "content": "d_{s}(u,v)" + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "text", + "content": ", that captures the similarity between two nodes " + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "inline_equation", + "content": "u,v" + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "text", + "content": " with respect to structure " + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "text", + "content": ". Furthermore, we assume that this structural measure function " + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "inline_equation", + "content": "d_{s}(u,v)" + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "text", + "content": " and the vertex set " + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 192, + 482, + 277 + ], + "type": "text", + "content": " form a metric space. A metric space is a fundamental mathematical object that allows for studying and measuring the relation among the elements in a given set, and it is formally defined as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "text", + "content": "Definition 1 (metric space). A metric space is a pair " + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q},d)" + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "text", + "content": " is a set and " + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "inline_equation", + "content": "d:\\mathcal{Q}\\times \\mathcal{Q}\\mapsto \\mathbb{R}" + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "text", + "content": " is a metric satisfying the following axioms (" + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "inline_equation", + "content": "x,y,z" + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "text", + "content": " are arbitrary points of " + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 131, + 284, + 483, + 321 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 331, + 376, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 331, + 376, + 403 + ], + "spans": [ + { + "bbox": [ + 239, + 331, + 376, + 403 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} (M 1) d (x, y) \\geq 0, \\\\ (M 2) d (x, x) = 0, \\\\ (M 3) d (x, y) > 0, x \\neq y, \\\\ (M 4) d (x, y) = d (y, x), \\\\ (M 5) d (x, y) + d (y, z) \\geq d (x, z). \\\\ \\end{array}", + "image_path": "db70adc96c6dde395270834d7e9c97308f18ad5d84b6304138b9358572e22da0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": "These properties ensure that the metric " + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "inline_equation", + "content": "d(.)" + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": " can be used to quantify the distance or dissimilarity between any two elements in the set " + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": ". In the context of graph learning, the metric space " + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "inline_equation", + "content": "(\\mathcal{V},d_s)" + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": " allows us to study the similarity between nodes in terms of the structural property " + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": " and measure the structural distance between different subgroups of the test set. One natural and commonly used example in the graph is the metric space formed by the vertex set and the shortest-path distance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "text", + "content": "Definition 2 (Structural Group Distance). Given a metric space " + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}, d_s)" + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "text", + "content": " with respect to structure " + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "text", + "content": " and two distinct sets (groups) " + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_1, \\mathcal{V}_2 \\subset \\mathcal{V}" + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "text", + "content": ", we define the structural distance " + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "inline_equation", + "content": "D_s" + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "text", + "content": " between an element " + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}" + }, + { + "bbox": [ + 131, + 506, + 482, + 544 + ], + "type": "text", + "content": " to a group as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 254, + 554, + 480, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 554, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 254, + 554, + 480, + 571 + ], + "type": "interline_equation", + "content": "D _ {s} (v, \\mathcal {V} _ {1}) = \\min _ {u \\in \\mathcal {V} _ {1}} d _ {s} (v, u) \\tag {3}", + "image_path": "1eae2de6ad137a9308cf09ed48b39839dedaff3e542a5a12751dc7d97acee971.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 580, + 369, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 369, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 369, + 593 + ], + "type": "text", + "content": "and the structural distance between the two groups as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 241, + 603, + 480, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 603, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 241, + 603, + 480, + 620 + ], + "type": "interline_equation", + "content": "D _ {s} \\left(\\mathcal {V} _ {1}, \\mathcal {V} _ {2}\\right) = \\max _ {v \\in \\mathcal {V} _ {1}} \\min _ {u \\in \\mathcal {V} _ {2}} d _ {s} (v, u) \\tag {4}", + "image_path": "28c7202c38a1dc3c03c4f7f19db2d8a097708035cea71f0656a51dc42232a9a7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": "It should be noted that the distance function in Def. 2 is not symmetrical, and Eq. 4 captures the maximal structural difference of vertices in " + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_1" + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_2" + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": ". To provide further insight into the concept of a structural subgroup, let's" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "consider a scenario where " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0\\subset \\mathcal{V}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " represents the training set and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "d_{s}(.)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " represents the shortest-path distance. In this context, we can partition the remaining vertices " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\backslash \\mathcal{V}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " into different structural subgroups based on their shortest-path distance to the nearest vertex in the training set. Specifically, we can define " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_k = \\{v\\in \\mathcal{V}\\backslash \\mathcal{V}_0|D_s(v,\\mathcal{V}_0) = k\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". However, it is important to emphasize that our analysis does not assume a specific partition scheme, and the framework remains adaptable to various scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 215, + 263, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 215, + 263, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 263, + 228 + ], + "type": "text", + "content": "3.3 Topology Awareness" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 281 + ], + "type": "text", + "content": "To characterize and formalize the topology awareness of GNNs, we introduce the concept of distortion, which measures the degree to which the graph structure is preserved in the embedding space. Mathematically, the distortion of a function between two general metric spaces is defined as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": "Definition 3 (distortion). Given two metric spaces " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q},d)" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q}',d')" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": " and a mapping " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "\\gamma :\\mathcal{Q}\\mapsto \\mathcal{Q}'" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q}(\\mathcal{Q}')" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": " is a set and " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "d(d^{\\prime})" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": " is a distance function)), " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": " is said to have a distortion " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "\\alpha \\geq 1" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": ", if there exists a constant " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "r > 0" + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 131, + 285, + 483, + 334 + ], + "type": "inline_equation", + "content": "\\forall u,v\\in \\mathcal{Q}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 334, + 388, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 334, + 388, + 346 + ], + "spans": [ + { + "bbox": [ + 227, + 334, + 388, + 346 + ], + "type": "interline_equation", + "content": "r d (u, v) \\leq d ^ {\\prime} (\\gamma (u), \\gamma (v)) \\leq \\alpha r d (u, v).", + "image_path": "c8a5acf44fdd5f7f6cac25e2902994e83141e2fddba786a2c0d42187a3114dc4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "content": "The definition above captures the maximum expansion and minimum contraction between the two metric spaces. In the extreme case when " + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "content": ", the mapping " + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "content": " is the so-called isometric mapping that perfectly preserves the distance between any two points in the space. In other words, for any two points " + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "content": " in the domain of the function, their distance " + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "inline_equation", + "content": "d(x,y)" + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "content": " in the original metric space is equal to the distance " + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "inline_equation", + "content": "d'(\\gamma(x), \\gamma(y))" + }, + { + "bbox": [ + 130, + 351, + 482, + 446 + ], + "type": "text", + "content": " in the target metric space (up to a constant scaling factor). If the distortion rate of the mapping is close to 1, then this mapping is said to have a low distortion rate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 604 + ], + "type": "text", + "content": "In the context of GNNs, evaluating distortion involves a straightforward comparison of embeddings in the embedding domain with the structural measures on the graph, as demonstrated in the case study. If the distortion rate between the graph domain and the embedding domain is close to 1, the resulting embedding would largely preserve the corresponding structural information from the graph space, thereby indicating high topology awareness. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. This concept of distortion is agnostic to specific structural configurations, enabling the study of diverse graph structures using appropriate distance functions. Moreover, its estimation can be readily achieved by comparing the embedding with the original structural measure, as illustrated later in the case study." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 619, + 234, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 619, + 234, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 619, + 234, + 631 + ], + "type": "text", + "content": "4 Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we present the main results derived from our proposed framework for studying the generalization and fairness of GNNs with respect to different" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "structural groups in a graph. We start by demonstrating the relationship between the generalization performance of the model and the structural distance between the subgroups and the training set (Theorem 1). We then show that such a structural relation can induce disparities in the generalization performance across different subgroups (Theorem 2). Finally, we connect these theoretical results to the recent deep learning spline line theory [2] to provide additional insights. Detailed proof of the results presented in this paper can be found in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 232, + 425, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 425, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 425, + 243 + ], + "type": "text", + "content": "4.1 Generalization Performance and Structural Distance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": "Theorem 1. Let " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}, d_s)" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": " be a metric space for structure " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": " be a given labelled training set and " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i \\subset \\mathcal{V} \\backslash \\mathcal{V}_0" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": " be an arbitrary test subgroup. Suppose " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": " is a GNN model with distortion " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": " with respect to the structural measure " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "d_s(:,.)" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 254, + 482, + 302 + ], + "type": "text", + "content": " are Lipschitz, we have that" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 312, + 481, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 312, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 197, + 312, + 481, + 326 + ], + "type": "interline_equation", + "content": "R ^ {\\mathcal {L}} \\left(g \\circ \\mathcal {M}, \\mathcal {V} _ {i}\\right) \\leq R ^ {\\mathcal {L}} \\left(g \\circ \\mathcal {M}, \\mathcal {V} _ {0}\\right) + \\mathcal {O} \\left(\\alpha D _ {s} \\left(\\mathcal {V} _ {i}, \\mathcal {V} _ {0}\\right)\\right). \\tag {5}", + "image_path": "dd05adb918031eada05a139bee26f9dbcfdd39067baf81fe55d821d3d4e34124.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 336, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 504 + ], + "type": "text", + "content": "Theorem 1 states that the generalization risk of a classifier with respect to a structural subgroup depends on the empirical training loss of the classifier, the structural distance between the subgroup and the training set, and the distortion (topology awareness) of GNNs. Such a relation provides a new perspective on understanding the link between topology awareness and generalization performance in GNNs. Prior research in this area has mainly focused on the connection between GNNs and the Weisfeiler-Lehman algorithm, emphasizing the ability of GNNs to differentiate various graph structures. However, Theorem 1 highlights a direct link between topology awareness and generalization performance, and a potential issue of fairness in GNNs, as it shows that the model's performance on different structural subgroups is influenced by their distance to the training set. Next, we extend the result of Theorem 1 to analyze the relation between the generalization of different structural subgroups and their distance to the training set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 524, + 465, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 465, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 465, + 537 + ], + "type": "text", + "content": "4.2 Unfair Generalization Performance on Structural Subgroups" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": "Theorem 2. Let " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}, d_s)" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " be a metric space for structure " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": ". Suppose " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " is a GNN model with small training loss on " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " and distortion " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " with respect to structural measure " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "d_s(:,.)" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " are smooth. Let " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i, \\mathcal{V}_j \\subset \\mathcal{V} \\backslash \\mathcal{V}_0" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " be two distinct subgroups in the local neighbourhood of the training set " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": ", then there exists constant " + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 131, + 546, + 483, + 607 + ], + "type": "text", + "content": " (independent of the choice of subgroups) such that if" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 617, + 367, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 617, + 367, + 630 + ], + "spans": [ + { + "bbox": [ + 249, + 617, + 367, + 630 + ], + "type": "interline_equation", + "content": "D _ {s} \\left(\\mathcal {V} _ {i}, \\mathcal {V} _ {0}\\right) \\geq \\alpha \\delta D _ {s} \\left(\\mathcal {V} _ {j}, \\mathcal {V} _ {0}\\right),", + "image_path": "be8395f813b482acf3ae1ec72792000ea6ae0f19b81d33d5880ceff422383eb9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 640, + 193, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 640, + 193, + 651 + ], + "spans": [ + { + "bbox": [ + 132, + 640, + 193, + 651 + ], + "type": "text", + "content": "we have that," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 236, + 652, + 378, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 652, + 378, + 666 + ], + "spans": [ + { + "bbox": [ + 236, + 652, + 378, + 666 + ], + "type": "interline_equation", + "content": "R ^ {\\mathcal {L}} (g \\circ \\mathcal {M}, \\mathcal {V} _ {i}) \\geq R ^ {\\mathcal {L}} (g \\circ \\mathcal {M}, \\mathcal {V} _ {j}).", + "image_path": "132b1ace0d0faca6ce531ff04fbdc3a66b99051e4b8364f1d9afdbbd1ffcd4be.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "Theorem 2 states that when comparing two subgroups, if the difference in their structural distance to the training set is larger than the threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "(\\delta \\alpha)" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ", then GNNs exhibit unfair generalization performance with respect to these groups. Specifically, the theorem states that GNNs generalize better on subgroups that are structurally closer to the training set, which leads to accuracy disparities among different subgroups. Additionally, the theorem notes that the topology awareness of GNNs (measured by the distortion rate " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ") plays a key role in determining the strength of this accuracy disparity. This has important implications for the use of GNNs in critical applications, as it suggests that increased topology awareness in GNNs can lead to improved generalization performance but also a greater potential for unfairness in terms of accuracy disparities among different structural subgroups. As such, it is important to carefully consider the trade-offs between topology awareness and fairness when designing new GNNs and applying them to real-world problems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 312, + 395, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 395, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 395, + 324 + ], + "type": "text", + "content": "4.3 Connection with Deep Learning Spline Theory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 299, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 299, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 299, + 665 + ], + "type": "text", + "content": "Our theoretical findings reveal that when considering the performance of the model from the viewpoint of the embedding space, the training set creates a \"valley\" effect on generalization performance, with the training vertex at the bottom of the valley. Representations that are closer to the embedding of the training set perform better, while those farther away perform worse. This pattern has an interesting connection with deep learning spline theory [2]. This theory suggests that in DNNs, the learning process for classification divides the input/latent space into small circles (referred to as power diagrams), where representations within each circle are grouped, and representations closer to the centre of the circle are predicted with more confidence. Our findings suggest that in GNNs, the mapping of test data to these power diagrams is related to the structural similarity between test and training data in the graph. This connection provides further validation for our derived results." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 372, + 468, + 522 + ], + "blocks": [ + { + "bbox": [ + 317, + 372, + 468, + 522 + ], + "lines": [ + { + "bbox": [ + 317, + 372, + 468, + 522 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 468, + 522 + ], + "type": "image", + "image_path": "81cad5d5b5b90ea693f888047329963422b316ff861f369dc6c95edcb6513a6f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "lines": [ + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "text", + "content": "Fig.2. The loss landscape in embedding space induced by different training sets. When vertices " + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "text", + "content": " are selected as the training set instead of " + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "text", + "content": ", the test losses are smaller as the average distance between the training set and the other vertices is smaller (the neighbourhoods of " + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "inline_equation", + "content": "h_d" + }, + { + "bbox": [ + 304, + 547, + 482, + 647 + ], + "type": "text", + "content": " cover embeddings of more vertices in the embedding space)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 397, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 397, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 397, + 129 + ], + "type": "text", + "content": "5 A Case Study on Shortest-Path Distance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "text", + "content": "In this section, we present a case study which utilizes the shortest path distance to validate the predictions and effectiveness of our proposed framework. We aim to demonstrate how our framework can be applied to a real-world scenario and how it can be used to understand the generalization performance of GNNs on different structural subgroups. The case study will provide insights and show the potential impact of our results on various critical applications of GNNs. In this case study, we aim to answer empirically the following questions: Given a GNN model that can preserve certain structure " + }, + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "text", + "content": " closed to 1), is the derived structural result observed on the generalization performance of the GNN model with respect to structural subgroups defined by " + }, + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 139, + 482, + 270 + ], + "type": "text", + "content": "? Furthermore, how can our outcomes be harnessed to effectively tackle real-world challenges?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 342 + ], + "type": "text", + "content": "Overall, our case study demonstrates that our proposed framework can be applied to a real-world scenario and provides useful insights into the performance of GNNs in different structural subgroups. It also shows that the results of our framework can be generalized for different train/test pairs. Moreover, we showcase its practical utility by employing the findings to address the cold start problem within graph active learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "text", + "content": "General Setup. We use Deep Graph Library [43] for our implementation and experiment with four representative GNN models, GCN [17], Sage [13], GAT [42] and GCNII [7]. In all experiments, we follow closely the common GNN experimental setup, and multiple independent trials are conducted for each experiment to obtain average results. We adopt five popular and publicly available node classification datasets: Cora [34], Citeseer [21], CoraFull [4], CoauthorCS [36] and Ogbn-arxiv [16]. Due to space limitations, we only show part of the experimental results in the main body of the paper. Detailed descriptions of the datasets, hyperparameters and remaining experimental results can be found in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": "Structural Subgroups. We examine the accuracy disparity of subgroups defined by the shortest path distance (spd). In order to directly investigate the effect of structural distance on the generalization bound (Theorem 1), we first split the test nodes into subgroups by their shortest-path distance to the training set. For each test node " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": ", its distance to the training set " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "d_v = D_{\\mathrm{spd}}(v, \\mathcal{V}_0)" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": " with the shortest path distance metric as defined in Def. 2. Then we group the test nodes according to this distance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Distance Awareness of the Selected GNNs. We first verify if the selected GNNs preserve (are aware of) graph distances (with a small distortion rate), by evaluating the relation between the graph distances of test vertices to the training set and the embedding distances (Euclidean distance) of vertex representations to the representations of the training set, i.e., the relation between " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{spd}}(v,\\mathcal{V}_0)" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{Euclidean}}(\\mathcal{M}(v),\\mathcal{M}(\\mathcal{V}_0))" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ". In this experiment, we train the GNN models on the default training set provided in the dataset. We extract vertices within the 5-hop neighbourhood of the training set and group the vertices based on their distances to the training set. 5-hop neighbourhood is chosen because the diameter of most real-life networks is around 5 and it gives a better visual" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "presentation. We then compute the average embedding distance of each group to the training set in the embedding space. Based on Def. 3, if an embedding has low distortion, we should expect to see linear relation with the slope being the scaling factor. As shown in Fig. 3, there exists a strong correlation between the graph distance and the embedding distance. The relation is near linear and the relative orders of distances are mostly preserved within the first few hops. This implies that the distortion " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is small with the selected GNNs, i.e., they are distance aware. It also indicates that the premise of the small distortion in our theoretical results can be commonly satisfied in practice." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 240, + 407, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 240, + 407, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 407, + 253 + ], + "type": "text", + "content": "5.1 Graph Distance and Generalization Performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 330 + ], + "type": "text", + "content": "Next, we study the relation between graph distance and generalization performance following the same settings as in the previous experiment. We compute the average prediction accuracy in each group of vertices. In Fig. 4, '0 hop' indicates the training set. We observe that the selected GNNs (with low distortion) indeed tend to generalize better on the group of vertices that are closer to the training set, which validates Theorem 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": "We further validate if the distance between different pairs of the training sets and the test group is related to the generalization performance of GNNs in the test group. We randomly sample " + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": " vertices from each class as the training set " + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": ". Then we evaluate the GNN model on the rest of the vertices " + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\backslash \\mathcal{V}_0" + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": ". We make " + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": " relatively small to enable a larger variance (among different trials) in the distance " + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)" + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": " (as given in Def. 2) between the training set and the rest of the vertices. We conduct 30 independent trials each selecting a different training set and ranking all the trials in descending order of the mean graph distance between the training set and the rest of the vertices. In Table 1, Group 1 contains the 10 trials with the largest graph distance " + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)" + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": " and Group 3 includes the 10 trials with the smallest graph distance " + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)" + }, + { + "bbox": [ + 130, + 331, + 482, + 486 + ], + "type": "text", + "content": ". The average classification accuracy and the variance (value in the bracket) are computed for each group." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "content": "As shown in Table 1, training GNNs on a training set with a smaller graph distance " + }, + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{spd}}(\\mathcal{V} \\backslash \\mathcal{V}_0, \\mathcal{V}_0)" + }, + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "content": " to the rest of the vertices, does generalize better, which validates our result and further enables us to use these insights for designing and developing new GNNs algorithms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 551, + 474, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 474, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 474, + 563 + ], + "type": "text", + "content": "5.2 Application in Cold Start Problem for Graph Active Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Active learning is a type of machine learning where the model is actively involved in the labelling process of the training data. In active learning, the model is initially trained on a small labelled dataset and then used to select the most informative examples from the unlabeled dataset to be labelled next. Unsatisfactory initial set selection can lead to slow learning progress and unfair performance, which is referred to as the cold start problem [11]. Effective resolution of this issue entails an initial set selection that adheres to the following criteria: 1) exclusively leveraging graph structure; 2) achieving robust generalization; and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 126, + 210, + 190 + ], + "blocks": [ + { + "bbox": [ + 135, + 126, + 210, + 190 + ], + "lines": [ + { + "bbox": [ + 135, + 126, + 210, + 190 + ], + "spans": [ + { + "bbox": [ + 135, + 126, + 210, + 190 + ], + "type": "image", + "image_path": "946e531828fad46dc27183d8a11b403593b79f2805de9fbd550ec7b2cf3ac5a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 191, + 208, + 202 + ], + "lines": [ + { + "bbox": [ + 145, + 191, + 208, + 202 + ], + "spans": [ + { + "bbox": [ + 145, + 191, + 208, + 202 + ], + "type": "text", + "content": "(a) GCN, Cora" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 126, + 296, + 191 + ], + "blocks": [ + { + "bbox": [ + 222, + 126, + 296, + 191 + ], + "lines": [ + { + "bbox": [ + 222, + 126, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 222, + 126, + 296, + 191 + ], + "type": "image", + "image_path": "8738f1566eeb92e9f5718d510477bcea0c7251885d9d7280a45710018f9a40ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 192, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 234, + 192, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 234, + 192, + 294, + 203 + ], + "type": "text", + "content": "(b) Sage, Cora" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 126, + 383, + 191 + ], + "blocks": [ + { + "bbox": [ + 309, + 126, + 383, + 191 + ], + "lines": [ + { + "bbox": [ + 309, + 126, + 383, + 191 + ], + "spans": [ + { + "bbox": [ + 309, + 126, + 383, + 191 + ], + "type": "image", + "image_path": "a835381639d7517d35d2c14b2fa3da45149337efad5d434e0ca13a663ec4f0f7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 192, + 381, + 203 + ], + "lines": [ + { + "bbox": [ + 320, + 192, + 381, + 203 + ], + "spans": [ + { + "bbox": [ + 320, + 192, + 381, + 203 + ], + "type": "text", + "content": "(c) GAT, Cora" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 396, + 126, + 470, + 191 + ], + "blocks": [ + { + "bbox": [ + 396, + 126, + 470, + 191 + ], + "lines": [ + { + "bbox": [ + 396, + 126, + 470, + 191 + ], + "spans": [ + { + "bbox": [ + 396, + 126, + 470, + 191 + ], + "type": "image", + "image_path": "4e9209739cca6e1a6356318383c88c548519cb90f423c12b3b4df02b07ded9cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 192, + 471, + 203 + ], + "lines": [ + { + "bbox": [ + 405, + 192, + 471, + 203 + ], + "spans": [ + { + "bbox": [ + 405, + 192, + 471, + 203 + ], + "type": "text", + "content": "(d) GCN2, Cora" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 219, + 209, + 284 + ], + "blocks": [ + { + "bbox": [ + 135, + 219, + 209, + 284 + ], + "lines": [ + { + "bbox": [ + 135, + 219, + 209, + 284 + ], + "spans": [ + { + "bbox": [ + 135, + 219, + 209, + 284 + ], + "type": "image", + "image_path": "6ce5e3918f3f657d548b573cf27a209d6ee92500da964e9b3006c2f6c6cafea7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 285, + 219, + 306 + ], + "lines": [ + { + "bbox": [ + 133, + 285, + 219, + 306 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 219, + 306 + ], + "type": "text", + "content": "(e) GCN, Coauthor CS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 222, + 220, + 296, + 284 + ], + "blocks": [ + { + "bbox": [ + 222, + 220, + 296, + 284 + ], + "lines": [ + { + "bbox": [ + 222, + 220, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 222, + 220, + 296, + 284 + ], + "type": "image", + "image_path": "123f5f88dfabbbc933b0af2f1b8ab30efae89afa73bd4d3db9f1b12607ef3676.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 285, + 305, + 306 + ], + "lines": [ + { + "bbox": [ + 220, + 285, + 305, + 306 + ], + "spans": [ + { + "bbox": [ + 220, + 285, + 305, + 306 + ], + "type": "text", + "content": "(f) Sage, Coauthor CS" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 220, + 383, + 284 + ], + "blocks": [ + { + "bbox": [ + 309, + 220, + 383, + 284 + ], + "lines": [ + { + "bbox": [ + 309, + 220, + 383, + 284 + ], + "spans": [ + { + "bbox": [ + 309, + 220, + 383, + 284 + ], + "type": "image", + "image_path": "f426896bffeba2d89cdf89c9722fe1e4ed86383f4ebb82342c9acf88f00fef2a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 285, + 392, + 306 + ], + "lines": [ + { + "bbox": [ + 309, + 285, + 392, + 306 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 392, + 306 + ], + "type": "text", + "content": "(g) GAT, Coauthor \nCS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 396, + 220, + 470, + 284 + ], + "blocks": [ + { + "bbox": [ + 396, + 220, + 470, + 284 + ], + "lines": [ + { + "bbox": [ + 396, + 220, + 470, + 284 + ], + "spans": [ + { + "bbox": [ + 396, + 220, + 470, + 284 + ], + "type": "image", + "image_path": "72b0fc676d728179f45ccd341544634676249e2a6b6695bb5a40ce957c28cd59.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 285, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 395, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 395, + 285, + 480, + 306 + ], + "type": "text", + "content": "(h) GCNII, Coauthor CS" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 148, + 409, + 302, + 487 + ], + "blocks": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "text", + "content": "Fig. 3. Graph distance vs. embedding distance. (a)-(d) are the relation between graph distance (shortest-path distance) and embedding distance (Euclidean) of the four representative GNNs on Cora. (e)-(h) are the relation between graph distance and embedding distance of the four representative GNNs on Coauthor CS. The highlighted regions represent the " + }, + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "text", + "content": " confidence interval for that regression. (a)-(h) all demonstrate a rather linear relation between graph distance and embedding distance." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 409, + 302, + 487 + ], + "lines": [ + { + "bbox": [ + 148, + 409, + 302, + 487 + ], + "spans": [ + { + "bbox": [ + 148, + 409, + 302, + 487 + ], + "type": "image", + "image_path": "1cd4281aedfa3b7694dcc00e692f29b10c2808b28d514f55e47017674cebcf2d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 490, + 241, + 501 + ], + "lines": [ + { + "bbox": [ + 205, + 490, + 241, + 501 + ], + "spans": [ + { + "bbox": [ + 205, + 490, + 241, + 501 + ], + "type": "text", + "content": "(a) Cora" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 515, + 482, + 571 + ], + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 571 + ], + "type": "text", + "content": "Fig.4. Graph distance vs. accuracy. (a) is the relation between graph distance (shortest-path distance) and the accuracy of four representative GNNs on Cora. (b) is the relation between graph distance and the accuracy of four representative GNNs on Coauthor CS. Both (a) and (b) show that the graph distance is indeed a good indicator for generalization performance when GNNs are distance aware." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 313, + 409, + 467, + 487 + ], + "blocks": [ + { + "bbox": [ + 313, + 409, + 467, + 487 + ], + "lines": [ + { + "bbox": [ + 313, + 409, + 467, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 467, + 487 + ], + "type": "image", + "image_path": "fdf0a3bbf809c30410ce869893d47eb38ad60924425fa770cc533a82bcdcf6c2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 490, + 423, + 501 + ], + "lines": [ + { + "bbox": [ + 353, + 490, + 423, + 501 + ], + "spans": [ + { + "bbox": [ + 353, + 490, + 423, + 501 + ], + "type": "text", + "content": "(b) Coauthor CS" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 594, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 628 + ], + "type": "text", + "content": "3) ensuring fairness (with minimal discrepancy in generalization performance) across distinct subgroups. We illustrate how our derived findings can be instrumental in the selection of an effective initial labelled dataset." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Our empirical results show that the selected GNNs are highly distance-aware, meaning that their generalization performance is closely related to the coverage of the training set. To ensure that the model is fair and effective, the initial" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 188, + 481, + 230 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "type": "text", + "content": "Table 1. Accuracy of 30 independent trials with different train/test pairs grouped by " + }, + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "type": "inline_equation", + "content": "D_{s}(\\mathcal{V}\\backslash \\mathcal{V}_{0},\\mathcal{V}_{0})" + }, + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "type": "text", + "content": ". Group 1 contains the 10 trials with the largest graph distance " + }, + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{spd}}(\\mathcal{V}\\backslash \\mathcal{V}_0,\\mathcal{V}_0)" + }, + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "type": "text", + "content": " and Group 3 includes the 10 trials with the smallest graph distance " + }, + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{spd}}(\\mathcal{V}\\backslash \\mathcal{V}_0,\\mathcal{V}_0)" + }, + { + "bbox": [ + 132, + 113, + 482, + 180 + ], + "type": "text", + "content": ". The average accuracy and variance of the 10 trials are recorded in the table. The results show that train/test pairs with closer structural similarity indeed have better generalization performance, as predicted by our theoretical results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 188, + 481, + 230 + ], + "lines": [ + { + "bbox": [ + 136, + 188, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 136, + 188, + 481, + 230 + ], + "type": "table", + "html": "
CoraCiteseer
GCNGATSageGCNIIGCNGATSageGCNII
Group 165.42 (5.03)68.03 (4.62)67.61 (5.02)61.42 (4.03)59.72 (5.41)56.78 (4.73)58.52 (4.85)55.46 (3.69)
Group 267.56 (4.23)70.94 (3.86)69.54 (4.84)66.42 (3.76)61.48 (5.15)60.33 (4.24)60.78 (3.60)58.42 (4.08)
Group 371.40 (3.59)73.52 (3.53)72.60 (4.23)72.18 (4.01)62.93 (4.14)64.63 (3.65)63.03 (4.36)63.23 (5.06)
", + "image_path": "d6b2a12bdd7827705ee4621c6f13fc147917d7fa0d2ebc84d4eea1933aea046d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 317, + 481, + 426 + ], + "blocks": [ + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "type": "text", + "content": "Table 2. Initial data labelling: performance comparison of different GNNs sampling strategies based on graph structure. Results are averaged among five trials. For each dataset, we select " + }, + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "type": "text", + "content": " of the total data as the initial set and evaluate on the rest of the vertices. Each entry corresponds to \"test accuracy (ACC)| maximum discrepancy (MD)\". " + }, + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "type": "text", + "content": " means the higher the better and " + }, + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 132, + 242, + 482, + 308 + ], + "type": "text", + "content": " means the lower the better. Bold letter indicates the best sampling strategy for the given dataset and GNN model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 317, + 481, + 426 + ], + "lines": [ + { + "bbox": [ + 136, + 317, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 136, + 317, + 481, + 426 + ], + "type": "table", + "html": "
GCNSage
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)
R37.08|13.2336.35|10.238.13|3.2357.75|16.2337.69|12.0337.24|11.208.56|4.2049.72|15.52
Degree40.87|12.5236.27|9.2810.68|2.3251.50|18.2338.04|11.8735.26|12.0511.56|3.0242.78|13.02
Centrality42.75|10.4138.45|8.159.49|2.0853.48|16.1142.18|10.1540.81|9.0512.56|3.0845.65|10.11
PageRank47.87|9.5242.27|8.2512.48|2.1055.50|15.2350.15|9.1443.26|9.1513.15|2.9855.16|10.74
OURS49.03|8.2350.98|7.3215.25|1.2361.75|14.3258.57|8.2948.80|7.0314.76|2.0161.63|8.82
GATGCNII
CoraCiteSeerCoraFullCoauthor CSCoraCiteSeerCoraFullCoauthor CS
ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)MD(↓)ACC(↑)|MD(↓)ACC(↑)|MD(↓)
R49.36|15.7843.53|16.279.55|3.0743.25|15.2031.53|14.0237.29|17.3210.98|3.0652.61|17.02
Degree54.08|13.5244.65|15.1710.50|2.7847.30|13.0432.19|13.8638.88|16.9810.40|2.7954.70|15.96
Centrality56.81|11.5248.15|13.0811.15|2.5248.50|13.1832.89|12.5739.98|16.0512.48|2.5256.88|15.18
PageRank60.91|11.0849.77|12.0712.94|2.1150.50|12.2334.04|12.0842.77|13.9514.56|2.1958.78|14.08
OURS62.76|10.5451.11|11.0713.39|1.8851.21|11.0235.61|11.0244.50|13.0415.19|1.9660.76|13.18
", + "image_path": "7265c3fc318d258418fbc93b45e956af0d999ec318db87980d59fae4950fab48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 489 + ], + "type": "text", + "content": "labelled set should consist of vertices that are in close proximity to the rest of the input graph (in graph distance), as illustrated in Fig. 2. This can be formulated into the following optimization problem:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 500, + 352, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 500, + 352, + 533 + ], + "spans": [ + { + "bbox": [ + 261, + 500, + 352, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\mathcal {V} _ {0} \\subset \\mathcal {V}} D _ {\\operatorname {s p d}} (\\mathcal {V} \\backslash \\mathcal {V} _ {0}, \\mathcal {V} _ {0}) \\\\ s. t. \\quad | \\mathcal {V} _ {0} | = k \\\\ \\end{array}", + "image_path": "7774f14c13f9f4487c099e67e4a281974378bcd30982339ce48da0eb72ff2e59.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is the given initial set size and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{s}(.)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is given in Def. 2. Intuitively, Eq.6 aims to select a subset of vertices that exhibit the closest structural similarity (as defined by the graph distance in this case) to the remaining vertices. This problem corresponds to the well-known k-center problem in graph theory [14], which is NP-hard. While finding the optimal solution may be intractable, efficient heuristics exist for tackling Eq. 6. Exploring the best heuristic is beyond the scope of this paper. Instead, for our experiment, we adopt the classic farthest-first traversal algorithm to obtain a solution. We refer to this sample selection procedure as the coverage-based sampling algorithm and provide a detailed description of the method used in the experiment within the supplementary material." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "We evaluate the initial set selected by our coverage-based sampling algorithm, comparing it to the existing baselines: (1) " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ": uniform random sampling [30], and importance sampling with graph properties (2) Degree [15], (3) Centrality, (4) Density and (5) PageRank [5]. We apply each strategy to select an initial dataset of " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " of the total data for labelling (training). We evaluate the performance of the learnt model with the rest of the vertices on two metrics: 1) classification accuracy, which reflects the generalization performance, and 2) the maximum discrepancy among structural groups [1,8], which reflects the fairness. Let " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " denote the mean classification accuracy of the vertex group that is " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "-hop away from " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". Then, the maximum discrepancy is computed as " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\max \\{|R_i - R_j||i,j\\in 1,\\dots,r\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " is the maximum hop number. We use " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "r = 5" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " to keep consistent with other experiments. The results in Table 2 show that coverage-based sampling achieves a significant improvement in model performance and fairness over the other strategies across different datasets, giving a better kick-start to the active learning process of GNNs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 311, + 312, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 312, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 312, + 324 + ], + "type": "text", + "content": "6 Conclusion and Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 334, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 479 + ], + "type": "text", + "content": "In this work, we study and investigate GNNs which are fundamental to many computer vision and machine learning tasks. In particular, we explore the relationship between the topology awareness and generalization performance of GNNs. We introduce a novel framework that connects the structural awareness of GNNs with approximate metric embedding, offering a fresh perspective on their generalization capabilities in semi-supervised node classification tasks. This structure-agnostic framework facilitates a deeper, more intuitive comprehension of GNN generalizability across varied graph structures. Through a case study centered on graph distance, we demonstrate that our theoretical findings regarding structural results are reflected in the practical generalization performance of GNNs. Additionally, our findings shed light on the cold start problem in graph active learning and could influence various significant GNN applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 492, + 309, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 309, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 309, + 504 + ], + "type": "text", + "content": "6.1 Limitation and Future Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "Our proposed framework introduces a novel angle on GNN generalization performance, yet it is not without limitations. The current approach interprets GNN embeddings in terms of metric mapping and describes the structure awareness of GNNs through distortion. This perspective, however, overlooks the specific dynamics that contribute to reduced distortion within GNNs. Investigating these underlying dynamics would significantly enrich this area of research. Moreover, our study primarily examines the transductive setting. Expanding this analysis to the inductive setting—where the model is trained on one graph but applied to another—would offer a more comprehensive view of GNN generalization capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgement. We would like to thank the anonymous reviewers for their helpful comments. JS and CW are supported by grants from Hong Kong RGC under the contracts HKU 17207621, 17203522 and C7004-22G (CRF)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 144, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 144, + 481, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 144, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 138, + 144, + 481, + 189 + ], + "type": "text", + "content": "1. Ali, J., Babaei, M., Chakraborty, A., Mirzasoleiman, B., Gummadi, K.P., Singla, A.: On the fairness of time-critical influence maximization in social networks. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 1541-1542. IEEE (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 190, + 481, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 481, + 234 + ], + "type": "text", + "content": "2. Balestriero, R., richard baraniuk: A spline theory of deep learning. In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 374-383. PMLR (10-15 Jul 2018), https://proceedings.mlr.press/v80/balestriero18b.htm1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 481, + 267 + ], + "type": "text", + "content": "3. Baranwal, A., Fountoulakis, K., Jagannath, A.: Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization. arXiv preprint arXiv:2102.06966 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 481, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 291 + ], + "type": "text", + "content": "4. Bojchevski, A., Gunnemann, S.: Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 292, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 481, + 313 + ], + "type": "text", + "content": "5. Cai, H., Zheng, V.W., Chang, K.C.C.: Active learning for graph embedding. arXiv preprint arXiv:1705.05085 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "type": "text", + "content": "6. Chen, J., Zhu, J., Song, L.: Stochastic training of graph convolutional networks with variance reduction. arXiv preprint arXiv:1710.10568 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 338, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 358 + ], + "type": "text", + "content": "7. Chen, M., Wei, Z., Huang, Z., Ding, B., Li, Y.: Simple and deep graph convolutional networks (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 360, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 481, + 392 + ], + "type": "text", + "content": "8. Dong, Y., Ma, J., Chen, C., Li, J.: Fairness in graph mining: A survey (2022). https://doi.org/10.48550/ARXIV.2204.09888, https://arxiv.org/abs/2204.09888" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 394, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 481, + 426 + ], + "type": "text", + "content": "9. Esser, P., Chennuru Vankadara, L., Ghoshdastidar, D.: Learning theory can (sometimes) explain generalisation in graph neural networks. Advances in Neural Information Processing Systems 34, 27043-27056 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 450 + ], + "type": "text", + "content": "0. Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for quantum chemistry (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 452, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 452, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 134, + 452, + 481, + 494 + ], + "type": "text", + "content": "11. Grimova, N., Macas, M., Gerla, V.: Addressing the cold start problem in active learning approach used for semi-automated sleep stages classification. In: 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). pp. 2249–2253. IEEE (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 495, + 481, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 495, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 134, + 495, + 481, + 517 + ], + "type": "text", + "content": "12. Hamilton, W.L.: Graph representation learning. Synthesis Lectures on Artificial Intelligence and Machine Learning 14(3), 1-159 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 518, + 481, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 518, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 134, + 518, + 481, + 540 + ], + "type": "text", + "content": "13. Hamilton, W.L., Ying, R., Leskovec, J.: Inductive representation learning on large graphs (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 542, + 481, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 542, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 134, + 542, + 481, + 574 + ], + "type": "text", + "content": "14. Hochbaum, D.S.: Approximating covering and packing problems: set cover, vertex cover, independent set, and related problems. In: Approximation algorithms for NP-hard problems, pp. 94-143 (1996)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 576, + 481, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 576, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 134, + 576, + 481, + 608 + ], + "type": "text", + "content": "15. Hu, S., Xiong, Z., Qu, M., Yuan, X., Côté, M.A., Liu, Z., Tang, J.: Graph policy network for transferable active learning on graphs. Advances in Neural Information Processing Systems 33, 10174-10185 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 609, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 609, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 609, + 481, + 642 + ], + "type": "text", + "content": "16. Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., Leskovec, J.: Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "17. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "text", + "content": "18. Lao, D., Yang, X., Wu, Q., Yan, J.: Variational inference for training graph neural networks in low-data regime through joint structure-label estimation. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 824-834 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "text", + "content": "19. Leman, A., Weisfeiler, B.: A reduction of a graph to a canonical form and an algebra arising during this reduction. Nauchno-Technicheskaya Informatiya 2(9), 12-16 (1968)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "20. Li, P., Wang, Y., Wang, H., Leskovec, J.: Distance encoding: Design provably more powerful neural networks for graph representation learning (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "21. Li, Q., Han, Z., Wu, X.M.: Deeper insights into graph convolutional networks for semi-supervised learning. In: Thirty-Second AAAI conference on artificial intelligence (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "22. Li, Y., Yu, R., Shahabi, C., Liu, Y.: Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "23. Liao, R., Urtasun, R., Zemel, R.: A {pac}-bayesian approach to generalization bounds for graph neural networks. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=TR-Nj6nFx42" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "24. Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with language priors (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "25. Ma, J., Deng, J., Mei, Q.: Subgroup generalization and fairness of graph neural networks. arXiv preprint arXiv:2106.15535 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "text", + "content": "26. Ma, J., Ma, Z., Chai, J., Mei, Q.: Partition-based active learning for graph neural networks. arXiv preprint arXiv:2201.09391 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "27. Morris, C., Ritzert, M., Fey, M., Hamilton, W.L., Lenssen, J.E., Rattan, G., Grohe, M.: Weisfeiler and leman go neural: Higher-order graph neural networks (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "28. Nishad, S., Agarwal, S., Bhattacharya, A., Ranu, S.: Graphreach: Position-aware graph neural network using reachability estimations (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "29. Oono, K., Suzuki, T.: Graph neural networks exponentially lose expressive power for node classification. arXiv preprint arXiv:1905.10947 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "30. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Chen, X., Wang, X.: A survey of deep active learning (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "31. Ren, P., Xiao, Y., Chang, X., Huang, P.Y., Li, Z., Gupta, B.B., Chen, X., Wang, X.: A survey of deep active learning. ACM computing surveys (CSUR) 54(9), 1-40 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "text", + "content": "32. Sato, R.: A survey on the expressive power of graph neural networks. arXiv preprint arXiv:2003.04078 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "33. Scarselli, F., Tsoi, A.C., Hagenbuchner, M.: The vapnik-chervonenkis dimension of graph and recursive neural networks. Neural Networks 108, 248-259 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "34. Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., Eliassi-Rad, T.: Collective classification in network data. AI magazine 29(3), 93-93 (2008)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "35. Sheng, G., Su, J., Huang, C., Wu, C.: Mspipe: Efficient temporal gnn training via staleness-aware pipeline. arXiv preprint arXiv:2402.15113 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "36. Sinha, A., Shen, Z., Song, Y., Ma, H., Eide, D., Hsu, B.J., Wang, K.: An overview of microsoft academic service (mas) and applications. In: Proceedings of the 24th international conference on world wide web. pp. 243-246 (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "37. Su, J., Marbach, P.: Structure of core-periphery communities. In: International Conference on Complex Networks and Their Applications. pp. 151-161. Springer (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "38. Su, J., Marbach, P.: Structural properties of corea" + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "periphery communities. Advances in Complex Systems (ACS) 26(06), 1-43 (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "J.Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 533 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "39. Su, J., Zou, D., Wu, C.: Pres: Toward scalable memory-based dynamic graph neural networks (2024), https://arxiv.org/abs/2402.04284" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "text", + "content": "40. Su, J., Zou, D., Zhang, Z., Wu, C.: Towards robust graph incremental learning on evolving graphs. In: International Conference on Machine Learning. pp. 32728-32748. PMLR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "41. Tsubaki, M., Tomii, K., Sese, J.: Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics 35(2), 309-318 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "42. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "43. Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y., et al.: Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "44. Wu, Q., Zhang, H., Yan, J., Wipf, D.: Handling distribution shifts on graphs: An invariance perspective (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "text", + "content": "45. Wu, Y., Xu, Y., Singh, A., Yang, Y., Dubrawski, A.: Active learning for graph neural networks via node feature propagation. arXiv preprint arXiv:1910.07567 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "text", + "content": "46. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "47. Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "48. Yang, C., Wu, Q., Wang, J., Yan, J.: Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps. arXiv preprint arXiv:2212.09034 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 412 + ], + "type": "text", + "content": "49. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "50. Yang, Z., Cohen, W.W., Salakhutdinov, R.: Revisiting semi-supervised learning with graph embeddings (2016). https://doi.org/10.48550/ARXIV.1603.08861, https://arxiv.org/abs/1603.08861" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 445, + 459, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 459, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 459, + 456 + ], + "type": "text", + "content": "51. You, J., Ying, R., Leskovec, J.: Position-aware graph neural networks (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "53. Zhou, K., Huang, X., Li, Y., Zha, D., Chen, R., Hu, X.: Towards deeper graph neural networks with differentiable group normalization (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "54. Zhu, G., Zhang, L., Jiang, Y., Dang, Y., Hou, H., Shen, P., Feng, M., Zhao, X., Miao, Q., Shah, S.A.A., Bennamoun, M.: Scene graph generation: A comprehensive survey (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Awareness Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_content_list.json b/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a323d8701a24f9fd3d9c34563857fc707581e05 --- /dev/null +++ b/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_content_list.json @@ -0,0 +1,1773 @@ +[ + { + "type": "text", + "text": "On the Utility of 3D Hand Poses for Action Recognition", + "text_level": 1, + "bbox": [ + 292, + 140, + 709, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Md Salman Shamil1, Dibyadip Chatterjee1, Fadime Sener2, Shugao Ma2, and Angela Yao1", + "bbox": [ + 269, + 210, + 730, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 National University of Singapore 2 Meta Reality Labs {salman, dibyadip, ayao}@comp.nus.edu.sg {famesener, shugao}@meta.com https://s-shamil.github.io/HandFormer/", + "bbox": [ + 344, + 253, + 656, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D hand pose is an underexplored modality for action recognition. Poses are compact yet informative and can greatly benefit applications with limited compute budgets. However, poses alone offer an incomplete understanding of actions, as they cannot fully capture objects and environments with which humans interact. We propose HandFormer, a novel multimodal transformer, to efficiently model hand-object interactions. HandFormer combines 3D hand poses at a high temporal resolution for fine-grained motion modeling with sparsely sampled RGB frames for encoding scene semantics. Observing the unique characteristics of hand poses, we temporally factorize hand modeling and represent each joint by its short-term trajectories. This factorized pose representation combined with sparse RGB samples is remarkably efficient and highly accurate. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods at $5 \\times$ fewer FLOPs. With RGB, we achieve new state-of-the-art performance on Assembly101 and H2O with significant improvements in egocentric action recognition.", + "bbox": [ + 261, + 364, + 738, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Skeleton-based action recognition $\\cdot$ 3D hand poses $\\cdot$ Multi-modal transformer", + "bbox": [ + 261, + 599, + 740, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 656, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The popularity of AR/VR headsets has driven interest in recognizing hand-object interactions, particularly through egocentric [14, 26] and multiview cameras [34, 51]. Such interactions are inherently fine-grained; recognizing them requires distinguishing subtle motions and object state changes. State-of-the-art methods for hand action recognition [23, 45, 47, 66] primarily rely on multi-or single-view RGB streams, which are computationally heavy and unsuitable for resource-constrained scenarios like AR/VR. Motivated by advancements in lightweight hand pose estimation methods leveraging monochrome cameras [27, 28, 46], and the evolution of low-dimensional sensors [36, 42] such as accelerometers, MMG, EMG, demonstrating real-time hand pose estimation, we advocate", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/541766225c4522e643c67e6585bbf26257512ede07be096f1ea02cfe6679d827.jpg", + "image_caption": [ + "Fig. 1: We densely sample 3D hand poses to understand fine-grained hand motions and sparsely sample RGB frames to capture the scene semantics. Our 3D hand poses are acquired from low-resolution monochrome cameras but can also come from wearable sensors, facilitating an efficient understanding of hand-object interactions. Video frames and hand poses in the figure are from Assembly101 [51]." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 781, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the use of 3D hand poses as an input modality for recognizing hand-object interactions. Hand poses are a compact yet informative representation that captures the motions and nuances of hand movements.", + "bbox": [ + 212, + 369, + 784, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing works on 3D pose-based action recognition have focused primarily on full-body skeletons [20,44,65,68]. Hand poses differ fundamentally from full-body skeletons. In full-body recognition datasets [41,52], the actions are predominantly static from a global perspective. The relative changes in joint or limb positions signify the action category, e.g., 'sitting down'. Conversely, hand joints typically move together for many actions and lack a static joint as a global reference [51], e.g., 'put down toy'. Full-body skeleton methods also benefit from modeling long-range spatiotemporal dependencies between joints [44], while this is less important for the hands as shown in Fig. 2 and in Sec. 3.", + "bbox": [ + 212, + 415, + 784, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the 3D pose alone cannot encode the hands' actions. Unlike the full-body case, where actions are self-contained by the sequence of poses, the hands are often manipulating objects [34, 46, 51]. Hand pose is excellent for identifying motions (verbs) but struggles with associated objects [51]. Therefore, supplementing pose data with visual context from images or videos is crucial for full semantic understanding. However, as noted earlier, using dense RGB frames contradicts our motivation for using hand poses.", + "bbox": [ + 212, + 551, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper introduces HandFormer, a novel and lightweight multimodal transformer that leverages dense 3D hand poses complemented with sparsely sampled RGB frames. To this end, we conceptualize an action as a sequence of short segments, termed micro-actions, which are analogous to words that form sentence-level complex actions. Each micro-action comprises a dense sequence of pose frames and a single RGB frame. As every hand joint moves in close spatial proximity, we encode the pose sequence from a Lagrangian view [49] and track each joint as an individual entity. HandFormer effectively models each micro-action whose features are then temporally aggregated for action classification. This formulation benefits from our observations on the spatiotemporal dynamics of hand movement and delivers strong performance. Additionally, limiting long-range spatiotemporal dependencies among joints and only using sparse RGB frame", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "features improves computational efficiency, making it well-suited for always-on, low-compute AR/VR applications.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are: (i) We analyze the differences between hand pose and full-body skeleton actions and design a novel pose sequence encoding that reflects hand-specific properties. (ii) We propose HandFormer that takes a sequence of dense 3D hand pose and sparse RGB frames as input in the form of micro-actions. (iii) HandFormer achieves state-of-the-art action recognition on H2O [34] and Assembly101 [51]. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods in Assembly101, incurring at least $5 \\times$ fewer FLOPs. (iv) We experimentally demonstrate how hand poses are crucial for multiview and egocentric action recognition.", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 337, + 387, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Action recognition. Video-based action recognition systems are well-developed with sophisticated 3D-CNN [9,23,57,67] or Video Transformer [2,3,43,47] architectures. However, they all bear significant computational expense for both feature extraction and motion modeling, either explicitly in the form of optical flow [9,33,55] or implicitly through the architecture [2,47]. Such designs are well-suited for high-facility applications but are not suitable for integration into lightweight systems. To this end, efficient video understanding is an active topic of research, with a focus on reducing expensive 3D operations [22,40,58,67], quadratic complexity of attention [32,43,47] and dropping tokens [4,21]. Additionally, densely-sampled RGB frames incur a high cost, yet limiting the temporal resolution will hinder fine-grained action understanding. We propose complementing 3D hand poses with sparsely-sampled RGB frames, developing a lightweight video understanding system.", + "bbox": [ + 212, + 371, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Skeleton-based Action Recognition. Skeleton-based action recognition has been tackled with hand-crafted features [60, 62], sequence models like RNNs and LSTMs [19, 69], and CNN-based methods that either employ temporal convolution on the pose sequence [56] or transform the skeleton data into pseudoimages to be processed with 2D or 3D convolutional networks [7, 20, 31]. Recent progress has been driven by GNN-based methods that exploit the graph structure of skeletal data to construct spatio-temporal graphs and perform graph convolutions [68, 70]. These methods often model functional links between joints that go beyond skeletal connectivity [39, 54] or expand the spatiotemporal receptive field [44]. Self-attention and transformer-based methods have also been proposed [48, 65, 71]. However, most existing methods are tailored for datasets involving full-body poses. Few works [24, 30, 37, 50] study hand poses but are restricted to simple gestures, recognizable without complex temporal modeling.", + "bbox": [ + 212, + 568, + 787, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fusing RGB with Skeleton. Skeletal data has been used for cropping images of body parts [12], for weighting RGB patches around regions of interest [5, 6], or for pooling CNN features [1, 8]. Projecting into common embedding space is done in [16], enabling pose distillation [15]. Multi-stream architectures are also designed with separate paths for RGB and skeleton having lateral connec", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89eca3f064d23587c0fdd50bc5047e89169b4718ca4e7379d4a58d99aed647b9.jpg", + "image_caption": [ + "Stand up" + ], + "image_footnote": [], + "bbox": [ + 222, + 181, + 264, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4854dfb504b50a8617f4ff20d3546cc7b4617e232960666444d5808d07a33c12.jpg", + "image_caption": [ + "Put on bag" + ], + "image_footnote": [], + "bbox": [ + 277, + 143, + 325, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e691409f9f81ee2def47f9d257d486d30f8c779640dcbd82165d358d0376828.jpg", + "image_caption": [ + "Sit down", + "Fig. 2: Comparing skeletal changes in full-body actions from NTU RGB+D 120 [41] (left) and hand actions from Assembly101 [51] (right). Two pose frames at interval $T$ are shown. $J_{j,t}$ indicates the 3D coordinate of joint $j$ at timestep $t$ , and $$ is the correlation between two such joints. Modeling the correlation between spatiotemporally distant joints can be informative for full-body poses but does not provide a useful action cue for hand poses." + ], + "image_footnote": [], + "bbox": [ + 277, + 207, + 323, + 258 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49650b0f05e7738b609bdb167cd0bb3e2c792ded9fa3c5bddb296200744cd313.jpg", + "image_caption": [ + "$\\leftarrow \\rightarrow \\rightarrow < J_{\\mathrm{H,0}}J_{\\mathrm{ET}}>$" + ], + "image_footnote": [], + "bbox": [ + 352, + 143, + 480, + 248 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dbe57dbbea64fd2cababef3f172ce5dfee47975988e570815a1e00177c984d97.jpg", + "image_caption": [ + "←→" + ], + "image_footnote": [], + "bbox": [ + 506, + 154, + 668, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8dda702072f573aa6a3d8fbbf3661b7ed2199189f1bec8c2ec040a74b816060b.jpg", + "image_caption": [ + "Pick up" + ], + "image_footnote": [], + "bbox": [ + 679, + 148, + 730, + 188 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ada46ecffa20204e520e1632ef3f414852d63775abcf0114a4f32976dea33fef.jpg", + "image_caption": [ + "Unscrew" + ], + "image_footnote": [], + "bbox": [ + 679, + 220, + 730, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f2bf0dc21d2efbfeb8adc3fa0d002dd6d5054614851679a0b50a12e9fea0d08.jpg", + "image_caption": [ + "Rotate" + ], + "image_footnote": [], + "bbox": [ + 733, + 185, + 784, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions between them [38, 63]. RGBPoseConv3D [20] is a two-stream model that employs 3D CNNs for both RGB and pose. Similar to skeleton-based action recognition, these multimodal approaches primarily focus on full-body poses. Some approaches simultaneously perform both hand pose estimation and action recognition, using pose data to supervise the training process [13, 64]. However, it is important to note that pose estimation deals with lower-level semantics than action recognition. Actions can often be inferred even when the estimated poses are not accurate [46].", + "bbox": [ + 212, + 392, + 787, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Modeling Full-body vs. Hand Skeletons", + "text_level": 1, + "bbox": [ + 215, + 536, + 633, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing skeleton-based action recognition datasets primarily consist of full-body poses where actions feature significant changes over time in limb positions relative to other body parts, which highly correlate to the action category. These changes can be captured through long-range spatiotemporal modeling using graph convolutions with large receptive fields [44] or self-attention [65]. Conversely, hand actions show diverse movement patterns, as hands can move in arbitrary directions, often without significant changes in articulation. We analyze pose sequences from NTU RGB+D 120 [41] and Assembly101 [51], calculating the distance covered by each joint and identifying the least and the most active joints. We observe that, for full-body poses, these two representative joints show a significant difference in the distances covered, while for hand poses, the difference is subtle. The Pearson correlation coefficient is 0.93 for hand poses (indicating highly coupled joints) and 0.33 for full-body poses. Please refer to Suppl. Sec. A for details regarding these computations and comparisons.", + "bbox": [ + 212, + 568, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An example showing the difference between body and hand poses is provided in Fig. 2, where we depict two frames separated by an interval of $T$ frames. For the full body skeleton, the head joint $J_{H}$ remains static, while the right hand joint $J_{F}$ moves upward. Conversely, for the hand, all the joints move,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 393, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/67d57c644999ed5a67f0c5cacd4a59432410b785efeeb0712f4f1315ba45e7c0.jpg", + "image_caption": [ + "Fig. 3: Overall architecture of HandFormer. An action segment is divided into $K$ micro-actions $\\{M_1, M_2 \\ldots M_K\\}$ . Each micro-action comprises a dense sequence of pose frames and a single RGB frame. The frame encoder $F$ and the trajectory encoder $T$ encode the RGB and the dense poses, respectively, after which they are passed to a Multimodal Tokenizer. The modality-mixed tokens are then fed to a Temporal Transformer. Dotted paths are optional and only required when RGB is used." + ], + "image_footnote": [], + "bbox": [ + 228, + 150, + 743, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "including the wrist joint $J_{W}$ and index fingertip $J_{I}$ . The spatiotemporal correlation between the head joint in the first frame and the hand joint in the last frame ( $$ ) is a crucial action cue that captures the relative structural change. However, such a correlation between the wrist and the fingertip for hand skeleton ( $$ ) does not provide a stronger cue than the wrist movement itself. To avoid such redundant spatiotemporally distant correlations, we propose to divide the hand pose sequence into micro-action blocks of fixed temporal length. This formulation allows for encoding short-term movements while enabling parameter sharing.", + "bbox": [ + 212, + 482, + 787, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, full-skeletal motion is a dominant feature in hand pose sequences, differentiating them from full-body poses. While the location and orientation of the human body can be trivial for most actions, this is not the case for hands. The hands do not conform to any particular 6D pose, and notably, they frequently and unpredictably alter their 6D poses over time. A comprehensive analysis of this phenomenon is provided in Suppl. Sec. A. To this end, we explicitly consider the global 6D poses of the hands during micro-action-based pose encoding.", + "bbox": [ + 212, + 619, + 787, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 HandFormer", + "text_level": 1, + "bbox": [ + 215, + 747, + 375, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 illustrates our proposed HandFormer, which consists of a sequence of micro-action blocks, a novel trajectory encoder, and a temporal aggregation module. The design of our HandFormer allows us to easily incorporate semantic context by sampling a single RGB frame from certain micro-actions.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Micro-actions", + "text_level": 1, + "bbox": [ + 215, + 146, + 375, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an action segment containing $\\mathcal{T}$ frames sampled at a certain fps, the input to our model comprises — $i$ ) a dense sequence of 3D hand poses, represented as $\\mathbf{S} = \\{P_1,P_2,\\dots ,P_{\\mathcal{T}}\\}$ , where $P_{t} = \\{P_{t}^{left},P_{t}^{right}\\} \\in \\mathbb{R}^{2\\times J\\times 3}$ signifying the 3D coordinates of $J$ keypoints in the left and right hands, respectively, and $ii)$ a sparse set of RGB frames sampled at intervals $\\varDelta f$ , denoted as $\\mathbf{V} = \\{I_1,I_{1 + \\varDelta f},\\ldots ,I_{\\mathcal{T}_r}\\}$ , where $I_{t}\\in \\mathbb{R}^{H\\times W\\times 3}$ are frame-wise RGB features and $\\mathcal{T}_r = 1 + \\left|\\frac{\\mathcal{T} - 1}{\\varDelta f}\\right|\\times \\varDelta f$ .", + "bbox": [ + 212, + 170, + 787, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We factorize the raw input into a sequence of $K$ micro-action blocks of length $N$ frames, obtained by shifting the window across the action segment with a stride of $R$ frames. Each block consists of two components — the initial appearance derived from the first RGB frame within the block and the hand motion characterized by the dense sequence of $N$ pose frames. To obtain a fixed length input containing $\\mathcal{T}' = (K - 1)\\times R + N$ pose frames, we perform linear interpolation for each joint along the temporal axis, transforming pose sequence $\\mathbf{S}$ having $\\mathcal{T}$ frames to $\\mathbf{S}' = \\{P_1',P_2',\\dots ,P_{\\mathcal{T}'}'\\}$ having $\\mathcal{T}'$ frames, where $P_t' \\in \\mathbb{R}^{2\\times J\\times 3}$ . Thus, we represent the input as a sequence of micro-actions $\\mathbf{M} = \\{M_1,M_2,\\dots ,M_K\\}$ , derived from $\\mathbf{V}$ and $\\mathbf{S}'$ through the following equations:", + "bbox": [ + 212, + 271, + 787, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM _ {k} = \\left[ M _ {k} ^ {\\mathrm {R G B}}, M _ {k} ^ {\\text {P o s e}} \\right] = \\left[ I _ {h (k)}, \\left\\{P _ {g (k) + i} ^ {\\prime} \\right\\} _ {i = 0} ^ {N - 1} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 431, + 785, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $g(k) = (k - 1)\\times R + 1$ denotes the first pose frame in $k$ -th Micro-action, and $h(k)$ determines the index of the nearest available RGB frame.", + "bbox": [ + 212, + 465, + 785, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The dense pose sequence in a micro-action captures fine-grained hand motion crucial for recognizing verbs, whereas a single RGB frame provides semantic context for recognizing objects. To extract features from micro-actions, we use a frame encoder $F$ and a trajectory encoder $T$ , which operate on RGB frames and pose sequences, respectively. Consequently, the RGB and pose features for the $k^{th}$ micro-action is given by-", + "bbox": [ + 212, + 497, + 787, + 587 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ f _ {k} ^ {\\mathrm {R G B}}, f _ {k} ^ {\\mathrm {P o s e}} \\right] = \\left[ F \\left(M _ {k} ^ {\\mathrm {R G B}}\\right), T \\left(M _ {k} ^ {\\mathrm {P o s e}}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 597, + 785, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{k}^{\\mathrm{RGB}}, f_{k}^{\\mathrm{pos}} \\in \\mathbb{R}^{d}$ , where $d$ denotes the common dimensionality of both RGB and pose embedding space.", + "bbox": [ + 212, + 626, + 785, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Trajectory Encoder", + "text_level": 1, + "bbox": [ + 215, + 679, + 423, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We devise a trajectory-based pose encoder to encode dense hand pose sequences within micro-actions, illustrated in Fig. 4. Each joint is represented by its trajectory of dimension $3 \\times N$ , encapsulating the sequence of 3D coordinates across the $N$ pose frames of a micro-action. This yields $2 \\times J$ feature vectors for the $J$ joints of two hands. Each joint's trajectory is passed through a TCN [35], whose parameters are shared for all joints. This produces $2 \\times J$ Local Trajectory Tokens. Additionally, the full-skeletal motion of the hand during the entire action is used as a reference through an additional token named Global Wrist Token. This token is generated by a separate TCN operating on the sequence of 6D", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7827b62edb38da93d51875c2a450e292076fc150ac7301369a31088c7b59593.jpg", + "image_caption": [ + "Fig. 4: Our Trajectory Encoder $T$ , which operates on micro-actions, derives tokens with trajectory-based features and performs self-attention to encode the pose sequence into a feature vector. The Single-Joint TCN is a Temporal Convolutional Network [35] that processes the trajectories of all the joints individually with shared parameters across all joints. Wrist-TCN takes an action-wide sequence of wrist location and hand orientation (6D pose) to produce a global reference token." + ], + "image_footnote": [], + "bbox": [ + 215, + 150, + 787, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "poses of hands, indicating wrist location and hand orientation. Subsequently, a self-attention layer is applied to these trajectory tokens, preserving the temporal dimension for subsequent stages. This iterative process culminates in spatiotemporal average pooling, summarizing the hand motion of the micro-action.", + "bbox": [ + 212, + 420, + 782, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Multimodal Tokenizer", + "text_level": 1, + "bbox": [ + 215, + 505, + 444, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section discusses incorporating sparsely sampled RGB frames into Hand-Former to better capture scene semantics. A single frame is sampled from each micro-action, generating an extended crop $(1.25\\times)$ around the hands. While the full image provides an overall scene context, the crop focuses specifically on hand-object interaction regions [10]. Features for both are separately generated using a pre-existing image encoder and then aggregated to enrich the hand-object interaction feature with scene context. The hand-object ROI crop can be obtained using an off-the-shelf HOI detector [53] or by using the corresponding hand pose projections. We opt for the latter.", + "bbox": [ + 212, + 532, + 784, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our multimodal tokenizer in Fig. 3 generates RGB and Pose tokens enhanced via multimodal interactions. This involves concatenating each frame feature and trajectory encoding, and projecting them into a shared PoseRGB feature space using an MLP. This PoseRGB feature is then split into two parts, which are added to the original frame feature and trajectory encoding, respectively.", + "bbox": [ + 212, + 667, + 784, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Temporal Transformer", + "text_level": 1, + "bbox": [ + 215, + 767, + 447, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The multimodal tokenizer provides RGB and pose tokens for each micro-action. Since an action segment consists of a sequence of micro-actions, these micro-action tokens are aggregated over time via a temporal transformer. A video $\\mathbf{V}$ ,", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "divided into $K$ micro-actions can be represented by two sets of tokens $\\{\\hat{f}_k^{\\mathrm{RGB}}\\}_{k=1}^K$ and $\\{\\hat{f}_k^{\\mathrm{Pose}}\\}_{k=1}^K$ , respectively, produced by the multimodal tokenizer. These $2 \\times K$ tokens of dimension $d$ form the input sequence of the temporal transformer. Positional encoding and modality embedding are added to each input token to indicate the temporal location and source modality. We use the fixed sine/cosine positional encoding [59], assigning the same position to two tokens from the same micro-action. Modality embeddings are learned and shared across the tokens from the same modality (RGB or Pose). Following standard practice [17,18], we preprocess an additional learnable class token $[\\mathrm{CLS}] \\in \\mathbb{R}^d$ , the output of which is then used for classifying actions.", + "bbox": [ + 212, + 143, + 787, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5 Learning Objectives", + "text_level": 1, + "bbox": [ + 215, + 320, + 429, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HandFormer is trained end-to-end for action recognition, supervised by a cross-entropy loss: $\\mathcal{L}_{cls} = -a_i\\log \\hat{a}_i$ , where $a_{i}$ represents the ground truth action label for the $i^{th}$ sample, and $\\hat{a}_i$ denotes the predicted action category. The input pose and RGB modalities provide complementary information regarding a scene, capturing motion and interacting objects, respectively. To effectively utilize this, we employ explicit verb and object supervision ( $\\mathcal{L}_{verb}$ , $\\mathcal{L}_{obj}$ ) via two additional learnable class tokens. Given that pose strongly correlates with verbs, whereas objects are identifiable from RGB frames alone, we let the verb class token attend exclusively to pose encodings and the object class token to frame features.", + "bbox": [ + 212, + 344, + 787, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Anticipation Loss. Hand pose sequence captures the primary sources of state changes during hand-object interactions. We posit that the visual state from an initial RGB frame, in combination with the subsequent hand pose sequence, is indicative of the visual state that results from the completion of the sequence. Therefore, given an RGB feature and the corresponding pose features from a micro-action, we force our model to anticipate the RGB feature for the next micro-action by minimizing an $L_{1}$ feature loss. This loss, inspired by existing efforts [25,61], quantifies the difference between the anticipated image feature and the true feature extracted from a frozen image encoder. Formally,", + "bbox": [ + 212, + 481, + 789, + 617 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a n t} = \\sum_ {k = 1} ^ {K - 1} \\left\\| \\Phi_ {\\text {a n t}} \\left(f _ {k} ^ {\\text {P o s e R G B}}\\right) - f _ {k + 1} ^ {\\text {R G B}} \\right\\| _ {1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 640, + 785, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varPhi_{\\mathrm{ant}}$ denotes a linear projection layer. Hence, the total loss is", + "bbox": [ + 214, + 689, + 710, + 705 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\lambda_ {1} \\mathcal {L} _ {v e r b} + \\lambda_ {2} \\mathcal {L} _ {o b j} + \\lambda_ {3} \\mathcal {L} _ {a n t} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 720, + 785, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3$ are hyperparameters used to balance the four losses.", + "bbox": [ + 214, + 741, + 710, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 777, + 375, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two publicly available hand-object interaction datasets. Assembly101 [51] is a large-scale multiview dataset that fea-", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tures videos of procedural activities for assembling and disassembling 101 take-aside toy vehicles. This dataset has 1380 fine-grained actions based on 24 verbs and 90 objects. It consists of 12 temporally synchronized views — eight static and four egocentric. The egocentric views exhibit low-resolution monochrome images, making it challenging even for human eyes to discern objects in view. Hence, we opt for fixed views as our RGB modality. 3D hand poses estimated using off-the-shelf MEgATrack [27] are provided by the dataset. H2O [34] features four participants performing 36 actions that involve 8 objects and 11 verbs. Ground truth 3D hand poses are provided. The dataset consists of 5 temporally synchronized RGB views — 4 fixed and 1 egocentric. We use RGB frames from the egocentric view only.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. Similar to [65], our pose input consists of 120 frames by setting $\\mathcal{T}' = 120$ . The number of frames per micro-action ( $N$ ) is 15. For multimodal experiments, we set the window stride $R = N$ and take $K = 8$ non-overlapping micro-actions, which allows us to sample 8 RGB frames for each video following [51]. However, we allow a $50\\%$ overlap between consecutive micro-actions for pose-only variants. A frozen ViT [18] or ResNet [29], followed by a learnable linear layer, is used as the frame encoder $F$ . More details are provided in Suppl. Sec. C. Each model is trained for 50 epochs using SGD with momentum 0.9, a batch size of 32, a learning rate of 0.025, and step LR decay with a factor of 0.1 at epochs {25, 40}. Loss hyperparameters $\\{\\lambda_1, \\lambda_2, \\lambda_3\\}$ are chosen to be {1.0, 1.0, 2.0}.", + "bbox": [ + 212, + 313, + 787, + 478 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Model Variants. We propose several variants of our model by adjusting the width $d$ and the number of layers $T_{n}$ of the transformer to balance efficiency and accuracy. Our default HandFormer, denoted HandFormer-B, has parameters $(d,T_n) = (256,2)$ . We introduce a larger variant, HandFormer-L, with $(d,T_{n}) = (512,4)$ . We also explore different configurations for the number of input joints $J$ per hand. Unless otherwise mentioned, we utilize all 21 joints per hand along with the base model denoted as HandFormer-B/21 while offering a highly efficient option utilizing only six joints per hand (five fingertips and the wrist) termed HandFormer-B/6. Additionally, HandFormer-X/J×N denotes a multimodal variant that utilizes $\\mathcal{N}$ RGB frames, where $\\mathcal{N} \\leq K$ .", + "bbox": [ + 212, + 479, + 789, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 214, + 650, + 542, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed architecture, we compare it against several baselines. For pose-only comparisons, we choose a graph-based network MS-G3D [44] and an attention-based method ISTA-Net [65] — the two best performing skeleton-based methods on Assembly101 [51], as reported by [65]. We also employ video baselines TSM [40] and SlowFast [23], emphasizing efficiency and high temporal resolution, respectively. Additionally, we include H2OTR [13], the state-of-the-art for the H2O dataset. For Assembly101, we replicate the results of the methods above using RGB frames from view 4 of the dataset, while results on H2O are acquired from the respective papers. Furthermore, on both datasets, we train and test RGBPoseConv3D [20], which is current state-of-the-art in multimodal action recognition with skeleton and RGB data.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 643, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8a48c7341c5311fccdaa4982f8e4c0696802cc01f929ce3b21735106d367ab79.jpg", + "table_caption": [ + "Table 1: Quantitative comparison with state-of-the-art methods on Assembly101 [51] and H2O [34]. For H2O, 6D object pose is used by ISTA-Net during training and inference and by H2OTR during training. 'MS-G3D + TSM' denotes a late fusion of the corresponding unimodal architectures. Reported GFLOPs include pose estimation costs in the context of Assembly101 at 15 fps. †For pose-only methods, pose estimation costs are common and excluded for simpler comparison." + ], + "table_footnote": [], + "table_body": "
MethodPoseRGBGFLOPsAssembly101H2O Action
ActionVerbObject
MS-G3D [44]✓✗\\( 21.2^† \\)28.7863.4637.2650.83
ISTA-Net [65]✓✗\\( 35.2^† \\)28.1462.7036.7789.09
SlowFast [23]✗✓65.7---77.69
TSM [40]✗✓33.035.2758.2747.45-
H2OTR [13]✗✓----90.90
RGBPoseConv3D [20]✓✓68.933.6161.9942.9083.47
MS-G3D + TSM✓✓66.239.7465.1251.12-
HandFormer-B/21×8✓✗\\( 4.2^† \\)28.8065.3336.2857.44
✗✓33.032.0755.6144.8984.71
✓✓47.641.0669.2351.1793.39
", + "bbox": [ + 243, + 239, + 754, + 397 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate three variants of our method by controlling the input modalities. As shown in Tab. 1, our unimodal pose-only model excels in verb recognition on Assembly101. In contrast, RGB-based methods benefit from object appearance and usually perform well for actions due to the strong object recognition.", + "bbox": [ + 212, + 421, + 782, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this context, the accuracy of ISTA-Net on H2O is not directly comparable to our method, as they also use the 6D object poses, while we only use hand poses as input. RGBPoseConv3D struggles to achieve satisfactory performance, particularly on Assembly101, suggesting that generalizing to hand poses is non-trivial for skeleton-based methods. Therefore, we combine two best-performing unimodal methods with late fusion (MS-G3D + TSM) to create a stronger baseline. Our model even outperforms this, indicating the effectiveness of our proposed multimodal fusion.", + "bbox": [ + 212, + 482, + 787, + 602 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Skeleton-based Action Recognition for Hands", + "text_level": 1, + "bbox": [ + 214, + 621, + 635, + 636 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The compositional nature of action classes in hand-object interaction videos allows us to break down the action into a verb and an object. Recognizing such actions from 3D hand poses is an ill-posed problem, as the hand skeletons lack explicit information about the interacting objects, which are also part of the action semantics. However, as the pose data completely captures the hand motion information, it can be reliably used for verb recognition. Therefore, we evaluate a pose-only version of our method on the verb recognition task and compare the performance and efficiency metrics with other state-of-the-art skeleton-based methods in Fig. 5. Our method uses significantly fewer GFLOPs due to our spatiotemporal factorization using micro-actions. With $J = 21$ , all our HandFormer variants outperform existing methods. The $J = 6$ variant is exceptionally efficient with comparable accuracy, notably surpassing the efficiency-focused Shift-GCN [11]. HandFormer-H/21 combines the HandFormer-B and HandFormer-L", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c8bbdbfd31fc2fbb0900f212d79be075e43c4f514dd24c16236e70b7b51bfb0d.jpg", + "image_caption": [ + "Fig.5: Comparison of skeleton-based methods for verb recognition on Assembly101 [51]. Our method achieves state-of-the-art performance while utilizing significantly fewer FLOPs." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 475, + 292 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4d598a8731ddaaa4af32e72798ed58dcd48f61e2a09d42f0c5ca07b84907520a.jpg", + "image_caption": [ + "Fig.6: Ablating number of RGB frames in HandFormer on Assembly101 [51]. With more RGB frames, verb recognition shows marginal gain, whereas object recognition shows improvement with diminishing returns." + ], + "image_footnote": [], + "bbox": [ + 535, + 150, + 782, + 292 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "variants and becomes our best-performing model, improving over MS-G3D by $2.04\\%$ while maintaining comparable FLOPs.", + "bbox": [ + 215, + 417, + 784, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 How many RGB frames are required?", + "text_level": 1, + "bbox": [ + 215, + 469, + 571, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In our model, the pose modality maintains a high temporal resolution to capture fine-grained hand movements, resulting in good verb recognition performance. On the contrary, RGB frames are primarily required to introduce semantic context for object recognition and does not necessitate a high temporal resolution like hand movements. The design of our model allows us to sample only a few RGB frames (as low as one) but still perform competitively at a reduced computational cost. Fig. 6 shows the impact of using more RGB frames for Assembly101 [51]. For this experiment, non-overlapping micro-actions are considered. Using only one RGB frame in HandFormer (35.46) outperforms the video model TSM (35.27), as shown in Tab. 1. This performance gain stems primarily from a significant improvement in object accuracy, with only a slight enhancement in verb accuracy. However, including more RGB frames shows a diminishing return, which is unsurprising as additional frames are expected to provide redundant information. These results are obtained with a simplified version of our model by setting $\\mathcal{L} = \\mathcal{L}_{cls}$ and bypassing the multimodal tokenizer.", + "bbox": [ + 215, + 493, + 785, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Can 3D hand pose be an efficient alternative to multi-view?", + "text_level": 1, + "bbox": [ + 215, + 739, + 751, + 755 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "While multiview action recognition benefits from precise hand-movement information in 3D space, processing all views with video models is expensive and highly redundant. In Tab. 2, we demonstrate that combining 3D hand pose with a single RGB view (view 4) achieves comparable performance to multiview action recognition on Assembly101 [51]. Specifically, our action recognition", + "bbox": [ + 215, + 763, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/59bbcbb4382232b1c4adf30201353c2b06c68406d44bb0faacb55ca77cd5ef3b.jpg", + "table_caption": [ + "Table 2: Multi-view Action Recognition on Assembly101 [51]. HandFormer with 'Single View+Pose' has better performance than 'Single View+Egocentric'. Also, verb performance is comparable to using all (8) RGB views." + ], + "table_footnote": [], + "table_body": "
ViewsActionVerbObject
Single View35.2758.2747.45
Single View + Egocentric37.7561.8049.43
Two Views41.9665.2253.26
All (8) Views47.5170.9957.73
Single View + Pose41.0669.2351.17
", + "bbox": [ + 218, + 253, + 472, + 335 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2a7d2cc06d7600bde6f49ab4c6f3e6a13797bdd0e2d1c68b391fae46a6847e40.jpg", + "table_caption": [ + "Table 3: Cross-view performance of HandFormer in Assembly101 [51] shows its generalization capability to unseen view 1, outperforming the video baseline which is trained on view 1 directly. Egocentric views are the source of hand poses here and, therefore, are included in the video models." + ], + "table_footnote": [], + "table_body": "
MethodTrain onTest onActionVerbObject
TSM [40]v4 + egov4 + ego37.7561.8049.43
v4 + egov1 + ego35.2759.5347.72
v1 + egov1 + ego36.2160.7848.52
Our Methodv4 + Posev4 + Pose41.0669.2351.17
v4 + Posev1 + Pose38.4367.8648.32
", + "bbox": [ + 486, + 253, + 779, + 334 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "performance ('Single View + Pose') matches the fusion of the two most informative views — view 1 and view 4. Notably, our verb recognition accuracy is on par with combining all 8 RGB views, though the latter incurs at least $5 \\times$ more FLOPs despite using the efficient TSM model. Additionally, using hand pose in combination with an RGB view ('Single View + Pose') outperforms directly using the egocentric videos ('Single View + Egocentric') from which the hand poses are derived. While fusing multiple RGB views improves accuracy by ensembling multiple complementary predictions, the computational overhead increases significantly. In contrast, our model processes hand pose and single-view RGB frames, enhancing efficiency by leveraging the less redundant and low-dimensional pose data.", + "bbox": [ + 212, + 364, + 785, + 532 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cross-view Generalization. 3D hand pose offers a unique opportunity for cross-view generalization because of its universality across different viewpoints. To evaluate the effectiveness of our method for unseen views, we train our model with frame-wise RGB features from view 4 and test it on view 1. We train TSM [40] as a baseline video model on both RGB views separately. As our method includes 3D hand poses obtained from the egocentric views, we include the egocentric videos in the TSM baselines for a fair comparison. As seen from Tab. 3, our method, trained on view 4, generalizes well on unseen view 1, outperforming the TSM model that was directly trained on view 1.", + "bbox": [ + 212, + 547, + 787, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Egocentric Action Recognition", + "text_level": 1, + "bbox": [ + 215, + 707, + 514, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Recognizing actions in egocentric videos is challenging due to camera motion and the occlusion of interacting objects by the hands. Additionally, in the case of Assembly101 [51], the egocentric cameras are similar to Oculus Quest VR headsets, which provide monochrome low-resolution frames. As a result, action recognition performance is significantly lower than fixed RGB views, as found in [51]. We address this challenging scenario with our proposed multimodal architecture and achieve state-of-the-art performance in egocentric action recognition on Assem", + "bbox": [ + 212, + 733, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c1f4a6e4591540a1a8810d622c78597218bbc0a27d9582a14c1618167499817a.jpg", + "table_caption": [ + "Table 4: Egocentric action recognition in Assembly101 [51]. TSM features from [51] are used as RGB frame features." + ], + "table_footnote": [], + "table_body": "
MethodActionVerbObject
TSM egocentric (fuse 4 views)33.8059.0046.50
Egocentric (e3) + Pose36.0765.5245.82
Egocentric (e4) + Pose35.5665.7945.20
Egocentric (e3+e4) + Pose38.0566.3247.86
", + "bbox": [ + 217, + 199, + 527, + 275 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/70fb96d834dbba458c9ce6eb107f592c9cced3fe768c8949da0df7295961fde0.jpg", + "table_caption": [ + "Table 6: Micro-action length ablation for verb recognition in Assembly101 [51]" + ], + "table_footnote": [], + "table_body": "
#Frames1153060120
Verb \nAccuracy(%)59.1263.7063.6863.5162.29
", + "bbox": [ + 217, + 330, + 514, + 377 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bf5a4a00afb7ec58f73eb6030356b4850c9c1a8ee2b7a04bb7536ceb273af54b.jpg", + "table_caption": [ + "Table 5: Keypoint ablation for verb recognition in Assembly101 [51]" + ], + "table_footnote": [], + "table_body": "
#JointsGlobal ReferenceVerb Accuracy (%)
21×64.17
21✓64.90
11✓64.77
6✓63.70
", + "bbox": [ + 558, + 199, + 771, + 276 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6db324321f4632412325c4208c877f4f2dc6c22d0f0b528814db0fbcc7c731b4.jpg", + "table_caption": [ + "Table 7: Ablating unimodal temporal aggregation for verb recognition in Assembly101 [51]" + ], + "table_footnote": [], + "table_body": "
Temp. Agg.TCNLSTMTransformer
Verb Acc. (%)62.9563.3463.70
", + "bbox": [ + 542, + 345, + 774, + 375 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "bly101. For this experiment, we used frame-wise TSM features provided by [51]. As depicted in Tab. 4, our model, using a single egocentric view ( $e3$ or $e4$ ) outperforms the fusion of four egocentric views as reported in [51]. Moreover, fusing $e3$ and $e4$ significantly enhances our model, resulting in a $4.25\\%$ increase in action accuracy over the baseline.", + "bbox": [ + 212, + 406, + 784, + 481 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.6 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 507, + 401, + 522 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Keypoints. Not all joints of the hands are equally informative for understanding hand actions. For instance, fingertips exhibit greater mobility compared to the inner joints. Moreover, from an egocentric viewpoint, certain joints are more prone to self-occlusion than others. In Tab. 5, we present the impact of incorporating varying numbers of joints on verb recognition within the Assembly101 dataset [51]. For the case of 6 joints, we consider only the wrist joint along with the five fingertips. To expand to 11 joints, we incorporate all the joints along the index and the thumb, which are least affected by self-occlusion. We also show the effect of including the Global Wrist Token, which acts as a reference to the global motion of the hands while encoding micro-actions.", + "bbox": [ + 212, + 536, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Micro-action length. As the resized input comprises a fixed number of pose frames, enlarging the window size for a micro-action decreases the number of micro-actions to aggregate, and vice-versa. In Tab. 6, we vary the micro-action length for verb recognition in the Assembly101 [51] using 6 joints per hand as input, i.e. fingertips and wrist joint. The input pose clip is temporally resized to $\\mathcal{T}' = 120$ before breaking into micro-actions. Lengths 1 and 120 represent two extreme versions with frame-based and trajectory-based encoding, respectively, while the others conform to our micro-action-based formulation.", + "bbox": [ + 212, + 703, + 787, + 824 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 643, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c7164a2bf5f1cee3b1926828848281464edc6ff09237eb591175d6430c5ba561.jpg", + "table_caption": [ + "Table 8: Ablating tokenization and different losses for action recognition" + ], + "table_footnote": [], + "table_body": "
Multimodal TokenizerFeature Ant. LossVerb & Object LossAction Accuracy (%)
Assembly101 [51]H2O [34]
XXX38.9885.95
✓XX40.1988.84
✓✓X40.2489.26
XX✓40.5690.50
✓✓✓41.0693.39
", + "bbox": [ + 245, + 170, + 759, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Temporal Aggregation. After extracting micro-action features, aggregation for action recognition can be done with any sequence model. In Tab. 7, we evaluate the effectiveness of different temporal aggregation methods for verb recognition in the Assembly101 dataset [51]. Here, we use 6 joints per hand as input, i.e., the fingertips and the wrist joint.", + "bbox": [ + 212, + 306, + 785, + 381 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Loss components. To assess the individual contributions of different components, we begin with a basic configuration. We then systematically introduce each element to understand its impact on the overall performance as observed in Tab. 8. Incorporating modality interaction between RGB and pose at the micro-action level through a multimodal tokenizer enhances action accuracy. The introduction of auxiliary losses also has a positive impact, resulting in an overall improvement of $2.08\\%$ for Assembly101 [51] and $7.44\\%$ for H2O [34].", + "bbox": [ + 212, + 383, + 785, + 489 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 523, + 357, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "With the growing interest in AR/VR and wearables, hand pose estimation has rapidly advanced, and holds promise as a compact and domain-independent modality to complement the visual input. To address the underexplored domain of using 3D hand pose as a modality for hand-object interaction recognition, we introduce HandFormer, a novel multimodal transformer that leverages dense sequences of 3D hand poses with sparsely sampled RGB frames to achieve state-of-the-art action recognition performance. Our model also reduces computational requirements, offering immediate significance across various low-resource applications in mobile devices.", + "bbox": [ + 212, + 565, + 785, + 700 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Our method relies on the availability of hand poses, which, if extracted from the visual modality with pose estimation tools [27, 28], can encounter out-of-view scenarios and produce noisy poses. Our experiments reveal that these estimated poses can still achieve good accuracy, yet further research can be conducted to explicitly address this phenomenon. We also assume that a uniform sampling of RGB frames from each micro-action should provide good representations for understanding the semantic context. However, not all frames are equally important in understanding the action. In such a case, adaptive frame sampling methods can be employed, which we leave for future work.", + "bbox": [ + 212, + 703, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported by A\\*STAR under its National Robotics Programme (NRP) (Award M23NBK0053).", + "bbox": [ + 215, + 178, + 784, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 234, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, D., Kim, S., Hong, H., Ko, B.C.: Star-transformer: a spatio-temporal cross attention transformer for human action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3330-3339 (2023)", + "2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)", + "3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)", + "4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)", + "5. Bruce, X., Liu, Y., Chan, K.C.: Multimodal fusion via teacher-student network for indoor action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3199-3207 (2021)", + "6. Bruce, X., Liu, Y., Zhang, X., Zhong, S.h., Chan, K.C.: Mmnet: A model-based multimodal network for human action recognition in rgb-d videos. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3522-3538 (2022)", + "7. Caetano, C., Sena, J., Brémond, F., Dos Santos, J.A., Schwartz, W.R.: Skelemotion: A new representation of skeleton joint sequences based on motion information for 3d action recognition. In: 2019 16th IEEE international conference on advanced video and signal based surveillance (AVSS). pp. 1-8. IEEE (2019)", + "8. Cao, C., Zhang, Y., Zhang, C., Lu, H.: Body joint guided 3-d deep convolutional descriptors for action recognition. IEEE transactions on cybernetics 48(3), 1095-1108 (2017)", + "9. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "0. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "11. Cheng, K., Zhang, Y., He, X., Chen, W., Cheng, J., Lu, H.: Skeleton-based action recognition with shift graph convolutional network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 183-192 (2020)", + "12. Chéron, G., Laptev, I., Schmid, C.: P-cnn: Pose-based cnn features for action recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 3218-3226 (2015)", + "13. Cho, H., Kim, C., Kim, J., Lee, S., Ismayilzada, E., Baek, S.: Transformer-based unified recognition of two hands manipulating objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4769-4778 (2023)" + ], + "bbox": [ + 218, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 643, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European conference on computer vision (ECCV). pp. 720-736 (2018)", + "15. Das, S., Dai, R., Yang, D., Bremond, F.: Vpn++: Rethinking video-pose embeddings for understanding activities of daily living. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9703-9717 (2021)", + "16. Das, S., Sharma, S., Dai, R., Bremond, F., Thonnat, M.: Vpn: Learning video-posed embedding for activities of daily living. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16. pp. 72–90. Springer (2020)", + "17. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "19. Du, Y., Wang, W., Wang, L.: Hierarchical recurrent neural network for skeleton based action recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1110-1118 (2015)", + "20. Duan, H., Zhao, Y., Chen, K., Lin, D., Dai, B.: Revisiting skeleton-based action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2969-2978 (2022)", + "21. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022)", + "22. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)", + "23. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)", + "24. Gan, M., Liu, J., He, Y., Chen, A., Ma, Q.: Keyframe selection via deep reinforcement learning for skeleton-based gesture recognition. IEEE Robotics and Automation Letters (2023)", + "25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)", + "26. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)", + "27. Han, S., Liu, B., Cabezas, R., Twigg, C.D., Zhang, P., Petkau, J., Yu, T.H., Tai, C.J., Akbay, M., Wang, Z., et al.: Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG) 39(4), 87-1 (2020)", + "28. Han, S., Wu, P.c., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., et al.: Umetrack: Unified multi-view end-to-end hand tracking for vr. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "30. Hou, J., Wang, G., Chen, X., Xue, J.H., Zhu, R., Yang, H.: Spatial-temporal attention res-tcn for skeleton-based dynamic hand gesture recognition. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)", + "31. Hou, Y., Li, Z., Wang, P., Li, W.: Skeleton optical spectra-based action recognition using convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology 28(3), 807-811 (2016)", + "32. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021)", + "33. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019)", + "34. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10138-10148 (2021)", + "35. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017)", + "36. Lee, C.J., Zhang, R., Agarwal, D., Yu, T.C., Gunda, V., Lopez, O., Kim, J., Yin, S., Deng, B., Li, K., et al.: Echowrist: Continuous hand pose tracking and hand-object interaction recognition using low-power active acoustic sensing on a wristband. arXiv preprint arXiv:2401.17409 (2024)", + "37. Li, C., Li, S., Gao, Y., Zhang, X., Li, W.: A two-stream neural network for pose-based hand gesture recognition. IEEE Transactions on Cognitive and Developmental Systems 14(4), 1594-1603 (2021)", + "38. Li, J., Xie, X., Pan, Q., Cao, Y., Zhao, Z., Shi, G.: Sgm-net: Skeleton-guided multimodal network for action recognition. Pattern Recognition 104, 107356 (2020)", + "39. Li, M., Chen, S., Chen, X., Zhang, Y., Wang, Y., Tian, Q.: Actional-structural graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3595-3603 (2019)", + "40. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)", + "41. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence 42(10), 2684-2701 (2019)", + "42. Liu, Y., Zhang, S., Gowda, M.: Neuropose: 3d hand pose tracking using emg wearables. In: Proceedings of the Web Conference 2021. pp. 1471-1482 (2021)", + "43. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)", + "44. Liu, Z., Zhang, H., Chen, Z., Wang, Z., Ouyang, W.: Disentangling and unifying graph convolutions for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 143-152 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Ma, J., Damen, D.: Hand-object interaction reasoning. In: 2022 18th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). pp. 1-8. IEEE (2022)", + "46. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12999-13008 (2023)", + "47. Patrick, M., Campbell, D., Asano, Y., Misra, I., Metze, F., Feichtenhofer, C., Vedaldi, A., Henriques, J.F.: Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems 34, 12493-12506 (2021)", + "48. Plizzari, C., Cannici, M., Matteucci, M.: Spatial temporal transformer network for skeleton-based action recognition. In: Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10-15, 2021, Proceedings, Part III. pp. 694-701. Springer (2021)", + "49. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3d pose and tracking for human action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 640-649 (2023)", + "50. Sabater, A., Alonso, I., Montesano, L., Murillo, A.C.: Domain and view-point agnostic hand action recognition. IEEE Robotics and Automation Letters 6(4), 7823-7830 (2021)", + "51. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022)", + "52. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016)", + "53. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9869-9878 (2020)", + "54. Shi, L., Zhang, Y., Cheng, J., Lu, H.: Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12026-12035 (2019)", + "55. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)", + "56. Soo Kim, T., Reiter, A.: Interpretable 3d human action analysis with temporal convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 20-28 (2017)", + "57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)", + "58. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)", + "59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M.S. Shamil et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Vemulapalli, R., Arrate, F., Chellappa, R.: Human action recognition by representing 3d skeletons as points in a lie group. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 588-595 (2014)", + "61. Vondrick, C., Pirsiavash, H., Torralba, A.: Anticipating visual representations from unlabeled video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 98-106 (2016)", + "62. Wang, J., Liu, Z., Wu, Y., Yuan, J.: Mining actionlet ensemble for action recognition with depth cameras. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 1290-1297. IEEE (2012)", + "63. Weiyao, X., Muqing, W., Min, Z., Ting, X.: Fusion of skeleton and rgb features for rgb-d human action recognition. IEEE Sensors Journal 21(17), 19157-19164 (2021)", + "64. Wen, Y., Pan, H., Yang, L., Pan, J., Komura, T., Wang, W.: Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21243-21253 (2023)", + "65. Wen, Y., Tang, Z., Pang, Y., Ding, B., Liu, M.: Interactive spatiotemporal token attention network for skeleton-based general interactive action recognition. arXiv preprint arXiv:2307.07469 (2023)", + "66. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13587-13597 (2022)", + "67. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European conference on computer vision (ECCV). pp. 305–321 (2018)", + "68. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)", + "69. Zhang, S., Liu, X., Xiao, J.: On geometric features for skeleton-based action recognition using multilayer LSTM networks. In: 2017 IEEE winter conference on applications of computer vision (WACV). pp. 148-157. IEEE (2017)", + "70. Zhang, X., Xu, C., Tian, X., Tao, D.: Graph edge convolutional neural networks for skeleton-based action recognition. IEEE transactions on neural networks and learning systems 31(8), 3047-3060 (2019)", + "71. Zhang, Y., Wu, B., Li, W., Duan, L., Gan, C.: Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 3229-3237 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 674 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "HandFormer", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_model.json b/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e29cb467095e2cc8e6b83d5669a93307ea7afd5c --- /dev/null +++ b/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_model.json @@ -0,0 +1,2636 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.294, + 0.141, + 0.71, + 0.187 + ], + "angle": 0, + "content": "On the Utility of 3D Hand Poses for Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.212, + 0.732, + 0.245 + ], + "angle": 0, + "content": "Md Salman Shamil1, Dibyadip Chatterjee1, Fadime Sener2, Shugao Ma2, and Angela Yao1" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.255, + 0.658, + 0.327 + ], + "angle": 0, + "content": "1 National University of Singapore 2 Meta Reality Labs {salman, dibyadip, ayao}@comp.nus.edu.sg {famesener, shugao}@meta.com https://s-shamil.github.io/HandFormer/" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.365, + 0.74, + 0.588 + ], + "angle": 0, + "content": "Abstract. 3D hand pose is an underexplored modality for action recognition. Poses are compact yet informative and can greatly benefit applications with limited compute budgets. However, poses alone offer an incomplete understanding of actions, as they cannot fully capture objects and environments with which humans interact. We propose HandFormer, a novel multimodal transformer, to efficiently model hand-object interactions. HandFormer combines 3D hand poses at a high temporal resolution for fine-grained motion modeling with sparsely sampled RGB frames for encoding scene semantics. Observing the unique characteristics of hand poses, we temporally factorize hand modeling and represent each joint by its short-term trajectories. This factorized pose representation combined with sparse RGB samples is remarkably efficient and highly accurate. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods at \\(5 \\times\\) fewer FLOPs. With RGB, we achieve new state-of-the-art performance on Assembly101 and H2O with significant improvements in egocentric action recognition." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.601, + 0.741, + 0.629 + ], + "angle": 0, + "content": "Keywords: Skeleton-based action recognition \\(\\cdot\\) 3D hand poses \\(\\cdot\\) Multi-modal transformer" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.657, + 0.377, + 0.673 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The popularity of AR/VR headsets has driven interest in recognizing hand-object interactions, particularly through egocentric [14, 26] and multiview cameras [34, 51]. Such interactions are inherently fine-grained; recognizing them requires distinguishing subtle motions and object state changes. State-of-the-art methods for hand action recognition [23, 45, 47, 66] primarily rely on multi-or single-view RGB streams, which are computationally heavy and unsuitable for resource-constrained scenarios like AR/VR. Motivated by advancements in lightweight hand pose estimation methods leveraging monochrome cameras [27, 28, 46], and the evolution of low-dimensional sensors [36, 42] such as accelerometers, MMG, EMG, demonstrating real-time hand pose estimation, we advocate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.782, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 1: We densely sample 3D hand poses to understand fine-grained hand motions and sparsely sample RGB frames to capture the scene semantics. Our 3D hand poses are acquired from low-resolution monochrome cameras but can also come from wearable sensors, facilitating an efficient understanding of hand-object interactions. Video frames and hand poses in the figure are from Assembly101 [51]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.414 + ], + "angle": 0, + "content": "the use of 3D hand poses as an input modality for recognizing hand-object interactions. Hand poses are a compact yet informative representation that captures the motions and nuances of hand movements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.552 + ], + "angle": 0, + "content": "Existing works on 3D pose-based action recognition have focused primarily on full-body skeletons [20,44,65,68]. Hand poses differ fundamentally from full-body skeletons. In full-body recognition datasets [41,52], the actions are predominantly static from a global perspective. The relative changes in joint or limb positions signify the action category, e.g., 'sitting down'. Conversely, hand joints typically move together for many actions and lack a static joint as a global reference [51], e.g., 'put down toy'. Full-body skeleton methods also benefit from modeling long-range spatiotemporal dependencies between joints [44], while this is less important for the hands as shown in Fig. 2 and in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.658 + ], + "angle": 0, + "content": "However, the 3D pose alone cannot encode the hands' actions. Unlike the full-body case, where actions are self-contained by the sequence of poses, the hands are often manipulating objects [34, 46, 51]. Hand pose is excellent for identifying motions (verbs) but struggles with associated objects [51]. Therefore, supplementing pose data with visual context from images or videos is crucial for full semantic understanding. However, as noted earlier, using dense RGB frames contradicts our motivation for using hand poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "This paper introduces HandFormer, a novel and lightweight multimodal transformer that leverages dense 3D hand poses complemented with sparsely sampled RGB frames. To this end, we conceptualize an action as a sequence of short segments, termed micro-actions, which are analogous to words that form sentence-level complex actions. Each micro-action comprises a dense sequence of pose frames and a single RGB frame. As every hand joint moves in close spatial proximity, we encode the pose sequence from a Lagrangian view [49] and track each joint as an individual entity. HandFormer effectively models each micro-action whose features are then temporally aggregated for action classification. This formulation benefits from our observations on the spatiotemporal dynamics of hand movement and delivers strong performance. Additionally, limiting long-range spatiotemporal dependencies among joints and only using sparse RGB frame" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "features improves computational efficiency, making it well-suited for always-on, low-compute AR/VR applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Our contributions are: (i) We analyze the differences between hand pose and full-body skeleton actions and design a novel pose sequence encoding that reflects hand-specific properties. (ii) We propose HandFormer that takes a sequence of dense 3D hand pose and sparse RGB frames as input in the form of micro-actions. (iii) HandFormer achieves state-of-the-art action recognition on H2O [34] and Assembly101 [51]. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods in Assembly101, incurring at least \\(5 \\times\\) fewer FLOPs. (iv) We experimentally demonstrate how hand poses are crucial for multiview and egocentric action recognition." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.338, + 0.388, + 0.356 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Video Action recognition. Video-based action recognition systems are well-developed with sophisticated 3D-CNN [9,23,57,67] or Video Transformer [2,3,43,47] architectures. However, they all bear significant computational expense for both feature extraction and motion modeling, either explicitly in the form of optical flow [9,33,55] or implicitly through the architecture [2,47]. Such designs are well-suited for high-facility applications but are not suitable for integration into lightweight systems. To this end, efficient video understanding is an active topic of research, with a focus on reducing expensive 3D operations [22,40,58,67], quadratic complexity of attention [32,43,47] and dropping tokens [4,21]. Additionally, densely-sampled RGB frames incur a high cost, yet limiting the temporal resolution will hinder fine-grained action understanding. We propose complementing 3D hand poses with sparsely-sampled RGB frames, developing a lightweight video understanding system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Skeleton-based Action Recognition. Skeleton-based action recognition has been tackled with hand-crafted features [60, 62], sequence models like RNNs and LSTMs [19, 69], and CNN-based methods that either employ temporal convolution on the pose sequence [56] or transform the skeleton data into pseudoimages to be processed with 2D or 3D convolutional networks [7, 20, 31]. Recent progress has been driven by GNN-based methods that exploit the graph structure of skeletal data to construct spatio-temporal graphs and perform graph convolutions [68, 70]. These methods often model functional links between joints that go beyond skeletal connectivity [39, 54] or expand the spatiotemporal receptive field [44]. Self-attention and transformer-based methods have also been proposed [48, 65, 71]. However, most existing methods are tailored for datasets involving full-body poses. Few works [24, 30, 37, 50] study hand poses but are restricted to simple gestures, recognizable without complex temporal modeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Fusing RGB with Skeleton. Skeletal data has been used for cropping images of body parts [12], for weighting RGB patches around regions of interest [5, 6], or for pooling CNN features [1, 8]. Projecting into common embedding space is done in [16], enabling pose distillation [15]. Multi-stream architectures are also designed with separate paths for RGB and skeleton having lateral connec" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.127 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.183, + 0.266, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.238, + 0.266, + 0.246 + ], + "angle": 0, + "content": "Stand up" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.145, + 0.326, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.193, + 0.326, + 0.203 + ], + "angle": 0, + "content": "Put on bag" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.208, + 0.325, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.259, + 0.322, + 0.267 + ], + "angle": 0, + "content": "Sit down" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.144, + 0.481, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.257, + 0.46, + 0.269 + ], + "angle": 0, + "content": "\\(\\leftarrow \\rightarrow \\rightarrow < J_{\\mathrm{H,0}}J_{\\mathrm{ET}}>\\)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.155, + 0.669, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.256, + 0.634, + 0.269 + ], + "angle": 0, + "content": "←→" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.15, + 0.732, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.189, + 0.723, + 0.199 + ], + "angle": 0, + "content": "Pick up" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.221, + 0.732, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.258, + 0.726, + 0.266 + ], + "angle": 0, + "content": "Unscrew" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.186, + 0.785, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.228, + 0.775, + 0.236 + ], + "angle": 0, + "content": "Rotate" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 2: Comparing skeletal changes in full-body actions from NTU RGB+D 120 [41] (left) and hand actions from Assembly101 [51] (right). Two pose frames at interval \\( T \\) are shown. \\( J_{j,t} \\) indicates the 3D coordinate of joint \\( j \\) at timestep \\( t \\), and \\( \\) is the correlation between two such joints. Modeling the correlation between spatiotemporally distant joints can be informative for full-body poses but does not provide a useful action cue for hand poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.515 + ], + "angle": 0, + "content": "tions between them [38, 63]. RGBPoseConv3D [20] is a two-stream model that employs 3D CNNs for both RGB and pose. Similar to skeleton-based action recognition, these multimodal approaches primarily focus on full-body poses. Some approaches simultaneously perform both hand pose estimation and action recognition, using pose data to supervise the training process [13, 64]. However, it is important to note that pose estimation deals with lower-level semantics than action recognition. Actions can often be inferred even when the estimated poses are not accurate [46]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.537, + 0.634, + 0.554 + ], + "angle": 0, + "content": "3 Modeling Full-body vs. Hand Skeletons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Existing skeleton-based action recognition datasets primarily consist of full-body poses where actions feature significant changes over time in limb positions relative to other body parts, which highly correlate to the action category. These changes can be captured through long-range spatiotemporal modeling using graph convolutions with large receptive fields [44] or self-attention [65]. Conversely, hand actions show diverse movement patterns, as hands can move in arbitrary directions, often without significant changes in articulation. We analyze pose sequences from NTU RGB+D 120 [41] and Assembly101 [51], calculating the distance covered by each joint and identifying the least and the most active joints. We observe that, for full-body poses, these two representative joints show a significant difference in the distances covered, while for hand poses, the difference is subtle. The Pearson correlation coefficient is 0.93 for hand poses (indicating highly coupled joints) and 0.33 for full-body poses. Please refer to Suppl. Sec. A for details regarding these computations and comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "An example showing the difference between body and hand poses is provided in Fig. 2, where we depict two frames separated by an interval of \\( T \\) frames. For the full body skeleton, the head joint \\( J_{H} \\) remains static, while the right hand joint \\( J_{F} \\) moves upward. Conversely, for the hand, all the joints move," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.151, + 0.744, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.369, + 0.788, + 0.453 + ], + "angle": 0, + "content": "Fig. 3: Overall architecture of HandFormer. An action segment is divided into \\( K \\) micro-actions \\( \\{M_1, M_2 \\ldots M_K\\} \\). Each micro-action comprises a dense sequence of pose frames and a single RGB frame. The frame encoder \\( F \\) and the trajectory encoder \\( T \\) encode the RGB and the dense poses, respectively, after which they are passed to a Multimodal Tokenizer. The modality-mixed tokens are then fed to a Temporal Transformer. Dotted paths are optional and only required when RGB is used." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.619 + ], + "angle": 0, + "content": "including the wrist joint \\( J_{W} \\) and index fingertip \\( J_{I} \\). The spatiotemporal correlation between the head joint in the first frame and the hand joint in the last frame (\\(\\)) is a crucial action cue that captures the relative structural change. However, such a correlation between the wrist and the fingertip for hand skeleton (\\(\\)) does not provide a stronger cue than the wrist movement itself. To avoid such redundant spatiotemporally distant correlations, we propose to divide the hand pose sequence into micro-action blocks of fixed temporal length. This formulation allows for encoding short-term movements while enabling parameter sharing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Moreover, full-skeletal motion is a dominant feature in hand pose sequences, differentiating them from full-body poses. While the location and orientation of the human body can be trivial for most actions, this is not the case for hands. The hands do not conform to any particular 6D pose, and notably, they frequently and unpredictably alter their 6D poses over time. A comprehensive analysis of this phenomenon is provided in Suppl. Sec. A. To this end, we explicitly consider the global 6D poses of the hands during micro-action-based pose encoding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.376, + 0.765 + ], + "angle": 0, + "content": "4 HandFormer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Fig. 3 illustrates our proposed HandFormer, which consists of a sequence of micro-action blocks, a novel trajectory encoder, and a temporal aggregation module. The design of our HandFormer allows us to easily incorporate semantic context by sampling a single RGB frame from certain micro-actions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.376, + 0.16 + ], + "angle": 0, + "content": "4.1 Micro-actions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.272 + ], + "angle": 0, + "content": "Given an action segment containing \\(\\mathcal{T}\\) frames sampled at a certain fps, the input to our model comprises — \\(i\\) ) a dense sequence of 3D hand poses, represented as \\(\\mathbf{S} = \\{P_1,P_2,\\dots ,P_{\\mathcal{T}}\\}\\), where \\(P_{t} = \\{P_{t}^{left},P_{t}^{right}\\} \\in \\mathbb{R}^{2\\times J\\times 3}\\) signifying the 3D coordinates of \\(J\\) keypoints in the left and right hands, respectively, and \\(ii)\\) a sparse set of RGB frames sampled at intervals \\(\\varDelta f\\), denoted as \\(\\mathbf{V} = \\{I_1,I_{1 + \\varDelta f},\\ldots ,I_{\\mathcal{T}_r}\\}\\), where \\(I_{t}\\in \\mathbb{R}^{H\\times W\\times 3}\\) are frame-wise RGB features and \\(\\mathcal{T}_r = 1 + \\left|\\frac{\\mathcal{T} - 1}{\\varDelta f}\\right|\\times \\varDelta f\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.422 + ], + "angle": 0, + "content": "We factorize the raw input into a sequence of \\( K \\) micro-action blocks of length \\( N \\) frames, obtained by shifting the window across the action segment with a stride of \\( R \\) frames. Each block consists of two components — the initial appearance derived from the first RGB frame within the block and the hand motion characterized by the dense sequence of \\( N \\) pose frames. To obtain a fixed length input containing \\( \\mathcal{T}' = (K - 1)\\times R + N \\) pose frames, we perform linear interpolation for each joint along the temporal axis, transforming pose sequence \\( \\mathbf{S} \\) having \\( \\mathcal{T} \\) frames to \\( \\mathbf{S}' = \\{P_1',P_2',\\dots ,P_{\\mathcal{T}'}'\\} \\) having \\( \\mathcal{T}' \\) frames, where \\( P_t' \\in \\mathbb{R}^{2\\times J\\times 3} \\). Thus, we represent the input as a sequence of micro-actions \\( \\mathbf{M} = \\{M_1,M_2,\\dots ,M_K\\} \\), derived from \\( \\mathbf{V} \\) and \\( \\mathbf{S}' \\) through the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.432, + 0.786, + 0.457 + ], + "angle": 0, + "content": "\\[\nM _ {k} = \\left[ M _ {k} ^ {\\mathrm {R G B}}, M _ {k} ^ {\\text {P o s e}} \\right] = \\left[ I _ {h (k)}, \\left\\{P _ {g (k) + i} ^ {\\prime} \\right\\} _ {i = 0} ^ {N - 1} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.786, + 0.497 + ], + "angle": 0, + "content": "where \\( g(k) = (k - 1)\\times R + 1 \\) denotes the first pose frame in \\( k \\)-th Micro-action, and \\( h(k) \\) determines the index of the nearest available RGB frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.588 + ], + "angle": 0, + "content": "The dense pose sequence in a micro-action captures fine-grained hand motion crucial for recognizing verbs, whereas a single RGB frame provides semantic context for recognizing objects. To extract features from micro-actions, we use a frame encoder \\( F \\) and a trajectory encoder \\( T \\), which operate on RGB frames and pose sequences, respectively. Consequently, the RGB and pose features for the \\( k^{th} \\) micro-action is given by-" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.598, + 0.786, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\left[ f _ {k} ^ {\\mathrm {R G B}}, f _ {k} ^ {\\mathrm {P o s e}} \\right] = \\left[ F \\left(M _ {k} ^ {\\mathrm {R G B}}\\right), T \\left(M _ {k} ^ {\\mathrm {P o s e}}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.787, + 0.658 + ], + "angle": 0, + "content": "where \\( f_{k}^{\\mathrm{RGB}}, f_{k}^{\\mathrm{pos}} \\in \\mathbb{R}^{d} \\), where \\( d \\) denotes the common dimensionality of both RGB and pose embedding space." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.424, + 0.696 + ], + "angle": 0, + "content": "4.2 Trajectory Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We devise a trajectory-based pose encoder to encode dense hand pose sequences within micro-actions, illustrated in Fig. 4. Each joint is represented by its trajectory of dimension \\(3 \\times N\\), encapsulating the sequence of 3D coordinates across the \\(N\\) pose frames of a micro-action. This yields \\(2 \\times J\\) feature vectors for the \\(J\\) joints of two hands. Each joint's trajectory is passed through a TCN [35], whose parameters are shared for all joints. This produces \\(2 \\times J\\) Local Trajectory Tokens. Additionally, the full-skeletal motion of the hand during the entire action is used as a reference through an additional token named Global Wrist Token. This token is generated by a separate TCN operating on the sequence of 6D" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.151, + 0.788, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.785, + 0.39 + ], + "angle": 0, + "content": "Fig. 4: Our Trajectory Encoder \\( T \\), which operates on micro-actions, derives tokens with trajectory-based features and performs self-attention to encode the pose sequence into a feature vector. The Single-Joint TCN is a Temporal Convolutional Network [35] that processes the trajectories of all the joints individually with shared parameters across all joints. Wrist-TCN takes an action-wide sequence of wrist location and hand orientation (6D pose) to produce a global reference token." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.482 + ], + "angle": 0, + "content": "poses of hands, indicating wrist location and hand orientation. Subsequently, a self-attention layer is applied to these trajectory tokens, preserving the temporal dimension for subsequent stages. This iterative process culminates in spatiotemporal average pooling, summarizing the hand motion of the micro-action." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.506, + 0.445, + 0.52 + ], + "angle": 0, + "content": "4.3 Multimodal Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.785, + 0.668 + ], + "angle": 0, + "content": "This section discusses incorporating sparsely sampled RGB frames into Hand-Former to better capture scene semantics. A single frame is sampled from each micro-action, generating an extended crop \\((1.25\\times)\\) around the hands. While the full image provides an overall scene context, the crop focuses specifically on hand-object interaction regions [10]. Features for both are separately generated using a pre-existing image encoder and then aggregated to enrich the hand-object interaction feature with scene context. The hand-object ROI crop can be obtained using an off-the-shelf HOI detector [53] or by using the corresponding hand pose projections. We opt for the latter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.785, + 0.745 + ], + "angle": 0, + "content": "Our multimodal tokenizer in Fig. 3 generates RGB and Pose tokens enhanced via multimodal interactions. This involves concatenating each frame feature and trajectory encoding, and projecting them into a shared PoseRGB feature space using an MLP. This PoseRGB feature is then split into two parts, which are added to the original frame feature and trajectory encoding, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.449, + 0.783 + ], + "angle": 0, + "content": "4.4 Temporal Transformer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The multimodal tokenizer provides RGB and pose tokens for each micro-action. Since an action segment consists of a sequence of micro-actions, these micro-action tokens are aggregated over time via a temporal transformer. A video \\(\\mathbf{V}\\)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.3 + ], + "angle": 0, + "content": "divided into \\(K\\) micro-actions can be represented by two sets of tokens \\(\\{\\hat{f}_k^{\\mathrm{RGB}}\\}_{k=1}^K\\) and \\(\\{\\hat{f}_k^{\\mathrm{Pose}}\\}_{k=1}^K\\), respectively, produced by the multimodal tokenizer. These \\(2 \\times K\\) tokens of dimension \\(d\\) form the input sequence of the temporal transformer. Positional encoding and modality embedding are added to each input token to indicate the temporal location and source modality. We use the fixed sine/cosine positional encoding [59], assigning the same position to two tokens from the same micro-action. Modality embeddings are learned and shared across the tokens from the same modality (RGB or Pose). Following standard practice [17,18], we preprocess an additional learnable class token \\([\\mathrm{CLS}] \\in \\mathbb{R}^d\\), the output of which is then used for classifying actions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.321, + 0.43, + 0.337 + ], + "angle": 0, + "content": "4.5 Learning Objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.789, + 0.481 + ], + "angle": 0, + "content": "HandFormer is trained end-to-end for action recognition, supervised by a cross-entropy loss: \\(\\mathcal{L}_{cls} = -a_i\\log \\hat{a}_i\\), where \\(a_{i}\\) represents the ground truth action label for the \\(i^{th}\\) sample, and \\(\\hat{a}_i\\) denotes the predicted action category. The input pose and RGB modalities provide complementary information regarding a scene, capturing motion and interacting objects, respectively. To effectively utilize this, we employ explicit verb and object supervision (\\(\\mathcal{L}_{verb}\\), \\(\\mathcal{L}_{obj}\\)) via two additional learnable class tokens. Given that pose strongly correlates with verbs, whereas objects are identifiable from RGB frames alone, we let the verb class token attend exclusively to pose encodings and the object class token to frame features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.79, + 0.618 + ], + "angle": 0, + "content": "Feature Anticipation Loss. Hand pose sequence captures the primary sources of state changes during hand-object interactions. We posit that the visual state from an initial RGB frame, in combination with the subsequent hand pose sequence, is indicative of the visual state that results from the completion of the sequence. Therefore, given an RGB feature and the corresponding pose features from a micro-action, we force our model to anticipate the RGB feature for the next micro-action by minimizing an \\( L_{1} \\) feature loss. This loss, inspired by existing efforts [25,61], quantifies the difference between the anticipated image feature and the true feature extracted from a frozen image encoder. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.641, + 0.787, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a n t} = \\sum_ {k = 1} ^ {K - 1} \\left\\| \\Phi_ {\\text {a n t}} \\left(f _ {k} ^ {\\text {P o s e R G B}}\\right) - f _ {k + 1} ^ {\\text {R G B}} \\right\\| _ {1} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.69, + 0.711, + 0.706 + ], + "angle": 0, + "content": "where \\(\\varPhi_{\\mathrm{ant}}\\) denotes a linear projection layer. Hence, the total loss is" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.722, + 0.786, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\lambda_ {1} \\mathcal {L} _ {v e r b} + \\lambda_ {2} \\mathcal {L} _ {o b j} + \\lambda_ {3} \\mathcal {L} _ {a n t} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.742, + 0.712, + 0.758 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3\\) are hyperparameters used to balance the four losses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.779, + 0.377, + 0.797 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two publicly available hand-object interaction datasets. Assembly101 [51] is a large-scale multiview dataset that fea-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "tures videos of procedural activities for assembling and disassembling 101 take-aside toy vehicles. This dataset has 1380 fine-grained actions based on 24 verbs and 90 objects. It consists of 12 temporally synchronized views — eight static and four egocentric. The egocentric views exhibit low-resolution monochrome images, making it challenging even for human eyes to discern objects in view. Hence, we opt for fixed views as our RGB modality. 3D hand poses estimated using off-the-shelf MEgATrack [27] are provided by the dataset. H2O [34] features four participants performing 36 actions that involve 8 objects and 11 verbs. Ground truth 3D hand poses are provided. The dataset consists of 5 temporally synchronized RGB views — 4 fixed and 1 egocentric. We use RGB frames from the egocentric view only." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.789, + 0.479 + ], + "angle": 0, + "content": "Implementation Details. Similar to [65], our pose input consists of 120 frames by setting \\(\\mathcal{T}' = 120\\). The number of frames per micro-action (\\(N\\)) is 15. For multimodal experiments, we set the window stride \\(R = N\\) and take \\(K = 8\\) non-overlapping micro-actions, which allows us to sample 8 RGB frames for each video following [51]. However, we allow a \\(50\\%\\) overlap between consecutive micro-actions for pose-only variants. A frozen ViT [18] or ResNet [29], followed by a learnable linear layer, is used as the frame encoder \\(F\\). More details are provided in Suppl. Sec. C. Each model is trained for 50 epochs using SGD with momentum 0.9, a batch size of 32, a learning rate of 0.025, and step LR decay with a factor of 0.1 at epochs {25, 40}. Loss hyperparameters \\(\\{\\lambda_1, \\lambda_2, \\lambda_3\\}\\) are chosen to be {1.0, 1.0, 2.0}." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.79, + 0.631 + ], + "angle": 0, + "content": "Model Variants. We propose several variants of our model by adjusting the width \\(d\\) and the number of layers \\(T_{n}\\) of the transformer to balance efficiency and accuracy. Our default HandFormer, denoted HandFormer-B, has parameters \\((d,T_n) = (256,2)\\). We introduce a larger variant, HandFormer-L, with \\((d,T_{n}) = (512,4)\\). We also explore different configurations for the number of input joints \\(J\\) per hand. Unless otherwise mentioned, we utilize all 21 joints per hand along with the base model denoted as HandFormer-B/21 while offering a highly efficient option utilizing only six joints per hand (five fingertips and the wrist) termed HandFormer-B/6. Additionally, HandFormer-X/J×N denotes a multimodal variant that utilizes \\(\\mathcal{N}\\) RGB frames, where \\(\\mathcal{N} \\leq K\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.543, + 0.667 + ], + "angle": 0, + "content": "5.1 Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.84 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed architecture, we compare it against several baselines. For pose-only comparisons, we choose a graph-based network MS-G3D [44] and an attention-based method ISTA-Net [65] — the two best performing skeleton-based methods on Assembly101 [51], as reported by [65]. We also employ video baselines TSM [40] and SlowFast [23], emphasizing efficiency and high temporal resolution, respectively. Additionally, we include H2OTR [13], the state-of-the-art for the H2O dataset. For Assembly101, we replicate the results of the methods above using RGB frames from view 4 of the dataset, while results on H2O are acquired from the respective papers. Furthermore, on both datasets, we train and test RGBPoseConv3D [20], which is current state-of-the-art in multimodal action recognition with skeleton and RGB data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with state-of-the-art methods on Assembly101 [51] and H2O [34]. For H2O, 6D object pose is used by ISTA-Net during training and inference and by H2OTR during training. 'MS-G3D + TSM' denotes a late fusion of the corresponding unimodal architectures. Reported GFLOPs include pose estimation costs in the context of Assembly101 at 15 fps. †For pose-only methods, pose estimation costs are common and excluded for simpler comparison." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.24, + 0.756, + 0.398 + ], + "angle": 0, + "content": "
MethodPoseRGBGFLOPsAssembly101H2O Action
ActionVerbObject
MS-G3D [44]✓✗\\( 21.2^† \\)28.7863.4637.2650.83
ISTA-Net [65]✓✗\\( 35.2^† \\)28.1462.7036.7789.09
SlowFast [23]✗✓65.7---77.69
TSM [40]✗✓33.035.2758.2747.45-
H2OTR [13]✗✓----90.90
RGBPoseConv3D [20]✓✓68.933.6161.9942.9083.47
MS-G3D + TSM✓✓66.239.7465.1251.12-
HandFormer-B/21×8✓✗\\( 4.2^† \\)28.8065.3336.2857.44
✗✓33.032.0755.6144.8984.71
✓✓47.641.0669.2351.1793.39
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.784, + 0.481 + ], + "angle": 0, + "content": "We evaluate three variants of our method by controlling the input modalities. As shown in Tab. 1, our unimodal pose-only model excels in verb recognition on Assembly101. In contrast, RGB-based methods benefit from object appearance and usually perform well for actions due to the strong object recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.603 + ], + "angle": 0, + "content": "In this context, the accuracy of ISTA-Net on H2O is not directly comparable to our method, as they also use the 6D object poses, while we only use hand poses as input. RGBPoseConv3D struggles to achieve satisfactory performance, particularly on Assembly101, suggesting that generalizing to hand poses is non-trivial for skeleton-based methods. Therefore, we combine two best-performing unimodal methods with late fusion (MS-G3D + TSM) to create a stronger baseline. Our model even outperforms this, indicating the effectiveness of our proposed multimodal fusion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.622, + 0.637, + 0.637 + ], + "angle": 0, + "content": "5.2 Skeleton-based Action Recognition for Hands" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The compositional nature of action classes in hand-object interaction videos allows us to break down the action into a verb and an object. Recognizing such actions from 3D hand poses is an ill-posed problem, as the hand skeletons lack explicit information about the interacting objects, which are also part of the action semantics. However, as the pose data completely captures the hand motion information, it can be reliably used for verb recognition. Therefore, we evaluate a pose-only version of our method on the verb recognition task and compare the performance and efficiency metrics with other state-of-the-art skeleton-based methods in Fig. 5. Our method uses significantly fewer GFLOPs due to our spatiotemporal factorization using micro-actions. With \\( J = 21 \\), all our HandFormer variants outperform existing methods. The \\( J = 6 \\) variant is exceptionally efficient with comparable accuracy, notably surpassing the efficiency-focused Shift-GCN [11]. HandFormer-H/21 combines the HandFormer-B and HandFormer-L" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.476, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.295, + 0.481, + 0.378 + ], + "angle": 0, + "content": "Fig.5: Comparison of skeleton-based methods for verb recognition on Assembly101 [51]. Our method achieves state-of-the-art performance while utilizing significantly fewer FLOPs." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.151, + 0.784, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.295, + 0.787, + 0.379 + ], + "angle": 0, + "content": "Fig.6: Ablating number of RGB frames in HandFormer on Assembly101 [51]. With more RGB frames, verb recognition shows marginal gain, whereas object recognition shows improvement with diminishing returns." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.419, + 0.785, + 0.451 + ], + "angle": 0, + "content": "variants and becomes our best-performing model, improving over MS-G3D by \\(2.04\\%\\) while maintaining comparable FLOPs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.573, + 0.485 + ], + "angle": 0, + "content": "5.3 How many RGB frames are required?" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.494, + 0.787, + 0.721 + ], + "angle": 0, + "content": "In our model, the pose modality maintains a high temporal resolution to capture fine-grained hand movements, resulting in good verb recognition performance. On the contrary, RGB frames are primarily required to introduce semantic context for object recognition and does not necessitate a high temporal resolution like hand movements. The design of our model allows us to sample only a few RGB frames (as low as one) but still perform competitively at a reduced computational cost. Fig. 6 shows the impact of using more RGB frames for Assembly101 [51]. For this experiment, non-overlapping micro-actions are considered. Using only one RGB frame in HandFormer (35.46) outperforms the video model TSM (35.27), as shown in Tab. 1. This performance gain stems primarily from a significant improvement in object accuracy, with only a slight enhancement in verb accuracy. However, including more RGB frames shows a diminishing return, which is unsurprising as additional frames are expected to provide redundant information. These results are obtained with a simplified version of our model by setting \\(\\mathcal{L} = \\mathcal{L}_{cls}\\) and bypassing the multimodal tokenizer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.753, + 0.756 + ], + "angle": 0, + "content": "5.4 Can 3D hand pose be an efficient alternative to multi-view?" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "While multiview action recognition benefits from precise hand-movement information in 3D space, processing all views with video models is expensive and highly redundant. In Tab. 2, we demonstrate that combining 3D hand pose with a single RGB view (view 4) achieves comparable performance to multiview action recognition on Assembly101 [51]. Specifically, our action recognition" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.473, + 0.241 + ], + "angle": 0, + "content": "Table 2: Multi-view Action Recognition on Assembly101 [51]. HandFormer with 'Single View+Pose' has better performance than 'Single View+Egocentric'. Also, verb performance is comparable to using all (8) RGB views." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.255, + 0.473, + 0.337 + ], + "angle": 0, + "content": "
ViewsActionVerbObject
Single View35.2758.2747.45
Single View + Egocentric37.7561.8049.43
Two Views41.9665.2253.26
All (8) Views47.5170.9957.73
Single View + Pose41.0669.2351.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.483, + 0.145, + 0.783, + 0.242 + ], + "angle": 0, + "content": "Table 3: Cross-view performance of HandFormer in Assembly101 [51] shows its generalization capability to unseen view 1, outperforming the video baseline which is trained on view 1 directly. Egocentric views are the source of hand poses here and, therefore, are included in the video models." + }, + { + "type": "table", + "bbox": [ + 0.488, + 0.255, + 0.781, + 0.335 + ], + "angle": 0, + "content": "
MethodTrain onTest onActionVerbObject
TSM [40]v4 + egov4 + ego37.7561.8049.43
v4 + egov1 + ego35.2759.5347.72
v1 + egov1 + ego36.2160.7848.52
Our Methodv4 + Posev4 + Pose41.0669.2351.17
v4 + Posev1 + Pose38.4367.8648.32
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.787, + 0.533 + ], + "angle": 0, + "content": "performance ('Single View + Pose') matches the fusion of the two most informative views — view 1 and view 4. Notably, our verb recognition accuracy is on par with combining all 8 RGB views, though the latter incurs at least \\(5 \\times\\) more FLOPs despite using the efficient TSM model. Additionally, using hand pose in combination with an RGB view ('Single View + Pose') outperforms directly using the egocentric videos ('Single View + Egocentric') from which the hand poses are derived. While fusing multiple RGB views improves accuracy by ensembling multiple complementary predictions, the computational overhead increases significantly. In contrast, our model processes hand pose and single-view RGB frames, enhancing efficiency by leveraging the less redundant and low-dimensional pose data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Cross-view Generalization. 3D hand pose offers a unique opportunity for cross-view generalization because of its universality across different viewpoints. To evaluate the effectiveness of our method for unseen views, we train our model with frame-wise RGB features from view 4 and test it on view 1. We train TSM [40] as a baseline video model on both RGB views separately. As our method includes 3D hand poses obtained from the egocentric views, we include the egocentric videos in the TSM baselines for a fair comparison. As seen from Tab. 3, our method, trained on view 4, generalizes well on unseen view 1, outperforming the TSM model that was directly trained on view 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.515, + 0.723 + ], + "angle": 0, + "content": "5.5 Egocentric Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Recognizing actions in egocentric videos is challenging due to camera motion and the occlusion of interacting objects by the hands. Additionally, in the case of Assembly101 [51], the egocentric cameras are similar to Oculus Quest VR headsets, which provide monochrome low-resolution frames. As a result, action recognition performance is significantly lower than fixed RGB views, as found in [51]. We address this challenging scenario with our proposed multimodal architecture and achieve state-of-the-art performance in egocentric action recognition on Assem" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.158, + 0.525, + 0.198 + ], + "angle": 0, + "content": "Table 4: Egocentric action recognition in Assembly101 [51]. TSM features from [51] are used as RGB frame features." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.2, + 0.528, + 0.276 + ], + "angle": 0, + "content": "
MethodActionVerbObject
TSM egocentric (fuse 4 views)33.8059.0046.50
Egocentric (e3) + Pose36.0765.5245.82
Egocentric (e4) + Pose35.5665.7945.20
Egocentric (e3+e4) + Pose38.0566.3247.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.304, + 0.512, + 0.331 + ], + "angle": 0, + "content": "Table 6: Micro-action length ablation for verb recognition in Assembly101 [51]" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.332, + 0.515, + 0.378 + ], + "angle": 0, + "content": "
#Frames1153060120
Verb \nAccuracy(%)59.1263.7063.6863.5162.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.548, + 0.158, + 0.776, + 0.199 + ], + "angle": 0, + "content": "Table 5: Keypoint ablation for verb recognition in Assembly101 [51]" + }, + { + "type": "table", + "bbox": [ + 0.559, + 0.2, + 0.772, + 0.277 + ], + "angle": 0, + "content": "
#JointsGlobal ReferenceVerb Accuracy (%)
21×64.17
21✓64.90
11✓64.77
6✓63.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.535, + 0.304, + 0.781, + 0.345 + ], + "angle": 0, + "content": "Table 7: Ablating unimodal temporal aggregation for verb recognition in Assembly101 [51]" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.347, + 0.776, + 0.376 + ], + "angle": 0, + "content": "
Temp. Agg.TCNLSTMTransformer
Verb Acc. (%)62.9563.3463.70
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.785, + 0.482 + ], + "angle": 0, + "content": "bly101. For this experiment, we used frame-wise TSM features provided by [51]. As depicted in Tab. 4, our model, using a single egocentric view (\\(e3\\) or \\(e4\\)) outperforms the fusion of four egocentric views as reported in [51]. Moreover, fusing \\(e3\\) and \\(e4\\) significantly enhances our model, resulting in a \\(4.25\\%\\) increase in action accuracy over the baseline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.402, + 0.523 + ], + "angle": 0, + "content": "5.6 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Keypoints. Not all joints of the hands are equally informative for understanding hand actions. For instance, fingertips exhibit greater mobility compared to the inner joints. Moreover, from an egocentric viewpoint, certain joints are more prone to self-occlusion than others. In Tab. 5, we present the impact of incorporating varying numbers of joints on verb recognition within the Assembly101 dataset [51]. For the case of 6 joints, we consider only the wrist joint along with the five fingertips. To expand to 11 joints, we incorporate all the joints along the index and the thumb, which are least affected by self-occlusion. We also show the effect of including the Global Wrist Token, which acts as a reference to the global motion of the hands while encoding micro-actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Micro-action length. As the resized input comprises a fixed number of pose frames, enlarging the window size for a micro-action decreases the number of micro-actions to aggregate, and vice-versa. In Tab. 6, we vary the micro-action length for verb recognition in the Assembly101 [51] using 6 joints per hand as input, i.e. fingertips and wrist joint. The input pose clip is temporally resized to \\(\\mathcal{T}' = 120\\) before breaking into micro-actions. Lengths 1 and 120 represent two extreme versions with frame-based and trajectory-based encoding, respectively, while the others conform to our micro-action-based formulation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.257, + 0.145, + 0.746, + 0.16 + ], + "angle": 0, + "content": "Table 8: Ablating tokenization and different losses for action recognition" + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.171, + 0.761, + 0.274 + ], + "angle": 0, + "content": "
Multimodal TokenizerFeature Ant. LossVerb & Object LossAction Accuracy (%)
Assembly101 [51]H2O [34]
XXX38.9885.95
✓XX40.1988.84
✓✓X40.2489.26
XX✓40.5690.50
✓✓✓41.0693.39
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.787, + 0.382 + ], + "angle": 0, + "content": "Temporal Aggregation. After extracting micro-action features, aggregation for action recognition can be done with any sequence model. In Tab. 7, we evaluate the effectiveness of different temporal aggregation methods for verb recognition in the Assembly101 dataset [51]. Here, we use 6 joints per hand as input, i.e., the fingertips and the wrist joint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.491 + ], + "angle": 0, + "content": "Loss components. To assess the individual contributions of different components, we begin with a basic configuration. We then systematically introduce each element to understand its impact on the overall performance as observed in Tab. 8. Incorporating modality interaction between RGB and pose at the micro-action level through a multimodal tokenizer enhances action accuracy. The introduction of auxiliary losses also has a positive impact, resulting in an overall improvement of \\(2.08\\%\\) for Assembly101 [51] and \\(7.44\\%\\) for H2O [34]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.359, + 0.54 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.787, + 0.702 + ], + "angle": 0, + "content": "With the growing interest in AR/VR and wearables, hand pose estimation has rapidly advanced, and holds promise as a compact and domain-independent modality to complement the visual input. To address the underexplored domain of using 3D hand pose as a modality for hand-object interaction recognition, we introduce HandFormer, a novel multimodal transformer that leverages dense sequences of 3D hand poses with sparsely sampled RGB frames to achieve state-of-the-art action recognition performance. Our model also reduces computational requirements, offering immediate significance across various low-resource applications in mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitations. Our method relies on the availability of hand poses, which, if extracted from the visual modality with pose estimation tools [27, 28], can encounter out-of-view scenarios and produce noisy poses. Our experiments reveal that these estimated poses can still achieve good accuracy, yet further research can be conducted to explicitly address this phenomenon. We also assume that a uniform sampling of RGB frames from each micro-action should provide good representations for understanding the semantic context. However, not all frames are equally important in understanding the action. In such a case, adaptive frame sampling methods can be employed, which we leave for future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.21 + ], + "angle": 0, + "content": "This research is supported by A\\*STAR under its National Robotics Programme (NRP) (Award M23NBK0053)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.235, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.322 + ], + "angle": 0, + "content": "1. Ahn, D., Kim, S., Hong, H., Ko, B.C.: Star-transformer: a spatio-temporal cross attention transformer for human action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3330-3339 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.786, + 0.365 + ], + "angle": 0, + "content": "2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.786, + 0.393 + ], + "angle": 0, + "content": "3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.786, + 0.421 + ], + "angle": 0, + "content": "4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.786, + 0.462 + ], + "angle": 0, + "content": "5. Bruce, X., Liu, Y., Chan, K.C.: Multimodal fusion via teacher-student network for indoor action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3199-3207 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.786, + 0.505 + ], + "angle": 0, + "content": "6. Bruce, X., Liu, Y., Zhang, X., Zhong, S.h., Chan, K.C.: Mmnet: A model-based multimodal network for human action recognition in rgb-d videos. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3522-3538 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.786, + 0.561 + ], + "angle": 0, + "content": "7. Caetano, C., Sena, J., Brémond, F., Dos Santos, J.A., Schwartz, W.R.: Skelemotion: A new representation of skeleton joint sequences based on motion information for 3d action recognition. In: 2019 16th IEEE international conference on advanced video and signal based surveillance (AVSS). pp. 1-8. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.786, + 0.602 + ], + "angle": 0, + "content": "8. Cao, C., Zhang, Y., Zhang, C., Lu, H.: Body joint guided 3-d deep convolutional descriptors for action recognition. IEEE transactions on cybernetics 48(3), 1095-1108 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.786, + 0.644 + ], + "angle": 0, + "content": "9. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.786, + 0.686 + ], + "angle": 0, + "content": "0. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.688, + 0.786, + 0.742 + ], + "angle": 0, + "content": "11. Cheng, K., Zhang, Y., He, X., Chen, W., Cheng, J., Lu, H.: Skeleton-based action recognition with shift graph convolutional network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 183-192 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.786, + 0.784 + ], + "angle": 0, + "content": "12. Chéron, G., Laptev, I., Schmid, C.: P-cnn: Pose-based cnn features for action recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 3218-3226 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "13. Cho, H., Kim, C., Kim, J., Lee, S., Ismayilzada, E., Baek, S.: Transformer-based unified recognition of two hands manipulating objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4769-4778 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.268, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "14. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European conference on computer vision (ECCV). pp. 720-736 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Das, S., Dai, R., Yang, D., Bremond, F.: Vpn++: Rethinking video-pose embeddings for understanding activities of daily living. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9703-9717 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "16. Das, S., Sharma, S., Dai, R., Bremond, F., Thonnat, M.: Vpn: Learning video-posed embedding for activities of daily living. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16. pp. 72–90. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "17. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.787, + 0.397 + ], + "angle": 0, + "content": "18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "19. Du, Y., Wang, W., Wang, L.: Hierarchical recurrent neural network for skeleton based action recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1110-1118 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.481 + ], + "angle": 0, + "content": "20. Duan, H., Zhao, Y., Chen, K., Lin, D., Dai, B.: Revisiting skeleton-based action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2969-2978 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.535 + ], + "angle": 0, + "content": "21. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "22. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "23. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "24. Gan, M., Liu, J., He, Y., Chen, A., Ma, Q.: Keyframe selection via deep reinforcement learning for skeleton-based gesture recognition. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.689 + ], + "angle": 0, + "content": "25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "26. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "27. Han, S., Liu, B., Cabezas, R., Twigg, C.D., Zhang, P., Petkau, J., Yu, T.H., Tai, C.J., Akbay, M., Wang, Z., et al.: Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG) 39(4), 87-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "28. Han, S., Wu, P.c., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., et al.: Umetrack: Unified multi-view end-to-end hand tracking for vr. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "30. Hou, J., Wang, G., Chen, X., Xue, J.H., Zhu, R., Yang, H.: Spatial-temporal attention res-tcn for skeleton-based dynamic hand gesture recognition. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "31. Hou, Y., Li, Z., Wang, P., Li, W.: Skeleton optical spectra-based action recognition using convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology 28(3), 807-811 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "32. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.36 + ], + "angle": 0, + "content": "33. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.402 + ], + "angle": 0, + "content": "34. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10138-10148 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "35. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.5 + ], + "angle": 0, + "content": "36. Lee, C.J., Zhang, R., Agarwal, D., Yu, T.C., Gunda, V., Lopez, O., Kim, J., Yin, S., Deng, B., Li, K., et al.: Echowrist: Continuous hand pose tracking and hand-object interaction recognition using low-power active acoustic sensing on a wristband. arXiv preprint arXiv:2401.17409 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.543 + ], + "angle": 0, + "content": "37. Li, C., Li, S., Gao, Y., Zhang, X., Li, W.: A two-stream neural network for pose-based hand gesture recognition. IEEE Transactions on Cognitive and Developmental Systems 14(4), 1594-1603 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.573 + ], + "angle": 0, + "content": "38. Li, J., Xie, X., Pan, Q., Cao, Y., Zhao, Z., Shi, G.: Sgm-net: Skeleton-guided multimodal network for action recognition. Pattern Recognition 104, 107356 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.628 + ], + "angle": 0, + "content": "39. Li, M., Chen, S., Chen, X., Zhang, Y., Wang, Y., Tian, Q.: Actional-structural graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3595-3603 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.67 + ], + "angle": 0, + "content": "40. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "41. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence 42(10), 2684-2701 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.742 + ], + "angle": 0, + "content": "42. Liu, Y., Zhang, S., Gowda, M.: Neuropose: 3d hand pose tracking using emg wearables. In: Proceedings of the Web Conference 2021. pp. 1471-1482 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.784 + ], + "angle": 0, + "content": "43. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "44. Liu, Z., Zhang, H., Chen, Z., Wang, Z., Ouyang, W.: Disentangling and unifying graph convolutions for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 143-152 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "M.S. Shamil et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Ma, J., Damen, D.: Hand-object interaction reasoning. In: 2022 18th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). pp. 1-8. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "46. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12999-13008 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.302 + ], + "angle": 0, + "content": "47. Patrick, M., Campbell, D., Asano, Y., Misra, I., Metze, F., Feichtenhofer, C., Vedaldi, A., Henriques, J.F.: Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems 34, 12493-12506 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.358 + ], + "angle": 0, + "content": "48. Plizzari, C., Cannici, M., Matteucci, M.: Spatial temporal transformer network for skeleton-based action recognition. In: Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10-15, 2021, Proceedings, Part III. pp. 694-701. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.415 + ], + "angle": 0, + "content": "49. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3d pose and tracking for human action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 640-649 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.457 + ], + "angle": 0, + "content": "50. Sabater, A., Alonso, I., Montesano, L., Murillo, A.C.: Domain and view-point agnostic hand action recognition. IEEE Robotics and Automation Letters 6(4), 7823-7830 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.514 + ], + "angle": 0, + "content": "51. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.556 + ], + "angle": 0, + "content": "52. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.787, + 0.599 + ], + "angle": 0, + "content": "53. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9869-9878 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.641 + ], + "angle": 0, + "content": "54. Shi, L., Zhang, Y., Cheng, J., Lu, H.: Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12026-12035 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.67 + ], + "angle": 0, + "content": "55. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.712 + ], + "angle": 0, + "content": "56. Soo Kim, T., Reiter, A.: Interpretable 3d human action analysis with temporal convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 20-28 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "58. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HandFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "60. Vemulapalli, R., Arrate, F., Chellappa, R.: Human action recognition by representing 3d skeletons as points in a lie group. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 588-595 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "61. Vondrick, C., Pirsiavash, H., Torralba, A.: Anticipating visual representations from unlabeled video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 98-106 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "62. Wang, J., Liu, Z., Wu, Y., Yuan, J.: Mining actionlet ensemble for action recognition with depth cameras. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 1290-1297. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.313 + ], + "angle": 0, + "content": "63. Weiyao, X., Muqing, W., Min, Z., Ting, X.: Fusion of skeleton and rgb features for rgb-d human action recognition. IEEE Sensors Journal 21(17), 19157-19164 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.37 + ], + "angle": 0, + "content": "64. Wen, Y., Pan, H., Yang, L., Pan, J., Komura, T., Wang, W.: Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21243-21253 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.411 + ], + "angle": 0, + "content": "65. Wen, Y., Tang, Z., Pang, Y., Ding, B., Liu, M.: Interactive spatiotemporal token attention network for skeleton-based general interactive action recognition. arXiv preprint arXiv:2307.07469 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.466 + ], + "angle": 0, + "content": "66. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13587-13597 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.508 + ], + "angle": 0, + "content": "67. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European conference on computer vision (ECCV). pp. 305–321 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "68. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "69. Zhang, S., Liu, X., Xiao, J.: On geometric features for skeleton-based action recognition using multilayer LSTM networks. In: 2017 IEEE winter conference on applications of computer vision (WACV). pp. 148-157. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.632 + ], + "angle": 0, + "content": "70. Zhang, X., Xu, C., Tian, X., Tao, D.: Graph edge convolutional neural networks for skeleton-based action recognition. IEEE transactions on neural networks and learning systems 31(8), 3047-3060 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.675 + ], + "angle": 0, + "content": "71. Zhang, Y., Wu, B., Li, W., Duan, L., Gan, C.: Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 3229-3237 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_origin.pdf b/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..808c0e5ff88e8d6c4b2cc4f9f029938587dbb8c3 --- /dev/null +++ b/2024/On the Utility of 3D Hand Poses for Action Recognition/23b1d504-2d99-41c2-8a52-059b2f2058b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28f9bac723326808bcd86922ca62e8d3ce234b5da05f52127a6d9e2a0e2cc791 +size 1335656 diff --git a/2024/On the Utility of 3D Hand Poses for Action Recognition/full.md b/2024/On the Utility of 3D Hand Poses for Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7fe3032ff40f4d0ea015954033dc72a5841c032c --- /dev/null +++ b/2024/On the Utility of 3D Hand Poses for Action Recognition/full.md @@ -0,0 +1,318 @@ +# On the Utility of 3D Hand Poses for Action Recognition + +Md Salman Shamil1, Dibyadip Chatterjee1, Fadime Sener2, Shugao Ma2, and Angela Yao1 + +1 National University of Singapore 2 Meta Reality Labs {salman, dibyadip, ayao}@comp.nus.edu.sg {famesener, shugao}@meta.com https://s-shamil.github.io/HandFormer/ + +Abstract. 3D hand pose is an underexplored modality for action recognition. Poses are compact yet informative and can greatly benefit applications with limited compute budgets. However, poses alone offer an incomplete understanding of actions, as they cannot fully capture objects and environments with which humans interact. We propose HandFormer, a novel multimodal transformer, to efficiently model hand-object interactions. HandFormer combines 3D hand poses at a high temporal resolution for fine-grained motion modeling with sparsely sampled RGB frames for encoding scene semantics. Observing the unique characteristics of hand poses, we temporally factorize hand modeling and represent each joint by its short-term trajectories. This factorized pose representation combined with sparse RGB samples is remarkably efficient and highly accurate. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods at $5 \times$ fewer FLOPs. With RGB, we achieve new state-of-the-art performance on Assembly101 and H2O with significant improvements in egocentric action recognition. + +Keywords: Skeleton-based action recognition $\cdot$ 3D hand poses $\cdot$ Multi-modal transformer + +# 1 Introduction + +The popularity of AR/VR headsets has driven interest in recognizing hand-object interactions, particularly through egocentric [14, 26] and multiview cameras [34, 51]. Such interactions are inherently fine-grained; recognizing them requires distinguishing subtle motions and object state changes. State-of-the-art methods for hand action recognition [23, 45, 47, 66] primarily rely on multi-or single-view RGB streams, which are computationally heavy and unsuitable for resource-constrained scenarios like AR/VR. Motivated by advancements in lightweight hand pose estimation methods leveraging monochrome cameras [27, 28, 46], and the evolution of low-dimensional sensors [36, 42] such as accelerometers, MMG, EMG, demonstrating real-time hand pose estimation, we advocate + +![](images/541766225c4522e643c67e6585bbf26257512ede07be096f1ea02cfe6679d827.jpg) +Fig. 1: We densely sample 3D hand poses to understand fine-grained hand motions and sparsely sample RGB frames to capture the scene semantics. Our 3D hand poses are acquired from low-resolution monochrome cameras but can also come from wearable sensors, facilitating an efficient understanding of hand-object interactions. Video frames and hand poses in the figure are from Assembly101 [51]. + +the use of 3D hand poses as an input modality for recognizing hand-object interactions. Hand poses are a compact yet informative representation that captures the motions and nuances of hand movements. + +Existing works on 3D pose-based action recognition have focused primarily on full-body skeletons [20,44,65,68]. Hand poses differ fundamentally from full-body skeletons. In full-body recognition datasets [41,52], the actions are predominantly static from a global perspective. The relative changes in joint or limb positions signify the action category, e.g., 'sitting down'. Conversely, hand joints typically move together for many actions and lack a static joint as a global reference [51], e.g., 'put down toy'. Full-body skeleton methods also benefit from modeling long-range spatiotemporal dependencies between joints [44], while this is less important for the hands as shown in Fig. 2 and in Sec. 3. + +However, the 3D pose alone cannot encode the hands' actions. Unlike the full-body case, where actions are self-contained by the sequence of poses, the hands are often manipulating objects [34, 46, 51]. Hand pose is excellent for identifying motions (verbs) but struggles with associated objects [51]. Therefore, supplementing pose data with visual context from images or videos is crucial for full semantic understanding. However, as noted earlier, using dense RGB frames contradicts our motivation for using hand poses. + +This paper introduces HandFormer, a novel and lightweight multimodal transformer that leverages dense 3D hand poses complemented with sparsely sampled RGB frames. To this end, we conceptualize an action as a sequence of short segments, termed micro-actions, which are analogous to words that form sentence-level complex actions. Each micro-action comprises a dense sequence of pose frames and a single RGB frame. As every hand joint moves in close spatial proximity, we encode the pose sequence from a Lagrangian view [49] and track each joint as an individual entity. HandFormer effectively models each micro-action whose features are then temporally aggregated for action classification. This formulation benefits from our observations on the spatiotemporal dynamics of hand movement and delivers strong performance. Additionally, limiting long-range spatiotemporal dependencies among joints and only using sparse RGB frame + +features improves computational efficiency, making it well-suited for always-on, low-compute AR/VR applications. + +Our contributions are: (i) We analyze the differences between hand pose and full-body skeleton actions and design a novel pose sequence encoding that reflects hand-specific properties. (ii) We propose HandFormer that takes a sequence of dense 3D hand pose and sparse RGB frames as input in the form of micro-actions. (iii) HandFormer achieves state-of-the-art action recognition on H2O [34] and Assembly101 [51]. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods in Assembly101, incurring at least $5 \times$ fewer FLOPs. (iv) We experimentally demonstrate how hand poses are crucial for multiview and egocentric action recognition. + +# 2 Related Work + +Video Action recognition. Video-based action recognition systems are well-developed with sophisticated 3D-CNN [9,23,57,67] or Video Transformer [2,3,43,47] architectures. However, they all bear significant computational expense for both feature extraction and motion modeling, either explicitly in the form of optical flow [9,33,55] or implicitly through the architecture [2,47]. Such designs are well-suited for high-facility applications but are not suitable for integration into lightweight systems. To this end, efficient video understanding is an active topic of research, with a focus on reducing expensive 3D operations [22,40,58,67], quadratic complexity of attention [32,43,47] and dropping tokens [4,21]. Additionally, densely-sampled RGB frames incur a high cost, yet limiting the temporal resolution will hinder fine-grained action understanding. We propose complementing 3D hand poses with sparsely-sampled RGB frames, developing a lightweight video understanding system. + +Skeleton-based Action Recognition. Skeleton-based action recognition has been tackled with hand-crafted features [60, 62], sequence models like RNNs and LSTMs [19, 69], and CNN-based methods that either employ temporal convolution on the pose sequence [56] or transform the skeleton data into pseudoimages to be processed with 2D or 3D convolutional networks [7, 20, 31]. Recent progress has been driven by GNN-based methods that exploit the graph structure of skeletal data to construct spatio-temporal graphs and perform graph convolutions [68, 70]. These methods often model functional links between joints that go beyond skeletal connectivity [39, 54] or expand the spatiotemporal receptive field [44]. Self-attention and transformer-based methods have also been proposed [48, 65, 71]. However, most existing methods are tailored for datasets involving full-body poses. Few works [24, 30, 37, 50] study hand poses but are restricted to simple gestures, recognizable without complex temporal modeling. + +Fusing RGB with Skeleton. Skeletal data has been used for cropping images of body parts [12], for weighting RGB patches around regions of interest [5, 6], or for pooling CNN features [1, 8]. Projecting into common embedding space is done in [16], enabling pose distillation [15]. Multi-stream architectures are also designed with separate paths for RGB and skeleton having lateral connec + +![](images/89eca3f064d23587c0fdd50bc5047e89169b4718ca4e7379d4a58d99aed647b9.jpg) +Stand up + +![](images/4854dfb504b50a8617f4ff20d3546cc7b4617e232960666444d5808d07a33c12.jpg) +Put on bag + +![](images/1e691409f9f81ee2def47f9d257d486d30f8c779640dcbd82165d358d0376828.jpg) +Sit down +Fig. 2: Comparing skeletal changes in full-body actions from NTU RGB+D 120 [41] (left) and hand actions from Assembly101 [51] (right). Two pose frames at interval $T$ are shown. $J_{j,t}$ indicates the 3D coordinate of joint $j$ at timestep $t$ , and $$ is the correlation between two such joints. Modeling the correlation between spatiotemporally distant joints can be informative for full-body poses but does not provide a useful action cue for hand poses. + +![](images/49650b0f05e7738b609bdb167cd0bb3e2c792ded9fa3c5bddb296200744cd313.jpg) +$\leftarrow \rightarrow \rightarrow < J_{\mathrm{H,0}}J_{\mathrm{ET}}>$ + +![](images/dbe57dbbea64fd2cababef3f172ce5dfee47975988e570815a1e00177c984d97.jpg) +←→ + +![](images/8dda702072f573aa6a3d8fbbf3661b7ed2199189f1bec8c2ec040a74b816060b.jpg) +Pick up + +![](images/ada46ecffa20204e520e1632ef3f414852d63775abcf0114a4f32976dea33fef.jpg) +Unscrew + +![](images/2f2bf0dc21d2efbfeb8adc3fa0d002dd6d5054614851679a0b50a12e9fea0d08.jpg) +Rotate + +tions between them [38, 63]. RGBPoseConv3D [20] is a two-stream model that employs 3D CNNs for both RGB and pose. Similar to skeleton-based action recognition, these multimodal approaches primarily focus on full-body poses. Some approaches simultaneously perform both hand pose estimation and action recognition, using pose data to supervise the training process [13, 64]. However, it is important to note that pose estimation deals with lower-level semantics than action recognition. Actions can often be inferred even when the estimated poses are not accurate [46]. + +# 3 Modeling Full-body vs. Hand Skeletons + +Existing skeleton-based action recognition datasets primarily consist of full-body poses where actions feature significant changes over time in limb positions relative to other body parts, which highly correlate to the action category. These changes can be captured through long-range spatiotemporal modeling using graph convolutions with large receptive fields [44] or self-attention [65]. Conversely, hand actions show diverse movement patterns, as hands can move in arbitrary directions, often without significant changes in articulation. We analyze pose sequences from NTU RGB+D 120 [41] and Assembly101 [51], calculating the distance covered by each joint and identifying the least and the most active joints. We observe that, for full-body poses, these two representative joints show a significant difference in the distances covered, while for hand poses, the difference is subtle. The Pearson correlation coefficient is 0.93 for hand poses (indicating highly coupled joints) and 0.33 for full-body poses. Please refer to Suppl. Sec. A for details regarding these computations and comparisons. + +An example showing the difference between body and hand poses is provided in Fig. 2, where we depict two frames separated by an interval of $T$ frames. For the full body skeleton, the head joint $J_{H}$ remains static, while the right hand joint $J_{F}$ moves upward. Conversely, for the hand, all the joints move, + +![](images/67d57c644999ed5a67f0c5cacd4a59432410b785efeeb0712f4f1315ba45e7c0.jpg) +Fig. 3: Overall architecture of HandFormer. An action segment is divided into $K$ micro-actions $\{M_1, M_2 \ldots M_K\}$ . Each micro-action comprises a dense sequence of pose frames and a single RGB frame. The frame encoder $F$ and the trajectory encoder $T$ encode the RGB and the dense poses, respectively, after which they are passed to a Multimodal Tokenizer. The modality-mixed tokens are then fed to a Temporal Transformer. Dotted paths are optional and only required when RGB is used. + +including the wrist joint $J_{W}$ and index fingertip $J_{I}$ . The spatiotemporal correlation between the head joint in the first frame and the hand joint in the last frame ( $$ ) is a crucial action cue that captures the relative structural change. However, such a correlation between the wrist and the fingertip for hand skeleton ( $$ ) does not provide a stronger cue than the wrist movement itself. To avoid such redundant spatiotemporally distant correlations, we propose to divide the hand pose sequence into micro-action blocks of fixed temporal length. This formulation allows for encoding short-term movements while enabling parameter sharing. + +Moreover, full-skeletal motion is a dominant feature in hand pose sequences, differentiating them from full-body poses. While the location and orientation of the human body can be trivial for most actions, this is not the case for hands. The hands do not conform to any particular 6D pose, and notably, they frequently and unpredictably alter their 6D poses over time. A comprehensive analysis of this phenomenon is provided in Suppl. Sec. A. To this end, we explicitly consider the global 6D poses of the hands during micro-action-based pose encoding. + +# 4 HandFormer + +Fig. 3 illustrates our proposed HandFormer, which consists of a sequence of micro-action blocks, a novel trajectory encoder, and a temporal aggregation module. The design of our HandFormer allows us to easily incorporate semantic context by sampling a single RGB frame from certain micro-actions. + +# 4.1 Micro-actions + +Given an action segment containing $\mathcal{T}$ frames sampled at a certain fps, the input to our model comprises — $i$ ) a dense sequence of 3D hand poses, represented as $\mathbf{S} = \{P_1,P_2,\dots ,P_{\mathcal{T}}\}$ , where $P_{t} = \{P_{t}^{left},P_{t}^{right}\} \in \mathbb{R}^{2\times J\times 3}$ signifying the 3D coordinates of $J$ keypoints in the left and right hands, respectively, and $ii)$ a sparse set of RGB frames sampled at intervals $\varDelta f$ , denoted as $\mathbf{V} = \{I_1,I_{1 + \varDelta f},\ldots ,I_{\mathcal{T}_r}\}$ , where $I_{t}\in \mathbb{R}^{H\times W\times 3}$ are frame-wise RGB features and $\mathcal{T}_r = 1 + \left|\frac{\mathcal{T} - 1}{\varDelta f}\right|\times \varDelta f$ . + +We factorize the raw input into a sequence of $K$ micro-action blocks of length $N$ frames, obtained by shifting the window across the action segment with a stride of $R$ frames. Each block consists of two components — the initial appearance derived from the first RGB frame within the block and the hand motion characterized by the dense sequence of $N$ pose frames. To obtain a fixed length input containing $\mathcal{T}' = (K - 1)\times R + N$ pose frames, we perform linear interpolation for each joint along the temporal axis, transforming pose sequence $\mathbf{S}$ having $\mathcal{T}$ frames to $\mathbf{S}' = \{P_1',P_2',\dots ,P_{\mathcal{T}'}'\}$ having $\mathcal{T}'$ frames, where $P_t' \in \mathbb{R}^{2\times J\times 3}$ . Thus, we represent the input as a sequence of micro-actions $\mathbf{M} = \{M_1,M_2,\dots ,M_K\}$ , derived from $\mathbf{V}$ and $\mathbf{S}'$ through the following equations: + +$$ +M _ {k} = \left[ M _ {k} ^ {\mathrm {R G B}}, M _ {k} ^ {\text {P o s e}} \right] = \left[ I _ {h (k)}, \left\{P _ {g (k) + i} ^ {\prime} \right\} _ {i = 0} ^ {N - 1} \right], \tag {1} +$$ + +where $g(k) = (k - 1)\times R + 1$ denotes the first pose frame in $k$ -th Micro-action, and $h(k)$ determines the index of the nearest available RGB frame. + +The dense pose sequence in a micro-action captures fine-grained hand motion crucial for recognizing verbs, whereas a single RGB frame provides semantic context for recognizing objects. To extract features from micro-actions, we use a frame encoder $F$ and a trajectory encoder $T$ , which operate on RGB frames and pose sequences, respectively. Consequently, the RGB and pose features for the $k^{th}$ micro-action is given by- + +$$ +\left[ f _ {k} ^ {\mathrm {R G B}}, f _ {k} ^ {\mathrm {P o s e}} \right] = \left[ F \left(M _ {k} ^ {\mathrm {R G B}}\right), T \left(M _ {k} ^ {\mathrm {P o s e}}\right) \right], \tag {2} +$$ + +where $f_{k}^{\mathrm{RGB}}, f_{k}^{\mathrm{pos}} \in \mathbb{R}^{d}$ , where $d$ denotes the common dimensionality of both RGB and pose embedding space. + +# 4.2 Trajectory Encoder + +We devise a trajectory-based pose encoder to encode dense hand pose sequences within micro-actions, illustrated in Fig. 4. Each joint is represented by its trajectory of dimension $3 \times N$ , encapsulating the sequence of 3D coordinates across the $N$ pose frames of a micro-action. This yields $2 \times J$ feature vectors for the $J$ joints of two hands. Each joint's trajectory is passed through a TCN [35], whose parameters are shared for all joints. This produces $2 \times J$ Local Trajectory Tokens. Additionally, the full-skeletal motion of the hand during the entire action is used as a reference through an additional token named Global Wrist Token. This token is generated by a separate TCN operating on the sequence of 6D + +![](images/c7827b62edb38da93d51875c2a450e292076fc150ac7301369a31088c7b59593.jpg) +Fig. 4: Our Trajectory Encoder $T$ , which operates on micro-actions, derives tokens with trajectory-based features and performs self-attention to encode the pose sequence into a feature vector. The Single-Joint TCN is a Temporal Convolutional Network [35] that processes the trajectories of all the joints individually with shared parameters across all joints. Wrist-TCN takes an action-wide sequence of wrist location and hand orientation (6D pose) to produce a global reference token. + +poses of hands, indicating wrist location and hand orientation. Subsequently, a self-attention layer is applied to these trajectory tokens, preserving the temporal dimension for subsequent stages. This iterative process culminates in spatiotemporal average pooling, summarizing the hand motion of the micro-action. + +# 4.3 Multimodal Tokenizer + +This section discusses incorporating sparsely sampled RGB frames into Hand-Former to better capture scene semantics. A single frame is sampled from each micro-action, generating an extended crop $(1.25\times)$ around the hands. While the full image provides an overall scene context, the crop focuses specifically on hand-object interaction regions [10]. Features for both are separately generated using a pre-existing image encoder and then aggregated to enrich the hand-object interaction feature with scene context. The hand-object ROI crop can be obtained using an off-the-shelf HOI detector [53] or by using the corresponding hand pose projections. We opt for the latter. + +Our multimodal tokenizer in Fig. 3 generates RGB and Pose tokens enhanced via multimodal interactions. This involves concatenating each frame feature and trajectory encoding, and projecting them into a shared PoseRGB feature space using an MLP. This PoseRGB feature is then split into two parts, which are added to the original frame feature and trajectory encoding, respectively. + +# 4.4 Temporal Transformer + +The multimodal tokenizer provides RGB and pose tokens for each micro-action. Since an action segment consists of a sequence of micro-actions, these micro-action tokens are aggregated over time via a temporal transformer. A video $\mathbf{V}$ , + +divided into $K$ micro-actions can be represented by two sets of tokens $\{\hat{f}_k^{\mathrm{RGB}}\}_{k=1}^K$ and $\{\hat{f}_k^{\mathrm{Pose}}\}_{k=1}^K$ , respectively, produced by the multimodal tokenizer. These $2 \times K$ tokens of dimension $d$ form the input sequence of the temporal transformer. Positional encoding and modality embedding are added to each input token to indicate the temporal location and source modality. We use the fixed sine/cosine positional encoding [59], assigning the same position to two tokens from the same micro-action. Modality embeddings are learned and shared across the tokens from the same modality (RGB or Pose). Following standard practice [17,18], we preprocess an additional learnable class token $[\mathrm{CLS}] \in \mathbb{R}^d$ , the output of which is then used for classifying actions. + +# 4.5 Learning Objectives + +HandFormer is trained end-to-end for action recognition, supervised by a cross-entropy loss: $\mathcal{L}_{cls} = -a_i\log \hat{a}_i$ , where $a_{i}$ represents the ground truth action label for the $i^{th}$ sample, and $\hat{a}_i$ denotes the predicted action category. The input pose and RGB modalities provide complementary information regarding a scene, capturing motion and interacting objects, respectively. To effectively utilize this, we employ explicit verb and object supervision ( $\mathcal{L}_{verb}$ , $\mathcal{L}_{obj}$ ) via two additional learnable class tokens. Given that pose strongly correlates with verbs, whereas objects are identifiable from RGB frames alone, we let the verb class token attend exclusively to pose encodings and the object class token to frame features. + +Feature Anticipation Loss. Hand pose sequence captures the primary sources of state changes during hand-object interactions. We posit that the visual state from an initial RGB frame, in combination with the subsequent hand pose sequence, is indicative of the visual state that results from the completion of the sequence. Therefore, given an RGB feature and the corresponding pose features from a micro-action, we force our model to anticipate the RGB feature for the next micro-action by minimizing an $L_{1}$ feature loss. This loss, inspired by existing efforts [25,61], quantifies the difference between the anticipated image feature and the true feature extracted from a frozen image encoder. Formally, + +$$ +\mathcal {L} _ {a n t} = \sum_ {k = 1} ^ {K - 1} \left\| \Phi_ {\text {a n t}} \left(f _ {k} ^ {\text {P o s e R G B}}\right) - f _ {k + 1} ^ {\text {R G B}} \right\| _ {1} \tag {3} +$$ + +where $\varPhi_{\mathrm{ant}}$ denotes a linear projection layer. Hence, the total loss is + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \lambda_ {1} \mathcal {L} _ {v e r b} + \lambda_ {2} \mathcal {L} _ {o b j} + \lambda_ {3} \mathcal {L} _ {a n t} \tag {4} +$$ + +where $\lambda_1, \lambda_2, \lambda_3$ are hyperparameters used to balance the four losses. + +# 5 Experiments + +Datasets. We conduct experiments on two publicly available hand-object interaction datasets. Assembly101 [51] is a large-scale multiview dataset that fea- + +tures videos of procedural activities for assembling and disassembling 101 take-aside toy vehicles. This dataset has 1380 fine-grained actions based on 24 verbs and 90 objects. It consists of 12 temporally synchronized views — eight static and four egocentric. The egocentric views exhibit low-resolution monochrome images, making it challenging even for human eyes to discern objects in view. Hence, we opt for fixed views as our RGB modality. 3D hand poses estimated using off-the-shelf MEgATrack [27] are provided by the dataset. H2O [34] features four participants performing 36 actions that involve 8 objects and 11 verbs. Ground truth 3D hand poses are provided. The dataset consists of 5 temporally synchronized RGB views — 4 fixed and 1 egocentric. We use RGB frames from the egocentric view only. + +Implementation Details. Similar to [65], our pose input consists of 120 frames by setting $\mathcal{T}' = 120$ . The number of frames per micro-action ( $N$ ) is 15. For multimodal experiments, we set the window stride $R = N$ and take $K = 8$ non-overlapping micro-actions, which allows us to sample 8 RGB frames for each video following [51]. However, we allow a $50\%$ overlap between consecutive micro-actions for pose-only variants. A frozen ViT [18] or ResNet [29], followed by a learnable linear layer, is used as the frame encoder $F$ . More details are provided in Suppl. Sec. C. Each model is trained for 50 epochs using SGD with momentum 0.9, a batch size of 32, a learning rate of 0.025, and step LR decay with a factor of 0.1 at epochs {25, 40}. Loss hyperparameters $\{\lambda_1, \lambda_2, \lambda_3\}$ are chosen to be {1.0, 1.0, 2.0}. + +Model Variants. We propose several variants of our model by adjusting the width $d$ and the number of layers $T_{n}$ of the transformer to balance efficiency and accuracy. Our default HandFormer, denoted HandFormer-B, has parameters $(d,T_n) = (256,2)$ . We introduce a larger variant, HandFormer-L, with $(d,T_{n}) = (512,4)$ . We also explore different configurations for the number of input joints $J$ per hand. Unless otherwise mentioned, we utilize all 21 joints per hand along with the base model denoted as HandFormer-B/21 while offering a highly efficient option utilizing only six joints per hand (five fingertips and the wrist) termed HandFormer-B/6. Additionally, HandFormer-X/J×N denotes a multimodal variant that utilizes $\mathcal{N}$ RGB frames, where $\mathcal{N} \leq K$ . + +# 5.1 Comparison with State-of-the-Art + +To evaluate the effectiveness of our proposed architecture, we compare it against several baselines. For pose-only comparisons, we choose a graph-based network MS-G3D [44] and an attention-based method ISTA-Net [65] — the two best performing skeleton-based methods on Assembly101 [51], as reported by [65]. We also employ video baselines TSM [40] and SlowFast [23], emphasizing efficiency and high temporal resolution, respectively. Additionally, we include H2OTR [13], the state-of-the-art for the H2O dataset. For Assembly101, we replicate the results of the methods above using RGB frames from view 4 of the dataset, while results on H2O are acquired from the respective papers. Furthermore, on both datasets, we train and test RGBPoseConv3D [20], which is current state-of-the-art in multimodal action recognition with skeleton and RGB data. + +Table 1: Quantitative comparison with state-of-the-art methods on Assembly101 [51] and H2O [34]. For H2O, 6D object pose is used by ISTA-Net during training and inference and by H2OTR during training. 'MS-G3D + TSM' denotes a late fusion of the corresponding unimodal architectures. Reported GFLOPs include pose estimation costs in the context of Assembly101 at 15 fps. †For pose-only methods, pose estimation costs are common and excluded for simpler comparison. + +
MethodPoseRGBGFLOPsAssembly101H2O Action
ActionVerbObject
MS-G3D [44]✓✗\( 21.2^† \)28.7863.4637.2650.83
ISTA-Net [65]✓✗\( 35.2^† \)28.1462.7036.7789.09
SlowFast [23]✗✓65.7---77.69
TSM [40]✗✓33.035.2758.2747.45-
H2OTR [13]✗✓----90.90
RGBPoseConv3D [20]✓✓68.933.6161.9942.9083.47
MS-G3D + TSM✓✓66.239.7465.1251.12-
HandFormer-B/21×8✓✗\( 4.2^† \)28.8065.3336.2857.44
✗✓33.032.0755.6144.8984.71
✓✓47.641.0669.2351.1793.39
+ +We evaluate three variants of our method by controlling the input modalities. As shown in Tab. 1, our unimodal pose-only model excels in verb recognition on Assembly101. In contrast, RGB-based methods benefit from object appearance and usually perform well for actions due to the strong object recognition. + +In this context, the accuracy of ISTA-Net on H2O is not directly comparable to our method, as they also use the 6D object poses, while we only use hand poses as input. RGBPoseConv3D struggles to achieve satisfactory performance, particularly on Assembly101, suggesting that generalizing to hand poses is non-trivial for skeleton-based methods. Therefore, we combine two best-performing unimodal methods with late fusion (MS-G3D + TSM) to create a stronger baseline. Our model even outperforms this, indicating the effectiveness of our proposed multimodal fusion. + +# 5.2 Skeleton-based Action Recognition for Hands + +The compositional nature of action classes in hand-object interaction videos allows us to break down the action into a verb and an object. Recognizing such actions from 3D hand poses is an ill-posed problem, as the hand skeletons lack explicit information about the interacting objects, which are also part of the action semantics. However, as the pose data completely captures the hand motion information, it can be reliably used for verb recognition. Therefore, we evaluate a pose-only version of our method on the verb recognition task and compare the performance and efficiency metrics with other state-of-the-art skeleton-based methods in Fig. 5. Our method uses significantly fewer GFLOPs due to our spatiotemporal factorization using micro-actions. With $J = 21$ , all our HandFormer variants outperform existing methods. The $J = 6$ variant is exceptionally efficient with comparable accuracy, notably surpassing the efficiency-focused Shift-GCN [11]. HandFormer-H/21 combines the HandFormer-B and HandFormer-L + +![](images/c8bbdbfd31fc2fbb0900f212d79be075e43c4f514dd24c16236e70b7b51bfb0d.jpg) +Fig.5: Comparison of skeleton-based methods for verb recognition on Assembly101 [51]. Our method achieves state-of-the-art performance while utilizing significantly fewer FLOPs. + +![](images/4d598a8731ddaaa4af32e72798ed58dcd48f61e2a09d42f0c5ca07b84907520a.jpg) +Fig.6: Ablating number of RGB frames in HandFormer on Assembly101 [51]. With more RGB frames, verb recognition shows marginal gain, whereas object recognition shows improvement with diminishing returns. + +variants and becomes our best-performing model, improving over MS-G3D by $2.04\%$ while maintaining comparable FLOPs. + +# 5.3 How many RGB frames are required? + +In our model, the pose modality maintains a high temporal resolution to capture fine-grained hand movements, resulting in good verb recognition performance. On the contrary, RGB frames are primarily required to introduce semantic context for object recognition and does not necessitate a high temporal resolution like hand movements. The design of our model allows us to sample only a few RGB frames (as low as one) but still perform competitively at a reduced computational cost. Fig. 6 shows the impact of using more RGB frames for Assembly101 [51]. For this experiment, non-overlapping micro-actions are considered. Using only one RGB frame in HandFormer (35.46) outperforms the video model TSM (35.27), as shown in Tab. 1. This performance gain stems primarily from a significant improvement in object accuracy, with only a slight enhancement in verb accuracy. However, including more RGB frames shows a diminishing return, which is unsurprising as additional frames are expected to provide redundant information. These results are obtained with a simplified version of our model by setting $\mathcal{L} = \mathcal{L}_{cls}$ and bypassing the multimodal tokenizer. + +# 5.4 Can 3D hand pose be an efficient alternative to multi-view? + +While multiview action recognition benefits from precise hand-movement information in 3D space, processing all views with video models is expensive and highly redundant. In Tab. 2, we demonstrate that combining 3D hand pose with a single RGB view (view 4) achieves comparable performance to multiview action recognition on Assembly101 [51]. Specifically, our action recognition + +Table 2: Multi-view Action Recognition on Assembly101 [51]. HandFormer with 'Single View+Pose' has better performance than 'Single View+Egocentric'. Also, verb performance is comparable to using all (8) RGB views. + +
ViewsActionVerbObject
Single View35.2758.2747.45
Single View + Egocentric37.7561.8049.43
Two Views41.9665.2253.26
All (8) Views47.5170.9957.73
Single View + Pose41.0669.2351.17
+ +Table 3: Cross-view performance of HandFormer in Assembly101 [51] shows its generalization capability to unseen view 1, outperforming the video baseline which is trained on view 1 directly. Egocentric views are the source of hand poses here and, therefore, are included in the video models. + +
MethodTrain onTest onActionVerbObject
TSM [40]v4 + egov4 + ego37.7561.8049.43
v4 + egov1 + ego35.2759.5347.72
v1 + egov1 + ego36.2160.7848.52
Our Methodv4 + Posev4 + Pose41.0669.2351.17
v4 + Posev1 + Pose38.4367.8648.32
+ +performance ('Single View + Pose') matches the fusion of the two most informative views — view 1 and view 4. Notably, our verb recognition accuracy is on par with combining all 8 RGB views, though the latter incurs at least $5 \times$ more FLOPs despite using the efficient TSM model. Additionally, using hand pose in combination with an RGB view ('Single View + Pose') outperforms directly using the egocentric videos ('Single View + Egocentric') from which the hand poses are derived. While fusing multiple RGB views improves accuracy by ensembling multiple complementary predictions, the computational overhead increases significantly. In contrast, our model processes hand pose and single-view RGB frames, enhancing efficiency by leveraging the less redundant and low-dimensional pose data. + +Cross-view Generalization. 3D hand pose offers a unique opportunity for cross-view generalization because of its universality across different viewpoints. To evaluate the effectiveness of our method for unseen views, we train our model with frame-wise RGB features from view 4 and test it on view 1. We train TSM [40] as a baseline video model on both RGB views separately. As our method includes 3D hand poses obtained from the egocentric views, we include the egocentric videos in the TSM baselines for a fair comparison. As seen from Tab. 3, our method, trained on view 4, generalizes well on unseen view 1, outperforming the TSM model that was directly trained on view 1. + +# 5.5 Egocentric Action Recognition + +Recognizing actions in egocentric videos is challenging due to camera motion and the occlusion of interacting objects by the hands. Additionally, in the case of Assembly101 [51], the egocentric cameras are similar to Oculus Quest VR headsets, which provide monochrome low-resolution frames. As a result, action recognition performance is significantly lower than fixed RGB views, as found in [51]. We address this challenging scenario with our proposed multimodal architecture and achieve state-of-the-art performance in egocentric action recognition on Assem + +Table 4: Egocentric action recognition in Assembly101 [51]. TSM features from [51] are used as RGB frame features. + +
MethodActionVerbObject
TSM egocentric (fuse 4 views)33.8059.0046.50
Egocentric (e3) + Pose36.0765.5245.82
Egocentric (e4) + Pose35.5665.7945.20
Egocentric (e3+e4) + Pose38.0566.3247.86
+ +Table 6: Micro-action length ablation for verb recognition in Assembly101 [51] + +
#Frames1153060120
Verb +Accuracy(%)59.1263.7063.6863.5162.29
+ +Table 5: Keypoint ablation for verb recognition in Assembly101 [51] + +
#JointsGlobal ReferenceVerb Accuracy (%)
21×64.17
21✓64.90
11✓64.77
6✓63.70
+ +Table 7: Ablating unimodal temporal aggregation for verb recognition in Assembly101 [51] + +
Temp. Agg.TCNLSTMTransformer
Verb Acc. (%)62.9563.3463.70
+ +bly101. For this experiment, we used frame-wise TSM features provided by [51]. As depicted in Tab. 4, our model, using a single egocentric view ( $e3$ or $e4$ ) outperforms the fusion of four egocentric views as reported in [51]. Moreover, fusing $e3$ and $e4$ significantly enhances our model, resulting in a $4.25\%$ increase in action accuracy over the baseline. + +# 5.6 Ablation Studies + +Keypoints. Not all joints of the hands are equally informative for understanding hand actions. For instance, fingertips exhibit greater mobility compared to the inner joints. Moreover, from an egocentric viewpoint, certain joints are more prone to self-occlusion than others. In Tab. 5, we present the impact of incorporating varying numbers of joints on verb recognition within the Assembly101 dataset [51]. For the case of 6 joints, we consider only the wrist joint along with the five fingertips. To expand to 11 joints, we incorporate all the joints along the index and the thumb, which are least affected by self-occlusion. We also show the effect of including the Global Wrist Token, which acts as a reference to the global motion of the hands while encoding micro-actions. + +Micro-action length. As the resized input comprises a fixed number of pose frames, enlarging the window size for a micro-action decreases the number of micro-actions to aggregate, and vice-versa. In Tab. 6, we vary the micro-action length for verb recognition in the Assembly101 [51] using 6 joints per hand as input, i.e. fingertips and wrist joint. The input pose clip is temporally resized to $\mathcal{T}' = 120$ before breaking into micro-actions. Lengths 1 and 120 represent two extreme versions with frame-based and trajectory-based encoding, respectively, while the others conform to our micro-action-based formulation. + +Table 8: Ablating tokenization and different losses for action recognition + +
Multimodal TokenizerFeature Ant. LossVerb & Object LossAction Accuracy (%)
Assembly101 [51]H2O [34]
XXX38.9885.95
✓XX40.1988.84
✓✓X40.2489.26
XX✓40.5690.50
✓✓✓41.0693.39
+ +Temporal Aggregation. After extracting micro-action features, aggregation for action recognition can be done with any sequence model. In Tab. 7, we evaluate the effectiveness of different temporal aggregation methods for verb recognition in the Assembly101 dataset [51]. Here, we use 6 joints per hand as input, i.e., the fingertips and the wrist joint. + +Loss components. To assess the individual contributions of different components, we begin with a basic configuration. We then systematically introduce each element to understand its impact on the overall performance as observed in Tab. 8. Incorporating modality interaction between RGB and pose at the micro-action level through a multimodal tokenizer enhances action accuracy. The introduction of auxiliary losses also has a positive impact, resulting in an overall improvement of $2.08\%$ for Assembly101 [51] and $7.44\%$ for H2O [34]. + +# 6 Conclusion + +With the growing interest in AR/VR and wearables, hand pose estimation has rapidly advanced, and holds promise as a compact and domain-independent modality to complement the visual input. To address the underexplored domain of using 3D hand pose as a modality for hand-object interaction recognition, we introduce HandFormer, a novel multimodal transformer that leverages dense sequences of 3D hand poses with sparsely sampled RGB frames to achieve state-of-the-art action recognition performance. Our model also reduces computational requirements, offering immediate significance across various low-resource applications in mobile devices. + +Limitations. Our method relies on the availability of hand poses, which, if extracted from the visual modality with pose estimation tools [27, 28], can encounter out-of-view scenarios and produce noisy poses. Our experiments reveal that these estimated poses can still achieve good accuracy, yet further research can be conducted to explicitly address this phenomenon. We also assume that a uniform sampling of RGB frames from each micro-action should provide good representations for understanding the semantic context. However, not all frames are equally important in understanding the action. In such a case, adaptive frame sampling methods can be employed, which we leave for future work. + +# Acknowledgements + +This research is supported by A\*STAR under its National Robotics Programme (NRP) (Award M23NBK0053). + +# References + +1. Ahn, D., Kim, S., Hong, H., Ko, B.C.: Star-transformer: a spatio-temporal cross attention transformer for human action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3330-3339 (2023) +2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021) +3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021) +4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022) +5. Bruce, X., Liu, Y., Chan, K.C.: Multimodal fusion via teacher-student network for indoor action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3199-3207 (2021) +6. Bruce, X., Liu, Y., Zhang, X., Zhong, S.h., Chan, K.C.: Mmnet: A model-based multimodal network for human action recognition in rgb-d videos. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3522-3538 (2022) +7. Caetano, C., Sena, J., Brémond, F., Dos Santos, J.A., Schwartz, W.R.: Skelemotion: A new representation of skeleton joint sequences based on motion information for 3d action recognition. In: 2019 16th IEEE international conference on advanced video and signal based surveillance (AVSS). pp. 1-8. IEEE (2019) +8. Cao, C., Zhang, Y., Zhang, C., Lu, H.: Body joint guided 3-d deep convolutional descriptors for action recognition. IEEE transactions on cybernetics 48(3), 1095-1108 (2017) +9. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +0. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +11. Cheng, K., Zhang, Y., He, X., Chen, W., Cheng, J., Lu, H.: Skeleton-based action recognition with shift graph convolutional network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 183-192 (2020) +12. Chéron, G., Laptev, I., Schmid, C.: P-cnn: Pose-based cnn features for action recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 3218-3226 (2015) +13. Cho, H., Kim, C., Kim, J., Lee, S., Ismayilzada, E., Baek, S.: Transformer-based unified recognition of two hands manipulating objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4769-4778 (2023) + +14. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European conference on computer vision (ECCV). pp. 720-736 (2018) +15. Das, S., Dai, R., Yang, D., Bremond, F.: Vpn++: Rethinking video-pose embeddings for understanding activities of daily living. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9703-9717 (2021) +16. Das, S., Sharma, S., Dai, R., Bremond, F., Thonnat, M.: Vpn: Learning video-posed embedding for activities of daily living. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16. pp. 72–90. Springer (2020) +17. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +19. Du, Y., Wang, W., Wang, L.: Hierarchical recurrent neural network for skeleton based action recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1110-1118 (2015) +20. Duan, H., Zhao, Y., Chen, K., Lin, D., Dai, B.: Revisiting skeleton-based action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2969-2978 (2022) +21. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022) +22. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020) +23. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) +24. Gan, M., Liu, J., He, Y., Chen, A., Ma, Q.: Keyframe selection via deep reinforcement learning for skeleton-based gesture recognition. IEEE Robotics and Automation Letters (2023) +25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021) +26. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022) +27. Han, S., Liu, B., Cabezas, R., Twigg, C.D., Zhang, P., Petkau, J., Yu, T.H., Tai, C.J., Akbay, M., Wang, Z., et al.: Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG) 39(4), 87-1 (2020) +28. Han, S., Wu, P.c., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., et al.: Umetrack: Unified multi-view end-to-end hand tracking for vr. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) + +29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +30. Hou, J., Wang, G., Chen, X., Xue, J.H., Zhu, R., Yang, H.: Spatial-temporal attention res-tcn for skeleton-based dynamic hand gesture recognition. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018) +31. Hou, Y., Li, Z., Wang, P., Li, W.: Skeleton optical spectra-based action recognition using convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology 28(3), 807-811 (2016) +32. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021) +33. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019) +34. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10138-10148 (2021) +35. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017) +36. Lee, C.J., Zhang, R., Agarwal, D., Yu, T.C., Gunda, V., Lopez, O., Kim, J., Yin, S., Deng, B., Li, K., et al.: Echowrist: Continuous hand pose tracking and hand-object interaction recognition using low-power active acoustic sensing on a wristband. arXiv preprint arXiv:2401.17409 (2024) +37. Li, C., Li, S., Gao, Y., Zhang, X., Li, W.: A two-stream neural network for pose-based hand gesture recognition. IEEE Transactions on Cognitive and Developmental Systems 14(4), 1594-1603 (2021) +38. Li, J., Xie, X., Pan, Q., Cao, Y., Zhao, Z., Shi, G.: Sgm-net: Skeleton-guided multimodal network for action recognition. Pattern Recognition 104, 107356 (2020) +39. Li, M., Chen, S., Chen, X., Zhang, Y., Wang, Y., Tian, Q.: Actional-structural graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3595-3603 (2019) +40. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019) +41. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence 42(10), 2684-2701 (2019) +42. Liu, Y., Zhang, S., Gowda, M.: Neuropose: 3d hand pose tracking using emg wearables. In: Proceedings of the Web Conference 2021. pp. 1471-1482 (2021) +43. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022) +44. Liu, Z., Zhang, H., Chen, Z., Wang, Z., Ouyang, W.: Disentangling and unifying graph convolutions for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 143-152 (2020) + +45. Ma, J., Damen, D.: Hand-object interaction reasoning. In: 2022 18th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). pp. 1-8. IEEE (2022) +46. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12999-13008 (2023) +47. Patrick, M., Campbell, D., Asano, Y., Misra, I., Metze, F., Feichtenhofer, C., Vedaldi, A., Henriques, J.F.: Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems 34, 12493-12506 (2021) +48. Plizzari, C., Cannici, M., Matteucci, M.: Spatial temporal transformer network for skeleton-based action recognition. In: Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10-15, 2021, Proceedings, Part III. pp. 694-701. Springer (2021) +49. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3d pose and tracking for human action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 640-649 (2023) +50. Sabater, A., Alonso, I., Montesano, L., Murillo, A.C.: Domain and view-point agnostic hand action recognition. IEEE Robotics and Automation Letters 6(4), 7823-7830 (2021) +51. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022) +52. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016) +53. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9869-9878 (2020) +54. Shi, L., Zhang, Y., Cheng, J., Lu, H.: Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12026-12035 (2019) +55. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014) +56. Soo Kim, T., Reiter, A.: Interpretable 3d human action analysis with temporal convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 20-28 (2017) +57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015) +58. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018) +59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) + +60. Vemulapalli, R., Arrate, F., Chellappa, R.: Human action recognition by representing 3d skeletons as points in a lie group. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 588-595 (2014) +61. Vondrick, C., Pirsiavash, H., Torralba, A.: Anticipating visual representations from unlabeled video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 98-106 (2016) +62. Wang, J., Liu, Z., Wu, Y., Yuan, J.: Mining actionlet ensemble for action recognition with depth cameras. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 1290-1297. IEEE (2012) +63. Weiyao, X., Muqing, W., Min, Z., Ting, X.: Fusion of skeleton and rgb features for rgb-d human action recognition. IEEE Sensors Journal 21(17), 19157-19164 (2021) +64. Wen, Y., Pan, H., Yang, L., Pan, J., Komura, T., Wang, W.: Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21243-21253 (2023) +65. Wen, Y., Tang, Z., Pang, Y., Ding, B., Liu, M.: Interactive spatiotemporal token attention network for skeleton-based general interactive action recognition. arXiv preprint arXiv:2307.07469 (2023) +66. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13587-13597 (2022) +67. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European conference on computer vision (ECCV). pp. 305–321 (2018) +68. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) +69. Zhang, S., Liu, X., Xiao, J.: On geometric features for skeleton-based action recognition using multilayer LSTM networks. In: 2017 IEEE winter conference on applications of computer vision (WACV). pp. 148-157. IEEE (2017) +70. Zhang, X., Xu, C., Tian, X., Tao, D.: Graph edge convolutional neural networks for skeleton-based action recognition. IEEE transactions on neural networks and learning systems 31(8), 3047-3060 (2019) +71. Zhang, Y., Wu, B., Li, W., Duan, L., Gan, C.: Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 3229-3237 (2021) \ No newline at end of file diff --git a/2024/On the Utility of 3D Hand Poses for Action Recognition/images.zip b/2024/On the Utility of 3D Hand Poses for Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f15ac861f7ae8a53043543b64c87e29deec6016a --- /dev/null +++ b/2024/On the Utility of 3D Hand Poses for Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bb63337411dae25e6be92f1084a71b14b5a8423eb4bcedd724e278090f1b1dd +size 359459 diff --git a/2024/On the Utility of 3D Hand Poses for Action Recognition/layout.json b/2024/On the Utility of 3D Hand Poses for Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d9db1b2079cbd33ed1cfc5d54dc5c26d4dbe1eb3 --- /dev/null +++ b/2024/On the Utility of 3D Hand Poses for Action Recognition/layout.json @@ -0,0 +1,10191 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 179, + 111, + 434, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 111, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 179, + 111, + 434, + 148 + ], + "type": "text", + "content": "On the Utility of 3D Hand Poses for Action Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "spans": [ + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "content": "Md Salman Shamil1, Dibyadip Chatterjee1, Fadime Sener2, Shugao Ma2, and Angela Yao1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 211, + 201, + 402, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 201, + 402, + 258 + ], + "spans": [ + { + "bbox": [ + 211, + 201, + 402, + 258 + ], + "type": "text", + "content": "1 National University of Singapore 2 Meta Reality Labs {salman, dibyadip, ayao}@comp.nus.edu.sg {famesener, shugao}@meta.com https://s-shamil.github.io/HandFormer/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 289, + 452, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 289, + 452, + 465 + ], + "spans": [ + { + "bbox": [ + 160, + 289, + 452, + 465 + ], + "type": "text", + "content": "Abstract. 3D hand pose is an underexplored modality for action recognition. Poses are compact yet informative and can greatly benefit applications with limited compute budgets. However, poses alone offer an incomplete understanding of actions, as they cannot fully capture objects and environments with which humans interact. We propose HandFormer, a novel multimodal transformer, to efficiently model hand-object interactions. HandFormer combines 3D hand poses at a high temporal resolution for fine-grained motion modeling with sparsely sampled RGB frames for encoding scene semantics. Observing the unique characteristics of hand poses, we temporally factorize hand modeling and represent each joint by its short-term trajectories. This factorized pose representation combined with sparse RGB samples is remarkably efficient and highly accurate. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods at " + }, + { + "bbox": [ + 160, + 289, + 452, + 465 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 160, + 289, + 452, + 465 + ], + "type": "text", + "content": " fewer FLOPs. With RGB, we achieve new state-of-the-art performance on Assembly101 and H2O with significant improvements in egocentric action recognition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 475, + 453, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 475, + 453, + 498 + ], + "spans": [ + { + "bbox": [ + 160, + 475, + 453, + 498 + ], + "type": "text", + "content": "Keywords: Skeleton-based action recognition " + }, + { + "bbox": [ + 160, + 475, + 453, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 475, + 453, + 498 + ], + "type": "text", + "content": " 3D hand poses " + }, + { + "bbox": [ + 160, + 475, + 453, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 475, + 453, + 498 + ], + "type": "text", + "content": " Multi-modal transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "The popularity of AR/VR headsets has driven interest in recognizing hand-object interactions, particularly through egocentric [14, 26] and multiview cameras [34, 51]. Such interactions are inherently fine-grained; recognizing them requires distinguishing subtle motions and object state changes. State-of-the-art methods for hand action recognition [23, 45, 47, 66] primarily rely on multi-or single-view RGB streams, which are computationally heavy and unsuitable for resource-constrained scenarios like AR/VR. Motivated by advancements in lightweight hand pose estimation methods leveraging monochrome cameras [27, 28, 46], and the evolution of low-dimensional sensors [36, 42] such as accelerometers, MMG, EMG, demonstrating real-time hand pose estimation, we advocate" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 478, + 203 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 203 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 203 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 203 + ], + "type": "image", + "image_path": "541766225c4522e643c67e6585bbf26257512ede07be096f1ea02cfe6679d827.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 212, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 269 + ], + "type": "text", + "content": "Fig. 1: We densely sample 3D hand poses to understand fine-grained hand motions and sparsely sample RGB frames to capture the scene semantics. Our 3D hand poses are acquired from low-resolution monochrome cameras but can also come from wearable sensors, facilitating an efficient understanding of hand-object interactions. Video frames and hand poses in the figure are from Assembly101 [51]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "type": "text", + "content": "the use of 3D hand poses as an input modality for recognizing hand-object interactions. Hand poses are a compact yet informative representation that captures the motions and nuances of hand movements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 437 + ], + "type": "text", + "content": "Existing works on 3D pose-based action recognition have focused primarily on full-body skeletons [20,44,65,68]. Hand poses differ fundamentally from full-body skeletons. In full-body recognition datasets [41,52], the actions are predominantly static from a global perspective. The relative changes in joint or limb positions signify the action category, e.g., 'sitting down'. Conversely, hand joints typically move together for many actions and lack a static joint as a global reference [51], e.g., 'put down toy'. Full-body skeleton methods also benefit from modeling long-range spatiotemporal dependencies between joints [44], while this is less important for the hands as shown in Fig. 2 and in Sec. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "content": "However, the 3D pose alone cannot encode the hands' actions. Unlike the full-body case, where actions are self-contained by the sequence of poses, the hands are often manipulating objects [34, 46, 51]. Hand pose is excellent for identifying motions (verbs) but struggles with associated objects [51]. Therefore, supplementing pose data with visual context from images or videos is crucial for full semantic understanding. However, as noted earlier, using dense RGB frames contradicts our motivation for using hand poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "This paper introduces HandFormer, a novel and lightweight multimodal transformer that leverages dense 3D hand poses complemented with sparsely sampled RGB frames. To this end, we conceptualize an action as a sequence of short segments, termed micro-actions, which are analogous to words that form sentence-level complex actions. Each micro-action comprises a dense sequence of pose frames and a single RGB frame. As every hand joint moves in close spatial proximity, we encode the pose sequence from a Lagrangian view [49] and track each joint as an individual entity. HandFormer effectively models each micro-action whose features are then temporally aggregated for action classification. This formulation benefits from our observations on the spatiotemporal dynamics of hand movement and delivers strong performance. Additionally, limiting long-range spatiotemporal dependencies among joints and only using sparse RGB frame" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "features improves computational efficiency, making it well-suited for always-on, low-compute AR/VR applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "Our contributions are: (i) We analyze the differences between hand pose and full-body skeleton actions and design a novel pose sequence encoding that reflects hand-specific properties. (ii) We propose HandFormer that takes a sequence of dense 3D hand pose and sparse RGB frames as input in the form of micro-actions. (iii) HandFormer achieves state-of-the-art action recognition on H2O [34] and Assembly101 [51]. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods in Assembly101, incurring at least " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " fewer FLOPs. (iv) We experimentally demonstrate how hand poses are crucial for multiview and egocentric action recognition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 267, + 237, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 237, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 237, + 281 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 449 + ], + "type": "text", + "content": "Video Action recognition. Video-based action recognition systems are well-developed with sophisticated 3D-CNN [9,23,57,67] or Video Transformer [2,3,43,47] architectures. However, they all bear significant computational expense for both feature extraction and motion modeling, either explicitly in the form of optical flow [9,33,55] or implicitly through the architecture [2,47]. Such designs are well-suited for high-facility applications but are not suitable for integration into lightweight systems. To this end, efficient video understanding is an active topic of research, with a focus on reducing expensive 3D operations [22,40,58,67], quadratic complexity of attention [32,43,47] and dropping tokens [4,21]. Additionally, densely-sampled RGB frames incur a high cost, yet limiting the temporal resolution will hinder fine-grained action understanding. We propose complementing 3D hand poses with sparsely-sampled RGB frames, developing a lightweight video understanding system." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 605 + ], + "type": "text", + "content": "Skeleton-based Action Recognition. Skeleton-based action recognition has been tackled with hand-crafted features [60, 62], sequence models like RNNs and LSTMs [19, 69], and CNN-based methods that either employ temporal convolution on the pose sequence [56] or transform the skeleton data into pseudoimages to be processed with 2D or 3D convolutional networks [7, 20, 31]. Recent progress has been driven by GNN-based methods that exploit the graph structure of skeletal data to construct spatio-temporal graphs and perform graph convolutions [68, 70]. These methods often model functional links between joints that go beyond skeletal connectivity [39, 54] or expand the spatiotemporal receptive field [44]. Self-attention and transformer-based methods have also been proposed [48, 65, 71]. However, most existing methods are tailored for datasets involving full-body poses. Few works [24, 30, 37, 50] study hand poses but are restricted to simple gestures, recognizable without complex temporal modeling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Fusing RGB with Skeleton. Skeletal data has been used for cropping images of body parts [12], for weighting RGB patches around regions of interest [5, 6], or for pooling CNN features [1, 8]. Projecting into common embedding space is done in [16], enabling pose distillation [15]. Multi-stream architectures are also designed with separate paths for RGB and skeleton having lateral connec" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 144, + 162, + 187 + ], + "blocks": [ + { + "bbox": [ + 136, + 144, + 162, + 187 + ], + "lines": [ + { + "bbox": [ + 136, + 144, + 162, + 187 + ], + "spans": [ + { + "bbox": [ + 136, + 144, + 162, + 187 + ], + "type": "image", + "image_path": "89eca3f064d23587c0fdd50bc5047e89169b4718ca4e7379d4a58d99aed647b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 188, + 162, + 194 + ], + "lines": [ + { + "bbox": [ + 137, + 188, + 162, + 194 + ], + "spans": [ + { + "bbox": [ + 137, + 188, + 162, + 194 + ], + "type": "text", + "content": "Stand up" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 114, + 199, + 152 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 199, + 152 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 199, + 152 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 199, + 152 + ], + "type": "image", + "image_path": "4854dfb504b50a8617f4ff20d3546cc7b4617e232960666444d5808d07a33c12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 152, + 199, + 160 + ], + "lines": [ + { + "bbox": [ + 170, + 152, + 199, + 160 + ], + "spans": [ + { + "bbox": [ + 170, + 152, + 199, + 160 + ], + "type": "text", + "content": "Put on bag" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 170, + 164, + 198, + 205 + ], + "blocks": [ + { + "bbox": [ + 170, + 164, + 198, + 205 + ], + "lines": [ + { + "bbox": [ + 170, + 164, + 198, + 205 + ], + "spans": [ + { + "bbox": [ + 170, + 164, + 198, + 205 + ], + "type": "image", + "image_path": "1e691409f9f81ee2def47f9d257d486d30f8c779640dcbd82165d358d0376828.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 205, + 197, + 211 + ], + "lines": [ + { + "bbox": [ + 171, + 205, + 197, + 211 + ], + "spans": [ + { + "bbox": [ + 171, + 205, + 197, + 211 + ], + "type": "text", + "content": "Sit down" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "text", + "content": "Fig. 2: Comparing skeletal changes in full-body actions from NTU RGB+D 120 [41] (left) and hand actions from Assembly101 [51] (right). Two pose frames at interval " + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "text", + "content": " are shown. " + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "inline_equation", + "content": "J_{j,t}" + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "text", + "content": " indicates the 3D coordinate of joint " + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 130, + 220, + 482, + 288 + ], + "type": "text", + "content": " is the correlation between two such joints. Modeling the correlation between spatiotemporally distant joints can be informative for full-body poses but does not provide a useful action cue for hand poses." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 216, + 114, + 294, + 197 + ], + "blocks": [ + { + "bbox": [ + 216, + 114, + 294, + 197 + ], + "lines": [ + { + "bbox": [ + 216, + 114, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 216, + 114, + 294, + 197 + ], + "type": "image", + "image_path": "49650b0f05e7738b609bdb167cd0bb3e2c792ded9fa3c5bddb296200744cd313.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 203, + 281, + 213 + ], + "lines": [ + { + "bbox": [ + 220, + 203, + 281, + 213 + ], + "spans": [ + { + "bbox": [ + 220, + 203, + 281, + 213 + ], + "type": "inline_equation", + "content": "\\leftarrow \\rightarrow \\rightarrow < J_{\\mathrm{H,0}}J_{\\mathrm{ET}}>" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 122, + 409, + 190 + ], + "blocks": [ + { + "bbox": [ + 310, + 122, + 409, + 190 + ], + "lines": [ + { + "bbox": [ + 310, + 122, + 409, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 122, + 409, + 190 + ], + "type": "image", + "image_path": "dbe57dbbea64fd2cababef3f172ce5dfee47975988e570815a1e00177c984d97.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 202, + 388, + 213 + ], + "lines": [ + { + "bbox": [ + 326, + 202, + 388, + 213 + ], + "spans": [ + { + "bbox": [ + 326, + 202, + 388, + 213 + ], + "type": "text", + "content": "←→" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 416, + 118, + 447, + 149 + ], + "blocks": [ + { + "bbox": [ + 416, + 118, + 447, + 149 + ], + "lines": [ + { + "bbox": [ + 416, + 118, + 447, + 149 + ], + "spans": [ + { + "bbox": [ + 416, + 118, + 447, + 149 + ], + "type": "image", + "image_path": "8dda702072f573aa6a3d8fbbf3661b7ed2199189f1bec8c2ec040a74b816060b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 149, + 442, + 157 + ], + "lines": [ + { + "bbox": [ + 421, + 149, + 442, + 157 + ], + "spans": [ + { + "bbox": [ + 421, + 149, + 442, + 157 + ], + "type": "text", + "content": "Pick up" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 416, + 175, + 447, + 204 + ], + "blocks": [ + { + "bbox": [ + 416, + 175, + 447, + 204 + ], + "lines": [ + { + "bbox": [ + 416, + 175, + 447, + 204 + ], + "spans": [ + { + "bbox": [ + 416, + 175, + 447, + 204 + ], + "type": "image", + "image_path": "ada46ecffa20204e520e1632ef3f414852d63775abcf0114a4f32976dea33fef.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 204, + 444, + 210 + ], + "lines": [ + { + "bbox": [ + 419, + 204, + 444, + 210 + ], + "spans": [ + { + "bbox": [ + 419, + 204, + 444, + 210 + ], + "type": "text", + "content": "Unscrew" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 449, + 147, + 480, + 177 + ], + "blocks": [ + { + "bbox": [ + 449, + 147, + 480, + 177 + ], + "lines": [ + { + "bbox": [ + 449, + 147, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 449, + 147, + 480, + 177 + ], + "type": "image", + "image_path": "2f2bf0dc21d2efbfeb8adc3fa0d002dd6d5054614851679a0b50a12e9fea0d08.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 180, + 474, + 186 + ], + "lines": [ + { + "bbox": [ + 455, + 180, + 474, + 186 + ], + "spans": [ + { + "bbox": [ + 455, + 180, + 474, + 186 + ], + "type": "text", + "content": "Rotate" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "text", + "content": "tions between them [38, 63]. RGBPoseConv3D [20] is a two-stream model that employs 3D CNNs for both RGB and pose. Similar to skeleton-based action recognition, these multimodal approaches primarily focus on full-body poses. Some approaches simultaneously perform both hand pose estimation and action recognition, using pose data to supervise the training process [13, 64]. However, it is important to note that pose estimation deals with lower-level semantics than action recognition. Actions can often be inferred even when the estimated poses are not accurate [46]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 425, + 388, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 388, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 388, + 438 + ], + "type": "text", + "content": "3 Modeling Full-body vs. Hand Skeletons" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 617 + ], + "type": "text", + "content": "Existing skeleton-based action recognition datasets primarily consist of full-body poses where actions feature significant changes over time in limb positions relative to other body parts, which highly correlate to the action category. These changes can be captured through long-range spatiotemporal modeling using graph convolutions with large receptive fields [44] or self-attention [65]. Conversely, hand actions show diverse movement patterns, as hands can move in arbitrary directions, often without significant changes in articulation. We analyze pose sequences from NTU RGB+D 120 [41] and Assembly101 [51], calculating the distance covered by each joint and identifying the least and the most active joints. We observe that, for full-body poses, these two representative joints show a significant difference in the distances covered, while for hand poses, the difference is subtle. The Pearson correlation coefficient is 0.93 for hand poses (indicating highly coupled joints) and 0.33 for full-body poses. Please refer to Suppl. Sec. A for details regarding these computations and comparisons." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "An example showing the difference between body and hand poses is provided in Fig. 2, where we depict two frames separated by an interval of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " frames. For the full body skeleton, the head joint " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "J_{H}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " remains static, while the right hand joint " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "J_{F}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " moves upward. Conversely, for the hand, all the joints move," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 100 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 119, + 455, + 281 + ], + "blocks": [ + { + "bbox": [ + 140, + 119, + 455, + 281 + ], + "lines": [ + { + "bbox": [ + 140, + 119, + 455, + 281 + ], + "spans": [ + { + "bbox": [ + 140, + 119, + 455, + 281 + ], + "type": "image", + "image_path": "67d57c644999ed5a67f0c5cacd4a59432410b785efeeb0712f4f1315ba45e7c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "text", + "content": "Fig. 3: Overall architecture of HandFormer. An action segment is divided into " + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "text", + "content": " micro-actions " + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "inline_equation", + "content": "\\{M_1, M_2 \\ldots M_K\\}" + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "text", + "content": ". Each micro-action comprises a dense sequence of pose frames and a single RGB frame. The frame encoder " + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "text", + "content": " and the trajectory encoder " + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 292, + 482, + 358 + ], + "type": "text", + "content": " encode the RGB and the dense poses, respectively, after which they are passed to a Multimodal Tokenizer. The modality-mixed tokens are then fed to a Temporal Transformer. Dotted paths are optional and only required when RGB is used." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": "including the wrist joint " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "J_{W}" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " and index fingertip " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "J_{I}" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": ". The spatiotemporal correlation between the head joint in the first frame and the hand joint in the last frame (" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": ") is a crucial action cue that captures the relative structural change. However, such a correlation between the wrist and the fingertip for hand skeleton (" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": ") does not provide a stronger cue than the wrist movement itself. To avoid such redundant spatiotemporally distant correlations, we propose to divide the hand pose sequence into micro-action blocks of fixed temporal length. This formulation allows for encoding short-term movements while enabling parameter sharing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "type": "text", + "content": "Moreover, full-skeletal motion is a dominant feature in hand pose sequences, differentiating them from full-body poses. While the location and orientation of the human body can be trivial for most actions, this is not the case for hands. The hands do not conform to any particular 6D pose, and notably, they frequently and unpredictably alter their 6D poses over time. A comprehensive analysis of this phenomenon is provided in Suppl. Sec. A. To this end, we explicitly consider the global 6D poses of the hands during micro-action-based pose encoding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "type": "text", + "content": "4 HandFormer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Fig. 3 illustrates our proposed HandFormer, which consists of a sequence of micro-action blocks, a novel trajectory encoder, and a temporal aggregation module. The design of our HandFormer allows us to easily incorporate semantic context by sampling a single RGB frame from certain micro-actions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 230, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 230, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 230, + 126 + ], + "type": "text", + "content": "4.1 Micro-actions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": "Given an action segment containing " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": " frames sampled at a certain fps, the input to our model comprises — " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": " ) a dense sequence of 3D hand poses, represented as " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\{P_1,P_2,\\dots ,P_{\\mathcal{T}}\\}" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "P_{t} = \\{P_{t}^{left},P_{t}^{right}\\} \\in \\mathbb{R}^{2\\times J\\times 3}" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": " signifying the 3D coordinates of " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": " keypoints in the left and right hands, respectively, and " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "ii)" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": " a sparse set of RGB frames sampled at intervals " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "\\varDelta f" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{I_1,I_{1 + \\varDelta f},\\ldots ,I_{\\mathcal{T}_r}\\}" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "I_{t}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": " are frame-wise RGB features and " + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_r = 1 + \\left|\\frac{\\mathcal{T} - 1}{\\varDelta f}\\right|\\times \\varDelta f" + }, + { + "bbox": [ + 130, + 135, + 482, + 215 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": "We factorize the raw input into a sequence of " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " micro-action blocks of length " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " frames, obtained by shifting the window across the action segment with a stride of " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " frames. Each block consists of two components — the initial appearance derived from the first RGB frame within the block and the hand motion characterized by the dense sequence of " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " pose frames. To obtain a fixed length input containing " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{T}' = (K - 1)\\times R + N" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " pose frames, we perform linear interpolation for each joint along the temporal axis, transforming pose sequence " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " having " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " frames to " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{S}' = \\{P_1',P_2',\\dots ,P_{\\mathcal{T}'}'\\}" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " having " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " frames, where " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "P_t' \\in \\mathbb{R}^{2\\times J\\times 3}" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": ". Thus, we represent the input as a sequence of micro-actions " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{M_1,M_2,\\dots ,M_K\\}" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{S}'" + }, + { + "bbox": [ + 130, + 215, + 482, + 334 + ], + "type": "text", + "content": " through the following equations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 342, + 481, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 342, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 198, + 342, + 481, + 361 + ], + "type": "interline_equation", + "content": "M _ {k} = \\left[ M _ {k} ^ {\\mathrm {R G B}}, M _ {k} ^ {\\text {P o s e}} \\right] = \\left[ I _ {h (k)}, \\left\\{P _ {g (k) + i} ^ {\\prime} \\right\\} _ {i = 0} ^ {N - 1} \\right], \\tag {1}", + "image_path": "ce759bc2a47eac676cf72bf59aafd659eb0b09d00c78d5b34ee4afb4a91892c9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "inline_equation", + "content": "g(k) = (k - 1)\\times R + 1" + }, + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "text", + "content": " denotes the first pose frame in " + }, + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "text", + "content": "-th Micro-action, and " + }, + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "inline_equation", + "content": "h(k)" + }, + { + "bbox": [ + 130, + 369, + 481, + 393 + ], + "type": "text", + "content": " determines the index of the nearest available RGB frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "text", + "content": "The dense pose sequence in a micro-action captures fine-grained hand motion crucial for recognizing verbs, whereas a single RGB frame provides semantic context for recognizing objects. To extract features from micro-actions, we use a frame encoder " + }, + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "text", + "content": " and a trajectory encoder " + }, + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "text", + "content": ", which operate on RGB frames and pose sequences, respectively. Consequently, the RGB and pose features for the " + }, + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 394, + 482, + 465 + ], + "type": "text", + "content": " micro-action is given by-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 473, + 481, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 473, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 216, + 473, + 481, + 488 + ], + "type": "interline_equation", + "content": "\\left[ f _ {k} ^ {\\mathrm {R G B}}, f _ {k} ^ {\\mathrm {P o s e}} \\right] = \\left[ F \\left(M _ {k} ^ {\\mathrm {R G B}}\\right), T \\left(M _ {k} ^ {\\mathrm {P o s e}}\\right) \\right], \\tag {2}", + "image_path": "2ac5cada1577e995aceeb4420148f91012e64e6b86390a2b0160bc933991afb6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 496, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 481, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 496, + 481, + 521 + ], + "type": "inline_equation", + "content": "f_{k}^{\\mathrm{RGB}}, f_{k}^{\\mathrm{pos}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 496, + 481, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 496, + 481, + 521 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 496, + 481, + 521 + ], + "type": "text", + "content": " denotes the common dimensionality of both RGB and pose embedding space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 538, + 259, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 259, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 259, + 551 + ], + "type": "text", + "content": "4.2 Trajectory Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We devise a trajectory-based pose encoder to encode dense hand pose sequences within micro-actions, illustrated in Fig. 4. Each joint is represented by its trajectory of dimension " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 \\times N" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", encapsulating the sequence of 3D coordinates across the " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " pose frames of a micro-action. This yields " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "2 \\times J" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " feature vectors for the " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " joints of two hands. Each joint's trajectory is passed through a TCN [35], whose parameters are shared for all joints. This produces " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "2 \\times J" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " Local Trajectory Tokens. Additionally, the full-skeletal motion of the hand during the entire action is used as a reference through an additional token named Global Wrist Token. This token is generated by a separate TCN operating on the sequence of 6D" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 119, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 119, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 119, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 119, + 482, + 232 + ], + "type": "image", + "image_path": "c7827b62edb38da93d51875c2a450e292076fc150ac7301369a31088c7b59593.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 308 + ], + "type": "text", + "content": "Fig. 4: Our Trajectory Encoder " + }, + { + "bbox": [ + 130, + 243, + 480, + 308 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 243, + 480, + 308 + ], + "type": "text", + "content": ", which operates on micro-actions, derives tokens with trajectory-based features and performs self-attention to encode the pose sequence into a feature vector. The Single-Joint TCN is a Temporal Convolutional Network [35] that processes the trajectories of all the joints individually with shared parameters across all joints. Wrist-TCN takes an action-wide sequence of wrist location and hand orientation (6D pose) to produce a global reference token." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 333, + 479, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 381 + ], + "type": "text", + "content": "poses of hands, indicating wrist location and hand orientation. Subsequently, a self-attention layer is applied to these trajectory tokens, preserving the temporal dimension for subsequent stages. This iterative process culminates in spatiotemporal average pooling, summarizing the hand motion of the micro-action." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 400, + 272, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 272, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 272, + 411 + ], + "type": "text", + "content": "4.3 Multimodal Tokenizer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "type": "text", + "content": "This section discusses incorporating sparsely sampled RGB frames into Hand-Former to better capture scene semantics. A single frame is sampled from each micro-action, generating an extended crop " + }, + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "type": "inline_equation", + "content": "(1.25\\times)" + }, + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "type": "text", + "content": " around the hands. While the full image provides an overall scene context, the crop focuses specifically on hand-object interaction regions [10]. Features for both are separately generated using a pre-existing image encoder and then aggregated to enrich the hand-object interaction feature with scene context. The hand-object ROI crop can be obtained using an off-the-shelf HOI detector [53] or by using the corresponding hand pose projections. We opt for the latter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 480, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 480, + 590 + ], + "type": "text", + "content": "Our multimodal tokenizer in Fig. 3 generates RGB and Pose tokens enhanced via multimodal interactions. This involves concatenating each frame feature and trajectory encoding, and projecting them into a shared PoseRGB feature space using an MLP. This PoseRGB feature is then split into two parts, which are added to the original frame feature and trajectory encoding, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 608, + 274, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 274, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 274, + 620 + ], + "type": "text", + "content": "4.4 Temporal Transformer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "The multimodal tokenizer provides RGB and pose tokens for each micro-action. Since an action segment consists of a sequence of micro-actions, these micro-action tokens are aggregated over time via a temporal transformer. A video " + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "content": "divided into " + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "content": " micro-actions can be represented by two sets of tokens " + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_k^{\\mathrm{RGB}}\\}_{k=1}^K" + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_k^{\\mathrm{Pose}}\\}_{k=1}^K" + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "content": ", respectively, produced by the multimodal tokenizer. These " + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "inline_equation", + "content": "2 \\times K" + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "content": " tokens of dimension " + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "content": " form the input sequence of the temporal transformer. Positional encoding and modality embedding are added to each input token to indicate the temporal location and source modality. We use the fixed sine/cosine positional encoding [59], assigning the same position to two tokens from the same micro-action. Modality embeddings are learned and shared across the tokens from the same modality (RGB or Pose). Following standard practice [17,18], we preprocess an additional learnable class token " + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLS}] \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 114, + 482, + 237 + ], + "type": "text", + "content": ", the output of which is then used for classifying actions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 254, + 263, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 263, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 263, + 266 + ], + "type": "text", + "content": "4.5 Learning Objectives" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "content": "HandFormer is trained end-to-end for action recognition, supervised by a cross-entropy loss: " + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls} = -a_i\\log \\hat{a}_i" + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "content": " represents the ground truth action label for the " + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "content": " sample, and " + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\hat{a}_i" + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "content": " denotes the predicted action category. The input pose and RGB modalities provide complementary information regarding a scene, capturing motion and interacting objects, respectively. To effectively utilize this, we employ explicit verb and object supervision (" + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{verb}" + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 130, + 273, + 482, + 380 + ], + "type": "text", + "content": ") via two additional learnable class tokens. Given that pose strongly correlates with verbs, whereas objects are identifiable from RGB frames alone, we let the verb class token attend exclusively to pose encodings and the object class token to frame features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 381, + 483, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 483, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 483, + 489 + ], + "type": "text", + "content": "Feature Anticipation Loss. Hand pose sequence captures the primary sources of state changes during hand-object interactions. We posit that the visual state from an initial RGB frame, in combination with the subsequent hand pose sequence, is indicative of the visual state that results from the completion of the sequence. Therefore, given an RGB feature and the corresponding pose features from a micro-action, we force our model to anticipate the RGB feature for the next micro-action by minimizing an " + }, + { + "bbox": [ + 130, + 381, + 483, + 489 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 381, + 483, + 489 + ], + "type": "text", + "content": " feature loss. This loss, inspired by existing efforts [25,61], quantifies the difference between the anticipated image feature and the true feature extracted from a frozen image encoder. Formally," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 222, + 507, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 507, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 222, + 507, + 481, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a n t} = \\sum_ {k = 1} ^ {K - 1} \\left\\| \\Phi_ {\\text {a n t}} \\left(f _ {k} ^ {\\text {P o s e R G B}}\\right) - f _ {k + 1} ^ {\\text {R G B}} \\right\\| _ {1} \\tag {3}", + "image_path": "54d5f671c03e10668c2f093883c68678524f4aa1bf57771886f26b766ee63071.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 546, + 435, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 435, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 435, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 546, + 435, + 559 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{ant}}" + }, + { + "bbox": [ + 131, + 546, + 435, + 559 + ], + "type": "text", + "content": " denotes a linear projection layer. Hence, the total loss is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 224, + 571, + 481, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 571, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 224, + 571, + 481, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\lambda_ {1} \\mathcal {L} _ {v e r b} + \\lambda_ {2} \\mathcal {L} _ {o b j} + \\lambda_ {3} \\mathcal {L} _ {a n t} \\tag {4}", + "image_path": "9ab067ca0eef79e255bfc417d7c7b90bcc8ad7e1b86c8468e5ac9c65dcde7ed4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 587, + 435, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 435, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 435, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 587, + 435, + 600 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 131, + 587, + 435, + 600 + ], + "type": "text", + "content": " are hyperparameters used to balance the four losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 616, + 230, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 230, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 230, + 631 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two publicly available hand-object interaction datasets. Assembly101 [51] is a large-scale multiview dataset that fea-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "tures videos of procedural activities for assembling and disassembling 101 take-aside toy vehicles. This dataset has 1380 fine-grained actions based on 24 verbs and 90 objects. It consists of 12 temporally synchronized views — eight static and four egocentric. The egocentric views exhibit low-resolution monochrome images, making it challenging even for human eyes to discern objects in view. Hence, we opt for fixed views as our RGB modality. 3D hand poses estimated using off-the-shelf MEgATrack [27] are provided by the dataset. H2O [34] features four participants performing 36 actions that involve 8 objects and 11 verbs. Ground truth 3D hand poses are provided. The dataset consists of 5 temporally synchronized RGB views — 4 fixed and 1 egocentric. We use RGB frames from the egocentric view only." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": "Implementation Details. Similar to [65], our pose input consists of 120 frames by setting " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{T}' = 120" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": ". The number of frames per micro-action (" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": ") is 15. For multimodal experiments, we set the window stride " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "R = N" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": " and take " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": " non-overlapping micro-actions, which allows us to sample 8 RGB frames for each video following [51]. However, we allow a " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": " overlap between consecutive micro-actions for pose-only variants. A frozen ViT [18] or ResNet [29], followed by a learnable linear layer, is used as the frame encoder " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": ". More details are provided in Suppl. Sec. C. Each model is trained for 50 epochs using SGD with momentum 0.9, a batch size of 32, a learning rate of 0.025, and step LR decay with a factor of 0.1 at epochs {25, 40}. Loss hyperparameters " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\{\\lambda_1, \\lambda_2, \\lambda_3\\}" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": " are chosen to be {1.0, 1.0, 2.0}." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": "Model Variants. We propose several variants of our model by adjusting the width " + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": " and the number of layers " + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": " of the transformer to balance efficiency and accuracy. Our default HandFormer, denoted HandFormer-B, has parameters " + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "inline_equation", + "content": "(d,T_n) = (256,2)" + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": ". We introduce a larger variant, HandFormer-L, with " + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "inline_equation", + "content": "(d,T_{n}) = (512,4)" + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": ". We also explore different configurations for the number of input joints " + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": " per hand. Unless otherwise mentioned, we utilize all 21 joints per hand along with the base model denoted as HandFormer-B/21 while offering a highly efficient option utilizing only six joints per hand (five fingertips and the wrist) termed HandFormer-B/6. Additionally, HandFormer-X/J×N denotes a multimodal variant that utilizes " + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": " RGB frames, where " + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{N} \\leq K" + }, + { + "bbox": [ + 130, + 380, + 483, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 515, + 332, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 332, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 332, + 528 + ], + "type": "text", + "content": "5.1 Comparison with State-of-the-Art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed architecture, we compare it against several baselines. For pose-only comparisons, we choose a graph-based network MS-G3D [44] and an attention-based method ISTA-Net [65] — the two best performing skeleton-based methods on Assembly101 [51], as reported by [65]. We also employ video baselines TSM [40] and SlowFast [23], emphasizing efficiency and high temporal resolution, respectively. Additionally, we include H2OTR [13], the state-of-the-art for the H2O dataset. For Assembly101, we replicate the results of the methods above using RGB frames from view 4 of the dataset, while results on H2O are acquired from the respective papers. Furthermore, on both datasets, we train and test RGBPoseConv3D [20], which is current state-of-the-art in multimodal action recognition with skeleton and RGB data." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 190, + 462, + 315 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with state-of-the-art methods on Assembly101 [51] and H2O [34]. For H2O, 6D object pose is used by ISTA-Net during training and inference and by H2OTR during training. 'MS-G3D + TSM' denotes a late fusion of the corresponding unimodal architectures. Reported GFLOPs include pose estimation costs in the context of Assembly101 at 15 fps. †For pose-only methods, pose estimation costs are common and excluded for simpler comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 190, + 462, + 315 + ], + "lines": [ + { + "bbox": [ + 149, + 190, + 462, + 315 + ], + "spans": [ + { + "bbox": [ + 149, + 190, + 462, + 315 + ], + "type": "table", + "html": "
MethodPoseRGBGFLOPsAssembly101H2O Action
ActionVerbObject
MS-G3D [44]✓✗\\( 21.2^† \\)28.7863.4637.2650.83
ISTA-Net [65]✓✗\\( 35.2^† \\)28.1462.7036.7789.09
SlowFast [23]✗✓65.7---77.69
TSM [40]✗✓33.035.2758.2747.45-
H2OTR [13]✗✓----90.90
RGBPoseConv3D [20]✓✓68.933.6161.9942.9083.47
MS-G3D + TSM✓✓66.239.7465.1251.12-
HandFormer-B/21×8✓✗\\( 4.2^† \\)28.8065.3336.2857.44
✗✓33.032.0755.6144.8984.71
✓✓47.641.0669.2351.1793.39
", + "image_path": "8a48c7341c5311fccdaa4982f8e4c0696802cc01f929ce3b21735106d367ab79.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 334, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 479, + 380 + ], + "type": "text", + "content": "We evaluate three variants of our method by controlling the input modalities. As shown in Tab. 1, our unimodal pose-only model excels in verb recognition on Assembly101. In contrast, RGB-based methods benefit from object appearance and usually perform well for actions due to the strong object recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 382, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 477 + ], + "type": "text", + "content": "In this context, the accuracy of ISTA-Net on H2O is not directly comparable to our method, as they also use the 6D object poses, while we only use hand poses as input. RGBPoseConv3D struggles to achieve satisfactory performance, particularly on Assembly101, suggesting that generalizing to hand poses is non-trivial for skeleton-based methods. Therefore, we combine two best-performing unimodal methods with late fusion (MS-G3D + TSM) to create a stronger baseline. Our model even outperforms this, indicating the effectiveness of our proposed multimodal fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 492, + 389, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 389, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 389, + 504 + ], + "type": "text", + "content": "5.2 Skeleton-based Action Recognition for Hands" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "The compositional nature of action classes in hand-object interaction videos allows us to break down the action into a verb and an object. Recognizing such actions from 3D hand poses is an ill-posed problem, as the hand skeletons lack explicit information about the interacting objects, which are also part of the action semantics. However, as the pose data completely captures the hand motion information, it can be reliably used for verb recognition. Therefore, we evaluate a pose-only version of our method on the verb recognition task and compare the performance and efficiency metrics with other state-of-the-art skeleton-based methods in Fig. 5. Our method uses significantly fewer GFLOPs due to our spatiotemporal factorization using micro-actions. With " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "J = 21" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", all our HandFormer variants outperform existing methods. The " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "J = 6" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " variant is exceptionally efficient with comparable accuracy, notably surpassing the efficiency-focused Shift-GCN [11]. HandFormer-H/21 combines the HandFormer-B and HandFormer-L" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 291, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 291, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 291, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 291, + 232 + ], + "type": "image", + "image_path": "c8bbdbfd31fc2fbb0900f212d79be075e43c4f514dd24c16236e70b7b51bfb0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 233, + 294, + 299 + ], + "lines": [ + { + "bbox": [ + 132, + 233, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 233, + 294, + 299 + ], + "type": "text", + "content": "Fig.5: Comparison of skeleton-based methods for verb recognition on Assembly101 [51]. Our method achieves state-of-the-art performance while utilizing significantly fewer FLOPs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 328, + 119, + 479, + 232 + ], + "blocks": [ + { + "bbox": [ + 328, + 119, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 328, + 119, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 328, + 119, + 479, + 232 + ], + "type": "image", + "image_path": "4d598a8731ddaaa4af32e72798ed58dcd48f61e2a09d42f0c5ca07b84907520a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 233, + 481, + 300 + ], + "lines": [ + { + "bbox": [ + 325, + 233, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 325, + 233, + 481, + 300 + ], + "type": "text", + "content": "Fig.6: Ablating number of RGB frames in HandFormer on Assembly101 [51]. With more RGB frames, verb recognition shows marginal gain, whereas object recognition shows improvement with diminishing returns." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 331, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 357 + ], + "type": "text", + "content": "variants and becomes our best-performing model, improving over MS-G3D by " + }, + { + "bbox": [ + 132, + 331, + 480, + 357 + ], + "type": "inline_equation", + "content": "2.04\\%" + }, + { + "bbox": [ + 132, + 331, + 480, + 357 + ], + "type": "text", + "content": " while maintaining comparable FLOPs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 372, + 350, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 350, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 350, + 384 + ], + "type": "text", + "content": "5.3 How many RGB frames are required?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 391, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 571 + ], + "type": "text", + "content": "In our model, the pose modality maintains a high temporal resolution to capture fine-grained hand movements, resulting in good verb recognition performance. On the contrary, RGB frames are primarily required to introduce semantic context for object recognition and does not necessitate a high temporal resolution like hand movements. The design of our model allows us to sample only a few RGB frames (as low as one) but still perform competitively at a reduced computational cost. Fig. 6 shows the impact of using more RGB frames for Assembly101 [51]. For this experiment, non-overlapping micro-actions are considered. Using only one RGB frame in HandFormer (35.46) outperforms the video model TSM (35.27), as shown in Tab. 1. This performance gain stems primarily from a significant improvement in object accuracy, with only a slight enhancement in verb accuracy. However, including more RGB frames shows a diminishing return, which is unsurprising as additional frames are expected to provide redundant information. These results are obtained with a simplified version of our model by setting " + }, + { + "bbox": [ + 132, + 391, + 481, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{cls}" + }, + { + "bbox": [ + 132, + 391, + 481, + 571 + ], + "type": "text", + "content": " and bypassing the multimodal tokenizer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 586, + 460, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 460, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 460, + 598 + ], + "type": "text", + "content": "5.4 Can 3D hand pose be an efficient alternative to multi-view?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 481, + 666 + ], + "type": "text", + "content": "While multiview action recognition benefits from precise hand-movement information in 3D space, processing all views with video models is expensive and highly redundant. In Tab. 2, we demonstrate that combining 3D hand pose with a single RGB view (view 4) achieves comparable performance to multiview action recognition on Assembly101 [51]. Specifically, our action recognition" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 201, + 289, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 289, + 190 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 289, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 289, + 190 + ], + "type": "text", + "content": "Table 2: Multi-view Action Recognition on Assembly101 [51]. HandFormer with 'Single View+Pose' has better performance than 'Single View+Egocentric'. Also, verb performance is comparable to using all (8) RGB views." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 201, + 289, + 266 + ], + "lines": [ + { + "bbox": [ + 134, + 201, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 134, + 201, + 289, + 266 + ], + "type": "table", + "html": "
ViewsActionVerbObject
Single View35.2758.2747.45
Single View + Egocentric37.7561.8049.43
Two Views41.9665.2253.26
All (8) Views47.5170.9957.73
Single View + Pose41.0669.2351.17
", + "image_path": "59bbcbb4382232b1c4adf30201353c2b06c68406d44bb0faacb55ca77cd5ef3b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 298, + 201, + 477, + 265 + ], + "blocks": [ + { + "bbox": [ + 295, + 114, + 479, + 191 + ], + "lines": [ + { + "bbox": [ + 295, + 114, + 479, + 191 + ], + "spans": [ + { + "bbox": [ + 295, + 114, + 479, + 191 + ], + "type": "text", + "content": "Table 3: Cross-view performance of HandFormer in Assembly101 [51] shows its generalization capability to unseen view 1, outperforming the video baseline which is trained on view 1 directly. Egocentric views are the source of hand poses here and, therefore, are included in the video models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 201, + 477, + 265 + ], + "lines": [ + { + "bbox": [ + 298, + 201, + 477, + 265 + ], + "spans": [ + { + "bbox": [ + 298, + 201, + 477, + 265 + ], + "type": "table", + "html": "
MethodTrain onTest onActionVerbObject
TSM [40]v4 + egov4 + ego37.7561.8049.43
v4 + egov1 + ego35.2759.5347.72
v1 + egov1 + ego36.2160.7848.52
Our Methodv4 + Posev4 + Pose41.0669.2351.17
v4 + Posev1 + Pose38.4367.8648.32
", + "image_path": "2a7d2cc06d7600bde6f49ab4c6f3e6a13797bdd0e2d1c68b391fae46a6847e40.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 289, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 422 + ], + "type": "text", + "content": "performance ('Single View + Pose') matches the fusion of the two most informative views — view 1 and view 4. Notably, our verb recognition accuracy is on par with combining all 8 RGB views, though the latter incurs at least " + }, + { + "bbox": [ + 130, + 289, + 481, + 422 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 130, + 289, + 481, + 422 + ], + "type": "text", + "content": " more FLOPs despite using the efficient TSM model. Additionally, using hand pose in combination with an RGB view ('Single View + Pose') outperforms directly using the egocentric videos ('Single View + Egocentric') from which the hand poses are derived. While fusing multiple RGB views improves accuracy by ensembling multiple complementary predictions, the computational overhead increases significantly. In contrast, our model processes hand pose and single-view RGB frames, enhancing efficiency by leveraging the less redundant and low-dimensional pose data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "type": "text", + "content": "Cross-view Generalization. 3D hand pose offers a unique opportunity for cross-view generalization because of its universality across different viewpoints. To evaluate the effectiveness of our method for unseen views, we train our model with frame-wise RGB features from view 4 and test it on view 1. We train TSM [40] as a baseline video model on both RGB views separately. As our method includes 3D hand poses obtained from the egocentric views, we include the egocentric videos in the TSM baselines for a fair comparison. As seen from Tab. 3, our method, trained on view 4, generalizes well on unseen view 1, outperforming the TSM model that was directly trained on view 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 560, + 315, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 315, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 315, + 572 + ], + "type": "text", + "content": "5.5 Egocentric Action Recognition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "Recognizing actions in egocentric videos is challenging due to camera motion and the occlusion of interacting objects by the hands. Additionally, in the case of Assembly101 [51], the egocentric cameras are similar to Oculus Quest VR headsets, which provide monochrome low-resolution frames. As a result, action recognition performance is significantly lower than fixed RGB views, as found in [51]. We address this challenging scenario with our proposed multimodal architecture and achieve state-of-the-art performance in egocentric action recognition on Assem" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 158, + 323, + 218 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 321, + 156 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 321, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 321, + 156 + ], + "type": "text", + "content": "Table 4: Egocentric action recognition in Assembly101 [51]. TSM features from [51] are used as RGB frame features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 158, + 323, + 218 + ], + "lines": [ + { + "bbox": [ + 133, + 158, + 323, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 323, + 218 + ], + "type": "table", + "html": "
MethodActionVerbObject
TSM egocentric (fuse 4 views)33.8059.0046.50
Egocentric (e3) + Pose36.0765.5245.82
Egocentric (e4) + Pose35.5665.7945.20
Egocentric (e3+e4) + Pose38.0566.3247.86
", + "image_path": "c1f4a6e4591540a1a8810d622c78597218bbc0a27d9582a14c1618167499817a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 262, + 315, + 299 + ], + "blocks": [ + { + "bbox": [ + 132, + 240, + 313, + 262 + ], + "lines": [ + { + "bbox": [ + 132, + 240, + 313, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 313, + 262 + ], + "type": "text", + "content": "Table 6: Micro-action length ablation for verb recognition in Assembly101 [51]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 262, + 315, + 299 + ], + "lines": [ + { + "bbox": [ + 133, + 262, + 315, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 262, + 315, + 299 + ], + "type": "table", + "html": "
#Frames1153060120
Verb \nAccuracy(%)59.1263.7063.6863.5162.29
", + "image_path": "70fb96d834dbba458c9ce6eb107f592c9cced3fe768c8949da0df7295961fde0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 342, + 158, + 472, + 219 + ], + "blocks": [ + { + "bbox": [ + 335, + 125, + 474, + 157 + ], + "lines": [ + { + "bbox": [ + 335, + 125, + 474, + 157 + ], + "spans": [ + { + "bbox": [ + 335, + 125, + 474, + 157 + ], + "type": "text", + "content": "Table 5: Keypoint ablation for verb recognition in Assembly101 [51]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 158, + 472, + 219 + ], + "lines": [ + { + "bbox": [ + 342, + 158, + 472, + 219 + ], + "spans": [ + { + "bbox": [ + 342, + 158, + 472, + 219 + ], + "type": "table", + "html": "
#JointsGlobal ReferenceVerb Accuracy (%)
21×64.17
21✓64.90
11✓64.77
6✓63.70
", + "image_path": "bf5a4a00afb7ec58f73eb6030356b4850c9c1a8ee2b7a04bb7536ceb273af54b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 332, + 274, + 474, + 297 + ], + "blocks": [ + { + "bbox": [ + 327, + 240, + 477, + 273 + ], + "lines": [ + { + "bbox": [ + 327, + 240, + 477, + 273 + ], + "spans": [ + { + "bbox": [ + 327, + 240, + 477, + 273 + ], + "type": "text", + "content": "Table 7: Ablating unimodal temporal aggregation for verb recognition in Assembly101 [51]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 274, + 474, + 297 + ], + "lines": [ + { + "bbox": [ + 332, + 274, + 474, + 297 + ], + "spans": [ + { + "bbox": [ + 332, + 274, + 474, + 297 + ], + "type": "table", + "html": "
Temp. Agg.TCNLSTMTransformer
Verb Acc. (%)62.9563.3463.70
", + "image_path": "6db324321f4632412325c4208c877f4f2dc6c22d0f0b528814db0fbcc7c731b4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "text", + "content": "bly101. For this experiment, we used frame-wise TSM features provided by [51]. As depicted in Tab. 4, our model, using a single egocentric view (" + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "inline_equation", + "content": "e3" + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "inline_equation", + "content": "e4" + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "text", + "content": ") outperforms the fusion of four egocentric views as reported in [51]. Moreover, fusing " + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "inline_equation", + "content": "e3" + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "inline_equation", + "content": "e4" + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "text", + "content": " significantly enhances our model, resulting in a " + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "inline_equation", + "content": "4.25\\%" + }, + { + "bbox": [ + 130, + 322, + 480, + 381 + ], + "type": "text", + "content": " increase in action accuracy over the baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 246, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 246, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 246, + 414 + ], + "type": "text", + "content": "5.6 Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 425, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 545 + ], + "type": "text", + "content": "Keypoints. Not all joints of the hands are equally informative for understanding hand actions. For instance, fingertips exhibit greater mobility compared to the inner joints. Moreover, from an egocentric viewpoint, certain joints are more prone to self-occlusion than others. In Tab. 5, we present the impact of incorporating varying numbers of joints on verb recognition within the Assembly101 dataset [51]. For the case of 6 joints, we consider only the wrist joint along with the five fingertips. To expand to 11 joints, we incorporate all the joints along the index and the thumb, which are least affected by self-occlusion. We also show the effect of including the Global Wrist Token, which acts as a reference to the global motion of the hands while encoding micro-actions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 557, + 482, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 653 + ], + "type": "text", + "content": "Micro-action length. As the resized input comprises a fixed number of pose frames, enlarging the window size for a micro-action decreases the number of micro-actions to aggregate, and vice-versa. In Tab. 6, we vary the micro-action length for verb recognition in the Assembly101 [51] using 6 joints per hand as input, i.e. fingertips and wrist joint. The input pose clip is temporally resized to " + }, + { + "bbox": [ + 130, + 557, + 482, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{T}' = 120" + }, + { + "bbox": [ + 130, + 557, + 482, + 653 + ], + "type": "text", + "content": " before breaking into micro-actions. Lengths 1 and 120 represent two extreme versions with frame-based and trajectory-based encoding, respectively, while the others conform to our micro-action-based formulation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 135, + 465, + 217 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 456, + 126 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 456, + 126 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 456, + 126 + ], + "type": "text", + "content": "Table 8: Ablating tokenization and different losses for action recognition" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 135, + 465, + 217 + ], + "lines": [ + { + "bbox": [ + 150, + 135, + 465, + 217 + ], + "spans": [ + { + "bbox": [ + 150, + 135, + 465, + 217 + ], + "type": "table", + "html": "
Multimodal TokenizerFeature Ant. LossVerb & Object LossAction Accuracy (%)
Assembly101 [51]H2O [34]
XXX38.9885.95
✓XX40.1988.84
✓✓X40.2489.26
XX✓40.5690.50
✓✓✓41.0693.39
", + "image_path": "c7164a2bf5f1cee3b1926828848281464edc6ff09237eb591175d6430c5ba561.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 481, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 302 + ], + "type": "text", + "content": "Temporal Aggregation. After extracting micro-action features, aggregation for action recognition can be done with any sequence model. In Tab. 7, we evaluate the effectiveness of different temporal aggregation methods for verb recognition in the Assembly101 dataset [51]. Here, we use 6 joints per hand as input, i.e., the fingertips and the wrist joint." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 481, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 388 + ], + "type": "text", + "content": "Loss components. To assess the individual contributions of different components, we begin with a basic configuration. We then systematically introduce each element to understand its impact on the overall performance as observed in Tab. 8. Incorporating modality interaction between RGB and pose at the micro-action level through a multimodal tokenizer enhances action accuracy. The introduction of auxiliary losses also has a positive impact, resulting in an overall improvement of " + }, + { + "bbox": [ + 130, + 304, + 481, + 388 + ], + "type": "inline_equation", + "content": "2.08\\%" + }, + { + "bbox": [ + 130, + 304, + 481, + 388 + ], + "type": "text", + "content": " for Assembly101 [51] and " + }, + { + "bbox": [ + 130, + 304, + 481, + 388 + ], + "type": "inline_equation", + "content": "7.44\\%" + }, + { + "bbox": [ + 130, + 304, + 481, + 388 + ], + "type": "text", + "content": " for H2O [34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 415, + 219, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 219, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 219, + 427 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 448, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 481, + 555 + ], + "type": "text", + "content": "With the growing interest in AR/VR and wearables, hand pose estimation has rapidly advanced, and holds promise as a compact and domain-independent modality to complement the visual input. To address the underexplored domain of using 3D hand pose as a modality for hand-object interaction recognition, we introduce HandFormer, a novel multimodal transformer that leverages dense sequences of 3D hand poses with sparsely sampled RGB frames to achieve state-of-the-art action recognition performance. Our model also reduces computational requirements, offering immediate significance across various low-resource applications in mobile devices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": "Limitations. Our method relies on the availability of hand poses, which, if extracted from the visual modality with pose estimation tools [27, 28], can encounter out-of-view scenarios and produce noisy poses. Our experiments reveal that these estimated poses can still achieve good accuracy, yet further research can be conducted to explicitly address this phenomenon. We also assume that a uniform sampling of RGB frames from each micro-action should provide good representations for understanding the semantic context. However, not all frames are equally important in understanding the action. In such a case, adaptive frame sampling methods can be employed, which we leave for future work." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "content": "This research is supported by A\\*STAR under its National Robotics Programme (NRP) (Award M23NBK0053)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "type": "text", + "content": "1. Ahn, D., Kim, S., Hong, H., Ko, B.C.: Star-transformer: a spatio-temporal cross attention transformer for human action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3330-3339 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "text", + "content": "2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 481, + 311 + ], + "type": "text", + "content": "3. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "type": "text", + "content": "4. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 335, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 481, + 365 + ], + "type": "text", + "content": "5. Bruce, X., Liu, Y., Chan, K.C.: Multimodal fusion via teacher-student network for indoor action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3199-3207 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 367, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 481, + 399 + ], + "type": "text", + "content": "6. Bruce, X., Liu, Y., Zhang, X., Zhong, S.h., Chan, K.C.: Mmnet: A model-based multimodal network for human action recognition in rgb-d videos. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3522-3538 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 481, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 444 + ], + "type": "text", + "content": "7. Caetano, C., Sena, J., Brémond, F., Dos Santos, J.A., Schwartz, W.R.: Skelemotion: A new representation of skeleton joint sequences based on motion information for 3d action recognition. In: 2019 16th IEEE international conference on advanced video and signal based surveillance (AVSS). pp. 1-8. IEEE (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 445, + 481, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 476 + ], + "type": "text", + "content": "8. Cao, C., Zhang, Y., Zhang, C., Lu, H.: Body joint guided 3-d deep convolutional descriptors for action recognition. IEEE transactions on cybernetics 48(3), 1095-1108 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "type": "text", + "content": "9. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 543 + ], + "type": "text", + "content": "0. Chatterjee, D., Sener, F., Ma, S., Yao, A.: Opening the vocabulary of egocentric actions. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 544, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 481, + 587 + ], + "type": "text", + "content": "11. Cheng, K., Zhang, Y., He, X., Chen, W., Cheng, J., Lu, H.: Skeleton-based action recognition with shift graph convolutional network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 183-192 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 589, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 481, + 620 + ], + "type": "text", + "content": "12. Chéron, G., Laptev, I., Schmid, C.: P-cnn: Pose-based cnn features for action recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 3218-3226 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "13. Cho, H., Kim, C., Kim, J., Lee, S., Ismayilzada, E., Baek, S.: Transformer-based unified recognition of two hands manipulating objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4769-4778 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 101 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "14. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European conference on computer vision (ECCV). pp. 720-736 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "text", + "content": "15. Das, S., Dai, R., Yang, D., Bremond, F.: Vpn++: Rethinking video-pose embeddings for understanding activities of daily living. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9703-9717 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "16. Das, S., Sharma, S., Dai, R., Bremond, F., Thonnat, M.: Vpn: Learning video-posed embedding for activities of daily living. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16. pp. 72–90. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "17. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 271, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 481, + 314 + ], + "type": "text", + "content": "18. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 347 + ], + "type": "text", + "content": "19. Du, Y., Wang, W., Wang, L.: Hierarchical recurrent neural network for skeleton based action recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1110-1118 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "20. Duan, H., Zhao, Y., Chen, K., Lin, D., Dai, B.: Revisiting skeleton-based action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2969-2978 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "text", + "content": "21. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "22. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "23. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "24. Gan, M., Liu, J., He, Y., Chen, A., Ma, Q.: Keyframe selection via deep reinforcement learning for skeleton-based gesture recognition. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 523, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 545 + ], + "type": "text", + "content": "25. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "26. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "27. Han, S., Liu, B., Cabezas, R., Twigg, C.D., Zhang, P., Petkau, J., Yu, T.H., Tai, C.J., Akbay, M., Wang, Z., et al.: Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG) 39(4), 87-1 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "28. Han, S., Wu, P.c., Zhang, Y., Liu, B., Zhang, L., Wang, Z., Si, W., Zhang, P., Cai, Y., Hodan, T., et al.: Umetrack: Unified multi-view end-to-end hand tracking for vr. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "30. Hou, J., Wang, G., Chen, X., Xue, J.H., Zhu, R., Yang, H.: Spatial-temporal attention res-tcn for skeleton-based dynamic hand gesture recognition. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "31. Hou, Y., Li, Z., Wang, P., Li, W.: Skeleton optical spectra-based action recognition using convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology 28(3), 807-811 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "32. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "type": "text", + "content": "33. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "34. Kwon, T., Tekin, B., Stühmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manipulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10138-10148 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "35. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "type": "text", + "content": "36. Lee, C.J., Zhang, R., Agarwal, D., Yu, T.C., Gunda, V., Lopez, O., Kim, J., Yin, S., Deng, B., Li, K., et al.: Echowrist: Continuous hand pose tracking and hand-object interaction recognition using low-power active acoustic sensing on a wristband. arXiv preprint arXiv:2401.17409 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "text", + "content": "37. Li, C., Li, S., Gao, Y., Zhang, X., Li, W.: A two-stream neural network for pose-based hand gesture recognition. IEEE Transactions on Cognitive and Developmental Systems 14(4), 1594-1603 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "text", + "content": "38. Li, J., Xie, X., Pan, Q., Cao, Y., Zhao, Z., Shi, G.: Sgm-net: Skeleton-guided multimodal network for action recognition. Pattern Recognition 104, 107356 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "text", + "content": "39. Li, M., Chen, S., Chen, X., Zhang, Y., Wang, Y., Tian, Q.: Actional-structural graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3595-3603 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "40. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "text", + "content": "41. Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence 42(10), 2684-2701 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 587 + ], + "type": "text", + "content": "42. Liu, Y., Zhang, S., Gowda, M.: Neuropose: 3d hand pose tracking using emg wearables. In: Proceedings of the Web Conference 2021. pp. 1471-1482 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "type": "text", + "content": "43. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "44. Liu, Z., Zhang, H., Chen, Z., Wang, Z., Ouyang, W.: Disentangling and unifying graph convolutions for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 143-152 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "45. Ma, J., Damen, D.: Hand-object interaction reasoning. In: 2022 18th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). pp. 1-8. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "46. Ohkawa, T., He, K., Sener, F., Hodan, T., Tran, L., Keskin, C.: Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12999-13008 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "text", + "content": "47. Patrick, M., Campbell, D., Asano, Y., Misra, I., Metze, F., Feichtenhofer, C., Vedaldi, A., Henriques, J.F.: Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems 34, 12493-12506 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 240, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 283 + ], + "type": "text", + "content": "48. Plizzari, C., Cannici, M., Matteucci, M.: Spatial temporal transformer network for skeleton-based action recognition. In: Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10-15, 2021, Proceedings, Part III. pp. 694-701. Springer (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 328 + ], + "type": "text", + "content": "49. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3d pose and tracking for human action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 640-649 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 330, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 361 + ], + "type": "text", + "content": "50. Sabater, A., Alonso, I., Montesano, L., Murillo, A.C.: Domain and view-point agnostic hand action recognition. IEEE Robotics and Automation Letters 6(4), 7823-7830 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "type": "text", + "content": "51. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21096-21106 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "text", + "content": "52. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "text", + "content": "53. Shan, D., Geng, J., Shu, M., Fouhey, D.F.: Understanding human hands in contact at internet scale. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9869-9878 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "type": "text", + "content": "54. Shi, L., Zhang, Y., Cheng, J., Lu, H.: Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12026-12035 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 509, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 530 + ], + "type": "text", + "content": "55. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "text", + "content": "56. Soo Kim, T., Reiter, A.: Interpretable 3d human action analysis with temporal convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 20-28 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "text", + "content": "58. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "M.S. Shamil et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 534 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "60. Vemulapalli, R., Arrate, F., Chellappa, R.: Human action recognition by representing 3d skeletons as points in a lie group. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 588-595 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "61. Vondrick, C., Pirsiavash, H., Torralba, A.: Anticipating visual representations from unlabeled video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 98-106 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "62. Wang, J., Liu, Z., Wu, Y., Yuan, J.: Mining actionlet ensemble for action recognition with depth cameras. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 1290-1297. IEEE (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "text", + "content": "63. Weiyao, X., Muqing, W., Min, Z., Ting, X.: Fusion of skeleton and rgb features for rgb-d human action recognition. IEEE Sensors Journal 21(17), 19157-19164 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "text", + "content": "64. Wen, Y., Pan, H., Yang, L., Pan, J., Komura, T., Wang, W.: Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21243-21253 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "65. Wen, Y., Tang, Z., Pang, Y., Ding, B., Liu, M.: Interactive spatiotemporal token attention network for skeleton-based general interactive action recognition. arXiv preprint arXiv:2307.07469 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "type": "text", + "content": "66. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13587-13597 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "67. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European conference on computer vision (ECCV). pp. 305–321 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "68. Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "69. Zhang, S., Liu, X., Xiao, J.: On geometric features for skeleton-based action recognition using multilayer LSTM networks. In: 2017 IEEE winter conference on applications of computer vision (WACV). pp. 148-157. IEEE (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "text", + "content": "70. Zhang, X., Xu, C., Tian, X., Tao, D.: Graph edge convolutional neural networks for skeleton-based action recognition. IEEE transactions on neural networks and learning systems 31(8), 3047-3060 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 500, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 534 + ], + "type": "text", + "content": "71. Zhang, Y., Wu, B., Li, W., Duan, L., Gan, C.: Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 3229-3237 (2021)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "HandFormer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_content_list.json b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c57c55cfe695dc4bb1e37e6847dcac0b734b3072 --- /dev/null +++ b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_content_list.json @@ -0,0 +1,1772 @@ +[ + { + "type": "text", + "text": "On the Viability of Monocular Depth Pre-training for Semantic Segmentation", + "text_level": 1, + "bbox": [ + 274, + 140, + 728, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong Lao1, Fengyu Yang2, Daniel Wang2, Hyoungseob Park2, Samuel Lu1, Alex Wong2, and Stefano Soatto1", + "bbox": [ + 227, + 210, + 776, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UCLA Vision Lab, Los Angeles, CA 90024, USA", + "bbox": [ + 328, + 253, + 669, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lao, soatto}@cs.ucla.edu, samuellu@ucla.edu", + "bbox": [ + 328, + 268, + 671, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Yale Vision Lab, New Haven, CT 06511, USA", + "bbox": [ + 328, + 282, + 661, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fengyu.yang, daniel.wang.dhw33, hyoungseob.park, alex.wong}@yale.edu", + "bbox": [ + 233, + 297, + 767, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The question of whether pre-training on geometric tasks is viable for downstream transfer to semantic tasks is important for two reasons, one practical and the other scientific. If the answer is positive, we may be able to reduce pre-training costs and bias from human annotators significantly. If the answer is negative, it may shed light on the role of embodiment in the emergence of language and other cognitive functions in evolutionary history. To frame the question in a way that is testable with current means, we pre-train a model on a geometric task, and test whether that can be used to prime a notion of \"object\" that enables inference of semantics as soon as symbols (labels) are assigned. We choose monocular depth prediction as the geometric task, and semantic segmentation as the downstream semantic task, and design a collection of empirical tests by exploring different forms of supervision, training pipelines, and data sources for both depth pre-training and semantic fine-tuning. We find that monocular depth is a viable form of pre-training for semantic segmentation, validated by improvements over common baselines. Based on the findings, we propose several possible mechanisms behind the improvements, including their relation to dataset size, resolution, architecture, in/out-of-domain source data, and validate them through a wide range of ablation studies. We also find that optical flow, which at first glance may seem as good as depth prediction since it optimizes the same photometric reprojection error, is considerably less effective, as it does not explicitly aim to infer the latent structure of the scene, but rather the raw phenomenology of temporally adjacent images. Code: https://github.com/donglao/DepthToSemantic.", + "bbox": [ + 259, + 339, + 743, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Depth estimation $\\cdot$ semantic segmentation $\\cdot$ pre-training", + "bbox": [ + 261, + 699, + 723, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 734, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We probe the following seemingly counter-intuitive hypothesis:", + "bbox": [ + 215, + 763, + 668, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can pre-training on a geometric task benefit a downstream semantic task? Geometric inference is often viewed as a low-level vision task requiring little abstraction that is needed for semantics [26]. For example, depth can be aquired through minimizing reprojection error, i.e. from multi-view or videos,", + "bbox": [ + 212, + 779, + 766, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "directly from range sensors. Both can be performed procedurally, without inductive learning, rendering depth a meaningless task for pre-training. However, induction is needed to infer one 3D scene among infinitely many compatible with the same 2D image. Therefore, if a model could solve this ill-posed problem, it would provide evidence of the viability of pre-training with little to no human intervention, which is important in specialized data domains for which little annotated data is publicly available.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on testing monocular depth as the pre-training geometric task, and semantic segmentation as the downstream semantic task. They are purposefully chosen: Training deep neural networks for semantic segmentation requires labor-intensive pixel-level annotation, so the choice of pre-training is essential to its performance. Existing studies have shown mixed results about the relationship between the two tasks. Taskonomy [57], a framework for measuring relationships between visual tasks, suggests that depth estimation is \"far\" from semantic segmentation, while recent work [17] shows that depth pre-training can beat \"closer\" tasks like image classification. Prior work [19, 23, 40, 42] has also shown improvement in semantic segmentation when incorporating depth. This mixed evidence motivates us to take a closer look at the underlying mechanism of how monocular depth may benefit semantic segmentation.", + "bbox": [ + 212, + 257, + 789, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another more subtle reason for exploring this hypothesis is that pre-training is often performed on heavily human-biased datasets [11,38], where the photographer who framed the picture meant to convey a particular concept (say, a cup), and therefore took care to make sure that the manifestation of the concept (the image) prominently features the object by choice of vantage point, illumination, and (lack of) occlusion. This bias is mitigated if data is not purposefully organized into \"shots.\" Unfortunately, existing datasets are mostly composed of purposefully framed shots which could obfuscate the analysis. We note that depth can be inferred without any semantic interpretation [26] regardless of whether the data is captured purposefully or randomly. With monocular depth as the pre-training task, there are two ways of reducing the aforementioned human selective bias: The first involves directly pre-training within the specific domain of interest, leveraging the simplicity of data gathering; The second way is scaling up pre-training by incorporating diverse sources of data, which is made possible by recent developments [33,41,54] on relative depth estimation.", + "bbox": [ + 212, + 441, + 789, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods. We formalize the main hypothesis in Sect. 3. Since it cannot be tested analytically without knowledge of the joint distribution of test images and labels, we propose an empirical testing protocol. We test on monocular depth models trained under multiple forms of supervision, including structure-from-motion, binocular stereo, and depth sensors. We then change the prediction head of the resulting network, either the final layer or the whole decoder, and fine-tune it for semantic segmentation (Fig. 1). We consider depth estimation as a viable pre-training option if it yields comparable improvements to downstream semantic segmentation tasks as other common pre-training practices, e.g. ImageNet classification. To this end, we design a series of controlled experiments to test the effect of choice of initialization (Tab. 1, Fig. 2), training with various datasets", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sizes (Fig. 3), choice of network component to be frozen and fine-tuned (Fig. 4), effect of resolution of training images (Fig. 5). Conclusions are drawn from both quantitative and qualitative (Fig. 6) results.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Findings. Pre-training for depth estimation improves the performance of downstream semantic segmentation across different experimental settings. Particularly, we show that depth estimation is indeed a viable pre-training option as compared to existing methods (Tab. 4). For example, compared to classification, using depth on average improves by $5.8\\%$ mIoU and $5.2\\%$ pixel accuracy on KITTI. As a sanity check, we test both a depth network pre-trained from scratch and one trained after ImageNet initialization, and both outperform classification-based pre-training in downstream semantic segmentation. To control the effect of our choice of architecture, we used our pre-trained encoder to initialize a standard semantic segmentation network [5]. We observed similar findings on Cityscapes and NYU-V2 regardless of how depth training is supervised. Inferring depth without explicit supervision typically involves minimizing the prediction error, just like optical flow. Somewhat surprisingly, not only does pre-training for depth outperform optical flow, but the latter is often worse than random initialization (Fig. 7). One may also argue that observed improvements mainly come from the availability of in-domain pre-training data for depth. To test this conjecture, we fine-tune a depth model [55] trained on large-scale out-of-domain data. Improvements in semantic segmentation reveal that when trained at scale, depth models show strong transferability to unseen downstream data domains.", + "bbox": [ + 212, + 193, + 787, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 512, + 387, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-training aims to learn a representation (function) of the test data that is maximally informative (sufficient), while providing some kind of complexity advantage. In our case, we measure complexity by the validation error after fine-tuning on limited amount of labeled data, which measures the inductive value of pretraining. The recent literature comprises a large variety of \"self-supervised\" methods that are purportedly task-agnostic. In reality, the task is specified indirectly by the choice of hand-designed nuisance transformations that leave the outcome of inference unchanged. Such transformations are sampled through data augmentation while the image identity holds constant (Contrastive Learning) [4,8,9,38], or reconstruction [3]. Group transformations organize the dataset into orbits, which contrastive learning tries to collapse onto its quotient, which is a maximal invariant. Such a maximal invariant is transferable to all and only tasks for which the chosen transformation is uninformative. For group transformation, the maximal invariant can, in theory, be computed in closed form [44]. In practice, contrastive learning are extended to non-group transformations, e.g. occlusions, as seen in language [2] and images [7]. All self-supervised methods boil down to hand-designed and quantized subsets of planar domain diffeomorphisms (discrete rotations, translations, scaling, reflections, etc.), range homeomorphisms (contrast, colormap transformations) and occlusion masks.", + "bbox": [ + 212, + 553, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6fcc8480d5bd6c2eb105573169c6b1000ee380fcd9a85ba87b3fe89a1323e831.jpg", + "image_caption": [ + "(a) Baseline" + ], + "image_footnote": [], + "bbox": [ + 267, + 140, + 385, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f63a1edecb0fd2377ec7220b4020b2d0b6fa897fa5bb1a55cd9d2ac497d7933b.jpg", + "image_caption": [ + "(b)", + "Fig. 1: Diagram for different pre-training and fine-tuning setups. (a) Common practice: pre-train the encoder, e.g. on ImageNet, attach a decoder, and fine-tune the network. (b) Our best practice: pre-train the network by monocular depth, and fine-tune for semantic segmentation. (c) Cross architecture: for fair comparisons with common practice, we pre-train by depth, replace the decoder, and fine-tune. (d) To test the quality of pre-trained encoders, we fix the encoders and fine-tune decoders only." + ], + "image_footnote": [], + "bbox": [ + 395, + 140, + 470, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eed0a2744226ecb4a62d10439189c1b3ba6250eda358a00f71453adfaee1c56a.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 477, + 138, + 594, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e2bd894af91c3ce860bc89c8d11b734c71cc5a2f5d9f94a2a7993a925c30f1fc.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 602, + 137, + 733, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our case, rather than hand-designing the nuisance transformations assumed to be shared among pre-training and fine-tuning tasks, we let the scene itself provide the needed supervision: images portend the same scene, either from the same timestamp (stereo) or adjacent in the temporal domain (video frames), so their variability defines the union of nuisance factors. These include domain deformations due to ego- and scene motion, range transformations due to changes in illumination, and occlusions. In addition to sharing nuisance variability, pretraining and fine-tuning tasks should ideally also share the hypothesis space. It may seem odd to choose a geometric task, where the hypothesis space is depth, to pre-train for a semantic task, where the hypothesis space is a discrete set of labels. However, due to the statistics of range images [21] and their similarity to the statistics of natural images [22], this is actually quite natural: A range map is a piecewise smooth function defined on the image domain, whereas a segmentation map is a piecewise constant function where the levels are mapped to arbitrary labels. As a result, the decoder for depth estimation can be easily modified for semantic segmentation. A discussion of this choice, specifically on the representational power of deterministic predictors, in Sect. 5. [19, 23] also utilize depth for semantic segmentation. [23] proposes pre-training on relative depth prediction, and [19, 20] utilizes self-supervised depth estimation on video sequences. Our experiments validate their findings. We further investigate whether features obtained purely from monocular depth improve semantic segmentation.", + "bbox": [ + 212, + 366, + 787, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular depth [16, 50] may use different supervision, either through additional sensors [14, 51, 52], or synthetic data [36, 47, 56], but none require human annotation. Some use regularizers with sparse seeds [35, 48, 49], or adopt pretrainings [41]. We design experiments agnostic to how depth models are trained, but also make comparisons across different forms of depth supervision (Tab. 2).", + "bbox": [ + 212, + 683, + 787, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Formalization", + "text_level": 1, + "bbox": [ + 215, + 780, + 387, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $x: D \\subset \\mathbb{R}^2 \\to \\{0, \\ldots, 255\\}^3$ be an image, where the domain $D$ is quantized into a lattice, $z: D \\to \\{1, \\ldots, Z\\}$ a depth map with $Z$ depth or disparity levels,", + "bbox": [ + 212, + 808, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $y: D \\to \\{1, \\ldots, K\\}$ a semantic segmentation map. In coordinates, each pixel in the lattice, $(i,j) \\in \\{1, \\ldots, N\\} \\times \\{1, \\ldots, M\\}$ is mapped to RGB intensities by $x(i,j)$ , a depth by $z(i,j)$ , and a label by $y(i,j)$ . Despite the discrete nature of the data and the hypothesis space, we relax them to the continuum by considering the vectors $\\mathbf{x} \\in \\mathbb{R}^{NM3}$ , $\\mathbf{y} \\in \\mathbb{R}^{NMK}$ and $\\mathbf{z} \\in \\mathbb{R}^{NM}$ . With a slight abuse of notation, $y \\in \\{1, \\ldots, K\\}$ denotes a single label and $\\bar{y} \\in \\mathbb{R}^K$ its embedding, often restricted to a one-hot encoding.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, consider a dataset $\\mathcal{D}_z = \\{\\mathbf{x}_t^i,\\mathbf{z}_t^i\\}_{i,t = 1}^{V,T_i}$ comprised of $V$ image sequences each of length $T_{i}$ . For the simplicity of the notations, in the case of multi-view stereo, we also consider multiple 2D image inputs as a \"sequence\" without loss of generality. In the case of supervised depth estimation, synchronized depth maps $z_{t}^{i}$ 's are measured by a range sensor. Typical datasets supporting depth estimation may include just image sequences or both modalities.", + "bbox": [ + 212, + 251, + 784, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training for monocular depth estimation yields a mapping $\\phi_w: \\mathbf{x} \\mapsto \\mathbf{z}$ , parametrized by weights $w$ in a neural network, via", + "bbox": [ + 212, + 343, + 782, + 373 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw = \\arg \\min _ {w, g _ {t}} \\sum_ {i, j, n, t} \\ell \\left(x _ {t + 1} ^ {n} (i, j), \\hat {x} _ {t} ^ {n} (i, j)\\right) + \\lambda \\sum_ {i, j, n, t} \\ell \\left(z _ {t} ^ {n} (i, j), \\hat {z} _ {t} ^ {n} (i, j)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 385, + 784, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{x}_t$ is the warping of an image $x$ from $t$ to $t + 1$ based on camera pose $g_t$ . Here we consider a generic formulation for different modalities of depth estimation. The first term in Eq. (1) measures reprojection error across frames, and the second term measures the distance between estimated depth values and the ground-truth from the range sensor. In the case of unsupervised depth estimation from videos, e.g. [16], only the reprojection loss is considered; while in the case of supervised depth estimation from single images, e.g. [37], $T_i = 1$ for all $i$ 's, so only the second term is minimized.", + "bbox": [ + 212, + 428, + 784, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The goal is to use these representations as encodings of the data to then learn a semantic segmentation map. In practice, the representations above are implemented by deep neural networks, that can be truncated at intermediate layers thus providing embedding spaces larger than the respective hypothesis spaces. We refer to the parts before and after this intermediate layer encoder and decoder, respectively. We overload the notation and refer to the encoding as $\\mathbf{h} = \\phi_w(\\mathbf{x})$ for both depth estimation and other pre-training methods, presumably with weights $w'$ , assuming they have the same encoder architecture. The goal of semantic segmentation is then to learn a parametrized map $\\psi_{w''}: \\mathbf{h} \\mapsto \\mathbf{y}$ using a small but fully supervised dataset $\\mathcal{D}_s = \\{\\mathbf{x}^n, \\mathbf{y}^n\\}_{n=1}^N$ , by minimizing some loss function or (pseudo-)distance in the hypothesis space $d(\\mathbf{y}, \\hat{\\mathbf{y}})$ , where", + "bbox": [ + 212, + 549, + 785, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {\\prime \\prime} = \\arg \\min _ {w} \\sum_ {n = 1} ^ {N} d \\left(\\mathbf {y} ^ {n}, \\psi_ {w} \\left(\\mathbf {h} ^ {n}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 729, + 784, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "plus customary regularizers. In the aggregate, we have a Markov chain: $\\mathbf{x} \\longrightarrow \\mathbf{h} = \\phi_w(\\mathbf{x}) \\longrightarrow \\mathbf{y} = \\psi_{w''}(\\mathbf{h}) = \\psi_{w''} \\circ \\phi_w(\\mathbf{x})$ for depth estimation, and $\\mathbf{x} \\longrightarrow \\hat{\\mathbf{h}} = \\phi_{w'}(\\mathbf{x}) \\longrightarrow \\mathbf{y} = \\psi_{w''} \\circ \\phi_{w'}(\\mathbf{x})$ for other pre-training methods. A representation obtained through a Markov chain is optimal (minimal sufficient) only if the", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/550462561bb8be7f2f453b434711b3b1bf22c0fdbbf63202b819b7907edc8ce6.jpg", + "table_caption": [ + "Table 1: Semantic segmentation accuracy on KITTI. Unsupervised depth as pre-training improves semantic segmentation accuracy under all settings. Our best practice (in blue) improves common practice (in purple) by $7.53\\%$ mIoU and $4.68\\%$ pixel accuracy. Freezing the encoder with ImageNet pre-training (in red) is worse than no pre-training (random initialization). DeepLabV3†: with ResNet50 encoder." + ], + "table_footnote": [], + "table_body": "
Fine-tune AllFreeze Encoder
ResNet18ResNet50DeepLabV3†ResNet18ResNet50
Pre-trainingmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.Acc
None41.3570.7544.6673.3721.9352.3241.2470.5237.7267.38
ImageNet45.1572.3944.6573.0643.3972.6633.3365.3432.0362.53
Depth-Rand46.0072.4349.9076.2843.4371.3443.0272.3845.7974.71
Depth50.2076.3950.9277.3443.7772.6846.5374.4246.5574.48
", + "bbox": [ + 236, + 226, + 771, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "intermediate variable $\\mathbf{h}$ or $\\hat{\\mathbf{h}}$ reduces the Information Bottleneck [45] to zero. In general, there is information loss, so we formalize the key question as whether the two Information Bottleneck Lagrangians satisfy the following:", + "bbox": [ + 212, + 354, + 787, + 402 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH (\\mathbf {y} | \\mathbf {h}) + \\beta I (\\mathbf {h}; \\mathbf {x}) \\stackrel {?} {\\leq} H (\\mathbf {y} | \\hat {\\mathbf {h}}) + \\beta^ {\\prime} I (\\hat {\\mathbf {h}}; \\mathbf {x}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 411, + 785, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\beta'$ are hyperparameters that can be optimized as part of the training process, and $I, H$ denotes the (Shannon) Mutual Information and cross-entropy respectively. If the above is satisfied, then pre-training for depth estimation is a viable option, or even better than pre-training with another method. It would be ideal if this question could be settled analytically. Unfortunately, this is not possible, but the formalization above suggests a protocol to settle it empirically.", + "bbox": [ + 212, + 445, + 787, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To test this empirically, we use the validation error on a supervised dataset $\\mathcal{D}_s$ as a proxy for residual information. We conduct fine-tuning under several configurations (Fig. 1): with respect to $w''$ using $\\mathcal{D}_s$ , i.e. yielding a comparison of the raw pre-trained back-bones (encoders, $w, w'$ ), or with respect to both $w''$ and $w$ (for depth estimation) or $w'$ (for other pre-training methods). Finally, all four resulting models can be compared with one obtained by training from scratch by optimizing a generic architecture with respect to $w''$ alone.", + "bbox": [ + 212, + 537, + 787, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 664, + 375, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Controlled Experiments with Few-shot Fine-tuning", + "text_level": 1, + "bbox": [ + 214, + 694, + 683, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first cover an extensive collection of controlled experiments and ablations to gain insights into the main hypothesis. We specifically conduct experiments under the few-shot setting, where only a small amount of labels are available for fine-tuning, to highlight the role of pre-training.", + "bbox": [ + 212, + 719, + 785, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KITTI [15] contains 93000 video frames for depth training with 200 densely annotated images for semantic segmentation. Segmentation results are evaluated by mean IoU (mIoU) and pixel-level accuracy (P.Acc). We randomly choose a small training set of 16 images and limit data augmentation to horizontal", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63ed67334ad45839306c3a0fcd33015c518c5652aa2830b79a9a7bb565dc3a1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 146, + 426, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2009dc9ac4121e32e212ab1db1bc7ad4bc35de0882060e3dbcbc35a2bfcc203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 146, + 566, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/708f8fe632a1972f66d4f0e7b1fae7780ed23cf248bd6ebe8ac1a8cc9f57badb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 146, + 705, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b57aa0b486123b48c5444ef9073ebad1c168801ee98a5f5ce79684c96e982bb.jpg", + "image_caption": [ + "Fig. 2: Comparison between different network initializations. Models initialized by depth pre-training (unsupervised) train faster and achieve higher final accuracy." + ], + "image_footnote": [], + "bbox": [ + 279, + 228, + 426, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1eb239e452bfacfa0deab7913bd57fec3a168e8c42b0fa1a728546a39730c224.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 229, + 565, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3610a031aa4f0fcaba0e6933c6a9ff73af11574dd04dd27df45b8d802b10385b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 229, + 705, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f3c3dcb5e590a9164029753c969b39c88c670981451ee881273f926b6bb44ae.jpg", + "image_caption": [ + "Fig. 3: Final accuracy vs different training set size. Under all training set sizes, our best practice constantly outperforms ImageNet pre-trained. Encoder: ResNet 18." + ], + "image_footnote": [], + "bbox": [ + 218, + 368, + 352, + 449 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f988e7db50399416bf7c87d3dd26a68c4b1f394f3696e2c1f2d1f9161fd8e30a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 368, + 496, + 449 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/56c3c460cb8c50a1e16342ff1a9c7debe1ffb66bcd9ef17721fd29c1dfa6ca4c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Depth Sup.mIoUP.Acc
Video46.0072.43
Stereo49.1174.58
Lidar52.7877.17
", + "bbox": [ + 570, + 381, + 743, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Forms of depth supervision matter. Direct supervision with Lidar works the best, followed by stereo (with known camera pose), and monocular video (camera pose unknown).", + "bbox": [ + 524, + 450, + 787, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "flips to highlight the impact of pre-training, except for Fig. 3 where we test on different training set partitions. We use Monodepth2 [16] for depth pre-training. For semantic segmentation, we replace the last layer of the decoder with a fully connected layer, using the finest scale of the multi-scale output. We test on ResNet18 and ResNet50 encoders due to their compatibility with various network architectures and widely public-available pre-trained models. Fig. 1 summarizes our experimental setup and Tab. 1 summarizes the outcomes. In all cases, depth pre-training improves segmentation accuracy.", + "bbox": [ + 212, + 551, + 787, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Full model. Fig. 2 shows the evolution of training loss and model accuracy. Depth pre-training outperforms ImageNet and random initialization. ImageNet pre-training slightly improves over random initialization on ResNet18, but shows almost identical performance to random initialization on ResNet50. Depth also speeds up training, taking $\\sim 5000$ iterations to converge, while ImageNet takes 15000 to 20000. Similar results on full-resolution are deferred to the Supp. Mat.", + "bbox": [ + 212, + 672, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different training set size. Fig. 3 shows that depth pre-training improves final segmentation scores over all dataset partitions (with ResNet 18). When training samples increase (e.g. 128), ImageNet and depth pre-training (from random initialization) are comparable to random initialization, but depth pretraining initialized with ImageNet yields the best results.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7989ba00cd01927a57550e824864c98edf572490419520ab6a170496b9e03d58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 146, + 426, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7859699f2b2dd1c5c74326a3051db54ef5e56edec978bccc8bc250f41d11a2ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 148, + 566, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27ca382456b6d0cd7090e57af860a5f4ea2ab26f137ca9238f5f68a24f71e8df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 146, + 705, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb00436f24d07ccbcb894ac30efe77179fb4a6a1ba28b7fa8f78b016b60ac088.jpg", + "image_caption": [ + "Fig. 4: Frozen encoder results. Using an encoder pre-trained by depth significantly outperforms one with random weights and one for ImageNet classification. Note that in this experiment, ImageNet pre-training performs worse than random initialization." + ], + "image_footnote": [], + "bbox": [ + 279, + 229, + 426, + 308 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18f43b5f8d558ab010a601bfabcf3977a52c16938b73db7e8f483f78a4647502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 232, + 565, + 308 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2215ad626162615c9f57b4aed9c67f819e705da8ddc1743bb6a457f79c316238.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 229, + 705, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different forms of depth supervision. Pre-training quality depends on the source of supervision. Training on monocular videos involves minimizing reprojection error, which requires joint estimation of depth and pose. Since pose estimation relies on sufficiently distinctive textures (large eigenvalues of the structure tensor of image gradients), the supervision signal is sparse. Conversely, with stereo images, one may omit the pose network when training. With depth sensors, training losses minimize error w.r.t. dense or semi-dense measured depth, offering stronger supervision. Tab. 2 shows that supervising with Lidar is the best, followed by stereo, and monocular video - all improving over ImageNet.", + "bbox": [ + 212, + 405, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frozen encoder. We freeze pre-trained encoders, and fine-tune the decoder only, testing the ability of features from pre-trained encoders to capture semantics. With both ResNet18 and ResNet50, encoder pre-trained for depth significantly outperforms random initialization and ImageNet pre-trained (Fig. 4). It is surprising that ImageNet pre-training is detrimental in this case (after a grid search over learning rates): worse than fixed random weights. This suggests that while classification is a semantic task, it removes semantic information about the scene due to the object-centric bias in datasets. ImageNet pre-training tends to favor image-level features, that may not capture object shape, making finetuning the decoder difficult for segmentation. We conjecture that these uncontrolled biases in ImageNet pre-training cause difficulties in directly predicting segmentation without fine-tuning the encoder.", + "bbox": [ + 212, + 547, + 787, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initializing with a pre-trained encoder only. To eliminate the effect of the depth-initialized decoder and only test the encoder, we replace the decoder by 'fresh' randomly initialized weights and fine-tune the whole network. Tab. 3 shows that depth pre-training outperforms ImageNet when the effect of pretraining is isolate to the encoder. This is also supported by neural activations in Fig. 6 where the regions activated after depth pre-training align well with semantic boundaries. Nonetheless, the decoder does play a role in segmentation", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/597977ce0ec739de430852cf341be3ed04218794acff87cfc543c836aff5f8f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-trainingmIoUP.Acc
ViT-1 Res. 18None41.3570.75
ImageNet45.1572.39
Depth (encoder only)46.6975.04
ViT-1 Res. 50None44.6673.37
ImageNet44.6573.06
Depth (encoder only)46.9973.57
ViT-1 Res. 18ImageNet57.5381.48
Depth (encoder only)58.1281.94
", + "bbox": [ + 218, + 147, + 472, + 281 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Initializing with depth encoder and random decoder. Initializing with the depth encoder and a random decoder outperforms ImageNet initialization, but is worse than initializing with both encoder and decoder from the depth network (see Tab. 1).", + "bbox": [ + 215, + 287, + 475, + 385 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ff8224c0f29b407897acb4ca4d8827419e6d3e6ab9ed2a1f3e239dfc142967e0.jpg", + "image_caption": [ + "Fig.5: Mismatch between object scales. ImageNet models are trained with a fixed input size and objects of interest tend to have a similar scale. However, objects in semantic segmentation dataset vary drastically in scale. Pre-training for depth provides robustness to scale change." + ], + "image_footnote": [], + "bbox": [ + 514, + 143, + 771, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "accuracy: Initializing the whole network with depth pre-training still performs the best, an advantage that is not afforded by a classification head.", + "bbox": [ + 212, + 415, + 782, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results on Vision Transformers. Given that ViTs [12] necessitate extensive pre-training, conducting comprehensive ablation studies on ViTs, as feasible with ResNet, becomes impractical due to data constraints. Thus, we conduct experiments using the DPT [41] depth model trained on a collection of datasets [33] (not including KITTI), and report the best results in Tab. 3 after an exhaustive search for the optimal learning rate. Notably, we identified the optimal learning rate for this experiment to be 5e-8, with larger learning rates yielding suboptimal results. This suggests that DPT inherently provides robust representations sufficient for segmentation, requiring minimal fine-tuning in comparison to CNNs.", + "bbox": [ + 212, + 446, + 784, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cross architecture. To test whether depth pre-training favors our particular choice of architecture, we use the same encoders to initialize DeepLab V3 and follow the same fine-tuning procedure as common practice, i.e., ImageNet initialization. Results are presented in Tab. 1. All pre-trainings significantly improve accuracy compared with random initialization. The depth model trained from scratch provides the same level of performance as ImageNet, while subsequent depth training after ImageNet pre-training leads to further improvements.", + "bbox": [ + 212, + 583, + 784, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Robustness to object scales. Given a fixed resolution, \"primary\" objects in object-centric data tend to have a similar scale. For example, ImageNet models are trained on $224 \\times 224$ , thus cars typically have a size of 100 to 200 pixels. In KITTI, however, cars appear at different scales, varying from a few to a few hundred pixels. Fig. 5 illustrates the scale mismatch between datasets. On the other hand, depth pre-training can be done in the same domain as segmentation, hence having robustness to object scales. We examine such robustness: Pretraining on one resolution, fine-tuning on another. Since higher resolution images contain smaller scales (yellow box in Fig. 5), pre-training on them should still work on smaller images. Pre-trained on smaller resolutions, however, should work", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/eee26b152085dbc0595ff981a9c3c25416c4e113fd32030e2f446c4e1bdf9fb1.jpg", + "image_caption": [ + "Fig.6: Neural activation and semantic segmentation result. We visualize the neural activation map for a shallow layer of ResNet 18 encoder trained from different initializations, and their corresponding segmentation results. Boundaries are better aligned to semantic boundaries in our model." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 772, + 224 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "poorly on larger images. Our results validate this conjecture: The former achieves a final mIoU of 48.54 (ImageNet: 45.15) while the latter diverges during training.", + "bbox": [ + 212, + 325, + 782, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Neural Activation. We visualize the activations of the ResNet18 encoder by Grad-CAM [43], which was originally designed for classification. We modify it for segmentation by inspecting the gradient of neural response to the summation of predicted labels for pixels instead of one single class label. We visualize shallow layers (before pooling) for high spatial resolution. Fig. 6 shows neural activation maps and segmentation outputs from depth pre-training align with semantic boundaries. This confirms not just the similarities between natural and range image statistics discussed in [22], but also the bias introduced by classification, as activations of ImageNet pre-trained encoder do not resemble object boundaries.", + "bbox": [ + 212, + 357, + 784, + 492 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison with optical flow. One hypothesis for the effectiveness of depth pre-training is that the process leverages the statistics of natural scenes where simply-connected components of the range map often correspond to semantically consistent regions. Thus, fine-tuning simply aligns the range of two piece-wise smooth functions. We challenge this hypothesis by trying optical flow, which also exhibits a piecewise-smooth range and is obtained by minimizing the same photometric error. We train optical flow on a siamese network with two shared-weight encoders. While both optical flow and depth capture multiply-connected object boundaries (Fig. 7), using encoders pre-trained for optical flow is detrimental. We conjecture that optical flow does not capture the stable inductive bias afforded by the static component of the underlying scene. Specifically, optical flow is compatible with an underlying 3D geometry only when the scene is rigid, but rigidity is not enforced when inferring optical flow. In contrast, depth forces recognition of rigidity and discards moving objects [29-31] as outliers, which then enables isolating them, also beneficial to semantic segmentation.", + "bbox": [ + 212, + 492, + 785, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison with other pre-training methods. In Tab. 4, for completeness, we report results on supervised pre-training by semantic segmentation on MS-COCO [34]. Unsurprisingly, pre-training on the same task with additional annotated data yields good performance. Depth estimation, despite not needing additional annotation, yields on-par performance. We also report results on masked autoencoding. We remove random rectangular regions from images, and the network aims to reconstruct the original image. Also known as ' inpainting' [1], it is considered an effective method for feature learning [39]. It yields", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9df5195851d1c662822c6e7075ce10b2f892508de1a2e1aa6866c8f876ac92f9.jpg", + "image_caption": [ + "Fig. 7: Depth helps, flow hurts. Although both are pre-trained by minimizing a photometric reconstruction error, monocular depth outperforms optical flow. This stems from the fact that inferring depth from a single image is ill-posed so the network learns inductive priors that are rich in semantics over the structures within a scene. In contrast, any discriminative features will support the correspondence search, so the flow network is not constrained to learning semantics, yielding poor fine-tuning accuracy." + ], + "image_footnote": [], + "bbox": [ + 217, + 138, + 514, + 218 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a3ba1cced9162a14101219726b1557cc58e81ec12538ee39eb487b228e95e75e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AllFreeze
mIoU↑mIoU↑
None41.35-41.24-
Flow38.47-2.8832.19-9.05
Depth-Rand46.004.6543.021.78
Depth50.208.8546.535.29
", + "bbox": [ + 519, + 136, + 766, + 223 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bc4d4dd71127875d57711a8707f648ac1cc171130c1a72c04a40c95534e5283e.jpg", + "table_caption": [ + "Table 4: Comparison with different pre-trainings. Encoder: ResNet50; Reconstruction: by inpainting randomly corrupted regions (masked autoencoding); Supervised Segmentation: trained on MS-COCO." + ], + "table_footnote": [], + "table_body": "
Pre-trainingmIoUP.AccPre-trainingmIoUP.Acc
Supervised Segmentation51.2874.88Contrastive (DINO)44.1971.36
Depth50.9277.34Optical Flow42.7271.80
Reconstruction (MAE)47.1874.16Contrastive (MOCO V2)37.0465.91
", + "bbox": [ + 233, + 385, + 769, + 445 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "inferior performance compared with depth. We conjecture that this is due to artificial rectangular masking which does not respect the natural image statistics, while monocular depth estimation yields occluded regions that border on objects' silhouettes. We also provide results initializing the encoder with contrastive learning (MOCO V2 [9] and DINO [4]). While they also remove bias in human annotation for classification, similar to supervised classification, they are still prone to the inherent inductive bias in pre-training data.", + "bbox": [ + 212, + 473, + 784, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Full-scale Fine-tuning on the Whole Dataset", + "text_level": 1, + "bbox": [ + 214, + 602, + 625, + 618 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We now extend experiments to fine-tuning on full-scale datasets, aiming to validate the findings drawn from the controlled experiments conducted on KITTI.", + "bbox": [ + 212, + 628, + 782, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cityscapes [10] contains 2975 training and 500 validation images. Each image has a resolution of $2048 \\times 1024$ densely labeled into 19 semantic classes. The dataset also has 20000 unlabeled stereo pairs with a disparity map, converted to depth via focal length and camera baseline. Like KITTI, Cityscapes is also an outdoor driving dataset. Here we minimize an L1 loss between depth estimates and depth computed from stereo. We modify the prediction head of DeepLabV3 to train for depth, then re-initialize the last layer of the decoder for semantic segmentation. We are unable to reproduce the original numbers. For a fair comparison, we retrained [6] using the discussed pre-training methods and finetuned under the same training protocol i.e. batch size, augmentations, schedule, etc. Tab. 5 summarizes the outcomes. We present not only the most favorable results from an extensive training process but also results from a controlled approach", + "bbox": [ + 212, + 659, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/103fa43a073919402ccb77ef3623b46d0a462f45d0a82cf4f1aac6060019833b.jpg", + "table_caption": [ + "Table 5: Segmentation accuracy on Cityscapes. Similar to KITTI, pre-training for depth improves segmentation accuracy. Interestingly, under the controlled settings with limited data augmentation and fewer fine-tuning epochs, the model achieves higher performance when pre-training on cropped $256 \\times 256$ patches." + ], + "table_footnote": [], + "table_body": "
FullControlled
Training mIoUP.Acc.Validation mIoUP.Acc.Training mIoUP.Acc.Validation mIoUP.Acc.
None76.1095.4163.9793.7673.8295.2660.4393.07
ImageNet81.8496.4670.4195.7576.7095.7161.8093.40
Depth83.4696.8073.1795.2477.4295.8262.5793.46
Depth-cropped86.8097.4172.2295.0179.9096.2465.0994.00
", + "bbox": [ + 261, + 212, + 741, + 313 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d182751251ca513e3f0bbac454397f6dcd95b2a6ddf6f5bbb99c3572735718bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 330, + 426, + 409 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4c04280ee8bdbcc57a030909706e06060d6b1e9dbe20301af63685ff7d8299cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 329, + 566, + 409 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7283be2ba11d177f7d65de6364b462a752685b4ce1cd8dc61e75baa6f9ecbc3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 329, + 705, + 409 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/aa256eaaafae5a18d3feaeb31da492f1ec839e5bb93d98f4c12e5685d519c912.jpg", + "image_caption": [ + "Fig. 8: Results on NYU-V2. Similar to KITTI, initializing with depth pre-trained weights trains faster and significantly improves semantic segmentation accuracy." + ], + "image_footnote": [], + "bbox": [ + 279, + 412, + 426, + 492 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3569bf16a9c6073a3cceb31301edae89cf51e6cf15f321966f2d7ebe15d56dfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 412, + 566, + 492 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6b834b614406ec5fb752514a7fda4cd82b2307f15e6bf88d73676886be64362b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 412, + 705, + 492 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "with restricted data augmentations and fewer training iterations. Remarkably, improvements remain consistent across both scenarios.", + "bbox": [ + 212, + 566, + 784, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "One may argue that since depth and semantic segmentation maps are both piece-wise smooth, adapting from depth is naturally easy if the model is aware of each pixel's relative position in the image. In order to test this statement, instead of pre-training for depth on the full image, we train depth on randomly cropped $256 \\times 256$ patches, and the model has no spatial awareness of the position of a patch in the image, so depth is purely estimated by local information. This practice (Depth-cropped) surprisingly improves semantic segmentation results under controlled settings, showing that using depth as pre-training goes beyond a simple mapping from one smooth function to another. Interestingly, this approach significantly improves training accuracy in the full setting but leads to a slight reduction in validation accuracy, suggesting a potential issue of overfitting. Future research is necessary to delve into this intriguing phenomenon. Another noteworthy observation is that when pre-trained using depth data, the model exhibits superior performance with a higher initial learning rate of 0.1, as opposed to 0.01 used for ImageNet initialization. Conversely, employing an initial learning rate of 0.1 with ImageNet weights can be detrimental and may", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "result in divergence. These findings suggest that pre-training the network with depth estimation may lead to a smoother local loss landscape.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "NYU-V2 [37] is an indoor dataset that contains 795 densely annotated images for training and 654 for testing. There are also 407024 unannotated frames with synchronized depth images captured by a Microsoft Kinect. Since the main hypothesis is agnostic to how depth is learned, we pre-trained for depth using ground-truth as supervision. Unlike outdoor driving, which commonly features sky on top, and road and vehicles in the middle of the image with largely planar camera motion, indoor scenes are characterized by more complex layouts with 6 DoF camera motion, yielding images that are even less likely to resemble the object-centric ones commonly observed in classification datasets. This may be why initializing the model with depth pre-trained weights significantly improves semantic segmentation accuracy with both ResNet18 and ResNet50 (see Fig. 8). Note that pre-training by depth yields faster convergence, similar to KITTI.", + "bbox": [ + 212, + 176, + 787, + 357 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Out-of-domain Transfer from Large-scale Pre-training", + "text_level": 1, + "bbox": [ + 214, + 378, + 705, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "One may conjecture that the advantage of pre-training with monocular depth stems from training within the same domain as downstream semantic segmentation. In practice, in-domain data collection for depth pre-training is indeed ideal since it does not require human labeling. In a scientific context, we are interested in testing this conjecture by investigating the transferability of depth models to tasks outside their original domain. However, common depth datasets are considerably small in scale, making fair comparisons with popular pre-trained models, e.g. MAE [18] and DINO v2 [38] that are trained on millions or even billions of images, infeasible. Fortunately, recent methods [33,41] suggest an alternative approach for learning monocular depth by training for relative depth instead of absolute depth, allowing for the integration of multiple data sources during training. Leveraging Depth Anything [55], a depth model trained on such scaled-up mixed datasets, we fine-tune for segmentation on three out-of-domain downstream datasets: ADE20k [59], PascalVOC [13], and CityScapes [10].", + "bbox": [ + 212, + 402, + 787, + 613 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The PascalVOC dataset comprises 10,582 fully annotated images for training purposes and an additional 1,449 for testing, covering a variety of 20 foreground object classes. On the other hand, ADE20k, a sizable dataset, includes 20,210 training images and 2,000 testing images across 150 different classes. The resolution of PascalVOC and ADE20k is $512 \\times 512$ and $896 \\times 896$ respectively. On both datasets, we try both fine-tuning the whole network and linear probing with a frozen encoder. Reported in Tab. 6, the results align with the findings on ViTs on the KITTI dataset (in Tab. 3), in which case the pre-training data also originates from out-of-domain sources. Notably, consistent improvements in semantic segmentation performance are observed across all datasets compared to the baseline initialization DINO v2 [38]. It is anticipated that the improvement achieved with a frozen encoder is more substantial than when fine-tuning the entire network, which is consistent with our previous results. This trend underscores that, when trained at a scale, depth models exhibit robust transferability to novel downstream data domains, just as other pre-training methods.", + "bbox": [ + 212, + 614, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bf3f2d532c0873a74629d32038946e500056a67cf818e98415783513ce78fd78.jpg", + "table_caption": [ + "Table 6: Out-of-domain transfer with large-scale pre-trainings. We compare Depth Anything with MAE and DINO v2 for semantic segmentation, reporting results (in mIoU) on both fine-tuning (ft) and linear probing (lin.). Depth Anything improves semantic segmentation across all datasets and settings. *: with DINO v2 initialization." + ], + "table_footnote": [], + "table_body": "
Pre-training# of pre-train dataADE20kPascalVOCCityScapes
ftlin.ftlin.ftlin.
MAE [18]1.28 million53.649.0-67.6-58.4
DINO v2 [38]142 million58.147.786.586.382.771.3
Depth Anything [55]63.5 million*59.752.387.787.384.874.8
", + "bbox": [ + 223, + 212, + 781, + 287 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 311, + 354, + 327 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Inferring depth only requires multiple images (e.g. videos or multiple viewpoints) of the same scene [25] or range sensing, both do not require human-induced priors and bias, unlike semantic tasks that rely entirely on induction: We can associate a label to an image because that image has something in common with some other image, portraying a different scene, that some annotator attached a particular label to. That inductive chain has to go through the head of human annotators, who are biased in ways that cannot be easily quantified and controlled. Depth from binocular or motion imagery does not require induction and can be performed ab-ovo. Learning a monocular model from such supervision eliminates the implicit selective bias from human annotators, yet our findings validate the main hypothesis that the inductive bias learned from such a \"human-free\" process transfers well to the downstream semantic segmentation task. Of course, if different supervisions are available, either for semantic [11], segmentation [27], or both [28], we want to incorporate that information.", + "bbox": [ + 212, + 340, + 787, + 551 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our hypothesis and findings are agnostic to how depth is attributed to a single image: One can perform pre-training using monocular videos, stereo, structure light, LIDAR, or even human guidance [58]. One of our pre-training minimizes the photometric reprojection error, used by many predictive and generative approaches. However, an unstructured displacement field is in general not compatible with a rigid motion. Only if this displacement field has the structure of an epipolar transformation [44] is the prediction task forced to encode the 3D structure of the scene. This may explain why video prediction is not as effective for pre-training despite many attempts [24,32,46,53].", + "bbox": [ + 212, + 551, + 787, + 688 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "One limitation of monocular depth estimation is that it may require a calibrated camera, so one cannot use generic videos harvested from the web. There is nothing in principle preventing us from using uncalibrated cameras, simply by adding the calibration matrix $K$ to the nuisance variables. While the necessary conditions for full Euclidean reconstruction are rarely satisfied in consumer videos (for instance, they require cyclo rotation around the optical axis, not just panning and tilting), the degrees of freedom that cannot be reconstructed are moot as they do not affect the reprojection. Moreover, recent progress in training for relative depth from mixed data sources [33, 41, 55] shows the potential to unlock a virtually unlimited volume of training data, warranting future research.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by ONR N00014-22-1-2252 and ARO W911NF-17-1-0304.", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 202, + 321, + 218 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bertalmio, M., Sapiro, G., Caselles, V., Ballester, C.: Image inpainting. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 417-424 (2000)", + "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "3. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017)", + "6. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017)", + "7. Chen, M., Artières, T., Denoyer, L.: Unsupervised object segmentation by redrawing. Advances in neural information processing systems 32 (2019)", + "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009)", + "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "13. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88(2), 303-338 (Jun 2010)", + "14. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019)", + "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + ], + "bbox": [ + 218, + 237, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth prediction (October 2019)", + "17. Goldblum, M., Souri, H., Ni, R., Shu, M., Prabhu, V., Somepalli, G., Chattopadhyay, P., Ibrahim, M., Bardes, A., Hoffman, J., et al.: Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems 36 (2024)", + "18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)", + "19. Hoyer, L., Dai, D., Chen, Y., Köring, A., Saha, S., Van Gool, L.: Three ways to improve semantic segmentation with self-supervised depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11130-11140 (2021)", + "20. Hoyer, L., Dai, D., Wang, Q., Chen, Y., Van Gool, L.: Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. arXiv preprint arXiv:2108.12545 [cs] (2021)", + "21. Huang, J., Lee, A.B., Mumford, D.: Statistics of range images. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662). vol. 1, pp. 324-331. IEEE (2000)", + "22. Huang, J., Mumford, D.: Statistics of natural images and models. In: Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149). vol. 1, pp. 541-547. IEEE (1999)", + "23. Jiang, H., Larsson, G., Shakhnarovich, M.M.G., Learned-Miller, E.: Self-supervised relative depth learning for urban scene understanding. In: Proceedings of the European conference on computer vision (eccv). pp. 19-35 (2018)", + "24. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020)", + "25. Julesz, B.: Binocular depth perception without familiarity cues: Random-dot stereo images with controlled spatial and temporal properties clarify problems in stereopsis. Science 145(3630), 356-362 (1964)", + "26. Julesz, B.: Foundations of cyclopean perception. (1971)", + "27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)", + "28. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020)", + "29. Lao, D., Hu, Z., Locatello, F., Yang, Y., Soatto, S.: Divided attention: Unsupervised multi-object discovery with contextually separated slots. arXiv preprint arXiv:2304.01430 (2023)", + "30. Lao, D., Sundaramoorthi, G.: Minimum delay moving object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4250-4259 (2017)", + "31. Lao, D., Sundaramoorthi, G.: Extending layered models to 3d motion. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Lao, D., Zhu, P., Wonka, P., Sundaramoorthi, G.: Flow-guided video inpainting with scene templates. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14599-14608 (2021)", + "33. Lasinger, K., Ranftl, R., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. arXiv preprint arXiv:1907.01341 (2019)", + "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "35. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: European Conference on Computer Vision. pp. 35-53. Springer (2022)", + "36. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020)", + "37. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)", + "38. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024), https://openreview.net/forum?id=a68SUt6zFt", + "39. Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T., Efros, A.A.: Context encoders: Feature learning by inpainting. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2536-2544 (2016)", + "40. Ramirez, P.Z., Tonioni, A., Salti, S., Stefano, L.D.: Learning across tasks and domains. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8110-8119 (2019)", + "41. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)", + "42. Saha, S., Obukhov, A., Paudel, D.P., Kanakis, M., Chen, Y., Georgoulis, S., Van Gool, L.: Learning to relate depth and semantics for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8197-8207 (2021)", + "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "44. Sundaramoorthi, G., Petersen, P., Varadarajan, V., Soatto, S.: On the set of images modulo viewpoint and contrast changes. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 832-839. IEEE (2009)", + "45. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method. arXiv preprint physics/0004057 (2000)", + "46. Wang, Y., Wu, J., Long, M., Tenenbaum, J.B.: Probabilistic video prediction from noisy data with a posterior confidence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10830-10839 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On Viability Monocular Depth Pre-train. Semantic Segmentation", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021)", + "48. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021)", + "49. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020)", + "50. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019)", + "51. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021)", + "52. Wu, Y., Liu, T.Y., Park, H., Soatto, S., Lao, D., Wong, A.: Augundo: Scaling up augmentations for monocular depth completion and estimation. In: European Conference on Computer Vision. Springer (2024)", + "53. Wu, Y., Gao, R., Park, J., Chen, Q.: Future video synthesis with object motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5539-5548 (2020)", + "54. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)", + "55. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)", + "56. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019)", + "57. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3712-3722 (2018)", + "58. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024)", + "59. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + ], + "bbox": [ + 215, + 147, + 785, + 674 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "D Lao et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_model.json b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ff28bfffd71ecaae754c1c97e43369b55ccf1d54 --- /dev/null +++ b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_model.json @@ -0,0 +1,2447 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.275, + 0.141, + 0.73, + 0.187 + ], + "angle": 0, + "content": "On the Viability of Monocular Depth Pre-training for Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.212, + 0.777, + 0.245 + ], + "angle": 0, + "content": "Dong Lao1, Fengyu Yang2, Daniel Wang2, Hyoungseob Park2, Samuel Lu1, Alex Wong2, and Stefano Soatto1" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.255, + 0.671, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) UCLA Vision Lab, Los Angeles, CA 90024, USA" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.27, + 0.672, + 0.282 + ], + "angle": 0, + "content": "{lao, soatto}@cs.ucla.edu, samuellu@ucla.edu" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.283, + 0.663, + 0.297 + ], + "angle": 0, + "content": "\\(^{2}\\) Yale Vision Lab, New Haven, CT 06511, USA" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.298, + 0.768, + 0.311 + ], + "angle": 0, + "content": "{fengyu.yang, daniel.wang.dhw33, hyoungseob.park, alex.wong}@yale.edu" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.34, + 0.744, + 0.687 + ], + "angle": 0, + "content": "Abstract. The question of whether pre-training on geometric tasks is viable for downstream transfer to semantic tasks is important for two reasons, one practical and the other scientific. If the answer is positive, we may be able to reduce pre-training costs and bias from human annotators significantly. If the answer is negative, it may shed light on the role of embodiment in the emergence of language and other cognitive functions in evolutionary history. To frame the question in a way that is testable with current means, we pre-train a model on a geometric task, and test whether that can be used to prime a notion of \"object\" that enables inference of semantics as soon as symbols (labels) are assigned. We choose monocular depth prediction as the geometric task, and semantic segmentation as the downstream semantic task, and design a collection of empirical tests by exploring different forms of supervision, training pipelines, and data sources for both depth pre-training and semantic fine-tuning. We find that monocular depth is a viable form of pre-training for semantic segmentation, validated by improvements over common baselines. Based on the findings, we propose several possible mechanisms behind the improvements, including their relation to dataset size, resolution, architecture, in/out-of-domain source data, and validate them through a wide range of ablation studies. We also find that optical flow, which at first glance may seem as good as depth prediction since it optimizes the same photometric reprojection error, is considerably less effective, as it does not explicitly aim to infer the latent structure of the scene, but rather the raw phenomenology of temporally adjacent images. Code: https://github.com/donglao/DepthToSemantic." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.7, + 0.725, + 0.714 + ], + "angle": 0, + "content": "Keywords: Depth estimation \\(\\cdot\\) semantic segmentation \\(\\cdot\\) pre-training" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.376, + 0.751 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.765, + 0.669, + 0.78 + ], + "angle": 0, + "content": "We probe the following seemingly counter-intuitive hypothesis:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.767, + 0.842 + ], + "angle": 0, + "content": "Can pre-training on a geometric task benefit a downstream semantic task? Geometric inference is often viewed as a low-level vision task requiring little abstraction that is needed for semantics [26]. For example, depth can be aquired through minimizing reprojection error, i.e. from multi-view or videos," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "directly from range sensors. Both can be performed procedurally, without inductive learning, rendering depth a meaningless task for pre-training. However, induction is needed to infer one 3D scene among infinitely many compatible with the same 2D image. Therefore, if a model could solve this ill-posed problem, it would provide evidence of the viability of pre-training with little to no human intervention, which is important in specialized data domains for which little annotated data is publicly available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.79, + 0.44 + ], + "angle": 0, + "content": "In this paper, we focus on testing monocular depth as the pre-training geometric task, and semantic segmentation as the downstream semantic task. They are purposefully chosen: Training deep neural networks for semantic segmentation requires labor-intensive pixel-level annotation, so the choice of pre-training is essential to its performance. Existing studies have shown mixed results about the relationship between the two tasks. Taskonomy [57], a framework for measuring relationships between visual tasks, suggests that depth estimation is \"far\" from semantic segmentation, while recent work [17] shows that depth pre-training can beat \"closer\" tasks like image classification. Prior work [19, 23, 40, 42] has also shown improvement in semantic segmentation when incorporating depth. This mixed evidence motivates us to take a closer look at the underlying mechanism of how monocular depth may benefit semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.79, + 0.672 + ], + "angle": 0, + "content": "Another more subtle reason for exploring this hypothesis is that pre-training is often performed on heavily human-biased datasets [11,38], where the photographer who framed the picture meant to convey a particular concept (say, a cup), and therefore took care to make sure that the manifestation of the concept (the image) prominently features the object by choice of vantage point, illumination, and (lack of) occlusion. This bias is mitigated if data is not purposefully organized into \"shots.\" Unfortunately, existing datasets are mostly composed of purposefully framed shots which could obfuscate the analysis. We note that depth can be inferred without any semantic interpretation [26] regardless of whether the data is captured purposefully or randomly. With monocular depth as the pre-training task, there are two ways of reducing the aforementioned human selective bias: The first involves directly pre-training within the specific domain of interest, leveraging the simplicity of data gathering; The second way is scaling up pre-training by incorporating diverse sources of data, which is made possible by recent developments [33,41,54] on relative depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Methods. We formalize the main hypothesis in Sect. 3. Since it cannot be tested analytically without knowledge of the joint distribution of test images and labels, we propose an empirical testing protocol. We test on monocular depth models trained under multiple forms of supervision, including structure-from-motion, binocular stereo, and depth sensors. We then change the prediction head of the resulting network, either the final layer or the whole decoder, and fine-tune it for semantic segmentation (Fig. 1). We consider depth estimation as a viable pre-training option if it yields comparable improvements to downstream semantic segmentation tasks as other common pre-training practices, e.g. ImageNet classification. To this end, we design a series of controlled experiments to test the effect of choice of initialization (Tab. 1, Fig. 2), training with various datasets" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "sizes (Fig. 3), choice of network component to be frozen and fine-tuned (Fig. 4), effect of resolution of training images (Fig. 5). Conclusions are drawn from both quantitative and qualitative (Fig. 6) results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.788, + 0.482 + ], + "angle": 0, + "content": "Findings. Pre-training for depth estimation improves the performance of downstream semantic segmentation across different experimental settings. Particularly, we show that depth estimation is indeed a viable pre-training option as compared to existing methods (Tab. 4). For example, compared to classification, using depth on average improves by \\(5.8\\%\\) mIoU and \\(5.2\\%\\) pixel accuracy on KITTI. As a sanity check, we test both a depth network pre-trained from scratch and one trained after ImageNet initialization, and both outperform classification-based pre-training in downstream semantic segmentation. To control the effect of our choice of architecture, we used our pre-trained encoder to initialize a standard semantic segmentation network [5]. We observed similar findings on Cityscapes and NYU-V2 regardless of how depth training is supervised. Inferring depth without explicit supervision typically involves minimizing the prediction error, just like optical flow. Somewhat surprisingly, not only does pre-training for depth outperform optical flow, but the latter is often worse than random initialization (Fig. 7). One may also argue that observed improvements mainly come from the availability of in-domain pre-training data for depth. To test this conjecture, we fine-tune a depth model [55] trained on large-scale out-of-domain data. Improvements in semantic segmentation reveal that when trained at scale, depth models show strong transferability to unseen downstream data domains." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.388, + 0.529 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Pre-training aims to learn a representation (function) of the test data that is maximally informative (sufficient), while providing some kind of complexity advantage. In our case, we measure complexity by the validation error after fine-tuning on limited amount of labeled data, which measures the inductive value of pretraining. The recent literature comprises a large variety of \"self-supervised\" methods that are purportedly task-agnostic. In reality, the task is specified indirectly by the choice of hand-designed nuisance transformations that leave the outcome of inference unchanged. Such transformations are sampled through data augmentation while the image identity holds constant (Contrastive Learning) [4,8,9,38], or reconstruction [3]. Group transformations organize the dataset into orbits, which contrastive learning tries to collapse onto its quotient, which is a maximal invariant. Such a maximal invariant is transferable to all and only tasks for which the chosen transformation is uninformative. For group transformation, the maximal invariant can, in theory, be computed in closed form [44]. In practice, contrastive learning are extended to non-group transformations, e.g. occlusions, as seen in language [2] and images [7]. All self-supervised methods boil down to hand-designed and quantized subsets of planar domain diffeomorphisms (discrete rotations, translations, scaling, reflections, etc.), range homeomorphisms (contrast, colormap transformations) and occlusion masks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.141, + 0.387, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.228, + 0.376, + 0.238 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.141, + 0.472, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.228, + 0.443, + 0.238 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.14, + 0.596, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.228, + 0.542, + 0.238 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.138, + 0.734, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.228, + 0.676, + 0.238 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 1: Diagram for different pre-training and fine-tuning setups. (a) Common practice: pre-train the encoder, e.g. on ImageNet, attach a decoder, and fine-tune the network. (b) Our best practice: pre-train the network by monocular depth, and fine-tune for semantic segmentation. (c) Cross architecture: for fair comparisons with common practice, we pre-train by depth, replace the decoder, and fine-tune. (d) To test the quality of pre-trained encoders, we fix the encoders and fine-tune decoders only." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.789, + 0.683 + ], + "angle": 0, + "content": "In our case, rather than hand-designing the nuisance transformations assumed to be shared among pre-training and fine-tuning tasks, we let the scene itself provide the needed supervision: images portend the same scene, either from the same timestamp (stereo) or adjacent in the temporal domain (video frames), so their variability defines the union of nuisance factors. These include domain deformations due to ego- and scene motion, range transformations due to changes in illumination, and occlusions. In addition to sharing nuisance variability, pretraining and fine-tuning tasks should ideally also share the hypothesis space. It may seem odd to choose a geometric task, where the hypothesis space is depth, to pre-train for a semantic task, where the hypothesis space is a discrete set of labels. However, due to the statistics of range images [21] and their similarity to the statistics of natural images [22], this is actually quite natural: A range map is a piecewise smooth function defined on the image domain, whereas a segmentation map is a piecewise constant function where the levels are mapped to arbitrary labels. As a result, the decoder for depth estimation can be easily modified for semantic segmentation. A discussion of this choice, specifically on the representational power of deterministic predictors, in Sect. 5. [19, 23] also utilize depth for semantic segmentation. [23] proposes pre-training on relative depth prediction, and [19, 20] utilizes self-supervised depth estimation on video sequences. Our experiments validate their findings. We further investigate whether features obtained purely from monocular depth improve semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Monocular depth [16, 50] may use different supervision, either through additional sensors [14, 51, 52], or synthetic data [36, 47, 56], but none require human annotation. Some use regularizers with sparse seeds [35, 48, 49], or adopt pretrainings [41]. We design experiments agnostic to how depth models are trained, but also make comparisons across different forms of depth supervision (Tab. 2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.781, + 0.388, + 0.797 + ], + "angle": 0, + "content": "3 Formalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Let \\( x: D \\subset \\mathbb{R}^2 \\to \\{0, \\ldots, 255\\}^3 \\) be an image, where the domain \\( D \\) is quantized into a lattice, \\( z: D \\to \\{1, \\ldots, Z\\} \\) a depth map with \\( Z \\) depth or disparity levels," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.296, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "and \\(y: D \\to \\{1, \\ldots, K\\}\\) a semantic segmentation map. In coordinates, each pixel in the lattice, \\((i,j) \\in \\{1, \\ldots, N\\} \\times \\{1, \\ldots, M\\}\\) is mapped to RGB intensities by \\(x(i,j)\\), a depth by \\(z(i,j)\\), and a label by \\(y(i,j)\\). Despite the discrete nature of the data and the hypothesis space, we relax them to the continuum by considering the vectors \\(\\mathbf{x} \\in \\mathbb{R}^{NM3}\\), \\(\\mathbf{y} \\in \\mathbb{R}^{NMK}\\) and \\(\\mathbf{z} \\in \\mathbb{R}^{NM}\\). With a slight abuse of notation, \\(y \\in \\{1, \\ldots, K\\}\\) denotes a single label and \\(\\bar{y} \\in \\mathbb{R}^K\\) its embedding, often restricted to a one-hot encoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.252, + 0.785, + 0.344 + ], + "angle": 0, + "content": "Now, consider a dataset \\( \\mathcal{D}_z = \\{\\mathbf{x}_t^i,\\mathbf{z}_t^i\\}_{i,t = 1}^{V,T_i} \\) comprised of \\( V \\) image sequences each of length \\( T_{i} \\). For the simplicity of the notations, in the case of multi-view stereo, we also consider multiple 2D image inputs as a \"sequence\" without loss of generality. In the case of supervised depth estimation, synchronized depth maps \\( z_{t}^{i} \\)'s are measured by a range sensor. Typical datasets supporting depth estimation may include just image sequences or both modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.784, + 0.374 + ], + "angle": 0, + "content": "Training for monocular depth estimation yields a mapping \\(\\phi_w: \\mathbf{x} \\mapsto \\mathbf{z}\\), parametrized by weights \\(w\\) in a neural network, via" + }, + { + "type": "equation", + "bbox": [ + 0.269, + 0.386, + 0.785, + 0.419 + ], + "angle": 0, + "content": "\\[\nw = \\arg \\min _ {w, g _ {t}} \\sum_ {i, j, n, t} \\ell \\left(x _ {t + 1} ^ {n} (i, j), \\hat {x} _ {t} ^ {n} (i, j)\\right) + \\lambda \\sum_ {i, j, n, t} \\ell \\left(z _ {t} ^ {n} (i, j), \\hat {z} _ {t} ^ {n} (i, j)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.785, + 0.548 + ], + "angle": 0, + "content": "where \\(\\hat{x}_t\\) is the warping of an image \\(x\\) from \\(t\\) to \\(t + 1\\) based on camera pose \\(g_t\\). Here we consider a generic formulation for different modalities of depth estimation. The first term in Eq. (1) measures reprojection error across frames, and the second term measures the distance between estimated depth values and the ground-truth from the range sensor. In the case of unsupervised depth estimation from videos, e.g. [16], only the reprojection loss is considered; while in the case of supervised depth estimation from single images, e.g. [37], \\(T_i = 1\\) for all \\(i\\)'s, so only the second term is minimized." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.716 + ], + "angle": 0, + "content": "The goal is to use these representations as encodings of the data to then learn a semantic segmentation map. In practice, the representations above are implemented by deep neural networks, that can be truncated at intermediate layers thus providing embedding spaces larger than the respective hypothesis spaces. We refer to the parts before and after this intermediate layer encoder and decoder, respectively. We overload the notation and refer to the encoding as \\(\\mathbf{h} = \\phi_w(\\mathbf{x})\\) for both depth estimation and other pre-training methods, presumably with weights \\(w'\\), assuming they have the same encoder architecture. The goal of semantic segmentation is then to learn a parametrized map \\(\\psi_{w''}: \\mathbf{h} \\mapsto \\mathbf{y}\\) using a small but fully supervised dataset \\(\\mathcal{D}_s = \\{\\mathbf{x}^n, \\mathbf{y}^n\\}_{n=1}^N\\), by minimizing some loss function or (pseudo-)distance in the hypothesis space \\(d(\\mathbf{y}, \\hat{\\mathbf{y}})\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.73, + 0.785, + 0.769 + ], + "angle": 0, + "content": "\\[\nw ^ {\\prime \\prime} = \\arg \\min _ {w} \\sum_ {n = 1} ^ {N} d \\left(\\mathbf {y} ^ {n}, \\psi_ {w} \\left(\\mathbf {h} ^ {n}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "plus customary regularizers. In the aggregate, we have a Markov chain: \\(\\mathbf{x} \\longrightarrow \\mathbf{h} = \\phi_w(\\mathbf{x}) \\longrightarrow \\mathbf{y} = \\psi_{w''}(\\mathbf{h}) = \\psi_{w''} \\circ \\phi_w(\\mathbf{x})\\) for depth estimation, and \\(\\mathbf{x} \\longrightarrow \\hat{\\mathbf{h}} = \\phi_{w'}(\\mathbf{x}) \\longrightarrow \\mathbf{y} = \\psi_{w''} \\circ \\phi_{w'}(\\mathbf{x})\\) for other pre-training methods. A representation obtained through a Markov chain is optimal (minimal sufficient) only if the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Semantic segmentation accuracy on KITTI. Unsupervised depth as pre-training improves semantic segmentation accuracy under all settings. Our best practice (in blue) improves common practice (in purple) by \\(7.53\\%\\) mIoU and \\(4.68\\%\\) pixel accuracy. Freezing the encoder with ImageNet pre-training (in red) is worse than no pre-training (random initialization). DeepLabV3†: with ResNet50 encoder." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.227, + 0.772, + 0.33 + ], + "angle": 0, + "content": "
Fine-tune AllFreeze Encoder
ResNet18ResNet50DeepLabV3†ResNet18ResNet50
Pre-trainingmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.Acc
None41.3570.7544.6673.3721.9352.3241.2470.5237.7267.38
ImageNet45.1572.3944.6573.0643.3972.6633.3365.3432.0362.53
Depth-Rand46.0072.4349.9076.2843.4371.3443.0272.3845.7974.71
Depth50.2076.3950.9277.3443.7772.6846.5374.4246.5574.48
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.403 + ], + "angle": 0, + "content": "intermediate variable \\(\\mathbf{h}\\) or \\(\\hat{\\mathbf{h}}\\) reduces the Information Bottleneck [45] to zero. In general, there is information loss, so we formalize the key question as whether the two Information Bottleneck Lagrangians satisfy the following:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.412, + 0.786, + 0.437 + ], + "angle": 0, + "content": "\\[\nH (\\mathbf {y} | \\mathbf {h}) + \\beta I (\\mathbf {h}; \\mathbf {x}) \\stackrel {?} {\\leq} H (\\mathbf {y} | \\hat {\\mathbf {h}}) + \\beta^ {\\prime} I (\\hat {\\mathbf {h}}; \\mathbf {x}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.537 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\beta'\\) are hyperparameters that can be optimized as part of the training process, and \\(I, H\\) denotes the (Shannon) Mutual Information and cross-entropy respectively. If the above is satisfied, then pre-training for depth estimation is a viable option, or even better than pre-training with another method. It would be ideal if this question could be settled analytically. Unfortunately, this is not possible, but the formalization above suggests a protocol to settle it empirically." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.643 + ], + "angle": 0, + "content": "To test this empirically, we use the validation error on a supervised dataset \\(\\mathcal{D}_s\\) as a proxy for residual information. We conduct fine-tuning under several configurations (Fig. 1): with respect to \\(w''\\) using \\(\\mathcal{D}_s\\), i.e. yielding a comparison of the raw pre-trained back-bones (encoders, \\(w, w'\\)), or with respect to both \\(w''\\) and \\(w\\) (for depth estimation) or \\(w'\\) (for other pre-training methods). Finally, all four resulting models can be compared with one obtained by training from scratch by optimizing a generic architecture with respect to \\(w''\\) alone." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.377, + 0.683 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.684, + 0.712 + ], + "angle": 0, + "content": "4.1 Controlled Experiments with Few-shot Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.78 + ], + "angle": 0, + "content": "We first cover an extensive collection of controlled experiments and ablations to gain insights into the main hypothesis. We specifically conduct experiments under the few-shot setting, where only a small amount of labels are available for fine-tuning, to highlight the role of pre-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "KITTI [15] contains 93000 video frames for depth training with 200 densely annotated images for semantic segmentation. Segmentation results are evaluated by mean IoU (mIoU) and pixel-level accuracy (P.Acc). We randomly choose a small training set of 16 images and limit data augmentation to horizontal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.147, + 0.428, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.147, + 0.567, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.147, + 0.707, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.229, + 0.427, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.23, + 0.566, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.23, + 0.706, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.323, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 2: Comparison between different network initializations. Models initialized by depth pre-training (unsupervised) train faster and achieve higher final accuracy." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.369, + 0.353, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.369, + 0.498, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.466, + 0.516, + 0.522 + ], + "angle": 0, + "content": "Fig. 3: Final accuracy vs different training set size. Under all training set sizes, our best practice constantly outperforms ImageNet pre-trained. Encoder: ResNet 18." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.382, + 0.744, + 0.442 + ], + "angle": 0, + "content": "
Depth Sup.mIoUP.Acc
Video46.0072.43
Stereo49.1174.58
Lidar52.7877.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.452, + 0.788, + 0.523 + ], + "angle": 0, + "content": "Table 2: Forms of depth supervision matter. Direct supervision with Lidar works the best, followed by stereo (with known camera pose), and monocular video (camera pose unknown)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.674 + ], + "angle": 0, + "content": "flips to highlight the impact of pre-training, except for Fig. 3 where we test on different training set partitions. We use Monodepth2 [16] for depth pre-training. For semantic segmentation, we replace the last layer of the decoder with a fully connected layer, using the finest scale of the multi-scale output. We test on ResNet18 and ResNet50 encoders due to their compatibility with various network architectures and widely public-available pre-trained models. Fig. 1 summarizes our experimental setup and Tab. 1 summarizes the outcomes. In all cases, depth pre-training improves segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Full model. Fig. 2 shows the evolution of training loss and model accuracy. Depth pre-training outperforms ImageNet and random initialization. ImageNet pre-training slightly improves over random initialization on ResNet18, but shows almost identical performance to random initialization on ResNet50. Depth also speeds up training, taking \\(\\sim 5000\\) iterations to converge, while ImageNet takes 15000 to 20000. Similar results on full-resolution are deferred to the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Different training set size. Fig. 3 shows that depth pre-training improves final segmentation scores over all dataset partitions (with ResNet 18). When training samples increase (e.g. 128), ImageNet and depth pre-training (from random initialization) are comparable to random initialization, but depth pretraining initialized with ImageNet yields the best results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.147, + 0.428, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.149, + 0.567, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.147, + 0.707, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.23, + 0.427, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.233, + 0.566, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.23, + 0.706, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 4: Frozen encoder results. Using an encoder pre-trained by depth significantly outperforms one with random weights and one for ImageNet classification. Note that in this experiment, ImageNet pre-training performs worse than random initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.543 + ], + "angle": 0, + "content": "Different forms of depth supervision. Pre-training quality depends on the source of supervision. Training on monocular videos involves minimizing reprojection error, which requires joint estimation of depth and pose. Since pose estimation relies on sufficiently distinctive textures (large eigenvalues of the structure tensor of image gradients), the supervision signal is sparse. Conversely, with stereo images, one may omit the pose network when training. With depth sensors, training losses minimize error w.r.t. dense or semi-dense measured depth, offering stronger supervision. Tab. 2 shows that supervising with Lidar is the best, followed by stereo, and monocular video - all improving over ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.789, + 0.73 + ], + "angle": 0, + "content": "Frozen encoder. We freeze pre-trained encoders, and fine-tune the decoder only, testing the ability of features from pre-trained encoders to capture semantics. With both ResNet18 and ResNet50, encoder pre-trained for depth significantly outperforms random initialization and ImageNet pre-trained (Fig. 4). It is surprising that ImageNet pre-training is detrimental in this case (after a grid search over learning rates): worse than fixed random weights. This suggests that while classification is a semantic task, it removes semantic information about the scene due to the object-centric bias in datasets. ImageNet pre-training tends to favor image-level features, that may not capture object shape, making finetuning the decoder difficult for segmentation. We conjecture that these uncontrolled biases in ImageNet pre-training cause difficulties in directly predicting segmentation without fine-tuning the encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Initializing with a pre-trained encoder only. To eliminate the effect of the depth-initialized decoder and only test the encoder, we replace the decoder by 'fresh' randomly initialized weights and fine-tune the whole network. Tab. 3 shows that depth pre-training outperforms ImageNet when the effect of pretraining is isolate to the encoder. This is also supported by neural activations in Fig. 6 where the regions activated after depth pre-training align well with semantic boundaries. Nonetheless, the decoder does play a role in segmentation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.296, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.148, + 0.473, + 0.282 + ], + "angle": 0, + "content": "
Pre-trainingmIoUP.Acc
ViT-1 Res. 18None41.3570.75
ImageNet45.1572.39
Depth (encoder only)46.6975.04
ViT-1 Res. 50None44.6673.37
ImageNet44.6573.06
Depth (encoder only)46.9973.57
ViT-1 Res. 18ImageNet57.5381.48
Depth (encoder only)58.1281.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.288, + 0.476, + 0.386 + ], + "angle": 0, + "content": "Table 3: Initializing with depth encoder and random decoder. Initializing with the depth encoder and a random decoder outperforms ImageNet initialization, but is worse than initializing with both encoder and decoder from the depth network (see Tab. 1)." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.145, + 0.772, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.288, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Fig.5: Mismatch between object scales. ImageNet models are trained with a fixed input size and objects of interest tend to have a similar scale. However, objects in semantic segmentation dataset vary drastically in scale. Pre-training for depth provides robustness to scale change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.784, + 0.446 + ], + "angle": 0, + "content": "accuracy: Initializing the whole network with depth pre-training still performs the best, an advantage that is not afforded by a classification head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.785, + 0.583 + ], + "angle": 0, + "content": "Results on Vision Transformers. Given that ViTs [12] necessitate extensive pre-training, conducting comprehensive ablation studies on ViTs, as feasible with ResNet, becomes impractical due to data constraints. Thus, we conduct experiments using the DPT [41] depth model trained on a collection of datasets [33] (not including KITTI), and report the best results in Tab. 3 after an exhaustive search for the optimal learning rate. Notably, we identified the optimal learning rate for this experiment to be 5e-8, with larger learning rates yielding suboptimal results. This suggests that DPT inherently provides robust representations sufficient for segmentation, requiring minimal fine-tuning in comparison to CNNs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.688 + ], + "angle": 0, + "content": "Cross architecture. To test whether depth pre-training favors our particular choice of architecture, we use the same encoders to initialize DeepLab V3 and follow the same fine-tuning procedure as common practice, i.e., ImageNet initialization. Results are presented in Tab. 1. All pre-trainings significantly improve accuracy compared with random initialization. The depth model trained from scratch provides the same level of performance as ImageNet, while subsequent depth training after ImageNet pre-training leads to further improvements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Robustness to object scales. Given a fixed resolution, \"primary\" objects in object-centric data tend to have a similar scale. For example, ImageNet models are trained on \\(224 \\times 224\\), thus cars typically have a size of 100 to 200 pixels. In KITTI, however, cars appear at different scales, varying from a few to a few hundred pixels. Fig. 5 illustrates the scale mismatch between datasets. On the other hand, depth pre-training can be done in the same domain as segmentation, hence having robustness to object scales. We examine such robustness: Pretraining on one resolution, fine-tuning on another. Since higher resolution images contain smaller scales (yellow box in Fig. 5), pre-training on them should still work on smaller images. Pre-trained on smaller resolutions, however, should work" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.774, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.296 + ], + "angle": 0, + "content": "Fig.6: Neural activation and semantic segmentation result. We visualize the neural activation map for a shallow layer of ResNet 18 encoder trained from different initializations, and their corresponding segmentation results. Boundaries are better aligned to semantic boundaries in our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.357 + ], + "angle": 0, + "content": "poorly on larger images. Our results validate this conjecture: The former achieves a final mIoU of 48.54 (ImageNet: 45.15) while the latter diverges during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.785, + 0.493 + ], + "angle": 0, + "content": "Neural Activation. We visualize the activations of the ResNet18 encoder by Grad-CAM [43], which was originally designed for classification. We modify it for segmentation by inspecting the gradient of neural response to the summation of predicted labels for pixels instead of one single class label. We visualize shallow layers (before pooling) for high spatial resolution. Fig. 6 shows neural activation maps and segmentation outputs from depth pre-training align with semantic boundaries. This confirms not just the similarities between natural and range image statistics discussed in [22], but also the bias introduced by classification, as activations of ImageNet pre-trained encoder do not resemble object boundaries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.786, + 0.719 + ], + "angle": 0, + "content": "Comparison with optical flow. One hypothesis for the effectiveness of depth pre-training is that the process leverages the statistics of natural scenes where simply-connected components of the range map often correspond to semantically consistent regions. Thus, fine-tuning simply aligns the range of two piece-wise smooth functions. We challenge this hypothesis by trying optical flow, which also exhibits a piecewise-smooth range and is obtained by minimizing the same photometric error. We train optical flow on a siamese network with two shared-weight encoders. While both optical flow and depth capture multiply-connected object boundaries (Fig. 7), using encoders pre-trained for optical flow is detrimental. We conjecture that optical flow does not capture the stable inductive bias afforded by the static component of the underlying scene. Specifically, optical flow is compatible with an underlying 3D geometry only when the scene is rigid, but rigidity is not enforced when inferring optical flow. In contrast, depth forces recognition of rigidity and discards moving objects [29-31] as outliers, which then enables isolating them, also beneficial to semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparison with other pre-training methods. In Tab. 4, for completeness, we report results on supervised pre-training by semantic segmentation on MS-COCO [34]. Unsurprisingly, pre-training on the same task with additional annotated data yields good performance. Depth estimation, despite not needing additional annotation, yields on-par performance. We also report results on masked autoencoding. We remove random rectangular regions from images, and the network aims to reconstruct the original image. Also known as ' inpainting' [1], it is considered an effective method for feature learning [39]. It yields" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.296, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.139, + 0.516, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.137, + 0.767, + 0.224 + ], + "angle": 0, + "content": "
AllFreeze
mIoU↑mIoU↑
None41.35-41.24-
Flow38.47-2.8832.19-9.05
Depth-Rand46.004.6543.021.78
Depth50.208.8546.535.29
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.234, + 0.785, + 0.318 + ], + "angle": 0, + "content": "Fig. 7: Depth helps, flow hurts. Although both are pre-trained by minimizing a photometric reconstruction error, monocular depth outperforms optical flow. This stems from the fact that inferring depth from a single image is ill-posed so the network learns inductive priors that are rich in semantics over the structures within a scene. In contrast, any discriminative features will support the correspondence search, so the flow network is not constrained to learning semantics, yielding poor fine-tuning accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.332, + 0.784, + 0.374 + ], + "angle": 0, + "content": "Table 4: Comparison with different pre-trainings. Encoder: ResNet50; Reconstruction: by inpainting randomly corrupted regions (masked autoencoding); Supervised Segmentation: trained on MS-COCO." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.386, + 0.771, + 0.446 + ], + "angle": 0, + "content": "
Pre-trainingmIoUP.AccPre-trainingmIoUP.Acc
Supervised Segmentation51.2874.88Contrastive (DINO)44.1971.36
Depth50.9277.34Optical Flow42.7271.80
Reconstruction (MAE)47.1874.16Contrastive (MOCO V2)37.0465.91
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.581 + ], + "angle": 0, + "content": "inferior performance compared with depth. We conjecture that this is due to artificial rectangular masking which does not respect the natural image statistics, while monocular depth estimation yields occluded regions that border on objects' silhouettes. We also provide results initializing the encoder with contrastive learning (MOCO V2 [9] and DINO [4]). While they also remove bias in human annotation for classification, similar to supervised classification, they are still prone to the inherent inductive bias in pre-training data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.603, + 0.627, + 0.619 + ], + "angle": 0, + "content": "4.2 Full-scale Fine-tuning on the Whole Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.658 + ], + "angle": 0, + "content": "We now extend experiments to fine-tuning on full-scale datasets, aiming to validate the findings drawn from the controlled experiments conducted on KITTI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Cityscapes [10] contains 2975 training and 500 validation images. Each image has a resolution of \\(2048 \\times 1024\\) densely labeled into 19 semantic classes. The dataset also has 20000 unlabeled stereo pairs with a disparity map, converted to depth via focal length and camera baseline. Like KITTI, Cityscapes is also an outdoor driving dataset. Here we minimize an L1 loss between depth estimates and depth computed from stereo. We modify the prediction head of DeepLabV3 to train for depth, then re-initialize the last layer of the decoder for semantic segmentation. We are unable to reproduce the original numbers. For a fair comparison, we retrained [6] using the discussed pre-training methods and finetuned under the same training protocol i.e. batch size, augmentations, schedule, etc. Tab. 5 summarizes the outcomes. We present not only the most favorable results from an extensive training process but also results from a controlled approach" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Segmentation accuracy on Cityscapes. Similar to KITTI, pre-training for depth improves segmentation accuracy. Interestingly, under the controlled settings with limited data augmentation and fewer fine-tuning epochs, the model achieves higher performance when pre-training on cropped \\(256 \\times 256\\) patches." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.213, + 0.742, + 0.314 + ], + "angle": 0, + "content": "
FullControlled
Training mIoUP.Acc.Validation mIoUP.Acc.Training mIoUP.Acc.Validation mIoUP.Acc.
None76.1095.4163.9793.7673.8295.2660.4393.07
ImageNet81.8496.4670.4195.7576.7095.7161.8093.40
Depth83.4696.8073.1795.2477.4295.8262.5793.46
Depth-cropped86.8097.4172.2295.0179.9096.2465.0994.00
" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.331, + 0.428, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.33, + 0.567, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.33, + 0.707, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.414, + 0.427, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.414, + 0.567, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.414, + 0.706, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.536 + ], + "angle": 0, + "content": "Fig. 8: Results on NYU-V2. Similar to KITTI, initializing with depth pre-trained weights trains faster and significantly improves semantic segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.597 + ], + "angle": 0, + "content": "with restricted data augmentations and fewer training iterations. Remarkably, improvements remain consistent across both scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "One may argue that since depth and semantic segmentation maps are both piece-wise smooth, adapting from depth is naturally easy if the model is aware of each pixel's relative position in the image. In order to test this statement, instead of pre-training for depth on the full image, we train depth on randomly cropped \\(256 \\times 256\\) patches, and the model has no spatial awareness of the position of a patch in the image, so depth is purely estimated by local information. This practice (Depth-cropped) surprisingly improves semantic segmentation results under controlled settings, showing that using depth as pre-training goes beyond a simple mapping from one smooth function to another. Interestingly, this approach significantly improves training accuracy in the full setting but leads to a slight reduction in validation accuracy, suggesting a potential issue of overfitting. Future research is necessary to delve into this intriguing phenomenon. Another noteworthy observation is that when pre-trained using depth data, the model exhibits superior performance with a higher initial learning rate of 0.1, as opposed to 0.01 used for ImageNet initialization. Conversely, employing an initial learning rate of 0.1 with ImageNet weights can be detrimental and may" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "result in divergence. These findings suggest that pre-training the network with depth estimation may lead to a smoother local loss landscape." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.358 + ], + "angle": 0, + "content": "NYU-V2 [37] is an indoor dataset that contains 795 densely annotated images for training and 654 for testing. There are also 407024 unannotated frames with synchronized depth images captured by a Microsoft Kinect. Since the main hypothesis is agnostic to how depth is learned, we pre-trained for depth using ground-truth as supervision. Unlike outdoor driving, which commonly features sky on top, and road and vehicles in the middle of the image with largely planar camera motion, indoor scenes are characterized by more complex layouts with 6 DoF camera motion, yielding images that are even less likely to resemble the object-centric ones commonly observed in classification datasets. This may be why initializing the model with depth pre-trained weights significantly improves semantic segmentation accuracy with both ResNet18 and ResNet50 (see Fig. 8). Note that pre-training by depth yields faster convergence, similar to KITTI." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.379, + 0.706, + 0.394 + ], + "angle": 0, + "content": "4.3 Out-of-domain Transfer from Large-scale Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.614 + ], + "angle": 0, + "content": "One may conjecture that the advantage of pre-training with monocular depth stems from training within the same domain as downstream semantic segmentation. In practice, in-domain data collection for depth pre-training is indeed ideal since it does not require human labeling. In a scientific context, we are interested in testing this conjecture by investigating the transferability of depth models to tasks outside their original domain. However, common depth datasets are considerably small in scale, making fair comparisons with popular pre-trained models, e.g. MAE [18] and DINO v2 [38] that are trained on millions or even billions of images, infeasible. Fortunately, recent methods [33,41] suggest an alternative approach for learning monocular depth by training for relative depth instead of absolute depth, allowing for the integration of multiple data sources during training. Leveraging Depth Anything [55], a depth model trained on such scaled-up mixed datasets, we fine-tune for segmentation on three out-of-domain downstream datasets: ADE20k [59], PascalVOC [13], and CityScapes [10]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The PascalVOC dataset comprises 10,582 fully annotated images for training purposes and an additional 1,449 for testing, covering a variety of 20 foreground object classes. On the other hand, ADE20k, a sizable dataset, includes 20,210 training images and 2,000 testing images across 150 different classes. The resolution of PascalVOC and ADE20k is \\(512 \\times 512\\) and \\(896 \\times 896\\) respectively. On both datasets, we try both fine-tuning the whole network and linear probing with a frozen encoder. Reported in Tab. 6, the results align with the findings on ViTs on the KITTI dataset (in Tab. 3), in which case the pre-training data also originates from out-of-domain sources. Notably, consistent improvements in semantic segmentation performance are observed across all datasets compared to the baseline initialization DINO v2 [38]. It is anticipated that the improvement achieved with a frozen encoder is more substantial than when fine-tuning the entire network, which is consistent with our previous results. This trend underscores that, when trained at a scale, depth models exhibit robust transferability to novel downstream data domains, just as other pre-training methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 6: Out-of-domain transfer with large-scale pre-trainings. We compare Depth Anything with MAE and DINO v2 for semantic segmentation, reporting results (in mIoU) on both fine-tuning (ft) and linear probing (lin.). Depth Anything improves semantic segmentation across all datasets and settings. *: with DINO v2 initialization." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.213, + 0.782, + 0.288 + ], + "angle": 0, + "content": "
Pre-training# of pre-train dataADE20kPascalVOCCityScapes
ftlin.ftlin.ftlin.
MAE [18]1.28 million53.649.0-67.6-58.4
DINO v2 [38]142 million58.147.786.586.382.771.3
Depth Anything [55]63.5 million*59.752.387.787.384.874.8
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.312, + 0.355, + 0.328 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.789, + 0.553 + ], + "angle": 0, + "content": "Inferring depth only requires multiple images (e.g. videos or multiple viewpoints) of the same scene [25] or range sensing, both do not require human-induced priors and bias, unlike semantic tasks that rely entirely on induction: We can associate a label to an image because that image has something in common with some other image, portraying a different scene, that some annotator attached a particular label to. That inductive chain has to go through the head of human annotators, who are biased in ways that cannot be easily quantified and controlled. Depth from binocular or motion imagery does not require induction and can be performed ab-ovo. Learning a monocular model from such supervision eliminates the implicit selective bias from human annotators, yet our findings validate the main hypothesis that the inductive bias learned from such a \"human-free\" process transfers well to the downstream semantic segmentation task. Of course, if different supervisions are available, either for semantic [11], segmentation [27], or both [28], we want to incorporate that information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Our hypothesis and findings are agnostic to how depth is attributed to a single image: One can perform pre-training using monocular videos, stereo, structure light, LIDAR, or even human guidance [58]. One of our pre-training minimizes the photometric reprojection error, used by many predictive and generative approaches. However, an unstructured displacement field is in general not compatible with a rigid motion. Only if this displacement field has the structure of an epipolar transformation [44] is the prediction task forced to encode the 3D structure of the scene. This may explain why video prediction is not as effective for pre-training despite many attempts [24,32,46,53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "One limitation of monocular depth estimation is that it may require a calibrated camera, so one cannot use generic videos harvested from the web. There is nothing in principle preventing us from using uncalibrated cameras, simply by adding the calibration matrix \\( K \\) to the nuisance variables. While the necessary conditions for full Euclidean reconstruction are rarely satisfied in consumer videos (for instance, they require cyclo rotation around the optical axis, not just panning and tilting), the degrees of freedom that cannot be reconstructed are moot as they do not affect the reprojection. Moreover, recent progress in training for relative depth from mixed data sources [33, 41, 55] shows the potential to unlock a virtually unlimited volume of training data, warranting future research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.296, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.177 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by ONR N00014-22-1-2252 and ARO W911NF-17-1-0304." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.203, + 0.323, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.279 + ], + "angle": 0, + "content": "1. Bertalmio, M., Sapiro, G., Caselles, V., Ballester, C.: Image inpainting. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 417-424 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.785, + 0.363 + ], + "angle": 0, + "content": "3. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.462 + ], + "angle": 0, + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.463, + 0.785, + 0.49 + ], + "angle": 0, + "content": "6. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.518 + ], + "angle": 0, + "content": "7. Chen, M., Artières, T., Denoyer, L.: Unsupervised object segmentation by redrawing. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.56 + ], + "angle": 0, + "content": "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.785, + 0.587 + ], + "angle": 0, + "content": "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.589, + 0.785, + 0.644 + ], + "angle": 0, + "content": "10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.645, + 0.785, + 0.672 + ], + "angle": 0, + "content": "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.673, + 0.785, + 0.728 + ], + "angle": 0, + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "13. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88(2), 303-338 (Jun 2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "14. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.238, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.127 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth prediction (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.23 + ], + "angle": 0, + "content": "17. Goldblum, M., Souri, H., Ni, R., Shu, M., Prabhu, V., Somepalli, G., Chattopadhyay, P., Ibrahim, M., Bardes, A., Hoffman, J., et al.: Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "19. Hoyer, L., Dai, D., Chen, Y., Köring, A., Saha, S., Van Gool, L.: Three ways to improve semantic segmentation with self-supervised depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11130-11140 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "20. Hoyer, L., Dai, D., Wang, Q., Chen, Y., Van Gool, L.: Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. arXiv preprint arXiv:2108.12545 [cs] (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "21. Huang, J., Lee, A.B., Mumford, D.: Statistics of range images. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662). vol. 1, pp. 324-331. IEEE (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "22. Huang, J., Mumford, D.: Statistics of natural images and models. In: Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149). vol. 1, pp. 541-547. IEEE (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "23. Jiang, H., Larsson, G., Shakhnarovich, M.M.G., Learned-Miller, E.: Self-supervised relative depth learning for urban scene understanding. In: Proceedings of the European conference on computer vision (eccv). pp. 19-35 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.549 + ], + "angle": 0, + "content": "24. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "25. Julesz, B.: Binocular depth perception without familiarity cues: Random-dot stereo images with controlled spatial and temporal properties clarify problems in stereopsis. Science 145(3630), 356-362 (1964)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.613, + 0.605 + ], + "angle": 0, + "content": "26. Julesz, B.: Foundations of cyclopean perception. (1971)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.659 + ], + "angle": 0, + "content": "27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.715 + ], + "angle": 0, + "content": "28. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "29. Lao, D., Hu, Z., Locatello, F., Yang, Y., Soatto, S.: Divided attention: Unsupervised multi-object discovery with contextually separated slots. arXiv preprint arXiv:2304.01430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "30. Lao, D., Sundaramoorthi, G.: Minimum delay moving object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4250-4259 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Lao, D., Sundaramoorthi, G.: Extending layered models to 3d motion. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Lao, D., Zhu, P., Wonka, P., Sundaramoorthi, G.: Flow-guided video inpainting with scene templates. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14599-14608 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "33. Lasinger, K., Ranftl, R., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. arXiv preprint arXiv:1907.01341 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.234, + 0.787, + 0.288 + ], + "angle": 0, + "content": "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.289, + 0.787, + 0.331 + ], + "angle": 0, + "content": "35. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: European Conference on Computer Vision. pp. 35-53. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "36. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.375, + 0.787, + 0.402 + ], + "angle": 0, + "content": "37. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.487 + ], + "angle": 0, + "content": "38. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024), https://openreview.net/forum?id=a68SUt6zFt" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "39. Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T., Efros, A.A.: Context encoders: Feature learning by inpainting. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2536-2544 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "40. Ramirez, P.Z., Tonioni, A., Salti, S., Stefano, L.D.: Learning across tasks and domains. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8110-8119 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.573, + 0.787, + 0.613 + ], + "angle": 0, + "content": "41. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.671 + ], + "angle": 0, + "content": "42. Saha, S., Obukhov, A., Paudel, D.P., Kanakis, M., Chen, Y., Georgoulis, S., Van Gool, L.: Learning to relate depth and semantics for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8197-8207 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.671, + 0.787, + 0.727 + ], + "angle": 0, + "content": "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.728, + 0.787, + 0.769 + ], + "angle": 0, + "content": "44. Sundaramoorthi, G., Petersen, P., Varadarajan, V., Soatto, S.: On the set of images modulo viewpoint and contrast changes. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 832-839. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "45. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method. arXiv preprint physics/0004057 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "46. Wang, Y., Wu, J., Long, M., Tenenbaum, J.B.: Probabilistic video prediction from noisy data with a posterior confidence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10830-10839 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "D Lao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "47. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "48. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "49. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "50. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "51. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "52. Wu, Y., Liu, T.Y., Park, H., Soatto, S., Lao, D., Wong, A.: Augundo: Scaling up augmentations for monocular depth completion and estimation. In: European Conference on Computer Vision. Springer (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "53. Wu, Y., Gao, R., Park, J., Chen, Q.: Future video synthesis with object motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5539-5548 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.466 + ], + "angle": 0, + "content": "54. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "55. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "56. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "57. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3712-3722 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.632 + ], + "angle": 0, + "content": "58. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "59. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_origin.pdf b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6ab77d8d99b0997a4f12d97cf3574890a8529dcb --- /dev/null +++ b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/a95034c0-2c81-4920-a8e3-13939ba6326a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:408ebb2636c3c438acc3d123cb70f5fa04c0d70c7f3ef704bf5b5a43d9edf208 +size 4106724 diff --git a/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/full.md b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1c4058c368050626bc3db2c3daa423ce478e8ec2 --- /dev/null +++ b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/full.md @@ -0,0 +1,300 @@ +# On the Viability of Monocular Depth Pre-training for Semantic Segmentation + +Dong Lao1, Fengyu Yang2, Daniel Wang2, Hyoungseob Park2, Samuel Lu1, Alex Wong2, and Stefano Soatto1 + +$^{1}$ UCLA Vision Lab, Los Angeles, CA 90024, USA + +{lao, soatto}@cs.ucla.edu, samuellu@ucla.edu + +$^{2}$ Yale Vision Lab, New Haven, CT 06511, USA + +{fengyu.yang, daniel.wang.dhw33, hyoungseob.park, alex.wong}@yale.edu + +Abstract. The question of whether pre-training on geometric tasks is viable for downstream transfer to semantic tasks is important for two reasons, one practical and the other scientific. If the answer is positive, we may be able to reduce pre-training costs and bias from human annotators significantly. If the answer is negative, it may shed light on the role of embodiment in the emergence of language and other cognitive functions in evolutionary history. To frame the question in a way that is testable with current means, we pre-train a model on a geometric task, and test whether that can be used to prime a notion of "object" that enables inference of semantics as soon as symbols (labels) are assigned. We choose monocular depth prediction as the geometric task, and semantic segmentation as the downstream semantic task, and design a collection of empirical tests by exploring different forms of supervision, training pipelines, and data sources for both depth pre-training and semantic fine-tuning. We find that monocular depth is a viable form of pre-training for semantic segmentation, validated by improvements over common baselines. Based on the findings, we propose several possible mechanisms behind the improvements, including their relation to dataset size, resolution, architecture, in/out-of-domain source data, and validate them through a wide range of ablation studies. We also find that optical flow, which at first glance may seem as good as depth prediction since it optimizes the same photometric reprojection error, is considerably less effective, as it does not explicitly aim to infer the latent structure of the scene, but rather the raw phenomenology of temporally adjacent images. Code: https://github.com/donglao/DepthToSemantic. + +Keywords: Depth estimation $\cdot$ semantic segmentation $\cdot$ pre-training + +# 1 Introduction + +We probe the following seemingly counter-intuitive hypothesis: + +Can pre-training on a geometric task benefit a downstream semantic task? Geometric inference is often viewed as a low-level vision task requiring little abstraction that is needed for semantics [26]. For example, depth can be aquired through minimizing reprojection error, i.e. from multi-view or videos, + +directly from range sensors. Both can be performed procedurally, without inductive learning, rendering depth a meaningless task for pre-training. However, induction is needed to infer one 3D scene among infinitely many compatible with the same 2D image. Therefore, if a model could solve this ill-posed problem, it would provide evidence of the viability of pre-training with little to no human intervention, which is important in specialized data domains for which little annotated data is publicly available. + +In this paper, we focus on testing monocular depth as the pre-training geometric task, and semantic segmentation as the downstream semantic task. They are purposefully chosen: Training deep neural networks for semantic segmentation requires labor-intensive pixel-level annotation, so the choice of pre-training is essential to its performance. Existing studies have shown mixed results about the relationship between the two tasks. Taskonomy [57], a framework for measuring relationships between visual tasks, suggests that depth estimation is "far" from semantic segmentation, while recent work [17] shows that depth pre-training can beat "closer" tasks like image classification. Prior work [19, 23, 40, 42] has also shown improvement in semantic segmentation when incorporating depth. This mixed evidence motivates us to take a closer look at the underlying mechanism of how monocular depth may benefit semantic segmentation. + +Another more subtle reason for exploring this hypothesis is that pre-training is often performed on heavily human-biased datasets [11,38], where the photographer who framed the picture meant to convey a particular concept (say, a cup), and therefore took care to make sure that the manifestation of the concept (the image) prominently features the object by choice of vantage point, illumination, and (lack of) occlusion. This bias is mitigated if data is not purposefully organized into "shots." Unfortunately, existing datasets are mostly composed of purposefully framed shots which could obfuscate the analysis. We note that depth can be inferred without any semantic interpretation [26] regardless of whether the data is captured purposefully or randomly. With monocular depth as the pre-training task, there are two ways of reducing the aforementioned human selective bias: The first involves directly pre-training within the specific domain of interest, leveraging the simplicity of data gathering; The second way is scaling up pre-training by incorporating diverse sources of data, which is made possible by recent developments [33,41,54] on relative depth estimation. + +Methods. We formalize the main hypothesis in Sect. 3. Since it cannot be tested analytically without knowledge of the joint distribution of test images and labels, we propose an empirical testing protocol. We test on monocular depth models trained under multiple forms of supervision, including structure-from-motion, binocular stereo, and depth sensors. We then change the prediction head of the resulting network, either the final layer or the whole decoder, and fine-tune it for semantic segmentation (Fig. 1). We consider depth estimation as a viable pre-training option if it yields comparable improvements to downstream semantic segmentation tasks as other common pre-training practices, e.g. ImageNet classification. To this end, we design a series of controlled experiments to test the effect of choice of initialization (Tab. 1, Fig. 2), training with various datasets + +sizes (Fig. 3), choice of network component to be frozen and fine-tuned (Fig. 4), effect of resolution of training images (Fig. 5). Conclusions are drawn from both quantitative and qualitative (Fig. 6) results. + +Findings. Pre-training for depth estimation improves the performance of downstream semantic segmentation across different experimental settings. Particularly, we show that depth estimation is indeed a viable pre-training option as compared to existing methods (Tab. 4). For example, compared to classification, using depth on average improves by $5.8\%$ mIoU and $5.2\%$ pixel accuracy on KITTI. As a sanity check, we test both a depth network pre-trained from scratch and one trained after ImageNet initialization, and both outperform classification-based pre-training in downstream semantic segmentation. To control the effect of our choice of architecture, we used our pre-trained encoder to initialize a standard semantic segmentation network [5]. We observed similar findings on Cityscapes and NYU-V2 regardless of how depth training is supervised. Inferring depth without explicit supervision typically involves minimizing the prediction error, just like optical flow. Somewhat surprisingly, not only does pre-training for depth outperform optical flow, but the latter is often worse than random initialization (Fig. 7). One may also argue that observed improvements mainly come from the availability of in-domain pre-training data for depth. To test this conjecture, we fine-tune a depth model [55] trained on large-scale out-of-domain data. Improvements in semantic segmentation reveal that when trained at scale, depth models show strong transferability to unseen downstream data domains. + +# 2 Related Work + +Pre-training aims to learn a representation (function) of the test data that is maximally informative (sufficient), while providing some kind of complexity advantage. In our case, we measure complexity by the validation error after fine-tuning on limited amount of labeled data, which measures the inductive value of pretraining. The recent literature comprises a large variety of "self-supervised" methods that are purportedly task-agnostic. In reality, the task is specified indirectly by the choice of hand-designed nuisance transformations that leave the outcome of inference unchanged. Such transformations are sampled through data augmentation while the image identity holds constant (Contrastive Learning) [4,8,9,38], or reconstruction [3]. Group transformations organize the dataset into orbits, which contrastive learning tries to collapse onto its quotient, which is a maximal invariant. Such a maximal invariant is transferable to all and only tasks for which the chosen transformation is uninformative. For group transformation, the maximal invariant can, in theory, be computed in closed form [44]. In practice, contrastive learning are extended to non-group transformations, e.g. occlusions, as seen in language [2] and images [7]. All self-supervised methods boil down to hand-designed and quantized subsets of planar domain diffeomorphisms (discrete rotations, translations, scaling, reflections, etc.), range homeomorphisms (contrast, colormap transformations) and occlusion masks. + +![](images/6fcc8480d5bd6c2eb105573169c6b1000ee380fcd9a85ba87b3fe89a1323e831.jpg) +(a) Baseline + +![](images/f63a1edecb0fd2377ec7220b4020b2d0b6fa897fa5bb1a55cd9d2ac497d7933b.jpg) +(b) +Fig. 1: Diagram for different pre-training and fine-tuning setups. (a) Common practice: pre-train the encoder, e.g. on ImageNet, attach a decoder, and fine-tune the network. (b) Our best practice: pre-train the network by monocular depth, and fine-tune for semantic segmentation. (c) Cross architecture: for fair comparisons with common practice, we pre-train by depth, replace the decoder, and fine-tune. (d) To test the quality of pre-trained encoders, we fix the encoders and fine-tune decoders only. + +![](images/eed0a2744226ecb4a62d10439189c1b3ba6250eda358a00f71453adfaee1c56a.jpg) +(c) + +![](images/e2bd894af91c3ce860bc89c8d11b734c71cc5a2f5d9f94a2a7993a925c30f1fc.jpg) +(d) + +In our case, rather than hand-designing the nuisance transformations assumed to be shared among pre-training and fine-tuning tasks, we let the scene itself provide the needed supervision: images portend the same scene, either from the same timestamp (stereo) or adjacent in the temporal domain (video frames), so their variability defines the union of nuisance factors. These include domain deformations due to ego- and scene motion, range transformations due to changes in illumination, and occlusions. In addition to sharing nuisance variability, pretraining and fine-tuning tasks should ideally also share the hypothesis space. It may seem odd to choose a geometric task, where the hypothesis space is depth, to pre-train for a semantic task, where the hypothesis space is a discrete set of labels. However, due to the statistics of range images [21] and their similarity to the statistics of natural images [22], this is actually quite natural: A range map is a piecewise smooth function defined on the image domain, whereas a segmentation map is a piecewise constant function where the levels are mapped to arbitrary labels. As a result, the decoder for depth estimation can be easily modified for semantic segmentation. A discussion of this choice, specifically on the representational power of deterministic predictors, in Sect. 5. [19, 23] also utilize depth for semantic segmentation. [23] proposes pre-training on relative depth prediction, and [19, 20] utilizes self-supervised depth estimation on video sequences. Our experiments validate their findings. We further investigate whether features obtained purely from monocular depth improve semantic segmentation. + +Monocular depth [16, 50] may use different supervision, either through additional sensors [14, 51, 52], or synthetic data [36, 47, 56], but none require human annotation. Some use regularizers with sparse seeds [35, 48, 49], or adopt pretrainings [41]. We design experiments agnostic to how depth models are trained, but also make comparisons across different forms of depth supervision (Tab. 2). + +# 3 Formalization + +Let $x: D \subset \mathbb{R}^2 \to \{0, \ldots, 255\}^3$ be an image, where the domain $D$ is quantized into a lattice, $z: D \to \{1, \ldots, Z\}$ a depth map with $Z$ depth or disparity levels, + +and $y: D \to \{1, \ldots, K\}$ a semantic segmentation map. In coordinates, each pixel in the lattice, $(i,j) \in \{1, \ldots, N\} \times \{1, \ldots, M\}$ is mapped to RGB intensities by $x(i,j)$ , a depth by $z(i,j)$ , and a label by $y(i,j)$ . Despite the discrete nature of the data and the hypothesis space, we relax them to the continuum by considering the vectors $\mathbf{x} \in \mathbb{R}^{NM3}$ , $\mathbf{y} \in \mathbb{R}^{NMK}$ and $\mathbf{z} \in \mathbb{R}^{NM}$ . With a slight abuse of notation, $y \in \{1, \ldots, K\}$ denotes a single label and $\bar{y} \in \mathbb{R}^K$ its embedding, often restricted to a one-hot encoding. + +Now, consider a dataset $\mathcal{D}_z = \{\mathbf{x}_t^i,\mathbf{z}_t^i\}_{i,t = 1}^{V,T_i}$ comprised of $V$ image sequences each of length $T_{i}$ . For the simplicity of the notations, in the case of multi-view stereo, we also consider multiple 2D image inputs as a "sequence" without loss of generality. In the case of supervised depth estimation, synchronized depth maps $z_{t}^{i}$ 's are measured by a range sensor. Typical datasets supporting depth estimation may include just image sequences or both modalities. + +Training for monocular depth estimation yields a mapping $\phi_w: \mathbf{x} \mapsto \mathbf{z}$ , parametrized by weights $w$ in a neural network, via + +$$ +w = \arg \min _ {w, g _ {t}} \sum_ {i, j, n, t} \ell \left(x _ {t + 1} ^ {n} (i, j), \hat {x} _ {t} ^ {n} (i, j)\right) + \lambda \sum_ {i, j, n, t} \ell \left(z _ {t} ^ {n} (i, j), \hat {z} _ {t} ^ {n} (i, j)\right) \tag {1} +$$ + +where $\hat{x}_t$ is the warping of an image $x$ from $t$ to $t + 1$ based on camera pose $g_t$ . Here we consider a generic formulation for different modalities of depth estimation. The first term in Eq. (1) measures reprojection error across frames, and the second term measures the distance between estimated depth values and the ground-truth from the range sensor. In the case of unsupervised depth estimation from videos, e.g. [16], only the reprojection loss is considered; while in the case of supervised depth estimation from single images, e.g. [37], $T_i = 1$ for all $i$ 's, so only the second term is minimized. + +The goal is to use these representations as encodings of the data to then learn a semantic segmentation map. In practice, the representations above are implemented by deep neural networks, that can be truncated at intermediate layers thus providing embedding spaces larger than the respective hypothesis spaces. We refer to the parts before and after this intermediate layer encoder and decoder, respectively. We overload the notation and refer to the encoding as $\mathbf{h} = \phi_w(\mathbf{x})$ for both depth estimation and other pre-training methods, presumably with weights $w'$ , assuming they have the same encoder architecture. The goal of semantic segmentation is then to learn a parametrized map $\psi_{w''}: \mathbf{h} \mapsto \mathbf{y}$ using a small but fully supervised dataset $\mathcal{D}_s = \{\mathbf{x}^n, \mathbf{y}^n\}_{n=1}^N$ , by minimizing some loss function or (pseudo-)distance in the hypothesis space $d(\mathbf{y}, \hat{\mathbf{y}})$ , where + +$$ +w ^ {\prime \prime} = \arg \min _ {w} \sum_ {n = 1} ^ {N} d \left(\mathbf {y} ^ {n}, \psi_ {w} \left(\mathbf {h} ^ {n}\right)\right) \tag {2} +$$ + +plus customary regularizers. In the aggregate, we have a Markov chain: $\mathbf{x} \longrightarrow \mathbf{h} = \phi_w(\mathbf{x}) \longrightarrow \mathbf{y} = \psi_{w''}(\mathbf{h}) = \psi_{w''} \circ \phi_w(\mathbf{x})$ for depth estimation, and $\mathbf{x} \longrightarrow \hat{\mathbf{h}} = \phi_{w'}(\mathbf{x}) \longrightarrow \mathbf{y} = \psi_{w''} \circ \phi_{w'}(\mathbf{x})$ for other pre-training methods. A representation obtained through a Markov chain is optimal (minimal sufficient) only if the + +Table 1: Semantic segmentation accuracy on KITTI. Unsupervised depth as pre-training improves semantic segmentation accuracy under all settings. Our best practice (in blue) improves common practice (in purple) by $7.53\%$ mIoU and $4.68\%$ pixel accuracy. Freezing the encoder with ImageNet pre-training (in red) is worse than no pre-training (random initialization). DeepLabV3†: with ResNet50 encoder. + +
Fine-tune AllFreeze Encoder
ResNet18ResNet50DeepLabV3†ResNet18ResNet50
Pre-trainingmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.Acc
None41.3570.7544.6673.3721.9352.3241.2470.5237.7267.38
ImageNet45.1572.3944.6573.0643.3972.6633.3365.3432.0362.53
Depth-Rand46.0072.4349.9076.2843.4371.3443.0272.3845.7974.71
Depth50.2076.3950.9277.3443.7772.6846.5374.4246.5574.48
+ +intermediate variable $\mathbf{h}$ or $\hat{\mathbf{h}}$ reduces the Information Bottleneck [45] to zero. In general, there is information loss, so we formalize the key question as whether the two Information Bottleneck Lagrangians satisfy the following: + +$$ +H (\mathbf {y} | \mathbf {h}) + \beta I (\mathbf {h}; \mathbf {x}) \stackrel {?} {\leq} H (\mathbf {y} | \hat {\mathbf {h}}) + \beta^ {\prime} I (\hat {\mathbf {h}}; \mathbf {x}) \tag {3} +$$ + +where $\beta$ and $\beta'$ are hyperparameters that can be optimized as part of the training process, and $I, H$ denotes the (Shannon) Mutual Information and cross-entropy respectively. If the above is satisfied, then pre-training for depth estimation is a viable option, or even better than pre-training with another method. It would be ideal if this question could be settled analytically. Unfortunately, this is not possible, but the formalization above suggests a protocol to settle it empirically. + +To test this empirically, we use the validation error on a supervised dataset $\mathcal{D}_s$ as a proxy for residual information. We conduct fine-tuning under several configurations (Fig. 1): with respect to $w''$ using $\mathcal{D}_s$ , i.e. yielding a comparison of the raw pre-trained back-bones (encoders, $w, w'$ ), or with respect to both $w''$ and $w$ (for depth estimation) or $w'$ (for other pre-training methods). Finally, all four resulting models can be compared with one obtained by training from scratch by optimizing a generic architecture with respect to $w''$ alone. + +# 4 Experiments + +# 4.1 Controlled Experiments with Few-shot Fine-tuning + +We first cover an extensive collection of controlled experiments and ablations to gain insights into the main hypothesis. We specifically conduct experiments under the few-shot setting, where only a small amount of labels are available for fine-tuning, to highlight the role of pre-training. + +KITTI [15] contains 93000 video frames for depth training with 200 densely annotated images for semantic segmentation. Segmentation results are evaluated by mean IoU (mIoU) and pixel-level accuracy (P.Acc). We randomly choose a small training set of 16 images and limit data augmentation to horizontal + +![](images/63ed67334ad45839306c3a0fcd33015c518c5652aa2830b79a9a7bb565dc3a1b.jpg) + +![](images/a2009dc9ac4121e32e212ab1db1bc7ad4bc35de0882060e3dbcbc35a2bfcc203.jpg) + +![](images/708f8fe632a1972f66d4f0e7b1fae7780ed23cf248bd6ebe8ac1a8cc9f57badb.jpg) + +![](images/5b57aa0b486123b48c5444ef9073ebad1c168801ee98a5f5ce79684c96e982bb.jpg) +Fig. 2: Comparison between different network initializations. Models initialized by depth pre-training (unsupervised) train faster and achieve higher final accuracy. + +![](images/1eb239e452bfacfa0deab7913bd57fec3a168e8c42b0fa1a728546a39730c224.jpg) + +![](images/3610a031aa4f0fcaba0e6933c6a9ff73af11574dd04dd27df45b8d802b10385b.jpg) + +![](images/2f3c3dcb5e590a9164029753c969b39c88c670981451ee881273f926b6bb44ae.jpg) +Fig. 3: Final accuracy vs different training set size. Under all training set sizes, our best practice constantly outperforms ImageNet pre-trained. Encoder: ResNet 18. + +![](images/f988e7db50399416bf7c87d3dd26a68c4b1f394f3696e2c1f2d1f9161fd8e30a.jpg) + +
Depth Sup.mIoUP.Acc
Video46.0072.43
Stereo49.1174.58
Lidar52.7877.17
+ +Table 2: Forms of depth supervision matter. Direct supervision with Lidar works the best, followed by stereo (with known camera pose), and monocular video (camera pose unknown). + +flips to highlight the impact of pre-training, except for Fig. 3 where we test on different training set partitions. We use Monodepth2 [16] for depth pre-training. For semantic segmentation, we replace the last layer of the decoder with a fully connected layer, using the finest scale of the multi-scale output. We test on ResNet18 and ResNet50 encoders due to their compatibility with various network architectures and widely public-available pre-trained models. Fig. 1 summarizes our experimental setup and Tab. 1 summarizes the outcomes. In all cases, depth pre-training improves segmentation accuracy. + +Full model. Fig. 2 shows the evolution of training loss and model accuracy. Depth pre-training outperforms ImageNet and random initialization. ImageNet pre-training slightly improves over random initialization on ResNet18, but shows almost identical performance to random initialization on ResNet50. Depth also speeds up training, taking $\sim 5000$ iterations to converge, while ImageNet takes 15000 to 20000. Similar results on full-resolution are deferred to the Supp. Mat. + +Different training set size. Fig. 3 shows that depth pre-training improves final segmentation scores over all dataset partitions (with ResNet 18). When training samples increase (e.g. 128), ImageNet and depth pre-training (from random initialization) are comparable to random initialization, but depth pretraining initialized with ImageNet yields the best results. + +![](images/7989ba00cd01927a57550e824864c98edf572490419520ab6a170496b9e03d58.jpg) + +![](images/7859699f2b2dd1c5c74326a3051db54ef5e56edec978bccc8bc250f41d11a2ff.jpg) + +![](images/27ca382456b6d0cd7090e57af860a5f4ea2ab26f137ca9238f5f68a24f71e8df.jpg) + +![](images/eb00436f24d07ccbcb894ac30efe77179fb4a6a1ba28b7fa8f78b016b60ac088.jpg) +Fig. 4: Frozen encoder results. Using an encoder pre-trained by depth significantly outperforms one with random weights and one for ImageNet classification. Note that in this experiment, ImageNet pre-training performs worse than random initialization. + +![](images/18f43b5f8d558ab010a601bfabcf3977a52c16938b73db7e8f483f78a4647502.jpg) + +![](images/2215ad626162615c9f57b4aed9c67f819e705da8ddc1743bb6a457f79c316238.jpg) + +Different forms of depth supervision. Pre-training quality depends on the source of supervision. Training on monocular videos involves minimizing reprojection error, which requires joint estimation of depth and pose. Since pose estimation relies on sufficiently distinctive textures (large eigenvalues of the structure tensor of image gradients), the supervision signal is sparse. Conversely, with stereo images, one may omit the pose network when training. With depth sensors, training losses minimize error w.r.t. dense or semi-dense measured depth, offering stronger supervision. Tab. 2 shows that supervising with Lidar is the best, followed by stereo, and monocular video - all improving over ImageNet. + +Frozen encoder. We freeze pre-trained encoders, and fine-tune the decoder only, testing the ability of features from pre-trained encoders to capture semantics. With both ResNet18 and ResNet50, encoder pre-trained for depth significantly outperforms random initialization and ImageNet pre-trained (Fig. 4). It is surprising that ImageNet pre-training is detrimental in this case (after a grid search over learning rates): worse than fixed random weights. This suggests that while classification is a semantic task, it removes semantic information about the scene due to the object-centric bias in datasets. ImageNet pre-training tends to favor image-level features, that may not capture object shape, making finetuning the decoder difficult for segmentation. We conjecture that these uncontrolled biases in ImageNet pre-training cause difficulties in directly predicting segmentation without fine-tuning the encoder. + +Initializing with a pre-trained encoder only. To eliminate the effect of the depth-initialized decoder and only test the encoder, we replace the decoder by 'fresh' randomly initialized weights and fine-tune the whole network. Tab. 3 shows that depth pre-training outperforms ImageNet when the effect of pretraining is isolate to the encoder. This is also supported by neural activations in Fig. 6 where the regions activated after depth pre-training align well with semantic boundaries. Nonetheless, the decoder does play a role in segmentation + +
Pre-trainingmIoUP.Acc
ViT-1 Res. 18None41.3570.75
ImageNet45.1572.39
Depth (encoder only)46.6975.04
ViT-1 Res. 50None44.6673.37
ImageNet44.6573.06
Depth (encoder only)46.9973.57
ViT-1 Res. 18ImageNet57.5381.48
Depth (encoder only)58.1281.94
+ +Table 3: Initializing with depth encoder and random decoder. Initializing with the depth encoder and a random decoder outperforms ImageNet initialization, but is worse than initializing with both encoder and decoder from the depth network (see Tab. 1). + +![](images/ff8224c0f29b407897acb4ca4d8827419e6d3e6ab9ed2a1f3e239dfc142967e0.jpg) +Fig.5: Mismatch between object scales. ImageNet models are trained with a fixed input size and objects of interest tend to have a similar scale. However, objects in semantic segmentation dataset vary drastically in scale. Pre-training for depth provides robustness to scale change. + +accuracy: Initializing the whole network with depth pre-training still performs the best, an advantage that is not afforded by a classification head. + +Results on Vision Transformers. Given that ViTs [12] necessitate extensive pre-training, conducting comprehensive ablation studies on ViTs, as feasible with ResNet, becomes impractical due to data constraints. Thus, we conduct experiments using the DPT [41] depth model trained on a collection of datasets [33] (not including KITTI), and report the best results in Tab. 3 after an exhaustive search for the optimal learning rate. Notably, we identified the optimal learning rate for this experiment to be 5e-8, with larger learning rates yielding suboptimal results. This suggests that DPT inherently provides robust representations sufficient for segmentation, requiring minimal fine-tuning in comparison to CNNs. + +Cross architecture. To test whether depth pre-training favors our particular choice of architecture, we use the same encoders to initialize DeepLab V3 and follow the same fine-tuning procedure as common practice, i.e., ImageNet initialization. Results are presented in Tab. 1. All pre-trainings significantly improve accuracy compared with random initialization. The depth model trained from scratch provides the same level of performance as ImageNet, while subsequent depth training after ImageNet pre-training leads to further improvements. + +Robustness to object scales. Given a fixed resolution, "primary" objects in object-centric data tend to have a similar scale. For example, ImageNet models are trained on $224 \times 224$ , thus cars typically have a size of 100 to 200 pixels. In KITTI, however, cars appear at different scales, varying from a few to a few hundred pixels. Fig. 5 illustrates the scale mismatch between datasets. On the other hand, depth pre-training can be done in the same domain as segmentation, hence having robustness to object scales. We examine such robustness: Pretraining on one resolution, fine-tuning on another. Since higher resolution images contain smaller scales (yellow box in Fig. 5), pre-training on them should still work on smaller images. Pre-trained on smaller resolutions, however, should work + +![](images/eee26b152085dbc0595ff981a9c3c25416c4e113fd32030e2f446c4e1bdf9fb1.jpg) +Fig.6: Neural activation and semantic segmentation result. We visualize the neural activation map for a shallow layer of ResNet 18 encoder trained from different initializations, and their corresponding segmentation results. Boundaries are better aligned to semantic boundaries in our model. + +poorly on larger images. Our results validate this conjecture: The former achieves a final mIoU of 48.54 (ImageNet: 45.15) while the latter diverges during training. + +Neural Activation. We visualize the activations of the ResNet18 encoder by Grad-CAM [43], which was originally designed for classification. We modify it for segmentation by inspecting the gradient of neural response to the summation of predicted labels for pixels instead of one single class label. We visualize shallow layers (before pooling) for high spatial resolution. Fig. 6 shows neural activation maps and segmentation outputs from depth pre-training align with semantic boundaries. This confirms not just the similarities between natural and range image statistics discussed in [22], but also the bias introduced by classification, as activations of ImageNet pre-trained encoder do not resemble object boundaries. + +Comparison with optical flow. One hypothesis for the effectiveness of depth pre-training is that the process leverages the statistics of natural scenes where simply-connected components of the range map often correspond to semantically consistent regions. Thus, fine-tuning simply aligns the range of two piece-wise smooth functions. We challenge this hypothesis by trying optical flow, which also exhibits a piecewise-smooth range and is obtained by minimizing the same photometric error. We train optical flow on a siamese network with two shared-weight encoders. While both optical flow and depth capture multiply-connected object boundaries (Fig. 7), using encoders pre-trained for optical flow is detrimental. We conjecture that optical flow does not capture the stable inductive bias afforded by the static component of the underlying scene. Specifically, optical flow is compatible with an underlying 3D geometry only when the scene is rigid, but rigidity is not enforced when inferring optical flow. In contrast, depth forces recognition of rigidity and discards moving objects [29-31] as outliers, which then enables isolating them, also beneficial to semantic segmentation. + +Comparison with other pre-training methods. In Tab. 4, for completeness, we report results on supervised pre-training by semantic segmentation on MS-COCO [34]. Unsurprisingly, pre-training on the same task with additional annotated data yields good performance. Depth estimation, despite not needing additional annotation, yields on-par performance. We also report results on masked autoencoding. We remove random rectangular regions from images, and the network aims to reconstruct the original image. Also known as ' inpainting' [1], it is considered an effective method for feature learning [39]. It yields + +![](images/9df5195851d1c662822c6e7075ce10b2f892508de1a2e1aa6866c8f876ac92f9.jpg) +Fig. 7: Depth helps, flow hurts. Although both are pre-trained by minimizing a photometric reconstruction error, monocular depth outperforms optical flow. This stems from the fact that inferring depth from a single image is ill-posed so the network learns inductive priors that are rich in semantics over the structures within a scene. In contrast, any discriminative features will support the correspondence search, so the flow network is not constrained to learning semantics, yielding poor fine-tuning accuracy. + +
AllFreeze
mIoU↑mIoU↑
None41.35-41.24-
Flow38.47-2.8832.19-9.05
Depth-Rand46.004.6543.021.78
Depth50.208.8546.535.29
+ +Table 4: Comparison with different pre-trainings. Encoder: ResNet50; Reconstruction: by inpainting randomly corrupted regions (masked autoencoding); Supervised Segmentation: trained on MS-COCO. + +
Pre-trainingmIoUP.AccPre-trainingmIoUP.Acc
Supervised Segmentation51.2874.88Contrastive (DINO)44.1971.36
Depth50.9277.34Optical Flow42.7271.80
Reconstruction (MAE)47.1874.16Contrastive (MOCO V2)37.0465.91
+ +inferior performance compared with depth. We conjecture that this is due to artificial rectangular masking which does not respect the natural image statistics, while monocular depth estimation yields occluded regions that border on objects' silhouettes. We also provide results initializing the encoder with contrastive learning (MOCO V2 [9] and DINO [4]). While they also remove bias in human annotation for classification, similar to supervised classification, they are still prone to the inherent inductive bias in pre-training data. + +# 4.2 Full-scale Fine-tuning on the Whole Dataset + +We now extend experiments to fine-tuning on full-scale datasets, aiming to validate the findings drawn from the controlled experiments conducted on KITTI. + +Cityscapes [10] contains 2975 training and 500 validation images. Each image has a resolution of $2048 \times 1024$ densely labeled into 19 semantic classes. The dataset also has 20000 unlabeled stereo pairs with a disparity map, converted to depth via focal length and camera baseline. Like KITTI, Cityscapes is also an outdoor driving dataset. Here we minimize an L1 loss between depth estimates and depth computed from stereo. We modify the prediction head of DeepLabV3 to train for depth, then re-initialize the last layer of the decoder for semantic segmentation. We are unable to reproduce the original numbers. For a fair comparison, we retrained [6] using the discussed pre-training methods and finetuned under the same training protocol i.e. batch size, augmentations, schedule, etc. Tab. 5 summarizes the outcomes. We present not only the most favorable results from an extensive training process but also results from a controlled approach + +Table 5: Segmentation accuracy on Cityscapes. Similar to KITTI, pre-training for depth improves segmentation accuracy. Interestingly, under the controlled settings with limited data augmentation and fewer fine-tuning epochs, the model achieves higher performance when pre-training on cropped $256 \times 256$ patches. + +
FullControlled
Training mIoUP.Acc.Validation mIoUP.Acc.Training mIoUP.Acc.Validation mIoUP.Acc.
None76.1095.4163.9793.7673.8295.2660.4393.07
ImageNet81.8496.4670.4195.7576.7095.7161.8093.40
Depth83.4696.8073.1795.2477.4295.8262.5793.46
Depth-cropped86.8097.4172.2295.0179.9096.2465.0994.00
+ +![](images/d182751251ca513e3f0bbac454397f6dcd95b2a6ddf6f5bbb99c3572735718bf.jpg) + +![](images/4c04280ee8bdbcc57a030909706e06060d6b1e9dbe20301af63685ff7d8299cf.jpg) + +![](images/7283be2ba11d177f7d65de6364b462a752685b4ce1cd8dc61e75baa6f9ecbc3c.jpg) + +![](images/aa256eaaafae5a18d3feaeb31da492f1ec839e5bb93d98f4c12e5685d519c912.jpg) +Fig. 8: Results on NYU-V2. Similar to KITTI, initializing with depth pre-trained weights trains faster and significantly improves semantic segmentation accuracy. + +![](images/3569bf16a9c6073a3cceb31301edae89cf51e6cf15f321966f2d7ebe15d56dfb.jpg) + +![](images/6b834b614406ec5fb752514a7fda4cd82b2307f15e6bf88d73676886be64362b.jpg) + +with restricted data augmentations and fewer training iterations. Remarkably, improvements remain consistent across both scenarios. + +One may argue that since depth and semantic segmentation maps are both piece-wise smooth, adapting from depth is naturally easy if the model is aware of each pixel's relative position in the image. In order to test this statement, instead of pre-training for depth on the full image, we train depth on randomly cropped $256 \times 256$ patches, and the model has no spatial awareness of the position of a patch in the image, so depth is purely estimated by local information. This practice (Depth-cropped) surprisingly improves semantic segmentation results under controlled settings, showing that using depth as pre-training goes beyond a simple mapping from one smooth function to another. Interestingly, this approach significantly improves training accuracy in the full setting but leads to a slight reduction in validation accuracy, suggesting a potential issue of overfitting. Future research is necessary to delve into this intriguing phenomenon. Another noteworthy observation is that when pre-trained using depth data, the model exhibits superior performance with a higher initial learning rate of 0.1, as opposed to 0.01 used for ImageNet initialization. Conversely, employing an initial learning rate of 0.1 with ImageNet weights can be detrimental and may + +result in divergence. These findings suggest that pre-training the network with depth estimation may lead to a smoother local loss landscape. + +NYU-V2 [37] is an indoor dataset that contains 795 densely annotated images for training and 654 for testing. There are also 407024 unannotated frames with synchronized depth images captured by a Microsoft Kinect. Since the main hypothesis is agnostic to how depth is learned, we pre-trained for depth using ground-truth as supervision. Unlike outdoor driving, which commonly features sky on top, and road and vehicles in the middle of the image with largely planar camera motion, indoor scenes are characterized by more complex layouts with 6 DoF camera motion, yielding images that are even less likely to resemble the object-centric ones commonly observed in classification datasets. This may be why initializing the model with depth pre-trained weights significantly improves semantic segmentation accuracy with both ResNet18 and ResNet50 (see Fig. 8). Note that pre-training by depth yields faster convergence, similar to KITTI. + +# 4.3 Out-of-domain Transfer from Large-scale Pre-training + +One may conjecture that the advantage of pre-training with monocular depth stems from training within the same domain as downstream semantic segmentation. In practice, in-domain data collection for depth pre-training is indeed ideal since it does not require human labeling. In a scientific context, we are interested in testing this conjecture by investigating the transferability of depth models to tasks outside their original domain. However, common depth datasets are considerably small in scale, making fair comparisons with popular pre-trained models, e.g. MAE [18] and DINO v2 [38] that are trained on millions or even billions of images, infeasible. Fortunately, recent methods [33,41] suggest an alternative approach for learning monocular depth by training for relative depth instead of absolute depth, allowing for the integration of multiple data sources during training. Leveraging Depth Anything [55], a depth model trained on such scaled-up mixed datasets, we fine-tune for segmentation on three out-of-domain downstream datasets: ADE20k [59], PascalVOC [13], and CityScapes [10]. + +The PascalVOC dataset comprises 10,582 fully annotated images for training purposes and an additional 1,449 for testing, covering a variety of 20 foreground object classes. On the other hand, ADE20k, a sizable dataset, includes 20,210 training images and 2,000 testing images across 150 different classes. The resolution of PascalVOC and ADE20k is $512 \times 512$ and $896 \times 896$ respectively. On both datasets, we try both fine-tuning the whole network and linear probing with a frozen encoder. Reported in Tab. 6, the results align with the findings on ViTs on the KITTI dataset (in Tab. 3), in which case the pre-training data also originates from out-of-domain sources. Notably, consistent improvements in semantic segmentation performance are observed across all datasets compared to the baseline initialization DINO v2 [38]. It is anticipated that the improvement achieved with a frozen encoder is more substantial than when fine-tuning the entire network, which is consistent with our previous results. This trend underscores that, when trained at a scale, depth models exhibit robust transferability to novel downstream data domains, just as other pre-training methods. + +Table 6: Out-of-domain transfer with large-scale pre-trainings. We compare Depth Anything with MAE and DINO v2 for semantic segmentation, reporting results (in mIoU) on both fine-tuning (ft) and linear probing (lin.). Depth Anything improves semantic segmentation across all datasets and settings. *: with DINO v2 initialization. + +
Pre-training# of pre-train dataADE20kPascalVOCCityScapes
ftlin.ftlin.ftlin.
MAE [18]1.28 million53.649.0-67.6-58.4
DINO v2 [38]142 million58.147.786.586.382.771.3
Depth Anything [55]63.5 million*59.752.387.787.384.874.8
+ +# 5 Discussion + +Inferring depth only requires multiple images (e.g. videos or multiple viewpoints) of the same scene [25] or range sensing, both do not require human-induced priors and bias, unlike semantic tasks that rely entirely on induction: We can associate a label to an image because that image has something in common with some other image, portraying a different scene, that some annotator attached a particular label to. That inductive chain has to go through the head of human annotators, who are biased in ways that cannot be easily quantified and controlled. Depth from binocular or motion imagery does not require induction and can be performed ab-ovo. Learning a monocular model from such supervision eliminates the implicit selective bias from human annotators, yet our findings validate the main hypothesis that the inductive bias learned from such a "human-free" process transfers well to the downstream semantic segmentation task. Of course, if different supervisions are available, either for semantic [11], segmentation [27], or both [28], we want to incorporate that information. + +Our hypothesis and findings are agnostic to how depth is attributed to a single image: One can perform pre-training using monocular videos, stereo, structure light, LIDAR, or even human guidance [58]. One of our pre-training minimizes the photometric reprojection error, used by many predictive and generative approaches. However, an unstructured displacement field is in general not compatible with a rigid motion. Only if this displacement field has the structure of an epipolar transformation [44] is the prediction task forced to encode the 3D structure of the scene. This may explain why video prediction is not as effective for pre-training despite many attempts [24,32,46,53]. + +One limitation of monocular depth estimation is that it may require a calibrated camera, so one cannot use generic videos harvested from the web. There is nothing in principle preventing us from using uncalibrated cameras, simply by adding the calibration matrix $K$ to the nuisance variables. While the necessary conditions for full Euclidean reconstruction are rarely satisfied in consumer videos (for instance, they require cyclo rotation around the optical axis, not just panning and tilting), the degrees of freedom that cannot be reconstructed are moot as they do not affect the reprojection. Moreover, recent progress in training for relative depth from mixed data sources [33, 41, 55] shows the potential to unlock a virtually unlimited volume of training data, warranting future research. + +Acknowledgements. This work was supported by ONR N00014-22-1-2252 and ARO W911NF-17-1-0304. + +# References + +1. Bertalmio, M., Sapiro, G., Caselles, V., Ballester, C.: Image inpainting. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 417-424 (2000) +2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +3. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019) +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017) +6. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017) +7. Chen, M., Artières, T., Denoyer, L.: Unsupervised object segmentation by redrawing. Advances in neural information processing systems 32 (2019) +8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009) +12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +13. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88(2), 303-338 (Jun 2010) +14. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019) +15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012) + +16. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth prediction (October 2019) +17. Goldblum, M., Souri, H., Ni, R., Shu, M., Prabhu, V., Somepalli, G., Chattopadhyay, P., Ibrahim, M., Bardes, A., Hoffman, J., et al.: Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems 36 (2024) +18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021) +19. Hoyer, L., Dai, D., Chen, Y., Köring, A., Saha, S., Van Gool, L.: Three ways to improve semantic segmentation with self-supervised depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11130-11140 (2021) +20. Hoyer, L., Dai, D., Wang, Q., Chen, Y., Van Gool, L.: Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. arXiv preprint arXiv:2108.12545 [cs] (2021) +21. Huang, J., Lee, A.B., Mumford, D.: Statistics of range images. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662). vol. 1, pp. 324-331. IEEE (2000) +22. Huang, J., Mumford, D.: Statistics of natural images and models. In: Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149). vol. 1, pp. 541-547. IEEE (1999) +23. Jiang, H., Larsson, G., Shakhnarovich, M.M.G., Learned-Miller, E.: Self-supervised relative depth learning for urban scene understanding. In: Proceedings of the European conference on computer vision (eccv). pp. 19-35 (2018) +24. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020) +25. Julesz, B.: Binocular depth perception without familiarity cues: Random-dot stereo images with controlled spatial and temporal properties clarify problems in stereopsis. Science 145(3630), 356-362 (1964) +26. Julesz, B.: Foundations of cyclopean perception. (1971) +27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023) +28. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020) +29. Lao, D., Hu, Z., Locatello, F., Yang, Y., Soatto, S.: Divided attention: Unsupervised multi-object discovery with contextually separated slots. arXiv preprint arXiv:2304.01430 (2023) +30. Lao, D., Sundaramoorthi, G.: Minimum delay moving object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4250-4259 (2017) +31. Lao, D., Sundaramoorthi, G.: Extending layered models to 3d motion. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018) + +32. Lao, D., Zhu, P., Wonka, P., Sundaramoorthi, G.: Flow-guided video inpainting with scene templates. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14599-14608 (2021) +33. Lasinger, K., Ranftl, R., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. arXiv preprint arXiv:1907.01341 (2019) +34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +35. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: European Conference on Computer Vision. pp. 35-53. Springer (2022) +36. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020) +37. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012) +38. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024), https://openreview.net/forum?id=a68SUt6zFt +39. Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T., Efros, A.A.: Context encoders: Feature learning by inpainting. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2536-2544 (2016) +40. Ramirez, P.Z., Tonioni, A., Salti, S., Stefano, L.D.: Learning across tasks and domains. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8110-8119 (2019) +41. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) +42. Saha, S., Obukhov, A., Paudel, D.P., Kanakis, M., Chen, Y., Georgoulis, S., Van Gool, L.: Learning to relate depth and semantics for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8197-8207 (2021) +43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +44. Sundaramoorthi, G., Petersen, P., Varadarajan, V., Soatto, S.: On the set of images modulo viewpoint and contrast changes. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 832-839. IEEE (2009) +45. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method. arXiv preprint physics/0004057 (2000) +46. Wang, Y., Wu, J., Long, M., Tenenbaum, J.B.: Probabilistic video prediction from noisy data with a posterior confidence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10830-10839 (2020) + +47. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021) +48. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021) +49. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020) +50. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019) +51. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021) +52. Wu, Y., Liu, T.Y., Park, H., Soatto, S., Lao, D., Wong, A.: Augundo: Scaling up augmentations for monocular depth completion and estimation. In: European Conference on Computer Vision. Springer (2024) +53. Wu, Y., Gao, R., Park, J., Chen, Q.: Future video synthesis with object motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5539-5548 (2020) +54. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) +55. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024) +56. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019) +57. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3712-3722 (2018) +58. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024) +59. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017) \ No newline at end of file diff --git a/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/images.zip b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..685e3f1527bd9abf12347a974abc69551b59f65f --- /dev/null +++ b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:727474dbd7898c5bd5bec78501e4e55b5177a01ea72f5a63f88ef85d60ca1f20 +size 493994 diff --git a/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/layout.json b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8739de54d40b0aae8426879f9eb8aaa6e534f2b4 --- /dev/null +++ b/2024/On the Viability of Monocular Depth Pre-training for Semantic Segmentation/layout.json @@ -0,0 +1,9012 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "text", + "content": "On the Viability of Monocular Depth Pre-training for Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 167, + 475, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 167, + 475, + 194 + ], + "spans": [ + { + "bbox": [ + 139, + 167, + 475, + 194 + ], + "type": "text", + "content": "Dong Lao1, Fengyu Yang2, Daniel Wang2, Hyoungseob Park2, Samuel Lu1, Alex Wong2, and Stefano Soatto1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 201, + 410, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 201, + 410, + 213 + ], + "spans": [ + { + "bbox": [ + 201, + 201, + 410, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 201, + 410, + 213 + ], + "type": "text", + "content": " UCLA Vision Lab, Los Angeles, CA 90024, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 213, + 411, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 213, + 411, + 223 + ], + "spans": [ + { + "bbox": [ + 201, + 213, + 411, + 223 + ], + "type": "text", + "content": "{lao, soatto}@cs.ucla.edu, samuellu@ucla.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 224, + 405, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 224, + 405, + 235 + ], + "spans": [ + { + "bbox": [ + 201, + 224, + 405, + 235 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 224, + 405, + 235 + ], + "type": "text", + "content": " Yale Vision Lab, New Haven, CT 06511, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 236, + 470, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 470, + 246 + ], + "type": "text", + "content": "{fengyu.yang, daniel.wang.dhw33, hyoungseob.park, alex.wong}@yale.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 269, + 455, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 269, + 455, + 544 + ], + "spans": [ + { + "bbox": [ + 159, + 269, + 455, + 544 + ], + "type": "text", + "content": "Abstract. The question of whether pre-training on geometric tasks is viable for downstream transfer to semantic tasks is important for two reasons, one practical and the other scientific. If the answer is positive, we may be able to reduce pre-training costs and bias from human annotators significantly. If the answer is negative, it may shed light on the role of embodiment in the emergence of language and other cognitive functions in evolutionary history. To frame the question in a way that is testable with current means, we pre-train a model on a geometric task, and test whether that can be used to prime a notion of \"object\" that enables inference of semantics as soon as symbols (labels) are assigned. We choose monocular depth prediction as the geometric task, and semantic segmentation as the downstream semantic task, and design a collection of empirical tests by exploring different forms of supervision, training pipelines, and data sources for both depth pre-training and semantic fine-tuning. We find that monocular depth is a viable form of pre-training for semantic segmentation, validated by improvements over common baselines. Based on the findings, we propose several possible mechanisms behind the improvements, including their relation to dataset size, resolution, architecture, in/out-of-domain source data, and validate them through a wide range of ablation studies. We also find that optical flow, which at first glance may seem as good as depth prediction since it optimizes the same photometric reprojection error, is considerably less effective, as it does not explicitly aim to infer the latent structure of the scene, but rather the raw phenomenology of temporally adjacent images. Code: https://github.com/donglao/DepthToSemantic." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 554, + 443, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 554, + 443, + 565 + ], + "spans": [ + { + "bbox": [ + 160, + 554, + 443, + 565 + ], + "type": "text", + "content": "Keywords: Depth estimation " + }, + { + "bbox": [ + 160, + 554, + 443, + 565 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 554, + 443, + 565 + ], + "type": "text", + "content": " semantic segmentation " + }, + { + "bbox": [ + 160, + 554, + 443, + 565 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 554, + 443, + 565 + ], + "type": "text", + "content": " pre-training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 605, + 409, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 409, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 409, + 617 + ], + "type": "text", + "content": "We probe the following seemingly counter-intuitive hypothesis:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 469, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 469, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 469, + 666 + ], + "type": "text", + "content": "Can pre-training on a geometric task benefit a downstream semantic task? Geometric inference is often viewed as a low-level vision task requiring little abstraction that is needed for semantics [26]. For example, depth can be aquired through minimizing reprojection error, i.e. from multi-view or videos," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "directly from range sensors. Both can be performed procedurally, without inductive learning, rendering depth a meaningless task for pre-training. However, induction is needed to infer one 3D scene among infinitely many compatible with the same 2D image. Therefore, if a model could solve this ill-posed problem, it would provide evidence of the viability of pre-training with little to no human intervention, which is important in specialized data domains for which little annotated data is publicly available." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 204, + 483, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 483, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 483, + 348 + ], + "type": "text", + "content": "In this paper, we focus on testing monocular depth as the pre-training geometric task, and semantic segmentation as the downstream semantic task. They are purposefully chosen: Training deep neural networks for semantic segmentation requires labor-intensive pixel-level annotation, so the choice of pre-training is essential to its performance. Existing studies have shown mixed results about the relationship between the two tasks. Taskonomy [57], a framework for measuring relationships between visual tasks, suggests that depth estimation is \"far\" from semantic segmentation, while recent work [17] shows that depth pre-training can beat \"closer\" tasks like image classification. Prior work [19, 23, 40, 42] has also shown improvement in semantic segmentation when incorporating depth. This mixed evidence motivates us to take a closer look at the underlying mechanism of how monocular depth may benefit semantic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 483, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 483, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 483, + 532 + ], + "type": "text", + "content": "Another more subtle reason for exploring this hypothesis is that pre-training is often performed on heavily human-biased datasets [11,38], where the photographer who framed the picture meant to convey a particular concept (say, a cup), and therefore took care to make sure that the manifestation of the concept (the image) prominently features the object by choice of vantage point, illumination, and (lack of) occlusion. This bias is mitigated if data is not purposefully organized into \"shots.\" Unfortunately, existing datasets are mostly composed of purposefully framed shots which could obfuscate the analysis. We note that depth can be inferred without any semantic interpretation [26] regardless of whether the data is captured purposefully or randomly. With monocular depth as the pre-training task, there are two ways of reducing the aforementioned human selective bias: The first involves directly pre-training within the specific domain of interest, leveraging the simplicity of data gathering; The second way is scaling up pre-training by incorporating diverse sources of data, which is made possible by recent developments [33,41,54] on relative depth estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "Methods. We formalize the main hypothesis in Sect. 3. Since it cannot be tested analytically without knowledge of the joint distribution of test images and labels, we propose an empirical testing protocol. We test on monocular depth models trained under multiple forms of supervision, including structure-from-motion, binocular stereo, and depth sensors. We then change the prediction head of the resulting network, either the final layer or the whole decoder, and fine-tune it for semantic segmentation (Fig. 1). We consider depth estimation as a viable pre-training option if it yields comparable improvements to downstream semantic segmentation tasks as other common pre-training practices, e.g. ImageNet classification. To this end, we design a series of controlled experiments to test the effect of choice of initialization (Tab. 1, Fig. 2), training with various datasets" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "sizes (Fig. 3), choice of network component to be frozen and fine-tuned (Fig. 4), effect of resolution of training images (Fig. 5). Conclusions are drawn from both quantitative and qualitative (Fig. 6) results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 482, + 381 + ], + "type": "text", + "content": "Findings. Pre-training for depth estimation improves the performance of downstream semantic segmentation across different experimental settings. Particularly, we show that depth estimation is indeed a viable pre-training option as compared to existing methods (Tab. 4). For example, compared to classification, using depth on average improves by " + }, + { + "bbox": [ + 130, + 153, + 482, + 381 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 130, + 153, + 482, + 381 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 130, + 153, + 482, + 381 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 130, + 153, + 482, + 381 + ], + "type": "text", + "content": " pixel accuracy on KITTI. As a sanity check, we test both a depth network pre-trained from scratch and one trained after ImageNet initialization, and both outperform classification-based pre-training in downstream semantic segmentation. To control the effect of our choice of architecture, we used our pre-trained encoder to initialize a standard semantic segmentation network [5]. We observed similar findings on Cityscapes and NYU-V2 regardless of how depth training is supervised. Inferring depth without explicit supervision typically involves minimizing the prediction error, just like optical flow. Somewhat surprisingly, not only does pre-training for depth outperform optical flow, but the latter is often worse than random initialization (Fig. 7). One may also argue that observed improvements mainly come from the availability of in-domain pre-training data for depth. To test this conjecture, we fine-tune a depth model [55] trained on large-scale out-of-domain data. Improvements in semantic segmentation reveal that when trained at scale, depth models show strong transferability to unseen downstream data domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 406, + 237, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 237, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 237, + 418 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "Pre-training aims to learn a representation (function) of the test data that is maximally informative (sufficient), while providing some kind of complexity advantage. In our case, we measure complexity by the validation error after fine-tuning on limited amount of labeled data, which measures the inductive value of pretraining. The recent literature comprises a large variety of \"self-supervised\" methods that are purportedly task-agnostic. In reality, the task is specified indirectly by the choice of hand-designed nuisance transformations that leave the outcome of inference unchanged. Such transformations are sampled through data augmentation while the image identity holds constant (Contrastive Learning) [4,8,9,38], or reconstruction [3]. Group transformations organize the dataset into orbits, which contrastive learning tries to collapse onto its quotient, which is a maximal invariant. Such a maximal invariant is transferable to all and only tasks for which the chosen transformation is uninformative. For group transformation, the maximal invariant can, in theory, be computed in closed form [44]. In practice, contrastive learning are extended to non-group transformations, e.g. occlusions, as seen in language [2] and images [7]. All self-supervised methods boil down to hand-designed and quantized subsets of planar domain diffeomorphisms (discrete rotations, translations, scaling, reflections, etc.), range homeomorphisms (contrast, colormap transformations) and occlusion masks." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 111, + 236, + 178 + ], + "blocks": [ + { + "bbox": [ + 164, + 111, + 236, + 178 + ], + "lines": [ + { + "bbox": [ + 164, + 111, + 236, + 178 + ], + "spans": [ + { + "bbox": [ + 164, + 111, + 236, + 178 + ], + "type": "image", + "image_path": "6fcc8480d5bd6c2eb105573169c6b1000ee380fcd9a85ba87b3fe89a1323e831.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 180, + 230, + 188 + ], + "lines": [ + { + "bbox": [ + 196, + 180, + 230, + 188 + ], + "spans": [ + { + "bbox": [ + 196, + 180, + 230, + 188 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 111, + 288, + 178 + ], + "blocks": [ + { + "bbox": [ + 242, + 111, + 288, + 178 + ], + "lines": [ + { + "bbox": [ + 242, + 111, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 242, + 111, + 288, + 178 + ], + "type": "image", + "image_path": "f63a1edecb0fd2377ec7220b4020b2d0b6fa897fa5bb1a55cd9d2ac497d7933b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 180, + 271, + 188 + ], + "lines": [ + { + "bbox": [ + 261, + 180, + 271, + 188 + ], + "spans": [ + { + "bbox": [ + 261, + 180, + 271, + 188 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 269 + ], + "type": "text", + "content": "Fig. 1: Diagram for different pre-training and fine-tuning setups. (a) Common practice: pre-train the encoder, e.g. on ImageNet, attach a decoder, and fine-tune the network. (b) Our best practice: pre-train the network by monocular depth, and fine-tune for semantic segmentation. (c) Cross architecture: for fair comparisons with common practice, we pre-train by depth, replace the decoder, and fine-tune. (d) To test the quality of pre-trained encoders, we fix the encoders and fine-tune decoders only." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 292, + 110, + 364, + 178 + ], + "blocks": [ + { + "bbox": [ + 292, + 110, + 364, + 178 + ], + "lines": [ + { + "bbox": [ + 292, + 110, + 364, + 178 + ], + "spans": [ + { + "bbox": [ + 292, + 110, + 364, + 178 + ], + "type": "image", + "image_path": "eed0a2744226ecb4a62d10439189c1b3ba6250eda358a00f71453adfaee1c56a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 180, + 331, + 188 + ], + "lines": [ + { + "bbox": [ + 322, + 180, + 331, + 188 + ], + "spans": [ + { + "bbox": [ + 322, + 180, + 331, + 188 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 369, + 109, + 449, + 178 + ], + "blocks": [ + { + "bbox": [ + 369, + 109, + 449, + 178 + ], + "lines": [ + { + "bbox": [ + 369, + 109, + 449, + 178 + ], + "spans": [ + { + "bbox": [ + 369, + 109, + 449, + 178 + ], + "type": "image", + "image_path": "e2bd894af91c3ce860bc89c8d11b734c71cc5a2f5d9f94a2a7993a925c30f1fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 180, + 413, + 188 + ], + "lines": [ + { + "bbox": [ + 403, + 180, + 413, + 188 + ], + "spans": [ + { + "bbox": [ + 403, + 180, + 413, + 188 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 290, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 540 + ], + "type": "text", + "content": "In our case, rather than hand-designing the nuisance transformations assumed to be shared among pre-training and fine-tuning tasks, we let the scene itself provide the needed supervision: images portend the same scene, either from the same timestamp (stereo) or adjacent in the temporal domain (video frames), so their variability defines the union of nuisance factors. These include domain deformations due to ego- and scene motion, range transformations due to changes in illumination, and occlusions. In addition to sharing nuisance variability, pretraining and fine-tuning tasks should ideally also share the hypothesis space. It may seem odd to choose a geometric task, where the hypothesis space is depth, to pre-train for a semantic task, where the hypothesis space is a discrete set of labels. However, due to the statistics of range images [21] and their similarity to the statistics of natural images [22], this is actually quite natural: A range map is a piecewise smooth function defined on the image domain, whereas a segmentation map is a piecewise constant function where the levels are mapped to arbitrary labels. As a result, the decoder for depth estimation can be easily modified for semantic segmentation. A discussion of this choice, specifically on the representational power of deterministic predictors, in Sect. 5. [19, 23] also utilize depth for semantic segmentation. [23] proposes pre-training on relative depth prediction, and [19, 20] utilizes self-supervised depth estimation on video sequences. Our experiments validate their findings. We further investigate whether features obtained purely from monocular depth improve semantic segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 601 + ], + "type": "text", + "content": "Monocular depth [16, 50] may use different supervision, either through additional sensors [14, 51, 52], or synthetic data [36, 47, 56], but none require human annotation. Some use regularizers with sparse seeds [35, 48, 49], or adopt pretrainings [41]. We design experiments agnostic to how depth models are trained, but also make comparisons across different forms of depth supervision (Tab. 2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 618, + 237, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 237, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 237, + 631 + ], + "type": "text", + "content": "3 Formalization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "x: D \\subset \\mathbb{R}^2 \\to \\{0, \\ldots, 255\\}^3" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " be an image, where the domain " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " is quantized into a lattice, " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "z: D \\to \\{1, \\ldots, Z\\}" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " a depth map with " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " depth or disparity levels," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "y: D \\to \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " a semantic segmentation map. In coordinates, each pixel in the lattice, " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "(i,j) \\in \\{1, \\ldots, N\\} \\times \\{1, \\ldots, M\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " is mapped to RGB intensities by " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "x(i,j)" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": ", a depth by " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "z(i,j)" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": ", and a label by " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "y(i,j)" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": ". Despite the discrete nature of the data and the hypothesis space, we relax them to the continuum by considering the vectors " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{NM3}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{NMK}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{NM}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": ". With a slight abuse of notation, " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "y \\in \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " denotes a single label and " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\bar{y} \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " its embedding, often restricted to a one-hot encoding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "text", + "content": "Now, consider a dataset " + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_z = \\{\\mathbf{x}_t^i,\\mathbf{z}_t^i\\}_{i,t = 1}^{V,T_i}" + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "text", + "content": " comprised of " + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "text", + "content": " image sequences each of length " + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "text", + "content": ". For the simplicity of the notations, in the case of multi-view stereo, we also consider multiple 2D image inputs as a \"sequence\" without loss of generality. In the case of supervised depth estimation, synchronized depth maps " + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "inline_equation", + "content": "z_{t}^{i}" + }, + { + "bbox": [ + 130, + 199, + 480, + 272 + ], + "type": "text", + "content": "'s are measured by a range sensor. Typical datasets supporting depth estimation may include just image sequences or both modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "text", + "content": "Training for monocular depth estimation yields a mapping " + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "inline_equation", + "content": "\\phi_w: \\mathbf{x} \\mapsto \\mathbf{z}" + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "text", + "content": ", parametrized by weights " + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 272, + 479, + 296 + ], + "type": "text", + "content": " in a neural network, via" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 305, + 480, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 305, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 164, + 305, + 480, + 331 + ], + "type": "interline_equation", + "content": "w = \\arg \\min _ {w, g _ {t}} \\sum_ {i, j, n, t} \\ell \\left(x _ {t + 1} ^ {n} (i, j), \\hat {x} _ {t} ^ {n} (i, j)\\right) + \\lambda \\sum_ {i, j, n, t} \\ell \\left(z _ {t} ^ {n} (i, j), \\hat {z} _ {t} ^ {n} (i, j)\\right) \\tag {1}", + "image_path": "a4074b42b425f38c1280c2b3b450edacd2578470040b4fec65462b2884ac2697.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": " is the warping of an image " + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": " based on camera pose " + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": ". Here we consider a generic formulation for different modalities of depth estimation. The first term in Eq. (1) measures reprojection error across frames, and the second term measures the distance between estimated depth values and the ground-truth from the range sensor. In the case of unsupervised depth estimation from videos, e.g. [16], only the reprojection loss is considered; while in the case of supervised depth estimation from single images, e.g. [37], " + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "inline_equation", + "content": "T_i = 1" + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 339, + 480, + 434 + ], + "type": "text", + "content": "'s, so only the second term is minimized." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "text", + "content": "The goal is to use these representations as encodings of the data to then learn a semantic segmentation map. In practice, the representations above are implemented by deep neural networks, that can be truncated at intermediate layers thus providing embedding spaces larger than the respective hypothesis spaces. We refer to the parts before and after this intermediate layer encoder and decoder, respectively. We overload the notation and refer to the encoding as " + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = \\phi_w(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "text", + "content": " for both depth estimation and other pre-training methods, presumably with weights " + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "inline_equation", + "content": "w'" + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "text", + "content": ", assuming they have the same encoder architecture. The goal of semantic segmentation is then to learn a parametrized map " + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "inline_equation", + "content": "\\psi_{w''}: \\mathbf{h} \\mapsto \\mathbf{y}" + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "text", + "content": " using a small but fully supervised dataset " + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = \\{\\mathbf{x}^n, \\mathbf{y}^n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "text", + "content": ", by minimizing some loss function or (pseudo-)distance in the hypothesis space " + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "inline_equation", + "content": "d(\\mathbf{y}, \\hat{\\mathbf{y}})" + }, + { + "bbox": [ + 130, + 435, + 481, + 567 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 578, + 480, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 238, + 578, + 480, + 609 + ], + "type": "interline_equation", + "content": "w ^ {\\prime \\prime} = \\arg \\min _ {w} \\sum_ {n = 1} ^ {N} d \\left(\\mathbf {y} ^ {n}, \\psi_ {w} \\left(\\mathbf {h} ^ {n}\\right)\\right) \\tag {2}", + "image_path": "ec1474b79b66dd737e5b5fbe676ed048fb55b1c534ce8453f60d18e4f7e75657.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "plus customary regularizers. In the aggregate, we have a Markov chain: " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\longrightarrow \\mathbf{h} = \\phi_w(\\mathbf{x}) \\longrightarrow \\mathbf{y} = \\psi_{w''}(\\mathbf{h}) = \\psi_{w''} \\circ \\phi_w(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " for depth estimation, and " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\longrightarrow \\hat{\\mathbf{h}} = \\phi_{w'}(\\mathbf{x}) \\longrightarrow \\mathbf{y} = \\psi_{w''} \\circ \\phi_{w'}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " for other pre-training methods. A representation obtained through a Markov chain is optimal (minimal sufficient) only if the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 179, + 472, + 261 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Semantic segmentation accuracy on KITTI. Unsupervised depth as pre-training improves semantic segmentation accuracy under all settings. Our best practice (in blue) improves common practice (in purple) by " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "7.53\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "4.68\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " pixel accuracy. Freezing the encoder with ImageNet pre-training (in red) is worse than no pre-training (random initialization). DeepLabV3†: with ResNet50 encoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 179, + 472, + 261 + ], + "lines": [ + { + "bbox": [ + 145, + 179, + 472, + 261 + ], + "spans": [ + { + "bbox": [ + 145, + 179, + 472, + 261 + ], + "type": "table", + "html": "
Fine-tune AllFreeze Encoder
ResNet18ResNet50DeepLabV3†ResNet18ResNet50
Pre-trainingmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.AccmIoUP.Acc
None41.3570.7544.6673.3721.9352.3241.2470.5237.7267.38
ImageNet45.1572.3944.6573.0643.3972.6633.3365.3432.0362.53
Depth-Rand46.0072.4349.9076.2843.4371.3443.0272.3845.7974.71
Depth50.2076.3950.9277.3443.7772.6846.5374.4246.5574.48
", + "image_path": "550462561bb8be7f2f453b434711b3b1bf22c0fdbbf63202b819b7907edc8ce6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 319 + ], + "type": "text", + "content": "intermediate variable " + }, + { + "bbox": [ + 130, + 281, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 130, + 281, + 482, + 319 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 281, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{h}}" + }, + { + "bbox": [ + 130, + 281, + 482, + 319 + ], + "type": "text", + "content": " reduces the Information Bottleneck [45] to zero. In general, there is information loss, so we formalize the key question as whether the two Information Bottleneck Lagrangians satisfy the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 326, + 481, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 216, + 326, + 481, + 346 + ], + "type": "interline_equation", + "content": "H (\\mathbf {y} | \\mathbf {h}) + \\beta I (\\mathbf {h}; \\mathbf {x}) \\stackrel {?} {\\leq} H (\\mathbf {y} | \\hat {\\mathbf {h}}) + \\beta^ {\\prime} I (\\hat {\\mathbf {h}}; \\mathbf {x}) \\tag {3}", + "image_path": "eba335e37feddc2be901f4cb144cfdecc0af8631d07b9f1057de5937726a26db.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\beta'" + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "text", + "content": " are hyperparameters that can be optimized as part of the training process, and " + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "inline_equation", + "content": "I, H" + }, + { + "bbox": [ + 130, + 353, + 482, + 425 + ], + "type": "text", + "content": " denotes the (Shannon) Mutual Information and cross-entropy respectively. If the above is satisfied, then pre-training for depth estimation is a viable option, or even better than pre-training with another method. It would be ideal if this question could be settled analytically. Unfortunately, this is not possible, but the formalization above suggests a protocol to settle it empirically." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": "To test this empirically, we use the validation error on a supervised dataset " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": " as a proxy for residual information. We conduct fine-tuning under several configurations (Fig. 1): with respect to " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "w''" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": ", i.e. yielding a comparison of the raw pre-trained back-bones (encoders, " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "w, w'" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": "), or with respect to both " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "w''" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": " (for depth estimation) or " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "w'" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": " (for other pre-training methods). Finally, all four resulting models can be compared with one obtained by training from scratch by optimizing a generic architecture with respect to " + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "inline_equation", + "content": "w''" + }, + { + "bbox": [ + 130, + 426, + 482, + 509 + ], + "type": "text", + "content": " alone." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 526, + 230, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 230, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 230, + 540 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 550, + 418, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 418, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 418, + 563 + ], + "type": "text", + "content": "4.1 Controlled Experiments with Few-shot Fine-tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "content": "We first cover an extensive collection of controlled experiments and ablations to gain insights into the main hypothesis. We specifically conduct experiments under the few-shot setting, where only a small amount of labels are available for fine-tuning, to highlight the role of pre-training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "KITTI [15] contains 93000 video frames for depth training with 200 densely annotated images for semantic segmentation. Segmentation results are evaluated by mean IoU (mIoU) and pixel-level accuracy (P.Acc). We randomly choose a small training set of 16 images and limit data augmentation to horizontal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 116, + 261, + 178 + ], + "blocks": [ + { + "bbox": [ + 171, + 116, + 261, + 178 + ], + "lines": [ + { + "bbox": [ + 171, + 116, + 261, + 178 + ], + "spans": [ + { + "bbox": [ + 171, + 116, + 261, + 178 + ], + "type": "image", + "image_path": "63ed67334ad45839306c3a0fcd33015c518c5652aa2830b79a9a7bb565dc3a1b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 267, + 116, + 347, + 178 + ], + "blocks": [ + { + "bbox": [ + 267, + 116, + 347, + 178 + ], + "lines": [ + { + "bbox": [ + 267, + 116, + 347, + 178 + ], + "spans": [ + { + "bbox": [ + 267, + 116, + 347, + 178 + ], + "type": "image", + "image_path": "a2009dc9ac4121e32e212ab1db1bc7ad4bc35de0882060e3dbcbc35a2bfcc203.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 352, + 116, + 432, + 178 + ], + "blocks": [ + { + "bbox": [ + 352, + 116, + 432, + 178 + ], + "lines": [ + { + "bbox": [ + 352, + 116, + 432, + 178 + ], + "spans": [ + { + "bbox": [ + 352, + 116, + 432, + 178 + ], + "type": "image", + "image_path": "708f8fe632a1972f66d4f0e7b1fae7780ed23cf248bd6ebe8ac1a8cc9f57badb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 171, + 181, + 261, + 244 + ], + "blocks": [ + { + "bbox": [ + 171, + 181, + 261, + 244 + ], + "lines": [ + { + "bbox": [ + 171, + 181, + 261, + 244 + ], + "spans": [ + { + "bbox": [ + 171, + 181, + 261, + 244 + ], + "type": "image", + "image_path": "5b57aa0b486123b48c5444ef9073ebad1c168801ee98a5f5ce79684c96e982bb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "type": "text", + "content": "Fig. 2: Comparison between different network initializations. Models initialized by depth pre-training (unsupervised) train faster and achieve higher final accuracy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 267, + 182, + 346, + 244 + ], + "blocks": [ + { + "bbox": [ + 267, + 182, + 346, + 244 + ], + "lines": [ + { + "bbox": [ + 267, + 182, + 346, + 244 + ], + "spans": [ + { + "bbox": [ + 267, + 182, + 346, + 244 + ], + "type": "image", + "image_path": "1eb239e452bfacfa0deab7913bd57fec3a168e8c42b0fa1a728546a39730c224.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 353, + 182, + 432, + 244 + ], + "blocks": [ + { + "bbox": [ + 353, + 182, + 432, + 244 + ], + "lines": [ + { + "bbox": [ + 353, + 182, + 432, + 244 + ], + "spans": [ + { + "bbox": [ + 353, + 182, + 432, + 244 + ], + "type": "image", + "image_path": "3610a031aa4f0fcaba0e6933c6a9ff73af11574dd04dd27df45b8d802b10385b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 134, + 292, + 216, + 356 + ], + "blocks": [ + { + "bbox": [ + 134, + 292, + 216, + 356 + ], + "lines": [ + { + "bbox": [ + 134, + 292, + 216, + 356 + ], + "spans": [ + { + "bbox": [ + 134, + 292, + 216, + 356 + ], + "type": "image", + "image_path": "2f3c3dcb5e590a9164029753c969b39c88c670981451ee881273f926b6bb44ae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 369, + 315, + 413 + ], + "lines": [ + { + "bbox": [ + 131, + 369, + 315, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 315, + 413 + ], + "type": "text", + "content": "Fig. 3: Final accuracy vs different training set size. Under all training set sizes, our best practice constantly outperforms ImageNet pre-trained. Encoder: ResNet 18." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 225, + 292, + 304, + 356 + ], + "blocks": [ + { + "bbox": [ + 225, + 292, + 304, + 356 + ], + "lines": [ + { + "bbox": [ + 225, + 292, + 304, + 356 + ], + "spans": [ + { + "bbox": [ + 225, + 292, + 304, + 356 + ], + "type": "image", + "image_path": "f988e7db50399416bf7c87d3dd26a68c4b1f394f3696e2c1f2d1f9161fd8e30a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 349, + 302, + 455, + 350 + ], + "blocks": [ + { + "bbox": [ + 349, + 302, + 455, + 350 + ], + "lines": [ + { + "bbox": [ + 349, + 302, + 455, + 350 + ], + "spans": [ + { + "bbox": [ + 349, + 302, + 455, + 350 + ], + "type": "table", + "html": "
Depth Sup.mIoUP.Acc
Video46.0072.43
Stereo49.1174.58
Lidar52.7877.17
", + "image_path": "56c3c460cb8c50a1e16342ff1a9c7debe1ffb66bcd9ef17721fd29c1dfa6ca4c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 357, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 321, + 357, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 321, + 357, + 482, + 414 + ], + "type": "text", + "content": "Table 2: Forms of depth supervision matter. Direct supervision with Lidar works the best, followed by stereo (with known camera pose), and monocular video (camera pose unknown)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "flips to highlight the impact of pre-training, except for Fig. 3 where we test on different training set partitions. We use Monodepth2 [16] for depth pre-training. For semantic segmentation, we replace the last layer of the decoder with a fully connected layer, using the finest scale of the multi-scale output. We test on ResNet18 and ResNet50 encoders due to their compatibility with various network architectures and widely public-available pre-trained models. Fig. 1 summarizes our experimental setup and Tab. 1 summarizes the outcomes. In all cases, depth pre-training improves segmentation accuracy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": "Full model. Fig. 2 shows the evolution of training loss and model accuracy. Depth pre-training outperforms ImageNet and random initialization. ImageNet pre-training slightly improves over random initialization on ResNet18, but shows almost identical performance to random initialization on ResNet50. Depth also speeds up training, taking " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\sim 5000" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " iterations to converge, while ImageNet takes 15000 to 20000. Similar results on full-resolution are deferred to the Supp. Mat." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Different training set size. Fig. 3 shows that depth pre-training improves final segmentation scores over all dataset partitions (with ResNet 18). When training samples increase (e.g. 128), ImageNet and depth pre-training (from random initialization) are comparable to random initialization, but depth pretraining initialized with ImageNet yields the best results." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 116, + 261, + 178 + ], + "blocks": [ + { + "bbox": [ + 171, + 116, + 261, + 178 + ], + "lines": [ + { + "bbox": [ + 171, + 116, + 261, + 178 + ], + "spans": [ + { + "bbox": [ + 171, + 116, + 261, + 178 + ], + "type": "image", + "image_path": "7989ba00cd01927a57550e824864c98edf572490419520ab6a170496b9e03d58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 267, + 118, + 347, + 178 + ], + "blocks": [ + { + "bbox": [ + 267, + 118, + 347, + 178 + ], + "lines": [ + { + "bbox": [ + 267, + 118, + 347, + 178 + ], + "spans": [ + { + "bbox": [ + 267, + 118, + 347, + 178 + ], + "type": "image", + "image_path": "7859699f2b2dd1c5c74326a3051db54ef5e56edec978bccc8bc250f41d11a2ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 352, + 116, + 432, + 178 + ], + "blocks": [ + { + "bbox": [ + 352, + 116, + 432, + 178 + ], + "lines": [ + { + "bbox": [ + 352, + 116, + 432, + 178 + ], + "spans": [ + { + "bbox": [ + 352, + 116, + 432, + 178 + ], + "type": "image", + "image_path": "27ca382456b6d0cd7090e57af860a5f4ea2ab26f137ca9238f5f68a24f71e8df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 171, + 182, + 261, + 244 + ], + "blocks": [ + { + "bbox": [ + 171, + 182, + 261, + 244 + ], + "lines": [ + { + "bbox": [ + 171, + 182, + 261, + 244 + ], + "spans": [ + { + "bbox": [ + 171, + 182, + 261, + 244 + ], + "type": "image", + "image_path": "eb00436f24d07ccbcb894ac30efe77179fb4a6a1ba28b7fa8f78b016b60ac088.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 289 + ], + "type": "text", + "content": "Fig. 4: Frozen encoder results. Using an encoder pre-trained by depth significantly outperforms one with random weights and one for ImageNet classification. Note that in this experiment, ImageNet pre-training performs worse than random initialization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 267, + 184, + 346, + 244 + ], + "blocks": [ + { + "bbox": [ + 267, + 184, + 346, + 244 + ], + "lines": [ + { + "bbox": [ + 267, + 184, + 346, + 244 + ], + "spans": [ + { + "bbox": [ + 267, + 184, + 346, + 244 + ], + "type": "image", + "image_path": "18f43b5f8d558ab010a601bfabcf3977a52c16938b73db7e8f483f78a4647502.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 352, + 182, + 432, + 244 + ], + "blocks": [ + { + "bbox": [ + 352, + 182, + 432, + 244 + ], + "lines": [ + { + "bbox": [ + 352, + 182, + 432, + 244 + ], + "spans": [ + { + "bbox": [ + 352, + 182, + 432, + 244 + ], + "type": "image", + "image_path": "2215ad626162615c9f57b4aed9c67f819e705da8ddc1743bb6a457f79c316238.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 430 + ], + "type": "text", + "content": "Different forms of depth supervision. Pre-training quality depends on the source of supervision. Training on monocular videos involves minimizing reprojection error, which requires joint estimation of depth and pose. Since pose estimation relies on sufficiently distinctive textures (large eigenvalues of the structure tensor of image gradients), the supervision signal is sparse. Conversely, with stereo images, one may omit the pose network when training. With depth sensors, training losses minimize error w.r.t. dense or semi-dense measured depth, offering stronger supervision. Tab. 2 shows that supervising with Lidar is the best, followed by stereo, and monocular video - all improving over ImageNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 434, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 578 + ], + "type": "text", + "content": "Frozen encoder. We freeze pre-trained encoders, and fine-tune the decoder only, testing the ability of features from pre-trained encoders to capture semantics. With both ResNet18 and ResNet50, encoder pre-trained for depth significantly outperforms random initialization and ImageNet pre-trained (Fig. 4). It is surprising that ImageNet pre-training is detrimental in this case (after a grid search over learning rates): worse than fixed random weights. This suggests that while classification is a semantic task, it removes semantic information about the scene due to the object-centric bias in datasets. ImageNet pre-training tends to favor image-level features, that may not capture object shape, making finetuning the decoder difficult for segmentation. We conjecture that these uncontrolled biases in ImageNet pre-training cause difficulties in directly predicting segmentation without fine-tuning the encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Initializing with a pre-trained encoder only. To eliminate the effect of the depth-initialized decoder and only test the encoder, we replace the decoder by 'fresh' randomly initialized weights and fine-tune the whole network. Tab. 3 shows that depth pre-training outperforms ImageNet when the effect of pretraining is isolate to the encoder. This is also supported by neural activations in Fig. 6 where the regions activated after depth pre-training align well with semantic boundaries. Nonetheless, the decoder does play a role in segmentation" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 117, + 289, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 289, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 289, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 289, + 223 + ], + "type": "table", + "html": "
Pre-trainingmIoUP.Acc
ViT-1 Res. 18None41.3570.75
ImageNet45.1572.39
Depth (encoder only)46.6975.04
ViT-1 Res. 50None44.6673.37
ImageNet44.6573.06
Depth (encoder only)46.9973.57
ViT-1 Res. 18ImageNet57.5381.48
Depth (encoder only)58.1281.94
", + "image_path": "597977ce0ec739de430852cf341be3ed04218794acff87cfc543c836aff5f8f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 228, + 291, + 305 + ], + "lines": [ + { + "bbox": [ + 132, + 228, + 291, + 305 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 291, + 305 + ], + "type": "text", + "content": "Table 3: Initializing with depth encoder and random decoder. Initializing with the depth encoder and a random decoder outperforms ImageNet initialization, but is worse than initializing with both encoder and decoder from the depth network (see Tab. 1)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 114, + 472, + 220 + ], + "blocks": [ + { + "bbox": [ + 315, + 114, + 472, + 220 + ], + "lines": [ + { + "bbox": [ + 315, + 114, + 472, + 220 + ], + "spans": [ + { + "bbox": [ + 315, + 114, + 472, + 220 + ], + "type": "image", + "image_path": "ff8224c0f29b407897acb4ca4d8827419e6d3e6ab9ed2a1f3e239dfc142967e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 228, + 482, + 305 + ], + "lines": [ + { + "bbox": [ + 304, + 228, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 482, + 305 + ], + "type": "text", + "content": "Fig.5: Mismatch between object scales. ImageNet models are trained with a fixed input size and objects of interest tend to have a similar scale. However, objects in semantic segmentation dataset vary drastically in scale. Pre-training for depth provides robustness to scale change." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 353 + ], + "type": "text", + "content": "accuracy: Initializing the whole network with depth pre-training still performs the best, an advantage that is not afforded by a classification head." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 461 + ], + "type": "text", + "content": "Results on Vision Transformers. Given that ViTs [12] necessitate extensive pre-training, conducting comprehensive ablation studies on ViTs, as feasible with ResNet, becomes impractical due to data constraints. Thus, we conduct experiments using the DPT [41] depth model trained on a collection of datasets [33] (not including KITTI), and report the best results in Tab. 3 after an exhaustive search for the optimal learning rate. Notably, we identified the optimal learning rate for this experiment to be 5e-8, with larger learning rates yielding suboptimal results. This suggests that DPT inherently provides robust representations sufficient for segmentation, requiring minimal fine-tuning in comparison to CNNs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 544 + ], + "type": "text", + "content": "Cross architecture. To test whether depth pre-training favors our particular choice of architecture, we use the same encoders to initialize DeepLab V3 and follow the same fine-tuning procedure as common practice, i.e., ImageNet initialization. Results are presented in Tab. 1. All pre-trainings significantly improve accuracy compared with random initialization. The depth model trained from scratch provides the same level of performance as ImageNet, while subsequent depth training after ImageNet pre-training leads to further improvements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "Robustness to object scales. Given a fixed resolution, \"primary\" objects in object-centric data tend to have a similar scale. For example, ImageNet models are trained on " + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": ", thus cars typically have a size of 100 to 200 pixels. In KITTI, however, cars appear at different scales, varying from a few to a few hundred pixels. Fig. 5 illustrates the scale mismatch between datasets. On the other hand, depth pre-training can be done in the same domain as segmentation, hence having robustness to object scales. We examine such robustness: Pretraining on one resolution, fine-tuning on another. Since higher resolution images contain smaller scales (yellow box in Fig. 5), pre-training on them should still work on smaller images. Pre-trained on smaller resolutions, however, should work" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 473, + 178 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 473, + 178 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 473, + 178 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 473, + 178 + ], + "type": "image", + "image_path": "eee26b152085dbc0595ff981a9c3c25416c4e113fd32030e2f446c4e1bdf9fb1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 190, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 234 + ], + "type": "text", + "content": "Fig.6: Neural activation and semantic segmentation result. We visualize the neural activation map for a shallow layer of ResNet 18 encoder trained from different initializations, and their corresponding segmentation results. Boundaries are better aligned to semantic boundaries in our model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "type": "text", + "content": "poorly on larger images. Our results validate this conjecture: The former achieves a final mIoU of 48.54 (ImageNet: 45.15) while the latter diverges during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 480, + 390 + ], + "type": "text", + "content": "Neural Activation. We visualize the activations of the ResNet18 encoder by Grad-CAM [43], which was originally designed for classification. We modify it for segmentation by inspecting the gradient of neural response to the summation of predicted labels for pixels instead of one single class label. We visualize shallow layers (before pooling) for high spatial resolution. Fig. 6 shows neural activation maps and segmentation outputs from depth pre-training align with semantic boundaries. This confirms not just the similarities between natural and range image statistics discussed in [22], but also the bias introduced by classification, as activations of ImageNet pre-trained encoder do not resemble object boundaries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 569 + ], + "type": "text", + "content": "Comparison with optical flow. One hypothesis for the effectiveness of depth pre-training is that the process leverages the statistics of natural scenes where simply-connected components of the range map often correspond to semantically consistent regions. Thus, fine-tuning simply aligns the range of two piece-wise smooth functions. We challenge this hypothesis by trying optical flow, which also exhibits a piecewise-smooth range and is obtained by minimizing the same photometric error. We train optical flow on a siamese network with two shared-weight encoders. While both optical flow and depth capture multiply-connected object boundaries (Fig. 7), using encoders pre-trained for optical flow is detrimental. We conjecture that optical flow does not capture the stable inductive bias afforded by the static component of the underlying scene. Specifically, optical flow is compatible with an underlying 3D geometry only when the scene is rigid, but rigidity is not enforced when inferring optical flow. In contrast, depth forces recognition of rigidity and discards moving objects [29-31] as outliers, which then enables isolating them, also beneficial to semantic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "Comparison with other pre-training methods. In Tab. 4, for completeness, we report results on supervised pre-training by semantic segmentation on MS-COCO [34]. Unsurprisingly, pre-training on the same task with additional annotated data yields good performance. Depth estimation, despite not needing additional annotation, yields on-par performance. We also report results on masked autoencoding. We remove random rectangular regions from images, and the network aims to reconstruct the original image. Also known as ' inpainting' [1], it is considered an effective method for feature learning [39]. It yields" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 110, + 315, + 173 + ], + "blocks": [ + { + "bbox": [ + 133, + 110, + 315, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 110, + 315, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 110, + 315, + 173 + ], + "type": "image", + "image_path": "9df5195851d1c662822c6e7075ce10b2f892508de1a2e1aa6866c8f876ac92f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 185, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 185, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 480, + 251 + ], + "type": "text", + "content": "Fig. 7: Depth helps, flow hurts. Although both are pre-trained by minimizing a photometric reconstruction error, monocular depth outperforms optical flow. This stems from the fact that inferring depth from a single image is ill-posed so the network learns inductive priors that are rich in semantics over the structures within a scene. In contrast, any discriminative features will support the correspondence search, so the flow network is not constrained to learning semantics, yielding poor fine-tuning accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 318, + 108, + 469, + 177 + ], + "blocks": [ + { + "bbox": [ + 318, + 108, + 469, + 177 + ], + "lines": [ + { + "bbox": [ + 318, + 108, + 469, + 177 + ], + "spans": [ + { + "bbox": [ + 318, + 108, + 469, + 177 + ], + "type": "table", + "html": "
AllFreeze
mIoU↑mIoU↑
None41.35-41.24-
Flow38.47-2.8832.19-9.05
Depth-Rand46.004.6543.021.78
Depth50.208.8546.535.29
", + "image_path": "a3ba1cced9162a14101219726b1557cc58e81ec12538ee39eb487b228e95e75e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 143, + 305, + 471, + 353 + ], + "blocks": [ + { + "bbox": [ + 130, + 262, + 479, + 296 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 296 + ], + "type": "text", + "content": "Table 4: Comparison with different pre-trainings. Encoder: ResNet50; Reconstruction: by inpainting randomly corrupted regions (masked autoencoding); Supervised Segmentation: trained on MS-COCO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 305, + 471, + 353 + ], + "lines": [ + { + "bbox": [ + 143, + 305, + 471, + 353 + ], + "spans": [ + { + "bbox": [ + 143, + 305, + 471, + 353 + ], + "type": "table", + "html": "
Pre-trainingmIoUP.AccPre-trainingmIoUP.Acc
Supervised Segmentation51.2874.88Contrastive (DINO)44.1971.36
Depth50.9277.34Optical Flow42.7271.80
Reconstruction (MAE)47.1874.16Contrastive (MOCO V2)37.0465.91
", + "image_path": "bc4d4dd71127875d57711a8707f648ac1cc171130c1a72c04a40c95534e5283e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 375, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 460 + ], + "type": "text", + "content": "inferior performance compared with depth. We conjecture that this is due to artificial rectangular masking which does not respect the natural image statistics, while monocular depth estimation yields occluded regions that border on objects' silhouettes. We also provide results initializing the encoder with contrastive learning (MOCO V2 [9] and DINO [4]). While they also remove bias in human annotation for classification, similar to supervised classification, they are still prone to the inherent inductive bias in pre-training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 477, + 383, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 383, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 383, + 490 + ], + "type": "text", + "content": "4.2 Full-scale Fine-tuning on the Whole Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "type": "text", + "content": "We now extend experiments to fine-tuning on full-scale datasets, aiming to validate the findings drawn from the controlled experiments conducted on KITTI." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "type": "text", + "content": "Cityscapes [10] contains 2975 training and 500 validation images. Each image has a resolution of " + }, + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 130, + 522, + 481, + 666 + ], + "type": "text", + "content": " densely labeled into 19 semantic classes. The dataset also has 20000 unlabeled stereo pairs with a disparity map, converted to depth via focal length and camera baseline. Like KITTI, Cityscapes is also an outdoor driving dataset. Here we minimize an L1 loss between depth estimates and depth computed from stereo. We modify the prediction head of DeepLabV3 to train for depth, then re-initialize the last layer of the decoder for semantic segmentation. We are unable to reproduce the original numbers. For a fair comparison, we retrained [6] using the discussed pre-training methods and finetuned under the same training protocol i.e. batch size, augmentations, schedule, etc. Tab. 5 summarizes the outcomes. We present not only the most favorable results from an extensive training process but also results from a controlled approach" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 168, + 454, + 248 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Segmentation accuracy on Cityscapes. Similar to KITTI, pre-training for depth improves segmentation accuracy. Interestingly, under the controlled settings with limited data augmentation and fewer fine-tuning epochs, the model achieves higher performance when pre-training on cropped " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " patches." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 168, + 454, + 248 + ], + "lines": [ + { + "bbox": [ + 160, + 168, + 454, + 248 + ], + "spans": [ + { + "bbox": [ + 160, + 168, + 454, + 248 + ], + "type": "table", + "html": "
FullControlled
Training mIoUP.Acc.Validation mIoUP.Acc.Training mIoUP.Acc.Validation mIoUP.Acc.
None76.1095.4163.9793.7673.8295.2660.4393.07
ImageNet81.8496.4670.4195.7576.7095.7161.8093.40
Depth83.4696.8073.1795.2477.4295.8262.5793.46
Depth-cropped86.8097.4172.2295.0179.9096.2465.0994.00
", + "image_path": "103fa43a073919402ccb77ef3623b46d0a462f45d0a82cf4f1aac6060019833b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 171, + 262, + 261, + 324 + ], + "blocks": [ + { + "bbox": [ + 171, + 262, + 261, + 324 + ], + "lines": [ + { + "bbox": [ + 171, + 262, + 261, + 324 + ], + "spans": [ + { + "bbox": [ + 171, + 262, + 261, + 324 + ], + "type": "image", + "image_path": "d182751251ca513e3f0bbac454397f6dcd95b2a6ddf6f5bbb99c3572735718bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 267, + 261, + 347, + 324 + ], + "blocks": [ + { + "bbox": [ + 267, + 261, + 347, + 324 + ], + "lines": [ + { + "bbox": [ + 267, + 261, + 347, + 324 + ], + "spans": [ + { + "bbox": [ + 267, + 261, + 347, + 324 + ], + "type": "image", + "image_path": "4c04280ee8bdbcc57a030909706e06060d6b1e9dbe20301af63685ff7d8299cf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 352, + 261, + 432, + 324 + ], + "blocks": [ + { + "bbox": [ + 352, + 261, + 432, + 324 + ], + "lines": [ + { + "bbox": [ + 352, + 261, + 432, + 324 + ], + "spans": [ + { + "bbox": [ + 352, + 261, + 432, + 324 + ], + "type": "image", + "image_path": "7283be2ba11d177f7d65de6364b462a752685b4ce1cd8dc61e75baa6f9ecbc3c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 171, + 327, + 261, + 390 + ], + "blocks": [ + { + "bbox": [ + 171, + 327, + 261, + 390 + ], + "lines": [ + { + "bbox": [ + 171, + 327, + 261, + 390 + ], + "spans": [ + { + "bbox": [ + 171, + 327, + 261, + 390 + ], + "type": "image", + "image_path": "aa256eaaafae5a18d3feaeb31da492f1ec839e5bb93d98f4c12e5685d519c912.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 401, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 424 + ], + "type": "text", + "content": "Fig. 8: Results on NYU-V2. Similar to KITTI, initializing with depth pre-trained weights trains faster and significantly improves semantic segmentation accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 267, + 327, + 347, + 390 + ], + "blocks": [ + { + "bbox": [ + 267, + 327, + 347, + 390 + ], + "lines": [ + { + "bbox": [ + 267, + 327, + 347, + 390 + ], + "spans": [ + { + "bbox": [ + 267, + 327, + 347, + 390 + ], + "type": "image", + "image_path": "3569bf16a9c6073a3cceb31301edae89cf51e6cf15f321966f2d7ebe15d56dfb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 352, + 327, + 432, + 390 + ], + "blocks": [ + { + "bbox": [ + 352, + 327, + 432, + 390 + ], + "lines": [ + { + "bbox": [ + 352, + 327, + 432, + 390 + ], + "spans": [ + { + "bbox": [ + 352, + 327, + 432, + 390 + ], + "type": "image", + "image_path": "6b834b614406ec5fb752514a7fda4cd82b2307f15e6bf88d73676886be64362b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 472 + ], + "type": "text", + "content": "with restricted data augmentations and fewer training iterations. Remarkably, improvements remain consistent across both scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "One may argue that since depth and semantic segmentation maps are both piece-wise smooth, adapting from depth is naturally easy if the model is aware of each pixel's relative position in the image. In order to test this statement, instead of pre-training for depth on the full image, we train depth on randomly cropped " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " patches, and the model has no spatial awareness of the position of a patch in the image, so depth is purely estimated by local information. This practice (Depth-cropped) surprisingly improves semantic segmentation results under controlled settings, showing that using depth as pre-training goes beyond a simple mapping from one smooth function to another. Interestingly, this approach significantly improves training accuracy in the full setting but leads to a slight reduction in validation accuracy, suggesting a potential issue of overfitting. Future research is necessary to delve into this intriguing phenomenon. Another noteworthy observation is that when pre-trained using depth data, the model exhibits superior performance with a higher initial learning rate of 0.1, as opposed to 0.01 used for ImageNet initialization. Conversely, employing an initial learning rate of 0.1 with ImageNet weights can be detrimental and may" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "result in divergence. These findings suggest that pre-training the network with depth estimation may lead to a smoother local loss landscape." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 283 + ], + "type": "text", + "content": "NYU-V2 [37] is an indoor dataset that contains 795 densely annotated images for training and 654 for testing. There are also 407024 unannotated frames with synchronized depth images captured by a Microsoft Kinect. Since the main hypothesis is agnostic to how depth is learned, we pre-trained for depth using ground-truth as supervision. Unlike outdoor driving, which commonly features sky on top, and road and vehicles in the middle of the image with largely planar camera motion, indoor scenes are characterized by more complex layouts with 6 DoF camera motion, yielding images that are even less likely to resemble the object-centric ones commonly observed in classification datasets. This may be why initializing the model with depth pre-trained weights significantly improves semantic segmentation accuracy with both ResNet18 and ResNet50 (see Fig. 8). Note that pre-training by depth yields faster convergence, similar to KITTI." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 300, + 432, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 432, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 432, + 312 + ], + "type": "text", + "content": "4.3 Out-of-domain Transfer from Large-scale Pre-training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 486 + ], + "type": "text", + "content": "One may conjecture that the advantage of pre-training with monocular depth stems from training within the same domain as downstream semantic segmentation. In practice, in-domain data collection for depth pre-training is indeed ideal since it does not require human labeling. In a scientific context, we are interested in testing this conjecture by investigating the transferability of depth models to tasks outside their original domain. However, common depth datasets are considerably small in scale, making fair comparisons with popular pre-trained models, e.g. MAE [18] and DINO v2 [38] that are trained on millions or even billions of images, infeasible. Fortunately, recent methods [33,41] suggest an alternative approach for learning monocular depth by training for relative depth instead of absolute depth, allowing for the integration of multiple data sources during training. Leveraging Depth Anything [55], a depth model trained on such scaled-up mixed datasets, we fine-tune for segmentation on three out-of-domain downstream datasets: ADE20k [59], PascalVOC [13], and CityScapes [10]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": "The PascalVOC dataset comprises 10,582 fully annotated images for training purposes and an additional 1,449 for testing, covering a variety of 20 foreground object classes. On the other hand, ADE20k, a sizable dataset, includes 20,210 training images and 2,000 testing images across 150 different classes. The resolution of PascalVOC and ADE20k is " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "896 \\times 896" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " respectively. On both datasets, we try both fine-tuning the whole network and linear probing with a frozen encoder. Reported in Tab. 6, the results align with the findings on ViTs on the KITTI dataset (in Tab. 3), in which case the pre-training data also originates from out-of-domain sources. Notably, consistent improvements in semantic segmentation performance are observed across all datasets compared to the baseline initialization DINO v2 [38]. It is anticipated that the improvement achieved with a frozen encoder is more substantial than when fine-tuning the entire network, which is consistent with our previous results. This trend underscores that, when trained at a scale, depth models exhibit robust transferability to novel downstream data domains, just as other pre-training methods." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 168, + 478, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 6: Out-of-domain transfer with large-scale pre-trainings. We compare Depth Anything with MAE and DINO v2 for semantic segmentation, reporting results (in mIoU) on both fine-tuning (ft) and linear probing (lin.). Depth Anything improves semantic segmentation across all datasets and settings. *: with DINO v2 initialization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 168, + 478, + 228 + ], + "lines": [ + { + "bbox": [ + 137, + 168, + 478, + 228 + ], + "spans": [ + { + "bbox": [ + 137, + 168, + 478, + 228 + ], + "type": "table", + "html": "
Pre-training# of pre-train dataADE20kPascalVOCCityScapes
ftlin.ftlin.ftlin.
MAE [18]1.28 million53.649.0-67.6-58.4
DINO v2 [38]142 million58.147.786.586.382.771.3
Depth Anything [55]63.5 million*59.752.387.787.384.874.8
", + "image_path": "bf3f2d532c0873a74629d32038946e500056a67cf818e98415783513ce78fd78.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 247, + 217, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 217, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 217, + 259 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 437 + ], + "type": "text", + "content": "Inferring depth only requires multiple images (e.g. videos or multiple viewpoints) of the same scene [25] or range sensing, both do not require human-induced priors and bias, unlike semantic tasks that rely entirely on induction: We can associate a label to an image because that image has something in common with some other image, portraying a different scene, that some annotator attached a particular label to. That inductive chain has to go through the head of human annotators, who are biased in ways that cannot be easily quantified and controlled. Depth from binocular or motion imagery does not require induction and can be performed ab-ovo. Learning a monocular model from such supervision eliminates the implicit selective bias from human annotators, yet our findings validate the main hypothesis that the inductive bias learned from such a \"human-free\" process transfers well to the downstream semantic segmentation task. Of course, if different supervisions are available, either for semantic [11], segmentation [27], or both [28], we want to incorporate that information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "text", + "content": "Our hypothesis and findings are agnostic to how depth is attributed to a single image: One can perform pre-training using monocular videos, stereo, structure light, LIDAR, or even human guidance [58]. One of our pre-training minimizes the photometric reprojection error, used by many predictive and generative approaches. However, an unstructured displacement field is in general not compatible with a rigid motion. Only if this displacement field has the structure of an epipolar transformation [44] is the prediction task forced to encode the 3D structure of the scene. This may explain why video prediction is not as effective for pre-training despite many attempts [24,32,46,53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "One limitation of monocular depth estimation is that it may require a calibrated camera, so one cannot use generic videos harvested from the web. There is nothing in principle preventing us from using uncalibrated cameras, simply by adding the calibration matrix " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " to the nuisance variables. While the necessary conditions for full Euclidean reconstruction are rarely satisfied in consumer videos (for instance, they require cyclo rotation around the optical axis, not just panning and tilting), the degrees of freedom that cannot be reconstructed are moot as they do not affect the reprojection. Moreover, recent progress in training for relative depth from mixed data sources [33, 41, 55] shows the potential to unlock a virtually unlimited volume of training data, warranting future research." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by ONR N00014-22-1-2252 and ARO W911NF-17-1-0304." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 188, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "type": "text", + "content": "1. Bertalmio, M., Sapiro, G., Caselles, V., Ballester, C.: Image inpainting. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 417-424 (2000)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "type": "text", + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "text", + "content": "3. Caron, M., Bojanowski, P., Mairal, J., Joulin, A.: Unsupervised pre-training of image features on non-curated data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "type": "text", + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834-848 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "type": "text", + "content": "6. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 480, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 410 + ], + "type": "text", + "content": "7. Chen, M., Artières, T., Denoyer, L.: Unsupervised object segmentation by redrawing. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "type": "text", + "content": "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "type": "text", + "content": "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 466, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 466, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 134, + 466, + 480, + 510 + ], + "type": "text", + "content": "10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 510, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 480, + 532 + ], + "type": "text", + "content": "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR09 (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 533, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 480, + 576 + ], + "type": "text", + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "text", + "content": "13. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88(2), 303-338 (Jun 2010)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "type": "text", + "content": "14. Fei, X., Wong, A., Soatto, S.: Geo-supervised visual depth prediction. IEEE Robotics and Automation Letters 4(2), 1661-1668 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth prediction (October 2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "17. Goldblum, M., Souri, H., Ni, R., Shu, M., Prabhu, V., Somepalli, G., Chattopadhyay, P., Ibrahim, M., Bardes, A., Hoffman, J., et al.: Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 215 + ], + "type": "text", + "content": "18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "19. Hoyer, L., Dai, D., Chen, Y., Köring, A., Saha, S., Van Gool, L.: Three ways to improve semantic segmentation with self-supervised depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11130-11140 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "20. Hoyer, L., Dai, D., Wang, Q., Chen, Y., Van Gool, L.: Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. arXiv preprint arXiv:2108.12545 [cs] (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "21. Huang, J., Lee, A.B., Mumford, D.: Statistics of range images. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662). vol. 1, pp. 324-331. IEEE (2000)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "22. Huang, J., Mumford, D.: Statistics of natural images and models. In: Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149). vol. 1, pp. 541-547. IEEE (1999)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "23. Jiang, H., Larsson, G., Shakhnarovich, M.M.G., Learned-Miller, E.: Self-supervised relative depth learning for urban scene understanding. In: Proceedings of the European conference on computer vision (eccv). pp. 19-35 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 434 + ], + "type": "text", + "content": "24. Jin, B., Hu, Y., Tang, Q., Niu, J., Shi, Z., Han, Y., Li, X.: Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4554-4563 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "25. Julesz, B.: Binocular depth perception without familiarity cues: Random-dot stereo images with controlled spatial and temporal properties clarify problems in stereopsis. Science 145(3630), 356-362 (1964)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 375, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 375, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 375, + 479 + ], + "type": "text", + "content": "26. Julesz, B.: Foundations of cyclopean perception. (1971)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 521 + ], + "type": "text", + "content": "27. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 566 + ], + "type": "text", + "content": "28. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "29. Lao, D., Hu, Z., Locatello, F., Yang, Y., Soatto, S.: Divided attention: Unsupervised multi-object discovery with contextually separated slots. arXiv preprint arXiv:2304.01430 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "30. Lao, D., Sundaramoorthi, G.: Minimum delay moving object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4250-4259 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Lao, D., Sundaramoorthi, G.: Extending layered models to 3d motion. In: Proceedings of the European conference on computer vision (ECCV). pp. 435-451 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Lao, D., Zhu, P., Wonka, P., Sundaramoorthi, G.: Flow-guided video inpainting with scene templates. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14599-14608 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "text", + "content": "33. Lasinger, K., Ranftl, R., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. arXiv preprint arXiv:1907.01341 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 481, + 228 + ], + "type": "text", + "content": "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 262 + ], + "type": "text", + "content": "35. Liu, T.Y., Agrawal, P., Chen, A., Hong, B.W., Wong, A.: Monitored distillation for positive congruent depth completion. In: European Conference on Computer Vision. pp. 35-53. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 481, + 296 + ], + "type": "text", + "content": "36. Lopez-Rodriguez, A., Busam, B., Mikolajczyk, K.: Project to adapt: Domain adaptation for depth completion from noisy and sparse sensor data. In: Proceedings of the Asian Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 297, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 318 + ], + "type": "text", + "content": "37. Nathan Silberman, Derek Hoiem, P.K., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 385 + ], + "type": "text", + "content": "38. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024), https://openreview.net/forum?id=a68SUt6zFt" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "text", + "content": "39. Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T., Efros, A.A.: Context encoders: Feature learning by inpainting. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2536-2544 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 481, + 453 + ], + "type": "text", + "content": "40. Ramirez, P.Z., Tonioni, A., Salti, S., Stefano, L.D.: Learning across tasks and domains. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8110-8119 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 485 + ], + "type": "text", + "content": "41. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 531 + ], + "type": "text", + "content": "42. Saha, S., Obukhov, A., Paudel, D.P., Kanakis, M., Chen, Y., Georgoulis, S., Van Gool, L.: Learning to relate depth and semantics for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8197-8207 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 531, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 481, + 575 + ], + "type": "text", + "content": "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 481, + 609 + ], + "type": "text", + "content": "44. Sundaramoorthi, G., Petersen, P., Varadarajan, V., Soatto, S.: On the set of images modulo viewpoint and contrast changes. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 832-839. IEEE (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 481, + 632 + ], + "type": "text", + "content": "45. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method. arXiv preprint physics/0004057 (2000)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "46. Wang, Y., Wu, J., Long, M., Tenenbaum, J.B.: Probabilistic video prediction from noisy data with a posterior confidence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10830-10839 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "On Viability Monocular Depth Pre-train. Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 534 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "47. Wong, A., Cicek, S., Soatto, S.: Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 1495-1502 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "48. Wong, A., Fei, X., Hong, B.W., Soatto, S.: An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters 6(2), 3120-3127 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "49. Wong, A., Fei, X., Tsuei, S., Soatto, S.: Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters 5(2), 1899-1906 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "50. Wong, A., Soatto, S.: Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5644-5653 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "51. Wong, A., Soatto, S.: Unsupervised depth completion with calibrated backprojection layers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12747-12756 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "52. Wu, Y., Liu, T.Y., Park, H., Soatto, S., Lao, D., Wong, A.: Augundo: Scaling up augmentations for monocular depth completion and estimation. In: European Conference on Computer Vision. Springer (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "53. Wu, Y., Gao, R., Park, J., Chen, Q.: Future video synthesis with object motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5539-5548 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "54. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "55. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "56. Yang, Y., Wong, A., Soatto, S.: Dense depth posterior (ddp) from single image and sparse range. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3353-3362 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "57. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3712-3722 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "text", + "content": "58. Zeng, Z., Wang, D., Yang, F., Park, H., Soatto, S., Lao, D., Wong, A.: Wordepth: Variational language prior for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9708-9719 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "59. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 633-641 (2017)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "D Lao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_content_list.json b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f2ba0f9a090811c9a5296fe807a4e12c95c3267 --- /dev/null +++ b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "text_level": 1, + "bbox": [ + 250, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Koh Jun Hao*, Sy-Tuyen Ho*, Ngoc-Bao Nguyen, and Ngai-Man Cheung", + "bbox": [ + 236, + 212, + 766, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Singapore University of Technology and Design (SUTD)", + "bbox": [ + 312, + 239, + 689, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{junhao_koh,ngaiman_cheung}@sutd.edu.sg", + "bbox": [ + 349, + 255, + 651, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Skip connections are fundamental architecture designs for modern deep neural networks (DNNs) such as CNNs and ViTs. While they help improve model performance significantly, we identify a vulnerability associated with skip connections to Model Inversion (MI) attacks, a type of privacy attack that aims to reconstruct private training data through abusive exploitation of a model. In this paper, as a pioneer work to understand how DNN architectures affect MI, we study the impact of skip connections on MI. We make the following discoveries: 1) Skip connections reinforce MI attacks and compromise data privacy. 2) Skip connections in the last stage are the most critical to attack. 3) RepVGG, an approach to remove skip connections in the inference-time architectures, could not mitigate the vulnerability to MI attacks. 4) Based on our findings, we propose MI-resilient architecture designs for the first time. Without bells and whistles, we show in extensive experiments that our MI-resilient architectures can outperform state-of-the-art (SOTA) defense methods in MI robustness. Furthermore, our MI-resilient architectures are complementary to existing MI defense methods. Our project is available at https://Pillowkoh.github.io/projects/RoLSS/", + "bbox": [ + 261, + 318, + 743, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Model Inversion $\\cdot$ Skip Connection $\\cdot$ Model Inversion Resilient Architecture", + "bbox": [ + 259, + 580, + 740, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 635, + 375, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As deep neural networks (DNNs) see growing deployment across various applications like face recognition [3,6,7,10,14,18,21,27,34] and healthcare [6,7,20,22, 23,34], concerns about the privacy implications of DNNs are on the rise. Many DNNs are trained on private and sensitive datasets. There is an increasing concern of potential leakage of information of these private training samples through malicious exploitation of the model.", + "bbox": [ + 212, + 666, + 787, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One particular privacy threat that has garnered growing attention is Model Inversion (MI) attacks. In MI attacks, adversaries seek to reconstruct private training samples by exploiting vulnerabilities in the model. For instance, an adversary with access to a face recognition model may abuse it to reconstruct the", + "bbox": [ + 212, + 756, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Co-first authors", + "bbox": [ + 232, + 825, + 354, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "private facial images of individuals from the model's training dataset. Following previous works [4, 11, 25, 39], we focus on reconstruction of images and use the face recognition as a running example.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research gap. Recently, there is an increasing interest to study MI and to understand the feasibility and extent of reconstructing private training samples from DNNs, from the MI attack and MI defense perspectives. However, previous studies have overlooked DNN architecture, and there is a lack of study to understand how DNN architecture designs affect MI (Tab. 1). In particular, MI has been formulated as an optimization problem to seek an image similar to that of an identity in the private training dataset. Commonly, the MI optimization", + "bbox": [ + 212, + 198, + 787, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "objective is formulated as maximization of likelihood under the model being attacked (target model). Several improved MI objectives have been proposed recently, e.g. logit maximization [25,38]. Meanwhile, various regularizations on the MI objective have been studied to improve the effectiveness of MI, e.g. prior loss to penalize unrealistic images [39]. In addition, various distributional priors leveraging generative models trained on public datasets have been proposed to guide the inversion (optimization) process during MI attacks [1, 4, 31, 38, 39]. Furthermore, regularizations on the training objective of the target model to reduce the correlation exploited by MI have been studied as methods to defend against MI attacks [26,32]. However, there is a lack of study to understand the effect of DNN architecture design on MI.", + "bbox": [ + 212, + 305, + 495, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: There is a lack of study to understand how DNN architecture designs affect MI. Previous MI studies are DNN architecture-agnostic, focusing on MI objective, effect of regularizing MI objective, effect of distributional prior based on generative modelling, and regularization on the training objective of the target model. Our work is a pioneer study to understand how DNN architectures affect MI attacks.", + "bbox": [ + 501, + 308, + 787, + 446 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/7af01c354325ac79c2a1cb78f98bc0acfdd527fc83e9fdd8fc68b6cc85473b4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MI objectiveEffect of regularizing MI objectiveEffect of distributional prior to guide MIRegularization on the training objective of target modelEffect of DNN architecture design on MI
MI [9]✓
GMI [39]✓✓
KEDMI [4]✓✓
VMI [31]✓
MIRROR [1]✓
PPA [28]✓
LOMMA [25]✓
PLGMI [38]✓✓
RLBMI [11]✓
BREPMI [15]✓
MID [32]✓
BiDO [26]✓
Ours✓
", + "bbox": [ + 506, + 449, + 782, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address this research gap and conduct the first study to understand how DNN architecture designs affect MI. We put our focus on skip connections [12, 13], a fundamental network design that facilitates the training of very deep neural networks. Skip connections mitigate the vanishing gradient problem during the training stage [12]. Meanwhile, many state-of-the-art (SOTA) MI attacks [4, 25, 28, 38, 39] require the use of gradients to guide the reconstruction of private training samples during the inversion stage. We hypothesize that skip connections facilitate flowing of gradient during inversion, reinforcing MI attacks and posing a considerable vulnerability to data privacy in DNNs (Fig. 1). To validate our hypothesis, we carefully design experiments to single out the effect of skip connections on MI attack performance. Our extensive experiments on SOTA networks (ResNet [12], DenseNet [13], MaxViT [30], EfficientNet [29]) against SOTA MI attacks (PPA [28], PLG-MI [38], LOMMA [25], KEDMI [4])", + "bbox": [ + 212, + 643, + 789, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d0d7eaaae399a5522b793a9f363074c537cb71cc64fc6fd5395cc76cd2a3b9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 143, + 769, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d5e20aa343d75e7ab7de49c0c3d3d5d195f26f9576bce1e492b99cd3339c965.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 241, + 769, + 342 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/15405a9b6068ac98d634a14c58456dd8835a9f439dc90027923e75fb59104755.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 343, + 769, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 1: (I) Illustration of MI attack on ResNet-like architecture (Sec. 3.1). This figure depicts the MI attack framework for SOTA white-box MI attacks [1,4,25, 28,31,38,39], which leverage a generative model $G(\\cdot)$ to exploit the target model via gradient descent and backpropagation. Specifically, for each iteration, $\\tilde{x} = G(w)$ is fed into the target model in the forward pass, and MI loss $\\mathcal{L}$ is computed. In the backward pass, gradients of $\\mathcal{L}$ are computed and back-propagated to obtain $\\partial \\mathcal{L} / \\partial w$ , which is used to update $w$ to achieve reconstruction of private training data. (II) Additive Skip Connection (Sec. 3.1). During MI attacks, skip connections allow gradients to bypass the residue module and enhance backpropagation. We hypothesize that this reinforces MI attacks. (III) Our study on skip connection removal (Sec. 3.2 and Sec. 3.3). To validate our hypothesis that skip connections could reinforce MI, we study the effect of skip connections on MI by removing skip connections within various stages of the target model. We study both additive and concatenative skip connections. (IV) Results of stage-wise skip connection removal study (Sec. 3.2 and Sec. 3.3). The sub-figures show that skip connections have a considerable effect on MI. For both additive and concatenative skip connections, we observe that removal of skip connections result in considerable degradation of MI attack accuracy. Furthermore, we observe that skip connections in the last stage have the most significant effect on MI. Best viewed in color with zooming in. consistently show that skip connections reinforce MI attacks in DNNs. Furthermore, we find that skip connections in the last stage have the most significant effect to MI attacks.", + "bbox": [ + 215, + 484, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate the MI vulnerability caused by skip connection, we analyze, RepVGG [5], an established reparameterization technique which converts a multibranch training-time architecture with skip connections to a plain, VGG-like", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inference-time architecture with skip connection removed. However, our analysis shows that, while RepVGG could enable an inference-time network without any skip connection, the gradients during MI attacks on this inference-time network are the same as that on the multi-branch training-time architecture with skip connections. Therefore, RepVGG could not mitigate the vulnerability.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To bridge the existing gap, we propose MI-resilient architecture designs based on our own findings. Specifically, as we find that the last stage's skip connections have the most significant effect to MI, we propose Removal of Last Stage Skip-Connection (RoLSS) as MI-resilient architecture designs. As our designs remove only the last stage's skip connections and keep other stages' skip connections intact, we could keep the impact on natural accuracy small in many cases. Building on top of RoLSS, we propose Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS) to recover the model's natural accuracy while maintaining competitive MI robustness. Our contributions are:", + "bbox": [ + 212, + 222, + 787, + 358 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct a pioneer study to understand how skip connections impact MI attacks and MI robustness. We design experiments to carefully single out the effect of skip connections on MI attack performance, accounting for the effect of natural accuracy in our analysis. Through extensive experiments spanning 4 SOTA MI attacks and 10 architectures, we validate that skip connections reinforce MI attacks and pose a considerable vulnerability to data privacy in DNNs (Sec. 3).", + "- Notably, we discover that skip connections in the last stage consistently are the most critical to MI attacks (Sec. 3 and Sec. 3.3).", + "- We analyze RepVGG, a well-established reparameterization technique that remove skip connections by decoupling training-time and inference-time architectures. Our analysis reveals that this approach could not mitigate vulnerability to MI attacks (Sec. 4).", + "- Based on our findings, we propose MI resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS). Extensive experiments show that our MI-resilient architectures can outperform SOTA defense methods in MI robustness (Sec. 5)." + ], + "bbox": [ + 225, + 364, + 787, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 657, + 388, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Inversion. The concept of MI was initially studied by Fredrikson et al. [9], who demonstrated that adversaries could employ machine learning to extract genomic and demographic information about patients from a medical imaging model. This work was later extended to facial recognition in [8]. Since then, several MI studies have been conducted to understand the feasibility and extent of reconstructing private training samples from DNNs [11,15,25,26,28,31,32,36, 38,39]. These studies encompass both MI attacks and MI defense perspectives. We summarize notable developments in Tab. 1. See Supp. for further discussion of related work. Despite considerable progress in MI research, there is a lack of study to understand the effect of DNN architecture design on MI.", + "bbox": [ + 212, + 688, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skip connections and DNNs Attacks. Skip connections are recognized as an effective approach to alleviate the vanishing gradient problem, allowing us to train very deep neural networks [12, 13]. There are a few works that study the effect of skip connections to adversarial attacks [2, 33] and backdoor attacks [35]. Differ from existing works, our study is the first to understand how DNN architectures affect model inversion (MI), a growing privacy attack. Our investigation reveals a distinctive aspect: while previous work observed that skip connections aids adversarial robustness [2], our work instead discovers that skip connections reinforce MI attacks. It is important to emphasize that the nature of MI attacks differs significantly from adversarial or backdoor attacks. Particularly, for adversarial attacks, the goal is to deceive the model into making incorrect predictions. For backdoor attacks, the goal is to implant malicious functionality in the model such that the model produces incorrect outputs when a specific attack trigger is present in the input. Importantly, adversarial attacks/backdoor attacks are not privacy attacks and they do not extract sensitive training data information from ML models. Rather, adversarial attacks/backdoor attacks aim to undermine model utility and robustness. Our work is the first to study the implications of skip connection on data privacy of ML models through the MI attacks.", + "bbox": [ + 212, + 146, + 787, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 An Investigation on the Skip Connection Vulnerability to Model Inversion Attacks", + "text_level": 1, + "bbox": [ + 214, + 452, + 781, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Skip connections and MI attacks", + "text_level": 1, + "bbox": [ + 214, + 497, + 532, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first discuss the potential effect of skip connections on MI in this sub-section. Then, the effect is validated in Sec. 3.2 and Sec. 3.3.", + "bbox": [ + 212, + 518, + 782, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MI and gradients. MI attacks are a data privacy threat. For a DNN model $T$ trained on a private training dataset $\\mathcal{D}_{priv}$ , the adversary tries to exploit sensitive training data $\\mathcal{D}_{priv}$ via the trained model $T$ . In most works, MI is formulated as the reconstruction of an input $\\tilde{x}$ which is most likely classified into an identity $y$ by $T$ . The model $T$ subject to MI attacks is called target model. We focus on white-box MI attack, which is the most popular and powerful MI attack [1,4,25,28,38,39]. Specifically, we follow previous works and assume attackers have access to the parameters, architectures, and outputs of the models [1,4,25,28,38,39].", + "bbox": [ + 212, + 549, + 787, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reconstruct a high-dimensional image $\\tilde{x}$ , some distributional priors have been proposed in SOTA MI to constrain the search space [4, 39]. The distributional prior is commonly encoded by a generative model $G(w)$ trained on a public dataset $\\mathcal{D}_{pub}$ which has no class intersection with $\\mathcal{D}_{priv}$ . MI attacks are commonly formulated as the following optimization:", + "bbox": [ + 212, + 685, + 787, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {*} = \\underset {w} {\\arg \\min } \\mathcal {L} (w; y, T) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 768, + 785, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}(w; y, T)$ is the MI loss which guides reconstruction of $\\tilde{x} = G(w)$ that is most likely to be classified by model $T$ as identity $y$ . Commonly, negative log-likelihood is used: $\\mathcal{L}(w; y, T) = -\\log \\mathbb{P}_T(y|G(w))$ , while other losses have been", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "proposed, e.g., logit-based [25]. In addition, other regularization can be included in $\\mathcal{L}$ , e.g. prior loss [39].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importantly, to solve the optimization in Eq. 1, gradient descent and back propagation are used by most SOTA white-box MI attacks [1, 4, 25, 28, 38, 39]: For each iteration, $G(w)$ is fed into $T$ in the forward pass, and $\\mathcal{L}$ is computed. In the backward pass, gradients of $\\mathcal{L}$ are computed in $T$ and back-propagated to obtain $\\partial \\mathcal{L} / \\partial w$ , which is used to update $w$ by the attackers.", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Skip connections could reinforce MI. Following the above discussion, backpropagation of gradients during MI inversion could have a considerable effect on the MI attack performance. Meanwhile, for conventional DNN training, skip connections are a fundamental architecture design that is effective in mitigating gradient vanishing during backpropagation. We hypothesize that skip connections could facilitate gradient backpropagation during MI attacks and reinforce MI, thereby compromising data privacy.", + "bbox": [ + 212, + 253, + 785, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, in a ResNet-like architecture, there are multiple ResNet blocks. Each ResNet block, with input $z_{i}$ and output $z_{i+1}$ , can be represented as: $z_{i+1} = z_{i} + g_{i}(z_{i})$ , including an additive skip connection and a residual module $g_{i}$ comprising multiple convolutions (Fig. 1). During MI inversion, the gradient backpropagates across a ResNet block as follows:", + "bbox": [ + 214, + 358, + 785, + 429 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial z _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\frac {\\partial z _ {i + 1}}{\\partial z _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\left(1 + \\frac {\\partial g _ {i}}{\\partial z _ {i}}\\right) = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} + \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\frac {\\partial g _ {i}}{\\partial z _ {i}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 439, + 785, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importantly, the first gradient component, $\\frac{\\partial\\mathcal{L}}{\\partial z_{i + 1}}$ , enabled by the skip connection, enhances backpropagation. We hypothesize that this reinforces MI attacks.", + "bbox": [ + 214, + 479, + 784, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Stage-wise skip connection removal study", + "text_level": 1, + "bbox": [ + 214, + 539, + 602, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we validate effect of skip connections on MI.", + "bbox": [ + 212, + 566, + 648, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MI setup. We conduct our analysis on ResNet-101 [12] and DenseNet-121 [13] as target models under the attack setup of SOTA MI attack method PPA [28]. We strictly follow PPA MI setups, where we use FaceScrub [24] as private dataset, $\\mathcal{D}_{priv}$ and attack all IDs as per PPA setup. Following previous SOTA MI works [1,4,15,25,28,31,38,39], we adopt attack accuracy (AttAcc), measured using an evaluation model, as the primary metric for assessing MI performance. Attack accuracy is defined as the percentage of reconstructed images correctly identified by the evaluation model with respect to the target ID. Specific MI attack configuration can be found in the Supp.", + "bbox": [ + 212, + 582, + 785, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct our study by removing skip connections from various stages of the architecture. Each time, skip connections from a specific stage are removed, while those in the remaining stages remain unchanged. Each architecture with removed skip connections is trained using $\\mathcal{D}_{priv}$ in exactly the same settings as the original unaltered architecture. In this study, we focus on two common skip connection mechanisms: additive and concatenative.", + "bbox": [ + 212, + 719, + 785, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additive skip connection removal. We investigate additive skip connections within individual stages of the ResNet-101 architecture. To remove additive skip", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/141fe3034297eaff3e79d290c75e4b95b2ce6129ad35154e61736d044caaed75.jpg", + "table_caption": [ + "Table 2: We strictly follow SOTA PPA [28] for the attack setup and evaluation. Here $\\mathcal{D}_{priv} =$ Facescrub [24], $\\mathcal{D}_{pub} =$ FFHQ [16]. Across architectures both additive and concatenative skip connections, we consistently observe that Skip connections in the last stage are the most critical to MI attacks, resulting in the most degradation in MI attack accuracy. $\\varDelta_{AttAcc}$ represents the reduction in attack accuracy when compared to \"Full\" setting." + ], + "table_footnote": [], + "table_body": "
ArchitectureSkip ConnectionsAttAccΔAttAcc
ResNet-34Full90.78-
Skip-1 Removed83.257.53
Skip-2 Removed77.9212.86
Skip-4 Removed80.6110.17
ResNet-50Full82.76-
Skip-1 Removed73.239.53
Skip-2 Removed78.773.99
Skip-4 Removed68.4414.32
ResNet-101Full83.00-
Skip-1 Removed74.818.19
Skip-2 Removed78.754.25
Skip-4 Removed58.6824.32
ResNet-152Full86.51-
Skip-1 Removed80.356.16
Skip-2 Removed69.0417.47
Skip-4 Removed68.4418.07
", + "bbox": [ + 222, + 229, + 480, + 421 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f8f2c49e8006b5aaf086d2e1612a7a8c94f291c5e9d0ea1ab2eee87cc8af2bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureSkip ConnectionsAttAccΔAttAcc
DenseNet-121Full88.11-
Skip-1 Removed61.7226.39
Skip-2 Removed55.2132.90
Skip-4 Removed24.4863.63
DenseNet-161Full74.86-
Skip-1 Removed55.3819.48
Skip-2 Removed59.2515.61
Skip-4 Removed20.7154.15
DenseNet-169Full77.15-
Skip-1 Removed60.9916.16
Skip-2 Removed51.8625.29
Skip-4 Removed6.7770.38
DenseNet-201Full77.62-
Skip-1 Removed57.4120.21
Skip-2 Removed46.6531.97
Skip-4 Removed20.7156.91
", + "bbox": [ + 514, + 231, + 777, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "connections in ResNet-like architectures, we ensure that outputs from the previous layers are not added to the subsequent layers during the feedforward process, as shown in Fig. 1-III. For ResNet-101, the resulting ResNet block encompasses a convolutional layer, $g_{i}$ , that consists of a single $1 \\times 1$ convolution, followed by a $3 \\times 3$ convolution, and lastly, another $1 \\times 1$ convolution layer, without a shortcut connection linking the input and output of the ResNet block.", + "bbox": [ + 212, + 436, + 787, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concatenative skip connection removal. We investigate concatenative skip connections within individual stages of the DenseNet-121 architecture. We remove the concatenative skip connections in a similar manner as the removal of additive skip connections (The details can be found in the Supp). DenseNet architectures contain DenseBlocks where input features are concatenated with the output features, before being fed into the next DenseBlock. When these features are merged through concatenation, each layer has direct access to the gradients from the loss function and the original input image. To remove these concatenative skip connections, we remove the process of concatenation and only pass the output feature of the current DenseBlock to the subsequent DenseBlock.", + "bbox": [ + 212, + 532, + 787, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skip connections reinforce MI. To benchmark our experiments, we utilize the original unaltered architecture to assess the performance of the modified architectures. For a fair comparison, we consider the strong correlation between natural accuracy and attack accuracy [39]. Thus, we compare these architectures at multiple checkpoints that achieve similar natural accuracy. When presenting our results, we denote the removal of all skip connections in the $N^{th}$ stage as \"Skip-N Removed\". ResNet and DenseNet consist of four stages. We examine removal of skip connections from stages 1, 2, and 4. The result of removing skip connections in stage 3 are not included in our study, as this stage contains many parameters, and its removal leads to a severe reduction in model accuracy.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our findings reveal that architectures with fewer skip connections impede MI attacks, leading to a decrease in MI attack accuracy. As depicted in Fig. 1-IV, both additive and concatenative skip connection studies consistently show that architectures labeled as \"Skip-N Removed\" exhibit low MI attack accuracy compared to the original architecture.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Removal of Last Stage Skip-Connection (RoLSS) is the most critical to MI attacks. Notably, we consistently observe that removing the skip connection in the last stage (i.e., \"Skip-4 Removed\") results in the most degradation in MI attack accuracy. We attribute this to the specific position of skip connections removed within the architecture. During gradient backpropagation in MI attack, gradients in earlier stages depend on those in later stages, as illustrated in Eq. 2. When skip connections in stage 4 (last stage) are removed, the degraded gradients in stage 4 permeate throughout the earlier stages of the architecture, resulting in the most degradation in MI attack accuracy. We further validate this observation on various architectures for both additive and concatenative skip connections in Tab. 2.", + "bbox": [ + 212, + 220, + 787, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Removing the skip connections leads to MI optimization converging with more false positives. As discussed in Eq. 1, MI adversaries aim to identify optimal $w^{*}$ that maximizes the likelihood $\\mathbb{P}_T(y|G(w))$ . We provide this key observation to understand why removing skip connections degrades MI attacks: When the skip connections are removed, latent variables with high likelihood $\\mathbb{P}_T(y|G(w))$ can still be identified by Eq. 1, but many $w^{*}$ are false positives. Consequently, this leads to a notable decrease in the accuracy of MI attacks.", + "bbox": [ + 212, + 388, + 488, + 613 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ea65c72d4324ddfe762a7148718c10e42e680de37994ca1240f528e3113b338.jpg", + "image_caption": [ + "Fig.2: MI convergence analysis. We compare histograms of likelihoods for original architecture (Full) and Removal of Last Stage Skip-Connection (RoLSS) architecture for ResNet-101 under PPA attack." + ], + "image_footnote": [], + "bbox": [ + 524, + 397, + 759, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This observation becomes evident when examining the likelihood distribution of \"Full\" and \"Skip-4 removed\" settings (see Fig. 2), which are similar and both very close to 1. This results imply that, with \"Skip-4 removed\" setting, Eq. 1 could still perform well to seek latent variables $w$ to maximize the likelihood $\\mathbb{P}_T(y|G(w))$ . However, despite the similarity in likelihood distributions, the attack accuracy of the \"Skip-4 removed\" setting is significantly lower than that of the \"Full\" setting (i.e., $24.32\\%$ ). This suggests that, due to the absence of skip connections, the gradients in the \"Skip-4 removed\" setting lead MI adversaries to exploit many $w^*$ that do not correspond to images resembling private data.", + "bbox": [ + 212, + 614, + 787, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Extensive validation of the MI vulnerability of skip connections", + "text_level": 1, + "bbox": [ + 212, + 771, + 785, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct extensive experiments to further validate the impact of skip connections to MI attacks. Tab. 3 summarize all MI setups in our validation ranging 10 architectures, 4 SOTA MI attacks, 3 datasets.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f3dcf08bd9244a8d04c595bd30666c003a8022c30505a72f0d800ebf9c0c279.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 359, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d891b93e06e2ef88e2cc6ddf5015300c639251256bf2c9da4fadb3e10532d1da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 143, + 500, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0d48092f1890a04e609c775d8d855d60b832c62ebddd60651473dfa145f52c11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 642, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f2ac7982cbf1e399cad481edfb8355b99e36686f404ebd70570be8cb4a356fb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 145, + 785, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/78bf9733f5fb78d7247bb0511eeee3abf94a7c01473eb3477f1ba359f24eb08f.jpg", + "image_caption": [ + "Fig. 3: Additional experiments to validate the impact of skip connections on MI attacks across various architecture designs, including networks with additive skip connections (ResNet-34/50/152 [12]), concatenative skip connections (DenseNet-161/169/201 [13]), and compact CNN (EfficientNet-B0 [29]). We strictly follow PPA [28] for MI setups. In all cases, a significant decrease in attack accuracy is consistently observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks." + ], + "image_footnote": [], + "bbox": [ + 215, + 273, + 357, + 386 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f5ec4b482d04bab664b39319660c61c6c74ac9ae32b391a87a266a1c1773723d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 273, + 500, + 386 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c86254f2d8e81304792eab393545d8b0866834bccf2e8c06758960d6e97f5ea5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 273, + 640, + 386 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6bba80b2492caa1fd1d0a0814c7df8c23b14aeec2a7c47891355528fddd2799d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 273, + 785, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experimental Setting. For a fair comparison, we follow the previous MI works [1, 4, 11, 15, 25, 26, 28, 31, 38, 39] to select Evaluation Metrics, Private Dataset, Public Dataset, and Data Preparation Protocol. The details can be found in Tab. 3. Additional details are presented in the Supp.", + "bbox": [ + 212, + 506, + 787, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Skip connections removal. We apply our finding of Removal of Last Stage Skip-Connection (RoLSS) for various architectures.", + "bbox": [ + 212, + 566, + 491, + 626 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following the previous MI works, we adopt natural accuracy (Acc) and Attack Accuracy (AttAcc) as the main evaluation metrics. The detailed description and additional qualitative results are presented in the Supp.", + "bbox": [ + 212, + 628, + 491, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experimental results on various architectures. We note that in Fig. 3, we focus on the range of natural accu", + "bbox": [ + 212, + 734, + 491, + 777 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/939c09d822c0e1e09ca1e1f48f21668e44d9e8679976328d39245d9d7114e9fb.jpg", + "table_caption": [ + "Table 3: The summary of our MI setups. We follow the exact the experiment setups of PPA [28]. For the other MI attacks, we follow the setups in [4, 25] for [4, 25, 39]. In total, our study spans 10 architectures, 4 MI attacks, 3 datasets." + ], + "table_footnote": [], + "table_body": "
ArchitecturesMI AttackPrivate Dataset
ResNet-34/50/101/152 [12]PPA [28]Facescrub [24]
DesnseNet-121/161/169/201 [13]
MaxViT-T [30]
EfficientNet-B0 [29]
ResNet-50/101 [12]Stanford Dogs [17]
DenseNet-121/169 [13]
IR152 [12]KEDMI [4]LOMMA [25]PLG-MI [38]CelebA [19]
", + "bbox": [ + 503, + 662, + 785, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "racy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. The results are consistent with our empirical study in Sec. 3, where the RoLSS of additive skip connections (e.g., ResNet-34/50/152/EfficientNet-B0) reduce the attack accuracy by around", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e854107c6902639fc2e74eb64401f65e2a8ab2fef463a600ff8a80a2bafbb867.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 143, + 750, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2d3795e78fd57b90c992f053dfab6d8ed127a35536e99038647c253541cf9b0a.jpg", + "image_caption": [ + "Fig. 4: Additional experiments on other SOTA MI attacks including KEDMI [4], LOMMA [25], and PLG-MI [38]. We follow the standard setup, where $T = \\mathrm{IR152}$ , $\\mathcal{D}_{priv} = \\mathrm{CelebA}$ , $\\mathcal{D}_{pub} = \\mathrm{CelebA/FFHQ}$ . Across all SOTA MI attacks, a consistent and notable reduction in attack accuracy is observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks.", + "Fig. 5: Additional experiments on the Stanford Dogs [17] dataset as $\\mathcal{D}_{priv}$ . The experiments are conducted under PPA [28] attacks across various architectures, including ResNet-50/101 and DenseNet-121/169. We strictly follow MI setups in PPA." + ], + "image_footnote": [], + "bbox": [ + 243, + 330, + 750, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "10% to 15% while the RoLSS of concatenative skip connections (e.g., DenseNet-161/169/201) reduce the attack accuracy by around 30% to 35%. Overall, in all cases including addictive connections (ResNet-34/50/152), concatenative skip connections (DenseNet-161/169/201), compact CNN (EfficientNet-B0), we consistently observe the significant drops in MI attack accuracy, ranging from 10% to 35%. This results further validate our findings in Sec. 3 that skip connections reinforce MI attacks. More results can be found in the Supp.", + "bbox": [ + 212, + 505, + 787, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental results on other SOTA MI attacks. Beside PPA attack [28], we validate our findings of RoLSS on other SOTA MI attacks including KEDMI [4], PLG-MI [38] and LOMMA [25]. The details for these MI attack can be found in the Tab. 3. Following previous setups, we use IR152 [12] as the architecture of the target classifier. The experimental results are presented in Fig. 4. In Fig. 4, we focus on the range of natural accuracy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. Across all these MI attacks, the results are consistent to those under PPA attack in Sec. 3.2, where the attack accuracy reduces significantly when the skip connections are removed regardless the effect from the natural accuracy.", + "bbox": [ + 212, + 611, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental results on other private datasets. In addition to Facescrub dataset [24] in the main study, we further validate our findings of RoLSS on other datasets, including CelebA [19] and Stanford Dogs [17]. For CelebA [19], we conduct our study on KEDMI [4]/LOMMA [25]/PLG-MI [38]. We use the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "standard setup of IR152 [12] as the architecture of target classifier. For Stanford Dogs, we conduct our study on PPA [28]. We use the setup of ResNet-50/101 [12] and DenseNet-121/169 [13] as the architecture of the target classifier. The experimental results are presented in Fig. 4 and Fig. 5 for CelebA and Stanford Dogs, respectively. The results in both datasets are consistent with the results for Facescrub, where the attack accuracy experiences a significantly reduction when the skip connection is removed regardless the effect from the natural accuracy.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Removing skip connection in inference-time architecture via RepVGG could not help mitigate vulnerability to MI", + "text_level": 1, + "bbox": [ + 214, + 279, + 741, + 333 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we analyze RepVGG [5], an established method to decouple the training-time and inference-time architectures through structural re-parameterization. RepVGG converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection) to accelerate inference speed, as shown in Fig. 6. As RepVGG removes skip connections in inference-time architecture, we seek to explore: Can RepVGG inference-time architecture mitigate vulnerability to MI attacks?", + "bbox": [ + 212, + 349, + 486, + 575 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We denote the training-time multibranch architecture by $T_{RepVGG}$ , and the inference-time plain architecture", + "bbox": [ + 212, + 578, + 486, + 625 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/41eb5af770a31595bed5f7be6e8cad80621252104ca7ffc4e1980940d8765cf5.jpg", + "image_caption": [ + "Fig.6: RepVGG [5] converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection). Through our analytical and empirical analysis, we show that despite the differences in architectures, the gradients across the blocks are similar (See our analysis in the Supp.). Consequently, even with the removal of skip connections, RepVGG cannot mitigate vulnerability to MI attacks." + ], + "image_footnote": [], + "bbox": [ + 522, + 329, + 764, + 454 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "by $T_{RepVGG}$ . Through our analytical and empirical analysis, we show that the gradients in $T_{RepVGG}$ under MI attacks and that in $T_{RepVGG}$ are the same. Therefore, removing skip connection in the inference-time architecture $T_{RepVGG}$ could not help mitigate vulnerability to MI. Our detailed analytical analysis can be found in the Supp.", + "bbox": [ + 212, + 625, + 787, + 703 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To empirically validate this, we assess the vulnerability of RepVGG inference-time architecture to MI attacks, comparing it to the training-time architecture. We apply SOTA MI attack, PPA [28], on both the training-time and inference-time RepVGG-A0/B3/D2 architectures. The results for other RepVGG architectures can be found in Supp. We strictly follow the training and inference conversion implementation from the original RepVGG source code for the target classifiers trained on the Facescrub dataset [24]. For the PPA attack [28], we follow the attack setup provided in the original PPA source code. Our results in Tab. 4 clearly show that, despite the removal of skip connections, the", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "inference-time architecture obtained via RepVGG remains as vulnerable to MI attacks as the training-time architecture.", + "text_level": 1, + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The skip connection removed architectures in Sec. 3 and those in the study of RepVGG in this section are fundamentally different. During training, RepVGG training-time architecture still receives gradients via skip connection branches. We show that the gradients in RepVGG inference-time architecture remain the same, including the skip connection branch's gradient components even though skip connections are absent in the inference-time architecture. In contrast, our study in Sec. 3 removes skip connections during training, eliminating gradients on skip connections. The study", + "bbox": [ + 212, + 178, + 491, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Experimental results of RepVGG [5] training-time and inference-time architectures. We strictly follow PPA [28] for the attack setup and evaluation. Here $\\mathcal{D}_{priv} =$ Facescrub, $\\mathcal{D}_{pub} =$ FFHQ. Despite the removal of skip connections, RepVGG inference-time architecture remains as vulnerable to MI attacks as the training-time architecture.", + "bbox": [ + 501, + 178, + 787, + 301 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3d82d1fc9fe717c3b164c32d93d10334cb25a35fed9ae58c288da0c7e4975eaa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureAcc↑AttAcc↓
RepVGG-A0Training-time94.9085.19
Inference-time94.9086.13
RepVGG-B3Training-time94.5580.69
Inference-time94.5580.61
RepVGG-D2Training-time92.0567.08
Inference-time92.0566.44
", + "bbox": [ + 514, + 304, + 774, + 416 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "demonstrates the causal effect of skip connections on MI attack accuracy.", + "bbox": [ + 212, + 419, + 743, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Model Inversion Resilient Architecture Design", + "text_level": 1, + "bbox": [ + 214, + 458, + 702, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our findings so far reveal that skip-connections reinforce MI attacks while existing reparameterization technique, RepVGG [5], to remove skip connection in the inference-time architecture cannot mitigate vulnerability to MI attack. To bridge this gap, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our RoLSS, SSF, and TTS are remarkably simple, maintaining the same training procedure as the original model. Applying RoLSS involves no additional hyperparameters, while SSF/TTS requires only one extra hyperparameter, making them easily applicable to various architectures. In contrast, the SOTA MI defense BiDO requires an extensive grid search for each architecture [26].", + "bbox": [ + 212, + 492, + 491, + 840 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. $\\Delta$ represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search.", + "bbox": [ + 501, + 479, + 787, + 618 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d70b64ff203e55e48d9f2508a46f34b4c11f9e36a0b135f45c9e82aadbf304e6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureDefenseAcc↑AttAcc↓Δ↑
ResNet-34No Def.94.6990.78-
BiDO91.6681.982.90
RoLSS (Ours)91.3871.865.72
SSF (Ours)94.2179.7922.90
TTS (Ours)94.4081.6531.48
ResNet-50No Def.94.5882.76-
BiDO91.1258.417.04
RoLSS (Ours)92.8968.448.47
SSF (Ours)93.0574.796.87
TTS (Ours)93.5677.215.44
ResNet-101No Def.94.8683.00-
BiDO90.3167.073.50
RoLSS (Ours)92.4058.689.89
SSF (Ours)93.7971.0611.16
TTS (Ours)94.1677.268.20
ResNet-152No Def.95.4386.51-
BiDO91.8058.147.82
RoLSS (Ours)93.0064.988.86
SSF (Ours)93.7970.719.63
TTS (Ours)93.9773.598.85
", + "bbox": [ + 514, + 618, + 774, + 847 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Removal of Last Stage Skip-Connection (RoLSS)", + "text_level": 1, + "bbox": [ + 214, + 146, + 663, + 162 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Sec. 3, our investigation reveals that removing skip connections in the last stage yields significant degradation in MI attack accuracy, suggesting a promising approach to improve MI robustness from architectural perspective. The MI defense results are presented in Tab. 6 and Tab. 5. Across architectures and skip connection mechanism, the results consistently show that removing the skip connections in last stage (i.e., RoLSS) can improve the MI robustness. Notably, our simple RoLSS achieves highly competitive MI robustness compared to the SOTA MI defense BiDO [26]. For instance, with ResNet-101, our RoLSS improves model accuracy by $2.09\\%$ , while the MI attack accuracy degrades by $8.49\\%$ , resulting in superior MI robustness compared to BiDO.", + "bbox": [ + 212, + 176, + 787, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Skip-Connection Scaling Factor (SSF)", + "text_level": 1, + "bbox": [ + 214, + 354, + 571, + 372 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We further propose SSF on top of RoLSS to improve natural accuracy of the model while maintaining competitive MI robustness. For additive skip connections, we introduce a scale factor $0 \\leq k \\leq 1$ for the signal on the skip connection of the last stage:", + "bbox": [ + 212, + 386, + 785, + 446 + ], + "page_idx": 12 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i + 1} = g _ {i} \\left(z _ {i}\\right) + k \\cdot z _ {i} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 450, + 785, + 465 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Details of SSF for concatenative skip connections can be found in Supp. Our SSF generalizes the skip connection, where $k = 1$ corresponds to the original skip connection, while $k = 0$ is similar to our skip connection removal study. With $k < 1$ , gradients can be limited during MI attack, and this could degrade MI.", + "bbox": [ + 212, + 476, + 491, + 611 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results are presented in Tab. 6 and Tab. 5. We apply SSF over RoLSS and set $k = 0.01$ for all concatenative skip connection architectures and $k = 0.2$ for all additive skip connection architectures in our experimental setups. Overall, our SSF further improves MI robustness beyond RoLSS and outperforms the SOTA MI defense BiDO [26] across various architectures. Notably, for DenseNet, SSF significantly aids in recovering model", + "bbox": [ + 212, + 613, + 491, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. $\\Delta$ represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search.", + "bbox": [ + 501, + 484, + 787, + 622 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/046fbf491251145c3d2ee1af7033b371bdb7eac086384d72c18f799d1a542cc6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureDefenseAcc↑AttAcc↓Δ↑
DenseNet-121No Def.94.6778.09-
RoLSS (Ours)86.8624.486.86
SSF (Ours)91.7356.327.35
DenseNet-161No Def.93.9374.86-
RoLSS (Ours)87.6720.718.65
SSF (Ours)93.7774.273.69
DenseNet-169No Def.94.2877.15-
RoLSS (Ours)72.146.773.18
SSF (Ours)92.9560.9912.15
DenseNet-201No Def.94.3277.62-
RoLSS (Ours)74.2510.243.36
SSF (Ours)93.0965.2110.09
", + "bbox": [ + 514, + 623, + 776, + 773 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "performance while still mitigating MI attacks, resulting in much improved MI robustness. For example, with DenseNet-201, SSF only incurs a $\\sim 1\\%$ drop in model accuracy while degrading MI attack accuracy by $\\sim 12\\%$ .", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Two-stage Training Scheme (TTS)", + "text_level": 1, + "bbox": [ + 215, + 146, + 542, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We further introduce a Two-stage Training Scheme (TTS) on top of RoLSS to improve model accuracy while still maintaining competitive MI robustness. Inspired by Transfer Learning literature [37], TTS consists of two training stages:", + "bbox": [ + 212, + 167, + 785, + 213 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Stage 1: We train model $T$ with full skip-connections architecture over $M$ epochs. This stage ensures the reasonable convergence of $\\theta_T$ with well-backpropagated gradients through the full skip connections architecture. Note that model parameters are far from optimum initially. With full skip connections in this stage, large gradients can be backpropagated in making larger parameter updates.", + "bbox": [ + 212, + 214, + 816, + 289 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Stage 2: We remove skip connections in the last stage, i.e. RoLSS, to create skip connection-removed architecture $T_{p}$ . Then, we continue to train $\\theta_{T_p}$ over $N$ epochs. The pre-trained parameters in Stage 1 serves as initialization for $\\theta_{T_p}$ , thereby aiding the enhanced convergence of $T_{p}$ .", + "bbox": [ + 212, + 289, + 785, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We build TTS on top of RoLSS. For a fair comparison, the total training epochs for both stages (i.e., $M + N$ ) match the total training epochs of the original model. Across all setups, we set $M = 5$ and $N = 95$ . The results in Tab. 5 demonstrate that TTS outperforms the SOTA MI defense BiDO [26]. Furthermore, TTS improves model accuracy while maintaining competitive MI robustness when compared to our RoLSS. For instance, in the ResNet-34 setup, TTS achieves similar MI attack accuracy as BiDO but maintains comparable model accuracy with No Def. model, achieving very competitive MI robustness.", + "bbox": [ + 212, + 349, + 785, + 470 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 489, + 357, + 507 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We conducted a pioneering study to examine the impact of DNN architecture on SOTA MI attacks. Our findings reveal that skip connections reinforce MI attacks, thereby jeopardizing data privacy. Through extensive MI setups, we find that the skip connections in the last stage is the most critical to MI attacks. Furthermore, our analytical and empirical analysis on RepVGG reveal that the removal of skip connections in the inference-time architecture could not help mitigate the MI vulnerability. Based on our own findings, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our MI-resilient architecture designs are remarkably simple to apply and achieve very competitive MI robustness compared to SOTA MI defense.", + "bbox": [ + 212, + 518, + 785, + 685 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 705, + 426, + 722 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programmes (AISG Award No.: AISG2-TC-2022-007); The Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). This material is based on the research/work support in part by the Changi General Hospital and Singapore University of Technology and Design, under the HealthTech Innovation Fund (HTIF Award No. CGH-SUTD-2021-004).", + "bbox": [ + 212, + 733, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. An, S., Tao, G., Xu, Q., Liu, Y., Shen, G., Yao, Y., Xu, J., Zhang, X.: Mirror: Model inversion for deep learning network with high fidelity. In: Proceedings of the 29th Network and Distributed System Security Symposium (2022)", + "2. Cazenavette, G., Murdock, C., Lucey, S.: Architectural adversarial robustness: The case for deep pursuit. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7150-7158 (2021)", + "3. Chang, X., Zhang, W., Qian, Y., Le Roux, J., Watanabe, S.: End-to-end multispeaker speech recognition with transformer. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6134-6138. IEEE (2020)", + "4. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16178-16187 (2021)", + "5. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021)", + "6. Dippel, J., Vogler, S., Hohne, J.: Towards fine-grained visual representations by combining contrastive learning with image reconstruction and attention-weighted pooling. arXiv preprint arXiv:2104.04323 (2021)", + "7. Dufumier, B., Gori, P., Victor, J., Grigis, A., Wessa, M., Brambilla, P., Favre, P., Polosan, M., Mcdonald, C., Piguet, C.M., et al.: Contrastive learning with continuous proxy meta-data for 3d mri classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 58-68. Springer (2021)", + "8. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. pp. 1322-1333 (2015)", + "9. Fredrikson, M., Lantz, E., Jha, S., Lin, S., Page, D., Ristenpart, T.: Privacy in pharmacogenetics: An end-to-end case study of personalized warfarin dosing. In: 23rd USENIX Security Symposium (USENIX Security 14). pp. 17-32 (2014)", + "0. Guo, J., Zhu, X., Zhao, C., Cao, D., Lei, Z., Li, S.Z.: Learning meta face recognition in unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6172 (2020)", + "1. Han, G., Choi, J., Lee, H., Kim, J.: Reinforcement learning-based black-box model inversion attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20504-20513 (2023)", + "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "3. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)", + "4. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5901-5910 (2020)" + ], + "bbox": [ + 225, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022)", + "16. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)", + "17. Khosla, A., Jayadevaprakash, N., Yao, B., Fei-Fei, L.: Novel dataset for fine-grained image categorization. In: First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition. Colorado Springs, CO (June 2011)", + "18. Krishna, G., Tran, C., Yu, J., Tewfik, A.H.: Speech recognition with no speech or with noisy speech. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1090-1094. IEEE (2019)", + "19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)", + "20. Luo, L., Xu, D., Chen, H., Wong, T.T., Heng, P.A.: Pseudo bias-balanced learning for debiased chest x-ray classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 621-631. Springer (2022)", + "21. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14234 (2021)", + "22. Mishra, S., Zhang, Y., Zhang, L., Zhang, T., Hu, X.S., Chen, D.Z.: Data-driven deep supervision for skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 721-731. Springer (2022)", + "23. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Radiological reports improve pretraining for localized imaging tasks on chest x-rays. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 647-657. Springer (2022)", + "24. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014)", + "25. Nguyen, N.B., Chandrasegaran, K., Abdollahzadeh, M., Cheung, N.M.: Rethinking model inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "26. Peng, X., Liu, F., Zhang, J., Lan, L., Ye, J., Liu, T., Han, B.: Bilateral dependency optimization: Defending against model-inversion attacks. In: KDD (2022)", + "27. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)", + "28. Struppek, L., Hintersdorf, D., Correia, A.D.A., Adler, A., Kersting, K.: Plug & play attacks: Towards robust and flexible model inversion attacks. arXiv preprint arXiv:2201.12179 (2022)", + "29. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)", + "30. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: European conference on computer vision. pp. 459-479. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Koh et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Wang, K.C., Fu, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021)", + "32. Wang, T., Zhang, Y., Jia, R.: Improving robustness to model inversion attacks via mutual information regularization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11666-11673 (2021)", + "33. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020)", + "34. Yang, J., Chen, H., Yan, J., Chen, X., Yao, J.: Towards better understanding and better generalization of few-shot classification in histology images with contrastive learning (2022)", + "35. Yang, S., Li, Y., Jiang, Y., Xia, S.T.: Backdoor defense via suppressing model shortcuts. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "36. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 225-240 (2019)", + "37. Yosinski, J., Clune, J., Bengio, Y., Lipson, H.: How transferable are features in deep neural networks? Advances in neural information processing systems 27 (2014)", + "38. Yuan, X., Chen, K., Zhang, J., Zhang, W., Yu, N., Zhang, Y.: Pseudo label-guided model inversion attack via conditional generative adversarial network. Thirty Seventh AAAI Conference on Artificial Intelligence (AAAI 23) (2023)", + "39. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 253-261 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 534 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On the Vulnerability of Skip Connections to Model Inversion Attacks", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_model.json b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6c9ac18a19e5d8f33647f3b4d14468c526349eb4 --- /dev/null +++ b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_model.json @@ -0,0 +1,2203 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.213, + 0.767, + 0.229 + ], + "angle": 0, + "content": "Koh Jun Hao*, Sy-Tuyen Ho*, Ngoc-Bao Nguyen, and Ngai-Man Cheung" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.24, + 0.69, + 0.255 + ], + "angle": 0, + "content": "Singapore University of Technology and Design (SUTD)" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.256, + 0.653, + 0.269 + ], + "angle": 0, + "content": "{junhao_koh,ngaiman_cheung}@sutd.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.319, + 0.744, + 0.57 + ], + "angle": 0, + "content": "Abstract. Skip connections are fundamental architecture designs for modern deep neural networks (DNNs) such as CNNs and ViTs. While they help improve model performance significantly, we identify a vulnerability associated with skip connections to Model Inversion (MI) attacks, a type of privacy attack that aims to reconstruct private training data through abusive exploitation of a model. In this paper, as a pioneer work to understand how DNN architectures affect MI, we study the impact of skip connections on MI. We make the following discoveries: 1) Skip connections reinforce MI attacks and compromise data privacy. 2) Skip connections in the last stage are the most critical to attack. 3) RepVGG, an approach to remove skip connections in the inference-time architectures, could not mitigate the vulnerability to MI attacks. 4) Based on our findings, we propose MI-resilient architecture designs for the first time. Without bells and whistles, we show in extensive experiments that our MI-resilient architectures can outperform state-of-the-art (SOTA) defense methods in MI robustness. Furthermore, our MI-resilient architectures are complementary to existing MI defense methods. Our project is available at https://Pillowkoh.github.io/projects/RoLSS/" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.582, + 0.741, + 0.61 + ], + "angle": 0, + "content": "Keywords: Model Inversion \\(\\cdot\\) Skip Connection \\(\\cdot\\) Model Inversion Resilient Architecture" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.377, + 0.651 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.756 + ], + "angle": 0, + "content": "As deep neural networks (DNNs) see growing deployment across various applications like face recognition [3,6,7,10,14,18,21,27,34] and healthcare [6,7,20,22, 23,34], concerns about the privacy implications of DNNs are on the rise. Many DNNs are trained on private and sensitive datasets. There is an increasing concern of potential leakage of information of these private training samples through malicious exploitation of the model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.817 + ], + "angle": 0, + "content": "One particular privacy threat that has garnered growing attention is Model Inversion (MI) attacks. In MI attacks, adversaries seek to reconstruct private training samples by exploiting vulnerabilities in the model. For instance, an adversary with access to a face recognition model may abuse it to reconstruct the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.355, + 0.839 + ], + "angle": 0, + "content": "* Co-first authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "private facial images of individuals from the model's training dataset. Following previous works [4, 11, 25, 39], we focus on reconstruction of images and use the face recognition as a running example." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Research gap. Recently, there is an increasing interest to study MI and to understand the feasibility and extent of reconstructing private training samples from DNNs, from the MI attack and MI defense perspectives. However, previous studies have overlooked DNN architecture, and there is a lack of study to understand how DNN architecture designs affect MI (Tab. 1). In particular, MI has been formulated as an optimization problem to seek an image similar to that of an identity in the private training dataset. Commonly, the MI optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.496, + 0.637 + ], + "angle": 0, + "content": "objective is formulated as maximization of likelihood under the model being attacked (target model). Several improved MI objectives have been proposed recently, e.g. logit maximization [25,38]. Meanwhile, various regularizations on the MI objective have been studied to improve the effectiveness of MI, e.g. prior loss to penalize unrealistic images [39]. In addition, various distributional priors leveraging generative models trained on public datasets have been proposed to guide the inversion (optimization) process during MI attacks [1, 4, 31, 38, 39]. Furthermore, regularizations on the training objective of the target model to reduce the correlation exploited by MI have been studied as methods to defend against MI attacks [26,32]. However, there is a lack of study to understand the effect of DNN architecture design on MI." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.309, + 0.789, + 0.448 + ], + "angle": 0, + "content": "Table 1: There is a lack of study to understand how DNN architecture designs affect MI. Previous MI studies are DNN architecture-agnostic, focusing on MI objective, effect of regularizing MI objective, effect of distributional prior based on generative modelling, and regularization on the training objective of the target model. Our work is a pioneer study to understand how DNN architectures affect MI attacks." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.45, + 0.784, + 0.631 + ], + "angle": 0, + "content": "
MI objectiveEffect of regularizing MI objectiveEffect of distributional prior to guide MIRegularization on the training objective of target modelEffect of DNN architecture design on MI
MI [9]✓
GMI [39]✓✓
KEDMI [4]✓✓
VMI [31]✓
MIRROR [1]✓
PPA [28]✓
LOMMA [25]✓
PLGMI [38]✓✓
RLBMI [11]✓
BREPMI [15]✓
MID [32]✓
BiDO [26]✓
Ours✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.79, + 0.844 + ], + "angle": 0, + "content": "In this paper, we address this research gap and conduct the first study to understand how DNN architecture designs affect MI. We put our focus on skip connections [12, 13], a fundamental network design that facilitates the training of very deep neural networks. Skip connections mitigate the vanishing gradient problem during the training stage [12]. Meanwhile, many state-of-the-art (SOTA) MI attacks [4, 25, 28, 38, 39] require the use of gradients to guide the reconstruction of private training samples during the inversion stage. We hypothesize that skip connections facilitate flowing of gradient during inversion, reinforcing MI attacks and posing a considerable vulnerability to data privacy in DNNs (Fig. 1). To validate our hypothesis, we carefully design experiments to single out the effect of skip connections on MI attack performance. Our extensive experiments on SOTA networks (ResNet [12], DenseNet [13], MaxViT [30], EfficientNet [29]) against SOTA MI attacks (PPA [28], PLG-MI [38], LOMMA [25], KEDMI [4])" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.145, + 0.77, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.242, + 0.77, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.344, + 0.77, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.485, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Fig. 1: (I) Illustration of MI attack on ResNet-like architecture (Sec. 3.1). This figure depicts the MI attack framework for SOTA white-box MI attacks [1,4,25, 28,31,38,39], which leverage a generative model \\( G(\\cdot) \\) to exploit the target model via gradient descent and backpropagation. Specifically, for each iteration, \\( \\tilde{x} = G(w) \\) is fed into the target model in the forward pass, and MI loss \\( \\mathcal{L} \\) is computed. In the backward pass, gradients of \\( \\mathcal{L} \\) are computed and back-propagated to obtain \\( \\partial \\mathcal{L} / \\partial w \\), which is used to update \\( w \\) to achieve reconstruction of private training data. (II) Additive Skip Connection (Sec. 3.1). During MI attacks, skip connections allow gradients to bypass the residue module and enhance backpropagation. We hypothesize that this reinforces MI attacks. (III) Our study on skip connection removal (Sec. 3.2 and Sec. 3.3). To validate our hypothesis that skip connections could reinforce MI, we study the effect of skip connections on MI by removing skip connections within various stages of the target model. We study both additive and concatenative skip connections. (IV) Results of stage-wise skip connection removal study (Sec. 3.2 and Sec. 3.3). The sub-figures show that skip connections have a considerable effect on MI. For both additive and concatenative skip connections, we observe that removal of skip connections result in considerable degradation of MI attack accuracy. Furthermore, we observe that skip connections in the last stage have the most significant effect on MI. Best viewed in color with zooming in. consistently show that skip connections reinforce MI attacks in DNNs. Furthermore, we find that skip connections in the last stage have the most significant effect to MI attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "To mitigate the MI vulnerability caused by skip connection, we analyze, RepVGG [5], an established reparameterization technique which converts a multibranch training-time architecture with skip connections to a plain, VGG-like" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "inference-time architecture with skip connection removed. However, our analysis shows that, while RepVGG could enable an inference-time network without any skip connection, the gradients during MI attacks on this inference-time network are the same as that on the multi-branch training-time architecture with skip connections. Therefore, RepVGG could not mitigate the vulnerability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.359 + ], + "angle": 0, + "content": "To bridge the existing gap, we propose MI-resilient architecture designs based on our own findings. Specifically, as we find that the last stage's skip connections have the most significant effect to MI, we propose Removal of Last Stage Skip-Connection (RoLSS) as MI-resilient architecture designs. As our designs remove only the last stage's skip connections and keep other stages' skip connections intact, we could keep the impact on natural accuracy small in many cases. Building on top of RoLSS, we propose Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS) to recover the model's natural accuracy while maintaining competitive MI robustness. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.366, + 0.788, + 0.472 + ], + "angle": 0, + "content": "- We conduct a pioneer study to understand how skip connections impact MI attacks and MI robustness. We design experiments to carefully single out the effect of skip connections on MI attack performance, accounting for the effect of natural accuracy in our analysis. Through extensive experiments spanning 4 SOTA MI attacks and 10 architectures, we validate that skip connections reinforce MI attacks and pose a considerable vulnerability to data privacy in DNNs (Sec. 3)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.473, + 0.788, + 0.502 + ], + "angle": 0, + "content": "- Notably, we discover that skip connections in the last stage consistently are the most critical to MI attacks (Sec. 3 and Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.788, + 0.562 + ], + "angle": 0, + "content": "- We analyze RepVGG, a well-established reparameterization technique that remove skip connections by decoupling training-time and inference-time architectures. Our analysis reveals that this approach could not mitigate vulnerability to MI attacks (Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.562, + 0.788, + 0.638 + ], + "angle": 0, + "content": "- Based on our findings, we propose MI resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS). Extensive experiments show that our MI-resilient architectures can outperform SOTA defense methods in MI robustness (Sec. 5)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.366, + 0.788, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.39, + 0.675 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Model Inversion. The concept of MI was initially studied by Fredrikson et al. [9], who demonstrated that adversaries could employ machine learning to extract genomic and demographic information about patients from a medical imaging model. This work was later extended to facial recognition in [8]. Since then, several MI studies have been conducted to understand the feasibility and extent of reconstructing private training samples from DNNs [11,15,25,26,28,31,32,36, 38,39]. These studies encompass both MI attacks and MI defense perspectives. We summarize notable developments in Tab. 1. See Supp. for further discussion of related work. Despite considerable progress in MI research, there is a lack of study to understand the effect of DNN architecture design on MI." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Skip connections and DNNs Attacks. Skip connections are recognized as an effective approach to alleviate the vanishing gradient problem, allowing us to train very deep neural networks [12, 13]. There are a few works that study the effect of skip connections to adversarial attacks [2, 33] and backdoor attacks [35]. Differ from existing works, our study is the first to understand how DNN architectures affect model inversion (MI), a growing privacy attack. Our investigation reveals a distinctive aspect: while previous work observed that skip connections aids adversarial robustness [2], our work instead discovers that skip connections reinforce MI attacks. It is important to emphasize that the nature of MI attacks differs significantly from adversarial or backdoor attacks. Particularly, for adversarial attacks, the goal is to deceive the model into making incorrect predictions. For backdoor attacks, the goal is to implant malicious functionality in the model such that the model produces incorrect outputs when a specific attack trigger is present in the input. Importantly, adversarial attacks/backdoor attacks are not privacy attacks and they do not extract sensitive training data information from ML models. Rather, adversarial attacks/backdoor attacks aim to undermine model utility and robustness. Our work is the first to study the implications of skip connection on data privacy of ML models through the MI attacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.453, + 0.782, + 0.487 + ], + "angle": 0, + "content": "3 An Investigation on the Skip Connection Vulnerability to Model Inversion Attacks" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.533, + 0.513 + ], + "angle": 0, + "content": "3.1 Skip connections and MI attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.784, + 0.548 + ], + "angle": 0, + "content": "We first discuss the potential effect of skip connections on MI in this sub-section. Then, the effect is validated in Sec. 3.2 and Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.685 + ], + "angle": 0, + "content": "MI and gradients. MI attacks are a data privacy threat. For a DNN model \\( T \\) trained on a private training dataset \\( \\mathcal{D}_{priv} \\), the adversary tries to exploit sensitive training data \\( \\mathcal{D}_{priv} \\) via the trained model \\( T \\). In most works, MI is formulated as the reconstruction of an input \\( \\tilde{x} \\) which is most likely classified into an identity \\( y \\) by \\( T \\). The model \\( T \\) subject to MI attacks is called target model. We focus on white-box MI attack, which is the most popular and powerful MI attack [1,4,25,28,38,39]. Specifically, we follow previous works and assume attackers have access to the parameters, architectures, and outputs of the models [1,4,25,28,38,39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.761 + ], + "angle": 0, + "content": "To reconstruct a high-dimensional image \\(\\tilde{x}\\), some distributional priors have been proposed in SOTA MI to constrain the search space [4, 39]. The distributional prior is commonly encoded by a generative model \\(G(w)\\) trained on a public dataset \\(\\mathcal{D}_{pub}\\) which has no class intersection with \\(\\mathcal{D}_{priv}\\). MI attacks are commonly formulated as the following optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.769, + 0.786, + 0.789 + ], + "angle": 0, + "content": "\\[\nw ^ {*} = \\underset {w} {\\arg \\min } \\mathcal {L} (w; y, T) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}(w; y, T)\\) is the MI loss which guides reconstruction of \\(\\tilde{x} = G(w)\\) that is most likely to be classified by model \\(T\\) as identity \\(y\\). Commonly, negative log-likelihood is used: \\(\\mathcal{L}(w; y, T) = -\\log \\mathbb{P}_T(y|G(w))\\), while other losses have been" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "proposed, e.g., logit-based [25]. In addition, other regularization can be included in \\(\\mathcal{L}\\), e.g. prior loss [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "Importantly, to solve the optimization in Eq. 1, gradient descent and back propagation are used by most SOTA white-box MI attacks [1, 4, 25, 28, 38, 39]: For each iteration, \\( G(w) \\) is fed into \\( T \\) in the forward pass, and \\( \\mathcal{L} \\) is computed. In the backward pass, gradients of \\( \\mathcal{L} \\) are computed in \\( T \\) and back-propagated to obtain \\( \\partial \\mathcal{L} / \\partial w \\), which is used to update \\( w \\) by the attackers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.787, + 0.359 + ], + "angle": 0, + "content": "Skip connections could reinforce MI. Following the above discussion, backpropagation of gradients during MI inversion could have a considerable effect on the MI attack performance. Meanwhile, for conventional DNN training, skip connections are a fundamental architecture design that is effective in mitigating gradient vanishing during backpropagation. We hypothesize that skip connections could facilitate gradient backpropagation during MI attacks and reinforce MI, thereby compromising data privacy." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.359, + 0.787, + 0.43 + ], + "angle": 0, + "content": "Specifically, in a ResNet-like architecture, there are multiple ResNet blocks. Each ResNet block, with input \\( z_{i} \\) and output \\( z_{i+1} \\), can be represented as: \\( z_{i+1} = z_{i} + g_{i}(z_{i}) \\), including an additive skip connection and a residual module \\( g_{i} \\) comprising multiple convolutions (Fig. 1). During MI inversion, the gradient backpropagates across a ResNet block as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.44, + 0.786, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial z _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\frac {\\partial z _ {i + 1}}{\\partial z _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\left(1 + \\frac {\\partial g _ {i}}{\\partial z _ {i}}\\right) = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} + \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\frac {\\partial g _ {i}}{\\partial z _ {i}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.481, + 0.785, + 0.515 + ], + "angle": 0, + "content": "Importantly, the first gradient component, \\(\\frac{\\partial\\mathcal{L}}{\\partial z_{i + 1}}\\), enabled by the skip connection, enhances backpropagation. We hypothesize that this reinforces MI attacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.54, + 0.604, + 0.555 + ], + "angle": 0, + "content": "3.2 Stage-wise skip connection removal study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.649, + 0.581 + ], + "angle": 0, + "content": "In this section, we validate effect of skip connections on MI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.718 + ], + "angle": 0, + "content": "MI setup. We conduct our analysis on ResNet-101 [12] and DenseNet-121 [13] as target models under the attack setup of SOTA MI attack method PPA [28]. We strictly follow PPA MI setups, where we use FaceScrub [24] as private dataset, \\( \\mathcal{D}_{priv} \\) and attack all IDs as per PPA setup. Following previous SOTA MI works [1,4,15,25,28,31,38,39], we adopt attack accuracy (AttAcc), measured using an evaluation model, as the primary metric for assessing MI performance. Attack accuracy is defined as the percentage of reconstructed images correctly identified by the evaluation model with respect to the target ID. Specific MI attack configuration can be found in the Supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.809 + ], + "angle": 0, + "content": "We conduct our study by removing skip connections from various stages of the architecture. Each time, skip connections from a specific stage are removed, while those in the remaining stages remain unchanged. Each architecture with removed skip connections is trained using \\( \\mathcal{D}_{priv} \\) in exactly the same settings as the original unaltered architecture. In this study, we focus on two common skip connection mechanisms: additive and concatenative." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Additive skip connection removal. We investigate additive skip connections within individual stages of the ResNet-101 architecture. To remove additive skip" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: We strictly follow SOTA PPA [28] for the attack setup and evaluation. Here \\(\\mathcal{D}_{priv} =\\) Facescrub [24], \\(\\mathcal{D}_{pub} =\\) FFHQ [16]. Across architectures both additive and concatenative skip connections, we consistently observe that Skip connections in the last stage are the most critical to MI attacks, resulting in the most degradation in MI attack accuracy. \\(\\varDelta_{AttAcc}\\) represents the reduction in attack accuracy when compared to \"Full\" setting." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.23, + 0.482, + 0.422 + ], + "angle": 0, + "content": "
ArchitectureSkip ConnectionsAttAccΔAttAcc
ResNet-34Full90.78-
Skip-1 Removed83.257.53
Skip-2 Removed77.9212.86
Skip-4 Removed80.6110.17
ResNet-50Full82.76-
Skip-1 Removed73.239.53
Skip-2 Removed78.773.99
Skip-4 Removed68.4414.32
ResNet-101Full83.00-
Skip-1 Removed74.818.19
Skip-2 Removed78.754.25
Skip-4 Removed58.6824.32
ResNet-152Full86.51-
Skip-1 Removed80.356.16
Skip-2 Removed69.0417.47
Skip-4 Removed68.4418.07
" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.232, + 0.779, + 0.421 + ], + "angle": 0, + "content": "
ArchitectureSkip ConnectionsAttAccΔAttAcc
DenseNet-121Full88.11-
Skip-1 Removed61.7226.39
Skip-2 Removed55.2132.90
Skip-4 Removed24.4863.63
DenseNet-161Full74.86-
Skip-1 Removed55.3819.48
Skip-2 Removed59.2515.61
Skip-4 Removed20.7154.15
DenseNet-169Full77.15-
Skip-1 Removed60.9916.16
Skip-2 Removed51.8625.29
Skip-4 Removed6.7770.38
DenseNet-201Full77.62-
Skip-1 Removed57.4120.21
Skip-2 Removed46.6531.97
Skip-4 Removed20.7156.91
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.529 + ], + "angle": 0, + "content": "connections in ResNet-like architectures, we ensure that outputs from the previous layers are not added to the subsequent layers during the feedforward process, as shown in Fig. 1-III. For ResNet-101, the resulting ResNet block encompasses a convolutional layer, \\( g_{i} \\), that consists of a single \\( 1 \\times 1 \\) convolution, followed by a \\( 3 \\times 3 \\) convolution, and lastly, another \\( 1 \\times 1 \\) convolution layer, without a shortcut connection linking the input and output of the ResNet block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Concatenative skip connection removal. We investigate concatenative skip connections within individual stages of the DenseNet-121 architecture. We remove the concatenative skip connections in a similar manner as the removal of additive skip connections (The details can be found in the Supp). DenseNet architectures contain DenseBlocks where input features are concatenated with the output features, before being fed into the next DenseBlock. When these features are merged through concatenation, each layer has direct access to the gradients from the loss function and the original input image. To remove these concatenative skip connections, we remove the process of concatenation and only pass the output feature of the current DenseBlock to the subsequent DenseBlock." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Skip connections reinforce MI. To benchmark our experiments, we utilize the original unaltered architecture to assess the performance of the modified architectures. For a fair comparison, we consider the strong correlation between natural accuracy and attack accuracy [39]. Thus, we compare these architectures at multiple checkpoints that achieve similar natural accuracy. When presenting our results, we denote the removal of all skip connections in the \\( N^{th} \\) stage as \"Skip-N Removed\". ResNet and DenseNet consist of four stages. We examine removal of skip connections from stages 1, 2, and 4. The result of removing skip connections in stage 3 are not included in our study, as this stage contains many parameters, and its removal leads to a severe reduction in model accuracy." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "Our findings reveal that architectures with fewer skip connections impede MI attacks, leading to a decrease in MI attack accuracy. As depicted in Fig. 1-IV, both additive and concatenative skip connection studies consistently show that architectures labeled as \"Skip-N Removed\" exhibit low MI attack accuracy compared to the original architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Removal of Last Stage Skip-Connection (RoLSS) is the most critical to MI attacks. Notably, we consistently observe that removing the skip connection in the last stage (i.e., \"Skip-4 Removed\") results in the most degradation in MI attack accuracy. We attribute this to the specific position of skip connections removed within the architecture. During gradient backpropagation in MI attack, gradients in earlier stages depend on those in later stages, as illustrated in Eq. 2. When skip connections in stage 4 (last stage) are removed, the degraded gradients in stage 4 permeate throughout the earlier stages of the architecture, resulting in the most degradation in MI attack accuracy. We further validate this observation on various architectures for both additive and concatenative skip connections in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.49, + 0.614 + ], + "angle": 0, + "content": "Removing the skip connections leads to MI optimization converging with more false positives. As discussed in Eq. 1, MI adversaries aim to identify optimal \\( w^{*} \\) that maximizes the likelihood \\( \\mathbb{P}_T(y|G(w)) \\). We provide this key observation to understand why removing skip connections degrades MI attacks: When the skip connections are removed, latent variables with high likelihood \\( \\mathbb{P}_T(y|G(w)) \\) can still be identified by Eq. 1, but many \\( w^{*} \\) are false positives. Consequently, this leads to a notable decrease in the accuracy of MI attacks." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.398, + 0.761, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.54, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Fig.2: MI convergence analysis. We compare histograms of likelihoods for original architecture (Full) and Removal of Last Stage Skip-Connection (RoLSS) architecture for ResNet-101 under PPA attack." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.753 + ], + "angle": 0, + "content": "This observation becomes evident when examining the likelihood distribution of \"Full\" and \"Skip-4 removed\" settings (see Fig. 2), which are similar and both very close to 1. This results imply that, with \"Skip-4 removed\" setting, Eq. 1 could still perform well to seek latent variables \\( w \\) to maximize the likelihood \\( \\mathbb{P}_T(y|G(w)) \\). However, despite the similarity in likelihood distributions, the attack accuracy of the \"Skip-4 removed\" setting is significantly lower than that of the \"Full\" setting (i.e., \\( 24.32\\% \\)). This suggests that, due to the absence of skip connections, the gradients in the \"Skip-4 removed\" setting lead MI adversaries to exploit many \\( w^* \\) that do not correspond to images resembling private data." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.788 + ], + "angle": 0, + "content": "3.3 Extensive validation of the MI vulnerability of skip connections" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We conduct extensive experiments to further validate the impact of skip connections to MI attacks. Tab. 3 summarize all MI setups in our validation ranging 10 architectures, 4 SOTA MI attacks, 3 datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.36, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.145, + 0.501, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.145, + 0.643, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.146, + 0.787, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.275, + 0.359, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.275, + 0.5, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.275, + 0.642, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.275, + 0.787, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 3: Additional experiments to validate the impact of skip connections on MI attacks across various architecture designs, including networks with additive skip connections (ResNet-34/50/152 [12]), concatenative skip connections (DenseNet-161/169/201 [13]), and compact CNN (EfficientNet-B0 [29]). We strictly follow PPA [28] for MI setups. In all cases, a significant decrease in attack accuracy is consistently observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Experimental Setting. For a fair comparison, we follow the previous MI works [1, 4, 11, 15, 25, 26, 28, 31, 38, 39] to select Evaluation Metrics, Private Dataset, Public Dataset, and Data Preparation Protocol. The details can be found in Tab. 3. Additional details are presented in the Supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.492, + 0.627 + ], + "angle": 0, + "content": "Skip connections removal. We apply our finding of Removal of Last Stage Skip-Connection (RoLSS) for various architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.492, + 0.734 + ], + "angle": 0, + "content": "Evaluation Metrics. Following the previous MI works, we adopt natural accuracy (Acc) and Attack Accuracy (AttAcc) as the main evaluation metrics. The detailed description and additional qualitative results are presented in the Supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.492, + 0.779 + ], + "angle": 0, + "content": "Experimental results on various architectures. We note that in Fig. 3, we focus on the range of natural accu" + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.58, + 0.788, + 0.662 + ], + "angle": 0, + "content": "Table 3: The summary of our MI setups. We follow the exact the experiment setups of PPA [28]. For the other MI attacks, we follow the setups in [4, 25] for [4, 25, 39]. In total, our study spans 10 architectures, 4 MI attacks, 3 datasets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.663, + 0.787, + 0.769 + ], + "angle": 0, + "content": "
ArchitecturesMI AttackPrivate Dataset
ResNet-34/50/101/152 [12]PPA [28]Facescrub [24]
DesnseNet-121/161/169/201 [13]
MaxViT-T [30]
EfficientNet-B0 [29]
ResNet-50/101 [12]Stanford Dogs [17]
DenseNet-121/169 [13]
IR152 [12]KEDMI [4]LOMMA [25]PLG-MI [38]CelebA [19]
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "racy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. The results are consistent with our empirical study in Sec. 3, where the RoLSS of additive skip connections (e.g., ResNet-34/50/152/EfficientNet-B0) reduce the attack accuracy by around" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.75, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.262, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 4: Additional experiments on other SOTA MI attacks including KEDMI [4], LOMMA [25], and PLG-MI [38]. We follow the standard setup, where \\( T = \\mathrm{IR152} \\), \\( \\mathcal{D}_{priv} = \\mathrm{CelebA} \\), \\( \\mathcal{D}_{pub} = \\mathrm{CelebA/FFHQ} \\). Across all SOTA MI attacks, a consistent and notable reduction in attack accuracy is observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.332, + 0.75, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.453, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Fig. 5: Additional experiments on the Stanford Dogs [17] dataset as \\(\\mathcal{D}_{priv}\\). The experiments are conducted under PPA [28] attacks across various architectures, including ResNet-50/101 and DenseNet-121/169. We strictly follow MI setups in PPA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.612 + ], + "angle": 0, + "content": "10% to 15% while the RoLSS of concatenative skip connections (e.g., DenseNet-161/169/201) reduce the attack accuracy by around 30% to 35%. Overall, in all cases including addictive connections (ResNet-34/50/152), concatenative skip connections (DenseNet-161/169/201), compact CNN (EfficientNet-B0), we consistently observe the significant drops in MI attack accuracy, ranging from 10% to 35%. This results further validate our findings in Sec. 3 that skip connections reinforce MI attacks. More results can be found in the Supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Experimental results on other SOTA MI attacks. Beside PPA attack [28], we validate our findings of RoLSS on other SOTA MI attacks including KEDMI [4], PLG-MI [38] and LOMMA [25]. The details for these MI attack can be found in the Tab. 3. Following previous setups, we use IR152 [12] as the architecture of the target classifier. The experimental results are presented in Fig. 4. In Fig. 4, we focus on the range of natural accuracy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. Across all these MI attacks, the results are consistent to those under PPA attack in Sec. 3.2, where the attack accuracy reduces significantly when the skip connections are removed regardless the effect from the natural accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Experimental results on other private datasets. In addition to Facescrub dataset [24] in the main study, we further validate our findings of RoLSS on other datasets, including CelebA [19] and Stanford Dogs [17]. For CelebA [19], we conduct our study on KEDMI [4]/LOMMA [25]/PLG-MI [38]. We use the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "standard setup of IR152 [12] as the architecture of target classifier. For Stanford Dogs, we conduct our study on PPA [28]. We use the setup of ResNet-50/101 [12] and DenseNet-121/169 [13] as the architecture of the target classifier. The experimental results are presented in Fig. 4 and Fig. 5 for CelebA and Stanford Dogs, respectively. The results in both datasets are consistent with the results for Facescrub, where the attack accuracy experiences a significantly reduction when the skip connection is removed regardless the effect from the natural accuracy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.28, + 0.742, + 0.334 + ], + "angle": 0, + "content": "4 Removing skip connection in inference-time architecture via RepVGG could not help mitigate vulnerability to MI" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.488, + 0.577 + ], + "angle": 0, + "content": "In this section, we analyze RepVGG [5], an established method to decouple the training-time and inference-time architectures through structural re-parameterization. RepVGG converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection) to accelerate inference speed, as shown in Fig. 6. As RepVGG removes skip connections in inference-time architecture, we seek to explore: Can RepVGG inference-time architecture mitigate vulnerability to MI attacks?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.488, + 0.625 + ], + "angle": 0, + "content": "We denote the training-time multibranch architecture by \\(T_{RepVGG}\\), and the inference-time plain architecture" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.33, + 0.765, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.465, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Fig.6: RepVGG [5] converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection). Through our analytical and empirical analysis, we show that despite the differences in architectures, the gradients across the blocks are similar (See our analysis in the Supp.). Consequently, even with the removal of skip connections, RepVGG cannot mitigate vulnerability to MI attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.704 + ], + "angle": 0, + "content": "by \\( T_{RepVGG} \\). Through our analytical and empirical analysis, we show that the gradients in \\( T_{RepVGG} \\) under MI attacks and that in \\( T_{RepVGG} \\) are the same. Therefore, removing skip connection in the inference-time architecture \\( T_{RepVGG} \\) could not help mitigate vulnerability to MI. Our detailed analytical analysis can be found in the Supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To empirically validate this, we assess the vulnerability of RepVGG inference-time architecture to MI attacks, comparing it to the training-time architecture. We apply SOTA MI attack, PPA [28], on both the training-time and inference-time RepVGG-A0/B3/D2 architectures. The results for other RepVGG architectures can be found in Supp. We strictly follow the training and inference conversion implementation from the original RepVGG source code for the target classifiers trained on the Facescrub dataset [24]. For the PPA attack [28], we follow the attack setup provided in the original PPA source code. Our results in Tab. 4 clearly show that, despite the removal of skip connections, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "inference-time architecture obtained via RepVGG remains as vulnerable to MI attacks as the training-time architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.493, + 0.419 + ], + "angle": 0, + "content": "The skip connection removed architectures in Sec. 3 and those in the study of RepVGG in this section are fundamentally different. During training, RepVGG training-time architecture still receives gradients via skip connection branches. We show that the gradients in RepVGG inference-time architecture remain the same, including the skip connection branch's gradient components even though skip connections are absent in the inference-time architecture. In contrast, our study in Sec. 3 removes skip connections during training, eliminating gradients on skip connections. The study" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.179, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Table 4: Experimental results of RepVGG [5] training-time and inference-time architectures. We strictly follow PPA [28] for the attack setup and evaluation. Here \\( \\mathcal{D}_{priv} = \\) Facescrub, \\( \\mathcal{D}_{pub} = \\) FFHQ. Despite the removal of skip connections, RepVGG inference-time architecture remains as vulnerable to MI attacks as the training-time architecture." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.305, + 0.776, + 0.417 + ], + "angle": 0, + "content": "
ArchitectureAcc↑AttAcc↓
RepVGG-A0Training-time94.9085.19
Inference-time94.9086.13
RepVGG-B3Training-time94.5580.69
Inference-time94.5580.61
RepVGG-D2Training-time92.0567.08
Inference-time92.0566.44
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.744, + 0.435 + ], + "angle": 0, + "content": "demonstrates the causal effect of skip connections on MI attack accuracy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.459, + 0.703, + 0.476 + ], + "angle": 0, + "content": "5 Model Inversion Resilient Architecture Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.493, + 0.842 + ], + "angle": 0, + "content": "Our findings so far reveal that skip-connections reinforce MI attacks while existing reparameterization technique, RepVGG [5], to remove skip connection in the inference-time architecture cannot mitigate vulnerability to MI attack. To bridge this gap, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our RoLSS, SSF, and TTS are remarkably simple, maintaining the same training procedure as the original model. Applying RoLSS involves no additional hyperparameters, while SSF/TTS requires only one extra hyperparameter, making them easily applicable to various architectures. In contrast, the SOTA MI defense BiDO requires an extensive grid search for each architecture [26]." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.48, + 0.788, + 0.619 + ], + "angle": 0, + "content": "Table 5: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. \\(\\Delta\\) represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.619, + 0.776, + 0.848 + ], + "angle": 0, + "content": "
ArchitectureDefenseAcc↑AttAcc↓Δ↑
ResNet-34No Def.94.6990.78-
BiDO91.6681.982.90
RoLSS (Ours)91.3871.865.72
SSF (Ours)94.2179.7922.90
TTS (Ours)94.4081.6531.48
ResNet-50No Def.94.5882.76-
BiDO91.1258.417.04
RoLSS (Ours)92.8968.448.47
SSF (Ours)93.0574.796.87
TTS (Ours)93.5677.215.44
ResNet-101No Def.94.8683.00-
BiDO90.3167.073.50
RoLSS (Ours)92.4058.689.89
SSF (Ours)93.7971.0611.16
TTS (Ours)94.1677.268.20
ResNet-152No Def.95.4386.51-
BiDO91.8058.147.82
RoLSS (Ours)93.0064.988.86
SSF (Ours)93.7970.719.63
TTS (Ours)93.9773.598.85
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.147, + 0.665, + 0.163 + ], + "angle": 0, + "content": "5.1 Removal of Last Stage Skip-Connection (RoLSS)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.329 + ], + "angle": 0, + "content": "In Sec. 3, our investigation reveals that removing skip connections in the last stage yields significant degradation in MI attack accuracy, suggesting a promising approach to improve MI robustness from architectural perspective. The MI defense results are presented in Tab. 6 and Tab. 5. Across architectures and skip connection mechanism, the results consistently show that removing the skip connections in last stage (i.e., RoLSS) can improve the MI robustness. Notably, our simple RoLSS achieves highly competitive MI robustness compared to the SOTA MI defense BiDO [26]. For instance, with ResNet-101, our RoLSS improves model accuracy by \\(2.09\\%\\), while the MI attack accuracy degrades by \\(8.49\\%\\), resulting in superior MI robustness compared to BiDO." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.356, + 0.573, + 0.373 + ], + "angle": 0, + "content": "5.2 Skip-Connection Scaling Factor (SSF)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.787, + 0.448 + ], + "angle": 0, + "content": "We further propose SSF on top of RoLSS to improve natural accuracy of the model while maintaining competitive MI robustness. For additive skip connections, we introduce a scale factor \\(0 \\leq k \\leq 1\\) for the signal on the skip connection of the last stage:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.451, + 0.786, + 0.467 + ], + "angle": 0, + "content": "\\[\nz _ {i + 1} = g _ {i} \\left(z _ {i}\\right) + k \\cdot z _ {i} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.492, + 0.612 + ], + "angle": 0, + "content": "Details of SSF for concatenative skip connections can be found in Supp. Our SSF generalizes the skip connection, where \\( k = 1 \\) corresponds to the original skip connection, while \\( k = 0 \\) is similar to our skip connection removal study. With \\( k < 1 \\), gradients can be limited during MI attack, and this could degrade MI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.493, + 0.795 + ], + "angle": 0, + "content": "The results are presented in Tab. 6 and Tab. 5. We apply SSF over RoLSS and set \\( k = 0.01 \\) for all concatenative skip connection architectures and \\( k = 0.2 \\) for all additive skip connection architectures in our experimental setups. Overall, our SSF further improves MI robustness beyond RoLSS and outperforms the SOTA MI defense BiDO [26] across various architectures. Notably, for DenseNet, SSF significantly aids in recovering model" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.485, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Table 6: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. \\(\\Delta\\) represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.624, + 0.777, + 0.774 + ], + "angle": 0, + "content": "
ArchitectureDefenseAcc↑AttAcc↓Δ↑
DenseNet-121No Def.94.6778.09-
RoLSS (Ours)86.8624.486.86
SSF (Ours)91.7356.327.35
DenseNet-161No Def.93.9374.86-
RoLSS (Ours)87.6720.718.65
SSF (Ours)93.7774.273.69
DenseNet-169No Def.94.2877.15-
RoLSS (Ours)72.146.773.18
SSF (Ours)92.9560.9912.15
DenseNet-201No Def.94.3277.62-
RoLSS (Ours)74.2510.243.36
SSF (Ours)93.0965.2110.09
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "performance while still mitigating MI attacks, resulting in much improved MI robustness. For example, with DenseNet-201, SSF only incurs a \\(\\sim 1\\%\\) drop in model accuracy while degrading MI attack accuracy by \\(\\sim 12\\%\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.544, + 0.163 + ], + "angle": 0, + "content": "5.3 Two-stage Training Scheme (TTS)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.786, + 0.214 + ], + "angle": 0, + "content": "We further introduce a Two-stage Training Scheme (TTS) on top of RoLSS to improve model accuracy while still maintaining competitive MI robustness. Inspired by Transfer Learning literature [37], TTS consists of two training stages:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.215, + 0.817, + 0.29 + ], + "angle": 0, + "content": "Stage 1: We train model \\( T \\) with full skip-connections architecture over \\( M \\) epochs. This stage ensures the reasonable convergence of \\( \\theta_T \\) with well-backpropagated gradients through the full skip connections architecture. Note that model parameters are far from optimum initially. With full skip connections in this stage, large gradients can be backpropagated in making larger parameter updates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.787, + 0.35 + ], + "angle": 0, + "content": "Stage 2: We remove skip connections in the last stage, i.e. RoLSS, to create skip connection-removed architecture \\( T_{p} \\). Then, we continue to train \\( \\theta_{T_p} \\) over \\( N \\) epochs. The pre-trained parameters in Stage 1 serves as initialization for \\( \\theta_{T_p} \\), thereby aiding the enhanced convergence of \\( T_{p} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.787, + 0.471 + ], + "angle": 0, + "content": "We build TTS on top of RoLSS. For a fair comparison, the total training epochs for both stages (i.e., \\( M + N \\)) match the total training epochs of the original model. Across all setups, we set \\( M = 5 \\) and \\( N = 95 \\). The results in Tab. 5 demonstrate that TTS outperforms the SOTA MI defense BiDO [26]. Furthermore, TTS improves model accuracy while maintaining competitive MI robustness when compared to our RoLSS. For instance, in the ResNet-34 setup, TTS achieves similar MI attack accuracy as BiDO but maintains comparable model accuracy with No Def. model, achieving very competitive MI robustness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.359, + 0.508 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.787, + 0.686 + ], + "angle": 0, + "content": "We conducted a pioneering study to examine the impact of DNN architecture on SOTA MI attacks. Our findings reveal that skip connections reinforce MI attacks, thereby jeopardizing data privacy. Through extensive MI setups, we find that the skip connections in the last stage is the most critical to MI attacks. Furthermore, our analytical and empirical analysis on RepVGG reveal that the removal of skip connections in the inference-time architecture could not help mitigate the MI vulnerability. Based on our own findings, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our MI-resilient architecture designs are remarkably simple to apply and achieve very competitive MI robustness compared to SOTA MI defense." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.706, + 0.427, + 0.723 + ], + "angle": 0, + "content": "7 Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programmes (AISG Award No.: AISG2-TC-2022-007); The Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). This material is based on the research/work support in part by the Changi General Hospital and Singapore University of Technology and Design, under the HealthTech Innovation Fund (HTIF Award No. CGH-SUTD-2021-004)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.218 + ], + "angle": 0, + "content": "1. An, S., Tao, G., Xu, Q., Liu, Y., Shen, G., Yao, Y., Xu, J., Zhang, X.: Mirror: Model inversion for deep learning network with high fidelity. In: Proceedings of the 29th Network and Distributed System Security Symposium (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.26 + ], + "angle": 0, + "content": "2. Cazenavette, G., Murdock, C., Lucey, S.: Architectural adversarial robustness: The case for deep pursuit. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7150-7158 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.787, + 0.314 + ], + "angle": 0, + "content": "3. Chang, X., Zhang, W., Qian, Y., Le Roux, J., Watanabe, S.: End-to-end multispeaker speech recognition with transformer. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6134-6138. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "4. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16178-16187 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "5. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.787, + 0.44 + ], + "angle": 0, + "content": "6. Dippel, J., Vogler, S., Hohne, J.: Towards fine-grained visual representations by combining contrastive learning with image reconstruction and attention-weighted pooling. arXiv preprint arXiv:2104.04323 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.787, + 0.522 + ], + "angle": 0, + "content": "7. Dufumier, B., Gori, P., Victor, J., Grigis, A., Wessa, M., Brambilla, P., Favre, P., Polosan, M., Mcdonald, C., Piguet, C.M., et al.: Contrastive learning with continuous proxy meta-data for 3d mri classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 58-68. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.787, + 0.578 + ], + "angle": 0, + "content": "8. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. pp. 1322-1333 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.62 + ], + "angle": 0, + "content": "9. Fredrikson, M., Lantz, E., Jha, S., Lin, S., Page, D., Ristenpart, T.: Privacy in pharmacogenetics: An end-to-end case study of personalized warfarin dosing. In: 23rd USENIX Security Symposium (USENIX Security 14). pp. 17-32 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.787, + 0.661 + ], + "angle": 0, + "content": "0. Guo, J., Zhu, X., Zhao, C., Cao, D., Lei, Z., Li, S.Z.: Learning meta face recognition in unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6172 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "1. Han, G., Choi, J., Lee, H., Kim, J.: Reinforcement learning-based black-box model inversion attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20504-20513 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.787, + 0.744 + ], + "angle": 0, + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "3. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "4. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5901-5910 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Koh et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "16. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.284 + ], + "angle": 0, + "content": "17. Khosla, A., Jayadevaprakash, N., Yao, B., Fei-Fei, L.: Novel dataset for fine-grained image categorization. In: First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition. Colorado Springs, CO (June 2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.325 + ], + "angle": 0, + "content": "18. Krishna, G., Tran, C., Yu, J., Tewfik, A.H.: Speech recognition with no speech or with noisy speech. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1090-1094. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.365 + ], + "angle": 0, + "content": "19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.787, + 0.407 + ], + "angle": 0, + "content": "20. Luo, L., Xu, D., Chen, H., Wong, T.T., Heng, P.A.: Pseudo bias-balanced learning for debiased chest x-ray classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 621-631. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14234 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.5 + ], + "angle": 0, + "content": "22. Mishra, S., Zhang, Y., Zhang, L., Zhang, T., Hu, X.S., Chen, D.Z.: Data-driven deep supervision for skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 721-731. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.555 + ], + "angle": 0, + "content": "23. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Radiological reports improve pretraining for localized imaging tasks on chest x-rays. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 647-657. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.597 + ], + "angle": 0, + "content": "24. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.651 + ], + "angle": 0, + "content": "25. Nguyen, N.B., Chandrasegaran, K., Abdollahzadeh, M., Cheung, N.M.: Rethinking model inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.677 + ], + "angle": 0, + "content": "26. Peng, X., Liu, F., Zhang, J., Lan, L., Ye, J., Liu, T., Han, B.: Bilateral dependency optimization: Defending against model-inversion attacks. In: KDD (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.718 + ], + "angle": 0, + "content": "27. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "28. Struppek, L., Hintersdorf, D., Correia, A.D.A., Adler, A., Kersting, K.: Plug & play attacks: Towards robust and flexible model inversion attacks. arXiv preprint arXiv:2201.12179 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "29. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "30. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: European conference on computer vision. pp. 459-479. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Wang, K.C., Fu, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "32. Wang, T., Zhang, Y., Jia, R.: Improving robustness to model inversion attacks via mutual information regularization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11666-11673 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "33. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "34. Yang, J., Chen, H., Yan, J., Chen, X., Yao, J.: Towards better understanding and better generalization of few-shot classification in histology images with contrastive learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "35. Yang, S., Li, Y., Jiang, Y., Xia, S.T.: Backdoor defense via suppressing model shortcuts. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.787, + 0.41 + ], + "angle": 0, + "content": "36. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 225-240 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.439 + ], + "angle": 0, + "content": "37. Yosinski, J., Clune, J., Bengio, Y., Lipson, H.: How transferable are features in deep neural networks? Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.44, + 0.787, + 0.481 + ], + "angle": 0, + "content": "38. Yuan, X., Chen, K., Zhang, J., Zhang, W., Yu, N., Zhang, Y.: Pseudo label-guided model inversion attack via conditional generative adversarial network. Thirty Seventh AAAI Conference on Artificial Intelligence (AAAI 23) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.535 + ], + "angle": 0, + "content": "39. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 253-261 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.535 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_origin.pdf b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..511be3997b854951bade07507e95940b00a9ed3e --- /dev/null +++ b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/ef086c62-5a84-4279-b346-33af023d90c1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a424d2442bd9eee2094d5809aaf704b7359572a12330f5aefff00840dff27423 +size 1638556 diff --git a/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/full.md b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0dfb26ade5cb4fd8a07e9fb66739f56bd534c20 --- /dev/null +++ b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/full.md @@ -0,0 +1,281 @@ +# On the Vulnerability of Skip Connections to Model Inversion Attacks + +Koh Jun Hao*, Sy-Tuyen Ho*, Ngoc-Bao Nguyen, and Ngai-Man Cheung + +Singapore University of Technology and Design (SUTD) + +{junhao_koh,ngaiman_cheung}@sutd.edu.sg + +Abstract. Skip connections are fundamental architecture designs for modern deep neural networks (DNNs) such as CNNs and ViTs. While they help improve model performance significantly, we identify a vulnerability associated with skip connections to Model Inversion (MI) attacks, a type of privacy attack that aims to reconstruct private training data through abusive exploitation of a model. In this paper, as a pioneer work to understand how DNN architectures affect MI, we study the impact of skip connections on MI. We make the following discoveries: 1) Skip connections reinforce MI attacks and compromise data privacy. 2) Skip connections in the last stage are the most critical to attack. 3) RepVGG, an approach to remove skip connections in the inference-time architectures, could not mitigate the vulnerability to MI attacks. 4) Based on our findings, we propose MI-resilient architecture designs for the first time. Without bells and whistles, we show in extensive experiments that our MI-resilient architectures can outperform state-of-the-art (SOTA) defense methods in MI robustness. Furthermore, our MI-resilient architectures are complementary to existing MI defense methods. Our project is available at https://Pillowkoh.github.io/projects/RoLSS/ + +Keywords: Model Inversion $\cdot$ Skip Connection $\cdot$ Model Inversion Resilient Architecture + +# 1 Introduction + +As deep neural networks (DNNs) see growing deployment across various applications like face recognition [3,6,7,10,14,18,21,27,34] and healthcare [6,7,20,22, 23,34], concerns about the privacy implications of DNNs are on the rise. Many DNNs are trained on private and sensitive datasets. There is an increasing concern of potential leakage of information of these private training samples through malicious exploitation of the model. + +One particular privacy threat that has garnered growing attention is Model Inversion (MI) attacks. In MI attacks, adversaries seek to reconstruct private training samples by exploiting vulnerabilities in the model. For instance, an adversary with access to a face recognition model may abuse it to reconstruct the + +private facial images of individuals from the model's training dataset. Following previous works [4, 11, 25, 39], we focus on reconstruction of images and use the face recognition as a running example. + +Research gap. Recently, there is an increasing interest to study MI and to understand the feasibility and extent of reconstructing private training samples from DNNs, from the MI attack and MI defense perspectives. However, previous studies have overlooked DNN architecture, and there is a lack of study to understand how DNN architecture designs affect MI (Tab. 1). In particular, MI has been formulated as an optimization problem to seek an image similar to that of an identity in the private training dataset. Commonly, the MI optimization + +objective is formulated as maximization of likelihood under the model being attacked (target model). Several improved MI objectives have been proposed recently, e.g. logit maximization [25,38]. Meanwhile, various regularizations on the MI objective have been studied to improve the effectiveness of MI, e.g. prior loss to penalize unrealistic images [39]. In addition, various distributional priors leveraging generative models trained on public datasets have been proposed to guide the inversion (optimization) process during MI attacks [1, 4, 31, 38, 39]. Furthermore, regularizations on the training objective of the target model to reduce the correlation exploited by MI have been studied as methods to defend against MI attacks [26,32]. However, there is a lack of study to understand the effect of DNN architecture design on MI. + +Table 1: There is a lack of study to understand how DNN architecture designs affect MI. Previous MI studies are DNN architecture-agnostic, focusing on MI objective, effect of regularizing MI objective, effect of distributional prior based on generative modelling, and regularization on the training objective of the target model. Our work is a pioneer study to understand how DNN architectures affect MI attacks. + +
MI objectiveEffect of regularizing MI objectiveEffect of distributional prior to guide MIRegularization on the training objective of target modelEffect of DNN architecture design on MI
MI [9]✓
GMI [39]✓✓
KEDMI [4]✓✓
VMI [31]✓
MIRROR [1]✓
PPA [28]✓
LOMMA [25]✓
PLGMI [38]✓✓
RLBMI [11]✓
BREPMI [15]✓
MID [32]✓
BiDO [26]✓
Ours✓
+ +In this paper, we address this research gap and conduct the first study to understand how DNN architecture designs affect MI. We put our focus on skip connections [12, 13], a fundamental network design that facilitates the training of very deep neural networks. Skip connections mitigate the vanishing gradient problem during the training stage [12]. Meanwhile, many state-of-the-art (SOTA) MI attacks [4, 25, 28, 38, 39] require the use of gradients to guide the reconstruction of private training samples during the inversion stage. We hypothesize that skip connections facilitate flowing of gradient during inversion, reinforcing MI attacks and posing a considerable vulnerability to data privacy in DNNs (Fig. 1). To validate our hypothesis, we carefully design experiments to single out the effect of skip connections on MI attack performance. Our extensive experiments on SOTA networks (ResNet [12], DenseNet [13], MaxViT [30], EfficientNet [29]) against SOTA MI attacks (PPA [28], PLG-MI [38], LOMMA [25], KEDMI [4]) + +![](images/9d0d7eaaae399a5522b793a9f363074c537cb71cc64fc6fd5395cc76cd2a3b9f.jpg) + +![](images/7d5e20aa343d75e7ab7de49c0c3d3d5d195f26f9576bce1e492b99cd3339c965.jpg) + +![](images/15405a9b6068ac98d634a14c58456dd8835a9f439dc90027923e75fb59104755.jpg) + +Fig. 1: (I) Illustration of MI attack on ResNet-like architecture (Sec. 3.1). This figure depicts the MI attack framework for SOTA white-box MI attacks [1,4,25, 28,31,38,39], which leverage a generative model $G(\cdot)$ to exploit the target model via gradient descent and backpropagation. Specifically, for each iteration, $\tilde{x} = G(w)$ is fed into the target model in the forward pass, and MI loss $\mathcal{L}$ is computed. In the backward pass, gradients of $\mathcal{L}$ are computed and back-propagated to obtain $\partial \mathcal{L} / \partial w$ , which is used to update $w$ to achieve reconstruction of private training data. (II) Additive Skip Connection (Sec. 3.1). During MI attacks, skip connections allow gradients to bypass the residue module and enhance backpropagation. We hypothesize that this reinforces MI attacks. (III) Our study on skip connection removal (Sec. 3.2 and Sec. 3.3). To validate our hypothesis that skip connections could reinforce MI, we study the effect of skip connections on MI by removing skip connections within various stages of the target model. We study both additive and concatenative skip connections. (IV) Results of stage-wise skip connection removal study (Sec. 3.2 and Sec. 3.3). The sub-figures show that skip connections have a considerable effect on MI. For both additive and concatenative skip connections, we observe that removal of skip connections result in considerable degradation of MI attack accuracy. Furthermore, we observe that skip connections in the last stage have the most significant effect on MI. Best viewed in color with zooming in. consistently show that skip connections reinforce MI attacks in DNNs. Furthermore, we find that skip connections in the last stage have the most significant effect to MI attacks. + +To mitigate the MI vulnerability caused by skip connection, we analyze, RepVGG [5], an established reparameterization technique which converts a multibranch training-time architecture with skip connections to a plain, VGG-like + +inference-time architecture with skip connection removed. However, our analysis shows that, while RepVGG could enable an inference-time network without any skip connection, the gradients during MI attacks on this inference-time network are the same as that on the multi-branch training-time architecture with skip connections. Therefore, RepVGG could not mitigate the vulnerability. + +To bridge the existing gap, we propose MI-resilient architecture designs based on our own findings. Specifically, as we find that the last stage's skip connections have the most significant effect to MI, we propose Removal of Last Stage Skip-Connection (RoLSS) as MI-resilient architecture designs. As our designs remove only the last stage's skip connections and keep other stages' skip connections intact, we could keep the impact on natural accuracy small in many cases. Building on top of RoLSS, we propose Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS) to recover the model's natural accuracy while maintaining competitive MI robustness. Our contributions are: + +- We conduct a pioneer study to understand how skip connections impact MI attacks and MI robustness. We design experiments to carefully single out the effect of skip connections on MI attack performance, accounting for the effect of natural accuracy in our analysis. Through extensive experiments spanning 4 SOTA MI attacks and 10 architectures, we validate that skip connections reinforce MI attacks and pose a considerable vulnerability to data privacy in DNNs (Sec. 3). +- Notably, we discover that skip connections in the last stage consistently are the most critical to MI attacks (Sec. 3 and Sec. 3.3). +- We analyze RepVGG, a well-established reparameterization technique that remove skip connections by decoupling training-time and inference-time architectures. Our analysis reveals that this approach could not mitigate vulnerability to MI attacks (Sec. 4). +- Based on our findings, we propose MI resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS). Extensive experiments show that our MI-resilient architectures can outperform SOTA defense methods in MI robustness (Sec. 5). + +# 2 Related Work + +Model Inversion. The concept of MI was initially studied by Fredrikson et al. [9], who demonstrated that adversaries could employ machine learning to extract genomic and demographic information about patients from a medical imaging model. This work was later extended to facial recognition in [8]. Since then, several MI studies have been conducted to understand the feasibility and extent of reconstructing private training samples from DNNs [11,15,25,26,28,31,32,36, 38,39]. These studies encompass both MI attacks and MI defense perspectives. We summarize notable developments in Tab. 1. See Supp. for further discussion of related work. Despite considerable progress in MI research, there is a lack of study to understand the effect of DNN architecture design on MI. + +Skip connections and DNNs Attacks. Skip connections are recognized as an effective approach to alleviate the vanishing gradient problem, allowing us to train very deep neural networks [12, 13]. There are a few works that study the effect of skip connections to adversarial attacks [2, 33] and backdoor attacks [35]. Differ from existing works, our study is the first to understand how DNN architectures affect model inversion (MI), a growing privacy attack. Our investigation reveals a distinctive aspect: while previous work observed that skip connections aids adversarial robustness [2], our work instead discovers that skip connections reinforce MI attacks. It is important to emphasize that the nature of MI attacks differs significantly from adversarial or backdoor attacks. Particularly, for adversarial attacks, the goal is to deceive the model into making incorrect predictions. For backdoor attacks, the goal is to implant malicious functionality in the model such that the model produces incorrect outputs when a specific attack trigger is present in the input. Importantly, adversarial attacks/backdoor attacks are not privacy attacks and they do not extract sensitive training data information from ML models. Rather, adversarial attacks/backdoor attacks aim to undermine model utility and robustness. Our work is the first to study the implications of skip connection on data privacy of ML models through the MI attacks. + +# 3 An Investigation on the Skip Connection Vulnerability to Model Inversion Attacks + +# 3.1 Skip connections and MI attacks + +We first discuss the potential effect of skip connections on MI in this sub-section. Then, the effect is validated in Sec. 3.2 and Sec. 3.3. + +MI and gradients. MI attacks are a data privacy threat. For a DNN model $T$ trained on a private training dataset $\mathcal{D}_{priv}$ , the adversary tries to exploit sensitive training data $\mathcal{D}_{priv}$ via the trained model $T$ . In most works, MI is formulated as the reconstruction of an input $\tilde{x}$ which is most likely classified into an identity $y$ by $T$ . The model $T$ subject to MI attacks is called target model. We focus on white-box MI attack, which is the most popular and powerful MI attack [1,4,25,28,38,39]. Specifically, we follow previous works and assume attackers have access to the parameters, architectures, and outputs of the models [1,4,25,28,38,39]. + +To reconstruct a high-dimensional image $\tilde{x}$ , some distributional priors have been proposed in SOTA MI to constrain the search space [4, 39]. The distributional prior is commonly encoded by a generative model $G(w)$ trained on a public dataset $\mathcal{D}_{pub}$ which has no class intersection with $\mathcal{D}_{priv}$ . MI attacks are commonly formulated as the following optimization: + +$$ +w ^ {*} = \underset {w} {\arg \min } \mathcal {L} (w; y, T) \tag {1} +$$ + +Here, $\mathcal{L}(w; y, T)$ is the MI loss which guides reconstruction of $\tilde{x} = G(w)$ that is most likely to be classified by model $T$ as identity $y$ . Commonly, negative log-likelihood is used: $\mathcal{L}(w; y, T) = -\log \mathbb{P}_T(y|G(w))$ , while other losses have been + +proposed, e.g., logit-based [25]. In addition, other regularization can be included in $\mathcal{L}$ , e.g. prior loss [39]. + +Importantly, to solve the optimization in Eq. 1, gradient descent and back propagation are used by most SOTA white-box MI attacks [1, 4, 25, 28, 38, 39]: For each iteration, $G(w)$ is fed into $T$ in the forward pass, and $\mathcal{L}$ is computed. In the backward pass, gradients of $\mathcal{L}$ are computed in $T$ and back-propagated to obtain $\partial \mathcal{L} / \partial w$ , which is used to update $w$ by the attackers. + +Skip connections could reinforce MI. Following the above discussion, backpropagation of gradients during MI inversion could have a considerable effect on the MI attack performance. Meanwhile, for conventional DNN training, skip connections are a fundamental architecture design that is effective in mitigating gradient vanishing during backpropagation. We hypothesize that skip connections could facilitate gradient backpropagation during MI attacks and reinforce MI, thereby compromising data privacy. + +Specifically, in a ResNet-like architecture, there are multiple ResNet blocks. Each ResNet block, with input $z_{i}$ and output $z_{i+1}$ , can be represented as: $z_{i+1} = z_{i} + g_{i}(z_{i})$ , including an additive skip connection and a residual module $g_{i}$ comprising multiple convolutions (Fig. 1). During MI inversion, the gradient backpropagates across a ResNet block as follows: + +$$ +\frac {\partial \mathcal {L}}{\partial z _ {i}} = \frac {\partial \mathcal {L}}{\partial z _ {i + 1}} \frac {\partial z _ {i + 1}}{\partial z _ {i}} = \frac {\partial \mathcal {L}}{\partial z _ {i + 1}} \left(1 + \frac {\partial g _ {i}}{\partial z _ {i}}\right) = \frac {\partial \mathcal {L}}{\partial z _ {i + 1}} + \frac {\partial \mathcal {L}}{\partial z _ {i + 1}} \frac {\partial g _ {i}}{\partial z _ {i}} \tag {2} +$$ + +Importantly, the first gradient component, $\frac{\partial\mathcal{L}}{\partial z_{i + 1}}$ , enabled by the skip connection, enhances backpropagation. We hypothesize that this reinforces MI attacks. + +# 3.2 Stage-wise skip connection removal study + +In this section, we validate effect of skip connections on MI. + +MI setup. We conduct our analysis on ResNet-101 [12] and DenseNet-121 [13] as target models under the attack setup of SOTA MI attack method PPA [28]. We strictly follow PPA MI setups, where we use FaceScrub [24] as private dataset, $\mathcal{D}_{priv}$ and attack all IDs as per PPA setup. Following previous SOTA MI works [1,4,15,25,28,31,38,39], we adopt attack accuracy (AttAcc), measured using an evaluation model, as the primary metric for assessing MI performance. Attack accuracy is defined as the percentage of reconstructed images correctly identified by the evaluation model with respect to the target ID. Specific MI attack configuration can be found in the Supp. + +We conduct our study by removing skip connections from various stages of the architecture. Each time, skip connections from a specific stage are removed, while those in the remaining stages remain unchanged. Each architecture with removed skip connections is trained using $\mathcal{D}_{priv}$ in exactly the same settings as the original unaltered architecture. In this study, we focus on two common skip connection mechanisms: additive and concatenative. + +Additive skip connection removal. We investigate additive skip connections within individual stages of the ResNet-101 architecture. To remove additive skip + +Table 2: We strictly follow SOTA PPA [28] for the attack setup and evaluation. Here $\mathcal{D}_{priv} =$ Facescrub [24], $\mathcal{D}_{pub} =$ FFHQ [16]. Across architectures both additive and concatenative skip connections, we consistently observe that Skip connections in the last stage are the most critical to MI attacks, resulting in the most degradation in MI attack accuracy. $\varDelta_{AttAcc}$ represents the reduction in attack accuracy when compared to "Full" setting. + +
ArchitectureSkip ConnectionsAttAccΔAttAcc
ResNet-34Full90.78-
Skip-1 Removed83.257.53
Skip-2 Removed77.9212.86
Skip-4 Removed80.6110.17
ResNet-50Full82.76-
Skip-1 Removed73.239.53
Skip-2 Removed78.773.99
Skip-4 Removed68.4414.32
ResNet-101Full83.00-
Skip-1 Removed74.818.19
Skip-2 Removed78.754.25
Skip-4 Removed58.6824.32
ResNet-152Full86.51-
Skip-1 Removed80.356.16
Skip-2 Removed69.0417.47
Skip-4 Removed68.4418.07
+ +
ArchitectureSkip ConnectionsAttAccΔAttAcc
DenseNet-121Full88.11-
Skip-1 Removed61.7226.39
Skip-2 Removed55.2132.90
Skip-4 Removed24.4863.63
DenseNet-161Full74.86-
Skip-1 Removed55.3819.48
Skip-2 Removed59.2515.61
Skip-4 Removed20.7154.15
DenseNet-169Full77.15-
Skip-1 Removed60.9916.16
Skip-2 Removed51.8625.29
Skip-4 Removed6.7770.38
DenseNet-201Full77.62-
Skip-1 Removed57.4120.21
Skip-2 Removed46.6531.97
Skip-4 Removed20.7156.91
+ +connections in ResNet-like architectures, we ensure that outputs from the previous layers are not added to the subsequent layers during the feedforward process, as shown in Fig. 1-III. For ResNet-101, the resulting ResNet block encompasses a convolutional layer, $g_{i}$ , that consists of a single $1 \times 1$ convolution, followed by a $3 \times 3$ convolution, and lastly, another $1 \times 1$ convolution layer, without a shortcut connection linking the input and output of the ResNet block. + +Concatenative skip connection removal. We investigate concatenative skip connections within individual stages of the DenseNet-121 architecture. We remove the concatenative skip connections in a similar manner as the removal of additive skip connections (The details can be found in the Supp). DenseNet architectures contain DenseBlocks where input features are concatenated with the output features, before being fed into the next DenseBlock. When these features are merged through concatenation, each layer has direct access to the gradients from the loss function and the original input image. To remove these concatenative skip connections, we remove the process of concatenation and only pass the output feature of the current DenseBlock to the subsequent DenseBlock. + +Skip connections reinforce MI. To benchmark our experiments, we utilize the original unaltered architecture to assess the performance of the modified architectures. For a fair comparison, we consider the strong correlation between natural accuracy and attack accuracy [39]. Thus, we compare these architectures at multiple checkpoints that achieve similar natural accuracy. When presenting our results, we denote the removal of all skip connections in the $N^{th}$ stage as "Skip-N Removed". ResNet and DenseNet consist of four stages. We examine removal of skip connections from stages 1, 2, and 4. The result of removing skip connections in stage 3 are not included in our study, as this stage contains many parameters, and its removal leads to a severe reduction in model accuracy. + +Our findings reveal that architectures with fewer skip connections impede MI attacks, leading to a decrease in MI attack accuracy. As depicted in Fig. 1-IV, both additive and concatenative skip connection studies consistently show that architectures labeled as "Skip-N Removed" exhibit low MI attack accuracy compared to the original architecture. + +Removal of Last Stage Skip-Connection (RoLSS) is the most critical to MI attacks. Notably, we consistently observe that removing the skip connection in the last stage (i.e., "Skip-4 Removed") results in the most degradation in MI attack accuracy. We attribute this to the specific position of skip connections removed within the architecture. During gradient backpropagation in MI attack, gradients in earlier stages depend on those in later stages, as illustrated in Eq. 2. When skip connections in stage 4 (last stage) are removed, the degraded gradients in stage 4 permeate throughout the earlier stages of the architecture, resulting in the most degradation in MI attack accuracy. We further validate this observation on various architectures for both additive and concatenative skip connections in Tab. 2. + +Removing the skip connections leads to MI optimization converging with more false positives. As discussed in Eq. 1, MI adversaries aim to identify optimal $w^{*}$ that maximizes the likelihood $\mathbb{P}_T(y|G(w))$ . We provide this key observation to understand why removing skip connections degrades MI attacks: When the skip connections are removed, latent variables with high likelihood $\mathbb{P}_T(y|G(w))$ can still be identified by Eq. 1, but many $w^{*}$ are false positives. Consequently, this leads to a notable decrease in the accuracy of MI attacks. + +![](images/2ea65c72d4324ddfe762a7148718c10e42e680de37994ca1240f528e3113b338.jpg) +Fig.2: MI convergence analysis. We compare histograms of likelihoods for original architecture (Full) and Removal of Last Stage Skip-Connection (RoLSS) architecture for ResNet-101 under PPA attack. + +This observation becomes evident when examining the likelihood distribution of "Full" and "Skip-4 removed" settings (see Fig. 2), which are similar and both very close to 1. This results imply that, with "Skip-4 removed" setting, Eq. 1 could still perform well to seek latent variables $w$ to maximize the likelihood $\mathbb{P}_T(y|G(w))$ . However, despite the similarity in likelihood distributions, the attack accuracy of the "Skip-4 removed" setting is significantly lower than that of the "Full" setting (i.e., $24.32\%$ ). This suggests that, due to the absence of skip connections, the gradients in the "Skip-4 removed" setting lead MI adversaries to exploit many $w^*$ that do not correspond to images resembling private data. + +# 3.3 Extensive validation of the MI vulnerability of skip connections + +We conduct extensive experiments to further validate the impact of skip connections to MI attacks. Tab. 3 summarize all MI setups in our validation ranging 10 architectures, 4 SOTA MI attacks, 3 datasets. + +![](images/0f3dcf08bd9244a8d04c595bd30666c003a8022c30505a72f0d800ebf9c0c279.jpg) + +![](images/d891b93e06e2ef88e2cc6ddf5015300c639251256bf2c9da4fadb3e10532d1da.jpg) + +![](images/0d48092f1890a04e609c775d8d855d60b832c62ebddd60651473dfa145f52c11.jpg) + +![](images/f2ac7982cbf1e399cad481edfb8355b99e36686f404ebd70570be8cb4a356fb4.jpg) + +![](images/78bf9733f5fb78d7247bb0511eeee3abf94a7c01473eb3477f1ba359f24eb08f.jpg) +Fig. 3: Additional experiments to validate the impact of skip connections on MI attacks across various architecture designs, including networks with additive skip connections (ResNet-34/50/152 [12]), concatenative skip connections (DenseNet-161/169/201 [13]), and compact CNN (EfficientNet-B0 [29]). We strictly follow PPA [28] for MI setups. In all cases, a significant decrease in attack accuracy is consistently observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks. + +![](images/f5ec4b482d04bab664b39319660c61c6c74ac9ae32b391a87a266a1c1773723d.jpg) + +![](images/c86254f2d8e81304792eab393545d8b0866834bccf2e8c06758960d6e97f5ea5.jpg) + +![](images/6bba80b2492caa1fd1d0a0814c7df8c23b14aeec2a7c47891355528fddd2799d.jpg) + +Experimental Setting. For a fair comparison, we follow the previous MI works [1, 4, 11, 15, 25, 26, 28, 31, 38, 39] to select Evaluation Metrics, Private Dataset, Public Dataset, and Data Preparation Protocol. The details can be found in Tab. 3. Additional details are presented in the Supp. + +Skip connections removal. We apply our finding of Removal of Last Stage Skip-Connection (RoLSS) for various architectures. + +Evaluation Metrics. Following the previous MI works, we adopt natural accuracy (Acc) and Attack Accuracy (AttAcc) as the main evaluation metrics. The detailed description and additional qualitative results are presented in the Supp. + +Experimental results on various architectures. We note that in Fig. 3, we focus on the range of natural accu + +Table 3: The summary of our MI setups. We follow the exact the experiment setups of PPA [28]. For the other MI attacks, we follow the setups in [4, 25] for [4, 25, 39]. In total, our study spans 10 architectures, 4 MI attacks, 3 datasets. + +
ArchitecturesMI AttackPrivate Dataset
ResNet-34/50/101/152 [12]PPA [28]Facescrub [24]
DesnseNet-121/161/169/201 [13]
MaxViT-T [30]
EfficientNet-B0 [29]
ResNet-50/101 [12]Stanford Dogs [17]
DenseNet-121/169 [13]
IR152 [12]KEDMI [4]LOMMA [25]PLG-MI [38]CelebA [19]
+ +racy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. The results are consistent with our empirical study in Sec. 3, where the RoLSS of additive skip connections (e.g., ResNet-34/50/152/EfficientNet-B0) reduce the attack accuracy by around + +![](images/e854107c6902639fc2e74eb64401f65e2a8ab2fef463a600ff8a80a2bafbb867.jpg) + +![](images/2d3795e78fd57b90c992f053dfab6d8ed127a35536e99038647c253541cf9b0a.jpg) +Fig. 4: Additional experiments on other SOTA MI attacks including KEDMI [4], LOMMA [25], and PLG-MI [38]. We follow the standard setup, where $T = \mathrm{IR152}$ , $\mathcal{D}_{priv} = \mathrm{CelebA}$ , $\mathcal{D}_{pub} = \mathrm{CelebA/FFHQ}$ . Across all SOTA MI attacks, a consistent and notable reduction in attack accuracy is observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks. +Fig. 5: Additional experiments on the Stanford Dogs [17] dataset as $\mathcal{D}_{priv}$ . The experiments are conducted under PPA [28] attacks across various architectures, including ResNet-50/101 and DenseNet-121/169. We strictly follow MI setups in PPA. + +10% to 15% while the RoLSS of concatenative skip connections (e.g., DenseNet-161/169/201) reduce the attack accuracy by around 30% to 35%. Overall, in all cases including addictive connections (ResNet-34/50/152), concatenative skip connections (DenseNet-161/169/201), compact CNN (EfficientNet-B0), we consistently observe the significant drops in MI attack accuracy, ranging from 10% to 35%. This results further validate our findings in Sec. 3 that skip connections reinforce MI attacks. More results can be found in the Supp. + +Experimental results on other SOTA MI attacks. Beside PPA attack [28], we validate our findings of RoLSS on other SOTA MI attacks including KEDMI [4], PLG-MI [38] and LOMMA [25]. The details for these MI attack can be found in the Tab. 3. Following previous setups, we use IR152 [12] as the architecture of the target classifier. The experimental results are presented in Fig. 4. In Fig. 4, we focus on the range of natural accuracy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. Across all these MI attacks, the results are consistent to those under PPA attack in Sec. 3.2, where the attack accuracy reduces significantly when the skip connections are removed regardless the effect from the natural accuracy. + +Experimental results on other private datasets. In addition to Facescrub dataset [24] in the main study, we further validate our findings of RoLSS on other datasets, including CelebA [19] and Stanford Dogs [17]. For CelebA [19], we conduct our study on KEDMI [4]/LOMMA [25]/PLG-MI [38]. We use the + +standard setup of IR152 [12] as the architecture of target classifier. For Stanford Dogs, we conduct our study on PPA [28]. We use the setup of ResNet-50/101 [12] and DenseNet-121/169 [13] as the architecture of the target classifier. The experimental results are presented in Fig. 4 and Fig. 5 for CelebA and Stanford Dogs, respectively. The results in both datasets are consistent with the results for Facescrub, where the attack accuracy experiences a significantly reduction when the skip connection is removed regardless the effect from the natural accuracy. + +# 4 Removing skip connection in inference-time architecture via RepVGG could not help mitigate vulnerability to MI + +In this section, we analyze RepVGG [5], an established method to decouple the training-time and inference-time architectures through structural re-parameterization. RepVGG converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection) to accelerate inference speed, as shown in Fig. 6. As RepVGG removes skip connections in inference-time architecture, we seek to explore: Can RepVGG inference-time architecture mitigate vulnerability to MI attacks? + +We denote the training-time multibranch architecture by $T_{RepVGG}$ , and the inference-time plain architecture + +![](images/41eb5af770a31595bed5f7be6e8cad80621252104ca7ffc4e1980940d8765cf5.jpg) +Fig.6: RepVGG [5] converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection). Through our analytical and empirical analysis, we show that despite the differences in architectures, the gradients across the blocks are similar (See our analysis in the Supp.). Consequently, even with the removal of skip connections, RepVGG cannot mitigate vulnerability to MI attacks. + +by $T_{RepVGG}$ . Through our analytical and empirical analysis, we show that the gradients in $T_{RepVGG}$ under MI attacks and that in $T_{RepVGG}$ are the same. Therefore, removing skip connection in the inference-time architecture $T_{RepVGG}$ could not help mitigate vulnerability to MI. Our detailed analytical analysis can be found in the Supp. + +To empirically validate this, we assess the vulnerability of RepVGG inference-time architecture to MI attacks, comparing it to the training-time architecture. We apply SOTA MI attack, PPA [28], on both the training-time and inference-time RepVGG-A0/B3/D2 architectures. The results for other RepVGG architectures can be found in Supp. We strictly follow the training and inference conversion implementation from the original RepVGG source code for the target classifiers trained on the Facescrub dataset [24]. For the PPA attack [28], we follow the attack setup provided in the original PPA source code. Our results in Tab. 4 clearly show that, despite the removal of skip connections, the + +# inference-time architecture obtained via RepVGG remains as vulnerable to MI attacks as the training-time architecture. + +The skip connection removed architectures in Sec. 3 and those in the study of RepVGG in this section are fundamentally different. During training, RepVGG training-time architecture still receives gradients via skip connection branches. We show that the gradients in RepVGG inference-time architecture remain the same, including the skip connection branch's gradient components even though skip connections are absent in the inference-time architecture. In contrast, our study in Sec. 3 removes skip connections during training, eliminating gradients on skip connections. The study + +Table 4: Experimental results of RepVGG [5] training-time and inference-time architectures. We strictly follow PPA [28] for the attack setup and evaluation. Here $\mathcal{D}_{priv} =$ Facescrub, $\mathcal{D}_{pub} =$ FFHQ. Despite the removal of skip connections, RepVGG inference-time architecture remains as vulnerable to MI attacks as the training-time architecture. + +
ArchitectureAcc↑AttAcc↓
RepVGG-A0Training-time94.9085.19
Inference-time94.9086.13
RepVGG-B3Training-time94.5580.69
Inference-time94.5580.61
RepVGG-D2Training-time92.0567.08
Inference-time92.0566.44
+ +demonstrates the causal effect of skip connections on MI attack accuracy. + +# 5 Model Inversion Resilient Architecture Design + +Our findings so far reveal that skip-connections reinforce MI attacks while existing reparameterization technique, RepVGG [5], to remove skip connection in the inference-time architecture cannot mitigate vulnerability to MI attack. To bridge this gap, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our RoLSS, SSF, and TTS are remarkably simple, maintaining the same training procedure as the original model. Applying RoLSS involves no additional hyperparameters, while SSF/TTS requires only one extra hyperparameter, making them easily applicable to various architectures. In contrast, the SOTA MI defense BiDO requires an extensive grid search for each architecture [26]. + +Table 5: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. $\Delta$ represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search. + +
ArchitectureDefenseAcc↑AttAcc↓Δ↑
ResNet-34No Def.94.6990.78-
BiDO91.6681.982.90
RoLSS (Ours)91.3871.865.72
SSF (Ours)94.2179.7922.90
TTS (Ours)94.4081.6531.48
ResNet-50No Def.94.5882.76-
BiDO91.1258.417.04
RoLSS (Ours)92.8968.448.47
SSF (Ours)93.0574.796.87
TTS (Ours)93.5677.215.44
ResNet-101No Def.94.8683.00-
BiDO90.3167.073.50
RoLSS (Ours)92.4058.689.89
SSF (Ours)93.7971.0611.16
TTS (Ours)94.1677.268.20
ResNet-152No Def.95.4386.51-
BiDO91.8058.147.82
RoLSS (Ours)93.0064.988.86
SSF (Ours)93.7970.719.63
TTS (Ours)93.9773.598.85
+ +# 5.1 Removal of Last Stage Skip-Connection (RoLSS) + +In Sec. 3, our investigation reveals that removing skip connections in the last stage yields significant degradation in MI attack accuracy, suggesting a promising approach to improve MI robustness from architectural perspective. The MI defense results are presented in Tab. 6 and Tab. 5. Across architectures and skip connection mechanism, the results consistently show that removing the skip connections in last stage (i.e., RoLSS) can improve the MI robustness. Notably, our simple RoLSS achieves highly competitive MI robustness compared to the SOTA MI defense BiDO [26]. For instance, with ResNet-101, our RoLSS improves model accuracy by $2.09\%$ , while the MI attack accuracy degrades by $8.49\%$ , resulting in superior MI robustness compared to BiDO. + +# 5.2 Skip-Connection Scaling Factor (SSF) + +We further propose SSF on top of RoLSS to improve natural accuracy of the model while maintaining competitive MI robustness. For additive skip connections, we introduce a scale factor $0 \leq k \leq 1$ for the signal on the skip connection of the last stage: + +$$ +z _ {i + 1} = g _ {i} \left(z _ {i}\right) + k \cdot z _ {i} \tag {3} +$$ + +Details of SSF for concatenative skip connections can be found in Supp. Our SSF generalizes the skip connection, where $k = 1$ corresponds to the original skip connection, while $k = 0$ is similar to our skip connection removal study. With $k < 1$ , gradients can be limited during MI attack, and this could degrade MI. + +The results are presented in Tab. 6 and Tab. 5. We apply SSF over RoLSS and set $k = 0.01$ for all concatenative skip connection architectures and $k = 0.2$ for all additive skip connection architectures in our experimental setups. Overall, our SSF further improves MI robustness beyond RoLSS and outperforms the SOTA MI defense BiDO [26] across various architectures. Notably, for DenseNet, SSF significantly aids in recovering model + +Table 6: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. $\Delta$ represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search. + +
ArchitectureDefenseAcc↑AttAcc↓Δ↑
DenseNet-121No Def.94.6778.09-
RoLSS (Ours)86.8624.486.86
SSF (Ours)91.7356.327.35
DenseNet-161No Def.93.9374.86-
RoLSS (Ours)87.6720.718.65
SSF (Ours)93.7774.273.69
DenseNet-169No Def.94.2877.15-
RoLSS (Ours)72.146.773.18
SSF (Ours)92.9560.9912.15
DenseNet-201No Def.94.3277.62-
RoLSS (Ours)74.2510.243.36
SSF (Ours)93.0965.2110.09
+ +performance while still mitigating MI attacks, resulting in much improved MI robustness. For example, with DenseNet-201, SSF only incurs a $\sim 1\%$ drop in model accuracy while degrading MI attack accuracy by $\sim 12\%$ . + +# 5.3 Two-stage Training Scheme (TTS) + +We further introduce a Two-stage Training Scheme (TTS) on top of RoLSS to improve model accuracy while still maintaining competitive MI robustness. Inspired by Transfer Learning literature [37], TTS consists of two training stages: + +Stage 1: We train model $T$ with full skip-connections architecture over $M$ epochs. This stage ensures the reasonable convergence of $\theta_T$ with well-backpropagated gradients through the full skip connections architecture. Note that model parameters are far from optimum initially. With full skip connections in this stage, large gradients can be backpropagated in making larger parameter updates. + +Stage 2: We remove skip connections in the last stage, i.e. RoLSS, to create skip connection-removed architecture $T_{p}$ . Then, we continue to train $\theta_{T_p}$ over $N$ epochs. The pre-trained parameters in Stage 1 serves as initialization for $\theta_{T_p}$ , thereby aiding the enhanced convergence of $T_{p}$ . + +We build TTS on top of RoLSS. For a fair comparison, the total training epochs for both stages (i.e., $M + N$ ) match the total training epochs of the original model. Across all setups, we set $M = 5$ and $N = 95$ . The results in Tab. 5 demonstrate that TTS outperforms the SOTA MI defense BiDO [26]. Furthermore, TTS improves model accuracy while maintaining competitive MI robustness when compared to our RoLSS. For instance, in the ResNet-34 setup, TTS achieves similar MI attack accuracy as BiDO but maintains comparable model accuracy with No Def. model, achieving very competitive MI robustness. + +# 6 Conclusion + +We conducted a pioneering study to examine the impact of DNN architecture on SOTA MI attacks. Our findings reveal that skip connections reinforce MI attacks, thereby jeopardizing data privacy. Through extensive MI setups, we find that the skip connections in the last stage is the most critical to MI attacks. Furthermore, our analytical and empirical analysis on RepVGG reveal that the removal of skip connections in the inference-time architecture could not help mitigate the MI vulnerability. Based on our own findings, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our MI-resilient architecture designs are remarkably simple to apply and achieve very competitive MI robustness compared to SOTA MI defense. + +# 7 Acknowledgement + +This research is supported by the National Research Foundation, Singapore under its AI Singapore Programmes (AISG Award No.: AISG2-TC-2022-007); The Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). This material is based on the research/work support in part by the Changi General Hospital and Singapore University of Technology and Design, under the HealthTech Innovation Fund (HTIF Award No. CGH-SUTD-2021-004). + +# References + +1. An, S., Tao, G., Xu, Q., Liu, Y., Shen, G., Yao, Y., Xu, J., Zhang, X.: Mirror: Model inversion for deep learning network with high fidelity. In: Proceedings of the 29th Network and Distributed System Security Symposium (2022) +2. Cazenavette, G., Murdock, C., Lucey, S.: Architectural adversarial robustness: The case for deep pursuit. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7150-7158 (2021) +3. Chang, X., Zhang, W., Qian, Y., Le Roux, J., Watanabe, S.: End-to-end multispeaker speech recognition with transformer. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6134-6138. IEEE (2020) +4. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16178-16187 (2021) +5. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021) +6. Dippel, J., Vogler, S., Hohne, J.: Towards fine-grained visual representations by combining contrastive learning with image reconstruction and attention-weighted pooling. arXiv preprint arXiv:2104.04323 (2021) +7. Dufumier, B., Gori, P., Victor, J., Grigis, A., Wessa, M., Brambilla, P., Favre, P., Polosan, M., Mcdonald, C., Piguet, C.M., et al.: Contrastive learning with continuous proxy meta-data for 3d mri classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 58-68. Springer (2021) +8. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. pp. 1322-1333 (2015) +9. Fredrikson, M., Lantz, E., Jha, S., Lin, S., Page, D., Ristenpart, T.: Privacy in pharmacogenetics: An end-to-end case study of personalized warfarin dosing. In: 23rd USENIX Security Symposium (USENIX Security 14). pp. 17-32 (2014) +0. Guo, J., Zhu, X., Zhao, C., Cao, D., Lei, Z., Li, S.Z.: Learning meta face recognition in unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6172 (2020) +1. Han, G., Choi, J., Lee, H., Kim, J.: Reinforcement learning-based black-box model inversion attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20504-20513 (2023) +2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +3. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017) +4. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5901-5910 (2020) + +15. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022) +16. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019) +17. Khosla, A., Jayadevaprakash, N., Yao, B., Fei-Fei, L.: Novel dataset for fine-grained image categorization. In: First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition. Colorado Springs, CO (June 2011) +18. Krishna, G., Tran, C., Yu, J., Tewfik, A.H.: Speech recognition with no speech or with noisy speech. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1090-1094. IEEE (2019) +19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015) +20. Luo, L., Xu, D., Chen, H., Wong, T.T., Heng, P.A.: Pseudo bias-balanced learning for debiased chest x-ray classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 621-631. Springer (2022) +21. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14234 (2021) +22. Mishra, S., Zhang, Y., Zhang, L., Zhang, T., Hu, X.S., Chen, D.Z.: Data-driven deep supervision for skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 721-731. Springer (2022) +23. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Radiological reports improve pretraining for localized imaging tasks on chest x-rays. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 647-657. Springer (2022) +24. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014) +25. Nguyen, N.B., Chandrasegaran, K., Abdollahzadeh, M., Cheung, N.M.: Rethinking model inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +26. Peng, X., Liu, F., Zhang, J., Lan, L., Ye, J., Liu, T., Han, B.: Bilateral dependency optimization: Defending against model-inversion attacks. In: KDD (2022) +27. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) +28. Struppek, L., Hintersdorf, D., Correia, A.D.A., Adler, A., Kersting, K.: Plug & play attacks: Towards robust and flexible model inversion attacks. arXiv preprint arXiv:2201.12179 (2022) +29. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019) +30. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: European conference on computer vision. pp. 459-479. Springer (2022) + +31. Wang, K.C., Fu, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021) +32. Wang, T., Zhang, Y., Jia, R.: Improving robustness to model inversion attacks via mutual information regularization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11666-11673 (2021) +33. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020) +34. Yang, J., Chen, H., Yan, J., Chen, X., Yao, J.: Towards better understanding and better generalization of few-shot classification in histology images with contrastive learning (2022) +35. Yang, S., Li, Y., Jiang, Y., Xia, S.T.: Backdoor defense via suppressing model shortcuts. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +36. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 225-240 (2019) +37. Yosinski, J., Clune, J., Bengio, Y., Lipson, H.: How transferable are features in deep neural networks? Advances in neural information processing systems 27 (2014) +38. Yuan, X., Chen, K., Zhang, J., Zhang, W., Yu, N., Zhang, Y.: Pseudo label-guided model inversion attack via conditional generative adversarial network. Thirty Seventh AAAI Conference on Artificial Intelligence (AAAI 23) (2023) +39. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 253-261 (2020) \ No newline at end of file diff --git a/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/images.zip b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7658cd54a86cadc13820f380ae9ffbdecfcdc403 --- /dev/null +++ b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8d3ddc2c136297a4826d53a9950ec9e03d4c38a79edcafb6f63d77a80f80a54 +size 451556 diff --git a/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/layout.json b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e9c3298cacb59e7342a04c69784ab3da9fb7eb34 --- /dev/null +++ b/2024/On the Vulnerability of Skip Connections to Model Inversion Attacks/layout.json @@ -0,0 +1,8692 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 168, + 469, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 168, + 469, + 181 + ], + "spans": [ + { + "bbox": [ + 145, + 168, + 469, + 181 + ], + "type": "text", + "content": "Koh Jun Hao*, Sy-Tuyen Ho*, Ngoc-Bao Nguyen, and Ngai-Man Cheung" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 191, + 190, + 422, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 190, + 422, + 201 + ], + "spans": [ + { + "bbox": [ + 191, + 190, + 422, + 201 + ], + "type": "text", + "content": "Singapore University of Technology and Design (SUTD)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 202, + 399, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 202, + 399, + 213 + ], + "spans": [ + { + "bbox": [ + 214, + 202, + 399, + 213 + ], + "type": "text", + "content": "{junhao_koh,ngaiman_cheung}@sutd.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 252, + 455, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 252, + 455, + 451 + ], + "spans": [ + { + "bbox": [ + 160, + 252, + 455, + 451 + ], + "type": "text", + "content": "Abstract. Skip connections are fundamental architecture designs for modern deep neural networks (DNNs) such as CNNs and ViTs. While they help improve model performance significantly, we identify a vulnerability associated with skip connections to Model Inversion (MI) attacks, a type of privacy attack that aims to reconstruct private training data through abusive exploitation of a model. In this paper, as a pioneer work to understand how DNN architectures affect MI, we study the impact of skip connections on MI. We make the following discoveries: 1) Skip connections reinforce MI attacks and compromise data privacy. 2) Skip connections in the last stage are the most critical to attack. 3) RepVGG, an approach to remove skip connections in the inference-time architectures, could not mitigate the vulnerability to MI attacks. 4) Based on our findings, we propose MI-resilient architecture designs for the first time. Without bells and whistles, we show in extensive experiments that our MI-resilient architectures can outperform state-of-the-art (SOTA) defense methods in MI robustness. Furthermore, our MI-resilient architectures are complementary to existing MI defense methods. Our project is available at https://Pillowkoh.github.io/projects/RoLSS/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "spans": [ + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "content": "Keywords: Model Inversion " + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "content": " Skip Connection " + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "content": " Model Inversion Resilient Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 503, + 230, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 230, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 230, + 515 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 598 + ], + "type": "text", + "content": "As deep neural networks (DNNs) see growing deployment across various applications like face recognition [3,6,7,10,14,18,21,27,34] and healthcare [6,7,20,22, 23,34], concerns about the privacy implications of DNNs are on the rise. Many DNNs are trained on private and sensitive datasets. There is an increasing concern of potential leakage of information of these private training samples through malicious exploitation of the model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": "One particular privacy threat that has garnered growing attention is Model Inversion (MI) attacks. In MI attacks, adversaries seek to reconstruct private training samples by exploiting vulnerabilities in the model. For instance, an adversary with access to a face recognition model may abuse it to reconstruct the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 217, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 217, + 664 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 217, + 664 + ], + "type": "text", + "content": "* Co-first authors" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "private facial images of individuals from the model's training dataset. Following previous works [4, 11, 25, 39], we focus on reconstruction of images and use the face recognition as a running example." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 157, + 482, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 482, + 242 + ], + "type": "text", + "content": "Research gap. Recently, there is an increasing interest to study MI and to understand the feasibility and extent of reconstructing private training samples from DNNs, from the MI attack and MI defense perspectives. However, previous studies have overlooked DNN architecture, and there is a lack of study to understand how DNN architecture designs affect MI (Tab. 1). In particular, MI has been formulated as an optimization problem to seek an image similar to that of an identity in the private training dataset. Commonly, the MI optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 303, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 303, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 303, + 504 + ], + "type": "text", + "content": "objective is formulated as maximization of likelihood under the model being attacked (target model). Several improved MI objectives have been proposed recently, e.g. logit maximization [25,38]. Meanwhile, various regularizations on the MI objective have been studied to improve the effectiveness of MI, e.g. prior loss to penalize unrealistic images [39]. In addition, various distributional priors leveraging generative models trained on public datasets have been proposed to guide the inversion (optimization) process during MI attacks [1, 4, 31, 38, 39]. Furthermore, regularizations on the training objective of the target model to reduce the correlation exploited by MI have been studied as methods to defend against MI attacks [26,32]. However, there is a lack of study to understand the effect of DNN architecture design on MI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 244, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 482, + 354 + ], + "type": "text", + "content": "Table 1: There is a lack of study to understand how DNN architecture designs affect MI. Previous MI studies are DNN architecture-agnostic, focusing on MI objective, effect of regularizing MI objective, effect of distributional prior based on generative modelling, and regularization on the training objective of the target model. Our work is a pioneer study to understand how DNN architectures affect MI attacks." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 356, + 479, + 499 + ], + "blocks": [ + { + "bbox": [ + 310, + 356, + 479, + 499 + ], + "lines": [ + { + "bbox": [ + 310, + 356, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 310, + 356, + 479, + 499 + ], + "type": "table", + "html": "
MI objectiveEffect of regularizing MI objectiveEffect of distributional prior to guide MIRegularization on the training objective of target modelEffect of DNN architecture design on MI
MI [9]✓
GMI [39]✓✓
KEDMI [4]✓✓
VMI [31]✓
MIRROR [1]✓
PPA [28]✓
LOMMA [25]✓
PLGMI [38]✓✓
RLBMI [11]✓
BREPMI [15]✓
MID [32]✓
BiDO [26]✓
Ours✓
", + "image_path": "7af01c354325ac79c2a1cb78f98bc0acfdd527fc83e9fdd8fc68b6cc85473b4e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 668 + ], + "type": "text", + "content": "In this paper, we address this research gap and conduct the first study to understand how DNN architecture designs affect MI. We put our focus on skip connections [12, 13], a fundamental network design that facilitates the training of very deep neural networks. Skip connections mitigate the vanishing gradient problem during the training stage [12]. Meanwhile, many state-of-the-art (SOTA) MI attacks [4, 25, 28, 38, 39] require the use of gradients to guide the reconstruction of private training samples during the inversion stage. We hypothesize that skip connections facilitate flowing of gradient during inversion, reinforcing MI attacks and posing a considerable vulnerability to data privacy in DNNs (Fig. 1). To validate our hypothesis, we carefully design experiments to single out the effect of skip connections on MI attack performance. Our extensive experiments on SOTA networks (ResNet [12], DenseNet [13], MaxViT [30], EfficientNet [29]) against SOTA MI attacks (PPA [28], PLG-MI [38], LOMMA [25], KEDMI [4])" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 471, + 190 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 471, + 190 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 471, + 190 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 471, + 190 + ], + "type": "image", + "image_path": "9d0d7eaaae399a5522b793a9f363074c537cb71cc64fc6fd5395cc76cd2a3b9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 191, + 471, + 271 + ], + "blocks": [ + { + "bbox": [ + 140, + 191, + 471, + 271 + ], + "lines": [ + { + "bbox": [ + 140, + 191, + 471, + 271 + ], + "spans": [ + { + "bbox": [ + 140, + 191, + 471, + 271 + ], + "type": "image", + "image_path": "7d5e20aa343d75e7ab7de49c0c3d3d5d195f26f9576bce1e492b99cd3339c965.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 272, + 471, + 373 + ], + "blocks": [ + { + "bbox": [ + 140, + 272, + 471, + 373 + ], + "lines": [ + { + "bbox": [ + 140, + 272, + 471, + 373 + ], + "spans": [ + { + "bbox": [ + 140, + 272, + 471, + 373 + ], + "type": "image", + "image_path": "15405a9b6068ac98d634a14c58456dd8835a9f439dc90027923e75fb59104755.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "content": "Fig. 1: (I) Illustration of MI attack on ResNet-like architecture (Sec. 3.1). This figure depicts the MI attack framework for SOTA white-box MI attacks [1,4,25, 28,31,38,39], which leverage a generative model " + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "content": " to exploit the target model via gradient descent and backpropagation. Specifically, for each iteration, " + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\tilde{x} = G(w)" + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "content": " is fed into the target model in the forward pass, and MI loss " + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "content": " is computed. In the backward pass, gradients of " + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "content": " are computed and back-propagated to obtain " + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L} / \\partial w" + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "content": ", which is used to update " + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 132, + 384, + 482, + 628 + ], + "type": "text", + "content": " to achieve reconstruction of private training data. (II) Additive Skip Connection (Sec. 3.1). During MI attacks, skip connections allow gradients to bypass the residue module and enhance backpropagation. We hypothesize that this reinforces MI attacks. (III) Our study on skip connection removal (Sec. 3.2 and Sec. 3.3). To validate our hypothesis that skip connections could reinforce MI, we study the effect of skip connections on MI by removing skip connections within various stages of the target model. We study both additive and concatenative skip connections. (IV) Results of stage-wise skip connection removal study (Sec. 3.2 and Sec. 3.3). The sub-figures show that skip connections have a considerable effect on MI. For both additive and concatenative skip connections, we observe that removal of skip connections result in considerable degradation of MI attack accuracy. Furthermore, we observe that skip connections in the last stage have the most significant effect on MI. Best viewed in color with zooming in. consistently show that skip connections reinforce MI attacks in DNNs. Furthermore, we find that skip connections in the last stage have the most significant effect to MI attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "To mitigate the MI vulnerability caused by skip connection, we analyze, RepVGG [5], an established reparameterization technique which converts a multibranch training-time architecture with skip connections to a plain, VGG-like" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "inference-time architecture with skip connection removed. However, our analysis shows that, while RepVGG could enable an inference-time network without any skip connection, the gradients during MI attacks on this inference-time network are the same as that on the multi-branch training-time architecture with skip connections. Therefore, RepVGG could not mitigate the vulnerability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 284 + ], + "type": "text", + "content": "To bridge the existing gap, we propose MI-resilient architecture designs based on our own findings. Specifically, as we find that the last stage's skip connections have the most significant effect to MI, we propose Removal of Last Stage Skip-Connection (RoLSS) as MI-resilient architecture designs. As our designs remove only the last stage's skip connections and keep other stages' skip connections intact, we could keep the impact on natural accuracy small in many cases. Building on top of RoLSS, we propose Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS) to recover the model's natural accuracy while maintaining competitive MI robustness. Our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 289, + 482, + 505 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 289, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 482, + 373 + ], + "type": "text", + "content": "- We conduct a pioneer study to understand how skip connections impact MI attacks and MI robustness. We design experiments to carefully single out the effect of skip connections on MI attack performance, accounting for the effect of natural accuracy in our analysis. Through extensive experiments spanning 4 SOTA MI attacks and 10 architectures, we validate that skip connections reinforce MI attacks and pose a considerable vulnerability to data privacy in DNNs (Sec. 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 374, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 482, + 397 + ], + "type": "text", + "content": "- Notably, we discover that skip connections in the last stage consistently are the most critical to MI attacks (Sec. 3 and Sec. 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 397, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 482, + 445 + ], + "type": "text", + "content": "- We analyze RepVGG, a well-established reparameterization technique that remove skip connections by decoupling training-time and inference-time architectures. Our analysis reveals that this approach could not mitigate vulnerability to MI attacks (Sec. 4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 445, + 482, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 482, + 505 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 482, + 505 + ], + "type": "text", + "content": "- Based on our findings, we propose MI resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF) and Two-Stage Training Scheme (TTS). Extensive experiments show that our MI-resilient architectures can outperform SOTA defense methods in MI robustness (Sec. 5)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 521, + 238, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 238, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 238, + 534 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": "Model Inversion. The concept of MI was initially studied by Fredrikson et al. [9], who demonstrated that adversaries could employ machine learning to extract genomic and demographic information about patients from a medical imaging model. This work was later extended to facial recognition in [8]. Since then, several MI studies have been conducted to understand the feasibility and extent of reconstructing private training samples from DNNs [11,15,25,26,28,31,32,36, 38,39]. These studies encompass both MI attacks and MI defense perspectives. We summarize notable developments in Tab. 1. See Supp. for further discussion of related work. Despite considerable progress in MI research, there is a lack of study to understand the effect of DNN architecture design on MI." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "Skip connections and DNNs Attacks. Skip connections are recognized as an effective approach to alleviate the vanishing gradient problem, allowing us to train very deep neural networks [12, 13]. There are a few works that study the effect of skip connections to adversarial attacks [2, 33] and backdoor attacks [35]. Differ from existing works, our study is the first to understand how DNN architectures affect model inversion (MI), a growing privacy attack. Our investigation reveals a distinctive aspect: while previous work observed that skip connections aids adversarial robustness [2], our work instead discovers that skip connections reinforce MI attacks. It is important to emphasize that the nature of MI attacks differs significantly from adversarial or backdoor attacks. Particularly, for adversarial attacks, the goal is to deceive the model into making incorrect predictions. For backdoor attacks, the goal is to implant malicious functionality in the model such that the model produces incorrect outputs when a specific attack trigger is present in the input. Importantly, adversarial attacks/backdoor attacks are not privacy attacks and they do not extract sensitive training data information from ML models. Rather, adversarial attacks/backdoor attacks aim to undermine model utility and robustness. Our work is the first to study the implications of skip connection on data privacy of ML models through the MI attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 358, + 478, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 478, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 478, + 385 + ], + "type": "text", + "content": "3 An Investigation on the Skip Connection Vulnerability to Model Inversion Attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 394, + 326, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 326, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 326, + 406 + ], + "type": "text", + "content": "3.1 Skip connections and MI attacks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 479, + 434 + ], + "type": "text", + "content": "We first discuss the potential effect of skip connections on MI in this sub-section. Then, the effect is validated in Sec. 3.2 and Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": "MI and gradients. MI attacks are a data privacy threat. For a DNN model " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": " trained on a private training dataset " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv}" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": ", the adversary tries to exploit sensitive training data " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv}" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": " via the trained model " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": ". In most works, MI is formulated as the reconstruction of an input " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": " which is most likely classified into an identity " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": ". The model " + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 435, + 482, + 542 + ], + "type": "text", + "content": " subject to MI attacks is called target model. We focus on white-box MI attack, which is the most popular and powerful MI attack [1,4,25,28,38,39]. Specifically, we follow previous works and assume attackers have access to the parameters, architectures, and outputs of the models [1,4,25,28,38,39]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "content": "To reconstruct a high-dimensional image " + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "content": ", some distributional priors have been proposed in SOTA MI to constrain the search space [4, 39]. The distributional prior is commonly encoded by a generative model " + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "inline_equation", + "content": "G(w)" + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "content": " trained on a public dataset " + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{pub}" + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "content": " which has no class intersection with " + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv}" + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "content": ". MI attacks are commonly formulated as the following optimization:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 255, + 609, + 481, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 609, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 255, + 609, + 481, + 624 + ], + "type": "interline_equation", + "content": "w ^ {*} = \\underset {w} {\\arg \\min } \\mathcal {L} (w; y, T) \\tag {1}", + "image_path": "3968212a6d4bb52953277fedf291bdb67e8d2cfa63b54040063ef34d077b25d3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(w; y, T)" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is the MI loss which guides reconstruction of " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{x} = G(w)" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " that is most likely to be classified by model " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " as identity " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ". Commonly, negative log-likelihood is used: " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(w; y, T) = -\\log \\mathbb{P}_T(y|G(w))" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", while other losses have been" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "proposed, e.g., logit-based [25]. In addition, other regularization can be included in " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", e.g. prior loss [39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "Importantly, to solve the optimization in Eq. 1, gradient descent and back propagation are used by most SOTA white-box MI attacks [1, 4, 25, 28, 38, 39]: For each iteration, " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "G(w)" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " is fed into " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " in the forward pass, and " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " is computed. In the backward pass, gradients of " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " are computed in " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " and back-propagated to obtain " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L} / \\partial w" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ", which is used to update " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " by the attackers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 481, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 481, + 284 + ], + "type": "text", + "content": "Skip connections could reinforce MI. Following the above discussion, backpropagation of gradients during MI inversion could have a considerable effect on the MI attack performance. Meanwhile, for conventional DNN training, skip connections are a fundamental architecture design that is effective in mitigating gradient vanishing during backpropagation. We hypothesize that skip connections could facilitate gradient backpropagation during MI attacks and reinforce MI, thereby compromising data privacy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "text", + "content": "Specifically, in a ResNet-like architecture, there are multiple ResNet blocks. Each ResNet block, with input " + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "text", + "content": " and output " + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "inline_equation", + "content": "z_{i+1}" + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "text", + "content": ", can be represented as: " + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "inline_equation", + "content": "z_{i+1} = z_{i} + g_{i}(z_{i})" + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "text", + "content": ", including an additive skip connection and a residual module " + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 131, + 284, + 481, + 340 + ], + "type": "text", + "content": " comprising multiple convolutions (Fig. 1). During MI inversion, the gradient backpropagates across a ResNet block as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 183, + 348, + 481, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 348, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 183, + 348, + 481, + 373 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial z _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\frac {\\partial z _ {i + 1}}{\\partial z _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\left(1 + \\frac {\\partial g _ {i}}{\\partial z _ {i}}\\right) = \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} + \\frac {\\partial \\mathcal {L}}{\\partial z _ {i + 1}} \\frac {\\partial g _ {i}}{\\partial z _ {i}} \\tag {2}", + "image_path": "fa66502735fa43511bb9aa9f9358506a64bf98f8bfca077ecd200238128baace.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 380, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 480, + 407 + ], + "type": "text", + "content": "Importantly, the first gradient component, " + }, + { + "bbox": [ + 131, + 380, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial z_{i + 1}}" + }, + { + "bbox": [ + 131, + 380, + 480, + 407 + ], + "type": "text", + "content": ", enabled by the skip connection, enhances backpropagation. We hypothesize that this reinforces MI attacks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 427, + 369, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 427, + 369, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 427, + 369, + 439 + ], + "type": "text", + "content": "3.2 Stage-wise skip connection removal study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 397, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 397, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 397, + 460 + ], + "type": "text", + "content": "In this section, we validate effect of skip connections on MI." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 461, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 568 + ], + "type": "text", + "content": "MI setup. We conduct our analysis on ResNet-101 [12] and DenseNet-121 [13] as target models under the attack setup of SOTA MI attack method PPA [28]. We strictly follow PPA MI setups, where we use FaceScrub [24] as private dataset, " + }, + { + "bbox": [ + 130, + 461, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv}" + }, + { + "bbox": [ + 130, + 461, + 481, + 568 + ], + "type": "text", + "content": " and attack all IDs as per PPA setup. Following previous SOTA MI works [1,4,15,25,28,31,38,39], we adopt attack accuracy (AttAcc), measured using an evaluation model, as the primary metric for assessing MI performance. Attack accuracy is defined as the percentage of reconstructed images correctly identified by the evaluation model with respect to the target ID. Specific MI attack configuration can be found in the Supp." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "text", + "content": "We conduct our study by removing skip connections from various stages of the architecture. Each time, skip connections from a specific stage are removed, while those in the remaining stages remain unchanged. Each architecture with removed skip connections is trained using " + }, + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv}" + }, + { + "bbox": [ + 130, + 570, + 481, + 640 + ], + "type": "text", + "content": " in exactly the same settings as the original unaltered architecture. In this study, we focus on two common skip connection mechanisms: additive and concatenative." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Additive skip connection removal. We investigate additive skip connections within individual stages of the ResNet-101 architecture. To remove additive skip" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 182, + 294, + 334 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: We strictly follow SOTA PPA [28] for the attack setup and evaluation. Here " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv} =" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " Facescrub [24], " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{pub} =" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " FFHQ [16]. Across architectures both additive and concatenative skip connections, we consistently observe that Skip connections in the last stage are the most critical to MI attacks, resulting in the most degradation in MI attack accuracy. " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\varDelta_{AttAcc}" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " represents the reduction in attack accuracy when compared to \"Full\" setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 182, + 294, + 334 + ], + "lines": [ + { + "bbox": [ + 136, + 182, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 136, + 182, + 294, + 334 + ], + "type": "table", + "html": "
ArchitectureSkip ConnectionsAttAccΔAttAcc
ResNet-34Full90.78-
Skip-1 Removed83.257.53
Skip-2 Removed77.9212.86
Skip-4 Removed80.6110.17
ResNet-50Full82.76-
Skip-1 Removed73.239.53
Skip-2 Removed78.773.99
Skip-4 Removed68.4414.32
ResNet-101Full83.00-
Skip-1 Removed74.818.19
Skip-2 Removed78.754.25
Skip-4 Removed58.6824.32
ResNet-152Full86.51-
Skip-1 Removed80.356.16
Skip-2 Removed69.0417.47
Skip-4 Removed68.4418.07
", + "image_path": "141fe3034297eaff3e79d290c75e4b95b2ce6129ad35154e61736d044caaed75.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 183, + 476, + 333 + ], + "blocks": [ + { + "bbox": [ + 315, + 183, + 476, + 333 + ], + "lines": [ + { + "bbox": [ + 315, + 183, + 476, + 333 + ], + "spans": [ + { + "bbox": [ + 315, + 183, + 476, + 333 + ], + "type": "table", + "html": "
ArchitectureSkip ConnectionsAttAccΔAttAcc
DenseNet-121Full88.11-
Skip-1 Removed61.7226.39
Skip-2 Removed55.2132.90
Skip-4 Removed24.4863.63
DenseNet-161Full74.86-
Skip-1 Removed55.3819.48
Skip-2 Removed59.2515.61
Skip-4 Removed20.7154.15
DenseNet-169Full77.15-
Skip-1 Removed60.9916.16
Skip-2 Removed51.8625.29
Skip-4 Removed6.7770.38
DenseNet-201Full77.62-
Skip-1 Removed57.4120.21
Skip-2 Removed46.6531.97
Skip-4 Removed20.7156.91
", + "image_path": "2f8f2c49e8006b5aaf086d2e1612a7a8c94f291c5e9d0ea1ab2eee87cc8af2bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": "connections in ResNet-like architectures, we ensure that outputs from the previous layers are not added to the subsequent layers during the feedforward process, as shown in Fig. 1-III. For ResNet-101, the resulting ResNet block encompasses a convolutional layer, " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": ", that consists of a single " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " convolution, followed by a " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " convolution, and lastly, another " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " convolution layer, without a shortcut connection linking the input and output of the ResNet block." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 542 + ], + "type": "text", + "content": "Concatenative skip connection removal. We investigate concatenative skip connections within individual stages of the DenseNet-121 architecture. We remove the concatenative skip connections in a similar manner as the removal of additive skip connections (The details can be found in the Supp). DenseNet architectures contain DenseBlocks where input features are concatenated with the output features, before being fed into the next DenseBlock. When these features are merged through concatenation, each layer has direct access to the gradients from the loss function and the original input image. To remove these concatenative skip connections, we remove the process of concatenation and only pass the output feature of the current DenseBlock to the subsequent DenseBlock." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Skip connections reinforce MI. To benchmark our experiments, we utilize the original unaltered architecture to assess the performance of the modified architectures. For a fair comparison, we consider the strong correlation between natural accuracy and attack accuracy [39]. Thus, we compare these architectures at multiple checkpoints that achieve similar natural accuracy. When presenting our results, we denote the removal of all skip connections in the " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "N^{th}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " stage as \"Skip-N Removed\". ResNet and DenseNet consist of four stages. We examine removal of skip connections from stages 1, 2, and 4. The result of removing skip connections in stage 3 are not included in our study, as this stage contains many parameters, and its removal leads to a severe reduction in model accuracy." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "Our findings reveal that architectures with fewer skip connections impede MI attacks, leading to a decrease in MI attack accuracy. As depicted in Fig. 1-IV, both additive and concatenative skip connection studies consistently show that architectures labeled as \"Skip-N Removed\" exhibit low MI attack accuracy compared to the original architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 307 + ], + "type": "text", + "content": "Removal of Last Stage Skip-Connection (RoLSS) is the most critical to MI attacks. Notably, we consistently observe that removing the skip connection in the last stage (i.e., \"Skip-4 Removed\") results in the most degradation in MI attack accuracy. We attribute this to the specific position of skip connections removed within the architecture. During gradient backpropagation in MI attack, gradients in earlier stages depend on those in later stages, as illustrated in Eq. 2. When skip connections in stage 4 (last stage) are removed, the degraded gradients in stage 4 permeate throughout the earlier stages of the architecture, resulting in the most degradation in MI attack accuracy. We further validate this observation on various architectures for both additive and concatenative skip connections in Tab. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "text", + "content": "Removing the skip connections leads to MI optimization converging with more false positives. As discussed in Eq. 1, MI adversaries aim to identify optimal " + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "inline_equation", + "content": "w^{*}" + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "text", + "content": " that maximizes the likelihood " + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_T(y|G(w))" + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "text", + "content": ". We provide this key observation to understand why removing skip connections degrades MI attacks: When the skip connections are removed, latent variables with high likelihood " + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_T(y|G(w))" + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "text", + "content": " can still be identified by Eq. 1, but many " + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "inline_equation", + "content": "w^{*}" + }, + { + "bbox": [ + 130, + 308, + 299, + 486 + ], + "type": "text", + "content": " are false positives. Consequently, this leads to a notable decrease in the accuracy of MI attacks." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 321, + 315, + 465, + 417 + ], + "blocks": [ + { + "bbox": [ + 321, + 315, + 465, + 417 + ], + "lines": [ + { + "bbox": [ + 321, + 315, + 465, + 417 + ], + "spans": [ + { + "bbox": [ + 321, + 315, + 465, + 417 + ], + "type": "image", + "image_path": "2ea65c72d4324ddfe762a7148718c10e42e680de37994ca1240f528e3113b338.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 427, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 482, + 483 + ], + "type": "text", + "content": "Fig.2: MI convergence analysis. We compare histograms of likelihoods for original architecture (Full) and Removal of Last Stage Skip-Connection (RoLSS) architecture for ResNet-101 under PPA attack." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": "This observation becomes evident when examining the likelihood distribution of \"Full\" and \"Skip-4 removed\" settings (see Fig. 2), which are similar and both very close to 1. This results imply that, with \"Skip-4 removed\" setting, Eq. 1 could still perform well to seek latent variables " + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": " to maximize the likelihood " + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_T(y|G(w))" + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": ". However, despite the similarity in likelihood distributions, the attack accuracy of the \"Skip-4 removed\" setting is significantly lower than that of the \"Full\" setting (i.e., " + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "inline_equation", + "content": "24.32\\%" + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": "). This suggests that, due to the absence of skip connections, the gradients in the \"Skip-4 removed\" setting lead MI adversaries to exploit many " + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "inline_equation", + "content": "w^*" + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": " that do not correspond to images resembling private data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 611, + 481, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 624 + ], + "type": "text", + "content": "3.3 Extensive validation of the MI vulnerability of skip connections" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "We conduct extensive experiments to further validate the impact of skip connections to MI attacks. Tab. 3 summarize all MI setups in our validation ranging 10 architectures, 4 SOTA MI attacks, 3 datasets." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 220, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 220, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 220, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 220, + 217 + ], + "type": "image", + "image_path": "0f3dcf08bd9244a8d04c595bd30666c003a8022c30505a72f0d800ebf9c0c279.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 306, + 217 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 306, + 217 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 306, + 217 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 306, + 217 + ], + "type": "image", + "image_path": "d891b93e06e2ef88e2cc6ddf5015300c639251256bf2c9da4fadb3e10532d1da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 393, + 217 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 393, + 217 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 393, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 393, + 217 + ], + "type": "image", + "image_path": "0d48092f1890a04e609c775d8d855d60b832c62ebddd60651473dfa145f52c11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 393, + 115, + 481, + 217 + ], + "blocks": [ + { + "bbox": [ + 393, + 115, + 481, + 217 + ], + "lines": [ + { + "bbox": [ + 393, + 115, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 393, + 115, + 481, + 217 + ], + "type": "image", + "image_path": "f2ac7982cbf1e399cad481edfb8355b99e36686f404ebd70570be8cb4a356fb4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 217, + 219, + 306 + ], + "blocks": [ + { + "bbox": [ + 132, + 217, + 219, + 306 + ], + "lines": [ + { + "bbox": [ + 132, + 217, + 219, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 219, + 306 + ], + "type": "image", + "image_path": "78bf9733f5fb78d7247bb0511eeee3abf94a7c01473eb3477f1ba359f24eb08f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 313, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 390 + ], + "type": "text", + "content": "Fig. 3: Additional experiments to validate the impact of skip connections on MI attacks across various architecture designs, including networks with additive skip connections (ResNet-34/50/152 [12]), concatenative skip connections (DenseNet-161/169/201 [13]), and compact CNN (EfficientNet-B0 [29]). We strictly follow PPA [28] for MI setups. In all cases, a significant decrease in attack accuracy is consistently observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 217, + 306, + 306 + ], + "blocks": [ + { + "bbox": [ + 220, + 217, + 306, + 306 + ], + "lines": [ + { + "bbox": [ + 220, + 217, + 306, + 306 + ], + "spans": [ + { + "bbox": [ + 220, + 217, + 306, + 306 + ], + "type": "image", + "image_path": "f5ec4b482d04bab664b39319660c61c6c74ac9ae32b391a87a266a1c1773723d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 217, + 392, + 306 + ], + "blocks": [ + { + "bbox": [ + 306, + 217, + 392, + 306 + ], + "lines": [ + { + "bbox": [ + 306, + 217, + 392, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 392, + 306 + ], + "type": "image", + "image_path": "c86254f2d8e81304792eab393545d8b0866834bccf2e8c06758960d6e97f5ea5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 217, + 481, + 306 + ], + "blocks": [ + { + "bbox": [ + 392, + 217, + 481, + 306 + ], + "lines": [ + { + "bbox": [ + 392, + 217, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 392, + 217, + 481, + 306 + ], + "type": "image", + "image_path": "6bba80b2492caa1fd1d0a0814c7df8c23b14aeec2a7c47891355528fddd2799d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 449 + ], + "type": "text", + "content": "Experimental Setting. For a fair comparison, we follow the previous MI works [1, 4, 11, 15, 25, 26, 28, 31, 38, 39] to select Evaluation Metrics, Private Dataset, Public Dataset, and Data Preparation Protocol. The details can be found in Tab. 3. Additional details are presented in the Supp." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 301, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 301, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 301, + 496 + ], + "type": "text", + "content": "Skip connections removal. We apply our finding of Removal of Last Stage Skip-Connection (RoLSS) for various architectures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 301, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 301, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 301, + 581 + ], + "type": "text", + "content": "Evaluation Metrics. Following the previous MI works, we adopt natural accuracy (Acc) and Attack Accuracy (AttAcc) as the main evaluation metrics. The detailed description and additional qualitative results are presented in the Supp." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 582, + 301, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 301, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 301, + 616 + ], + "type": "text", + "content": "Experimental results on various architectures. We note that in Fig. 3, we focus on the range of natural accu" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 308, + 525, + 481, + 609 + ], + "blocks": [ + { + "bbox": [ + 307, + 459, + 482, + 524 + ], + "lines": [ + { + "bbox": [ + 307, + 459, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 482, + 524 + ], + "type": "text", + "content": "Table 3: The summary of our MI setups. We follow the exact the experiment setups of PPA [28]. For the other MI attacks, we follow the setups in [4, 25] for [4, 25, 39]. In total, our study spans 10 architectures, 4 MI attacks, 3 datasets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 525, + 481, + 609 + ], + "lines": [ + { + "bbox": [ + 308, + 525, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 481, + 609 + ], + "type": "table", + "html": "
ArchitecturesMI AttackPrivate Dataset
ResNet-34/50/101/152 [12]PPA [28]Facescrub [24]
DesnseNet-121/161/169/201 [13]
MaxViT-T [30]
EfficientNet-B0 [29]
ResNet-50/101 [12]Stanford Dogs [17]
DenseNet-121/169 [13]
IR152 [12]KEDMI [4]LOMMA [25]PLG-MI [38]CelebA [19]
", + "image_path": "939c09d822c0e1e09ca1e1f48f21668e44d9e8679976328d39245d9d7114e9fb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "racy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. The results are consistent with our empirical study in Sec. 3, where the RoLSS of additive skip connections (e.g., ResNet-34/50/152/EfficientNet-B0) reduce the attack accuracy by around" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 459, + 205 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 459, + 205 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 459, + 205 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 459, + 205 + ], + "type": "image", + "image_path": "e854107c6902639fc2e74eb64401f65e2a8ab2fef463a600ff8a80a2bafbb867.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 262, + 459, + 354 + ], + "blocks": [ + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "type": "text", + "content": "Fig. 4: Additional experiments on other SOTA MI attacks including KEDMI [4], LOMMA [25], and PLG-MI [38]. We follow the standard setup, where " + }, + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "type": "inline_equation", + "content": "T = \\mathrm{IR152}" + }, + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv} = \\mathrm{CelebA}" + }, + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{pub} = \\mathrm{CelebA/FFHQ}" + }, + { + "bbox": [ + 131, + 207, + 482, + 262 + ], + "type": "text", + "content": ". Across all SOTA MI attacks, a consistent and notable reduction in attack accuracy is observed when skip connections are removed in the last stage, demonstrating that skip connections reinforce MI attacks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 262, + 459, + 354 + ], + "lines": [ + { + "bbox": [ + 149, + 262, + 459, + 354 + ], + "spans": [ + { + "bbox": [ + 149, + 262, + 459, + 354 + ], + "type": "image", + "image_path": "2d3795e78fd57b90c992f053dfab6d8ed127a35536e99038647c253541cf9b0a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 358, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 131, + 358, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 482, + 392 + ], + "type": "text", + "content": "Fig. 5: Additional experiments on the Stanford Dogs [17] dataset as " + }, + { + "bbox": [ + 131, + 358, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv}" + }, + { + "bbox": [ + 131, + 358, + 482, + 392 + ], + "type": "text", + "content": ". The experiments are conducted under PPA [28] attacks across various architectures, including ResNet-50/101 and DenseNet-121/169. We strictly follow MI setups in PPA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 484 + ], + "type": "text", + "content": "10% to 15% while the RoLSS of concatenative skip connections (e.g., DenseNet-161/169/201) reduce the attack accuracy by around 30% to 35%. Overall, in all cases including addictive connections (ResNet-34/50/152), concatenative skip connections (DenseNet-161/169/201), compact CNN (EfficientNet-B0), we consistently observe the significant drops in MI attack accuracy, ranging from 10% to 35%. This results further validate our findings in Sec. 3 that skip connections reinforce MI attacks. More results can be found in the Supp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": "Experimental results on other SOTA MI attacks. Beside PPA attack [28], we validate our findings of RoLSS on other SOTA MI attacks including KEDMI [4], PLG-MI [38] and LOMMA [25]. The details for these MI attack can be found in the Tab. 3. Following previous setups, we use IR152 [12] as the architecture of the target classifier. The experimental results are presented in Fig. 4. In Fig. 4, we focus on the range of natural accuracy where the two setups overlap, allowing us to observe changes in attack accuracy at the same natural accuracy level. Across all these MI attacks, the results are consistent to those under PPA attack in Sec. 3.2, where the attack accuracy reduces significantly when the skip connections are removed regardless the effect from the natural accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Experimental results on other private datasets. In addition to Facescrub dataset [24] in the main study, we further validate our findings of RoLSS on other datasets, including CelebA [19] and Stanford Dogs [17]. For CelebA [19], we conduct our study on KEDMI [4]/LOMMA [25]/PLG-MI [38]. We use the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "standard setup of IR152 [12] as the architecture of target classifier. For Stanford Dogs, we conduct our study on PPA [28]. We use the setup of ResNet-50/101 [12] and DenseNet-121/169 [13] as the architecture of the target classifier. The experimental results are presented in Fig. 4 and Fig. 5 for CelebA and Stanford Dogs, respectively. The results in both datasets are consistent with the results for Facescrub, where the attack accuracy experiences a significantly reduction when the skip connection is removed regardless the effect from the natural accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 221, + 454, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 221, + 454, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 221, + 454, + 264 + ], + "type": "text", + "content": "4 Removing skip connection in inference-time architecture via RepVGG could not help mitigate vulnerability to MI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 298, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 298, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 298, + 456 + ], + "type": "text", + "content": "In this section, we analyze RepVGG [5], an established method to decouple the training-time and inference-time architectures through structural re-parameterization. RepVGG converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection) to accelerate inference speed, as shown in Fig. 6. As RepVGG removes skip connections in inference-time architecture, we seek to explore: Can RepVGG inference-time architecture mitigate vulnerability to MI attacks?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 458, + 298, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 298, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 298, + 495 + ], + "type": "text", + "content": "We denote the training-time multibranch architecture by " + }, + { + "bbox": [ + 130, + 458, + 298, + 495 + ], + "type": "inline_equation", + "content": "T_{RepVGG}" + }, + { + "bbox": [ + 130, + 458, + 298, + 495 + ], + "type": "text", + "content": ", and the inference-time plain architecture" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 261, + 468, + 360 + ], + "blocks": [ + { + "bbox": [ + 320, + 261, + 468, + 360 + ], + "lines": [ + { + "bbox": [ + 320, + 261, + 468, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 261, + 468, + 360 + ], + "type": "image", + "image_path": "41eb5af770a31595bed5f7be6e8cad80621252104ca7ffc4e1980940d8765cf5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 368, + 482, + 488 + ], + "lines": [ + { + "bbox": [ + 304, + 368, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 482, + 488 + ], + "type": "text", + "content": "Fig.6: RepVGG [5] converts a training-time multi-branch block (with skip connection) into an inference-time plain convolutional layer (without skip connection). Through our analytical and empirical analysis, we show that despite the differences in architectures, the gradients across the blocks are similar (See our analysis in the Supp.). Consequently, even with the removal of skip connections, RepVGG cannot mitigate vulnerability to MI attacks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "inline_equation", + "content": "T_{RepVGG}" + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "text", + "content": ". Through our analytical and empirical analysis, we show that the gradients in " + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "inline_equation", + "content": "T_{RepVGG}" + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "text", + "content": " under MI attacks and that in " + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "inline_equation", + "content": "T_{RepVGG}" + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "text", + "content": " are the same. Therefore, removing skip connection in the inference-time architecture " + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "inline_equation", + "content": "T_{RepVGG}" + }, + { + "bbox": [ + 130, + 495, + 482, + 557 + ], + "type": "text", + "content": " could not help mitigate vulnerability to MI. Our detailed analytical analysis can be found in the Supp." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "To empirically validate this, we assess the vulnerability of RepVGG inference-time architecture to MI attacks, comparing it to the training-time architecture. We apply SOTA MI attack, PPA [28], on both the training-time and inference-time RepVGG-A0/B3/D2 architectures. The results for other RepVGG architectures can be found in Supp. We strictly follow the training and inference conversion implementation from the original RepVGG source code for the target classifiers trained on the Facescrub dataset [24]. For the PPA attack [28], we follow the attack setup provided in the original PPA source code. Our results in Tab. 4 clearly show that, despite the removal of skip connections, the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 90, + 447, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "inference-time architecture obtained via RepVGG remains as vulnerable to MI attacks as the training-time architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 301, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 301, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 301, + 331 + ], + "type": "text", + "content": "The skip connection removed architectures in Sec. 3 and those in the study of RepVGG in this section are fundamentally different. During training, RepVGG training-time architecture still receives gradients via skip connection branches. We show that the gradients in RepVGG inference-time architecture remain the same, including the skip connection branch's gradient components even though skip connections are absent in the inference-time architecture. In contrast, our study in Sec. 3 removes skip connections during training, eliminating gradients on skip connections. The study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 141, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 482, + 239 + ], + "type": "text", + "content": "Table 4: Experimental results of RepVGG [5] training-time and inference-time architectures. We strictly follow PPA [28] for the attack setup and evaluation. Here " + }, + { + "bbox": [ + 307, + 141, + 482, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{priv} =" + }, + { + "bbox": [ + 307, + 141, + 482, + 239 + ], + "type": "text", + "content": " Facescrub, " + }, + { + "bbox": [ + 307, + 141, + 482, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{pub} =" + }, + { + "bbox": [ + 307, + 141, + 482, + 239 + ], + "type": "text", + "content": " FFHQ. Despite the removal of skip connections, RepVGG inference-time architecture remains as vulnerable to MI attacks as the training-time architecture." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 241, + 474, + 330 + ], + "blocks": [ + { + "bbox": [ + 315, + 241, + 474, + 330 + ], + "lines": [ + { + "bbox": [ + 315, + 241, + 474, + 330 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 474, + 330 + ], + "type": "table", + "html": "
ArchitectureAcc↑AttAcc↓
RepVGG-A0Training-time94.9085.19
Inference-time94.9086.13
RepVGG-B3Training-time94.5580.69
Inference-time94.5580.61
RepVGG-D2Training-time92.0567.08
Inference-time92.0566.44
", + "image_path": "3d82d1fc9fe717c3b164c32d93d10334cb25a35fed9ae58c288da0c7e4975eaa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 332, + 455, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 455, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 455, + 344 + ], + "type": "text", + "content": "demonstrates the causal effect of skip connections on MI attack accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 363, + 430, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 363, + 430, + 376 + ], + "spans": [ + { + "bbox": [ + 131, + 363, + 430, + 376 + ], + "type": "text", + "content": "5 Model Inversion Resilient Architecture Design" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 301, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 301, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 301, + 666 + ], + "type": "text", + "content": "Our findings so far reveal that skip-connections reinforce MI attacks while existing reparameterization technique, RepVGG [5], to remove skip connection in the inference-time architecture cannot mitigate vulnerability to MI attack. To bridge this gap, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our RoLSS, SSF, and TTS are remarkably simple, maintaining the same training procedure as the original model. Applying RoLSS involves no additional hyperparameters, while SSF/TTS requires only one extra hyperparameter, making them easily applicable to various architectures. In contrast, the SOTA MI defense BiDO requires an extensive grid search for each architecture [26]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 380, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 482, + 490 + ], + "type": "text", + "content": "Table 5: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. " + }, + { + "bbox": [ + 307, + 380, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 307, + 380, + 482, + 490 + ], + "type": "text", + "content": " represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 490, + 474, + 671 + ], + "blocks": [ + { + "bbox": [ + 315, + 490, + 474, + 671 + ], + "lines": [ + { + "bbox": [ + 315, + 490, + 474, + 671 + ], + "spans": [ + { + "bbox": [ + 315, + 490, + 474, + 671 + ], + "type": "table", + "html": "
ArchitectureDefenseAcc↑AttAcc↓Δ↑
ResNet-34No Def.94.6990.78-
BiDO91.6681.982.90
RoLSS (Ours)91.3871.865.72
SSF (Ours)94.2179.7922.90
TTS (Ours)94.4081.6531.48
ResNet-50No Def.94.5882.76-
BiDO91.1258.417.04
RoLSS (Ours)92.8968.448.47
SSF (Ours)93.0574.796.87
TTS (Ours)93.5677.215.44
ResNet-101No Def.94.8683.00-
BiDO90.3167.073.50
RoLSS (Ours)92.4058.689.89
SSF (Ours)93.7971.0611.16
TTS (Ours)94.1677.268.20
ResNet-152No Def.95.4386.51-
BiDO91.8058.147.82
RoLSS (Ours)93.0064.988.86
SSF (Ours)93.7970.719.63
TTS (Ours)93.9773.598.85
", + "image_path": "d70b64ff203e55e48d9f2508a46f34b4c11f9e36a0b135f45c9e82aadbf304e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 406, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 406, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 406, + 129 + ], + "type": "text", + "content": "5.1 Removal of Last Stage Skip-Connection (RoLSS)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": "In Sec. 3, our investigation reveals that removing skip connections in the last stage yields significant degradation in MI attack accuracy, suggesting a promising approach to improve MI robustness from architectural perspective. The MI defense results are presented in Tab. 6 and Tab. 5. Across architectures and skip connection mechanism, the results consistently show that removing the skip connections in last stage (i.e., RoLSS) can improve the MI robustness. Notably, our simple RoLSS achieves highly competitive MI robustness compared to the SOTA MI defense BiDO [26]. For instance, with ResNet-101, our RoLSS improves model accuracy by " + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "inline_equation", + "content": "2.09\\%" + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": ", while the MI attack accuracy degrades by " + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "inline_equation", + "content": "8.49\\%" + }, + { + "bbox": [ + 130, + 140, + 482, + 260 + ], + "type": "text", + "content": ", resulting in superior MI robustness compared to BiDO." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 281, + 350, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 281, + 350, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 350, + 295 + ], + "type": "text", + "content": "5.2 Skip-Connection Scaling Factor (SSF)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 481, + 354 + ], + "type": "text", + "content": "We further propose SSF on top of RoLSS to improve natural accuracy of the model while maintaining competitive MI robustness. For additive skip connections, we introduce a scale factor " + }, + { + "bbox": [ + 130, + 306, + 481, + 354 + ], + "type": "inline_equation", + "content": "0 \\leq k \\leq 1" + }, + { + "bbox": [ + 130, + 306, + 481, + 354 + ], + "type": "text", + "content": " for the signal on the skip connection of the last stage:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 357, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 357, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 261, + 357, + 481, + 369 + ], + "type": "interline_equation", + "content": "z _ {i + 1} = g _ {i} \\left(z _ {i}\\right) + k \\cdot z _ {i} \\tag {3}", + "image_path": "30567a62782c435d096c75213ee1f12fcbbe1a1a06ed0ef4e84c32d8e4f9cf79.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "text", + "content": "Details of SSF for concatenative skip connections can be found in Supp. Our SSF generalizes the skip connection, where " + }, + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "text", + "content": " corresponds to the original skip connection, while " + }, + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "text", + "content": " is similar to our skip connection removal study. With " + }, + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "inline_equation", + "content": "k < 1" + }, + { + "bbox": [ + 130, + 377, + 301, + 484 + ], + "type": "text", + "content": ", gradients can be limited during MI attack, and this could degrade MI." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 301, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 301, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 301, + 629 + ], + "type": "text", + "content": "The results are presented in Tab. 6 and Tab. 5. We apply SSF over RoLSS and set " + }, + { + "bbox": [ + 130, + 486, + 301, + 629 + ], + "type": "inline_equation", + "content": "k = 0.01" + }, + { + "bbox": [ + 130, + 486, + 301, + 629 + ], + "type": "text", + "content": " for all concatenative skip connection architectures and " + }, + { + "bbox": [ + 130, + 486, + 301, + 629 + ], + "type": "inline_equation", + "content": "k = 0.2" + }, + { + "bbox": [ + 130, + 486, + 301, + 629 + ], + "type": "text", + "content": " for all additive skip connection architectures in our experimental setups. Overall, our SSF further improves MI robustness beyond RoLSS and outperforms the SOTA MI defense BiDO [26] across various architectures. Notably, for DenseNet, SSF significantly aids in recovering model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 384, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 482, + 493 + ], + "type": "text", + "content": "Table 6: Our simple RoLSS outperforms SOTA MI defense BiDO [26]. Our further proposed SSF and TTS help recover Acc while offer competitive MI robustness. " + }, + { + "bbox": [ + 307, + 384, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 307, + 384, + 482, + 493 + ], + "type": "text", + "content": " represents the ratio of attack accuracy drop to natural accuracy drop. We could not compare with unsupported BiDO architectures (i.e., DenseNet), as BiDO requires extensive hyperparameter grid search." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 494, + 475, + 613 + ], + "blocks": [ + { + "bbox": [ + 315, + 494, + 475, + 613 + ], + "lines": [ + { + "bbox": [ + 315, + 494, + 475, + 613 + ], + "spans": [ + { + "bbox": [ + 315, + 494, + 475, + 613 + ], + "type": "table", + "html": "
ArchitectureDefenseAcc↑AttAcc↓Δ↑
DenseNet-121No Def.94.6778.09-
RoLSS (Ours)86.8624.486.86
SSF (Ours)91.7356.327.35
DenseNet-161No Def.93.9374.86-
RoLSS (Ours)87.6720.718.65
SSF (Ours)93.7774.273.69
DenseNet-169No Def.94.2877.15-
RoLSS (Ours)72.146.773.18
SSF (Ours)92.9560.9912.15
DenseNet-201No Def.94.3277.62-
RoLSS (Ours)74.2510.243.36
SSF (Ours)93.0965.2110.09
", + "image_path": "046fbf491251145c3d2ee1af7033b371bdb7eac086384d72c18f799d1a542cc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "performance while still mitigating MI attacks, resulting in much improved MI robustness. For example, with DenseNet-201, SSF only incurs a " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " drop in model accuracy while degrading MI attack accuracy by " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\sim 12\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "text", + "content": "5.3 Two-stage Training Scheme (TTS)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 481, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 481, + 169 + ], + "type": "text", + "content": "We further introduce a Two-stage Training Scheme (TTS) on top of RoLSS to improve model accuracy while still maintaining competitive MI robustness. Inspired by Transfer Learning literature [37], TTS consists of two training stages:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "text", + "content": "Stage 1: We train model " + }, + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "text", + "content": " with full skip-connections architecture over " + }, + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "text", + "content": " epochs. This stage ensures the reasonable convergence of " + }, + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "inline_equation", + "content": "\\theta_T" + }, + { + "bbox": [ + 130, + 170, + 500, + 229 + ], + "type": "text", + "content": " with well-backpropagated gradients through the full skip connections architecture. Note that model parameters are far from optimum initially. With full skip connections in this stage, large gradients can be backpropagated in making larger parameter updates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "text", + "content": "Stage 2: We remove skip connections in the last stage, i.e. RoLSS, to create skip connection-removed architecture " + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "text", + "content": ". Then, we continue to train " + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\theta_{T_p}" + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "text", + "content": " epochs. The pre-trained parameters in Stage 1 serves as initialization for " + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\theta_{T_p}" + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "text", + "content": ", thereby aiding the enhanced convergence of " + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 130, + 229, + 481, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "text", + "content": "We build TTS on top of RoLSS. For a fair comparison, the total training epochs for both stages (i.e., " + }, + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "inline_equation", + "content": "M + N" + }, + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "text", + "content": ") match the total training epochs of the original model. Across all setups, we set " + }, + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "inline_equation", + "content": "M = 5" + }, + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "inline_equation", + "content": "N = 95" + }, + { + "bbox": [ + 130, + 277, + 481, + 373 + ], + "type": "text", + "content": ". The results in Tab. 5 demonstrate that TTS outperforms the SOTA MI defense BiDO [26]. Furthermore, TTS improves model accuracy while maintaining competitive MI robustness when compared to our RoLSS. For instance, in the ResNet-34 setup, TTS achieves similar MI attack accuracy as BiDO but maintains comparable model accuracy with No Def. model, achieving very competitive MI robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 388, + 219, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 219, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 219, + 402 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 481, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 481, + 543 + ], + "type": "text", + "content": "We conducted a pioneering study to examine the impact of DNN architecture on SOTA MI attacks. Our findings reveal that skip connections reinforce MI attacks, thereby jeopardizing data privacy. Through extensive MI setups, we find that the skip connections in the last stage is the most critical to MI attacks. Furthermore, our analytical and empirical analysis on RepVGG reveal that the removal of skip connections in the inference-time architecture could not help mitigate the MI vulnerability. Based on our own findings, we propose MI-resilient architecture designs for the first time, including: Removal of Last Stage Skip-Connection (RoLSS), Skip-Connection Scaling Factor (SSF), and Two-stage Training Scheme (TTS). Our MI-resilient architecture designs are remarkably simple to apply and achieve very competitive MI robustness compared to SOTA MI defense." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 559, + 261, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 261, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 261, + 572 + ], + "type": "text", + "content": "7 Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programmes (AISG Award No.: AISG2-TC-2022-007); The Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). This material is based on the research/work support in part by the Changi General Hospital and Singapore University of Technology and Design, under the HealthTech Innovation Fund (HTIF Award No. CGH-SUTD-2021-004)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. An, S., Tao, G., Xu, Q., Liu, Y., Shen, G., Yao, Y., Xu, J., Zhang, X.: Mirror: Model inversion for deep learning network with high fidelity. In: Proceedings of the 29th Network and Distributed System Security Symposium (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "text", + "content": "2. Cazenavette, G., Murdock, C., Lucey, S.: Architectural adversarial robustness: The case for deep pursuit. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7150-7158 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "type": "text", + "content": "3. Chang, X., Zhang, W., Qian, Y., Le Roux, J., Watanabe, S.: End-to-end multispeaker speech recognition with transformer. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6134-6138. IEEE (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 282 + ], + "type": "text", + "content": "4. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16178-16187 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "text", + "content": "5. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "text", + "content": "6. Dippel, J., Vogler, S., Hohne, J.: Towards fine-grained visual representations by combining contrastive learning with image reconstruction and attention-weighted pooling. arXiv preprint arXiv:2104.04323 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 413 + ], + "type": "text", + "content": "7. Dufumier, B., Gori, P., Victor, J., Grigis, A., Wessa, M., Brambilla, P., Favre, P., Polosan, M., Mcdonald, C., Piguet, C.M., et al.: Contrastive learning with continuous proxy meta-data for 3d mri classification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 58-68. Springer (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 413, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 481, + 457 + ], + "type": "text", + "content": "8. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. pp. 1322-1333 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 457, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 491 + ], + "type": "text", + "content": "9. Fredrikson, M., Lantz, E., Jha, S., Lin, S., Page, D., Ristenpart, T.: Privacy in pharmacogenetics: An end-to-end case study of personalized warfarin dosing. In: 23rd USENIX Security Symposium (USENIX Security 14). pp. 17-32 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "text", + "content": "0. Guo, J., Zhu, X., Zhao, C., Cao, D., Lei, Z., Li, S.Z.: Learning meta face recognition in unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6172 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "text", + "content": "1. Han, G., Choi, J., Lee, H., Kim, J.: Reinforcement learning-based black-box model inversion attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20504-20513 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 589 + ], + "type": "text", + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "text", + "content": "3. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "4. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5901-5910 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "16. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 224 + ], + "type": "text", + "content": "17. Khosla, A., Jayadevaprakash, N., Yao, B., Fei-Fei, L.: Novel dataset for fine-grained image categorization. In: First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition. Colorado Springs, CO (June 2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "text", + "content": "18. Krishna, G., Tran, C., Yu, J., Tewfik, A.H.: Speech recognition with no speech or with noisy speech. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1090-1094. IEEE (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "text", + "content": "19. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "text", + "content": "20. Luo, L., Xu, D., Chen, H., Wong, T.T., Heng, P.A.: Pseudo bias-balanced learning for debiased chest x-ray classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 621-631. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "text", + "content": "21. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14234 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 396 + ], + "type": "text", + "content": "22. Mishra, S., Zhang, Y., Zhang, L., Zhang, T., Hu, X.S., Chen, D.Z.: Data-driven deep supervision for skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 721-731. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 439 + ], + "type": "text", + "content": "23. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Radiological reports improve pretraining for localized imaging tasks on chest x-rays. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 647-657. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "24. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "type": "text", + "content": "25. Nguyen, N.B., Chandrasegaran, K., Abdollahzadeh, M., Cheung, N.M.: Rethinking model inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "text", + "content": "26. Peng, X., Liu, F., Zhang, J., Lan, L., Ye, J., Liu, T., Han, B.: Bilateral dependency optimization: Defending against model-inversion attacks. In: KDD (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "27. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "28. Struppek, L., Hintersdorf, D., Correia, A.D.A., Adler, A., Kersting, K.: Plug & play attacks: Towards robust and flexible model inversion attacks. arXiv preprint arXiv:2201.12179 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "29. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "30. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: European conference on computer vision. pp. 459-479. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Koh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 423 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Wang, K.C., Fu, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "32. Wang, T., Zhang, Y., Jia, R.: Improving robustness to model inversion attacks via mutual information regularization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11666-11673 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "type": "text", + "content": "33. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "text", + "content": "34. Yang, J., Chen, H., Yan, J., Chen, X., Yao, J.: Towards better understanding and better generalization of few-shot classification in histology images with contrastive learning (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "text", + "content": "35. Yang, S., Li, Y., Jiang, Y., Xia, S.T.: Backdoor defense via suppressing model shortcuts. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 481, + 324 + ], + "type": "text", + "content": "36. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. In: Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security. pp. 225-240 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 347 + ], + "type": "text", + "content": "37. Yosinski, J., Clune, J., Bengio, Y., Lipson, H.: How transferable are features in deep neural networks? Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 481, + 380 + ], + "type": "text", + "content": "38. Yuan, X., Chen, K., Zhang, J., Zhang, W., Yu, N., Zhang, Y.: Pseudo label-guided model inversion attack via conditional generative adversarial network. Thirty Seventh AAAI Conference on Artificial Intelligence (AAAI 23) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 423 + ], + "type": "text", + "content": "39. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 253-261 (2020)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "On the Vulnerability of Skip Connections to Model Inversion Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_content_list.json b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5352e4c6d11d65de740b327c2ebc488c0afd1202 --- /dev/null +++ b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_content_list.json @@ -0,0 +1,1755 @@ +[ + { + "type": "text", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "text_level": 1, + "bbox": [ + 263, + 140, + 738, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeonseong Kim $^{\\text{①}}$ , Sung-Hoon Yoon $^{\\text{②}}$ , Minseok Kim $^{\\text{③}}$ , and Kuk-Jin Yoon $^{\\text{④}}$", + "bbox": [ + 220, + 212, + 781, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Intelligence Lab., KAIST, Korea {brian617,yoon307,alstjrx1x1,kjyoon}@kaist.ac.kr", + "bbox": [ + 313, + 239, + 687, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. LiDAR semantic segmentation is important for understanding the surrounding environment in autonomous driving. Existing methods assume closed-set situations with the same training and testing label space. However, in the real world, unknown classes not encountered during training may appear during testing, making it difficult to apply existing methodologies. In this paper, we propose a novel on-the-fly category discovery method for LiDAR semantic segmentation, aiming to classify and segment both unknown and known classes instantaneously during test time, achieved solely by learning with known classes in training. To embed instant segmentation capability in an inductive setting, we adopt a hash coding-based model with an expandable prediction space as a baseline. Based on this, dual prototypical learning is proposed to enhance the recognition of the known classes by reducing the sensitivity to intra-class variance. Additionally, we propose a novel mixing-based category learning framework based on representation mixing to improve the discovery capability of unknown classes. The proposed mixing-based framework effectively models out-of-distribution representations and learns to semantically group them during training while distinguishing them from in-distribution representations. Extensive experiments on SemanticKITTI and SemanticPOSS datasets demonstrate the superiority of the proposed method compared to the baselines. The code is available at https://github.com/hskim617/OCDSS.", + "bbox": [ + 261, + 305, + 738, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Autonomous driving $\\cdot$ LiDAR semantic segmentation $\\cdot$ On-the-fly category discovery", + "bbox": [ + 261, + 623, + 738, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 678, + 374, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR semantic segmentation is crucial for recognizing the surrounding environment in autonomous driving and has recently made significant progress [2, 11, 12, 26, 30, 38, 39, 50, 59, 65]. Although these models demonstrate promising results across various datasets, their effectiveness is constrained to closed-set settings, as they are assumed and trained within such scenarios. In real-world scenarios, however, unknown classes are bound to appear during testing, making it incapable of predicting unknown classes for closed-set models as in Fig. 1a.", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 218, + 825, + 364, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09ed99855aaa56f7a326ba3ca593a407832ca8277c28237e84051c843a8f6af9.jpg", + "image_caption": [ + "Fig. 1: Illustration of training and testing scenarios for on-the-fly category discovery in LiDAR semantic segmentation and relevant tasks. The goal of each task is (a) recognizing known classes, (b) further identifying and rejecting unknown classes, (c) learning clustering using unlabeled, unknown class data, and (d) cognizing unknown classes that are not identified during training while recognizing known classes." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 781, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, open-set recognition (OSR) [47], which aims to filter out unknown classes has been explored in diverse fields, including image classification [3,5,9,49,58], semantic segmentation [7,31,37,57], and 3D semantic segmentation [8,35]. Considering the nature of the human cognitive system that can cognize novel visual concepts based on understanding, simply rejecting unknown classes as in OSR (Fig. 1b) is unsatisfactory.", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To narrow the gap between humans and machines by learning novel concepts, novel category discovery (NCD) [15, 23, 24, 28, 63] and on-the-fly category discovery (OCD) [14] are emerging as new research fields and attracting attention. Given labeled data of known classes and unlabeled data of unknown classes, the goal of NCD is to cluster the unlabeled data with the same semantics, aiming to cognize them as in Fig. 1c. Therefore, NCD is formulated as transductive learning in which data of unknown classes are available, making it incapable of predicting unknown classes not identified during training. On the other hand, OCD aims for a more realistic scenario by learning solely from data of known classes and inductively cognizing unknown classes at test time. Furthermore, during testing, the model should independently classify each test sample instantaneously, regardless of the distribution of the test data.", + "bbox": [ + 212, + 553, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we broaden the scope of OCD to LiDAR point clouds and introduce a new problem, on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). As shown in Fig. 1d, our goal is to classify and segment both known and unknown classes instantaneously in an inductive manner during test time, achieved solely by learning with known classes in training. To achieve the goal, one straightforward solution is adopting the previous approach in image classification [14] and applying it to the LiDAR semantic segmentation.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we implemented the hash coding-based model proposed by [14] as a baseline for LiDAR point clouds. This model utilizes hash codes as class descriptors, generated by binarizing each dimension of the learned representation based on its sign. However, extending it to the LiDAR domain entails two limitations.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, due to LiDAR's sparse nature and lack of texture information, hash coding in learned representations is more sensitive to intra-class variance. In practice, we observed that the method for reducing sensitivity to intra-class variance by disentangling sign and magnitude [14] had shown unsatisfactory results in the LiDAR domain (refer Fig. 4a). Secondly, prior work relies on representation learning from known classes to discover unknown classes. Without a dedicated mechanism for discovering unknown classes, only a tentative expectation that the learned representations will form semantically meaningful groups for unknown classes is possible. This reliance, combined with the aforementioned limitation, complicates OCDSS.", + "bbox": [ + 212, + 215, + 784, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, we propose a novel learning method for OCDSS that learns representations robust to intra-class variance in hash coding and suitable for unknown class discovery. First, we propose dual prototypical learning (DPL) that constrain the representation space by introducing extra hash prototypes, effectively reducing the sensitivity of hash codes to intra-class variance. Specifically, the hash prototypes enforce additional constraints to ensure that representations within the same intra-class share the same hash code, dramatically enhancing recognition capability for known classes. Second, we propose a novel mixing-based category learning (MCL) framework to improve the ability to cognize unknown classes. The proposed framework consists of three modules: a mixing module, a discrimination module, and a grouping module. The mixing module models out-of-distribution (OOD) representation through representation mixing and obtains mixed representation to mimic unknown classes during training. The discrimination module then enhances the model's OOD discrimination ability by learning to distinguish the representation of known classes from mixed representations. In addition, the grouping module develops the model's clustering capability for OOD representations by learning to cluster mixed representations within the same semantics. Meanwhile, we integrate the hash prototypes of DPL during MCL to fully exploit the discovering capability for unknown classes. Finally, we introduce an evaluation protocol for OCDSS and evaluate our approach on SemanticKITTI [4] and SemanticPOSS [42] datasets. Extensive experiments demonstrate the superiority of our method compared to baselines, while ablation studies show the effectiveness of each component.", + "bbox": [ + 212, + 377, + 785, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions are four-fold: (I) We introduce a new problem of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). (II) We propose dual prototypical learning that effectively learns representations that are less sensitive to intra-class variation in hash coding. (III) We present a novel mixing-based category learning framework, enabling the model to discover unknown classes. (IV) We introduce a new evaluation protocol for OCDSS and demonstrate the superiority of our method through extensive experiments.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiDAR Semantic Segmentation. LiDAR semantic segmentation is the task of inferring semantic labels to individual 3D points within the LiDAR point cloud. These approaches can be broadly categorized into point-based, projection-based, and voxel-based according to the way they represent and process point clouds. Point-based methods [26,36,44,51] directly operate on points in line with the foundational work PointNet [43] by using multi-layer perceptrons. Despite their advantages, such as minimal information loss and few parameters, applying these methods to large-scale LiDAR point clouds demands heavy computational resources and memory. Projection-based methods have been proposed to address efficiency and leverage advanced 2D convolutional neural network architectures. These methods project LiDAR points onto a range image [1,2,10,13,19,30,39,55, 56] using spherical projection or a bird-eye-view [59]. However, the projection of 3D points into 2D representation might lose some geometric information. Voxel-based methods represent point clouds as 3D voxels and process them through 3D convolution [66] or sparse convolution [11, 12, 20, 38, 50, 65]. In particular, sparse convolution achieves high performance with less computation, so we adopt MinkowskiNet [12] as our backbone architecture.", + "bbox": [ + 212, + 174, + 787, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-set Recognition. Open-set recognition (OSR) was first introduced in [47], extending the closed-set assumption to realistic scenarios where unknown classes appear during testing. As OSR aims to detect unknown classes, previous studies estimate the unknown probability based on extreme value theory [5] or estimate uncertainty through maximum softmax probability [25], maximum logit [3], or Bayesian inference approximation [16, 33]. Other approaches include training a redundancy classifier for unknown classes [54, 64] or using metric learning [9, 22, 48]. Meanwhile, generative model-based approaches employ generative adversarial networks to synthesize unknown class samples [18, 40] or use image reconstruction error for the criterion of determining unknown classes [41, 49, 58]. Unlike these methods for object recognition, studies for semantic segmentation [7, 31, 37, 57], panoptic segmentation [17, 27] in open-set situations have also been proposed. Recently, open-set semantic segmentation for LiDAR point clouds has emerged [8, 35], extending an open-set problem to the LiDAR domain. While OSR methods are aware of unknown classes, they focus on identification, whereas we aim to go beyond achieving semantically cognizing them.", + "bbox": [ + 212, + 431, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Novel Category Discovery. Novel category discovery (NCD), first formalized in [24], is another relevant task that aims to cognize unknown classes in unlabeled data by transferring the knowledge from known classes in labeled data. After DTC [24] proposed deep transfer clustering for NCD, several studies were suggested, achieving significant progress [15, 21, 23, 28, 60, 62, 63]. These approaches first learn representations using supervised learning from labeled known classes, and then discover unknown classes by transferring knowledge through pseudolabeling on unknown classes based on pair-wise similarity [23, 28, 60, 62, 63] or cluster assignment [15, 21]. Additionally, EUMS [61] and NOPS [45] extend the NCD problem to semantic segmentation in images and LiDAR point clouds, respectively. Nevertheless, NCD usually assumes the number of unknown classes", + "bbox": [ + 212, + 674, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as prior knowledge [15, 23, 28, 60, 62, 63], and presuppose all unlabeled data as unknown classes. Hence, generalized category discovery (GCD) was introduced where unlabeled data contain both known and unknown classes [6, 52]. However, as pointed out in [14], two limitations in prior settings still exist: 1) transductive learning and 2) offline evaluation. In prior settings, category discovery is formulated as transductive learning in which data of unknown classes are available during training. Also, offline evaluation constrains the results depending on the entire test data distribution, making instant feedback on each test sample impossible. Therefore, [14] proposed a new task named on-the-fly category discovery (OCD), which aims to recognize known classes and cognize unknown classes instantaneously in an inductive manner. In this study, we tackle the challenging OCD for LiDAR semantic segmentation, which instantaneously segments LiDAR point clouds from both known and unknown classes while learning exclusively from known classes.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 388, + 330, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 3.1, we first elaborate on our problem formulation of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). In Sec. 3.2, we introduce a hash coding-based baseline model for OCDSS that embeds expandable prediction space and instant inference capability within an inductive setting. Then, in Sec. 3.3, we describe our dual prototypical learning for reducing the sensitivity of representation in hash coding, improving the closed-set performance. In Sec. 3.4, we propose a mixing-based category learning framework to enable the model to possess category discovery capabilities for both known and unknown classes. Fig. 2 shows the overview of the proposed method.", + "bbox": [ + 212, + 429, + 787, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 595, + 439, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we formulate the problem of OCDSS, which aims to recognize known classes and cognize unknown classes during testing in an inductive manner. Let $D_{S} = \\{P_{S}\\}$ be a support set used for training and $D_{Q} = \\{P_{Q}\\}$ be a query set for testing. Then, we have 3D point clouds $P_{S} = \\{(x,y)\\} \\in \\mathbb{R}^{3} \\times Y_{S}$ and $P_{Q} = \\{(x,y)\\} \\in \\mathbb{R}^{3} \\times Y_{Q}$ , representing a single LiDAR scan from $D_{S}$ and $D_{Q}$ , respectively. Here, $x$ denotes each point, and $y$ is the corresponding label. $Y_{S}$ and $Y_{Q}$ are the label spaces of support and query sets. Following the setting of OCD [14], $Y_{S}$ defines the known classes observed during training, while $Y_{Q}$ contains both known and unknown classes that appear at testing, i.e. $Y_{S} \\subset Y_{Q}$ . We denote the number of known and unknown classes as $C_k = |Y_S|$ and $C_u = |Y_Q\\backslash Y_S|$ , where we assume $C_u$ is unknown. This is different from existing NCD approaches [15,23,28,45,60,62,63] that assume $C_u$ as a known prior. The goal is to train a model using only $D_{S}$ so that the model can inductively classify and segment point clouds in $D_{Q}$ instantaneously.", + "bbox": [ + 212, + 628, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/24583f59c1bcc33370058ce4388d61fd51813753391fbe005e1a441351288a41.jpg", + "image_caption": [ + "(b) Discrimination Module" + ], + "image_footnote": [ + ": Mixed Representation O:Origin Representation" + ], + "bbox": [ + 222, + 146, + 777, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2f335596e5dffca1f7ac12673f6d604ae2927f6cb773217e7713939c79e39a0.jpg", + "image_caption": [ + "(a) Mixing Module", + "Fig. 2: Overview of the proposed framework. Given the LiDAR point cloud, the model extracts the features through feature extractor $\\mathcal{E}$ , which are then mapped into the prediction space to obtain the representations by the projection head $\\phi$ . Through dual prototypical learning (DPL), the model learns representations suitable for hash code-based clustering, enhancing its recognition ability for known classes. Besides, mixing-based category learning (MCL) improves the model's cognition capability for unknown classes through three modules: mixing module, discrimination module, and grouping module. The final prediction is obtained by the hash code-based clustering." + ], + "image_footnote": [], + "bbox": [ + 241, + 263, + 395, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ae4aa88713eaacb26edc580f3dafde6db11540748ec9a581e8bb513302143fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 268, + 524, + 347 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/683bedbca81f632b7171987c71a7ae038df792982cfa149f3fdcd790c669dd24.jpg", + "image_caption": [ + "(c) Grouping Module" + ], + "image_footnote": [], + "bbox": [ + 547, + 268, + 648, + 347 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/728bfcb94efba52df3e479f8b1f9a42829fd0e8fb6ee39fa3da1d042142278b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 292, + 767, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Hash Coding-based Baseline Model", + "text_level": 1, + "bbox": [ + 214, + 505, + 553, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first introduce a hash coding-based baseline model for OCDSS. To embed the expandable prediction space and instant inference capability while learning the model inductively, we adopt the basic pipeline of hash coding-based model from [14]. Formally, our baseline model consists of a feature extractor $\\mathcal{E}$ and a projection head $\\phi$ . Given point cloud $P_{S} = \\{(x,y)\\}_{i=1}^{N}$ , the feature extractor first extracts the feature representation $f_{i}$ for each point $x_{i}$ . The projection head maps $f_{i}$ to the prediction space, obtaining the projected representation $p_{i} = \\phi(f_{i}) \\in \\mathbb{R}^{L}$ , which is used for the hash coding. Here, $L$ is the dimension of the hash code. Then, the hash code for point $x_{i}$ is obtained by $\\text{hash}(\\phi(\\mathcal{E}(x_{i})))$ , where $\\text{hash}$ is a hashing operation. For an arbitrary vector $\\mathbf{v} = [v_{1}, \\ldots, v_{l}, \\ldots, v_{L}] \\in \\mathbb{R}^{L}$ , the hashing operation is defined as,", + "bbox": [ + 212, + 527, + 787, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh a s h (\\mathbf {v}) = \\left[ v _ {1} ^ {*}, \\dots , v _ {l} ^ {*}, \\dots , v _ {L} ^ {*} \\right], v _ {l} ^ {*} = \\left\\{ \\begin{array}{l l} 1, & v _ {l} \\geq 0 \\\\ 0, & v _ {l} < 0 \\end{array} . \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 700, + 785, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To predict classes of points from $P_{Q}$ during testing, we use the hash code as a class descriptor where points with the same code are considered to be in the same class. This hash coding-based inference allows the model to have an expanded label space to predict new classes beyond the known classes in an inductive setting. Moreover, it also has the advantage of being able to immediately predict the result of each test sample individually.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For representation learning of the baseline model, [14] used supervised contrastive learning [29]; however, there is a lack of clear design choices for representation learning in LiDAR point clouds. For example, in the case of contrastive learning where pair-wise computation between samples is required, it is not suitable due to its high demand for computation and memory. Instead, we use a standard cross-entropy loss on the known classes. To learn a more discriminative representation, we employ parametric prototypes $\\mathcal{G} = \\{g_1, g_2, \\dots, g_{C_k}\\} \\in \\mathbb{R}^{C_k \\times L}$ for known classes as in [7] and replace the classifier as cosine similarity between the prototype and the projected representation $p_i$ . Specifically, we compute the distance based on the cosine similarity as,", + "bbox": [ + 214, + 146, + 787, + 297 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {c o s} \\left(p _ {i}, g _ {j}\\right) = 1 - \\frac {p _ {i} \\cdot g _ {j}}{\\| p _ {i} \\| \\| g _ {j} \\|}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 304, + 785, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, let us define $\\sigma_{\\mathcal{G}}(p_i,y_i)$ that computes the softmax probability of $p_i$ with respect to $\\mathcal{G}$ as follows:", + "bbox": [ + 214, + 339, + 785, + 369 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\mathcal {G}} \\left(p _ {i}, y _ {i}\\right) = \\frac {\\exp \\left(- d _ {\\text {c o s}} \\left(p _ {i} , g _ {y _ {i}}\\right) / \\tau_ {\\text {c o s}}\\right)}{\\sum_ {j = 1} ^ {C _ {k}} \\exp \\left(- d _ {\\text {c o s}} \\left(p _ {i} , g _ {j}\\right) / \\tau_ {\\text {c o s}}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 377, + 785, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $g_{y_i}$ denotes the prototype corresponding to the label $y_i$ and $\\tau_{cos}$ is a temperature parameter. The loss function for the baseline model is defined as,", + "bbox": [ + 214, + 419, + 785, + 450 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b a s e}} = \\sum_ {i} - \\log \\sigma_ {\\mathcal {G}} \\left(p _ {i}, y _ {i}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 457, + 785, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Dual Prototypical Learning", + "text_level": 1, + "bbox": [ + 214, + 503, + 488, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The baseline model is trained to generate similar representations for points belonging to the same class. However, during inference, points are classified into different classes if they do not share the same hash code, which can be sensitive to intra-class variance. Due to the sparse nature and lack of texture information in LiDAR, hash coding in learned representations becomes more sensitive to intra-class variance compared to images. To address this issue, we propose dual prototypical learning (DPL) to learn the representation robust to intra-class variance in hash coding. Specifically, DPL employs extra parametric prototypes to constrain the representation space, denoted as $\\mathcal{H} = \\{h_1, h_2, \\ldots, h_{C_k}\\} \\in \\mathbb{R}^{C_k \\times L}$ , referred to as hash prototypes. Similar to $\\mathcal{G}$ , we calculate the distance between the projected representation $p_i$ and the hash prototypes. In this case, based on the fact that the sign of each dimension in representation determines the hash code, we utilize the L2 distance between signs instead of cosine similarity. In practice, for stable optimization, we use the hyperbolic tangent function, tanh, instead of the sign operation as,", + "bbox": [ + 214, + 525, + 787, + 752 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {h a s h} \\left(p _ {i}, h _ {j}\\right) = \\left\\| t a n h \\left(p _ {i}\\right) - t a n h \\left(h _ {j}\\right) \\right\\| _ {2} ^ {2} / L. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 760, + 785, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to $\\sigma_{\\mathcal{G}}$ in Eq. (3), we define softmax probability $\\sigma_{\\mathcal{H}}(p_i, y_i)$ as follows:", + "bbox": [ + 214, + 784, + 763, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\mathcal {H}} \\left(p _ {i}, y _ {i}\\right) = \\frac {\\exp \\left(- d _ {\\text {h a s h}} \\left(p _ {i} , h _ {y _ {i}}\\right) / \\tau_ {\\text {h a s h}}\\right)}{\\sum_ {j = 1} ^ {C _ {k}} \\exp \\left(- d _ {\\text {h a s h}} \\left(p _ {i} , h _ {j}\\right) / \\tau_ {\\text {h a s h}}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 806, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $h_{y_i}$ denotes the hash prototype corresponding to the label $y_i$ and $\\tau_{hash}$ is a temperature parameter. The prototypical learning loss based on hash prototypes is defined as,", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h a s h}} = \\sum_ {i} - \\log \\sigma_ {\\mathcal {H}} \\left(p _ {i}, y _ {i}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 191, + 785, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Along with $\\mathcal{L}_{base}$ , the final loss function of dual prototypical learning is as follows:", + "bbox": [ + 215, + 224, + 782, + 239 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D P L} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {h a s h}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 257, + 785, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the proposed DPL, we impose additional constraints on representations to make hash codes less sensitive to intra-class variance, improving the model's recognition capability.", + "bbox": [ + 214, + 277, + 782, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Mixing-based Category Learning", + "text_level": 1, + "bbox": [ + 215, + 344, + 532, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Along with the recognition, our goal is to train the model to cognize unknown classes through semantic understanding. Our basic idea to achieve this goal is to mimic unknown classes during training and learn the model to distinguish and group them from known classes for a comprehensive understanding of unknown semantics. To this end, we propose mixing-based category learning (MCL), which consists of a mixing module, a discrimination module, and a grouping module.", + "bbox": [ + 212, + 367, + 784, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mixing Module. We mix the representations of different known classes to mimic the unknown classes during training. Our intention in using representation mixing is that 1) it can effectively model OOD representations, as stated in [64], and 2) it can further mimic diverse unknown classes by modeling OOD representations that share common semantics depending on which classes are mixed. Specifically, we use manifold mixup technique [53] to mix two feature representations $f_{i}$ and $f_{j}$ of different classes $y_{i}$ and $y_{j}$ , where $y_{i} \\neq y_{j}$ . The mixed representation $\\overline{f_{i,j}}$ is calculated as,", + "bbox": [ + 212, + 458, + 784, + 580 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{f _ {i , j}}} = r f _ {i} + (1 - r) f _ {j}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 588, + 784, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $r$ is a mixing ratio. We can also define the imaginary corresponding label $\\overline{y_{i,j}}$ of $\\overline{f_{i,j}}$ , referred to as a mix label. Note that if $(i,j) \\neq (k,l)$ , then $\\overline{y_{i,j}} \\neq \\overline{y_{k,l}}$ . These mixed representations are considered as OOD representations in the following modules.", + "bbox": [ + 212, + 614, + 782, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discrimination Module. In the discrimination module (DM), we train the model to distinguish mixed representations from prototypes. In particular, we learn parametric prototypes $\\mathcal{G}$ and $\\mathcal{H}$ to be discriminative from the mixed representation. To this end, we first project $\\overline{f_{i,j}}$ into projection space and obtain $\\phi(\\overline{f_{i,j}})$ . Then, we enforce $\\phi(\\overline{f_{i,j}})$ not to be close to the prototype of any specific class in the form of entropy maximization as follows:", + "bbox": [ + 214, + 675, + 784, + 765 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s c}} = \\sum_ {i, j} (\\sum_ {k} \\sigma_ {\\mathcal {G}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\log \\sigma_ {\\mathcal {G}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\tag {10} \\\\ + \\sum_ {k} \\sigma_ {\\mathcal {H}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\log \\sigma_ {\\mathcal {H}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right)). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 776, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Notably, we stop the gradient of $\\phi(f_{i,j})$ and only optimize prototypes. In addition, we incorporate mixed representations that are mixed after being projected into the $\\mathcal{L}_{disc}$ to enhance discrimination for more diverse OOD representations.", + "bbox": [ + 214, + 146, + 787, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Grouping Module. We propose a grouping module (GM) to enhance the model's understanding of the underlying semantics. Since the mixed representations might share common semantics depending on which classes are mixed, we can group them using the mix labels. For mathematical brevity, we omit the summation over $i$ and $j$ here. Specifically, we employ contrastive loss [29] and propose contrastive learning on mixed representation based on mixed labels:", + "bbox": [ + 212, + 191, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g p} = \\frac {- 1}{| P (i , j) |} \\sum_ {p \\in P (i, j)} \\log \\sigma_ {A (i, j)} \\left(\\frac {- d _ {c o s} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right) , \\phi \\left(\\overline {{f _ {p}}}\\right)\\right)}{\\tau_ {c o s}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 294, + 785, + 334 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $A(i,j)$ denotes all mixed representations excluding $\\phi(\\overline{f_{i,j}})$ and $P(i,j)$ is all positive pairs of $\\phi(\\overline{f_{i,j}})$ that share the same mix label. $\\sigma_{(A(i,j))}$ refers to a softmax operation with respect to $A(i,j)$ . In practice, we perform sub-sampling on $A(i,j)$ to reduce the computation from heavy pair-wise operations. Additionally, we impose the following constraint so that mixed representations with the same mix label share the hash code as,", + "bbox": [ + 214, + 347, + 787, + 439 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g p - h a s h} = \\log \\sum_ {p \\in P (i, j)} \\exp \\left(d _ {h a s h} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), \\phi \\left(\\overline {{f _ {p}}}\\right)\\right) / \\tau_ {h a s h}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 450, + 785, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The loss function for the proposed MCL is defined as,", + "bbox": [ + 215, + 494, + 604, + 508 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M C L} = \\mathcal {L} _ {\\text {d i s c}} + \\lambda \\left(\\mathcal {L} _ {g p} + \\mathcal {L} _ {g p - h a s h}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 523, + 784, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where, $\\lambda$ is the weight for $\\mathcal{L}_{gp}$ and $\\mathcal{L}_{gp - hash}$ .", + "bbox": [ + 214, + 550, + 535, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, the total loss function of our method is,", + "bbox": [ + 238, + 566, + 588, + 579 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {D P L} + \\mathcal {L} _ {M C L}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 593, + 784, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Evaluation Protocol", + "text_level": 1, + "bbox": [ + 215, + 631, + 446, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Data Split", + "text_level": 1, + "bbox": [ + 215, + 662, + 349, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on two datasets, SemanticKITTI [4] and SemanticPOSS [42]. SemanticKITTI consists of 43,552 frames with point-wise annotations of 19 classes. We follow the official split and use sequences 00 to 07 and 09 to 10 with 19,130 frames for training, and sequence 08 with 4,071 frames for validation. SemanticPOSS contains 2,988 frames with point-wise annotations of 13 classes. We use sequence 03 for validation, while the remaining sequences are used for training, following the official split.", + "bbox": [ + 212, + 688, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Scenarios. We designed three scenarios considering various real-world situations. More precisely, we defined two unknown class types that can exist in the real world: observed-unknown and unobserved-unknown. Observed-unknown", + "bbox": [ + 214, + 795, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e67822f23c1d64112b7ae195a6e40142b46c0c06384434daf27ec2fbaab1e937.jpg", + "table_caption": [ + "Table 1: Data splits of scenario-A on SemanticPOSS and SemanticKITTI datasets. The $\\mathrm{POSS}_i$ and the $\\mathrm{KITTI}_i$ denotes the $i^{th}$ split." + ], + "table_footnote": [], + "table_body": "
SplitUnknown classesSplitUnknown classes
POSS0trashcan, traffic sign, bike, plantsKITTI0bicycle, other-vehicle, trunk, terrain, vegetation
POSS1pole, person, buildingKITTI1bicyclist, traffic-sign, other-ground, fence, road
POSS2rider, fence, groundKITTI2motorcyclist, motorcycle, pole, car, sidewalk
POSS3cone/stone, trunk, carKITTI3person, truck, parking, building
", + "bbox": [ + 228, + 196, + 779, + 267 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "class refers to cases where data was captured during acquisition but remains unlabeled. An example could be cases where instances are not included in the pre-defined category set or when it is challenging to define a class at the time of annotation, resulting in their assignment as a void or an ignore label. On the other hand, unobserved-unknown class corresponds to cases where it first appears during testing without being captured during data acquisition. We then constructed three scenarios based on the presence of each unknown class type.", + "bbox": [ + 212, + 306, + 784, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Scenario-A: In scenario-A, we assess the discovery capability for observed-unknown classes. Following the previous open-set 3D semantic segmentation [8, 35] literature that addresses the observed-unknown classes in testing, we set the labels of unknown classes to be void and ignore them during training. Depending on the choice of the unknown class, we made four splits each for SemanticKITTI and SemanticPOSS, as shown in Tab. 1. The splits are selected considering the diversity of unknown classes and the relationship between classes. We provide details on the selection process in the Supplementary Material.", + "bbox": [ + 212, + 412, + 784, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Scenario-B: In scenario-B, we evaluate the capability to discover unobserved-unknown classes that do not appear in the training data. Specifically, we select relatively infrequent classes from SemanticKITTI, {other-vehicle, bicyclist}, as unknown classes to prevent a significant decrease in the amount of training data. Frames containing any points of these classes are then excluded from training data. As a result, 13,501 out of 19,130 frames are used for training. We did not use SemanticPOSS, as all classes appear in most frames. The evaluation is conducted for all classes, including unobserved-unknown classes, using the same validation data as the official benchmark.", + "bbox": [ + 212, + 534, + 785, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Scenario-C: Both observed-unknown and unobserved-unknown classes are considered in scenario-C. In SemanticKITTI, we select {other-vehicle, parking} as observed-unknown classes, and {bicyclist} as an unobserved-unknown class. Frames containing any points of the unobserved-unknown class are excluded from training data, resulting in 17,954 frames for training, and observed-unknown classes are ignored during training. The evaluation proceeds as in scenario-B.", + "bbox": [ + 212, + 670, + 785, + 760 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2Baselines", + "text_level": 1, + "bbox": [ + 215, + 784, + 339, + 797 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Since OCDSS necessitates inductive learning with known classes while enabling instant inference during testing, many previous approaches are impractical un", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "der our setting. Clustering-based methods do not provide instant inference and demand a substantial amount of memory due to the large number of points in the LiDAR point cloud, making them more impractical. Previous NCD method [45] is also unable to learn under our setting where unknown classes are unavailable. Therefore, following the previous OCD approach [14], we implement MLDG [34] and SMILE [14], representative OCD methods in image classification, in the 3D LiDAR points domain as baselines.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baseline. The baseline is the hash coding-based model introduced in Sec. 3.2. MLDG. MLDG is a meta-learning method for domain generalization. Since OCDSS is more like generalizing to unknown classes from known classes, we simulate introducing unknown classes during training. At each training iteration, we randomly divide known classes into two groups, using one group for meta-training and the other for meta-testing, and then applying MLDG to Baseline. SMILE. SMILE proposes sign-magnitude disentanglement to better capture the class-level semantics. Following [14], we modified the projection head $\\phi$ by replacing the final single MLP layer with two separate MLP layers for sign and magnitude predictions, and subsequently applying SMILE.", + "bbox": [ + 212, + 252, + 785, + 404 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We use the same backbone network and optimization strategy for all baselines. Specifically, MinkUNet34 [12] is used as the feature extractor $\\mathcal{E}$ . For the projection head $\\phi$ , we use multi-layer perceptrons. We use SGD optimizer [46] with learning rate 0.01, momentum 0.9, and weight decay $1e - 4$ with the batch size of 4. We use fixed mixing ratio $r = 0.1$ , $\\tau_{cos} = 0.07$ , $\\tau_{hash} = 0.14$ , and $L = 12$ . More details on baselines and implementation are provided in Supplementary Material.", + "bbox": [ + 212, + 404, + 787, + 508 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 532, + 419, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We adopt Greedy-Hungarian [15] and Strict-Hungarian [52] evaluation protocols by following [14]. During testing, the class descriptor from the hash code directly forms clusters. Then, we match these clusters with ground truth using the Hungarian matching algorithm [32]. Clusters that are not matched are regarded as misclassified. In Greedy-Hungarian, we divide the points into two groups, known and unknown classes, based on the ground truth and then perform Hungarian matching separately within each group. On the other hand, Strict-Hungarian conducts Hungarian matching for all clusters without distinguishing between known and unknown classes. The mean IoU (mIoU) is used for evaluating semantic segmentation performance.", + "bbox": [ + 212, + 559, + 787, + 712 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experimental Results", + "text_level": 1, + "bbox": [ + 214, + 734, + 462, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Comparison", + "text_level": 1, + "bbox": [ + 214, + 768, + 362, + 782 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results. We conduct comparative experiments with baselines on all proposed scenarios. We report the results on scenario-A in Tab. 2, while the results on scenario-B and scenario-C are reported in Tab. 3. As shown in Tables. 2", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/87c6fe5e51c432a0c1fe242483d50cc8c0f257915f7e0b6910f8d91b03b12ba0.jpg", + "table_caption": [ + "Table 2: On-the-fly category discovery results on SemanticPOSS (left) and SemanticKITTI (right) within Scenario-A. Best in bold." + ], + "table_footnote": [], + "table_body": "
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
SplitMethodUnknownKnownAllUnknownKnownAllSplitMethodUnknownKnownAllUnknownKnownAll
POSS0Baseline9.2028.5922.6311.4132.7126.16KITTI0Baseline10.4725.3421.4313.2230.2825.79
MLDG6.3334.6925.967.5540.5430.39MLDG12.6523.1720.4020.5927.3425.56
SMILE5.7940.3429.719.0556.5241.91SMILE9.9428.0923.3121.6638.9634.41
Ours17.8849.6039.8420.3056.5445.39Ours22.5545.9139.7732.9355.1249.28
POSS1Baseline13.2123.1620.8720.0130.1127.78KITTI1Baseline13.5123.7221.0420.4625.1323.90
MLDG15.7228.9725.9123.6635.6532.89MLDG8.9928.8723.6412.7129.7625.27
SMILE15.2032.4728.4818.2246.4039.89SMILE6.5936.6128.7122.6343.3237.88
Ours26.6844.1140.0939.2850.4747.89Ours22.2338.7034.3636.2642.6440.96
POSS2Baseline21.0319.8220.1028.5127.1527.46KITTI2Baseline10.0326.7522.3512.5428.4524.27
MLDG10.0430.1725.5219.8233.4030.27MLDG9.5030.1524.7213.5831.3326.66
SMILE9.0031.6426.4225.2738.2435.24SMILE11.1231.0925.8425.7539.7936.10
Ours24.5241.4037.5139.5946.7145.07Ours14.9838.3932.2323.3942.9037.77
POSS3Baseline4.4728.9223.2818.1228.6326.21KITTI3Baseline12.9326.1223.3421.6927.2726.10
MLDG2.2433.1226.0015.4333.9629.68MLDG7.8025.4421.7216.0027.5125.09
SMILE1.8936.6028.5924.0437.3834.30SMILE19.2130.9728.4932.7336.6035.79
Ours17.0951.3943.4849.1952.5451.77Ours17.7445.1539.3827.9447.8943.69
AvgBaseline11.9825.1221.7219.5129.6526.90AvgBaseline11.7425.4822.0416.9827.7825.02
MLDG8.5831.7425.8516.6235.8930.81MLDG9.7426.9122.6215.7228.9925.65
SMILE7.9735.2628.3019.1544.6437.84SMILE11.7231.6926.5925.6939.6736.05
Ours21.5446.6340.2337.0951.5747.53Ours19.3842.0436.4430.1347.1442.93
", + "bbox": [ + 220, + 198, + 784, + 414 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4b473f90f16e1d694b039d9763398495b41b1322bce980abbb8342107860f516.jpg", + "table_caption": [ + "Table 3: On-the-fly category discovery results on SemanticKITTI within Scenario-B (left) and Scenario-C (right). Best in bold." + ], + "table_footnote": [], + "table_body": "
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAllMethodUnknownKnownAllUnknownKnownAll
Baseline14.6324.8923.8133.6425.9826.79Baseline11.7730.6127.6319.1431.0729.19
MLDG25.8025.6825.7040.7525.4927.10MLDG11.8033.7730.3029.4734.7033.87
SMILE13.1031.6129.6747.6732.5534.14SMILE11.0841.7436.9051.8643.8945.15
Ours21.7347.8345.0842.1248.9648.24Ours16.6647.2942.4537.1947.7646.09
", + "bbox": [ + 225, + 507, + 782, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and 3, in all scenarios, the baseline model exhibits diminished recognition ability for known classes due to its sensitive representation in hash coding, while the MLDG and SMILE demonstrate improved recognition capabilities compared to the baseline. Compared to all baselines, the proposed method achieves superior performance for known classes, demonstrating the effectiveness of the proposed representation learning for OCDSS, including DPL and MCL. Furthermore, the proposed method's comparative performance in the known classes compared to the oracle (All: $46.16\\%$ in SemanticPOSS, All: $49.23\\%$ in SemanticKITTI) indicates its superior recognition ability. MLDG and SMILE exhibit lower results for unknown classes in Strict-Hungarian except for MLDG in scenario-B. This implies that these models tend to become overly focused on the semantics of known classes, making them particularly confused when faced with unknown classes. MLDG performing well for unknown classes in scenario-B is speculated to be the simulation of unknown classes during training, which enhances gen", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/64fc7562fadf2f91ab337567778f5c5ed39e692822014c2410a5ac56e9c3866a.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparison results on $\\mathrm{POSS}_3$ within scenario-A (top), scenario-B (middle), and scenario-C (bottom). Best viewed when zoomed in with colors." + ], + "image_footnote": [], + "bbox": [ + 217, + 135, + 787, + 324 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "eralization to unobserved-unknown classes. However, this approach offers only limited performance improvements for known classes. In the case of SMILE, its favorable performance on unknown classes in the Greedy-Hungarian suggests that it understands the semantics of unknown classes well. Yet, the poor results in the Strict-Hungarian imply that SMILE frequently confuses known and unknown classes. Overall, the proposed method outperforms baselines in both known and unknown classes in scenario-A and scenario-C. In scenario-B as well, our proposed method consistently outperforms in all classes average, further validating the efficacy of our approach.", + "bbox": [ + 212, + 392, + 787, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 3, we visualize the qualitative comparison results on scenario-A (top), scenario-B (middle), and scenario-C (bottom). While MLDG and SMILE fail to predict the known and unknown classes correctly, the proposed method not only successfully recognizes the known classes but also cognizes the unknown classes (top: plant/car, middle: terrain/other-vehicle, bottom: vegetation/bicyclist). This implies that the proposed method learns the better representations for OCDSS through the DPL and MCL. Additional qualitative results are provided in Supplementary Material.", + "bbox": [ + 212, + 527, + 787, + 650 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 674, + 401, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed method, we conduct an ablation study as depicted in Fig. 4. Here, we set the baseline as the model introduced in Sec. 3.2. The experiments are conducted on the SemanticPOSS within scenario-A, and the results are averaged over four splits. Please refer to the Supplementary Material for the discussion with class-wise results and analysis on the hash code dimension $L$ , mixing ratio $r$ , weight $\\lambda$ , and sub-sampling in GM. DPL helps to recognize known classes. Comparing the baseline and baseline+DPL, we observe a significant improvement in performance for the known classes. When the model is trained only with the baseline, hash codes from the", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d57aa633c0bbdfcd46498159525d6ba95e96d08b74271956cfea8f989217ee6f.jpg", + "image_caption": [ + "(a)", + "Fig. 4: (a) Results of using dual prototypical learning (DPL) with the baseline and SMILE. (b) Ablation study of the proposed method. Base refers to the baseline model." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 470, + 223 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/3a4a916bc3b7550ff31d974354b221682773eb45c9bb87f355847e456ba3e3f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAll
(a) Base11.9825.1221.7219.5129.6526.90
(b) Base+DPL5.3340.0831.3317.8946.4338.96
(c) Base+DPL+DM16.4848.2440.2031.3954.8348.58
(d) Base+DPL+DM+GM21.5446.6340.3637.0951.5747.53
", + "bbox": [ + 475, + 164, + 784, + 224 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 617, + 226, + 638, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "learned representations are sensitive to intra-class variation, which is supported by the poor performance for known classes. Conversely, adding hash prototypes through DPL imposes additional constraints on the representation, making it more suitable for hash code-based clustering. To further analyze the effects of DPL, we conduct an experiment by integrating the DPL with the baseline and SMILE [14]. As shown in Fig. 4a, when DPL is used together, the performance for known classes of both baseline and SMILE is significantly improved in both Strict- and Greedy-Hungarian. This implies that the proposed DPL effectively improves the robustness of learned representation under intra-class variation. The proposed DPL benefits both baseline and SMILE, demonstrating its effectiveness in a hash-code-based framework for addressing the OCDSS problem.", + "bbox": [ + 212, + 309, + 784, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "MCL helps to cognize unknown classes. To excavate the ability to discover unknown classes, we proposed mixing-based category learning based on DPL. With DM, the prototypes become discriminative for ID and OOD. This enhancement leads to improved performance not only for known classes but also for unknown classes. The proposed GM further boosts the performance of unknown classes. Comparing the results of using (c) DM and (d) DM+GM in Fig. 4b, we observe approximately $5\\%$ and $5.7\\%$ gains in Strict- and Greedy-Hungarian, respectively. Although performance on known classes experiences a slight reduction, considering the inherent challenge of distinguishing unknown classes that the model has not seen during training, this underscores the necessity of GM.", + "bbox": [ + 212, + 476, + 784, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 652, + 359, + 669 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a new problem named on-the-fly category discovery for LiDAR semantic segmentation. Along with the baseline model based on hash coding, we propose dual prototypical learning to enhance the model's recognition ability. We further design the mixing-based category learning to inductively transfer recognition knowledge from known classes to unknown classes. Extensive experiments demonstrate the effectiveness of the proposed method for recognizing known classes and discovering unknown classes. While the proposed method effectively addresses the problem, performance disparities between the known and unknown classes remain. In this respect, exploring suitable representation learning for discovering unknown classes presents a promising research direction.", + "bbox": [ + 212, + 689, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the Technology Innovation Program (1415187329, 20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636).", + "bbox": [ + 212, + 175, + 787, + 252 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alonso, I., Riazuelo, L., Montesano, L., Murillo, A.C.: 3d-mininet: Learning a 2d representation from point clouds for fast and efficient 3d lidar semantic segmentation. IEEE Robotics and Automation Letters 5(4), 5432-5439 (2020)", + "2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023)", + "3. Basart, S., Mantas, M., Mohammadreza, M., Jacob, S., Dawn, S.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (2022)", + "4. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)", + "5. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016)", + "6. Cao, K., Brbic, M., Leskovec, J.: Open-world semi-supervised learning. In: International Conference on Learning Representations (ICLR) (2022)", + "7. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021)", + "8. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 318-334 (2022)", + "9. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021)", + "10. Chen, K., Oldja, R., Smolyanskiy, N., Birchfield, S., Popov, A., Wehr, D., Eden, I., Pehserl, J.: Mvlidarnet: Real-time multi-class scene understanding for autonomous driving using multiple views. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2288-2294. IEEE (2020)", + "11. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12547-12556 (2021)", + "12. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)" + ], + "bbox": [ + 218, + 301, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: International Symposium on Visual Computing. pp. 207-222. Springer (2020)", + "14. Du, R., Chang, D., Liang, K., Hospedales, T., Song, Y.Z., Ma, Z.: On-the-fly category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11691-11700 (2023)", + "15. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021)", + "16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016)", + "17. Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Navab, N., Busam, B., Tombari, F.: Segmenting known objects and unseen unknowns without prior knowledge. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 19321-19332 (2023)", + "18. Ge, Z., Demyanov, S., Chen, Z., Garnavi, R.: Generative openmax for multi-class open set classification. In: British Machine Vision Conference. BMVA Press (2017)", + "19. Gerdzhev, M., Razani, R., Taghavi, E., Bingbing, L.: Tornado-net: multiview total variation semantic segmentation with diamond inception module. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9543-9549. IEEE (2021)", + "20. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)", + "21. Gu, P., Zhang, C., Xu, R., He, X.: Class-relation knowledge distillation for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16474-16483 (2023)", + "22. Guo, Y., Camporese, G., Yang, W., Sperduti, A., Ballan, L.: Conditional variational capsule network for open set recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 103-111 (2021)", + "23. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. In: International Conference on Learning Representations (2019)", + "24. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019)", + "25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1", + "26. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11108-11117 (2020)", + "27. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021)", + "28. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)", + "30. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)", + "31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)", + "32. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)", + "33. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017)", + "34. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)", + "35. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9425-9434 (2023)", + "36. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems 31 (2018)", + "37. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019)", + "38. Lu, T., Ding, X., Liu, H., Wu, G., Wang, L.: Link: Linear kernel for lidar-based 3d perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1105-1115 (2023)", + "39. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)", + "40. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 613-628 (2018)", + "41. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019)", + "42. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693. IEEE (2020)", + "43. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)", + "44. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "45. Riz, L., Saltori, C., Ricci, E., Poiesi, F.: Novel class discovery for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9393-9402 (2023)", + "46. Ruder, S.: An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747 (2016)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)", + "48. Shu, Y., Shi, Y., Wang, Y., Huang, T., Tian, Y.: P-odn: Prototype-based open deep network for open set recognition. Scientific reports 10(1), 7146 (2020)", + "49. Sun, X., Yang, Z., Zhang, C., Ling, K.V., Peng, G.: Conditional gaussian distribution learning for open set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13480-13489 (2020)", + "50. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European conference on computer vision. pp. 685-702. Springer (2020)", + "51. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)", + "52. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022)", + "53. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019)", + "54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021)", + "55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 1887-1893. IEEE (2018)", + "56. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 4376-4382. IEEE (2019)", + "57. Xia, Y., Zhang, Y., Liu, F., Shen, W., Yuille, A.L.: Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 145-161. Springer (2020)", + "58. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019)", + "59. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020)", + "60. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021)", + "61. Zhao, Y., Zhong, Z., Sebe, N., Lee, G.H.: Novel class discovery in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4340-4349 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021)", + "63. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021)", + "64. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2021)", + "65. Zhou, H., Zhu, X., Song, X., Ma, Y., Wang, Z., Li, H., Lin, D.: Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550 (2020)", + "66. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 369 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "On-the-fly Category Discovery for LiDAR Semantic Segmentation", + "bbox": [ + 292, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_model.json b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b752e8a77cb4b2d3df6081bc9129584807c06a28 --- /dev/null +++ b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_model.json @@ -0,0 +1,2493 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.264, + 0.141, + 0.74, + 0.187 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.213, + 0.782, + 0.228 + ], + "angle": 0, + "content": "Hyeonseong Kim\\(^{\\text{①}}\\), Sung-Hoon Yoon\\(^{\\text{②}}\\), Minseok Kim\\(^{\\text{③}}\\), and Kuk-Jin Yoon\\(^{\\text{④}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.24, + 0.688, + 0.268 + ], + "angle": 0, + "content": "Visual Intelligence Lab., KAIST, Korea {brian617,yoon307,alstjrx1x1,kjyoon}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.306, + 0.74, + 0.611 + ], + "angle": 0, + "content": "Abstract. LiDAR semantic segmentation is important for understanding the surrounding environment in autonomous driving. Existing methods assume closed-set situations with the same training and testing label space. However, in the real world, unknown classes not encountered during training may appear during testing, making it difficult to apply existing methodologies. In this paper, we propose a novel on-the-fly category discovery method for LiDAR semantic segmentation, aiming to classify and segment both unknown and known classes instantaneously during test time, achieved solely by learning with known classes in training. To embed instant segmentation capability in an inductive setting, we adopt a hash coding-based model with an expandable prediction space as a baseline. Based on this, dual prototypical learning is proposed to enhance the recognition of the known classes by reducing the sensitivity to intra-class variance. Additionally, we propose a novel mixing-based category learning framework based on representation mixing to improve the discovery capability of unknown classes. The proposed mixing-based framework effectively models out-of-distribution representations and learns to semantically group them during training while distinguishing them from in-distribution representations. Extensive experiments on SemanticKITTI and SemanticPOSS datasets demonstrate the superiority of the proposed method compared to the baselines. The code is available at https://github.com/hskim617/OCDSS." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.624, + 0.74, + 0.653 + ], + "angle": 0, + "content": "Keywords: Autonomous driving \\(\\cdot\\) LiDAR semantic segmentation \\(\\cdot\\) On-the-fly category discovery" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.679, + 0.375, + 0.695 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.817 + ], + "angle": 0, + "content": "LiDAR semantic segmentation is crucial for recognizing the surrounding environment in autonomous driving and has recently made significant progress [2, 11, 12, 26, 30, 38, 39, 50, 59, 65]. Although these models demonstrate promising results across various datasets, their effectiveness is constrained to closed-set settings, as they are assumed and trained within such scenarios. In real-world scenarios, however, unknown classes are bound to appear during testing, making it incapable of predicting unknown classes for closed-set models as in Fig. 1a." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.365, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.782, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 1: Illustration of training and testing scenarios for on-the-fly category discovery in LiDAR semantic segmentation and relevant tasks. The goal of each task is (a) recognizing known classes, (b) further identifying and rejecting unknown classes, (c) learning clustering using unlabeled, unknown class data, and (d) cognizing unknown classes that are not identified during training while recognizing known classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.553 + ], + "angle": 0, + "content": "To address this challenge, open-set recognition (OSR) [47], which aims to filter out unknown classes has been explored in diverse fields, including image classification [3,5,9,49,58], semantic segmentation [7,31,37,57], and 3D semantic segmentation [8,35]. Considering the nature of the human cognitive system that can cognize novel visual concepts based on understanding, simply rejecting unknown classes as in OSR (Fig. 1b) is unsatisfactory." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.734 + ], + "angle": 0, + "content": "To narrow the gap between humans and machines by learning novel concepts, novel category discovery (NCD) [15, 23, 24, 28, 63] and on-the-fly category discovery (OCD) [14] are emerging as new research fields and attracting attention. Given labeled data of known classes and unlabeled data of unknown classes, the goal of NCD is to cluster the unlabeled data with the same semantics, aiming to cognize them as in Fig. 1c. Therefore, NCD is formulated as transductive learning in which data of unknown classes are available, making it incapable of predicting unknown classes not identified during training. On the other hand, OCD aims for a more realistic scenario by learning solely from data of known classes and inductively cognizing unknown classes at test time. Furthermore, during testing, the model should independently classify each test sample instantaneously, regardless of the distribution of the test data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we broaden the scope of OCD to LiDAR point clouds and introduce a new problem, on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). As shown in Fig. 1d, our goal is to classify and segment both known and unknown classes instantaneously in an inductive manner during test time, achieved solely by learning with known classes in training. To achieve the goal, one straightforward solution is adopting the previous approach in image classification [14] and applying it to the LiDAR semantic segmentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Therefore, we implemented the hash coding-based model proposed by [14] as a baseline for LiDAR point clouds. This model utilizes hash codes as class descriptors, generated by binarizing each dimension of the learned representation based on its sign. However, extending it to the LiDAR domain entails two limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.217, + 0.785, + 0.368 + ], + "angle": 0, + "content": "Firstly, due to LiDAR's sparse nature and lack of texture information, hash coding in learned representations is more sensitive to intra-class variance. In practice, we observed that the method for reducing sensitivity to intra-class variance by disentangling sign and magnitude [14] had shown unsatisfactory results in the LiDAR domain (refer Fig. 4a). Secondly, prior work relies on representation learning from known classes to discover unknown classes. Without a dedicated mechanism for discovering unknown classes, only a tentative expectation that the learned representations will form semantically meaningful groups for unknown classes is possible. This reliance, combined with the aforementioned limitation, complicates OCDSS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.787, + 0.726 + ], + "angle": 0, + "content": "Therefore, we propose a novel learning method for OCDSS that learns representations robust to intra-class variance in hash coding and suitable for unknown class discovery. First, we propose dual prototypical learning (DPL) that constrain the representation space by introducing extra hash prototypes, effectively reducing the sensitivity of hash codes to intra-class variance. Specifically, the hash prototypes enforce additional constraints to ensure that representations within the same intra-class share the same hash code, dramatically enhancing recognition capability for known classes. Second, we propose a novel mixing-based category learning (MCL) framework to improve the ability to cognize unknown classes. The proposed framework consists of three modules: a mixing module, a discrimination module, and a grouping module. The mixing module models out-of-distribution (OOD) representation through representation mixing and obtains mixed representation to mimic unknown classes during training. The discrimination module then enhances the model's OOD discrimination ability by learning to distinguish the representation of known classes from mixed representations. In addition, the grouping module develops the model's clustering capability for OOD representations by learning to cluster mixed representations within the same semantics. Meanwhile, we integrate the hash prototypes of DPL during MCL to fully exploit the discovering capability for unknown classes. Finally, we introduce an evaluation protocol for OCDSS and evaluate our approach on SemanticKITTI [4] and SemanticPOSS [42] datasets. Extensive experiments demonstrate the superiority of our method compared to baselines, while ablation studies show the effectiveness of each component." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our main contributions are four-fold: (I) We introduce a new problem of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). (II) We propose dual prototypical learning that effectively learns representations that are less sensitive to intra-class variation in hash coding. (III) We present a novel mixing-based category learning framework, enabling the model to discover unknown classes. (IV) We introduce a new evaluation protocol for OCDSS and demonstrate the superiority of our method through extensive experiments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.432 + ], + "angle": 0, + "content": "LiDAR Semantic Segmentation. LiDAR semantic segmentation is the task of inferring semantic labels to individual 3D points within the LiDAR point cloud. These approaches can be broadly categorized into point-based, projection-based, and voxel-based according to the way they represent and process point clouds. Point-based methods [26,36,44,51] directly operate on points in line with the foundational work PointNet [43] by using multi-layer perceptrons. Despite their advantages, such as minimal information loss and few parameters, applying these methods to large-scale LiDAR point clouds demands heavy computational resources and memory. Projection-based methods have been proposed to address efficiency and leverage advanced 2D convolutional neural network architectures. These methods project LiDAR points onto a range image [1,2,10,13,19,30,39,55, 56] using spherical projection or a bird-eye-view [59]. However, the projection of 3D points into 2D representation might lose some geometric information. Voxel-based methods represent point clouds as 3D voxels and process them through 3D convolution [66] or sparse convolution [11, 12, 20, 38, 50, 65]. In particular, sparse convolution achieves high performance with less computation, so we adopt MinkowskiNet [12] as our backbone architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Open-set Recognition. Open-set recognition (OSR) was first introduced in [47], extending the closed-set assumption to realistic scenarios where unknown classes appear during testing. As OSR aims to detect unknown classes, previous studies estimate the unknown probability based on extreme value theory [5] or estimate uncertainty through maximum softmax probability [25], maximum logit [3], or Bayesian inference approximation [16, 33]. Other approaches include training a redundancy classifier for unknown classes [54, 64] or using metric learning [9, 22, 48]. Meanwhile, generative model-based approaches employ generative adversarial networks to synthesize unknown class samples [18, 40] or use image reconstruction error for the criterion of determining unknown classes [41, 49, 58]. Unlike these methods for object recognition, studies for semantic segmentation [7, 31, 37, 57], panoptic segmentation [17, 27] in open-set situations have also been proposed. Recently, open-set semantic segmentation for LiDAR point clouds has emerged [8, 35], extending an open-set problem to the LiDAR domain. While OSR methods are aware of unknown classes, they focus on identification, whereas we aim to go beyond achieving semantically cognizing them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Novel Category Discovery. Novel category discovery (NCD), first formalized in [24], is another relevant task that aims to cognize unknown classes in unlabeled data by transferring the knowledge from known classes in labeled data. After DTC [24] proposed deep transfer clustering for NCD, several studies were suggested, achieving significant progress [15, 21, 23, 28, 60, 62, 63]. These approaches first learn representations using supervised learning from labeled known classes, and then discover unknown classes by transferring knowledge through pseudolabeling on unknown classes based on pair-wise similarity [23, 28, 60, 62, 63] or cluster assignment [15, 21]. Additionally, EUMS [61] and NOPS [45] extend the NCD problem to semantic segmentation in images and LiDAR point clouds, respectively. Nevertheless, NCD usually assumes the number of unknown classes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "as prior knowledge [15, 23, 28, 60, 62, 63], and presuppose all unlabeled data as unknown classes. Hence, generalized category discovery (GCD) was introduced where unlabeled data contain both known and unknown classes [6, 52]. However, as pointed out in [14], two limitations in prior settings still exist: 1) transductive learning and 2) offline evaluation. In prior settings, category discovery is formulated as transductive learning in which data of unknown classes are available during training. Also, offline evaluation constrains the results depending on the entire test data distribution, making instant feedback on each test sample impossible. Therefore, [14] proposed a new task named on-the-fly category discovery (OCD), which aims to recognize known classes and cognize unknown classes instantaneously in an inductive manner. In this study, we tackle the challenging OCD for LiDAR semantic segmentation, which instantaneously segments LiDAR point clouds from both known and unknown classes while learning exclusively from known classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.39, + 0.331, + 0.406 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.567 + ], + "angle": 0, + "content": "In Sec. 3.1, we first elaborate on our problem formulation of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). In Sec. 3.2, we introduce a hash coding-based baseline model for OCDSS that embeds expandable prediction space and instant inference capability within an inductive setting. Then, in Sec. 3.3, we describe our dual prototypical learning for reducing the sensitivity of representation in hash coding, improving the closed-set performance. In Sec. 3.4, we propose a mixing-based category learning framework to enable the model to possess category discovery capabilities for both known and unknown classes. Fig. 2 shows the overview of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.44, + 0.61 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In this section, we formulate the problem of OCDSS, which aims to recognize known classes and cognize unknown classes during testing in an inductive manner. Let \\( D_{S} = \\{P_{S}\\} \\) be a support set used for training and \\( D_{Q} = \\{P_{Q}\\} \\) be a query set for testing. Then, we have 3D point clouds \\( P_{S} = \\{(x,y)\\} \\in \\mathbb{R}^{3} \\times Y_{S} \\) and \\( P_{Q} = \\{(x,y)\\} \\in \\mathbb{R}^{3} \\times Y_{Q} \\), representing a single LiDAR scan from \\( D_{S} \\) and \\( D_{Q} \\), respectively. Here, \\( x \\) denotes each point, and \\( y \\) is the corresponding label. \\( Y_{S} \\) and \\( Y_{Q} \\) are the label spaces of support and query sets. Following the setting of OCD [14], \\( Y_{S} \\) defines the known classes observed during training, while \\( Y_{Q} \\) contains both known and unknown classes that appear at testing, i.e. \\( Y_{S} \\subset Y_{Q} \\). We denote the number of known and unknown classes as \\( C_k = |Y_S| \\) and \\( C_u = |Y_Q\\backslash Y_S| \\), where we assume \\( C_u \\) is unknown. This is different from existing NCD approaches [15,23,28,45,60,62,63] that assume \\( C_u \\) as a known prior. The goal is to train a model using only \\( D_{S} \\) so that the model can inductively classify and segment point clouds in \\( D_{Q} \\) instantaneously." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.778, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.257, + 0.346, + 0.267 + ], + "angle": 0, + "content": "(a) Mixing Module" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.265, + 0.396, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.258, + 0.522, + 0.267 + ], + "angle": 0, + "content": "(b) Discrimination Module" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.269, + 0.526, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.258, + 0.635, + 0.268 + ], + "angle": 0, + "content": "(c) Grouping Module" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.27, + 0.649, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.664, + 0.27, + 0.769, + 0.289 + ], + "angle": 0, + "content": ": Mixed Representation O:Origin Representation" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.293, + 0.769, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed framework. Given the LiDAR point cloud, the model extracts the features through feature extractor \\(\\mathcal{E}\\), which are then mapped into the prediction space to obtain the representations by the projection head \\(\\phi\\). Through dual prototypical learning (DPL), the model learns representations suitable for hash code-based clustering, enhancing its recognition ability for known classes. Besides, mixing-based category learning (MCL) improves the model's cognition capability for unknown classes through three modules: mixing module, discrimination module, and grouping module. The final prediction is obtained by the hash code-based clustering." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.506, + 0.555, + 0.521 + ], + "angle": 0, + "content": "3.2 Hash Coding-based Baseline Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.789, + 0.695 + ], + "angle": 0, + "content": "We first introduce a hash coding-based baseline model for OCDSS. To embed the expandable prediction space and instant inference capability while learning the model inductively, we adopt the basic pipeline of hash coding-based model from [14]. Formally, our baseline model consists of a feature extractor \\(\\mathcal{E}\\) and a projection head \\(\\phi\\). Given point cloud \\(P_{S} = \\{(x,y)\\}_{i=1}^{N}\\), the feature extractor first extracts the feature representation \\(f_{i}\\) for each point \\(x_{i}\\). The projection head maps \\(f_{i}\\) to the prediction space, obtaining the projected representation \\(p_{i} = \\phi(f_{i}) \\in \\mathbb{R}^{L}\\), which is used for the hash coding. Here, \\(L\\) is the dimension of the hash code. Then, the hash code for point \\(x_{i}\\) is obtained by \\(\\text{hash}(\\phi(\\mathcal{E}(x_{i})))\\), where \\(\\text{hash}\\) is a hashing operation. For an arbitrary vector \\(\\mathbf{v} = [v_{1}, \\ldots, v_{l}, \\ldots, v_{L}] \\in \\mathbb{R}^{L}\\), the hashing operation is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "\\[\nh a s h (\\mathbf {v}) = \\left[ v _ {1} ^ {*}, \\dots , v _ {l} ^ {*}, \\dots , v _ {L} ^ {*} \\right], v _ {l} ^ {*} = \\left\\{ \\begin{array}{l l} 1, & v _ {l} \\geq 0 \\\\ 0, & v _ {l} < 0 \\end{array} . \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To predict classes of points from \\( P_{Q} \\) during testing, we use the hash code as a class descriptor where points with the same code are considered to be in the same class. This hash coding-based inference allows the model to have an expanded label space to predict new classes beyond the known classes in an inductive setting. Moreover, it also has the advantage of being able to immediately predict the result of each test sample individually." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "For representation learning of the baseline model, [14] used supervised contrastive learning [29]; however, there is a lack of clear design choices for representation learning in LiDAR point clouds. For example, in the case of contrastive learning where pair-wise computation between samples is required, it is not suitable due to its high demand for computation and memory. Instead, we use a standard cross-entropy loss on the known classes. To learn a more discriminative representation, we employ parametric prototypes \\(\\mathcal{G} = \\{g_1, g_2, \\dots, g_{C_k}\\} \\in \\mathbb{R}^{C_k \\times L}\\) for known classes as in [7] and replace the classifier as cosine similarity between the prototype and the projected representation \\(p_i\\). Specifically, we compute the distance based on the cosine similarity as," + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.305, + 0.786, + 0.335 + ], + "angle": 0, + "content": "\\[\nd _ {c o s} \\left(p _ {i}, g _ {j}\\right) = 1 - \\frac {p _ {i} \\cdot g _ {j}}{\\| p _ {i} \\| \\| g _ {j} \\|}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.34, + 0.787, + 0.371 + ], + "angle": 0, + "content": "Then, let us define \\(\\sigma_{\\mathcal{G}}(p_i,y_i)\\) that computes the softmax probability of \\(p_i\\) with respect to \\(\\mathcal{G}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.378, + 0.786, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\mathcal {G}} \\left(p _ {i}, y _ {i}\\right) = \\frac {\\exp \\left(- d _ {\\text {c o s}} \\left(p _ {i} , g _ {y _ {i}}\\right) / \\tau_ {\\text {c o s}}\\right)}{\\sum_ {j = 1} ^ {C _ {k}} \\exp \\left(- d _ {\\text {c o s}} \\left(p _ {i} , g _ {j}\\right) / \\tau_ {\\text {c o s}}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.42, + 0.787, + 0.451 + ], + "angle": 0, + "content": "where \\( g_{y_i} \\) denotes the prototype corresponding to the label \\( y_i \\) and \\( \\tau_{cos} \\) is a temperature parameter. The loss function for the baseline model is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.458, + 0.786, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b a s e}} = \\sum_ {i} - \\log \\sigma_ {\\mathcal {G}} \\left(p _ {i}, y _ {i}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.505, + 0.49, + 0.521 + ], + "angle": 0, + "content": "3.3 Dual Prototypical Learning" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.526, + 0.788, + 0.753 + ], + "angle": 0, + "content": "The baseline model is trained to generate similar representations for points belonging to the same class. However, during inference, points are classified into different classes if they do not share the same hash code, which can be sensitive to intra-class variance. Due to the sparse nature and lack of texture information in LiDAR, hash coding in learned representations becomes more sensitive to intra-class variance compared to images. To address this issue, we propose dual prototypical learning (DPL) to learn the representation robust to intra-class variance in hash coding. Specifically, DPL employs extra parametric prototypes to constrain the representation space, denoted as \\(\\mathcal{H} = \\{h_1, h_2, \\ldots, h_{C_k}\\} \\in \\mathbb{R}^{C_k \\times L}\\), referred to as hash prototypes. Similar to \\(\\mathcal{G}\\), we calculate the distance between the projected representation \\(p_i\\) and the hash prototypes. In this case, based on the fact that the sign of each dimension in representation determines the hash code, we utilize the L2 distance between signs instead of cosine similarity. In practice, for stable optimization, we use the hyperbolic tangent function, tanh, instead of the sign operation as," + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.761, + 0.786, + 0.78 + ], + "angle": 0, + "content": "\\[\nd _ {h a s h} \\left(p _ {i}, h _ {j}\\right) = \\left\\| t a n h \\left(p _ {i}\\right) - t a n h \\left(h _ {j}\\right) \\right\\| _ {2} ^ {2} / L. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.785, + 0.764, + 0.801 + ], + "angle": 0, + "content": "Similar to \\(\\sigma_{\\mathcal{G}}\\) in Eq. (3), we define softmax probability \\(\\sigma_{\\mathcal{H}}(p_i, y_i)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.808, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\mathcal {H}} \\left(p _ {i}, y _ {i}\\right) = \\frac {\\exp \\left(- d _ {\\text {h a s h}} \\left(p _ {i} , h _ {y _ {i}}\\right) / \\tau_ {\\text {h a s h}}\\right)}{\\sum_ {j = 1} ^ {C _ {k}} \\exp \\left(- d _ {\\text {h a s h}} \\left(p _ {i} , h _ {j}\\right) / \\tau_ {\\text {h a s h}}\\right)}, \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "where \\( h_{y_i} \\) denotes the hash prototype corresponding to the label \\( y_i \\) and \\( \\tau_{hash} \\) is a temperature parameter. The prototypical learning loss based on hash prototypes is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.192, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h a s h}} = \\sum_ {i} - \\log \\sigma_ {\\mathcal {H}} \\left(p _ {i}, y _ {i}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.226, + 0.784, + 0.241 + ], + "angle": 0, + "content": "Along with \\(\\mathcal{L}_{base}\\), the final loss function of dual prototypical learning is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.258, + 0.786, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D P L} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {h a s h}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.784, + 0.324 + ], + "angle": 0, + "content": "With the proposed DPL, we impose additional constraints on representations to make hash codes less sensitive to intra-class variance, improving the model's recognition capability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.345, + 0.533, + 0.361 + ], + "angle": 0, + "content": "3.4 Mixing-based Category Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.785, + 0.459 + ], + "angle": 0, + "content": "Along with the recognition, our goal is to train the model to cognize unknown classes through semantic understanding. Our basic idea to achieve this goal is to mimic unknown classes during training and learn the model to distinguish and group them from known classes for a comprehensive understanding of unknown semantics. To this end, we propose mixing-based category learning (MCL), which consists of a mixing module, a discrimination module, and a grouping module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Mixing Module. We mix the representations of different known classes to mimic the unknown classes during training. Our intention in using representation mixing is that 1) it can effectively model OOD representations, as stated in [64], and 2) it can further mimic diverse unknown classes by modeling OOD representations that share common semantics depending on which classes are mixed. Specifically, we use manifold mixup technique [53] to mix two feature representations \\( f_{i} \\) and \\( f_{j} \\) of different classes \\( y_{i} \\) and \\( y_{j} \\), where \\( y_{i} \\neq y_{j} \\). The mixed representation \\( \\overline{f_{i,j}} \\) is calculated as," + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.589, + 0.785, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\overline {{f _ {i , j}}} = r f _ {i} + (1 - r) f _ {j}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.784, + 0.675 + ], + "angle": 0, + "content": "where \\( r \\) is a mixing ratio. We can also define the imaginary corresponding label \\( \\overline{y_{i,j}} \\) of \\( \\overline{f_{i,j}} \\), referred to as a mix label. Note that if \\( (i,j) \\neq (k,l) \\), then \\( \\overline{y_{i,j}} \\neq \\overline{y_{k,l}} \\). These mixed representations are considered as OOD representations in the following modules." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.676, + 0.785, + 0.766 + ], + "angle": 0, + "content": "Discrimination Module. In the discrimination module (DM), we train the model to distinguish mixed representations from prototypes. In particular, we learn parametric prototypes \\(\\mathcal{G}\\) and \\(\\mathcal{H}\\) to be discriminative from the mixed representation. To this end, we first project \\(\\overline{f_{i,j}}\\) into projection space and obtain \\(\\phi(\\overline{f_{i,j}})\\). Then, we enforce \\(\\phi(\\overline{f_{i,j}})\\) not to be close to the prototype of any specific class in the form of entropy maximization as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.777, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s c}} = \\sum_ {i, j} (\\sum_ {k} \\sigma_ {\\mathcal {G}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\log \\sigma_ {\\mathcal {G}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\tag {10} \\\\ + \\sum_ {k} \\sigma_ {\\mathcal {H}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\log \\sigma_ {\\mathcal {H}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right)). \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.192 + ], + "angle": 0, + "content": "Notably, we stop the gradient of \\(\\phi(f_{i,j})\\) and only optimize prototypes. In addition, we incorporate mixed representations that are mixed after being projected into the \\(\\mathcal{L}_{disc}\\) to enhance discrimination for more diverse OOD representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.789, + 0.284 + ], + "angle": 0, + "content": "Grouping Module. We propose a grouping module (GM) to enhance the model's understanding of the underlying semantics. Since the mixed representations might share common semantics depending on which classes are mixed, we can group them using the mix labels. For mathematical brevity, we omit the summation over \\( i \\) and \\( j \\) here. Specifically, we employ contrastive loss [29] and propose contrastive learning on mixed representation based on mixed labels:" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.295, + 0.786, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g p} = \\frac {- 1}{| P (i , j) |} \\sum_ {p \\in P (i, j)} \\log \\sigma_ {A (i, j)} \\left(\\frac {- d _ {c o s} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right) , \\phi \\left(\\overline {{f _ {p}}}\\right)\\right)}{\\tau_ {c o s}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.348, + 0.789, + 0.44 + ], + "angle": 0, + "content": "where \\( A(i,j) \\) denotes all mixed representations excluding \\( \\phi(\\overline{f_{i,j}}) \\) and \\( P(i,j) \\) is all positive pairs of \\( \\phi(\\overline{f_{i,j}}) \\) that share the same mix label. \\( \\sigma_{(A(i,j))} \\) refers to a softmax operation with respect to \\( A(i,j) \\). In practice, we perform sub-sampling on \\( A(i,j) \\) to reduce the computation from heavy pair-wise operations. Additionally, we impose the following constraint so that mixed representations with the same mix label share the hash code as," + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.452, + 0.786, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g p - h a s h} = \\log \\sum_ {p \\in P (i, j)} \\exp \\left(d _ {h a s h} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), \\phi \\left(\\overline {{f _ {p}}}\\right)\\right) / \\tau_ {h a s h}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.495, + 0.606, + 0.51 + ], + "angle": 0, + "content": "The loss function for the proposed MCL is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.524, + 0.785, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M C L} = \\mathcal {L} _ {\\text {d i s c}} + \\lambda \\left(\\mathcal {L} _ {g p} + \\mathcal {L} _ {g p - h a s h}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.551, + 0.536, + 0.566 + ], + "angle": 0, + "content": "where, \\(\\lambda\\) is the weight for \\(\\mathcal{L}_{gp}\\) and \\(\\mathcal{L}_{gp - hash}\\)." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.567, + 0.589, + 0.58 + ], + "angle": 0, + "content": "Finally, the total loss function of our method is," + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.594, + 0.785, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {D P L} + \\mathcal {L} _ {M C L}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.447, + 0.648 + ], + "angle": 0, + "content": "4 Evaluation Protocol" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.35, + 0.68 + ], + "angle": 0, + "content": "4.1 Data Split" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on two datasets, SemanticKITTI [4] and SemanticPOSS [42]. SemanticKITTI consists of 43,552 frames with point-wise annotations of 19 classes. We follow the official split and use sequences 00 to 07 and 09 to 10 with 19,130 frames for training, and sequence 08 with 4,071 frames for validation. SemanticPOSS contains 2,988 frames with point-wise annotations of 13 classes. We use sequence 03 for validation, while the remaining sequences are used for training, following the official split." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Scenarios. We designed three scenarios considering various real-world situations. More precisely, we defined two unknown class types that can exist in the real world: observed-unknown and unobserved-unknown. Observed-unknown" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Data splits of scenario-A on SemanticPOSS and SemanticKITTI datasets. The \\(\\mathrm{POSS}_i\\) and the \\(\\mathrm{KITTI}_i\\) denotes the \\(i^{th}\\) split." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.198, + 0.78, + 0.268 + ], + "angle": 0, + "content": "
SplitUnknown classesSplitUnknown classes
POSS0trashcan, traffic sign, bike, plantsKITTI0bicycle, other-vehicle, trunk, terrain, vegetation
POSS1pole, person, buildingKITTI1bicyclist, traffic-sign, other-ground, fence, road
POSS2rider, fence, groundKITTI2motorcyclist, motorcycle, pole, car, sidewalk
POSS3cone/stone, trunk, carKITTI3person, truck, parking, building
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.785, + 0.413 + ], + "angle": 0, + "content": "class refers to cases where data was captured during acquisition but remains unlabeled. An example could be cases where instances are not included in the pre-defined category set or when it is challenging to define a class at the time of annotation, resulting in their assignment as a void or an ignore label. On the other hand, unobserved-unknown class corresponds to cases where it first appears during testing without being captured during data acquisition. We then constructed three scenarios based on the presence of each unknown class type." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.785, + 0.534 + ], + "angle": 0, + "content": "Scenario-A: In scenario-A, we assess the discovery capability for observed-unknown classes. Following the previous open-set 3D semantic segmentation [8, 35] literature that addresses the observed-unknown classes in testing, we set the labels of unknown classes to be void and ignore them during training. Depending on the choice of the unknown class, we made four splits each for SemanticKITTI and SemanticPOSS, as shown in Tab. 1. The splits are selected considering the diversity of unknown classes and the relationship between classes. We provide details on the selection process in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.787, + 0.669 + ], + "angle": 0, + "content": "Scenario-B: In scenario-B, we evaluate the capability to discover unobserved-unknown classes that do not appear in the training data. Specifically, we select relatively infrequent classes from SemanticKITTI, {other-vehicle, bicyclist}, as unknown classes to prevent a significant decrease in the amount of training data. Frames containing any points of these classes are then excluded from training data. As a result, 13,501 out of 19,130 frames are used for training. We did not use SemanticPOSS, as all classes appear in most frames. The evaluation is conducted for all classes, including unobserved-unknown classes, using the same validation data as the official benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.787, + 0.761 + ], + "angle": 0, + "content": "Scenario-C: Both observed-unknown and unobserved-unknown classes are considered in scenario-C. In SemanticKITTI, we select {other-vehicle, parking} as observed-unknown classes, and {bicyclist} as an unobserved-unknown class. Frames containing any points of the unobserved-unknown class are excluded from training data, resulting in 17,954 frames for training, and observed-unknown classes are ignored during training. The evaluation proceeds as in scenario-B." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.341, + 0.799 + ], + "angle": 0, + "content": "4.2Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Since OCDSS necessitates inductive learning with known classes while enabling instant inference during testing, many previous approaches are impractical un" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "der our setting. Clustering-based methods do not provide instant inference and demand a substantial amount of memory due to the large number of points in the LiDAR point cloud, making them more impractical. Previous NCD method [45] is also unable to learn under our setting where unknown classes are unavailable. Therefore, following the previous OCD approach [14], we implement MLDG [34] and SMILE [14], representative OCD methods in image classification, in the 3D LiDAR points domain as baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.786, + 0.405 + ], + "angle": 0, + "content": "Baseline. The baseline is the hash coding-based model introduced in Sec. 3.2. MLDG. MLDG is a meta-learning method for domain generalization. Since OCDSS is more like generalizing to unknown classes from known classes, we simulate introducing unknown classes during training. At each training iteration, we randomly divide known classes into two groups, using one group for meta-training and the other for meta-testing, and then applying MLDG to Baseline. SMILE. SMILE proposes sign-magnitude disentanglement to better capture the class-level semantics. Following [14], we modified the projection head \\(\\phi\\) by replacing the final single MLP layer with two separate MLP layers for sign and magnitude predictions, and subsequently applying SMILE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Implementation Details. We use the same backbone network and optimization strategy for all baselines. Specifically, MinkUNet34 [12] is used as the feature extractor \\(\\mathcal{E}\\). For the projection head \\(\\phi\\), we use multi-layer perceptrons. We use SGD optimizer [46] with learning rate 0.01, momentum 0.9, and weight decay \\(1e - 4\\) with the batch size of 4. We use fixed mixing ratio \\(r = 0.1\\), \\(\\tau_{cos} = 0.07\\), \\(\\tau_{hash} = 0.14\\), and \\(L = 12\\). More details on baselines and implementation are provided in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.534, + 0.421, + 0.548 + ], + "angle": 0, + "content": "4.3 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.713 + ], + "angle": 0, + "content": "We adopt Greedy-Hungarian [15] and Strict-Hungarian [52] evaluation protocols by following [14]. During testing, the class descriptor from the hash code directly forms clusters. Then, we match these clusters with ground truth using the Hungarian matching algorithm [32]. Clusters that are not matched are regarded as misclassified. In Greedy-Hungarian, we divide the points into two groups, known and unknown classes, based on the ground truth and then perform Hungarian matching separately within each group. On the other hand, Strict-Hungarian conducts Hungarian matching for all clusters without distinguishing between known and unknown classes. The mean IoU (mIoU) is used for evaluating semantic segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.736, + 0.463, + 0.753 + ], + "angle": 0, + "content": "5 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.363, + 0.784 + ], + "angle": 0, + "content": "5.1 Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Quantitative Results. We conduct comparative experiments with baselines on all proposed scenarios. We report the results on scenario-A in Tab. 2, while the results on scenario-B and scenario-C are reported in Tab. 3. As shown in Tables. 2" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 2: On-the-fly category discovery results on SemanticPOSS (left) and SemanticKITTI (right) within Scenario-A. Best in bold." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.199, + 0.785, + 0.415 + ], + "angle": 0, + "content": "
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
SplitMethodUnknownKnownAllUnknownKnownAllSplitMethodUnknownKnownAllUnknownKnownAll
POSS0Baseline9.2028.5922.6311.4132.7126.16KITTI0Baseline10.4725.3421.4313.2230.2825.79
MLDG6.3334.6925.967.5540.5430.39MLDG12.6523.1720.4020.5927.3425.56
SMILE5.7940.3429.719.0556.5241.91SMILE9.9428.0923.3121.6638.9634.41
Ours17.8849.6039.8420.3056.5445.39Ours22.5545.9139.7732.9355.1249.28
POSS1Baseline13.2123.1620.8720.0130.1127.78KITTI1Baseline13.5123.7221.0420.4625.1323.90
MLDG15.7228.9725.9123.6635.6532.89MLDG8.9928.8723.6412.7129.7625.27
SMILE15.2032.4728.4818.2246.4039.89SMILE6.5936.6128.7122.6343.3237.88
Ours26.6844.1140.0939.2850.4747.89Ours22.2338.7034.3636.2642.6440.96
POSS2Baseline21.0319.8220.1028.5127.1527.46KITTI2Baseline10.0326.7522.3512.5428.4524.27
MLDG10.0430.1725.5219.8233.4030.27MLDG9.5030.1524.7213.5831.3326.66
SMILE9.0031.6426.4225.2738.2435.24SMILE11.1231.0925.8425.7539.7936.10
Ours24.5241.4037.5139.5946.7145.07Ours14.9838.3932.2323.3942.9037.77
POSS3Baseline4.4728.9223.2818.1228.6326.21KITTI3Baseline12.9326.1223.3421.6927.2726.10
MLDG2.2433.1226.0015.4333.9629.68MLDG7.8025.4421.7216.0027.5125.09
SMILE1.8936.6028.5924.0437.3834.30SMILE19.2130.9728.4932.7336.6035.79
Ours17.0951.3943.4849.1952.5451.77Ours17.7445.1539.3827.9447.8943.69
AvgBaseline11.9825.1221.7219.5129.6526.90AvgBaseline11.7425.4822.0416.9827.7825.02
MLDG8.5831.7425.8516.6235.8930.81MLDG9.7426.9122.6215.7228.9925.65
SMILE7.9735.2628.3019.1544.6437.84SMILE11.7231.6926.5925.6939.6736.05
Ours21.5446.6340.2337.0951.5747.53Ours19.3842.0436.4430.1347.1442.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "Table 3: On-the-fly category discovery results on SemanticKITTI within Scenario-B (left) and Scenario-C (right). Best in bold." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.508, + 0.784, + 0.582 + ], + "angle": 0, + "content": "
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAllMethodUnknownKnownAllUnknownKnownAll
Baseline14.6324.8923.8133.6425.9826.79Baseline11.7730.6127.6319.1431.0729.19
MLDG25.8025.6825.7040.7525.4927.10MLDG11.8033.7730.3029.4734.7033.87
SMILE13.1031.6129.6747.6732.5534.14SMILE11.0841.7436.9051.8643.8945.15
Ours21.7347.8345.0842.1248.9648.24Ours16.6647.2942.4537.1947.7646.09
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "and 3, in all scenarios, the baseline model exhibits diminished recognition ability for known classes due to its sensitive representation in hash coding, while the MLDG and SMILE demonstrate improved recognition capabilities compared to the baseline. Compared to all baselines, the proposed method achieves superior performance for known classes, demonstrating the effectiveness of the proposed representation learning for OCDSS, including DPL and MCL. Furthermore, the proposed method's comparative performance in the known classes compared to the oracle (All: \\(46.16\\%\\) in SemanticPOSS, All: \\(49.23\\%\\) in SemanticKITTI) indicates its superior recognition ability. MLDG and SMILE exhibit lower results for unknown classes in Strict-Hungarian except for MLDG in scenario-B. This implies that these models tend to become overly focused on the semantics of known classes, making them particularly confused when faced with unknown classes. MLDG performing well for unknown classes in scenario-B is speculated to be the simulation of unknown classes during training, which enhances gen" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.136, + 0.788, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.333, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison results on \\(\\mathrm{POSS}_3\\) within scenario-A (top), scenario-B (middle), and scenario-C (bottom). Best viewed when zoomed in with colors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.528 + ], + "angle": 0, + "content": "eralization to unobserved-unknown classes. However, this approach offers only limited performance improvements for known classes. In the case of SMILE, its favorable performance on unknown classes in the Greedy-Hungarian suggests that it understands the semantics of unknown classes well. Yet, the poor results in the Strict-Hungarian imply that SMILE frequently confuses known and unknown classes. Overall, the proposed method outperforms baselines in both known and unknown classes in scenario-A and scenario-C. In scenario-B as well, our proposed method consistently outperforms in all classes average, further validating the efficacy of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.789, + 0.651 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 3, we visualize the qualitative comparison results on scenario-A (top), scenario-B (middle), and scenario-C (bottom). While MLDG and SMILE fail to predict the known and unknown classes correctly, the proposed method not only successfully recognizes the known classes but also cognizes the unknown classes (top: plant/car, middle: terrain/other-vehicle, bottom: vegetation/bicyclist). This implies that the proposed method learns the better representations for OCDSS through the DPL and MCL. Additional qualitative results are provided in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.402, + 0.689 + ], + "angle": 0, + "content": "5.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed method, we conduct an ablation study as depicted in Fig. 4. Here, we set the baseline as the model introduced in Sec. 3.2. The experiments are conducted on the SemanticPOSS within scenario-A, and the results are averaged over four splits. Please refer to the Supplementary Material for the discussion with class-wise results and analysis on the hash code dimension \\( L \\), mixing ratio \\( r \\), weight \\( \\lambda \\), and sub-sampling in GM. DPL helps to recognize known classes. Comparing the baseline and baseline+DPL, we observe a significant improvement in performance for the known classes. When the model is trained only with the baseline, hash codes from the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.472, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.227, + 0.357, + 0.24 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.476, + 0.165, + 0.785, + 0.226 + ], + "angle": 0, + "content": "
Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAll
(a) Base11.9825.1221.7219.5129.6526.90
(b) Base+DPL5.3340.0831.3317.8946.4338.96
(c) Base+DPL+DM16.4848.2440.2031.3954.8348.58
(d) Base+DPL+DM+GM21.5446.6340.3637.0951.5747.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.619, + 0.227, + 0.64, + 0.239 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.25, + 0.784, + 0.278 + ], + "angle": 0, + "content": "Fig. 4: (a) Results of using dual prototypical learning (DPL) with the baseline and SMILE. (b) Ablation study of the proposed method. Base refers to the baseline model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.476 + ], + "angle": 0, + "content": "learned representations are sensitive to intra-class variation, which is supported by the poor performance for known classes. Conversely, adding hash prototypes through DPL imposes additional constraints on the representation, making it more suitable for hash code-based clustering. To further analyze the effects of DPL, we conduct an experiment by integrating the DPL with the baseline and SMILE [14]. As shown in Fig. 4a, when DPL is used together, the performance for known classes of both baseline and SMILE is significantly improved in both Strict- and Greedy-Hungarian. This implies that the proposed DPL effectively improves the robustness of learned representation under intra-class variation. The proposed DPL benefits both baseline and SMILE, demonstrating its effectiveness in a hash-code-based framework for addressing the OCDSS problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.627 + ], + "angle": 0, + "content": "MCL helps to cognize unknown classes. To excavate the ability to discover unknown classes, we proposed mixing-based category learning based on DPL. With DM, the prototypes become discriminative for ID and OOD. This enhancement leads to improved performance not only for known classes but also for unknown classes. The proposed GM further boosts the performance of unknown classes. Comparing the results of using (c) DM and (d) DM+GM in Fig. 4b, we observe approximately \\(5\\%\\) and \\(5.7\\%\\) gains in Strict- and Greedy-Hungarian, respectively. Although performance on known classes experiences a slight reduction, considering the inherent challenge of distinguishing unknown classes that the model has not seen during training, this underscores the necessity of GM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.36, + 0.67 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we introduce a new problem named on-the-fly category discovery for LiDAR semantic segmentation. Along with the baseline model based on hash coding, we propose dual prototypical learning to enhance the model's recognition ability. We further design the mixing-based category learning to inductively transfer recognition knowledge from known classes to unknown classes. Extensive experiments demonstrate the effectiveness of the proposed method for recognizing known classes and discovering unknown classes. While the proposed method effectively addresses the problem, performance disparities between the known and unknown classes remain. In this respect, exploring suitable representation learning for discovering unknown classes presents a promising research direction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.253 + ], + "angle": 0, + "content": "This work was supported by the Technology Innovation Program (1415187329, 20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.345 + ], + "angle": 0, + "content": "1. Alonso, I., Riazuelo, L., Montesano, L., Murillo, A.C.: 3d-mininet: Learning a 2d representation from point clouds for fast and efficient 3d lidar semantic segmentation. IEEE Robotics and Automation Letters 5(4), 5432-5439 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.4 + ], + "angle": 0, + "content": "2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.441 + ], + "angle": 0, + "content": "3. Basart, S., Mantas, M., Mohammadreza, M., Jacob, S., Dawn, S.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.496 + ], + "angle": 0, + "content": "4. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.525 + ], + "angle": 0, + "content": "5. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.551 + ], + "angle": 0, + "content": "6. Cao, K., Brbic, M., Leskovec, J.: Open-world semi-supervised learning. In: International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.785, + 0.592 + ], + "angle": 0, + "content": "7. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.648 + ], + "angle": 0, + "content": "8. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 318-334 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.689 + ], + "angle": 0, + "content": "9. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.689, + 0.785, + 0.744 + ], + "angle": 0, + "content": "10. Chen, K., Oldja, R., Smolyanskiy, N., Birchfield, S., Popov, A., Wehr, D., Eden, I., Pehserl, J.: Mvlidarnet: Real-time multi-class scene understanding for autonomous driving using multiple views. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2288-2294. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "11. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12547-12556 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.303, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: International Symposium on Visual Computing. pp. 207-222. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "14. Du, R., Chang, D., Liang, K., Hospedales, T., Song, Y.Z., Ma, Z.: On-the-fly category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11691-11700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "15. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.37 + ], + "angle": 0, + "content": "17. Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Navab, N., Busam, B., Tombari, F.: Segmenting known objects and unseen unknowns without prior knowledge. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 19321-19332 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "18. Ge, Z., Demyanov, S., Chen, Z., Garnavi, R.: Generative openmax for multi-class open set classification. In: British Machine Vision Conference. BMVA Press (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.399, + 0.785, + 0.453 + ], + "angle": 0, + "content": "19. Gerdzhev, M., Razani, R., Taghavi, E., Bingbing, L.: Tornado-net: multiview total variation semantic segmentation with diamond inception module. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9543-9549. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "20. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "21. Gu, P., Zhang, C., Xu, R., He, X.: Class-relation knowledge distillation for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16474-16483 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.578 + ], + "angle": 0, + "content": "22. Guo, Y., Camporese, G., Yang, W., Sperduti, A., Ballan, L.: Conditional variational capsule network for open set recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 103-111 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "23. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "24. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.758 + ], + "angle": 0, + "content": "26. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11108-11117 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "27. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "28. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "30. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "32. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "33. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "34. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "35. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9425-9434 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.787, + 0.447 + ], + "angle": 0, + "content": "36. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "37. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "38. Lu, T., Ding, X., Liu, H., Wu, G., Wang, L.: Link: Linear kernel for lidar-based 3d perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1105-1115 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "39. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.61 + ], + "angle": 0, + "content": "40. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 613-628 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.787, + 0.651 + ], + "angle": 0, + "content": "41. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.787, + 0.692 + ], + "angle": 0, + "content": "42. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.692, + 0.787, + 0.732 + ], + "angle": 0, + "content": "43. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "44. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "45. Riz, L., Saltori, C., Ricci, E., Poiesi, F.: Novel class discovery for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9393-9402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Ruder, S.: An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "H. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.785, + 0.219 + ], + "angle": 0, + "content": "48. Shu, Y., Shi, Y., Wang, Y., Huang, T., Tian, Y.: P-odn: Prototype-based open deep network for open set recognition. Scientific reports 10(1), 7146 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.261 + ], + "angle": 0, + "content": "49. Sun, X., Yang, Z., Zhang, C., Ling, K.V., Peng, G.: Conditional gaussian distribution learning for open set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13480-13489 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.303 + ], + "angle": 0, + "content": "50. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European conference on computer vision. pp. 685-702. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.347 + ], + "angle": 0, + "content": "51. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.348, + 0.785, + 0.389 + ], + "angle": 0, + "content": "52. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.785, + 0.432 + ], + "angle": 0, + "content": "53. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.785, + 0.53 + ], + "angle": 0, + "content": "55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 1887-1893. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.586 + ], + "angle": 0, + "content": "56. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 4376-4382. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.785, + 0.642 + ], + "angle": 0, + "content": "57. Xia, Y., Zhang, Y., Liu, F., Shen, W., Yuille, A.L.: Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 145-161. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.699 + ], + "angle": 0, + "content": "58. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.755 + ], + "angle": 0, + "content": "59. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "60. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "61. Zhao, Y., Zhong, Z., Sebe, N., Lee, G.H.: Novel class discovery in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4340-4349 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.293, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "62. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "63. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "64. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "65. Zhou, H., Zhu, X., Song, X., Ma, Y., Wang, Z., Li, H., Lin, D.: Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "66. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_origin.pdf b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1d8ffcc86cc6b7a52b029b9eaedf569435fde191 --- /dev/null +++ b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/c9a23254-3e06-4a70-aa78-c2eed722dc95_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19aa5d408bda601cf6c13d01decc5512d9327a479085f68aee4cef3f8200ec26 +size 7526843 diff --git a/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/full.md b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b90f6a7dddd23f0c785150c43f7620c6aff3fb0d --- /dev/null +++ b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/full.md @@ -0,0 +1,319 @@ +# On-the-fly Category Discovery for LiDAR Semantic Segmentation + +Hyeonseong Kim $^{\text{①}}$ , Sung-Hoon Yoon $^{\text{②}}$ , Minseok Kim $^{\text{③}}$ , and Kuk-Jin Yoon $^{\text{④}}$ + +Visual Intelligence Lab., KAIST, Korea {brian617,yoon307,alstjrx1x1,kjyoon}@kaist.ac.kr + +Abstract. LiDAR semantic segmentation is important for understanding the surrounding environment in autonomous driving. Existing methods assume closed-set situations with the same training and testing label space. However, in the real world, unknown classes not encountered during training may appear during testing, making it difficult to apply existing methodologies. In this paper, we propose a novel on-the-fly category discovery method for LiDAR semantic segmentation, aiming to classify and segment both unknown and known classes instantaneously during test time, achieved solely by learning with known classes in training. To embed instant segmentation capability in an inductive setting, we adopt a hash coding-based model with an expandable prediction space as a baseline. Based on this, dual prototypical learning is proposed to enhance the recognition of the known classes by reducing the sensitivity to intra-class variance. Additionally, we propose a novel mixing-based category learning framework based on representation mixing to improve the discovery capability of unknown classes. The proposed mixing-based framework effectively models out-of-distribution representations and learns to semantically group them during training while distinguishing them from in-distribution representations. Extensive experiments on SemanticKITTI and SemanticPOSS datasets demonstrate the superiority of the proposed method compared to the baselines. The code is available at https://github.com/hskim617/OCDSS. + +Keywords: Autonomous driving $\cdot$ LiDAR semantic segmentation $\cdot$ On-the-fly category discovery + +# 1 Introduction + +LiDAR semantic segmentation is crucial for recognizing the surrounding environment in autonomous driving and has recently made significant progress [2, 11, 12, 26, 30, 38, 39, 50, 59, 65]. Although these models demonstrate promising results across various datasets, their effectiveness is constrained to closed-set settings, as they are assumed and trained within such scenarios. In real-world scenarios, however, unknown classes are bound to appear during testing, making it incapable of predicting unknown classes for closed-set models as in Fig. 1a. + +![](images/09ed99855aaa56f7a326ba3ca593a407832ca8277c28237e84051c843a8f6af9.jpg) +Fig. 1: Illustration of training and testing scenarios for on-the-fly category discovery in LiDAR semantic segmentation and relevant tasks. The goal of each task is (a) recognizing known classes, (b) further identifying and rejecting unknown classes, (c) learning clustering using unlabeled, unknown class data, and (d) cognizing unknown classes that are not identified during training while recognizing known classes. + +To address this challenge, open-set recognition (OSR) [47], which aims to filter out unknown classes has been explored in diverse fields, including image classification [3,5,9,49,58], semantic segmentation [7,31,37,57], and 3D semantic segmentation [8,35]. Considering the nature of the human cognitive system that can cognize novel visual concepts based on understanding, simply rejecting unknown classes as in OSR (Fig. 1b) is unsatisfactory. + +To narrow the gap between humans and machines by learning novel concepts, novel category discovery (NCD) [15, 23, 24, 28, 63] and on-the-fly category discovery (OCD) [14] are emerging as new research fields and attracting attention. Given labeled data of known classes and unlabeled data of unknown classes, the goal of NCD is to cluster the unlabeled data with the same semantics, aiming to cognize them as in Fig. 1c. Therefore, NCD is formulated as transductive learning in which data of unknown classes are available, making it incapable of predicting unknown classes not identified during training. On the other hand, OCD aims for a more realistic scenario by learning solely from data of known classes and inductively cognizing unknown classes at test time. Furthermore, during testing, the model should independently classify each test sample instantaneously, regardless of the distribution of the test data. + +In this paper, we broaden the scope of OCD to LiDAR point clouds and introduce a new problem, on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). As shown in Fig. 1d, our goal is to classify and segment both known and unknown classes instantaneously in an inductive manner during test time, achieved solely by learning with known classes in training. To achieve the goal, one straightforward solution is adopting the previous approach in image classification [14] and applying it to the LiDAR semantic segmentation. + +Therefore, we implemented the hash coding-based model proposed by [14] as a baseline for LiDAR point clouds. This model utilizes hash codes as class descriptors, generated by binarizing each dimension of the learned representation based on its sign. However, extending it to the LiDAR domain entails two limitations. + +Firstly, due to LiDAR's sparse nature and lack of texture information, hash coding in learned representations is more sensitive to intra-class variance. In practice, we observed that the method for reducing sensitivity to intra-class variance by disentangling sign and magnitude [14] had shown unsatisfactory results in the LiDAR domain (refer Fig. 4a). Secondly, prior work relies on representation learning from known classes to discover unknown classes. Without a dedicated mechanism for discovering unknown classes, only a tentative expectation that the learned representations will form semantically meaningful groups for unknown classes is possible. This reliance, combined with the aforementioned limitation, complicates OCDSS. + +Therefore, we propose a novel learning method for OCDSS that learns representations robust to intra-class variance in hash coding and suitable for unknown class discovery. First, we propose dual prototypical learning (DPL) that constrain the representation space by introducing extra hash prototypes, effectively reducing the sensitivity of hash codes to intra-class variance. Specifically, the hash prototypes enforce additional constraints to ensure that representations within the same intra-class share the same hash code, dramatically enhancing recognition capability for known classes. Second, we propose a novel mixing-based category learning (MCL) framework to improve the ability to cognize unknown classes. The proposed framework consists of three modules: a mixing module, a discrimination module, and a grouping module. The mixing module models out-of-distribution (OOD) representation through representation mixing and obtains mixed representation to mimic unknown classes during training. The discrimination module then enhances the model's OOD discrimination ability by learning to distinguish the representation of known classes from mixed representations. In addition, the grouping module develops the model's clustering capability for OOD representations by learning to cluster mixed representations within the same semantics. Meanwhile, we integrate the hash prototypes of DPL during MCL to fully exploit the discovering capability for unknown classes. Finally, we introduce an evaluation protocol for OCDSS and evaluate our approach on SemanticKITTI [4] and SemanticPOSS [42] datasets. Extensive experiments demonstrate the superiority of our method compared to baselines, while ablation studies show the effectiveness of each component. + +Our main contributions are four-fold: (I) We introduce a new problem of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). (II) We propose dual prototypical learning that effectively learns representations that are less sensitive to intra-class variation in hash coding. (III) We present a novel mixing-based category learning framework, enabling the model to discover unknown classes. (IV) We introduce a new evaluation protocol for OCDSS and demonstrate the superiority of our method through extensive experiments. + +# 2 Related Work + +LiDAR Semantic Segmentation. LiDAR semantic segmentation is the task of inferring semantic labels to individual 3D points within the LiDAR point cloud. These approaches can be broadly categorized into point-based, projection-based, and voxel-based according to the way they represent and process point clouds. Point-based methods [26,36,44,51] directly operate on points in line with the foundational work PointNet [43] by using multi-layer perceptrons. Despite their advantages, such as minimal information loss and few parameters, applying these methods to large-scale LiDAR point clouds demands heavy computational resources and memory. Projection-based methods have been proposed to address efficiency and leverage advanced 2D convolutional neural network architectures. These methods project LiDAR points onto a range image [1,2,10,13,19,30,39,55, 56] using spherical projection or a bird-eye-view [59]. However, the projection of 3D points into 2D representation might lose some geometric information. Voxel-based methods represent point clouds as 3D voxels and process them through 3D convolution [66] or sparse convolution [11, 12, 20, 38, 50, 65]. In particular, sparse convolution achieves high performance with less computation, so we adopt MinkowskiNet [12] as our backbone architecture. + +Open-set Recognition. Open-set recognition (OSR) was first introduced in [47], extending the closed-set assumption to realistic scenarios where unknown classes appear during testing. As OSR aims to detect unknown classes, previous studies estimate the unknown probability based on extreme value theory [5] or estimate uncertainty through maximum softmax probability [25], maximum logit [3], or Bayesian inference approximation [16, 33]. Other approaches include training a redundancy classifier for unknown classes [54, 64] or using metric learning [9, 22, 48]. Meanwhile, generative model-based approaches employ generative adversarial networks to synthesize unknown class samples [18, 40] or use image reconstruction error for the criterion of determining unknown classes [41, 49, 58]. Unlike these methods for object recognition, studies for semantic segmentation [7, 31, 37, 57], panoptic segmentation [17, 27] in open-set situations have also been proposed. Recently, open-set semantic segmentation for LiDAR point clouds has emerged [8, 35], extending an open-set problem to the LiDAR domain. While OSR methods are aware of unknown classes, they focus on identification, whereas we aim to go beyond achieving semantically cognizing them. + +Novel Category Discovery. Novel category discovery (NCD), first formalized in [24], is another relevant task that aims to cognize unknown classes in unlabeled data by transferring the knowledge from known classes in labeled data. After DTC [24] proposed deep transfer clustering for NCD, several studies were suggested, achieving significant progress [15, 21, 23, 28, 60, 62, 63]. These approaches first learn representations using supervised learning from labeled known classes, and then discover unknown classes by transferring knowledge through pseudolabeling on unknown classes based on pair-wise similarity [23, 28, 60, 62, 63] or cluster assignment [15, 21]. Additionally, EUMS [61] and NOPS [45] extend the NCD problem to semantic segmentation in images and LiDAR point clouds, respectively. Nevertheless, NCD usually assumes the number of unknown classes + +as prior knowledge [15, 23, 28, 60, 62, 63], and presuppose all unlabeled data as unknown classes. Hence, generalized category discovery (GCD) was introduced where unlabeled data contain both known and unknown classes [6, 52]. However, as pointed out in [14], two limitations in prior settings still exist: 1) transductive learning and 2) offline evaluation. In prior settings, category discovery is formulated as transductive learning in which data of unknown classes are available during training. Also, offline evaluation constrains the results depending on the entire test data distribution, making instant feedback on each test sample impossible. Therefore, [14] proposed a new task named on-the-fly category discovery (OCD), which aims to recognize known classes and cognize unknown classes instantaneously in an inductive manner. In this study, we tackle the challenging OCD for LiDAR semantic segmentation, which instantaneously segments LiDAR point clouds from both known and unknown classes while learning exclusively from known classes. + +# 3 Method + +In Sec. 3.1, we first elaborate on our problem formulation of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). In Sec. 3.2, we introduce a hash coding-based baseline model for OCDSS that embeds expandable prediction space and instant inference capability within an inductive setting. Then, in Sec. 3.3, we describe our dual prototypical learning for reducing the sensitivity of representation in hash coding, improving the closed-set performance. In Sec. 3.4, we propose a mixing-based category learning framework to enable the model to possess category discovery capabilities for both known and unknown classes. Fig. 2 shows the overview of the proposed method. + +# 3.1 Problem Formulation + +In this section, we formulate the problem of OCDSS, which aims to recognize known classes and cognize unknown classes during testing in an inductive manner. Let $D_{S} = \{P_{S}\}$ be a support set used for training and $D_{Q} = \{P_{Q}\}$ be a query set for testing. Then, we have 3D point clouds $P_{S} = \{(x,y)\} \in \mathbb{R}^{3} \times Y_{S}$ and $P_{Q} = \{(x,y)\} \in \mathbb{R}^{3} \times Y_{Q}$ , representing a single LiDAR scan from $D_{S}$ and $D_{Q}$ , respectively. Here, $x$ denotes each point, and $y$ is the corresponding label. $Y_{S}$ and $Y_{Q}$ are the label spaces of support and query sets. Following the setting of OCD [14], $Y_{S}$ defines the known classes observed during training, while $Y_{Q}$ contains both known and unknown classes that appear at testing, i.e. $Y_{S} \subset Y_{Q}$ . We denote the number of known and unknown classes as $C_k = |Y_S|$ and $C_u = |Y_Q\backslash Y_S|$ , where we assume $C_u$ is unknown. This is different from existing NCD approaches [15,23,28,45,60,62,63] that assume $C_u$ as a known prior. The goal is to train a model using only $D_{S}$ so that the model can inductively classify and segment point clouds in $D_{Q}$ instantaneously. + +(b) Discrimination Module +![](images/24583f59c1bcc33370058ce4388d61fd51813753391fbe005e1a441351288a41.jpg) +: Mixed Representation O:Origin Representation + +![](images/b2f335596e5dffca1f7ac12673f6d604ae2927f6cb773217e7713939c79e39a0.jpg) +(a) Mixing Module +Fig. 2: Overview of the proposed framework. Given the LiDAR point cloud, the model extracts the features through feature extractor $\mathcal{E}$ , which are then mapped into the prediction space to obtain the representations by the projection head $\phi$ . Through dual prototypical learning (DPL), the model learns representations suitable for hash code-based clustering, enhancing its recognition ability for known classes. Besides, mixing-based category learning (MCL) improves the model's cognition capability for unknown classes through three modules: mixing module, discrimination module, and grouping module. The final prediction is obtained by the hash code-based clustering. + +![](images/7ae4aa88713eaacb26edc580f3dafde6db11540748ec9a581e8bb513302143fb.jpg) + +![](images/683bedbca81f632b7171987c71a7ae038df792982cfa149f3fdcd790c669dd24.jpg) +(c) Grouping Module + +![](images/728bfcb94efba52df3e479f8b1f9a42829fd0e8fb6ee39fa3da1d042142278b1.jpg) + +# 3.2 Hash Coding-based Baseline Model + +We first introduce a hash coding-based baseline model for OCDSS. To embed the expandable prediction space and instant inference capability while learning the model inductively, we adopt the basic pipeline of hash coding-based model from [14]. Formally, our baseline model consists of a feature extractor $\mathcal{E}$ and a projection head $\phi$ . Given point cloud $P_{S} = \{(x,y)\}_{i=1}^{N}$ , the feature extractor first extracts the feature representation $f_{i}$ for each point $x_{i}$ . The projection head maps $f_{i}$ to the prediction space, obtaining the projected representation $p_{i} = \phi(f_{i}) \in \mathbb{R}^{L}$ , which is used for the hash coding. Here, $L$ is the dimension of the hash code. Then, the hash code for point $x_{i}$ is obtained by $\text{hash}(\phi(\mathcal{E}(x_{i})))$ , where $\text{hash}$ is a hashing operation. For an arbitrary vector $\mathbf{v} = [v_{1}, \ldots, v_{l}, \ldots, v_{L}] \in \mathbb{R}^{L}$ , the hashing operation is defined as, + +$$ +h a s h (\mathbf {v}) = \left[ v _ {1} ^ {*}, \dots , v _ {l} ^ {*}, \dots , v _ {L} ^ {*} \right], v _ {l} ^ {*} = \left\{ \begin{array}{l l} 1, & v _ {l} \geq 0 \\ 0, & v _ {l} < 0 \end{array} . \right. \tag {1} +$$ + +To predict classes of points from $P_{Q}$ during testing, we use the hash code as a class descriptor where points with the same code are considered to be in the same class. This hash coding-based inference allows the model to have an expanded label space to predict new classes beyond the known classes in an inductive setting. Moreover, it also has the advantage of being able to immediately predict the result of each test sample individually. + +For representation learning of the baseline model, [14] used supervised contrastive learning [29]; however, there is a lack of clear design choices for representation learning in LiDAR point clouds. For example, in the case of contrastive learning where pair-wise computation between samples is required, it is not suitable due to its high demand for computation and memory. Instead, we use a standard cross-entropy loss on the known classes. To learn a more discriminative representation, we employ parametric prototypes $\mathcal{G} = \{g_1, g_2, \dots, g_{C_k}\} \in \mathbb{R}^{C_k \times L}$ for known classes as in [7] and replace the classifier as cosine similarity between the prototype and the projected representation $p_i$ . Specifically, we compute the distance based on the cosine similarity as, + +$$ +d _ {c o s} \left(p _ {i}, g _ {j}\right) = 1 - \frac {p _ {i} \cdot g _ {j}}{\| p _ {i} \| \| g _ {j} \|}. \tag {2} +$$ + +Then, let us define $\sigma_{\mathcal{G}}(p_i,y_i)$ that computes the softmax probability of $p_i$ with respect to $\mathcal{G}$ as follows: + +$$ +\sigma_ {\mathcal {G}} \left(p _ {i}, y _ {i}\right) = \frac {\exp \left(- d _ {\text {c o s}} \left(p _ {i} , g _ {y _ {i}}\right) / \tau_ {\text {c o s}}\right)}{\sum_ {j = 1} ^ {C _ {k}} \exp \left(- d _ {\text {c o s}} \left(p _ {i} , g _ {j}\right) / \tau_ {\text {c o s}}\right)}, \tag {3} +$$ + +where $g_{y_i}$ denotes the prototype corresponding to the label $y_i$ and $\tau_{cos}$ is a temperature parameter. The loss function for the baseline model is defined as, + +$$ +\mathcal {L} _ {\text {b a s e}} = \sum_ {i} - \log \sigma_ {\mathcal {G}} \left(p _ {i}, y _ {i}\right). \tag {4} +$$ + +# 3.3 Dual Prototypical Learning + +The baseline model is trained to generate similar representations for points belonging to the same class. However, during inference, points are classified into different classes if they do not share the same hash code, which can be sensitive to intra-class variance. Due to the sparse nature and lack of texture information in LiDAR, hash coding in learned representations becomes more sensitive to intra-class variance compared to images. To address this issue, we propose dual prototypical learning (DPL) to learn the representation robust to intra-class variance in hash coding. Specifically, DPL employs extra parametric prototypes to constrain the representation space, denoted as $\mathcal{H} = \{h_1, h_2, \ldots, h_{C_k}\} \in \mathbb{R}^{C_k \times L}$ , referred to as hash prototypes. Similar to $\mathcal{G}$ , we calculate the distance between the projected representation $p_i$ and the hash prototypes. In this case, based on the fact that the sign of each dimension in representation determines the hash code, we utilize the L2 distance between signs instead of cosine similarity. In practice, for stable optimization, we use the hyperbolic tangent function, tanh, instead of the sign operation as, + +$$ +d _ {h a s h} \left(p _ {i}, h _ {j}\right) = \left\| t a n h \left(p _ {i}\right) - t a n h \left(h _ {j}\right) \right\| _ {2} ^ {2} / L. \tag {5} +$$ + +Similar to $\sigma_{\mathcal{G}}$ in Eq. (3), we define softmax probability $\sigma_{\mathcal{H}}(p_i, y_i)$ as follows: + +$$ +\sigma_ {\mathcal {H}} \left(p _ {i}, y _ {i}\right) = \frac {\exp \left(- d _ {\text {h a s h}} \left(p _ {i} , h _ {y _ {i}}\right) / \tau_ {\text {h a s h}}\right)}{\sum_ {j = 1} ^ {C _ {k}} \exp \left(- d _ {\text {h a s h}} \left(p _ {i} , h _ {j}\right) / \tau_ {\text {h a s h}}\right)}, \tag {6} +$$ + +where $h_{y_i}$ denotes the hash prototype corresponding to the label $y_i$ and $\tau_{hash}$ is a temperature parameter. The prototypical learning loss based on hash prototypes is defined as, + +$$ +\mathcal {L} _ {\text {h a s h}} = \sum_ {i} - \log \sigma_ {\mathcal {H}} \left(p _ {i}, y _ {i}\right). \tag {7} +$$ + +Along with $\mathcal{L}_{base}$ , the final loss function of dual prototypical learning is as follows: + +$$ +\mathcal {L} _ {D P L} = \mathcal {L} _ {\text {b a s e}} + \mathcal {L} _ {\text {h a s h}}. \tag {8} +$$ + +With the proposed DPL, we impose additional constraints on representations to make hash codes less sensitive to intra-class variance, improving the model's recognition capability. + +# 3.4 Mixing-based Category Learning + +Along with the recognition, our goal is to train the model to cognize unknown classes through semantic understanding. Our basic idea to achieve this goal is to mimic unknown classes during training and learn the model to distinguish and group them from known classes for a comprehensive understanding of unknown semantics. To this end, we propose mixing-based category learning (MCL), which consists of a mixing module, a discrimination module, and a grouping module. + +Mixing Module. We mix the representations of different known classes to mimic the unknown classes during training. Our intention in using representation mixing is that 1) it can effectively model OOD representations, as stated in [64], and 2) it can further mimic diverse unknown classes by modeling OOD representations that share common semantics depending on which classes are mixed. Specifically, we use manifold mixup technique [53] to mix two feature representations $f_{i}$ and $f_{j}$ of different classes $y_{i}$ and $y_{j}$ , where $y_{i} \neq y_{j}$ . The mixed representation $\overline{f_{i,j}}$ is calculated as, + +$$ +\overline {{f _ {i , j}}} = r f _ {i} + (1 - r) f _ {j}, \tag {9} +$$ + +where $r$ is a mixing ratio. We can also define the imaginary corresponding label $\overline{y_{i,j}}$ of $\overline{f_{i,j}}$ , referred to as a mix label. Note that if $(i,j) \neq (k,l)$ , then $\overline{y_{i,j}} \neq \overline{y_{k,l}}$ . These mixed representations are considered as OOD representations in the following modules. + +Discrimination Module. In the discrimination module (DM), we train the model to distinguish mixed representations from prototypes. In particular, we learn parametric prototypes $\mathcal{G}$ and $\mathcal{H}$ to be discriminative from the mixed representation. To this end, we first project $\overline{f_{i,j}}$ into projection space and obtain $\phi(\overline{f_{i,j}})$ . Then, we enforce $\phi(\overline{f_{i,j}})$ not to be close to the prototype of any specific class in the form of entropy maximization as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i s c}} = \sum_ {i, j} (\sum_ {k} \sigma_ {\mathcal {G}} \left(\phi \left(\overline {{f _ {i , j}}}\right), k\right) \log \sigma_ {\mathcal {G}} \left(\phi \left(\overline {{f _ {i , j}}}\right), k\right) \tag {10} \\ + \sum_ {k} \sigma_ {\mathcal {H}} \left(\phi \left(\overline {{f _ {i , j}}}\right), k\right) \log \sigma_ {\mathcal {H}} \left(\phi \left(\overline {{f _ {i , j}}}\right), k\right)). \\ \end{array} +$$ + +Notably, we stop the gradient of $\phi(f_{i,j})$ and only optimize prototypes. In addition, we incorporate mixed representations that are mixed after being projected into the $\mathcal{L}_{disc}$ to enhance discrimination for more diverse OOD representations. + +Grouping Module. We propose a grouping module (GM) to enhance the model's understanding of the underlying semantics. Since the mixed representations might share common semantics depending on which classes are mixed, we can group them using the mix labels. For mathematical brevity, we omit the summation over $i$ and $j$ here. Specifically, we employ contrastive loss [29] and propose contrastive learning on mixed representation based on mixed labels: + +$$ +\mathcal {L} _ {g p} = \frac {- 1}{| P (i , j) |} \sum_ {p \in P (i, j)} \log \sigma_ {A (i, j)} \left(\frac {- d _ {c o s} \left(\phi \left(\overline {{f _ {i , j}}}\right) , \phi \left(\overline {{f _ {p}}}\right)\right)}{\tau_ {c o s}}\right), \tag {11} +$$ + +where $A(i,j)$ denotes all mixed representations excluding $\phi(\overline{f_{i,j}})$ and $P(i,j)$ is all positive pairs of $\phi(\overline{f_{i,j}})$ that share the same mix label. $\sigma_{(A(i,j))}$ refers to a softmax operation with respect to $A(i,j)$ . In practice, we perform sub-sampling on $A(i,j)$ to reduce the computation from heavy pair-wise operations. Additionally, we impose the following constraint so that mixed representations with the same mix label share the hash code as, + +$$ +\mathcal {L} _ {g p - h a s h} = \log \sum_ {p \in P (i, j)} \exp \left(d _ {h a s h} \left(\phi \left(\overline {{f _ {i , j}}}\right), \phi \left(\overline {{f _ {p}}}\right)\right) / \tau_ {h a s h}\right). \tag {12} +$$ + +The loss function for the proposed MCL is defined as, + +$$ +\mathcal {L} _ {M C L} = \mathcal {L} _ {\text {d i s c}} + \lambda \left(\mathcal {L} _ {g p} + \mathcal {L} _ {g p - h a s h}\right), \tag {13} +$$ + +where, $\lambda$ is the weight for $\mathcal{L}_{gp}$ and $\mathcal{L}_{gp - hash}$ . + +Finally, the total loss function of our method is, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {D P L} + \mathcal {L} _ {M C L}. \tag {14} +$$ + +# 4 Evaluation Protocol + +# 4.1 Data Split + +Datasets. We evaluate our method on two datasets, SemanticKITTI [4] and SemanticPOSS [42]. SemanticKITTI consists of 43,552 frames with point-wise annotations of 19 classes. We follow the official split and use sequences 00 to 07 and 09 to 10 with 19,130 frames for training, and sequence 08 with 4,071 frames for validation. SemanticPOSS contains 2,988 frames with point-wise annotations of 13 classes. We use sequence 03 for validation, while the remaining sequences are used for training, following the official split. + +Scenarios. We designed three scenarios considering various real-world situations. More precisely, we defined two unknown class types that can exist in the real world: observed-unknown and unobserved-unknown. Observed-unknown + +Table 1: Data splits of scenario-A on SemanticPOSS and SemanticKITTI datasets. The $\mathrm{POSS}_i$ and the $\mathrm{KITTI}_i$ denotes the $i^{th}$ split. + +
SplitUnknown classesSplitUnknown classes
POSS0trashcan, traffic sign, bike, plantsKITTI0bicycle, other-vehicle, trunk, terrain, vegetation
POSS1pole, person, buildingKITTI1bicyclist, traffic-sign, other-ground, fence, road
POSS2rider, fence, groundKITTI2motorcyclist, motorcycle, pole, car, sidewalk
POSS3cone/stone, trunk, carKITTI3person, truck, parking, building
+ +class refers to cases where data was captured during acquisition but remains unlabeled. An example could be cases where instances are not included in the pre-defined category set or when it is challenging to define a class at the time of annotation, resulting in their assignment as a void or an ignore label. On the other hand, unobserved-unknown class corresponds to cases where it first appears during testing without being captured during data acquisition. We then constructed three scenarios based on the presence of each unknown class type. + +Scenario-A: In scenario-A, we assess the discovery capability for observed-unknown classes. Following the previous open-set 3D semantic segmentation [8, 35] literature that addresses the observed-unknown classes in testing, we set the labels of unknown classes to be void and ignore them during training. Depending on the choice of the unknown class, we made four splits each for SemanticKITTI and SemanticPOSS, as shown in Tab. 1. The splits are selected considering the diversity of unknown classes and the relationship between classes. We provide details on the selection process in the Supplementary Material. + +Scenario-B: In scenario-B, we evaluate the capability to discover unobserved-unknown classes that do not appear in the training data. Specifically, we select relatively infrequent classes from SemanticKITTI, {other-vehicle, bicyclist}, as unknown classes to prevent a significant decrease in the amount of training data. Frames containing any points of these classes are then excluded from training data. As a result, 13,501 out of 19,130 frames are used for training. We did not use SemanticPOSS, as all classes appear in most frames. The evaluation is conducted for all classes, including unobserved-unknown classes, using the same validation data as the official benchmark. + +Scenario-C: Both observed-unknown and unobserved-unknown classes are considered in scenario-C. In SemanticKITTI, we select {other-vehicle, parking} as observed-unknown classes, and {bicyclist} as an unobserved-unknown class. Frames containing any points of the unobserved-unknown class are excluded from training data, resulting in 17,954 frames for training, and observed-unknown classes are ignored during training. The evaluation proceeds as in scenario-B. + +# 4.2Baselines + +Since OCDSS necessitates inductive learning with known classes while enabling instant inference during testing, many previous approaches are impractical un + +der our setting. Clustering-based methods do not provide instant inference and demand a substantial amount of memory due to the large number of points in the LiDAR point cloud, making them more impractical. Previous NCD method [45] is also unable to learn under our setting where unknown classes are unavailable. Therefore, following the previous OCD approach [14], we implement MLDG [34] and SMILE [14], representative OCD methods in image classification, in the 3D LiDAR points domain as baselines. + +Baseline. The baseline is the hash coding-based model introduced in Sec. 3.2. MLDG. MLDG is a meta-learning method for domain generalization. Since OCDSS is more like generalizing to unknown classes from known classes, we simulate introducing unknown classes during training. At each training iteration, we randomly divide known classes into two groups, using one group for meta-training and the other for meta-testing, and then applying MLDG to Baseline. SMILE. SMILE proposes sign-magnitude disentanglement to better capture the class-level semantics. Following [14], we modified the projection head $\phi$ by replacing the final single MLP layer with two separate MLP layers for sign and magnitude predictions, and subsequently applying SMILE. + +Implementation Details. We use the same backbone network and optimization strategy for all baselines. Specifically, MinkUNet34 [12] is used as the feature extractor $\mathcal{E}$ . For the projection head $\phi$ , we use multi-layer perceptrons. We use SGD optimizer [46] with learning rate 0.01, momentum 0.9, and weight decay $1e - 4$ with the batch size of 4. We use fixed mixing ratio $r = 0.1$ , $\tau_{cos} = 0.07$ , $\tau_{hash} = 0.14$ , and $L = 12$ . More details on baselines and implementation are provided in Supplementary Material. + +# 4.3 Evaluation Metrics + +We adopt Greedy-Hungarian [15] and Strict-Hungarian [52] evaluation protocols by following [14]. During testing, the class descriptor from the hash code directly forms clusters. Then, we match these clusters with ground truth using the Hungarian matching algorithm [32]. Clusters that are not matched are regarded as misclassified. In Greedy-Hungarian, we divide the points into two groups, known and unknown classes, based on the ground truth and then perform Hungarian matching separately within each group. On the other hand, Strict-Hungarian conducts Hungarian matching for all clusters without distinguishing between known and unknown classes. The mean IoU (mIoU) is used for evaluating semantic segmentation performance. + +# 5 Experimental Results + +# 5.1 Comparison + +Quantitative Results. We conduct comparative experiments with baselines on all proposed scenarios. We report the results on scenario-A in Tab. 2, while the results on scenario-B and scenario-C are reported in Tab. 3. As shown in Tables. 2 + +Table 2: On-the-fly category discovery results on SemanticPOSS (left) and SemanticKITTI (right) within Scenario-A. Best in bold. + +
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
SplitMethodUnknownKnownAllUnknownKnownAllSplitMethodUnknownKnownAllUnknownKnownAll
POSS0Baseline9.2028.5922.6311.4132.7126.16KITTI0Baseline10.4725.3421.4313.2230.2825.79
MLDG6.3334.6925.967.5540.5430.39MLDG12.6523.1720.4020.5927.3425.56
SMILE5.7940.3429.719.0556.5241.91SMILE9.9428.0923.3121.6638.9634.41
Ours17.8849.6039.8420.3056.5445.39Ours22.5545.9139.7732.9355.1249.28
POSS1Baseline13.2123.1620.8720.0130.1127.78KITTI1Baseline13.5123.7221.0420.4625.1323.90
MLDG15.7228.9725.9123.6635.6532.89MLDG8.9928.8723.6412.7129.7625.27
SMILE15.2032.4728.4818.2246.4039.89SMILE6.5936.6128.7122.6343.3237.88
Ours26.6844.1140.0939.2850.4747.89Ours22.2338.7034.3636.2642.6440.96
POSS2Baseline21.0319.8220.1028.5127.1527.46KITTI2Baseline10.0326.7522.3512.5428.4524.27
MLDG10.0430.1725.5219.8233.4030.27MLDG9.5030.1524.7213.5831.3326.66
SMILE9.0031.6426.4225.2738.2435.24SMILE11.1231.0925.8425.7539.7936.10
Ours24.5241.4037.5139.5946.7145.07Ours14.9838.3932.2323.3942.9037.77
POSS3Baseline4.4728.9223.2818.1228.6326.21KITTI3Baseline12.9326.1223.3421.6927.2726.10
MLDG2.2433.1226.0015.4333.9629.68MLDG7.8025.4421.7216.0027.5125.09
SMILE1.8936.6028.5924.0437.3834.30SMILE19.2130.9728.4932.7336.6035.79
Ours17.0951.3943.4849.1952.5451.77Ours17.7445.1539.3827.9447.8943.69
AvgBaseline11.9825.1221.7219.5129.6526.90AvgBaseline11.7425.4822.0416.9827.7825.02
MLDG8.5831.7425.8516.6235.8930.81MLDG9.7426.9122.6215.7228.9925.65
SMILE7.9735.2628.3019.1544.6437.84SMILE11.7231.6926.5925.6939.6736.05
Ours21.5446.6340.2337.0951.5747.53Ours19.3842.0436.4430.1347.1442.93
+ +Table 3: On-the-fly category discovery results on SemanticKITTI within Scenario-B (left) and Scenario-C (right). Best in bold. + +
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAllMethodUnknownKnownAllUnknownKnownAll
Baseline14.6324.8923.8133.6425.9826.79Baseline11.7730.6127.6319.1431.0729.19
MLDG25.8025.6825.7040.7525.4927.10MLDG11.8033.7730.3029.4734.7033.87
SMILE13.1031.6129.6747.6732.5534.14SMILE11.0841.7436.9051.8643.8945.15
Ours21.7347.8345.0842.1248.9648.24Ours16.6647.2942.4537.1947.7646.09
+ +and 3, in all scenarios, the baseline model exhibits diminished recognition ability for known classes due to its sensitive representation in hash coding, while the MLDG and SMILE demonstrate improved recognition capabilities compared to the baseline. Compared to all baselines, the proposed method achieves superior performance for known classes, demonstrating the effectiveness of the proposed representation learning for OCDSS, including DPL and MCL. Furthermore, the proposed method's comparative performance in the known classes compared to the oracle (All: $46.16\%$ in SemanticPOSS, All: $49.23\%$ in SemanticKITTI) indicates its superior recognition ability. MLDG and SMILE exhibit lower results for unknown classes in Strict-Hungarian except for MLDG in scenario-B. This implies that these models tend to become overly focused on the semantics of known classes, making them particularly confused when faced with unknown classes. MLDG performing well for unknown classes in scenario-B is speculated to be the simulation of unknown classes during training, which enhances gen + +![](images/64fc7562fadf2f91ab337567778f5c5ed39e692822014c2410a5ac56e9c3866a.jpg) +Fig. 3: Qualitative comparison results on $\mathrm{POSS}_3$ within scenario-A (top), scenario-B (middle), and scenario-C (bottom). Best viewed when zoomed in with colors. + +eralization to unobserved-unknown classes. However, this approach offers only limited performance improvements for known classes. In the case of SMILE, its favorable performance on unknown classes in the Greedy-Hungarian suggests that it understands the semantics of unknown classes well. Yet, the poor results in the Strict-Hungarian imply that SMILE frequently confuses known and unknown classes. Overall, the proposed method outperforms baselines in both known and unknown classes in scenario-A and scenario-C. In scenario-B as well, our proposed method consistently outperforms in all classes average, further validating the efficacy of our approach. + +Qualitative Results. In Fig. 3, we visualize the qualitative comparison results on scenario-A (top), scenario-B (middle), and scenario-C (bottom). While MLDG and SMILE fail to predict the known and unknown classes correctly, the proposed method not only successfully recognizes the known classes but also cognizes the unknown classes (top: plant/car, middle: terrain/other-vehicle, bottom: vegetation/bicyclist). This implies that the proposed method learns the better representations for OCDSS through the DPL and MCL. Additional qualitative results are provided in Supplementary Material. + +# 5.2 Ablation Studies + +To demonstrate the effectiveness of our proposed method, we conduct an ablation study as depicted in Fig. 4. Here, we set the baseline as the model introduced in Sec. 3.2. The experiments are conducted on the SemanticPOSS within scenario-A, and the results are averaged over four splits. Please refer to the Supplementary Material for the discussion with class-wise results and analysis on the hash code dimension $L$ , mixing ratio $r$ , weight $\lambda$ , and sub-sampling in GM. DPL helps to recognize known classes. Comparing the baseline and baseline+DPL, we observe a significant improvement in performance for the known classes. When the model is trained only with the baseline, hash codes from the + +![](images/d57aa633c0bbdfcd46498159525d6ba95e96d08b74271956cfea8f989217ee6f.jpg) +(a) +Fig. 4: (a) Results of using dual prototypical learning (DPL) with the baseline and SMILE. (b) Ablation study of the proposed method. Base refers to the baseline model. + +
Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAll
(a) Base11.9825.1221.7219.5129.6526.90
(b) Base+DPL5.3340.0831.3317.8946.4338.96
(c) Base+DPL+DM16.4848.2440.2031.3954.8348.58
(d) Base+DPL+DM+GM21.5446.6340.3637.0951.5747.53
+ +(b) + +learned representations are sensitive to intra-class variation, which is supported by the poor performance for known classes. Conversely, adding hash prototypes through DPL imposes additional constraints on the representation, making it more suitable for hash code-based clustering. To further analyze the effects of DPL, we conduct an experiment by integrating the DPL with the baseline and SMILE [14]. As shown in Fig. 4a, when DPL is used together, the performance for known classes of both baseline and SMILE is significantly improved in both Strict- and Greedy-Hungarian. This implies that the proposed DPL effectively improves the robustness of learned representation under intra-class variation. The proposed DPL benefits both baseline and SMILE, demonstrating its effectiveness in a hash-code-based framework for addressing the OCDSS problem. + +MCL helps to cognize unknown classes. To excavate the ability to discover unknown classes, we proposed mixing-based category learning based on DPL. With DM, the prototypes become discriminative for ID and OOD. This enhancement leads to improved performance not only for known classes but also for unknown classes. The proposed GM further boosts the performance of unknown classes. Comparing the results of using (c) DM and (d) DM+GM in Fig. 4b, we observe approximately $5\%$ and $5.7\%$ gains in Strict- and Greedy-Hungarian, respectively. Although performance on known classes experiences a slight reduction, considering the inherent challenge of distinguishing unknown classes that the model has not seen during training, this underscores the necessity of GM. + +# 6 Conclusion + +In this paper, we introduce a new problem named on-the-fly category discovery for LiDAR semantic segmentation. Along with the baseline model based on hash coding, we propose dual prototypical learning to enhance the model's recognition ability. We further design the mixing-based category learning to inductively transfer recognition knowledge from known classes to unknown classes. Extensive experiments demonstrate the effectiveness of the proposed method for recognizing known classes and discovering unknown classes. While the proposed method effectively addresses the problem, performance disparities between the known and unknown classes remain. In this respect, exploring suitable representation learning for discovering unknown classes presents a promising research direction. + +# Acknowledgments + +This work was supported by the Technology Innovation Program (1415187329, 20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636). + +# References + +1. Alonso, I., Riazuelo, L., Montesano, L., Murillo, A.C.: 3d-mininet: Learning a 2d representation from point clouds for fast and efficient 3d lidar semantic segmentation. IEEE Robotics and Automation Letters 5(4), 5432-5439 (2020) +2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023) +3. Basart, S., Mantas, M., Mohammadreza, M., Jacob, S., Dawn, S.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (2022) +4. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019) +5. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016) +6. Cao, K., Brbic, M., Leskovec, J.: Open-world semi-supervised learning. In: International Conference on Learning Representations (ICLR) (2022) +7. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021) +8. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 318-334 (2022) +9. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021) +10. Chen, K., Oldja, R., Smolyanskiy, N., Birchfield, S., Popov, A., Wehr, D., Eden, I., Pehserl, J.: Mvlidarnet: Real-time multi-class scene understanding for autonomous driving using multiple views. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2288-2294. IEEE (2020) +11. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12547-12556 (2021) +12. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019) + +13. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: International Symposium on Visual Computing. pp. 207-222. Springer (2020) +14. Du, R., Chang, D., Liang, K., Hospedales, T., Song, Y.Z., Ma, Z.: On-the-fly category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11691-11700 (2023) +15. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021) +16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016) +17. Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Navab, N., Busam, B., Tombari, F.: Segmenting known objects and unseen unknowns without prior knowledge. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 19321-19332 (2023) +18. Ge, Z., Demyanov, S., Chen, Z., Garnavi, R.: Generative openmax for multi-class open set classification. In: British Machine Vision Conference. BMVA Press (2017) +19. Gerdzhev, M., Razani, R., Taghavi, E., Bingbing, L.: Tornado-net: multiview total variation semantic segmentation with diamond inception module. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9543-9549. IEEE (2021) +20. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018) +21. Gu, P., Zhang, C., Xu, R., He, X.: Class-relation knowledge distillation for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16474-16483 (2023) +22. Guo, Y., Camporese, G., Yang, W., Sperduti, A., Ballan, L.: Conditional variational capsule network for open set recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 103-111 (2021) +23. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. In: International Conference on Learning Representations (2019) +24. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019) +25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1 +26. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11108-11117 (2020) +27. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021) +28. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021) + +29. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020) +30. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023) +31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021) +32. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) +33. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017) +34. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) +35. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9425-9434 (2023) +36. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems 31 (2018) +37. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019) +38. Lu, T., Ding, X., Liu, H., Wu, G., Wang, L.: Link: Linear kernel for lidar-based 3d perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1105-1115 (2023) +39. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019) +40. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 613-628 (2018) +41. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019) +42. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693. IEEE (2020) +43. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) +44. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +45. Riz, L., Saltori, C., Ricci, E., Poiesi, F.: Novel class discovery for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9393-9402 (2023) +46. Ruder, S.: An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747 (2016) + +47. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012) +48. Shu, Y., Shi, Y., Wang, Y., Huang, T., Tian, Y.: P-odn: Prototype-based open deep network for open set recognition. Scientific reports 10(1), 7146 (2020) +49. Sun, X., Yang, Z., Zhang, C., Ling, K.V., Peng, G.: Conditional gaussian distribution learning for open set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13480-13489 (2020) +50. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European conference on computer vision. pp. 685-702. Springer (2020) +51. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019) +52. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022) +53. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019) +54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021) +55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 1887-1893. IEEE (2018) +56. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 4376-4382. IEEE (2019) +57. Xia, Y., Zhang, Y., Liu, F., Shen, W., Yuille, A.L.: Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 145-161. Springer (2020) +58. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019) +59. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020) +60. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021) +61. Zhao, Y., Zhong, Z., Sebe, N., Lee, G.H.: Novel class discovery in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4340-4349 (2022) + +62. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021) +63. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021) +64. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2021) +65. Zhou, H., Zhu, X., Song, X., Ma, Y., Wang, Z., Li, H., Lin, D.: Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550 (2020) +66. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018) \ No newline at end of file diff --git a/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/images.zip b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..99b71e3001f3a65e7225efd12f2cbacd23d64077 --- /dev/null +++ b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a37a946301fb48ef47d33f3d3924ebecb00307a4624ff5189ed9a05b21e2eacd +size 621976 diff --git a/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/layout.json b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e74b7d7738b58e44e99a46b0c77addcd27f266a2 --- /dev/null +++ b/2024/On-the-fly Category Discovery for LiDAR Semantic Segmentation/layout.json @@ -0,0 +1,9751 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 111, + 452, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 111, + 452, + 148 + ], + "spans": [ + { + "bbox": [ + 161, + 111, + 452, + 148 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "text", + "content": "Hyeonseong Kim" + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "inline_equation", + "content": "^{\\text{①}}" + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "text", + "content": ", Sung-Hoon Yoon" + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "inline_equation", + "content": "^{\\text{②}}" + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "text", + "content": ", Minseok Kim" + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "inline_equation", + "content": "^{\\text{③}}" + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "text", + "content": ", and Kuk-Jin Yoon" + }, + { + "bbox": [ + 135, + 168, + 478, + 180 + ], + "type": "inline_equation", + "content": "^{\\text{④}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 192, + 190, + 421, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 190, + 421, + 212 + ], + "spans": [ + { + "bbox": [ + 192, + 190, + 421, + 212 + ], + "type": "text", + "content": "Visual Intelligence Lab., KAIST, Korea {brian617,yoon307,alstjrx1x1,kjyoon}@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 242, + 452, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 242, + 452, + 483 + ], + "spans": [ + { + "bbox": [ + 160, + 242, + 452, + 483 + ], + "type": "text", + "content": "Abstract. LiDAR semantic segmentation is important for understanding the surrounding environment in autonomous driving. Existing methods assume closed-set situations with the same training and testing label space. However, in the real world, unknown classes not encountered during training may appear during testing, making it difficult to apply existing methodologies. In this paper, we propose a novel on-the-fly category discovery method for LiDAR semantic segmentation, aiming to classify and segment both unknown and known classes instantaneously during test time, achieved solely by learning with known classes in training. To embed instant segmentation capability in an inductive setting, we adopt a hash coding-based model with an expandable prediction space as a baseline. Based on this, dual prototypical learning is proposed to enhance the recognition of the known classes by reducing the sensitivity to intra-class variance. Additionally, we propose a novel mixing-based category learning framework based on representation mixing to improve the discovery capability of unknown classes. The proposed mixing-based framework effectively models out-of-distribution representations and learns to semantically group them during training while distinguishing them from in-distribution representations. Extensive experiments on SemanticKITTI and SemanticPOSS datasets demonstrate the superiority of the proposed method compared to the baselines. The code is available at https://github.com/hskim617/OCDSS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 494, + 452, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 494, + 452, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 494, + 452, + 517 + ], + "type": "text", + "content": "Keywords: Autonomous driving " + }, + { + "bbox": [ + 160, + 494, + 452, + 517 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 452, + 517 + ], + "type": "text", + "content": " LiDAR semantic segmentation " + }, + { + "bbox": [ + 160, + 494, + 452, + 517 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 452, + 517 + ], + "type": "text", + "content": " On-the-fly category discovery" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 537, + 229, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 537, + 229, + 550 + ], + "spans": [ + { + "bbox": [ + 133, + 537, + 229, + 550 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "LiDAR semantic segmentation is crucial for recognizing the surrounding environment in autonomous driving and has recently made significant progress [2, 11, 12, 26, 30, 38, 39, 50, 59, 65]. Although these models demonstrate promising results across various datasets, their effectiveness is constrained to closed-set settings, as they are assumed and trained within such scenarios. In real-world scenarios, however, unknown classes are bound to appear during testing, making it incapable of predicting unknown classes for closed-set models as in Fig. 1a." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 223, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 223, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 223, + 665 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 478, + 277 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 478, + 277 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 478, + 277 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 478, + 277 + ], + "type": "image", + "image_path": "09ed99855aaa56f7a326ba3ca593a407832ca8277c28237e84051c843a8f6af9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 342 + ], + "type": "text", + "content": "Fig. 1: Illustration of training and testing scenarios for on-the-fly category discovery in LiDAR semantic segmentation and relevant tasks. The goal of each task is (a) recognizing known classes, (b) further identifying and rejecting unknown classes, (c) learning clustering using unlabeled, unknown class data, and (d) cognizing unknown classes that are not identified during training while recognizing known classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "To address this challenge, open-set recognition (OSR) [47], which aims to filter out unknown classes has been explored in diverse fields, including image classification [3,5,9,49,58], semantic segmentation [7,31,37,57], and 3D semantic segmentation [8,35]. Considering the nature of the human cognitive system that can cognize novel visual concepts based on understanding, simply rejecting unknown classes as in OSR (Fig. 1b) is unsatisfactory." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 581 + ], + "type": "text", + "content": "To narrow the gap between humans and machines by learning novel concepts, novel category discovery (NCD) [15, 23, 24, 28, 63] and on-the-fly category discovery (OCD) [14] are emerging as new research fields and attracting attention. Given labeled data of known classes and unlabeled data of unknown classes, the goal of NCD is to cluster the unlabeled data with the same semantics, aiming to cognize them as in Fig. 1c. Therefore, NCD is formulated as transductive learning in which data of unknown classes are available, making it incapable of predicting unknown classes not identified during training. On the other hand, OCD aims for a more realistic scenario by learning solely from data of known classes and inductively cognizing unknown classes at test time. Furthermore, during testing, the model should independently classify each test sample instantaneously, regardless of the distribution of the test data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we broaden the scope of OCD to LiDAR point clouds and introduce a new problem, on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). As shown in Fig. 1d, our goal is to classify and segment both known and unknown classes instantaneously in an inductive manner during test time, achieved solely by learning with known classes in training. To achieve the goal, one straightforward solution is adopting the previous approach in image classification [14] and applying it to the LiDAR semantic segmentation." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Therefore, we implemented the hash coding-based model proposed by [14] as a baseline for LiDAR point clouds. This model utilizes hash codes as class descriptors, generated by binarizing each dimension of the learned representation based on its sign. However, extending it to the LiDAR domain entails two limitations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 171, + 480, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 480, + 291 + ], + "type": "text", + "content": "Firstly, due to LiDAR's sparse nature and lack of texture information, hash coding in learned representations is more sensitive to intra-class variance. In practice, we observed that the method for reducing sensitivity to intra-class variance by disentangling sign and magnitude [14] had shown unsatisfactory results in the LiDAR domain (refer Fig. 4a). Secondly, prior work relies on representation learning from known classes to discover unknown classes. Without a dedicated mechanism for discovering unknown classes, only a tentative expectation that the learned representations will form semantically meaningful groups for unknown classes is possible. This reliance, combined with the aforementioned limitation, complicates OCDSS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 299, + 481, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 481, + 574 + ], + "type": "text", + "content": "Therefore, we propose a novel learning method for OCDSS that learns representations robust to intra-class variance in hash coding and suitable for unknown class discovery. First, we propose dual prototypical learning (DPL) that constrain the representation space by introducing extra hash prototypes, effectively reducing the sensitivity of hash codes to intra-class variance. Specifically, the hash prototypes enforce additional constraints to ensure that representations within the same intra-class share the same hash code, dramatically enhancing recognition capability for known classes. Second, we propose a novel mixing-based category learning (MCL) framework to improve the ability to cognize unknown classes. The proposed framework consists of three modules: a mixing module, a discrimination module, and a grouping module. The mixing module models out-of-distribution (OOD) representation through representation mixing and obtains mixed representation to mimic unknown classes during training. The discrimination module then enhances the model's OOD discrimination ability by learning to distinguish the representation of known classes from mixed representations. In addition, the grouping module develops the model's clustering capability for OOD representations by learning to cluster mixed representations within the same semantics. Meanwhile, we integrate the hash prototypes of DPL during MCL to fully exploit the discovering capability for unknown classes. Finally, we introduce an evaluation protocol for OCDSS and evaluate our approach on SemanticKITTI [4] and SemanticPOSS [42] datasets. Extensive experiments demonstrate the superiority of our method compared to baselines, while ablation studies show the effectiveness of each component." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Our main contributions are four-fold: (I) We introduce a new problem of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). (II) We propose dual prototypical learning that effectively learns representations that are less sensitive to intra-class variation in hash coding. (III) We present a novel mixing-based category learning framework, enabling the model to discover unknown classes. (IV) We introduce a new evaluation protocol for OCDSS and demonstrate the superiority of our method through extensive experiments." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 342 + ], + "type": "text", + "content": "LiDAR Semantic Segmentation. LiDAR semantic segmentation is the task of inferring semantic labels to individual 3D points within the LiDAR point cloud. These approaches can be broadly categorized into point-based, projection-based, and voxel-based according to the way they represent and process point clouds. Point-based methods [26,36,44,51] directly operate on points in line with the foundational work PointNet [43] by using multi-layer perceptrons. Despite their advantages, such as minimal information loss and few parameters, applying these methods to large-scale LiDAR point clouds demands heavy computational resources and memory. Projection-based methods have been proposed to address efficiency and leverage advanced 2D convolutional neural network architectures. These methods project LiDAR points onto a range image [1,2,10,13,19,30,39,55, 56] using spherical projection or a bird-eye-view [59]. However, the projection of 3D points into 2D representation might lose some geometric information. Voxel-based methods represent point clouds as 3D voxels and process them through 3D convolution [66] or sparse convolution [11, 12, 20, 38, 50, 65]. In particular, sparse convolution achieves high performance with less computation, so we adopt MinkowskiNet [12] as our backbone architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 533 + ], + "type": "text", + "content": "Open-set Recognition. Open-set recognition (OSR) was first introduced in [47], extending the closed-set assumption to realistic scenarios where unknown classes appear during testing. As OSR aims to detect unknown classes, previous studies estimate the unknown probability based on extreme value theory [5] or estimate uncertainty through maximum softmax probability [25], maximum logit [3], or Bayesian inference approximation [16, 33]. Other approaches include training a redundancy classifier for unknown classes [54, 64] or using metric learning [9, 22, 48]. Meanwhile, generative model-based approaches employ generative adversarial networks to synthesize unknown class samples [18, 40] or use image reconstruction error for the criterion of determining unknown classes [41, 49, 58]. Unlike these methods for object recognition, studies for semantic segmentation [7, 31, 37, 57], panoptic segmentation [17, 27] in open-set situations have also been proposed. Recently, open-set semantic segmentation for LiDAR point clouds has emerged [8, 35], extending an open-set problem to the LiDAR domain. While OSR methods are aware of unknown classes, they focus on identification, whereas we aim to go beyond achieving semantically cognizing them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "content": "Novel Category Discovery. Novel category discovery (NCD), first formalized in [24], is another relevant task that aims to cognize unknown classes in unlabeled data by transferring the knowledge from known classes in labeled data. After DTC [24] proposed deep transfer clustering for NCD, several studies were suggested, achieving significant progress [15, 21, 23, 28, 60, 62, 63]. These approaches first learn representations using supervised learning from labeled known classes, and then discover unknown classes by transferring knowledge through pseudolabeling on unknown classes based on pair-wise similarity [23, 28, 60, 62, 63] or cluster assignment [15, 21]. Additionally, EUMS [61] and NOPS [45] extend the NCD problem to semantic segmentation in images and LiDAR point clouds, respectively. Nevertheless, NCD usually assumes the number of unknown classes" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "as prior knowledge [15, 23, 28, 60, 62, 63], and presuppose all unlabeled data as unknown classes. Hence, generalized category discovery (GCD) was introduced where unlabeled data contain both known and unknown classes [6, 52]. However, as pointed out in [14], two limitations in prior settings still exist: 1) transductive learning and 2) offline evaluation. In prior settings, category discovery is formulated as transductive learning in which data of unknown classes are available during training. Also, offline evaluation constrains the results depending on the entire test data distribution, making instant feedback on each test sample impossible. Therefore, [14] proposed a new task named on-the-fly category discovery (OCD), which aims to recognize known classes and cognize unknown classes instantaneously in an inductive manner. In this study, we tackle the challenging OCD for LiDAR semantic segmentation, which instantaneously segments LiDAR point clouds from both known and unknown classes while learning exclusively from known classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 308, + 202, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 202, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 202, + 321 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": "In Sec. 3.1, we first elaborate on our problem formulation of on-the-fly category discovery for LiDAR semantic segmentation (OCDSS). In Sec. 3.2, we introduce a hash coding-based baseline model for OCDSS that embeds expandable prediction space and instant inference capability within an inductive setting. Then, in Sec. 3.3, we describe our dual prototypical learning for reducing the sensitivity of representation in hash coding, improving the closed-set performance. In Sec. 3.4, we propose a mixing-based category learning framework to enable the model to possess category discovery capabilities for both known and unknown classes. Fig. 2 shows the overview of the proposed method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 472, + 269, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 269, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 269, + 483 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": "In this section, we formulate the problem of OCDSS, which aims to recognize known classes and cognize unknown classes during testing in an inductive manner. Let " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{S} = \\{P_{S}\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " be a support set used for training and " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{Q} = \\{P_{Q}\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " be a query set for testing. Then, we have 3D point clouds " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "P_{S} = \\{(x,y)\\} \\in \\mathbb{R}^{3} \\times Y_{S}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "P_{Q} = \\{(x,y)\\} \\in \\mathbb{R}^{3} \\times Y_{Q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": ", representing a single LiDAR scan from " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{Q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": ", respectively. Here, " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " denotes each point, and " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " is the corresponding label. " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y_{S}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y_{Q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " are the label spaces of support and query sets. Following the setting of OCD [14], " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y_{S}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " defines the known classes observed during training, while " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y_{Q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " contains both known and unknown classes that appear at testing, i.e. " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "Y_{S} \\subset Y_{Q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": ". We denote the number of known and unknown classes as " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "C_k = |Y_S|" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "C_u = |Y_Q\\backslash Y_S|" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": ", where we assume " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "C_u" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " is unknown. This is different from existing NCD approaches [15,23,28,45,60,62,63] that assume " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "C_u" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " as a known prior. The goal is to train a model using only " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " so that the model can inductively classify and segment point clouds in " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{Q}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " instantaneously." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 476, + 202 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 476, + 202 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 476, + 202 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 476, + 202 + ], + "type": "image", + "image_path": "24583f59c1bcc33370058ce4388d61fd51813753391fbe005e1a441351288a41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 204, + 319, + 211 + ], + "lines": [ + { + "bbox": [ + 252, + 204, + 319, + 211 + ], + "spans": [ + { + "bbox": [ + 252, + 204, + 319, + 211 + ], + "type": "text", + "content": "(b) Discrimination Module" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 213, + 470, + 228 + ], + "lines": [ + { + "bbox": [ + 406, + 213, + 470, + 228 + ], + "spans": [ + { + "bbox": [ + 406, + 213, + 470, + 228 + ], + "type": "text", + "content": ": Mixed Representation O:Origin Representation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 148, + 209, + 242, + 278 + ], + "blocks": [ + { + "bbox": [ + 164, + 203, + 211, + 211 + ], + "lines": [ + { + "bbox": [ + 164, + 203, + 211, + 211 + ], + "spans": [ + { + "bbox": [ + 164, + 203, + 211, + 211 + ], + "type": "text", + "content": "(a) Mixing Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 209, + 242, + 278 + ], + "lines": [ + { + "bbox": [ + 148, + 209, + 242, + 278 + ], + "spans": [ + { + "bbox": [ + 148, + 209, + 242, + 278 + ], + "type": "image", + "image_path": "b2f335596e5dffca1f7ac12673f6d604ae2927f6cb773217e7713939c79e39a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 290, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 380 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed framework. Given the LiDAR point cloud, the model extracts the features through feature extractor " + }, + { + "bbox": [ + 130, + 290, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 290, + 482, + 380 + ], + "type": "text", + "content": ", which are then mapped into the prediction space to obtain the representations by the projection head " + }, + { + "bbox": [ + 130, + 290, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 290, + 482, + 380 + ], + "type": "text", + "content": ". Through dual prototypical learning (DPL), the model learns representations suitable for hash code-based clustering, enhancing its recognition ability for known classes. Besides, mixing-based category learning (MCL) improves the model's cognition capability for unknown classes through three modules: mixing module, discrimination module, and grouping module. The final prediction is obtained by the hash code-based clustering." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 260, + 213, + 321, + 275 + ], + "blocks": [ + { + "bbox": [ + 260, + 213, + 321, + 275 + ], + "lines": [ + { + "bbox": [ + 260, + 213, + 321, + 275 + ], + "spans": [ + { + "bbox": [ + 260, + 213, + 321, + 275 + ], + "type": "image", + "image_path": "7ae4aa88713eaacb26edc580f3dafde6db11540748ec9a581e8bb513302143fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 335, + 213, + 397, + 275 + ], + "blocks": [ + { + "bbox": [ + 335, + 204, + 388, + 212 + ], + "lines": [ + { + "bbox": [ + 335, + 204, + 388, + 212 + ], + "spans": [ + { + "bbox": [ + 335, + 204, + 388, + 212 + ], + "type": "text", + "content": "(c) Grouping Module" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 213, + 397, + 275 + ], + "lines": [ + { + "bbox": [ + 335, + 213, + 397, + 275 + ], + "spans": [ + { + "bbox": [ + 335, + 213, + 397, + 275 + ], + "type": "image", + "image_path": "683bedbca81f632b7171987c71a7ae038df792982cfa149f3fdcd790c669dd24.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 406, + 232, + 470, + 263 + ], + "blocks": [ + { + "bbox": [ + 406, + 232, + 470, + 263 + ], + "lines": [ + { + "bbox": [ + 406, + 232, + 470, + 263 + ], + "spans": [ + { + "bbox": [ + 406, + 232, + 470, + 263 + ], + "type": "image", + "image_path": "728bfcb94efba52df3e479f8b1f9a42829fd0e8fb6ee39fa3da1d042142278b1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 400, + 339, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 400, + 339, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 400, + 339, + 412 + ], + "type": "text", + "content": "3.2 Hash Coding-based Baseline Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": "We first introduce a hash coding-based baseline model for OCDSS. To embed the expandable prediction space and instant inference capability while learning the model inductively, we adopt the basic pipeline of hash coding-based model from [14]. Formally, our baseline model consists of a feature extractor " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": " and a projection head " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": ". Given point cloud " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "P_{S} = \\{(x,y)\\}_{i=1}^{N}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": ", the feature extractor first extracts the feature representation " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": " for each point " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": ". The projection head maps " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": " to the prediction space, obtaining the projected representation " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "p_{i} = \\phi(f_{i}) \\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": ", which is used for the hash coding. Here, " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": " is the dimension of the hash code. Then, the hash code for point " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\text{hash}(\\phi(\\mathcal{E}(x_{i})))" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\text{hash}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": " is a hashing operation. For an arbitrary vector " + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = [v_{1}, \\ldots, v_{l}, \\ldots, v_{L}] \\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 130, + 418, + 482, + 550 + ], + "type": "text", + "content": ", the hashing operation is defined as," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 197, + 555, + 481, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 197, + 555, + 481, + 588 + ], + "type": "interline_equation", + "content": "h a s h (\\mathbf {v}) = \\left[ v _ {1} ^ {*}, \\dots , v _ {l} ^ {*}, \\dots , v _ {L} ^ {*} \\right], v _ {l} ^ {*} = \\left\\{ \\begin{array}{l l} 1, & v _ {l} \\geq 0 \\\\ 0, & v _ {l} < 0 \\end{array} . \\right. \\tag {1}", + "image_path": "944ce98aa7cff2cd15bde42a3c27582b28d625b5e3799ad0e86df9d405f228f5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To predict classes of points from " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{Q}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " during testing, we use the hash code as a class descriptor where points with the same code are considered to be in the same class. This hash coding-based inference allows the model to have an expanded label space to predict new classes beyond the known classes in an inductive setting. Moreover, it also has the advantage of being able to immediately predict the result of each test sample individually." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 236 + ], + "type": "text", + "content": "For representation learning of the baseline model, [14] used supervised contrastive learning [29]; however, there is a lack of clear design choices for representation learning in LiDAR point clouds. For example, in the case of contrastive learning where pair-wise computation between samples is required, it is not suitable due to its high demand for computation and memory. Instead, we use a standard cross-entropy loss on the known classes. To learn a more discriminative representation, we employ parametric prototypes " + }, + { + "bbox": [ + 131, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{g_1, g_2, \\dots, g_{C_k}\\} \\in \\mathbb{R}^{C_k \\times L}" + }, + { + "bbox": [ + 131, + 116, + 482, + 236 + ], + "type": "text", + "content": " for known classes as in [7] and replace the classifier as cosine similarity between the prototype and the projected representation " + }, + { + "bbox": [ + 131, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 131, + 116, + 482, + 236 + ], + "type": "text", + "content": ". Specifically, we compute the distance based on the cosine similarity as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 247, + 241, + 481, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 241, + 481, + 265 + ], + "spans": [ + { + "bbox": [ + 247, + 241, + 481, + 265 + ], + "type": "interline_equation", + "content": "d _ {c o s} \\left(p _ {i}, g _ {j}\\right) = 1 - \\frac {p _ {i} \\cdot g _ {j}}{\\| p _ {i} \\| \\| g _ {j} \\|}. \\tag {2}", + "image_path": "0720b67aea9123ebccaf65ea8e232f0c69273d6d03604cd4f4aed72c3c88b74c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "text", + "content": "Then, let us define " + }, + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{G}}(p_i,y_i)" + }, + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "text", + "content": " that computes the softmax probability of " + }, + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 269, + 481, + 293 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 299, + 481, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 299, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 215, + 299, + 481, + 328 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\mathcal {G}} \\left(p _ {i}, y _ {i}\\right) = \\frac {\\exp \\left(- d _ {\\text {c o s}} \\left(p _ {i} , g _ {y _ {i}}\\right) / \\tau_ {\\text {c o s}}\\right)}{\\sum_ {j = 1} ^ {C _ {k}} \\exp \\left(- d _ {\\text {c o s}} \\left(p _ {i} , g _ {j}\\right) / \\tau_ {\\text {c o s}}\\right)}, \\tag {3}", + "image_path": "4a22cb6769745c85f795e08cff30f4251eed27d92879dd49cb687ef769e99639.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "inline_equation", + "content": "g_{y_i}" + }, + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "text", + "content": " denotes the prototype corresponding to the label " + }, + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "inline_equation", + "content": "\\tau_{cos}" + }, + { + "bbox": [ + 131, + 332, + 481, + 357 + ], + "type": "text", + "content": " is a temperature parameter. The loss function for the baseline model is defined as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 362, + 481, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 362, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 247, + 362, + 481, + 386 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b a s e}} = \\sum_ {i} - \\log \\sigma_ {\\mathcal {G}} \\left(p _ {i}, y _ {i}\\right). \\tag {4}", + "image_path": "b53342b695a79ecae742e8b58295afd31197715c87f0f14e0bf39d7b96895033.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 399, + 299, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 299, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 299, + 412 + ], + "type": "text", + "content": "3.3 Dual Prototypical Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "text", + "content": "The baseline model is trained to generate similar representations for points belonging to the same class. However, during inference, points are classified into different classes if they do not share the same hash code, which can be sensitive to intra-class variance. Due to the sparse nature and lack of texture information in LiDAR, hash coding in learned representations becomes more sensitive to intra-class variance compared to images. To address this issue, we propose dual prototypical learning (DPL) to learn the representation robust to intra-class variance in hash coding. Specifically, DPL employs extra parametric prototypes to constrain the representation space, denoted as " + }, + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = \\{h_1, h_2, \\ldots, h_{C_k}\\} \\in \\mathbb{R}^{C_k \\times L}" + }, + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "text", + "content": ", referred to as hash prototypes. Similar to " + }, + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "text", + "content": ", we calculate the distance between the projected representation " + }, + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 131, + 416, + 482, + 596 + ], + "type": "text", + "content": " and the hash prototypes. In this case, based on the fact that the sign of each dimension in representation determines the hash code, we utilize the L2 distance between signs instead of cosine similarity. In practice, for stable optimization, we use the hyperbolic tangent function, tanh, instead of the sign operation as," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 215, + 602, + 481, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 602, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 215, + 602, + 481, + 617 + ], + "type": "interline_equation", + "content": "d _ {h a s h} \\left(p _ {i}, h _ {j}\\right) = \\left\\| t a n h \\left(p _ {i}\\right) - t a n h \\left(h _ {j}\\right) \\right\\| _ {2} ^ {2} / L. \\tag {5}", + "image_path": "94641bddbe31071cfadb9e6cc5f9d54033b64d6c494247d4c058f2a1bdf0ad2e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 621, + 467, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 467, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 467, + 634 + ], + "type": "text", + "content": "Similar to " + }, + { + "bbox": [ + 131, + 621, + 467, + 634 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{G}}" + }, + { + "bbox": [ + 131, + 621, + 467, + 634 + ], + "type": "text", + "content": " in Eq. (3), we define softmax probability " + }, + { + "bbox": [ + 131, + 621, + 467, + 634 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{H}}(p_i, y_i)" + }, + { + "bbox": [ + 131, + 621, + 467, + 634 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 207, + 639, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 639, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 207, + 639, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\mathcal {H}} \\left(p _ {i}, y _ {i}\\right) = \\frac {\\exp \\left(- d _ {\\text {h a s h}} \\left(p _ {i} , h _ {y _ {i}}\\right) / \\tau_ {\\text {h a s h}}\\right)}{\\sum_ {j = 1} ^ {C _ {k}} \\exp \\left(- d _ {\\text {h a s h}} \\left(p _ {i} , h _ {j}\\right) / \\tau_ {\\text {h a s h}}\\right)}, \\tag {6}", + "image_path": "678f9ca2de7161ad89db1ea600478fdd120038aecb687ce652aa0374bf2e683c.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 448, + 103 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "h_{y_i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " denotes the hash prototype corresponding to the label " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\tau_{hash}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " is a temperature parameter. The prototypical learning loss based on hash prototypes is defined as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 245, + 152, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 152, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 245, + 152, + 481, + 175 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h a s h}} = \\sum_ {i} - \\log \\sigma_ {\\mathcal {H}} \\left(p _ {i}, y _ {i}\\right). \\tag {7}", + "image_path": "be0302d2f365245c5cae3e285992f473ee91a8ad7b82af506796a882782ef2b0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 479, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 479, + 190 + ], + "type": "text", + "content": "Along with " + }, + { + "bbox": [ + 132, + 178, + 479, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{base}" + }, + { + "bbox": [ + 132, + 178, + 479, + 190 + ], + "type": "text", + "content": ", the final loss function of dual prototypical learning is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 204, + 481, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 204, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 256, + 204, + 481, + 216 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D P L} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {h a s h}}. \\tag {8}", + "image_path": "a21e45542438d84cbaf775e4d0218ee3f18e42a251ee906c542d75f79ef97d1e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 220, + 479, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 479, + 256 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 479, + 256 + ], + "type": "text", + "content": "With the proposed DPL, we impose additional constraints on representations to make hash codes less sensitive to intra-class variance, improving the model's recognition capability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 273, + 326, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 326, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 326, + 285 + ], + "type": "text", + "content": "3.4 Mixing-based Category Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 291, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 480, + 363 + ], + "type": "text", + "content": "Along with the recognition, our goal is to train the model to cognize unknown classes through semantic understanding. Our basic idea to achieve this goal is to mimic unknown classes during training and learn the model to distinguish and group them from known classes for a comprehensive understanding of unknown semantics. To this end, we propose mixing-based category learning (MCL), which consists of a mixing module, a discrimination module, and a grouping module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "content": "Mixing Module. We mix the representations of different known classes to mimic the unknown classes during training. Our intention in using representation mixing is that 1) it can effectively model OOD representations, as stated in [64], and 2) it can further mimic diverse unknown classes by modeling OOD representations that share common semantics depending on which classes are mixed. Specifically, we use manifold mixup technique [53] to mix two feature representations " + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "content": " of different classes " + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "inline_equation", + "content": "y_{i} \\neq y_{j}" + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "content": ". The mixed representation " + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "inline_equation", + "content": "\\overline{f_{i,j}}" + }, + { + "bbox": [ + 130, + 363, + 480, + 460 + ], + "type": "text", + "content": " is calculated as," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 260, + 466, + 480, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 466, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 260, + 466, + 480, + 480 + ], + "type": "interline_equation", + "content": "\\overline {{f _ {i , j}}} = r f _ {i} + (1 - r) f _ {j}, \\tag {9}", + "image_path": "28e3409dd4931de672fd8e87c6e36e28f959abb964d6c644bb051e2f79799464.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": " is a mixing ratio. We can also define the imaginary corresponding label " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "\\overline{y_{i,j}}" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "\\overline{f_{i,j}}" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": ", referred to as a mix label. Note that if " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "(i,j) \\neq (k,l)" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "\\overline{y_{i,j}} \\neq \\overline{y_{k,l}}" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": ". These mixed representations are considered as OOD representations in the following modules." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "text", + "content": "Discrimination Module. In the discrimination module (DM), we train the model to distinguish mixed representations from prototypes. In particular, we learn parametric prototypes " + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "text", + "content": " to be discriminative from the mixed representation. To this end, we first project " + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\overline{f_{i,j}}" + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "text", + "content": " into projection space and obtain " + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\phi(\\overline{f_{i,j}})" + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "text", + "content": ". Then, we enforce " + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "inline_equation", + "content": "\\phi(\\overline{f_{i,j}})" + }, + { + "bbox": [ + 131, + 535, + 480, + 606 + ], + "type": "text", + "content": " not to be close to the prototype of any specific class in the form of entropy maximization as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 207, + 615, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 615, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 207, + 615, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d i s c}} = \\sum_ {i, j} (\\sum_ {k} \\sigma_ {\\mathcal {G}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\log \\sigma_ {\\mathcal {G}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\tag {10} \\\\ + \\sum_ {k} \\sigma_ {\\mathcal {H}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right) \\log \\sigma_ {\\mathcal {H}} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), k\\right)). \\\\ \\end{array}", + "image_path": "91605956da0986edbc83c58040379130a667783826f3b59330919ee87ca258a7.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": "Notably, we stop the gradient of " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\phi(f_{i,j})" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": " and only optimize prototypes. In addition, we incorporate mixed representations that are mixed after being projected into the " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{disc}" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": " to enhance discrimination for more diverse OOD representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "text", + "content": "Grouping Module. We propose a grouping module (GM) to enhance the model's understanding of the underlying semantics. Since the mixed representations might share common semantics depending on which classes are mixed, we can group them using the mix labels. For mathematical brevity, we omit the summation over " + }, + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "text", + "content": " here. Specifically, we employ contrastive loss [29] and propose contrastive learning on mixed representation based on mixed labels:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 233, + 481, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 233, + 481, + 265 + ], + "spans": [ + { + "bbox": [ + 186, + 233, + 481, + 265 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g p} = \\frac {- 1}{| P (i , j) |} \\sum_ {p \\in P (i, j)} \\log \\sigma_ {A (i, j)} \\left(\\frac {- d _ {c o s} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right) , \\phi \\left(\\overline {{f _ {p}}}\\right)\\right)}{\\tau_ {c o s}}\\right), \\tag {11}", + "image_path": "1b7877195731869baee027209985a3b45a888f4e7451023d007afb8994bed4f6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "A(i,j)" + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": " denotes all mixed representations excluding " + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\phi(\\overline{f_{i,j}})" + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "P(i,j)" + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": " is all positive pairs of " + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\phi(\\overline{f_{i,j}})" + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": " that share the same mix label. " + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\sigma_{(A(i,j))}" + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": " refers to a softmax operation with respect to " + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "A(i,j)" + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": ". In practice, we perform sub-sampling on " + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "inline_equation", + "content": "A(i,j)" + }, + { + "bbox": [ + 131, + 275, + 482, + 348 + ], + "type": "text", + "content": " to reduce the computation from heavy pair-wise operations. Additionally, we impose the following constraint so that mixed representations with the same mix label share the hash code as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 188, + 357, + 481, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 357, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 188, + 357, + 481, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g p - h a s h} = \\log \\sum_ {p \\in P (i, j)} \\exp \\left(d _ {h a s h} \\left(\\phi \\left(\\overline {{f _ {i , j}}}\\right), \\phi \\left(\\overline {{f _ {p}}}\\right)\\right) / \\tau_ {h a s h}\\right). \\tag {12}", + "image_path": "271ef2aa99fe7ae65b2f32187f1d85683c3ab4747bb882c8dbe703fef2aa21a5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 392, + 370, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 370, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 370, + 403 + ], + "type": "text", + "content": "The loss function for the proposed MCL is defined as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 415, + 480, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 415, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 228, + 415, + 480, + 427 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M C L} = \\mathcal {L} _ {\\text {d i s c}} + \\lambda \\left(\\mathcal {L} _ {g p} + \\mathcal {L} _ {g p - h a s h}\\right), \\tag {13}", + "image_path": "6728d66eaa0a0a9f454bc945a0b286c5d5785bf3e50da684a7b4921bb0adfa75.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "text", + "content": " is the weight for " + }, + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gp}" + }, + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gp - hash}" + }, + { + "bbox": [ + 131, + 436, + 328, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 449, + 360, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 449, + 360, + 459 + ], + "spans": [ + { + "bbox": [ + 146, + 449, + 360, + 459 + ], + "type": "text", + "content": "Finally, the total loss function of our method is," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 254, + 470, + 480, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 470, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 254, + 470, + 480, + 482 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {D P L} + \\mathcal {L} _ {M C L}. \\tag {14}", + "image_path": "9f84717e9237bc51105771c899793f114dd05a49f854af17f4075aae32a4351e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 500, + 273, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 273, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 273, + 513 + ], + "type": "text", + "content": "4 Evaluation Protocol" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 525, + 214, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 214, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 214, + 538 + ], + "type": "text", + "content": "4.1 Data Split" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": "Datasets. We evaluate our method on two datasets, SemanticKITTI [4] and SemanticPOSS [42]. SemanticKITTI consists of 43,552 frames with point-wise annotations of 19 classes. We follow the official split and use sequences 00 to 07 and 09 to 10 with 19,130 frames for training, and sequence 08 with 4,071 frames for validation. SemanticPOSS contains 2,988 frames with point-wise annotations of 13 classes. We use sequence 03 for validation, while the remaining sequences are used for training, following the official split." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 482, + 665 + ], + "type": "text", + "content": "Scenarios. We designed three scenarios considering various real-world situations. More precisely, we defined two unknown class types that can exist in the real world: observed-unknown and unobserved-unknown. Observed-unknown" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 448, + 103 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 156, + 477, + 212 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Data splits of scenario-A on SemanticPOSS and SemanticKITTI datasets. The " + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS}_i" + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{KITTI}_i" + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": " split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 156, + 477, + 212 + ], + "lines": [ + { + "bbox": [ + 140, + 156, + 477, + 212 + ], + "spans": [ + { + "bbox": [ + 140, + 156, + 477, + 212 + ], + "type": "table", + "html": "
SplitUnknown classesSplitUnknown classes
POSS0trashcan, traffic sign, bike, plantsKITTI0bicycle, other-vehicle, trunk, terrain, vegetation
POSS1pole, person, buildingKITTI1bicyclist, traffic-sign, other-ground, fence, road
POSS2rider, fence, groundKITTI2motorcyclist, motorcycle, pole, car, sidewalk
POSS3cone/stone, trunk, carKITTI3person, truck, parking, building
", + "image_path": "e67822f23c1d64112b7ae195a6e40142b46c0c06384434daf27ec2fbaab1e937.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 327 + ], + "type": "text", + "content": "class refers to cases where data was captured during acquisition but remains unlabeled. An example could be cases where instances are not included in the pre-defined category set or when it is challenging to define a class at the time of annotation, resulting in their assignment as a void or an ignore label. On the other hand, unobserved-unknown class corresponds to cases where it first appears during testing without being captured during data acquisition. We then constructed three scenarios based on the presence of each unknown class type." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 422 + ], + "type": "text", + "content": "Scenario-A: In scenario-A, we assess the discovery capability for observed-unknown classes. Following the previous open-set 3D semantic segmentation [8, 35] literature that addresses the observed-unknown classes in testing, we set the labels of unknown classes to be void and ignore them during training. Depending on the choice of the unknown class, we made four splits each for SemanticKITTI and SemanticPOSS, as shown in Tab. 1. The splits are selected considering the diversity of unknown classes and the relationship between classes. We provide details on the selection process in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 529 + ], + "type": "text", + "content": "Scenario-B: In scenario-B, we evaluate the capability to discover unobserved-unknown classes that do not appear in the training data. Specifically, we select relatively infrequent classes from SemanticKITTI, {other-vehicle, bicyclist}, as unknown classes to prevent a significant decrease in the amount of training data. Frames containing any points of these classes are then excluded from training data. As a result, 13,501 out of 19,130 frames are used for training. We did not use SemanticPOSS, as all classes appear in most frames. The evaluation is conducted for all classes, including unobserved-unknown classes, using the same validation data as the official benchmark." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 531, + 481, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 481, + 602 + ], + "type": "text", + "content": "Scenario-C: Both observed-unknown and unobserved-unknown classes are considered in scenario-C. In SemanticKITTI, we select {other-vehicle, parking} as observed-unknown classes, and {bicyclist} as an unobserved-unknown class. Frames containing any points of the unobserved-unknown class are excluded from training data, resulting in 17,954 frames for training, and observed-unknown classes are ignored during training. The evaluation proceeds as in scenario-B." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 621, + 208, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 208, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 208, + 632 + ], + "type": "text", + "content": "4.2Baselines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Since OCDSS necessitates inductive learning with known classes while enabling instant inference during testing, many previous approaches are impractical un" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "der our setting. Clustering-based methods do not provide instant inference and demand a substantial amount of memory due to the large number of points in the LiDAR point cloud, making them more impractical. Previous NCD method [45] is also unable to learn under our setting where unknown classes are unavailable. Therefore, following the previous OCD approach [14], we implement MLDG [34] and SMILE [14], representative OCD methods in image classification, in the 3D LiDAR points domain as baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 481, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 320 + ], + "type": "text", + "content": "Baseline. The baseline is the hash coding-based model introduced in Sec. 3.2. MLDG. MLDG is a meta-learning method for domain generalization. Since OCDSS is more like generalizing to unknown classes from known classes, we simulate introducing unknown classes during training. At each training iteration, we randomly divide known classes into two groups, using one group for meta-training and the other for meta-testing, and then applying MLDG to Baseline. SMILE. SMILE proposes sign-magnitude disentanglement to better capture the class-level semantics. Following [14], we modified the projection head " + }, + { + "bbox": [ + 130, + 200, + 481, + 320 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 200, + 481, + 320 + ], + "type": "text", + "content": " by replacing the final single MLP layer with two separate MLP layers for sign and magnitude predictions, and subsequently applying SMILE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": "Implementation Details. We use the same backbone network and optimization strategy for all baselines. Specifically, MinkUNet34 [12] is used as the feature extractor " + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": ". For the projection head " + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": ", we use multi-layer perceptrons. We use SGD optimizer [46] with learning rate 0.01, momentum 0.9, and weight decay " + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": " with the batch size of 4. We use fixed mixing ratio " + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "inline_equation", + "content": "r = 0.1" + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\tau_{cos} = 0.07" + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\tau_{hash} = 0.14" + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "inline_equation", + "content": "L = 12" + }, + { + "bbox": [ + 130, + 320, + 482, + 403 + ], + "type": "text", + "content": ". More details on baselines and implementation are provided in Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 422, + 257, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 257, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 257, + 434 + ], + "type": "text", + "content": "4.3 Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 443, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 564 + ], + "type": "text", + "content": "We adopt Greedy-Hungarian [15] and Strict-Hungarian [52] evaluation protocols by following [14]. During testing, the class descriptor from the hash code directly forms clusters. Then, we match these clusters with ground truth using the Hungarian matching algorithm [32]. Clusters that are not matched are regarded as misclassified. In Greedy-Hungarian, we divide the points into two groups, known and unknown classes, based on the ground truth and then perform Hungarian matching separately within each group. On the other hand, Strict-Hungarian conducts Hungarian matching for all clusters without distinguishing between known and unknown classes. The mean IoU (mIoU) is used for evaluating semantic segmentation performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 582, + 283, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 283, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 283, + 596 + ], + "type": "text", + "content": "5 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 609, + 222, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 222, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 222, + 620 + ], + "type": "text", + "content": "5.1 Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Quantitative Results. We conduct comparative experiments with baselines on all proposed scenarios. We report the results on scenario-A in Tab. 2, while the results on scenario-B and scenario-C are reported in Tab. 3. As shown in Tables. 2" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 480, + 328 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: On-the-fly category discovery results on SemanticPOSS (left) and SemanticKITTI (right) within Scenario-A. Best in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 480, + 328 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 480, + 328 + ], + "type": "table", + "html": "
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
SplitMethodUnknownKnownAllUnknownKnownAllSplitMethodUnknownKnownAllUnknownKnownAll
POSS0Baseline9.2028.5922.6311.4132.7126.16KITTI0Baseline10.4725.3421.4313.2230.2825.79
MLDG6.3334.6925.967.5540.5430.39MLDG12.6523.1720.4020.5927.3425.56
SMILE5.7940.3429.719.0556.5241.91SMILE9.9428.0923.3121.6638.9634.41
Ours17.8849.6039.8420.3056.5445.39Ours22.5545.9139.7732.9355.1249.28
POSS1Baseline13.2123.1620.8720.0130.1127.78KITTI1Baseline13.5123.7221.0420.4625.1323.90
MLDG15.7228.9725.9123.6635.6532.89MLDG8.9928.8723.6412.7129.7625.27
SMILE15.2032.4728.4818.2246.4039.89SMILE6.5936.6128.7122.6343.3237.88
Ours26.6844.1140.0939.2850.4747.89Ours22.2338.7034.3636.2642.6440.96
POSS2Baseline21.0319.8220.1028.5127.1527.46KITTI2Baseline10.0326.7522.3512.5428.4524.27
MLDG10.0430.1725.5219.8233.4030.27MLDG9.5030.1524.7213.5831.3326.66
SMILE9.0031.6426.4225.2738.2435.24SMILE11.1231.0925.8425.7539.7936.10
Ours24.5241.4037.5139.5946.7145.07Ours14.9838.3932.2323.3942.9037.77
POSS3Baseline4.4728.9223.2818.1228.6326.21KITTI3Baseline12.9326.1223.3421.6927.2726.10
MLDG2.2433.1226.0015.4333.9629.68MLDG7.8025.4421.7216.0027.5125.09
SMILE1.8936.6028.5924.0437.3834.30SMILE19.2130.9728.4932.7336.6035.79
Ours17.0951.3943.4849.1952.5451.77Ours17.7445.1539.3827.9447.8943.69
AvgBaseline11.9825.1221.7219.5129.6526.90AvgBaseline11.7425.4822.0416.9827.7825.02
MLDG8.5831.7425.8516.6235.8930.81MLDG9.7426.9122.6215.7228.9925.65
SMILE7.9735.2628.3019.1544.6437.84SMILE11.7231.6926.5925.6939.6736.05
Ours21.5446.6340.2337.0951.5747.53Ours19.3842.0436.4430.1347.1442.93
", + "image_path": "87c6fe5e51c432a0c1fe242483d50cc8c0f257915f7e0b6910f8d91b03b12ba0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 402, + 479, + 460 + ], + "blocks": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "text", + "content": "Table 3: On-the-fly category discovery results on SemanticKITTI within Scenario-B (left) and Scenario-C (right). Best in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 402, + 479, + 460 + ], + "lines": [ + { + "bbox": [ + 138, + 402, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 479, + 460 + ], + "type": "table", + "html": "
Strict-Hungarian (%)Greedy-Hungarian (%)Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAllMethodUnknownKnownAllUnknownKnownAll
Baseline14.6324.8923.8133.6425.9826.79Baseline11.7730.6127.6319.1431.0729.19
MLDG25.8025.6825.7040.7525.4927.10MLDG11.8033.7730.3029.4734.7033.87
SMILE13.1031.6129.6747.6732.5534.14SMILE11.0841.7436.9051.8643.8945.15
Ours21.7347.8345.0842.1248.9648.24Ours16.6647.2942.4537.1947.7646.09
", + "image_path": "4b473f90f16e1d694b039d9763398495b41b1322bce980abbb8342107860f516.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "and 3, in all scenarios, the baseline model exhibits diminished recognition ability for known classes due to its sensitive representation in hash coding, while the MLDG and SMILE demonstrate improved recognition capabilities compared to the baseline. Compared to all baselines, the proposed method achieves superior performance for known classes, demonstrating the effectiveness of the proposed representation learning for OCDSS, including DPL and MCL. Furthermore, the proposed method's comparative performance in the known classes compared to the oracle (All: " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "46.16\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in SemanticPOSS, All: " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "49.23\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in SemanticKITTI) indicates its superior recognition ability. MLDG and SMILE exhibit lower results for unknown classes in Strict-Hungarian except for MLDG in scenario-B. This implies that these models tend to become overly focused on the semantics of known classes, making them particularly confused when faced with unknown classes. MLDG performing well for unknown classes in scenario-B is speculated to be the simulation of unknown classes during training, which enhances gen" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 107, + 482, + 257 + ], + "blocks": [ + { + "bbox": [ + 133, + 107, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 133, + 107, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 107, + 482, + 257 + ], + "type": "image", + "image_path": "64fc7562fadf2f91ab337567778f5c5ed39e692822014c2410a5ac56e9c3866a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 263, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 132, + 263, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 482, + 285 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison results on " + }, + { + "bbox": [ + 132, + 263, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS}_3" + }, + { + "bbox": [ + 132, + 263, + 482, + 285 + ], + "type": "text", + "content": " within scenario-A (top), scenario-B (middle), and scenario-C (bottom). Best viewed when zoomed in with colors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 311, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 418 + ], + "type": "text", + "content": "eralization to unobserved-unknown classes. However, this approach offers only limited performance improvements for known classes. In the case of SMILE, its favorable performance on unknown classes in the Greedy-Hungarian suggests that it understands the semantics of unknown classes well. Yet, the poor results in the Strict-Hungarian imply that SMILE frequently confuses known and unknown classes. Overall, the proposed method outperforms baselines in both known and unknown classes in scenario-A and scenario-C. In scenario-B as well, our proposed method consistently outperforms in all classes average, further validating the efficacy of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 418, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 515 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 3, we visualize the qualitative comparison results on scenario-A (top), scenario-B (middle), and scenario-C (bottom). While MLDG and SMILE fail to predict the known and unknown classes correctly, the proposed method not only successfully recognizes the known classes but also cognizes the unknown classes (top: plant/car, middle: terrain/other-vehicle, bottom: vegetation/bicyclist). This implies that the proposed method learns the better representations for OCDSS through the DPL and MCL. Additional qualitative results are provided in Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 534, + 246, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 246, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 246, + 545 + ], + "type": "text", + "content": "5.2 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed method, we conduct an ablation study as depicted in Fig. 4. Here, we set the baseline as the model introduced in Sec. 3.2. The experiments are conducted on the SemanticPOSS within scenario-A, and the results are averaged over four splits. Please refer to the Supplementary Material for the discussion with class-wise results and analysis on the hash code dimension " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": ", mixing ratio " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": ", weight " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": ", and sub-sampling in GM. DPL helps to recognize known classes. Comparing the baseline and baseline+DPL, we observe a significant improvement in performance for the known classes. When the model is trained only with the baseline, hash codes from the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 288, + 177 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 288, + 177 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 288, + 177 + ], + "type": "image", + "image_path": "d57aa633c0bbdfcd46498159525d6ba95e96d08b74271956cfea8f989217ee6f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 179, + 218, + 190 + ], + "lines": [ + { + "bbox": [ + 206, + 179, + 218, + 190 + ], + "spans": [ + { + "bbox": [ + 206, + 179, + 218, + 190 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 198, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 479, + 220 + ], + "type": "text", + "content": "Fig. 4: (a) Results of using dual prototypical learning (DPL) with the baseline and SMILE. (b) Ablation study of the proposed method. Base refers to the baseline model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 291, + 130, + 480, + 178 + ], + "blocks": [ + { + "bbox": [ + 291, + 130, + 480, + 178 + ], + "lines": [ + { + "bbox": [ + 291, + 130, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 291, + 130, + 480, + 178 + ], + "type": "table", + "html": "
Strict-Hungarian (%)Greedy-Hungarian (%)
MethodUnknownKnownAllUnknownKnownAll
(a) Base11.9825.1221.7219.5129.6526.90
(b) Base+DPL5.3340.0831.3317.8946.4338.96
(c) Base+DPL+DM16.4848.2440.2031.3954.8348.58
(d) Base+DPL+DM+GM21.5446.6340.3637.0951.5747.53
", + "image_path": "3a4a916bc3b7550ff31d974354b221682773eb45c9bb87f355847e456ba3e3f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 378, + 179, + 391, + 189 + ], + "lines": [ + { + "bbox": [ + 378, + 179, + 391, + 189 + ], + "spans": [ + { + "bbox": [ + 378, + 179, + 391, + 189 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 245, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 376 + ], + "type": "text", + "content": "learned representations are sensitive to intra-class variation, which is supported by the poor performance for known classes. Conversely, adding hash prototypes through DPL imposes additional constraints on the representation, making it more suitable for hash code-based clustering. To further analyze the effects of DPL, we conduct an experiment by integrating the DPL with the baseline and SMILE [14]. As shown in Fig. 4a, when DPL is used together, the performance for known classes of both baseline and SMILE is significantly improved in both Strict- and Greedy-Hungarian. This implies that the proposed DPL effectively improves the robustness of learned representation under intra-class variation. The proposed DPL benefits both baseline and SMILE, demonstrating its effectiveness in a hash-code-based framework for addressing the OCDSS problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "content": "MCL helps to cognize unknown classes. To excavate the ability to discover unknown classes, we proposed mixing-based category learning based on DPL. With DM, the prototypes become discriminative for ID and OOD. This enhancement leads to improved performance not only for known classes but also for unknown classes. The proposed GM further boosts the performance of unknown classes. Comparing the results of using (c) DM and (d) DM+GM in Fig. 4b, we observe approximately " + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "content": " gains in Strict- and Greedy-Hungarian, respectively. Although performance on known classes experiences a slight reduction, considering the inherent challenge of distinguishing unknown classes that the model has not seen during training, this underscores the necessity of GM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we introduce a new problem named on-the-fly category discovery for LiDAR semantic segmentation. Along with the baseline model based on hash coding, we propose dual prototypical learning to enhance the model's recognition ability. We further design the mixing-based category learning to inductively transfer recognition knowledge from known classes to unknown classes. Extensive experiments demonstrate the effectiveness of the proposed method for recognizing known classes and discovering unknown classes. While the proposed method effectively addresses the problem, performance disparities between the known and unknown classes remain. In this respect, exploring suitable representation learning for discovering unknown classes presents a promising research direction." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "content": "This work was supported by the Technology Innovation Program (1415187329, 20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 239, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 239, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 273 + ], + "type": "text", + "content": "1. Alonso, I., Riazuelo, L., Montesano, L., Murillo, A.C.: 3d-mininet: Learning a 2d representation from point clouds for fast and efficient 3d lidar semantic segmentation. IEEE Robotics and Automation Letters 5(4), 5432-5439 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 274, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 316 + ], + "type": "text", + "content": "2. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5240-5250 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "text", + "content": "3. Basart, S., Mantas, M., Mohammadreza, M., Jacob, S., Dawn, S.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 350, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 392 + ], + "type": "text", + "content": "4. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 393, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 415 + ], + "type": "text", + "content": "5. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "type": "text", + "content": "6. Cao, K., Brbic, M., Leskovec, J.: Open-world semi-supervised learning. In: International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 468 + ], + "type": "text", + "content": "7. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 469, + 480, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 513 + ], + "type": "text", + "content": "8. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 318-334 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "type": "text", + "content": "9. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 545, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 480, + 589 + ], + "type": "text", + "content": "10. Chen, K., Oldja, R., Smolyanskiy, N., Birchfield, S., Popov, A., Wehr, D., Eden, I., Pehserl, J.: Mvlidarnet: Real-time multi-class scene understanding for autonomous driving using multiple views. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 2288-2294. IEEE (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 632 + ], + "type": "text", + "content": "11. Cheng, R., Razani, R., Taghavi, E., Li, E., Liu, B.: 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12547-12556 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "12. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "13. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: International Symposium on Visual Computing. pp. 207-222. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "14. Du, R., Chang, D., Liang, K., Hospedales, T., Song, Y.Z., Ma, Z.: On-the-fly category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11691-11700 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "type": "text", + "content": "15. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "text", + "content": "16. Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. pp. 1050-1059. PMLR (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 293 + ], + "type": "text", + "content": "17. Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Navab, N., Busam, B., Tombari, F.: Segmenting known objects and unseen unknowns without prior knowledge. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 19321-19332 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "18. Ge, Z., Demyanov, S., Chen, Z., Garnavi, R.: Generative openmax for multi-class open set classification. In: British Machine Vision Conference. BMVA Press (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 316, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 316, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 316, + 480, + 358 + ], + "type": "text", + "content": "19. Gerdzhev, M., Razani, R., Taghavi, E., Bingbing, L.: Tornado-net: multiview total variation semantic segmentation with diamond inception module. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9543-9549. IEEE (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "20. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "21. Gu, P., Zhang, C., Xu, R., He, X.: Class-relation knowledge distillation for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16474-16483 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 457 + ], + "type": "text", + "content": "22. Guo, Y., Camporese, G., Yang, W., Sperduti, A., Ballan, L.: Conditional variational capsule network for open set recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 103-111 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "text", + "content": "23. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "24. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "text", + "content": "26. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11108-11117 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "27. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "28. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "30. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 214 + ], + "type": "text", + "content": "31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "text", + "content": "32. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": "33. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 300 + ], + "type": "text", + "content": "34. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 481, + 332 + ], + "type": "text", + "content": "35. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9425-9434 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 332, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 481, + 354 + ], + "type": "text", + "content": "36. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "type": "text", + "content": "37. Lis, K., Nakka, K., Fua, P., Salzmann, M.: Detecting the unexpected via image resynthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2152-2161 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "text", + "content": "38. Lu, T., Ding, X., Liu, H., Wu, G., Wang, L.: Link: Linear kernel for lidar-based 3d perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1105-1115 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 450 + ], + "type": "text", + "content": "39. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 483 + ], + "type": "text", + "content": "40. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 613-628 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 483, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 515 + ], + "type": "text", + "content": "41. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 515, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 481, + 548 + ], + "type": "text", + "content": "42. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693. IEEE (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 548, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 481, + 579 + ], + "type": "text", + "content": "43. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 481, + 611 + ], + "type": "text", + "content": "44. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 643 + ], + "type": "text", + "content": "45. Riz, L., Saltori, C., Ricci, E., Poiesi, F.: Novel class discovery for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9393-9402 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "text", + "content": "46. Ruder, S.: An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "47. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 480, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 480, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 480, + 173 + ], + "type": "text", + "content": "48. Shu, Y., Shi, Y., Wang, Y., Huang, T., Tian, Y.: P-odn: Prototype-based open deep network for open set recognition. Scientific reports 10(1), 7146 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "text", + "content": "49. Sun, X., Yang, Z., Zhang, C., Ling, K.V., Peng, G.: Conditional gaussian distribution learning for open set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13480-13489 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "text", + "content": "50. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European conference on computer vision. pp. 685-702. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "51. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 275, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 480, + 308 + ], + "type": "text", + "content": "52. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 342 + ], + "type": "text", + "content": "53. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 419 + ], + "type": "text", + "content": "55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 1887-1893. IEEE (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 464 + ], + "type": "text", + "content": "56. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 4376-4382. IEEE (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 465, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 480, + 508 + ], + "type": "text", + "content": "57. Xia, Y., Zhang, Y., Liu, F., Shen, W., Yuille, A.L.: Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 145-161. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 553 + ], + "type": "text", + "content": "58. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "text", + "content": "59. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "60. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "61. Zhao, Y., Zhong, Z., Sebe, N., Lee, G.H.: Novel class discovery in semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4340-4349 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "H. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 293 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "62. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "63. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "64. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "text", + "content": "65. Zhou, H., Zhu, X., Song, X., Ma, Y., Wang, Z., Li, H., Lin, D.: Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "66. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 179, + 91, + 447, + 102 + ], + "type": "text", + "content": "On-the-fly Category Discovery for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_content_list.json b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18417a245cf82de84ca32191923588fd49e89e1a --- /dev/null +++ b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gang Dai $^{1\\star}$ , Yifan Zhang $^{2,3\\star}$ , Quhui Ke $^{1}$ , Qiangya Guo $^{1}$ , and Shuangping Huang $^{1,4\\dagger}$", + "bbox": [ + 287, + 210, + 712, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ South China University of Technology \n $^{2}$ National University of Singapore, $^{3}$ Skywork AI, $^{4}$ Pazhou Laboratory \n $\\{\\mathrm{eedaigang}@mail., \\mathrm{eehsp@}\\}$ scut.edu.cn, yifan.zhang@u.nus.edu", + "bbox": [ + 267, + 253, + 730, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing handwritten text generation methods often require more than ten handwriting samples as style references. However, in practical applications, users tend to prefer a handwriting generation model that operates with just a single reference sample for its convenience and efficiency. This approach, known as \"one-shot generation\", significantly simplifies the process but poses a significant challenge due to the difficulty of accurately capturing a writer's style from a single sample, especially when extracting fine details from the characters' edges amidst sparse foreground and undesired background noise. To address this problem, we propose a One-shot Diffusion Mimicker (One-DM) to generate handwritten text that can mimic any calligraphic style with only one reference sample. Inspired by the fact that high-frequency information of the individual sample often contains distinct style patterns (e.g., character slant and letter joining), we develop a novel style-enhanced module to improve the style extraction by incorporating high-frequency components from a single sample. We then fuse the style features with the text content as a merged condition for guiding the diffusion model to produce high-quality handwritten text images. Extensive experiments demonstrate that our method can successfully generate handwriting scripts with just one sample reference in multiple languages, even outperforming previous methods using over ten samples. Our source code is available at https://github.com/dailenson/One-DM.", + "bbox": [ + 261, + 334, + 738, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Handwritten Text Generation $\\cdot$ One-Shot Generation", + "bbox": [ + 261, + 652, + 702, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 691, + 374, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the digital age, handwriting text generation blends the personalization of traditional handwriting with the efficiency of automated processes, offering a digital format to preserve the authenticity of individual handwriting. This task aims to automatically generate the desired handwritten text images that not only correspond to specific text content, but also emulate the calligraphic style of a given exemplar writer (e.g., character slant, cursive join, stroke thickness, and", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Authors contributed equally; † Corresponding author", + "bbox": [ + 217, + 824, + 588, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c0fe1ae5b0af618aa0a891aa277118b2e2e3c2328b433d0772e64f21898650a.jpg", + "image_caption": [ + "Fig. 1: User experience comparisons between one-shot and few-shot handwritten text generation methods. It reveals that one-shot setting leads to a better user experience." + ], + "image_footnote": [], + "bbox": [ + 248, + 143, + 493, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3168de4542c0ee588ddfeb45d78150b9050c1ccf0605bd8be842fa719533cf59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 143, + 759, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ink color). This provides great convenience for people with hand impairments and also contributes to accelerating the process of handwritten font design.", + "bbox": [ + 212, + 325, + 782, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works [4,8,11-13,24,39] introduce generative adversarial networks (GANs) [5,14,32] for handwritten text generation. For instance, ScrabbleGAN [11] leverages random noises as style inputs and conditions content inputs on character-level labels, enabling the synthesis of handwritten words with randomly sampled styles. In a recent study [9], diffusion models [31,42,53] such as denoising diffusion probabilistic models (DDPM) [18] have demonstrated even higher quality of image generation compared to GANs. This motivates several attempts, such as WordStylist [37], GC-DDPM [10], and CTIG-DM [57], to condition the denoising process on the fixed writer ID to generate handwritten text images with controllable styles. However, the major limitation of these methods [10,11,37,57] is that they are unable to mimic the unseen writers' handwriting styles.", + "bbox": [ + 212, + 357, + 792, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To imitate any given handwriting style, some previous methods [4, 24, 39], known as few-shot generation, require users to provide a few samples (typically 15) as style references. They employ a style encoder to extract writing styles from the given samples, thus offering flexible control over the style of the generated samples. However, the traditional few-shot generation pipeline is inconvenient, inefficient, and time-consuming, as shown in Figure 1. Users prefer methods that only require a single sample as a style reference, known as one-shot generation methods, because they are more convenient, efficient, and easy-to-use. Our goal is to investigate the more challenging one-shot generation task that holds significant practical value, striving to achieve high-quality handwritten text image generation with desired styles and contents.", + "bbox": [ + 212, + 523, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary challenge of one-shot generation task lies in accurately extracting a user's handwriting style from just one style reference image. As illustrated in Figure 2, characters, being abstract symbols, typically occupy only a minor portion of the reference image. Besides, the reference image is often cluttered with noisy background information, which poses a significant obstacle in extracting the individual handwriting style. Previous one-shot generation methods [8,12,13] simply follow the architectural design of few-shot approaches, using a vanilla CNN encoder to directly extract the handwriting style from a single sample. The extracted style is then combined with text contents and input into a CNN decoder to generate the desired handwritten images. These methods ex", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/702e2806fd347d3dcba7f96df7856799792949a5d80b59480d0938cffbaff0b0.jpg", + "image_caption": [ + "Fig. 2: Handwritten text samples and corresponding high-frequency components. We find that high-frequency components have more pronounced character contours, clearly showcasing the style patterns, such as character slant and cursive connections." + ], + "image_footnote": [], + "bbox": [ + 305, + 150, + 707, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hibit a limited performance in emulating handwriting styles, due to their poor ability of style extraction.", + "bbox": [ + 212, + 340, + 782, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above challenges, our key idea revolves around utilizing the high-frequency information of the sample to enhance the extraction of the handwriting style. As depicted in Figure 2, the high-frequency information encompasses the overall contours of the handwritten text, allowing for a clear observation of key style patterns such as the text slant, letter spacing, and cursive connections. Hence, the incorporation of high-frequency information facilitates a more effective handwriting style extraction.", + "bbox": [ + 212, + 371, + 784, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In light of the above insight, we propose a One-shot Diffusion Mimicker (One-DM) for handwritten text generation, which is simultaneously guided by desired style and arbitrary content. More specifically, we first develop a style-enhanced module to process both a style reference image and its high-frequency components in parallel. Considering that the reference image often contains background noise, we design a gate mechanism to suppress the inflow of background noise. Regarding the high-frequency components, such as character slant and ligature that show clearer style patterns, we employ a contrastive learning framework [25,54,55] to further obtain discriminative style features, guiding the handwritten text synthesis with both realistic and diverse styles. Subsequently, the style features extracted from both branches are adaptively fused with the specific content prototype in a style-content fusion module. Finally, the seamlessly integrated style and content features act as conditional inputs, guiding the denoising process for the progressive synthesis of stylized handwritten text images. It is worth noting that our One-DM effectively mimics a user's writing style with only one reference sample through our style-enhanced module, surpassing few-shot methods in producing higher-quality stylized handwritings.", + "bbox": [ + 212, + 478, + 785, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sum up, our contributions are as follows:", + "bbox": [ + 240, + 734, + 558, + 750 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel diffusion model for stylized handwritten text generation, which only requires a single reference sample as style input, and imitates its writing style to generate handwritten text with arbitrary content.", + "- We introduce the high-frequency components of the reference sample to enhance the extraction of handwriting style. The proposed style-enhanced" + ], + "bbox": [ + 225, + 763, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "module can effectively capture the writing style patterns and suppress the interference of background noise.", + "bbox": [ + 241, + 146, + 785, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Extensive experiments on handwriting datasets in English, Chinese, and Japanese demonstrate that our approach with a single style reference even outperforms previous methods with 15x-more references.", + "bbox": [ + 225, + 176, + 785, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 247, + 387, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Handwriting generation. Handwritten text is typically stored in two formats: online trajectory form or offline image form. Online handwriting generation methods [1, 7, 28, 44, 46, 56] often employ Recurrent Neural Networks (RNNs) [1, 6, 28, 44-46, 56], transformer decoders [7], or diffusion models [34, 40] to progressively generate writing trajectories. Unlike online methods, offline generation methods have the advantage of not requiring additional trajectory supervision information and can generate realistic handwritten text with stroke width and ink color, which online methods cannot produce.", + "bbox": [ + 212, + 280, + 784, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous offline handwriting generation methods [2, 4, 8, 11, 20, 24, 35, 39] in deep learning predominantly rely on Generative Adversarial Networks (GANs). Early works [2, 11] condition the generative process on the word embeddings [2] or concatenated letter-tokens [11] to synthesize handwritten word images. However, these methods struggle to flexibly control the writing style. Thus, few-shot approaches [4, 24, 39] which rely on 15 style references are introduced. For instance, GANwriting [24] utilizes a CNN encoder [16, 50, 51] to extract a user's handwriting style from a few samples, which are then combined with specific text content to generate handwritings in the desired style. In its follow-up work [23], the synthesized samples are demonstrated to assist in training more robust handwritten text recognizers. Further, HWT [4] employs a transformer encoder [29, 47] to extract rich style patterns from reference samples, enhancing the performance in style mimicry. Recently, VATr [39] uses images of symbols as content representations, enabling the generation of out-of-charset characters.", + "bbox": [ + 212, + 402, + 785, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concurrently, one-shot generation methods [8, 12, 13] are proposed. Despite their ability to mimic handwriting styles with only a single sample, these techniques still lag behind few-shot methods in terms of stylized generation results. Additionally, previous handwriting generation methods [2, 4, 8, 11-13, 24] often rely on complex content representations, such as recurrent embeddings [2] and letter-level tokens [4, 8, 11-13, 24]. SLOGAN [35] extracts textual contents from easily obtainable printed images. However, it faces challenges in generalizing to unseen writing styles due to its fixed writer ID. Similarly, some diffusion-based [10, 37, 57] methods condition the denoising process on fixed style ID and are unable to mimic styles that they have not previously encountered. In contrast, our One-DM effectively obtains style information from one style sample and thus can generate handwritings with arbitrary styles. Due to space constraints, we discuss more related works in supplementary, including diffusion methods for general image generation, contrastive-based method [48], and frequency-based methods [30, 38].", + "bbox": [ + 212, + 613, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c2a04fd867a48f0f68059db58c5c9dea90a83773872be2b1f5435c6398ca6766.jpg", + "image_caption": [ + "Fig. 3: Overview of the proposed method. The style reference initially passes through a high-pass filter to extract its high-frequency components. Subsequently, the spatial and the high-frequency style encoders independently extract style features $F_{spa}$ and $F_{fre}$ from the style reference and its high-frequency information, respectively. $F_{spa}$ , after being filtered through a gate mechanism, is fused with $F_{fre}$ and content features $E$ in the fusion module. The merged feature then serves as a condition input to guide the diffusion generation process." + ], + "image_footnote": [], + "bbox": [ + 246, + 135, + 767, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 412, + 328, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem statement. We aim to synthesize handwritten text images that are controlled by both textual content and handwriting style. Given any word string $\\mathcal{A}$ and a single style sample $I_{s}$ from a writer $w_{s}$ , the generated handwritten word image $X_{s}$ should replicate the unique handwriting style of $w_{s}$ while maintaining the content of $\\mathcal{A}$ , where the textual content $\\mathcal{A} = \\{a_i\\}_{i=1}^L$ spans a length $L$ and $a_{i}$ is from a character set without any restrictions.", + "bbox": [ + 212, + 449, + 787, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this task, we incorporate high-frequency information to enhance the extraction of writing styles. Existing methods [4,8,12,13,24,39] typically use a vanilla CNN or transformer encoder to directly process style images, often resulting in undesired style extraction. In contrast, we introduce a novel One-shot diffusion mimicker (One-DM) by innovating Laplacian high-frequency extraction and a gating mechanism. Our One-DM can effectively capture style features from a single reference while suppressing background noise.", + "bbox": [ + 212, + 541, + 787, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overall Scheme", + "text_level": 1, + "bbox": [ + 215, + 674, + 390, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our high-level idea focuses on incorporating high-frequency information from style reference images to enhance the extraction of style patterns. A straightforward implementation involves using a vanilla transformer encoder to extract style features from both the style image and its corresponding high-frequency image. This naive solution encounters two main issues: (1) the lack of efficient supervision objective still makes it challenging to accurately learn the writer's style patterns from the high-frequency image, and (2) the style features captured from the original image still retain undesirable noise backgrounds, which can adversely affect subsequent image generation performance.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the above-mentioned issues, we develop a more effective method, shown in Figure 3. Our method consists of a style-enhanced module, a content encoder, a style-content fusion module, and a conditional diffusion module. Initially, we utilize the Laplacian kernel as a high-frequency filter to extract the high-frequency components from the style reference. Then, two parallel style encoders are employed to simultaneously extract the corresponding style features from both style reference and its high-frequency information. Since undesirable background noise is often present in style references, we design a gate mechanism to facilitate the transmission of informative style information while mitigating noise transmission. The style patterns in the high-frequency components are comparatively cleaner and more distinct, facilitating the observation of individual styles, such as character slant. For the observation, we propose a contrastive learning objective, termed LaplacianNCE $\\mathcal{L}_{lapNCE}$ , to enforce the more discriminative style learning from high-frequency components.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding content guidance, we render the given string $\\mathcal{A}$ into a unifont image, as done in VATr [39]. Briefly, Unifont's key advantage is that it covers all Unicode characters, allowing our method to transform any user inputs into corresponding images. We further feed render results into a content encoder that combines a ResNet18 [16] with a Transformer encoder. This process first involves using a ResNet18 to handle each character image in parallel, followed by concatenating them to form word sequence features. The Transformer encoder then processes these features to extract an informative content feature $E = \\{e_i\\}_{i=1}^L \\in \\mathbb{R}^{L \\times c}$ with a global context, where $c$ is the channel dimension. After obtaining style features and content guidance, we seamlessly fusion them using a style-content fusion module. The fused results then guide the denoising process of the conditional diffusion model to synthesize the desired handwritten text images. The denoising process is supervised by the reconstruction loss $\\mathcal{L}_{rec}$ .", + "bbox": [ + 212, + 358, + 787, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To summarize, the overall training objective of our method combines all two loss functions:", + "bbox": [ + 212, + 554, + 785, + 582 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {l a p N C E} + \\mathcal {L} _ {r e c}.\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 584, + 578, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Style-enhanced Module", + "text_level": 1, + "bbox": [ + 214, + 619, + 455, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose the style-enhanced module to enhance style extraction by incorporating high-frequency components $H_{s} \\in \\mathbb{R}^{h \\times w \\times c}$ from a reference image $I_{s} \\in \\mathbb{R}^{h \\times w \\times c}$ since clearer style patterns are presented in high-frequency components like character slanting and shape connections (cf. Figure 2). As shown in Figure 3, we use a Laplacian kernel as a high-frequency filter to extract $H_{s}$ from $I_{s}$ . The Laplacian kernel excels in extracting high-frequency information without the need for Fast Fourier Transform (FFT) and parametric separation in the frequency domain. Then, two style encoders, $\\mathcal{E}_{spa}$ and $\\mathcal{E}_{fre}$ , each a combination of CNN and transformer, process the $I_{s}$ and $H_{s}$ , respectively. This independent processing leads to the extraction of distinct style features: $F_{spa} = \\{f_{spa}^{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d \\times c}$ from $\\mathcal{E}_{spa}$ and $F_{fre} = \\{f_{fre}^{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d \\times c}$ from $\\mathcal{E}_{fre}$ , where $d = h \\times w$ . While structurally identical, $\\mathcal{E}_{spa}$ and $\\mathcal{E}_{fre}$ do not share weights with each other. Then, the proposed $\\mathcal{L}_{lapNCE}$ forces $\\mathcal{E}_{fre}$ to focus on extracting discriminative style features", + "bbox": [ + 212, + 642, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from $H_{s}$ . A gate mechanism is designed to selectively filter out background noise from the reference style features, allowing only meaningful style patterns to pass.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Laplacian Contrastive Learning. The goal of the proposed $\\mathcal{L}_{\\text{lapNCE}}$ is to guide the high-frequency style encoder $\\mathcal{E}_{\\text{fre}}$ in learning more discriminative style features from the high-frequency information. Thus, we propose to bring closer extracted style features $F_{\\text{fre}}$ belonging to the same writer, while distancing those from different writers. We formulate our $\\mathcal{L}_{\\text{lapNCE}}$ as follows:", + "bbox": [ + 212, + 176, + 784, + 253 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a p N C E}} = - \\frac {1}{N} \\sum_ {i \\in M} \\frac {1}{| P (i) |} \\sum_ {p \\in P (i)} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} / \\tau\\right)}{\\sum_ {a \\in A (i)} \\exp \\left(z _ {i} \\cdot z _ {a} / \\tau\\right)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 263, + 784, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In detail, $i \\in M = \\{1 \\dots N\\}$ is the index of any element in the mini-batch with size $N$ and $A(i) = M \\setminus \\{i\\}$ is other indices distinct from $i$ . $z_{i}$ is an anchor sample belonging to writer $w_{i}$ and $P(i) = \\{p \\in A(i) : w_{p} = w_{i}\\}$ is its in-batch positive sample set and the other $A(i) \\setminus P(i)$ is its negative set. Here, $z = Proj(F_{fre})$ , where $Proj$ is a learnable multi-layer perceptron (MLP), $\\tau$ is a scalar temperature parameter and the $\\cdot$ symbol denotes the inner product.", + "bbox": [ + 212, + 309, + 782, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gate mechanism. As illustrated in Figure 2, the stroke areas of characters in reference images are typically sparse, with background noise interfering with the extraction of character style features. To address this challenge, we propose a gate mechanism to selectively filter the information of the sample $I_{s}$ , as shown in Figure 3. Specifically, the extracted sample style features $F_{spa}$ are fed into a gate layer, consisting of a learnable fully connected layer and followed by a sigmoid activation, to obtain the corresponding gate units $W = \\{w_{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d}$ . Each unit $w_{i}$ determines the pass rate for the corresponding $f_{spa}^{i}$ , allowing for a higher pass rate where $w_{i}$ is larger. This design effectively enables informative style features $\\hat{F}_{spa} = \\{\\hat{f}_{spa}^{i}\\}_{i=1}^{d}$ to be extracted while suppressing extraneous background noise, where $\\hat{f}_{spa}^{i} = f_{spa}^{i} \\cdot w_{i}$ .", + "bbox": [ + 212, + 400, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Style-content Fusion Module", + "text_level": 1, + "bbox": [ + 214, + 592, + 498, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Upon acquiring textual content feature $E$ and two style features $\\hat{F}_{spa}$ and $F_{fre}$ , we integrate all features within two multi-head attention mechanisms to guide the denoising generation process of the diffusion model, as shown in Figure 3. Specifically, the first cross-attention module uses the textual content $E$ as queries to identify the most relevant style information in the style reference, thereby inferring the style attributes corresponding to each character. For instance, if the textual content is 'a', it prioritizes searching for style features of characters like 'a', 'b', 'd', 'g' in the style reference, due to these characters appearing similar looped structures, implying more comparable style attributes. This process (cross-attention in Figure 3) is represented as:", + "bbox": [ + 212, + 617, + 782, + 768 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nO = \\operatorname {A t t e n} _ {1} \\left(Q _ {1} = E, K _ {1} = V _ {1} = \\hat {F} _ {s p a} + F _ {f r e}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 781, + 784, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subsequently, we obtain the initial fusion embedding between content and style guidance by simply summing $O$ and $E$ . The merged intermediate vector is then", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "employed as the query, key, and value in the self-attention mechanism to facilitate comprehensive interaction of information. Finally, the blended embedding $g$ serves as the condition of the diffusion process. The second multi-head attention (self-attention in Figure 3) is defined as:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ng = \\operatorname {A t t e n} _ {2} \\left(Q _ {2} = K _ {2} = V _ {2} = O + E\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 219, + 785, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Conditional Diffusion Model", + "text_level": 1, + "bbox": [ + 215, + 255, + 496, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The goal of the conditional diffusion model $p_{\\theta}$ is to generate realistic images of handwritten text, guided by acquired conditions $g$ . Specifically, as shown in Figure 3, under the guidance of $g$ , $p_{\\theta}$ executes a denoising generative process with $T$ steps, starting from a sampled Gaussian noise $x_{T}$ and progressively denoises to obtain the desired handwritten text $x_{0}$ :", + "bbox": [ + 212, + 280, + 787, + 356 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0} \\mid g\\right) = \\int p _ {\\theta} \\left(x _ {0: T} \\mid g\\right) d _ {x _ {1: T}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 366, + 785, + 396 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0: T} \\mid g\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, g\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 412, + 785, + 452 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, g\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, g, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, g, t\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 463, + 784, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The denoising process aims to learn how to reverse a predefined forward process, as described in DDPM [18]. The forward process is modeled as a fixed Markov chain, where noise conforming to a normal distribution is incrementally added to $x_{t - 1}$ to derive $x_{t}$ . This can be mathematically expressed as:", + "bbox": [ + 212, + 484, + 787, + 546 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 558, + 785, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the noise is characterized by a variance schedule $\\beta_{t}$ . During training, a variational bound on the maximum likelihood objective is applied to guide the generation process to recover $x_0$ from the standard Gaussian noise $x_{T}$ conditioned on $g$ . We give the training objective as follows:", + "bbox": [ + 212, + 585, + 787, + 647 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {t, q} \\| \\mu_ {t} \\left(x _ {t}, x _ {0}\\right) - \\mu_ {\\theta} \\left(x _ {t}, g, t\\right) \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 657, + 785, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mu_t(x_t, x_0)$ is the mean of the forward process posterior $q(x_t | x_{t-1})$ , which has a closed-form solution [18]. We put more details in supplementary material.", + "bbox": [ + 212, + 685, + 787, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 738, + 375, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 768, + 447, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset. To evaluate our One-DM in handwritten text generation, we use the widely-used handwriting dataset IAM [36] and CVL [27].IAM consists of 62,857 English word images from 500 unique writers. Following [4,24,39], we use words", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from 339 writers for training and the remaining 161 writers for testing. CVL includes texts by 310 writers in English and German. We use the English portion, totaling 84,514 words, and follow CVL's standard split, with 283 writers for training and 27 for testing. Throughout all experiments, we resize images to 64 pixels in height, preserving their aspect ratio.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. We use the Fréchet Inception Distance (FID) [17] and Geometry Score (GS) [26] to assess the generation quality, following the settings of [4, 24, 39]. We also conduct user studies to quantify the subjective quality of the generated handwritten text images in supplementary.", + "bbox": [ + 212, + 220, + 784, + 281 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. In our all experiments, we only use single style reference sample. We first train our model for 700 epochs (batch size of 384) with classifier-free guidance strategy [19] under the guidance scale of 0.25, and then fine-tune the model for 4500 iterations (batch size of 128) using a text recognizer [41] with a CTC loss [15, 21, 58], on four RTX3090 GPUs. The fine-tuning process forces our One-DM to generate readable text with accurate content. The optimizer is AdamW [33], with a learning rate of $10^{-4}$ . During inference, each style sample is randomly sampled from the target-writer. To speed up the sampling, we use the denoising diffusion implicit model (DDIM) [43] with 50 steps. More details are provided in supplementary material.", + "bbox": [ + 212, + 282, + 785, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Compared methods. We compare our One-DM with state-of-the-art handwritten text generation methods, including GAN-based methods (i.e., GANwriting [24], HWT [4], VATr [39], TS-GAN [8], HiGAN+ [13]), and diffusion-based methods (i.e. GC-DDPM [10] and WordStylist [37]). For a fair comparison, we configure all methods to generate images with a height of 64 pixels in supplementary. Additionally, in supplementary, we compare a variant of One-DM that generates 32-pixel high images with official VATr [39] and HWT [4].", + "bbox": [ + 212, + 433, + 785, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 560, + 372, + 573 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stylish Handwritten Text Generation. Initially, we assess our One-DM for producing styled handwritten text images, aiming to replicate both the style and content in the generated images. Following [4, 24, 39], we first calculate FID between generated and real samples for each writer separately and finally average them. In line with the previous works [4, 24, 39], our experiments on the IAM dataset are divided into four distinct scenarios: IV-S, IV-U, OOV-S, OOV-U. Among these four scenarios, OOV-U represents the most challenging case where both the target style and the words are entirely unseen during training. For the CVL dataset, we directly report the results of all methods on the test set.", + "bbox": [ + 212, + 583, + 785, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first report the quantitative results on the IAM dataset in Table 1. We can observe that our One-DM outperforms all the competitors in all settings. Notably, it significantly exceeds one-shot methods in all scenarios. Impressively, our One-DM also holds a substantial advantage over few-shot methods (GAN-writing [24], HWT [4], VATr [39]), which use 15-x more reference samples for style guidance, in IV-S and OOV-S settings. Even in the most challenging OOV-U scenario, our One-DM leads the second-best, VATr, by a large margin (102.75 vs. 108.76), demonstrating the superior performance of our One-DM in stylized", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/33eea6d346ef1c0d8a872185fd1a49b0308e42bc4aac877f456cfc484a93eb82.jpg", + "table_caption": [ + "Table 1: Comparisons with state-of-the-art methods on styled and style-agnostic handwritten text generation in the IAM dataset. Note that, all methods are trained on the same training set used in GANwriting [24], HWT [4], and VATr [39]." + ], + "table_footnote": [], + "table_body": "
MethodShotStyloid EvaluationStyle-agnostic
IV-SIV-UOOV-SOOV-UFID↓GS↓
TS-GAN [8]One118.56128.75127.11136.6720.654.88×10-2
GANwriting [24]Few120.07124.30125.87130.6828.375.67×10-2
HiGAN+ [13]One117.33116.95121.55121.4822.952.06×10-2
GC-DDPM [10]One99.86105.73112.52118.3919.051.31×10-2
WordStylist [37]One98.10104.27109.45115.5218.582.85×10-2
HWT [4]Few109.25106.90116.55113.5218.994.41×10-3
VATr [39]Few103.75101.73111.64108.7616.031.74×10-2
Ours (One-DM)One89.4798.3693.30102.7515.731.98×10-3
", + "bbox": [ + 299, + 191, + 705, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "handwritten text generation. Similarly, our method outperforms HWT and VATr on the CVL dataset, achieving the lowest FID score, as shown in Table 4.", + "bbox": [ + 212, + 344, + 782, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We provide qualitative results to intuitively explain the benefit of our One-DM, in Figure 4. GANwriting struggles to capture the style patterns of reference samples, such as character slant and occasionally produces unclear character shapes. HiGAN+ more consistently generates characters with correct content, but the character spacing within the generated words lacks realism. WordStylist typically produces images with noticeable background noise. HWT and VATr can produce satisfactory handwritten words in terms of content accuracy and style mimicry; however, their downside is the tendency to create smoother character appearances. Compared to HWT and VATr, our synthesized samples excel in more authentic character ink color and stroke thickness. However, some generated samples by our One-DM are visibly different in ink color. We provide comprehensive explanations in supplementary. We further display more qualitative comparisons between our method and few-shot methods in Figure 5.", + "bbox": [ + 212, + 375, + 784, + 570 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Style-agnostic Handwritten Text Generation. We further evaluate our One-DM to generate realistic handwritten text images, irrespective of style imitation. For this purpose, we compute the FID and GS on the IAM test set, under the same conditions as applied to ScrabbleGAN [11] (FID: 20.72, GS: $2.56 \\times 10^{-2}$ ), capable of generating handwritten texts with randomly sampled styles. Specifically, each method generates 25k random samples to calculate FID against 25k samples from the test set, and 5k random samples for GS calculation in comparison with 5k test set samples. As presented in Table 1, our One-DM achieves the best results in both FID and GS metrics, further demonstrating its capability to generate higher-quality handwritten text images.", + "bbox": [ + 212, + 571, + 785, + 723 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Analysis", + "text_level": 1, + "bbox": [ + 214, + 742, + 333, + 757 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to analyze our One-DM. More analyses are provided in supplementary, including generalization evaluation on different style backgrounds, generation quality assessment through OCR performance, failure case analysis, and the effects of different designs (e.g., high-frequency filter, style-content fusion mechanism, and style input sample length).", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/15223f376dd427af09f5ef18083c9c37e518b62127ba366890e57a90b0165621.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Style examplesSee us next him unless some eight have the be is unassailable of our boxInour there they on as had inside are shaw have us God is unprouctand tuning which oil to you will the them may the rivers Rim Fishand to is all Common Wood Lessis of ther in work and it hean of
GANw.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*HiGAN+The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*WordS.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
HWTThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
VATrThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*OursThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
", + "bbox": [ + 259, + 143, + 748, + 366 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 4: Qualitative comparisons between our method with state-of-the-art methods on handwritten text generation with both specific textual content and desired handwriting style in the IAM dataset. We utilize the identical guiding text, 'The greatest test of courage on earth is to bear defeat without losing heart,' across all handwriting generation methods, directing them to produce text in varied styles. Better zoom in $200\\%$ . * denotes one-shot methods, while others are few-shot methods.", + "bbox": [ + 212, + 375, + 789, + 455 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/682a7fa689dbb818439afdebc95d3a073e71104b7f45e0cc40fdcee9ed6b1b3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Targethopetherelatersuccess
Generated Sampleshopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
", + "bbox": [ + 240, + 468, + 472, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 5: Each row shows results from our One-DM and few-shot methods on IAM dataset; readers are invited to identify our method. The answer is at the paper's end.", + "bbox": [ + 214, + 566, + 491, + 623 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/eac177df89240900fdd53803de76fdc1a0d63bf4112ef69e467b4621729a750e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TextStyle A··· interpolation··· Style B
plantsplants plnph plnch plnns plnns plnns plnns
cloudcloud cloud cloud cloud cloud cloud cloud
stickstick stick stick stick stick stick stick
photophoto photo photo photo photo photo photo
vaporVapor vapor vapor vapor vapor vapor vapor vapor
listslists lists lists lists lists lists lists
", + "bbox": [ + 524, + 470, + 772, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 6: We provide style interpolation results generated by our One-DM. Different individual writing styles are extracted from the IAM test dataset.", + "bbox": [ + 508, + 566, + 787, + 621 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative evaluation of Laplacian branch and gate mechanism. We conduct various ablation experiments on the IAM dataset to evaluate the effect of distinct components within our approach. We provide the quantitative result in Table 2. We find that: (1) The inclusion of both Laplacian branch and gate mechanism enhances the quality of the generated handwritten text images, improving FID by 3.92 and 2.71, respectively. (2) Integrating the Laplacian branch with the gate mechanism further boosts the generative performance.", + "bbox": [ + 212, + 640, + 787, + 746 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative evaluation of Laplacian branch and gate mechanism. To further analyze each module in our One-DM, we conduct visual ablation experiments. As shown in Table 2, we can observe that firstly, after adding a gate mechanism, background noise can be somewhat suppressed, resulting in characters with relatively clean backgrounds. Then, the independent addition of the Laplacian branch helps the model learn cursive connections and other", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1774ce3c6f3db2c6fc6f8aebdbe62df48343662b3d80eb5fcb152d71fe1577d7.jpg", + "table_caption": [ + "Table 2: Ablation study. Effect of the Laplacian branch and gate mechanism on the IAM dataset under the OOV-U setting, as done in [4]. In the middle, we showcase the generated samples of each component." + ], + "table_footnote": [], + "table_body": "
Base Εfre GateStyle samplesFID ↓
would like
✓cloudfoul108.44
✓cloudfoul105.73
✓cloudfoul104.52
✓cloudfoul102.75
", + "bbox": [ + 241, + 224, + 483, + 363 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b3f1a9819762c0aaa440f4fef71d7b275be53eb7bf53f8401b9695d9fc4fd8cf.jpg", + "table_caption": [ + "Table 3: Effect of each part of our Laplacian branch on IAM dataset under OOV-U setting." + ], + "table_footnote": [], + "table_body": "
HsLapNCEFID ↓
108.44
✓106.16
✓107.58
✓✓104.52
", + "bbox": [ + 584, + 189, + 702, + 262 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/273a1979ec18c427c5eb8bb82e54853d500bd4537f5ab7fa0e30647d79a39354.jpg", + "table_caption": [ + "Table 4: Comparisons with competitors on the CVL dataset." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓
HWT [4]58.22
VATr [39]54.44
Ours (One-DM)51.78
", + "bbox": [ + 576, + 299, + 710, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "style patterns. Finally, our method integrated the Laplacian branch and gate mechanism, which can generate the highest quality handwritten text images.", + "bbox": [ + 212, + 381, + 784, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Discussions of the Laplacian branch. Our Laplacian branch consists of two key components: utilizing high-frequency images $H_{s}$ and the Laplacian contrastive learning loss $\\mathcal{L}_{lapNCE}$ . In the previous ablation study, they are always combined. We further conduct exploration experiments to explain why they cannot be separated. As reported in Table 3, combining $H_{s}$ and $\\mathcal{L}_{lapNCE}$ maximizes effectiveness; separating them significantly reduces performance. Without the guidance of $\\mathcal{L}_{lapNCE}$ , extracting discriminative features from $H_{s}$ is challenging. Likewise, directly applying $\\mathcal{L}_{lapNCE}$ on original images leads to an undesired style feature extraction, as original images have less clear style patterns than $H_{s}$ .", + "bbox": [ + 212, + 411, + 784, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Discussions about learning style from a single reference. We are quite surprised that One-DM, with just a single reference sample, even exceeds the generative performance of few-shot methods. We provide the potential reason analysis below. Firstly, One-DM learns a meaningful style latent space, wherein new styles can be generated based on seen styles (cf. Figure 6). Then, through our style-enhanced module, One-DM effectively extracts the style feature from a single example and maps it to a position close to the exemplary writer in the feature space, thereby producing high-quality styled handwritten text images.", + "bbox": [ + 212, + 547, + 785, + 667 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Comparisons with SOTA Industrial Methods", + "text_level": 1, + "bbox": [ + 214, + 689, + 632, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To highlight the superiority of our method, we compare One-DM with leading industrial image generation methods that are trained on tremendously large datasets (including numerous text-centric images), including two prominent text-to-image methods, DALL-E3 [3] and Stable Diffusion [42] (SD), and two popular style transfer methods, Artbreeder1 and IP-Adapter2 (IP-A.), on the IAM dataset. Further experiment details are in supplementary material.", + "bbox": [ + 212, + 712, + 784, + 804 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "1 https://www.artbreeder.com/", + "bbox": [ + 217, + 810, + 442, + 824 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "2 https://github.com/tencent-ailab/IP-Adapter", + "bbox": [ + 218, + 825, + 568, + 839 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/805e3e0f4589677c79e85db22cc210f5c31488d5e288123592d6156ca553e840.jpg", + "image_caption": [ + "Fig. 7: (a) Qualitative comparisons between our method with SOTA industrial image generation methods, including DALL-E3 [3], Stable Diffusion [42], Artbreeder $^1$ and IP-Adapter $^2$ on handwritten text generation with both controllable styles and contents. (b) Qualitative comparisons with Fzshouji $^3$ on ICDAR-2013 competition database [52]." + ], + "image_footnote": [], + "bbox": [ + 230, + 146, + 763, + 246 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/36a29fe8864d814cbd14c89163702eb26e2f7dc58604529cc4bf2c0f1f14ae83.jpg", + "table_caption": [ + "Table 5: Quantitative comparisons with competitors on styled handwritten character generation in terms of FID." + ], + "table_footnote": [], + "table_body": "
MethodChineseJapanese
GANwriting [24]116.49111.86
HWT [4]165.74148.66
VATr [39]139.91124.98
WordStylist [37]34.61101.93
Ours (One-DM)27.2495.43
", + "bbox": [ + 243, + 362, + 452, + 448 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/68b6e1f75d5508c0c57049b301e78ddb992e5479114d096e008d793eff7eea84.jpg", + "table_caption": [ + "Table 6: Qualitative comparisons with VATr [39] on Chinese dataset." + ], + "table_footnote": [], + "table_body": "
GTVATrDiffusion Step
0123103050
艾思艾艾艾艾
布布布布布
琳洁淋淋淋淋
昂昂昂昂昂昂
", + "bbox": [ + 531, + 352, + 759, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Figure 7(a), our method excels industrial methods in style mimicry and content preservation. The performance of IP-A. is the poorest, often producing distorted images. Artbreeder can replicate the color of the strokes from style samples, but fails in content preservation. DALL-E3 and SD generate characters with accurate content, but often mismatch style details with references, such as character spacing and stroke width, with SD often generating extra backgrounds. Besides, we compare Fzshouji3, an advanced industrial method designed for Chinese handwriting generation. As shown in Figure 7(b), our method outperforms Fzshouji in replicating character details and ink color.", + "bbox": [ + 212, + 463, + 787, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Applications to Other Languages", + "text_level": 1, + "bbox": [ + 215, + 623, + 532, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we evaluate whether One-DM can be used to generate languages other than English. We further conduct experiments in the Chinese (i.e. ICDAR2013 competition database [52]) and Japanese(i.e. UP_Kuchibue database [22]) datasets. We use FID to assess generated samples of each writer, then average them. More experiment details can be seen in supplementary material.", + "bbox": [ + 212, + 648, + 784, + 724 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For the Chinese handwritten character generation task, as shown in Table 5, we find that our One-DM outperforms the second-best method by a large margin, achieving a 7.37 lower FID. From Figure 8 (a), we can observe that our One-DM generates characters that closely match the target images in geometric shape and character slant. In contrast, handwritings from HWT and VATr exhibit noticeable artifacts like blurring and collapsed structures. GANwriting tends to", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "3 https://www.fzshouji.com/make", + "bbox": [ + 217, + 823, + 460, + 840 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1533947c780ae3abe8b5c355723de45060f1dc2460e499931786197299abaf9f.jpg", + "image_caption": [ + "Fig. 8: Comparisons with GANwriting [24], HWT [4], VATr [39], and WordStylist [37] for styled handwritten character generation on Chinese and Japanese scripts. The red boxes highlight failures of structure preservation, while blue boxes highlight comparisons between the style patterns of targets and generated characters." + ], + "image_footnote": [], + "bbox": [ + 282, + 146, + 723, + 303 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "miss strokes in its handwritings. WordStylist sometimes struggles to accurately mimic style patterns and tends to generate characters with incorrect radicals. Table 5 and Figure 8 (b) further verify the effectiveness of One-DM for Japanese handwriting generation. We also achieve the lowest FID score, and our generated Japanese samples excel in both content preservation and style imitation.", + "bbox": [ + 212, + 369, + 784, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We further investigate why diffusion-based methods (Our One-DM and Word-Stylist) that require only one single sample outperform few-shot GAN-based methods (i.e., GANwriting, HWT, and VATr) so significantly in generating Chinese and Japanese characters. The lower performance of GAN-based methods on Chinese and Japanese characters may stem from their vanilla convolutional architectures struggling with Chinese and Japanese characters' complex geometry, as noted in [49]. In contrast, our One-DM breaks down the generation of Chinese and Japanese characters into simpler steps. For instance, as demonstrated in Table 6, during the early stages of the diffusion generation process, the model first attempts to generate a rough Chinese handwritten character. It then continues to refine the writing style (e.g., character shape and stroke color) under the condition guidance, until it synthesizes handwritings that are satisfactory.", + "bbox": [ + 212, + 445, + 787, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 648, + 359, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel One-DM for handwritten text generation, requiring only one style reference to produce realistic handwritten text images4. We enhance style extraction by incorporating high-frequency components from the style reference. For high-frequency components with distinct style patterns, we employ laplacian contrastive learning to capture more discriminative style features. Moreover, a gate mechanism improves the transfer of informative features from the reference, reducing background noise. Our One-DM outperforms few-shot methods in multiple language scripts. In the future, we aim to explore the potential of One-DM in font generation and vector font creation tasks.", + "bbox": [ + 212, + 680, + 787, + 816 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "4 In Figure 5, the second from bottom row belongs to our method.", + "bbox": [ + 217, + 824, + 663, + 839 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016).", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 247, + 323, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aksan, E., Pece, F., Hilliges, O.: Deepwriting: Making digital ink editable via deep generative modeling. In: ACM Conf. Human Factors Comput. Sys. pp. 1-14 (2018)", + "2. Alonso, E., Moysset, B., Messina, R.: Adversarial generation of handwritten text images conditioned on sequences. In: ICDAR. pp. 481-486 (2019)", + "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. 2(3), 8 (2023)", + "4. Bhunia, A.K., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S., Shah, M.: Hand-writing transformers. In: ICCV. pp. 1086-1094 (2021)", + "5. Cao, J., Mo, L., Zhang, Y., Jia, K., Shen, C., Tan, M.: Multi-marginal WassersteinianGAN. In: NeurIPS. vol. 32 (2019)", + "6. Chen, Z., Yang, D., Liang, J., Liu, X., Wang, Y., Peng, Z., Huang, S.: Complex handwriting trajectory recovery: Evaluation metrics and algorithm. In: ACCV. pp. 1060-1076 (2022)", + "7. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: CVPR. pp. 5977-5986 (2023)", + "8. Davis, B.L., Morse, B.S., Price, B.L., Tensmeyer, C., Wigington, C., Jain, R.: Text and style conditioned GAN for the generation of offline-handwriting lines. In: BMVC (2020)", + "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021)", + "0. Ding, H., Luan, B., Gui, D., Chen, K., Huo, Q.: Improving handwrittenOCR with training samples generated by glyph conditional denoising diffusion probabilistic model. In: ICDAR. pp. 20-37 (2023)", + "1. Fogel, S., Averbuch-Elor, H., Cohen, S., Mazor, S., Litman, R.: Scrabblegan: Semi-supervised varying length handwritten text generation. In: CVPR. pp. 4324-4333 (2020)", + "2. Gan, J., Wang, W.: Higan: Handwriting imitation conditioned on arbitrary-length texts and disentangled styles. In: AAAI. pp. 7484-7492 (2021)", + "3. Gan, J., Wang, W., Leng, J., Gao, X.: Higan+: Handwriting imitation gan with disentangled representations. ACM TOG 42(1), 1-17 (2022)", + "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. pp. 2672-2680 (2014)", + "5. Graves, A., Fernández, S., Gomez, F., Schmidhuber, J.: Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In: ICML. pp. 369-376 (2006)", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + ], + "bbox": [ + 225, + 280, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in neural information processing systems (2017)", + "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. vol. 33, pp. 6840-6851 (2020)", + "19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: arXiv (2022)", + "20. Huang, H., Yang, D., Dai, G., Han, Z., Wang, Y., Lam, K.M., Yang, F., Huang, S., Liu, Y., He, M.: Agtgan: Unpaired image translation for photographic ancient character generation. In: ACM MM. pp. 5456-5467 (2022)", + "21. Huang, S., Luo, Y., Zhuang, Z., Yu, J.G., He, M., Wang, Y.: Context-aware selective label smoothing for calibrating sequence recognition model. In: ACM MM (2021)", + "22. Jaeger, S., Nakagawa, M.: Two on-line japanese character databases in unipen format. In: ICDAR. pp. 566-570 (2001)", + "23. Kang, L., Riba, P., Rusinol, M., Fornes, A., Villegas, M.: Content and style aware generation of text-line images for handwriting recognition. IEEE TPAMI 44(12), 8846-8860 (2021)", + "24. Kang, L., Riba, P., Wang, Y., Rusinol, M., Fornés, A., Villegas, M.: Ganwriting: content-conditioned generation of styled handwritten word images. In: ECCV. pp. 273-289 (2020)", + "25. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: NeurIPS. pp. 18661-18673 (2020)", + "26. Khrulkov, V., Oseledets, I.: Geometry score: A method for comparing generative adversarial networks. In: International conference on machine learning. pp. 2621-2629. PMLR (2018)", + "27. Kleber, F., Fiel, S., Diem, M., Sablatnig, R.: Cvl-database: An off-line database for writer retrieval, writer identification and word spotting. In: ICDAR. pp. 560-564 (2013)", + "28. Kotani, A., Tellex, S., Tompkin, J.: Generating handwriting via decoupled style descriptors. In: ECCV. pp. 764-780 (2020)", + "29. Li, D., Chen, G., Wu, X., Yu, Z., Tan, M.: Face anti-spoofing with cross-stage relation enhancement and spoof material perception. Neural Networks 175, 106275 (2024)", + "30. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR. pp. 5141-5150 (2021)", + "31. Liu, Z., Chen, K., Zhang, Y., Han, J., Hong, L., Xu, H., Li, Z., Yeung, D.Y., Kwok, J.: Geom-erasing: Geometry-driven removal of implicit concept in diffusion models. arXiv (2023)", + "32. Liu, Z., Jia, W., Yang, M., Luo, P., Guo, Y., Tan, M.: Deep view synthesis via self-consistent generative network. IEEE TMM 24, 451-465 (2021)", + "33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017)", + "34. Luhman, T., Luhman, E.: Diffusion models for handwriting generation. arXiv (2020)", + "35. Luo, C., Zhu, Y., Jin, L., Li, Z., Peng, D.: Slogan: handwriting style synthesis for arbitrary-length and out-of-vocabulary text. IEEE Trans. Neural Networks Learn. Syst. (2022)", + "36. Marti, U.V., Bunke, H.: The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition 5, 39-46 (2002)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Nikolaidou, K., Retsinas, G., Christlein, V., Seuret, M., Sfikas, G., Smith, E.B., Mokayed, H., Liwicki, M.: Wordstylist: Structured verbatim handwritten text generation with latent diffusion models. In: ICDAR. pp. 384-401 (2023)", + "38. Pan, Z., Ji, Z., Liu, X., Bai, J., Liu, C.L.: Visa: Visual and semantic alignment for robust scene text recognition. In: ICDAR. pp. 223-242 (2023)", + "39. Pippi, V., Cascianelli, S., Cucchiara, R.: Handwritten text generation from visual archetypes. In: CVPR. pp. 22458-22467 (2023)", + "40. Ren, M.S., Zhang, Y.M., Wang, Q.F., Yin, F., Liu, C.L.: Diff-writer: A diffusion model-based stylized online handwritten chinese character generator. In: International Conference on Neural Information Processing. pp. 86-100 (2023)", + "41. Retsinas, G., Sfikas, G., Gatos, B., Nikou, C.: Best practices for a handwritten text recognition system. In: International Workshop on Document Analysis Systems. pp. 247-259 (2022)", + "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)", + "44. Tang, S., Lian, Z.: Write like you: Synthesizing your cursive online chinese handwriting via metric-based meta learning. In: Comput. Graph. Forum. pp. 141-151 (2021)", + "45. Tang, S., Xia, Z., Lian, Z., Tang, Y., Xiao, J.: Fontrnn: Generating large-scale chinese fonts via recurrent neural network. In: Comput. Graph. Forum. pp. 567-577 (2019)", + "46. Tolosana, R., Delgado-Santos, P., Perez-Uribe, A., Vera-Rodriguez, R., Fierrez, J., Morales, A.: Deepwritesyn: On-line handwriting synthesis via deep short-term representations. In: AAAI. pp. 600-608 (2021)", + "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "48. Xie, X., Fu, L., Zhang, Z., Wang, Z., Bai, X.: Toward understanding wordart: Corner-guided transformer for scene text recognition. In: ECCV. pp. 5456-5467 (2022)", + "49. Xie, Y., Chen, X., Sun, L., Lu, Y.: Dg-font: Deformable generative networks for unsupervised font generation. In: CVPR (2021)", + "50. Yang, Y., Liu, D., Zhang, S., Deng, Z., Huang, Z., Tan, M.: Hilo: Detailed and robust 3d clothed human reconstruction with high-and low-frequency information of parametric models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10671-10681 (2024)", + "51. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15901-15911 (2023)", + "52. Yin, F., Wang, Q.F., Zhang, X.Y., Liu, C.L.: Icdar 2013 chinese handwriting recognition competition. In: International Conference on Document Analysis and Recognition. pp. 1464-1470 (2013)", + "53. Zhang, Y., Hooi, B.: Hipa: Enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv (2023)", + "54. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS. pp. 34077-34090 (2022)", + "55. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: NeurIPS. pp. 29848-29860 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation", + "bbox": [ + 243, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Zhao, B., Tao, J., Yang, M., Tian, Z., Fan, C., Bai, Y.: Deep imitator: Handwriting calligraphy imitation via deep attention networks. Pattern Recognition 104, 107080 (2020)", + "57. Zhu, Y., Li, Z., Wang, T., He, M., Yao, C.: Conditional text image generation with diffusion models. In: CVPR. pp. 14235-14245 (2023)", + "58. Zhuang, Z., Liu, Z., Lam, K.M., Huang, S., Dai, G.: A new semi-automatic annotation model via semantic boundary estimation for scene text detection. In: ICDAR. pp. 257-273 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 258 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G. Dai et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_model.json b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7170a2814a8b6c54e2ac534133ef58141d8f3119 --- /dev/null +++ b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_model.json @@ -0,0 +1,2447 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.185 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.212, + 0.713, + 0.244 + ], + "angle": 0, + "content": "Gang Dai\\(^{1\\star}\\), Yifan Zhang\\(^{2,3\\star}\\), Quhui Ke\\(^{1}\\), Qiangya Guo\\(^{1}\\), and Shuangping Huang\\(^{1,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.255, + 0.732, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\)South China University of Technology \n\\(^{2}\\)National University of Singapore, \\(^{3}\\)Skywork AI, \\(^{4}\\)Pazhou Laboratory \n\\(\\{\\mathrm{eedaigang}@mail., \\mathrm{eehsp@}\\}\\) scut.edu.cn, yifan.zhang@u.nus.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.335, + 0.74, + 0.64 + ], + "angle": 0, + "content": "Abstract. Existing handwritten text generation methods often require more than ten handwriting samples as style references. However, in practical applications, users tend to prefer a handwriting generation model that operates with just a single reference sample for its convenience and efficiency. This approach, known as \"one-shot generation\", significantly simplifies the process but poses a significant challenge due to the difficulty of accurately capturing a writer's style from a single sample, especially when extracting fine details from the characters' edges amidst sparse foreground and undesired background noise. To address this problem, we propose a One-shot Diffusion Mimicker (One-DM) to generate handwritten text that can mimic any calligraphic style with only one reference sample. Inspired by the fact that high-frequency information of the individual sample often contains distinct style patterns (e.g., character slant and letter joining), we develop a novel style-enhanced module to improve the style extraction by incorporating high-frequency components from a single sample. We then fuse the style features with the text content as a merged condition for guiding the diffusion model to produce high-quality handwritten text images. Extensive experiments demonstrate that our method can successfully generate handwriting scripts with just one sample reference in multiple languages, even outperforming previous methods using over ten samples. Our source code is available at https://github.com/dailenson/One-DM." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.653, + 0.704, + 0.667 + ], + "angle": 0, + "content": "Keywords: Handwritten Text Generation \\(\\cdot\\) One-Shot Generation" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.693, + 0.375, + 0.709 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In the digital age, handwriting text generation blends the personalization of traditional handwriting with the efficiency of automated processes, offering a digital format to preserve the authenticity of individual handwriting. This task aims to automatically generate the desired handwritten text images that not only correspond to specific text content, but also emulate the calligraphic style of a given exemplar writer (e.g., character slant, cursive join, stroke thickness, and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.589, + 0.84 + ], + "angle": 0, + "content": "* Authors contributed equally; † Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.144, + 0.495, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.145, + 0.761, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 1: User experience comparisons between one-shot and few-shot handwritten text generation methods. It reveals that one-shot setting leads to a better user experience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.357 + ], + "angle": 0, + "content": "ink color). This provides great convenience for people with hand impairments and also contributes to accelerating the process of handwritten font design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.794, + 0.522 + ], + "angle": 0, + "content": "Previous works [4,8,11-13,24,39] introduce generative adversarial networks (GANs) [5,14,32] for handwritten text generation. For instance, ScrabbleGAN [11] leverages random noises as style inputs and conditions content inputs on character-level labels, enabling the synthesis of handwritten words with randomly sampled styles. In a recent study [9], diffusion models [31,42,53] such as denoising diffusion probabilistic models (DDPM) [18] have demonstrated even higher quality of image generation compared to GANs. This motivates several attempts, such as WordStylist [37], GC-DDPM [10], and CTIG-DM [57], to condition the denoising process on the fixed writer ID to generate handwritten text images with controllable styles. However, the major limitation of these methods [10,11,37,57] is that they are unable to mimic the unseen writers' handwriting styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.689 + ], + "angle": 0, + "content": "To imitate any given handwriting style, some previous methods [4, 24, 39], known as few-shot generation, require users to provide a few samples (typically 15) as style references. They employ a style encoder to extract writing styles from the given samples, thus offering flexible control over the style of the generated samples. However, the traditional few-shot generation pipeline is inconvenient, inefficient, and time-consuming, as shown in Figure 1. Users prefer methods that only require a single sample as a style reference, known as one-shot generation methods, because they are more convenient, efficient, and easy-to-use. Our goal is to investigate the more challenging one-shot generation task that holds significant practical value, striving to achieve high-quality handwritten text image generation with desired styles and contents." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The primary challenge of one-shot generation task lies in accurately extracting a user's handwriting style from just one style reference image. As illustrated in Figure 2, characters, being abstract symbols, typically occupy only a minor portion of the reference image. Besides, the reference image is often cluttered with noisy background information, which poses a significant obstacle in extracting the individual handwriting style. Previous one-shot generation methods [8,12,13] simply follow the architectural design of few-shot approaches, using a vanilla CNN encoder to directly extract the handwriting style from a single sample. The extracted style is then combined with text contents and input into a CNN decoder to generate the desired handwritten images. These methods ex" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.151, + 0.709, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.285, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Fig. 2: Handwritten text samples and corresponding high-frequency components. We find that high-frequency components have more pronounced character contours, clearly showcasing the style patterns, such as character slant and cursive connections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.784, + 0.371 + ], + "angle": 0, + "content": "hibit a limited performance in emulating handwriting styles, due to their poor ability of style extraction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.476 + ], + "angle": 0, + "content": "To address the above challenges, our key idea revolves around utilizing the high-frequency information of the sample to enhance the extraction of the handwriting style. As depicted in Figure 2, the high-frequency information encompasses the overall contours of the handwritten text, allowing for a clear observation of key style patterns such as the text slant, letter spacing, and cursive connections. Hence, the incorporation of high-frequency information facilitates a more effective handwriting style extraction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.786, + 0.735 + ], + "angle": 0, + "content": "In light of the above insight, we propose a One-shot Diffusion Mimicker (One-DM) for handwritten text generation, which is simultaneously guided by desired style and arbitrary content. More specifically, we first develop a style-enhanced module to process both a style reference image and its high-frequency components in parallel. Considering that the reference image often contains background noise, we design a gate mechanism to suppress the inflow of background noise. Regarding the high-frequency components, such as character slant and ligature that show clearer style patterns, we employ a contrastive learning framework [25,54,55] to further obtain discriminative style features, guiding the handwritten text synthesis with both realistic and diverse styles. Subsequently, the style features extracted from both branches are adaptively fused with the specific content prototype in a style-content fusion module. Finally, the seamlessly integrated style and content features act as conditional inputs, guiding the denoising process for the progressive synthesis of stylized handwritten text images. It is worth noting that our One-DM effectively mimics a user's writing style with only one reference sample through our style-enhanced module, surpassing few-shot methods in producing higher-quality stylized handwritings." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.736, + 0.56, + 0.751 + ], + "angle": 0, + "content": "To sum up, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.764, + 0.784, + 0.809 + ], + "angle": 0, + "content": "- We propose a novel diffusion model for stylized handwritten text generation, which only requires a single reference sample as style input, and imitates its writing style to generate handwritten text with arbitrary content." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- We introduce the high-frequency components of the reference sample to enhance the extraction of handwriting style. The proposed style-enhanced" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.764, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.147, + 0.786, + 0.177 + ], + "angle": 0, + "content": "module can effectively capture the writing style patterns and suppress the interference of background noise." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.787, + 0.223 + ], + "angle": 0, + "content": "- Extensive experiments on handwriting datasets in English, Chinese, and Japanese demonstrate that our approach with a single style reference even outperforms previous methods with 15x-more references." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.248, + 0.388, + 0.264 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.785, + 0.401 + ], + "angle": 0, + "content": "Handwriting generation. Handwritten text is typically stored in two formats: online trajectory form or offline image form. Online handwriting generation methods [1, 7, 28, 44, 46, 56] often employ Recurrent Neural Networks (RNNs) [1, 6, 28, 44-46, 56], transformer decoders [7], or diffusion models [34, 40] to progressively generate writing trajectories. Unlike online methods, offline generation methods have the advantage of not requiring additional trajectory supervision information and can generate realistic handwritten text with stroke width and ink color, which online methods cannot produce." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.614 + ], + "angle": 0, + "content": "Previous offline handwriting generation methods [2, 4, 8, 11, 20, 24, 35, 39] in deep learning predominantly rely on Generative Adversarial Networks (GANs). Early works [2, 11] condition the generative process on the word embeddings [2] or concatenated letter-tokens [11] to synthesize handwritten word images. However, these methods struggle to flexibly control the writing style. Thus, few-shot approaches [4, 24, 39] which rely on 15 style references are introduced. For instance, GANwriting [24] utilizes a CNN encoder [16, 50, 51] to extract a user's handwriting style from a few samples, which are then combined with specific text content to generate handwritings in the desired style. In its follow-up work [23], the synthesized samples are demonstrated to assist in training more robust handwritten text recognizers. Further, HWT [4] employs a transformer encoder [29, 47] to extract rich style patterns from reference samples, enhancing the performance in style mimicry. Recently, VATr [39] uses images of symbols as content representations, enabling the generation of out-of-charset characters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Concurrently, one-shot generation methods [8, 12, 13] are proposed. Despite their ability to mimic handwriting styles with only a single sample, these techniques still lag behind few-shot methods in terms of stylized generation results. Additionally, previous handwriting generation methods [2, 4, 8, 11-13, 24] often rely on complex content representations, such as recurrent embeddings [2] and letter-level tokens [4, 8, 11-13, 24]. SLOGAN [35] extracts textual contents from easily obtainable printed images. However, it faces challenges in generalizing to unseen writing styles due to its fixed writer ID. Similarly, some diffusion-based [10, 37, 57] methods condition the denoising process on fixed style ID and are unable to mimic styles that they have not previously encountered. In contrast, our One-DM effectively obtains style information from one style sample and thus can generate handwritings with arbitrary styles. Due to space constraints, we discuss more related works in supplementary, including diffusion methods for general image generation, contrastive-based method [48], and frequency-based methods [30, 38]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.136, + 0.768, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.304, + 0.788, + 0.402 + ], + "angle": 0, + "content": "Fig. 3: Overview of the proposed method. The style reference initially passes through a high-pass filter to extract its high-frequency components. Subsequently, the spatial and the high-frequency style encoders independently extract style features \\( F_{spa} \\) and \\( F_{fre} \\) from the style reference and its high-frequency information, respectively. \\( F_{spa} \\), after being filtered through a gate mechanism, is fused with \\( F_{fre} \\) and content features \\( E \\) in the fusion module. The merged feature then serves as a condition input to guide the diffusion generation process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.33, + 0.429 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.54 + ], + "angle": 0, + "content": "Problem statement. We aim to synthesize handwritten text images that are controlled by both textual content and handwriting style. Given any word string \\(\\mathcal{A}\\) and a single style sample \\(I_{s}\\) from a writer \\(w_{s}\\), the generated handwritten word image \\(X_{s}\\) should replicate the unique handwriting style of \\(w_{s}\\) while maintaining the content of \\(\\mathcal{A}\\), where the textual content \\(\\mathcal{A} = \\{a_i\\}_{i=1}^L\\) spans a length \\(L\\) and \\(a_{i}\\) is from a character set without any restrictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.648 + ], + "angle": 0, + "content": "To address this task, we incorporate high-frequency information to enhance the extraction of writing styles. Existing methods [4,8,12,13,24,39] typically use a vanilla CNN or transformer encoder to directly process style images, often resulting in undesired style extraction. In contrast, we introduce a novel One-shot diffusion mimicker (One-DM) by innovating Laplacian high-frequency extraction and a gating mechanism. Our One-DM can effectively capture style features from a single reference while suppressing background noise." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.391, + 0.688 + ], + "angle": 0, + "content": "3.1 Overall Scheme" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our high-level idea focuses on incorporating high-frequency information from style reference images to enhance the extraction of style patterns. A straightforward implementation involves using a vanilla transformer encoder to extract style features from both the style image and its corresponding high-frequency image. This naive solution encounters two main issues: (1) the lack of efficient supervision objective still makes it challenging to accurately learn the writer's style patterns from the high-frequency image, and (2) the style features captured from the original image still retain undesirable noise backgrounds, which can adversely affect subsequent image generation performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "To address the above-mentioned issues, we develop a more effective method, shown in Figure 3. Our method consists of a style-enhanced module, a content encoder, a style-content fusion module, and a conditional diffusion module. Initially, we utilize the Laplacian kernel as a high-frequency filter to extract the high-frequency components from the style reference. Then, two parallel style encoders are employed to simultaneously extract the corresponding style features from both style reference and its high-frequency information. Since undesirable background noise is often present in style references, we design a gate mechanism to facilitate the transmission of informative style information while mitigating noise transmission. The style patterns in the high-frequency components are comparatively cleaner and more distinct, facilitating the observation of individual styles, such as character slant. For the observation, we propose a contrastive learning objective, termed LaplacianNCE \\(\\mathcal{L}_{lapNCE}\\), to enforce the more discriminative style learning from high-frequency components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Regarding content guidance, we render the given string \\(\\mathcal{A}\\) into a unifont image, as done in VATr [39]. Briefly, Unifont's key advantage is that it covers all Unicode characters, allowing our method to transform any user inputs into corresponding images. We further feed render results into a content encoder that combines a ResNet18 [16] with a Transformer encoder. This process first involves using a ResNet18 to handle each character image in parallel, followed by concatenating them to form word sequence features. The Transformer encoder then processes these features to extract an informative content feature \\(E = \\{e_i\\}_{i=1}^L \\in \\mathbb{R}^{L \\times c}\\) with a global context, where \\(c\\) is the channel dimension. After obtaining style features and content guidance, we seamlessly fusion them using a style-content fusion module. The fused results then guide the denoising process of the conditional diffusion model to synthesize the desired handwritten text images. The denoising process is supervised by the reconstruction loss \\(\\mathcal{L}_{rec}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.787, + 0.583 + ], + "angle": 0, + "content": "To summarize, the overall training objective of our method combines all two loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.585, + 0.579, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {l a p N C E} + \\mathcal {L} _ {r e c}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.62, + 0.457, + 0.635 + ], + "angle": 0, + "content": "3.2 Style-enhanced Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.844 + ], + "angle": 0, + "content": "We propose the style-enhanced module to enhance style extraction by incorporating high-frequency components \\( H_{s} \\in \\mathbb{R}^{h \\times w \\times c} \\) from a reference image \\( I_{s} \\in \\mathbb{R}^{h \\times w \\times c} \\) since clearer style patterns are presented in high-frequency components like character slanting and shape connections (cf. Figure 2). As shown in Figure 3, we use a Laplacian kernel as a high-frequency filter to extract \\( H_{s} \\) from \\( I_{s} \\). The Laplacian kernel excels in extracting high-frequency information without the need for Fast Fourier Transform (FFT) and parametric separation in the frequency domain. Then, two style encoders, \\( \\mathcal{E}_{spa} \\) and \\( \\mathcal{E}_{fre} \\), each a combination of CNN and transformer, process the \\( I_{s} \\) and \\( H_{s} \\), respectively. This independent processing leads to the extraction of distinct style features: \\( F_{spa} = \\{f_{spa}^{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d \\times c} \\) from \\( \\mathcal{E}_{spa} \\) and \\( F_{fre} = \\{f_{fre}^{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d \\times c} \\) from \\( \\mathcal{E}_{fre} \\), where \\( d = h \\times w \\). While structurally identical, \\( \\mathcal{E}_{spa} \\) and \\( \\mathcal{E}_{fre} \\) do not share weights with each other. Then, the proposed \\( \\mathcal{L}_{lapNCE} \\) forces \\( \\mathcal{E}_{fre} \\) to focus on extracting discriminative style features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "from \\( H_{s} \\). A gate mechanism is designed to selectively filter out background noise from the reference style features, allowing only meaningful style patterns to pass." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.254 + ], + "angle": 0, + "content": "Laplacian Contrastive Learning. The goal of the proposed \\(\\mathcal{L}_{\\text{lapNCE}}\\) is to guide the high-frequency style encoder \\(\\mathcal{E}_{\\text{fre}}\\) in learning more discriminative style features from the high-frequency information. Thus, we propose to bring closer extracted style features \\(F_{\\text{fre}}\\) belonging to the same writer, while distancing those from different writers. We formulate our \\(\\mathcal{L}_{\\text{lapNCE}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.265, + 0.785, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a p N C E}} = - \\frac {1}{N} \\sum_ {i \\in M} \\frac {1}{| P (i) |} \\sum_ {p \\in P (i)} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} / \\tau\\right)}{\\sum_ {a \\in A (i)} \\exp \\left(z _ {i} \\cdot z _ {a} / \\tau\\right)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.784, + 0.4 + ], + "angle": 0, + "content": "In detail, \\( i \\in M = \\{1 \\dots N\\} \\) is the index of any element in the mini-batch with size \\( N \\) and \\( A(i) = M \\setminus \\{i\\} \\) is other indices distinct from \\( i \\). \\( z_{i} \\) is an anchor sample belonging to writer \\( w_{i} \\) and \\( P(i) = \\{p \\in A(i) : w_{p} = w_{i}\\} \\) is its in-batch positive sample set and the other \\( A(i) \\setminus P(i) \\) is its negative set. Here, \\( z = Proj(F_{fre}) \\), where \\( Proj \\) is a learnable multi-layer perceptron (MLP), \\( \\tau \\) is a scalar temperature parameter and the \\( \\cdot \\) symbol denotes the inner product." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.785, + 0.572 + ], + "angle": 0, + "content": "Gate mechanism. As illustrated in Figure 2, the stroke areas of characters in reference images are typically sparse, with background noise interfering with the extraction of character style features. To address this challenge, we propose a gate mechanism to selectively filter the information of the sample \\( I_{s} \\), as shown in Figure 3. Specifically, the extracted sample style features \\( F_{spa} \\) are fed into a gate layer, consisting of a learnable fully connected layer and followed by a sigmoid activation, to obtain the corresponding gate units \\( W = \\{w_{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d} \\). Each unit \\( w_{i} \\) determines the pass rate for the corresponding \\( f_{spa}^{i} \\), allowing for a higher pass rate where \\( w_{i} \\) is larger. This design effectively enables informative style features \\( \\hat{F}_{spa} = \\{\\hat{f}_{spa}^{i}\\}_{i=1}^{d} \\) to be extracted while suppressing extraneous background noise, where \\( \\hat{f}_{spa}^{i} = f_{spa}^{i} \\cdot w_{i} \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.593, + 0.499, + 0.607 + ], + "angle": 0, + "content": "3.3 Style-content Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.784, + 0.77 + ], + "angle": 0, + "content": "Upon acquiring textual content feature \\( E \\) and two style features \\( \\hat{F}_{spa} \\) and \\( F_{fre} \\), we integrate all features within two multi-head attention mechanisms to guide the denoising generation process of the diffusion model, as shown in Figure 3. Specifically, the first cross-attention module uses the textual content \\( E \\) as queries to identify the most relevant style information in the style reference, thereby inferring the style attributes corresponding to each character. For instance, if the textual content is 'a', it prioritizes searching for style features of characters like 'a', 'b', 'd', 'g' in the style reference, due to these characters appearing similar looped structures, implying more comparable style attributes. This process (cross-attention in Figure 3) is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.782, + 0.785, + 0.799 + ], + "angle": 0, + "content": "\\[\nO = \\operatorname {A t t e n} _ {1} \\left(Q _ {1} = E, K _ {1} = V _ {1} = \\hat {F} _ {s p a} + F _ {f r e}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Subsequently, we obtain the initial fusion embedding between content and style guidance by simply summing \\( O \\) and \\( E \\). The merged intermediate vector is then" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "employed as the query, key, and value in the self-attention mechanism to facilitate comprehensive interaction of information. Finally, the blended embedding \\( g \\) serves as the condition of the diffusion process. The second multi-head attention (self-attention in Figure 3) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.22, + 0.786, + 0.236 + ], + "angle": 0, + "content": "\\[\ng = \\operatorname {A t t e n} _ {2} \\left(Q _ {2} = K _ {2} = V _ {2} = O + E\\right). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.256, + 0.498, + 0.271 + ], + "angle": 0, + "content": "3.4 Conditional Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.789, + 0.357 + ], + "angle": 0, + "content": "The goal of the conditional diffusion model \\( p_{\\theta} \\) is to generate realistic images of handwritten text, guided by acquired conditions \\( g \\). Specifically, as shown in Figure 3, under the guidance of \\( g \\), \\( p_{\\theta} \\) executes a denoising generative process with \\( T \\) steps, starting from a sampled Gaussian noise \\( x_{T} \\) and progressively denoises to obtain the desired handwritten text \\( x_{0} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.367, + 0.786, + 0.397 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0} \\mid g\\right) = \\int p _ {\\theta} \\left(x _ {0: T} \\mid g\\right) d _ {x _ {1: T}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.413, + 0.786, + 0.453 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0: T} \\mid g\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, g\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.464, + 0.785, + 0.481 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, g\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, g, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, g, t\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.547 + ], + "angle": 0, + "content": "The denoising process aims to learn how to reverse a predefined forward process, as described in DDPM [18]. The forward process is modeled as a fixed Markov chain, where noise conforming to a normal distribution is incrementally added to \\( x_{t - 1} \\) to derive \\( x_{t} \\). This can be mathematically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.559, + 0.786, + 0.577 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.648 + ], + "angle": 0, + "content": "where the noise is characterized by a variance schedule \\(\\beta_{t}\\). During training, a variational bound on the maximum likelihood objective is applied to guide the generation process to recover \\(x_0\\) from the standard Gaussian noise \\(x_{T}\\) conditioned on \\(g\\). We give the training objective as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.658, + 0.786, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {t, q} \\| \\mu_ {t} \\left(x _ {t}, x _ {0}\\right) - \\mu_ {\\theta} \\left(x _ {t}, g, t\\right) \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.717 + ], + "angle": 0, + "content": "where \\(\\mu_t(x_t, x_0)\\) is the mean of the forward process posterior \\(q(x_t | x_{t-1})\\), which has a closed-form solution [18]. We put more details in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.377, + 0.757 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.448, + 0.787 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Dataset. To evaluate our One-DM in handwritten text generation, we use the widely-used handwriting dataset IAM [36] and CVL [27].IAM consists of 62,857 English word images from 500 unique writers. Following [4,24,39], we use words" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "from 339 writers for training and the remaining 161 writers for testing. CVL includes texts by 310 writers in English and German. We use the English portion, totaling 84,514 words, and follow CVL's standard split, with 283 writers for training and 27 for testing. Throughout all experiments, we resize images to 64 pixels in height, preserving their aspect ratio." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.785, + 0.282 + ], + "angle": 0, + "content": "Evaluation metrics. We use the Fréchet Inception Distance (FID) [17] and Geometry Score (GS) [26] to assess the generation quality, following the settings of [4, 24, 39]. We also conduct user studies to quantify the subjective quality of the generated handwritten text images in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.433 + ], + "angle": 0, + "content": "Implementation details. In our all experiments, we only use single style reference sample. We first train our model for 700 epochs (batch size of 384) with classifier-free guidance strategy [19] under the guidance scale of 0.25, and then fine-tune the model for 4500 iterations (batch size of 128) using a text recognizer [41] with a CTC loss [15, 21, 58], on four RTX3090 GPUs. The fine-tuning process forces our One-DM to generate readable text with accurate content. The optimizer is AdamW [33], with a learning rate of \\(10^{-4}\\). During inference, each style sample is randomly sampled from the target-writer. To speed up the sampling, we use the denoising diffusion implicit model (DDIM) [43] with 50 steps. More details are provided in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.541 + ], + "angle": 0, + "content": "Compared methods. We compare our One-DM with state-of-the-art handwritten text generation methods, including GAN-based methods (i.e., GANwriting [24], HWT [4], VATr [39], TS-GAN [8], HiGAN+ [13]), and diffusion-based methods (i.e. GC-DDPM [10] and WordStylist [37]). For a fair comparison, we configure all methods to generate images with a height of 64 pixels in supplementary. Additionally, in supplementary, we compare a variant of One-DM that generates 32-pixel high images with official VATr [39] and HWT [4]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.561, + 0.373, + 0.574 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.718 + ], + "angle": 0, + "content": "Stylish Handwritten Text Generation. Initially, we assess our One-DM for producing styled handwritten text images, aiming to replicate both the style and content in the generated images. Following [4, 24, 39], we first calculate FID between generated and real samples for each writer separately and finally average them. In line with the previous works [4, 24, 39], our experiments on the IAM dataset are divided into four distinct scenarios: IV-S, IV-U, OOV-S, OOV-U. Among these four scenarios, OOV-U represents the most challenging case where both the target style and the words are entirely unseen during training. For the CVL dataset, we directly report the results of all methods on the test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We first report the quantitative results on the IAM dataset in Table 1. We can observe that our One-DM outperforms all the competitors in all settings. Notably, it significantly exceeds one-shot methods in all scenarios. Impressively, our One-DM also holds a substantial advantage over few-shot methods (GAN-writing [24], HWT [4], VATr [39]), which use 15-x more reference samples for style guidance, in IV-S and OOV-S settings. Even in the most challenging OOV-U scenario, our One-DM leads the second-best, VATr, by a large margin (102.75 vs. 108.76), demonstrating the superior performance of our One-DM in stylized" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Comparisons with state-of-the-art methods on styled and style-agnostic handwritten text generation in the IAM dataset. Note that, all methods are trained on the same training set used in GANwriting [24], HWT [4], and VATr [39]." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.193, + 0.706, + 0.327 + ], + "angle": 0, + "content": "
MethodShotStyloid EvaluationStyle-agnostic
IV-SIV-UOOV-SOOV-UFID↓GS↓
TS-GAN [8]One118.56128.75127.11136.6720.654.88×10-2
GANwriting [24]Few120.07124.30125.87130.6828.375.67×10-2
HiGAN+ [13]One117.33116.95121.55121.4822.952.06×10-2
GC-DDPM [10]One99.86105.73112.52118.3919.051.31×10-2
WordStylist [37]One98.10104.27109.45115.5218.582.85×10-2
HWT [4]Few109.25106.90116.55113.5218.994.41×10-3
VATr [39]Few103.75101.73111.64108.7616.031.74×10-2
Ours (One-DM)One89.4798.3693.30102.7515.731.98×10-3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.784, + 0.375 + ], + "angle": 0, + "content": "handwritten text generation. Similarly, our method outperforms HWT and VATr on the CVL dataset, achieving the lowest FID score, as shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.785, + 0.571 + ], + "angle": 0, + "content": "We provide qualitative results to intuitively explain the benefit of our One-DM, in Figure 4. GANwriting struggles to capture the style patterns of reference samples, such as character slant and occasionally produces unclear character shapes. HiGAN+ more consistently generates characters with correct content, but the character spacing within the generated words lacks realism. WordStylist typically produces images with noticeable background noise. HWT and VATr can produce satisfactory handwritten words in terms of content accuracy and style mimicry; however, their downside is the tendency to create smoother character appearances. Compared to HWT and VATr, our synthesized samples excel in more authentic character ink color and stroke thickness. However, some generated samples by our One-DM are visibly different in ink color. We provide comprehensive explanations in supplementary. We further display more qualitative comparisons between our method and few-shot methods in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.787, + 0.724 + ], + "angle": 0, + "content": "Style-agnostic Handwritten Text Generation. We further evaluate our One-DM to generate realistic handwritten text images, irrespective of style imitation. For this purpose, we compute the FID and GS on the IAM test set, under the same conditions as applied to ScrabbleGAN [11] (FID: 20.72, GS: \\(2.56 \\times 10^{-2}\\)), capable of generating handwritten texts with randomly sampled styles. Specifically, each method generates 25k random samples to calculate FID against 25k samples from the test set, and 5k random samples for GS calculation in comparison with 5k test set samples. As presented in Table 1, our One-DM achieves the best results in both FID and GS metrics, further demonstrating its capability to generate higher-quality handwritten text images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.743, + 0.334, + 0.758 + ], + "angle": 0, + "content": "4.3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to analyze our One-DM. More analyses are provided in supplementary, including generalization evaluation on different style backgrounds, generation quality assessment through OCR performance, failure case analysis, and the effects of different designs (e.g., high-frequency filter, style-content fusion mechanism, and style input sample length)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.144, + 0.749, + 0.367 + ], + "angle": 0, + "content": "
Style examplesSee us next him unless some eight have the be is unassailable of our boxInour there they on as had inside are shaw have us God is unprouctand tuning which oil to you will the them may the rivers Rim Fishand to is all Common Wood Lessis of ther in work and it hean of
GANw.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*HiGAN+The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*WordS.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
HWTThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
VATrThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*OursThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.375, + 0.79, + 0.457 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons between our method with state-of-the-art methods on handwritten text generation with both specific textual content and desired handwriting style in the IAM dataset. We utilize the identical guiding text, 'The greatest test of courage on earth is to bear defeat without losing heart,' across all handwriting generation methods, directing them to produce text in varied styles. Better zoom in \\(200\\%\\). * denotes one-shot methods, while others are few-shot methods." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.469, + 0.473, + 0.564 + ], + "angle": 0, + "content": "
Targethopetherelatersuccess
Generated Sampleshopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.568, + 0.493, + 0.624 + ], + "angle": 0, + "content": "Fig. 5: Each row shows results from our One-DM and few-shot methods on IAM dataset; readers are invited to identify our method. The answer is at the paper's end." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.472, + 0.773, + 0.563 + ], + "angle": 0, + "content": "
TextStyle A··· interpolation··· Style B
plantsplants plnph plnch plnns plnns plnns plnns
cloudcloud cloud cloud cloud cloud cloud cloud
stickstick stick stick stick stick stick stick
photophoto photo photo photo photo photo photo
vaporVapor vapor vapor vapor vapor vapor vapor vapor
listslists lists lists lists lists lists lists
" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.567, + 0.788, + 0.622 + ], + "angle": 0, + "content": "Fig. 6: We provide style interpolation results generated by our One-DM. Different individual writing styles are extracted from the IAM test dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Quantitative evaluation of Laplacian branch and gate mechanism. We conduct various ablation experiments on the IAM dataset to evaluate the effect of distinct components within our approach. We provide the quantitative result in Table 2. We find that: (1) The inclusion of both Laplacian branch and gate mechanism enhances the quality of the generated handwritten text images, improving FID by 3.92 and 2.71, respectively. (2) Integrating the Laplacian branch with the gate mechanism further boosts the generative performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative evaluation of Laplacian branch and gate mechanism. To further analyze each module in our One-DM, we conduct visual ablation experiments. As shown in Table 2, we can observe that firstly, after adding a gate mechanism, background noise can be somewhat suppressed, resulting in characters with relatively clean backgrounds. Then, the independent addition of the Laplacian branch helps the model learn cursive connections and other" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.51, + 0.215 + ], + "angle": 0, + "content": "Table 2: Ablation study. Effect of the Laplacian branch and gate mechanism on the IAM dataset under the OOV-U setting, as done in [4]. In the middle, we showcase the generated samples of each component." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.226, + 0.485, + 0.364 + ], + "angle": 0, + "content": "
Base Εfre GateStyle samplesFID ↓
would like
✓cloudfoul108.44
✓cloudfoul105.73
✓cloudfoul104.52
✓cloudfoul102.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.145, + 0.755, + 0.187 + ], + "angle": 0, + "content": "Table 3: Effect of each part of our Laplacian branch on IAM dataset under OOV-U setting." + }, + { + "type": "table", + "bbox": [ + 0.586, + 0.19, + 0.703, + 0.263 + ], + "angle": 0, + "content": "
HsLapNCEFID ↓
108.44
✓106.16
✓107.58
✓✓104.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.533, + 0.27, + 0.753, + 0.297 + ], + "angle": 0, + "content": "Table 4: Comparisons with competitors on the CVL dataset." + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.3, + 0.712, + 0.362 + ], + "angle": 0, + "content": "
MethodFID ↓
HWT [4]58.22
VATr [39]54.44
Ours (One-DM)51.78
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.412 + ], + "angle": 0, + "content": "style patterns. Finally, our method integrated the Laplacian branch and gate mechanism, which can generate the highest quality handwritten text images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.548 + ], + "angle": 0, + "content": "Discussions of the Laplacian branch. Our Laplacian branch consists of two key components: utilizing high-frequency images \\( H_{s} \\) and the Laplacian contrastive learning loss \\( \\mathcal{L}_{lapNCE} \\). In the previous ablation study, they are always combined. We further conduct exploration experiments to explain why they cannot be separated. As reported in Table 3, combining \\( H_{s} \\) and \\( \\mathcal{L}_{lapNCE} \\) maximizes effectiveness; separating them significantly reduces performance. Without the guidance of \\( \\mathcal{L}_{lapNCE} \\), extracting discriminative features from \\( H_{s} \\) is challenging. Likewise, directly applying \\( \\mathcal{L}_{lapNCE} \\) on original images leads to an undesired style feature extraction, as original images have less clear style patterns than \\( H_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.787, + 0.669 + ], + "angle": 0, + "content": "Discussions about learning style from a single reference. We are quite surprised that One-DM, with just a single reference sample, even exceeds the generative performance of few-shot methods. We provide the potential reason analysis below. Firstly, One-DM learns a meaningful style latent space, wherein new styles can be generated based on seen styles (cf. Figure 6). Then, through our style-enhanced module, One-DM effectively extracts the style feature from a single example and maps it to a position close to the exemplary writer in the feature space, thereby producing high-quality styled handwritten text images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.69, + 0.633, + 0.705 + ], + "angle": 0, + "content": "4.4 Comparisons with SOTA Industrial Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.785, + 0.805 + ], + "angle": 0, + "content": "To highlight the superiority of our method, we compare One-DM with leading industrial image generation methods that are trained on tremendously large datasets (including numerous text-centric images), including two prominent text-to-image methods, DALL-E3 [3] and Stable Diffusion [42] (SD), and two popular style transfer methods, Artbreeder1 and IP-Adapter2 (IP-A.), on the IAM dataset. Further experiment details are in supplementary material." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.444, + 0.825 + ], + "angle": 0, + "content": "1 https://www.artbreeder.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.569, + 0.84 + ], + "angle": 0, + "content": "2 https://github.com/tencent-ailab/IP-Adapter" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.569, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.147, + 0.764, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.252, + 0.785, + 0.308 + ], + "angle": 0, + "content": "Fig. 7: (a) Qualitative comparisons between our method with SOTA industrial image generation methods, including DALL-E3 [3], Stable Diffusion [42], Artbreeder\\(^1\\) and IP-Adapter\\(^2\\) on handwritten text generation with both controllable styles and contents. (b) Qualitative comparisons with Fzshouji\\(^3\\) on ICDAR-2013 competition database [52]." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.314, + 0.482, + 0.356 + ], + "angle": 0, + "content": "Table 5: Quantitative comparisons with competitors on styled handwritten character generation in terms of FID." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.363, + 0.454, + 0.449 + ], + "angle": 0, + "content": "
MethodChineseJapanese
GANwriting [24]116.49111.86
HWT [4]165.74148.66
VATr [39]139.91124.98
WordStylist [37]34.61101.93
Ours (One-DM)27.2495.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.316, + 0.78, + 0.344 + ], + "angle": 0, + "content": "Table 6: Qualitative comparisons with VATr [39] on Chinese dataset." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.353, + 0.761, + 0.451 + ], + "angle": 0, + "content": "
GTVATrDiffusion Step
0123103050
艾思艾艾艾艾
布布布布布
琳洁淋淋淋淋
昂昂昂昂昂昂
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.601 + ], + "angle": 0, + "content": "As shown in Figure 7(a), our method excels industrial methods in style mimicry and content preservation. The performance of IP-A. is the poorest, often producing distorted images. Artbreeder can replicate the color of the strokes from style samples, but fails in content preservation. DALL-E3 and SD generate characters with accurate content, but often mismatch style details with references, such as character spacing and stroke width, with SD often generating extra backgrounds. Besides, we compare Fzshouji3, an advanced industrial method designed for Chinese handwriting generation. As shown in Figure 7(b), our method outperforms Fzshouji in replicating character details and ink color." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.534, + 0.64 + ], + "angle": 0, + "content": "4.5 Applications to Other Languages" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.785, + 0.725 + ], + "angle": 0, + "content": "In this section, we evaluate whether One-DM can be used to generate languages other than English. We further conduct experiments in the Chinese (i.e. ICDAR2013 competition database [52]) and Japanese(i.e. UP_Kuchibue database [22]) datasets. We use FID to assess generated samples of each writer, then average them. More experiment details can be seen in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.817 + ], + "angle": 0, + "content": "For the Chinese handwritten character generation task, as shown in Table 5, we find that our One-DM outperforms the second-best method by a large margin, achieving a 7.37 lower FID. From Figure 8 (a), we can observe that our One-DM generates characters that closely match the target images in geometric shape and character slant. In contrast, handwritings from HWT and VATr exhibit noticeable artifacts like blurring and collapsed structures. GANwriting tends to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.461, + 0.841 + ], + "angle": 0, + "content": "3 https://www.fzshouji.com/make" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.147, + 0.725, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.787, + 0.363 + ], + "angle": 0, + "content": "Fig. 8: Comparisons with GANwriting [24], HWT [4], VATr [39], and WordStylist [37] for styled handwritten character generation on Chinese and Japanese scripts. The red boxes highlight failures of structure preservation, while blue boxes highlight comparisons between the style patterns of targets and generated characters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.445 + ], + "angle": 0, + "content": "miss strokes in its handwritings. WordStylist sometimes struggles to accurately mimic style patterns and tends to generate characters with incorrect radicals. Table 5 and Figure 8 (b) further verify the effectiveness of One-DM for Japanese handwriting generation. We also achieve the lowest FID score, and our generated Japanese samples excel in both content preservation and style imitation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.627 + ], + "angle": 0, + "content": "We further investigate why diffusion-based methods (Our One-DM and Word-Stylist) that require only one single sample outperform few-shot GAN-based methods (i.e., GANwriting, HWT, and VATr) so significantly in generating Chinese and Japanese characters. The lower performance of GAN-based methods on Chinese and Japanese characters may stem from their vanilla convolutional architectures struggling with Chinese and Japanese characters' complex geometry, as noted in [49]. In contrast, our One-DM breaks down the generation of Chinese and Japanese characters into simpler steps. For instance, as demonstrated in Table 6, during the early stages of the diffusion generation process, the model first attempts to generate a rough Chinese handwritten character. It then continues to refine the writing style (e.g., character shape and stroke color) under the condition guidance, until it synthesizes handwritings that are satisfactory." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.36, + 0.667 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In this paper, we introduce a novel One-DM for handwritten text generation, requiring only one style reference to produce realistic handwritten text images4. We enhance style extraction by incorporating high-frequency components from the style reference. For high-frequency components with distinct style patterns, we employ laplacian contrastive learning to capture more discriminative style features. Moreover, a gate mechanism improves the transfer of informative features from the reference, reducing background noise. Our One-DM outperforms few-shot methods in multiple language scripts. In the future, we aim to explore the potential of One-DM in font generation and vector font creation tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.665, + 0.84 + ], + "angle": 0, + "content": "4 In Figure 5, the second from bottom row belongs to our method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Acknowledgments The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.248, + 0.324, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.309 + ], + "angle": 0, + "content": "1. Aksan, E., Pece, F., Hilliges, O.: Deepwriting: Making digital ink editable via deep generative modeling. In: ACM Conf. Human Factors Comput. Sys. pp. 1-14 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "2. Alonso, E., Moysset, B., Messina, R.: Adversarial generation of handwritten text images conditioned on sequences. In: ICDAR. pp. 481-486 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. 2(3), 8 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.38, + 0.785, + 0.407 + ], + "angle": 0, + "content": "4. Bhunia, A.K., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S., Shah, M.: Hand-writing transformers. In: ICCV. pp. 1086-1094 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.785, + 0.435 + ], + "angle": 0, + "content": "5. Cao, J., Mo, L., Zhang, Y., Jia, K., Shen, C., Tan, M.: Multi-marginal WassersteinianGAN. In: NeurIPS. vol. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.436, + 0.785, + 0.476 + ], + "angle": 0, + "content": "6. Chen, Z., Yang, D., Liang, J., Liu, X., Wang, Y., Peng, Z., Huang, S.: Complex handwriting trajectory recovery: Evaluation metrics and algorithm. In: ACCV. pp. 1060-1076 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.785, + 0.518 + ], + "angle": 0, + "content": "7. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: CVPR. pp. 5977-5986 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.56 + ], + "angle": 0, + "content": "8. Davis, B.L., Morse, B.S., Price, B.L., Tensmeyer, C., Wigington, C., Jain, R.: Text and style conditioned GAN for the generation of offline-handwriting lines. In: BMVC (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.785, + 0.589 + ], + "angle": 0, + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.631 + ], + "angle": 0, + "content": "0. Ding, H., Luan, B., Gui, D., Chen, K., Huo, Q.: Improving handwrittenOCR with training samples generated by glyph conditional denoising diffusion probabilistic model. In: ICDAR. pp. 20-37 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.672 + ], + "angle": 0, + "content": "1. Fogel, S., Averbuch-Elor, H., Cohen, S., Mazor, S., Litman, R.: Scrabblegan: Semi-supervised varying length handwritten text generation. In: CVPR. pp. 4324-4333 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.701 + ], + "angle": 0, + "content": "2. Gan, J., Wang, W.: Higan: Handwriting imitation conditioned on arbitrary-length texts and disentangled styles. In: AAAI. pp. 7484-7492 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.729 + ], + "angle": 0, + "content": "3. Gan, J., Wang, W., Leng, J., Gao, X.: Higan+: Handwriting imitation gan with disentangled representations. ACM TOG 42(1), 1-17 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. pp. 2672-2680 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "5. Graves, A., Fernández, S., Gomez, F., Schmidhuber, J.: Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In: ICML. pp. 369-376 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in neural information processing systems (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. vol. 33, pp. 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.717, + 0.231 + ], + "angle": 0, + "content": "19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "20. Huang, H., Yang, D., Dai, G., Han, Z., Wang, Y., Lam, K.M., Yang, F., Huang, S., Liu, Y., He, M.: Agtgan: Unpaired image translation for photographic ancient character generation. In: ACM MM. pp. 5456-5467 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "21. Huang, S., Luo, Y., Zhuang, Z., Yu, J.G., He, M., Wang, Y.: Context-aware selective label smoothing for calibrating sequence recognition model. In: ACM MM (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "22. Jaeger, S., Nakagawa, M.: Two on-line japanese character databases in unipen format. In: ICDAR. pp. 566-570 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "23. Kang, L., Riba, P., Rusinol, M., Fornes, A., Villegas, M.: Content and style aware generation of text-line images for handwriting recognition. IEEE TPAMI 44(12), 8846-8860 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "24. Kang, L., Riba, P., Wang, Y., Rusinol, M., Fornés, A., Villegas, M.: Ganwriting: content-conditioned generation of styled handwritten word images. In: ECCV. pp. 273-289 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "25. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: NeurIPS. pp. 18661-18673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "26. Khrulkov, V., Oseledets, I.: Geometry score: A method for comparing generative adversarial networks. In: International conference on machine learning. pp. 2621-2629. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.536 + ], + "angle": 0, + "content": "27. Kleber, F., Fiel, S., Diem, M., Sablatnig, R.: Cvl-database: An off-line database for writer retrieval, writer identification and word spotting. In: ICDAR. pp. 560-564 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Kotani, A., Tellex, S., Tompkin, J.: Generating handwriting via decoupled style descriptors. In: ECCV. pp. 764-780 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "29. Li, D., Chen, G., Wu, X., Yu, Z., Tan, M.: Face anti-spoofing with cross-stage relation enhancement and spoof material perception. Neural Networks 175, 106275 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "30. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR. pp. 5141-5150 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "31. Liu, Z., Chen, K., Zhang, Y., Han, J., Hong, L., Xu, H., Li, Z., Yeung, D.Y., Kwok, J.: Geom-erasing: Geometry-driven removal of implicit concept in diffusion models. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "32. Liu, Z., Jia, W., Yang, M., Luo, P., Guo, Y., Tan, M.: Deep view synthesis via self-consistent generative network. IEEE TMM 24, 451-465 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.73 + ], + "angle": 0, + "content": "33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "34. Luhman, T., Luhman, E.: Diffusion models for handwriting generation. arXiv (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "35. Luo, C., Zhu, Y., Jin, L., Li, Z., Peng, D.: Slogan: handwriting style synthesis for arbitrary-length and out-of-vocabulary text. IEEE Trans. Neural Networks Learn. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Marti, U.V., Bunke, H.: The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition 5, 39-46 (2002)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "37. Nikolaidou, K., Retsinas, G., Christlein, V., Seuret, M., Sfikas, G., Smith, E.B., Mokayed, H., Liwicki, M.: Wordstylist: Structured verbatim handwritten text generation with latent diffusion models. In: ICDAR. pp. 384-401 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "38. Pan, Z., Ji, Z., Liu, X., Bai, J., Liu, C.L.: Visa: Visual and semantic alignment for robust scene text recognition. In: ICDAR. pp. 223-242 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "39. Pippi, V., Cascianelli, S., Cucchiara, R.: Handwritten text generation from visual archetypes. In: CVPR. pp. 22458-22467 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "40. Ren, M.S., Zhang, Y.M., Wang, Q.F., Yin, F., Liu, C.L.: Diff-writer: A diffusion model-based stylized online handwritten chinese character generator. In: International Conference on Neural Information Processing. pp. 86-100 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "41. Retsinas, G., Sfikas, G., Gatos, B., Nikou, C.: Best practices for a handwritten text recognition system. In: International Workshop on Document Analysis Systems. pp. 247-259 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.37 + ], + "angle": 0, + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "44. Tang, S., Lian, Z.: Write like you: Synthesizing your cursive online chinese handwriting via metric-based meta learning. In: Comput. Graph. Forum. pp. 141-151 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "45. Tang, S., Xia, Z., Lian, Z., Tang, Y., Xiao, J.: Fontrnn: Generating large-scale chinese fonts via recurrent neural network. In: Comput. Graph. Forum. pp. 567-577 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "46. Tolosana, R., Delgado-Santos, P., Perez-Uribe, A., Vera-Rodriguez, R., Fierrez, J., Morales, A.: Deepwritesyn: On-line handwriting synthesis via deep short-term representations. In: AAAI. pp. 600-608 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "48. Xie, X., Fu, L., Zhang, Z., Wang, Z., Bai, X.: Toward understanding wordart: Corner-guided transformer for scene text recognition. In: ECCV. pp. 5456-5467 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "49. Xie, Y., Chen, X., Sun, L., Lu, Y.: Dg-font: Deformable generative networks for unsupervised font generation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.647 + ], + "angle": 0, + "content": "50. Yang, Y., Liu, D., Zhang, S., Deng, Z., Huang, Z., Tan, M.: Hilo: Detailed and robust 3d clothed human reconstruction with high-and low-frequency information of parametric models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10671-10681 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.701 + ], + "angle": 0, + "content": "51. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15901-15911 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "52. Yin, F., Wang, Q.F., Zhang, X.Y., Liu, C.L.: Icdar 2013 chinese handwriting recognition competition. In: International Conference on Document Analysis and Recognition. pp. 1464-1470 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "53. Zhang, Y., Hooi, B.: Hipa: Enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "54. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS. pp. 34077-34090 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "55. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: NeurIPS. pp. 29848-29860 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "G. Dai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "56. Zhao, B., Tao, J., Yang, M., Tian, Z., Fan, C., Bai, Y.: Deep imitator: Handwriting calligraphy imitation via deep attention networks. Pattern Recognition 104, 107080 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "57. Zhu, Y., Li, Z., Wang, T., He, M., Yao, C.: Conditional text image generation with diffusion models. In: CVPR. pp. 14235-14245 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.26 + ], + "angle": 0, + "content": "58. Zhuang, Z., Liu, Z., Lam, K.M., Huang, S., Dai, G.: A new semi-automatic annotation model via semantic boundary estimation for scene text detection. In: ICDAR. pp. 257-273 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.26 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_origin.pdf b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dc917becf845bacbd8279665b671f14b50160cea --- /dev/null +++ b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/d4f3a781-4384-437f-82b5-ab4810bd79af_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cc57a5f5acd22c08a6bbda117f5939dd3fde610b4f1c2b4b08786a7b44c5ce0 +size 3021323 diff --git a/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/full.md b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c0543dc7ef87e47ab2e0fee56b5c4d6701f6cf6 --- /dev/null +++ b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/full.md @@ -0,0 +1,307 @@ +# One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation + +Gang Dai $^{1\star}$ , Yifan Zhang $^{2,3\star}$ , Quhui Ke $^{1}$ , Qiangya Guo $^{1}$ , and Shuangping Huang $^{1,4\dagger}$ + +$^{1}$ South China University of Technology + $^{2}$ National University of Singapore, $^{3}$ Skywork AI, $^{4}$ Pazhou Laboratory + $\{\mathrm{eedaigang}@mail., \mathrm{eehsp@}\}$ scut.edu.cn, yifan.zhang@u.nus.edu + +Abstract. Existing handwritten text generation methods often require more than ten handwriting samples as style references. However, in practical applications, users tend to prefer a handwriting generation model that operates with just a single reference sample for its convenience and efficiency. This approach, known as "one-shot generation", significantly simplifies the process but poses a significant challenge due to the difficulty of accurately capturing a writer's style from a single sample, especially when extracting fine details from the characters' edges amidst sparse foreground and undesired background noise. To address this problem, we propose a One-shot Diffusion Mimicker (One-DM) to generate handwritten text that can mimic any calligraphic style with only one reference sample. Inspired by the fact that high-frequency information of the individual sample often contains distinct style patterns (e.g., character slant and letter joining), we develop a novel style-enhanced module to improve the style extraction by incorporating high-frequency components from a single sample. We then fuse the style features with the text content as a merged condition for guiding the diffusion model to produce high-quality handwritten text images. Extensive experiments demonstrate that our method can successfully generate handwriting scripts with just one sample reference in multiple languages, even outperforming previous methods using over ten samples. Our source code is available at https://github.com/dailenson/One-DM. + +Keywords: Handwritten Text Generation $\cdot$ One-Shot Generation + +# 1 Introduction + +In the digital age, handwriting text generation blends the personalization of traditional handwriting with the efficiency of automated processes, offering a digital format to preserve the authenticity of individual handwriting. This task aims to automatically generate the desired handwritten text images that not only correspond to specific text content, but also emulate the calligraphic style of a given exemplar writer (e.g., character slant, cursive join, stroke thickness, and + +![](images/1c0fe1ae5b0af618aa0a891aa277118b2e2e3c2328b433d0772e64f21898650a.jpg) +Fig. 1: User experience comparisons between one-shot and few-shot handwritten text generation methods. It reveals that one-shot setting leads to a better user experience. + +![](images/3168de4542c0ee588ddfeb45d78150b9050c1ccf0605bd8be842fa719533cf59.jpg) + +ink color). This provides great convenience for people with hand impairments and also contributes to accelerating the process of handwritten font design. + +Previous works [4,8,11-13,24,39] introduce generative adversarial networks (GANs) [5,14,32] for handwritten text generation. For instance, ScrabbleGAN [11] leverages random noises as style inputs and conditions content inputs on character-level labels, enabling the synthesis of handwritten words with randomly sampled styles. In a recent study [9], diffusion models [31,42,53] such as denoising diffusion probabilistic models (DDPM) [18] have demonstrated even higher quality of image generation compared to GANs. This motivates several attempts, such as WordStylist [37], GC-DDPM [10], and CTIG-DM [57], to condition the denoising process on the fixed writer ID to generate handwritten text images with controllable styles. However, the major limitation of these methods [10,11,37,57] is that they are unable to mimic the unseen writers' handwriting styles. + +To imitate any given handwriting style, some previous methods [4, 24, 39], known as few-shot generation, require users to provide a few samples (typically 15) as style references. They employ a style encoder to extract writing styles from the given samples, thus offering flexible control over the style of the generated samples. However, the traditional few-shot generation pipeline is inconvenient, inefficient, and time-consuming, as shown in Figure 1. Users prefer methods that only require a single sample as a style reference, known as one-shot generation methods, because they are more convenient, efficient, and easy-to-use. Our goal is to investigate the more challenging one-shot generation task that holds significant practical value, striving to achieve high-quality handwritten text image generation with desired styles and contents. + +The primary challenge of one-shot generation task lies in accurately extracting a user's handwriting style from just one style reference image. As illustrated in Figure 2, characters, being abstract symbols, typically occupy only a minor portion of the reference image. Besides, the reference image is often cluttered with noisy background information, which poses a significant obstacle in extracting the individual handwriting style. Previous one-shot generation methods [8,12,13] simply follow the architectural design of few-shot approaches, using a vanilla CNN encoder to directly extract the handwriting style from a single sample. The extracted style is then combined with text contents and input into a CNN decoder to generate the desired handwritten images. These methods ex + +![](images/702e2806fd347d3dcba7f96df7856799792949a5d80b59480d0938cffbaff0b0.jpg) +Fig. 2: Handwritten text samples and corresponding high-frequency components. We find that high-frequency components have more pronounced character contours, clearly showcasing the style patterns, such as character slant and cursive connections. + +hibit a limited performance in emulating handwriting styles, due to their poor ability of style extraction. + +To address the above challenges, our key idea revolves around utilizing the high-frequency information of the sample to enhance the extraction of the handwriting style. As depicted in Figure 2, the high-frequency information encompasses the overall contours of the handwritten text, allowing for a clear observation of key style patterns such as the text slant, letter spacing, and cursive connections. Hence, the incorporation of high-frequency information facilitates a more effective handwriting style extraction. + +In light of the above insight, we propose a One-shot Diffusion Mimicker (One-DM) for handwritten text generation, which is simultaneously guided by desired style and arbitrary content. More specifically, we first develop a style-enhanced module to process both a style reference image and its high-frequency components in parallel. Considering that the reference image often contains background noise, we design a gate mechanism to suppress the inflow of background noise. Regarding the high-frequency components, such as character slant and ligature that show clearer style patterns, we employ a contrastive learning framework [25,54,55] to further obtain discriminative style features, guiding the handwritten text synthesis with both realistic and diverse styles. Subsequently, the style features extracted from both branches are adaptively fused with the specific content prototype in a style-content fusion module. Finally, the seamlessly integrated style and content features act as conditional inputs, guiding the denoising process for the progressive synthesis of stylized handwritten text images. It is worth noting that our One-DM effectively mimics a user's writing style with only one reference sample through our style-enhanced module, surpassing few-shot methods in producing higher-quality stylized handwritings. + +To sum up, our contributions are as follows: + +- We propose a novel diffusion model for stylized handwritten text generation, which only requires a single reference sample as style input, and imitates its writing style to generate handwritten text with arbitrary content. +- We introduce the high-frequency components of the reference sample to enhance the extraction of handwriting style. The proposed style-enhanced + +module can effectively capture the writing style patterns and suppress the interference of background noise. + +- Extensive experiments on handwriting datasets in English, Chinese, and Japanese demonstrate that our approach with a single style reference even outperforms previous methods with 15x-more references. + +# 2 Related Work + +Handwriting generation. Handwritten text is typically stored in two formats: online trajectory form or offline image form. Online handwriting generation methods [1, 7, 28, 44, 46, 56] often employ Recurrent Neural Networks (RNNs) [1, 6, 28, 44-46, 56], transformer decoders [7], or diffusion models [34, 40] to progressively generate writing trajectories. Unlike online methods, offline generation methods have the advantage of not requiring additional trajectory supervision information and can generate realistic handwritten text with stroke width and ink color, which online methods cannot produce. + +Previous offline handwriting generation methods [2, 4, 8, 11, 20, 24, 35, 39] in deep learning predominantly rely on Generative Adversarial Networks (GANs). Early works [2, 11] condition the generative process on the word embeddings [2] or concatenated letter-tokens [11] to synthesize handwritten word images. However, these methods struggle to flexibly control the writing style. Thus, few-shot approaches [4, 24, 39] which rely on 15 style references are introduced. For instance, GANwriting [24] utilizes a CNN encoder [16, 50, 51] to extract a user's handwriting style from a few samples, which are then combined with specific text content to generate handwritings in the desired style. In its follow-up work [23], the synthesized samples are demonstrated to assist in training more robust handwritten text recognizers. Further, HWT [4] employs a transformer encoder [29, 47] to extract rich style patterns from reference samples, enhancing the performance in style mimicry. Recently, VATr [39] uses images of symbols as content representations, enabling the generation of out-of-charset characters. + +Concurrently, one-shot generation methods [8, 12, 13] are proposed. Despite their ability to mimic handwriting styles with only a single sample, these techniques still lag behind few-shot methods in terms of stylized generation results. Additionally, previous handwriting generation methods [2, 4, 8, 11-13, 24] often rely on complex content representations, such as recurrent embeddings [2] and letter-level tokens [4, 8, 11-13, 24]. SLOGAN [35] extracts textual contents from easily obtainable printed images. However, it faces challenges in generalizing to unseen writing styles due to its fixed writer ID. Similarly, some diffusion-based [10, 37, 57] methods condition the denoising process on fixed style ID and are unable to mimic styles that they have not previously encountered. In contrast, our One-DM effectively obtains style information from one style sample and thus can generate handwritings with arbitrary styles. Due to space constraints, we discuss more related works in supplementary, including diffusion methods for general image generation, contrastive-based method [48], and frequency-based methods [30, 38]. + +![](images/c2a04fd867a48f0f68059db58c5c9dea90a83773872be2b1f5435c6398ca6766.jpg) +Fig. 3: Overview of the proposed method. The style reference initially passes through a high-pass filter to extract its high-frequency components. Subsequently, the spatial and the high-frequency style encoders independently extract style features $F_{spa}$ and $F_{fre}$ from the style reference and its high-frequency information, respectively. $F_{spa}$ , after being filtered through a gate mechanism, is fused with $F_{fre}$ and content features $E$ in the fusion module. The merged feature then serves as a condition input to guide the diffusion generation process. + +# 3 Method + +Problem statement. We aim to synthesize handwritten text images that are controlled by both textual content and handwriting style. Given any word string $\mathcal{A}$ and a single style sample $I_{s}$ from a writer $w_{s}$ , the generated handwritten word image $X_{s}$ should replicate the unique handwriting style of $w_{s}$ while maintaining the content of $\mathcal{A}$ , where the textual content $\mathcal{A} = \{a_i\}_{i=1}^L$ spans a length $L$ and $a_{i}$ is from a character set without any restrictions. + +To address this task, we incorporate high-frequency information to enhance the extraction of writing styles. Existing methods [4,8,12,13,24,39] typically use a vanilla CNN or transformer encoder to directly process style images, often resulting in undesired style extraction. In contrast, we introduce a novel One-shot diffusion mimicker (One-DM) by innovating Laplacian high-frequency extraction and a gating mechanism. Our One-DM can effectively capture style features from a single reference while suppressing background noise. + +# 3.1 Overall Scheme + +Our high-level idea focuses on incorporating high-frequency information from style reference images to enhance the extraction of style patterns. A straightforward implementation involves using a vanilla transformer encoder to extract style features from both the style image and its corresponding high-frequency image. This naive solution encounters two main issues: (1) the lack of efficient supervision objective still makes it challenging to accurately learn the writer's style patterns from the high-frequency image, and (2) the style features captured from the original image still retain undesirable noise backgrounds, which can adversely affect subsequent image generation performance. + +To address the above-mentioned issues, we develop a more effective method, shown in Figure 3. Our method consists of a style-enhanced module, a content encoder, a style-content fusion module, and a conditional diffusion module. Initially, we utilize the Laplacian kernel as a high-frequency filter to extract the high-frequency components from the style reference. Then, two parallel style encoders are employed to simultaneously extract the corresponding style features from both style reference and its high-frequency information. Since undesirable background noise is often present in style references, we design a gate mechanism to facilitate the transmission of informative style information while mitigating noise transmission. The style patterns in the high-frequency components are comparatively cleaner and more distinct, facilitating the observation of individual styles, such as character slant. For the observation, we propose a contrastive learning objective, termed LaplacianNCE $\mathcal{L}_{lapNCE}$ , to enforce the more discriminative style learning from high-frequency components. + +Regarding content guidance, we render the given string $\mathcal{A}$ into a unifont image, as done in VATr [39]. Briefly, Unifont's key advantage is that it covers all Unicode characters, allowing our method to transform any user inputs into corresponding images. We further feed render results into a content encoder that combines a ResNet18 [16] with a Transformer encoder. This process first involves using a ResNet18 to handle each character image in parallel, followed by concatenating them to form word sequence features. The Transformer encoder then processes these features to extract an informative content feature $E = \{e_i\}_{i=1}^L \in \mathbb{R}^{L \times c}$ with a global context, where $c$ is the channel dimension. After obtaining style features and content guidance, we seamlessly fusion them using a style-content fusion module. The fused results then guide the denoising process of the conditional diffusion model to synthesize the desired handwritten text images. The denoising process is supervised by the reconstruction loss $\mathcal{L}_{rec}$ . + +To summarize, the overall training objective of our method combines all two loss functions: + +$$ +\mathcal {L} = \mathcal {L} _ {l a p N C E} + \mathcal {L} _ {r e c}. +$$ + +# 3.2 Style-enhanced Module + +We propose the style-enhanced module to enhance style extraction by incorporating high-frequency components $H_{s} \in \mathbb{R}^{h \times w \times c}$ from a reference image $I_{s} \in \mathbb{R}^{h \times w \times c}$ since clearer style patterns are presented in high-frequency components like character slanting and shape connections (cf. Figure 2). As shown in Figure 3, we use a Laplacian kernel as a high-frequency filter to extract $H_{s}$ from $I_{s}$ . The Laplacian kernel excels in extracting high-frequency information without the need for Fast Fourier Transform (FFT) and parametric separation in the frequency domain. Then, two style encoders, $\mathcal{E}_{spa}$ and $\mathcal{E}_{fre}$ , each a combination of CNN and transformer, process the $I_{s}$ and $H_{s}$ , respectively. This independent processing leads to the extraction of distinct style features: $F_{spa} = \{f_{spa}^{i}\}_{i=1}^{d} \in \mathbb{R}^{d \times c}$ from $\mathcal{E}_{spa}$ and $F_{fre} = \{f_{fre}^{i}\}_{i=1}^{d} \in \mathbb{R}^{d \times c}$ from $\mathcal{E}_{fre}$ , where $d = h \times w$ . While structurally identical, $\mathcal{E}_{spa}$ and $\mathcal{E}_{fre}$ do not share weights with each other. Then, the proposed $\mathcal{L}_{lapNCE}$ forces $\mathcal{E}_{fre}$ to focus on extracting discriminative style features + +from $H_{s}$ . A gate mechanism is designed to selectively filter out background noise from the reference style features, allowing only meaningful style patterns to pass. + +Laplacian Contrastive Learning. The goal of the proposed $\mathcal{L}_{\text{lapNCE}}$ is to guide the high-frequency style encoder $\mathcal{E}_{\text{fre}}$ in learning more discriminative style features from the high-frequency information. Thus, we propose to bring closer extracted style features $F_{\text{fre}}$ belonging to the same writer, while distancing those from different writers. We formulate our $\mathcal{L}_{\text{lapNCE}}$ as follows: + +$$ +\mathcal {L} _ {\text {l a p N C E}} = - \frac {1}{N} \sum_ {i \in M} \frac {1}{| P (i) |} \sum_ {p \in P (i)} \log \frac {\exp \left(z _ {i} \cdot z _ {p} / \tau\right)}{\sum_ {a \in A (i)} \exp \left(z _ {i} \cdot z _ {a} / \tau\right)}. \tag {1} +$$ + +In detail, $i \in M = \{1 \dots N\}$ is the index of any element in the mini-batch with size $N$ and $A(i) = M \setminus \{i\}$ is other indices distinct from $i$ . $z_{i}$ is an anchor sample belonging to writer $w_{i}$ and $P(i) = \{p \in A(i) : w_{p} = w_{i}\}$ is its in-batch positive sample set and the other $A(i) \setminus P(i)$ is its negative set. Here, $z = Proj(F_{fre})$ , where $Proj$ is a learnable multi-layer perceptron (MLP), $\tau$ is a scalar temperature parameter and the $\cdot$ symbol denotes the inner product. + +Gate mechanism. As illustrated in Figure 2, the stroke areas of characters in reference images are typically sparse, with background noise interfering with the extraction of character style features. To address this challenge, we propose a gate mechanism to selectively filter the information of the sample $I_{s}$ , as shown in Figure 3. Specifically, the extracted sample style features $F_{spa}$ are fed into a gate layer, consisting of a learnable fully connected layer and followed by a sigmoid activation, to obtain the corresponding gate units $W = \{w_{i}\}_{i=1}^{d} \in \mathbb{R}^{d}$ . Each unit $w_{i}$ determines the pass rate for the corresponding $f_{spa}^{i}$ , allowing for a higher pass rate where $w_{i}$ is larger. This design effectively enables informative style features $\hat{F}_{spa} = \{\hat{f}_{spa}^{i}\}_{i=1}^{d}$ to be extracted while suppressing extraneous background noise, where $\hat{f}_{spa}^{i} = f_{spa}^{i} \cdot w_{i}$ . + +# 3.3 Style-content Fusion Module + +Upon acquiring textual content feature $E$ and two style features $\hat{F}_{spa}$ and $F_{fre}$ , we integrate all features within two multi-head attention mechanisms to guide the denoising generation process of the diffusion model, as shown in Figure 3. Specifically, the first cross-attention module uses the textual content $E$ as queries to identify the most relevant style information in the style reference, thereby inferring the style attributes corresponding to each character. For instance, if the textual content is 'a', it prioritizes searching for style features of characters like 'a', 'b', 'd', 'g' in the style reference, due to these characters appearing similar looped structures, implying more comparable style attributes. This process (cross-attention in Figure 3) is represented as: + +$$ +O = \operatorname {A t t e n} _ {1} \left(Q _ {1} = E, K _ {1} = V _ {1} = \hat {F} _ {s p a} + F _ {f r e}\right). \tag {2} +$$ + +Subsequently, we obtain the initial fusion embedding between content and style guidance by simply summing $O$ and $E$ . The merged intermediate vector is then + +employed as the query, key, and value in the self-attention mechanism to facilitate comprehensive interaction of information. Finally, the blended embedding $g$ serves as the condition of the diffusion process. The second multi-head attention (self-attention in Figure 3) is defined as: + +$$ +g = \operatorname {A t t e n} _ {2} \left(Q _ {2} = K _ {2} = V _ {2} = O + E\right). \tag {3} +$$ + +# 3.4 Conditional Diffusion Model + +The goal of the conditional diffusion model $p_{\theta}$ is to generate realistic images of handwritten text, guided by acquired conditions $g$ . Specifically, as shown in Figure 3, under the guidance of $g$ , $p_{\theta}$ executes a denoising generative process with $T$ steps, starting from a sampled Gaussian noise $x_{T}$ and progressively denoises to obtain the desired handwritten text $x_{0}$ : + +$$ +p _ {\theta} \left(x _ {0} \mid g\right) = \int p _ {\theta} \left(x _ {0: T} \mid g\right) d _ {x _ {1: T}}, \tag {4} +$$ + +$$ +p _ {\theta} \left(x _ {0: T} \mid g\right) = p \left(x _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, g\right), \tag {5} +$$ + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, g\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, g, t\right), \Sigma_ {\theta} \left(x _ {t}, g, t\right)\right). \tag {6} +$$ + +The denoising process aims to learn how to reverse a predefined forward process, as described in DDPM [18]. The forward process is modeled as a fixed Markov chain, where noise conforming to a normal distribution is incrementally added to $x_{t - 1}$ to derive $x_{t}$ . This can be mathematically expressed as: + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {7} +$$ + +where the noise is characterized by a variance schedule $\beta_{t}$ . During training, a variational bound on the maximum likelihood objective is applied to guide the generation process to recover $x_0$ from the standard Gaussian noise $x_{T}$ conditioned on $g$ . We give the training objective as follows: + +$$ +\mathcal {L} _ {r e c} = \mathbb {E} _ {t, q} \| \mu_ {t} \left(x _ {t}, x _ {0}\right) - \mu_ {\theta} \left(x _ {t}, g, t\right) \| _ {2} ^ {2}, \tag {8} +$$ + +where $\mu_t(x_t, x_0)$ is the mean of the forward process posterior $q(x_t | x_{t-1})$ , which has a closed-form solution [18]. We put more details in supplementary material. + +# 4 Experiments + +# 4.1 Experimental Settings + +Dataset. To evaluate our One-DM in handwritten text generation, we use the widely-used handwriting dataset IAM [36] and CVL [27].IAM consists of 62,857 English word images from 500 unique writers. Following [4,24,39], we use words + +from 339 writers for training and the remaining 161 writers for testing. CVL includes texts by 310 writers in English and German. We use the English portion, totaling 84,514 words, and follow CVL's standard split, with 283 writers for training and 27 for testing. Throughout all experiments, we resize images to 64 pixels in height, preserving their aspect ratio. + +Evaluation metrics. We use the Fréchet Inception Distance (FID) [17] and Geometry Score (GS) [26] to assess the generation quality, following the settings of [4, 24, 39]. We also conduct user studies to quantify the subjective quality of the generated handwritten text images in supplementary. + +Implementation details. In our all experiments, we only use single style reference sample. We first train our model for 700 epochs (batch size of 384) with classifier-free guidance strategy [19] under the guidance scale of 0.25, and then fine-tune the model for 4500 iterations (batch size of 128) using a text recognizer [41] with a CTC loss [15, 21, 58], on four RTX3090 GPUs. The fine-tuning process forces our One-DM to generate readable text with accurate content. The optimizer is AdamW [33], with a learning rate of $10^{-4}$ . During inference, each style sample is randomly sampled from the target-writer. To speed up the sampling, we use the denoising diffusion implicit model (DDIM) [43] with 50 steps. More details are provided in supplementary material. + +Compared methods. We compare our One-DM with state-of-the-art handwritten text generation methods, including GAN-based methods (i.e., GANwriting [24], HWT [4], VATr [39], TS-GAN [8], HiGAN+ [13]), and diffusion-based methods (i.e. GC-DDPM [10] and WordStylist [37]). For a fair comparison, we configure all methods to generate images with a height of 64 pixels in supplementary. Additionally, in supplementary, we compare a variant of One-DM that generates 32-pixel high images with official VATr [39] and HWT [4]. + +# 4.2 Main Results + +Stylish Handwritten Text Generation. Initially, we assess our One-DM for producing styled handwritten text images, aiming to replicate both the style and content in the generated images. Following [4, 24, 39], we first calculate FID between generated and real samples for each writer separately and finally average them. In line with the previous works [4, 24, 39], our experiments on the IAM dataset are divided into four distinct scenarios: IV-S, IV-U, OOV-S, OOV-U. Among these four scenarios, OOV-U represents the most challenging case where both the target style and the words are entirely unseen during training. For the CVL dataset, we directly report the results of all methods on the test set. + +We first report the quantitative results on the IAM dataset in Table 1. We can observe that our One-DM outperforms all the competitors in all settings. Notably, it significantly exceeds one-shot methods in all scenarios. Impressively, our One-DM also holds a substantial advantage over few-shot methods (GAN-writing [24], HWT [4], VATr [39]), which use 15-x more reference samples for style guidance, in IV-S and OOV-S settings. Even in the most challenging OOV-U scenario, our One-DM leads the second-best, VATr, by a large margin (102.75 vs. 108.76), demonstrating the superior performance of our One-DM in stylized + +Table 1: Comparisons with state-of-the-art methods on styled and style-agnostic handwritten text generation in the IAM dataset. Note that, all methods are trained on the same training set used in GANwriting [24], HWT [4], and VATr [39]. + +
MethodShotStyloid EvaluationStyle-agnostic
IV-SIV-UOOV-SOOV-UFID↓GS↓
TS-GAN [8]One118.56128.75127.11136.6720.654.88×10-2
GANwriting [24]Few120.07124.30125.87130.6828.375.67×10-2
HiGAN+ [13]One117.33116.95121.55121.4822.952.06×10-2
GC-DDPM [10]One99.86105.73112.52118.3919.051.31×10-2
WordStylist [37]One98.10104.27109.45115.5218.582.85×10-2
HWT [4]Few109.25106.90116.55113.5218.994.41×10-3
VATr [39]Few103.75101.73111.64108.7616.031.74×10-2
Ours (One-DM)One89.4798.3693.30102.7515.731.98×10-3
+ +handwritten text generation. Similarly, our method outperforms HWT and VATr on the CVL dataset, achieving the lowest FID score, as shown in Table 4. + +We provide qualitative results to intuitively explain the benefit of our One-DM, in Figure 4. GANwriting struggles to capture the style patterns of reference samples, such as character slant and occasionally produces unclear character shapes. HiGAN+ more consistently generates characters with correct content, but the character spacing within the generated words lacks realism. WordStylist typically produces images with noticeable background noise. HWT and VATr can produce satisfactory handwritten words in terms of content accuracy and style mimicry; however, their downside is the tendency to create smoother character appearances. Compared to HWT and VATr, our synthesized samples excel in more authentic character ink color and stroke thickness. However, some generated samples by our One-DM are visibly different in ink color. We provide comprehensive explanations in supplementary. We further display more qualitative comparisons between our method and few-shot methods in Figure 5. + +Style-agnostic Handwritten Text Generation. We further evaluate our One-DM to generate realistic handwritten text images, irrespective of style imitation. For this purpose, we compute the FID and GS on the IAM test set, under the same conditions as applied to ScrabbleGAN [11] (FID: 20.72, GS: $2.56 \times 10^{-2}$ ), capable of generating handwritten texts with randomly sampled styles. Specifically, each method generates 25k random samples to calculate FID against 25k samples from the test set, and 5k random samples for GS calculation in comparison with 5k test set samples. As presented in Table 1, our One-DM achieves the best results in both FID and GS metrics, further demonstrating its capability to generate higher-quality handwritten text images. + +# 4.3 Analysis + +In this section, we conduct ablation studies to analyze our One-DM. More analyses are provided in supplementary, including generalization evaluation on different style backgrounds, generation quality assessment through OCR performance, failure case analysis, and the effects of different designs (e.g., high-frequency filter, style-content fusion mechanism, and style input sample length). + +
Style examplesSee us next him unless some eight have the be is unassailable of our boxInour there they on as had inside are shaw have us God is unprouctand tuning which oil to you will the them may the rivers Rim Fishand to is all Common Wood Lessis of ther in work and it hean of
GANw.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*HiGAN+The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*WordS.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
HWTThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
VATrThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*OursThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
+ +Fig. 4: Qualitative comparisons between our method with state-of-the-art methods on handwritten text generation with both specific textual content and desired handwriting style in the IAM dataset. We utilize the identical guiding text, 'The greatest test of courage on earth is to bear defeat without losing heart,' across all handwriting generation methods, directing them to produce text in varied styles. Better zoom in $200\%$ . * denotes one-shot methods, while others are few-shot methods. + +
Targethopetherelatersuccess
Generated Sampleshopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
+ +Fig. 5: Each row shows results from our One-DM and few-shot methods on IAM dataset; readers are invited to identify our method. The answer is at the paper's end. + +
TextStyle A··· interpolation··· Style B
plantsplants plnph plnch plnns plnns plnns plnns
cloudcloud cloud cloud cloud cloud cloud cloud
stickstick stick stick stick stick stick stick
photophoto photo photo photo photo photo photo
vaporVapor vapor vapor vapor vapor vapor vapor vapor
listslists lists lists lists lists lists lists
+ +Fig. 6: We provide style interpolation results generated by our One-DM. Different individual writing styles are extracted from the IAM test dataset. + +Quantitative evaluation of Laplacian branch and gate mechanism. We conduct various ablation experiments on the IAM dataset to evaluate the effect of distinct components within our approach. We provide the quantitative result in Table 2. We find that: (1) The inclusion of both Laplacian branch and gate mechanism enhances the quality of the generated handwritten text images, improving FID by 3.92 and 2.71, respectively. (2) Integrating the Laplacian branch with the gate mechanism further boosts the generative performance. + +Qualitative evaluation of Laplacian branch and gate mechanism. To further analyze each module in our One-DM, we conduct visual ablation experiments. As shown in Table 2, we can observe that firstly, after adding a gate mechanism, background noise can be somewhat suppressed, resulting in characters with relatively clean backgrounds. Then, the independent addition of the Laplacian branch helps the model learn cursive connections and other + +Table 2: Ablation study. Effect of the Laplacian branch and gate mechanism on the IAM dataset under the OOV-U setting, as done in [4]. In the middle, we showcase the generated samples of each component. + +
Base Εfre GateStyle samplesFID ↓
would like
✓cloudfoul108.44
✓cloudfoul105.73
✓cloudfoul104.52
✓cloudfoul102.75
+ +Table 3: Effect of each part of our Laplacian branch on IAM dataset under OOV-U setting. + +
HsLapNCEFID ↓
108.44
✓106.16
✓107.58
✓✓104.52
+ +Table 4: Comparisons with competitors on the CVL dataset. + +
MethodFID ↓
HWT [4]58.22
VATr [39]54.44
Ours (One-DM)51.78
+ +style patterns. Finally, our method integrated the Laplacian branch and gate mechanism, which can generate the highest quality handwritten text images. + +Discussions of the Laplacian branch. Our Laplacian branch consists of two key components: utilizing high-frequency images $H_{s}$ and the Laplacian contrastive learning loss $\mathcal{L}_{lapNCE}$ . In the previous ablation study, they are always combined. We further conduct exploration experiments to explain why they cannot be separated. As reported in Table 3, combining $H_{s}$ and $\mathcal{L}_{lapNCE}$ maximizes effectiveness; separating them significantly reduces performance. Without the guidance of $\mathcal{L}_{lapNCE}$ , extracting discriminative features from $H_{s}$ is challenging. Likewise, directly applying $\mathcal{L}_{lapNCE}$ on original images leads to an undesired style feature extraction, as original images have less clear style patterns than $H_{s}$ . + +Discussions about learning style from a single reference. We are quite surprised that One-DM, with just a single reference sample, even exceeds the generative performance of few-shot methods. We provide the potential reason analysis below. Firstly, One-DM learns a meaningful style latent space, wherein new styles can be generated based on seen styles (cf. Figure 6). Then, through our style-enhanced module, One-DM effectively extracts the style feature from a single example and maps it to a position close to the exemplary writer in the feature space, thereby producing high-quality styled handwritten text images. + +# 4.4 Comparisons with SOTA Industrial Methods + +To highlight the superiority of our method, we compare One-DM with leading industrial image generation methods that are trained on tremendously large datasets (including numerous text-centric images), including two prominent text-to-image methods, DALL-E3 [3] and Stable Diffusion [42] (SD), and two popular style transfer methods, Artbreeder1 and IP-Adapter2 (IP-A.), on the IAM dataset. Further experiment details are in supplementary material. + +![](images/805e3e0f4589677c79e85db22cc210f5c31488d5e288123592d6156ca553e840.jpg) +Fig. 7: (a) Qualitative comparisons between our method with SOTA industrial image generation methods, including DALL-E3 [3], Stable Diffusion [42], Artbreeder $^1$ and IP-Adapter $^2$ on handwritten text generation with both controllable styles and contents. (b) Qualitative comparisons with Fzshouji $^3$ on ICDAR-2013 competition database [52]. + +Table 5: Quantitative comparisons with competitors on styled handwritten character generation in terms of FID. + +
MethodChineseJapanese
GANwriting [24]116.49111.86
HWT [4]165.74148.66
VATr [39]139.91124.98
WordStylist [37]34.61101.93
Ours (One-DM)27.2495.43
+ +Table 6: Qualitative comparisons with VATr [39] on Chinese dataset. + +
GTVATrDiffusion Step
0123103050
艾思艾艾艾艾
布布布布布
琳洁淋淋淋淋
昂昂昂昂昂昂
+ +As shown in Figure 7(a), our method excels industrial methods in style mimicry and content preservation. The performance of IP-A. is the poorest, often producing distorted images. Artbreeder can replicate the color of the strokes from style samples, but fails in content preservation. DALL-E3 and SD generate characters with accurate content, but often mismatch style details with references, such as character spacing and stroke width, with SD often generating extra backgrounds. Besides, we compare Fzshouji3, an advanced industrial method designed for Chinese handwriting generation. As shown in Figure 7(b), our method outperforms Fzshouji in replicating character details and ink color. + +# 4.5 Applications to Other Languages + +In this section, we evaluate whether One-DM can be used to generate languages other than English. We further conduct experiments in the Chinese (i.e. ICDAR2013 competition database [52]) and Japanese(i.e. UP_Kuchibue database [22]) datasets. We use FID to assess generated samples of each writer, then average them. More experiment details can be seen in supplementary material. + +For the Chinese handwritten character generation task, as shown in Table 5, we find that our One-DM outperforms the second-best method by a large margin, achieving a 7.37 lower FID. From Figure 8 (a), we can observe that our One-DM generates characters that closely match the target images in geometric shape and character slant. In contrast, handwritings from HWT and VATr exhibit noticeable artifacts like blurring and collapsed structures. GANwriting tends to + +![](images/1533947c780ae3abe8b5c355723de45060f1dc2460e499931786197299abaf9f.jpg) +Fig. 8: Comparisons with GANwriting [24], HWT [4], VATr [39], and WordStylist [37] for styled handwritten character generation on Chinese and Japanese scripts. The red boxes highlight failures of structure preservation, while blue boxes highlight comparisons between the style patterns of targets and generated characters. + +miss strokes in its handwritings. WordStylist sometimes struggles to accurately mimic style patterns and tends to generate characters with incorrect radicals. Table 5 and Figure 8 (b) further verify the effectiveness of One-DM for Japanese handwriting generation. We also achieve the lowest FID score, and our generated Japanese samples excel in both content preservation and style imitation. + +We further investigate why diffusion-based methods (Our One-DM and Word-Stylist) that require only one single sample outperform few-shot GAN-based methods (i.e., GANwriting, HWT, and VATr) so significantly in generating Chinese and Japanese characters. The lower performance of GAN-based methods on Chinese and Japanese characters may stem from their vanilla convolutional architectures struggling with Chinese and Japanese characters' complex geometry, as noted in [49]. In contrast, our One-DM breaks down the generation of Chinese and Japanese characters into simpler steps. For instance, as demonstrated in Table 6, during the early stages of the diffusion generation process, the model first attempts to generate a rough Chinese handwritten character. It then continues to refine the writing style (e.g., character shape and stroke color) under the condition guidance, until it synthesizes handwritings that are satisfactory. + +# 5 Conclusion + +In this paper, we introduce a novel One-DM for handwritten text generation, requiring only one style reference to produce realistic handwritten text images4. We enhance style extraction by incorporating high-frequency components from the style reference. For high-frequency components with distinct style patterns, we employ laplacian contrastive learning to capture more discriminative style features. Moreover, a gate mechanism improves the transfer of informative features from the reference, reducing background noise. Our One-DM outperforms few-shot methods in multiple language scripts. In the future, we aim to explore the potential of One-DM in font generation and vector font creation tasks. + +Acknowledgments The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016). + +# References + +1. Aksan, E., Pece, F., Hilliges, O.: Deepwriting: Making digital ink editable via deep generative modeling. In: ACM Conf. Human Factors Comput. Sys. pp. 1-14 (2018) +2. Alonso, E., Moysset, B., Messina, R.: Adversarial generation of handwritten text images conditioned on sequences. In: ICDAR. pp. 481-486 (2019) +3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. 2(3), 8 (2023) +4. Bhunia, A.K., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S., Shah, M.: Hand-writing transformers. In: ICCV. pp. 1086-1094 (2021) +5. Cao, J., Mo, L., Zhang, Y., Jia, K., Shen, C., Tan, M.: Multi-marginal WassersteinianGAN. In: NeurIPS. vol. 32 (2019) +6. Chen, Z., Yang, D., Liang, J., Liu, X., Wang, Y., Peng, Z., Huang, S.: Complex handwriting trajectory recovery: Evaluation metrics and algorithm. In: ACCV. pp. 1060-1076 (2022) +7. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: CVPR. pp. 5977-5986 (2023) +8. Davis, B.L., Morse, B.S., Price, B.L., Tensmeyer, C., Wigington, C., Jain, R.: Text and style conditioned GAN for the generation of offline-handwriting lines. In: BMVC (2020) +9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021) +0. Ding, H., Luan, B., Gui, D., Chen, K., Huo, Q.: Improving handwrittenOCR with training samples generated by glyph conditional denoising diffusion probabilistic model. In: ICDAR. pp. 20-37 (2023) +1. Fogel, S., Averbuch-Elor, H., Cohen, S., Mazor, S., Litman, R.: Scrabblegan: Semi-supervised varying length handwritten text generation. In: CVPR. pp. 4324-4333 (2020) +2. Gan, J., Wang, W.: Higan: Handwriting imitation conditioned on arbitrary-length texts and disentangled styles. In: AAAI. pp. 7484-7492 (2021) +3. Gan, J., Wang, W., Leng, J., Gao, X.: Higan+: Handwriting imitation gan with disentangled representations. ACM TOG 42(1), 1-17 (2022) +4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. pp. 2672-2680 (2014) +5. Graves, A., Fernández, S., Gomez, F., Schmidhuber, J.: Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In: ICML. pp. 369-376 (2006) +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) + +17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in neural information processing systems (2017) +18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. vol. 33, pp. 6840-6851 (2020) +19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: arXiv (2022) +20. Huang, H., Yang, D., Dai, G., Han, Z., Wang, Y., Lam, K.M., Yang, F., Huang, S., Liu, Y., He, M.: Agtgan: Unpaired image translation for photographic ancient character generation. In: ACM MM. pp. 5456-5467 (2022) +21. Huang, S., Luo, Y., Zhuang, Z., Yu, J.G., He, M., Wang, Y.: Context-aware selective label smoothing for calibrating sequence recognition model. In: ACM MM (2021) +22. Jaeger, S., Nakagawa, M.: Two on-line japanese character databases in unipen format. In: ICDAR. pp. 566-570 (2001) +23. Kang, L., Riba, P., Rusinol, M., Fornes, A., Villegas, M.: Content and style aware generation of text-line images for handwriting recognition. IEEE TPAMI 44(12), 8846-8860 (2021) +24. Kang, L., Riba, P., Wang, Y., Rusinol, M., Fornés, A., Villegas, M.: Ganwriting: content-conditioned generation of styled handwritten word images. In: ECCV. pp. 273-289 (2020) +25. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: NeurIPS. pp. 18661-18673 (2020) +26. Khrulkov, V., Oseledets, I.: Geometry score: A method for comparing generative adversarial networks. In: International conference on machine learning. pp. 2621-2629. PMLR (2018) +27. Kleber, F., Fiel, S., Diem, M., Sablatnig, R.: Cvl-database: An off-line database for writer retrieval, writer identification and word spotting. In: ICDAR. pp. 560-564 (2013) +28. Kotani, A., Tellex, S., Tompkin, J.: Generating handwriting via decoupled style descriptors. In: ECCV. pp. 764-780 (2020) +29. Li, D., Chen, G., Wu, X., Yu, Z., Tan, M.: Face anti-spoofing with cross-stage relation enhancement and spoof material perception. Neural Networks 175, 106275 (2024) +30. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR. pp. 5141-5150 (2021) +31. Liu, Z., Chen, K., Zhang, Y., Han, J., Hong, L., Xu, H., Li, Z., Yeung, D.Y., Kwok, J.: Geom-erasing: Geometry-driven removal of implicit concept in diffusion models. arXiv (2023) +32. Liu, Z., Jia, W., Yang, M., Luo, P., Guo, Y., Tan, M.: Deep view synthesis via self-consistent generative network. IEEE TMM 24, 451-465 (2021) +33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017) +34. Luhman, T., Luhman, E.: Diffusion models for handwriting generation. arXiv (2020) +35. Luo, C., Zhu, Y., Jin, L., Li, Z., Peng, D.: Slogan: handwriting style synthesis for arbitrary-length and out-of-vocabulary text. IEEE Trans. Neural Networks Learn. Syst. (2022) +36. Marti, U.V., Bunke, H.: The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition 5, 39-46 (2002) + +37. Nikolaidou, K., Retsinas, G., Christlein, V., Seuret, M., Sfikas, G., Smith, E.B., Mokayed, H., Liwicki, M.: Wordstylist: Structured verbatim handwritten text generation with latent diffusion models. In: ICDAR. pp. 384-401 (2023) +38. Pan, Z., Ji, Z., Liu, X., Bai, J., Liu, C.L.: Visa: Visual and semantic alignment for robust scene text recognition. In: ICDAR. pp. 223-242 (2023) +39. Pippi, V., Cascianelli, S., Cucchiara, R.: Handwritten text generation from visual archetypes. In: CVPR. pp. 22458-22467 (2023) +40. Ren, M.S., Zhang, Y.M., Wang, Q.F., Yin, F., Liu, C.L.: Diff-writer: A diffusion model-based stylized online handwritten chinese character generator. In: International Conference on Neural Information Processing. pp. 86-100 (2023) +41. Retsinas, G., Sfikas, G., Gatos, B., Nikou, C.: Best practices for a handwritten text recognition system. In: International Workshop on Document Analysis Systems. pp. 247-259 (2022) +42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) +44. Tang, S., Lian, Z.: Write like you: Synthesizing your cursive online chinese handwriting via metric-based meta learning. In: Comput. Graph. Forum. pp. 141-151 (2021) +45. Tang, S., Xia, Z., Lian, Z., Tang, Y., Xiao, J.: Fontrnn: Generating large-scale chinese fonts via recurrent neural network. In: Comput. Graph. Forum. pp. 567-577 (2019) +46. Tolosana, R., Delgado-Santos, P., Perez-Uribe, A., Vera-Rodriguez, R., Fierrez, J., Morales, A.: Deepwritesyn: On-line handwriting synthesis via deep short-term representations. In: AAAI. pp. 600-608 (2021) +47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +48. Xie, X., Fu, L., Zhang, Z., Wang, Z., Bai, X.: Toward understanding wordart: Corner-guided transformer for scene text recognition. In: ECCV. pp. 5456-5467 (2022) +49. Xie, Y., Chen, X., Sun, L., Lu, Y.: Dg-font: Deformable generative networks for unsupervised font generation. In: CVPR (2021) +50. Yang, Y., Liu, D., Zhang, S., Deng, Z., Huang, Z., Tan, M.: Hilo: Detailed and robust 3d clothed human reconstruction with high-and low-frequency information of parametric models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10671-10681 (2024) +51. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15901-15911 (2023) +52. Yin, F., Wang, Q.F., Zhang, X.Y., Liu, C.L.: Icdar 2013 chinese handwriting recognition competition. In: International Conference on Document Analysis and Recognition. pp. 1464-1470 (2013) +53. Zhang, Y., Hooi, B.: Hipa: Enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv (2023) +54. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS. pp. 34077-34090 (2022) +55. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: NeurIPS. pp. 29848-29860 (2021) + +56. Zhao, B., Tao, J., Yang, M., Tian, Z., Fan, C., Bai, Y.: Deep imitator: Handwriting calligraphy imitation via deep attention networks. Pattern Recognition 104, 107080 (2020) +57. Zhu, Y., Li, Z., Wang, T., He, M., Yao, C.: Conditional text image generation with diffusion models. In: CVPR. pp. 14235-14245 (2023) +58. Zhuang, Z., Liu, Z., Lam, K.M., Huang, S., Dai, G.: A new semi-automatic annotation model via semantic boundary estimation for scene text detection. In: ICDAR. pp. 257-273 (2021) \ No newline at end of file diff --git a/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/images.zip b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b6a66713539c079ec3453622336ae655ee28660 --- /dev/null +++ b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72a8a70252da8f958f974a963ae26467e5c3b58a763409b9387454688c7b62ca +size 506022 diff --git a/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/layout.json b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a9b694ff7a37d1e9527a7ba645a53740d3ca6224 --- /dev/null +++ b/2024/One-Shot Diffusion Mimicker for Handwritten Text Generation/layout.json @@ -0,0 +1,9641 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 146 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 146 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "spans": [ + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "text", + "content": "Gang Dai" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "text", + "content": ", Yifan Zhang" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "inline_equation", + "content": "^{2,3\\star}" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "text", + "content": ", Quhui Ke" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "text", + "content": ", Qiangya Guo" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "text", + "content": ", and Shuangping Huang" + }, + { + "bbox": [ + 176, + 167, + 436, + 193 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "spans": [ + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "text", + "content": "South China University of Technology \n" + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "text", + "content": "National University of Singapore, " + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "text", + "content": "Skywork AI, " + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "text", + "content": "Pazhou Laboratory \n" + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{eedaigang}@mail., \\mathrm{eehsp@}\\}" + }, + { + "bbox": [ + 164, + 201, + 447, + 236 + ], + "type": "text", + "content": " scut.edu.cn, yifan.zhang@u.nus.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 265, + 452, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 265, + 452, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 265, + 452, + 506 + ], + "type": "text", + "content": "Abstract. Existing handwritten text generation methods often require more than ten handwriting samples as style references. However, in practical applications, users tend to prefer a handwriting generation model that operates with just a single reference sample for its convenience and efficiency. This approach, known as \"one-shot generation\", significantly simplifies the process but poses a significant challenge due to the difficulty of accurately capturing a writer's style from a single sample, especially when extracting fine details from the characters' edges amidst sparse foreground and undesired background noise. To address this problem, we propose a One-shot Diffusion Mimicker (One-DM) to generate handwritten text that can mimic any calligraphic style with only one reference sample. Inspired by the fact that high-frequency information of the individual sample often contains distinct style patterns (e.g., character slant and letter joining), we develop a novel style-enhanced module to improve the style extraction by incorporating high-frequency components from a single sample. We then fuse the style features with the text content as a merged condition for guiding the diffusion model to produce high-quality handwritten text images. Extensive experiments demonstrate that our method can successfully generate handwriting scripts with just one sample reference in multiple languages, even outperforming previous methods using over ten samples. Our source code is available at https://github.com/dailenson/One-DM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 517, + 430, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 517, + 430, + 528 + ], + "spans": [ + { + "bbox": [ + 160, + 517, + 430, + 528 + ], + "type": "text", + "content": "Keywords: Handwritten Text Generation " + }, + { + "bbox": [ + 160, + 517, + 430, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 517, + 430, + 528 + ], + "type": "text", + "content": " One-Shot Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 548, + 229, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 548, + 229, + 561 + ], + "spans": [ + { + "bbox": [ + 133, + 548, + 229, + 561 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "In the digital age, handwriting text generation blends the personalization of traditional handwriting with the efficiency of automated processes, offering a digital format to preserve the authenticity of individual handwriting. This task aims to automatically generate the desired handwritten text images that not only correspond to specific text content, but also emulate the calligraphic style of a given exemplar writer (e.g., character slant, cursive join, stroke thickness, and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 360, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 360, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 360, + 665 + ], + "type": "text", + "content": "* Authors contributed equally; † Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 114, + 302, + 219 + ], + "blocks": [ + { + "bbox": [ + 152, + 114, + 302, + 219 + ], + "lines": [ + { + "bbox": [ + 152, + 114, + 302, + 219 + ], + "spans": [ + { + "bbox": [ + 152, + 114, + 302, + 219 + ], + "type": "image", + "image_path": "1c0fe1ae5b0af618aa0a891aa277118b2e2e3c2328b433d0772e64f21898650a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 480, + 248 + ], + "type": "text", + "content": "Fig. 1: User experience comparisons between one-shot and few-shot handwritten text generation methods. It reveals that one-shot setting leads to a better user experience." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 465, + 218 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 465, + 218 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 465, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 465, + 218 + ], + "type": "image", + "image_path": "3168de4542c0ee588ddfeb45d78150b9050c1ccf0605bd8be842fa719533cf59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 282 + ], + "type": "text", + "content": "ink color). This provides great convenience for people with hand impairments and also contributes to accelerating the process of handwritten font design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 485, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 485, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 485, + 413 + ], + "type": "text", + "content": "Previous works [4,8,11-13,24,39] introduce generative adversarial networks (GANs) [5,14,32] for handwritten text generation. For instance, ScrabbleGAN [11] leverages random noises as style inputs and conditions content inputs on character-level labels, enabling the synthesis of handwritten words with randomly sampled styles. In a recent study [9], diffusion models [31,42,53] such as denoising diffusion probabilistic models (DDPM) [18] have demonstrated even higher quality of image generation compared to GANs. This motivates several attempts, such as WordStylist [37], GC-DDPM [10], and CTIG-DM [57], to condition the denoising process on the fixed writer ID to generate handwritten text images with controllable styles. However, the major limitation of these methods [10,11,37,57] is that they are unable to mimic the unseen writers' handwriting styles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "type": "text", + "content": "To imitate any given handwriting style, some previous methods [4, 24, 39], known as few-shot generation, require users to provide a few samples (typically 15) as style references. They employ a style encoder to extract writing styles from the given samples, thus offering flexible control over the style of the generated samples. However, the traditional few-shot generation pipeline is inconvenient, inefficient, and time-consuming, as shown in Figure 1. Users prefer methods that only require a single sample as a style reference, known as one-shot generation methods, because they are more convenient, efficient, and easy-to-use. Our goal is to investigate the more challenging one-shot generation task that holds significant practical value, striving to achieve high-quality handwritten text image generation with desired styles and contents." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "The primary challenge of one-shot generation task lies in accurately extracting a user's handwriting style from just one style reference image. As illustrated in Figure 2, characters, being abstract symbols, typically occupy only a minor portion of the reference image. Besides, the reference image is often cluttered with noisy background information, which poses a significant obstacle in extracting the individual handwriting style. Previous one-shot generation methods [8,12,13] simply follow the architectural design of few-shot approaches, using a vanilla CNN encoder to directly extract the handwriting style from a single sample. The extracted style is then combined with text contents and input into a CNN decoder to generate the desired handwritten images. These methods ex" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 119, + 433, + 216 + ], + "blocks": [ + { + "bbox": [ + 187, + 119, + 433, + 216 + ], + "lines": [ + { + "bbox": [ + 187, + 119, + 433, + 216 + ], + "spans": [ + { + "bbox": [ + 187, + 119, + 433, + 216 + ], + "type": "image", + "image_path": "702e2806fd347d3dcba7f96df7856799792949a5d80b59480d0938cffbaff0b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 225, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 225, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 480, + 258 + ], + "type": "text", + "content": "Fig. 2: Handwritten text samples and corresponding high-frequency components. We find that high-frequency components have more pronounced character contours, clearly showcasing the style patterns, such as character slant and cursive connections." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "content": "hibit a limited performance in emulating handwriting styles, due to their poor ability of style extraction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 376 + ], + "type": "text", + "content": "To address the above challenges, our key idea revolves around utilizing the high-frequency information of the sample to enhance the extraction of the handwriting style. As depicted in Figure 2, the high-frequency information encompasses the overall contours of the handwritten text, allowing for a clear observation of key style patterns such as the text slant, letter spacing, and cursive connections. Hence, the incorporation of high-frequency information facilitates a more effective handwriting style extraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 582 + ], + "type": "text", + "content": "In light of the above insight, we propose a One-shot Diffusion Mimicker (One-DM) for handwritten text generation, which is simultaneously guided by desired style and arbitrary content. More specifically, we first develop a style-enhanced module to process both a style reference image and its high-frequency components in parallel. Considering that the reference image often contains background noise, we design a gate mechanism to suppress the inflow of background noise. Regarding the high-frequency components, such as character slant and ligature that show clearer style patterns, we employ a contrastive learning framework [25,54,55] to further obtain discriminative style features, guiding the handwritten text synthesis with both realistic and diverse styles. Subsequently, the style features extracted from both branches are adaptively fused with the specific content prototype in a style-content fusion module. Finally, the seamlessly integrated style and content features act as conditional inputs, guiding the denoising process for the progressive synthesis of stylized handwritten text images. It is worth noting that our One-DM effectively mimics a user's writing style with only one reference sample through our style-enhanced module, surpassing few-shot methods in producing higher-quality stylized handwritings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 582, + 342, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 582, + 342, + 594 + ], + "spans": [ + { + "bbox": [ + 147, + 582, + 342, + 594 + ], + "type": "text", + "content": "To sum up, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 605, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 605, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 605, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 138, + 605, + 479, + 640 + ], + "type": "text", + "content": "- We propose a novel diffusion model for stylized handwritten text generation, which only requires a single reference sample as style input, and imitates its writing style to generate handwritten text with arbitrary content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 480, + 665 + ], + "type": "text", + "content": "- We introduce the high-frequency components of the reference sample to enhance the extraction of handwriting style. The proposed style-enhanced" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 481, + 140 + ], + "type": "text", + "content": "module can effectively capture the writing style patterns and suppress the interference of background noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 176 + ], + "type": "text", + "content": "- Extensive experiments on handwriting datasets in English, Chinese, and Japanese demonstrate that our approach with a single style reference even outperforms previous methods with 15x-more references." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 237, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 237, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 237, + 209 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 480, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 480, + 317 + ], + "type": "text", + "content": "Handwriting generation. Handwritten text is typically stored in two formats: online trajectory form or offline image form. Online handwriting generation methods [1, 7, 28, 44, 46, 56] often employ Recurrent Neural Networks (RNNs) [1, 6, 28, 44-46, 56], transformer decoders [7], or diffusion models [34, 40] to progressively generate writing trajectories. Unlike online methods, offline generation methods have the advantage of not requiring additional trajectory supervision information and can generate realistic handwritten text with stroke width and ink color, which online methods cannot produce." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 481, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 486 + ], + "type": "text", + "content": "Previous offline handwriting generation methods [2, 4, 8, 11, 20, 24, 35, 39] in deep learning predominantly rely on Generative Adversarial Networks (GANs). Early works [2, 11] condition the generative process on the word embeddings [2] or concatenated letter-tokens [11] to synthesize handwritten word images. However, these methods struggle to flexibly control the writing style. Thus, few-shot approaches [4, 24, 39] which rely on 15 style references are introduced. For instance, GANwriting [24] utilizes a CNN encoder [16, 50, 51] to extract a user's handwriting style from a few samples, which are then combined with specific text content to generate handwritings in the desired style. In its follow-up work [23], the synthesized samples are demonstrated to assist in training more robust handwritten text recognizers. Further, HWT [4] employs a transformer encoder [29, 47] to extract rich style patterns from reference samples, enhancing the performance in style mimicry. Recently, VATr [39] uses images of symbols as content representations, enabling the generation of out-of-charset characters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "Concurrently, one-shot generation methods [8, 12, 13] are proposed. Despite their ability to mimic handwriting styles with only a single sample, these techniques still lag behind few-shot methods in terms of stylized generation results. Additionally, previous handwriting generation methods [2, 4, 8, 11-13, 24] often rely on complex content representations, such as recurrent embeddings [2] and letter-level tokens [4, 8, 11-13, 24]. SLOGAN [35] extracts textual contents from easily obtainable printed images. However, it faces challenges in generalizing to unseen writing styles due to its fixed writer ID. Similarly, some diffusion-based [10, 37, 57] methods condition the denoising process on fixed style ID and are unable to mimic styles that they have not previously encountered. In contrast, our One-DM effectively obtains style information from one style sample and thus can generate handwritings with arbitrary styles. Due to space constraints, we discuss more related works in supplementary, including diffusion methods for general image generation, contrastive-based method [48], and frequency-based methods [30, 38]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 107, + 470, + 235 + ], + "blocks": [ + { + "bbox": [ + 151, + 107, + 470, + 235 + ], + "lines": [ + { + "bbox": [ + 151, + 107, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 151, + 107, + 470, + 235 + ], + "type": "image", + "image_path": "c2a04fd867a48f0f68059db58c5c9dea90a83773872be2b1f5435c6398ca6766.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "text", + "content": "Fig. 3: Overview of the proposed method. The style reference initially passes through a high-pass filter to extract its high-frequency components. Subsequently, the spatial and the high-frequency style encoders independently extract style features " + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "inline_equation", + "content": "F_{spa}" + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "inline_equation", + "content": "F_{fre}" + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "text", + "content": " from the style reference and its high-frequency information, respectively. " + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "inline_equation", + "content": "F_{spa}" + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "text", + "content": ", after being filtered through a gate mechanism, is fused with " + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "inline_equation", + "content": "F_{fre}" + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "text", + "content": " and content features " + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 240, + 482, + 318 + ], + "type": "text", + "content": " in the fusion module. The merged feature then serves as a condition input to guide the diffusion generation process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 327, + 201, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 201, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 201, + 339 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": "Problem statement. We aim to synthesize handwritten text images that are controlled by both textual content and handwriting style. Given any word string " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": " and a single style sample " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": " from a writer " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": ", the generated handwritten word image " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": " should replicate the unique handwriting style of " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": " while maintaining the content of " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": ", where the textual content " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{a_i\\}_{i=1}^L" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": " spans a length " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 130, + 356, + 482, + 427 + ], + "type": "text", + "content": " is from a character set without any restrictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": "To address this task, we incorporate high-frequency information to enhance the extraction of writing styles. Existing methods [4,8,12,13,24,39] typically use a vanilla CNN or transformer encoder to directly process style images, often resulting in undesired style extraction. In contrast, we introduce a novel One-shot diffusion mimicker (One-DM) by innovating Laplacian high-frequency extraction and a gating mechanism. Our One-DM can effectively capture style features from a single reference while suppressing background noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 534, + 239, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 239, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 239, + 544 + ], + "type": "text", + "content": "3.1 Overall Scheme" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Our high-level idea focuses on incorporating high-frequency information from style reference images to enhance the extraction of style patterns. A straightforward implementation involves using a vanilla transformer encoder to extract style features from both the style image and its corresponding high-frequency image. This naive solution encounters two main issues: (1) the lack of efficient supervision objective still makes it challenging to accurately learn the writer's style patterns from the high-frequency image, and (2) the style features captured from the original image still retain undesirable noise backgrounds, which can adversely affect subsequent image generation performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "To address the above-mentioned issues, we develop a more effective method, shown in Figure 3. Our method consists of a style-enhanced module, a content encoder, a style-content fusion module, and a conditional diffusion module. Initially, we utilize the Laplacian kernel as a high-frequency filter to extract the high-frequency components from the style reference. Then, two parallel style encoders are employed to simultaneously extract the corresponding style features from both style reference and its high-frequency information. Since undesirable background noise is often present in style references, we design a gate mechanism to facilitate the transmission of informative style information while mitigating noise transmission. The style patterns in the high-frequency components are comparatively cleaner and more distinct, facilitating the observation of individual styles, such as character slant. For the observation, we propose a contrastive learning objective, termed LaplacianNCE " + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lapNCE}" + }, + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": ", to enforce the more discriminative style learning from high-frequency components." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "text", + "content": "Regarding content guidance, we render the given string " + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "text", + "content": " into a unifont image, as done in VATr [39]. Briefly, Unifont's key advantage is that it covers all Unicode characters, allowing our method to transform any user inputs into corresponding images. We further feed render results into a content encoder that combines a ResNet18 [16] with a Transformer encoder. This process first involves using a ResNet18 to handle each character image in parallel, followed by concatenating them to form word sequence features. The Transformer encoder then processes these features to extract an informative content feature " + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "inline_equation", + "content": "E = \\{e_i\\}_{i=1}^L \\in \\mathbb{R}^{L \\times c}" + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "text", + "content": " with a global context, where " + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "text", + "content": " is the channel dimension. After obtaining style features and content guidance, we seamlessly fusion them using a style-content fusion module. The fused results then guide the denoising process of the conditional diffusion model to synthesize the desired handwritten text images. The denoising process is supervised by the reconstruction loss " + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 130, + 284, + 482, + 439 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 439, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 481, + 461 + ], + "type": "text", + "content": "To summarize, the overall training objective of our method combines all two loss functions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 259, + 463, + 354, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 463, + 354, + 476 + ], + "spans": [ + { + "bbox": [ + 259, + 463, + 354, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {l a p N C E} + \\mathcal {L} _ {r e c}.", + "image_path": "05252025217775d4555783a602f76fbf5f18dcc9f1baa3a263bbe1d48a12e589.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 491, + 279, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 279, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 279, + 502 + ], + "type": "text", + "content": "3.2 Style-enhanced Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": "We propose the style-enhanced module to enhance style extraction by incorporating high-frequency components " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "H_{s} \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " from a reference image " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "I_{s} \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " since clearer style patterns are presented in high-frequency components like character slanting and shape connections (cf. Figure 2). As shown in Figure 3, we use a Laplacian kernel as a high-frequency filter to extract " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": ". The Laplacian kernel excels in extracting high-frequency information without the need for Fast Fourier Transform (FFT) and parametric separation in the frequency domain. Then, two style encoders, " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{spa}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{fre}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": ", each a combination of CNN and transformer, process the " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": ", respectively. This independent processing leads to the extraction of distinct style features: " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{spa} = \\{f_{spa}^{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d \\times c}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{spa}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{fre} = \\{f_{fre}^{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d \\times c}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{fre}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "d = h \\times w" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": ". While structurally identical, " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{spa}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{fre}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " do not share weights with each other. Then, the proposed " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lapNCE}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " forces " + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{fre}" + }, + { + "bbox": [ + 130, + 509, + 482, + 668 + ], + "type": "text", + "content": " to focus on extracting discriminative style features" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ". A gate mechanism is designed to selectively filter out background noise from the reference style features, allowing only meaningful style patterns to pass." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": "Laplacian Contrastive Learning. The goal of the proposed " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{lapNCE}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " is to guide the high-frequency style encoder " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text{fre}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " in learning more discriminative style features from the high-frequency information. Thus, we propose to bring closer extracted style features " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "F_{\\text{fre}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " belonging to the same writer, while distancing those from different writers. We formulate our " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{lapNCE}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 209, + 480, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 209, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 180, + 209, + 480, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a p N C E}} = - \\frac {1}{N} \\sum_ {i \\in M} \\frac {1}{| P (i) |} \\sum_ {p \\in P (i)} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} / \\tau\\right)}{\\sum_ {a \\in A (i)} \\exp \\left(z _ {i} \\cdot z _ {a} / \\tau\\right)}. \\tag {1}", + "image_path": "bdf987a75178ff78e5d4f0bc4d3ae6313193e5de9e55065e6cdb11e901b71eaf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": "In detail, " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "i \\in M = \\{1 \\dots N\\}" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " is the index of any element in the mini-batch with size " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "A(i) = M \\setminus \\{i\\}" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " is other indices distinct from " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " is an anchor sample belonging to writer " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "P(i) = \\{p \\in A(i) : w_{p} = w_{i}\\}" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " is its in-batch positive sample set and the other " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "A(i) \\setminus P(i)" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " is its negative set. Here, " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "z = Proj(F_{fre})" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "Proj" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " is a learnable multi-layer perceptron (MLP), " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " is a scalar temperature parameter and the " + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": " symbol denotes the inner product." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": "Gate mechanism. As illustrated in Figure 2, the stroke areas of characters in reference images are typically sparse, with background noise interfering with the extraction of character style features. To address this challenge, we propose a gate mechanism to selectively filter the information of the sample " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": ", as shown in Figure 3. Specifically, the extracted sample style features " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "F_{spa}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": " are fed into a gate layer, consisting of a learnable fully connected layer and followed by a sigmoid activation, to obtain the corresponding gate units " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "W = \\{w_{i}\\}_{i=1}^{d} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": ". Each unit " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": " determines the pass rate for the corresponding " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "f_{spa}^{i}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": ", allowing for a higher pass rate where " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": " is larger. This design effectively enables informative style features " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\hat{F}_{spa} = \\{\\hat{f}_{spa}^{i}\\}_{i=1}^{d}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": " to be extracted while suppressing extraneous background noise, where " + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\hat{f}_{spa}^{i} = f_{spa}^{i} \\cdot w_{i}" + }, + { + "bbox": [ + 130, + 317, + 480, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 469, + 305, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 469, + 305, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 469, + 305, + 480 + ], + "type": "text", + "content": "3.3 Style-content Fusion Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "text", + "content": "Upon acquiring textual content feature " + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "text", + "content": " and two style features " + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "inline_equation", + "content": "\\hat{F}_{spa}" + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "inline_equation", + "content": "F_{fre}" + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "text", + "content": ", we integrate all features within two multi-head attention mechanisms to guide the denoising generation process of the diffusion model, as shown in Figure 3. Specifically, the first cross-attention module uses the textual content " + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 489, + 479, + 609 + ], + "type": "text", + "content": " as queries to identify the most relevant style information in the style reference, thereby inferring the style attributes corresponding to each character. For instance, if the textual content is 'a', it prioritizes searching for style features of characters like 'a', 'b', 'd', 'g' in the style reference, due to these characters appearing similar looped structures, implying more comparable style attributes. This process (cross-attention in Figure 3) is represented as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 207, + 619, + 480, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 619, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 207, + 619, + 480, + 632 + ], + "type": "interline_equation", + "content": "O = \\operatorname {A t t e n} _ {1} \\left(Q _ {1} = E, K _ {1} = V _ {1} = \\hat {F} _ {s p a} + F _ {f r e}\\right). \\tag {2}", + "image_path": "aebbc2f84285a7f5c796379b48e67ac74a6f04b47e389f94ce0d5749163793ba.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Subsequently, we obtain the initial fusion embedding between content and style guidance by simply summing " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": ". The merged intermediate vector is then" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 101 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "employed as the query, key, and value in the self-attention mechanism to facilitate comprehensive interaction of information. Finally, the blended embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " serves as the condition of the diffusion process. The second multi-head attention (self-attention in Figure 3) is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 174, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 174, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 225, + 174, + 481, + 186 + ], + "type": "interline_equation", + "content": "g = \\operatorname {A t t e n} _ {2} \\left(Q _ {2} = K _ {2} = V _ {2} = O + E\\right). \\tag {3}", + "image_path": "b43afc7bcec57f00f1348920f592071ddcffe70cfeda442e12d83bba7e90e33b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 202, + 304, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 202, + 304, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 202, + 304, + 214 + ], + "type": "text", + "content": "3.4 Conditional Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": "The goal of the conditional diffusion model " + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": " is to generate realistic images of handwritten text, guided by acquired conditions " + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": ". Specifically, as shown in Figure 3, under the guidance of " + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": " executes a denoising generative process with " + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": " steps, starting from a sampled Gaussian noise " + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": " and progressively denoises to obtain the desired handwritten text " + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 243, + 290, + 481, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 290, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 243, + 290, + 481, + 314 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0} \\mid g\\right) = \\int p _ {\\theta} \\left(x _ {0: T} \\mid g\\right) d _ {x _ {1: T}}, \\tag {4}", + "image_path": "62fec1ac62ec34e572a110bec981385f5e60d91930f51eae7a6a328c64c6c093.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 327, + 481, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 327, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 228, + 327, + 481, + 358 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0: T} \\mid g\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, g\\right), \\tag {5}", + "image_path": "79a9e7eea86894bd0efe6f8b133bd5f09aa36ad403b64ab505ec4406f36e3507.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 202, + 367, + 480, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 367, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 202, + 367, + 480, + 380 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, g\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, g, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, g, t\\right)\\right). \\tag {6}", + "image_path": "e2a52cc89b01fd959871ad3a54a0b1ce1352922826e7b39380d41451cb14b2b5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "content": "The denoising process aims to learn how to reverse a predefined forward process, as described in DDPM [18]. The forward process is modeled as a fixed Markov chain, where noise conforming to a normal distribution is incrementally added to " + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "content": " to derive " + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "content": ". This can be mathematically expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 222, + 442, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 442, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 222, + 442, + 481, + 456 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {7}", + "image_path": "c030606d2cba270e35e1222aa93893c5690bcf0aaa37e118a645748bfc6fe0d3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": "where the noise is characterized by a variance schedule " + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": ". During training, a variational bound on the maximum likelihood objective is applied to guide the generation process to recover " + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": " from the standard Gaussian noise " + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": ". We give the training objective as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 224, + 521, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 521, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 224, + 521, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\mathbb {E} _ {t, q} \\| \\mu_ {t} \\left(x _ {t}, x _ {0}\\right) - \\mu_ {\\theta} \\left(x _ {t}, g, t\\right) \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "d3ef82e0004b249913ded23c3ed5ce8f1e2560bdec20c5cc5426da90d2ed8fee.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 543, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 543, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mu_t(x_t, x_0)" + }, + { + "bbox": [ + 130, + 543, + 482, + 567 + ], + "type": "text", + "content": " is the mean of the forward process posterior " + }, + { + "bbox": [ + 130, + 543, + 482, + 567 + ], + "type": "inline_equation", + "content": "q(x_t | x_{t-1})" + }, + { + "bbox": [ + 130, + 543, + 482, + 567 + ], + "type": "text", + "content": ", which has a closed-form solution [18]. We put more details in supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 585, + 230, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 230, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 230, + 599 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 609, + 274, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 274, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 274, + 623 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Dataset. To evaluate our One-DM in handwritten text generation, we use the widely-used handwriting dataset IAM [36] and CVL [27].IAM consists of 62,857 English word images from 500 unique writers. Following [4,24,39], we use words" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "from 339 writers for training and the remaining 161 writers for testing. CVL includes texts by 310 writers in English and German. We use the English portion, totaling 84,514 words, and follow CVL's standard split, with 283 writers for training and 27 for testing. Throughout all experiments, we resize images to 64 pixels in height, preserving their aspect ratio." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 480, + 223 + ], + "type": "text", + "content": "Evaluation metrics. We use the Fréchet Inception Distance (FID) [17] and Geometry Score (GS) [26] to assess the generation quality, following the settings of [4, 24, 39]. We also conduct user studies to quantify the subjective quality of the generated handwritten text images in supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 342 + ], + "type": "text", + "content": "Implementation details. In our all experiments, we only use single style reference sample. We first train our model for 700 epochs (batch size of 384) with classifier-free guidance strategy [19] under the guidance scale of 0.25, and then fine-tune the model for 4500 iterations (batch size of 128) using a text recognizer [41] with a CTC loss [15, 21, 58], on four RTX3090 GPUs. The fine-tuning process forces our One-DM to generate readable text with accurate content. The optimizer is AdamW [33], with a learning rate of " + }, + { + "bbox": [ + 130, + 224, + 481, + 342 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 224, + 481, + 342 + ], + "type": "text", + "content": ". During inference, each style sample is randomly sampled from the target-writer. To speed up the sampling, we use the denoising diffusion implicit model (DDIM) [43] with 50 steps. More details are provided in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 481, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 428 + ], + "type": "text", + "content": "Compared methods. We compare our One-DM with state-of-the-art handwritten text generation methods, including GAN-based methods (i.e., GANwriting [24], HWT [4], VATr [39], TS-GAN [8], HiGAN+ [13]), and diffusion-based methods (i.e. GC-DDPM [10] and WordStylist [37]). For a fair comparison, we configure all methods to generate images with a height of 64 pixels in supplementary. Additionally, in supplementary, we compare a variant of One-DM that generates 32-pixel high images with official VATr [39] and HWT [4]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 444, + 228, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 228, + 454 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 228, + 454 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 568 + ], + "type": "text", + "content": "Stylish Handwritten Text Generation. Initially, we assess our One-DM for producing styled handwritten text images, aiming to replicate both the style and content in the generated images. Following [4, 24, 39], we first calculate FID between generated and real samples for each writer separately and finally average them. In line with the previous works [4, 24, 39], our experiments on the IAM dataset are divided into four distinct scenarios: IV-S, IV-U, OOV-S, OOV-U. Among these four scenarios, OOV-U represents the most challenging case where both the target style and the words are entirely unseen during training. For the CVL dataset, we directly report the results of all methods on the test set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "We first report the quantitative results on the IAM dataset in Table 1. We can observe that our One-DM outperforms all the competitors in all settings. Notably, it significantly exceeds one-shot methods in all scenarios. Impressively, our One-DM also holds a substantial advantage over few-shot methods (GAN-writing [24], HWT [4], VATr [39]), which use 15-x more reference samples for style guidance, in IV-S and OOV-S settings. Even in the most challenging OOV-U scenario, our One-DM leads the second-best, VATr, by a large margin (102.75 vs. 108.76), demonstrating the superior performance of our One-DM in stylized" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 183, + 152, + 432, + 258 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparisons with state-of-the-art methods on styled and style-agnostic handwritten text generation in the IAM dataset. Note that, all methods are trained on the same training set used in GANwriting [24], HWT [4], and VATr [39]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 152, + 432, + 258 + ], + "lines": [ + { + "bbox": [ + 183, + 152, + 432, + 258 + ], + "spans": [ + { + "bbox": [ + 183, + 152, + 432, + 258 + ], + "type": "table", + "html": "
MethodShotStyloid EvaluationStyle-agnostic
IV-SIV-UOOV-SOOV-UFID↓GS↓
TS-GAN [8]One118.56128.75127.11136.6720.654.88×10-2
GANwriting [24]Few120.07124.30125.87130.6828.375.67×10-2
HiGAN+ [13]One117.33116.95121.55121.4822.952.06×10-2
GC-DDPM [10]One99.86105.73112.52118.3919.051.31×10-2
WordStylist [37]One98.10104.27109.45115.5218.582.85×10-2
HWT [4]Few109.25106.90116.55113.5218.994.41×10-3
VATr [39]Few103.75101.73111.64108.7616.031.74×10-2
Ours (One-DM)One89.4798.3693.30102.7515.731.98×10-3
", + "image_path": "33eea6d346ef1c0d8a872185fd1a49b0308e42bc4aac877f456cfc484a93eb82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 273, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 479, + 297 + ], + "type": "text", + "content": "handwritten text generation. Similarly, our method outperforms HWT and VATr on the CVL dataset, achieving the lowest FID score, as shown in Table 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 480, + 452 + ], + "type": "text", + "content": "We provide qualitative results to intuitively explain the benefit of our One-DM, in Figure 4. GANwriting struggles to capture the style patterns of reference samples, such as character slant and occasionally produces unclear character shapes. HiGAN+ more consistently generates characters with correct content, but the character spacing within the generated words lacks realism. WordStylist typically produces images with noticeable background noise. HWT and VATr can produce satisfactory handwritten words in terms of content accuracy and style mimicry; however, their downside is the tendency to create smoother character appearances. Compared to HWT and VATr, our synthesized samples excel in more authentic character ink color and stroke thickness. However, some generated samples by our One-DM are visibly different in ink color. We provide comprehensive explanations in supplementary. We further display more qualitative comparisons between our method and few-shot methods in Figure 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 453, + 481, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 573 + ], + "type": "text", + "content": "Style-agnostic Handwritten Text Generation. We further evaluate our One-DM to generate realistic handwritten text images, irrespective of style imitation. For this purpose, we compute the FID and GS on the IAM test set, under the same conditions as applied to ScrabbleGAN [11] (FID: 20.72, GS: " + }, + { + "bbox": [ + 130, + 453, + 481, + 573 + ], + "type": "inline_equation", + "content": "2.56 \\times 10^{-2}" + }, + { + "bbox": [ + 130, + 453, + 481, + 573 + ], + "type": "text", + "content": "), capable of generating handwritten texts with randomly sampled styles. Specifically, each method generates 25k random samples to calculate FID against 25k samples from the test set, and 5k random samples for GS calculation in comparison with 5k test set samples. As presented in Table 1, our One-DM achieves the best results in both FID and GS metrics, further demonstrating its capability to generate higher-quality handwritten text images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 588, + 204, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 588, + 204, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 588, + 204, + 600 + ], + "type": "text", + "content": "4.3 Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to analyze our One-DM. More analyses are provided in supplementary, including generalization evaluation on different style backgrounds, generation quality assessment through OCR performance, failure case analysis, and the effects of different designs (e.g., high-frequency filter, style-content fusion mechanism, and style input sample length)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 114, + 458, + 290 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 458, + 290 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 458, + 290 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 458, + 290 + ], + "type": "table", + "html": "
Style examplesSee us next him unless some eight have the be is unassailable of our boxInour there they on as had inside are shaw have us God is unprouctand tuning which oil to you will the them may the rivers Rim Fishand to is all Common Wood Lessis of ther in work and it hean of
GANw.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*HiGAN+The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*WordS.The greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
HWTThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
VATrThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
*OursThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heartThe greatest test of courage on earth is to bear defeat without losing heart
", + "image_path": "15223f376dd427af09f5ef18083c9c37e518b62127ba366890e57a90b0165621.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 483, + 361 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 483, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 483, + 361 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons between our method with state-of-the-art methods on handwritten text generation with both specific textual content and desired handwriting style in the IAM dataset. We utilize the identical guiding text, 'The greatest test of courage on earth is to bear defeat without losing heart,' across all handwriting generation methods, directing them to produce text in varied styles. Better zoom in " + }, + { + "bbox": [ + 130, + 297, + 483, + 361 + ], + "type": "inline_equation", + "content": "200\\%" + }, + { + "bbox": [ + 130, + 297, + 483, + 361 + ], + "type": "text", + "content": ". * denotes one-shot methods, while others are few-shot methods." + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 147, + 371, + 289, + 446 + ], + "blocks": [ + { + "bbox": [ + 147, + 371, + 289, + 446 + ], + "lines": [ + { + "bbox": [ + 147, + 371, + 289, + 446 + ], + "spans": [ + { + "bbox": [ + 147, + 371, + 289, + 446 + ], + "type": "table", + "html": "
Targethopetherelatersuccess
Generated Sampleshopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
hopetherelatersuccess
", + "image_path": "682a7fa689dbb818439afdebc95d3a073e71104b7f45e0cc40fdcee9ed6b1b3c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 449, + 301, + 494 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 301, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 301, + 494 + ], + "type": "text", + "content": "Fig. 5: Each row shows results from our One-DM and few-shot methods on IAM dataset; readers are invited to identify our method. The answer is at the paper's end." + } + ] + } + ], + "index": 5, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 321, + 373, + 473, + 445 + ], + "blocks": [ + { + "bbox": [ + 321, + 373, + 473, + 445 + ], + "lines": [ + { + "bbox": [ + 321, + 373, + 473, + 445 + ], + "spans": [ + { + "bbox": [ + 321, + 373, + 473, + 445 + ], + "type": "table", + "html": "
TextStyle A··· interpolation··· Style B
plantsplants plnph plnch plnns plnns plnns plnns
cloudcloud cloud cloud cloud cloud cloud cloud
stickstick stick stick stick stick stick stick
photophoto photo photo photo photo photo photo
vaporVapor vapor vapor vapor vapor vapor vapor vapor
listslists lists lists lists lists lists lists
", + "image_path": "eac177df89240900fdd53803de76fdc1a0d63bf4112ef69e467b4621729a750e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 449, + 482, + 492 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 449, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 311, + 449, + 482, + 492 + ], + "type": "text", + "content": "Fig. 6: We provide style interpolation results generated by our One-DM. Different individual writing styles are extracted from the IAM test dataset." + } + ] + } + ], + "index": 7, + "type": "text" + }, + { + "bbox": [ + 130, + 507, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 591 + ], + "type": "text", + "content": "Quantitative evaluation of Laplacian branch and gate mechanism. We conduct various ablation experiments on the IAM dataset to evaluate the effect of distinct components within our approach. We provide the quantitative result in Table 2. We find that: (1) The inclusion of both Laplacian branch and gate mechanism enhances the quality of the generated handwritten text images, improving FID by 3.92 and 2.71, respectively. (2) Integrating the Laplacian branch with the gate mechanism further boosts the generative performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Qualitative evaluation of Laplacian branch and gate mechanism. To further analyze each module in our One-DM, we conduct visual ablation experiments. As shown in Table 2, we can observe that firstly, after adding a gate mechanism, background noise can be somewhat suppressed, resulting in characters with relatively clean backgrounds. Then, the independent addition of the Laplacian branch helps the model learn cursive connections and other" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 178, + 296, + 288 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 312, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 312, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 312, + 170 + ], + "type": "text", + "content": "Table 2: Ablation study. Effect of the Laplacian branch and gate mechanism on the IAM dataset under the OOV-U setting, as done in [4]. In the middle, we showcase the generated samples of each component." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 178, + 296, + 288 + ], + "lines": [ + { + "bbox": [ + 148, + 178, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 148, + 178, + 296, + 288 + ], + "type": "table", + "html": "
Base Εfre GateStyle samplesFID ↓
would like
✓cloudfoul108.44
✓cloudfoul105.73
✓cloudfoul104.52
✓cloudfoul102.75
", + "image_path": "1774ce3c6f3db2c6fc6f8aebdbe62df48343662b3d80eb5fcb152d71fe1577d7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 358, + 150, + 430, + 208 + ], + "blocks": [ + { + "bbox": [ + 325, + 114, + 462, + 148 + ], + "lines": [ + { + "bbox": [ + 325, + 114, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 325, + 114, + 462, + 148 + ], + "type": "text", + "content": "Table 3: Effect of each part of our Laplacian branch on IAM dataset under OOV-U setting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 358, + 150, + 430, + 208 + ], + "lines": [ + { + "bbox": [ + 358, + 150, + 430, + 208 + ], + "spans": [ + { + "bbox": [ + 358, + 150, + 430, + 208 + ], + "type": "table", + "html": "
HsLapNCEFID ↓
108.44
✓106.16
✓107.58
✓✓104.52
", + "image_path": "b3f1a9819762c0aaa440f4fef71d7b275be53eb7bf53f8401b9695d9fc4fd8cf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 353, + 237, + 435, + 286 + ], + "blocks": [ + { + "bbox": [ + 326, + 213, + 460, + 235 + ], + "lines": [ + { + "bbox": [ + 326, + 213, + 460, + 235 + ], + "spans": [ + { + "bbox": [ + 326, + 213, + 460, + 235 + ], + "type": "text", + "content": "Table 4: Comparisons with competitors on the CVL dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 237, + 435, + 286 + ], + "lines": [ + { + "bbox": [ + 353, + 237, + 435, + 286 + ], + "spans": [ + { + "bbox": [ + 353, + 237, + 435, + 286 + ], + "type": "table", + "html": "
MethodFID ↓
HWT [4]58.22
VATr [39]54.44
Ours (One-DM)51.78
", + "image_path": "273a1979ec18c427c5eb8bb82e54853d500bd4537f5ab7fa0e30647d79a39354.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 302, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 326 + ], + "type": "text", + "content": "style patterns. Finally, our method integrated the Laplacian branch and gate mechanism, which can generate the highest quality handwritten text images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": "Discussions of the Laplacian branch. Our Laplacian branch consists of two key components: utilizing high-frequency images " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": " and the Laplacian contrastive learning loss " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lapNCE}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": ". In the previous ablation study, they are always combined. We further conduct exploration experiments to explain why they cannot be separated. As reported in Table 3, combining " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lapNCE}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": " maximizes effectiveness; separating them significantly reduces performance. Without the guidance of " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lapNCE}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": ", extracting discriminative features from " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": " is challenging. Likewise, directly applying " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lapNCE}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": " on original images leads to an undesired style feature extraction, as original images have less clear style patterns than " + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 326, + 480, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 434, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 481, + 529 + ], + "type": "text", + "content": "Discussions about learning style from a single reference. We are quite surprised that One-DM, with just a single reference sample, even exceeds the generative performance of few-shot methods. We provide the potential reason analysis below. Firstly, One-DM learns a meaningful style latent space, wherein new styles can be generated based on seen styles (cf. Figure 6). Then, through our style-enhanced module, One-DM effectively extracts the style feature from a single example and maps it to a position close to the exemplary writer in the feature space, thereby producing high-quality styled handwritten text images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 546, + 387, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 387, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 387, + 558 + ], + "type": "text", + "content": "4.4 Comparisons with SOTA Industrial Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 564, + 480, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 480, + 637 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 480, + 637 + ], + "type": "text", + "content": "To highlight the superiority of our method, we compare One-DM with leading industrial image generation methods that are trained on tremendously large datasets (including numerous text-centric images), including two prominent text-to-image methods, DALL-E3 [3] and Stable Diffusion [42] (SD), and two popular style transfer methods, Artbreeder1 and IP-Adapter2 (IP-A.), on the IAM dataset. Further experiment details are in supplementary material." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 271, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 271, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 271, + 653 + ], + "type": "text", + "content": "1 https://www.artbreeder.com/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 654, + 348, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 348, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 348, + 665 + ], + "type": "text", + "content": "2 https://github.com/tencent-ailab/IP-Adapter" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 116, + 467, + 195 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 467, + 195 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 467, + 195 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 467, + 195 + ], + "type": "image", + "image_path": "805e3e0f4589677c79e85db22cc210f5c31488d5e288123592d6156ca553e840.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "type": "text", + "content": "Fig. 7: (a) Qualitative comparisons between our method with SOTA industrial image generation methods, including DALL-E3 [3], Stable Diffusion [42], Artbreeder" + }, + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "type": "text", + "content": " and IP-Adapter" + }, + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "type": "text", + "content": " on handwritten text generation with both controllable styles and contents. (b) Qualitative comparisons with Fzshouji" + }, + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 131, + 199, + 480, + 243 + ], + "type": "text", + "content": " on ICDAR-2013 competition database [52]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 149, + 287, + 277, + 355 + ], + "blocks": [ + { + "bbox": [ + 131, + 248, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 131, + 248, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 294, + 281 + ], + "type": "text", + "content": "Table 5: Quantitative comparisons with competitors on styled handwritten character generation in terms of FID." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 287, + 277, + 355 + ], + "lines": [ + { + "bbox": [ + 149, + 287, + 277, + 355 + ], + "spans": [ + { + "bbox": [ + 149, + 287, + 277, + 355 + ], + "type": "table", + "html": "
MethodChineseJapanese
GANwriting [24]116.49111.86
HWT [4]165.74148.66
VATr [39]139.91124.98
WordStylist [37]34.61101.93
Ours (One-DM)27.2495.43
", + "image_path": "36a29fe8864d814cbd14c89163702eb26e2f7dc58604529cc4bf2c0f1f14ae83.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 325, + 279, + 465, + 357 + ], + "blocks": [ + { + "bbox": [ + 312, + 250, + 477, + 272 + ], + "lines": [ + { + "bbox": [ + 312, + 250, + 477, + 272 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 477, + 272 + ], + "type": "text", + "content": "Table 6: Qualitative comparisons with VATr [39] on Chinese dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 279, + 465, + 357 + ], + "lines": [ + { + "bbox": [ + 325, + 279, + 465, + 357 + ], + "spans": [ + { + "bbox": [ + 325, + 279, + 465, + 357 + ], + "type": "table", + "html": "
GTVATrDiffusion Step
0123103050
艾思艾艾艾艾
布布布布布
琳洁淋淋淋淋
昂昂昂昂昂昂
", + "image_path": "68b6e1f75d5508c0c57049b301e78ddb992e5479114d096e008d793eff7eea84.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": "As shown in Figure 7(a), our method excels industrial methods in style mimicry and content preservation. The performance of IP-A. is the poorest, often producing distorted images. Artbreeder can replicate the color of the strokes from style samples, but fails in content preservation. DALL-E3 and SD generate characters with accurate content, but often mismatch style details with references, such as character spacing and stroke width, with SD often generating extra backgrounds. Besides, we compare Fzshouji3, an advanced industrial method designed for Chinese handwriting generation. As shown in Figure 7(b), our method outperforms Fzshouji in replicating character details and ink color." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 494, + 326, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 326, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 326, + 506 + ], + "type": "text", + "content": "4.5 Applications to Other Languages" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 514, + 480, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 480, + 574 + ], + "type": "text", + "content": "In this section, we evaluate whether One-DM can be used to generate languages other than English. We further conduct experiments in the Chinese (i.e. ICDAR2013 competition database [52]) and Japanese(i.e. UP_Kuchibue database [22]) datasets. We use FID to assess generated samples of each writer, then average them. More experiment details can be seen in supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "For the Chinese handwritten character generation task, as shown in Table 5, we find that our One-DM outperforms the second-best method by a large margin, achieving a 7.37 lower FID. From Figure 8 (a), we can observe that our One-DM generates characters that closely match the target images in geometric shape and character slant. In contrast, handwritings from HWT and VATr exhibit noticeable artifacts like blurring and collapsed structures. GANwriting tends to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 282, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 282, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 282, + 666 + ], + "type": "text", + "content": "3 https://www.fzshouji.com/make" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 116, + 443, + 240 + ], + "blocks": [ + { + "bbox": [ + 173, + 116, + 443, + 240 + ], + "lines": [ + { + "bbox": [ + 173, + 116, + 443, + 240 + ], + "spans": [ + { + "bbox": [ + 173, + 116, + 443, + 240 + ], + "type": "image", + "image_path": "1533947c780ae3abe8b5c355723de45060f1dc2460e499931786197299abaf9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 481, + 287 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 287 + ], + "type": "text", + "content": "Fig. 8: Comparisons with GANwriting [24], HWT [4], VATr [39], and WordStylist [37] for styled handwritten character generation on Chinese and Japanese scripts. The red boxes highlight failures of structure preservation, while blue boxes highlight comparisons between the style patterns of targets and generated characters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "text", + "content": "miss strokes in its handwritings. WordStylist sometimes struggles to accurately mimic style patterns and tends to generate characters with incorrect radicals. Table 5 and Figure 8 (b) further verify the effectiveness of One-DM for Japanese handwriting generation. We also achieve the lowest FID score, and our generated Japanese samples excel in both content preservation and style imitation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 496 + ], + "type": "text", + "content": "We further investigate why diffusion-based methods (Our One-DM and Word-Stylist) that require only one single sample outperform few-shot GAN-based methods (i.e., GANwriting, HWT, and VATr) so significantly in generating Chinese and Japanese characters. The lower performance of GAN-based methods on Chinese and Japanese characters may stem from their vanilla convolutional architectures struggling with Chinese and Japanese characters' complex geometry, as noted in [49]. In contrast, our One-DM breaks down the generation of Chinese and Japanese characters into simpler steps. For instance, as demonstrated in Table 6, during the early stages of the diffusion generation process, the model first attempts to generate a rough Chinese handwritten character. It then continues to refine the writing style (e.g., character shape and stroke color) under the condition guidance, until it synthesizes handwritings that are satisfactory." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": "In this paper, we introduce a novel One-DM for handwritten text generation, requiring only one style reference to produce realistic handwritten text images4. We enhance style extraction by incorporating high-frequency components from the style reference. For high-frequency components with distinct style patterns, we employ laplacian contrastive learning to capture more discriminative style features. Moreover, a gate mechanism improves the transfer of informative features from the reference, reducing background noise. Our One-DM outperforms few-shot methods in multiple language scripts. In the future, we aim to explore the potential of One-DM in font generation and vector font creation tasks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 406, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 406, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 406, + 665 + ], + "type": "text", + "content": "4 In Figure 5, the second from bottom row belongs to our method." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Acknowledgments The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 196, + 198, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 198, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 198, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 222, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 222, + 480, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 244 + ], + "type": "text", + "content": "1. Aksan, E., Pece, F., Hilliges, O.: Deepwriting: Making digital ink editable via deep generative modeling. In: ACM Conf. Human Factors Comput. Sys. pp. 1-14 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "2. Alonso, E., Moysset, B., Messina, R.: Adversarial generation of handwritten text images conditioned on sequences. In: ICDAR. pp. 481-486 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 300 + ], + "type": "text", + "content": "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. 2(3), 8 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 322 + ], + "type": "text", + "content": "4. Bhunia, A.K., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S., Shah, M.: Hand-writing transformers. In: ICCV. pp. 1086-1094 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "text", + "content": "5. Cao, J., Mo, L., Zhang, Y., Jia, K., Shen, C., Tan, M.: Multi-marginal WassersteinianGAN. In: NeurIPS. vol. 32 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 345, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 480, + 376 + ], + "type": "text", + "content": "6. Chen, Z., Yang, D., Liang, J., Liu, X., Wang, Y., Peng, Z., Huang, S.: Complex handwriting trajectory recovery: Evaluation metrics and algorithm. In: ACCV. pp. 1060-1076 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 378, + 480, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 480, + 410 + ], + "type": "text", + "content": "7. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: CVPR. pp. 5977-5986 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 443 + ], + "type": "text", + "content": "8. Davis, B.L., Morse, B.S., Price, B.L., Tensmeyer, C., Wigington, C., Jain, R.: Text and style conditioned GAN for the generation of offline-handwriting lines. In: BMVC (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 445, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 466 + ], + "type": "text", + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS. pp. 8780-8794 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "text", + "content": "0. Ding, H., Luan, B., Gui, D., Chen, K., Huo, Q.: Improving handwrittenOCR with training samples generated by glyph conditional denoising diffusion probabilistic model. In: ICDAR. pp. 20-37 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 500, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 532 + ], + "type": "text", + "content": "1. Fogel, S., Averbuch-Elor, H., Cohen, S., Mazor, S., Litman, R.: Scrabblegan: Semi-supervised varying length handwritten text generation. In: CVPR. pp. 4324-4333 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "text", + "content": "2. Gan, J., Wang, W.: Higan: Handwriting imitation conditioned on arbitrary-length texts and disentangled styles. In: AAAI. pp. 7484-7492 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "type": "text", + "content": "3. Gan, J., Wang, W., Leng, J., Gao, X.: Higan+: Handwriting imitation gan with disentangled representations. ACM TOG 42(1), 1-17 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "text", + "content": "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. pp. 2672-2680 (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "5. Graves, A., Fernández, S., Gomez, F., Schmidhuber, J.: Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In: ICML. pp. 369-376 (2006)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in neural information processing systems (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. vol. 33, pp. 6840-6851 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 438, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 438, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 438, + 182 + ], + "type": "text", + "content": "19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: arXiv (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "type": "text", + "content": "20. Huang, H., Yang, D., Dai, G., Han, Z., Wang, Y., Lam, K.M., Yang, F., Huang, S., Liu, Y., He, M.: Agtgan: Unpaired image translation for photographic ancient character generation. In: ACM MM. pp. 5456-5467 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "21. Huang, S., Luo, Y., Zhuang, Z., Yu, J.G., He, M., Wang, Y.: Context-aware selective label smoothing for calibrating sequence recognition model. In: ACM MM (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "22. Jaeger, S., Nakagawa, M.: Two on-line japanese character databases in unipen format. In: ICDAR. pp. 566-570 (2001)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "23. Kang, L., Riba, P., Rusinol, M., Fornes, A., Villegas, M.: Content and style aware generation of text-line images for handwriting recognition. IEEE TPAMI 44(12), 8846-8860 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "24. Kang, L., Riba, P., Wang, Y., Rusinol, M., Fornés, A., Villegas, M.: Ganwriting: content-conditioned generation of styled handwritten word images. In: ECCV. pp. 273-289 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "25. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: NeurIPS. pp. 18661-18673 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "26. Khrulkov, V., Oseledets, I.: Geometry score: A method for comparing generative adversarial networks. In: International conference on machine learning. pp. 2621-2629. PMLR (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "text", + "content": "27. Kleber, F., Fiel, S., Diem, M., Sablatnig, R.: Cvl-database: An off-line database for writer retrieval, writer identification and word spotting. In: ICDAR. pp. 560-564 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "28. Kotani, A., Tellex, S., Tompkin, J.: Generating handwriting via decoupled style descriptors. In: ECCV. pp. 764-780 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "29. Li, D., Chen, G., Wu, X., Yu, Z., Tan, M.: Face anti-spoofing with cross-stage relation enhancement and spoof material perception. Neural Networks 175, 106275 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "30. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR. pp. 5141-5150 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "31. Liu, Z., Chen, K., Zhang, Y., Han, J., Hong, L., Xu, H., Li, Z., Yeung, D.Y., Kwok, J.: Geom-erasing: Geometry-driven removal of implicit concept in diffusion models. arXiv (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "32. Liu, Z., Jia, W., Yang, M., Luo, P., Guo, Y., Tan, M.: Deep view synthesis via self-consistent generative network. IEEE TMM 24, 451-465 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 578 + ], + "type": "text", + "content": "33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "34. Luhman, T., Luhman, E.: Diffusion models for handwriting generation. arXiv (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "35. Luo, C., Zhu, Y., Jin, L., Li, Z., Peng, D.: Slogan: handwriting style synthesis for arbitrary-length and out-of-vocabulary text. IEEE Trans. Neural Networks Learn. Syst. (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "36. Marti, U.V., Bunke, H.: The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition 5, 39-46 (2002)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "37. Nikolaidou, K., Retsinas, G., Christlein, V., Seuret, M., Sfikas, G., Smith, E.B., Mokayed, H., Liwicki, M.: Wordstylist: Structured verbatim handwritten text generation with latent diffusion models. In: ICDAR. pp. 384-401 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "38. Pan, Z., Ji, Z., Liu, X., Bai, J., Liu, C.L.: Visa: Visual and semantic alignment for robust scene text recognition. In: ICDAR. pp. 223-242 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "39. Pippi, V., Cascianelli, S., Cucchiara, R.: Handwritten text generation from visual archetypes. In: CVPR. pp. 22458-22467 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "40. Ren, M.S., Zhang, Y.M., Wang, Q.F., Yin, F., Liu, C.L.: Diff-writer: A diffusion model-based stylized online handwritten chinese character generator. In: International Conference on Neural Information Processing. pp. 86-100 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "41. Retsinas, G., Sfikas, G., Gatos, B., Nikou, C.: Best practices for a handwritten text recognition system. In: International Workshop on Document Analysis Systems. pp. 247-259 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 293 + ], + "type": "text", + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "44. Tang, S., Lian, Z.: Write like you: Synthesizing your cursive online chinese handwriting via metric-based meta learning. In: Comput. Graph. Forum. pp. 141-151 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "45. Tang, S., Xia, Z., Lian, Z., Tang, Y., Xiao, J.: Fontrnn: Generating large-scale chinese fonts via recurrent neural network. In: Comput. Graph. Forum. pp. 567-577 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "46. Tolosana, R., Delgado-Santos, P., Perez-Uribe, A., Vera-Rodriguez, R., Fierrez, J., Morales, A.: Deepwritesyn: On-line handwriting synthesis via deep short-term representations. In: AAAI. pp. 600-608 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "48. Xie, X., Fu, L., Zhang, Z., Wang, Z., Bai, X.: Toward understanding wordart: Corner-guided transformer for scene text recognition. In: ECCV. pp. 5456-5467 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "49. Xie, Y., Chen, X., Sun, L., Lu, Y.: Dg-font: Deformable generative networks for unsupervised font generation. In: CVPR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "type": "text", + "content": "50. Yang, Y., Liu, D., Zhang, S., Deng, Z., Huang, Z., Tan, M.: Hilo: Detailed and robust 3d clothed human reconstruction with high-and low-frequency information of parametric models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10671-10681 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "text", + "content": "51. Yang, Y., Zhang, S., Huang, Z., Zhang, Y., Tan, M.: Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15901-15911 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "52. Yin, F., Wang, Q.F., Zhang, X.Y., Liu, C.L.: Icdar 2013 chinese handwriting recognition competition. In: International Conference on Document Analysis and Recognition. pp. 1464-1470 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "text", + "content": "53. Zhang, Y., Hooi, B.: Hipa: Enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "54. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: NeurIPS. pp. 34077-34090 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "55. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: NeurIPS. pp. 29848-29860 (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 101 + ], + "type": "text", + "content": "One-DM: One-Shot Diffusion Mimicker for Handwritten Text Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 205 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "56. Zhao, B., Tao, J., Yang, M., Tian, Z., Fan, C., Bai, Y.: Deep imitator: Handwriting calligraphy imitation via deep attention networks. Pattern Recognition 104, 107080 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "57. Zhu, Y., Li, Z., Wang, T., He, M., Yao, C.: Conditional text image generation with diffusion models. In: CVPR. pp. 14235-14245 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "58. Zhuang, Z., Liu, Z., Lam, K.M., Huang, S., Dai, G.: A new semi-automatic annotation model via semantic boundary estimation for scene text detection. In: ICDAR. pp. 257-273 (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "G. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_content_list.json b/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f9c826821a70dc5aa62fd6b91f516fbc4faaf3a9 --- /dev/null +++ b/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_content_list.json @@ -0,0 +1,1765 @@ +[ + { + "type": "text", + "text": "One-stage Prompt-based Continual Learning", + "text_level": 1, + "bbox": [ + 246, + 141, + 754, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngeun Kim, Yuhang Li, and Priyadarshini Panda", + "bbox": [ + 292, + 189, + 710, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yale University, CT, USA {youngun.kim,yuhang.li, priya.panda}@yale.edu", + "bbox": [ + 326, + 217, + 676, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Prompt-based Continual Learning (PCL) has gained considerable attention as a promising continual learning solution because it achieves state-of-the-art performance while preventing privacy violations and memory overhead problems. Nonetheless, existing PCL approaches face significant computational burdens because of two Vision Transformer (ViT) feed-forward stages; one is for the query ViT that generates a prompt query to select prompts inside a prompt pool; the other one is a backbone ViT that mixes information between selected prompts and image tokens. To address this, we introduce a one-stage PCL framework by directly using the intermediate layer's token embedding as a prompt query. This design removes the need for an additional feed-forward stage for query ViT, resulting in $\\sim 50\\%$ computational cost reduction for both training and inference with marginal accuracy drop $(\\leq 1\\%)$ . We further introduce a Query-Pool Regularization (QR) loss that regulates the relationship between the prompt query and the prompt pool to improve representation power. The QR loss is only applied during training time, so there is no computational overhead at inference from the QR loss. With the QR loss, our approach maintains $\\sim 50\\%$ computational cost reduction during inference as well as outperforms the prior two-stage PCL methods by $\\sim 1.4\\%$ on public class-incremental continual learning benchmarks including CIFAR-100, ImageNet-R, and DomainNet.", + "bbox": [ + 261, + 282, + 738, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Efficient learning $\\cdot$ Continual learning $\\cdot$ Transfer learning", + "bbox": [ + 261, + 601, + 730, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 642, + 375, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training models effectively and efficiently on a continuous stream of data presents a significant practical hurdle. A straightforward approach would entail accumulating both prior and new data and then updating the model using this comprehensive dataset. However, as data volume grows, fully retraining a model on such extensive data becomes increasingly impractical [11, 25]. Additionally, storing past data can raise privacy issues, such as those highlighted by the EU General Data Protection Regulation (GDPR) [36]. An alternative solution is to adapt the model based solely on currently incoming data, eschewing any access to past data. This paradigm is termed as rehearsal-free continual learning [7,9,32,33,39-41], and the primary goal is to diminish the effects of catastrophic forgetting on previously acquired data.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ecd3893e7bd2e1ff20b0da9997f2338260a744ee86f04ffdc7c54dc37700af0d.jpg", + "image_caption": [ + "Fig. 1: Difference between prior PCL work and ours. Prior PCL work (Left) has two feed-forward stages for (1) a query function (ViT) to select input-specific prompts and (2) a backbone ViT layer to perform prompt learning with the selected prompts and image tokens. On the other hand, our one-stage PCL framework (Middle) uses an intermediate layer's token embedding as a prompt query so that it requires only one backbone ViT feed-forward stage. As a result, our method reduces GFLOPs by $\\sim 50\\%$ compared to prior work while maintaining accuracy (Right)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 413, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9a3e8773bd867ca47ccef4cb8ed9fae8233334a92e9f91b27f57b7ad8e5e504.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 143, + 604, + 286 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0aa1ec17fcf570a5a36dbe5753eae61a8128aee44ab809b81a8499aad7b13ff2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 148, + 781, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among the rehearsal-free continual learning methods, Prompt-based Continual Learning (PCL) stands out as it has demonstrated state-of-the-art performance in image classification tasks, even surpassing rehearsal-based methods [33, 37, 39, 40]. PCL utilizes a pre-trained Vision Transformer (ViT) [8] and refines the model by training learnable tokens on the given data. PCL adopts a prompt pool-based training scheme where different prompts are selected and trained for each continual learning stage. This strategy enables a model to learn the information of the training data in a sequential manner with less memory overhead, as the prompt pool requires minimal resources.", + "bbox": [ + 212, + 444, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although PCL methods show state-of-the-art performance, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices [12,26,38]. Specifically, the PCL method requires two-stage ViT feedforward steps. One is for the query function that generates a prompt query. The other one is a backbone ViT that mixes information between selected prompts and input image tokens. We refer to this approach as a two-stage PCL method, illustrated in Fig. 1 Left.", + "bbox": [ + 212, + 582, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this limitation, we propose a one-stage PCL framework with only one ViT feed-forward stage (Fig. 1 Middle), called OS-Prompt. Rather than deploying a separate ViT feed-forward phase to generate a prompt query, we directly use the intermediate layer's token embedding as a query. This is based on our observation that early layers show marginal shifts in the feature space during continual prompt learning (details are in Section 4.1). This observation enables us to use the intermediate token embedding as a prompt query which requires consistent representation across continual learning to minimize catastrophic forgetting. Surprisingly, OS-Prompt shows a marginal performance drop (less than $1\\%$ ) while saving $\\sim 50\\%$ of computational cost (Fig. 1 Right).", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As our OS-Prompt uses the intermediate layer's token embedding as a query instead of the last layer's token embedding (which is used in prior PCL methods), there is a slight performance drop due to the lack of representation power. To address this, we introduce a Query-Pool Regularization (QR) loss, which is designed to enhance the representation power of a prompt pool by utilizing the token embedding from the last layer. Importantly, the QR loss is applied only during training, ensuring no added computational burden during inference. We refer to our enhanced one-stage framework with QR loss as OS-Prompt++. Our OS-Prompt++ bridges the accuracy gap, performing better than OS-Prompt.", + "bbox": [ + 212, + 146, + 782, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our contribution can be summarized as follows: (1) We raise a drawback of the current PCL methods — high computational cost, particularly due to the two distinct ViT feed-forward stages. As a solution to the computational inefficiency in existing PCL techniques, we propose OS-Prompt that reduces the computational cost by nearly $50\\%$ without any significant performance drop. (2) To counter the slight performance degradation observed in our one-stage PCL framework, we introduce a QR loss. This ensures that the prompt pool maintains similarity with token embedding from both intermediate and final layers. Notably, our QR loss avoids any additional computational overhead during inference. (3) We conduct experiments with rehearsal-free continual learning setting on CIFAR-100 [21], ImageNet-R [16], and DomainNet [27] benchmarks, outperforming the previous SOTA method CodaPrompt [33] by $\\sim 1.4\\%$ with $\\sim 50\\%$ computational cost saving.", + "bbox": [ + 212, + 282, + 784, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 500, + 387, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Continual Learning", + "text_level": 1, + "bbox": [ + 215, + 530, + 421, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a decade, continual learning has been explored as an important research topic, concentrating on the progressive adaptation of models over successive tasks or datasets. One of the representative methods to address continual learning is the regularization-based method [1, 23, 44]. By adding a regularization loss between the current model and the previous model, these methods aim to minimize catastrophic forgetting. However, their performance is relatively low on challenging datasets compared to the other continual learning methods. An alternative approach proposes to expand the network architecture as it progresses through different continual learning stages [22, 30, 31, 42]. Though these typically surpass the results of regularization-based methods, they come at the expense of significant memory consumption due to the increased parameters. Rehearsal-based continual learning has introduced a component for archiving previous data [2, 5, 6, 13, 29]. By leveraging accumulated data during the subsequent stages, they often outperform other continual learning methods. However, saving images causes memory overhead and privacy problems. Given the inherent limitations, a growing interest is observed in rehearsal-free continual learning. These are crafted to address catastrophic forgetting without access to past data. Most of them propose methods to generate rehearsal images [7, 9, 32, 41], but generating images is resource-heavy and time-consuming. Recently, within", + "bbox": [ + 212, + 553, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 433, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the rehearsal-free domain, Prompt-based Continual Learning (PCL) methods have gained considerable attention because of their performance and memory efficiency. By utilizing a pre-trained ViT model, they train only a small set of parameters called prompts to maintain the information against catastrophic forgetting. For example, L2P [40] utilizes a prompt pool, selecting the appropriate prompts for the given image. Extending this principle, DualPrompt [39] proposes general prompts for encompassing knowledge across multiple continual learning stages. Advancing this domain further, [33] facilitates end-to-end training of the prompt pool, achieving state-of-the-art performance. However, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices. Our work aims to resolve such computational complexity overhead in PCL. Furthermore, Tang et al. [35] have proposed a prompt-based continual learning that generates prompts based on the intermediate features from the transformer. However, this prior work focuses on maximizing continual learning performance without a pre-trained model, while ours aims to leverage the power of pre-trained models. Also, the prior work focuses on the prompt generation method, while we propose a framework compatible with different prompt learning techniques.", + "bbox": [ + 212, + 146, + 787, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Prompt-based Learning", + "text_level": 1, + "bbox": [ + 215, + 441, + 457, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The efficient fine-tuning method for large pre-trained models has shown their practical benefits across various machine learning tasks [14, 18, 28, 45-47]. Instead of adjusting all parameters within neural architectures, the emphasis has shifted to leveraging a minimal set of weights for optimal transfer outcomes. In alignment with this, multiple methodologies [3, 30] have integrated a streamlined bottleneck component within the transformer framework, thus constraining gradient evaluations to select parameters. Strategies like TinyTL [3] and BitFit [43] advocate for bias modifications in the fine-tuning process. In a recent shift, prompt-based learning [19, 20] captures task-specific knowledge with much smaller additional parameters than the other fine-tuning methods [40]. Also, prompt-based learning only requires storing several prompt tokens, which are easy to plug-and-play, and hence, they are used to construct a prompt pool for recent continual learning methods.", + "bbox": [ + 212, + 468, + 787, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 689, + 370, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Setting", + "text_level": 1, + "bbox": [ + 215, + 722, + 401, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a continual learning setting, a model is trained on $T$ continual learning stages with dataset $D = \\{D_{1},D_{2},\\dots,D_{T}\\}$ , where $D_{t}$ is the data provided in $t$ -th stage. Our problem is based on an image recognition task, so each data $D_{t}$ consists of pairs of images and labels. Also, data from the previous tasks is not accessible for future tasks. Following the previous rehearsal-free continual learning settings [33, 39], we focus on the class-incremental continual learning setting where task", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "identity is unknown at inference. This is a challenging scenario compared to others such as task-incremental continual learning where the task labels are provided for both training and test phases [17]. For the experiments, we split classes into $T$ chunks with continual learning benchmarks including CIFAR-100 [21] and ImageNet-R [16].", + "bbox": [ + 212, + 146, + 782, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Two-stage Prompt-based Continual Learning", + "text_level": 1, + "bbox": [ + 214, + 250, + 630, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our framework, we focus on improving efficiency through a new query selection rather than changing the way prompts are formed from the prompt pool with the given query, a common focus in earlier PCL studies [33, 39, 40]. For a fair comparison with earlier work, we follow the overall PCL framework established in previous studies. Given that our contribution complements the prompt formation technique, our method can seamlessly work with future works that propose a stronger prompt generation method.", + "bbox": [ + 212, + 280, + 784, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The PCL framework selects the input-aware prompts from the layer-wise prompt pool and adds them to the backbone ViT to consider task knowledge. The $l$ -th layer has the prompt pool $P_{l} = \\{k_{l}^{1}:p_{l}^{1},\\dots,k_{l}^{M}:p_{l}^{M}\\}$ which contains $M$ prompt components $p\\in \\mathbb{R}^{L_p\\times D}$ and the corresponding key $k\\in \\mathbb{R}^D$ . Here, $L_{p}$ and $D$ stand for the prompt length and the feature dimension, respectively.", + "bbox": [ + 214, + 386, + 784, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prior PCL method consists of two feed-forward stages. In the first stage, a prompt query $q \\in \\mathbb{R}^D$ is extracted from pre-trained query ViT $Q(\\cdot)$ , utilizing the [CLS] token from the final layer.", + "bbox": [ + 214, + 462, + 784, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq = Q (x) _ {[ C L S ]}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 523, + 784, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $x$ is the given RGB image input. The extracted prompt query is used to form a prompt $\\phi_{l} \\in \\mathbb{R}^{L_{p} \\times D}$ from a prompt pool $P_{l}$ , which can be formulated as:", + "bbox": [ + 214, + 551, + 782, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {l} = g (q, P _ {l}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 602, + 784, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prompt formation function $g(\\cdot)$ has been a major contribution in the previous literature. L2P [40] and DualPrompt [39] select prompts having top-N similarity (e.g. cosine similarity) between query $q$ and prompt key $k_{l}$ . The recent state-of-the-art CodaPrompt [33] conducts a weighted summation of prompt components based on their similarity to enable end-to-end training.", + "bbox": [ + 212, + 626, + 782, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The obtained prompt $\\phi_l$ for layer $l$ is given to backbone ViT layer $f_l$ with input token embedding $x_l$ . This is the second feed-forward stage of the prior PCL methods.", + "bbox": [ + 212, + 702, + 782, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {l + 1} = f _ {l} \\left(x _ {l}, \\phi_ {l}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 752, + 784, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following DualPrompt [39] and CodaPrompt [33], we use prefix-tuning to add the information of a prompt $\\phi_l$ inside $f_{l}$ . The prefix-tuning splits prompt $\\phi_l$ into $[\\phi_k,\\phi_v]\\in \\mathbb{R}^{\\frac{L_p}{2}\\times D}$ , and then prepends them to the key and value inside the self-attention block of ViT. We utilize Multi-Head Self-Attention (MHSA) [8]", + "bbox": [ + 212, + 776, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 431, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3b6423ad60c4e9b747562e696bc5bf64cb4b7541c18d2e274f7bdab2ae65f4e6.jpg", + "image_caption": [ + "Fig. 2: We measure layer-wise feature distances between token embeddings of the model after training on task 1 and when a new task is learned. Each column represents the embeddings after training a model on Tasks 3, 5, 7, and 9. For instance, the left figure represents the distance between the token embeddings of Task 1 and those learned when Task 3 is completed. We train prompts using CodaPrompt [33] and use $1 - \\text{CosSim}(x, y)$ to measure the layer-wise distance on the training dataset. We use a CIFAR100 10-task setting. We provide more examples in the Supplementary Materials." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "like the prior PCL work, which computes the outputs from multiple single-head attention blocks.", + "bbox": [ + 212, + 392, + 785, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM H S A \\left(x, \\phi_ {k}, \\phi_ {v}\\right) = C o n c a t [ h e a d _ {1}, \\dots , h e a d _ {H} ] W _ {o}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 433, + 785, + 450 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {h e a d} _ {i} = \\operatorname {A t t e n t i o n} \\left(x W _ {q} ^ {i}, \\left[ \\phi_ {k}; x \\right] W _ {k} ^ {i}, \\left[ \\phi_ {v}; x \\right] W _ {v} ^ {i}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 474, + 785, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $W_{o}, W_{q}, W_{k}, W_{v}$ are the projection matrices. This prefix-tuning method is applied to the first 5 layers of backbone ViT. The leftover layers conduct a standard MHSA without prompts.", + "bbox": [ + 212, + 497, + 784, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, the prior PCL approach requires two-stage ViT feedforward (Eq. 1 and Eq. 3), which doubles computational cost. In our work, we aim to improve the computational efficiency of the PCL framework without degrading the performance.", + "bbox": [ + 212, + 542, + 785, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 One-stage Prompt-based Continual Learning", + "text_level": 1, + "bbox": [ + 214, + 626, + 684, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose to reduce the computational cost of PCL by restructuring two-step feedforward stages. To this end, we propose a new one-stage PCL framework called OS-Prompt. Instead of using a separate feedforward step to compute the prompt query $q$ from Eq. 1, we take a token embedding from the intermediate layer of a backbone ViT as the query (illustrated in Fig. 1).", + "bbox": [ + 212, + 657, + 787, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 How Stable are Token Embeddings across Continual Learning?", + "text_level": 1, + "bbox": [ + 214, + 753, + 779, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first ensure the validity of using intermediate token embedding as a prompt query. The original two-stage design employs a frozen pre-trained ViT, ensuring consistent prompt query representation throughout continual learning. It is essential to maintain a consistent (or similar) prompt query representation because", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "changes in prompt query would bring catastrophic forgetting. In our approach, token embeddings in the backbone ViT continually change as prompt tokens are updated during learning. For instance, after training on tasks 1 and 2, prompts have different values, resulting in varying token embeddings for an identical image. Consequently, it is crucial to assess if token embeddings sustain a consistent representation throughout the continual learning.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address the concern, we validate the deviation of the token embedding as prompt continual learning progresses. In Fig. 2, we show how this change happens at each layer during continuous learning tasks. From our study, two main observations can be highlighted: (1) When we add prompts, there is a larger difference in the deeper layers. For example, layers $1 \\sim 5$ have a small difference ( $\\leq 0.1$ ). However, the last layer shows a bigger change ( $\\geq 0.1$ ). (2) As learning continues, the earlier layers remain relatively stable, but the deeper layers change more. This observation concludes that although prompt tokens are included, the token embedding of early layers shows minor changes during training. Therefore, using token embeddings from these earlier layers would give a stable and consistent representation throughout the continual learning stages.", + "bbox": [ + 212, + 237, + 787, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 One-stage PCL Framework", + "text_level": 1, + "bbox": [ + 215, + 422, + 485, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Building on these observations, we employ the token embedding from the early layers as a prompt query to generate layer-wise prompt tokens. For a fair comparison, we implement prompts across layers 1 to 5, in line with prior work. The proposed OS-Prompt framework is illustrated in Fig. 3. For each layer, given the input token embedding, we directly use the [CLS] token as the prompt query. The original query selection equation (Eq. 1) can be reformulated as:", + "bbox": [ + 212, + 445, + 782, + 536 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq _ {l} = x _ {l _ {[ C L S ]}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 459, + 545, + 784, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the provided query $q_{l}$ , we generate a prompt from the prompt pool following the state-of-the-art CodaPrompt [33]. It is worth highlighting that our primary contribution lies not in introducing a new prompt generation technique but in presenting a more efficient framework for query selection. We first measure the cosine similarity $\\gamma(\\cdot)$ between $q_{l}$ and keys $\\{k_{l}^{1}, \\dots, k_{l}^{M}\\}$ , and then we perform a weighted summation of the corresponding value $p_{l}^{m}$ (i.e. prompt) based on the similarity.", + "bbox": [ + 212, + 568, + 787, + 672 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {l} = \\sum_ {m} \\gamma \\left(q _ {l}, k _ {l} ^ {m}\\right) p _ {l} ^ {m}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 674, + 784, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The generated prompt is then prepended to the image tokens. Notably, since we produce a prompt query without the need for an extra ViT feedforward, the computational overhead is reduced by approximately $50\\%$ for both training and inference.", + "bbox": [ + 212, + 707, + 782, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Query-Pool Regularization Loss", + "text_level": 1, + "bbox": [ + 215, + 787, + 522, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our OS-Prompt relies on intermediate token embeddings. As a result, the prompt query exhibits diminished representational capacity compared to previous PCL", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 431, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/777cd9e5a205306d527d4f9741d79c11c7e253c0101e074c379fc524c21faa31.jpg", + "image_caption": [ + "Fig. 3: Our OS-Prompt (OS-Prompt++) framework. An image passes through the backbone ViT layers to get the final prediction. From layer 1 to layer 5, we preprocess prompt tokens to the image tokens, which can be obtained in the following progress: ① We first compute the image token embedding from the previous layer. ② We use [CLS] token as a prompt query used for measuring cosine similarity between prompt keys inside the prompt pool. ③ Based on the similarity, we do weighted sum values to obtain prompt tokens. ④ To further improve the accuracy, we present OS-Prompt++. We integrate the query-pool regularization loss (dotted line), enabling the prompt pool to capture a stronger representation from the reference ViT." + ], + "image_footnote": [], + "bbox": [ + 271, + 142, + 727, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "approaches that utilize the [CLS] token from the final layer. This reduced capacity for representation brings performance degradation. To mitigate this, we introduce a Query-Pool Regularization (QR) loss. The QR loss ensures that the query-pool relationship becomes similar to that of the final layer's [CLS] token, thereby improving representation power. We extract the [CLS] token from the last layer of a reference ViT architecture $R(\\cdot)$ and employ it as our reference prompt query $r \\in \\mathbb{R}^{1 \\times D}$ where $D$ is the feature dimension:", + "bbox": [ + 212, + 479, + 787, + 585 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr = R (x). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 462, + 597, + 784, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let $K_{l} \\in \\mathbb{R}^{M \\times D}$ be the matrix representation of prompt keys at layer $l$ . We then define a similarity matrix $A_{query}^{l} \\in \\mathbb{R}^{M \\times 1}$ to capture the relationship between query $q_{l} \\in \\mathbb{R}^{1 \\times D}$ (from Eq. 6) and prompt keys. Similarly, we compute the similarity matrix $A_{ref}^{l}$ to represent the relationship between the reference prompt query $r$ and the prompt keys.", + "bbox": [ + 212, + 623, + 787, + 702 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nA _ {q u e r y} ^ {l} = S o f t m a x (\\frac {K _ {l} q _ {l} ^ {T}}{| | K _ {l} | | _ {2} | | q _ {l} | | _ {2}}) \\qquad A _ {r e f} ^ {l} = S o f t m a x (\\frac {K _ {l} r ^ {T}}{| | K _ {l} | | _ {2} | | r | | _ {2}}). \\qquad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 710, + 785, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We apply Softmax() to measure the relative similarity among query-key pairs, considering that $q$ and $r$ show distinct feature distributions originating from different layers. The QR loss penalizes the distance between two similarity matrices.", + "bbox": [ + 212, + 753, + 787, + 800 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {Q R} = \\sum_ {l} | | A _ {\\text {q u e r y}} ^ {l} - A _ {\\text {r e f}} ^ {l} | | _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 811, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1cfc3dab425fa3b188320e2a78c413f9a07bd17107a49fe4673981afcca09640.jpg", + "table_caption": [ + "Table 1: Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). $A_N$ represents the average accuracy across tasks, and $F_N$ indicates the mean forgetting rate. $\\uparrow$ and $\\downarrow$ indicate whether a metric is better with a higher or lower value, respectively. We average values over five runs with different seeds." + ], + "table_footnote": [], + "table_body": "
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
ER71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
LwF74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
OS-Prompt++77.07 ± 0.152.23 ± 0.1875.67 ± 0.401.27 ± 0.1073.77 ± 0.190.79 ± 0.07
", + "bbox": [ + 246, + 226, + 750, + 356 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6bdc05d6c285b58533fa2cdbabbfbc50e2de3cec8747153dfdeb4d0b1e3b0c3b.jpg", + "table_caption": [ + "Table 2: Accuracy comparison on CIFAR-100 10-task setting." + ], + "table_footnote": [], + "table_body": "
MethodAN(↑)FN(↓)
UB89.30-
ER76.20 ± 1.048.50 ± 0.37
Deep L2P84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CodaPrompt86.25 ± 0.741.67 ± 0.26
OS-Prompt86.42 ± 0.611.64 ± 0.14
OS-Prompt++86.68 ± 0.671.18 ± 0.21
", + "bbox": [ + 230, + 410, + 431, + 489 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0cd7c954be4d43b10e1fd4fd3281ed6bd08366fae230ecee7483f26fa005dd5a.jpg", + "table_caption": [ + "Table 3: GFLOPs comparison of PCL works. We provide relative cost $(\\%)$ with respect to L2P." + ], + "table_footnote": [], + "table_body": "
MethodTraining GFLOPsInference GFLOPs
L2P52.8 (100%)35.1 (100%)
Deep L2P52.8 (100%)35.1 (100%)
DualPrompt52.8 (100%)35.1 (100%)
CodaPrompt52.8 (100%)35.1 (100%)
OS-Prompt35.4 (66.7%)17.6 (50.1%)
OS-Prompt++52.8 (100%)17.6 (50.1%)
", + "bbox": [ + 473, + 410, + 753, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The QR loss is added to the cross-entropy loss $\\mathcal{L}_{CE}$ for classification. The total loss function can be written as:", + "bbox": [ + 212, + 518, + 784, + 546 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\lambda \\mathcal {L} _ {Q R}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 559, + 784, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter to balance between two loss terms. Importantly, the QR loss is applied during the training phase, ensuring that there is no computational overhead during inference. Note that we only train prompts within a prompt pool while freezing the other weight parameters.", + "bbox": [ + 212, + 584, + 787, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 666, + 375, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experiment Setting", + "text_level": 1, + "bbox": [ + 214, + 695, + 426, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We utilize the Split CIFAR-100 [21] and Split ImageNet-R [15] benchmarks for class-incremental continual learning. The Split CIFAR-100 partitions the original CIFAR-100 dataset into 10 distinct tasks, each comprising 10 classes. The Split ImageNet-R benchmark is an adaptation of ImageNet-R, encompassing diverse styles, including cartoon, graffiti, and origami. For our experiments, we segment ImageNet-R into 5, 10, or 20 distinct class groupings. Given its substantial intra-class diversity, the ImageNet-R benchmark is viewed as a particularly challenging benchmark. We also provide experiments on DomainNet [27], a", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 431, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "large-scale domain incremental learning dataset. The dataset consists of around 0.6 million images distributed across 345 classes. We use 5 different domains (Clipart $\\rightarrow$ Real $\\rightarrow$ Infograph $\\rightarrow$ Sketch $\\rightarrow$ Painting) as a continual learning task where each task consists of 69 classes. The DomainNet results are provided in the Supplementary Materials.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Details. We performed our experiments using the ViT-B/16 [8] pre-trained on ImageNet-1k, a standard backbone in earlier PCL research. To ensure a fair comparison with prior work, we maintain the same prompt length (8) and number of prompt components (100) as used in CodaPrompt. Like CodaPrompt, we divide the total prompt components into the number of tasks, and provide the partial component for each task. In task $N$ , the prompt components from task $1 \\sim N$ are frozen, and we only train the key and prompt components from task $N$ . We split $20\\%$ of the training set for hyperparameter $\\lambda$ tuning in Eq. 11. Our experimental setup is based on the PyTorch. The experiments utilize four Nvidia RTX2080ti GPUs. For robustness, we conducted our benchmarks with five different permutations of task class order, presenting both the average and standard deviation of the results. More detailed information can be found in the Supplementary Materials.", + "bbox": [ + 212, + 222, + 787, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use two metrics for evaluation: (1) Average final accuracy $A_N$ , which measures the overall accuracy across N tasks. [33,39,40] (2) Average forgetting $F_N$ , which tracks local performance drop over N tasks [24,33,34]. Note, $A_N$ is mainly used for performance comparison in literature.", + "bbox": [ + 212, + 420, + 785, + 481 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Comparison with prior PCL works", + "text_level": 1, + "bbox": [ + 214, + 507, + 547, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our OS-Prompt with prior continual learning methods. This includes non-PCL methods such as ER [6] and LWF [23], as well as PCL methods like L2P [40], DualPrompt [39], and CodaPrompt [33]. We present both the upper bound (UB) performance and results from fine-tuning (FT). UB refers to training a model with standard supervised training using all data (no continual learning). FT indicates that a model undergoes sequential training with continual learning data without incorporating prompt learning. Moreover, we report L2P/Deep L2P implementation from CodaPrompt [33], both in its original form and improved version by applying prompt pool through layers 1 to 5. In our comparisons, OS-Prompt represents our one-shot prompt framework without the QR loss, while OS-Prompt++ is with the QR loss.", + "bbox": [ + 212, + 537, + 787, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 shows the results on ImageNet-R. Our OS-Prompt indicates only a slight performance drop ( $\\leq 1\\%$ ) across the 5-task, 10-task, and 20-task settings. The reason for the performance drop could be the reduced representational capacity of a prompt query from the intermediate token embedding. However, OS-Prompt++ effectively counters this limitation by incorporating the QR loss. Such an observation underscores the significance of the relationship between the query and the prompt pool in PCL. OS-prompt++ shows slight performance improvement across all scenarios. This implies a prompt query that contains task information helps to enhance representation in the prompt pool, suggesting", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fddc9dad5e0a61ec60a1f8d45d7ebc571fcea1a26f6df5215cbd20d0d25fed33.jpg", + "table_caption": [ + "Table 4: Relative training computational cost with respect to ER." + ], + "table_footnote": [], + "table_body": "
CL strategyMethodTraining Complexity
Replay-basedER1
Regularization-basedLWF4/3
Prompt-basedL2P1
Prompt-basedDualPrompt1
Prompt-basedCodaPrompt1
Prompt-basedOS-Prompt (Ours)2/3
Prompt-basedOS-Prompt++ (Ours)1
", + "bbox": [ + 220, + 185, + 509, + 286 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a840d9a843763fb090de418c53bda3586283cc9af370cb99ac936164b6c814cb.jpg", + "table_caption": [ + "Table 5: ImageNet-R 10-task results with unsupervised pre-trained model [4]." + ], + "table_footnote": [], + "table_body": "
MethodAN(↑)FN(↓)
ER60.43 ± 1.1613.30 ± 0.12
LwF62.73 ± 1.134.32 ± 0.63
L2P60.32 ± 0.562.30 ± 0.11
DualPrompt61.77 ± 0.612.31 ± 0.23
CodaPrompt67.61 ± 0.192.23 ± 0.29
OS-Prompt67.52 ± 0.342.32 ± 0.13
OS-Prompt++67.92 ± 0.422.19 ± 0.26
", + "bbox": [ + 535, + 184, + 781, + 287 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "that exploring methods to integrate task information inside the prompt selection process could be an interesting research direction in PCL. Table 2 provides results on 10-task CIFAR-100, which shows a similar trend with ImageNet-R.", + "bbox": [ + 212, + 316, + 782, + 362 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Table 3, we provide GFLOPs for various PCL methods, rounding GFLOPs to one decimal place. It is worth noting that while prior PCL methods might have marginally different GFLOP values, they are close enough to appear identical in the table. This implies that different prompt formation schemes proposed in prior works do not substantially impact GFLOPs. The table shows OS-Prompt operates at $66.7\\%$ and $50.1\\%$ of the GFLOPs, during training and inference, respectively, relative to prior methods. While OS-Prompt++ maintains a training GFLOPs count comparable to", + "bbox": [ + 212, + 363, + 517, + 589 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f627728d5724d88c29b20fb157107406b148f3ddbe26e47fa8b11df005fbd2e2.jpg", + "image_caption": [ + "Fig. 4: Comparison of latency across PCL methods on different GPUs." + ], + "image_footnote": [], + "bbox": [ + 529, + 396, + 782, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "earlier work due to the added feedforward process in the reference ViT, it does not employ the reference ViT during inference, bringing it down to $50.1\\%$ GFLOPs. In Fig. 4, we compare the inference GPU latency of previous PCL methods with our OS-Prompt across three distinct GPU setups. All measurements are taken using a batch size of 32. Similar to the trend observed in FLOPs, our approach reduces the latency by $\\sim 50\\%$ .", + "bbox": [ + 212, + 589, + 787, + 679 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Discussion on Training Computational Cost", + "text_level": 1, + "bbox": [ + 214, + 705, + 622, + 722 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we provide an in-depth discussion of the training computational cost of PCL methods. Following [10], Table 4 presents the relative training complexity of each method with respect to ER [6], which is the simple baseline with standard gradient-based training.", + "bbox": [ + 212, + 733, + 782, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Here, we would like to clarify the training computational cost of prompt learning. In general neural network training, the forward-backward computational cost ratio is approximately 1:2. This is due to gradient backpropagation", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 431, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$\\left(\\frac{dL}{da_l} = W_{l + 1}\\frac{dL}{da_{l + 1}}\\right)$ and weight-updating $\\left(\\frac{dL}{dW_l} = \\frac{dL}{da_{l + 1}} a_l\\right)$ , where $L$ represents the loss, $W$ denotes the weight parameter, $a$ is the activation, and $l$ is the layer index. In prompt learning, the forward-backward computational cost ratio is approximately 1:1. This is in contrast to general neural network training, as only a small fraction of the weight parameters (less than $1\\%$ ) are updated.", + "bbox": [ + 212, + 143, + 787, + 223 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Bearing this in mind, we present the observations derived from the table, assuming that all methods employ the same architecture. (1) Previous PCL (L2P, DualPrompt, CodaPrompt) consists of two steps: First, the query ViT requires only feedforward without backpropagation, Second, the backbone ViT requires feedforward-backward training with prompt tuning. This results in the relative training computational cost $1 \\left( = \\frac{\\text{QueryViT}_{fw}(1) + \\text{BackboneViT}_{fw}(1) + \\text{BackboneViT}_{bw}(1)}{\\text{ER}_{fw}(1) + \\text{ER}_{bw}(2)} \\right)$ . Here, $fw$ and $bw$ denote forward and backward propagation respectively. (2) Our OS-Prompt requires one feedforward-backward step like standard training. Therefore, relative training cost becomes $\\frac{2}{3} \\left( = \\frac{\\text{BackboneViT}_{fw}(1) + \\text{BackboneViT}_{bw}(1)}{\\text{ER}_{fw}(1) + \\text{ER}_{bw}(2)} \\right)$ . This shows the potential of our OS-Prompt on online continual learning. (3) Our OS-Prompt++ has a reference ViT to perform the regularization, bringing 1 relative training computational cost with respect to ER.", + "bbox": [ + 212, + 223, + 794, + 415 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Impact of Unsupervised Pre-trained Weights", + "text_level": 1, + "bbox": [ + 215, + 436, + 632, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "One underlying assumption of existing PCL methods is their reliance on supervised pretraining on ImageNet-1k. While this pre-training has a substantial impact on model performance, it may not always be feasible for a general continual learning task. One possible solution is using unsupervised pre-trained models. To explore this, we use DINO pre-trained weights [4] instead of ImageNet-1k pre-trained weights, and compare the performance across ER, LwF, and PCL works. We set the other experimental settings as identical. We report the results with ImageNet-R 10-task in Table 5. Compared to the ImageNet-1k pre-trained model (with supervised training), performances across all methods degrade. This observation suggests that the backbone model plays a crucial role in PCL. Moreover, similar to the ImageNet-1k pre-trained model, CodaPrompt and our OS-Prompt outperform the other methods by a significant margin. This indicates that our method continues to perform well with various backbones.", + "bbox": [ + 212, + 460, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Analysis of Design Components", + "text_level": 1, + "bbox": [ + 215, + 679, + 522, + 695 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "QR Loss Design. To understand the effect of different components in our proposed QR loss (Eq. 10), we conducted an ablation study with different settings, summarized in Table 6. There are two main components in the QR loss: Cosine Similarity and Softmax, and we provide the accuracy of different four combinations. Without Cosine Similarity and Softmax, our framework yields a performance of $75.00\\%$ for $A_{N}$ . Adding Cosine Similarity or Softmax improves the performance, suggesting their collaborative role in enhancing the model's performance in our QR loss design. We also provide the sensitivity analysis of our method with respect to the hyperparameter $\\lambda$ in Table 7. Across the three", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/45e976f8e84ad191de8903a8c8a3b34e8033987862937f3e2f2a61dbe46918b6.jpg", + "table_caption": [ + "Table 6: Analysis of QR loss design on 10-task ImageNet-R setting." + ], + "table_footnote": [], + "table_body": "
CosSim SoftmaxAN(↑)FN(↓)
✓75.00 ± 0.531.68 ± 0.12
75.47 ± 0.421.38 ± 0.16
✓75.51 ± 0.331.28 ± 0.06
✓75.67 ± 0.401.27 ± 0.10
", + "bbox": [ + 222, + 184, + 483, + 253 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a0af50cd6f9ff49a818b13de55fbf4428b0ec66dd1e96afc46fe3a1bfeac75c5.jpg", + "table_caption": [ + "Table 7: Hyperparameter sensitivity study of $\\lambda$ on ImageNet-R 5/10/20-task." + ], + "table_footnote": [], + "table_body": "
λTask-5Task-10Task-20
1e-577.03 ± 0.1075.63 ± 0.3973.63 ± 0.21
5e-577.02 ± 0.1375.62 ± 0.4173.62 ± 0.19
1e-477.07 ± 0.1575.67 ± 0.4073.77 ± 0.19
5e-477.13 ± 0.2475.68 ± 0.3873.68 ± 0.17
", + "bbox": [ + 509, + 184, + 781, + 252 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b2c63dc9d14f81e3ae7306c59a02bccfe8844649d1e5342720f1f1b238492e30.jpg", + "image_caption": [ + "Fig. 5: Analysis of accuracy $A_N$ with respect to prompt components (Left) and prompt length (Right). We use 10-taks ImageNet-R setting." + ], + "image_footnote": [], + "bbox": [ + 232, + 268, + 390, + 367 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f23e8a4243edee892875192a0ffc089383e3bd5b9e8628339e53d492b9bae222.jpg", + "image_caption": [ + "Fig.6: Trade-off between GFLOPs and accuracy within a reference ViT." + ], + "image_footnote": [], + "bbox": [ + 405, + 268, + 562, + 368 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/af6f6432b03f12338a895483944c790af25cdaa85d480daa8485efd18395809a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 268, + 782, + 367 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "distinct task configurations on ImageNet-R, we note that the performance fluctuations are minimal, despite the variation in $\\lambda$ values. These results underline the robustness of our method with respect to $\\lambda$ value, suggesting that our approach is not sensitive to hyperparameter.", + "bbox": [ + 212, + 454, + 784, + 515 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Prompt Design. We further examine the impact of both the number of prompt components and the prompt length on accuracy. In Fig. 5 (left), we measure accuracy across configurations with $\\{10,20,50,100,200,500\\}$ prompts. The OS-prompt demonstrates a consistent accuracy enhancement as the number of prompts increases. Notably, OS-prompt++ reaches a performance plateau after just 50 prompts. Our ablation study of prompt length, presented in Fig. 5 (right), reveals that ours maintains stable accuracy across various prompt lengths.", + "bbox": [ + 212, + 515, + 799, + 621 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.6 Accuracy-Efficiency Trade-off within a Reference ViT", + "text_level": 1, + "bbox": [ + 214, + 646, + 702, + 661 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "During the training of OS-prompt++, our method does not improve the energy efficiency (Note, we achieve $\\sim 50\\%$ computational saving during inference). This arises from our approach of extracting the reference prompt query $r$ from the final layer of the reference ViT. To further enhance energy efficiency during the OS-prompt++ training, we delve into the trade-off between accuracy and efficiency within the Reference ViT. Instead of relying on the last layer's [CLS] token embedding for the reference prompt query, we opt for intermediate token embeddings. Since our prompt pool is applied in layers 1 to 5 of the backbone ViT, we utilize the intermediate token embeddings from layers deeper than 5 within the reference ViT. Fig. 6 illustrates this trade-off with respect to the layer index where we get the reference prompt. Our findings show that util", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/12e2dccc6968af54314382a748e34c8e4be813e6c45e0e99ece564544947913c.jpg", + "table_caption": [ + "Table 8: Analysis on prompt formation methods. Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). $A_{N}$ represents the average accuracy across tasks, and $F_{N}$ indicates the mean forgetting rate. We average values over five runs with different seeds." + ], + "table_footnote": [], + "table_body": "
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
OS-Prompt (L2P)74.61 ± 0.292.29 ± 0.2173.43 ± 0.601.42 ± 0.1071.42 ± 0.240.95 ± 0.06
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
OS-Prompt (Dual)74.65 ± 0.152.06 ± 0.2072.57 ± 0.231.13 ± 0.0370.55 ± 0.480.81 ± 0.10
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt (Coda)75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
", + "bbox": [ + 269, + 212, + 727, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "layer 8, there is only a slight increase in accuracy, which suggests a potential to reduce GFLOPs without performance drop.", + "bbox": [ + 212, + 337, + 782, + 368 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.7 Analysis on Prompt Formation Strategy", + "text_level": 1, + "bbox": [ + 215, + 387, + 591, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here, we delve into the impact of varying prompt formation strategies on performance. For comparative insight, we present results from two previous PCL methodologies: L2P and DualPrompt. While L2P selects 5 prompts from the internal prompt pool, DualPrompt distinguishes between general and task-specific prompts. Aside from the distinct prompt formation strategies, the configurations remain consistent. In Table 8, we adopt the prompt formation strategies proposed in L2P and Dual, resulting in configurations denoted as OS-Prompt (L2P) and OS-Prompt (Dual). Our findings reveal the following: (1) The effectiveness of OS-prompt varies based on the prompt formation technique employed. Notably, OS-Prompt (L2P) and OS-Prompt (Dual) yield lower accuracy than our original approach, which relies on CodaPrompt. This suggests that our performance could benefit from refined prompt formation techniques in future iterations. (2) The OS-prompt framework, when integrated with L2P and Dual prompt formation strategies, outperforms the original Deep L2P and DualPrompt, a trend not observed with CodaPrompt. This may indicate that the prompt formation strategies of L2P and DualPrompt, which rely on hard matching (i.e., top-k), are more resilient than CodaPrompt's. Conversely, CodaPrompt employs soft matching, utilizing a weighted summation of all prompts based on proximity.", + "bbox": [ + 212, + 410, + 787, + 684 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 704, + 359, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce the OS-Prompt framework where we improve the efficiency of the conventional two-stage PCL in a simple yet effective manner. One of the potential limitations of our OS-Prompt++ is, in comparison to OS-Prompt (our light version), it introduces an increase in training computational cost from the reference ViT feedforward. This computational efficiency during training becomes important in the context of addressing the online continual learning problem [10] where the model needs rapid training on the streaming data.", + "bbox": [ + 212, + 733, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), and the DoE MMICC center SEA-CROGS (Award #DE-SC0023198)", + "bbox": [ + 215, + 180, + 784, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 268, + 321, + 282 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: ECCV (2018) 3", + "2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: NeurIPS (2020) 3", + "3. Cai, H., Gan, C., Zhu, L., Han, S.: Tinytl: Reduce memory, not parameters for efficient on-device learning. Advances in Neural Information Processing Systems 33, 11285-11297 (2020) 4", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 11, 12", + "5. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018) 3", + "6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 3, 10, 11", + "7. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 1, 3", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 2, 5, 10", + "9. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 1, 3", + "0. Ghunaim, Y., Bibi, A., Alhamoud, K., Alfarra, M., Al Kader Hammoud, H.A., Prabhu, A., Torr, P.H., Ghanem, B.: Real-time evaluation in online continual learning: A new hope. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11888-11897 (2023) 11, 14", + "1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020) 1", + "12. Harun, M.Y., Gallardo, J., Hayes, T.L., Kanan, C.: How efficient are today's continual learning algorithms? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2430-2435 (2023) 2", + "13. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: ICRA (2019) 3", + "14. He, X., Li, C., Zhang, P., Yang, J., Wang, X.E.: Parameter-efficient fine-tuning for vision transformers. arXiv preprint arXiv:2203.16329 (2022) 4" + ], + "bbox": [ + 218, + 303, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. arXiv preprint arXiv:2006.16241 (2020) 9", + "16. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV. pp. 8340-8349 (2021) 3, 5", + "17. Hsu, Y.C., Liu, Y.C., Ramasamy, A., Kira, Z.: Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488 (2018) 5", + "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4", + "19. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022) 4", + "20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4", + "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 3, 5, 9", + "22. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: ICML. pp. 3925-3934. PMLR (2019) 3", + "23. Li, Z., Hoiem, D.: Learning without forgetting. TPAMI 40(12), 2935-2947 (2017) 3, 10", + "24. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS (2017) 10", + "25. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022) 1", + "26. Pellegrini, L., Lomonaco, V., Graffieti, G., Maltoni, D.: Continual learning at the edge: Real-time training on smartphone devices. arXiv preprint arXiv:2105.13127 (2021) 2", + "27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 3, 9", + "28. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018) 4", + "29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 3", + "30. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) 3, 4", + "31. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: ICML. pp. 4548-4557 (2018) 3", + "32. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 1, 3" + ], + "bbox": [ + 215, + 146, + 785, + 838 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Kim, Y. Li, P. Panda.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) 1, 2, 3, 4, 5, 6, 7, 10", + "34. Smith, J.S., Tian, J., Halbe, S., Hsu, Y.C., Kira, Z.: A closer look at rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2409-2419 (2023) 10", + "35. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023) 4", + "36. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017) 1", + "37. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 2", + "38. Wang, Z., Zhan, Z., Gong, Y., Yuan, G., Niu, W., Jian, T., Ren, B., Ioannidis, S., Wang, Y., Dy, J.: Sparcl: Sparse continual learning on the edge. Advances in Neural Information Processing Systems 35, 20366-20380 (2022) 2", + "39. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022) 1, 2, 4, 5, 10", + "40. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. CVPR (2022) 1, 2, 4, 5, 10", + "41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020) 1, 3", + "42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017) 3", + "43. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) 4", + "44. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017) 3", + "45. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 698-714. Springer (2020) 4", + "46. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021) 4", + "47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4" + ], + "bbox": [ + 212, + 146, + 787, + 797 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "One-stage Prompt-based Continual Learning", + "bbox": [ + 431, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_model.json b/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d2e377bbba710c4729e20d858c1c3953c007d899 --- /dev/null +++ b/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_model.json @@ -0,0 +1,2313 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.142, + 0.756, + 0.163 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.19, + 0.711, + 0.206 + ], + "angle": 0, + "content": "Youngeun Kim, Yuhang Li, and Priyadarshini Panda" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.218, + 0.677, + 0.247 + ], + "angle": 0, + "content": "Yale University, CT, USA {youngun.kim,yuhang.li, priya.panda}@yale.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.283, + 0.74, + 0.587 + ], + "angle": 0, + "content": "Abstract. Prompt-based Continual Learning (PCL) has gained considerable attention as a promising continual learning solution because it achieves state-of-the-art performance while preventing privacy violations and memory overhead problems. Nonetheless, existing PCL approaches face significant computational burdens because of two Vision Transformer (ViT) feed-forward stages; one is for the query ViT that generates a prompt query to select prompts inside a prompt pool; the other one is a backbone ViT that mixes information between selected prompts and image tokens. To address this, we introduce a one-stage PCL framework by directly using the intermediate layer's token embedding as a prompt query. This design removes the need for an additional feed-forward stage for query ViT, resulting in \\(\\sim 50\\%\\) computational cost reduction for both training and inference with marginal accuracy drop \\((\\leq 1\\%)\\). We further introduce a Query-Pool Regularization (QR) loss that regulates the relationship between the prompt query and the prompt pool to improve representation power. The QR loss is only applied during training time, so there is no computational overhead at inference from the QR loss. With the QR loss, our approach maintains \\(\\sim 50\\%\\) computational cost reduction during inference as well as outperforms the prior two-stage PCL methods by \\(\\sim 1.4\\%\\) on public class-incremental continual learning benchmarks including CIFAR-100, ImageNet-R, and DomainNet." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.602, + 0.731, + 0.617 + ], + "angle": 0, + "content": "Keywords: Efficient learning \\(\\cdot\\) Continual learning \\(\\cdot\\) Transfer learning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.643, + 0.377, + 0.658 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training models effectively and efficiently on a continuous stream of data presents a significant practical hurdle. A straightforward approach would entail accumulating both prior and new data and then updating the model using this comprehensive dataset. However, as data volume grows, fully retraining a model on such extensive data becomes increasingly impractical [11, 25]. Additionally, storing past data can raise privacy issues, such as those highlighted by the EU General Data Protection Regulation (GDPR) [36]. An alternative solution is to adapt the model based solely on currently incoming data, eschewing any access to past data. This paradigm is termed as rehearsal-free continual learning [7,9,32,33,39-41], and the primary goal is to diminish the effects of catastrophic forgetting on previously acquired data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.414, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.145, + 0.605, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.15, + 0.782, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 1: Difference between prior PCL work and ours. Prior PCL work (Left) has two feed-forward stages for (1) a query function (ViT) to select input-specific prompts and (2) a backbone ViT layer to perform prompt learning with the selected prompts and image tokens. On the other hand, our one-stage PCL framework (Middle) uses an intermediate layer's token embedding as a prompt query so that it requires only one backbone ViT feed-forward stage. As a result, our method reduces GFLOPs by \\(\\sim 50\\%\\) compared to prior work while maintaining accuracy (Right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Among the rehearsal-free continual learning methods, Prompt-based Continual Learning (PCL) stands out as it has demonstrated state-of-the-art performance in image classification tasks, even surpassing rehearsal-based methods [33, 37, 39, 40]. PCL utilizes a pre-trained Vision Transformer (ViT) [8] and refines the model by training learnable tokens on the given data. PCL adopts a prompt pool-based training scheme where different prompts are selected and trained for each continual learning stage. This strategy enables a model to learn the information of the training data in a sequential manner with less memory overhead, as the prompt pool requires minimal resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Although PCL methods show state-of-the-art performance, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices [12,26,38]. Specifically, the PCL method requires two-stage ViT feedforward steps. One is for the query function that generates a prompt query. The other one is a backbone ViT that mixes information between selected prompts and input image tokens. We refer to this approach as a two-stage PCL method, illustrated in Fig. 1 Left." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address this limitation, we propose a one-stage PCL framework with only one ViT feed-forward stage (Fig. 1 Middle), called OS-Prompt. Rather than deploying a separate ViT feed-forward phase to generate a prompt query, we directly use the intermediate layer's token embedding as a query. This is based on our observation that early layers show marginal shifts in the feature space during continual prompt learning (details are in Section 4.1). This observation enables us to use the intermediate token embedding as a prompt query which requires consistent representation across continual learning to minimize catastrophic forgetting. Surprisingly, OS-Prompt shows a marginal performance drop (less than \\(1\\%\\)) while saving \\(\\sim 50\\%\\) of computational cost (Fig. 1 Right)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.434, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.282 + ], + "angle": 0, + "content": "As our OS-Prompt uses the intermediate layer's token embedding as a query instead of the last layer's token embedding (which is used in prior PCL methods), there is a slight performance drop due to the lack of representation power. To address this, we introduce a Query-Pool Regularization (QR) loss, which is designed to enhance the representation power of a prompt pool by utilizing the token embedding from the last layer. Importantly, the QR loss is applied only during training, ensuring no added computational burden during inference. We refer to our enhanced one-stage framework with QR loss as OS-Prompt++. Our OS-Prompt++ bridges the accuracy gap, performing better than OS-Prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.479 + ], + "angle": 0, + "content": "Overall, our contribution can be summarized as follows: (1) We raise a drawback of the current PCL methods — high computational cost, particularly due to the two distinct ViT feed-forward stages. As a solution to the computational inefficiency in existing PCL techniques, we propose OS-Prompt that reduces the computational cost by nearly \\(50\\%\\) without any significant performance drop. (2) To counter the slight performance degradation observed in our one-stage PCL framework, we introduce a QR loss. This ensures that the prompt pool maintains similarity with token embedding from both intermediate and final layers. Notably, our QR loss avoids any additional computational overhead during inference. (3) We conduct experiments with rehearsal-free continual learning setting on CIFAR-100 [21], ImageNet-R [16], and DomainNet [27] benchmarks, outperforming the previous SOTA method CodaPrompt [33] by \\(\\sim 1.4\\%\\) with \\(\\sim 50\\%\\) computational cost saving." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.388, + 0.516 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.422, + 0.546 + ], + "angle": 0, + "content": "2.1 Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For a decade, continual learning has been explored as an important research topic, concentrating on the progressive adaptation of models over successive tasks or datasets. One of the representative methods to address continual learning is the regularization-based method [1, 23, 44]. By adding a regularization loss between the current model and the previous model, these methods aim to minimize catastrophic forgetting. However, their performance is relatively low on challenging datasets compared to the other continual learning methods. An alternative approach proposes to expand the network architecture as it progresses through different continual learning stages [22, 30, 31, 42]. Though these typically surpass the results of regularization-based methods, they come at the expense of significant memory consumption due to the increased parameters. Rehearsal-based continual learning has introduced a component for archiving previous data [2, 5, 6, 13, 29]. By leveraging accumulated data during the subsequent stages, they often outperform other continual learning methods. However, saving images causes memory overhead and privacy problems. Given the inherent limitations, a growing interest is observed in rehearsal-free continual learning. These are crafted to address catastrophic forgetting without access to past data. Most of them propose methods to generate rehearsal images [7, 9, 32, 41], but generating images is resource-heavy and time-consuming. Recently, within" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.421 + ], + "angle": 0, + "content": "the rehearsal-free domain, Prompt-based Continual Learning (PCL) methods have gained considerable attention because of their performance and memory efficiency. By utilizing a pre-trained ViT model, they train only a small set of parameters called prompts to maintain the information against catastrophic forgetting. For example, L2P [40] utilizes a prompt pool, selecting the appropriate prompts for the given image. Extending this principle, DualPrompt [39] proposes general prompts for encompassing knowledge across multiple continual learning stages. Advancing this domain further, [33] facilitates end-to-end training of the prompt pool, achieving state-of-the-art performance. However, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices. Our work aims to resolve such computational complexity overhead in PCL. Furthermore, Tang et al. [35] have proposed a prompt-based continual learning that generates prompts based on the intermediate features from the transformer. However, this prior work focuses on maximizing continual learning performance without a pre-trained model, while ours aims to leverage the power of pre-trained models. Also, the prior work focuses on the prompt generation method, while we propose a framework compatible with different prompt learning techniques." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.442, + 0.459, + 0.459 + ], + "angle": 0, + "content": "2.2 Prompt-based Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.668 + ], + "angle": 0, + "content": "The efficient fine-tuning method for large pre-trained models has shown their practical benefits across various machine learning tasks [14, 18, 28, 45-47]. Instead of adjusting all parameters within neural architectures, the emphasis has shifted to leveraging a minimal set of weights for optimal transfer outcomes. In alignment with this, multiple methodologies [3, 30] have integrated a streamlined bottleneck component within the transformer framework, thus constraining gradient evaluations to select parameters. Strategies like TinyTL [3] and BitFit [43] advocate for bias modifications in the fine-tuning process. In a recent shift, prompt-based learning [19, 20] captures task-specific knowledge with much smaller additional parameters than the other fine-tuning methods [40]. Also, prompt-based learning only requires storing several prompt tokens, which are easy to plug-and-play, and hence, they are used to construct a prompt pool for recent continual learning methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.371, + 0.708 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.402, + 0.74 + ], + "angle": 0, + "content": "3.1 Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In a continual learning setting, a model is trained on \\( T \\) continual learning stages with dataset \\( D = \\{D_{1},D_{2},\\dots,D_{T}\\} \\), where \\( D_{t} \\) is the data provided in \\( t \\)-th stage. Our problem is based on an image recognition task, so each data \\( D_{t} \\) consists of pairs of images and labels. Also, data from the previous tasks is not accessible for future tasks. Following the previous rehearsal-free continual learning settings [33, 39], we focus on the class-incremental continual learning setting where task" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.433, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.224 + ], + "angle": 0, + "content": "identity is unknown at inference. This is a challenging scenario compared to others such as task-incremental continual learning where the task labels are provided for both training and test phases [17]. For the experiments, we split classes into \\( T \\) chunks with continual learning benchmarks including CIFAR-100 [21] and ImageNet-R [16]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.25, + 0.632, + 0.266 + ], + "angle": 0, + "content": "3.2 Two-stage Prompt-based Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.785, + 0.386 + ], + "angle": 0, + "content": "In our framework, we focus on improving efficiency through a new query selection rather than changing the way prompts are formed from the prompt pool with the given query, a common focus in earlier PCL studies [33, 39, 40]. For a fair comparison with earlier work, we follow the overall PCL framework established in previous studies. Given that our contribution complements the prompt formation technique, our method can seamlessly work with future works that propose a stronger prompt generation method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.387, + 0.785, + 0.463 + ], + "angle": 0, + "content": "The PCL framework selects the input-aware prompts from the layer-wise prompt pool and adds them to the backbone ViT to consider task knowledge. The \\(l\\)-th layer has the prompt pool \\(P_{l} = \\{k_{l}^{1}:p_{l}^{1},\\dots,k_{l}^{M}:p_{l}^{M}\\}\\) which contains \\(M\\) prompt components \\(p\\in \\mathbb{R}^{L_p\\times D}\\) and the corresponding key \\(k\\in \\mathbb{R}^D\\). Here, \\(L_{p}\\) and \\(D\\) stand for the prompt length and the feature dimension, respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.463, + 0.785, + 0.509 + ], + "angle": 0, + "content": "The prior PCL method consists of two feed-forward stages. In the first stage, a prompt query \\( q \\in \\mathbb{R}^D \\) is extracted from pre-trained query ViT \\( Q(\\cdot) \\), utilizing the [CLS] token from the final layer." + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.525, + 0.785, + 0.541 + ], + "angle": 0, + "content": "\\[\nq = Q (x) _ {[ C L S ]}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.553, + 0.784, + 0.584 + ], + "angle": 0, + "content": "Here \\( x \\) is the given RGB image input. The extracted prompt query is used to form a prompt \\( \\phi_{l} \\in \\mathbb{R}^{L_{p} \\times D} \\) from a prompt pool \\( P_{l} \\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.603, + 0.785, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\phi_ {l} = g (q, P _ {l}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.784, + 0.702 + ], + "angle": 0, + "content": "The prompt formation function \\( g(\\cdot) \\) has been a major contribution in the previous literature. L2P [40] and DualPrompt [39] select prompts having top-N similarity (e.g. cosine similarity) between query \\( q \\) and prompt key \\( k_{l} \\). The recent state-of-the-art CodaPrompt [33] conducts a weighted summation of prompt components based on their similarity to enable end-to-end training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.784, + 0.747 + ], + "angle": 0, + "content": "The obtained prompt \\(\\phi_l\\) for layer \\(l\\) is given to backbone ViT layer \\(f_l\\) with input token embedding \\(x_l\\). This is the second feed-forward stage of the prior PCL methods." + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.753, + 0.785, + 0.767 + ], + "angle": 0, + "content": "\\[\nx _ {l + 1} = f _ {l} \\left(x _ {l}, \\phi_ {l}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Following DualPrompt [39] and CodaPrompt [33], we use prefix-tuning to add the information of a prompt \\(\\phi_l\\) inside \\(f_{l}\\). The prefix-tuning splits prompt \\(\\phi_l\\) into \\([\\phi_k,\\phi_v]\\in \\mathbb{R}^{\\frac{L_p}{2}\\times D}\\), and then prepends them to the key and value inside the self-attention block of ViT. We utilize Multi-Head Self-Attention (MHSA) [8]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.782, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.266, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 2: We measure layer-wise feature distances between token embeddings of the model after training on task 1 and when a new task is learned. Each column represents the embeddings after training a model on Tasks 3, 5, 7, and 9. For instance, the left figure represents the distance between the token embeddings of Task 1 and those learned when Task 3 is completed. We train prompts using CodaPrompt [33] and use \\(1 - \\text{CosSim}(x, y)\\) to measure the layer-wise distance on the training dataset. We use a CIFAR100 10-task setting. We provide more examples in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.786, + 0.423 + ], + "angle": 0, + "content": "like the prior PCL work, which computes the outputs from multiple single-head attention blocks." + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.434, + 0.786, + 0.452 + ], + "angle": 0, + "content": "\\[\nM H S A \\left(x, \\phi_ {k}, \\phi_ {v}\\right) = C o n c a t [ h e a d _ {1}, \\dots , h e a d _ {H} ] W _ {o}. \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.475, + 0.786, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\operatorname {h e a d} _ {i} = \\operatorname {A t t e n t i o n} \\left(x W _ {q} ^ {i}, \\left[ \\phi_ {k}; x \\right] W _ {k} ^ {i}, \\left[ \\phi_ {v}; x \\right] W _ {v} ^ {i}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.785, + 0.543 + ], + "angle": 0, + "content": "Here, \\( W_{o}, W_{q}, W_{k}, W_{v} \\) are the projection matrices. This prefix-tuning method is applied to the first 5 layers of backbone ViT. The leftover layers conduct a standard MHSA without prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.787, + 0.605 + ], + "angle": 0, + "content": "However, the prior PCL approach requires two-stage ViT feedforward (Eq. 1 and Eq. 3), which doubles computational cost. In our work, we aim to improve the computational efficiency of the PCL framework without degrading the performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.685, + 0.646 + ], + "angle": 0, + "content": "4 One-stage Prompt-based Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.735 + ], + "angle": 0, + "content": "We propose to reduce the computational cost of PCL by restructuring two-step feedforward stages. To this end, we propose a new one-stage PCL framework called OS-Prompt. Instead of using a separate feedforward step to compute the prompt query \\( q \\) from Eq. 1, we take a token embedding from the intermediate layer of a backbone ViT as the query (illustrated in Fig. 1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.78, + 0.772 + ], + "angle": 0, + "content": "4.1 How Stable are Token Embeddings across Continual Learning?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We first ensure the validity of using intermediate token embedding as a prompt query. The original two-stage design employs a frozen pre-trained ViT, ensuring consistent prompt query representation throughout continual learning. It is essential to maintain a consistent (or similar) prompt query representation because" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.433, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "changes in prompt query would bring catastrophic forgetting. In our approach, token embeddings in the backbone ViT continually change as prompt tokens are updated during learning. For instance, after training on tasks 1 and 2, prompts have different values, resulting in varying token embeddings for an identical image. Consequently, it is crucial to assess if token embeddings sustain a consistent representation throughout the continual learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.404 + ], + "angle": 0, + "content": "To address the concern, we validate the deviation of the token embedding as prompt continual learning progresses. In Fig. 2, we show how this change happens at each layer during continuous learning tasks. From our study, two main observations can be highlighted: (1) When we add prompts, there is a larger difference in the deeper layers. For example, layers \\(1 \\sim 5\\) have a small difference (\\(\\leq 0.1\\)). However, the last layer shows a bigger change (\\(\\geq 0.1\\)). (2) As learning continues, the earlier layers remain relatively stable, but the deeper layers change more. This observation concludes that although prompt tokens are included, the token embedding of early layers shows minor changes during training. Therefore, using token embeddings from these earlier layers would give a stable and consistent representation throughout the continual learning stages." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.423, + 0.486, + 0.438 + ], + "angle": 0, + "content": "4.2 One-stage PCL Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.784, + 0.537 + ], + "angle": 0, + "content": "Building on these observations, we employ the token embedding from the early layers as a prompt query to generate layer-wise prompt tokens. For a fair comparison, we implement prompts across layers 1 to 5, in line with prior work. The proposed OS-Prompt framework is illustrated in Fig. 3. For each layer, given the input token embedding, we directly use the [CLS] token as the prompt query. The original query selection equation (Eq. 1) can be reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.46, + 0.546, + 0.785, + 0.563 + ], + "angle": 0, + "content": "\\[\nq _ {l} = x _ {l _ {[ C L S ]}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Using the provided query \\( q_{l} \\), we generate a prompt from the prompt pool following the state-of-the-art CodaPrompt [33]. It is worth highlighting that our primary contribution lies not in introducing a new prompt generation technique but in presenting a more efficient framework for query selection. We first measure the cosine similarity \\( \\gamma(\\cdot) \\) between \\( q_{l} \\) and keys \\( \\{k_{l}^{1}, \\dots, k_{l}^{M}\\} \\), and then we perform a weighted summation of the corresponding value \\( p_{l}^{m} \\) (i.e. prompt) based on the similarity." + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.675, + 0.785, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\phi_ {l} = \\sum_ {m} \\gamma \\left(q _ {l}, k _ {l} ^ {m}\\right) p _ {l} ^ {m}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.784, + 0.768 + ], + "angle": 0, + "content": "The generated prompt is then prepended to the image tokens. Notably, since we produce a prompt query without the need for an extra ViT feedforward, the computational overhead is reduced by approximately \\(50\\%\\) for both training and inference." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.524, + 0.803 + ], + "angle": 0, + "content": "4.3 Query-Pool Regularization Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Our OS-Prompt relies on intermediate token embeddings. As a result, the prompt query exhibits diminished representational capacity compared to previous PCL" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.143, + 0.728, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 3: Our OS-Prompt (OS-Prompt++) framework. An image passes through the backbone ViT layers to get the final prediction. From layer 1 to layer 5, we preprocess prompt tokens to the image tokens, which can be obtained in the following progress: ① We first compute the image token embedding from the previous layer. ② We use [CLS] token as a prompt query used for measuring cosine similarity between prompt keys inside the prompt pool. ③ Based on the similarity, we do weighted sum values to obtain prompt tokens. ④ To further improve the accuracy, we present OS-Prompt++. We integrate the query-pool regularization loss (dotted line), enabling the prompt pool to capture a stronger representation from the reference ViT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.587 + ], + "angle": 0, + "content": "approaches that utilize the [CLS] token from the final layer. This reduced capacity for representation brings performance degradation. To mitigate this, we introduce a Query-Pool Regularization (QR) loss. The QR loss ensures that the query-pool relationship becomes similar to that of the final layer's [CLS] token, thereby improving representation power. We extract the [CLS] token from the last layer of a reference ViT architecture \\( R(\\cdot) \\) and employ it as our reference prompt query \\( r \\in \\mathbb{R}^{1 \\times D} \\) where \\( D \\) is the feature dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.464, + 0.598, + 0.785, + 0.614 + ], + "angle": 0, + "content": "\\[\nr = R (x). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Let \\( K_{l} \\in \\mathbb{R}^{M \\times D} \\) be the matrix representation of prompt keys at layer \\( l \\). We then define a similarity matrix \\( A_{query}^{l} \\in \\mathbb{R}^{M \\times 1} \\) to capture the relationship between query \\( q_{l} \\in \\mathbb{R}^{1 \\times D} \\) (from Eq. 6) and prompt keys. Similarly, we compute the similarity matrix \\( A_{ref}^{l} \\) to represent the relationship between the reference prompt query \\( r \\) and the prompt keys." + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.711, + 0.786, + 0.748 + ], + "angle": 0, + "content": "\\[\nA _ {q u e r y} ^ {l} = S o f t m a x (\\frac {K _ {l} q _ {l} ^ {T}}{| | K _ {l} | | _ {2} | | q _ {l} | | _ {2}}) \\qquad A _ {r e f} ^ {l} = S o f t m a x (\\frac {K _ {l} r ^ {T}}{| | K _ {l} | | _ {2} | | r | | _ {2}}). \\qquad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.801 + ], + "angle": 0, + "content": "We apply Softmax() to measure the relative similarity among query-key pairs, considering that \\( q \\) and \\( r \\) show distinct feature distributions originating from different layers. The QR loss penalizes the distance between two similarity matrices." + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.812, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {Q R} = \\sum_ {l} | | A _ {\\text {q u e r y}} ^ {l} - A _ {\\text {r e f}} ^ {l} | | _ {2} ^ {2}. \\tag {10}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.433, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). \\( A_N \\) represents the average accuracy across tasks, and \\( F_N \\) indicates the mean forgetting rate. \\( \\uparrow \\) and \\( \\downarrow \\) indicate whether a metric is better with a higher or lower value, respectively. We average values over five runs with different seeds." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.227, + 0.751, + 0.357 + ], + "angle": 0, + "content": "
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
ER71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
LwF74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
OS-Prompt++77.07 ± 0.152.23 ± 0.1875.67 ± 0.401.27 ± 0.1073.77 ± 0.190.79 ± 0.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.37, + 0.446, + 0.398 + ], + "angle": 0, + "content": "Table 2: Accuracy comparison on CIFAR-100 10-task setting." + }, + { + "type": "table_caption", + "bbox": [ + 0.455, + 0.369, + 0.776, + 0.397 + ], + "angle": 0, + "content": "Table 3: GFLOPs comparison of PCL works. We provide relative cost \\((\\%)\\) with respect to L2P." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.411, + 0.432, + 0.491 + ], + "angle": 0, + "content": "
MethodAN(↑)FN(↓)
UB89.30-
ER76.20 ± 1.048.50 ± 0.37
Deep L2P84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CodaPrompt86.25 ± 0.741.67 ± 0.26
OS-Prompt86.42 ± 0.611.64 ± 0.14
OS-Prompt++86.68 ± 0.671.18 ± 0.21
" + }, + { + "type": "table", + "bbox": [ + 0.474, + 0.411, + 0.754, + 0.492 + ], + "angle": 0, + "content": "
MethodTraining GFLOPsInference GFLOPs
L2P52.8 (100%)35.1 (100%)
Deep L2P52.8 (100%)35.1 (100%)
DualPrompt52.8 (100%)35.1 (100%)
CodaPrompt52.8 (100%)35.1 (100%)
OS-Prompt35.4 (66.7%)17.6 (50.1%)
OS-Prompt++52.8 (100%)17.6 (50.1%)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.785, + 0.547 + ], + "angle": 0, + "content": "The QR loss is added to the cross-entropy loss \\(\\mathcal{L}_{CE}\\) for classification. The total loss function can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.56, + 0.785, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\lambda \\mathcal {L} _ {Q R}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.645 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter to balance between two loss terms. Importantly, the QR loss is applied during the training phase, ensuring that there is no computational overhead during inference. Note that we only train prompts within a prompt pool while freezing the other weight parameters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.667, + 0.377, + 0.684 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.427, + 0.712 + ], + "angle": 0, + "content": "5.1 Experiment Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dataset. We utilize the Split CIFAR-100 [21] and Split ImageNet-R [15] benchmarks for class-incremental continual learning. The Split CIFAR-100 partitions the original CIFAR-100 dataset into 10 distinct tasks, each comprising 10 classes. The Split ImageNet-R benchmark is an adaptation of ImageNet-R, encompassing diverse styles, including cartoon, graffiti, and origami. For our experiments, we segment ImageNet-R into 5, 10, or 20 distinct class groupings. Given its substantial intra-class diversity, the ImageNet-R benchmark is viewed as a particularly challenging benchmark. We also provide experiments on DomainNet [27], a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "large-scale domain incremental learning dataset. The dataset consists of around 0.6 million images distributed across 345 classes. We use 5 different domains (Clipart \\(\\rightarrow\\) Real \\(\\rightarrow\\) Infograph \\(\\rightarrow\\) Sketch \\(\\rightarrow\\) Painting) as a continual learning task where each task consists of 69 classes. The DomainNet results are provided in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Experimental Details. We performed our experiments using the ViT-B/16 [8] pre-trained on ImageNet-1k, a standard backbone in earlier PCL research. To ensure a fair comparison with prior work, we maintain the same prompt length (8) and number of prompt components (100) as used in CodaPrompt. Like CodaPrompt, we divide the total prompt components into the number of tasks, and provide the partial component for each task. In task \\( N \\), the prompt components from task \\( 1 \\sim N \\) are frozen, and we only train the key and prompt components from task \\( N \\). We split \\( 20\\% \\) of the training set for hyperparameter \\( \\lambda \\) tuning in Eq. 11. Our experimental setup is based on the PyTorch. The experiments utilize four Nvidia RTX2080ti GPUs. For robustness, we conducted our benchmarks with five different permutations of task class order, presenting both the average and standard deviation of the results. More detailed information can be found in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.787, + 0.482 + ], + "angle": 0, + "content": "Evaluation Metrics. We use two metrics for evaluation: (1) Average final accuracy \\( A_N \\), which measures the overall accuracy across N tasks. [33,39,40] (2) Average forgetting \\( F_N \\), which tracks local performance drop over N tasks [24,33,34]. Note, \\( A_N \\) is mainly used for performance comparison in literature." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.508, + 0.549, + 0.525 + ], + "angle": 0, + "content": "5.2 Comparison with prior PCL works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.704 + ], + "angle": 0, + "content": "We compare our OS-Prompt with prior continual learning methods. This includes non-PCL methods such as ER [6] and LWF [23], as well as PCL methods like L2P [40], DualPrompt [39], and CodaPrompt [33]. We present both the upper bound (UB) performance and results from fine-tuning (FT). UB refers to training a model with standard supervised training using all data (no continual learning). FT indicates that a model undergoes sequential training with continual learning data without incorporating prompt learning. Moreover, we report L2P/Deep L2P implementation from CodaPrompt [33], both in its original form and improved version by applying prompt pool through layers 1 to 5. In our comparisons, OS-Prompt represents our one-shot prompt framework without the QR loss, while OS-Prompt++ is with the QR loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Table 1 shows the results on ImageNet-R. Our OS-Prompt indicates only a slight performance drop (\\(\\leq 1\\%\\)) across the 5-task, 10-task, and 20-task settings. The reason for the performance drop could be the reduced representational capacity of a prompt query from the intermediate token embedding. However, OS-Prompt++ effectively counters this limitation by incorporating the QR loss. Such an observation underscores the significance of the relationship between the query and the prompt pool in PCL. OS-prompt++ shows slight performance improvement across all scenarios. This implies a prompt query that contains task information helps to enhance representation in the prompt pool, suggesting" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.433, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.146, + 0.514, + 0.174 + ], + "angle": 0, + "content": "Table 4: Relative training computational cost with respect to ER." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.186, + 0.511, + 0.287 + ], + "angle": 0, + "content": "
CL strategyMethodTraining Complexity
Replay-basedER1
Regularization-basedLWF4/3
Prompt-basedL2P1
Prompt-basedDualPrompt1
Prompt-basedCodaPrompt1
Prompt-basedOS-Prompt (Ours)2/3
Prompt-basedOS-Prompt++ (Ours)1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.145, + 0.796, + 0.174 + ], + "angle": 0, + "content": "Table 5: ImageNet-R 10-task results with unsupervised pre-trained model [4]." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.185, + 0.782, + 0.288 + ], + "angle": 0, + "content": "
MethodAN(↑)FN(↓)
ER60.43 ± 1.1613.30 ± 0.12
LwF62.73 ± 1.134.32 ± 0.63
L2P60.32 ± 0.562.30 ± 0.11
DualPrompt61.77 ± 0.612.31 ± 0.23
CodaPrompt67.61 ± 0.192.23 ± 0.29
OS-Prompt67.52 ± 0.342.32 ± 0.13
OS-Prompt++67.92 ± 0.422.19 ± 0.26
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.784, + 0.363 + ], + "angle": 0, + "content": "that exploring methods to integrate task information inside the prompt selection process could be an interesting research direction in PCL. Table 2 provides results on 10-task CIFAR-100, which shows a similar trend with ImageNet-R." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.518, + 0.59 + ], + "angle": 0, + "content": "In Table 3, we provide GFLOPs for various PCL methods, rounding GFLOPs to one decimal place. It is worth noting that while prior PCL methods might have marginally different GFLOP values, they are close enough to appear identical in the table. This implies that different prompt formation schemes proposed in prior works do not substantially impact GFLOPs. The table shows OS-Prompt operates at \\(66.7\\%\\) and \\(50.1\\%\\) of the GFLOPs, during training and inference, respectively, relative to prior methods. While OS-Prompt++ maintains a training GFLOPs count comparable to" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.397, + 0.784, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.533, + 0.788, + 0.561 + ], + "angle": 0, + "content": "Fig. 4: Comparison of latency across PCL methods on different GPUs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.68 + ], + "angle": 0, + "content": "earlier work due to the added feedforward process in the reference ViT, it does not employ the reference ViT during inference, bringing it down to \\(50.1\\%\\) GFLOPs. In Fig. 4, we compare the inference GPU latency of previous PCL methods with our OS-Prompt across three distinct GPU setups. All measurements are taken using a batch size of 32. Similar to the trend observed in FLOPs, our approach reduces the latency by \\(\\sim 50\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.706, + 0.624, + 0.723 + ], + "angle": 0, + "content": "5.3 Discussion on Training Computational Cost" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.794 + ], + "angle": 0, + "content": "In this section, we provide an in-depth discussion of the training computational cost of PCL methods. Following [10], Table 4 presents the relative training complexity of each method with respect to ER [6], which is the simple baseline with standard gradient-based training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.786, + 0.842 + ], + "angle": 0, + "content": "Here, we would like to clarify the training computational cost of prompt learning. In general neural network training, the forward-backward computational cost ratio is approximately 1:2. This is due to gradient backpropagation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.224 + ], + "angle": 0, + "content": "\\(\\left(\\frac{dL}{da_l} = W_{l + 1}\\frac{dL}{da_{l + 1}}\\right)\\) and weight-updating \\(\\left(\\frac{dL}{dW_l} = \\frac{dL}{da_{l + 1}} a_l\\right)\\), where \\(L\\) represents the loss, \\(W\\) denotes the weight parameter, \\(a\\) is the activation, and \\(l\\) is the layer index. In prompt learning, the forward-backward computational cost ratio is approximately 1:1. This is in contrast to general neural network training, as only a small fraction of the weight parameters (less than \\(1\\%\\)) are updated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.795, + 0.416 + ], + "angle": 0, + "content": "Bearing this in mind, we present the observations derived from the table, assuming that all methods employ the same architecture. (1) Previous PCL (L2P, DualPrompt, CodaPrompt) consists of two steps: First, the query ViT requires only feedforward without backpropagation, Second, the backbone ViT requires feedforward-backward training with prompt tuning. This results in the relative training computational cost \\( 1 \\left( = \\frac{\\text{QueryViT}_{fw}(1) + \\text{BackboneViT}_{fw}(1) + \\text{BackboneViT}_{bw}(1)}{\\text{ER}_{fw}(1) + \\text{ER}_{bw}(2)} \\right) \\). Here, \\( fw \\) and \\( bw \\) denote forward and backward propagation respectively. (2) Our OS-Prompt requires one feedforward-backward step like standard training. Therefore, relative training cost becomes \\( \\frac{2}{3} \\left( = \\frac{\\text{BackboneViT}_{fw}(1) + \\text{BackboneViT}_{bw}(1)}{\\text{ER}_{fw}(1) + \\text{ER}_{bw}(2)} \\right) \\). This shows the potential of our OS-Prompt on online continual learning. (3) Our OS-Prompt++ has a reference ViT to perform the regularization, bringing 1 relative training computational cost with respect to ER." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.437, + 0.633, + 0.452 + ], + "angle": 0, + "content": "5.4 Impact of Unsupervised Pre-trained Weights" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.789, + 0.658 + ], + "angle": 0, + "content": "One underlying assumption of existing PCL methods is their reliance on supervised pretraining on ImageNet-1k. While this pre-training has a substantial impact on model performance, it may not always be feasible for a general continual learning task. One possible solution is using unsupervised pre-trained models. To explore this, we use DINO pre-trained weights [4] instead of ImageNet-1k pre-trained weights, and compare the performance across ER, LwF, and PCL works. We set the other experimental settings as identical. We report the results with ImageNet-R 10-task in Table 5. Compared to the ImageNet-1k pre-trained model (with supervised training), performances across all methods degrade. This observation suggests that the backbone model plays a crucial role in PCL. Moreover, similar to the ImageNet-1k pre-trained model, CodaPrompt and our OS-Prompt outperform the other methods by a significant margin. This indicates that our method continues to perform well with various backbones." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.524, + 0.696 + ], + "angle": 0, + "content": "5.5 Analysis of Design Components" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "QR Loss Design. To understand the effect of different components in our proposed QR loss (Eq. 10), we conducted an ablation study with different settings, summarized in Table 6. There are two main components in the QR loss: Cosine Similarity and Softmax, and we provide the accuracy of different four combinations. Without Cosine Similarity and Softmax, our framework yields a performance of \\(75.00\\%\\) for \\(A_{N}\\). Adding Cosine Similarity or Softmax improves the performance, suggesting their collaborative role in enhancing the model's performance in our QR loss design. We also provide the sensitivity analysis of our method with respect to the hyperparameter \\(\\lambda\\) in Table 7. Across the three" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.433, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.492, + 0.173 + ], + "angle": 0, + "content": "Table 6: Analysis of QR loss design on 10-task ImageNet-R setting." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "
CosSim SoftmaxAN(↑)FN(↓)
✓75.00 ± 0.531.68 ± 0.12
75.47 ± 0.421.38 ± 0.16
✓75.51 ± 0.331.28 ± 0.06
✓75.67 ± 0.401.27 ± 0.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.145, + 0.79, + 0.173 + ], + "angle": 0, + "content": "Table 7: Hyperparameter sensitivity study of \\( \\lambda \\) on ImageNet-R 5/10/20-task." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.185, + 0.782, + 0.253 + ], + "angle": 0, + "content": "
λTask-5Task-10Task-20
1e-577.03 ± 0.1075.63 ± 0.3973.63 ± 0.21
5e-577.02 ± 0.1375.62 ± 0.4173.62 ± 0.19
1e-477.07 ± 0.1575.67 ± 0.4073.77 ± 0.19
5e-477.13 ± 0.2475.68 ± 0.3873.68 ± 0.17
" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.27, + 0.391, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.269, + 0.563, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.269, + 0.784, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.382, + 0.584, + 0.424 + ], + "angle": 0, + "content": "Fig. 5: Analysis of accuracy \\( A_N \\) with respect to prompt components (Left) and prompt length (Right). We use 10-taks ImageNet-R setting." + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.382, + 0.787, + 0.423 + ], + "angle": 0, + "content": "Fig.6: Trade-off between GFLOPs and accuracy within a reference ViT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.785, + 0.516 + ], + "angle": 0, + "content": "distinct task configurations on ImageNet-R, we note that the performance fluctuations are minimal, despite the variation in \\(\\lambda\\) values. These results underline the robustness of our method with respect to \\(\\lambda\\) value, suggesting that our approach is not sensitive to hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.8, + 0.622 + ], + "angle": 0, + "content": "Prompt Design. We further examine the impact of both the number of prompt components and the prompt length on accuracy. In Fig. 5 (left), we measure accuracy across configurations with \\(\\{10,20,50,100,200,500\\}\\) prompts. The OS-prompt demonstrates a consistent accuracy enhancement as the number of prompts increases. Notably, OS-prompt++ reaches a performance plateau after just 50 prompts. Our ablation study of prompt length, presented in Fig. 5 (right), reveals that ours maintains stable accuracy across various prompt lengths." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.647, + 0.704, + 0.662 + ], + "angle": 0, + "content": "5.6 Accuracy-Efficiency Trade-off within a Reference ViT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "During the training of OS-prompt++, our method does not improve the energy efficiency (Note, we achieve \\(\\sim 50\\%\\) computational saving during inference). This arises from our approach of extracting the reference prompt query \\(r\\) from the final layer of the reference ViT. To further enhance energy efficiency during the OS-prompt++ training, we delve into the trade-off between accuracy and efficiency within the Reference ViT. Instead of relying on the last layer's [CLS] token embedding for the reference prompt query, we opt for intermediate token embeddings. Since our prompt pool is applied in layers 1 to 5 of the backbone ViT, we utilize the intermediate token embeddings from layers deeper than 5 within the reference ViT. Fig. 6 illustrates this trade-off with respect to the layer index where we get the reference prompt. Our findings show that util" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 8: Analysis on prompt formation methods. Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). \\( A_{N} \\) represents the average accuracy across tasks, and \\( F_{N} \\) indicates the mean forgetting rate. We average values over five runs with different seeds." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.213, + 0.728, + 0.312 + ], + "angle": 0, + "content": "
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
OS-Prompt (L2P)74.61 ± 0.292.29 ± 0.2173.43 ± 0.601.42 ± 0.1071.42 ± 0.240.95 ± 0.06
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
OS-Prompt (Dual)74.65 ± 0.152.06 ± 0.2072.57 ± 0.231.13 ± 0.0370.55 ± 0.480.81 ± 0.10
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt (Coda)75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.784, + 0.369 + ], + "angle": 0, + "content": "layer 8, there is only a slight increase in accuracy, which suggests a potential to reduce GFLOPs without performance drop." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.388, + 0.593, + 0.404 + ], + "angle": 0, + "content": "5.7 Analysis on Prompt Formation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Here, we delve into the impact of varying prompt formation strategies on performance. For comparative insight, we present results from two previous PCL methodologies: L2P and DualPrompt. While L2P selects 5 prompts from the internal prompt pool, DualPrompt distinguishes between general and task-specific prompts. Aside from the distinct prompt formation strategies, the configurations remain consistent. In Table 8, we adopt the prompt formation strategies proposed in L2P and Dual, resulting in configurations denoted as OS-Prompt (L2P) and OS-Prompt (Dual). Our findings reveal the following: (1) The effectiveness of OS-prompt varies based on the prompt formation technique employed. Notably, OS-Prompt (L2P) and OS-Prompt (Dual) yield lower accuracy than our original approach, which relies on CodaPrompt. This suggests that our performance could benefit from refined prompt formation techniques in future iterations. (2) The OS-prompt framework, when integrated with L2P and Dual prompt formation strategies, outperforms the original Deep L2P and DualPrompt, a trend not observed with CodaPrompt. This may indicate that the prompt formation strategies of L2P and DualPrompt, which rely on hard matching (i.e., top-k), are more resilient than CodaPrompt's. Conversely, CodaPrompt employs soft matching, utilizing a weighted summation of all prompts based on proximity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.705, + 0.36, + 0.721 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We introduce the OS-Prompt framework where we improve the efficiency of the conventional two-stage PCL in a simple yet effective manner. One of the potential limitations of our OS-Prompt++ is, in comparison to OS-Prompt (our light version), it introduces an increase in training computational cost from the reference ViT feedforward. This computational efficiency during training becomes important in the context of addressing the online continual learning problem [10] where the model needs rapid training on the streaming data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.433, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.181, + 0.785, + 0.241 + ], + "angle": 0, + "content": "This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), and the DoE MMICC center SEA-CROGS (Award #DE-SC0023198)" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.269, + 0.323, + 0.284 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.332 + ], + "angle": 0, + "content": "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: ECCV (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.36 + ], + "angle": 0, + "content": "2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: NeurIPS (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.403 + ], + "angle": 0, + "content": "3. Cai, H., Gan, C., Zhu, L., Han, S.: Tinytl: Reduce memory, not parameters for efficient on-device learning. Advances in Neural Information Processing Systems 33, 11285-11297 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.404, + 0.785, + 0.457 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "5. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.785, + 0.529 + ], + "angle": 0, + "content": "6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.785, + 0.572 + ], + "angle": 0, + "content": "7. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.613 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 2, 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.656 + ], + "angle": 0, + "content": "9. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.712 + ], + "angle": 0, + "content": "0. Ghunaim, Y., Bibi, A., Alhamoud, K., Alfarra, M., Al Kader Hammoud, H.A., Prabhu, A., Torr, P.H., Ghanem, B.: Real-time evaluation in online continual learning: A new hope. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11888-11897 (2023) 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.741 + ], + "angle": 0, + "content": "1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.783 + ], + "angle": 0, + "content": "12. Harun, M.Y., Gallardo, J., Hayes, T.L., Kanan, C.: How efficient are today's continual learning algorithms? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2430-2435 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.811 + ], + "angle": 0, + "content": "13. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: ICRA (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. He, X., Li, C., Zhang, P., Yang, J., Wang, X.E.: Parameter-efficient fine-tuning for vision transformers. arXiv preprint arXiv:2203.16329 (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.304, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Y. Kim, Y. Li, P. Panda." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "15. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. arXiv preprint arXiv:2006.16241 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "16. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV. pp. 8340-8349 (2021) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "17. Hsu, Y.C., Liu, Y.C., Ramasamy, A., Kira, Z.: Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.284, + 0.787, + 0.324 + ], + "angle": 0, + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "19. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 3, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "22. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: ICML. pp. 3925-3934. PMLR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "23. Li, Z., Hoiem, D.: Learning without forgetting. TPAMI 40(12), 2935-2947 (2017) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.542 + ], + "angle": 0, + "content": "24. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.569 + ], + "angle": 0, + "content": "25. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "26. Pellegrini, L., Lomonaco, V., Graffieti, G., Maltoni, D.: Continual learning at the edge: Real-time training on smartphone devices. arXiv preprint arXiv:2105.13127 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.651 + ], + "angle": 0, + "content": "27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "28. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "30. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.786 + ], + "angle": 0, + "content": "31. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: ICML. pp. 4548-4557 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.839 + ], + "angle": 0, + "content": "32. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.433, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "One-stage Prompt-based Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.218 + ], + "angle": 0, + "content": "33. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) 1, 2, 3, 4, 5, 6, 7, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "34. Smith, J.S., Tian, J., Halbe, S., Hsu, Y.C., Kira, Z.: A closer look at rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2409-2419 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "35. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "36. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.383 + ], + "angle": 0, + "content": "37. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.786, + 0.425 + ], + "angle": 0, + "content": "38. Wang, Z., Zhan, Z., Gong, Y., Yuan, G., Niu, W., Jian, T., Ren, B., Ioannidis, S., Wang, Y., Dy, J.: Sparcl: Sparse continual learning on the edge. Advances in Neural Information Processing Systems 35, 20366-20380 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.48 + ], + "angle": 0, + "content": "39. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022) 1, 2, 4, 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.786, + 0.522 + ], + "angle": 0, + "content": "40. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. CVPR (2022) 1, 2, 4, 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.578 + ], + "angle": 0, + "content": "41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.605 + ], + "angle": 0, + "content": "42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.786, + 0.646 + ], + "angle": 0, + "content": "43. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.786, + 0.674 + ], + "angle": 0, + "content": "44. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.786, + 0.73 + ], + "angle": 0, + "content": "45. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 698-714. Springer (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.771 + ], + "angle": 0, + "content": "46. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.799 + ], + "angle": 0, + "content": "47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.799 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_origin.pdf b/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b52f70a8762f4405e280c67c87214eb2280951f --- /dev/null +++ b/2024/One-stage Prompt-based Continual Learning/e0ea330f-e0aa-49f8-91ec-761371947c04_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4831deb79111004d2d074ced128e1c06090735ee3dae2d516b6338cd27e3488e +size 787437 diff --git a/2024/One-stage Prompt-based Continual Learning/full.md b/2024/One-stage Prompt-based Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..229c820c0d93a82e2e19883ea7231d74e5d6f884 --- /dev/null +++ b/2024/One-stage Prompt-based Continual Learning/full.md @@ -0,0 +1,315 @@ +# One-stage Prompt-based Continual Learning + +Youngeun Kim, Yuhang Li, and Priyadarshini Panda + +Yale University, CT, USA {youngun.kim,yuhang.li, priya.panda}@yale.edu + +Abstract. Prompt-based Continual Learning (PCL) has gained considerable attention as a promising continual learning solution because it achieves state-of-the-art performance while preventing privacy violations and memory overhead problems. Nonetheless, existing PCL approaches face significant computational burdens because of two Vision Transformer (ViT) feed-forward stages; one is for the query ViT that generates a prompt query to select prompts inside a prompt pool; the other one is a backbone ViT that mixes information between selected prompts and image tokens. To address this, we introduce a one-stage PCL framework by directly using the intermediate layer's token embedding as a prompt query. This design removes the need for an additional feed-forward stage for query ViT, resulting in $\sim 50\%$ computational cost reduction for both training and inference with marginal accuracy drop $(\leq 1\%)$ . We further introduce a Query-Pool Regularization (QR) loss that regulates the relationship between the prompt query and the prompt pool to improve representation power. The QR loss is only applied during training time, so there is no computational overhead at inference from the QR loss. With the QR loss, our approach maintains $\sim 50\%$ computational cost reduction during inference as well as outperforms the prior two-stage PCL methods by $\sim 1.4\%$ on public class-incremental continual learning benchmarks including CIFAR-100, ImageNet-R, and DomainNet. + +Keywords: Efficient learning $\cdot$ Continual learning $\cdot$ Transfer learning + +# 1 Introduction + +Training models effectively and efficiently on a continuous stream of data presents a significant practical hurdle. A straightforward approach would entail accumulating both prior and new data and then updating the model using this comprehensive dataset. However, as data volume grows, fully retraining a model on such extensive data becomes increasingly impractical [11, 25]. Additionally, storing past data can raise privacy issues, such as those highlighted by the EU General Data Protection Regulation (GDPR) [36]. An alternative solution is to adapt the model based solely on currently incoming data, eschewing any access to past data. This paradigm is termed as rehearsal-free continual learning [7,9,32,33,39-41], and the primary goal is to diminish the effects of catastrophic forgetting on previously acquired data. + +![](images/ecd3893e7bd2e1ff20b0da9997f2338260a744ee86f04ffdc7c54dc37700af0d.jpg) +Fig. 1: Difference between prior PCL work and ours. Prior PCL work (Left) has two feed-forward stages for (1) a query function (ViT) to select input-specific prompts and (2) a backbone ViT layer to perform prompt learning with the selected prompts and image tokens. On the other hand, our one-stage PCL framework (Middle) uses an intermediate layer's token embedding as a prompt query so that it requires only one backbone ViT feed-forward stage. As a result, our method reduces GFLOPs by $\sim 50\%$ compared to prior work while maintaining accuracy (Right). + +![](images/a9a3e8773bd867ca47ccef4cb8ed9fae8233334a92e9f91b27f57b7ad8e5e504.jpg) + +![](images/0aa1ec17fcf570a5a36dbe5753eae61a8128aee44ab809b81a8499aad7b13ff2.jpg) + +Among the rehearsal-free continual learning methods, Prompt-based Continual Learning (PCL) stands out as it has demonstrated state-of-the-art performance in image classification tasks, even surpassing rehearsal-based methods [33, 37, 39, 40]. PCL utilizes a pre-trained Vision Transformer (ViT) [8] and refines the model by training learnable tokens on the given data. PCL adopts a prompt pool-based training scheme where different prompts are selected and trained for each continual learning stage. This strategy enables a model to learn the information of the training data in a sequential manner with less memory overhead, as the prompt pool requires minimal resources. + +Although PCL methods show state-of-the-art performance, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices [12,26,38]. Specifically, the PCL method requires two-stage ViT feedforward steps. One is for the query function that generates a prompt query. The other one is a backbone ViT that mixes information between selected prompts and input image tokens. We refer to this approach as a two-stage PCL method, illustrated in Fig. 1 Left. + +To address this limitation, we propose a one-stage PCL framework with only one ViT feed-forward stage (Fig. 1 Middle), called OS-Prompt. Rather than deploying a separate ViT feed-forward phase to generate a prompt query, we directly use the intermediate layer's token embedding as a query. This is based on our observation that early layers show marginal shifts in the feature space during continual prompt learning (details are in Section 4.1). This observation enables us to use the intermediate token embedding as a prompt query which requires consistent representation across continual learning to minimize catastrophic forgetting. Surprisingly, OS-Prompt shows a marginal performance drop (less than $1\%$ ) while saving $\sim 50\%$ of computational cost (Fig. 1 Right). + +As our OS-Prompt uses the intermediate layer's token embedding as a query instead of the last layer's token embedding (which is used in prior PCL methods), there is a slight performance drop due to the lack of representation power. To address this, we introduce a Query-Pool Regularization (QR) loss, which is designed to enhance the representation power of a prompt pool by utilizing the token embedding from the last layer. Importantly, the QR loss is applied only during training, ensuring no added computational burden during inference. We refer to our enhanced one-stage framework with QR loss as OS-Prompt++. Our OS-Prompt++ bridges the accuracy gap, performing better than OS-Prompt. + +Overall, our contribution can be summarized as follows: (1) We raise a drawback of the current PCL methods — high computational cost, particularly due to the two distinct ViT feed-forward stages. As a solution to the computational inefficiency in existing PCL techniques, we propose OS-Prompt that reduces the computational cost by nearly $50\%$ without any significant performance drop. (2) To counter the slight performance degradation observed in our one-stage PCL framework, we introduce a QR loss. This ensures that the prompt pool maintains similarity with token embedding from both intermediate and final layers. Notably, our QR loss avoids any additional computational overhead during inference. (3) We conduct experiments with rehearsal-free continual learning setting on CIFAR-100 [21], ImageNet-R [16], and DomainNet [27] benchmarks, outperforming the previous SOTA method CodaPrompt [33] by $\sim 1.4\%$ with $\sim 50\%$ computational cost saving. + +# 2 Related Work + +# 2.1 Continual Learning + +For a decade, continual learning has been explored as an important research topic, concentrating on the progressive adaptation of models over successive tasks or datasets. One of the representative methods to address continual learning is the regularization-based method [1, 23, 44]. By adding a regularization loss between the current model and the previous model, these methods aim to minimize catastrophic forgetting. However, their performance is relatively low on challenging datasets compared to the other continual learning methods. An alternative approach proposes to expand the network architecture as it progresses through different continual learning stages [22, 30, 31, 42]. Though these typically surpass the results of regularization-based methods, they come at the expense of significant memory consumption due to the increased parameters. Rehearsal-based continual learning has introduced a component for archiving previous data [2, 5, 6, 13, 29]. By leveraging accumulated data during the subsequent stages, they often outperform other continual learning methods. However, saving images causes memory overhead and privacy problems. Given the inherent limitations, a growing interest is observed in rehearsal-free continual learning. These are crafted to address catastrophic forgetting without access to past data. Most of them propose methods to generate rehearsal images [7, 9, 32, 41], but generating images is resource-heavy and time-consuming. Recently, within + +the rehearsal-free domain, Prompt-based Continual Learning (PCL) methods have gained considerable attention because of their performance and memory efficiency. By utilizing a pre-trained ViT model, they train only a small set of parameters called prompts to maintain the information against catastrophic forgetting. For example, L2P [40] utilizes a prompt pool, selecting the appropriate prompts for the given image. Extending this principle, DualPrompt [39] proposes general prompts for encompassing knowledge across multiple continual learning stages. Advancing this domain further, [33] facilitates end-to-end training of the prompt pool, achieving state-of-the-art performance. However, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices. Our work aims to resolve such computational complexity overhead in PCL. Furthermore, Tang et al. [35] have proposed a prompt-based continual learning that generates prompts based on the intermediate features from the transformer. However, this prior work focuses on maximizing continual learning performance without a pre-trained model, while ours aims to leverage the power of pre-trained models. Also, the prior work focuses on the prompt generation method, while we propose a framework compatible with different prompt learning techniques. + +# 2.2 Prompt-based Learning + +The efficient fine-tuning method for large pre-trained models has shown their practical benefits across various machine learning tasks [14, 18, 28, 45-47]. Instead of adjusting all parameters within neural architectures, the emphasis has shifted to leveraging a minimal set of weights for optimal transfer outcomes. In alignment with this, multiple methodologies [3, 30] have integrated a streamlined bottleneck component within the transformer framework, thus constraining gradient evaluations to select parameters. Strategies like TinyTL [3] and BitFit [43] advocate for bias modifications in the fine-tuning process. In a recent shift, prompt-based learning [19, 20] captures task-specific knowledge with much smaller additional parameters than the other fine-tuning methods [40]. Also, prompt-based learning only requires storing several prompt tokens, which are easy to plug-and-play, and hence, they are used to construct a prompt pool for recent continual learning methods. + +# 3 Preliminary + +# 3.1 Problem Setting + +In a continual learning setting, a model is trained on $T$ continual learning stages with dataset $D = \{D_{1},D_{2},\dots,D_{T}\}$ , where $D_{t}$ is the data provided in $t$ -th stage. Our problem is based on an image recognition task, so each data $D_{t}$ consists of pairs of images and labels. Also, data from the previous tasks is not accessible for future tasks. Following the previous rehearsal-free continual learning settings [33, 39], we focus on the class-incremental continual learning setting where task + +identity is unknown at inference. This is a challenging scenario compared to others such as task-incremental continual learning where the task labels are provided for both training and test phases [17]. For the experiments, we split classes into $T$ chunks with continual learning benchmarks including CIFAR-100 [21] and ImageNet-R [16]. + +# 3.2 Two-stage Prompt-based Continual Learning + +In our framework, we focus on improving efficiency through a new query selection rather than changing the way prompts are formed from the prompt pool with the given query, a common focus in earlier PCL studies [33, 39, 40]. For a fair comparison with earlier work, we follow the overall PCL framework established in previous studies. Given that our contribution complements the prompt formation technique, our method can seamlessly work with future works that propose a stronger prompt generation method. + +The PCL framework selects the input-aware prompts from the layer-wise prompt pool and adds them to the backbone ViT to consider task knowledge. The $l$ -th layer has the prompt pool $P_{l} = \{k_{l}^{1}:p_{l}^{1},\dots,k_{l}^{M}:p_{l}^{M}\}$ which contains $M$ prompt components $p\in \mathbb{R}^{L_p\times D}$ and the corresponding key $k\in \mathbb{R}^D$ . Here, $L_{p}$ and $D$ stand for the prompt length and the feature dimension, respectively. + +The prior PCL method consists of two feed-forward stages. In the first stage, a prompt query $q \in \mathbb{R}^D$ is extracted from pre-trained query ViT $Q(\cdot)$ , utilizing the [CLS] token from the final layer. + +$$ +q = Q (x) _ {[ C L S ]}. \tag {1} +$$ + +Here $x$ is the given RGB image input. The extracted prompt query is used to form a prompt $\phi_{l} \in \mathbb{R}^{L_{p} \times D}$ from a prompt pool $P_{l}$ , which can be formulated as: + +$$ +\phi_ {l} = g (q, P _ {l}). \tag {2} +$$ + +The prompt formation function $g(\cdot)$ has been a major contribution in the previous literature. L2P [40] and DualPrompt [39] select prompts having top-N similarity (e.g. cosine similarity) between query $q$ and prompt key $k_{l}$ . The recent state-of-the-art CodaPrompt [33] conducts a weighted summation of prompt components based on their similarity to enable end-to-end training. + +The obtained prompt $\phi_l$ for layer $l$ is given to backbone ViT layer $f_l$ with input token embedding $x_l$ . This is the second feed-forward stage of the prior PCL methods. + +$$ +x _ {l + 1} = f _ {l} \left(x _ {l}, \phi_ {l}\right). \tag {3} +$$ + +Following DualPrompt [39] and CodaPrompt [33], we use prefix-tuning to add the information of a prompt $\phi_l$ inside $f_{l}$ . The prefix-tuning splits prompt $\phi_l$ into $[\phi_k,\phi_v]\in \mathbb{R}^{\frac{L_p}{2}\times D}$ , and then prepends them to the key and value inside the self-attention block of ViT. We utilize Multi-Head Self-Attention (MHSA) [8] + +![](images/3b6423ad60c4e9b747562e696bc5bf64cb4b7541c18d2e274f7bdab2ae65f4e6.jpg) +Fig. 2: We measure layer-wise feature distances between token embeddings of the model after training on task 1 and when a new task is learned. Each column represents the embeddings after training a model on Tasks 3, 5, 7, and 9. For instance, the left figure represents the distance between the token embeddings of Task 1 and those learned when Task 3 is completed. We train prompts using CodaPrompt [33] and use $1 - \text{CosSim}(x, y)$ to measure the layer-wise distance on the training dataset. We use a CIFAR100 10-task setting. We provide more examples in the Supplementary Materials. + +like the prior PCL work, which computes the outputs from multiple single-head attention blocks. + +$$ +M H S A \left(x, \phi_ {k}, \phi_ {v}\right) = C o n c a t [ h e a d _ {1}, \dots , h e a d _ {H} ] W _ {o}. \tag {4} +$$ + +$$ +\operatorname {h e a d} _ {i} = \operatorname {A t t e n t i o n} \left(x W _ {q} ^ {i}, \left[ \phi_ {k}; x \right] W _ {k} ^ {i}, \left[ \phi_ {v}; x \right] W _ {v} ^ {i}\right). \tag {5} +$$ + +Here, $W_{o}, W_{q}, W_{k}, W_{v}$ are the projection matrices. This prefix-tuning method is applied to the first 5 layers of backbone ViT. The leftover layers conduct a standard MHSA without prompts. + +However, the prior PCL approach requires two-stage ViT feedforward (Eq. 1 and Eq. 3), which doubles computational cost. In our work, we aim to improve the computational efficiency of the PCL framework without degrading the performance. + +# 4 One-stage Prompt-based Continual Learning + +We propose to reduce the computational cost of PCL by restructuring two-step feedforward stages. To this end, we propose a new one-stage PCL framework called OS-Prompt. Instead of using a separate feedforward step to compute the prompt query $q$ from Eq. 1, we take a token embedding from the intermediate layer of a backbone ViT as the query (illustrated in Fig. 1). + +# 4.1 How Stable are Token Embeddings across Continual Learning? + +We first ensure the validity of using intermediate token embedding as a prompt query. The original two-stage design employs a frozen pre-trained ViT, ensuring consistent prompt query representation throughout continual learning. It is essential to maintain a consistent (or similar) prompt query representation because + +changes in prompt query would bring catastrophic forgetting. In our approach, token embeddings in the backbone ViT continually change as prompt tokens are updated during learning. For instance, after training on tasks 1 and 2, prompts have different values, resulting in varying token embeddings for an identical image. Consequently, it is crucial to assess if token embeddings sustain a consistent representation throughout the continual learning. + +To address the concern, we validate the deviation of the token embedding as prompt continual learning progresses. In Fig. 2, we show how this change happens at each layer during continuous learning tasks. From our study, two main observations can be highlighted: (1) When we add prompts, there is a larger difference in the deeper layers. For example, layers $1 \sim 5$ have a small difference ( $\leq 0.1$ ). However, the last layer shows a bigger change ( $\geq 0.1$ ). (2) As learning continues, the earlier layers remain relatively stable, but the deeper layers change more. This observation concludes that although prompt tokens are included, the token embedding of early layers shows minor changes during training. Therefore, using token embeddings from these earlier layers would give a stable and consistent representation throughout the continual learning stages. + +# 4.2 One-stage PCL Framework + +Building on these observations, we employ the token embedding from the early layers as a prompt query to generate layer-wise prompt tokens. For a fair comparison, we implement prompts across layers 1 to 5, in line with prior work. The proposed OS-Prompt framework is illustrated in Fig. 3. For each layer, given the input token embedding, we directly use the [CLS] token as the prompt query. The original query selection equation (Eq. 1) can be reformulated as: + +$$ +q _ {l} = x _ {l _ {[ C L S ]}}. \tag {6} +$$ + +Using the provided query $q_{l}$ , we generate a prompt from the prompt pool following the state-of-the-art CodaPrompt [33]. It is worth highlighting that our primary contribution lies not in introducing a new prompt generation technique but in presenting a more efficient framework for query selection. We first measure the cosine similarity $\gamma(\cdot)$ between $q_{l}$ and keys $\{k_{l}^{1}, \dots, k_{l}^{M}\}$ , and then we perform a weighted summation of the corresponding value $p_{l}^{m}$ (i.e. prompt) based on the similarity. + +$$ +\phi_ {l} = \sum_ {m} \gamma \left(q _ {l}, k _ {l} ^ {m}\right) p _ {l} ^ {m}. \tag {7} +$$ + +The generated prompt is then prepended to the image tokens. Notably, since we produce a prompt query without the need for an extra ViT feedforward, the computational overhead is reduced by approximately $50\%$ for both training and inference. + +# 4.3 Query-Pool Regularization Loss + +Our OS-Prompt relies on intermediate token embeddings. As a result, the prompt query exhibits diminished representational capacity compared to previous PCL + +![](images/777cd9e5a205306d527d4f9741d79c11c7e253c0101e074c379fc524c21faa31.jpg) +Fig. 3: Our OS-Prompt (OS-Prompt++) framework. An image passes through the backbone ViT layers to get the final prediction. From layer 1 to layer 5, we preprocess prompt tokens to the image tokens, which can be obtained in the following progress: ① We first compute the image token embedding from the previous layer. ② We use [CLS] token as a prompt query used for measuring cosine similarity between prompt keys inside the prompt pool. ③ Based on the similarity, we do weighted sum values to obtain prompt tokens. ④ To further improve the accuracy, we present OS-Prompt++. We integrate the query-pool regularization loss (dotted line), enabling the prompt pool to capture a stronger representation from the reference ViT. + +approaches that utilize the [CLS] token from the final layer. This reduced capacity for representation brings performance degradation. To mitigate this, we introduce a Query-Pool Regularization (QR) loss. The QR loss ensures that the query-pool relationship becomes similar to that of the final layer's [CLS] token, thereby improving representation power. We extract the [CLS] token from the last layer of a reference ViT architecture $R(\cdot)$ and employ it as our reference prompt query $r \in \mathbb{R}^{1 \times D}$ where $D$ is the feature dimension: + +$$ +r = R (x). \tag {8} +$$ + +Let $K_{l} \in \mathbb{R}^{M \times D}$ be the matrix representation of prompt keys at layer $l$ . We then define a similarity matrix $A_{query}^{l} \in \mathbb{R}^{M \times 1}$ to capture the relationship between query $q_{l} \in \mathbb{R}^{1 \times D}$ (from Eq. 6) and prompt keys. Similarly, we compute the similarity matrix $A_{ref}^{l}$ to represent the relationship between the reference prompt query $r$ and the prompt keys. + +$$ +A _ {q u e r y} ^ {l} = S o f t m a x (\frac {K _ {l} q _ {l} ^ {T}}{| | K _ {l} | | _ {2} | | q _ {l} | | _ {2}}) \qquad A _ {r e f} ^ {l} = S o f t m a x (\frac {K _ {l} r ^ {T}}{| | K _ {l} | | _ {2} | | r | | _ {2}}). \qquad (9) +$$ + +We apply Softmax() to measure the relative similarity among query-key pairs, considering that $q$ and $r$ show distinct feature distributions originating from different layers. The QR loss penalizes the distance between two similarity matrices. + +$$ +\mathcal {L} _ {Q R} = \sum_ {l} | | A _ {\text {q u e r y}} ^ {l} - A _ {\text {r e f}} ^ {l} | | _ {2} ^ {2}. \tag {10} +$$ + +Table 1: Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). $A_N$ represents the average accuracy across tasks, and $F_N$ indicates the mean forgetting rate. $\uparrow$ and $\downarrow$ indicate whether a metric is better with a higher or lower value, respectively. We average values over five runs with different seeds. + +
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
ER71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
LwF74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
OS-Prompt++77.07 ± 0.152.23 ± 0.1875.67 ± 0.401.27 ± 0.1073.77 ± 0.190.79 ± 0.07
+ +Table 2: Accuracy comparison on CIFAR-100 10-task setting. + +
MethodAN(↑)FN(↓)
UB89.30-
ER76.20 ± 1.048.50 ± 0.37
Deep L2P84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CodaPrompt86.25 ± 0.741.67 ± 0.26
OS-Prompt86.42 ± 0.611.64 ± 0.14
OS-Prompt++86.68 ± 0.671.18 ± 0.21
+ +Table 3: GFLOPs comparison of PCL works. We provide relative cost $(\%)$ with respect to L2P. + +
MethodTraining GFLOPsInference GFLOPs
L2P52.8 (100%)35.1 (100%)
Deep L2P52.8 (100%)35.1 (100%)
DualPrompt52.8 (100%)35.1 (100%)
CodaPrompt52.8 (100%)35.1 (100%)
OS-Prompt35.4 (66.7%)17.6 (50.1%)
OS-Prompt++52.8 (100%)17.6 (50.1%)
+ +The QR loss is added to the cross-entropy loss $\mathcal{L}_{CE}$ for classification. The total loss function can be written as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {C E} + \lambda \mathcal {L} _ {Q R}, \tag {11} +$$ + +where $\lambda$ is a hyperparameter to balance between two loss terms. Importantly, the QR loss is applied during the training phase, ensuring that there is no computational overhead during inference. Note that we only train prompts within a prompt pool while freezing the other weight parameters. + +# 5 Experiments + +# 5.1 Experiment Setting + +Dataset. We utilize the Split CIFAR-100 [21] and Split ImageNet-R [15] benchmarks for class-incremental continual learning. The Split CIFAR-100 partitions the original CIFAR-100 dataset into 10 distinct tasks, each comprising 10 classes. The Split ImageNet-R benchmark is an adaptation of ImageNet-R, encompassing diverse styles, including cartoon, graffiti, and origami. For our experiments, we segment ImageNet-R into 5, 10, or 20 distinct class groupings. Given its substantial intra-class diversity, the ImageNet-R benchmark is viewed as a particularly challenging benchmark. We also provide experiments on DomainNet [27], a + +large-scale domain incremental learning dataset. The dataset consists of around 0.6 million images distributed across 345 classes. We use 5 different domains (Clipart $\rightarrow$ Real $\rightarrow$ Infograph $\rightarrow$ Sketch $\rightarrow$ Painting) as a continual learning task where each task consists of 69 classes. The DomainNet results are provided in the Supplementary Materials. + +Experimental Details. We performed our experiments using the ViT-B/16 [8] pre-trained on ImageNet-1k, a standard backbone in earlier PCL research. To ensure a fair comparison with prior work, we maintain the same prompt length (8) and number of prompt components (100) as used in CodaPrompt. Like CodaPrompt, we divide the total prompt components into the number of tasks, and provide the partial component for each task. In task $N$ , the prompt components from task $1 \sim N$ are frozen, and we only train the key and prompt components from task $N$ . We split $20\%$ of the training set for hyperparameter $\lambda$ tuning in Eq. 11. Our experimental setup is based on the PyTorch. The experiments utilize four Nvidia RTX2080ti GPUs. For robustness, we conducted our benchmarks with five different permutations of task class order, presenting both the average and standard deviation of the results. More detailed information can be found in the Supplementary Materials. + +Evaluation Metrics. We use two metrics for evaluation: (1) Average final accuracy $A_N$ , which measures the overall accuracy across N tasks. [33,39,40] (2) Average forgetting $F_N$ , which tracks local performance drop over N tasks [24,33,34]. Note, $A_N$ is mainly used for performance comparison in literature. + +# 5.2 Comparison with prior PCL works + +We compare our OS-Prompt with prior continual learning methods. This includes non-PCL methods such as ER [6] and LWF [23], as well as PCL methods like L2P [40], DualPrompt [39], and CodaPrompt [33]. We present both the upper bound (UB) performance and results from fine-tuning (FT). UB refers to training a model with standard supervised training using all data (no continual learning). FT indicates that a model undergoes sequential training with continual learning data without incorporating prompt learning. Moreover, we report L2P/Deep L2P implementation from CodaPrompt [33], both in its original form and improved version by applying prompt pool through layers 1 to 5. In our comparisons, OS-Prompt represents our one-shot prompt framework without the QR loss, while OS-Prompt++ is with the QR loss. + +Table 1 shows the results on ImageNet-R. Our OS-Prompt indicates only a slight performance drop ( $\leq 1\%$ ) across the 5-task, 10-task, and 20-task settings. The reason for the performance drop could be the reduced representational capacity of a prompt query from the intermediate token embedding. However, OS-Prompt++ effectively counters this limitation by incorporating the QR loss. Such an observation underscores the significance of the relationship between the query and the prompt pool in PCL. OS-prompt++ shows slight performance improvement across all scenarios. This implies a prompt query that contains task information helps to enhance representation in the prompt pool, suggesting + +Table 4: Relative training computational cost with respect to ER. + +
CL strategyMethodTraining Complexity
Replay-basedER1
Regularization-basedLWF4/3
Prompt-basedL2P1
Prompt-basedDualPrompt1
Prompt-basedCodaPrompt1
Prompt-basedOS-Prompt (Ours)2/3
Prompt-basedOS-Prompt++ (Ours)1
+ +Table 5: ImageNet-R 10-task results with unsupervised pre-trained model [4]. + +
MethodAN(↑)FN(↓)
ER60.43 ± 1.1613.30 ± 0.12
LwF62.73 ± 1.134.32 ± 0.63
L2P60.32 ± 0.562.30 ± 0.11
DualPrompt61.77 ± 0.612.31 ± 0.23
CodaPrompt67.61 ± 0.192.23 ± 0.29
OS-Prompt67.52 ± 0.342.32 ± 0.13
OS-Prompt++67.92 ± 0.422.19 ± 0.26
+ +that exploring methods to integrate task information inside the prompt selection process could be an interesting research direction in PCL. Table 2 provides results on 10-task CIFAR-100, which shows a similar trend with ImageNet-R. + +In Table 3, we provide GFLOPs for various PCL methods, rounding GFLOPs to one decimal place. It is worth noting that while prior PCL methods might have marginally different GFLOP values, they are close enough to appear identical in the table. This implies that different prompt formation schemes proposed in prior works do not substantially impact GFLOPs. The table shows OS-Prompt operates at $66.7\%$ and $50.1\%$ of the GFLOPs, during training and inference, respectively, relative to prior methods. While OS-Prompt++ maintains a training GFLOPs count comparable to + +![](images/f627728d5724d88c29b20fb157107406b148f3ddbe26e47fa8b11df005fbd2e2.jpg) +Fig. 4: Comparison of latency across PCL methods on different GPUs. + +earlier work due to the added feedforward process in the reference ViT, it does not employ the reference ViT during inference, bringing it down to $50.1\%$ GFLOPs. In Fig. 4, we compare the inference GPU latency of previous PCL methods with our OS-Prompt across three distinct GPU setups. All measurements are taken using a batch size of 32. Similar to the trend observed in FLOPs, our approach reduces the latency by $\sim 50\%$ . + +# 5.3 Discussion on Training Computational Cost + +In this section, we provide an in-depth discussion of the training computational cost of PCL methods. Following [10], Table 4 presents the relative training complexity of each method with respect to ER [6], which is the simple baseline with standard gradient-based training. + +Here, we would like to clarify the training computational cost of prompt learning. In general neural network training, the forward-backward computational cost ratio is approximately 1:2. This is due to gradient backpropagation + +$\left(\frac{dL}{da_l} = W_{l + 1}\frac{dL}{da_{l + 1}}\right)$ and weight-updating $\left(\frac{dL}{dW_l} = \frac{dL}{da_{l + 1}} a_l\right)$ , where $L$ represents the loss, $W$ denotes the weight parameter, $a$ is the activation, and $l$ is the layer index. In prompt learning, the forward-backward computational cost ratio is approximately 1:1. This is in contrast to general neural network training, as only a small fraction of the weight parameters (less than $1\%$ ) are updated. + +Bearing this in mind, we present the observations derived from the table, assuming that all methods employ the same architecture. (1) Previous PCL (L2P, DualPrompt, CodaPrompt) consists of two steps: First, the query ViT requires only feedforward without backpropagation, Second, the backbone ViT requires feedforward-backward training with prompt tuning. This results in the relative training computational cost $1 \left( = \frac{\text{QueryViT}_{fw}(1) + \text{BackboneViT}_{fw}(1) + \text{BackboneViT}_{bw}(1)}{\text{ER}_{fw}(1) + \text{ER}_{bw}(2)} \right)$ . Here, $fw$ and $bw$ denote forward and backward propagation respectively. (2) Our OS-Prompt requires one feedforward-backward step like standard training. Therefore, relative training cost becomes $\frac{2}{3} \left( = \frac{\text{BackboneViT}_{fw}(1) + \text{BackboneViT}_{bw}(1)}{\text{ER}_{fw}(1) + \text{ER}_{bw}(2)} \right)$ . This shows the potential of our OS-Prompt on online continual learning. (3) Our OS-Prompt++ has a reference ViT to perform the regularization, bringing 1 relative training computational cost with respect to ER. + +# 5.4 Impact of Unsupervised Pre-trained Weights + +One underlying assumption of existing PCL methods is their reliance on supervised pretraining on ImageNet-1k. While this pre-training has a substantial impact on model performance, it may not always be feasible for a general continual learning task. One possible solution is using unsupervised pre-trained models. To explore this, we use DINO pre-trained weights [4] instead of ImageNet-1k pre-trained weights, and compare the performance across ER, LwF, and PCL works. We set the other experimental settings as identical. We report the results with ImageNet-R 10-task in Table 5. Compared to the ImageNet-1k pre-trained model (with supervised training), performances across all methods degrade. This observation suggests that the backbone model plays a crucial role in PCL. Moreover, similar to the ImageNet-1k pre-trained model, CodaPrompt and our OS-Prompt outperform the other methods by a significant margin. This indicates that our method continues to perform well with various backbones. + +# 5.5 Analysis of Design Components + +QR Loss Design. To understand the effect of different components in our proposed QR loss (Eq. 10), we conducted an ablation study with different settings, summarized in Table 6. There are two main components in the QR loss: Cosine Similarity and Softmax, and we provide the accuracy of different four combinations. Without Cosine Similarity and Softmax, our framework yields a performance of $75.00\%$ for $A_{N}$ . Adding Cosine Similarity or Softmax improves the performance, suggesting their collaborative role in enhancing the model's performance in our QR loss design. We also provide the sensitivity analysis of our method with respect to the hyperparameter $\lambda$ in Table 7. Across the three + +Table 6: Analysis of QR loss design on 10-task ImageNet-R setting. + +
CosSim SoftmaxAN(↑)FN(↓)
✓75.00 ± 0.531.68 ± 0.12
75.47 ± 0.421.38 ± 0.16
✓75.51 ± 0.331.28 ± 0.06
✓75.67 ± 0.401.27 ± 0.10
+ +Table 7: Hyperparameter sensitivity study of $\lambda$ on ImageNet-R 5/10/20-task. + +
λTask-5Task-10Task-20
1e-577.03 ± 0.1075.63 ± 0.3973.63 ± 0.21
5e-577.02 ± 0.1375.62 ± 0.4173.62 ± 0.19
1e-477.07 ± 0.1575.67 ± 0.4073.77 ± 0.19
5e-477.13 ± 0.2475.68 ± 0.3873.68 ± 0.17
+ +![](images/b2c63dc9d14f81e3ae7306c59a02bccfe8844649d1e5342720f1f1b238492e30.jpg) +Fig. 5: Analysis of accuracy $A_N$ with respect to prompt components (Left) and prompt length (Right). We use 10-taks ImageNet-R setting. + +![](images/f23e8a4243edee892875192a0ffc089383e3bd5b9e8628339e53d492b9bae222.jpg) +Fig.6: Trade-off between GFLOPs and accuracy within a reference ViT. + +![](images/af6f6432b03f12338a895483944c790af25cdaa85d480daa8485efd18395809a.jpg) + +distinct task configurations on ImageNet-R, we note that the performance fluctuations are minimal, despite the variation in $\lambda$ values. These results underline the robustness of our method with respect to $\lambda$ value, suggesting that our approach is not sensitive to hyperparameter. + +Prompt Design. We further examine the impact of both the number of prompt components and the prompt length on accuracy. In Fig. 5 (left), we measure accuracy across configurations with $\{10,20,50,100,200,500\}$ prompts. The OS-prompt demonstrates a consistent accuracy enhancement as the number of prompts increases. Notably, OS-prompt++ reaches a performance plateau after just 50 prompts. Our ablation study of prompt length, presented in Fig. 5 (right), reveals that ours maintains stable accuracy across various prompt lengths. + +# 5.6 Accuracy-Efficiency Trade-off within a Reference ViT + +During the training of OS-prompt++, our method does not improve the energy efficiency (Note, we achieve $\sim 50\%$ computational saving during inference). This arises from our approach of extracting the reference prompt query $r$ from the final layer of the reference ViT. To further enhance energy efficiency during the OS-prompt++ training, we delve into the trade-off between accuracy and efficiency within the Reference ViT. Instead of relying on the last layer's [CLS] token embedding for the reference prompt query, we opt for intermediate token embeddings. Since our prompt pool is applied in layers 1 to 5 of the backbone ViT, we utilize the intermediate token embeddings from layers deeper than 5 within the reference ViT. Fig. 6 illustrates this trade-off with respect to the layer index where we get the reference prompt. Our findings show that util + +Table 8: Analysis on prompt formation methods. Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). $A_{N}$ represents the average accuracy across tasks, and $F_{N}$ indicates the mean forgetting rate. We average values over five runs with different seeds. + +
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
OS-Prompt (L2P)74.61 ± 0.292.29 ± 0.2173.43 ± 0.601.42 ± 0.1071.42 ± 0.240.95 ± 0.06
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
OS-Prompt (Dual)74.65 ± 0.152.06 ± 0.2072.57 ± 0.231.13 ± 0.0370.55 ± 0.480.81 ± 0.10
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt (Coda)75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
+ +layer 8, there is only a slight increase in accuracy, which suggests a potential to reduce GFLOPs without performance drop. + +# 5.7 Analysis on Prompt Formation Strategy + +Here, we delve into the impact of varying prompt formation strategies on performance. For comparative insight, we present results from two previous PCL methodologies: L2P and DualPrompt. While L2P selects 5 prompts from the internal prompt pool, DualPrompt distinguishes between general and task-specific prompts. Aside from the distinct prompt formation strategies, the configurations remain consistent. In Table 8, we adopt the prompt formation strategies proposed in L2P and Dual, resulting in configurations denoted as OS-Prompt (L2P) and OS-Prompt (Dual). Our findings reveal the following: (1) The effectiveness of OS-prompt varies based on the prompt formation technique employed. Notably, OS-Prompt (L2P) and OS-Prompt (Dual) yield lower accuracy than our original approach, which relies on CodaPrompt. This suggests that our performance could benefit from refined prompt formation techniques in future iterations. (2) The OS-prompt framework, when integrated with L2P and Dual prompt formation strategies, outperforms the original Deep L2P and DualPrompt, a trend not observed with CodaPrompt. This may indicate that the prompt formation strategies of L2P and DualPrompt, which rely on hard matching (i.e., top-k), are more resilient than CodaPrompt's. Conversely, CodaPrompt employs soft matching, utilizing a weighted summation of all prompts based on proximity. + +# 6 Conclusion + +We introduce the OS-Prompt framework where we improve the efficiency of the conventional two-stage PCL in a simple yet effective manner. One of the potential limitations of our OS-Prompt++ is, in comparison to OS-Prompt (our light version), it introduces an increase in training computational cost from the reference ViT feedforward. This computational efficiency during training becomes important in the context of addressing the online continual learning problem [10] where the model needs rapid training on the streaming data. + +# Acknowledgment + +This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), and the DoE MMICC center SEA-CROGS (Award #DE-SC0023198) + +# References + +1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: ECCV (2018) 3 +2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: NeurIPS (2020) 3 +3. Cai, H., Gan, C., Zhu, L., Han, S.: Tinytl: Reduce memory, not parameters for efficient on-device learning. Advances in Neural Information Processing Systems 33, 11285-11297 (2020) 4 +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 11, 12 +5. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018) 3 +6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 3, 10, 11 +7. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 1, 3 +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 2, 5, 10 +9. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 1, 3 +0. Ghunaim, Y., Bibi, A., Alhamoud, K., Alfarra, M., Al Kader Hammoud, H.A., Prabhu, A., Torr, P.H., Ghanem, B.: Real-time evaluation in online continual learning: A new hope. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11888-11897 (2023) 11, 14 +1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020) 1 +12. Harun, M.Y., Gallardo, J., Hayes, T.L., Kanan, C.: How efficient are today's continual learning algorithms? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2430-2435 (2023) 2 +13. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: ICRA (2019) 3 +14. He, X., Li, C., Zhang, P., Yang, J., Wang, X.E.: Parameter-efficient fine-tuning for vision transformers. arXiv preprint arXiv:2203.16329 (2022) 4 + +15. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. arXiv preprint arXiv:2006.16241 (2020) 9 +16. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV. pp. 8340-8349 (2021) 3, 5 +17. Hsu, Y.C., Liu, Y.C., Ramasamy, A., Kira, Z.: Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488 (2018) 5 +18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4 +19. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022) 4 +20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4 +21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 3, 5, 9 +22. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: ICML. pp. 3925-3934. PMLR (2019) 3 +23. Li, Z., Hoiem, D.: Learning without forgetting. TPAMI 40(12), 2935-2947 (2017) 3, 10 +24. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS (2017) 10 +25. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022) 1 +26. Pellegrini, L., Lomonaco, V., Graffieti, G., Maltoni, D.: Continual learning at the edge: Real-time training on smartphone devices. arXiv preprint arXiv:2105.13127 (2021) 2 +27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 3, 9 +28. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018) 4 +29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 3 +30. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) 3, 4 +31. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: ICML. pp. 4548-4557 (2018) 3 +32. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 1, 3 + +33. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) 1, 2, 3, 4, 5, 6, 7, 10 +34. Smith, J.S., Tian, J., Halbe, S., Hsu, Y.C., Kira, Z.: A closer look at rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2409-2419 (2023) 10 +35. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023) 4 +36. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017) 1 +37. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 2 +38. Wang, Z., Zhan, Z., Gong, Y., Yuan, G., Niu, W., Jian, T., Ren, B., Ioannidis, S., Wang, Y., Dy, J.: Sparcl: Sparse continual learning on the edge. Advances in Neural Information Processing Systems 35, 20366-20380 (2022) 2 +39. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022) 1, 2, 4, 5, 10 +40. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. CVPR (2022) 1, 2, 4, 5, 10 +41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020) 1, 3 +42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017) 3 +43. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) 4 +44. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017) 3 +45. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 698-714. Springer (2020) 4 +46. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021) 4 +47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4 \ No newline at end of file diff --git a/2024/One-stage Prompt-based Continual Learning/images.zip b/2024/One-stage Prompt-based Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc031543ed8c4cef2c5e7226a19e97288f167f97 --- /dev/null +++ b/2024/One-stage Prompt-based Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd24757a865b27d3fa977f5283c302a1c50fb3a38f4356ff96692feb325589d9 +size 475037 diff --git a/2024/One-stage Prompt-based Continual Learning/layout.json b/2024/One-stage Prompt-based Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0576c59158e0e49717fe926ff1883657cb293fc5 --- /dev/null +++ b/2024/One-stage Prompt-based Continual Learning/layout.json @@ -0,0 +1,9399 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 112, + 462, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 112, + 462, + 129 + ], + "spans": [ + { + "bbox": [ + 151, + 112, + 462, + 129 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 150, + 435, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 150, + 435, + 163 + ], + "spans": [ + { + "bbox": [ + 179, + 150, + 435, + 163 + ], + "type": "text", + "content": "Youngeun Kim, Yuhang Li, and Priyadarshini Panda" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 200, + 172, + 414, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 172, + 414, + 195 + ], + "spans": [ + { + "bbox": [ + 200, + 172, + 414, + 195 + ], + "type": "text", + "content": "Yale University, CT, USA {youngun.kim,yuhang.li, priya.panda}@yale.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "spans": [ + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "text", + "content": "Abstract. Prompt-based Continual Learning (PCL) has gained considerable attention as a promising continual learning solution because it achieves state-of-the-art performance while preventing privacy violations and memory overhead problems. Nonetheless, existing PCL approaches face significant computational burdens because of two Vision Transformer (ViT) feed-forward stages; one is for the query ViT that generates a prompt query to select prompts inside a prompt pool; the other one is a backbone ViT that mixes information between selected prompts and image tokens. To address this, we introduce a one-stage PCL framework by directly using the intermediate layer's token embedding as a prompt query. This design removes the need for an additional feed-forward stage for query ViT, resulting in " + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "text", + "content": " computational cost reduction for both training and inference with marginal accuracy drop " + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "inline_equation", + "content": "(\\leq 1\\%)" + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "text", + "content": ". We further introduce a Query-Pool Regularization (QR) loss that regulates the relationship between the prompt query and the prompt pool to improve representation power. The QR loss is only applied during training time, so there is no computational overhead at inference from the QR loss. With the QR loss, our approach maintains " + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "text", + "content": " computational cost reduction during inference as well as outperforms the prior two-stage PCL methods by " + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "inline_equation", + "content": "\\sim 1.4\\%" + }, + { + "bbox": [ + 160, + 224, + 452, + 464 + ], + "type": "text", + "content": " on public class-incremental continual learning benchmarks including CIFAR-100, ImageNet-R, and DomainNet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 476, + 447, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 476, + 447, + 488 + ], + "spans": [ + { + "bbox": [ + 160, + 476, + 447, + 488 + ], + "type": "text", + "content": "Keywords: Efficient learning " + }, + { + "bbox": [ + 160, + 476, + 447, + 488 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 476, + 447, + 488 + ], + "type": "text", + "content": " Continual learning " + }, + { + "bbox": [ + 160, + 476, + 447, + 488 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 476, + 447, + 488 + ], + "type": "text", + "content": " Transfer learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 509, + 230, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 509, + 230, + 521 + ], + "spans": [ + { + "bbox": [ + 133, + 509, + 230, + 521 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Training models effectively and efficiently on a continuous stream of data presents a significant practical hurdle. A straightforward approach would entail accumulating both prior and new data and then updating the model using this comprehensive dataset. However, as data volume grows, fully retraining a model on such extensive data becomes increasingly impractical [11, 25]. Additionally, storing past data can raise privacy issues, such as those highlighted by the EU General Data Protection Regulation (GDPR) [36]. An alternative solution is to adapt the model based solely on currently incoming data, eschewing any access to past data. This paradigm is termed as rehearsal-free continual learning [7,9,32,33,39-41], and the primary goal is to diminish the effects of catastrophic forgetting on previously acquired data." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 253, + 228 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 253, + 228 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 253, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 253, + 228 + ], + "type": "image", + "image_path": "ecd3893e7bd2e1ff20b0da9997f2338260a744ee86f04ffdc7c54dc37700af0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 249, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 327 + ], + "type": "text", + "content": "Fig. 1: Difference between prior PCL work and ours. Prior PCL work (Left) has two feed-forward stages for (1) a query function (ViT) to select input-specific prompts and (2) a backbone ViT layer to perform prompt learning with the selected prompts and image tokens. On the other hand, our one-stage PCL framework (Middle) uses an intermediate layer's token embedding as a prompt query so that it requires only one backbone ViT feed-forward stage. As a result, our method reduces GFLOPs by " + }, + { + "bbox": [ + 130, + 249, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 130, + 249, + 482, + 327 + ], + "type": "text", + "content": " compared to prior work while maintaining accuracy (Right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 370, + 227 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 370, + 227 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 370, + 227 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 370, + 227 + ], + "type": "image", + "image_path": "a9a3e8773bd867ca47ccef4cb8ed9fae8233334a92e9f91b27f57b7ad8e5e504.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 373, + 118, + 478, + 227 + ], + "blocks": [ + { + "bbox": [ + 373, + 118, + 478, + 227 + ], + "lines": [ + { + "bbox": [ + 373, + 118, + 478, + 227 + ], + "spans": [ + { + "bbox": [ + 373, + 118, + 478, + 227 + ], + "type": "image", + "image_path": "0aa1ec17fcf570a5a36dbe5753eae61a8128aee44ab809b81a8499aad7b13ff2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": "Among the rehearsal-free continual learning methods, Prompt-based Continual Learning (PCL) stands out as it has demonstrated state-of-the-art performance in image classification tasks, even surpassing rehearsal-based methods [33, 37, 39, 40]. PCL utilizes a pre-trained Vision Transformer (ViT) [8] and refines the model by training learnable tokens on the given data. PCL adopts a prompt pool-based training scheme where different prompts are selected and trained for each continual learning stage. This strategy enables a model to learn the information of the training data in a sequential manner with less memory overhead, as the prompt pool requires minimal resources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "content": "Although PCL methods show state-of-the-art performance, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices [12,26,38]. Specifically, the PCL method requires two-stage ViT feedforward steps. One is for the query function that generates a prompt query. The other one is a backbone ViT that mixes information between selected prompts and input image tokens. We refer to this approach as a two-stage PCL method, illustrated in Fig. 1 Left." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "To address this limitation, we propose a one-stage PCL framework with only one ViT feed-forward stage (Fig. 1 Middle), called OS-Prompt. Rather than deploying a separate ViT feed-forward phase to generate a prompt query, we directly use the intermediate layer's token embedding as a query. This is based on our observation that early layers show marginal shifts in the feature space during continual prompt learning (details are in Section 4.1). This observation enables us to use the intermediate token embedding as a prompt query which requires consistent representation across continual learning to minimize catastrophic forgetting. Surprisingly, OS-Prompt shows a marginal performance drop (less than " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ") while saving " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " of computational cost (Fig. 1 Right)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "type": "text", + "content": "As our OS-Prompt uses the intermediate layer's token embedding as a query instead of the last layer's token embedding (which is used in prior PCL methods), there is a slight performance drop due to the lack of representation power. To address this, we introduce a Query-Pool Regularization (QR) loss, which is designed to enhance the representation power of a prompt pool by utilizing the token embedding from the last layer. Importantly, the QR loss is applied only during training, ensuring no added computational burden during inference. We refer to our enhanced one-stage framework with QR loss as OS-Prompt++. Our OS-Prompt++ bridges the accuracy gap, performing better than OS-Prompt." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "text", + "content": "Overall, our contribution can be summarized as follows: (1) We raise a drawback of the current PCL methods — high computational cost, particularly due to the two distinct ViT feed-forward stages. As a solution to the computational inefficiency in existing PCL techniques, we propose OS-Prompt that reduces the computational cost by nearly " + }, + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "text", + "content": " without any significant performance drop. (2) To counter the slight performance degradation observed in our one-stage PCL framework, we introduce a QR loss. This ensures that the prompt pool maintains similarity with token embedding from both intermediate and final layers. Notably, our QR loss avoids any additional computational overhead during inference. (3) We conduct experiments with rehearsal-free continual learning setting on CIFAR-100 [21], ImageNet-R [16], and DomainNet [27] benchmarks, outperforming the previous SOTA method CodaPrompt [33] by " + }, + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "inline_equation", + "content": "\\sim 1.4\\%" + }, + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 130, + 224, + 480, + 379 + ], + "type": "text", + "content": " computational cost saving." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 396, + 237, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 237, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 237, + 408 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 420, + 258, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 258, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 258, + 432 + ], + "type": "text", + "content": "2.1 Continual Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "For a decade, continual learning has been explored as an important research topic, concentrating on the progressive adaptation of models over successive tasks or datasets. One of the representative methods to address continual learning is the regularization-based method [1, 23, 44]. By adding a regularization loss between the current model and the previous model, these methods aim to minimize catastrophic forgetting. However, their performance is relatively low on challenging datasets compared to the other continual learning methods. An alternative approach proposes to expand the network architecture as it progresses through different continual learning stages [22, 30, 31, 42]. Though these typically surpass the results of regularization-based methods, they come at the expense of significant memory consumption due to the increased parameters. Rehearsal-based continual learning has introduced a component for archiving previous data [2, 5, 6, 13, 29]. By leveraging accumulated data during the subsequent stages, they often outperform other continual learning methods. However, saving images causes memory overhead and privacy problems. Given the inherent limitations, a growing interest is observed in rehearsal-free continual learning. These are crafted to address catastrophic forgetting without access to past data. Most of them propose methods to generate rehearsal images [7, 9, 32, 41], but generating images is resource-heavy and time-consuming. Recently, within" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 265, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 265, + 91, + 447, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 333 + ], + "type": "text", + "content": "the rehearsal-free domain, Prompt-based Continual Learning (PCL) methods have gained considerable attention because of their performance and memory efficiency. By utilizing a pre-trained ViT model, they train only a small set of parameters called prompts to maintain the information against catastrophic forgetting. For example, L2P [40] utilizes a prompt pool, selecting the appropriate prompts for the given image. Extending this principle, DualPrompt [39] proposes general prompts for encompassing knowledge across multiple continual learning stages. Advancing this domain further, [33] facilitates end-to-end training of the prompt pool, achieving state-of-the-art performance. However, huge computational costs from the two ViT feed-forward stages make the model difficult to deploy into resource-constrained devices. Our work aims to resolve such computational complexity overhead in PCL. Furthermore, Tang et al. [35] have proposed a prompt-based continual learning that generates prompts based on the intermediate features from the transformer. However, this prior work focuses on maximizing continual learning performance without a pre-trained model, while ours aims to leverage the power of pre-trained models. Also, the prior work focuses on the prompt generation method, while we propose a framework compatible with different prompt learning techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 350, + 280, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 280, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 280, + 363 + ], + "type": "text", + "content": "2.2 Prompt-based Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 371, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 529 + ], + "type": "text", + "content": "The efficient fine-tuning method for large pre-trained models has shown their practical benefits across various machine learning tasks [14, 18, 28, 45-47]. Instead of adjusting all parameters within neural architectures, the emphasis has shifted to leveraging a minimal set of weights for optimal transfer outcomes. In alignment with this, multiple methodologies [3, 30] have integrated a streamlined bottleneck component within the transformer framework, thus constraining gradient evaluations to select parameters. Strategies like TinyTL [3] and BitFit [43] advocate for bias modifications in the fine-tuning process. In a recent shift, prompt-based learning [19, 20] captures task-specific knowledge with much smaller additional parameters than the other fine-tuning methods [40]. Also, prompt-based learning only requires storing several prompt tokens, which are easy to plug-and-play, and hence, they are used to construct a prompt pool for recent continual learning methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 546, + 227, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 227, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 227, + 560 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 572, + 246, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 246, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 246, + 586 + ], + "type": "text", + "content": "3.1 Problem Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In a continual learning setting, a model is trained on " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " continual learning stages with dataset " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "D = \\{D_{1},D_{2},\\dots,D_{T}\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the data provided in " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "-th stage. Our problem is based on an image recognition task, so each data " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " consists of pairs of images and labels. Also, data from the previous tasks is not accessible for future tasks. Following the previous rehearsal-free continual learning settings [33, 39], we focus on the class-incremental continual learning setting where task" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "text", + "content": "identity is unknown at inference. This is a challenging scenario compared to others such as task-incremental continual learning where the task labels are provided for both training and test phases [17]. For the experiments, we split classes into " + }, + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "text", + "content": " chunks with continual learning benchmarks including CIFAR-100 [21] and ImageNet-R [16]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 198, + 386, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 386, + 210 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 386, + 210 + ], + "type": "text", + "content": "3.2 Two-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 222, + 480, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 480, + 305 + ], + "type": "text", + "content": "In our framework, we focus on improving efficiency through a new query selection rather than changing the way prompts are formed from the prompt pool with the given query, a common focus in earlier PCL studies [33, 39, 40]. For a fair comparison with earlier work, we follow the overall PCL framework established in previous studies. Given that our contribution complements the prompt formation technique, our method can seamlessly work with future works that propose a stronger prompt generation method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": "The PCL framework selects the input-aware prompts from the layer-wise prompt pool and adds them to the backbone ViT to consider task knowledge. The " + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": "-th layer has the prompt pool " + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "inline_equation", + "content": "P_{l} = \\{k_{l}^{1}:p_{l}^{1},\\dots,k_{l}^{M}:p_{l}^{M}\\}" + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": " which contains " + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": " prompt components " + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "inline_equation", + "content": "p\\in \\mathbb{R}^{L_p\\times D}" + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": " and the corresponding key " + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "inline_equation", + "content": "k\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 306, + 480, + 366 + ], + "type": "text", + "content": " stand for the prompt length and the feature dimension, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 366, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 366, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 480, + 403 + ], + "type": "text", + "content": "The prior PCL method consists of two feed-forward stages. In the first stage, a prompt query " + }, + { + "bbox": [ + 131, + 366, + 480, + 403 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 131, + 366, + 480, + 403 + ], + "type": "text", + "content": " is extracted from pre-trained query ViT " + }, + { + "bbox": [ + 131, + 366, + 480, + 403 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 131, + 366, + 480, + 403 + ], + "type": "text", + "content": ", utilizing the [CLS] token from the final layer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 274, + 415, + 480, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 415, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 274, + 415, + 480, + 428 + ], + "type": "interline_equation", + "content": "q = Q (x) _ {[ C L S ]}. \\tag {1}", + "image_path": "c79bbbe10d2073d62c39c0cfa1b2e7ad90df5ce70adec0f9f0acd52ba3e56bed.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "text", + "content": " is the given RGB image input. The extracted prompt query is used to form a prompt " + }, + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "inline_equation", + "content": "\\phi_{l} \\in \\mathbb{R}^{L_{p} \\times D}" + }, + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "text", + "content": " from a prompt pool " + }, + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 131, + 437, + 479, + 462 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 279, + 477, + 480, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 477, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 279, + 477, + 480, + 488 + ], + "type": "interline_equation", + "content": "\\phi_ {l} = g (q, P _ {l}). \\tag {2}", + "image_path": "b26ebf50b0d1d0e3bf751530553368733e0089eb6d5a1ce73541681f11232abc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "text", + "content": "The prompt formation function " + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "text", + "content": " has been a major contribution in the previous literature. L2P [40] and DualPrompt [39] select prompts having top-N similarity (e.g. cosine similarity) between query " + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "text", + "content": " and prompt key " + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "text", + "content": ". The recent state-of-the-art CodaPrompt [33] conducts a weighted summation of prompt components based on their similarity to enable end-to-end training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "text", + "content": "The obtained prompt " + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "text", + "content": " for layer " + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "text", + "content": " is given to backbone ViT layer " + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "inline_equation", + "content": "f_l" + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "text", + "content": " with input token embedding " + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "inline_equation", + "content": "x_l" + }, + { + "bbox": [ + 130, + 556, + 479, + 591 + ], + "type": "text", + "content": ". This is the second feed-forward stage of the prior PCL methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 270, + 596, + 480, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 596, + 480, + 607 + ], + "spans": [ + { + "bbox": [ + 270, + 596, + 480, + 607 + ], + "type": "interline_equation", + "content": "x _ {l + 1} = f _ {l} \\left(x _ {l}, \\phi_ {l}\\right). \\tag {3}", + "image_path": "33cc796c7a6e201ca65a595cd4200c9a28d23e2d9becb70189f7d144dabde73d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "text", + "content": "Following DualPrompt [39] and CodaPrompt [33], we use prefix-tuning to add the information of a prompt " + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "text", + "content": " inside " + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "text", + "content": ". The prefix-tuning splits prompt " + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "inline_equation", + "content": "[\\phi_k,\\phi_v]\\in \\mathbb{R}^{\\frac{L_p}{2}\\times D}" + }, + { + "bbox": [ + 130, + 615, + 480, + 666 + ], + "type": "text", + "content": ", and then prepends them to the key and value inside the self-attention block of ViT. We utilize Multi-Head Self-Attention (MHSA) [8]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 190 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 190 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 190 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 190 + ], + "type": "image", + "image_path": "3b6423ad60c4e9b747562e696bc5bf64cb4b7541c18d2e274f7bdab2ae65f4e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 210, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 131, + 210, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 482, + 288 + ], + "type": "text", + "content": "Fig. 2: We measure layer-wise feature distances between token embeddings of the model after training on task 1 and when a new task is learned. Each column represents the embeddings after training a model on Tasks 3, 5, 7, and 9. For instance, the left figure represents the distance between the token embeddings of Task 1 and those learned when Task 3 is completed. We train prompts using CodaPrompt [33] and use " + }, + { + "bbox": [ + 131, + 210, + 482, + 288 + ], + "type": "inline_equation", + "content": "1 - \\text{CosSim}(x, y)" + }, + { + "bbox": [ + 131, + 210, + 482, + 288 + ], + "type": "text", + "content": " to measure the layer-wise distance on the training dataset. We use a CIFAR100 10-task setting. We provide more examples in the Supplementary Materials." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 311, + 481, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 481, + 335 + ], + "type": "text", + "content": "like the prior PCL work, which computes the outputs from multiple single-head attention blocks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 196, + 343, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 343, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 196, + 343, + 481, + 357 + ], + "type": "interline_equation", + "content": "M H S A \\left(x, \\phi_ {k}, \\phi_ {v}\\right) = C o n c a t [ h e a d _ {1}, \\dots , h e a d _ {H} ] W _ {o}. \\tag {4}", + "image_path": "bd04f225d83c9b4ce25e66f84377044c2b3373f7e29189379fbcd1dc3bfc9d23.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 203, + 376, + 481, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 376, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 203, + 376, + 481, + 390 + ], + "type": "interline_equation", + "content": "\\operatorname {h e a d} _ {i} = \\operatorname {A t t e n t i o n} \\left(x W _ {q} ^ {i}, \\left[ \\phi_ {k}; x \\right] W _ {k} ^ {i}, \\left[ \\phi_ {v}; x \\right] W _ {v} ^ {i}\\right). \\tag {5}", + "image_path": "f9cdd5871e947605df7bb697a6c3515c0a7a9ced80dd0eca6b49858cdabbeb7f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 394, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 480, + 430 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 394, + 480, + 430 + ], + "type": "inline_equation", + "content": "W_{o}, W_{q}, W_{k}, W_{v}" + }, + { + "bbox": [ + 130, + 394, + 480, + 430 + ], + "type": "text", + "content": " are the projection matrices. This prefix-tuning method is applied to the first 5 layers of backbone ViT. The leftover layers conduct a standard MHSA without prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "type": "text", + "content": "However, the prior PCL approach requires two-stage ViT feedforward (Eq. 1 and Eq. 3), which doubles computational cost. In our work, we aim to improve the computational efficiency of the PCL framework without degrading the performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 496, + 419, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 419, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 419, + 511 + ], + "type": "text", + "content": "4 One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "We propose to reduce the computational cost of PCL by restructuring two-step feedforward stages. To this end, we propose a new one-stage PCL framework called OS-Prompt. Instead of using a separate feedforward step to compute the prompt query " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " from Eq. 1, we take a token embedding from the intermediate layer of a backbone ViT as the query (illustrated in Fig. 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 597, + 477, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 477, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 477, + 611 + ], + "type": "text", + "content": "4.1 How Stable are Token Embeddings across Continual Learning?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "We first ensure the validity of using intermediate token embedding as a prompt query. The original two-stage design employs a frozen pre-trained ViT, ensuring consistent prompt query representation throughout continual learning. It is essential to maintain a consistent (or similar) prompt query representation because" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "changes in prompt query would bring catastrophic forgetting. In our approach, token embeddings in the backbone ViT continually change as prompt tokens are updated during learning. For instance, after training on tasks 1 and 2, prompts have different values, resulting in varying token embeddings for an identical image. Consequently, it is crucial to assess if token embeddings sustain a consistent representation throughout the continual learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "text", + "content": "To address the concern, we validate the deviation of the token embedding as prompt continual learning progresses. In Fig. 2, we show how this change happens at each layer during continuous learning tasks. From our study, two main observations can be highlighted: (1) When we add prompts, there is a larger difference in the deeper layers. For example, layers " + }, + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "inline_equation", + "content": "1 \\sim 5" + }, + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "text", + "content": " have a small difference (" + }, + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\leq 0.1" + }, + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "text", + "content": "). However, the last layer shows a bigger change (" + }, + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\geq 0.1" + }, + { + "bbox": [ + 130, + 188, + 482, + 319 + ], + "type": "text", + "content": "). (2) As learning continues, the earlier layers remain relatively stable, but the deeper layers change more. This observation concludes that although prompt tokens are included, the token embedding of early layers shows minor changes during training. Therefore, using token embeddings from these earlier layers would give a stable and consistent representation throughout the continual learning stages." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 335, + 297, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 297, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 297, + 346 + ], + "type": "text", + "content": "4.2 One-stage PCL Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 479, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 479, + 425 + ], + "type": "text", + "content": "Building on these observations, we employ the token embedding from the early layers as a prompt query to generate layer-wise prompt tokens. For a fair comparison, we implement prompts across layers 1 to 5, in line with prior work. The proposed OS-Prompt framework is illustrated in Fig. 3. For each layer, given the input token embedding, we directly use the [CLS] token as the prompt query. The original query selection equation (Eq. 1) can be reformulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 281, + 432, + 480, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 432, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 281, + 432, + 480, + 445 + ], + "type": "interline_equation", + "content": "q _ {l} = x _ {l _ {[ C L S ]}}. \\tag {6}", + "image_path": "e237f9ea960e0733df4e04dc4df9f5d2fbf22574ced49bd122b6aee4977d2572.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": "Using the provided query " + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "inline_equation", + "content": "q_{l}" + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": ", we generate a prompt from the prompt pool following the state-of-the-art CodaPrompt [33]. It is worth highlighting that our primary contribution lies not in introducing a new prompt generation technique but in presenting a more efficient framework for query selection. We first measure the cosine similarity " + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\gamma(\\cdot)" + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "inline_equation", + "content": "q_{l}" + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\{k_{l}^{1}, \\dots, k_{l}^{M}\\}" + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": ", and then we perform a weighted summation of the corresponding value " + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "inline_equation", + "content": "p_{l}^{m}" + }, + { + "bbox": [ + 130, + 450, + 482, + 533 + ], + "type": "text", + "content": " (i.e. prompt) based on the similarity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 261, + 534, + 480, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 534, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 261, + 534, + 480, + 557 + ], + "type": "interline_equation", + "content": "\\phi_ {l} = \\sum_ {m} \\gamma \\left(q _ {l}, k _ {l} ^ {m}\\right) p _ {l} ^ {m}. \\tag {7}", + "image_path": "8d2e889020464eabd3a008f3039ed54dcc8842abe071469da54fa390ea039f95.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 560, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 479, + 608 + ], + "type": "text", + "content": "The generated prompt is then prepended to the image tokens. Notably, since we produce a prompt query without the need for an extra ViT feedforward, the computational overhead is reduced by approximately " + }, + { + "bbox": [ + 130, + 560, + 479, + 608 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 560, + 479, + 608 + ], + "type": "text", + "content": " for both training and inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 320, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 320, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 320, + 635 + ], + "type": "text", + "content": "4.3 Query-Pool Regularization Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Our OS-Prompt relies on intermediate token embeddings. As a result, the prompt query exhibits diminished representational capacity compared to previous PCL" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 113, + 445, + 238 + ], + "blocks": [ + { + "bbox": [ + 166, + 113, + 445, + 238 + ], + "lines": [ + { + "bbox": [ + 166, + 113, + 445, + 238 + ], + "spans": [ + { + "bbox": [ + 166, + 113, + 445, + 238 + ], + "type": "image", + "image_path": "777cd9e5a205306d527d4f9741d79c11c7e253c0101e074c379fc524c21faa31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 357 + ], + "type": "text", + "content": "Fig. 3: Our OS-Prompt (OS-Prompt++) framework. An image passes through the backbone ViT layers to get the final prediction. From layer 1 to layer 5, we preprocess prompt tokens to the image tokens, which can be obtained in the following progress: ① We first compute the image token embedding from the previous layer. ② We use [CLS] token as a prompt query used for measuring cosine similarity between prompt keys inside the prompt pool. ③ Based on the similarity, we do weighted sum values to obtain prompt tokens. ④ To further improve the accuracy, we present OS-Prompt++. We integrate the query-pool regularization loss (dotted line), enabling the prompt pool to capture a stronger representation from the reference ViT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "content": "approaches that utilize the [CLS] token from the final layer. This reduced capacity for representation brings performance degradation. To mitigate this, we introduce a Query-Pool Regularization (QR) loss. The QR loss ensures that the query-pool relationship becomes similar to that of the final layer's [CLS] token, thereby improving representation power. We extract the [CLS] token from the last layer of a reference ViT architecture " + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "content": " and employ it as our reference prompt query " + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "content": " is the feature dimension:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 283, + 473, + 480, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 473, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 283, + 473, + 480, + 486 + ], + "type": "interline_equation", + "content": "r = R (x). \\tag {8}", + "image_path": "1e759209a98092735a0957f85d0e2a99b4896f62e687ae088f26ad2d4abadfd0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "inline_equation", + "content": "K_{l} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "content": " be the matrix representation of prompt keys at layer " + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "content": ". We then define a similarity matrix " + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "inline_equation", + "content": "A_{query}^{l} \\in \\mathbb{R}^{M \\times 1}" + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "content": " to capture the relationship between query " + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "inline_equation", + "content": "q_{l} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "content": " (from Eq. 6) and prompt keys. Similarly, we compute the similarity matrix " + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "inline_equation", + "content": "A_{ref}^{l}" + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "content": " to represent the relationship between the reference prompt query " + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 494, + 482, + 556 + ], + "type": "text", + "content": " and the prompt keys." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 563, + 481, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 563, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 149, + 563, + 481, + 592 + ], + "type": "interline_equation", + "content": "A _ {q u e r y} ^ {l} = S o f t m a x (\\frac {K _ {l} q _ {l} ^ {T}}{| | K _ {l} | | _ {2} | | q _ {l} | | _ {2}}) \\qquad A _ {r e f} ^ {l} = S o f t m a x (\\frac {K _ {l} r ^ {T}}{| | K _ {l} | | _ {2} | | r | | _ {2}}). \\qquad (9)", + "image_path": "3abd3fe9de24ef3056219eab1f457eda4f6601032eb8cd97a277602d825e00d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 597, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 634 + ], + "type": "text", + "content": "We apply Softmax() to measure the relative similarity among query-key pairs, considering that " + }, + { + "bbox": [ + 130, + 597, + 482, + 634 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 597, + 482, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 597, + 482, + 634 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 597, + 482, + 634 + ], + "type": "text", + "content": " show distinct feature distributions originating from different layers. The QR loss penalizes the distance between two similarity matrices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 240, + 643, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 643, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 240, + 643, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {Q R} = \\sum_ {l} | | A _ {\\text {q u e r y}} ^ {l} - A _ {\\text {r e f}} ^ {l} | | _ {2} ^ {2}. \\tag {10}", + "image_path": "d4955240cc52f8c3c401aa0a37c9e6014a73d4554eff4993cbf87239905c6505.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 179, + 459, + 282 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " represents the average accuracy across tasks, and " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " indicates the mean forgetting rate. " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " indicate whether a metric is better with a higher or lower value, respectively. We average values over five runs with different seeds." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 179, + 459, + 282 + ], + "lines": [ + { + "bbox": [ + 151, + 179, + 459, + 282 + ], + "spans": [ + { + "bbox": [ + 151, + 179, + 459, + 282 + ], + "type": "table", + "html": "
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
ER71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
LwF74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
OS-Prompt++77.07 ± 0.152.23 ± 0.1875.67 ± 0.401.27 ± 0.1073.77 ± 0.190.79 ± 0.07
", + "image_path": "1cfc3dab425fa3b188320e2a78c413f9a07bd17107a49fe4673981afcca09640.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 141, + 325, + 264, + 388 + ], + "blocks": [ + { + "bbox": [ + 132, + 293, + 272, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 272, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 272, + 315 + ], + "type": "text", + "content": "Table 2: Accuracy comparison on CIFAR-100 10-task setting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 325, + 264, + 388 + ], + "lines": [ + { + "bbox": [ + 141, + 325, + 264, + 388 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 264, + 388 + ], + "type": "table", + "html": "
MethodAN(↑)FN(↓)
UB89.30-
ER76.20 ± 1.048.50 ± 0.37
Deep L2P84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CodaPrompt86.25 ± 0.741.67 ± 0.26
OS-Prompt86.42 ± 0.611.64 ± 0.14
OS-Prompt++86.68 ± 0.671.18 ± 0.21
", + "image_path": "6bdc05d6c285b58533fa2cdbabbfbc50e2de3cec8747153dfdeb4d0b1e3b0c3b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 290, + 325, + 461, + 389 + ], + "blocks": [ + { + "bbox": [ + 278, + 292, + 474, + 314 + ], + "lines": [ + { + "bbox": [ + 278, + 292, + 474, + 314 + ], + "spans": [ + { + "bbox": [ + 278, + 292, + 474, + 314 + ], + "type": "text", + "content": "Table 3: GFLOPs comparison of PCL works. We provide relative cost " + }, + { + "bbox": [ + 278, + 292, + 474, + 314 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 278, + 292, + 474, + 314 + ], + "type": "text", + "content": " with respect to L2P." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 325, + 461, + 389 + ], + "lines": [ + { + "bbox": [ + 290, + 325, + 461, + 389 + ], + "spans": [ + { + "bbox": [ + 290, + 325, + 461, + 389 + ], + "type": "table", + "html": "
MethodTraining GFLOPsInference GFLOPs
L2P52.8 (100%)35.1 (100%)
Deep L2P52.8 (100%)35.1 (100%)
DualPrompt52.8 (100%)35.1 (100%)
CodaPrompt52.8 (100%)35.1 (100%)
OS-Prompt35.4 (66.7%)17.6 (50.1%)
OS-Prompt++52.8 (100%)17.6 (50.1%)
", + "image_path": "0cd7c954be4d43b10e1fd4fd3281ed6bd08366fae230ecee7483f26fa005dd5a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 433 + ], + "type": "text", + "content": "The QR loss is added to the cross-entropy loss " + }, + { + "bbox": [ + 130, + 411, + 480, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 130, + 411, + 480, + 433 + ], + "type": "text", + "content": " for classification. The total loss function can be written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 255, + 443, + 480, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 443, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 255, + 443, + 480, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\lambda \\mathcal {L} _ {Q R}, \\tag {11}", + "image_path": "6a36e3d14f9781740143e44e4ee5243cd966be727713d582952ef0cacb322028.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 463, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 463, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 463, + 482, + 510 + ], + "type": "text", + "content": " is a hyperparameter to balance between two loss terms. Importantly, the QR loss is applied during the training phase, ensuring that there is no computational overhead during inference. Note that we only train prompts within a prompt pool while freezing the other weight parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 528, + 230, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 230, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 230, + 541 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 551, + 261, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 261, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 261, + 563 + ], + "type": "text", + "content": "5.1 Experiment Setting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Dataset. We utilize the Split CIFAR-100 [21] and Split ImageNet-R [15] benchmarks for class-incremental continual learning. The Split CIFAR-100 partitions the original CIFAR-100 dataset into 10 distinct tasks, each comprising 10 classes. The Split ImageNet-R benchmark is an adaptation of ImageNet-R, encompassing diverse styles, including cartoon, graffiti, and origami. For our experiments, we segment ImageNet-R into 5, 10, or 20 distinct class groupings. Given its substantial intra-class diversity, the ImageNet-R benchmark is viewed as a particularly challenging benchmark. We also provide experiments on DomainNet [27], a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "large-scale domain incremental learning dataset. The dataset consists of around 0.6 million images distributed across 345 classes. We use 5 different domains (Clipart " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " Real " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " Infograph " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " Sketch " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " Painting) as a continual learning task where each task consists of 69 classes. The DomainNet results are provided in the Supplementary Materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "content": "Experimental Details. We performed our experiments using the ViT-B/16 [8] pre-trained on ImageNet-1k, a standard backbone in earlier PCL research. To ensure a fair comparison with prior work, we maintain the same prompt length (8) and number of prompt components (100) as used in CodaPrompt. Like CodaPrompt, we divide the total prompt components into the number of tasks, and provide the partial component for each task. In task " + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "content": ", the prompt components from task " + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "inline_equation", + "content": "1 \\sim N" + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "content": " are frozen, and we only train the key and prompt components from task " + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "content": ". We split " + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "content": " of the training set for hyperparameter " + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "content": " tuning in Eq. 11. Our experimental setup is based on the PyTorch. The experiments utilize four Nvidia RTX2080ti GPUs. For robustness, we conducted our benchmarks with five different permutations of task class order, presenting both the average and standard deviation of the results. More detailed information can be found in the Supplementary Materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "text", + "content": "Evaluation Metrics. We use two metrics for evaluation: (1) Average final accuracy " + }, + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "text", + "content": ", which measures the overall accuracy across N tasks. [33,39,40] (2) Average forgetting " + }, + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "text", + "content": ", which tracks local performance drop over N tasks [24,33,34]. Note, " + }, + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 333, + 481, + 381 + ], + "type": "text", + "content": " is mainly used for performance comparison in literature." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 402, + 335, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 335, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 335, + 415 + ], + "type": "text", + "content": "5.2 Comparison with prior PCL works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "content": "We compare our OS-Prompt with prior continual learning methods. This includes non-PCL methods such as ER [6] and LWF [23], as well as PCL methods like L2P [40], DualPrompt [39], and CodaPrompt [33]. We present both the upper bound (UB) performance and results from fine-tuning (FT). UB refers to training a model with standard supervised training using all data (no continual learning). FT indicates that a model undergoes sequential training with continual learning data without incorporating prompt learning. Moreover, we report L2P/Deep L2P implementation from CodaPrompt [33], both in its original form and improved version by applying prompt pool through layers 1 to 5. In our comparisons, OS-Prompt represents our one-shot prompt framework without the QR loss, while OS-Prompt++ is with the QR loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Table 1 shows the results on ImageNet-R. Our OS-Prompt indicates only a slight performance drop (" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\leq 1\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ") across the 5-task, 10-task, and 20-task settings. The reason for the performance drop could be the reduced representational capacity of a prompt query from the intermediate token embedding. However, OS-Prompt++ effectively counters this limitation by incorporating the QR loss. Such an observation underscores the significance of the relationship between the query and the prompt pool in PCL. OS-prompt++ shows slight performance improvement across all scenarios. This implies a prompt query that contains task information helps to enhance representation in the prompt pool, suggesting" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 147, + 312, + 227 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 314, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 314, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 314, + 137 + ], + "type": "text", + "content": "Table 4: Relative training computational cost with respect to ER." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 147, + 312, + 227 + ], + "lines": [ + { + "bbox": [ + 135, + 147, + 312, + 227 + ], + "spans": [ + { + "bbox": [ + 135, + 147, + 312, + 227 + ], + "type": "table", + "html": "
CL strategyMethodTraining Complexity
Replay-basedER1
Regularization-basedLWF4/3
Prompt-basedL2P1
Prompt-basedDualPrompt1
Prompt-basedCodaPrompt1
Prompt-basedOS-Prompt (Ours)2/3
Prompt-basedOS-Prompt++ (Ours)1
", + "image_path": "fddc9dad5e0a61ec60a1f8d45d7ebc571fcea1a26f6df5215cbd20d0d25fed33.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 146, + 478, + 228 + ], + "blocks": [ + { + "bbox": [ + 317, + 114, + 487, + 137 + ], + "lines": [ + { + "bbox": [ + 317, + 114, + 487, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 487, + 137 + ], + "type": "text", + "content": "Table 5: ImageNet-R 10-task results with unsupervised pre-trained model [4]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 146, + 478, + 228 + ], + "lines": [ + { + "bbox": [ + 328, + 146, + 478, + 228 + ], + "spans": [ + { + "bbox": [ + 328, + 146, + 478, + 228 + ], + "type": "table", + "html": "
MethodAN(↑)FN(↓)
ER60.43 ± 1.1613.30 ± 0.12
LwF62.73 ± 1.134.32 ± 0.63
L2P60.32 ± 0.562.30 ± 0.11
DualPrompt61.77 ± 0.612.31 ± 0.23
CodaPrompt67.61 ± 0.192.23 ± 0.29
OS-Prompt67.52 ± 0.342.32 ± 0.13
OS-Prompt++67.92 ± 0.422.19 ± 0.26
", + "image_path": "a840d9a843763fb090de418c53bda3586283cc9af370cb99ac936164b6c814cb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 479, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 287 + ], + "type": "text", + "content": "that exploring methods to integrate task information inside the prompt selection process could be an interesting research direction in PCL. Table 2 provides results on 10-task CIFAR-100, which shows a similar trend with ImageNet-R." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 288, + 317, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 317, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 317, + 467 + ], + "type": "text", + "content": "In Table 3, we provide GFLOPs for various PCL methods, rounding GFLOPs to one decimal place. It is worth noting that while prior PCL methods might have marginally different GFLOP values, they are close enough to appear identical in the table. This implies that different prompt formation schemes proposed in prior works do not substantially impact GFLOPs. The table shows OS-Prompt operates at " + }, + { + "bbox": [ + 130, + 288, + 317, + 467 + ], + "type": "inline_equation", + "content": "66.7\\%" + }, + { + "bbox": [ + 130, + 288, + 317, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 288, + 317, + 467 + ], + "type": "inline_equation", + "content": "50.1\\%" + }, + { + "bbox": [ + 130, + 288, + 317, + 467 + ], + "type": "text", + "content": " of the GFLOPs, during training and inference, respectively, relative to prior methods. While OS-Prompt++ maintains a training GFLOPs count comparable to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 324, + 314, + 479, + 411 + ], + "blocks": [ + { + "bbox": [ + 324, + 314, + 479, + 411 + ], + "lines": [ + { + "bbox": [ + 324, + 314, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 324, + 314, + 479, + 411 + ], + "type": "image", + "image_path": "f627728d5724d88c29b20fb157107406b148f3ddbe26e47fa8b11df005fbd2e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 422, + 482, + 444 + ], + "lines": [ + { + "bbox": [ + 321, + 422, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 482, + 444 + ], + "type": "text", + "content": "Fig. 4: Comparison of latency across PCL methods on different GPUs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "text", + "content": "earlier work due to the added feedforward process in the reference ViT, it does not employ the reference ViT during inference, bringing it down to " + }, + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "inline_equation", + "content": "50.1\\%" + }, + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "text", + "content": " GFLOPs. In Fig. 4, we compare the inference GPU latency of previous PCL methods with our OS-Prompt across three distinct GPU setups. All measurements are taken using a batch size of 32. Similar to the trend observed in FLOPs, our approach reduces the latency by " + }, + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 559, + 381, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 381, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 381, + 572 + ], + "type": "text", + "content": "5.3 Discussion on Training Computational Cost" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "text", + "content": "In this section, we provide an in-depth discussion of the training computational cost of PCL methods. Following [10], Table 4 presents the relative training complexity of each method with respect to ER [6], which is the simple baseline with standard gradient-based training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Here, we would like to clarify the training computational cost of prompt learning. In general neural network training, the forward-backward computational cost ratio is approximately 1:2. This is due to gradient backpropagation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 448, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\left(\\frac{dL}{da_l} = W_{l + 1}\\frac{dL}{da_{l + 1}}\\right)" + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "content": " and weight-updating " + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\left(\\frac{dL}{dW_l} = \\frac{dL}{da_{l + 1}} a_l\\right)" + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "content": " represents the loss, " + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "content": " denotes the weight parameter, " + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "content": " is the activation, and " + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "content": " is the layer index. In prompt learning, the forward-backward computational cost ratio is approximately 1:1. This is in contrast to general neural network training, as only a small fraction of the weight parameters (less than " + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 177 + ], + "type": "text", + "content": ") are updated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "text", + "content": "Bearing this in mind, we present the observations derived from the table, assuming that all methods employ the same architecture. (1) Previous PCL (L2P, DualPrompt, CodaPrompt) consists of two steps: First, the query ViT requires only feedforward without backpropagation, Second, the backbone ViT requires feedforward-backward training with prompt tuning. This results in the relative training computational cost " + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "inline_equation", + "content": "1 \\left( = \\frac{\\text{QueryViT}_{fw}(1) + \\text{BackboneViT}_{fw}(1) + \\text{BackboneViT}_{bw}(1)}{\\text{ER}_{fw}(1) + \\text{ER}_{bw}(2)} \\right)" + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "inline_equation", + "content": "fw" + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "inline_equation", + "content": "bw" + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "text", + "content": " denote forward and backward propagation respectively. (2) Our OS-Prompt requires one feedforward-backward step like standard training. Therefore, relative training cost becomes " + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "inline_equation", + "content": "\\frac{2}{3} \\left( = \\frac{\\text{BackboneViT}_{fw}(1) + \\text{BackboneViT}_{bw}(1)}{\\text{ER}_{fw}(1) + \\text{ER}_{bw}(2)} \\right)" + }, + { + "bbox": [ + 130, + 177, + 486, + 329 + ], + "type": "text", + "content": ". This shows the potential of our OS-Prompt on online continual learning. (3) Our OS-Prompt++ has a reference ViT to perform the regularization, bringing 1 relative training computational cost with respect to ER." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 346, + 387, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 387, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 387, + 357 + ], + "type": "text", + "content": "5.4 Impact of Unsupervised Pre-trained Weights" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 521 + ], + "type": "text", + "content": "One underlying assumption of existing PCL methods is their reliance on supervised pretraining on ImageNet-1k. While this pre-training has a substantial impact on model performance, it may not always be feasible for a general continual learning task. One possible solution is using unsupervised pre-trained models. To explore this, we use DINO pre-trained weights [4] instead of ImageNet-1k pre-trained weights, and compare the performance across ER, LwF, and PCL works. We set the other experimental settings as identical. We report the results with ImageNet-R 10-task in Table 5. Compared to the ImageNet-1k pre-trained model (with supervised training), performances across all methods degrade. This observation suggests that the backbone model plays a crucial role in PCL. Moreover, similar to the ImageNet-1k pre-trained model, CodaPrompt and our OS-Prompt outperform the other methods by a significant margin. This indicates that our method continues to perform well with various backbones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 538, + 320, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 320, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 320, + 551 + ], + "type": "text", + "content": "5.5 Analysis of Design Components" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "QR Loss Design. To understand the effect of different components in our proposed QR loss (Eq. 10), we conducted an ablation study with different settings, summarized in Table 6. There are two main components in the QR loss: Cosine Similarity and Softmax, and we provide the accuracy of different four combinations. Without Cosine Similarity and Softmax, our framework yields a performance of " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "75.00\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{N}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". Adding Cosine Similarity or Softmax improves the performance, suggesting their collaborative role in enhancing the model's performance in our QR loss design. We also provide the sensitivity analysis of our method with respect to the hyperparameter " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " in Table 7. Across the three" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 146, + 296, + 201 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "type": "text", + "content": "Table 6: Analysis of QR loss design on 10-task ImageNet-R setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 146, + 296, + 201 + ], + "lines": [ + { + "bbox": [ + 136, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 136, + 146, + 296, + 201 + ], + "type": "table", + "html": "
CosSim SoftmaxAN(↑)FN(↓)
✓75.00 ± 0.531.68 ± 0.12
75.47 ± 0.421.38 ± 0.16
✓75.51 ± 0.331.28 ± 0.06
✓75.67 ± 0.401.27 ± 0.10
", + "image_path": "45e976f8e84ad191de8903a8c8a3b34e8033987862937f3e2f2a61dbe46918b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 312, + 146, + 478, + 200 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 483, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 483, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 483, + 137 + ], + "type": "text", + "content": "Table 7: Hyperparameter sensitivity study of " + }, + { + "bbox": [ + 307, + 114, + 483, + 137 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 307, + 114, + 483, + 137 + ], + "type": "text", + "content": " on ImageNet-R 5/10/20-task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 146, + 478, + 200 + ], + "lines": [ + { + "bbox": [ + 312, + 146, + 478, + 200 + ], + "spans": [ + { + "bbox": [ + 312, + 146, + 478, + 200 + ], + "type": "table", + "html": "
λTask-5Task-10Task-20
1e-577.03 ± 0.1075.63 ± 0.3973.63 ± 0.21
5e-577.02 ± 0.1375.62 ± 0.4173.62 ± 0.19
1e-477.07 ± 0.1575.67 ± 0.4073.77 ± 0.19
5e-477.13 ± 0.2475.68 ± 0.3873.68 ± 0.17
", + "image_path": "a0af50cd6f9ff49a818b13de55fbf4428b0ec66dd1e96afc46fe3a1bfeac75c5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 142, + 213, + 239, + 291 + ], + "blocks": [ + { + "bbox": [ + 142, + 213, + 239, + 291 + ], + "lines": [ + { + "bbox": [ + 142, + 213, + 239, + 291 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 239, + 291 + ], + "type": "image", + "image_path": "b2c63dc9d14f81e3ae7306c59a02bccfe8844649d1e5342720f1f1b238492e30.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 302, + 357, + 335 + ], + "lines": [ + { + "bbox": [ + 131, + 302, + 357, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 302, + 357, + 335 + ], + "type": "text", + "content": "Fig. 5: Analysis of accuracy " + }, + { + "bbox": [ + 131, + 302, + 357, + 335 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 131, + 302, + 357, + 335 + ], + "type": "text", + "content": " with respect to prompt components (Left) and prompt length (Right). We use 10-taks ImageNet-R setting." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 248, + 213, + 344, + 292 + ], + "blocks": [ + { + "bbox": [ + 248, + 213, + 344, + 292 + ], + "lines": [ + { + "bbox": [ + 248, + 213, + 344, + 292 + ], + "spans": [ + { + "bbox": [ + 248, + 213, + 344, + 292 + ], + "type": "image", + "image_path": "f23e8a4243edee892875192a0ffc089383e3bd5b9e8628339e53d492b9bae222.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 302, + 481, + 335 + ], + "lines": [ + { + "bbox": [ + 364, + 302, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 364, + 302, + 481, + 335 + ], + "type": "text", + "content": "Fig.6: Trade-off between GFLOPs and accuracy within a reference ViT." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 366, + 213, + 479, + 291 + ], + "blocks": [ + { + "bbox": [ + 366, + 213, + 479, + 291 + ], + "lines": [ + { + "bbox": [ + 366, + 213, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 366, + 213, + 479, + 291 + ], + "type": "image", + "image_path": "af6f6432b03f12338a895483944c790af25cdaa85d480daa8485efd18395809a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 360, + 480, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 480, + 408 + ], + "type": "text", + "content": "distinct task configurations on ImageNet-R, we note that the performance fluctuations are minimal, despite the variation in " + }, + { + "bbox": [ + 130, + 360, + 480, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 360, + 480, + 408 + ], + "type": "text", + "content": " values. These results underline the robustness of our method with respect to " + }, + { + "bbox": [ + 130, + 360, + 480, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 360, + 480, + 408 + ], + "type": "text", + "content": " value, suggesting that our approach is not sensitive to hyperparameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 408, + 489, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 489, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 489, + 492 + ], + "type": "text", + "content": "Prompt Design. We further examine the impact of both the number of prompt components and the prompt length on accuracy. In Fig. 5 (left), we measure accuracy across configurations with " + }, + { + "bbox": [ + 130, + 408, + 489, + 492 + ], + "type": "inline_equation", + "content": "\\{10,20,50,100,200,500\\}" + }, + { + "bbox": [ + 130, + 408, + 489, + 492 + ], + "type": "text", + "content": " prompts. The OS-prompt demonstrates a consistent accuracy enhancement as the number of prompts increases. Notably, OS-prompt++ reaches a performance plateau after just 50 prompts. Our ablation study of prompt length, presented in Fig. 5 (right), reveals that ours maintains stable accuracy across various prompt lengths." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 512, + 430, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 512, + 430, + 524 + ], + "spans": [ + { + "bbox": [ + 131, + 512, + 430, + 524 + ], + "type": "text", + "content": "5.6 Accuracy-Efficiency Trade-off within a Reference ViT" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "During the training of OS-prompt++, our method does not improve the energy efficiency (Note, we achieve " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " computational saving during inference). This arises from our approach of extracting the reference prompt query " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " from the final layer of the reference ViT. To further enhance energy efficiency during the OS-prompt++ training, we delve into the trade-off between accuracy and efficiency within the Reference ViT. Instead of relying on the last layer's [CLS] token embedding for the reference prompt query, we opt for intermediate token embeddings. Since our prompt pool is applied in layers 1 to 5 of the backbone ViT, we utilize the intermediate token embeddings from layers deeper than 5 within the reference ViT. Fig. 6 illustrates this trade-off with respect to the layer index where we get the reference prompt. Our findings show that util" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 168, + 445, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 8: Analysis on prompt formation methods. Results on ImageNet-R for different task configurations: 5 tasks (40 classes/task), 10 tasks (20 classes/task), and 20 tasks (10 classes/task). " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "A_{N}" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " represents the average accuracy across tasks, and " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "F_{N}" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " indicates the mean forgetting rate. We average values over five runs with different seeds." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 168, + 445, + 247 + ], + "lines": [ + { + "bbox": [ + 165, + 168, + 445, + 247 + ], + "spans": [ + { + "bbox": [ + 165, + 168, + 445, + 247 + ], + "type": "table", + "html": "
SettingTask-5Task-10Task-20
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
Deep L2P73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
OS-Prompt (L2P)74.61 ± 0.292.29 ± 0.2173.43 ± 0.601.42 ± 0.1071.42 ± 0.240.95 ± 0.06
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
OS-Prompt (Dual)74.65 ± 0.152.06 ± 0.2072.57 ± 0.231.13 ± 0.0370.55 ± 0.480.81 ± 0.10
CodaPrompt76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
OS-Prompt (Coda)75.74 ± 0.583.32 ± 0.3174.58 ± 0.561.92 ± 0.1572.00 ± 0.601.09 ± 0.11
", + "image_path": "12e2dccc6968af54314382a748e34c8e4be813e6c45e0e99ece564544947913c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 479, + 292 + ], + "type": "text", + "content": "layer 8, there is only a slight increase in accuracy, which suggests a potential to reduce GFLOPs without performance drop." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 307, + 362, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 362, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 362, + 319 + ], + "type": "text", + "content": "5.7 Analysis on Prompt Formation Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 325, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 542 + ], + "type": "text", + "content": "Here, we delve into the impact of varying prompt formation strategies on performance. For comparative insight, we present results from two previous PCL methodologies: L2P and DualPrompt. While L2P selects 5 prompts from the internal prompt pool, DualPrompt distinguishes between general and task-specific prompts. Aside from the distinct prompt formation strategies, the configurations remain consistent. In Table 8, we adopt the prompt formation strategies proposed in L2P and Dual, resulting in configurations denoted as OS-Prompt (L2P) and OS-Prompt (Dual). Our findings reveal the following: (1) The effectiveness of OS-prompt varies based on the prompt formation technique employed. Notably, OS-Prompt (L2P) and OS-Prompt (Dual) yield lower accuracy than our original approach, which relies on CodaPrompt. This suggests that our performance could benefit from refined prompt formation techniques in future iterations. (2) The OS-prompt framework, when integrated with L2P and Dual prompt formation strategies, outperforms the original Deep L2P and DualPrompt, a trend not observed with CodaPrompt. This may indicate that the prompt formation strategies of L2P and DualPrompt, which rely on hard matching (i.e., top-k), are more resilient than CodaPrompt's. Conversely, CodaPrompt employs soft matching, utilizing a weighted summation of all prompts based on proximity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "We introduce the OS-Prompt framework where we improve the efficiency of the conventional two-stage PCL in a simple yet effective manner. One of the potential limitations of our OS-Prompt++ is, in comparison to OS-Prompt (our light version), it introduces an increase in training computational cost from the reference ViT feedforward. This computational efficiency during training becomes important in the context of addressing the online continual learning problem [10] where the model needs rapid training on the streaming data." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 480, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 480, + 190 + ], + "type": "text", + "content": "This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), and the DoE MMICC center SEA-CROGS (Award #DE-SC0023198)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 213, + 197, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 197, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 197, + 224 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 240, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "text", + "content": "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: ECCV (2018) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 264, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 285 + ], + "type": "text", + "content": "2. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: NeurIPS (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "type": "text", + "content": "3. Cai, H., Gan, C., Zhu, L., Han, S.: Tinytl: Reduce memory, not parameters for efficient on-device learning. Advances in Neural Information Processing Systems 33, 11285-11297 (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 319, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 361 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 11, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "type": "text", + "content": "5. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 387, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 480, + 418 + ], + "type": "text", + "content": "6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) 3, 10, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 420, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 480, + 453 + ], + "type": "text", + "content": "7. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 454, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 485 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) 2, 5, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 519 + ], + "type": "text", + "content": "9. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 563 + ], + "type": "text", + "content": "0. Ghunaim, Y., Bibi, A., Alhamoud, K., Alfarra, M., Al Kader Hammoud, H.A., Prabhu, A., Torr, P.H., Ghanem, B.: Real-time evaluation in online continual learning: A new hope. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11888-11897 (2023) 11, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 586 + ], + "type": "text", + "content": "1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "type": "text", + "content": "12. Harun, M.Y., Gallardo, J., Hayes, T.L., Kanan, C.: How efficient are today's continual learning algorithms? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2430-2435 (2023) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 621, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 642 + ], + "type": "text", + "content": "13. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: ICRA (2019) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "14. He, X., Li, C., Zhang, P., Yang, J., Wang, X.E.: Parameter-efficient fine-tuning for vision transformers. arXiv preprint arXiv:2203.16329 (2022) 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "15. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., Gilmer, J.: The many faces of robustness: A critical analysis of out-of-distribution generalization. arXiv preprint arXiv:2006.16241 (2020) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "16. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV. pp. 8340-8349 (2021) 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 224 + ], + "type": "text", + "content": "17. Hsu, Y.C., Liu, Y.C., Ramasamy, A., Kira, Z.: Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488 (2018) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 224, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 481, + 256 + ], + "type": "text", + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "text", + "content": "19. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "20. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 333, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 354 + ], + "type": "text", + "content": "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 3, 5, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "22. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: ICML. pp. 3925-3934. PMLR (2019) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "23. Li, Z., Hoiem, D.: Learning without forgetting. TPAMI 40(12), 2935-2947 (2017) 3, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "text", + "content": "24. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. NeurIPS (2017) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "type": "text", + "content": "25. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "26. Pellegrini, L., Lomonaco, V., Graffieti, G., Maltoni, D.: Continual learning at the edge: Real-time training on smartphone devices. arXiv preprint arXiv:2105.13127 (2021) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "text", + "content": "27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 3, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "text", + "content": "28. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "text", + "content": "29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR. pp. 2001-2010 (2017) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "30. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "31. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: ICML. pp. 4548-4557 (2018) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 664 + ], + "type": "text", + "content": "32. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) 1, 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Y. Kim, Y. Li, P. Panda." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 632 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "text", + "content": "33. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) 1, 2, 3, 4, 5, 6, 7, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "34. Smith, J.S., Tian, J., Halbe, S., Hsu, Y.C., Kira, Z.: A closer look at rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2409-2419 (2023) 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "35. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1706-1716 (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "36. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "37. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "text", + "content": "38. Wang, Z., Zhan, Z., Gong, Y., Yuan, G., Niu, W., Jian, T., Ren, B., Ioannidis, S., Wang, Y., Dy, J.: Sparcl: Sparse continual learning on the edge. Advances in Neural Information Processing Systems 35, 20366-20380 (2022) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "text", + "content": "39. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022) 1, 2, 4, 5, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "40. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. CVPR (2022) 1, 2, 4, 5, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "text", + "content": "41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020) 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "43. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "type": "text", + "content": "44. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 533, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 578 + ], + "type": "text", + "content": "45. Zhang, J.O., Sax, A., Zamir, A., Guibas, L., Malik, J.: Side-tuning: a baseline for network adaptation via additive side networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 698-714. Springer (2020) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "46. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 264, + 91, + 447, + 102 + ], + "type": "text", + "content": "One-stage Prompt-based Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_content_list.json b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..77ab4d93aaf30d5838fff90c0513ded788fd1215 --- /dev/null +++ b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_content_list.json @@ -0,0 +1,1718 @@ +[ + { + "type": "text", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Guo $^{1,2,\\dagger}$ , Yuan Gao $^{1,\\dagger}$ , Yuxu Lu $^{3}$ , Huilin Zhu $^{1,2}$ , Ryan Wen Liu $^{1(\\boxtimes)}$ , and Shengfeng He $^{2(\\boxtimes)}$", + "bbox": [ + 240, + 210, + 761, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Wuhan University of Technology, Wuhan, China \nwenliu@whut.edu.cn", + "bbox": [ + 331, + 253, + 669, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Singapore Management University, Singapore shengfenghe@smu.edu.sg", + "bbox": [ + 341, + 282, + 661, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The Hong Kong Polytechnic University, Hong Kong, China \n† Equal Contribution \nhttps://github.com/gy65896/OneRestore", + "bbox": [ + 297, + 310, + 705, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In real-world scenarios, image impairments often manifest as composite degradations, presenting a complex interplay of elements such as low light, haze, rain, and snow. Despite this reality, existing restoration methods typically target isolated degradation types, thereby falling short in environments where multiple degrading factors coexist. To bridge this gap, our study proposes a versatile imaging model that consolidates four physical corruption paradigms to accurately represent complex, composite degradation scenarios. In this context, we propose OneRestore, a novel transformer-based framework designed for adaptive, controllable scene restoration. The proposed framework leverages a unique cross-attention mechanism, merging degraded scene descriptors with image features, allowing for nuanced restoration. Our model allows versatile input scene descriptors, ranging from manual text embeddings to automatic extractions based on visual attributes. Our methodology is further enhanced through a composite degradation restoration loss, using extra degraded images as negative samples to fortify model constraints. Comparative results on synthetic and real-world datasets demonstrate OneRestore as a superior solution, significantly advancing the state-of-the-art in addressing complex, composite degradations.", + "bbox": [ + 261, + 385, + 743, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image restoration is an essential task that entails the recovery of high-quality visuals from compromised inputs. The fidelity of these restored images is vital for the subsequent downstream tasks, such as robot navigation and autonomous driving. Considerable progress has been made in addressing single degradation scenarios by a single model (i.e., One-to-One models, see Fig. 2a), such as low-light conditions [16, 49, 59], haze [21, 41, 57], rain [11, 14, 31], and snow [4, 8, 38]. These methods have achieved remarkable success within their intended contexts. Nevertheless, real-world conditions often involve unpredictable and variable composite degradations that can severely impact image quality and clarity [28]. The", + "bbox": [ + 212, + 704, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce0911d14d93dc0caa79befdfb59a29e29097f3b39cd25b2871ee9ab6c5784fa.jpg", + "image_caption": [ + "Fig. 1: Our OneRestore allows fully controllable image restoration using scene descriptors derived from both automatic visual attribute extraction (top) and manual text embeddings (bottom)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "occurrence of mixed degrading conditions in one image can lead to intricate interactions, posing significant challenges to One-to-One solutions [22,52,54,60]. The necessity for seamless switching and decision-making across specialized methods for each condition proves to be inefficient and ineffective in practical applications. Therefore, there is a growing imperative for robust, all-encompassing image restoration approaches that can universally address multiple forms of degradation through a singular, dynamic framework.", + "bbox": [ + 212, + 393, + 787, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research has been directed toward One-to-Many models, which address various individual degradation factors within a single framework. These models are categorized into: those with partial parameter-sharing (One-to-Many $_p$ ) [2,19,28,45] and those with full parameter-sharing (One-to-Many $_f$ ) [6,26,35,42, 43]. One-to-Many $_p$ (left of Fig. 2b) restoration methods typically employ models with either multiple heads and tails or a single tail to address various types of image impairments. However, the complexity and size of these models tend to scale with the diversity of degradations they aim to correct, leading to increased training costs and parameter counts. Furthermore, when images suffer from a combination of degrading factors, these models may underperform. Some One-to-Many $_p$ techniques also depend on user intervention to select appropriate restoration weights, which is impractical for automated or dynamic settings. On the other hand, One-to-Many $_f$ (right of Fig. 2b) approaches train models using a direct learning strategy, where inputs are images with different individual degradations and the targets are their clear counterparts. However, this direct training approach is designed predominantly for single-category degradation issues, it does not adequately address the complex dynamics that arise when multiple types of impairments affect an image simultaneously. As a result, such methods can falter in practical applications where user control and precise directionality of restoration are crucial. For instance, in cases where users seek to improve low-light imagery, the model might inadvertently prioritize dehazing actions, deviating from user intent.", + "bbox": [ + 212, + 507, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d3bf3cb80d8a45e50e8cda9e7b379be45966ec8f4f6e29b3033ae91509144a2f.jpg", + "image_caption": [ + "(a) One-to-One" + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 372, + 261 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89c731f4eeff7a0e89a4ba6f770a22f2ca71c8b2ceaaac1535c8230c029872f7.jpg", + "image_caption": [ + "(b) One-to-Many" + ], + "image_footnote": [], + "bbox": [ + 375, + 145, + 589, + 260 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9673f91c9c1157ad1d621b7a5deefc002ef0a3d892847725de2faec3d48fea31.jpg", + "image_caption": [ + "(c) One-to-Composite", + "Fig.2: Structural comparison between One-to-One, One-to-Many, and our One-to-Composite image restoration methods." + ], + "image_footnote": [], + "bbox": [ + 591, + 145, + 784, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing from the preceding discussion, we delineate the primary challenges in developing image restoration models for composite degradations in real-world conditions:", + "bbox": [ + 212, + 339, + 782, + 383 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- How can the model be designed to adaptively discern the specific degradation afflicting the input image?", + "- What approaches enable a model to be trained efficiently across a spectrum of degradations while enhancing user control over the restoration process?", + "- How can the recovery process be optimized that not only approach the quality of clear images but also maintain distinction from various degraded forms?" + ], + "bbox": [ + 225, + 391, + 784, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In response to the challenges outlined, we present a unified model that builds upon established physical corruption models for degradations such as low light, haze, rain, and snow. This model effectively captures the complexity of real-world degradation conditions and forms the basis for constructing our Composite Degradation Dataset (CDD-11), which serves both training and evaluation purposes. To address the issue of composite degradation, our work introduces OneRestore, a transformer-based universal restoration architecture. This novel framework distinguishes itself from conventional methods by embodying a Oneto-Composite strategy, as depicted in Fig. 2c, which allows for precise and controllable outcomes (see Fig. 1). It employs an innovative cross-attention block that fuses scene descriptors, which can be manual text embeddings or derived automatically from visual attributes. We further enhance the model's robustness by integrating a composite degradation restoration loss, using additional degraded images as negative samples to establish stringent lower-bound constraints.", + "bbox": [ + 212, + 487, + 787, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key contributions are summarized as follows:", + "bbox": [ + 238, + 699, + 596, + 713 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first attempt to unify composite degradations, creating a versatile imaging model that simulates a range of degradation types. This innovation underpins the development of our comprehensive Composite Degradation Dataset.", + "- We propose a universal framework that introduces a novel level of controllability in image restoration, employing a cross-attention mechanism that integrates scene descriptors derived from either manual text embeddings or automated visual attribute extractions." + ], + "bbox": [ + 225, + 719, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We tailor a composite degradation restoration loss, enhancing the model's ability to discern between various forms of image degradation.", + "bbox": [ + 223, + 146, + 782, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Our universal model not only sets a new state-of-the-art for composite dataset performance but also competes favorably with One-to-One models in single-degradation scenarios.", + "bbox": [ + 225, + 176, + 784, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 239, + 387, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Complex Degradation Image Restoration. Recent advancements in image restoration have introduced a variety of solutions for complexly degraded images. A notable direction is the partial parameter-sharing One-to-Many approach, employing architectures with multiple heads and tails, and a single backbone, to address various degradation factors. For instance, Li et al. [28] develop an all-weather network that restores scenes affected by haze, rain, and snow, using specialized encoders for each condition and a shared decoder. Chen et al. [2] introduce a transformer-based model for deraining, denoising, and super-resolving images, leveraging a multi-head and multi-tail design. Wang et al. [45] present a knowledge assignment strategy to differentiate and process rain and snow degradations adaptively. While these models represent strides in complex degradation image restoration, they stop short of being truly universal One-to-Many networks. With only partial weight sharing, they address different degradation types through distinct network pathways.", + "bbox": [ + 212, + 266, + 787, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conversely, universal image restoration has emerged with the One-to-Many methods sharing all network weights, thereby uniformly mitigating multiple degradation types. Examples include the unified two-stage knowledge learning framework by Chen et al. [6], which efficiently processes dehazing, deraining, and desnoring with a singular pretrained weight set. Li et al. [26] present a network capable of comprehensive image restoration without the need for pre-identifying degradation levels. Additionally, Ozdenizci et al. [35] introduce a conditional denoising diffusion model tailored for patch-based image restoration under varied weather conditions. However, due to the significant differences and mutual interference between various degradation factors, this type of method can only solve specific types of degradation and cannot achieve image restoration in composite degradation scenarios. To this end, we propose a universal One-to-Composite image restoration method that allows the model to adapt to arbitrary combinations of composite degradations.", + "bbox": [ + 212, + 478, + 787, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Attribute Control. Visual attributes are pivotal in defining image features and enhancing the interpretability of models. They enrich feature representations and provide a hierarchical framework for image analysis. Early research highlighted visual attributes' transformative impact on image generation, recognition, and retrieval [13, 24, 50]. More recent studies underscore their utility in improving object recognition and detection, particularly in data-scarce environments [1, 33, 48, 61]. Visual attributes have also been leveraged for zero-shot learning applications, advancing classification tasks beyond the constraints of labeled datasets [39, 44]. In our study, we harness visual attributes to delineate the intricate degradation patterns present in real-world imaging conditions,", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "enhancing the pertinence and efficacy of image restoration techniques. It is noteworthy that conventional discriminators, when tasked with identifying image degradation, are typically confined to image data, precluding autonomous feature control. Our approach, by contrast, accommodates both images and text descriptions as inputs, thereby facilitating the automatic or manual determination of degradation characteristics.", + "bbox": [ + 212, + 146, + 782, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Composite Degradation Formulation", + "text_level": 1, + "bbox": [ + 215, + 258, + 606, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Imaging Formulation", + "text_level": 1, + "bbox": [ + 215, + 290, + 434, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In real-world scenarios, images are often affected by multiple degrading factors simultaneously, resulting in suboptimal performance from restoration methods designed for isolated degradations. To address the broad spectrum of degradation scenarios, we introduce a versatile imaging model that encapsulates all degradations, including low light, haze, rain, and snow. This model aims to narrow the disparity between simulated and authentic real-world data. Given a degraded image $I(x)$ and its corresponding clear state $J(x)$ , the proposed imaging model is mathematically articulated as", + "bbox": [ + 212, + 313, + 784, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI (x) = \\mathcal {P} _ {h} \\left(\\mathcal {P} _ {r s} \\left(\\mathcal {P} _ {l} (J (x)))\\right), \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 446, + 785, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_l, \\mathcal{P}_{rs}$ , and $\\mathcal{P}_h$ denote the low light, rain/snow, and haze degradations, respectively. Given that concurrent rain and snow conditions are infrequent, they are not addressed within the scope of this work. A comprehensive exposition of the proposed imaging model is detailed below.", + "bbox": [ + 212, + 470, + 785, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Low-Light Conditions. Inspired by Retinex theory, low-light degraded image $I_{l}(x)$ can be generated by", + "bbox": [ + 212, + 532, + 785, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l} (x) = \\mathcal {P} _ {l} (J (x)) = \\frac {J (x)}{L (x)} L (x) ^ {\\gamma} + \\varepsilon . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 573, + 785, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we use a darkening coefficient $\\gamma \\in [2,3]$ as the brightness controller to adjust the illumination map $L(x)$ generated via LIME [17]. To simulate poor light conditions realistically, we add the Gaussian noise $\\varepsilon$ with mean $= 0$ and variance $\\in [0.03,0.08]$ .", + "bbox": [ + 212, + 614, + 782, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rain/Snow Streaks. Following the methodologies outlined in [5, 27], our process superimposes rain or snow streaks onto the image before synthesizing haze. To introduce rain streaks $\\mathcal{R}$ , we produce the rain-degraded image $I_{rs}(x)$ by overlaying $\\mathcal{P}_{rs}$ onto the low-light image $I_l(x)$ , which is expressed as", + "bbox": [ + 212, + 675, + 785, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {r s} (x) = \\mathcal {P} _ {r s} \\left(I _ {l} (x)\\right) = I _ {l} (x) + \\mathcal {R}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 747, + 785, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For snow streaks $S$ , the snow-degraded image is defined as", + "bbox": [ + 215, + 772, + 638, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {r s} (x) = \\mathcal {P} _ {r s} \\left(I _ {l} (x)\\right) = I _ {l} (x) (1 - \\mathcal {S}) + M (x) \\mathcal {S}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 800, + 785, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $M(x)$ being the chromatic aberration map.", + "bbox": [ + 215, + 824, + 563, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/daf44e536a16f742dade2485957e8ba3a71419b22b4d49cc02457034b5a2fea6.jpg", + "image_caption": [ + "(a) Overall Pipeline" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 486, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5133cc4a36aac790ceb6d30ac7a73882055fdfc11d2766e109f5e5b6921c18d8.jpg", + "image_caption": [ + "(b) Scene Descriptor Generation", + "Fig. 3: Architecture of our OneRestore. (a) Overall pipeline, where 32, 64, 128, and 256 represent the number of channels. (b) Scene descriptor generation, where scene descriptors are fed into each (c) Scene Descriptor-guided Transformer Block (SDTB) by manual text embeddings or automatic extraction based on visual attributes." + ], + "image_footnote": [], + "bbox": [ + 488, + 142, + 689, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03612244224db9547ef4f2d7a3d07af8efcde1069e026f0ad7c810984d074af4.jpg", + "image_caption": [ + "(c)SDTB" + ], + "image_footnote": [], + "bbox": [ + 694, + 143, + 785, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Haze Degradations. We incorporate haze degradation by employing the atmospheric scattering model as", + "bbox": [ + 212, + 397, + 787, + 428 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI (x) = \\mathcal {P} _ {h} \\left(I _ {r s} (x)\\right) = I _ {r s} (x) t + A (1 - t), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 440, + 785, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t$ denotes the transmission map and $A$ represents the atmospheric light. The transmission map $t$ is defined by the exponential decay of light, $t = e^{-\\beta d(x)}$ , where $\\beta$ represents the haze density coefficient and $d(x)$ is the scene depth (estimated by MegaDepth [29]). The haze density coefficient $\\beta$ is varied within the range of [1.0, 2.0] and the atmospheric light $A$ is constrained to [0.6, 0.9] for realistic simulation.", + "bbox": [ + 212, + 469, + 787, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Composite Degradation Dataset", + "text_level": 1, + "bbox": [ + 214, + 585, + 527, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Drawing on Eq. (1), we have developed the Composite Degradation Dataset (CDD-11), encompassing 11 categories of image degradations and their clear counterparts. These degraded samples include low (low-light), haze, rain, snow, low+haze, low+rain, low+snow, haze+rain, haze+snow, low+haze+rain, and low+haze+snow. From the RAISE database [10], we selected 1,383 high-resolution clear images for producing 11 composite degradations. They are resized to a uniform resolution of $1080 \\times 720$ . The overall dataset is split into 13,013 image pairs for training and 2,200 for testing.", + "bbox": [ + 212, + 613, + 789, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 OneRestore", + "text_level": 1, + "bbox": [ + 214, + 760, + 366, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The encoder-decoder network, validated as an effective architectural paradigm for image restoration [7, 12, 18], serves as the overall structure for our model, as depicted in Fig. 3a. Building upon the encoder-decoder structure, we devise", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "an innovative Scene Descriptor-guided Transformer Block (SDTB) as the basic unit for feature extraction (Fig. 3c). Our approach diverges from existing transformer-based methods [4,7,41] by concurrently integrating image features and scene descriptors into the SDTB basic block for supporting fine-grained image restoration. In particular, the generation of scene descriptors ranges from manual text embeddings to automatic extraction based on visual attributes, as illustrated in Fig. 3b. Further details on scene descriptor generation and SDTB are elaborated in Sections 4.1 and 4.2. For capturing multi-scale representations, the encoder employs three downsampling operations, mirrored by the decoder's upsampling operations. To be specific, we utilize \"Max-Pooling + Conv\" and \"Bilinear Interpolation + Conv\" for downsampling and upsampling, respectively. Simultaneously, three skip connections and a global residual enhance model convergence by bridging shallow and deep features. Additional details regarding the configuration of our OneRestore can be found in supplementary materials.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ultimately, the overall loss $\\mathcal{L}$ for model training is defined as follows", + "bbox": [ + 240, + 358, + 733, + 372 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {1} ^ {s} (J, \\hat {J}) + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {M}} (J, \\hat {J}) + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {c}} (J, \\hat {J}, I, \\{I _ {o} \\}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 380, + 784, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_1^s$ , $\\mathcal{L}_{\\mathrm{M}}$ , and $\\mathcal{L}_{\\mathrm{c}}$ represent the smooth $l_{1}$ loss [15], Multi-Scale Structural Similarity (MS-SSIM) loss, and the proposed composite degradation restoration loss (see Sec. 4.3), respectively, $\\alpha_{1-3}$ denote the penalty coefficients, $J$ , $\\hat{J}$ , $I$ , and $\\{I_o\\}$ are the positive, model output anchor, model input negative, and other negatives, respectively.", + "bbox": [ + 214, + 405, + 787, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Scene Descriptor Generation", + "text_level": 1, + "bbox": [ + 214, + 501, + 501, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To augment the controllability of our OneRestore, we introduce the scene description embedding as an additional input to the model, alongside the degraded image. As delineated in Fig. 3b, the generation of the scene descriptor offers two selectable modes: the manual mode $①$ and the automatic mode $②$ . In the manual mode, user involvement is necessitated to supply scene description text, generating the corresponding text embedding employed as an input. Conversely, the automatic mode involves the extraction of visual attributes from the image to generate visual embedding, estimating the most proximate text embedding. To quantify the difference between visual and text embeddings and train optimal text and visual embedders, we introduce cosine cross-entropy loss for model weight optimization.", + "bbox": [ + 212, + 522, + 787, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text Embedder. For the text embedding generation task, we employ a set of 5 scene description texts as input to generate 12 text embeddings. Initially, GloVe [36] is leveraged to generate initial text embeddings for 1 clear scene and 4 scenes with single degradation types. Subsequently, 7 additional composite degradation text embeddings (i.e., low+ haze, low+ rain, low+ snow, haze+ rain, haze+ snow, low+ haze+ rain, and low+ haze+ snow) are generated by averaging the element values in the corresponding text embeddings. A Multi-Layer Perception (MLP) is then employed to refine the 12 text embeddings.", + "bbox": [ + 212, + 689, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Embedder. Taking inspiration from the work of Saini et al. [39], we first resize the input image to $224 \\times 224$ and extract image features using a", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ResNet-18 [20] network without average pooling. The initial weights of ResNet-18 are pre-trained on the ImageNet dataset. Subsequently, a single convolutional layer, followed by a dropout layer and a linear layer, is introduced to generate the final visual embedding. By calculating the cosine similarity between visual embeddings and all text embeddings, and selecting the one with the highest similarity, we can automatically generate scene descriptors to achieve image restoration based on visual attribute control.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cosine Cross-Entropy. Similar to [39], the score of visual embedding and each text embedding is obtained by \"Cosine Similarity + Softmax\". Given $e_v$ and $e_t$ being the visual and text embeddings, the similarity score $S(e_v, e_t)$ can be given by", + "bbox": [ + 212, + 252, + 784, + 313 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\cos \\left(e _ {v}, e _ {t}\\right) = \\delta \\cdot \\frac {e _ {v} \\cdot e _ {t} ^ {\\top}}{\\| e _ {v} \\| \\| e _ {t} \\|}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 323, + 784, + 368 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS \\left(e _ {v}, e _ {t}\\right) = \\frac {e ^ {\\cos \\left(e _ {v} , e _ {t}\\right)}}{\\sum_ {t _ {i} = 1} ^ {N _ {t}} e ^ {\\cos \\left(e _ {v} , e _ {t _ {i}}\\right)}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 354, + 784, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\delta$ and $N_{t}$ being the temperature factor and the number of text embeddings, respectively. Based on the calculated $S(e_v, e_t)$ , we use the cross-entropy loss for model training.", + "bbox": [ + 212, + 407, + 782, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Scene Descriptor-guided Transformer Block", + "text_level": 1, + "bbox": [ + 214, + 476, + 620, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further harness the capabilities of scene description embeddings in image restoration tasks and facilitate a more adaptive and nuanced restoration process, we introduce the innovative Scene Descriptor-guided Transformer Block (SDTB), illustrated in Fig. 3c. The SDTB consists of three primary components, namely Scene Descriptor-guided Cross-Attention (SDCA), Self-Attention (SA), and Feed-Forward Network (FFN). Diverging from traditional self-attention mechanisms, the proposed SDCA module utilizes a query generated from scene descriptors instead of queries generated from image features. The mathematical representation of our SDCA can be written as", + "bbox": [ + 212, + 501, + 784, + 636 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S D C A} \\left(\\mathbf {Q} _ {t}, \\mathbf {K}, \\mathbf {V}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t} \\cdot \\mathbf {K} ^ {\\top}}{\\lambda}\\right) \\mathbf {V}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 648, + 784, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ is a temperature factor, $\\mathbf{Q}_t$ represents the scene descriptor-generated query matrix, $\\mathbf{K}$ and $\\mathbf{V}$ denote the image feature-generated key and value matrices, respectively. $\\mathbf{Q}_t$ is obtained from the input scene description embedding $e_t$ via a linear layer. During the process of matrix multiplication between $\\mathbf{Q}_t$ and $\\mathbf{K}$ . To ensure consistency in the number of tokens between $\\mathbf{K}$ and $\\mathbf{Q}_t$ , we resize the original image before performing the reshaping operation to generate $\\mathbf{K}$ . By incorporating the SDCA module, each transformer block in OneRestore can effectively integrate scene descriptors and image features for targeted image restoration. As for the SA and FFN modules, we adopt structures akin to those in Restormer [52]. More details of our SDTB are introduced in the supplement.", + "bbox": [ + 212, + 689, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13540744a2f4e1a7cae699f1a7cd0e6895e301f0fabbac1cd87ca92d4d2722f3.jpg", + "image_caption": [ + "Fig. 4: Illustration of the composite degradation restoration loss." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 785, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Composite Degradation Restoration Loss", + "text_level": 1, + "bbox": [ + 215, + 316, + 602, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Using contrastive loss has been demonstrated as an effective way of improving image restoration performance [47, 57]. Essentially, contrastive loss views the degraded input image as a negative sample, the clear image as a positive sample, and the restored image as an anchor. This loss function works to minimize the distance between the anchor and the positive sample while maximizing its distance from the negative sample. In the composite degradation image restoration task, the traditional contrastive loss, however, may inadvertently reduce the distance between the anchor and other degraded images. To make the anchor closer to the positive sample while maintaining a clear separation from all degraded negatives, we design a composite degradation restoration loss, illustrated in Fig. 4. Considering $I$ , $I_o$ , $J$ , and $\\hat{J}$ as the input negative, other negatives, positive, and anchor, the proposed loss can be written as", + "bbox": [ + 212, + 343, + 787, + 512 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} (J, \\hat {J}, I, \\left\\{I _ {o} \\right\\}) = \\sum_ {k = 1} ^ {K} \\xi_ {k} \\frac {\\mathcal {L} _ {1} \\left(V _ {k} (\\hat {J}) , V _ {k} (\\hat {J})\\right)}{\\xi_ {c} \\mathcal {L} _ {1} \\left(V _ {k} (\\hat {J}) , V _ {k} (I)\\right) + \\sum_ {o = 1} ^ {O} \\xi_ {o} \\mathcal {L} _ {1} \\left(V _ {k} (I _ {o}), V _ {k} (\\hat {J})\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 258, + 522, + 785, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $V_{k}(\\cdot)$ denotes the extraction operation of $k$ -th hidden feature from the VGG-16 pre-trained on the ImageNet, $\\mathcal{L}_1$ represents the $l_1$ loss, $K$ and $O$ are the numbers of used VGG-16 feature layers and other degraded negatives, $\\xi_{k}, \\xi_{c}$ , and $\\xi_{o}$ denote the hyper-parameters, respectively. In this work, the number of other degraded negatives $O$ is 10. We use the output features of 3-rd, 8-th, and 15-th layers from VGG-16 and set $K = 3$ . Furthermore, we consider that the features of each feature layer of each degraded negative have equal importance. Therefore, we set $\\xi_{c} = \\xi_{o} = \\frac{1}{11}$ and $\\xi_{k} = \\frac{1}{3}$ . According to our ablation study, the proposed loss confers a significant advantage over using a single input degradation as a negative sample, which can easily capture positive sample features.", + "bbox": [ + 212, + 569, + 787, + 722 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cfda75b32790de40772a8637d1084cd3ed21b9c980d7982b9e95012ca4ca99aa.jpg", + "table_caption": [ + "Table 1: Settings of model training. \"lr\" denotes the learning rate." + ], + "table_footnote": [], + "table_body": "
TermsText/Visual EmbedderOneRestore
Epoch200120
Initial lr0.00010.0002
lr Decay0.5/50 epochs0.5/20 epochs
Batch Size2564
", + "bbox": [ + 217, + 773, + 785, + 837 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/57b06a0d0d95d167d47e2676e0560c85d84c547849b37a60d4460c78c1ce3269.jpg", + "table_caption": [ + "Table 2: Comparison of quantitative results on CDD-11 dataset. OneRestore† means using the corresponding scene description text as additional input, and WGWSNet requires the scene type as model input. Red, green, and blue indicate the best, second-best, and third-best results, respectively." + ], + "table_footnote": [], + "table_body": "
TypesMethodsVenue & YearPSNR ↑SSIM ↑#Params
Input16.000.6008-
One-to-OneMIRNet [53]ECCV202025.970.847431.79M
MPRNet [54]CVPR202125.470.855515.74M
MIRNetv2 [55]TPAMI202225.370.83355.86M
Restormer [52]CVPR202226.990.864626.13M
DGUNet [34]CVPR202226.920.855917.33M
NAFNet [3]ECCV202224.130.796417.11M
SRUDC [40]ICCV202327.640.86006.80M
Fourmer [58]ICML202323.440.78850.55M
OKNet [9]AAAI202426.330.86054.72M
One-to-ManyAirNet [26]CVPR202223.750.81408.93M
TransWeather [42]CVPR202223.130.781021.90M
WeatherDiff [35]TPAMI202322.490.798582.96M
PromptIR [37]NIPS202325.900.849938.45M
WGWSNet [62]CVPR202326.960.862625.76M
One-to-CompositeOneRestore28.470.87845.98M
OneRestore†28.720.88215.98M
", + "bbox": [ + 215, + 213, + 789, + 406 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 434, + 375, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 215, + 465, + 433, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Our OneRestore is implemented by PyTorch 1.12.0 and trained on a PC with 2 AMD EPYC 7543 32-Core Processors and 8 NVIDIA L40 GPUs. We use the Adam with exponential decay rates being $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ for optimization. The training parameter settings for text/visual embedder and OneRestore are shown in Table 1. In particular, all images used for training are cropped into $256\\times 256$ image patches with a sampling stride of 200 and randomly flipped by 0, 90, 180, and 270 degrees to generate $312\\mathrm{k}$ patch pairs. More details about training and inference can be found in the supplement.", + "bbox": [ + 212, + 491, + 787, + 612 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Datasets. To facilitate a thorough comparison, we execute image restoration experiments across various datasets. These include the Composite Degradation Dataset (CDD-11) crafted by us, established One-to-One benchmarks, and authentic real-world datasets. The standard One-to-One benchmarks are designated for four principal tasks: the LOw-Light dataset (LOL) [46], the REalistic Single Image DEhazing Outdoor Training Set (RESIDE-OTS) [25], the Rain1200 dataset [56], and the Snow100k dataset [30]. Owing to the dearth of specialized datasets for real-world composite degradation, we have curated a selection from the LIME low-light enhancement dataset [17], the RESIDE Realworld Task-driven Testing Set (RESIDE-RTTS) [25], Yang's rainy dataset [51], and the realistically snowy Snow100k dataset [25].", + "bbox": [ + 212, + 613, + 787, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Competitors and Evaluation Metrics. We compare our OneRestore with 9 One-to-One image restoration methods (MIRNet [53], MPRNet [54], MIRNetv2 [55], Restormer [52], DGUNet [34], NAFNet [3], SRUDC [40], Fourmer [58], and OKNet [9]) and 5 One-to-Many image restoration methods (AirNet [26],", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "TransWeather [42], WeatherDiff [35], PromptIR [37], and WGWSNet [62]). In particular, all compared state-of-the-art methods are retrained on our CDD-11 train set. For PromptIR and WGWSNet, we adjust the number of prompts and the weather-specific parameters of each layer for corresponding to our task, respectively. Furthermore, we adopt the Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM) as evaluation metrics.", + "bbox": [ + 212, + 145, + 787, + 238 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Comparison with SOTAs", + "text_level": 1, + "bbox": [ + 214, + 257, + 468, + 273 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The quantitative evaluation results on our CDD-11 test dataset are displayed in Table 2. OneRestore† with text embedder achieves the highest performance, followed by OneRestore with visual embedder as the second-best. Both OneRestore† and OneRestore demonstrate substantial leads over the third-ranked methods in terms of both PSNR and SSIM. Further detailed restoration performance comparison of our method and the One-to-Many methods in each degradation scenario is illustrated in Fig. 5 using radar charts. Notably, image restoration for composite degradation scenarios proves challenging. By incorporating scene description em", + "bbox": [ + 212, + 282, + 540, + 507 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2f3bc6b251e00659ff9072372f5dd158b5a3e13641f7cda6abdf13592a185b96.jpg", + "image_caption": [ + "Fig.5: Comparison of quantitative results for different degradation scenarios on CDD-11 dataset." + ], + "image_footnote": [], + "bbox": [ + 550, + 311, + 785, + 424 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "beddings into image restoration models and utilizing the proposed composite degradation restoration loss, our method robustly reconstructs clear images across various situations. For compelling evidence, Fig. 6 provides the visual comparison of image restoration in two composite degradation samples. It is evident that SRUDC, OKNet, AirNet, and WGWSNet are unable to suppress all degenerative interferences, while Restormer and PromptIR are prone to color distortion and noise residual. In contrast, the proposed OneRestore can produce more natural results and fully preserve image textures.", + "bbox": [ + 212, + 508, + 787, + 628 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, Fig. 7 displays the image restoration results in two real-world scenarios. Compared with existing One-to-Many image restoration methods [26,35, 37], the proposed OneRestore can simultaneously focus on multiple degradation factors and robustly reconstruct scene details. More results on CDD-11, realworld scenarios, and other One-to-One benchmarks of low-light, hazy, rainy, and snowy scenes are included in supplementary materials.", + "bbox": [ + 212, + 628, + 787, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 739, + 390, + 756 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of Network Modules. Table 3 reports the quantitative evaluation of different module configurations. While the Feed-Forward Network (FFN) with convolution as the basic unit serves as a crucial module in the transformer network, it alone fails to generate satisfactory reconstruction results. The ordinary Self-Attention module (SA) significantly strengthens network performance,", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b218a344f70fced369512260e59c87b18705cdc09eac3fdaa87955ec1a89956d.jpg", + "image_caption": [ + "Fig. 6: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) synthetic samples." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 785, + 272 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/344e23f352e4b9baa19335a5f028e1aefaad89bfa4d796b6923b483d4ae9bdea.jpg", + "image_caption": [ + "Fig. 7: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) samples in real-world scenarios. * represents the utilization of original weights published in the author's code." + ], + "image_footnote": [], + "bbox": [ + 217, + 328, + 785, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "but this method regrets that the model fails to be flexibly controlled. Compared to depending only on SA, relying exclusively on Scene Description-guided Cross-Attention (SDCA) utilizing scene description embedding yields better outcomes. This further confirms that addressing complex composite degradations can be done more effectively and flexibly by integrating degradation scene descriptors into the model. Ultimately, the best performance can be achieved by using a combination of SDCA, SA, and FFN.", + "bbox": [ + 212, + 505, + 787, + 611 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Selection of Description Embedding Strategies. In this section, we compare the image restoration performance by using different strategies to generate scene descriptor, including Classifier (utilizing the embedding before the last linear layer of the retrained AlexNet [23] as the scene descriptor), Visual Embedder (employing visual embedding for model training), Visual Embedder‡ (deriving text embedding via visual embedder), and Text Embedder (employing text embedding for model training). The quantitative evaluation results are presented in Table 4. It is observed that employing Classifier and Visual Embedder to generate scene descriptors", + "bbox": [ + 212, + 613, + 539, + 840 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1975813ece73deb6c035bf2ef205dcb59bbb05a99dde20ac023e1beaf6f728aa.jpg", + "image_caption": [ + "Fig.8: Comparison of image restoration on different description embedding strategies." + ], + "image_footnote": [], + "bbox": [ + 555, + 643, + 782, + 750 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e38c18451995c77ee745d876990221508dca8b1b15c084bad65552e72cb814b8.jpg", + "table_caption": [ + "Table 3: Ablation study for different model configurations. SDCA, SA, and FFN denote the scene description-guided cross-attention, self-attention, and feed-forward network, respectively." + ], + "table_footnote": [], + "table_body": "
SDCASAFFNPSNR ↑SSIM ↑Controllability
✓24.810.8607
✓✓27.190.8697
✓✓27.930.8767✓
✓✓✓28.720.8821✓
", + "bbox": [ + 217, + 199, + 785, + 258 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8cbc7b31a0ae2f37c7326a6ccfacb78e5e528a8bd96cba6918db0a56202e2540.jpg", + "image_caption": [ + "Fig. 9: Comparison of image restoration on $low + haze + rain$ (top) and $low + haze + snow$ (bottom) synthetic samples by using different texts." + ], + "image_footnote": [], + "bbox": [ + 217, + 279, + 785, + 352 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "yields comparable outcomes. However, these approaches have a drawback in that each image has a distinct embedding as input, making it challenging for the model to recognize the cause of degradation. Fig. 8 displays a classic case comparison using different embedding strategies, where the abnormal artifacts of the results produced by Classifier and Visual Embedder are caused by the model's fuzzy identification of degradation factors. Moreover, the restoration model lacks controllability. The discrepancy between Visual Embedder $\\ddagger$ and Text Embedder stems from errors in scene descriptor estimation. Visual Embedder $\\ddagger$ has an accuracy of $97.55\\%$ on the test dataset, with misestimation occurring when some degradation types are insignificant. For instance, a haze+snow scene may be evaluated as a haze or snow scene. In contrast, the text embedding-based approach is superior in describing degradation scenes and recovering images, using a fixed number of scene descriptors. This allows the model to employ degraded scene descriptors as switches to achieve controlled restoration.", + "bbox": [ + 212, + 428, + 787, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Loss Functions. Table 5 presents a performance comparison of different loss functions. Notably, the integration of the MS-SSIM loss is instrumental in elevating SSIM scores at the cost of reduced PSNR performance. In contrast, the CL loss demonstrates minimal impact, providing only marginal improvements in SSIM and PSNR. Upon replacing CL with CDRL, a substantial enhancement is observed in both PSNR and SSIM. The proposed CDRL loss yields an output that closely resembles the clear image while avoiding other forms of degradation.", + "bbox": [ + 212, + 643, + 558, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2c5fd263c7c705af5b4b73760ba662c7b5ba5e1ef6fcddf6a42b9cd9253ee00c.jpg", + "image_caption": [ + "Fig.10: Visualization of visual embeddings based on t-SNE." + ], + "image_footnote": [], + "bbox": [ + 568, + 674, + 787, + 776 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8d69424b2dc0020dd2562979f6e71ff3b161eac6b6361010618569b8257f98a0.jpg", + "table_caption": [ + "Table 4: Ablation study for description embedding strategies." + ], + "table_footnote": [], + "table_body": "
ModelsPSNR ↑SSIM ↑Controllability
Classifier28.190.8783
Visual Embedder28.240.8781
Visual Embedder‡28.470.8784
Text Embedder28.720.8821✓
", + "bbox": [ + 217, + 171, + 785, + 233 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1f727791b321d53d5c53922e0a614a583ce1ad391d27503bd9793ce9320c970b.jpg", + "table_caption": [ + "Table 5: Ablation study for different loss functions. CL and CDRL denote the ordinary contrastive loss and the proposed composite degradation restoration loss, respectively." + ], + "table_footnote": [], + "table_body": "
Smooth l1MS-SSIMCLCDRLPSNR ↑SSIM ↑
✓28.160.8633
✓✓27.540.8708
✓✓✓27.610.8723
✓✓✓28.720.8821
", + "bbox": [ + 217, + 286, + 789, + 347 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Model Control via Scene Descriptor", + "text_level": 1, + "bbox": [ + 215, + 373, + 555, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To demonstrate the controllability of the scene descriptor in our OneRestore, we conduct experiments on two intricate synthetic degradation scenarios, as depicted in Fig. 9. By manually adjusting specific text to generate corresponding degraded scene description embeddings, our model can selectively focus on different degradation factors, achieving targeted restorations. Our approach excels in controllable restoration without the need for multiple sets of weight screening, making it more concise compared to other One-to-Many methods that rely on partial parameter sharing. Meanwhile, our method can support adaptive restoration through the estimation of text embeddings based on visual attribute extraction. Fig. 10 employs the t-SNE [32] to visualize the embeddings generated by the visual embedder on the CDD-11 train and test sets. Obviously, the constructed visual embedder can clearly distinguish between different scenarios.", + "bbox": [ + 212, + 396, + 787, + 578 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 598, + 359, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our proposed OneRestore framework represents a significant step forward in image restoration, synthesizing a range of degradation patterns to accurately simulate complex scenarios. The innovative use of a scene descriptor-guided cross-attention block within a transformer-based model facilitates an adaptable and fine-grained restoration process, yielding results that surpass prior methods in both synthetic and real-world datasets. However, despite its advancements, OneRestore exhibits limitations in processing extremely high-density corruption scenarios and high-complex corruption scenarios containing unconsidered degradations, where the model's predictive capabilities can be challenged. Future work will aim to enhance the robustness of OneRestore against such extreme conditions and to reduce computational overhead, further extending its applicability and efficiency. Our findings underscore the potential of universal scene restoration, with the caveat that continued refinement is essential for tackling the full spectrum of real-world image degradation.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This project is supported by the Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097 and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bertasius, G., Torresani, L.: Cobe: Contextualized object embeddings from narrated instructional video. In: NeurIPS. pp. 15133-15145 (2020)", + "2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021)", + "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33 (2022)", + "4. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP. pp. 1-5 (2023)", + "5. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: CVPR. pp. 4196-4205 (2021)", + "6. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022)", + "7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR. pp. 5896-5905 (2023)", + "8. Chen, Z., Sun, Y., Bi, X., Yue, J.: Lightweight image de-snowing: A better trade-off between network capacity and performance. Neural Networks 165, 896-908 (2023)", + "9. Cui, Y., Ren, W., Knoll, A.: Omni-kernel network for image restoration. In: AAAI. vol. 38, pp. 1426-1434 (2024)", + "0. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: Raise: A raw images dataset for digital image forensics. In: MMSys. pp. 219-224 (2015)", + "1. Du, Y., Deng, J., Zheng, Y., Dong, J., He, S.: Dsdnet: Toward single image deraining with self-paced curricular dual stimulations. CVIU 230, 103657 (2023)", + "2. Du, Y., Han, G., Tan, Y., Xiao, C., He, S.: Blind image denoising via dynamic dual learning. IEEE TMM 23, 2139-2152 (2020)", + "3. Ferrari, V., Zisserman, A.: Learning visual attributes. In: NeurIPS. pp. 433-440 (2007)", + "4. Fu, X., Xiao, J., Zhu, Y., Liu, A., Wu, F., Zha, Z.J.: Continual image deraining with hypergraph convolutional networks. IEEE TPAMI 45(8), 9534-9551 (2023)", + "5. Girshick, R.: Fast r-cnn. In: ICCV. pp. 1440-1448 (2015)", + "6. Guo, X., Hu, Q.: Low-light image enhancement via breaking down the darkness. IJCV 131(1), 48-66 (2023)", + "7. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016)", + "8. Guo, Y., Gao, Y., Liu, W., Lu, Y., Qu, J., He, S., Ren, W.: Scenet: Self-paced semicurricular attention network for non-homogeneous image dehazing. In: CVPRW. pp. 1884-1893 (2023)" + ], + "bbox": [ + 225, + 260, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Han, J., Li, W., Fang, P., Sun, C., Hong, J., Armin, M.A., Petersson, L., Li, H.: Blind image decomposition. In: ECCV. pp. 218-237 (2022)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "21. Hoang, T., Zhang, H., Yazdani, A., Monga, V.: Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In: CVPR. pp. 1670-1679 (2023)", + "22. Jiao, J., Tu, W.C., Liu, D., He, S., Lau, R.W.H., Huang, T.S.: Formnet: Formatted learning for image restoration. IEEE TIP 29, 6302-6314 (2020)", + "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. CACM 60(6), 84-90 (2017)", + "24. Kumar, N., Berg, A., Belhumeur, P.N., Nayar, S.: Describable visual attributes for face verification and image search. IEEE TPAMI 33(10), 1962-1977 (2011)", + "25. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP 28(1), 492-505 (2018)", + "26. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022)", + "27. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR. pp. 1633-1642 (2019)", + "28. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020)", + "29. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR. pp. 2041-2050 (2018)", + "30. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP 27(6), 3064-3073 (2018)", + "31. Luo, Y., Huang, Q., Ling, J., Lin, K., Zhou, T.: Local and global knowledge distillation with direction-enhanced contrastive learning for single-image deraining. Knowledge-Based Syst. 268, 110480 (2023)", + "32. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. JMLR 9(11), 2579-2605 (2008)", + "33. Meng, Q., Shin'ichi, S.: Adinet: Attribute driven incremental network for retinal image classification. In: CVPR. pp. 4033-4042 (2020)", + "34. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: CVPR. pp. 17399-17410 (2022)", + "35. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI 45(8), 10346-10357 (2023)", + "36. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014)", + "37. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: PromptIR: Prompting for all-in-one image restoration. NeurIPS 36 (2024)", + "38. Quan, Y., Tan, X., Huang, Y., Xu, Y., Ji, H.: Image desnowing via deep invertible separation. IEEE TCSVT 33(7), 3133-3144 (2023)", + "39. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: CVPR. pp. 13658-13667 (2022)", + "40. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: ICCV. pp. 12580-12589 (2023)", + "41. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP 32, 1927-1941 (2023)", + "42. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Wang, C., Zheng, Z., Quan, R., Sun, Y., Yang, Y.: Context-aware pretraining for efficient blind image decomposition. In: CVPR. pp. 18186-18195 (2023)", + "44. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: CVPR. pp. 11197-11206 (2023)", + "45. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR. pp. 3677-3686 (2023)", + "46. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: BMVC. pp. 1-12 (2018)", + "47. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR. pp. 10551-10560 (2021)", + "48. Xu, W., Xian, Y., Wang, J., Schiele, B., Akata, Z.: Attribute prototype network for zero-shot learning. In: NeurIPS. pp. 21969-21980 (2020)", + "49. Xu, X., Wang, R., Lu, J.: Low-light image enhancement via structure modeling and guidance. In: CVPR. pp. 9893-9903 (2023)", + "50. Yan, X., Yang, J., Sohn, K., Lee, H.: Attribute2image: Conditional image generation from visual attributes. In: ECCV. pp. 776-791 (2016)", + "51. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017)", + "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022)", + "53. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV. pp. 492-511 (2020)", + "54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR. pp. 14821-14831 (2021)", + "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for fast image restoration and enhancement. IEEE TPAMI 45(2), 1934-1948 (2022)", + "56. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018)", + "57. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR. pp. 5785-5794 (2023)", + "58. Zhou, M., Huang, J., Guo, C.L., Li, C.: Fourmer: An efficient global modeling paradigm for image restoration. In: ICML. pp. 42589-42601 (2023)", + "59. Zhou, M., Wu, X., Wei, X., Xiang, T., Fang, B., Kwong, S.: Low-light enhancement method based on a retina model for structure preservation. IEEE TMM pp. 1-13 (2023)", + "60. Zhou, Y., Xu, Y., Du, Y., Wen, Q., He, S.: Propulse: Learning progressive encoders of latent semantics in gans for photo upsampling. IEEE TIP 31, 1230-1242 (2022)", + "61. Zhu, Y., Min, W., Jiang, S.: Attribute-guided feature learning for few-shot image recognition. IEEE TMM 23, 1200-1209 (2020)", + "62. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 784 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OneRestore: A Universal Restoration Framework for Composite Degradation", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_model.json b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_model.json new file mode 100644 index 0000000000000000000000000000000000000000..55353650d8984912cb2340a3c339d3a5225edaf1 --- /dev/null +++ b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_model.json @@ -0,0 +1,2533 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.222, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.212, + 0.763, + 0.245 + ], + "angle": 0, + "content": "Yu Guo\\(^{1,2,\\dagger}\\), Yuan Gao\\(^{1,\\dagger}\\), Yuxu Lu\\(^{3}\\), Huilin Zhu\\(^{1,2}\\), Ryan Wen Liu\\(^{1(\\boxtimes)}\\), and Shengfeng He\\(^{2(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.255, + 0.671, + 0.282 + ], + "angle": 0, + "content": "1 Wuhan University of Technology, Wuhan, China \nwenliu@whut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.283, + 0.662, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) Singapore Management University, Singapore shengfenghe@smu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.311, + 0.706, + 0.354 + ], + "angle": 0, + "content": "3 The Hong Kong Polytechnic University, Hong Kong, China \n† Equal Contribution \nhttps://github.com/gy65896/OneRestore" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.386, + 0.744, + 0.651 + ], + "angle": 0, + "content": "Abstract. In real-world scenarios, image impairments often manifest as composite degradations, presenting a complex interplay of elements such as low light, haze, rain, and snow. Despite this reality, existing restoration methods typically target isolated degradation types, thereby falling short in environments where multiple degrading factors coexist. To bridge this gap, our study proposes a versatile imaging model that consolidates four physical corruption paradigms to accurately represent complex, composite degradation scenarios. In this context, we propose OneRestore, a novel transformer-based framework designed for adaptive, controllable scene restoration. The proposed framework leverages a unique cross-attention mechanism, merging degraded scene descriptors with image features, allowing for nuanced restoration. Our model allows versatile input scene descriptors, ranging from manual text embeddings to automatic extractions based on visual attributes. Our methodology is further enhanced through a composite degradation restoration loss, using extra degraded images as negative samples to fortify model constraints. Comparative results on synthetic and real-world datasets demonstrate OneRestore as a superior solution, significantly advancing the state-of-the-art in addressing complex, composite degradations." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Image restoration is an essential task that entails the recovery of high-quality visuals from compromised inputs. The fidelity of these restored images is vital for the subsequent downstream tasks, such as robot navigation and autonomous driving. Considerable progress has been made in addressing single degradation scenarios by a single model (i.e., One-to-One models, see Fig. 2a), such as low-light conditions [16, 49, 59], haze [21, 41, 57], rain [11, 14, 31], and snow [4, 8, 38]. These methods have achieved remarkable success within their intended contexts. Nevertheless, real-world conditions often involve unpredictable and variable composite degradations that can severely impact image quality and clarity [28]. The" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.788, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 1: Our OneRestore allows fully controllable image restoration using scene descriptors derived from both automatic visual attribute extraction (top) and manual text embeddings (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.502 + ], + "angle": 0, + "content": "occurrence of mixed degrading conditions in one image can lead to intricate interactions, posing significant challenges to One-to-One solutions [22,52,54,60]. The necessity for seamless switching and decision-making across specialized methods for each condition proves to be inefficient and ineffective in practical applications. Therefore, there is a growing imperative for robust, all-encompassing image restoration approaches that can universally address multiple forms of degradation through a singular, dynamic framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Recent research has been directed toward One-to-Many models, which address various individual degradation factors within a single framework. These models are categorized into: those with partial parameter-sharing (One-to-Many\\(_p\\)) [2,19,28,45] and those with full parameter-sharing (One-to-Many\\(_f\\)) [6,26,35,42, 43]. One-to-Many\\(_p\\) (left of Fig. 2b) restoration methods typically employ models with either multiple heads and tails or a single tail to address various types of image impairments. However, the complexity and size of these models tend to scale with the diversity of degradations they aim to correct, leading to increased training costs and parameter counts. Furthermore, when images suffer from a combination of degrading factors, these models may underperform. Some One-to-Many\\(_p\\) techniques also depend on user intervention to select appropriate restoration weights, which is impractical for automated or dynamic settings. On the other hand, One-to-Many\\(_f\\) (right of Fig. 2b) approaches train models using a direct learning strategy, where inputs are images with different individual degradations and the targets are their clear counterparts. However, this direct training approach is designed predominantly for single-category degradation issues, it does not adequately address the complex dynamics that arise when multiple types of impairments affect an image simultaneously. As a result, such methods can falter in practical applications where user control and precise directionality of restoration are crucial. For instance, in cases where users seek to improve low-light imagery, the model might inadvertently prioritize dehazing actions, deviating from user intent." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.373, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.263, + 0.337, + 0.274 + ], + "angle": 0, + "content": "(a) One-to-One" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.146, + 0.59, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.263, + 0.527, + 0.275 + ], + "angle": 0, + "content": "(b) One-to-Many" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.146, + 0.785, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.263, + 0.746, + 0.275 + ], + "angle": 0, + "content": "(c) One-to-Composite" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.285, + 0.784, + 0.313 + ], + "angle": 0, + "content": "Fig.2: Structural comparison between One-to-One, One-to-Many, and our One-to-Composite image restoration methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.385 + ], + "angle": 0, + "content": "Drawing from the preceding discussion, we delineate the primary challenges in developing image restoration models for composite degradations in real-world conditions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.422 + ], + "angle": 0, + "content": "- How can the model be designed to adaptively discern the specific degradation afflicting the input image?" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.422, + 0.785, + 0.451 + ], + "angle": 0, + "content": "- What approaches enable a model to be trained efficiently across a spectrum of degradations while enhancing user control over the restoration process?" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.481 + ], + "angle": 0, + "content": "- How can the recovery process be optimized that not only approach the quality of clear images but also maintain distinction from various degraded forms?" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.699 + ], + "angle": 0, + "content": "In response to the challenges outlined, we present a unified model that builds upon established physical corruption models for degradations such as low light, haze, rain, and snow. This model effectively captures the complexity of real-world degradation conditions and forms the basis for constructing our Composite Degradation Dataset (CDD-11), which serves both training and evaluation purposes. To address the issue of composite degradation, our work introduces OneRestore, a transformer-based universal restoration architecture. This novel framework distinguishes itself from conventional methods by embodying a Oneto-Composite strategy, as depicted in Fig. 2c, which allows for precise and controllable outcomes (see Fig. 1). It employs an innovative cross-attention block that fuses scene descriptors, which can be manual text embeddings or derived automatically from visual attributes. We further enhance the model's robustness by integrating a composite degradation restoration loss, using additional degraded images as negative samples to establish stringent lower-bound constraints." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.7, + 0.598, + 0.714 + ], + "angle": 0, + "content": "Our key contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.72, + 0.785, + 0.779 + ], + "angle": 0, + "content": "- We present the first attempt to unify composite degradations, creating a versatile imaging model that simulates a range of degradation types. This innovation underpins the development of our comprehensive Composite Degradation Dataset." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.785, + 0.839 + ], + "angle": 0, + "content": "- We propose a universal framework that introduces a novel level of controllability in image restoration, employing a cross-attention mechanism that integrates scene descriptors derived from either manual text embeddings or automated visual attribute extractions." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.72, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "- We tailor a composite degradation restoration loss, enhancing the model's ability to discern between various forms of image degradation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.22 + ], + "angle": 0, + "content": "- Our universal model not only sets a new state-of-the-art for composite dataset performance but also competes favorably with One-to-One models in single-degradation scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.24, + 0.388, + 0.256 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Complex Degradation Image Restoration. Recent advancements in image restoration have introduced a variety of solutions for complexly degraded images. A notable direction is the partial parameter-sharing One-to-Many approach, employing architectures with multiple heads and tails, and a single backbone, to address various degradation factors. For instance, Li et al. [28] develop an all-weather network that restores scenes affected by haze, rain, and snow, using specialized encoders for each condition and a shared decoder. Chen et al. [2] introduce a transformer-based model for deraining, denoising, and super-resolving images, leveraging a multi-head and multi-tail design. Wang et al. [45] present a knowledge assignment strategy to differentiate and process rain and snow degradations adaptively. While these models represent strides in complex degradation image restoration, they stop short of being truly universal One-to-Many networks. With only partial weight sharing, they address different degradation types through distinct network pathways." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Conversely, universal image restoration has emerged with the One-to-Many methods sharing all network weights, thereby uniformly mitigating multiple degradation types. Examples include the unified two-stage knowledge learning framework by Chen et al. [6], which efficiently processes dehazing, deraining, and desnoring with a singular pretrained weight set. Li et al. [26] present a network capable of comprehensive image restoration without the need for pre-identifying degradation levels. Additionally, Ozdenizci et al. [35] introduce a conditional denoising diffusion model tailored for patch-based image restoration under varied weather conditions. However, due to the significant differences and mutual interference between various degradation factors, this type of method can only solve specific types of degradation and cannot achieve image restoration in composite degradation scenarios. To this end, we propose a universal One-to-Composite image restoration method that allows the model to adapt to arbitrary combinations of composite degradations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Visual Attribute Control. Visual attributes are pivotal in defining image features and enhancing the interpretability of models. They enrich feature representations and provide a hierarchical framework for image analysis. Early research highlighted visual attributes' transformative impact on image generation, recognition, and retrieval [13, 24, 50]. More recent studies underscore their utility in improving object recognition and detection, particularly in data-scarce environments [1, 33, 48, 61]. Visual attributes have also been leveraged for zero-shot learning applications, advancing classification tasks beyond the constraints of labeled datasets [39, 44]. In our study, we harness visual attributes to delineate the intricate degradation patterns present in real-world imaging conditions," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.239 + ], + "angle": 0, + "content": "enhancing the pertinence and efficacy of image restoration techniques. It is noteworthy that conventional discriminators, when tasked with identifying image degradation, are typically confined to image data, precluding autonomous feature control. Our approach, by contrast, accommodates both images and text descriptions as inputs, thereby facilitating the automatic or manual determination of degradation characteristics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.607, + 0.278 + ], + "angle": 0, + "content": "3 Composite Degradation Formulation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.291, + 0.436, + 0.306 + ], + "angle": 0, + "content": "3.1 Imaging Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.435 + ], + "angle": 0, + "content": "In real-world scenarios, images are often affected by multiple degrading factors simultaneously, resulting in suboptimal performance from restoration methods designed for isolated degradations. To address the broad spectrum of degradation scenarios, we introduce a versatile imaging model that encapsulates all degradations, including low light, haze, rain, and snow. This model aims to narrow the disparity between simulated and authentic real-world data. Given a degraded image \\( I(x) \\) and its corresponding clear state \\( J(x) \\), the proposed imaging model is mathematically articulated as" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.447, + 0.786, + 0.463 + ], + "angle": 0, + "content": "\\[\nI (x) = \\mathcal {P} _ {h} \\left(\\mathcal {P} _ {r s} \\left(\\mathcal {P} _ {l} (J (x)))\\right), \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.787, + 0.532 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_l, \\mathcal{P}_{rs}\\), and \\(\\mathcal{P}_h\\) denote the low light, rain/snow, and haze degradations, respectively. Given that concurrent rain and snow conditions are infrequent, they are not addressed within the scope of this work. A comprehensive exposition of the proposed imaging model is detailed below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.786, + 0.564 + ], + "angle": 0, + "content": "Low-Light Conditions. Inspired by Retinex theory, low-light degraded image \\( I_{l}(x) \\) can be generated by" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.574, + 0.786, + 0.606 + ], + "angle": 0, + "content": "\\[\nI _ {l} (x) = \\mathcal {P} _ {l} (J (x)) = \\frac {J (x)}{L (x)} L (x) ^ {\\gamma} + \\varepsilon . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.784, + 0.675 + ], + "angle": 0, + "content": "Here, we use a darkening coefficient \\(\\gamma \\in [2,3]\\) as the brightness controller to adjust the illumination map \\(L(x)\\) generated via LIME [17]. To simulate poor light conditions realistically, we add the Gaussian noise \\(\\varepsilon\\) with mean \\(= 0\\) and variance \\(\\in [0.03,0.08]\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.786, + 0.737 + ], + "angle": 0, + "content": "Rain/Snow Streaks. Following the methodologies outlined in [5, 27], our process superimposes rain or snow streaks onto the image before synthesizing haze. To introduce rain streaks \\(\\mathcal{R}\\), we produce the rain-degraded image \\(I_{rs}(x)\\) by overlaying \\(\\mathcal{P}_{rs}\\) onto the low-light image \\(I_l(x)\\), which is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.748, + 0.786, + 0.763 + ], + "angle": 0, + "content": "\\[\nI _ {r s} (x) = \\mathcal {P} _ {r s} \\left(I _ {l} (x)\\right) = I _ {l} (x) + \\mathcal {R}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.773, + 0.64, + 0.789 + ], + "angle": 0, + "content": "For snow streaks \\(S\\), the snow-degraded image is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.801, + 0.786, + 0.815 + ], + "angle": 0, + "content": "\\[\nI _ {r s} (x) = \\mathcal {P} _ {r s} \\left(I _ {l} (x)\\right) = I _ {l} (x) (1 - \\mathcal {S}) + M (x) \\mathcal {S}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.565, + 0.841 + ], + "angle": 0, + "content": "with \\( M(x) \\) being the chromatic aberration map." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.487, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.288, + 0.404, + 0.3 + ], + "angle": 0, + "content": "(a) Overall Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.143, + 0.691, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.289, + 0.675, + 0.3 + ], + "angle": 0, + "content": "(b) Scene Descriptor Generation" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.144, + 0.787, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.288, + 0.765, + 0.3 + ], + "angle": 0, + "content": "(c)SDTB" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 3: Architecture of our OneRestore. (a) Overall pipeline, where 32, 64, 128, and 256 represent the number of channels. (b) Scene descriptor generation, where scene descriptors are fed into each (c) Scene Descriptor-guided Transformer Block (SDTB) by manual text embeddings or automatic extraction based on visual attributes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.789, + 0.429 + ], + "angle": 0, + "content": "Haze Degradations. We incorporate haze degradation by employing the atmospheric scattering model as" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.441, + 0.786, + 0.459 + ], + "angle": 0, + "content": "\\[\nI (x) = \\mathcal {P} _ {h} \\left(I _ {r s} (x)\\right) = I _ {r s} (x) t + A (1 - t), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.561 + ], + "angle": 0, + "content": "where \\( t \\) denotes the transmission map and \\( A \\) represents the atmospheric light. The transmission map \\( t \\) is defined by the exponential decay of light, \\( t = e^{-\\beta d(x)} \\), where \\( \\beta \\) represents the haze density coefficient and \\( d(x) \\) is the scene depth (estimated by MegaDepth [29]). The haze density coefficient \\( \\beta \\) is varied within the range of [1.0, 2.0] and the atmospheric light \\( A \\) is constrained to [0.6, 0.9] for realistic simulation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.586, + 0.528, + 0.602 + ], + "angle": 0, + "content": "3.2 Composite Degradation Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.79, + 0.734 + ], + "angle": 0, + "content": "Drawing on Eq. (1), we have developed the Composite Degradation Dataset (CDD-11), encompassing 11 categories of image degradations and their clear counterparts. These degraded samples include low (low-light), haze, rain, snow, low+haze, low+rain, low+snow, haze+rain, haze+snow, low+haze+rain, and low+haze+snow. From the RAISE database [10], we selected 1,383 high-resolution clear images for producing 11 composite degradations. They are resized to a uniform resolution of \\(1080 \\times 720\\). The overall dataset is split into 13,013 image pairs for training and 2,200 for testing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.761, + 0.367, + 0.777 + ], + "angle": 0, + "content": "4 OneRestore" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The encoder-decoder network, validated as an effective architectural paradigm for image restoration [7, 12, 18], serves as the overall structure for our model, as depicted in Fig. 3a. Building upon the encoder-decoder structure, we devise" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "an innovative Scene Descriptor-guided Transformer Block (SDTB) as the basic unit for feature extraction (Fig. 3c). Our approach diverges from existing transformer-based methods [4,7,41] by concurrently integrating image features and scene descriptors into the SDTB basic block for supporting fine-grained image restoration. In particular, the generation of scene descriptors ranges from manual text embeddings to automatic extraction based on visual attributes, as illustrated in Fig. 3b. Further details on scene descriptor generation and SDTB are elaborated in Sections 4.1 and 4.2. For capturing multi-scale representations, the encoder employs three downsampling operations, mirrored by the decoder's upsampling operations. To be specific, we utilize \"Max-Pooling + Conv\" and \"Bilinear Interpolation + Conv\" for downsampling and upsampling, respectively. Simultaneously, three skip connections and a global residual enhance model convergence by bridging shallow and deep features. Additional details regarding the configuration of our OneRestore can be found in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.359, + 0.735, + 0.373 + ], + "angle": 0, + "content": "Ultimately, the overall loss \\(\\mathcal{L}\\) for model training is defined as follows" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.381, + 0.785, + 0.399 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {1} ^ {s} (J, \\hat {J}) + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {M}} (J, \\hat {J}) + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {c}} (J, \\hat {J}, I, \\{I _ {o} \\}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.406, + 0.788, + 0.482 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_1^s\\), \\(\\mathcal{L}_{\\mathrm{M}}\\), and \\(\\mathcal{L}_{\\mathrm{c}}\\) represent the smooth \\(l_{1}\\) loss [15], Multi-Scale Structural Similarity (MS-SSIM) loss, and the proposed composite degradation restoration loss (see Sec. 4.3), respectively, \\(\\alpha_{1-3}\\) denote the penalty coefficients, \\(J\\), \\(\\hat{J}\\), \\(I\\), and \\(\\{I_o\\}\\) are the positive, model output anchor, model input negative, and other negatives, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.502, + 0.502, + 0.517 + ], + "angle": 0, + "content": "4.1 Scene Descriptor Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.689 + ], + "angle": 0, + "content": "To augment the controllability of our OneRestore, we introduce the scene description embedding as an additional input to the model, alongside the degraded image. As delineated in Fig. 3b, the generation of the scene descriptor offers two selectable modes: the manual mode \\(①\\) and the automatic mode \\(②\\). In the manual mode, user involvement is necessitated to supply scene description text, generating the corresponding text embedding employed as an input. Conversely, the automatic mode involves the extraction of visual attributes from the image to generate visual embedding, estimating the most proximate text embedding. To quantify the difference between visual and text embeddings and train optimal text and visual embedders, we introduce cosine cross-entropy loss for model weight optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Text Embedder. For the text embedding generation task, we employ a set of 5 scene description texts as input to generate 12 text embeddings. Initially, GloVe [36] is leveraged to generate initial text embeddings for 1 clear scene and 4 scenes with single degradation types. Subsequently, 7 additional composite degradation text embeddings (i.e., low+ haze, low+ rain, low+ snow, haze+ rain, haze+ snow, low+ haze+ rain, and low+ haze+ snow) are generated by averaging the element values in the corresponding text embeddings. A Multi-Layer Perception (MLP) is then employed to refine the 12 text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Visual Embedder. Taking inspiration from the work of Saini et al. [39], we first resize the input image to \\(224 \\times 224\\) and extract image features using a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "ResNet-18 [20] network without average pooling. The initial weights of ResNet-18 are pre-trained on the ImageNet dataset. Subsequently, a single convolutional layer, followed by a dropout layer and a linear layer, is introduced to generate the final visual embedding. By calculating the cosine similarity between visual embeddings and all text embeddings, and selecting the one with the highest similarity, we can automatically generate scene descriptors to achieve image restoration based on visual attribute control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Cosine Cross-Entropy. Similar to [39], the score of visual embedding and each text embedding is obtained by \"Cosine Similarity + Softmax\". Given \\( e_v \\) and \\( e_t \\) being the visual and text embeddings, the similarity score \\( S(e_v, e_t) \\) can be given by" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.324, + 0.785, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\cos \\left(e _ {v}, e _ {t}\\right) = \\delta \\cdot \\frac {e _ {v} \\cdot e _ {t} ^ {\\top}}{\\| e _ {v} \\| \\| e _ {t} \\|}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.356, + 0.785, + 0.398 + ], + "angle": 0, + "content": "\\[\nS \\left(e _ {v}, e _ {t}\\right) = \\frac {e ^ {\\cos \\left(e _ {v} , e _ {t}\\right)}}{\\sum_ {t _ {i} = 1} ^ {N _ {t}} e ^ {\\cos \\left(e _ {v} , e _ {t _ {i}}\\right)}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.784, + 0.454 + ], + "angle": 0, + "content": "with \\(\\delta\\) and \\(N_{t}\\) being the temperature factor and the number of text embeddings, respectively. Based on the calculated \\(S(e_v, e_t)\\), we use the cross-entropy loss for model training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.477, + 0.622, + 0.492 + ], + "angle": 0, + "content": "4.2 Scene Descriptor-guided Transformer Block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.785, + 0.637 + ], + "angle": 0, + "content": "To further harness the capabilities of scene description embeddings in image restoration tasks and facilitate a more adaptive and nuanced restoration process, we introduce the innovative Scene Descriptor-guided Transformer Block (SDTB), illustrated in Fig. 3c. The SDTB consists of three primary components, namely Scene Descriptor-guided Cross-Attention (SDCA), Self-Attention (SA), and Feed-Forward Network (FFN). Diverging from traditional self-attention mechanisms, the proposed SDCA module utilizes a query generated from scene descriptors instead of queries generated from image features. The mathematical representation of our SDCA can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.649, + 0.785, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S D C A} \\left(\\mathbf {Q} _ {t}, \\mathbf {K}, \\mathbf {V}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t} \\cdot \\mathbf {K} ^ {\\top}}{\\lambda}\\right) \\mathbf {V}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.841 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a temperature factor, \\(\\mathbf{Q}_t\\) represents the scene descriptor-generated query matrix, \\(\\mathbf{K}\\) and \\(\\mathbf{V}\\) denote the image feature-generated key and value matrices, respectively. \\(\\mathbf{Q}_t\\) is obtained from the input scene description embedding \\(e_t\\) via a linear layer. During the process of matrix multiplication between \\(\\mathbf{Q}_t\\) and \\(\\mathbf{K}\\). To ensure consistency in the number of tokens between \\(\\mathbf{K}\\) and \\(\\mathbf{Q}_t\\), we resize the original image before performing the reshaping operation to generate \\(\\mathbf{K}\\). By incorporating the SDCA module, each transformer block in OneRestore can effectively integrate scene descriptors and image features for targeted image restoration. As for the SA and FFN modules, we adopt structures akin to those in Restormer [52]. More details of our SDTB are introduced in the supplement." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.273, + 0.719, + 0.288 + ], + "angle": 0, + "content": "Fig. 4: Illustration of the composite degradation restoration loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.318, + 0.603, + 0.334 + ], + "angle": 0, + "content": "4.3 Composite Degradation Restoration Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Using contrastive loss has been demonstrated as an effective way of improving image restoration performance [47, 57]. Essentially, contrastive loss views the degraded input image as a negative sample, the clear image as a positive sample, and the restored image as an anchor. This loss function works to minimize the distance between the anchor and the positive sample while maximizing its distance from the negative sample. In the composite degradation image restoration task, the traditional contrastive loss, however, may inadvertently reduce the distance between the anchor and other degraded images. To make the anchor closer to the positive sample while maintaining a clear separation from all degraded negatives, we design a composite degradation restoration loss, illustrated in Fig. 4. Considering \\( I \\), \\( I_o \\), \\( J \\), and \\( \\hat{J} \\) as the input negative, other negatives, positive, and anchor, the proposed loss can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.259, + 0.523, + 0.786, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} (J, \\hat {J}, I, \\left\\{I _ {o} \\right\\}) = \\sum_ {k = 1} ^ {K} \\xi_ {k} \\frac {\\mathcal {L} _ {1} \\left(V _ {k} (\\hat {J}) , V _ {k} (\\hat {J})\\right)}{\\xi_ {c} \\mathcal {L} _ {1} \\left(V _ {k} (\\hat {J}) , V _ {k} (I)\\right) + \\sum_ {o = 1} ^ {O} \\xi_ {o} \\mathcal {L} _ {1} \\left(V _ {k} (I _ {o}), V _ {k} (\\hat {J})\\right)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.723 + ], + "angle": 0, + "content": "where \\( V_{k}(\\cdot) \\) denotes the extraction operation of \\( k \\)-th hidden feature from the VGG-16 pre-trained on the ImageNet, \\( \\mathcal{L}_1 \\) represents the \\( l_1 \\) loss, \\( K \\) and \\( O \\) are the numbers of used VGG-16 feature layers and other degraded negatives, \\( \\xi_{k}, \\xi_{c} \\), and \\( \\xi_{o} \\) denote the hyper-parameters, respectively. In this work, the number of other degraded negatives \\( O \\) is 10. We use the output features of 3-rd, 8-th, and 15-th layers from VGG-16 and set \\( K = 3 \\). Furthermore, we consider that the features of each feature layer of each degraded negative have equal importance. Therefore, we set \\( \\xi_{c} = \\xi_{o} = \\frac{1}{11} \\) and \\( \\xi_{k} = \\frac{1}{3} \\). According to our ablation study, the proposed loss confers a significant advantage over using a single input degradation as a negative sample, which can easily capture positive sample features." + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.748, + 0.725, + 0.763 + ], + "angle": 0, + "content": "Table 1: Settings of model training. \"lr\" denotes the learning rate." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.775, + 0.787, + 0.838 + ], + "angle": 0, + "content": "
TermsText/Visual EmbedderOneRestore
Epoch200120
Initial lr0.00010.0002
lr Decay0.5/50 epochs0.5/20 epochs
Batch Size2564
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Comparison of quantitative results on CDD-11 dataset. OneRestore† means using the corresponding scene description text as additional input, and WGWSNet requires the scene type as model input. Red, green, and blue indicate the best, second-best, and third-best results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.214, + 0.79, + 0.407 + ], + "angle": 0, + "content": "
TypesMethodsVenue & YearPSNR ↑SSIM ↑#Params
Input16.000.6008-
One-to-OneMIRNet [53]ECCV202025.970.847431.79M
MPRNet [54]CVPR202125.470.855515.74M
MIRNetv2 [55]TPAMI202225.370.83355.86M
Restormer [52]CVPR202226.990.864626.13M
DGUNet [34]CVPR202226.920.855917.33M
NAFNet [3]ECCV202224.130.796417.11M
SRUDC [40]ICCV202327.640.86006.80M
Fourmer [58]ICML202323.440.78850.55M
OKNet [9]AAAI202426.330.86054.72M
One-to-ManyAirNet [26]CVPR202223.750.81408.93M
TransWeather [42]CVPR202223.130.781021.90M
WeatherDiff [35]TPAMI202322.490.798582.96M
PromptIR [37]NIPS202325.900.849938.45M
WGWSNet [62]CVPR202326.960.862625.76M
One-to-CompositeOneRestore28.470.87845.98M
OneRestore†28.720.88215.98M
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.435, + 0.377, + 0.453 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.434, + 0.484 + ], + "angle": 0, + "content": "5.1 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Implementation Details. Our OneRestore is implemented by PyTorch 1.12.0 and trained on a PC with 2 AMD EPYC 7543 32-Core Processors and 8 NVIDIA L40 GPUs. We use the Adam with exponential decay rates being \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\) for optimization. The training parameter settings for text/visual embedder and OneRestore are shown in Table 1. In particular, all images used for training are cropped into \\(256\\times 256\\) image patches with a sampling stride of 200 and randomly flipped by 0, 90, 180, and 270 degrees to generate \\(312\\mathrm{k}\\) patch pairs. More details about training and inference can be found in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Evaluation Datasets. To facilitate a thorough comparison, we execute image restoration experiments across various datasets. These include the Composite Degradation Dataset (CDD-11) crafted by us, established One-to-One benchmarks, and authentic real-world datasets. The standard One-to-One benchmarks are designated for four principal tasks: the LOw-Light dataset (LOL) [46], the REalistic Single Image DEhazing Outdoor Training Set (RESIDE-OTS) [25], the Rain1200 dataset [56], and the Snow100k dataset [30]. Owing to the dearth of specialized datasets for real-world composite degradation, we have curated a selection from the LIME low-light enhancement dataset [17], the RESIDE Realworld Task-driven Testing Set (RESIDE-RTTS) [25], Yang's rainy dataset [51], and the realistically snowy Snow100k dataset [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Competitors and Evaluation Metrics. We compare our OneRestore with 9 One-to-One image restoration methods (MIRNet [53], MPRNet [54], MIRNetv2 [55], Restormer [52], DGUNet [34], NAFNet [3], SRUDC [40], Fourmer [58], and OKNet [9]) and 5 One-to-Many image restoration methods (AirNet [26]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "TransWeather [42], WeatherDiff [35], PromptIR [37], and WGWSNet [62]). In particular, all compared state-of-the-art methods are retrained on our CDD-11 train set. For PromptIR and WGWSNet, we adjust the number of prompts and the weather-specific parameters of each layer for corresponding to our task, respectively. Furthermore, we adopt the Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM) as evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.258, + 0.469, + 0.274 + ], + "angle": 0, + "content": "5.2 Comparison with SOTAs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.541, + 0.508 + ], + "angle": 0, + "content": "The quantitative evaluation results on our CDD-11 test dataset are displayed in Table 2. OneRestore† with text embedder achieves the highest performance, followed by OneRestore with visual embedder as the second-best. Both OneRestore† and OneRestore demonstrate substantial leads over the third-ranked methods in terms of both PSNR and SSIM. Further detailed restoration performance comparison of our method and the One-to-Many methods in each degradation scenario is illustrated in Fig. 5 using radar charts. Notably, image restoration for composite degradation scenarios proves challenging. By incorporating scene description em" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.313, + 0.786, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.435, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig.5: Comparison of quantitative results for different degradation scenarios on CDD-11 dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.629 + ], + "angle": 0, + "content": "beddings into image restoration models and utilizing the proposed composite degradation restoration loss, our method robustly reconstructs clear images across various situations. For compelling evidence, Fig. 6 provides the visual comparison of image restoration in two composite degradation samples. It is evident that SRUDC, OKNet, AirNet, and WGWSNet are unable to suppress all degenerative interferences, while Restormer and PromptIR are prone to color distortion and noise residual. In contrast, the proposed OneRestore can produce more natural results and fully preserve image textures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Moreover, Fig. 7 displays the image restoration results in two real-world scenarios. Compared with existing One-to-Many image restoration methods [26,35, 37], the proposed OneRestore can simultaneously focus on multiple degradation factors and robustly reconstruct scene details. More results on CDD-11, realworld scenarios, and other One-to-One benchmarks of low-light, hazy, rainy, and snowy scenes are included in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.391, + 0.757 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Effectiveness of Network Modules. Table 3 reports the quantitative evaluation of different module configurations. While the Feed-Forward Network (FFN) with convolution as the basic unit serves as a crucial module in the transformer network, it alone fails to generate satisfactory reconstruction results. The ordinary Self-Attention module (SA) significantly strengthens network performance," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.787, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.283, + 0.789, + 0.311 + ], + "angle": 0, + "content": "Fig. 6: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) synthetic samples." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.329, + 0.787, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.43, + 0.789, + 0.473 + ], + "angle": 0, + "content": "Fig. 7: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) samples in real-world scenarios. * represents the utilization of original weights published in the author's code." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.612 + ], + "angle": 0, + "content": "but this method regrets that the model fails to be flexibly controlled. Compared to depending only on SA, relying exclusively on Scene Description-guided Cross-Attention (SDCA) utilizing scene description embedding yields better outcomes. This further confirms that addressing complex composite degradations can be done more effectively and flexibly by integrating degradation scene descriptors into the model. Ultimately, the best performance can be achieved by using a combination of SDCA, SA, and FFN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.54, + 0.842 + ], + "angle": 0, + "content": "Selection of Description Embedding Strategies. In this section, we compare the image restoration performance by using different strategies to generate scene descriptor, including Classifier (utilizing the embedding before the last linear layer of the retrained AlexNet [23] as the scene descriptor), Visual Embedder (employing visual embedding for model training), Visual Embedder‡ (deriving text embedding via visual embedder), and Text Embedder (employing text embedding for model training). The quantitative evaluation results are presented in Table 4. It is observed that employing Classifier and Visual Embedder to generate scene descriptors" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.645, + 0.783, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.762, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Fig.8: Comparison of image restoration on different description embedding strategies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Ablation study for different model configurations. SDCA, SA, and FFN denote the scene description-guided cross-attention, self-attention, and feed-forward network, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.2, + 0.787, + 0.26 + ], + "angle": 0, + "content": "
SDCASAFFNPSNR ↑SSIM ↑Controllability
✓24.810.8607
✓✓27.190.8697
✓✓27.930.8767✓
✓✓✓28.720.8821✓
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.28, + 0.787, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.363, + 0.79, + 0.392 + ], + "angle": 0, + "content": "Fig. 9: Comparison of image restoration on \\( low + haze + rain \\) (top) and \\( low + haze + snow \\) (bottom) synthetic samples by using different texts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.789, + 0.64 + ], + "angle": 0, + "content": "yields comparable outcomes. However, these approaches have a drawback in that each image has a distinct embedding as input, making it challenging for the model to recognize the cause of degradation. Fig. 8 displays a classic case comparison using different embedding strategies, where the abnormal artifacts of the results produced by Classifier and Visual Embedder are caused by the model's fuzzy identification of degradation factors. Moreover, the restoration model lacks controllability. The discrepancy between Visual Embedder\\(\\ddagger\\) and Text Embedder stems from errors in scene descriptor estimation. Visual Embedder\\(\\ddagger\\) has an accuracy of \\(97.55\\%\\) on the test dataset, with misestimation occurring when some degradation types are insignificant. For instance, a haze+snow scene may be evaluated as a haze or snow scene. In contrast, the text embedding-based approach is superior in describing degradation scenes and recovering images, using a fixed number of scene descriptors. This allows the model to employ degraded scene descriptors as switches to achieve controlled restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.559, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of Loss Functions. Table 5 presents a performance comparison of different loss functions. Notably, the integration of the MS-SSIM loss is instrumental in elevating SSIM scores at the cost of reduced PSNR performance. In contrast, the CL loss demonstrates minimal impact, providing only marginal improvements in SSIM and PSNR. Upon replacing CL with CDRL, a substantial enhancement is observed in both PSNR and SSIM. The proposed CDRL loss yields an output that closely resembles the clear image while avoiding other forms of degradation." + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.675, + 0.789, + 0.777 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.786, + 0.788, + 0.814 + ], + "angle": 0, + "content": "Fig.10: Visualization of visual embeddings based on t-SNE." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.292, + 0.145, + 0.71, + 0.16 + ], + "angle": 0, + "content": "Table 4: Ablation study for description embedding strategies." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.172, + 0.787, + 0.234 + ], + "angle": 0, + "content": "
ModelsPSNR ↑SSIM ↑Controllability
Classifier28.190.8783
Visual Embedder28.240.8781
Visual Embedder‡28.470.8784
Text Embedder28.720.8821✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.246, + 0.785, + 0.274 + ], + "angle": 0, + "content": "Table 5: Ablation study for different loss functions. CL and CDRL denote the ordinary contrastive loss and the proposed composite degradation restoration loss, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.287, + 0.79, + 0.348 + ], + "angle": 0, + "content": "
Smooth l1MS-SSIMCLCDRLPSNR ↑SSIM ↑
✓28.160.8633
✓✓27.540.8708
✓✓✓27.610.8723
✓✓✓28.720.8821
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.374, + 0.557, + 0.389 + ], + "angle": 0, + "content": "5.4 Model Control via Scene Descriptor" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.579 + ], + "angle": 0, + "content": "To demonstrate the controllability of the scene descriptor in our OneRestore, we conduct experiments on two intricate synthetic degradation scenarios, as depicted in Fig. 9. By manually adjusting specific text to generate corresponding degraded scene description embeddings, our model can selectively focus on different degradation factors, achieving targeted restorations. Our approach excels in controllable restoration without the need for multiple sets of weight screening, making it more concise compared to other One-to-Many methods that rely on partial parameter sharing. Meanwhile, our method can support adaptive restoration through the estimation of text embeddings based on visual attribute extraction. Fig. 10 employs the t-SNE [32] to visualize the embeddings generated by the visual embedder on the CDD-11 train and test sets. Obviously, the constructed visual embedder can clearly distinguish between different scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.36, + 0.615 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our proposed OneRestore framework represents a significant step forward in image restoration, synthesizing a range of degradation patterns to accurately simulate complex scenarios. The innovative use of a scene descriptor-guided cross-attention block within a transformer-based model facilitates an adaptable and fine-grained restoration process, yielding results that surpass prior methods in both synthetic and real-world datasets. However, despite its advancements, OneRestore exhibits limitations in processing extremely high-density corruption scenarios and high-complex corruption scenarios containing unconsidered degradations, where the model's predictive capabilities can be challenged. Future work will aim to enhance the robustness of OneRestore against such extreme conditions and to reduce computational overhead, further extending its applicability and efficiency. Our findings underscore the potential of universal scene restoration, with the caveat that continued refinement is essential for tackling the full spectrum of real-world image degradation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgement. This project is supported by the Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097 and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.288 + ], + "angle": 0, + "content": "1. Bertasius, G., Torresani, L.: Cobe: Contextualized object embeddings from narrated instructional video. In: NeurIPS. pp. 15133-15145 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.328 + ], + "angle": 0, + "content": "2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.357 + ], + "angle": 0, + "content": "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.785, + 0.398 + ], + "angle": 0, + "content": "4. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP. pp. 1-5 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.4, + 0.785, + 0.453 + ], + "angle": 0, + "content": "5. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: CVPR. pp. 4196-4205 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.509 + ], + "angle": 0, + "content": "6. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.536 + ], + "angle": 0, + "content": "7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR. pp. 5896-5905 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.564 + ], + "angle": 0, + "content": "8. Chen, Z., Sun, Y., Bi, X., Yue, J.: Lightweight image de-snowing: A better trade-off between network capacity and performance. Neural Networks 165, 896-908 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.785, + 0.592 + ], + "angle": 0, + "content": "9. Cui, Y., Ren, W., Knoll, A.: Omni-kernel network for image restoration. In: AAAI. vol. 38, pp. 1426-1434 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.62 + ], + "angle": 0, + "content": "0. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: Raise: A raw images dataset for digital image forensics. In: MMSys. pp. 219-224 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.646 + ], + "angle": 0, + "content": "1. Du, Y., Deng, J., Zheng, Y., Dong, J., He, S.: Dsdnet: Toward single image deraining with self-paced curricular dual stimulations. CVIU 230, 103657 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.674 + ], + "angle": 0, + "content": "2. Du, Y., Han, G., Tan, Y., Xiao, C., He, S.: Blind image denoising via dynamic dual learning. IEEE TMM 23, 2139-2152 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.702 + ], + "angle": 0, + "content": "3. Ferrari, V., Zisserman, A.: Learning visual attributes. In: NeurIPS. pp. 433-440 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. Fu, X., Xiao, J., Zhu, Y., Liu, A., Wu, F., Zha, Z.J.: Continual image deraining with hypergraph convolutional networks. IEEE TPAMI 45(8), 9534-9551 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.623, + 0.743 + ], + "angle": 0, + "content": "5. Girshick, R.: Fast r-cnn. In: ICCV. pp. 1440-1448 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "6. Guo, X., Hu, Q.: Low-light image enhancement via breaking down the darkness. IJCV 131(1), 48-66 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "7. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Guo, Y., Gao, Y., Liu, W., Lu, Y., Qu, J., He, S., Ren, W.: Scenet: Self-paced semicurricular attention network for non-homogeneous image dehazing. In: CVPRW. pp. 1884-1893 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Han, J., Li, W., Fang, P., Sun, C., Hong, J., Armin, M.A., Petersson, L., Li, H.: Blind image decomposition. In: ECCV. pp. 218-237 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "21. Hoang, T., Zhang, H., Yazdani, A., Monga, V.: Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In: CVPR. pp. 1670-1679 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.271 + ], + "angle": 0, + "content": "22. Jiao, J., Tu, W.C., Liu, D., He, S., Lau, R.W.H., Huang, T.S.: Formnet: Formatted learning for image restoration. IEEE TIP 29, 6302-6314 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. CACM 60(6), 84-90 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "24. Kumar, N., Berg, A., Belhumeur, P.N., Nayar, S.: Describable visual attributes for face verification and image search. IEEE TPAMI 33(10), 1962-1977 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "25. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP 28(1), 492-505 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.379 + ], + "angle": 0, + "content": "26. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.406 + ], + "angle": 0, + "content": "27. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR. pp. 1633-1642 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "28. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "29. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR. pp. 2041-2050 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "30. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP 27(6), 3064-3073 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "31. Luo, Y., Huang, Q., Ling, J., Lin, K., Zhou, T.: Local and global knowledge distillation with direction-enhanced contrastive learning for single-image deraining. Knowledge-Based Syst. 268, 110480 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.554 + ], + "angle": 0, + "content": "32. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. JMLR 9(11), 2579-2605 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.582 + ], + "angle": 0, + "content": "33. Meng, Q., Shin'ichi, S.: Adinet: Attribute driven incremental network for retinal image classification. In: CVPR. pp. 4033-4042 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "34. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: CVPR. pp. 17399-17410 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "35. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI 45(8), 10346-10357 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "36. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.691 + ], + "angle": 0, + "content": "37. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: PromptIR: Prompting for all-in-one image restoration. NeurIPS 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "38. Quan, Y., Tan, X., Huang, Y., Xu, Y., Ji, H.: Image desnowing via deep invertible separation. IEEE TCSVT 33(7), 3133-3144 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "39. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: CVPR. pp. 13658-13667 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.772 + ], + "angle": 0, + "content": "40. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: ICCV. pp. 12580-12589 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "41. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP 32, 1927-1941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "43. Wang, C., Zheng, Z., Quan, R., Sun, Y., Yang, Y.: Context-aware pretraining for efficient blind image decomposition. In: CVPR. pp. 18186-18195 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "44. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: CVPR. pp. 11197-11206 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "45. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR. pp. 3677-3686 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "46. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: BMVC. pp. 1-12 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "47. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR. pp. 10551-10560 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "48. Xu, W., Xian, Y., Wang, J., Schiele, B., Akata, Z.: Attribute prototype network for zero-shot learning. In: NeurIPS. pp. 21969-21980 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.356 + ], + "angle": 0, + "content": "49. Xu, X., Wang, R., Lu, J.: Low-light image enhancement via structure modeling and guidance. In: CVPR. pp. 9893-9903 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "50. Yan, X., Yang, J., Sohn, K., Lee, H.: Attribute2image: Conditional image generation from visual attributes. In: ECCV. pp. 776-791 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "51. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "53. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV. pp. 492-511 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR. pp. 14821-14831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for fast image restoration and enhancement. IEEE TPAMI 45(2), 1934-1948 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "56. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "57. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR. pp. 5785-5794 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "58. Zhou, M., Huang, J., Guo, C.L., Li, C.: Fourmer: An efficient global modeling paradigm for image restoration. In: ICML. pp. 42589-42601 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "59. Zhou, M., Wu, X., Wei, X., Xiang, T., Fang, B., Kwong, S.: Low-light enhancement method based on a retina model for structure preservation. IEEE TMM pp. 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "60. Zhou, Y., Xu, Y., Du, Y., Wen, Q., He, S.: Propulse: Learning progressive encoders of latent semantics in gans for photo upsampling. IEEE TIP 31, 1230-1242 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "61. Zhu, Y., Min, W., Jiang, S.: Attribute-guided feature learning for few-shot image recognition. IEEE TMM 23, 1200-1209 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "62. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_origin.pdf b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e5a9162308474ad3f44bc8a8cc2e7757153c335 --- /dev/null +++ b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/4c19b25d-ef9c-4ce6-aa25-4370cb382413_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48a9704ddcec3aa41810ea3677c5c4873f065c70a09ff6ffc212e8c426c699cc +size 8075488 diff --git a/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/full.md b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..74fd0832ba0d90d24df6eac5ef3654152e80b5ac --- /dev/null +++ b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/full.md @@ -0,0 +1,324 @@ +# OneRestore: A Universal Restoration Framework for Composite Degradation + +Yu Guo $^{1,2,\dagger}$ , Yuan Gao $^{1,\dagger}$ , Yuxu Lu $^{3}$ , Huilin Zhu $^{1,2}$ , Ryan Wen Liu $^{1(\boxtimes)}$ , and Shengfeng He $^{2(\boxtimes)}$ + +1 Wuhan University of Technology, Wuhan, China +wenliu@whut.edu.cn + +$^{2}$ Singapore Management University, Singapore shengfenghe@smu.edu.sg + +3 The Hong Kong Polytechnic University, Hong Kong, China +† Equal Contribution +https://github.com/gy65896/OneRestore + +Abstract. In real-world scenarios, image impairments often manifest as composite degradations, presenting a complex interplay of elements such as low light, haze, rain, and snow. Despite this reality, existing restoration methods typically target isolated degradation types, thereby falling short in environments where multiple degrading factors coexist. To bridge this gap, our study proposes a versatile imaging model that consolidates four physical corruption paradigms to accurately represent complex, composite degradation scenarios. In this context, we propose OneRestore, a novel transformer-based framework designed for adaptive, controllable scene restoration. The proposed framework leverages a unique cross-attention mechanism, merging degraded scene descriptors with image features, allowing for nuanced restoration. Our model allows versatile input scene descriptors, ranging from manual text embeddings to automatic extractions based on visual attributes. Our methodology is further enhanced through a composite degradation restoration loss, using extra degraded images as negative samples to fortify model constraints. Comparative results on synthetic and real-world datasets demonstrate OneRestore as a superior solution, significantly advancing the state-of-the-art in addressing complex, composite degradations. + +# 1 Introduction + +Image restoration is an essential task that entails the recovery of high-quality visuals from compromised inputs. The fidelity of these restored images is vital for the subsequent downstream tasks, such as robot navigation and autonomous driving. Considerable progress has been made in addressing single degradation scenarios by a single model (i.e., One-to-One models, see Fig. 2a), such as low-light conditions [16, 49, 59], haze [21, 41, 57], rain [11, 14, 31], and snow [4, 8, 38]. These methods have achieved remarkable success within their intended contexts. Nevertheless, real-world conditions often involve unpredictable and variable composite degradations that can severely impact image quality and clarity [28]. The + +![](images/ce0911d14d93dc0caa79befdfb59a29e29097f3b39cd25b2871ee9ab6c5784fa.jpg) +Fig. 1: Our OneRestore allows fully controllable image restoration using scene descriptors derived from both automatic visual attribute extraction (top) and manual text embeddings (bottom). + +occurrence of mixed degrading conditions in one image can lead to intricate interactions, posing significant challenges to One-to-One solutions [22,52,54,60]. The necessity for seamless switching and decision-making across specialized methods for each condition proves to be inefficient and ineffective in practical applications. Therefore, there is a growing imperative for robust, all-encompassing image restoration approaches that can universally address multiple forms of degradation through a singular, dynamic framework. + +Recent research has been directed toward One-to-Many models, which address various individual degradation factors within a single framework. These models are categorized into: those with partial parameter-sharing (One-to-Many $_p$ ) [2,19,28,45] and those with full parameter-sharing (One-to-Many $_f$ ) [6,26,35,42, 43]. One-to-Many $_p$ (left of Fig. 2b) restoration methods typically employ models with either multiple heads and tails or a single tail to address various types of image impairments. However, the complexity and size of these models tend to scale with the diversity of degradations they aim to correct, leading to increased training costs and parameter counts. Furthermore, when images suffer from a combination of degrading factors, these models may underperform. Some One-to-Many $_p$ techniques also depend on user intervention to select appropriate restoration weights, which is impractical for automated or dynamic settings. On the other hand, One-to-Many $_f$ (right of Fig. 2b) approaches train models using a direct learning strategy, where inputs are images with different individual degradations and the targets are their clear counterparts. However, this direct training approach is designed predominantly for single-category degradation issues, it does not adequately address the complex dynamics that arise when multiple types of impairments affect an image simultaneously. As a result, such methods can falter in practical applications where user control and precise directionality of restoration are crucial. For instance, in cases where users seek to improve low-light imagery, the model might inadvertently prioritize dehazing actions, deviating from user intent. + +![](images/d3bf3cb80d8a45e50e8cda9e7b379be45966ec8f4f6e29b3033ae91509144a2f.jpg) +(a) One-to-One + +![](images/89c731f4eeff7a0e89a4ba6f770a22f2ca71c8b2ceaaac1535c8230c029872f7.jpg) +(b) One-to-Many + +![](images/9673f91c9c1157ad1d621b7a5deefc002ef0a3d892847725de2faec3d48fea31.jpg) +(c) One-to-Composite +Fig.2: Structural comparison between One-to-One, One-to-Many, and our One-to-Composite image restoration methods. + +Drawing from the preceding discussion, we delineate the primary challenges in developing image restoration models for composite degradations in real-world conditions: + +- How can the model be designed to adaptively discern the specific degradation afflicting the input image? +- What approaches enable a model to be trained efficiently across a spectrum of degradations while enhancing user control over the restoration process? +- How can the recovery process be optimized that not only approach the quality of clear images but also maintain distinction from various degraded forms? + +In response to the challenges outlined, we present a unified model that builds upon established physical corruption models for degradations such as low light, haze, rain, and snow. This model effectively captures the complexity of real-world degradation conditions and forms the basis for constructing our Composite Degradation Dataset (CDD-11), which serves both training and evaluation purposes. To address the issue of composite degradation, our work introduces OneRestore, a transformer-based universal restoration architecture. This novel framework distinguishes itself from conventional methods by embodying a Oneto-Composite strategy, as depicted in Fig. 2c, which allows for precise and controllable outcomes (see Fig. 1). It employs an innovative cross-attention block that fuses scene descriptors, which can be manual text embeddings or derived automatically from visual attributes. We further enhance the model's robustness by integrating a composite degradation restoration loss, using additional degraded images as negative samples to establish stringent lower-bound constraints. + +Our key contributions are summarized as follows: + +- We present the first attempt to unify composite degradations, creating a versatile imaging model that simulates a range of degradation types. This innovation underpins the development of our comprehensive Composite Degradation Dataset. +- We propose a universal framework that introduces a novel level of controllability in image restoration, employing a cross-attention mechanism that integrates scene descriptors derived from either manual text embeddings or automated visual attribute extractions. + +- We tailor a composite degradation restoration loss, enhancing the model's ability to discern between various forms of image degradation. + +- Our universal model not only sets a new state-of-the-art for composite dataset performance but also competes favorably with One-to-One models in single-degradation scenarios. + +# 2 Related Work + +Complex Degradation Image Restoration. Recent advancements in image restoration have introduced a variety of solutions for complexly degraded images. A notable direction is the partial parameter-sharing One-to-Many approach, employing architectures with multiple heads and tails, and a single backbone, to address various degradation factors. For instance, Li et al. [28] develop an all-weather network that restores scenes affected by haze, rain, and snow, using specialized encoders for each condition and a shared decoder. Chen et al. [2] introduce a transformer-based model for deraining, denoising, and super-resolving images, leveraging a multi-head and multi-tail design. Wang et al. [45] present a knowledge assignment strategy to differentiate and process rain and snow degradations adaptively. While these models represent strides in complex degradation image restoration, they stop short of being truly universal One-to-Many networks. With only partial weight sharing, they address different degradation types through distinct network pathways. + +Conversely, universal image restoration has emerged with the One-to-Many methods sharing all network weights, thereby uniformly mitigating multiple degradation types. Examples include the unified two-stage knowledge learning framework by Chen et al. [6], which efficiently processes dehazing, deraining, and desnoring with a singular pretrained weight set. Li et al. [26] present a network capable of comprehensive image restoration without the need for pre-identifying degradation levels. Additionally, Ozdenizci et al. [35] introduce a conditional denoising diffusion model tailored for patch-based image restoration under varied weather conditions. However, due to the significant differences and mutual interference between various degradation factors, this type of method can only solve specific types of degradation and cannot achieve image restoration in composite degradation scenarios. To this end, we propose a universal One-to-Composite image restoration method that allows the model to adapt to arbitrary combinations of composite degradations. + +Visual Attribute Control. Visual attributes are pivotal in defining image features and enhancing the interpretability of models. They enrich feature representations and provide a hierarchical framework for image analysis. Early research highlighted visual attributes' transformative impact on image generation, recognition, and retrieval [13, 24, 50]. More recent studies underscore their utility in improving object recognition and detection, particularly in data-scarce environments [1, 33, 48, 61]. Visual attributes have also been leveraged for zero-shot learning applications, advancing classification tasks beyond the constraints of labeled datasets [39, 44]. In our study, we harness visual attributes to delineate the intricate degradation patterns present in real-world imaging conditions, + +enhancing the pertinence and efficacy of image restoration techniques. It is noteworthy that conventional discriminators, when tasked with identifying image degradation, are typically confined to image data, precluding autonomous feature control. Our approach, by contrast, accommodates both images and text descriptions as inputs, thereby facilitating the automatic or manual determination of degradation characteristics. + +# 3 Composite Degradation Formulation + +# 3.1 Imaging Formulation + +In real-world scenarios, images are often affected by multiple degrading factors simultaneously, resulting in suboptimal performance from restoration methods designed for isolated degradations. To address the broad spectrum of degradation scenarios, we introduce a versatile imaging model that encapsulates all degradations, including low light, haze, rain, and snow. This model aims to narrow the disparity between simulated and authentic real-world data. Given a degraded image $I(x)$ and its corresponding clear state $J(x)$ , the proposed imaging model is mathematically articulated as + +$$ +I (x) = \mathcal {P} _ {h} \left(\mathcal {P} _ {r s} \left(\mathcal {P} _ {l} (J (x)))\right), \right. \tag {1} +$$ + +where $\mathcal{P}_l, \mathcal{P}_{rs}$ , and $\mathcal{P}_h$ denote the low light, rain/snow, and haze degradations, respectively. Given that concurrent rain and snow conditions are infrequent, they are not addressed within the scope of this work. A comprehensive exposition of the proposed imaging model is detailed below. + +Low-Light Conditions. Inspired by Retinex theory, low-light degraded image $I_{l}(x)$ can be generated by + +$$ +I _ {l} (x) = \mathcal {P} _ {l} (J (x)) = \frac {J (x)}{L (x)} L (x) ^ {\gamma} + \varepsilon . \tag {2} +$$ + +Here, we use a darkening coefficient $\gamma \in [2,3]$ as the brightness controller to adjust the illumination map $L(x)$ generated via LIME [17]. To simulate poor light conditions realistically, we add the Gaussian noise $\varepsilon$ with mean $= 0$ and variance $\in [0.03,0.08]$ . + +Rain/Snow Streaks. Following the methodologies outlined in [5, 27], our process superimposes rain or snow streaks onto the image before synthesizing haze. To introduce rain streaks $\mathcal{R}$ , we produce the rain-degraded image $I_{rs}(x)$ by overlaying $\mathcal{P}_{rs}$ onto the low-light image $I_l(x)$ , which is expressed as + +$$ +I _ {r s} (x) = \mathcal {P} _ {r s} \left(I _ {l} (x)\right) = I _ {l} (x) + \mathcal {R}. \tag {3} +$$ + +For snow streaks $S$ , the snow-degraded image is defined as + +$$ +I _ {r s} (x) = \mathcal {P} _ {r s} \left(I _ {l} (x)\right) = I _ {l} (x) (1 - \mathcal {S}) + M (x) \mathcal {S}, \tag {4} +$$ + +with $M(x)$ being the chromatic aberration map. + +![](images/daf44e536a16f742dade2485957e8ba3a71419b22b4d49cc02457034b5a2fea6.jpg) +(a) Overall Pipeline + +![](images/5133cc4a36aac790ceb6d30ac7a73882055fdfc11d2766e109f5e5b6921c18d8.jpg) +(b) Scene Descriptor Generation +Fig. 3: Architecture of our OneRestore. (a) Overall pipeline, where 32, 64, 128, and 256 represent the number of channels. (b) Scene descriptor generation, where scene descriptors are fed into each (c) Scene Descriptor-guided Transformer Block (SDTB) by manual text embeddings or automatic extraction based on visual attributes. + +![](images/03612244224db9547ef4f2d7a3d07af8efcde1069e026f0ad7c810984d074af4.jpg) +(c)SDTB + +Haze Degradations. We incorporate haze degradation by employing the atmospheric scattering model as + +$$ +I (x) = \mathcal {P} _ {h} \left(I _ {r s} (x)\right) = I _ {r s} (x) t + A (1 - t), \tag {5} +$$ + +where $t$ denotes the transmission map and $A$ represents the atmospheric light. The transmission map $t$ is defined by the exponential decay of light, $t = e^{-\beta d(x)}$ , where $\beta$ represents the haze density coefficient and $d(x)$ is the scene depth (estimated by MegaDepth [29]). The haze density coefficient $\beta$ is varied within the range of [1.0, 2.0] and the atmospheric light $A$ is constrained to [0.6, 0.9] for realistic simulation. + +# 3.2 Composite Degradation Dataset + +Drawing on Eq. (1), we have developed the Composite Degradation Dataset (CDD-11), encompassing 11 categories of image degradations and their clear counterparts. These degraded samples include low (low-light), haze, rain, snow, low+haze, low+rain, low+snow, haze+rain, haze+snow, low+haze+rain, and low+haze+snow. From the RAISE database [10], we selected 1,383 high-resolution clear images for producing 11 composite degradations. They are resized to a uniform resolution of $1080 \times 720$ . The overall dataset is split into 13,013 image pairs for training and 2,200 for testing. + +# 4 OneRestore + +The encoder-decoder network, validated as an effective architectural paradigm for image restoration [7, 12, 18], serves as the overall structure for our model, as depicted in Fig. 3a. Building upon the encoder-decoder structure, we devise + +an innovative Scene Descriptor-guided Transformer Block (SDTB) as the basic unit for feature extraction (Fig. 3c). Our approach diverges from existing transformer-based methods [4,7,41] by concurrently integrating image features and scene descriptors into the SDTB basic block for supporting fine-grained image restoration. In particular, the generation of scene descriptors ranges from manual text embeddings to automatic extraction based on visual attributes, as illustrated in Fig. 3b. Further details on scene descriptor generation and SDTB are elaborated in Sections 4.1 and 4.2. For capturing multi-scale representations, the encoder employs three downsampling operations, mirrored by the decoder's upsampling operations. To be specific, we utilize "Max-Pooling + Conv" and "Bilinear Interpolation + Conv" for downsampling and upsampling, respectively. Simultaneously, three skip connections and a global residual enhance model convergence by bridging shallow and deep features. Additional details regarding the configuration of our OneRestore can be found in supplementary materials. + +Ultimately, the overall loss $\mathcal{L}$ for model training is defined as follows + +$$ +\mathcal {L} = \alpha_ {1} \mathcal {L} _ {1} ^ {s} (J, \hat {J}) + \alpha_ {2} \mathcal {L} _ {\mathrm {M}} (J, \hat {J}) + \alpha_ {3} \mathcal {L} _ {\mathrm {c}} (J, \hat {J}, I, \{I _ {o} \}), \tag {6} +$$ + +where $\mathcal{L}_1^s$ , $\mathcal{L}_{\mathrm{M}}$ , and $\mathcal{L}_{\mathrm{c}}$ represent the smooth $l_{1}$ loss [15], Multi-Scale Structural Similarity (MS-SSIM) loss, and the proposed composite degradation restoration loss (see Sec. 4.3), respectively, $\alpha_{1-3}$ denote the penalty coefficients, $J$ , $\hat{J}$ , $I$ , and $\{I_o\}$ are the positive, model output anchor, model input negative, and other negatives, respectively. + +# 4.1 Scene Descriptor Generation + +To augment the controllability of our OneRestore, we introduce the scene description embedding as an additional input to the model, alongside the degraded image. As delineated in Fig. 3b, the generation of the scene descriptor offers two selectable modes: the manual mode $①$ and the automatic mode $②$ . In the manual mode, user involvement is necessitated to supply scene description text, generating the corresponding text embedding employed as an input. Conversely, the automatic mode involves the extraction of visual attributes from the image to generate visual embedding, estimating the most proximate text embedding. To quantify the difference between visual and text embeddings and train optimal text and visual embedders, we introduce cosine cross-entropy loss for model weight optimization. + +Text Embedder. For the text embedding generation task, we employ a set of 5 scene description texts as input to generate 12 text embeddings. Initially, GloVe [36] is leveraged to generate initial text embeddings for 1 clear scene and 4 scenes with single degradation types. Subsequently, 7 additional composite degradation text embeddings (i.e., low+ haze, low+ rain, low+ snow, haze+ rain, haze+ snow, low+ haze+ rain, and low+ haze+ snow) are generated by averaging the element values in the corresponding text embeddings. A Multi-Layer Perception (MLP) is then employed to refine the 12 text embeddings. + +Visual Embedder. Taking inspiration from the work of Saini et al. [39], we first resize the input image to $224 \times 224$ and extract image features using a + +ResNet-18 [20] network without average pooling. The initial weights of ResNet-18 are pre-trained on the ImageNet dataset. Subsequently, a single convolutional layer, followed by a dropout layer and a linear layer, is introduced to generate the final visual embedding. By calculating the cosine similarity between visual embeddings and all text embeddings, and selecting the one with the highest similarity, we can automatically generate scene descriptors to achieve image restoration based on visual attribute control. + +Cosine Cross-Entropy. Similar to [39], the score of visual embedding and each text embedding is obtained by "Cosine Similarity + Softmax". Given $e_v$ and $e_t$ being the visual and text embeddings, the similarity score $S(e_v, e_t)$ can be given by + +$$ +\cos \left(e _ {v}, e _ {t}\right) = \delta \cdot \frac {e _ {v} \cdot e _ {t} ^ {\top}}{\| e _ {v} \| \| e _ {t} \|}, \tag {7} +$$ + +$$ +S \left(e _ {v}, e _ {t}\right) = \frac {e ^ {\cos \left(e _ {v} , e _ {t}\right)}}{\sum_ {t _ {i} = 1} ^ {N _ {t}} e ^ {\cos \left(e _ {v} , e _ {t _ {i}}\right)}}, \tag {7} +$$ + +with $\delta$ and $N_{t}$ being the temperature factor and the number of text embeddings, respectively. Based on the calculated $S(e_v, e_t)$ , we use the cross-entropy loss for model training. + +# 4.2 Scene Descriptor-guided Transformer Block + +To further harness the capabilities of scene description embeddings in image restoration tasks and facilitate a more adaptive and nuanced restoration process, we introduce the innovative Scene Descriptor-guided Transformer Block (SDTB), illustrated in Fig. 3c. The SDTB consists of three primary components, namely Scene Descriptor-guided Cross-Attention (SDCA), Self-Attention (SA), and Feed-Forward Network (FFN). Diverging from traditional self-attention mechanisms, the proposed SDCA module utilizes a query generated from scene descriptors instead of queries generated from image features. The mathematical representation of our SDCA can be written as + +$$ +\operatorname {S D C A} \left(\mathbf {Q} _ {t}, \mathbf {K}, \mathbf {V}\right) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {t} \cdot \mathbf {K} ^ {\top}}{\lambda}\right) \mathbf {V}, \tag {8} +$$ + +where $\lambda$ is a temperature factor, $\mathbf{Q}_t$ represents the scene descriptor-generated query matrix, $\mathbf{K}$ and $\mathbf{V}$ denote the image feature-generated key and value matrices, respectively. $\mathbf{Q}_t$ is obtained from the input scene description embedding $e_t$ via a linear layer. During the process of matrix multiplication between $\mathbf{Q}_t$ and $\mathbf{K}$ . To ensure consistency in the number of tokens between $\mathbf{K}$ and $\mathbf{Q}_t$ , we resize the original image before performing the reshaping operation to generate $\mathbf{K}$ . By incorporating the SDCA module, each transformer block in OneRestore can effectively integrate scene descriptors and image features for targeted image restoration. As for the SA and FFN modules, we adopt structures akin to those in Restormer [52]. More details of our SDTB are introduced in the supplement. + +![](images/13540744a2f4e1a7cae699f1a7cd0e6895e301f0fabbac1cd87ca92d4d2722f3.jpg) +Fig. 4: Illustration of the composite degradation restoration loss. + +# 4.3 Composite Degradation Restoration Loss + +Using contrastive loss has been demonstrated as an effective way of improving image restoration performance [47, 57]. Essentially, contrastive loss views the degraded input image as a negative sample, the clear image as a positive sample, and the restored image as an anchor. This loss function works to minimize the distance between the anchor and the positive sample while maximizing its distance from the negative sample. In the composite degradation image restoration task, the traditional contrastive loss, however, may inadvertently reduce the distance between the anchor and other degraded images. To make the anchor closer to the positive sample while maintaining a clear separation from all degraded negatives, we design a composite degradation restoration loss, illustrated in Fig. 4. Considering $I$ , $I_o$ , $J$ , and $\hat{J}$ as the input negative, other negatives, positive, and anchor, the proposed loss can be written as + +$$ +\mathcal {L} _ {\mathrm {c}} (J, \hat {J}, I, \left\{I _ {o} \right\}) = \sum_ {k = 1} ^ {K} \xi_ {k} \frac {\mathcal {L} _ {1} \left(V _ {k} (\hat {J}) , V _ {k} (\hat {J})\right)}{\xi_ {c} \mathcal {L} _ {1} \left(V _ {k} (\hat {J}) , V _ {k} (I)\right) + \sum_ {o = 1} ^ {O} \xi_ {o} \mathcal {L} _ {1} \left(V _ {k} (I _ {o}), V _ {k} (\hat {J})\right)}, \tag {9} +$$ + +where $V_{k}(\cdot)$ denotes the extraction operation of $k$ -th hidden feature from the VGG-16 pre-trained on the ImageNet, $\mathcal{L}_1$ represents the $l_1$ loss, $K$ and $O$ are the numbers of used VGG-16 feature layers and other degraded negatives, $\xi_{k}, \xi_{c}$ , and $\xi_{o}$ denote the hyper-parameters, respectively. In this work, the number of other degraded negatives $O$ is 10. We use the output features of 3-rd, 8-th, and 15-th layers from VGG-16 and set $K = 3$ . Furthermore, we consider that the features of each feature layer of each degraded negative have equal importance. Therefore, we set $\xi_{c} = \xi_{o} = \frac{1}{11}$ and $\xi_{k} = \frac{1}{3}$ . According to our ablation study, the proposed loss confers a significant advantage over using a single input degradation as a negative sample, which can easily capture positive sample features. + +Table 1: Settings of model training. "lr" denotes the learning rate. + +
TermsText/Visual EmbedderOneRestore
Epoch200120
Initial lr0.00010.0002
lr Decay0.5/50 epochs0.5/20 epochs
Batch Size2564
+ +Table 2: Comparison of quantitative results on CDD-11 dataset. OneRestore† means using the corresponding scene description text as additional input, and WGWSNet requires the scene type as model input. Red, green, and blue indicate the best, second-best, and third-best results, respectively. + +
TypesMethodsVenue & YearPSNR ↑SSIM ↑#Params
Input16.000.6008-
One-to-OneMIRNet [53]ECCV202025.970.847431.79M
MPRNet [54]CVPR202125.470.855515.74M
MIRNetv2 [55]TPAMI202225.370.83355.86M
Restormer [52]CVPR202226.990.864626.13M
DGUNet [34]CVPR202226.920.855917.33M
NAFNet [3]ECCV202224.130.796417.11M
SRUDC [40]ICCV202327.640.86006.80M
Fourmer [58]ICML202323.440.78850.55M
OKNet [9]AAAI202426.330.86054.72M
One-to-ManyAirNet [26]CVPR202223.750.81408.93M
TransWeather [42]CVPR202223.130.781021.90M
WeatherDiff [35]TPAMI202322.490.798582.96M
PromptIR [37]NIPS202325.900.849938.45M
WGWSNet [62]CVPR202326.960.862625.76M
One-to-CompositeOneRestore28.470.87845.98M
OneRestore†28.720.88215.98M
+ +# 5 Experiments + +# 5.1 Experiment Settings + +Implementation Details. Our OneRestore is implemented by PyTorch 1.12.0 and trained on a PC with 2 AMD EPYC 7543 32-Core Processors and 8 NVIDIA L40 GPUs. We use the Adam with exponential decay rates being $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ for optimization. The training parameter settings for text/visual embedder and OneRestore are shown in Table 1. In particular, all images used for training are cropped into $256\times 256$ image patches with a sampling stride of 200 and randomly flipped by 0, 90, 180, and 270 degrees to generate $312\mathrm{k}$ patch pairs. More details about training and inference can be found in the supplement. + +Evaluation Datasets. To facilitate a thorough comparison, we execute image restoration experiments across various datasets. These include the Composite Degradation Dataset (CDD-11) crafted by us, established One-to-One benchmarks, and authentic real-world datasets. The standard One-to-One benchmarks are designated for four principal tasks: the LOw-Light dataset (LOL) [46], the REalistic Single Image DEhazing Outdoor Training Set (RESIDE-OTS) [25], the Rain1200 dataset [56], and the Snow100k dataset [30]. Owing to the dearth of specialized datasets for real-world composite degradation, we have curated a selection from the LIME low-light enhancement dataset [17], the RESIDE Realworld Task-driven Testing Set (RESIDE-RTTS) [25], Yang's rainy dataset [51], and the realistically snowy Snow100k dataset [25]. + +Competitors and Evaluation Metrics. We compare our OneRestore with 9 One-to-One image restoration methods (MIRNet [53], MPRNet [54], MIRNetv2 [55], Restormer [52], DGUNet [34], NAFNet [3], SRUDC [40], Fourmer [58], and OKNet [9]) and 5 One-to-Many image restoration methods (AirNet [26], + +TransWeather [42], WeatherDiff [35], PromptIR [37], and WGWSNet [62]). In particular, all compared state-of-the-art methods are retrained on our CDD-11 train set. For PromptIR and WGWSNet, we adjust the number of prompts and the weather-specific parameters of each layer for corresponding to our task, respectively. Furthermore, we adopt the Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM) as evaluation metrics. + +# 5.2 Comparison with SOTAs + +The quantitative evaluation results on our CDD-11 test dataset are displayed in Table 2. OneRestore† with text embedder achieves the highest performance, followed by OneRestore with visual embedder as the second-best. Both OneRestore† and OneRestore demonstrate substantial leads over the third-ranked methods in terms of both PSNR and SSIM. Further detailed restoration performance comparison of our method and the One-to-Many methods in each degradation scenario is illustrated in Fig. 5 using radar charts. Notably, image restoration for composite degradation scenarios proves challenging. By incorporating scene description em + +![](images/2f3bc6b251e00659ff9072372f5dd158b5a3e13641f7cda6abdf13592a185b96.jpg) +Fig.5: Comparison of quantitative results for different degradation scenarios on CDD-11 dataset. + +beddings into image restoration models and utilizing the proposed composite degradation restoration loss, our method robustly reconstructs clear images across various situations. For compelling evidence, Fig. 6 provides the visual comparison of image restoration in two composite degradation samples. It is evident that SRUDC, OKNet, AirNet, and WGWSNet are unable to suppress all degenerative interferences, while Restormer and PromptIR are prone to color distortion and noise residual. In contrast, the proposed OneRestore can produce more natural results and fully preserve image textures. + +Moreover, Fig. 7 displays the image restoration results in two real-world scenarios. Compared with existing One-to-Many image restoration methods [26,35, 37], the proposed OneRestore can simultaneously focus on multiple degradation factors and robustly reconstruct scene details. More results on CDD-11, realworld scenarios, and other One-to-One benchmarks of low-light, hazy, rainy, and snowy scenes are included in supplementary materials. + +# 5.3 Ablation Study + +Effectiveness of Network Modules. Table 3 reports the quantitative evaluation of different module configurations. While the Feed-Forward Network (FFN) with convolution as the basic unit serves as a crucial module in the transformer network, it alone fails to generate satisfactory reconstruction results. The ordinary Self-Attention module (SA) significantly strengthens network performance, + +![](images/b218a344f70fced369512260e59c87b18705cdc09eac3fdaa87955ec1a89956d.jpg) +Fig. 6: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) synthetic samples. + +![](images/344e23f352e4b9baa19335a5f028e1aefaad89bfa4d796b6923b483d4ae9bdea.jpg) +Fig. 7: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) samples in real-world scenarios. * represents the utilization of original weights published in the author's code. + +but this method regrets that the model fails to be flexibly controlled. Compared to depending only on SA, relying exclusively on Scene Description-guided Cross-Attention (SDCA) utilizing scene description embedding yields better outcomes. This further confirms that addressing complex composite degradations can be done more effectively and flexibly by integrating degradation scene descriptors into the model. Ultimately, the best performance can be achieved by using a combination of SDCA, SA, and FFN. + +Selection of Description Embedding Strategies. In this section, we compare the image restoration performance by using different strategies to generate scene descriptor, including Classifier (utilizing the embedding before the last linear layer of the retrained AlexNet [23] as the scene descriptor), Visual Embedder (employing visual embedding for model training), Visual Embedder‡ (deriving text embedding via visual embedder), and Text Embedder (employing text embedding for model training). The quantitative evaluation results are presented in Table 4. It is observed that employing Classifier and Visual Embedder to generate scene descriptors + +![](images/1975813ece73deb6c035bf2ef205dcb59bbb05a99dde20ac023e1beaf6f728aa.jpg) +Fig.8: Comparison of image restoration on different description embedding strategies. + +Table 3: Ablation study for different model configurations. SDCA, SA, and FFN denote the scene description-guided cross-attention, self-attention, and feed-forward network, respectively. + +
SDCASAFFNPSNR ↑SSIM ↑Controllability
✓24.810.8607
✓✓27.190.8697
✓✓27.930.8767✓
✓✓✓28.720.8821✓
+ +![](images/8cbc7b31a0ae2f37c7326a6ccfacb78e5e528a8bd96cba6918db0a56202e2540.jpg) +Fig. 9: Comparison of image restoration on $low + haze + rain$ (top) and $low + haze + snow$ (bottom) synthetic samples by using different texts. + +yields comparable outcomes. However, these approaches have a drawback in that each image has a distinct embedding as input, making it challenging for the model to recognize the cause of degradation. Fig. 8 displays a classic case comparison using different embedding strategies, where the abnormal artifacts of the results produced by Classifier and Visual Embedder are caused by the model's fuzzy identification of degradation factors. Moreover, the restoration model lacks controllability. The discrepancy between Visual Embedder $\ddagger$ and Text Embedder stems from errors in scene descriptor estimation. Visual Embedder $\ddagger$ has an accuracy of $97.55\%$ on the test dataset, with misestimation occurring when some degradation types are insignificant. For instance, a haze+snow scene may be evaluated as a haze or snow scene. In contrast, the text embedding-based approach is superior in describing degradation scenes and recovering images, using a fixed number of scene descriptors. This allows the model to employ degraded scene descriptors as switches to achieve controlled restoration. + +Effectiveness of Loss Functions. Table 5 presents a performance comparison of different loss functions. Notably, the integration of the MS-SSIM loss is instrumental in elevating SSIM scores at the cost of reduced PSNR performance. In contrast, the CL loss demonstrates minimal impact, providing only marginal improvements in SSIM and PSNR. Upon replacing CL with CDRL, a substantial enhancement is observed in both PSNR and SSIM. The proposed CDRL loss yields an output that closely resembles the clear image while avoiding other forms of degradation. + +![](images/2c5fd263c7c705af5b4b73760ba662c7b5ba5e1ef6fcddf6a42b9cd9253ee00c.jpg) +Fig.10: Visualization of visual embeddings based on t-SNE. + +Table 4: Ablation study for description embedding strategies. + +
ModelsPSNR ↑SSIM ↑Controllability
Classifier28.190.8783
Visual Embedder28.240.8781
Visual Embedder‡28.470.8784
Text Embedder28.720.8821✓
+ +Table 5: Ablation study for different loss functions. CL and CDRL denote the ordinary contrastive loss and the proposed composite degradation restoration loss, respectively. + +
Smooth l1MS-SSIMCLCDRLPSNR ↑SSIM ↑
✓28.160.8633
✓✓27.540.8708
✓✓✓27.610.8723
✓✓✓28.720.8821
+ +# 5.4 Model Control via Scene Descriptor + +To demonstrate the controllability of the scene descriptor in our OneRestore, we conduct experiments on two intricate synthetic degradation scenarios, as depicted in Fig. 9. By manually adjusting specific text to generate corresponding degraded scene description embeddings, our model can selectively focus on different degradation factors, achieving targeted restorations. Our approach excels in controllable restoration without the need for multiple sets of weight screening, making it more concise compared to other One-to-Many methods that rely on partial parameter sharing. Meanwhile, our method can support adaptive restoration through the estimation of text embeddings based on visual attribute extraction. Fig. 10 employs the t-SNE [32] to visualize the embeddings generated by the visual embedder on the CDD-11 train and test sets. Obviously, the constructed visual embedder can clearly distinguish between different scenarios. + +# 6 Conclusion + +Our proposed OneRestore framework represents a significant step forward in image restoration, synthesizing a range of degradation patterns to accurately simulate complex scenarios. The innovative use of a scene descriptor-guided cross-attention block within a transformer-based model facilitates an adaptable and fine-grained restoration process, yielding results that surpass prior methods in both synthetic and real-world datasets. However, despite its advancements, OneRestore exhibits limitations in processing extremely high-density corruption scenarios and high-complex corruption scenarios containing unconsidered degradations, where the model's predictive capabilities can be challenged. Future work will aim to enhance the robustness of OneRestore against such extreme conditions and to reduce computational overhead, further extending its applicability and efficiency. Our findings underscore the potential of universal scene restoration, with the caveat that continued refinement is essential for tackling the full spectrum of real-world image degradation. + +Acknowledgement. This project is supported by the Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097 and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011. + +# References + +1. Bertasius, G., Torresani, L.: Cobe: Contextualized object embeddings from narrated instructional video. In: NeurIPS. pp. 15133-15145 (2020) +2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021) +3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33 (2022) +4. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP. pp. 1-5 (2023) +5. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: CVPR. pp. 4196-4205 (2021) +6. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022) +7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR. pp. 5896-5905 (2023) +8. Chen, Z., Sun, Y., Bi, X., Yue, J.: Lightweight image de-snowing: A better trade-off between network capacity and performance. Neural Networks 165, 896-908 (2023) +9. Cui, Y., Ren, W., Knoll, A.: Omni-kernel network for image restoration. In: AAAI. vol. 38, pp. 1426-1434 (2024) +0. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: Raise: A raw images dataset for digital image forensics. In: MMSys. pp. 219-224 (2015) +1. Du, Y., Deng, J., Zheng, Y., Dong, J., He, S.: Dsdnet: Toward single image deraining with self-paced curricular dual stimulations. CVIU 230, 103657 (2023) +2. Du, Y., Han, G., Tan, Y., Xiao, C., He, S.: Blind image denoising via dynamic dual learning. IEEE TMM 23, 2139-2152 (2020) +3. Ferrari, V., Zisserman, A.: Learning visual attributes. In: NeurIPS. pp. 433-440 (2007) +4. Fu, X., Xiao, J., Zhu, Y., Liu, A., Wu, F., Zha, Z.J.: Continual image deraining with hypergraph convolutional networks. IEEE TPAMI 45(8), 9534-9551 (2023) +5. Girshick, R.: Fast r-cnn. In: ICCV. pp. 1440-1448 (2015) +6. Guo, X., Hu, Q.: Low-light image enhancement via breaking down the darkness. IJCV 131(1), 48-66 (2023) +7. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016) +8. Guo, Y., Gao, Y., Liu, W., Lu, Y., Qu, J., He, S., Ren, W.: Scenet: Self-paced semicurricular attention network for non-homogeneous image dehazing. In: CVPRW. pp. 1884-1893 (2023) + +19. Han, J., Li, W., Fang, P., Sun, C., Hong, J., Armin, M.A., Petersson, L., Li, H.: Blind image decomposition. In: ECCV. pp. 218-237 (2022) +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +21. Hoang, T., Zhang, H., Yazdani, A., Monga, V.: Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In: CVPR. pp. 1670-1679 (2023) +22. Jiao, J., Tu, W.C., Liu, D., He, S., Lau, R.W.H., Huang, T.S.: Formnet: Formatted learning for image restoration. IEEE TIP 29, 6302-6314 (2020) +23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. CACM 60(6), 84-90 (2017) +24. Kumar, N., Berg, A., Belhumeur, P.N., Nayar, S.: Describable visual attributes for face verification and image search. IEEE TPAMI 33(10), 1962-1977 (2011) +25. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP 28(1), 492-505 (2018) +26. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022) +27. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR. pp. 1633-1642 (2019) +28. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020) +29. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR. pp. 2041-2050 (2018) +30. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP 27(6), 3064-3073 (2018) +31. Luo, Y., Huang, Q., Ling, J., Lin, K., Zhou, T.: Local and global knowledge distillation with direction-enhanced contrastive learning for single-image deraining. Knowledge-Based Syst. 268, 110480 (2023) +32. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. JMLR 9(11), 2579-2605 (2008) +33. Meng, Q., Shin'ichi, S.: Adinet: Attribute driven incremental network for retinal image classification. In: CVPR. pp. 4033-4042 (2020) +34. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: CVPR. pp. 17399-17410 (2022) +35. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI 45(8), 10346-10357 (2023) +36. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014) +37. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: PromptIR: Prompting for all-in-one image restoration. NeurIPS 36 (2024) +38. Quan, Y., Tan, X., Huang, Y., Xu, Y., Ji, H.: Image desnowing via deep invertible separation. IEEE TCSVT 33(7), 3133-3144 (2023) +39. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: CVPR. pp. 13658-13667 (2022) +40. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: ICCV. pp. 12580-12589 (2023) +41. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP 32, 1927-1941 (2023) +42. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022) + +43. Wang, C., Zheng, Z., Quan, R., Sun, Y., Yang, Y.: Context-aware pretraining for efficient blind image decomposition. In: CVPR. pp. 18186-18195 (2023) +44. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: CVPR. pp. 11197-11206 (2023) +45. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR. pp. 3677-3686 (2023) +46. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: BMVC. pp. 1-12 (2018) +47. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR. pp. 10551-10560 (2021) +48. Xu, W., Xian, Y., Wang, J., Schiele, B., Akata, Z.: Attribute prototype network for zero-shot learning. In: NeurIPS. pp. 21969-21980 (2020) +49. Xu, X., Wang, R., Lu, J.: Low-light image enhancement via structure modeling and guidance. In: CVPR. pp. 9893-9903 (2023) +50. Yan, X., Yang, J., Sohn, K., Lee, H.: Attribute2image: Conditional image generation from visual attributes. In: ECCV. pp. 776-791 (2016) +51. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017) +52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022) +53. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV. pp. 492-511 (2020) +54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR. pp. 14821-14831 (2021) +55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for fast image restoration and enhancement. IEEE TPAMI 45(2), 1934-1948 (2022) +56. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018) +57. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR. pp. 5785-5794 (2023) +58. Zhou, M., Huang, J., Guo, C.L., Li, C.: Fourmer: An efficient global modeling paradigm for image restoration. In: ICML. pp. 42589-42601 (2023) +59. Zhou, M., Wu, X., Wei, X., Xiang, T., Fang, B., Kwong, S.: Low-light enhancement method based on a retina model for structure preservation. IEEE TMM pp. 1-13 (2023) +60. Zhou, Y., Xu, Y., Du, Y., Wen, Q., He, S.: Propulse: Learning progressive encoders of latent semantics in gans for photo upsampling. IEEE TIP 31, 1230-1242 (2022) +61. Zhu, Y., Min, W., Jiang, S.: Attribute-guided feature learning for few-shot image recognition. IEEE TMM 23, 1200-1209 (2020) +62. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023) \ No newline at end of file diff --git a/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/images.zip b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f0094de2c83af16d4b90e37f5dd943d7dd36413b --- /dev/null +++ b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f09fe4d294eccf25e5f8629ff2b1524ae29b730599819a842dcde81b593ccad +size 597676 diff --git a/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/layout.json b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b7d9e09b7b90ad392880e930cd47b31a1d8fa697 --- /dev/null +++ b/2024/OneRestore_ A Universal Restoration Framework for Composite Degradation/layout.json @@ -0,0 +1,9651 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "spans": [ + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "text", + "content": "Yu Guo" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Yuan Gao" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Yuxu Lu" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Huilin Zhu" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Ryan Wen Liu" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1(\\boxtimes)}" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "text", + "content": ", and Shengfeng He" + }, + { + "bbox": [ + 147, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{2(\\boxtimes)}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 203, + 201, + 410, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 201, + 410, + 223 + ], + "spans": [ + { + "bbox": [ + 203, + 201, + 410, + 223 + ], + "type": "text", + "content": "1 Wuhan University of Technology, Wuhan, China \nwenliu@whut.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 224, + 405, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 224, + 405, + 246 + ], + "spans": [ + { + "bbox": [ + 209, + 224, + 405, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 224, + 405, + 246 + ], + "type": "text", + "content": " Singapore Management University, Singapore shengfenghe@smu.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 246, + 432, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 246, + 432, + 280 + ], + "spans": [ + { + "bbox": [ + 182, + 246, + 432, + 280 + ], + "type": "text", + "content": "3 The Hong Kong Polytechnic University, Hong Kong, China \n† Equal Contribution \nhttps://github.com/gy65896/OneRestore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 305, + 455, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 305, + 455, + 515 + ], + "spans": [ + { + "bbox": [ + 160, + 305, + 455, + 515 + ], + "type": "text", + "content": "Abstract. In real-world scenarios, image impairments often manifest as composite degradations, presenting a complex interplay of elements such as low light, haze, rain, and snow. Despite this reality, existing restoration methods typically target isolated degradation types, thereby falling short in environments where multiple degrading factors coexist. To bridge this gap, our study proposes a versatile imaging model that consolidates four physical corruption paradigms to accurately represent complex, composite degradation scenarios. In this context, we propose OneRestore, a novel transformer-based framework designed for adaptive, controllable scene restoration. The proposed framework leverages a unique cross-attention mechanism, merging degraded scene descriptors with image features, allowing for nuanced restoration. Our model allows versatile input scene descriptors, ranging from manual text embeddings to automatic extractions based on visual attributes. Our methodology is further enhanced through a composite degradation restoration loss, using extra degraded images as negative samples to fortify model constraints. Comparative results on synthetic and real-world datasets demonstrate OneRestore as a superior solution, significantly advancing the state-of-the-art in addressing complex, composite degradations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": "Image restoration is an essential task that entails the recovery of high-quality visuals from compromised inputs. The fidelity of these restored images is vital for the subsequent downstream tasks, such as robot navigation and autonomous driving. Considerable progress has been made in addressing single degradation scenarios by a single model (i.e., One-to-One models, see Fig. 2a), such as low-light conditions [16, 49, 59], haze [21, 41, 57], rain [11, 14, 31], and snow [4, 8, 38]. These methods have achieved remarkable success within their intended contexts. Nevertheless, real-world conditions often involve unpredictable and variable composite degradations that can severely impact image quality and clarity [28]. The" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 237 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 237 + ], + "type": "image", + "image_path": "ce0911d14d93dc0caa79befdfb59a29e29097f3b39cd25b2871ee9ab6c5784fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "type": "text", + "content": "Fig. 1: Our OneRestore allows fully controllable image restoration using scene descriptors derived from both automatic visual attribute extraction (top) and manual text embeddings (bottom)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 397 + ], + "type": "text", + "content": "occurrence of mixed degrading conditions in one image can lead to intricate interactions, posing significant challenges to One-to-One solutions [22,52,54,60]. The necessity for seamless switching and decision-making across specialized methods for each condition proves to be inefficient and ineffective in practical applications. Therefore, there is a growing imperative for robust, all-encompassing image restoration approaches that can universally address multiple forms of degradation through a singular, dynamic framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "text", + "content": "Recent research has been directed toward One-to-Many models, which address various individual degradation factors within a single framework. These models are categorized into: those with partial parameter-sharing (One-to-Many" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "inline_equation", + "content": "_p" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "text", + "content": ") [2,19,28,45] and those with full parameter-sharing (One-to-Many" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "inline_equation", + "content": "_f" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "text", + "content": ") [6,26,35,42, 43]. One-to-Many" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "inline_equation", + "content": "_p" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "text", + "content": " (left of Fig. 2b) restoration methods typically employ models with either multiple heads and tails or a single tail to address various types of image impairments. However, the complexity and size of these models tend to scale with the diversity of degradations they aim to correct, leading to increased training costs and parameter counts. Furthermore, when images suffer from a combination of degrading factors, these models may underperform. Some One-to-Many" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "inline_equation", + "content": "_p" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "text", + "content": " techniques also depend on user intervention to select appropriate restoration weights, which is impractical for automated or dynamic settings. On the other hand, One-to-Many" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "inline_equation", + "content": "_f" + }, + { + "bbox": [ + 130, + 402, + 483, + 666 + ], + "type": "text", + "content": " (right of Fig. 2b) approaches train models using a direct learning strategy, where inputs are images with different individual degradations and the targets are their clear counterparts. However, this direct training approach is designed predominantly for single-category degradation issues, it does not adequately address the complex dynamics that arise when multiple types of impairments affect an image simultaneously. As a result, such methods can falter in practical applications where user control and precise directionality of restoration are crucial. For instance, in cases where users seek to improve low-light imagery, the model might inadvertently prioritize dehazing actions, deviating from user intent." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 228, + 207 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 228, + 207 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 228, + 207 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 228, + 207 + ], + "type": "image", + "image_path": "d3bf3cb80d8a45e50e8cda9e7b379be45966ec8f4f6e29b3033ae91509144a2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 208, + 206, + 217 + ], + "lines": [ + { + "bbox": [ + 156, + 208, + 206, + 217 + ], + "spans": [ + { + "bbox": [ + 156, + 208, + 206, + 217 + ], + "type": "text", + "content": "(a) One-to-One" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 115, + 361, + 206 + ], + "blocks": [ + { + "bbox": [ + 230, + 115, + 361, + 206 + ], + "lines": [ + { + "bbox": [ + 230, + 115, + 361, + 206 + ], + "spans": [ + { + "bbox": [ + 230, + 115, + 361, + 206 + ], + "type": "image", + "image_path": "89c731f4eeff7a0e89a4ba6f770a22f2ca71c8b2ceaaac1535c8230c029872f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 208, + 322, + 217 + ], + "lines": [ + { + "bbox": [ + 268, + 208, + 322, + 217 + ], + "spans": [ + { + "bbox": [ + 268, + 208, + 322, + 217 + ], + "type": "text", + "content": "(b) One-to-Many" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 115, + 480, + 207 + ], + "blocks": [ + { + "bbox": [ + 362, + 115, + 480, + 207 + ], + "lines": [ + { + "bbox": [ + 362, + 115, + 480, + 207 + ], + "spans": [ + { + "bbox": [ + 362, + 115, + 480, + 207 + ], + "type": "image", + "image_path": "9673f91c9c1157ad1d621b7a5deefc002ef0a3d892847725de2faec3d48fea31.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 208, + 456, + 217 + ], + "lines": [ + { + "bbox": [ + 386, + 208, + 456, + 217 + ], + "spans": [ + { + "bbox": [ + 386, + 208, + 456, + 217 + ], + "type": "text", + "content": "(c) One-to-Composite" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 225, + 479, + 247 + ], + "lines": [ + { + "bbox": [ + 131, + 225, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 479, + 247 + ], + "type": "text", + "content": "Fig.2: Structural comparison between One-to-One, One-to-Many, and our One-to-Composite image restoration methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "content": "Drawing from the preceding discussion, we delineate the primary challenges in developing image restoration models for composite degradations in real-world conditions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 310, + 480, + 380 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 138, + 310, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 334 + ], + "type": "text", + "content": "- How can the model be designed to adaptively discern the specific degradation afflicting the input image?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 334, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 480, + 357 + ], + "type": "text", + "content": "- What approaches enable a model to be trained efficiently across a spectrum of degradations while enhancing user control over the restoration process?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 357, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 380 + ], + "type": "text", + "content": "- How can the recovery process be optimized that not only approach the quality of clear images but also maintain distinction from various degraded forms?" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 386, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 553 + ], + "type": "text", + "content": "In response to the challenges outlined, we present a unified model that builds upon established physical corruption models for degradations such as low light, haze, rain, and snow. This model effectively captures the complexity of real-world degradation conditions and forms the basis for constructing our Composite Degradation Dataset (CDD-11), which serves both training and evaluation purposes. To address the issue of composite degradation, our work introduces OneRestore, a transformer-based universal restoration architecture. This novel framework distinguishes itself from conventional methods by embodying a Oneto-Composite strategy, as depicted in Fig. 2c, which allows for precise and controllable outcomes (see Fig. 1). It employs an innovative cross-attention block that fuses scene descriptors, which can be manual text embeddings or derived automatically from visual attributes. We further enhance the model's robustness by integrating a composite degradation restoration loss, using additional degraded images as negative samples to establish stringent lower-bound constraints." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 146, + 554, + 365, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 554, + 365, + 565 + ], + "spans": [ + { + "bbox": [ + 146, + 554, + 365, + 565 + ], + "type": "text", + "content": "Our key contributions are summarized as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 570, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 570, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 570, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 138, + 570, + 480, + 616 + ], + "type": "text", + "content": "- We present the first attempt to unify composite degradations, creating a versatile imaging model that simulates a range of degradation types. This innovation underpins the development of our comprehensive Composite Degradation Dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 617, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 480, + 664 + ], + "type": "text", + "content": "- We propose a universal framework that introduces a novel level of controllability in image restoration, employing a cross-attention mechanism that integrates scene descriptors derived from either manual text embeddings or automated visual attribute extractions." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "content": "- We tailor a composite degradation restoration loss, enhancing the model's ability to discern between various forms of image degradation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 174 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 174 + ], + "type": "text", + "content": "- Our universal model not only sets a new state-of-the-art for composite dataset performance but also competes favorably with One-to-One models in single-degradation scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 190, + 237, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 237, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 237, + 202 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 211, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 377 + ], + "type": "text", + "content": "Complex Degradation Image Restoration. Recent advancements in image restoration have introduced a variety of solutions for complexly degraded images. A notable direction is the partial parameter-sharing One-to-Many approach, employing architectures with multiple heads and tails, and a single backbone, to address various degradation factors. For instance, Li et al. [28] develop an all-weather network that restores scenes affected by haze, rain, and snow, using specialized encoders for each condition and a shared decoder. Chen et al. [2] introduce a transformer-based model for deraining, denoising, and super-resolving images, leveraging a multi-head and multi-tail design. Wang et al. [45] present a knowledge assignment strategy to differentiate and process rain and snow degradations adaptively. While these models represent strides in complex degradation image restoration, they stop short of being truly universal One-to-Many networks. With only partial weight sharing, they address different degradation types through distinct network pathways." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 545 + ], + "type": "text", + "content": "Conversely, universal image restoration has emerged with the One-to-Many methods sharing all network weights, thereby uniformly mitigating multiple degradation types. Examples include the unified two-stage knowledge learning framework by Chen et al. [6], which efficiently processes dehazing, deraining, and desnoring with a singular pretrained weight set. Li et al. [26] present a network capable of comprehensive image restoration without the need for pre-identifying degradation levels. Additionally, Ozdenizci et al. [35] introduce a conditional denoising diffusion model tailored for patch-based image restoration under varied weather conditions. However, due to the significant differences and mutual interference between various degradation factors, this type of method can only solve specific types of degradation and cannot achieve image restoration in composite degradation scenarios. To this end, we propose a universal One-to-Composite image restoration method that allows the model to adapt to arbitrary combinations of composite degradations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Visual Attribute Control. Visual attributes are pivotal in defining image features and enhancing the interpretability of models. They enrich feature representations and provide a hierarchical framework for image analysis. Early research highlighted visual attributes' transformative impact on image generation, recognition, and retrieval [13, 24, 50]. More recent studies underscore their utility in improving object recognition and detection, particularly in data-scarce environments [1, 33, 48, 61]. Visual attributes have also been leveraged for zero-shot learning applications, advancing classification tasks beyond the constraints of labeled datasets [39, 44]. In our study, we harness visual attributes to delineate the intricate degradation patterns present in real-world imaging conditions," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": "enhancing the pertinence and efficacy of image restoration techniques. It is noteworthy that conventional discriminators, when tasked with identifying image degradation, are typically confined to image data, precluding autonomous feature control. Our approach, by contrast, accommodates both images and text descriptions as inputs, thereby facilitating the automatic or manual determination of degradation characteristics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 371, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 371, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 371, + 220 + ], + "type": "text", + "content": "3 Composite Degradation Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 230, + 266, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 266, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 266, + 242 + ], + "type": "text", + "content": "3.1 Imaging Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 344 + ], + "type": "text", + "content": "In real-world scenarios, images are often affected by multiple degrading factors simultaneously, resulting in suboptimal performance from restoration methods designed for isolated degradations. To address the broad spectrum of degradation scenarios, we introduce a versatile imaging model that encapsulates all degradations, including low light, haze, rain, and snow. This model aims to narrow the disparity between simulated and authentic real-world data. Given a degraded image " + }, + { + "bbox": [ + 130, + 248, + 480, + 344 + ], + "type": "inline_equation", + "content": "I(x)" + }, + { + "bbox": [ + 130, + 248, + 480, + 344 + ], + "type": "text", + "content": " and its corresponding clear state " + }, + { + "bbox": [ + 130, + 248, + 480, + 344 + ], + "type": "inline_equation", + "content": "J(x)" + }, + { + "bbox": [ + 130, + 248, + 480, + 344 + ], + "type": "text", + "content": ", the proposed imaging model is mathematically articulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 354, + 481, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 354, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 249, + 354, + 481, + 366 + ], + "type": "interline_equation", + "content": "I (x) = \\mathcal {P} _ {h} \\left(\\mathcal {P} _ {r s} \\left(\\mathcal {P} _ {l} (J (x)))\\right), \\right. \\tag {1}", + "image_path": "79b0e7e9ac5e8c6aa967c264b971b562e724e18cdfc6ca9535732823de867691.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l, \\mathcal{P}_{rs}" + }, + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 130, + 373, + 481, + 421 + ], + "type": "text", + "content": " denote the low light, rain/snow, and haze degradations, respectively. Given that concurrent rain and snow conditions are infrequent, they are not addressed within the scope of this work. A comprehensive exposition of the proposed imaging model is detailed below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "type": "text", + "content": "Low-Light Conditions. Inspired by Retinex theory, low-light degraded image " + }, + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "type": "inline_equation", + "content": "I_{l}(x)" + }, + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "type": "text", + "content": " can be generated by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 454, + 481, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 454, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 231, + 454, + 481, + 479 + ], + "type": "interline_equation", + "content": "I _ {l} (x) = \\mathcal {P} _ {l} (J (x)) = \\frac {J (x)}{L (x)} L (x) ^ {\\gamma} + \\varepsilon . \\tag {2}", + "image_path": "7fa684f10b03923c85a0af912af934bed64a76de2a7d61e860ecf3c6ae7f357e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": "Here, we use a darkening coefficient " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "\\gamma \\in [2,3]" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": " as the brightness controller to adjust the illumination map " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "L(x)" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": " generated via LIME [17]. To simulate poor light conditions realistically, we add the Gaussian noise " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": " with mean " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "inline_equation", + "content": "\\in [0.03,0.08]" + }, + { + "bbox": [ + 130, + 487, + 479, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "text", + "content": "Rain/Snow Streaks. Following the methodologies outlined in [5, 27], our process superimposes rain or snow streaks onto the image before synthesizing haze. To introduce rain streaks " + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "text", + "content": ", we produce the rain-degraded image " + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "inline_equation", + "content": "I_{rs}(x)" + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "text", + "content": " by overlaying " + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{rs}" + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "text", + "content": " onto the low-light image " + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "inline_equation", + "content": "I_l(x)" + }, + { + "bbox": [ + 130, + 535, + 481, + 583 + ], + "type": "text", + "content": ", which is expressed as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 236, + 592, + 481, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 592, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 236, + 592, + 481, + 604 + ], + "type": "interline_equation", + "content": "I _ {r s} (x) = \\mathcal {P} _ {r s} \\left(I _ {l} (x)\\right) = I _ {l} (x) + \\mathcal {R}. \\tag {3}", + "image_path": "fb91a2fbd992a553ea73b3bd02851dfffcf500ebd7ca16a1954cbfd32392ad40.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 612, + 391, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 391, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 391, + 624 + ], + "type": "text", + "content": "For snow streaks " + }, + { + "bbox": [ + 132, + 612, + 391, + 624 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 612, + 391, + 624 + ], + "type": "text", + "content": ", the snow-degraded image is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 210, + 634, + 481, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 634, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 210, + 634, + 481, + 645 + ], + "type": "interline_equation", + "content": "I _ {r s} (x) = \\mathcal {P} _ {r s} \\left(I _ {l} (x)\\right) = I _ {l} (x) (1 - \\mathcal {S}) + M (x) \\mathcal {S}, \\tag {4}", + "image_path": "339528e16da5431d672172e6ae0bc61d0b09d57461ab4f7d8b3df400f89241d1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 653, + 345, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 345, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 345, + 666 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 132, + 653, + 345, + 666 + ], + "type": "inline_equation", + "content": "M(x)" + }, + { + "bbox": [ + 132, + 653, + 345, + 666 + ], + "type": "text", + "content": " being the chromatic aberration map." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 298, + 225 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 298, + 225 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 298, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 298, + 225 + ], + "type": "image", + "image_path": "daf44e536a16f742dade2485957e8ba3a71419b22b4d49cc02457034b5a2fea6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 228, + 247, + 237 + ], + "lines": [ + { + "bbox": [ + 186, + 228, + 247, + 237 + ], + "spans": [ + { + "bbox": [ + 186, + 228, + 247, + 237 + ], + "type": "text", + "content": "(a) Overall Pipeline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 113, + 422, + 227 + ], + "blocks": [ + { + "bbox": [ + 299, + 113, + 422, + 227 + ], + "lines": [ + { + "bbox": [ + 299, + 113, + 422, + 227 + ], + "spans": [ + { + "bbox": [ + 299, + 113, + 422, + 227 + ], + "type": "image", + "image_path": "5133cc4a36aac790ceb6d30ac7a73882055fdfc11d2766e109f5e5b6921c18d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 228, + 413, + 237 + ], + "lines": [ + { + "bbox": [ + 304, + 228, + 413, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 413, + 237 + ], + "type": "text", + "content": "(b) Scene Descriptor Generation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 246, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 291 + ], + "type": "text", + "content": "Fig. 3: Architecture of our OneRestore. (a) Overall pipeline, where 32, 64, 128, and 256 represent the number of channels. (b) Scene descriptor generation, where scene descriptors are fed into each (c) Scene Descriptor-guided Transformer Block (SDTB) by manual text embeddings or automatic extraction based on visual attributes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 114, + 481, + 227 + ], + "blocks": [ + { + "bbox": [ + 425, + 114, + 481, + 227 + ], + "lines": [ + { + "bbox": [ + 425, + 114, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 425, + 114, + 481, + 227 + ], + "type": "image", + "image_path": "03612244224db9547ef4f2d7a3d07af8efcde1069e026f0ad7c810984d074af4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 228, + 468, + 237 + ], + "lines": [ + { + "bbox": [ + 437, + 228, + 468, + 237 + ], + "spans": [ + { + "bbox": [ + 437, + 228, + 468, + 237 + ], + "type": "text", + "content": "(c)SDTB" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 339 + ], + "type": "text", + "content": "Haze Degradations. We incorporate haze degradation by employing the atmospheric scattering model as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 219, + 349, + 481, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 349, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 219, + 349, + 481, + 363 + ], + "type": "interline_equation", + "content": "I (x) = \\mathcal {P} _ {h} \\left(I _ {r s} (x)\\right) = I _ {r s} (x) t + A (1 - t), \\tag {5}", + "image_path": "2fa5ea78581c415422e87bd2578226b39f62ecd4c3f6335591e0b1a0bfecebb1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " denotes the transmission map and " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " represents the atmospheric light. The transmission map " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " is defined by the exponential decay of light, " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "t = e^{-\\beta d(x)}" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " represents the haze density coefficient and " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "d(x)" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " is the scene depth (estimated by MegaDepth [29]). The haze density coefficient " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " is varied within the range of [1.0, 2.0] and the atmospheric light " + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 372, + 482, + 444 + ], + "type": "text", + "content": " is constrained to [0.6, 0.9] for realistic simulation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 464, + 323, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 323, + 476 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 323, + 476 + ], + "type": "text", + "content": "3.2 Composite Degradation Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 486, + 483, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 483, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 483, + 581 + ], + "type": "text", + "content": "Drawing on Eq. (1), we have developed the Composite Degradation Dataset (CDD-11), encompassing 11 categories of image degradations and their clear counterparts. These degraded samples include low (low-light), haze, rain, snow, low+haze, low+rain, low+snow, haze+rain, haze+snow, low+haze+rain, and low+haze+snow. From the RAISE database [10], we selected 1,383 high-resolution clear images for producing 11 composite degradations. They are resized to a uniform resolution of " + }, + { + "bbox": [ + 130, + 486, + 483, + 581 + ], + "type": "inline_equation", + "content": "1080 \\times 720" + }, + { + "bbox": [ + 130, + 486, + 483, + 581 + ], + "type": "text", + "content": ". The overall dataset is split into 13,013 image pairs for training and 2,200 for testing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 602, + 224, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 602, + 224, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 602, + 224, + 615 + ], + "type": "text", + "content": "4 OneRestore" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "The encoder-decoder network, validated as an effective architectural paradigm for image restoration [7, 12, 18], serves as the overall structure for our model, as depicted in Fig. 3a. Building upon the encoder-decoder structure, we devise" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "an innovative Scene Descriptor-guided Transformer Block (SDTB) as the basic unit for feature extraction (Fig. 3c). Our approach diverges from existing transformer-based methods [4,7,41] by concurrently integrating image features and scene descriptors into the SDTB basic block for supporting fine-grained image restoration. In particular, the generation of scene descriptors ranges from manual text embeddings to automatic extraction based on visual attributes, as illustrated in Fig. 3b. Further details on scene descriptor generation and SDTB are elaborated in Sections 4.1 and 4.2. For capturing multi-scale representations, the encoder employs three downsampling operations, mirrored by the decoder's upsampling operations. To be specific, we utilize \"Max-Pooling + Conv\" and \"Bilinear Interpolation + Conv\" for downsampling and upsampling, respectively. Simultaneously, three skip connections and a global residual enhance model convergence by bridging shallow and deep features. Additional details regarding the configuration of our OneRestore can be found in supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 284, + 449, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 284, + 449, + 295 + ], + "spans": [ + { + "bbox": [ + 147, + 284, + 449, + 295 + ], + "type": "text", + "content": "Ultimately, the overall loss " + }, + { + "bbox": [ + 147, + 284, + 449, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 147, + 284, + 449, + 295 + ], + "type": "text", + "content": " for model training is defined as follows" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 301, + 480, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 301, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 196, + 301, + 480, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {1} ^ {s} (J, \\hat {J}) + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {M}} (J, \\hat {J}) + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {c}} (J, \\hat {J}, I, \\{I _ {o} \\}), \\tag {6}", + "image_path": "b04c882d9327463efee521588befce147aa9854c9d380cadcf99868ef8967be4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1^s" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}}" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": " represent the smooth " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": " loss [15], Multi-Scale Structural Similarity (MS-SSIM) loss, and the proposed composite degradation restoration loss (see Sec. 4.3), respectively, " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\alpha_{1-3}" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": " denote the penalty coefficients, " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\hat{J}" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\{I_o\\}" + }, + { + "bbox": [ + 131, + 321, + 482, + 381 + ], + "type": "text", + "content": " are the positive, model output anchor, model input negative, and other negatives, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 397, + 307, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 397, + 307, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 397, + 307, + 409 + ], + "type": "text", + "content": "4.1 Scene Descriptor Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 545 + ], + "type": "text", + "content": "To augment the controllability of our OneRestore, we introduce the scene description embedding as an additional input to the model, alongside the degraded image. As delineated in Fig. 3b, the generation of the scene descriptor offers two selectable modes: the manual mode " + }, + { + "bbox": [ + 130, + 414, + 482, + 545 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 130, + 414, + 482, + 545 + ], + "type": "text", + "content": " and the automatic mode " + }, + { + "bbox": [ + 130, + 414, + 482, + 545 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 130, + 414, + 482, + 545 + ], + "type": "text", + "content": ". In the manual mode, user involvement is necessitated to supply scene description text, generating the corresponding text embedding employed as an input. Conversely, the automatic mode involves the extraction of visual attributes from the image to generate visual embedding, estimating the most proximate text embedding. To quantify the difference between visual and text embeddings and train optimal text and visual embedders, we introduce cosine cross-entropy loss for model weight optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 641 + ], + "type": "text", + "content": "Text Embedder. For the text embedding generation task, we employ a set of 5 scene description texts as input to generate 12 text embeddings. Initially, GloVe [36] is leveraged to generate initial text embeddings for 1 clear scene and 4 scenes with single degradation types. Subsequently, 7 additional composite degradation text embeddings (i.e., low+ haze, low+ rain, low+ snow, haze+ rain, haze+ snow, low+ haze+ rain, and low+ haze+ snow) are generated by averaging the element values in the corresponding text embeddings. A Multi-Layer Perception (MLP) is then employed to refine the 12 text embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Visual Embedder. Taking inspiration from the work of Saini et al. [39], we first resize the input image to " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " and extract image features using a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "ResNet-18 [20] network without average pooling. The initial weights of ResNet-18 are pre-trained on the ImageNet dataset. Subsequently, a single convolutional layer, followed by a dropout layer and a linear layer, is introduced to generate the final visual embedding. By calculating the cosine similarity between visual embeddings and all text embeddings, and selecting the one with the highest similarity, we can automatically generate scene descriptors to achieve image restoration based on visual attribute control." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "content": "Cosine Cross-Entropy. Similar to [39], the score of visual embedding and each text embedding is obtained by \"Cosine Similarity + Softmax\". Given " + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "content": " being the visual and text embeddings, the similarity score " + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "inline_equation", + "content": "S(e_v, e_t)" + }, + { + "bbox": [ + 130, + 200, + 480, + 248 + ], + "type": "text", + "content": " can be given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 256, + 480, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 256, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 244, + 256, + 480, + 292 + ], + "type": "interline_equation", + "content": "\\cos \\left(e _ {v}, e _ {t}\\right) = \\delta \\cdot \\frac {e _ {v} \\cdot e _ {t} ^ {\\top}}{\\| e _ {v} \\| \\| e _ {t} \\|}, \\tag {7}", + "image_path": "1fa5921b22c717396df84c91467734faeeaac8364385434952dc410ed6dc01b6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 281, + 480, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 281, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 244, + 281, + 480, + 315 + ], + "type": "interline_equation", + "content": "S \\left(e _ {v}, e _ {t}\\right) = \\frac {e ^ {\\cos \\left(e _ {v} , e _ {t}\\right)}}{\\sum_ {t _ {i} = 1} ^ {N _ {t}} e ^ {\\cos \\left(e _ {v} , e _ {t _ {i}}\\right)}}, \\tag {7}", + "image_path": "fd4131cb037e1b7bc04990f78326fe985a989d23484d132b1e90e8838c5f84fd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": " being the temperature factor and the number of text embeddings, respectively. Based on the calculated " + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "inline_equation", + "content": "S(e_v, e_t)" + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": ", we use the cross-entropy loss for model training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 377, + 380, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 380, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 380, + 389 + ], + "type": "text", + "content": "4.2 Scene Descriptor-guided Transformer Block" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 397, + 480, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 480, + 504 + ], + "type": "text", + "content": "To further harness the capabilities of scene description embeddings in image restoration tasks and facilitate a more adaptive and nuanced restoration process, we introduce the innovative Scene Descriptor-guided Transformer Block (SDTB), illustrated in Fig. 3c. The SDTB consists of three primary components, namely Scene Descriptor-guided Cross-Attention (SDCA), Self-Attention (SA), and Feed-Forward Network (FFN). Diverging from traditional self-attention mechanisms, the proposed SDCA module utilizes a query generated from scene descriptors instead of queries generated from image features. The mathematical representation of our SDCA can be written as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 216, + 514, + 480, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 514, + 480, + 538 + ], + "spans": [ + { + "bbox": [ + 216, + 514, + 480, + 538 + ], + "type": "interline_equation", + "content": "\\operatorname {S D C A} \\left(\\mathbf {Q} _ {t}, \\mathbf {K}, \\mathbf {V}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t} \\cdot \\mathbf {K} ^ {\\top}}{\\lambda}\\right) \\mathbf {V}, \\tag {8}", + "image_path": "be58bf7884e581d61f75b9adb1e403cda3430f16ca17f71e1514ad602e155f00.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " is a temperature factor, " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " represents the scene descriptor-generated query matrix, " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " denote the image feature-generated key and value matrices, respectively. " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " is obtained from the input scene description embedding " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " via a linear layer. During the process of matrix multiplication between " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": ". To ensure consistency in the number of tokens between " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": ", we resize the original image before performing the reshaping operation to generate " + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": ". By incorporating the SDCA module, each transformer block in OneRestore can effectively integrate scene descriptors and image features for targeted image restoration. As for the SA and FFN modules, we adopt structures akin to those in Restormer [52]. More details of our SDTB are introduced in the supplement." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 481, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 208 + ], + "type": "image", + "image_path": "13540744a2f4e1a7cae699f1a7cd0e6895e301f0fabbac1cd87ca92d4d2722f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 216, + 440, + 228 + ], + "lines": [ + { + "bbox": [ + 173, + 216, + 440, + 228 + ], + "spans": [ + { + "bbox": [ + 173, + 216, + 440, + 228 + ], + "type": "text", + "content": "Fig. 4: Illustration of the composite degradation restoration loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 251, + 369, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 369, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 369, + 264 + ], + "type": "text", + "content": "4.3 Composite Degradation Restoration Loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "text", + "content": "Using contrastive loss has been demonstrated as an effective way of improving image restoration performance [47, 57]. Essentially, contrastive loss views the degraded input image as a negative sample, the clear image as a positive sample, and the restored image as an anchor. This loss function works to minimize the distance between the anchor and the positive sample while maximizing its distance from the negative sample. In the composite degradation image restoration task, the traditional contrastive loss, however, may inadvertently reduce the distance between the anchor and other degraded images. To make the anchor closer to the positive sample while maintaining a clear separation from all degraded negatives, we design a composite degradation restoration loss, illustrated in Fig. 4. Considering " + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "inline_equation", + "content": "I_o" + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\hat{J}" + }, + { + "bbox": [ + 130, + 272, + 482, + 406 + ], + "type": "text", + "content": " as the input negative, other negatives, positive, and anchor, the proposed loss can be written as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 158, + 414, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 414, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 158, + 414, + 481, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} (J, \\hat {J}, I, \\left\\{I _ {o} \\right\\}) = \\sum_ {k = 1} ^ {K} \\xi_ {k} \\frac {\\mathcal {L} _ {1} \\left(V _ {k} (\\hat {J}) , V _ {k} (\\hat {J})\\right)}{\\xi_ {c} \\mathcal {L} _ {1} \\left(V _ {k} (\\hat {J}) , V _ {k} (I)\\right) + \\sum_ {o = 1} ^ {O} \\xi_ {o} \\mathcal {L} _ {1} \\left(V _ {k} (I _ {o}), V _ {k} (\\hat {J})\\right)}, \\tag {9}", + "image_path": "59496c23e3675381bff0cb0d83f0b6e1a173c53b91c07049fe5ad9f7121c93b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "V_{k}(\\cdot)" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " denotes the extraction operation of " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": "-th hidden feature from the VGG-16 pre-trained on the ImageNet, " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " loss, " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " are the numbers of used VGG-16 feature layers and other degraded negatives, " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\xi_{k}, \\xi_{c}" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\xi_{o}" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " denote the hyper-parameters, respectively. In this work, the number of other degraded negatives " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " is 10. We use the output features of 3-rd, 8-th, and 15-th layers from VGG-16 and set " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": ". Furthermore, we consider that the features of each feature layer of each degraded negative have equal importance. Therefore, we set " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\xi_{c} = \\xi_{o} = \\frac{1}{11}" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\xi_{k} = \\frac{1}{3}" + }, + { + "bbox": [ + 130, + 451, + 482, + 572 + ], + "type": "text", + "content": ". According to our ablation study, the proposed loss confers a significant advantage over using a single input degradation as a negative sample, which can easily capture positive sample features." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 133, + 613, + 481, + 663 + ], + "blocks": [ + { + "bbox": [ + 169, + 592, + 443, + 604 + ], + "lines": [ + { + "bbox": [ + 169, + 592, + 443, + 604 + ], + "spans": [ + { + "bbox": [ + 169, + 592, + 443, + 604 + ], + "type": "text", + "content": "Table 1: Settings of model training. \"lr\" denotes the learning rate." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 613, + 481, + 663 + ], + "lines": [ + { + "bbox": [ + 133, + 613, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 133, + 613, + 481, + 663 + ], + "type": "table", + "html": "
TermsText/Visual EmbedderOneRestore
Epoch200120
Initial lr0.00010.0002
lr Decay0.5/50 epochs0.5/20 epochs
Batch Size2564
", + "image_path": "cfda75b32790de40772a8637d1084cd3ed21b9c980d7982b9e95012ca4ca99aa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 169, + 483, + 322 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Comparison of quantitative results on CDD-11 dataset. OneRestore† means using the corresponding scene description text as additional input, and WGWSNet requires the scene type as model input. Red, green, and blue indicate the best, second-best, and third-best results, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 169, + 483, + 322 + ], + "lines": [ + { + "bbox": [ + 132, + 169, + 483, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 169, + 483, + 322 + ], + "type": "table", + "html": "
TypesMethodsVenue & YearPSNR ↑SSIM ↑#Params
Input16.000.6008-
One-to-OneMIRNet [53]ECCV202025.970.847431.79M
MPRNet [54]CVPR202125.470.855515.74M
MIRNetv2 [55]TPAMI202225.370.83355.86M
Restormer [52]CVPR202226.990.864626.13M
DGUNet [34]CVPR202226.920.855917.33M
NAFNet [3]ECCV202224.130.796417.11M
SRUDC [40]ICCV202327.640.86006.80M
Fourmer [58]ICML202323.440.78850.55M
OKNet [9]AAAI202426.330.86054.72M
One-to-ManyAirNet [26]CVPR202223.750.81408.93M
TransWeather [42]CVPR202223.130.781021.90M
WeatherDiff [35]TPAMI202322.490.798582.96M
PromptIR [37]NIPS202325.900.849938.45M
WGWSNet [62]CVPR202326.960.862625.76M
One-to-CompositeOneRestore28.470.87845.98M
OneRestore†28.720.88215.98M
", + "image_path": "57b06a0d0d95d167d47e2676e0560c85d84c547849b37a60d4460c78c1ce3269.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 344, + 230, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 230, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 230, + 358 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 369, + 265, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 265, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 265, + 383 + ], + "type": "text", + "content": "5.1 Experiment Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": "Implementation Details. Our OneRestore is implemented by PyTorch 1.12.0 and trained on a PC with 2 AMD EPYC 7543 32-Core Processors and 8 NVIDIA L40 GPUs. We use the Adam with exponential decay rates being " + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": " for optimization. The training parameter settings for text/visual embedder and OneRestore are shown in Table 1. In particular, all images used for training are cropped into " + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": " image patches with a sampling stride of 200 and randomly flipped by 0, 90, 180, and 270 degrees to generate " + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "inline_equation", + "content": "312\\mathrm{k}" + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": " patch pairs. More details about training and inference can be found in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "content": "Evaluation Datasets. To facilitate a thorough comparison, we execute image restoration experiments across various datasets. These include the Composite Degradation Dataset (CDD-11) crafted by us, established One-to-One benchmarks, and authentic real-world datasets. The standard One-to-One benchmarks are designated for four principal tasks: the LOw-Light dataset (LOL) [46], the REalistic Single Image DEhazing Outdoor Training Set (RESIDE-OTS) [25], the Rain1200 dataset [56], and the Snow100k dataset [30]. Owing to the dearth of specialized datasets for real-world composite degradation, we have curated a selection from the LIME low-light enhancement dataset [17], the RESIDE Realworld Task-driven Testing Set (RESIDE-RTTS) [25], Yang's rainy dataset [51], and the realistically snowy Snow100k dataset [25]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Competitors and Evaluation Metrics. We compare our OneRestore with 9 One-to-One image restoration methods (MIRNet [53], MPRNet [54], MIRNetv2 [55], Restormer [52], DGUNet [34], NAFNet [3], SRUDC [40], Fourmer [58], and OKNet [9]) and 5 One-to-Many image restoration methods (AirNet [26]," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": "TransWeather [42], WeatherDiff [35], PromptIR [37], and WGWSNet [62]). In particular, all compared state-of-the-art methods are retrained on our CDD-11 train set. For PromptIR and WGWSNet, we adjust the number of prompts and the weather-specific parameters of each layer for corresponding to our task, respectively. Furthermore, we adopt the Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM) as evaluation metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 204, + 287, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 204, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 204, + 287, + 217 + ], + "type": "text", + "content": "5.2 Comparison with SOTAs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 331, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 331, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 331, + 402 + ], + "type": "text", + "content": "The quantitative evaluation results on our CDD-11 test dataset are displayed in Table 2. OneRestore† with text embedder achieves the highest performance, followed by OneRestore with visual embedder as the second-best. Both OneRestore† and OneRestore demonstrate substantial leads over the third-ranked methods in terms of both PSNR and SSIM. Further detailed restoration performance comparison of our method and the One-to-Many methods in each degradation scenario is illustrated in Fig. 5 using radar charts. Notably, image restoration for composite degradation scenarios proves challenging. By incorporating scene description em" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 337, + 247, + 481, + 336 + ], + "blocks": [ + { + "bbox": [ + 337, + 247, + 481, + 336 + ], + "lines": [ + { + "bbox": [ + 337, + 247, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 337, + 247, + 481, + 336 + ], + "type": "image", + "image_path": "2f3bc6b251e00659ff9072372f5dd158b5a3e13641f7cda6abdf13592a185b96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 344, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 335, + 344, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 335, + 344, + 482, + 376 + ], + "type": "text", + "content": "Fig.5: Comparison of quantitative results for different degradation scenarios on CDD-11 dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 498 + ], + "type": "text", + "content": "beddings into image restoration models and utilizing the proposed composite degradation restoration loss, our method robustly reconstructs clear images across various situations. For compelling evidence, Fig. 6 provides the visual comparison of image restoration in two composite degradation samples. It is evident that SRUDC, OKNet, AirNet, and WGWSNet are unable to suppress all degenerative interferences, while Restormer and PromptIR are prone to color distortion and noise residual. In contrast, the proposed OneRestore can produce more natural results and fully preserve image textures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 571 + ], + "type": "text", + "content": "Moreover, Fig. 7 displays the image restoration results in two real-world scenarios. Compared with existing One-to-Many image restoration methods [26,35, 37], the proposed OneRestore can simultaneously focus on multiple degradation factors and robustly reconstruct scene details. More results on CDD-11, realworld scenarios, and other One-to-One benchmarks of low-light, hazy, rainy, and snowy scenes are included in supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 239, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 239, + 599 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 239, + 599 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Effectiveness of Network Modules. Table 3 reports the quantitative evaluation of different module configurations. While the Feed-Forward Network (FFN) with convolution as the basic unit serves as a crucial module in the transformer network, it alone fails to generate satisfactory reconstruction results. The ordinary Self-Attention module (SA) significantly strengthens network performance," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 481, + 216 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 481, + 216 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 481, + 216 + ], + "type": "image", + "image_path": "b218a344f70fced369512260e59c87b18705cdc09eac3fdaa87955ec1a89956d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 224, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 132, + 224, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 482, + 246 + ], + "type": "text", + "content": "Fig. 6: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) synthetic samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 260, + 481, + 331 + ], + "blocks": [ + { + "bbox": [ + 133, + 260, + 481, + 331 + ], + "lines": [ + { + "bbox": [ + 133, + 260, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 481, + 331 + ], + "type": "image", + "image_path": "344e23f352e4b9baa19335a5f028e1aefaad89bfa4d796b6923b483d4ae9bdea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 340, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 132, + 340, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 482, + 374 + ], + "type": "text", + "content": "Fig. 7: Comparison of image restoration on low+haze+rain (top) and low+haze+snow (bottom) samples in real-world scenarios. * represents the utilization of original weights published in the author's code." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 484 + ], + "type": "text", + "content": "but this method regrets that the model fails to be flexibly controlled. Compared to depending only on SA, relying exclusively on Scene Description-guided Cross-Attention (SDCA) utilizing scene description embedding yields better outcomes. This further confirms that addressing complex composite degradations can be done more effectively and flexibly by integrating degradation scene descriptors into the model. Ultimately, the best performance can be achieved by using a combination of SDCA, SA, and FFN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 330, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 330, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 330, + 666 + ], + "type": "text", + "content": "Selection of Description Embedding Strategies. In this section, we compare the image restoration performance by using different strategies to generate scene descriptor, including Classifier (utilizing the embedding before the last linear layer of the retrained AlexNet [23] as the scene descriptor), Visual Embedder (employing visual embedding for model training), Visual Embedder‡ (deriving text embedding via visual embedder), and Text Embedder (employing text embedding for model training). The quantitative evaluation results are presented in Table 4. It is observed that employing Classifier and Visual Embedder to generate scene descriptors" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 510, + 479, + 594 + ], + "blocks": [ + { + "bbox": [ + 340, + 510, + 479, + 594 + ], + "lines": [ + { + "bbox": [ + 340, + 510, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 340, + 510, + 479, + 594 + ], + "type": "image", + "image_path": "1975813ece73deb6c035bf2ef205dcb59bbb05a99dde20ac023e1beaf6f728aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 603, + 482, + 636 + ], + "lines": [ + { + "bbox": [ + 335, + 603, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 335, + 603, + 482, + 636 + ], + "type": "text", + "content": "Fig.8: Comparison of image restoration on different description embedding strategies." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 158, + 481, + 205 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Ablation study for different model configurations. SDCA, SA, and FFN denote the scene description-guided cross-attention, self-attention, and feed-forward network, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 158, + 481, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 158, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 481, + 205 + ], + "type": "table", + "html": "
SDCASAFFNPSNR ↑SSIM ↑Controllability
✓24.810.8607
✓✓27.190.8697
✓✓27.930.8767✓
✓✓✓28.720.8821✓
", + "image_path": "e38c18451995c77ee745d876990221508dca8b1b15c084bad65552e72cb814b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 221, + 481, + 279 + ], + "blocks": [ + { + "bbox": [ + 133, + 221, + 481, + 279 + ], + "lines": [ + { + "bbox": [ + 133, + 221, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 221, + 481, + 279 + ], + "type": "image", + "image_path": "8cbc7b31a0ae2f37c7326a6ccfacb78e5e528a8bd96cba6918db0a56202e2540.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 287, + 483, + 310 + ], + "lines": [ + { + "bbox": [ + 131, + 287, + 483, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 483, + 310 + ], + "type": "text", + "content": "Fig. 9: Comparison of image restoration on " + }, + { + "bbox": [ + 131, + 287, + 483, + 310 + ], + "type": "inline_equation", + "content": "low + haze + rain" + }, + { + "bbox": [ + 131, + 287, + 483, + 310 + ], + "type": "text", + "content": " (top) and " + }, + { + "bbox": [ + 131, + 287, + 483, + 310 + ], + "type": "inline_equation", + "content": "low + haze + snow" + }, + { + "bbox": [ + 131, + 287, + 483, + 310 + ], + "type": "text", + "content": " (bottom) synthetic samples by using different texts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "text", + "content": "yields comparable outcomes. However, these approaches have a drawback in that each image has a distinct embedding as input, making it challenging for the model to recognize the cause of degradation. Fig. 8 displays a classic case comparison using different embedding strategies, where the abnormal artifacts of the results produced by Classifier and Visual Embedder are caused by the model's fuzzy identification of degradation factors. Moreover, the restoration model lacks controllability. The discrepancy between Visual Embedder" + }, + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "text", + "content": " and Text Embedder stems from errors in scene descriptor estimation. Visual Embedder" + }, + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "text", + "content": " has an accuracy of " + }, + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "inline_equation", + "content": "97.55\\%" + }, + { + "bbox": [ + 130, + 339, + 482, + 506 + ], + "type": "text", + "content": " on the test dataset, with misestimation occurring when some degradation types are insignificant. For instance, a haze+snow scene may be evaluated as a haze or snow scene. In contrast, the text embedding-based approach is superior in describing degradation scenes and recovering images, using a fixed number of scene descriptors. This allows the model to employ degraded scene descriptors as switches to achieve controlled restoration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 342, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 342, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 342, + 665 + ], + "type": "text", + "content": "Effectiveness of Loss Functions. Table 5 presents a performance comparison of different loss functions. Notably, the integration of the MS-SSIM loss is instrumental in elevating SSIM scores at the cost of reduced PSNR performance. In contrast, the CL loss demonstrates minimal impact, providing only marginal improvements in SSIM and PSNR. Upon replacing CL with CDRL, a substantial enhancement is observed in both PSNR and SSIM. The proposed CDRL loss yields an output that closely resembles the clear image while avoiding other forms of degradation." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 348, + 534, + 482, + 615 + ], + "blocks": [ + { + "bbox": [ + 348, + 534, + 482, + 615 + ], + "lines": [ + { + "bbox": [ + 348, + 534, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 348, + 534, + 482, + 615 + ], + "type": "image", + "image_path": "2c5fd263c7c705af5b4b73760ba662c7b5ba5e1ef6fcddf6a42b9cd9253ee00c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 622, + 482, + 644 + ], + "lines": [ + { + "bbox": [ + 347, + 622, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 347, + 622, + 482, + 644 + ], + "type": "text", + "content": "Fig.10: Visualization of visual embeddings based on t-SNE." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 136, + 481, + 185 + ], + "blocks": [ + { + "bbox": [ + 178, + 114, + 434, + 126 + ], + "lines": [ + { + "bbox": [ + 178, + 114, + 434, + 126 + ], + "spans": [ + { + "bbox": [ + 178, + 114, + 434, + 126 + ], + "type": "text", + "content": "Table 4: Ablation study for description embedding strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 136, + 481, + 185 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 481, + 185 + ], + "type": "table", + "html": "
ModelsPSNR ↑SSIM ↑Controllability
Classifier28.190.8783
Visual Embedder28.240.8781
Visual Embedder‡28.470.8784
Text Embedder28.720.8821✓
", + "image_path": "8d69424b2dc0020dd2562979f6e71ff3b161eac6b6361010618569b8257f98a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 227, + 483, + 275 + ], + "blocks": [ + { + "bbox": [ + 132, + 194, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 217 + ], + "type": "text", + "content": "Table 5: Ablation study for different loss functions. CL and CDRL denote the ordinary contrastive loss and the proposed composite degradation restoration loss, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 227, + 483, + 275 + ], + "lines": [ + { + "bbox": [ + 133, + 227, + 483, + 275 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 483, + 275 + ], + "type": "table", + "html": "
Smooth l1MS-SSIMCLCDRLPSNR ↑SSIM ↑
✓28.160.8633
✓✓27.540.8708
✓✓✓27.610.8723
✓✓✓28.720.8821
", + "image_path": "1f727791b321d53d5c53922e0a614a583ce1ad391d27503bd9793ce9320c970b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 296, + 340, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 340, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 340, + 308 + ], + "type": "text", + "content": "5.4 Model Control via Scene Descriptor" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "text", + "content": "To demonstrate the controllability of the scene descriptor in our OneRestore, we conduct experiments on two intricate synthetic degradation scenarios, as depicted in Fig. 9. By manually adjusting specific text to generate corresponding degraded scene description embeddings, our model can selectively focus on different degradation factors, achieving targeted restorations. Our approach excels in controllable restoration without the need for multiple sets of weight screening, making it more concise compared to other One-to-Many methods that rely on partial parameter sharing. Meanwhile, our method can support adaptive restoration through the estimation of text embeddings based on visual attribute extraction. Fig. 10 employs the t-SNE [32] to visualize the embeddings generated by the visual embedder on the CDD-11 train and test sets. Obviously, the constructed visual embedder can clearly distinguish between different scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 474, + 220, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 220, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 220, + 487 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Our proposed OneRestore framework represents a significant step forward in image restoration, synthesizing a range of degradation patterns to accurately simulate complex scenarios. The innovative use of a scene descriptor-guided cross-attention block within a transformer-based model facilitates an adaptable and fine-grained restoration process, yielding results that surpass prior methods in both synthetic and real-world datasets. However, despite its advancements, OneRestore exhibits limitations in processing extremely high-density corruption scenarios and high-complex corruption scenarios containing unconsidered degradations, where the model's predictive capabilities can be challenged. Future work will aim to enhance the robustness of OneRestore against such extreme conditions and to reduce computational overhead, further extending its applicability and efficiency. Our findings underscore the potential of universal scene restoration, with the caveat that continued refinement is essential for tackling the full spectrum of real-world image degradation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgement. This project is supported by the Guangdong Natural Science Funds for Distinguished Young Scholar under Grant 2023B1515020097 and National Research Foundation Singapore under the AI Singapore Programme under Grant AISG3-GV-2023-011." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "text", + "content": "1. Bertasius, G., Torresani, L.: Cobe: Contextualized object embeddings from narrated instructional video. In: NeurIPS. pp. 15133-15145 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "2. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "text", + "content": "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 480, + 315 + ], + "type": "text", + "content": "4. Chen, S., Ye, T., Liu, Y., Liao, T., Jiang, J., Chen, E., Chen, P.: Msp-former: Multi-scale projection transformer for single image desnowing. In: ICASSP. pp. 1-5 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 358 + ], + "type": "text", + "content": "5. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: CVPR. pp. 4196-4205 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 360, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 403 + ], + "type": "text", + "content": "6. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 424 + ], + "type": "text", + "content": "7. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR. pp. 5896-5905 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 446 + ], + "type": "text", + "content": "8. Chen, Z., Sun, Y., Bi, X., Yue, J.: Lightweight image de-snowing: A better trade-off between network capacity and performance. Neural Networks 165, 896-908 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 448, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 480, + 468 + ], + "type": "text", + "content": "9. Cui, Y., Ren, W., Knoll, A.: Omni-kernel network for image restoration. In: AAAI. vol. 38, pp. 1426-1434 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 491 + ], + "type": "text", + "content": "0. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: Raise: A raw images dataset for digital image forensics. In: MMSys. pp. 219-224 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "1. Du, Y., Deng, J., Zheng, Y., Dong, J., He, S.: Dsdnet: Toward single image deraining with self-paced curricular dual stimulations. CVIU 230, 103657 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "type": "text", + "content": "2. Du, Y., Han, G., Tan, Y., Xiao, C., He, S.: Blind image denoising via dynamic dual learning. IEEE TMM 23, 2139-2152 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 535, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 555 + ], + "type": "text", + "content": "3. Ferrari, V., Zisserman, A.: Learning visual attributes. In: NeurIPS. pp. 433-440 (2007)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "type": "text", + "content": "4. Fu, X., Xiao, J., Zhu, Y., Liu, A., Wu, F., Zha, Z.J.: Continual image deraining with hypergraph convolutional networks. IEEE TPAMI 45(8), 9534-9551 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 381, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 381, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 381, + 588 + ], + "type": "text", + "content": "5. Girshick, R.: Fast r-cnn. In: ICCV. pp. 1440-1448 (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "6. Guo, X., Hu, Q.: Low-light image enhancement via breaking down the darkness. IJCV 131(1), 48-66 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "7. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "8. Guo, Y., Gao, Y., Liu, W., Lu, Y., Qu, J., He, S., Ren, W.: Scenet: Self-paced semicurricular attention network for non-homogeneous image dehazing. In: CVPRW. pp. 1884-1893 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. Han, J., Li, W., Fang, P., Sun, C., Hong, J., Armin, M.A., Petersson, L., Li, H.: Blind image decomposition. In: ECCV. pp. 218-237 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "21. Hoang, T., Zhang, H., Yazdani, A., Monga, V.: Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In: CVPR. pp. 1670-1679 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "text", + "content": "22. Jiao, J., Tu, W.C., Liu, D., He, S., Lau, R.W.H., Huang, T.S.: Formnet: Formatted learning for image restoration. IEEE TIP 29, 6302-6314 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "text", + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. CACM 60(6), 84-90 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "text", + "content": "24. Kumar, N., Berg, A., Belhumeur, P.N., Nayar, S.: Describable visual attributes for face verification and image search. IEEE TPAMI 33(10), 1962-1977 (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "25. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP 28(1), 492-505 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "26. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "text", + "content": "27. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR. pp. 1633-1642 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "28. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "text", + "content": "29. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR. pp. 2041-2050 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "30. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP 27(6), 3064-3073 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "31. Luo, Y., Huang, Q., Ling, J., Lin, K., Zhou, T.: Local and global knowledge distillation with direction-enhanced contrastive learning for single-image deraining. Knowledge-Based Syst. 268, 110480 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "text", + "content": "32. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. JMLR 9(11), 2579-2605 (2008)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "text", + "content": "33. Meng, Q., Shin'ichi, S.: Adinet: Attribute driven incremental network for retinal image classification. In: CVPR. pp. 4033-4042 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "34. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: CVPR. pp. 17399-17410 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "35. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI 45(8), 10346-10357 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "36. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. pp. 1532-1543 (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "text", + "content": "37. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: PromptIR: Prompting for all-in-one image restoration. NeurIPS 36 (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "38. Quan, Y., Tan, X., Huang, Y., Xu, Y., Ji, H.: Image desnowing via deep invertible separation. IEEE TCSVT 33(7), 3133-3144 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "39. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: CVPR. pp. 13658-13667 (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "40. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: ICCV. pp. 12580-12589 (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "41. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP 32, 1927-1941 (2023)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "42. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 621 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "43. Wang, C., Zheng, Z., Quan, R., Sun, Y., Yang, Y.: Context-aware pretraining for efficient blind image decomposition. In: CVPR. pp. 18186-18195 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "text", + "content": "44. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: CVPR. pp. 11197-11206 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "text", + "content": "45. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR. pp. 3677-3686 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "text", + "content": "46. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: BMVC. pp. 1-12 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "47. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR. pp. 10551-10560 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "48. Xu, W., Xian, Y., Wang, J., Schiele, B., Akata, Z.: Attribute prototype network for zero-shot learning. In: NeurIPS. pp. 21969-21980 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "49. Xu, X., Wang, R., Lu, J.: Low-light image enhancement via structure modeling and guidance. In: CVPR. pp. 9893-9903 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "50. Yan, X., Yang, J., Sohn, K., Lee, H.: Attribute2image: Conditional image generation from visual attributes. In: ECCV. pp. 776-791 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "51. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "52. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "53. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV. pp. 492-511 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "54. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR. pp. 14821-14831 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for fast image restoration and enhancement. IEEE TPAMI 45(2), 1934-1948 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "56. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "57. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR. pp. 5785-5794 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "58. Zhou, M., Huang, J., Guo, C.L., Li, C.: Fourmer: An efficient global modeling paradigm for image restoration. In: ICML. pp. 42589-42601 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "59. Zhou, M., Wu, X., Wei, X., Xiang, T., Fang, B., Kwong, S.: Low-light enhancement method based on a retina model for structure preservation. IEEE TMM pp. 1-13 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "60. Zhou, Y., Xu, Y., Du, Y., Wen, Q., He, S.: Propulse: Learning progressive encoders of latent semantics in gans for photo upsampling. IEEE TIP 31, 1230-1242 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "61. Zhu, Y., Min, W., Jiang, S.: Attribute-guided feature learning for few-shot image recognition. IEEE TMM 23, 1200-1209 (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "62. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "OneRestore: A Universal Restoration Framework for Composite Degradation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_content_list.json b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02c42237b477882592bc85687426d884ca7b41b4 --- /dev/null +++ b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_content_list.json @@ -0,0 +1,1654 @@ +[ + { + "type": "text", + "text": "OneTrack: Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers", + "text_level": 1, + "bbox": [ + 240, + 140, + 763, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qitai Wang $^{1,2}$ , Jiawei He $^{2}$ , Yuntao Chen $^{3}$ , and Zhaoxiang Zhang $^{1,2,3,4}$", + "bbox": [ + 218, + 233, + 781, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Future Technology, University of Chinese Academy of Sciences (UCAS)", + "bbox": [ + 225, + 261, + 777, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA) {wangqitai2020,zhaoxiang.zhang}@ia.ac.cn,jwhe2024@gmail.com", + "bbox": [ + 228, + 277, + 771, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Center for Artificial Intelligence and Robotics, HKISI, CAS chenyuntao08@gmail.com", + "bbox": [ + 297, + 304, + 705, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 351, + 330, + 650, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing end-to-end trackers for vision-based 3D perception suffer from performance degradation due to the conflict between detection and tracking tasks. In this work, we get to the bottom of this conflict, which was vaguely attributed to incompatible task-specific object features previously. We find the conflict between the two tasks lies in their partially conflicted classification gradients, which stems from their subtle difference in positive sample assignments. Based on this observation, we propose to coordinate those conflicted gradients from object queries with contradicted polarity in the two tasks. We also dynamically split all object queries into four groups based on their polarity in the two tasks. Attention between query sets with conflicted positive sample assignments is masked. The tracking classification loss is modified to suppress inaccurate predictions. To this end, we propose OneTrack, the first one-stage joint detection and tracking model that bridges the gap between detection and tracking under a unified object feature representation. On the nuScenes camera-based object tracking benchmark, OneTrack outperforms previous works by $6.9\\%$ AMOTA on the validation set and by $3.1\\%$ AMOTA on the test set.", + "bbox": [ + 259, + 383, + 743, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Tracking $\\cdot$ Camera-based Detection and Tracking $\\cdot$ End-to-end Tracking $\\cdot$ Gradient Coordination", + "bbox": [ + 259, + 647, + 740, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 702, + 375, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reasoning about the location and trajectory of surrounding objects is a fundamental task for autonomous driving and robotic navigation systems. Due to the robustness and low cost of cameras, vision-based 3D detection and tracking have received widespread attention from the research community [14,22,23,26,36,37]. Recently, several approaches [21, 31, 48] have been proposed to improve 3D multi-object tracking (MOT) by jointly optimizing the detection and tracking pipeline. Based on transformer architecture, those end-to-end 3D trackers use", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/d337e36d1dafa074696ba345466c9ff306c474823b3aa0dc3d92487a22bf11f2.jpg", + "table_caption": [ + "Table 1: Comparison with previous works trying to solve the conflict between detection and tracking. *: DQTrack relies on frozen pre-trained detectors when compared with state-of-the-art trackers." + ], + "table_footnote": [], + "table_body": "
Methodsettinge2eone-stagedetection supervisiontrain from scratch
MOTRv2 [50]2D××××
MOTRv3 [45]2D✓✓×✓
DQTrack [21]3D✓×✓✓*
OneTrack3D✓✓✓✓
", + "bbox": [ + 299, + 198, + 699, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "track queries to track objects, being free from the dependence on offline detectors and post-processing operations.", + "bbox": [ + 212, + 333, + 782, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, most end-to-end 3D trackers suffer from the optimization conflict between detection and tracking, a problem also widely recognized in 2D MOT studies [45, 50]. This conflict results in the inferior precision and recall of end-to-end trackers compared to their baseline detectors. Previous works roughly attribute the conflict between two tasks to differences in required features for detection and tracking and attempted to alleviate this problem. Some of them train their model as detectors before the tracking training process [31, 33] or take an off-the-shelf detector as guidance [50]. Some separate the detection and association stages as well as their object feature representation to avoid conflict, such as DQTrack [21]. Although much progress has been made, training a one-stage end-to-end tracker from scratch remains challenging. It is also challenging to train trackers jointly under both detection and tracking supervision. We summarize recent works in Table 1. In this work, we aim to accomplish the true joint training of detection and tracking from scratch with a one-stage model and unified feature representation for both tasks.", + "bbox": [ + 212, + 366, + 787, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct a pilot study to investigate how and where the two tasks conflict during model optimization. As illustrated in Table 2, only jointly supervising the detector with an extra classification loss computed under the ground truth assignments for tracking will lead to severe performance degradation of detectors. Therefore, we conclude that the conflict between the two tasks originates from their conflicting classification gradients.", + "bbox": [ + 212, + 595, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we precisely identify and coordinate the conflicted classification gradient between the two tasks, which originates from their partially conflicted positive sample definition. We propose OneTrack, the first model capable of achieving the true joint training of detection and tracking in a unified, one-stage model. OneTrack follows the \"tracking by track queries\" paradigm. During the training process, we assign ground truth annotations to object queries twice based on positive sample definitions for detection or tracking. We dynamically categorize all object queries into four groups based on their polarity in two tasks. We have two lightweight classification heads in OneTrack which are respectively responsible for classifying detection or tracking positive samples. Queries that are", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/9be26db0d9e9b5b926d95bc1ac7a702c299deb9588e5258e4a4ef1f66d7f9e7f.jpg", + "table_caption": [ + "Table 2: Pilot study on supervising 3D detector under additional ground truth assignments for tracking. The stronger color means more degradation in performance. Experiments indicate the conflict between detection and tracking tasks lies in their classification supervision." + ], + "table_footnote": [], + "table_body": "
Detection ClassificationDetection RegressionTracking ClassificationTracking RegressionAMOTA↑IDS↓NDS↑mAP↑
✓✓××--0.5700.479
××✓✓0.305121--
✓✓✓×0.4043630.5120.432
✓✓×✓0.41785120.5660.473
✓✓✓✓0.4052550.5280.428
", + "bbox": [ + 215, + 212, + 802, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "positive samples only for one task will only backpropagate classification gradients in the respective head to reduce conflicted gradients. We dynamically mask all attention between conflicted positive queries in each decoder layer to prevent unexpected query competition. We also adjust the tracking loss to suppress inaccurate predictions.", + "bbox": [ + 212, + 361, + 782, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 240, + 438, + 581, + 452 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal that the conflict between detection and tracking tasks lies in their diverged definition of positive samples and their partially conflicted classification gradients.", + "- We propose OneTrack, the first model capable of completely addressing the conflict between detection and tracking under a unified object feature representation. OneTrack can be trained as both a detector and a tracker from scratch in a single training stage and can perform tracking and detection in a one-stage fashion. The detection performance of OneTrack is comparable to detectors, while its tracking performance surpasses all previous 3D trackers.", + "- Our method establishes a new state-of-the-art on nuScenes MOT benchmark, surpassing previous methods by over $3.1\\%$ AMOTA." + ], + "bbox": [ + 225, + 460, + 784, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 648, + 395, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Query Propagation in MOT", + "text_level": 1, + "bbox": [ + 215, + 679, + 493, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early works in 2D or 3D MOT most follow the \"tracking by detection\" paradigm [4] and associate detection results across frames through post-processing [1, 2, 8, 9, 12, 13, 16, 19, 25, 35, 40-42, 44, 46, 49]. Based on DETR [6] which introduced object queries to detect objects, MOTR [47] and Trackformer [30] extended the concept of object query to track query by propagating object queries across frames. By tracking objects through track queries, MOTR and Trackformer can perform end-to-end joint 2D detection and tracking without post-processing. However, they seriously suffer from the conflict between detection and tracking. MOTRv2 [50] and MOTRv3 [45] concentrate on this conflict. MOTRv2 proposes", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using an off-the-shelf detector to guide the tracker, which is effective but damages the end-to-end fashion of the tracker. MOTRv3 proposes to balance the label assignment between detection queries and track queries to achieve a better tradeoff between the detection and tracking performance of the model. MUTR3D [48] and PF-Track [31] applied the \"tracking by track queries\" paradigm to 3D MOT. MUTR3D [48] performs camera-based tracking in an end-to-end fashion by introducing 3D track queries to model spatial and appearance coherent tracks. PF-Track [31] further refine the tracks and predict the trajectories of objects with cross-object and cross-frame attention between object queries. Recently, DQTrack [21] proposes to separate the detection process and association process of trackers to avoid their conflicts in object feature representation. None of the previous joint detection and tracking approaches can truly bridge the gap between detection and tracking with a one-stage model and a unified object feature representation. Neither can they train the model from scratch and get rid of the reliance on pre-trained detectors while detecting and tracking objects with high precision and recall performance.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Camera-based 3D MOT", + "text_level": 1, + "bbox": [ + 215, + 417, + 460, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera-based 3D MOT has gained remarkable progress recently thanks to the developments in camera-based 3D detection and depth estimation [11, 14, 15, 22, 23, 26, 27, 36-38, 43, 51]. Early methods in camera-based 3D MOT perform tracking in 2D first and then lift the tracks to 3D space [34]. CC-3DT [10] fuses multi-view object features to enhance associating objects across different views. QD-3DT [13] perform 2D association first and enhance the instance association with the depth order and motions of 3D objects. TripletTrack [29] extracts local object feature embeddings and motion descriptors with CNN or LSTM to measure the affinity between objects. MUTR3D [48], PF-Track [31], and DQTrack [21] recently established new state-of-the-art tracking performance by applying the transformer-based tracking architecture in 3D.", + "bbox": [ + 212, + 450, + 787, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 648, + 330, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The model design of OneTrack is conceptually concise. We add an additional lightweight classification branch to a basic transformer-based tracker. Both ground truth assignments for the detection and tracking tasks are employed to supervise the model. We begin by reviewing the design of transformer-based trackers in Sec. 3.1. Then in Sec. 3.2, we propose to identify the conflicted positive sample assignments between detection and tracking through query grouping and prevent the contradicted gradient from back-propagating. In Sec. 3.3, we introduce the dynamic attention mask to further prevent gradient conflict between the two tasks. Additionally, the tracking classification loss is modified to suppress inaccurate boxes predicted by track queries, as described in Sec. 3.4.", + "bbox": [ + 212, + 688, + 792, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Reviewing the \"Tracking by Track Queries\" Paradigm", + "text_level": 1, + "bbox": [ + 215, + 146, + 702, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As previous transformer-based trackers [21,30,31,45,47,48,50], OneTrack detect and track 3D objects with object queries. At each timestamp $t$ , given multiview images $\\mathbf{I}$ , the 3D detection task aims to predict a set of 3D bounding boxes $\\mathbf{B} = \\{\\mathbf{b}^1, \\mathbf{b}^2, \\dots, \\mathbf{b}^n\\}$ and their confidence scores $\\mathbf{S} = \\{s^1, s^2, \\dots, s^n\\}$ to capture 3D objects. Meanwhile, 3D MOT requires a consistent ID for each object across frames. We represent the outputs for the 3D MOT task by $\\{\\mathbf{B}, \\mathbf{S}, \\mathbf{ID}\\}$ , including the predicted boxes, confidence scores, and ID of predictions.", + "bbox": [ + 212, + 167, + 787, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In each frame, OneTrack first receives a set of object queries $\\mathbf{Q} = \\{\\mathbf{q}^1,\\mathbf{q}^2,\\dots \\}$ . Following previous works [21, 31, 47, 48], $\\mathbf{Q}$ consists of track queries propagated from the previous frame and a fixed number of new detection queries $\\mathbf{Q}^d$ . The query propagation process can be formulated as follows:", + "bbox": [ + 212, + 273, + 787, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {t} = \\hat {\\mathbf {Q}} _ {t - 1} \\cup \\mathbf {Q} ^ {d}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 342, + 784, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{Q}}_{t - 1}$ represents the propagated queries from the last frame. We denote $\\mathbf{Q}_t$ as $\\mathbf{Q}$ in the subsequent descriptions for simplicity.", + "bbox": [ + 212, + 369, + 784, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In each frame, given the multi-view image features $\\mathbf{F}$ extracted with image encoder, OneTrack utilizes a transformer decoder to extract object features. The decoder comprises six transformer decoder layers. In the $i^{\\mathrm{th}}$ layer $\\mathbf{Q}$ is updated as follows:", + "bbox": [ + 212, + 400, + 785, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\operatorname {D e c o d e r L a y e r} _ {i} (\\mathbf {F}, \\mathbf {Q}, \\mathbf {M}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 469, + 784, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{M}$ represents the attention mask for the self-attention process in decoder layers. In end-to-end trackers, attention between all object queries is usually not masked. Then the object queries are fed into the classification and regression heads to generate prediction boxes. Ground truths are then assigned to the predictions for the calculation of classification loss and regression loss. Recent works [21,31] usually preserve historical track queries from a few past frames as tracking memory. These stored queries serve as references to the tracking process in the current frame.", + "bbox": [ + 212, + 492, + 787, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of associating boxes across frames with post-processing techniques, the \"tracking by track queries\" paradigm tracks objects with track queries propagated across frames. Therefore during the training of trackers, a ground truth instance is assigned to a specific track query in future frames once they are associated in the current frame, regardless of the accuracy of the assigned query's predictions in future frames. We denote these ground truth-query pairs as \"locked\" [45] pairs, and the remaining ground truths and queries are denoted as \"free\" ground truths or queries. In contrast, for detectors, positive samples are simply defined as the most precise prediction in each frame. Hence, a positive sample for detection may be classified as a negative sample in tracking and be suppressed, and vice versa, as illustrated in Fig.1. This originates from ground truth objects being assigned to different object queries in the two tasks. This conflict will lead to conflicted classification gradients between the two tasks, hindering the convergence of the model. We found that those conflicted gradients are primarily responsible for the precision and recall degradation of end-to-end trackers.", + "bbox": [ + 212, + 614, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f12da4d09ecadf1a37c3a42eb4d7c09506729af94a8352e6fb93e14db1faad9.jpg", + "image_caption": [ + "Fig. 1: Difference in positive sample definition for query-based detector and tracker. With object query propagated from the last frame, the end-to-end tracker ensures that queries assigned as positive samples in the previous frame (like $\\mathrm{Q}_{\\mathrm{C}}^{\\prime}$ ) maintain their pairing with the same ground truth in the current frame. Meanwhile, the detector always takes the most precise predictions (like $\\mathrm{Q}_{\\mathrm{X}}$ ) as positive samples. In such case, the polarity of $\\mathrm{Q}_{\\mathrm{C}}^{\\prime}$ and $\\mathrm{Q}_{\\mathrm{X}}$ as training samples in the two tasks is opposite." + ], + "image_footnote": [], + "bbox": [ + 316, + 147, + 694, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Solving the Conflict Between Detection and Tracking via Gradient Coordination", + "text_level": 1, + "bbox": [ + 214, + 454, + 740, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first try a trivial and straightforward solution to avoid the classification gradient conflict between the two tasks. We attach two classification heads to the model, one for detection classification and the other for tracking classification. Since there is no obvious conflict in box regression for the detection or tracking tasks, we use a shared head for regressing 3D boxes in the two tasks. Then all predicted boxes are assigned to ground truths twice following the positive sample definition in detection or tracking as follows:", + "bbox": [ + 212, + 496, + 787, + 603 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} _ {\\mathrm {d e t}} = \\arg \\min _ {\\sigma} \\sum_ {k} ^ {N} \\mathcal {L} _ {\\operatorname {m a t c h}} \\left(\\mathbf {g} ^ {k}, \\mathbf {b} ^ {\\sigma (k)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 614, + 785, + 656 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} _ {\\mathrm {t r k}} ^ {\\text {f r e e}} = \\arg \\min _ {\\sigma} \\sum_ {k} ^ {M} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\mathbf {g} ^ {k}, \\mathbf {b} ^ {\\sigma (k)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 659, + 785, + 700 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} _ {\\mathrm {t r k}} = \\hat {\\sigma} _ {\\mathrm {t r k}} ^ {\\mathrm {f r e e}} \\cup \\sigma_ {\\mathrm {t r k}} ^ {\\mathrm {l o c k e d}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 700, + 785, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\{\\mathbf{g}^k\\}_{k=1}^N$ is $N$ ground truths(GTs), in which $M$ of them are free GTs for tracking. $\\{\\mathbf{b}^i\\}_{i=1}^n$ is the set of predicted boxes. $\\sigma$ is the possible permutation of $N$ elements, following the notation in [6]. For example, $(\\mathbf{g}^k, \\mathbf{b}^{\\sigma(k)})$ means the $k^{th}$ ground truth box is assigned to the $\\sigma(k)^{th}$ predicted box. $\\mathcal{L}_{\\mathrm{match}}$ denotes the matching cost between a ground truth and a prediction from an object query. $\\hat{\\sigma}_{\\mathrm{det}}$ and $\\hat{\\sigma}_{\\mathrm{trk}}$ are the ground truth assignments for detection and tracking, respectively. $\\sigma_{\\mathrm{trk}}^{\\mathrm{locked}}$ stands for the locked ground truth and query pairs inherited", + "bbox": [ + 214, + 733, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from the last frame. Classification losses and regression losses are then computed for gradient back-propagation in each decoder layer.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, we found that merely separating the classification heads for two tasks and supervising the model under both detection and tracking ground truth assignments can neither obtain a satisfying detector nor tracker, as shown in Table 4. Clearly, simply separating the classification heads cannot solve the conflict between the sample classification in detection and tracking. This is because as previously mentioned, a part of positive samples for training detectors will be assigned as negative samples for training trackers, and vice versa. Therefore even when the classification heads of two tasks are separated, classification gradients on those samples will still be conflicted when back-propagated to the decoder layers. To address this, we propose to decompose the classification gradients in both tasks to identify and avoid their partial conflict on those conflicted samples. We refer to this design as classification gradient coordination between the two tasks.", + "bbox": [ + 212, + 176, + 784, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We dynamically categorize object queries into four query groups in each decode layer based on their polarity in both tasks: positive for both tasks; positive for detection and negative for tracking; negative for detection and positive for tracking; and negative for both tasks. We denote those four groups of queries as $\\mathbf{Q}^{\\mathrm{pp}}$ , $\\mathbf{Q}^{\\mathrm{pn}}$ , $\\mathbf{Q}^{\\mathrm{np}}$ , and $\\mathbf{Q}^{\\mathrm{nn}}$ . This proposed query grouping can be illustrated as follows:", + "bbox": [ + 212, + 373, + 784, + 462 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {\\mathrm {p p}} = \\left\\{q ^ {k} | k \\in \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\in \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 476, + 784, + 491 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {\\mathrm {p n}} = \\left\\{q ^ {k} | k \\in \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\notin \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 496, + 784, + 512 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {\\mathrm {n p}} = \\left\\{q ^ {k} | k \\notin \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\in \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 518, + 784, + 532 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {\\mathrm {n n}} = \\left\\{q ^ {k} \\mid k \\notin \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\notin \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 539, + 784, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Through this, the classification gradients of the two tasks can also be decomposed as the sum of detection or tracking classification gradients on each group of queries. The classification gradient conflict only lies in the second and third groups of queries, as illustrated in Fig. 2(a). Therefore we propose to remove the gradient of $\\mathbf{Q}^{\\mathrm{pn}}$ contributed to the tracking classification loss and the gradient of $\\mathbf{Q}^{\\mathrm{np}}$ contributed to the detection classification loss, as illustrated in Fig. 2(b). Through this, the partial classification gradient conflict of the two tasks on $\\mathbf{Q}^{\\mathrm{pn}}$ and $\\mathbf{Q}^{\\mathrm{np}}$ is avoided. $\\mathbf{Q}^{\\mathrm{pp}}$ and $\\mathbf{Q}^{\\mathrm{nn}}$ still contribute gradient in both heads since the partial classification gradients upon them are not contradictory.", + "bbox": [ + 212, + 566, + 784, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Dynamic Attention Mask", + "text_level": 1, + "bbox": [ + 214, + 724, + 472, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following previous end-to-end trackers [21, 30, 31, 45, 47, 48, 50], OneTrack generates predictions in a set-prediction fashion. Hence the model is encouraged to suppress redundant predictions regarding the same ground truth object through self-attention between object queries in each decoder layer. As previously mentioned, $\\mathbf{Q}^{\\mathrm{pn}}$ and $\\mathbf{Q}^{\\mathrm{np}}$ originates from the same ground truth object being assigned to different object queries in the two tasks. To avoid unexpected competition", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8bdaa410e6798843dd3da7acc3b3f2ffc02b4024ba17812da56df5e856fbcae1.jpg", + "image_caption": [ + "Fig. 2: Backpropagation of Classification Gradients. Gradient coordination prevents the classification gradient conflict between detection and tracking. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 276, + 147, + 723, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between task-specific queries regarding the same object, in each layer of the decoder, we dynamically mask the attention between $\\mathbf{Q}^{\\mathrm{pn}}$ and $\\mathbf{Q}^{\\mathrm{np}}$ . Attention between all other query groups is not masked. Through this, $\\mathbf{Q}^{\\mathrm{pn}}$ and $\\mathbf{Q}^{\\mathrm{np}}$ are not directly exposed to each other during the forward process.", + "bbox": [ + 212, + 431, + 787, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We utilize the ground truth assignment results in each decoder layer to generate the self-attention mask for the next layer. Before the first decoder layer where detection queries are initialized as empty queries, we use distances between their initial reference points and centers of ground truths to perform bipartite matching. After each decoder layer, we perform bipartite matching based on matching costs between ground truths and predicted 3D boxes defined in [37]. We employ the Hungarian algorithm [17] for bipartite matching. As for outputs, the detection results of OneTrack include predictions from $\\mathbf{Q}^{\\mathrm{pp}}$ , $\\mathbf{Q}^{\\mathrm{pn}}$ and $\\mathbf{Q}^{\\mathrm{nn}}$ . The tracking results include predictions from $\\mathbf{Q}^{\\mathrm{pp}}$ , $\\mathbf{Q}^{\\mathrm{pn}}$ and $\\mathbf{Q}^{\\mathrm{nn}}$ . We illustrate the pipeline of OneTrack in Fig. 3.", + "bbox": [ + 212, + 491, + 787, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During inference, we categorize object queries into four groups based on the detection and tracking confidence scores of their predictions. Queries with high detection or tracking confidence scores over $T_{p}$ are classified as positive samples for the respective tasks. We set $T_{p} = 0.2$ in our experiments.", + "bbox": [ + 212, + 643, + 787, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At last, in both the training or inference process, we propagate the Top- $K$ object queries with high confidence scores from both detection and tracking results to the next frame as $\\hat{\\mathbf{Q}}_t$ . In particular, we will still propagate the \"locked\" track queries to the future frames even if they are not in the Top- $K$ object queries. This is for addressing challenging tracking scenarios, such as an object being occluded for several frames. We propagate \"locked\" track queries through this for up to $\\mathrm{A_{age}} = 3$ frames into future frames.", + "bbox": [ + 212, + 704, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following [31,36], we also preserve the propagated object queries as the memory queue for the detection or tracking process for up to 4 frames. The mem", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57a7132ff371c2f733db751d0b580166e2d013ab1032d9adada585c64d0924ee.jpg", + "image_caption": [ + "Fig. 3: Training pipeline of OneTrack. For each decoder layer in OneTrack, We identify object queries into four groups based on their positivity as training samples in detection and tracking tasks. Task-specific queries that are positive samples only for one task will only backpropagate gradients in the respective task-specific classification head to prevent conflicted classification gradients. We dynamically mask all attention across positive sample queries only for detection or tracking. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 779, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ory queues for the detection and tracking processes are updated independently. Those memory queries are provided to the decoder for performing hybrid attention proposed by [36]. As keys and values for the hybrid attention layer, detection and tracking memory queries are masked for $\\mathbf{Q}^{\\mathrm{np}}$ and $\\mathbf{Q}^{\\mathrm{pn}}$ , respectively. We show the design of decoder layers in OneTrack in Fig. 4.", + "bbox": [ + 212, + 455, + 787, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As demonstrated in Table 4, with the proposed dynamic hybrid-attention mask, the true joint training of detection and tracking can be achieved. This allows for training detection and tracking in a single training stage and inferring them in a single forward pass with a unified model. This feature of OneTrack significantly reduces its training cost and inference latency.", + "bbox": [ + 212, + 532, + 787, + 607 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Modified Tracking Classification Loss", + "text_level": 1, + "bbox": [ + 214, + 631, + 568, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In addition to introducing the true joint training of detection and tracking to reduce the detection capability degradation of trackers, we also adjusted the tracking classification loss to suppress the in fact inaccurate predictions. As previously mentioned, previous end-to-end trackers regard track queries whose predictions are associated with ground truths as positive samples equally, regardless of the precision of their current predictions. This encourages the tracker to assign high confidence to the in fact inaccurate predictions from track queries. Therefore we propose to restrain the positive assignment of those in fact inaccurate predictions.", + "bbox": [ + 212, + 657, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In previous end-to-end trackers, the assignment result of a ground truth object $g$ can be denoted as a one-hot vector: $A_{g}^{n\\times 1} = [0,\\dots ,1,\\dots ,0]$ , where $n$ is the number of object queries and the ground truth object is exclusively", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ff3a3d84509c258c6433f8b8b2e37f0483d22552795a5b86c9e464f0eb238912.jpg", + "image_caption": [ + "Fig. 4: Design of decoder layers in OneTrack. In addition to current object queries, queries in detection memory and tracking memory are fed to the hybrid attention layer as keys and values. As keys and values, $\\mathbf{Q}^{\\mathrm{pn}}$ and detection memory are masked to $\\mathbf{Q}^{\\mathrm{np}}$ . As keys and values, $\\mathbf{Q}^{\\mathrm{np}}$ and tracking memory are masked to $\\mathbf{Q}^{\\mathrm{pn}}$ ." + ], + "image_footnote": [], + "bbox": [ + 299, + 143, + 692, + 353 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "assigned to the $i^{th}$ object query. This assignment is unaffected by the matching cost between the $i^{th}$ object query and $g$ if they have been associated in recent frames.", + "bbox": [ + 212, + 449, + 787, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In OneTrack, we propose to suppress the assignment weight between ground truths and queries that are locked but have high matching costs. We adjust the assignment result of locked ground truth $g$ as follows:", + "bbox": [ + 212, + 494, + 787, + 542 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nA _ {g} ^ {n \\times 1} = [ 0, \\dots , W _ {g i}, \\dots , 0 ], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 551, + 785, + 571 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nW _ {g i} = - \\frac {1}{\\gamma} * \\operatorname {M a x} \\left(\\left(\\operatorname {d i s t} (g, q ^ {i}) - 0. 5\\right), 0\\right) + 1, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 573, + 785, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathrm{dist}(g, q^i)$ represents the distance between the centers of ground truth box $g$ and the predicted box of the locked query $q^i$ . $\\gamma$ is a hyper-parameter greater than zero. The assigned weight $W_{gi}$ between $g$ and $q^i$ decreases linearly from 1.0 as $\\mathrm{dist}(g, q^i)$ increases from $0.5\\mathrm{m}$ . The assignment weight serves as the weighting factor for calculating box regression loss and the targets for predicting tracking confidence score. Instead of Focal Loss [24], we employ the Quality Focal Loss proposed by [20] to supervise the tracking classification head, given the continuous $0\\sim 1$ confidence label.", + "bbox": [ + 212, + 614, + 787, + 737 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 761, + 375, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we first introduce our detailed experimental setup. Then we provide comparisons with recent works on nuScenes camera-based 3D tracking benchmark. Ablation studies on each component are presented in the last part.", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/498c1d6fdb2d504113ffc7537b1b117f21f703b29bc043c6b20c5eccd92fce5f.jpg", + "table_caption": [ + "Table 3: Compare with recent works on nuScenes dataset. $\\dagger$ : trained from scratch. OneTrack-F / OneTrack-S: OneTrack trained under full-resolution / half-resolution as described in Sec. 4.2." + ], + "table_footnote": [], + "table_body": "
Backbonee2eResolutionAMOTA↑AMOTP↓MOTA↑RECALL↑IDS↓
Validation Split
QD3DT [13]R101✓1600×9000.2421.5180.21839.9%5646
MUTR3D [48]R101✓1600×9000.2941.4980.26742.7%3822
CC-3DT [10]R101×1600×6400.4291.2570.38553.4%2219
PF-Track-S [31]V2-99✓800×3200.4081.3430.37650.7%166
PF-Track-F [31]V2-99✓1600×6400.4791.2270.43559.0%181
DQTrack [21]V2-99✓800×3200.4461.251--1193
OneTrack-S†V2-99✓800×3200.4921.1220.40958.8%315
OneTrack-F†V2-99✓1600×6400.5481.0880.47961.8%389
Test Split
QD3DT [13]R101✓1600×6400.2171.5500.19837.5%6856
MUTR3D [48]R101✓1600×6400.2701.4940.23541.1%6018
SRCN3D [32]V2-99×1600×6400.3981.3170.35953.8%3334
CC-3DT [10]R101×1600×6400.4101.2740.35753.8%3334
PF-Track [31]V2-99✓1600×6400.4341.2520.37853.8%249
DQTrack [21]V2-99✓1600×6400.5231.0960.44462.2%1204
OneTrack-F†V2-99✓1600×6400.5541.0210.46160.8%481
", + "bbox": [ + 243, + 199, + 754, + 449 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 215, + 477, + 439, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate our proposed method on the nuScenes dataset. nuScenes [5] dataset is a large-scale autonomous driving benchmark containing 1000 multi-modal videos. Videos recorded by six cameras are divided into 700, 150, and 100 scenes for training, validation, and testing, respectively. 3D box annotations of 10 object classes are provided for keyframes at $2\\mathrm{Hz}$ . For nuScenes 3D tracking benchmark, we report AMOTA, AMOTP [39], MOTA [3], Recall and identity switches (IDS). For the 3D detection task, we report mean Average Precision (mAP) and nuScenes Detection Score (NDS).", + "bbox": [ + 212, + 503, + 787, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 647, + 457, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our implementation is mainly based on StreamPETR [36]. We define two implementation settings: The full-resolution setting for state-of-the-art comparison and the half-resolution setting for ablation studies. We conduct experiments using V2-99 [18] as the image backbone. OneTrack is trained using AdamW [28] optimizer with a batch size of 8 and a base learning rate of 4e-4. The cosine annealing policy is employed to adjust the learning rate. We use 644 initialized detection queries in each frame and propagate the Top-256 object queries into the next frame. When training, we record all matched ground truth and object query pairs as \"locked\" pairs for label assignment. An object query matched with a ground truth will be recorded as a track query. As previously mentioned, a \"locked\" pair will be preserved in memory for up to $A_{\\mathrm{age}} = 3$ frames, even if", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b6bdb9bdbf58bec8d2b192cba3914ada0319d10cee6b55b35908c321299f64b9.jpg", + "table_caption": [ + "Table 4: Ablation study on components breakdown." + ], + "table_footnote": [], + "table_body": "
Training taskInitialized from pre-trained detectorGradient coordinationDynamic High-cost assign masksuppressAMOTA↑IDS↓NDS↑mAP↑
detection------0.5700.479
tracking×--×0.3051210.4550.313
tracking✓--×0.3754660.5030.409
joint××××0.4252460.5170.410
joint×✓××0.4412310.5550.462
joint×✓✓×0.4692610.5750.484
joint×✓✓✓0.4923150.5660.478
", + "bbox": [ + 220, + 171, + 784, + 292 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the respective track query is not in the Top- $K$ object queries in each frame. For track queries that failed to predict confident prediction over $\\mathrm{A}_{\\mathrm{age}}$ frames, its \"locked\" ground truth will be assigned to other \"free\" object queries. During inference, we initialize an object query as a track query if it predicts a 3D box with confidence over $T_{\\mathrm{conf}} = 0.4$ . Additionally, all track queries are preserved and propagated across frames for up to $\\mathrm{A}_{\\mathrm{age}} = 3$ frames. Predictions from the same track query propagated across frames will be assigned the same tracking ID. We output all detection predictions with confidence scores over 0.05 and all tracking predictions with confidence scores over 0.3. We introduce other setting-specific training configurations as follows.", + "bbox": [ + 212, + 321, + 785, + 473 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "OneTrack-F: We crop input images from the initial resolution of $1600 \\times 900$ to $1600 \\times 640$ . We then conduct end-to-end training for 48 epochs. OneTrack-S: We first crop the input images to the resolution of $1600 \\times 640$ and downsample them to $800 \\times 320$ . Models are trained for 24 epochs.", + "bbox": [ + 212, + 473, + 785, + 535 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 215, + 556, + 552, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Table 3, OneTrack outperforms all previous trackers in both the validation and test set of nuScenes dataset by a significant margin. On the validation set, OneTrack surpasses previous state-of-the-art PF-Track [31] by $10.2\\%$ AMOTA. On the test set, OneTrack surpasses DQTrack [21] by $3.3\\%$ AMOTA. Notably, OneTrack reports fewer than half of the identity switches compared to those reported by DQTrack. It is also worth mentioning that OneTrack is the only model in Table 3 that is trained from scratch.", + "bbox": [ + 212, + 580, + 787, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablations", + "text_level": 1, + "bbox": [ + 215, + 709, + 341, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablation studies on each proposed component with OneTrack-S on the nuScenes validation set.", + "bbox": [ + 212, + 734, + 785, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Components breakdown. The breakdown analysis of proposed components is presented in Table 4. Here \"det\" or \"trk\" in the first row stands for only training the model for detection or tracking tasks. And \"joint\" stands for jointly training the model for both detection and tracking tasks, with an extra classification head attached to the model. As shown in Table 4, the model trained solely", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/05c7a253a57041435d97f5902819b770cbc10364af1d2b03952b707895c5a2ae.jpg", + "image_caption": [ + "Fig. 5: Ablation on hyper-parameters. (a) On parameter $\\gamma$ in the modified tracking losses. (b) On the max preserved frames of track queries $A_{\\mathrm{age}}$ . (c) On the score threshold $T_{\\mathrm{conf}}$ for track query initialization." + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 411, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/39a1941da2f64dc7cac4c2da3d61991fa2d78216e77ecf9e10764e5c189518b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 145, + 604, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3aaa808cd8f770f592e01716dfb9d8db137d4c291894b48f0dcc006ff9519da5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 145, + 800, + 255 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "for the tracking task suffers from severe degradation in detection performance, which results in their inferior detection performance compared to the detector. Initializing the tracker from a detector checkpoint or trivially supervising the model with both detection and tracking losses can only slightly alleviate this problem. Our proposed components including classification gradient coordination, dynamic attention mask, and high-cost tracking assignment suppress, are all validated as highly effective in achieving the true joint training of detection and tracking tasks. It is noteworthy that OneTrack without modifying its tracking loss achieves an even slightly better detection performance than the model trained solely as a detector. This observation suggests the potential for further enhancing the detection process through the integrated tracking process in a unified model.", + "bbox": [ + 212, + 339, + 787, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Modified tracking loss. In Fig. 5(a), we conduct ablation studies on the hyper-parameter $\\gamma$ in the modified tracking loss for suppressing positive assignments of inaccurate predictions. OneTrack achieves optimal tracking performance when $\\gamma$ is set to 7.5. While the detection performance of OneTrack shows improvement as $\\gamma$ increases, this gain becomes limited when $\\gamma$ exceeds 7.5.", + "bbox": [ + 212, + 520, + 787, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Track age. We conduct ablation studies on the maximum preservation age $A_{\\mathrm{age}}$ of track queries. As illustrated in Fig. 5(b), a longer track age will lead to more identity switches and decreased detection performance. We speculate that this is because propagating too many outdated track queries will mislead the object feature extraction process in the current frame.", + "bbox": [ + 212, + 597, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Confidence threshold for track query initialization. We assess the impact of $T_{\\mathrm{conf}}$ in Fig. 5(c). The AMOTA performance will slightly increase as $T_{\\mathrm{conf}}$ increases, which results from the more accurate tracking predictions on average. However, a higher $T_{\\mathrm{conf}}$ reduces the initialization of track queries from detection queries with low confidence. This limitation hinders tracking objects under poor observation conditions, resulting in a higher number of identity switch cases.", + "bbox": [ + 212, + 672, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Separated classification heads. As shown in Table 5, when its two classification heads share weights except for the final linear layer, the tracking and detection performance of OneTrack is slightly worse than when the classification heads are independent. We speculate that this is because the separation of classification heads leads to the easier convergence of both heads.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b64a2bcd6e184a4d794ee78e68e9b8794e9fb929a74bbcd8a0c39be065be900c.jpg", + "table_caption": [ + "Table 5: Ablation on the separated classification heads for two tasks." + ], + "table_footnote": [], + "table_body": "
Cls headsAMOTA↑IDS↓NDS↑mAP↑
shared49.03760.5590.473
separated49.23150.5660.478
", + "bbox": [ + 313, + 170, + 683, + 229 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c086ce47ead7e5bfe2df88efc5eb5e9d2861e7e60a289b1ffa58f6826c56ff08.jpg", + "table_caption": [ + "Table 6: Computational cost comparison with recent works. All training time costs and inference latencies are measured on a single A100 GPU. *: PF-Track is first trained as a detector and then as a tracker for 12 epochs each. We train DQTrack and OneTrack from scratch for 24 epochs." + ], + "table_footnote": [], + "table_body": "
MethodBackboneResolutionTraining time costInference latency
PF-TrackV2-99800×32034h+82h*124.4ms
DQTrackV2-99800×320176h115.2ms
OneTrackV2-99800×32087h83.7ms
", + "bbox": [ + 308, + 309, + 689, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Computational cost. To ensure fair comparison, we apply gradient checkpointing [7] on the decoder layers of all trackers. We measure their training cost and inference latency on a single A100 GPU. As shown in Table 6, the training cost of OneTrack is only half that of previous state-of-the-art trackers. The inference latency of OneTrack is also significantly lower. We attribute the low computational cost of OneTrack to its elegant model design. We do not extend the tracker with various plug-in modules [31] or separate the detection and tracking stages of the tracker [21], which will introduce additional computational burdens. Notably, OneTrack is trained from scratch while PF-Track and DQTrack rely on extra training stages for achieving high tracking performance.", + "bbox": [ + 212, + 422, + 787, + 574 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 594, + 359, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we delve into the optimization conflict between detection and tracking in end-to-end trackers, which arises from the partially conflicted classification gradients between two tasks. Based on this observation, we propose to identify the conflicted positive samples between two tasks through query grouping and solve the conflicted gradients with the proposed gradient coordination, dynamic attention mask, and suppression tracking positive samples of low quality. On the nuScenes dataset, OneTrack outperforms all previous end-to-end trackers.", + "bbox": [ + 212, + 623, + 787, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 750, + 393, + 767 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ali, A., Jalil, A., Niu, J., Zhao, X., Rathore, S., Ahmed, J., Aksam Iftikhar, M.: Visual object tracking—classical and contemporary approaches. Frontiers of Computer Science 10, 167–188 (2016)", + "2. Benbarka, N., Schröder, J., Zell, A.: Score refinement for confidence-based 3d multi-object tracking. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8083-8090. IEEE (2021)", + "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing 2008, 1-10 (2008)", + "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016)", + "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)", + "6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "7. Chen, T., Xu, B., Zhang, C., Guestrin, C.: Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174 (2016)", + "8. Chiu, H.k., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: ICRA (2021)", + "9. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4870-4880 (2023)", + "0. Fischer, T., Yang, Y., Kumar, S., Sun, M., Yu, F.: CC-3DT: panoramic 3d object tracking via cross-camera fusion. In: CoRL. Proceedings of Machine Learning Research, vol. 205, pp. 2294–2305. PMLR (2022)", + "1. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science 18(5), 185706 (2024)", + "2. He, J., Huang, Z., Wang, N., Zhang, Z.: Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5299-5309 (2021)", + "3. Hu, H.N., Yang, Y.H., Fischer, T., Darrell, T., Yu, F., Sun, M.: Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "4. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)", + "5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)", + "6. Kim, A., Osep, A., Leal-Taixe, L.: Eagermot: 3d multi-object tracking via sensor fusion. In: 2021 IEEE International conference on Robotics and Automation (ICRA). pp. 11315-11321 (2021)", + "7. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)", + "19. Li, J., Ding, Y., Wei, H.L., Zhang, Y., Lin, W.: Simpletrack: Rethinking and improving the jde approach for multi-object tracking. Sensors 22(15), 5863 (2022)", + "20. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020)", + "21. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18302-18311 (2023)", + "22. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023)", + "23. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)", + "24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "25. Liu, C., Chen, X.F., Bo, C.J., Wang, D.: Long-term visual tracking: review and experimental comparison. Machine Intelligence Research 19(6), 512-530 (2022)", + "26. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)", + "27. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023)", + "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "29. Marinello, N., Proesmans, M., Van Gool, L.: Tripletrack: 3d object tracking using triplet embeddings and lstm. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4499-4509 (2022)", + "30. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8844-8854 (2022)", + "31. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17928-17938 (2023)", + "32. Shi, Y., Shen, J., Sun, Y., Wang, Y., Li, J., Sun, S., Jiang, K., Yang, D.: Srcn3d: Sparse r-cnn 3d surround-view camera object detection and tracking for autonomous driving. arXiv preprint arXiv:2206.14451 (2022)", + "33. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10860-10869 (2021)", + "34. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: ICCV. pp. 10840-10849. IEEE (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Wang, Q., Chen, Y., Pang, Z., Wang, N., Zhang, Z.: Immortal tracker: Tracklet never dies. arXiv preprint arXiv:2111.13672 (2021)", + "36. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3621-3631 (2023)", + "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)", + "38. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5096-5105 (2023)", + "39. Weng, X., Kitani, K.: A baseline for 3d multi-object tracking. arXiv preprint arXiv:1907.03961 1(2), 6 (2019)", + "40. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. IROS (2020)", + "41. Weng, X., Wang, Y., Man, Y., Kitani, K.: Gnn3dmot: Graph neural network for 3d multi-object tracking with multi-feature learning. arXiv preprint arXiv:2006.07327 (2020)", + "42. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: 2017 IEEE international conference on image processing (ICIP). pp. 3645-3649. IEEE (2017)", + "43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)", + "44. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. CVPR (2021)", + "45. Yu, E., Wang, T., Li, Z., Zhang, Y., Zhang, X., Tao, W.: Motrv3: Releasefetch supervision for end-to-end multi-object tracking. arXiv preprint arXiv:2305.14298 (2023)", + "46. Zaech, J.N., Liniger, A., Dai, D., Danelljan, M., Van Gool, L.: Learnable online graph representations for 3d multi-object tracking. IEEE Robotics and Automation Letters 7(2), 5103-5110 (2022)", + "47. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022)", + "48. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4537-4546 (2022)", + "49. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision 129, 3069-3087 (2021)", + "50. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22056-22065 (2023)", + "51. Zhao, H., Zhang, J., Chen, Z., Yuan, B., Tao, D.: On robust cross-view consistency in self-supervised monocular depth estimation. Machine Intelligence Research 21(3), 495-513 (2024)" + ], + "bbox": [ + 212, + 146, + 785, + 811 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OneTrack", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_model.json b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..505bcba591bca596d6f0f200d447b03fbe5397e0 --- /dev/null +++ b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_model.json @@ -0,0 +1,2258 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.241, + 0.141, + 0.764, + 0.207 + ], + "angle": 0, + "content": "OneTrack: Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.234, + 0.782, + 0.252 + ], + "angle": 0, + "content": "Qitai Wang\\(^{1,2}\\), Jiawei He\\(^{2}\\), Yuntao Chen\\(^{3}\\), and Zhaoxiang Zhang\\(^{1,2,3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.262, + 0.778, + 0.278 + ], + "angle": 0, + "content": "1 School of Future Technology, University of Chinese Academy of Sciences (UCAS)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.278, + 0.772, + 0.305 + ], + "angle": 0, + "content": "\\(^{2}\\) NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA) {wangqitai2020,zhaoxiang.zhang}@ia.ac.cn,jwhe2024@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.305, + 0.707, + 0.332 + ], + "angle": 0, + "content": "3 Center for Artificial Intelligence and Robotics, HKISI, CAS chenyuntao08@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.332, + 0.651, + 0.347 + ], + "angle": 0, + "content": "4 Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.385, + 0.744, + 0.635 + ], + "angle": 0, + "content": "Abstract. Existing end-to-end trackers for vision-based 3D perception suffer from performance degradation due to the conflict between detection and tracking tasks. In this work, we get to the bottom of this conflict, which was vaguely attributed to incompatible task-specific object features previously. We find the conflict between the two tasks lies in their partially conflicted classification gradients, which stems from their subtle difference in positive sample assignments. Based on this observation, we propose to coordinate those conflicted gradients from object queries with contradicted polarity in the two tasks. We also dynamically split all object queries into four groups based on their polarity in the two tasks. Attention between query sets with conflicted positive sample assignments is masked. The tracking classification loss is modified to suppress inaccurate predictions. To this end, we propose OneTrack, the first one-stage joint detection and tracking model that bridges the gap between detection and tracking under a unified object feature representation. On the nuScenes camera-based object tracking benchmark, OneTrack outperforms previous works by \\(6.9\\%\\) AMOTA on the validation set and by \\(3.1\\%\\) AMOTA on the test set." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.648, + 0.741, + 0.676 + ], + "angle": 0, + "content": "Keywords: 3D Tracking \\(\\cdot\\) Camera-based Detection and Tracking \\(\\cdot\\) End-to-end Tracking \\(\\cdot\\) Gradient Coordination" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.376, + 0.719 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Reasoning about the location and trajectory of surrounding objects is a fundamental task for autonomous driving and robotic navigation systems. Due to the robustness and low cost of cameras, vision-based 3D detection and tracking have received widespread attention from the research community [14,22,23,26,36,37]. Recently, several approaches [21, 31, 48] have been proposed to improve 3D multi-object tracking (MOT) by jointly optimizing the detection and tracking pipeline. Based on transformer architecture, those end-to-end 3D trackers use" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison with previous works trying to solve the conflict between detection and tracking. *: DQTrack relies on frozen pre-trained detectors when compared with state-of-the-art trackers." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.199, + 0.7, + 0.3 + ], + "angle": 0, + "content": "
Methodsettinge2eone-stagedetection supervisiontrain from scratch
MOTRv2 [50]2D××××
MOTRv3 [45]2D✓✓×✓
DQTrack [21]3D✓×✓✓*
OneTrack3D✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.784, + 0.364 + ], + "angle": 0, + "content": "track queries to track objects, being free from the dependence on offline detectors and post-processing operations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.593 + ], + "angle": 0, + "content": "However, most end-to-end 3D trackers suffer from the optimization conflict between detection and tracking, a problem also widely recognized in 2D MOT studies [45, 50]. This conflict results in the inferior precision and recall of end-to-end trackers compared to their baseline detectors. Previous works roughly attribute the conflict between two tasks to differences in required features for detection and tracking and attempted to alleviate this problem. Some of them train their model as detectors before the tracking training process [31, 33] or take an off-the-shelf detector as guidance [50]. Some separate the detection and association stages as well as their object feature representation to avoid conflict, such as DQTrack [21]. Although much progress has been made, training a one-stage end-to-end tracker from scratch remains challenging. It is also challenging to train trackers jointly under both detection and tracking supervision. We summarize recent works in Table 1. In this work, we aim to accomplish the true joint training of detection and tracking from scratch with a one-stage model and unified feature representation for both tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.687 + ], + "angle": 0, + "content": "We conduct a pilot study to investigate how and where the two tasks conflict during model optimization. As illustrated in Table 2, only jointly supervising the detector with an extra classification loss computed under the ground truth assignments for tracking will lead to severe performance degradation of detectors. Therefore, we conclude that the conflict between the two tasks originates from their conflicting classification gradients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we precisely identify and coordinate the conflicted classification gradient between the two tasks, which originates from their partially conflicted positive sample definition. We propose OneTrack, the first model capable of achieving the true joint training of detection and tracking in a unified, one-stage model. OneTrack follows the \"tracking by track queries\" paradigm. During the training process, we assign ground truth annotations to object queries twice based on positive sample definitions for detection or tracking. We dynamically categorize all object queries into four groups based on their polarity in two tasks. We have two lightweight classification heads in OneTrack which are respectively responsible for classifying detection or tracking positive samples. Queries that are" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Pilot study on supervising 3D detector under additional ground truth assignments for tracking. The stronger color means more degradation in performance. Experiments indicate the conflict between detection and tracking tasks lies in their classification supervision." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.213, + 0.803, + 0.334 + ], + "angle": 0, + "content": "
Detection ClassificationDetection RegressionTracking ClassificationTracking RegressionAMOTA↑IDS↓NDS↑mAP↑
✓✓××--0.5700.479
××✓✓0.305121--
✓✓✓×0.4043630.5120.432
✓✓×✓0.41785120.5660.473
✓✓✓✓0.4052550.5280.428
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.784, + 0.437 + ], + "angle": 0, + "content": "positive samples only for one task will only backpropagate classification gradients in the respective head to reduce conflicted gradients. We dynamically mask all attention between conflicted positive queries in each decoder layer to prevent unexpected query competition. We also adjust the tracking loss to suppress inaccurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.439, + 0.583, + 0.453 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.462, + 0.784, + 0.506 + ], + "angle": 0, + "content": "- We reveal that the conflict between detection and tracking tasks lies in their diverged definition of positive samples and their partially conflicted classification gradients." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.596 + ], + "angle": 0, + "content": "- We propose OneTrack, the first model capable of completely addressing the conflict between detection and tracking under a unified object feature representation. OneTrack can be trained as both a detector and a tracker from scratch in a single training stage and can perform tracking and detection in a one-stage fashion. The detection performance of OneTrack is comparable to detectors, while its tracking performance surpasses all previous 3D trackers." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.598, + 0.784, + 0.627 + ], + "angle": 0, + "content": "- Our method establishes a new state-of-the-art on nuScenes MOT benchmark, surpassing previous methods by over \\(3.1\\%\\) AMOTA." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.396, + 0.665 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.494, + 0.696 + ], + "angle": 0, + "content": "2.1 Query Propagation in MOT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Early works in 2D or 3D MOT most follow the \"tracking by detection\" paradigm [4] and associate detection results across frames through post-processing [1, 2, 8, 9, 12, 13, 16, 19, 25, 35, 40-42, 44, 46, 49]. Based on DETR [6] which introduced object queries to detect objects, MOTR [47] and Trackformer [30] extended the concept of object query to track query by propagating object queries across frames. By tracking objects through track queries, MOTR and Trackformer can perform end-to-end joint 2D detection and tracking without post-processing. However, they seriously suffer from the conflict between detection and tracking. MOTRv2 [50] and MOTRv3 [45] concentrate on this conflict. MOTRv2 proposes" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "using an off-the-shelf detector to guide the tracker, which is effective but damages the end-to-end fashion of the tracker. MOTRv3 proposes to balance the label assignment between detection queries and track queries to achieve a better tradeoff between the detection and tracking performance of the model. MUTR3D [48] and PF-Track [31] applied the \"tracking by track queries\" paradigm to 3D MOT. MUTR3D [48] performs camera-based tracking in an end-to-end fashion by introducing 3D track queries to model spatial and appearance coherent tracks. PF-Track [31] further refine the tracks and predict the trajectories of objects with cross-object and cross-frame attention between object queries. Recently, DQTrack [21] proposes to separate the detection process and association process of trackers to avoid their conflicts in object feature representation. None of the previous joint detection and tracking approaches can truly bridge the gap between detection and tracking with a one-stage model and a unified object feature representation. Neither can they train the model from scratch and get rid of the reliance on pre-trained detectors while detecting and tracking objects with high precision and recall performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.419, + 0.461, + 0.433 + ], + "angle": 0, + "content": "2.2 Camera-based 3D MOT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.789, + 0.619 + ], + "angle": 0, + "content": "Camera-based 3D MOT has gained remarkable progress recently thanks to the developments in camera-based 3D detection and depth estimation [11, 14, 15, 22, 23, 26, 27, 36-38, 43, 51]. Early methods in camera-based 3D MOT perform tracking in 2D first and then lift the tracks to 3D space [34]. CC-3DT [10] fuses multi-view object features to enhance associating objects across different views. QD-3DT [13] perform 2D association first and enhance the instance association with the depth order and motions of 3D objects. TripletTrack [29] extracts local object feature embeddings and motion descriptors with CNN or LSTM to measure the affinity between objects. MUTR3D [48], PF-Track [31], and DQTrack [21] recently established new state-of-the-art tracking performance by applying the transformer-based tracking architecture in 3D." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.331, + 0.667 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.794, + 0.841 + ], + "angle": 0, + "content": "The model design of OneTrack is conceptually concise. We add an additional lightweight classification branch to a basic transformer-based tracker. Both ground truth assignments for the detection and tracking tasks are employed to supervise the model. We begin by reviewing the design of transformer-based trackers in Sec. 3.1. Then in Sec. 3.2, we propose to identify the conflicted positive sample assignments between detection and tracking through query grouping and prevent the contradicted gradient from back-propagating. In Sec. 3.3, we introduce the dynamic attention mask to further prevent gradient conflict between the two tasks. Additionally, the tracking classification loss is modified to suppress inaccurate boxes predicted by track queries, as described in Sec. 3.4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.704, + 0.163 + ], + "angle": 0, + "content": "3.1 Reviewing the \"Tracking by Track Queries\" Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.273 + ], + "angle": 0, + "content": "As previous transformer-based trackers [21,30,31,45,47,48,50], OneTrack detect and track 3D objects with object queries. At each timestamp \\( t \\), given multiview images \\( \\mathbf{I} \\), the 3D detection task aims to predict a set of 3D bounding boxes \\( \\mathbf{B} = \\{\\mathbf{b}^1, \\mathbf{b}^2, \\dots, \\mathbf{b}^n\\} \\) and their confidence scores \\( \\mathbf{S} = \\{s^1, s^2, \\dots, s^n\\} \\) to capture 3D objects. Meanwhile, 3D MOT requires a consistent ID for each object across frames. We represent the outputs for the 3D MOT task by \\( \\{\\mathbf{B}, \\mathbf{S}, \\mathbf{ID}\\} \\), including the predicted boxes, confidence scores, and ID of predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.336 + ], + "angle": 0, + "content": "In each frame, OneTrack first receives a set of object queries \\(\\mathbf{Q} = \\{\\mathbf{q}^1,\\mathbf{q}^2,\\dots \\}\\). Following previous works [21, 31, 47, 48], \\(\\mathbf{Q}\\) consists of track queries propagated from the previous frame and a fixed number of new detection queries \\(\\mathbf{Q}^d\\). The query propagation process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.343, + 0.785, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {t} = \\hat {\\mathbf {Q}} _ {t - 1} \\cup \\mathbf {Q} ^ {d}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.4 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{Q}}_{t - 1}\\) represents the propagated queries from the last frame. We denote \\(\\mathbf{Q}_t\\) as \\(\\mathbf{Q}\\) in the subsequent descriptions for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.787, + 0.459 + ], + "angle": 0, + "content": "In each frame, given the multi-view image features \\(\\mathbf{F}\\) extracted with image encoder, OneTrack utilizes a transformer decoder to extract object features. The decoder comprises six transformer decoder layers. In the \\(i^{\\mathrm{th}}\\) layer \\(\\mathbf{Q}\\) is updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.47, + 0.785, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\operatorname {D e c o d e r L a y e r} _ {i} (\\mathbf {F}, \\mathbf {Q}, \\mathbf {M}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{M}\\) represents the attention mask for the self-attention process in decoder layers. In end-to-end trackers, attention between all object queries is usually not masked. Then the object queries are fed into the classification and regression heads to generate prediction boxes. Ground truths are then assigned to the predictions for the calculation of classification loss and regression loss. Recent works [21,31] usually preserve historical track queries from a few past frames as tracking memory. These stored queries serve as references to the tracking process in the current frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Instead of associating boxes across frames with post-processing techniques, the \"tracking by track queries\" paradigm tracks objects with track queries propagated across frames. Therefore during the training of trackers, a ground truth instance is assigned to a specific track query in future frames once they are associated in the current frame, regardless of the accuracy of the assigned query's predictions in future frames. We denote these ground truth-query pairs as \"locked\" [45] pairs, and the remaining ground truths and queries are denoted as \"free\" ground truths or queries. In contrast, for detectors, positive samples are simply defined as the most precise prediction in each frame. Hence, a positive sample for detection may be classified as a negative sample in tracking and be suppressed, and vice versa, as illustrated in Fig.1. This originates from ground truth objects being assigned to different object queries in the two tasks. This conflict will lead to conflicted classification gradients between the two tasks, hindering the convergence of the model. We found that those conflicted gradients are primarily responsible for the precision and recall degradation of end-to-end trackers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.148, + 0.696, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 1: Difference in positive sample definition for query-based detector and tracker. With object query propagated from the last frame, the end-to-end tracker ensures that queries assigned as positive samples in the previous frame (like \\( \\mathrm{Q}_{\\mathrm{C}}^{\\prime} \\)) maintain their pairing with the same ground truth in the current frame. Meanwhile, the detector always takes the most precise predictions (like \\( \\mathrm{Q}_{\\mathrm{X}} \\)) as positive samples. In such case, the polarity of \\( \\mathrm{Q}_{\\mathrm{C}}^{\\prime} \\) and \\( \\mathrm{Q}_{\\mathrm{X}} \\) as training samples in the two tasks is opposite." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.455, + 0.741, + 0.484 + ], + "angle": 0, + "content": "3.2 Solving the Conflict Between Detection and Tracking via Gradient Coordination" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.789, + 0.604 + ], + "angle": 0, + "content": "We first try a trivial and straightforward solution to avoid the classification gradient conflict between the two tasks. We attach two classification heads to the model, one for detection classification and the other for tracking classification. Since there is no obvious conflict in box regression for the detection or tracking tasks, we use a shared head for regressing 3D boxes in the two tasks. Then all predicted boxes are assigned to ground truths twice following the positive sample definition in detection or tracking as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.615, + 0.787, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} _ {\\mathrm {d e t}} = \\arg \\min _ {\\sigma} \\sum_ {k} ^ {N} \\mathcal {L} _ {\\operatorname {m a t c h}} \\left(\\mathbf {g} ^ {k}, \\mathbf {b} ^ {\\sigma (k)}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.66, + 0.786, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} _ {\\mathrm {t r k}} ^ {\\text {f r e e}} = \\arg \\min _ {\\sigma} \\sum_ {k} ^ {M} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\mathbf {g} ^ {k}, \\mathbf {b} ^ {\\sigma (k)}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.702, + 0.787, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} _ {\\mathrm {t r k}} = \\hat {\\sigma} _ {\\mathrm {t r k}} ^ {\\mathrm {f r e e}} \\cup \\sigma_ {\\mathrm {t r k}} ^ {\\mathrm {l o c k e d}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Where \\(\\{\\mathbf{g}^k\\}_{k=1}^N\\) is \\(N\\) ground truths(GTs), in which \\(M\\) of them are free GTs for tracking. \\(\\{\\mathbf{b}^i\\}_{i=1}^n\\) is the set of predicted boxes. \\(\\sigma\\) is the possible permutation of \\(N\\) elements, following the notation in [6]. For example, \\((\\mathbf{g}^k, \\mathbf{b}^{\\sigma(k)})\\) means the \\(k^{th}\\) ground truth box is assigned to the \\(\\sigma(k)^{th}\\) predicted box. \\(\\mathcal{L}_{\\mathrm{match}}\\) denotes the matching cost between a ground truth and a prediction from an object query. \\(\\hat{\\sigma}_{\\mathrm{det}}\\) and \\(\\hat{\\sigma}_{\\mathrm{trk}}\\) are the ground truth assignments for detection and tracking, respectively. \\(\\sigma_{\\mathrm{trk}}^{\\mathrm{locked}}\\) stands for the locked ground truth and query pairs inherited" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "from the last frame. Classification losses and regression losses are then computed for gradient back-propagation in each decoder layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.372 + ], + "angle": 0, + "content": "However, we found that merely separating the classification heads for two tasks and supervising the model under both detection and tracking ground truth assignments can neither obtain a satisfying detector nor tracker, as shown in Table 4. Clearly, simply separating the classification heads cannot solve the conflict between the sample classification in detection and tracking. This is because as previously mentioned, a part of positive samples for training detectors will be assigned as negative samples for training trackers, and vice versa. Therefore even when the classification heads of two tasks are separated, classification gradients on those samples will still be conflicted when back-propagated to the decoder layers. To address this, we propose to decompose the classification gradients in both tasks to identify and avoid their partial conflict on those conflicted samples. We refer to this design as classification gradient coordination between the two tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.463 + ], + "angle": 0, + "content": "We dynamically categorize object queries into four query groups in each decode layer based on their polarity in both tasks: positive for both tasks; positive for detection and negative for tracking; negative for detection and positive for tracking; and negative for both tasks. We denote those four groups of queries as \\(\\mathbf{Q}^{\\mathrm{pp}}\\), \\(\\mathbf{Q}^{\\mathrm{pn}}\\), \\(\\mathbf{Q}^{\\mathrm{np}}\\), and \\(\\mathbf{Q}^{\\mathrm{nn}}\\). This proposed query grouping can be illustrated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.477, + 0.785, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {\\mathrm {p p}} = \\left\\{q ^ {k} | k \\in \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\in \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.497, + 0.785, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {\\mathrm {p n}} = \\left\\{q ^ {k} | k \\in \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\notin \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.519, + 0.785, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {\\mathrm {n p}} = \\left\\{q ^ {k} | k \\notin \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\in \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.54, + 0.785, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {\\mathrm {n n}} = \\left\\{q ^ {k} \\mid k \\notin \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\notin \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Through this, the classification gradients of the two tasks can also be decomposed as the sum of detection or tracking classification gradients on each group of queries. The classification gradient conflict only lies in the second and third groups of queries, as illustrated in Fig. 2(a). Therefore we propose to remove the gradient of \\(\\mathbf{Q}^{\\mathrm{pn}}\\) contributed to the tracking classification loss and the gradient of \\(\\mathbf{Q}^{\\mathrm{np}}\\) contributed to the detection classification loss, as illustrated in Fig. 2(b). Through this, the partial classification gradient conflict of the two tasks on \\(\\mathbf{Q}^{\\mathrm{pn}}\\) and \\(\\mathbf{Q}^{\\mathrm{np}}\\) is avoided. \\(\\mathbf{Q}^{\\mathrm{pp}}\\) and \\(\\mathbf{Q}^{\\mathrm{nn}}\\) still contribute gradient in both heads since the partial classification gradients upon them are not contradictory." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.473, + 0.74 + ], + "angle": 0, + "content": "3.3 Dynamic Attention Mask" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Following previous end-to-end trackers [21, 30, 31, 45, 47, 48, 50], OneTrack generates predictions in a set-prediction fashion. Hence the model is encouraged to suppress redundant predictions regarding the same ground truth object through self-attention between object queries in each decoder layer. As previously mentioned, \\(\\mathbf{Q}^{\\mathrm{pn}}\\) and \\(\\mathbf{Q}^{\\mathrm{np}}\\) originates from the same ground truth object being assigned to different object queries in the two tasks. To avoid unexpected competition" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.148, + 0.725, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Fig. 2: Backpropagation of Classification Gradients. Gradient coordination prevents the classification gradient conflict between detection and tracking. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.492 + ], + "angle": 0, + "content": "between task-specific queries regarding the same object, in each layer of the decoder, we dynamically mask the attention between \\(\\mathbf{Q}^{\\mathrm{pn}}\\) and \\(\\mathbf{Q}^{\\mathrm{np}}\\). Attention between all other query groups is not masked. Through this, \\(\\mathbf{Q}^{\\mathrm{pn}}\\) and \\(\\mathbf{Q}^{\\mathrm{np}}\\) are not directly exposed to each other during the forward process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.643 + ], + "angle": 0, + "content": "We utilize the ground truth assignment results in each decoder layer to generate the self-attention mask for the next layer. Before the first decoder layer where detection queries are initialized as empty queries, we use distances between their initial reference points and centers of ground truths to perform bipartite matching. After each decoder layer, we perform bipartite matching based on matching costs between ground truths and predicted 3D boxes defined in [37]. We employ the Hungarian algorithm [17] for bipartite matching. As for outputs, the detection results of OneTrack include predictions from \\(\\mathbf{Q}^{\\mathrm{pp}}\\), \\(\\mathbf{Q}^{\\mathrm{pn}}\\) and \\(\\mathbf{Q}^{\\mathrm{nn}}\\). The tracking results include predictions from \\(\\mathbf{Q}^{\\mathrm{pp}}\\), \\(\\mathbf{Q}^{\\mathrm{pn}}\\) and \\(\\mathbf{Q}^{\\mathrm{nn}}\\). We illustrate the pipeline of OneTrack in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.704 + ], + "angle": 0, + "content": "During inference, we categorize object queries into four groups based on the detection and tracking confidence scores of their predictions. Queries with high detection or tracking confidence scores over \\( T_{p} \\) are classified as positive samples for the respective tasks. We set \\( T_{p} = 0.2 \\) in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.81 + ], + "angle": 0, + "content": "At last, in both the training or inference process, we propagate the Top- \\(K\\) object queries with high confidence scores from both detection and tracking results to the next frame as \\(\\hat{\\mathbf{Q}}_t\\). In particular, we will still propagate the \"locked\" track queries to the future frames even if they are not in the Top- \\(K\\) object queries. This is for addressing challenging tracking scenarios, such as an object being occluded for several frames. We propagate \"locked\" track queries through this for up to \\(\\mathrm{A_{age}} = 3\\) frames into future frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Following [31,36], we also preserve the propagated object queries as the memory queue for the detection or tracking process for up to 4 frames. The mem" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.78, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 3: Training pipeline of OneTrack. For each decoder layer in OneTrack, We identify object queries into four groups based on their positivity as training samples in detection and tracking tasks. Task-specific queries that are positive samples only for one task will only backpropagate gradients in the respective task-specific classification head to prevent conflicted classification gradients. We dynamically mask all attention across positive sample queries only for detection or tracking. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.532 + ], + "angle": 0, + "content": "ory queues for the detection and tracking processes are updated independently. Those memory queries are provided to the decoder for performing hybrid attention proposed by [36]. As keys and values for the hybrid attention layer, detection and tracking memory queries are masked for \\(\\mathbf{Q}^{\\mathrm{np}}\\) and \\(\\mathbf{Q}^{\\mathrm{pn}}\\), respectively. We show the design of decoder layers in OneTrack in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.608 + ], + "angle": 0, + "content": "As demonstrated in Table 4, with the proposed dynamic hybrid-attention mask, the true joint training of detection and tracking can be achieved. This allows for training detection and tracking in a single training stage and inferring them in a single forward pass with a unified model. This feature of OneTrack significantly reduces its training cost and inference latency." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.632, + 0.57, + 0.646 + ], + "angle": 0, + "content": "3.4 Modified Tracking Classification Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In addition to introducing the true joint training of detection and tracking to reduce the detection capability degradation of trackers, we also adjusted the tracking classification loss to suppress the in fact inaccurate predictions. As previously mentioned, previous end-to-end trackers regard track queries whose predictions are associated with ground truths as positive samples equally, regardless of the precision of their current predictions. This encourages the tracker to assign high confidence to the in fact inaccurate predictions from track queries. Therefore we propose to restrain the positive assignment of those in fact inaccurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In previous end-to-end trackers, the assignment result of a ground truth object \\( g \\) can be denoted as a one-hot vector: \\( A_{g}^{n\\times 1} = [0,\\dots ,1,\\dots ,0] \\), where \\( n \\) is the number of object queries and the ground truth object is exclusively" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.144, + 0.694, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.364, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig. 4: Design of decoder layers in OneTrack. In addition to current object queries, queries in detection memory and tracking memory are fed to the hybrid attention layer as keys and values. As keys and values, \\(\\mathbf{Q}^{\\mathrm{pn}}\\) and detection memory are masked to \\(\\mathbf{Q}^{\\mathrm{np}}\\). As keys and values, \\(\\mathbf{Q}^{\\mathrm{np}}\\) and tracking memory are masked to \\(\\mathbf{Q}^{\\mathrm{pn}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.495 + ], + "angle": 0, + "content": "assigned to the \\( i^{th} \\) object query. This assignment is unaffected by the matching cost between the \\( i^{th} \\) object query and \\( g \\) if they have been associated in recent frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.543 + ], + "angle": 0, + "content": "In OneTrack, we propose to suppress the assignment weight between ground truths and queries that are locked but have high matching costs. We adjust the assignment result of locked ground truth \\( g \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.552, + 0.786, + 0.572 + ], + "angle": 0, + "content": "\\[\nA _ {g} ^ {n \\times 1} = [ 0, \\dots , W _ {g i}, \\dots , 0 ], \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.574, + 0.786, + 0.605 + ], + "angle": 0, + "content": "\\[\nW _ {g i} = - \\frac {1}{\\gamma} * \\operatorname {M a x} \\left(\\left(\\operatorname {d i s t} (g, q ^ {i}) - 0. 5\\right), 0\\right) + 1, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.738 + ], + "angle": 0, + "content": "where \\(\\mathrm{dist}(g, q^i)\\) represents the distance between the centers of ground truth box \\(g\\) and the predicted box of the locked query \\(q^i\\). \\(\\gamma\\) is a hyper-parameter greater than zero. The assigned weight \\(W_{gi}\\) between \\(g\\) and \\(q^i\\) decreases linearly from 1.0 as \\(\\mathrm{dist}(g, q^i)\\) increases from \\(0.5\\mathrm{m}\\). The assignment weight serves as the weighting factor for calculating box regression loss and the targets for predicting tracking confidence score. Instead of Focal Loss [24], we employ the Quality Focal Loss proposed by [20] to supervise the tracking classification head, given the continuous \\(0\\sim 1\\) confidence label." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.377, + 0.78 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In this section, we first introduce our detailed experimental setup. Then we provide comparisons with recent works on nuScenes camera-based 3D tracking benchmark. Ablation studies on each component are presented in the last part." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Compare with recent works on nuScenes dataset. \\(\\dagger\\): trained from scratch. OneTrack-F / OneTrack-S: OneTrack trained under full-resolution / half-resolution as described in Sec. 4.2." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.2, + 0.755, + 0.45 + ], + "angle": 0, + "content": "
Backbonee2eResolutionAMOTA↑AMOTP↓MOTA↑RECALL↑IDS↓
Validation Split
QD3DT [13]R101✓1600×9000.2421.5180.21839.9%5646
MUTR3D [48]R101✓1600×9000.2941.4980.26742.7%3822
CC-3DT [10]R101×1600×6400.4291.2570.38553.4%2219
PF-Track-S [31]V2-99✓800×3200.4081.3430.37650.7%166
PF-Track-F [31]V2-99✓1600×6400.4791.2270.43559.0%181
DQTrack [21]V2-99✓800×3200.4461.251--1193
OneTrack-S†V2-99✓800×3200.4921.1220.40958.8%315
OneTrack-F†V2-99✓1600×6400.5481.0880.47961.8%389
Test Split
QD3DT [13]R101✓1600×6400.2171.5500.19837.5%6856
MUTR3D [48]R101✓1600×6400.2701.4940.23541.1%6018
SRCN3D [32]V2-99×1600×6400.3981.3170.35953.8%3334
CC-3DT [10]R101×1600×6400.4101.2740.35753.8%3334
PF-Track [31]V2-99✓1600×6400.4341.2520.37853.8%249
DQTrack [21]V2-99✓1600×6400.5231.0960.44462.2%1204
OneTrack-F†V2-99✓1600×6400.5541.0210.46160.8%481
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.478, + 0.44, + 0.492 + ], + "angle": 0, + "content": "4.1 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.625 + ], + "angle": 0, + "content": "We evaluate our proposed method on the nuScenes dataset. nuScenes [5] dataset is a large-scale autonomous driving benchmark containing 1000 multi-modal videos. Videos recorded by six cameras are divided into 700, 150, and 100 scenes for training, validation, and testing, respectively. 3D box annotations of 10 object classes are provided for keyframes at \\(2\\mathrm{Hz}\\). For nuScenes 3D tracking benchmark, we report AMOTA, AMOTP [39], MOTA [3], Recall and identity switches (IDS). For the 3D detection task, we report mean Average Precision (mAP) and nuScenes Detection Score (NDS)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.458, + 0.664 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Our implementation is mainly based on StreamPETR [36]. We define two implementation settings: The full-resolution setting for state-of-the-art comparison and the half-resolution setting for ablation studies. We conduct experiments using V2-99 [18] as the image backbone. OneTrack is trained using AdamW [28] optimizer with a batch size of 8 and a base learning rate of 4e-4. The cosine annealing policy is employed to adjust the learning rate. We use 644 initialized detection queries in each frame and propagate the Top-256 object queries into the next frame. When training, we record all matched ground truth and object query pairs as \"locked\" pairs for label assignment. An object query matched with a ground truth will be recorded as a track query. As previously mentioned, a \"locked\" pair will be preserved in memory for up to \\( A_{\\mathrm{age}} = 3 \\) frames, even if" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.302, + 0.145, + 0.699, + 0.16 + ], + "angle": 0, + "content": "Table 4: Ablation study on components breakdown." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.172, + 0.785, + 0.294 + ], + "angle": 0, + "content": "
Training taskInitialized from pre-trained detectorGradient coordinationDynamic High-cost assign masksuppressAMOTA↑IDS↓NDS↑mAP↑
detection------0.5700.479
tracking×--×0.3051210.4550.313
tracking✓--×0.3754660.5030.409
joint××××0.4252460.5170.410
joint×✓××0.4412310.5550.462
joint×✓✓×0.4692610.5750.484
joint×✓✓✓0.4923150.5660.478
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.787, + 0.474 + ], + "angle": 0, + "content": "the respective track query is not in the Top- \\( K \\) object queries in each frame. For track queries that failed to predict confident prediction over \\( \\mathrm{A}_{\\mathrm{age}} \\) frames, its \"locked\" ground truth will be assigned to other \"free\" object queries. During inference, we initialize an object query as a track query if it predicts a 3D box with confidence over \\( T_{\\mathrm{conf}} = 0.4 \\). Additionally, all track queries are preserved and propagated across frames for up to \\( \\mathrm{A}_{\\mathrm{age}} = 3 \\) frames. Predictions from the same track query propagated across frames will be assigned the same tracking ID. We output all detection predictions with confidence scores over 0.05 and all tracking predictions with confidence scores over 0.3. We introduce other setting-specific training configurations as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.787, + 0.536 + ], + "angle": 0, + "content": "OneTrack-F: We crop input images from the initial resolution of \\(1600 \\times 900\\) to \\(1600 \\times 640\\). We then conduct end-to-end training for 48 epochs. OneTrack-S: We first crop the input images to the resolution of \\(1600 \\times 640\\) and downsample them to \\(800 \\times 320\\). Models are trained for 24 epochs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.553, + 0.573 + ], + "angle": 0, + "content": "4.3 Comparisons with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.689 + ], + "angle": 0, + "content": "As shown in Table 3, OneTrack outperforms all previous trackers in both the validation and test set of nuScenes dataset by a significant margin. On the validation set, OneTrack surpasses previous state-of-the-art PF-Track [31] by \\(10.2\\%\\) AMOTA. On the test set, OneTrack surpasses DQTrack [21] by \\(3.3\\%\\) AMOTA. Notably, OneTrack reports fewer than half of the identity switches compared to those reported by DQTrack. It is also worth mentioning that OneTrack is the only model in Table 3 that is trained from scratch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.343, + 0.725 + ], + "angle": 0, + "content": "4.4 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.763 + ], + "angle": 0, + "content": "We conduct ablation studies on each proposed component with OneTrack-S on the nuScenes validation set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Components breakdown. The breakdown analysis of proposed components is presented in Table 4. Here \"det\" or \"trk\" in the first row stands for only training the model for detection or tracking tasks. And \"joint\" stands for jointly training the model for both detection and tracking tasks, with an extra classification head attached to the model. As shown in Table 4, the model trained solely" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.146, + 0.413, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.146, + 0.605, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.146, + 0.801, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.308 + ], + "angle": 0, + "content": "Fig. 5: Ablation on hyper-parameters. (a) On parameter \\(\\gamma\\) in the modified tracking losses. (b) On the max preserved frames of track queries \\(A_{\\mathrm{age}}\\). (c) On the score threshold \\(T_{\\mathrm{conf}}\\) for track query initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.519 + ], + "angle": 0, + "content": "for the tracking task suffers from severe degradation in detection performance, which results in their inferior detection performance compared to the detector. Initializing the tracker from a detector checkpoint or trivially supervising the model with both detection and tracking losses can only slightly alleviate this problem. Our proposed components including classification gradient coordination, dynamic attention mask, and high-cost tracking assignment suppress, are all validated as highly effective in achieving the true joint training of detection and tracking tasks. It is noteworthy that OneTrack without modifying its tracking loss achieves an even slightly better detection performance than the model trained solely as a detector. This observation suggests the potential for further enhancing the detection process through the integrated tracking process in a unified model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Modified tracking loss. In Fig. 5(a), we conduct ablation studies on the hyper-parameter \\(\\gamma\\) in the modified tracking loss for suppressing positive assignments of inaccurate predictions. OneTrack achieves optimal tracking performance when \\(\\gamma\\) is set to 7.5. While the detection performance of OneTrack shows improvement as \\(\\gamma\\) increases, this gain becomes limited when \\(\\gamma\\) exceeds 7.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Track age. We conduct ablation studies on the maximum preservation age \\( A_{\\mathrm{age}} \\) of track queries. As illustrated in Fig. 5(b), a longer track age will lead to more identity switches and decreased detection performance. We speculate that this is because propagating too many outdated track queries will mislead the object feature extraction process in the current frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Confidence threshold for track query initialization. We assess the impact of \\( T_{\\mathrm{conf}} \\) in Fig. 5(c). The AMOTA performance will slightly increase as \\( T_{\\mathrm{conf}} \\) increases, which results from the more accurate tracking predictions on average. However, a higher \\( T_{\\mathrm{conf}} \\) reduces the initialization of track queries from detection queries with low confidence. This limitation hinders tracking objects under poor observation conditions, resulting in a higher number of identity switch cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Separated classification heads. As shown in Table 5, when its two classification heads share weights except for the final linear layer, the tracking and detection performance of OneTrack is slightly worse than when the classification heads are independent. We speculate that this is because the separation of classification heads leads to the easier convergence of both heads." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.145, + 0.764, + 0.159 + ], + "angle": 0, + "content": "Table 5: Ablation on the separated classification heads for two tasks." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.171, + 0.684, + 0.23 + ], + "angle": 0, + "content": "
Cls headsAMOTA↑IDS↓NDS↑mAP↑
shared49.03760.5590.473
separated49.23150.5660.478
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Table 6: Computational cost comparison with recent works. All training time costs and inference latencies are measured on a single A100 GPU. *: PF-Track is first trained as a detector and then as a tracker for 12 epochs each. We train DQTrack and OneTrack from scratch for 24 epochs." + }, + { + "type": "table", + "bbox": [ + 0.309, + 0.31, + 0.691, + 0.398 + ], + "angle": 0, + "content": "
MethodBackboneResolutionTraining time costInference latency
PF-TrackV2-99800×32034h+82h*124.4ms
DQTrackV2-99800×320176h115.2ms
OneTrackV2-99800×32087h83.7ms
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Computational cost. To ensure fair comparison, we apply gradient checkpointing [7] on the decoder layers of all trackers. We measure their training cost and inference latency on a single A100 GPU. As shown in Table 6, the training cost of OneTrack is only half that of previous state-of-the-art trackers. The inference latency of OneTrack is also significantly lower. We attribute the low computational cost of OneTrack to its elegant model design. We do not extend the tracker with various plug-in modules [31] or separate the detection and tracking stages of the tracker [21], which will introduce additional computational burdens. Notably, OneTrack is trained from scratch while PF-Track and DQTrack rely on extra training stages for achieving high tracking performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.36, + 0.611 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.73 + ], + "angle": 0, + "content": "In this work, we delve into the optimization conflict between detection and tracking in end-to-end trackers, which arises from the partially conflicted classification gradients between two tasks. Based on this observation, we propose to identify the conflicted positive samples between two tasks through query grouping and solve the conflicted gradients with the proposed gradient coordination, dynamic attention mask, and suppression tracking positive samples of low quality. On the nuScenes dataset, OneTrack outperforms all previous end-to-end trackers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.395, + 0.768 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Ali, A., Jalil, A., Niu, J., Zhao, X., Rathore, S., Ahmed, J., Aksam Iftikhar, M.: Visual object tracking—classical and contemporary approaches. Frontiers of Computer Science 10, 167–188 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.263 + ], + "angle": 0, + "content": "2. Benbarka, N., Schröder, J., Zell, A.: Score refinement for confidence-based 3d multi-object tracking. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8083-8090. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.305 + ], + "angle": 0, + "content": "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing 2008, 1-10 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.785, + 0.348 + ], + "angle": 0, + "content": "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "7. Chen, T., Xu, B., Zhang, C., Guestrin, C.: Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.488 + ], + "angle": 0, + "content": "8. Chiu, H.k., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: ICRA (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.531 + ], + "angle": 0, + "content": "9. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4870-4880 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "0. Fischer, T., Yang, Y., Kumar, S., Sun, M., Yu, F.: CC-3DT: panoramic 3d object tracking via cross-camera fusion. In: CoRL. Proceedings of Machine Learning Research, vol. 205, pp. 2294–2305. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.602 + ], + "angle": 0, + "content": "1. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science 18(5), 185706 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.785, + 0.657 + ], + "angle": 0, + "content": "2. He, J., Huang, Z., Wang, N., Zhang, Z.: Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5299-5309 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.699 + ], + "angle": 0, + "content": "3. Hu, H.N., Yang, Y.H., Fischer, T., Darrell, T., Yu, F., Sun, M.: Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.728 + ], + "angle": 0, + "content": "4. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Kim, A., Osep, A., Leal-Taixe, L.: Eagermot: 3d multi-object tracking via sensor fusion. In: 2021 IEEE International conference on Robotics and Automation (ICRA). pp. 11315-11321 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Q. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.204 + ], + "angle": 0, + "content": "18. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "19. Li, J., Ding, Y., Wei, H.L., Zhang, Y., Lin, W.: Simpletrack: Rethinking and improving the jde approach for multi-object tracking. Sensors 22(15), 5863 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.285 + ], + "angle": 0, + "content": "20. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.325 + ], + "angle": 0, + "content": "21. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18302-18311 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.367 + ], + "angle": 0, + "content": "22. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.787, + 0.42 + ], + "angle": 0, + "content": "23. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.462 + ], + "angle": 0, + "content": "24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.787, + 0.488 + ], + "angle": 0, + "content": "25. Liu, C., Chen, X.F., Bo, C.J., Wang, D.: Long-term visual tracking: review and experimental comparison. Machine Intelligence Research 19(6), 512-530 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "26. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.57 + ], + "angle": 0, + "content": "27. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.57, + 0.787, + 0.596 + ], + "angle": 0, + "content": "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.787, + 0.637 + ], + "angle": 0, + "content": "29. Marinello, N., Proesmans, M., Van Gool, L.: Tripletrack: 3d object tracking using triplet embeddings and lstm. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4499-4509 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "30. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8844-8854 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.733 + ], + "angle": 0, + "content": "31. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17928-17938 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.787, + 0.773 + ], + "angle": 0, + "content": "32. Shi, Y., Shen, J., Sun, Y., Wang, Y., Li, J., Sun, S., Jiang, K., Yang, D.: Srcn3d: Sparse r-cnn 3d surround-view camera object detection and tracking for autonomous driving. arXiv preprint arXiv:2206.14451 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.787, + 0.813 + ], + "angle": 0, + "content": "33. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10860-10869 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "34. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: ICCV. pp. 10840-10849. IEEE (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.664, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OneTrack" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "35. Wang, Q., Chen, Y., Pang, Z., Wang, N., Zhang, Z.: Immortal tracker: Tracklet never dies. arXiv preprint arXiv:2111.13672 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "36. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3621-3631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "38. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5096-5105 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "39. Weng, X., Kitani, K.: A baseline for 3d multi-object tracking. arXiv preprint arXiv:1907.03961 1(2), 6 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "40. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. IROS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.785, + 0.396 + ], + "angle": 0, + "content": "41. Weng, X., Wang, Y., Man, Y., Kitani, K.: Gnn3dmot: Graph neural network for 3d multi-object tracking with multi-feature learning. arXiv preprint arXiv:2006.07327 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "42. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: 2017 IEEE international conference on image processing (ICIP). pp. 3645-3649. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.44, + 0.785, + 0.494 + ], + "angle": 0, + "content": "43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "44. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "45. Yu, E., Wang, T., Li, Z., Zhang, Y., Zhang, X., Tao, W.: Motrv3: Releasefetch supervision for end-to-end multi-object tracking. arXiv preprint arXiv:2305.14298 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "46. Zaech, J.N., Liniger, A., Dai, D., Danelljan, M., Van Gool, L.: Learnable online graph representations for 3d multi-object tracking. IEEE Robotics and Automation Letters 7(2), 5103-5110 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "47. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "48. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4537-4546 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "49. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision 129, 3069-3087 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "50. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22056-22065 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "51. Zhao, H., Zhang, J., Chen, Z., Yuan, B., Tao, D.: On robust cross-view consistency in self-supervised monocular depth estimation. Machine Intelligence Research 21(3), 495-513 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_origin.pdf b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f1d5fc87c9960e898d0554ed92b6b12c7e7060c7 --- /dev/null +++ b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/72eb6377-51dc-4c32-b71b-98f8eb5e37dc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ea419004e513c1b140723fc31e701dac96bb7e60311058a790b1490ebe13522 +size 681113 diff --git a/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/full.md b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bd78c4e9a599c1a49e5fb9dbf679eccb85501865 --- /dev/null +++ b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/full.md @@ -0,0 +1,298 @@ +# OneTrack: Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers + +Qitai Wang $^{1,2}$ , Jiawei He $^{2}$ , Yuntao Chen $^{3}$ , and Zhaoxiang Zhang $^{1,2,3,4}$ + +1 School of Future Technology, University of Chinese Academy of Sciences (UCAS) + +$^{2}$ NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA) {wangqitai2020,zhaoxiang.zhang}@ia.ac.cn,jwhe2024@gmail.com + +3 Center for Artificial Intelligence and Robotics, HKISI, CAS chenyuntao08@gmail.com + +4 Shanghai Artificial Intelligence Laboratory + +Abstract. Existing end-to-end trackers for vision-based 3D perception suffer from performance degradation due to the conflict between detection and tracking tasks. In this work, we get to the bottom of this conflict, which was vaguely attributed to incompatible task-specific object features previously. We find the conflict between the two tasks lies in their partially conflicted classification gradients, which stems from their subtle difference in positive sample assignments. Based on this observation, we propose to coordinate those conflicted gradients from object queries with contradicted polarity in the two tasks. We also dynamically split all object queries into four groups based on their polarity in the two tasks. Attention between query sets with conflicted positive sample assignments is masked. The tracking classification loss is modified to suppress inaccurate predictions. To this end, we propose OneTrack, the first one-stage joint detection and tracking model that bridges the gap between detection and tracking under a unified object feature representation. On the nuScenes camera-based object tracking benchmark, OneTrack outperforms previous works by $6.9\%$ AMOTA on the validation set and by $3.1\%$ AMOTA on the test set. + +Keywords: 3D Tracking $\cdot$ Camera-based Detection and Tracking $\cdot$ End-to-end Tracking $\cdot$ Gradient Coordination + +# 1 Introduction + +Reasoning about the location and trajectory of surrounding objects is a fundamental task for autonomous driving and robotic navigation systems. Due to the robustness and low cost of cameras, vision-based 3D detection and tracking have received widespread attention from the research community [14,22,23,26,36,37]. Recently, several approaches [21, 31, 48] have been proposed to improve 3D multi-object tracking (MOT) by jointly optimizing the detection and tracking pipeline. Based on transformer architecture, those end-to-end 3D trackers use + +Table 1: Comparison with previous works trying to solve the conflict between detection and tracking. *: DQTrack relies on frozen pre-trained detectors when compared with state-of-the-art trackers. + +
Methodsettinge2eone-stagedetection supervisiontrain from scratch
MOTRv2 [50]2D××××
MOTRv3 [45]2D✓✓×✓
DQTrack [21]3D✓×✓✓*
OneTrack3D✓✓✓✓
+ +track queries to track objects, being free from the dependence on offline detectors and post-processing operations. + +However, most end-to-end 3D trackers suffer from the optimization conflict between detection and tracking, a problem also widely recognized in 2D MOT studies [45, 50]. This conflict results in the inferior precision and recall of end-to-end trackers compared to their baseline detectors. Previous works roughly attribute the conflict between two tasks to differences in required features for detection and tracking and attempted to alleviate this problem. Some of them train their model as detectors before the tracking training process [31, 33] or take an off-the-shelf detector as guidance [50]. Some separate the detection and association stages as well as their object feature representation to avoid conflict, such as DQTrack [21]. Although much progress has been made, training a one-stage end-to-end tracker from scratch remains challenging. It is also challenging to train trackers jointly under both detection and tracking supervision. We summarize recent works in Table 1. In this work, we aim to accomplish the true joint training of detection and tracking from scratch with a one-stage model and unified feature representation for both tasks. + +We conduct a pilot study to investigate how and where the two tasks conflict during model optimization. As illustrated in Table 2, only jointly supervising the detector with an extra classification loss computed under the ground truth assignments for tracking will lead to severe performance degradation of detectors. Therefore, we conclude that the conflict between the two tasks originates from their conflicting classification gradients. + +In this paper, we precisely identify and coordinate the conflicted classification gradient between the two tasks, which originates from their partially conflicted positive sample definition. We propose OneTrack, the first model capable of achieving the true joint training of detection and tracking in a unified, one-stage model. OneTrack follows the "tracking by track queries" paradigm. During the training process, we assign ground truth annotations to object queries twice based on positive sample definitions for detection or tracking. We dynamically categorize all object queries into four groups based on their polarity in two tasks. We have two lightweight classification heads in OneTrack which are respectively responsible for classifying detection or tracking positive samples. Queries that are + +Table 2: Pilot study on supervising 3D detector under additional ground truth assignments for tracking. The stronger color means more degradation in performance. Experiments indicate the conflict between detection and tracking tasks lies in their classification supervision. + +
Detection ClassificationDetection RegressionTracking ClassificationTracking RegressionAMOTA↑IDS↓NDS↑mAP↑
✓✓××--0.5700.479
××✓✓0.305121--
✓✓✓×0.4043630.5120.432
✓✓×✓0.41785120.5660.473
✓✓✓✓0.4052550.5280.428
+ +positive samples only for one task will only backpropagate classification gradients in the respective head to reduce conflicted gradients. We dynamically mask all attention between conflicted positive queries in each decoder layer to prevent unexpected query competition. We also adjust the tracking loss to suppress inaccurate predictions. + +To summarize, our contributions are as follows: + +- We reveal that the conflict between detection and tracking tasks lies in their diverged definition of positive samples and their partially conflicted classification gradients. +- We propose OneTrack, the first model capable of completely addressing the conflict between detection and tracking under a unified object feature representation. OneTrack can be trained as both a detector and a tracker from scratch in a single training stage and can perform tracking and detection in a one-stage fashion. The detection performance of OneTrack is comparable to detectors, while its tracking performance surpasses all previous 3D trackers. +- Our method establishes a new state-of-the-art on nuScenes MOT benchmark, surpassing previous methods by over $3.1\%$ AMOTA. + +# 2 Related Works + +# 2.1 Query Propagation in MOT + +Early works in 2D or 3D MOT most follow the "tracking by detection" paradigm [4] and associate detection results across frames through post-processing [1, 2, 8, 9, 12, 13, 16, 19, 25, 35, 40-42, 44, 46, 49]. Based on DETR [6] which introduced object queries to detect objects, MOTR [47] and Trackformer [30] extended the concept of object query to track query by propagating object queries across frames. By tracking objects through track queries, MOTR and Trackformer can perform end-to-end joint 2D detection and tracking without post-processing. However, they seriously suffer from the conflict between detection and tracking. MOTRv2 [50] and MOTRv3 [45] concentrate on this conflict. MOTRv2 proposes + +using an off-the-shelf detector to guide the tracker, which is effective but damages the end-to-end fashion of the tracker. MOTRv3 proposes to balance the label assignment between detection queries and track queries to achieve a better tradeoff between the detection and tracking performance of the model. MUTR3D [48] and PF-Track [31] applied the "tracking by track queries" paradigm to 3D MOT. MUTR3D [48] performs camera-based tracking in an end-to-end fashion by introducing 3D track queries to model spatial and appearance coherent tracks. PF-Track [31] further refine the tracks and predict the trajectories of objects with cross-object and cross-frame attention between object queries. Recently, DQTrack [21] proposes to separate the detection process and association process of trackers to avoid their conflicts in object feature representation. None of the previous joint detection and tracking approaches can truly bridge the gap between detection and tracking with a one-stage model and a unified object feature representation. Neither can they train the model from scratch and get rid of the reliance on pre-trained detectors while detecting and tracking objects with high precision and recall performance. + +# 2.2 Camera-based 3D MOT + +Camera-based 3D MOT has gained remarkable progress recently thanks to the developments in camera-based 3D detection and depth estimation [11, 14, 15, 22, 23, 26, 27, 36-38, 43, 51]. Early methods in camera-based 3D MOT perform tracking in 2D first and then lift the tracks to 3D space [34]. CC-3DT [10] fuses multi-view object features to enhance associating objects across different views. QD-3DT [13] perform 2D association first and enhance the instance association with the depth order and motions of 3D objects. TripletTrack [29] extracts local object feature embeddings and motion descriptors with CNN or LSTM to measure the affinity between objects. MUTR3D [48], PF-Track [31], and DQTrack [21] recently established new state-of-the-art tracking performance by applying the transformer-based tracking architecture in 3D. + +# 3 Method + +The model design of OneTrack is conceptually concise. We add an additional lightweight classification branch to a basic transformer-based tracker. Both ground truth assignments for the detection and tracking tasks are employed to supervise the model. We begin by reviewing the design of transformer-based trackers in Sec. 3.1. Then in Sec. 3.2, we propose to identify the conflicted positive sample assignments between detection and tracking through query grouping and prevent the contradicted gradient from back-propagating. In Sec. 3.3, we introduce the dynamic attention mask to further prevent gradient conflict between the two tasks. Additionally, the tracking classification loss is modified to suppress inaccurate boxes predicted by track queries, as described in Sec. 3.4. + +# 3.1 Reviewing the "Tracking by Track Queries" Paradigm + +As previous transformer-based trackers [21,30,31,45,47,48,50], OneTrack detect and track 3D objects with object queries. At each timestamp $t$ , given multiview images $\mathbf{I}$ , the 3D detection task aims to predict a set of 3D bounding boxes $\mathbf{B} = \{\mathbf{b}^1, \mathbf{b}^2, \dots, \mathbf{b}^n\}$ and their confidence scores $\mathbf{S} = \{s^1, s^2, \dots, s^n\}$ to capture 3D objects. Meanwhile, 3D MOT requires a consistent ID for each object across frames. We represent the outputs for the 3D MOT task by $\{\mathbf{B}, \mathbf{S}, \mathbf{ID}\}$ , including the predicted boxes, confidence scores, and ID of predictions. + +In each frame, OneTrack first receives a set of object queries $\mathbf{Q} = \{\mathbf{q}^1,\mathbf{q}^2,\dots \}$ . Following previous works [21, 31, 47, 48], $\mathbf{Q}$ consists of track queries propagated from the previous frame and a fixed number of new detection queries $\mathbf{Q}^d$ . The query propagation process can be formulated as follows: + +$$ +\mathbf {Q} _ {t} = \hat {\mathbf {Q}} _ {t - 1} \cup \mathbf {Q} ^ {d}, \tag {1} +$$ + +where $\hat{\mathbf{Q}}_{t - 1}$ represents the propagated queries from the last frame. We denote $\mathbf{Q}_t$ as $\mathbf{Q}$ in the subsequent descriptions for simplicity. + +In each frame, given the multi-view image features $\mathbf{F}$ extracted with image encoder, OneTrack utilizes a transformer decoder to extract object features. The decoder comprises six transformer decoder layers. In the $i^{\mathrm{th}}$ layer $\mathbf{Q}$ is updated as follows: + +$$ +\mathbf {Q} = \operatorname {D e c o d e r L a y e r} _ {i} (\mathbf {F}, \mathbf {Q}, \mathbf {M}). \tag {2} +$$ + +Here, $\mathbf{M}$ represents the attention mask for the self-attention process in decoder layers. In end-to-end trackers, attention between all object queries is usually not masked. Then the object queries are fed into the classification and regression heads to generate prediction boxes. Ground truths are then assigned to the predictions for the calculation of classification loss and regression loss. Recent works [21,31] usually preserve historical track queries from a few past frames as tracking memory. These stored queries serve as references to the tracking process in the current frame. + +Instead of associating boxes across frames with post-processing techniques, the "tracking by track queries" paradigm tracks objects with track queries propagated across frames. Therefore during the training of trackers, a ground truth instance is assigned to a specific track query in future frames once they are associated in the current frame, regardless of the accuracy of the assigned query's predictions in future frames. We denote these ground truth-query pairs as "locked" [45] pairs, and the remaining ground truths and queries are denoted as "free" ground truths or queries. In contrast, for detectors, positive samples are simply defined as the most precise prediction in each frame. Hence, a positive sample for detection may be classified as a negative sample in tracking and be suppressed, and vice versa, as illustrated in Fig.1. This originates from ground truth objects being assigned to different object queries in the two tasks. This conflict will lead to conflicted classification gradients between the two tasks, hindering the convergence of the model. We found that those conflicted gradients are primarily responsible for the precision and recall degradation of end-to-end trackers. + +![](images/1f12da4d09ecadf1a37c3a42eb4d7c09506729af94a8352e6fb93e14db1faad9.jpg) +Fig. 1: Difference in positive sample definition for query-based detector and tracker. With object query propagated from the last frame, the end-to-end tracker ensures that queries assigned as positive samples in the previous frame (like $\mathrm{Q}_{\mathrm{C}}^{\prime}$ ) maintain their pairing with the same ground truth in the current frame. Meanwhile, the detector always takes the most precise predictions (like $\mathrm{Q}_{\mathrm{X}}$ ) as positive samples. In such case, the polarity of $\mathrm{Q}_{\mathrm{C}}^{\prime}$ and $\mathrm{Q}_{\mathrm{X}}$ as training samples in the two tasks is opposite. + +# 3.2 Solving the Conflict Between Detection and Tracking via Gradient Coordination + +We first try a trivial and straightforward solution to avoid the classification gradient conflict between the two tasks. We attach two classification heads to the model, one for detection classification and the other for tracking classification. Since there is no obvious conflict in box regression for the detection or tracking tasks, we use a shared head for regressing 3D boxes in the two tasks. Then all predicted boxes are assigned to ground truths twice following the positive sample definition in detection or tracking as follows: + +$$ +\hat {\sigma} _ {\mathrm {d e t}} = \arg \min _ {\sigma} \sum_ {k} ^ {N} \mathcal {L} _ {\operatorname {m a t c h}} \left(\mathbf {g} ^ {k}, \mathbf {b} ^ {\sigma (k)}\right), \tag {3} +$$ + +$$ +\hat {\sigma} _ {\mathrm {t r k}} ^ {\text {f r e e}} = \arg \min _ {\sigma} \sum_ {k} ^ {M} \mathcal {L} _ {\text {m a t c h}} \left(\mathbf {g} ^ {k}, \mathbf {b} ^ {\sigma (k)}\right), \tag {4} +$$ + +$$ +\hat {\sigma} _ {\mathrm {t r k}} = \hat {\sigma} _ {\mathrm {t r k}} ^ {\mathrm {f r e e}} \cup \sigma_ {\mathrm {t r k}} ^ {\mathrm {l o c k e d}}, \tag {5} +$$ + +Where $\{\mathbf{g}^k\}_{k=1}^N$ is $N$ ground truths(GTs), in which $M$ of them are free GTs for tracking. $\{\mathbf{b}^i\}_{i=1}^n$ is the set of predicted boxes. $\sigma$ is the possible permutation of $N$ elements, following the notation in [6]. For example, $(\mathbf{g}^k, \mathbf{b}^{\sigma(k)})$ means the $k^{th}$ ground truth box is assigned to the $\sigma(k)^{th}$ predicted box. $\mathcal{L}_{\mathrm{match}}$ denotes the matching cost between a ground truth and a prediction from an object query. $\hat{\sigma}_{\mathrm{det}}$ and $\hat{\sigma}_{\mathrm{trk}}$ are the ground truth assignments for detection and tracking, respectively. $\sigma_{\mathrm{trk}}^{\mathrm{locked}}$ stands for the locked ground truth and query pairs inherited + +from the last frame. Classification losses and regression losses are then computed for gradient back-propagation in each decoder layer. + +However, we found that merely separating the classification heads for two tasks and supervising the model under both detection and tracking ground truth assignments can neither obtain a satisfying detector nor tracker, as shown in Table 4. Clearly, simply separating the classification heads cannot solve the conflict between the sample classification in detection and tracking. This is because as previously mentioned, a part of positive samples for training detectors will be assigned as negative samples for training trackers, and vice versa. Therefore even when the classification heads of two tasks are separated, classification gradients on those samples will still be conflicted when back-propagated to the decoder layers. To address this, we propose to decompose the classification gradients in both tasks to identify and avoid their partial conflict on those conflicted samples. We refer to this design as classification gradient coordination between the two tasks. + +We dynamically categorize object queries into four query groups in each decode layer based on their polarity in both tasks: positive for both tasks; positive for detection and negative for tracking; negative for detection and positive for tracking; and negative for both tasks. We denote those four groups of queries as $\mathbf{Q}^{\mathrm{pp}}$ , $\mathbf{Q}^{\mathrm{pn}}$ , $\mathbf{Q}^{\mathrm{np}}$ , and $\mathbf{Q}^{\mathrm{nn}}$ . This proposed query grouping can be illustrated as follows: + +$$ +\mathbf {Q} ^ {\mathrm {p p}} = \left\{q ^ {k} | k \in \hat {\sigma} _ {\mathrm {d e t}}, k \in \hat {\sigma} _ {\mathrm {t r k}} \right\}, \tag {6} +$$ + +$$ +\mathbf {Q} ^ {\mathrm {p n}} = \left\{q ^ {k} | k \in \hat {\sigma} _ {\mathrm {d e t}}, k \notin \hat {\sigma} _ {\mathrm {t r k}} \right\}, \tag {7} +$$ + +$$ +\mathbf {Q} ^ {\mathrm {n p}} = \left\{q ^ {k} | k \notin \hat {\sigma} _ {\mathrm {d e t}}, k \in \hat {\sigma} _ {\mathrm {t r k}} \right\}, \tag {8} +$$ + +$$ +\mathbf {Q} ^ {\mathrm {n n}} = \left\{q ^ {k} \mid k \notin \hat {\sigma} _ {\mathrm {d e t}}, k \notin \hat {\sigma} _ {\mathrm {t r k}} \right\}. \tag {9} +$$ + +Through this, the classification gradients of the two tasks can also be decomposed as the sum of detection or tracking classification gradients on each group of queries. The classification gradient conflict only lies in the second and third groups of queries, as illustrated in Fig. 2(a). Therefore we propose to remove the gradient of $\mathbf{Q}^{\mathrm{pn}}$ contributed to the tracking classification loss and the gradient of $\mathbf{Q}^{\mathrm{np}}$ contributed to the detection classification loss, as illustrated in Fig. 2(b). Through this, the partial classification gradient conflict of the two tasks on $\mathbf{Q}^{\mathrm{pn}}$ and $\mathbf{Q}^{\mathrm{np}}$ is avoided. $\mathbf{Q}^{\mathrm{pp}}$ and $\mathbf{Q}^{\mathrm{nn}}$ still contribute gradient in both heads since the partial classification gradients upon them are not contradictory. + +# 3.3 Dynamic Attention Mask + +Following previous end-to-end trackers [21, 30, 31, 45, 47, 48, 50], OneTrack generates predictions in a set-prediction fashion. Hence the model is encouraged to suppress redundant predictions regarding the same ground truth object through self-attention between object queries in each decoder layer. As previously mentioned, $\mathbf{Q}^{\mathrm{pn}}$ and $\mathbf{Q}^{\mathrm{np}}$ originates from the same ground truth object being assigned to different object queries in the two tasks. To avoid unexpected competition + +![](images/8bdaa410e6798843dd3da7acc3b3f2ffc02b4024ba17812da56df5e856fbcae1.jpg) +Fig. 2: Backpropagation of Classification Gradients. Gradient coordination prevents the classification gradient conflict between detection and tracking. Best viewed in color. + +between task-specific queries regarding the same object, in each layer of the decoder, we dynamically mask the attention between $\mathbf{Q}^{\mathrm{pn}}$ and $\mathbf{Q}^{\mathrm{np}}$ . Attention between all other query groups is not masked. Through this, $\mathbf{Q}^{\mathrm{pn}}$ and $\mathbf{Q}^{\mathrm{np}}$ are not directly exposed to each other during the forward process. + +We utilize the ground truth assignment results in each decoder layer to generate the self-attention mask for the next layer. Before the first decoder layer where detection queries are initialized as empty queries, we use distances between their initial reference points and centers of ground truths to perform bipartite matching. After each decoder layer, we perform bipartite matching based on matching costs between ground truths and predicted 3D boxes defined in [37]. We employ the Hungarian algorithm [17] for bipartite matching. As for outputs, the detection results of OneTrack include predictions from $\mathbf{Q}^{\mathrm{pp}}$ , $\mathbf{Q}^{\mathrm{pn}}$ and $\mathbf{Q}^{\mathrm{nn}}$ . The tracking results include predictions from $\mathbf{Q}^{\mathrm{pp}}$ , $\mathbf{Q}^{\mathrm{pn}}$ and $\mathbf{Q}^{\mathrm{nn}}$ . We illustrate the pipeline of OneTrack in Fig. 3. + +During inference, we categorize object queries into four groups based on the detection and tracking confidence scores of their predictions. Queries with high detection or tracking confidence scores over $T_{p}$ are classified as positive samples for the respective tasks. We set $T_{p} = 0.2$ in our experiments. + +At last, in both the training or inference process, we propagate the Top- $K$ object queries with high confidence scores from both detection and tracking results to the next frame as $\hat{\mathbf{Q}}_t$ . In particular, we will still propagate the "locked" track queries to the future frames even if they are not in the Top- $K$ object queries. This is for addressing challenging tracking scenarios, such as an object being occluded for several frames. We propagate "locked" track queries through this for up to $\mathrm{A_{age}} = 3$ frames into future frames. + +Following [31,36], we also preserve the propagated object queries as the memory queue for the detection or tracking process for up to 4 frames. The mem + +![](images/57a7132ff371c2f733db751d0b580166e2d013ab1032d9adada585c64d0924ee.jpg) +Fig. 3: Training pipeline of OneTrack. For each decoder layer in OneTrack, We identify object queries into four groups based on their positivity as training samples in detection and tracking tasks. Task-specific queries that are positive samples only for one task will only backpropagate gradients in the respective task-specific classification head to prevent conflicted classification gradients. We dynamically mask all attention across positive sample queries only for detection or tracking. Best viewed in color. + +ory queues for the detection and tracking processes are updated independently. Those memory queries are provided to the decoder for performing hybrid attention proposed by [36]. As keys and values for the hybrid attention layer, detection and tracking memory queries are masked for $\mathbf{Q}^{\mathrm{np}}$ and $\mathbf{Q}^{\mathrm{pn}}$ , respectively. We show the design of decoder layers in OneTrack in Fig. 4. + +As demonstrated in Table 4, with the proposed dynamic hybrid-attention mask, the true joint training of detection and tracking can be achieved. This allows for training detection and tracking in a single training stage and inferring them in a single forward pass with a unified model. This feature of OneTrack significantly reduces its training cost and inference latency. + +# 3.4 Modified Tracking Classification Loss + +In addition to introducing the true joint training of detection and tracking to reduce the detection capability degradation of trackers, we also adjusted the tracking classification loss to suppress the in fact inaccurate predictions. As previously mentioned, previous end-to-end trackers regard track queries whose predictions are associated with ground truths as positive samples equally, regardless of the precision of their current predictions. This encourages the tracker to assign high confidence to the in fact inaccurate predictions from track queries. Therefore we propose to restrain the positive assignment of those in fact inaccurate predictions. + +In previous end-to-end trackers, the assignment result of a ground truth object $g$ can be denoted as a one-hot vector: $A_{g}^{n\times 1} = [0,\dots ,1,\dots ,0]$ , where $n$ is the number of object queries and the ground truth object is exclusively + +![](images/ff3a3d84509c258c6433f8b8b2e37f0483d22552795a5b86c9e464f0eb238912.jpg) +Fig. 4: Design of decoder layers in OneTrack. In addition to current object queries, queries in detection memory and tracking memory are fed to the hybrid attention layer as keys and values. As keys and values, $\mathbf{Q}^{\mathrm{pn}}$ and detection memory are masked to $\mathbf{Q}^{\mathrm{np}}$ . As keys and values, $\mathbf{Q}^{\mathrm{np}}$ and tracking memory are masked to $\mathbf{Q}^{\mathrm{pn}}$ . + +assigned to the $i^{th}$ object query. This assignment is unaffected by the matching cost between the $i^{th}$ object query and $g$ if they have been associated in recent frames. + +In OneTrack, we propose to suppress the assignment weight between ground truths and queries that are locked but have high matching costs. We adjust the assignment result of locked ground truth $g$ as follows: + +$$ +A _ {g} ^ {n \times 1} = [ 0, \dots , W _ {g i}, \dots , 0 ], \tag {10} +$$ + +$$ +W _ {g i} = - \frac {1}{\gamma} * \operatorname {M a x} \left(\left(\operatorname {d i s t} (g, q ^ {i}) - 0. 5\right), 0\right) + 1, \tag {11} +$$ + +where $\mathrm{dist}(g, q^i)$ represents the distance between the centers of ground truth box $g$ and the predicted box of the locked query $q^i$ . $\gamma$ is a hyper-parameter greater than zero. The assigned weight $W_{gi}$ between $g$ and $q^i$ decreases linearly from 1.0 as $\mathrm{dist}(g, q^i)$ increases from $0.5\mathrm{m}$ . The assignment weight serves as the weighting factor for calculating box regression loss and the targets for predicting tracking confidence score. Instead of Focal Loss [24], we employ the Quality Focal Loss proposed by [20] to supervise the tracking classification head, given the continuous $0\sim 1$ confidence label. + +# 4 Experiments + +In this section, we first introduce our detailed experimental setup. Then we provide comparisons with recent works on nuScenes camera-based 3D tracking benchmark. Ablation studies on each component are presented in the last part. + +Table 3: Compare with recent works on nuScenes dataset. $\dagger$ : trained from scratch. OneTrack-F / OneTrack-S: OneTrack trained under full-resolution / half-resolution as described in Sec. 4.2. + +
Backbonee2eResolutionAMOTA↑AMOTP↓MOTA↑RECALL↑IDS↓
Validation Split
QD3DT [13]R101✓1600×9000.2421.5180.21839.9%5646
MUTR3D [48]R101✓1600×9000.2941.4980.26742.7%3822
CC-3DT [10]R101×1600×6400.4291.2570.38553.4%2219
PF-Track-S [31]V2-99✓800×3200.4081.3430.37650.7%166
PF-Track-F [31]V2-99✓1600×6400.4791.2270.43559.0%181
DQTrack [21]V2-99✓800×3200.4461.251--1193
OneTrack-S†V2-99✓800×3200.4921.1220.40958.8%315
OneTrack-F†V2-99✓1600×6400.5481.0880.47961.8%389
Test Split
QD3DT [13]R101✓1600×6400.2171.5500.19837.5%6856
MUTR3D [48]R101✓1600×6400.2701.4940.23541.1%6018
SRCN3D [32]V2-99×1600×6400.3981.3170.35953.8%3334
CC-3DT [10]R101×1600×6400.4101.2740.35753.8%3334
PF-Track [31]V2-99✓1600×6400.4341.2520.37853.8%249
DQTrack [21]V2-99✓1600×6400.5231.0960.44462.2%1204
OneTrack-F†V2-99✓1600×6400.5541.0210.46160.8%481
+ +# 4.1 Datasets and Metrics + +We evaluate our proposed method on the nuScenes dataset. nuScenes [5] dataset is a large-scale autonomous driving benchmark containing 1000 multi-modal videos. Videos recorded by six cameras are divided into 700, 150, and 100 scenes for training, validation, and testing, respectively. 3D box annotations of 10 object classes are provided for keyframes at $2\mathrm{Hz}$ . For nuScenes 3D tracking benchmark, we report AMOTA, AMOTP [39], MOTA [3], Recall and identity switches (IDS). For the 3D detection task, we report mean Average Precision (mAP) and nuScenes Detection Score (NDS). + +# 4.2 Implementation Details + +Our implementation is mainly based on StreamPETR [36]. We define two implementation settings: The full-resolution setting for state-of-the-art comparison and the half-resolution setting for ablation studies. We conduct experiments using V2-99 [18] as the image backbone. OneTrack is trained using AdamW [28] optimizer with a batch size of 8 and a base learning rate of 4e-4. The cosine annealing policy is employed to adjust the learning rate. We use 644 initialized detection queries in each frame and propagate the Top-256 object queries into the next frame. When training, we record all matched ground truth and object query pairs as "locked" pairs for label assignment. An object query matched with a ground truth will be recorded as a track query. As previously mentioned, a "locked" pair will be preserved in memory for up to $A_{\mathrm{age}} = 3$ frames, even if + +Table 4: Ablation study on components breakdown. + +
Training taskInitialized from pre-trained detectorGradient coordinationDynamic High-cost assign masksuppressAMOTA↑IDS↓NDS↑mAP↑
detection------0.5700.479
tracking×--×0.3051210.4550.313
tracking✓--×0.3754660.5030.409
joint××××0.4252460.5170.410
joint×✓××0.4412310.5550.462
joint×✓✓×0.4692610.5750.484
joint×✓✓✓0.4923150.5660.478
+ +the respective track query is not in the Top- $K$ object queries in each frame. For track queries that failed to predict confident prediction over $\mathrm{A}_{\mathrm{age}}$ frames, its "locked" ground truth will be assigned to other "free" object queries. During inference, we initialize an object query as a track query if it predicts a 3D box with confidence over $T_{\mathrm{conf}} = 0.4$ . Additionally, all track queries are preserved and propagated across frames for up to $\mathrm{A}_{\mathrm{age}} = 3$ frames. Predictions from the same track query propagated across frames will be assigned the same tracking ID. We output all detection predictions with confidence scores over 0.05 and all tracking predictions with confidence scores over 0.3. We introduce other setting-specific training configurations as follows. + +OneTrack-F: We crop input images from the initial resolution of $1600 \times 900$ to $1600 \times 640$ . We then conduct end-to-end training for 48 epochs. OneTrack-S: We first crop the input images to the resolution of $1600 \times 640$ and downsample them to $800 \times 320$ . Models are trained for 24 epochs. + +# 4.3 Comparisons with State-of-the-arts + +As shown in Table 3, OneTrack outperforms all previous trackers in both the validation and test set of nuScenes dataset by a significant margin. On the validation set, OneTrack surpasses previous state-of-the-art PF-Track [31] by $10.2\%$ AMOTA. On the test set, OneTrack surpasses DQTrack [21] by $3.3\%$ AMOTA. Notably, OneTrack reports fewer than half of the identity switches compared to those reported by DQTrack. It is also worth mentioning that OneTrack is the only model in Table 3 that is trained from scratch. + +# 4.4 Ablations + +We conduct ablation studies on each proposed component with OneTrack-S on the nuScenes validation set. + +Components breakdown. The breakdown analysis of proposed components is presented in Table 4. Here "det" or "trk" in the first row stands for only training the model for detection or tracking tasks. And "joint" stands for jointly training the model for both detection and tracking tasks, with an extra classification head attached to the model. As shown in Table 4, the model trained solely + +![](images/05c7a253a57041435d97f5902819b770cbc10364af1d2b03952b707895c5a2ae.jpg) +Fig. 5: Ablation on hyper-parameters. (a) On parameter $\gamma$ in the modified tracking losses. (b) On the max preserved frames of track queries $A_{\mathrm{age}}$ . (c) On the score threshold $T_{\mathrm{conf}}$ for track query initialization. + +![](images/39a1941da2f64dc7cac4c2da3d61991fa2d78216e77ecf9e10764e5c189518b9.jpg) + +![](images/3aaa808cd8f770f592e01716dfb9d8db137d4c291894b48f0dcc006ff9519da5.jpg) + +for the tracking task suffers from severe degradation in detection performance, which results in their inferior detection performance compared to the detector. Initializing the tracker from a detector checkpoint or trivially supervising the model with both detection and tracking losses can only slightly alleviate this problem. Our proposed components including classification gradient coordination, dynamic attention mask, and high-cost tracking assignment suppress, are all validated as highly effective in achieving the true joint training of detection and tracking tasks. It is noteworthy that OneTrack without modifying its tracking loss achieves an even slightly better detection performance than the model trained solely as a detector. This observation suggests the potential for further enhancing the detection process through the integrated tracking process in a unified model. + +Modified tracking loss. In Fig. 5(a), we conduct ablation studies on the hyper-parameter $\gamma$ in the modified tracking loss for suppressing positive assignments of inaccurate predictions. OneTrack achieves optimal tracking performance when $\gamma$ is set to 7.5. While the detection performance of OneTrack shows improvement as $\gamma$ increases, this gain becomes limited when $\gamma$ exceeds 7.5. + +Track age. We conduct ablation studies on the maximum preservation age $A_{\mathrm{age}}$ of track queries. As illustrated in Fig. 5(b), a longer track age will lead to more identity switches and decreased detection performance. We speculate that this is because propagating too many outdated track queries will mislead the object feature extraction process in the current frame. + +Confidence threshold for track query initialization. We assess the impact of $T_{\mathrm{conf}}$ in Fig. 5(c). The AMOTA performance will slightly increase as $T_{\mathrm{conf}}$ increases, which results from the more accurate tracking predictions on average. However, a higher $T_{\mathrm{conf}}$ reduces the initialization of track queries from detection queries with low confidence. This limitation hinders tracking objects under poor observation conditions, resulting in a higher number of identity switch cases. + +Separated classification heads. As shown in Table 5, when its two classification heads share weights except for the final linear layer, the tracking and detection performance of OneTrack is slightly worse than when the classification heads are independent. We speculate that this is because the separation of classification heads leads to the easier convergence of both heads. + +Table 5: Ablation on the separated classification heads for two tasks. + +
Cls headsAMOTA↑IDS↓NDS↑mAP↑
shared49.03760.5590.473
separated49.23150.5660.478
+ +Table 6: Computational cost comparison with recent works. All training time costs and inference latencies are measured on a single A100 GPU. *: PF-Track is first trained as a detector and then as a tracker for 12 epochs each. We train DQTrack and OneTrack from scratch for 24 epochs. + +
MethodBackboneResolutionTraining time costInference latency
PF-TrackV2-99800×32034h+82h*124.4ms
DQTrackV2-99800×320176h115.2ms
OneTrackV2-99800×32087h83.7ms
+ +Computational cost. To ensure fair comparison, we apply gradient checkpointing [7] on the decoder layers of all trackers. We measure their training cost and inference latency on a single A100 GPU. As shown in Table 6, the training cost of OneTrack is only half that of previous state-of-the-art trackers. The inference latency of OneTrack is also significantly lower. We attribute the low computational cost of OneTrack to its elegant model design. We do not extend the tracker with various plug-in modules [31] or separate the detection and tracking stages of the tracker [21], which will introduce additional computational burdens. Notably, OneTrack is trained from scratch while PF-Track and DQTrack rely on extra training stages for achieving high tracking performance. + +# 5 Conclusion + +In this work, we delve into the optimization conflict between detection and tracking in end-to-end trackers, which arises from the partially conflicted classification gradients between two tasks. Based on this observation, we propose to identify the conflicted positive samples between two tasks through query grouping and solve the conflicted gradients with the proposed gradient coordination, dynamic attention mask, and suppression tracking positive samples of low quality. On the nuScenes dataset, OneTrack outperforms all previous end-to-end trackers. + +# Acknowledgement + +This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program. + +# References + +1. Ali, A., Jalil, A., Niu, J., Zhao, X., Rathore, S., Ahmed, J., Aksam Iftikhar, M.: Visual object tracking—classical and contemporary approaches. Frontiers of Computer Science 10, 167–188 (2016) +2. Benbarka, N., Schröder, J., Zell, A.: Score refinement for confidence-based 3d multi-object tracking. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8083-8090. IEEE (2021) +3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing 2008, 1-10 (2008) +4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016) +5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) +6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +7. Chen, T., Xu, B., Zhang, C., Guestrin, C.: Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174 (2016) +8. Chiu, H.k., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: ICRA (2021) +9. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4870-4880 (2023) +0. Fischer, T., Yang, Y., Kumar, S., Sun, M., Yu, F.: CC-3DT: panoramic 3d object tracking via cross-camera fusion. In: CoRL. Proceedings of Machine Learning Research, vol. 205, pp. 2294–2305. PMLR (2022) +1. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science 18(5), 185706 (2024) +2. He, J., Huang, Z., Wang, N., Zhang, Z.: Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5299-5309 (2021) +3. Hu, H.N., Yang, Y.H., Fischer, T., Darrell, T., Yu, F., Sun, M.: Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +4. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022) +5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) +6. Kim, A., Osep, A., Leal-Taixe, L.: Eagermot: 3d multi-object tracking via sensor fusion. In: 2021 IEEE International conference on Robotics and Automation (ICRA). pp. 11315-11321 (2021) +7. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) + +18. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) +19. Li, J., Ding, Y., Wei, H.L., Zhang, Y., Lin, W.: Simpletrack: Rethinking and improving the jde approach for multi-object tracking. Sensors 22(15), 5863 (2022) +20. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020) +21. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18302-18311 (2023) +22. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023) +23. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022) +24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +25. Liu, C., Chen, X.F., Bo, C.J., Wang, D.: Long-term visual tracking: review and experimental comparison. Machine Intelligence Research 19(6), 512-530 (2022) +26. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022) +27. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023) +28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +29. Marinello, N., Proesmans, M., Van Gool, L.: Tripletrack: 3d object tracking using triplet embeddings and lstm. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4499-4509 (2022) +30. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8844-8854 (2022) +31. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17928-17938 (2023) +32. Shi, Y., Shen, J., Sun, Y., Wang, Y., Li, J., Sun, S., Jiang, K., Yang, D.: Srcn3d: Sparse r-cnn 3d surround-view camera object detection and tracking for autonomous driving. arXiv preprint arXiv:2206.14451 (2022) +33. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10860-10869 (2021) +34. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: ICCV. pp. 10840-10849. IEEE (2021) + +35. Wang, Q., Chen, Y., Pang, Z., Wang, N., Zhang, Z.: Immortal tracker: Tracklet never dies. arXiv preprint arXiv:2111.13672 (2021) +36. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3621-3631 (2023) +37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022) +38. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5096-5105 (2023) +39. Weng, X., Kitani, K.: A baseline for 3d multi-object tracking. arXiv preprint arXiv:1907.03961 1(2), 6 (2019) +40. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. IROS (2020) +41. Weng, X., Wang, Y., Man, Y., Kitani, K.: Gnn3dmot: Graph neural network for 3d multi-object tracking with multi-feature learning. arXiv preprint arXiv:2006.07327 (2020) +42. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: 2017 IEEE international conference on image processing (ICIP). pp. 3645-3649. IEEE (2017) +43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023) +44. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. CVPR (2021) +45. Yu, E., Wang, T., Li, Z., Zhang, Y., Zhang, X., Tao, W.: Motrv3: Releasefetch supervision for end-to-end multi-object tracking. arXiv preprint arXiv:2305.14298 (2023) +46. Zaech, J.N., Liniger, A., Dai, D., Danelljan, M., Van Gool, L.: Learnable online graph representations for 3d multi-object tracking. IEEE Robotics and Automation Letters 7(2), 5103-5110 (2022) +47. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022) +48. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4537-4546 (2022) +49. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision 129, 3069-3087 (2021) +50. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22056-22065 (2023) +51. Zhao, H., Zhang, J., Chen, Z., Yuan, B., Tao, D.: On robust cross-view consistency in self-supervised monocular depth estimation. Machine Intelligence Research 21(3), 495-513 (2024) \ No newline at end of file diff --git a/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/images.zip b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b7baf2575436e3b3cbc004dbc9eea8982c91c0c --- /dev/null +++ b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:874dc2802202d58889b2f81aa9dcec08e91c2e2f9f726d5f0bef2877068bb67c +size 476758 diff --git a/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/layout.json b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b68140c3552cbaccde3b18ffbbc3b14cb110349c --- /dev/null +++ b/2024/OneTrack_ Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers/layout.json @@ -0,0 +1,9295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 111, + 467, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 111, + 467, + 163 + ], + "spans": [ + { + "bbox": [ + 147, + 111, + 467, + 163 + ], + "type": "text", + "content": "OneTrack: Demystifying the Conflict Between Detection and Tracking in End-to-End 3D Trackers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "spans": [ + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "text", + "content": "Qitai Wang" + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "text", + "content": ", Jiawei He" + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "text", + "content": ", Yuntao Chen" + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "text", + "content": ", and Zhaoxiang Zhang" + }, + { + "bbox": [ + 134, + 185, + 478, + 199 + ], + "type": "inline_equation", + "content": "^{1,2,3,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 207, + 476, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 476, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 476, + 220 + ], + "type": "text", + "content": "1 School of Future Technology, University of Chinese Academy of Sciences (UCAS)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 220, + 472, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 220, + 472, + 241 + ], + "spans": [ + { + "bbox": [ + 140, + 220, + 472, + 241 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 220, + 472, + 241 + ], + "type": "text", + "content": " NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA) {wangqitai2020,zhaoxiang.zhang}@ia.ac.cn,jwhe2024@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 241, + 432, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 241, + 432, + 262 + ], + "spans": [ + { + "bbox": [ + 182, + 241, + 432, + 262 + ], + "type": "text", + "content": "3 Center for Artificial Intelligence and Robotics, HKISI, CAS chenyuntao08@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 262, + 398, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 262, + 398, + 274 + ], + "spans": [ + { + "bbox": [ + 215, + 262, + 398, + 274 + ], + "type": "text", + "content": "4 Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 304, + 455, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 304, + 455, + 502 + ], + "spans": [ + { + "bbox": [ + 159, + 304, + 455, + 502 + ], + "type": "text", + "content": "Abstract. Existing end-to-end trackers for vision-based 3D perception suffer from performance degradation due to the conflict between detection and tracking tasks. In this work, we get to the bottom of this conflict, which was vaguely attributed to incompatible task-specific object features previously. We find the conflict between the two tasks lies in their partially conflicted classification gradients, which stems from their subtle difference in positive sample assignments. Based on this observation, we propose to coordinate those conflicted gradients from object queries with contradicted polarity in the two tasks. We also dynamically split all object queries into four groups based on their polarity in the two tasks. Attention between query sets with conflicted positive sample assignments is masked. The tracking classification loss is modified to suppress inaccurate predictions. To this end, we propose OneTrack, the first one-stage joint detection and tracking model that bridges the gap between detection and tracking under a unified object feature representation. On the nuScenes camera-based object tracking benchmark, OneTrack outperforms previous works by " + }, + { + "bbox": [ + 159, + 304, + 455, + 502 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 159, + 304, + 455, + 502 + ], + "type": "text", + "content": " AMOTA on the validation set and by " + }, + { + "bbox": [ + 159, + 304, + 455, + 502 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 159, + 304, + 455, + 502 + ], + "type": "text", + "content": " AMOTA on the test set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "spans": [ + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "text", + "content": "Keywords: 3D Tracking " + }, + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "text", + "content": " Camera-based Detection and Tracking " + }, + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "text", + "content": " End-to-end Tracking " + }, + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 513, + 453, + 535 + ], + "type": "text", + "content": " Gradient Coordination" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Reasoning about the location and trajectory of surrounding objects is a fundamental task for autonomous driving and robotic navigation systems. Due to the robustness and low cost of cameras, vision-based 3D detection and tracking have received widespread attention from the research community [14,22,23,26,36,37]. Recently, several approaches [21, 31, 48] have been proposed to improve 3D multi-object tracking (MOT) by jointly optimizing the detection and tracking pipeline. Based on transformer architecture, those end-to-end 3D trackers use" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 183, + 157, + 428, + 237 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison with previous works trying to solve the conflict between detection and tracking. *: DQTrack relies on frozen pre-trained detectors when compared with state-of-the-art trackers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 157, + 428, + 237 + ], + "lines": [ + { + "bbox": [ + 183, + 157, + 428, + 237 + ], + "spans": [ + { + "bbox": [ + 183, + 157, + 428, + 237 + ], + "type": "table", + "html": "
Methodsettinge2eone-stagedetection supervisiontrain from scratch
MOTRv2 [50]2D××××
MOTRv3 [45]2D✓✓×✓
DQTrack [21]3D✓×✓✓*
OneTrack3D✓✓✓✓
", + "image_path": "d337e36d1dafa074696ba345466c9ff306c474823b3aa0dc3d92487a22bf11f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 479, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 479, + 288 + ], + "type": "text", + "content": "track queries to track objects, being free from the dependence on offline detectors and post-processing operations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 469 + ], + "type": "text", + "content": "However, most end-to-end 3D trackers suffer from the optimization conflict between detection and tracking, a problem also widely recognized in 2D MOT studies [45, 50]. This conflict results in the inferior precision and recall of end-to-end trackers compared to their baseline detectors. Previous works roughly attribute the conflict between two tasks to differences in required features for detection and tracking and attempted to alleviate this problem. Some of them train their model as detectors before the tracking training process [31, 33] or take an off-the-shelf detector as guidance [50]. Some separate the detection and association stages as well as their object feature representation to avoid conflict, such as DQTrack [21]. Although much progress has been made, training a one-stage end-to-end tracker from scratch remains challenging. It is also challenging to train trackers jointly under both detection and tracking supervision. We summarize recent works in Table 1. In this work, we aim to accomplish the true joint training of detection and tracking from scratch with a one-stage model and unified feature representation for both tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": "We conduct a pilot study to investigate how and where the two tasks conflict during model optimization. As illustrated in Table 2, only jointly supervising the detector with an extra classification loss computed under the ground truth assignments for tracking will lead to severe performance degradation of detectors. Therefore, we conclude that the conflict between the two tasks originates from their conflicting classification gradients." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we precisely identify and coordinate the conflicted classification gradient between the two tasks, which originates from their partially conflicted positive sample definition. We propose OneTrack, the first model capable of achieving the true joint training of detection and tracking in a unified, one-stage model. OneTrack follows the \"tracking by track queries\" paradigm. During the training process, we assign ground truth annotations to object queries twice based on positive sample definitions for detection or tracking. We dynamically categorize all object queries into four groups based on their polarity in two tasks. We have two lightweight classification heads in OneTrack which are respectively responsible for classifying detection or tracking positive samples. Queries that are" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 491, + 264 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Pilot study on supervising 3D detector under additional ground truth assignments for tracking. The stronger color means more degradation in performance. Experiments indicate the conflict between detection and tracking tasks lies in their classification supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 491, + 264 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 491, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 491, + 264 + ], + "type": "table", + "html": "
Detection ClassificationDetection RegressionTracking ClassificationTracking RegressionAMOTA↑IDS↓NDS↑mAP↑
✓✓××--0.5700.479
××✓✓0.305121--
✓✓✓×0.4043630.5120.432
✓✓×✓0.41785120.5660.473
✓✓✓✓0.4052550.5280.428
", + "image_path": "9be26db0d9e9b5b926d95bc1ac7a702c299deb9588e5258e4a4ef1f66d7f9e7f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 286, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 479, + 346 + ], + "type": "text", + "content": "positive samples only for one task will only backpropagate classification gradients in the respective head to reduce conflicted gradients. We dynamically mask all attention between conflicted positive queries in each decoder layer to prevent unexpected query competition. We also adjust the tracking loss to suppress inaccurate predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 347, + 356, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 347, + 356, + 358 + ], + "spans": [ + { + "bbox": [ + 147, + 347, + 356, + 358 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 365, + 480, + 496 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 365, + 479, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 479, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 479, + 400 + ], + "type": "text", + "content": "- We reveal that the conflict between detection and tracking tasks lies in their diverged definition of positive samples and their partially conflicted classification gradients." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 401, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 472 + ], + "type": "text", + "content": "- We propose OneTrack, the first model capable of completely addressing the conflict between detection and tracking under a unified object feature representation. OneTrack can be trained as both a detector and a tracker from scratch in a single training stage and can perform tracking and detection in a one-stage fashion. The detection performance of OneTrack is comparable to detectors, while its tracking performance surpasses all previous 3D trackers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 473, + 479, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 473, + 479, + 496 + ], + "spans": [ + { + "bbox": [ + 138, + 473, + 479, + 496 + ], + "type": "text", + "content": "- Our method establishes a new state-of-the-art on nuScenes MOT benchmark, surpassing previous methods by over " + }, + { + "bbox": [ + 138, + 473, + 479, + 496 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 138, + 473, + 479, + 496 + ], + "type": "text", + "content": " AMOTA." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 514, + 242, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 242, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 242, + 526 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 538, + 302, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 302, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 302, + 551 + ], + "type": "text", + "content": "2.1 Query Propagation in MOT" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Early works in 2D or 3D MOT most follow the \"tracking by detection\" paradigm [4] and associate detection results across frames through post-processing [1, 2, 8, 9, 12, 13, 16, 19, 25, 35, 40-42, 44, 46, 49]. Based on DETR [6] which introduced object queries to detect objects, MOTR [47] and Trackformer [30] extended the concept of object query to track query by propagating object queries across frames. By tracking objects through track queries, MOTR and Trackformer can perform end-to-end joint 2D detection and tracking without post-processing. However, they seriously suffer from the conflict between detection and tracking. MOTRv2 [50] and MOTRv3 [45] concentrate on this conflict. MOTRv2 proposes" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "using an off-the-shelf detector to guide the tracker, which is effective but damages the end-to-end fashion of the tracker. MOTRv3 proposes to balance the label assignment between detection queries and track queries to achieve a better tradeoff between the detection and tracking performance of the model. MUTR3D [48] and PF-Track [31] applied the \"tracking by track queries\" paradigm to 3D MOT. MUTR3D [48] performs camera-based tracking in an end-to-end fashion by introducing 3D track queries to model spatial and appearance coherent tracks. PF-Track [31] further refine the tracks and predict the trajectories of objects with cross-object and cross-frame attention between object queries. Recently, DQTrack [21] proposes to separate the detection process and association process of trackers to avoid their conflicts in object feature representation. None of the previous joint detection and tracking approaches can truly bridge the gap between detection and tracking with a one-stage model and a unified object feature representation. Neither can they train the model from scratch and get rid of the reliance on pre-trained detectors while detecting and tracking objects with high precision and recall performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 331, + 282, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 282, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 282, + 342 + ], + "type": "text", + "content": "2.2 Camera-based 3D MOT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 357, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 490 + ], + "type": "text", + "content": "Camera-based 3D MOT has gained remarkable progress recently thanks to the developments in camera-based 3D detection and depth estimation [11, 14, 15, 22, 23, 26, 27, 36-38, 43, 51]. Early methods in camera-based 3D MOT perform tracking in 2D first and then lift the tracks to 3D space [34]. CC-3DT [10] fuses multi-view object features to enhance associating objects across different views. QD-3DT [13] perform 2D association first and enhance the instance association with the depth order and motions of 3D objects. TripletTrack [29] extracts local object feature embeddings and motion descriptors with CNN or LSTM to measure the affinity between objects. MUTR3D [48], PF-Track [31], and DQTrack [21] recently established new state-of-the-art tracking performance by applying the transformer-based tracking architecture in 3D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 202, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 202, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 202, + 528 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 485, + 666 + ], + "type": "text", + "content": "The model design of OneTrack is conceptually concise. We add an additional lightweight classification branch to a basic transformer-based tracker. Both ground truth assignments for the detection and tracking tasks are employed to supervise the model. We begin by reviewing the design of transformer-based trackers in Sec. 3.1. Then in Sec. 3.2, we propose to identify the conflicted positive sample assignments between detection and tracking through query grouping and prevent the contradicted gradient from back-propagating. In Sec. 3.3, we introduce the dynamic attention mask to further prevent gradient conflict between the two tasks. Additionally, the tracking classification loss is modified to suppress inaccurate boxes predicted by track queries, as described in Sec. 3.4." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 430, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 430, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 430, + 129 + ], + "type": "text", + "content": "3.1 Reviewing the \"Tracking by Track Queries\" Paradigm" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "text", + "content": "As previous transformer-based trackers [21,30,31,45,47,48,50], OneTrack detect and track 3D objects with object queries. At each timestamp " + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "text", + "content": ", given multiview images " + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "text", + "content": ", the 3D detection task aims to predict a set of 3D bounding boxes " + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = \\{\\mathbf{b}^1, \\mathbf{b}^2, \\dots, \\mathbf{b}^n\\}" + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "text", + "content": " and their confidence scores " + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\{s^1, s^2, \\dots, s^n\\}" + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "text", + "content": " to capture 3D objects. Meanwhile, 3D MOT requires a consistent ID for each object across frames. We represent the outputs for the 3D MOT task by " + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}, \\mathbf{S}, \\mathbf{ID}\\}" + }, + { + "bbox": [ + 130, + 133, + 482, + 216 + ], + "type": "text", + "content": ", including the predicted boxes, confidence scores, and ID of predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "content": "In each frame, OneTrack first receives a set of object queries " + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\{\\mathbf{q}^1,\\mathbf{q}^2,\\dots \\}" + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "content": ". Following previous works [21, 31, 47, 48], " + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "content": " consists of track queries propagated from the previous frame and a fixed number of new detection queries " + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^d" + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "content": ". The query propagation process can be formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 269, + 271, + 480, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 271, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 269, + 271, + 480, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {t} = \\hat {\\mathbf {Q}} _ {t - 1} \\cup \\mathbf {Q} ^ {d}, \\tag {1}", + "image_path": "a1d1896c6f707c0d5c75ef05f5f2791f53d10e5a4cdf06f627b91feff092ffd1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Q}}_{t - 1}" + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "content": " represents the propagated queries from the last frame. We denote " + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "content": " in the subsequent descriptions for simplicity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "text", + "content": "In each frame, given the multi-view image features " + }, + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "text", + "content": " extracted with image encoder, OneTrack utilizes a transformer decoder to extract object features. The decoder comprises six transformer decoder layers. In the " + }, + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "text", + "content": " layer " + }, + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 317, + 481, + 363 + ], + "type": "text", + "content": " is updated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 372, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 372, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 241, + 372, + 480, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\operatorname {D e c o d e r L a y e r} _ {i} (\\mathbf {F}, \\mathbf {Q}, \\mathbf {M}). \\tag {2}", + "image_path": "06212fffd0e9d840c067316e9c101722cd5ba0d452dcde87d4dd74099f5f6229.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 390, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 485 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 390, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 390, + 482, + 485 + ], + "type": "text", + "content": " represents the attention mask for the self-attention process in decoder layers. In end-to-end trackers, attention between all object queries is usually not masked. Then the object queries are fed into the classification and regression heads to generate prediction boxes. Ground truths are then assigned to the predictions for the calculation of classification loss and regression loss. Recent works [21,31] usually preserve historical track queries from a few past frames as tracking memory. These stored queries serve as references to the tracking process in the current frame." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": "Instead of associating boxes across frames with post-processing techniques, the \"tracking by track queries\" paradigm tracks objects with track queries propagated across frames. Therefore during the training of trackers, a ground truth instance is assigned to a specific track query in future frames once they are associated in the current frame, regardless of the accuracy of the assigned query's predictions in future frames. We denote these ground truth-query pairs as \"locked\" [45] pairs, and the remaining ground truths and queries are denoted as \"free\" ground truths or queries. In contrast, for detectors, positive samples are simply defined as the most precise prediction in each frame. Hence, a positive sample for detection may be classified as a negative sample in tracking and be suppressed, and vice versa, as illustrated in Fig.1. This originates from ground truth objects being assigned to different object queries in the two tasks. This conflict will lead to conflicted classification gradients between the two tasks, hindering the convergence of the model. We found that those conflicted gradients are primarily responsible for the precision and recall degradation of end-to-end trackers." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 117, + 425, + 261 + ], + "blocks": [ + { + "bbox": [ + 194, + 117, + 425, + 261 + ], + "lines": [ + { + "bbox": [ + 194, + 117, + 425, + 261 + ], + "spans": [ + { + "bbox": [ + 194, + 117, + 425, + 261 + ], + "type": "image", + "image_path": "1f12da4d09ecadf1a37c3a42eb4d7c09506729af94a8352e6fb93e14db1faad9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: Difference in positive sample definition for query-based detector and tracker. With object query propagated from the last frame, the end-to-end tracker ensures that queries assigned as positive samples in the previous frame (like " + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}_{\\mathrm{C}}^{\\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "text", + "content": ") maintain their pairing with the same ground truth in the current frame. Meanwhile, the detector always takes the most precise predictions (like " + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}_{\\mathrm{X}}" + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "text", + "content": ") as positive samples. In such case, the polarity of " + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}_{\\mathrm{C}}^{\\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}_{\\mathrm{X}}" + }, + { + "bbox": [ + 130, + 269, + 482, + 338 + ], + "type": "text", + "content": " as training samples in the two tasks is opposite." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 360, + 453, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 360, + 453, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 360, + 453, + 383 + ], + "type": "text", + "content": "3.2 Solving the Conflict Between Detection and Tracking via Gradient Coordination" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": "We first try a trivial and straightforward solution to avoid the classification gradient conflict between the two tasks. We attach two classification heads to the model, one for detection classification and the other for tracking classification. Since there is no obvious conflict in box regression for the detection or tracking tasks, we use a shared head for regressing 3D boxes in the two tasks. Then all predicted boxes are assigned to ground truths twice following the positive sample definition in detection or tracking as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 226, + 487, + 481, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 226, + 487, + 481, + 520 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} _ {\\mathrm {d e t}} = \\arg \\min _ {\\sigma} \\sum_ {k} ^ {N} \\mathcal {L} _ {\\operatorname {m a t c h}} \\left(\\mathbf {g} ^ {k}, \\mathbf {b} ^ {\\sigma (k)}\\right), \\tag {3}", + "image_path": "ec373199af4364021851af93e205f169d2bcc158dcb00720b30814fa7025a626.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 225, + 522, + 481, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 225, + 522, + 481, + 555 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} _ {\\mathrm {t r k}} ^ {\\text {f r e e}} = \\arg \\min _ {\\sigma} \\sum_ {k} ^ {M} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\mathbf {g} ^ {k}, \\mathbf {b} ^ {\\sigma (k)}\\right), \\tag {4}", + "image_path": "74696a0fdf64c00d862826b37c116c309c5df363cffcb15563f1ec64cbb76037.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 555, + 481, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 555, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 227, + 555, + 481, + 572 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} _ {\\mathrm {t r k}} = \\hat {\\sigma} _ {\\mathrm {t r k}} ^ {\\mathrm {f r e e}} \\cup \\sigma_ {\\mathrm {t r k}} ^ {\\mathrm {l o c k e d}}, \\tag {5}", + "image_path": "efeb50c319a399b5a25d80735586af78898402016aa15c3f499fdec2b43a536a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{g}^k\\}_{k=1}^N" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " ground truths(GTs), in which " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " of them are free GTs for tracking. " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{b}^i\\}_{i=1}^n" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " is the set of predicted boxes. " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " is the possible permutation of " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " elements, following the notation in [6]. For example, " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\mathbf{g}^k, \\mathbf{b}^{\\sigma(k)})" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " means the " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " ground truth box is assigned to the " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\sigma(k)^{th}" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " predicted box. " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " denotes the matching cost between a ground truth and a prediction from an object query. " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}_{\\mathrm{det}}" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}_{\\mathrm{trk}}" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " are the ground truth assignments for detection and tracking, respectively. " + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{trk}}^{\\mathrm{locked}}" + }, + { + "bbox": [ + 131, + 581, + 482, + 668 + ], + "type": "text", + "content": " stands for the locked ground truth and query pairs inherited" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "from the last frame. Classification losses and regression losses are then computed for gradient back-propagation in each decoder layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 294 + ], + "type": "text", + "content": "However, we found that merely separating the classification heads for two tasks and supervising the model under both detection and tracking ground truth assignments can neither obtain a satisfying detector nor tracker, as shown in Table 4. Clearly, simply separating the classification heads cannot solve the conflict between the sample classification in detection and tracking. This is because as previously mentioned, a part of positive samples for training detectors will be assigned as negative samples for training trackers, and vice versa. Therefore even when the classification heads of two tasks are separated, classification gradients on those samples will still be conflicted when back-propagated to the decoder layers. To address this, we propose to decompose the classification gradients in both tasks to identify and avoid their partial conflict on those conflicted samples. We refer to this design as classification gradient coordination between the two tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "text", + "content": "We dynamically categorize object queries into four query groups in each decode layer based on their polarity in both tasks: positive for both tasks; positive for detection and negative for tracking; negative for detection and positive for tracking; and negative for both tasks. We denote those four groups of queries as " + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pp}}" + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{nn}}" + }, + { + "bbox": [ + 130, + 296, + 480, + 366 + ], + "type": "text", + "content": ". This proposed query grouping can be illustrated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 377, + 480, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 377, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 242, + 377, + 480, + 389 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {\\mathrm {p p}} = \\left\\{q ^ {k} | k \\in \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\in \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {6}", + "image_path": "d2e37acb1dd64a50f68b128e8bb105cc31c1abac8d9d2c9b24cb5e75b7220dd9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 393, + 480, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 393, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 242, + 393, + 480, + 406 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {\\mathrm {p n}} = \\left\\{q ^ {k} | k \\in \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\notin \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {7}", + "image_path": "31cfa769a1bcd92c165a0bc17e8dc7e73884ebf9e37190778938d65563cde9c8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 411, + 480, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 411, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 242, + 411, + 480, + 422 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {\\mathrm {n p}} = \\left\\{q ^ {k} | k \\notin \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\in \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}, \\tag {8}", + "image_path": "f3f1f25f7512cc53c9876ae004ff61f7e336f82cdfb010400907061376fb07cd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 427, + 480, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 427, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 242, + 427, + 480, + 438 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {\\mathrm {n n}} = \\left\\{q ^ {k} \\mid k \\notin \\hat {\\sigma} _ {\\mathrm {d e t}}, k \\notin \\hat {\\sigma} _ {\\mathrm {t r k}} \\right\\}. \\tag {9}", + "image_path": "5942e8e255e127c7c084e6f0219beb20749500ab37a805bdee60217d7d34ef4f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": "Through this, the classification gradients of the two tasks can also be decomposed as the sum of detection or tracking classification gradients on each group of queries. The classification gradient conflict only lies in the second and third groups of queries, as illustrated in Fig. 2(a). Therefore we propose to remove the gradient of " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " contributed to the tracking classification loss and the gradient of " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " contributed to the detection classification loss, as illustrated in Fig. 2(b). Through this, the partial classification gradient conflict of the two tasks on " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " is avoided. " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pp}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{nn}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 556 + ], + "type": "text", + "content": " still contribute gradient in both heads since the partial classification gradients upon them are not contradictory." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 574, + 289, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 289, + 586 + ], + "type": "text", + "content": "3.3 Dynamic Attention Mask" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "Following previous end-to-end trackers [21, 30, 31, 45, 47, 48, 50], OneTrack generates predictions in a set-prediction fashion. Hence the model is encouraged to suppress redundant predictions regarding the same ground truth object through self-attention between object queries in each decoder layer. As previously mentioned, " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " originates from the same ground truth object being assigned to different object queries in the two tasks. To avoid unexpected competition" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 117, + 443, + 273 + ], + "blocks": [ + { + "bbox": [ + 169, + 117, + 443, + 273 + ], + "lines": [ + { + "bbox": [ + 169, + 117, + 443, + 273 + ], + "spans": [ + { + "bbox": [ + 169, + 117, + 443, + 273 + ], + "type": "image", + "image_path": "8bdaa410e6798843dd3da7acc3b3f2ffc02b4024ba17812da56df5e856fbcae1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "content": "Fig. 2: Backpropagation of Classification Gradients. Gradient coordination prevents the classification gradient conflict between detection and tracking. Best viewed in color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "content": "between task-specific queries regarding the same object, in each layer of the decoder, we dynamically mask the attention between " + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "content": ". Attention between all other query groups is not masked. Through this, " + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "content": " are not directly exposed to each other during the forward process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": "We utilize the ground truth assignment results in each decoder layer to generate the self-attention mask for the next layer. Before the first decoder layer where detection queries are initialized as empty queries, we use distances between their initial reference points and centers of ground truths to perform bipartite matching. After each decoder layer, we perform bipartite matching based on matching costs between ground truths and predicted 3D boxes defined in [37]. We employ the Hungarian algorithm [17] for bipartite matching. As for outputs, the detection results of OneTrack include predictions from " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pp}}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{nn}}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": ". The tracking results include predictions from " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pp}}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{nn}}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": ". We illustrate the pipeline of OneTrack in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "content": "During inference, we categorize object queries into four groups based on the detection and tracking confidence scores of their predictions. Queries with high detection or tracking confidence scores over " + }, + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "content": " are classified as positive samples for the respective tasks. We set " + }, + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "inline_equation", + "content": "T_{p} = 0.2" + }, + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": "At last, in both the training or inference process, we propagate the Top- " + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": " object queries with high confidence scores from both detection and tracking results to the next frame as " + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Q}}_t" + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": ". In particular, we will still propagate the \"locked\" track queries to the future frames even if they are not in the Top- " + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": " object queries. This is for addressing challenging tracking scenarios, such as an object being occluded for several frames. We propagate \"locked\" track queries through this for up to " + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{A_{age}} = 3" + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": " frames into future frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Following [31,36], we also preserve the propagated object queries as the memory queue for the detection or tracking process for up to 4 frames. The mem" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 477, + 257 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 477, + 257 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 477, + 257 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 477, + 257 + ], + "type": "image", + "image_path": "57a7132ff371c2f733db751d0b580166e2d013ab1032d9adada585c64d0924ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 338 + ], + "type": "text", + "content": "Fig. 3: Training pipeline of OneTrack. For each decoder layer in OneTrack, We identify object queries into four groups based on their positivity as training samples in detection and tracking tasks. Task-specific queries that are positive samples only for one task will only backpropagate gradients in the respective task-specific classification head to prevent conflicted classification gradients. We dynamically mask all attention across positive sample queries only for detection or tracking. Best viewed in color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "text", + "content": "ory queues for the detection and tracking processes are updated independently. Those memory queries are provided to the decoder for performing hybrid attention proposed by [36]. As keys and values for the hybrid attention layer, detection and tracking memory queries are masked for " + }, + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "text", + "content": ", respectively. We show the design of decoder layers in OneTrack in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "type": "text", + "content": "As demonstrated in Table 4, with the proposed dynamic hybrid-attention mask, the true joint training of detection and tracking can be achieved. This allows for training detection and tracking in a single training stage and inferring them in a single forward pass with a unified model. This feature of OneTrack significantly reduces its training cost and inference latency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 500, + 348, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 348, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 348, + 511 + ], + "type": "text", + "content": "3.4 Modified Tracking Classification Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": "In addition to introducing the true joint training of detection and tracking to reduce the detection capability degradation of trackers, we also adjusted the tracking classification loss to suppress the in fact inaccurate predictions. As previously mentioned, previous end-to-end trackers regard track queries whose predictions are associated with ground truths as positive samples equally, regardless of the precision of their current predictions. This encourages the tracker to assign high confidence to the in fact inaccurate predictions from track queries. Therefore we propose to restrain the positive assignment of those in fact inaccurate predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "In previous end-to-end trackers, the assignment result of a ground truth object " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " can be denoted as a one-hot vector: " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{g}^{n\\times 1} = [0,\\dots ,1,\\dots ,0]" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " is the number of object queries and the ground truth object is exclusively" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 183, + 114, + 424, + 280 + ], + "blocks": [ + { + "bbox": [ + 183, + 114, + 424, + 280 + ], + "lines": [ + { + "bbox": [ + 183, + 114, + 424, + 280 + ], + "spans": [ + { + "bbox": [ + 183, + 114, + 424, + 280 + ], + "type": "image", + "image_path": "ff3a3d84509c258c6433f8b8b2e37f0483d22552795a5b86c9e464f0eb238912.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "text", + "content": "Fig. 4: Design of decoder layers in OneTrack. In addition to current object queries, queries in detection memory and tracking memory are fed to the hybrid attention layer as keys and values. As keys and values, " + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "text", + "content": " and detection memory are masked to " + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "text", + "content": ". As keys and values, " + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{np}}" + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "text", + "content": " and tracking memory are masked to " + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{pn}}" + }, + { + "bbox": [ + 131, + 288, + 482, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "content": "assigned to the " + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "content": " object query. This assignment is unaffected by the matching cost between the " + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "content": " object query and " + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "content": " if they have been associated in recent frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 430 + ], + "type": "text", + "content": "In OneTrack, we propose to suppress the assignment weight between ground truths and queries that are locked but have high matching costs. We adjust the assignment result of locked ground truth " + }, + { + "bbox": [ + 130, + 392, + 482, + 430 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 392, + 482, + 430 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 208, + 437, + 481, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 437, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 208, + 437, + 481, + 453 + ], + "type": "interline_equation", + "content": "A _ {g} ^ {n \\times 1} = [ 0, \\dots , W _ {g i}, \\dots , 0 ], \\tag {10}", + "image_path": "decf5f81805f4e2c6fcb9d43099e48877959ce5a0d114f44291047afc5600500.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 454, + 481, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 454, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 214, + 454, + 481, + 479 + ], + "type": "interline_equation", + "content": "W _ {g i} = - \\frac {1}{\\gamma} * \\operatorname {M a x} \\left(\\left(\\operatorname {d i s t} (g, q ^ {i}) - 0. 5\\right), 0\\right) + 1, \\tag {11}", + "image_path": "daa722744495499cba16a237ae3dc7f685dbfd519d1781ba92880aa6783256ce.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(g, q^i)" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " represents the distance between the centers of ground truth box " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " and the predicted box of the locked query " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "q^i" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " is a hyper-parameter greater than zero. The assigned weight " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "W_{gi}" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "q^i" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " decreases linearly from 1.0 as " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(g, q^i)" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " increases from " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": ". The assignment weight serves as the weighting factor for calculating box regression loss and the targets for predicting tracking confidence score. Instead of Focal Loss [24], we employ the Quality Focal Loss proposed by [20] to supervise the tracking classification head, given the continuous " + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "inline_equation", + "content": "0\\sim 1" + }, + { + "bbox": [ + 130, + 487, + 482, + 584 + ], + "type": "text", + "content": " confidence label." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 603, + 230, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 230, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 230, + 617 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "In this section, we first introduce our detailed experimental setup. Then we provide comparisons with recent works on nuScenes camera-based 3D tracking benchmark. Ablation studies on each component are presented in the last part." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 158, + 462, + 356 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Compare with recent works on nuScenes dataset. " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": ": trained from scratch. OneTrack-F / OneTrack-S: OneTrack trained under full-resolution / half-resolution as described in Sec. 4.2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 158, + 462, + 356 + ], + "lines": [ + { + "bbox": [ + 149, + 158, + 462, + 356 + ], + "spans": [ + { + "bbox": [ + 149, + 158, + 462, + 356 + ], + "type": "table", + "html": "
Backbonee2eResolutionAMOTA↑AMOTP↓MOTA↑RECALL↑IDS↓
Validation Split
QD3DT [13]R101✓1600×9000.2421.5180.21839.9%5646
MUTR3D [48]R101✓1600×9000.2941.4980.26742.7%3822
CC-3DT [10]R101×1600×6400.4291.2570.38553.4%2219
PF-Track-S [31]V2-99✓800×3200.4081.3430.37650.7%166
PF-Track-F [31]V2-99✓1600×6400.4791.2270.43559.0%181
DQTrack [21]V2-99✓800×3200.4461.251--1193
OneTrack-S†V2-99✓800×3200.4921.1220.40958.8%315
OneTrack-F†V2-99✓1600×6400.5481.0880.47961.8%389
Test Split
QD3DT [13]R101✓1600×6400.2171.5500.19837.5%6856
MUTR3D [48]R101✓1600×6400.2701.4940.23541.1%6018
SRCN3D [32]V2-99×1600×6400.3981.3170.35953.8%3334
CC-3DT [10]R101×1600×6400.4101.2740.35753.8%3334
PF-Track [31]V2-99✓1600×6400.4341.2520.37853.8%249
DQTrack [21]V2-99✓1600×6400.5231.0960.44462.2%1204
OneTrack-F†V2-99✓1600×6400.5541.0210.46160.8%481
", + "image_path": "498c1d6fdb2d504113ffc7537b1b117f21f703b29bc043c6b20c5eccd92fce5f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 378, + 269, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 269, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 269, + 389 + ], + "type": "text", + "content": "4.1 Datasets and Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": "We evaluate our proposed method on the nuScenes dataset. nuScenes [5] dataset is a large-scale autonomous driving benchmark containing 1000 multi-modal videos. Videos recorded by six cameras are divided into 700, 150, and 100 scenes for training, validation, and testing, respectively. 3D box annotations of 10 object classes are provided for keyframes at " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": ". For nuScenes 3D tracking benchmark, we report AMOTA, AMOTP [39], MOTA [3], Recall and identity switches (IDS). For the 3D detection task, we report mean Average Precision (mAP) and nuScenes Detection Score (NDS)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 513, + 280, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 280, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 280, + 525 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "Our implementation is mainly based on StreamPETR [36]. We define two implementation settings: The full-resolution setting for state-of-the-art comparison and the half-resolution setting for ablation studies. We conduct experiments using V2-99 [18] as the image backbone. OneTrack is trained using AdamW [28] optimizer with a batch size of 8 and a base learning rate of 4e-4. The cosine annealing policy is employed to adjust the learning rate. We use 644 initialized detection queries in each frame and propagate the Top-256 object queries into the next frame. When training, we record all matched ground truth and object query pairs as \"locked\" pairs for label assignment. An object query matched with a ground truth will be recorded as a track query. As previously mentioned, a \"locked\" pair will be preserved in memory for up to " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{age}} = 3" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " frames, even if" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 136, + 480, + 232 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 427, + 126 + ], + "type": "text", + "content": "Table 4: Ablation study on components breakdown." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 136, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 135, + 136, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 136, + 480, + 232 + ], + "type": "table", + "html": "
Training taskInitialized from pre-trained detectorGradient coordinationDynamic High-cost assign masksuppressAMOTA↑IDS↓NDS↑mAP↑
detection------0.5700.479
tracking×--×0.3051210.4550.313
tracking✓--×0.3754660.5030.409
joint××××0.4252460.5170.410
joint×✓××0.4412310.5550.462
joint×✓✓×0.4692610.5750.484
joint×✓✓✓0.4923150.5660.478
", + "image_path": "b6bdb9bdbf58bec8d2b192cba3914ada0319d10cee6b55b35908c321299f64b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "text", + "content": "the respective track query is not in the Top- " + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "text", + "content": " object queries in each frame. For track queries that failed to predict confident prediction over " + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\mathrm{age}}" + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "text", + "content": " frames, its \"locked\" ground truth will be assigned to other \"free\" object queries. During inference, we initialize an object query as a track query if it predicts a 3D box with confidence over " + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{conf}} = 0.4" + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "text", + "content": ". Additionally, all track queries are preserved and propagated across frames for up to " + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\mathrm{age}} = 3" + }, + { + "bbox": [ + 130, + 255, + 481, + 375 + ], + "type": "text", + "content": " frames. Predictions from the same track query propagated across frames will be assigned the same tracking ID. We output all detection predictions with confidence scores over 0.05 and all tracking predictions with confidence scores over 0.3. We introduce other setting-specific training configurations as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "text", + "content": "OneTrack-F: We crop input images from the initial resolution of " + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "inline_equation", + "content": "1600 \\times 900" + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "inline_equation", + "content": "1600 \\times 640" + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "text", + "content": ". We then conduct end-to-end training for 48 epochs. OneTrack-S: We first crop the input images to the resolution of " + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "inline_equation", + "content": "1600 \\times 640" + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "text", + "content": " and downsample them to " + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "inline_equation", + "content": "800 \\times 320" + }, + { + "bbox": [ + 130, + 375, + 481, + 424 + ], + "type": "text", + "content": ". Models are trained for 24 epochs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 441, + 338, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 338, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 338, + 453 + ], + "type": "text", + "content": "4.3 Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "text", + "content": "As shown in Table 3, OneTrack outperforms all previous trackers in both the validation and test set of nuScenes dataset by a significant margin. On the validation set, OneTrack surpasses previous state-of-the-art PF-Track [31] by " + }, + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "text", + "content": " AMOTA. On the test set, OneTrack surpasses DQTrack [21] by " + }, + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 545 + ], + "type": "text", + "content": " AMOTA. Notably, OneTrack reports fewer than half of the identity switches compared to those reported by DQTrack. It is also worth mentioning that OneTrack is the only model in Table 3 that is trained from scratch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 562, + 209, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 209, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 209, + 574 + ], + "type": "text", + "content": "4.4 Ablations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 604 + ], + "type": "text", + "content": "We conduct ablation studies on each proposed component with OneTrack-S on the nuScenes validation set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Components breakdown. The breakdown analysis of proposed components is presented in Table 4. Here \"det\" or \"trk\" in the first row stands for only training the model for detection or tracking tasks. And \"joint\" stands for jointly training the model for both detection and tracking tasks, with an extra classification head attached to the model. As shown in Table 4, the model trained solely" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 252, + 202 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 252, + 202 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 252, + 202 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 252, + 202 + ], + "type": "image", + "image_path": "05c7a253a57041435d97f5902819b770cbc10364af1d2b03952b707895c5a2ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "type": "text", + "content": "Fig. 5: Ablation on hyper-parameters. (a) On parameter " + }, + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "type": "text", + "content": " in the modified tracking losses. (b) On the max preserved frames of track queries " + }, + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{age}}" + }, + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "type": "text", + "content": ". (c) On the score threshold " + }, + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{conf}}" + }, + { + "bbox": [ + 130, + 210, + 482, + 243 + ], + "type": "text", + "content": " for track query initialization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 115, + 370, + 202 + ], + "blocks": [ + { + "bbox": [ + 255, + 115, + 370, + 202 + ], + "lines": [ + { + "bbox": [ + 255, + 115, + 370, + 202 + ], + "spans": [ + { + "bbox": [ + 255, + 115, + 370, + 202 + ], + "type": "image", + "image_path": "39a1941da2f64dc7cac4c2da3d61991fa2d78216e77ecf9e10764e5c189518b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 375, + 115, + 490, + 202 + ], + "blocks": [ + { + "bbox": [ + 375, + 115, + 490, + 202 + ], + "lines": [ + { + "bbox": [ + 375, + 115, + 490, + 202 + ], + "spans": [ + { + "bbox": [ + 375, + 115, + 490, + 202 + ], + "type": "image", + "image_path": "3aaa808cd8f770f592e01716dfb9d8db137d4c291894b48f0dcc006ff9519da5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 411 + ], + "type": "text", + "content": "for the tracking task suffers from severe degradation in detection performance, which results in their inferior detection performance compared to the detector. Initializing the tracker from a detector checkpoint or trivially supervising the model with both detection and tracking losses can only slightly alleviate this problem. Our proposed components including classification gradient coordination, dynamic attention mask, and high-cost tracking assignment suppress, are all validated as highly effective in achieving the true joint training of detection and tracking tasks. It is noteworthy that OneTrack without modifying its tracking loss achieves an even slightly better detection performance than the model trained solely as a detector. This observation suggests the potential for further enhancing the detection process through the integrated tracking process in a unified model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "text", + "content": "Modified tracking loss. In Fig. 5(a), we conduct ablation studies on the hyper-parameter " + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "text", + "content": " in the modified tracking loss for suppressing positive assignments of inaccurate predictions. OneTrack achieves optimal tracking performance when " + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "text", + "content": " is set to 7.5. While the detection performance of OneTrack shows improvement as " + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "text", + "content": " increases, this gain becomes limited when " + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 412, + 482, + 472 + ], + "type": "text", + "content": " exceeds 7.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 533 + ], + "type": "text", + "content": "Track age. We conduct ablation studies on the maximum preservation age " + }, + { + "bbox": [ + 130, + 473, + 482, + 533 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{age}}" + }, + { + "bbox": [ + 130, + 473, + 482, + 533 + ], + "type": "text", + "content": " of track queries. As illustrated in Fig. 5(b), a longer track age will lead to more identity switches and decreased detection performance. We speculate that this is because propagating too many outdated track queries will mislead the object feature extraction process in the current frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": "Confidence threshold for track query initialization. We assess the impact of " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{conf}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " in Fig. 5(c). The AMOTA performance will slightly increase as " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{conf}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " increases, which results from the more accurate tracking predictions on average. However, a higher " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{conf}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " reduces the initialization of track queries from detection queries with low confidence. This limitation hinders tracking objects under poor observation conditions, resulting in a higher number of identity switch cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Separated classification heads. As shown in Table 5, when its two classification heads share weights except for the final linear layer, the tracking and detection performance of OneTrack is slightly worse than when the classification heads are independent. We speculate that this is because the separation of classification heads leads to the easier convergence of both heads." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 135, + 418, + 182 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 467, + 125 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 467, + 125 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 467, + 125 + ], + "type": "text", + "content": "Table 5: Ablation on the separated classification heads for two tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 135, + 418, + 182 + ], + "lines": [ + { + "bbox": [ + 192, + 135, + 418, + 182 + ], + "spans": [ + { + "bbox": [ + 192, + 135, + 418, + 182 + ], + "type": "table", + "html": "
Cls headsAMOTA↑IDS↓NDS↑mAP↑
shared49.03760.5590.473
separated49.23150.5660.478
", + "image_path": "b64a2bcd6e184a4d794ee78e68e9b8794e9fb929a74bbcd8a0c39be065be900c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 189, + 245, + 422, + 315 + ], + "blocks": [ + { + "bbox": [ + 130, + 191, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 236 + ], + "type": "text", + "content": "Table 6: Computational cost comparison with recent works. All training time costs and inference latencies are measured on a single A100 GPU. *: PF-Track is first trained as a detector and then as a tracker for 12 epochs each. We train DQTrack and OneTrack from scratch for 24 epochs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 245, + 422, + 315 + ], + "lines": [ + { + "bbox": [ + 189, + 245, + 422, + 315 + ], + "spans": [ + { + "bbox": [ + 189, + 245, + 422, + 315 + ], + "type": "table", + "html": "
MethodBackboneResolutionTraining time costInference latency
PF-TrackV2-99800×32034h+82h*124.4ms
DQTrackV2-99800×320176h115.2ms
OneTrackV2-99800×32087h83.7ms
", + "image_path": "c086ce47ead7e5bfe2df88efc5eb5e9d2861e7e60a289b1ffa58f6826c56ff08.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 455 + ], + "type": "text", + "content": "Computational cost. To ensure fair comparison, we apply gradient checkpointing [7] on the decoder layers of all trackers. We measure their training cost and inference latency on a single A100 GPU. As shown in Table 6, the training cost of OneTrack is only half that of previous state-of-the-art trackers. The inference latency of OneTrack is also significantly lower. We attribute the low computational cost of OneTrack to its elegant model design. We do not extend the tracker with various plug-in modules [31] or separate the detection and tracking stages of the tracker [21], which will introduce additional computational burdens. Notably, OneTrack is trained from scratch while PF-Track and DQTrack rely on extra training stages for achieving high tracking performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 471, + 220, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 220, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 220, + 483 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "type": "text", + "content": "In this work, we delve into the optimization conflict between detection and tracking in end-to-end trackers, which arises from the partially conflicted classification gradients between two tasks. Based on this observation, we propose to identify the conflicted positive samples between two tasks through query grouping and solve the conflicted gradients with the proposed gradient coordination, dynamic attention mask, and suppression tracking positive samples of low quality. On the nuScenes dataset, OneTrack outperforms all previous end-to-end trackers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 241, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 241, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 241, + 608 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Ali, A., Jalil, A., Niu, J., Zhao, X., Rathore, S., Ahmed, J., Aksam Iftikhar, M.: Visual object tracking—classical and contemporary approaches. Frontiers of Computer Science 10, 167–188 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "text", + "content": "2. Benbarka, N., Schröder, J., Zell, A.: Score refinement for confidence-based 3d multi-object tracking. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8083-8090. IEEE (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "type": "text", + "content": "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing 2008, 1-10 (2008)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "text", + "content": "4. Bewley, A., Ge, Z., Ott, L., Ramos, F., Upcroft, B.: Simple online and realtime tracking. In: 2016 IEEE international conference on image processing (ICIP). pp. 3464-3468. IEEE (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "text", + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "6. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "text", + "content": "7. Chen, T., Xu, B., Zhang, C., Guestrin, C.: Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 386 + ], + "type": "text", + "content": "8. Chiu, H.k., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: ICRA (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "text", + "content": "9. Chu, P., Wang, J., You, Q., Ling, H., Liu, Z.: Transmot: Spatial-temporal graph transformer for multiple object tracking. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4870-4880 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "text", + "content": "0. Fischer, T., Yang, Y., Kumar, S., Sun, M., Yu, F.: CC-3DT: panoramic 3d object tracking via cross-camera fusion. In: CoRL. Proceedings of Machine Learning Research, vol. 205, pp. 2294–2305. PMLR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 454, + 480, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 476 + ], + "type": "text", + "content": "1. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science 18(5), 185706 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 477, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 480, + 520 + ], + "type": "text", + "content": "2. He, J., Huang, Z., Wang, N., Zhang, Z.: Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5299-5309 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "text", + "content": "3. Hu, H.N., Yang, Y.H., Fischer, T., Darrell, T., Yu, F., Sun, M.: Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 554, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 576 + ], + "type": "text", + "content": "4. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "text", + "content": "5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "6. Kim, A., Osep, A., Leal-Taixe, L.: Eagermot: 3d multi-object tracking via sensor fusion. In: 2021 IEEE International conference on Robotics and Automation (ICRA). pp. 11315-11321 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "text", + "content": "18. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "text", + "content": "19. Li, J., Ding, Y., Wei, H.L., Zhang, Y., Lin, W.: Simpletrack: Rethinking and improving the jde approach for multi-object tracking. Sensors 22(15), 5863 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "text", + "content": "20. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems 33, 21002-21012 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "text", + "content": "21. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18302-18311 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "type": "text", + "content": "22. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 481, + 332 + ], + "type": "text", + "content": "23. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "text", + "content": "24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "text", + "content": "25. Liu, C., Chen, X.F., Bo, C.J., Wang, D.: Long-term visual tracking: review and experimental comparison. Machine Intelligence Research 19(6), 512-530 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "26. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "text", + "content": "27. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3262-3272 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 451, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 481, + 472 + ], + "type": "text", + "content": "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "29. Marinello, N., Proesmans, M., Van Gool, L.: Tripletrack: 3d object tracking using triplet embeddings and lstm. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4499-4509 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "30. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multi-object tracking with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8844-8854 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "type": "text", + "content": "31. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17928-17938 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "type": "text", + "content": "32. Shi, Y., Shen, J., Sun, Y., Wang, Y., Li, J., Sun, S., Jiang, K., Yang, D.: Srcn3d: Sparse r-cnn 3d surround-view camera object detection and tracking for autonomous driving. arXiv preprint arXiv:2206.14451 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 612, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 643 + ], + "type": "text", + "content": "33. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10860-10869 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "34. Tokmakov, P., Li, J., Burgard, W., Gaidon, A.: Learning to track with object permanence. In: ICCV. pp. 10840-10849. IEEE (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Q. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 643 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "35. Wang, Q., Chen, Y., Pang, Z., Wang, N., Zhang, Z.: Immortal tracker: Tracklet never dies. arXiv preprint arXiv:2111.13672 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "text", + "content": "36. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3621-3631 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "text", + "content": "38. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5096-5105 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "type": "text", + "content": "39. Weng, X., Kitani, K.: A baseline for 3d multi-object tracking. arXiv preprint arXiv:1907.03961 1(2), 6 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 480, + 281 + ], + "type": "text", + "content": "40. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. IROS (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 282, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 480, + 313 + ], + "type": "text", + "content": "41. Weng, X., Wang, Y., Man, Y., Kitani, K.: Gnn3dmot: Graph neural network for 3d multi-object tracking with multi-feature learning. arXiv preprint arXiv:2006.07327 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 480, + 346 + ], + "type": "text", + "content": "42. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: 2017 IEEE international conference on image processing (ICIP). pp. 3645-3649. IEEE (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 348, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 348, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 348, + 480, + 391 + ], + "type": "text", + "content": "43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "text", + "content": "44. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "type": "text", + "content": "45. Yu, E., Wang, T., Li, Z., Zhang, Y., Zhang, X., Tao, W.: Motrv3: Releasefetch supervision for end-to-end multi-object tracking. arXiv preprint arXiv:2305.14298 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "type": "text", + "content": "46. Zaech, J.N., Liniger, A., Dai, D., Danelljan, M., Van Gool, L.: Learnable online graph representations for 3d multi-object tracking. IEEE Robotics and Automation Letters 7(2), 5103-5110 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "text", + "content": "47. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision. pp. 659-675. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 480, + 544 + ], + "type": "text", + "content": "48. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4537-4546 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 480, + 577 + ], + "type": "text", + "content": "49. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision 129, 3069-3087 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "type": "text", + "content": "50. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22056-22065 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "51. Zhao, H., Zhang, J., Chen, Z., Yuan, B., Tao, D.: On robust cross-view consistency in self-supervised monocular depth estimation. Machine Intelligence Research 21(3), 495-513 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "OneTrack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_content_list.json b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a8e34010d3fcd7f9098fc767580f26db0857b64 --- /dev/null +++ b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_content_list.json @@ -0,0 +1,1863 @@ +[ + { + "type": "text", + "text": "OneVOS: Unifying Video Object Segmentation with All-in-One Transformer Framework", + "text_level": 1, + "bbox": [ + 235, + 140, + 769, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wanyun Li $^{1,*}$ , Pinxue Guo $^{2,*}$ , Xinyu Zhou $^{1}$ , Lingyi Hong $^{1}$ , Yangji He $^{1}$ , Xiangyu Zheng $^{1}$ , Wei Zhang $^{1,\\dagger}$ , and Wenqiang Zhang $^{1,3,\\dagger}$", + "bbox": [ + 243, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai, China", + "bbox": [ + 285, + 253, + 712, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai Engineering Research Center of AI & Robotics, Academy for Engineering & Technology, Fudan University, Shanghai, China \n $^{3}$ Engineering Research Center of AI & Robotics, Ministry of Education, Academy for Engineering & Technology, Fudan University, Shanghai, China {wyli22,pxguo21}@m.fudan.edu.cn, {weizh, wqzhang}@fudan.edu.cn", + "bbox": [ + 248, + 282, + 753, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Contemporary Video Object Segmentation (VOS) approaches typically consist of feature extraction, matching, memory management, and multiple objects aggregation. Recent advanced models either employ a discrete modeling for these components in a sequential manner, or optimize a combined pipeline through substructure aggregation. However, these existing explicit staged approaches prevent the VOS framework from being optimized as a unified whole, leading to the limited capacity and suboptimal performance in tackling complex videos. In this paper, we propose OneVOS, a novel framework that unifies the core components of VOS with All-in-One Transformer. Specifically, to unify all aforementioned modules into a vision transformer, we model all the features of frames, masks and memory for multiple objects as transformer tokens, and integrally accomplish feature extraction, matching and memory management of multiple objects through the flexible attention mechanism. Furthermore, a Unidirectional Hybrid Attention is proposed through a double decoupling of the original attention operation, to rectify semantic errors and ambiguities of stored tokens in OneVOS framework. Finally, to alleviate the storage burden and expedite inference, we propose the Dynamic Token Selector, which unveils the working mechanism of OneVOS and naturally leads to a more efficient version of OneVOS. Extensive experiments demonstrate the superiority of OneVOS, achieving state-of-the-art performance across 7 datasets, particularly exceeding in complex LVOS and MOSE datasets with $70.1\\%$ and $66.4\\%$ $J\\& F$ scores, surpassing previous state-of-the-art methods by $4.2\\%$ and $7.0\\%$ , respectively. Code is available at: https://github.com/L599wy/OneVOS.", + "bbox": [ + 261, + 378, + 743, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 758, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised Video Object Segmentation (VOS) [2, 22, 30, 37, 44, 45, 63, 64] is a fundamental task in video analysis [18, 19], aiming to consistently seg-", + "bbox": [ + 212, + 786, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding authors.", + "bbox": [ + 230, + 825, + 553, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3746337afe18d228069598e659319365bafa067089b36f3435bd0f27c82862e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 145, + 496, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2598b97b547d3b20c53e025f402918a65e127eeb22d20b3607133be2fac464b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 759, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f75c54cb4d79e617056685e3ed2911dacd00b82dc7bb8a84818ce82d4b161a7.jpg", + "image_caption": [ + "(b) Unify Matching and Objects Aggregation" + ], + "image_footnote": [], + "bbox": [ + 248, + 210, + 501, + 286 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cf56c97d1d60892b6b2d8633288ffba70e01d8090a44b0ab8376c379b3239f43.jpg", + "image_caption": [ + "(d) Integrate Core Modules Al in One Transformer", + "Fig. 1: Four different modeling for semi-supervised video object segmentation pipelines. (a) Traditional discrete modeling of VOS. (b) Unified multiple object modeling in one reference. (c) Unified single-object feature extraction and matching. (d) OneVOS: Integrates all core modules with All-in-One Transformer." + ], + "image_footnote": [], + "bbox": [ + 509, + 209, + 759, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ment objects throughout a video based on initial masks. Advancements in VOS primarily focus on the full utilization of available information. Among them, Memory-based methods [7,9,40,45] have emerged as the mainstream due to the effectiveness in storing and leveraging abundant spatio-temporal features. Its foundational paradigm, as depicted in Fig. 1(a) [16,40,51,55], encompasses four split key stages including feature extraction, matching, memory management, and multi-object aggregation, which have been widely adopted as a baseline in a number of methods [23,28,50,56].", + "bbox": [ + 212, + 371, + 784, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, some investigations demonstrate the advantages of integration modeling compared with highly discrete modeling. AOT [66] and DEAOT [68] (Fig. 1(b)) emphasize the shortage of single-object inference compared to joint modeling of multiple objects, particularly in exploring contextual and robust feature representations. Meanwhile, SimVOS [59] (Fig. 1(c)) reveals limitations in the separation of feature extraction and matching processes for capturing dynamic object-aware features. However, existing methods solely focus on combining or aggregating substructures of the VOS pipeline, resulting in their suboptimal solutions. For example, AOT relies on sophisticated designing matching modules which fails to leverage superior pre-training weights, while the simplistic single-object framework of SimVOS encounters inefficiencies and the inapplicability of memory makes it unable to utilize rich features of the processed frames.", + "bbox": [ + 212, + 492, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we introduce an essentially novel framework distinct from previous methods, aiming to develop a unified end-to-end differentiable VOS model for global optimization. Based on this, the All-in-One Transformer for Video Object Segmentation (OneVOS) is proposed, which models and unifies all important modules into a Vision Transformer. Specifically, we first leverage the flexible attention mechanism to conduct feature extraction and mutual matching between the reference and current frames. To effectively capture multi-object relationships, we introduce a mask embedding layer that enriches features with multi-object modeling insights. Moreover, in the memory design, we propose an external layer-wise token memory to enhance the coherence understanding of videos by storing and providing spatio-temporal features for subsequent frames.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the feature extraction and matching of multiple objects performing simultaneously, the reference tokens output from each layer of Transformer for memory update could be represented by a weighted linear combination of tokens from memory, reference frame and current frame, which may lead to a significant semantic errors and ambiguity to serving as features stored in memory. To address this, we further propose a Unidirectional Hybrid Attention (Uni-Hybrid Attention) through a double decoupling of the original attention operation. It guarantees the versatility of stored features and enhances memory performance. Notably, as the Uni-Hybrid Attention is derived from naive attention mechanism, our framework is still able to utilize pre-trained weights from various Vision Transformers, thereby enhancing its generality and flexibility.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this way, the proposed OneVOS framework implements end-to-end modeling and global optimization, exhibiting superior performance compared to previous models. To further strengthen its advantages, we conduct an in-depth exploration of the behavior within the attention layer of OneVOS to enhance its efficiency, which is particularly crucial in resource-constrained and long-term video tasks. Concretely, we introduce the Dynamic Token Selector (DTS), capable of dynamically selecting partial tokens from each layer for cross-attention and memory storage, trained solely under segmentation loss without additional constraints. With DTS integrated into every layer of the transformer, we analyze the token selection distribution of DTS and the attention weight distribution, uncovering a significant insight: unlike the traditional sequential paradigm, OneVOS adopts a cross-layer segmentation mode that dynamically shifts focus between feature extraction and matching across different layers. Based on this, we propose a natural acceleration scheme involving designing distinct memory capacities for each layer and layer-wise memory updates via tokens selected by DTS. This approach enhances both speed and accuracy across multiple datasets.", + "bbox": [ + 212, + 313, + 787, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conducted comprehensive experiments across seven datasets, and OneVOS consistently achieves state-of-the-art performance across all of them. Notably, OneVOS demonstrates a substantial performance advantage enven in more challenging scenarios, achieving $71.1\\%$ $\\mathcal{I}\\& \\mathcal{F}$ on LVOS validation and $67.2\\%$ $\\mathcal{I}\\& \\mathcal{F}$ on MOSE validation. The main contributions are summarized as follows:", + "bbox": [ + 212, + 555, + 789, + 630 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose All-in-One Transformer framework for Video Object Segmentation (OneVOS), integrating feature extraction, matching, memory management, and objects aggregation into a vision transformer, which can be optimized globally. This provides a fundamentally novel paradigm in VOS.", + "- We introduce the Unidirectional Hybrid Attention mechanism to facilitate multi-object feature extraction and matching, ensuring tokens are devoid of semantic ambiguity for efficient memory storage.", + "- We propose the Dynamic Tokens Selector, which not only unveils the internal working mechanism but also leads to a more efficient OneVOS, based on the exploration and the resulting insight.", + "- OneVOS achieves state-of-the-art performance across 7 datasets, particularly excelling in more complex LVOS and MOSE datasets with $70.1\\%$ and $66.4\\%$ $J\\& F$ , surpassing previous best methods by $4.2\\%$ and $7.0\\%$ , respectively." + ], + "bbox": [ + 225, + 642, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 366, + 114, + 732, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Video object segmentation", + "text_level": 1, + "bbox": [ + 215, + 181, + 482, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-supervised VOS methods can be divided into four types: The earlier fine-tuning methods [1, 2, 42, 54, 60] learn general segmentation features via offline pre-training, then enhance object-specific representations by online fine-tuning. Then, propagation-based methods [10, 29, 33, 44, 46, 57] utilize optical flow to propagate the mask from the previous frame to the next one. Subsequent methods optimized it by keyframes enhancement [34], advanced tracking techniques [4, 10, 31, 62], and re-identification mechanisms [38]. Matching-based methods [6, 30, 53, 65, 67] identify the object of the current frame based on the correlation with the reference frame in the embedding space.", + "bbox": [ + 212, + 210, + 785, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, Memory-based approaches [23,28,45,50] store features of segmented frames in a memory bank to provide a set of features for segmentation, which has become the dominant method in SVOS. Following that, methods such as HMMN [51], SSTVOS [16], and XMem [7] enhance performance by utilizing multi-scale features, introducing Transformer for matching, and building long-term memory. These methods typically adhere to a traditional discrete modeling and single-object architecture. More recently, the multi-object framework [66], [68] is proposed to address the inherent limitations of the previous single-object while still maintaining a sophisticated matching module for high performance. [59] presents a simplified VOS framework for joint feature extraction and matching which adopts single-object modeling and discards the vital memory design in VOS. Differently, our OneVOS, innovatively unifies feature extraction, matching, memory construction, and multi-objective aggregation all in one Transformer. This design not only streamlines the segmentation process but also enhances the dynamic interaction among different stages of SVOS.", + "bbox": [ + 212, + 347, + 787, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Visual Object Tracking", + "text_level": 1, + "bbox": [ + 215, + 599, + 454, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video object tracking (VOT) [3,27,35,70] shares similarities with the task setting of SVOS, but it primarily utilizes bounding boxes for object localization instead of detailed segmentation masks. In VOT, there has been a paradigm shift from traditional Siamese-based methods towards leveraging Transformer layers for enriched relational modeling between the template and search regions. Pioneering works like TransT [5] introduce a two-stream attention mechanism for robust bi-directional information exchange, while innovative methods like MixFormer [12] and AiATrack [21] adopt asymmetric attention strategies to minimize distractions from irrelevant objects. Notably, OSTrack [69] proposes a unified attention model, promoting seamless feature integration and relation modeling. Nevertheless, the distinct multi-object settings and difference in target information representation, coupled with the critical role of memory in VOS, render unification within the VOS field more challenging. This paper endeavors to tackle this challenge.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the Fig. 2, in a given video sequence, VOS focuses on segmenting the current $t$ -th frame $f_{t} \\in \\mathbb{R}^{H \\times W \\times 3}$ using the reference frame $f_{t-1} \\in \\mathbb{R}^{H \\times W \\times 3}$ with its mask and the history features stored in Memory. In this section, we aim to construct an unified framework at a global scale to tackle the challenges of VOS. We commence by introducing the OneVOS framework in Sec. 3.1, which employs Vision Transformer to integrate core elements in VOS. And then Sec. 3.2 provides a detailed exposition on the novel Unidirectional Hybrid Attention mechanism. In addition, in Sec. 3.3, we explore the working mechanism of attention within OneVOS through the proposed Dynamic Token Selector in order to introduce the efficient version of OneVOS.", + "bbox": [ + 212, + 176, + 787, + 325 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a9da2db1d58dedbded6f8e879caf3feeb187f20792ccb416c101db48fed8b752.jpg", + "image_caption": [ + "(a) Hybrid Attention Block" + ], + "image_footnote": [], + "bbox": [ + 250, + 330, + 472, + 441 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dbe76f2aa51d65433abdb0e0adec0aaaff1b6c919531c8ef84a566e1c1b9b0e3.jpg", + "image_caption": [ + "(b) Block with Dynamic Token Selector (DTS)" + ], + "image_footnote": [], + "bbox": [ + 475, + 330, + 751, + 436 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35e1eab2a4cb32f940bfe4f581ec8d0c98b7ed15375bf18eb82fb696bdef8076.jpg", + "image_caption": [ + "(c) Overview", + "Fig. 2: (a) Hybrid Attention Block. (b) Blocks with Dynamic Token Selector. (c) Overview of OneVOS framework equipped with All-in-One Transformer. The reference mask is initially encoded into mask embeddings. Patch embeddings are generated from both the reference and current frames, which are then processed by the Transformer. Within each Transformer layer, mask embeddings are integrated with reference tokens. Then they are selected by Dynamic Token Selector and concatenated with tokens stored in memory and current frame fed into the Hybrid Attention Module. Finally, the features of the current frame output from Transformer are subsequently decoded into multi-object masks." + ], + "image_footnote": [], + "bbox": [ + 251, + 458, + 751, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 OneVOS Framework", + "text_level": 1, + "bbox": [ + 215, + 768, + 429, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask Embedding. To capture complex target interactions and minimize redundant single-target inferences, motivated by [66], we utilize the Mask Embedding Layer (as Patch Embedding Layer in Transformer) to encode reference", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 364, + 113, + 732, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mask into Mask Embedding denoted as $m_{ref} \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P} \\times C}$ through a single convolutional layer. Here, $C$ represents the embedding dimension, with $P$ as the stride of the mask embedding. Utilizing the local relationship modeling capability of the convolution operation, the discriminability of multiple targets is enhanced, which will then be integrated with visual feature map in Transformer before attention computation to integrating mask embeddings to the immediate value tokens of each layer, thereby effectively promoting the model to perform multiple targets inference.", + "bbox": [ + 212, + 143, + 787, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All-in-One Transformer. Then, we adopt a transformer architecture to unify the core modules of VOS. It first generates the non-overlapping patch embeddings $e_t^0 \\in \\mathbb{R}^{N \\times C}$ and $e_{ref}^0 \\in \\mathbb{R}^{N \\times C}$ for the current and reference frames, where $P$ is the patch resolution same as the mask embedding, thus $N = HW / P^2$ . Then, the concatenation of $[e_{ref}^0; e_t^0]$ along the spatial dimension serves as the initial input for the backbone, facilitating adaptive multi-object feature extraction and target information propagation through the stacked layers. In each layer, we first integrate mask embedding with visual feature map as shown in Eq 1:", + "bbox": [ + 214, + 272, + 787, + 396 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} \\right\\} = \\left\\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} + m _ {r e f} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 401, + 785, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Q$ , $K$ , and $V$ are query, key, and value matrices aligning with the standard attention mechanism [15], and $l \\in \\{0, 1, \\dots, L\\}$ represents the layer. Then, these object-aware reference tokens, memory tokens, and tokens of current frames undergo adaptive feature extraction and target information propagation by applying the attention mechanism of the Transformer. The output $a^l = [a_{ref}^l; a_t^l]$ of layer $l$ after the attention operation is as follows:", + "bbox": [ + 214, + 425, + 787, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\na ^ {l} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot V = \\operatorname {S o f t m a x} \\left(\\frac {\\left[ Q _ {r e f} ^ {l} ; Q _ {t} ^ {l} \\right] \\left[ K _ {M} ^ {l} ; K _ {r e f} ^ {l} ; K _ {t} ^ {l} \\right]}{\\sqrt {d _ {k}}}\\right) \\cdot \\left[ V _ {M} ^ {l}; V _ {r e f} ^ {l}; V _ {t} ^ {l} \\right], \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 532, + 785, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $K_M^l, V_M^l$ are features stored in memory. Supposing $W = QK^{T}$ is the matrix of similarity weights, $a_{t}^{l}$ can be written as:", + "bbox": [ + 214, + 570, + 787, + 601 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\na _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 607, + 784, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $[W_{t,ref}V_{ref}]$ and $[W_{t,M}V_M]$ are responsible for aggregating the inter-frame features from memory and reference frame to current frame (matching), and $[W_{t,t}V_t]$ aggregating the intra-frame features (feature extraction) for current frame, respectively. Thus, dynamic coordination between feature extraction and matching is achieved through the attention weights learned adaptively.", + "bbox": [ + 214, + 627, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Token Memory. In addition, since the target information and the reference frame are integrated within our framework, we can store intermediate features directly into Token Memory without necessitating additional post-processing fusion steps. Given the semantic variance of intermediate tokens in the reference frame across layers, we introduce a token storage mechanism to save essential Transformer tokens instead of whole frames as follows:", + "bbox": [ + 214, + 702, + 787, + 785 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM = \\left\\{K _ {M} ^ {l}, V _ {M} ^ {l} \\mid \\in \\{0, 1, \\dots L \\} \\right\\} = \\left\\{K _ {\\text {r e f s}} ^ {l}, V _ {\\text {r e f s}} ^ {l} \\mid \\in \\{0, 1, \\dots L \\} \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 789, + 785, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This method allows these tokens to be involved only in the attention modules instead of multiple inferences, greatly reducing inference overhead.", + "bbox": [ + 214, + 811, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/395d5876de5b3d53d08a5979d97bf55681b9cc3befe6001eb255259c4857b0f1.jpg", + "image_caption": [ + "Fig. 3: (a) Visualization of attention distribution of target areas in OneVOS, showing the alternation between self-attention and cross-attention across layers. (b) Analysis of the maximal attention weights reveals a pattern of alternation between feature extraction via self-attention (notably in layers 3 and 5) and matching via cross-attention (as in layers 2 and 8). (c) The proportional analysis of the attention weights' sum, normalized to one, underscores the non-sequential prioritization across different layers between self-attention and cross-attention." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 496, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7af79e823c1599a2fa52ddfdb957dc04831c27f423e3b0a230ac5a4e07a85ec2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 146, + 640, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3284748381c29e9423cff0ade4ca7de693bca5519c9ca89200d9015b3645517.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 148, + 784, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask Decoding. Finally, the output of All-in-One Transformer, $e_t^L$ is input to the lightweight Feature Pyramid Network (FPN) for Mask Decoding. These features are then gradually upgraded to the original scale and decoded into multi-object masks.", + "bbox": [ + 214, + 340, + 784, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Unidirectional Hybrid Attention", + "text_level": 1, + "bbox": [ + 215, + 406, + 529, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Building upon our newly developed pipeline, all modules are integrated into a transformer as in Fig. 2(c). However, obtaining the stored features by employing original attention mechanism may lead to semantic errors and ambiguity since they are represented incorrectly. To address this challenge, our proposed Unidirectional Hybrid Attention (Uni-Hybrid Attention) is strategically designed through dual-level attention decoupling as shown in Fig. 2(a).", + "bbox": [ + 212, + 431, + 784, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Firstly, reference features calculated by naive attention operation to all the tokens are denoted as $a_{ref}^{l} = W_{ref,M}V_{M} + W_{ref,ref}V_{ref} + W_{refs,t}V_{t}$ . A key limitation in the computations arises where tokens of reference frame containing target information (as in Eq. (1)) are represented by tokens of current frame that without target information, causing semantic errors in the features stored in the memory which are essential for storing spatio-temporal target information for subsequent frames. Therefore, to alleviate this situation, our first modification decoupling attention from the current token to the reference token. The revised expression is:", + "bbox": [ + 212, + 513, + 785, + 626 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na _ {r e f} ^ {l} = W _ {r e f, M} V _ {M} + W _ {r e f, r e f} V _ {r e f}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 635, + 784, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subsequently, the attention computation between reference features and memory tokens is decoupled, serving a dual purpose: 1) Avoiding Information Overlap and Ambiguity: Memory tokens, enriched with mask embeddings, encapsulate target information from prior frames. Integrating these embeddings with those of the reference frame risks overlapping and confusing object cues. 2) Preserving Temporal Specificity: Every video frame contains essential temporal intricacies crucial for segmentation precision. Attention decoupling preserves the reference frame's temporal uniqueness, critical for accurate and context-aware segmentation. Therefore, we decouple attention between memory tokens and those from the reference frame, and the final unidirectional hybrid attention formulations are:", + "bbox": [ + 212, + 661, + 787, + 797 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na _ {r e f} ^ {l} = W _ {r e f, r e f} V _ {r e f} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 805, + 785, + 830 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 827, + 624, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 366, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a6fbaeba53cff3baf1fb75b5821301ee718e1254ca7a8df13b2b9954fd8b878.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 410, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc9e9bbfc14e091f1395a056962e2a9c36b2365f7d8b6ec68b98362c77ac418c.jpg", + "image_caption": [ + "Fig. 4: (a) Visualization of the selected tokens of reference frame by DTS in different Transformer layers. (b) Adaptive selection ratios learned by DTS in different layers without any constraints." + ], + "image_footnote": [], + "bbox": [ + 413, + 143, + 602, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5b83b667266c087cdaca6abda1d6bfc44ea014f069fcbf754d72e26c6201b4a.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 612, + 152, + 785, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Towards Efficient OneVOS", + "text_level": 1, + "bbox": [ + 214, + 306, + 480, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Utilizing the advanced All-in-One Transformer equipped with Uni-Hybrid Attention, OneVOS achieves outstanding performance across various datasets. However, the integration process of these modules in OneVOS still lacks explainability, hindering the rational design of a more efficient version. Therefore, this section delves into its working mechanism to propose the efficient OneVOS.", + "bbox": [ + 212, + 330, + 782, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Exploration and Insight. Within the Transformer architecture, attention weights determine the influence of each token in the input sequence on the output which is useful for interpreting Transformer behavior. By examining the attention weights of tokens in the target area of the current frame across each layer, we compare their focus in reference frames or themselves. As presented in Fig. 3, qualitative results in 3(a) shown that a fluctuating focus between reference frames (cross-attention) and themselves (self-attention), Subsequently, Fig. 3 (b) and (c) provide quantitative evidence of a complex alternation pattern, which deviates from the sequential feature extraction and matching process of traditional discrete VOS modeling or uniform distribution in each layer. Instead, these findings lead us to a natural insight: OneVOS may learn a dynamic layer-wise integration strategy, preferentially executing multi-object feature extraction or matching in each layer according to the semantic context.", + "bbox": [ + 212, + 404, + 784, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic Token Selector. Based on above insights and motivated by [49], we introduce a Dynamic Token Selector (DTS) into the OneVOS architecture, as depicted in Fig. 2(b). This module adaptively selects a subset of tokens from the reference frame for cross-attention and storage. The selection process, as defined in Eq. 7, combines global context and local token information:", + "bbox": [ + 212, + 575, + 784, + 646 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p = \\operatorname {S o f t m a x} (\\operatorname {M L P} \\left(\\left[ \\operatorname {M a x P o o l} \\left(z _ {r e f} ^ {l}\\right); z _ {r e f} ^ {l}\\right]\\right)) \\\\ z _ {r e f} ^ {l} = \\operatorname {M L P} \\left(\\operatorname {N o r m} \\left(e _ {r e f} ^ {l}\\right)\\right) \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 654, + 784, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $z_{ref}^{l}$ denotes the local feature encoded of each token, while [MaxPool( $z_{ref}^{l}$ )] aggregates the overall context of the reference frame. $p \\in \\mathbb{R}^{N \\times 2}$ represents the probabilities associated with each token. Specifically, $p_{i,0}$ indicates the probability of dropping $i$ -th token while $p_{i,1}$ indicates for selection.", + "bbox": [ + 212, + 700, + 782, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This design aims to enable the network to autonomously determine the optimal token selection ratio for high-performance segmentation. Consequently, it is integrated into an All-in-One Transformer, as depicted in Fig. 2(a). Moreover, it is exclusively trained under segmentation loss without additional constraints. By observing the proportion of tokens selected by DTS in each transformer layer, as illustrated in Fig. 4(a) and Fig. 4(b), a crucial insight emerges. The model automatically allocates varying", + "bbox": [ + 212, + 756, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66175493c3afb3bd5e120b006f4ecb74974800cb63a048ace280568936aba7e6.jpg", + "image_caption": [ + "Fig. 5: Visualization dynamic memory updates of OneVOS utilizing DTS. The selected tokens in red boxes are derived output from DTS as Fig. 4(a)." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 787, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "selection ratios to different layers, and even assigns a selection ratio of zero or one to specific layers. This observation underscores the focus of DTS on certain layers for feature extraction (selecting a proportion of reference frame tokens as zero) and on specific layers for matching (selecting a proportion of reference frame tokens as one), presenting a characteristic of alternating focus across layers.", + "bbox": [ + 212, + 294, + 782, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Efficient Version of One VOS. Building upon the above observations, a straightforward acceleration method is naturally introduced. This includes layer-wise memory capacity and adaptive token update mechanism via DTS, as shown in Fig. 5. Specifically, we initialize different capacities for each layer of Token Memory, with the token capacity formula as follows:", + "bbox": [ + 212, + 369, + 784, + 438 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nc _ {l} = r _ {l} * N * C a p. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 439, + 784, + 454 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $r_l$ represents the layer-specific selection ratio determined by DTS during training, as illustrated in Fig. 4(b). Additionally, $Cap$ is a hyper-parameter denoting the frame capacity (typically set to 3 or 5). Consequently, tokens are incrementally added to the memory based on these selection ratios until reaching the capacity limit of each layer. Subsequently, the top-k tokens, determined by the current attention weights $W_{t,ref}$ , are retained, followed by the continuous addition of tokens to memory. Aligned with the segmentation mode of OneVOS, this method achieves improvements in accuracy and speed across multiple datasets.", + "bbox": [ + 212, + 458, + 787, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 593, + 366, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 628, + 455, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One-stream VOS uses ConvVIT-base [20] as the backbone, initialized with ImageNet-1k [13] pre-training weights. The feature resolution of the mask decoding is increased from 1/16 to $1/4$ and the channel dimension is reduced from 768 to 384.", + "bbox": [ + 212, + 652, + 784, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training details. Following the common settings of VOS [8,66], the training phase is divided into two stages: pre-training on synthetic video sequences generated from still image datasets [11,17,24,39,52], and main training on standard VOS dataset (DAVIS 17 and YouTube VOS in default, the model mark with \"*\" uses MOSE in main training as well). The loss function is a 0.5:0.5 combination of bootstrapped cross-entropy loss and soft Jaccard loss [43]. The learning rate of two stages is 2e-4 and 1e-4, respectively, utilizing $4 \\times 3090$ s GPUs. Our batchsize is 4, with a gradient accumulation of 2 and a sequence length of 5, running each training phase for 200,000 iterations. In the training of DTS, we employ the Gumbel-Softmax [32] on its output $p$ to generate a one-hot selection vector, indicating whether each token is selected (assigned a value of 1) or", + "bbox": [ + 212, + 700, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 366, + 114, + 732, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/70be2ebb49364ca343aa4975ec24b72e75f4ce09d303b9954cbc0186deb8b217.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Long-term VideosComplicated Videos
MethodLVOS VALLVOS TESTMethodMOSEJ&FJF
J&FJFJ&FJF
CFBI [65]50.045.055.144.840.549.0XMem [7]x56.452.160.6
AOT-L [66]61.359.453.654.148.759.5DropSeg [58]x52.248.056.4
STCN [9]45.841.150.545.841.650.0R50-AOT-L [66]x58.153.962.3
RDE [36]52.947.758.149.044.453.5R50-DeAOT-L [68]x58.954.563.3
XMem [7]50.045.554.449.545.253.7SwinB-AOT-L [66]x60.056.164.0
AOT-B [66]56.951.861.954.449.359.4OneVOSx62.358.166.6
DeAOT [68]63.757.669.858.152.963.4RDE [36]✓48.844.652.9
R50-AOT-L [66]63.858.169.661.056.165.9STCN [9]✓50.846.655.0
R50-DeAOT-L [68]65.959.971.958.053.262.8XMem [7]✓57.653.362.0
Swin-AOT-L [66]65.660.875.161.556.474.8DeAOT [68]✓59.455.163.8
OneVOS-D70.164.775.560.155.668.0OneVOS-D✓61.757.566.0
OneVOS67.462.072.963.859.570.0OneVOS✓66.462.270.6
OneVOS*71.166.076.264.160.068.2OneVOS*✓67.263.071.5
", + "bbox": [ + 220, + 143, + 782, + 315 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Quantitative comparison of long-duration video datasets (LVOS validation and test) and complex datasets, MOSE. $\\checkmark$ indicates MOSE used for training. OneVOS( $\\checkmark$ ) only employs MOSE in main training, whereas OneVOS\\* $(\\checkmark)$ utilizes DAVIS, YouTube, and MOSE.", + "bbox": [ + 215, + 325, + 787, + 383 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4e4bfc47311d1b17d3f11651efb96d034901e79d348c89b6fd74091428fb2ffb.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging short-term videos (MOSE and DAVIS 17)." + ], + "image_footnote": [], + "bbox": [ + 217, + 397, + 785, + 520 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "not (value of 0). Throughout training, we record the selection rate, observing that in the later stages of training, the selection rate of each layer gradually stabilizes.", + "bbox": [ + 215, + 565, + 784, + 594 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Inference details. OneVOS utilizes basic memory storage, storing tokens at intervals of 5 or 10 frames and updates using the first-in-first-out queue when it is full, except for the initial frame tokens. Conversely, OneVOS-D employs dynamic memory, defaulting to retain a maximum of 520 top-k tokens when full. The maximum frame-level capacity $Cap$ varies by dataset, set to $3/5/8$ . For fair comparison in LVOS, we strictly limit memory to under 6.", + "bbox": [ + 215, + 598, + 784, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets and Metrics. We evaluate the performance of OneVOS on 4 standard VOS datasets, including DAVIS 16 val [47], DAVIS 17 val [48], DAVIS 17 test [48], and YouTube 2019 [61], as well as 3 more challenging datasets, MOSE [14], LVOS val, and LVOS test [26]. The evaluation metrics include region similarity $\\mathcal{I}$ , boundary accuracy $\\mathcal{F}$ , and their average $\\mathcal{J} \\& \\mathcal{F}$ , and FPS for speed. For YouTube-VOS, we also report them from the seen $\\mathcal{J}_s$ , $\\mathcal{F}_s$ and unseen $\\mathcal{J}_u$ , $\\mathcal{F}_u$ categories.", + "bbox": [ + 215, + 686, + 785, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 215, + 789, + 614, + 806 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Complicated datasets. Table 1 tabulates the quantitative results on the LVOS val, LVOS test, and MOSE. OneVOS demonstrates superior performance across multi-", + "bbox": [ + 215, + 811, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d14b4e8a58aa6e2c7b1b221c8b77c677cd3a5a81a593307948d092e86d4e8b78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDAVIS 16 VALDAVIS 17 VALDAVIS 17 TESTYOUTUBE 19 VAL
J&FJFJ&FJFFPSJ&FJFGJsFsJuFu
CFBI+ [67]89.988.791.182.980.185.75.678.074.481.682.681.786.277.185.2
HMMN [51]90.889.692.084.781.987.55.078.674.782.582.581.786.177.385.0
DropSeg [58]92.190.993.386.583.589.5----83.482.987.377.785.6
STCN [9]91.690.892.585.482.288.620.276.172.779.682.781.185.478.285.9
XMem [7]91.590.492.786.282.989.522.681.077.484.585.584.388.680.388.6
Swin-AOTL [66]92.090.793.385.482.488.412.181.277.385.184.584.088.878.486.7
Swin-DeAOTL [68]92.991.194.786.283.189.215.482.878.986.786.185.390.280.488.6
ISVOS [56]92.691.593.787.183.790.5-82.879.386.286.185.289.780.788.9
MiVOS' [8]91.089.692.484.581.787.411.278.674.982.282.480.684.778.186.4
STCN' [9]91.790.493.085.382.088.620.277.874.381.384.282.687.079.487.7
XMem' [7]92.090.793.287.784.091.422.681.277.684.785.884.889.280.388.8
SimVOS [59]92.991.394.488.085.091.04.980.476.184.684.283.1-79.1-
OneVOS92.791.094.388.584.692.412.084.880.788.986.184.989.880.689.1
OneVOS-D93.191.494.888.885.292.416.083.980.087.685.084.589.279.187.3
OneVOS*92.891.093.788.384.492.112.085.381.389.386.385.089.981.289.1
", + "bbox": [ + 217, + 143, + 802, + 323 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Quantitative comparisons on the DAVIS 2016 val, DAVIS 2017 val, and YouTube-VOS 2019 val split. The methods Use ' as a superscript indicates adding the BL30k [8] dataset for training.", + "bbox": [ + 215, + 335, + 787, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ple settings and datasets. Notably, OneVOS-D excels in the long video dataset LVOS, achieving an impressive $70.1\\%$ $J\\& F$ score under identical training conditions, demonstrating enhanced robustness in long-term videos. On the MOSE dataset, results of OneVOS are particularly noteworthy, surpassing the highest results reported in the paper by nearly $7\\%$ (66.4% vs 59.4%) trained in the same setting. When trained on YouTube, DAVIS, and MOSE together in the main training, OneVOS* achieves state-of-the-art performance across all three datasets.", + "bbox": [ + 215, + 385, + 785, + 482 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Common datasets. Table 2 tabulates the quantitative results on the four standard datasets. The results clearly highlight the superiority of OneVOS. OneVOS-D achieves the best $93.1\\%$ in DAVIS 2016 val and $88.8\\%$ in DAVIS 2017 val respectively, setting new state-of-the-art performance. OneVOS and OneVOS* showcase remark improvement in DAVIS 2017 test and YouTube-VOS 2019 val. This proves that OneVOS has excellent segmentation capabilities in both regular and challenging scenarios, demonstrating its practicability.", + "bbox": [ + 215, + 487, + 785, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results. We present visualization results of OneVOS comparing against the state-of-the-art method XMem [7] and Swin-AOTL [36]. The results for the short-term datasets and complicated long-term videos are shown in Fig. 6 and Fig. 7, respectively. OneVOS consistently achieves finer and higher-quality segmentation on the short-term datasets. Notably, it showcases superior segmentation capabilities in long-term scenarios involving both background interference and small targets.", + "bbox": [ + 215, + 589, + 787, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 217, + 707, + 387, + 722 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we conduct a comprehensive analysis of the OneVOS framework and the efficacy of its core designs on both common and the more challenging datasets.", + "bbox": [ + 215, + 728, + 785, + 756 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3 details an incremental ablation study highlighting the significance of key design elements in OneVOS. Starting with a baseline where ConvVIT replaces the backbone in a top discrete multi-model framework [66] (line 1). Furthermore, to negate any bias stemming from varying modeling approaches toward the backbone, we employ an optimal model based on Swin Transformer [41] as an advanced baseline (line 2), providing a stringent benchmark to validate the effectiveness of OneVOS.", + "bbox": [ + 215, + 757, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 367, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f198866c59fc4b761c44b125f64729abd57f731ea2b2bb4f66a3bec8cbed0f5a.jpg", + "image_caption": [ + "Fig. 7: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging long-term videos (LVOS)." + ], + "image_footnote": [], + "bbox": [ + 225, + 147, + 782, + 452 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "All-in-One Modeling. We initially evaluate the efficacy of All-in-One Modeling by analyzing lines 1, 2, and 3 of Table 3. The outcomes highlight the substantial benefits of All-in-One Modeling. Our model registers a $2.45\\%$ improvement in performance on DAVIS 17 over the enhanced baseline. Moreover, in long video datasets like LVOS, even without any specialized module designs and using only the basic self-attention mechanism, our model exhibits only a minor decrease in performance, demonstrating robust performance of Unification modeling on complex long video datasets.", + "bbox": [ + 212, + 505, + 787, + 603 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Unidirectional Hybrid Attention. In our unidirectional hybrid attention module analysis, shown in lines 3 to 6 of Table 3. Notably, through an examination of lines 4 and 6 we observed that the integration of dual decoupling significantly enhances model performance on both DAVIS 17 val and LVOS datasets. However, Sole decoupling from the current to reference frame results in a performance drop. We analyze that the attributed to the increased focus of reference frame on memory tokens, thereby amplifying the ambiguity in mask embedding and hindering accurate target identification. On the other hand, decoupling solely from memory tokens to reference frame yields a marginal performance enhancement in DAVIS 17 val. However, this improvement is limited as the target features with objects information are represented by tokens from current frame without targets information, failing to provide a consistent feature representation across the video sequence so the performance drops in long videos. These findings highlight the necessity of dual decoupling is indispensable, and their joint effect promotes optimal model performance.", + "bbox": [ + 212, + 609, + 787, + 805 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dynamic Tokens Selector. The comparison results between lines 6 and 7 of Table 3 highlights the role of DTS. DTS not only probes the segmentation mechanisms", + "bbox": [ + 215, + 811, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e76502f6d277285b9109ed5d7f315ff73dd686bf09a592dd560df315aa356ce0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAll-in-OneDecoup1Decoup2DTSDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
182.879.785.960.856.465.1
285.482.488.465.660.875.1
3✓87.584.390.764.960.369.5
4✓✓87.184.090.161.055.766.2
5✓✓87.684.490.862.357.167.5
6✓✓✓88.584.692.467.462.072.9
7✓✓✓✓88.885.292.470.164.775.5
", + "bbox": [ + 245, + 143, + 756, + 250 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/10a0d368f32b09d04a48dec32b5de551de9672ce969b82bcf663a9b16740b4ff.jpg", + "table_caption": [ + "Table 3: Ablation on Core Design of OneVOS. (1-2) baseline: multi-object discrete method with ConvVIT or Swin Transformer [41] as backbone. (3-7): OneVOS with ConvVIT: 'All-in-One': All-in-One modeling. 'Decoup1': decouple attention from current to reference, 'Decoup2': decouple attention between memory and reference. 'DTS': DTS with dynamic memory." + ], + "table_footnote": [], + "table_body": "
MethodDAVIS17 VALMOSE
J&FJFJ&FJF
first 6 layers87.683.791.460.956.865.1
last 6 layers88.084.491.659.155.163.1
all layers88.584.692.762.358.166.6
", + "bbox": [ + 230, + 347, + 500, + 421 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/14a2010c088b765b1390dd7d5f47a9ba987b6e341635e4edb6f3f6df76ed8946.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
DTS-t87.283.690.866.461.271.6
DTS-Lr86.783.190.360.254.965.5
DTS88.885.292.470.164.775.5
", + "bbox": [ + 517, + 347, + 764, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: From left to right: (a) Ablations on different mask embedding adding layers and (b) Ablation of different DTS designs in OneVOS.", + "bbox": [ + 214, + 433, + 784, + 460 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "within OneVOS but, when integrated into the network, it bolsters the learning of this paradigm. It contributes to a more efficient OneVOS by adaptively selecting tokens that align with the model's feature extraction and matching mechanisms, culminating in state-of-the-art performance on both datasets.", + "bbox": [ + 214, + 470, + 787, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Mask Embedding Add Layers. Table. 4(a) explores the effect of integrating mask embedding at different transformer layers. Results show that, by integrating mask embeddings to the immediate value tokens of each layer, the model can capture the global and local semantics of targets at different levels, thereby effectively promoting the model to perform multiple targets inference. Notably, embedding in the first six layers proved more beneficial than in the latter six, this indicates that early interaction between the current and reference frames without target information may negatively impact the feature extraction of the current frame.", + "bbox": [ + 214, + 534, + 787, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "DTS design. The role of DTS is to judiciously select tokens by amalgamating each token feature of the reference frame with the global features of all tokens. Table 4(b) explores the effect of various DTS configurations, including the method using pooled tokens from the current frame as global information (DTS-t), revealing that guiding selection with the current frame may hinder the generalization of OneVOS across frames, consistent with our initial attention decoupling. Besides, Implementing a loss function to maintain a fixed 0.5 selection rate per layer $(\\mathrm{DTS} - L_r)$ significantly reduces performance, underscoring that rigid selection criteria detrimentally affect the capabilities of OneVOS for adaptive matching and feature extraction.", + "bbox": [ + 214, + 651, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Dynamic Memory Capacity. Various memory capacities (Cap in Eq. 8) impact accuracy, inference speed, and storage use. The performance of OneVOS under different memory capacities is validated in Table 5(a). The results reveal that while memory plays a crucial role in handling complex videos, OneVOS achieves robust performance", + "bbox": [ + 214, + 782, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 366, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2750de26832921275f13ea8c0c67a2b33e14bae7e9697b96c9364b3c60995702.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CapD17-VD17-TFPSMem(G)
187.583.216.61.72
288.084.413.72.20
388.484.612.92.47
488.584.612.32.69
588.584.812.02.94
", + "bbox": [ + 220, + 143, + 776, + 220 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 5: Ablations on 3 specific module design of All-in-One framework. From left to right: (a) different $Cap$ of dynamic memory, where D17-V and D17-T denote the $J\\&F$ metrics of DAVIS-17 VAL and TEST datasets, respectively. (b) Different top-k tokens for dynamic memory, and (c) Different Transformer backbones of OneVOS.", + "bbox": [ + 212, + 229, + 785, + 287 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "even with just two frames or a single frame in common scenarios, thanks to its unified framework and efficient architecture. Optimal performance is observed when the memory capacity reaches 4 or 5. This underscores the importance of memory in enabling OneVOS to effectively handle longer and more intricate videos.", + "bbox": [ + 212, + 300, + 784, + 356 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Number of Retained Nodes. Table. 5(b) focuses on retaining a set number of top-k nodes in token memory, based on attention weights. Results show that after reaching a per-layer threshold, keeping approximately 512 nodes struck an optimal balance between information depth and computational efficiency.", + "bbox": [ + 212, + 364, + 785, + 422 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Backbone. In Table. 5(c), we assess Vision Transformers (ViT) [15] pre-trained with Masked Autoencoders (MAE) [25] and ConvVIT (base and large variants) as backbones for OneVOS. With the VIT-B backbone OneVOS still outperforms most of the methods on D17 datasets. However, a key limitation of ViT was identified in its direct 16x downsampling approach, which resulted in a significant loss of multi-scale detail information. This is particularly detrimental for high-precision, pixel-level VOS tasks. On the other hand, the ConvVIT model, which employs simple convolutional operations for its patch embedding process, was found to effectively preserve multi-scale information. Additionally, the increased complexity of the ConvVIT-large variant further enhanced the model's performance, underscoring the importance of architectural design in capturing intricate details necessary for accurate VOS.", + "bbox": [ + 212, + 431, + 787, + 584 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 631, + 359, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces OneVOS, a novel semi-supervised video object segmentation framework that globally integrates feature extraction, matching, memory management, and object aggregation within a single Transformer architecture. By employing the All-in-One Transformer with specialized Unidirectional Hybrid Attention modules, we significantly enhance the practicality and performance of OneVOS. Furthermore, to propose a rational, efficient version of OneVOS, we investigate the mechanism of the All-in-One Transformer via the proposed Dynamic Token Selector, revealing its layer-crossing segmentation mode for multi-object feature extraction and matching, leading to a naturally efficient version with dynamic memory. OneVOS demonstrates state-of-the-art performance on 7 datasets, particularly excelling in complex scenarios. We anticipate that this work will draw greater attention to the unified VOS modeling.", + "bbox": [ + 212, + 686, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300).", + "bbox": [ + 215, + 176, + 828, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 241, + 321, + 257 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bhat, G., Lawin, F.J., Danelljan, M., Robinson, A., Felsberg, M., Van Gool, L., Timofte, R.: Learning what to learn for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 777-794. Springer (2020)", + "2. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)", + "3. Chen, B., Li, P., Bai, L., Qiao, L., Shen, Q., Li, B., Gan, W., Wu, W., Ouyang, W.: Backbone is all your need: A simplified architecture for visual object tracking. In: European Conference on Computer Vision. pp. 375-392. Springer (2022)", + "4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020)", + "5. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8126-8135 (2021)", + "6. Chen, Y., Pont-Tuset, J., Montes, A., Van Gool, L.: Blazingly fast video object segmentation with pixel-wise metric learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1189-1198 (2018)", + "7. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022)", + "8. Cheng, H.K., Tai, Y.W., Tang, C.K.: Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5559-5568 (2021)", + "9. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021)", + "10. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018)", + "11. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence 37(3), 569-582 (2014)", + "12. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13608-13618 (2022)", + "13. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + ], + "bbox": [ + 218, + 272, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 364, + 113, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023)", + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "16. Duke, B., Ahmed, A., Wolf, C., Aarabi, P., Taylor, G.W.: Sstvos: Sparse spatiotemporal transformers for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5912-5921 (2021)", + "17. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)", + "18. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023)", + "19. Gao, J., Qian, X., Wang, Y., Xiao, T., He, T., Zhang, Z., Fu, Y.: Coarse-to-fine amodal segmentation with shape prior. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1262-1271 (2023)", + "20. Gao, P., Ma, T., Li, H., Lin, Z., Dai, J., Qiao, Y.: Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892 (2022)", + "21. Gao, S., Zhou, C., Ma, C., Wang, X., Yuan, J.: Aiatrack: Attention in attention for transformer visual tracking. In: European Conference on Computer Vision. pp. 146-164. Springer (2022)", + "22. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024)", + "23. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022)", + "24. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)", + "25. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "26. Hong, L., Chen, W., Liu, Z., Zhang, W., Guo, P., Chen, Z., Zhang, W.: Lvos: A benchmark for long-term video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13480-13492 (2023)", + "27. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024)", + "28. Hong, L., Zhang, W., Chen, L., Zhang, W., Fan, J.: Adaptive selection of reference frames for video object segmentation. IEEE Transactions on Image Processing 31, 1057-1071 (2021)", + "29. Hu, P., Wang, G., Kong, X., Kuen, J., Tan, Y.P.: Motion-guided cascaded refinement network for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1400-1409 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018)", + "31. Huang, X., Xu, J., Tai, Y.W., Tang, C.K.: Fast video object segmentation with temporal aggregation network and dynamic template matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8879-8889 (2020)", + "32. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)", + "33. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8953-8962 (2019)", + "34. Khoreva, A., Benenson, R., Ilg, E., Brox, T., Schiele, B.: Lucid data dreaming for video object segmentation. International Journal of Computer Vision 127(9), 1175-1197 (2019)", + "35. Kristan, M., Leonardis, A., Matas, J., Felsberg, M., Pflugfelder, R., Cehovin Zajc, L., Vojir, T., Bhat, G., Lukezic, A., Eldesokey, A., et al.: The sixth visual object tracking vot2018 challenge results. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)", + "36. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022)", + "37. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024)", + "38. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018)", + "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "40. Lin, Z., Yang, T., Li, M., Wang, Z., Yuan, C., Jiang, W., Liu, W.: Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1362-1372 (2022)", + "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "42. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018)", + "43. Nowozin, S.: Optimal decisions from probabilistic models: the intersection-over-union case. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 548-555 (2014)", + "44. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 366, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019)", + "46. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)", + "47. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)", + "48. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)", + "49. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021)", + "50. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 629-645. Springer (2020)", + "51. Seong, H., Oh, S.W., Lee, J.Y., Lee, S., Lee, S., Kim, E.: Hierarchical memory matching network for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12889-12898 (2021)", + "52. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015)", + "53. Voigtlaender, P., Chai, Y., Schroff, F., Adam, H., Leibe, B., Chen, L.C.: Feelvos: Fast end-to-end embedding learning for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9481-9490 (2019)", + "54. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017)", + "55. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021)", + "56. Wang, J., Chen, D., Wu, Z., Luo, C., Tang, C., Dai, X., Zhao, Y., Xie, Y., Yuan, L., Jiang, Y.G.: Look before you match: Instance understanding matters in video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2268-2278 (2023)", + "57. Wang, W., Shen, J., Porikli, F., Yang, R.: Semi-supervised video object segmentation with super-trajectories. IEEE transactions on pattern analysis and machine intelligence 41(4), 985-998 (2018)", + "58. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14561-14571 (2023)", + "59. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13879-13889 (2023)", + "60. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)", + "62. Xu, S., Liu, D., Bao, L., Liu, W., Zhou, P.: Mhp-vos: Multiple hypotheses propagation for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 314-323 (2019)", + "63. Yan, S., Xu, X., Hong, L., Chen, W., Zhang, W., Zhang, W.: Panovos: Bridging non-panoramic and panoramic views with transformer for video segmentation. arXiv preprint arXiv:2309.12303 (2023)", + "64. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024)", + "65. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V. pp. 332-348. Springer (2020)", + "66. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "67. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4701-4712 (2021)", + "68. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "69. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357. Springer (2022)", + "70. Zhou, X., Guo, P., Hong, L., Li, J., Zhang, W., Ge, W., Zhang, W.: Reading relevant feature from global representation memory for visual object tracking. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 590 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "OneVOS: All-in-One Transformer Framework For VOS", + "bbox": [ + 366, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_model.json b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6b84df3801f690616bd5138512b257c0a11a3337 --- /dev/null +++ b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_model.json @@ -0,0 +1,2691 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.141, + 0.77, + 0.187 + ], + "angle": 0, + "content": "OneVOS: Unifying Video Object Segmentation with All-in-One Transformer Framework" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.212, + 0.761, + 0.246 + ], + "angle": 0, + "content": "Wanyun Li\\(^{1,*}\\), Pinxue Guo\\(^{2,*}\\), Xinyu Zhou\\(^{1}\\), Lingyi Hong\\(^{1}\\), Yangji He\\(^{1}\\), Xiangyu Zheng\\(^{1}\\), Wei Zhang\\(^{1,\\dagger}\\), and Wenqiang Zhang\\(^{1,3,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.255, + 0.714, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.283, + 0.754, + 0.354 + ], + "angle": 0, + "content": "\\(^{2}\\) Shanghai Engineering Research Center of AI & Robotics, Academy for Engineering & Technology, Fudan University, Shanghai, China \n\\(^{3}\\) Engineering Research Center of AI & Robotics, Ministry of Education, Academy for Engineering & Technology, Fudan University, Shanghai, China {wyli22,pxguo21}@m.fudan.edu.cn, {weizh, wqzhang}@fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.379, + 0.744, + 0.726 + ], + "angle": 0, + "content": "Abstract. Contemporary Video Object Segmentation (VOS) approaches typically consist of feature extraction, matching, memory management, and multiple objects aggregation. Recent advanced models either employ a discrete modeling for these components in a sequential manner, or optimize a combined pipeline through substructure aggregation. However, these existing explicit staged approaches prevent the VOS framework from being optimized as a unified whole, leading to the limited capacity and suboptimal performance in tackling complex videos. In this paper, we propose OneVOS, a novel framework that unifies the core components of VOS with All-in-One Transformer. Specifically, to unify all aforementioned modules into a vision transformer, we model all the features of frames, masks and memory for multiple objects as transformer tokens, and integrally accomplish feature extraction, matching and memory management of multiple objects through the flexible attention mechanism. Furthermore, a Unidirectional Hybrid Attention is proposed through a double decoupling of the original attention operation, to rectify semantic errors and ambiguities of stored tokens in OneVOS framework. Finally, to alleviate the storage burden and expedite inference, we propose the Dynamic Token Selector, which unveils the working mechanism of OneVOS and naturally leads to a more efficient version of OneVOS. Extensive experiments demonstrate the superiority of OneVOS, achieving state-of-the-art performance across 7 datasets, particularly exceeding in complex LVOS and MOSE datasets with \\(70.1\\%\\) and \\(66.4\\%\\) \\(J\\& F\\) scores, surpassing previous state-of-the-art methods by \\(4.2\\%\\) and \\(7.0\\%\\), respectively. Code is available at: https://github.com/L599wy/OneVOS." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.76, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Semi-supervised Video Object Segmentation (VOS) [2, 22, 30, 37, 44, 45, 63, 64] is a fundamental task in video analysis [18, 19], aiming to consistently seg-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.555, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.498, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.146, + 0.761, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.211, + 0.503, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.289, + 0.453, + 0.298 + ], + "angle": 0, + "content": "(b) Unify Matching and Objects Aggregation" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.21, + 0.761, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.288, + 0.738, + 0.297 + ], + "angle": 0, + "content": "(d) Integrate Core Modules Al in One Transformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.785, + 0.364 + ], + "angle": 0, + "content": "Fig. 1: Four different modeling for semi-supervised video object segmentation pipelines. (a) Traditional discrete modeling of VOS. (b) Unified multiple object modeling in one reference. (c) Unified single-object feature extraction and matching. (d) OneVOS: Integrates all core modules with All-in-One Transformer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.493 + ], + "angle": 0, + "content": "ment objects throughout a video based on initial masks. Advancements in VOS primarily focus on the full utilization of available information. Among them, Memory-based methods [7,9,40,45] have emerged as the mainstream due to the effectiveness in storing and leveraging abundant spatio-temporal features. Its foundational paradigm, as depicted in Fig. 1(a) [16,40,51,55], encompasses four split key stages including feature extraction, matching, memory management, and multi-object aggregation, which have been widely adopted as a baseline in a number of methods [23,28,50,56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Recently, some investigations demonstrate the advantages of integration modeling compared with highly discrete modeling. AOT [66] and DEAOT [68] (Fig. 1(b)) emphasize the shortage of single-object inference compared to joint modeling of multiple objects, particularly in exploring contextual and robust feature representations. Meanwhile, SimVOS [59] (Fig. 1(c)) reveals limitations in the separation of feature extraction and matching processes for capturing dynamic object-aware features. However, existing methods solely focus on combining or aggregating substructures of the VOS pipeline, resulting in their suboptimal solutions. For example, AOT relies on sophisticated designing matching modules which fails to leverage superior pre-training weights, while the simplistic single-object framework of SimVOS encounters inefficiencies and the inapplicability of memory makes it unable to utilize rich features of the processed frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Therefore, we introduce an essentially novel framework distinct from previous methods, aiming to develop a unified end-to-end differentiable VOS model for global optimization. Based on this, the All-in-One Transformer for Video Object Segmentation (OneVOS) is proposed, which models and unifies all important modules into a Vision Transformer. Specifically, we first leverage the flexible attention mechanism to conduct feature extraction and mutual matching between the reference and current frames. To effectively capture multi-object relationships, we introduce a mask embedding layer that enriches features with multi-object modeling insights. Moreover, in the memory design, we propose an external layer-wise token memory to enhance the coherence understanding of videos by storing and providing spatio-temporal features for subsequent frames." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "As the feature extraction and matching of multiple objects performing simultaneously, the reference tokens output from each layer of Transformer for memory update could be represented by a weighted linear combination of tokens from memory, reference frame and current frame, which may lead to a significant semantic errors and ambiguity to serving as features stored in memory. To address this, we further propose a Unidirectional Hybrid Attention (Uni-Hybrid Attention) through a double decoupling of the original attention operation. It guarantees the versatility of stored features and enhances memory performance. Notably, as the Uni-Hybrid Attention is derived from naive attention mechanism, our framework is still able to utilize pre-trained weights from various Vision Transformers, thereby enhancing its generality and flexibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.555 + ], + "angle": 0, + "content": "In this way, the proposed OneVOS framework implements end-to-end modeling and global optimization, exhibiting superior performance compared to previous models. To further strengthen its advantages, we conduct an in-depth exploration of the behavior within the attention layer of OneVOS to enhance its efficiency, which is particularly crucial in resource-constrained and long-term video tasks. Concretely, we introduce the Dynamic Token Selector (DTS), capable of dynamically selecting partial tokens from each layer for cross-attention and memory storage, trained solely under segmentation loss without additional constraints. With DTS integrated into every layer of the transformer, we analyze the token selection distribution of DTS and the attention weight distribution, uncovering a significant insight: unlike the traditional sequential paradigm, OneVOS adopts a cross-layer segmentation mode that dynamically shifts focus between feature extraction and matching across different layers. Based on this, we propose a natural acceleration scheme involving designing distinct memory capacities for each layer and layer-wise memory updates via tokens selected by DTS. This approach enhances both speed and accuracy across multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.79, + 0.631 + ], + "angle": 0, + "content": "We conducted comprehensive experiments across seven datasets, and OneVOS consistently achieves state-of-the-art performance across all of them. Notably, OneVOS demonstrates a substantial performance advantage enven in more challenging scenarios, achieving \\(71.1\\%\\) \\(\\mathcal{I}\\& \\mathcal{F}\\) on LVOS validation and \\(67.2\\%\\) \\(\\mathcal{I}\\& \\mathcal{F}\\) on MOSE validation. The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.703 + ], + "angle": 0, + "content": "- We propose All-in-One Transformer framework for Video Object Segmentation (OneVOS), integrating feature extraction, matching, memory management, and objects aggregation into a vision transformer, which can be optimized globally. This provides a fundamentally novel paradigm in VOS." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.704, + 0.787, + 0.749 + ], + "angle": 0, + "content": "- We introduce the Unidirectional Hybrid Attention mechanism to facilitate multi-object feature extraction and matching, ensuring tokens are devoid of semantic ambiguity for efficient memory storage." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.75, + 0.787, + 0.794 + ], + "angle": 0, + "content": "- We propose the Dynamic Tokens Selector, which not only unveils the internal working mechanism but also leads to a more efficient OneVOS, based on the exploration and the resulting insight." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- OneVOS achieves state-of-the-art performance across 7 datasets, particularly excelling in more complex LVOS and MOSE datasets with \\(70.1\\%\\) and \\(66.4\\%\\) \\(J\\& F\\), surpassing previous best methods by \\(4.2\\%\\) and \\(7.0\\%\\), respectively." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.328, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.183, + 0.483, + 0.198 + ], + "angle": 0, + "content": "2.1 Video object segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.787, + 0.346 + ], + "angle": 0, + "content": "Semi-supervised VOS methods can be divided into four types: The earlier fine-tuning methods [1, 2, 42, 54, 60] learn general segmentation features via offline pre-training, then enhance object-specific representations by online fine-tuning. Then, propagation-based methods [10, 29, 33, 44, 46, 57] utilize optical flow to propagate the mask from the previous frame to the next one. Subsequent methods optimized it by keyframes enhancement [34], advanced tracking techniques [4, 10, 31, 62], and re-identification mechanisms [38]. Matching-based methods [6, 30, 53, 65, 67] identify the object of the current frame based on the correlation with the reference frame in the embedding space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Recently, Memory-based approaches [23,28,45,50] store features of segmented frames in a memory bank to provide a set of features for segmentation, which has become the dominant method in SVOS. Following that, methods such as HMMN [51], SSTVOS [16], and XMem [7] enhance performance by utilizing multi-scale features, introducing Transformer for matching, and building long-term memory. These methods typically adhere to a traditional discrete modeling and single-object architecture. More recently, the multi-object framework [66], [68] is proposed to address the inherent limitations of the previous single-object while still maintaining a sophisticated matching module for high performance. [59] presents a simplified VOS framework for joint feature extraction and matching which adopts single-object modeling and discards the vital memory design in VOS. Differently, our OneVOS, innovatively unifies feature extraction, matching, memory construction, and multi-objective aggregation all in one Transformer. This design not only streamlines the segmentation process but also enhances the dynamic interaction among different stages of SVOS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.455, + 0.617 + ], + "angle": 0, + "content": "2.2 Visual Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Video object tracking (VOT) [3,27,35,70] shares similarities with the task setting of SVOS, but it primarily utilizes bounding boxes for object localization instead of detailed segmentation masks. In VOT, there has been a paradigm shift from traditional Siamese-based methods towards leveraging Transformer layers for enriched relational modeling between the template and search regions. Pioneering works like TransT [5] introduce a two-stream attention mechanism for robust bi-directional information exchange, while innovative methods like MixFormer [12] and AiATrack [21] adopt asymmetric attention strategies to minimize distractions from irrelevant objects. Notably, OSTrack [69] proposes a unified attention model, promoting seamless feature integration and relation modeling. Nevertheless, the distinct multi-object settings and difference in target information representation, coupled with the critical role of memory in VOS, render unification within the VOS field more challenging. This paper endeavors to tackle this challenge." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.114, + 0.733, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.327 + ], + "angle": 0, + "content": "As shown in the Fig. 2, in a given video sequence, VOS focuses on segmenting the current \\( t \\)-th frame \\( f_{t} \\in \\mathbb{R}^{H \\times W \\times 3} \\) using the reference frame \\( f_{t-1} \\in \\mathbb{R}^{H \\times W \\times 3} \\) with its mask and the history features stored in Memory. In this section, we aim to construct an unified framework at a global scale to tackle the challenges of VOS. We commence by introducing the OneVOS framework in Sec. 3.1, which employs Vision Transformer to integrate core elements in VOS. And then Sec. 3.2 provides a detailed exposition on the novel Unidirectional Hybrid Attention mechanism. In addition, in Sec. 3.3, we explore the working mechanism of attention within OneVOS through the proposed Dynamic Token Selector in order to introduce the efficient version of OneVOS." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.332, + 0.473, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.444, + 0.413, + 0.453 + ], + "angle": 0, + "content": "(a) Hybrid Attention Block" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.332, + 0.753, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.444, + 0.704, + 0.454 + ], + "angle": 0, + "content": "(b) Block with Dynamic Token Selector (DTS)" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.459, + 0.753, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.592, + 0.527, + 0.602 + ], + "angle": 0, + "content": "(c) Overview" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.613, + 0.789, + 0.738 + ], + "angle": 0, + "content": "Fig. 2: (a) Hybrid Attention Block. (b) Blocks with Dynamic Token Selector. (c) Overview of OneVOS framework equipped with All-in-One Transformer. The reference mask is initially encoded into mask embeddings. Patch embeddings are generated from both the reference and current frames, which are then processed by the Transformer. Within each Transformer layer, mask embeddings are integrated with reference tokens. Then they are selected by Dynamic Token Selector and concatenated with tokens stored in memory and current frame fed into the Hybrid Attention Module. Finally, the features of the current frame output from Transformer are subsequently decoded into multi-object masks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.431, + 0.785 + ], + "angle": 0, + "content": "3.1 OneVOS Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Mask Embedding. To capture complex target interactions and minimize redundant single-target inferences, motivated by [66], we utilize the Mask Embedding Layer (as Patch Embedding Layer in Transformer) to encode reference" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.268 + ], + "angle": 0, + "content": "mask into Mask Embedding denoted as \\( m_{ref} \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P} \\times C} \\) through a single convolutional layer. Here, \\( C \\) represents the embedding dimension, with \\( P \\) as the stride of the mask embedding. Utilizing the local relationship modeling capability of the convolution operation, the discriminability of multiple targets is enhanced, which will then be integrated with visual feature map in Transformer before attention computation to integrating mask embeddings to the immediate value tokens of each layer, thereby effectively promoting the model to perform multiple targets inference." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.273, + 0.789, + 0.397 + ], + "angle": 0, + "content": "All-in-One Transformer. Then, we adopt a transformer architecture to unify the core modules of VOS. It first generates the non-overlapping patch embeddings \\( e_t^0 \\in \\mathbb{R}^{N \\times C} \\) and \\( e_{ref}^0 \\in \\mathbb{R}^{N \\times C} \\) for the current and reference frames, where \\( P \\) is the patch resolution same as the mask embedding, thus \\( N = HW / P^2 \\). Then, the concatenation of \\( [e_{ref}^0; e_t^0] \\) along the spatial dimension serves as the initial input for the backbone, facilitating adaptive multi-object feature extraction and target information propagation through the stacked layers. In each layer, we first integrate mask embedding with visual feature map as shown in Eq 1:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.402, + 0.786, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\left\\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} \\right\\} = \\left\\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} + m _ {r e f} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.426, + 0.788, + 0.517 + ], + "angle": 0, + "content": "where \\( Q \\), \\( K \\), and \\( V \\) are query, key, and value matrices aligning with the standard attention mechanism [15], and \\( l \\in \\{0, 1, \\dots, L\\} \\) represents the layer. Then, these object-aware reference tokens, memory tokens, and tokens of current frames undergo adaptive feature extraction and target information propagation by applying the attention mechanism of the Transformer. The output \\( a^l = [a_{ref}^l; a_t^l] \\) of layer \\( l \\) after the attention operation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.533, + 0.786, + 0.567 + ], + "angle": 0, + "content": "\\[\na ^ {l} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot V = \\operatorname {S o f t m a x} \\left(\\frac {\\left[ Q _ {r e f} ^ {l} ; Q _ {t} ^ {l} \\right] \\left[ K _ {M} ^ {l} ; K _ {r e f} ^ {l} ; K _ {t} ^ {l} \\right]}{\\sqrt {d _ {k}}}\\right) \\cdot \\left[ V _ {M} ^ {l}; V _ {r e f} ^ {l}; V _ {t} ^ {l} \\right], \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.571, + 0.788, + 0.602 + ], + "angle": 0, + "content": "where \\( K_M^l, V_M^l \\) are features stored in memory. Supposing \\( W = QK^{T} \\) is the matrix of similarity weights, \\( a_{t}^{l} \\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.608, + 0.785, + 0.623 + ], + "angle": 0, + "content": "\\[\na _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.628, + 0.788, + 0.698 + ], + "angle": 0, + "content": "where \\([W_{t,ref}V_{ref}]\\) and \\([W_{t,M}V_M]\\) are responsible for aggregating the inter-frame features from memory and reference frame to current frame (matching), and \\([W_{t,t}V_t]\\) aggregating the intra-frame features (feature extraction) for current frame, respectively. Thus, dynamic coordination between feature extraction and matching is achieved through the attention weights learned adaptively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Token Memory. In addition, since the target information and the reference frame are integrated within our framework, we can store intermediate features directly into Token Memory without necessitating additional post-processing fusion steps. Given the semantic variance of intermediate tokens in the reference frame across layers, we introduce a token storage mechanism to save essential Transformer tokens instead of whole frames as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.79, + 0.786, + 0.808 + ], + "angle": 0, + "content": "\\[\nM = \\left\\{K _ {M} ^ {l}, V _ {M} ^ {l} \\mid \\in \\{0, 1, \\dots L \\} \\right\\} = \\left\\{K _ {\\text {r e f s}} ^ {l}, V _ {\\text {r e f s}} ^ {l} \\mid \\in \\{0, 1, \\dots L \\} \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.812, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This method allows these tokens to be involved only in the attention modules instead of multiple inferences, greatly reducing inference overhead." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.497, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.147, + 0.641, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.149, + 0.785, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.237, + 0.785, + 0.333 + ], + "angle": 0, + "content": "Fig. 3: (a) Visualization of attention distribution of target areas in OneVOS, showing the alternation between self-attention and cross-attention across layers. (b) Analysis of the maximal attention weights reveals a pattern of alternation between feature extraction via self-attention (notably in layers 3 and 5) and matching via cross-attention (as in layers 2 and 8). (c) The proportional analysis of the attention weights' sum, normalized to one, underscores the non-sequential prioritization across different layers between self-attention and cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.342, + 0.785, + 0.385 + ], + "angle": 0, + "content": "Mask Decoding. Finally, the output of All-in-One Transformer, \\( e_t^L \\) is input to the lightweight Feature Pyramid Network (FPN) for Mask Decoding. These features are then gradually upgraded to the original scale and decoded into multi-object masks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.53, + 0.423 + ], + "angle": 0, + "content": "3.2 Unidirectional Hybrid Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.516 + ], + "angle": 0, + "content": "Building upon our newly developed pipeline, all modules are integrated into a transformer as in Fig. 2(c). However, obtaining the stored features by employing original attention mechanism may lead to semantic errors and ambiguity since they are represented incorrectly. To address this challenge, our proposed Unidirectional Hybrid Attention (Uni-Hybrid Attention) is strategically designed through dual-level attention decoupling as shown in Fig. 2(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.787, + 0.627 + ], + "angle": 0, + "content": "Firstly, reference features calculated by naive attention operation to all the tokens are denoted as \\( a_{ref}^{l} = W_{ref,M}V_{M} + W_{ref,ref}V_{ref} + W_{refs,t}V_{t} \\). A key limitation in the computations arises where tokens of reference frame containing target information (as in Eq. (1)) are represented by tokens of current frame that without target information, causing semantic errors in the features stored in the memory which are essential for storing spatio-temporal target information for subsequent frames. Therefore, to alleviate this situation, our first modification decoupling attention from the current token to the reference token. The revised expression is:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.636, + 0.785, + 0.653 + ], + "angle": 0, + "content": "\\[\na _ {r e f} ^ {l} = W _ {r e f, M} V _ {M} + W _ {r e f, r e f} V _ {r e f}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.788, + 0.799 + ], + "angle": 0, + "content": "Subsequently, the attention computation between reference features and memory tokens is decoupled, serving a dual purpose: 1) Avoiding Information Overlap and Ambiguity: Memory tokens, enriched with mask embeddings, encapsulate target information from prior frames. Integrating these embeddings with those of the reference frame risks overlapping and confusing object cues. 2) Preserving Temporal Specificity: Every video frame contains essential temporal intricacies crucial for segmentation precision. Attention decoupling preserves the reference frame's temporal uniqueness, critical for accurate and context-aware segmentation. Therefore, we decouple attention between memory tokens and those from the reference frame, and the final unidirectional hybrid attention formulations are:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.806, + 0.786, + 0.831 + ], + "angle": 0, + "content": "\\[\na _ {r e f} ^ {l} = W _ {r e f, r e f} V _ {r e f} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.828, + 0.625, + 0.843 + ], + "angle": 0, + "content": "\\[\na _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}.\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.411, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.145, + 0.603, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.236, + 0.417, + 0.246 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.154, + 0.787, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.236, + 0.712, + 0.246 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.257, + 0.785, + 0.299 + ], + "angle": 0, + "content": "Fig. 4: (a) Visualization of the selected tokens of reference frame by DTS in different Transformer layers. (b) Adaptive selection ratios learned by DTS in different layers without any constraints." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.482, + 0.321 + ], + "angle": 0, + "content": "3.3 Towards Efficient OneVOS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.4 + ], + "angle": 0, + "content": "Utilizing the advanced All-in-One Transformer equipped with Uni-Hybrid Attention, OneVOS achieves outstanding performance across various datasets. However, the integration process of these modules in OneVOS still lacks explainability, hindering the rational design of a more efficient version. Therefore, this section delves into its working mechanism to propose the efficient OneVOS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.572 + ], + "angle": 0, + "content": "Exploration and Insight. Within the Transformer architecture, attention weights determine the influence of each token in the input sequence on the output which is useful for interpreting Transformer behavior. By examining the attention weights of tokens in the target area of the current frame across each layer, we compare their focus in reference frames or themselves. As presented in Fig. 3, qualitative results in 3(a) shown that a fluctuating focus between reference frames (cross-attention) and themselves (self-attention), Subsequently, Fig. 3 (b) and (c) provide quantitative evidence of a complex alternation pattern, which deviates from the sequential feature extraction and matching process of traditional discrete VOS modeling or uniform distribution in each layer. Instead, these findings lead us to a natural insight: OneVOS may learn a dynamic layer-wise integration strategy, preferentially executing multi-object feature extraction or matching in each layer according to the semantic context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.785, + 0.647 + ], + "angle": 0, + "content": "Dynamic Token Selector. Based on above insights and motivated by [49], we introduce a Dynamic Token Selector (DTS) into the OneVOS architecture, as depicted in Fig. 2(b). This module adaptively selects a subset of tokens from the reference frame for cross-attention and storage. The selection process, as defined in Eq. 7, combines global context and local token information:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.655, + 0.785, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p = \\operatorname {S o f t m a x} (\\operatorname {M L P} \\left(\\left[ \\operatorname {M a x P o o l} \\left(z _ {r e f} ^ {l}\\right); z _ {r e f} ^ {l}\\right]\\right)) \\\\ z _ {r e f} ^ {l} = \\operatorname {M L P} \\left(\\operatorname {N o r m} \\left(e _ {r e f} ^ {l}\\right)\\right) \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.784, + 0.757 + ], + "angle": 0, + "content": "where \\( z_{ref}^{l} \\) denotes the local feature encoded of each token, while [MaxPool(\\( z_{ref}^{l} \\))] aggregates the overall context of the reference frame. \\( p \\in \\mathbb{R}^{N \\times 2} \\) represents the probabilities associated with each token. Specifically, \\( p_{i,0} \\) indicates the probability of dropping \\( i \\)-th token while \\( p_{i,1} \\) indicates for selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.785, + 0.84 + ], + "angle": 0, + "content": "This design aims to enable the network to autonomously determine the optimal token selection ratio for high-performance segmentation. Consequently, it is integrated into an All-in-One Transformer, as depicted in Fig. 2(a). Moreover, it is exclusively trained under segmentation loss without additional constraints. By observing the proportion of tokens selected by DTS in each transformer layer, as illustrated in Fig. 4(a) and Fig. 4(b), a crucial insight emerges. The model automatically allocates varying" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.788, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.258, + 0.785, + 0.286 + ], + "angle": 0, + "content": "Fig. 5: Visualization dynamic memory updates of OneVOS utilizing DTS. The selected tokens in red boxes are derived output from DTS as Fig. 4(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.784, + 0.365 + ], + "angle": 0, + "content": "selection ratios to different layers, and even assigns a selection ratio of zero or one to specific layers. This observation underscores the focus of DTS on certain layers for feature extraction (selecting a proportion of reference frame tokens as zero) and on specific layers for matching (selecting a proportion of reference frame tokens as one), presenting a characteristic of alternating focus across layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.439 + ], + "angle": 0, + "content": "Efficient Version of One VOS. Building upon the above observations, a straightforward acceleration method is naturally introduced. This includes layer-wise memory capacity and adaptive token update mechanism via DTS, as shown in Fig. 5. Specifically, we initialize different capacities for each layer of Token Memory, with the token capacity formula as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.44, + 0.785, + 0.455 + ], + "angle": 0, + "content": "\\[\nc _ {l} = r _ {l} * N * C a p. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Here, \\( r_l \\) represents the layer-specific selection ratio determined by DTS during training, as illustrated in Fig. 4(b). Additionally, \\( Cap \\) is a hyper-parameter denoting the frame capacity (typically set to 3 or 5). Consequently, tokens are incrementally added to the memory based on these selection ratios until reaching the capacity limit of each layer. Subsequently, the top-k tokens, determined by the current attention weights \\( W_{t,ref} \\), are retained, followed by the continuous addition of tokens to memory. Aligned with the segmentation mode of OneVOS, this method achieves improvements in accuracy and speed across multiple datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.367, + 0.611 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.457, + 0.644 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.785, + 0.696 + ], + "angle": 0, + "content": "One-stream VOS uses ConvVIT-base [20] as the backbone, initialized with ImageNet-1k [13] pre-training weights. The feature resolution of the mask decoding is increased from 1/16 to \\(1/4\\) and the channel dimension is reduced from 768 to 384." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Training details. Following the common settings of VOS [8,66], the training phase is divided into two stages: pre-training on synthetic video sequences generated from still image datasets [11,17,24,39,52], and main training on standard VOS dataset (DAVIS 17 and YouTube VOS in default, the model mark with \"*\" uses MOSE in main training as well). The loss function is a 0.5:0.5 combination of bootstrapped cross-entropy loss and soft Jaccard loss [43]. The learning rate of two stages is 2e-4 and 1e-4, respectively, utilizing \\(4 \\times 3090\\)s GPUs. Our batchsize is 4, with a gradient accumulation of 2 and a sequence length of 5, running each training phase for 200,000 iterations. In the training of DTS, we employ the Gumbel-Softmax [32] on its output \\(p\\) to generate a one-hot selection vector, indicating whether each token is selected (assigned a value of 1) or" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.145, + 0.784, + 0.316 + ], + "angle": 0, + "content": "
Long-term VideosComplicated Videos
MethodLVOS VALLVOS TESTMethodMOSEJ&FJF
J&FJFJ&FJF
CFBI [65]50.045.055.144.840.549.0XMem [7]x56.452.160.6
AOT-L [66]61.359.453.654.148.759.5DropSeg [58]x52.248.056.4
STCN [9]45.841.150.545.841.650.0R50-AOT-L [66]x58.153.962.3
RDE [36]52.947.758.149.044.453.5R50-DeAOT-L [68]x58.954.563.3
XMem [7]50.045.554.449.545.253.7SwinB-AOT-L [66]x60.056.164.0
AOT-B [66]56.951.861.954.449.359.4OneVOSx62.358.166.6
DeAOT [68]63.757.669.858.152.963.4RDE [36]✓48.844.652.9
R50-AOT-L [66]63.858.169.661.056.165.9STCN [9]✓50.846.655.0
R50-DeAOT-L [68]65.959.971.958.053.262.8XMem [7]✓57.653.362.0
Swin-AOT-L [66]65.660.875.161.556.474.8DeAOT [68]✓59.455.163.8
OneVOS-D70.164.775.560.155.668.0OneVOS-D✓61.757.566.0
OneVOS67.462.072.963.859.570.0OneVOS✓66.462.270.6
OneVOS*71.166.076.264.160.068.2OneVOS*✓67.263.071.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.327, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of long-duration video datasets (LVOS validation and test) and complex datasets, MOSE. \\(\\checkmark\\) indicates MOSE used for training. OneVOS( \\(\\checkmark\\) ) only employs MOSE in main training, whereas OneVOS\\* \\((\\checkmark)\\) utilizes DAVIS, YouTube, and MOSE." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.398, + 0.787, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.53, + 0.785, + 0.559 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging short-term videos (MOSE and DAVIS 17)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.566, + 0.785, + 0.595 + ], + "angle": 0, + "content": "not (value of 0). Throughout training, we record the selection rate, observing that in the later stages of training, the selection rate of each layer gradually stabilizes." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.599, + 0.785, + 0.682 + ], + "angle": 0, + "content": "Inference details. OneVOS utilizes basic memory storage, storing tokens at intervals of 5 or 10 frames and updates using the first-in-first-out queue when it is full, except for the initial frame tokens. Conversely, OneVOS-D employs dynamic memory, defaulting to retain a maximum of 520 top-k tokens when full. The maximum frame-level capacity \\( Cap \\) varies by dataset, set to \\( 3/5/8 \\). For fair comparison in LVOS, we strictly limit memory to under 6." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.687, + 0.787, + 0.772 + ], + "angle": 0, + "content": "Datasets and Metrics. We evaluate the performance of OneVOS on 4 standard VOS datasets, including DAVIS 16 val [47], DAVIS 17 val [48], DAVIS 17 test [48], and YouTube 2019 [61], as well as 3 more challenging datasets, MOSE [14], LVOS val, and LVOS test [26]. The evaluation metrics include region similarity \\(\\mathcal{I}\\), boundary accuracy \\(\\mathcal{F}\\), and their average \\(\\mathcal{J} \\& \\mathcal{F}\\), and FPS for speed. For YouTube-VOS, we also report them from the seen \\(\\mathcal{J}_s\\), \\(\\mathcal{F}_s\\) and unseen \\(\\mathcal{J}_u\\), \\(\\mathcal{F}_u\\) categories." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.79, + 0.616, + 0.807 + ], + "angle": 0, + "content": "4.2 Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Complicated datasets. Table 1 tabulates the quantitative results on the LVOS val, LVOS test, and MOSE. OneVOS demonstrates superior performance across multi-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.145, + 0.803, + 0.324 + ], + "angle": 0, + "content": "
MethodDAVIS 16 VALDAVIS 17 VALDAVIS 17 TESTYOUTUBE 19 VAL
J&FJFJ&FJFFPSJ&FJFGJsFsJuFu
CFBI+ [67]89.988.791.182.980.185.75.678.074.481.682.681.786.277.185.2
HMMN [51]90.889.692.084.781.987.55.078.674.782.582.581.786.177.385.0
DropSeg [58]92.190.993.386.583.589.5----83.482.987.377.785.6
STCN [9]91.690.892.585.482.288.620.276.172.779.682.781.185.478.285.9
XMem [7]91.590.492.786.282.989.522.681.077.484.585.584.388.680.388.6
Swin-AOTL [66]92.090.793.385.482.488.412.181.277.385.184.584.088.878.486.7
Swin-DeAOTL [68]92.991.194.786.283.189.215.482.878.986.786.185.390.280.488.6
ISVOS [56]92.691.593.787.183.790.5-82.879.386.286.185.289.780.788.9
MiVOS' [8]91.089.692.484.581.787.411.278.674.982.282.480.684.778.186.4
STCN' [9]91.790.493.085.382.088.620.277.874.381.384.282.687.079.487.7
XMem' [7]92.090.793.287.784.091.422.681.277.684.785.884.889.280.388.8
SimVOS [59]92.991.394.488.085.091.04.980.476.184.684.283.1-79.1-
OneVOS92.791.094.388.584.692.412.084.880.788.986.184.989.880.689.1
OneVOS-D93.191.494.888.885.292.416.083.980.087.685.084.589.279.187.3
OneVOS*92.891.093.788.384.492.112.085.381.389.386.385.089.981.289.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.336, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons on the DAVIS 2016 val, DAVIS 2017 val, and YouTube-VOS 2019 val split. The methods Use ' as a superscript indicates adding the BL30k [8] dataset for training." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.386, + 0.787, + 0.483 + ], + "angle": 0, + "content": "ple settings and datasets. Notably, OneVOS-D excels in the long video dataset LVOS, achieving an impressive \\(70.1\\%\\) \\(J\\& F\\) score under identical training conditions, demonstrating enhanced robustness in long-term videos. On the MOSE dataset, results of OneVOS are particularly noteworthy, surpassing the highest results reported in the paper by nearly \\(7\\%\\) (66.4% vs 59.4%) trained in the same setting. When trained on YouTube, DAVIS, and MOSE together in the main training, OneVOS* achieves state-of-the-art performance across all three datasets." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.586 + ], + "angle": 0, + "content": "Common datasets. Table 2 tabulates the quantitative results on the four standard datasets. The results clearly highlight the superiority of OneVOS. OneVOS-D achieves the best \\(93.1\\%\\) in DAVIS 2016 val and \\(88.8\\%\\) in DAVIS 2017 val respectively, setting new state-of-the-art performance. OneVOS and OneVOS* showcase remark improvement in DAVIS 2017 test and YouTube-VOS 2019 val. This proves that OneVOS has excellent segmentation capabilities in both regular and challenging scenarios, demonstrating its practicability." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.59, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Qualitative Results. We present visualization results of OneVOS comparing against the state-of-the-art method XMem [7] and Swin-AOTL [36]. The results for the short-term datasets and complicated long-term videos are shown in Fig. 6 and Fig. 7, respectively. OneVOS consistently achieves finer and higher-quality segmentation on the short-term datasets. Notably, it showcases superior segmentation capabilities in long-term scenarios involving both background interference and small targets." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.708, + 0.388, + 0.723 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.729, + 0.786, + 0.757 + ], + "angle": 0, + "content": "In this section, we conduct a comprehensive analysis of the OneVOS framework and the efficacy of its core designs on both common and the more challenging datasets." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Table 3 details an incremental ablation study highlighting the significance of key design elements in OneVOS. Starting with a baseline where ConvVIT replaces the backbone in a top discrete multi-model framework [66] (line 1). Furthermore, to negate any bias stemming from varying modeling approaches toward the backbone, we employ an optimal model based on Swin Transformer [41] as an advanced baseline (line 2), providing a stringent benchmark to validate the effectiveness of OneVOS." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.148, + 0.784, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.468, + 0.785, + 0.497 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging long-term videos (LVOS)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.604 + ], + "angle": 0, + "content": "All-in-One Modeling. We initially evaluate the efficacy of All-in-One Modeling by analyzing lines 1, 2, and 3 of Table 3. The outcomes highlight the substantial benefits of All-in-One Modeling. Our model registers a \\(2.45\\%\\) improvement in performance on DAVIS 17 over the enhanced baseline. Moreover, in long video datasets like LVOS, even without any specialized module designs and using only the basic self-attention mechanism, our model exhibits only a minor decrease in performance, demonstrating robust performance of Unification modeling on complex long video datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Unidirectional Hybrid Attention. In our unidirectional hybrid attention module analysis, shown in lines 3 to 6 of Table 3. Notably, through an examination of lines 4 and 6 we observed that the integration of dual decoupling significantly enhances model performance on both DAVIS 17 val and LVOS datasets. However, Sole decoupling from the current to reference frame results in a performance drop. We analyze that the attributed to the increased focus of reference frame on memory tokens, thereby amplifying the ambiguity in mask embedding and hindering accurate target identification. On the other hand, decoupling solely from memory tokens to reference frame yields a marginal performance enhancement in DAVIS 17 val. However, this improvement is limited as the target features with objects information are represented by tokens from current frame without targets information, failing to provide a consistent feature representation across the video sequence so the performance drops in long videos. These findings highlight the necessity of dual decoupling is indispensable, and their joint effect promotes optimal model performance." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.812, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Dynamic Tokens Selector. The comparison results between lines 6 and 7 of Table 3 highlights the role of DTS. DTS not only probes the segmentation mechanisms" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.145, + 0.758, + 0.25 + ], + "angle": 0, + "content": "
MethodAll-in-OneDecoup1Decoup2DTSDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
182.879.785.960.856.465.1
285.482.488.465.660.875.1
3✓87.584.390.764.960.369.5
4✓✓87.184.090.161.055.766.2
5✓✓87.684.490.862.357.167.5
6✓✓✓88.584.692.467.462.072.9
7✓✓✓✓88.885.292.470.164.775.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.261, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Table 3: Ablation on Core Design of OneVOS. (1-2) baseline: multi-object discrete method with ConvVIT or Swin Transformer [41] as backbone. (3-7): OneVOS with ConvVIT: 'All-in-One': All-in-One modeling. 'Decoup1': decouple attention from current to reference, 'Decoup2': decouple attention between memory and reference. 'DTS': DTS with dynamic memory." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.348, + 0.501, + 0.422 + ], + "angle": 0, + "content": "
MethodDAVIS17 VALMOSE
J&FJFJ&FJF
first 6 layers87.683.791.460.956.865.1
last 6 layers88.084.491.659.155.163.1
all layers88.584.692.762.358.166.6
" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.348, + 0.765, + 0.422 + ], + "angle": 0, + "content": "
MethodDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
DTS-t87.283.690.866.461.271.6
DTS-Lr86.783.190.360.254.965.5
DTS88.885.292.470.164.775.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.434, + 0.785, + 0.462 + ], + "angle": 0, + "content": "Table 4: From left to right: (a) Ablations on different mask embedding adding layers and (b) Ablation of different DTS designs in OneVOS." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.472, + 0.788, + 0.528 + ], + "angle": 0, + "content": "within OneVOS but, when integrated into the network, it bolsters the learning of this paradigm. It contributes to a more efficient OneVOS by adaptively selecting tokens that align with the model's feature extraction and matching mechanisms, culminating in state-of-the-art performance on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.535, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Mask Embedding Add Layers. Table. 4(a) explores the effect of integrating mask embedding at different transformer layers. Results show that, by integrating mask embeddings to the immediate value tokens of each layer, the model can capture the global and local semantics of targets at different levels, thereby effectively promoting the model to perform multiple targets inference. Notably, embedding in the first six layers proved more beneficial than in the latter six, this indicates that early interaction between the current and reference frames without target information may negatively impact the feature extraction of the current frame." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.652, + 0.788, + 0.779 + ], + "angle": 0, + "content": "DTS design. The role of DTS is to judiciously select tokens by amalgamating each token feature of the reference frame with the global features of all tokens. Table 4(b) explores the effect of various DTS configurations, including the method using pooled tokens from the current frame as global information (DTS-t), revealing that guiding selection with the current frame may hinder the generalization of OneVOS across frames, consistent with our initial attention decoupling. Besides, Implementing a loss function to maintain a fixed 0.5 selection rate per layer \\((\\mathrm{DTS} - L_r)\\) significantly reduces performance, underscoring that rigid selection criteria detrimentally affect the capabilities of OneVOS for adaptive matching and feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.784, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dynamic Memory Capacity. Various memory capacities (Cap in Eq. 8) impact accuracy, inference speed, and storage use. The performance of OneVOS under different memory capacities is validated in Table 5(a). The results reveal that while memory plays a crucial role in handling complex videos, OneVOS achieves robust performance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.145, + 0.777, + 0.221 + ], + "angle": 0, + "content": "
CapD17-VD17-TFPSMem(G)
187.583.216.61.72
288.084.413.72.20
388.484.612.92.47
488.584.612.32.69
588.584.812.02.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.231, + 0.787, + 0.288 + ], + "angle": 0, + "content": "Table 5: Ablations on 3 specific module design of All-in-One framework. From left to right: (a) different \\(Cap\\) of dynamic memory, where D17-V and D17-T denote the \\(J\\&F\\) metrics of DAVIS-17 VAL and TEST datasets, respectively. (b) Different top-k tokens for dynamic memory, and (c) Different Transformer backbones of OneVOS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.785, + 0.357 + ], + "angle": 0, + "content": "even with just two frames or a single frame in common scenarios, thanks to its unified framework and efficient architecture. Optimal performance is observed when the memory capacity reaches 4 or 5. This underscores the importance of memory in enabling OneVOS to effectively handle longer and more intricate videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.787, + 0.423 + ], + "angle": 0, + "content": "Number of Retained Nodes. Table. 5(b) focuses on retaining a set number of top-k nodes in token memory, based on attention weights. Results show that after reaching a per-layer threshold, keeping approximately 512 nodes struck an optimal balance between information depth and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Backbone. In Table. 5(c), we assess Vision Transformers (ViT) [15] pre-trained with Masked Autoencoders (MAE) [25] and ConvVIT (base and large variants) as backbones for OneVOS. With the VIT-B backbone OneVOS still outperforms most of the methods on D17 datasets. However, a key limitation of ViT was identified in its direct 16x downsampling approach, which resulted in a significant loss of multi-scale detail information. This is particularly detrimental for high-precision, pixel-level VOS tasks. On the other hand, the ConvVIT model, which employs simple convolutional operations for its patch embedding process, was found to effectively preserve multi-scale information. Additionally, the increased complexity of the ConvVIT-large variant further enhanced the model's performance, underscoring the importance of architectural design in capturing intricate details necessary for accurate VOS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.36, + 0.648 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper introduces OneVOS, a novel semi-supervised video object segmentation framework that globally integrates feature extraction, matching, memory management, and object aggregation within a single Transformer architecture. By employing the All-in-One Transformer with specialized Unidirectional Hybrid Attention modules, we significantly enhance the practicality and performance of OneVOS. Furthermore, to propose a rational, efficient version of OneVOS, we investigate the mechanism of the All-in-One Transformer via the proposed Dynamic Token Selector, revealing its layer-crossing segmentation mode for multi-object feature extraction and matching, leading to a naturally efficient version with dynamic memory. OneVOS demonstrates state-of-the-art performance on 7 datasets, particularly excelling in complex scenarios. We anticipate that this work will draw greater attention to the unified VOS modeling." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.177, + 0.83, + 0.22 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.242, + 0.323, + 0.258 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.786, + 0.328 + ], + "angle": 0, + "content": "1. Bhat, G., Lawin, F.J., Danelljan, M., Robinson, A., Felsberg, M., Van Gool, L., Timofte, R.: Learning what to learn for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 777-794. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.37 + ], + "angle": 0, + "content": "2. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.786, + 0.412 + ], + "angle": 0, + "content": "3. Chen, B., Li, P., Bai, L., Qiao, L., Shen, Q., Li, B., Gan, W., Wu, W., Ouyang, W.: Backbone is all your need: A simplified architecture for visual object tracking. In: European Conference on Computer Vision. pp. 375-392. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.786, + 0.453 + ], + "angle": 0, + "content": "4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.786, + 0.495 + ], + "angle": 0, + "content": "5. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8126-8135 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.786, + 0.536 + ], + "angle": 0, + "content": "6. Chen, Y., Pont-Tuset, J., Montes, A., Van Gool, L.: Blazingly fast video object segmentation with pixel-wise metric learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1189-1198 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.786, + 0.578 + ], + "angle": 0, + "content": "7. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.786, + 0.632 + ], + "angle": 0, + "content": "8. Cheng, H.K., Tai, Y.W., Tang, C.K.: Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5559-5568 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.786, + 0.674 + ], + "angle": 0, + "content": "9. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.675, + 0.786, + 0.716 + ], + "angle": 0, + "content": "10. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.717, + 0.786, + 0.757 + ], + "angle": 0, + "content": "11. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence 37(3), 569-582 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "12. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13608-13618 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "13. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.273, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "16. Duke, B., Ahmed, A., Wolf, C., Aarabi, P., Taylor, G.W.: Sstvos: Sparse spatiotemporal transformers for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5912-5921 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "17. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "18. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "19. Gao, J., Qian, X., Wang, Y., Xiao, T., He, T., Zhang, Z., Fu, Y.: Coarse-to-fine amodal segmentation with shape prior. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1262-1271 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.453 + ], + "angle": 0, + "content": "20. Gao, P., Ma, T., Li, H., Lin, Z., Dai, J., Qiao, Y.: Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "21. Gao, S., Zhou, C., Ma, C., Wang, X., Yuan, J.: Aiatrack: Attention in attention for transformer visual tracking. In: European Conference on Computer Vision. pp. 146-164. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "22. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "23. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.619 + ], + "angle": 0, + "content": "24. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.661 + ], + "angle": 0, + "content": "25. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "26. Hong, L., Chen, W., Liu, Z., Zhang, W., Guo, P., Chen, Z., Zhang, W.: Lvos: A benchmark for long-term video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13480-13492 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.758 + ], + "angle": 0, + "content": "27. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "28. Hong, L., Zhang, W., Chen, L., Zhang, W., Fan, J.: Adaptive selection of reference frames for video object segmentation. IEEE Transactions on Image Processing 31, 1057-1071 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "29. Hu, P., Wang, G., Kong, X., Kuen, J., Tan, Y.P.: Motion-guided cascaded refinement network for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1400-1409 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.245 + ], + "angle": 0, + "content": "31. Huang, X., Xu, J., Tai, Y.W., Tang, C.K.: Fast video object segmentation with temporal aggregation network and dynamic template matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8879-8889 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.273 + ], + "angle": 0, + "content": "32. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "33. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8953-8962 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "34. Khoreva, A., Benenson, R., Ilg, E., Brox, T., Schiele, B.: Lucid data dreaming for video object segmentation. International Journal of Computer Vision 127(9), 1175-1197 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.425 + ], + "angle": 0, + "content": "35. Kristan, M., Leonardis, A., Matas, J., Felsberg, M., Pflugfelder, R., Cehovin Zajc, L., Vojir, T., Bhat, G., Lukezic, A., Eldesokey, A., et al.: The sixth visual object tracking vot2018 challenge results. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "36. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "37. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "38. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.787, + 0.66 + ], + "angle": 0, + "content": "40. Lin, Z., Yang, T., Li, M., Wang, Z., Yuan, C., Jiang, W., Liu, W.: Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1362-1372 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.716 + ], + "angle": 0, + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "42. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "43. Nowozin, S.: Optimal decisions from probabilistic models: the intersection-over-union case. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 548-555 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "44. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "45. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "46. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.285 + ], + "angle": 0, + "content": "47. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.325 + ], + "angle": 0, + "content": "48. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.365 + ], + "angle": 0, + "content": "49. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.787, + 0.407 + ], + "angle": 0, + "content": "50. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 629-645. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.447 + ], + "angle": 0, + "content": "51. Seong, H., Oh, S.W., Lee, J.Y., Lee, S., Lee, S., Kim, E.: Hierarchical memory matching network for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12889-12898 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "52. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.542 + ], + "angle": 0, + "content": "53. Voigtlaender, P., Chai, Y., Schroff, F., Adam, H., Leibe, B., Chen, L.C.: Feelvos: Fast end-to-end embedding learning for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9481-9490 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.569 + ], + "angle": 0, + "content": "54. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "55. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.664 + ], + "angle": 0, + "content": "56. Wang, J., Chen, D., Wu, Z., Luo, C., Tang, C., Dai, X., Zhao, Y., Xie, Y., Yuan, L., Jiang, Y.G.: Look before you match: Instance understanding matters in video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2268-2278 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "57. Wang, W., Shen, J., Porikli, F., Yang, R.: Semi-supervised video object segmentation with super-trajectories. IEEE transactions on pattern analysis and machine intelligence 41(4), 985-998 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.759 + ], + "angle": 0, + "content": "58. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14561-14571 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "59. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13879-13889 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "60. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "OneVOS: All-in-One Transformer Framework For VOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "61. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "62. Xu, S., Liu, D., Bao, L., Liu, W., Zhou, P.: Mhp-vos: Multiple hypotheses propagation for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 314-323 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "63. Yan, S., Xu, X., Hong, L., Chen, W., Zhang, W., Zhang, W.: Panovos: Bridging non-panoramic and panoramic views with transformer for video segmentation. arXiv preprint arXiv:2309.12303 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "64. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.384 + ], + "angle": 0, + "content": "65. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V. pp. 332-348. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "66. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "67. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4701-4712 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.508 + ], + "angle": 0, + "content": "68. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "69. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "70. Zhou, X., Guo, P., Hong, L., Li, J., Zhang, W., Ge, W., Zhang, W.: Reading relevant feature from global representation memory for visual object tracking. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_origin.pdf b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cea7571b1a35e6199e431cf3e9ed03c3fbd5375d --- /dev/null +++ b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/746728e7-d87a-494d-8a5a-df2168951889_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a8592938cabb3865701b3dacb584a340fe2dfb44bf4210bf50729ab0c91e70c +size 1679886 diff --git a/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/full.md b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4e3c3abc5ed61d4377e4d14d35f7b32fde4cf9f --- /dev/null +++ b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/full.md @@ -0,0 +1,332 @@ +# OneVOS: Unifying Video Object Segmentation with All-in-One Transformer Framework + +Wanyun Li $^{1,*}$ , Pinxue Guo $^{2,*}$ , Xinyu Zhou $^{1}$ , Lingyi Hong $^{1}$ , Yangji He $^{1}$ , Xiangyu Zheng $^{1}$ , Wei Zhang $^{1,\dagger}$ , and Wenqiang Zhang $^{1,3,\dagger}$ + +$^{1}$ Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai, China + +$^{2}$ Shanghai Engineering Research Center of AI & Robotics, Academy for Engineering & Technology, Fudan University, Shanghai, China + $^{3}$ Engineering Research Center of AI & Robotics, Ministry of Education, Academy for Engineering & Technology, Fudan University, Shanghai, China {wyli22,pxguo21}@m.fudan.edu.cn, {weizh, wqzhang}@fudan.edu.cn + +Abstract. Contemporary Video Object Segmentation (VOS) approaches typically consist of feature extraction, matching, memory management, and multiple objects aggregation. Recent advanced models either employ a discrete modeling for these components in a sequential manner, or optimize a combined pipeline through substructure aggregation. However, these existing explicit staged approaches prevent the VOS framework from being optimized as a unified whole, leading to the limited capacity and suboptimal performance in tackling complex videos. In this paper, we propose OneVOS, a novel framework that unifies the core components of VOS with All-in-One Transformer. Specifically, to unify all aforementioned modules into a vision transformer, we model all the features of frames, masks and memory for multiple objects as transformer tokens, and integrally accomplish feature extraction, matching and memory management of multiple objects through the flexible attention mechanism. Furthermore, a Unidirectional Hybrid Attention is proposed through a double decoupling of the original attention operation, to rectify semantic errors and ambiguities of stored tokens in OneVOS framework. Finally, to alleviate the storage burden and expedite inference, we propose the Dynamic Token Selector, which unveils the working mechanism of OneVOS and naturally leads to a more efficient version of OneVOS. Extensive experiments demonstrate the superiority of OneVOS, achieving state-of-the-art performance across 7 datasets, particularly exceeding in complex LVOS and MOSE datasets with $70.1\%$ and $66.4\%$ $J\& F$ scores, surpassing previous state-of-the-art methods by $4.2\%$ and $7.0\%$ , respectively. Code is available at: https://github.com/L599wy/OneVOS. + +# 1 Introduction + +Semi-supervised Video Object Segmentation (VOS) [2, 22, 30, 37, 44, 45, 63, 64] is a fundamental task in video analysis [18, 19], aiming to consistently seg- + +![](images/f3746337afe18d228069598e659319365bafa067089b36f3435bd0f27c82862e.jpg) + +![](images/2598b97b547d3b20c53e025f402918a65e127eeb22d20b3607133be2fac464b1.jpg) + +![](images/0f75c54cb4d79e617056685e3ed2911dacd00b82dc7bb8a84818ce82d4b161a7.jpg) +(b) Unify Matching and Objects Aggregation + +![](images/cf56c97d1d60892b6b2d8633288ffba70e01d8090a44b0ab8376c379b3239f43.jpg) +(d) Integrate Core Modules Al in One Transformer +Fig. 1: Four different modeling for semi-supervised video object segmentation pipelines. (a) Traditional discrete modeling of VOS. (b) Unified multiple object modeling in one reference. (c) Unified single-object feature extraction and matching. (d) OneVOS: Integrates all core modules with All-in-One Transformer. + +ment objects throughout a video based on initial masks. Advancements in VOS primarily focus on the full utilization of available information. Among them, Memory-based methods [7,9,40,45] have emerged as the mainstream due to the effectiveness in storing and leveraging abundant spatio-temporal features. Its foundational paradigm, as depicted in Fig. 1(a) [16,40,51,55], encompasses four split key stages including feature extraction, matching, memory management, and multi-object aggregation, which have been widely adopted as a baseline in a number of methods [23,28,50,56]. + +Recently, some investigations demonstrate the advantages of integration modeling compared with highly discrete modeling. AOT [66] and DEAOT [68] (Fig. 1(b)) emphasize the shortage of single-object inference compared to joint modeling of multiple objects, particularly in exploring contextual and robust feature representations. Meanwhile, SimVOS [59] (Fig. 1(c)) reveals limitations in the separation of feature extraction and matching processes for capturing dynamic object-aware features. However, existing methods solely focus on combining or aggregating substructures of the VOS pipeline, resulting in their suboptimal solutions. For example, AOT relies on sophisticated designing matching modules which fails to leverage superior pre-training weights, while the simplistic single-object framework of SimVOS encounters inefficiencies and the inapplicability of memory makes it unable to utilize rich features of the processed frames. + +Therefore, we introduce an essentially novel framework distinct from previous methods, aiming to develop a unified end-to-end differentiable VOS model for global optimization. Based on this, the All-in-One Transformer for Video Object Segmentation (OneVOS) is proposed, which models and unifies all important modules into a Vision Transformer. Specifically, we first leverage the flexible attention mechanism to conduct feature extraction and mutual matching between the reference and current frames. To effectively capture multi-object relationships, we introduce a mask embedding layer that enriches features with multi-object modeling insights. Moreover, in the memory design, we propose an external layer-wise token memory to enhance the coherence understanding of videos by storing and providing spatio-temporal features for subsequent frames. + +As the feature extraction and matching of multiple objects performing simultaneously, the reference tokens output from each layer of Transformer for memory update could be represented by a weighted linear combination of tokens from memory, reference frame and current frame, which may lead to a significant semantic errors and ambiguity to serving as features stored in memory. To address this, we further propose a Unidirectional Hybrid Attention (Uni-Hybrid Attention) through a double decoupling of the original attention operation. It guarantees the versatility of stored features and enhances memory performance. Notably, as the Uni-Hybrid Attention is derived from naive attention mechanism, our framework is still able to utilize pre-trained weights from various Vision Transformers, thereby enhancing its generality and flexibility. + +In this way, the proposed OneVOS framework implements end-to-end modeling and global optimization, exhibiting superior performance compared to previous models. To further strengthen its advantages, we conduct an in-depth exploration of the behavior within the attention layer of OneVOS to enhance its efficiency, which is particularly crucial in resource-constrained and long-term video tasks. Concretely, we introduce the Dynamic Token Selector (DTS), capable of dynamically selecting partial tokens from each layer for cross-attention and memory storage, trained solely under segmentation loss without additional constraints. With DTS integrated into every layer of the transformer, we analyze the token selection distribution of DTS and the attention weight distribution, uncovering a significant insight: unlike the traditional sequential paradigm, OneVOS adopts a cross-layer segmentation mode that dynamically shifts focus between feature extraction and matching across different layers. Based on this, we propose a natural acceleration scheme involving designing distinct memory capacities for each layer and layer-wise memory updates via tokens selected by DTS. This approach enhances both speed and accuracy across multiple datasets. + +We conducted comprehensive experiments across seven datasets, and OneVOS consistently achieves state-of-the-art performance across all of them. Notably, OneVOS demonstrates a substantial performance advantage enven in more challenging scenarios, achieving $71.1\%$ $\mathcal{I}\& \mathcal{F}$ on LVOS validation and $67.2\%$ $\mathcal{I}\& \mathcal{F}$ on MOSE validation. The main contributions are summarized as follows: + +- We propose All-in-One Transformer framework for Video Object Segmentation (OneVOS), integrating feature extraction, matching, memory management, and objects aggregation into a vision transformer, which can be optimized globally. This provides a fundamentally novel paradigm in VOS. +- We introduce the Unidirectional Hybrid Attention mechanism to facilitate multi-object feature extraction and matching, ensuring tokens are devoid of semantic ambiguity for efficient memory storage. +- We propose the Dynamic Tokens Selector, which not only unveils the internal working mechanism but also leads to a more efficient OneVOS, based on the exploration and the resulting insight. +- OneVOS achieves state-of-the-art performance across 7 datasets, particularly excelling in more complex LVOS and MOSE datasets with $70.1\%$ and $66.4\%$ $J\& F$ , surpassing previous best methods by $4.2\%$ and $7.0\%$ , respectively. + +# 2 Related work + +# 2.1 Video object segmentation + +Semi-supervised VOS methods can be divided into four types: The earlier fine-tuning methods [1, 2, 42, 54, 60] learn general segmentation features via offline pre-training, then enhance object-specific representations by online fine-tuning. Then, propagation-based methods [10, 29, 33, 44, 46, 57] utilize optical flow to propagate the mask from the previous frame to the next one. Subsequent methods optimized it by keyframes enhancement [34], advanced tracking techniques [4, 10, 31, 62], and re-identification mechanisms [38]. Matching-based methods [6, 30, 53, 65, 67] identify the object of the current frame based on the correlation with the reference frame in the embedding space. + +Recently, Memory-based approaches [23,28,45,50] store features of segmented frames in a memory bank to provide a set of features for segmentation, which has become the dominant method in SVOS. Following that, methods such as HMMN [51], SSTVOS [16], and XMem [7] enhance performance by utilizing multi-scale features, introducing Transformer for matching, and building long-term memory. These methods typically adhere to a traditional discrete modeling and single-object architecture. More recently, the multi-object framework [66], [68] is proposed to address the inherent limitations of the previous single-object while still maintaining a sophisticated matching module for high performance. [59] presents a simplified VOS framework for joint feature extraction and matching which adopts single-object modeling and discards the vital memory design in VOS. Differently, our OneVOS, innovatively unifies feature extraction, matching, memory construction, and multi-objective aggregation all in one Transformer. This design not only streamlines the segmentation process but also enhances the dynamic interaction among different stages of SVOS. + +# 2.2 Visual Object Tracking + +Video object tracking (VOT) [3,27,35,70] shares similarities with the task setting of SVOS, but it primarily utilizes bounding boxes for object localization instead of detailed segmentation masks. In VOT, there has been a paradigm shift from traditional Siamese-based methods towards leveraging Transformer layers for enriched relational modeling between the template and search regions. Pioneering works like TransT [5] introduce a two-stream attention mechanism for robust bi-directional information exchange, while innovative methods like MixFormer [12] and AiATrack [21] adopt asymmetric attention strategies to minimize distractions from irrelevant objects. Notably, OSTrack [69] proposes a unified attention model, promoting seamless feature integration and relation modeling. Nevertheless, the distinct multi-object settings and difference in target information representation, coupled with the critical role of memory in VOS, render unification within the VOS field more challenging. This paper endeavors to tackle this challenge. + +# 3 Method + +As shown in the Fig. 2, in a given video sequence, VOS focuses on segmenting the current $t$ -th frame $f_{t} \in \mathbb{R}^{H \times W \times 3}$ using the reference frame $f_{t-1} \in \mathbb{R}^{H \times W \times 3}$ with its mask and the history features stored in Memory. In this section, we aim to construct an unified framework at a global scale to tackle the challenges of VOS. We commence by introducing the OneVOS framework in Sec. 3.1, which employs Vision Transformer to integrate core elements in VOS. And then Sec. 3.2 provides a detailed exposition on the novel Unidirectional Hybrid Attention mechanism. In addition, in Sec. 3.3, we explore the working mechanism of attention within OneVOS through the proposed Dynamic Token Selector in order to introduce the efficient version of OneVOS. + +![](images/a9da2db1d58dedbded6f8e879caf3feeb187f20792ccb416c101db48fed8b752.jpg) +(a) Hybrid Attention Block + +![](images/dbe76f2aa51d65433abdb0e0adec0aaaff1b6c919531c8ef84a566e1c1b9b0e3.jpg) +(b) Block with Dynamic Token Selector (DTS) + +![](images/35e1eab2a4cb32f940bfe4f581ec8d0c98b7ed15375bf18eb82fb696bdef8076.jpg) +(c) Overview +Fig. 2: (a) Hybrid Attention Block. (b) Blocks with Dynamic Token Selector. (c) Overview of OneVOS framework equipped with All-in-One Transformer. The reference mask is initially encoded into mask embeddings. Patch embeddings are generated from both the reference and current frames, which are then processed by the Transformer. Within each Transformer layer, mask embeddings are integrated with reference tokens. Then they are selected by Dynamic Token Selector and concatenated with tokens stored in memory and current frame fed into the Hybrid Attention Module. Finally, the features of the current frame output from Transformer are subsequently decoded into multi-object masks. + +# 3.1 OneVOS Framework + +Mask Embedding. To capture complex target interactions and minimize redundant single-target inferences, motivated by [66], we utilize the Mask Embedding Layer (as Patch Embedding Layer in Transformer) to encode reference + +mask into Mask Embedding denoted as $m_{ref} \in \mathbb{R}^{\frac{H}{P} \times \frac{W}{P} \times C}$ through a single convolutional layer. Here, $C$ represents the embedding dimension, with $P$ as the stride of the mask embedding. Utilizing the local relationship modeling capability of the convolution operation, the discriminability of multiple targets is enhanced, which will then be integrated with visual feature map in Transformer before attention computation to integrating mask embeddings to the immediate value tokens of each layer, thereby effectively promoting the model to perform multiple targets inference. + +All-in-One Transformer. Then, we adopt a transformer architecture to unify the core modules of VOS. It first generates the non-overlapping patch embeddings $e_t^0 \in \mathbb{R}^{N \times C}$ and $e_{ref}^0 \in \mathbb{R}^{N \times C}$ for the current and reference frames, where $P$ is the patch resolution same as the mask embedding, thus $N = HW / P^2$ . Then, the concatenation of $[e_{ref}^0; e_t^0]$ along the spatial dimension serves as the initial input for the backbone, facilitating adaptive multi-object feature extraction and target information propagation through the stacked layers. In each layer, we first integrate mask embedding with visual feature map as shown in Eq 1: + +$$ +\left\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} \right\} = \left\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} + m _ {r e f} \right\}, \tag {1} +$$ + +where $Q$ , $K$ , and $V$ are query, key, and value matrices aligning with the standard attention mechanism [15], and $l \in \{0, 1, \dots, L\}$ represents the layer. Then, these object-aware reference tokens, memory tokens, and tokens of current frames undergo adaptive feature extraction and target information propagation by applying the attention mechanism of the Transformer. The output $a^l = [a_{ref}^l; a_t^l]$ of layer $l$ after the attention operation is as follows: + +$$ +a ^ {l} = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) \cdot V = \operatorname {S o f t m a x} \left(\frac {\left[ Q _ {r e f} ^ {l} ; Q _ {t} ^ {l} \right] \left[ K _ {M} ^ {l} ; K _ {r e f} ^ {l} ; K _ {t} ^ {l} \right]}{\sqrt {d _ {k}}}\right) \cdot \left[ V _ {M} ^ {l}; V _ {r e f} ^ {l}; V _ {t} ^ {l} \right], \quad (2) +$$ + +where $K_M^l, V_M^l$ are features stored in memory. Supposing $W = QK^{T}$ is the matrix of similarity weights, $a_{t}^{l}$ can be written as: + +$$ +a _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}, \tag {3} +$$ + +where $[W_{t,ref}V_{ref}]$ and $[W_{t,M}V_M]$ are responsible for aggregating the inter-frame features from memory and reference frame to current frame (matching), and $[W_{t,t}V_t]$ aggregating the intra-frame features (feature extraction) for current frame, respectively. Thus, dynamic coordination between feature extraction and matching is achieved through the attention weights learned adaptively. + +Token Memory. In addition, since the target information and the reference frame are integrated within our framework, we can store intermediate features directly into Token Memory without necessitating additional post-processing fusion steps. Given the semantic variance of intermediate tokens in the reference frame across layers, we introduce a token storage mechanism to save essential Transformer tokens instead of whole frames as follows: + +$$ +M = \left\{K _ {M} ^ {l}, V _ {M} ^ {l} \mid \in \{0, 1, \dots L \} \right\} = \left\{K _ {\text {r e f s}} ^ {l}, V _ {\text {r e f s}} ^ {l} \mid \in \{0, 1, \dots L \} \right\}, \tag {4} +$$ + +This method allows these tokens to be involved only in the attention modules instead of multiple inferences, greatly reducing inference overhead. + +![](images/395d5876de5b3d53d08a5979d97bf55681b9cc3befe6001eb255259c4857b0f1.jpg) +Fig. 3: (a) Visualization of attention distribution of target areas in OneVOS, showing the alternation between self-attention and cross-attention across layers. (b) Analysis of the maximal attention weights reveals a pattern of alternation between feature extraction via self-attention (notably in layers 3 and 5) and matching via cross-attention (as in layers 2 and 8). (c) The proportional analysis of the attention weights' sum, normalized to one, underscores the non-sequential prioritization across different layers between self-attention and cross-attention. + +![](images/7af79e823c1599a2fa52ddfdb957dc04831c27f423e3b0a230ac5a4e07a85ec2.jpg) + +![](images/c3284748381c29e9423cff0ade4ca7de693bca5519c9ca89200d9015b3645517.jpg) + +Mask Decoding. Finally, the output of All-in-One Transformer, $e_t^L$ is input to the lightweight Feature Pyramid Network (FPN) for Mask Decoding. These features are then gradually upgraded to the original scale and decoded into multi-object masks. + +# 3.2 Unidirectional Hybrid Attention + +Building upon our newly developed pipeline, all modules are integrated into a transformer as in Fig. 2(c). However, obtaining the stored features by employing original attention mechanism may lead to semantic errors and ambiguity since they are represented incorrectly. To address this challenge, our proposed Unidirectional Hybrid Attention (Uni-Hybrid Attention) is strategically designed through dual-level attention decoupling as shown in Fig. 2(a). + +Firstly, reference features calculated by naive attention operation to all the tokens are denoted as $a_{ref}^{l} = W_{ref,M}V_{M} + W_{ref,ref}V_{ref} + W_{refs,t}V_{t}$ . A key limitation in the computations arises where tokens of reference frame containing target information (as in Eq. (1)) are represented by tokens of current frame that without target information, causing semantic errors in the features stored in the memory which are essential for storing spatio-temporal target information for subsequent frames. Therefore, to alleviate this situation, our first modification decoupling attention from the current token to the reference token. The revised expression is: + +$$ +a _ {r e f} ^ {l} = W _ {r e f, M} V _ {M} + W _ {r e f, r e f} V _ {r e f}. \tag {5} +$$ + +Subsequently, the attention computation between reference features and memory tokens is decoupled, serving a dual purpose: 1) Avoiding Information Overlap and Ambiguity: Memory tokens, enriched with mask embeddings, encapsulate target information from prior frames. Integrating these embeddings with those of the reference frame risks overlapping and confusing object cues. 2) Preserving Temporal Specificity: Every video frame contains essential temporal intricacies crucial for segmentation precision. Attention decoupling preserves the reference frame's temporal uniqueness, critical for accurate and context-aware segmentation. Therefore, we decouple attention between memory tokens and those from the reference frame, and the final unidirectional hybrid attention formulations are: + +$$ +a _ {r e f} ^ {l} = W _ {r e f, r e f} V _ {r e f} \tag {6} +$$ + +$$ +a _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}. +$$ + +![](images/1a6fbaeba53cff3baf1fb75b5821301ee718e1254ca7a8df13b2b9954fd8b878.jpg) +(a) + +![](images/cc9e9bbfc14e091f1395a056962e2a9c36b2365f7d8b6ec68b98362c77ac418c.jpg) +Fig. 4: (a) Visualization of the selected tokens of reference frame by DTS in different Transformer layers. (b) Adaptive selection ratios learned by DTS in different layers without any constraints. + +![](images/c5b83b667266c087cdaca6abda1d6bfc44ea014f069fcbf754d72e26c6201b4a.jpg) +(b) + +# 3.3 Towards Efficient OneVOS + +Utilizing the advanced All-in-One Transformer equipped with Uni-Hybrid Attention, OneVOS achieves outstanding performance across various datasets. However, the integration process of these modules in OneVOS still lacks explainability, hindering the rational design of a more efficient version. Therefore, this section delves into its working mechanism to propose the efficient OneVOS. + +Exploration and Insight. Within the Transformer architecture, attention weights determine the influence of each token in the input sequence on the output which is useful for interpreting Transformer behavior. By examining the attention weights of tokens in the target area of the current frame across each layer, we compare their focus in reference frames or themselves. As presented in Fig. 3, qualitative results in 3(a) shown that a fluctuating focus between reference frames (cross-attention) and themselves (self-attention), Subsequently, Fig. 3 (b) and (c) provide quantitative evidence of a complex alternation pattern, which deviates from the sequential feature extraction and matching process of traditional discrete VOS modeling or uniform distribution in each layer. Instead, these findings lead us to a natural insight: OneVOS may learn a dynamic layer-wise integration strategy, preferentially executing multi-object feature extraction or matching in each layer according to the semantic context. + +Dynamic Token Selector. Based on above insights and motivated by [49], we introduce a Dynamic Token Selector (DTS) into the OneVOS architecture, as depicted in Fig. 2(b). This module adaptively selects a subset of tokens from the reference frame for cross-attention and storage. The selection process, as defined in Eq. 7, combines global context and local token information: + +$$ +\begin{array}{l} p = \operatorname {S o f t m a x} (\operatorname {M L P} \left(\left[ \operatorname {M a x P o o l} \left(z _ {r e f} ^ {l}\right); z _ {r e f} ^ {l}\right]\right)) \\ z _ {r e f} ^ {l} = \operatorname {M L P} \left(\operatorname {N o r m} \left(e _ {r e f} ^ {l}\right)\right) \tag {7} \\ \end{array} +$$ + +where $z_{ref}^{l}$ denotes the local feature encoded of each token, while [MaxPool( $z_{ref}^{l}$ )] aggregates the overall context of the reference frame. $p \in \mathbb{R}^{N \times 2}$ represents the probabilities associated with each token. Specifically, $p_{i,0}$ indicates the probability of dropping $i$ -th token while $p_{i,1}$ indicates for selection. + +This design aims to enable the network to autonomously determine the optimal token selection ratio for high-performance segmentation. Consequently, it is integrated into an All-in-One Transformer, as depicted in Fig. 2(a). Moreover, it is exclusively trained under segmentation loss without additional constraints. By observing the proportion of tokens selected by DTS in each transformer layer, as illustrated in Fig. 4(a) and Fig. 4(b), a crucial insight emerges. The model automatically allocates varying + +![](images/66175493c3afb3bd5e120b006f4ecb74974800cb63a048ace280568936aba7e6.jpg) +Fig. 5: Visualization dynamic memory updates of OneVOS utilizing DTS. The selected tokens in red boxes are derived output from DTS as Fig. 4(a). + +selection ratios to different layers, and even assigns a selection ratio of zero or one to specific layers. This observation underscores the focus of DTS on certain layers for feature extraction (selecting a proportion of reference frame tokens as zero) and on specific layers for matching (selecting a proportion of reference frame tokens as one), presenting a characteristic of alternating focus across layers. + +Efficient Version of One VOS. Building upon the above observations, a straightforward acceleration method is naturally introduced. This includes layer-wise memory capacity and adaptive token update mechanism via DTS, as shown in Fig. 5. Specifically, we initialize different capacities for each layer of Token Memory, with the token capacity formula as follows: + +$$ +c _ {l} = r _ {l} * N * C a p. \tag {8} +$$ + +Here, $r_l$ represents the layer-specific selection ratio determined by DTS during training, as illustrated in Fig. 4(b). Additionally, $Cap$ is a hyper-parameter denoting the frame capacity (typically set to 3 or 5). Consequently, tokens are incrementally added to the memory based on these selection ratios until reaching the capacity limit of each layer. Subsequently, the top-k tokens, determined by the current attention weights $W_{t,ref}$ , are retained, followed by the continuous addition of tokens to memory. Aligned with the segmentation mode of OneVOS, this method achieves improvements in accuracy and speed across multiple datasets. + +# 4 Experiment + +# 4.1 Implementation Details + +One-stream VOS uses ConvVIT-base [20] as the backbone, initialized with ImageNet-1k [13] pre-training weights. The feature resolution of the mask decoding is increased from 1/16 to $1/4$ and the channel dimension is reduced from 768 to 384. + +Training details. Following the common settings of VOS [8,66], the training phase is divided into two stages: pre-training on synthetic video sequences generated from still image datasets [11,17,24,39,52], and main training on standard VOS dataset (DAVIS 17 and YouTube VOS in default, the model mark with "*" uses MOSE in main training as well). The loss function is a 0.5:0.5 combination of bootstrapped cross-entropy loss and soft Jaccard loss [43]. The learning rate of two stages is 2e-4 and 1e-4, respectively, utilizing $4 \times 3090$ s GPUs. Our batchsize is 4, with a gradient accumulation of 2 and a sequence length of 5, running each training phase for 200,000 iterations. In the training of DTS, we employ the Gumbel-Softmax [32] on its output $p$ to generate a one-hot selection vector, indicating whether each token is selected (assigned a value of 1) or + +
Long-term VideosComplicated Videos
MethodLVOS VALLVOS TESTMethodMOSEJ&FJF
J&FJFJ&FJF
CFBI [65]50.045.055.144.840.549.0XMem [7]x56.452.160.6
AOT-L [66]61.359.453.654.148.759.5DropSeg [58]x52.248.056.4
STCN [9]45.841.150.545.841.650.0R50-AOT-L [66]x58.153.962.3
RDE [36]52.947.758.149.044.453.5R50-DeAOT-L [68]x58.954.563.3
XMem [7]50.045.554.449.545.253.7SwinB-AOT-L [66]x60.056.164.0
AOT-B [66]56.951.861.954.449.359.4OneVOSx62.358.166.6
DeAOT [68]63.757.669.858.152.963.4RDE [36]✓48.844.652.9
R50-AOT-L [66]63.858.169.661.056.165.9STCN [9]✓50.846.655.0
R50-DeAOT-L [68]65.959.971.958.053.262.8XMem [7]✓57.653.362.0
Swin-AOT-L [66]65.660.875.161.556.474.8DeAOT [68]✓59.455.163.8
OneVOS-D70.164.775.560.155.668.0OneVOS-D✓61.757.566.0
OneVOS67.462.072.963.859.570.0OneVOS✓66.462.270.6
OneVOS*71.166.076.264.160.068.2OneVOS*✓67.263.071.5
+ +Table 1: Quantitative comparison of long-duration video datasets (LVOS validation and test) and complex datasets, MOSE. $\checkmark$ indicates MOSE used for training. OneVOS( $\checkmark$ ) only employs MOSE in main training, whereas OneVOS\* $(\checkmark)$ utilizes DAVIS, YouTube, and MOSE. + +![](images/4e4bfc47311d1b17d3f11651efb96d034901e79d348c89b6fd74091428fb2ffb.jpg) +Fig. 6: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging short-term videos (MOSE and DAVIS 17). + +not (value of 0). Throughout training, we record the selection rate, observing that in the later stages of training, the selection rate of each layer gradually stabilizes. + +Inference details. OneVOS utilizes basic memory storage, storing tokens at intervals of 5 or 10 frames and updates using the first-in-first-out queue when it is full, except for the initial frame tokens. Conversely, OneVOS-D employs dynamic memory, defaulting to retain a maximum of 520 top-k tokens when full. The maximum frame-level capacity $Cap$ varies by dataset, set to $3/5/8$ . For fair comparison in LVOS, we strictly limit memory to under 6. + +Datasets and Metrics. We evaluate the performance of OneVOS on 4 standard VOS datasets, including DAVIS 16 val [47], DAVIS 17 val [48], DAVIS 17 test [48], and YouTube 2019 [61], as well as 3 more challenging datasets, MOSE [14], LVOS val, and LVOS test [26]. The evaluation metrics include region similarity $\mathcal{I}$ , boundary accuracy $\mathcal{F}$ , and their average $\mathcal{J} \& \mathcal{F}$ , and FPS for speed. For YouTube-VOS, we also report them from the seen $\mathcal{J}_s$ , $\mathcal{F}_s$ and unseen $\mathcal{J}_u$ , $\mathcal{F}_u$ categories. + +# 4.2 Comparison with State-of-the-art Methods + +Complicated datasets. Table 1 tabulates the quantitative results on the LVOS val, LVOS test, and MOSE. OneVOS demonstrates superior performance across multi- + +
MethodDAVIS 16 VALDAVIS 17 VALDAVIS 17 TESTYOUTUBE 19 VAL
J&FJFJ&FJFFPSJ&FJFGJsFsJuFu
CFBI+ [67]89.988.791.182.980.185.75.678.074.481.682.681.786.277.185.2
HMMN [51]90.889.692.084.781.987.55.078.674.782.582.581.786.177.385.0
DropSeg [58]92.190.993.386.583.589.5----83.482.987.377.785.6
STCN [9]91.690.892.585.482.288.620.276.172.779.682.781.185.478.285.9
XMem [7]91.590.492.786.282.989.522.681.077.484.585.584.388.680.388.6
Swin-AOTL [66]92.090.793.385.482.488.412.181.277.385.184.584.088.878.486.7
Swin-DeAOTL [68]92.991.194.786.283.189.215.482.878.986.786.185.390.280.488.6
ISVOS [56]92.691.593.787.183.790.5-82.879.386.286.185.289.780.788.9
MiVOS' [8]91.089.692.484.581.787.411.278.674.982.282.480.684.778.186.4
STCN' [9]91.790.493.085.382.088.620.277.874.381.384.282.687.079.487.7
XMem' [7]92.090.793.287.784.091.422.681.277.684.785.884.889.280.388.8
SimVOS [59]92.991.394.488.085.091.04.980.476.184.684.283.1-79.1-
OneVOS92.791.094.388.584.692.412.084.880.788.986.184.989.880.689.1
OneVOS-D93.191.494.888.885.292.416.083.980.087.685.084.589.279.187.3
OneVOS*92.891.093.788.384.492.112.085.381.389.386.385.089.981.289.1
+ +Table 2: Quantitative comparisons on the DAVIS 2016 val, DAVIS 2017 val, and YouTube-VOS 2019 val split. The methods Use ' as a superscript indicates adding the BL30k [8] dataset for training. + +ple settings and datasets. Notably, OneVOS-D excels in the long video dataset LVOS, achieving an impressive $70.1\%$ $J\& F$ score under identical training conditions, demonstrating enhanced robustness in long-term videos. On the MOSE dataset, results of OneVOS are particularly noteworthy, surpassing the highest results reported in the paper by nearly $7\%$ (66.4% vs 59.4%) trained in the same setting. When trained on YouTube, DAVIS, and MOSE together in the main training, OneVOS* achieves state-of-the-art performance across all three datasets. + +Common datasets. Table 2 tabulates the quantitative results on the four standard datasets. The results clearly highlight the superiority of OneVOS. OneVOS-D achieves the best $93.1\%$ in DAVIS 2016 val and $88.8\%$ in DAVIS 2017 val respectively, setting new state-of-the-art performance. OneVOS and OneVOS* showcase remark improvement in DAVIS 2017 test and YouTube-VOS 2019 val. This proves that OneVOS has excellent segmentation capabilities in both regular and challenging scenarios, demonstrating its practicability. + +Qualitative Results. We present visualization results of OneVOS comparing against the state-of-the-art method XMem [7] and Swin-AOTL [36]. The results for the short-term datasets and complicated long-term videos are shown in Fig. 6 and Fig. 7, respectively. OneVOS consistently achieves finer and higher-quality segmentation on the short-term datasets. Notably, it showcases superior segmentation capabilities in long-term scenarios involving both background interference and small targets. + +# 4.3 Ablation Study + +In this section, we conduct a comprehensive analysis of the OneVOS framework and the efficacy of its core designs on both common and the more challenging datasets. + +Table 3 details an incremental ablation study highlighting the significance of key design elements in OneVOS. Starting with a baseline where ConvVIT replaces the backbone in a top discrete multi-model framework [66] (line 1). Furthermore, to negate any bias stemming from varying modeling approaches toward the backbone, we employ an optimal model based on Swin Transformer [41] as an advanced baseline (line 2), providing a stringent benchmark to validate the effectiveness of OneVOS. + +![](images/f198866c59fc4b761c44b125f64729abd57f731ea2b2bb4f66a3bec8cbed0f5a.jpg) +Fig. 7: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging long-term videos (LVOS). + +All-in-One Modeling. We initially evaluate the efficacy of All-in-One Modeling by analyzing lines 1, 2, and 3 of Table 3. The outcomes highlight the substantial benefits of All-in-One Modeling. Our model registers a $2.45\%$ improvement in performance on DAVIS 17 over the enhanced baseline. Moreover, in long video datasets like LVOS, even without any specialized module designs and using only the basic self-attention mechanism, our model exhibits only a minor decrease in performance, demonstrating robust performance of Unification modeling on complex long video datasets. + +Unidirectional Hybrid Attention. In our unidirectional hybrid attention module analysis, shown in lines 3 to 6 of Table 3. Notably, through an examination of lines 4 and 6 we observed that the integration of dual decoupling significantly enhances model performance on both DAVIS 17 val and LVOS datasets. However, Sole decoupling from the current to reference frame results in a performance drop. We analyze that the attributed to the increased focus of reference frame on memory tokens, thereby amplifying the ambiguity in mask embedding and hindering accurate target identification. On the other hand, decoupling solely from memory tokens to reference frame yields a marginal performance enhancement in DAVIS 17 val. However, this improvement is limited as the target features with objects information are represented by tokens from current frame without targets information, failing to provide a consistent feature representation across the video sequence so the performance drops in long videos. These findings highlight the necessity of dual decoupling is indispensable, and their joint effect promotes optimal model performance. + +Dynamic Tokens Selector. The comparison results between lines 6 and 7 of Table 3 highlights the role of DTS. DTS not only probes the segmentation mechanisms + +
MethodAll-in-OneDecoup1Decoup2DTSDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
182.879.785.960.856.465.1
285.482.488.465.660.875.1
3✓87.584.390.764.960.369.5
4✓✓87.184.090.161.055.766.2
5✓✓87.684.490.862.357.167.5
6✓✓✓88.584.692.467.462.072.9
7✓✓✓✓88.885.292.470.164.775.5
+ +Table 3: Ablation on Core Design of OneVOS. (1-2) baseline: multi-object discrete method with ConvVIT or Swin Transformer [41] as backbone. (3-7): OneVOS with ConvVIT: 'All-in-One': All-in-One modeling. 'Decoup1': decouple attention from current to reference, 'Decoup2': decouple attention between memory and reference. 'DTS': DTS with dynamic memory. + +
MethodDAVIS17 VALMOSE
J&FJFJ&FJF
first 6 layers87.683.791.460.956.865.1
last 6 layers88.084.491.659.155.163.1
all layers88.584.692.762.358.166.6
+ +
MethodDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
DTS-t87.283.690.866.461.271.6
DTS-Lr86.783.190.360.254.965.5
DTS88.885.292.470.164.775.5
+ +Table 4: From left to right: (a) Ablations on different mask embedding adding layers and (b) Ablation of different DTS designs in OneVOS. + +within OneVOS but, when integrated into the network, it bolsters the learning of this paradigm. It contributes to a more efficient OneVOS by adaptively selecting tokens that align with the model's feature extraction and matching mechanisms, culminating in state-of-the-art performance on both datasets. + +Mask Embedding Add Layers. Table. 4(a) explores the effect of integrating mask embedding at different transformer layers. Results show that, by integrating mask embeddings to the immediate value tokens of each layer, the model can capture the global and local semantics of targets at different levels, thereby effectively promoting the model to perform multiple targets inference. Notably, embedding in the first six layers proved more beneficial than in the latter six, this indicates that early interaction between the current and reference frames without target information may negatively impact the feature extraction of the current frame. + +DTS design. The role of DTS is to judiciously select tokens by amalgamating each token feature of the reference frame with the global features of all tokens. Table 4(b) explores the effect of various DTS configurations, including the method using pooled tokens from the current frame as global information (DTS-t), revealing that guiding selection with the current frame may hinder the generalization of OneVOS across frames, consistent with our initial attention decoupling. Besides, Implementing a loss function to maintain a fixed 0.5 selection rate per layer $(\mathrm{DTS} - L_r)$ significantly reduces performance, underscoring that rigid selection criteria detrimentally affect the capabilities of OneVOS for adaptive matching and feature extraction. + +Dynamic Memory Capacity. Various memory capacities (Cap in Eq. 8) impact accuracy, inference speed, and storage use. The performance of OneVOS under different memory capacities is validated in Table 5(a). The results reveal that while memory plays a crucial role in handling complex videos, OneVOS achieves robust performance + +
CapD17-VD17-TFPSMem(G)
187.583.216.61.72
288.084.413.72.20
388.484.612.92.47
488.584.612.32.69
588.584.812.02.94
+ +Table 5: Ablations on 3 specific module design of All-in-One framework. From left to right: (a) different $Cap$ of dynamic memory, where D17-V and D17-T denote the $J\&F$ metrics of DAVIS-17 VAL and TEST datasets, respectively. (b) Different top-k tokens for dynamic memory, and (c) Different Transformer backbones of OneVOS. + +even with just two frames or a single frame in common scenarios, thanks to its unified framework and efficient architecture. Optimal performance is observed when the memory capacity reaches 4 or 5. This underscores the importance of memory in enabling OneVOS to effectively handle longer and more intricate videos. + +Number of Retained Nodes. Table. 5(b) focuses on retaining a set number of top-k nodes in token memory, based on attention weights. Results show that after reaching a per-layer threshold, keeping approximately 512 nodes struck an optimal balance between information depth and computational efficiency. + +Backbone. In Table. 5(c), we assess Vision Transformers (ViT) [15] pre-trained with Masked Autoencoders (MAE) [25] and ConvVIT (base and large variants) as backbones for OneVOS. With the VIT-B backbone OneVOS still outperforms most of the methods on D17 datasets. However, a key limitation of ViT was identified in its direct 16x downsampling approach, which resulted in a significant loss of multi-scale detail information. This is particularly detrimental for high-precision, pixel-level VOS tasks. On the other hand, the ConvVIT model, which employs simple convolutional operations for its patch embedding process, was found to effectively preserve multi-scale information. Additionally, the increased complexity of the ConvVIT-large variant further enhanced the model's performance, underscoring the importance of architectural design in capturing intricate details necessary for accurate VOS. + +# 5 Conclusion + +This paper introduces OneVOS, a novel semi-supervised video object segmentation framework that globally integrates feature extraction, matching, memory management, and object aggregation within a single Transformer architecture. By employing the All-in-One Transformer with specialized Unidirectional Hybrid Attention modules, we significantly enhance the practicality and performance of OneVOS. Furthermore, to propose a rational, efficient version of OneVOS, we investigate the mechanism of the All-in-One Transformer via the proposed Dynamic Token Selector, revealing its layer-crossing segmentation mode for multi-object feature extraction and matching, leading to a naturally efficient version with dynamic memory. OneVOS demonstrates state-of-the-art performance on 7 datasets, particularly excelling in complex scenarios. We anticipate that this work will draw greater attention to the unified VOS modeling. + +# Acknowledgements + +This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300). + +# References + +1. Bhat, G., Lawin, F.J., Danelljan, M., Robinson, A., Felsberg, M., Van Gool, L., Timofte, R.: Learning what to learn for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 777-794. Springer (2020) +2. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017) +3. Chen, B., Li, P., Bai, L., Qiao, L., Shen, Q., Li, B., Gan, W., Wu, W., Ouyang, W.: Backbone is all your need: A simplified architecture for visual object tracking. In: European Conference on Computer Vision. pp. 375-392. Springer (2022) +4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020) +5. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8126-8135 (2021) +6. Chen, Y., Pont-Tuset, J., Montes, A., Van Gool, L.: Blazingly fast video object segmentation with pixel-wise metric learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1189-1198 (2018) +7. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022) +8. Cheng, H.K., Tai, Y.W., Tang, C.K.: Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5559-5568 (2021) +9. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021) +10. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018) +11. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence 37(3), 569-582 (2014) +12. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13608-13618 (2022) +13. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) + +14. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023) +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +16. Duke, B., Ahmed, A., Wolf, C., Aarabi, P., Taylor, G.W.: Sstvos: Sparse spatiotemporal transformers for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5912-5921 (2021) +17. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010) +18. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023) +19. Gao, J., Qian, X., Wang, Y., Xiao, T., He, T., Zhang, Z., Fu, Y.: Coarse-to-fine amodal segmentation with shape prior. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1262-1271 (2023) +20. Gao, P., Ma, T., Li, H., Lin, Z., Dai, J., Qiao, Y.: Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892 (2022) +21. Gao, S., Zhou, C., Ma, C., Wang, X., Yuan, J.: Aiatrack: Attention in attention for transformer visual tracking. In: European Conference on Computer Vision. pp. 146-164. Springer (2022) +22. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024) +23. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022) +24. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011) +25. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +26. Hong, L., Chen, W., Liu, Z., Zhang, W., Guo, P., Chen, Z., Zhang, W.: Lvos: A benchmark for long-term video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13480-13492 (2023) +27. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024) +28. Hong, L., Zhang, W., Chen, L., Zhang, W., Fan, J.: Adaptive selection of reference frames for video object segmentation. IEEE Transactions on Image Processing 31, 1057-1071 (2021) +29. Hu, P., Wang, G., Kong, X., Kuen, J., Tan, Y.P.: Motion-guided cascaded refinement network for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1400-1409 (2018) + +30. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018) +31. Huang, X., Xu, J., Tai, Y.W., Tang, C.K.: Fast video object segmentation with temporal aggregation network and dynamic template matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8879-8889 (2020) +32. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016) +33. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8953-8962 (2019) +34. Khoreva, A., Benenson, R., Ilg, E., Brox, T., Schiele, B.: Lucid data dreaming for video object segmentation. International Journal of Computer Vision 127(9), 1175-1197 (2019) +35. Kristan, M., Leonardis, A., Matas, J., Felsberg, M., Pflugfelder, R., Cehovin Zajc, L., Vojir, T., Bhat, G., Lukezic, A., Eldesokey, A., et al.: The sixth visual object tracking vot2018 challenge results. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018) +36. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022) +37. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024) +38. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018) +39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +40. Lin, Z., Yang, T., Li, M., Wang, Z., Yuan, C., Jiang, W., Liu, W.: Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1362-1372 (2022) +41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +42. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018) +43. Nowozin, S.: Optimal decisions from probabilistic models: the intersection-over-union case. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 548-555 (2014) +44. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018) + +45. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019) +46. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017) +47. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) +48. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) +49. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021) +50. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 629-645. Springer (2020) +51. Seong, H., Oh, S.W., Lee, J.Y., Lee, S., Lee, S., Kim, E.: Hierarchical memory matching network for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12889-12898 (2021) +52. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015) +53. Voigtlaender, P., Chai, Y., Schroff, F., Adam, H., Leibe, B., Chen, L.C.: Feelvos: Fast end-to-end embedding learning for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9481-9490 (2019) +54. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017) +55. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021) +56. Wang, J., Chen, D., Wu, Z., Luo, C., Tang, C., Dai, X., Zhao, Y., Xie, Y., Yuan, L., Jiang, Y.G.: Look before you match: Instance understanding matters in video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2268-2278 (2023) +57. Wang, W., Shen, J., Porikli, F., Yang, R.: Semi-supervised video object segmentation with super-trajectories. IEEE transactions on pattern analysis and machine intelligence 41(4), 985-998 (2018) +58. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14561-14571 (2023) +59. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13879-13889 (2023) +60. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018) + +61. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) +62. Xu, S., Liu, D., Bao, L., Liu, W., Zhou, P.: Mhp-vos: Multiple hypotheses propagation for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 314-323 (2019) +63. Yan, S., Xu, X., Hong, L., Chen, W., Zhang, W., Zhang, W.: Panovos: Bridging non-panoramic and panoramic views with transformer for video segmentation. arXiv preprint arXiv:2309.12303 (2023) +64. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024) +65. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V. pp. 332-348. Springer (2020) +66. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +67. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4701-4712 (2021) +68. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) +69. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357. Springer (2022) +70. Zhou, X., Guo, P., Hong, L., Li, J., Zhang, W., Ge, W., Zhang, W.: Reading relevant feature from global representation memory for visual object tracking. Advances in Neural Information Processing Systems 36 (2024) \ No newline at end of file diff --git a/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/images.zip b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ef7d191ff9b88ecb6d9ed095b4e1b27d9015b3c --- /dev/null +++ b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a24ee96a5e42dd8cb5ce35a9ab46f9df8dfea67dc574ad1aa7e76482b272d521 +size 780386 diff --git a/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/layout.json b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bba2ede2fbc4caf5bf0a362a31e1d6eaffbc5b5e --- /dev/null +++ b/2024/OneVOS_ Unifying Video Object Segmentation with All-in-One Transformer Framework/layout.json @@ -0,0 +1,9913 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 111, + 471, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 111, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 144, + 111, + 471, + 148 + ], + "type": "text", + "content": "OneVOS: Unifying Video Object Segmentation with All-in-One Transformer Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": "Wanyun Li" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Pinxue Guo" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Xinyu Zhou" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Lingyi Hong" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Yangji He" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Xiangyu Zheng" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Wei Zhang" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Wenqiang Zhang" + }, + { + "bbox": [ + 149, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,3,\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "spans": [ + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 201, + 436, + 224 + ], + "type": "text", + "content": " Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 224, + 461, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 224, + 461, + 280 + ], + "spans": [ + { + "bbox": [ + 152, + 224, + 461, + 280 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 224, + 461, + 280 + ], + "type": "text", + "content": " Shanghai Engineering Research Center of AI & Robotics, Academy for Engineering & Technology, Fudan University, Shanghai, China \n" + }, + { + "bbox": [ + 152, + 224, + 461, + 280 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 152, + 224, + 461, + 280 + ], + "type": "text", + "content": " Engineering Research Center of AI & Robotics, Ministry of Education, Academy for Engineering & Technology, Fudan University, Shanghai, China {wyli22,pxguo21}@m.fudan.edu.cn, {weizh, wqzhang}@fudan.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "spans": [ + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "text", + "content": "Abstract. Contemporary Video Object Segmentation (VOS) approaches typically consist of feature extraction, matching, memory management, and multiple objects aggregation. Recent advanced models either employ a discrete modeling for these components in a sequential manner, or optimize a combined pipeline through substructure aggregation. However, these existing explicit staged approaches prevent the VOS framework from being optimized as a unified whole, leading to the limited capacity and suboptimal performance in tackling complex videos. In this paper, we propose OneVOS, a novel framework that unifies the core components of VOS with All-in-One Transformer. Specifically, to unify all aforementioned modules into a vision transformer, we model all the features of frames, masks and memory for multiple objects as transformer tokens, and integrally accomplish feature extraction, matching and memory management of multiple objects through the flexible attention mechanism. Furthermore, a Unidirectional Hybrid Attention is proposed through a double decoupling of the original attention operation, to rectify semantic errors and ambiguities of stored tokens in OneVOS framework. Finally, to alleviate the storage burden and expedite inference, we propose the Dynamic Token Selector, which unveils the working mechanism of OneVOS and naturally leads to a more efficient version of OneVOS. Extensive experiments demonstrate the superiority of OneVOS, achieving state-of-the-art performance across 7 datasets, particularly exceeding in complex LVOS and MOSE datasets with " + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "inline_equation", + "content": "70.1\\%" + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "inline_equation", + "content": "66.4\\%" + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "inline_equation", + "content": "J\\& F" + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "text", + "content": " scores, surpassing previous state-of-the-art methods by " + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 160, + 300, + 455, + 574 + ], + "type": "text", + "content": ", respectively. Code is available at: https://github.com/L599wy/OneVOS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 623, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 648 + ], + "type": "text", + "content": "Semi-supervised Video Object Segmentation (VOS) [2, 22, 30, 37, 44, 45, 63, 64] is a fundamental task in video analysis [18, 19], aiming to consistently seg-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 339, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 339, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 339, + 666 + ], + "type": "text", + "content": "* Equal contribution † Corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 304, + 163 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 304, + 163 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 304, + 163 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 304, + 163 + ], + "type": "image", + "image_path": "f3746337afe18d228069598e659319365bafa067089b36f3435bd0f27c82862e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 465, + 163 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 465, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 465, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 465, + 163 + ], + "type": "image", + "image_path": "2598b97b547d3b20c53e025f402918a65e127eeb22d20b3607133be2fac464b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 152, + 167, + 307, + 227 + ], + "blocks": [ + { + "bbox": [ + 152, + 167, + 307, + 227 + ], + "lines": [ + { + "bbox": [ + 152, + 167, + 307, + 227 + ], + "spans": [ + { + "bbox": [ + 152, + 167, + 307, + 227 + ], + "type": "image", + "image_path": "0f75c54cb4d79e617056685e3ed2911dacd00b82dc7bb8a84818ce82d4b161a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 228, + 277, + 236 + ], + "lines": [ + { + "bbox": [ + 184, + 228, + 277, + 236 + ], + "spans": [ + { + "bbox": [ + 184, + 228, + 277, + 236 + ], + "type": "text", + "content": "(b) Unify Matching and Objects Aggregation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 166, + 465, + 226 + ], + "blocks": [ + { + "bbox": [ + 312, + 166, + 465, + 226 + ], + "lines": [ + { + "bbox": [ + 312, + 166, + 465, + 226 + ], + "spans": [ + { + "bbox": [ + 312, + 166, + 465, + 226 + ], + "type": "image", + "image_path": "cf56c97d1d60892b6b2d8633288ffba70e01d8090a44b0ab8376c379b3239f43.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 228, + 451, + 235 + ], + "lines": [ + { + "bbox": [ + 347, + 228, + 451, + 235 + ], + "spans": [ + { + "bbox": [ + 347, + 228, + 451, + 235 + ], + "type": "text", + "content": "(d) Integrate Core Modules Al in One Transformer" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 244, + 480, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 480, + 288 + ], + "type": "text", + "content": "Fig. 1: Four different modeling for semi-supervised video object segmentation pipelines. (a) Traditional discrete modeling of VOS. (b) Unified multiple object modeling in one reference. (c) Unified single-object feature extraction and matching. (d) OneVOS: Integrates all core modules with All-in-One Transformer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 294, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 390 + ], + "type": "text", + "content": "ment objects throughout a video based on initial masks. Advancements in VOS primarily focus on the full utilization of available information. Among them, Memory-based methods [7,9,40,45] have emerged as the mainstream due to the effectiveness in storing and leveraging abundant spatio-temporal features. Its foundational paradigm, as depicted in Fig. 1(a) [16,40,51,55], encompasses four split key stages including feature extraction, matching, memory management, and multi-object aggregation, which have been widely adopted as a baseline in a number of methods [23,28,50,56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 533 + ], + "type": "text", + "content": "Recently, some investigations demonstrate the advantages of integration modeling compared with highly discrete modeling. AOT [66] and DEAOT [68] (Fig. 1(b)) emphasize the shortage of single-object inference compared to joint modeling of multiple objects, particularly in exploring contextual and robust feature representations. Meanwhile, SimVOS [59] (Fig. 1(c)) reveals limitations in the separation of feature extraction and matching processes for capturing dynamic object-aware features. However, existing methods solely focus on combining or aggregating substructures of the VOS pipeline, resulting in their suboptimal solutions. For example, AOT relies on sophisticated designing matching modules which fails to leverage superior pre-training weights, while the simplistic single-object framework of SimVOS encounters inefficiencies and the inapplicability of memory makes it unable to utilize rich features of the processed frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Therefore, we introduce an essentially novel framework distinct from previous methods, aiming to develop a unified end-to-end differentiable VOS model for global optimization. Based on this, the All-in-One Transformer for Video Object Segmentation (OneVOS) is proposed, which models and unifies all important modules into a Vision Transformer. Specifically, we first leverage the flexible attention mechanism to conduct feature extraction and mutual matching between the reference and current frames. To effectively capture multi-object relationships, we introduce a mask embedding layer that enriches features with multi-object modeling insights. Moreover, in the memory design, we propose an external layer-wise token memory to enhance the coherence understanding of videos by storing and providing spatio-temporal features for subsequent frames." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "As the feature extraction and matching of multiple objects performing simultaneously, the reference tokens output from each layer of Transformer for memory update could be represented by a weighted linear combination of tokens from memory, reference frame and current frame, which may lead to a significant semantic errors and ambiguity to serving as features stored in memory. To address this, we further propose a Unidirectional Hybrid Attention (Uni-Hybrid Attention) through a double decoupling of the original attention operation. It guarantees the versatility of stored features and enhances memory performance. Notably, as the Uni-Hybrid Attention is derived from naive attention mechanism, our framework is still able to utilize pre-trained weights from various Vision Transformers, thereby enhancing its generality and flexibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 439 + ], + "type": "text", + "content": "In this way, the proposed OneVOS framework implements end-to-end modeling and global optimization, exhibiting superior performance compared to previous models. To further strengthen its advantages, we conduct an in-depth exploration of the behavior within the attention layer of OneVOS to enhance its efficiency, which is particularly crucial in resource-constrained and long-term video tasks. Concretely, we introduce the Dynamic Token Selector (DTS), capable of dynamically selecting partial tokens from each layer for cross-attention and memory storage, trained solely under segmentation loss without additional constraints. With DTS integrated into every layer of the transformer, we analyze the token selection distribution of DTS and the attention weight distribution, uncovering a significant insight: unlike the traditional sequential paradigm, OneVOS adopts a cross-layer segmentation mode that dynamically shifts focus between feature extraction and matching across different layers. Based on this, we propose a natural acceleration scheme involving designing distinct memory capacities for each layer and layer-wise memory updates via tokens selected by DTS. This approach enhances both speed and accuracy across multiple datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "text", + "content": "We conducted comprehensive experiments across seven datasets, and OneVOS consistently achieves state-of-the-art performance across all of them. Notably, OneVOS demonstrates a substantial performance advantage enven in more challenging scenarios, achieving " + }, + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "inline_equation", + "content": "71.1\\%" + }, + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{I}\\& \\mathcal{F}" + }, + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "text", + "content": " on LVOS validation and " + }, + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "inline_equation", + "content": "67.2\\%" + }, + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{I}\\& \\mathcal{F}" + }, + { + "bbox": [ + 130, + 440, + 483, + 499 + ], + "type": "text", + "content": " on MOSE validation. The main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 509, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 509, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 556 + ], + "type": "text", + "content": "- We propose All-in-One Transformer framework for Video Object Segmentation (OneVOS), integrating feature extraction, matching, memory management, and objects aggregation into a vision transformer, which can be optimized globally. This provides a fundamentally novel paradigm in VOS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 557, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 481, + 593 + ], + "type": "text", + "content": "- We introduce the Unidirectional Hybrid Attention mechanism to facilitate multi-object feature extraction and matching, ensuring tokens are devoid of semantic ambiguity for efficient memory storage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 594, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 481, + 628 + ], + "type": "text", + "content": "- We propose the Dynamic Tokens Selector, which not only unveils the internal working mechanism but also leads to a more efficient OneVOS, based on the exploration and the resulting insight." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "text", + "content": "- OneVOS achieves state-of-the-art performance across 7 datasets, particularly excelling in more complex LVOS and MOSE datasets with " + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "70.1\\%" + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "66.4\\%" + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "J\\& F" + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "text", + "content": ", surpassing previous best methods by " + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 138, + 630, + 481, + 665 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 144, + 295, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 295, + 156 + ], + "type": "text", + "content": "2.1 Video object segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 167, + 481, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 481, + 274 + ], + "type": "text", + "content": "Semi-supervised VOS methods can be divided into four types: The earlier fine-tuning methods [1, 2, 42, 54, 60] learn general segmentation features via offline pre-training, then enhance object-specific representations by online fine-tuning. Then, propagation-based methods [10, 29, 33, 44, 46, 57] utilize optical flow to propagate the mask from the previous frame to the next one. Subsequent methods optimized it by keyframes enhancement [34], advanced tracking techniques [4, 10, 31, 62], and re-identification mechanisms [38]. Matching-based methods [6, 30, 53, 65, 67] identify the object of the current frame based on the correlation with the reference frame in the embedding space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 275, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 455 + ], + "type": "text", + "content": "Recently, Memory-based approaches [23,28,45,50] store features of segmented frames in a memory bank to provide a set of features for segmentation, which has become the dominant method in SVOS. Following that, methods such as HMMN [51], SSTVOS [16], and XMem [7] enhance performance by utilizing multi-scale features, introducing Transformer for matching, and building long-term memory. These methods typically adhere to a traditional discrete modeling and single-object architecture. More recently, the multi-object framework [66], [68] is proposed to address the inherent limitations of the previous single-object while still maintaining a sophisticated matching module for high performance. [59] presents a simplified VOS framework for joint feature extraction and matching which adopts single-object modeling and discards the vital memory design in VOS. Differently, our OneVOS, innovatively unifies feature extraction, matching, memory construction, and multi-objective aggregation all in one Transformer. This design not only streamlines the segmentation process but also enhances the dynamic interaction among different stages of SVOS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 475, + 278, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 278, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 278, + 488 + ], + "type": "text", + "content": "2.2 Visual Object Tracking" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Video object tracking (VOT) [3,27,35,70] shares similarities with the task setting of SVOS, but it primarily utilizes bounding boxes for object localization instead of detailed segmentation masks. In VOT, there has been a paradigm shift from traditional Siamese-based methods towards leveraging Transformer layers for enriched relational modeling between the template and search regions. Pioneering works like TransT [5] introduce a two-stream attention mechanism for robust bi-directional information exchange, while innovative methods like MixFormer [12] and AiATrack [21] adopt asymmetric attention strategies to minimize distractions from irrelevant objects. Notably, OSTrack [69] proposes a unified attention model, promoting seamless feature integration and relation modeling. Nevertheless, the distinct multi-object settings and difference in target information representation, coupled with the critical role of memory in VOS, render unification within the VOS field more challenging. This paper endeavors to tackle this challenge." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "text", + "content": "As shown in the Fig. 2, in a given video sequence, VOS focuses on segmenting the current " + }, + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "text", + "content": "-th frame " + }, + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "inline_equation", + "content": "f_{t} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "text", + "content": " using the reference frame " + }, + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "inline_equation", + "content": "f_{t-1} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 140, + 482, + 258 + ], + "type": "text", + "content": " with its mask and the history features stored in Memory. In this section, we aim to construct an unified framework at a global scale to tackle the challenges of VOS. We commence by introducing the OneVOS framework in Sec. 3.1, which employs Vision Transformer to integrate core elements in VOS. And then Sec. 3.2 provides a detailed exposition on the novel Unidirectional Hybrid Attention mechanism. In addition, in Sec. 3.3, we explore the working mechanism of attention within OneVOS through the proposed Dynamic Token Selector in order to introduce the efficient version of OneVOS." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 153, + 262, + 289, + 350 + ], + "blocks": [ + { + "bbox": [ + 153, + 262, + 289, + 350 + ], + "lines": [ + { + "bbox": [ + 153, + 262, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 153, + 262, + 289, + 350 + ], + "type": "image", + "image_path": "a9da2db1d58dedbded6f8e879caf3feeb187f20792ccb416c101db48fed8b752.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 351, + 252, + 358 + ], + "lines": [ + { + "bbox": [ + 184, + 351, + 252, + 358 + ], + "spans": [ + { + "bbox": [ + 184, + 351, + 252, + 358 + ], + "type": "text", + "content": "(a) Hybrid Attention Block" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 291, + 262, + 460, + 346 + ], + "blocks": [ + { + "bbox": [ + 291, + 262, + 460, + 346 + ], + "lines": [ + { + "bbox": [ + 291, + 262, + 460, + 346 + ], + "spans": [ + { + "bbox": [ + 291, + 262, + 460, + 346 + ], + "type": "image", + "image_path": "dbe76f2aa51d65433abdb0e0adec0aaaff1b6c919531c8ef84a566e1c1b9b0e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 351, + 430, + 359 + ], + "lines": [ + { + "bbox": [ + 315, + 351, + 430, + 359 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 430, + 359 + ], + "type": "text", + "content": "(b) Block with Dynamic Token Selector (DTS)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 154, + 363, + 460, + 468 + ], + "blocks": [ + { + "bbox": [ + 154, + 363, + 460, + 468 + ], + "lines": [ + { + "bbox": [ + 154, + 363, + 460, + 468 + ], + "spans": [ + { + "bbox": [ + 154, + 363, + 460, + 468 + ], + "type": "image", + "image_path": "35e1eab2a4cb32f940bfe4f581ec8d0c98b7ed15375bf18eb82fb696bdef8076.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 468, + 322, + 476 + ], + "lines": [ + { + "bbox": [ + 284, + 468, + 322, + 476 + ], + "spans": [ + { + "bbox": [ + 284, + 468, + 322, + 476 + ], + "type": "text", + "content": "(c) Overview" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 485, + 482, + 584 + ], + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 584 + ], + "type": "text", + "content": "Fig. 2: (a) Hybrid Attention Block. (b) Blocks with Dynamic Token Selector. (c) Overview of OneVOS framework equipped with All-in-One Transformer. The reference mask is initially encoded into mask embeddings. Patch embeddings are generated from both the reference and current frames, which are then processed by the Transformer. Within each Transformer layer, mask embeddings are integrated with reference tokens. Then they are selected by Dynamic Token Selector and concatenated with tokens stored in memory and current frame fed into the Hybrid Attention Module. Finally, the features of the current frame output from Transformer are subsequently decoded into multi-object masks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 263, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 263, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 263, + 621 + ], + "type": "text", + "content": "3.1 OneVOS Framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Mask Embedding. To capture complex target interactions and minimize redundant single-target inferences, motivated by [66], we utilize the Mask Embedding Layer (as Patch Embedding Layer in Transformer) to encode reference" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 90, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 90, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 90, + 448, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "text", + "content": "mask into Mask Embedding denoted as " + }, + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "inline_equation", + "content": "m_{ref} \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P} \\times C}" + }, + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "text", + "content": " through a single convolutional layer. Here, " + }, + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "text", + "content": " represents the embedding dimension, with " + }, + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 114, + 482, + 212 + ], + "type": "text", + "content": " as the stride of the mask embedding. Utilizing the local relationship modeling capability of the convolution operation, the discriminability of multiple targets is enhanced, which will then be integrated with visual feature map in Transformer before attention computation to integrating mask embeddings to the immediate value tokens of each layer, thereby effectively promoting the model to perform multiple targets inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "text", + "content": "All-in-One Transformer. Then, we adopt a transformer architecture to unify the core modules of VOS. It first generates the non-overlapping patch embeddings " + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "inline_equation", + "content": "e_t^0 \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "inline_equation", + "content": "e_{ref}^0 \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "text", + "content": " for the current and reference frames, where " + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "text", + "content": " is the patch resolution same as the mask embedding, thus " + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "inline_equation", + "content": "N = HW / P^2" + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "text", + "content": ". Then, the concatenation of " + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "inline_equation", + "content": "[e_{ref}^0; e_t^0]" + }, + { + "bbox": [ + 131, + 216, + 482, + 314 + ], + "type": "text", + "content": " along the spatial dimension serves as the initial input for the backbone, facilitating adaptive multi-object feature extraction and target information propagation through the stacked layers. In each layer, we first integrate mask embedding with visual feature map as shown in Eq 1:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 318, + 481, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 318, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 211, + 318, + 481, + 332 + ], + "type": "interline_equation", + "content": "\\left\\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} \\right\\} = \\left\\{Q _ {r e f} ^ {l}, K _ {r e f} ^ {l}, V _ {r e f} ^ {l} + m _ {r e f} \\right\\}, \\tag {1}", + "image_path": "6ebfadd17a5b9f5ca73e2b379636bcd86b731aedda6c81e3a2390a375ce5d373.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "content": " are query, key, and value matrices aligning with the standard attention mechanism [15], and " + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "inline_equation", + "content": "l \\in \\{0, 1, \\dots, L\\}" + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "content": " represents the layer. Then, these object-aware reference tokens, memory tokens, and tokens of current frames undergo adaptive feature extraction and target information propagation by applying the attention mechanism of the Transformer. The output " + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "inline_equation", + "content": "a^l = [a_{ref}^l; a_t^l]" + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "content": " of layer " + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 337, + 482, + 409 + ], + "type": "text", + "content": " after the attention operation is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 422, + 481, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 422, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 142, + 422, + 481, + 449 + ], + "type": "interline_equation", + "content": "a ^ {l} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot V = \\operatorname {S o f t m a x} \\left(\\frac {\\left[ Q _ {r e f} ^ {l} ; Q _ {t} ^ {l} \\right] \\left[ K _ {M} ^ {l} ; K _ {r e f} ^ {l} ; K _ {t} ^ {l} \\right]}{\\sqrt {d _ {k}}}\\right) \\cdot \\left[ V _ {M} ^ {l}; V _ {r e f} ^ {l}; V _ {t} ^ {l} \\right], \\quad (2)", + "image_path": "fb636c77ecd885449f5fc543e15ec0805b6e2ad5190e349c473939d7c623d4e6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "inline_equation", + "content": "K_M^l, V_M^l" + }, + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "text", + "content": " are features stored in memory. Supposing " + }, + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "inline_equation", + "content": "W = QK^{T}" + }, + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "text", + "content": " is the matrix of similarity weights, " + }, + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "inline_equation", + "content": "a_{t}^{l}" + }, + { + "bbox": [ + 131, + 452, + 482, + 476 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 481, + 480, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 481, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 231, + 481, + 480, + 493 + ], + "type": "interline_equation", + "content": "a _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}, \\tag {3}", + "image_path": "877c9775755fd0fbb0cc674c5aee13e57d3fe70771a47390eb813a432846efed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "inline_equation", + "content": "[W_{t,ref}V_{ref}]" + }, + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "inline_equation", + "content": "[W_{t,M}V_M]" + }, + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "text", + "content": " are responsible for aggregating the inter-frame features from memory and reference frame to current frame (matching), and " + }, + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "inline_equation", + "content": "[W_{t,t}V_t]" + }, + { + "bbox": [ + 131, + 497, + 482, + 552 + ], + "type": "text", + "content": " aggregating the intra-frame features (feature extraction) for current frame, respectively. Thus, dynamic coordination between feature extraction and matching is achieved through the attention weights learned adaptively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 556, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 482, + 622 + ], + "type": "text", + "content": "Token Memory. In addition, since the target information and the reference frame are integrated within our framework, we can store intermediate features directly into Token Memory without necessitating additional post-processing fusion steps. Given the semantic variance of intermediate tokens in the reference frame across layers, we introduce a token storage mechanism to save essential Transformer tokens instead of whole frames as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 173, + 625, + 481, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 625, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 173, + 625, + 481, + 639 + ], + "type": "interline_equation", + "content": "M = \\left\\{K _ {M} ^ {l}, V _ {M} ^ {l} \\mid \\in \\{0, 1, \\dots L \\} \\right\\} = \\left\\{K _ {\\text {r e f s}} ^ {l}, V _ {\\text {r e f s}} ^ {l} \\mid \\in \\{0, 1, \\dots L \\} \\right\\}, \\tag {4}", + "image_path": "bf0c5084f34227f454b9d4a54ccaa5436b8a5f614cdf2bb5251f69b0f649e8f9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 643, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 643, + 482, + 665 + ], + "type": "text", + "content": "This method allows these tokens to be involved only in the attention modules instead of multiple inferences, greatly reducing inference overhead." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 304, + 175 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 304, + 175 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 304, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 304, + 175 + ], + "type": "image", + "image_path": "395d5876de5b3d53d08a5979d97bf55681b9cc3befe6001eb255259c4857b0f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 187, + 480, + 263 + ], + "lines": [ + { + "bbox": [ + 130, + 187, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 480, + 263 + ], + "type": "text", + "content": "Fig. 3: (a) Visualization of attention distribution of target areas in OneVOS, showing the alternation between self-attention and cross-attention across layers. (b) Analysis of the maximal attention weights reveals a pattern of alternation between feature extraction via self-attention (notably in layers 3 and 5) and matching via cross-attention (as in layers 2 and 8). (c) The proportional analysis of the attention weights' sum, normalized to one, underscores the non-sequential prioritization across different layers between self-attention and cross-attention." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 116, + 392, + 176 + ], + "blocks": [ + { + "bbox": [ + 307, + 116, + 392, + 176 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 392, + 176 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 392, + 176 + ], + "type": "image", + "image_path": "7af79e823c1599a2fa52ddfdb957dc04831c27f423e3b0a230ac5a4e07a85ec2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 395, + 118, + 480, + 176 + ], + "blocks": [ + { + "bbox": [ + 395, + 118, + 480, + 176 + ], + "lines": [ + { + "bbox": [ + 395, + 118, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 395, + 118, + 480, + 176 + ], + "type": "image", + "image_path": "c3284748381c29e9423cff0ade4ca7de693bca5519c9ca89200d9015b3645517.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 270, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 480, + 304 + ], + "type": "text", + "content": "Mask Decoding. Finally, the output of All-in-One Transformer, " + }, + { + "bbox": [ + 131, + 270, + 480, + 304 + ], + "type": "inline_equation", + "content": "e_t^L" + }, + { + "bbox": [ + 131, + 270, + 480, + 304 + ], + "type": "text", + "content": " is input to the lightweight Feature Pyramid Network (FPN) for Mask Decoding. These features are then gradually upgraded to the original scale and decoded into multi-object masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 322, + 324, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 324, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 324, + 335 + ], + "type": "text", + "content": "3.2 Unidirectional Hybrid Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 480, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 408 + ], + "type": "text", + "content": "Building upon our newly developed pipeline, all modules are integrated into a transformer as in Fig. 2(c). However, obtaining the stored features by employing original attention mechanism may lead to semantic errors and ambiguity since they are represented incorrectly. To address this challenge, our proposed Unidirectional Hybrid Attention (Uni-Hybrid Attention) is strategically designed through dual-level attention decoupling as shown in Fig. 2(a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 407, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 496 + ], + "type": "text", + "content": "Firstly, reference features calculated by naive attention operation to all the tokens are denoted as " + }, + { + "bbox": [ + 130, + 407, + 481, + 496 + ], + "type": "inline_equation", + "content": "a_{ref}^{l} = W_{ref,M}V_{M} + W_{ref,ref}V_{ref} + W_{refs,t}V_{t}" + }, + { + "bbox": [ + 130, + 407, + 481, + 496 + ], + "type": "text", + "content": ". A key limitation in the computations arises where tokens of reference frame containing target information (as in Eq. (1)) are represented by tokens of current frame that without target information, causing semantic errors in the features stored in the memory which are essential for storing spatio-temporal target information for subsequent frames. Therefore, to alleviate this situation, our first modification decoupling attention from the current token to the reference token. The revised expression is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 235, + 503, + 480, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 503, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 235, + 503, + 480, + 517 + ], + "type": "interline_equation", + "content": "a _ {r e f} ^ {l} = W _ {r e f, M} V _ {M} + W _ {r e f, r e f} V _ {r e f}. \\tag {5}", + "image_path": "8a5b569557333e50d5cbfb4c128e6ee89d8e7be6c3d6c4ff8c73379908f257fe.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 524, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 482, + 632 + ], + "type": "text", + "content": "Subsequently, the attention computation between reference features and memory tokens is decoupled, serving a dual purpose: 1) Avoiding Information Overlap and Ambiguity: Memory tokens, enriched with mask embeddings, encapsulate target information from prior frames. Integrating these embeddings with those of the reference frame risks overlapping and confusing object cues. 2) Preserving Temporal Specificity: Every video frame contains essential temporal intricacies crucial for segmentation precision. Attention decoupling preserves the reference frame's temporal uniqueness, critical for accurate and context-aware segmentation. Therefore, we decouple attention between memory tokens and those from the reference frame, and the final unidirectional hybrid attention formulations are:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 230, + 638, + 481, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 638, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 230, + 638, + 481, + 658 + ], + "type": "interline_equation", + "content": "a _ {r e f} ^ {l} = W _ {r e f, r e f} V _ {r e f} \\tag {6}", + "image_path": "d785f0f855b81b56db8726facc5f369ffccacdcb325da317e0f1ee315c46004d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 231, + 655, + 382, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 655, + 382, + 667 + ], + "spans": [ + { + "bbox": [ + 231, + 655, + 382, + 667 + ], + "type": "interline_equation", + "content": "a _ {t} ^ {l} = W _ {t, M} V _ {M} + W _ {t, r e f} V _ {r e f} + W _ {t, t} V _ {t}.", + "image_path": "7fac4260b56eb839829296da87487375d26705bba3f78f57aee70c7c5fc5bf9a.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 251, + 184 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 251, + 184 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 251, + 184 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 251, + 184 + ], + "type": "image", + "image_path": "1a6fbaeba53cff3baf1fb75b5821301ee718e1254ca7a8df13b2b9954fd8b878.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 186, + 255, + 194 + ], + "lines": [ + { + "bbox": [ + 247, + 186, + 255, + 194 + ], + "spans": [ + { + "bbox": [ + 247, + 186, + 255, + 194 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 114, + 369, + 184 + ], + "blocks": [ + { + "bbox": [ + 253, + 114, + 369, + 184 + ], + "lines": [ + { + "bbox": [ + 253, + 114, + 369, + 184 + ], + "spans": [ + { + "bbox": [ + 253, + 114, + 369, + 184 + ], + "type": "image", + "image_path": "cc9e9bbfc14e091f1395a056962e2a9c36b2365f7d8b6ec68b98362c77ac418c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 203, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 480, + 236 + ], + "type": "text", + "content": "Fig. 4: (a) Visualization of the selected tokens of reference frame by DTS in different Transformer layers. (b) Adaptive selection ratios learned by DTS in different layers without any constraints." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 375, + 121, + 481, + 184 + ], + "blocks": [ + { + "bbox": [ + 375, + 121, + 481, + 184 + ], + "lines": [ + { + "bbox": [ + 375, + 121, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 375, + 121, + 481, + 184 + ], + "type": "image", + "image_path": "c5b83b667266c087cdaca6abda1d6bfc44ea014f069fcbf754d72e26c6201b4a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 186, + 435, + 194 + ], + "lines": [ + { + "bbox": [ + 427, + 186, + 435, + 194 + ], + "spans": [ + { + "bbox": [ + 427, + 186, + 435, + 194 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 243, + 294, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 294, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 294, + 254 + ], + "type": "text", + "content": "3.3 Towards Efficient OneVOS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 262, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 316 + ], + "type": "text", + "content": "Utilizing the advanced All-in-One Transformer equipped with Uni-Hybrid Attention, OneVOS achieves outstanding performance across various datasets. However, the integration process of these modules in OneVOS still lacks explainability, hindering the rational design of a more efficient version. Therefore, this section delves into its working mechanism to propose the efficient OneVOS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 320, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 453 + ], + "type": "text", + "content": "Exploration and Insight. Within the Transformer architecture, attention weights determine the influence of each token in the input sequence on the output which is useful for interpreting Transformer behavior. By examining the attention weights of tokens in the target area of the current frame across each layer, we compare their focus in reference frames or themselves. As presented in Fig. 3, qualitative results in 3(a) shown that a fluctuating focus between reference frames (cross-attention) and themselves (self-attention), Subsequently, Fig. 3 (b) and (c) provide quantitative evidence of a complex alternation pattern, which deviates from the sequential feature extraction and matching process of traditional discrete VOS modeling or uniform distribution in each layer. Instead, these findings lead us to a natural insight: OneVOS may learn a dynamic layer-wise integration strategy, preferentially executing multi-object feature extraction or matching in each layer according to the semantic context." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 456, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 480, + 512 + ], + "type": "text", + "content": "Dynamic Token Selector. Based on above insights and motivated by [49], we introduce a Dynamic Token Selector (DTS) into the OneVOS architecture, as depicted in Fig. 2(b). This module adaptively selects a subset of tokens from the reference frame for cross-attention and storage. The selection process, as defined in Eq. 7, combines global context and local token information:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 218, + 518, + 480, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 518, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 218, + 518, + 480, + 548 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p = \\operatorname {S o f t m a x} (\\operatorname {M L P} \\left(\\left[ \\operatorname {M a x P o o l} \\left(z _ {r e f} ^ {l}\\right); z _ {r e f} ^ {l}\\right]\\right)) \\\\ z _ {r e f} ^ {l} = \\operatorname {M L P} \\left(\\operatorname {N o r m} \\left(e _ {r e f} ^ {l}\\right)\\right) \\tag {7} \\\\ \\end{array}", + "image_path": "c2bf1f02a9b2fa1846f5692b1f8f8e96950b45b9e382e91551ec548289cf831b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "inline_equation", + "content": "z_{ref}^{l}" + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "content": " denotes the local feature encoded of each token, while [MaxPool(" + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "inline_equation", + "content": "z_{ref}^{l}" + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "content": ")] aggregates the overall context of the reference frame. " + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{N \\times 2}" + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "content": " represents the probabilities associated with each token. Specifically, " + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "inline_equation", + "content": "p_{i,0}" + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "content": " indicates the probability of dropping " + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "content": "-th token while " + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "inline_equation", + "content": "p_{i,1}" + }, + { + "bbox": [ + 130, + 555, + 479, + 599 + ], + "type": "text", + "content": " indicates for selection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 599, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 480, + 665 + ], + "type": "text", + "content": "This design aims to enable the network to autonomously determine the optimal token selection ratio for high-performance segmentation. Consequently, it is integrated into an All-in-One Transformer, as depicted in Fig. 2(a). Moreover, it is exclusively trained under segmentation loss without additional constraints. By observing the proportion of tokens selected by DTS in each transformer layer, as illustrated in Fig. 4(a) and Fig. 4(b), a crucial insight emerges. The model automatically allocates varying" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 482, + 194 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 482, + 194 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 482, + 194 + ], + "type": "image", + "image_path": "66175493c3afb3bd5e120b006f4ecb74974800cb63a048ace280568936aba7e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 204, + 480, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 226 + ], + "type": "text", + "content": "Fig. 5: Visualization dynamic memory updates of OneVOS utilizing DTS. The selected tokens in red boxes are derived output from DTS as Fig. 4(a)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 233, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 479, + 289 + ], + "type": "text", + "content": "selection ratios to different layers, and even assigns a selection ratio of zero or one to specific layers. This observation underscores the focus of DTS on certain layers for feature extraction (selecting a proportion of reference frame tokens as zero) and on specific layers for matching (selecting a proportion of reference frame tokens as one), presenting a characteristic of alternating focus across layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 480, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 347 + ], + "type": "text", + "content": "Efficient Version of One VOS. Building upon the above observations, a straightforward acceleration method is naturally introduced. This includes layer-wise memory capacity and adaptive token update mechanism via DTS, as shown in Fig. 5. Specifically, we initialize different capacities for each layer of Token Memory, with the token capacity formula as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 269, + 348, + 480, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 348, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 269, + 348, + 480, + 360 + ], + "type": "interline_equation", + "content": "c _ {l} = r _ {l} * N * C a p. \\tag {8}", + "image_path": "0314635280c97df2ae8d6b40471e5161208bf47b3db50757fccdb1770fa9d823.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "inline_equation", + "content": "r_l" + }, + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "text", + "content": " represents the layer-specific selection ratio determined by DTS during training, as illustrated in Fig. 4(b). Additionally, " + }, + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "inline_equation", + "content": "Cap" + }, + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "text", + "content": " is a hyper-parameter denoting the frame capacity (typically set to 3 or 5). Consequently, tokens are incrementally added to the memory based on these selection ratios until reaching the capacity limit of each layer. Subsequently, the top-k tokens, determined by the current attention weights " + }, + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "inline_equation", + "content": "W_{t,ref}" + }, + { + "bbox": [ + 130, + 363, + 482, + 451 + ], + "type": "text", + "content": ", are retained, followed by the continuous addition of tokens to memory. Aligned with the segmentation mode of OneVOS, this method achieves improvements in accuracy and speed across multiple datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 470, + 224, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 224, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 224, + 483 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 517, + 480, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 480, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 480, + 551 + ], + "type": "text", + "content": "One-stream VOS uses ConvVIT-base [20] as the backbone, initialized with ImageNet-1k [13] pre-training weights. The feature resolution of the mask decoding is increased from 1/16 to " + }, + { + "bbox": [ + 130, + 517, + 480, + 551 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 517, + 480, + 551 + ], + "type": "text", + "content": " and the channel dimension is reduced from 768 to 384." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 555, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 665 + ], + "type": "text", + "content": "Training details. Following the common settings of VOS [8,66], the training phase is divided into two stages: pre-training on synthetic video sequences generated from still image datasets [11,17,24,39,52], and main training on standard VOS dataset (DAVIS 17 and YouTube VOS in default, the model mark with \"*\" uses MOSE in main training as well). The loss function is a 0.5:0.5 combination of bootstrapped cross-entropy loss and soft Jaccard loss [43]. The learning rate of two stages is 2e-4 and 1e-4, respectively, utilizing " + }, + { + "bbox": [ + 130, + 555, + 482, + 665 + ], + "type": "inline_equation", + "content": "4 \\times 3090" + }, + { + "bbox": [ + 130, + 555, + 482, + 665 + ], + "type": "text", + "content": "s GPUs. Our batchsize is 4, with a gradient accumulation of 2 and a sequence length of 5, running each training phase for 200,000 iterations. In the training of DTS, we employ the Gumbel-Softmax [32] on its output " + }, + { + "bbox": [ + 130, + 555, + 482, + 665 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 555, + 482, + 665 + ], + "type": "text", + "content": " to generate a one-hot selection vector, indicating whether each token is selected (assigned a value of 1) or" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 448, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 479, + 250 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 250 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 250 + ], + "type": "table", + "html": "
Long-term VideosComplicated Videos
MethodLVOS VALLVOS TESTMethodMOSEJ&FJF
J&FJFJ&FJF
CFBI [65]50.045.055.144.840.549.0XMem [7]x56.452.160.6
AOT-L [66]61.359.453.654.148.759.5DropSeg [58]x52.248.056.4
STCN [9]45.841.150.545.841.650.0R50-AOT-L [66]x58.153.962.3
RDE [36]52.947.758.149.044.453.5R50-DeAOT-L [68]x58.954.563.3
XMem [7]50.045.554.449.545.253.7SwinB-AOT-L [66]x60.056.164.0
AOT-B [66]56.951.861.954.449.359.4OneVOSx62.358.166.6
DeAOT [68]63.757.669.858.152.963.4RDE [36]✓48.844.652.9
R50-AOT-L [66]63.858.169.661.056.165.9STCN [9]✓50.846.655.0
R50-DeAOT-L [68]65.959.971.958.053.262.8XMem [7]✓57.653.362.0
Swin-AOT-L [66]65.660.875.161.556.474.8DeAOT [68]✓59.455.163.8
OneVOS-D70.164.775.560.155.668.0OneVOS-D✓61.757.566.0
OneVOS67.462.072.963.859.570.0OneVOS✓66.462.270.6
OneVOS*71.166.076.264.160.068.2OneVOS*✓67.263.071.5
", + "image_path": "70be2ebb49364ca343aa4975ec24b72e75f4ce09d303b9954cbc0186deb8b217.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of long-duration video datasets (LVOS validation and test) and complex datasets, MOSE. " + }, + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "type": "text", + "content": " indicates MOSE used for training. OneVOS( " + }, + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "type": "text", + "content": " ) only employs MOSE in main training, whereas OneVOS\\* " + }, + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "type": "inline_equation", + "content": "(\\checkmark)" + }, + { + "bbox": [ + 132, + 258, + 482, + 304 + ], + "type": "text", + "content": " utilizes DAVIS, YouTube, and MOSE." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 315, + 481, + 412 + ], + "blocks": [ + { + "bbox": [ + 133, + 315, + 481, + 412 + ], + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 412 + ], + "type": "image", + "image_path": "4e4bfc47311d1b17d3f11651efb96d034901e79d348c89b6fd74091428fb2ffb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 419, + 480, + 442 + ], + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 442 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging short-term videos (MOSE and DAVIS 17)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 448, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 480, + 471 + ], + "type": "text", + "content": "not (value of 0). Throughout training, we record the selection rate, observing that in the later stages of training, the selection rate of each layer gradually stabilizes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 474, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 480, + 540 + ], + "type": "text", + "content": "Inference details. OneVOS utilizes basic memory storage, storing tokens at intervals of 5 or 10 frames and updates using the first-in-first-out queue when it is full, except for the initial frame tokens. Conversely, OneVOS-D employs dynamic memory, defaulting to retain a maximum of 520 top-k tokens when full. The maximum frame-level capacity " + }, + { + "bbox": [ + 132, + 474, + 480, + 540 + ], + "type": "inline_equation", + "content": "Cap" + }, + { + "bbox": [ + 132, + 474, + 480, + 540 + ], + "type": "text", + "content": " varies by dataset, set to " + }, + { + "bbox": [ + 132, + 474, + 480, + 540 + ], + "type": "inline_equation", + "content": "3/5/8" + }, + { + "bbox": [ + 132, + 474, + 480, + 540 + ], + "type": "text", + "content": ". For fair comparison in LVOS, we strictly limit memory to under 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": "Datasets and Metrics. We evaluate the performance of OneVOS on 4 standard VOS datasets, including DAVIS 16 val [47], DAVIS 17 val [48], DAVIS 17 test [48], and YouTube 2019 [61], as well as 3 more challenging datasets, MOSE [14], LVOS val, and LVOS test [26]. The evaluation metrics include region similarity " + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": ", boundary accuracy " + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": ", and their average " + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": ", and FPS for speed. For YouTube-VOS, we also report them from the seen " + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_s" + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": " and unseen " + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_u" + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_u" + }, + { + "bbox": [ + 132, + 544, + 481, + 611 + ], + "type": "text", + "content": " categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 625, + 376, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 625, + 376, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 376, + 639 + ], + "type": "text", + "content": "4.2 Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "Complicated datasets. Table 1 tabulates the quantitative results on the LVOS val, LVOS test, and MOSE. OneVOS demonstrates superior performance across multi-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 491, + 256 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 491, + 256 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 491, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 491, + 256 + ], + "type": "table", + "html": "
MethodDAVIS 16 VALDAVIS 17 VALDAVIS 17 TESTYOUTUBE 19 VAL
J&FJFJ&FJFFPSJ&FJFGJsFsJuFu
CFBI+ [67]89.988.791.182.980.185.75.678.074.481.682.681.786.277.185.2
HMMN [51]90.889.692.084.781.987.55.078.674.782.582.581.786.177.385.0
DropSeg [58]92.190.993.386.583.589.5----83.482.987.377.785.6
STCN [9]91.690.892.585.482.288.620.276.172.779.682.781.185.478.285.9
XMem [7]91.590.492.786.282.989.522.681.077.484.585.584.388.680.388.6
Swin-AOTL [66]92.090.793.385.482.488.412.181.277.385.184.584.088.878.486.7
Swin-DeAOTL [68]92.991.194.786.283.189.215.482.878.986.786.185.390.280.488.6
ISVOS [56]92.691.593.787.183.790.5-82.879.386.286.185.289.780.788.9
MiVOS' [8]91.089.692.484.581.787.411.278.674.982.282.480.684.778.186.4
STCN' [9]91.790.493.085.382.088.620.277.874.381.384.282.687.079.487.7
XMem' [7]92.090.793.287.784.091.422.681.277.684.785.884.889.280.388.8
SimVOS [59]92.991.394.488.085.091.04.980.476.184.684.283.1-79.1-
OneVOS92.791.094.388.584.692.412.084.880.788.986.184.989.880.689.1
OneVOS-D93.191.494.888.885.292.416.083.980.087.685.084.589.279.187.3
OneVOS*92.891.093.788.384.492.112.085.381.389.386.385.089.981.289.1
", + "image_path": "d14b4e8a58aa6e2c7b1b221c8b77c677cd3a5a81a593307948d092e86d4e8b78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 266, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 132, + 266, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 482, + 300 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons on the DAVIS 2016 val, DAVIS 2017 val, and YouTube-VOS 2019 val split. The methods Use ' as a superscript indicates adding the BL30k [8] dataset for training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "type": "text", + "content": "ple settings and datasets. Notably, OneVOS-D excels in the long video dataset LVOS, achieving an impressive " + }, + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "type": "inline_equation", + "content": "70.1\\%" + }, + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "type": "inline_equation", + "content": "J\\& F" + }, + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "type": "text", + "content": " score under identical training conditions, demonstrating enhanced robustness in long-term videos. On the MOSE dataset, results of OneVOS are particularly noteworthy, surpassing the highest results reported in the paper by nearly " + }, + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 132, + 305, + 481, + 382 + ], + "type": "text", + "content": " (66.4% vs 59.4%) trained in the same setting. When trained on YouTube, DAVIS, and MOSE together in the main training, OneVOS* achieves state-of-the-art performance across all three datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 386, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 464 + ], + "type": "text", + "content": "Common datasets. Table 2 tabulates the quantitative results on the four standard datasets. The results clearly highlight the superiority of OneVOS. OneVOS-D achieves the best " + }, + { + "bbox": [ + 132, + 386, + 481, + 464 + ], + "type": "inline_equation", + "content": "93.1\\%" + }, + { + "bbox": [ + 132, + 386, + 481, + 464 + ], + "type": "text", + "content": " in DAVIS 2016 val and " + }, + { + "bbox": [ + 132, + 386, + 481, + 464 + ], + "type": "inline_equation", + "content": "88.8\\%" + }, + { + "bbox": [ + 132, + 386, + 481, + 464 + ], + "type": "text", + "content": " in DAVIS 2017 val respectively, setting new state-of-the-art performance. OneVOS and OneVOS* showcase remark improvement in DAVIS 2017 test and YouTube-VOS 2019 val. This proves that OneVOS has excellent segmentation capabilities in both regular and challenging scenarios, demonstrating its practicability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 467, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 482, + 534 + ], + "type": "text", + "content": "Qualitative Results. We present visualization results of OneVOS comparing against the state-of-the-art method XMem [7] and Swin-AOTL [36]. The results for the short-term datasets and complicated long-term videos are shown in Fig. 6 and Fig. 7, respectively. OneVOS consistently achieves finer and higher-quality segmentation on the short-term datasets. Notably, it showcases superior segmentation capabilities in long-term scenarios involving both background interference and small targets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 560, + 237, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 560, + 237, + 572 + ], + "spans": [ + { + "bbox": [ + 133, + 560, + 237, + 572 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "text", + "content": "In this section, we conduct a comprehensive analysis of the OneVOS framework and the efficacy of its core designs on both common and the more challenging datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "type": "text", + "content": "Table 3 details an incremental ablation study highlighting the significance of key design elements in OneVOS. Starting with a baseline where ConvVIT replaces the backbone in a top discrete multi-model framework [66] (line 1). Furthermore, to negate any bias stemming from varying modeling approaches toward the backbone, we employ an optimal model based on Swin Transformer [41] as an advanced baseline (line 2), providing a stringent benchmark to validate the effectiveness of OneVOS." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 117, + 479, + 358 + ], + "blocks": [ + { + "bbox": [ + 138, + 117, + 479, + 358 + ], + "lines": [ + { + "bbox": [ + 138, + 117, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 117, + 479, + 358 + ], + "type": "image", + "image_path": "f198866c59fc4b761c44b125f64729abd57f731ea2b2bb4f66a3bec8cbed0f5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 370, + 480, + 393 + ], + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 393 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparisons between the state-of-the-art models XMem and Swin-AOTL in challenging long-term videos (LVOS)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 400, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 478 + ], + "type": "text", + "content": "All-in-One Modeling. We initially evaluate the efficacy of All-in-One Modeling by analyzing lines 1, 2, and 3 of Table 3. The outcomes highlight the substantial benefits of All-in-One Modeling. Our model registers a " + }, + { + "bbox": [ + 130, + 400, + 482, + 478 + ], + "type": "inline_equation", + "content": "2.45\\%" + }, + { + "bbox": [ + 130, + 400, + 482, + 478 + ], + "type": "text", + "content": " improvement in performance on DAVIS 17 over the enhanced baseline. Moreover, in long video datasets like LVOS, even without any specialized module designs and using only the basic self-attention mechanism, our model exhibits only a minor decrease in performance, demonstrating robust performance of Unification modeling on complex long video datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 483, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 638 + ], + "type": "text", + "content": "Unidirectional Hybrid Attention. In our unidirectional hybrid attention module analysis, shown in lines 3 to 6 of Table 3. Notably, through an examination of lines 4 and 6 we observed that the integration of dual decoupling significantly enhances model performance on both DAVIS 17 val and LVOS datasets. However, Sole decoupling from the current to reference frame results in a performance drop. We analyze that the attributed to the increased focus of reference frame on memory tokens, thereby amplifying the ambiguity in mask embedding and hindering accurate target identification. On the other hand, decoupling solely from memory tokens to reference frame yields a marginal performance enhancement in DAVIS 17 val. However, this improvement is limited as the target features with objects information are represented by tokens from current frame without targets information, failing to provide a consistent feature representation across the video sequence so the performance drops in long videos. These findings highlight the necessity of dual decoupling is indispensable, and their joint effect promotes optimal model performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "type": "text", + "content": "Dynamic Tokens Selector. The comparison results between lines 6 and 7 of Table 3 highlights the role of DTS. DTS not only probes the segmentation mechanisms" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 114, + 463, + 198 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 463, + 198 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 463, + 198 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 463, + 198 + ], + "type": "table", + "html": "
MethodAll-in-OneDecoup1Decoup2DTSDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
182.879.785.960.856.465.1
285.482.488.465.660.875.1
3✓87.584.390.764.960.369.5
4✓✓87.184.090.161.055.766.2
5✓✓87.684.490.862.357.167.5
6✓✓✓88.584.692.467.462.072.9
7✓✓✓✓88.885.292.470.164.775.5
", + "image_path": "e76502f6d277285b9109ed5d7f315ff73dd686bf09a592dd560df315aa356ce0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 141, + 275, + 306, + 334 + ], + "blocks": [ + { + "bbox": [ + 131, + 206, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 206, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 482, + 262 + ], + "type": "text", + "content": "Table 3: Ablation on Core Design of OneVOS. (1-2) baseline: multi-object discrete method with ConvVIT or Swin Transformer [41] as backbone. (3-7): OneVOS with ConvVIT: 'All-in-One': All-in-One modeling. 'Decoup1': decouple attention from current to reference, 'Decoup2': decouple attention between memory and reference. 'DTS': DTS with dynamic memory." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 275, + 306, + 334 + ], + "lines": [ + { + "bbox": [ + 141, + 275, + 306, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 306, + 334 + ], + "type": "table", + "html": "
MethodDAVIS17 VALMOSE
J&FJFJ&FJF
first 6 layers87.683.791.460.956.865.1
last 6 layers88.084.491.659.155.163.1
all layers88.584.692.762.358.166.6
", + "image_path": "10a0d368f32b09d04a48dec32b5de551de9672ce969b82bcf663a9b16740b4ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 275, + 468, + 334 + ], + "blocks": [ + { + "bbox": [ + 317, + 275, + 468, + 334 + ], + "lines": [ + { + "bbox": [ + 317, + 275, + 468, + 334 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 468, + 334 + ], + "type": "table", + "html": "
MethodDAVIS 17 VALLVOS VAL
J&FJFJ&FJF
DTS-t87.283.690.866.461.271.6
DTS-Lr86.783.190.360.254.965.5
DTS88.885.292.470.164.775.5
", + "image_path": "14a2010c088b765b1390dd7d5f47a9ba987b6e341635e4edb6f3f6df76ed8946.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 343, + 480, + 365 + ], + "lines": [ + { + "bbox": [ + 131, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 343, + 480, + 365 + ], + "type": "text", + "content": "Table 4: From left to right: (a) Ablations on different mask embedding adding layers and (b) Ablation of different DTS designs in OneVOS." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 373, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 482, + 418 + ], + "type": "text", + "content": "within OneVOS but, when integrated into the network, it bolsters the learning of this paradigm. It contributes to a more efficient OneVOS by adaptively selecting tokens that align with the model's feature extraction and matching mechanisms, culminating in state-of-the-art performance on both datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 423, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 482, + 511 + ], + "type": "text", + "content": "Mask Embedding Add Layers. Table. 4(a) explores the effect of integrating mask embedding at different transformer layers. Results show that, by integrating mask embeddings to the immediate value tokens of each layer, the model can capture the global and local semantics of targets at different levels, thereby effectively promoting the model to perform multiple targets inference. Notably, embedding in the first six layers proved more beneficial than in the latter six, this indicates that early interaction between the current and reference frames without target information may negatively impact the feature extraction of the current frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 516, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 516, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 516, + 482, + 616 + ], + "type": "text", + "content": "DTS design. The role of DTS is to judiciously select tokens by amalgamating each token feature of the reference frame with the global features of all tokens. Table 4(b) explores the effect of various DTS configurations, including the method using pooled tokens from the current frame as global information (DTS-t), revealing that guiding selection with the current frame may hinder the generalization of OneVOS across frames, consistent with our initial attention decoupling. Besides, Implementing a loss function to maintain a fixed 0.5 selection rate per layer " + }, + { + "bbox": [ + 131, + 516, + 482, + 616 + ], + "type": "inline_equation", + "content": "(\\mathrm{DTS} - L_r)" + }, + { + "bbox": [ + 131, + 516, + 482, + 616 + ], + "type": "text", + "content": " significantly reduces performance, underscoring that rigid selection criteria detrimentally affect the capabilities of OneVOS for adaptive matching and feature extraction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 620, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 482, + 666 + ], + "type": "text", + "content": "Dynamic Memory Capacity. Various memory capacities (Cap in Eq. 8) impact accuracy, inference speed, and storage use. The performance of OneVOS under different memory capacities is validated in Table 5(a). The results reveal that while memory plays a crucial role in handling complex videos, OneVOS achieves robust performance" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 475, + 175 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 475, + 175 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 475, + 175 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 475, + 175 + ], + "type": "table", + "html": "
CapD17-VD17-TFPSMem(G)
187.583.216.61.72
288.084.413.72.20
388.484.612.92.47
488.584.612.32.69
588.584.812.02.94
", + "image_path": "2750de26832921275f13ea8c0c67a2b33e14bae7e9697b96c9364b3c60995702.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 182, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 228 + ], + "type": "text", + "content": "Table 5: Ablations on 3 specific module design of All-in-One framework. From left to right: (a) different " + }, + { + "bbox": [ + 130, + 182, + 481, + 228 + ], + "type": "inline_equation", + "content": "Cap" + }, + { + "bbox": [ + 130, + 182, + 481, + 228 + ], + "type": "text", + "content": " of dynamic memory, where D17-V and D17-T denote the " + }, + { + "bbox": [ + 130, + 182, + 481, + 228 + ], + "type": "inline_equation", + "content": "J\\&F" + }, + { + "bbox": [ + 130, + 182, + 481, + 228 + ], + "type": "text", + "content": " metrics of DAVIS-17 VAL and TEST datasets, respectively. (b) Different top-k tokens for dynamic memory, and (c) Different Transformer backbones of OneVOS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 238, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 480, + 282 + ], + "type": "text", + "content": "even with just two frames or a single frame in common scenarios, thanks to its unified framework and efficient architecture. Optimal performance is observed when the memory capacity reaches 4 or 5. This underscores the importance of memory in enabling OneVOS to effectively handle longer and more intricate videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 481, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 335 + ], + "type": "text", + "content": "Number of Retained Nodes. Table. 5(b) focuses on retaining a set number of top-k nodes in token memory, based on attention weights. Results show that after reaching a per-layer threshold, keeping approximately 512 nodes struck an optimal balance between information depth and computational efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 463 + ], + "type": "text", + "content": "Backbone. In Table. 5(c), we assess Vision Transformers (ViT) [15] pre-trained with Masked Autoencoders (MAE) [25] and ConvVIT (base and large variants) as backbones for OneVOS. With the VIT-B backbone OneVOS still outperforms most of the methods on D17 datasets. However, a key limitation of ViT was identified in its direct 16x downsampling approach, which resulted in a significant loss of multi-scale detail information. This is particularly detrimental for high-precision, pixel-level VOS tasks. On the other hand, the ConvVIT model, which employs simple convolutional operations for its patch embedding process, was found to effectively preserve multi-scale information. Additionally, the increased complexity of the ConvVIT-large variant further enhanced the model's performance, underscoring the importance of architectural design in capturing intricate details necessary for accurate VOS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 666 + ], + "type": "text", + "content": "This paper introduces OneVOS, a novel semi-supervised video object segmentation framework that globally integrates feature extraction, matching, memory management, and object aggregation within a single Transformer architecture. By employing the All-in-One Transformer with specialized Unidirectional Hybrid Attention modules, we significantly enhance the practicality and performance of OneVOS. Furthermore, to propose a rational, efficient version of OneVOS, we investigate the mechanism of the All-in-One Transformer via the proposed Dynamic Token Selector, revealing its layer-crossing segmentation mode for multi-object feature extraction and matching, leading to a naturally efficient version with dynamic memory. OneVOS demonstrates state-of-the-art performance on 7 datasets, particularly excelling in complex scenarios. We anticipate that this work will draw greater attention to the unified VOS modeling." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 507, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 507, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 507, + 174 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 191, + 197, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 191, + 197, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 197, + 204 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 216, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 216, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 481, + 259 + ], + "type": "text", + "content": "1. Bhat, G., Lawin, F.J., Danelljan, M., Robinson, A., Felsberg, M., Van Gool, L., Timofte, R.: Learning what to learn for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 777-794. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 261, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 293 + ], + "type": "text", + "content": "2. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "text", + "content": "3. Chen, B., Li, P., Bai, L., Qiao, L., Shen, Q., Li, B., Gan, W., Wu, W., Ouyang, W.: Backbone is all your need: A simplified architecture for visual object tracking. In: European Conference on Computer Vision. pp. 375-392. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 481, + 358 + ], + "type": "text", + "content": "4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 359, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 481, + 392 + ], + "type": "text", + "content": "5. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8126-8135 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 424 + ], + "type": "text", + "content": "6. Chen, Y., Pont-Tuset, J., Montes, A., Van Gool, L.: Blazingly fast video object segmentation with pixel-wise metric learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1189-1198 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 425, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 481, + 457 + ], + "type": "text", + "content": "7. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 458, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 500 + ], + "type": "text", + "content": "8. Cheng, H.K., Tai, Y.W., Tang, C.K.: Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5559-5568 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 533 + ], + "type": "text", + "content": "9. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 534, + 481, + 567 + ], + "type": "text", + "content": "10. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 481, + 599 + ], + "type": "text", + "content": "11. Cheng, M.M., Mitra, N.J., Huang, X., Torr, P.H., Hu, S.M.: Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence 37(3), 569-582 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "12. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13608-13618 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "13. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "text", + "content": "16. Duke, B., Ahmed, A., Wolf, C., Aarabi, P., Taylor, G.W.: Sstvos: Sparse spatiotemporal transformers for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5912-5921 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "17. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "18. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "19. Gao, J., Qian, X., Wang, Y., Xiao, T., He, T., Zhang, Z., Fu, Y.: Coarse-to-fine amodal segmentation with shape prior. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1262-1271 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "type": "text", + "content": "20. Gao, P., Ma, T., Li, H., Lin, Z., Dai, J., Qiao, Y.: Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "21. Gao, S., Zhou, C., Ma, C., Wang, X., Yuan, J.: Aiatrack: Attention in attention for transformer visual tracking. In: European Conference on Computer Vision. pp. 146-164. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "22. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "23. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "text", + "content": "24. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 523 + ], + "type": "text", + "content": "25. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "26. Hong, L., Chen, W., Liu, Z., Zhang, W., Guo, P., Chen, Z., Zhang, W.: Lvos: A benchmark for long-term video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13480-13492 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "text", + "content": "27. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "28. Hong, L., Zhang, W., Chen, L., Zhang, W., Fan, J.: Adaptive selection of reference frames for video object segmentation. IEEE Transactions on Image Processing 31, 1057-1071 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "29. Hu, P., Wang, G., Kong, X., Kuen, J., Tan, Y.P.: Motion-guided cascaded refinement network for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1400-1409 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Hu, Y.T., Huang, J.B., Schwing, A.G.: Videomatch: Matching based video object segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 54-70 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "text", + "content": "31. Huang, X., Xu, J., Tai, Y.W., Tang, C.K.: Fast video object segmentation with temporal aggregation network and dynamic template matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8879-8889 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 216 + ], + "type": "text", + "content": "32. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 258 + ], + "type": "text", + "content": "33. Johnander, J., Danelljan, M., Brissman, E., Khan, F.S., Felsberg, M.: A generative appearance model for end-to-end video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8953-8962 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "text", + "content": "34. Khoreva, A., Benenson, R., Ilg, E., Brox, T., Schiele, B.: Lucid data dreaming for video object segmentation. International Journal of Computer Vision 127(9), 1175-1197 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 336 + ], + "type": "text", + "content": "35. Kristan, M., Leonardis, A., Matas, J., Felsberg, M., Pflugfelder, R., Cehovin Zajc, L., Vojir, T., Bhat, G., Lukezic, A., Eldesokey, A., et al.: The sixth visual object tracking vot2018 challenge results. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "type": "text", + "content": "36. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "text", + "content": "37. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "text", + "content": "38. Li, X., Loy, C.C.: Video object segmentation with joint re-identification and attention-aware mask propagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 90-105 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 479 + ], + "type": "text", + "content": "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 522 + ], + "type": "text", + "content": "40. Lin, Z., Yang, T., Li, M., Wang, Z., Yuan, C., Jiang, W., Liu, W.: Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1362-1372 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 567 + ], + "type": "text", + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "type": "text", + "content": "42. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "type": "text", + "content": "43. Nowozin, S.: Optimal decisions from probabilistic models: the intersection-over-union case. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 548-555 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "44. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "45. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "46. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "text", + "content": "47. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "text", + "content": "48. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "text", + "content": "49. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "text", + "content": "50. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 629-645. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "text", + "content": "51. Seong, H., Oh, S.W., Lee, J.Y., Lee, S., Lee, S., Kim, E.: Hierarchical memory matching network for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12889-12898 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "52. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 429 + ], + "type": "text", + "content": "53. Voigtlaender, P., Chai, Y., Schroff, F., Adam, H., Leibe, B., Chen, L.C.: Feelvos: Fast end-to-end embedding learning for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9481-9490 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "type": "text", + "content": "54. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "55. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "text", + "content": "56. Wang, J., Chen, D., Wu, Z., Luo, C., Tang, C., Dai, X., Zhao, Y., Xie, Y., Yuan, L., Jiang, Y.G.: Look before you match: Instance understanding matters in video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2268-2278 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "57. Wang, W., Shen, J., Porikli, F., Yang, R.: Semi-supervised video object segmentation with super-trajectories. IEEE transactions on pattern analysis and machine intelligence 41(4), 985-998 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "text", + "content": "58. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14561-14571 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "59. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13879-13889 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "60. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 468 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "61. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "62. Xu, S., Liu, D., Bao, L., Liu, W., Zhou, P.: Mhp-vos: Multiple hypotheses propagation for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 314-323 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "63. Yan, S., Xu, X., Hong, L., Chen, W., Zhang, W., Zhang, W.: Panovos: Bridging non-panoramic and panoramic views with transformer for video segmentation. arXiv preprint arXiv:2309.12303 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "text", + "content": "64. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "type": "text", + "content": "65. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V. pp. 332-348. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "66. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "67. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(9), 4701-4712 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "text", + "content": "68. Yang, Z., Yang, Y.: Decoupling features in hierarchical propagation for video object segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "69. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "70. Zhou, X., Guo, P., Hong, L., Li, J., Zhang, W., Ge, W., Zhang, W.: Reading relevant feature from global representation memory for visual object tracking. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 447, + 101 + ], + "type": "text", + "content": "OneVOS: All-in-One Transformer Framework For VOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_content_list.json b/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fbaa09c22ceff5755309ee2136ef532fcb60a0a4 --- /dev/null +++ b/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_content_list.json @@ -0,0 +1,1655 @@ +[ + { + "type": "text", + "text": "Online Continuous Generalized Category Discovery", + "text_level": 1, + "bbox": [ + 269, + 140, + 733, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keon-Hee Park1, Hakyung Lee2, Kyungwoo Song2†, and Gyeong-Moon Park1†", + "bbox": [ + 274, + 210, + 727, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Kyung Hee University, Yongin, Republic of Korea {khpark, gmpark}@ku.ac.kr", + "bbox": [ + 326, + 253, + 673, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Yonsei University, Seoul, Republic of Korea {hakyunglee0417, kyungwoo.song}@yonsei.ac.kr", + "bbox": [ + 330, + 282, + 673, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. With the advancement of deep neural networks in computer vision, artificial intelligence (AI) is widely employed in real-world applications. However, AI still faces limitations in mimicking high-level human capabilities, such as novel category discovery, for practical use. While some methods utilizing offline continual learning have been proposed for novel category discovery, they neglect the continuity of data streams in real-world settings. In this work, we introduce Online Continuous Generalized Category Discovery (OCGCD), which considers the dynamic nature of data streams where data can be created and deleted in real time. Additionally, we propose a novel method, DEAN, Discovery via Energy guidance and feature Augmentation, which can discover novel categories in an online manner through energy-guided discovery and facilitate discriminative learning via energy-based contrastive loss. Furthermore, DEAN effectively pseudo-labels unlabeled data through variance-based feature augmentation. Experimental results demonstrate that our proposed DEAN achieves outstanding performance in proposed OCGCD scenario. The implementation code is available at https://github.com/KHU-AGI/OCGCD.", + "bbox": [ + 261, + 349, + 743, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Online Continual Learning $\\cdot$ Generalized Category Discovery $\\cdot$ Energy-Guided Discovery $\\cdot$ Variance-based Feature Augmentation", + "bbox": [ + 259, + 612, + 740, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 690, + 375, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid progress of deep neural networks in computer vision has greatly facilitated the widespread adoption of artificial intelligence into real-world applications. Nevertheless, the networks still have challenges in embodying high-level human capabilities such as object recognition and grouping. To emulate these sophisticated human abilities, there has been a lot of interest in the field of category discovery. Specifically, generalized category discovery (GCD) [40] demands", + "bbox": [ + 212, + 723, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 217, + 823, + 387, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/184125e7f3dba4888e0400978e84885bda23fd4536f509dabc7e170d2d8ea667.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 145, + 454, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4dafa45173861cdfd4c6f145b1996db797c0fe53d32488d56ae5b09ea596467f.jpg", + "image_caption": [ + "(a) Experiment of prior methods in an online manner on CUB200." + ], + "image_footnote": [], + "bbox": [ + 218, + 233, + 454, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c918eeb5a4da590fabc540a08cd5dec57cc4d20bedee7d748a2691c6be28b240.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 150, + 787, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1389858bf68b881018020a15319c4abc3ba03a9f0f2ecd036ee36d978dc9b438.jpg", + "image_caption": [ + "(b) The description of the proposed OCGCD.", + "Fig. 1: Figure 1a shows that existing methods recorded poor performance in online training, suggesting that prior methods cannot handle online continual learning. Figure 1b shows the proposed scenario, OCGCD. As our scenario assumes batch-wise online learning, the model suffers severe forgetting and poor novel category discovery." + ], + "image_footnote": [], + "bbox": [ + 470, + 233, + 785, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "advanced recognition and grouping skills to discover unknown categories within unlabeled datasets. In general, existing methods for GCD rely on an impractical assumption that they can access both labeled and unlabeled datasets simultaneously during training, which prevents these methods from meeting real-world application demands.", + "bbox": [ + 212, + 455, + 787, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this context, a new scenario of continuous category discovery [19,22,37,42, 44] has emerged to address the unrealistic assumption by adopting offline continual learning. Despite its great advantage not requiring the assumption, these offline-based methods have inherent limitations in handling the continuous nature of data streams in the real world. Current approaches to offline continual learning primarily focus on static data streams, thus failing to adequately accommodate the dynamic nature of data streams where data can be created and deleted in an online manner. To verify this limitation, we conducted an empirical analysis using the CUB200 [41] dataset, dividing it into 160 labeled categories and 40 unlabeled categories. We evaluated the performance of existing methods [22,37] for continuous category discovery in an online manner. As shown in Figure 1a, we observed that existing methods [22,37] exhibited inferior performance in identifying novel categories, thereby resulting in poor overall performances. This indicates that the offline-based GCD methods have difficulty in clustering novel data effectively in the online learning scenario.", + "bbox": [ + 212, + 535, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this observation, in this paper, we introduce a novel online learning scenario in continuous generalized category discovery, called Online Continuous Generalized Category Discovery (OCGCD). As shown in Figure 1b, OCGCD involves batch-wise training of unlabeled data in an online continual manner without any prior knowledge of the unknown data. Therefore, our novel", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 385, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scenario closely replicates real-world data streams characterized by dynamic creation and elimination of data during training [14]. Here we summarize the key challenges of our novel OCGCD scenario as follows: 1) Challenging category discovery. The model struggles to retain the knowledge of trained categories because of the continuous nature of the data, making novel category discovery challenging. 2) Noisy pseudo-labeling. Online learning restricts access to the entire training dataset, leading to inaccurate category assignments and the generation of severe noise pseudo-labels for unlabeled data. 3) Severe catastrophic forgetting. Online learning erases observed data after training, resulting in severe forgetting of previously acquired knowledge [33]. Hence, our novel OCGCD scenario requires addressing the aforementioned challenges to achieve successful learning outcomes.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the challenges of OCGCD, we propose a novel approach called DEAN, Discovery via Energy Guidance and Feature Augmentation, which utilizes the energy score [28] to discover unknown data and further discriminate whether the data is seen or unseen and employs feature augmentation for better pseudo-labeling. By comparing the energy values of known and unknown data, DEAN effectively identifies unknown data among them, facilitating the clustering of novel categories. On top of that, we introduce an energy-based contrastive loss to enhance learning from unknown samples, promoting the acquisition of more discriminative knowledge during online learning. In addition, we propose a novel variance-based feature augmentation, a simple yet effective approach for accurate novel clustering and pseudo-labeling. In our scenario, since the model can access only batch-wise training data at once, it struggles to cluster data accurately. Variance-based feature augmentation (VFA) temporally extends the number of feature vectors through variational information to promote accurate clustering. To mitigate catastrophic forgetting, we employ parameter-efficient tuning, especially LoRA [17] in our framework while preserving the pre-trained weights. Extensive experiments demonstrate that our proposed method, DEAN, enables the network to discover novel categories more effectively in the OCGCD scenario than existing methods.", + "bbox": [ + 212, + 328, + 787, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize the main contributions of this paper as follows:", + "bbox": [ + 240, + 616, + 689, + 631 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel Online Continuous Generalized Category Discovery (OCGCD) scenario that effectively reflects the online characteristics of real-world data. To tackle this scenario, we propose a novel method called Discovery via Energy Guidance and Feature Augmentation (DEAN).", + "- To the best of our knowledge, this is the first work to introduce the energy score for novel category discovery. We propose the energy-based contrastive loss to enhance online learning of discovered unknown data.", + "- For effective pseudo-labeling, we propose a new variance-based feature augmentation, called VFA. The proposed VFA enhances the sample clustering, leading to the improvement of pseudo-label quality.", + "- Comprehensive experiments demonstrate that our novel method achieves significant performance improvements in the OCGCD scenario compared to state-of-the-art models." + ], + "bbox": [ + 225, + 642, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Novel Category Discovery", + "text_level": 1, + "bbox": [ + 215, + 176, + 478, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Novel Category Discovery (NCD) focuses on the task of identifying and categorizing unlabeled data [39]. To effectively discover and classify data, this field requires advanced object recognition and clustering skills. Most of the NCD methods fall into two training scenarios: disjoint training and joint training. The disjoint training scenario, also known as a two-stage approach, starts by pre-training the model on a labeled dataset, followed by fine-tuning on an unlabeled dataset from a different domain [13, 15, 16, 29]. The joint training scenario adopts a more integrated approach, simultaneously utilizing both labeled and unlabeled datasets during the training process [7, 12, 18, 46, 47]. While this scenario outperforms disjoint training approach [29], its dependence on the availability of labeled data makes it less feasible in real-world applications. Accordingly, more research has been directed towards designing scenarios where they mitigate constraints of NCD for better alignment with real-world settings. Generalized Category Discovery (GCD) [40] broadens the setting of NCD by allowing test data to contain both known and unknown classes, in contrast to NCD which focuses on exclusively unknown classes. Moreover, to better represent real-world scenarios, several methods [19, 22, 37, 42, 44] have been proposed that integrate continual learning into category discovery. However, existing methods predominantly utilizes an offline continual learning framework, which remains an impractical assumption for real-world applications.", + "bbox": [ + 212, + 202, + 787, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose an online version of continuous GCD that tackles the challenges of online learning through energy-guided clustering and energy-based contrastive loss. Energy-guided clustering utilizes the trained classifier to extract energy from each sample, efficiently discovering novel categories. Moreover, applying energy-based contrastive loss enables the model to effectively learn feature knowledge, thereby enhancing the capacity of the model in online continual learning.", + "bbox": [ + 212, + 505, + 787, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Online Continual Learning", + "text_level": 1, + "bbox": [ + 215, + 618, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Online Continual Learning (OCL) is a more realistic continual learning scenario where the model learns from data arriving in small, sequential batches in real-time, without access to previously seen batches [30]. A key aspect of continual learning is to mitigate catastrophic forgetting and balance existing knowledge while adapting to new knowledge [8, 32, 33]. The dynamic distribution of online setting intensifies this challenge, and various approaches have been introduced to address this issue. MIR [1] proposed a memory retrieval technique for replay, selecting mostly interfered samples. OCM [10] deals with forgetting by maximizing mutual information between past and current data to learn more transferable features. i-Blurry [23] introduces a realistic setting that combines disjoint and blurry class distributions, mitigating forgetting by employing sample-wise importance sampling. Furthermore, Si-Blurry [34] proposed a new blurry task scenario, acknowledging the stochastic properties in real-world data distributions.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we propose a novel setting that integrates online continual learning with generalized category discovery. By emphasizing the online learning, OCGCD presents a new perspective in the field of category discovery. It is a more realistic and challenging scenario than existing approaches in GCD, taking into account the continuous flow of real-world data.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Noise Label", + "text_level": 1, + "bbox": [ + 214, + 243, + 361, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deep neural networks can overfit to noisy label data, leading to a substantial decrease in the robustness and generalization ability [2, 43]. Consequently, tackling noisy label data becomes essential, especially in category discovery where pseudo-labeling of unlabeled data plays a pivotal role in identifying new categories. Recent studies have primarily focused on discriminating between clean data and noisy label data. DivideMix [26] use a Gaussian Mixture Model (GMM) generated from the training dataset to differentiate between clean and noisy data, while AugDesc [35] applies various data augmentations to the training dataset to amplify the distinction. SplitNet [21] introduces a compact learnable module that separates clean and noisy label data.", + "bbox": [ + 212, + 267, + 787, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In OCGCD, learning occurs without prior knowledge of novel categories, relying only on the current batch-wise data to predict the number of clusters and generate pseudo-labels based on this prediction. This scenario makes it difficult to apply existing techniques that depend on utilizing the training dataset or employing learnable modules. Therefore, we propose Variance-based Feature Augmentation (VFA) to effectively separate clean and noisy labels in our framework. VFA can diversely augment (re-generate) the feature vectors of given data, enabling the creation of discriminative clusters for accurate pseudo-labeling. Unlike CGCD [22], which requires training additional modules for the discovery, VFA facilitates end-to-end training without the additional modules or extra learning.", + "bbox": [ + 212, + 419, + 789, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Online Continuous Generalized Category Discovery", + "text_level": 1, + "bbox": [ + 214, + 590, + 750, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation and Method Overview", + "text_level": 1, + "bbox": [ + 214, + 622, + 629, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Formulation. In online continuous GCD, the training dataset $\\mathcal{D} = \\{\\mathcal{D}^{base},\\mathcal{D}^{inc}\\}$ is given sequentially, where $\\mathcal{D}^{base} = \\{(x_i,y_i)\\in \\mathcal{X}\\times \\mathcal{Y}_l\\}_{i = 1}^{|D^{base}|}$ is the labeled dataset for the base session and $\\mathcal{D}^{inc} = \\{(x_i,y_i)\\in \\mathcal{X}\\times \\mathcal{Y}_u\\}_{i = 1}^{|D^{inc}|}$ is the unlabeled dataset for the incremental session. The proposed novel scenario assumes an overlap between the labeled category set $\\mathcal{V}_l$ and the unlabeled category set $\\mathcal{V}_u$ , as it is a sub-category of the generalized category discovery. In this setting, the network trains on $\\mathcal{D}^{base}$ and on $\\mathcal{D}^{inc}$ in an offline and online manner, respectively. Our proposed online continuous GCD aims to facilitate the discovery of novel categories in real-world scenarios while preserving the clustering knowledge obtained from the offline training of the labeled dataset.", + "bbox": [ + 212, + 646, + 787, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Method Overview. Figure 2 shows an overview of the proposed method, where the pre-trained vision transformer [6] via self-supervised learning serves as the", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2469bdf2f8a83daea607ee94603a102b607d121b41e0f330ae960e63db22fdfe.jpg", + "image_caption": [ + "Fig. 2: Overall process of the proposed DEAN framework. The energy-guided discovery splits unlabeled data into known, seen, and unseen data for better novel category discovery, while variance-based feature augmentation enhances the clustering of unseen data. $\\mathcal{L}_{ec}$ facilitates better discriminative learning in the online continual learning." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "backbone network. We initially train the network using the labeled dataset in the base session (Section 3.2). We introduce energy-guided discovery to identify unknown samples from the given data. Our energy-guided discovery involves a two-step process to determine whether data belong to known or unknown categories and whether the unknown data belong to previously seen or unseen categories (Section 3.3). We split the data into known, seen, and unseen categories through energy-guided discovery. For pseudo-labeling of unseen data, we propose variance-based feature augmentation to improve the clustering of unseen data, resulting in better pseudo-labeling (Section 3.4). Subsequently, the network trains pseudo-labeled data in an online manner using an energy-based contrastive loss for better discriminative learning (Section 3.5).", + "bbox": [ + 212, + 422, + 787, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Supervised Learning at Base Session", + "text_level": 1, + "bbox": [ + 215, + 608, + 565, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the base session, we train the offline model $\\theta^{off}$ consisting of the feature extractor $f^{off}(\\cdot):\\mathcal{X}\\to \\mathbb{R}^d$ and classifier head $g^{off}(\\cdot):\\mathbb{R}^d\\rightarrow \\mathcal{Y}_l$ in an offline manner. As the offline network $\\theta^{off} = \\{f^{off},g^{off}\\}$ trains on the labeled dataset during the base session, we employ the cross-entropy loss for supervised learning. While existing novel category discovery methods often utilize self-supervised [11] and supervised [20] contrastive loss for representation learning without a classifier head, our approach utilizes a classifier for clustering. The formulation of the base session training is as follows:", + "bbox": [ + 212, + 628, + 787, + 751 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = - \\mathbb {E} _ {(\\mathcal {X}, \\mathcal {Y} _ {l}) \\sim \\mathcal {D} ^ {b a s e}} \\frac {1}{| \\mathcal {D} ^ {b a s e} |} \\sum_ {i = 1} ^ {| \\mathcal {D} ^ {b a s e} |} y _ {i} \\cdot \\log \\delta (g ^ {\\text {o f f}} (f ^ {\\text {o f f}} (x _ {i}))), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 756, + 785, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta(\\cdot)$ is the softmax function. We train the model to capture representation knowledge during the base session training. After training the base session, we", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a06de2bd6aa5851dd24a2615dc12694c5fbcecd03d7f3b181a0ebcdb05407e6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 460, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6651a230f898248f0056e6c40976e2efd380444e632109993318f18b3a90e3b2.jpg", + "image_caption": [ + "(a) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using fine-split of CGCD.", + "(c) Comparing normal distributions of known and unknown samples on CUB200 using fine-split of CGCD." + ], + "image_footnote": [], + "bbox": [ + 217, + 325, + 460, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/423210ebdc5c836cb61062f186d84d91b4f848a9b95c54a3de3d635a61a7f800.jpg", + "image_caption": [ + "(b) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using energy scores." + ], + "image_footnote": [], + "bbox": [ + 537, + 143, + 785, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d9bb272b821e9069b802168bec71a34348acd5b39d884d799d17b5797798278.jpg", + "image_caption": [ + "(d) Comparing normal distributions of known and unknown samples on CUB200 using energy scores.", + "Fig. 3: Validating the effectiveness of energy scores for novel category discovery without prior knowledge about novel categories by comparison with existing methods." + ], + "image_footnote": [], + "bbox": [ + 537, + 325, + 785, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "freeze the offline model and utilize the model at the discovery step to distinguish between known and unknown data, enabling accurate discovery. In addition, we initialize the online model, denoted as $\\theta^{on}$ , based on the parameters of the offline model for the online continual learning on the unlabeled dataset.", + "bbox": [ + 212, + 575, + 787, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Energy-Guided Discovery to Identify Unknowns", + "text_level": 1, + "bbox": [ + 214, + 661, + 658, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The discovery step aims to discover new categories within the provided batchwise data. While CGCD, which assumes the absence of prior knowledge about novel categories, introduced a two-stage split process that depends on learning an additional split network to detect novel data, it has some limitations in our proposed scenario. CGCD includes the requirement for extra parameters and additional learning stage. Moreover, in our proposed online scenario where the model cannot access the entire training data, the split network faces difficulty in identifying novel samples due to severe forgetting of the split network. Liu et al. [28] proposed a method for detecting out-of-distribution (OOD) data by analyzing the difference in energy scores between in-distribution (ID) and OOD", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data. We conducted experiments to compare the two-stage split process from CGCD with the energy-based discovery method to identify trained categories (known to the model) and untrained categories (unknown to the model).", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 3, we found that the two-stage split of the CGCD struggled to identify both known and unknown categories in our proposed online learning scenario. Online learning with batch-wise data led to severe forgetting in the split network, resulting in poor detection of unknown samples. In contrast, energy-based discovery showed better performance in novel category discovery without extra parameters compared to CGCD. Moreover, as it does not require an additional learning phase, energy-based discovery enables end-to-end training. Inspired by this observation, we propose an energy-guided discovery approach for novel category discovery. To the best of our knowledge, this is the first work to utilize the energy score for novel category discovery.", + "bbox": [ + 212, + 193, + 784, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the incremental session, the network trains unlabeled batch-wise data $B^{t} = \\{x_{1}, x_{2}, \\ldots, x_{|B^{t}|}\\}$ , where $t$ is the index of the batch. We classify unlabeled data $B^{t}$ into known data, seen data, and unseen data using both the $\\theta^{off}$ and $\\theta^{on}$ through a two-stage process. In the first stage, we split the data into known and unknown categories. We obtain logits $z$ using the offline model $\\theta^{off}(\\cdot) = g^{off}(f^{off}(\\cdot))$ and calculate the energy scores using these logits. The formulation for calculating the energy scores is as follows:", + "bbox": [ + 214, + 347, + 785, + 454 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {t} = \\theta^ {\\text {o f f}} \\left(B ^ {t}\\right) \\in \\mathbb {R} ^ {\\left| B ^ {t} \\right| \\times \\left| \\mathcal {Y} _ {l} \\right|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 469, + 785, + 489 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {t} = \\left\\{e _ {i} ^ {t} \\mid e _ {i} ^ {t} = - \\log \\sum_ {j = 1} ^ {| \\mathcal {Y} _ {i} |} \\exp \\left(z _ {i, j} ^ {t}\\right) \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 494, + 785, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\exp (\\cdot)$ denotes an exponential function, $e^t$ denotes a set of energy scores, and $z_{i,j}$ denotes the logit from $i$ -th sample corresponding to the $j$ -th class. We divide the energy scores by the Gaussian Mixture Model (GMM). We initially set the GMM to have two clusters and use the energy score of each sample as input for clustering. We then label the cluster with a low energy score as \"known\" and the cluster with a high energy score as \"unknown\". This division allows us to split the unlabeled data into known and unknown data. In the second stage, we divide the unknown data into seen and unseen categories. Seen data comprise the data belonging to observed novel categories during the incremental session, while unseen data belong to undiscovered categories. To split the unknown data, we utilize the online model $\\theta^{on}(\\cdot) = g^{on}(f^{on}(\\cdot))$ to extract energy scores. The process in the second stage is the same as the first stage, where we split the unknown data into seen and unseen categories based on their energy scores. For the initial batch of incremental sessions, the online model is identical to the offline model. Additionally, since the initial batch data is the first data of the incremental session, any data classified as unknown is assumed to be unseen data. The proposed energy-guided discovery splits unlabeled data into known, seen, and unseen categories effectively. Unlike prior methods which require sufficient data, it can identify novel categories with batch-wise data.", + "bbox": [ + 214, + 551, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 385, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Pseudo-Labeling via Variance-based Feature Augmentation", + "text_level": 1, + "bbox": [ + 215, + 146, + 750, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After the discovery step, we pseudo-label known, seen, and unseen data to assign labels to each sample. For known and seen data, where the model has already learned knowledge for the categories, we utilize predictions from the offline model for known data and from the online model for seen data to assign pseudo-labels, respectively. Conversely, as the unseen data belong to undiscovered categories, the online model cannot predict proper categories for them and needs to discover novel categories. Moreover, due to the given batch-wise data, discovering novel categories to assign pseudo labels is challenging.", + "bbox": [ + 212, + 172, + 787, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address this challenge, we propose variance-based feature augmentation, which temporally augments feature vectors using variational knowledge from unseen data to acquire accurate novel clusters. We obtain feature vectors of the unseen data $B_{u}^{t}$ using the online feature extractor $f^{on}(\\cdot)$ denoted as $h_u^t = f^{on}(B_u^t)$ . We calculate the variance of the unseen feature vectors and augment unseen feature vectors based on the variance. We employ each feature vector to augment features by the Gaussian distribution $\\hat{h}_u^t = \\{\\hat{h}_u\\sim \\mathcal{N}(h_u,\\sigma_u^2)\\} ,h_u\\in h_u^t$ . The formulation of the variance-based feature augmentation is as follows:", + "bbox": [ + 214, + 292, + 787, + 414 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nh _ {a u g} ^ {t} = \\{h _ {u} ^ {t}; \\hat {h} _ {u} ^ {t} \\}, \\quad \\hat {h} _ {u} ^ {t} = \\{\\hat {h} _ {u} \\sim \\mathcal {N} (h _ {u}, \\sigma_ {u} ^ {2}) \\} \\in \\mathbb {R} ^ {(\\left| B _ {u} ^ {t} \\right| \\cdot K) \\times d}, \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 425, + 785, + 446 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{h}_u$ and $\\sigma_u$ represent augmented feature vectors and the standard deviation of the unseen feature vectors, respectively. $K$ denotes the hyperparameter for the number of augmented feature vectors, and $h_{aug}^t$ denotes the concatenation of unseen feature vectors and augmented feature vectors. Novel categories are discovered based on $h_{aug}^t$ . Augmented feature vectors expand the feature space with variational knowledge, which helps the model discover novel clusters.", + "bbox": [ + 212, + 458, + 787, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We employ affinity propagation [9] to assign pseudo-labels to unseen data. Affinity propagation is a non-parametric clustering algorithm that does not require a predefined number of clusters, using the similarity matrix of the data. Given the absence of prior knowledge of novel categories, we extend the online classifier with the estimated number of novel clusters obtained by affinity propagation. The proposed variance-based feature augmentation is a simple yet effective approach that helps the model assign effective pseudo-labels based on accurate novel category discovery.", + "bbox": [ + 212, + 550, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Online Continual Learning with Pseudo-Labeled Data", + "text_level": 1, + "bbox": [ + 214, + 693, + 705, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For online continual learning, we employ parameter-efficient tuning to learn novel representation knowledge while preserving previously learned knowledge, as it effectively mitigates catastrophic forgetting. We use cross-entropy loss with pseudo-labeled data in an online manner to capture novel knowledge. However, adapting the model with batch-wise data raises recency bias, where sufficient training through the given batch-wise data impedes discriminative knowledge. To address this, we propose an energy-based contrastive loss $(\\mathcal{L}_{ec})$ for effective discriminative learning. The energy-based contrastive loss focuses on inducing", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the online model to capture the knowledge of novel categories. It amplifies the energy from the known classifier while reducing the energy from the novel classifier. By selectively applying $\\mathcal{L}_{ec}$ to novel data $B_{n}^{t}$ including seen and unseen data, it helps the online model learn novel discriminative knowledge. The formulation of $\\mathcal{L}_{ec}$ is as follows:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E} (f (x); g) = - \\log \\sum_ {i = 1} ^ {| \\mathcal {Y} |} \\exp \\left(g _ {i} (f (x))\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 234, + 785, + 277 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e c} = \\frac {1}{| B _ {n} ^ {t} |} \\sum_ {(x _ {n}) \\sim B _ {n} ^ {t}} \\log \\left(1 + \\frac {\\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {n e w})}{\\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {o l d})}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 301, + 785, + 342 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\operatorname{E}(\\cdot ;\\cdot)$ and $|\\mathcal{V}|$ denote energy function and the number of nodes at the classifier, respectively, $g_{i}(\\cdot)$ represents the $i$ -th node from the given classifier $g$ , and $g^{old}$ and $g^{new}$ denote the nodes corresponding to known categories and discovered novel categories in the online classifier $g^{on}(\\cdot)$ , respectively. Eq. 5 represents the calculation of the energy score given the feature vector and classifier. The total loss $\\mathcal{L}_{inc}$ for the incremental session can be summarized as follows:", + "bbox": [ + 214, + 352, + 787, + 444 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n c} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {e c}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 458, + 785, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{ec}$ represents energy-based contrastive loss, and $\\mathcal{L}_{ce}$ denotes cross-entropy loss with pseudo-labeled data.", + "bbox": [ + 214, + 484, + 789, + 515 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 539, + 375, + 556 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 571, + 428, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines and Implementation Details. Recently, many approaches have emerged in novel category discovery. However, many of them focus on self-supervised learning with non-parametric classifiers (e.g. K-means) [42] and rarely consider continual learning [45]. Some methods that consider continual learning assume prior knowledge for novel category discovery. Since the proposed EGD and fine-split from CGCD require a classifier to estimate the number of novel categories, non-parametric classifiers cannot be used. Additionally, existing methods not designed for continual learning are unsuitable baselines due to severe forgetting of known classes. For a fair comparison, we mainly compared our method with FRoST [37] and CGCD [22], which focus on continual learning and utilize parametric classifiers for novel category discovery.", + "bbox": [ + 212, + 598, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We used ViT-B-16 [6] pre-trained on DINO-ImageNet [3] as the backbone network for all methods, including ours. We trained our method 30 epochs for the base session and 15 epochs for the incremental session. We utilized the AdamW optimizer with a weight decay of 1e-4 and a learning rate of 1e-3. We trained our model using a single RTX 3090 GPU with a batch size set to 64.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We evaluated our method with state-of-the-art (SoTA) continuous category discovery methods on three datasets: CUB200 [41], FGVC-Aircraft [31], and CIFAR-100 [24]. For continual learning, we partitioned CUB200 and FGVC-Aircraft following the same split configuration in CGCD. For CIFAR-100, we divided 80 classes for known classes and 20 classes for novel classes. To follow the GCD setting, we divided the training samples into labeled and unlabeled per class, maintaining a ratio of 0.8 for the labeled to 0.2 for the unlabeled.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluated the methods by measuring clustering accuracy on test dataset $\\mathcal{D}$ . We utilized the hungarian algorithm [25] to obtain the optimal permutation $h^*$ that aligns the prediction $y_i^*$ from the model with the ground truth label $y_i$ :", + "bbox": [ + 212, + 252, + 782, + 313 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {*} = \\arg \\min _ {h} \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h \\left(y _ {i} ^ {*}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 325, + 784, + 367 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $|\\mathcal{D}|$ is the size of the test dataset. Clustering accuracy on $\\mathcal{D}$ is defined as follows:", + "bbox": [ + 212, + 377, + 782, + 407 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nM = \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h ^ {*} \\left(y _ {i} ^ {*}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 417, + 784, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We employed this metric to measure the clustering accuracy of all classes ( $M_{all}$ ), known classes ( $M_{old}$ ), and unknown classes ( $M_{new}$ ) in the cases of test dataset $\\mathcal{D}^{base}$ and $\\mathcal{D}^{inc}$ , respectively. Following continuous GCD approaches to evaluate the ability to maintain performance for known classes [22, 44], we employed the forgetting metric $F = M_{old}^{base} - M_{old}^{inc}$ , where $M_{old}^{base}$ , $M_{old}^{inc}$ refers to known class clustering accuracy at the base session and incremental session. A higher clustering accuracy metric is better, and a lower forgetting metric is better. This ensures the model is effective for real-world applications.", + "bbox": [ + 212, + 470, + 782, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, we evaluate the estimated pseudo-labels to validate their effectiveness in pseudo-labeling. Pseudo-labeling for unlabeled data is crucial as the model utilizes pseudo-labels to interpret the unlabeled data. The pseudo-labeling accuracy on the dataset $\\mathcal{D}$ is defined as follows:", + "bbox": [ + 212, + 592, + 782, + 651 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nM = \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h ^ {*} \\left(\\tilde {y} _ {i}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 664, + 784, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\tilde{y}_i$ denotes the pseudo-label of the $i$ -th sample.", + "bbox": [ + 214, + 715, + 598, + 732 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparison with Baselines", + "text_level": 1, + "bbox": [ + 214, + 753, + 486, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compared the clustering performance of our proposed method, DEAN, with SoTA approaches, as shown in Table 1. Since we considered an online scenario for category discovery, traditionally approached using offline methods, there is no previous work for comparison. Therefore, for a fair comparison, our study adapts", + "bbox": [ + 212, + 779, + 782, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/73b12beb49ff9c948f0eaa6dd4a5433f54997083847e736c161cdadfeabc3f7f.jpg", + "table_caption": [ + "Table 1: Clustering performance comparison on CUB200, FGVC-Aircraft, and CIFAR-100. Best performance is highlighted in bold. All the baselines utilize the fine-split method from CGCD to handle the absence of prior knowledge." + ], + "table_footnote": [], + "table_body": "
MethodCUB200FGVC-AircraftCIFAR-100
MallMoldMnewF↓MallMoldMnewF↓MallMoldMnewF↓
Supervised72.5170.6879.6611.9166.2563.6276.7611.0381.3479.4888.8010.07
Fine-Tuning1.961.942.0180.522.061.305.1073.353.683.245.4586.26
FRoST [ECCV' 22]23.9828.725.4550.8514.4415.659.6048.316.136.215.8069.24
CGCD[ICCV' 23]40.4047.8811.1835.6114.0815.677.7560.002.642.762.1586.74
DEAN w/ FS55.5765.3717.2917.6240.3249.145.1025.4352.2762.1112.9027.55
DEAN (Ours)59.5666.3333.1116.2646.4453.5617.9920.9762.3470.3430.3519.17
", + "bbox": [ + 223, + 198, + 787, + 342 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6ddf3724fd900dd449ccd851e0ddc9b3cf773ae6db00c4d982f75ac6d015f9da.jpg", + "table_caption": [ + "Table 2: Ablation experiment on CUB200. The baseline denotes fine-tuning with the fine split from the CGCD." + ], + "table_footnote": [], + "table_body": "
AblationCUB200
MallMoldNewF↓
Baseline1.961.942.0180.52
+ LoRA49.3360.356.2422.24
+ EGD54.9762.9823.6719.61
+ EC Loss57.3764.9827.6317.61
+ VFA (Ours)59.5666.3333.1116.26
", + "bbox": [ + 223, + 412, + 504, + 537 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6dd8acbd88e15d10aade891c7d282c91859dcaff3f55c92a1a14b9d05a3c7e3e.jpg", + "table_caption": [ + "Table 3: Ablation study for the hyperparameter K of variance-based feature augmentation on CUB200." + ], + "table_footnote": [], + "table_body": "
AblationCUB200
\\(M_{all}^{PS}\\)\\(M_{old}^{PS}\\)\\(M_{new}^{PS}\\)
K=057.3185.1946.47
K=164.2788.9653.74
K=375.8793.4364.48
K=576.9394.0565.75
K=775.4493.4164.07
K=975.2293.563.36
", + "bbox": [ + 535, + 410, + 787, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "offline continuous NCD and GCD methods into OCGCD scenarios. The experiment includes two SoTA continuous category discovery methods: FRoST [37] and CGCD [22]. In addition, we include a supervised learning method as an upper bound and a fine-tuning method as a performance baseline. Note that in our proposed scenario, OCGCD, all baselines are not provided with prior knowledge about novel categories. Instead, we employed fine-split method from CGCD [22] which consists of two stages: initial split and fine split. During the initial split stage, unlabeled data is roughly divided into known and unknown categories based on cosine similarity. In the fine split stage, a split model is trained to distinguish between known and unknown data. The fine-split method estimates the number of novel categories conveying information about novel categories to the comparison methods.", + "bbox": [ + 212, + 566, + 787, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results show that our method, DEAN, recorded a robust clustering performance, significantly surpassing existing continuous NCD and GCD methods. DEAN recorded consistently high metrics across $M_{all}$ , $M_{old}$ , and $M_{new}$ while keeping forgetting metric $F$ substantially low. This denotes our method can learn novel knowledge without forgetting. Moreover, it is noteworthy that DEAN w/ FS which denotes the adoption of fine-split from CGCD into DEAN instead", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9c72b61ec52edae09619d959e4b0bdcea9ebe7876725793e68b22e7f36b8a89e.jpg", + "table_caption": [ + "Table 4: Ablation study for the cluster accuracy of pseudo-labels." + ], + "table_footnote": [], + "table_body": "
MethodCUB200
\\( M_{all}^{PS} \\)\\( M_{old}^{PS} \\)\\( M_{new}^{PS} \\)
CGCD [ICCV' 23]61.8881.0758.68
DEAN w/o (\\( \\mathcal{L}_{ec} \\), VFA)57.2385.1145.80
DEAN w/o VFA57.3185.1946.47
DEAN (Ours)76.9394.0565.75
", + "bbox": [ + 223, + 185, + 488, + 282 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3489abb60324c93ef2eda66b2f7d1c357b747f8af2cf28c2dc23a39204f8896b.jpg", + "table_caption": [ + "Table 5: Ablation experiment of parameter-efficient tuning on CUB200." + ], + "table_footnote": [], + "table_body": "
MethodCUB200
\\(M_{all}\\)\\(M_{old}\\)\\(M_{new}\\)\\(F\\downarrow\\)
DEAN(Ours)Adapter [4]56.9664.7526.5317.64
Prefix [27]58.5466.2928.2816.30
LoRA [17]59.5666.3333.1116.26
", + "bbox": [ + 516, + 184, + 787, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of energy-guided discovery (EGD) also showed notable performance compared to the baselines. But, DEAN achieved significant performance gains across all datasets as the proposed EGD is applied. In contrast, existing continual methods showed shortcomings in clustering performance. Specifically, we observed extremely low clustering performance on CIFAR-100 dataset, low $M_{all}$ , $M_{old}$ , and $M_{new}$ values, coupled with high $F$ . On CIFAR-100 dataset, DEAN surpassed the most recent method CGCD assuming no prior knowledge for novel categories by $59.7\\%$ , $67.58\\%$ , $28.2\\%$ , and $67.57\\%$ for $M_{all}$ , $M_{old}$ , $M_{new}$ , and $F$ .", + "bbox": [ + 212, + 311, + 785, + 431 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Such observations show that while current GCD methods are vulnerable to online settings, DEAN can manage the balance required in the OCGCD scheme of retaining old knowledge while integrating the new. We believe this results shows the suitability of DEAN for real-world applications, where data is never static. We empirically confirmed that DEAN can evolve with the data it encounters, overcoming the challenges presented in the OCGCD scheme.", + "bbox": [ + 212, + 431, + 785, + 523 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 542, + 388, + 559 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The experimental results in Section 4.2 show that our DEAN achieved significant performance in OCGCD scenarios. Through further experiments on the CUB200, we provide a detailed analysis of the components proposed in our method.", + "bbox": [ + 212, + 566, + 785, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Components. We validate the effectiveness of adding each component of our method to the baseline model. The baseline denotes fine-tuning the backbone model with a data-split following settings in CGCD. As shown in Table 2, results reveal that all components have an important impact on performance. Employing LoRA to baseline shows significant improvement in $M_{old}$ , $F$ . This shows that our application of LoRA on the frozen model at incremental session effectively mitigates forgetting. While LoRA shows only a slight increase in $M_{new}$ , additionally employing energy-guided discovery (EGD) and EC Loss markedly improves $M_{new}$ . This confirms that adopting energy guidance will enhance the ability of the model to differentiate between old and new categories. Finally, the integration of VFA further boosts the model performance on $M_{new}$ . This improvement reveals the importance of accurate pseudo-labeling since VFA is a technique applied after training. Each component is crucial, incrementally building upon the last to solidify the model's performance in our online category discovery scenario.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hyperparameter Analysis of VFA. Variance-based Feature Augmentation (VFA) is our proposed method of temporarily augmenting feature vectors to acquire accurate pseudo-labeling. Our analysis evaluated the effect of hyperparameter $K$ , the number of augmented feature vectors, on the accuracy of pseudolabels. From Table 3 we observed that employment of VFA is beneficial, increasing novel category pseudo-label accuracy from $46.47\\%$ to $65.75\\%$ . Among VFA employment, we observed that $K = 5$ yields the best performance across all metrics. When $K$ is less than 5, the smaller $K$ worsened performance. Conversely, values of $K$ greater than 5 did not benefit the model. We analyzed that since VFA augments features based on variance, as the number of $K$ increases, unseen samples can be clustered together without being distinguished.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Pseudo-Labeling Accuracy. Accurate pseudo-labeling for clustering is important because it directly influences the performance of the model in incremental learning. Shown in Table 4, our proposed method, DEAN, significantly enhanced the clustering accuracy of pseudo-labels, achieving a $13.12\\%$ improvement in $M_{all}^{PS}$ over the most recent method CGCD. Through the ablation study, we investigated the contributions of each step in DEAN. When DEAN adopts only parameter-efficient tuning, excluding EC loss and VFA, we observed a notable declination in $M_{new}^{PS}$ . While each ablation step showed gradually accuracy enhancement, the application of VFA elevates the performance of DEAN above CGCD. Thus, the experimental result revealed that VFA played a vital role in distinguishing novel categories and confirming the effectiveness of our method.", + "bbox": [ + 212, + 313, + 787, + 478 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Parameter Efficient Tuning. The ablation study in Table 5 assesses the impact of parameter-efficient tuning (PET) methods on the CUB200 dataset. Our method can employ various tuning methods such as Adapter [4] or Prefix-tuning [27], and in this experiment, we compared modules with LoRA [17], Adapter and Prefix-tuning. We set the size of all PET modules size to 5 and integrated them into the last 5 layers of ViT. Overall improvements are shown by adopting PET over existing methods reported in Table 1. Among PET methods, LoRA showed the highest clustering accuracy. The performance enhancement of LoRA may stem from its fast adaptation, which is highly effective in online continual learning scenarios that often depend on fewer samples for training.", + "bbox": [ + 212, + 479, + 787, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 655, + 359, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel scenario for category discovery, Online Continuous Generalized Category Discovery, which accounts for the dynamic nature of real-world data streams, leading to severe forgetting and poor novel category discovery. In addition, we proposed a novel framework, DEAN, which utilizes energy-based novel category discovery for the first time in this field and introduces variance-based feature augmentation to enhance accurate pseudo-labeling. Our DEAN demonstrated promising performance in the proposed novel scenario. However, DEAN could not guarantee inference performance midway through training. In future work, we plan to explore an online continual framework that can ensure inference performance during training through fast adaptation.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)).", + "bbox": [ + 212, + 179, + 787, + 303 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 327, + 323, + 343 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aljundi, R., Belilovsky, E., Tuytelaars, T., Charlin, L., Caccia, M., Lin, M., Page-Caccia, L.: Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32 (2019)", + "2. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: International conference on machine learning. pp. 233-242. PMLR (2017)", + "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)", + "4. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)", + "5. Choe, S.A., Shin, A.H., Park, K.H., Choi, J., Park, G.M.: Open-set domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23943-23953 (2024)", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)", + "7. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021)", + "8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999)", + "9. Frey, B.J., Dueck, D.: Clustering by passing messages between data points. science 315(5814), 972-976 (2007)", + "10. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: International Conference on Machine Learning. pp. 8109-8126. PMLR (2022)", + "11. Gutmann, M., Hyvärinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)" + ], + "bbox": [ + 217, + 361, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. arXiv preprint arXiv:2002.05714 (2020)", + "13. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019)", + "14. He, J., Mao, R., Shao, Z., Zhu, F.: Incremental learning in online scenario. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13926-13935 (2020)", + "15. Hsu, Y.C., Lv, Z., Kira, Z.: Learning to cluster in order to transfer across domains and tasks. arXiv preprint arXiv:1711.10125 (2017)", + "16. Hsu, Y.C., Lv, Z., Schlosser, J., Odom, P., Kira, Z.: Multi-class classification without multi-class labels. arXiv preprint arXiv:1901.00544 (2019)", + "17. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)", + "18. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021)", + "19. Joseph, K., Paul, S., Aggarwal, G., Biswas, S., Rai, P., Han, K., Balasubramanian, V.N.: Novel class discovery without forgetting. In: European Conference on Computer Vision. pp. 570-586. Springer (2022)", + "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)", + "21. Kim, D., Ryoo, K., Cho, H., Kim, S.: Splitnet: learnable clean-noisy label splitting for learning with noisy labels. arXiv preprint arXiv:2211.11753 (2022)", + "22. Kim, H., Suh, S., Kim, D., Jeong, D., Cho, H., Kim, J.: Proxy anchor-based unsupervised learning for continuous generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16688-16697 (2023)", + "23. Koh, H., Kim, D., Ha, J.W., Choi, J.: Online continual learning on class incremental blurry task configuration with anytime inference. arXiv preprint arXiv:2110.10031 (2021)", + "24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "25. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)", + "26. Li, J., Socher, R., Hoi, S.C.: Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020)", + "27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597 (2021)", + "28. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020)", + "29. Liu, Y., Tuytelaars, T.: Residual tuning: Toward novel category discovery without labels. IEEE Transactions on Neural Networks and Learning Systems (2022)", + "30. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022)" + ], + "bbox": [ + 217, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K.-H. Park et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)", + "32. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995)", + "33. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)", + "34. Moon, J.Y., Park, K.H., Kim, J.U., Park, G.M.: Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11731-11741 (2023)", + "35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021)", + "36. Park, K.H., Song, K., Park, G.M.: Pre-trained vision and language transformers are few-shot incremental learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23881-23890 (2024)", + "37. Roy, S., Liu, M., Zhong, Z., Sebe, N., Ricci, E.: Class-incremental novel class discovery. In: European Conference on Computer Vision. pp. 317-333. Springer (2022)", + "38. Seo, J., Lee, S.H., Lee, T.Y., Moon, S., Park, G.M.: Generative unlearning for any identity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9151-9161 (2024)", + "39. Troisemaine, C., Lemaire, V., Gosselin, S., Reiffers-Masson, A., Flocon-Cholet, J., Vaton, S.: Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028 (2023)", + "40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022)", + "41. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "42. Wu, Y., Chi, Z., Wang, Y., Feng, S.: Metagcd: Learning to continually learn in generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1655-1665 (2023)", + "43. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)", + "44. Zhang, X., Jiang, J., Feng, Y., Wu, Z.F., Zhao, X., Wan, H., Tang, M., Jin, R., Gao, Y.: Grow and merge: A unified framework for continuous categories discovery. Advances in Neural Information Processing Systems 35, 27455-27468 (2022)", + "45. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021)", + "46. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021)", + "47. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 834 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Online Continuous Generalized Category Discovery", + "bbox": [ + 388, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_model.json b/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cba40f9c1ac17c1cfae026b012315c608b4d4f1c --- /dev/null +++ b/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_model.json @@ -0,0 +1,2258 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.27, + 0.141, + 0.734, + 0.187 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.212, + 0.728, + 0.244 + ], + "angle": 0, + "content": "Keon-Hee Park1, Hakyung Lee2, Kyungwoo Song2†, and Gyeong-Moon Park1†" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.255, + 0.674, + 0.283 + ], + "angle": 0, + "content": "1 Kyung Hee University, Yongin, Republic of Korea {khpark, gmpark}@ku.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.283, + 0.674, + 0.311 + ], + "angle": 0, + "content": "2 Yonsei University, Seoul, Republic of Korea {hakyunglee0417, kyungwoo.song}@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.35, + 0.744, + 0.601 + ], + "angle": 0, + "content": "Abstract. With the advancement of deep neural networks in computer vision, artificial intelligence (AI) is widely employed in real-world applications. However, AI still faces limitations in mimicking high-level human capabilities, such as novel category discovery, for practical use. While some methods utilizing offline continual learning have been proposed for novel category discovery, they neglect the continuity of data streams in real-world settings. In this work, we introduce Online Continuous Generalized Category Discovery (OCGCD), which considers the dynamic nature of data streams where data can be created and deleted in real time. Additionally, we propose a novel method, DEAN, Discovery via Energy guidance and feature Augmentation, which can discover novel categories in an online manner through energy-guided discovery and facilitate discriminative learning via energy-based contrastive loss. Furthermore, DEAN effectively pseudo-labels unlabeled data through variance-based feature augmentation. Experimental results demonstrate that our proposed DEAN achieves outstanding performance in proposed OCGCD scenario. The implementation code is available at https://github.com/KHU-AGI/OCGCD." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.613, + 0.741, + 0.643 + ], + "angle": 0, + "content": "Keywords: Online Continual Learning \\(\\cdot\\) Generalized Category Discovery \\(\\cdot\\) Energy-Guided Discovery \\(\\cdot\\) Variance-based Feature Augmentation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.377, + 0.708 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.817 + ], + "angle": 0, + "content": "The rapid progress of deep neural networks in computer vision has greatly facilitated the widespread adoption of artificial intelligence into real-world applications. Nevertheless, the networks still have challenges in embodying high-level human capabilities such as object recognition and grouping. To emulate these sophisticated human abilities, there has been a lot of interest in the field of category discovery. Specifically, generalized category discovery (GCD) [40] demands" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.388, + 0.84 + ], + "angle": 0, + "content": "† Corresponding authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.127 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.455, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.234, + 0.455, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.332, + 0.458, + 0.353 + ], + "angle": 0, + "content": "(a) Experiment of prior methods in an online manner on CUB200." + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.151, + 0.788, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.234, + 0.787, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.332, + 0.761, + 0.344 + ], + "angle": 0, + "content": "(b) The description of the proposed OCGCD." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 1: Figure 1a shows that existing methods recorded poor performance in online training, suggesting that prior methods cannot handle online continual learning. Figure 1b shows the proposed scenario, OCGCD. As our scenario assumes batch-wise online learning, the model suffers severe forgetting and poor novel category discovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.532 + ], + "angle": 0, + "content": "advanced recognition and grouping skills to discover unknown categories within unlabeled datasets. In general, existing methods for GCD rely on an impractical assumption that they can access both labeled and unlabeled datasets simultaneously during training, which prevents these methods from meeting real-world application demands." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.789, + 0.763 + ], + "angle": 0, + "content": "In this context, a new scenario of continuous category discovery [19,22,37,42, 44] has emerged to address the unrealistic assumption by adopting offline continual learning. Despite its great advantage not requiring the assumption, these offline-based methods have inherent limitations in handling the continuous nature of data streams in the real world. Current approaches to offline continual learning primarily focus on static data streams, thus failing to adequately accommodate the dynamic nature of data streams where data can be created and deleted in an online manner. To verify this limitation, we conducted an empirical analysis using the CUB200 [41] dataset, dividing it into 160 labeled categories and 40 unlabeled categories. We evaluated the performance of existing methods [22,37] for continuous category discovery in an online manner. As shown in Figure 1a, we observed that existing methods [22,37] exhibited inferior performance in identifying novel categories, thereby resulting in poor overall performances. This indicates that the offline-based GCD methods have difficulty in clustering novel data effectively in the online learning scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In light of this observation, in this paper, we introduce a novel online learning scenario in continuous generalized category discovery, called Online Continuous Generalized Category Discovery (OCGCD). As shown in Figure 1b, OCGCD involves batch-wise training of unlabeled data in an online continual manner without any prior knowledge of the unknown data. Therefore, our novel" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "scenario closely replicates real-world data streams characterized by dynamic creation and elimination of data during training [14]. Here we summarize the key challenges of our novel OCGCD scenario as follows: 1) Challenging category discovery. The model struggles to retain the knowledge of trained categories because of the continuous nature of the data, making novel category discovery challenging. 2) Noisy pseudo-labeling. Online learning restricts access to the entire training dataset, leading to inaccurate category assignments and the generation of severe noise pseudo-labels for unlabeled data. 3) Severe catastrophic forgetting. Online learning erases observed data after training, resulting in severe forgetting of previously acquired knowledge [33]. Hence, our novel OCGCD scenario requires addressing the aforementioned challenges to achieve successful learning outcomes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.617 + ], + "angle": 0, + "content": "To address the challenges of OCGCD, we propose a novel approach called DEAN, Discovery via Energy Guidance and Feature Augmentation, which utilizes the energy score [28] to discover unknown data and further discriminate whether the data is seen or unseen and employs feature augmentation for better pseudo-labeling. By comparing the energy values of known and unknown data, DEAN effectively identifies unknown data among them, facilitating the clustering of novel categories. On top of that, we introduce an energy-based contrastive loss to enhance learning from unknown samples, promoting the acquisition of more discriminative knowledge during online learning. In addition, we propose a novel variance-based feature augmentation, a simple yet effective approach for accurate novel clustering and pseudo-labeling. In our scenario, since the model can access only batch-wise training data at once, it struggles to cluster data accurately. Variance-based feature augmentation (VFA) temporally extends the number of feature vectors through variational information to promote accurate clustering. To mitigate catastrophic forgetting, we employ parameter-efficient tuning, especially LoRA [17] in our framework while preserving the pre-trained weights. Extensive experiments demonstrate that our proposed method, DEAN, enables the network to discover novel categories more effectively in the OCGCD scenario than existing methods." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.617, + 0.691, + 0.632 + ], + "angle": 0, + "content": "We summarize the main contributions of this paper as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.643, + 0.785, + 0.703 + ], + "angle": 0, + "content": "- We introduce a novel Online Continuous Generalized Category Discovery (OCGCD) scenario that effectively reflects the online characteristics of real-world data. To tackle this scenario, we propose a novel method called Discovery via Energy Guidance and Feature Augmentation (DEAN)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.748 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first work to introduce the energy score for novel category discovery. We propose the energy-based contrastive loss to enhance online learning of discovered unknown data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.75, + 0.785, + 0.794 + ], + "angle": 0, + "content": "- For effective pseudo-labeling, we propose a new variance-based feature augmentation, called VFA. The proposed VFA enhances the sample clustering, leading to the improvement of pseudo-label quality." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.796, + 0.785, + 0.839 + ], + "angle": 0, + "content": "- Comprehensive experiments demonstrate that our novel method achieves significant performance improvements in the OCGCD scenario compared to state-of-the-art models." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.643, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.479, + 0.194 + ], + "angle": 0, + "content": "2.1 Novel Category Discovery" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.505 + ], + "angle": 0, + "content": "Novel Category Discovery (NCD) focuses on the task of identifying and categorizing unlabeled data [39]. To effectively discover and classify data, this field requires advanced object recognition and clustering skills. Most of the NCD methods fall into two training scenarios: disjoint training and joint training. The disjoint training scenario, also known as a two-stage approach, starts by pre-training the model on a labeled dataset, followed by fine-tuning on an unlabeled dataset from a different domain [13, 15, 16, 29]. The joint training scenario adopts a more integrated approach, simultaneously utilizing both labeled and unlabeled datasets during the training process [7, 12, 18, 46, 47]. While this scenario outperforms disjoint training approach [29], its dependence on the availability of labeled data makes it less feasible in real-world applications. Accordingly, more research has been directed towards designing scenarios where they mitigate constraints of NCD for better alignment with real-world settings. Generalized Category Discovery (GCD) [40] broadens the setting of NCD by allowing test data to contain both known and unknown classes, in contrast to NCD which focuses on exclusively unknown classes. Moreover, to better represent real-world scenarios, several methods [19, 22, 37, 42, 44] have been proposed that integrate continual learning into category discovery. However, existing methods predominantly utilizes an offline continual learning framework, which remains an impractical assumption for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.597 + ], + "angle": 0, + "content": "We propose an online version of continuous GCD that tackles the challenges of online learning through energy-guided clustering and energy-based contrastive loss. Energy-guided clustering utilizes the trained classifier to extract energy from each sample, efficiently discovering novel categories. Moreover, applying energy-based contrastive loss enables the model to effectively learn feature knowledge, thereby enhancing the capacity of the model in online continual learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.483, + 0.635 + ], + "angle": 0, + "content": "2.2 Online Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Online Continual Learning (OCL) is a more realistic continual learning scenario where the model learns from data arriving in small, sequential batches in real-time, without access to previously seen batches [30]. A key aspect of continual learning is to mitigate catastrophic forgetting and balance existing knowledge while adapting to new knowledge [8, 32, 33]. The dynamic distribution of online setting intensifies this challenge, and various approaches have been introduced to address this issue. MIR [1] proposed a memory retrieval technique for replay, selecting mostly interfered samples. OCM [10] deals with forgetting by maximizing mutual information between past and current data to learn more transferable features. i-Blurry [23] introduces a realistic setting that combines disjoint and blurry class distributions, mitigating forgetting by employing sample-wise importance sampling. Furthermore, Si-Blurry [34] proposed a new blurry task scenario, acknowledging the stochastic properties in real-world data distributions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "In this paper, we propose a novel setting that integrates online continual learning with generalized category discovery. By emphasizing the online learning, OCGCD presents a new perspective in the field of category discovery. It is a more realistic and challenging scenario than existing approaches in GCD, taking into account the continuous flow of real-world data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.244, + 0.362, + 0.258 + ], + "angle": 0, + "content": "2.3 Noise Label" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.789, + 0.418 + ], + "angle": 0, + "content": "Deep neural networks can overfit to noisy label data, leading to a substantial decrease in the robustness and generalization ability [2, 43]. Consequently, tackling noisy label data becomes essential, especially in category discovery where pseudo-labeling of unlabeled data plays a pivotal role in identifying new categories. Recent studies have primarily focused on discriminating between clean data and noisy label data. DivideMix [26] use a Gaussian Mixture Model (GMM) generated from the training dataset to differentiate between clean and noisy data, while AugDesc [35] applies various data augmentations to the training dataset to amplify the distinction. SplitNet [21] introduces a compact learnable module that separates clean and noisy label data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.79, + 0.571 + ], + "angle": 0, + "content": "In OCGCD, learning occurs without prior knowledge of novel categories, relying only on the current batch-wise data to predict the number of clusters and generate pseudo-labels based on this prediction. This scenario makes it difficult to apply existing techniques that depend on utilizing the training dataset or employing learnable modules. Therefore, we propose Variance-based Feature Augmentation (VFA) to effectively separate clean and noisy labels in our framework. VFA can diversely augment (re-generate) the feature vectors of given data, enabling the creation of discriminative clusters for accurate pseudo-labeling. Unlike CGCD [22], which requires training additional modules for the discovery, VFA facilitates end-to-end training without the additional modules or extra learning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.75, + 0.61 + ], + "angle": 0, + "content": "3 Online Continuous Generalized Category Discovery" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.623, + 0.63, + 0.637 + ], + "angle": 0, + "content": "3.1 Problem Formulation and Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Problem Formulation. In online continuous GCD, the training dataset \\(\\mathcal{D} = \\{\\mathcal{D}^{base},\\mathcal{D}^{inc}\\}\\) is given sequentially, where \\(\\mathcal{D}^{base} = \\{(x_i,y_i)\\in \\mathcal{X}\\times \\mathcal{Y}_l\\}_{i = 1}^{|D^{base}|}\\) is the labeled dataset for the base session and \\(\\mathcal{D}^{inc} = \\{(x_i,y_i)\\in \\mathcal{X}\\times \\mathcal{Y}_u\\}_{i = 1}^{|D^{inc}|}\\) is the unlabeled dataset for the incremental session. The proposed novel scenario assumes an overlap between the labeled category set \\(\\mathcal{V}_l\\) and the unlabeled category set \\(\\mathcal{V}_u\\), as it is a sub-category of the generalized category discovery. In this setting, the network trains on \\(\\mathcal{D}^{base}\\) and on \\(\\mathcal{D}^{inc}\\) in an offline and online manner, respectively. Our proposed online continuous GCD aims to facilitate the discovery of novel categories in real-world scenarios while preserving the clustering knowledge obtained from the offline training of the labeled dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Method Overview. Figure 2 shows an overview of the proposed method, where the pre-trained vision transformer [6] via self-supervised learning serves as the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 2: Overall process of the proposed DEAN framework. The energy-guided discovery splits unlabeled data into known, seen, and unseen data for better novel category discovery, while variance-based feature augmentation enhances the clustering of unseen data. \\(\\mathcal{L}_{ec}\\) facilitates better discriminative learning in the online continual learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.789, + 0.592 + ], + "angle": 0, + "content": "backbone network. We initially train the network using the labeled dataset in the base session (Section 3.2). We introduce energy-guided discovery to identify unknown samples from the given data. Our energy-guided discovery involves a two-step process to determine whether data belong to known or unknown categories and whether the unknown data belong to previously seen or unseen categories (Section 3.3). We split the data into known, seen, and unseen categories through energy-guided discovery. For pseudo-labeling of unseen data, we propose variance-based feature augmentation to improve the clustering of unseen data, resulting in better pseudo-labeling (Section 3.4). Subsequently, the network trains pseudo-labeled data in an online manner using an energy-based contrastive loss for better discriminative learning (Section 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.566, + 0.625 + ], + "angle": 0, + "content": "3.2 Supervised Learning at Base Session" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.752 + ], + "angle": 0, + "content": "During the base session, we train the offline model \\(\\theta^{off}\\) consisting of the feature extractor \\(f^{off}(\\cdot):\\mathcal{X}\\to \\mathbb{R}^d\\) and classifier head \\(g^{off}(\\cdot):\\mathbb{R}^d\\rightarrow \\mathcal{Y}_l\\) in an offline manner. As the offline network \\(\\theta^{off} = \\{f^{off},g^{off}\\}\\) trains on the labeled dataset during the base session, we employ the cross-entropy loss for supervised learning. While existing novel category discovery methods often utilize self-supervised [11] and supervised [20] contrastive loss for representation learning without a classifier head, our approach utilizes a classifier for clustering. The formulation of the base session training is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.757, + 0.787, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = - \\mathbb {E} _ {(\\mathcal {X}, \\mathcal {Y} _ {l}) \\sim \\mathcal {D} ^ {b a s e}} \\frac {1}{| \\mathcal {D} ^ {b a s e} |} \\sum_ {i = 1} ^ {| \\mathcal {D} ^ {b a s e} |} y _ {i} \\cdot \\log \\delta (g ^ {\\text {o f f}} (f ^ {\\text {o f f}} (x _ {i}))), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\delta(\\cdot)\\) is the softmax function. We train the model to capture representation knowledge during the base session training. After training the base session, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.462, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.293, + 0.465, + 0.325 + ], + "angle": 0, + "content": "(a) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using fine-split of CGCD." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.327, + 0.462, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.473, + 0.466, + 0.506 + ], + "angle": 0, + "content": "(c) Comparing normal distributions of known and unknown samples on CUB200 using fine-split of CGCD." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.144, + 0.787, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.293, + 0.787, + 0.325 + ], + "angle": 0, + "content": "(b) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using energy scores." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.327, + 0.787, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.473, + 0.787, + 0.506 + ], + "angle": 0, + "content": "(d) Comparing normal distributions of known and unknown samples on CUB200 using energy scores." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Fig. 3: Validating the effectiveness of energy scores for novel category discovery without prior knowledge about novel categories by comparison with existing methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.639 + ], + "angle": 0, + "content": "freeze the offline model and utilize the model at the discovery step to distinguish between known and unknown data, enabling accurate discovery. In addition, we initialize the online model, denoted as \\(\\theta^{on}\\), based on the parameters of the offline model for the online continual learning on the unlabeled dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.662, + 0.66, + 0.679 + ], + "angle": 0, + "content": "3.3 Energy-Guided Discovery to Identify Unknowns" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.844 + ], + "angle": 0, + "content": "The discovery step aims to discover new categories within the provided batchwise data. While CGCD, which assumes the absence of prior knowledge about novel categories, introduced a two-stage split process that depends on learning an additional split network to detect novel data, it has some limitations in our proposed scenario. CGCD includes the requirement for extra parameters and additional learning stage. Moreover, in our proposed online scenario where the model cannot access the entire training data, the split network faces difficulty in identifying novel samples due to severe forgetting of the split network. Liu et al. [28] proposed a method for detecting out-of-distribution (OOD) data by analyzing the difference in energy scores between in-distribution (ID) and OOD" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.128 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "data. We conducted experiments to compare the two-stage split process from CGCD with the energy-based discovery method to identify trained categories (known to the model) and untrained categories (unknown to the model)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.785, + 0.346 + ], + "angle": 0, + "content": "In Figure 3, we found that the two-stage split of the CGCD struggled to identify both known and unknown categories in our proposed online learning scenario. Online learning with batch-wise data led to severe forgetting in the split network, resulting in poor detection of unknown samples. In contrast, energy-based discovery showed better performance in novel category discovery without extra parameters compared to CGCD. Moreover, as it does not require an additional learning phase, energy-based discovery enables end-to-end training. Inspired by this observation, we propose an energy-guided discovery approach for novel category discovery. To the best of our knowledge, this is the first work to utilize the energy score for novel category discovery." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.348, + 0.787, + 0.455 + ], + "angle": 0, + "content": "During the incremental session, the network trains unlabeled batch-wise data \\( B^{t} = \\{x_{1}, x_{2}, \\ldots, x_{|B^{t}|}\\} \\), where \\( t \\) is the index of the batch. We classify unlabeled data \\( B^{t} \\) into known data, seen data, and unseen data using both the \\( \\theta^{off} \\) and \\( \\theta^{on} \\) through a two-stage process. In the first stage, we split the data into known and unknown categories. We obtain logits \\( z \\) using the offline model \\( \\theta^{off}(\\cdot) = g^{off}(f^{off}(\\cdot)) \\) and calculate the energy scores using these logits. The formulation for calculating the energy scores is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.47, + 0.786, + 0.49 + ], + "angle": 0, + "content": "\\[\nz ^ {t} = \\theta^ {\\text {o f f}} \\left(B ^ {t}\\right) \\in \\mathbb {R} ^ {\\left| B ^ {t} \\right| \\times \\left| \\mathcal {Y} _ {l} \\right|}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.495, + 0.786, + 0.538 + ], + "angle": 0, + "content": "\\[\ne ^ {t} = \\left\\{e _ {i} ^ {t} \\mid e _ {i} ^ {t} = - \\log \\sum_ {j = 1} ^ {| \\mathcal {Y} _ {i} |} \\exp \\left(z _ {i, j} ^ {t}\\right) \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.553, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\exp (\\cdot)\\) denotes an exponential function, \\(e^t\\) denotes a set of energy scores, and \\(z_{i,j}\\) denotes the logit from \\(i\\)-th sample corresponding to the \\(j\\)-th class. We divide the energy scores by the Gaussian Mixture Model (GMM). We initially set the GMM to have two clusters and use the energy score of each sample as input for clustering. We then label the cluster with a low energy score as \"known\" and the cluster with a high energy score as \"unknown\". This division allows us to split the unlabeled data into known and unknown data. In the second stage, we divide the unknown data into seen and unseen categories. Seen data comprise the data belonging to observed novel categories during the incremental session, while unseen data belong to undiscovered categories. To split the unknown data, we utilize the online model \\(\\theta^{on}(\\cdot) = g^{on}(f^{on}(\\cdot))\\) to extract energy scores. The process in the second stage is the same as the first stage, where we split the unknown data into seen and unseen categories based on their energy scores. For the initial batch of incremental sessions, the online model is identical to the offline model. Additionally, since the initial batch data is the first data of the incremental session, any data classified as unknown is assumed to be unseen data. The proposed energy-guided discovery splits unlabeled data into known, seen, and unseen categories effectively. Unlike prior methods which require sufficient data, it can identify novel categories with batch-wise data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.75, + 0.163 + ], + "angle": 0, + "content": "3.4 Pseudo-Labeling via Variance-based Feature Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.293 + ], + "angle": 0, + "content": "After the discovery step, we pseudo-label known, seen, and unseen data to assign labels to each sample. For known and seen data, where the model has already learned knowledge for the categories, we utilize predictions from the offline model for known data and from the online model for seen data to assign pseudo-labels, respectively. Conversely, as the unseen data belong to undiscovered categories, the online model cannot predict proper categories for them and needs to discover novel categories. Moreover, due to the given batch-wise data, discovering novel categories to assign pseudo labels is challenging." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.294, + 0.789, + 0.415 + ], + "angle": 0, + "content": "To address this challenge, we propose variance-based feature augmentation, which temporally augments feature vectors using variational knowledge from unseen data to acquire accurate novel clusters. We obtain feature vectors of the unseen data \\( B_{u}^{t} \\) using the online feature extractor \\( f^{on}(\\cdot) \\) denoted as \\( h_u^t = f^{on}(B_u^t) \\). We calculate the variance of the unseen feature vectors and augment unseen feature vectors based on the variance. We employ each feature vector to augment features by the Gaussian distribution \\( \\hat{h}_u^t = \\{\\hat{h}_u\\sim \\mathcal{N}(h_u,\\sigma_u^2)\\} ,h_u\\in h_u^t \\). The formulation of the variance-based feature augmentation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.426, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\nh _ {a u g} ^ {t} = \\{h _ {u} ^ {t}; \\hat {h} _ {u} ^ {t} \\}, \\quad \\hat {h} _ {u} ^ {t} = \\{\\hat {h} _ {u} \\sim \\mathcal {N} (h _ {u}, \\sigma_ {u} ^ {2}) \\} \\in \\mathbb {R} ^ {(\\left| B _ {u} ^ {t} \\right| \\cdot K) \\times d}, \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.55 + ], + "angle": 0, + "content": "where \\(\\hat{h}_u\\) and \\(\\sigma_u\\) represent augmented feature vectors and the standard deviation of the unseen feature vectors, respectively. \\(K\\) denotes the hyperparameter for the number of augmented feature vectors, and \\(h_{aug}^t\\) denotes the concatenation of unseen feature vectors and augmented feature vectors. Novel categories are discovered based on \\(h_{aug}^t\\). Augmented feature vectors expand the feature space with variational knowledge, which helps the model discover novel clusters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.673 + ], + "angle": 0, + "content": "We employ affinity propagation [9] to assign pseudo-labels to unseen data. Affinity propagation is a non-parametric clustering algorithm that does not require a predefined number of clusters, using the similarity matrix of the data. Given the absence of prior knowledge of novel categories, we extend the online classifier with the estimated number of novel clusters obtained by affinity propagation. The proposed variance-based feature augmentation is a simple yet effective approach that helps the model assign effective pseudo-labels based on accurate novel category discovery." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.707, + 0.71 + ], + "angle": 0, + "content": "3.5 Online Continual Learning with Pseudo-Labeled Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For online continual learning, we employ parameter-efficient tuning to learn novel representation knowledge while preserving previously learned knowledge, as it effectively mitigates catastrophic forgetting. We use cross-entropy loss with pseudo-labeled data in an online manner to capture novel knowledge. However, adapting the model with batch-wise data raises recency bias, where sufficient training through the given batch-wise data impedes discriminative knowledge. To address this, we propose an energy-based contrastive loss \\((\\mathcal{L}_{ec})\\) for effective discriminative learning. The energy-based contrastive loss focuses on inducing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "the online model to capture the knowledge of novel categories. It amplifies the energy from the known classifier while reducing the energy from the novel classifier. By selectively applying \\(\\mathcal{L}_{ec}\\) to novel data \\(B_{n}^{t}\\) including seen and unseen data, it helps the online model learn novel discriminative knowledge. The formulation of \\(\\mathcal{L}_{ec}\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.235, + 0.786, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E} (f (x); g) = - \\log \\sum_ {i = 1} ^ {| \\mathcal {Y} |} \\exp \\left(g _ {i} (f (x))\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.303, + 0.786, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e c} = \\frac {1}{| B _ {n} ^ {t} |} \\sum_ {(x _ {n}) \\sim B _ {n} ^ {t}} \\log \\left(1 + \\frac {\\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {n e w})}{\\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {o l d})}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.353, + 0.788, + 0.445 + ], + "angle": 0, + "content": "where \\(\\operatorname{E}(\\cdot ;\\cdot)\\) and \\(|\\mathcal{V}|\\) denote energy function and the number of nodes at the classifier, respectively, \\(g_{i}(\\cdot)\\) represents the \\(i\\)-th node from the given classifier \\(g\\), and \\(g^{old}\\) and \\(g^{new}\\) denote the nodes corresponding to known categories and discovered novel categories in the online classifier \\(g^{on}(\\cdot)\\), respectively. Eq. 5 represents the calculation of the energy score given the feature vector and classifier. The total loss \\(\\mathcal{L}_{inc}\\) for the incremental session can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.459, + 0.786, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n c} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {e c}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.486, + 0.79, + 0.516 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{ec}\\) represents energy-based contrastive loss, and \\(\\mathcal{L}_{ce}\\) denotes cross-entropy loss with pseudo-labeled data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.54, + 0.377, + 0.557 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.429, + 0.589 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Baselines and Implementation Details. Recently, many approaches have emerged in novel category discovery. However, many of them focus on self-supervised learning with non-parametric classifiers (e.g. K-means) [42] and rarely consider continual learning [45]. Some methods that consider continual learning assume prior knowledge for novel category discovery. Since the proposed EGD and fine-split from CGCD require a classifier to estimate the number of novel categories, non-parametric classifiers cannot be used. Additionally, existing methods not designed for continual learning are unsuitable baselines due to severe forgetting of known classes. For a fair comparison, we mainly compared our method with FRoST [37] and CGCD [22], which focus on continual learning and utilize parametric classifiers for novel category discovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We used ViT-B-16 [6] pre-trained on DINO-ImageNet [3] as the backbone network for all methods, including ours. We trained our method 30 epochs for the base session and 15 epochs for the incremental session. We utilized the AdamW optimizer with a weight decay of 1e-4 and a learning rate of 1e-3. We trained our model using a single RTX 3090 GPU with a batch size set to 64." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Datasets. We evaluated our method with state-of-the-art (SoTA) continuous category discovery methods on three datasets: CUB200 [41], FGVC-Aircraft [31], and CIFAR-100 [24]. For continual learning, we partitioned CUB200 and FGVC-Aircraft following the same split configuration in CGCD. For CIFAR-100, we divided 80 classes for known classes and 20 classes for novel classes. To follow the GCD setting, we divided the training samples into labeled and unlabeled per class, maintaining a ratio of 0.8 for the labeled to 0.2 for the unlabeled." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.314 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluated the methods by measuring clustering accuracy on test dataset \\(\\mathcal{D}\\). We utilized the hungarian algorithm [25] to obtain the optimal permutation \\(h^*\\) that aligns the prediction \\(y_i^*\\) from the model with the ground truth label \\(y_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.326, + 0.785, + 0.368 + ], + "angle": 0, + "content": "\\[\nh ^ {*} = \\arg \\min _ {h} \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h \\left(y _ {i} ^ {*}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.784, + 0.409 + ], + "angle": 0, + "content": "where \\( |\\mathcal{D}| \\) is the size of the test dataset. Clustering accuracy on \\( \\mathcal{D} \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.419, + 0.785, + 0.461 + ], + "angle": 0, + "content": "\\[\nM = \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h ^ {*} \\left(y _ {i} ^ {*}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.784, + 0.592 + ], + "angle": 0, + "content": "We employed this metric to measure the clustering accuracy of all classes (\\(M_{all}\\)), known classes (\\(M_{old}\\)), and unknown classes (\\(M_{new}\\)) in the cases of test dataset \\(\\mathcal{D}^{base}\\) and \\(\\mathcal{D}^{inc}\\), respectively. Following continuous GCD approaches to evaluate the ability to maintain performance for known classes [22, 44], we employed the forgetting metric \\(F = M_{old}^{base} - M_{old}^{inc}\\), where \\(M_{old}^{base}\\), \\(M_{old}^{inc}\\) refers to known class clustering accuracy at the base session and incremental session. A higher clustering accuracy metric is better, and a lower forgetting metric is better. This ensures the model is effective for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.784, + 0.652 + ], + "angle": 0, + "content": "Moreover, we evaluate the estimated pseudo-labels to validate their effectiveness in pseudo-labeling. Pseudo-labeling for unlabeled data is crucial as the model utilizes pseudo-labels to interpret the unlabeled data. The pseudo-labeling accuracy on the dataset \\(\\mathcal{D}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.665, + 0.785, + 0.707 + ], + "angle": 0, + "content": "\\[\nM = \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h ^ {*} \\left(\\tilde {y} _ {i}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.717, + 0.599, + 0.733 + ], + "angle": 0, + "content": "where \\(\\tilde{y}_i\\) denotes the pseudo-label of the \\(i\\)-th sample." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.487, + 0.77 + ], + "angle": 0, + "content": "4.2 Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.842 + ], + "angle": 0, + "content": "We compared the clustering performance of our proposed method, DEAN, with SoTA approaches, as shown in Table 1. Since we considered an online scenario for category discovery, traditionally approached using offline methods, there is no previous work for comparison. Therefore, for a fair comparison, our study adapts" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Clustering performance comparison on CUB200, FGVC-Aircraft, and CIFAR-100. Best performance is highlighted in bold. All the baselines utilize the fine-split method from CGCD to handle the absence of prior knowledge." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.199, + 0.789, + 0.343 + ], + "angle": 0, + "content": "
MethodCUB200FGVC-AircraftCIFAR-100
MallMoldMnewF↓MallMoldMnewF↓MallMoldMnewF↓
Supervised72.5170.6879.6611.9166.2563.6276.7611.0381.3479.4888.8010.07
Fine-Tuning1.961.942.0180.522.061.305.1073.353.683.245.4586.26
FRoST [ECCV' 22]23.9828.725.4550.8514.4415.659.6048.316.136.215.8069.24
CGCD[ICCV' 23]40.4047.8811.1835.6114.0815.677.7560.002.642.762.1586.74
DEAN w/ FS55.5765.3717.2917.6240.3249.145.1025.4352.2762.1112.9027.55
DEAN (Ours)59.5666.3333.1116.2646.4453.5617.9920.9762.3470.3430.3519.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.357, + 0.505, + 0.4 + ], + "angle": 0, + "content": "Table 2: Ablation experiment on CUB200. The baseline denotes fine-tuning with the fine split from the CGCD." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.413, + 0.505, + 0.539 + ], + "angle": 0, + "content": "
AblationCUB200
MallMoldNewF↓
Baseline1.961.942.0180.52
+ LoRA49.3360.356.2422.24
+ EGD54.9762.9823.6719.61
+ EC Loss57.3764.9827.6317.61
+ VFA (Ours)59.5666.3333.1116.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.357, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Table 3: Ablation study for the hyperparameter K of variance-based feature augmentation on CUB200." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.411, + 0.789, + 0.54 + ], + "angle": 0, + "content": "
AblationCUB200
\\(M_{all}^{PS}\\)\\(M_{old}^{PS}\\)\\(M_{new}^{PS}\\)
K=057.3185.1946.47
K=164.2788.9653.74
K=375.8793.4364.48
K=576.9394.0565.75
K=775.4493.4164.07
K=975.2293.563.36
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.749 + ], + "angle": 0, + "content": "offline continuous NCD and GCD methods into OCGCD scenarios. The experiment includes two SoTA continuous category discovery methods: FRoST [37] and CGCD [22]. In addition, we include a supervised learning method as an upper bound and a fine-tuning method as a performance baseline. Note that in our proposed scenario, OCGCD, all baselines are not provided with prior knowledge about novel categories. Instead, we employed fine-split method from CGCD [22] which consists of two stages: initial split and fine split. During the initial split stage, unlabeled data is roughly divided into known and unknown categories based on cosine similarity. In the fine split stage, a split model is trained to distinguish between known and unknown data. The fine-split method estimates the number of novel categories conveying information about novel categories to the comparison methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The results show that our method, DEAN, recorded a robust clustering performance, significantly surpassing existing continuous NCD and GCD methods. DEAN recorded consistently high metrics across \\( M_{all} \\), \\( M_{old} \\), and \\( M_{new} \\) while keeping forgetting metric \\( F \\) substantially low. This denotes our method can learn novel knowledge without forgetting. Moreover, it is noteworthy that DEAN w/ FS which denotes the adoption of fine-split from CGCD into DEAN instead" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.146, + 0.493, + 0.174 + ], + "angle": 0, + "content": "Table 4: Ablation study for the cluster accuracy of pseudo-labels." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.186, + 0.49, + 0.284 + ], + "angle": 0, + "content": "
MethodCUB200
\\( M_{all}^{PS} \\)\\( M_{old}^{PS} \\)\\( M_{new}^{PS} \\)
CGCD [ICCV' 23]61.8881.0758.68
DEAN w/o (\\( \\mathcal{L}_{ec} \\), VFA)57.2385.1145.80
DEAN w/o VFA57.3185.1946.47
DEAN (Ours)76.9394.0565.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 5: Ablation experiment of parameter-efficient tuning on CUB200." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.185, + 0.789, + 0.284 + ], + "angle": 0, + "content": "
MethodCUB200
\\(M_{all}\\)\\(M_{old}\\)\\(M_{new}\\)\\(F\\downarrow\\)
DEAN(Ours)Adapter [4]56.9664.7526.5317.64
Prefix [27]58.5466.2928.2816.30
LoRA [17]59.5666.3333.1116.26
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.432 + ], + "angle": 0, + "content": "of energy-guided discovery (EGD) also showed notable performance compared to the baselines. But, DEAN achieved significant performance gains across all datasets as the proposed EGD is applied. In contrast, existing continual methods showed shortcomings in clustering performance. Specifically, we observed extremely low clustering performance on CIFAR-100 dataset, low \\( M_{all} \\), \\( M_{old} \\), and \\( M_{new} \\) values, coupled with high \\( F \\). On CIFAR-100 dataset, DEAN surpassed the most recent method CGCD assuming no prior knowledge for novel categories by \\( 59.7\\% \\), \\( 67.58\\% \\), \\( 28.2\\% \\), and \\( 67.57\\% \\) for \\( M_{all} \\), \\( M_{old} \\), \\( M_{new} \\), and \\( F \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Such observations show that while current GCD methods are vulnerable to online settings, DEAN can manage the balance required in the OCGCD scheme of retaining old knowledge while integrating the new. We believe this results shows the suitability of DEAN for real-world applications, where data is never static. We empirically confirmed that DEAN can evolve with the data it encounters, overcoming the challenges presented in the OCGCD scheme." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.39, + 0.56 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.613 + ], + "angle": 0, + "content": "The experimental results in Section 4.2 show that our DEAN achieved significant performance in OCGCD scenarios. Through further experiments on the CUB200, we provide a detailed analysis of the components proposed in our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Components. We validate the effectiveness of adding each component of our method to the baseline model. The baseline denotes fine-tuning the backbone model with a data-split following settings in CGCD. As shown in Table 2, results reveal that all components have an important impact on performance. Employing LoRA to baseline shows significant improvement in \\( M_{old} \\), \\( F \\). This shows that our application of LoRA on the frozen model at incremental session effectively mitigates forgetting. While LoRA shows only a slight increase in \\( M_{new} \\), additionally employing energy-guided discovery (EGD) and EC Loss markedly improves \\( M_{new} \\). This confirms that adopting energy guidance will enhance the ability of the model to differentiate between old and new categories. Finally, the integration of VFA further boosts the model performance on \\( M_{new} \\). This improvement reveals the importance of accurate pseudo-labeling since VFA is a technique applied after training. Each component is crucial, incrementally building upon the last to solidify the model's performance in our online category discovery scenario." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Hyperparameter Analysis of VFA. Variance-based Feature Augmentation (VFA) is our proposed method of temporarily augmenting feature vectors to acquire accurate pseudo-labeling. Our analysis evaluated the effect of hyperparameter \\( K \\), the number of augmented feature vectors, on the accuracy of pseudolabels. From Table 3 we observed that employment of VFA is beneficial, increasing novel category pseudo-label accuracy from \\( 46.47\\% \\) to \\( 65.75\\% \\). Among VFA employment, we observed that \\( K = 5 \\) yields the best performance across all metrics. When \\( K \\) is less than 5, the smaller \\( K \\) worsened performance. Conversely, values of \\( K \\) greater than 5 did not benefit the model. We analyzed that since VFA augments features based on variance, as the number of \\( K \\) increases, unseen samples can be clustered together without being distinguished." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.479 + ], + "angle": 0, + "content": "Pseudo-Labeling Accuracy. Accurate pseudo-labeling for clustering is important because it directly influences the performance of the model in incremental learning. Shown in Table 4, our proposed method, DEAN, significantly enhanced the clustering accuracy of pseudo-labels, achieving a \\(13.12\\%\\) improvement in \\(M_{all}^{PS}\\) over the most recent method CGCD. Through the ablation study, we investigated the contributions of each step in DEAN. When DEAN adopts only parameter-efficient tuning, excluding EC loss and VFA, we observed a notable declination in \\(M_{new}^{PS}\\). While each ablation step showed gradually accuracy enhancement, the application of VFA elevates the performance of DEAN above CGCD. Thus, the experimental result revealed that VFA played a vital role in distinguishing novel categories and confirming the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Parameter Efficient Tuning. The ablation study in Table 5 assesses the impact of parameter-efficient tuning (PET) methods on the CUB200 dataset. Our method can employ various tuning methods such as Adapter [4] or Prefix-tuning [27], and in this experiment, we compared modules with LoRA [17], Adapter and Prefix-tuning. We set the size of all PET modules size to 5 and integrated them into the last 5 layers of ViT. Overall improvements are shown by adopting PET over existing methods reported in Table 1. Among PET methods, LoRA showed the highest clustering accuracy. The performance enhancement of LoRA may stem from its fast adaptation, which is highly effective in online continual learning scenarios that often depend on fewer samples for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.36, + 0.673 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we introduced a novel scenario for category discovery, Online Continuous Generalized Category Discovery, which accounts for the dynamic nature of real-world data streams, leading to severe forgetting and poor novel category discovery. In addition, we proposed a novel framework, DEAN, which utilizes energy-based novel category discovery for the first time in this field and introduces variance-based feature augmentation to enhance accurate pseudo-labeling. Our DEAN demonstrated promising performance in the proposed novel scenario. However, DEAN could not guarantee inference performance midway through training. In future work, we plan to explore an online continual framework that can ensure inference performance during training through fast adaptation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.304 + ], + "angle": 0, + "content": "This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University))." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.328, + 0.324, + 0.344 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.362, + 0.787, + 0.405 + ], + "angle": 0, + "content": "1. Aljundi, R., Belilovsky, E., Tuytelaars, T., Charlin, L., Caccia, M., Lin, M., Page-Caccia, L.: Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.406, + 0.787, + 0.46 + ], + "angle": 0, + "content": "2. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: International conference on machine learning. pp. 233-242. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.461, + 0.787, + 0.503 + ], + "angle": 0, + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.504, + 0.787, + 0.545 + ], + "angle": 0, + "content": "4. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.546, + 0.787, + 0.588 + ], + "angle": 0, + "content": "5. Choe, S.A., Shin, A.H., Park, K.H., Choi, J., Park, G.M.: Open-set domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23943-23953 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.589, + 0.787, + 0.643 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.644, + 0.787, + 0.685 + ], + "angle": 0, + "content": "7. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.687, + 0.787, + 0.714 + ], + "angle": 0, + "content": "8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.715, + 0.787, + 0.742 + ], + "angle": 0, + "content": "9. Frey, B.J., Dueck, D.: Clustering by passing messages between data points. science 315(5814), 972-976 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.786, + 0.784 + ], + "angle": 0, + "content": "10. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: International Conference on Machine Learning. pp. 8109-8126. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Gutmann, M., Hyvärinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.362, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.127 + ], + "angle": 0, + "content": "K.-H. Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "12. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. arXiv preprint arXiv:2002.05714 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "13. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "14. He, J., Mao, R., Shao, Z., Zhu, F.: Incremental learning in online scenario. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13926-13935 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "15. Hsu, Y.C., Lv, Z., Kira, Z.: Learning to cluster in order to transfer across domains and tasks. arXiv preprint arXiv:1711.10125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "16. Hsu, Y.C., Lv, Z., Schlosser, J., Odom, P., Kira, Z.: Multi-class classification without multi-class labels. arXiv preprint arXiv:1901.00544 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "17. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "18. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "19. Joseph, K., Paul, S., Aggarwal, G., Biswas, S., Rai, P., Han, K., Balasubramanian, V.N.: Novel class discovery without forgetting. In: European Conference on Computer Vision. pp. 570-586. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "21. Kim, D., Ryoo, K., Cho, H., Kim, S.: Splitnet: learnable clean-noisy label splitting for learning with noisy labels. arXiv preprint arXiv:2211.11753 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.522, + 0.785, + 0.577 + ], + "angle": 0, + "content": "22. Kim, H., Suh, S., Kim, D., Jeong, D., Cho, H., Kim, J.: Proxy anchor-based unsupervised learning for continuous generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16688-16697 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "23. Koh, H., Kim, D., Ha, J.W., Choi, J.: Online continual learning on class incremental blurry task configuration with anytime inference. arXiv preprint arXiv:2110.10031 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "25. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "26. Li, J., Socher, R., Hoi, S.C.: Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.702, + 0.785, + 0.758 + ], + "angle": 0, + "content": "27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "28. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "29. Liu, Y., Tuytelaars, T.: Residual tuning: Toward novel category discovery without labels. IEEE Transactions on Neural Networks and Learning Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "30. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.39, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Continuous Generalized Category Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "31. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.229 + ], + "angle": 0, + "content": "32. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.27 + ], + "angle": 0, + "content": "33. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.323 + ], + "angle": 0, + "content": "34. Moon, J.Y., Park, K.H., Kim, J.U., Park, G.M.: Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11731-11741 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.365 + ], + "angle": 0, + "content": "35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.404 + ], + "angle": 0, + "content": "36. Park, K.H., Song, K., Park, G.M.: Pre-trained vision and language transformers are few-shot incremental learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23881-23890 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.431 + ], + "angle": 0, + "content": "37. Roy, S., Liu, M., Zhong, Z., Sebe, N., Ricci, E.: Class-incremental novel class discovery. In: European Conference on Computer Vision. pp. 317-333. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.471 + ], + "angle": 0, + "content": "38. Seo, J., Lee, S.H., Lee, T.Y., Moon, S., Park, G.M.: Generative unlearning for any identity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9151-9161 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.512 + ], + "angle": 0, + "content": "39. Troisemaine, C., Lemaire, V., Gosselin, S., Reiffers-Masson, A., Flocon-Cholet, J., Vaton, S.: Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.551 + ], + "angle": 0, + "content": "40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.578 + ], + "angle": 0, + "content": "41. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.618 + ], + "angle": 0, + "content": "42. Wu, Y., Chi, Z., Wang, Y., Feng, S.: Metagcd: Learning to continually learn in generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1655-1665 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.658 + ], + "angle": 0, + "content": "43. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.699 + ], + "angle": 0, + "content": "44. Zhang, X., Jiang, J., Feng, Y., Wu, Z.F., Zhao, X., Wan, H., Tang, M., Jin, R., Gao, Y.: Grow and merge: A unified framework for continuous categories discovery. Advances in Neural Information Processing Systems 35, 27455-27468 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.739 + ], + "angle": 0, + "content": "45. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.739, + 0.788, + 0.779 + ], + "angle": 0, + "content": "46. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.835 + ], + "angle": 0, + "content": "47. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.835 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_origin.pdf b/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45cac4bbbc31732dddd2e380cf979f7b7b89e2b5 --- /dev/null +++ b/2024/Online Continuous Generalized Category Discovery/4795a798-1c3d-4a13-b7b6-e1ac6f809f5a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:022ace9659307c2e308befab3ff10555c7cac9b56e26330e19028055256de644 +size 3976469 diff --git a/2024/Online Continuous Generalized Category Discovery/full.md b/2024/Online Continuous Generalized Category Discovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e0dee7341ebb42c6eaab00ddd7d696c933a6eaa1 --- /dev/null +++ b/2024/Online Continuous Generalized Category Discovery/full.md @@ -0,0 +1,293 @@ +# Online Continuous Generalized Category Discovery + +Keon-Hee Park1, Hakyung Lee2, Kyungwoo Song2†, and Gyeong-Moon Park1† + +1 Kyung Hee University, Yongin, Republic of Korea {khpark, gmpark}@ku.ac.kr + +2 Yonsei University, Seoul, Republic of Korea {hakyunglee0417, kyungwoo.song}@yonsei.ac.kr + +Abstract. With the advancement of deep neural networks in computer vision, artificial intelligence (AI) is widely employed in real-world applications. However, AI still faces limitations in mimicking high-level human capabilities, such as novel category discovery, for practical use. While some methods utilizing offline continual learning have been proposed for novel category discovery, they neglect the continuity of data streams in real-world settings. In this work, we introduce Online Continuous Generalized Category Discovery (OCGCD), which considers the dynamic nature of data streams where data can be created and deleted in real time. Additionally, we propose a novel method, DEAN, Discovery via Energy guidance and feature Augmentation, which can discover novel categories in an online manner through energy-guided discovery and facilitate discriminative learning via energy-based contrastive loss. Furthermore, DEAN effectively pseudo-labels unlabeled data through variance-based feature augmentation. Experimental results demonstrate that our proposed DEAN achieves outstanding performance in proposed OCGCD scenario. The implementation code is available at https://github.com/KHU-AGI/OCGCD. + +Keywords: Online Continual Learning $\cdot$ Generalized Category Discovery $\cdot$ Energy-Guided Discovery $\cdot$ Variance-based Feature Augmentation + +# 1 Introduction + +The rapid progress of deep neural networks in computer vision has greatly facilitated the widespread adoption of artificial intelligence into real-world applications. Nevertheless, the networks still have challenges in embodying high-level human capabilities such as object recognition and grouping. To emulate these sophisticated human abilities, there has been a lot of interest in the field of category discovery. Specifically, generalized category discovery (GCD) [40] demands + +![](images/184125e7f3dba4888e0400978e84885bda23fd4536f509dabc7e170d2d8ea667.jpg) + +![](images/4dafa45173861cdfd4c6f145b1996db797c0fe53d32488d56ae5b09ea596467f.jpg) +(a) Experiment of prior methods in an online manner on CUB200. + +![](images/c918eeb5a4da590fabc540a08cd5dec57cc4d20bedee7d748a2691c6be28b240.jpg) + +![](images/1389858bf68b881018020a15319c4abc3ba03a9f0f2ecd036ee36d978dc9b438.jpg) +(b) The description of the proposed OCGCD. +Fig. 1: Figure 1a shows that existing methods recorded poor performance in online training, suggesting that prior methods cannot handle online continual learning. Figure 1b shows the proposed scenario, OCGCD. As our scenario assumes batch-wise online learning, the model suffers severe forgetting and poor novel category discovery. + +advanced recognition and grouping skills to discover unknown categories within unlabeled datasets. In general, existing methods for GCD rely on an impractical assumption that they can access both labeled and unlabeled datasets simultaneously during training, which prevents these methods from meeting real-world application demands. + +In this context, a new scenario of continuous category discovery [19,22,37,42, 44] has emerged to address the unrealistic assumption by adopting offline continual learning. Despite its great advantage not requiring the assumption, these offline-based methods have inherent limitations in handling the continuous nature of data streams in the real world. Current approaches to offline continual learning primarily focus on static data streams, thus failing to adequately accommodate the dynamic nature of data streams where data can be created and deleted in an online manner. To verify this limitation, we conducted an empirical analysis using the CUB200 [41] dataset, dividing it into 160 labeled categories and 40 unlabeled categories. We evaluated the performance of existing methods [22,37] for continuous category discovery in an online manner. As shown in Figure 1a, we observed that existing methods [22,37] exhibited inferior performance in identifying novel categories, thereby resulting in poor overall performances. This indicates that the offline-based GCD methods have difficulty in clustering novel data effectively in the online learning scenario. + +In light of this observation, in this paper, we introduce a novel online learning scenario in continuous generalized category discovery, called Online Continuous Generalized Category Discovery (OCGCD). As shown in Figure 1b, OCGCD involves batch-wise training of unlabeled data in an online continual manner without any prior knowledge of the unknown data. Therefore, our novel + +scenario closely replicates real-world data streams characterized by dynamic creation and elimination of data during training [14]. Here we summarize the key challenges of our novel OCGCD scenario as follows: 1) Challenging category discovery. The model struggles to retain the knowledge of trained categories because of the continuous nature of the data, making novel category discovery challenging. 2) Noisy pseudo-labeling. Online learning restricts access to the entire training dataset, leading to inaccurate category assignments and the generation of severe noise pseudo-labels for unlabeled data. 3) Severe catastrophic forgetting. Online learning erases observed data after training, resulting in severe forgetting of previously acquired knowledge [33]. Hence, our novel OCGCD scenario requires addressing the aforementioned challenges to achieve successful learning outcomes. + +To address the challenges of OCGCD, we propose a novel approach called DEAN, Discovery via Energy Guidance and Feature Augmentation, which utilizes the energy score [28] to discover unknown data and further discriminate whether the data is seen or unseen and employs feature augmentation for better pseudo-labeling. By comparing the energy values of known and unknown data, DEAN effectively identifies unknown data among them, facilitating the clustering of novel categories. On top of that, we introduce an energy-based contrastive loss to enhance learning from unknown samples, promoting the acquisition of more discriminative knowledge during online learning. In addition, we propose a novel variance-based feature augmentation, a simple yet effective approach for accurate novel clustering and pseudo-labeling. In our scenario, since the model can access only batch-wise training data at once, it struggles to cluster data accurately. Variance-based feature augmentation (VFA) temporally extends the number of feature vectors through variational information to promote accurate clustering. To mitigate catastrophic forgetting, we employ parameter-efficient tuning, especially LoRA [17] in our framework while preserving the pre-trained weights. Extensive experiments demonstrate that our proposed method, DEAN, enables the network to discover novel categories more effectively in the OCGCD scenario than existing methods. + +We summarize the main contributions of this paper as follows: + +- We introduce a novel Online Continuous Generalized Category Discovery (OCGCD) scenario that effectively reflects the online characteristics of real-world data. To tackle this scenario, we propose a novel method called Discovery via Energy Guidance and Feature Augmentation (DEAN). +- To the best of our knowledge, this is the first work to introduce the energy score for novel category discovery. We propose the energy-based contrastive loss to enhance online learning of discovered unknown data. +- For effective pseudo-labeling, we propose a new variance-based feature augmentation, called VFA. The proposed VFA enhances the sample clustering, leading to the improvement of pseudo-label quality. +- Comprehensive experiments demonstrate that our novel method achieves significant performance improvements in the OCGCD scenario compared to state-of-the-art models. + +# 2 Related Work + +# 2.1 Novel Category Discovery + +Novel Category Discovery (NCD) focuses on the task of identifying and categorizing unlabeled data [39]. To effectively discover and classify data, this field requires advanced object recognition and clustering skills. Most of the NCD methods fall into two training scenarios: disjoint training and joint training. The disjoint training scenario, also known as a two-stage approach, starts by pre-training the model on a labeled dataset, followed by fine-tuning on an unlabeled dataset from a different domain [13, 15, 16, 29]. The joint training scenario adopts a more integrated approach, simultaneously utilizing both labeled and unlabeled datasets during the training process [7, 12, 18, 46, 47]. While this scenario outperforms disjoint training approach [29], its dependence on the availability of labeled data makes it less feasible in real-world applications. Accordingly, more research has been directed towards designing scenarios where they mitigate constraints of NCD for better alignment with real-world settings. Generalized Category Discovery (GCD) [40] broadens the setting of NCD by allowing test data to contain both known and unknown classes, in contrast to NCD which focuses on exclusively unknown classes. Moreover, to better represent real-world scenarios, several methods [19, 22, 37, 42, 44] have been proposed that integrate continual learning into category discovery. However, existing methods predominantly utilizes an offline continual learning framework, which remains an impractical assumption for real-world applications. + +We propose an online version of continuous GCD that tackles the challenges of online learning through energy-guided clustering and energy-based contrastive loss. Energy-guided clustering utilizes the trained classifier to extract energy from each sample, efficiently discovering novel categories. Moreover, applying energy-based contrastive loss enables the model to effectively learn feature knowledge, thereby enhancing the capacity of the model in online continual learning. + +# 2.2 Online Continual Learning + +Online Continual Learning (OCL) is a more realistic continual learning scenario where the model learns from data arriving in small, sequential batches in real-time, without access to previously seen batches [30]. A key aspect of continual learning is to mitigate catastrophic forgetting and balance existing knowledge while adapting to new knowledge [8, 32, 33]. The dynamic distribution of online setting intensifies this challenge, and various approaches have been introduced to address this issue. MIR [1] proposed a memory retrieval technique for replay, selecting mostly interfered samples. OCM [10] deals with forgetting by maximizing mutual information between past and current data to learn more transferable features. i-Blurry [23] introduces a realistic setting that combines disjoint and blurry class distributions, mitigating forgetting by employing sample-wise importance sampling. Furthermore, Si-Blurry [34] proposed a new blurry task scenario, acknowledging the stochastic properties in real-world data distributions. + +In this paper, we propose a novel setting that integrates online continual learning with generalized category discovery. By emphasizing the online learning, OCGCD presents a new perspective in the field of category discovery. It is a more realistic and challenging scenario than existing approaches in GCD, taking into account the continuous flow of real-world data. + +# 2.3 Noise Label + +Deep neural networks can overfit to noisy label data, leading to a substantial decrease in the robustness and generalization ability [2, 43]. Consequently, tackling noisy label data becomes essential, especially in category discovery where pseudo-labeling of unlabeled data plays a pivotal role in identifying new categories. Recent studies have primarily focused on discriminating between clean data and noisy label data. DivideMix [26] use a Gaussian Mixture Model (GMM) generated from the training dataset to differentiate between clean and noisy data, while AugDesc [35] applies various data augmentations to the training dataset to amplify the distinction. SplitNet [21] introduces a compact learnable module that separates clean and noisy label data. + +In OCGCD, learning occurs without prior knowledge of novel categories, relying only on the current batch-wise data to predict the number of clusters and generate pseudo-labels based on this prediction. This scenario makes it difficult to apply existing techniques that depend on utilizing the training dataset or employing learnable modules. Therefore, we propose Variance-based Feature Augmentation (VFA) to effectively separate clean and noisy labels in our framework. VFA can diversely augment (re-generate) the feature vectors of given data, enabling the creation of discriminative clusters for accurate pseudo-labeling. Unlike CGCD [22], which requires training additional modules for the discovery, VFA facilitates end-to-end training without the additional modules or extra learning. + +# 3 Online Continuous Generalized Category Discovery + +# 3.1 Problem Formulation and Method Overview + +Problem Formulation. In online continuous GCD, the training dataset $\mathcal{D} = \{\mathcal{D}^{base},\mathcal{D}^{inc}\}$ is given sequentially, where $\mathcal{D}^{base} = \{(x_i,y_i)\in \mathcal{X}\times \mathcal{Y}_l\}_{i = 1}^{|D^{base}|}$ is the labeled dataset for the base session and $\mathcal{D}^{inc} = \{(x_i,y_i)\in \mathcal{X}\times \mathcal{Y}_u\}_{i = 1}^{|D^{inc}|}$ is the unlabeled dataset for the incremental session. The proposed novel scenario assumes an overlap between the labeled category set $\mathcal{V}_l$ and the unlabeled category set $\mathcal{V}_u$ , as it is a sub-category of the generalized category discovery. In this setting, the network trains on $\mathcal{D}^{base}$ and on $\mathcal{D}^{inc}$ in an offline and online manner, respectively. Our proposed online continuous GCD aims to facilitate the discovery of novel categories in real-world scenarios while preserving the clustering knowledge obtained from the offline training of the labeled dataset. + +Method Overview. Figure 2 shows an overview of the proposed method, where the pre-trained vision transformer [6] via self-supervised learning serves as the + +![](images/2469bdf2f8a83daea607ee94603a102b607d121b41e0f330ae960e63db22fdfe.jpg) +Fig. 2: Overall process of the proposed DEAN framework. The energy-guided discovery splits unlabeled data into known, seen, and unseen data for better novel category discovery, while variance-based feature augmentation enhances the clustering of unseen data. $\mathcal{L}_{ec}$ facilitates better discriminative learning in the online continual learning. + +backbone network. We initially train the network using the labeled dataset in the base session (Section 3.2). We introduce energy-guided discovery to identify unknown samples from the given data. Our energy-guided discovery involves a two-step process to determine whether data belong to known or unknown categories and whether the unknown data belong to previously seen or unseen categories (Section 3.3). We split the data into known, seen, and unseen categories through energy-guided discovery. For pseudo-labeling of unseen data, we propose variance-based feature augmentation to improve the clustering of unseen data, resulting in better pseudo-labeling (Section 3.4). Subsequently, the network trains pseudo-labeled data in an online manner using an energy-based contrastive loss for better discriminative learning (Section 3.5). + +# 3.2 Supervised Learning at Base Session + +During the base session, we train the offline model $\theta^{off}$ consisting of the feature extractor $f^{off}(\cdot):\mathcal{X}\to \mathbb{R}^d$ and classifier head $g^{off}(\cdot):\mathbb{R}^d\rightarrow \mathcal{Y}_l$ in an offline manner. As the offline network $\theta^{off} = \{f^{off},g^{off}\}$ trains on the labeled dataset during the base session, we employ the cross-entropy loss for supervised learning. While existing novel category discovery methods often utilize self-supervised [11] and supervised [20] contrastive loss for representation learning without a classifier head, our approach utilizes a classifier for clustering. The formulation of the base session training is as follows: + +$$ +\mathcal {L} _ {c e} = - \mathbb {E} _ {(\mathcal {X}, \mathcal {Y} _ {l}) \sim \mathcal {D} ^ {b a s e}} \frac {1}{| \mathcal {D} ^ {b a s e} |} \sum_ {i = 1} ^ {| \mathcal {D} ^ {b a s e} |} y _ {i} \cdot \log \delta (g ^ {\text {o f f}} (f ^ {\text {o f f}} (x _ {i}))), \tag {1} +$$ + +where $\delta(\cdot)$ is the softmax function. We train the model to capture representation knowledge during the base session training. After training the base session, we + +![](images/a06de2bd6aa5851dd24a2615dc12694c5fbcecd03d7f3b181a0ebcdb05407e6d.jpg) + +![](images/6651a230f898248f0056e6c40976e2efd380444e632109993318f18b3a90e3b2.jpg) +(a) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using fine-split of CGCD. +(c) Comparing normal distributions of known and unknown samples on CUB200 using fine-split of CGCD. + +![](images/423210ebdc5c836cb61062f186d84d91b4f848a9b95c54a3de3d635a61a7f800.jpg) +(b) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using energy scores. + +![](images/6d9bb272b821e9069b802168bec71a34348acd5b39d884d799d17b5797798278.jpg) +(d) Comparing normal distributions of known and unknown samples on CUB200 using energy scores. +Fig. 3: Validating the effectiveness of energy scores for novel category discovery without prior knowledge about novel categories by comparison with existing methods. + +freeze the offline model and utilize the model at the discovery step to distinguish between known and unknown data, enabling accurate discovery. In addition, we initialize the online model, denoted as $\theta^{on}$ , based on the parameters of the offline model for the online continual learning on the unlabeled dataset. + +# 3.3 Energy-Guided Discovery to Identify Unknowns + +The discovery step aims to discover new categories within the provided batchwise data. While CGCD, which assumes the absence of prior knowledge about novel categories, introduced a two-stage split process that depends on learning an additional split network to detect novel data, it has some limitations in our proposed scenario. CGCD includes the requirement for extra parameters and additional learning stage. Moreover, in our proposed online scenario where the model cannot access the entire training data, the split network faces difficulty in identifying novel samples due to severe forgetting of the split network. Liu et al. [28] proposed a method for detecting out-of-distribution (OOD) data by analyzing the difference in energy scores between in-distribution (ID) and OOD + +data. We conducted experiments to compare the two-stage split process from CGCD with the energy-based discovery method to identify trained categories (known to the model) and untrained categories (unknown to the model). + +In Figure 3, we found that the two-stage split of the CGCD struggled to identify both known and unknown categories in our proposed online learning scenario. Online learning with batch-wise data led to severe forgetting in the split network, resulting in poor detection of unknown samples. In contrast, energy-based discovery showed better performance in novel category discovery without extra parameters compared to CGCD. Moreover, as it does not require an additional learning phase, energy-based discovery enables end-to-end training. Inspired by this observation, we propose an energy-guided discovery approach for novel category discovery. To the best of our knowledge, this is the first work to utilize the energy score for novel category discovery. + +During the incremental session, the network trains unlabeled batch-wise data $B^{t} = \{x_{1}, x_{2}, \ldots, x_{|B^{t}|}\}$ , where $t$ is the index of the batch. We classify unlabeled data $B^{t}$ into known data, seen data, and unseen data using both the $\theta^{off}$ and $\theta^{on}$ through a two-stage process. In the first stage, we split the data into known and unknown categories. We obtain logits $z$ using the offline model $\theta^{off}(\cdot) = g^{off}(f^{off}(\cdot))$ and calculate the energy scores using these logits. The formulation for calculating the energy scores is as follows: + +$$ +z ^ {t} = \theta^ {\text {o f f}} \left(B ^ {t}\right) \in \mathbb {R} ^ {\left| B ^ {t} \right| \times \left| \mathcal {Y} _ {l} \right|}, \tag {2} +$$ + +$$ +e ^ {t} = \left\{e _ {i} ^ {t} \mid e _ {i} ^ {t} = - \log \sum_ {j = 1} ^ {| \mathcal {Y} _ {i} |} \exp \left(z _ {i, j} ^ {t}\right) \right\}, \tag {3} +$$ + +where $\exp (\cdot)$ denotes an exponential function, $e^t$ denotes a set of energy scores, and $z_{i,j}$ denotes the logit from $i$ -th sample corresponding to the $j$ -th class. We divide the energy scores by the Gaussian Mixture Model (GMM). We initially set the GMM to have two clusters and use the energy score of each sample as input for clustering. We then label the cluster with a low energy score as "known" and the cluster with a high energy score as "unknown". This division allows us to split the unlabeled data into known and unknown data. In the second stage, we divide the unknown data into seen and unseen categories. Seen data comprise the data belonging to observed novel categories during the incremental session, while unseen data belong to undiscovered categories. To split the unknown data, we utilize the online model $\theta^{on}(\cdot) = g^{on}(f^{on}(\cdot))$ to extract energy scores. The process in the second stage is the same as the first stage, where we split the unknown data into seen and unseen categories based on their energy scores. For the initial batch of incremental sessions, the online model is identical to the offline model. Additionally, since the initial batch data is the first data of the incremental session, any data classified as unknown is assumed to be unseen data. The proposed energy-guided discovery splits unlabeled data into known, seen, and unseen categories effectively. Unlike prior methods which require sufficient data, it can identify novel categories with batch-wise data. + +# 3.4 Pseudo-Labeling via Variance-based Feature Augmentation + +After the discovery step, we pseudo-label known, seen, and unseen data to assign labels to each sample. For known and seen data, where the model has already learned knowledge for the categories, we utilize predictions from the offline model for known data and from the online model for seen data to assign pseudo-labels, respectively. Conversely, as the unseen data belong to undiscovered categories, the online model cannot predict proper categories for them and needs to discover novel categories. Moreover, due to the given batch-wise data, discovering novel categories to assign pseudo labels is challenging. + +To address this challenge, we propose variance-based feature augmentation, which temporally augments feature vectors using variational knowledge from unseen data to acquire accurate novel clusters. We obtain feature vectors of the unseen data $B_{u}^{t}$ using the online feature extractor $f^{on}(\cdot)$ denoted as $h_u^t = f^{on}(B_u^t)$ . We calculate the variance of the unseen feature vectors and augment unseen feature vectors based on the variance. We employ each feature vector to augment features by the Gaussian distribution $\hat{h}_u^t = \{\hat{h}_u\sim \mathcal{N}(h_u,\sigma_u^2)\} ,h_u\in h_u^t$ . The formulation of the variance-based feature augmentation is as follows: + +$$ +h _ {a u g} ^ {t} = \{h _ {u} ^ {t}; \hat {h} _ {u} ^ {t} \}, \quad \hat {h} _ {u} ^ {t} = \{\hat {h} _ {u} \sim \mathcal {N} (h _ {u}, \sigma_ {u} ^ {2}) \} \in \mathbb {R} ^ {(\left| B _ {u} ^ {t} \right| \cdot K) \times d}, \qquad (4) +$$ + +where $\hat{h}_u$ and $\sigma_u$ represent augmented feature vectors and the standard deviation of the unseen feature vectors, respectively. $K$ denotes the hyperparameter for the number of augmented feature vectors, and $h_{aug}^t$ denotes the concatenation of unseen feature vectors and augmented feature vectors. Novel categories are discovered based on $h_{aug}^t$ . Augmented feature vectors expand the feature space with variational knowledge, which helps the model discover novel clusters. + +We employ affinity propagation [9] to assign pseudo-labels to unseen data. Affinity propagation is a non-parametric clustering algorithm that does not require a predefined number of clusters, using the similarity matrix of the data. Given the absence of prior knowledge of novel categories, we extend the online classifier with the estimated number of novel clusters obtained by affinity propagation. The proposed variance-based feature augmentation is a simple yet effective approach that helps the model assign effective pseudo-labels based on accurate novel category discovery. + +# 3.5 Online Continual Learning with Pseudo-Labeled Data + +For online continual learning, we employ parameter-efficient tuning to learn novel representation knowledge while preserving previously learned knowledge, as it effectively mitigates catastrophic forgetting. We use cross-entropy loss with pseudo-labeled data in an online manner to capture novel knowledge. However, adapting the model with batch-wise data raises recency bias, where sufficient training through the given batch-wise data impedes discriminative knowledge. To address this, we propose an energy-based contrastive loss $(\mathcal{L}_{ec})$ for effective discriminative learning. The energy-based contrastive loss focuses on inducing + +the online model to capture the knowledge of novel categories. It amplifies the energy from the known classifier while reducing the energy from the novel classifier. By selectively applying $\mathcal{L}_{ec}$ to novel data $B_{n}^{t}$ including seen and unseen data, it helps the online model learn novel discriminative knowledge. The formulation of $\mathcal{L}_{ec}$ is as follows: + +$$ +\operatorname {E} (f (x); g) = - \log \sum_ {i = 1} ^ {| \mathcal {Y} |} \exp \left(g _ {i} (f (x))\right), \tag {5} +$$ + +$$ +\mathcal {L} _ {e c} = \frac {1}{| B _ {n} ^ {t} |} \sum_ {(x _ {n}) \sim B _ {n} ^ {t}} \log \left(1 + \frac {\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {n e w})}{\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {o l d})}\right), \tag {6} +$$ + +where $\operatorname{E}(\cdot ;\cdot)$ and $|\mathcal{V}|$ denote energy function and the number of nodes at the classifier, respectively, $g_{i}(\cdot)$ represents the $i$ -th node from the given classifier $g$ , and $g^{old}$ and $g^{new}$ denote the nodes corresponding to known categories and discovered novel categories in the online classifier $g^{on}(\cdot)$ , respectively. Eq. 5 represents the calculation of the energy score given the feature vector and classifier. The total loss $\mathcal{L}_{inc}$ for the incremental session can be summarized as follows: + +$$ +\mathcal {L} _ {i n c} = \mathcal {L} _ {c e} + \mathcal {L} _ {e c}, \tag {7} +$$ + +where $\mathcal{L}_{ec}$ represents energy-based contrastive loss, and $\mathcal{L}_{ce}$ denotes cross-entropy loss with pseudo-labeled data. + +# 4 Experiments + +# 4.1 Experimental Setup + +Baselines and Implementation Details. Recently, many approaches have emerged in novel category discovery. However, many of them focus on self-supervised learning with non-parametric classifiers (e.g. K-means) [42] and rarely consider continual learning [45]. Some methods that consider continual learning assume prior knowledge for novel category discovery. Since the proposed EGD and fine-split from CGCD require a classifier to estimate the number of novel categories, non-parametric classifiers cannot be used. Additionally, existing methods not designed for continual learning are unsuitable baselines due to severe forgetting of known classes. For a fair comparison, we mainly compared our method with FRoST [37] and CGCD [22], which focus on continual learning and utilize parametric classifiers for novel category discovery. + +We used ViT-B-16 [6] pre-trained on DINO-ImageNet [3] as the backbone network for all methods, including ours. We trained our method 30 epochs for the base session and 15 epochs for the incremental session. We utilized the AdamW optimizer with a weight decay of 1e-4 and a learning rate of 1e-3. We trained our model using a single RTX 3090 GPU with a batch size set to 64. + +Datasets. We evaluated our method with state-of-the-art (SoTA) continuous category discovery methods on three datasets: CUB200 [41], FGVC-Aircraft [31], and CIFAR-100 [24]. For continual learning, we partitioned CUB200 and FGVC-Aircraft following the same split configuration in CGCD. For CIFAR-100, we divided 80 classes for known classes and 20 classes for novel classes. To follow the GCD setting, we divided the training samples into labeled and unlabeled per class, maintaining a ratio of 0.8 for the labeled to 0.2 for the unlabeled. + +Evaluation Metrics. We evaluated the methods by measuring clustering accuracy on test dataset $\mathcal{D}$ . We utilized the hungarian algorithm [25] to obtain the optimal permutation $h^*$ that aligns the prediction $y_i^*$ from the model with the ground truth label $y_i$ : + +$$ +h ^ {*} = \arg \min _ {h} \frac {1}{| \mathcal {D} |} \sum_ {i = 1} ^ {| \mathcal {D} |} \mathbb {I} \left(y _ {i} = h \left(y _ {i} ^ {*}\right)\right), \tag {8} +$$ + +where $|\mathcal{D}|$ is the size of the test dataset. Clustering accuracy on $\mathcal{D}$ is defined as follows: + +$$ +M = \frac {1}{| \mathcal {D} |} \sum_ {i = 1} ^ {| \mathcal {D} |} \mathbb {I} \left(y _ {i} = h ^ {*} \left(y _ {i} ^ {*}\right)\right). \tag {9} +$$ + +We employed this metric to measure the clustering accuracy of all classes ( $M_{all}$ ), known classes ( $M_{old}$ ), and unknown classes ( $M_{new}$ ) in the cases of test dataset $\mathcal{D}^{base}$ and $\mathcal{D}^{inc}$ , respectively. Following continuous GCD approaches to evaluate the ability to maintain performance for known classes [22, 44], we employed the forgetting metric $F = M_{old}^{base} - M_{old}^{inc}$ , where $M_{old}^{base}$ , $M_{old}^{inc}$ refers to known class clustering accuracy at the base session and incremental session. A higher clustering accuracy metric is better, and a lower forgetting metric is better. This ensures the model is effective for real-world applications. + +Moreover, we evaluate the estimated pseudo-labels to validate their effectiveness in pseudo-labeling. Pseudo-labeling for unlabeled data is crucial as the model utilizes pseudo-labels to interpret the unlabeled data. The pseudo-labeling accuracy on the dataset $\mathcal{D}$ is defined as follows: + +$$ +M = \frac {1}{| \mathcal {D} |} \sum_ {i = 1} ^ {| \mathcal {D} |} \mathbb {I} \left(y _ {i} = h ^ {*} \left(\tilde {y} _ {i}\right)\right), \tag {10} +$$ + +where $\tilde{y}_i$ denotes the pseudo-label of the $i$ -th sample. + +# 4.2 Comparison with Baselines + +We compared the clustering performance of our proposed method, DEAN, with SoTA approaches, as shown in Table 1. Since we considered an online scenario for category discovery, traditionally approached using offline methods, there is no previous work for comparison. Therefore, for a fair comparison, our study adapts + +Table 1: Clustering performance comparison on CUB200, FGVC-Aircraft, and CIFAR-100. Best performance is highlighted in bold. All the baselines utilize the fine-split method from CGCD to handle the absence of prior knowledge. + +
MethodCUB200FGVC-AircraftCIFAR-100
MallMoldMnewF↓MallMoldMnewF↓MallMoldMnewF↓
Supervised72.5170.6879.6611.9166.2563.6276.7611.0381.3479.4888.8010.07
Fine-Tuning1.961.942.0180.522.061.305.1073.353.683.245.4586.26
FRoST [ECCV' 22]23.9828.725.4550.8514.4415.659.6048.316.136.215.8069.24
CGCD[ICCV' 23]40.4047.8811.1835.6114.0815.677.7560.002.642.762.1586.74
DEAN w/ FS55.5765.3717.2917.6240.3249.145.1025.4352.2762.1112.9027.55
DEAN (Ours)59.5666.3333.1116.2646.4453.5617.9920.9762.3470.3430.3519.17
+ +Table 2: Ablation experiment on CUB200. The baseline denotes fine-tuning with the fine split from the CGCD. + +
AblationCUB200
MallMoldNewF↓
Baseline1.961.942.0180.52
+ LoRA49.3360.356.2422.24
+ EGD54.9762.9823.6719.61
+ EC Loss57.3764.9827.6317.61
+ VFA (Ours)59.5666.3333.1116.26
+ +Table 3: Ablation study for the hyperparameter K of variance-based feature augmentation on CUB200. + +
AblationCUB200
\(M_{all}^{PS}\)\(M_{old}^{PS}\)\(M_{new}^{PS}\)
K=057.3185.1946.47
K=164.2788.9653.74
K=375.8793.4364.48
K=576.9394.0565.75
K=775.4493.4164.07
K=975.2293.563.36
+ +offline continuous NCD and GCD methods into OCGCD scenarios. The experiment includes two SoTA continuous category discovery methods: FRoST [37] and CGCD [22]. In addition, we include a supervised learning method as an upper bound and a fine-tuning method as a performance baseline. Note that in our proposed scenario, OCGCD, all baselines are not provided with prior knowledge about novel categories. Instead, we employed fine-split method from CGCD [22] which consists of two stages: initial split and fine split. During the initial split stage, unlabeled data is roughly divided into known and unknown categories based on cosine similarity. In the fine split stage, a split model is trained to distinguish between known and unknown data. The fine-split method estimates the number of novel categories conveying information about novel categories to the comparison methods. + +The results show that our method, DEAN, recorded a robust clustering performance, significantly surpassing existing continuous NCD and GCD methods. DEAN recorded consistently high metrics across $M_{all}$ , $M_{old}$ , and $M_{new}$ while keeping forgetting metric $F$ substantially low. This denotes our method can learn novel knowledge without forgetting. Moreover, it is noteworthy that DEAN w/ FS which denotes the adoption of fine-split from CGCD into DEAN instead + +Table 4: Ablation study for the cluster accuracy of pseudo-labels. + +
MethodCUB200
\( M_{all}^{PS} \)\( M_{old}^{PS} \)\( M_{new}^{PS} \)
CGCD [ICCV' 23]61.8881.0758.68
DEAN w/o (\( \mathcal{L}_{ec} \), VFA)57.2385.1145.80
DEAN w/o VFA57.3185.1946.47
DEAN (Ours)76.9394.0565.75
+ +Table 5: Ablation experiment of parameter-efficient tuning on CUB200. + +
MethodCUB200
\(M_{all}\)\(M_{old}\)\(M_{new}\)\(F\downarrow\)
DEAN(Ours)Adapter [4]56.9664.7526.5317.64
Prefix [27]58.5466.2928.2816.30
LoRA [17]59.5666.3333.1116.26
+ +of energy-guided discovery (EGD) also showed notable performance compared to the baselines. But, DEAN achieved significant performance gains across all datasets as the proposed EGD is applied. In contrast, existing continual methods showed shortcomings in clustering performance. Specifically, we observed extremely low clustering performance on CIFAR-100 dataset, low $M_{all}$ , $M_{old}$ , and $M_{new}$ values, coupled with high $F$ . On CIFAR-100 dataset, DEAN surpassed the most recent method CGCD assuming no prior knowledge for novel categories by $59.7\%$ , $67.58\%$ , $28.2\%$ , and $67.57\%$ for $M_{all}$ , $M_{old}$ , $M_{new}$ , and $F$ . + +Such observations show that while current GCD methods are vulnerable to online settings, DEAN can manage the balance required in the OCGCD scheme of retaining old knowledge while integrating the new. We believe this results shows the suitability of DEAN for real-world applications, where data is never static. We empirically confirmed that DEAN can evolve with the data it encounters, overcoming the challenges presented in the OCGCD scheme. + +# 4.3 Ablation Study + +The experimental results in Section 4.2 show that our DEAN achieved significant performance in OCGCD scenarios. Through further experiments on the CUB200, we provide a detailed analysis of the components proposed in our method. + +Effectiveness of Components. We validate the effectiveness of adding each component of our method to the baseline model. The baseline denotes fine-tuning the backbone model with a data-split following settings in CGCD. As shown in Table 2, results reveal that all components have an important impact on performance. Employing LoRA to baseline shows significant improvement in $M_{old}$ , $F$ . This shows that our application of LoRA on the frozen model at incremental session effectively mitigates forgetting. While LoRA shows only a slight increase in $M_{new}$ , additionally employing energy-guided discovery (EGD) and EC Loss markedly improves $M_{new}$ . This confirms that adopting energy guidance will enhance the ability of the model to differentiate between old and new categories. Finally, the integration of VFA further boosts the model performance on $M_{new}$ . This improvement reveals the importance of accurate pseudo-labeling since VFA is a technique applied after training. Each component is crucial, incrementally building upon the last to solidify the model's performance in our online category discovery scenario. + +Hyperparameter Analysis of VFA. Variance-based Feature Augmentation (VFA) is our proposed method of temporarily augmenting feature vectors to acquire accurate pseudo-labeling. Our analysis evaluated the effect of hyperparameter $K$ , the number of augmented feature vectors, on the accuracy of pseudolabels. From Table 3 we observed that employment of VFA is beneficial, increasing novel category pseudo-label accuracy from $46.47\%$ to $65.75\%$ . Among VFA employment, we observed that $K = 5$ yields the best performance across all metrics. When $K$ is less than 5, the smaller $K$ worsened performance. Conversely, values of $K$ greater than 5 did not benefit the model. We analyzed that since VFA augments features based on variance, as the number of $K$ increases, unseen samples can be clustered together without being distinguished. + +Pseudo-Labeling Accuracy. Accurate pseudo-labeling for clustering is important because it directly influences the performance of the model in incremental learning. Shown in Table 4, our proposed method, DEAN, significantly enhanced the clustering accuracy of pseudo-labels, achieving a $13.12\%$ improvement in $M_{all}^{PS}$ over the most recent method CGCD. Through the ablation study, we investigated the contributions of each step in DEAN. When DEAN adopts only parameter-efficient tuning, excluding EC loss and VFA, we observed a notable declination in $M_{new}^{PS}$ . While each ablation step showed gradually accuracy enhancement, the application of VFA elevates the performance of DEAN above CGCD. Thus, the experimental result revealed that VFA played a vital role in distinguishing novel categories and confirming the effectiveness of our method. + +Parameter Efficient Tuning. The ablation study in Table 5 assesses the impact of parameter-efficient tuning (PET) methods on the CUB200 dataset. Our method can employ various tuning methods such as Adapter [4] or Prefix-tuning [27], and in this experiment, we compared modules with LoRA [17], Adapter and Prefix-tuning. We set the size of all PET modules size to 5 and integrated them into the last 5 layers of ViT. Overall improvements are shown by adopting PET over existing methods reported in Table 1. Among PET methods, LoRA showed the highest clustering accuracy. The performance enhancement of LoRA may stem from its fast adaptation, which is highly effective in online continual learning scenarios that often depend on fewer samples for training. + +# 5 Conclusion + +In this paper, we introduced a novel scenario for category discovery, Online Continuous Generalized Category Discovery, which accounts for the dynamic nature of real-world data streams, leading to severe forgetting and poor novel category discovery. In addition, we proposed a novel framework, DEAN, which utilizes energy-based novel category discovery for the first time in this field and introduces variance-based feature augmentation to enhance accurate pseudo-labeling. Our DEAN demonstrated promising performance in the proposed novel scenario. However, DEAN could not guarantee inference performance midway through training. In future work, we plan to explore an online continual framework that can ensure inference performance during training through fast adaptation. + +# Acknowledgements + +This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)). + +# References + +1. Aljundi, R., Belilovsky, E., Tuytelaars, T., Charlin, L., Caccia, M., Lin, M., Page-Caccia, L.: Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32 (2019) +2. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: International conference on machine learning. pp. 233-242. PMLR (2017) +3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) +4. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) +5. Choe, S.A., Shin, A.H., Park, K.H., Choi, J., Park, G.M.: Open-set domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23943-23953 (2024) +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) +7. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021) +8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999) +9. Frey, B.J., Dueck, D.: Clustering by passing messages between data points. science 315(5814), 972-976 (2007) +10. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: International Conference on Machine Learning. pp. 8109-8126. PMLR (2022) +11. Gutmann, M., Hyvärinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010) + +12. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. arXiv preprint arXiv:2002.05714 (2020) +13. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019) +14. He, J., Mao, R., Shao, Z., Zhu, F.: Incremental learning in online scenario. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13926-13935 (2020) +15. Hsu, Y.C., Lv, Z., Kira, Z.: Learning to cluster in order to transfer across domains and tasks. arXiv preprint arXiv:1711.10125 (2017) +16. Hsu, Y.C., Lv, Z., Schlosser, J., Odom, P., Kira, Z.: Multi-class classification without multi-class labels. arXiv preprint arXiv:1901.00544 (2019) +17. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) +18. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021) +19. Joseph, K., Paul, S., Aggarwal, G., Biswas, S., Rai, P., Han, K., Balasubramanian, V.N.: Novel class discovery without forgetting. In: European Conference on Computer Vision. pp. 570-586. Springer (2022) +20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020) +21. Kim, D., Ryoo, K., Cho, H., Kim, S.: Splitnet: learnable clean-noisy label splitting for learning with noisy labels. arXiv preprint arXiv:2211.11753 (2022) +22. Kim, H., Suh, S., Kim, D., Jeong, D., Cho, H., Kim, J.: Proxy anchor-based unsupervised learning for continuous generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16688-16697 (2023) +23. Koh, H., Kim, D., Ha, J.W., Choi, J.: Online continual learning on class incremental blurry task configuration with anytime inference. arXiv preprint arXiv:2110.10031 (2021) +24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +25. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) +26. Li, J., Socher, R., Hoi, S.C.: Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020) +27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597 (2021) +28. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020) +29. Liu, Y., Tuytelaars, T.: Residual tuning: Toward novel category discovery without labels. IEEE Transactions on Neural Networks and Learning Systems (2022) +30. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022) + +31. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) +32. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995) +33. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989) +34. Moon, J.Y., Park, K.H., Kim, J.U., Park, G.M.: Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11731-11741 (2023) +35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021) +36. Park, K.H., Song, K., Park, G.M.: Pre-trained vision and language transformers are few-shot incremental learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23881-23890 (2024) +37. Roy, S., Liu, M., Zhong, Z., Sebe, N., Ricci, E.: Class-incremental novel class discovery. In: European Conference on Computer Vision. pp. 317-333. Springer (2022) +38. Seo, J., Lee, S.H., Lee, T.Y., Moon, S., Park, G.M.: Generative unlearning for any identity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9151-9161 (2024) +39. Troisemaine, C., Lemaire, V., Gosselin, S., Reiffers-Masson, A., Flocon-Cholet, J., Vaton, S.: Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028 (2023) +40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022) +41. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +42. Wu, Y., Chi, Z., Wang, Y., Feng, S.: Metagcd: Learning to continually learn in generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1655-1665 (2023) +43. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021) +44. Zhang, X., Jiang, J., Feng, Y., Wu, Z.F., Zhao, X., Wan, H., Tang, M., Jin, R., Gao, Y.: Grow and merge: A unified framework for continuous categories discovery. Advances in Neural Information Processing Systems 35, 27455-27468 (2022) +45. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021) +46. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021) +47. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021) \ No newline at end of file diff --git a/2024/Online Continuous Generalized Category Discovery/images.zip b/2024/Online Continuous Generalized Category Discovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0832e4e6e354935f6ce2325276b115d1acd448a --- /dev/null +++ b/2024/Online Continuous Generalized Category Discovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7997c3d7dec028c7d88df0683d44eb1a671d1a759218eaa4574b0ce25da76a +size 374340 diff --git a/2024/Online Continuous Generalized Category Discovery/layout.json b/2024/Online Continuous Generalized Category Discovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a855e9769bfb208de1f50e8e00ff975bdd59d933 --- /dev/null +++ b/2024/Online Continuous Generalized Category Discovery/layout.json @@ -0,0 +1,9026 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "spans": [ + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": "Keon-Hee Park1, Hakyung Lee2, Kyungwoo Song2†, and Gyeong-Moon Park1†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 200, + 201, + 412, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 201, + 412, + 224 + ], + "spans": [ + { + "bbox": [ + 200, + 201, + 412, + 224 + ], + "type": "text", + "content": "1 Kyung Hee University, Yongin, Republic of Korea {khpark, gmpark}@ku.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 224, + 412, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 224, + 412, + 246 + ], + "spans": [ + { + "bbox": [ + 202, + 224, + 412, + 246 + ], + "type": "text", + "content": "2 Yonsei University, Seoul, Republic of Korea {hakyunglee0417, kyungwoo.song}@yonsei.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 277, + 455, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 277, + 455, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 277, + 455, + 475 + ], + "type": "text", + "content": "Abstract. With the advancement of deep neural networks in computer vision, artificial intelligence (AI) is widely employed in real-world applications. However, AI still faces limitations in mimicking high-level human capabilities, such as novel category discovery, for practical use. While some methods utilizing offline continual learning have been proposed for novel category discovery, they neglect the continuity of data streams in real-world settings. In this work, we introduce Online Continuous Generalized Category Discovery (OCGCD), which considers the dynamic nature of data streams where data can be created and deleted in real time. Additionally, we propose a novel method, DEAN, Discovery via Energy guidance and feature Augmentation, which can discover novel categories in an online manner through energy-guided discovery and facilitate discriminative learning via energy-based contrastive loss. Furthermore, DEAN effectively pseudo-labels unlabeled data through variance-based feature augmentation. Experimental results demonstrate that our proposed DEAN achieves outstanding performance in proposed OCGCD scenario. The implementation code is available at https://github.com/KHU-AGI/OCGCD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "spans": [ + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "text", + "content": "Keywords: Online Continual Learning " + }, + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "text", + "content": " Generalized Category Discovery " + }, + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "text", + "content": " Energy-Guided Discovery " + }, + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 485, + 453, + 509 + ], + "type": "text", + "content": " Variance-based Feature Augmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 547, + 230, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 230, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 230, + 560 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 573, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 647 + ], + "type": "text", + "content": "The rapid progress of deep neural networks in computer vision has greatly facilitated the widespread adoption of artificial intelligence into real-world applications. Nevertheless, the networks still have challenges in embodying high-level human capabilities such as object recognition and grouping. To emulate these sophisticated human abilities, there has been a lot of interest in the field of category discovery. Specifically, generalized category discovery (GCD) [40] demands" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 652, + 237, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 237, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 237, + 665 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 278, + 185 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 278, + 185 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 278, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 278, + 185 + ], + "type": "image", + "image_path": "184125e7f3dba4888e0400978e84885bda23fd4536f509dabc7e170d2d8ea667.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 185, + 278, + 258 + ], + "blocks": [ + { + "bbox": [ + 134, + 185, + 278, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 185, + 278, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 185, + 278, + 258 + ], + "type": "image", + "image_path": "4dafa45173861cdfd4c6f145b1996db797c0fe53d32488d56ae5b09ea596467f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 262, + 280, + 279 + ], + "lines": [ + { + "bbox": [ + 132, + 262, + 280, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 280, + 279 + ], + "type": "text", + "content": "(a) Experiment of prior methods in an online manner on CUB200." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 288, + 119, + 482, + 177 + ], + "blocks": [ + { + "bbox": [ + 288, + 119, + 482, + 177 + ], + "lines": [ + { + "bbox": [ + 288, + 119, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 288, + 119, + 482, + 177 + ], + "type": "image", + "image_path": "c918eeb5a4da590fabc540a08cd5dec57cc4d20bedee7d748a2691c6be28b240.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 288, + 185, + 481, + 261 + ], + "blocks": [ + { + "bbox": [ + 288, + 185, + 481, + 261 + ], + "lines": [ + { + "bbox": [ + 288, + 185, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 288, + 185, + 481, + 261 + ], + "type": "image", + "image_path": "1389858bf68b881018020a15319c4abc3ba03a9f0f2ecd036ee36d978dc9b438.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 262, + 465, + 272 + ], + "lines": [ + { + "bbox": [ + 302, + 262, + 465, + 272 + ], + "spans": [ + { + "bbox": [ + 302, + 262, + 465, + 272 + ], + "type": "text", + "content": "(b) The description of the proposed OCGCD." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 289, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 334 + ], + "type": "text", + "content": "Fig. 1: Figure 1a shows that existing methods recorded poor performance in online training, suggesting that prior methods cannot handle online continual learning. Figure 1b shows the proposed scenario, OCGCD. As our scenario assumes batch-wise online learning, the model suffers severe forgetting and poor novel category discovery." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 421 + ], + "type": "text", + "content": "advanced recognition and grouping skills to discover unknown categories within unlabeled datasets. In general, existing methods for GCD rely on an impractical assumption that they can access both labeled and unlabeled datasets simultaneously during training, which prevents these methods from meeting real-world application demands." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 424, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 604 + ], + "type": "text", + "content": "In this context, a new scenario of continuous category discovery [19,22,37,42, 44] has emerged to address the unrealistic assumption by adopting offline continual learning. Despite its great advantage not requiring the assumption, these offline-based methods have inherent limitations in handling the continuous nature of data streams in the real world. Current approaches to offline continual learning primarily focus on static data streams, thus failing to adequately accommodate the dynamic nature of data streams where data can be created and deleted in an online manner. To verify this limitation, we conducted an empirical analysis using the CUB200 [41] dataset, dividing it into 160 labeled categories and 40 unlabeled categories. We evaluated the performance of existing methods [22,37] for continuous category discovery in an online manner. As shown in Figure 1a, we observed that existing methods [22,37] exhibited inferior performance in identifying novel categories, thereby resulting in poor overall performances. This indicates that the offline-based GCD methods have difficulty in clustering novel data effectively in the online learning scenario." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In light of this observation, in this paper, we introduce a novel online learning scenario in continuous generalized category discovery, called Online Continuous Generalized Category Discovery (OCGCD). As shown in Figure 1b, OCGCD involves batch-wise training of unlabeled data in an online continual manner without any prior knowledge of the unknown data. Therefore, our novel" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "scenario closely replicates real-world data streams characterized by dynamic creation and elimination of data during training [14]. Here we summarize the key challenges of our novel OCGCD scenario as follows: 1) Challenging category discovery. The model struggles to retain the knowledge of trained categories because of the continuous nature of the data, making novel category discovery challenging. 2) Noisy pseudo-labeling. Online learning restricts access to the entire training dataset, leading to inaccurate category assignments and the generation of severe noise pseudo-labels for unlabeled data. 3) Severe catastrophic forgetting. Online learning erases observed data after training, resulting in severe forgetting of previously acquired knowledge [33]. Hence, our novel OCGCD scenario requires addressing the aforementioned challenges to achieve successful learning outcomes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 488 + ], + "type": "text", + "content": "To address the challenges of OCGCD, we propose a novel approach called DEAN, Discovery via Energy Guidance and Feature Augmentation, which utilizes the energy score [28] to discover unknown data and further discriminate whether the data is seen or unseen and employs feature augmentation for better pseudo-labeling. By comparing the energy values of known and unknown data, DEAN effectively identifies unknown data among them, facilitating the clustering of novel categories. On top of that, we introduce an energy-based contrastive loss to enhance learning from unknown samples, promoting the acquisition of more discriminative knowledge during online learning. In addition, we propose a novel variance-based feature augmentation, a simple yet effective approach for accurate novel clustering and pseudo-labeling. In our scenario, since the model can access only batch-wise training data at once, it struggles to cluster data accurately. Variance-based feature augmentation (VFA) temporally extends the number of feature vectors through variational information to promote accurate clustering. To mitigate catastrophic forgetting, we employ parameter-efficient tuning, especially LoRA [17] in our framework while preserving the pre-trained weights. Extensive experiments demonstrate that our proposed method, DEAN, enables the network to discover novel categories more effectively in the OCGCD scenario than existing methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 488, + 422, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 488, + 422, + 500 + ], + "spans": [ + { + "bbox": [ + 147, + 488, + 422, + 500 + ], + "type": "text", + "content": "We summarize the main contributions of this paper as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 509, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 509, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 480, + 556 + ], + "type": "text", + "content": "- We introduce a novel Online Continuous Generalized Category Discovery (OCGCD) scenario that effectively reflects the online characteristics of real-world data. To tackle this scenario, we propose a novel method called Discovery via Energy Guidance and Feature Augmentation (DEAN)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 557, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 592 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first work to introduce the energy score for novel category discovery. We propose the energy-based contrastive loss to enhance online learning of discovered unknown data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 594, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 480, + 628 + ], + "type": "text", + "content": "- For effective pseudo-labeling, we propose a new variance-based feature augmentation, called VFA. The proposed VFA enhances the sample clustering, leading to the improvement of pseudo-label quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 630, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 630, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 630, + 480, + 664 + ], + "type": "text", + "content": "- Comprehensive experiments demonstrate that our novel method achieves significant performance improvements in the OCGCD scenario compared to state-of-the-art models." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 293, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 293, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 293, + 153 + ], + "type": "text", + "content": "2.1 Novel Category Discovery" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 399 + ], + "type": "text", + "content": "Novel Category Discovery (NCD) focuses on the task of identifying and categorizing unlabeled data [39]. To effectively discover and classify data, this field requires advanced object recognition and clustering skills. Most of the NCD methods fall into two training scenarios: disjoint training and joint training. The disjoint training scenario, also known as a two-stage approach, starts by pre-training the model on a labeled dataset, followed by fine-tuning on an unlabeled dataset from a different domain [13, 15, 16, 29]. The joint training scenario adopts a more integrated approach, simultaneously utilizing both labeled and unlabeled datasets during the training process [7, 12, 18, 46, 47]. While this scenario outperforms disjoint training approach [29], its dependence on the availability of labeled data makes it less feasible in real-world applications. Accordingly, more research has been directed towards designing scenarios where they mitigate constraints of NCD for better alignment with real-world settings. Generalized Category Discovery (GCD) [40] broadens the setting of NCD by allowing test data to contain both known and unknown classes, in contrast to NCD which focuses on exclusively unknown classes. Moreover, to better represent real-world scenarios, several methods [19, 22, 37, 42, 44] have been proposed that integrate continual learning into category discovery. However, existing methods predominantly utilizes an offline continual learning framework, which remains an impractical assumption for real-world applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": "We propose an online version of continuous GCD that tackles the challenges of online learning through energy-guided clustering and energy-based contrastive loss. Energy-guided clustering utilizes the trained classifier to extract energy from each sample, efficiently discovering novel categories. Moreover, applying energy-based contrastive loss enables the model to effectively learn feature knowledge, thereby enhancing the capacity of the model in online continual learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 490, + 295, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 295, + 502 + ], + "type": "text", + "content": "2.2 Online Continual Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Online Continual Learning (OCL) is a more realistic continual learning scenario where the model learns from data arriving in small, sequential batches in real-time, without access to previously seen batches [30]. A key aspect of continual learning is to mitigate catastrophic forgetting and balance existing knowledge while adapting to new knowledge [8, 32, 33]. The dynamic distribution of online setting intensifies this challenge, and various approaches have been introduced to address this issue. MIR [1] proposed a memory retrieval technique for replay, selecting mostly interfered samples. OCM [10] deals with forgetting by maximizing mutual information between past and current data to learn more transferable features. i-Blurry [23] introduces a realistic setting that combines disjoint and blurry class distributions, mitigating forgetting by employing sample-wise importance sampling. Furthermore, Si-Blurry [34] proposed a new blurry task scenario, acknowledging the stochastic properties in real-world data distributions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "In this paper, we propose a novel setting that integrates online continual learning with generalized category discovery. By emphasizing the online learning, OCGCD presents a new perspective in the field of category discovery. It is a more realistic and challenging scenario than existing approaches in GCD, taking into account the continuous flow of real-world data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 193, + 221, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 221, + 204 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 221, + 204 + ], + "type": "text", + "content": "2.3 Noise Label" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "type": "text", + "content": "Deep neural networks can overfit to noisy label data, leading to a substantial decrease in the robustness and generalization ability [2, 43]. Consequently, tackling noisy label data becomes essential, especially in category discovery where pseudo-labeling of unlabeled data plays a pivotal role in identifying new categories. Recent studies have primarily focused on discriminating between clean data and noisy label data. DivideMix [26] use a Gaussian Mixture Model (GMM) generated from the training dataset to differentiate between clean and noisy data, while AugDesc [35] applies various data augmentations to the training dataset to amplify the distinction. SplitNet [21] introduces a compact learnable module that separates clean and noisy label data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 483, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 483, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 483, + 452 + ], + "type": "text", + "content": "In OCGCD, learning occurs without prior knowledge of novel categories, relying only on the current batch-wise data to predict the number of clusters and generate pseudo-labels based on this prediction. This scenario makes it difficult to apply existing techniques that depend on utilizing the training dataset or employing learnable modules. Therefore, we propose Variance-based Feature Augmentation (VFA) to effectively separate clean and noisy labels in our framework. VFA can diversely augment (re-generate) the feature vectors of given data, enabling the creation of discriminative clusters for accurate pseudo-labeling. Unlike CGCD [22], which requires training additional modules for the discovery, VFA facilitates end-to-end training without the additional modules or extra learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 468, + 459, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 459, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 459, + 483 + ], + "type": "text", + "content": "3 Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 493, + 385, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 493, + 385, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 493, + 385, + 504 + ], + "type": "text", + "content": "3.1 Problem Formulation and Method Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": "Problem Formulation. In online continuous GCD, the training dataset " + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathcal{D}^{base},\\mathcal{D}^{inc}\\}" + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": " is given sequentially, where " + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{base} = \\{(x_i,y_i)\\in \\mathcal{X}\\times \\mathcal{Y}_l\\}_{i = 1}^{|D^{base}|}" + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": " is the labeled dataset for the base session and " + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{inc} = \\{(x_i,y_i)\\in \\mathcal{X}\\times \\mathcal{Y}_u\\}_{i = 1}^{|D^{inc}|}" + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": " is the unlabeled dataset for the incremental session. The proposed novel scenario assumes an overlap between the labeled category set " + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_l" + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": " and the unlabeled category set " + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_u" + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": ", as it is a sub-category of the generalized category discovery. In this setting, the network trains on " + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{base}" + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": " and on " + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{inc}" + }, + { + "bbox": [ + 130, + 512, + 482, + 641 + ], + "type": "text", + "content": " in an offline and online manner, respectively. Our proposed online continuous GCD aims to facilitate the discovery of novel categories in real-world scenarios while preserving the clustering knowledge obtained from the offline training of the labeled dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Method Overview. Figure 2 shows an overview of the proposed method, where the pre-trained vision transformer [6] via self-supervised learning serves as the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 261 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 261 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 261 + ], + "type": "image", + "image_path": "2469bdf2f8a83daea607ee94603a102b607d121b41e0f330ae960e63db22fdfe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "text", + "content": "Fig. 2: Overall process of the proposed DEAN framework. The energy-guided discovery splits unlabeled data into known, seen, and unseen data for better novel category discovery, while variance-based feature augmentation enhances the clustering of unseen data. " + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ec}" + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "text", + "content": " facilitates better discriminative learning in the online continual learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 335, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 468 + ], + "type": "text", + "content": "backbone network. We initially train the network using the labeled dataset in the base session (Section 3.2). We introduce energy-guided discovery to identify unknown samples from the given data. Our energy-guided discovery involves a two-step process to determine whether data belong to known or unknown categories and whether the unknown data belong to previously seen or unseen categories (Section 3.3). We split the data into known, seen, and unseen categories through energy-guided discovery. For pseudo-labeling of unseen data, we propose variance-based feature augmentation to improve the clustering of unseen data, resulting in better pseudo-labeling (Section 3.4). Subsequently, the network trains pseudo-labeled data in an online manner using an energy-based contrastive loss for better discriminative learning (Section 3.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 482, + 346, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 346, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 346, + 495 + ], + "type": "text", + "content": "3.2 Supervised Learning at Base Session" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "text", + "content": "During the base session, we train the offline model " + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\theta^{off}" + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "text", + "content": " consisting of the feature extractor " + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "inline_equation", + "content": "f^{off}(\\cdot):\\mathcal{X}\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "text", + "content": " and classifier head " + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "inline_equation", + "content": "g^{off}(\\cdot):\\mathbb{R}^d\\rightarrow \\mathcal{Y}_l" + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "text", + "content": " in an offline manner. As the offline network " + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\theta^{off} = \\{f^{off},g^{off}\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 595 + ], + "type": "text", + "content": " trains on the labeled dataset during the base session, we employ the cross-entropy loss for supervised learning. While existing novel category discovery methods often utilize self-supervised [11] and supervised [20] contrastive loss for representation learning without a classifier head, our approach utilizes a classifier for clustering. The formulation of the base session training is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 178, + 599, + 481, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 599, + 481, + 637 + ], + "spans": [ + { + "bbox": [ + 178, + 599, + 481, + 637 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = - \\mathbb {E} _ {(\\mathcal {X}, \\mathcal {Y} _ {l}) \\sim \\mathcal {D} ^ {b a s e}} \\frac {1}{| \\mathcal {D} ^ {b a s e} |} \\sum_ {i = 1} ^ {| \\mathcal {D} ^ {b a s e} |} y _ {i} \\cdot \\log \\delta (g ^ {\\text {o f f}} (f ^ {\\text {o f f}} (x _ {i}))), \\tag {1}", + "image_path": "d7f1b7fd62f6104955dd982d0514c18aac8b055372d49d68131b707a9b184134.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\delta(\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the softmax function. We train the model to capture representation knowledge during the base session training. After training the base session, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 282, + 229 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 282, + 229 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 282, + 229 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 282, + 229 + ], + "type": "image", + "image_path": "a06de2bd6aa5851dd24a2615dc12694c5fbcecd03d7f3b181a0ebcdb05407e6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 258, + 282, + 371 + ], + "blocks": [ + { + "bbox": [ + 133, + 232, + 284, + 257 + ], + "lines": [ + { + "bbox": [ + 133, + 232, + 284, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 232, + 284, + 257 + ], + "type": "text", + "content": "(a) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using fine-split of CGCD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 258, + 282, + 371 + ], + "lines": [ + { + "bbox": [ + 133, + 258, + 282, + 371 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 282, + 371 + ], + "type": "image", + "image_path": "6651a230f898248f0056e6c40976e2efd380444e632109993318f18b3a90e3b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 374, + 285, + 400 + ], + "lines": [ + { + "bbox": [ + 132, + 374, + 285, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 285, + 400 + ], + "type": "text", + "content": "(c) Comparing normal distributions of known and unknown samples on CUB200 using fine-split of CGCD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 114, + 481, + 231 + ], + "blocks": [ + { + "bbox": [ + 329, + 114, + 481, + 231 + ], + "lines": [ + { + "bbox": [ + 329, + 114, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 329, + 114, + 481, + 231 + ], + "type": "image", + "image_path": "423210ebdc5c836cb61062f186d84d91b4f848a9b95c54a3de3d635a61a7f800.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 232, + 481, + 257 + ], + "lines": [ + { + "bbox": [ + 329, + 232, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 329, + 232, + 481, + 257 + ], + "type": "text", + "content": "(b) Comparing normal distributions of known and unknown samples on FGVC-Aircraft using energy scores." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 329, + 258, + 481, + 373 + ], + "blocks": [ + { + "bbox": [ + 329, + 258, + 481, + 373 + ], + "lines": [ + { + "bbox": [ + 329, + 258, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 329, + 258, + 481, + 373 + ], + "type": "image", + "image_path": "6d9bb272b821e9069b802168bec71a34348acd5b39d884d799d17b5797798278.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 374, + 481, + 400 + ], + "lines": [ + { + "bbox": [ + 329, + 374, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 329, + 374, + 481, + 400 + ], + "type": "text", + "content": "(d) Comparing normal distributions of known and unknown samples on CUB200 using energy scores." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 409, + 482, + 433 + ], + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 433 + ], + "type": "text", + "content": "Fig. 3: Validating the effectiveness of energy scores for novel category discovery without prior knowledge about novel categories by comparison with existing methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 456, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 506 + ], + "type": "text", + "content": "freeze the offline model and utilize the model at the discovery step to distinguish between known and unknown data, enabling accurate discovery. In addition, we initialize the online model, denoted as " + }, + { + "bbox": [ + 130, + 456, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\theta^{on}" + }, + { + "bbox": [ + 130, + 456, + 482, + 506 + ], + "type": "text", + "content": ", based on the parameters of the offline model for the online continual learning on the unlabeled dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 524, + 403, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 524, + 403, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 524, + 403, + 537 + ], + "type": "text", + "content": "3.3 Energy-Guided Discovery to Identify Unknowns" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "The discovery step aims to discover new categories within the provided batchwise data. While CGCD, which assumes the absence of prior knowledge about novel categories, introduced a two-stage split process that depends on learning an additional split network to detect novel data, it has some limitations in our proposed scenario. CGCD includes the requirement for extra parameters and additional learning stage. Moreover, in our proposed online scenario where the model cannot access the entire training data, the split network faces difficulty in identifying novel samples due to severe forgetting of the split network. Liu et al. [28] proposed a method for detecting out-of-distribution (OOD) data by analyzing the difference in energy scores between in-distribution (ID) and OOD" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 448, + 102 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "data. We conducted experiments to compare the two-stage split process from CGCD with the energy-based discovery method to identify trained categories (known to the model) and untrained categories (unknown to the model)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 480, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 480, + 274 + ], + "type": "text", + "content": "In Figure 3, we found that the two-stage split of the CGCD struggled to identify both known and unknown categories in our proposed online learning scenario. Online learning with batch-wise data led to severe forgetting in the split network, resulting in poor detection of unknown samples. In contrast, energy-based discovery showed better performance in novel category discovery without extra parameters compared to CGCD. Moreover, as it does not require an additional learning phase, energy-based discovery enables end-to-end training. Inspired by this observation, we propose an energy-guided discovery approach for novel category discovery. To the best of our knowledge, this is the first work to utilize the energy score for novel category discovery." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": "During the incremental session, the network trains unlabeled batch-wise data " + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "inline_equation", + "content": "B^{t} = \\{x_{1}, x_{2}, \\ldots, x_{|B^{t}|}\\}" + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": " is the index of the batch. We classify unlabeled data " + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "inline_equation", + "content": "B^{t}" + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": " into known data, seen data, and unseen data using both the " + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "inline_equation", + "content": "\\theta^{off}" + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "inline_equation", + "content": "\\theta^{on}" + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": " through a two-stage process. In the first stage, we split the data into known and unknown categories. We obtain logits " + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": " using the offline model " + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "inline_equation", + "content": "\\theta^{off}(\\cdot) = g^{off}(f^{off}(\\cdot))" + }, + { + "bbox": [ + 131, + 275, + 481, + 360 + ], + "type": "text", + "content": " and calculate the energy scores using these logits. The formulation for calculating the energy scores is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 372, + 481, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 372, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 236, + 372, + 481, + 388 + ], + "type": "interline_equation", + "content": "z ^ {t} = \\theta^ {\\text {o f f}} \\left(B ^ {t}\\right) \\in \\mathbb {R} ^ {\\left| B ^ {t} \\right| \\times \\left| \\mathcal {Y} _ {l} \\right|}, \\tag {2}", + "image_path": "68d7b0945d78fc070b18bab43fc61e183cb4cbe22cb41ddd8efdf4754e4476d0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 392, + 481, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 392, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 236, + 392, + 481, + 426 + ], + "type": "interline_equation", + "content": "e ^ {t} = \\left\\{e _ {i} ^ {t} \\mid e _ {i} ^ {t} = - \\log \\sum_ {j = 1} ^ {| \\mathcal {Y} _ {i} |} \\exp \\left(z _ {i, j} ^ {t}\\right) \\right\\}, \\tag {3}", + "image_path": "56fd390077b4b684f44cc8821f84dbff1b3e9024f6ac7d29ae089106b4301170.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\exp (\\cdot)" + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "content": " denotes an exponential function, " + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "e^t" + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "content": " denotes a set of energy scores, and " + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{i,j}" + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "content": " denotes the logit from " + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "content": "-th sample corresponding to the " + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "content": "-th class. We divide the energy scores by the Gaussian Mixture Model (GMM). We initially set the GMM to have two clusters and use the energy score of each sample as input for clustering. We then label the cluster with a low energy score as \"known\" and the cluster with a high energy score as \"unknown\". This division allows us to split the unlabeled data into known and unknown data. In the second stage, we divide the unknown data into seen and unseen categories. Seen data comprise the data belonging to observed novel categories during the incremental session, while unseen data belong to undiscovered categories. To split the unknown data, we utilize the online model " + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta^{on}(\\cdot) = g^{on}(f^{on}(\\cdot))" + }, + { + "bbox": [ + 131, + 437, + 482, + 666 + ], + "type": "text", + "content": " to extract energy scores. The process in the second stage is the same as the first stage, where we split the unknown data into seen and unseen categories based on their energy scores. For the initial batch of incremental sessions, the online model is identical to the offline model. Additionally, since the initial batch data is the first data of the incremental session, any data classified as unknown is assumed to be unseen data. The proposed energy-guided discovery splits unlabeled data into known, seen, and unseen categories effectively. Unlike prior methods which require sufficient data, it can identify novel categories with batch-wise data." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 459, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 459, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 459, + 129 + ], + "type": "text", + "content": "3.4 Pseudo-Labeling via Variance-based Feature Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 232 + ], + "type": "text", + "content": "After the discovery step, we pseudo-label known, seen, and unseen data to assign labels to each sample. For known and seen data, where the model has already learned knowledge for the categories, we utilize predictions from the offline model for known data and from the online model for seen data to assign pseudo-labels, respectively. Conversely, as the unseen data belong to undiscovered categories, the online model cannot predict proper categories for them and needs to discover novel categories. Moreover, due to the given batch-wise data, discovering novel categories to assign pseudo labels is challenging." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "text", + "content": "To address this challenge, we propose variance-based feature augmentation, which temporally augments feature vectors using variational knowledge from unseen data to acquire accurate novel clusters. We obtain feature vectors of the unseen data " + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "inline_equation", + "content": "B_{u}^{t}" + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "text", + "content": " using the online feature extractor " + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "inline_equation", + "content": "f^{on}(\\cdot)" + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "inline_equation", + "content": "h_u^t = f^{on}(B_u^t)" + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "text", + "content": ". We calculate the variance of the unseen feature vectors and augment unseen feature vectors based on the variance. We employ each feature vector to augment features by the Gaussian distribution " + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "inline_equation", + "content": "\\hat{h}_u^t = \\{\\hat{h}_u\\sim \\mathcal{N}(h_u,\\sigma_u^2)\\} ,h_u\\in h_u^t" + }, + { + "bbox": [ + 131, + 232, + 482, + 328 + ], + "type": "text", + "content": ". The formulation of the variance-based feature augmentation is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 337, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 337, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 181, + 337, + 481, + 354 + ], + "type": "interline_equation", + "content": "h _ {a u g} ^ {t} = \\{h _ {u} ^ {t}; \\hat {h} _ {u} ^ {t} \\}, \\quad \\hat {h} _ {u} ^ {t} = \\{\\hat {h} _ {u} \\sim \\mathcal {N} (h _ {u}, \\sigma_ {u} ^ {2}) \\} \\in \\mathbb {R} ^ {(\\left| B _ {u} ^ {t} \\right| \\cdot K) \\times d}, \\qquad (4)", + "image_path": "713796f11a579e05056eb032550170e91347b72b48a140d0b14e4d1129cc6ca3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\hat{h}_u" + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\sigma_u" + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "content": " represent augmented feature vectors and the standard deviation of the unseen feature vectors, respectively. " + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "content": " denotes the hyperparameter for the number of augmented feature vectors, and " + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "inline_equation", + "content": "h_{aug}^t" + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "content": " denotes the concatenation of unseen feature vectors and augmented feature vectors. Novel categories are discovered based on " + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "inline_equation", + "content": "h_{aug}^t" + }, + { + "bbox": [ + 130, + 363, + 482, + 435 + ], + "type": "text", + "content": ". Augmented feature vectors expand the feature space with variational knowledge, which helps the model discover novel clusters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": "We employ affinity propagation [9] to assign pseudo-labels to unseen data. Affinity propagation is a non-parametric clustering algorithm that does not require a predefined number of clusters, using the similarity matrix of the data. Given the absence of prior knowledge of novel categories, we extend the online classifier with the estimated number of novel clusters obtained by affinity propagation. The proposed variance-based feature augmentation is a simple yet effective approach that helps the model assign effective pseudo-labels based on accurate novel category discovery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 549, + 432, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 432, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 432, + 562 + ], + "type": "text", + "content": "3.5 Online Continual Learning with Pseudo-Labeled Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "For online continual learning, we employ parameter-efficient tuning to learn novel representation knowledge while preserving previously learned knowledge, as it effectively mitigates catastrophic forgetting. We use cross-entropy loss with pseudo-labeled data in an online manner to capture novel knowledge. However, adapting the model with batch-wise data raises recency bias, where sufficient training through the given batch-wise data impedes discriminative knowledge. To address this, we propose an energy-based contrastive loss " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ec})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " for effective discriminative learning. The energy-based contrastive loss focuses on inducing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "the online model to capture the knowledge of novel categories. It amplifies the energy from the known classifier while reducing the energy from the novel classifier. By selectively applying " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ec}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " to novel data " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "B_{n}^{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " including seen and unseen data, it helps the online model learn novel discriminative knowledge. The formulation of " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ec}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 186, + 481, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 186, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 228, + 186, + 481, + 220 + ], + "type": "interline_equation", + "content": "\\operatorname {E} (f (x); g) = - \\log \\sum_ {i = 1} ^ {| \\mathcal {Y} |} \\exp \\left(g _ {i} (f (x))\\right), \\tag {5}", + "image_path": "8b2817d1ef8035d1d3a463eaa79ec657af56b330db77429d21cb99fae52160c7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 239, + 481, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 239, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 200, + 239, + 481, + 271 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e c} = \\frac {1}{| B _ {n} ^ {t} |} \\sum_ {(x _ {n}) \\sim B _ {n} ^ {t}} \\log \\left(1 + \\frac {\\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {n e w})}{\\operatorname {E} (f ^ {o n} (x _ {n}) ; g ^ {o l d})}\\right), \\tag {6}", + "image_path": "4e335e8d8f92d35f344f44bca1be30771bc3899cadc1a33da278d4783e31acf2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\operatorname{E}(\\cdot ;\\cdot)" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}|" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " denote energy function and the number of nodes at the classifier, respectively, " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "g_{i}(\\cdot)" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": "-th node from the given classifier " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "g^{old}" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "g^{new}" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " denote the nodes corresponding to known categories and discovered novel categories in the online classifier " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "g^{on}(\\cdot)" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": ", respectively. Eq. 5 represents the calculation of the energy score given the feature vector and classifier. The total loss " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inc}" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " for the incremental session can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 269, + 363, + 481, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 363, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 269, + 363, + 481, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n c} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {e c}, \\tag {7}", + "image_path": "7587f800e4602560b9c3c717823959bc05038adc3b4b53b630f2a535aa529c33.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 384, + 483, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 483, + 408 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 483, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 384, + 483, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ec}" + }, + { + "bbox": [ + 131, + 384, + 483, + 408 + ], + "type": "text", + "content": " represents energy-based contrastive loss, and " + }, + { + "bbox": [ + 131, + 384, + 483, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 131, + 384, + 483, + 408 + ], + "type": "text", + "content": " denotes cross-entropy loss with pseudo-labeled data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 427, + 230, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 230, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 230, + 441 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 453, + 262, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 262, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 262, + 466 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "content": "Baselines and Implementation Details. Recently, many approaches have emerged in novel category discovery. However, many of them focus on self-supervised learning with non-parametric classifiers (e.g. K-means) [42] and rarely consider continual learning [45]. Some methods that consider continual learning assume prior knowledge for novel category discovery. Since the proposed EGD and fine-split from CGCD require a classifier to estimate the number of novel categories, non-parametric classifiers cannot be used. Additionally, existing methods not designed for continual learning are unsuitable baselines due to severe forgetting of known classes. For a fair comparison, we mainly compared our method with FRoST [37] and CGCD [22], which focus on continual learning and utilize parametric classifiers for novel category discovery." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "We used ViT-B-16 [6] pre-trained on DINO-ImageNet [3] as the backbone network for all methods, including ours. We trained our method 30 epochs for the base session and 15 epochs for the incremental session. We utilized the AdamW optimizer with a weight decay of 1e-4 and a learning rate of 1e-3. We trained our model using a single RTX 3090 GPU with a batch size set to 64." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "Datasets. We evaluated our method with state-of-the-art (SoTA) continuous category discovery methods on three datasets: CUB200 [41], FGVC-Aircraft [31], and CIFAR-100 [24]. For continual learning, we partitioned CUB200 and FGVC-Aircraft following the same split configuration in CGCD. For CIFAR-100, we divided 80 classes for known classes and 20 classes for novel classes. To follow the GCD setting, we divided the training samples into labeled and unlabeled per class, maintaining a ratio of 0.8 for the labeled to 0.2 for the unlabeled." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluated the methods by measuring clustering accuracy on test dataset " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": ". We utilized the hungarian algorithm [25] to obtain the optimal permutation " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "h^*" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": " that aligns the prediction " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "y_i^*" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": " from the model with the ground truth label " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 258, + 480, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 258, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 231, + 258, + 480, + 291 + ], + "type": "interline_equation", + "content": "h ^ {*} = \\arg \\min _ {h} \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h \\left(y _ {i} ^ {*}\\right)\\right), \\tag {8}", + "image_path": "d51da3fdb738dca768ca5dfbb0dedeb4218a1f90067341f5ced051e84db2472e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 479, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 299, + 479, + 323 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}|" + }, + { + "bbox": [ + 130, + 299, + 479, + 323 + ], + "type": "text", + "content": " is the size of the test dataset. Clustering accuracy on " + }, + { + "bbox": [ + 130, + 299, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 299, + 479, + 323 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 245, + 331, + 480, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 331, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 245, + 331, + 480, + 365 + ], + "type": "interline_equation", + "content": "M = \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h ^ {*} \\left(y _ {i} ^ {*}\\right)\\right). \\tag {9}", + "image_path": "ff82fbdebed3e5a14f3b04fa8acc2908ef99ce62f65e3112a9683943dbe00e00.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": "We employed this metric to measure the clustering accuracy of all classes (" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "M_{all}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": "), known classes (" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "M_{old}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": "), and unknown classes (" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "M_{new}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": ") in the cases of test dataset " + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{base}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{inc}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": ", respectively. Following continuous GCD approaches to evaluate the ability to maintain performance for known classes [22, 44], we employed the forgetting metric " + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "F = M_{old}^{base} - M_{old}^{inc}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "M_{old}^{base}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "inline_equation", + "content": "M_{old}^{inc}" + }, + { + "bbox": [ + 130, + 373, + 479, + 468 + ], + "type": "text", + "content": " refers to known class clustering accuracy at the base session and incremental session. A higher clustering accuracy metric is better, and a lower forgetting metric is better. This ensures the model is effective for real-world applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 469, + 479, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 479, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 479, + 516 + ], + "type": "text", + "content": "Moreover, we evaluate the estimated pseudo-labels to validate their effectiveness in pseudo-labeling. Pseudo-labeling for unlabeled data is crucial as the model utilizes pseudo-labels to interpret the unlabeled data. The pseudo-labeling accuracy on the dataset " + }, + { + "bbox": [ + 130, + 469, + 479, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 469, + 479, + 516 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 526, + 480, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 526, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 246, + 526, + 480, + 559 + ], + "type": "interline_equation", + "content": "M = \\frac {1}{| \\mathcal {D} |} \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\mathbb {I} \\left(y _ {i} = h ^ {*} \\left(\\tilde {y} _ {i}\\right)\\right), \\tag {10}", + "image_path": "2693351e6712568ecbdeac9ed942f1ca872cdb8bc44abaa499aaef415b61acb1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 567, + 366, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 567, + 366, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 567, + 366, + 580 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 567, + 366, + 580 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i" + }, + { + "bbox": [ + 131, + 567, + 366, + 580 + ], + "type": "text", + "content": " denotes the pseudo-label of the " + }, + { + "bbox": [ + 131, + 567, + 366, + 580 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 567, + 366, + 580 + ], + "type": "text", + "content": "-th sample." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 597, + 298, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 298, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 298, + 609 + ], + "type": "text", + "content": "4.2 Comparison with Baselines" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "content": "We compared the clustering performance of our proposed method, DEAN, with SoTA approaches, as shown in Table 1. Since we considered an online scenario for category discovery, traditionally approached using offline methods, there is no previous work for comparison. Therefore, for a fair comparison, our study adapts" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 447, + 103 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 157, + 482, + 271 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Clustering performance comparison on CUB200, FGVC-Aircraft, and CIFAR-100. Best performance is highlighted in bold. All the baselines utilize the fine-split method from CGCD to handle the absence of prior knowledge." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 157, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 137, + 157, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 137, + 157, + 482, + 271 + ], + "type": "table", + "html": "
MethodCUB200FGVC-AircraftCIFAR-100
MallMoldMnewF↓MallMoldMnewF↓MallMoldMnewF↓
Supervised72.5170.6879.6611.9166.2563.6276.7611.0381.3479.4888.8010.07
Fine-Tuning1.961.942.0180.522.061.305.1073.353.683.245.4586.26
FRoST [ECCV' 22]23.9828.725.4550.8514.4415.659.6048.316.136.215.8069.24
CGCD[ICCV' 23]40.4047.8811.1835.6114.0815.677.7560.002.642.762.1586.74
DEAN w/ FS55.5765.3717.2917.6240.3249.145.1025.4352.2762.1112.9027.55
DEAN (Ours)59.5666.3333.1116.2646.4453.5617.9920.9762.3470.3430.3519.17
", + "image_path": "73b12beb49ff9c948f0eaa6dd4a5433f54997083847e736c161cdadfeabc3f7f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 137, + 327, + 309, + 426 + ], + "blocks": [ + { + "bbox": [ + 132, + 282, + 309, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 282, + 309, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 309, + 316 + ], + "type": "text", + "content": "Table 2: Ablation experiment on CUB200. The baseline denotes fine-tuning with the fine split from the CGCD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 327, + 309, + 426 + ], + "lines": [ + { + "bbox": [ + 137, + 327, + 309, + 426 + ], + "spans": [ + { + "bbox": [ + 137, + 327, + 309, + 426 + ], + "type": "table", + "html": "
AblationCUB200
MallMoldNewF↓
Baseline1.961.942.0180.52
+ LoRA49.3360.356.2422.24
+ EGD54.9762.9823.6719.61
+ EC Loss57.3764.9827.6317.61
+ VFA (Ours)59.5666.3333.1116.26
", + "image_path": "6ddf3724fd900dd449ccd851e0ddc9b3cf773ae6db00c4d982f75ac6d015f9da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 325, + 482, + 427 + ], + "blocks": [ + { + "bbox": [ + 322, + 282, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 322, + 282, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 322, + 282, + 482, + 316 + ], + "type": "text", + "content": "Table 3: Ablation study for the hyperparameter K of variance-based feature augmentation on CUB200." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 325, + 482, + 427 + ], + "lines": [ + { + "bbox": [ + 328, + 325, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 328, + 325, + 482, + 427 + ], + "type": "table", + "html": "
AblationCUB200
\\(M_{all}^{PS}\\)\\(M_{old}^{PS}\\)\\(M_{new}^{PS}\\)
K=057.3185.1946.47
K=164.2788.9653.74
K=375.8793.4364.48
K=576.9394.0565.75
K=775.4493.4164.07
K=975.2293.563.36
", + "image_path": "6dd8acbd88e15d10aade891c7d282c91859dcaff3f55c92a1a14b9d05a3c7e3e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "content": "offline continuous NCD and GCD methods into OCGCD scenarios. The experiment includes two SoTA continuous category discovery methods: FRoST [37] and CGCD [22]. In addition, we include a supervised learning method as an upper bound and a fine-tuning method as a performance baseline. Note that in our proposed scenario, OCGCD, all baselines are not provided with prior knowledge about novel categories. Instead, we employed fine-split method from CGCD [22] which consists of two stages: initial split and fine split. During the initial split stage, unlabeled data is roughly divided into known and unknown categories based on cosine similarity. In the fine split stage, a split model is trained to distinguish between known and unknown data. The fine-split method estimates the number of novel categories conveying information about novel categories to the comparison methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The results show that our method, DEAN, recorded a robust clustering performance, significantly surpassing existing continuous NCD and GCD methods. DEAN recorded consistently high metrics across " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{all}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{old}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{new}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " while keeping forgetting metric " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " substantially low. This denotes our method can learn novel knowledge without forgetting. Moreover, it is noteworthy that DEAN w/ FS which denotes the adoption of fine-split from CGCD into DEAN instead" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 147, + 299, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 301, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 301, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 301, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study for the cluster accuracy of pseudo-labels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 147, + 299, + 224 + ], + "lines": [ + { + "bbox": [ + 137, + 147, + 299, + 224 + ], + "spans": [ + { + "bbox": [ + 137, + 147, + 299, + 224 + ], + "type": "table", + "html": "
MethodCUB200
\\( M_{all}^{PS} \\)\\( M_{old}^{PS} \\)\\( M_{new}^{PS} \\)
CGCD [ICCV' 23]61.8881.0758.68
DEAN w/o (\\( \\mathcal{L}_{ec} \\), VFA)57.2385.1145.80
DEAN w/o VFA57.3185.1946.47
DEAN (Ours)76.9394.0565.75
", + "image_path": "9c72b61ec52edae09619d959e4b0bdcea9ebe7876725793e68b22e7f36b8a89e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 316, + 146, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 5: Ablation experiment of parameter-efficient tuning on CUB200." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 146, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 316, + 146, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 146, + 482, + 224 + ], + "type": "table", + "html": "
MethodCUB200
\\(M_{all}\\)\\(M_{old}\\)\\(M_{new}\\)\\(F\\downarrow\\)
DEAN(Ours)Adapter [4]56.9664.7526.5317.64
Prefix [27]58.5466.2928.2816.30
LoRA [17]59.5666.3333.1116.26
", + "image_path": "3489abb60324c93ef2eda66b2f7d1c357b747f8af2cf28c2dc23a39204f8896b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": "of energy-guided discovery (EGD) also showed notable performance compared to the baselines. But, DEAN achieved significant performance gains across all datasets as the proposed EGD is applied. In contrast, existing continual methods showed shortcomings in clustering performance. Specifically, we observed extremely low clustering performance on CIFAR-100 dataset, low " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "M_{all}" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "M_{old}" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "M_{new}" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": " values, coupled with high " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ". On CIFAR-100 dataset, DEAN surpassed the most recent method CGCD assuming no prior knowledge for novel categories by " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "59.7\\%" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "67.58\\%" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "28.2\\%" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "67.57\\%" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "M_{all}" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "M_{old}" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "M_{new}" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 247, + 481, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 415 + ], + "type": "text", + "content": "Such observations show that while current GCD methods are vulnerable to online settings, DEAN can manage the balance required in the OCGCD scheme of retaining old knowledge while integrating the new. We believe this results shows the suitability of DEAN for real-world applications, where data is never static. We empirically confirmed that DEAN can evolve with the data it encounters, overcoming the challenges presented in the OCGCD scheme." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 430, + 238, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 238, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 238, + 443 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 485 + ], + "type": "text", + "content": "The experimental results in Section 4.2 show that our DEAN achieved significant performance in OCGCD scenarios. Through further experiments on the CUB200, we provide a detailed analysis of the components proposed in our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of Components. We validate the effectiveness of adding each component of our method to the baseline model. The baseline denotes fine-tuning the backbone model with a data-split following settings in CGCD. As shown in Table 2, results reveal that all components have an important impact on performance. Employing LoRA to baseline shows significant improvement in " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{old}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". This shows that our application of LoRA on the frozen model at incremental session effectively mitigates forgetting. While LoRA shows only a slight increase in " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{new}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", additionally employing energy-guided discovery (EGD) and EC Loss markedly improves " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{new}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". This confirms that adopting energy guidance will enhance the ability of the model to differentiate between old and new categories. Finally, the integration of VFA further boosts the model performance on " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{new}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". This improvement reveals the importance of accurate pseudo-labeling since VFA is a technique applied after training. Each component is crucial, incrementally building upon the last to solidify the model's performance in our online category discovery scenario." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "Hyperparameter Analysis of VFA. Variance-based Feature Augmentation (VFA) is our proposed method of temporarily augmenting feature vectors to acquire accurate pseudo-labeling. Our analysis evaluated the effect of hyperparameter " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ", the number of augmented feature vectors, on the accuracy of pseudolabels. From Table 3 we observed that employment of VFA is beneficial, increasing novel category pseudo-label accuracy from " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "46.47\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "65.75\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ". Among VFA employment, we observed that " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " yields the best performance across all metrics. When " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " is less than 5, the smaller " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " worsened performance. Conversely, values of " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " greater than 5 did not benefit the model. We analyzed that since VFA augments features based on variance, as the number of " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " increases, unseen samples can be clustered together without being distinguished." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": "Pseudo-Labeling Accuracy. Accurate pseudo-labeling for clustering is important because it directly influences the performance of the model in incremental learning. Shown in Table 4, our proposed method, DEAN, significantly enhanced the clustering accuracy of pseudo-labels, achieving a " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "13.12\\%" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": " improvement in " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "M_{all}^{PS}" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": " over the most recent method CGCD. Through the ablation study, we investigated the contributions of each step in DEAN. When DEAN adopts only parameter-efficient tuning, excluding EC loss and VFA, we observed a notable declination in " + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "inline_equation", + "content": "M_{new}^{PS}" + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": ". While each ablation step showed gradually accuracy enhancement, the application of VFA elevates the performance of DEAN above CGCD. Thus, the experimental result revealed that VFA played a vital role in distinguishing novel categories and confirming the effectiveness of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 499 + ], + "type": "text", + "content": "Parameter Efficient Tuning. The ablation study in Table 5 assesses the impact of parameter-efficient tuning (PET) methods on the CUB200 dataset. Our method can employ various tuning methods such as Adapter [4] or Prefix-tuning [27], and in this experiment, we compared modules with LoRA [17], Adapter and Prefix-tuning. We set the size of all PET modules size to 5 and integrated them into the last 5 layers of ViT. Overall improvements are shown by adopting PET over existing methods reported in Table 1. Among PET methods, LoRA showed the highest clustering accuracy. The performance enhancement of LoRA may stem from its fast adaptation, which is highly effective in online continual learning scenarios that often depend on fewer samples for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 519, + 220, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 220, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 220, + 533 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we introduced a novel scenario for category discovery, Online Continuous Generalized Category Discovery, which accounts for the dynamic nature of real-world data streams, leading to severe forgetting and poor novel category discovery. In addition, we proposed a novel framework, DEAN, which utilizes energy-based novel category discovery for the first time in this field and introduces variance-based feature augmentation to enhance accurate pseudo-labeling. Our DEAN demonstrated promising performance in the proposed novel scenario. However, DEAN could not guarantee inference performance midway through training. In future work, we plan to explore an online continual framework that can ensure inference performance during training through fast adaptation." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 240 + ], + "type": "text", + "content": "This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University))." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 259, + 198, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 198, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 198, + 272 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 286, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 136, + 286, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 286, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 136, + 286, + 481, + 320 + ], + "type": "text", + "content": "1. Aljundi, R., Belilovsky, E., Tuytelaars, T., Charlin, L., Caccia, M., Lin, M., Page-Caccia, L.: Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 321, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 321, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 136, + 321, + 481, + 364 + ], + "type": "text", + "content": "2. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: International conference on machine learning. pp. 233-242. PMLR (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 365, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 365, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 136, + 365, + 481, + 398 + ], + "type": "text", + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 399, + 481, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 399, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 136, + 399, + 481, + 431 + ], + "type": "text", + "content": "4. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 432, + 481, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 432, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 136, + 432, + 481, + 465 + ], + "type": "text", + "content": "5. Choe, S.A., Shin, A.H., Park, K.H., Choi, J., Park, G.M.: Open-set domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23943-23953 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 466, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 466, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 136, + 466, + 481, + 509 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 136, + 510, + 481, + 542 + ], + "type": "text", + "content": "7. Fini, E., Sangineto, E., Lathuiliere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284-9292 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 544, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 544, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 136, + 544, + 481, + 565 + ], + "type": "text", + "content": "8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences 3(4), 128-135 (1999)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 136, + 566, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 566, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 136, + 566, + 481, + 587 + ], + "type": "text", + "content": "9. Frey, B.J., Dueck, D.: Clustering by passing messages between data points. science 315(5814), 972-976 (2007)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 481, + 620 + ], + "type": "text", + "content": "10. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: International Conference on Machine Learning. pp. 8109-8126. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "text", + "content": "11. Gutmann, M., Hyvärinen, A.: Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 297-304. JMLR Workshop and Conference Proceedings (2010)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "12. Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Automatically discovering and learning new visual categories with ranking statistics. arXiv preprint arXiv:2002.05714 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "13. Han, K., Vedaldi, A., Zisserman, A.: Learning to discover novel visual categories via deep transfer clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8401-8409 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "14. He, J., Mao, R., Shao, Z., Zhu, F.: Incremental learning in online scenario. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13926-13935 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "15. Hsu, Y.C., Lv, Z., Kira, Z.: Learning to cluster in order to transfer across domains and tasks. arXiv preprint arXiv:1711.10125 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "type": "text", + "content": "16. Hsu, Y.C., Lv, Z., Schlosser, J., Odom, P., Kira, Z.: Multi-class classification without multi-class labels. arXiv preprint arXiv:1901.00544 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "text", + "content": "17. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 326 + ], + "type": "text", + "content": "18. Jia, X., Han, K., Zhu, Y., Green, B.: Joint representation learning and novel category discovery on single-and multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 610-619 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 480, + 358 + ], + "type": "text", + "content": "19. Joseph, K., Paul, S., Aggarwal, G., Biswas, S., Rai, P., Han, K., Balasubramanian, V.N.: Novel class discovery without forgetting. In: European Conference on Computer Vision. pp. 570-586. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 480, + 392 + ], + "type": "text", + "content": "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "text", + "content": "21. Kim, D., Ryoo, K., Cho, H., Kim, S.: Splitnet: learnable clean-noisy label splitting for learning with noisy labels. arXiv preprint arXiv:2211.11753 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 413, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 413, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 413, + 480, + 456 + ], + "type": "text", + "content": "22. Kim, H., Suh, S., Kim, D., Jeong, D., Cho, H., Kim, J.: Proxy anchor-based unsupervised learning for continuous generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16688-16697 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "type": "text", + "content": "23. Koh, H., Kim, D., Ha, J.W., Choi, J.: Online continual learning on class incremental blurry task configuration with anytime inference. arXiv preprint arXiv:2110.10031 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 490, + 480, + 511 + ], + "type": "text", + "content": "24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "type": "text", + "content": "25. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "type": "text", + "content": "26. Li, J., Socher, R., Hoi, S.C.: Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 555, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 480, + 600 + ], + "type": "text", + "content": "27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "type": "text", + "content": "28. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "text", + "content": "29. Liu, Y., Tuytelaars, T.: Residual tuning: Toward novel category discovery without labels. IEEE Transactions on Neural Networks and Learning Systems (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "text", + "content": "30. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "K.-H. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 661 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "31. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 181 + ], + "type": "text", + "content": "32. McClelland, J.L., McNaughton, B.L., O'Reilly, R.C.: Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review 102(3), 419 (1995)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 213 + ], + "type": "text", + "content": "33. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 482, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 255 + ], + "type": "text", + "content": "34. Moon, J.Y., Park, K.H., Kim, J.U., Park, G.M.: Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11731-11741 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 255, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 289 + ], + "type": "text", + "content": "35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 319 + ], + "type": "text", + "content": "36. Park, K.H., Song, K., Park, G.M.: Pre-trained vision and language transformers are few-shot incremental learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23881-23890 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 482, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 341 + ], + "type": "text", + "content": "37. Roy, S., Liu, M., Zhong, Z., Sebe, N., Ricci, E.: Class-incremental novel class discovery. In: European Conference on Computer Vision. pp. 317-333. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 341, + 482, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 373 + ], + "type": "text", + "content": "38. Seo, J., Lee, S.H., Lee, T.Y., Moon, S., Park, G.M.: Generative unlearning for any identity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9151-9161 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 373, + 482, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 405 + ], + "type": "text", + "content": "39. Troisemaine, C., Lemaire, V., Gosselin, S., Reiffers-Masson, A., Flocon-Cholet, J., Vaton, S.: Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 405, + 482, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 436 + ], + "type": "text", + "content": "40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492-7501 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 436, + 482, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 457 + ], + "type": "text", + "content": "41. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 457, + 482, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 489 + ], + "type": "text", + "content": "42. Wu, Y., Chi, Z., Wang, Y., Feng, S.: Metagcd: Learning to continually learn in generalized category discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1655-1665 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 489, + 482, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 521 + ], + "type": "text", + "content": "43. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 521, + 482, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 553 + ], + "type": "text", + "content": "44. Zhang, X., Jiang, J., Feng, Y., Wu, Z.F., Zhao, X., Wan, H., Tang, M., Jin, R., Gao, Y.: Grow and merge: A unified framework for continuous categories discovery. Advances in Neural Information Processing Systems 35, 27455-27468 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 553, + 482, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 585 + ], + "type": "text", + "content": "45. Zhao, B., Han, K.: Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. Advances in Neural Information Processing Systems 34, 22982-22994 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 585, + 482, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 482, + 616 + ], + "type": "text", + "content": "46. Zhong, Z., Fini, E., Roy, S., Luo, Z., Ricci, E., Sebe, N.: Neighborhood contrastive learning for novel class discovery. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10867-10875 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 616, + 482, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 661 + ], + "type": "text", + "content": "47. Zhong, Z., Zhu, L., Luo, Z., Li, S., Yang, Y., Sebe, N.: Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9462-9470 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Continuous Generalized Category Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_content_list.json b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..24f74886575598d9fe6182402cac2f52eecd3ad2 --- /dev/null +++ b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "text_level": 1, + "bbox": [ + 261, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngkil Song\\*", + "bbox": [ + 233, + 212, + 362, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongkeun Kim*", + "bbox": [ + 388, + 213, + 519, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minsu Cho", + "bbox": [ + 547, + 213, + 642, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suha Kwak", + "bbox": [ + 671, + 213, + 767, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pohang University of Science and Technology (POSTECH), South Korea {songyk, kdk1563, mscho, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/MATR/", + "bbox": [ + 258, + 239, + 745, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Online temporal action localization (On-TAL) is the task of identifying multiple action instances given a streaming video. Since existing methods take as input only a video segment of fixed size per iteration, they are limited in considering long-term context and require tuning the segment size carefully. To overcome these limitations, we propose memory-augmented transformer (MATR). MATR utilizes the memory queue that selectively preserves the past segment features, allowing to leverage long-term context for inference. We also propose a novel action localization method that observes the current input segment to predict the end time of the ongoing action and accesses the memory queue to estimate the start time of the action. Our method outperformed existing methods on two datasets, THUMOS14 and MUSES, surpassing not only TAL methods in the online setting but also some offline TAL methods.", + "bbox": [ + 261, + 320, + 743, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Temporal action localization $\\cdot$ Online video understanding", + "bbox": [ + 261, + 513, + 740, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 556, + 375, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Today, video is one of the most popular types of media, with tons of videos being created and uploaded to online platforms like YouTube and TikTok every second. As a result, the demand for video understanding is growing rapidly. In particular, dealing with untrimmed videos is a key to practical video understanding since most videos in the real world are not trimmed for individual events in advance. Temporal action localization (TAL) has been extensively studied in this context [10, 17, 21, 30, 38, 39, 43-45, 47, 49], with the aim of detecting each action instance in the input untrimmed video by predicting its start time, end time, and action class.", + "bbox": [ + 212, + 588, + 787, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, online TAL (On-TAL), i.e., TAL for streaming video [14], has gained increasing attention thanks to its potential applications in video surveillance, sports video analysis, and video summarization. Unlike TAL which demands the entire video as input, On-TAL aims to detect action instances using only input frames observed so far in an online manner. Moreover, once action instances are predicted, it is not allowed to modify the results afterward.", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 217, + 824, + 364, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da375160703c66cee8a846f459f1e75e56a5aa3c7ec8a63d40229d01134fd1d1.jpg", + "image_caption": [ + "Fig. 1: A conceptual illustration of our model. Our model selectively stores information about previously occurred actions in the memory queue from streaming video. When the end of the action is detected in the current input segment, the model retrieves information from the memory queue to locate the start of the action." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A straightforward solution to this challenging task is to first conduct online action detection (OAD) [7,35], i.e., frame-by-frame action classification for streaming video, and then build action instances by aggregating the frame-wise action predictions. In this sense, earlier methods for On-TAL [14,31] follow this approach by applying OAD and refining per-frame predictions using past observations. However, this OAD-based approach exploits only frame-level supervision for training, which is not optimal because On-TAL aims to predict action instances in the form of time intervals. To alleviate this issue, online anchor transformer (OAT) has been recently proposed [15], which adopts sliding window approach and anchor-based method to utilize instance-level supervision. Although OAT improved performance remarkably, however, it still has several drawbacks: Since it takes a fixed-size segment of input video at each iteration and does not consider long-term contexts, its capability is limited in detecting long action instances, and its performance is sensitive to hyperparameters like the size of the input segment.", + "bbox": [ + 212, + 412, + 787, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To resolve the above issues, we propose a new end-to-end architecture for OntAL, dubbed memory-augmented Transformer (MATR). At the heart of MATR lies in memory queue, which selectively stores past segment features so that the model exploits long-term context for inference. Thanks to the memory queue that contains long-term context and reduces dependence on the segment size, our model accurately localizes even long-term action instances without the need for careful hyperparameter tuning for each dataset.", + "bbox": [ + 212, + 641, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we propose a novel approach to action instance localization using the memory queue as illustrated in Fig. 1. First, it detects the end of an action using the current segment features and then scans past segment features in the memory queue to find the action start corresponding to the detected action end. To achieve this, we adopt two Transformer decoders: one for end detection and the other for start detection. Specifically, we adopt learnable queries as input to", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the decoders so that each of them learns to localize action boundaries through the attention mechanism of Transformer [8,32]. In addition, motivated by recent advances on object detection [6,37], MATR decouples action classification and localization to further improve the performance. This is realized in MATR by using distinct queries for the two subtasks.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed method is evaluated on two datasets, THUMOS14 [13] and MUSES [22]. Our model achieves the state of the art on both datasets in the online TAL setting. In addition, surprisingly, it is also comparable to existing offline TAL methods although it does not employ offline non-maximum suppression (NMS). In summary, the contribution of this paper is three-fold:", + "bbox": [ + 212, + 222, + 782, + 297 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose MATR, a new end-to-end architecture for On-TAL, leveraging memory queue to utilize long-term context and enabling precise localization of action instances with less dependence on dataset-specific hyperparameter.", + "- We introduce a new action instance localization method that first identifies action end using the input segments and scans past information in the memory queue to find the action start. Moreover, our method adopts distinct queries to separate information for action classification and localization.", + "- Our method outperformed the existing On-TAL methods on the two benchmarks. Also, extensive ablation studies demonstrate the contribution of each component of the proposed method." + ], + "bbox": [ + 225, + 311, + 784, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 488, + 385, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Action Localization. Temporal action localization (TAL) [5,10,17, 21,23,30,38,39,43-45,47,49] which identifies action instances in time and classifying their actions, has given much attention on the video research community. Earlier attempts employ two-stage approaches that first generate action proposals, then classify action class and refine action boundaries [1,9,12,18,19,24,47]. Recent studies have primarily dealt with widening the range of video context through convolutional neural networks [28], graph neural networks [39,43], Gaussian kernel [25], and utilizing global context [49]. To extend temporal context and enable end-to-end TAL, Transformer [8,32] has recently been adopted [6,23,44]. These approaches incorporate temporal deformable attention module [23] and multi-scale features [44] to improve computational efficiency and performance.", + "bbox": [ + 212, + 522, + 782, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online Video Understanding. Online action detection (OAD), which classifies action classes in a frame-wise manner, is first proposed by Geest et al. [7]. To incorporate long-term context, recurrent neural network [40] and Transformer [35] are utilized. These methods predict actions for future frames and use this information to refine current frame predictions. Yang et al. [42] further improve the performance by using category exemplars and capturing long-term dependencies within a video segment. On the other hand, Shou et al. [29] proposes online detection of action start (ODAS), which aims to detect action start as early as possible. Gao et al. [11] further improves ODAS by adopting reinforcement learning.", + "bbox": [ + 212, + 688, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online Temporal Action Localization. Towards practical event understanding, online temporal action localization (On-TAL) has been proposed recently. Kang et al. [14] first propose On-TAL, which localizes action boundaries and classifies action classes without accessing future frames and offline post-processing. Earlier methods [14,31] adopt OAD framework to classify per-frame action class and group them into action instances to solve On-TAL. Specifically, Kang et al. [14] utilize Markov decision process to aware prior contexts to group per-frame predictions. SimOn [31] is proposed as an end-to-end On-TAL method that leverages past visual context and action probabilities for accurate action prediction at the current frame. Kim et al. [15] propose a sliding window scheme and an anchor-based method that adopts Transformer [32] for On-TAL. However, it only uses a fixed input segment which is highly dependent on the action length distribution of each dataset. Unlike the previous methods, we utilize a memory queue for long-term context, which also reduces the dependence on the input segment size.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Difference between OAD and On-TAL. OAD [2,33,35,40-42,46] focus on predicting action classes at the frame-level, which limits the ability to distinguish boundaries of overlapping actions. On the other hand, On-TAL [14,15,31] predicts entire action instances rather than individual frames, enabling it to precisely identify the start and end times of actions, making it effective for applications requiring instance-level action understanding, like sports video analysis. Using Memory for Video Understanding. Prior work in video understanding employs memory to store contextual information, in the forms of long-term and short-term memory [2,33,36,41]: short-term memory captures fine details, and long-term memory stores compressed information of past contexts. Specifically, Wang et al. [33] divides long-term memory into several groups and compresses each group into a single vector. Stream Buffer [2] gradually reduces temporal dimensions of the memory using a compression module. However, unlike video classification tasks such as action recognition [36] and OAD [2,33,41,46], which requires to store representative information of past frames in the memory for classifying the action of the current frame, TAL needs to preserve the temporal positional information of past frames in the memory to accurately predict the exact time. In that sense, MATR preserves temporal information in memory to accurately localize actions and effectively stores information by using flag tokens. A comparison with the OAD memory modules is shown in Table 4.", + "bbox": [ + 212, + 373, + 787, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 700, + 426, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider an untrimmed video $\\mathbf{V} = \\{\\mathbf{v}_i\\}_{i=1}^T$ with $T$ frames and $M$ action instances $\\varPsi=\\{(s_m,e_m,c_m)\\}_{m=1}^{M}$ , where each instance is represented by its start time $s_m$ , end time $e_m$ and action class $c_m$ . Unlike temporal action localization (TAL), which allows a model to use whole video sequence as input, online temporal action localization (On-TAL) forces a model to predict action instances using only input frames seen until the current timestamp. Note that action instances predicted at previous iterations cannot be modified or deleted afterwards.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 113, + 673, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a6d7ed0afe60bf9567f5d531cf41a14e4f748475a10ec1f6ed7e1cbb0e37d60.jpg", + "image_caption": [ + "Fig. 2: Overall architecture of MATR. MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 784, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads. In line with previous research, the current segment of the input streaming video is given as a unit input, and its frame-level features, referred to as segment features, are extracted by a video backbone network and a linear projection layer (Sec. 3.1). The segment features are then fed to the memory-augmented video encoder, which encodes temporal context between frames in the current segment and stores the segment features into the memory (Sec. 3.2). The instance decoding module localizes action instances via two Transformer decoders: the end decoder and the start decoder. Specifically, the end decoder references the encoded segment features to locate the action end around the current time, and then the start decoder refers to the memory queue to find the action start based on the past information stored in the memory queue. Queries for each instance consist of a class query for action classification and a boundary query for action localization (Sec. 3.3). The outputs of the instance decoding module are used as inputs to the prediction heads, which consist of end prediction head, start prediction head, and action classification head (Sec. 3.4). The entire model is trained in an end-to-end manner (Sec. 3.5).", + "bbox": [ + 212, + 381, + 787, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Feature Extractor", + "text_level": 1, + "bbox": [ + 215, + 700, + 410, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a segment comprising $L_{s}$ consecutive video frames as input, a pretrained backbone [4, 34] followed by a linear projection layer extracts the segment features $\\mathbf{X}_t = \\{\\mathbf{x}_i\\}_{t - L_s + 1}^t\\in \\mathbb{R}^{L_s\\times D}$ for each frame within the segment, following previous work [14, 15]. Inspired by Kim et al. [15], MATR employs a sliding window scheme, which predicts multiple action instances by moving frame by frame along the temporal axis. Note that the window size is the input segment size.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Memory-Augmented Video Encoder", + "text_level": 1, + "bbox": [ + 215, + 146, + 560, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At each time step, the input segment features are fed to two modules of the memory-augmented video encoder, the segment encoder and the memory update module. The segment encoder encodes the temporal context across input segment features, while the memory update module selectively stores the segment features and updates the memory queue.", + "bbox": [ + 212, + 172, + 785, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segment Encoder. The segment encoder is a standard Transformer encoder composed of self-attention layers and a feed forward network (FFN). The combination of segment features and a learnable flag token is fed into the encoder, and transformed to queries, keys, and values for self-attention. Sinusoidal positional encoding $\\mathbf{S}_{\\mathrm{pos}} \\in \\mathbb{R}^{(L_s + 1) \\times D}$ is added to the queries and keys. Among the output embeddings, the encoded flag token is fed into the memory update module and the encoded segment features that integrate temporal context across input frames are used as input to the end decoder.", + "bbox": [ + 212, + 247, + 787, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Memory Update Module. In an online setting, where an input is a streaming video, efficiently storing information of past frames and accessing it effectively is crucial for detecting action instances, in particular those of long-term actions whose temporal extents are beyond the size of input segments. To this end, we adopt the memory queue for storing information of past input segments in a first-in-first-out (FIFO) manner. Additionally, we propose an efficient memory update method employing the flag token. When the input segment features pass through the segment encoder, the flag token is concatenated into the input segment features and then employed in the flag prediction head. It is trained to predict [FLAG], which identifies whether an input segment is relevant to action instances, so that only relevant segments are stored in the memory. [FLAG] = 1 when there are overlapping frames between the input segment and action instances, and 0 otherwise. During training, the ground-truth [FLAG] is utilized. During inference, [FLAG] = 1 if sigmoid(\\hat{g}) > \\theta where $\\hat{g}$ is the output logit of the flag prediction head and $\\theta$ is a predefined threshold, and 0 otherwise. The input segment feature is added to the memory queue when [FLAG] = 1 and discards it when [FLAG] = 0. If the memory queue is full, the oldest segment feature is purged.", + "bbox": [ + 212, + 368, + 787, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Instance Decoding Module", + "text_level": 1, + "bbox": [ + 215, + 648, + 482, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The instance decoding module localizes and classifies action instances by leveraging the encoded segment features and the memory queue through the attention mechanism of Transformer. Given the encoded segment features that have short-term temporal contexts across input segment and the memory queue that stores long-term contexts for actions that are potentially ongoing, a set of $2N$ instance queries $\\mathbf{Q} = \\{\\mathbf{Q}_{\\mathrm{class}}; \\mathbf{Q}_{\\mathrm{bound}}\\} \\in \\mathbb{R}^{2N \\times D}$ is trained to generate $N$ action instances for each input segment. Half of these queries pertain to predicting the action class (class query; $\\mathbf{Q}_{\\mathrm{class}} \\in \\mathbb{R}^{N \\times D}$ ), while the other half is used for predicting the start and end timestamps (boundary query; $\\mathbf{Q}_{\\mathrm{bound}} \\in \\mathbb{R}^{N \\times D}$ ). The class and boundary query pairs for the same instance share the same positional embedding $\\mathbf{E}_{\\mathrm{pos}} \\in \\mathbb{R}^{N \\times D}$ to accurately identify and differentiate between", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57feb6d1ef320ad3bda0e3452736157218e717846a8a0095b857edcbe7be4792.jpg", + "image_caption": [ + "Fig. 3: Detailed architecture of the instance decoding module." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 784, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "instances; $\\tilde{\\mathbf{Q}} = \\{\\mathbf{Q}_{\\mathrm{class}} + \\mathbf{E}_{\\mathrm{pos}};\\mathbf{Q}_{\\mathrm{bound}} + \\mathbf{E}_{\\mathrm{pos}}\\}$ . Unlike the previous Transformer-based On-TAL method [15] that predicts start and end offset for each instance simultaneously, our method employs separate Transformer decoders for predicting start offset and end offset. As shown in Fig. 3, the start decoder and end decoders share the same architecture but utilizes different information to predict action start and action end, respectively.", + "bbox": [ + 212, + 362, + 784, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "End Decoder. Given the encoded segment features, instance queries $\\mathbf{Q}$ is trained to generate the output embeddings used for detecting action end near the current timestamp. To this end, we adopt a Transformer [3] composed of multi-head self-attention layers, multi-head cross-attention layers, and FFNs. The detailed architecture is illustrated in Fig. 3. The outputs of the end decoder, which consist of end boundary embeddings and class embeddings, are then used as inputs to the start decoder and the prediction heads.", + "bbox": [ + 212, + 455, + 785, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Start Decoder. The start decoder takes the output embeddings from the end decoder and locates the corresponding action start by utilizing the memory queue. First, the memory queue is concatenated with the current segment features. The concatenated memory features are used as long-term context in the start decoder. Then, 2D temporal positional encodings are added to the memory features, and the results are used as the key and value for the cross-attention layers of the start decoder. The overall architecture of the start decoder is the same as the end decoder (Fig. 3), but it uses the memory features rather than the encoded segment features. The output embeddings of the start decoder, start boundary embeddings and class embeddings, are used as inputs to the start prediction and the action classification head, respectively.", + "bbox": [ + 212, + 565, + 785, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, we introduce two techniques to efficiently and effectively utilize the memory queue. First, we conduct a $50\\%$ uniform sampling to the memory features which enables efficient use of memory since adjacent frames contain similar information. Second, to expand the scope of positional encoding for streaming video, whose duration is unpredictable, we separate the temporal positional encoding into the two parts: relative segment position and relative frame position. A relative segment position is the position of a segment within the memory rela", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tive to the current segment. Meanwhile, a relative frame position is the position of a frame relative to the most recent frame within the same segment.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Prediction Heads", + "text_level": 1, + "bbox": [ + 214, + 198, + 405, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To generate $N$ action instances $\\{(\\hat{s}_i,\\hat{e}_i,\\hat{c}_i)\\}_{i = 1}^N$ from the output of the instance decoding module, we use three prediction heads, which are end prediction head, start prediction head, and action classification head. Each of the prediction heads is composed of a 2-layer FFN.", + "bbox": [ + 212, + 220, + 782, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "End Prediction Head. The end prediction head estimates the offset between the end of the target action and the current timestamp. The boundary embeddings from the end decoder are fed to the 2-layer FFN, which in turn estimates the end offset $\\{\\hat{u}_i\\}_{i=1}^N \\in \\mathbb{R}^N$ .", + "bbox": [ + 212, + 282, + 784, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Start Prediction Head. The start prediction head estimates the offset between the start of the target action and the current timestamp. Unlike the end prediction, which finds the end offset around the input segment, the start prediction needs to perform offset regression over a relatively wider range. To narrow down the scope of the start offset regression, we divide the time horizon into $L_{m} + 2$ regions: the region preceding the coverage of memory, $L_{m}$ regions covered by the memory, and the region corresponding to the current input segment. In that sense, the start prediction head consists of the region classification and the offset regression, each composed of a 2-layer FFN. They take the boundary embeddings from the start decoder as input to predict the start time. The region classification identifies the regions to which the start time is assigned, represented as $\\{\\hat{o}_i|\\hat{o}_i = \\mathrm{argmax}(\\hat{r}_i)\\}_{i=1}^N \\in \\mathbb{R}^N$ . Here, $\\{\\hat{r}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (L_m + 2)}$ are the output logits from the region classification head. Following this, the offset regression head is utilized to estimate the offset within the regions, denoted as $\\{\\hat{v}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (L_m + 2)}$ . Note that the start offset $\\hat{v}_i$ is calculated for all $L_{m} + 2$ regions, and the offset from the identified region $\\{\\hat{v}_i(\\hat{o}_i)\\}_{i=1}^N \\in \\mathbb{R}^N$ is used in inference. By combining start classification result $\\hat{o}_i$ and start regression result $\\hat{v}_i(\\hat{o}_i)$ , the start time is predicted. The effectiveness of the proposed start prediction head is validated in Table 5.", + "bbox": [ + 212, + 343, + 785, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Action Classification Head. For action class prediction, we utilize the class embeddings from both the end decoder and the start decoder. These embeddings are concatenated and fed into the action classification head to derive class probabilities $\\{\\hat{p}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (C+1)}$ , where $C$ is the number of action class.", + "bbox": [ + 212, + 628, + 782, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Action Instance Prediction. Finally, our model makes $N$ action proposals $\\{(\\hat{s}_i,\\hat{e}_i,\\hat{c}_i)\\}_{i = 1}^N$ at time $t$ by", + "bbox": [ + 214, + 690, + 782, + 720 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {i} = t - \\left(\\hat {o} _ {i} + \\hat {v} _ {i} \\left(\\hat {o} _ {i}\\right)\\right) \\times L _ {s}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 732, + 782, + 746 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {e} _ {i} = t + \\hat {u} _ {i} \\times L _ {s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 750, + 784, + 765 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} _ {i} = \\operatorname {a r g m a x} \\left(\\hat {p} _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 768, + 784, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $L_{s}$ is the segment length. Since the sliding window scheme generates $N$ action proposals at each timestamp, post-processing is crucial for performance improvement by removing redundant or overlapped action instances.", + "bbox": [ + 214, + 794, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Non-maximum suppression (NMS) is applied to the action proposals at each timestamp, subsequently removing those highly overlapped with the proposals generated from the past. To prevent more reliable prediction made in the future from being removed, instances where the predicted end time is beyond the current timestamp, $t < \\hat{e}_i$ , are also removed.", + "bbox": [ + 214, + 146, + 787, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training Objective", + "text_level": 1, + "bbox": [ + 215, + 241, + 418, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "At timestamp $t$ , the model is trained to detect action instances whose end times are in the range $[t - T_d + 1, t + T_a]$ , where $T_d$ and $T_a$ are hyperparameters. Then, hungarian algorithm matches the action proposals and the ground-truth action instances with the lowest matching cost. The matching cost of the ground-truth group $i$ and the proposal $\\sigma(i)$ is given by:", + "bbox": [ + 214, + 263, + 782, + 339 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nC _ {i, \\sigma (i)} = \\hat {p} _ {\\sigma (i)} \\left(y _ {i}\\right) + \\operatorname {I o U} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 347, + 785, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\sigma$ is the permutation of $N$ action proposals, $\\hat{p}_i$ is the class probabilities of the $i$ -th proposal, $b_i = \\{s_i, e_i\\}$ is the ground-truth action boundary, and $\\hat{b}_i = \\{\\hat{s}_i, \\hat{e}_i\\}$ is the predicted action boundary. We adopt the focal loss [20] for the action classification, the cross entropy loss for the start region classification, the $\\ell_1$ loss for both the start offset regression and the end offset regression. The action classification loss $L_{\\mathrm{class}}$ , the start prediction loss $L_{\\mathrm{start}}$ , and the end prediction loss $L_{\\mathrm{end}}$ are defined as follow:", + "bbox": [ + 214, + 371, + 787, + 479 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c l a s s}} = \\sum_ {i = 1} ^ {N} \\operatorname {F L} \\left(\\hat {p} _ {i}, y _ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 487, + 785, + 526 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s t a r t}} = \\sum_ {i = 1} ^ {N _ {\\text {m a t c h}}} \\left\\{\\mathrm {C E} \\left(\\operatorname {S o f t m a x} \\left(\\hat {r} _ {i}\\right), r _ {i}\\right) + \\left| \\hat {v} _ {i} \\left(\\hat {o} _ {i}\\right) - v _ {i} \\right| \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 531, + 784, + 570 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {e n d}} = \\sum_ {i = 1} ^ {N _ {\\text {m a t c h}}} | \\hat {u} _ {i} - u _ {i} |, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 575, + 784, + 614 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $y_{i}, r_{i}, v_{i}, u_{i}$ is the ground-truth action class, start region, start offset, and end offset, respectively.", + "bbox": [ + 214, + 621, + 782, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To provide instance-level supervision and facilitate the connection between start and end time prediction, we employ the DIoU loss [48] following Action-Former [44]:", + "bbox": [ + 214, + 651, + 782, + 696 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {d i o u}} = \\sum_ {i = 1} ^ {N _ {\\mathrm {m a t c h}}} 1 - \\operatorname {I o U} (\\hat {b} _ {i}, b _ {i}) + \\frac {\\rho^ {2} (\\hat {c} _ {i} , c _ {i})}{d _ {i} ^ {2}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 698, + 784, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\rho(\\cdot, \\cdot)$ is the Euclidean distance between two points, $\\hat{c}_i$ and $c_i$ are the center of the proposal and ground-truth instance, and $d_i$ is the smallest enclosing 1-dim box length.", + "bbox": [ + 214, + 741, + 782, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To train the memory update module in an end-to-end manner, we employ the flag loss for training the flag prediction head:", + "bbox": [ + 214, + 786, + 782, + 816 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {f l a g}} = \\operatorname {B C E} (\\operatorname {S i g m o i d} (\\hat {g}), g), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 825, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{g}$ and $g$ is the predicted logit of flag token and ground-truth [FLAG] respectively.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our model is trained with six losses simultaneously in an end-to-end manner. The total loss is calculated as follows:", + "bbox": [ + 214, + 176, + 782, + 205 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {c l a s s}} + L _ {\\text {s t a r t}} + L _ {\\text {e n d}} + L _ {\\text {d i o u}} + L _ {\\text {f l a g}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 220, + 782, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Note that all loss coefficients are set to 1 and loss balancing is not required.", + "bbox": [ + 214, + 244, + 759, + 261 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 284, + 375, + 301 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 214, + 315, + 439, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method and previous methods on two On-TAL benchmarks: THUMOS14 [13] and MUSES [22]. THUMOS14 contains 200 videos for training and 213 videos for testing with 20 action classes, while MUSES has 2,587 videos for training and 1,110 videos for testing with 25 action classes. MUSES consists of multi-shot action instances, which makes action localization more challenging.", + "bbox": [ + 212, + 339, + 784, + 430 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hyperparameters. We use a frozen two-stream TSN [34] pretrained on Kinetics [4] to extract RGB and flow features for THUMOS14, and an I3D [4] trained on Kinetics [4] utilizing only RGB features for MUSES, following previous OntAL work [14, 15, 31]. The segment size $L_{s}$ is set to 64 in THUMOS14 and 75 in MUSES. The dimension of the segment feature is set to $D = 1024$ for both datasets. For the memory-augmented video encoder, we stack 3 Transformer layers with 8 attention heads for both datasets. The flag threshold $\\theta$ is set to 0.5. The memory size $L_{m}$ is set to 7 for THUMOS14, while 15 for MUSES. For the instance decoding module, we stack 5 Transformer layers with 4 attention heads for both datasets. The number of the class query and the boundary query $N = 10$ for THUMOS14 and $N = 6$ for MUSES.", + "bbox": [ + 212, + 431, + 784, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training. During training, we use Adam optimizer [16], with the initial learning rate of 1e-8 with CosineAnnealing scheduler [26]. The batch size is set to 64 for THUMOS14, while 75 for MUSES. The focal loss coefficient is $\\alpha = 0.25$ and $\\gamma = 2$ for THUMOS14 and $\\alpha = 1$ and $\\gamma = 5$ for MUSES. All loss coefficients are set to 1. More details are listed in the supplementary material (Sec. D).", + "bbox": [ + 214, + 597, + 784, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison with the State of the Art", + "text_level": 1, + "bbox": [ + 214, + 694, + 575, + 709 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our method with previous online and offline TAL methods on THUMOS14 and MUSES dataset. As presented in Table 1, our method outperforms all previous On-TAL methods in both benchmarks by a substantial margin: $4.9\\%$ of average mAP in THUMOS14 and $0.7\\%$ of average mAP in MUSES. The results show that MATR utilizing the memory queue is more effective than previous OAD-based methods [14,31,33,46] and instance-level method OAT-OSN [15]. We also evaluate two OAD methods utilizing memory [33,46] in On-TAL setting by grouping their frame-level predictions into action instances.", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9b7bb9d9f01b3b44740f66c677e85374f6eb323abb995cbef82664a93d698a56.jpg", + "table_caption": [ + "Table 1: Comparison with On-TAL methods on THUMOS14 and MUSES dataset. The results are reported in mAP measure (%). **' indicates that the result is validated using the On-TAL ground truth from classification annotation." + ], + "table_footnote": [], + "table_body": "
MethodTHUMOS14MUSES
backbone0.30.40.50.60.7Averagebackbone0.30.40.50.60.7
Offline TAL
G-TAD [39]TSN [34]54.547.640.230.823.439.9I3D [4]19.114.811.17.44.7
ContextLoc [49]I3D68.363.854.341.826.250.9------
P-GCN [43]I3D63.657.849.1---I3D19.917.113.19.75.4
MUSES [22]I3D68.964.056.946.331.053.4I3D25.922.618.915.010.6
ActionFormer [44]I3D82.177.871.059.443.966.8------
TriDet [27]I3D83.680.172.962.447.469.3------
OAD-based Online TAL
TeSTra [46]TSN35.629.221.413.47.621.4------
MAT [33]I3D36.027.519.212.36.420.3------
CAG-QIL [14]TSN44.737.629.821.914.529.7I3D8.56.54.22.81.9
SimOn [31]TSN54.345.035.023.314.634.4------
SimOn* [31]TSN57.047.537.326.616.036.9------
Instance-level Online TAL
OAT-Naive [15]TSN57.650.643.030.015.739.4I3D20.316.612.97.73.6
OAT-OSN [15]TSN63.056.747.136.320.044.6I3D22.118.514.28.94.7
MATRTSN70.362.752.138.623.749.5I3D23.519.314.39.45.7
", + "bbox": [ + 218, + 198, + 787, + 410 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "However, their performance is far lower than MATR, which verifies the effectiveness of the proposed memory queue of MATR. We also compare our model with several offline TAL methods. Although there is still a performance gap between online and offline methods, our model achieves comparable performance to offline methods, and even outperforms previous work [39, 43]. The performance on the MUSES dataset is lower compared to THUMOS14 and the reason is twofold: (1) Action instances in MUSES are captured in multi-shots, making it challenging to detect action boundaries, and (2) there are confusing action classes such as \"quarrel\" and \"fight,\" making action classification hard. These make MUSES challenging and limit the upper-bound of performance on it.", + "bbox": [ + 212, + 439, + 787, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 614, + 401, + 628 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We perform ablation studies on THUMOS14 and MUSES dataset to show the effectiveness of the proposed modules.", + "bbox": [ + 212, + 642, + 785, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation on the Proposed Modules. In Table 2, we conduct an ablation study to verify the effectiveness of each component of our model. In Table 2(a), focuses on the components of the memory-augmented video, both removing the flag token and the segment encoder led to a performance drop. The results show the efficacy of selectively storing the past information and the importance of temporal context across the input segment. In Table 2(b), we use a single decoder and predict both start and end at once, which results significant performance drop, from $49.5\\mathrm{mAP}$ to $42.7\\mathrm{mAP}$ . The result demonstrates the contribution of our new action localization method, using two separate decoders for the start and the end prediction. We also ablate new query design, sampling the memory for the start decoder, and positional embedding for instance queries. The results", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f47519afe84525e2db92d59394d0b3342c04a0930a5b2d57352adb9ff7689ecd.jpg", + "table_caption": [ + "Table 2: Ablation of the proposed modules on THUMOS14 dataset. The results are reported in mAP measure $(\\%)$ ." + ], + "table_footnote": [], + "table_body": "
Method0.30.40.50.60.7Average
Ours70.362.752.138.623.749.5
(a) Memory-augmented video encoder
w/o flag token67.361.050.136.522.347.4
w/o segment encoder65.859.349.936.321.846.6
(b) Instance decoding module
single decoder65.156.445.030.516.742.7
w/o splitting query66.261.450.938.023.247.9
w/o sampling66.159.650.639.922.547.2
w/o pos embedding67.260.549.836.322.747.3
(c) Training objective
w/o DIoU loss62.253.443.231.017.441.4
", + "bbox": [ + 225, + 199, + 514, + 349 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/231cfde0532c43f6d9c0deb0ee4aa02adf5fc92ae4e970e9adaf91a8959989b6.jpg", + "table_caption": [ + "Table 3: Analysis of the maximum length of the memory queue." + ], + "table_footnote": [], + "table_body": "
Memory size0.30.40.50.60.7Average
(a) THUMOS14 dataset
w/o mem65.858.947.636.920.846.0
167.259.950.737.922.747.7
367.361.951.937.922.648.3
770.362.752.138.623.749.5
1167.960.649.837.823.247.9
1566.960.150.738.424.148.0
1967.060.152.540.224.649.1
(b) MUSES dataset
w/o mem22.317.613.19.14.613.3
123.018.013.58.64.913.6
323.118.313.98.95.113.9
723.518.113.58.94.913.8
1122.918.714.09.45.414.1
1523.519.314.39.45.714.4
1922.718.913.89.65.514.1
", + "bbox": [ + 524, + 185, + 771, + 357 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d6f9e438d01ee68067b7fe725ec66f4da3403645e67ad91647c559ef978fa61c.jpg", + "table_caption": [ + "Table 4: Comparison between different memory modules on the THUMOS14." + ], + "table_footnote": [], + "table_body": "
Memory moduleBackboneSegment sizeInference timefpsMemory parametersAverage mAP
MAT [33]TSN [34]64191.9ms5.240.1M46.9
E2E-LOAD [2]TSN64196.1ms5.153.1M47.9
MATRTSN64167.1ms6.024.0M49.5
", + "bbox": [ + 233, + 397, + 767, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "show that all design choices of the proposed components are important for action instance localization. Lastly, training our model without DIoU loss shows a drastic performance drop, which demonstrates the importance of instance-level supervision for On-TAL (Table 2(c)).", + "bbox": [ + 212, + 476, + 784, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Impact of the Memory Queue Size. We conduct an ablation study on the different memory queue sizes. As shown in Table 3, employing the memory queue enhances the performance compared to the without memory case, and generally the performance improves as the memory queue size increases. With an adequately large memory queue size covering the top $99\\%$ of the instances in the training split, such as 7 for the THUMOS14 and 5 for the MUSES, the model shows robust performance and the performance is higher than the previous state-of-the-art model [15]. The memory queue size of 7 and 15 shows the best on THUMOS14 and MUSES, respectively.", + "bbox": [ + 212, + 537, + 785, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison of the proposed memory queue with OAD methods using memory. We compare our memory module with previous memory modules from OAD tasks [2, 33] by replacing the memory-augmented video encoder (Sec. 3.2) of our model. For a fair comparison, hidden dimension of the attention blocks for all methods is set to 1024. As shown in Table 4, MATR outperforms the existing memory modules in terms of both mAP and space-time complexity. While existing memory modules use multiple attention blocks, our memory module consists of a segment encoder and a flag prediction head, making it superior in terms of parameters and inference speed.", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis of the Segment Size. As shown in Fig. 4, our model is less sensitive to the input segment size, which should be carefully tuned across various", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0adb606a4a25c04d57fd640bb28efd7652879108e51a660b08ed545786c666e1.jpg", + "image_caption": [ + "Fig.4: Average mAP $(\\%)$ versus the segment size in THUMOS14. Red line represents the result of our model and blue line represents the result of OAT-OSN." + ], + "image_footnote": [], + "bbox": [ + 220, + 155, + 442, + 281 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0c2c06e58ff1701cf5f83bd7d1feca1b3c901168f4dace0faf6224c068122190.jpg", + "table_caption": [ + "Table 5: Ablation on the start and the end prediction heads. Reg and CIs refer to the offset regression and the region classification respectively." + ], + "table_footnote": [], + "table_body": "
StartEnd0.30.40.50.60.7Average
RegReg64.259.050.037.822.746.7
Reg + ClsReg70.362.752.138.623.749.5
Reg + ClsReg + Cls64.058.950.238.323.246.9
", + "bbox": [ + 467, + 195, + 776, + 243 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ed83e36216de42f43bfec490e3711ca1f133b194f30b42f1f52b98f9508f5c3e.jpg", + "table_caption": [ + "Table 6: Ablation on the memory compression factor on the THUMOS14 dataset." + ], + "table_footnote": [], + "table_body": "
Compress factor0.30.40.50.60.7Average
None (Ours)70.362.752.138.623.749.5
267.359.949.436.922.347.7
468.161.552.739.724.549.3
867.561.352.038.823.848.7
", + "bbox": [ + 464, + 290, + 781, + 363 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0be7ae2891cd75f8ce7dacd9f09848e79ca7c3dd2413685309acfb19d85681df.jpg", + "table_caption": [ + "Table 7: Comparison of the inference time and # parameters on the THUMOS14." + ], + "table_footnote": [], + "table_body": "
MethodBackboneSegment sizeInference timefps# ParametersAverage mAP
OAT-OSN [15]TSN [34]64163.7ms6.1128.7M44.6
MATRTSN64167.1ms6.0192.8M49.5
MATRTSN1653.8ms18.6192.8M46.3
", + "bbox": [ + 251, + 409, + 750, + 459 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "datasets. While the performance of OAT-OSN drops significantly from $44.6\\mathrm{mAP}$ to $25.8\\mathrm{mAP}$ when the segment size is reduced from 64 to 8, the performance drop of ours is only $9.1\\%$ .", + "bbox": [ + 212, + 488, + 785, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis of Start and End Prediction Heads. We investigate various offset prediction methods for the start prediction and the end prediction (Table 5). For the first method, start offset regression without region classification, the offset is normalized using the offset statistics from the training dataset. The second approach, which adopts region classification and offset regression for the start prediction, and only offset regression for the end prediction, exhibits the best performance. The reason for performance drop when using region classification for the end prediction is that frame-wise end classification is a complex problem, adversely affects the performance.", + "bbox": [ + 212, + 535, + 787, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis of Memory Compression Factor. We also conduct an experiment on efficiently storing input segment features in the memory by compressing them. We utilize a 1D convolution layer to compress the number of segment features. As shown in Table 6, $4 \\times$ compression yields competitive results with efficient memory usage. However, we decide not to use this method due to the efficient utilization of memory already achieved through the flag token and avoid to use additional hyperparameters that require optimization for each dataset.", + "bbox": [ + 212, + 672, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Inference Time. As shown in Table 7, when using the same backbone and the segment size, the inference time of MATR is almost equal to that of OAT-OSN, yet its average mAP is significantly better. Moreover, by reducing the segment length to 16, MATR outperforms OAT-OSN in terms of both mAP and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/85c2b19333401bd42aaf9ce423d6ced38564b16eeef31d5b4daebd5add361f7e.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of MATR and OAT-OSN [15] on the THUMOS14. Generated time is the timestamp when the predicted instance is generated." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 329 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the inference time. MATR uses more parameters due to the separation of the decoder into the start decoder and the end decoders.", + "bbox": [ + 212, + 398, + 787, + 430 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 454, + 421, + 469 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 5 illustrates the predictions of our model and OAT-OSN [15]. The results show that MATR has the capability to efficiently detect action instances by identifying intervals immediately after the current action ends. Additionally, the results show that our model is able to classify and localize action instances more precisely than OAT-OSN.", + "bbox": [ + 212, + 479, + 787, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 580, + 359, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we address online temporal action localization by introducing a new model, MATR, which leverages the memory queue to exploit long-term context. We also propose a new action instance localization method based on Transformer, which separately identifies action starts and action ends. As a result, MATR not only surpasses existing online TAL methods but also demonstrates comparable performance to offline TAL methods. The code base will be open to the public to promote future research on On-TAL.", + "bbox": [ + 212, + 613, + 787, + 718 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations: When the memory queue contains multiple action instances, there is a risk of action instances being matched with incorrect start points. Additionally, when storing the input segment into the memory queue, MATR only considers whether the input segment is relevant to action instances but does not leverage the relationship with the past context stored in the memory. To address these issues, investigating a method to utilize the information stored in the memory queue when storing the input segment features, could lead to more effective memory storage and utilization.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the NRF grant and the IITP grant funded by Ministry of Science and ICT, Korea (RS-2019-II191906, RS-2021-II212068, RS-2022-II220290, NRF-2021R1A2C3012728).", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 212, + 323, + 228 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Buch, S., Escorcia, V., Shen, C., Ghanem, B., Carlos Niebles, J.: Sst: Single-stream temporal action proposals. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2911-2920 (2017)", + "2. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. Proc. IEEE International Conference on Computer Vision (ICCV) (2023)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020)", + "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "5. Chen, P., Gan, C., Shen, G., Huang, W., Zeng, R., Tan, M.: Relation attention for temporal action localization. IEEE Transactions on Multimedia 22(10), 2723-2733 (2019)", + "6. Cheng, F., Bertasius, G.: Tallformer: Temporal action localization with a long-memory transformer. In: Proc. European Conference on Computer Vision (ECCV). pp. 503-521. Springer (2022)", + "7. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Proc. European Conference on Computer Vision (ECCV). pp. 269-284. Springer (2016)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. International Conference on Learning Representations (ICLR) (2021), https://openreview.net/forum?id=YicbFdNTTy", + "9. Escorcia, V., Caba Heilbron, F., Niebles, J.C., Ghanem, B.: Daps: Deep action proposals for action understanding. In: Proc. European Conference on Computer Vision (ECCV). pp. 768-784. Springer (2016)", + "0. Gao, J., Yang, Z., Chen, K., Sun, C., Nevatia, R.: Turn tap: Temporal unit regression network for temporal action proposals. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3628-3636 (2017)", + "1. Gao, M., Xu, M., Davis, L.S., Socher, R., Xiong, C.: Startnet: Online detection of action start in untrimmed videos. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5542-5551 (2019)", + "2. Heilbron, F.C., Niebles, J.C., Ghanem, B.: Fast temporal activity proposals for efficient detection of human actions in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1914-1923 (2016)", + "3. Jiang, Y.G., Liu, J., Roshan Zamir, A., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/ (2014)", + "4. Kang, H., Kim, K., Ko, Y., Kim, S.J.: Cag-qil: Context-aware actionness grouping via q imitation learning for online temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13729-13738 (2021)" + ], + "bbox": [ + 225, + 239, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Proc. European Conference on Computer Vision (ECCV). pp. 653-669. Springer (2022)", + "16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proc. International Conference on Learning Representations (ICLR) (2015)", + "17. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3320-3329 (2021)", + "18. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3889-3898 (2019)", + "19. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: Proc. European Conference on Computer Vision (ECCV). pp. 3-19 (2018)", + "20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)", + "21. Liu, Q., Wang, Z.: Progressive boundary refinement network for temporal action detection. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 11612-11619 (2020)", + "22. Liu, X., Hu, Y., Bai, S., Ding, F., Bai, X., Torr, P.H.: Multi-shot temporal event localization: a benchmark. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12596-12606 (2021)", + "23. Liu, X., Wang, Q., Hu, Y., Tang, X., Zhang, S., Bai, S., Bai, X.: End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing 31, 5427-5441 (2022)", + "24. Liu, Y., Ma, L., Zhang, Y., Liu, W., Chang, S.F.: Multi-granularity generator for temporal action proposal. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3604-3613 (2019)", + "25. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Gaussian temporal awareness networks for action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 344-353 (2019)", + "26. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "27. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18857-18866 (2023)", + "28. Shou, Z., Chan, J., Zareian, A., Miyazawa, K., Chang, S.F.: Cdc: Convolutional-de-convolutional networks for precise temporal action localization in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5734-5743 (2017)", + "29. Shou, Z., Pan, J., Chan, J., Miyazawa, K., Mansour, H., Vetro, A., Giro-i Nieto, X., Chang, S.F.: Online detection of action start in untrimmed, streaming videos. In: Proc. European Conference on Computer Vision (ECCV). pp. 534-551 (2018)", + "30. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1049-1058 (2016)", + "31. Tang, T.N., Park, J., Kim, K., Sohn, K.: Simon: A simple framework for online temporal action localization. arXiv preprint arXiv:2211.04905 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. Advances in Neural Information Processing Systems (NeurIPS). pp. 5998-6008 (2017)", + "33. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13824-13835 (2023)", + "34. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proc. European Conference on Computer Vision (ECCV). pp. 20-36. Springer (2016)", + "35. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7565-7575 (2021)", + "36. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13587-13597 (2022)", + "37. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10186-10195 (2020)", + "38. Xu, M., Perez Rua, J.M., Zhu, X., Ghanem, B., Martinez, B.: Low-fidelity video encoder optimization for temporal action localization. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 9923-9935 (2021)", + "39. Xu, M., Zhao, C., Rojas, D.S., Thabet, A., Ghanem, B.: G-tad: Sub-graph localization for temporal action detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10156-10165 (2020)", + "40. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5532-5541 (2019)", + "41. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long shortterm transformer for online action detection. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 1086-1099 (2021)", + "42. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3160-3169 (2022)", + "43. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7094-7103 (2019)", + "44. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 492-510. Springer (2022)", + "45. Zhao, C., Thabet, A.K., Ghanem, B.: Video self-stitching graph network for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13658-13667 (2021)", + "46. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 485-502. Springer (2022)", + "47. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2914-2923 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Online Temporal Action Localization with Memory-Augmented Transformer", + "bbox": [ + 222, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., Ren, D.: Distance-iou loss: Faster and better learning for bounding box regression. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 12993-13000 (2020)", + "49. Zhu, Z., Tang, W., Wang, L., Zheng, N., Hua, G.: Enriching local and global contexts for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13516-13525 (2021)" + ], + "bbox": [ + 217, + 146, + 785, + 231 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_model.json b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..20f81b1eec3ad496f31253c0657959bfd6d6a14a --- /dev/null +++ b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_model.json @@ -0,0 +1,2326 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.262, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.213, + 0.363, + 0.228 + ], + "angle": 0, + "content": "Youngkil Song\\*" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.214, + 0.52, + 0.228 + ], + "angle": 0, + "content": "Dongkeun Kim*" + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.214, + 0.643, + 0.227 + ], + "angle": 0, + "content": "Minsu Cho" + }, + { + "type": "text", + "bbox": [ + 0.672, + 0.214, + 0.768, + 0.227 + ], + "angle": 0, + "content": "Suha Kwak" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.24, + 0.746, + 0.283 + ], + "angle": 0, + "content": "Pohang University of Science and Technology (POSTECH), South Korea {songyk, kdk1563, mscho, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/MATR/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.321, + 0.744, + 0.502 + ], + "angle": 0, + "content": "Abstract. Online temporal action localization (On-TAL) is the task of identifying multiple action instances given a streaming video. Since existing methods take as input only a video segment of fixed size per iteration, they are limited in considering long-term context and require tuning the segment size carefully. To overcome these limitations, we propose memory-augmented transformer (MATR). MATR utilizes the memory queue that selectively preserves the past segment features, allowing to leverage long-term context for inference. We also propose a novel action localization method that observes the current input segment to predict the end time of the ongoing action and accesses the memory queue to estimate the start time of the action. Our method outperformed existing methods on two datasets, THUMOS14 and MUSES, surpassing not only TAL methods in the online setting but also some offline TAL methods." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.515, + 0.741, + 0.53 + ], + "angle": 0, + "content": "Keywords: Temporal action localization \\(\\cdot\\) Online video understanding" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.377, + 0.572 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.723 + ], + "angle": 0, + "content": "Today, video is one of the most popular types of media, with tons of videos being created and uploaded to online platforms like YouTube and TikTok every second. As a result, the demand for video understanding is growing rapidly. In particular, dealing with untrimmed videos is a key to practical video understanding since most videos in the real world are not trimmed for individual events in advance. Temporal action localization (TAL) has been extensively studied in this context [10, 17, 21, 30, 38, 39, 43-45, 47, 49], with the aim of detecting each action instance in the input untrimmed video by predicting its start time, end time, and action class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Recently, online TAL (On-TAL), i.e., TAL for streaming video [14], has gained increasing attention thanks to its potential applications in video surveillance, sports video analysis, and video summarization. Unlike TAL which demands the entire video as input, On-TAL aims to detect action instances using only input frames observed so far in an online manner. Moreover, once action instances are predicted, it is not allowed to modify the results afterward." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.365, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig. 1: A conceptual illustration of our model. Our model selectively stores information about previously occurred actions in the memory queue from streaming video. When the end of the action is detected in the current input segment, the model retrieves information from the memory queue to locate the start of the action." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.789, + 0.64 + ], + "angle": 0, + "content": "A straightforward solution to this challenging task is to first conduct online action detection (OAD) [7,35], i.e., frame-by-frame action classification for streaming video, and then build action instances by aggregating the frame-wise action predictions. In this sense, earlier methods for On-TAL [14,31] follow this approach by applying OAD and refining per-frame predictions using past observations. However, this OAD-based approach exploits only frame-level supervision for training, which is not optimal because On-TAL aims to predict action instances in the form of time intervals. To alleviate this issue, online anchor transformer (OAT) has been recently proposed [15], which adopts sliding window approach and anchor-based method to utilize instance-level supervision. Although OAT improved performance remarkably, however, it still has several drawbacks: Since it takes a fixed-size segment of input video at each iteration and does not consider long-term contexts, its capability is limited in detecting long action instances, and its performance is sensitive to hyperparameters like the size of the input segment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.789, + 0.748 + ], + "angle": 0, + "content": "To resolve the above issues, we propose a new end-to-end architecture for OntAL, dubbed memory-augmented Transformer (MATR). At the heart of MATR lies in memory queue, which selectively stores past segment features so that the model exploits long-term context for inference. Thanks to the memory queue that contains long-term context and reduces dependence on the segment size, our model accurately localizes even long-term action instances without the need for careful hyperparameter tuning for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Moreover, we propose a novel approach to action instance localization using the memory queue as illustrated in Fig. 1. First, it detects the end of an action using the current segment features and then scans past segment features in the memory queue to find the action start corresponding to the detected action end. To achieve this, we adopt two Transformer decoders: one for end detection and the other for start detection. Specifically, we adopt learnable queries as input to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.221 + ], + "angle": 0, + "content": "the decoders so that each of them learns to localize action boundaries through the attention mechanism of Transformer [8,32]. In addition, motivated by recent advances on object detection [6,37], MATR decouples action classification and localization to further improve the performance. This is realized in MATR by using distinct queries for the two subtasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.298 + ], + "angle": 0, + "content": "The proposed method is evaluated on two datasets, THUMOS14 [13] and MUSES [22]. Our model achieves the state of the art on both datasets in the online TAL setting. In addition, surprisingly, it is also comparable to existing offline TAL methods although it does not employ offline non-maximum suppression (NMS). In summary, the contribution of this paper is three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.313, + 0.784, + 0.357 + ], + "angle": 0, + "content": "- We propose MATR, a new end-to-end architecture for On-TAL, leveraging memory queue to utilize long-term context and enabling precise localization of action instances with less dependence on dataset-specific hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.359, + 0.785, + 0.417 + ], + "angle": 0, + "content": "- We introduce a new action instance localization method that first identifies action end using the input segments and scans past information in the memory queue to find the action start. Moreover, our method adopts distinct queries to separate information for action classification and localization." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.42, + 0.784, + 0.463 + ], + "angle": 0, + "content": "- Our method outperformed the existing On-TAL methods on the two benchmarks. Also, extensive ablation studies demonstrate the contribution of each component of the proposed method." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.313, + 0.785, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.387, + 0.505 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.688 + ], + "angle": 0, + "content": "Temporal Action Localization. Temporal action localization (TAL) [5,10,17, 21,23,30,38,39,43-45,47,49] which identifies action instances in time and classifying their actions, has given much attention on the video research community. Earlier attempts employ two-stage approaches that first generate action proposals, then classify action class and refine action boundaries [1,9,12,18,19,24,47]. Recent studies have primarily dealt with widening the range of video context through convolutional neural networks [28], graph neural networks [39,43], Gaussian kernel [25], and utilizing global context [49]. To extend temporal context and enable end-to-end TAL, Transformer [8,32] has recently been adopted [6,23,44]. These approaches incorporate temporal deformable attention module [23] and multi-scale features [44] to improve computational efficiency and performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Online Video Understanding. Online action detection (OAD), which classifies action classes in a frame-wise manner, is first proposed by Geest et al. [7]. To incorporate long-term context, recurrent neural network [40] and Transformer [35] are utilized. These methods predict actions for future frames and use this information to refine current frame predictions. Yang et al. [42] further improve the performance by using category exemplars and capturing long-term dependencies within a video segment. On the other hand, Shou et al. [29] proposes online detection of action start (ODAS), which aims to detect action start as early as possible. Gao et al. [11] further improves ODAS by adopting reinforcement learning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Online Temporal Action Localization. Towards practical event understanding, online temporal action localization (On-TAL) has been proposed recently. Kang et al. [14] first propose On-TAL, which localizes action boundaries and classifies action classes without accessing future frames and offline post-processing. Earlier methods [14,31] adopt OAD framework to classify per-frame action class and group them into action instances to solve On-TAL. Specifically, Kang et al. [14] utilize Markov decision process to aware prior contexts to group per-frame predictions. SimOn [31] is proposed as an end-to-end On-TAL method that leverages past visual context and action probabilities for accurate action prediction at the current frame. Kim et al. [15] propose a sliding window scheme and an anchor-based method that adopts Transformer [32] for On-TAL. However, it only uses a fixed input segment which is highly dependent on the action length distribution of each dataset. Unlike the previous methods, we utilize a memory queue for long-term context, which also reduces the dependence on the input segment size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Difference between OAD and On-TAL. OAD [2,33,35,40-42,46] focus on predicting action classes at the frame-level, which limits the ability to distinguish boundaries of overlapping actions. On the other hand, On-TAL [14,15,31] predicts entire action instances rather than individual frames, enabling it to precisely identify the start and end times of actions, making it effective for applications requiring instance-level action understanding, like sports video analysis. Using Memory for Video Understanding. Prior work in video understanding employs memory to store contextual information, in the forms of long-term and short-term memory [2,33,36,41]: short-term memory captures fine details, and long-term memory stores compressed information of past contexts. Specifically, Wang et al. [33] divides long-term memory into several groups and compresses each group into a single vector. Stream Buffer [2] gradually reduces temporal dimensions of the memory using a compression module. However, unlike video classification tasks such as action recognition [36] and OAD [2,33,41,46], which requires to store representative information of past frames in the memory for classifying the action of the current frame, TAL needs to preserve the temporal positional information of past frames in the memory to accurately predict the exact time. In that sense, MATR preserves temporal information in memory to accurately localize actions and effectively stores information by using flag tokens. A comparison with the OAD memory modules is shown in Table 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.427, + 0.719 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Consider an untrimmed video \\(\\mathbf{V} = \\{\\mathbf{v}_i\\}_{i=1}^T\\) with \\(T\\) frames and \\(M\\) action instances \\(\\varPsi=\\{(s_m,e_m,c_m)\\}_{m=1}^{M}\\), where each instance is represented by its start time \\(s_m\\), end time \\(e_m\\) and action class \\(c_m\\). Unlike temporal action localization (TAL), which allows a model to use whole video sequence as input, online temporal action localization (On-TAL) forces a model to predict action instances using only input frames seen until the current timestamp. Note that action instances predicted at previous iterations cannot be modified or deleted afterwards." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.785, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.314, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Fig. 2: Overall architecture of MATR. MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.789, + 0.655 + ], + "angle": 0, + "content": "As shown in Fig. 2, MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads. In line with previous research, the current segment of the input streaming video is given as a unit input, and its frame-level features, referred to as segment features, are extracted by a video backbone network and a linear projection layer (Sec. 3.1). The segment features are then fed to the memory-augmented video encoder, which encodes temporal context between frames in the current segment and stores the segment features into the memory (Sec. 3.2). The instance decoding module localizes action instances via two Transformer decoders: the end decoder and the start decoder. Specifically, the end decoder references the encoded segment features to locate the action end around the current time, and then the start decoder refers to the memory queue to find the action start based on the past information stored in the memory queue. Queries for each instance consist of a class query for action classification and a boundary query for action localization (Sec. 3.3). The outputs of the instance decoding module are used as inputs to the prediction heads, which consist of end prediction head, start prediction head, and action classification head (Sec. 3.4). The entire model is trained in an end-to-end manner (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.411, + 0.714 + ], + "angle": 0, + "content": "3.1 Feature Extractor" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given a segment comprising \\(L_{s}\\) consecutive video frames as input, a pretrained backbone [4, 34] followed by a linear projection layer extracts the segment features \\(\\mathbf{X}_t = \\{\\mathbf{x}_i\\}_{t - L_s + 1}^t\\in \\mathbb{R}^{L_s\\times D}\\) for each frame within the segment, following previous work [14, 15]. Inspired by Kim et al. [15], MATR employs a sliding window scheme, which predicts multiple action instances by moving frame by frame along the temporal axis. Note that the window size is the input segment size." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.561, + 0.163 + ], + "angle": 0, + "content": "3.2 Memory-Augmented Video Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.787, + 0.248 + ], + "angle": 0, + "content": "At each time step, the input segment features are fed to two modules of the memory-augmented video encoder, the segment encoder and the memory update module. The segment encoder encodes the temporal context across input segment features, while the memory update module selectively stores the segment features and updates the memory queue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Segment Encoder. The segment encoder is a standard Transformer encoder composed of self-attention layers and a feed forward network (FFN). The combination of segment features and a learnable flag token is fed into the encoder, and transformed to queries, keys, and values for self-attention. Sinusoidal positional encoding \\(\\mathbf{S}_{\\mathrm{pos}} \\in \\mathbb{R}^{(L_s + 1) \\times D}\\) is added to the queries and keys. Among the output embeddings, the encoded flag token is fed into the memory update module and the encoded segment features that integrate temporal context across input frames are used as input to the end decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.789, + 0.626 + ], + "angle": 0, + "content": "Memory Update Module. In an online setting, where an input is a streaming video, efficiently storing information of past frames and accessing it effectively is crucial for detecting action instances, in particular those of long-term actions whose temporal extents are beyond the size of input segments. To this end, we adopt the memory queue for storing information of past input segments in a first-in-first-out (FIFO) manner. Additionally, we propose an efficient memory update method employing the flag token. When the input segment features pass through the segment encoder, the flag token is concatenated into the input segment features and then employed in the flag prediction head. It is trained to predict [FLAG], which identifies whether an input segment is relevant to action instances, so that only relevant segments are stored in the memory. [FLAG] = 1 when there are overlapping frames between the input segment and action instances, and 0 otherwise. During training, the ground-truth [FLAG] is utilized. During inference, [FLAG] = 1 if sigmoid(\\hat{g}) > \\theta where \\(\\hat{g}\\) is the output logit of the flag prediction head and \\(\\theta\\) is a predefined threshold, and 0 otherwise. The input segment feature is added to the memory queue when [FLAG] = 1 and discards it when [FLAG] = 0. If the memory queue is full, the oldest segment feature is purged." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.483, + 0.664 + ], + "angle": 0, + "content": "3.3 Instance Decoding Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The instance decoding module localizes and classifies action instances by leveraging the encoded segment features and the memory queue through the attention mechanism of Transformer. Given the encoded segment features that have short-term temporal contexts across input segment and the memory queue that stores long-term contexts for actions that are potentially ongoing, a set of \\(2N\\) instance queries \\(\\mathbf{Q} = \\{\\mathbf{Q}_{\\mathrm{class}}; \\mathbf{Q}_{\\mathrm{bound}}\\} \\in \\mathbb{R}^{2N \\times D}\\) is trained to generate \\(N\\) action instances for each input segment. Half of these queries pertain to predicting the action class (class query; \\(\\mathbf{Q}_{\\mathrm{class}} \\in \\mathbb{R}^{N \\times D}\\)), while the other half is used for predicting the start and end timestamps (boundary query; \\(\\mathbf{Q}_{\\mathrm{bound}} \\in \\mathbb{R}^{N \\times D}\\)). The class and boundary query pairs for the same instance share the same positional embedding \\(\\mathbf{E}_{\\mathrm{pos}} \\in \\mathbb{R}^{N \\times D}\\) to accurately identify and differentiate between" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.314, + 0.709, + 0.328 + ], + "angle": 0, + "content": "Fig. 3: Detailed architecture of the instance decoding module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.455 + ], + "angle": 0, + "content": "instances; \\(\\tilde{\\mathbf{Q}} = \\{\\mathbf{Q}_{\\mathrm{class}} + \\mathbf{E}_{\\mathrm{pos}};\\mathbf{Q}_{\\mathrm{bound}} + \\mathbf{E}_{\\mathrm{pos}}\\}\\). Unlike the previous Transformer-based On-TAL method [15] that predicts start and end offset for each instance simultaneously, our method employs separate Transformer decoders for predicting start offset and end offset. As shown in Fig. 3, the start decoder and end decoders share the same architecture but utilizes different information to predict action start and action end, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.563 + ], + "angle": 0, + "content": "End Decoder. Given the encoded segment features, instance queries \\(\\mathbf{Q}\\) is trained to generate the output embeddings used for detecting action end near the current timestamp. To this end, we adopt a Transformer [3] composed of multi-head self-attention layers, multi-head cross-attention layers, and FFNs. The detailed architecture is illustrated in Fig. 3. The outputs of the end decoder, which consist of end boundary embeddings and class embeddings, are then used as inputs to the start decoder and the prediction heads." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.787, + 0.732 + ], + "angle": 0, + "content": "Start Decoder. The start decoder takes the output embeddings from the end decoder and locates the corresponding action start by utilizing the memory queue. First, the memory queue is concatenated with the current segment features. The concatenated memory features are used as long-term context in the start decoder. Then, 2D temporal positional encodings are added to the memory features, and the results are used as the key and value for the cross-attention layers of the start decoder. The overall architecture of the start decoder is the same as the end decoder (Fig. 3), but it uses the memory features rather than the encoded segment features. The output embeddings of the start decoder, start boundary embeddings and class embeddings, are used as inputs to the start prediction and the action classification head, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Moreover, we introduce two techniques to efficiently and effectively utilize the memory queue. First, we conduct a \\(50\\%\\) uniform sampling to the memory features which enables efficient use of memory since adjacent frames contain similar information. Second, to expand the scope of positional encoding for streaming video, whose duration is unpredictable, we separate the temporal positional encoding into the two parts: relative segment position and relative frame position. A relative segment position is the position of a segment within the memory rela" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "tive to the current segment. Meanwhile, a relative frame position is the position of a frame relative to the most recent frame within the same segment." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.199, + 0.406, + 0.213 + ], + "angle": 0, + "content": "3.4 Prediction Heads" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.784, + 0.282 + ], + "angle": 0, + "content": "To generate \\(N\\) action instances \\(\\{(\\hat{s}_i,\\hat{e}_i,\\hat{c}_i)\\}_{i = 1}^N\\) from the output of the instance decoding module, we use three prediction heads, which are end prediction head, start prediction head, and action classification head. Each of the prediction heads is composed of a 2-layer FFN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.344 + ], + "angle": 0, + "content": "End Prediction Head. The end prediction head estimates the offset between the end of the target action and the current timestamp. The boundary embeddings from the end decoder are fed to the 2-layer FFN, which in turn estimates the end offset \\(\\{\\hat{u}_i\\}_{i=1}^N \\in \\mathbb{R}^N\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.787, + 0.629 + ], + "angle": 0, + "content": "Start Prediction Head. The start prediction head estimates the offset between the start of the target action and the current timestamp. Unlike the end prediction, which finds the end offset around the input segment, the start prediction needs to perform offset regression over a relatively wider range. To narrow down the scope of the start offset regression, we divide the time horizon into \\( L_{m} + 2 \\) regions: the region preceding the coverage of memory, \\( L_{m} \\) regions covered by the memory, and the region corresponding to the current input segment. In that sense, the start prediction head consists of the region classification and the offset regression, each composed of a 2-layer FFN. They take the boundary embeddings from the start decoder as input to predict the start time. The region classification identifies the regions to which the start time is assigned, represented as \\( \\{\\hat{o}_i|\\hat{o}_i = \\mathrm{argmax}(\\hat{r}_i)\\}_{i=1}^N \\in \\mathbb{R}^N \\). Here, \\( \\{\\hat{r}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (L_m + 2)} \\) are the output logits from the region classification head. Following this, the offset regression head is utilized to estimate the offset within the regions, denoted as \\( \\{\\hat{v}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (L_m + 2)} \\). Note that the start offset \\( \\hat{v}_i \\) is calculated for all \\( L_{m} + 2 \\) regions, and the offset from the identified region \\( \\{\\hat{v}_i(\\hat{o}_i)\\}_{i=1}^N \\in \\mathbb{R}^N \\) is used in inference. By combining start classification result \\( \\hat{o}_i \\) and start regression result \\( \\hat{v}_i(\\hat{o}_i) \\), the start time is predicted. The effectiveness of the proposed start prediction head is validated in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.784, + 0.691 + ], + "angle": 0, + "content": "Action Classification Head. For action class prediction, we utilize the class embeddings from both the end decoder and the start decoder. These embeddings are concatenated and fed into the action classification head to derive class probabilities \\(\\{\\hat{p}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (C+1)}\\), where \\(C\\) is the number of action class." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.691, + 0.784, + 0.722 + ], + "angle": 0, + "content": "Action Instance Prediction. Finally, our model makes \\(N\\) action proposals \\(\\{(\\hat{s}_i,\\hat{e}_i,\\hat{c}_i)\\}_{i = 1}^N\\) at time \\(t\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.733, + 0.784, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {i} = t - \\left(\\hat {o} _ {i} + \\hat {v} _ {i} \\left(\\hat {o} _ {i}\\right)\\right) \\times L _ {s}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.751, + 0.785, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\hat {e} _ {i} = t + \\hat {u} _ {i} \\times L _ {s}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.77, + 0.785, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\hat {c} _ {i} = \\operatorname {a r g m a x} \\left(\\hat {p} _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\( L_{s} \\) is the segment length. Since the sliding window scheme generates \\( N \\) action proposals at each timestamp, post-processing is crucial for performance improvement by removing redundant or overlapped action instances." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Non-maximum suppression (NMS) is applied to the action proposals at each timestamp, subsequently removing those highly overlapped with the proposals generated from the past. To prevent more reliable prediction made in the future from being removed, instances where the predicted end time is beyond the current timestamp, \\( t < \\hat{e}_i \\), are also removed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.242, + 0.419, + 0.258 + ], + "angle": 0, + "content": "3.5 Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.264, + 0.784, + 0.34 + ], + "angle": 0, + "content": "At timestamp \\( t \\), the model is trained to detect action instances whose end times are in the range \\( [t - T_d + 1, t + T_a] \\), where \\( T_d \\) and \\( T_a \\) are hyperparameters. Then, hungarian algorithm matches the action proposals and the ground-truth action instances with the lowest matching cost. The matching cost of the ground-truth group \\( i \\) and the proposal \\( \\sigma(i) \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.348, + 0.786, + 0.367 + ], + "angle": 0, + "content": "\\[\nC _ {i, \\sigma (i)} = \\hat {p} _ {\\sigma (i)} \\left(y _ {i}\\right) + \\operatorname {I o U} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.372, + 0.788, + 0.48 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the permutation of \\(N\\) action proposals, \\(\\hat{p}_i\\) is the class probabilities of the \\(i\\)-th proposal, \\(b_i = \\{s_i, e_i\\}\\) is the ground-truth action boundary, and \\(\\hat{b}_i = \\{\\hat{s}_i, \\hat{e}_i\\}\\) is the predicted action boundary. We adopt the focal loss [20] for the action classification, the cross entropy loss for the start region classification, the \\(\\ell_1\\) loss for both the start offset regression and the end offset regression. The action classification loss \\(L_{\\mathrm{class}}\\), the start prediction loss \\(L_{\\mathrm{start}}\\), and the end prediction loss \\(L_{\\mathrm{end}}\\) are defined as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.488, + 0.786, + 0.527 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c l a s s}} = \\sum_ {i = 1} ^ {N} \\operatorname {F L} \\left(\\hat {p} _ {i}, y _ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.532, + 0.785, + 0.571 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s t a r t}} = \\sum_ {i = 1} ^ {N _ {\\text {m a t c h}}} \\left\\{\\mathrm {C E} \\left(\\operatorname {S o f t m a x} \\left(\\hat {r} _ {i}\\right), r _ {i}\\right) + \\left| \\hat {v} _ {i} \\left(\\hat {o} _ {i}\\right) - v _ {i} \\right| \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.576, + 0.785, + 0.616 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {e n d}} = \\sum_ {i = 1} ^ {N _ {\\text {m a t c h}}} | \\hat {u} _ {i} - u _ {i} |, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.622, + 0.784, + 0.651 + ], + "angle": 0, + "content": "where \\( y_{i}, r_{i}, v_{i}, u_{i} \\) is the ground-truth action class, start region, start offset, and end offset, respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.652, + 0.784, + 0.697 + ], + "angle": 0, + "content": "To provide instance-level supervision and facilitate the connection between start and end time prediction, we employ the DIoU loss [48] following Action-Former [44]:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.699, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {d i o u}} = \\sum_ {i = 1} ^ {N _ {\\mathrm {m a t c h}}} 1 - \\operatorname {I o U} (\\hat {b} _ {i}, b _ {i}) + \\frac {\\rho^ {2} (\\hat {c} _ {i} , c _ {i})}{d _ {i} ^ {2}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.742, + 0.784, + 0.786 + ], + "angle": 0, + "content": "where \\(\\rho(\\cdot, \\cdot)\\) is the Euclidean distance between two points, \\(\\hat{c}_i\\) and \\(c_i\\) are the center of the proposal and ground-truth instance, and \\(d_i\\) is the smallest enclosing 1-dim box length." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.787, + 0.784, + 0.817 + ], + "angle": 0, + "content": "To train the memory update module in an end-to-end manner, we employ the flag loss for training the flag prediction head:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.826, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {f l a g}} = \\operatorname {B C E} (\\operatorname {S i g m o i d} (\\hat {g}), g), \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "where \\(\\hat{g}\\) and \\(g\\) is the predicted logit of flag token and ground-truth [FLAG] respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.178, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Our model is trained with six losses simultaneously in an end-to-end manner. The total loss is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.221, + 0.784, + 0.235 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {c l a s s}} + L _ {\\text {s t a r t}} + L _ {\\text {e n d}} + L _ {\\text {d i o u}} + L _ {\\text {f l a g}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.246, + 0.761, + 0.262 + ], + "angle": 0, + "content": "Note that all loss coefficients are set to 1 and loss balancing is not required." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.285, + 0.376, + 0.303 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.316, + 0.44, + 0.332 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.785, + 0.431 + ], + "angle": 0, + "content": "Datasets. We evaluate our method and previous methods on two On-TAL benchmarks: THUMOS14 [13] and MUSES [22]. THUMOS14 contains 200 videos for training and 213 videos for testing with 20 action classes, while MUSES has 2,587 videos for training and 1,110 videos for testing with 25 action classes. MUSES consists of multi-shot action instances, which makes action localization more challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Hyperparameters. We use a frozen two-stream TSN [34] pretrained on Kinetics [4] to extract RGB and flow features for THUMOS14, and an I3D [4] trained on Kinetics [4] utilizing only RGB features for MUSES, following previous OntAL work [14, 15, 31]. The segment size \\( L_{s} \\) is set to 64 in THUMOS14 and 75 in MUSES. The dimension of the segment feature is set to \\( D = 1024 \\) for both datasets. For the memory-augmented video encoder, we stack 3 Transformer layers with 8 attention heads for both datasets. The flag threshold \\( \\theta \\) is set to 0.5. The memory size \\( L_{m} \\) is set to 7 for THUMOS14, while 15 for MUSES. For the instance decoding module, we stack 5 Transformer layers with 4 attention heads for both datasets. The number of the class query and the boundary query \\( N = 10 \\) for THUMOS14 and \\( N = 6 \\) for MUSES." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.598, + 0.785, + 0.673 + ], + "angle": 0, + "content": "Training. During training, we use Adam optimizer [16], with the initial learning rate of 1e-8 with CosineAnnealing scheduler [26]. The batch size is set to 64 for THUMOS14, while 75 for MUSES. The focal loss coefficient is \\(\\alpha = 0.25\\) and \\(\\gamma = 2\\) for THUMOS14 and \\(\\alpha = 1\\) and \\(\\gamma = 5\\) for MUSES. All loss coefficients are set to 1. More details are listed in the supplementary material (Sec. D)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.576, + 0.71 + ], + "angle": 0, + "content": "4.2 Comparison with the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We compare our method with previous online and offline TAL methods on THUMOS14 and MUSES dataset. As presented in Table 1, our method outperforms all previous On-TAL methods in both benchmarks by a substantial margin: \\(4.9\\%\\) of average mAP in THUMOS14 and \\(0.7\\%\\) of average mAP in MUSES. The results show that MATR utilizing the memory queue is more effective than previous OAD-based methods [14,31,33,46] and instance-level method OAT-OSN [15]. We also evaluate two OAD methods utilizing memory [33,46] in On-TAL setting by grouping their frame-level predictions into action instances." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison with On-TAL methods on THUMOS14 and MUSES dataset. The results are reported in mAP measure (%). **' indicates that the result is validated using the On-TAL ground truth from classification annotation." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.789, + 0.411 + ], + "angle": 0, + "content": "
MethodTHUMOS14MUSES
backbone0.30.40.50.60.7Averagebackbone0.30.40.50.60.7
Offline TAL
G-TAD [39]TSN [34]54.547.640.230.823.439.9I3D [4]19.114.811.17.44.7
ContextLoc [49]I3D68.363.854.341.826.250.9------
P-GCN [43]I3D63.657.849.1---I3D19.917.113.19.75.4
MUSES [22]I3D68.964.056.946.331.053.4I3D25.922.618.915.010.6
ActionFormer [44]I3D82.177.871.059.443.966.8------
TriDet [27]I3D83.680.172.962.447.469.3------
OAD-based Online TAL
TeSTra [46]TSN35.629.221.413.47.621.4------
MAT [33]I3D36.027.519.212.36.420.3------
CAG-QIL [14]TSN44.737.629.821.914.529.7I3D8.56.54.22.81.9
SimOn [31]TSN54.345.035.023.314.634.4------
SimOn* [31]TSN57.047.537.326.616.036.9------
Instance-level Online TAL
OAT-Naive [15]TSN57.650.643.030.015.739.4I3D20.316.612.97.73.6
OAT-OSN [15]TSN63.056.747.136.320.044.6I3D22.118.514.28.94.7
MATRTSN70.362.752.138.623.749.5I3D23.519.314.39.45.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.592 + ], + "angle": 0, + "content": "However, their performance is far lower than MATR, which verifies the effectiveness of the proposed memory queue of MATR. We also compare our model with several offline TAL methods. Although there is still a performance gap between online and offline methods, our model achieves comparable performance to offline methods, and even outperforms previous work [39, 43]. The performance on the MUSES dataset is lower compared to THUMOS14 and the reason is twofold: (1) Action instances in MUSES are captured in multi-shots, making it challenging to detect action boundaries, and (2) there are confusing action classes such as \"quarrel\" and \"fight,\" making action classification hard. These make MUSES challenging and limit the upper-bound of performance on it." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.403, + 0.63 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.673 + ], + "angle": 0, + "content": "We perform ablation studies on THUMOS14 and MUSES dataset to show the effectiveness of the proposed modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Ablation on the Proposed Modules. In Table 2, we conduct an ablation study to verify the effectiveness of each component of our model. In Table 2(a), focuses on the components of the memory-augmented video, both removing the flag token and the segment encoder led to a performance drop. The results show the efficacy of selectively storing the past information and the importance of temporal context across the input segment. In Table 2(b), we use a single decoder and predict both start and end at once, which results significant performance drop, from \\(49.5\\mathrm{mAP}\\) to \\(42.7\\mathrm{mAP}\\). The result demonstrates the contribution of our new action localization method, using two separate decoders for the start and the end prediction. We also ablate new query design, sampling the memory for the start decoder, and positional embedding for instance queries. The results" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "table_caption", + "bbox": [ + 0.224, + 0.158, + 0.511, + 0.199 + ], + "angle": 0, + "content": "Table 2: Ablation of the proposed modules on THUMOS14 dataset. The results are reported in mAP measure \\((\\%)\\)." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.2, + 0.516, + 0.35 + ], + "angle": 0, + "content": "
Method0.30.40.50.60.7Average
Ours70.362.752.138.623.749.5
(a) Memory-augmented video encoder
w/o flag token67.361.050.136.522.347.4
w/o segment encoder65.859.349.936.321.846.6
(b) Instance decoding module
single decoder65.156.445.030.516.742.7
w/o splitting query66.261.450.938.023.247.9
w/o sampling66.159.650.639.922.547.2
w/o pos embedding67.260.549.836.322.747.3
(c) Training objective
w/o DIoU loss62.253.443.231.017.441.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.158, + 0.777, + 0.185 + ], + "angle": 0, + "content": "Table 3: Analysis of the maximum length of the memory queue." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.186, + 0.772, + 0.358 + ], + "angle": 0, + "content": "
Memory size0.30.40.50.60.7Average
(a) THUMOS14 dataset
w/o mem65.858.947.636.920.846.0
167.259.950.737.922.747.7
367.361.951.937.922.648.3
770.362.752.138.623.749.5
1167.960.649.837.823.247.9
1566.960.150.738.424.148.0
1967.060.152.540.224.649.1
(b) MUSES dataset
w/o mem22.317.613.19.14.613.3
123.018.013.58.64.913.6
323.118.313.98.95.113.9
723.518.113.58.94.913.8
1122.918.714.09.45.414.1
1523.519.314.39.45.714.4
1922.718.913.89.65.514.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.372, + 0.762, + 0.386 + ], + "angle": 0, + "content": "Table 4: Comparison between different memory modules on the THUMOS14." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.398, + 0.769, + 0.448 + ], + "angle": 0, + "content": "
Memory moduleBackboneSegment sizeInference timefpsMemory parametersAverage mAP
MAT [33]TSN [34]64191.9ms5.240.1M46.9
E2E-LOAD [2]TSN64196.1ms5.153.1M47.9
MATRTSN64167.1ms6.024.0M49.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.537 + ], + "angle": 0, + "content": "show that all design choices of the proposed components are important for action instance localization. Lastly, training our model without DIoU loss shows a drastic performance drop, which demonstrates the importance of instance-level supervision for On-TAL (Table 2(c))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Impact of the Memory Queue Size. We conduct an ablation study on the different memory queue sizes. As shown in Table 3, employing the memory queue enhances the performance compared to the without memory case, and generally the performance improves as the memory queue size increases. With an adequately large memory queue size covering the top \\(99\\%\\) of the instances in the training split, such as 7 for the THUMOS14 and 5 for the MUSES, the model shows robust performance and the performance is higher than the previous state-of-the-art model [15]. The memory queue size of 7 and 15 shows the best on THUMOS14 and MUSES, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Comparison of the proposed memory queue with OAD methods using memory. We compare our memory module with previous memory modules from OAD tasks [2, 33] by replacing the memory-augmented video encoder (Sec. 3.2) of our model. For a fair comparison, hidden dimension of the attention blocks for all methods is set to 1024. As shown in Table 4, MATR outperforms the existing memory modules in terms of both mAP and space-time complexity. While existing memory modules use multiple attention blocks, our memory module consists of a segment encoder and a flag prediction head, making it superior in terms of parameters and inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Analysis of the Segment Size. As shown in Fig. 4, our model is less sensitive to the input segment size, which should be carefully tuned across various" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.156, + 0.444, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.297, + 0.449, + 0.366 + ], + "angle": 0, + "content": "Fig.4: Average mAP \\((\\%)\\) versus the segment size in THUMOS14. Red line represents the result of our model and blue line represents the result of OAT-OSN." + }, + { + "type": "table_caption", + "bbox": [ + 0.459, + 0.154, + 0.788, + 0.195 + ], + "angle": 0, + "content": "Table 5: Ablation on the start and the end prediction heads. Reg and CIs refer to the offset regression and the region classification respectively." + }, + { + "type": "table", + "bbox": [ + 0.468, + 0.196, + 0.777, + 0.244 + ], + "angle": 0, + "content": "
StartEnd0.30.40.50.60.7Average
RegReg64.259.050.037.822.746.7
Reg + ClsReg70.362.752.138.623.749.5
Reg + ClsReg + Cls64.058.950.238.323.246.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.459, + 0.263, + 0.786, + 0.289 + ], + "angle": 0, + "content": "Table 6: Ablation on the memory compression factor on the THUMOS14 dataset." + }, + { + "type": "table", + "bbox": [ + 0.465, + 0.291, + 0.782, + 0.364 + ], + "angle": 0, + "content": "
Compress factor0.30.40.50.60.7Average
None (Ours)70.362.752.138.623.749.5
267.359.949.436.922.347.7
468.161.552.739.724.549.3
867.561.352.038.823.848.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.224, + 0.384, + 0.777, + 0.399 + ], + "angle": 0, + "content": "Table 7: Comparison of the inference time and # parameters on the THUMOS14." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.41, + 0.751, + 0.46 + ], + "angle": 0, + "content": "
MethodBackboneSegment sizeInference timefps# ParametersAverage mAP
OAT-OSN [15]TSN [34]64163.7ms6.1128.7M44.6
MATRTSN64167.1ms6.0192.8M49.5
MATRTSN1653.8ms18.6192.8M46.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.787, + 0.534 + ], + "angle": 0, + "content": "datasets. While the performance of OAT-OSN drops significantly from \\(44.6\\mathrm{mAP}\\) to \\(25.8\\mathrm{mAP}\\) when the segment size is reduced from 64 to 8, the performance drop of ours is only \\(9.1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Analysis of Start and End Prediction Heads. We investigate various offset prediction methods for the start prediction and the end prediction (Table 5). For the first method, start offset regression without region classification, the offset is normalized using the offset statistics from the training dataset. The second approach, which adopts region classification and offset regression for the start prediction, and only offset regression for the end prediction, exhibits the best performance. The reason for performance drop when using region classification for the end prediction is that frame-wise end classification is a complex problem, adversely affects the performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Analysis of Memory Compression Factor. We also conduct an experiment on efficiently storing input segment features in the memory by compressing them. We utilize a 1D convolution layer to compress the number of segment features. As shown in Table 6, \\(4 \\times\\) compression yields competitive results with efficient memory usage. However, we decide not to use this method due to the efficient utilization of memory already achieved through the flag token and avoid to use additional hyperparameters that require optimization for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Inference Time. As shown in Table 7, when using the same backbone and the segment size, the inference time of MATR is almost equal to that of OAT-OSN, yet its average mAP is significantly better. Moreover, by reducing the segment length to 16, MATR outperforms OAT-OSN in terms of both mAP and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.371 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of MATR and OAT-OSN [15] on the THUMOS14. Generated time is the timestamp when the predicted instance is generated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.431 + ], + "angle": 0, + "content": "the inference time. MATR uses more parameters due to the separation of the decoder into the start decoder and the end decoders." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.422, + 0.47 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.789, + 0.557 + ], + "angle": 0, + "content": "Fig. 5 illustrates the predictions of our model and OAT-OSN [15]. The results show that MATR has the capability to efficiently detect action instances by identifying intervals immediately after the current action ends. Additionally, the results show that our model is able to classify and localize action instances more precisely than OAT-OSN." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.36, + 0.597 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.719 + ], + "angle": 0, + "content": "In this paper, we address online temporal action localization by introducing a new model, MATR, which leverages the memory queue to exploit long-term context. We also propose a new action instance localization method based on Transformer, which separately identifies action starts and action ends. As a result, MATR not only surpasses existing online TAL methods but also demonstrates comparable performance to offline TAL methods. The code base will be open to the public to promote future research on On-TAL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Limitations: When the memory queue contains multiple action instances, there is a risk of action instances being matched with incorrect start points. Additionally, when storing the input segment into the memory queue, MATR only considers whether the input segment is relevant to action instances but does not leverage the relationship with the past context stored in the memory. To address these issues, investigating a method to utilize the information stored in the memory queue when storing the input segment features, could lead to more effective memory storage and utilization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the NRF grant and the IITP grant funded by Ministry of Science and ICT, Korea (RS-2019-II191906, RS-2021-II212068, RS-2022-II220290, NRF-2021R1A2C3012728)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.213, + 0.324, + 0.229 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.281 + ], + "angle": 0, + "content": "1. Buch, S., Escorcia, V., Shen, C., Ghanem, B., Carlos Niebles, J.: Sst: Single-stream temporal action proposals. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2911-2920 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. Proc. IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.785, + 0.363 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.404 + ], + "angle": 0, + "content": "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.444 + ], + "angle": 0, + "content": "5. Chen, P., Gan, C., Shen, G., Huang, W., Zeng, R., Tan, M.: Relation attention for temporal action localization. IEEE Transactions on Multimedia 22(10), 2723-2733 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.485 + ], + "angle": 0, + "content": "6. Cheng, F., Bertasius, G.: Tallformer: Temporal action localization with a long-memory transformer. In: Proc. European Conference on Computer Vision (ECCV). pp. 503-521. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.526 + ], + "angle": 0, + "content": "7. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Proc. European Conference on Computer Vision (ECCV). pp. 269-284. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.594 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. International Conference on Learning Representations (ICLR) (2021), https://openreview.net/forum?id=YicbFdNTTy" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.785, + 0.636 + ], + "angle": 0, + "content": "9. Escorcia, V., Caba Heilbron, F., Niebles, J.C., Ghanem, B.: Daps: Deep action proposals for action understanding. In: Proc. European Conference on Computer Vision (ECCV). pp. 768-784. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.676 + ], + "angle": 0, + "content": "0. Gao, J., Yang, Z., Chen, K., Sun, C., Nevatia, R.: Turn tap: Temporal unit regression network for temporal action proposals. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3628-3636 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "1. Gao, M., Xu, M., Davis, L.S., Socher, R., Xiong, C.: Startnet: Online detection of action start in untrimmed videos. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5542-5551 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. Heilbron, F.C., Niebles, J.C., Ghanem, B.: Fast temporal activity proposals for efficient detection of human actions in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1914-1923 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Jiang, Y.G., Liu, J., Roshan Zamir, A., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/ (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Kang, H., Kim, K., Ko, Y., Kim, S.J.: Cag-qil: Context-aware actionness grouping via q imitation learning for online temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13729-13738 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Proc. European Conference on Computer Vision (ECCV). pp. 653-669. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proc. International Conference on Learning Representations (ICLR) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.272 + ], + "angle": 0, + "content": "17. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3320-3329 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "18. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3889-3898 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "19. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: Proc. European Conference on Computer Vision (ECCV). pp. 3-19 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "21. Liu, Q., Wang, Z.: Progressive boundary refinement network for temporal action detection. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 11612-11619 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "22. Liu, X., Hu, Y., Bai, S., Ding, F., Bai, X., Torr, P.H.: Multi-shot temporal event localization: a benchmark. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12596-12606 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "23. Liu, X., Wang, Q., Hu, Y., Tang, X., Zhang, S., Bai, S., Bai, X.: End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing 31, 5427-5441 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "24. Liu, Y., Ma, L., Zhang, Y., Liu, W., Chang, S.F.: Multi-granularity generator for temporal action proposal. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3604-3613 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "25. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Gaussian temporal awareness networks for action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 344-353 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "27. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18857-18866 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.73 + ], + "angle": 0, + "content": "28. Shou, Z., Chan, J., Zareian, A., Miyazawa, K., Chang, S.F.: Cdc: Convolutional-de-convolutional networks for precise temporal action localization in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5734-5743 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "29. Shou, Z., Pan, J., Chan, J., Miyazawa, K., Mansour, H., Vetro, A., Giro-i Nieto, X., Chang, S.F.: Online detection of action start in untrimmed, streaming videos. In: Proc. European Conference on Computer Vision (ECCV). pp. 534-551 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "30. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1049-1058 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Tang, T.N., Park, J., Kim, K., Sohn, K.: Simon: A simple framework for online temporal action localization. arXiv preprint arXiv:2211.04905 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.223, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. Advances in Neural Information Processing Systems (NeurIPS). pp. 5998-6008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "33. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13824-13835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.275 + ], + "angle": 0, + "content": "34. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proc. European Conference on Computer Vision (ECCV). pp. 20-36. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.277, + 0.785, + 0.317 + ], + "angle": 0, + "content": "35. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7565-7575 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.373 + ], + "angle": 0, + "content": "36. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13587-13597 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.416 + ], + "angle": 0, + "content": "37. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10186-10195 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.459 + ], + "angle": 0, + "content": "38. Xu, M., Perez Rua, J.M., Zhu, X., Ghanem, B., Martinez, B.: Low-fidelity video encoder optimization for temporal action localization. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 9923-9935 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "39. Xu, M., Zhao, C., Rojas, D.S., Thabet, A., Ghanem, B.: G-tad: Sub-graph localization for temporal action detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10156-10165 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "40. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5532-5541 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.585 + ], + "angle": 0, + "content": "41. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long shortterm transformer for online action detection. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 1086-1099 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "42. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3160-3169 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.67 + ], + "angle": 0, + "content": "43. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7094-7103 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.712 + ], + "angle": 0, + "content": "44. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 492-510. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "45. Zhao, C., Thabet, A.K., Ghanem, B.: Video self-stitching graph network for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13658-13667 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "46. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 485-502. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2914-2923 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., Ren, D.: Distance-iou loss: Faster and better learning for bounding box regression. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 12993-13000 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "49. Zhu, Z., Tang, W., Wang, L., Zheng, N., Hua, G.: Enriching local and global contexts for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13516-13525 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_origin.pdf b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dbb2190eb0496408143d511253f6f61283cbf358 --- /dev/null +++ b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/37bd7329-2735-4908-b7bb-a4a18c449d84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6fe0b5c6fa58348ecb5cf3219ab99a3a44649d5a0c04dbaa3f19afcd2a311a2 +size 1173927 diff --git a/2024/Online Temporal Action Localization with Memory-Augmented Transformer/full.md b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3d428615a08b3bff3ff81c791713ff775aba4208 --- /dev/null +++ b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/full.md @@ -0,0 +1,303 @@ +# Online Temporal Action Localization with Memory-Augmented Transformer + +Youngkil Song\* + +Dongkeun Kim* + +Minsu Cho + +Suha Kwak + +Pohang University of Science and Technology (POSTECH), South Korea {songyk, kdk1563, mscho, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/MATR/ + +Abstract. Online temporal action localization (On-TAL) is the task of identifying multiple action instances given a streaming video. Since existing methods take as input only a video segment of fixed size per iteration, they are limited in considering long-term context and require tuning the segment size carefully. To overcome these limitations, we propose memory-augmented transformer (MATR). MATR utilizes the memory queue that selectively preserves the past segment features, allowing to leverage long-term context for inference. We also propose a novel action localization method that observes the current input segment to predict the end time of the ongoing action and accesses the memory queue to estimate the start time of the action. Our method outperformed existing methods on two datasets, THUMOS14 and MUSES, surpassing not only TAL methods in the online setting but also some offline TAL methods. + +Keywords: Temporal action localization $\cdot$ Online video understanding + +# 1 Introduction + +Today, video is one of the most popular types of media, with tons of videos being created and uploaded to online platforms like YouTube and TikTok every second. As a result, the demand for video understanding is growing rapidly. In particular, dealing with untrimmed videos is a key to practical video understanding since most videos in the real world are not trimmed for individual events in advance. Temporal action localization (TAL) has been extensively studied in this context [10, 17, 21, 30, 38, 39, 43-45, 47, 49], with the aim of detecting each action instance in the input untrimmed video by predicting its start time, end time, and action class. + +Recently, online TAL (On-TAL), i.e., TAL for streaming video [14], has gained increasing attention thanks to its potential applications in video surveillance, sports video analysis, and video summarization. Unlike TAL which demands the entire video as input, On-TAL aims to detect action instances using only input frames observed so far in an online manner. Moreover, once action instances are predicted, it is not allowed to modify the results afterward. + +![](images/da375160703c66cee8a846f459f1e75e56a5aa3c7ec8a63d40229d01134fd1d1.jpg) +Fig. 1: A conceptual illustration of our model. Our model selectively stores information about previously occurred actions in the memory queue from streaming video. When the end of the action is detected in the current input segment, the model retrieves information from the memory queue to locate the start of the action. + +A straightforward solution to this challenging task is to first conduct online action detection (OAD) [7,35], i.e., frame-by-frame action classification for streaming video, and then build action instances by aggregating the frame-wise action predictions. In this sense, earlier methods for On-TAL [14,31] follow this approach by applying OAD and refining per-frame predictions using past observations. However, this OAD-based approach exploits only frame-level supervision for training, which is not optimal because On-TAL aims to predict action instances in the form of time intervals. To alleviate this issue, online anchor transformer (OAT) has been recently proposed [15], which adopts sliding window approach and anchor-based method to utilize instance-level supervision. Although OAT improved performance remarkably, however, it still has several drawbacks: Since it takes a fixed-size segment of input video at each iteration and does not consider long-term contexts, its capability is limited in detecting long action instances, and its performance is sensitive to hyperparameters like the size of the input segment. + +To resolve the above issues, we propose a new end-to-end architecture for OntAL, dubbed memory-augmented Transformer (MATR). At the heart of MATR lies in memory queue, which selectively stores past segment features so that the model exploits long-term context for inference. Thanks to the memory queue that contains long-term context and reduces dependence on the segment size, our model accurately localizes even long-term action instances without the need for careful hyperparameter tuning for each dataset. + +Moreover, we propose a novel approach to action instance localization using the memory queue as illustrated in Fig. 1. First, it detects the end of an action using the current segment features and then scans past segment features in the memory queue to find the action start corresponding to the detected action end. To achieve this, we adopt two Transformer decoders: one for end detection and the other for start detection. Specifically, we adopt learnable queries as input to + +the decoders so that each of them learns to localize action boundaries through the attention mechanism of Transformer [8,32]. In addition, motivated by recent advances on object detection [6,37], MATR decouples action classification and localization to further improve the performance. This is realized in MATR by using distinct queries for the two subtasks. + +The proposed method is evaluated on two datasets, THUMOS14 [13] and MUSES [22]. Our model achieves the state of the art on both datasets in the online TAL setting. In addition, surprisingly, it is also comparable to existing offline TAL methods although it does not employ offline non-maximum suppression (NMS). In summary, the contribution of this paper is three-fold: + +- We propose MATR, a new end-to-end architecture for On-TAL, leveraging memory queue to utilize long-term context and enabling precise localization of action instances with less dependence on dataset-specific hyperparameter. +- We introduce a new action instance localization method that first identifies action end using the input segments and scans past information in the memory queue to find the action start. Moreover, our method adopts distinct queries to separate information for action classification and localization. +- Our method outperformed the existing On-TAL methods on the two benchmarks. Also, extensive ablation studies demonstrate the contribution of each component of the proposed method. + +# 2 Related Work + +Temporal Action Localization. Temporal action localization (TAL) [5,10,17, 21,23,30,38,39,43-45,47,49] which identifies action instances in time and classifying their actions, has given much attention on the video research community. Earlier attempts employ two-stage approaches that first generate action proposals, then classify action class and refine action boundaries [1,9,12,18,19,24,47]. Recent studies have primarily dealt with widening the range of video context through convolutional neural networks [28], graph neural networks [39,43], Gaussian kernel [25], and utilizing global context [49]. To extend temporal context and enable end-to-end TAL, Transformer [8,32] has recently been adopted [6,23,44]. These approaches incorporate temporal deformable attention module [23] and multi-scale features [44] to improve computational efficiency and performance. + +Online Video Understanding. Online action detection (OAD), which classifies action classes in a frame-wise manner, is first proposed by Geest et al. [7]. To incorporate long-term context, recurrent neural network [40] and Transformer [35] are utilized. These methods predict actions for future frames and use this information to refine current frame predictions. Yang et al. [42] further improve the performance by using category exemplars and capturing long-term dependencies within a video segment. On the other hand, Shou et al. [29] proposes online detection of action start (ODAS), which aims to detect action start as early as possible. Gao et al. [11] further improves ODAS by adopting reinforcement learning. + +Online Temporal Action Localization. Towards practical event understanding, online temporal action localization (On-TAL) has been proposed recently. Kang et al. [14] first propose On-TAL, which localizes action boundaries and classifies action classes without accessing future frames and offline post-processing. Earlier methods [14,31] adopt OAD framework to classify per-frame action class and group them into action instances to solve On-TAL. Specifically, Kang et al. [14] utilize Markov decision process to aware prior contexts to group per-frame predictions. SimOn [31] is proposed as an end-to-end On-TAL method that leverages past visual context and action probabilities for accurate action prediction at the current frame. Kim et al. [15] propose a sliding window scheme and an anchor-based method that adopts Transformer [32] for On-TAL. However, it only uses a fixed input segment which is highly dependent on the action length distribution of each dataset. Unlike the previous methods, we utilize a memory queue for long-term context, which also reduces the dependence on the input segment size. + +Difference between OAD and On-TAL. OAD [2,33,35,40-42,46] focus on predicting action classes at the frame-level, which limits the ability to distinguish boundaries of overlapping actions. On the other hand, On-TAL [14,15,31] predicts entire action instances rather than individual frames, enabling it to precisely identify the start and end times of actions, making it effective for applications requiring instance-level action understanding, like sports video analysis. Using Memory for Video Understanding. Prior work in video understanding employs memory to store contextual information, in the forms of long-term and short-term memory [2,33,36,41]: short-term memory captures fine details, and long-term memory stores compressed information of past contexts. Specifically, Wang et al. [33] divides long-term memory into several groups and compresses each group into a single vector. Stream Buffer [2] gradually reduces temporal dimensions of the memory using a compression module. However, unlike video classification tasks such as action recognition [36] and OAD [2,33,41,46], which requires to store representative information of past frames in the memory for classifying the action of the current frame, TAL needs to preserve the temporal positional information of past frames in the memory to accurately predict the exact time. In that sense, MATR preserves temporal information in memory to accurately localize actions and effectively stores information by using flag tokens. A comparison with the OAD memory modules is shown in Table 4. + +# 3 Proposed Method + +Consider an untrimmed video $\mathbf{V} = \{\mathbf{v}_i\}_{i=1}^T$ with $T$ frames and $M$ action instances $\varPsi=\{(s_m,e_m,c_m)\}_{m=1}^{M}$ , where each instance is represented by its start time $s_m$ , end time $e_m$ and action class $c_m$ . Unlike temporal action localization (TAL), which allows a model to use whole video sequence as input, online temporal action localization (On-TAL) forces a model to predict action instances using only input frames seen until the current timestamp. Note that action instances predicted at previous iterations cannot be modified or deleted afterwards. + +![](images/7a6d7ed0afe60bf9567f5d531cf41a14e4f748475a10ec1f6ed7e1cbb0e37d60.jpg) +Fig. 2: Overall architecture of MATR. MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads. + +As shown in Fig. 2, MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads. In line with previous research, the current segment of the input streaming video is given as a unit input, and its frame-level features, referred to as segment features, are extracted by a video backbone network and a linear projection layer (Sec. 3.1). The segment features are then fed to the memory-augmented video encoder, which encodes temporal context between frames in the current segment and stores the segment features into the memory (Sec. 3.2). The instance decoding module localizes action instances via two Transformer decoders: the end decoder and the start decoder. Specifically, the end decoder references the encoded segment features to locate the action end around the current time, and then the start decoder refers to the memory queue to find the action start based on the past information stored in the memory queue. Queries for each instance consist of a class query for action classification and a boundary query for action localization (Sec. 3.3). The outputs of the instance decoding module are used as inputs to the prediction heads, which consist of end prediction head, start prediction head, and action classification head (Sec. 3.4). The entire model is trained in an end-to-end manner (Sec. 3.5). + +# 3.1 Feature Extractor + +Given a segment comprising $L_{s}$ consecutive video frames as input, a pretrained backbone [4, 34] followed by a linear projection layer extracts the segment features $\mathbf{X}_t = \{\mathbf{x}_i\}_{t - L_s + 1}^t\in \mathbb{R}^{L_s\times D}$ for each frame within the segment, following previous work [14, 15]. Inspired by Kim et al. [15], MATR employs a sliding window scheme, which predicts multiple action instances by moving frame by frame along the temporal axis. Note that the window size is the input segment size. + +# 3.2 Memory-Augmented Video Encoder + +At each time step, the input segment features are fed to two modules of the memory-augmented video encoder, the segment encoder and the memory update module. The segment encoder encodes the temporal context across input segment features, while the memory update module selectively stores the segment features and updates the memory queue. + +Segment Encoder. The segment encoder is a standard Transformer encoder composed of self-attention layers and a feed forward network (FFN). The combination of segment features and a learnable flag token is fed into the encoder, and transformed to queries, keys, and values for self-attention. Sinusoidal positional encoding $\mathbf{S}_{\mathrm{pos}} \in \mathbb{R}^{(L_s + 1) \times D}$ is added to the queries and keys. Among the output embeddings, the encoded flag token is fed into the memory update module and the encoded segment features that integrate temporal context across input frames are used as input to the end decoder. + +Memory Update Module. In an online setting, where an input is a streaming video, efficiently storing information of past frames and accessing it effectively is crucial for detecting action instances, in particular those of long-term actions whose temporal extents are beyond the size of input segments. To this end, we adopt the memory queue for storing information of past input segments in a first-in-first-out (FIFO) manner. Additionally, we propose an efficient memory update method employing the flag token. When the input segment features pass through the segment encoder, the flag token is concatenated into the input segment features and then employed in the flag prediction head. It is trained to predict [FLAG], which identifies whether an input segment is relevant to action instances, so that only relevant segments are stored in the memory. [FLAG] = 1 when there are overlapping frames between the input segment and action instances, and 0 otherwise. During training, the ground-truth [FLAG] is utilized. During inference, [FLAG] = 1 if sigmoid(\hat{g}) > \theta where $\hat{g}$ is the output logit of the flag prediction head and $\theta$ is a predefined threshold, and 0 otherwise. The input segment feature is added to the memory queue when [FLAG] = 1 and discards it when [FLAG] = 0. If the memory queue is full, the oldest segment feature is purged. + +# 3.3 Instance Decoding Module + +The instance decoding module localizes and classifies action instances by leveraging the encoded segment features and the memory queue through the attention mechanism of Transformer. Given the encoded segment features that have short-term temporal contexts across input segment and the memory queue that stores long-term contexts for actions that are potentially ongoing, a set of $2N$ instance queries $\mathbf{Q} = \{\mathbf{Q}_{\mathrm{class}}; \mathbf{Q}_{\mathrm{bound}}\} \in \mathbb{R}^{2N \times D}$ is trained to generate $N$ action instances for each input segment. Half of these queries pertain to predicting the action class (class query; $\mathbf{Q}_{\mathrm{class}} \in \mathbb{R}^{N \times D}$ ), while the other half is used for predicting the start and end timestamps (boundary query; $\mathbf{Q}_{\mathrm{bound}} \in \mathbb{R}^{N \times D}$ ). The class and boundary query pairs for the same instance share the same positional embedding $\mathbf{E}_{\mathrm{pos}} \in \mathbb{R}^{N \times D}$ to accurately identify and differentiate between + +![](images/57feb6d1ef320ad3bda0e3452736157218e717846a8a0095b857edcbe7be4792.jpg) +Fig. 3: Detailed architecture of the instance decoding module. + +instances; $\tilde{\mathbf{Q}} = \{\mathbf{Q}_{\mathrm{class}} + \mathbf{E}_{\mathrm{pos}};\mathbf{Q}_{\mathrm{bound}} + \mathbf{E}_{\mathrm{pos}}\}$ . Unlike the previous Transformer-based On-TAL method [15] that predicts start and end offset for each instance simultaneously, our method employs separate Transformer decoders for predicting start offset and end offset. As shown in Fig. 3, the start decoder and end decoders share the same architecture but utilizes different information to predict action start and action end, respectively. + +End Decoder. Given the encoded segment features, instance queries $\mathbf{Q}$ is trained to generate the output embeddings used for detecting action end near the current timestamp. To this end, we adopt a Transformer [3] composed of multi-head self-attention layers, multi-head cross-attention layers, and FFNs. The detailed architecture is illustrated in Fig. 3. The outputs of the end decoder, which consist of end boundary embeddings and class embeddings, are then used as inputs to the start decoder and the prediction heads. + +Start Decoder. The start decoder takes the output embeddings from the end decoder and locates the corresponding action start by utilizing the memory queue. First, the memory queue is concatenated with the current segment features. The concatenated memory features are used as long-term context in the start decoder. Then, 2D temporal positional encodings are added to the memory features, and the results are used as the key and value for the cross-attention layers of the start decoder. The overall architecture of the start decoder is the same as the end decoder (Fig. 3), but it uses the memory features rather than the encoded segment features. The output embeddings of the start decoder, start boundary embeddings and class embeddings, are used as inputs to the start prediction and the action classification head, respectively. + +Moreover, we introduce two techniques to efficiently and effectively utilize the memory queue. First, we conduct a $50\%$ uniform sampling to the memory features which enables efficient use of memory since adjacent frames contain similar information. Second, to expand the scope of positional encoding for streaming video, whose duration is unpredictable, we separate the temporal positional encoding into the two parts: relative segment position and relative frame position. A relative segment position is the position of a segment within the memory rela + +tive to the current segment. Meanwhile, a relative frame position is the position of a frame relative to the most recent frame within the same segment. + +# 3.4 Prediction Heads + +To generate $N$ action instances $\{(\hat{s}_i,\hat{e}_i,\hat{c}_i)\}_{i = 1}^N$ from the output of the instance decoding module, we use three prediction heads, which are end prediction head, start prediction head, and action classification head. Each of the prediction heads is composed of a 2-layer FFN. + +End Prediction Head. The end prediction head estimates the offset between the end of the target action and the current timestamp. The boundary embeddings from the end decoder are fed to the 2-layer FFN, which in turn estimates the end offset $\{\hat{u}_i\}_{i=1}^N \in \mathbb{R}^N$ . + +Start Prediction Head. The start prediction head estimates the offset between the start of the target action and the current timestamp. Unlike the end prediction, which finds the end offset around the input segment, the start prediction needs to perform offset regression over a relatively wider range. To narrow down the scope of the start offset regression, we divide the time horizon into $L_{m} + 2$ regions: the region preceding the coverage of memory, $L_{m}$ regions covered by the memory, and the region corresponding to the current input segment. In that sense, the start prediction head consists of the region classification and the offset regression, each composed of a 2-layer FFN. They take the boundary embeddings from the start decoder as input to predict the start time. The region classification identifies the regions to which the start time is assigned, represented as $\{\hat{o}_i|\hat{o}_i = \mathrm{argmax}(\hat{r}_i)\}_{i=1}^N \in \mathbb{R}^N$ . Here, $\{\hat{r}_i\}_{i=1}^N \in \mathbb{R}^{N \times (L_m + 2)}$ are the output logits from the region classification head. Following this, the offset regression head is utilized to estimate the offset within the regions, denoted as $\{\hat{v}_i\}_{i=1}^N \in \mathbb{R}^{N \times (L_m + 2)}$ . Note that the start offset $\hat{v}_i$ is calculated for all $L_{m} + 2$ regions, and the offset from the identified region $\{\hat{v}_i(\hat{o}_i)\}_{i=1}^N \in \mathbb{R}^N$ is used in inference. By combining start classification result $\hat{o}_i$ and start regression result $\hat{v}_i(\hat{o}_i)$ , the start time is predicted. The effectiveness of the proposed start prediction head is validated in Table 5. + +Action Classification Head. For action class prediction, we utilize the class embeddings from both the end decoder and the start decoder. These embeddings are concatenated and fed into the action classification head to derive class probabilities $\{\hat{p}_i\}_{i=1}^N \in \mathbb{R}^{N \times (C+1)}$ , where $C$ is the number of action class. + +Action Instance Prediction. Finally, our model makes $N$ action proposals $\{(\hat{s}_i,\hat{e}_i,\hat{c}_i)\}_{i = 1}^N$ at time $t$ by + +$$ +\hat {s} _ {i} = t - \left(\hat {o} _ {i} + \hat {v} _ {i} \left(\hat {o} _ {i}\right)\right) \times L _ {s}, \tag {1} +$$ + +$$ +\hat {e} _ {i} = t + \hat {u} _ {i} \times L _ {s}, \tag {2} +$$ + +$$ +\hat {c} _ {i} = \operatorname {a r g m a x} \left(\hat {p} _ {i}\right), \tag {3} +$$ + +where $L_{s}$ is the segment length. Since the sliding window scheme generates $N$ action proposals at each timestamp, post-processing is crucial for performance improvement by removing redundant or overlapped action instances. + +Non-maximum suppression (NMS) is applied to the action proposals at each timestamp, subsequently removing those highly overlapped with the proposals generated from the past. To prevent more reliable prediction made in the future from being removed, instances where the predicted end time is beyond the current timestamp, $t < \hat{e}_i$ , are also removed. + +# 3.5 Training Objective + +At timestamp $t$ , the model is trained to detect action instances whose end times are in the range $[t - T_d + 1, t + T_a]$ , where $T_d$ and $T_a$ are hyperparameters. Then, hungarian algorithm matches the action proposals and the ground-truth action instances with the lowest matching cost. The matching cost of the ground-truth group $i$ and the proposal $\sigma(i)$ is given by: + +$$ +C _ {i, \sigma (i)} = \hat {p} _ {\sigma (i)} \left(y _ {i}\right) + \operatorname {I o U} \left(b _ {i}, \hat {b} _ {\sigma (i)}\right), \tag {4} +$$ + +where $\sigma$ is the permutation of $N$ action proposals, $\hat{p}_i$ is the class probabilities of the $i$ -th proposal, $b_i = \{s_i, e_i\}$ is the ground-truth action boundary, and $\hat{b}_i = \{\hat{s}_i, \hat{e}_i\}$ is the predicted action boundary. We adopt the focal loss [20] for the action classification, the cross entropy loss for the start region classification, the $\ell_1$ loss for both the start offset regression and the end offset regression. The action classification loss $L_{\mathrm{class}}$ , the start prediction loss $L_{\mathrm{start}}$ , and the end prediction loss $L_{\mathrm{end}}$ are defined as follow: + +$$ +L _ {\text {c l a s s}} = \sum_ {i = 1} ^ {N} \operatorname {F L} \left(\hat {p} _ {i}, y _ {i}\right), \tag {5} +$$ + +$$ +L _ {\text {s t a r t}} = \sum_ {i = 1} ^ {N _ {\text {m a t c h}}} \left\{\mathrm {C E} \left(\operatorname {S o f t m a x} \left(\hat {r} _ {i}\right), r _ {i}\right) + \left| \hat {v} _ {i} \left(\hat {o} _ {i}\right) - v _ {i} \right| \right\}, \tag {6} +$$ + +$$ +L _ {\text {e n d}} = \sum_ {i = 1} ^ {N _ {\text {m a t c h}}} | \hat {u} _ {i} - u _ {i} |, \tag {7} +$$ + +where $y_{i}, r_{i}, v_{i}, u_{i}$ is the ground-truth action class, start region, start offset, and end offset, respectively. + +To provide instance-level supervision and facilitate the connection between start and end time prediction, we employ the DIoU loss [48] following Action-Former [44]: + +$$ +L _ {\mathrm {d i o u}} = \sum_ {i = 1} ^ {N _ {\mathrm {m a t c h}}} 1 - \operatorname {I o U} (\hat {b} _ {i}, b _ {i}) + \frac {\rho^ {2} (\hat {c} _ {i} , c _ {i})}{d _ {i} ^ {2}}, \tag {8} +$$ + +where $\rho(\cdot, \cdot)$ is the Euclidean distance between two points, $\hat{c}_i$ and $c_i$ are the center of the proposal and ground-truth instance, and $d_i$ is the smallest enclosing 1-dim box length. + +To train the memory update module in an end-to-end manner, we employ the flag loss for training the flag prediction head: + +$$ +L _ {\text {f l a g}} = \operatorname {B C E} (\operatorname {S i g m o i d} (\hat {g}), g), \tag {9} +$$ + +where $\hat{g}$ and $g$ is the predicted logit of flag token and ground-truth [FLAG] respectively. + +Our model is trained with six losses simultaneously in an end-to-end manner. The total loss is calculated as follows: + +$$ +L = L _ {\text {c l a s s}} + L _ {\text {s t a r t}} + L _ {\text {e n d}} + L _ {\text {d i o u}} + L _ {\text {f l a g}}. \tag {10} +$$ + +Note that all loss coefficients are set to 1 and loss balancing is not required. + +# 4 Experiments + +# 4.1 Experimental Setting + +Datasets. We evaluate our method and previous methods on two On-TAL benchmarks: THUMOS14 [13] and MUSES [22]. THUMOS14 contains 200 videos for training and 213 videos for testing with 20 action classes, while MUSES has 2,587 videos for training and 1,110 videos for testing with 25 action classes. MUSES consists of multi-shot action instances, which makes action localization more challenging. + +Hyperparameters. We use a frozen two-stream TSN [34] pretrained on Kinetics [4] to extract RGB and flow features for THUMOS14, and an I3D [4] trained on Kinetics [4] utilizing only RGB features for MUSES, following previous OntAL work [14, 15, 31]. The segment size $L_{s}$ is set to 64 in THUMOS14 and 75 in MUSES. The dimension of the segment feature is set to $D = 1024$ for both datasets. For the memory-augmented video encoder, we stack 3 Transformer layers with 8 attention heads for both datasets. The flag threshold $\theta$ is set to 0.5. The memory size $L_{m}$ is set to 7 for THUMOS14, while 15 for MUSES. For the instance decoding module, we stack 5 Transformer layers with 4 attention heads for both datasets. The number of the class query and the boundary query $N = 10$ for THUMOS14 and $N = 6$ for MUSES. + +Training. During training, we use Adam optimizer [16], with the initial learning rate of 1e-8 with CosineAnnealing scheduler [26]. The batch size is set to 64 for THUMOS14, while 75 for MUSES. The focal loss coefficient is $\alpha = 0.25$ and $\gamma = 2$ for THUMOS14 and $\alpha = 1$ and $\gamma = 5$ for MUSES. All loss coefficients are set to 1. More details are listed in the supplementary material (Sec. D). + +# 4.2 Comparison with the State of the Art + +We compare our method with previous online and offline TAL methods on THUMOS14 and MUSES dataset. As presented in Table 1, our method outperforms all previous On-TAL methods in both benchmarks by a substantial margin: $4.9\%$ of average mAP in THUMOS14 and $0.7\%$ of average mAP in MUSES. The results show that MATR utilizing the memory queue is more effective than previous OAD-based methods [14,31,33,46] and instance-level method OAT-OSN [15]. We also evaluate two OAD methods utilizing memory [33,46] in On-TAL setting by grouping their frame-level predictions into action instances. + +Table 1: Comparison with On-TAL methods on THUMOS14 and MUSES dataset. The results are reported in mAP measure (%). **' indicates that the result is validated using the On-TAL ground truth from classification annotation. + +
MethodTHUMOS14MUSES
backbone0.30.40.50.60.7Averagebackbone0.30.40.50.60.7
Offline TAL
G-TAD [39]TSN [34]54.547.640.230.823.439.9I3D [4]19.114.811.17.44.7
ContextLoc [49]I3D68.363.854.341.826.250.9------
P-GCN [43]I3D63.657.849.1---I3D19.917.113.19.75.4
MUSES [22]I3D68.964.056.946.331.053.4I3D25.922.618.915.010.6
ActionFormer [44]I3D82.177.871.059.443.966.8------
TriDet [27]I3D83.680.172.962.447.469.3------
OAD-based Online TAL
TeSTra [46]TSN35.629.221.413.47.621.4------
MAT [33]I3D36.027.519.212.36.420.3------
CAG-QIL [14]TSN44.737.629.821.914.529.7I3D8.56.54.22.81.9
SimOn [31]TSN54.345.035.023.314.634.4------
SimOn* [31]TSN57.047.537.326.616.036.9------
Instance-level Online TAL
OAT-Naive [15]TSN57.650.643.030.015.739.4I3D20.316.612.97.73.6
OAT-OSN [15]TSN63.056.747.136.320.044.6I3D22.118.514.28.94.7
MATRTSN70.362.752.138.623.749.5I3D23.519.314.39.45.7
+ +However, their performance is far lower than MATR, which verifies the effectiveness of the proposed memory queue of MATR. We also compare our model with several offline TAL methods. Although there is still a performance gap between online and offline methods, our model achieves comparable performance to offline methods, and even outperforms previous work [39, 43]. The performance on the MUSES dataset is lower compared to THUMOS14 and the reason is twofold: (1) Action instances in MUSES are captured in multi-shots, making it challenging to detect action boundaries, and (2) there are confusing action classes such as "quarrel" and "fight," making action classification hard. These make MUSES challenging and limit the upper-bound of performance on it. + +# 4.3 Ablation Studies + +We perform ablation studies on THUMOS14 and MUSES dataset to show the effectiveness of the proposed modules. + +Ablation on the Proposed Modules. In Table 2, we conduct an ablation study to verify the effectiveness of each component of our model. In Table 2(a), focuses on the components of the memory-augmented video, both removing the flag token and the segment encoder led to a performance drop. The results show the efficacy of selectively storing the past information and the importance of temporal context across the input segment. In Table 2(b), we use a single decoder and predict both start and end at once, which results significant performance drop, from $49.5\mathrm{mAP}$ to $42.7\mathrm{mAP}$ . The result demonstrates the contribution of our new action localization method, using two separate decoders for the start and the end prediction. We also ablate new query design, sampling the memory for the start decoder, and positional embedding for instance queries. The results + +Table 2: Ablation of the proposed modules on THUMOS14 dataset. The results are reported in mAP measure $(\%)$ . + +
Method0.30.40.50.60.7Average
Ours70.362.752.138.623.749.5
(a) Memory-augmented video encoder
w/o flag token67.361.050.136.522.347.4
w/o segment encoder65.859.349.936.321.846.6
(b) Instance decoding module
single decoder65.156.445.030.516.742.7
w/o splitting query66.261.450.938.023.247.9
w/o sampling66.159.650.639.922.547.2
w/o pos embedding67.260.549.836.322.747.3
(c) Training objective
w/o DIoU loss62.253.443.231.017.441.4
+ +Table 3: Analysis of the maximum length of the memory queue. + +
Memory size0.30.40.50.60.7Average
(a) THUMOS14 dataset
w/o mem65.858.947.636.920.846.0
167.259.950.737.922.747.7
367.361.951.937.922.648.3
770.362.752.138.623.749.5
1167.960.649.837.823.247.9
1566.960.150.738.424.148.0
1967.060.152.540.224.649.1
(b) MUSES dataset
w/o mem22.317.613.19.14.613.3
123.018.013.58.64.913.6
323.118.313.98.95.113.9
723.518.113.58.94.913.8
1122.918.714.09.45.414.1
1523.519.314.39.45.714.4
1922.718.913.89.65.514.1
+ +Table 4: Comparison between different memory modules on the THUMOS14. + +
Memory moduleBackboneSegment sizeInference timefpsMemory parametersAverage mAP
MAT [33]TSN [34]64191.9ms5.240.1M46.9
E2E-LOAD [2]TSN64196.1ms5.153.1M47.9
MATRTSN64167.1ms6.024.0M49.5
+ +show that all design choices of the proposed components are important for action instance localization. Lastly, training our model without DIoU loss shows a drastic performance drop, which demonstrates the importance of instance-level supervision for On-TAL (Table 2(c)). + +Impact of the Memory Queue Size. We conduct an ablation study on the different memory queue sizes. As shown in Table 3, employing the memory queue enhances the performance compared to the without memory case, and generally the performance improves as the memory queue size increases. With an adequately large memory queue size covering the top $99\%$ of the instances in the training split, such as 7 for the THUMOS14 and 5 for the MUSES, the model shows robust performance and the performance is higher than the previous state-of-the-art model [15]. The memory queue size of 7 and 15 shows the best on THUMOS14 and MUSES, respectively. + +Comparison of the proposed memory queue with OAD methods using memory. We compare our memory module with previous memory modules from OAD tasks [2, 33] by replacing the memory-augmented video encoder (Sec. 3.2) of our model. For a fair comparison, hidden dimension of the attention blocks for all methods is set to 1024. As shown in Table 4, MATR outperforms the existing memory modules in terms of both mAP and space-time complexity. While existing memory modules use multiple attention blocks, our memory module consists of a segment encoder and a flag prediction head, making it superior in terms of parameters and inference speed. + +Analysis of the Segment Size. As shown in Fig. 4, our model is less sensitive to the input segment size, which should be carefully tuned across various + +![](images/0adb606a4a25c04d57fd640bb28efd7652879108e51a660b08ed545786c666e1.jpg) +Fig.4: Average mAP $(\%)$ versus the segment size in THUMOS14. Red line represents the result of our model and blue line represents the result of OAT-OSN. + +Table 5: Ablation on the start and the end prediction heads. Reg and CIs refer to the offset regression and the region classification respectively. + +
StartEnd0.30.40.50.60.7Average
RegReg64.259.050.037.822.746.7
Reg + ClsReg70.362.752.138.623.749.5
Reg + ClsReg + Cls64.058.950.238.323.246.9
+ +Table 6: Ablation on the memory compression factor on the THUMOS14 dataset. + +
Compress factor0.30.40.50.60.7Average
None (Ours)70.362.752.138.623.749.5
267.359.949.436.922.347.7
468.161.552.739.724.549.3
867.561.352.038.823.848.7
+ +Table 7: Comparison of the inference time and # parameters on the THUMOS14. + +
MethodBackboneSegment sizeInference timefps# ParametersAverage mAP
OAT-OSN [15]TSN [34]64163.7ms6.1128.7M44.6
MATRTSN64167.1ms6.0192.8M49.5
MATRTSN1653.8ms18.6192.8M46.3
+ +datasets. While the performance of OAT-OSN drops significantly from $44.6\mathrm{mAP}$ to $25.8\mathrm{mAP}$ when the segment size is reduced from 64 to 8, the performance drop of ours is only $9.1\%$ . + +Analysis of Start and End Prediction Heads. We investigate various offset prediction methods for the start prediction and the end prediction (Table 5). For the first method, start offset regression without region classification, the offset is normalized using the offset statistics from the training dataset. The second approach, which adopts region classification and offset regression for the start prediction, and only offset regression for the end prediction, exhibits the best performance. The reason for performance drop when using region classification for the end prediction is that frame-wise end classification is a complex problem, adversely affects the performance. + +Analysis of Memory Compression Factor. We also conduct an experiment on efficiently storing input segment features in the memory by compressing them. We utilize a 1D convolution layer to compress the number of segment features. As shown in Table 6, $4 \times$ compression yields competitive results with efficient memory usage. However, we decide not to use this method due to the efficient utilization of memory already achieved through the flag token and avoid to use additional hyperparameters that require optimization for each dataset. + +Inference Time. As shown in Table 7, when using the same backbone and the segment size, the inference time of MATR is almost equal to that of OAT-OSN, yet its average mAP is significantly better. Moreover, by reducing the segment length to 16, MATR outperforms OAT-OSN in terms of both mAP and + +![](images/85c2b19333401bd42aaf9ce423d6ced38564b16eeef31d5b4daebd5add361f7e.jpg) +Fig. 5: Qualitative results of MATR and OAT-OSN [15] on the THUMOS14. Generated time is the timestamp when the predicted instance is generated. + +the inference time. MATR uses more parameters due to the separation of the decoder into the start decoder and the end decoders. + +# 4.4 Qualitative Results + +Fig. 5 illustrates the predictions of our model and OAT-OSN [15]. The results show that MATR has the capability to efficiently detect action instances by identifying intervals immediately after the current action ends. Additionally, the results show that our model is able to classify and localize action instances more precisely than OAT-OSN. + +# 5 Conclusion + +In this paper, we address online temporal action localization by introducing a new model, MATR, which leverages the memory queue to exploit long-term context. We also propose a new action instance localization method based on Transformer, which separately identifies action starts and action ends. As a result, MATR not only surpasses existing online TAL methods but also demonstrates comparable performance to offline TAL methods. The code base will be open to the public to promote future research on On-TAL. + +Limitations: When the memory queue contains multiple action instances, there is a risk of action instances being matched with incorrect start points. Additionally, when storing the input segment into the memory queue, MATR only considers whether the input segment is relevant to action instances but does not leverage the relationship with the past context stored in the memory. To address these issues, investigating a method to utilize the information stored in the memory queue when storing the input segment features, could lead to more effective memory storage and utilization. + +Acknowledgement. This work was supported by the NRF grant and the IITP grant funded by Ministry of Science and ICT, Korea (RS-2019-II191906, RS-2021-II212068, RS-2022-II220290, NRF-2021R1A2C3012728). + +# References + +1. Buch, S., Escorcia, V., Shen, C., Ghanem, B., Carlos Niebles, J.: Sst: Single-stream temporal action proposals. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2911-2920 (2017) +2. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. Proc. IEEE International Conference on Computer Vision (ICCV) (2023) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020) +4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +5. Chen, P., Gan, C., Shen, G., Huang, W., Zeng, R., Tan, M.: Relation attention for temporal action localization. IEEE Transactions on Multimedia 22(10), 2723-2733 (2019) +6. Cheng, F., Bertasius, G.: Tallformer: Temporal action localization with a long-memory transformer. In: Proc. European Conference on Computer Vision (ECCV). pp. 503-521. Springer (2022) +7. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Proc. European Conference on Computer Vision (ECCV). pp. 269-284. Springer (2016) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. International Conference on Learning Representations (ICLR) (2021), https://openreview.net/forum?id=YicbFdNTTy +9. Escorcia, V., Caba Heilbron, F., Niebles, J.C., Ghanem, B.: Daps: Deep action proposals for action understanding. In: Proc. European Conference on Computer Vision (ECCV). pp. 768-784. Springer (2016) +0. Gao, J., Yang, Z., Chen, K., Sun, C., Nevatia, R.: Turn tap: Temporal unit regression network for temporal action proposals. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3628-3636 (2017) +1. Gao, M., Xu, M., Davis, L.S., Socher, R., Xiong, C.: Startnet: Online detection of action start in untrimmed videos. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5542-5551 (2019) +2. Heilbron, F.C., Niebles, J.C., Ghanem, B.: Fast temporal activity proposals for efficient detection of human actions in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1914-1923 (2016) +3. Jiang, Y.G., Liu, J., Roshan Zamir, A., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/ (2014) +4. Kang, H., Kim, K., Ko, Y., Kim, S.J.: Cag-qil: Context-aware actionness grouping via q imitation learning for online temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13729-13738 (2021) + +15. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Proc. European Conference on Computer Vision (ECCV). pp. 653-669. Springer (2022) +16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proc. International Conference on Learning Representations (ICLR) (2015) +17. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3320-3329 (2021) +18. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3889-3898 (2019) +19. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: Proc. European Conference on Computer Vision (ECCV). pp. 3-19 (2018) +20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017) +21. Liu, Q., Wang, Z.: Progressive boundary refinement network for temporal action detection. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 11612-11619 (2020) +22. Liu, X., Hu, Y., Bai, S., Ding, F., Bai, X., Torr, P.H.: Multi-shot temporal event localization: a benchmark. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12596-12606 (2021) +23. Liu, X., Wang, Q., Hu, Y., Tang, X., Zhang, S., Bai, S., Bai, X.: End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing 31, 5427-5441 (2022) +24. Liu, Y., Ma, L., Zhang, Y., Liu, W., Chang, S.F.: Multi-granularity generator for temporal action proposal. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3604-3613 (2019) +25. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Gaussian temporal awareness networks for action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 344-353 (2019) +26. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +27. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18857-18866 (2023) +28. Shou, Z., Chan, J., Zareian, A., Miyazawa, K., Chang, S.F.: Cdc: Convolutional-de-convolutional networks for precise temporal action localization in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5734-5743 (2017) +29. Shou, Z., Pan, J., Chan, J., Miyazawa, K., Mansour, H., Vetro, A., Giro-i Nieto, X., Chang, S.F.: Online detection of action start in untrimmed, streaming videos. In: Proc. European Conference on Computer Vision (ECCV). pp. 534-551 (2018) +30. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1049-1058 (2016) +31. Tang, T.N., Park, J., Kim, K., Sohn, K.: Simon: A simple framework for online temporal action localization. arXiv preprint arXiv:2211.04905 (2022) + +32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. Advances in Neural Information Processing Systems (NeurIPS). pp. 5998-6008 (2017) +33. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13824-13835 (2023) +34. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proc. European Conference on Computer Vision (ECCV). pp. 20-36. Springer (2016) +35. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7565-7575 (2021) +36. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13587-13597 (2022) +37. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10186-10195 (2020) +38. Xu, M., Perez Rua, J.M., Zhu, X., Ghanem, B., Martinez, B.: Low-fidelity video encoder optimization for temporal action localization. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 9923-9935 (2021) +39. Xu, M., Zhao, C., Rojas, D.S., Thabet, A., Ghanem, B.: G-tad: Sub-graph localization for temporal action detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10156-10165 (2020) +40. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5532-5541 (2019) +41. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long shortterm transformer for online action detection. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 1086-1099 (2021) +42. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3160-3169 (2022) +43. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7094-7103 (2019) +44. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 492-510. Springer (2022) +45. Zhao, C., Thabet, A.K., Ghanem, B.: Video self-stitching graph network for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13658-13667 (2021) +46. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 485-502. Springer (2022) +47. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2914-2923 (2017) + +48. Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., Ren, D.: Distance-iou loss: Faster and better learning for bounding box regression. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 12993-13000 (2020) +49. Zhu, Z., Tang, W., Wang, L., Zheng, N., Hua, G.: Enriching local and global contexts for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13516-13525 (2021) \ No newline at end of file diff --git a/2024/Online Temporal Action Localization with Memory-Augmented Transformer/images.zip b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..00508115a6c5d724619c268dd4948076c3810e47 --- /dev/null +++ b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d097ddaa3a8b1e3fd48714942572ef3dce3b836cfc5ed101238479af75a9945 +size 522082 diff --git a/2024/Online Temporal Action Localization with Memory-Augmented Transformer/layout.json b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b49fae0621a6091035f1484990da79d9cf24e242 --- /dev/null +++ b/2024/Online Temporal Action Localization with Memory-Augmented Transformer/layout.json @@ -0,0 +1,8874 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 160, + 111, + 455, + 148 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 168, + 222, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 168, + 222, + 180 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 222, + 180 + ], + "type": "text", + "content": "Youngkil Song\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 238, + 169, + 318, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 169, + 318, + 180 + ], + "spans": [ + { + "bbox": [ + 238, + 169, + 318, + 180 + ], + "type": "text", + "content": "Dongkeun Kim*" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 335, + 169, + 393, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 169, + 393, + 179 + ], + "spans": [ + { + "bbox": [ + 335, + 169, + 393, + 179 + ], + "type": "text", + "content": "Minsu Cho" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 411, + 169, + 470, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 169, + 470, + 179 + ], + "spans": [ + { + "bbox": [ + 411, + 169, + 470, + 179 + ], + "type": "text", + "content": "Suha Kwak" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 158, + 190, + 456, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 190, + 456, + 224 + ], + "spans": [ + { + "bbox": [ + 158, + 190, + 456, + 224 + ], + "type": "text", + "content": "Pohang University of Science and Technology (POSTECH), South Korea {songyk, kdk1563, mscho, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/MATR/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 254, + 455, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 254, + 455, + 397 + ], + "spans": [ + { + "bbox": [ + 160, + 254, + 455, + 397 + ], + "type": "text", + "content": "Abstract. Online temporal action localization (On-TAL) is the task of identifying multiple action instances given a streaming video. Since existing methods take as input only a video segment of fixed size per iteration, they are limited in considering long-term context and require tuning the segment size carefully. To overcome these limitations, we propose memory-augmented transformer (MATR). MATR utilizes the memory queue that selectively preserves the past segment features, allowing to leverage long-term context for inference. We also propose a novel action localization method that observes the current input segment to predict the end time of the ongoing action and accesses the memory queue to estimate the start time of the action. Our method outperformed existing methods on two datasets, THUMOS14 and MUSES, surpassing not only TAL methods in the online setting but also some offline TAL methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 407, + 453, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 407, + 453, + 419 + ], + "spans": [ + { + "bbox": [ + 160, + 407, + 453, + 419 + ], + "type": "text", + "content": "Keywords: Temporal action localization " + }, + { + "bbox": [ + 160, + 407, + 453, + 419 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 407, + 453, + 419 + ], + "type": "text", + "content": " Online video understanding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 441, + 230, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 230, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 230, + 453 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 466, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 572 + ], + "type": "text", + "content": "Today, video is one of the most popular types of media, with tons of videos being created and uploaded to online platforms like YouTube and TikTok every second. As a result, the demand for video understanding is growing rapidly. In particular, dealing with untrimmed videos is a key to practical video understanding since most videos in the real world are not trimmed for individual events in advance. Temporal action localization (TAL) has been extensively studied in this context [10, 17, 21, 30, 38, 39, 43-45, 47, 49], with the aim of detecting each action instance in the input untrimmed video by predicting its start time, end time, and action class." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Recently, online TAL (On-TAL), i.e., TAL for streaming video [14], has gained increasing attention thanks to its potential applications in video surveillance, sports video analysis, and video summarization. Unlike TAL which demands the entire video as input, On-TAL aims to detect action instances using only input frames observed so far in an online manner. Moreover, once action instances are predicted, it is not allowed to modify the results afterward." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 223, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 223, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 223, + 665 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 247 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 247 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 247 + ], + "type": "image", + "image_path": "da375160703c66cee8a846f459f1e75e56a5aa3c7ec8a63d40229d01134fd1d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 301 + ], + "type": "text", + "content": "Fig. 1: A conceptual illustration of our model. Our model selectively stores information about previously occurred actions in the memory queue from streaming video. When the end of the action is detected in the current input segment, the model retrieves information from the memory queue to locate the start of the action." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 506 + ], + "type": "text", + "content": "A straightforward solution to this challenging task is to first conduct online action detection (OAD) [7,35], i.e., frame-by-frame action classification for streaming video, and then build action instances by aggregating the frame-wise action predictions. In this sense, earlier methods for On-TAL [14,31] follow this approach by applying OAD and refining per-frame predictions using past observations. However, this OAD-based approach exploits only frame-level supervision for training, which is not optimal because On-TAL aims to predict action instances in the form of time intervals. To alleviate this issue, online anchor transformer (OAT) has been recently proposed [15], which adopts sliding window approach and anchor-based method to utilize instance-level supervision. Although OAT improved performance remarkably, however, it still has several drawbacks: Since it takes a fixed-size segment of input video at each iteration and does not consider long-term contexts, its capability is limited in detecting long action instances, and its performance is sensitive to hyperparameters like the size of the input segment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 508, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 592 + ], + "type": "text", + "content": "To resolve the above issues, we propose a new end-to-end architecture for OntAL, dubbed memory-augmented Transformer (MATR). At the heart of MATR lies in memory queue, which selectively stores past segment features so that the model exploits long-term context for inference. Thanks to the memory queue that contains long-term context and reduces dependence on the segment size, our model accurately localizes even long-term action instances without the need for careful hyperparameter tuning for each dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Moreover, we propose a novel approach to action instance localization using the memory queue as illustrated in Fig. 1. First, it detects the end of an action using the current segment features and then scans past segment features in the memory queue to find the action start corresponding to the detected action end. To achieve this, we adopt two Transformer decoders: one for end detection and the other for start detection. Specifically, we adopt learnable queries as input to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "the decoders so that each of them learns to localize action boundaries through the attention mechanism of Transformer [8,32]. In addition, motivated by recent advances on object detection [6,37], MATR decouples action classification and localization to further improve the performance. This is realized in MATR by using distinct queries for the two subtasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 236 + ], + "type": "text", + "content": "The proposed method is evaluated on two datasets, THUMOS14 [13] and MUSES [22]. Our model achieves the state of the art on both datasets in the online TAL setting. In addition, surprisingly, it is also comparable to existing offline TAL methods although it does not employ offline non-maximum suppression (NMS). In summary, the contribution of this paper is three-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 247, + 480, + 366 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 247, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 479, + 282 + ], + "type": "text", + "content": "- We propose MATR, a new end-to-end architecture for On-TAL, leveraging memory queue to utilize long-term context and enabling precise localization of action instances with less dependence on dataset-specific hyperparameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 284, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 284, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 139, + 284, + 480, + 330 + ], + "type": "text", + "content": "- We introduce a new action instance localization method that first identifies action end using the input segments and scans past information in the memory queue to find the action start. Moreover, our method adopts distinct queries to separate information for action classification and localization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 332, + 479, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 332, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 139, + 332, + 479, + 366 + ], + "type": "text", + "content": "- Our method outperformed the existing On-TAL methods on the two benchmarks. Also, extensive ablation studies demonstrate the contribution of each component of the proposed method." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 387, + 236, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 236, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 236, + 399 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 544 + ], + "type": "text", + "content": "Temporal Action Localization. Temporal action localization (TAL) [5,10,17, 21,23,30,38,39,43-45,47,49] which identifies action instances in time and classifying their actions, has given much attention on the video research community. Earlier attempts employ two-stage approaches that first generate action proposals, then classify action class and refine action boundaries [1,9,12,18,19,24,47]. Recent studies have primarily dealt with widening the range of video context through convolutional neural networks [28], graph neural networks [39,43], Gaussian kernel [25], and utilizing global context [49]. To extend temporal context and enable end-to-end TAL, Transformer [8,32] has recently been adopted [6,23,44]. These approaches incorporate temporal deformable attention module [23] and multi-scale features [44] to improve computational efficiency and performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 665 + ], + "type": "text", + "content": "Online Video Understanding. Online action detection (OAD), which classifies action classes in a frame-wise manner, is first proposed by Geest et al. [7]. To incorporate long-term context, recurrent neural network [40] and Transformer [35] are utilized. These methods predict actions for future frames and use this information to refine current frame predictions. Yang et al. [42] further improve the performance by using category exemplars and capturing long-term dependencies within a video segment. On the other hand, Shou et al. [29] proposes online detection of action start (ODAS), which aims to detect action start as early as possible. Gao et al. [11] further improves ODAS by adopting reinforcement learning." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "Online Temporal Action Localization. Towards practical event understanding, online temporal action localization (On-TAL) has been proposed recently. Kang et al. [14] first propose On-TAL, which localizes action boundaries and classifies action classes without accessing future frames and offline post-processing. Earlier methods [14,31] adopt OAD framework to classify per-frame action class and group them into action instances to solve On-TAL. Specifically, Kang et al. [14] utilize Markov decision process to aware prior contexts to group per-frame predictions. SimOn [31] is proposed as an end-to-end On-TAL method that leverages past visual context and action probabilities for accurate action prediction at the current frame. Kim et al. [15] propose a sliding window scheme and an anchor-based method that adopts Transformer [32] for On-TAL. However, it only uses a fixed input segment which is highly dependent on the action length distribution of each dataset. Unlike the previous methods, we utilize a memory queue for long-term context, which also reduces the dependence on the input segment size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 537 + ], + "type": "text", + "content": "Difference between OAD and On-TAL. OAD [2,33,35,40-42,46] focus on predicting action classes at the frame-level, which limits the ability to distinguish boundaries of overlapping actions. On the other hand, On-TAL [14,15,31] predicts entire action instances rather than individual frames, enabling it to precisely identify the start and end times of actions, making it effective for applications requiring instance-level action understanding, like sports video analysis. Using Memory for Video Understanding. Prior work in video understanding employs memory to store contextual information, in the forms of long-term and short-term memory [2,33,36,41]: short-term memory captures fine details, and long-term memory stores compressed information of past contexts. Specifically, Wang et al. [33] divides long-term memory into several groups and compresses each group into a single vector. Stream Buffer [2] gradually reduces temporal dimensions of the memory using a compression module. However, unlike video classification tasks such as action recognition [36] and OAD [2,33,41,46], which requires to store representative information of past frames in the memory for classifying the action of the current frame, TAL needs to preserve the temporal positional information of past frames in the memory to accurately predict the exact time. In that sense, MATR preserves temporal information in memory to accurately localize actions and effectively stores information by using flag tokens. A comparison with the OAD memory modules is shown in Table 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 555, + 261, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 261, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 261, + 569 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Consider an untrimmed video " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\mathbf{v}_i\\}_{i=1}^T" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " action instances " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varPsi=\\{(s_m,e_m,c_m)\\}_{m=1}^{M}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", where each instance is represented by its start time " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", end time " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "e_m" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and action class " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_m" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". Unlike temporal action localization (TAL), which allows a model to use whole video sequence as input, online temporal action localization (On-TAL) forces a model to predict action instances using only input frames seen until the current timestamp. Note that action instances predicted at previous iterations cannot be modified or deleted afterwards." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 480, + 240 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 480, + 240 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 480, + 240 + ], + "type": "image", + "image_path": "7a6d7ed0afe60bf9567f5d531cf41a14e4f748475a10ec1f6ed7e1cbb0e37d60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 248, + 480, + 271 + ], + "lines": [ + { + "bbox": [ + 131, + 248, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 480, + 271 + ], + "type": "text", + "content": "Fig. 2: Overall architecture of MATR. MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 302, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 518 + ], + "type": "text", + "content": "As shown in Fig. 2, MATR consists of four parts: feature extractor, memory-augmented video encoder, instance decoding module, and prediction heads. In line with previous research, the current segment of the input streaming video is given as a unit input, and its frame-level features, referred to as segment features, are extracted by a video backbone network and a linear projection layer (Sec. 3.1). The segment features are then fed to the memory-augmented video encoder, which encodes temporal context between frames in the current segment and stores the segment features into the memory (Sec. 3.2). The instance decoding module localizes action instances via two Transformer decoders: the end decoder and the start decoder. Specifically, the end decoder references the encoded segment features to locate the action end around the current time, and then the start decoder refers to the memory queue to find the action start based on the past information stored in the memory queue. Queries for each instance consist of a class query for action classification and a boundary query for action localization (Sec. 3.3). The outputs of the instance decoding module are used as inputs to the prediction heads, which consist of end prediction head, start prediction head, and action classification head (Sec. 3.4). The entire model is trained in an end-to-end manner (Sec. 3.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 555, + 251, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 251, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 251, + 565 + ], + "type": "text", + "content": "3.1 Feature Extractor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Given a segment comprising " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " consecutive video frames as input, a pretrained backbone [4, 34] followed by a linear projection layer extracts the segment features " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t = \\{\\mathbf{x}_i\\}_{t - L_s + 1}^t\\in \\mathbb{R}^{L_s\\times D}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for each frame within the segment, following previous work [14, 15]. Inspired by Kim et al. [15], MATR employs a sliding window scheme, which predicts multiple action instances by moving frame by frame along the temporal axis. Note that the window size is the input segment size." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 448, + 102 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 343, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 343, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 343, + 129 + ], + "type": "text", + "content": "3.2 Memory-Augmented Video Encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 196 + ], + "type": "text", + "content": "At each time step, the input segment features are fed to two modules of the memory-augmented video encoder, the segment encoder and the memory update module. The segment encoder encodes the temporal context across input segment features, while the memory update module selectively stores the segment features and updates the memory queue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 196, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 482, + 291 + ], + "type": "text", + "content": "Segment Encoder. The segment encoder is a standard Transformer encoder composed of self-attention layers and a feed forward network (FFN). The combination of segment features and a learnable flag token is fed into the encoder, and transformed to queries, keys, and values for self-attention. Sinusoidal positional encoding " + }, + { + "bbox": [ + 130, + 196, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{pos}} \\in \\mathbb{R}^{(L_s + 1) \\times D}" + }, + { + "bbox": [ + 130, + 196, + 482, + 291 + ], + "type": "text", + "content": " is added to the queries and keys. Among the output embeddings, the encoded flag token is fed into the memory update module and the encoded segment features that integrate temporal context across input frames are used as input to the end decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "text", + "content": "Memory Update Module. In an online setting, where an input is a streaming video, efficiently storing information of past frames and accessing it effectively is crucial for detecting action instances, in particular those of long-term actions whose temporal extents are beyond the size of input segments. To this end, we adopt the memory queue for storing information of past input segments in a first-in-first-out (FIFO) manner. Additionally, we propose an efficient memory update method employing the flag token. When the input segment features pass through the segment encoder, the flag token is concatenated into the input segment features and then employed in the flag prediction head. It is trained to predict [FLAG], which identifies whether an input segment is relevant to action instances, so that only relevant segments are stored in the memory. [FLAG] = 1 when there are overlapping frames between the input segment and action instances, and 0 otherwise. During training, the ground-truth [FLAG] is utilized. During inference, [FLAG] = 1 if sigmoid(\\hat{g}) > \\theta where " + }, + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "text", + "content": " is the output logit of the flag prediction head and " + }, + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "text", + "content": " is a predefined threshold, and 0 otherwise. The input segment feature is added to the memory queue when [FLAG] = 1 and discards it when [FLAG] = 0. If the memory queue is full, the oldest segment feature is purged." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 295, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 295, + 525 + ], + "type": "text", + "content": "3.3 Instance Decoding Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The instance decoding module localizes and classifies action instances by leveraging the encoded segment features and the memory queue through the attention mechanism of Transformer. Given the encoded segment features that have short-term temporal contexts across input segment and the memory queue that stores long-term contexts for actions that are potentially ongoing, a set of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " instance queries " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\{\\mathbf{Q}_{\\mathrm{class}}; \\mathbf{Q}_{\\mathrm{bound}}\\} \\in \\mathbb{R}^{2N \\times D}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " is trained to generate " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " action instances for each input segment. Half of these queries pertain to predicting the action class (class query; " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{class}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "), while the other half is used for predicting the start and end timestamps (boundary query; " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{bound}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "). The class and boundary query pairs for the same instance share the same positional embedding " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pos}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to accurately identify and differentiate between" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 480, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 240 + ], + "type": "image", + "image_path": "57feb6d1ef320ad3bda0e3452736157218e717846a8a0095b857edcbe7be4792.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 248, + 433, + 259 + ], + "lines": [ + { + "bbox": [ + 178, + 248, + 433, + 259 + ], + "spans": [ + { + "bbox": [ + 178, + 248, + 433, + 259 + ], + "type": "text", + "content": "Fig. 3: Detailed architecture of the instance decoding module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 287, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 360 + ], + "type": "text", + "content": "instances; " + }, + { + "bbox": [ + 130, + 287, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}} = \\{\\mathbf{Q}_{\\mathrm{class}} + \\mathbf{E}_{\\mathrm{pos}};\\mathbf{Q}_{\\mathrm{bound}} + \\mathbf{E}_{\\mathrm{pos}}\\}" + }, + { + "bbox": [ + 130, + 287, + 480, + 360 + ], + "type": "text", + "content": ". Unlike the previous Transformer-based On-TAL method [15] that predicts start and end offset for each instance simultaneously, our method employs separate Transformer decoders for predicting start offset and end offset. As shown in Fig. 3, the start decoder and end decoders share the same architecture but utilizes different information to predict action start and action end, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 481, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 445 + ], + "type": "text", + "content": "End Decoder. Given the encoded segment features, instance queries " + }, + { + "bbox": [ + 130, + 361, + 481, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 361, + 481, + 445 + ], + "type": "text", + "content": " is trained to generate the output embeddings used for detecting action end near the current timestamp. To this end, we adopt a Transformer [3] composed of multi-head self-attention layers, multi-head cross-attention layers, and FFNs. The detailed architecture is illustrated in Fig. 3. The outputs of the end decoder, which consist of end boundary embeddings and class embeddings, are then used as inputs to the start decoder and the prediction heads." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 481, + 579 + ], + "type": "text", + "content": "Start Decoder. The start decoder takes the output embeddings from the end decoder and locates the corresponding action start by utilizing the memory queue. First, the memory queue is concatenated with the current segment features. The concatenated memory features are used as long-term context in the start decoder. Then, 2D temporal positional encodings are added to the memory features, and the results are used as the key and value for the cross-attention layers of the start decoder. The overall architecture of the start decoder is the same as the end decoder (Fig. 3), but it uses the memory features rather than the encoded segment features. The output embeddings of the start decoder, start boundary embeddings and class embeddings, are used as inputs to the start prediction and the action classification head, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Moreover, we introduce two techniques to efficiently and effectively utilize the memory queue. First, we conduct a " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " uniform sampling to the memory features which enables efficient use of memory since adjacent frames contain similar information. Second, to expand the scope of positional encoding for streaming video, whose duration is unpredictable, we separate the temporal positional encoding into the two parts: relative segment position and relative frame position. A relative segment position is the position of a segment within the memory rela" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 448, + 102 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "tive to the current segment. Meanwhile, a relative frame position is the position of a frame relative to the most recent frame within the same segment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 157, + 248, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 157, + 248, + 168 + ], + "spans": [ + { + "bbox": [ + 131, + 157, + 248, + 168 + ], + "type": "text", + "content": "3.4 Prediction Heads" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 479, + 223 + ], + "type": "text", + "content": "To generate " + }, + { + "bbox": [ + 130, + 175, + 479, + 223 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 175, + 479, + 223 + ], + "type": "text", + "content": " action instances " + }, + { + "bbox": [ + 130, + 175, + 479, + 223 + ], + "type": "inline_equation", + "content": "\\{(\\hat{s}_i,\\hat{e}_i,\\hat{c}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 175, + 479, + 223 + ], + "type": "text", + "content": " from the output of the instance decoding module, we use three prediction heads, which are end prediction head, start prediction head, and action classification head. Each of the prediction heads is composed of a 2-layer FFN." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 272 + ], + "type": "text", + "content": "End Prediction Head. The end prediction head estimates the offset between the end of the target action and the current timestamp. The boundary embeddings from the end decoder are fed to the 2-layer FFN, which in turn estimates the end offset " + }, + { + "bbox": [ + 130, + 224, + 480, + 272 + ], + "type": "inline_equation", + "content": "\\{\\hat{u}_i\\}_{i=1}^N \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 224, + 480, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": "Start Prediction Head. The start prediction head estimates the offset between the start of the target action and the current timestamp. Unlike the end prediction, which finds the end offset around the input segment, the start prediction needs to perform offset regression over a relatively wider range. To narrow down the scope of the start offset regression, we divide the time horizon into " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "L_{m} + 2" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": " regions: the region preceding the coverage of memory, " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": " regions covered by the memory, and the region corresponding to the current input segment. In that sense, the start prediction head consists of the region classification and the offset regression, each composed of a 2-layer FFN. They take the boundary embeddings from the start decoder as input to predict the start time. The region classification identifies the regions to which the start time is assigned, represented as " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\{\\hat{o}_i|\\hat{o}_i = \\mathrm{argmax}(\\hat{r}_i)\\}_{i=1}^N \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\{\\hat{r}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (L_m + 2)}" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": " are the output logits from the region classification head. Following this, the offset regression head is utilized to estimate the offset within the regions, denoted as " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\{\\hat{v}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (L_m + 2)}" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": ". Note that the start offset " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\hat{v}_i" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": " is calculated for all " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "L_{m} + 2" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": " regions, and the offset from the identified region " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\{\\hat{v}_i(\\hat{o}_i)\\}_{i=1}^N \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": " is used in inference. By combining start classification result " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\hat{o}_i" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": " and start regression result " + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\hat{v}_i(\\hat{o}_i)" + }, + { + "bbox": [ + 130, + 272, + 481, + 498 + ], + "type": "text", + "content": ", the start time is predicted. The effectiveness of the proposed start prediction head is validated in Table 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 479, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 547 + ], + "type": "text", + "content": "Action Classification Head. For action class prediction, we utilize the class embeddings from both the end decoder and the start decoder. These embeddings are concatenated and fed into the action classification head to derive class probabilities " + }, + { + "bbox": [ + 130, + 498, + 479, + 547 + ], + "type": "inline_equation", + "content": "\\{\\hat{p}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (C+1)}" + }, + { + "bbox": [ + 130, + 498, + 479, + 547 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 498, + 479, + 547 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 498, + 479, + 547 + ], + "type": "text", + "content": " is the number of action class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "text", + "content": "Action Instance Prediction. Finally, our model makes " + }, + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "text", + "content": " action proposals " + }, + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "inline_equation", + "content": "\\{(\\hat{s}_i,\\hat{e}_i,\\hat{c}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 547, + 479, + 571 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 580, + 479, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 580, + 479, + 591 + ], + "spans": [ + { + "bbox": [ + 249, + 580, + 479, + 591 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {i} = t - \\left(\\hat {o} _ {i} + \\hat {v} _ {i} \\left(\\hat {o} _ {i}\\right)\\right) \\times L _ {s}, \\tag {1}", + "image_path": "31a824eee58f4c3834347044635c3ff0c587107628dbe592c764b68d15add1ab.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 249, + 594, + 480, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 594, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 249, + 594, + 480, + 606 + ], + "type": "interline_equation", + "content": "\\hat {e} _ {i} = t + \\hat {u} _ {i} \\times L _ {s}, \\tag {2}", + "image_path": "6df6c4330b746f8f61ec45b8bf22e1c682d40799823e4939bf51e7c206484243.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 609, + 480, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 609, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 249, + 609, + 480, + 621 + ], + "type": "interline_equation", + "content": "\\hat {c} _ {i} = \\operatorname {a r g m a x} \\left(\\hat {p} _ {i}\\right), \\tag {3}", + "image_path": "764e2175a2156eb29e59170bdf0f79e2c69c28340eb314a7fe97bfe090a30957.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": " is the segment length. Since the sliding window scheme generates " + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": " action proposals at each timestamp, post-processing is crucial for performance improvement by removing redundant or overlapped action instances." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": "Non-maximum suppression (NMS) is applied to the action proposals at each timestamp, subsequently removing those highly overlapped with the proposals generated from the past. To prevent more reliable prediction made in the future from being removed, instances where the predicted end time is beyond the current timestamp, " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "t < \\hat{e}_i" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": ", are also removed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 191, + 256, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 191, + 256, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 256, + 204 + ], + "type": "text", + "content": "3.5 Training Objective" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "content": "At timestamp " + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "content": ", the model is trained to detect action instances whose end times are in the range " + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "inline_equation", + "content": "[t - T_d + 1, t + T_a]" + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "inline_equation", + "content": "T_d" + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "inline_equation", + "content": "T_a" + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "content": " are hyperparameters. Then, hungarian algorithm matches the action proposals and the ground-truth action instances with the lowest matching cost. The matching cost of the ground-truth group " + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "content": " and the proposal " + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "inline_equation", + "content": "\\sigma(i)" + }, + { + "bbox": [ + 131, + 209, + 479, + 269 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 275, + 481, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 275, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 235, + 275, + 481, + 290 + ], + "type": "interline_equation", + "content": "C _ {i, \\sigma (i)} = \\hat {p} _ {\\sigma (i)} \\left(y _ {i}\\right) + \\operatorname {I o U} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right), \\tag {4}", + "image_path": "a14354ea9f8a2da6b7424f69bba1cb2fd286cff9df44df69e0660b1ff520749b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": " is the permutation of " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": " action proposals, " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": " is the class probabilities of the " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": "-th proposal, " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "b_i = \\{s_i, e_i\\}" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": " is the ground-truth action boundary, and " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\hat{b}_i = \\{\\hat{s}_i, \\hat{e}_i\\}" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": " is the predicted action boundary. We adopt the focal loss [20] for the action classification, the cross entropy loss for the start region classification, the " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": " loss for both the start offset regression and the end offset regression. The action classification loss " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{class}}" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": ", the start prediction loss " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{start}}" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": ", and the end prediction loss " + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{end}}" + }, + { + "bbox": [ + 131, + 294, + 482, + 380 + ], + "type": "text", + "content": " are defined as follow:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 196, + 386, + 481, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 386, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 196, + 386, + 481, + 417 + ], + "type": "interline_equation", + "content": "L _ {\\text {c l a s s}} = \\sum_ {i = 1} ^ {N} \\operatorname {F L} \\left(\\hat {p} _ {i}, y _ {i}\\right), \\tag {5}", + "image_path": "d4b2c61423fc82ffc26fc7cc289d04908d421d63de89d2962a39b88df857fa75.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 196, + 421, + 480, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 421, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 196, + 421, + 480, + 452 + ], + "type": "interline_equation", + "content": "L _ {\\text {s t a r t}} = \\sum_ {i = 1} ^ {N _ {\\text {m a t c h}}} \\left\\{\\mathrm {C E} \\left(\\operatorname {S o f t m a x} \\left(\\hat {r} _ {i}\\right), r _ {i}\\right) + \\left| \\hat {v} _ {i} \\left(\\hat {o} _ {i}\\right) - v _ {i} \\right| \\right\\}, \\tag {6}", + "image_path": "089dc16680535c243071b1990641d12f92ae233eb05aaaa41d6c675ae187297f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 456, + 480, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 456, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 200, + 456, + 480, + 487 + ], + "type": "interline_equation", + "content": "L _ {\\text {e n d}} = \\sum_ {i = 1} ^ {N _ {\\text {m a t c h}}} | \\hat {u} _ {i} - u _ {i} |, \\tag {7}", + "image_path": "911e704f7b167305854be996213847241002157dad0038ae3345368c7f343a2d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 492, + 479, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 479, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 479, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 492, + 479, + 515 + ], + "type": "inline_equation", + "content": "y_{i}, r_{i}, v_{i}, u_{i}" + }, + { + "bbox": [ + 131, + 492, + 479, + 515 + ], + "type": "text", + "content": " is the ground-truth action class, start region, start offset, and end offset, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 516, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 516, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 516, + 479, + 552 + ], + "type": "text", + "content": "To provide instance-level supervision and facilitate the connection between start and end time prediction, we employ the DIoU loss [48] following Action-Former [44]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 219, + 553, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 553, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 219, + 553, + 480, + 583 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {d i o u}} = \\sum_ {i = 1} ^ {N _ {\\mathrm {m a t c h}}} 1 - \\operatorname {I o U} (\\hat {b} _ {i}, b _ {i}) + \\frac {\\rho^ {2} (\\hat {c} _ {i} , c _ {i})}{d _ {i} ^ {2}}, \\tag {8}", + "image_path": "c996d9a1b24cbfd2eb03e12e69a6527db9ae4bbde10498790dfe731f1e1ddff9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot, \\cdot)" + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "content": " is the Euclidean distance between two points, " + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "content": " are the center of the proposal and ground-truth instance, and " + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "content": " is the smallest enclosing 1-dim box length." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "type": "text", + "content": "To train the memory update module in an end-to-end manner, we employ the flag loss for training the flag prediction head:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 246, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 246, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "L _ {\\text {f l a g}} = \\operatorname {B C E} (\\operatorname {S i g m o i d} (\\hat {g}), g), \\tag {9}", + "image_path": "353424edfe3677943a12d841a6c7beaeb57d0722986b8dca801610a58fefdd10.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 448, + 103 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " is the predicted logit of flag token and ground-truth [FLAG] respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 140, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 140, + 479, + 163 + ], + "type": "text", + "content": "Our model is trained with six losses simultaneously in an end-to-end manner. The total loss is calculated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 175, + 479, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 175, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 217, + 175, + 479, + 186 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {c l a s s}} + L _ {\\text {s t a r t}} + L _ {\\text {e n d}} + L _ {\\text {d i o u}} + L _ {\\text {f l a g}}. \\tag {10}", + "image_path": "87a99100a7ad7f12ba5f18f9bb654601c4c9e3a8508b5733c53d7d169821fe5f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 194, + 465, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 465, + 207 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 465, + 207 + ], + "type": "text", + "content": "Note that all loss coefficients are set to 1 and loss balancing is not required." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 225, + 230, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 230, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 230, + 239 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 250, + 269, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 250, + 269, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 250, + 269, + 262 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "content": "Datasets. We evaluate our method and previous methods on two On-TAL benchmarks: THUMOS14 [13] and MUSES [22]. THUMOS14 contains 200 videos for training and 213 videos for testing with 20 action classes, while MUSES has 2,587 videos for training and 1,110 videos for testing with 25 action classes. MUSES consists of multi-shot action instances, which makes action localization more challenging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "content": "Hyperparameters. We use a frozen two-stream TSN [34] pretrained on Kinetics [4] to extract RGB and flow features for THUMOS14, and an I3D [4] trained on Kinetics [4] utilizing only RGB features for MUSES, following previous OntAL work [14, 15, 31]. The segment size " + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "content": " is set to 64 in THUMOS14 and 75 in MUSES. The dimension of the segment feature is set to " + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "inline_equation", + "content": "D = 1024" + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "content": " for both datasets. For the memory-augmented video encoder, we stack 3 Transformer layers with 8 attention heads for both datasets. The flag threshold " + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "content": " is set to 0.5. The memory size " + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "content": " is set to 7 for THUMOS14, while 15 for MUSES. For the instance decoding module, we stack 5 Transformer layers with 4 attention heads for both datasets. The number of the class query and the boundary query " + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "content": " for THUMOS14 and " + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 130, + 342, + 480, + 472 + ], + "type": "text", + "content": " for MUSES." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "text", + "content": "Training. During training, we use Adam optimizer [16], with the initial learning rate of 1e-8 with CosineAnnealing scheduler [26]. The batch size is set to 64 for THUMOS14, while 75 for MUSES. The focal loss coefficient is " + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\alpha = 0.25" + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "text", + "content": " for THUMOS14 and " + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\gamma = 5" + }, + { + "bbox": [ + 131, + 473, + 480, + 533 + ], + "type": "text", + "content": " for MUSES. All loss coefficients are set to 1. More details are listed in the supplementary material (Sec. D)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 550, + 352, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 352, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 352, + 562 + ], + "type": "text", + "content": "4.2 Comparison with the State of the Art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "We compare our method with previous online and offline TAL methods on THUMOS14 and MUSES dataset. As presented in Table 1, our method outperforms all previous On-TAL methods in both benchmarks by a substantial margin: " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " of average mAP in THUMOS14 and " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " of average mAP in MUSES. The results show that MATR utilizing the memory queue is more effective than previous OAD-based methods [14,31,33,46] and instance-level method OAT-OSN [15]. We also evaluate two OAD methods utilizing memory [33,46] in On-TAL setting by grouping their frame-level predictions into action instances." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 482, + 325 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison with On-TAL methods on THUMOS14 and MUSES dataset. The results are reported in mAP measure (%). **' indicates that the result is validated using the On-TAL ground truth from classification annotation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 482, + 325 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 482, + 325 + ], + "type": "table", + "html": "
MethodTHUMOS14MUSES
backbone0.30.40.50.60.7Averagebackbone0.30.40.50.60.7
Offline TAL
G-TAD [39]TSN [34]54.547.640.230.823.439.9I3D [4]19.114.811.17.44.7
ContextLoc [49]I3D68.363.854.341.826.250.9------
P-GCN [43]I3D63.657.849.1---I3D19.917.113.19.75.4
MUSES [22]I3D68.964.056.946.331.053.4I3D25.922.618.915.010.6
ActionFormer [44]I3D82.177.871.059.443.966.8------
TriDet [27]I3D83.680.172.962.447.469.3------
OAD-based Online TAL
TeSTra [46]TSN35.629.221.413.47.621.4------
MAT [33]I3D36.027.519.212.36.420.3------
CAG-QIL [14]TSN44.737.629.821.914.529.7I3D8.56.54.22.81.9
SimOn [31]TSN54.345.035.023.314.634.4------
SimOn* [31]TSN57.047.537.326.616.036.9------
Instance-level Online TAL
OAT-Naive [15]TSN57.650.643.030.015.739.4I3D20.316.612.97.73.6
OAT-OSN [15]TSN63.056.747.136.320.044.6I3D22.118.514.28.94.7
MATRTSN70.362.752.138.623.749.5I3D23.519.314.39.45.7
", + "image_path": "9b7bb9d9f01b3b44740f66c677e85374f6eb323abb995cbef82664a93d698a56.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 348, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 468 + ], + "type": "text", + "content": "However, their performance is far lower than MATR, which verifies the effectiveness of the proposed memory queue of MATR. We also compare our model with several offline TAL methods. Although there is still a performance gap between online and offline methods, our model achieves comparable performance to offline methods, and even outperforms previous work [39, 43]. The performance on the MUSES dataset is lower compared to THUMOS14 and the reason is twofold: (1) Action instances in MUSES are captured in multi-shots, making it challenging to detect action boundaries, and (2) there are confusing action classes such as \"quarrel\" and \"fight,\" making action classification hard. These make MUSES challenging and limit the upper-bound of performance on it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 246, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 246, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 246, + 498 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "type": "text", + "content": "We perform ablation studies on THUMOS14 and MUSES dataset to show the effectiveness of the proposed modules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Ablation on the Proposed Modules. In Table 2, we conduct an ablation study to verify the effectiveness of each component of our model. In Table 2(a), focuses on the components of the memory-augmented video, both removing the flag token and the segment encoder led to a performance drop. The results show the efficacy of selectively storing the past information and the importance of temporal context across the input segment. In Table 2(b), we use a single decoder and predict both start and end at once, which results significant performance drop, from " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "49.5\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "42.7\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". The result demonstrates the contribution of our new action localization method, using two separate decoders for the start and the end prediction. We also ablate new query design, sampling the memory for the start decoder, and positional embedding for instance queries. The results" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 158, + 315, + 277 + ], + "blocks": [ + { + "bbox": [ + 137, + 125, + 312, + 157 + ], + "lines": [ + { + "bbox": [ + 137, + 125, + 312, + 157 + ], + "spans": [ + { + "bbox": [ + 137, + 125, + 312, + 157 + ], + "type": "text", + "content": "Table 2: Ablation of the proposed modules on THUMOS14 dataset. The results are reported in mAP measure " + }, + { + "bbox": [ + 137, + 125, + 312, + 157 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 137, + 125, + 312, + 157 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 158, + 315, + 277 + ], + "lines": [ + { + "bbox": [ + 138, + 158, + 315, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 158, + 315, + 277 + ], + "type": "table", + "html": "
Method0.30.40.50.60.7Average
Ours70.362.752.138.623.749.5
(a) Memory-augmented video encoder
w/o flag token67.361.050.136.522.347.4
w/o segment encoder65.859.349.936.321.846.6
(b) Instance decoding module
single decoder65.156.445.030.516.742.7
w/o splitting query66.261.450.938.023.247.9
w/o sampling66.159.650.639.922.547.2
w/o pos embedding67.260.549.836.322.747.3
(c) Training objective
w/o DIoU loss62.253.443.231.017.441.4
", + "image_path": "f47519afe84525e2db92d59394d0b3342c04a0930a5b2d57352adb9ff7689ecd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 321, + 147, + 472, + 283 + ], + "blocks": [ + { + "bbox": [ + 318, + 125, + 475, + 146 + ], + "lines": [ + { + "bbox": [ + 318, + 125, + 475, + 146 + ], + "spans": [ + { + "bbox": [ + 318, + 125, + 475, + 146 + ], + "type": "text", + "content": "Table 3: Analysis of the maximum length of the memory queue." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 147, + 472, + 283 + ], + "lines": [ + { + "bbox": [ + 321, + 147, + 472, + 283 + ], + "spans": [ + { + "bbox": [ + 321, + 147, + 472, + 283 + ], + "type": "table", + "html": "
Memory size0.30.40.50.60.7Average
(a) THUMOS14 dataset
w/o mem65.858.947.636.920.846.0
167.259.950.737.922.747.7
367.361.951.937.922.648.3
770.362.752.138.623.749.5
1167.960.649.837.823.247.9
1566.960.150.738.424.148.0
1967.060.152.540.224.649.1
(b) MUSES dataset
w/o mem22.317.613.19.14.613.3
123.018.013.58.64.913.6
323.118.313.98.95.113.9
723.518.113.58.94.913.8
1122.918.714.09.45.414.1
1523.519.314.39.45.714.4
1922.718.913.89.65.514.1
", + "image_path": "231cfde0532c43f6d9c0deb0ee4aa02adf5fc92ae4e970e9adaf91a8959989b6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 143, + 315, + 470, + 354 + ], + "blocks": [ + { + "bbox": [ + 146, + 294, + 466, + 305 + ], + "lines": [ + { + "bbox": [ + 146, + 294, + 466, + 305 + ], + "spans": [ + { + "bbox": [ + 146, + 294, + 466, + 305 + ], + "type": "text", + "content": "Table 4: Comparison between different memory modules on the THUMOS14." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 315, + 470, + 354 + ], + "lines": [ + { + "bbox": [ + 143, + 315, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 143, + 315, + 470, + 354 + ], + "type": "table", + "html": "
Memory moduleBackboneSegment sizeInference timefpsMemory parametersAverage mAP
MAT [33]TSN [34]64191.9ms5.240.1M46.9
E2E-LOAD [2]TSN64196.1ms5.153.1M47.9
MATRTSN64167.1ms6.024.0M49.5
", + "image_path": "d6f9e438d01ee68067b7fe725ec66f4da3403645e67ad91647c559ef978fa61c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 377, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 425 + ], + "type": "text", + "content": "show that all design choices of the proposed components are important for action instance localization. Lastly, training our model without DIoU loss shows a drastic performance drop, which demonstrates the importance of instance-level supervision for On-TAL (Table 2(c))." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "type": "text", + "content": "Impact of the Memory Queue Size. We conduct an ablation study on the different memory queue sizes. As shown in Table 3, employing the memory queue enhances the performance compared to the without memory case, and generally the performance improves as the memory queue size increases. With an adequately large memory queue size covering the top " + }, + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "type": "text", + "content": " of the instances in the training split, such as 7 for the THUMOS14 and 5 for the MUSES, the model shows robust performance and the performance is higher than the previous state-of-the-art model [15]. The memory queue size of 7 and 15 shows the best on THUMOS14 and MUSES, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "Comparison of the proposed memory queue with OAD methods using memory. We compare our memory module with previous memory modules from OAD tasks [2, 33] by replacing the memory-augmented video encoder (Sec. 3.2) of our model. For a fair comparison, hidden dimension of the attention blocks for all methods is set to 1024. As shown in Table 4, MATR outperforms the existing memory modules in terms of both mAP and space-time complexity. While existing memory modules use multiple attention blocks, our memory module consists of a segment encoder and a flag prediction head, making it superior in terms of parameters and inference speed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Analysis of the Segment Size. As shown in Fig. 4, our model is less sensitive to the input segment size, which should be carefully tuned across various" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 123, + 271, + 223 + ], + "blocks": [ + { + "bbox": [ + 135, + 123, + 271, + 223 + ], + "lines": [ + { + "bbox": [ + 135, + 123, + 271, + 223 + ], + "spans": [ + { + "bbox": [ + 135, + 123, + 271, + 223 + ], + "type": "image", + "image_path": "0adb606a4a25c04d57fd640bb28efd7652879108e51a660b08ed545786c666e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 235, + 274, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 235, + 274, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 274, + 289 + ], + "type": "text", + "content": "Fig.4: Average mAP " + }, + { + "bbox": [ + 132, + 235, + 274, + 289 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 235, + 274, + 289 + ], + "type": "text", + "content": " versus the segment size in THUMOS14. Red line represents the result of our model and blue line represents the result of OAT-OSN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 286, + 155, + 475, + 193 + ], + "blocks": [ + { + "bbox": [ + 280, + 121, + 482, + 154 + ], + "lines": [ + { + "bbox": [ + 280, + 121, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 280, + 121, + 482, + 154 + ], + "type": "text", + "content": "Table 5: Ablation on the start and the end prediction heads. Reg and CIs refer to the offset regression and the region classification respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 286, + 155, + 475, + 193 + ], + "lines": [ + { + "bbox": [ + 286, + 155, + 475, + 193 + ], + "spans": [ + { + "bbox": [ + 286, + 155, + 475, + 193 + ], + "type": "table", + "html": "
StartEnd0.30.40.50.60.7Average
RegReg64.259.050.037.822.746.7
Reg + ClsReg70.362.752.138.623.749.5
Reg + ClsReg + Cls64.058.950.238.323.246.9
", + "image_path": "0c2c06e58ff1701cf5f83bd7d1feca1b3c901168f4dace0faf6224c068122190.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 284, + 230, + 478, + 288 + ], + "blocks": [ + { + "bbox": [ + 280, + 208, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 280, + 208, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 280, + 208, + 481, + 228 + ], + "type": "text", + "content": "Table 6: Ablation on the memory compression factor on the THUMOS14 dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 284, + 230, + 478, + 288 + ], + "lines": [ + { + "bbox": [ + 284, + 230, + 478, + 288 + ], + "spans": [ + { + "bbox": [ + 284, + 230, + 478, + 288 + ], + "type": "table", + "html": "
Compress factor0.30.40.50.60.7Average
None (Ours)70.362.752.138.623.749.5
267.359.949.436.922.347.7
468.161.552.739.724.549.3
867.561.352.038.823.848.7
", + "image_path": "ed83e36216de42f43bfec490e3711ca1f133b194f30b42f1f52b98f9508f5c3e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 154, + 324, + 459, + 364 + ], + "blocks": [ + { + "bbox": [ + 137, + 304, + 475, + 316 + ], + "lines": [ + { + "bbox": [ + 137, + 304, + 475, + 316 + ], + "spans": [ + { + "bbox": [ + 137, + 304, + 475, + 316 + ], + "type": "text", + "content": "Table 7: Comparison of the inference time and # parameters on the THUMOS14." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 324, + 459, + 364 + ], + "lines": [ + { + "bbox": [ + 154, + 324, + 459, + 364 + ], + "spans": [ + { + "bbox": [ + 154, + 324, + 459, + 364 + ], + "type": "table", + "html": "
MethodBackboneSegment sizeInference timefps# ParametersAverage mAP
OAT-OSN [15]TSN [34]64163.7ms6.1128.7M44.6
MATRTSN64167.1ms6.0192.8M49.5
MATRTSN1653.8ms18.6192.8M46.3
", + "image_path": "0be7ae2891cd75f8ce7dacd9f09848e79ca7c3dd2413685309acfb19d85681df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "text", + "content": "datasets. While the performance of OAT-OSN drops significantly from " + }, + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "inline_equation", + "content": "44.6\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "inline_equation", + "content": "25.8\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "text", + "content": " when the segment size is reduced from 64 to 8, the performance drop of ours is only " + }, + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "inline_equation", + "content": "9.1\\%" + }, + { + "bbox": [ + 130, + 387, + 481, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 532 + ], + "type": "text", + "content": "Analysis of Start and End Prediction Heads. We investigate various offset prediction methods for the start prediction and the end prediction (Table 5). For the first method, start offset regression without region classification, the offset is normalized using the offset statistics from the training dataset. The second approach, which adopts region classification and offset regression for the start prediction, and only offset regression for the end prediction, exhibits the best performance. The reason for performance drop when using region classification for the end prediction is that frame-wise end classification is a complex problem, adversely affects the performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "Analysis of Memory Compression Factor. We also conduct an experiment on efficiently storing input segment features in the memory by compressing them. We utilize a 1D convolution layer to compress the number of segment features. As shown in Table 6, " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": " compression yields competitive results with efficient memory usage. However, we decide not to use this method due to the efficient utilization of memory already achieved through the flag token and avoid to use additional hyperparameters that require optimization for each dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Inference Time. As shown in Table 7, when using the same backbone and the segment size, the inference time of MATR is almost equal to that of OAT-OSN, yet its average mAP is significantly better. Moreover, by reducing the segment length to 16, MATR outperforms OAT-OSN in terms of both mAP and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 90, + 447, + 102 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 261 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 261 + ], + "type": "image", + "image_path": "85c2b19333401bd42aaf9ce423d6ced38564b16eeef31d5b4daebd5add361f7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of MATR and OAT-OSN [15] on the THUMOS14. Generated time is the timestamp when the predicted instance is generated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 341 + ], + "type": "text", + "content": "the inference time. MATR uses more parameters due to the separation of the decoder into the start decoder and the end decoders." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 360, + 258, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 258, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 258, + 372 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": "Fig. 5 illustrates the predictions of our model and OAT-OSN [15]. The results show that MATR has the capability to efficiently detect action instances by identifying intervals immediately after the current action ends. Additionally, the results show that our model is able to classify and localize action instances more precisely than OAT-OSN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "content": "In this paper, we address online temporal action localization by introducing a new model, MATR, which leverages the memory queue to exploit long-term context. We also propose a new action instance localization method based on Transformer, which separately identifies action starts and action ends. As a result, MATR not only surpasses existing online TAL methods but also demonstrates comparable performance to offline TAL methods. The code base will be open to the public to promote future research on On-TAL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Limitations: When the memory queue contains multiple action instances, there is a risk of action instances being matched with incorrect start points. Additionally, when storing the input segment into the memory queue, MATR only considers whether the input segment is relevant to action instances but does not leverage the relationship with the past context stored in the memory. To address these issues, investigating a method to utilize the information stored in the memory queue when storing the input segment features, could lead to more effective memory storage and utilization." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the NRF grant and the IITP grant funded by Ministry of Science and ICT, Korea (RS-2019-II191906, RS-2021-II212068, RS-2022-II220290, NRF-2021R1A2C3012728)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 168, + 198, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 198, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 198, + 181 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 190, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "type": "text", + "content": "1. Buch, S., Escorcia, V., Shen, C., Ghanem, B., Carlos Niebles, J.: Sst: Single-stream temporal action proposals. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2911-2920 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "2. Cao, S., Luo, W., Wang, B., Zhang, W., Ma, L.: E2e-load: End-to-end long-form online action detection. Proc. IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "text", + "content": "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 351 + ], + "type": "text", + "content": "5. Chen, P., Gan, C., Shen, G., Huang, W., Zeng, R., Tan, M.: Relation attention for temporal action localization. IEEE Transactions on Multimedia 22(10), 2723-2733 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "type": "text", + "content": "6. Cheng, F., Bertasius, G.: Tallformer: Temporal action localization with a long-memory transformer. In: Proc. European Conference on Computer Vision (ECCV). pp. 503-521. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 416 + ], + "type": "text", + "content": "7. De Geest, R., Gavves, E., Ghodrati, A., Li, Z., Snoek, C., Tuytelaars, T.: Online action detection. In: Proc. European Conference on Computer Vision (ECCV). pp. 269-284. Springer (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 480, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 470 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. International Conference on Learning Representations (ICLR) (2021), https://openreview.net/forum?id=YicbFdNTTy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "text", + "content": "9. Escorcia, V., Caba Heilbron, F., Niebles, J.C., Ghanem, B.: Daps: Deep action proposals for action understanding. In: Proc. European Conference on Computer Vision (ECCV). pp. 768-784. Springer (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "type": "text", + "content": "0. Gao, J., Yang, Z., Chen, K., Sun, C., Nevatia, R.: Turn tap: Temporal unit regression network for temporal action proposals. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3628-3636 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "1. Gao, M., Xu, M., Davis, L.S., Socher, R., Xiong, C.: Startnet: Online detection of action start in untrimmed videos. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5542-5551 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "2. Heilbron, F.C., Niebles, J.C., Ghanem, B.: Fast temporal activity proposals for efficient detection of human actions in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1914-1923 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "3. Jiang, Y.G., Liu, J., Roshan Zamir, A., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/ (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Kang, H., Kim, K., Ko, Y., Kim, S.J.: Cag-qil: Context-aware actionness grouping via q imitation learning for online temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13729-13738 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Kim, Y.H., Kang, H., Kim, S.J.: A sliding window scheme for online temporal action localization. In: Proc. European Conference on Computer Vision (ECCV). pp. 653-669. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "16. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Proc. International Conference on Learning Representations (ICLR) (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "type": "text", + "content": "17. Lin, C., Xu, C., Luo, D., Wang, Y., Tai, Y., Wang, C., Li, J., Huang, F., Fu, Y.: Learning salient boundary feature for anchor-free temporal action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3320-3329 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "text", + "content": "18. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 3889-3898 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "19. Lin, T., Zhao, X., Su, H., Wang, C., Yang, M.: Bsn: Boundary sensitive network for temporal action proposal generation. In: Proc. European Conference on Computer Vision (ECCV). pp. 3-19 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "21. Liu, Q., Wang, Z.: Progressive boundary refinement network for temporal action detection. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 11612-11619 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "22. Liu, X., Hu, Y., Bai, S., Ding, F., Bai, X., Torr, P.H.: Multi-shot temporal event localization: a benchmark. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12596-12606 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "23. Liu, X., Wang, Q., Hu, Y., Tang, X., Zhang, S., Bai, S., Bai, X.: End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing 31, 5427-5441 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "24. Liu, Y., Ma, L., Zhang, Y., Liu, W., Chang, S.F.: Multi-granularity generator for temporal action proposal. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3604-3613 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "25. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Gaussian temporal awareness networks for action localization. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 344-353 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "27. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18857-18866 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "type": "text", + "content": "28. Shou, Z., Chan, J., Zareian, A., Miyazawa, K., Chang, S.F.: Cdc: Convolutional-de-convolutional networks for precise temporal action localization in untrimmed videos. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5734-5743 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "29. Shou, Z., Pan, J., Chan, J., Miyazawa, K., Mansour, H., Vetro, A., Giro-i Nieto, X., Chang, S.F.: Online detection of action start in untrimmed, streaming videos. In: Proc. European Conference on Computer Vision (ECCV). pp. 534-551 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "30. Shou, Z., Wang, D., Chang, S.F.: Temporal action localization in untrimmed videos via multi-stage cnns. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1049-1058 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "31. Tang, T.N., Park, J., Kim, K., Sohn, K.: Simon: A simple framework for online temporal action localization. arXiv preprint arXiv:2211.04905 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "32. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proc. Advances in Neural Information Processing Systems (NeurIPS). pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "33. Wang, J., Chen, G., Huang, Y., Wang, L., Lu, T.: Memory-and-anticipation transformer for online action understanding. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13824-13835 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "34. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proc. European Conference on Computer Vision (ECCV). pp. 20-36. Springer (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "type": "text", + "content": "35. Wang, X., Zhang, S., Qing, Z., Shao, Y., Zuo, Z., Gao, C., Sang, N.: Oadtr: Online action detection with transformers. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7565-7575 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 295 + ], + "type": "text", + "content": "36. Wu, C.Y., Li, Y., Mangalam, K., Fan, H., Xiong, B., Malik, J., Feichtenhofer, C.: Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13587-13597 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "text", + "content": "37. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10186-10195 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "text", + "content": "38. Xu, M., Perez Rua, J.M., Zhu, X., Ghanem, B., Martinez, B.: Low-fidelity video encoder optimization for temporal action localization. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 9923-9935 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "39. Xu, M., Zhao, C., Rojas, D.S., Thabet, A., Ghanem, B.: G-tad: Sub-graph localization for temporal action detection. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10156-10165 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "40. Xu, M., Gao, M., Chen, Y.T., Davis, L.S., Crandall, D.J.: Temporal recurrent networks for online action detection. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 5532-5541 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "text", + "content": "41. Xu, M., Xiong, Y., Chen, H., Li, X., Xia, W., Tu, Z., Soatto, S.: Long shortterm transformer for online action detection. Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 1086-1099 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "text", + "content": "42. Yang, L., Han, J., Zhang, D.: Colar: Effective and efficient online action detection by consulting exemplars. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3160-3169 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "43. Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 7094-7103 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "text", + "content": "44. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 492-510. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "45. Zhao, C., Thabet, A.K., Ghanem, B.: Video self-stitching graph network for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13658-13667 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "46. Zhao, Y., Krahenbuhl, P.: Real-time online video detection with temporal smoothing transformers. In: Proc. European Conference on Computer Vision (ECCV). pp. 485-502. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "47. Zhao, Y., Xiong, Y., Wang, L., Wu, Z., Tang, X., Lin, D.: Temporal action detection with structured segment networks. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 2914-2923 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 136, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Temporal Action Localization with Memory-Augmented Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., Ren, D.: Distance-iou loss: Faster and better learning for bounding box regression. In: Proc. AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 12993-13000 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "49. Zhu, Z., Tang, W., Wang, L., Zheng, N., Hua, G.: Enriching local and global contexts for temporal action localization. In: Proc. IEEE International Conference on Computer Vision (ICCV). pp. 13516-13525 (2021)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Youngkil Song, Dongkeun Kim, Minsu Cho, and Suha Kwak" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_content_list.json b/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..824da8c22ee010c790b8acfd1f8a25a384428b90 --- /dev/null +++ b/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_content_list.json @@ -0,0 +1,1850 @@ +[ + { + "type": "text", + "text": "Online Vectorized HD Map Construction using Geometry", + "text_level": 1, + "bbox": [ + 233, + 138, + 769, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhixin Zhang $^{1}$ Yiyuan Zhang $^{2}$ Xiaohan Ding $^{3}$ Fusheng Jin $^{1*}$ Xiangyu Yue $^{2}$", + "bbox": [ + 218, + 210, + 782, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Beijing Institute of Technology \n $^{2}$ The Chinese University of Hong Kong \n $^{3}$ Tencent AI Lab \nzhangzhixin@bit.edu.cn, yiyuanzhang.ai@gmail.com \nhttps://invictus717.github.io/GeMap/", + "bbox": [ + 243, + 239, + 756, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Online vectorized High-Definition (HD) map construction is critical for downstream prediction and planning. Recent efforts have built strong baselines for this task, however, geometric shapes and relations of instances in road systems are still under-explored, such as parallelism, perpendicular, rectangle-shape, etc. In our work, we propose GeMap (Geometry Map), which end-to-end learns Euclidean shapes and relations of map instances beyond fundamental perception. Specifically, we design a geometric loss based on angle and magnitude clues, robust to rigid transformations of driving scenarios. To address the limitations of the vanilla attention mechanism in learning geometry, we propose to decouple self-attention to handle Euclidean shapes and relations independently. GeMap achieves new state-of-the-art performance on the nuScenes and Argoverse 2 datasets. Remarkably, it reaches a $71.8\\%$ mAP on the large-scale Argoverse 2 dataset, outperforming MapTRv2 by $+4.4\\%$ and surpassing the $70\\%$ mAP threshold for the first time. Code is available at https://github.com/cnzzx/GeMap.", + "bbox": [ + 261, + 337, + 738, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: HD Map Construction $\\cdot$ Geometry Representation $\\cdot$ Geometry-Decoupled Attention", + "bbox": [ + 261, + 573, + 756, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 628, + 374, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vectorized HD maps provide structured environmental information for autonomous vehicles and have been widely adopted in downstream tasks, such as trajectory forecasting [3,21,55] and planning [6,37]. Online Vectorized HD Map Construction can significantly reduce the need for labor-intensive annotations and facilitate real-time updates in autonomous driving [4,18,22,53].", + "bbox": [ + 217, + 662, + 799, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of Bird's-Eye-View (BEV) representation and multimodal learning [5, 11, 19, 28, 50–52], online HD map construction has achieved significant advancements [10, 20, 23]. Early works [14, 32, 35, 47] formulate HD map construction as a dense prediction task. However, these methods generate maps in image format, which is redundant for representing sparse map instances. Then,", + "bbox": [ + 217, + 739, + 785, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 824, + 380, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92ca623e88b2fd84784cd2f0faeff2fb84f6d63fc8b5e7df9c1904839468bbca.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 241, + 146, + 485, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9315f629727cff8d2588bc14c9ea6f6d592cb5f056478a1c96a75230a01dd8e1.jpg", + "image_caption": [ + "(b)", + "Fig. 1: Geometric Invariance. (a) As the ego vehicle moves, after rotation $(\\leftarrow)$ and translation $(\\rightarrow)$ , the shape of the crossing and the parallelism between lanes remain unchanged, which indicates the invariant property of geometry to rigid transformations. (b) Absolute coordinates are vulnerable to rotation and translation, however, our G-representation is invariant, which is more suitable to capture geometric properties." + ], + "image_footnote": [], + "bbox": [ + 493, + 143, + 767, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a more compact map formulation was introduced to minimize redundancy, albeit at the cost of adding time-consuming post-processing steps [18]. In response to this, recent works [4, 22, 26, 36] attempt to end-to-end construct vectorized HD maps to avoid extensive post-processing. These methods typically sample points from map instances and represent each instance as a polyline [4, 22, 26] or parameterized curve [36].", + "bbox": [ + 212, + 411, + 787, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observe that transportation road systems exhibit significant geometric characteristics (Figure 2a), such as parallel lanes, perpendicular crossings, equal lane widths, etc. However, these geometric properties of shapes and relations between map instances have not been fully explored. Meanwhile, there are two notable limitations among existing methods [4,22,26,36] which can be alleviated by leveraging these geometric properties. 1) Excessive dependency on absolute coordinates: as the ego-vehicle moves, instances experience rotations and translations, as depicted in Figure 1a. However, widely adopted representations such as polylines and parameterized curves [8,36] are inherently sensitive to rotation and translation changes. 2) Objective conflicts of vanilla attention mechanism (detail in § 3.5) struggles to learn diverse shapes and relations, even though geometric properties such as rectangle shape, parallelism, and perpendicular relationships are commonly found in driving scenarios, as illustrated in Figure 2a. We believe that incorporating these geometric properties significantly enhances the precision and efficiency of online HD map construction.", + "bbox": [ + 212, + 505, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above two limitations, we introduce a novel geometric representation that captures shapes of individual map instances and relations between different instances as illustrated in Figure 2b and 2c, referred to as G-Representation. It enhances the vanilla representation of map instances by incorporating a translation- and rotation-invariant representation that effectively leverages instance geometry. The local structures of map features are encoded using displacement vectors. These vectors are computed from the absolute coor", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/07f2e2057c54d24fd03550135c8637d1125bb1c2b21b7de3e8b2120739a04938.jpg", + "image_caption": [ + "Rectangle-shaped" + ], + "image_footnote": [], + "bbox": [ + 256, + 147, + 385, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cd70cba1aee6886e85e2ca7b39a9d9d9f03aecd4bf3fe395749cfcd4fe399f55.jpg", + "image_caption": [ + "Parallelism" + ], + "image_footnote": [], + "bbox": [ + 405, + 145, + 491, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/827a7e3e5aea6c3fd98b4a49a721fb3992a6ada01dd99a0fc93d5c9ad1a2777a.jpg", + "image_caption": [ + "Perpendicular" + ], + "image_footnote": [], + "bbox": [ + 506, + 152, + 617, + 219 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b99d966c30efcbee61524067991e4180597ec7a55fca6056c5f3012b6dfeb21c.jpg", + "image_caption": [ + "Equal Lane Width" + ], + "image_footnote": [], + "bbox": [ + 629, + 154, + 751, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48b1d99a7f95457c0eac6ee3ba5b37f8ec8fcee0e5b0da77fe9af0fcca134dce.jpg", + "image_caption": [ + "(a)", + "Single Map Instance", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 251, + 263, + 496, + 321 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4e59fee30e6ad9c38c4c1f05a4b71b84719fb92349c0d4454851400ba57daad.jpg", + "image_caption": [ + "Cross Map Instances", + "(c)", + "Fig. 2: Geometric properties and G-Representation. (a) Geometry in the transportation road system. (b) (c) We propose to model the geometric properties of a single map instance and multiple instances, with magnitude $d$ and angle $\\alpha$ ." + ], + "image_footnote": [], + "bbox": [ + 521, + 258, + 617, + 325 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e21ada3d7040117a4a781a088eff69fc8da64d4715a51785f479701d03750b9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 265, + 738, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dinates of polyline points to effectively represent the relative positions and orientations of adjacent points. To quantify these features within Euclidean space, we employ both the magnitudes of the displacement vectors and the angles formed between these vectors. Based on G-Representation, we propose the Euclidean Shape Clues to represent the shapes of individual map instances and the Euclidean Relation Clues to represent the inter-instance relations in Euclidean space.", + "bbox": [ + 212, + 444, + 787, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "G-Representation is a simple yet effective method to address the abovementioned two limitations: 1) Geometric Invariance, by concentrating on the relative relationships between points within and across map instances, it inherently attains translation and rotation invariance as illustrated in Figure 1b. This enhances its robustness against variations in data collection and equips it to effectively handle different coordinate systems. 2) Euclidean Modeling, it captures the inherent geometry in transportation road systems. Diverse shapes and relation geometry can be simplified as enumerations of magnitudes and directions. For instance, in Euclidean space, parallel lanes can be easily modeled as approximate directions $(\\alpha \\approx 0)$ .", + "bbox": [ + 212, + 551, + 787, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon G-Representation, we propose a framework named GeMap for HD map construction. A BEV encoder is used to extract features from multiview input images, while a geometry-decoupled decoder is employed to focus on geometric aspects. Specifically, we adapt the attention mechanism [42] and propose Geometry-Decoupled Attention (GDA). GDA sequentially applies attention to queries belonging to the same instance and attention to queries across different instances. This can significantly boost the geometry learning of key points of shapes and relations between map instances. Furthermore, we propose an objective function named Euclidean Loss to optimize G-Representation.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 348, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we transform the conventional polylines of the ground truth map into our G-Representation. In this way, the model gets optimized to better understand the magnitude of displacement vectors and the angles between them, thereby facilitating a more effective learning of geometric properties.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experiments on nuScenes and Argoverse 2 datasets demonstrate the effectiveness of GeMap. We reach new state-of-the-art performances on both datasets. With camera images only, GeMap achieves $69.4\\%$ and $71.8\\%$ mAP on the nuScenes and Argovserse 2, respectively. Visualization results (§ 4.4) further demonstrate the better perception of shape and relation, alleviation of occlusion, and robustness to rigid transformations of GeMap.", + "bbox": [ + 212, + 207, + 792, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our contributions are summarized as the following:", + "bbox": [ + 238, + 297, + 609, + 313 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose G-Representation which harnesses critical geometric properties of rotation and translation invariance in autonomous driving scenarios, opening up new research avenues within the field.", + "- We introduce GeMap, a novel framework incorporating geometry-decoupled attention and Euclidean Loss function, specifically designed to learn the intrinsic geometry of online HD maps.", + "- GeMap achieves new state-of-the-art results in HD map construction on the nuScenes and Argoverse 2 datasets, notably surpassing the $70\\%$ mAP on the large-scale Argoverse 2 dataset for the first time." + ], + "bbox": [ + 223, + 325, + 782, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 488, + 387, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Online HD Map Construction", + "text_level": 1, + "bbox": [ + 215, + 521, + 508, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditionally, HD map construction has required labor-intensive manual or semi-automatic annotations [14,33]. To streamline this, recent studies [20,31,34,35,54] have focused on online construction, approaching HD maps as a dense prediction challenge. Innovations such as MetaBEV [9] aim to mitigate sensor issues, while MVNet [45] uses historical data for improved semantic consistency. The trend towards automatic vectorization of HD maps is spearheaded by works such as HDMapNet [18], which fuses camera and LiDAR inputs in BEV space and is advanced by end-to-end solutions such as VectorMapNet [26] and MapTR [22], leveraging Transformer-based models. StreamMapNet [48] and PivotNet [4] build upon this, modifying attention mechanisms for better performance. Our contribution starts from the strengths of the Transformer architecture, optimizing it with a decoupled self-attention block for enhanced geometric processing.", + "bbox": [ + 212, + 547, + 787, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Cross-view BEV Learning", + "text_level": 1, + "bbox": [ + 215, + 753, + 477, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The conversion of Perspective View (PV) camera images into a unified BEV space is a significant challenge for autonomous driving systems. Previous studies such as [19,35] use depth estimation from monocular images for this transformation, others [2, 20, 28] have developed PV-to-BEV conversion methods without", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "explicit depths. However, these methods can distort shapes and geometric relations, leading to inaccuracies. Our approach integrates geometric supervision in prediction, addressing these geometric inconsistencies while acknowledging the value of depth information.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Geometric Instance Modeling", + "text_level": 1, + "bbox": [ + 214, + 229, + 508, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional vectorization approaches model these instances as polylines [22, 26, 38, 43] or polynomial curves [7, 36, 39, 41], with some incorporating Bézier Curves for improved fitting [7, 36]. However, such methods often overlook geometric properties including shapes, parallelism, perpendicular, and etc. In 3D Lane Detection, some works [13, 17, 25] also explore simple geometry priors of lanes such as \"equal lane width\". However, they are not generalizable to more complex HD map instances. Existing works on modeling geometry properties of map instances mainly address rasterization-based shape loss [41, 49] or edge loss [22], and they fail to preserve cross-instance relationships. PivotNet [4] offers an architectural advancement with its line-aware point decoder, yet it still does not fully address the complexity of instance geometry.", + "bbox": [ + 212, + 255, + 787, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the current under-exploration of diverse geometric properties in shape and relation, we propose utilizing Euclidean shape and relation losses and decoupling the traditional self-attention module to empower the model with a more robust understanding of instance geometry.", + "bbox": [ + 212, + 422, + 785, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 506, + 330, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 214, + 537, + 361, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The architecture of our method is illustrated in Figure 3. Input images are denoted by $\\mathbb{I} = \\{\\pmb{I}_i\\}_{i=1}^{N_c}$ , where $\\pmb{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}$ and $N_c$ is the number of cameras. The output is a set of $N$ map instances $\\mathbb{M} = \\{\\pmb{L}_i\\}_{i=1}^N$ , where each map instance is represented by a polyline $\\pmb{L}_i \\in \\mathbb{R}^{N_v \\times 2}$ , i.e., an ordered sequence of $N_v$ two-dimensional points. As shown in Figure 3, the commonly adopted pipeline comprises three steps. 1) A BEV feature extractor processes the multi-view images. 2) The extracted BEV features are fed into a map decoder, which predicts the map instances based on the BEV features. 3) To optimize the pipeline, the predicted map instances are compared against the ground truth. Beyond point-to-point comparison, G-representations of predictions and the ground truth are computed and the difference is measured by $L^1$ loss.", + "bbox": [ + 212, + 564, + 787, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Architecture Overview", + "text_level": 1, + "bbox": [ + 214, + 753, + 450, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BEV Feature Extractor. A shared vision backbone in different views and parameterized PV-to-BEV transformation network are employed to aggregate features from various perspectives. In our default configuration, we utilize ResNet [12] as the PV backbone and GKT [2] as the PV-to-BEV transformation network.", + "bbox": [ + 212, + 779, + 792, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 348, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2157971095d7ce5ba6b9ed1002ad3c234a091c316a7ad66e8e84f682244c9602.jpg", + "image_caption": [ + "Fig. 3: Illustration of GeMap. First, PV images are transformed into BEV features, then a Geometry-Decoupled Decoder outputs the vectorized HD Map. In each block of the decoder, queries are first processed by Euclidean shape and relation attention, which focuses on geometric relevance. Finally, predictions are enhanced in G-Representations by shape and relation constraint." + ], + "image_footnote": [], + "bbox": [ + 254, + 143, + 750, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometry-Decoupled Decoder. We adopt a Transformer decoder to predict polylines. The decoder obtains $N \\times N_v$ queries that represent points and processes them via self-attention and aggregate BEV features via deformable cross-attention [56]. Finally, a prediction head is used to convert queries to polylines $\\{\\hat{L}_i\\}_{i=1}^N$ . We utilize a multi-layer perceptron as the polyline prediction head.", + "bbox": [ + 212, + 506, + 787, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometric Loss. For training, we convert both the predicted polylines and ground truth into the proposed G-Representation. Based on that, we let the model optimize the predicted magnitudes of displacement vectors and angles between displacement vectors to match the ground truth.", + "bbox": [ + 212, + 582, + 787, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following subsections, we first introduce Geometric Representation in § 3.3 since it is the core of our framework. Based on that, we introduce Geometric Loss in § 3.4 and Geometry-Decoupled Decoder in § 3.5, respectively.", + "bbox": [ + 212, + 642, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Geometric Representation", + "text_level": 1, + "bbox": [ + 214, + 707, + 478, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Euclidean Shape Clues. We first introduce the representation of shapes of individual map instances. For each instance, we describe the local geometry with displacement vectors between neighboring points which are computed as:", + "bbox": [ + 212, + 729, + 787, + 776 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {u} ^ {i} = \\boldsymbol {L} _ {i, u + 1} - \\boldsymbol {L} _ {i, u} \\quad (u \\in \\{1, 2, \\dots , N _ {v} \\}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 784, + 785, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we define $\\pmb{L}_{i,N_v + 1} \\coloneqq \\pmb{L}_{i,1}$ to unify the geometric formulation of closed and open polylines.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fbcdb00e68bbf3e8240394691f1b3d7948f085f310d4230157ea632eb96eea42.jpg", + "image_caption": [ + "Fig. 4: Euclidean Shape Clues. Magnitudes of displacement vectors and angles between neighboring vectors indicate shape clues and are utilized to compute shape loss. The right part shows how to connect Euclidean Shape Clues to shape geometry." + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 743, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These displacement vectors are sufficient to represent the shape of a map instance and such a representation is invariant to translation transformations. However, we would like to note that this representation is vulnerable to rotations, which might prevent the model from learning robust geometry. To solve this problem, we propose to represent the shape with magnitudes of displacement vectors and angles between consecutive displacement vectors, as illustrated in Figure 4. Specifically, the $i$ -th instance is denoted by $N_v$ angle values and $N_v$ magnitude values. Let $u$ be the index, the $u$ -th angle value and the magnitude value of the $i$ -th instance are computed as:", + "bbox": [ + 212, + 303, + 784, + 439 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {u} ^ {i} = \\left\\langle \\boldsymbol {v} _ {u} ^ {i}, \\boldsymbol {v} _ {u + 1} ^ {i} \\right\\rangle , d _ {u} ^ {i} = \\left\\| \\boldsymbol {v} _ {u} ^ {i} \\right\\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 446, + 784, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\langle \\cdot \\rangle$ denotes the angle between two vectors and similarly $\\pmb{v}_{N_v + 1}^i \\coloneqq \\pmb{v}_1^i$ .", + "bbox": [ + 212, + 470, + 756, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The proposed representation inherently captures common geometric patterns, such as parallelism, right angles, and proper line width, by translating them into corresponding numerical patterns within this representation. For example, in Figure 4, a rectangle is characterized by one $90^{\\circ}$ angle and equal magnitudes of opposite displacement vectors; in addition, a straight line is represented by $0^{\\circ}$ angles. Beyond discussed regular cases, the Euclidean shape clues are capable of handling more complex shapes, as illustrated in scenarios (a) and (c) of Figure 6 and discussed in § 4.4.", + "bbox": [ + 212, + 487, + 785, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Euclidean Relation Clues. Having highlighted the translation and rotation invariance of G-Representation and its advantages in representing the geometric shapes of individual map instances, we further introduce its ability to capture the relations between two map instances, e.g., parallelism and perpendicular, in Euclidean space. Specifically, given the vanilla representations and displacement vectors of the $i$ -th and $j$ -th map instance, we represent the relation between instance $i$ and $j$ with 1) angles between each pair of their respective displacement vectors, and 2) magnitudes of displacement vectors between each pair of points, as illustrated in Figure 5. Formally,", + "bbox": [ + 212, + 626, + 782, + 762 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {u, v} ^ {i, j} = \\left\\langle \\boldsymbol {v} _ {u} ^ {i}, \\boldsymbol {v} _ {v} ^ {j} \\right\\rangle , d _ {u, v} ^ {i, j} = \\left\\| \\boldsymbol {L} _ {i, u} - \\boldsymbol {L} _ {j, v} \\right\\| _ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 768, + 784, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This representation is also translation- and rotation-invariant. Similar to Euclidean Shape Clues, it inherently captures common relations between map instances by translating them into corresponding numerical patterns. For example,", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 348, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9982cddd7e67134a42fafcf015de569f554aea3ee509115cea600b499053cfa1.jpg", + "image_caption": [ + "Fig. 5: Euclidean Relation Clues. Angles between pairs of displacement vectors on different polylines, and magnitudes of displacement vectors between point pairs indicate relation clues." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 743, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 shows that the perpendicular relation can be directly represented by a $90^{\\circ}$ angle, and the distance between two parallel polylines, which may correspond to the width of a lane, is naturally represented by magnitudes of displacement vector between point pairs. Scenarios (a) and (b) of Figure 6 show the effectiveness of Euclidean Relation Clues and more discussion can be viewed in $\\S 4.4$ .", + "bbox": [ + 212, + 327, + 782, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Euclidean Loss and Objectives", + "text_level": 1, + "bbox": [ + 214, + 422, + 511, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We transform both the ground truth data and the model's predicted polylines from their original format into G-Representation. This conversion allows the model to independently optimize the angles between displacement vectors and the magnitudes of displacement vectors to more accurately align with the ground truth. The proposed Euclidean Loss is composed of two parts that measure how accurately the model predicts the shape of individual map instances and the inter-instance relations, respectively, which are denoted by $\\mathcal{L}_{\\mathrm{shp}}$ and $\\mathcal{L}_{\\mathrm{rel}}$ . The Euclidean Loss can be computed as the following:", + "bbox": [ + 212, + 446, + 787, + 568 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E u c}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {s h p}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {r e l}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 578, + 785, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters, whose effects on the model performances are evaluated in Figure 7a. Specifically,", + "bbox": [ + 212, + 604, + 785, + 635 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} w _ {i, j} = 1 - \\left(\\min _ {u, v} \\left\\{d _ {u, v} ^ {i, j} \\right\\} / \\sqrt {2}\\right) ^ {p}, \\quad \\mathcal {L} _ {\\mathrm {s h p}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N _ {v}} \\left| \\hat {d} _ {j} ^ {i} - d _ {j} ^ {i} \\right| + \\ell \\left(\\hat {\\alpha} _ {j} ^ {i}, \\alpha_ {j} ^ {i}\\right), \\tag {5} \\\\ \\mathcal {L} _ {\\mathrm {r e l}} = \\sum_ {i = 1} ^ {N} \\sum_ {j > i} ^ {N} w _ {i, j} \\sum_ {u = 1} ^ {N _ {v}} \\sum_ {v = 1} ^ {N _ {v}} \\left(| \\hat {d} _ {u, v} ^ {i, j} - d _ {u, v} ^ {i, j} | + \\ell (\\hat {\\alpha} _ {u, v} ^ {i, j}, \\alpha_ {u, v} ^ {i, j})\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 643, + 785, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\ell(\\cdot)$ is a function based on $L^1$ loss. We avoid using the inverse trigonometric function to directly compute the angles, but use sine and cosine values instead:", + "bbox": [ + 212, + 742, + 782, + 773 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (\\hat {\\alpha}, \\alpha) = | \\cos (\\hat {\\alpha}) - \\cos (\\alpha) | + | \\sin (\\hat {\\alpha}) - \\sin (\\alpha) |. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 782, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, the distance might influence the relation strength of instance pairs. For example, if two instances are far from each other, the relation between them", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "might be weak. For this reason, we further adopt $w_{i,j}$ to punish weakly related instance pairs and discuss it in § 4.5. According to the best experimental results, we treat all instance pairs equally in other experiments.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following common practices [22,48], we also use focal loss [24] $\\mathcal{L}_{\\mathrm{cls}}$ for classification. For polyline regression, we adopt point-to-point loss and edge direction loss, denoted by $\\mathcal{L}_{\\mathrm{pts}}$ and $\\mathcal{L}_{\\mathrm{dir}}$ respectively. Following [23,36], we further adopt segmentation loss $\\mathcal{L}_{\\mathrm{seg}}$ and depth estimation loss $\\mathcal{L}_{\\mathrm{dep}}$ . These losses are detailed in appendix § A.1. The overall loss function can be written as:", + "bbox": [ + 212, + 191, + 784, + 267 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\cdot \\mathcal {L} _ {\\text {E u c}} + \\beta_ {1} \\cdot \\mathcal {L} _ {\\text {c l s}} + \\beta_ {2} \\cdot \\mathcal {L} _ {\\text {p t s}} + \\beta_ {3} \\cdot \\mathcal {L} _ {\\text {d i r}} + \\beta_ {4} \\cdot \\mathcal {L} _ {\\text {s e g}} + \\beta_ {5} \\cdot \\mathcal {L} _ {\\text {d e p}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 273, + 784, + 290 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where hyperparameters are discussed in Figure 7b and appendix Table A1.", + "bbox": [ + 214, + 296, + 751, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Geometry-Decoupled Decoder", + "text_level": 1, + "bbox": [ + 214, + 330, + 511, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the Transformer decoder, we obtain the $N \\times N_v$ fused queries by adding $N$ instance queries to $N_v$ point queries, which is suggested in [22]. We denote input tokens by $\\pmb{E} \\in \\mathbb{R}^{N_A \\times D}$ , where $D$ is the feature dimension and $N_A = N \\cdot N_v$ is the number of fused queries. Intuitively, each fused query corresponds to a point on a predicted map instance. Self-Attention (SA) is formulated by:", + "bbox": [ + 212, + 349, + 782, + 425 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} (\\boldsymbol {E}, \\boldsymbol {M}) = \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\boldsymbol {E} \\boldsymbol {W} ^ {q}\\right) \\left(\\boldsymbol {E} \\boldsymbol {W} ^ {k}\\right) ^ {\\top}}{\\sqrt {D _ {k}}} \\odot \\boldsymbol {M}\\right) \\boldsymbol {E} \\boldsymbol {W} ^ {v}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 431, + 784, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\pmb{W}^q, \\pmb{W}^k, \\pmb{W}^v \\in \\mathbb{R}^{D \\times D_k}$ are linear projection matrices, $\\pmb{M} \\in \\mathbb{R}^{N_A \\times N_A}$ is the attention mask, and $\\odot$ is the Hadamard product. The vanilla SA computes relations between every pair of tokens.", + "bbox": [ + 212, + 470, + 782, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nevertheless, the geometry of shape and relation pertains to distinct subsets of tokens. For any given map instance, its shape is intimately related to tokens representing that instance's points. Precise shape geometry capture requires the model to discern token correlations specific to an instance while avoiding interference from tokens of unrelated instances. Conversely, for relation geometry modeling, it is beneficial to isolate token correlations that span across different instances, rather than those confined within a single instance.", + "bbox": [ + 212, + 518, + 784, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Therefore, we present Geometry-Decoupled Attention (GDA). First, we multiply a binary mask $M$ to the computed attention map so that the tokens of a map instance are aggregated according to the tokens within the same map instances only, which allows the model to adjust points' positions according to learned shape geometry. Denote the index of map instance that the $i$ -th token belongs to as $\\mathcal{I}_i$ . For example, tokens $1,2,\\ldots,N_v$ belong to the first instance so that $\\mathcal{I}_1,\\mathcal{I}_2,\\ldots,\\mathcal{I}_{N_v}=1$ . Tokens $N_v+1,N_v+2,\\ldots,2\\cdot N_v$ belong to the second instance so that $\\mathcal{I}_{N_v+1},\\mathcal{I}_{N_v+2},\\ldots,\\mathcal{I}_{2N_v}=2$ . With this notation, the binary mask $M$ can be simply constructed by:", + "bbox": [ + 212, + 623, + 784, + 758 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {i, j} ^ {\\mathrm {s h p}} = \\left\\{ \\begin{array}{l l} 1 & , \\mathcal {I} _ {i} = \\mathcal {I} _ {j} \\\\ 0 & , \\mathcal {I} _ {i} \\neq \\mathcal {I} _ {j} \\end{array} , \\quad \\boldsymbol {M} _ {i, j} ^ {\\mathrm {r e l}} = \\left\\{ \\begin{array}{l l} 1 & , \\mathcal {I} _ {i} \\neq \\mathcal {I} _ {j} \\\\ 0 & , \\mathcal {I} _ {i} = \\mathcal {I} _ {j} \\end{array} . \\right. \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 766, + 784, + 805 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The second attention is expected to model relations between tokens of different map instances.", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 348, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5ca897c04608f8b761da6ea79714f9be443a6b0cd3ed96cef312d96f10be2d84.jpg", + "table_caption": [ + "Table 1: Comparison on the nuScenes dataset, GeMap reaches a new state-of-the-art performance. \"EB0\", \"R50\", \"PP\", \"Sec\", \"Swin-T\", and \"V2-99\" denote EfficientNet-B0 [40], ResNet50 [12], PointPillars [15], SECOND [46], Swin Transformer Tiny [27], and VoVNetV2-99 [16] respectively. Methods with two backbones utilize both camera and LiDAR inputs. \"Dns. Loss\" denotes whether any dense prediction (e.g. semantic segmentation) loss is adopted. The best result is highlighted in **bold**. We reproduce all methods on a single RTX3090 GPU to test FPS for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneDns. Loss\\( AP_{div}(\\uparrow) \\)\\( AP_{ped}(\\uparrow) \\)\\( AP_{bnd}(\\uparrow) \\)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5042.551.444.146.05.3
MapTR [ICLR'23] [22]R5059.856.260.158.719.8
MapVR [NeurIPS'23] [49]R5061.855.059.458.819.8
GeMap [Ours]R5065.1 ↑ 3.359.8 ↑ 4.863.2 ↑ 3.862.7 ↑ 3.919.0
HDMapNet [ICRA'22] [18]EB0✓14.421.733.023.00.7
PivotNet [ICCV'23] [4]R50✓58.853.859.657.49.5
BeMapNet [CVPR'23] [36]R50✓66.762.665.164.86.6
MapTRv2 [Arxiv'23] [23]R50✓68.368.169.768.715.0
GeMap [Ours]R50✓69.8 ↑ 1.567.171.4 ↑ 1.769.4 ↑ 0.715.8
HDMapNet [ICRA'22] [18]EB0 & PP29.616.346.731.00.7
VectorMapNet [ICML'23] [26]R50 & PP60.148.253.053.7-
MapTR [ICLR'23] [22]R50 & Sec62.355.969.362.56.7
MapVR [NeurIPS'23] [49]R50 & PP62.760.467.263.5-
GeMap [Ours]R50 & Sec66.3 ↑ 3.662.2 ↑ 1.871.1 ↑ 3.966.5 ↑ 3.07.3
MapTRv2 [Arxiv'23] [23]R50 & Sec✓65.666.574.869.06.6
GeMap [Ours]R50 & Sec✓69.8 ↑ 4.268.0 ↑ 1.573.470.4 ↑ 1.46.8
MapTRv2 [Arxiv'23] [23]V2-99✓73.771.475.073.410.1
GeMap [Ours]Swin-T✓72.870.472.872.011.4
GeMap [Ours]V2-99✓76.0 ↑ 2.374.3 ↑ 2.977.7 ↑ 2.776.0 ↑ 2.610.8
", + "bbox": [ + 233, + 255, + 764, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 532, + 375, + 549 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setups", + "text_level": 1, + "bbox": [ + 215, + 574, + 434, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. To evaluate GeMap, we conduct experiments on the nuScenes dataset [1], a widely adopted large-scale autonomous driving dataset that includes 1,000 scenes captured by six RGB cameras with a 360-degree field of view, and provides precise annotations from LiDAR point clouds for HD map construction. For comparability with prior research [18,22], we focus on three static categories of map instances: pedestrian crossings, lane dividers, and road boundaries. With dataset splits provided by BeMapNet [36], we also evaluate GeMaps under three weather conditions: sunny, cloudy, and rainy. Additionally, we use the Argoverse 2 dataset [44] as another benchmark, which consists of approximately 108,000 frames, each providing images from seven cameras. We evaluate the performance by Average Precision (AP) [18,22] with thresholds at 0.5, 1.0, and 1.5 meters.", + "bbox": [ + 212, + 611, + 808, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. GeMap leverages 8 NVIDIA RTX 3090 GPUs for training. We adopt AdamW [30] as the optimizer and utilize Cosine Annealing with a linear warm-up phase [29] as the learning rate scheduler. For more details on hyperparameters, please refer to appendix § A.2.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5842196eb144cfc75ac3bae0387b5843ab6aa6722e0dd93ef05f43a2bcf9c6ab.jpg", + "table_caption": [ + "Table 2: Following BeMapNet [36], we also compare GeMap on the nuScenes dataset under different weather conditions." + ], + "table_footnote": [], + "table_body": "
MethodsBackbone Dns. LossmAPsun(↑)mAPcld(↑)mAPrny(↑)mAPavg(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5043.844.136.641.55.3
MapTR [ICLR'23] [22]R5062.160.552.858.419.8
GeMap [Ours]R5066.0 ↑ 3.964.3 ↑ 3.854.4 ↑ 1.661.5 ↑ 3.119.0
BeMapNet [CVPR'23] [36]R50 √67.367.556.663.86.6
GeMap [Ours]R50 √73.1 ↑ 5.871.0 ↑ 3.559.3 ↑ 2.767.8 ↑ 4.015.8
", + "bbox": [ + 233, + 184, + 763, + 268 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/065e5ef768ad58cb92af48647f11fbf41f3598f61d8abeb56dcea91693faf77e.jpg", + "table_caption": [ + "Table 3: Comparison on Argoverse 2 dataset. GeMap demonstrates significant performance improvements over previous methods." + ], + "table_footnote": [], + "table_body": "
MethodsBackbone Dns. Loss\\( AP_{div}(\\uparrow) \\)\\( AP_{ped}(\\uparrow) \\)\\( AP_{bnd}(\\uparrow) \\)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5036.138.339.237.9-
GeMap [Ours]R5067.6 ↑ 31.559.3 ↑ 21.064.7 ↑ 25.563.9 ↑ 26.016.7
HMapNet [ICRA'22] [18]EB0✓5.713.137.618.8
MapTRv2 [Arxiv'23] [23]R50✓72.162.967.167.4
GeMap [Ours]R50✓75.7 ↑ 3.669.2 ↑ 6.370.5 ↑ 3.471.8 ↑ 4.4
", + "bbox": [ + 233, + 330, + 763, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 450, + 372, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on nuScenes. As delineated in Table 1, GeMap delivers a state-of-the-art performance, achieving a mean Average Precision (mAP) of $69.4\\%$ using the camera-only setup and ResNet50 [12] backbone. GeMap particularly improves the precision for identifying dividers and boundaries, with enhancements of $+1.5\\%$ and $+1.7\\%$ , respectively. Importantly, these performance gains do not come at a high cost of efficiency, as GeMap sustains inference speeds on par with and even exceeds the previously established state-of-the-art, measured in frames per second (FPS). With more powerful vision backbones, GeMap can achieve significantly improved performance. Specifically, VoVNetV2-99 [16] and Swin-T [27] outperform the ResNet50 baseline by $+6.6\\%$ and $+2.6\\%$ in mAP, respectively. Also, GeMap achieves significantly improved performance with extra LiDAR inputs, which indicates the generalizability of GeMap to multi-modality settings. Moreover, we also compare performance under different weather conditions in Table 2. GeMap exceeds previous works under sunny, cloudy, and rainy scenarios, which shows the potential of geometry to raise model robustness to varied weather conditions.", + "bbox": [ + 212, + 501, + 785, + 743 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on Argoverse 2. Extending our evaluation to the Argoverse 2 dataset, as presented in Table 3, GeMap also presents a new SOTA performance of $71.8\\%$ mAP, outperforming the highly advanced MapTRv2 model by $+4.4\\%$ . This achievement not only demonstrates the effectiveness of GeMap on a different dataset but also emphasizes the adaptability and precision of GeMap in the evolving landscape of autonomous driving technologies.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 346, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4ddb0c04186a8894eeae1d61e2c0dfe3c390751b7b26757cb519b85faf7cc554.jpg", + "table_caption": [ + "Table 4: Ablation study on the nuScenes dataset. We train the model for 24 epochs. * denotes replacing GDA with 2 layers of vanilla self-attention." + ], + "table_footnote": [], + "table_body": "
Method\\( {\\mathcal{L}}_{\\text{shp }}\\left( {\\S {3.3}}\\right) {\\mathcal{L}}_{\\text{rel }}\\left( {\\S {3.3}}\\right) \\)GDA (§ 3.5)\\( {\\mathrm{{AP}}}_{div}\\left( \\uparrow \\right) \\)\\( {\\mathrm{{AP}}}_{ped}\\left( \\uparrow \\right) \\)\\( {\\mathrm{{AP}}}_{bnd}\\left( \\uparrow \\right) \\)mAP(↑)
Baseline49.544.753.749.3
+ Decoupled Attention✓53.446.653.551.2
+ Single Euclidean Loss✓51.045.452.749.7
51.743.453.049.4
+ Euclidean Loss✓✓51.543.951.148.8
+ Single Euclidean Loss with Decoupled Attention✓✓54.048.253.151.8
✓✓54.747.355.352.4
Replace GDA with 2-SA✓✓\\( 2 - SA^* \\)53.546.254.451.4
Full✓✓✓53.6 ↑ 4.149.2 ↑ 4.554.8 ↑ 1.152.6 ↑ 3.3
", + "bbox": [ + 233, + 184, + 763, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 340, + 388, + 356 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Components Ablation. The ablation experiments, detailed in Table 4, involved training the model for a reduced duration of 24 epochs without dense prediction losses, to facilitate efficient analysis. These experiments confirm the significant role of GDA, which alone increases mAP by $+1.9\\%$ . Moreover, applying the full suite of components results in a further mAP enhancement of $+1.4\\%$ . A notable discovery is the detrimental effect on model performance when Euclidean Loss is applied without GDA, which leads to a $0.5\\%$ reduction in mAP. This reinforces our position that conventional self-attention mechanisms are insufficient for encoding a variety of geometric properties.", + "bbox": [ + 212, + 367, + 785, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of GDA. In the proposed GDA (3.5), we decouple shape and relation learning, sequentially applying Euclidean shape and relation attention. In addition, we also attempt to intuitively double the self-attention layers in one block, referred to as \"2-SA\". As shown in Table 4, GDA outperforms vanilla self-attention by $+1.2\\%$ mAP, indicating the superiority of GDA when combined with Euclidean Loss.", + "bbox": [ + 212, + 503, + 785, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Visualization Analysis", + "text_level": 1, + "bbox": [ + 215, + 617, + 444, + 633 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "More Precise Shape Awareness. As shown in Figure 6(a), the shape geometry allows for an inherent description of the triangle feature, in contrast to baselines which do not discern it as effectively. A more irregularly shaped pedestrian crossing can be viewed in scenario (c) of Figure 6 and GeMap provides more precise construction than other methods.", + "bbox": [ + 212, + 643, + 785, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Better Relation Awareness. With the help of Euclidean Relation Clues, GeMap can better infer parallelism and perpendicular in scenario (a) of Figure 6. More interestingly, in scenario (b), the strong sunlight makes the divider highlighted hard to recognize. However, according to the understanding of lane width patterns, GeMap has the sense that there should be a divider.", + "bbox": [ + 212, + 719, + 785, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alleviating Occlusion Issues. Figure 6(c) demonstrates that shape geometry alleviates occlusions in partially visible instances. It showcases a black car that obscures part of a pedestrian crossing, while, GeMap successfully deduces the", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/613ff0d6cc99954666acbab67f745803be1f04fea88dc1a5e6ad965b81431896.jpg", + "image_caption": [ + "Fig. 6: Visualization results. Hard predictions are highlighted, such as complex triangular road boundary (a), strong sunlight (b), partially observed (c), and obscured situations (d). These challenging cases indicate the superiority and robustness of GeMap." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 443 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "overall structure of the crossing within a complex shape. In contrast, these baseline models [4, 22, 23] struggle with such complex shape recovery under similar occlusion conditions.", + "bbox": [ + 212, + 523, + 782, + 568 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Enhanced Robustness to Rotational Transformations. Figure 6(d) showcases that the BEV map is highly tilted by turning, which is extensively obscured by vehicular traffic. GeMap more adeptly maintains the integrity of lane width and parallelism, which are key geometry. This underscores the superior robustness of our geometric constructs against rotational distortions.", + "bbox": [ + 212, + 569, + 784, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Hyperparameter Experiments", + "text_level": 1, + "bbox": [ + 214, + 665, + 508, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparative Impact of Euclidean Losses. Figure 7a shows that GeMap exhibits larger sensitivity to the shape loss than to the relation loss, as evidenced by the steeper change of the performance curve. The optimal result is obtained when the shape and relation losses are relatively balanced by weights.", + "bbox": [ + 212, + 688, + 784, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Optimization of Euclidean Loss Weighting. The performance of the model, as demonstrated in Figure 7b, peaks when the weight of the Euclidean Loss, $\\lambda$ , is set to $5 \\times 10^{-3}$ . Notably, model performance degrades with a too large or too small $\\lambda$ , which informed us to use a $\\lambda$ value of $5 \\times 10^{-3}$ for our experiments.", + "bbox": [ + 212, + 750, + 784, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Distance Weighting. Experimental results of distance weighting are presented in Table 5. We change the order of distance $(p)$ in Equation 5 and $p\\in \\{1,2,4\\}$", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 348, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c8848dbd709d1278610f4228219e8e6f6d28ea9c757035d13827c202780ba3fb.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 500, + 285 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/67515aafdedc2b0aa1b1799256f4835fcc81c6a2cbcb431c8b8a7f7ee992f880.jpg", + "image_caption": [ + "(b)", + "Fig. 7: Hyper-parameters. The best performance is reached with balanced Euclidean shape loss and relation loss." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 759, + 285 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "corresponds to \"Linear\", \"Square\" and \"4th Power\" respectively. Moreover, we also try to treat instance pairs equally and it delivers the best performance.", + "bbox": [ + 212, + 364, + 784, + 396 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9b715b1839df8aee28575253092865ab2cf4702952ebfe2a1e3cb1fd1221c54f.jpg", + "table_caption": [ + "Table 5: Distance weighting on the nuScenes dataset." + ], + "table_footnote": [], + "table_body": "
StrategyAP_div(↑)AP_ped(↑)AP_bnd(↑)mAP(↑)
Linear53.546.656.252.1
Square53.247.955.252.1
4th Power53.247.455.852.1
Equal53.649.254.852.6
", + "bbox": [ + 359, + 446, + 638, + 522 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 568, + 359, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we realize significant shape and relation geometry inherent in HD map instances and propose the GeMap. GeMap includes the integration of Euclidean shape and relation losses for auxiliary supervision. To further refine the model's awareness of diverse geometry, we introduce the Geometry-Decoupled Attention mechanism. GeMap has achieved state-of-the-art performances on both the nuScenes and Argoverse 2 datasets, underscoring its effectiveness. Despite these promising results, the current application of geometry remains fundamental, and future research could focus on more sophisticated representations or enhanced geometric patterns. Furthermore, the application of geometry extends beyond HD map construction, offering potential solutions to occlusion challenges in other autonomous driving tasks. We anticipate that these findings will inspire further research.", + "bbox": [ + 212, + 598, + 785, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62272045, No. 8326014), The Shun Hing Institute of Advanced Engineering (No. 8115074), and CUHK Direct Grants (No. 4055190).", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "2. Chen, S., Cheng, T., Wang, X., Meng, W., Zhang, Q., Liu, W.: Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584 (2022)", + "3. Deo, N., Wolff, E., Beijbom, O.: Multimodal trajectory prediction conditioned on lane-graph traversals. In: Conference on Robot Learning. pp. 203-212. PMLR (2022)", + "4. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3672-3682 (2023)", + "5. Ding, X., Zhang, Y., Ge, Y., Zhao, S., Song, L., Yue, X., Shan, Y.: Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5513-5524 (2024)", + "6. Espinoza, J.L.V., Liniger, A., Schwarting, W., Rus, D., Van Gool, L.: Deep interactive motion prediction and planning: Playing games with motion prediction models. In: Learning for Dynamics and Control Conference. pp. 1006-1019. PMLR (2022)", + "7. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking efficient lane detection via curve modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17062-17070 (2022)", + "8. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020)", + "9. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for 3d detection and map segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8721-8731 (2023)", + "0. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023)", + "1. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26584-26595 (2024)", + "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "3. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023)" + ], + "bbox": [ + 225, + 179, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 348, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Jiao, J.: Machine learning assisted high-definition map creation. In: 2018 IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC). vol. 1, pp. 367-373. IEEE (2018)", + "15. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019)", + "16. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)", + "17. Li, C., Shi, J., Wang, Y., Cheng, G.: Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4370-4379 (2022)", + "18. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: An online hd map construction and evaluation framework. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 4628-4634. IEEE (2022)", + "19. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023)", + "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)", + "21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 541-556. Springer (2020)", + "22. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: MapTR: Structured modeling and learning for online vectorized HD map construction. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=k7p_YA07yE", + "23. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736 (2023)", + "24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "25. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1765-1772 (2022)", + "26. Liu, Y., Yuan, T., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: International Conference on Machine Learning. pp. 22352-22369. PMLR (2023)", + "27. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "31. Loukkal, A., Grandvalet, Y., Drummond, T., Li, Y.: Driving among flatmobiles: Bird-eye-view occupancy grids from a monocular camera for holistic trajectory planning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 51-60 (2021)", + "32. Lu, C., van de Molengraft, M.J.G., Dubbelman, G.: Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters 4(2), 445-452 (2019)", + "33. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4227-4236 (2021)", + "34. Pan, B., Sun, J., Leung, H.Y.T., Andonian, A., Zhou, B.: Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters 5(3), 4867-4873 (2020)", + "35. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)", + "36. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13218-13228 (2023)", + "37. Scheel, O., Bergamini, L., Wolczyk, M., Osiński, B., Ondruska, P.: Urban driver: Learning to drive from real-world demonstrations using policy gradients. In: Conference on Robot Learning. pp. 718-728. PMLR (2022)", + "38. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Keep your eyes on the lane: Real-time attention-guided lane detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 294-302 (2021)", + "39. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Polylanenet: Lane estimation via deep polynomial regression. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6150-6156. IEEE (2021)", + "40. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)", + "41. Van Gansbeke, W., De Brabandere, B., Neven, D., Proesmans, M., Van Gool, L.: End-to-end lane detection through differentiable least-squares fitting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)", + "42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "43. Wang, J., Ma, Y., Huang, S., Hui, T., Wang, F., Qian, C., Zhang, T.: A keypoint-based global association network for lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1392-1401 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Online Vectorized HD Map Construction using Geometry", + "bbox": [ + 348, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)", + "45. Xie, Z., Pang, Z., Wang, Y.X.: Mv-map: Offboard hd-map generation with multiview consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8658-8668 (2023)", + "46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "47. Yang, W., Li, Q., Liu, W., Yu, Y., Ma, Y., He, S., Pan, J.: Projecting your view attentively: Monocular road scene layout estimation via cross-view transformation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15536-15545 (2021)", + "48. Yuan, T., Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Streammapnet: Streaming mapping network for vectorized online hd map construction. arXiv preprint arXiv:2308.12570 (2023)", + "49. Zhang, G., Lin, J., Wu, S., Luo, Z., Xue, Y., Lu, S., Wang, Z., et al.: Online map vectorization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems 36 (2024)", + "50. Zhang, Y., Ding, X., Gong, K., Ge, Y., Shan, Y., Yue, X.: Multimodal pathway: Improve transformers with irrelevant data from other modalities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6108-6117 (2024)", + "51. Zhang, Y., Gong, K., Zhang, K., Li, H., Qiao, Y., Ouyang, W., Yue, X.: Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802 (2023)", + "52. Zhang, Y., Li, H., Liu, J., Yue, X.: Explore the limits of omni-modal pretraining at scale. arXiv preprint arXiv:2406.09412 (2024)", + "53. Zhang, Z., Zhang, Y., Ding, X., Jin, F., Yue, X.: Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341 (2023)", + "54. Zhou, B., Krahenbuhl, P.: Cross-view transformers for real-time map-view semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13760-13769 (2022)", + "55. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022)", + "56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 672 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_model.json b/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82de7243c16464f744eb776c3c0c049e25793042 --- /dev/null +++ b/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_model.json @@ -0,0 +1,2623 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.14, + 0.771, + 0.187 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.212, + 0.783, + 0.228 + ], + "angle": 0, + "content": "Zhixin Zhang\\(^{1}\\) Yiyuan Zhang\\(^{2}\\) Xiaohan Ding\\(^{3}\\) Fusheng Jin\\(^{1*}\\) Xiangyu Yue\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.24, + 0.758, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Beijing Institute of Technology \n\\(^{2}\\) The Chinese University of Hong Kong \n\\(^{3}\\)Tencent AI Lab \nzhangzhixin@bit.edu.cn, yiyuanzhang.ai@gmail.com \nhttps://invictus717.github.io/GeMap/" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.338, + 0.74, + 0.56 + ], + "angle": 0, + "content": "Abstract. Online vectorized High-Definition (HD) map construction is critical for downstream prediction and planning. Recent efforts have built strong baselines for this task, however, geometric shapes and relations of instances in road systems are still under-explored, such as parallelism, perpendicular, rectangle-shape, etc. In our work, we propose GeMap (Geometry Map), which end-to-end learns Euclidean shapes and relations of map instances beyond fundamental perception. Specifically, we design a geometric loss based on angle and magnitude clues, robust to rigid transformations of driving scenarios. To address the limitations of the vanilla attention mechanism in learning geometry, we propose to decouple self-attention to handle Euclidean shapes and relations independently. GeMap achieves new state-of-the-art performance on the nuScenes and Argoverse 2 datasets. Remarkably, it reaches a \\(71.8\\%\\) mAP on the large-scale Argoverse 2 dataset, outperforming MapTRv2 by \\(+4.4\\%\\) and surpassing the \\(70\\%\\) mAP threshold for the first time. Code is available at https://github.com/cnzzx/GeMap." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.574, + 0.757, + 0.601 + ], + "angle": 0, + "content": "Keywords: HD Map Construction \\(\\cdot\\) Geometry Representation \\(\\cdot\\) Geometry-Decoupled Attention" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.629, + 0.375, + 0.645 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.664, + 0.8, + 0.739 + ], + "angle": 0, + "content": "Vectorized HD maps provide structured environmental information for autonomous vehicles and have been widely adopted in downstream tasks, such as trajectory forecasting [3,21,55] and planning [6,37]. Online Vectorized HD Map Construction can significantly reduce the need for labor-intensive annotations and facilitate real-time updates in autonomous driving [4,18,22,53]." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.74, + 0.787, + 0.815 + ], + "angle": 0, + "content": "With the development of Bird's-Eye-View (BEV) representation and multimodal learning [5, 11, 19, 28, 50–52], online HD map construction has achieved significant advancements [10, 20, 23]. Early works [14, 32, 35, 47] formulate HD map construction as a dense prediction task. However, these methods generate maps in image format, which is redundant for representing sparse map instances. Then," + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.382, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.147, + 0.486, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.283, + 0.371, + 0.296 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.145, + 0.769, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.283, + 0.637, + 0.296 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.307, + 0.789, + 0.378 + ], + "angle": 0, + "content": "Fig. 1: Geometric Invariance. (a) As the ego vehicle moves, after rotation \\((\\leftarrow)\\) and translation \\((\\rightarrow)\\), the shape of the crossing and the parallelism between lanes remain unchanged, which indicates the invariant property of geometry to rigid transformations. (b) Absolute coordinates are vulnerable to rotation and translation, however, our G-representation is invariant, which is more suitable to capture geometric properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.504 + ], + "angle": 0, + "content": "a more compact map formulation was introduced to minimize redundancy, albeit at the cost of adding time-consuming post-processing steps [18]. In response to this, recent works [4, 22, 26, 36] attempt to end-to-end construct vectorized HD maps to avoid extensive post-processing. These methods typically sample points from map instances and represent each instance as a polyline [4, 22, 26] or parameterized curve [36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.734 + ], + "angle": 0, + "content": "We observe that transportation road systems exhibit significant geometric characteristics (Figure 2a), such as parallel lanes, perpendicular crossings, equal lane widths, etc. However, these geometric properties of shapes and relations between map instances have not been fully explored. Meanwhile, there are two notable limitations among existing methods [4,22,26,36] which can be alleviated by leveraging these geometric properties. 1) Excessive dependency on absolute coordinates: as the ego-vehicle moves, instances experience rotations and translations, as depicted in Figure 1a. However, widely adopted representations such as polylines and parameterized curves [8,36] are inherently sensitive to rotation and translation changes. 2) Objective conflicts of vanilla attention mechanism (detail in § 3.5) struggles to learn diverse shapes and relations, even though geometric properties such as rectangle shape, parallelism, and perpendicular relationships are commonly found in driving scenarios, as illustrated in Figure 2a. We believe that incorporating these geometric properties significantly enhances the precision and efficiency of online HD map construction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To address the above two limitations, we introduce a novel geometric representation that captures shapes of individual map instances and relations between different instances as illustrated in Figure 2b and 2c, referred to as G-Representation. It enhances the vanilla representation of map instances by incorporating a translation- and rotation-invariant representation that effectively leverages instance geometry. The local structures of map features are encoded using displacement vectors. These vectors are computed from the absolute coor" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.148, + 0.387, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.232, + 0.371, + 0.245 + ], + "angle": 0, + "content": "Rectangle-shaped" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.146, + 0.493, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.232, + 0.481, + 0.244 + ], + "angle": 0, + "content": "Parallelism" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.153, + 0.619, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.232, + 0.605, + 0.245 + ], + "angle": 0, + "content": "Perpendicular" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.155, + 0.753, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.232, + 0.748, + 0.245 + ], + "angle": 0, + "content": "Equal Lane Width" + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.248, + 0.51, + 0.259 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.264, + 0.498, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.331, + 0.435, + 0.344 + ], + "angle": 0, + "content": "Single Map Instance" + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.347, + 0.381, + 0.359 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.26, + 0.619, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.332, + 0.692, + 0.344 + ], + "angle": 0, + "content": "Cross Map Instances" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.266, + 0.739, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.347, + 0.639, + 0.359 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.37, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: Geometric properties and G-Representation. (a) Geometry in the transportation road system. (b) (c) We propose to model the geometric properties of a single map instance and multiple instances, with magnitude \\( d \\) and angle \\( \\alpha \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.55 + ], + "angle": 0, + "content": "dinates of polyline points to effectively represent the relative positions and orientations of adjacent points. To quantify these features within Euclidean space, we employ both the magnitudes of the displacement vectors and the angles formed between these vectors. Based on G-Representation, we propose the Euclidean Shape Clues to represent the shapes of individual map instances and the Euclidean Relation Clues to represent the inter-instance relations in Euclidean space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.703 + ], + "angle": 0, + "content": "G-Representation is a simple yet effective method to address the abovementioned two limitations: 1) Geometric Invariance, by concentrating on the relative relationships between points within and across map instances, it inherently attains translation and rotation invariance as illustrated in Figure 1b. This enhances its robustness against variations in data collection and equips it to effectively handle different coordinate systems. 2) Euclidean Modeling, it captures the inherent geometry in transportation road systems. Diverse shapes and relation geometry can be simplified as enumerations of magnitudes and directions. For instance, in Euclidean space, parallel lanes can be easily modeled as approximate directions \\((\\alpha \\approx 0)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Building upon G-Representation, we propose a framework named GeMap for HD map construction. A BEV encoder is used to extract features from multiview input images, while a geometry-decoupled decoder is employed to focus on geometric aspects. Specifically, we adapt the attention mechanism [42] and propose Geometry-Decoupled Attention (GDA). GDA sequentially applies attention to queries belonging to the same instance and attention to queries across different instances. This can significantly boost the geometry learning of key points of shapes and relations between map instances. Furthermore, we propose an objective function named Euclidean Loss to optimize G-Representation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "Specifically, we transform the conventional polylines of the ground truth map into our G-Representation. In this way, the model gets optimized to better understand the magnitude of displacement vectors and the angles between them, thereby facilitating a more effective learning of geometric properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.794, + 0.297 + ], + "angle": 0, + "content": "Experiments on nuScenes and Argoverse 2 datasets demonstrate the effectiveness of GeMap. We reach new state-of-the-art performances on both datasets. With camera images only, GeMap achieves \\(69.4\\%\\) and \\(71.8\\%\\) mAP on the nuScenes and Argovserse 2, respectively. Visualization results (§ 4.4) further demonstrate the better perception of shape and relation, alleviation of occlusion, and robustness to rigid transformations of GeMap." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.299, + 0.611, + 0.314 + ], + "angle": 0, + "content": "Our contributions are summarized as the following:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.327, + 0.784, + 0.371 + ], + "angle": 0, + "content": "- We propose G-Representation which harnesses critical geometric properties of rotation and translation invariance in autonomous driving scenarios, opening up new research avenues within the field." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.373, + 0.784, + 0.416 + ], + "angle": 0, + "content": "- We introduce GeMap, a novel framework incorporating geometry-decoupled attention and Euclidean Loss function, specifically designed to learn the intrinsic geometry of online HD maps." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.419, + 0.784, + 0.462 + ], + "angle": 0, + "content": "- GeMap achieves new state-of-the-art results in HD map construction on the nuScenes and Argoverse 2 datasets, notably surpassing the \\(70\\%\\) mAP on the large-scale Argoverse 2 dataset for the first time." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.327, + 0.784, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.388, + 0.505 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.509, + 0.538 + ], + "angle": 0, + "content": "2.1 Online HD Map Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Traditionally, HD map construction has required labor-intensive manual or semi-automatic annotations [14,33]. To streamline this, recent studies [20,31,34,35,54] have focused on online construction, approaching HD maps as a dense prediction challenge. Innovations such as MetaBEV [9] aim to mitigate sensor issues, while MVNet [45] uses historical data for improved semantic consistency. The trend towards automatic vectorization of HD maps is spearheaded by works such as HDMapNet [18], which fuses camera and LiDAR inputs in BEV space and is advanced by end-to-end solutions such as VectorMapNet [26] and MapTR [22], leveraging Transformer-based models. StreamMapNet [48] and PivotNet [4] build upon this, modifying attention mechanisms for better performance. Our contribution starts from the strengths of the Transformer architecture, optimizing it with a decoupled self-attention block for enhanced geometric processing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.478, + 0.769 + ], + "angle": 0, + "content": "2.2 Cross-view BEV Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "The conversion of Perspective View (PV) camera images into a unified BEV space is a significant challenge for autonomous driving systems. Previous studies such as [19,35] use depth estimation from monocular images for this transformation, others [2, 20, 28] have developed PV-to-BEV conversion methods without" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "explicit depths. However, these methods can distort shapes and geometric relations, leading to inaccuracies. Our approach integrates geometric supervision in prediction, addressing these geometric inconsistencies while acknowledging the value of depth information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.231, + 0.509, + 0.246 + ], + "angle": 0, + "content": "2.3 Geometric Instance Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Traditional vectorization approaches model these instances as polylines [22, 26, 38, 43] or polynomial curves [7, 36, 39, 41], with some incorporating Bézier Curves for improved fitting [7, 36]. However, such methods often overlook geometric properties including shapes, parallelism, perpendicular, and etc. In 3D Lane Detection, some works [13, 17, 25] also explore simple geometry priors of lanes such as \"equal lane width\". However, they are not generalizable to more complex HD map instances. Existing works on modeling geometry properties of map instances mainly address rasterization-based shape loss [41, 49] or edge loss [22], and they fail to preserve cross-instance relationships. PivotNet [4] offers an architectural advancement with its line-aware point decoder, yet it still does not fully address the complexity of instance geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.787, + 0.483 + ], + "angle": 0, + "content": "Given the current under-exploration of diverse geometric properties in shape and relation, we propose utilizing Euclidean shape and relation losses and decoupling the traditional self-attention module to empower the model with a more robust understanding of instance geometry." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.331, + 0.523 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.539, + 0.362, + 0.555 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.732 + ], + "angle": 0, + "content": "The architecture of our method is illustrated in Figure 3. Input images are denoted by \\(\\mathbb{I} = \\{\\pmb{I}_i\\}_{i=1}^{N_c}\\), where \\(\\pmb{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}\\) and \\(N_c\\) is the number of cameras. The output is a set of \\(N\\) map instances \\(\\mathbb{M} = \\{\\pmb{L}_i\\}_{i=1}^N\\), where each map instance is represented by a polyline \\(\\pmb{L}_i \\in \\mathbb{R}^{N_v \\times 2}\\), i.e., an ordered sequence of \\(N_v\\) two-dimensional points. As shown in Figure 3, the commonly adopted pipeline comprises three steps. 1) A BEV feature extractor processes the multi-view images. 2) The extracted BEV features are fed into a map decoder, which predicts the map instances based on the BEV features. 3) To optimize the pipeline, the predicted map instances are compared against the ground truth. Beyond point-to-point comparison, G-representations of predictions and the ground truth are computed and the difference is measured by \\(L^1\\) loss." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.754, + 0.452, + 0.769 + ], + "angle": 0, + "content": "3.2 Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.794, + 0.842 + ], + "angle": 0, + "content": "BEV Feature Extractor. A shared vision backbone in different views and parameterized PV-to-BEV transformation network are employed to aggregate features from various perspectives. In our default configuration, we utilize ResNet [12] as the PV backbone and GKT [2] as the PV-to-BEV transformation network." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.144, + 0.75, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 3: Illustration of GeMap. First, PV images are transformed into BEV features, then a Geometry-Decoupled Decoder outputs the vectorized HD Map. In each block of the decoder, queries are first processed by Euclidean shape and relation attention, which focuses on geometric relevance. Finally, predictions are enhanced in G-Representations by shape and relation constraint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Geometry-Decoupled Decoder. We adopt a Transformer decoder to predict polylines. The decoder obtains \\( N \\times N_v \\) queries that represent points and processes them via self-attention and aggregate BEV features via deformable cross-attention [56]. Finally, a prediction head is used to convert queries to polylines \\( \\{\\hat{L}_i\\}_{i=1}^N \\). We utilize a multi-layer perceptron as the polyline prediction head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Geometric Loss. For training, we convert both the predicted polylines and ground truth into the proposed G-Representation. Based on that, we let the model optimize the predicted magnitudes of displacement vectors and angles between displacement vectors to match the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.689 + ], + "angle": 0, + "content": "In the following subsections, we first introduce Geometric Representation in § 3.3 since it is the core of our framework. Based on that, we introduce Geometric Loss in § 3.4 and Geometry-Decoupled Decoder in § 3.5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.48, + 0.724 + ], + "angle": 0, + "content": "3.3 Geometric Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Euclidean Shape Clues. We first introduce the representation of shapes of individual map instances. For each instance, we describe the local geometry with displacement vectors between neighboring points which are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.785, + 0.787, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {u} ^ {i} = \\boldsymbol {L} _ {i, u + 1} - \\boldsymbol {L} _ {i, u} \\quad (u \\in \\{1, 2, \\dots , N _ {v} \\}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "where we define \\( \\pmb{L}_{i,N_v + 1} \\coloneqq \\pmb{L}_{i,1} \\) to unify the geometric formulation of closed and open polylines." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.144, + 0.744, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.235, + 0.785, + 0.278 + ], + "angle": 0, + "content": "Fig. 4: Euclidean Shape Clues. Magnitudes of displacement vectors and angles between neighboring vectors indicate shape clues and are utilized to compute shape loss. The right part shows how to connect Euclidean Shape Clues to shape geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.785, + 0.44 + ], + "angle": 0, + "content": "These displacement vectors are sufficient to represent the shape of a map instance and such a representation is invariant to translation transformations. However, we would like to note that this representation is vulnerable to rotations, which might prevent the model from learning robust geometry. To solve this problem, we propose to represent the shape with magnitudes of displacement vectors and angles between consecutive displacement vectors, as illustrated in Figure 4. Specifically, the \\(i\\)-th instance is denoted by \\(N_v\\) angle values and \\(N_v\\) magnitude values. Let \\(u\\) be the index, the \\(u\\)-th angle value and the magnitude value of the \\(i\\)-th instance are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.447, + 0.785, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {u} ^ {i} = \\left\\langle \\boldsymbol {v} _ {u} ^ {i}, \\boldsymbol {v} _ {u + 1} ^ {i} \\right\\rangle , d _ {u} ^ {i} = \\left\\| \\boldsymbol {v} _ {u} ^ {i} \\right\\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.757, + 0.488 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot \\rangle\\) denotes the angle between two vectors and similarly \\(\\pmb{v}_{N_v + 1}^i \\coloneqq \\pmb{v}_1^i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.608 + ], + "angle": 0, + "content": "The proposed representation inherently captures common geometric patterns, such as parallelism, right angles, and proper line width, by translating them into corresponding numerical patterns within this representation. For example, in Figure 4, a rectangle is characterized by one \\(90^{\\circ}\\) angle and equal magnitudes of opposite displacement vectors; in addition, a straight line is represented by \\(0^{\\circ}\\) angles. Beyond discussed regular cases, the Euclidean shape clues are capable of handling more complex shapes, as illustrated in scenarios (a) and (c) of Figure 6 and discussed in § 4.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.784, + 0.763 + ], + "angle": 0, + "content": "Euclidean Relation Clues. Having highlighted the translation and rotation invariance of G-Representation and its advantages in representing the geometric shapes of individual map instances, we further introduce its ability to capture the relations between two map instances, e.g., parallelism and perpendicular, in Euclidean space. Specifically, given the vanilla representations and displacement vectors of the \\(i\\)-th and \\(j\\)-th map instance, we represent the relation between instance \\(i\\) and \\(j\\) with 1) angles between each pair of their respective displacement vectors, and 2) magnitudes of displacement vectors between each pair of points, as illustrated in Figure 5. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.77, + 0.785, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {u, v} ^ {i, j} = \\left\\langle \\boldsymbol {v} _ {u} ^ {i}, \\boldsymbol {v} _ {v} ^ {j} \\right\\rangle , d _ {u, v} ^ {i, j} = \\left\\| \\boldsymbol {L} _ {i, u} - \\boldsymbol {L} _ {j, v} \\right\\| _ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "This representation is also translation- and rotation-invariant. Similar to Euclidean Shape Clues, it inherently captures common relations between map instances by translating them into corresponding numerical patterns. For example," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.144, + 0.745, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Fig. 5: Euclidean Relation Clues. Angles between pairs of displacement vectors on different polylines, and magnitudes of displacement vectors between point pairs indicate relation clues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.404 + ], + "angle": 0, + "content": "Figure 5 shows that the perpendicular relation can be directly represented by a \\(90^{\\circ}\\) angle, and the distance between two parallel polylines, which may correspond to the width of a lane, is naturally represented by magnitudes of displacement vector between point pairs. Scenarios (a) and (b) of Figure 6 show the effectiveness of Euclidean Relation Clues and more discussion can be viewed in \\(\\S 4.4\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.424, + 0.513, + 0.44 + ], + "angle": 0, + "content": "3.4 Euclidean Loss and Objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.569 + ], + "angle": 0, + "content": "We transform both the ground truth data and the model's predicted polylines from their original format into G-Representation. This conversion allows the model to independently optimize the angles between displacement vectors and the magnitudes of displacement vectors to more accurately align with the ground truth. The proposed Euclidean Loss is composed of two parts that measure how accurately the model predicts the shape of individual map instances and the inter-instance relations, respectively, which are denoted by \\(\\mathcal{L}_{\\mathrm{shp}}\\) and \\(\\mathcal{L}_{\\mathrm{rel}}\\). The Euclidean Loss can be computed as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.579, + 0.787, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E u c}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {s h p}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {r e l}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.787, + 0.636 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters, whose effects on the model performances are evaluated in Figure 7a. Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.645, + 0.786, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} w _ {i, j} = 1 - \\left(\\min _ {u, v} \\left\\{d _ {u, v} ^ {i, j} \\right\\} / \\sqrt {2}\\right) ^ {p}, \\quad \\mathcal {L} _ {\\mathrm {s h p}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N _ {v}} \\left| \\hat {d} _ {j} ^ {i} - d _ {j} ^ {i} \\right| + \\ell \\left(\\hat {\\alpha} _ {j} ^ {i}, \\alpha_ {j} ^ {i}\\right), \\tag {5} \\\\ \\mathcal {L} _ {\\mathrm {r e l}} = \\sum_ {i = 1} ^ {N} \\sum_ {j > i} ^ {N} w _ {i, j} \\sum_ {u = 1} ^ {N _ {v}} \\sum_ {v = 1} ^ {N _ {v}} \\left(| \\hat {d} _ {u, v} ^ {i, j} - d _ {u, v} ^ {i, j} | + \\ell (\\hat {\\alpha} _ {u, v} ^ {i, j}, \\alpha_ {u, v} ^ {i, j})\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.784, + 0.774 + ], + "angle": 0, + "content": "where \\(\\ell(\\cdot)\\) is a function based on \\(L^1\\) loss. We avoid using the inverse trigonometric function to directly compute the angles, but use sine and cosine values instead:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.784, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\ell (\\hat {\\alpha}, \\alpha) = | \\cos (\\hat {\\alpha}) - \\cos (\\alpha) | + | \\sin (\\hat {\\alpha}) - \\sin (\\alpha) |. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Moreover, the distance might influence the relation strength of instance pairs. For example, if two instances are far from each other, the relation between them" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "might be weak. For this reason, we further adopt \\( w_{i,j} \\) to punish weakly related instance pairs and discuss it in § 4.5. According to the best experimental results, we treat all instance pairs equally in other experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.268 + ], + "angle": 0, + "content": "Following common practices [22,48], we also use focal loss [24] \\(\\mathcal{L}_{\\mathrm{cls}}\\) for classification. For polyline regression, we adopt point-to-point loss and edge direction loss, denoted by \\(\\mathcal{L}_{\\mathrm{pts}}\\) and \\(\\mathcal{L}_{\\mathrm{dir}}\\) respectively. Following [23,36], we further adopt segmentation loss \\(\\mathcal{L}_{\\mathrm{seg}}\\) and depth estimation loss \\(\\mathcal{L}_{\\mathrm{dep}}\\). These losses are detailed in appendix § A.1. The overall loss function can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.275, + 0.785, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\cdot \\mathcal {L} _ {\\text {E u c}} + \\beta_ {1} \\cdot \\mathcal {L} _ {\\text {c l s}} + \\beta_ {2} \\cdot \\mathcal {L} _ {\\text {p t s}} + \\beta_ {3} \\cdot \\mathcal {L} _ {\\text {d i r}} + \\beta_ {4} \\cdot \\mathcal {L} _ {\\text {s e g}} + \\beta_ {5} \\cdot \\mathcal {L} _ {\\text {d e p}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.297, + 0.753, + 0.312 + ], + "angle": 0, + "content": "where hyperparameters are discussed in Figure 7b and appendix Table A1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.331, + 0.512, + 0.346 + ], + "angle": 0, + "content": "3.5 Geometry-Decoupled Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.784, + 0.426 + ], + "angle": 0, + "content": "In the Transformer decoder, we obtain the \\(N \\times N_v\\) fused queries by adding \\(N\\) instance queries to \\(N_v\\) point queries, which is suggested in [22]. We denote input tokens by \\(\\pmb{E} \\in \\mathbb{R}^{N_A \\times D}\\), where \\(D\\) is the feature dimension and \\(N_A = N \\cdot N_v\\) is the number of fused queries. Intuitively, each fused query corresponds to a point on a predicted map instance. Self-Attention (SA) is formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.433, + 0.785, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} (\\boldsymbol {E}, \\boldsymbol {M}) = \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\boldsymbol {E} \\boldsymbol {W} ^ {q}\\right) \\left(\\boldsymbol {E} \\boldsymbol {W} ^ {k}\\right) ^ {\\top}}{\\sqrt {D _ {k}}} \\odot \\boldsymbol {M}\\right) \\boldsymbol {E} \\boldsymbol {W} ^ {v}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.784, + 0.517 + ], + "angle": 0, + "content": "where \\(\\pmb{W}^q, \\pmb{W}^k, \\pmb{W}^v \\in \\mathbb{R}^{D \\times D_k}\\) are linear projection matrices, \\(\\pmb{M} \\in \\mathbb{R}^{N_A \\times N_A}\\) is the attention mask, and \\(\\odot\\) is the Hadamard product. The vanilla SA computes relations between every pair of tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.785, + 0.623 + ], + "angle": 0, + "content": "Nevertheless, the geometry of shape and relation pertains to distinct subsets of tokens. For any given map instance, its shape is intimately related to tokens representing that instance's points. Precise shape geometry capture requires the model to discern token correlations specific to an instance while avoiding interference from tokens of unrelated instances. Conversely, for relation geometry modeling, it is beneficial to isolate token correlations that span across different instances, rather than those confined within a single instance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.76 + ], + "angle": 0, + "content": "Therefore, we present Geometry-Decoupled Attention (GDA). First, we multiply a binary mask \\(M\\) to the computed attention map so that the tokens of a map instance are aggregated according to the tokens within the same map instances only, which allows the model to adjust points' positions according to learned shape geometry. Denote the index of map instance that the \\(i\\)-th token belongs to as \\(\\mathcal{I}_i\\). For example, tokens \\(1,2,\\ldots,N_v\\) belong to the first instance so that \\(\\mathcal{I}_1,\\mathcal{I}_2,\\ldots,\\mathcal{I}_{N_v}=1\\). Tokens \\(N_v+1,N_v+2,\\ldots,2\\cdot N_v\\) belong to the second instance so that \\(\\mathcal{I}_{N_v+1},\\mathcal{I}_{N_v+2},\\ldots,\\mathcal{I}_{2N_v}=2\\). With this notation, the binary mask \\(M\\) can be simply constructed by:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.767, + 0.785, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {i, j} ^ {\\mathrm {s h p}} = \\left\\{ \\begin{array}{l l} 1 & , \\mathcal {I} _ {i} = \\mathcal {I} _ {j} \\\\ 0 & , \\mathcal {I} _ {i} \\neq \\mathcal {I} _ {j} \\end{array} , \\quad \\boldsymbol {M} _ {i, j} ^ {\\mathrm {r e l}} = \\left\\{ \\begin{array}{l l} 1 & , \\mathcal {I} _ {i} \\neq \\mathcal {I} _ {j} \\\\ 0 & , \\mathcal {I} _ {i} = \\mathcal {I} _ {j} \\end{array} . \\right. \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The second attention is expected to model relations between tokens of different map instances." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 1: Comparison on the nuScenes dataset, GeMap reaches a new state-of-the-art performance. \"EB0\", \"R50\", \"PP\", \"Sec\", \"Swin-T\", and \"V2-99\" denote EfficientNet-B0 [40], ResNet50 [12], PointPillars [15], SECOND [46], Swin Transformer Tiny [27], and VoVNetV2-99 [16] respectively. Methods with two backbones utilize both camera and LiDAR inputs. \"Dns. Loss\" denotes whether any dense prediction (e.g. semantic segmentation) loss is adopted. The best result is highlighted in **bold**. We reproduce all methods on a single RTX3090 GPU to test FPS for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.256, + 0.765, + 0.501 + ], + "angle": 0, + "content": "
MethodsBackboneDns. Loss\\( AP_{div}(\\uparrow) \\)\\( AP_{ped}(\\uparrow) \\)\\( AP_{bnd}(\\uparrow) \\)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5042.551.444.146.05.3
MapTR [ICLR'23] [22]R5059.856.260.158.719.8
MapVR [NeurIPS'23] [49]R5061.855.059.458.819.8
GeMap [Ours]R5065.1 ↑ 3.359.8 ↑ 4.863.2 ↑ 3.862.7 ↑ 3.919.0
HDMapNet [ICRA'22] [18]EB0✓14.421.733.023.00.7
PivotNet [ICCV'23] [4]R50✓58.853.859.657.49.5
BeMapNet [CVPR'23] [36]R50✓66.762.665.164.86.6
MapTRv2 [Arxiv'23] [23]R50✓68.368.169.768.715.0
GeMap [Ours]R50✓69.8 ↑ 1.567.171.4 ↑ 1.769.4 ↑ 0.715.8
HDMapNet [ICRA'22] [18]EB0 & PP29.616.346.731.00.7
VectorMapNet [ICML'23] [26]R50 & PP60.148.253.053.7-
MapTR [ICLR'23] [22]R50 & Sec62.355.969.362.56.7
MapVR [NeurIPS'23] [49]R50 & PP62.760.467.263.5-
GeMap [Ours]R50 & Sec66.3 ↑ 3.662.2 ↑ 1.871.1 ↑ 3.966.5 ↑ 3.07.3
MapTRv2 [Arxiv'23] [23]R50 & Sec✓65.666.574.869.06.6
GeMap [Ours]R50 & Sec✓69.8 ↑ 4.268.0 ↑ 1.573.470.4 ↑ 1.46.8
MapTRv2 [Arxiv'23] [23]V2-99✓73.771.475.073.410.1
GeMap [Ours]Swin-T✓72.870.472.872.011.4
GeMap [Ours]V2-99✓76.0 ↑ 2.374.3 ↑ 2.977.7 ↑ 2.776.0 ↑ 2.610.8
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.377, + 0.55 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.436, + 0.592 + ], + "angle": 0, + "content": "4.1 Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.809, + 0.778 + ], + "angle": 0, + "content": "Datasets. To evaluate GeMap, we conduct experiments on the nuScenes dataset [1], a widely adopted large-scale autonomous driving dataset that includes 1,000 scenes captured by six RGB cameras with a 360-degree field of view, and provides precise annotations from LiDAR point clouds for HD map construction. For comparability with prior research [18,22], we focus on three static categories of map instances: pedestrian crossings, lane dividers, and road boundaries. With dataset splits provided by BeMapNet [36], we also evaluate GeMaps under three weather conditions: sunny, cloudy, and rainy. Additionally, we use the Argoverse 2 dataset [44] as another benchmark, which consists of approximately 108,000 frames, each providing images from seven cameras. We evaluate the performance by Average Precision (AP) [18,22] with thresholds at 0.5, 1.0, and 1.5 meters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. GeMap leverages 8 NVIDIA RTX 3090 GPUs for training. We adopt AdamW [30] as the optimizer and utilize Cosine Annealing with a linear warm-up phase [29] as the learning rate scheduler. For more details on hyperparameters, please refer to appendix § A.2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 2: Following BeMapNet [36], we also compare GeMap on the nuScenes dataset under different weather conditions." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.185, + 0.764, + 0.269 + ], + "angle": 0, + "content": "
MethodsBackbone Dns. LossmAPsun(↑)mAPcld(↑)mAPrny(↑)mAPavg(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5043.844.136.641.55.3
MapTR [ICLR'23] [22]R5062.160.552.858.419.8
GeMap [Ours]R5066.0 ↑ 3.964.3 ↑ 3.854.4 ↑ 1.661.5 ↑ 3.119.0
BeMapNet [CVPR'23] [36]R50 √67.367.556.663.86.6
GeMap [Ours]R50 √73.1 ↑ 5.871.0 ↑ 3.559.3 ↑ 2.767.8 ↑ 4.015.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.292, + 0.784, + 0.32 + ], + "angle": 0, + "content": "Table 3: Comparison on Argoverse 2 dataset. GeMap demonstrates significant performance improvements over previous methods." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.332, + 0.764, + 0.413 + ], + "angle": 0, + "content": "
MethodsBackbone Dns. Loss\\( AP_{div}(\\uparrow) \\)\\( AP_{ped}(\\uparrow) \\)\\( AP_{bnd}(\\uparrow) \\)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5036.138.339.237.9-
GeMap [Ours]R5067.6 ↑ 31.559.3 ↑ 21.064.7 ↑ 25.563.9 ↑ 26.016.7
HMapNet [ICRA'22] [18]EB0✓5.713.137.618.8
MapTRv2 [Arxiv'23] [23]R50✓72.162.967.167.4
GeMap [Ours]R50✓75.7 ↑ 3.669.2 ↑ 6.370.5 ↑ 3.471.8 ↑ 4.4
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.373, + 0.466 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.787, + 0.744 + ], + "angle": 0, + "content": "Results on nuScenes. As delineated in Table 1, GeMap delivers a state-of-the-art performance, achieving a mean Average Precision (mAP) of \\(69.4\\%\\) using the camera-only setup and ResNet50 [12] backbone. GeMap particularly improves the precision for identifying dividers and boundaries, with enhancements of \\(+1.5\\%\\) and \\(+1.7\\%\\), respectively. Importantly, these performance gains do not come at a high cost of efficiency, as GeMap sustains inference speeds on par with and even exceeds the previously established state-of-the-art, measured in frames per second (FPS). With more powerful vision backbones, GeMap can achieve significantly improved performance. Specifically, VoVNetV2-99 [16] and Swin-T [27] outperform the ResNet50 baseline by \\(+6.6\\%\\) and \\(+2.6\\%\\) in mAP, respectively. Also, GeMap achieves significantly improved performance with extra LiDAR inputs, which indicates the generalizability of GeMap to multi-modality settings. Moreover, we also compare performance under different weather conditions in Table 2. GeMap exceeds previous works under sunny, cloudy, and rainy scenarios, which shows the potential of geometry to raise model robustness to varied weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results on Argoverse 2. Extending our evaluation to the Argoverse 2 dataset, as presented in Table 3, GeMap also presents a new SOTA performance of \\(71.8\\%\\) mAP, outperforming the highly advanced MapTRv2 model by \\(+4.4\\%\\). This achievement not only demonstrates the effectiveness of GeMap on a different dataset but also emphasizes the adaptability and precision of GeMap in the evolving landscape of autonomous driving technologies." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 4: Ablation study on the nuScenes dataset. We train the model for 24 epochs. * denotes replacing GDA with 2 layers of vanilla self-attention." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.185, + 0.764, + 0.313 + ], + "angle": 0, + "content": "
Method\\( {\\mathcal{L}}_{\\text{shp }}\\left( {\\S {3.3}}\\right) {\\mathcal{L}}_{\\text{rel }}\\left( {\\S {3.3}}\\right) \\)GDA (§ 3.5)\\( {\\mathrm{{AP}}}_{div}\\left( \\uparrow \\right) \\)\\( {\\mathrm{{AP}}}_{ped}\\left( \\uparrow \\right) \\)\\( {\\mathrm{{AP}}}_{bnd}\\left( \\uparrow \\right) \\)mAP(↑)
Baseline49.544.753.749.3
+ Decoupled Attention✓53.446.653.551.2
+ Single Euclidean Loss✓51.045.452.749.7
51.743.453.049.4
+ Euclidean Loss✓✓51.543.951.148.8
+ Single Euclidean Loss with Decoupled Attention✓✓54.048.253.151.8
✓✓54.747.355.352.4
Replace GDA with 2-SA✓✓\\( 2 - SA^* \\)53.546.254.451.4
Full✓✓✓53.6 ↑ 4.149.2 ↑ 4.554.8 ↑ 1.152.6 ↑ 3.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.342, + 0.39, + 0.357 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.787, + 0.503 + ], + "angle": 0, + "content": "Components Ablation. The ablation experiments, detailed in Table 4, involved training the model for a reduced duration of 24 epochs without dense prediction losses, to facilitate efficient analysis. These experiments confirm the significant role of GDA, which alone increases mAP by \\(+1.9\\%\\). Moreover, applying the full suite of components results in a further mAP enhancement of \\(+1.4\\%\\). A notable discovery is the detrimental effect on model performance when Euclidean Loss is applied without GDA, which leads to a \\(0.5\\%\\) reduction in mAP. This reinforces our position that conventional self-attention mechanisms are insufficient for encoding a variety of geometric properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.787, + 0.593 + ], + "angle": 0, + "content": "Effectiveness of GDA. In the proposed GDA (3.5), we decouple shape and relation learning, sequentially applying Euclidean shape and relation attention. In addition, we also attempt to intuitively double the self-attention layers in one block, referred to as \"2-SA\". As shown in Table 4, GDA outperforms vanilla self-attention by \\(+1.2\\%\\) mAP, indicating the superiority of GDA when combined with Euclidean Loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.446, + 0.634 + ], + "angle": 0, + "content": "4.4 Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.718 + ], + "angle": 0, + "content": "More Precise Shape Awareness. As shown in Figure 6(a), the shape geometry allows for an inherent description of the triangle feature, in contrast to baselines which do not discern it as effectively. A more irregularly shaped pedestrian crossing can be viewed in scenario (c) of Figure 6 and GeMap provides more precise construction than other methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Better Relation Awareness. With the help of Euclidean Relation Clues, GeMap can better infer parallelism and perpendicular in scenario (a) of Figure 6. More interestingly, in scenario (b), the strong sunlight makes the divider highlighted hard to recognize. However, according to the understanding of lane width patterns, GeMap has the sense that there should be a divider." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Alleviating Occlusion Issues. Figure 6(c) demonstrates that shape geometry alleviates occlusions in partially visible instances. It showcases a black car that obscures part of a pedestrian crossing, while, GeMap successfully deduces the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.455, + 0.785, + 0.497 + ], + "angle": 0, + "content": "Fig. 6: Visualization results. Hard predictions are highlighted, such as complex triangular road boundary (a), strong sunlight (b), partially observed (c), and obscured situations (d). These challenging cases indicate the superiority and robustness of GeMap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.784, + 0.569 + ], + "angle": 0, + "content": "overall structure of the crossing within a complex shape. In contrast, these baseline models [4, 22, 23] struggle with such complex shape recovery under similar occlusion conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.785, + 0.646 + ], + "angle": 0, + "content": "Enhanced Robustness to Rotational Transformations. Figure 6(d) showcases that the BEV map is highly tilted by turning, which is extensively obscured by vehicular traffic. GeMap more adeptly maintains the integrity of lane width and parallelism, which are key geometry. This underscores the superior robustness of our geometric constructs against rotational distortions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.509, + 0.683 + ], + "angle": 0, + "content": "4.5 Hyperparameter Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Comparative Impact of Euclidean Losses. Figure 7a shows that GeMap exhibits larger sensitivity to the shape loss than to the relation loss, as evidenced by the steeper change of the performance curve. The optimal result is obtained when the shape and relation losses are relatively balanced by weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.81 + ], + "angle": 0, + "content": "Optimization of Euclidean Loss Weighting. The performance of the model, as demonstrated in Figure 7b, peaks when the weight of the Euclidean Loss, \\(\\lambda\\), is set to \\(5 \\times 10^{-3}\\). Notably, model performance degrades with a too large or too small \\(\\lambda\\), which informed us to use a \\(\\lambda\\) value of \\(5 \\times 10^{-3}\\) for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.842 + ], + "angle": 0, + "content": "Distance Weighting. Experimental results of distance weighting are presented in Table 5. We change the order of distance \\((p)\\) in Equation 5 and \\(p\\in \\{1,2,4\\}\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.501, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.288, + 0.38, + 0.299 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.145, + 0.761, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.288, + 0.639, + 0.299 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Fig. 7: Hyper-parameters. The best performance is reached with balanced Euclidean shape loss and relation loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.785, + 0.397 + ], + "angle": 0, + "content": "corresponds to \"Linear\", \"Square\" and \"4th Power\" respectively. Moreover, we also try to treat instance pairs equally and it delivers the best performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.32, + 0.421, + 0.68, + 0.435 + ], + "angle": 0, + "content": "Table 5: Distance weighting on the nuScenes dataset." + }, + { + "type": "table", + "bbox": [ + 0.36, + 0.448, + 0.639, + 0.523 + ], + "angle": 0, + "content": "
StrategyAP_div(↑)AP_ped(↑)AP_bnd(↑)mAP(↑)
Linear53.546.656.252.1
Square53.247.955.252.1
4th Power53.247.455.852.1
Equal53.649.254.852.6
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.36, + 0.586 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.779 + ], + "angle": 0, + "content": "In this paper, we realize significant shape and relation geometry inherent in HD map instances and propose the GeMap. GeMap includes the integration of Euclidean shape and relation losses for auxiliary supervision. To further refine the model's awareness of diverse geometry, we introduce the Geometry-Decoupled Attention mechanism. GeMap has achieved state-of-the-art performances on both the nuScenes and Argoverse 2 datasets, underscoring its effectiveness. Despite these promising results, the current application of geometry remains fundamental, and future research could focus on more sophisticated representations or enhanced geometric patterns. Furthermore, the application of geometry extends beyond HD map construction, offering potential solutions to occlusion challenges in other autonomous driving tasks. We anticipate that these findings will inspire further research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62272045, No. 8326014), The Shun Hing Institute of Advanced Engineering (No. 8115074), and CUHK Direct Grants (No. 4055190)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.786, + 0.236 + ], + "angle": 0, + "content": "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.786, + 0.278 + ], + "angle": 0, + "content": "2. Chen, S., Cheng, T., Wang, X., Meng, W., Zhang, Q., Liu, W.: Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.786, + 0.32 + ], + "angle": 0, + "content": "3. Deo, N., Wolff, E., Beijbom, O.: Multimodal trajectory prediction conditioned on lane-graph traversals. In: Conference on Robot Learning. pp. 203-212. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.322, + 0.786, + 0.363 + ], + "angle": 0, + "content": "4. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3672-3682 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.786, + 0.418 + ], + "angle": 0, + "content": "5. Ding, X., Zhang, Y., Ge, Y., Zhao, S., Song, L., Yue, X., Shan, Y.: Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5513-5524 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.786, + 0.475 + ], + "angle": 0, + "content": "6. Espinoza, J.L.V., Liniger, A., Schwarting, W., Rus, D., Van Gool, L.: Deep interactive motion prediction and planning: Playing games with motion prediction models. In: Learning for Dynamics and Control Conference. pp. 1006-1019. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.786, + 0.517 + ], + "angle": 0, + "content": "7. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking efficient lane detection via curve modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17062-17070 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.786, + 0.574 + ], + "angle": 0, + "content": "8. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.786, + 0.629 + ], + "angle": 0, + "content": "9. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for 3d detection and map segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8721-8731 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.786, + 0.685 + ], + "angle": 0, + "content": "0. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.786, + 0.741 + ], + "angle": 0, + "content": "1. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26584-26595 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.786, + 0.784 + ], + "angle": 0, + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "3. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Jiao, J.: Machine learning assisted high-definition map creation. In: 2018 IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC). vol. 1, pp. 367-373. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.23 + ], + "angle": 0, + "content": "15. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "16. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "17. Li, C., Shi, J., Wang, Y., Cheng, G.: Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4370-4379 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "18. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: An online hd map construction and evaluation framework. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 4628-4634. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.365, + 0.785, + 0.406 + ], + "angle": 0, + "content": "19. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.406, + 0.785, + 0.461 + ], + "angle": 0, + "content": "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.785, + 0.514 + ], + "angle": 0, + "content": "21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 541-556. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.785, + 0.57 + ], + "angle": 0, + "content": "22. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: MapTR: Structured modeling and learning for online vectorized HD map construction. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=k7p_YA07yE" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "23. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.785, + 0.691 + ], + "angle": 0, + "content": "25. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1765-1772 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "26. Liu, Y., Yuan, T., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: International Conference on Machine Learning. pp. 22352-22369. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.785, + 0.786 + ], + "angle": 0, + "content": "27. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Vectorized HD Map Construction using Geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.259 + ], + "angle": 0, + "content": "31. Loukkal, A., Grandvalet, Y., Drummond, T., Li, Y.: Driving among flatmobiles: Bird-eye-view occupancy grids from a monocular camera for holistic trajectory planning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 51-60 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "32. Lu, C., van de Molengraft, M.J.G., Dubbelman, G.: Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters 4(2), 445-452 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.356 + ], + "angle": 0, + "content": "33. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4227-4236 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "34. Pan, B., Sun, J., Leung, H.Y.T., Andonian, A., Zhou, B.: Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters 5(3), 4867-4873 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.453 + ], + "angle": 0, + "content": "35. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "36. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13218-13228 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "37. Scheel, O., Bergamini, L., Wolczyk, M., Osiński, B., Ondruska, P.: Urban driver: Learning to drive from real-world demonstrations using policy gradients. In: Conference on Robot Learning. pp. 718-728. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.591 + ], + "angle": 0, + "content": "38. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Keep your eyes on the lane: Real-time attention-guided lane detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 294-302 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.647 + ], + "angle": 0, + "content": "39. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Polylanenet: Lane estimation via deep polynomial regression. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6150-6156. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "40. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.744 + ], + "angle": 0, + "content": "41. Van Gansbeke, W., De Brabandere, B., Neven, D., Proesmans, M., Van Gool, L.: End-to-end lane detection through differentiable least-squares fitting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Wang, J., Ma, Y., Huang, S., Hui, T., Wang, F., Qian, C., Zhang, T.: A keypoint-based global association network for lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1392-1401 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "45. Xie, Z., Pang, Z., Wang, Y.X.: Mv-map: Offboard hd-map generation with multiview consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8658-8668 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.314 + ], + "angle": 0, + "content": "47. Yang, W., Li, Q., Liu, W., Yu, Y., Ma, Y., He, S., Pan, J.: Projecting your view attentively: Monocular road scene layout estimation via cross-view transformation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15536-15545 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "48. Yuan, T., Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Streammapnet: Streaming mapping network for vectorized online hd map construction. arXiv preprint arXiv:2308.12570 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "49. Zhang, G., Lin, J., Wu, S., Luo, Z., Xue, Y., Lu, S., Wang, Z., et al.: Online map vectorization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.453 + ], + "angle": 0, + "content": "50. Zhang, Y., Ding, X., Gong, K., Ge, Y., Shan, Y., Yue, X.: Multimodal pathway: Improve transformers with irrelevant data from other modalities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6108-6117 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.494 + ], + "angle": 0, + "content": "51. Zhang, Y., Gong, K., Zhang, K., Li, H., Qiao, Y., Ouyang, W., Yue, X.: Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "52. Zhang, Y., Li, H., Liu, J., Yue, X.: Explore the limits of omni-modal pretraining at scale. arXiv preprint arXiv:2406.09412 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "53. Zhang, Z., Zhang, Y., Ding, X., Jin, F., Yue, X.: Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "54. Zhou, B., Krahenbuhl, P.: Cross-view transformers for real-time map-view semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13760-13769 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.632 + ], + "angle": 0, + "content": "55. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_origin.pdf b/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..731da19f46830661f883cfde22ccce1d6931da63 --- /dev/null +++ b/2024/Online Vectorized HD Map Construction using Geometry/a879db55-840f-4c3f-9c6c-d74f1330f90f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68be614472dc2b5aef7a199ab71bf040845506200da8d1a8fff77ef585afa7ae +size 3960208 diff --git a/2024/Online Vectorized HD Map Construction using Geometry/full.md b/2024/Online Vectorized HD Map Construction using Geometry/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3c0ecd75c9e981dafb209d58f49cf4f89322ddbe --- /dev/null +++ b/2024/Online Vectorized HD Map Construction using Geometry/full.md @@ -0,0 +1,337 @@ +# Online Vectorized HD Map Construction using Geometry + +Zhixin Zhang $^{1}$ Yiyuan Zhang $^{2}$ Xiaohan Ding $^{3}$ Fusheng Jin $^{1*}$ Xiangyu Yue $^{2}$ + +$^{1}$ School of Computer Science and Technology, Beijing Institute of Technology + $^{2}$ The Chinese University of Hong Kong + $^{3}$ Tencent AI Lab +zhangzhixin@bit.edu.cn, yiyuanzhang.ai@gmail.com +https://invictus717.github.io/GeMap/ + +Abstract. Online vectorized High-Definition (HD) map construction is critical for downstream prediction and planning. Recent efforts have built strong baselines for this task, however, geometric shapes and relations of instances in road systems are still under-explored, such as parallelism, perpendicular, rectangle-shape, etc. In our work, we propose GeMap (Geometry Map), which end-to-end learns Euclidean shapes and relations of map instances beyond fundamental perception. Specifically, we design a geometric loss based on angle and magnitude clues, robust to rigid transformations of driving scenarios. To address the limitations of the vanilla attention mechanism in learning geometry, we propose to decouple self-attention to handle Euclidean shapes and relations independently. GeMap achieves new state-of-the-art performance on the nuScenes and Argoverse 2 datasets. Remarkably, it reaches a $71.8\%$ mAP on the large-scale Argoverse 2 dataset, outperforming MapTRv2 by $+4.4\%$ and surpassing the $70\%$ mAP threshold for the first time. Code is available at https://github.com/cnzzx/GeMap. + +Keywords: HD Map Construction $\cdot$ Geometry Representation $\cdot$ Geometry-Decoupled Attention + +# 1 Introduction + +Vectorized HD maps provide structured environmental information for autonomous vehicles and have been widely adopted in downstream tasks, such as trajectory forecasting [3,21,55] and planning [6,37]. Online Vectorized HD Map Construction can significantly reduce the need for labor-intensive annotations and facilitate real-time updates in autonomous driving [4,18,22,53]. + +With the development of Bird's-Eye-View (BEV) representation and multimodal learning [5, 11, 19, 28, 50–52], online HD map construction has achieved significant advancements [10, 20, 23]. Early works [14, 32, 35, 47] formulate HD map construction as a dense prediction task. However, these methods generate maps in image format, which is redundant for representing sparse map instances. Then, + +![](images/92ca623e88b2fd84784cd2f0faeff2fb84f6d63fc8b5e7df9c1904839468bbca.jpg) +(a) + +![](images/9315f629727cff8d2588bc14c9ea6f6d592cb5f056478a1c96a75230a01dd8e1.jpg) +(b) +Fig. 1: Geometric Invariance. (a) As the ego vehicle moves, after rotation $(\leftarrow)$ and translation $(\rightarrow)$ , the shape of the crossing and the parallelism between lanes remain unchanged, which indicates the invariant property of geometry to rigid transformations. (b) Absolute coordinates are vulnerable to rotation and translation, however, our G-representation is invariant, which is more suitable to capture geometric properties. + +a more compact map formulation was introduced to minimize redundancy, albeit at the cost of adding time-consuming post-processing steps [18]. In response to this, recent works [4, 22, 26, 36] attempt to end-to-end construct vectorized HD maps to avoid extensive post-processing. These methods typically sample points from map instances and represent each instance as a polyline [4, 22, 26] or parameterized curve [36]. + +We observe that transportation road systems exhibit significant geometric characteristics (Figure 2a), such as parallel lanes, perpendicular crossings, equal lane widths, etc. However, these geometric properties of shapes and relations between map instances have not been fully explored. Meanwhile, there are two notable limitations among existing methods [4,22,26,36] which can be alleviated by leveraging these geometric properties. 1) Excessive dependency on absolute coordinates: as the ego-vehicle moves, instances experience rotations and translations, as depicted in Figure 1a. However, widely adopted representations such as polylines and parameterized curves [8,36] are inherently sensitive to rotation and translation changes. 2) Objective conflicts of vanilla attention mechanism (detail in § 3.5) struggles to learn diverse shapes and relations, even though geometric properties such as rectangle shape, parallelism, and perpendicular relationships are commonly found in driving scenarios, as illustrated in Figure 2a. We believe that incorporating these geometric properties significantly enhances the precision and efficiency of online HD map construction. + +To address the above two limitations, we introduce a novel geometric representation that captures shapes of individual map instances and relations between different instances as illustrated in Figure 2b and 2c, referred to as G-Representation. It enhances the vanilla representation of map instances by incorporating a translation- and rotation-invariant representation that effectively leverages instance geometry. The local structures of map features are encoded using displacement vectors. These vectors are computed from the absolute coor + +![](images/07f2e2057c54d24fd03550135c8637d1125bb1c2b21b7de3e8b2120739a04938.jpg) +Rectangle-shaped + +![](images/cd70cba1aee6886e85e2ca7b39a9d9d9f03aecd4bf3fe395749cfcd4fe399f55.jpg) +Parallelism + +![](images/827a7e3e5aea6c3fd98b4a49a721fb3992a6ada01dd99a0fc93d5c9ad1a2777a.jpg) +Perpendicular + +![](images/b99d966c30efcbee61524067991e4180597ec7a55fca6056c5f3012b6dfeb21c.jpg) +Equal Lane Width + +![](images/48b1d99a7f95457c0eac6ee3ba5b37f8ec8fcee0e5b0da77fe9af0fcca134dce.jpg) +(a) +Single Map Instance +(b) + +![](images/b4e59fee30e6ad9c38c4c1f05a4b71b84719fb92349c0d4454851400ba57daad.jpg) +Cross Map Instances +(c) +Fig. 2: Geometric properties and G-Representation. (a) Geometry in the transportation road system. (b) (c) We propose to model the geometric properties of a single map instance and multiple instances, with magnitude $d$ and angle $\alpha$ . + +![](images/e21ada3d7040117a4a781a088eff69fc8da64d4715a51785f479701d03750b9d.jpg) + +dinates of polyline points to effectively represent the relative positions and orientations of adjacent points. To quantify these features within Euclidean space, we employ both the magnitudes of the displacement vectors and the angles formed between these vectors. Based on G-Representation, we propose the Euclidean Shape Clues to represent the shapes of individual map instances and the Euclidean Relation Clues to represent the inter-instance relations in Euclidean space. + +G-Representation is a simple yet effective method to address the abovementioned two limitations: 1) Geometric Invariance, by concentrating on the relative relationships between points within and across map instances, it inherently attains translation and rotation invariance as illustrated in Figure 1b. This enhances its robustness against variations in data collection and equips it to effectively handle different coordinate systems. 2) Euclidean Modeling, it captures the inherent geometry in transportation road systems. Diverse shapes and relation geometry can be simplified as enumerations of magnitudes and directions. For instance, in Euclidean space, parallel lanes can be easily modeled as approximate directions $(\alpha \approx 0)$ . + +Building upon G-Representation, we propose a framework named GeMap for HD map construction. A BEV encoder is used to extract features from multiview input images, while a geometry-decoupled decoder is employed to focus on geometric aspects. Specifically, we adapt the attention mechanism [42] and propose Geometry-Decoupled Attention (GDA). GDA sequentially applies attention to queries belonging to the same instance and attention to queries across different instances. This can significantly boost the geometry learning of key points of shapes and relations between map instances. Furthermore, we propose an objective function named Euclidean Loss to optimize G-Representation. + +Specifically, we transform the conventional polylines of the ground truth map into our G-Representation. In this way, the model gets optimized to better understand the magnitude of displacement vectors and the angles between them, thereby facilitating a more effective learning of geometric properties. + +Experiments on nuScenes and Argoverse 2 datasets demonstrate the effectiveness of GeMap. We reach new state-of-the-art performances on both datasets. With camera images only, GeMap achieves $69.4\%$ and $71.8\%$ mAP on the nuScenes and Argovserse 2, respectively. Visualization results (§ 4.4) further demonstrate the better perception of shape and relation, alleviation of occlusion, and robustness to rigid transformations of GeMap. + +Our contributions are summarized as the following: + +- We propose G-Representation which harnesses critical geometric properties of rotation and translation invariance in autonomous driving scenarios, opening up new research avenues within the field. +- We introduce GeMap, a novel framework incorporating geometry-decoupled attention and Euclidean Loss function, specifically designed to learn the intrinsic geometry of online HD maps. +- GeMap achieves new state-of-the-art results in HD map construction on the nuScenes and Argoverse 2 datasets, notably surpassing the $70\%$ mAP on the large-scale Argoverse 2 dataset for the first time. + +# 2 Related Work + +# 2.1 Online HD Map Construction + +Traditionally, HD map construction has required labor-intensive manual or semi-automatic annotations [14,33]. To streamline this, recent studies [20,31,34,35,54] have focused on online construction, approaching HD maps as a dense prediction challenge. Innovations such as MetaBEV [9] aim to mitigate sensor issues, while MVNet [45] uses historical data for improved semantic consistency. The trend towards automatic vectorization of HD maps is spearheaded by works such as HDMapNet [18], which fuses camera and LiDAR inputs in BEV space and is advanced by end-to-end solutions such as VectorMapNet [26] and MapTR [22], leveraging Transformer-based models. StreamMapNet [48] and PivotNet [4] build upon this, modifying attention mechanisms for better performance. Our contribution starts from the strengths of the Transformer architecture, optimizing it with a decoupled self-attention block for enhanced geometric processing. + +# 2.2 Cross-view BEV Learning + +The conversion of Perspective View (PV) camera images into a unified BEV space is a significant challenge for autonomous driving systems. Previous studies such as [19,35] use depth estimation from monocular images for this transformation, others [2, 20, 28] have developed PV-to-BEV conversion methods without + +explicit depths. However, these methods can distort shapes and geometric relations, leading to inaccuracies. Our approach integrates geometric supervision in prediction, addressing these geometric inconsistencies while acknowledging the value of depth information. + +# 2.3 Geometric Instance Modeling + +Traditional vectorization approaches model these instances as polylines [22, 26, 38, 43] or polynomial curves [7, 36, 39, 41], with some incorporating Bézier Curves for improved fitting [7, 36]. However, such methods often overlook geometric properties including shapes, parallelism, perpendicular, and etc. In 3D Lane Detection, some works [13, 17, 25] also explore simple geometry priors of lanes such as "equal lane width". However, they are not generalizable to more complex HD map instances. Existing works on modeling geometry properties of map instances mainly address rasterization-based shape loss [41, 49] or edge loss [22], and they fail to preserve cross-instance relationships. PivotNet [4] offers an architectural advancement with its line-aware point decoder, yet it still does not fully address the complexity of instance geometry. + +Given the current under-exploration of diverse geometric properties in shape and relation, we propose utilizing Euclidean shape and relation losses and decoupling the traditional self-attention module to empower the model with a more robust understanding of instance geometry. + +# 3 Method + +# 3.1 Preliminary + +The architecture of our method is illustrated in Figure 3. Input images are denoted by $\mathbb{I} = \{\pmb{I}_i\}_{i=1}^{N_c}$ , where $\pmb{I}_i \in \mathbb{R}^{H \times W \times 3}$ and $N_c$ is the number of cameras. The output is a set of $N$ map instances $\mathbb{M} = \{\pmb{L}_i\}_{i=1}^N$ , where each map instance is represented by a polyline $\pmb{L}_i \in \mathbb{R}^{N_v \times 2}$ , i.e., an ordered sequence of $N_v$ two-dimensional points. As shown in Figure 3, the commonly adopted pipeline comprises three steps. 1) A BEV feature extractor processes the multi-view images. 2) The extracted BEV features are fed into a map decoder, which predicts the map instances based on the BEV features. 3) To optimize the pipeline, the predicted map instances are compared against the ground truth. Beyond point-to-point comparison, G-representations of predictions and the ground truth are computed and the difference is measured by $L^1$ loss. + +# 3.2 Architecture Overview + +BEV Feature Extractor. A shared vision backbone in different views and parameterized PV-to-BEV transformation network are employed to aggregate features from various perspectives. In our default configuration, we utilize ResNet [12] as the PV backbone and GKT [2] as the PV-to-BEV transformation network. + +![](images/2157971095d7ce5ba6b9ed1002ad3c234a091c316a7ad66e8e84f682244c9602.jpg) +Fig. 3: Illustration of GeMap. First, PV images are transformed into BEV features, then a Geometry-Decoupled Decoder outputs the vectorized HD Map. In each block of the decoder, queries are first processed by Euclidean shape and relation attention, which focuses on geometric relevance. Finally, predictions are enhanced in G-Representations by shape and relation constraint. + +Geometry-Decoupled Decoder. We adopt a Transformer decoder to predict polylines. The decoder obtains $N \times N_v$ queries that represent points and processes them via self-attention and aggregate BEV features via deformable cross-attention [56]. Finally, a prediction head is used to convert queries to polylines $\{\hat{L}_i\}_{i=1}^N$ . We utilize a multi-layer perceptron as the polyline prediction head. + +Geometric Loss. For training, we convert both the predicted polylines and ground truth into the proposed G-Representation. Based on that, we let the model optimize the predicted magnitudes of displacement vectors and angles between displacement vectors to match the ground truth. + +In the following subsections, we first introduce Geometric Representation in § 3.3 since it is the core of our framework. Based on that, we introduce Geometric Loss in § 3.4 and Geometry-Decoupled Decoder in § 3.5, respectively. + +# 3.3 Geometric Representation + +Euclidean Shape Clues. We first introduce the representation of shapes of individual map instances. For each instance, we describe the local geometry with displacement vectors between neighboring points which are computed as: + +$$ +\boldsymbol {v} _ {u} ^ {i} = \boldsymbol {L} _ {i, u + 1} - \boldsymbol {L} _ {i, u} \quad (u \in \{1, 2, \dots , N _ {v} \}), \tag {1} +$$ + +where we define $\pmb{L}_{i,N_v + 1} \coloneqq \pmb{L}_{i,1}$ to unify the geometric formulation of closed and open polylines. + +![](images/fbcdb00e68bbf3e8240394691f1b3d7948f085f310d4230157ea632eb96eea42.jpg) +Fig. 4: Euclidean Shape Clues. Magnitudes of displacement vectors and angles between neighboring vectors indicate shape clues and are utilized to compute shape loss. The right part shows how to connect Euclidean Shape Clues to shape geometry. + +These displacement vectors are sufficient to represent the shape of a map instance and such a representation is invariant to translation transformations. However, we would like to note that this representation is vulnerable to rotations, which might prevent the model from learning robust geometry. To solve this problem, we propose to represent the shape with magnitudes of displacement vectors and angles between consecutive displacement vectors, as illustrated in Figure 4. Specifically, the $i$ -th instance is denoted by $N_v$ angle values and $N_v$ magnitude values. Let $u$ be the index, the $u$ -th angle value and the magnitude value of the $i$ -th instance are computed as: + +$$ +\alpha_ {u} ^ {i} = \left\langle \boldsymbol {v} _ {u} ^ {i}, \boldsymbol {v} _ {u + 1} ^ {i} \right\rangle , d _ {u} ^ {i} = \left\| \boldsymbol {v} _ {u} ^ {i} \right\| _ {2}, \tag {2} +$$ + +where $\langle \cdot \rangle$ denotes the angle between two vectors and similarly $\pmb{v}_{N_v + 1}^i \coloneqq \pmb{v}_1^i$ . + +The proposed representation inherently captures common geometric patterns, such as parallelism, right angles, and proper line width, by translating them into corresponding numerical patterns within this representation. For example, in Figure 4, a rectangle is characterized by one $90^{\circ}$ angle and equal magnitudes of opposite displacement vectors; in addition, a straight line is represented by $0^{\circ}$ angles. Beyond discussed regular cases, the Euclidean shape clues are capable of handling more complex shapes, as illustrated in scenarios (a) and (c) of Figure 6 and discussed in § 4.4. + +Euclidean Relation Clues. Having highlighted the translation and rotation invariance of G-Representation and its advantages in representing the geometric shapes of individual map instances, we further introduce its ability to capture the relations between two map instances, e.g., parallelism and perpendicular, in Euclidean space. Specifically, given the vanilla representations and displacement vectors of the $i$ -th and $j$ -th map instance, we represent the relation between instance $i$ and $j$ with 1) angles between each pair of their respective displacement vectors, and 2) magnitudes of displacement vectors between each pair of points, as illustrated in Figure 5. Formally, + +$$ +\alpha_ {u, v} ^ {i, j} = \left\langle \boldsymbol {v} _ {u} ^ {i}, \boldsymbol {v} _ {v} ^ {j} \right\rangle , d _ {u, v} ^ {i, j} = \left\| \boldsymbol {L} _ {i, u} - \boldsymbol {L} _ {j, v} \right\| _ {2}. \tag {3} +$$ + +This representation is also translation- and rotation-invariant. Similar to Euclidean Shape Clues, it inherently captures common relations between map instances by translating them into corresponding numerical patterns. For example, + +![](images/9982cddd7e67134a42fafcf015de569f554aea3ee509115cea600b499053cfa1.jpg) +Fig. 5: Euclidean Relation Clues. Angles between pairs of displacement vectors on different polylines, and magnitudes of displacement vectors between point pairs indicate relation clues. + +Figure 5 shows that the perpendicular relation can be directly represented by a $90^{\circ}$ angle, and the distance between two parallel polylines, which may correspond to the width of a lane, is naturally represented by magnitudes of displacement vector between point pairs. Scenarios (a) and (b) of Figure 6 show the effectiveness of Euclidean Relation Clues and more discussion can be viewed in $\S 4.4$ . + +# 3.4 Euclidean Loss and Objectives + +We transform both the ground truth data and the model's predicted polylines from their original format into G-Representation. This conversion allows the model to independently optimize the angles between displacement vectors and the magnitudes of displacement vectors to more accurately align with the ground truth. The proposed Euclidean Loss is composed of two parts that measure how accurately the model predicts the shape of individual map instances and the inter-instance relations, respectively, which are denoted by $\mathcal{L}_{\mathrm{shp}}$ and $\mathcal{L}_{\mathrm{rel}}$ . The Euclidean Loss can be computed as the following: + +$$ +\mathcal {L} _ {\mathrm {E u c}} = \lambda_ {1} \cdot \mathcal {L} _ {\mathrm {s h p}} + \lambda_ {2} \cdot \mathcal {L} _ {\mathrm {r e l}}, \tag {4} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters, whose effects on the model performances are evaluated in Figure 7a. Specifically, + +$$ +\begin{array}{l} w _ {i, j} = 1 - \left(\min _ {u, v} \left\{d _ {u, v} ^ {i, j} \right\} / \sqrt {2}\right) ^ {p}, \quad \mathcal {L} _ {\mathrm {s h p}} = \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N _ {v}} \left| \hat {d} _ {j} ^ {i} - d _ {j} ^ {i} \right| + \ell \left(\hat {\alpha} _ {j} ^ {i}, \alpha_ {j} ^ {i}\right), \tag {5} \\ \mathcal {L} _ {\mathrm {r e l}} = \sum_ {i = 1} ^ {N} \sum_ {j > i} ^ {N} w _ {i, j} \sum_ {u = 1} ^ {N _ {v}} \sum_ {v = 1} ^ {N _ {v}} \left(| \hat {d} _ {u, v} ^ {i, j} - d _ {u, v} ^ {i, j} | + \ell (\hat {\alpha} _ {u, v} ^ {i, j}, \alpha_ {u, v} ^ {i, j})\right), \\ \end{array} +$$ + +where $\ell(\cdot)$ is a function based on $L^1$ loss. We avoid using the inverse trigonometric function to directly compute the angles, but use sine and cosine values instead: + +$$ +\ell (\hat {\alpha}, \alpha) = | \cos (\hat {\alpha}) - \cos (\alpha) | + | \sin (\hat {\alpha}) - \sin (\alpha) |. \tag {6} +$$ + +Moreover, the distance might influence the relation strength of instance pairs. For example, if two instances are far from each other, the relation between them + +might be weak. For this reason, we further adopt $w_{i,j}$ to punish weakly related instance pairs and discuss it in § 4.5. According to the best experimental results, we treat all instance pairs equally in other experiments. + +Following common practices [22,48], we also use focal loss [24] $\mathcal{L}_{\mathrm{cls}}$ for classification. For polyline regression, we adopt point-to-point loss and edge direction loss, denoted by $\mathcal{L}_{\mathrm{pts}}$ and $\mathcal{L}_{\mathrm{dir}}$ respectively. Following [23,36], we further adopt segmentation loss $\mathcal{L}_{\mathrm{seg}}$ and depth estimation loss $\mathcal{L}_{\mathrm{dep}}$ . These losses are detailed in appendix § A.1. The overall loss function can be written as: + +$$ +\mathcal {L} = \lambda \cdot \mathcal {L} _ {\text {E u c}} + \beta_ {1} \cdot \mathcal {L} _ {\text {c l s}} + \beta_ {2} \cdot \mathcal {L} _ {\text {p t s}} + \beta_ {3} \cdot \mathcal {L} _ {\text {d i r}} + \beta_ {4} \cdot \mathcal {L} _ {\text {s e g}} + \beta_ {5} \cdot \mathcal {L} _ {\text {d e p}}. \tag {7} +$$ + +where hyperparameters are discussed in Figure 7b and appendix Table A1. + +# 3.5 Geometry-Decoupled Decoder + +In the Transformer decoder, we obtain the $N \times N_v$ fused queries by adding $N$ instance queries to $N_v$ point queries, which is suggested in [22]. We denote input tokens by $\pmb{E} \in \mathbb{R}^{N_A \times D}$ , where $D$ is the feature dimension and $N_A = N \cdot N_v$ is the number of fused queries. Intuitively, each fused query corresponds to a point on a predicted map instance. Self-Attention (SA) is formulated by: + +$$ +\operatorname {S A} (\boldsymbol {E}, \boldsymbol {M}) = \operatorname {S o f t m a x} \left(\frac {\left(\boldsymbol {E} \boldsymbol {W} ^ {q}\right) \left(\boldsymbol {E} \boldsymbol {W} ^ {k}\right) ^ {\top}}{\sqrt {D _ {k}}} \odot \boldsymbol {M}\right) \boldsymbol {E} \boldsymbol {W} ^ {v}, \tag {8} +$$ + +where $\pmb{W}^q, \pmb{W}^k, \pmb{W}^v \in \mathbb{R}^{D \times D_k}$ are linear projection matrices, $\pmb{M} \in \mathbb{R}^{N_A \times N_A}$ is the attention mask, and $\odot$ is the Hadamard product. The vanilla SA computes relations between every pair of tokens. + +Nevertheless, the geometry of shape and relation pertains to distinct subsets of tokens. For any given map instance, its shape is intimately related to tokens representing that instance's points. Precise shape geometry capture requires the model to discern token correlations specific to an instance while avoiding interference from tokens of unrelated instances. Conversely, for relation geometry modeling, it is beneficial to isolate token correlations that span across different instances, rather than those confined within a single instance. + +Therefore, we present Geometry-Decoupled Attention (GDA). First, we multiply a binary mask $M$ to the computed attention map so that the tokens of a map instance are aggregated according to the tokens within the same map instances only, which allows the model to adjust points' positions according to learned shape geometry. Denote the index of map instance that the $i$ -th token belongs to as $\mathcal{I}_i$ . For example, tokens $1,2,\ldots,N_v$ belong to the first instance so that $\mathcal{I}_1,\mathcal{I}_2,\ldots,\mathcal{I}_{N_v}=1$ . Tokens $N_v+1,N_v+2,\ldots,2\cdot N_v$ belong to the second instance so that $\mathcal{I}_{N_v+1},\mathcal{I}_{N_v+2},\ldots,\mathcal{I}_{2N_v}=2$ . With this notation, the binary mask $M$ can be simply constructed by: + +$$ +\boldsymbol {M} _ {i, j} ^ {\mathrm {s h p}} = \left\{ \begin{array}{l l} 1 & , \mathcal {I} _ {i} = \mathcal {I} _ {j} \\ 0 & , \mathcal {I} _ {i} \neq \mathcal {I} _ {j} \end{array} , \quad \boldsymbol {M} _ {i, j} ^ {\mathrm {r e l}} = \left\{ \begin{array}{l l} 1 & , \mathcal {I} _ {i} \neq \mathcal {I} _ {j} \\ 0 & , \mathcal {I} _ {i} = \mathcal {I} _ {j} \end{array} . \right. \right. \tag {9} +$$ + +The second attention is expected to model relations between tokens of different map instances. + +Table 1: Comparison on the nuScenes dataset, GeMap reaches a new state-of-the-art performance. "EB0", "R50", "PP", "Sec", "Swin-T", and "V2-99" denote EfficientNet-B0 [40], ResNet50 [12], PointPillars [15], SECOND [46], Swin Transformer Tiny [27], and VoVNetV2-99 [16] respectively. Methods with two backbones utilize both camera and LiDAR inputs. "Dns. Loss" denotes whether any dense prediction (e.g. semantic segmentation) loss is adopted. The best result is highlighted in **bold**. We reproduce all methods on a single RTX3090 GPU to test FPS for fair comparison. + +
MethodsBackboneDns. Loss\( AP_{div}(\uparrow) \)\( AP_{ped}(\uparrow) \)\( AP_{bnd}(\uparrow) \)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5042.551.444.146.05.3
MapTR [ICLR'23] [22]R5059.856.260.158.719.8
MapVR [NeurIPS'23] [49]R5061.855.059.458.819.8
GeMap [Ours]R5065.1 ↑ 3.359.8 ↑ 4.863.2 ↑ 3.862.7 ↑ 3.919.0
HDMapNet [ICRA'22] [18]EB0✓14.421.733.023.00.7
PivotNet [ICCV'23] [4]R50✓58.853.859.657.49.5
BeMapNet [CVPR'23] [36]R50✓66.762.665.164.86.6
MapTRv2 [Arxiv'23] [23]R50✓68.368.169.768.715.0
GeMap [Ours]R50✓69.8 ↑ 1.567.171.4 ↑ 1.769.4 ↑ 0.715.8
HDMapNet [ICRA'22] [18]EB0 & PP29.616.346.731.00.7
VectorMapNet [ICML'23] [26]R50 & PP60.148.253.053.7-
MapTR [ICLR'23] [22]R50 & Sec62.355.969.362.56.7
MapVR [NeurIPS'23] [49]R50 & PP62.760.467.263.5-
GeMap [Ours]R50 & Sec66.3 ↑ 3.662.2 ↑ 1.871.1 ↑ 3.966.5 ↑ 3.07.3
MapTRv2 [Arxiv'23] [23]R50 & Sec✓65.666.574.869.06.6
GeMap [Ours]R50 & Sec✓69.8 ↑ 4.268.0 ↑ 1.573.470.4 ↑ 1.46.8
MapTRv2 [Arxiv'23] [23]V2-99✓73.771.475.073.410.1
GeMap [Ours]Swin-T✓72.870.472.872.011.4
GeMap [Ours]V2-99✓76.0 ↑ 2.374.3 ↑ 2.977.7 ↑ 2.776.0 ↑ 2.610.8
+ +# 4 Experiments + +# 4.1 Experimental Setups + +Datasets. To evaluate GeMap, we conduct experiments on the nuScenes dataset [1], a widely adopted large-scale autonomous driving dataset that includes 1,000 scenes captured by six RGB cameras with a 360-degree field of view, and provides precise annotations from LiDAR point clouds for HD map construction. For comparability with prior research [18,22], we focus on three static categories of map instances: pedestrian crossings, lane dividers, and road boundaries. With dataset splits provided by BeMapNet [36], we also evaluate GeMaps under three weather conditions: sunny, cloudy, and rainy. Additionally, we use the Argoverse 2 dataset [44] as another benchmark, which consists of approximately 108,000 frames, each providing images from seven cameras. We evaluate the performance by Average Precision (AP) [18,22] with thresholds at 0.5, 1.0, and 1.5 meters. + +Implementation Details. GeMap leverages 8 NVIDIA RTX 3090 GPUs for training. We adopt AdamW [30] as the optimizer and utilize Cosine Annealing with a linear warm-up phase [29] as the learning rate scheduler. For more details on hyperparameters, please refer to appendix § A.2. + +Table 2: Following BeMapNet [36], we also compare GeMap on the nuScenes dataset under different weather conditions. + +
MethodsBackbone Dns. LossmAPsun(↑)mAPcld(↑)mAPrny(↑)mAPavg(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5043.844.136.641.55.3
MapTR [ICLR'23] [22]R5062.160.552.858.419.8
GeMap [Ours]R5066.0 ↑ 3.964.3 ↑ 3.854.4 ↑ 1.661.5 ↑ 3.119.0
BeMapNet [CVPR'23] [36]R50 √67.367.556.663.86.6
GeMap [Ours]R50 √73.1 ↑ 5.871.0 ↑ 3.559.3 ↑ 2.767.8 ↑ 4.015.8
+ +Table 3: Comparison on Argoverse 2 dataset. GeMap demonstrates significant performance improvements over previous methods. + +
MethodsBackbone Dns. Loss\( AP_{div}(\uparrow) \)\( AP_{ped}(\uparrow) \)\( AP_{bnd}(\uparrow) \)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5036.138.339.237.9-
GeMap [Ours]R5067.6 ↑ 31.559.3 ↑ 21.064.7 ↑ 25.563.9 ↑ 26.016.7
HMapNet [ICRA'22] [18]EB0✓5.713.137.618.8
MapTRv2 [Arxiv'23] [23]R50✓72.162.967.167.4
GeMap [Ours]R50✓75.7 ↑ 3.669.2 ↑ 6.370.5 ↑ 3.471.8 ↑ 4.4
+ +# 4.2 Main Results + +Results on nuScenes. As delineated in Table 1, GeMap delivers a state-of-the-art performance, achieving a mean Average Precision (mAP) of $69.4\%$ using the camera-only setup and ResNet50 [12] backbone. GeMap particularly improves the precision for identifying dividers and boundaries, with enhancements of $+1.5\%$ and $+1.7\%$ , respectively. Importantly, these performance gains do not come at a high cost of efficiency, as GeMap sustains inference speeds on par with and even exceeds the previously established state-of-the-art, measured in frames per second (FPS). With more powerful vision backbones, GeMap can achieve significantly improved performance. Specifically, VoVNetV2-99 [16] and Swin-T [27] outperform the ResNet50 baseline by $+6.6\%$ and $+2.6\%$ in mAP, respectively. Also, GeMap achieves significantly improved performance with extra LiDAR inputs, which indicates the generalizability of GeMap to multi-modality settings. Moreover, we also compare performance under different weather conditions in Table 2. GeMap exceeds previous works under sunny, cloudy, and rainy scenarios, which shows the potential of geometry to raise model robustness to varied weather conditions. + +Results on Argoverse 2. Extending our evaluation to the Argoverse 2 dataset, as presented in Table 3, GeMap also presents a new SOTA performance of $71.8\%$ mAP, outperforming the highly advanced MapTRv2 model by $+4.4\%$ . This achievement not only demonstrates the effectiveness of GeMap on a different dataset but also emphasizes the adaptability and precision of GeMap in the evolving landscape of autonomous driving technologies. + +Table 4: Ablation study on the nuScenes dataset. We train the model for 24 epochs. * denotes replacing GDA with 2 layers of vanilla self-attention. + +
Method\( {\mathcal{L}}_{\text{shp }}\left( {\S {3.3}}\right) {\mathcal{L}}_{\text{rel }}\left( {\S {3.3}}\right) \)GDA (§ 3.5)\( {\mathrm{{AP}}}_{div}\left( \uparrow \right) \)\( {\mathrm{{AP}}}_{ped}\left( \uparrow \right) \)\( {\mathrm{{AP}}}_{bnd}\left( \uparrow \right) \)mAP(↑)
Baseline49.544.753.749.3
+ Decoupled Attention✓53.446.653.551.2
+ Single Euclidean Loss✓51.045.452.749.7
51.743.453.049.4
+ Euclidean Loss✓✓51.543.951.148.8
+ Single Euclidean Loss with Decoupled Attention✓✓54.048.253.151.8
✓✓54.747.355.352.4
Replace GDA with 2-SA✓✓\( 2 - SA^* \)53.546.254.451.4
Full✓✓✓53.6 ↑ 4.149.2 ↑ 4.554.8 ↑ 1.152.6 ↑ 3.3
+ +# 4.3 Ablation Study + +Components Ablation. The ablation experiments, detailed in Table 4, involved training the model for a reduced duration of 24 epochs without dense prediction losses, to facilitate efficient analysis. These experiments confirm the significant role of GDA, which alone increases mAP by $+1.9\%$ . Moreover, applying the full suite of components results in a further mAP enhancement of $+1.4\%$ . A notable discovery is the detrimental effect on model performance when Euclidean Loss is applied without GDA, which leads to a $0.5\%$ reduction in mAP. This reinforces our position that conventional self-attention mechanisms are insufficient for encoding a variety of geometric properties. + +Effectiveness of GDA. In the proposed GDA (3.5), we decouple shape and relation learning, sequentially applying Euclidean shape and relation attention. In addition, we also attempt to intuitively double the self-attention layers in one block, referred to as "2-SA". As shown in Table 4, GDA outperforms vanilla self-attention by $+1.2\%$ mAP, indicating the superiority of GDA when combined with Euclidean Loss. + +# 4.4 Visualization Analysis + +More Precise Shape Awareness. As shown in Figure 6(a), the shape geometry allows for an inherent description of the triangle feature, in contrast to baselines which do not discern it as effectively. A more irregularly shaped pedestrian crossing can be viewed in scenario (c) of Figure 6 and GeMap provides more precise construction than other methods. + +Better Relation Awareness. With the help of Euclidean Relation Clues, GeMap can better infer parallelism and perpendicular in scenario (a) of Figure 6. More interestingly, in scenario (b), the strong sunlight makes the divider highlighted hard to recognize. However, according to the understanding of lane width patterns, GeMap has the sense that there should be a divider. + +Alleviating Occlusion Issues. Figure 6(c) demonstrates that shape geometry alleviates occlusions in partially visible instances. It showcases a black car that obscures part of a pedestrian crossing, while, GeMap successfully deduces the + +![](images/613ff0d6cc99954666acbab67f745803be1f04fea88dc1a5e6ad965b81431896.jpg) +Fig. 6: Visualization results. Hard predictions are highlighted, such as complex triangular road boundary (a), strong sunlight (b), partially observed (c), and obscured situations (d). These challenging cases indicate the superiority and robustness of GeMap. + +overall structure of the crossing within a complex shape. In contrast, these baseline models [4, 22, 23] struggle with such complex shape recovery under similar occlusion conditions. + +Enhanced Robustness to Rotational Transformations. Figure 6(d) showcases that the BEV map is highly tilted by turning, which is extensively obscured by vehicular traffic. GeMap more adeptly maintains the integrity of lane width and parallelism, which are key geometry. This underscores the superior robustness of our geometric constructs against rotational distortions. + +# 4.5 Hyperparameter Experiments + +Comparative Impact of Euclidean Losses. Figure 7a shows that GeMap exhibits larger sensitivity to the shape loss than to the relation loss, as evidenced by the steeper change of the performance curve. The optimal result is obtained when the shape and relation losses are relatively balanced by weights. + +Optimization of Euclidean Loss Weighting. The performance of the model, as demonstrated in Figure 7b, peaks when the weight of the Euclidean Loss, $\lambda$ , is set to $5 \times 10^{-3}$ . Notably, model performance degrades with a too large or too small $\lambda$ , which informed us to use a $\lambda$ value of $5 \times 10^{-3}$ for our experiments. + +Distance Weighting. Experimental results of distance weighting are presented in Table 5. We change the order of distance $(p)$ in Equation 5 and $p\in \{1,2,4\}$ + +![](images/c8848dbd709d1278610f4228219e8e6f6d28ea9c757035d13827c202780ba3fb.jpg) +(a) + +![](images/67515aafdedc2b0aa1b1799256f4835fcc81c6a2cbcb431c8b8a7f7ee992f880.jpg) +(b) +Fig. 7: Hyper-parameters. The best performance is reached with balanced Euclidean shape loss and relation loss. + +corresponds to "Linear", "Square" and "4th Power" respectively. Moreover, we also try to treat instance pairs equally and it delivers the best performance. + +Table 5: Distance weighting on the nuScenes dataset. + +
StrategyAP_div(↑)AP_ped(↑)AP_bnd(↑)mAP(↑)
Linear53.546.656.252.1
Square53.247.955.252.1
4th Power53.247.455.852.1
Equal53.649.254.852.6
+ +# 5 Conclusion + +In this paper, we realize significant shape and relation geometry inherent in HD map instances and propose the GeMap. GeMap includes the integration of Euclidean shape and relation losses for auxiliary supervision. To further refine the model's awareness of diverse geometry, we introduce the Geometry-Decoupled Attention mechanism. GeMap has achieved state-of-the-art performances on both the nuScenes and Argoverse 2 datasets, underscoring its effectiveness. Despite these promising results, the current application of geometry remains fundamental, and future research could focus on more sophisticated representations or enhanced geometric patterns. Furthermore, the application of geometry extends beyond HD map construction, offering potential solutions to occlusion challenges in other autonomous driving tasks. We anticipate that these findings will inspire further research. + +Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62272045, No. 8326014), The Shun Hing Institute of Advanced Engineering (No. 8115074), and CUHK Direct Grants (No. 4055190). + +# References + +1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +2. Chen, S., Cheng, T., Wang, X., Meng, W., Zhang, Q., Liu, W.: Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584 (2022) +3. Deo, N., Wolff, E., Beijbom, O.: Multimodal trajectory prediction conditioned on lane-graph traversals. In: Conference on Robot Learning. pp. 203-212. PMLR (2022) +4. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3672-3682 (2023) +5. Ding, X., Zhang, Y., Ge, Y., Zhao, S., Song, L., Yue, X., Shan, Y.: Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5513-5524 (2024) +6. Espinoza, J.L.V., Liniger, A., Schwarting, W., Rus, D., Van Gool, L.: Deep interactive motion prediction and planning: Playing games with motion prediction models. In: Learning for Dynamics and Control Conference. pp. 1006-1019. PMLR (2022) +7. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking efficient lane detection via curve modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17062-17070 (2022) +8. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020) +9. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for 3d detection and map segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8721-8731 (2023) +0. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023) +1. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26584-26595 (2024) +2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +3. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023) + +14. Jiao, J.: Machine learning assisted high-definition map creation. In: 2018 IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC). vol. 1, pp. 367-373. IEEE (2018) +15. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019) +16. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) +17. Li, C., Shi, J., Wang, Y., Cheng, G.: Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4370-4379 (2022) +18. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: An online hd map construction and evaluation framework. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 4628-4634. IEEE (2022) +19. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023) +20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022) +21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 541-556. Springer (2020) +22. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: MapTR: Structured modeling and learning for online vectorized HD map construction. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=k7p_YA07yE +23. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736 (2023) +24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +25. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1765-1772 (2022) +26. Liu, Y., Yuan, T., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: International Conference on Machine Learning. pp. 22352-22369. PMLR (2023) +27. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023) + +29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +31. Loukkal, A., Grandvalet, Y., Drummond, T., Li, Y.: Driving among flatmobiles: Bird-eye-view occupancy grids from a monocular camera for holistic trajectory planning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 51-60 (2021) +32. Lu, C., van de Molengraft, M.J.G., Dubbelman, G.: Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters 4(2), 445-452 (2019) +33. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4227-4236 (2021) +34. Pan, B., Sun, J., Leung, H.Y.T., Andonian, A., Zhou, B.: Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters 5(3), 4867-4873 (2020) +35. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020) +36. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13218-13228 (2023) +37. Scheel, O., Bergamini, L., Wolczyk, M., Osiński, B., Ondruska, P.: Urban driver: Learning to drive from real-world demonstrations using policy gradients. In: Conference on Robot Learning. pp. 718-728. PMLR (2022) +38. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Keep your eyes on the lane: Real-time attention-guided lane detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 294-302 (2021) +39. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Polylanenet: Lane estimation via deep polynomial regression. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6150-6156. IEEE (2021) +40. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019) +41. Van Gansbeke, W., De Brabandere, B., Neven, D., Proesmans, M., Van Gool, L.: End-to-end lane detection through differentiable least-squares fitting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) +42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +43. Wang, J., Ma, Y., Huang, S., Hui, T., Wang, F., Qian, C., Zhang, T.: A keypoint-based global association network for lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1392-1401 (2022) + +44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023) +45. Xie, Z., Pang, Z., Wang, Y.X.: Mv-map: Offboard hd-map generation with multiview consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8658-8668 (2023) +46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +47. Yang, W., Li, Q., Liu, W., Yu, Y., Ma, Y., He, S., Pan, J.: Projecting your view attentively: Monocular road scene layout estimation via cross-view transformation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15536-15545 (2021) +48. Yuan, T., Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Streammapnet: Streaming mapping network for vectorized online hd map construction. arXiv preprint arXiv:2308.12570 (2023) +49. Zhang, G., Lin, J., Wu, S., Luo, Z., Xue, Y., Lu, S., Wang, Z., et al.: Online map vectorization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems 36 (2024) +50. Zhang, Y., Ding, X., Gong, K., Ge, Y., Shan, Y., Yue, X.: Multimodal pathway: Improve transformers with irrelevant data from other modalities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6108-6117 (2024) +51. Zhang, Y., Gong, K., Zhang, K., Li, H., Qiao, Y., Ouyang, W., Yue, X.: Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802 (2023) +52. Zhang, Y., Li, H., Liu, J., Yue, X.: Explore the limits of omni-modal pretraining at scale. arXiv preprint arXiv:2406.09412 (2024) +53. Zhang, Z., Zhang, Y., Ding, X., Jin, F., Yue, X.: Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341 (2023) +54. Zhou, B., Krahenbuhl, P.: Cross-view transformers for real-time map-view semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13760-13769 (2022) +55. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022) +56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) \ No newline at end of file diff --git a/2024/Online Vectorized HD Map Construction using Geometry/images.zip b/2024/Online Vectorized HD Map Construction using Geometry/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..510175ab80759cb2916d628e674c98f40ba24715 --- /dev/null +++ b/2024/Online Vectorized HD Map Construction using Geometry/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d89ce46a45a2e318bc33781ee3e11849ebdc8c8ad089c27bf28e85cee9eba2e1 +size 593225 diff --git a/2024/Online Vectorized HD Map Construction using Geometry/layout.json b/2024/Online Vectorized HD Map Construction using Geometry/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..873d0f191d8c5e96398b058c5d126973cc40ecee --- /dev/null +++ b/2024/Online Vectorized HD Map Construction using Geometry/layout.json @@ -0,0 +1,9988 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 110, + 471, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 110, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 110, + 471, + 148 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": "Zhixin Zhang" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": " Yiyuan Zhang" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": " Xiaohan Ding" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": " Fusheng Jin" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "text", + "content": " Xiangyu Yue" + }, + { + "bbox": [ + 134, + 167, + 479, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "spans": [ + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "type": "text", + "content": "School of Computer Science and Technology, Beijing Institute of Technology \n" + }, + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "type": "text", + "content": " The Chinese University of Hong Kong \n" + }, + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 190, + 463, + 236 + ], + "type": "text", + "content": "Tencent AI Lab \nzhangzhixin@bit.edu.cn, yiyuanzhang.ai@gmail.com \nhttps://invictus717.github.io/GeMap/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "text", + "content": "Abstract. Online vectorized High-Definition (HD) map construction is critical for downstream prediction and planning. Recent efforts have built strong baselines for this task, however, geometric shapes and relations of instances in road systems are still under-explored, such as parallelism, perpendicular, rectangle-shape, etc. In our work, we propose GeMap (Geometry Map), which end-to-end learns Euclidean shapes and relations of map instances beyond fundamental perception. Specifically, we design a geometric loss based on angle and magnitude clues, robust to rigid transformations of driving scenarios. To address the limitations of the vanilla attention mechanism in learning geometry, we propose to decouple self-attention to handle Euclidean shapes and relations independently. GeMap achieves new state-of-the-art performance on the nuScenes and Argoverse 2 datasets. Remarkably, it reaches a " + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "text", + "content": " mAP on the large-scale Argoverse 2 dataset, outperforming MapTRv2 by " + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "inline_equation", + "content": "+4.4\\%" + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "text", + "content": " and surpassing the " + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "text", + "content": " mAP threshold for the first time. Code is available at https://github.com/cnzzx/GeMap." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 454, + 463, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 454, + 463, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 454, + 463, + 475 + ], + "type": "text", + "content": "Keywords: HD Map Construction " + }, + { + "bbox": [ + 160, + 454, + 463, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 454, + 463, + 475 + ], + "type": "text", + "content": " Geometry Representation " + }, + { + "bbox": [ + 160, + 454, + 463, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 454, + 463, + 475 + ], + "type": "text", + "content": " Geometry-Decoupled Attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 525, + 489, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 489, + 585 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 489, + 585 + ], + "type": "text", + "content": "Vectorized HD maps provide structured environmental information for autonomous vehicles and have been widely adopted in downstream tasks, such as trajectory forecasting [3,21,55] and planning [6,37]. Online Vectorized HD Map Construction can significantly reduce the need for labor-intensive annotations and facilitate real-time updates in autonomous driving [4,18,22,53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 586, + 481, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 586, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 133, + 586, + 481, + 645 + ], + "type": "text", + "content": "With the development of Bird's-Eye-View (BEV) representation and multimodal learning [5, 11, 19, 28, 50–52], online HD map construction has achieved significant advancements [10, 20, 23]. Early works [14, 32, 35, 47] formulate HD map construction as a dense prediction task. However, these methods generate maps in image format, which is redundant for representing sparse map instances. Then," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 233, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 233, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 233, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 116, + 297, + 220 + ], + "blocks": [ + { + "bbox": [ + 148, + 116, + 297, + 220 + ], + "lines": [ + { + "bbox": [ + 148, + 116, + 297, + 220 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 297, + 220 + ], + "type": "image", + "image_path": "92ca623e88b2fd84784cd2f0faeff2fb84f6d63fc8b5e7df9c1904839468bbca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 224, + 227, + 234 + ], + "lines": [ + { + "bbox": [ + 213, + 224, + 227, + 234 + ], + "spans": [ + { + "bbox": [ + 213, + 224, + 227, + 234 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 114, + 470, + 222 + ], + "blocks": [ + { + "bbox": [ + 302, + 114, + 470, + 222 + ], + "lines": [ + { + "bbox": [ + 302, + 114, + 470, + 222 + ], + "spans": [ + { + "bbox": [ + 302, + 114, + 470, + 222 + ], + "type": "image", + "image_path": "9315f629727cff8d2588bc14c9ea6f6d592cb5f056478a1c96a75230a01dd8e1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 224, + 389, + 234 + ], + "lines": [ + { + "bbox": [ + 376, + 224, + 389, + 234 + ], + "spans": [ + { + "bbox": [ + 376, + 224, + 389, + 234 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 243, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 131, + 243, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 482, + 299 + ], + "type": "text", + "content": "Fig. 1: Geometric Invariance. (a) As the ego vehicle moves, after rotation " + }, + { + "bbox": [ + 131, + 243, + 482, + 299 + ], + "type": "inline_equation", + "content": "(\\leftarrow)" + }, + { + "bbox": [ + 131, + 243, + 482, + 299 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 131, + 243, + 482, + 299 + ], + "type": "inline_equation", + "content": "(\\rightarrow)" + }, + { + "bbox": [ + 131, + 243, + 482, + 299 + ], + "type": "text", + "content": ", the shape of the crossing and the parallelism between lanes remain unchanged, which indicates the invariant property of geometry to rigid transformations. (b) Absolute coordinates are vulnerable to rotation and translation, however, our G-representation is invariant, which is more suitable to capture geometric properties." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 399 + ], + "type": "text", + "content": "a more compact map formulation was introduced to minimize redundancy, albeit at the cost of adding time-consuming post-processing steps [18]. In response to this, recent works [4, 22, 26, 36] attempt to end-to-end construct vectorized HD maps to avoid extensive post-processing. These methods typically sample points from map instances and represent each instance as a polyline [4, 22, 26] or parameterized curve [36]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 400, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 581 + ], + "type": "text", + "content": "We observe that transportation road systems exhibit significant geometric characteristics (Figure 2a), such as parallel lanes, perpendicular crossings, equal lane widths, etc. However, these geometric properties of shapes and relations between map instances have not been fully explored. Meanwhile, there are two notable limitations among existing methods [4,22,26,36] which can be alleviated by leveraging these geometric properties. 1) Excessive dependency on absolute coordinates: as the ego-vehicle moves, instances experience rotations and translations, as depicted in Figure 1a. However, widely adopted representations such as polylines and parameterized curves [8,36] are inherently sensitive to rotation and translation changes. 2) Objective conflicts of vanilla attention mechanism (detail in § 3.5) struggles to learn diverse shapes and relations, even though geometric properties such as rectangle shape, parallelism, and perpendicular relationships are commonly found in driving scenarios, as illustrated in Figure 2a. We believe that incorporating these geometric properties significantly enhances the precision and efficiency of online HD map construction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To address the above two limitations, we introduce a novel geometric representation that captures shapes of individual map instances and relations between different instances as illustrated in Figure 2b and 2c, referred to as G-Representation. It enhances the vanilla representation of map instances by incorporating a translation- and rotation-invariant representation that effectively leverages instance geometry. The local structures of map features are encoded using displacement vectors. These vectors are computed from the absolute coor" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 117, + 236, + 178 + ], + "blocks": [ + { + "bbox": [ + 157, + 117, + 236, + 178 + ], + "lines": [ + { + "bbox": [ + 157, + 117, + 236, + 178 + ], + "spans": [ + { + "bbox": [ + 157, + 117, + 236, + 178 + ], + "type": "image", + "image_path": "07f2e2057c54d24fd03550135c8637d1125bb1c2b21b7de3e8b2120739a04938.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 183, + 227, + 194 + ], + "lines": [ + { + "bbox": [ + 159, + 183, + 227, + 194 + ], + "spans": [ + { + "bbox": [ + 159, + 183, + 227, + 194 + ], + "type": "text", + "content": "Rectangle-shaped" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 115, + 301, + 180 + ], + "blocks": [ + { + "bbox": [ + 248, + 115, + 301, + 180 + ], + "lines": [ + { + "bbox": [ + 248, + 115, + 301, + 180 + ], + "spans": [ + { + "bbox": [ + 248, + 115, + 301, + 180 + ], + "type": "image", + "image_path": "cd70cba1aee6886e85e2ca7b39a9d9d9f03aecd4bf3fe395749cfcd4fe399f55.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 183, + 294, + 193 + ], + "lines": [ + { + "bbox": [ + 252, + 183, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 252, + 183, + 294, + 193 + ], + "type": "text", + "content": "Parallelism" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 121, + 378, + 174 + ], + "blocks": [ + { + "bbox": [ + 310, + 121, + 378, + 174 + ], + "lines": [ + { + "bbox": [ + 310, + 121, + 378, + 174 + ], + "spans": [ + { + "bbox": [ + 310, + 121, + 378, + 174 + ], + "type": "image", + "image_path": "827a7e3e5aea6c3fd98b4a49a721fb3992a6ada01dd99a0fc93d5c9ad1a2777a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 183, + 370, + 194 + ], + "lines": [ + { + "bbox": [ + 318, + 183, + 370, + 194 + ], + "spans": [ + { + "bbox": [ + 318, + 183, + 370, + 194 + ], + "type": "text", + "content": "Perpendicular" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 385, + 122, + 460, + 171 + ], + "blocks": [ + { + "bbox": [ + 385, + 122, + 460, + 171 + ], + "lines": [ + { + "bbox": [ + 385, + 122, + 460, + 171 + ], + "spans": [ + { + "bbox": [ + 385, + 122, + 460, + 171 + ], + "type": "image", + "image_path": "b99d966c30efcbee61524067991e4180597ec7a55fca6056c5f3012b6dfeb21c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 183, + 457, + 194 + ], + "lines": [ + { + "bbox": [ + 392, + 183, + 457, + 194 + ], + "spans": [ + { + "bbox": [ + 392, + 183, + 457, + 194 + ], + "type": "text", + "content": "Equal Lane Width" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 154, + 209, + 304, + 255 + ], + "blocks": [ + { + "bbox": [ + 299, + 196, + 312, + 205 + ], + "lines": [ + { + "bbox": [ + 299, + 196, + 312, + 205 + ], + "spans": [ + { + "bbox": [ + 299, + 196, + 312, + 205 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 209, + 304, + 255 + ], + "lines": [ + { + "bbox": [ + 154, + 209, + 304, + 255 + ], + "spans": [ + { + "bbox": [ + 154, + 209, + 304, + 255 + ], + "type": "image", + "image_path": "48b1d99a7f95457c0eac6ee3ba5b37f8ec8fcee0e5b0da77fe9af0fcca134dce.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 262, + 266, + 272 + ], + "lines": [ + { + "bbox": [ + 192, + 262, + 266, + 272 + ], + "spans": [ + { + "bbox": [ + 192, + 262, + 266, + 272 + ], + "type": "text", + "content": "Single Map Instance" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 274, + 233, + 284 + ], + "lines": [ + { + "bbox": [ + 220, + 274, + 233, + 284 + ], + "spans": [ + { + "bbox": [ + 220, + 274, + 233, + 284 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 205, + 378, + 258 + ], + "blocks": [ + { + "bbox": [ + 319, + 205, + 378, + 258 + ], + "lines": [ + { + "bbox": [ + 319, + 205, + 378, + 258 + ], + "spans": [ + { + "bbox": [ + 319, + 205, + 378, + 258 + ], + "type": "image", + "image_path": "b4e59fee30e6ad9c38c4c1f05a4b71b84719fb92349c0d4454851400ba57daad.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 262, + 423, + 272 + ], + "lines": [ + { + "bbox": [ + 347, + 262, + 423, + 272 + ], + "spans": [ + { + "bbox": [ + 347, + 262, + 423, + 272 + ], + "type": "text", + "content": "Cross Map Instances" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 274, + 391, + 284 + ], + "lines": [ + { + "bbox": [ + 378, + 274, + 391, + 284 + ], + "spans": [ + { + "bbox": [ + 378, + 274, + 391, + 284 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 293, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 326 + ], + "type": "text", + "content": "Fig. 2: Geometric properties and G-Representation. (a) Geometry in the transportation road system. (b) (c) We propose to model the geometric properties of a single map instance and multiple instances, with magnitude " + }, + { + "bbox": [ + 131, + 293, + 482, + 326 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 293, + 482, + 326 + ], + "type": "text", + "content": " and angle " + }, + { + "bbox": [ + 131, + 293, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 293, + 482, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 391, + 210, + 452, + 255 + ], + "blocks": [ + { + "bbox": [ + 391, + 210, + 452, + 255 + ], + "lines": [ + { + "bbox": [ + 391, + 210, + 452, + 255 + ], + "spans": [ + { + "bbox": [ + 391, + 210, + 452, + 255 + ], + "type": "image", + "image_path": "e21ada3d7040117a4a781a088eff69fc8da64d4715a51785f479701d03750b9d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 352, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 435 + ], + "type": "text", + "content": "dinates of polyline points to effectively represent the relative positions and orientations of adjacent points. To quantify these features within Euclidean space, we employ both the magnitudes of the displacement vectors and the angles formed between these vectors. Based on G-Representation, we propose the Euclidean Shape Clues to represent the shapes of individual map instances and the Euclidean Relation Clues to represent the inter-instance relations in Euclidean space." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "content": "G-Representation is a simple yet effective method to address the abovementioned two limitations: 1) Geometric Invariance, by concentrating on the relative relationships between points within and across map instances, it inherently attains translation and rotation invariance as illustrated in Figure 1b. This enhances its robustness against variations in data collection and equips it to effectively handle different coordinate systems. 2) Euclidean Modeling, it captures the inherent geometry in transportation road systems. Diverse shapes and relation geometry can be simplified as enumerations of magnitudes and directions. For instance, in Euclidean space, parallel lanes can be easily modeled as approximate directions " + }, + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "inline_equation", + "content": "(\\alpha \\approx 0)" + }, + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Building upon G-Representation, we propose a framework named GeMap for HD map construction. A BEV encoder is used to extract features from multiview input images, while a geometry-decoupled decoder is employed to focus on geometric aspects. Specifically, we adapt the attention mechanism [42] and propose Geometry-Decoupled Attention (GDA). GDA sequentially applies attention to queries belonging to the same instance and attention to queries across different instances. This can significantly boost the geometry learning of key points of shapes and relations between map instances. Furthermore, we propose an objective function named Euclidean Loss to optimize G-Representation." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Specifically, we transform the conventional polylines of the ground truth map into our G-Representation. In this way, the model gets optimized to better understand the magnitude of displacement vectors and the angles between them, thereby facilitating a more effective learning of geometric properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 485, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 485, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 485, + 235 + ], + "type": "text", + "content": "Experiments on nuScenes and Argoverse 2 datasets demonstrate the effectiveness of GeMap. We reach new state-of-the-art performances on both datasets. With camera images only, GeMap achieves " + }, + { + "bbox": [ + 130, + 164, + 485, + 235 + ], + "type": "inline_equation", + "content": "69.4\\%" + }, + { + "bbox": [ + 130, + 164, + 485, + 235 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 164, + 485, + 235 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 130, + 164, + 485, + 235 + ], + "type": "text", + "content": " mAP on the nuScenes and Argovserse 2, respectively. Visualization results (§ 4.4) further demonstrate the better perception of shape and relation, alleviation of occlusion, and robustness to rigid transformations of GeMap." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 236, + 373, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 236, + 373, + 248 + ], + "spans": [ + { + "bbox": [ + 146, + 236, + 373, + 248 + ], + "type": "text", + "content": "Our contributions are summarized as the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 258, + 479, + 365 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 258, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 479, + 293 + ], + "type": "text", + "content": "- We propose G-Representation which harnesses critical geometric properties of rotation and translation invariance in autonomous driving scenarios, opening up new research avenues within the field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 295, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 479, + 329 + ], + "type": "text", + "content": "- We introduce GeMap, a novel framework incorporating geometry-decoupled attention and Euclidean Loss function, specifically designed to learn the intrinsic geometry of online HD maps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 331, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 331, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 137, + 331, + 479, + 365 + ], + "type": "text", + "content": "- GeMap achieves new state-of-the-art results in HD map construction on the nuScenes and Argoverse 2 datasets, notably surpassing the " + }, + { + "bbox": [ + 137, + 331, + 479, + 365 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 137, + 331, + 479, + 365 + ], + "type": "text", + "content": " mAP on the large-scale Argoverse 2 dataset for the first time." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 237, + 399 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 311, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 311, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 311, + 426 + ], + "type": "text", + "content": "2.1 Online HD Map Construction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 434, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 578 + ], + "type": "text", + "content": "Traditionally, HD map construction has required labor-intensive manual or semi-automatic annotations [14,33]. To streamline this, recent studies [20,31,34,35,54] have focused on online construction, approaching HD maps as a dense prediction challenge. Innovations such as MetaBEV [9] aim to mitigate sensor issues, while MVNet [45] uses historical data for improved semantic consistency. The trend towards automatic vectorization of HD maps is spearheaded by works such as HDMapNet [18], which fuses camera and LiDAR inputs in BEV space and is advanced by end-to-end solutions such as VectorMapNet [26] and MapTR [22], leveraging Transformer-based models. StreamMapNet [48] and PivotNet [4] build upon this, modifying attention mechanisms for better performance. Our contribution starts from the strengths of the Transformer architecture, optimizing it with a decoupled self-attention block for enhanced geometric processing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 597, + 292, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 292, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 292, + 609 + ], + "type": "text", + "content": "2.2 Cross-view BEV Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "The conversion of Perspective View (PV) camera images into a unified BEV space is a significant challenge for autonomous driving systems. Previous studies such as [19,35] use depth estimation from monocular images for this transformation, others [2, 20, 28] have developed PV-to-BEV conversion methods without" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "explicit depths. However, these methods can distort shapes and geometric relations, leading to inaccuracies. Our approach integrates geometric supervision in prediction, addressing these geometric inconsistencies while acknowledging the value of depth information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 311, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 311, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 311, + 194 + ], + "type": "text", + "content": "2.3 Geometric Instance Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 202, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 334 + ], + "type": "text", + "content": "Traditional vectorization approaches model these instances as polylines [22, 26, 38, 43] or polynomial curves [7, 36, 39, 41], with some incorporating Bézier Curves for improved fitting [7, 36]. However, such methods often overlook geometric properties including shapes, parallelism, perpendicular, and etc. In 3D Lane Detection, some works [13, 17, 25] also explore simple geometry priors of lanes such as \"equal lane width\". However, they are not generalizable to more complex HD map instances. Existing works on modeling geometry properties of map instances mainly address rasterization-based shape loss [41, 49] or edge loss [22], and they fail to preserve cross-instance relationships. PivotNet [4] offers an architectural advancement with its line-aware point decoder, yet it still does not fully address the complexity of instance geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 481, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 481, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 481, + 382 + ], + "type": "text", + "content": "Given the current under-exploration of diverse geometric properties in shape and relation, we propose utilizing Euclidean shape and relation losses and decoupling the traditional self-attention module to empower the model with a more robust understanding of instance geometry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 401, + 202, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 202, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 202, + 414 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 426, + 221, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 221, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 221, + 439 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": "The architecture of our method is illustrated in Figure 3. Input images are denoted by " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbb{I} = \\{\\pmb{I}_i\\}_{i=1}^{N_c}" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": " is the number of cameras. The output is a set of " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": " map instances " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathbb{M} = \\{\\pmb{L}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": ", where each map instance is represented by a polyline " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{L}_i \\in \\mathbb{R}^{N_v \\times 2}" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": ", i.e., an ordered sequence of " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": " two-dimensional points. As shown in Figure 3, the commonly adopted pipeline comprises three steps. 1) A BEV feature extractor processes the multi-view images. 2) The extracted BEV features are fed into a map decoder, which predicts the map instances based on the BEV features. 3) To optimize the pipeline, the predicted map instances are compared against the ground truth. Beyond point-to-point comparison, G-representations of predictions and the ground truth are computed and the difference is measured by " + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "inline_equation", + "content": "L^1" + }, + { + "bbox": [ + 130, + 447, + 482, + 579 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 597, + 276, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 276, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 276, + 609 + ], + "type": "text", + "content": "3.2 Architecture Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 485, + 666 + ], + "type": "text", + "content": "BEV Feature Extractor. A shared vision backbone in different views and parameterized PV-to-BEV transformation network are employed to aggregate features from various perspectives. In our default configuration, we utilize ResNet [12] as the PV backbone and GKT [2] as the PV-to-BEV transformation network." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 447, + 103 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 114, + 459, + 315 + ], + "blocks": [ + { + "bbox": [ + 156, + 114, + 459, + 315 + ], + "lines": [ + { + "bbox": [ + 156, + 114, + 459, + 315 + ], + "spans": [ + { + "bbox": [ + 156, + 114, + 459, + 315 + ], + "type": "image", + "image_path": "2157971095d7ce5ba6b9ed1002ad3c234a091c316a7ad66e8e84f682244c9602.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 380 + ], + "type": "text", + "content": "Fig. 3: Illustration of GeMap. First, PV images are transformed into BEV features, then a Geometry-Decoupled Decoder outputs the vectorized HD Map. In each block of the decoder, queries are first processed by Euclidean shape and relation attention, which focuses on geometric relevance. Finally, predictions are enhanced in G-Representations by shape and relation constraint." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 401, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 461 + ], + "type": "text", + "content": "Geometry-Decoupled Decoder. We adopt a Transformer decoder to predict polylines. The decoder obtains " + }, + { + "bbox": [ + 130, + 401, + 482, + 461 + ], + "type": "inline_equation", + "content": "N \\times N_v" + }, + { + "bbox": [ + 130, + 401, + 482, + 461 + ], + "type": "text", + "content": " queries that represent points and processes them via self-attention and aggregate BEV features via deformable cross-attention [56]. Finally, a prediction head is used to convert queries to polylines " + }, + { + "bbox": [ + 130, + 401, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\{\\hat{L}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 401, + 482, + 461 + ], + "type": "text", + "content": ". We utilize a multi-layer perceptron as the polyline prediction head." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 508 + ], + "type": "text", + "content": "Geometric Loss. For training, we convert both the predicted polylines and ground truth into the proposed G-Representation. Based on that, we let the model optimize the predicted magnitudes of displacement vectors and angles between displacement vectors to match the ground truth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 545 + ], + "type": "text", + "content": "In the following subsections, we first introduce Geometric Representation in § 3.3 since it is the core of our framework. Based on that, we introduce Geometric Loss in § 3.4 and Geometry-Decoupled Decoder in § 3.5, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 560, + 293, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 293, + 573 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 293, + 573 + ], + "type": "text", + "content": "3.3 Geometric Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 578, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 615 + ], + "type": "text", + "content": "Euclidean Shape Clues. We first introduce the representation of shapes of individual map instances. For each instance, we describe the local geometry with displacement vectors between neighboring points which are computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 215, + 621, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 621, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 215, + 621, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {u} ^ {i} = \\boldsymbol {L} _ {i, u + 1} - \\boldsymbol {L} _ {i, u} \\quad (u \\in \\{1, 2, \\dots , N _ {v} \\}), \\tag {1}", + "image_path": "501b9e0c7e1d6f59927ae1f4be03a910cf4e187e46468466dbaa3d243c0fe587.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where we define " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{L}_{i,N_v + 1} \\coloneqq \\pmb{L}_{i,1}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " to unify the geometric formulation of closed and open polylines." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 455, + 178 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 178 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 178 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 178 + ], + "type": "image", + "image_path": "fbcdb00e68bbf3e8240394691f1b3d7948f085f310d4230157ea632eb96eea42.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 186, + 480, + 220 + ], + "lines": [ + { + "bbox": [ + 131, + 186, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 480, + 220 + ], + "type": "text", + "content": "Fig. 4: Euclidean Shape Clues. Magnitudes of displacement vectors and angles between neighboring vectors indicate shape clues and are utilized to compute shape loss. The right part shows how to connect Euclidean Shape Clues to shape geometry." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "content": "These displacement vectors are sufficient to represent the shape of a map instance and such a representation is invariant to translation transformations. However, we would like to note that this representation is vulnerable to rotations, which might prevent the model from learning robust geometry. To solve this problem, we propose to represent the shape with magnitudes of displacement vectors and angles between consecutive displacement vectors, as illustrated in Figure 4. Specifically, the " + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "content": "-th instance is denoted by " + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "content": " angle values and " + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "content": " magnitude values. Let " + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "content": " be the index, the " + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "content": "-th angle value and the magnitude value of the " + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 240, + 480, + 348 + ], + "type": "text", + "content": "-th instance are computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 354, + 480, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 354, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 238, + 354, + 480, + 369 + ], + "type": "interline_equation", + "content": "\\alpha_ {u} ^ {i} = \\left\\langle \\boldsymbol {v} _ {u} ^ {i}, \\boldsymbol {v} _ {u + 1} ^ {i} \\right\\rangle , d _ {u} ^ {i} = \\left\\| \\boldsymbol {v} _ {u} ^ {i} \\right\\| _ {2}, \\tag {2}", + "image_path": "5c337639e52599eec225ad75282e5688fdd662d40289dedfb2fa21993c21b2ef.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 463, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 463, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 463, + 386 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 373, + 463, + 386 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 130, + 373, + 463, + 386 + ], + "type": "text", + "content": " denotes the angle between two vectors and similarly " + }, + { + "bbox": [ + 130, + 373, + 463, + 386 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{N_v + 1}^i \\coloneqq \\pmb{v}_1^i" + }, + { + "bbox": [ + 130, + 373, + 463, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 386, + 481, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 481 + ], + "type": "text", + "content": "The proposed representation inherently captures common geometric patterns, such as parallelism, right angles, and proper line width, by translating them into corresponding numerical patterns within this representation. For example, in Figure 4, a rectangle is characterized by one " + }, + { + "bbox": [ + 130, + 386, + 481, + 481 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 130, + 386, + 481, + 481 + ], + "type": "text", + "content": " angle and equal magnitudes of opposite displacement vectors; in addition, a straight line is represented by " + }, + { + "bbox": [ + 130, + 386, + 481, + 481 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 130, + 386, + 481, + 481 + ], + "type": "text", + "content": " angles. Beyond discussed regular cases, the Euclidean shape clues are capable of handling more complex shapes, as illustrated in scenarios (a) and (c) of Figure 6 and discussed in § 4.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "text", + "content": "Euclidean Relation Clues. Having highlighted the translation and rotation invariance of G-Representation and its advantages in representing the geometric shapes of individual map instances, we further introduce its ability to capture the relations between two map instances, e.g., parallelism and perpendicular, in Euclidean space. Specifically, given the vanilla representations and displacement vectors of the " + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "text", + "content": "-th map instance, we represent the relation between instance " + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 496, + 479, + 604 + ], + "type": "text", + "content": " with 1) angles between each pair of their respective displacement vectors, and 2) magnitudes of displacement vectors between each pair of points, as illustrated in Figure 5. Formally," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 217, + 609, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 609, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 217, + 609, + 480, + 624 + ], + "type": "interline_equation", + "content": "\\alpha_ {u, v} ^ {i, j} = \\left\\langle \\boldsymbol {v} _ {u} ^ {i}, \\boldsymbol {v} _ {v} ^ {j} \\right\\rangle , d _ {u, v} ^ {i, j} = \\left\\| \\boldsymbol {L} _ {i, u} - \\boldsymbol {L} _ {j, v} \\right\\| _ {2}. \\tag {3}", + "image_path": "9b3a5aac467eab1ff55ea9601850e97d39878e5b8d7c807a3c58d06a54c7e552.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "This representation is also translation- and rotation-invariant. Similar to Euclidean Shape Clues, it inherently captures common relations between map instances by translating them into corresponding numerical patterns. For example," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 455, + 196 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 455, + 196 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 455, + 196 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 455, + 196 + ], + "type": "image", + "image_path": "9982cddd7e67134a42fafcf015de569f554aea3ee509115cea600b499053cfa1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "type": "text", + "content": "Fig. 5: Euclidean Relation Clues. Angles between pairs of displacement vectors on different polylines, and magnitudes of displacement vectors between point pairs indicate relation clues." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "content": "Figure 5 shows that the perpendicular relation can be directly represented by a " + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "content": " angle, and the distance between two parallel polylines, which may correspond to the width of a lane, is naturally represented by magnitudes of displacement vector between point pairs. Scenarios (a) and (b) of Figure 6 show the effectiveness of Euclidean Relation Clues and more discussion can be viewed in " + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\S 4.4" + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 335, + 313, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 313, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 313, + 348 + ], + "type": "text", + "content": "3.4 Euclidean Loss and Objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "content": "We transform both the ground truth data and the model's predicted polylines from their original format into G-Representation. This conversion allows the model to independently optimize the angles between displacement vectors and the magnitudes of displacement vectors to more accurately align with the ground truth. The proposed Euclidean Loss is composed of two parts that measure how accurately the model predicts the shape of individual map instances and the inter-instance relations, respectively, which are denoted by " + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shp}}" + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "content": ". The Euclidean Loss can be computed as the following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 245, + 458, + 481, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 458, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 245, + 458, + 481, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E u c}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {s h p}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {r e l}}, \\tag {4}", + "image_path": "346b6320596b2147790c3d60576cc1237561b6f6abfff03048058431ba31e734.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "text", + "content": " are hyperparameters, whose effects on the model performances are evaluated in Figure 7a. Specifically," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 148, + 510, + 481, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 510, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 148, + 510, + 481, + 581 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} w _ {i, j} = 1 - \\left(\\min _ {u, v} \\left\\{d _ {u, v} ^ {i, j} \\right\\} / \\sqrt {2}\\right) ^ {p}, \\quad \\mathcal {L} _ {\\mathrm {s h p}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N _ {v}} \\left| \\hat {d} _ {j} ^ {i} - d _ {j} ^ {i} \\right| + \\ell \\left(\\hat {\\alpha} _ {j} ^ {i}, \\alpha_ {j} ^ {i}\\right), \\tag {5} \\\\ \\mathcal {L} _ {\\mathrm {r e l}} = \\sum_ {i = 1} ^ {N} \\sum_ {j > i} ^ {N} w _ {i, j} \\sum_ {u = 1} ^ {N _ {v}} \\sum_ {v = 1} ^ {N _ {v}} \\left(| \\hat {d} _ {u, v} ^ {i, j} - d _ {u, v} ^ {i, j} | + \\ell (\\hat {\\alpha} _ {u, v} ^ {i, j}, \\alpha_ {u, v} ^ {i, j})\\right), \\\\ \\end{array}", + "image_path": "896a7cbcd91f450b8ffa43f3281553c6bc6f90627407d7ec683f34726c96f3de.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 588, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 479, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 588, + 479, + 613 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot)" + }, + { + "bbox": [ + 130, + 588, + 479, + 613 + ], + "type": "text", + "content": " is a function based on " + }, + { + "bbox": [ + 130, + 588, + 479, + 613 + ], + "type": "inline_equation", + "content": "L^1" + }, + { + "bbox": [ + 130, + 588, + 479, + 613 + ], + "type": "text", + "content": " loss. We avoid using the inverse trigonometric function to directly compute the angles, but use sine and cosine values instead:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 201, + 620, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 620, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 201, + 620, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\ell (\\hat {\\alpha}, \\alpha) = | \\cos (\\hat {\\alpha}) - \\cos (\\alpha) | + | \\sin (\\hat {\\alpha}) - \\sin (\\alpha) |. \\tag {6}", + "image_path": "b21348205482118db528e5aa8c196ba3967fa4277068b1e38a39fec7895ea1b4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Moreover, the distance might influence the relation strength of instance pairs. For example, if two instances are far from each other, the relation between them" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "might be weak. For this reason, we further adopt " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "w_{i,j}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " to punish weakly related instance pairs and discuss it in § 4.5. According to the best experimental results, we treat all instance pairs equally in other experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": "Following common practices [22,48], we also use focal loss [24] " + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": " for classification. For polyline regression, we adopt point-to-point loss and edge direction loss, denoted by " + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pts}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": " respectively. Following [23,36], we further adopt segmentation loss " + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": " and depth estimation loss " + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dep}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": ". These losses are detailed in appendix § A.1. The overall loss function can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 217, + 480, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 217, + 480, + 230 + ], + "spans": [ + { + "bbox": [ + 149, + 217, + 480, + 230 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\cdot \\mathcal {L} _ {\\text {E u c}} + \\beta_ {1} \\cdot \\mathcal {L} _ {\\text {c l s}} + \\beta_ {2} \\cdot \\mathcal {L} _ {\\text {p t s}} + \\beta_ {3} \\cdot \\mathcal {L} _ {\\text {d i r}} + \\beta_ {4} \\cdot \\mathcal {L} _ {\\text {s e g}} + \\beta_ {5} \\cdot \\mathcal {L} _ {\\text {d e p}}. \\tag {7}", + "image_path": "2084b5f356052ef039caa7f48c80caa5cc50fa42b8769f8698c0b461e64f9dc8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 235, + 460, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 235, + 460, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 460, + 247 + ], + "type": "text", + "content": "where hyperparameters are discussed in Figure 7b and appendix Table A1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 262, + 313, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 313, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 313, + 274 + ], + "type": "text", + "content": "3.5 Geometry-Decoupled Decoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "content": "In the Transformer decoder, we obtain the " + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "inline_equation", + "content": "N \\times N_v" + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "content": " fused queries by adding " + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "content": " instance queries to " + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "content": " point queries, which is suggested in [22]. We denote input tokens by " + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "inline_equation", + "content": "\\pmb{E} \\in \\mathbb{R}^{N_A \\times D}" + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "content": " is the feature dimension and " + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "inline_equation", + "content": "N_A = N \\cdot N_v" + }, + { + "bbox": [ + 130, + 277, + 479, + 337 + ], + "type": "text", + "content": " is the number of fused queries. Intuitively, each fused query corresponds to a point on a predicted map instance. Self-Attention (SA) is formulated by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 184, + 342, + 480, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 342, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 184, + 342, + 480, + 369 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} (\\boldsymbol {E}, \\boldsymbol {M}) = \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\boldsymbol {E} \\boldsymbol {W} ^ {q}\\right) \\left(\\boldsymbol {E} \\boldsymbol {W} ^ {k}\\right) ^ {\\top}}{\\sqrt {D _ {k}}} \\odot \\boldsymbol {M}\\right) \\boldsymbol {E} \\boldsymbol {W} ^ {v}, \\tag {8}", + "image_path": "a5aaea17a95dc12d7996378d6e66b5c464af44b1fb848480b13a16a144a59ce9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{W}^q, \\pmb{W}^k, \\pmb{W}^v \\in \\mathbb{R}^{D \\times D_k}" + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "text", + "content": " are linear projection matrices, " + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{M} \\in \\mathbb{R}^{N_A \\times N_A}" + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "text", + "content": " is the attention mask, and " + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "text", + "content": " is the Hadamard product. The vanilla SA computes relations between every pair of tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 411, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 493 + ], + "type": "text", + "content": "Nevertheless, the geometry of shape and relation pertains to distinct subsets of tokens. For any given map instance, its shape is intimately related to tokens representing that instance's points. Precise shape geometry capture requires the model to discern token correlations specific to an instance while avoiding interference from tokens of unrelated instances. Conversely, for relation geometry modeling, it is beneficial to isolate token correlations that span across different instances, rather than those confined within a single instance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": "Therefore, we present Geometry-Decoupled Attention (GDA). First, we multiply a binary mask " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": " to the computed attention map so that the tokens of a map instance are aggregated according to the tokens within the same map instances only, which allows the model to adjust points' positions according to learned shape geometry. Denote the index of map instance that the " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": "-th token belongs to as " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": ". For example, tokens " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "1,2,\\ldots,N_v" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": " belong to the first instance so that " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1,\\mathcal{I}_2,\\ldots,\\mathcal{I}_{N_v}=1" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": ". Tokens " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "N_v+1,N_v+2,\\ldots,2\\cdot N_v" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": " belong to the second instance so that " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{N_v+1},\\mathcal{I}_{N_v+2},\\ldots,\\mathcal{I}_{2N_v}=2" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": ". With this notation, the binary mask " + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": " can be simply constructed by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 201, + 607, + 480, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 607, + 480, + 638 + ], + "spans": [ + { + "bbox": [ + 201, + 607, + 480, + 638 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {i, j} ^ {\\mathrm {s h p}} = \\left\\{ \\begin{array}{l l} 1 & , \\mathcal {I} _ {i} = \\mathcal {I} _ {j} \\\\ 0 & , \\mathcal {I} _ {i} \\neq \\mathcal {I} _ {j} \\end{array} , \\quad \\boldsymbol {M} _ {i, j} ^ {\\mathrm {r e l}} = \\left\\{ \\begin{array}{l l} 1 & , \\mathcal {I} _ {i} \\neq \\mathcal {I} _ {j} \\\\ 0 & , \\mathcal {I} _ {i} = \\mathcal {I} _ {j} \\end{array} . \\right. \\right. \\tag {9}", + "image_path": "6ef43dc03eb975ef381fbaf084da908f2970c56ffc8b61ec647c6dd0ce530656.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "The second attention is expected to model relations between tokens of different map instances." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 202, + 468, + 396 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 1: Comparison on the nuScenes dataset, GeMap reaches a new state-of-the-art performance. \"EB0\", \"R50\", \"PP\", \"Sec\", \"Swin-T\", and \"V2-99\" denote EfficientNet-B0 [40], ResNet50 [12], PointPillars [15], SECOND [46], Swin Transformer Tiny [27], and VoVNetV2-99 [16] respectively. Methods with two backbones utilize both camera and LiDAR inputs. \"Dns. Loss\" denotes whether any dense prediction (e.g. semantic segmentation) loss is adopted. The best result is highlighted in **bold**. We reproduce all methods on a single RTX3090 GPU to test FPS for fair comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 202, + 468, + 396 + ], + "lines": [ + { + "bbox": [ + 143, + 202, + 468, + 396 + ], + "spans": [ + { + "bbox": [ + 143, + 202, + 468, + 396 + ], + "type": "table", + "html": "
MethodsBackboneDns. Loss\\( AP_{div}(\\uparrow) \\)\\( AP_{ped}(\\uparrow) \\)\\( AP_{bnd}(\\uparrow) \\)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5042.551.444.146.05.3
MapTR [ICLR'23] [22]R5059.856.260.158.719.8
MapVR [NeurIPS'23] [49]R5061.855.059.458.819.8
GeMap [Ours]R5065.1 ↑ 3.359.8 ↑ 4.863.2 ↑ 3.862.7 ↑ 3.919.0
HDMapNet [ICRA'22] [18]EB0✓14.421.733.023.00.7
PivotNet [ICCV'23] [4]R50✓58.853.859.657.49.5
BeMapNet [CVPR'23] [36]R50✓66.762.665.164.86.6
MapTRv2 [Arxiv'23] [23]R50✓68.368.169.768.715.0
GeMap [Ours]R50✓69.8 ↑ 1.567.171.4 ↑ 1.769.4 ↑ 0.715.8
HDMapNet [ICRA'22] [18]EB0 & PP29.616.346.731.00.7
VectorMapNet [ICML'23] [26]R50 & PP60.148.253.053.7-
MapTR [ICLR'23] [22]R50 & Sec62.355.969.362.56.7
MapVR [NeurIPS'23] [49]R50 & PP62.760.467.263.5-
GeMap [Ours]R50 & Sec66.3 ↑ 3.662.2 ↑ 1.871.1 ↑ 3.966.5 ↑ 3.07.3
MapTRv2 [Arxiv'23] [23]R50 & Sec✓65.666.574.869.06.6
GeMap [Ours]R50 & Sec✓69.8 ↑ 4.268.0 ↑ 1.573.470.4 ↑ 1.46.8
MapTRv2 [Arxiv'23] [23]V2-99✓73.771.475.073.410.1
GeMap [Ours]Swin-T✓72.870.472.872.011.4
GeMap [Ours]V2-99✓76.0 ↑ 2.374.3 ↑ 2.977.7 ↑ 2.776.0 ↑ 2.610.8
", + "image_path": "5ca897c04608f8b761da6ea79714f9be443a6b0cd3ed96cef312d96f10be2d84.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 422, + 230, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 230, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 230, + 435 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 455, + 266, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 266, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 266, + 468 + ], + "type": "text", + "content": "4.1 Experimental Setups" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 484, + 495, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 495, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 495, + 616 + ], + "type": "text", + "content": "Datasets. To evaluate GeMap, we conduct experiments on the nuScenes dataset [1], a widely adopted large-scale autonomous driving dataset that includes 1,000 scenes captured by six RGB cameras with a 360-degree field of view, and provides precise annotations from LiDAR point clouds for HD map construction. For comparability with prior research [18,22], we focus on three static categories of map instances: pedestrian crossings, lane dividers, and road boundaries. With dataset splits provided by BeMapNet [36], we also evaluate GeMaps under three weather conditions: sunny, cloudy, and rainy. Additionally, we use the Argoverse 2 dataset [44] as another benchmark, which consists of approximately 108,000 frames, each providing images from seven cameras. We evaluate the performance by Average Precision (AP) [18,22] with thresholds at 0.5, 1.0, and 1.5 meters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. GeMap leverages 8 NVIDIA RTX 3090 GPUs for training. We adopt AdamW [30] as the optimizer and utilize Cosine Annealing with a linear warm-up phase [29] as the learning rate scheduler. For more details on hyperparameters, please refer to appendix § A.2." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 146, + 467, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Following BeMapNet [36], we also compare GeMap on the nuScenes dataset under different weather conditions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 146, + 467, + 213 + ], + "lines": [ + { + "bbox": [ + 143, + 146, + 467, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 467, + 213 + ], + "type": "table", + "html": "
MethodsBackbone Dns. LossmAPsun(↑)mAPcld(↑)mAPrny(↑)mAPavg(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5043.844.136.641.55.3
MapTR [ICLR'23] [22]R5062.160.552.858.419.8
GeMap [Ours]R5066.0 ↑ 3.964.3 ↑ 3.854.4 ↑ 1.661.5 ↑ 3.119.0
BeMapNet [CVPR'23] [36]R50 √67.367.556.663.86.6
GeMap [Ours]R50 √73.1 ↑ 5.871.0 ↑ 3.559.3 ↑ 2.767.8 ↑ 4.015.8
", + "image_path": "5842196eb144cfc75ac3bae0387b5843ab6aa6722e0dd93ef05f43a2bcf9c6ab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 143, + 262, + 467, + 327 + ], + "blocks": [ + { + "bbox": [ + 132, + 231, + 479, + 253 + ], + "lines": [ + { + "bbox": [ + 132, + 231, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 479, + 253 + ], + "type": "text", + "content": "Table 3: Comparison on Argoverse 2 dataset. GeMap demonstrates significant performance improvements over previous methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 262, + 467, + 327 + ], + "lines": [ + { + "bbox": [ + 143, + 262, + 467, + 327 + ], + "spans": [ + { + "bbox": [ + 143, + 262, + 467, + 327 + ], + "type": "table", + "html": "
MethodsBackbone Dns. Loss\\( AP_{div}(\\uparrow) \\)\\( AP_{ped}(\\uparrow) \\)\\( AP_{bnd}(\\uparrow) \\)mAP(↑)FPS(↑)
VectorMapNet [ICML'23] [26]R5036.138.339.237.9-
GeMap [Ours]R5067.6 ↑ 31.559.3 ↑ 21.064.7 ↑ 25.563.9 ↑ 26.016.7
HMapNet [ICRA'22] [18]EB0✓5.713.137.618.8
MapTRv2 [Arxiv'23] [23]R50✓72.162.967.167.4
GeMap [Ours]R50✓75.7 ↑ 3.669.2 ↑ 6.370.5 ↑ 3.471.8 ↑ 4.4
", + "image_path": "065e5ef768ad58cb92af48647f11fbf41f3598f61d8abeb56dcea91693faf77e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 357, + 228, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 228, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 228, + 369 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "text", + "content": "Results on nuScenes. As delineated in Table 1, GeMap delivers a state-of-the-art performance, achieving a mean Average Precision (mAP) of " + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "inline_equation", + "content": "69.4\\%" + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "text", + "content": " using the camera-only setup and ResNet50 [12] backbone. GeMap particularly improves the precision for identifying dividers and boundaries, with enhancements of " + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "text", + "content": ", respectively. Importantly, these performance gains do not come at a high cost of efficiency, as GeMap sustains inference speeds on par with and even exceeds the previously established state-of-the-art, measured in frames per second (FPS). With more powerful vision backbones, GeMap can achieve significantly improved performance. Specifically, VoVNetV2-99 [16] and Swin-T [27] outperform the ResNet50 baseline by " + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "inline_equation", + "content": "+6.6\\%" + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 130, + 397, + 481, + 589 + ], + "type": "text", + "content": " in mAP, respectively. Also, GeMap achieves significantly improved performance with extra LiDAR inputs, which indicates the generalizability of GeMap to multi-modality settings. Moreover, we also compare performance under different weather conditions in Table 2. GeMap exceeds previous works under sunny, cloudy, and rainy scenarios, which shows the potential of geometry to raise model robustness to varied weather conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Results on Argoverse 2. Extending our evaluation to the Argoverse 2 dataset, as presented in Table 3, GeMap also presents a new SOTA performance of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " mAP, outperforming the highly advanced MapTRv2 model by " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "+4.4\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". This achievement not only demonstrates the effectiveness of GeMap on a different dataset but also emphasizes the adaptability and precision of GeMap in the evolving landscape of autonomous driving technologies." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 146, + 467, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study on the nuScenes dataset. We train the model for 24 epochs. * denotes replacing GDA with 2 layers of vanilla self-attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 146, + 467, + 247 + ], + "lines": [ + { + "bbox": [ + 143, + 146, + 467, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 467, + 247 + ], + "type": "table", + "html": "
Method\\( {\\mathcal{L}}_{\\text{shp }}\\left( {\\S {3.3}}\\right) {\\mathcal{L}}_{\\text{rel }}\\left( {\\S {3.3}}\\right) \\)GDA (§ 3.5)\\( {\\mathrm{{AP}}}_{div}\\left( \\uparrow \\right) \\)\\( {\\mathrm{{AP}}}_{ped}\\left( \\uparrow \\right) \\)\\( {\\mathrm{{AP}}}_{bnd}\\left( \\uparrow \\right) \\)mAP(↑)
Baseline49.544.753.749.3
+ Decoupled Attention✓53.446.653.551.2
+ Single Euclidean Loss✓51.045.452.749.7
51.743.453.049.4
+ Euclidean Loss✓✓51.543.951.148.8
+ Single Euclidean Loss with Decoupled Attention✓✓54.048.253.151.8
✓✓54.747.355.352.4
Replace GDA with 2-SA✓✓\\( 2 - SA^* \\)53.546.254.451.4
Full✓✓✓53.6 ↑ 4.149.2 ↑ 4.554.8 ↑ 1.152.6 ↑ 3.3
", + "image_path": "4ddb0c04186a8894eeae1d61e2c0dfe3c390751b7b26757cb519b85faf7cc554.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 270, + 238, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 238, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 238, + 282 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "text", + "content": "Components Ablation. The ablation experiments, detailed in Table 4, involved training the model for a reduced duration of 24 epochs without dense prediction losses, to facilitate efficient analysis. These experiments confirm the significant role of GDA, which alone increases mAP by " + }, + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "inline_equation", + "content": "+1.9\\%" + }, + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "text", + "content": ". Moreover, applying the full suite of components results in a further mAP enhancement of " + }, + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "inline_equation", + "content": "+1.4\\%" + }, + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "text", + "content": ". A notable discovery is the detrimental effect on model performance when Euclidean Loss is applied without GDA, which leads to a " + }, + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 291, + 481, + 398 + ], + "type": "text", + "content": " reduction in mAP. This reinforces our position that conventional self-attention mechanisms are insufficient for encoding a variety of geometric properties." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 481, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 469 + ], + "type": "text", + "content": "Effectiveness of GDA. In the proposed GDA (3.5), we decouple shape and relation learning, sequentially applying Euclidean shape and relation attention. In addition, we also attempt to intuitively double the self-attention layers in one block, referred to as \"2-SA\". As shown in Table 4, GDA outperforms vanilla self-attention by " + }, + { + "bbox": [ + 130, + 399, + 481, + 469 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 469 + ], + "type": "text", + "content": " mAP, indicating the superiority of GDA when combined with Euclidean Loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 489, + 272, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 272, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 272, + 502 + ], + "type": "text", + "content": "4.4 Visualization Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 568 + ], + "type": "text", + "content": "More Precise Shape Awareness. As shown in Figure 6(a), the shape geometry allows for an inherent description of the triangle feature, in contrast to baselines which do not discern it as effectively. A more irregularly shaped pedestrian crossing can be viewed in scenario (c) of Figure 6 and GeMap provides more precise construction than other methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "type": "text", + "content": "Better Relation Awareness. With the help of Euclidean Relation Clues, GeMap can better infer parallelism and perpendicular in scenario (a) of Figure 6. More interestingly, in scenario (b), the strong sunlight makes the divider highlighted hard to recognize. However, according to the understanding of lane width patterns, GeMap has the sense that there should be a divider." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Alleviating Occlusion Issues. Figure 6(c) demonstrates that shape geometry alleviates occlusions in partially visible instances. It showcases a black car that obscures part of a pedestrian crossing, while, GeMap successfully deduces the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 351 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 351 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 351 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 351 + ], + "type": "image", + "image_path": "613ff0d6cc99954666acbab67f745803be1f04fea88dc1a5e6ad965b81431896.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 360, + 480, + 393 + ], + "lines": [ + { + "bbox": [ + 130, + 360, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 480, + 393 + ], + "type": "text", + "content": "Fig. 6: Visualization results. Hard predictions are highlighted, such as complex triangular road boundary (a), strong sunlight (b), partially observed (c), and obscured situations (d). These challenging cases indicate the superiority and robustness of GeMap." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "content": "overall structure of the crossing within a complex shape. In contrast, these baseline models [4, 22, 23] struggle with such complex shape recovery under similar occlusion conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 451, + 480, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 480, + 511 + ], + "type": "text", + "content": "Enhanced Robustness to Rotational Transformations. Figure 6(d) showcases that the BEV map is highly tilted by turning, which is extensively obscured by vehicular traffic. GeMap more adeptly maintains the integrity of lane width and parallelism, which are key geometry. This underscores the superior robustness of our geometric constructs against rotational distortions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 527, + 311, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 311, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 311, + 540 + ], + "type": "text", + "content": "4.5 Hyperparameter Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 593 + ], + "type": "text", + "content": "Comparative Impact of Euclidean Losses. Figure 7a shows that GeMap exhibits larger sensitivity to the shape loss than to the relation loss, as evidenced by the steeper change of the performance curve. The optimal result is obtained when the shape and relation losses are relatively balanced by weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": "Optimization of Euclidean Loss Weighting. The performance of the model, as demonstrated in Figure 7b, peaks when the weight of the Euclidean Loss, " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": ", is set to " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": ". Notably, model performance degrades with a too large or too small " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": ", which informed us to use a " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": " value of " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": " for our experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "Distance Weighting. Experimental results of distance weighting are presented in Table 5. We change the order of distance " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": " in Equation 5 and " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "p\\in \\{1,2,4\\}" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 448, + 102 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 306, + 226 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 306, + 226 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 306, + 226 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 306, + 226 + ], + "type": "image", + "image_path": "c8848dbd709d1278610f4228219e8e6f6d28ea9c757035d13827c202780ba3fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 228, + 232, + 236 + ], + "lines": [ + { + "bbox": [ + 220, + 228, + 232, + 236 + ], + "spans": [ + { + "bbox": [ + 220, + 228, + 232, + 236 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 465, + 226 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 465, + 226 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 465, + 226 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 465, + 226 + ], + "type": "image", + "image_path": "67515aafdedc2b0aa1b1799256f4835fcc81c6a2cbcb431c8b8a7f7ee992f880.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 228, + 391, + 236 + ], + "lines": [ + { + "bbox": [ + 378, + 228, + 391, + 236 + ], + "spans": [ + { + "bbox": [ + 378, + 228, + 391, + 236 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 245, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 268 + ], + "type": "text", + "content": "Fig. 7: Hyper-parameters. The best performance is reached with balanced Euclidean shape loss and relation loss." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 314 + ], + "type": "text", + "content": "corresponds to \"Linear\", \"Square\" and \"4th Power\" respectively. Moreover, we also try to treat instance pairs equally and it delivers the best performance." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 220, + 354, + 391, + 414 + ], + "blocks": [ + { + "bbox": [ + 195, + 333, + 416, + 344 + ], + "lines": [ + { + "bbox": [ + 195, + 333, + 416, + 344 + ], + "spans": [ + { + "bbox": [ + 195, + 333, + 416, + 344 + ], + "type": "text", + "content": "Table 5: Distance weighting on the nuScenes dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 220, + 354, + 391, + 414 + ], + "lines": [ + { + "bbox": [ + 220, + 354, + 391, + 414 + ], + "spans": [ + { + "bbox": [ + 220, + 354, + 391, + 414 + ], + "type": "table", + "html": "
StrategyAP_div(↑)AP_ped(↑)AP_bnd(↑)mAP(↑)
Linear53.546.656.252.1
Square53.247.955.252.1
4th Power53.247.455.852.1
Equal53.649.254.852.6
", + "image_path": "9b715b1839df8aee28575253092865ab2cf4702952ebfe2a1e3cb1fd1221c54f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 450, + 220, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 220, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 220, + 464 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "type": "text", + "content": "In this paper, we realize significant shape and relation geometry inherent in HD map instances and propose the GeMap. GeMap includes the integration of Euclidean shape and relation losses for auxiliary supervision. To further refine the model's awareness of diverse geometry, we introduce the Geometry-Decoupled Attention mechanism. GeMap has achieved state-of-the-art performances on both the nuScenes and Argoverse 2 datasets, underscoring its effectiveness. Despite these promising results, the current application of geometry remains fundamental, and future research could focus on more sophisticated representations or enhanced geometric patterns. Furthermore, the application of geometry extends beyond HD map construction, offering potential solutions to occlusion challenges in other autonomous driving tasks. We anticipate that these findings will inspire further research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62272045, No. 8326014), The Shun Hing Institute of Advanced Engineering (No. 8115074), and CUHK Direct Grants (No. 4055190)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "type": "text", + "content": "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "text", + "content": "2. Chen, S., Cheng, T., Wang, X., Meng, W., Zhang, Q., Liu, W.: Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "text", + "content": "3. Deo, N., Wolff, E., Beijbom, O.: Multimodal trajectory prediction conditioned on lane-graph traversals. In: Conference on Robot Learning. pp. 203-212. PMLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 481, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 287 + ], + "type": "text", + "content": "4. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3672-3682 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "text", + "content": "5. Ding, X., Zhang, Y., Ge, Y., Zhao, S., Song, L., Yue, X., Shan, Y.: Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5513-5524 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 376 + ], + "type": "text", + "content": "6. Espinoza, J.L.V., Liniger, A., Schwarting, W., Rus, D., Van Gool, L.: Deep interactive motion prediction and planning: Playing games with motion prediction models. In: Learning for Dynamics and Control Conference. pp. 1006-1019. PMLR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "text", + "content": "7. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking efficient lane detection via curve modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17062-17070 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 410, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 454 + ], + "type": "text", + "content": "8. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 455, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 481, + 498 + ], + "type": "text", + "content": "9. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for 3d detection and map segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8721-8731 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 542 + ], + "type": "text", + "content": "0. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 544, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 586 + ], + "type": "text", + "content": "1. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26584-26595 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 620 + ], + "type": "text", + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "3. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "14. Jiao, J.: Machine learning assisted high-definition map creation. In: 2018 IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC). vol. 1, pp. 367-373. IEEE (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "15. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "text", + "content": "16. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "type": "text", + "content": "17. Li, C., Shi, J., Wang, Y., Cheng, G.: Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4370-4379 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "text", + "content": "18. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: An online hd map construction and evaluation framework. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 4628-4634. IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 321 + ], + "type": "text", + "content": "19. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1477-1485 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 321, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 321, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 321, + 480, + 365 + ], + "type": "text", + "content": "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 480, + 407 + ], + "type": "text", + "content": "21. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 541-556. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 407, + 480, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 451 + ], + "type": "text", + "content": "22. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: MapTR: Structured modeling and learning for online vectorized HD map construction. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=k7p_YA07yE" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 483 + ], + "type": "text", + "content": "23. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 480, + 515 + ], + "type": "text", + "content": "24. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 515, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 480, + 547 + ], + "type": "text", + "content": "25. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1765-1772 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 547, + 480, + 579 + ], + "type": "text", + "content": "26. Liu, Y., Yuan, T., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: International Conference on Machine Learning. pp. 22352-22369. PMLR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 579, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 480, + 622 + ], + "type": "text", + "content": "27. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "text", + "content": "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 205 + ], + "type": "text", + "content": "31. Loukkal, A., Grandvalet, Y., Drummond, T., Li, Y.: Driving among flatmobiles: Bird-eye-view occupancy grids from a monocular camera for holistic trajectory planning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 51-60 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "32. Lu, C., van de Molengraft, M.J.G., Dubbelman, G.: Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters 4(2), 445-452 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 281 + ], + "type": "text", + "content": "33. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4227-4236 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "34. Pan, B., Sun, J., Leung, H.Y.T., Andonian, A., Zhou, B.: Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters 5(3), 4867-4873 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "text", + "content": "35. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "36. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13218-13228 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "37. Scheel, O., Bergamini, L., Wolczyk, M., Osiński, B., Ondruska, P.: Urban driver: Learning to drive from real-world demonstrations using policy gradients. In: Conference on Robot Learning. pp. 718-728. PMLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 425, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 468 + ], + "type": "text", + "content": "38. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Keep your eyes on the lane: Real-time attention-guided lane detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 294-302 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "type": "text", + "content": "39. Tabelini, L., Berriel, R., Paixao, T.M., Badue, C., De Souza, A.F., Oliveira-Santos, T.: Polylanenet: Lane estimation via deep polynomial regression. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6150-6156. IEEE (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "40. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 589 + ], + "type": "text", + "content": "41. Van Gansbeke, W., De Brabandere, B., Neven, D., Proesmans, M., Van Gool, L.: End-to-end lane detection through differentiable least-squares fitting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "43. Wang, J., Ma, Y., Huang, S., Hui, T., Wang, F., Qian, C., Zhang, T.: A keypoint-based global association network for lane detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1392-1401 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Vectorized HD Map Construction using Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 533 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "45. Xie, Z., Pang, Z., Wang, Y.X.: Mv-map: Offboard hd-map generation with multiview consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8658-8668 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "type": "text", + "content": "47. Yang, W., Li, Q., Liu, W., Yu, Y., Ma, Y., He, S., Pan, J.: Projecting your view attentively: Monocular road scene layout estimation via cross-view transformation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15536-15545 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "48. Yuan, T., Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Streammapnet: Streaming mapping network for vectorized online hd map construction. arXiv preprint arXiv:2308.12570 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "49. Zhang, G., Lin, J., Wu, S., Luo, Z., Xue, Y., Lu, S., Wang, Z., et al.: Online map vectorization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "50. Zhang, Y., Ding, X., Gong, K., Ge, Y., Shan, Y., Yue, X.: Multimodal pathway: Improve transformers with irrelevant data from other modalities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6108-6117 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 391 + ], + "type": "text", + "content": "51. Zhang, Y., Gong, K., Zhang, K., Li, H., Qiao, Y., Ouyang, W., Yue, X.: Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "52. Zhang, Y., Li, H., Liu, J., Yue, X.: Explore the limits of omni-modal pretraining at scale. arXiv preprint arXiv:2406.09412 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "53. Zhang, Z., Zhang, Y., Ding, X., Jin, F., Yue, X.: Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "54. Zhou, B., Krahenbuhl, P.: Cross-view transformers for real-time map-view semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13760-13769 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "55. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "56. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Z. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_content_list.json b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44e0521c5c31afea1db80abaa7c7bebd3d1e7929 --- /dev/null +++ b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_content_list.json @@ -0,0 +1,1710 @@ +[ + { + "type": "text", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "text_level": 1, + "bbox": [ + 269, + 140, + 733, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zefan $\\mathrm{Qu}^{1}$ , Xinyang Jiang $^{\\dagger 2}$ , Yifan Yang $^{2}$ , Dongsheng Li $^{2}$ , and Cairong Zhao $^{\\dagger 1}$", + "bbox": [ + 225, + 212, + 776, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tongji University, Shanghai, China", + "bbox": [ + 375, + 255, + 627, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qzf, zhaocairong}@tongji.edu.cn", + "bbox": [ + 375, + 270, + 625, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Microsoft Research Asia, Shanghai, China", + "bbox": [ + 352, + 282, + 648, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xinyangjiang,yifanyang,dongsheng.li}@microsoft.com", + "bbox": [ + 303, + 297, + 699, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Low latency rates are crucial for online video-based applications, such as video conferencing and cloud gaming, which make improving video quality in online scenarios increasingly important. However, existing quality enhancement methods are limited by slow inference speed and the requirement for temporal information contained in future frames, making it challenging to deploy them directly in online tasks. In this paper, we propose a novel method, STLVQE, specifically designed to address the rarely studied online video quality enhancement (Online-VQE) problem. Our STLVQE designs a new VQE framework which contains a Module-Agnostic Feature Extractor that greatly reduces the redundant computations and redesign the propagation, alignment, and enhancement module of the network. A Spatial-Temporal Look-up Tables (STL) is proposed, which extracts spatial-temporal information in videos while saving substantial inference time. To the best of our knowledge, we are the first to exploit the LUT structure to extract temporal information in video tasks. Extensive experiments on the MFQE 2.0 dataset demonstrate our STLVQE achieves a satisfactory performance-speed trade-off.", + "bbox": [ + 259, + 354, + 743, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Online Video Quality Enhancement $\\cdot$ Look-up Table $\\cdot$ Video Compression Artifact Reduction", + "bbox": [ + 259, + 604, + 740, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 686, + 375, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of social media, users are increasingly seeking high-quality and real-time video content. In recent times, video-based applications such as video conferencing, webcasting, and cloud gaming have increasingly demanded low latency requirements. This has resulted in higher video compression rates, which can cause a noticeable decrease in video quality and an overall reduction in the Quality of Experience (QoE). As a result, online Video Quality", + "bbox": [ + 212, + 722, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 230, + 825, + 397, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46cbabc8dca024c6e69d705a1453a69b132be3d152381f10cf3e3a5a40511c68.jpg", + "image_caption": [ + "Fig. 1: Comparison of $\\Delta$ PSNR/ $\\Delta$ SSIM, average runtime per frame(720P video) and parameters on MFQE 2.0 test set at $\\mathrm{QP} = 37$ . Our STLVQE method achieves a great trade-off in enhancement performance and inference speed." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 504, + 310 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d95bcd21a4980bdc6cf6629ea7404945634c0bd848d02a5f936420e19366356.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 784, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Enhancement (Online-VQE), which improves the quality of real-time streaming video under online conditions and minimizing video defects caused by compression algorithms (e.g., H.264/AVC [27] and H.265/HEVC [24]) has become a significant problem. However, Online-VQE poses two main challenges compared to the conventional offline VQE approach. Firstly, it needs to process high-resolution videos in real-time to ensure a smooth viewing experience. Secondly, online video processing techniques [15, 18, 22] can only rely on the current and previous frames for inference due to the uncontrolled latency caused by waiting for future frames, which can result in overall video delay.", + "bbox": [ + 212, + 397, + 787, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, numerous neural network-based techniques for enhancing the quality of compressed video have been developed, resulting in significant advancements in this field. These methods can be broadly classified into two categories, namely single-image based VQE method [4, 12, 29, 30] and multi-frame based VQE method [2, 6, 10, 20, 36]. Single-image based approaches are relatively faster in inference speed, but they ignore the significance of temporal information in video [31]. On the other hand, multi-frame based VQE achieves state-of-the-art performance by utilizing temporal information. However, most of these approaches are not suitable for online scenarios because they rely on subsequent frames as reference frames and have unsatisfactory inference speeds. Therefore, developing a new online-VQE method that satisfies the aforementioned two constraints while effectively leveraging temporal information is essential.", + "bbox": [ + 212, + 536, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reasons that limit the speed and latency of the above temporal VQE methods can be summarized in the following two points: First, a high level of computational redundancy exists within these methods, and these homogeneous computations increase the network's inference time. Second, most of these methods employ complex and computationally intensive models, which sacrificing inference speed and latency for some of the enhancement capabilities. In this work, a more applicable VQE method for online scenarios are proposed by advancing these two perspectives.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As discussed in previous works [1], modern VQE methods exploit temporal information by constructing a neural network containing three main components: the propagation, alignment and enhancement module. In all three modules, similar feature extraction operations on the original image may be involved, e.g., the reference frame selection [10, 31] in propagation module, the optical flow calculation [1, 2] in alignment module and the final result obtaining in the enhancement module [6, 10]. Most of the state-of-the-art methods adopts distinct feature extractors in different components and obtaining latent feature for each component separately, causing large computation redundancy. In this paper, we design a new VQE network framework containing a Module-Agnostic Feature Extractor (MAFE), where the extracted latent feature are uniformly share across three modules. This design allows the network to greatly reduce the redundant feature computation steps and speed up the inference of the network.", + "bbox": [ + 212, + 145, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When processing high-resolution videos in VQE task, the complex network of the sota method will greatly increase the computation times of the network, thus affecting the online deployment of the method. Therefore, we in our proposed framework: (1) Redesign the structure of the three modules with a light-weight convolutional structure. (2) Propose a Spatial-Temporal Look-up Tables (ST-LUTs) that extend conventional LUTs [13] to conduct queries on both spatial and temporal dimensions, which allows the LUTs to simultaneously capture information in both temporal and spatial dimensions, facilitating efficient extraction of temporal information. To the best of our knowledge, ST-LUTs is the first approach to employ a LUT structure for the extraction of temporal information.", + "bbox": [ + 212, + 343, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The major contributions are summarized as follows: (1) We propose the first online-VQE method STLVQE that achieves real-time process speed, which reduces redundant feature extraction computations while redesigning the propagation, alignment and enhancement modules of the temporal VQE network. (2) A Spatial-Temporal Look-up Tables structure is first time proposed for fully exploiting the temporal and spatial information in the video. (3) Extensive experiments on MFQE 2.0 dataset demonstrate that our approach achieves a comparable speed-performance trade-off. We also test most of the previous VQE methods on prevalent GPU to provide performance and speed benchmarks, facilitating future research in this area.", + "bbox": [ + 212, + 494, + 787, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 670, + 397, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compressed Video Quality Enhancement. Recently, numerous works have focused on the degradation of the network video quality due to the lossy video compression algorithms. Existing Compressed VQE work can be broadly classified into two categories, which are single-image based method [4, 5, 8, 11, 12, 17, 30, 34] and multi-frame based method [2, 6, 7, 10, 20, 28, 31, 36].", + "bbox": [ + 212, + 703, + 787, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the single-image based methods, some of the early work [4,8,11,17,29,34] are mainly used to enhance the quality of JPEG compressed images, which can also be applied directly to compressed videos to process each frame independently. [5, 12, 30] utilize certain characteristics of the video compression algo", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rhythm or compression parameters to improve the video quality. For example, DS-CNN [30] and QE-CNN [29] use distinct models to enhance the video processed by different compression coding.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the multi-frame based methods, MFQE 1.0 [31] and MFQE 2.0 [10] were proposed to use two adjacent high quality frames(PQF) and compute the optical flow between them to enhance the current frame. STDF [6] uses deformation convolution for an implicit alignment. RFDA [36] extracts long-range temporal information from the video via the recursive fusion module. BasicVSR [1] and BasicVSR++ [2] utilize a bidirectional propagation network and a new alignment scheme is proposed in BasicVSR++. All these methods demonstrate the significance of temporal information for solving the compressed VQE problem.", + "bbox": [ + 212, + 191, + 784, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As far as we know, there is still no compression VQE method for online scenario that has been proposed so far. The above-mentioned methods cannot be directly applied to webcasting environments due to the limitations of the two online restriction mentioned above.", + "bbox": [ + 212, + 313, + 784, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Look-up Table. Look-up Tables (LUTs) have proven to be highly effective in reducing computational power and time consumption during the inference phase of algorithms by replacing complex computation with fast storage space retrieval operations. However, the use of LUTs directly is not always flexible as it requires manual adjustment by experts to accurately process images. To overcome this limitation, recent attempts have been made to improve LUTs. [25, 33] combine multiple LUTs and learn the weights through neural networks to summing up these LUT results. SR-LUT [13] replaces a small receptive field neural network with a 4D-LUT and achieves good speed-performance trade-offs on SR tasks. SP-LUT [21] and MuLUT [16] makes improvements to the SR-LUT by expanding its receptive field through parallel and serial multiple LUT tables. In this work, we utilize LUTs to extract information in both temporal and spatial dimensions, bringing its advantages to the online VQE task.", + "bbox": [ + 212, + 373, + 785, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 588, + 330, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 214, + 614, + 341, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a stream of compressed frames without an explicit ending $\\{x_0, \\ldots, x_i, \\ldots\\}$ , where $x_i \\in \\mathbb{R}^{C \\times H \\times W}$ with input channel $C$ , height $H$ and width $W$ , we sequentially enhance each compressed frame to obtain the target sequence $\\{y_0, \\ldots, y_i, \\ldots\\}$ . Due to the low latency requirement of Online-VQE, obtaining information from future frames to enhance $x_i$ is unacceptable. We only use past and current frames $\\{x_0, x_1, \\ldots, x_i\\}$ to predict $y_i$ , reducing the video's delay time. Additionally, to avoid stuttering issues, each frame's inference speed must be fast enough.", + "bbox": [ + 212, + 636, + 784, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall flow of our proposed network during inference period is illustrated in Fig. 2 (The difference between the training and the inference phase is illustrated in Subsec. 3.4 and the figure in appendix). First, a Module-Agnostic Feature Extractor (MAFE) is utilized to extract the feature $F_{\\text{eat}_i}$ of the current enhancing frame $x_i$ as follows,", + "bbox": [ + 212, + 742, + 784, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF e a t _ {i} = M A F E \\left(x _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 824, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26385d2ba90d424f1d876ac9095939b18e6358def4e14f07edf284979a24d154.jpg", + "image_caption": [ + "Fig. 2: The framework of STLVQE (inference phase), which consists of a Module-Agnostic Feature Extractor and three main modules. In the inference phase, the propagation module selects the reference frame and accesses the relevant information, which is sent to the alignment module for temporal alignment, and finally the aligned reference and enhancing frames are input to the enhancement module for the final results." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 781, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\text{Feat}_i$ will be shared by the three subsequent modules, thus reducing the multiple time-consuming extraction steps and fully exploiting the latent frame feature.", + "bbox": [ + 212, + 397, + 784, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, in the propagation module, $Feat_{i}$ is stored in a structure called Temporal Cache, which will provide the temporal information in subsequent modules based on the selection of the reference frames $x_{ref} = \\{x_{ref1}, x_{ref2}\\}$ without external computation. In the alignment module, $Feat_{i}$ is concatenated with $Feat_{ref}$ , the reference frame feature obtained from the propagation module, to deform $x_{ref}$ with a lightweight offset predictor. In the enhancement module, $Feat_{i}$ is used to perform a spatial complement, which fully exploits the extracted features and augments the ST-LUTs by increasing its receptive fields.", + "bbox": [ + 212, + 441, + 784, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Under our proposed new framework, only once feature extraction computation for a single frame is required for enhancing each frame. The extracted features are simultaneously computed with reference frame alignment and preliminary compensation in the current frame space.", + "bbox": [ + 212, + 564, + 785, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 The Propagation Module", + "text_level": 1, + "bbox": [ + 214, + 647, + 470, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The propagation module in the VQE task is responsible for extracting and transmitting temporal information within the video. Previous methods have explored various approaches, including designing additional structures [10] and increasing the number of selected reference frames [6, 36] to improve network performance, which greatly increase inference time and computational costs of the model. The architecture of the propagation module is illustrated in Fig. 3. The compression configuration parameter quantization parameter (QP) value is utilized as an indicator to select fewer but higher quality frames among multiple candidate frames without extra computation. Specifically, a Reference Frame Window (RFW) which contains $W$ frames $\\{x_{i-W}\\dots x_{i-1}\\}$ is set. Then, two frames with the smallest QP values are selected as $x_{ref}$ for the subsequent module.", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 230, + 113, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4dee20de2c65bd311fda64ee50378b0e747581a802da31f2b17234971fa8794b.jpg", + "image_caption": [ + "Fig. 3: Reference Frame Window and Temporal Cache in the propagation module." + ], + "image_footnote": [], + "bbox": [ + 305, + 145, + 697, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To mitigate the issue of additional feature extraction operations in previous works, we introduce a cache-like structure known as the Temporal Cache (TC). The TC stores reusable features of past frames that were already extracted by the MAFE in previous process, thereby eliminating the need to re-extract features for each subsequent frame. In particular, when enhancing the i-th frame $x_{i}$ , STLVQE caches the intermediate feature map $Fea t_{i}$ extracted by the alignment module, as well as the enhanced frame, in the information cache (TC). During reference frame selection, the features and enhanced result of $x_{ref1}$ and $x_{ref2}$ are directly retrieved from TC without extra feature extraction process.", + "bbox": [ + 212, + 367, + 787, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 The Alignment Module", + "text_level": 1, + "bbox": [ + 215, + 523, + 455, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The alignment module is a crucial component in video processing tasks that handles spatial misalignments between frames, which plays a vital role in effectively utilizing the temporal information embedded in the video, especially when the video content is subject to large motion. STLVQE includes a lightweight offset predictor to replace the optical flow or non-local computation of most VQE methods, which significantly reduce the computation cost.", + "bbox": [ + 212, + 547, + 787, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The offset predictor is responsible for predicting spatial offsets which align correlated pixels in the reference with pixels in $\\tilde{x}_i$ . As introduced in the last section, the latent features $Feat_{ref}$ and spatially compensated frames $\\tilde{x}_{ref}$ at each processing step are stored in TC, which are retrieved by RFW to enhance future frames. Given the feature of a reference frame $\\tilde{x}_{ref}$ retrieved from the TC, as well as the feature of the current frame $\\tilde{x}_i$ , the offset predictor predicts 9 spatial offsets for each pixel in $\\tilde{x}_i$ :", + "bbox": [ + 212, + 638, + 787, + 743 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\no _ {r e f} = C _ {\\text {o f f s e t}} \\left(\\left[ \\operatorname {F e a t} _ {i}, \\operatorname {F e a t} _ {r e f} ^ {d} \\right]\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 752, + 785, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the matrix concatenation is denoted by $[\\cdot]$ , and each offset value in $o_{ref}$ represents the position shift of a correlated pixel in $\\tilde{x}_{ref}$ relative to a pixel in $\\tilde{x}_i$ .", + "bbox": [ + 212, + 779, + 785, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The temporal corresponding pixels in the reference frames $x_{ref}$ are spatially aligned with the $3 \\times 3$ pixel offsets $o_{ref}$ and bilinear interpolation. Each set of", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "obtained correlated pixel values of a pixel in $\\tilde{x}_i$ are re-arranged into a $3 \\times 3$ patch, which are then stacked together to obtain the deformed reference frame denoted as $\\tilde{x}_{ref}^d$ with the size of $N \\times C \\times 3H \\times 3W$ .", + "bbox": [ + 212, + 146, + 782, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 The Enhancement Module", + "text_level": 1, + "bbox": [ + 214, + 212, + 480, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The enhancement module contains a spatial compensation network and a ST-LUTs network. To fully utilize the sharing latent feature $Feat_{i}$ , a spatial compensation network reuses $Feat_{i}$ to generate a spatial residual, which compensates the low quality frame $x_{i}$ to obtain preliminary spatial enhancement before ST-LUTs:", + "bbox": [ + 212, + 234, + 782, + 306 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} _ {i} = x _ {i} + C _ {\\text {c o m p}} (\\text {F e a t} _ {i}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 311, + 784, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $C_{comp}$ are 3 shallow convolutional layers. Since the subsequent enhancement module takes the obtained compensated frame $\\tilde{x}_i$ as input, spatial compensation network can boost the enhancement performance by expanding the receptive field of the ST-LUTs in the enhancement module. To match the resolution of $\\tilde{x}_{ref}^{d}$ , $\\tilde{x}_i$ is upsampled by repeating each pixel in $\\tilde{x}_i$ to form a 3x3 patch, then each pixel in the up-sampled $\\tilde{x}_i$ is correlated with the pixel at the same position in $\\tilde{x}_{ref}^{d}$ . This is achieved by the unfold and fold operations with different strides as follows:", + "bbox": [ + 212, + 330, + 784, + 450 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} _ {i} ^ {u p} = \\operatorname {F o l d} (\\operatorname {U n f o l d} (\\tilde {x} _ {i}, \\text {s t r i d e} = 1), \\text {s t r i d e} = 3), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 462, + 784, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the up-sampled input frame $\\tilde{x}_i^{up}$ and deformed reference frame $\\tilde{x}_{ref}^d$ together serve as the input of the ST-LUTs network for the final quality enhancement.", + "bbox": [ + 212, + 486, + 782, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instead of deep neural network, STLVQE adopts a novel Spatial-Temporal Look-up Tables in the enhancement module during inference to replace complex convolutional computation used in training stage with much faster direct memory access, and conducts table queries on both spatial and temporal perspective.", + "bbox": [ + 212, + 517, + 782, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantization. Due to the limited storage space of the LUT, the size of each pixel in the input compensated frame $\\tilde{x}_i^{up}$ and the deformed frames $\\tilde{x}_{ref}^d$ needs to be controlled to fit within a suitable range of bits. Because the range of values in $\\tilde{x}_i^{up}$ and $\\tilde{x}_{ref}^d$ can vary dynamically and the gradient needs to be passed through the quantization operation during end-to-end training, it's not practical to simply quantize the matrices using a fixed quantization hyper-parameter. The enhancement module adopts a Learnable Quantization layer (LSQ) [9] to quantize the input of the enhancement module as follows:", + "bbox": [ + 212, + 577, + 784, + 696 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {x} = \\left\\lfloor \\operatorname {c l i p} \\left(\\tilde {x} / s, - Q _ {N}, Q _ {P}\\right) \\right], \\quad \\hat {x} = \\bar {x} \\times s, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 714, + 784, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\text{clip}(z, r_1, r_2)$ returns matrix $z$ with values below $r_1$ set to $r_1$ and values above $r_2$ set to $r_2$ , and the $\\lfloor z \\rfloor$ rounds $z$ to the nearest integer. The step size parameter $s$ can be optimized along with the network's training. $\\overline{x}$ is the indexes used in the LUTs store and retrieval stages, and $\\hat{x}$ is a quantized representation of the data with the same scale as $x$ . After quantization, the deep enhancement network and the corresponding LUTs can be constructed with the quantized pixel values as input.", + "bbox": [ + 212, + 734, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 230, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e72c8caf4e5426fa770ce31b16d6670bfe8e9e5eb09e0024a0c6df5e3335f99.jpg", + "image_caption": [ + "Fig. 4: The specific schematic of our Spatial-Temporal Look-up Tables, which consists of two parts: Temporal-LUT (left side) and Spatial-LUT (right side)." + ], + "image_footnote": [], + "bbox": [ + 251, + 142, + 748, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Deep Enhancement Network Training. In the training phase, a neural network based enhancement module is optimized, which takes the compensated frame $\\hat{x}_i^{up}$ and reference frames $\\hat{x}_{ref}^d$ as input and output the final enhancement results. To train an enhancement network that is suitable for LUT conversion, the receptive field size of the deep network typically cannot exceed 6 (Larger receptive fields will result in a higher storage occupancy, refer to the appendix for details). Therefore, each branch of the enhancement network, which will be replaced by ST-LUTs during the inference stage, consists of 1 convolutional layer with a kernel size of 4 or 6 and 10 convolutional layers with $1\\times 1$ kernel. The specific design of the network will be illustrated in the appendix.", + "bbox": [ + 212, + 431, + 787, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ST-LUTs Construction & Replacement. We generate LUTs by permuting all input patches of the receptive field size, and storing all the input/output mappings of the trained deep enhancement network in mapping tables. During inference, we retrieve the outputs of the deep enhancement network directly from the LUTs, replacing slow and computationally-intensive convolution with faster memory access and retain the other components in STLVQE. Due to the input patch size restriction, conventional LUTs are not suitable for inter-frame queries in the temporal dimension. To address these issues, we firstly propose the STLUTs, which employ multiple LUTs to conduct parallel retrieval with different sets of query pixels in both spatial and temporal dimensions.", + "bbox": [ + 212, + 582, + 787, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, given the quantized compensated frames $\\hat{x}_i^{up}$ and the aligned frames $\\hat{x}_{ref}^d$ as input, the look-up tables are constructed as shown in Fig. 4, where two sets of LUTs named S-LUTs and T-LUTs are constructed for extracting spatial and temporal information respectively.", + "bbox": [ + 212, + 734, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For S-LUTs, a rotation ensemble operation [13] is used during the training process to further expand the LUT receptive field. As shown in Fig. 4, during training, input images are rotated by 0, 90, 180, and 270 degrees, and the results", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are reverse-rotated and summed up. This approach considers all pixels in a $3 \\times 3$ receptive field simultaneously. The result of S-LUT $R_{i}^{S}$ is expressed as follows:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i} ^ {S} = \\frac {1}{4} \\sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \\left(f ^ {S} \\left(R o t _ {j} \\left(\\hat {x} _ {i} ^ {u p}\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 186, + 785, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $Rot_{j}$ and $Rot_{j}^{-1}$ are the rotation and reserve-rotation operations, and $f^{S}$ is the spatial enhancement network or the S-LUTs retrieval.", + "bbox": [ + 212, + 239, + 782, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For T-LUTs, the pixels at the same location of the enhancing and reference frames will be strongly correlated after alignment, and jointly considering these values can make full use of the temporal information in the video. As shown on the left in Fig. 4, we establish two T-LUTs: T-LUT-1 take the same position pixels of $\\hat{x}_i^{up}$ and $\\hat{x}_{ref}^d$ as index with the rotation ensemble operation, while T-LUT-2 uses the four edge pixels of each kernel in the $\\hat{x}_{ref}^d$ as indexes to retrieve the surrounding information of the target pixels. The result of T-LUTs $R_i^T$ can be expressed as follows:", + "bbox": [ + 212, + 271, + 784, + 393 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R _ {i} ^ {T 1} = \\frac {1}{4} \\sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \\left(f ^ {T 1} \\left(R o t _ {j} \\left(\\left[ \\hat {x} _ {i} ^ {u p}, \\hat {x} _ {r e f} ^ {d} \\right]\\right)\\right)\\right), \\tag {7} \\\\ R _ {i} ^ {T 2} = f ^ {T 2} (\\hat {x} _ {r e f} ^ {d}), \\quad R _ {i} ^ {T} = R _ {i} ^ {T 1} + R _ {i} ^ {T 2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 406, + 784, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $f^{T1}$ and $f^{T2}$ are the temporal enhancement network or the T-LUTs retrieval. Then we can acquire the results of ST-LUTs and get the final enhanced image $y_{i}$ as follows:", + "bbox": [ + 212, + 478, + 782, + 523 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i} = R _ {i} ^ {S} + R _ {i} ^ {T}, \\quad y _ {i} = \\tilde {x} _ {i} + R _ {i}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 534, + 784, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training Scheme", + "text_level": 1, + "bbox": [ + 214, + 571, + 400, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We perform the training of the STLVQE in two stages with an end-to-end fashion. Firstly, the Charbonnier Loss [3] is utilized to optimize the model as follows:", + "bbox": [ + 212, + 595, + 782, + 626 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c h a r} = \\sqrt {\\left(y _ {t} - \\bar {y} _ {t}\\right) ^ {2} + \\epsilon}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 640, + 784, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $y_{t}$ and $\\overline{y_t}$ is the enhanced and ground truth frame at time $t$ , and $\\epsilon$ is set to $10^{-6}$ in our paper. Secondly, to further enhance the enhancement capability, we perform fine-tuning using L2 loss with a smaller learning rate as follows:", + "bbox": [ + 212, + 662, + 782, + 708 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M S E} = \\left\\| \\left(y _ {t} - \\bar {y} _ {t}\\right) ^ {2} \\right\\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 715, + 784, + 737 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 755, + 375, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 785, + 333, + 799 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following [6, 10, 36], we conduct our experiments on the MFQE 2.0 dataset. It contains totally 126 sequences videos with multi resolutions: SIF(352 × 240),", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 230, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5b69e6eab104892474a4a1848e5f522660bd0bef7579a7e48ed5482269ac6d9a.jpg", + "table_caption": [ + "Table 1: Overall comparison for $\\Delta$ PSNR(dB), $\\Delta$ SSIM( $x$ ) $^{-2}$ ), Latency(ms) and FPS over test sequences of MFQE 2.0 Dataset at QP=37. All methods were retested on a NVIDIA RTX 3090 in frame-to-frame manner. Video resolution: Class A(2560\\times1600), Class B(1980\\times1080), Class C(832\\times480), Class D(416\\times240), Class E(1280\\times720)." + ], + "table_footnote": [], + "table_body": "
ApproachHigh-Latency VQE MethodsLow-Latency VQE Methods
MFQE 1.0 [31]PSTQE [7]Li et al. [17]DS-CNN [29]DnCNN [34]DCAD [26]STLVQE (Ours)
MetricsΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%
ATraffic0.50/0.901120.10.64/1.045.6K0.29/0.60385.50.29/0.60374.60.24/0.57294.90.31/0.67153.70.41/0.9496.3
PeopleOnStreet0.80/1.371.00.48/0.922.60.42/0.852.70.41/0.823.40.50/0.956.50.68/1.3510.4
BKimono0.50/1.13628.10.69/1.362.9K0.28/0.780.25/0.750.24/0.750.24/0.750.24/0.750.28/0.780.28/0.780.38/1.0548.6
ParkScene0.39/1.030.42/1.150.49/1.210.15/0.480.23/0.580.15/0.500.14/0.500.14/0.500.16/0.500.16/0.500.16/0.500.33/0.8020.6
Cactus0.44/0.882.00.62/1.150.25/0.480.25/0.480.24/0.580.20/0.480.20/0.380.26/0.580.28/0.500.28/0.500.43/0.8548.6
BQTerrace0.27/0.480.50/0.870.50/0.870.30/0.680.30/0.680.26/0.480.25/0.580.25/0.580.28/0.650.31/0.680.31/0.680.42/0.8320.6
BasketballDrive0.41/0.800.60/1.040.60/1.040.30/0.680.30/0.680.28/0.650.25/0.650.25/0.650.28/0.650.28/0.650.28/0.650.39/0.9411.5
CRaceHorses0.34/0.55229.80.40/0.880.28/0.650.33/0.880.27/0.630.27/0.630.25/0.650.28/0.680.26/0.680.34/0.880.50/1.2587.0
BQMall0.51/1.03229.80.74/1.44611.60.33/0.8839.50.33/0.8036.70.28/0.6826.30.34/0.8814.30.34/1.15
PartyScene0.22/0.7310.40.51/1.461.80.13/0.4525.30.17/0.5827.20.13/0.4838.00.16/0.4869.90.37/1.37
BasketballDrill0.48/0.901.66/1.270.66/1.270.38/0.880.38/0.880.35/0.680.33/0.680.33/0.680.39/0.780.39/0.780.39/0.780.44/1.089.1
DRaceHorses0.51/1.13164.30.60/1.44203.70.33/0.830.32/0.750.32/0.750.31/0.730.31/0.730.32/0.830.32/0.750.47/1.059.1
BQSquare-0.01/0.15164.30.79/1.140.09/0.250.09/0.250.20/0.380.20/0.380.13/0.180.13/0.187.00.20/0.380.35/1.37109.9
BlowingBubbles0.39/1.2032.20.62/1.950.21/0.680.21/0.680.23/0.680.23/0.680.18/0.580.18/0.58142.90.22/0.65285.70.37/1.23
BasketballPass0.63/1.380.85/1.750.85/1.750.34/0.850.34/0.780.34/0.780.34/0.780.31/0.750.35/0.850.35/0.850.35/0.850.48/1.28128
EFourPeople0.66/0.85345.40.95/1.121.3K0.45/0.700.46/0.700.46/0.700.39/0.600.42/0.600.51/0.780.51/0.780.69/1.0224.7
Johnny0.55/0.554.70.75/0.850.850.40/0.6087.40.38/0.4081.60.32/0.4064.80.41/0.500.93/0.6440.5
KristenAndSara0.66/0.750.66/0.750.93/0.910.49/0.680.49/0.680.48/0.600.42/0.600.42/0.600.52/0.700.52/0.700.52/0.700.74/0.84128
Average0.46/0.88-0.69/1.25-0.30/0.66-0.30/0.63-0.26/0.58-0.32/0.67-0.49/1.03-
", + "bbox": [ + 215, + 212, + 787, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CIF(352 × 288), NTSC(720 × 486), 4CIF(704 × 576), 240P(416 × 240), 360P(640 × 360), 480P(832 × 280), 720P(1280 × 720), 1080P(1920 × 1080) and WQXGA (2560 × 1600). Same as [10], the 108 videos in the dataset are used as the training set and 18 videos are used as the test set. All videos were compressed by HM16.5 software in LDP mode with QP values of 22, 27, 32, 37 and 42.", + "bbox": [ + 212, + 412, + 789, + 491 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 508, + 457, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our proposed network is implemented based on PyTorch framework and MMEditing toolbox [23]. During the training phase, we set batch size to 8 with $180 \\times 180$ clips from the raw videos and the corresponding compressed videos as training samples. The random flip and rotation augmentation are utilized to fully exploit the information in the samples. The Adam optimizer [14] and Cosine Annealing scheme [19] are utilized for optimizing. The total iterations numbers and initial learning rate of the first and second stages is set to $200\\mathrm{K}$ , $1 \\times 10^{-4}$ and $100\\mathrm{K}$ , $1 \\times 10^{-6}$ . Our network is trained on 8 Telsa V100 GPUs.", + "bbox": [ + 212, + 532, + 787, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During the inference phase, ST-LUTs will replace the corresponding convolutional layers. To evaluate the model's performance in an online setting, we conduct inference on the test set frame-by-frame. Following the same settings as [10], we report quality enhancement on Y-channel in YUV/YCbCr space. The Peak Singal-to-Noise Ratio(ΔPSNR) and Structural Similarity(ΔSSIM) will be used as metrics to judge the model's enhancement capability. We further evaluate the latency(LT) and frame per second(FPS) of the models which is crucial to the online-VQE:", + "bbox": [ + 212, + 652, + 787, + 773 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL T = T _ {\\text {p r o c e s s}} + T _ {\\text {w a i t}}, \\quad T _ {\\text {w a i t}} = N _ {f} / F P S, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 782, + 785, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $T_{process}$ refers to the single-frame inference time and $T_{wait}$ refers to the future frames waiting time required for some video based methods, $N_f$ represents", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cc4f13063cefd500005b12e272f9eafb3cc9c8610ab87b4cb4106d1c5c62488a.jpg", + "table_caption": [ + "Table 2: Latency on 720P video and averaged $\\Delta$ PSNR/ $\\Delta$ SSIM comparison on the test sequences of MFQE 2.0 Dataset at different QP values. FV-Required means the full compressed video is required for enhancement, thus latency cannot be calculated." + ], + "table_footnote": [], + "table_body": "
MethodTypeLT(ms)StorageGFLOPsQP37QP42QP32QP27QP22
MFQEv1 [31]Bi-MF345.41788K-0.46/0.880.44/1.300.43/0.580.40/0.340.31/0.19
MFQEv2 [10]Bi-MF320.1255K7880.56/1.090.59/1.650.52/0.680.49/0.420.46/0.27
STDF [6]Bi-MF85.6487K2930.65/1.18-/-0.64/0.770.59/0.470.51/0.27
PSTQE [7]Bi-MF1249.8218K6790.69/1.250.69/1.860.67/0.830.63/0.520.55/0.29
RFDA [36]Bi-MF312.81250K-0.91/1.620.82/2.200.87/1.070.82/0.680.76/0.42
BasicVSR++ [2]Bi-MFFV-Required7300K4230.95/1.80-/-0.89/1.25-/--/-
RNAN [35]SF4329.48960K69790.44/0.95-/-0.41/0.62-/--/-
ARCNN [8]SF25.1118K1080.23/0.450.29/0.960.18/0.190.18/0.140.14/0.08
DnCNN [34]SF64.8558K5160.26/0.580.22/0.770.26/0.350.27/0.240.29/0.18
DS-CNN [29]SF81.6948K4370.30/0.630.31/1.010.27/0.380.27/0.230.25/0.15
Li et al. [17]SF87.4464K4340.30/0.660.32/1.050.28/0.370.30/0.280.30/0.19
DCAD [26]SF34.3299K2750.32/0.670.32/1.090.32/0.440.32/0.300.31/0.19
STLVQEUni-MF24.7398K+228M980.49/1.030.50/1.600.45/0.640.41/0.400.34/0.24
", + "bbox": [ + 217, + 198, + 784, + 340 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the number of future frames used. Note that the LT metric ignores the frame reception and decoding time, which is fair for different methods, and we set $FPS = 30$ in Eq. 11 as the general online video standard to facilitate comparison.", + "bbox": [ + 214, + 369, + 787, + 417 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison with State of the art Methods", + "text_level": 1, + "bbox": [ + 214, + 436, + 616, + 454 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this subsection, we present the results of our extensive experiments on various QP compressed videos in Tab. 1 and Tab. 2. Our proposed STLVQE network is compared with some widely used or state-of-art single-frame methods [8, 17, 26, 29, 34, 35], as well as some efficient multi-frame methods [6, 7, 10, 31]. In online environment where video is streamed in real-time, Bi-direction Multi-Frame (Bi-MF) method necessitates waiting for future frames that have not yet been received in current time. This waiting period introduces uncontrollable latency to the stream. Consequently, as outlined in previous online works [18, 22, 32], the Bi-MF method can not be directly applied to online scenarios and is predominantly utilized for offline processing. As a result, both Bi-MF and slow single frame methods $(>100\\mathrm{ms})$ are placed in the first part in Tab. 1 and Tab. 2 for performance comparison, and demonstrate our favorable trade-off between performance and inference speed. All experiments were conducted on an NVIDIA RTX 3090 using a frame-to-frame style.", + "bbox": [ + 212, + 462, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Latency and FPS. Tab. 1 presents the latency and FPS metrics for each VQE approach. We note that while low-latency VQE methods can process each frame at a lower latency and reduce video delay, they often struggle to handle high-resolution videos in real-time, which can lead to stuttering and reduced smoothness. Among the methods listed, only STLVQE can achieve FPS above 40 on 720P resolution videos, making it suitable for most online scenarios. Compared to the second fastest method DCAD, STLVQE's processing time on 720p video is 9.6 milliseconds faster and outperforms it by a considerable margin, improving PSNR and SSIM metrics by 0.17dB and $0.36 \\times 10^{-2}$ , respectively.", + "bbox": [ + 212, + 674, + 787, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 2 shows the latency of some large-scale high-latency temporal VQE methods. In contrast, high-latency VQE methods require waiting for future", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 230, + 113, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/740f685a70059ef27c94b88227a643ac21765737ec2d16c738899e9355fb8de6.jpg", + "table_caption": [ + "Table 3: Ablation study of Temporal Cache(TC), Reference Frame Window(RFW), Deform Operation(DF), and Preliminary Complement Residual(PC). The latter three sections share the features extracted by MAFE." + ], + "table_footnote": [], + "table_body": "
RFWTCDFPCΔP/ΔSTimeParams(K)
✓✓0.06/0.0110.85<1
✓✓✓0.31/0.5121.85352
✓✓✓0.39/0.8720.12149
✓✓✓0.49/1.0345.61398
✓✓✓0.46/0.9824.32398
✓✓✓✓0.49/1.0324.65398
", + "bbox": [ + 344, + 198, + 660, + 301 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "frames or taking a long time for inference, which can cause uncontrollable video delay. The latency of the two high-latency methods in Tab. 1 we evaluated exceeds $400\\mathrm{ms}$ , much higher than the latency caused by STLVQE. Moreover, when considering only the FPS metric, STLVQE achieves a significantly faster speed than the single-frame VQE method RNAN(FPS = 0.2) by a factor of 175 on 720P video. Compared with MFQE 1.0, our method is 8.6 times faster with a better enhancement performance. Additionally, STLVQE has a more significant speed advantage on higher resolution videos such as 1080P and 2K.", + "bbox": [ + 212, + 327, + 785, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Enhancement performance. In Tab. 1 and Tab. 2, we present the performance of STLVQE compared to other commonly used or sota VQE methods for each video at $\\mathrm{QP} = 37$ and for different QP compressed videos, respectively. Our method outperforms all the listed low-latency methods. In comparison with the high-latency VQE method, STLVQE reaches higher level with RNAN at $\\mathrm{QP} = 37$ and $\\mathrm{QP} = 32$ , with only $2.6\\%$ and $0.6\\%$ of RNAN's number of parameters and latency time (720P), respectively. The $\\Delta$ SSIM of STLVQE is higher than that of the MFQE method for all five QPs, indicating that our method produces an enhanced image with a more stable image structure. The specific visualization results and analysis can be found in Qualitative Comparison section. STDF is a robust VQE method, but we believe that a stable latency and higher frame rate are more important for enhancing the user's viewing experience.", + "bbox": [ + 212, + 449, + 785, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Memory Size & GPU Consumption. As shown in Tab. 2, STLVQE has space occupation of 398KB for the model and 228MB for the LUT (utilizing less than $5\\%$ memory of a low end GPU like GTX 1060), achieving real-time processing capability. Since most deep networks are replaced with ST-LUTs, inferencing STLVQE only needs 98 GFLOPs on a 720P frame. Note that GPU computing unit are more expensive than the memory.", + "bbox": [ + 212, + 628, + 785, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 739, + 390, + 757 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed new VQE framework, we perform ablation experiments and specific analysis on three components. The results of the runtime per frame is computed on 720P videos at $\\mathrm{QP} = 37$ .", + "bbox": [ + 212, + 763, + 785, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of the propagation module. The results are shown in the 5th and 6th row of Tab. 3. We found that selecting frames based on QP val-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6d3ec5921213555a8175319b118cb2622f944bd3cca3d5c22bdba3b8fa0b4ccf.jpg", + "table_caption": [ + "Table 4: Ablation study of STLUTs and the CNN enhancement network. The runtime is computed on a 720P video." + ], + "table_footnote": [], + "table_body": "
Enhancement ModuleΔP/ΔSTime(ms)Params(K)
S-LUT0.39/0.8323.52398
S-LUT + T-LUT-10.45/1.0224.12398
ST-LUTs0.47/1.0224.65398
ST-LUTs (on x̂i)0.49/1.0324.65398
CNN Network0.50/1.08147.02442
", + "bbox": [ + 316, + 184, + 686, + 273 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ues in RFW can bring a 0.03dB $\\Delta$ PSNR performance improvement compared to randomly selecting frames in the RFW, without adding parameters to the network and only increasing the inference time by $0.33\\mathrm{ms}$ for comparing QP values of each frame. Furthermore, as shown in rows 4 and 6 of Tab. 3, the use of the TC structure can significantly reduce the inference time by $21.0~\\mathrm{ms}$ , which is consumed by the repeated computation of the reference frame information. We achieve this reduction in time by merely occupying the GPU memory with information of several frames in the RFW.", + "bbox": [ + 212, + 305, + 787, + 426 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of the alignment module. We mainly validate the effect of the deform operation of the reference fram in the Alignment module. As shown in the first row of Tab. 3, when only using the ST-LUTs structure and discard the whole alignment module and PC operation, the result only has a $0.06\\mathrm{dB}$ PSNR improvement compared to HEVC video. When the Deform operation is applied, the network can capture more accurate temporal information, and the $\\Delta$ PSNR reaches $0.31~\\mathrm{dB}$ . Due to the lightweight convolutional layers, the alignment module only adds $11.6~\\mathrm{ms}$ of inference time to the network.", + "bbox": [ + 212, + 428, + 787, + 547 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of the enhancement module. As shown in row 3 of Tab. 3, using PC without the Deform operation achieves a significant $\\Delta$ PSNR of 0.39 dB, demonstrating the effectiveness of PC in providing initial compensation for spatial information and extending the receptive field, ultimately enhancing the performance of ST-LUTs. Finally, after combining two parts, the network achieves 0.49 dB $\\Delta$ PSNR. PC alone contributes 0.18 dB to the network and adds only 2.16 ms of inference time.", + "bbox": [ + 212, + 550, + 787, + 656 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tab. 4 provides an analysis of the impact of each LUT in the enhancement module. In the first three rows of the table, we attempted to enhance the original image $x_{i}$ with the results of different LUT combinations. Results in the first and second rows show that using only the S-LUT with T-LUT-1 does not fully improve the network due to T-LUT-1 not utilizing all pixels within a kernel when learning the alignment offset. However, when T-LUT-2 is used for network training along with T-LUT-1, the reference frame can be better aligned, resulting in a $\\Delta$ PSNR improvement of 0.08 dB. The 4th row shows the enhancing result on $\\tilde{x}_{i}$ , reducing the learning difficulty of the ST-LUTs, which improving the performance by 0.02 dB. Compared to the CNN layers replaced by LUTs, ST-LUTs sacrifice only 0.01 dB $\\Delta$ PSNR performance while reducing the number of parameters by 44K and the inference time by 122.37 ms.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 230, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/efb0ef0e07b8df0c0315bc1d9b5af99ab281b1d1b5a33aa3a4461799027fa0b4.jpg", + "image_caption": [ + "Fig. 5: Qualitative results at QP 37. The STLVQE method successfully enhances the left leg of the athlete, the face of the old man, the railings of the bridge and the eyebrow of the woman in each frame, respectively." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 785, + 354 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The above ablation studies have sufficiently demonstrated that the MAFE structure of shared features in STLVQE with the redesigned three new modules can sufficiently reduce the computational redundancy of the network and obtain satisfactory inference speed-performance trade-offs.", + "bbox": [ + 212, + 433, + 787, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 215, + 513, + 457, + 529 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 5 provides 4 examples of the qualitative results of our network on compressed video at $QP = 37$ . The quality of these images deteriorates significantly in compressed scenes, showing a variety of image defects (e.g. blurring legs, missing bridge railings). With the help of temporal information, STLVQE achieves better visual results than single-frame methods with significantly reduced runtime. Compared to large-scale multi-frame methods like MFQE 2.0, STLVQE can achieve competitive capabilities in detail and structural restoration.", + "bbox": [ + 212, + 535, + 787, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 660, + 359, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a lightweight and rapid method called STLVQE based on the constraints of online video quality enhancement tasks. STLVQE greatly reduce the computation redundancy and uses a Spatial-Temporal Look-up Table structure to extract spatio-temporal information from the video with minimal time consumption. The extensive experiments prove that our proposed method can outperform most low-latency VQE methods, achieve competitive performance with high-latency VQE methods which cannot be directly utilized on Online-VQE tasks. Furthermore, STLVQE achieves real-time processing of 720P resolution videos, demonstrating a good trade-off between enhancement performance and inference speed.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Fund of China (62076184, 61976158, 61976160, 62076182, 62276190), in part by Fundamental Research Funds for the Central Universities and State Key Laboratory of Integrated Services Networks (Xidian University), in part by Shanghai Innovation Action Project of Science and Technology (20511100700) and Shanghai Natural Science Foundation (22ZR1466700).", + "bbox": [ + 212, + 172, + 787, + 263 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 297 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021) 3, 4", + "2. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022) 2, 3, 4, 11", + "3. Charbonnier, P., Blanc-Feraud, L., Aubert, G., Barlaud, M.: Two deterministic half-quadratic regularization algorithms for computed imaging. In: Proceedings of 1st international conference on image processing. vol. 2, pp. 168-172. IEEE (1994) 9", + "4. Chen, H., He, X., Qing, L., Xiong, S., Nguyen, T.Q.: Dpw-sdnet: Dual pixel-wavelet domain deep cnns for soft decoding of JPEG-compressed images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 711-720 (2018) 2, 3", + "5. Dai, Y., Liu, D., Wu, F.: A convolutional neural network approach for post-processing in hevc intra coding. In: MultiMedia Modeling: 23rd International Conference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23. pp. 28-39. Springer (2017) 3", + "6. Deng, J., Wang, L., Pu, S., Zhuo, C.: Spatio-temporal deformable convolution for compressed video quality enhancement. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 10696-10703 (2020) 2, 3, 4, 5, 9, 11", + "7. Ding, Q., Shen, L., Yu, L., Yang, H., Xu, M.: Patch-wise spatial-temporal quality enhancement for hevc compressed video. IEEE Transactions on Image Processing 30, 6459-6472 (2021) 3, 10, 11", + "8. Dong, C., Deng, Y., Loy, C.C., Tang, X.: Compression artifacts reduction by a deep convolutional network. In: Proceedings of the IEEE international conference on computer vision. pp. 576-584 (2015) 3, 11", + "9. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. arXiv preprint arXiv:1902.08153 (2019) 7", + "10. Guan, Z., Xing, Q., Xu, M., Yang, R., Liu, T., Wang, Z.: Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE transactions on pattern analysis and machine intelligence 43(3), 949-963 (2019) 2, 3, 4, 5, 9, 10, 11", + "11. Guo, J., Chao, H.: Building dual-domain representations for compression artifacts reduction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 628-644. Springer (2016) 3" + ], + "bbox": [ + 217, + 308, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Jin, Z., An, P., Yang, C., Shen, L.: Quality enhancement for intra frame coding via cnns: An adversarial approach. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1368-1372. IEEE (2018) 2, 3", + "13. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 691-700 (2021) 3, 4, 8", + "14. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10", + "15. Kondratyuk, D., Yuan, L., Li, Y., Zhang, L., Tan, M., Brown, M., Gong, B.: Movinets: Mobile video networks for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16020-16030 (2021) 2", + "16. Li, J., Chen, C., Cheng, Z., Xiong, Z.: Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVIII. pp. 238-256. Springer (2022) 4", + "17. Li, K., Bare, B., Yan, B.: An efficient deep convolutional neural networks model for compressed image deblocking. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 1320-1325. IEEE (2017) 3, 10, 11", + "18. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019) 2, 11", + "19. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10", + "20. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep kalman filtering network for video compression artifact reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018) 2, 3", + "21. Ma, C., Zhang, J., Zhou, J., Lu, J.: Learning series-parallel lookup tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 305-321. Springer (2022) 4", + "22. Maggioni, M., Huang, Y., Li, C., Xiao, S., Fu, Z., Song, F.: Efficient multistage video denoising with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3466-3475 (2021) 2, 11", + "23. MMEditing Contributors: MMEditing: OpenMMLab image and video editing toolbox. https://github.com/open-mmlab/mmediting (2022) 10", + "24. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012) 2", + "25. Wang, T., Li, Y., Peng, J., Ma, Y., Wang, X., Song, F., Yan, Y.: Real-time image enhancer via learnable spatial-aware 3d lookup tables. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2471-2480 (2021) 4", + "26. Wang, T., Chen, M., Chao, H.: A novel deep learning-based method of improving coding efficiency from the decoder-end for hevc. In: 2017 Data Compression Conference (DCC). pp. 410-419. IEEE (2017) 10, 11", + "27. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003) 2" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Qu et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Xu, Y., Gao, L., Tian, K., Zhou, S., Sun, H.: Non-local convlstm for video compression artifact reduction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7043-7052 (2019) 3", + "29. Yang, R., Xu, M., Liu, T., Wang, Z., Guan, Z.: Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology 29(7), 2039-2054 (2018) 2, 3, 4, 10, 11", + "30. Yang, R., Xu, M., Wang, Z.: Decoder-side hevc quality enhancement with scalable convolutional neural network. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 817-822. IEEE (2017) 2, 3, 4", + "31. Yang, R., Xu, M., Wang, Z., Li, T.: Multi-frame quality enhancement for compressed video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6664-6673 (2018) 2, 3, 4, 10, 11", + "32. Yin, G., Qu, Z., Jiang, X., Jiang, S., Han, Z., Zheng, N., Yang, H., Liu, X., Yang, Y., Li, D., et al.: Online streaming video super-resolution with convolutional lookup table. IEEE Transactions on Image Processing 33, 2305-2317 (2024) 11", + "33. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 4", + "34. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 3, 10, 11", + "35. Zhang, Y., Li, K., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082 (2019) 11", + "36. Zhao, M., Xu, Y., Zhou, S.: Recursive fusion and deformable spatiotemporal attention for video compression artifact reduction. In: Proceedings of the 29th ACM international conference on multimedia. pp. 5646-5654 (2021) 2, 3, 4, 5, 9, 11" + ], + "bbox": [ + 215, + 146, + 787, + 507 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_model.json b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be4661fd636cfe3518aa3e7e6d9aef4064e6d4d6 --- /dev/null +++ b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_model.json @@ -0,0 +1,2126 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.271, + 0.141, + 0.734, + 0.187 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.213, + 0.777, + 0.243 + ], + "angle": 0, + "content": "Zefan \\(\\mathrm{Qu}^{1}\\), Xinyang Jiang\\(^{\\dagger 2}\\), Yifan Yang\\(^{2}\\), Dongsheng Li\\(^{2}\\), and Cairong Zhao\\(^{\\dagger 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.256, + 0.628, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Tongji University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.271, + 0.627, + 0.284 + ], + "angle": 0, + "content": "{qzf, zhaocairong}@tongji.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.284, + 0.649, + 0.298 + ], + "angle": 0, + "content": "2 Microsoft Research Asia, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.299, + 0.7, + 0.312 + ], + "angle": 0, + "content": "{xinyangjiang,yifanyang,dongsheng.li}@microsoft.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.355, + 0.744, + 0.593 + ], + "angle": 0, + "content": "Abstract. Low latency rates are crucial for online video-based applications, such as video conferencing and cloud gaming, which make improving video quality in online scenarios increasingly important. However, existing quality enhancement methods are limited by slow inference speed and the requirement for temporal information contained in future frames, making it challenging to deploy them directly in online tasks. In this paper, we propose a novel method, STLVQE, specifically designed to address the rarely studied online video quality enhancement (Online-VQE) problem. Our STLVQE designs a new VQE framework which contains a Module-Agnostic Feature Extractor that greatly reduces the redundant computations and redesign the propagation, alignment, and enhancement module of the network. A Spatial-Temporal Look-up Tables (STL) is proposed, which extracts spatial-temporal information in videos while saving substantial inference time. To the best of our knowledge, we are the first to exploit the LUT structure to extract temporal information in video tasks. Extensive experiments on the MFQE 2.0 dataset demonstrate our STLVQE achieves a satisfactory performance-speed trade-off." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.605, + 0.741, + 0.633 + ], + "angle": 0, + "content": "Keywords: Online Video Quality Enhancement \\(\\cdot\\) Look-up Table \\(\\cdot\\) Video Compression Artifact Reduction" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.377, + 0.703 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.816 + ], + "angle": 0, + "content": "With the rapid development of social media, users are increasingly seeking high-quality and real-time video content. In recent times, video-based applications such as video conferencing, webcasting, and cloud gaming have increasingly demanded low latency requirements. This has resulted in higher video compression rates, which can cause a noticeable decrease in video quality and an overall reduction in the Quality of Experience (QoE). As a result, online Video Quality" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.398, + 0.841 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.333, + 0.129 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.505, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.785, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.322, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Fig. 1: Comparison of \\(\\Delta\\)PSNR/ \\(\\Delta\\)SSIM, average runtime per frame(720P video) and parameters on MFQE 2.0 test set at \\(\\mathrm{QP} = 37\\). Our STLVQE method achieves a great trade-off in enhancement performance and inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.789, + 0.534 + ], + "angle": 0, + "content": "Enhancement (Online-VQE), which improves the quality of real-time streaming video under online conditions and minimizing video defects caused by compression algorithms (e.g., H.264/AVC [27] and H.265/HEVC [24]) has become a significant problem. However, Online-VQE poses two main challenges compared to the conventional offline VQE approach. Firstly, it needs to process high-resolution videos in real-time to ensure a smooth viewing experience. Secondly, online video processing techniques [15, 18, 22] can only rely on the current and previous frames for inference due to the uncontrolled latency caused by waiting for future frames, which can result in overall video delay." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.718 + ], + "angle": 0, + "content": "In recent years, numerous neural network-based techniques for enhancing the quality of compressed video have been developed, resulting in significant advancements in this field. These methods can be broadly classified into two categories, namely single-image based VQE method [4, 12, 29, 30] and multi-frame based VQE method [2, 6, 10, 20, 36]. Single-image based approaches are relatively faster in inference speed, but they ignore the significance of temporal information in video [31]. On the other hand, multi-frame based VQE achieves state-of-the-art performance by utilizing temporal information. However, most of these approaches are not suitable for online scenarios because they rely on subsequent frames as reference frames and have unsatisfactory inference speeds. Therefore, developing a new online-VQE method that satisfies the aforementioned two constraints while effectively leveraging temporal information is essential." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Reasons that limit the speed and latency of the above temporal VQE methods can be summarized in the following two points: First, a high level of computational redundancy exists within these methods, and these homogeneous computations increase the network's inference time. Second, most of these methods employ complex and computationally intensive models, which sacrificing inference speed and latency for some of the enhancement capabilities. In this work, a more applicable VQE method for online scenarios are proposed by advancing these two perspectives." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.344 + ], + "angle": 0, + "content": "As discussed in previous works [1], modern VQE methods exploit temporal information by constructing a neural network containing three main components: the propagation, alignment and enhancement module. In all three modules, similar feature extraction operations on the original image may be involved, e.g., the reference frame selection [10, 31] in propagation module, the optical flow calculation [1, 2] in alignment module and the final result obtaining in the enhancement module [6, 10]. Most of the state-of-the-art methods adopts distinct feature extractors in different components and obtaining latent feature for each component separately, causing large computation redundancy. In this paper, we design a new VQE network framework containing a Module-Agnostic Feature Extractor (MAFE), where the extracted latent feature are uniformly share across three modules. This design allows the network to greatly reduce the redundant feature computation steps and speed up the inference of the network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.496 + ], + "angle": 0, + "content": "When processing high-resolution videos in VQE task, the complex network of the sota method will greatly increase the computation times of the network, thus affecting the online deployment of the method. Therefore, we in our proposed framework: (1) Redesign the structure of the three modules with a light-weight convolutional structure. (2) Propose a Spatial-Temporal Look-up Tables (ST-LUTs) that extend conventional LUTs [13] to conduct queries on both spatial and temporal dimensions, which allows the LUTs to simultaneously capture information in both temporal and spatial dimensions, facilitating efficient extraction of temporal information. To the best of our knowledge, ST-LUTs is the first approach to employ a LUT structure for the extraction of temporal information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.648 + ], + "angle": 0, + "content": "The major contributions are summarized as follows: (1) We propose the first online-VQE method STLVQE that achieves real-time process speed, which reduces redundant feature extraction computations while redesigning the propagation, alignment and enhancement modules of the temporal VQE network. (2) A Spatial-Temporal Look-up Tables structure is first time proposed for fully exploiting the temporal and spatial information in the video. (3) Extensive experiments on MFQE 2.0 dataset demonstrate that our approach achieves a comparable speed-performance trade-off. We also test most of the previous VQE methods on prevalent GPU to provide performance and speed benchmarks, facilitating future research in this area." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.398, + 0.687 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.781 + ], + "angle": 0, + "content": "Compressed Video Quality Enhancement. Recently, numerous works have focused on the degradation of the network video quality due to the lossy video compression algorithms. Existing Compressed VQE work can be broadly classified into two categories, which are single-image based method [4, 5, 8, 11, 12, 17, 30, 34] and multi-frame based method [2, 6, 7, 10, 20, 28, 31, 36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For the single-image based methods, some of the early work [4,8,11,17,29,34] are mainly used to enhance the quality of JPEG compressed images, which can also be applied directly to compressed videos to process each frame independently. [5, 12, 30] utilize certain characteristics of the video compression algo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.129 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "rhythm or compression parameters to improve the video quality. For example, DS-CNN [30] and QE-CNN [29] use distinct models to enhance the video processed by different compression coding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.313 + ], + "angle": 0, + "content": "For the multi-frame based methods, MFQE 1.0 [31] and MFQE 2.0 [10] were proposed to use two adjacent high quality frames(PQF) and compute the optical flow between them to enhance the current frame. STDF [6] uses deformation convolution for an implicit alignment. RFDA [36] extracts long-range temporal information from the video via the recursive fusion module. BasicVSR [1] and BasicVSR++ [2] utilize a bidirectional propagation network and a new alignment scheme is proposed in BasicVSR++. All these methods demonstrate the significance of temporal information for solving the compressed VQE problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.372 + ], + "angle": 0, + "content": "As far as we know, there is still no compression VQE method for online scenario that has been proposed so far. The above-mentioned methods cannot be directly applied to webcasting environments due to the limitations of the two online restriction mentioned above." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.787, + 0.57 + ], + "angle": 0, + "content": "Look-up Table. Look-up Tables (LUTs) have proven to be highly effective in reducing computational power and time consumption during the inference phase of algorithms by replacing complex computation with fast storage space retrieval operations. However, the use of LUTs directly is not always flexible as it requires manual adjustment by experts to accurately process images. To overcome this limitation, recent attempts have been made to improve LUTs. [25, 33] combine multiple LUTs and learn the weights through neural networks to summing up these LUT results. SR-LUT [13] replaces a small receptive field neural network with a 4D-LUT and achieves good speed-performance trade-offs on SR tasks. SP-LUT [21] and MuLUT [16] makes improvements to the SR-LUT by expanding its receptive field through parallel and serial multiple LUT tables. In this work, we utilize LUTs to extract information in both temporal and spatial dimensions, bringing its advantages to the online VQE task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.331, + 0.604 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.342, + 0.63 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.785, + 0.742 + ], + "angle": 0, + "content": "Given a stream of compressed frames without an explicit ending \\(\\{x_0, \\ldots, x_i, \\ldots\\}\\), where \\(x_i \\in \\mathbb{R}^{C \\times H \\times W}\\) with input channel \\(C\\), height \\(H\\) and width \\(W\\), we sequentially enhance each compressed frame to obtain the target sequence \\(\\{y_0, \\ldots, y_i, \\ldots\\}\\). Due to the low latency requirement of Online-VQE, obtaining information from future frames to enhance \\(x_i\\) is unacceptable. We only use past and current frames \\(\\{x_0, x_1, \\ldots, x_i\\}\\) to predict \\(y_i\\), reducing the video's delay time. Additionally, to avoid stuttering issues, each frame's inference speed must be fast enough." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.785, + 0.819 + ], + "angle": 0, + "content": "The overall flow of our proposed network during inference period is illustrated in Fig. 2 (The difference between the training and the inference phase is illustrated in Subsec. 3.4 and the figure in appendix). First, a Module-Agnostic Feature Extractor (MAFE) is utilized to extract the feature \\( F_{\\text{eat}_i} \\) of the current enhancing frame \\( x_i \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.825, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nF e a t _ {i} = M A F E \\left(x _ {i}\\right), \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.145, + 0.782, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.298, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 2: The framework of STLVQE (inference phase), which consists of a Module-Agnostic Feature Extractor and three main modules. In the inference phase, the propagation module selects the reference frame and accesses the relevant information, which is sent to the alignment module for temporal alignment, and finally the aligned reference and enhancing frames are input to the enhancement module for the final results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.442 + ], + "angle": 0, + "content": "where the \\( \\text{Feat}_i \\) will be shared by the three subsequent modules, thus reducing the multiple time-consuming extraction steps and fully exploiting the latent frame feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.564 + ], + "angle": 0, + "content": "Specifically, in the propagation module, \\(Feat_{i}\\) is stored in a structure called Temporal Cache, which will provide the temporal information in subsequent modules based on the selection of the reference frames \\(x_{ref} = \\{x_{ref1}, x_{ref2}\\}\\) without external computation. In the alignment module, \\(Feat_{i}\\) is concatenated with \\(Feat_{ref}\\), the reference frame feature obtained from the propagation module, to deform \\(x_{ref}\\) with a lightweight offset predictor. In the enhancement module, \\(Feat_{i}\\) is used to perform a spatial complement, which fully exploits the extracted features and augments the ST-LUTs by increasing its receptive fields." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.787, + 0.626 + ], + "angle": 0, + "content": "Under our proposed new framework, only once feature extraction computation for a single frame is required for enhancing each frame. The extracted features are simultaneously computed with reference frame alignment and preliminary compensation in the current frame space." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.648, + 0.472, + 0.664 + ], + "angle": 0, + "content": "3.2 The Propagation Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.844 + ], + "angle": 0, + "content": "The propagation module in the VQE task is responsible for extracting and transmitting temporal information within the video. Previous methods have explored various approaches, including designing additional structures [10] and increasing the number of selected reference frames [6, 36] to improve network performance, which greatly increase inference time and computational costs of the model. The architecture of the propagation module is illustrated in Fig. 3. The compression configuration parameter quantization parameter (QP) value is utilized as an indicator to select fewer but higher quality frames among multiple candidate frames without extra computation. Specifically, a Reference Frame Window (RFW) which contains \\( W \\) frames \\( \\{x_{i-W}\\dots x_{i-1}\\} \\) is set. Then, two frames with the smallest QP values are selected as \\( x_{ref} \\) for the subsequent module." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.129 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.146, + 0.699, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.326, + 0.777, + 0.34 + ], + "angle": 0, + "content": "Fig. 3: Reference Frame Window and Temporal Cache in the propagation module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.504 + ], + "angle": 0, + "content": "To mitigate the issue of additional feature extraction operations in previous works, we introduce a cache-like structure known as the Temporal Cache (TC). The TC stores reusable features of past frames that were already extracted by the MAFE in previous process, thereby eliminating the need to re-extract features for each subsequent frame. In particular, when enhancing the i-th frame \\( x_{i} \\), STLVQE caches the intermediate feature map \\( Fea t_{i} \\) extracted by the alignment module, as well as the enhanced frame, in the information cache (TC). During reference frame selection, the features and enhanced result of \\( x_{ref1} \\) and \\( x_{ref2} \\) are directly retrieved from TC without extra feature extraction process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.525, + 0.456, + 0.54 + ], + "angle": 0, + "content": "3.3 The Alignment Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.638 + ], + "angle": 0, + "content": "The alignment module is a crucial component in video processing tasks that handles spatial misalignments between frames, which plays a vital role in effectively utilizing the temporal information embedded in the video, especially when the video content is subject to large motion. STLVQE includes a lightweight offset predictor to replace the optical flow or non-local computation of most VQE methods, which significantly reduce the computation cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.744 + ], + "angle": 0, + "content": "The offset predictor is responsible for predicting spatial offsets which align correlated pixels in the reference with pixels in \\(\\tilde{x}_i\\). As introduced in the last section, the latent features \\(Feat_{ref}\\) and spatially compensated frames \\(\\tilde{x}_{ref}\\) at each processing step are stored in TC, which are retrieved by RFW to enhance future frames. Given the feature of a reference frame \\(\\tilde{x}_{ref}\\) retrieved from the TC, as well as the feature of the current frame \\(\\tilde{x}_i\\), the offset predictor predicts 9 spatial offsets for each pixel in \\(\\tilde{x}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.753, + 0.786, + 0.772 + ], + "angle": 0, + "content": "\\[\no _ {r e f} = C _ {\\text {o f f s e t}} \\left(\\left[ \\operatorname {F e a t} _ {i}, \\operatorname {F e a t} _ {r e f} ^ {d} \\right]\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.811 + ], + "angle": 0, + "content": "where the matrix concatenation is denoted by \\([\\cdot]\\), and each offset value in \\(o_{ref}\\) represents the position shift of a correlated pixel in \\(\\tilde{x}_{ref}\\) relative to a pixel in \\(\\tilde{x}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The temporal corresponding pixels in the reference frames \\( x_{ref} \\) are spatially aligned with the \\( 3 \\times 3 \\) pixel offsets \\( o_{ref} \\) and bilinear interpolation. Each set of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.196 + ], + "angle": 0, + "content": "obtained correlated pixel values of a pixel in \\(\\tilde{x}_i\\) are re-arranged into a \\(3 \\times 3\\) patch, which are then stacked together to obtain the deformed reference frame denoted as \\(\\tilde{x}_{ref}^d\\) with the size of \\(N \\times C \\times 3H \\times 3W\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.213, + 0.482, + 0.227 + ], + "angle": 0, + "content": "3.4 The Enhancement Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.784, + 0.308 + ], + "angle": 0, + "content": "The enhancement module contains a spatial compensation network and a ST-LUTs network. To fully utilize the sharing latent feature \\(Feat_{i}\\), a spatial compensation network reuses \\(Feat_{i}\\) to generate a spatial residual, which compensates the low quality frame \\(x_{i}\\) to obtain preliminary spatial enhancement before ST-LUTs:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.312, + 0.785, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} _ {i} = x _ {i} + C _ {\\text {c o m p}} (\\text {F e a t} _ {i}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.785, + 0.452 + ], + "angle": 0, + "content": "where \\( C_{comp} \\) are 3 shallow convolutional layers. Since the subsequent enhancement module takes the obtained compensated frame \\( \\tilde{x}_i \\) as input, spatial compensation network can boost the enhancement performance by expanding the receptive field of the ST-LUTs in the enhancement module. To match the resolution of \\( \\tilde{x}_{ref}^{d} \\), \\( \\tilde{x}_i \\) is upsampled by repeating each pixel in \\( \\tilde{x}_i \\) to form a 3x3 patch, then each pixel in the up-sampled \\( \\tilde{x}_i \\) is correlated with the pixel at the same position in \\( \\tilde{x}_{ref}^{d} \\). This is achieved by the unfold and fold operations with different strides as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.463, + 0.785, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} _ {i} ^ {u p} = \\operatorname {F o l d} (\\operatorname {U n f o l d} (\\tilde {x} _ {i}, \\text {s t r i d e} = 1), \\text {s t r i d e} = 3), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.784, + 0.517 + ], + "angle": 0, + "content": "the up-sampled input frame \\(\\tilde{x}_i^{up}\\) and deformed reference frame \\(\\tilde{x}_{ref}^d\\) together serve as the input of the ST-LUTs network for the final quality enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.784, + 0.577 + ], + "angle": 0, + "content": "Instead of deep neural network, STLVQE adopts a novel Spatial-Temporal Look-up Tables in the enhancement module during inference to replace complex convolutional computation used in training stage with much faster direct memory access, and conducts table queries on both spatial and temporal perspective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.785, + 0.698 + ], + "angle": 0, + "content": "Quantization. Due to the limited storage space of the LUT, the size of each pixel in the input compensated frame \\(\\tilde{x}_i^{up}\\) and the deformed frames \\(\\tilde{x}_{ref}^d\\) needs to be controlled to fit within a suitable range of bits. Because the range of values in \\(\\tilde{x}_i^{up}\\) and \\(\\tilde{x}_{ref}^d\\) can vary dynamically and the gradient needs to be passed through the quantization operation during end-to-end training, it's not practical to simply quantize the matrices using a fixed quantization hyper-parameter. The enhancement module adopts a Learnable Quantization layer (LSQ) [9] to quantize the input of the enhancement module as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.715, + 0.785, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\bar {x} = \\left\\lfloor \\operatorname {c l i p} \\left(\\tilde {x} / s, - Q _ {N}, Q _ {P}\\right) \\right], \\quad \\hat {x} = \\bar {x} \\times s, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\( \\text{clip}(z, r_1, r_2) \\) returns matrix \\( z \\) with values below \\( r_1 \\) set to \\( r_1 \\) and values above \\( r_2 \\) set to \\( r_2 \\), and the \\( \\lfloor z \\rfloor \\) rounds \\( z \\) to the nearest integer. The step size parameter \\( s \\) can be optimized along with the network's training. \\( \\overline{x} \\) is the indexes used in the LUTs store and retrieval stages, and \\( \\hat{x} \\) is a quantized representation of the data with the same scale as \\( x \\). After quantization, the deep enhancement network and the corresponding LUTs can be constructed with the quantized pixel values as input." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.143, + 0.749, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Fig. 4: The specific schematic of our Spatial-Temporal Look-up Tables, which consists of two parts: Temporal-LUT (left side) and Spatial-LUT (right side)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.789, + 0.583 + ], + "angle": 0, + "content": "Deep Enhancement Network Training. In the training phase, a neural network based enhancement module is optimized, which takes the compensated frame \\(\\hat{x}_i^{up}\\) and reference frames \\(\\hat{x}_{ref}^d\\) as input and output the final enhancement results. To train an enhancement network that is suitable for LUT conversion, the receptive field size of the deep network typically cannot exceed 6 (Larger receptive fields will result in a higher storage occupancy, refer to the appendix for details). Therefore, each branch of the enhancement network, which will be replaced by ST-LUTs during the inference stage, consists of 1 convolutional layer with a kernel size of 4 or 6 and 10 convolutional layers with \\(1\\times 1\\) kernel. The specific design of the network will be illustrated in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.734 + ], + "angle": 0, + "content": "ST-LUTs Construction & Replacement. We generate LUTs by permuting all input patches of the receptive field size, and storing all the input/output mappings of the trained deep enhancement network in mapping tables. During inference, we retrieve the outputs of the deep enhancement network directly from the LUTs, replacing slow and computationally-intensive convolution with faster memory access and retain the other components in STLVQE. Due to the input patch size restriction, conventional LUTs are not suitable for inter-frame queries in the temporal dimension. To address these issues, we firstly propose the STLUTs, which employ multiple LUTs to conduct parallel retrieval with different sets of query pixels in both spatial and temporal dimensions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.795 + ], + "angle": 0, + "content": "Specifically, given the quantized compensated frames \\(\\hat{x}_i^{up}\\) and the aligned frames \\(\\hat{x}_{ref}^d\\) as input, the look-up tables are constructed as shown in Fig. 4, where two sets of LUTs named S-LUTs and T-LUTs are constructed for extracting spatial and temporal information respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "For S-LUTs, a rotation ensemble operation [13] is used during the training process to further expand the LUT receptive field. As shown in Fig. 4, during training, input images are rotated by 0, 90, 180, and 270 degrees, and the results" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "are reverse-rotated and summed up. This approach considers all pixels in a \\(3 \\times 3\\) receptive field simultaneously. The result of S-LUT \\(R_{i}^{S}\\) is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.188, + 0.786, + 0.23 + ], + "angle": 0, + "content": "\\[\nR _ {i} ^ {S} = \\frac {1}{4} \\sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \\left(f ^ {S} \\left(R o t _ {j} \\left(\\hat {x} _ {i} ^ {u p}\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.784, + 0.27 + ], + "angle": 0, + "content": "where \\( Rot_{j} \\) and \\( Rot_{j}^{-1} \\) are the rotation and reserve-rotation operations, and \\( f^{S} \\) is the spatial enhancement network or the S-LUTs retrieval." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.785, + 0.395 + ], + "angle": 0, + "content": "For T-LUTs, the pixels at the same location of the enhancing and reference frames will be strongly correlated after alignment, and jointly considering these values can make full use of the temporal information in the video. As shown on the left in Fig. 4, we establish two T-LUTs: T-LUT-1 take the same position pixels of \\(\\hat{x}_i^{up}\\) and \\(\\hat{x}_{ref}^d\\) as index with the rotation ensemble operation, while T-LUT-2 uses the four edge pixels of each kernel in the \\(\\hat{x}_{ref}^d\\) as indexes to retrieve the surrounding information of the target pixels. The result of T-LUTs \\(R_i^T\\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.407, + 0.785, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R _ {i} ^ {T 1} = \\frac {1}{4} \\sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \\left(f ^ {T 1} \\left(R o t _ {j} \\left(\\left[ \\hat {x} _ {i} ^ {u p}, \\hat {x} _ {r e f} ^ {d} \\right]\\right)\\right)\\right), \\tag {7} \\\\ R _ {i} ^ {T 2} = f ^ {T 2} (\\hat {x} _ {r e f} ^ {d}), \\quad R _ {i} ^ {T} = R _ {i} ^ {T 1} + R _ {i} ^ {T 2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.784, + 0.525 + ], + "angle": 0, + "content": "where \\( f^{T1} \\) and \\( f^{T2} \\) are the temporal enhancement network or the T-LUTs retrieval. Then we can acquire the results of ST-LUTs and get the final enhanced image \\( y_{i} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.535, + 0.785, + 0.553 + ], + "angle": 0, + "content": "\\[\nR _ {i} = R _ {i} ^ {S} + R _ {i} ^ {T}, \\quad y _ {i} = \\tilde {x} _ {i} + R _ {i}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.572, + 0.401, + 0.587 + ], + "angle": 0, + "content": "3.5 Training Scheme" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.784, + 0.627 + ], + "angle": 0, + "content": "We perform the training of the STLVQE in two stages with an end-to-end fashion. Firstly, the Charbonnier Loss [3] is utilized to optimize the model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.641, + 0.785, + 0.658 + ], + "angle": 0, + "content": "\\[\nL _ {c h a r} = \\sqrt {\\left(y _ {t} - \\bar {y} _ {t}\\right) ^ {2} + \\epsilon}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.784, + 0.709 + ], + "angle": 0, + "content": "where \\( y_{t} \\) and \\( \\overline{y_t} \\) is the enhanced and ground truth frame at time \\( t \\), and \\( \\epsilon \\) is set to \\( 10^{-6} \\) in our paper. Secondly, to further enhance the enhancement capability, we perform fine-tuning using L2 loss with a smaller learning rate as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.717, + 0.785, + 0.738 + ], + "angle": 0, + "content": "\\[\nL _ {M S E} = \\left\\| \\left(y _ {t} - \\bar {y} _ {t}\\right) ^ {2} \\right\\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.376, + 0.773 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.334, + 0.8 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Following [6, 10, 36], we conduct our experiments on the MFQE 2.0 dataset. It contains totally 126 sequences videos with multi resolutions: SIF(352 × 240)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.129 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Overall comparison for \\(\\Delta\\) PSNR(dB), \\(\\Delta\\) SSIM(\\(x\\))\\(^{-2}\\)), Latency(ms) and FPS over test sequences of MFQE 2.0 Dataset at QP=37. All methods were retested on a NVIDIA RTX 3090 in frame-to-frame manner. Video resolution: Class A(2560\\times1600), Class B(1980\\times1080), Class C(832\\times480), Class D(416\\times240), Class E(1280\\times720)." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.213, + 0.789, + 0.387 + ], + "angle": 0, + "content": "
ApproachHigh-Latency VQE MethodsLow-Latency VQE Methods
MFQE 1.0 [31]PSTQE [7]Li et al. [17]DS-CNN [29]DnCNN [34]DCAD [26]STLVQE (Ours)
MetricsΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%
ATraffic0.50/0.901120.10.64/1.045.6K0.29/0.60385.50.29/0.60374.60.24/0.57294.90.31/0.67153.70.41/0.9496.3
PeopleOnStreet0.80/1.371.00.48/0.922.60.42/0.852.70.41/0.823.40.50/0.956.50.68/1.3510.4
BKimono0.50/1.13628.10.69/1.362.9K0.28/0.780.25/0.750.24/0.750.24/0.750.24/0.750.28/0.780.28/0.780.38/1.0548.6
ParkScene0.39/1.030.42/1.150.49/1.210.15/0.480.23/0.580.15/0.500.14/0.500.14/0.500.16/0.500.16/0.500.16/0.500.33/0.8020.6
Cactus0.44/0.882.00.62/1.150.25/0.480.25/0.480.24/0.580.20/0.480.20/0.380.26/0.580.28/0.500.28/0.500.43/0.8548.6
BQTerrace0.27/0.480.50/0.870.50/0.870.30/0.680.30/0.680.26/0.480.25/0.580.25/0.580.28/0.650.31/0.680.31/0.680.42/0.8320.6
BasketballDrive0.41/0.800.60/1.040.60/1.040.30/0.680.30/0.680.28/0.650.25/0.650.25/0.650.28/0.650.28/0.650.28/0.650.39/0.9411.5
CRaceHorses0.34/0.55229.80.40/0.880.28/0.650.33/0.880.27/0.630.27/0.630.25/0.650.28/0.680.26/0.680.34/0.880.50/1.2587.0
BQMall0.51/1.03229.80.74/1.44611.60.33/0.8839.50.33/0.8036.70.28/0.6826.30.34/0.8814.30.34/1.15
PartyScene0.22/0.7310.40.51/1.461.80.13/0.4525.30.17/0.5827.20.13/0.4838.00.16/0.4869.90.37/1.37
BasketballDrill0.48/0.901.66/1.270.66/1.270.38/0.880.38/0.880.35/0.680.33/0.680.33/0.680.39/0.780.39/0.780.39/0.780.44/1.089.1
DRaceHorses0.51/1.13164.30.60/1.44203.70.33/0.830.32/0.750.32/0.750.31/0.730.31/0.730.32/0.830.32/0.750.47/1.059.1
BQSquare-0.01/0.15164.30.79/1.140.09/0.250.09/0.250.20/0.380.20/0.380.13/0.180.13/0.187.00.20/0.380.35/1.37109.9
BlowingBubbles0.39/1.2032.20.62/1.950.21/0.680.21/0.680.23/0.680.23/0.680.18/0.580.18/0.58142.90.22/0.65285.70.37/1.23
BasketballPass0.63/1.380.85/1.750.85/1.750.34/0.850.34/0.780.34/0.780.34/0.780.31/0.750.35/0.850.35/0.850.35/0.850.48/1.28128
EFourPeople0.66/0.85345.40.95/1.121.3K0.45/0.700.46/0.700.46/0.700.39/0.600.42/0.600.51/0.780.51/0.780.69/1.0224.7
Johnny0.55/0.554.70.75/0.850.850.40/0.6087.40.38/0.4081.60.32/0.4064.80.41/0.500.93/0.6440.5
KristenAndSara0.66/0.750.66/0.750.93/0.910.49/0.680.49/0.680.48/0.600.42/0.600.42/0.600.52/0.700.52/0.700.52/0.700.74/0.84128
Average0.46/0.88-0.69/1.25-0.30/0.66-0.30/0.63-0.26/0.58-0.32/0.67-0.49/1.03-
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.79, + 0.492 + ], + "angle": 0, + "content": "CIF(352 × 288), NTSC(720 × 486), 4CIF(704 × 576), 240P(416 × 240), 360P(640 × 360), 480P(832 × 280), 720P(1280 × 720), 1080P(1920 × 1080) and WQXGA (2560 × 1600). Same as [10], the 108 videos in the dataset are used as the training set and 18 videos are used as the test set. All videos were compressed by HM16.5 software in LDP mode with QP values of 22, 27, 32, 37 and 42." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.51, + 0.458, + 0.525 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.789, + 0.653 + ], + "angle": 0, + "content": "Our proposed network is implemented based on PyTorch framework and MMEditing toolbox [23]. During the training phase, we set batch size to 8 with \\(180 \\times 180\\) clips from the raw videos and the corresponding compressed videos as training samples. The random flip and rotation augmentation are utilized to fully exploit the information in the samples. The Adam optimizer [14] and Cosine Annealing scheme [19] are utilized for optimizing. The total iterations numbers and initial learning rate of the first and second stages is set to \\(200\\mathrm{K}\\), \\(1 \\times 10^{-4}\\) and \\(100\\mathrm{K}\\), \\(1 \\times 10^{-6}\\). Our network is trained on 8 Telsa V100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.789, + 0.775 + ], + "angle": 0, + "content": "During the inference phase, ST-LUTs will replace the corresponding convolutional layers. To evaluate the model's performance in an online setting, we conduct inference on the test set frame-by-frame. Following the same settings as [10], we report quality enhancement on Y-channel in YUV/YCbCr space. The Peak Singal-to-Noise Ratio(ΔPSNR) and Structural Similarity(ΔSSIM) will be used as metrics to judge the model's enhancement capability. We further evaluate the latency(LT) and frame per second(FPS) of the models which is crucial to the online-VQE:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.784, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\nL T = T _ {\\text {p r o c e s s}} + T _ {\\text {w a i t}}, \\quad T _ {\\text {w a i t}} = N _ {f} / F P S, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\( T_{process} \\) refers to the single-frame inference time and \\( T_{wait} \\) refers to the future frames waiting time required for some video based methods, \\( N_f \\) represents" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 2: Latency on 720P video and averaged \\(\\Delta\\)PSNR/ \\(\\Delta\\)SSIM comparison on the test sequences of MFQE 2.0 Dataset at different QP values. FV-Required means the full compressed video is required for enhancement, thus latency cannot be calculated." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.785, + 0.342 + ], + "angle": 0, + "content": "
MethodTypeLT(ms)StorageGFLOPsQP37QP42QP32QP27QP22
MFQEv1 [31]Bi-MF345.41788K-0.46/0.880.44/1.300.43/0.580.40/0.340.31/0.19
MFQEv2 [10]Bi-MF320.1255K7880.56/1.090.59/1.650.52/0.680.49/0.420.46/0.27
STDF [6]Bi-MF85.6487K2930.65/1.18-/-0.64/0.770.59/0.470.51/0.27
PSTQE [7]Bi-MF1249.8218K6790.69/1.250.69/1.860.67/0.830.63/0.520.55/0.29
RFDA [36]Bi-MF312.81250K-0.91/1.620.82/2.200.87/1.070.82/0.680.76/0.42
BasicVSR++ [2]Bi-MFFV-Required7300K4230.95/1.80-/-0.89/1.25-/--/-
RNAN [35]SF4329.48960K69790.44/0.95-/-0.41/0.62-/--/-
ARCNN [8]SF25.1118K1080.23/0.450.29/0.960.18/0.190.18/0.140.14/0.08
DnCNN [34]SF64.8558K5160.26/0.580.22/0.770.26/0.350.27/0.240.29/0.18
DS-CNN [29]SF81.6948K4370.30/0.630.31/1.010.27/0.380.27/0.230.25/0.15
Li et al. [17]SF87.4464K4340.30/0.660.32/1.050.28/0.370.30/0.280.30/0.19
DCAD [26]SF34.3299K2750.32/0.670.32/1.090.32/0.440.32/0.300.31/0.19
STLVQEUni-MF24.7398K+228M980.49/1.030.50/1.600.45/0.640.41/0.400.34/0.24
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.371, + 0.788, + 0.418 + ], + "angle": 0, + "content": "the number of future frames used. Note that the LT metric ignores the frame reception and decoding time, which is fair for different methods, and we set \\( FPS = 30 \\) in Eq. 11 as the general online video standard to facilitate comparison." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.617, + 0.455 + ], + "angle": 0, + "content": "4.3 Comparison with State of the art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.674 + ], + "angle": 0, + "content": "In this subsection, we present the results of our extensive experiments on various QP compressed videos in Tab. 1 and Tab. 2. Our proposed STLVQE network is compared with some widely used or state-of-art single-frame methods [8, 17, 26, 29, 34, 35], as well as some efficient multi-frame methods [6, 7, 10, 31]. In online environment where video is streamed in real-time, Bi-direction Multi-Frame (Bi-MF) method necessitates waiting for future frames that have not yet been received in current time. This waiting period introduces uncontrollable latency to the stream. Consequently, as outlined in previous online works [18, 22, 32], the Bi-MF method can not be directly applied to online scenarios and is predominantly utilized for offline processing. As a result, both Bi-MF and slow single frame methods \\((>100\\mathrm{ms})\\) are placed in the first part in Tab. 1 and Tab. 2 for performance comparison, and demonstrate our favorable trade-off between performance and inference speed. All experiments were conducted on an NVIDIA RTX 3090 using a frame-to-frame style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Latency and FPS. Tab. 1 presents the latency and FPS metrics for each VQE approach. We note that while low-latency VQE methods can process each frame at a lower latency and reduce video delay, they often struggle to handle high-resolution videos in real-time, which can lead to stuttering and reduced smoothness. Among the methods listed, only STLVQE can achieve FPS above 40 on 720P resolution videos, making it suitable for most online scenarios. Compared to the second fastest method DCAD, STLVQE's processing time on 720p video is 9.6 milliseconds faster and outperforms it by a considerable margin, improving PSNR and SSIM metrics by 0.17dB and \\(0.36 \\times 10^{-2}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Tab. 2 shows the latency of some large-scale high-latency temporal VQE methods. In contrast, high-latency VQE methods require waiting for future" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.129 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Ablation study of Temporal Cache(TC), Reference Frame Window(RFW), Deform Operation(DF), and Preliminary Complement Residual(PC). The latter three sections share the features extracted by MAFE." + }, + { + "type": "table", + "bbox": [ + 0.345, + 0.199, + 0.661, + 0.302 + ], + "angle": 0, + "content": "
RFWTCDFPCΔP/ΔSTimeParams(K)
✓✓0.06/0.0110.85<1
✓✓✓0.31/0.5121.85352
✓✓✓0.39/0.8720.12149
✓✓✓0.49/1.0345.61398
✓✓✓0.46/0.9824.32398
✓✓✓✓0.49/1.0324.65398
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.448 + ], + "angle": 0, + "content": "frames or taking a long time for inference, which can cause uncontrollable video delay. The latency of the two high-latency methods in Tab. 1 we evaluated exceeds \\(400\\mathrm{ms}\\), much higher than the latency caused by STLVQE. Moreover, when considering only the FPS metric, STLVQE achieves a significantly faster speed than the single-frame VQE method RNAN(FPS = 0.2) by a factor of 175 on 720P video. Compared with MFQE 1.0, our method is 8.6 times faster with a better enhancement performance. Additionally, STLVQE has a more significant speed advantage on higher resolution videos such as 1080P and 2K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.787, + 0.629 + ], + "angle": 0, + "content": "Enhancement performance. In Tab. 1 and Tab. 2, we present the performance of STLVQE compared to other commonly used or sota VQE methods for each video at \\(\\mathrm{QP} = 37\\) and for different QP compressed videos, respectively. Our method outperforms all the listed low-latency methods. In comparison with the high-latency VQE method, STLVQE reaches higher level with RNAN at \\(\\mathrm{QP} = 37\\) and \\(\\mathrm{QP} = 32\\), with only \\(2.6\\%\\) and \\(0.6\\%\\) of RNAN's number of parameters and latency time (720P), respectively. The \\(\\Delta\\)SSIM of STLVQE is higher than that of the MFQE method for all five QPs, indicating that our method produces an enhanced image with a more stable image structure. The specific visualization results and analysis can be found in Qualitative Comparison section. STDF is a robust VQE method, but we believe that a stable latency and higher frame rate are more important for enhancing the user's viewing experience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.721 + ], + "angle": 0, + "content": "Memory Size & GPU Consumption. As shown in Tab. 2, STLVQE has space occupation of 398KB for the model and 228MB for the LUT (utilizing less than \\(5\\%\\) memory of a low end GPU like GTX 1060), achieving real-time processing capability. Since most deep networks are replaced with ST-LUTs, inferencing STLVQE only needs 98 GFLOPs on a 720P frame. Note that GPU computing unit are more expensive than the memory." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.391, + 0.758 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.81 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed new VQE framework, we perform ablation experiments and specific analysis on three components. The results of the runtime per frame is computed on 720P videos at \\(\\mathrm{QP} = 37\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of the propagation module. The results are shown in the 5th and 6th row of Tab. 3. We found that selecting frames based on QP val-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: Ablation study of STLUTs and the CNN enhancement network. The runtime is computed on a 720P video." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.185, + 0.687, + 0.274 + ], + "angle": 0, + "content": "
Enhancement ModuleΔP/ΔSTime(ms)Params(K)
S-LUT0.39/0.8323.52398
S-LUT + T-LUT-10.45/1.0224.12398
ST-LUTs0.47/1.0224.65398
ST-LUTs (on x̂i)0.49/1.0324.65398
CNN Network0.50/1.08147.02442
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.788, + 0.427 + ], + "angle": 0, + "content": "ues in RFW can bring a 0.03dB \\(\\Delta\\)PSNR performance improvement compared to randomly selecting frames in the RFW, without adding parameters to the network and only increasing the inference time by \\(0.33\\mathrm{ms}\\) for comparing QP values of each frame. Furthermore, as shown in rows 4 and 6 of Tab. 3, the use of the TC structure can significantly reduce the inference time by \\(21.0~\\mathrm{ms}\\), which is consumed by the repeated computation of the reference frame information. We achieve this reduction in time by merely occupying the GPU memory with information of several frames in the RFW." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Effectiveness of the alignment module. We mainly validate the effect of the deform operation of the reference fram in the Alignment module. As shown in the first row of Tab. 3, when only using the ST-LUTs structure and discard the whole alignment module and PC operation, the result only has a \\(0.06\\mathrm{dB}\\) PSNR improvement compared to HEVC video. When the Deform operation is applied, the network can capture more accurate temporal information, and the \\(\\Delta\\)PSNR reaches \\(0.31~\\mathrm{dB}\\). Due to the lightweight convolutional layers, the alignment module only adds \\(11.6~\\mathrm{ms}\\) of inference time to the network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Effectiveness of the enhancement module. As shown in row 3 of Tab. 3, using PC without the Deform operation achieves a significant \\(\\Delta\\)PSNR of 0.39 dB, demonstrating the effectiveness of PC in providing initial compensation for spatial information and extending the receptive field, ultimately enhancing the performance of ST-LUTs. Finally, after combining two parts, the network achieves 0.49 dB \\(\\Delta\\)PSNR. PC alone contributes 0.18 dB to the network and adds only 2.16 ms of inference time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Tab. 4 provides an analysis of the impact of each LUT in the enhancement module. In the first three rows of the table, we attempted to enhance the original image \\( x_{i} \\) with the results of different LUT combinations. Results in the first and second rows show that using only the S-LUT with T-LUT-1 does not fully improve the network due to T-LUT-1 not utilizing all pixels within a kernel when learning the alignment offset. However, when T-LUT-2 is used for network training along with T-LUT-1, the reference frame can be better aligned, resulting in a \\( \\Delta \\)PSNR improvement of 0.08 dB. The 4th row shows the enhancing result on \\( \\tilde{x}_{i} \\), reducing the learning difficulty of the ST-LUTs, which improving the performance by 0.02 dB. Compared to the CNN layers replaced by LUTs, ST-LUTs sacrifice only 0.01 dB \\( \\Delta \\)PSNR performance while reducing the number of parameters by 44K and the inference time by 122.37 ms." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.787, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results at QP 37. The STLVQE method successfully enhances the left leg of the athlete, the face of the old man, the railings of the bridge and the eyebrow of the woman in each frame, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.496 + ], + "angle": 0, + "content": "The above ablation studies have sufficiently demonstrated that the MAFE structure of shared features in STLVQE with the redesigned three new modules can sufficiently reduce the computational redundancy of the network and obtain satisfactory inference speed-performance trade-offs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.459, + 0.53 + ], + "angle": 0, + "content": "4.5 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Fig. 5 provides 4 examples of the qualitative results of our network on compressed video at \\( QP = 37 \\). The quality of these images deteriorates significantly in compressed scenes, showing a variety of image defects (e.g. blurring legs, missing bridge railings). With the help of temporal information, STLVQE achieves better visual results than single-frame methods with significantly reduced runtime. Compared to large-scale multi-frame methods like MFQE 2.0, STLVQE can achieve competitive capabilities in detail and structural restoration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.36, + 0.677 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a lightweight and rapid method called STLVQE based on the constraints of online video quality enhancement tasks. STLVQE greatly reduce the computation redundancy and uses a Spatial-Temporal Look-up Table structure to extract spatio-temporal information from the video with minimal time consumption. The extensive experiments prove that our proposed method can outperform most low-latency VQE methods, achieve competitive performance with high-latency VQE methods which cannot be directly utilized on Online-VQE tasks. Furthermore, STLVQE achieves real-time processing of 720P resolution videos, demonstrating a good trade-off between enhancement performance and inference speed." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.264 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Fund of China (62076184, 61976158, 61976160, 62076182, 62276190), in part by Fundamental Research Funds for the Central Universities and State Key Laboratory of Integrated Services Networks (Xidian University), in part by Shanghai Innovation Action Project of Science and Technology (20511100700) and Shanghai Natural Science Foundation (22ZR1466700)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.283, + 0.323, + 0.299 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.309, + 0.787, + 0.365 + ], + "angle": 0, + "content": "1. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.787, + 0.42 + ], + "angle": 0, + "content": "2. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022) 2, 3, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.472 + ], + "angle": 0, + "content": "3. Charbonnier, P., Blanc-Feraud, L., Aubert, G., Barlaud, M.: Two deterministic half-quadratic regularization algorithms for computed imaging. In: Proceedings of 1st international conference on image processing. vol. 2, pp. 168-172. IEEE (1994) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.787, + 0.529 + ], + "angle": 0, + "content": "4. Chen, H., He, X., Qing, L., Xiong, S., Nguyen, T.Q.: Dpw-sdnet: Dual pixel-wavelet domain deep cnns for soft decoding of JPEG-compressed images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 711-720 (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.787, + 0.583 + ], + "angle": 0, + "content": "5. Dai, Y., Liu, D., Wu, F.: A convolutional neural network approach for post-processing in hevc intra coding. In: MultiMedia Modeling: 23rd International Conference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23. pp. 28-39. Springer (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.787, + 0.624 + ], + "angle": 0, + "content": "6. Deng, J., Wang, L., Pu, S., Zhuo, C.: Spatio-temporal deformable convolution for compressed video quality enhancement. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 10696-10703 (2020) 2, 3, 4, 5, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.787, + 0.663 + ], + "angle": 0, + "content": "7. Ding, Q., Shen, L., Yu, L., Yang, H., Xu, M.: Patch-wise spatial-temporal quality enhancement for hevc compressed video. IEEE Transactions on Image Processing 30, 6459-6472 (2021) 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "8. Dong, C., Deng, Y., Loy, C.C., Tang, X.: Compression artifacts reduction by a deep convolutional network. In: Proceedings of the IEEE international conference on computer vision. pp. 576-584 (2015) 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.787, + 0.731 + ], + "angle": 0, + "content": "9. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. arXiv preprint arXiv:1902.08153 (2019) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.732, + 0.787, + 0.784 + ], + "angle": 0, + "content": "10. Guan, Z., Xing, Q., Xu, M., Yang, R., Liu, T., Wang, Z.: Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE transactions on pattern analysis and machine intelligence 43(3), 949-963 (2019) 2, 3, 4, 5, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Guo, J., Chao, H.: Building dual-domain representations for compression artifacts reduction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 628-644. Springer (2016) 3" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.309, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Qu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.202 + ], + "angle": 0, + "content": "12. Jin, Z., An, P., Yang, C., Shen, L.: Quality enhancement for intra frame coding via cnns: An adversarial approach. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1368-1372. IEEE (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "13. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 691-700 (2021) 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.787, + 0.271 + ], + "angle": 0, + "content": "14. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.325 + ], + "angle": 0, + "content": "15. Kondratyuk, D., Yuan, L., Li, Y., Zhang, L., Tan, M., Brown, M., Gong, B.: Movinets: Mobile video networks for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16020-16030 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "16. Li, J., Chen, C., Cheng, Z., Xiong, Z.: Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVIII. pp. 238-256. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "17. Li, K., Bare, B., Yan, B.: An efficient deep convolutional neural networks model for compressed image deblocking. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 1320-1325. IEEE (2017) 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "18. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.787, + 0.487 + ], + "angle": 0, + "content": "19. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.487, + 0.787, + 0.528 + ], + "angle": 0, + "content": "20. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep kalman filtering network for video compression artifact reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.528, + 0.787, + 0.582 + ], + "angle": 0, + "content": "21. Ma, C., Zhang, J., Zhou, J., Lu, J.: Learning series-parallel lookup tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 305-321. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.582, + 0.787, + 0.636 + ], + "angle": 0, + "content": "22. Maggioni, M., Huang, Y., Li, C., Xiao, S., Fu, Z., Song, F.: Efficient multistage video denoising with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3466-3475 (2021) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.636, + 0.787, + 0.664 + ], + "angle": 0, + "content": "23. MMEditing Contributors: MMEditing: OpenMMLab image and video editing toolbox. https://github.com/open-mmlab/mmediting (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "24. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.787, + 0.757 + ], + "angle": 0, + "content": "25. Wang, T., Li, Y., Peng, J., Ma, Y., Wang, X., Song, F., Yan, Y.: Real-time image enhancer via learnable spatial-aware 3d lookup tables. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2471-2480 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.787, + 0.8 + ], + "angle": 0, + "content": "26. Wang, T., Chen, M., Chao, H.: A novel deep learning-based method of improving coding efficiency from the decoder-end for hevc. In: 2017 Data Compression Conference (DCC). pp. 410-419. IEEE (2017) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "27. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003) 2" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "28. Xu, Y., Gao, L., Tian, K., Zhou, S., Sun, H.: Non-local convlstm for video compression artifact reduction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7043-7052 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "29. Yang, R., Xu, M., Liu, T., Wang, Z., Guan, Z.: Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology 29(7), 2039-2054 (2018) 2, 3, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.273 + ], + "angle": 0, + "content": "30. Yang, R., Xu, M., Wang, Z.: Decoder-side hevc quality enhancement with scalable convolutional neural network. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 817-822. IEEE (2017) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "31. Yang, R., Xu, M., Wang, Z., Li, T.: Multi-frame quality enhancement for compressed video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6664-6673 (2018) 2, 3, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.356 + ], + "angle": 0, + "content": "32. Yin, G., Qu, Z., Jiang, X., Jiang, S., Han, Z., Zheng, N., Yang, H., Liu, X., Yang, Y., Li, D., et al.: Online streaming video super-resolution with convolutional lookup table. IEEE Transactions on Image Processing 33, 2305-2317 (2024) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.398 + ], + "angle": 0, + "content": "33. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "34. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.467 + ], + "angle": 0, + "content": "35. Zhang, Y., Li, K., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082 (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "36. Zhao, M., Xu, Y., Zhou, S.: Recursive fusion and deformable spatiotemporal attention for video compression artifact reduction. In: Proceedings of the 29th ACM international conference on multimedia. pp. 5646-5654 (2021) 2, 3, 4, 5, 9, 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_origin.pdf b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4039e4bc760b0b2f54b2523e25177c3012ac7b2a --- /dev/null +++ b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/fc3a51b4-06f4-47db-9ed9-b257d7d9d465_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcf231a4e61224f945b6af39615e9290c2c941ae2a83954887519996f2a5baf6 +size 942568 diff --git a/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/full.md b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02506887949e6460c23435b932b926998476447c --- /dev/null +++ b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/full.md @@ -0,0 +1,291 @@ +# Online Video Quality Enhancement with Spatial-Temporal Look-up Tables + +Zefan $\mathrm{Qu}^{1}$ , Xinyang Jiang $^{\dagger 2}$ , Yifan Yang $^{2}$ , Dongsheng Li $^{2}$ , and Cairong Zhao $^{\dagger 1}$ + +$^{1}$ Tongji University, Shanghai, China + +{qzf, zhaocairong}@tongji.edu.cn + +2 Microsoft Research Asia, Shanghai, China + +{xinyangjiang,yifanyang,dongsheng.li}@microsoft.com + +Abstract. Low latency rates are crucial for online video-based applications, such as video conferencing and cloud gaming, which make improving video quality in online scenarios increasingly important. However, existing quality enhancement methods are limited by slow inference speed and the requirement for temporal information contained in future frames, making it challenging to deploy them directly in online tasks. In this paper, we propose a novel method, STLVQE, specifically designed to address the rarely studied online video quality enhancement (Online-VQE) problem. Our STLVQE designs a new VQE framework which contains a Module-Agnostic Feature Extractor that greatly reduces the redundant computations and redesign the propagation, alignment, and enhancement module of the network. A Spatial-Temporal Look-up Tables (STL) is proposed, which extracts spatial-temporal information in videos while saving substantial inference time. To the best of our knowledge, we are the first to exploit the LUT structure to extract temporal information in video tasks. Extensive experiments on the MFQE 2.0 dataset demonstrate our STLVQE achieves a satisfactory performance-speed trade-off. + +Keywords: Online Video Quality Enhancement $\cdot$ Look-up Table $\cdot$ Video Compression Artifact Reduction + +# 1 Introduction + +With the rapid development of social media, users are increasingly seeking high-quality and real-time video content. In recent times, video-based applications such as video conferencing, webcasting, and cloud gaming have increasingly demanded low latency requirements. This has resulted in higher video compression rates, which can cause a noticeable decrease in video quality and an overall reduction in the Quality of Experience (QoE). As a result, online Video Quality + +![](images/46cbabc8dca024c6e69d705a1453a69b132be3d152381f10cf3e3a5a40511c68.jpg) +Fig. 1: Comparison of $\Delta$ PSNR/ $\Delta$ SSIM, average runtime per frame(720P video) and parameters on MFQE 2.0 test set at $\mathrm{QP} = 37$ . Our STLVQE method achieves a great trade-off in enhancement performance and inference speed. + +![](images/4d95bcd21a4980bdc6cf6629ea7404945634c0bd848d02a5f936420e19366356.jpg) + +Enhancement (Online-VQE), which improves the quality of real-time streaming video under online conditions and minimizing video defects caused by compression algorithms (e.g., H.264/AVC [27] and H.265/HEVC [24]) has become a significant problem. However, Online-VQE poses two main challenges compared to the conventional offline VQE approach. Firstly, it needs to process high-resolution videos in real-time to ensure a smooth viewing experience. Secondly, online video processing techniques [15, 18, 22] can only rely on the current and previous frames for inference due to the uncontrolled latency caused by waiting for future frames, which can result in overall video delay. + +In recent years, numerous neural network-based techniques for enhancing the quality of compressed video have been developed, resulting in significant advancements in this field. These methods can be broadly classified into two categories, namely single-image based VQE method [4, 12, 29, 30] and multi-frame based VQE method [2, 6, 10, 20, 36]. Single-image based approaches are relatively faster in inference speed, but they ignore the significance of temporal information in video [31]. On the other hand, multi-frame based VQE achieves state-of-the-art performance by utilizing temporal information. However, most of these approaches are not suitable for online scenarios because they rely on subsequent frames as reference frames and have unsatisfactory inference speeds. Therefore, developing a new online-VQE method that satisfies the aforementioned two constraints while effectively leveraging temporal information is essential. + +Reasons that limit the speed and latency of the above temporal VQE methods can be summarized in the following two points: First, a high level of computational redundancy exists within these methods, and these homogeneous computations increase the network's inference time. Second, most of these methods employ complex and computationally intensive models, which sacrificing inference speed and latency for some of the enhancement capabilities. In this work, a more applicable VQE method for online scenarios are proposed by advancing these two perspectives. + +As discussed in previous works [1], modern VQE methods exploit temporal information by constructing a neural network containing three main components: the propagation, alignment and enhancement module. In all three modules, similar feature extraction operations on the original image may be involved, e.g., the reference frame selection [10, 31] in propagation module, the optical flow calculation [1, 2] in alignment module and the final result obtaining in the enhancement module [6, 10]. Most of the state-of-the-art methods adopts distinct feature extractors in different components and obtaining latent feature for each component separately, causing large computation redundancy. In this paper, we design a new VQE network framework containing a Module-Agnostic Feature Extractor (MAFE), where the extracted latent feature are uniformly share across three modules. This design allows the network to greatly reduce the redundant feature computation steps and speed up the inference of the network. + +When processing high-resolution videos in VQE task, the complex network of the sota method will greatly increase the computation times of the network, thus affecting the online deployment of the method. Therefore, we in our proposed framework: (1) Redesign the structure of the three modules with a light-weight convolutional structure. (2) Propose a Spatial-Temporal Look-up Tables (ST-LUTs) that extend conventional LUTs [13] to conduct queries on both spatial and temporal dimensions, which allows the LUTs to simultaneously capture information in both temporal and spatial dimensions, facilitating efficient extraction of temporal information. To the best of our knowledge, ST-LUTs is the first approach to employ a LUT structure for the extraction of temporal information. + +The major contributions are summarized as follows: (1) We propose the first online-VQE method STLVQE that achieves real-time process speed, which reduces redundant feature extraction computations while redesigning the propagation, alignment and enhancement modules of the temporal VQE network. (2) A Spatial-Temporal Look-up Tables structure is first time proposed for fully exploiting the temporal and spatial information in the video. (3) Extensive experiments on MFQE 2.0 dataset demonstrate that our approach achieves a comparable speed-performance trade-off. We also test most of the previous VQE methods on prevalent GPU to provide performance and speed benchmarks, facilitating future research in this area. + +# 2 Related Works + +Compressed Video Quality Enhancement. Recently, numerous works have focused on the degradation of the network video quality due to the lossy video compression algorithms. Existing Compressed VQE work can be broadly classified into two categories, which are single-image based method [4, 5, 8, 11, 12, 17, 30, 34] and multi-frame based method [2, 6, 7, 10, 20, 28, 31, 36]. + +For the single-image based methods, some of the early work [4,8,11,17,29,34] are mainly used to enhance the quality of JPEG compressed images, which can also be applied directly to compressed videos to process each frame independently. [5, 12, 30] utilize certain characteristics of the video compression algo + +rhythm or compression parameters to improve the video quality. For example, DS-CNN [30] and QE-CNN [29] use distinct models to enhance the video processed by different compression coding. + +For the multi-frame based methods, MFQE 1.0 [31] and MFQE 2.0 [10] were proposed to use two adjacent high quality frames(PQF) and compute the optical flow between them to enhance the current frame. STDF [6] uses deformation convolution for an implicit alignment. RFDA [36] extracts long-range temporal information from the video via the recursive fusion module. BasicVSR [1] and BasicVSR++ [2] utilize a bidirectional propagation network and a new alignment scheme is proposed in BasicVSR++. All these methods demonstrate the significance of temporal information for solving the compressed VQE problem. + +As far as we know, there is still no compression VQE method for online scenario that has been proposed so far. The above-mentioned methods cannot be directly applied to webcasting environments due to the limitations of the two online restriction mentioned above. + +Look-up Table. Look-up Tables (LUTs) have proven to be highly effective in reducing computational power and time consumption during the inference phase of algorithms by replacing complex computation with fast storage space retrieval operations. However, the use of LUTs directly is not always flexible as it requires manual adjustment by experts to accurately process images. To overcome this limitation, recent attempts have been made to improve LUTs. [25, 33] combine multiple LUTs and learn the weights through neural networks to summing up these LUT results. SR-LUT [13] replaces a small receptive field neural network with a 4D-LUT and achieves good speed-performance trade-offs on SR tasks. SP-LUT [21] and MuLUT [16] makes improvements to the SR-LUT by expanding its receptive field through parallel and serial multiple LUT tables. In this work, we utilize LUTs to extract information in both temporal and spatial dimensions, bringing its advantages to the online VQE task. + +# 3 Method + +# 3.1 Overview + +Given a stream of compressed frames without an explicit ending $\{x_0, \ldots, x_i, \ldots\}$ , where $x_i \in \mathbb{R}^{C \times H \times W}$ with input channel $C$ , height $H$ and width $W$ , we sequentially enhance each compressed frame to obtain the target sequence $\{y_0, \ldots, y_i, \ldots\}$ . Due to the low latency requirement of Online-VQE, obtaining information from future frames to enhance $x_i$ is unacceptable. We only use past and current frames $\{x_0, x_1, \ldots, x_i\}$ to predict $y_i$ , reducing the video's delay time. Additionally, to avoid stuttering issues, each frame's inference speed must be fast enough. + +The overall flow of our proposed network during inference period is illustrated in Fig. 2 (The difference between the training and the inference phase is illustrated in Subsec. 3.4 and the figure in appendix). First, a Module-Agnostic Feature Extractor (MAFE) is utilized to extract the feature $F_{\text{eat}_i}$ of the current enhancing frame $x_i$ as follows, + +$$ +F e a t _ {i} = M A F E \left(x _ {i}\right), \tag {1} +$$ + +![](images/26385d2ba90d424f1d876ac9095939b18e6358def4e14f07edf284979a24d154.jpg) +Fig. 2: The framework of STLVQE (inference phase), which consists of a Module-Agnostic Feature Extractor and three main modules. In the inference phase, the propagation module selects the reference frame and accesses the relevant information, which is sent to the alignment module for temporal alignment, and finally the aligned reference and enhancing frames are input to the enhancement module for the final results. + +where the $\text{Feat}_i$ will be shared by the three subsequent modules, thus reducing the multiple time-consuming extraction steps and fully exploiting the latent frame feature. + +Specifically, in the propagation module, $Feat_{i}$ is stored in a structure called Temporal Cache, which will provide the temporal information in subsequent modules based on the selection of the reference frames $x_{ref} = \{x_{ref1}, x_{ref2}\}$ without external computation. In the alignment module, $Feat_{i}$ is concatenated with $Feat_{ref}$ , the reference frame feature obtained from the propagation module, to deform $x_{ref}$ with a lightweight offset predictor. In the enhancement module, $Feat_{i}$ is used to perform a spatial complement, which fully exploits the extracted features and augments the ST-LUTs by increasing its receptive fields. + +Under our proposed new framework, only once feature extraction computation for a single frame is required for enhancing each frame. The extracted features are simultaneously computed with reference frame alignment and preliminary compensation in the current frame space. + +# 3.2 The Propagation Module + +The propagation module in the VQE task is responsible for extracting and transmitting temporal information within the video. Previous methods have explored various approaches, including designing additional structures [10] and increasing the number of selected reference frames [6, 36] to improve network performance, which greatly increase inference time and computational costs of the model. The architecture of the propagation module is illustrated in Fig. 3. The compression configuration parameter quantization parameter (QP) value is utilized as an indicator to select fewer but higher quality frames among multiple candidate frames without extra computation. Specifically, a Reference Frame Window (RFW) which contains $W$ frames $\{x_{i-W}\dots x_{i-1}\}$ is set. Then, two frames with the smallest QP values are selected as $x_{ref}$ for the subsequent module. + +![](images/4dee20de2c65bd311fda64ee50378b0e747581a802da31f2b17234971fa8794b.jpg) +Fig. 3: Reference Frame Window and Temporal Cache in the propagation module. + +To mitigate the issue of additional feature extraction operations in previous works, we introduce a cache-like structure known as the Temporal Cache (TC). The TC stores reusable features of past frames that were already extracted by the MAFE in previous process, thereby eliminating the need to re-extract features for each subsequent frame. In particular, when enhancing the i-th frame $x_{i}$ , STLVQE caches the intermediate feature map $Fea t_{i}$ extracted by the alignment module, as well as the enhanced frame, in the information cache (TC). During reference frame selection, the features and enhanced result of $x_{ref1}$ and $x_{ref2}$ are directly retrieved from TC without extra feature extraction process. + +# 3.3 The Alignment Module + +The alignment module is a crucial component in video processing tasks that handles spatial misalignments between frames, which plays a vital role in effectively utilizing the temporal information embedded in the video, especially when the video content is subject to large motion. STLVQE includes a lightweight offset predictor to replace the optical flow or non-local computation of most VQE methods, which significantly reduce the computation cost. + +The offset predictor is responsible for predicting spatial offsets which align correlated pixels in the reference with pixels in $\tilde{x}_i$ . As introduced in the last section, the latent features $Feat_{ref}$ and spatially compensated frames $\tilde{x}_{ref}$ at each processing step are stored in TC, which are retrieved by RFW to enhance future frames. Given the feature of a reference frame $\tilde{x}_{ref}$ retrieved from the TC, as well as the feature of the current frame $\tilde{x}_i$ , the offset predictor predicts 9 spatial offsets for each pixel in $\tilde{x}_i$ : + +$$ +o _ {r e f} = C _ {\text {o f f s e t}} \left(\left[ \operatorname {F e a t} _ {i}, \operatorname {F e a t} _ {r e f} ^ {d} \right]\right) \tag {2} +$$ + +where the matrix concatenation is denoted by $[\cdot]$ , and each offset value in $o_{ref}$ represents the position shift of a correlated pixel in $\tilde{x}_{ref}$ relative to a pixel in $\tilde{x}_i$ . + +The temporal corresponding pixels in the reference frames $x_{ref}$ are spatially aligned with the $3 \times 3$ pixel offsets $o_{ref}$ and bilinear interpolation. Each set of + +obtained correlated pixel values of a pixel in $\tilde{x}_i$ are re-arranged into a $3 \times 3$ patch, which are then stacked together to obtain the deformed reference frame denoted as $\tilde{x}_{ref}^d$ with the size of $N \times C \times 3H \times 3W$ . + +# 3.4 The Enhancement Module + +The enhancement module contains a spatial compensation network and a ST-LUTs network. To fully utilize the sharing latent feature $Feat_{i}$ , a spatial compensation network reuses $Feat_{i}$ to generate a spatial residual, which compensates the low quality frame $x_{i}$ to obtain preliminary spatial enhancement before ST-LUTs: + +$$ +\tilde {x} _ {i} = x _ {i} + C _ {\text {c o m p}} (\text {F e a t} _ {i}) \tag {3} +$$ + +where $C_{comp}$ are 3 shallow convolutional layers. Since the subsequent enhancement module takes the obtained compensated frame $\tilde{x}_i$ as input, spatial compensation network can boost the enhancement performance by expanding the receptive field of the ST-LUTs in the enhancement module. To match the resolution of $\tilde{x}_{ref}^{d}$ , $\tilde{x}_i$ is upsampled by repeating each pixel in $\tilde{x}_i$ to form a 3x3 patch, then each pixel in the up-sampled $\tilde{x}_i$ is correlated with the pixel at the same position in $\tilde{x}_{ref}^{d}$ . This is achieved by the unfold and fold operations with different strides as follows: + +$$ +\tilde {x} _ {i} ^ {u p} = \operatorname {F o l d} (\operatorname {U n f o l d} (\tilde {x} _ {i}, \text {s t r i d e} = 1), \text {s t r i d e} = 3), \tag {4} +$$ + +the up-sampled input frame $\tilde{x}_i^{up}$ and deformed reference frame $\tilde{x}_{ref}^d$ together serve as the input of the ST-LUTs network for the final quality enhancement. + +Instead of deep neural network, STLVQE adopts a novel Spatial-Temporal Look-up Tables in the enhancement module during inference to replace complex convolutional computation used in training stage with much faster direct memory access, and conducts table queries on both spatial and temporal perspective. + +Quantization. Due to the limited storage space of the LUT, the size of each pixel in the input compensated frame $\tilde{x}_i^{up}$ and the deformed frames $\tilde{x}_{ref}^d$ needs to be controlled to fit within a suitable range of bits. Because the range of values in $\tilde{x}_i^{up}$ and $\tilde{x}_{ref}^d$ can vary dynamically and the gradient needs to be passed through the quantization operation during end-to-end training, it's not practical to simply quantize the matrices using a fixed quantization hyper-parameter. The enhancement module adopts a Learnable Quantization layer (LSQ) [9] to quantize the input of the enhancement module as follows: + +$$ +\bar {x} = \left\lfloor \operatorname {c l i p} \left(\tilde {x} / s, - Q _ {N}, Q _ {P}\right) \right], \quad \hat {x} = \bar {x} \times s, \tag {5} +$$ + +where $\text{clip}(z, r_1, r_2)$ returns matrix $z$ with values below $r_1$ set to $r_1$ and values above $r_2$ set to $r_2$ , and the $\lfloor z \rfloor$ rounds $z$ to the nearest integer. The step size parameter $s$ can be optimized along with the network's training. $\overline{x}$ is the indexes used in the LUTs store and retrieval stages, and $\hat{x}$ is a quantized representation of the data with the same scale as $x$ . After quantization, the deep enhancement network and the corresponding LUTs can be constructed with the quantized pixel values as input. + +![](images/1e72c8caf4e5426fa770ce31b16d6670bfe8e9e5eb09e0024a0c6df5e3335f99.jpg) +Fig. 4: The specific schematic of our Spatial-Temporal Look-up Tables, which consists of two parts: Temporal-LUT (left side) and Spatial-LUT (right side). + +Deep Enhancement Network Training. In the training phase, a neural network based enhancement module is optimized, which takes the compensated frame $\hat{x}_i^{up}$ and reference frames $\hat{x}_{ref}^d$ as input and output the final enhancement results. To train an enhancement network that is suitable for LUT conversion, the receptive field size of the deep network typically cannot exceed 6 (Larger receptive fields will result in a higher storage occupancy, refer to the appendix for details). Therefore, each branch of the enhancement network, which will be replaced by ST-LUTs during the inference stage, consists of 1 convolutional layer with a kernel size of 4 or 6 and 10 convolutional layers with $1\times 1$ kernel. The specific design of the network will be illustrated in the appendix. + +ST-LUTs Construction & Replacement. We generate LUTs by permuting all input patches of the receptive field size, and storing all the input/output mappings of the trained deep enhancement network in mapping tables. During inference, we retrieve the outputs of the deep enhancement network directly from the LUTs, replacing slow and computationally-intensive convolution with faster memory access and retain the other components in STLVQE. Due to the input patch size restriction, conventional LUTs are not suitable for inter-frame queries in the temporal dimension. To address these issues, we firstly propose the STLUTs, which employ multiple LUTs to conduct parallel retrieval with different sets of query pixels in both spatial and temporal dimensions. + +Specifically, given the quantized compensated frames $\hat{x}_i^{up}$ and the aligned frames $\hat{x}_{ref}^d$ as input, the look-up tables are constructed as shown in Fig. 4, where two sets of LUTs named S-LUTs and T-LUTs are constructed for extracting spatial and temporal information respectively. + +For S-LUTs, a rotation ensemble operation [13] is used during the training process to further expand the LUT receptive field. As shown in Fig. 4, during training, input images are rotated by 0, 90, 180, and 270 degrees, and the results + +are reverse-rotated and summed up. This approach considers all pixels in a $3 \times 3$ receptive field simultaneously. The result of S-LUT $R_{i}^{S}$ is expressed as follows: + +$$ +R _ {i} ^ {S} = \frac {1}{4} \sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \left(f ^ {S} \left(R o t _ {j} \left(\hat {x} _ {i} ^ {u p}\right)\right)\right), \tag {6} +$$ + +where $Rot_{j}$ and $Rot_{j}^{-1}$ are the rotation and reserve-rotation operations, and $f^{S}$ is the spatial enhancement network or the S-LUTs retrieval. + +For T-LUTs, the pixels at the same location of the enhancing and reference frames will be strongly correlated after alignment, and jointly considering these values can make full use of the temporal information in the video. As shown on the left in Fig. 4, we establish two T-LUTs: T-LUT-1 take the same position pixels of $\hat{x}_i^{up}$ and $\hat{x}_{ref}^d$ as index with the rotation ensemble operation, while T-LUT-2 uses the four edge pixels of each kernel in the $\hat{x}_{ref}^d$ as indexes to retrieve the surrounding information of the target pixels. The result of T-LUTs $R_i^T$ can be expressed as follows: + +$$ +\begin{array}{l} R _ {i} ^ {T 1} = \frac {1}{4} \sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \left(f ^ {T 1} \left(R o t _ {j} \left(\left[ \hat {x} _ {i} ^ {u p}, \hat {x} _ {r e f} ^ {d} \right]\right)\right)\right), \tag {7} \\ R _ {i} ^ {T 2} = f ^ {T 2} (\hat {x} _ {r e f} ^ {d}), \quad R _ {i} ^ {T} = R _ {i} ^ {T 1} + R _ {i} ^ {T 2}, \\ \end{array} +$$ + +where $f^{T1}$ and $f^{T2}$ are the temporal enhancement network or the T-LUTs retrieval. Then we can acquire the results of ST-LUTs and get the final enhanced image $y_{i}$ as follows: + +$$ +R _ {i} = R _ {i} ^ {S} + R _ {i} ^ {T}, \quad y _ {i} = \tilde {x} _ {i} + R _ {i}. \tag {8} +$$ + +# 3.5 Training Scheme + +We perform the training of the STLVQE in two stages with an end-to-end fashion. Firstly, the Charbonnier Loss [3] is utilized to optimize the model as follows: + +$$ +L _ {c h a r} = \sqrt {\left(y _ {t} - \bar {y} _ {t}\right) ^ {2} + \epsilon}, \tag {9} +$$ + +where $y_{t}$ and $\overline{y_t}$ is the enhanced and ground truth frame at time $t$ , and $\epsilon$ is set to $10^{-6}$ in our paper. Secondly, to further enhance the enhancement capability, we perform fine-tuning using L2 loss with a smaller learning rate as follows: + +$$ +L _ {M S E} = \left\| \left(y _ {t} - \bar {y} _ {t}\right) ^ {2} \right\| _ {2} ^ {2}. \tag {10} +$$ + +# 4 Experiments + +# 4.1 Datasets + +Following [6, 10, 36], we conduct our experiments on the MFQE 2.0 dataset. It contains totally 126 sequences videos with multi resolutions: SIF(352 × 240), + +Table 1: Overall comparison for $\Delta$ PSNR(dB), $\Delta$ SSIM( $x$ ) $^{-2}$ ), Latency(ms) and FPS over test sequences of MFQE 2.0 Dataset at QP=37. All methods were retested on a NVIDIA RTX 3090 in frame-to-frame manner. Video resolution: Class A(2560\times1600), Class B(1980\times1080), Class C(832\times480), Class D(416\times240), Class E(1280\times720). + +
ApproachHigh-Latency VQE MethodsLow-Latency VQE Methods
MFQE 1.0 [31]PSTQE [7]Li et al. [17]DS-CNN [29]DnCNN [34]DCAD [26]STLVQE (Ours)
MetricsΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%
ATraffic0.50/0.901120.10.64/1.045.6K0.29/0.60385.50.29/0.60374.60.24/0.57294.90.31/0.67153.70.41/0.9496.3
PeopleOnStreet0.80/1.371.00.48/0.922.60.42/0.852.70.41/0.823.40.50/0.956.50.68/1.3510.4
BKimono0.50/1.13628.10.69/1.362.9K0.28/0.780.25/0.750.24/0.750.24/0.750.24/0.750.28/0.780.28/0.780.38/1.0548.6
ParkScene0.39/1.030.42/1.150.49/1.210.15/0.480.23/0.580.15/0.500.14/0.500.14/0.500.16/0.500.16/0.500.16/0.500.33/0.8020.6
Cactus0.44/0.882.00.62/1.150.25/0.480.25/0.480.24/0.580.20/0.480.20/0.380.26/0.580.28/0.500.28/0.500.43/0.8548.6
BQTerrace0.27/0.480.50/0.870.50/0.870.30/0.680.30/0.680.26/0.480.25/0.580.25/0.580.28/0.650.31/0.680.31/0.680.42/0.8320.6
BasketballDrive0.41/0.800.60/1.040.60/1.040.30/0.680.30/0.680.28/0.650.25/0.650.25/0.650.28/0.650.28/0.650.28/0.650.39/0.9411.5
CRaceHorses0.34/0.55229.80.40/0.880.28/0.650.33/0.880.27/0.630.27/0.630.25/0.650.28/0.680.26/0.680.34/0.880.50/1.2587.0
BQMall0.51/1.03229.80.74/1.44611.60.33/0.8839.50.33/0.8036.70.28/0.6826.30.34/0.8814.30.34/1.15
PartyScene0.22/0.7310.40.51/1.461.80.13/0.4525.30.17/0.5827.20.13/0.4838.00.16/0.4869.90.37/1.37
BasketballDrill0.48/0.901.66/1.270.66/1.270.38/0.880.38/0.880.35/0.680.33/0.680.33/0.680.39/0.780.39/0.780.39/0.780.44/1.089.1
DRaceHorses0.51/1.13164.30.60/1.44203.70.33/0.830.32/0.750.32/0.750.31/0.730.31/0.730.32/0.830.32/0.750.47/1.059.1
BQSquare-0.01/0.15164.30.79/1.140.09/0.250.09/0.250.20/0.380.20/0.380.13/0.180.13/0.187.00.20/0.380.35/1.37109.9
BlowingBubbles0.39/1.2032.20.62/1.950.21/0.680.21/0.680.23/0.680.23/0.680.18/0.580.18/0.58142.90.22/0.65285.70.37/1.23
BasketballPass0.63/1.380.85/1.750.85/1.750.34/0.850.34/0.780.34/0.780.34/0.780.31/0.750.35/0.850.35/0.850.35/0.850.48/1.28128
EFourPeople0.66/0.85345.40.95/1.121.3K0.45/0.700.46/0.700.46/0.700.39/0.600.42/0.600.51/0.780.51/0.780.69/1.0224.7
Johnny0.55/0.554.70.75/0.850.850.40/0.6087.40.38/0.4081.60.32/0.4064.80.41/0.500.93/0.6440.5
KristenAndSara0.66/0.750.66/0.750.93/0.910.49/0.680.49/0.680.48/0.600.42/0.600.42/0.600.52/0.700.52/0.700.52/0.700.74/0.84128
Average0.46/0.88-0.69/1.25-0.30/0.66-0.30/0.63-0.26/0.58-0.32/0.67-0.49/1.03-
+ +CIF(352 × 288), NTSC(720 × 486), 4CIF(704 × 576), 240P(416 × 240), 360P(640 × 360), 480P(832 × 280), 720P(1280 × 720), 1080P(1920 × 1080) and WQXGA (2560 × 1600). Same as [10], the 108 videos in the dataset are used as the training set and 18 videos are used as the test set. All videos were compressed by HM16.5 software in LDP mode with QP values of 22, 27, 32, 37 and 42. + +# 4.2 Implementation Details + +Our proposed network is implemented based on PyTorch framework and MMEditing toolbox [23]. During the training phase, we set batch size to 8 with $180 \times 180$ clips from the raw videos and the corresponding compressed videos as training samples. The random flip and rotation augmentation are utilized to fully exploit the information in the samples. The Adam optimizer [14] and Cosine Annealing scheme [19] are utilized for optimizing. The total iterations numbers and initial learning rate of the first and second stages is set to $200\mathrm{K}$ , $1 \times 10^{-4}$ and $100\mathrm{K}$ , $1 \times 10^{-6}$ . Our network is trained on 8 Telsa V100 GPUs. + +During the inference phase, ST-LUTs will replace the corresponding convolutional layers. To evaluate the model's performance in an online setting, we conduct inference on the test set frame-by-frame. Following the same settings as [10], we report quality enhancement on Y-channel in YUV/YCbCr space. The Peak Singal-to-Noise Ratio(ΔPSNR) and Structural Similarity(ΔSSIM) will be used as metrics to judge the model's enhancement capability. We further evaluate the latency(LT) and frame per second(FPS) of the models which is crucial to the online-VQE: + +$$ +L T = T _ {\text {p r o c e s s}} + T _ {\text {w a i t}}, \quad T _ {\text {w a i t}} = N _ {f} / F P S, \tag {11} +$$ + +where $T_{process}$ refers to the single-frame inference time and $T_{wait}$ refers to the future frames waiting time required for some video based methods, $N_f$ represents + +Table 2: Latency on 720P video and averaged $\Delta$ PSNR/ $\Delta$ SSIM comparison on the test sequences of MFQE 2.0 Dataset at different QP values. FV-Required means the full compressed video is required for enhancement, thus latency cannot be calculated. + +
MethodTypeLT(ms)StorageGFLOPsQP37QP42QP32QP27QP22
MFQEv1 [31]Bi-MF345.41788K-0.46/0.880.44/1.300.43/0.580.40/0.340.31/0.19
MFQEv2 [10]Bi-MF320.1255K7880.56/1.090.59/1.650.52/0.680.49/0.420.46/0.27
STDF [6]Bi-MF85.6487K2930.65/1.18-/-0.64/0.770.59/0.470.51/0.27
PSTQE [7]Bi-MF1249.8218K6790.69/1.250.69/1.860.67/0.830.63/0.520.55/0.29
RFDA [36]Bi-MF312.81250K-0.91/1.620.82/2.200.87/1.070.82/0.680.76/0.42
BasicVSR++ [2]Bi-MFFV-Required7300K4230.95/1.80-/-0.89/1.25-/--/-
RNAN [35]SF4329.48960K69790.44/0.95-/-0.41/0.62-/--/-
ARCNN [8]SF25.1118K1080.23/0.450.29/0.960.18/0.190.18/0.140.14/0.08
DnCNN [34]SF64.8558K5160.26/0.580.22/0.770.26/0.350.27/0.240.29/0.18
DS-CNN [29]SF81.6948K4370.30/0.630.31/1.010.27/0.380.27/0.230.25/0.15
Li et al. [17]SF87.4464K4340.30/0.660.32/1.050.28/0.370.30/0.280.30/0.19
DCAD [26]SF34.3299K2750.32/0.670.32/1.090.32/0.440.32/0.300.31/0.19
STLVQEUni-MF24.7398K+228M980.49/1.030.50/1.600.45/0.640.41/0.400.34/0.24
+ +the number of future frames used. Note that the LT metric ignores the frame reception and decoding time, which is fair for different methods, and we set $FPS = 30$ in Eq. 11 as the general online video standard to facilitate comparison. + +# 4.3 Comparison with State of the art Methods + +In this subsection, we present the results of our extensive experiments on various QP compressed videos in Tab. 1 and Tab. 2. Our proposed STLVQE network is compared with some widely used or state-of-art single-frame methods [8, 17, 26, 29, 34, 35], as well as some efficient multi-frame methods [6, 7, 10, 31]. In online environment where video is streamed in real-time, Bi-direction Multi-Frame (Bi-MF) method necessitates waiting for future frames that have not yet been received in current time. This waiting period introduces uncontrollable latency to the stream. Consequently, as outlined in previous online works [18, 22, 32], the Bi-MF method can not be directly applied to online scenarios and is predominantly utilized for offline processing. As a result, both Bi-MF and slow single frame methods $(>100\mathrm{ms})$ are placed in the first part in Tab. 1 and Tab. 2 for performance comparison, and demonstrate our favorable trade-off between performance and inference speed. All experiments were conducted on an NVIDIA RTX 3090 using a frame-to-frame style. + +Latency and FPS. Tab. 1 presents the latency and FPS metrics for each VQE approach. We note that while low-latency VQE methods can process each frame at a lower latency and reduce video delay, they often struggle to handle high-resolution videos in real-time, which can lead to stuttering and reduced smoothness. Among the methods listed, only STLVQE can achieve FPS above 40 on 720P resolution videos, making it suitable for most online scenarios. Compared to the second fastest method DCAD, STLVQE's processing time on 720p video is 9.6 milliseconds faster and outperforms it by a considerable margin, improving PSNR and SSIM metrics by 0.17dB and $0.36 \times 10^{-2}$ , respectively. + +Tab. 2 shows the latency of some large-scale high-latency temporal VQE methods. In contrast, high-latency VQE methods require waiting for future + +Table 3: Ablation study of Temporal Cache(TC), Reference Frame Window(RFW), Deform Operation(DF), and Preliminary Complement Residual(PC). The latter three sections share the features extracted by MAFE. + +
RFWTCDFPCΔP/ΔSTimeParams(K)
✓✓0.06/0.0110.85<1
✓✓✓0.31/0.5121.85352
✓✓✓0.39/0.8720.12149
✓✓✓0.49/1.0345.61398
✓✓✓0.46/0.9824.32398
✓✓✓✓0.49/1.0324.65398
+ +frames or taking a long time for inference, which can cause uncontrollable video delay. The latency of the two high-latency methods in Tab. 1 we evaluated exceeds $400\mathrm{ms}$ , much higher than the latency caused by STLVQE. Moreover, when considering only the FPS metric, STLVQE achieves a significantly faster speed than the single-frame VQE method RNAN(FPS = 0.2) by a factor of 175 on 720P video. Compared with MFQE 1.0, our method is 8.6 times faster with a better enhancement performance. Additionally, STLVQE has a more significant speed advantage on higher resolution videos such as 1080P and 2K. + +Enhancement performance. In Tab. 1 and Tab. 2, we present the performance of STLVQE compared to other commonly used or sota VQE methods for each video at $\mathrm{QP} = 37$ and for different QP compressed videos, respectively. Our method outperforms all the listed low-latency methods. In comparison with the high-latency VQE method, STLVQE reaches higher level with RNAN at $\mathrm{QP} = 37$ and $\mathrm{QP} = 32$ , with only $2.6\%$ and $0.6\%$ of RNAN's number of parameters and latency time (720P), respectively. The $\Delta$ SSIM of STLVQE is higher than that of the MFQE method for all five QPs, indicating that our method produces an enhanced image with a more stable image structure. The specific visualization results and analysis can be found in Qualitative Comparison section. STDF is a robust VQE method, but we believe that a stable latency and higher frame rate are more important for enhancing the user's viewing experience. + +Memory Size & GPU Consumption. As shown in Tab. 2, STLVQE has space occupation of 398KB for the model and 228MB for the LUT (utilizing less than $5\%$ memory of a low end GPU like GTX 1060), achieving real-time processing capability. Since most deep networks are replaced with ST-LUTs, inferencing STLVQE only needs 98 GFLOPs on a 720P frame. Note that GPU computing unit are more expensive than the memory. + +# 4.4 Ablation Study + +To validate the effectiveness of our proposed new VQE framework, we perform ablation experiments and specific analysis on three components. The results of the runtime per frame is computed on 720P videos at $\mathrm{QP} = 37$ . + +Effectiveness of the propagation module. The results are shown in the 5th and 6th row of Tab. 3. We found that selecting frames based on QP val- + +Table 4: Ablation study of STLUTs and the CNN enhancement network. The runtime is computed on a 720P video. + +
Enhancement ModuleΔP/ΔSTime(ms)Params(K)
S-LUT0.39/0.8323.52398
S-LUT + T-LUT-10.45/1.0224.12398
ST-LUTs0.47/1.0224.65398
ST-LUTs (on x̂i)0.49/1.0324.65398
CNN Network0.50/1.08147.02442
+ +ues in RFW can bring a 0.03dB $\Delta$ PSNR performance improvement compared to randomly selecting frames in the RFW, without adding parameters to the network and only increasing the inference time by $0.33\mathrm{ms}$ for comparing QP values of each frame. Furthermore, as shown in rows 4 and 6 of Tab. 3, the use of the TC structure can significantly reduce the inference time by $21.0~\mathrm{ms}$ , which is consumed by the repeated computation of the reference frame information. We achieve this reduction in time by merely occupying the GPU memory with information of several frames in the RFW. + +Effectiveness of the alignment module. We mainly validate the effect of the deform operation of the reference fram in the Alignment module. As shown in the first row of Tab. 3, when only using the ST-LUTs structure and discard the whole alignment module and PC operation, the result only has a $0.06\mathrm{dB}$ PSNR improvement compared to HEVC video. When the Deform operation is applied, the network can capture more accurate temporal information, and the $\Delta$ PSNR reaches $0.31~\mathrm{dB}$ . Due to the lightweight convolutional layers, the alignment module only adds $11.6~\mathrm{ms}$ of inference time to the network. + +Effectiveness of the enhancement module. As shown in row 3 of Tab. 3, using PC without the Deform operation achieves a significant $\Delta$ PSNR of 0.39 dB, demonstrating the effectiveness of PC in providing initial compensation for spatial information and extending the receptive field, ultimately enhancing the performance of ST-LUTs. Finally, after combining two parts, the network achieves 0.49 dB $\Delta$ PSNR. PC alone contributes 0.18 dB to the network and adds only 2.16 ms of inference time. + +Tab. 4 provides an analysis of the impact of each LUT in the enhancement module. In the first three rows of the table, we attempted to enhance the original image $x_{i}$ with the results of different LUT combinations. Results in the first and second rows show that using only the S-LUT with T-LUT-1 does not fully improve the network due to T-LUT-1 not utilizing all pixels within a kernel when learning the alignment offset. However, when T-LUT-2 is used for network training along with T-LUT-1, the reference frame can be better aligned, resulting in a $\Delta$ PSNR improvement of 0.08 dB. The 4th row shows the enhancing result on $\tilde{x}_{i}$ , reducing the learning difficulty of the ST-LUTs, which improving the performance by 0.02 dB. Compared to the CNN layers replaced by LUTs, ST-LUTs sacrifice only 0.01 dB $\Delta$ PSNR performance while reducing the number of parameters by 44K and the inference time by 122.37 ms. + +![](images/efb0ef0e07b8df0c0315bc1d9b5af99ab281b1d1b5a33aa3a4461799027fa0b4.jpg) +Fig. 5: Qualitative results at QP 37. The STLVQE method successfully enhances the left leg of the athlete, the face of the old man, the railings of the bridge and the eyebrow of the woman in each frame, respectively. + +The above ablation studies have sufficiently demonstrated that the MAFE structure of shared features in STLVQE with the redesigned three new modules can sufficiently reduce the computational redundancy of the network and obtain satisfactory inference speed-performance trade-offs. + +# 4.5 Qualitative Comparison + +Fig. 5 provides 4 examples of the qualitative results of our network on compressed video at $QP = 37$ . The quality of these images deteriorates significantly in compressed scenes, showing a variety of image defects (e.g. blurring legs, missing bridge railings). With the help of temporal information, STLVQE achieves better visual results than single-frame methods with significantly reduced runtime. Compared to large-scale multi-frame methods like MFQE 2.0, STLVQE can achieve competitive capabilities in detail and structural restoration. + +# 5 Conclusion + +In this paper, we propose a lightweight and rapid method called STLVQE based on the constraints of online video quality enhancement tasks. STLVQE greatly reduce the computation redundancy and uses a Spatial-Temporal Look-up Table structure to extract spatio-temporal information from the video with minimal time consumption. The extensive experiments prove that our proposed method can outperform most low-latency VQE methods, achieve competitive performance with high-latency VQE methods which cannot be directly utilized on Online-VQE tasks. Furthermore, STLVQE achieves real-time processing of 720P resolution videos, demonstrating a good trade-off between enhancement performance and inference speed. + +# Acknowledgements + +This work was supported by National Natural Science Fund of China (62076184, 61976158, 61976160, 62076182, 62276190), in part by Fundamental Research Funds for the Central Universities and State Key Laboratory of Integrated Services Networks (Xidian University), in part by Shanghai Innovation Action Project of Science and Technology (20511100700) and Shanghai Natural Science Foundation (22ZR1466700). + +# References + +1. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021) 3, 4 +2. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022) 2, 3, 4, 11 +3. Charbonnier, P., Blanc-Feraud, L., Aubert, G., Barlaud, M.: Two deterministic half-quadratic regularization algorithms for computed imaging. In: Proceedings of 1st international conference on image processing. vol. 2, pp. 168-172. IEEE (1994) 9 +4. Chen, H., He, X., Qing, L., Xiong, S., Nguyen, T.Q.: Dpw-sdnet: Dual pixel-wavelet domain deep cnns for soft decoding of JPEG-compressed images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 711-720 (2018) 2, 3 +5. Dai, Y., Liu, D., Wu, F.: A convolutional neural network approach for post-processing in hevc intra coding. In: MultiMedia Modeling: 23rd International Conference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23. pp. 28-39. Springer (2017) 3 +6. Deng, J., Wang, L., Pu, S., Zhuo, C.: Spatio-temporal deformable convolution for compressed video quality enhancement. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 10696-10703 (2020) 2, 3, 4, 5, 9, 11 +7. Ding, Q., Shen, L., Yu, L., Yang, H., Xu, M.: Patch-wise spatial-temporal quality enhancement for hevc compressed video. IEEE Transactions on Image Processing 30, 6459-6472 (2021) 3, 10, 11 +8. Dong, C., Deng, Y., Loy, C.C., Tang, X.: Compression artifacts reduction by a deep convolutional network. In: Proceedings of the IEEE international conference on computer vision. pp. 576-584 (2015) 3, 11 +9. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. arXiv preprint arXiv:1902.08153 (2019) 7 +10. Guan, Z., Xing, Q., Xu, M., Yang, R., Liu, T., Wang, Z.: Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE transactions on pattern analysis and machine intelligence 43(3), 949-963 (2019) 2, 3, 4, 5, 9, 10, 11 +11. Guo, J., Chao, H.: Building dual-domain representations for compression artifacts reduction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 628-644. Springer (2016) 3 + +12. Jin, Z., An, P., Yang, C., Shen, L.: Quality enhancement for intra frame coding via cnns: An adversarial approach. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1368-1372. IEEE (2018) 2, 3 +13. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 691-700 (2021) 3, 4, 8 +14. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10 +15. Kondratyuk, D., Yuan, L., Li, Y., Zhang, L., Tan, M., Brown, M., Gong, B.: Movinets: Mobile video networks for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16020-16030 (2021) 2 +16. Li, J., Chen, C., Cheng, Z., Xiong, Z.: Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVIII. pp. 238-256. Springer (2022) 4 +17. Li, K., Bare, B., Yan, B.: An efficient deep convolutional neural networks model for compressed image deblocking. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 1320-1325. IEEE (2017) 3, 10, 11 +18. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019) 2, 11 +19. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10 +20. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep kalman filtering network for video compression artifact reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018) 2, 3 +21. Ma, C., Zhang, J., Zhou, J., Lu, J.: Learning series-parallel lookup tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 305-321. Springer (2022) 4 +22. Maggioni, M., Huang, Y., Li, C., Xiao, S., Fu, Z., Song, F.: Efficient multistage video denoising with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3466-3475 (2021) 2, 11 +23. MMEditing Contributors: MMEditing: OpenMMLab image and video editing toolbox. https://github.com/open-mmlab/mmediting (2022) 10 +24. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012) 2 +25. Wang, T., Li, Y., Peng, J., Ma, Y., Wang, X., Song, F., Yan, Y.: Real-time image enhancer via learnable spatial-aware 3d lookup tables. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2471-2480 (2021) 4 +26. Wang, T., Chen, M., Chao, H.: A novel deep learning-based method of improving coding efficiency from the decoder-end for hevc. In: 2017 Data Compression Conference (DCC). pp. 410-419. IEEE (2017) 10, 11 +27. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003) 2 + +28. Xu, Y., Gao, L., Tian, K., Zhou, S., Sun, H.: Non-local convlstm for video compression artifact reduction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7043-7052 (2019) 3 +29. Yang, R., Xu, M., Liu, T., Wang, Z., Guan, Z.: Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology 29(7), 2039-2054 (2018) 2, 3, 4, 10, 11 +30. Yang, R., Xu, M., Wang, Z.: Decoder-side hevc quality enhancement with scalable convolutional neural network. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 817-822. IEEE (2017) 2, 3, 4 +31. Yang, R., Xu, M., Wang, Z., Li, T.: Multi-frame quality enhancement for compressed video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6664-6673 (2018) 2, 3, 4, 10, 11 +32. Yin, G., Qu, Z., Jiang, X., Jiang, S., Han, Z., Zheng, N., Yang, H., Liu, X., Yang, Y., Li, D., et al.: Online streaming video super-resolution with convolutional lookup table. IEEE Transactions on Image Processing 33, 2305-2317 (2024) 11 +33. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 4 +34. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 3, 10, 11 +35. Zhang, Y., Li, K., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082 (2019) 11 +36. Zhao, M., Xu, Y., Zhou, S.: Recursive fusion and deformable spatiotemporal attention for video compression artifact reduction. In: Proceedings of the 29th ACM international conference on multimedia. pp. 5646-5654 (2021) 2, 3, 4, 5, 9, 11 \ No newline at end of file diff --git a/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/images.zip b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8042f01c1c3640dee420a9c8787b4802d5e6dae0 --- /dev/null +++ b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbb71960985a1999892c7d0b04c55f20995ecb4bafcc4890b923e97b0d7fdd12 +size 529357 diff --git a/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/layout.json b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..edaba88f7bf54336e3a2578358d763fe47f2f3da --- /dev/null +++ b/2024/Online Video Quality Enhancement with Spatial-Temporal Look-up Tables/layout.json @@ -0,0 +1,9381 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "text", + "content": "Zefan " + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{Qu}^{1}" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "text", + "content": ", Xinyang Jiang" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{\\dagger 2}" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "text", + "content": ", Yifan Yang" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "text", + "content": ", Dongsheng Li" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "text", + "content": ", and Cairong Zhao" + }, + { + "bbox": [ + 138, + 168, + 475, + 192 + ], + "type": "inline_equation", + "content": "^{\\dagger 1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 202, + 384, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 202, + 384, + 213 + ], + "spans": [ + { + "bbox": [ + 230, + 202, + 384, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 230, + 202, + 384, + 213 + ], + "type": "text", + "content": " Tongji University, Shanghai, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 214, + 383, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 214, + 383, + 224 + ], + "spans": [ + { + "bbox": [ + 230, + 214, + 383, + 224 + ], + "type": "text", + "content": "{qzf, zhaocairong}@tongji.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 216, + 224, + 397, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 224, + 397, + 236 + ], + "spans": [ + { + "bbox": [ + 216, + 224, + 397, + 236 + ], + "type": "text", + "content": "2 Microsoft Research Asia, Shanghai, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 236, + 428, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 236, + 428, + 247 + ], + "spans": [ + { + "bbox": [ + 186, + 236, + 428, + 247 + ], + "type": "text", + "content": "{xinyangjiang,yifanyang,dongsheng.li}@microsoft.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 281, + 455, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 281, + 455, + 469 + ], + "spans": [ + { + "bbox": [ + 159, + 281, + 455, + 469 + ], + "type": "text", + "content": "Abstract. Low latency rates are crucial for online video-based applications, such as video conferencing and cloud gaming, which make improving video quality in online scenarios increasingly important. However, existing quality enhancement methods are limited by slow inference speed and the requirement for temporal information contained in future frames, making it challenging to deploy them directly in online tasks. In this paper, we propose a novel method, STLVQE, specifically designed to address the rarely studied online video quality enhancement (Online-VQE) problem. Our STLVQE designs a new VQE framework which contains a Module-Agnostic Feature Extractor that greatly reduces the redundant computations and redesign the propagation, alignment, and enhancement module of the network. A Spatial-Temporal Look-up Tables (STL) is proposed, which extracts spatial-temporal information in videos while saving substantial inference time. To the best of our knowledge, we are the first to exploit the LUT structure to extract temporal information in video tasks. Extensive experiments on the MFQE 2.0 dataset demonstrate our STLVQE achieves a satisfactory performance-speed trade-off." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "spans": [ + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "content": "Keywords: Online Video Quality Enhancement " + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "content": " Look-up Table " + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "content": " Video Compression Artifact Reduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 544, + 230, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 230, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 230, + 556 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 572, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 646 + ], + "type": "text", + "content": "With the rapid development of social media, users are increasingly seeking high-quality and real-time video content. In recent times, video-based applications such as video conferencing, webcasting, and cloud gaming have increasingly demanded low latency requirements. This has resulted in higher video compression rates, which can cause a noticeable decrease in video quality and an overall reduction in the Quality of Experience (QoE). As a result, online Video Quality" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 309, + 246 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 246 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 246 + ], + "type": "image", + "image_path": "46cbabc8dca024c6e69d705a1453a69b132be3d152381f10cf3e3a5a40511c68.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "text", + "content": "Fig. 1: Comparison of " + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "text", + "content": "PSNR/ " + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "text", + "content": "SSIM, average runtime per frame(720P video) and parameters on MFQE 2.0 test set at " + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{QP} = 37" + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "text", + "content": ". Our STLVQE method achieves a great trade-off in enhancement performance and inference speed." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 480, + 246 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 480, + 246 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 480, + 246 + ], + "type": "image", + "image_path": "4d95bcd21a4980bdc6cf6629ea7404945634c0bd848d02a5f936420e19366356.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 315, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 422 + ], + "type": "text", + "content": "Enhancement (Online-VQE), which improves the quality of real-time streaming video under online conditions and minimizing video defects caused by compression algorithms (e.g., H.264/AVC [27] and H.265/HEVC [24]) has become a significant problem. However, Online-VQE poses two main challenges compared to the conventional offline VQE approach. Firstly, it needs to process high-resolution videos in real-time to ensure a smooth viewing experience. Secondly, online video processing techniques [15, 18, 22] can only rely on the current and previous frames for inference due to the uncontrolled latency caused by waiting for future frames, which can result in overall video delay." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": "In recent years, numerous neural network-based techniques for enhancing the quality of compressed video have been developed, resulting in significant advancements in this field. These methods can be broadly classified into two categories, namely single-image based VQE method [4, 12, 29, 30] and multi-frame based VQE method [2, 6, 10, 20, 36]. Single-image based approaches are relatively faster in inference speed, but they ignore the significance of temporal information in video [31]. On the other hand, multi-frame based VQE achieves state-of-the-art performance by utilizing temporal information. However, most of these approaches are not suitable for online scenarios because they rely on subsequent frames as reference frames and have unsatisfactory inference speeds. Therefore, developing a new online-VQE method that satisfies the aforementioned two constraints while effectively leveraging temporal information is essential." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Reasons that limit the speed and latency of the above temporal VQE methods can be summarized in the following two points: First, a high level of computational redundancy exists within these methods, and these homogeneous computations increase the network's inference time. Second, most of these methods employ complex and computationally intensive models, which sacrificing inference speed and latency for some of the enhancement capabilities. In this work, a more applicable VQE method for online scenarios are proposed by advancing these two perspectives." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": "As discussed in previous works [1], modern VQE methods exploit temporal information by constructing a neural network containing three main components: the propagation, alignment and enhancement module. In all three modules, similar feature extraction operations on the original image may be involved, e.g., the reference frame selection [10, 31] in propagation module, the optical flow calculation [1, 2] in alignment module and the final result obtaining in the enhancement module [6, 10]. Most of the state-of-the-art methods adopts distinct feature extractors in different components and obtaining latent feature for each component separately, causing large computation redundancy. In this paper, we design a new VQE network framework containing a Module-Agnostic Feature Extractor (MAFE), where the extracted latent feature are uniformly share across three modules. This design allows the network to greatly reduce the redundant feature computation steps and speed up the inference of the network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "content": "When processing high-resolution videos in VQE task, the complex network of the sota method will greatly increase the computation times of the network, thus affecting the online deployment of the method. Therefore, we in our proposed framework: (1) Redesign the structure of the three modules with a light-weight convolutional structure. (2) Propose a Spatial-Temporal Look-up Tables (ST-LUTs) that extend conventional LUTs [13] to conduct queries on both spatial and temporal dimensions, which allows the LUTs to simultaneously capture information in both temporal and spatial dimensions, facilitating efficient extraction of temporal information. To the best of our knowledge, ST-LUTs is the first approach to employ a LUT structure for the extraction of temporal information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": "The major contributions are summarized as follows: (1) We propose the first online-VQE method STLVQE that achieves real-time process speed, which reduces redundant feature extraction computations while redesigning the propagation, alignment and enhancement modules of the temporal VQE network. (2) A Spatial-Temporal Look-up Tables structure is first time proposed for fully exploiting the temporal and spatial information in the video. (3) Extensive experiments on MFQE 2.0 dataset demonstrate that our approach achieves a comparable speed-performance trade-off. We also test most of the previous VQE methods on prevalent GPU to provide performance and speed benchmarks, facilitating future research in this area." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 531, + 243, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 243, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 243, + 544 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 618 + ], + "type": "text", + "content": "Compressed Video Quality Enhancement. Recently, numerous works have focused on the degradation of the network video quality due to the lossy video compression algorithms. Existing Compressed VQE work can be broadly classified into two categories, which are single-image based method [4, 5, 8, 11, 12, 17, 30, 34] and multi-frame based method [2, 6, 7, 10, 20, 28, 31, 36]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "For the single-image based methods, some of the early work [4,8,11,17,29,34] are mainly used to enhance the quality of JPEG compressed images, which can also be applied directly to compressed videos to process each frame independently. [5, 12, 30] utilize certain characteristics of the video compression algo" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "rhythm or compression parameters to improve the video quality. For example, DS-CNN [30] and QE-CNN [29] use distinct models to enhance the video processed by different compression coding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 247 + ], + "type": "text", + "content": "For the multi-frame based methods, MFQE 1.0 [31] and MFQE 2.0 [10] were proposed to use two adjacent high quality frames(PQF) and compute the optical flow between them to enhance the current frame. STDF [6] uses deformation convolution for an implicit alignment. RFDA [36] extracts long-range temporal information from the video via the recursive fusion module. BasicVSR [1] and BasicVSR++ [2] utilize a bidirectional propagation network and a new alignment scheme is proposed in BasicVSR++. All these methods demonstrate the significance of temporal information for solving the compressed VQE problem." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 480, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 294 + ], + "type": "text", + "content": "As far as we know, there is still no compression VQE method for online scenario that has been proposed so far. The above-mentioned methods cannot be directly applied to webcasting environments due to the limitations of the two online restriction mentioned above." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 296, + 481, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 481, + 451 + ], + "type": "text", + "content": "Look-up Table. Look-up Tables (LUTs) have proven to be highly effective in reducing computational power and time consumption during the inference phase of algorithms by replacing complex computation with fast storage space retrieval operations. However, the use of LUTs directly is not always flexible as it requires manual adjustment by experts to accurately process images. To overcome this limitation, recent attempts have been made to improve LUTs. [25, 33] combine multiple LUTs and learn the weights through neural networks to summing up these LUT results. SR-LUT [13] replaces a small receptive field neural network with a 4D-LUT and achieves good speed-performance trade-offs on SR tasks. SP-LUT [21] and MuLUT [16] makes improvements to the SR-LUT by expanding its receptive field through parallel and serial multiple LUT tables. In this work, we utilize LUTs to extract information in both temporal and spatial dimensions, bringing its advantages to the online VQE task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 466, + 202, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 202, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 202, + 478 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 487, + 209, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 209, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 209, + 498 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": "Given a stream of compressed frames without an explicit ending " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "\\{x_0, \\ldots, x_i, \\ldots\\}" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": " with input channel " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": ", height " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": ", we sequentially enhance each compressed frame to obtain the target sequence " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "\\{y_0, \\ldots, y_i, \\ldots\\}" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": ". Due to the low latency requirement of Online-VQE, obtaining information from future frames to enhance " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": " is unacceptable. We only use past and current frames " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "\\{x_0, x_1, \\ldots, x_i\\}" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 130, + 504, + 480, + 587 + ], + "type": "text", + "content": ", reducing the video's delay time. Additionally, to avoid stuttering issues, each frame's inference speed must be fast enough." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 588, + 480, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 480, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 480, + 648 + ], + "type": "text", + "content": "The overall flow of our proposed network during inference period is illustrated in Fig. 2 (The difference between the training and the inference phase is illustrated in Subsec. 3.4 and the figure in appendix). First, a Module-Agnostic Feature Extractor (MAFE) is utilized to extract the feature " + }, + { + "bbox": [ + 130, + 588, + 480, + 648 + ], + "type": "inline_equation", + "content": "F_{\\text{eat}_i}" + }, + { + "bbox": [ + 130, + 588, + 480, + 648 + ], + "type": "text", + "content": " of the current enhancing frame " + }, + { + "bbox": [ + 130, + 588, + 480, + 648 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 588, + 480, + 648 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 259, + 653, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 259, + 653, + 481, + 666 + ], + "type": "interline_equation", + "content": "F e a t _ {i} = M A F E \\left(x _ {i}\\right), \\tag {1}", + "image_path": "d08da2e3e2c8642dd0636ceeb461879bcf6d2d65465251e72e696d0017c253f6.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 478, + 228 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 478, + 228 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 478, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 478, + 228 + ], + "type": "image", + "image_path": "26385d2ba90d424f1d876ac9095939b18e6358def4e14f07edf284979a24d154.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 482, + 291 + ], + "type": "text", + "content": "Fig. 2: The framework of STLVQE (inference phase), which consists of a Module-Agnostic Feature Extractor and three main modules. In the inference phase, the propagation module selects the reference frame and accesses the relevant information, which is sent to the alignment module for temporal alignment, and finally the aligned reference and enhancing frames are input to the enhancement module for the final results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 315, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 350 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 315, + 480, + 350 + ], + "type": "inline_equation", + "content": "\\text{Feat}_i" + }, + { + "bbox": [ + 130, + 315, + 480, + 350 + ], + "type": "text", + "content": " will be shared by the three subsequent modules, thus reducing the multiple time-consuming extraction steps and fully exploiting the latent frame feature." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "content": "Specifically, in the propagation module, " + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "inline_equation", + "content": "Feat_{i}" + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "content": " is stored in a structure called Temporal Cache, which will provide the temporal information in subsequent modules based on the selection of the reference frames " + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "inline_equation", + "content": "x_{ref} = \\{x_{ref1}, x_{ref2}\\}" + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "content": " without external computation. In the alignment module, " + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "inline_equation", + "content": "Feat_{i}" + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "inline_equation", + "content": "Feat_{ref}" + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "content": ", the reference frame feature obtained from the propagation module, to deform " + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "inline_equation", + "content": "x_{ref}" + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "content": " with a lightweight offset predictor. In the enhancement module, " + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "inline_equation", + "content": "Feat_{i}" + }, + { + "bbox": [ + 130, + 350, + 480, + 446 + ], + "type": "text", + "content": " is used to perform a spatial complement, which fully exploits the extracted features and augments the ST-LUTs by increasing its receptive fields." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 481, + 495 + ], + "type": "text", + "content": "Under our proposed new framework, only once feature extraction computation for a single frame is required for enhancing each frame. The extracted features are simultaneously computed with reference frame alignment and preliminary compensation in the current frame space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 513, + 288, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 288, + 525 + ], + "type": "text", + "content": "3.2 The Propagation Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "The propagation module in the VQE task is responsible for extracting and transmitting temporal information within the video. Previous methods have explored various approaches, including designing additional structures [10] and increasing the number of selected reference frames [6, 36] to improve network performance, which greatly increase inference time and computational costs of the model. The architecture of the propagation module is illustrated in Fig. 3. The compression configuration parameter quantization parameter (QP) value is utilized as an indicator to select fewer but higher quality frames among multiple candidate frames without extra computation. Specifically, a Reference Frame Window (RFW) which contains " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{x_{i-W}\\dots x_{i-1}\\}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is set. Then, two frames with the smallest QP values are selected as " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "x_{ref}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " for the subsequent module." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 115, + 427, + 251 + ], + "blocks": [ + { + "bbox": [ + 187, + 115, + 427, + 251 + ], + "lines": [ + { + "bbox": [ + 187, + 115, + 427, + 251 + ], + "spans": [ + { + "bbox": [ + 187, + 115, + 427, + 251 + ], + "type": "image", + "image_path": "4dee20de2c65bd311fda64ee50378b0e747581a802da31f2b17234971fa8794b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 258, + 475, + 269 + ], + "lines": [ + { + "bbox": [ + 137, + 258, + 475, + 269 + ], + "spans": [ + { + "bbox": [ + 137, + 258, + 475, + 269 + ], + "type": "text", + "content": "Fig. 3: Reference Frame Window and Temporal Cache in the propagation module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "content": "To mitigate the issue of additional feature extraction operations in previous works, we introduce a cache-like structure known as the Temporal Cache (TC). The TC stores reusable features of past frames that were already extracted by the MAFE in previous process, thereby eliminating the need to re-extract features for each subsequent frame. In particular, when enhancing the i-th frame " + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "content": ", STLVQE caches the intermediate feature map " + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "inline_equation", + "content": "Fea t_{i}" + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "content": " extracted by the alignment module, as well as the enhanced frame, in the information cache (TC). During reference frame selection, the features and enhanced result of " + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "inline_equation", + "content": "x_{ref1}" + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "inline_equation", + "content": "x_{ref2}" + }, + { + "bbox": [ + 130, + 291, + 482, + 399 + ], + "type": "text", + "content": " are directly retrieved from TC without extra feature extraction process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 415, + 279, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 279, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 279, + 427 + ], + "type": "text", + "content": "3.3 The Alignment Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 505 + ], + "type": "text", + "content": "The alignment module is a crucial component in video processing tasks that handles spatial misalignments between frames, which plays a vital role in effectively utilizing the temporal information embedded in the video, especially when the video content is subject to large motion. STLVQE includes a lightweight offset predictor to replace the optical flow or non-local computation of most VQE methods, which significantly reduce the computation cost." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "content": "The offset predictor is responsible for predicting spatial offsets which align correlated pixels in the reference with pixels in " + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "content": ". As introduced in the last section, the latent features " + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "inline_equation", + "content": "Feat_{ref}" + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "content": " and spatially compensated frames " + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}" + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "content": " at each processing step are stored in TC, which are retrieved by RFW to enhance future frames. Given the feature of a reference frame " + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}" + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "content": " retrieved from the TC, as well as the feature of the current frame " + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "content": ", the offset predictor predicts 9 spatial offsets for each pixel in " + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 506, + 482, + 589 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 229, + 596, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 596, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 229, + 596, + 481, + 611 + ], + "type": "interline_equation", + "content": "o _ {r e f} = C _ {\\text {o f f s e t}} \\left(\\left[ \\operatorname {F e a t} _ {i}, \\operatorname {F e a t} _ {r e f} ^ {d} \\right]\\right) \\tag {2}", + "image_path": "a8c8265b654387d32f33a8c8543b0a07080003dfdc5d40789d150554febecfeb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "content": "where the matrix concatenation is denoted by " + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "content": ", and each offset value in " + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "inline_equation", + "content": "o_{ref}" + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "content": " represents the position shift of a correlated pixel in " + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}" + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "content": " relative to a pixel in " + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "The temporal corresponding pixels in the reference frames " + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "x_{ref}" + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": " are spatially aligned with the " + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": " pixel offsets " + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "o_{ref}" + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": " and bilinear interpolation. Each set of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "text", + "content": "obtained correlated pixel values of a pixel in " + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "text", + "content": " are re-arranged into a " + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "text", + "content": " patch, which are then stacked together to obtain the deformed reference frame denoted as " + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "text", + "content": " with the size of " + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "inline_equation", + "content": "N \\times C \\times 3H \\times 3W" + }, + { + "bbox": [ + 130, + 116, + 479, + 155 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 168, + 294, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 168, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 131, + 168, + 294, + 179 + ], + "type": "text", + "content": "3.4 The Enhancement Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "text", + "content": "The enhancement module contains a spatial compensation network and a ST-LUTs network. To fully utilize the sharing latent feature " + }, + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "inline_equation", + "content": "Feat_{i}" + }, + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "text", + "content": ", a spatial compensation network reuses " + }, + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "inline_equation", + "content": "Feat_{i}" + }, + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "text", + "content": " to generate a spatial residual, which compensates the low quality frame " + }, + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 186, + 479, + 243 + ], + "type": "text", + "content": " to obtain preliminary spatial enhancement before ST-LUTs:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 247, + 480, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 247, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 250, + 247, + 480, + 258 + ], + "type": "interline_equation", + "content": "\\tilde {x} _ {i} = x _ {i} + C _ {\\text {c o m p}} (\\text {F e a t} _ {i}) \\tag {3}", + "image_path": "6691f3bc42d55f4a20506528dd37153fbf1a957b76e48a92d9537681d487f5f3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "inline_equation", + "content": "C_{comp}" + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": " are 3 shallow convolutional layers. Since the subsequent enhancement module takes the obtained compensated frame " + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": " as input, spatial compensation network can boost the enhancement performance by expanding the receptive field of the ST-LUTs in the enhancement module. To match the resolution of " + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}^{d}" + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": " is upsampled by repeating each pixel in " + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": " to form a 3x3 patch, then each pixel in the up-sampled " + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": " is correlated with the pixel at the same position in " + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}^{d}" + }, + { + "bbox": [ + 130, + 262, + 480, + 357 + ], + "type": "text", + "content": ". This is achieved by the unfold and fold operations with different strides as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 366, + 480, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 366, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 203, + 366, + 480, + 379 + ], + "type": "interline_equation", + "content": "\\tilde {x} _ {i} ^ {u p} = \\operatorname {F o l d} (\\operatorname {U n f o l d} (\\tilde {x} _ {i}, \\text {s t r i d e} = 1), \\text {s t r i d e} = 3), \\tag {4}", + "image_path": "2e3ab7aeb5046ab8ffa2f5ca60c5ecb0f3d4946c25d2fc00a733ce4632ce1b92.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 385, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 479, + 409 + ], + "type": "text", + "content": "the up-sampled input frame " + }, + { + "bbox": [ + 130, + 385, + 479, + 409 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^{up}" + }, + { + "bbox": [ + 130, + 385, + 479, + 409 + ], + "type": "text", + "content": " and deformed reference frame " + }, + { + "bbox": [ + 130, + 385, + 479, + 409 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 385, + 479, + 409 + ], + "type": "text", + "content": " together serve as the input of the ST-LUTs network for the final quality enhancement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 410, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 479, + 456 + ], + "type": "text", + "content": "Instead of deep neural network, STLVQE adopts a novel Spatial-Temporal Look-up Tables in the enhancement module during inference to replace complex convolutional computation used in training stage with much faster direct memory access, and conducts table queries on both spatial and temporal perspective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "text", + "content": "Quantization. Due to the limited storage space of the LUT, the size of each pixel in the input compensated frame " + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^{up}" + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "text", + "content": " and the deformed frames " + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "text", + "content": " needs to be controlled to fit within a suitable range of bits. Because the range of values in " + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^{up}" + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 457, + 480, + 552 + ], + "type": "text", + "content": " can vary dynamically and the gradient needs to be passed through the quantization operation during end-to-end training, it's not practical to simply quantize the matrices using a fixed quantization hyper-parameter. The enhancement module adopts a Learnable Quantization layer (LSQ) [9] to quantize the input of the enhancement module as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 216, + 566, + 480, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 566, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 216, + 566, + 480, + 578 + ], + "type": "interline_equation", + "content": "\\bar {x} = \\left\\lfloor \\operatorname {c l i p} \\left(\\tilde {x} / s, - Q _ {N}, Q _ {P}\\right) \\right], \\quad \\hat {x} = \\bar {x} \\times s, \\tag {5}", + "image_path": "eb35b73d3ba062d41459f9c7554f10b2961d519622e40a09dc4368933fe551c4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\text{clip}(z, r_1, r_2)" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " returns matrix " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " with values below " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " set to " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " and values above " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " set to " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\lfloor z \\rfloor" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " rounds " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " to the nearest integer. The step size parameter " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " can be optimized along with the network's training. " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\overline{x}" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " is the indexes used in the LUTs store and retrieval stages, and " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " is a quantized representation of the data with the same scale as " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": ". After quantization, the deep enhancement network and the corresponding LUTs can be constructed with the quantized pixel values as input." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 448, + 103 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 113, + 458, + 289 + ], + "blocks": [ + { + "bbox": [ + 154, + 113, + 458, + 289 + ], + "lines": [ + { + "bbox": [ + 154, + 113, + 458, + 289 + ], + "spans": [ + { + "bbox": [ + 154, + 113, + 458, + 289 + ], + "type": "image", + "image_path": "1e72c8caf4e5426fa770ce31b16d6670bfe8e9e5eb09e0024a0c6df5e3335f99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 295, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 319 + ], + "type": "text", + "content": "Fig. 4: The specific schematic of our Spatial-Temporal Look-up Tables, which consists of two parts: Temporal-LUT (left side) and Spatial-LUT (right side)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "text", + "content": "Deep Enhancement Network Training. In the training phase, a neural network based enhancement module is optimized, which takes the compensated frame " + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{up}" + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "text", + "content": " and reference frames " + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\hat{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "text", + "content": " as input and output the final enhancement results. To train an enhancement network that is suitable for LUT conversion, the receptive field size of the deep network typically cannot exceed 6 (Larger receptive fields will result in a higher storage occupancy, refer to the appendix for details). Therefore, each branch of the enhancement network, which will be replaced by ST-LUTs during the inference stage, consists of 1 convolutional layer with a kernel size of 4 or 6 and 10 convolutional layers with " + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 342, + 482, + 461 + ], + "type": "text", + "content": " kernel. The specific design of the network will be illustrated in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 581 + ], + "type": "text", + "content": "ST-LUTs Construction & Replacement. We generate LUTs by permuting all input patches of the receptive field size, and storing all the input/output mappings of the trained deep enhancement network in mapping tables. During inference, we retrieve the outputs of the deep enhancement network directly from the LUTs, replacing slow and computationally-intensive convolution with faster memory access and retain the other components in STLVQE. Due to the input patch size restriction, conventional LUTs are not suitable for inter-frame queries in the temporal dimension. To address these issues, we firstly propose the STLUTs, which employ multiple LUTs to conduct parallel retrieval with different sets of query pixels in both spatial and temporal dimensions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": "Specifically, given the quantized compensated frames " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{up}" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " and the aligned frames " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " as input, the look-up tables are constructed as shown in Fig. 4, where two sets of LUTs named S-LUTs and T-LUTs are constructed for extracting spatial and temporal information respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "For S-LUTs, a rotation ensemble operation [13] is used during the training process to further expand the LUT receptive field. As shown in Fig. 4, during training, input images are rotated by 0, 90, 180, and 270 degrees, and the results" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "are reverse-rotated and summed up. This approach considers all pixels in a " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " receptive field simultaneously. The result of S-LUT " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "R_{i}^{S}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " is expressed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 148, + 481, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 148, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 230, + 148, + 481, + 182 + ], + "type": "interline_equation", + "content": "R _ {i} ^ {S} = \\frac {1}{4} \\sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \\left(f ^ {S} \\left(R o t _ {j} \\left(\\hat {x} _ {i} ^ {u p}\\right)\\right)\\right), \\tag {6}", + "image_path": "633227ad5c74d1ee217eb794968726bc73c0f4d64417ae2b059d0173ac122969.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "inline_equation", + "content": "Rot_{j}" + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "inline_equation", + "content": "Rot_{j}^{-1}" + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "text", + "content": " are the rotation and reserve-rotation operations, and " + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "inline_equation", + "content": "f^{S}" + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "text", + "content": " is the spatial enhancement network or the S-LUTs retrieval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "text", + "content": "For T-LUTs, the pixels at the same location of the enhancing and reference frames will be strongly correlated after alignment, and jointly considering these values can make full use of the temporal information in the video. As shown on the left in Fig. 4, we establish two T-LUTs: T-LUT-1 take the same position pixels of " + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{up}" + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\hat{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "text", + "content": " as index with the rotation ensemble operation, while T-LUT-2 uses the four edge pixels of each kernel in the " + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\hat{x}_{ref}^d" + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "text", + "content": " as indexes to retrieve the surrounding information of the target pixels. The result of T-LUTs " + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "inline_equation", + "content": "R_i^T" + }, + { + "bbox": [ + 130, + 215, + 480, + 312 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 322, + 480, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 322, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 212, + 322, + 480, + 371 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R _ {i} ^ {T 1} = \\frac {1}{4} \\sum_ {j = 0} ^ {4} R o t _ {j} ^ {- 1} \\left(f ^ {T 1} \\left(R o t _ {j} \\left(\\left[ \\hat {x} _ {i} ^ {u p}, \\hat {x} _ {r e f} ^ {d} \\right]\\right)\\right)\\right), \\tag {7} \\\\ R _ {i} ^ {T 2} = f ^ {T 2} (\\hat {x} _ {r e f} ^ {d}), \\quad R _ {i} ^ {T} = R _ {i} ^ {T 1} + R _ {i} ^ {T 2}, \\\\ \\end{array}", + "image_path": "8c41022e8da929aebd8ef0ed2c774af77fcf777510dbc17c3e407c3dff32298c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "f^{T1}" + }, + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "f^{T2}" + }, + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "text", + "content": " are the temporal enhancement network or the T-LUTs retrieval. Then we can acquire the results of ST-LUTs and get the final enhanced image " + }, + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 379, + 479, + 415 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 423, + 480, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 423, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 233, + 423, + 480, + 437 + ], + "type": "interline_equation", + "content": "R _ {i} = R _ {i} ^ {S} + R _ {i} ^ {T}, \\quad y _ {i} = \\tilde {x} _ {i} + R _ {i}. \\tag {8}", + "image_path": "b29af39adc01b9643d4e6d12b765c2bc7e098032f74fe1115e2d25653dcd8f6b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 453, + 245, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 245, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 245, + 464 + ], + "type": "text", + "content": "3.5 Training Scheme" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 472, + 479, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 479, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 479, + 496 + ], + "type": "text", + "content": "We perform the training of the STLVQE in two stages with an end-to-end fashion. Firstly, the Charbonnier Loss [3] is utilized to optimize the model as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 253, + 507, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 507, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 253, + 507, + 480, + 521 + ], + "type": "interline_equation", + "content": "L _ {c h a r} = \\sqrt {\\left(y _ {t} - \\bar {y} _ {t}\\right) ^ {2} + \\epsilon}, \\tag {9}", + "image_path": "06c720de31c8fc8411b151cd0a086cea98d9488ba1b1e9a7ca435d01b2d3e1b4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\overline{y_t}" + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "text", + "content": " is the enhanced and ground truth frame at time " + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 525, + 479, + 561 + ], + "type": "text", + "content": " in our paper. Secondly, to further enhance the enhancement capability, we perform fine-tuning using L2 loss with a smaller learning rate as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 255, + 567, + 480, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 567, + 480, + 584 + ], + "spans": [ + { + "bbox": [ + 255, + 567, + 480, + 584 + ], + "type": "interline_equation", + "content": "L _ {M S E} = \\left\\| \\left(y _ {t} - \\bar {y} _ {t}\\right) ^ {2} \\right\\| _ {2} ^ {2}. \\tag {10}", + "image_path": "833f6076b837439c5ce7c8abf3aa7ccf41f815a4b2236143af18966be51f69a2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 598, + 230, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 230, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 230, + 612 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 622, + 204, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 204, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 204, + 633 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "Following [6, 10, 36], we conduct our experiments on the MFQE 2.0 dataset. It contains totally 126 sequences videos with multi resolutions: SIF(352 × 240)," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 103 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 482, + 306 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Overall comparison for " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " PSNR(dB), " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " SSIM(" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": ")" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "^{-2}" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "), Latency(ms) and FPS over test sequences of MFQE 2.0 Dataset at QP=37. All methods were retested on a NVIDIA RTX 3090 in frame-to-frame manner. Video resolution: Class A(2560\\times1600), Class B(1980\\times1080), Class C(832\\times480), Class D(416\\times240), Class E(1280\\times720)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 482, + 306 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 482, + 306 + ], + "type": "table", + "html": "
ApproachHigh-Latency VQE MethodsLow-Latency VQE Methods
MFQE 1.0 [31]PSTQE [7]Li et al. [17]DS-CNN [29]DnCNN [34]DCAD [26]STLVQE (Ours)
MetricsΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%ΔP/ΔSLT/FP%
ATraffic0.50/0.901120.10.64/1.045.6K0.29/0.60385.50.29/0.60374.60.24/0.57294.90.31/0.67153.70.41/0.9496.3
PeopleOnStreet0.80/1.371.00.48/0.922.60.42/0.852.70.41/0.823.40.50/0.956.50.68/1.3510.4
BKimono0.50/1.13628.10.69/1.362.9K0.28/0.780.25/0.750.24/0.750.24/0.750.24/0.750.28/0.780.28/0.780.38/1.0548.6
ParkScene0.39/1.030.42/1.150.49/1.210.15/0.480.23/0.580.15/0.500.14/0.500.14/0.500.16/0.500.16/0.500.16/0.500.33/0.8020.6
Cactus0.44/0.882.00.62/1.150.25/0.480.25/0.480.24/0.580.20/0.480.20/0.380.26/0.580.28/0.500.28/0.500.43/0.8548.6
BQTerrace0.27/0.480.50/0.870.50/0.870.30/0.680.30/0.680.26/0.480.25/0.580.25/0.580.28/0.650.31/0.680.31/0.680.42/0.8320.6
BasketballDrive0.41/0.800.60/1.040.60/1.040.30/0.680.30/0.680.28/0.650.25/0.650.25/0.650.28/0.650.28/0.650.28/0.650.39/0.9411.5
CRaceHorses0.34/0.55229.80.40/0.880.28/0.650.33/0.880.27/0.630.27/0.630.25/0.650.28/0.680.26/0.680.34/0.880.50/1.2587.0
BQMall0.51/1.03229.80.74/1.44611.60.33/0.8839.50.33/0.8036.70.28/0.6826.30.34/0.8814.30.34/1.15
PartyScene0.22/0.7310.40.51/1.461.80.13/0.4525.30.17/0.5827.20.13/0.4838.00.16/0.4869.90.37/1.37
BasketballDrill0.48/0.901.66/1.270.66/1.270.38/0.880.38/0.880.35/0.680.33/0.680.33/0.680.39/0.780.39/0.780.39/0.780.44/1.089.1
DRaceHorses0.51/1.13164.30.60/1.44203.70.33/0.830.32/0.750.32/0.750.31/0.730.31/0.730.32/0.830.32/0.750.47/1.059.1
BQSquare-0.01/0.15164.30.79/1.140.09/0.250.09/0.250.20/0.380.20/0.380.13/0.180.13/0.187.00.20/0.380.35/1.37109.9
BlowingBubbles0.39/1.2032.20.62/1.950.21/0.680.21/0.680.23/0.680.23/0.680.18/0.580.18/0.58142.90.22/0.65285.70.37/1.23
BasketballPass0.63/1.380.85/1.750.85/1.750.34/0.850.34/0.780.34/0.780.34/0.780.31/0.750.35/0.850.35/0.850.35/0.850.48/1.28128
EFourPeople0.66/0.85345.40.95/1.121.3K0.45/0.700.46/0.700.46/0.700.39/0.600.42/0.600.51/0.780.51/0.780.69/1.0224.7
Johnny0.55/0.554.70.75/0.850.850.40/0.6087.40.38/0.4081.60.32/0.4064.80.41/0.500.93/0.6440.5
KristenAndSara0.66/0.750.66/0.750.93/0.910.49/0.680.49/0.680.48/0.600.42/0.600.42/0.600.52/0.700.52/0.700.52/0.700.74/0.84128
Average0.46/0.88-0.69/1.25-0.30/0.66-0.30/0.63-0.26/0.58-0.32/0.67-0.49/1.03-
", + "image_path": "5b69e6eab104892474a4a1848e5f522660bd0bef7579a7e48ed5482269ac6d9a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 483, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 483, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 483, + 389 + ], + "type": "text", + "content": "CIF(352 × 288), NTSC(720 × 486), 4CIF(704 × 576), 240P(416 × 240), 360P(640 × 360), 480P(832 × 280), 720P(1280 × 720), 1080P(1920 × 1080) and WQXGA (2560 × 1600). Same as [10], the 108 videos in the dataset are used as the training set and 18 videos are used as the test set. All videos were compressed by HM16.5 software in LDP mode with QP values of 22, 27, 32, 37 and 42." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 403, + 280, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 280, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 280, + 415 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "text", + "content": "Our proposed network is implemented based on PyTorch framework and MMEditing toolbox [23]. During the training phase, we set batch size to 8 with " + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "inline_equation", + "content": "180 \\times 180" + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "text", + "content": " clips from the raw videos and the corresponding compressed videos as training samples. The random flip and rotation augmentation are utilized to fully exploit the information in the samples. The Adam optimizer [14] and Cosine Annealing scheme [19] are utilized for optimizing. The total iterations numbers and initial learning rate of the first and second stages is set to " + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "inline_equation", + "content": "200\\mathrm{K}" + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}" + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 130, + 422, + 482, + 517 + ], + "type": "text", + "content": ". Our network is trained on 8 Telsa V100 GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 613 + ], + "type": "text", + "content": "During the inference phase, ST-LUTs will replace the corresponding convolutional layers. To evaluate the model's performance in an online setting, we conduct inference on the test set frame-by-frame. Following the same settings as [10], we report quality enhancement on Y-channel in YUV/YCbCr space. The Peak Singal-to-Noise Ratio(ΔPSNR) and Structural Similarity(ΔSSIM) will be used as metrics to judge the model's enhancement capability. We further evaluate the latency(LT) and frame per second(FPS) of the models which is crucial to the online-VQE:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 211, + 620, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 620, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 211, + 620, + 481, + 634 + ], + "type": "interline_equation", + "content": "L T = T _ {\\text {p r o c e s s}} + T _ {\\text {w a i t}}, \\quad T _ {\\text {w a i t}} = N _ {f} / F P S, \\tag {11}", + "image_path": "2687d96410b646a43f532cee700a6af9f2d3d67333015645b12c1c055b48badc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{process}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " refers to the single-frame inference time and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{wait}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " refers to the future frames waiting time required for some video based methods, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_f" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " represents" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 480, + 270 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 2: Latency on 720P video and averaged " + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "PSNR/ " + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "SSIM comparison on the test sequences of MFQE 2.0 Dataset at different QP values. FV-Required means the full compressed video is required for enhancement, thus latency cannot be calculated." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 480, + 270 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 480, + 270 + ], + "type": "table", + "html": "
MethodTypeLT(ms)StorageGFLOPsQP37QP42QP32QP27QP22
MFQEv1 [31]Bi-MF345.41788K-0.46/0.880.44/1.300.43/0.580.40/0.340.31/0.19
MFQEv2 [10]Bi-MF320.1255K7880.56/1.090.59/1.650.52/0.680.49/0.420.46/0.27
STDF [6]Bi-MF85.6487K2930.65/1.18-/-0.64/0.770.59/0.470.51/0.27
PSTQE [7]Bi-MF1249.8218K6790.69/1.250.69/1.860.67/0.830.63/0.520.55/0.29
RFDA [36]Bi-MF312.81250K-0.91/1.620.82/2.200.87/1.070.82/0.680.76/0.42
BasicVSR++ [2]Bi-MFFV-Required7300K4230.95/1.80-/-0.89/1.25-/--/-
RNAN [35]SF4329.48960K69790.44/0.95-/-0.41/0.62-/--/-
ARCNN [8]SF25.1118K1080.23/0.450.29/0.960.18/0.190.18/0.140.14/0.08
DnCNN [34]SF64.8558K5160.26/0.580.22/0.770.26/0.350.27/0.240.29/0.18
DS-CNN [29]SF81.6948K4370.30/0.630.31/1.010.27/0.380.27/0.230.25/0.15
Li et al. [17]SF87.4464K4340.30/0.660.32/1.050.28/0.370.30/0.280.30/0.19
DCAD [26]SF34.3299K2750.32/0.670.32/1.090.32/0.440.32/0.300.31/0.19
STLVQEUni-MF24.7398K+228M980.49/1.030.50/1.600.45/0.640.41/0.400.34/0.24
", + "image_path": "cc4f13063cefd500005b12e272f9eafb3cc9c8610ab87b4cb4106d1c5c62488a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": "the number of future frames used. Note that the LT metric ignores the frame reception and decoding time, which is fair for different methods, and we set " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "FPS = 30" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": " in Eq. 11 as the general online video standard to facilitate comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 346, + 377, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 377, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 377, + 360 + ], + "type": "text", + "content": "4.3 Comparison with State of the art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 533 + ], + "type": "text", + "content": "In this subsection, we present the results of our extensive experiments on various QP compressed videos in Tab. 1 and Tab. 2. Our proposed STLVQE network is compared with some widely used or state-of-art single-frame methods [8, 17, 26, 29, 34, 35], as well as some efficient multi-frame methods [6, 7, 10, 31]. In online environment where video is streamed in real-time, Bi-direction Multi-Frame (Bi-MF) method necessitates waiting for future frames that have not yet been received in current time. This waiting period introduces uncontrollable latency to the stream. Consequently, as outlined in previous online works [18, 22, 32], the Bi-MF method can not be directly applied to online scenarios and is predominantly utilized for offline processing. As a result, both Bi-MF and slow single frame methods " + }, + { + "bbox": [ + 130, + 366, + 482, + 533 + ], + "type": "inline_equation", + "content": "(>100\\mathrm{ms})" + }, + { + "bbox": [ + 130, + 366, + 482, + 533 + ], + "type": "text", + "content": " are placed in the first part in Tab. 1 and Tab. 2 for performance comparison, and demonstrate our favorable trade-off between performance and inference speed. All experiments were conducted on an NVIDIA RTX 3090 using a frame-to-frame style." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "text", + "content": "Latency and FPS. Tab. 1 presents the latency and FPS metrics for each VQE approach. We note that while low-latency VQE methods can process each frame at a lower latency and reduce video delay, they often struggle to handle high-resolution videos in real-time, which can lead to stuttering and reduced smoothness. Among the methods listed, only STLVQE can achieve FPS above 40 on 720P resolution videos, making it suitable for most online scenarios. Compared to the second fastest method DCAD, STLVQE's processing time on 720p video is 9.6 milliseconds faster and outperforms it by a considerable margin, improving PSNR and SSIM metrics by 0.17dB and " + }, + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "inline_equation", + "content": "0.36 \\times 10^{-2}" + }, + { + "bbox": [ + 130, + 534, + 482, + 641 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Tab. 2 shows the latency of some large-scale high-latency temporal VQE methods. In contrast, high-latency VQE methods require waiting for future" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 211, + 157, + 404, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Ablation study of Temporal Cache(TC), Reference Frame Window(RFW), Deform Operation(DF), and Preliminary Complement Residual(PC). The latter three sections share the features extracted by MAFE." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 211, + 157, + 404, + 239 + ], + "lines": [ + { + "bbox": [ + 211, + 157, + 404, + 239 + ], + "spans": [ + { + "bbox": [ + 211, + 157, + 404, + 239 + ], + "type": "table", + "html": "
RFWTCDFPCΔP/ΔSTimeParams(K)
✓✓0.06/0.0110.85<1
✓✓✓0.31/0.5121.85352
✓✓✓0.39/0.8720.12149
✓✓✓0.49/1.0345.61398
✓✓✓0.46/0.9824.32398
✓✓✓✓0.49/1.0324.65398
", + "image_path": "740f685a70059ef27c94b88227a643ac21765737ec2d16c738899e9355fb8de6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 354 + ], + "type": "text", + "content": "frames or taking a long time for inference, which can cause uncontrollable video delay. The latency of the two high-latency methods in Tab. 1 we evaluated exceeds " + }, + { + "bbox": [ + 130, + 259, + 481, + 354 + ], + "type": "inline_equation", + "content": "400\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 259, + 481, + 354 + ], + "type": "text", + "content": ", much higher than the latency caused by STLVQE. Moreover, when considering only the FPS metric, STLVQE achieves a significantly faster speed than the single-frame VQE method RNAN(FPS = 0.2) by a factor of 175 on 720P video. Compared with MFQE 1.0, our method is 8.6 times faster with a better enhancement performance. Additionally, STLVQE has a more significant speed advantage on higher resolution videos such as 1080P and 2K." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "content": "Enhancement performance. In Tab. 1 and Tab. 2, we present the performance of STLVQE compared to other commonly used or sota VQE methods for each video at " + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{QP} = 37" + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "content": " and for different QP compressed videos, respectively. Our method outperforms all the listed low-latency methods. In comparison with the high-latency VQE method, STLVQE reaches higher level with RNAN at " + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{QP} = 37" + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{QP} = 32" + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "content": ", with only " + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "content": " of RNAN's number of parameters and latency time (720P), respectively. The " + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 356, + 481, + 498 + ], + "type": "text", + "content": "SSIM of STLVQE is higher than that of the MFQE method for all five QPs, indicating that our method produces an enhanced image with a more stable image structure. The specific visualization results and analysis can be found in Qualitative Comparison section. STDF is a robust VQE method, but we believe that a stable latency and higher frame rate are more important for enhancing the user's viewing experience." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 571 + ], + "type": "text", + "content": "Memory Size & GPU Consumption. As shown in Tab. 2, STLVQE has space occupation of 398KB for the model and 228MB for the LUT (utilizing less than " + }, + { + "bbox": [ + 130, + 498, + 481, + 571 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 498, + 481, + 571 + ], + "type": "text", + "content": " memory of a low end GPU like GTX 1060), achieving real-time processing capability. Since most deep networks are replaced with ST-LUTs, inferencing STLVQE only needs 98 GFLOPs on a 720P frame. Note that GPU computing unit are more expensive than the memory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 586, + 239, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 239, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 239, + 600 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed new VQE framework, we perform ablation experiments and specific analysis on three components. The results of the runtime per frame is computed on 720P videos at " + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{QP} = 37" + }, + { + "bbox": [ + 130, + 605, + 481, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of the propagation module. The results are shown in the 5th and 6th row of Tab. 3. We found that selecting frames based on QP val-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 102 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 194, + 146, + 420, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study of STLUTs and the CNN enhancement network. The runtime is computed on a 720P video." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 146, + 420, + 217 + ], + "lines": [ + { + "bbox": [ + 194, + 146, + 420, + 217 + ], + "spans": [ + { + "bbox": [ + 194, + 146, + 420, + 217 + ], + "type": "table", + "html": "
Enhancement ModuleΔP/ΔSTime(ms)Params(K)
S-LUT0.39/0.8323.52398
S-LUT + T-LUT-10.45/1.0224.12398
ST-LUTs0.47/1.0224.65398
ST-LUTs (on x̂i)0.49/1.0324.65398
CNN Network0.50/1.08147.02442
", + "image_path": "6d3ec5921213555a8175319b118cb2622f944bd3cca3d5c22bdba3b8fa0b4ccf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "text", + "content": "ues in RFW can bring a 0.03dB " + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "text", + "content": "PSNR performance improvement compared to randomly selecting frames in the RFW, without adding parameters to the network and only increasing the inference time by " + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "inline_equation", + "content": "0.33\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "text", + "content": " for comparing QP values of each frame. Furthermore, as shown in rows 4 and 6 of Tab. 3, the use of the TC structure can significantly reduce the inference time by " + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "inline_equation", + "content": "21.0~\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 242, + 482, + 338 + ], + "type": "text", + "content": ", which is consumed by the repeated computation of the reference frame information. We achieve this reduction in time by merely occupying the GPU memory with information of several frames in the RFW." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "content": "Effectiveness of the alignment module. We mainly validate the effect of the deform operation of the reference fram in the Alignment module. As shown in the first row of Tab. 3, when only using the ST-LUTs structure and discard the whole alignment module and PC operation, the result only has a " + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "inline_equation", + "content": "0.06\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "content": " PSNR improvement compared to HEVC video. When the Deform operation is applied, the network can capture more accurate temporal information, and the " + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "content": "PSNR reaches " + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "inline_equation", + "content": "0.31~\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "content": ". Due to the lightweight convolutional layers, the alignment module only adds " + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "inline_equation", + "content": "11.6~\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "content": " of inference time to the network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "text", + "content": "Effectiveness of the enhancement module. As shown in row 3 of Tab. 3, using PC without the Deform operation achieves a significant " + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "text", + "content": "PSNR of 0.39 dB, demonstrating the effectiveness of PC in providing initial compensation for spatial information and extending the receptive field, ultimately enhancing the performance of ST-LUTs. Finally, after combining two parts, the network achieves 0.49 dB " + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "text", + "content": "PSNR. PC alone contributes 0.18 dB to the network and adds only 2.16 ms of inference time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Tab. 4 provides an analysis of the impact of each LUT in the enhancement module. In the first three rows of the table, we attempted to enhance the original image " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " with the results of different LUT combinations. Results in the first and second rows show that using only the S-LUT with T-LUT-1 does not fully improve the network due to T-LUT-1 not utilizing all pixels within a kernel when learning the alignment offset. However, when T-LUT-2 is used for network training along with T-LUT-1, the reference frame can be better aligned, resulting in a " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "PSNR improvement of 0.08 dB. The 4th row shows the enhancing result on " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{i}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", reducing the learning difficulty of the ST-LUTs, which improving the performance by 0.02 dB. Compared to the CNN layers replaced by LUTs, ST-LUTs sacrifice only 0.01 dB " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "PSNR performance while reducing the number of parameters by 44K and the inference time by 122.37 ms." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 481, + 281 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 281 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 281 + ], + "type": "image", + "image_path": "efb0ef0e07b8df0c0315bc1d9b5af99ab281b1d1b5a33aa3a4461799027fa0b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 323 + ], + "type": "text", + "content": "Fig. 5: Qualitative results at QP 37. The STLVQE method successfully enhances the left leg of the athlete, the face of the old man, the railings of the bridge and the eyebrow of the woman in each frame, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "content": "The above ablation studies have sufficiently demonstrated that the MAFE structure of shared features in STLVQE with the redesigned three new modules can sufficiently reduce the computational redundancy of the network and obtain satisfactory inference speed-performance trade-offs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 280, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 280, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 280, + 419 + ], + "type": "text", + "content": "4.5 Qualitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "text", + "content": "Fig. 5 provides 4 examples of the qualitative results of our network on compressed video at " + }, + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "inline_equation", + "content": "QP = 37" + }, + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "text", + "content": ". The quality of these images deteriorates significantly in compressed scenes, showing a variety of image defects (e.g. blurring legs, missing bridge railings). With the help of temporal information, STLVQE achieves better visual results than single-frame methods with significantly reduced runtime. Compared to large-scale multi-frame methods like MFQE 2.0, STLVQE can achieve competitive capabilities in detail and structural restoration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a lightweight and rapid method called STLVQE based on the constraints of online video quality enhancement tasks. STLVQE greatly reduce the computation redundancy and uses a Spatial-Temporal Look-up Table structure to extract spatio-temporal information from the video with minimal time consumption. The extensive experiments prove that our proposed method can outperform most low-latency VQE methods, achieve competitive performance with high-latency VQE methods which cannot be directly utilized on Online-VQE tasks. Furthermore, STLVQE achieves real-time processing of 720P resolution videos, demonstrating a good trade-off between enhancement performance and inference speed." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": "This work was supported by National Natural Science Fund of China (62076184, 61976158, 61976160, 62076182, 62276190), in part by Fundamental Research Funds for the Central Universities and State Key Laboratory of Integrated Services Networks (Xidian University), in part by Shanghai Innovation Action Project of Science and Technology (20511100700) and Shanghai Natural Science Foundation (22ZR1466700)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 236 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 244, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 244, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 244, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 244, + 481, + 289 + ], + "type": "text", + "content": "1. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021) 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 289, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 332 + ], + "type": "text", + "content": "2. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5972-5981 (2022) 2, 3, 4, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 332, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 373 + ], + "type": "text", + "content": "3. Charbonnier, P., Blanc-Feraud, L., Aubert, G., Barlaud, M.: Two deterministic half-quadratic regularization algorithms for computed imaging. In: Proceedings of 1st international conference on image processing. vol. 2, pp. 168-172. IEEE (1994) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 375, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 481, + 418 + ], + "type": "text", + "content": "4. Chen, H., He, X., Qing, L., Xiong, S., Nguyen, T.Q.: Dpw-sdnet: Dual pixel-wavelet domain deep cnns for soft decoding of JPEG-compressed images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 711-720 (2018) 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 418, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 461 + ], + "type": "text", + "content": "5. Dai, Y., Liu, D., Wu, F.: A convolutional neural network approach for post-processing in hevc intra coding. In: MultiMedia Modeling: 23rd International Conference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23. pp. 28-39. Springer (2017) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 461, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 481, + 494 + ], + "type": "text", + "content": "6. Deng, J., Wang, L., Pu, S., Zhuo, C.: Spatio-temporal deformable convolution for compressed video quality enhancement. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 10696-10703 (2020) 2, 3, 4, 5, 9, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 494, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 481, + 525 + ], + "type": "text", + "content": "7. Ding, Q., Shen, L., Yu, L., Yang, H., Xu, M.: Patch-wise spatial-temporal quality enhancement for hevc compressed video. IEEE Transactions on Image Processing 30, 6459-6472 (2021) 3, 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "text", + "content": "8. Dong, C., Deng, Y., Loy, C.C., Tang, X.: Compression artifacts reduction by a deep convolutional network. In: Proceedings of the IEEE international conference on computer vision. pp. 576-584 (2015) 3, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "type": "text", + "content": "9. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. arXiv preprint arXiv:1902.08153 (2019) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 579, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 481, + 620 + ], + "type": "text", + "content": "10. Guan, Z., Xing, Q., Xu, M., Yang, R., Liu, T., Wang, Z.: Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE transactions on pattern analysis and machine intelligence 43(3), 949-963 (2019) 2, 3, 4, 5, 9, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "text", + "content": "11. Guo, J., Chao, H.: Building dual-domain representations for compression artifacts reduction. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 628-644. Springer (2016) 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 159 + ], + "type": "text", + "content": "12. Jin, Z., An, P., Yang, C., Shen, L.: Quality enhancement for intra frame coding via cnns: An adversarial approach. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1368-1372. IEEE (2018) 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "13. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 691-700 (2021) 3, 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 214 + ], + "type": "text", + "content": "14. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 257 + ], + "type": "text", + "content": "15. Kondratyuk, D., Yuan, L., Li, Y., Zhang, L., Tan, M., Brown, M., Gong, B.: Movinets: Mobile video networks for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16020-16030 (2021) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "text", + "content": "16. Li, J., Chen, C., Cheng, Z., Xiong, Z.: Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVIII. pp. 238-256. Springer (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 332 + ], + "type": "text", + "content": "17. Li, K., Bare, B., Yan, B.: An efficient deep convolutional neural networks model for compressed image deblocking. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 1320-1325. IEEE (2017) 3, 10, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 481, + 365 + ], + "type": "text", + "content": "18. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019) 2, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 481, + 385 + ], + "type": "text", + "content": "19. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 385, + 481, + 418 + ], + "type": "text", + "content": "20. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep kalman filtering network for video compression artifact reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 418, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 481, + 460 + ], + "type": "text", + "content": "21. Ma, C., Zhang, J., Zhou, J., Lu, J.: Learning series-parallel lookup tables for efficient image super-resolution. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII. pp. 305-321. Springer (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 460, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 460, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 133, + 460, + 481, + 503 + ], + "type": "text", + "content": "22. Maggioni, M., Huang, Y., Li, C., Xiao, S., Fu, Z., Song, F.: Efficient multistage video denoising with recurrent spatio-temporal fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3466-3475 (2021) 2, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 503, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 503, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 503, + 481, + 525 + ], + "type": "text", + "content": "23. MMEditing Contributors: MMEditing: OpenMMLab image and video editing toolbox. https://github.com/open-mmlab/mmediting (2022) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 481, + 557 + ], + "type": "text", + "content": "24. Sullivan, G.J., Ohm, J.R., Han, W.J., Wiegand, T.: Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology 22(12), 1649-1668 (2012) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 557, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 481, + 599 + ], + "type": "text", + "content": "25. Wang, T., Li, Y., Peng, J., Ma, Y., Wang, X., Song, F., Yan, Y.: Real-time image enhancer via learnable spatial-aware 3d lookup tables. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2471-2480 (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 600, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 481, + 633 + ], + "type": "text", + "content": "26. Wang, T., Chen, M., Chao, H.: A novel deep learning-based method of improving coding efficiency from the decoder-end for hevc. In: 2017 Data Compression Conference (DCC). pp. 410-419. IEEE (2017) 10, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "type": "text", + "content": "27. Wiegand, T., Sullivan, G.J., Bjontegaard, G., Luthra, A.: Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology 13(7), 560-576 (2003) 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 402 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "28. Xu, Y., Gao, L., Tian, K., Zhou, S., Sun, H.: Non-local convlstm for video compression artifact reduction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7043-7052 (2019) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "29. Yang, R., Xu, M., Liu, T., Wang, Z., Guan, Z.: Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology 29(7), 2039-2054 (2018) 2, 3, 4, 10, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "text", + "content": "30. Yang, R., Xu, M., Wang, Z.: Decoder-side hevc quality enhancement with scalable convolutional neural network. In: 2017 IEEE International Conference on Multimedia and Expo (ICME). pp. 817-822. IEEE (2017) 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "31. Yang, R., Xu, M., Wang, Z., Li, T.: Multi-frame quality enhancement for compressed video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6664-6673 (2018) 2, 3, 4, 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "32. Yin, G., Qu, Z., Jiang, X., Jiang, S., Han, Z., Zheng, N., Yang, H., Liu, X., Yang, Y., Li, D., et al.: Online streaming video super-resolution with convolutional lookup table. IEEE Transactions on Image Processing 33, 2305-2317 (2024) 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "text", + "content": "33. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "34. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 3, 10, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "35. Zhang, Y., Li, K., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082 (2019) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "36. Zhao, M., Xu, Y., Zhou, S.: Recursive fusion and deformable spatiotemporal attention for video compression artifact reduction. In: Proceedings of the 29th ACM international conference on multimedia. pp. 5646-5654 (2021) 2, 3, 4, 5, 9, 11" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "Online Video Quality Enhancement with Spatial-Temporal Look-up Tables" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_content_list.json b/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aba6e7f5c65128151a05247078d988d8fafed1de --- /dev/null +++ b/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_content_list.json @@ -0,0 +1,1769 @@ +[ + { + "type": "text", + "text": "Online Zero-Shot Classification with CLIP", + "text_level": 1, + "bbox": [ + 259, + 141, + 743, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Qian $^{1}$ and Juhua Hu $^{2}$", + "bbox": [ + 393, + 189, + 609, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Alibaba Group, Bellevue, WA 98004, USA \n2 School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA \nqi.qian@alibaba-inc.com, juhuah@uw.edu", + "bbox": [ + 325, + 215, + 673, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision-language pre-training such as CLIP enables zero-shot transfer that can classify images according to the candidate class names. While CLIP demonstrates an impressive zero-shot performance on diverse downstream tasks, the distribution from the target data has not been leveraged sufficiently. In this work, we study a novel online zero-shot transfer scenario, where each image arrives in a random order for classification and is visited only once to obtain prediction immediately without storing its representation. Compared with the vanilla zero-shot classification, the proposed framework preserves its flexibility for online service while considering the statistics of the arrived images as the side information to capture the distribution of target data, which can help improve the performance of real-world applications. To tackle the challenge of effective online optimization, we first develop online label learning to model the target data distribution. Then, the proxy of each class in the vision space is further optimized with the proposed online proxy learning method to mitigate the modality gap between images and text. The convergence of both online strategies can be theoretically guaranteed. By combining the predicted label from the online label learning and proxy learning, our online zero-shot transfer method (OnZeta) achieves $78.94\\%$ accuracy on ImageNet without accessing the entire data set. Moreover, extensive experiments on other 13 downstream tasks with different vision encoders show a more than $3\\%$ improvement on average, which demonstrates the effectiveness of our proposal.", + "bbox": [ + 261, + 301, + 743, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Online learning $\\cdot$ Zero-shot classification $\\cdot$ CLIP", + "bbox": [ + 261, + 632, + 669, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 667, + 375, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language pre-training has attracted much attention recently due to its impressive zero-shot transfer performance on various downstream tasks. The desired property mainly comes from aligning the vision and text space. For example, one of the most prevalent pre-training methods, i.e., CLIP [17], consists of distinct vision and text encoders for learning image and text representations, respectively. These encoders are optimized by minimizing a contrastive loss defined on image-text pairs. The loss aims to pull images and their corresponding text descriptions together while pushing the irrelevant text or images away [15].", + "bbox": [ + 212, + 696, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6c7bd7fc911e1039ee20cfd016219ef8d5d01bfed9ffdcc7c4133fcb6b5e486.jpg", + "image_caption": [ + "Fig. 1: Illustration of the proposed online zero-shot transfer method (OnZeta). Blue and orange lines denote the inference in the text and vision space, respectively. By incorporating the predictions from online label learning and online proxy learning, OnZeta can leverage the biased prediction from the text space to reduce the variance in the target vision space in an online manner." + ], + "image_footnote": [], + "bbox": [ + 308, + 143, + 694, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After the alignment between vision and text, a novel zero-shot classification paradigm emerges to transfer knowledge in pre-trained models to diverse tasks. Concretely, given all class names from the target task, each class can be represented by a class proxy that applies the text encoder in CLIP for the corresponding class name, which is also termed a text proxy. When an image arrives, its representation is extracted by the vision encoder of CLIP. Then, the label can be assigned by the 1-nearest neighbor (1-NN) classifier with the set of text proxies as the reference. Without any fine-tuning, this simple strategy achieves $77.02\\%$ accuracy on ImageNet [18] using the pre-trained ViT [4] as the vision encoder.", + "bbox": [ + 212, + 371, + 787, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the success of CLIP, many methods have been developed to explore different side information from the target task to further improve the transfer performance. Firstly, if a few labeled images are available from the target task, the zero-shot transfer can be formulated as a few-shot learning problem and the input text prompt can be learned with the labeled vision data by fine-tuning [26]. Secondly, if the computational resource is too limited to do fine-tuning, the labeled data can be cached as a vision classifier to classify images directly [24]. However, supervised information is often rare, while unsupervised information is more accessible. Therefore, many efforts have been devoted to exploiting the unlabeled images to enhance the zero-shot performance. [19] generates multiple augmented examples from a given image, and then learns the specific text prompt for each image as few-shot learning accordingly. [16] considers reconstructing the class proxy in the vision space, i.e., vision proxy, by leveraging a set of unlabeled images in proxy learning. While [16] shows superior zero-shot performance, it optimizes proxies in an offline manner. Due to the privacy issue, gathering or keeping a set of unlabeled images is still challenging in some real-world applications.", + "bbox": [ + 212, + 522, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this work, we investigate a novel practical scenario for zero-shot transfer, termed online zero-shot transfer. Concretely, when streaming images arrive as in conventional zero-shot learning, the model has to classify the image immediately without refining. After that, the class proxy can be updated but", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 398, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/121f387901d4124e3440254aca3c273c8ed4ef6a511d5a4c5faca2af3d5e32c6.jpg", + "image_caption": [ + "Fig. 2: Top 4 predicted labels with corresponding probabilities from baseline zero-shot method by CLIP and our proposal." + ], + "image_footnote": [], + "bbox": [ + 284, + 146, + 715, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the representations of arrived images cannot be kept. Unlike [16] that can access the whole unlabeled set, this online setting is way more challenging, where only statistics of seen images can be leveraged for optimization and each image is visited only once. Thereafter, the standard iterative optimization in [16] is not applicable, while the problem (i.e., online service to classify each unlabeled image on the fly without storing data) is ubiquitous in real-world applications (e.g., mobiles and robots).", + "bbox": [ + 212, + 318, + 784, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the online service challenge, we first consider exploring the data distribution of the target task to improve the prediction from the text proxy. Note that the vanilla zero-shot prediction focuses on individual images and ignores the distribution over the entire data set. To capture the distribution of all data, we develop an online label learning algorithm to balance the assignment between different classes globally. In addition, the class proxy from the text (i.e., text proxy) can be biased for the vision data due to the modality gap, and thus we propose online proxy learning to obtain the class proxy in the vision space (i.e., vision proxy) directly in an online manner. Finally, the prediction from both the text space and vision space can be combined to help reduce the variance of learning in the vision space as illustrated in the framework of our online zero-shot transfer (OnZeta) shown in Fig. 1. More importantly, with online optimization, our proposal OnZeta can leverage the information from target data effectively and efficiently, which helps correct the prediction compared to the baseline zero-shot method. For example, Fig. 2 shows that an image of \"Husky\" will be misclassified by vanilla CLIP. By considering the distribution over the whole data set and incorporating the learned vision proxy, our method can obtain the right label in an online manner. The main contributions of this work can be summarized as follows.", + "bbox": [ + 212, + 425, + 787, + 710 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Considering the practical need for online service (i.e., classifying each image on the fly without storage) and improving the performance of vanilla zero-shot classification, we study a novel setting, that is, online zero-shot transfer, using pre-trained vision-language models.", + "- To address the online service challenge, we propose an online label learning method to trace the distribution of assigned labels over the entire data in an online manner. Moreover, an online vision proxy learning method is proposed" + ], + "bbox": [ + 225, + 732, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Online Zero-Shot Classification with CLIP", + "bbox": [ + 447, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to further improve the class proxy in vision space for classification. The convergence of both methods can be theoretically guaranteed.", + "bbox": [ + 241, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Experiments on 14 downstream tasks demonstrate the effectiveness of our proposal OnZeta. Concretely, when visiting each image only once, OnZeta outperforms the baseline by a margin of more than $3\\%$ on average without storing any unlabeled data.", + "bbox": [ + 225, + 176, + 784, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 260, + 387, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP consists of two encoders (i.e., vision and text), and the optimization can be conducted on each of them individually. In this section, we briefly review research efforts in these two directions.", + "bbox": [ + 212, + 292, + 784, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Optimization with Text Encoder", + "text_level": 1, + "bbox": [ + 215, + 359, + 529, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP demonstrates that the zero-shot performance heavily depends on the text prompts for generating appropriate class proxies, and many methods investigate learning the optimal text prompt as input to the text encoder. CoOp [26] marks the input text prompts as learnable variables and optimizes them via few-shot learning with a limited number of labeled images for each class. CoCoOp [25] further includes the vision representation as the conditional context for prompt learning. TPT [19] relaxes the requirement of labeled examples and optimizes the consistency between different augmentations from the same image for prompt learning. However, it has to optimize the prompt for each image with 64 augmentations, which is inefficient for computation and memory, making it infeasible for real-time applications as the online scenario in this work.", + "bbox": [ + 212, + 385, + 785, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Optimization with Vision Encoder", + "text_level": 1, + "bbox": [ + 215, + 573, + 542, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While prompt learning can improve the performance on the target data, recent research shows that optimizing the vision representation can be more effective for classification [6,16]. CLIP-Adapter [6] introduces an additional adapter network after the vision encoder and optimizes the vision representation with few-shot learning. Although a similar adapter can be attached for the text encoder, the empirical study shows that optimizing the vision space is sufficient for transfer and can achieve better performance than either the version with only a text adapter or that with both vision and text adapters. [22] shows that fine-tuning the vision classifier directly with few-shot learning can achieve better performance than CoOp. Recently, the analysis in [16] demonstrates that the optimal class proxy for visual classification lives in the vision space. To mitigate the modality gap, [16] proposes proxy learning that aims to learn class proxy for each class in vision space. However, it requires a set of unlabeled data and will learn proxies iteratively. In this work, we also aim to refine the class proxy in the vision space, but with the learned labels in an online manner, which is more applicable for online real applications.", + "bbox": [ + 212, + 598, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 397, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Online Zero-shot Transfer", + "text_level": 1, + "bbox": [ + 215, + 143, + 501, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Zero-shot Transfer in CLIP", + "text_level": 1, + "bbox": [ + 215, + 176, + 488, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a data set consisting of image-text pairs as $\\{I_i, T_i\\}_{i=1}^n$ , where $I_i$ denotes an image and $T_i$ is the corresponding text such as class names or captions. CLIP trains vision and text encoders jointly to align the vision subspace and its corresponding text subspace. Let $f(\\cdot)$ and $g(\\cdot)$ denote the vision encoder and text encoder, respectively. The representations for images and text can be extracted as $\\mathbf{x}_i = f(I_i)$ and $\\mathbf{z}_i = g(T_i)$ .", + "bbox": [ + 212, + 202, + 787, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After pre-training two encoders, zero-shot classification can be implemented by the 1-nearest neighbor (1-NN) classifier. Concretely, given the class names of the target task, the proxy of $j$ -th class $\\mathbf{z}_j$ can be obtained by applying the text encoder for the prompt as \"a photo of a {class name}\". Then, the image $I_i$ will be classified as", + "bbox": [ + 212, + 294, + 787, + 369 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} = \\arg \\max _ {j} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 380, + 785, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the conventional zero-shot paradigm demonstrates an impressive transfer performance, the side information from the target task has not been explored sufficiently. Recent work shows that with only a set of unlabeled target data, the zero-shot performance can be improved dramatically [16]. In this work, we further relax the requirement of unlabeled target data and investigate a novel online zero-shot transfer scenario, where each unlabeled image arrives in an online manner without storing.", + "bbox": [ + 212, + 415, + 787, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Online Label Learning", + "text_level": 1, + "bbox": [ + 215, + 544, + 447, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we develop an online label learning method to leverage the information from arrived images to capture the distribution over the entire set.", + "bbox": [ + 212, + 571, + 785, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given text proxy $\\mathbf{z}_j$ , the predicted distribution over classes for $I_i$ can be computed as", + "bbox": [ + 212, + 601, + 785, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {j} / \\tau_ {T}\\right)}{\\sum_ {k} ^ {C} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {k} / \\tau_ {T}\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 643, + 785, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{T}$ is the temperature optimized by CLIP.", + "bbox": [ + 214, + 689, + 566, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If the whole image set is available, the label can be refined by considering the distribution over classes. Let $\\{p_i\\}$ denote the learnable labels, and the optimization problem with $C$ classes can be cast as", + "bbox": [ + 212, + 705, + 785, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {p _ {i} \\in \\Delta} \\frac {1}{n} \\sum_ {i} D _ {K L} \\left(p _ {i} \\| q _ {i}\\right) \\quad s. t. \\quad \\forall j, \\quad \\frac {1}{n} \\sum_ {i} p _ {i, j} \\geq \\frac {\\alpha}{C} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 762, + 785, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i = [p_{i,1}, \\ldots, p_{i,C}]$ , $q_i = [q_{i,1}, \\ldots, q_{i,C}]$ and $\\Delta$ denotes the simplex as $\\Delta = \\{p_i \\in \\mathbb{R}^C | \\sum_j^C p_{i,j} = 1; \\forall j, p_{i,j} \\geq 0\\}$ . The constraint bounds the minimal", + "bbox": [ + 214, + 806, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Online Zero-Shot Classification with CLIP", + "bbox": [ + 446, + 114, + 732, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Online Label Learning (OnLab)" + ], + "code_body": "1: Input: Unlabeled image set $\\{I_i\\}$ , class names $\\{T_j\\}$ , pre-trained CLIP encoders $(f, g)$ , dual variables $\\{\\rho_j\\}$ , ratio $\\alpha$ , temperature $\\tau_T$ , initial learning rate $c_{\\rho}$ \n2: Obtain representation of text proxy as $\\mathbf{z}_j = g(T_j)$ \n3: Initialize $\\forall j, \\rho_j = 0$ \n4: while the $i$ -th image arrives do \n5: Obtain vision representation as $\\mathbf{x}_i = f(I_i)$ \n6: Observe the probability $p_i^*$ by Eqn. 6 \n7: Have the prediction as $y_i = \\arg \\max_j p_{i,j}^*$ \n8: Update dual variables as in Eqn. 7 \n9: end while", + "bbox": [ + 222, + 166, + 785, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "size of each class to avoid collapsing and $\\alpha \\in [0,1]$ is the ratio for balancing the assignment between different classes. $\\alpha = 1$ denotes a well-balanced distribution that each class has the same proportion of images, while $\\alpha = 0$ keeps the prediction of the original zero-shot classification from $q_{i}$ .", + "bbox": [ + 212, + 335, + 782, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The problem can be solved effectively given all examples, but optimizing the objective with streaming images is challenging. To address the online problem, we introduce the Lagrange dual variables [2] $\\{\\rho_j\\}$ for constraints and rewrite the problem as", + "bbox": [ + 212, + 396, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\rho , p _ {i}) = \\max _ {\\rho_ {j} \\geq 0} \\min _ {p _ {i} \\in \\Delta} \\frac {1}{n} \\sum_ {i} D _ {K L} (p _ {i} | | q _ {i}) - \\sum_ {j} ^ {C} \\rho_ {j} \\left(\\frac {1}{n} \\sum_ {i} p _ {i, j} - \\frac {\\alpha}{C}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 464, + 784, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When the $i$ -th example arrives for prediction, the sub-problem with fixed dual variables becomes", + "bbox": [ + 212, + 513, + 782, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\prime} \\left(p _ {i}\\right) = \\min _ {p _ {i} \\in \\Delta} D _ {K L} \\left(p _ {i} \\mid \\mid q _ {i}\\right) - \\sum_ {j} ^ {C} \\rho_ {j} ^ {i - 1} p _ {i, j} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 551, + 784, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\rho_j^{i - 1}$ indicates the dual variable used for the $(i - 1)$ -th arrived image.", + "bbox": [ + 212, + 603, + 759, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fortunately, this problem has the closed-form solution as follows.", + "bbox": [ + 238, + 619, + 707, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 1 The optimal solution to the problem in Eqn. 5 is", + "bbox": [ + 214, + 641, + 686, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} ^ {*} = \\frac {q _ {i , j} \\exp \\left(\\rho_ {j} ^ {i - 1}\\right)}{\\sum_ {k} ^ {C} q _ {i , k} \\exp \\left(\\rho_ {k} ^ {i - 1}\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 664, + 784, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. It is from the K.K.T. condition [2].", + "bbox": [ + 215, + 710, + 519, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark Proposition 1 shows that the dual variable helps balance the assignment between classes. When $\\rho = 0$ without optimization, it degenerates to the original prediction from CLIP.", + "bbox": [ + 212, + 739, + 782, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the prediction for the image, the dual variable can be updated by gradient ascent as", + "bbox": [ + 212, + 785, + 782, + 815 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {j} ^ {i} = \\max \\left\\{0, \\rho_ {j} ^ {i - 1} - \\eta_ {\\rho} ^ {i} \\left(p _ {i, j} ^ {*} - \\alpha / C\\right) \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 821, + 784, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 398, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\eta_{\\rho}^{i}$ is the learning rate for the dual variable. When $\\{\\rho_j\\}$ is initialized as 0 and $\\alpha = 0$ , dual variables will not be updated, which leads to the original prediction.", + "bbox": [ + 212, + 145, + 782, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our online label learning (OnLab) algorithm is summarized in Alg. 1. The prediction will be obtained by a close-form solution and the updating for dual variables is also efficient, which makes the proposed method applicable for real-time applications.", + "bbox": [ + 212, + 191, + 782, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the efficiency, the effectiveness of OnLab can be theoretically guaranteed in the following theorem. The detailed proof can be found in the supplementary.", + "bbox": [ + 212, + 252, + 782, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 1 By running Alg. 1 for solving the problem in Eqn. 4 with the step size $\\eta_{\\rho}^{i} = c_{\\rho} / \\sqrt{i}$ , where $c_{\\rho}$ is a constant, the convergence can be guaranteed as", + "bbox": [ + 215, + 305, + 782, + 338 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{n} \\max _ {\\rho} \\sum_ {i} \\mathcal {L} \\left(\\rho , p _ {i} ^ {*}\\right) - \\min _ {p _ {i}} \\sum_ {i} \\mathcal {L} \\left(\\rho^ {i - 1}, p _ {i}\\right) \\leq \\mathcal {O} \\left(\\frac {1}{\\sqrt {n}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 345, + 785, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Online Proxy Learning", + "text_level": 1, + "bbox": [ + 215, + 400, + 450, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After optimizing the label by online label learning, we consider reconstructing the class proxy in the vision space to reduce the modality gap from the text proxy as suggested in [16]. Given the whole set of images, vision proxy learning can be cast as an optimization problem", + "bbox": [ + 212, + 422, + 782, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {w}} \\mathcal {L} (\\mathbf {w}) = - \\sum_ {i} \\sum_ {j} p _ {i, j} ^ {*} \\log \\left(p _ {i, j} ^ {\\prime}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 493, + 785, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_{i,j}^{*}$ is the output of online label learning and", + "bbox": [ + 215, + 534, + 589, + 551 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} ^ {\\prime} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j} / \\tau_ {I}\\right)}{\\sum_ {k} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {k} / \\tau_ {I}\\right)} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 561, + 785, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbf{w}_j$ denotes the learnable vision proxy for the $j$ -th class. $\\tau_I$ is the temperature that should be larger than $\\tau_T$ to obtain the class proxy in the vision space [16].", + "bbox": [ + 212, + 603, + 782, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When solving the problem in an online manner, only a single example will be received at each iteration. Therefore, the vision proxy will be updated according to the gradient from the $i$ -th example as", + "bbox": [ + 212, + 633, + 782, + 678 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} ^ {i} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\mathbf {w} ^ {i - 1} - \\eta_ {w} ^ {i} \\nabla \\mathcal {L} \\left(\\mathbf {w} ^ {i - 1}, \\mathbf {x} _ {i}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 688, + 785, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\Pi(\\cdot)$ projects the updated vision proxy to the unit norm. Note that the loss function is convex in $\\mathbf{w}$ , and the convergence can be guaranteed by the standard online learning theory [7].", + "bbox": [ + 212, + 713, + 782, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 2 By updating $\\mathbf{w}$ as in Eqn. 11 and setting $\\eta_w^i = c_w / \\sqrt{i}$ , where $c_{w}$ is a constant, the regret can be bounded as", + "bbox": [ + 215, + 768, + 782, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{n} \\sum_ {i} \\mathcal {L} \\left(\\mathbf {w} ^ {i - 1}, \\mathbf {x} _ {i}\\right) - \\min _ {\\mathbf {w}} \\mathcal {L} (\\mathbf {w}) \\leq \\mathcal {O} \\left(\\frac {1}{\\sqrt {n}}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 808, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Online Zero-Shot Classification with CLIP", + "bbox": [ + 446, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Online Zero-shot Transfer (OnZeta)" + ], + "code_body": "1: Input: Unlabeled image set $\\{I_i\\}$ , class names $\\{T_j\\}$ , pre-trained CLIP encoders $(f, g)$ , dual variables $\\{\\rho_j\\}$ , vision proxy $\\{\\mathbf{w}_j\\}$ , ratio $\\alpha$ , temperature $\\tau_T$ and $\\tau_I$ , initial learning rate $c_{\\rho}$ and $c_w$ , weight parameter $\\beta$ \n2: Obtain representation of text proxy as $\\mathbf{z}_j = g(T_j)$ \n3: Initialize $\\forall j, \\rho_j = 0$ and $\\mathbf{w}_j = \\mathbf{z}_j$ \n4: while the $i$ -th image arrives do \n5: Obtain vision representation as $\\mathbf{x}_i = f(I_i)$ \n6: Observe the probability $p_i^*$ by Eqn. 6 \n7: Observe the probability $p_i'$ by Eqn. 10 \n8: Have the prediction as $y_i = \\arg \\max_j \\tilde{p}_{i,j}$ where $\\tilde{p}_i$ is from Eqn. 13 \n9: Update dual variables as in Eqn. 7 \n10: Update vision proxy as in Eqn. 11 \n11: end while", + "bbox": [ + 222, + 166, + 785, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since images arrive in a random order and only a single image will be leveraged for updating $\\mathbf{w}$ , the variance in learning can be large, especially for the first few iterations. Unlike offline optimization in [16], where the performance is evaluated after the final iteration, online learning has to accumulate the accuracy of predictions from different timestamps. Therefore, we consider combining the pseudo label $p_i^*$ from online label learning in text space to stabilize the prediction.", + "bbox": [ + 212, + 378, + 782, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the $i$ -th example, assume that $\\hat{p}_i$ is the ground-truth distribution in the vision space. Since $p_i^*$ is mainly from the text proxy, it is a biased estimation from the text space. While $p_i'$ is estimated by the vision proxy, the variance can be large due to the online updating. Therefore, these predictions can be mixed to trade-off bias from text space and variance from vision space. Concretely, let", + "bbox": [ + 212, + 486, + 784, + 561 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {i} = \\lambda_ {i} p _ {i} ^ {\\prime} + (1 - \\lambda_ {i}) p _ {i} ^ {*} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 574, + 784, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "then we have", + "bbox": [ + 215, + 604, + 313, + 616 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E [ \\| \\tilde {p} _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} ] \\\\ = E \\left[ \\lambda_ {i} ^ {2} \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} + (1 - \\lambda_ {i}) ^ {2} \\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2} + 2 \\lambda (1 - \\lambda) \\langle p _ {i} ^ {\\prime} - \\hat {p} _ {i}, p _ {i} ^ {*} - \\hat {p} _ {i} \\rangle \\right] \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 631, + 784, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the sake of simplicity, we assume that $p_i'$ is an unbiased estimation for $\\hat{p}_i$ such that $E[p_i'] = \\hat{p}_i$ . Then", + "bbox": [ + 212, + 681, + 782, + 712 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nE _ {w} \\left[ \\| \\tilde {p} _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\right] = \\lambda_ {i} ^ {2} \\underbrace {E \\left[ \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\right]} _ {\\text {v a r i a n c e}} + (1 - \\lambda_ {i}) ^ {2} \\underbrace {\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2}} _ {\\text {b i a s}} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 722, + 784, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To minimize the approximation error, the trade-off ratio $\\lambda_{i}$ has the optimal solution as", + "bbox": [ + 212, + 770, + 782, + 799 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i} ^ {*} = \\frac {\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2}}{\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2} + E [ \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} ]} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 809, + 784, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 398, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We find that the bias can be considered as a constant but the variance will be reduced with the learning of $\\mathbf{w}$ , which implies a monotonically increasing function for combining the pseudo labels from different spaces. While various weight functions for $\\lambda$ can be developed according to the analysis, for the $i$ -th iteration, we empirically set", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i} = \\beta \\sqrt {i / n} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 238, + 785, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\beta \\in [0,1]$ is a constant. The combination strategy helps online transfer when the vision proxy is not trained sufficiently.", + "bbox": [ + 212, + 268, + 785, + 299 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Incorporating online label learning and online proxy learning, the proposed online zero-shot transfer (OnZeta) algorithm is summarized in Alg. 2. All operations in the proposed method can be computed efficiently without significantly increasing the cost of conventional zero-shot transfer.", + "bbox": [ + 212, + 301, + 785, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 395, + 375, + 412 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the proposed method, we conduct experiments on 14 diverse downstream tasks. Considering that text prompts are essential for obtaining the appropriate text proxy for zero-shot transfer, we follow the common practice of including an ensemble of 7 text prompts as suggested in [17] to generate the text proxy for baseline and our method. The initial learning rates for dual variables and vision proxy are set as $c_{\\rho} = 20$ and $c_{w} = 0.5$ , respectively. The learning rate will be decayed as $\\eta_{\\rho}^{i} = c_{\\rho} / \\sqrt{i}$ and $\\eta_w^i = c_w / \\sqrt{i}$ where $i$ denotes the order of the incoming image. Unlabeled images will arrive in a random order. For the temperature, we fix $\\tau_T = 0.01$ that is the optimized value in CLIP and $\\tau_I = 0.04$ as suggested in [16]. All experiments are implemented on a single V100 GPU.", + "bbox": [ + 212, + 436, + 787, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 621, + 388, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The main parameters in our method are the ratio for class assignment $\\alpha$ and the ratio for pseudo label combination $\\beta$ . In this subsection, we will investigate the influence of these parameters along with the setting of online zero-shot transfer. The ablation experiments are conducted on ImageNet with ResNet-50 [8] as vision encoder in CLIP.", + "bbox": [ + 212, + 656, + 787, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of $\\alpha$ $\\alpha$ is the ratio to capture the distribution over the entire data set as in Eqn. 3. While the vanilla zero-shot transfer ignores the side information from the arrived examples, the proposed online label learning can leverage the information of distribution by $\\alpha$ . Since $\\alpha \\in [0,1]$ , we vary $\\alpha$ in $\\{0,0.4,0.6,0.8,1\\}$ and summarize the results in Tab. 1.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Online Zero-Shot Classification with CLIP", + "bbox": [ + 446, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/32daa4409868a49195f7eb823094287b1d551dc02294f4abda66e8f7d0354bc8.jpg", + "table_caption": [ + "Table 1: Effect of $\\alpha$ in Eqn. 3 for online label learning on ImageNet with ResNet-50 as vision encoder. \"Min class\" denotes the proportion of examples in the smallest class." + ], + "table_footnote": [], + "table_body": "
αAcc(%)#Min class(%)
060.3213.7
0.460.4118.5
0.660.6323.0
0.861.2033.3
161.9639.0
", + "bbox": [ + 256, + 261, + 455, + 354 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cc944ef65f3a7865e6f00d1e563f9803e7ea33de4f74b134f8ec0b2b8a1b96a8.jpg", + "image_caption": [ + "Fig.3: Illustration of data distribution over 1,000 classes with different $\\alpha$ on ImageNet. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 478, + 148, + 779, + 319 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "First, according to our analysis, $\\alpha = 0$ will not update dual variables and implies the accuracy of baseline zero-shot transfer as in CLIP. Then, increasing $\\alpha$ can consistently improve the accuracy over the baseline, which shows the efficacy of online label learning. Moreover, the size of the smallest class is also increased as $\\alpha$ increases and it shows that our proposed method OnLab with only the online label learning can adjust the class assignment effectively.", + "bbox": [ + 212, + 378, + 784, + 469 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To further investigate the data distribution derived by different $\\alpha$ , we compare the size of all 1,000 classes in Fig. 3. Evidently, the proposed method is capable of balancing the distribution in an online manner. Note that even with $\\alpha = 1$ , the obtained distribution is not perfectly balanced, which is from the approximation error in the online scenario as analyzed in Theorem 1. We find that our method is not sensitive to $\\alpha$ and fix $\\alpha = 1$ in the following experiments if not specified.", + "bbox": [ + 212, + 470, + 785, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Effect of $\\beta$ With the predictions from text and vision space in online zero-shot transfer, a combination strategy with $\\beta$ in Eqn. 17 can trade-off between the bias and variance. Tab. 2 reports the performance with different $\\beta$ .", + "bbox": [ + 214, + 595, + 785, + 643 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/701b101dbcc7bc34491b9e0d96148b04b84232baac29eb378c16885466044e20.jpg", + "table_caption": [ + "Table 2: Effect of $\\beta$ in Eqn. 17 for online proxy Table 3: Comparison of strate-learning on ImageNet with ResNet-50 as vision gies for obtaining the combination ratio $\\lambda$ in Eqn. 13." + ], + "table_footnote": [], + "table_body": "
β00.20.40.60.81
Acc(%)61.9662.1762.2362.3862.5061.84
", + "bbox": [ + 248, + 696, + 560, + 729 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e38ab53cf04e4f1e14ebfb90241220cae30c6afd0bec667ae11fd17c1d965781.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
βFixedDynamic
Acc(%)62.2862.50
", + "bbox": [ + 589, + 696, + 761, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "When $\\beta = 0$ , only the label from the online label learning (i.e., OnLab) is adopted for prediction. If mixing the label from the vision proxy, the accuracy improves. The mixture prediction outperforms OnLab by $0.54\\%$ with $\\beta = 0.8$ and is $2.18\\%$ better than the baseline CLIP. If excluding the label from text space at the end of data streaming as $\\beta = 1$ , the performance degenerates due to the large variance from the vision space. The phenomenon illustrates the challenge", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 398, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of online learning that the vision proxy may not be well learned with a limited number of arrived images and the pseudo label from text proxy can help reduce the variance.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "According to our analysis in Section 3.3, a dynamic ratio for the bias-variance trade-off is more effective for the proposed online learning scenario. To verify the claim, we compare the dynamic ratio to the fixed one in Tab. 3. The value of fixed $\\lambda$ is searched in [0, 1] by a step size of 0.1 and the best result is reported.", + "bbox": [ + 212, + 191, + 784, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "While the fixed $\\lambda$ still improves the performance from the online label learning, a monotonically increasing $\\lambda$ captures the training dynamics better and shows an additional gain of $0.22\\%$ . It confirms our analysis about the bias-variance trade-off.", + "bbox": [ + 212, + 252, + 784, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of iterations in optimization Our method has to learn the dual variables and vision proxy with streaming data and those parameters may not be learned well at the early stage of online learning. To investigate the influence of iterations in online learning, we report the accumulated accuracy of different iterations in one epoch on ImageNet in Tab. 4.", + "bbox": [ + 212, + 333, + 782, + 407 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3fe4ecb98abf634aba7c16f1263c3bcf6a58e5d082f3e5543e642533f82a32d1.jpg", + "table_caption": [ + "Table 4: Comparison of accu- Table 5: Effect of epochs for optimization on Imracy $(\\%)$ with different itera-ageNet. Prediction from the last epoch is evaluated tions on ImageNet. for OnZeta after running multiple epochs." + ], + "table_footnote": [], + "table_body": "
#Iters5k10k20k
Acc(%)60.9661.5061.71
30k40k50k
62.0262.4162.50
", + "bbox": [ + 243, + 489, + 416, + 549 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2af5c567c4d7cf7e1f7130741668f964de0fe42995e85bf1c74e4b34cc196ee7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Epoch1(online learning)2345
Acc(%)62.5063.2463.3563.3963.46
", + "bbox": [ + 436, + 489, + 772, + 520 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "First, we can observe that even with 5,000 iterations for learning, the accuracy over 1,000 classes is already better than the baseline. With the arrival of more images, the performance increases steadily. It is because that the convergence rate for dual variables and vision proxy is $\\mathcal{O}(1 / \\sqrt{n})$ as analyzed in Theorems 1-2. It implies that these variables can reach a reasonable performance with a limited number of iterations, while it is challenging to obtain the optimal solution as discussed in the following experiment.", + "bbox": [ + 212, + 578, + 784, + 683 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of epochs in optimization In the proposed online zero-shot transfer problem, all examples can be visited only once when they arrive. The setting is prevalent in real-world applications but very challenging for optimization. This ablation experiment aims to investigate the performance gain when each example can be visited multiple times.", + "bbox": [ + 212, + 704, + 784, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Concretely, we run the proposed method on multiple epochs of data, where images arrive in a random order within each epoch. For multiple epochs, the prediction from the last one is reported for evaluation. Tab. 5 compares the performance of multiple epochs to that of a single epoch in the online scenario.", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Online Zero-Shot Classification with CLIP", + "bbox": [ + 446, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "First, we can observe that with an additional epoch, the performance of OnZeta can be boosted by $0.74\\%$ , which demonstrates the challenge of online zero-shot transfer. When increasing the number of epochs, the performance can be further improved and it achieves $63.46\\%$ with 5-epoch learning. The result approaches the performance of learning with the entire data set as in [16] that shows a $63.74\\%$ accuracy when visiting all data by 2,000 epochs. The competitive performance of OnZeta confirms the effectiveness of the proposed online learning algorithm. More experiments can be found in the supplementary.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparison on ImageNet", + "text_level": 1, + "bbox": [ + 215, + 285, + 472, + 301 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "After the ablation study, we compare our method OnZeta to state-of-the-art methods on ImageNet. Multiple vision encoders from CLIP are adopted for evaluation, which include ResNet-50, ViT-B/32, ViT-B/16, ViT-L/14, and ViT-L/14@336. Considering that images come in a random order in our online setting, the proposed method is repeated by 5 times and the averaged result is reported for the comparison. We note that the standard deviation is quite small and also report the best result among different runs by *. The same parameters are shared by different vision encoders in our method. Tab. 6 summarizes the comparison, where the performance of InMaP [16] is marked by gray due to the application of the whole unlabeled data set. All other methods only access a single image at each iteration. Our OnLab with only the online label learning is denoted as \"OnZeta0.5\" and the vanilla zero-shot transfer in Eqn. 1 is \"Baseline\".", + "bbox": [ + 212, + 306, + 787, + 488 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5008fa4e580905e02ca6cdc268edd8488d7e340b3087d84185b23dcb38559326.jpg", + "table_caption": [ + "Table 6: Comparison of accuracy (%) on ImageNet with different vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration." + ], + "table_footnote": [], + "table_body": "
Vision encoderBaseline [17]TPT [19]OnZeta0.5OnZeta*0.5OnZetaOnZeta*InMaP [16]
ResNet-5060.3260.7461.8961.9662.4262.5063.74
ViT-B/3263.77-65.0665.1865.7165.8067.29
ViT-B/1668.7568.9870.1870.2370.8770.9972.55
ViT-L/1475.96-76.9877.0777.8477.9479.29
ViT-L/1433677.02-77.9378.0578.7578.9480.21
", + "bbox": [ + 220, + 594, + 785, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "First, by capturing the overall distribution using only the proposed online label learning component, $\\mathrm{OnZeta}_{0.5}$ already gains about $1\\%$ for all vision encoders over the baseline. It demonstrates that the side information from the arrived images is useful for improving the zero-shot transfer. By learning the vision proxy and combining the predictions from both the text and vision spaces, OnZeta shows about $1\\%$ additional improvement over $\\mathrm{OnZeta}_{0.5}$ , which illustrates that online proxy learning is complementary to online label learning for zero-shot classification. Moreover, OnZeta obtains more gains with a large backbone, showing its potential for large models.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 397, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Then, we can observe that the prompt learning method TPT outperforms the baseline with an ensemble of artificial text prompts, which indicates the importance of the appropriate prompt for zero-shot transfer. However, OnZeta surpasses TPT by $1.76\\%$ on ResNet-50 and $2.01\\%$ on ViT-B/16. It is because that we learn a better class proxy in the vision space for visual classification, which helps reduce the modality gap and is more effective than optimizing the text prompt in the text space. The observation is consistent with the analysis in [16]. Moreover, compared with InMaP that can access the whole unlabeled set, OnZeta is only about $1\\%$ worse using different vision encoders. As analyzed in the ablation study, online learning that visits each example once without storing is more challenging than the offline method with multiple iterations over the entire set. Nevertheless, online learning is practical for real-world applications. Finally, all vision encoders share the same parameters in experiments. It implies that our method OnZeta is not sensitive to the selection of vision encoder and is applicable in different configurations.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Comparison on Other 13 Downstream Tasks", + "text_level": 1, + "bbox": [ + 214, + 393, + 627, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Besides the comparison on ImageNet, we conduct experiments on other 13 downstream tasks to evaluate the performance of our method. The benchmark data sets for zero-shot transfer are adopted in the comparison, which includes Aircraft [12], Caltech101 [5], Stanford Cars [10], CIFAR-10 [11], CIFAR-100 [11], CUB200-2011 [21], Describable Textures Dataset (DTD) [3], EuroSAT [9], Flowers [13], Food101 [1], Oxford-IIIT Pet (Pets) [14], Sun397 [23], and UCF101 [20]. A broad range of downstream tasks are involved by these diverse data sets, e.g., low-resolution image classification, fine-grained visual categorization (FGVC), land cover classification with satellite images, scene categorization, texture recognition, etc. These tasks contain fewer images than ImageNet, which increases the challenge of our online transfer with the convergence rate of $\\mathcal{O}(1 / \\sqrt{n})$ . Therefore, we reduce $\\beta$ from 0.8 to 0.4 for all tasks to exploit the label from the text space. For data sets that the baseline already achieves satisfied performance, e.g., Caltech101 and CIFAR-10 with ViT, online label learning can be skipped by letting $\\alpha = 0$ , while $\\alpha = 0.4$ can provide a slight improvement. Other parameters on the rest data sets remain the same as those for ImageNet. Two distinct vision encoders, i.e., ResNet-50 and ViT-B/16, are applied for evaluation. The performance averaged over 5 random trials is reported for our method.", + "bbox": [ + 212, + 417, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The comparison is summarized in Tab. 7. A similar phenomenon as ImageNet can be observed on these downstream tasks. Concretely, online label learning, i.e., $\\mathrm{OnZeta}_{0.5}$ , already improves the performance over baselines, where the gain is $3.12\\%$ on ResNet and $2.58\\%$ on ViT, respectively. Compared with the baseline, only streaming images are leveraged by our method, and the representation of each arrived image is not stored, which preserves the flexibility of zero-shot transfer but captures the distribution over the entire data set in an online manner. In addition, learning the vision proxy with the pseudo label from the text space further improves the performance with different vision encoders, which demonstrates the effectiveness of learning the class proxy directly in the target", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Online Zero-Shot Classification with CLIP", + "bbox": [ + 446, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/178ea5c7b36eed8f75d9bc3ad55c6f354f9fa74787af6853dc9f7769170559a7.jpg", + "table_caption": [ + "Table 7: Comparison of accuracy (\\%) on 13 downstream tasks with ResNet-50 and ViT-B/16 as vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration." + ], + "table_footnote": [], + "table_body": "
AircraftCaltechCarsCifar10Cifar100CUBDTDEuroSATFlowersFoodPetsSUNUCF101Avg.
ResNet-50:
Baseline16.6286.0056.3173.1540.6041.3741.1326.9062.9774.1081.8559.2555.5655.06
TPT17.5887.0258.46---40.8428.3362.6974.8884.4961.4660.82-
OnZeta0.518.2986.1758.5775.6246.9645.6843.2433.7962.4276.4183.0561.6460.1257.84
OnZeta*0.518.5186.4558.9175.7547.2245.9343.7434.3263.1776.5383.5961.8160.3858.18
OnZeta18.3186.2659.4075.8447.5546.0643.3833.9963.1676.8383.7762.3261.0058.30
OnZeta*18.5486.4159.7175.9547.7146.3443.9134.3563.9576.9784.6662.4061.3858.64
InMaP18.9686.7363.3078.8449.2649.1744.8635.8866.6878.3687.9363.8263.3660.55
ViT-B/16:
Baseline23.1393.5166.2991.0967.2949.3645.0950.1766.9984.4387.0065.6865.2165.79
TPT24.7894.1666.87---47.7542.4468.9884.6787.7965.5068.04-
OnZeta0.527.1993.6567.9991.1369.8552.8147.5156.2068.4985.8388.4468.0668.5068.13
OnZeta*0.527.4893.7968.1191.1870.0253.0947.8756.4269.0685.9788.7268.2768.8668.37
OnZeta27.5793.6968.7491.3170.3953.6648.0456.5269.3786.2189.0168.8769.5868.69
OnZeta*28.2993.8369.0391.3570.5253.8048.5856.7469.6386.3589.3269.0169.9468.95
InMaP28.3593.5973.0093.2772.5157.8750.7763.9871.2887.7692.9470.8572.0671.40
", + "bbox": [ + 218, + 239, + 787, + 414 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "vision space. Moreover, OnZeta outperforms TPT on 9 out of 10 (only 10 tasks are available in TPT's original paper) data sets with different vision encoders. Compared with TPT that has to optimize the text prompt with multiple augmentations of the image by back-propagation through the text encoder, OnZeta only contains the forward pass to extract the representation of the image, which is more efficient and is capable of online deployment. By comparing the best performance of the proposed method (i.e., indicated by * in the table) to the averaged one, we can find that the gap averaged over 13 data sets is only about $0.3\\%$ . While the images arrive in a random order, the performance of the proposed method can be theoretically guaranteed as indicated in Theorems 1-2, which is robust to the order of input images. Finally, most data sets share the same parameters and it implies that OnZeta is not sensitive to hyper-parameters and is applicable for different tasks.", + "bbox": [ + 212, + 441, + 787, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 657, + 359, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While CLIP demonstrates impressive zero-shot transfer performance, the information from the target data has not been explored sufficiently. In this work, we investigate an online zero-shot scenario, where a random image arrives at each iteration and the model has to predict its label without storing its representation. By capturing the overall label distribution and refining the class proxy in an online manner, our method OnZeta demonstrates a substantial improvement over the vanilla zero-shot transfer on diverse downstream tasks with different vision encoders. In addition, the operations in the proposed method can be implemented efficiently for real-time applications. Exploring more effective algorithms for challenging online zero-shot transfer can be our future work.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 398, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014)", + "2. Boyd, S.P., Vandenberghe, L.: Convex Optimization. Cambridge University Press (2014)", + "3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014)", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)", + "5. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004)", + "6. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. CoRR abs/2110.04544 (2021)", + "7. Hazan, E.: Introduction to online convex optimization. CoRR abs/1909.05207 (2019)", + "8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)", + "9. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens. 12(7), 2217-2226 (2019)", + "0. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013)", + "1. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "2. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)", + "3. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008)", + "4. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012)", + "5. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019)", + "6. Qian, Q., Xu, Y., Hu, J.: Intra-modal proxy learning for zero-shot visual categorization with clip. In: NeurIPS (2023)", + "7. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)", + "8. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015)", + "9. Shu, M., Nie, W., Huang, D., Yu, Z., Goldstein, T., Anandkumar, A., Xiao, C.: Test-time prompt tuning for zero-shot generalization in vision-language models. In: NeurIPS (2022)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Online Zero-Shot Classification with CLIP", + "bbox": [ + 446, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR abs/1212.0402 (2012)", + "21. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "22. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. In: CVPR. pp. 7949-7961. IEEE (2022)", + "23. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010)", + "24. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of CLIP for few-shot classification. In: ECCV. vol. 13695, pp. 493-510. Springer (2022)", + "25. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16795-16804. IEEE (2022)", + "26. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vis. 130(9), 2337-2348 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 383 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Qian and J. Hu", + "bbox": [ + 271, + 114, + 397, + 128 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_model.json b/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4ba394f51e8c660852af48ea91750a0eaeb5a648 --- /dev/null +++ b/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_model.json @@ -0,0 +1,2102 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.142, + 0.744, + 0.162 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.19, + 0.61, + 0.206 + ], + "angle": 0, + "content": "Qi Qian\\(^{1}\\) and Juhua Hu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.217, + 0.674, + 0.274 + ], + "angle": 0, + "content": "1 Alibaba Group, Bellevue, WA 98004, USA \n2 School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA \nqi.qian@alibaba-inc.com, juhuah@uw.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.302, + 0.744, + 0.621 + ], + "angle": 0, + "content": "Abstract. Vision-language pre-training such as CLIP enables zero-shot transfer that can classify images according to the candidate class names. While CLIP demonstrates an impressive zero-shot performance on diverse downstream tasks, the distribution from the target data has not been leveraged sufficiently. In this work, we study a novel online zero-shot transfer scenario, where each image arrives in a random order for classification and is visited only once to obtain prediction immediately without storing its representation. Compared with the vanilla zero-shot classification, the proposed framework preserves its flexibility for online service while considering the statistics of the arrived images as the side information to capture the distribution of target data, which can help improve the performance of real-world applications. To tackle the challenge of effective online optimization, we first develop online label learning to model the target data distribution. Then, the proxy of each class in the vision space is further optimized with the proposed online proxy learning method to mitigate the modality gap between images and text. The convergence of both online strategies can be theoretically guaranteed. By combining the predicted label from the online label learning and proxy learning, our online zero-shot transfer method (OnZeta) achieves \\(78.94\\%\\) accuracy on ImageNet without accessing the entire data set. Moreover, extensive experiments on other 13 downstream tasks with different vision encoders show a more than \\(3\\%\\) improvement on average, which demonstrates the effectiveness of our proposal." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.633, + 0.671, + 0.647 + ], + "angle": 0, + "content": "Keywords: Online learning \\(\\cdot\\) Zero-shot classification \\(\\cdot\\) CLIP" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.669, + 0.377, + 0.685 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Vision-language pre-training has attracted much attention recently due to its impressive zero-shot transfer performance on various downstream tasks. The desired property mainly comes from aligning the vision and text space. For example, one of the most prevalent pre-training methods, i.e., CLIP [17], consists of distinct vision and text encoders for learning image and text representations, respectively. These encoders are optimized by minimizing a contrastive loss defined on image-text pairs. The loss aims to pull images and their corresponding text descriptions together while pushing the irrelevant text or images away [15]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.399, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.145, + 0.695, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the proposed online zero-shot transfer method (OnZeta). Blue and orange lines denote the inference in the text and vision space, respectively. By incorporating the predictions from online label learning and online proxy learning, OnZeta can leverage the biased prediction from the text space to reduce the variance in the target vision space in an online manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.522 + ], + "angle": 0, + "content": "After the alignment between vision and text, a novel zero-shot classification paradigm emerges to transfer knowledge in pre-trained models to diverse tasks. Concretely, given all class names from the target task, each class can be represented by a class proxy that applies the text encoder in CLIP for the corresponding class name, which is also termed a text proxy. When an image arrives, its representation is extracted by the vision encoder of CLIP. Then, the label can be assigned by the 1-nearest neighbor (1-NN) classifier with the set of text proxies as the reference. Without any fine-tuning, this simple strategy achieves \\(77.02\\%\\) accuracy on ImageNet [18] using the pre-trained ViT [4] as the vision encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Since the success of CLIP, many methods have been developed to explore different side information from the target task to further improve the transfer performance. Firstly, if a few labeled images are available from the target task, the zero-shot transfer can be formulated as a few-shot learning problem and the input text prompt can be learned with the labeled vision data by fine-tuning [26]. Secondly, if the computational resource is too limited to do fine-tuning, the labeled data can be cached as a vision classifier to classify images directly [24]. However, supervised information is often rare, while unsupervised information is more accessible. Therefore, many efforts have been devoted to exploiting the unlabeled images to enhance the zero-shot performance. [19] generates multiple augmented examples from a given image, and then learns the specific text prompt for each image as few-shot learning accordingly. [16] considers reconstructing the class proxy in the vision space, i.e., vision proxy, by leveraging a set of unlabeled images in proxy learning. While [16] shows superior zero-shot performance, it optimizes proxies in an offline manner. Due to the privacy issue, gathering or keeping a set of unlabeled images is still challenging in some real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Therefore, in this work, we investigate a novel practical scenario for zero-shot transfer, termed online zero-shot transfer. Concretely, when streaming images arrive as in conventional zero-shot learning, the model has to classify the image immediately without refining. After that, the class proxy can be updated but" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.147, + 0.716, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.258, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Fig. 2: Top 4 predicted labels with corresponding probabilities from baseline zero-shot method by CLIP and our proposal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.785, + 0.424 + ], + "angle": 0, + "content": "the representations of arrived images cannot be kept. Unlike [16] that can access the whole unlabeled set, this online setting is way more challenging, where only statistics of seen images can be leveraged for optimization and each image is visited only once. Thereafter, the standard iterative optimization in [16] is not applicable, while the problem (i.e., online service to classify each unlabeled image on the fly without storing data) is ubiquitous in real-world applications (e.g., mobiles and robots)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.712 + ], + "angle": 0, + "content": "To address the online service challenge, we first consider exploring the data distribution of the target task to improve the prediction from the text proxy. Note that the vanilla zero-shot prediction focuses on individual images and ignores the distribution over the entire data set. To capture the distribution of all data, we develop an online label learning algorithm to balance the assignment between different classes globally. In addition, the class proxy from the text (i.e., text proxy) can be biased for the vision data due to the modality gap, and thus we propose online proxy learning to obtain the class proxy in the vision space (i.e., vision proxy) directly in an online manner. Finally, the prediction from both the text space and vision space can be combined to help reduce the variance of learning in the vision space as illustrated in the framework of our online zero-shot transfer (OnZeta) shown in Fig. 1. More importantly, with online optimization, our proposal OnZeta can leverage the information from target data effectively and efficiently, which helps correct the prediction compared to the baseline zero-shot method. For example, Fig. 2 shows that an image of \"Husky\" will be misclassified by vanilla CLIP. By considering the distribution over the whole data set and incorporating the learned vision proxy, our method can obtain the right label in an online manner. The main contributions of this work can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.733, + 0.784, + 0.793 + ], + "angle": 0, + "content": "- Considering the practical need for online service (i.e., classifying each image on the fly without storage) and improving the performance of vanilla zero-shot classification, we study a novel setting, that is, online zero-shot transfer, using pre-trained vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- To address the online service challenge, we propose an online label learning method to trace the distribution of assigned labels over the entire data in an online manner. Moreover, an online vision proxy learning method is proposed" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.733, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "to further improve the class proxy in vision space for classification. The convergence of both methods can be theoretically guaranteed." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.785, + 0.237 + ], + "angle": 0, + "content": "- Experiments on 14 downstream tasks demonstrate the effectiveness of our proposal OnZeta. Concretely, when visiting each image only once, OnZeta outperforms the baseline by a margin of more than \\(3\\%\\) on average without storing any unlabeled data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.261, + 0.388, + 0.276 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.785, + 0.337 + ], + "angle": 0, + "content": "CLIP consists of two encoders (i.e., vision and text), and the optimization can be conducted on each of them individually. In this section, we briefly review research efforts in these two directions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.361, + 0.531, + 0.375 + ], + "angle": 0, + "content": "2.1 Optimization with Text Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.787, + 0.552 + ], + "angle": 0, + "content": "CLIP demonstrates that the zero-shot performance heavily depends on the text prompts for generating appropriate class proxies, and many methods investigate learning the optimal text prompt as input to the text encoder. CoOp [26] marks the input text prompts as learnable variables and optimizes them via few-shot learning with a limited number of labeled images for each class. CoCoOp [25] further includes the vision representation as the conditional context for prompt learning. TPT [19] relaxes the requirement of labeled examples and optimizes the consistency between different augmentations from the same image for prompt learning. However, it has to optimize the prompt for each image with 64 augmentations, which is inefficient for computation and memory, making it infeasible for real-time applications as the online scenario in this work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.544, + 0.588 + ], + "angle": 0, + "content": "2.2 Optimization with Vision Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.841 + ], + "angle": 0, + "content": "While prompt learning can improve the performance on the target data, recent research shows that optimizing the vision representation can be more effective for classification [6,16]. CLIP-Adapter [6] introduces an additional adapter network after the vision encoder and optimizes the vision representation with few-shot learning. Although a similar adapter can be attached for the text encoder, the empirical study shows that optimizing the vision space is sufficient for transfer and can achieve better performance than either the version with only a text adapter or that with both vision and text adapters. [22] shows that fine-tuning the vision classifier directly with few-shot learning can achieve better performance than CoOp. Recently, the analysis in [16] demonstrates that the optimal class proxy for visual classification lives in the vision space. To mitigate the modality gap, [16] proposes proxy learning that aims to learn class proxy for each class in vision space. However, it requires a set of unlabeled data and will learn proxies iteratively. In this work, we also aim to refine the class proxy in the vision space, but with the learned labels in an online manner, which is more applicable for online real applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.447, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.503, + 0.161 + ], + "angle": 0, + "content": "3 Online Zero-shot Transfer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.178, + 0.489, + 0.192 + ], + "angle": 0, + "content": "3.1 Zero-shot Transfer in CLIP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.295 + ], + "angle": 0, + "content": "Given a data set consisting of image-text pairs as \\(\\{I_i, T_i\\}_{i=1}^n\\), where \\(I_i\\) denotes an image and \\(T_i\\) is the corresponding text such as class names or captions. CLIP trains vision and text encoders jointly to align the vision subspace and its corresponding text subspace. Let \\(f(\\cdot)\\) and \\(g(\\cdot)\\) denote the vision encoder and text encoder, respectively. The representations for images and text can be extracted as \\(\\mathbf{x}_i = f(I_i)\\) and \\(\\mathbf{z}_i = g(T_i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.37 + ], + "angle": 0, + "content": "After pre-training two encoders, zero-shot classification can be implemented by the 1-nearest neighbor (1-NN) classifier. Concretely, given the class names of the target task, the proxy of \\( j \\)-th class \\( \\mathbf{z}_j \\) can be obtained by applying the text encoder for the prompt as \"a photo of a {class name}\". Then, the image \\( I_i \\) will be classified as" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.381, + 0.786, + 0.406 + ], + "angle": 0, + "content": "\\[\ny _ {i} = \\arg \\max _ {j} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.522 + ], + "angle": 0, + "content": "While the conventional zero-shot paradigm demonstrates an impressive transfer performance, the side information from the target task has not been explored sufficiently. Recent work shows that with only a set of unlabeled target data, the zero-shot performance can be improved dramatically [16]. In this work, we further relax the requirement of unlabeled target data and investigate a novel online zero-shot transfer scenario, where each unlabeled image arrives in an online manner without storing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.449, + 0.561 + ], + "angle": 0, + "content": "3.2 Online Label Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.787, + 0.601 + ], + "angle": 0, + "content": "First, we develop an online label learning method to leverage the information from arrived images to capture the distribution over the entire set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.787, + 0.632 + ], + "angle": 0, + "content": "Given text proxy \\(\\mathbf{z}_j\\), the predicted distribution over classes for \\(I_i\\) can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.644, + 0.786, + 0.681 + ], + "angle": 0, + "content": "\\[\nq _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {j} / \\tau_ {T}\\right)}{\\sum_ {k} ^ {C} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {k} / \\tau_ {T}\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.69, + 0.568, + 0.705 + ], + "angle": 0, + "content": "where \\(\\tau_{T}\\) is the temperature optimized by CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.787, + 0.75 + ], + "angle": 0, + "content": "If the whole image set is available, the label can be refined by considering the distribution over classes. Let \\(\\{p_i\\}\\) denote the learnable labels, and the optimization problem with \\(C\\) classes can be cast as" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.763, + 0.786, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\min _ {p _ {i} \\in \\Delta} \\frac {1}{n} \\sum_ {i} D _ {K L} \\left(p _ {i} \\| q _ {i}\\right) \\quad s. t. \\quad \\forall j, \\quad \\frac {1}{n} \\sum_ {i} p _ {i, j} \\geq \\frac {\\alpha}{C} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.808, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( p_i = [p_{i,1}, \\ldots, p_{i,C}] \\), \\( q_i = [q_{i,1}, \\ldots, q_{i,C}] \\) and \\( \\Delta \\) denotes the simplex as \\( \\Delta = \\{p_i \\in \\mathbb{R}^C | \\sum_j^C p_{i,j} = 1; \\forall j, p_{i,j} \\geq 0\\} \\). The constraint bounds the minimal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.399, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.56, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Online Label Learning (OnLab)" + }, + { + "type": "algorithm", + "bbox": [ + 0.223, + 0.167, + 0.787, + 0.307 + ], + "angle": 0, + "content": "1: Input: Unlabeled image set \\(\\{I_i\\}\\), class names \\(\\{T_j\\}\\), pre-trained CLIP encoders \\((f, g)\\), dual variables \\(\\{\\rho_j\\}\\), ratio \\(\\alpha\\), temperature \\(\\tau_T\\), initial learning rate \\(c_{\\rho}\\) \n2: Obtain representation of text proxy as \\(\\mathbf{z}_j = g(T_j)\\) \n3: Initialize \\(\\forall j, \\rho_j = 0\\) \n4: while the \\(i\\)-th image arrives do \n5: Obtain vision representation as \\(\\mathbf{x}_i = f(I_i)\\) \n6: Observe the probability \\(p_i^*\\) by Eqn. 6 \n7: Have the prediction as \\(y_i = \\arg \\max_j p_{i,j}^*\\) \n8: Update dual variables as in Eqn. 7 \n9: end while" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.396 + ], + "angle": 0, + "content": "size of each class to avoid collapsing and \\(\\alpha \\in [0,1]\\) is the ratio for balancing the assignment between different classes. \\(\\alpha = 1\\) denotes a well-balanced distribution that each class has the same proportion of images, while \\(\\alpha = 0\\) keeps the prediction of the original zero-shot classification from \\(q_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.456 + ], + "angle": 0, + "content": "The problem can be solved effectively given all examples, but optimizing the objective with streaming images is challenging. To address the online problem, we introduce the Lagrange dual variables [2] \\(\\{\\rho_j\\}\\) for constraints and rewrite the problem as" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.465, + 0.785, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\rho , p _ {i}) = \\max _ {\\rho_ {j} \\geq 0} \\min _ {p _ {i} \\in \\Delta} \\frac {1}{n} \\sum_ {i} D _ {K L} (p _ {i} | | q _ {i}) - \\sum_ {j} ^ {C} \\rho_ {j} \\left(\\frac {1}{n} \\sum_ {i} p _ {i, j} - \\frac {\\alpha}{C}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.784, + 0.545 + ], + "angle": 0, + "content": "When the \\(i\\)-th example arrives for prediction, the sub-problem with fixed dual variables becomes" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.552, + 0.785, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\prime} \\left(p _ {i}\\right) = \\min _ {p _ {i} \\in \\Delta} D _ {K L} \\left(p _ {i} \\mid \\mid q _ {i}\\right) - \\sum_ {j} ^ {C} \\rho_ {j} ^ {i - 1} p _ {i, j} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.76, + 0.621 + ], + "angle": 0, + "content": "where \\(\\rho_j^{i - 1}\\) indicates the dual variable used for the \\((i - 1)\\)-th arrived image." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.621, + 0.708, + 0.634 + ], + "angle": 0, + "content": "Fortunately, this problem has the closed-form solution as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.642, + 0.687, + 0.658 + ], + "angle": 0, + "content": "Proposition 1 The optimal solution to the problem in Eqn. 5 is" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.665, + 0.785, + 0.705 + ], + "angle": 0, + "content": "\\[\np _ {i, j} ^ {*} = \\frac {q _ {i , j} \\exp \\left(\\rho_ {j} ^ {i - 1}\\right)}{\\sum_ {k} ^ {C} q _ {i , k} \\exp \\left(\\rho_ {k} ^ {i - 1}\\right)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.712, + 0.521, + 0.728 + ], + "angle": 0, + "content": "Proof. It is from the K.K.T. condition [2]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.784, + 0.784 + ], + "angle": 0, + "content": "Remark Proposition 1 shows that the dual variable helps balance the assignment between classes. When \\(\\rho = 0\\) without optimization, it degenerates to the original prediction from CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.784, + 0.816 + ], + "angle": 0, + "content": "With the prediction for the image, the dual variable can be updated by gradient ascent as" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.823, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\rho_ {j} ^ {i} = \\max \\left\\{0, \\rho_ {j} ^ {i - 1} - \\eta_ {\\rho} ^ {i} \\left(p _ {i, j} ^ {*} - \\alpha / C\\right) \\right\\} \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.447, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.191 + ], + "angle": 0, + "content": "where \\(\\eta_{\\rho}^{i}\\) is the learning rate for the dual variable. When \\(\\{\\rho_j\\}\\) is initialized as 0 and \\(\\alpha = 0\\), dual variables will not be updated, which leads to the original prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.784, + 0.252 + ], + "angle": 0, + "content": "Our online label learning (OnLab) algorithm is summarized in Alg. 1. The prediction will be obtained by a close-form solution and the updating for dual variables is also efficient, which makes the proposed method applicable for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.299 + ], + "angle": 0, + "content": "Besides the efficiency, the effectiveness of OnLab can be theoretically guaranteed in the following theorem. The detailed proof can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.306, + 0.784, + 0.339 + ], + "angle": 0, + "content": "Theorem 1 By running Alg. 1 for solving the problem in Eqn. 4 with the step size \\(\\eta_{\\rho}^{i} = c_{\\rho} / \\sqrt{i}\\), where \\(c_{\\rho}\\) is a constant, the convergence can be guaranteed as" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.347, + 0.786, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{n} \\max _ {\\rho} \\sum_ {i} \\mathcal {L} \\left(\\rho , p _ {i} ^ {*}\\right) - \\min _ {p _ {i}} \\sum_ {i} \\mathcal {L} \\left(\\rho^ {i - 1}, p _ {i}\\right) \\leq \\mathcal {O} \\left(\\frac {1}{\\sqrt {n}}\\right) \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.401, + 0.452, + 0.417 + ], + "angle": 0, + "content": "3.3 Online Proxy Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.784, + 0.485 + ], + "angle": 0, + "content": "After optimizing the label by online label learning, we consider reconstructing the class proxy in the vision space to reduce the modality gap from the text proxy as suggested in [16]. Given the whole set of images, vision proxy learning can be cast as an optimization problem" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.494, + 0.786, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {w}} \\mathcal {L} (\\mathbf {w}) = - \\sum_ {i} \\sum_ {j} p _ {i, j} ^ {*} \\log \\left(p _ {i, j} ^ {\\prime}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.535, + 0.591, + 0.552 + ], + "angle": 0, + "content": "where \\(p_{i,j}^{*}\\) is the output of online label learning and" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.562, + 0.786, + 0.597 + ], + "angle": 0, + "content": "\\[\np _ {i, j} ^ {\\prime} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j} / \\tau_ {I}\\right)}{\\sum_ {k} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {k} / \\tau_ {I}\\right)} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.784, + 0.634 + ], + "angle": 0, + "content": "\\(\\mathbf{w}_j\\) denotes the learnable vision proxy for the \\(j\\)-th class. \\(\\tau_I\\) is the temperature that should be larger than \\(\\tau_T\\) to obtain the class proxy in the vision space [16]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.784, + 0.679 + ], + "angle": 0, + "content": "When solving the problem in an online manner, only a single example will be received at each iteration. Therefore, the vision proxy will be updated according to the gradient from the \\(i\\)-th example as" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.689, + 0.786, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} ^ {i} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\mathbf {w} ^ {i - 1} - \\eta_ {w} ^ {i} \\nabla \\mathcal {L} \\left(\\mathbf {w} ^ {i - 1}, \\mathbf {x} _ {i}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.714, + 0.784, + 0.761 + ], + "angle": 0, + "content": "where \\(\\Pi(\\cdot)\\) projects the updated vision proxy to the unit norm. Note that the loss function is convex in \\(\\mathbf{w}\\), and the convergence can be guaranteed by the standard online learning theory [7]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.769, + 0.784, + 0.801 + ], + "angle": 0, + "content": "Theorem 2 By updating \\(\\mathbf{w}\\) as in Eqn. 11 and setting \\(\\eta_w^i = c_w / \\sqrt{i}\\), where \\(c_{w}\\) is a constant, the regret can be bounded as" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.809, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{n} \\sum_ {i} \\mathcal {L} \\left(\\mathbf {w} ^ {i - 1}, \\mathbf {x} _ {i}\\right) - \\min _ {\\mathbf {w}} \\mathcal {L} (\\mathbf {w}) \\leq \\mathcal {O} \\left(\\frac {1}{\\sqrt {n}}\\right) \\tag {12}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.399, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.592, + 0.163 + ], + "angle": 0, + "content": "Algorithm 2 Online Zero-shot Transfer (OnZeta)" + }, + { + "type": "algorithm", + "bbox": [ + 0.223, + 0.167, + 0.787, + 0.347 + ], + "angle": 0, + "content": "1: Input: Unlabeled image set \\(\\{I_i\\}\\), class names \\(\\{T_j\\}\\), pre-trained CLIP encoders \\((f, g)\\), dual variables \\(\\{\\rho_j\\}\\), vision proxy \\(\\{\\mathbf{w}_j\\}\\), ratio \\(\\alpha\\), temperature \\(\\tau_T\\) and \\(\\tau_I\\), initial learning rate \\(c_{\\rho}\\) and \\(c_w\\), weight parameter \\(\\beta\\) \n2: Obtain representation of text proxy as \\(\\mathbf{z}_j = g(T_j)\\) \n3: Initialize \\(\\forall j, \\rho_j = 0\\) and \\(\\mathbf{w}_j = \\mathbf{z}_j\\) \n4: while the \\(i\\)-th image arrives do \n5: Obtain vision representation as \\(\\mathbf{x}_i = f(I_i)\\) \n6: Observe the probability \\(p_i^*\\) by Eqn. 6 \n7: Observe the probability \\(p_i'\\) by Eqn. 10 \n8: Have the prediction as \\(y_i = \\arg \\max_j \\tilde{p}_{i,j}\\) where \\(\\tilde{p}_i\\) is from Eqn. 13 \n9: Update dual variables as in Eqn. 7 \n10: Update vision proxy as in Eqn. 11 \n11: end while" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.784, + 0.485 + ], + "angle": 0, + "content": "Since images arrive in a random order and only a single image will be leveraged for updating \\(\\mathbf{w}\\), the variance in learning can be large, especially for the first few iterations. Unlike offline optimization in [16], where the performance is evaluated after the final iteration, online learning has to accumulate the accuracy of predictions from different timestamps. Therefore, we consider combining the pseudo label \\(p_i^*\\) from online label learning in text space to stabilize the prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.562 + ], + "angle": 0, + "content": "For the \\(i\\)-th example, assume that \\(\\hat{p}_i\\) is the ground-truth distribution in the vision space. Since \\(p_i^*\\) is mainly from the text proxy, it is a biased estimation from the text space. While \\(p_i'\\) is estimated by the vision proxy, the variance can be large due to the online updating. Therefore, these predictions can be mixed to trade-off bias from text space and variance from vision space. Concretely, let" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.575, + 0.785, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {i} = \\lambda_ {i} p _ {i} ^ {\\prime} + (1 - \\lambda_ {i}) p _ {i} ^ {*} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.605, + 0.315, + 0.617 + ], + "angle": 0, + "content": "then we have" + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.632, + 0.785, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E [ \\| \\tilde {p} _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} ] \\\\ = E \\left[ \\lambda_ {i} ^ {2} \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} + (1 - \\lambda_ {i}) ^ {2} \\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2} + 2 \\lambda (1 - \\lambda) \\langle p _ {i} ^ {\\prime} - \\hat {p} _ {i}, p _ {i} ^ {*} - \\hat {p} _ {i} \\rangle \\right] \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.784, + 0.713 + ], + "angle": 0, + "content": "For the sake of simplicity, we assume that \\( p_i' \\) is an unbiased estimation for \\( \\hat{p}_i \\) such that \\( E[p_i'] = \\hat{p}_i \\). Then" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.723, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\nE _ {w} \\left[ \\| \\tilde {p} _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\right] = \\lambda_ {i} ^ {2} \\underbrace {E \\left[ \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\right]} _ {\\text {v a r i a n c e}} + (1 - \\lambda_ {i}) ^ {2} \\underbrace {\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2}} _ {\\text {b i a s}} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.771, + 0.784, + 0.8 + ], + "angle": 0, + "content": "To minimize the approximation error, the trade-off ratio \\(\\lambda_{i}\\) has the optimal solution as" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.81, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i} ^ {*} = \\frac {\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2}}{\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2} + E [ \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} ]} \\tag {16}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.447, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "We find that the bias can be considered as a constant but the variance will be reduced with the learning of \\(\\mathbf{w}\\), which implies a monotonically increasing function for combining the pseudo labels from different spaces. While various weight functions for \\(\\lambda\\) can be developed according to the analysis, for the \\(i\\)-th iteration, we empirically set" + }, + { + "type": "equation", + "bbox": [ + 0.454, + 0.239, + 0.786, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i} = \\beta \\sqrt {i / n} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.786, + 0.3 + ], + "angle": 0, + "content": "where \\(\\beta \\in [0,1]\\) is a constant. The combination strategy helps online transfer when the vision proxy is not trained sufficiently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.787, + 0.362 + ], + "angle": 0, + "content": "Incorporating online label learning and online proxy learning, the proposed online zero-shot transfer (OnZeta) algorithm is summarized in Alg. 2. All operations in the proposed method can be computed efficiently without significantly increasing the cost of conventional zero-shot transfer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.396, + 0.376, + 0.414 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.789, + 0.592 + ], + "angle": 0, + "content": "To evaluate the proposed method, we conduct experiments on 14 diverse downstream tasks. Considering that text prompts are essential for obtaining the appropriate text proxy for zero-shot transfer, we follow the common practice of including an ensemble of 7 text prompts as suggested in [17] to generate the text proxy for baseline and our method. The initial learning rates for dual variables and vision proxy are set as \\( c_{\\rho} = 20 \\) and \\( c_{w} = 0.5 \\), respectively. The learning rate will be decayed as \\( \\eta_{\\rho}^{i} = c_{\\rho} / \\sqrt{i} \\) and \\( \\eta_w^i = c_w / \\sqrt{i} \\) where \\( i \\) denotes the order of the incoming image. Unlabeled images will arrive in a random order. For the temperature, we fix \\( \\tau_T = 0.01 \\) that is the optimized value in CLIP and \\( \\tau_I = 0.04 \\) as suggested in [16]. All experiments are implemented on a single V100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.39, + 0.637 + ], + "angle": 0, + "content": "4.1 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.732 + ], + "angle": 0, + "content": "The main parameters in our method are the ratio for class assignment \\(\\alpha\\) and the ratio for pseudo label combination \\(\\beta\\). In this subsection, we will investigate the influence of these parameters along with the setting of online zero-shot transfer. The ablation experiments are conducted on ImageNet with ResNet-50 [8] as vision encoder in CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effect of \\(\\alpha\\) \\(\\alpha\\) is the ratio to capture the distribution over the entire data set as in Eqn. 3. While the vanilla zero-shot transfer ignores the side information from the arrived examples, the proposed online label learning can leverage the information of distribution by \\(\\alpha\\). Since \\(\\alpha \\in [0,1]\\), we vary \\(\\alpha\\) in \\(\\{0,0.4,0.6,0.8,1\\}\\) and summarize the results in Tab. 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.399, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.177, + 0.472, + 0.259 + ], + "angle": 0, + "content": "Table 1: Effect of \\(\\alpha\\) in Eqn. 3 for online label learning on ImageNet with ResNet-50 as vision encoder. \"Min class\" denotes the proportion of examples in the smallest class." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.262, + 0.456, + 0.356 + ], + "angle": 0, + "content": "
αAcc(%)#Min class(%)
060.3213.7
0.460.4118.5
0.660.6323.0
0.861.2033.3
161.9639.0
" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.149, + 0.78, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.333, + 0.782, + 0.374 + ], + "angle": 0, + "content": "Fig.3: Illustration of data distribution over 1,000 classes with different \\(\\alpha\\) on ImageNet. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.47 + ], + "angle": 0, + "content": "First, according to our analysis, \\(\\alpha = 0\\) will not update dual variables and implies the accuracy of baseline zero-shot transfer as in CLIP. Then, increasing \\(\\alpha\\) can consistently improve the accuracy over the baseline, which shows the efficacy of online label learning. Moreover, the size of the smallest class is also increased as \\(\\alpha\\) increases and it shows that our proposed method OnLab with only the online label learning can adjust the class assignment effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.787, + 0.577 + ], + "angle": 0, + "content": "To further investigate the data distribution derived by different \\(\\alpha\\), we compare the size of all 1,000 classes in Fig. 3. Evidently, the proposed method is capable of balancing the distribution in an online manner. Note that even with \\(\\alpha = 1\\), the obtained distribution is not perfectly balanced, which is from the approximation error in the online scenario as analyzed in Theorem 1. We find that our method is not sensitive to \\(\\alpha\\) and fix \\(\\alpha = 1\\) in the following experiments if not specified." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.597, + 0.787, + 0.645 + ], + "angle": 0, + "content": "Effect of \\(\\beta\\) With the predictions from text and vision space in online zero-shot transfer, a combination strategy with \\(\\beta\\) in Eqn. 17 can trade-off between the bias and variance. Tab. 2 reports the performance with different \\(\\beta\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.654, + 0.783, + 0.695 + ], + "angle": 0, + "content": "Table 2: Effect of \\(\\beta\\) in Eqn. 17 for online proxy Table 3: Comparison of strate-learning on ImageNet with ResNet-50 as vision gies for obtaining the combination ratio \\(\\lambda\\) in Eqn. 13." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.697, + 0.561, + 0.731 + ], + "angle": 0, + "content": "
β00.20.40.60.81
Acc(%)61.9662.1762.2362.3862.5061.84
" + }, + { + "type": "table", + "bbox": [ + 0.591, + 0.697, + 0.763, + 0.731 + ], + "angle": 0, + "content": "
βFixedDynamic
Acc(%)62.2862.50
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "When \\(\\beta = 0\\), only the label from the online label learning (i.e., OnLab) is adopted for prediction. If mixing the label from the vision proxy, the accuracy improves. The mixture prediction outperforms OnLab by \\(0.54\\%\\) with \\(\\beta = 0.8\\) and is \\(2.18\\%\\) better than the baseline CLIP. If excluding the label from text space at the end of data streaming as \\(\\beta = 1\\), the performance degenerates due to the large variance from the vision space. The phenomenon illustrates the challenge" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.447, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "of online learning that the vision proxy may not be well learned with a limited number of arrived images and the pseudo label from text proxy can help reduce the variance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.253 + ], + "angle": 0, + "content": "According to our analysis in Section 3.3, a dynamic ratio for the bias-variance trade-off is more effective for the proposed online learning scenario. To verify the claim, we compare the dynamic ratio to the fixed one in Tab. 3. The value of fixed \\(\\lambda\\) is searched in [0, 1] by a step size of 0.1 and the best result is reported." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.312 + ], + "angle": 0, + "content": "While the fixed \\(\\lambda\\) still improves the performance from the online label learning, a monotonically increasing \\(\\lambda\\) captures the training dynamics better and shows an additional gain of \\(0.22\\%\\). It confirms our analysis about the bias-variance trade-off." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.784, + 0.409 + ], + "angle": 0, + "content": "Effect of iterations in optimization Our method has to learn the dual variables and vision proxy with streaming data and those parameters may not be learned well at the early stage of online learning. To investigate the influence of iterations in online learning, we report the accumulated accuracy of different iterations in one epoch on ImageNet in Tab. 4." + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.435, + 0.774, + 0.477 + ], + "angle": 0, + "content": "Table 4: Comparison of accu- Table 5: Effect of epochs for optimization on Imracy \\((\\%)\\) with different itera-ageNet. Prediction from the last epoch is evaluated tions on ImageNet. for OnZeta after running multiple epochs." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.49, + 0.417, + 0.55 + ], + "angle": 0, + "content": "
#Iters5k10k20k
Acc(%)60.9661.5061.71
30k40k50k
62.0262.4162.50
" + }, + { + "type": "table", + "bbox": [ + 0.437, + 0.49, + 0.774, + 0.521 + ], + "angle": 0, + "content": "
#Epoch1(online learning)2345
Acc(%)62.5063.2463.3563.3963.46
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.684 + ], + "angle": 0, + "content": "First, we can observe that even with 5,000 iterations for learning, the accuracy over 1,000 classes is already better than the baseline. With the arrival of more images, the performance increases steadily. It is because that the convergence rate for dual variables and vision proxy is \\(\\mathcal{O}(1 / \\sqrt{n})\\) as analyzed in Theorems 1-2. It implies that these variables can reach a reasonable performance with a limited number of iterations, while it is challenging to obtain the optimal solution as discussed in the following experiment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.78 + ], + "angle": 0, + "content": "Effect of epochs in optimization In the proposed online zero-shot transfer problem, all examples can be visited only once when they arrive. The setting is prevalent in real-world applications but very challenging for optimization. This ablation experiment aims to investigate the performance gain when each example can be visited multiple times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Concretely, we run the proposed method on multiple epochs of data, where images arrive in a random order within each epoch. For multiple epochs, the prediction from the last one is reported for evaluation. Tab. 5 compares the performance of multiple epochs to that of a single epoch in the online scenario." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "First, we can observe that with an additional epoch, the performance of OnZeta can be boosted by \\(0.74\\%\\), which demonstrates the challenge of online zero-shot transfer. When increasing the number of epochs, the performance can be further improved and it achieves \\(63.46\\%\\) with 5-epoch learning. The result approaches the performance of learning with the entire data set as in [16] that shows a \\(63.74\\%\\) accuracy when visiting all data by 2,000 epochs. The competitive performance of OnZeta confirms the effectiveness of the proposed online learning algorithm. More experiments can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.286, + 0.473, + 0.302 + ], + "angle": 0, + "content": "4.2 Comparison on ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.489 + ], + "angle": 0, + "content": "After the ablation study, we compare our method OnZeta to state-of-the-art methods on ImageNet. Multiple vision encoders from CLIP are adopted for evaluation, which include ResNet-50, ViT-B/32, ViT-B/16, ViT-L/14, and ViT-L/14@336. Considering that images come in a random order in our online setting, the proposed method is repeated by 5 times and the averaged result is reported for the comparison. We note that the standard deviation is quite small and also report the best result among different runs by *. The same parameters are shared by different vision encoders in our method. Tab. 6 summarizes the comparison, where the performance of InMaP [16] is marked by gray due to the application of the whole unlabeled data set. All other methods only access a single image at each iteration. Our OnLab with only the online label learning is denoted as \"OnZeta0.5\" and the vanilla zero-shot transfer in Eqn. 1 is \"Baseline\"." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.512, + 0.789, + 0.584 + ], + "angle": 0, + "content": "Table 6: Comparison of accuracy (%) on ImageNet with different vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.595, + 0.787, + 0.678 + ], + "angle": 0, + "content": "
Vision encoderBaseline [17]TPT [19]OnZeta0.5OnZeta*0.5OnZetaOnZeta*InMaP [16]
ResNet-5060.3260.7461.8961.9662.4262.5063.74
ViT-B/3263.77-65.0665.1865.7165.8067.29
ViT-B/1668.7568.9870.1870.2370.8770.9972.55
ViT-L/1475.96-76.9877.0777.8477.9479.29
ViT-L/1433677.02-77.9378.0578.7578.9480.21
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "First, by capturing the overall distribution using only the proposed online label learning component, \\(\\mathrm{OnZeta}_{0.5}\\) already gains about \\(1\\%\\) for all vision encoders over the baseline. It demonstrates that the side information from the arrived images is useful for improving the zero-shot transfer. By learning the vision proxy and combining the predictions from both the text and vision spaces, OnZeta shows about \\(1\\%\\) additional improvement over \\(\\mathrm{OnZeta}_{0.5}\\), which illustrates that online proxy learning is complementary to online label learning for zero-shot classification. Moreover, OnZeta obtains more gains with a large backbone, showing its potential for large models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.447, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Then, we can observe that the prompt learning method TPT outperforms the baseline with an ensemble of artificial text prompts, which indicates the importance of the appropriate prompt for zero-shot transfer. However, OnZeta surpasses TPT by \\(1.76\\%\\) on ResNet-50 and \\(2.01\\%\\) on ViT-B/16. It is because that we learn a better class proxy in the vision space for visual classification, which helps reduce the modality gap and is more effective than optimizing the text prompt in the text space. The observation is consistent with the analysis in [16]. Moreover, compared with InMaP that can access the whole unlabeled set, OnZeta is only about \\(1\\%\\) worse using different vision encoders. As analyzed in the ablation study, online learning that visits each example once without storing is more challenging than the offline method with multiple iterations over the entire set. Nevertheless, online learning is practical for real-world applications. Finally, all vision encoders share the same parameters in experiments. It implies that our method OnZeta is not sensitive to the selection of vision encoder and is applicable in different configurations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.394, + 0.628, + 0.409 + ], + "angle": 0, + "content": "4.3 Comparison on Other 13 Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Besides the comparison on ImageNet, we conduct experiments on other 13 downstream tasks to evaluate the performance of our method. The benchmark data sets for zero-shot transfer are adopted in the comparison, which includes Aircraft [12], Caltech101 [5], Stanford Cars [10], CIFAR-10 [11], CIFAR-100 [11], CUB200-2011 [21], Describable Textures Dataset (DTD) [3], EuroSAT [9], Flowers [13], Food101 [1], Oxford-IIIT Pet (Pets) [14], Sun397 [23], and UCF101 [20]. A broad range of downstream tasks are involved by these diverse data sets, e.g., low-resolution image classification, fine-grained visual categorization (FGVC), land cover classification with satellite images, scene categorization, texture recognition, etc. These tasks contain fewer images than ImageNet, which increases the challenge of our online transfer with the convergence rate of \\(\\mathcal{O}(1 / \\sqrt{n})\\). Therefore, we reduce \\(\\beta\\) from 0.8 to 0.4 for all tasks to exploit the label from the text space. For data sets that the baseline already achieves satisfied performance, e.g., Caltech101 and CIFAR-10 with ViT, online label learning can be skipped by letting \\(\\alpha = 0\\), while \\(\\alpha = 0.4\\) can provide a slight improvement. Other parameters on the rest data sets remain the same as those for ImageNet. Two distinct vision encoders, i.e., ResNet-50 and ViT-B/16, are applied for evaluation. The performance averaged over 5 random trials is reported for our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The comparison is summarized in Tab. 7. A similar phenomenon as ImageNet can be observed on these downstream tasks. Concretely, online label learning, i.e., \\(\\mathrm{OnZeta}_{0.5}\\), already improves the performance over baselines, where the gain is \\(3.12\\%\\) on ResNet and \\(2.58\\%\\) on ViT, respectively. Compared with the baseline, only streaming images are leveraged by our method, and the representation of each arrived image is not stored, which preserves the flexibility of zero-shot transfer but captures the distribution over the entire data set in an online manner. In addition, learning the vision proxy with the pseudo label from the text space further improves the performance with different vision encoders, which demonstrates the effectiveness of learning the class proxy directly in the target" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.399, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 7: Comparison of accuracy (\\%) on 13 downstream tasks with ResNet-50 and ViT-B/16 as vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.241, + 0.788, + 0.415 + ], + "angle": 0, + "content": "
AircraftCaltechCarsCifar10Cifar100CUBDTDEuroSATFlowersFoodPetsSUNUCF101Avg.
ResNet-50:
Baseline16.6286.0056.3173.1540.6041.3741.1326.9062.9774.1081.8559.2555.5655.06
TPT17.5887.0258.46---40.8428.3362.6974.8884.4961.4660.82-
OnZeta0.518.2986.1758.5775.6246.9645.6843.2433.7962.4276.4183.0561.6460.1257.84
OnZeta*0.518.5186.4558.9175.7547.2245.9343.7434.3263.1776.5383.5961.8160.3858.18
OnZeta18.3186.2659.4075.8447.5546.0643.3833.9963.1676.8383.7762.3261.0058.30
OnZeta*18.5486.4159.7175.9547.7146.3443.9134.3563.9576.9784.6662.4061.3858.64
InMaP18.9686.7363.3078.8449.2649.1744.8635.8866.6878.3687.9363.8263.3660.55
ViT-B/16:
Baseline23.1393.5166.2991.0967.2949.3645.0950.1766.9984.4387.0065.6865.2165.79
TPT24.7894.1666.87---47.7542.4468.9884.6787.7965.5068.04-
OnZeta0.527.1993.6567.9991.1369.8552.8147.5156.2068.4985.8388.4468.0668.5068.13
OnZeta*0.527.4893.7968.1191.1870.0253.0947.8756.4269.0685.9788.7268.2768.8668.37
OnZeta27.5793.6968.7491.3170.3953.6648.0456.5269.3786.2189.0168.8769.5868.69
OnZeta*28.2993.8369.0391.3570.5253.8048.5856.7469.6386.3589.3269.0169.9468.95
InMaP28.3593.5973.0093.2772.5157.8750.7763.9871.2887.7692.9470.8572.0671.40
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.789, + 0.639 + ], + "angle": 0, + "content": "vision space. Moreover, OnZeta outperforms TPT on 9 out of 10 (only 10 tasks are available in TPT's original paper) data sets with different vision encoders. Compared with TPT that has to optimize the text prompt with multiple augmentations of the image by back-propagation through the text encoder, OnZeta only contains the forward pass to extract the representation of the image, which is more efficient and is capable of online deployment. By comparing the best performance of the proposed method (i.e., indicated by * in the table) to the averaged one, we can find that the gap averaged over 13 data sets is only about \\(0.3\\%\\). While the images arrive in a random order, the performance of the proposed method can be theoretically guaranteed as indicated in Theorems 1-2, which is robust to the order of input images. Finally, most data sets share the same parameters and it implies that OnZeta is not sensitive to hyper-parameters and is applicable for different tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.36, + 0.675 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While CLIP demonstrates impressive zero-shot transfer performance, the information from the target data has not been explored sufficiently. In this work, we investigate an online zero-shot scenario, where a random image arrives at each iteration and the model has to predict its label without storing its representation. By capturing the overall label distribution and refining the class proxy in an online manner, our method OnZeta demonstrates a substantial improvement over the vanilla zero-shot transfer on diverse downstream tasks with different vision encoders. In addition, the operations in the proposed method can be implemented efficiently for real-time applications. Exploring more effective algorithms for challenging online zero-shot transfer can be our future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.447, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Online Zero-Shot Classification with CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.249 + ], + "angle": 0, + "content": "2. Boyd, S.P., Vandenberghe, L.: Convex Optimization. Cambridge University Press (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.278 + ], + "angle": 0, + "content": "3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.28, + 0.785, + 0.333 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.785, + 0.375 + ], + "angle": 0, + "content": "5. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.417 + ], + "angle": 0, + "content": "6. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. CoRR abs/2110.04544 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.445 + ], + "angle": 0, + "content": "7. Hazan, E.: Introduction to online convex optimization. CoRR abs/1909.05207 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.474 + ], + "angle": 0, + "content": "8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.516 + ], + "angle": 0, + "content": "9. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens. 12(7), 2217-2226 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.544 + ], + "angle": 0, + "content": "0. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.572 + ], + "angle": 0, + "content": "1. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.6 + ], + "angle": 0, + "content": "2. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.629 + ], + "angle": 0, + "content": "3. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.657 + ], + "angle": 0, + "content": "4. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.685 + ], + "angle": 0, + "content": "5. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.713 + ], + "angle": 0, + "content": "6. Qian, Q., Xu, Y., Hu, J.: Intra-modal proxy learning for zero-shot visual categorization with clip. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.755 + ], + "angle": 0, + "content": "7. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.797 + ], + "angle": 0, + "content": "8. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Shu, M., Nie, W., Huang, D., Yu, Z., Goldstein, T., Anandkumar, A., Xiao, C.: Test-time prompt tuning for zero-shot generalization in vision-language models. In: NeurIPS (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.129 + ], + "angle": 0, + "content": "Q. Qian and J. Hu" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR abs/1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "21. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.244 + ], + "angle": 0, + "content": "22. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. In: CVPR. pp. 7949-7961. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "23. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "24. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of CLIP for few-shot classification. In: ECCV. vol. 13695, pp. 493-510. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "25. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16795-16804. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.384 + ], + "angle": 0, + "content": "26. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vis. 130(9), 2337-2348 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_origin.pdf b/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eeb0cbf1bad8b2ebe2c8a816849746ef1a37c8a0 --- /dev/null +++ b/2024/Online Zero-Shot Classification with CLIP/e360d63f-14d8-422a-8b7f-da89cce40c64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:959da3ea2e37a37da3de8f7949ac2877b9f11ee5b266f1ba4eb3f697247da90a +size 768819 diff --git a/2024/Online Zero-Shot Classification with CLIP/full.md b/2024/Online Zero-Shot Classification with CLIP/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c73d705f5ee6f19a83946ca918e21ed26e8fc5f8 --- /dev/null +++ b/2024/Online Zero-Shot Classification with CLIP/full.md @@ -0,0 +1,331 @@ +# Online Zero-Shot Classification with CLIP + +Qi Qian $^{1}$ and Juhua Hu $^{2}$ + +1 Alibaba Group, Bellevue, WA 98004, USA +2 School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA +qi.qian@alibaba-inc.com, juhuah@uw.edu + +Abstract. Vision-language pre-training such as CLIP enables zero-shot transfer that can classify images according to the candidate class names. While CLIP demonstrates an impressive zero-shot performance on diverse downstream tasks, the distribution from the target data has not been leveraged sufficiently. In this work, we study a novel online zero-shot transfer scenario, where each image arrives in a random order for classification and is visited only once to obtain prediction immediately without storing its representation. Compared with the vanilla zero-shot classification, the proposed framework preserves its flexibility for online service while considering the statistics of the arrived images as the side information to capture the distribution of target data, which can help improve the performance of real-world applications. To tackle the challenge of effective online optimization, we first develop online label learning to model the target data distribution. Then, the proxy of each class in the vision space is further optimized with the proposed online proxy learning method to mitigate the modality gap between images and text. The convergence of both online strategies can be theoretically guaranteed. By combining the predicted label from the online label learning and proxy learning, our online zero-shot transfer method (OnZeta) achieves $78.94\%$ accuracy on ImageNet without accessing the entire data set. Moreover, extensive experiments on other 13 downstream tasks with different vision encoders show a more than $3\%$ improvement on average, which demonstrates the effectiveness of our proposal. + +Keywords: Online learning $\cdot$ Zero-shot classification $\cdot$ CLIP + +# 1 Introduction + +Vision-language pre-training has attracted much attention recently due to its impressive zero-shot transfer performance on various downstream tasks. The desired property mainly comes from aligning the vision and text space. For example, one of the most prevalent pre-training methods, i.e., CLIP [17], consists of distinct vision and text encoders for learning image and text representations, respectively. These encoders are optimized by minimizing a contrastive loss defined on image-text pairs. The loss aims to pull images and their corresponding text descriptions together while pushing the irrelevant text or images away [15]. + +![](images/e6c7bd7fc911e1039ee20cfd016219ef8d5d01bfed9ffdcc7c4133fcb6b5e486.jpg) +Fig. 1: Illustration of the proposed online zero-shot transfer method (OnZeta). Blue and orange lines denote the inference in the text and vision space, respectively. By incorporating the predictions from online label learning and online proxy learning, OnZeta can leverage the biased prediction from the text space to reduce the variance in the target vision space in an online manner. + +After the alignment between vision and text, a novel zero-shot classification paradigm emerges to transfer knowledge in pre-trained models to diverse tasks. Concretely, given all class names from the target task, each class can be represented by a class proxy that applies the text encoder in CLIP for the corresponding class name, which is also termed a text proxy. When an image arrives, its representation is extracted by the vision encoder of CLIP. Then, the label can be assigned by the 1-nearest neighbor (1-NN) classifier with the set of text proxies as the reference. Without any fine-tuning, this simple strategy achieves $77.02\%$ accuracy on ImageNet [18] using the pre-trained ViT [4] as the vision encoder. + +Since the success of CLIP, many methods have been developed to explore different side information from the target task to further improve the transfer performance. Firstly, if a few labeled images are available from the target task, the zero-shot transfer can be formulated as a few-shot learning problem and the input text prompt can be learned with the labeled vision data by fine-tuning [26]. Secondly, if the computational resource is too limited to do fine-tuning, the labeled data can be cached as a vision classifier to classify images directly [24]. However, supervised information is often rare, while unsupervised information is more accessible. Therefore, many efforts have been devoted to exploiting the unlabeled images to enhance the zero-shot performance. [19] generates multiple augmented examples from a given image, and then learns the specific text prompt for each image as few-shot learning accordingly. [16] considers reconstructing the class proxy in the vision space, i.e., vision proxy, by leveraging a set of unlabeled images in proxy learning. While [16] shows superior zero-shot performance, it optimizes proxies in an offline manner. Due to the privacy issue, gathering or keeping a set of unlabeled images is still challenging in some real-world applications. + +Therefore, in this work, we investigate a novel practical scenario for zero-shot transfer, termed online zero-shot transfer. Concretely, when streaming images arrive as in conventional zero-shot learning, the model has to classify the image immediately without refining. After that, the class proxy can be updated but + +![](images/121f387901d4124e3440254aca3c273c8ed4ef6a511d5a4c5faca2af3d5e32c6.jpg) +Fig. 2: Top 4 predicted labels with corresponding probabilities from baseline zero-shot method by CLIP and our proposal. + +the representations of arrived images cannot be kept. Unlike [16] that can access the whole unlabeled set, this online setting is way more challenging, where only statistics of seen images can be leveraged for optimization and each image is visited only once. Thereafter, the standard iterative optimization in [16] is not applicable, while the problem (i.e., online service to classify each unlabeled image on the fly without storing data) is ubiquitous in real-world applications (e.g., mobiles and robots). + +To address the online service challenge, we first consider exploring the data distribution of the target task to improve the prediction from the text proxy. Note that the vanilla zero-shot prediction focuses on individual images and ignores the distribution over the entire data set. To capture the distribution of all data, we develop an online label learning algorithm to balance the assignment between different classes globally. In addition, the class proxy from the text (i.e., text proxy) can be biased for the vision data due to the modality gap, and thus we propose online proxy learning to obtain the class proxy in the vision space (i.e., vision proxy) directly in an online manner. Finally, the prediction from both the text space and vision space can be combined to help reduce the variance of learning in the vision space as illustrated in the framework of our online zero-shot transfer (OnZeta) shown in Fig. 1. More importantly, with online optimization, our proposal OnZeta can leverage the information from target data effectively and efficiently, which helps correct the prediction compared to the baseline zero-shot method. For example, Fig. 2 shows that an image of "Husky" will be misclassified by vanilla CLIP. By considering the distribution over the whole data set and incorporating the learned vision proxy, our method can obtain the right label in an online manner. The main contributions of this work can be summarized as follows. + +- Considering the practical need for online service (i.e., classifying each image on the fly without storage) and improving the performance of vanilla zero-shot classification, we study a novel setting, that is, online zero-shot transfer, using pre-trained vision-language models. +- To address the online service challenge, we propose an online label learning method to trace the distribution of assigned labels over the entire data in an online manner. Moreover, an online vision proxy learning method is proposed + +to further improve the class proxy in vision space for classification. The convergence of both methods can be theoretically guaranteed. + +- Experiments on 14 downstream tasks demonstrate the effectiveness of our proposal OnZeta. Concretely, when visiting each image only once, OnZeta outperforms the baseline by a margin of more than $3\%$ on average without storing any unlabeled data. + +# 2 Related Work + +CLIP consists of two encoders (i.e., vision and text), and the optimization can be conducted on each of them individually. In this section, we briefly review research efforts in these two directions. + +# 2.1 Optimization with Text Encoder + +CLIP demonstrates that the zero-shot performance heavily depends on the text prompts for generating appropriate class proxies, and many methods investigate learning the optimal text prompt as input to the text encoder. CoOp [26] marks the input text prompts as learnable variables and optimizes them via few-shot learning with a limited number of labeled images for each class. CoCoOp [25] further includes the vision representation as the conditional context for prompt learning. TPT [19] relaxes the requirement of labeled examples and optimizes the consistency between different augmentations from the same image for prompt learning. However, it has to optimize the prompt for each image with 64 augmentations, which is inefficient for computation and memory, making it infeasible for real-time applications as the online scenario in this work. + +# 2.2 Optimization with Vision Encoder + +While prompt learning can improve the performance on the target data, recent research shows that optimizing the vision representation can be more effective for classification [6,16]. CLIP-Adapter [6] introduces an additional adapter network after the vision encoder and optimizes the vision representation with few-shot learning. Although a similar adapter can be attached for the text encoder, the empirical study shows that optimizing the vision space is sufficient for transfer and can achieve better performance than either the version with only a text adapter or that with both vision and text adapters. [22] shows that fine-tuning the vision classifier directly with few-shot learning can achieve better performance than CoOp. Recently, the analysis in [16] demonstrates that the optimal class proxy for visual classification lives in the vision space. To mitigate the modality gap, [16] proposes proxy learning that aims to learn class proxy for each class in vision space. However, it requires a set of unlabeled data and will learn proxies iteratively. In this work, we also aim to refine the class proxy in the vision space, but with the learned labels in an online manner, which is more applicable for online real applications. + +# 3 Online Zero-shot Transfer + +# 3.1 Zero-shot Transfer in CLIP + +Given a data set consisting of image-text pairs as $\{I_i, T_i\}_{i=1}^n$ , where $I_i$ denotes an image and $T_i$ is the corresponding text such as class names or captions. CLIP trains vision and text encoders jointly to align the vision subspace and its corresponding text subspace. Let $f(\cdot)$ and $g(\cdot)$ denote the vision encoder and text encoder, respectively. The representations for images and text can be extracted as $\mathbf{x}_i = f(I_i)$ and $\mathbf{z}_i = g(T_i)$ . + +After pre-training two encoders, zero-shot classification can be implemented by the 1-nearest neighbor (1-NN) classifier. Concretely, given the class names of the target task, the proxy of $j$ -th class $\mathbf{z}_j$ can be obtained by applying the text encoder for the prompt as "a photo of a {class name}". Then, the image $I_i$ will be classified as + +$$ +y _ {i} = \arg \max _ {j} \mathbf {x} _ {i} ^ {\top} \mathbf {z} _ {j} \tag {1} +$$ + +While the conventional zero-shot paradigm demonstrates an impressive transfer performance, the side information from the target task has not been explored sufficiently. Recent work shows that with only a set of unlabeled target data, the zero-shot performance can be improved dramatically [16]. In this work, we further relax the requirement of unlabeled target data and investigate a novel online zero-shot transfer scenario, where each unlabeled image arrives in an online manner without storing. + +# 3.2 Online Label Learning + +First, we develop an online label learning method to leverage the information from arrived images to capture the distribution over the entire set. + +Given text proxy $\mathbf{z}_j$ , the predicted distribution over classes for $I_i$ can be computed as + +$$ +q _ {i, j} = \frac {\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {z} _ {j} / \tau_ {T}\right)}{\sum_ {k} ^ {C} \exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {z} _ {k} / \tau_ {T}\right)} \tag {2} +$$ + +where $\tau_{T}$ is the temperature optimized by CLIP. + +If the whole image set is available, the label can be refined by considering the distribution over classes. Let $\{p_i\}$ denote the learnable labels, and the optimization problem with $C$ classes can be cast as + +$$ +\min _ {p _ {i} \in \Delta} \frac {1}{n} \sum_ {i} D _ {K L} \left(p _ {i} \| q _ {i}\right) \quad s. t. \quad \forall j, \quad \frac {1}{n} \sum_ {i} p _ {i, j} \geq \frac {\alpha}{C} \tag {3} +$$ + +where $p_i = [p_{i,1}, \ldots, p_{i,C}]$ , $q_i = [q_{i,1}, \ldots, q_{i,C}]$ and $\Delta$ denotes the simplex as $\Delta = \{p_i \in \mathbb{R}^C | \sum_j^C p_{i,j} = 1; \forall j, p_{i,j} \geq 0\}$ . The constraint bounds the minimal + +Algorithm 1 Online Label Learning (OnLab) +1: Input: Unlabeled image set $\{I_i\}$ , class names $\{T_j\}$ , pre-trained CLIP encoders $(f, g)$ , dual variables $\{\rho_j\}$ , ratio $\alpha$ , temperature $\tau_T$ , initial learning rate $c_{\rho}$ +2: Obtain representation of text proxy as $\mathbf{z}_j = g(T_j)$ +3: Initialize $\forall j, \rho_j = 0$ +4: while the $i$ -th image arrives do +5: Obtain vision representation as $\mathbf{x}_i = f(I_i)$ +6: Observe the probability $p_i^*$ by Eqn. 6 +7: Have the prediction as $y_i = \arg \max_j p_{i,j}^*$ +8: Update dual variables as in Eqn. 7 +9: end while + +size of each class to avoid collapsing and $\alpha \in [0,1]$ is the ratio for balancing the assignment between different classes. $\alpha = 1$ denotes a well-balanced distribution that each class has the same proportion of images, while $\alpha = 0$ keeps the prediction of the original zero-shot classification from $q_{i}$ . + +The problem can be solved effectively given all examples, but optimizing the objective with streaming images is challenging. To address the online problem, we introduce the Lagrange dual variables [2] $\{\rho_j\}$ for constraints and rewrite the problem as + +$$ +\mathcal {L} (\rho , p _ {i}) = \max _ {\rho_ {j} \geq 0} \min _ {p _ {i} \in \Delta} \frac {1}{n} \sum_ {i} D _ {K L} (p _ {i} | | q _ {i}) - \sum_ {j} ^ {C} \rho_ {j} \left(\frac {1}{n} \sum_ {i} p _ {i, j} - \frac {\alpha}{C}\right) \tag {4} +$$ + +When the $i$ -th example arrives for prediction, the sub-problem with fixed dual variables becomes + +$$ +\mathcal {L} ^ {\prime} \left(p _ {i}\right) = \min _ {p _ {i} \in \Delta} D _ {K L} \left(p _ {i} \mid \mid q _ {i}\right) - \sum_ {j} ^ {C} \rho_ {j} ^ {i - 1} p _ {i, j} \tag {5} +$$ + +where $\rho_j^{i - 1}$ indicates the dual variable used for the $(i - 1)$ -th arrived image. + +Fortunately, this problem has the closed-form solution as follows. + +Proposition 1 The optimal solution to the problem in Eqn. 5 is + +$$ +p _ {i, j} ^ {*} = \frac {q _ {i , j} \exp \left(\rho_ {j} ^ {i - 1}\right)}{\sum_ {k} ^ {C} q _ {i , k} \exp \left(\rho_ {k} ^ {i - 1}\right)} \tag {6} +$$ + +Proof. It is from the K.K.T. condition [2]. + +Remark Proposition 1 shows that the dual variable helps balance the assignment between classes. When $\rho = 0$ without optimization, it degenerates to the original prediction from CLIP. + +With the prediction for the image, the dual variable can be updated by gradient ascent as + +$$ +\rho_ {j} ^ {i} = \max \left\{0, \rho_ {j} ^ {i - 1} - \eta_ {\rho} ^ {i} \left(p _ {i, j} ^ {*} - \alpha / C\right) \right\} \tag {7} +$$ + +where $\eta_{\rho}^{i}$ is the learning rate for the dual variable. When $\{\rho_j\}$ is initialized as 0 and $\alpha = 0$ , dual variables will not be updated, which leads to the original prediction. + +Our online label learning (OnLab) algorithm is summarized in Alg. 1. The prediction will be obtained by a close-form solution and the updating for dual variables is also efficient, which makes the proposed method applicable for real-time applications. + +Besides the efficiency, the effectiveness of OnLab can be theoretically guaranteed in the following theorem. The detailed proof can be found in the supplementary. + +Theorem 1 By running Alg. 1 for solving the problem in Eqn. 4 with the step size $\eta_{\rho}^{i} = c_{\rho} / \sqrt{i}$ , where $c_{\rho}$ is a constant, the convergence can be guaranteed as + +$$ +\frac {1}{n} \max _ {\rho} \sum_ {i} \mathcal {L} \left(\rho , p _ {i} ^ {*}\right) - \min _ {p _ {i}} \sum_ {i} \mathcal {L} \left(\rho^ {i - 1}, p _ {i}\right) \leq \mathcal {O} \left(\frac {1}{\sqrt {n}}\right) \tag {8} +$$ + +# 3.3 Online Proxy Learning + +After optimizing the label by online label learning, we consider reconstructing the class proxy in the vision space to reduce the modality gap from the text proxy as suggested in [16]. Given the whole set of images, vision proxy learning can be cast as an optimization problem + +$$ +\min _ {\mathbf {w}} \mathcal {L} (\mathbf {w}) = - \sum_ {i} \sum_ {j} p _ {i, j} ^ {*} \log \left(p _ {i, j} ^ {\prime}\right) \tag {9} +$$ + +where $p_{i,j}^{*}$ is the output of online label learning and + +$$ +p _ {i, j} ^ {\prime} = \frac {\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {j} / \tau_ {I}\right)}{\sum_ {k} \exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {k} / \tau_ {I}\right)} \tag {10} +$$ + +$\mathbf{w}_j$ denotes the learnable vision proxy for the $j$ -th class. $\tau_I$ is the temperature that should be larger than $\tau_T$ to obtain the class proxy in the vision space [16]. + +When solving the problem in an online manner, only a single example will be received at each iteration. Therefore, the vision proxy will be updated according to the gradient from the $i$ -th example as + +$$ +\mathbf {w} ^ {i} = \Pi_ {\| \mathbf {w} \| _ {2} = 1} \left(\mathbf {w} ^ {i - 1} - \eta_ {w} ^ {i} \nabla \mathcal {L} \left(\mathbf {w} ^ {i - 1}, \mathbf {x} _ {i}\right)\right) \tag {11} +$$ + +where $\Pi(\cdot)$ projects the updated vision proxy to the unit norm. Note that the loss function is convex in $\mathbf{w}$ , and the convergence can be guaranteed by the standard online learning theory [7]. + +Theorem 2 By updating $\mathbf{w}$ as in Eqn. 11 and setting $\eta_w^i = c_w / \sqrt{i}$ , where $c_{w}$ is a constant, the regret can be bounded as + +$$ +\frac {1}{n} \sum_ {i} \mathcal {L} \left(\mathbf {w} ^ {i - 1}, \mathbf {x} _ {i}\right) - \min _ {\mathbf {w}} \mathcal {L} (\mathbf {w}) \leq \mathcal {O} \left(\frac {1}{\sqrt {n}}\right) \tag {12} +$$ + +Algorithm 2 Online Zero-shot Transfer (OnZeta) +1: Input: Unlabeled image set $\{I_i\}$ , class names $\{T_j\}$ , pre-trained CLIP encoders $(f, g)$ , dual variables $\{\rho_j\}$ , vision proxy $\{\mathbf{w}_j\}$ , ratio $\alpha$ , temperature $\tau_T$ and $\tau_I$ , initial learning rate $c_{\rho}$ and $c_w$ , weight parameter $\beta$ +2: Obtain representation of text proxy as $\mathbf{z}_j = g(T_j)$ +3: Initialize $\forall j, \rho_j = 0$ and $\mathbf{w}_j = \mathbf{z}_j$ +4: while the $i$ -th image arrives do +5: Obtain vision representation as $\mathbf{x}_i = f(I_i)$ +6: Observe the probability $p_i^*$ by Eqn. 6 +7: Observe the probability $p_i'$ by Eqn. 10 +8: Have the prediction as $y_i = \arg \max_j \tilde{p}_{i,j}$ where $\tilde{p}_i$ is from Eqn. 13 +9: Update dual variables as in Eqn. 7 +10: Update vision proxy as in Eqn. 11 +11: end while + +Since images arrive in a random order and only a single image will be leveraged for updating $\mathbf{w}$ , the variance in learning can be large, especially for the first few iterations. Unlike offline optimization in [16], where the performance is evaluated after the final iteration, online learning has to accumulate the accuracy of predictions from different timestamps. Therefore, we consider combining the pseudo label $p_i^*$ from online label learning in text space to stabilize the prediction. + +For the $i$ -th example, assume that $\hat{p}_i$ is the ground-truth distribution in the vision space. Since $p_i^*$ is mainly from the text proxy, it is a biased estimation from the text space. While $p_i'$ is estimated by the vision proxy, the variance can be large due to the online updating. Therefore, these predictions can be mixed to trade-off bias from text space and variance from vision space. Concretely, let + +$$ +\tilde {p} _ {i} = \lambda_ {i} p _ {i} ^ {\prime} + (1 - \lambda_ {i}) p _ {i} ^ {*} \tag {13} +$$ + +then we have + +$$ +\begin{array}{l} E [ \| \tilde {p} _ {i} - \hat {p} _ {i} \| _ {2} ^ {2} ] \\ = E \left[ \lambda_ {i} ^ {2} \| p _ {i} ^ {\prime} - \hat {p} _ {i} \| _ {2} ^ {2} + (1 - \lambda_ {i}) ^ {2} \| p _ {i} ^ {*} - \hat {p} _ {i} \| _ {2} ^ {2} + 2 \lambda (1 - \lambda) \langle p _ {i} ^ {\prime} - \hat {p} _ {i}, p _ {i} ^ {*} - \hat {p} _ {i} \rangle \right] \tag {14} \\ \end{array} +$$ + +For the sake of simplicity, we assume that $p_i'$ is an unbiased estimation for $\hat{p}_i$ such that $E[p_i'] = \hat{p}_i$ . Then + +$$ +E _ {w} \left[ \| \tilde {p} _ {i} - \hat {p} _ {i} \| _ {2} ^ {2} \right] = \lambda_ {i} ^ {2} \underbrace {E \left[ \| p _ {i} ^ {\prime} - \hat {p} _ {i} \| _ {2} ^ {2} \right]} _ {\text {v a r i a n c e}} + (1 - \lambda_ {i}) ^ {2} \underbrace {\| p _ {i} ^ {*} - \hat {p} _ {i} \| _ {2} ^ {2}} _ {\text {b i a s}} \tag {15} +$$ + +To minimize the approximation error, the trade-off ratio $\lambda_{i}$ has the optimal solution as + +$$ +\lambda_ {i} ^ {*} = \frac {\| p _ {i} ^ {*} - \hat {p} _ {i} \| _ {2} ^ {2}}{\| p _ {i} ^ {*} - \hat {p} _ {i} \| _ {2} ^ {2} + E [ \| p _ {i} ^ {\prime} - \hat {p} _ {i} \| _ {2} ^ {2} ]} \tag {16} +$$ + +We find that the bias can be considered as a constant but the variance will be reduced with the learning of $\mathbf{w}$ , which implies a monotonically increasing function for combining the pseudo labels from different spaces. While various weight functions for $\lambda$ can be developed according to the analysis, for the $i$ -th iteration, we empirically set + +$$ +\lambda_ {i} = \beta \sqrt {i / n} \tag {17} +$$ + +where $\beta \in [0,1]$ is a constant. The combination strategy helps online transfer when the vision proxy is not trained sufficiently. + +Incorporating online label learning and online proxy learning, the proposed online zero-shot transfer (OnZeta) algorithm is summarized in Alg. 2. All operations in the proposed method can be computed efficiently without significantly increasing the cost of conventional zero-shot transfer. + +# 4 Experiments + +To evaluate the proposed method, we conduct experiments on 14 diverse downstream tasks. Considering that text prompts are essential for obtaining the appropriate text proxy for zero-shot transfer, we follow the common practice of including an ensemble of 7 text prompts as suggested in [17] to generate the text proxy for baseline and our method. The initial learning rates for dual variables and vision proxy are set as $c_{\rho} = 20$ and $c_{w} = 0.5$ , respectively. The learning rate will be decayed as $\eta_{\rho}^{i} = c_{\rho} / \sqrt{i}$ and $\eta_w^i = c_w / \sqrt{i}$ where $i$ denotes the order of the incoming image. Unlabeled images will arrive in a random order. For the temperature, we fix $\tau_T = 0.01$ that is the optimized value in CLIP and $\tau_I = 0.04$ as suggested in [16]. All experiments are implemented on a single V100 GPU. + +# 4.1 Ablation Study + +The main parameters in our method are the ratio for class assignment $\alpha$ and the ratio for pseudo label combination $\beta$ . In this subsection, we will investigate the influence of these parameters along with the setting of online zero-shot transfer. The ablation experiments are conducted on ImageNet with ResNet-50 [8] as vision encoder in CLIP. + +Effect of $\alpha$ $\alpha$ is the ratio to capture the distribution over the entire data set as in Eqn. 3. While the vanilla zero-shot transfer ignores the side information from the arrived examples, the proposed online label learning can leverage the information of distribution by $\alpha$ . Since $\alpha \in [0,1]$ , we vary $\alpha$ in $\{0,0.4,0.6,0.8,1\}$ and summarize the results in Tab. 1. + +Table 1: Effect of $\alpha$ in Eqn. 3 for online label learning on ImageNet with ResNet-50 as vision encoder. "Min class" denotes the proportion of examples in the smallest class. + +
αAcc(%)#Min class(%)
060.3213.7
0.460.4118.5
0.660.6323.0
0.861.2033.3
161.9639.0
+ +![](images/cc944ef65f3a7865e6f00d1e563f9803e7ea33de4f74b134f8ec0b2b8a1b96a8.jpg) +Fig.3: Illustration of data distribution over 1,000 classes with different $\alpha$ on ImageNet. Best viewed in color. + +First, according to our analysis, $\alpha = 0$ will not update dual variables and implies the accuracy of baseline zero-shot transfer as in CLIP. Then, increasing $\alpha$ can consistently improve the accuracy over the baseline, which shows the efficacy of online label learning. Moreover, the size of the smallest class is also increased as $\alpha$ increases and it shows that our proposed method OnLab with only the online label learning can adjust the class assignment effectively. + +To further investigate the data distribution derived by different $\alpha$ , we compare the size of all 1,000 classes in Fig. 3. Evidently, the proposed method is capable of balancing the distribution in an online manner. Note that even with $\alpha = 1$ , the obtained distribution is not perfectly balanced, which is from the approximation error in the online scenario as analyzed in Theorem 1. We find that our method is not sensitive to $\alpha$ and fix $\alpha = 1$ in the following experiments if not specified. + +Effect of $\beta$ With the predictions from text and vision space in online zero-shot transfer, a combination strategy with $\beta$ in Eqn. 17 can trade-off between the bias and variance. Tab. 2 reports the performance with different $\beta$ . + +Table 2: Effect of $\beta$ in Eqn. 17 for online proxy Table 3: Comparison of strate-learning on ImageNet with ResNet-50 as vision gies for obtaining the combination ratio $\lambda$ in Eqn. 13. + +
β00.20.40.60.81
Acc(%)61.9662.1762.2362.3862.5061.84
+ +
βFixedDynamic
Acc(%)62.2862.50
+ +When $\beta = 0$ , only the label from the online label learning (i.e., OnLab) is adopted for prediction. If mixing the label from the vision proxy, the accuracy improves. The mixture prediction outperforms OnLab by $0.54\%$ with $\beta = 0.8$ and is $2.18\%$ better than the baseline CLIP. If excluding the label from text space at the end of data streaming as $\beta = 1$ , the performance degenerates due to the large variance from the vision space. The phenomenon illustrates the challenge + +of online learning that the vision proxy may not be well learned with a limited number of arrived images and the pseudo label from text proxy can help reduce the variance. + +According to our analysis in Section 3.3, a dynamic ratio for the bias-variance trade-off is more effective for the proposed online learning scenario. To verify the claim, we compare the dynamic ratio to the fixed one in Tab. 3. The value of fixed $\lambda$ is searched in [0, 1] by a step size of 0.1 and the best result is reported. + +While the fixed $\lambda$ still improves the performance from the online label learning, a monotonically increasing $\lambda$ captures the training dynamics better and shows an additional gain of $0.22\%$ . It confirms our analysis about the bias-variance trade-off. + +Effect of iterations in optimization Our method has to learn the dual variables and vision proxy with streaming data and those parameters may not be learned well at the early stage of online learning. To investigate the influence of iterations in online learning, we report the accumulated accuracy of different iterations in one epoch on ImageNet in Tab. 4. + +Table 4: Comparison of accu- Table 5: Effect of epochs for optimization on Imracy $(\%)$ with different itera-ageNet. Prediction from the last epoch is evaluated tions on ImageNet. for OnZeta after running multiple epochs. + +
#Iters5k10k20k
Acc(%)60.9661.5061.71
30k40k50k
62.0262.4162.50
+ +
#Epoch1(online learning)2345
Acc(%)62.5063.2463.3563.3963.46
+ +First, we can observe that even with 5,000 iterations for learning, the accuracy over 1,000 classes is already better than the baseline. With the arrival of more images, the performance increases steadily. It is because that the convergence rate for dual variables and vision proxy is $\mathcal{O}(1 / \sqrt{n})$ as analyzed in Theorems 1-2. It implies that these variables can reach a reasonable performance with a limited number of iterations, while it is challenging to obtain the optimal solution as discussed in the following experiment. + +Effect of epochs in optimization In the proposed online zero-shot transfer problem, all examples can be visited only once when they arrive. The setting is prevalent in real-world applications but very challenging for optimization. This ablation experiment aims to investigate the performance gain when each example can be visited multiple times. + +Concretely, we run the proposed method on multiple epochs of data, where images arrive in a random order within each epoch. For multiple epochs, the prediction from the last one is reported for evaluation. Tab. 5 compares the performance of multiple epochs to that of a single epoch in the online scenario. + +First, we can observe that with an additional epoch, the performance of OnZeta can be boosted by $0.74\%$ , which demonstrates the challenge of online zero-shot transfer. When increasing the number of epochs, the performance can be further improved and it achieves $63.46\%$ with 5-epoch learning. The result approaches the performance of learning with the entire data set as in [16] that shows a $63.74\%$ accuracy when visiting all data by 2,000 epochs. The competitive performance of OnZeta confirms the effectiveness of the proposed online learning algorithm. More experiments can be found in the supplementary. + +# 4.2 Comparison on ImageNet + +After the ablation study, we compare our method OnZeta to state-of-the-art methods on ImageNet. Multiple vision encoders from CLIP are adopted for evaluation, which include ResNet-50, ViT-B/32, ViT-B/16, ViT-L/14, and ViT-L/14@336. Considering that images come in a random order in our online setting, the proposed method is repeated by 5 times and the averaged result is reported for the comparison. We note that the standard deviation is quite small and also report the best result among different runs by *. The same parameters are shared by different vision encoders in our method. Tab. 6 summarizes the comparison, where the performance of InMaP [16] is marked by gray due to the application of the whole unlabeled data set. All other methods only access a single image at each iteration. Our OnLab with only the online label learning is denoted as "OnZeta0.5" and the vanilla zero-shot transfer in Eqn. 1 is "Baseline". + +Table 6: Comparison of accuracy (%) on ImageNet with different vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration. + +
Vision encoderBaseline [17]TPT [19]OnZeta0.5OnZeta*0.5OnZetaOnZeta*InMaP [16]
ResNet-5060.3260.7461.8961.9662.4262.5063.74
ViT-B/3263.77-65.0665.1865.7165.8067.29
ViT-B/1668.7568.9870.1870.2370.8770.9972.55
ViT-L/1475.96-76.9877.0777.8477.9479.29
ViT-L/1433677.02-77.9378.0578.7578.9480.21
+ +First, by capturing the overall distribution using only the proposed online label learning component, $\mathrm{OnZeta}_{0.5}$ already gains about $1\%$ for all vision encoders over the baseline. It demonstrates that the side information from the arrived images is useful for improving the zero-shot transfer. By learning the vision proxy and combining the predictions from both the text and vision spaces, OnZeta shows about $1\%$ additional improvement over $\mathrm{OnZeta}_{0.5}$ , which illustrates that online proxy learning is complementary to online label learning for zero-shot classification. Moreover, OnZeta obtains more gains with a large backbone, showing its potential for large models. + +Then, we can observe that the prompt learning method TPT outperforms the baseline with an ensemble of artificial text prompts, which indicates the importance of the appropriate prompt for zero-shot transfer. However, OnZeta surpasses TPT by $1.76\%$ on ResNet-50 and $2.01\%$ on ViT-B/16. It is because that we learn a better class proxy in the vision space for visual classification, which helps reduce the modality gap and is more effective than optimizing the text prompt in the text space. The observation is consistent with the analysis in [16]. Moreover, compared with InMaP that can access the whole unlabeled set, OnZeta is only about $1\%$ worse using different vision encoders. As analyzed in the ablation study, online learning that visits each example once without storing is more challenging than the offline method with multiple iterations over the entire set. Nevertheless, online learning is practical for real-world applications. Finally, all vision encoders share the same parameters in experiments. It implies that our method OnZeta is not sensitive to the selection of vision encoder and is applicable in different configurations. + +# 4.3 Comparison on Other 13 Downstream Tasks + +Besides the comparison on ImageNet, we conduct experiments on other 13 downstream tasks to evaluate the performance of our method. The benchmark data sets for zero-shot transfer are adopted in the comparison, which includes Aircraft [12], Caltech101 [5], Stanford Cars [10], CIFAR-10 [11], CIFAR-100 [11], CUB200-2011 [21], Describable Textures Dataset (DTD) [3], EuroSAT [9], Flowers [13], Food101 [1], Oxford-IIIT Pet (Pets) [14], Sun397 [23], and UCF101 [20]. A broad range of downstream tasks are involved by these diverse data sets, e.g., low-resolution image classification, fine-grained visual categorization (FGVC), land cover classification with satellite images, scene categorization, texture recognition, etc. These tasks contain fewer images than ImageNet, which increases the challenge of our online transfer with the convergence rate of $\mathcal{O}(1 / \sqrt{n})$ . Therefore, we reduce $\beta$ from 0.8 to 0.4 for all tasks to exploit the label from the text space. For data sets that the baseline already achieves satisfied performance, e.g., Caltech101 and CIFAR-10 with ViT, online label learning can be skipped by letting $\alpha = 0$ , while $\alpha = 0.4$ can provide a slight improvement. Other parameters on the rest data sets remain the same as those for ImageNet. Two distinct vision encoders, i.e., ResNet-50 and ViT-B/16, are applied for evaluation. The performance averaged over 5 random trials is reported for our method. + +The comparison is summarized in Tab. 7. A similar phenomenon as ImageNet can be observed on these downstream tasks. Concretely, online label learning, i.e., $\mathrm{OnZeta}_{0.5}$ , already improves the performance over baselines, where the gain is $3.12\%$ on ResNet and $2.58\%$ on ViT, respectively. Compared with the baseline, only streaming images are leveraged by our method, and the representation of each arrived image is not stored, which preserves the flexibility of zero-shot transfer but captures the distribution over the entire data set in an online manner. In addition, learning the vision proxy with the pseudo label from the text space further improves the performance with different vision encoders, which demonstrates the effectiveness of learning the class proxy directly in the target + +Table 7: Comparison of accuracy (\%) on 13 downstream tasks with ResNet-50 and ViT-B/16 as vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration. + +
AircraftCaltechCarsCifar10Cifar100CUBDTDEuroSATFlowersFoodPetsSUNUCF101Avg.
ResNet-50:
Baseline16.6286.0056.3173.1540.6041.3741.1326.9062.9774.1081.8559.2555.5655.06
TPT17.5887.0258.46---40.8428.3362.6974.8884.4961.4660.82-
OnZeta0.518.2986.1758.5775.6246.9645.6843.2433.7962.4276.4183.0561.6460.1257.84
OnZeta*0.518.5186.4558.9175.7547.2245.9343.7434.3263.1776.5383.5961.8160.3858.18
OnZeta18.3186.2659.4075.8447.5546.0643.3833.9963.1676.8383.7762.3261.0058.30
OnZeta*18.5486.4159.7175.9547.7146.3443.9134.3563.9576.9784.6662.4061.3858.64
InMaP18.9686.7363.3078.8449.2649.1744.8635.8866.6878.3687.9363.8263.3660.55
ViT-B/16:
Baseline23.1393.5166.2991.0967.2949.3645.0950.1766.9984.4387.0065.6865.2165.79
TPT24.7894.1666.87---47.7542.4468.9884.6787.7965.5068.04-
OnZeta0.527.1993.6567.9991.1369.8552.8147.5156.2068.4985.8388.4468.0668.5068.13
OnZeta*0.527.4893.7968.1191.1870.0253.0947.8756.4269.0685.9788.7268.2768.8668.37
OnZeta27.5793.6968.7491.3170.3953.6648.0456.5269.3786.2189.0168.8769.5868.69
OnZeta*28.2993.8369.0391.3570.5253.8048.5856.7469.6386.3589.3269.0169.9468.95
InMaP28.3593.5973.0093.2772.5157.8750.7763.9871.2887.7692.9470.8572.0671.40
+ +vision space. Moreover, OnZeta outperforms TPT on 9 out of 10 (only 10 tasks are available in TPT's original paper) data sets with different vision encoders. Compared with TPT that has to optimize the text prompt with multiple augmentations of the image by back-propagation through the text encoder, OnZeta only contains the forward pass to extract the representation of the image, which is more efficient and is capable of online deployment. By comparing the best performance of the proposed method (i.e., indicated by * in the table) to the averaged one, we can find that the gap averaged over 13 data sets is only about $0.3\%$ . While the images arrive in a random order, the performance of the proposed method can be theoretically guaranteed as indicated in Theorems 1-2, which is robust to the order of input images. Finally, most data sets share the same parameters and it implies that OnZeta is not sensitive to hyper-parameters and is applicable for different tasks. + +# 5 Conclusion + +While CLIP demonstrates impressive zero-shot transfer performance, the information from the target data has not been explored sufficiently. In this work, we investigate an online zero-shot scenario, where a random image arrives at each iteration and the model has to predict its label without storing its representation. By capturing the overall label distribution and refining the class proxy in an online manner, our method OnZeta demonstrates a substantial improvement over the vanilla zero-shot transfer on diverse downstream tasks with different vision encoders. In addition, the operations in the proposed method can be implemented efficiently for real-time applications. Exploring more effective algorithms for challenging online zero-shot transfer can be our future work. + +# References + +1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014) +2. Boyd, S.P., Vandenberghe, L.: Convex Optimization. Cambridge University Press (2014) +3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014) +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021) +5. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004) +6. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. CoRR abs/2110.04544 (2021) +7. Hazan, E.: Introduction to online convex optimization. CoRR abs/1909.05207 (2019) +8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016) +9. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens. 12(7), 2217-2226 (2019) +0. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013) +1. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +2. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) +3. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008) +4. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012) +5. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019) +6. Qian, Q., Xu, Y., Hu, J.: Intra-modal proxy learning for zero-shot visual categorization with clip. In: NeurIPS (2023) +7. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) +8. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015) +9. Shu, M., Nie, W., Huang, D., Yu, Z., Goldstein, T., Anandkumar, A., Xiao, C.: Test-time prompt tuning for zero-shot generalization in vision-language models. In: NeurIPS (2022) + +20. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR abs/1212.0402 (2012) +21. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +22. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. In: CVPR. pp. 7949-7961. IEEE (2022) +23. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010) +24. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of CLIP for few-shot classification. In: ECCV. vol. 13695, pp. 493-510. Springer (2022) +25. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16795-16804. IEEE (2022) +26. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vis. 130(9), 2337-2348 (2022) \ No newline at end of file diff --git a/2024/Online Zero-Shot Classification with CLIP/images.zip b/2024/Online Zero-Shot Classification with CLIP/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..254189eab62a88d0fdaf0542524377941bf4b248 --- /dev/null +++ b/2024/Online Zero-Shot Classification with CLIP/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0868d905954caed04957754d4243d5ff77710752d1c3b28cb82a2faebc837ca9 +size 365567 diff --git a/2024/Online Zero-Shot Classification with CLIP/layout.json b/2024/Online Zero-Shot Classification with CLIP/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8478bd3cff40b4880772312808ede20cd07f00dc --- /dev/null +++ b/2024/Online Zero-Shot Classification with CLIP/layout.json @@ -0,0 +1,9609 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 112, + 455, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 112, + 455, + 128 + ], + "spans": [ + { + "bbox": [ + 159, + 112, + 455, + 128 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 241, + 150, + 373, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 150, + 373, + 163 + ], + "spans": [ + { + "bbox": [ + 241, + 150, + 373, + 163 + ], + "type": "text", + "content": "Qi Qian" + }, + { + "bbox": [ + 241, + 150, + 373, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 241, + 150, + 373, + 163 + ], + "type": "text", + "content": " and Juhua Hu" + }, + { + "bbox": [ + 241, + 150, + 373, + 163 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 199, + 171, + 412, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 171, + 412, + 217 + ], + "spans": [ + { + "bbox": [ + 199, + 171, + 412, + 217 + ], + "type": "text", + "content": "1 Alibaba Group, Bellevue, WA 98004, USA \n2 School of Engineering and Technology, University of Washington, Tacoma, WA 98402, USA \nqi.qian@alibaba-inc.com, juhuah@uw.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 239, + 455, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 455, + 491 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 455, + 491 + ], + "type": "text", + "content": "Abstract. Vision-language pre-training such as CLIP enables zero-shot transfer that can classify images according to the candidate class names. While CLIP demonstrates an impressive zero-shot performance on diverse downstream tasks, the distribution from the target data has not been leveraged sufficiently. In this work, we study a novel online zero-shot transfer scenario, where each image arrives in a random order for classification and is visited only once to obtain prediction immediately without storing its representation. Compared with the vanilla zero-shot classification, the proposed framework preserves its flexibility for online service while considering the statistics of the arrived images as the side information to capture the distribution of target data, which can help improve the performance of real-world applications. To tackle the challenge of effective online optimization, we first develop online label learning to model the target data distribution. Then, the proxy of each class in the vision space is further optimized with the proposed online proxy learning method to mitigate the modality gap between images and text. The convergence of both online strategies can be theoretically guaranteed. By combining the predicted label from the online label learning and proxy learning, our online zero-shot transfer method (OnZeta) achieves " + }, + { + "bbox": [ + 160, + 239, + 455, + 491 + ], + "type": "inline_equation", + "content": "78.94\\%" + }, + { + "bbox": [ + 160, + 239, + 455, + 491 + ], + "type": "text", + "content": " accuracy on ImageNet without accessing the entire data set. Moreover, extensive experiments on other 13 downstream tasks with different vision encoders show a more than " + }, + { + "bbox": [ + 160, + 239, + 455, + 491 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 160, + 239, + 455, + 491 + ], + "type": "text", + "content": " improvement on average, which demonstrates the effectiveness of our proposal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 501, + 410, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 501, + 410, + 512 + ], + "spans": [ + { + "bbox": [ + 160, + 501, + 410, + 512 + ], + "type": "text", + "content": "Keywords: Online learning " + }, + { + "bbox": [ + 160, + 501, + 410, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 501, + 410, + 512 + ], + "type": "text", + "content": " Zero-shot classification " + }, + { + "bbox": [ + 160, + 501, + 410, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 501, + 410, + 512 + ], + "type": "text", + "content": " CLIP" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 230, + 542 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": "Vision-language pre-training has attracted much attention recently due to its impressive zero-shot transfer performance on various downstream tasks. The desired property mainly comes from aligning the vision and text space. For example, one of the most prevalent pre-training methods, i.e., CLIP [17], consists of distinct vision and text encoders for learning image and text representations, respectively. These encoders are optimized by minimizing a contrastive loss defined on image-text pairs. The loss aims to pull images and their corresponding text descriptions together while pushing the irrelevant text or images away [15]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 114, + 425, + 205 + ], + "blocks": [ + { + "bbox": [ + 189, + 114, + 425, + 205 + ], + "lines": [ + { + "bbox": [ + 189, + 114, + 425, + 205 + ], + "spans": [ + { + "bbox": [ + 189, + 114, + 425, + 205 + ], + "type": "image", + "image_path": "e6c7bd7fc911e1039ee20cfd016219ef8d5d01bfed9ffdcc7c4133fcb6b5e486.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 214, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 270 + ], + "type": "text", + "content": "Fig. 1: Illustration of the proposed online zero-shot transfer method (OnZeta). Blue and orange lines denote the inference in the text and vision space, respectively. By incorporating the predictions from online label learning and online proxy learning, OnZeta can leverage the biased prediction from the text space to reduce the variance in the target vision space in an online manner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 294, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 413 + ], + "type": "text", + "content": "After the alignment between vision and text, a novel zero-shot classification paradigm emerges to transfer knowledge in pre-trained models to diverse tasks. Concretely, given all class names from the target task, each class can be represented by a class proxy that applies the text encoder in CLIP for the corresponding class name, which is also termed a text proxy. When an image arrives, its representation is extracted by the vision encoder of CLIP. Then, the label can be assigned by the 1-nearest neighbor (1-NN) classifier with the set of text proxies as the reference. Without any fine-tuning, this simple strategy achieves " + }, + { + "bbox": [ + 130, + 294, + 482, + 413 + ], + "type": "inline_equation", + "content": "77.02\\%" + }, + { + "bbox": [ + 130, + 294, + 482, + 413 + ], + "type": "text", + "content": " accuracy on ImageNet [18] using the pre-trained ViT [4] as the vision encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 617 + ], + "type": "text", + "content": "Since the success of CLIP, many methods have been developed to explore different side information from the target task to further improve the transfer performance. Firstly, if a few labeled images are available from the target task, the zero-shot transfer can be formulated as a few-shot learning problem and the input text prompt can be learned with the labeled vision data by fine-tuning [26]. Secondly, if the computational resource is too limited to do fine-tuning, the labeled data can be cached as a vision classifier to classify images directly [24]. However, supervised information is often rare, while unsupervised information is more accessible. Therefore, many efforts have been devoted to exploiting the unlabeled images to enhance the zero-shot performance. [19] generates multiple augmented examples from a given image, and then learns the specific text prompt for each image as few-shot learning accordingly. [16] considers reconstructing the class proxy in the vision space, i.e., vision proxy, by leveraging a set of unlabeled images in proxy learning. While [16] shows superior zero-shot performance, it optimizes proxies in an offline manner. Due to the privacy issue, gathering or keeping a set of unlabeled images is still challenging in some real-world applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Therefore, in this work, we investigate a novel practical scenario for zero-shot transfer, termed online zero-shot transfer. Concretely, when streaming images arrive as in conventional zero-shot learning, the model has to classify the image immediately without refining. After that, the class proxy can be updated but" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 174, + 116, + 438, + 193 + ], + "blocks": [ + { + "bbox": [ + 174, + 116, + 438, + 193 + ], + "lines": [ + { + "bbox": [ + 174, + 116, + 438, + 193 + ], + "spans": [ + { + "bbox": [ + 174, + 116, + 438, + 193 + ], + "type": "image", + "image_path": "121f387901d4124e3440254aca3c273c8ed4ef6a511d5a4c5faca2af3d5e32c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 204, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 225 + ], + "type": "text", + "content": "Fig. 2: Top 4 predicted labels with corresponding probabilities from baseline zero-shot method by CLIP and our proposal." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 252, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 480, + 335 + ], + "type": "text", + "content": "the representations of arrived images cannot be kept. Unlike [16] that can access the whole unlabeled set, this online setting is way more challenging, where only statistics of seen images can be leveraged for optimization and each image is visited only once. Thereafter, the standard iterative optimization in [16] is not applicable, while the problem (i.e., online service to classify each unlabeled image on the fly without storing data) is ubiquitous in real-world applications (e.g., mobiles and robots)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 563 + ], + "type": "text", + "content": "To address the online service challenge, we first consider exploring the data distribution of the target task to improve the prediction from the text proxy. Note that the vanilla zero-shot prediction focuses on individual images and ignores the distribution over the entire data set. To capture the distribution of all data, we develop an online label learning algorithm to balance the assignment between different classes globally. In addition, the class proxy from the text (i.e., text proxy) can be biased for the vision data due to the modality gap, and thus we propose online proxy learning to obtain the class proxy in the vision space (i.e., vision proxy) directly in an online manner. Finally, the prediction from both the text space and vision space can be combined to help reduce the variance of learning in the vision space as illustrated in the framework of our online zero-shot transfer (OnZeta) shown in Fig. 1. More importantly, with online optimization, our proposal OnZeta can leverage the information from target data effectively and efficiently, which helps correct the prediction compared to the baseline zero-shot method. For example, Fig. 2 shows that an image of \"Husky\" will be misclassified by vanilla CLIP. By considering the distribution over the whole data set and incorporating the learned vision proxy, our method can obtain the right label in an online manner. The main contributions of this work can be summarized as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 580, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 580, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 479, + 628 + ], + "type": "text", + "content": "- Considering the practical need for online service (i.e., classifying each image on the fly without storage) and improving the performance of vanilla zero-shot classification, we study a novel setting, that is, online zero-shot transfer, using pre-trained vision-language models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "text", + "content": "- To address the online service challenge, we propose an online label learning method to trace the distribution of assigned labels over the entire data in an online manner. Moreover, an online vision proxy learning method is proposed" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 101 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "content": "to further improve the class proxy in vision space for classification. The convergence of both methods can be theoretically guaranteed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 187 + ], + "type": "text", + "content": "- Experiments on 14 downstream tasks demonstrate the effectiveness of our proposal OnZeta. Concretely, when visiting each image only once, OnZeta outperforms the baseline by a margin of more than " + }, + { + "bbox": [ + 138, + 140, + 480, + 187 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 138, + 140, + 480, + 187 + ], + "type": "text", + "content": " on average without storing any unlabeled data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 237, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 237, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 237, + 218 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 232, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 480, + 266 + ], + "type": "text", + "content": "CLIP consists of two encoders (i.e., vision and text), and the optimization can be conducted on each of them individually. In this section, we briefly review research efforts in these two directions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 324, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 324, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 324, + 297 + ], + "type": "text", + "content": "2.1 Optimization with Text Encoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 305, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 481, + 437 + ], + "type": "text", + "content": "CLIP demonstrates that the zero-shot performance heavily depends on the text prompts for generating appropriate class proxies, and many methods investigate learning the optimal text prompt as input to the text encoder. CoOp [26] marks the input text prompts as learnable variables and optimizes them via few-shot learning with a limited number of labeled images for each class. CoCoOp [25] further includes the vision representation as the conditional context for prompt learning. TPT [19] relaxes the requirement of labeled examples and optimizes the consistency between different augmentations from the same image for prompt learning. However, it has to optimize the prompt for each image with 64 augmentations, which is inefficient for computation and memory, making it infeasible for real-time applications as the online scenario in this work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 454, + 332, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 332, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 332, + 465 + ], + "type": "text", + "content": "2.2 Optimization with Vision Encoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "content": "While prompt learning can improve the performance on the target data, recent research shows that optimizing the vision representation can be more effective for classification [6,16]. CLIP-Adapter [6] introduces an additional adapter network after the vision encoder and optimizes the vision representation with few-shot learning. Although a similar adapter can be attached for the text encoder, the empirical study shows that optimizing the vision space is sufficient for transfer and can achieve better performance than either the version with only a text adapter or that with both vision and text adapters. [22] shows that fine-tuning the vision classifier directly with few-shot learning can achieve better performance than CoOp. Recently, the analysis in [16] demonstrates that the optimal class proxy for visual classification lives in the vision space. To mitigate the modality gap, [16] proposes proxy learning that aims to learn class proxy for each class in vision space. However, it requires a set of unlabeled data and will learn proxies iteratively. In this work, we also aim to refine the class proxy in the vision space, but with the learned labels in an online manner, which is more applicable for online real applications." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 307, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 307, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 307, + 127 + ], + "type": "text", + "content": "3 Online Zero-shot Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 299, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 299, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 299, + 152 + ], + "type": "text", + "content": "3.1 Zero-shot Transfer in CLIP" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": "Given a data set consisting of image-text pairs as " + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "inline_equation", + "content": "\\{I_i, T_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": " denotes an image and " + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": " is the corresponding text such as class names or captions. CLIP trains vision and text encoders jointly to align the vision subspace and its corresponding text subspace. Let " + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": " denote the vision encoder and text encoder, respectively. The representations for images and text can be extracted as " + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f(I_i)" + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i = g(T_i)" + }, + { + "bbox": [ + 130, + 160, + 482, + 233 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "text", + "content": "After pre-training two encoders, zero-shot classification can be implemented by the 1-nearest neighbor (1-NN) classifier. Concretely, given the class names of the target task, the proxy of " + }, + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "text", + "content": "-th class " + }, + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "text", + "content": " can be obtained by applying the text encoder for the prompt as \"a photo of a {class name}\". Then, the image " + }, + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 233, + 482, + 293 + ], + "type": "text", + "content": " will be classified as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 301, + 481, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 301, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 264, + 301, + 481, + 321 + ], + "type": "interline_equation", + "content": "y _ {i} = \\arg \\max _ {j} \\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\tag {1}", + "image_path": "943c5ba88ed6af19aadc00e65ab08d0c70ee55d23d7ce1ded88c42b94f9db733.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 329, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 413 + ], + "type": "text", + "content": "While the conventional zero-shot paradigm demonstrates an impressive transfer performance, the side information from the target task has not been explored sufficiently. Recent work shows that with only a set of unlabeled target data, the zero-shot performance can be improved dramatically [16]. In this work, we further relax the requirement of unlabeled target data and investigate a novel online zero-shot transfer scenario, where each unlabeled image arrives in an online manner without storing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 431, + 274, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 274, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 274, + 444 + ], + "type": "text", + "content": "3.2 Online Label Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 453, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 475 + ], + "type": "text", + "content": "First, we develop an online label learning method to leverage the information from arrived images to capture the distribution over the entire set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 476, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 481, + 500 + ], + "type": "text", + "content": "Given text proxy " + }, + { + "bbox": [ + 130, + 476, + 481, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 130, + 476, + 481, + 500 + ], + "type": "text", + "content": ", the predicted distribution over classes for " + }, + { + "bbox": [ + 130, + 476, + 481, + 500 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 476, + 481, + 500 + ], + "type": "text", + "content": " can be computed as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 510, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 510, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 249, + 510, + 481, + 539 + ], + "type": "interline_equation", + "content": "q _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {j} / \\tau_ {T}\\right)}{\\sum_ {k} ^ {C} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {z} _ {k} / \\tau_ {T}\\right)} \\tag {2}", + "image_path": "42e387e9b7e1c3b59937ef1557eed71337bfb610a2e39873022d422bf13082f3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 546, + 347, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 347, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 347, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 546, + 347, + 558 + ], + "type": "inline_equation", + "content": "\\tau_{T}" + }, + { + "bbox": [ + 131, + 546, + 347, + 558 + ], + "type": "text", + "content": " is the temperature optimized by CLIP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 559, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 481, + 594 + ], + "type": "text", + "content": "If the whole image set is available, the label can be refined by considering the distribution over classes. Let " + }, + { + "bbox": [ + 130, + 559, + 481, + 594 + ], + "type": "inline_equation", + "content": "\\{p_i\\}" + }, + { + "bbox": [ + 130, + 559, + 481, + 594 + ], + "type": "text", + "content": " denote the learnable labels, and the optimization problem with " + }, + { + "bbox": [ + 130, + 559, + 481, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 559, + 481, + 594 + ], + "type": "text", + "content": " classes can be cast as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 189, + 604, + 481, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 604, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 189, + 604, + 481, + 631 + ], + "type": "interline_equation", + "content": "\\min _ {p _ {i} \\in \\Delta} \\frac {1}{n} \\sum_ {i} D _ {K L} \\left(p _ {i} \\| q _ {i}\\right) \\quad s. t. \\quad \\forall j, \\quad \\frac {1}{n} \\sum_ {i} p _ {i, j} \\geq \\frac {\\alpha}{C} \\tag {3}", + "image_path": "f4afe8263d7294db231c97a8f1613cff662d581c45a3ac70db1bf262f9af7743.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "p_i = [p_{i,1}, \\ldots, p_{i,C}]" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "q_i = [q_{i,1}, \\ldots, q_{i,C}]" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": " denotes the simplex as " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\Delta = \\{p_i \\in \\mathbb{R}^C | \\sum_j^C p_{i,j} = 1; \\forall j, p_{i,j} \\geq 0\\}" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": ". The constraint bounds the minimal" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 273, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 273, + 91, + 448, + 101 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 136, + 132, + 481, + 243 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 342, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 342, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 342, + 129 + ], + "type": "text", + "content": "Algorithm 1 Online Label Learning (OnLab)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "lines": [ + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": "1: Input: Unlabeled image set " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\{I_i\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": ", class names " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\{T_j\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": ", pre-trained CLIP encoders " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "(f, g)" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": ", dual variables " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\{\\rho_j\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": ", ratio " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": ", temperature " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\tau_T" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": ", initial learning rate " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "c_{\\rho}" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": " \n2: Obtain representation of text proxy as " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j = g(T_j)" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": " \n3: Initialize " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\forall j, \\rho_j = 0" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": " \n4: while the " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": "-th image arrives do \n5: Obtain vision representation as " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f(I_i)" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": " \n6: Observe the probability " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "p_i^*" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": " by Eqn. 6 \n7: Have the prediction as " + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "inline_equation", + "content": "y_i = \\arg \\max_j p_{i,j}^*" + }, + { + "bbox": [ + 136, + 132, + 481, + 243 + ], + "type": "text", + "content": " \n8: Update dual variables as in Eqn. 7 \n9: end while" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "content": "size of each class to avoid collapsing and " + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "content": " is the ratio for balancing the assignment between different classes. " + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "content": " denotes a well-balanced distribution that each class has the same proportion of images, while " + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "content": " keeps the prediction of the original zero-shot classification from " + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "type": "text", + "content": "The problem can be solved effectively given all examples, but optimizing the objective with streaming images is challenging. To address the online problem, we introduce the Lagrange dual variables [2] " + }, + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\{\\rho_j\\}" + }, + { + "bbox": [ + 130, + 314, + 480, + 361 + ], + "type": "text", + "content": " for constraints and rewrite the problem as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 168, + 368, + 480, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 368, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 168, + 368, + 480, + 402 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\rho , p _ {i}) = \\max _ {\\rho_ {j} \\geq 0} \\min _ {p _ {i} \\in \\Delta} \\frac {1}{n} \\sum_ {i} D _ {K L} (p _ {i} | | q _ {i}) - \\sum_ {j} ^ {C} \\rho_ {j} \\left(\\frac {1}{n} \\sum_ {i} p _ {i, j} - \\frac {\\alpha}{C}\\right) \\tag {4}", + "image_path": "8e11b57aa520664e82191a93d271d96634db83530903dfd03d4f59d8a1ad6577.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": "When the " + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": "-th example arrives for prediction, the sub-problem with fixed dual variables becomes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 437, + 480, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 437, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 216, + 437, + 480, + 470 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\prime} \\left(p _ {i}\\right) = \\min _ {p _ {i} \\in \\Delta} D _ {K L} \\left(p _ {i} \\mid \\mid q _ {i}\\right) - \\sum_ {j} ^ {C} \\rho_ {j} ^ {i - 1} p _ {i, j} \\tag {5}", + "image_path": "abea6e839ec6cabc77e0baa42ce7d260fc15df417c2e34e388242586b6257660.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 478, + 465, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 465, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 465, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 478, + 465, + 491 + ], + "type": "inline_equation", + "content": "\\rho_j^{i - 1}" + }, + { + "bbox": [ + 130, + 478, + 465, + 491 + ], + "type": "text", + "content": " indicates the dual variable used for the " + }, + { + "bbox": [ + 130, + 478, + 465, + 491 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 130, + 478, + 465, + 491 + ], + "type": "text", + "content": "-th arrived image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 491, + 433, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 491, + 433, + 502 + ], + "spans": [ + { + "bbox": [ + 146, + 491, + 433, + 502 + ], + "type": "text", + "content": "Fortunately, this problem has the closed-form solution as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 508, + 420, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 420, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 420, + 521 + ], + "type": "text", + "content": "Proposition 1 The optimal solution to the problem in Eqn. 5 is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 247, + 526, + 480, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 526, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 247, + 526, + 480, + 558 + ], + "type": "interline_equation", + "content": "p _ {i, j} ^ {*} = \\frac {q _ {i , j} \\exp \\left(\\rho_ {j} ^ {i - 1}\\right)}{\\sum_ {k} ^ {C} q _ {i , k} \\exp \\left(\\rho_ {k} ^ {i - 1}\\right)} \\tag {6}", + "image_path": "aa572ad3db920f54550d0c137868b2ad1ed7c1f42e911d40c8c4f7ee25462191.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 563, + 318, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 318, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 318, + 576 + ], + "type": "text", + "content": "Proof. It is from the K.K.T. condition [2]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 586, + 479, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 479, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 479, + 620 + ], + "type": "text", + "content": "Remark Proposition 1 shows that the dual variable helps balance the assignment between classes. When " + }, + { + "bbox": [ + 130, + 586, + 479, + 620 + ], + "type": "inline_equation", + "content": "\\rho = 0" + }, + { + "bbox": [ + 130, + 586, + 479, + 620 + ], + "type": "text", + "content": " without optimization, it degenerates to the original prediction from CLIP." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 622, + 479, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 479, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 479, + 646 + ], + "type": "text", + "content": "With the prediction for the image, the dual variable can be updated by gradient ascent as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 224, + 651, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 651, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 224, + 651, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\rho_ {j} ^ {i} = \\max \\left\\{0, \\rho_ {j} ^ {i - 1} - \\eta_ {\\rho} ^ {i} \\left(p _ {i, j} ^ {*} - \\alpha / C\\right) \\right\\} \\tag {7}", + "image_path": "a9c8ff769fb5a7d8e482dd2d847278847b04b27ca48bf8bb8657344f9b4dd88d.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\eta_{\\rho}^{i}" + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": " is the learning rate for the dual variable. When " + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\{\\rho_j\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": " is initialized as 0 and " + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": ", dual variables will not be updated, which leads to the original prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 199 + ], + "type": "text", + "content": "Our online label learning (OnLab) algorithm is summarized in Alg. 1. The prediction will be obtained by a close-form solution and the updating for dual variables is also efficient, which makes the proposed method applicable for real-time applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 236 + ], + "type": "text", + "content": "Besides the efficiency, the effectiveness of OnLab can be theoretically guaranteed in the following theorem. The detailed proof can be found in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 242, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 479, + 268 + ], + "type": "text", + "content": "Theorem 1 By running Alg. 1 for solving the problem in Eqn. 4 with the step size " + }, + { + "bbox": [ + 132, + 242, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\eta_{\\rho}^{i} = c_{\\rho} / \\sqrt{i}" + }, + { + "bbox": [ + 132, + 242, + 479, + 268 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 242, + 479, + 268 + ], + "type": "inline_equation", + "content": "c_{\\rho}" + }, + { + "bbox": [ + 132, + 242, + 479, + 268 + ], + "type": "text", + "content": " is a constant, the convergence can be guaranteed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 274, + 481, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 274, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 194, + 274, + 481, + 301 + ], + "type": "interline_equation", + "content": "\\frac {1}{n} \\max _ {\\rho} \\sum_ {i} \\mathcal {L} \\left(\\rho , p _ {i} ^ {*}\\right) - \\min _ {p _ {i}} \\sum_ {i} \\mathcal {L} \\left(\\rho^ {i - 1}, p _ {i}\\right) \\leq \\mathcal {O} \\left(\\frac {1}{\\sqrt {n}}\\right) \\tag {8}", + "image_path": "0db477de14967f3fc4b7e445f9b3f58e6858565395b8b64b653756a014a405f4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 317, + 276, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 317, + 276, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 317, + 276, + 330 + ], + "type": "text", + "content": "3.3 Online Proxy Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 479, + 384 + ], + "type": "text", + "content": "After optimizing the label by online label learning, we consider reconstructing the class proxy in the vision space to reduce the modality gap from the text proxy as suggested in [16]. Given the whole set of images, vision proxy learning can be cast as an optimization problem" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 230, + 391, + 481, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 391, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 230, + 391, + 481, + 416 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {w}} \\mathcal {L} (\\mathbf {w}) = - \\sum_ {i} \\sum_ {j} p _ {i, j} ^ {*} \\log \\left(p _ {i, j} ^ {\\prime}\\right) \\tag {9}", + "image_path": "f43dd5c97e2f0ba6443019ccf763ddb38c0e480d85604abcca37ae278a84d5a5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 423, + 361, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 361, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 361, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 423, + 361, + 437 + ], + "type": "inline_equation", + "content": "p_{i,j}^{*}" + }, + { + "bbox": [ + 132, + 423, + 361, + 437 + ], + "type": "text", + "content": " is the output of online label learning and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 445, + 481, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 445, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 249, + 445, + 481, + 472 + ], + "type": "interline_equation", + "content": "p _ {i, j} ^ {\\prime} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j} / \\tau_ {I}\\right)}{\\sum_ {k} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {k} / \\tau_ {I}\\right)} \\tag {10}", + "image_path": "7c2f1ef30e0b15dcb4190566f7075f18cb75192950e828ad5f5e07852f763f84.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j" + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "text", + "content": " denotes the learnable vision proxy for the " + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "text", + "content": "-th class. " + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\tau_I" + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "text", + "content": " is the temperature that should be larger than " + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "inline_equation", + "content": "\\tau_T" + }, + { + "bbox": [ + 130, + 478, + 479, + 502 + ], + "type": "text", + "content": " to obtain the class proxy in the vision space [16]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 502, + 479, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 479, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 479, + 537 + ], + "type": "text", + "content": "When solving the problem in an online manner, only a single example will be received at each iteration. Therefore, the vision proxy will be updated according to the gradient from the " + }, + { + "bbox": [ + 130, + 502, + 479, + 537 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 502, + 479, + 537 + ], + "type": "text", + "content": "-th example as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 216, + 545, + 481, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 545, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 216, + 545, + 481, + 559 + ], + "type": "interline_equation", + "content": "\\mathbf {w} ^ {i} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\mathbf {w} ^ {i - 1} - \\eta_ {w} ^ {i} \\nabla \\mathcal {L} \\left(\\mathbf {w} ^ {i - 1}, \\mathbf {x} _ {i}\\right)\\right) \\tag {11}", + "image_path": "75c54bbd4afc5e32c01b708baa862a14264e2d69221ff6022a530c894fec8f01.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 565, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 479, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 565, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 130, + 565, + 479, + 602 + ], + "type": "text", + "content": " projects the updated vision proxy to the unit norm. Note that the loss function is convex in " + }, + { + "bbox": [ + 130, + 565, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 565, + 479, + 602 + ], + "type": "text", + "content": ", and the convergence can be guaranteed by the standard online learning theory [7]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "text", + "content": "Theorem 2 By updating " + }, + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "text", + "content": " as in Eqn. 11 and setting " + }, + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "inline_equation", + "content": "\\eta_w^i = c_w / \\sqrt{i}" + }, + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "inline_equation", + "content": "c_{w}" + }, + { + "bbox": [ + 132, + 609, + 479, + 634 + ], + "type": "text", + "content": " is a constant, the regret can be bounded as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 216, + 640, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 640, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 216, + 640, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\frac {1}{n} \\sum_ {i} \\mathcal {L} \\left(\\mathbf {w} ^ {i - 1}, \\mathbf {x} _ {i}\\right) - \\min _ {\\mathbf {w}} \\mathcal {L} (\\mathbf {w}) \\leq \\mathcal {O} \\left(\\frac {1}{\\sqrt {n}}\\right) \\tag {12}", + "image_path": "21300a23def179be5dfa741a3d652a65ff697cedf6d70edfb19aac8f55f83ff2.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 273, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 273, + 91, + 448, + 101 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 136, + 132, + 481, + 274 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 362, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 362, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 362, + 129 + ], + "type": "text", + "content": "Algorithm 2 Online Zero-shot Transfer (OnZeta)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "lines": [ + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": "1: Input: Unlabeled image set " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\{I_i\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", class names " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\{T_j\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", pre-trained CLIP encoders " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "(f, g)" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", dual variables " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\{\\rho_j\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", vision proxy " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_j\\}" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", ratio " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", temperature " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\tau_T" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\tau_I" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", initial learning rate " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "c_{\\rho}" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "c_w" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": ", weight parameter " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " \n2: Obtain representation of text proxy as " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j = g(T_j)" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " \n3: Initialize " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\forall j, \\rho_j = 0" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j = \\mathbf{z}_j" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " \n4: while the " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": "-th image arrives do \n5: Obtain vision representation as " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f(I_i)" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " \n6: Observe the probability " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "p_i^*" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " by Eqn. 6 \n7: Observe the probability " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "p_i'" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " by Eqn. 10 \n8: Have the prediction as " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "y_i = \\arg \\max_j \\tilde{p}_{i,j}" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\tilde{p}_i" + }, + { + "bbox": [ + 136, + 132, + 481, + 274 + ], + "type": "text", + "content": " is from Eqn. 13 \n9: Update dual variables as in Eqn. 7 \n10: Update vision proxy as in Eqn. 11 \n11: end while" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 300, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 384 + ], + "type": "text", + "content": "Since images arrive in a random order and only a single image will be leveraged for updating " + }, + { + "bbox": [ + 130, + 300, + 479, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 300, + 479, + 384 + ], + "type": "text", + "content": ", the variance in learning can be large, especially for the first few iterations. Unlike offline optimization in [16], where the performance is evaluated after the final iteration, online learning has to accumulate the accuracy of predictions from different timestamps. Therefore, we consider combining the pseudo label " + }, + { + "bbox": [ + 130, + 300, + 479, + 384 + ], + "type": "inline_equation", + "content": "p_i^*" + }, + { + "bbox": [ + 130, + 300, + 479, + 384 + ], + "type": "text", + "content": " from online label learning in text space to stabilize the prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "text", + "content": "For the " + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "text", + "content": "-th example, assume that " + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "text", + "content": " is the ground-truth distribution in the vision space. Since " + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "inline_equation", + "content": "p_i^*" + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "text", + "content": " is mainly from the text proxy, it is a biased estimation from the text space. While " + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "inline_equation", + "content": "p_i'" + }, + { + "bbox": [ + 130, + 385, + 480, + 445 + ], + "type": "text", + "content": " is estimated by the vision proxy, the variance can be large due to the online updating. Therefore, these predictions can be mixed to trade-off bias from text space and variance from vision space. Concretely, let" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 254, + 455, + 480, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 455, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 254, + 455, + 480, + 469 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {i} = \\lambda_ {i} p _ {i} ^ {\\prime} + (1 - \\lambda_ {i}) p _ {i} ^ {*} \\tag {13}", + "image_path": "6da648c4dcf490564f99c6e5093a505b089ca6b8ad06a05601d30bff03b06b10.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 479, + 192, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 192, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 192, + 488 + ], + "type": "text", + "content": "then we have" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 500, + 480, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 500, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 147, + 500, + 480, + 530 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E [ \\| \\tilde {p} _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} ] \\\\ = E \\left[ \\lambda_ {i} ^ {2} \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} + (1 - \\lambda_ {i}) ^ {2} \\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2} + 2 \\lambda (1 - \\lambda) \\langle p _ {i} ^ {\\prime} - \\hat {p} _ {i}, p _ {i} ^ {*} - \\hat {p} _ {i} \\rangle \\right] \\tag {14} \\\\ \\end{array}", + "image_path": "5edf1e3eac06a19a6c9644bdec0ab049149d96b0f716cb97aded4541c2a47028.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "text", + "content": "For the sake of simplicity, we assume that " + }, + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "inline_equation", + "content": "p_i'" + }, + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "text", + "content": " is an unbiased estimation for " + }, + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "inline_equation", + "content": "E[p_i'] = \\hat{p}_i" + }, + { + "bbox": [ + 130, + 540, + 479, + 564 + ], + "type": "text", + "content": ". Then" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 183, + 572, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 572, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 183, + 572, + 480, + 601 + ], + "type": "interline_equation", + "content": "E _ {w} \\left[ \\| \\tilde {p} _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\right] = \\lambda_ {i} ^ {2} \\underbrace {E \\left[ \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\right]} _ {\\text {v a r i a n c e}} + (1 - \\lambda_ {i}) ^ {2} \\underbrace {\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2}} _ {\\text {b i a s}} \\tag {15}", + "image_path": "0809803256b98b25dbd3e89f9d28b6e318f01074bc6e536a6db61a704d2ba4e7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 610, + 479, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 479, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 479, + 633 + ], + "type": "text", + "content": "To minimize the approximation error, the trade-off ratio " + }, + { + "bbox": [ + 130, + 610, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 130, + 610, + 479, + 633 + ], + "type": "text", + "content": " has the optimal solution as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 232, + 641, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 641, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 232, + 641, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\lambda_ {i} ^ {*} = \\frac {\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2}}{\\| p _ {i} ^ {*} - \\hat {p} _ {i} \\| _ {2} ^ {2} + E [ \\| p _ {i} ^ {\\prime} - \\hat {p} _ {i} \\| _ {2} ^ {2} ]} \\tag {16}", + "image_path": "a7c819e57dac3cc5861ca5c9496c2f9e73f1b4c8052d09629ba23fdba878802d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "We find that the bias can be considered as a constant but the variance will be reduced with the learning of " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ", which implies a monotonically increasing function for combining the pseudo labels from different spaces. While various weight functions for " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " can be developed according to the analysis, for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "-th iteration, we empirically set" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 277, + 189, + 481, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 189, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 277, + 189, + 481, + 201 + ], + "type": "interline_equation", + "content": "\\lambda_ {i} = \\beta \\sqrt {i / n} \\tag {17}", + "image_path": "f63349cef6781d833d1b1d61944eab1a1b4cfcb30483454651ed1b79f9a27fac.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 481, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 213, + 481, + 237 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 130, + 213, + 481, + 237 + ], + "type": "text", + "content": " is a constant. The combination strategy helps online transfer when the vision proxy is not trained sufficiently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 481, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 481, + 286 + ], + "type": "text", + "content": "Incorporating online label learning and online proxy learning, the proposed online zero-shot transfer (OnZeta) algorithm is summarized in Alg. 2. All operations in the proposed method can be computed efficiently without significantly increasing the cost of conventional zero-shot transfer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 313, + 230, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 230, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 230, + 327 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": "To evaluate the proposed method, we conduct experiments on 14 diverse downstream tasks. Considering that text prompts are essential for obtaining the appropriate text proxy for zero-shot transfer, we follow the common practice of including an ensemble of 7 text prompts as suggested in [17] to generate the text proxy for baseline and our method. The initial learning rates for dual variables and vision proxy are set as " + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "inline_equation", + "content": "c_{\\rho} = 20" + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "inline_equation", + "content": "c_{w} = 0.5" + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": ", respectively. The learning rate will be decayed as " + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\eta_{\\rho}^{i} = c_{\\rho} / \\sqrt{i}" + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\eta_w^i = c_w / \\sqrt{i}" + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": " denotes the order of the incoming image. Unlabeled images will arrive in a random order. For the temperature, we fix " + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\tau_T = 0.01" + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": " that is the optimized value in CLIP and " + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\tau_I = 0.04" + }, + { + "bbox": [ + 130, + 346, + 482, + 468 + ], + "type": "text", + "content": " as suggested in [16]. All experiments are implemented on a single V100 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 492, + 238, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 238, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 238, + 504 + ], + "type": "text", + "content": "4.1 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "text", + "content": "The main parameters in our method are the ratio for class assignment " + }, + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "text", + "content": " and the ratio for pseudo label combination " + }, + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "text", + "content": ". In this subsection, we will investigate the influence of these parameters along with the setting of online zero-shot transfer. The ablation experiments are conducted on ImageNet with ResNet-50 [8] as vision encoder in CLIP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " is the ratio to capture the distribution over the entire data set as in Eqn. 3. While the vanilla zero-shot transfer ignores the side information from the arrived examples, the proposed online label learning can leverage the information of distribution by " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", we vary " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\{0,0.4,0.6,0.8,1\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " and summarize the results in Tab. 1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 207, + 279, + 281 + ], + "blocks": [ + { + "bbox": [ + 146, + 140, + 288, + 205 + ], + "lines": [ + { + "bbox": [ + 146, + 140, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 146, + 140, + 288, + 205 + ], + "type": "text", + "content": "Table 1: Effect of " + }, + { + "bbox": [ + 146, + 140, + 288, + 205 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 146, + 140, + 288, + 205 + ], + "type": "text", + "content": " in Eqn. 3 for online label learning on ImageNet with ResNet-50 as vision encoder. \"Min class\" denotes the proportion of examples in the smallest class." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 207, + 279, + 281 + ], + "lines": [ + { + "bbox": [ + 157, + 207, + 279, + 281 + ], + "spans": [ + { + "bbox": [ + 157, + 207, + 279, + 281 + ], + "type": "table", + "html": "
αAcc(%)#Min class(%)
060.3213.7
0.460.4118.5
0.660.6323.0
0.861.2033.3
161.9639.0
", + "image_path": "32daa4409868a49195f7eb823094287b1d551dc02294f4abda66e8f7d0354bc8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 293, + 118, + 477, + 253 + ], + "blocks": [ + { + "bbox": [ + 293, + 118, + 477, + 253 + ], + "lines": [ + { + "bbox": [ + 293, + 118, + 477, + 253 + ], + "spans": [ + { + "bbox": [ + 293, + 118, + 477, + 253 + ], + "type": "image", + "image_path": "cc944ef65f3a7865e6f00d1e563f9803e7ea33de4f74b134f8ec0b2b8a1b96a8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 263, + 478, + 296 + ], + "lines": [ + { + "bbox": [ + 288, + 263, + 478, + 296 + ], + "spans": [ + { + "bbox": [ + 288, + 263, + 478, + 296 + ], + "type": "text", + "content": "Fig.3: Illustration of data distribution over 1,000 classes with different " + }, + { + "bbox": [ + 288, + 263, + 478, + 296 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 288, + 263, + 478, + 296 + ], + "type": "text", + "content": " on ImageNet. Best viewed in color." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "content": "First, according to our analysis, " + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "content": " will not update dual variables and implies the accuracy of baseline zero-shot transfer as in CLIP. Then, increasing " + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "content": " can consistently improve the accuracy over the baseline, which shows the efficacy of online label learning. Moreover, the size of the smallest class is also increased as " + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 300, + 480, + 372 + ], + "type": "text", + "content": " increases and it shows that our proposed method OnLab with only the online label learning can adjust the class assignment effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "text", + "content": "To further investigate the data distribution derived by different " + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "text", + "content": ", we compare the size of all 1,000 classes in Fig. 3. Evidently, the proposed method is capable of balancing the distribution in an online manner. Note that even with " + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "text", + "content": ", the obtained distribution is not perfectly balanced, which is from the approximation error in the online scenario as analyzed in Theorem 1. We find that our method is not sensitive to " + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "text", + "content": " and fix " + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 373, + 481, + 456 + ], + "type": "text", + "content": " in the following experiments if not specified." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "text", + "content": " With the predictions from text and vision space in online zero-shot transfer, a combination strategy with " + }, + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "text", + "content": " in Eqn. 17 can trade-off between the bias and variance. Tab. 2 reports the performance with different " + }, + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 472, + 481, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 152, + 552, + 343, + 578 + ], + "blocks": [ + { + "bbox": [ + 146, + 517, + 479, + 550 + ], + "lines": [ + { + "bbox": [ + 146, + 517, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 146, + 517, + 479, + 550 + ], + "type": "text", + "content": "Table 2: Effect of " + }, + { + "bbox": [ + 146, + 517, + 479, + 550 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 146, + 517, + 479, + 550 + ], + "type": "text", + "content": " in Eqn. 17 for online proxy Table 3: Comparison of strate-learning on ImageNet with ResNet-50 as vision gies for obtaining the combination ratio " + }, + { + "bbox": [ + 146, + 517, + 479, + 550 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 146, + 517, + 479, + 550 + ], + "type": "text", + "content": " in Eqn. 13." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 552, + 343, + 578 + ], + "lines": [ + { + "bbox": [ + 152, + 552, + 343, + 578 + ], + "spans": [ + { + "bbox": [ + 152, + 552, + 343, + 578 + ], + "type": "table", + "html": "
β00.20.40.60.81
Acc(%)61.9662.1762.2362.3862.5061.84
", + "image_path": "701b101dbcc7bc34491b9e0d96148b04b84232baac29eb378c16885466044e20.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 361, + 552, + 466, + 578 + ], + "blocks": [ + { + "bbox": [ + 361, + 552, + 466, + 578 + ], + "lines": [ + { + "bbox": [ + 361, + 552, + 466, + 578 + ], + "spans": [ + { + "bbox": [ + 361, + 552, + 466, + 578 + ], + "type": "table", + "html": "
βFixedDynamic
Acc(%)62.2862.50
", + "image_path": "e38ab53cf04e4f1e14ebfb90241220cae30c6afd0bec667ae11fd17c1d965781.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", only the label from the online label learning (i.e., OnLab) is adopted for prediction. If mixing the label from the vision proxy, the accuracy improves. The mixture prediction outperforms OnLab by " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.54\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta = 0.8" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and is " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.18\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " better than the baseline CLIP. If excluding the label from text space at the end of data streaming as " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", the performance degenerates due to the large variance from the vision space. The phenomenon illustrates the challenge" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "of online learning that the vision proxy may not be well learned with a limited number of arrived images and the pseudo label from text proxy can help reduce the variance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "type": "text", + "content": "According to our analysis in Section 3.3, a dynamic ratio for the bias-variance trade-off is more effective for the proposed online learning scenario. To verify the claim, we compare the dynamic ratio to the fixed one in Tab. 3. The value of fixed " + }, + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 152, + 480, + 200 + ], + "type": "text", + "content": " is searched in [0, 1] by a step size of 0.1 and the best result is reported." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "text", + "content": "While the fixed " + }, + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "text", + "content": " still improves the performance from the online label learning, a monotonically increasing " + }, + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "text", + "content": " captures the training dynamics better and shows an additional gain of " + }, + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "inline_equation", + "content": "0.22\\%" + }, + { + "bbox": [ + 130, + 200, + 480, + 247 + ], + "type": "text", + "content": ". It confirms our analysis about the bias-variance trade-off." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 264, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 479, + 323 + ], + "type": "text", + "content": "Effect of iterations in optimization Our method has to learn the dual variables and vision proxy with streaming data and those parameters may not be learned well at the early stage of online learning. To investigate the influence of iterations in online learning, we report the accumulated accuracy of different iterations in one epoch on ImageNet in Tab. 4." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 149, + 388, + 255, + 435 + ], + "blocks": [ + { + "bbox": [ + 138, + 344, + 473, + 377 + ], + "lines": [ + { + "bbox": [ + 138, + 344, + 473, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 473, + 377 + ], + "type": "text", + "content": "Table 4: Comparison of accu- Table 5: Effect of epochs for optimization on Imracy " + }, + { + "bbox": [ + 138, + 344, + 473, + 377 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 138, + 344, + 473, + 377 + ], + "type": "text", + "content": " with different itera-ageNet. Prediction from the last epoch is evaluated tions on ImageNet. for OnZeta after running multiple epochs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 388, + 255, + 435 + ], + "lines": [ + { + "bbox": [ + 149, + 388, + 255, + 435 + ], + "spans": [ + { + "bbox": [ + 149, + 388, + 255, + 435 + ], + "type": "table", + "html": "
#Iters5k10k20k
Acc(%)60.9661.5061.71
30k40k50k
62.0262.4162.50
", + "image_path": "3fe4ecb98abf634aba7c16f1263c3bcf6a58e5d082f3e5543e642533f82a32d1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 267, + 388, + 473, + 412 + ], + "blocks": [ + { + "bbox": [ + 267, + 388, + 473, + 412 + ], + "lines": [ + { + "bbox": [ + 267, + 388, + 473, + 412 + ], + "spans": [ + { + "bbox": [ + 267, + 388, + 473, + 412 + ], + "type": "table", + "html": "
#Epoch1(online learning)2345
Acc(%)62.5063.2463.3563.3963.46
", + "image_path": "2af5c567c4d7cf7e1f7130741668f964de0fe42995e85bf1c74e4b34cc196ee7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 458, + 480, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 541 + ], + "type": "text", + "content": "First, we can observe that even with 5,000 iterations for learning, the accuracy over 1,000 classes is already better than the baseline. With the arrival of more images, the performance increases steadily. It is because that the convergence rate for dual variables and vision proxy is " + }, + { + "bbox": [ + 130, + 458, + 480, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1 / \\sqrt{n})" + }, + { + "bbox": [ + 130, + 458, + 480, + 541 + ], + "type": "text", + "content": " as analyzed in Theorems 1-2. It implies that these variables can reach a reasonable performance with a limited number of iterations, while it is challenging to obtain the optimal solution as discussed in the following experiment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "type": "text", + "content": "Effect of epochs in optimization In the proposed online zero-shot transfer problem, all examples can be visited only once when they arrive. The setting is prevalent in real-world applications but very challenging for optimization. This ablation experiment aims to investigate the performance gain when each example can be visited multiple times." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Concretely, we run the proposed method on multiple epochs of data, where images arrive in a random order within each epoch. For multiple epochs, the prediction from the last one is reported for evaluation. Tab. 5 compares the performance of multiple epochs to that of a single epoch in the online scenario." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "First, we can observe that with an additional epoch, the performance of OnZeta can be boosted by " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "0.74\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", which demonstrates the challenge of online zero-shot transfer. When increasing the number of epochs, the performance can be further improved and it achieves " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "63.46\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " with 5-epoch learning. The result approaches the performance of learning with the entire data set as in [16] that shows a " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "63.74\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " accuracy when visiting all data by 2,000 epochs. The competitive performance of OnZeta confirms the effectiveness of the proposed online learning algorithm. More experiments can be found in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 226, + 289, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 289, + 239 + ], + "type": "text", + "content": "4.2 Comparison on ImageNet" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 387 + ], + "type": "text", + "content": "After the ablation study, we compare our method OnZeta to state-of-the-art methods on ImageNet. Multiple vision encoders from CLIP are adopted for evaluation, which include ResNet-50, ViT-B/32, ViT-B/16, ViT-L/14, and ViT-L/14@336. Considering that images come in a random order in our online setting, the proposed method is repeated by 5 times and the averaged result is reported for the comparison. We note that the standard deviation is quite small and also report the best result among different runs by *. The same parameters are shared by different vision encoders in our method. Tab. 6 summarizes the comparison, where the performance of InMaP [16] is marked by gray due to the application of the whole unlabeled data set. All other methods only access a single image at each iteration. Our OnLab with only the online label learning is denoted as \"OnZeta0.5\" and the vanilla zero-shot transfer in Eqn. 1 is \"Baseline\"." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 471, + 481, + 536 + ], + "blocks": [ + { + "bbox": [ + 130, + 405, + 482, + 462 + ], + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 462 + ], + "type": "text", + "content": "Table 6: Comparison of accuracy (%) on ImageNet with different vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 471, + 481, + 536 + ], + "lines": [ + { + "bbox": [ + 135, + 471, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 135, + 471, + 481, + 536 + ], + "type": "table", + "html": "
Vision encoderBaseline [17]TPT [19]OnZeta0.5OnZeta*0.5OnZetaOnZeta*InMaP [16]
ResNet-5060.3260.7461.8961.9662.4262.5063.74
ViT-B/3263.77-65.0665.1865.7165.8067.29
ViT-B/1668.7568.9870.1870.2370.8770.9972.55
ViT-L/1475.96-76.9877.0777.8477.9479.29
ViT-L/1433677.02-77.9378.0578.7578.9480.21
", + "image_path": "5008fa4e580905e02ca6cdc268edd8488d7e340b3087d84185b23dcb38559326.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "First, by capturing the overall distribution using only the proposed online label learning component, " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{OnZeta}_{0.5}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " already gains about " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " for all vision encoders over the baseline. It demonstrates that the side information from the arrived images is useful for improving the zero-shot transfer. By learning the vision proxy and combining the predictions from both the text and vision spaces, OnZeta shows about " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " additional improvement over " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{OnZeta}_{0.5}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", which illustrates that online proxy learning is complementary to online label learning for zero-shot classification. Moreover, OnZeta obtains more gains with a large backbone, showing its potential for large models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "Then, we can observe that the prompt learning method TPT outperforms the baseline with an ensemble of artificial text prompts, which indicates the importance of the appropriate prompt for zero-shot transfer. However, OnZeta surpasses TPT by " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "1.76\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " on ResNet-50 and " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "2.01\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " on ViT-B/16. It is because that we learn a better class proxy in the vision space for visual classification, which helps reduce the modality gap and is more effective than optimizing the text prompt in the text space. The observation is consistent with the analysis in [16]. Moreover, compared with InMaP that can access the whole unlabeled set, OnZeta is only about " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " worse using different vision encoders. As analyzed in the ablation study, online learning that visits each example once without storing is more challenging than the offline method with multiple iterations over the entire set. Nevertheless, online learning is practical for real-world applications. Finally, all vision encoders share the same parameters in experiments. It implies that our method OnZeta is not sensitive to the selection of vision encoder and is applicable in different configurations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 312, + 384, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 384, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 384, + 323 + ], + "type": "text", + "content": "4.3 Comparison on Other 13 Downstream Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "content": "Besides the comparison on ImageNet, we conduct experiments on other 13 downstream tasks to evaluate the performance of our method. The benchmark data sets for zero-shot transfer are adopted in the comparison, which includes Aircraft [12], Caltech101 [5], Stanford Cars [10], CIFAR-10 [11], CIFAR-100 [11], CUB200-2011 [21], Describable Textures Dataset (DTD) [3], EuroSAT [9], Flowers [13], Food101 [1], Oxford-IIIT Pet (Pets) [14], Sun397 [23], and UCF101 [20]. A broad range of downstream tasks are involved by these diverse data sets, e.g., low-resolution image classification, fine-grained visual categorization (FGVC), land cover classification with satellite images, scene categorization, texture recognition, etc. These tasks contain fewer images than ImageNet, which increases the challenge of our online transfer with the convergence rate of " + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1 / \\sqrt{n})" + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "content": ". Therefore, we reduce " + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "content": " from 0.8 to 0.4 for all tasks to exploit the label from the text space. For data sets that the baseline already achieves satisfied performance, e.g., Caltech101 and CIFAR-10 with ViT, online label learning can be skipped by letting " + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\alpha = 0.4" + }, + { + "bbox": [ + 130, + 331, + 482, + 545 + ], + "type": "text", + "content": " can provide a slight improvement. Other parameters on the rest data sets remain the same as those for ImageNet. Two distinct vision encoders, i.e., ResNet-50 and ViT-B/16, are applied for evaluation. The performance averaged over 5 random trials is reported for our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "The comparison is summarized in Tab. 7. A similar phenomenon as ImageNet can be observed on these downstream tasks. Concretely, online label learning, i.e., " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{OnZeta}_{0.5}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", already improves the performance over baselines, where the gain is " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.12\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " on ResNet and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.58\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " on ViT, respectively. Compared with the baseline, only streaming images are leveraged by our method, and the representation of each arrived image is not stored, which preserves the flexibility of zero-shot transfer but captures the distribution over the entire data set in an online manner. In addition, learning the vision proxy with the pseudo label from the text space further improves the performance with different vision encoders, which demonstrates the effectiveness of learning the class proxy directly in the target" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 190, + 482, + 328 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 7: Comparison of accuracy (\\%) on 13 downstream tasks with ResNet-50 and ViT-B/16 as vision encoders in CLIP. The overall best performance is in bold. “-” denotes that the result is unavailable in their original papers. The averaged performance over 5 random trials is reported for our method, where the best one is denoted by *. InMaP is marked by gray since it leverages the whole unlabeled image set at each iteration." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 190, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 134, + 190, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 134, + 190, + 482, + 328 + ], + "type": "table", + "html": "
AircraftCaltechCarsCifar10Cifar100CUBDTDEuroSATFlowersFoodPetsSUNUCF101Avg.
ResNet-50:
Baseline16.6286.0056.3173.1540.6041.3741.1326.9062.9774.1081.8559.2555.5655.06
TPT17.5887.0258.46---40.8428.3362.6974.8884.4961.4660.82-
OnZeta0.518.2986.1758.5775.6246.9645.6843.2433.7962.4276.4183.0561.6460.1257.84
OnZeta*0.518.5186.4558.9175.7547.2245.9343.7434.3263.1776.5383.5961.8160.3858.18
OnZeta18.3186.2659.4075.8447.5546.0643.3833.9963.1676.8383.7762.3261.0058.30
OnZeta*18.5486.4159.7175.9547.7146.3443.9134.3563.9576.9784.6662.4061.3858.64
InMaP18.9686.7363.3078.8449.2649.1744.8635.8866.6878.3687.9363.8263.3660.55
ViT-B/16:
Baseline23.1393.5166.2991.0967.2949.3645.0950.1766.9984.4387.0065.6865.2165.79
TPT24.7894.1666.87---47.7542.4468.9884.6787.7965.5068.04-
OnZeta0.527.1993.6567.9991.1369.8552.8147.5156.2068.4985.8388.4468.0668.5068.13
OnZeta*0.527.4893.7968.1191.1870.0253.0947.8756.4269.0685.9788.7268.2768.8668.37
OnZeta27.5793.6968.7491.3170.3953.6648.0456.5269.3786.2189.0168.8769.5868.69
OnZeta*28.2993.8369.0391.3570.5253.8048.5856.7469.6386.3589.3269.0169.9468.95
InMaP28.3593.5973.0093.2772.5157.8750.7763.9871.2887.7692.9470.8572.0671.40
", + "image_path": "178ea5c7b36eed8f75d9bc3ad55c6f354f9fa74787af6853dc9f7769170559a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 506 + ], + "type": "text", + "content": "vision space. Moreover, OnZeta outperforms TPT on 9 out of 10 (only 10 tasks are available in TPT's original paper) data sets with different vision encoders. Compared with TPT that has to optimize the text prompt with multiple augmentations of the image by back-propagation through the text encoder, OnZeta only contains the forward pass to extract the representation of the image, which is more efficient and is capable of online deployment. By comparing the best performance of the proposed method (i.e., indicated by * in the table) to the averaged one, we can find that the gap averaged over 13 data sets is only about " + }, + { + "bbox": [ + 130, + 350, + 482, + 506 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 350, + 482, + 506 + ], + "type": "text", + "content": ". While the images arrive in a random order, the performance of the proposed method can be theoretically guaranteed as indicated in Theorems 1-2, which is robust to the order of input images. Finally, most data sets share the same parameters and it implies that OnZeta is not sensitive to hyper-parameters and is applicable for different tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 521, + 220, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 220, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 220, + 534 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "While CLIP demonstrates impressive zero-shot transfer performance, the information from the target data has not been explored sufficiently. In this work, we investigate an online zero-shot scenario, where a random image arrives at each iteration and the model has to predict its label without storing its representation. By capturing the overall label distribution and refining the class proxy in an online manner, our method OnZeta demonstrates a substantial improvement over the vanilla zero-shot transfer on diverse downstream tasks with different vision encoders. In addition, the operations in the proposed method can be implemented efficiently for real-time applications. Exploring more effective algorithms for challenging online zero-shot transfer can be our future work." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 244, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Bossard, L., Guillaumeun, M., Gool, L.V.: Food-101 - mining discriminative components with random forests. In: ECCV. Lecture Notes in Computer Science, vol. 8694, pp. 446-461. Springer (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 197 + ], + "type": "text", + "content": "2. Boyd, S.P., Vandenberghe, L.: Convex Optimization. Cambridge University Press (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "text", + "content": "3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: CVPR. pp. 3606-3613 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 221, + 480, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 221, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 221, + 480, + 263 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 480, + 297 + ], + "type": "text", + "content": "5. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In: CVPR workshop. pp. 178-178. IEEE (2004)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "text", + "content": "6. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. CoRR abs/2110.04544 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 352 + ], + "type": "text", + "content": "7. Hazan, E.: Introduction to online convex optimization. CoRR abs/1909.05207 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "text", + "content": "8. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 408 + ], + "type": "text", + "content": "9. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens. 12(7), 2217-2226 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "text", + "content": "0. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: ICCV workshop. pp. 554-561 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "text", + "content": "1. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "text", + "content": "2. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "text", + "content": "3. Nilsback, M., Zisserman, A.: Automated flower classification over a large number of classes. In: ICVGIP. pp. 722-729. IEEE Computer Society (2008)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 499, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 520 + ], + "type": "text", + "content": "4. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498-3505. IEEE (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "text", + "content": "5. Qian, Q., Shang, L., Sun, B., Hu, J., Li, H., Jin, R.: Softtriple loss: Deep metric learning without triplet sampling. In: ICCV. pp. 6449-6457. IEEE (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "type": "text", + "content": "6. Qian, Q., Xu, Y., Hu, J.: Intra-modal proxy learning for zero-shot visual categorization with clip. In: NeurIPS (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "type": "text", + "content": "7. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "type": "text", + "content": "8. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.S., Berg, A.C., Fei-Fei, L.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "9. Shu, M., Nie, W., Huang, D., Yu, Z., Goldstein, T., Anandkumar, A., Xiao, C.: Test-time prompt tuning for zero-shot generalization in vision-language models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 273, + 91, + 447, + 101 + ], + "type": "text", + "content": "Online Zero-Shot Classification with CLIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 304 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "20. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR abs/1212.0402 (2012)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "21. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 193 + ], + "type": "text", + "content": "22. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero-shot models. In: CVPR. pp. 7949-7961. IEEE (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "23. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: SUN database: Large-scale scene recognition from abbey to zoo. In: CVPR. pp. 3485-3492. IEEE Computer Society (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "24. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of CLIP for few-shot classification. In: ECCV. vol. 13695, pp. 493-510. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "25. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16795-16804. IEEE (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "text", + "content": "26. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vis. 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 102 + ], + "type": "text", + "content": "Q. Qian and J. Hu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_content_list.json b/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5624949a7fc1badaa5315270eed9e496bc47a0f4 --- /dev/null +++ b/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_content_list.json @@ -0,0 +1,1765 @@ +[ + { + "type": "text", + "text": "Open Panoramic Segmentation", + "text_level": 1, + "bbox": [ + 323, + 141, + 679, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junwei Zheng $^{1}$ , Ruiping Liu $^{1}$ , Yufan Chen $^{1}$ , Kunyu Peng $^{1}$ , Chengzhi Wu $^{1}$ , Kailun Yang $^{2}$ , Jiaming Zhang $^{1,\\dagger}$ , and Rainer Stiefelhagen $^{1}$", + "bbox": [ + 321, + 188, + 679, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Karlsruhe Institute of Technology \n $^{2}$ Hunan University", + "bbox": [ + 380, + 246, + 620, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Panoramic images, capturing a $360^{\\circ}$ field of view (FoV), encompass omnidirectional spatial information crucial for scene understanding. However, it is not only costly to obtain training-sufficient dense-annotated panoramas but also application-restricted when training models in a close-vocabulary setting. To tackle this problem, in this work, we define a new task termed Open Panoramic Segmentation (OPS), where models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain, enabling the zero-shot open panoramic semantic segmentation ability of models. Moreover, we propose a model named OOOPS with a Deformable Adapter Network (DAN), which significantly improves zero-shot panoramic semantic segmentation performance. To further enhance the distortion-aware modeling ability from the pinhole source domain, we propose a novel data augmentation method called Random Equirectangular Projection (RERP) which is specifically designed to address object deformations in advance. Surpassing other state-of-the-art open-vocabulary semantic segmentation approaches, a remarkable performance boost on three panoramic datasets, WildPASS, Stanford2D3D, and Matterport3D, proves the effectiveness of our proposed OOOPS model with RERP on the OPS task, especially $+2.2\\%$ on outdoor WildPASS and $+2.4\\%$ mIoU on indoor Stanford2D3D. The source code is publicly available at OPS.", + "bbox": [ + 261, + 306, + 743, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 636, + 375, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panoramic imaging systems [26, 27] have advanced significantly in recent years, which has fostered a wide variety of panoramic vision applications [1, 17]. Due to the comprehensive $360^{\\circ}$ Field of View (FoV), omnidirectional panoramas provide more informative visual cues when perceiving surroundings in a broad spectrum of scene understanding tasks [3, 36, 51, 52, 54, 79], enabling a more complete and immersive capture of environmental data, which is crucial for in-depth scene understanding. This wide perspective surpasses the limited scope of pinhole images, significantly enhancing the capability of computer vision systems to perceive and interpret surroundings in a variety of applications. While the benefits of utilizing panoramic images in computer vision applications are apparent compared", + "bbox": [ + 212, + 665, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Correspondence: jiaming.zhang@kit.edu", + "bbox": [ + 230, + 824, + 511, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9944832e53237798be69d5ca51307d5e129be6a789df41bbc1f109e8b19903ae.jpg", + "image_caption": [ + "(a) The challenge of wider FoV. Performance (mIoU / open mIoU) drops when FoV increases." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 498, + 214 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5ee89e04e53003570a9fdc65f0a3771b2b85b1be7edfbcc37974a025792c7b6.jpg", + "image_caption": [ + "(b) Close-vocabulary vs. Open-vocabulary segmentation on panoramas. While limited categories are recognized in a close-vocabulary setting, all objects in $360^{\\circ}$ can be covered in an open-vocabulary setting." + ], + "image_footnote": [], + "bbox": [ + 218, + 237, + 500, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1dbbd406fe1a711b5b77fb9494ea9d1c15795e07be5c8e8e54e96fef582feb1f.jpg", + "image_caption": [ + "(c) The paradigm of Open Panoramic Segmentation (OPS) task.", + "Fig. 1: (a) The challenge of existing state-of-the-art segmentation models. (b) The limitation of categories in traditional close-vocabulary panoramic segmentation tasks. (c) Our newly defined Open Panoramic Segmentation (OPS) task aims at tackling the above challenges. OPS consists of three important elements: Open the FoV targeted at the challenge of $360^{\\circ}$ FoV, Open the Vocabulary targeted at the drawback of close-vocabulary panoramic segmentation and Open the Domain targeted at the challenge of scarcity of panoramic labels." + ], + "image_footnote": [], + "bbox": [ + 504, + 148, + 782, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with pinhole images, it is imperative to give increasing consideration to some noteworthy challenges as follows: (1) The challenge of wider FoV. For example, Fig. 1a illustrates the degraded performance of state-of-the-art open-vocabulary semantic segmentation approaches [9,68,80] with an increasing FoV, i.e., from pinhole images to $360^{\\circ}$ panoramic images. Over $12\\%$ mIoU performance degradation can be observed, indicating the challenges introduced by the large divergence of semantic and structural information across narrow-wide imagery. (2) The limitation of categories. The traditional close-vocabulary segmentation task paradigm [74,82] only provides a limited amount of labeled categories, which cannot handle the inestimable amount of categories in real-world applications. Fig. 1b illustrates the difference between close- and open-vocabulary segmentation. Compared to the close-vocabulary setting (first row) the open-vocabulary setting (second row) is not limited by the number of categories of the datasets. Only four predefined categories (highlighted in different colors) are recognized in the close-vocabulary setting in contrast to the open one, where every pixel of the panorama has its own semantic meaning even though the categories are not annotated in the dataset.", + "bbox": [ + 212, + 507, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further release the great potential of panoramic images three critical problems are eager to be solved: How to obtain a holistic perception via a single image? How to break the barrier of limited recognizable categories in existing panoramic datasets so that downstream vision applications can benefit from unrestricted informative visual hints? How to deal with the scarcity of", + "bbox": [ + 212, + 763, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "panoramic labels? Based on the aforementioned three questions, we propose a new task named Open Panoramic Segmentation (OPS) that comprehensively addresses these challenges, aiming to better leverage the advantages brought by panoramic images. The new task paradigm is shown in Fig. 1c. The open panoramic segmentation task consists of three important elements considering the three problems. The concept of \"Open\" is three-fold: omnidirectional panoramic images (Open the FoV), an unrestricted range of recognizable classes (Open the Vocabulary), and additionally, models are trained using pinhole images in the source domain while evaluated with panoramic images in the target domain (Open the Domain). Given that densely annotated pinhole segmentation labels incur lower costs compared to panoramic ones [74], it is cost-efficient to open different domains. Note that OPS is different from Domain Adaptation (DA). In the context of training in DA, the utilization of data encompasses both the source and target domains while in OPS, exclusive reliance is placed on data originating solely from the source domain during the whole training process.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Apart from the new task, we propose a new model named OOOPS to address the aforementioned three openness-related challenges in the OPS task. It consists of a frozen CLIP model [49] and a key component named Deformable Adapter Network (DAN), which serves two critical purposes: (1) efficiently adapting the frozen CLIP model to the panoramic segmentation task and (2) addressing object deformation and image distortion inherent in panoramas. More specifically, a novel Deformable Adapter Operator (DAO), the key component of DAN, is designed to tackle panoramic distortion. To advance the modeling capabilities for distortion awareness in the pinhole source domain, we further introduce Random Equirectangular Projection (RERP) explicitly crafted to tackle object deformations and image distortions. A pinhole image is divided into four image patches that are randomly shuffled. Equirectangular Projection, one of the most common methods for mapping a globe into a panoramic plane, is then introduced in the shuffled image. Our OOOPS model with RERP outperforms other state-of-the-art open-vocabulary segmentation methods with $+2.2\\%$ , $+2.4\\%$ , and $+0.6\\%$ mIoU on WildPASS [76], Stanford2D3D [2], and Matterport3D [6], respectively.", + "bbox": [ + 212, + 397, + 787, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, we present the following contributions:", + "bbox": [ + 238, + 648, + 630, + 664 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new task termed open panoramic segmentation, OPS for short, including Open FoV, Open Vocabulary and Open Domain. Models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain.", + "- We put forward a model called OOOPS with the aim to address three openness-related challenges at once. A Deformable Adapter Network (DAN) is proposed to transfer the zero-shot learning ability of a frozen CLIP model from the pinhole domain to different panoramic domains.", + "- A novel data augmentation strategy named Random Equirectangular Projection (RERP), which is specifically designed for the proposed OPS" + ], + "bbox": [ + 225, + 672, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task, further boosts the accuracy of the OOOPS model, achieving state-of-the-art performance in the open panoramic segmentation task.", + "bbox": [ + 241, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- For benchmarking OPS, we conduct comprehensive evaluations on both indoor and outdoor datasets (WildPASS, Stanford2D3D and Matterport3D), involving more than 10 close- and open-vocabulary segmentation models.", + "bbox": [ + 225, + 176, + 785, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 243, + 387, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Open-vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 273, + 599, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the success of vision-language pre-training models, Open-Vocabulary Semantic Segmentation (OVSS) has been recently explored to achieve open-set pixel-wise understanding [7, 11, 12, 18, 19, 29, 31, 32, 34, 34, 37, 39, 40, 44, 53, 56, 58, 59, 61, 64, 70, 71]. Most works leverage the open-vocabulary capabilities of CLIP [49] to reason about relations between dense pixel semantics and arbitrary class labels [9, 12, 32, 38, 61, 71]. In [78, 89], multiple datasets are unified by transferring heterogeneous labels into text embeddings for training to achieve OVSS. In [35, 67], image-caption datasets are employed for learning alignment between visual regions and text entities. To reduce the computational costs of using a frozen pre-trained CLIP model for identifying novel classes, latest studies [21, 69] have also proposed efficient architectures for fast OVSS. SAN [68, 69] has designed an asymmetric input resolution strategy to decouple mask proposal prediction and attention bias prediction applied to CLIP for mask class recognition. Furthermore, open-vocabulary panoptic segmentation has been tackled in [8, 12, 13, 34, 45, 66, 66, 71, 80] to render universal image understanding. In this work, we focus on the open-vocabulary zero-shot learning on panoramas instead of pinhole images discussed in the aforementioned works.", + "bbox": [ + 212, + 297, + 792, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Panoramic Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 575, + 547, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Panoramic semantic segmentation enables more informative decision-making and interaction within complex and dynamic environments, advancing the capabilities of intelligent systems for navigation and perception [4, 16, 24, 25, 28, 33, 42, 46, 75]. PASS [72, 73] achieves panoramic segmentation on annular images, whereas 360BEV [54] tackles panoramic semantic mapping. ECANets [76] address omnidirectional semantic segmentation with multi-source omni-supervised learning to enhance generalization. Zhang et al. [81] propose to use attention-augmented designs for panoramic representation learning. Guttikonda et al. [20] propose a single-frame panoramic semantic segmentation approach using multimodal spherical images. Hu et al. [22] propose a distortion convolution module for panoramic semantic segmentation. An important cluster of works revisits panoramic semantic segmentation from an unsupervised domain adaptation perspective [23,30,41,81,82,86,87], by adapting from label-rich pinhole images to the panoramic images. Unlike these works, we propose a method to tackle open-vocabulary panoramic semantic segmentation for unconstrained surrounding understanding.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Deformable Neural Networks", + "text_level": 1, + "bbox": [ + 215, + 146, + 501, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the realm of deformable neural networks, notable advancements have been made through various iterations. The pioneering work [10] introduces the concept of deformable convolution, laying the foundation for subsequent developments. Building upon this foundation, DCNv2 [90] further refines deformable convolution by introducing a modulation mechanism that expands the scope of deformation modeling. The evolution of deformable neural networks continues with the introduction of DCNv3 [57], which combines the group convolution with DCNv2 for the sake of imitating multi-head mechanism of Transformer [55]. DCNv4 [65] aims to speedup DCNv3, achieving a more efficient model. Additionally, Trans4PASS [82,83] proposes Deformable Patch Embedding (DPE) and Deformable MLP (DMLP) to handle the distortion that occurs in panoramas. Incorporating the principles of deformable convolution into vision transformers [14], Xia et al. propose DeformableViT [62,63], extending the applicability of deformable neural networks to the domain of transformer architectures. After rethinking a series of DCN designs, we propose DAO to improve the model's ability to handle image distortion in the OPS task.", + "bbox": [ + 212, + 170, + 787, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 435, + 380, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Open Panoramic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 465, + 517, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Open Panoramic Segmentation (OPS) task is proposed to deal with three challenging problems: ① the narrow FoV, ② the restricted range of classes, and ③ the scarcity of panoramic labels. OPS gives three answers to the aforementioned issues: ① Open the FoV, ② Open the Vocabulary, and ③ Open the Domain. The OPS task paradigm is shown in Fig. 1c. Models are trained in the narrow-FoV pinhole source domain in an open-vocabulary setting while evaluated in the wide-FoV panoramic target domain.", + "bbox": [ + 212, + 491, + 787, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Model Architecture", + "text_level": 1, + "bbox": [ + 215, + 618, + 426, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A foundation model [49] can be efficiently transferred to downstream tasks by using an adapter [68]. To improve the modeling ability for panoramas, we design the OOOPS model. As shown in Fig. 2, it comprises a frozen CLIP model and a proposed Deformable Adapter Network (DAN) that incorporates multiple Transformer Layers and the novel DAO, which will be detailed in Sec. 3.3. Feature fusion takes place between the intermediate layers of CLIP and DAN. One of the two outputs from DAN consists of mask proposals, while the other serves as deep supervised guidance for CLIP to generate proposal logits. In the training phase, pinhole images are forwarded into OOOPS, generating mask proposals and proposal logits for the loss calculation. In the inference phase, panoramas are forwarded into OOOPS to generate segmentation predictions via the multiplication of mask proposals and the corresponding proposal logits. The frozen CLIP is necessary for the zero-shot learning ability of OOOPS.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/96240dd66b057b5eea2695ecf8ab59c7653cff69998751351a5c6e81494ba866.jpg", + "image_caption": [ + "Fig. 2: Overview of the OOOPS model architecture. It consists of a frozen CLIP model and a Deformable Adapter Network (DAN) which includes Transformer Layers and the proposed DAO." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Deformable Adapter Network", + "text_level": 1, + "bbox": [ + 215, + 441, + 508, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Deformable Adapter Network is a combination of multiple Transformer [55] Layers and the proposed DAO. Since distortion exists in panoramic images, which is a big challenge when utilizing informative panoramas, we delve deeper into the deformable design [57] and sampling methods like APES [60], MateRobot [85], proposing DAO to tackle the image distortions and object deformations of panoramas. To provide a detailed rethinking process of the deformable convolution design, we introduce the preliminary of the DCN series as follows.", + "bbox": [ + 212, + 464, + 787, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Revisiting DCN Series. The seminal work DCN [10] can enable the traditional CNN with spatial deformation-aware ability. Given a convolutional kernel of $K$ sampling locations, let $\\mathbf{w}_k$ and $\\mathbf{p}_k$ denote the weight and pre-specified offset for the $k$ -th location, respectively. For example, $K = 9$ and $\\mathbf{p}_k \\in \\{(1,1), \\dots, (-1, -1)\\}$ defines a $3 \\times 3$ convolutional kernel of dilation 1. Let $\\mathbf{x}(\\mathbf{p})$ and $\\mathbf{y}(\\mathbf{p})$ denote the features at location $\\mathbf{p}$ from the input feature maps $\\mathbf{x}$ and output feature maps $\\mathbf{y}$ , respectively. DCN is formulated as:", + "bbox": [ + 212, + 569, + 787, + 676 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} (\\mathbf {p}) = \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {k} \\mathbf {x} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {k}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 684, + 785, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Delta \\mathbf{p}_k$ is the learnable offset for the $k$ -th location. Although DCN is capable of capturing spatial deformation, every sampling location is treated equally when calculating the local features. DCNv2 [90] is proposed with an additional term called modulation scalar. Specifically, DCNv2 can be formulated as:", + "bbox": [ + 212, + 732, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} (\\mathbf {p}) = \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {k} \\mathbf {m} _ {k} \\mathbf {x} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {k}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 801, + 785, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{m}_k$ is the learnable modulation scalar for the $k$ -th location. Inspired by Transformer, DCNv3 [57] is proposed with a grouping operation, further boosting the deformation-aware ability of DCNv2. DCNv3 can be addressed with the following formula:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} (\\mathbf {p}) = \\sum_ {g = 1} ^ {G} \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {g} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {g k}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 217, + 785, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $G$ denotes the total number of aggregation groups. DCNv4 [65] is similar to DCNv3, achieving similar performance while significantly reducing the runtime.", + "bbox": [ + 214, + 270, + 787, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f06c2a86499b210dcd1eacf22acc69164e03fbc1dd56f82f5f5913404a1e637b.jpg", + "image_caption": [ + "Fig. 3: Salient map generation in DAO." + ], + "image_footnote": [], + "bbox": [ + 357, + 330, + 647, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deformable Adapter Operator (DAO). When dealing with distortions in panoramas, DCNv3 and DCNv4 fall short of meeting the distortion-aware requirements. This is illustrated by the visualization of Fig. 6 in Sec. 4.5. Therefore, DAO is proposed to tackle the distortion problem in panoramic images using the following expression:", + "bbox": [ + 212, + 534, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} (\\mathbf {p}) = \\mathbf {s} (\\mathbf {p}) \\sum_ {g = 1} ^ {G} \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {g} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {g k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 619, + 785, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{s}(\\mathbf{p})$ is the learnable salient scalar at location $\\mathbf{p}$ . Inheriting from DCNv3, DAO comes up with an additional term called salient scalar indicating the importance of each pixel within the whole panorama. It's worth noting that DCNv3 and DCNv4 share the same mathematical expression, but DCNv3 is more robust according to our experiments before designing DAO. Therefore, we adopt DCNv3 as a part of DAO rather than DCNv4. Illustrated in Fig. 2, the feature map outputted by DCNv3 passes through the Patch Similarity Layer, Normalization Layer, and Standard Deviation Layer sequentially to form a salient map. The intuition for such a design is straightforward. The salient pixels in an image are those that differ significantly from their neighboring pixels, e.g., edge pixels. If all pixels within an image patch are different, the standard deviation of the", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8357fdae817e0482fcaef067b2002045730139dec0591a682b911a14d7cbaa62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 143, + 413, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a89652183d1d5894817bbf5376ab43d72310c9add4d9018828b1b24a11131f62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 205, + 405, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34f27cedfb3013359b313cdd70477ac0779ffcfc05bbdeb66082d67e9348efbb.jpg", + "image_caption": [ + "(a) Equirectangular Projection (ERP) maps a globe into a panorama. It's clear that strong distortion occurs in the panorama, especially on object edges in the horizontal direction." + ], + "image_footnote": [], + "bbox": [ + 238, + 232, + 491, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9df6f7374ef486d09cff92e1638f00aae16e2a5daab186c82eef4380a1420a81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 143, + 589, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32aeedbff43f2742c95dc8624148e698b2877a25b42a6214db3d2f7f1cfeff71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 143, + 674, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e378c1784ab7463c61bbff1685d866a257219ea04f85e8641d388a8382f2d5a6.jpg", + "image_caption": [ + "Fig. 4: (a) Visualization of ERP on a panoramic image and (b) Our proposed RERP on pinhole images." + ], + "image_footnote": [], + "bbox": [ + 509, + 208, + 589, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0eaf5c01de9eeed8815723df3ff32c6c0732a6171a46e45f37ba7a7b821e454f.jpg", + "image_caption": [ + "(b) Visualization of ERP and RERP of pinhole images. First column: pinhole images. Second column: ERP on pinhole images. Third column: RERP on pinhole images." + ], + "image_footnote": [], + "bbox": [ + 593, + 207, + 674, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b814a83bf59069de5a5b863c893dfb42b267022716d7216b0ba6491ee1b7c17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 143, + 761, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12937113d2c42facd50178b70956f57af3e93762ce4e218f71ec156d35af2ff5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 207, + 761, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pixel similarity for this patch is higher than the one containing similar pixels, resulting in a higher salient scalar. Fig. 3 gives a more detailed explanation of the salient map generation. Given a feature map, DAO first calculates the cosine similarity between the center pixel and all pixels within a kernel, e.g., 9 pixels within a $3 \\times 3$ kernel in Fig. 3, resulting in a 9-dimensional cosine similarity vector. Softmax normalization is then applied to the vector. Afterward, DAO calculates the standard deviation of this normalized cosine similarity vector, indicating the importance of the center pixel. By traversing every pixel across the entire feature map, a salient map is generated for the sake of enhancing the salient pixels, which are usually the edge pixels of an image where strong panoramic distortion occurs.", + "bbox": [ + 212, + 381, + 787, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Random Equirectangular Projection", + "text_level": 1, + "bbox": [ + 214, + 566, + 560, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Equirectangular Projection [50] (ERP) is one of the most common methods for mapping a globe into a panoramic plane, which transforms spherical coordinates into planar coordinates expressed as follows:", + "bbox": [ + 212, + 590, + 785, + 636 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nx = R \\left(\\lambda - \\lambda_ {0}\\right) \\cos \\left(\\varphi_ {1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 645, + 784, + 662 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ny = R \\left(\\varphi - \\varphi_ {0}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 664, + 784, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ and $\\varphi$ are the longitude and latitude of the location to project, respectively. $\\varphi_{1}$ are the standard parallels. $\\lambda_0$ and $\\varphi_0$ are the central meridian and central parallel of the map, respectively. $R$ is the radius of the globe. $x$ denotes the horizontal coordinate, while $y$ represents the vertical coordinate of the projected location on the map. Fig. 4a visualizes the ERP on a panoramic plane. It can be observed that after equirectangular projection strong distortion occurs in the panorama, e.g., a straight line is transformed into a curve. To further boost the performance, we propose Random Equirectangular Projection (RERP) on pinhole images since the OPS task requires the model to train with pinhole images instead of panoramas.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We divide the pinhole image into four parts and randomly shuffle the image patches. Afterward, we apply equirectangular projection to the distortion-free pinhole image. Fig. 4b visualizes pinhole images after RERP. The first column is the pinhole images without any data augmentation. The second column is ERP on pinhole images without random shuffling. The last column is our proposed RERP. It can be observed that the panorama-like distortion also occurs in the pinhole images after RERP. Random shuffling is employed to enhance robustness and promote generalization.", + "bbox": [ + 218, + 145, + 787, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 218, + 287, + 374, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 218, + 319, + 333, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train our model using the pinhole general-object COCOStuff-164k [5] while evaluating the model performance on three different panoramic datasets: panoramic street-view WildPASS [76], panoramic indoor-scene Stanford2D3D [2] and panoramic indoor-scene Matterport3D [6].", + "bbox": [ + 218, + 342, + 812, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "COCOStuff-164k. The pinhole dataset comprises 164k images across 171 annotated classes, partitioned into training, validation, and test sets, with 118k, 5k, and 41k images, respectively. In our experiments, we utilize the entire 118k-image training set as the training data by default.", + "bbox": [ + 218, + 402, + 785, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "WildPASS. The panoramic street-view dataset contains 2,500 panoramas in 65 cities from all continents Asia, Europe, Africa, Oceania, and North and South America (excluding only Antarctica) with 8 dense-annotated classes. All images have a $70^{\\circ} \\times 360^{\\circ}$ FoV in the size of $400 \\times 2048$ .", + "bbox": [ + 218, + 463, + 785, + 522 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stanford2D3D. The panoramic indoor dataset consists of 1,413 images in the size of $512 \\times 1024$ and 13 object classes. All results are averaged over 3 cross-validation folds.", + "bbox": [ + 218, + 523, + 785, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matterport3D. The panoramic indoor dataset includes 10,800 panoramic views from 194,400 RGB-D images. According to the data preprocessing pipeline of 360BEV [54], a subset of 772 images in the size of $512 \\times 1024$ with 20 classes is used to evaluate the segmentation performance.", + "bbox": [ + 218, + 568, + 785, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Metrics", + "text_level": 1, + "bbox": [ + 218, + 650, + 323, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Close-vocabulary Metric. Mean Intersection over Union (mIoU) is used as the evaluation metric in close-vocabulary semantic segmentation tasks. Specifically, IoU is calculated as the area of overlap between the predicted segmentation and the ground truth, divided by the area of union between them. mIoU is a hard metric designed for close-vocabulary segmentation ignoring the semantic meanings of similar classes referring to the open-vocabulary setting.", + "bbox": [ + 218, + 672, + 785, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Open-vocabulary Metric. Open mean Intersection over Union [88], open mIoU for short, is a soft metric compared to mIoU since the similarity scores between different classes are also taken into account, which are calculated by the WordNet approach [43]. In this work, we will report both close and open metrics for a comprehensive understanding of all models.", + "bbox": [ + 218, + 763, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/dda5503e70337909b4f5c8276767944cd43d39a150eaea0cf615464abf7c3313.jpg", + "table_caption": [ + "Table 1: Results on WildPASS. The mIoU, open mIoU, and per-class IoU are in percentage (%). #PARAMs is the number of learnable parameters in millions." + ], + "table_footnote": [], + "table_body": "
Network#PARAMSmIoUOpen mIoUCarRoadSidewalkCrosswalkCurbpersonTruckBus
Close-Vocabulary
Fast-SCNN [48]1.124.8-45.960.031.79.717.16.014.213.8
PSPNet18 [84]17.528.2-58.266.828.413.019.26.218.215.6
SwiftNet [47]11.830.0-55.764.129.216.222.98.521.122.2
DenseASPP [77]8.333.2-51.169.138.116.426.38.727.428.4
ERF-PSPNet [73]2.534.0-66.370.536.56.424.19.426.532.0
PSPNet50 [84]53.346.1-80.074.951.723.931.419.838.948.1
DANet [15]47.447.2-74.872.249.928.923.825.451.950.6
OOSS (ERF-PSPNet) [74]2.556.1-87.279.360.828.038.154.548.852.2
PASS (ERF-PSPNet) [73]2.564.7-87.380.061.471.149.972.237.557.9
ERF-PSPNet (omni-sup') [76]2.566.8-90.582.765.670.551.558.262.053.1
ECANet (with attention) [76]2.667.7-90.483.768.467.852.161.454.563.3
ECANet (with fusion) [76]2.669.0-90.685.768.067.952.166.059.362.3
Open-Vocabulary
SAN [68]8.455.656.984.381.969.610.12.872.254.969.1
CAT-seg [9]59.555.757.184.181.570.010.23.072.554.869.3
OpenSeeD [80]65.455.857.284.381.669.710.33.073.654.969.1
OOOPS (ours, w/o RERP)8.757.058.584.981.668.715.18.471.355.370.4
OOOPS (ours, w/ RERP)8.758.059.784.981.668.719.48.373.854.972.2
", + "bbox": [ + 228, + 188, + 782, + 414 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 454, + 455, + 469 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We utilize CLIP ViT-B/16 as the frozen CLIP model. All models are trained with the pinhole general-object COCOStuff-164k training set with $4 \\times \\mathrm{A}40$ GPUs while evaluated on three panoramic datasets which are in totally different domains compared with the pinhole dataset. The AdamW optimizer is applied with an initial learning rate of 0.0001, weight decay of 0.0001, batch size of 32, and a total of 60,000 training iterations. Throughout the training, the learning rate follows a polynomial schedule with a power of 0.9. Data augmentations, including random image resizing within the short-side range of [320, 1024] and a crop size of $640 \\times 640$ are applied. Additionally, when verifying the effectiveness of the proposed RERP, it is applied as an extra data augmentation. Following the practice of SAN [68], apart from the supervision of mask recognition with cross-entropy loss, we utilize the dice loss and binary cross-entropy loss for mask proposal generation.", + "bbox": [ + 212, + 477, + 787, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Comparison with the State of the Art", + "text_level": 1, + "bbox": [ + 215, + 695, + 576, + 710 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on WildPASS. Comparisons with both close- and open-vocabulary state-of-the-art methods on WildPASS are listed in Table 1. It is not surprising that ECANet (with fusion) [76] achieves the best performance among all methods since it is a close-vocabulary method, meaning that ECANet is only able to predict the classes predefined by the dataset. However, our OOOPS model is qualified for predicting classes that are not annotated in the dataset. The OOOPS method (with RERP) surpasses other state-of-the-art open-vocabulary panoramic segmentation methods with over $+2.2\\%$ mIoU while introducing", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d5d995f006c1007f0e6a53716257988c635e6044cfc30b8070d4c849eec7e24f.jpg", + "image_caption": [ + "Fig. 5: (a) Comparison on the WildPASS dataset and (b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings." + ], + "image_footnote": [ + "(a) Comparison with baseline method SAN [68] on the WildPASS dataset with predefined 8 classes. First row: ground truth. Second row: prediction from SAN. Third row: prediction from OOOPS with RERP." + ], + "bbox": [ + 217, + 143, + 496, + 233 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6d1a70ea15719e6c11daca8a1148376fe15be37ed7e72f6e7e884c63a3a9c1c0.jpg", + "image_caption": [], + "image_footnote": [ + "(b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings. First row: RGB. Second row: prediction with close vocabularies. Third row: prediction with open vocabularies. Zoom in for a better view." + ], + "bbox": [ + 504, + 143, + 784, + 233 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "only a minor increment of 0.3 million learnable parameters compared to the lightweight baseline SAN [68], offering a promising step towards bridging the gap between close- and open-vocabulary settings. Moreover, our OOOPS method enhanced by RERP demonstrates significant advancements in challenging categories such as crosswalk and curb, achieving $+9.3\\%$ and $+5.5\\%$ mIoU compared with the baseline SAN, respectively. It is worth noting that in some categories, e.g., sidewalk, person, and bus, our proposed OOOPS (both with and without RERP) even outperforms the models trained in a close-vocabulary setting. All open-vocabulary methods achieve better performance with open mIoU metric compared to the close mIoU since similar semantic meanings are also considered in the calculation. Fig. 1a showcases the performance degradation on both metrics. Even though all methods have degraded performance when FoV increases, our OOOPS achieves a smaller drop $(-10.0\\%)$ while others have a larger performance decrease $(-12.0\\%)$ , indicating the effectiveness of the OOOPS model.", + "bbox": [ + 212, + 367, + 787, + 579 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 5a visualizes the prediction of the challenging curb category from baseline SAN and our proposed OOOPS with RERP. From top to bottom are the ground truth, prediction from SAN, and prediction from OOOPS with RERP, respectively. The baseline SAN cannot recognize the challenging curb category while our proposed OOOPS with RERP is capable of detecting the curb in the middle of the panorama highlighted in gray. Fig. 5b showcases the visualization of the prediction from OOOPS on WildPASS in close- and open-vocabulary settings. From top to bottom are input RGB, prediction from OOOPS with 8 predefined close vocabularies from WildPASS, and prediction from OOOPS with unlimited open vocabularies. It is evident that our OOOPS with RERP is capable of tackling unseen categories in the open-vocabulary setting. Every pixel on the panorama is assigned a meaningful semantic label. For example, \"mailbox\" is not a predefined category from WildPASS. While existing models can not recognize the mailbox under a close-vocabulary setting, the rare classes can be easily detected by our model in an open-vocabulary one, indicating the great potential of open-vocabulary panoramic segmentation for parsing unconstrained traffic surroundings.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/97019023d5e34c99681f23af5ec363fd8e5f88aed1656823fc13cffb6634646c.jpg", + "table_caption": [ + "Table 2: Results on Stanford2D3D. The mIoU, open mIoU, and per-class IoU are in percentage (\\%). #PARAMs is the number of learnable parameters in millions." + ], + "table_footnote": [], + "table_body": "
Network#PARAMsmIoUOpen mIoUBeamBoardBookcaseCeilingChairClutterColumnDoorFloorSofaTableWallWindow
Close-Vocabulary
Trans4PASS [82]28.252.1-11.460.651.181.859.737.923.554.688.834.653.968.151.0
360BEV [54]28.454.3-06.063.655.582.762.739.916.458.890.845.257.171.157.6
Open-Vocabulary
SAN [68]8.438.539.50.251.539.666.552.110.70.045.049.331.849.250.053.4
CAT-seg [9]59.538.639.60.051.642.366.851.48.00.047.450.731.247.551.453.5
OpenSeeD [80]65.438.740.00.051.842.067.051.68.40.247.550.031.447.151.753.6
OOOPS (w/o RERP)8.739.541.00.052.344.167.153.79.70.050.152.830.547.252.253.8
OOOPS (w/ RERP)8.741.142.60.050.744.868.851.88.10.048.271.133.647.255.454.5
", + "bbox": [ + 223, + 191, + 782, + 348 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5ae218e7f7d60f156392cf66d9eb8c6eef53217a440c1e317c5460a2dc5afca5.jpg", + "table_caption": [ + "Table 3: Results on Matterport3D. The mIoU, open mIoU, and per-class IoU are in percentage (\\%). #PARAMs is the number of learnable parameters in millions." + ], + "table_footnote": [], + "table_body": "
Network≈PARAMsinotOpen mIoUWallFloorChairDoorTablePictureFurnitureObjectsWindowSofbBedSinkStarsCeilingToiletMirrorshowerBathtubCounterShelving
Close-Vocabulary
Trans4PASS+ [83]28.542.6-63.479.139.140.332.836.031.031.537.544.063.220.641.877.640.724.323.758.334.332.9
360BEV [54]28.446.4-64.183.145.845.038.041.132.335.140.648.769.825.147.880.246.028.722.360.138.634.8
Open-Vocabulary
SAN [68]8.430.531.454.476.118.034.122.321.113.01.732.734.251.011.019.974.223.412.031.631.919.028.4
CAT-seg [9]59.530.231.154.374.817.034.121.920.812.61.932.934.050.110.718.873.522.411.631.932.919.228.5
OpenSecD [80]65.430.631.654.276.318.134.322.521.013.02.132.834.150.911.220.274.123.512.231.131.819.128.6
OOOPS (w/o RERP)8.731.132.455.277.017.135.725.017.513.52.333.032.451.812.620.474.522.610.834.136.819.429.7
OOOPS (w/ RERP)8.731.232.555.476.418.434.823.921.213.22.633.334.552.912.120.374.628.212.531.929.319.428.6
", + "bbox": [ + 222, + 421, + 782, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on Stanford2D3D. In Table 2, the comparison is conducted between state-of-the-art methods on the Stanford2D3D dataset. All reported results are averaged over 3 cross-validation folds. Without RERP, OOOPS achieves $39.5\\%$ mIoU, outperforming other open-vocabulary methods like SAN [68], CAT-seg [9], and OpenSeeD [80], which score between $38.5\\%$ and $38.7\\%$ . With RERP, OOOPS's mIoU increases to $41.1\\%$ with a $+2.4\\%$ gain, narrowing the gap with advanced close-vocabulary methods.", + "bbox": [ + 212, + 582, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on Matterport3D. In Table 3, the comparison is conducted between state-of-the-art methods on the Matterport3D dataset. Our OOOPS method showcases consistent performance. Without augmentation, OOOPS achieves mIoU of $31.1\\%$ , slightly outperforming other open-vocabulary methods like SAN [68] $(30.5\\%)$ , CAT-seg [9] $(30.2\\%)$ , and OpenSeeD [80] $(30.6\\%)$ . With augmentation, OOOPS's mIoU marginally increases to $31.2\\%$ , further reinforcing its lead over these open-vocabulary methods. Both variants of OOOPS still lag behind close-vocabulary methods like Trans4PASS+ [83] and 360BEV [54], which however, are trained in a supervised fashion in a close-vocabulary setting on panoramas and have mIoUs ranging from $42.6\\%$ to $46.4\\%$ .", + "bbox": [ + 212, + 688, + 795, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3a9acf26bcb2c2345a11bd001df00da8e9eb83811bdd7991a5cbb61720317c48.jpg", + "table_caption": [ + "Table 4: Task difference between Domain Adaptation (DA) and the proposed OPS." + ], + "table_footnote": [], + "table_body": "
TaskDomain#ClassesmIoU (%)
SourceTargetWildPASSMatterport3DStanford2D3D
DA✓✓limited72.247.255.4
OPS✓✗unlimited58.031.241.1
", + "bbox": [ + 222, + 170, + 784, + 224 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(a) Ablation study of OOOPS components on the WildPASS Dataset. The mIoU is reported in percentage $(\\%)$ . #PARAMs stands for the number of learnable parameters in millions.", + "bbox": [ + 215, + 266, + 498, + 309 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b) Ablation study of different numbers of shuffling patches on WildPASS, Matterport3D and Stanford2D3D datasets. The mIoU is reported in percentage $(\\%)$ .", + "bbox": [ + 503, + 267, + 785, + 310 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b897021ddfb07d27b2e5315df1f6ce4de1ace7d6e344acb532b5bfa9f16120a2.jpg", + "table_caption": [ + "Table 5: Ablation study of OOOPS components and the number of shuffling patches." + ], + "table_footnote": [], + "table_body": "
MethodERPRERP#PARAMsmIoU
SAN [68]✘✘8.455.6
SAN [68]✓✘8.456.0
+ DCN [68]✘✓8.456.5 (+0.9)
+ DCN✘✘8.655.6
+ DCN✓✘8.656.0
+ DCN✘✓8.656.6 (+1.0)
+ DCNv2✘✘8.755.7
+ DCNv2✓✘8.756.4
+ DCNv2✘✓8.756.6 (+0.9)
+ DCNv3✘✘8.755.8
+ DCNv3✓✘8.756.4
+ DCNv3✘✓8.756.8 (+1.0)
+ DCNv4✘✘8.755.8
+ DCNv4✓✘8.756.4
+ DCNv4✘✓8.756.8 (+1.0)
+ DAO (ours)✘✘8.757.0
+ DAO (ours)✓✘8.757.6
+ DAO (ours)✘✓8.758.0 (+1.0)
", + "bbox": [ + 232, + 320, + 485, + 436 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/88cf2b1e700bc10e5bbc325f1b50cfce194aa3c604195de4b692a84b5df0565f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mIoU (%) #Patches Dataset4 9 16 25
WildPASS58.0 58.5 58.2 58.0
Matterport3D31.2 31.6 31.4 31.0
Stanford2D3D41.1 41.9 41.5 41.1
", + "bbox": [ + 509, + 321, + 784, + 436 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results in DA Subtask. The proposed OPS task differs from Domain Adaptation since no data from the target domain is available to models in OPS while models have access to the data in DA during training. Moreover, models are required to recognize an unlimited amount of classes in OPS while only a limited number of classes can be recognized in DA. DA is a special case of OPS, meaning that models in OPS can be easily adapted to DA and achieve better performance. The converse doesn't hold. Table 4 showcases the task difference between OPS and DA. It can be observed that over $+14\\%$ gain in mIoU is achieved when adapting our OOOPS model in the DA subtask.", + "bbox": [ + 212, + 462, + 787, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 619, + 390, + 635 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We replace DAO in OOOPS, as illustrated in Fig. 2, with DCN [10], DCNv2 [90], DCNv3 [57] and DCNv4 [65]. When referring to the component substitution without data augmentation in Table 5a, it is clear that our DAO surpasses other counterparts with a maximum performance gain of $+1.4\\%$ mIoU. As for the component substitution with ERP, all models achieve a boost in the range of $+0.3\\%$ to $+0.7\\%$ compared with the one without data augmentation. The gain is even bigger considering the component substitution with RERP, achieving $+1.0\\%$ mIoU. It is worth noting that RERP is able to boost the model performance stably among all models. The best model performance is achieved by the combination of DAO and RERP, yielding $58.0\\%$ mIoU with only a slight increment in parameters. Table 5b showcases the influence of different numbers of shuffling patches. The performance is not proportional to the number of patches, indicating the importance of moderate content continuity during training.", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7d5096a80cb57a68af065098d6f774bc40acd6312e3c2f1dd0f08756f7b19cbb.jpg", + "image_caption": [ + "(a) Deformable offsets without RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 498, + 287 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/03505442d6f0e8b1e5367307847b1b80fcb74392fefd5aaf81cd3c8d0c69a639.jpg", + "image_caption": [ + "(b) Deformable offsets with RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO.", + "Fig. 6: Visualization of deformable offsets. The green points $\\bullet$ are sampling locations. The red points $\\bullet$ are deformable offsets in 2 levels, indicating a deformable receptive field (e.g., each level has a $3 \\times 3$ kernel size, resulting in $(3 \\times 3)^{2} = 81$ red points)." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 785, + 287 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To further illustrate the deformable ability of DAO and the superiority of RERP when tackling the distortion of panoramas, we visualize the deformable offset in Fig. 6. Fig. 6a presents the deformable offsets without RERP. It can be observed that DAO is better aware of the panoramic distortion compared with the rest in the DCN series. Since the strongest distortion occurs on the edge of objects, DAO focuses more on edge pixels for better distortion-aware feature extraction. The standard deviation of patch similarity in DAO implies the salience of the image patch center. By multiplicatively combining salient scalars and features, significant pixels, such as those representing object edges, are enhanced in their representation. In Fig. 6b, it is evident that the inclusion of RERP improves the distortion information capturing ability of all models in the DCN series, as opposed to their counterparts without RERP. This observation is consistent with the results in Table 5a. The incorporation of both DAO and RERP in OOOPS showcases a remarkable distortion-aware modeling capability.", + "bbox": [ + 212, + 411, + 789, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 674, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we formulate the novel task of Open Panoramic Segmentation (OPS) for unfolding holistic scene understanding, which reaches Open FoV, Open Vocabulary, and Open Domain in surrounding scene segmentation. We have designed the Deformable Adapter Network (DAN) and Random Equirectangular Projection (RERP) for enhancing distortion-aware segmentation capacities of the novel proposed OOOPS model. Extensive experiments on WildPASS, Stanford2D3D, and Matterport3D benchmarks show that the proposed method significantly elevates the performance of state-of-the-art panoramic segmenters.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the Ministry of Science, Research and the Arts of Baden-Württemberg (MWK) through the Cooperative Graduate School Accessibility through AI-based Assistive Technology (KATE) under Grant BW6-03, in part by BMBF through a fellowship within the IFI programme of DAAD, in part by the Helmholtz Association Initiative and Networking Fund on the HAICORE@KIT and HOREKA@KIT partition, in part by the National Key RD Program under Grant 2022YFB4701400, and in part by Hangzhou SurImage Technology Company Ltd.", + "bbox": [ + 212, + 174, + 787, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 316, + 321, + 333 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022)", + "2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)", + "3. Athwale, A., Afrasiyabi, A., Lagüie, J., Shili, I., Ahmad, O., Lalonde, J.F.: DarSwin: Distortion aware radial swim transformer. In: ICCV (2023)", + "4. Berenguel-Baeta, B., Bermudez-Cameo, J., Guerrero, J.J.: FreDSNet: Joint monocular depth and semantic segmentation with fast fourier convolutions from single panoramas. In: ICRA (2023)", + "5. Caesar, H., Uijlings, J.R.R., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. CVPR (2018)", + "6. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. 3DV (2017)", + "7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023)", + "8. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. arXiv preprint arXiv:2303.11324 (2023)", + "9. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.: CAT-Seg: Cost aggregation for open-vocabulary semantic segmentation. In: CVPR (2023)", + "0. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: ICCV (2017)", + "1. Dao, S.D., Shi, H., Phung, D., Cai, J.: Class enhancement losses with pseudo labels for open-vocabulary semantic segmentation. TMM (2023)", + "2. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023)", + "3. Dong, B., Gu, J., Han, J., Xu, H., Zuo, W.: Towards universal vision-language omni-supervised segmentation. arXiv preprint arXiv:2303.06547 (2023)", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "5. Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., Lu, H.: Dual attention network for scene segmentation. In: CVPR (2019)" + ], + "bbox": [ + 225, + 347, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhou, X., Geiger, A., Liao, Y.: PanopticNeRF-360: Panoramic 3D-to-2D label transfer in urban scenes. arXiv preprint arXiv:2309.10815 (2023)", + "17. Gao, S., Yang, K., Shi, H., Wang, K., Bai, J.: Review on panoramic imaging and its applications in scene understanding. TIM (2022)", + "18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022)", + "19. Guo, J., Wang, Q., Gao, Y., Jiang, X., Tang, X., Hu, Y., Zhang, B.: MVP-SEG: Multi-view prompt learning for open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.06957 (2023)", + "20. Guttikonda, S., Rambach, J.: Single frame semantic segmentation using multimodal spherical images. In: WACV (2024)", + "21. Han, K., Liu, Y., Liew, J.H., Ding, H., Wei, Y., Liu, J., Wang, Y., Tang, Y., Yang, Y., Feng, J., Zhao, Y.: Global knowledge calibration for fast open-vocabulary segmentation. arXiv preprint arXiv:2303.09181 (2023)", + "22. Hu, X., An, Y., Shao, C., Hu, H.: Distortion convolution module for semantic segmentation of panoramic images based on the image-forming principle. TIM (2022)", + "23. Jang, S., Na, J., Oh, D.: DaDA: Distortion-aware domain adaptation for unsupervised semantic segmentation. In: NeurIPS (2022)", + "24. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Towards complete surrounding understanding via unsupervised contrastive learning. In: IV (2021)", + "25. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Insights into surrounding parsing for mobile agents via unsupervised contrastive learning. T-ITS (2023)", + "26. Jiang, Q., Gao, S., Gao, Y., Yang, K., Yi, Z., Shi, H., Sun, L., Wang, K.: Minimalist and high-quality panoramic imaging with PSF-aware transformers. TIP (2024)", + "27. Jiang, Q., Shi, H., Sun, L., Gao, S., Yang, K., Wang, K.: Annular computational imaging: Capture clear panoramic images through simple lens. TCI (2022)", + "28. Jiang, W., Wu, Y.: DFNet: Semantic segmentation on panoramic images with dynamic loss weights and residual fusion block. In: ICRA (2018)", + "29. Jiayun, L., Khandelwal, S., Sigal, L., Li, B.: Plug-and-play, dense-label-free extraction of open-vocabulary semantic segmentation from vision-language models. arXiv preprint arXiv:2311.17095 (2023)", + "30. Kim, J., Jeong, S., Sohn, K.: PASTS: Toward effective distilling transformer for panoramic semantic segmentation. In: ICIP (2022)", + "31. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)", + "32. Li, J., Chen, P., Qian, S., Jia, J.: TagCLIP: Improving discrimination ability of open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.07547 (2023)", + "33. Li, X., Wu, T., Qi, Z., Wang, G., Shan, Y., Li, X.: SGAT4PASS: Spherical geometry-aware transformer for panoramic semantic segmentation. In: IJCAI (2023)", + "34. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023)", + "35. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Mar-culescu, D.: Open-vocabulary semantic segmentation with mask-adapted CLIP. In: CVPR (2023)", + "36. Ling, Z., Xing, Z., Zhou, X., Cao, M., Zhou, G.: PanoSwin: A pano-style swim transformer for panorama understanding. In: CVPR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Liu, Y., Ge, P., Liu, Q., Huang, D.: Multi-grained cross-modal alignment for learning open-vocabulary semantic segmentation from text supervision. arXiv preprint arXiv:2403.03707 (2024)", + "38. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023)", + "39. Ma, C., Yang, Y.H., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: BMVC (2022)", + "40. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096 (2023)", + "41. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: DensePASS: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: ITSC (2021)", + "42. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: ECCV (2022)", + "43. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995)", + "44. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H.S., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023)", + "45. Oin, J., Wu, J., Yan, P., Li, M., Ren, Y., Xiao, X., Wang, Y., Wang, R., Wen, S., Pan, X., Wang, X.: FreeSeg: Unified, universal and open-vocabulary image segmentation. In: CVPR (2023)", + "46. Orhan, S., Bastanlar, Y.: Semantic segmentation of outdoor panoramic images. SIVP (2021)", + "47. Orsic, M., Kreso, I., Bevandic, P., Segvic, S.: In defense of pre-trained ImageNet architectures for real-time semantic segmentation of road-driving images. In: CVPR (2019)", + "48. Poudel, R.P.K., Liwicki, S., Cipolla, R.: Fast-SCNN: Fast semantic segmentation network. In: BMVC (2019)", + "49. Radford, A., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "50. Ray, B., Jung, J., Larabi, M.: A low-complexity video encoder for equirectangular projected 360 video content. In: ICASSP (2018)", + "51. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: PanoFormer: Panorama transformer for indoor $360^{\\circ}$ depth estimation. In: ECCV (2022)", + "52. Shi, H., Zhou, Y., Yang, K., Yin, X., Wang, Z., Ye, Y., Yin, Z., Meng, S., Li, P., Wang, K.: PanoFlow: learning $360^{\\circ}$ optical flow for surrounding temporal understanding. T-ITS (2023)", + "53. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. arXiv preprint arXiv:2306.13631 (2023)", + "54. Teng, Z., Zhang, J., Yang, K., Peng, K., Shi, H., Reifs, S., Cao, K., Stiefelhagen, R.: 360BEV: Panoramic semantic mapping for indoor bird's-eye view. In: WACV (2024)", + "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "56. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "57. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., Wang, X., Qiao, Y.: InternImage: Exploring large-scale vision foundation models with deformable convolutions. In: CVPR (2023)", + "58. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. arXiv preprint arXiv:2307.00764 (2023)", + "59. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: OV-PARTS: Towards open-vocabulary part segmentation. arXiv preprint arXiv:2310.05107 (2023)", + "60. Wu, C., Zheng, J., Pfrommer, J., Beyerer, J.: Attention-based point cloud edge sampling. In: CVPR (2023)", + "61. Wysocza'nska, M., Ramamonjisoa, M., Trzci'nski, T., Simeoni, O.: CLIP-DIY: CLIP dense inference yields open-vocabulary semantic segmentation for-free. arXiv preprint arXiv:2309.14289 (2023)", + "62. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022)", + "63. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: DAT++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430 (2023)", + "64. Xie, B., Cao, J., Xie, J., Khan, F.S., Pang, Y.: SED: A simple encoder-decoder for open-vocabulary semantic segmentation. arXiv preprint arXiv:2311.15537 (2023)", + "65. Xiong, Y., Li, Z., Chen, Y., Wang, F., Zhu, X., Luo, J., Wang, W., Lu, T., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications. arXiv preprint arXiv:2401.06197 (2024)", + "66. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)", + "67. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023)", + "68. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: SAN: Side adapter network for open-vocabulary semantic segmentation. TPAMI (2023)", + "69. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023)", + "70. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022)", + "71. Xu, X., Xiong, T., Ding, Z., Tu, Z.: MasQCLIP for open-vocabulary universal image segmentation. In: ICCV (2023)", + "72. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Huang, X., Sun, D., Wang, K.: Can we PASS beyond the field of view? Panoramic annular semantic segmentation for real-world surrounding perception. In: IV (2019)", + "73. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Wang, K.: PASS: Panoramic annular semantic segmentation. T-ITS (2020)", + "74. Yang, K., Hu, X., Fang, Y., Wang, K., Stiefelhagen, R.: Omnisupervised omnidirectional semantic segmentation. T-ITS (2022)", + "75. Yang, K., Hu, X., Stiefelhagen, R.: Is context-aware CNN ready for the surroundings? Panoramic semantic segmentation in the wild. TIP (2021)", + "76. Yang, K., Zhang, J., Reif, S., Hu, X., Stiefelhagen, R.: Capturing omni-range context for omnidirectional segmentation. In: CVPR (2021)", + "77. Yang, M., Yu, K., Zhang, C., Li, Z., Yang, K.: DenseASPP for semantic segmentation in street scenes. In: CVPR (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Zheng et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "78. Yin, W., Liu, Y., Shen, C., Hengel, A.v.d., Sun, B.: The devil is in the labels: Semantic segmentation from sentences. arXiv preprint arXiv:2202.02002 (2022)", + "79. Yu, H., He, L., Jian, B., Feng, W., Liu, S.: PanelNet: Understanding 360 indoor environment via panel representation. In: CVPR (2023)", + "80. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Gao, J., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV (2023)", + "81. Zhang, J., Ma, C., Yang, K., Roitberg, A., Peng, K., Stiefelhagen, R.: Transfer beyond the field of view: Dense panoramic semantic segmentation via unsupervised domain adaptation. T-ITS (2022)", + "82. Zhang, J., Yang, K., Ma, C., Reif, S., Peng, K., Stiefelhagen, R.: Bending reality: Distortion-aware transformers for adapting to panoramic semantic segmentation. In: CVPR (2022)", + "83. Zhang, J., Yang, K., Shi, H., Reif, S., Peng, K., Ma, C., Fu, H., Torr, P.H.S., Wang, K., Stiefelhagen, R.: Behind every domain there is a shift: Adapting distortion-aware vision transformers for panoramic semantic segmentation. TPAMI (2024)", + "84. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017)", + "85. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Materobot: Material recognition in wearable robotics for people with visual impairments. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024)", + "86. Zheng, X., Pan, T., Luo, Y., Wang, L.: Look at the neighbor: Distortion-aware unsupervised domain adaptation for panoramic semantic segmentation. In: ICCV (2023)", + "87. Zheng, X., Zhu, J., Liu, Y., Cao, Z., Fu, C., Wang, L.: Both style and distortion matter: Dual-path unsupervised domain adaptation for panoramic semantic segmentation. In: CVPR (2023)", + "88. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentaion. arXiv preprint arXiv:2311.03352 (2023)", + "89. Zhou, Q., Liu, Y., Yu, C., Li, J., Wang, Z., Wang, F.: LMSeg: Language-guided multi-dataset segmentation. In: ICLR (2023)", + "90. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable ConvNets V2: More deformable, better results. In: CVPR (2019)" + ], + "bbox": [ + 212, + 146, + 784, + 604 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Open Panoramic Segmentation", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_model.json b/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b35c45fad4174828e2009860de19b1b0a8e2f48 --- /dev/null +++ b/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_model.json @@ -0,0 +1,2823 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.324, + 0.142, + 0.68, + 0.163 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.189, + 0.681, + 0.237 + ], + "angle": 0, + "content": "Junwei Zheng\\(^{1}\\), Ruiping Liu\\(^{1}\\), Yufan Chen\\(^{1}\\), Kunyu Peng\\(^{1}\\), Chengzhi Wu\\(^{1}\\), Kailun Yang\\(^{2}\\), Jiaming Zhang\\(^{1,\\dagger}\\), and Rainer Stiefelhagen\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.381, + 0.247, + 0.621, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\) Karlsruhe Institute of Technology \n\\(^{2}\\) Hunan University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.308, + 0.744, + 0.613 + ], + "angle": 0, + "content": "Abstract. Panoramic images, capturing a \\(360^{\\circ}\\) field of view (FoV), encompass omnidirectional spatial information crucial for scene understanding. However, it is not only costly to obtain training-sufficient dense-annotated panoramas but also application-restricted when training models in a close-vocabulary setting. To tackle this problem, in this work, we define a new task termed Open Panoramic Segmentation (OPS), where models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain, enabling the zero-shot open panoramic semantic segmentation ability of models. Moreover, we propose a model named OOOPS with a Deformable Adapter Network (DAN), which significantly improves zero-shot panoramic semantic segmentation performance. To further enhance the distortion-aware modeling ability from the pinhole source domain, we propose a novel data augmentation method called Random Equirectangular Projection (RERP) which is specifically designed to address object deformations in advance. Surpassing other state-of-the-art open-vocabulary semantic segmentation approaches, a remarkable performance boost on three panoramic datasets, WildPASS, Stanford2D3D, and Matterport3D, proves the effectiveness of our proposed OOOPS model with RERP on the OPS task, especially \\(+2.2\\%\\) on outdoor WildPASS and \\(+2.4\\%\\) mIoU on indoor Stanford2D3D. The source code is publicly available at OPS." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.637, + 0.376, + 0.652 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Panoramic imaging systems [26, 27] have advanced significantly in recent years, which has fostered a wide variety of panoramic vision applications [1, 17]. Due to the comprehensive \\(360^{\\circ}\\) Field of View (FoV), omnidirectional panoramas provide more informative visual cues when perceiving surroundings in a broad spectrum of scene understanding tasks [3, 36, 51, 52, 54, 79], enabling a more complete and immersive capture of environmental data, which is crucial for in-depth scene understanding. This wide perspective surpasses the limited scope of pinhole images, significantly enhancing the capability of computer vision systems to perceive and interpret surroundings in a variety of applications. While the benefits of utilizing panoramic images in computer vision applications are apparent compared" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.513, + 0.84 + ], + "angle": 0, + "content": "† Correspondence: jiaming.zhang@kit.edu" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.499, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.216, + 0.5, + 0.237 + ], + "angle": 0, + "content": "(a) The challenge of wider FoV. Performance (mIoU / open mIoU) drops when FoV increases." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.238, + 0.501, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.317, + 0.5, + 0.369 + ], + "angle": 0, + "content": "(b) Close-vocabulary vs. Open-vocabulary segmentation on panoramas. While limited categories are recognized in a close-vocabulary setting, all objects in \\(360^{\\circ}\\) can be covered in an open-vocabulary setting." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.15, + 0.783, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.342, + 0.784, + 0.364 + ], + "angle": 0, + "content": "(c) The paradigm of Open Panoramic Segmentation (OPS) task." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.478 + ], + "angle": 0, + "content": "Fig. 1: (a) The challenge of existing state-of-the-art segmentation models. (b) The limitation of categories in traditional close-vocabulary panoramic segmentation tasks. (c) Our newly defined Open Panoramic Segmentation (OPS) task aims at tackling the above challenges. OPS consists of three important elements: Open the FoV targeted at the challenge of \\(360^{\\circ}\\) FoV, Open the Vocabulary targeted at the drawback of close-vocabulary panoramic segmentation and Open the Domain targeted at the challenge of scarcity of panoramic labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.764 + ], + "angle": 0, + "content": "with pinhole images, it is imperative to give increasing consideration to some noteworthy challenges as follows: (1) The challenge of wider FoV. For example, Fig. 1a illustrates the degraded performance of state-of-the-art open-vocabulary semantic segmentation approaches [9,68,80] with an increasing FoV, i.e., from pinhole images to \\(360^{\\circ}\\) panoramic images. Over \\(12\\%\\) mIoU performance degradation can be observed, indicating the challenges introduced by the large divergence of semantic and structural information across narrow-wide imagery. (2) The limitation of categories. The traditional close-vocabulary segmentation task paradigm [74,82] only provides a limited amount of labeled categories, which cannot handle the inestimable amount of categories in real-world applications. Fig. 1b illustrates the difference between close- and open-vocabulary segmentation. Compared to the close-vocabulary setting (first row) the open-vocabulary setting (second row) is not limited by the number of categories of the datasets. Only four predefined categories (highlighted in different colors) are recognized in the close-vocabulary setting in contrast to the open one, where every pixel of the panorama has its own semantic meaning even though the categories are not annotated in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.79, + 0.842 + ], + "angle": 0, + "content": "To further release the great potential of panoramic images three critical problems are eager to be solved: How to obtain a holistic perception via a single image? How to break the barrier of limited recognizable categories in existing panoramic datasets so that downstream vision applications can benefit from unrestricted informative visual hints? How to deal with the scarcity of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.523, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "panoramic labels? Based on the aforementioned three questions, we propose a new task named Open Panoramic Segmentation (OPS) that comprehensively addresses these challenges, aiming to better leverage the advantages brought by panoramic images. The new task paradigm is shown in Fig. 1c. The open panoramic segmentation task consists of three important elements considering the three problems. The concept of \"Open\" is three-fold: omnidirectional panoramic images (Open the FoV), an unrestricted range of recognizable classes (Open the Vocabulary), and additionally, models are trained using pinhole images in the source domain while evaluated with panoramic images in the target domain (Open the Domain). Given that densely annotated pinhole segmentation labels incur lower costs compared to panoramic ones [74], it is cost-efficient to open different domains. Note that OPS is different from Domain Adaptation (DA). In the context of training in DA, the utilization of data encompasses both the source and target domains while in OPS, exclusive reliance is placed on data originating solely from the source domain during the whole training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Apart from the new task, we propose a new model named OOOPS to address the aforementioned three openness-related challenges in the OPS task. It consists of a frozen CLIP model [49] and a key component named Deformable Adapter Network (DAN), which serves two critical purposes: (1) efficiently adapting the frozen CLIP model to the panoramic segmentation task and (2) addressing object deformation and image distortion inherent in panoramas. More specifically, a novel Deformable Adapter Operator (DAO), the key component of DAN, is designed to tackle panoramic distortion. To advance the modeling capabilities for distortion awareness in the pinhole source domain, we further introduce Random Equirectangular Projection (RERP) explicitly crafted to tackle object deformations and image distortions. A pinhole image is divided into four image patches that are randomly shuffled. Equirectangular Projection, one of the most common methods for mapping a globe into a panoramic plane, is then introduced in the shuffled image. Our OOOPS model with RERP outperforms other state-of-the-art open-vocabulary segmentation methods with \\(+2.2\\%\\), \\(+2.4\\%\\), and \\(+0.6\\%\\) mIoU on WildPASS [76], Stanford2D3D [2], and Matterport3D [6], respectively." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.65, + 0.631, + 0.665 + ], + "angle": 0, + "content": "To summarize, we present the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.674, + 0.787, + 0.749 + ], + "angle": 0, + "content": "- We introduce a new task termed open panoramic segmentation, OPS for short, including Open FoV, Open Vocabulary and Open Domain. Models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.75, + 0.787, + 0.809 + ], + "angle": 0, + "content": "- We put forward a model called OOOPS with the aim to address three openness-related challenges at once. A Deformable Adapter Network (DAN) is proposed to transfer the zero-shot learning ability of a frozen CLIP model from the pinhole domain to different panoramic domains." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- A novel data augmentation strategy named Random Equirectangular Projection (RERP), which is specifically designed for the proposed OPS" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "task, further boosts the accuracy of the OOOPS model, achieving state-of-the-art performance in the open panoramic segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.786, + 0.222 + ], + "angle": 0, + "content": "- For benchmarking OPS, we conduct comprehensive evaluations on both indoor and outdoor datasets (WildPASS, Stanford2D3D and Matterport3D), involving more than 10 close- and open-vocabulary segmentation models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.244, + 0.388, + 0.26 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.275, + 0.601, + 0.29 + ], + "angle": 0, + "content": "2.1 Open-vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.794, + 0.555 + ], + "angle": 0, + "content": "Motivated by the success of vision-language pre-training models, Open-Vocabulary Semantic Segmentation (OVSS) has been recently explored to achieve open-set pixel-wise understanding [7, 11, 12, 18, 19, 29, 31, 32, 34, 34, 37, 39, 40, 44, 53, 56, 58, 59, 61, 64, 70, 71]. Most works leverage the open-vocabulary capabilities of CLIP [49] to reason about relations between dense pixel semantics and arbitrary class labels [9, 12, 32, 38, 61, 71]. In [78, 89], multiple datasets are unified by transferring heterogeneous labels into text embeddings for training to achieve OVSS. In [35, 67], image-caption datasets are employed for learning alignment between visual regions and text entities. To reduce the computational costs of using a frozen pre-trained CLIP model for identifying novel classes, latest studies [21, 69] have also proposed efficient architectures for fast OVSS. SAN [68, 69] has designed an asymmetric input resolution strategy to decouple mask proposal prediction and attention bias prediction applied to CLIP for mask class recognition. Furthermore, open-vocabulary panoptic segmentation has been tackled in [8, 12, 13, 34, 45, 66, 66, 71, 80] to render universal image understanding. In this work, we focus on the open-vocabulary zero-shot learning on panoramas instead of pinhole images discussed in the aforementioned works." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.576, + 0.549, + 0.59 + ], + "angle": 0, + "content": "2.2 Panoramic Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Panoramic semantic segmentation enables more informative decision-making and interaction within complex and dynamic environments, advancing the capabilities of intelligent systems for navigation and perception [4, 16, 24, 25, 28, 33, 42, 46, 75]. PASS [72, 73] achieves panoramic segmentation on annular images, whereas 360BEV [54] tackles panoramic semantic mapping. ECANets [76] address omnidirectional semantic segmentation with multi-source omni-supervised learning to enhance generalization. Zhang et al. [81] propose to use attention-augmented designs for panoramic representation learning. Guttikonda et al. [20] propose a single-frame panoramic semantic segmentation approach using multimodal spherical images. Hu et al. [22] propose a distortion convolution module for panoramic semantic segmentation. An important cluster of works revisits panoramic semantic segmentation from an unsupervised domain adaptation perspective [23,30,41,81,82,86,87], by adapting from label-rich pinhole images to the panoramic images. Unlike these works, we propose a method to tackle open-vocabulary panoramic semantic segmentation for unconstrained surrounding understanding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.523, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.503, + 0.162 + ], + "angle": 0, + "content": "2.3 Deformable Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.414 + ], + "angle": 0, + "content": "In the realm of deformable neural networks, notable advancements have been made through various iterations. The pioneering work [10] introduces the concept of deformable convolution, laying the foundation for subsequent developments. Building upon this foundation, DCNv2 [90] further refines deformable convolution by introducing a modulation mechanism that expands the scope of deformation modeling. The evolution of deformable neural networks continues with the introduction of DCNv3 [57], which combines the group convolution with DCNv2 for the sake of imitating multi-head mechanism of Transformer [55]. DCNv4 [65] aims to speedup DCNv3, achieving a more efficient model. Additionally, Trans4PASS [82,83] proposes Deformable Patch Embedding (DPE) and Deformable MLP (DMLP) to handle the distortion that occurs in panoramas. Incorporating the principles of deformable convolution into vision transformers [14], Xia et al. propose DeformableViT [62,63], extending the applicability of deformable neural networks to the domain of transformer architectures. After rethinking a series of DCN designs, we propose DAO to improve the model's ability to handle image distortion in the OPS task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.381, + 0.454 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.518, + 0.483 + ], + "angle": 0, + "content": "3.1 Open Panoramic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Open Panoramic Segmentation (OPS) task is proposed to deal with three challenging problems: ① the narrow FoV, ② the restricted range of classes, and ③ the scarcity of panoramic labels. OPS gives three answers to the aforementioned issues: ① Open the FoV, ② Open the Vocabulary, and ③ Open the Domain. The OPS task paradigm is shown in Fig. 1c. Models are trained in the narrow-FoV pinhole source domain in an open-vocabulary setting while evaluated in the wide-FoV panoramic target domain." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.427, + 0.634 + ], + "angle": 0, + "content": "3.2 Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "A foundation model [49] can be efficiently transferred to downstream tasks by using an adapter [68]. To improve the modeling ability for panoramas, we design the OOOPS model. As shown in Fig. 2, it comprises a frozen CLIP model and a proposed Deformable Adapter Network (DAN) that incorporates multiple Transformer Layers and the novel DAO, which will be detailed in Sec. 3.3. Feature fusion takes place between the intermediate layers of CLIP and DAN. One of the two outputs from DAN consists of mask proposals, while the other serves as deep supervised guidance for CLIP to generate proposal logits. In the training phase, pinhole images are forwarded into OOOPS, generating mask proposals and proposal logits for the loss calculation. In the inference phase, panoramas are forwarded into OOOPS to generate segmentation predictions via the multiplication of mask proposals and the corresponding proposal logits. The frozen CLIP is necessary for the zero-shot learning ability of OOOPS." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Fig. 2: Overview of the OOOPS model architecture. It consists of a frozen CLIP model and a Deformable Adapter Network (DAN) which includes Transformer Layers and the proposed DAO." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.442, + 0.509, + 0.459 + ], + "angle": 0, + "content": "3.3 Deformable Adapter Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.57 + ], + "angle": 0, + "content": "The Deformable Adapter Network is a combination of multiple Transformer [55] Layers and the proposed DAO. Since distortion exists in panoramic images, which is a big challenge when utilizing informative panoramas, we delve deeper into the deformable design [57] and sampling methods like APES [60], MateRobot [85], proposing DAO to tackle the image distortions and object deformations of panoramas. To provide a detailed rethinking process of the deformable convolution design, we introduce the preliminary of the DCN series as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Revisiting DCN Series. The seminal work DCN [10] can enable the traditional CNN with spatial deformation-aware ability. Given a convolutional kernel of \\(K\\) sampling locations, let \\(\\mathbf{w}_k\\) and \\(\\mathbf{p}_k\\) denote the weight and pre-specified offset for the \\(k\\)-th location, respectively. For example, \\(K = 9\\) and \\(\\mathbf{p}_k \\in \\{(1,1), \\dots, (-1, -1)\\}\\) defines a \\(3 \\times 3\\) convolutional kernel of dilation 1. Let \\(\\mathbf{x}(\\mathbf{p})\\) and \\(\\mathbf{y}(\\mathbf{p})\\) denote the features at location \\(\\mathbf{p}\\) from the input feature maps \\(\\mathbf{x}\\) and output feature maps \\(\\mathbf{y}\\), respectively. DCN is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.685, + 0.787, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} (\\mathbf {p}) = \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {k} \\mathbf {x} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {k}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.795 + ], + "angle": 0, + "content": "where \\(\\Delta \\mathbf{p}_k\\) is the learnable offset for the \\(k\\)-th location. Although DCN is capable of capturing spatial deformation, every sampling location is treated equally when calculating the local features. DCNv2 [90] is proposed with an additional term called modulation scalar. Specifically, DCNv2 can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.802, + 0.787, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} (\\mathbf {p}) = \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {k} \\mathbf {m} _ {k} \\mathbf {x} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {k}\\right), \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.523, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}_k\\) is the learnable modulation scalar for the \\(k\\)-th location. Inspired by Transformer, DCNv3 [57] is proposed with a grouping operation, further boosting the deformation-aware ability of DCNv2. DCNv3 can be addressed with the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.218, + 0.786, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} (\\mathbf {p}) = \\sum_ {g = 1} ^ {G} \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {g} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {g k}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.271, + 0.788, + 0.303 + ], + "angle": 0, + "content": "where \\( G \\) denotes the total number of aggregation groups. DCNv4 [65] is similar to DCNv3, achieving similar performance while significantly reducing the runtime." + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.331, + 0.648, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.49, + 0.637, + 0.506 + ], + "angle": 0, + "content": "Fig. 3: Salient map generation in DAO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Deformable Adapter Operator (DAO). When dealing with distortions in panoramas, DCNv3 and DCNv4 fall short of meeting the distortion-aware requirements. This is illustrated by the visualization of Fig. 6 in Sec. 4.5. Therefore, DAO is proposed to tackle the distortion problem in panoramic images using the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.621, + 0.787, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} (\\mathbf {p}) = \\mathbf {s} (\\mathbf {p}) \\sum_ {g = 1} ^ {G} \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {g} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {g k}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}(\\mathbf{p})\\) is the learnable salient scalar at location \\(\\mathbf{p}\\). Inheriting from DCNv3, DAO comes up with an additional term called salient scalar indicating the importance of each pixel within the whole panorama. It's worth noting that DCNv3 and DCNv4 share the same mathematical expression, but DCNv3 is more robust according to our experiments before designing DAO. Therefore, we adopt DCNv3 as a part of DAO rather than DCNv4. Illustrated in Fig. 2, the feature map outputted by DCNv3 passes through the Patch Similarity Layer, Normalization Layer, and Standard Deviation Layer sequentially to form a salient map. The intuition for such a design is straightforward. The salient pixels in an image are those that differ significantly from their neighboring pixels, e.g., edge pixels. If all pixels within an image patch are different, the standard deviation of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.145, + 0.414, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.206, + 0.406, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.233, + 0.493, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.273, + 0.496, + 0.316 + ], + "angle": 0, + "content": "(a) Equirectangular Projection (ERP) maps a globe into a panorama. It's clear that strong distortion occurs in the panorama, especially on object edges in the horizontal direction." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.145, + 0.591, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.145, + 0.676, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.209, + 0.591, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.208, + 0.676, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.145, + 0.762, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.208, + 0.762, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.273, + 0.766, + 0.316 + ], + "angle": 0, + "content": "(b) Visualization of ERP and RERP of pinhole images. First column: pinhole images. Second column: ERP on pinhole images. Third column: RERP on pinhole images." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.326, + 0.785, + 0.355 + ], + "angle": 0, + "content": "Fig. 4: (a) Visualization of ERP on a panoramic image and (b) Our proposed RERP on pinhole images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.549 + ], + "angle": 0, + "content": "pixel similarity for this patch is higher than the one containing similar pixels, resulting in a higher salient scalar. Fig. 3 gives a more detailed explanation of the salient map generation. Given a feature map, DAO first calculates the cosine similarity between the center pixel and all pixels within a kernel, e.g., 9 pixels within a \\(3 \\times 3\\) kernel in Fig. 3, resulting in a 9-dimensional cosine similarity vector. Softmax normalization is then applied to the vector. Afterward, DAO calculates the standard deviation of this normalized cosine similarity vector, indicating the importance of the center pixel. By traversing every pixel across the entire feature map, a salient map is generated for the sake of enhancing the salient pixels, which are usually the edge pixels of an image where strong panoramic distortion occurs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.568, + 0.562, + 0.584 + ], + "angle": 0, + "content": "3.4 Random Equirectangular Projection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.637 + ], + "angle": 0, + "content": "The Equirectangular Projection [50] (ERP) is one of the most common methods for mapping a globe into a panoramic plane, which transforms spherical coordinates into planar coordinates expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.646, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\nx = R \\left(\\lambda - \\lambda_ {0}\\right) \\cos \\left(\\varphi_ {1}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.665, + 0.785, + 0.681 + ], + "angle": 0, + "content": "\\[\ny = R \\left(\\varphi - \\varphi_ {0}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(\\varphi\\) are the longitude and latitude of the location to project, respectively. \\(\\varphi_{1}\\) are the standard parallels. \\(\\lambda_0\\) and \\(\\varphi_0\\) are the central meridian and central parallel of the map, respectively. \\(R\\) is the radius of the globe. \\(x\\) denotes the horizontal coordinate, while \\(y\\) represents the vertical coordinate of the projected location on the map. Fig. 4a visualizes the ERP on a panoramic plane. It can be observed that after equirectangular projection strong distortion occurs in the panorama, e.g., a straight line is transformed into a curve. To further boost the performance, we propose Random Equirectangular Projection (RERP) on pinhole images since the OPS task requires the model to train with pinhole images instead of panoramas." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.524, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.788, + 0.268 + ], + "angle": 0, + "content": "We divide the pinhole image into four parts and randomly shuffle the image patches. Afterward, we apply equirectangular projection to the distortion-free pinhole image. Fig. 4b visualizes pinhole images after RERP. The first column is the pinhole images without any data augmentation. The second column is ERP on pinhole images without random shuffling. The last column is our proposed RERP. It can be observed that the panorama-like distortion also occurs in the pinhole images after RERP. Random shuffling is employed to enhance robustness and promote generalization." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.289, + 0.375, + 0.308 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.32, + 0.334, + 0.334 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.343, + 0.813, + 0.403 + ], + "angle": 0, + "content": "We train our model using the pinhole general-object COCOStuff-164k [5] while evaluating the model performance on three different panoramic datasets: panoramic street-view WildPASS [76], panoramic indoor-scene Stanford2D3D [2] and panoramic indoor-scene Matterport3D [6]." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.404, + 0.786, + 0.463 + ], + "angle": 0, + "content": "COCOStuff-164k. The pinhole dataset comprises 164k images across 171 annotated classes, partitioned into training, validation, and test sets, with 118k, 5k, and 41k images, respectively. In our experiments, we utilize the entire 118k-image training set as the training data by default." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.464, + 0.786, + 0.523 + ], + "angle": 0, + "content": "WildPASS. The panoramic street-view dataset contains 2,500 panoramas in 65 cities from all continents Asia, Europe, Africa, Oceania, and North and South America (excluding only Antarctica) with 8 dense-annotated classes. All images have a \\(70^{\\circ} \\times 360^{\\circ}\\) FoV in the size of \\(400 \\times 2048\\)." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.524, + 0.786, + 0.568 + ], + "angle": 0, + "content": "Stanford2D3D. The panoramic indoor dataset consists of 1,413 images in the size of \\(512 \\times 1024\\) and 13 object classes. All results are averaged over 3 cross-validation folds." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.569, + 0.786, + 0.629 + ], + "angle": 0, + "content": "Matterport3D. The panoramic indoor dataset includes 10,800 panoramic views from 194,400 RGB-D images. According to the data preprocessing pipeline of 360BEV [54], a subset of 772 images in the size of \\(512 \\times 1024\\) with 20 classes is used to evaluate the segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.651, + 0.325, + 0.665 + ], + "angle": 0, + "content": "4.2 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.674, + 0.786, + 0.764 + ], + "angle": 0, + "content": "Close-vocabulary Metric. Mean Intersection over Union (mIoU) is used as the evaluation metric in close-vocabulary semantic segmentation tasks. Specifically, IoU is calculated as the area of overlap between the predicted segmentation and the ground truth, divided by the area of union between them. mIoU is a hard metric designed for close-vocabulary segmentation ignoring the semantic meanings of similar classes referring to the open-vocabulary setting." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Open-vocabulary Metric. Open mean Intersection over Union [88], open mIoU for short, is a soft metric compared to mIoU since the similarity scores between different classes are also taken into account, which are calculated by the WordNet approach [43]. In this work, we will report both close and open metrics for a comprehensive understanding of all models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 1: Results on WildPASS. The mIoU, open mIoU, and per-class IoU are in percentage (%). #PARAMs is the number of learnable parameters in millions." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.189, + 0.784, + 0.415 + ], + "angle": 0, + "content": "
Network#PARAMSmIoUOpen mIoUCarRoadSidewalkCrosswalkCurbpersonTruckBus
Close-Vocabulary
Fast-SCNN [48]1.124.8-45.960.031.79.717.16.014.213.8
PSPNet18 [84]17.528.2-58.266.828.413.019.26.218.215.6
SwiftNet [47]11.830.0-55.764.129.216.222.98.521.122.2
DenseASPP [77]8.333.2-51.169.138.116.426.38.727.428.4
ERF-PSPNet [73]2.534.0-66.370.536.56.424.19.426.532.0
PSPNet50 [84]53.346.1-80.074.951.723.931.419.838.948.1
DANet [15]47.447.2-74.872.249.928.923.825.451.950.6
OOSS (ERF-PSPNet) [74]2.556.1-87.279.360.828.038.154.548.852.2
PASS (ERF-PSPNet) [73]2.564.7-87.380.061.471.149.972.237.557.9
ERF-PSPNet (omni-sup') [76]2.566.8-90.582.765.670.551.558.262.053.1
ECANet (with attention) [76]2.667.7-90.483.768.467.852.161.454.563.3
ECANet (with fusion) [76]2.669.0-90.685.768.067.952.166.059.362.3
Open-Vocabulary
SAN [68]8.455.656.984.381.969.610.12.872.254.969.1
CAT-seg [9]59.555.757.184.181.570.010.23.072.554.869.3
OpenSeeD [80]65.455.857.284.381.669.710.33.073.654.969.1
OOOPS (ours, w/o RERP)8.757.058.584.981.668.715.18.471.355.370.4
OOOPS (ours, w/ RERP)8.758.059.784.981.668.719.48.373.854.972.2
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.457, + 0.47 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.678 + ], + "angle": 0, + "content": "We utilize CLIP ViT-B/16 as the frozen CLIP model. All models are trained with the pinhole general-object COCOStuff-164k training set with \\(4 \\times \\mathrm{A}40\\) GPUs while evaluated on three panoramic datasets which are in totally different domains compared with the pinhole dataset. The AdamW optimizer is applied with an initial learning rate of 0.0001, weight decay of 0.0001, batch size of 32, and a total of 60,000 training iterations. Throughout the training, the learning rate follows a polynomial schedule with a power of 0.9. Data augmentations, including random image resizing within the short-side range of [320, 1024] and a crop size of \\(640 \\times 640\\) are applied. Additionally, when verifying the effectiveness of the proposed RERP, it is applied as an extra data augmentation. Following the practice of SAN [68], apart from the supervision of mask recognition with cross-entropy loss, we utilize the dice loss and binary cross-entropy loss for mask proposal generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.696, + 0.577, + 0.712 + ], + "angle": 0, + "content": "4.4 Comparison with the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Results on WildPASS. Comparisons with both close- and open-vocabulary state-of-the-art methods on WildPASS are listed in Table 1. It is not surprising that ECANet (with fusion) [76] achieves the best performance among all methods since it is a close-vocabulary method, meaning that ECANet is only able to predict the classes predefined by the dataset. However, our OOOPS model is qualified for predicting classes that are not annotated in the dataset. The OOOPS method (with RERP) surpasses other state-of-the-art open-vocabulary panoramic segmentation methods with over \\(+2.2\\%\\) mIoU while introducing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.524, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.498, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.217, + 0.237, + 0.498, + 0.289 + ], + "angle": 0, + "content": "(a) Comparison with baseline method SAN [68] on the WildPASS dataset with predefined 8 classes. First row: ground truth. Second row: prediction from SAN. Third row: prediction from OOOPS with RERP." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.145, + 0.785, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.505, + 0.237, + 0.787, + 0.289 + ], + "angle": 0, + "content": "(b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings. First row: RGB. Second row: prediction with close vocabularies. Third row: prediction with open vocabularies. Zoom in for a better view." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.301, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Fig. 5: (a) Comparison on the WildPASS dataset and (b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.58 + ], + "angle": 0, + "content": "only a minor increment of 0.3 million learnable parameters compared to the lightweight baseline SAN [68], offering a promising step towards bridging the gap between close- and open-vocabulary settings. Moreover, our OOOPS method enhanced by RERP demonstrates significant advancements in challenging categories such as crosswalk and curb, achieving \\(+9.3\\%\\) and \\(+5.5\\%\\) mIoU compared with the baseline SAN, respectively. It is worth noting that in some categories, e.g., sidewalk, person, and bus, our proposed OOOPS (both with and without RERP) even outperforms the models trained in a close-vocabulary setting. All open-vocabulary methods achieve better performance with open mIoU metric compared to the close mIoU since similar semantic meanings are also considered in the calculation. Fig. 1a showcases the performance degradation on both metrics. Even though all methods have degraded performance when FoV increases, our OOOPS achieves a smaller drop \\((-10.0\\%)\\) while others have a larger performance decrease \\((-12.0\\%)\\), indicating the effectiveness of the OOOPS model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Fig. 5a visualizes the prediction of the challenging curb category from baseline SAN and our proposed OOOPS with RERP. From top to bottom are the ground truth, prediction from SAN, and prediction from OOOPS with RERP, respectively. The baseline SAN cannot recognize the challenging curb category while our proposed OOOPS with RERP is capable of detecting the curb in the middle of the panorama highlighted in gray. Fig. 5b showcases the visualization of the prediction from OOOPS on WildPASS in close- and open-vocabulary settings. From top to bottom are input RGB, prediction from OOOPS with 8 predefined close vocabularies from WildPASS, and prediction from OOOPS with unlimited open vocabularies. It is evident that our OOOPS with RERP is capable of tackling unseen categories in the open-vocabulary setting. Every pixel on the panorama is assigned a meaningful semantic label. For example, \"mailbox\" is not a predefined category from WildPASS. While existing models can not recognize the mailbox under a close-vocabulary setting, the rare classes can be easily detected by our model in an open-vocabulary one, indicating the great potential of open-vocabulary panoramic segmentation for parsing unconstrained traffic surroundings." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Results on Stanford2D3D. The mIoU, open mIoU, and per-class IoU are in percentage (\\%). #PARAMs is the number of learnable parameters in millions." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.192, + 0.784, + 0.349 + ], + "angle": 0, + "content": "
Network#PARAMsmIoUOpen mIoUBeamBoardBookcaseCeilingChairClutterColumnDoorFloorSofaTableWallWindow
Close-Vocabulary
Trans4PASS [82]28.252.1-11.460.651.181.859.737.923.554.688.834.653.968.151.0
360BEV [54]28.454.3-06.063.655.582.762.739.916.458.890.845.257.171.157.6
Open-Vocabulary
SAN [68]8.438.539.50.251.539.666.552.110.70.045.049.331.849.250.053.4
CAT-seg [9]59.538.639.60.051.642.366.851.48.00.047.450.731.247.551.453.5
OpenSeeD [80]65.438.740.00.051.842.067.051.68.40.247.550.031.447.151.753.6
OOOPS (w/o RERP)8.739.541.00.052.344.167.153.79.70.050.152.830.547.252.253.8
OOOPS (w/ RERP)8.741.142.60.050.744.868.851.88.10.048.271.133.647.255.454.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.376, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Table 3: Results on Matterport3D. The mIoU, open mIoU, and per-class IoU are in percentage (\\%). #PARAMs is the number of learnable parameters in millions." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.422, + 0.784, + 0.54 + ], + "angle": 0, + "content": "
Network≈PARAMsinotOpen mIoUWallFloorChairDoorTablePictureFurnitureObjectsWindowSofbBedSinkStarsCeilingToiletMirrorshowerBathtubCounterShelving
Close-Vocabulary
Trans4PASS+ [83]28.542.6-63.479.139.140.332.836.031.031.537.544.063.220.641.877.640.724.323.758.334.332.9
360BEV [54]28.446.4-64.183.145.845.038.041.132.335.140.648.769.825.147.880.246.028.722.360.138.634.8
Open-Vocabulary
SAN [68]8.430.531.454.476.118.034.122.321.113.01.732.734.251.011.019.974.223.412.031.631.919.028.4
CAT-seg [9]59.530.231.154.374.817.034.121.920.812.61.932.934.050.110.718.873.522.411.631.932.919.228.5
OpenSecD [80]65.430.631.654.276.318.134.322.521.013.02.132.834.150.911.220.274.123.512.231.131.819.128.6
OOOPS (w/o RERP)8.731.132.455.277.017.135.725.017.513.52.333.032.451.812.620.474.522.610.834.136.819.429.7
OOOPS (w/ RERP)8.731.232.555.476.418.434.823.921.213.22.633.334.552.912.120.374.628.212.531.929.319.428.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Results on Stanford2D3D. In Table 2, the comparison is conducted between state-of-the-art methods on the Stanford2D3D dataset. All reported results are averaged over 3 cross-validation folds. Without RERP, OOOPS achieves \\(39.5\\%\\) mIoU, outperforming other open-vocabulary methods like SAN [68], CAT-seg [9], and OpenSeeD [80], which score between \\(38.5\\%\\) and \\(38.7\\%\\). With RERP, OOOPS's mIoU increases to \\(41.1\\%\\) with a \\(+2.4\\%\\) gain, narrowing the gap with advanced close-vocabulary methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.796, + 0.841 + ], + "angle": 0, + "content": "Results on Matterport3D. In Table 3, the comparison is conducted between state-of-the-art methods on the Matterport3D dataset. Our OOOPS method showcases consistent performance. Without augmentation, OOOPS achieves mIoU of \\(31.1\\%\\), slightly outperforming other open-vocabulary methods like SAN [68] \\((30.5\\%)\\), CAT-seg [9] \\((30.2\\%)\\), and OpenSeeD [80] \\((30.6\\%)\\). With augmentation, OOOPS's mIoU marginally increases to \\(31.2\\%\\), further reinforcing its lead over these open-vocabulary methods. Both variants of OOOPS still lag behind close-vocabulary methods like Trans4PASS+ [83] and 360BEV [54], which however, are trained in a supervised fashion in a close-vocabulary setting on panoramas and have mIoUs ranging from \\(42.6\\%\\) to \\(46.4\\%\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.523, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.16 + ], + "angle": 0, + "content": "Table 4: Task difference between Domain Adaptation (DA) and the proposed OPS." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.171, + 0.785, + 0.226 + ], + "angle": 0, + "content": "
TaskDomain#ClassesmIoU (%)
SourceTargetWildPASSMatterport3DStanford2D3D
DA✓✓limited72.247.255.4
OPS✓✗unlimited58.031.241.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.237, + 0.784, + 0.251 + ], + "angle": 0, + "content": "Table 5: Ablation study of OOOPS components and the number of shuffling patches." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.267, + 0.499, + 0.31 + ], + "angle": 0, + "content": "(a) Ablation study of OOOPS components on the WildPASS Dataset. The mIoU is reported in percentage \\((\\%)\\). #PARAMs stands for the number of learnable parameters in millions." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.268, + 0.787, + 0.311 + ], + "angle": 0, + "content": "(b) Ablation study of different numbers of shuffling patches on WildPASS, Matterport3D and Stanford2D3D datasets. The mIoU is reported in percentage \\((\\%)\\)." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.321, + 0.486, + 0.438 + ], + "angle": 0, + "content": "
MethodERPRERP#PARAMsmIoU
SAN [68]✘✘8.455.6
SAN [68]✓✘8.456.0
+ DCN [68]✘✓8.456.5 (+0.9)
+ DCN✘✘8.655.6
+ DCN✓✘8.656.0
+ DCN✘✓8.656.6 (+1.0)
+ DCNv2✘✘8.755.7
+ DCNv2✓✘8.756.4
+ DCNv2✘✓8.756.6 (+0.9)
+ DCNv3✘✘8.755.8
+ DCNv3✓✘8.756.4
+ DCNv3✘✓8.756.8 (+1.0)
+ DCNv4✘✘8.755.8
+ DCNv4✓✘8.756.4
+ DCNv4✘✓8.756.8 (+1.0)
+ DAO (ours)✘✘8.757.0
+ DAO (ours)✓✘8.757.6
+ DAO (ours)✘✓8.758.0 (+1.0)
" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.322, + 0.785, + 0.437 + ], + "angle": 0, + "content": "
mIoU (%) #Patches Dataset4 9 16 25
WildPASS58.0 58.5 58.2 58.0
Matterport3D31.2 31.6 31.4 31.0
Stanford2D3D41.1 41.9 41.5 41.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Results in DA Subtask. The proposed OPS task differs from Domain Adaptation since no data from the target domain is available to models in OPS while models have access to the data in DA during training. Moreover, models are required to recognize an unlimited amount of classes in OPS while only a limited number of classes can be recognized in DA. DA is a special case of OPS, meaning that models in OPS can be easily adapted to DA and achieve better performance. The converse doesn't hold. Table 4 showcases the task difference between OPS and DA. It can be observed that over \\(+14\\%\\) gain in mIoU is achieved when adapting our OOOPS model in the DA subtask." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.391, + 0.636 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We replace DAO in OOOPS, as illustrated in Fig. 2, with DCN [10], DCNv2 [90], DCNv3 [57] and DCNv4 [65]. When referring to the component substitution without data augmentation in Table 5a, it is clear that our DAO surpasses other counterparts with a maximum performance gain of \\(+1.4\\%\\) mIoU. As for the component substitution with ERP, all models achieve a boost in the range of \\(+0.3\\%\\) to \\(+0.7\\%\\) compared with the one without data augmentation. The gain is even bigger considering the component substitution with RERP, achieving \\(+1.0\\%\\) mIoU. It is worth noting that RERP is able to boost the model performance stably among all models. The best model performance is achieved by the combination of DAO and RERP, yielding \\(58.0\\%\\) mIoU with only a slight increment in parameters. Table 5b showcases the influence of different numbers of shuffling patches. The performance is not proportional to the number of patches, indicating the importance of moderate content continuity during training." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.499, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.291, + 0.499, + 0.325 + ], + "angle": 0, + "content": "(a) Deformable offsets without RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.787, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.291, + 0.787, + 0.325 + ], + "angle": 0, + "content": "(b) Deformable offsets with RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 6: Visualization of deformable offsets. The green points \\(\\bullet\\) are sampling locations. The red points \\(\\bullet\\) are deformable offsets in 2 levels, indicating a deformable receptive field (e.g., each level has a \\(3 \\times 3\\) kernel size, resulting in \\((3 \\times 3)^{2} = 81\\) red points)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.79, + 0.624 + ], + "angle": 0, + "content": "To further illustrate the deformable ability of DAO and the superiority of RERP when tackling the distortion of panoramas, we visualize the deformable offset in Fig. 6. Fig. 6a presents the deformable offsets without RERP. It can be observed that DAO is better aware of the panoramic distortion compared with the rest in the DCN series. Since the strongest distortion occurs on the edge of objects, DAO focuses more on edge pixels for better distortion-aware feature extraction. The standard deviation of patch similarity in DAO implies the salience of the image patch center. By multiplicatively combining salient scalars and features, significant pixels, such as those representing object edges, are enhanced in their representation. In Fig. 6b, it is evident that the inclusion of RERP improves the distortion information capturing ability of all models in the DCN series, as opposed to their counterparts without RERP. This observation is consistent with the results in Table 5a. The incorporation of both DAO and RERP in OOOPS showcases a remarkable distortion-aware modeling capability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.36, + 0.691 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this work, we formulate the novel task of Open Panoramic Segmentation (OPS) for unfolding holistic scene understanding, which reaches Open FoV, Open Vocabulary, and Open Domain in surrounding scene segmentation. We have designed the Deformable Adapter Network (DAN) and Random Equirectangular Projection (RERP) for enhancing distortion-aware segmentation capacities of the novel proposed OOOPS model. Extensive experiments on WildPASS, Stanford2D3D, and Matterport3D benchmarks show that the proposed method significantly elevates the performance of state-of-the-art panoramic segmenters." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.523, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.298 + ], + "angle": 0, + "content": "This work was supported in part by the Ministry of Science, Research and the Arts of Baden-Württemberg (MWK) through the Cooperative Graduate School Accessibility through AI-based Assistive Technology (KATE) under Grant BW6-03, in part by BMBF through a fellowship within the IFI programme of DAAD, in part by the Helmholtz Association Initiative and Networking Fund on the HAICORE@KIT and HOREKA@KIT partition, in part by the National Key RD Program under Grant 2022YFB4701400, and in part by Hangzhou SurImage Technology Company Ltd." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.318, + 0.323, + 0.334 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.348, + 0.785, + 0.389 + ], + "angle": 0, + "content": "1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.785, + 0.416 + ], + "angle": 0, + "content": "2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.417, + 0.785, + 0.443 + ], + "angle": 0, + "content": "3. Athwale, A., Afrasiyabi, A., Lagüie, J., Shili, I., Ahmad, O., Lalonde, J.F.: DarSwin: Distortion aware radial swim transformer. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.785, + 0.485 + ], + "angle": 0, + "content": "4. Berenguel-Baeta, B., Bermudez-Cameo, J., Guerrero, J.J.: FreDSNet: Joint monocular depth and semantic segmentation with fast fourier convolutions from single panoramas. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.512 + ], + "angle": 0, + "content": "5. Caesar, H., Uijlings, J.R.R., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.785, + 0.553 + ], + "angle": 0, + "content": "6. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. 3DV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.785, + 0.594 + ], + "angle": 0, + "content": "7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.622 + ], + "angle": 0, + "content": "8. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. arXiv preprint arXiv:2303.11324 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.649 + ], + "angle": 0, + "content": "9. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.: CAT-Seg: Cost aggregation for open-vocabulary semantic segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.676 + ], + "angle": 0, + "content": "0. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.703 + ], + "angle": 0, + "content": "1. Dao, S.D., Shi, H., Phung, D., Cai, J.: Class enhancement losses with pseudo labels for open-vocabulary semantic segmentation. TMM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.731 + ], + "angle": 0, + "content": "2. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.758 + ], + "angle": 0, + "content": "3. Dong, B., Gu, J., Han, J., Xu, H., Zuo, W.: Towards universal vision-language omni-supervised segmentation. arXiv preprint arXiv:2303.06547 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., Lu, H.: Dual attention network for scene segmentation. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.348, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhou, X., Geiger, A., Liao, Y.: PanopticNeRF-360: Panoramic 3D-to-2D label transfer in urban scenes. arXiv preprint arXiv:2309.10815 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Gao, S., Yang, K., Shi, H., Wang, K., Bai, J.: Review on panoramic imaging and its applications in scene understanding. TIM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "19. Guo, J., Wang, Q., Gao, Y., Jiang, X., Tang, X., Hu, Y., Zhang, B.: MVP-SEG: Multi-view prompt learning for open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.06957 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "20. Guttikonda, S., Rambach, J.: Single frame semantic segmentation using multimodal spherical images. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "21. Han, K., Liu, Y., Liew, J.H., Ding, H., Wei, Y., Liu, J., Wang, Y., Tang, Y., Yang, Y., Feng, J., Zhao, Y.: Global knowledge calibration for fast open-vocabulary segmentation. arXiv preprint arXiv:2303.09181 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "22. Hu, X., An, Y., Shao, C., Hu, H.: Distortion convolution module for semantic segmentation of panoramic images based on the image-forming principle. TIM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "23. Jang, S., Na, J., Oh, D.: DaDA: Distortion-aware domain adaptation for unsupervised semantic segmentation. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "24. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Towards complete surrounding understanding via unsupervised contrastive learning. In: IV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.502 + ], + "angle": 0, + "content": "25. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Insights into surrounding parsing for mobile agents via unsupervised contrastive learning. T-ITS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "26. Jiang, Q., Gao, S., Gao, Y., Yang, K., Yi, Z., Shi, H., Sun, L., Wang, K.: Minimalist and high-quality panoramic imaging with PSF-aware transformers. TIP (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.556 + ], + "angle": 0, + "content": "27. Jiang, Q., Shi, H., Sun, L., Gao, S., Yang, K., Wang, K.: Annular computational imaging: Capture clear panoramic images through simple lens. TCI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.583 + ], + "angle": 0, + "content": "28. Jiang, W., Wu, Y.: DFNet: Semantic segmentation on panoramic images with dynamic loss weights and residual fusion block. In: ICRA (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.624 + ], + "angle": 0, + "content": "29. Jiayun, L., Khandelwal, S., Sigal, L., Li, B.: Plug-and-play, dense-label-free extraction of open-vocabulary semantic segmentation from vision-language models. arXiv preprint arXiv:2311.17095 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.651 + ], + "angle": 0, + "content": "30. Kim, J., Jeong, S., Sohn, K.: PASTS: Toward effective distilling transformer for panoramic semantic segmentation. In: ICIP (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "31. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.704 + ], + "angle": 0, + "content": "32. Li, J., Chen, P., Qian, S., Jia, J.: TagCLIP: Improving discrimination ability of open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.07547 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.746 + ], + "angle": 0, + "content": "33. Li, X., Wu, T., Qi, Z., Wang, G., Shan, Y., Li, X.: SGAT4PASS: Spherical geometry-aware transformer for panoramic semantic segmentation. In: IJCAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "34. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "35. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Mar-culescu, D.: Open-vocabulary semantic segmentation with mask-adapted CLIP. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Ling, Z., Xing, Z., Zhou, X., Cao, M., Zhou, G.: PanoSwin: A pano-style swim transformer for panorama understanding. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.524, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Liu, Y., Ge, P., Liu, Q., Huang, D.: Multi-grained cross-modal alignment for learning open-vocabulary semantic segmentation from text supervision. arXiv preprint arXiv:2403.03707 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "38. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.245 + ], + "angle": 0, + "content": "39. Ma, C., Yang, Y.H., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: BMVC (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "40. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.787, + 0.328 + ], + "angle": 0, + "content": "41. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: DensePASS: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: ITSC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.786, + 0.356 + ], + "angle": 0, + "content": "42. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.383 + ], + "angle": 0, + "content": "43. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "44. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H.S., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "45. Oin, J., Wu, J., Yan, P., Li, M., Ren, Y., Xiao, X., Wang, Y., Wang, R., Wen, S., Pan, X., Wang, X.: FreeSeg: Unified, universal and open-vocabulary image segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.786, + 0.494 + ], + "angle": 0, + "content": "46. Orhan, S., Bastanlar, Y.: Semantic segmentation of outdoor panoramic images. SIVP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "47. Orsic, M., Kreso, I., Bevandic, P., Segvic, S.: In defense of pre-trained ImageNet architectures for real-time semantic segmentation of road-driving images. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.563 + ], + "angle": 0, + "content": "48. Poudel, R.P.K., Liwicki, S., Cipolla, R.: Fast-SCNN: Fast semantic segmentation network. In: BMVC (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "49. Radford, A., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "50. Ray, B., Jung, J., Larabi, M.: A low-complexity video encoder for equirectangular projected 360 video content. In: ICASSP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.646 + ], + "angle": 0, + "content": "51. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: PanoFormer: Panorama transformer for indoor \\(360^{\\circ}\\) depth estimation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "52. Shi, H., Zhou, Y., Yang, K., Yin, X., Wang, Z., Ye, Y., Yin, Z., Meng, S., Li, P., Wang, K.: PanoFlow: learning \\(360^{\\circ}\\) optical flow for surrounding temporal understanding. T-ITS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.729 + ], + "angle": 0, + "content": "53. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. arXiv preprint arXiv:2306.13631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "54. Teng, Z., Zhang, J., Yang, K., Peng, K., Shi, H., Reifs, S., Cao, K., Stiefelhagen, R.: 360BEV: Panoramic semantic mapping for indoor bird's-eye view. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "56. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "J. Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "57. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., Wang, X., Qiao, Y.: InternImage: Exploring large-scale vision foundation models with deformable convolutions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "58. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. arXiv preprint arXiv:2307.00764 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "59. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: OV-PARTS: Towards open-vocabulary part segmentation. arXiv preprint arXiv:2310.05107 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "60. Wu, C., Zheng, J., Pfrommer, J., Beyerer, J.: Attention-based point cloud edge sampling. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "61. Wysocza'nska, M., Ramamonjisoa, M., Trzci'nski, T., Simeoni, O.: CLIP-DIY: CLIP dense inference yields open-vocabulary semantic segmentation for-free. arXiv preprint arXiv:2309.14289 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "62. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.384 + ], + "angle": 0, + "content": "63. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: DAT++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "64. Xie, B., Cao, J., Xie, J., Khan, F.S., Pang, Y.: SED: A simple encoder-decoder for open-vocabulary semantic segmentation. arXiv preprint arXiv:2311.15537 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.466 + ], + "angle": 0, + "content": "65. Xiong, Y., Li, Z., Chen, Y., Wang, F., Zhu, X., Luo, J., Wang, W., Lu, T., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications. arXiv preprint arXiv:2401.06197 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "66. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "67. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "68. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: SAN: Side adapter network for open-vocabulary semantic segmentation. TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "69. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "70. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "71. Xu, X., Xiong, T., Ding, Z., Tu, Z.: MasQCLIP for open-vocabulary universal image segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "72. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Huang, X., Sun, D., Wang, K.: Can we PASS beyond the field of view? Panoramic annular semantic segmentation for real-world surrounding perception. In: IV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "73. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Wang, K.: PASS: Panoramic annular semantic segmentation. T-ITS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "74. Yang, K., Hu, X., Fang, Y., Wang, K., Stiefelhagen, R.: Omnisupervised omnidirectional semantic segmentation. T-ITS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "75. Yang, K., Hu, X., Stiefelhagen, R.: Is context-aware CNN ready for the surroundings? Panoramic semantic segmentation in the wild. TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "76. Yang, K., Zhang, J., Reif, S., Hu, X., Stiefelhagen, R.: Capturing omni-range context for omnidirectional segmentation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "77. Yang, M., Yu, K., Zhang, C., Li, Z., Yang, K.: DenseASPP for semantic segmentation in street scenes. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.524, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Panoramic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "78. Yin, W., Liu, Y., Shen, C., Hengel, A.v.d., Sun, B.: The devil is in the labels: Semantic segmentation from sentences. arXiv preprint arXiv:2202.02002 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "79. Yu, H., He, L., Jian, B., Feng, W., Liu, S.: PanelNet: Understanding 360 indoor environment via panel representation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.785, + 0.231 + ], + "angle": 0, + "content": "80. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Gao, J., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "81. Zhang, J., Ma, C., Yang, K., Roitberg, A., Peng, K., Stiefelhagen, R.: Transfer beyond the field of view: Dense panoramic semantic segmentation via unsupervised domain adaptation. T-ITS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.785, + 0.313 + ], + "angle": 0, + "content": "82. Zhang, J., Yang, K., Ma, C., Reif, S., Peng, K., Stiefelhagen, R.: Bending reality: Distortion-aware transformers for adapting to panoramic semantic segmentation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.355 + ], + "angle": 0, + "content": "83. Zhang, J., Yang, K., Shi, H., Reif, S., Peng, K., Ma, C., Fu, H., Torr, P.H.S., Wang, K., Stiefelhagen, R.: Behind every domain there is a shift: Adapting distortion-aware vision transformers for panoramic semantic segmentation. TPAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "84. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.785, + 0.424 + ], + "angle": 0, + "content": "85. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Materobot: Material recognition in wearable robotics for people with visual impairments. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.785, + 0.466 + ], + "angle": 0, + "content": "86. Zheng, X., Pan, T., Luo, Y., Wang, L.: Look at the neighbor: Distortion-aware unsupervised domain adaptation for panoramic semantic segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.508 + ], + "angle": 0, + "content": "87. Zheng, X., Zhu, J., Liu, Y., Cao, Z., Fu, C., Wang, L.: Both style and distortion matter: Dual-path unsupervised domain adaptation for panoramic semantic segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "88. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentaion. arXiv preprint arXiv:2311.03352 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "89. Zhou, Q., Liu, Y., Yu, C., Li, J., Wang, Z., Wang, F.: LMSeg: Language-guided multi-dataset segmentation. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.785, + 0.605 + ], + "angle": 0, + "content": "90. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable ConvNets V2: More deformable, better results. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.605 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_origin.pdf b/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a94489d7cd2fcc20bd51e0454027573c33514a80 --- /dev/null +++ b/2024/Open Panoramic Segmentation/ca3f58a1-ca6a-495a-a62e-1bbf8e57e89a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d35fb2e914e97e261719e27ba02bc74802ebe84c051cb0a469a5b7047d65b4c +size 9437329 diff --git a/2024/Open Panoramic Segmentation/full.md b/2024/Open Panoramic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ccd0504ea60bda2ab1720a4353333e3ed748e7b --- /dev/null +++ b/2024/Open Panoramic Segmentation/full.md @@ -0,0 +1,330 @@ +# Open Panoramic Segmentation + +Junwei Zheng $^{1}$ , Ruiping Liu $^{1}$ , Yufan Chen $^{1}$ , Kunyu Peng $^{1}$ , Chengzhi Wu $^{1}$ , Kailun Yang $^{2}$ , Jiaming Zhang $^{1,\dagger}$ , and Rainer Stiefelhagen $^{1}$ + +$^{1}$ Karlsruhe Institute of Technology + $^{2}$ Hunan University + +Abstract. Panoramic images, capturing a $360^{\circ}$ field of view (FoV), encompass omnidirectional spatial information crucial for scene understanding. However, it is not only costly to obtain training-sufficient dense-annotated panoramas but also application-restricted when training models in a close-vocabulary setting. To tackle this problem, in this work, we define a new task termed Open Panoramic Segmentation (OPS), where models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain, enabling the zero-shot open panoramic semantic segmentation ability of models. Moreover, we propose a model named OOOPS with a Deformable Adapter Network (DAN), which significantly improves zero-shot panoramic semantic segmentation performance. To further enhance the distortion-aware modeling ability from the pinhole source domain, we propose a novel data augmentation method called Random Equirectangular Projection (RERP) which is specifically designed to address object deformations in advance. Surpassing other state-of-the-art open-vocabulary semantic segmentation approaches, a remarkable performance boost on three panoramic datasets, WildPASS, Stanford2D3D, and Matterport3D, proves the effectiveness of our proposed OOOPS model with RERP on the OPS task, especially $+2.2\%$ on outdoor WildPASS and $+2.4\%$ mIoU on indoor Stanford2D3D. The source code is publicly available at OPS. + +# 1 Introduction + +Panoramic imaging systems [26, 27] have advanced significantly in recent years, which has fostered a wide variety of panoramic vision applications [1, 17]. Due to the comprehensive $360^{\circ}$ Field of View (FoV), omnidirectional panoramas provide more informative visual cues when perceiving surroundings in a broad spectrum of scene understanding tasks [3, 36, 51, 52, 54, 79], enabling a more complete and immersive capture of environmental data, which is crucial for in-depth scene understanding. This wide perspective surpasses the limited scope of pinhole images, significantly enhancing the capability of computer vision systems to perceive and interpret surroundings in a variety of applications. While the benefits of utilizing panoramic images in computer vision applications are apparent compared + +![](images/9944832e53237798be69d5ca51307d5e129be6a789df41bbc1f109e8b19903ae.jpg) +(a) The challenge of wider FoV. Performance (mIoU / open mIoU) drops when FoV increases. + +![](images/e5ee89e04e53003570a9fdc65f0a3771b2b85b1be7edfbcc37974a025792c7b6.jpg) +(b) Close-vocabulary vs. Open-vocabulary segmentation on panoramas. While limited categories are recognized in a close-vocabulary setting, all objects in $360^{\circ}$ can be covered in an open-vocabulary setting. + +![](images/1dbbd406fe1a711b5b77fb9494ea9d1c15795e07be5c8e8e54e96fef582feb1f.jpg) +(c) The paradigm of Open Panoramic Segmentation (OPS) task. +Fig. 1: (a) The challenge of existing state-of-the-art segmentation models. (b) The limitation of categories in traditional close-vocabulary panoramic segmentation tasks. (c) Our newly defined Open Panoramic Segmentation (OPS) task aims at tackling the above challenges. OPS consists of three important elements: Open the FoV targeted at the challenge of $360^{\circ}$ FoV, Open the Vocabulary targeted at the drawback of close-vocabulary panoramic segmentation and Open the Domain targeted at the challenge of scarcity of panoramic labels. + +with pinhole images, it is imperative to give increasing consideration to some noteworthy challenges as follows: (1) The challenge of wider FoV. For example, Fig. 1a illustrates the degraded performance of state-of-the-art open-vocabulary semantic segmentation approaches [9,68,80] with an increasing FoV, i.e., from pinhole images to $360^{\circ}$ panoramic images. Over $12\%$ mIoU performance degradation can be observed, indicating the challenges introduced by the large divergence of semantic and structural information across narrow-wide imagery. (2) The limitation of categories. The traditional close-vocabulary segmentation task paradigm [74,82] only provides a limited amount of labeled categories, which cannot handle the inestimable amount of categories in real-world applications. Fig. 1b illustrates the difference between close- and open-vocabulary segmentation. Compared to the close-vocabulary setting (first row) the open-vocabulary setting (second row) is not limited by the number of categories of the datasets. Only four predefined categories (highlighted in different colors) are recognized in the close-vocabulary setting in contrast to the open one, where every pixel of the panorama has its own semantic meaning even though the categories are not annotated in the dataset. + +To further release the great potential of panoramic images three critical problems are eager to be solved: How to obtain a holistic perception via a single image? How to break the barrier of limited recognizable categories in existing panoramic datasets so that downstream vision applications can benefit from unrestricted informative visual hints? How to deal with the scarcity of + +panoramic labels? Based on the aforementioned three questions, we propose a new task named Open Panoramic Segmentation (OPS) that comprehensively addresses these challenges, aiming to better leverage the advantages brought by panoramic images. The new task paradigm is shown in Fig. 1c. The open panoramic segmentation task consists of three important elements considering the three problems. The concept of "Open" is three-fold: omnidirectional panoramic images (Open the FoV), an unrestricted range of recognizable classes (Open the Vocabulary), and additionally, models are trained using pinhole images in the source domain while evaluated with panoramic images in the target domain (Open the Domain). Given that densely annotated pinhole segmentation labels incur lower costs compared to panoramic ones [74], it is cost-efficient to open different domains. Note that OPS is different from Domain Adaptation (DA). In the context of training in DA, the utilization of data encompasses both the source and target domains while in OPS, exclusive reliance is placed on data originating solely from the source domain during the whole training process. + +Apart from the new task, we propose a new model named OOOPS to address the aforementioned three openness-related challenges in the OPS task. It consists of a frozen CLIP model [49] and a key component named Deformable Adapter Network (DAN), which serves two critical purposes: (1) efficiently adapting the frozen CLIP model to the panoramic segmentation task and (2) addressing object deformation and image distortion inherent in panoramas. More specifically, a novel Deformable Adapter Operator (DAO), the key component of DAN, is designed to tackle panoramic distortion. To advance the modeling capabilities for distortion awareness in the pinhole source domain, we further introduce Random Equirectangular Projection (RERP) explicitly crafted to tackle object deformations and image distortions. A pinhole image is divided into four image patches that are randomly shuffled. Equirectangular Projection, one of the most common methods for mapping a globe into a panoramic plane, is then introduced in the shuffled image. Our OOOPS model with RERP outperforms other state-of-the-art open-vocabulary segmentation methods with $+2.2\%$ , $+2.4\%$ , and $+0.6\%$ mIoU on WildPASS [76], Stanford2D3D [2], and Matterport3D [6], respectively. + +To summarize, we present the following contributions: + +- We introduce a new task termed open panoramic segmentation, OPS for short, including Open FoV, Open Vocabulary and Open Domain. Models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain. +- We put forward a model called OOOPS with the aim to address three openness-related challenges at once. A Deformable Adapter Network (DAN) is proposed to transfer the zero-shot learning ability of a frozen CLIP model from the pinhole domain to different panoramic domains. +- A novel data augmentation strategy named Random Equirectangular Projection (RERP), which is specifically designed for the proposed OPS + +task, further boosts the accuracy of the OOOPS model, achieving state-of-the-art performance in the open panoramic segmentation task. + +- For benchmarking OPS, we conduct comprehensive evaluations on both indoor and outdoor datasets (WildPASS, Stanford2D3D and Matterport3D), involving more than 10 close- and open-vocabulary segmentation models. + +# 2 Related Work + +# 2.1 Open-vocabulary Semantic Segmentation + +Motivated by the success of vision-language pre-training models, Open-Vocabulary Semantic Segmentation (OVSS) has been recently explored to achieve open-set pixel-wise understanding [7, 11, 12, 18, 19, 29, 31, 32, 34, 34, 37, 39, 40, 44, 53, 56, 58, 59, 61, 64, 70, 71]. Most works leverage the open-vocabulary capabilities of CLIP [49] to reason about relations between dense pixel semantics and arbitrary class labels [9, 12, 32, 38, 61, 71]. In [78, 89], multiple datasets are unified by transferring heterogeneous labels into text embeddings for training to achieve OVSS. In [35, 67], image-caption datasets are employed for learning alignment between visual regions and text entities. To reduce the computational costs of using a frozen pre-trained CLIP model for identifying novel classes, latest studies [21, 69] have also proposed efficient architectures for fast OVSS. SAN [68, 69] has designed an asymmetric input resolution strategy to decouple mask proposal prediction and attention bias prediction applied to CLIP for mask class recognition. Furthermore, open-vocabulary panoptic segmentation has been tackled in [8, 12, 13, 34, 45, 66, 66, 71, 80] to render universal image understanding. In this work, we focus on the open-vocabulary zero-shot learning on panoramas instead of pinhole images discussed in the aforementioned works. + +# 2.2 Panoramic Semantic Segmentation + +Panoramic semantic segmentation enables more informative decision-making and interaction within complex and dynamic environments, advancing the capabilities of intelligent systems for navigation and perception [4, 16, 24, 25, 28, 33, 42, 46, 75]. PASS [72, 73] achieves panoramic segmentation on annular images, whereas 360BEV [54] tackles panoramic semantic mapping. ECANets [76] address omnidirectional semantic segmentation with multi-source omni-supervised learning to enhance generalization. Zhang et al. [81] propose to use attention-augmented designs for panoramic representation learning. Guttikonda et al. [20] propose a single-frame panoramic semantic segmentation approach using multimodal spherical images. Hu et al. [22] propose a distortion convolution module for panoramic semantic segmentation. An important cluster of works revisits panoramic semantic segmentation from an unsupervised domain adaptation perspective [23,30,41,81,82,86,87], by adapting from label-rich pinhole images to the panoramic images. Unlike these works, we propose a method to tackle open-vocabulary panoramic semantic segmentation for unconstrained surrounding understanding. + +# 2.3 Deformable Neural Networks + +In the realm of deformable neural networks, notable advancements have been made through various iterations. The pioneering work [10] introduces the concept of deformable convolution, laying the foundation for subsequent developments. Building upon this foundation, DCNv2 [90] further refines deformable convolution by introducing a modulation mechanism that expands the scope of deformation modeling. The evolution of deformable neural networks continues with the introduction of DCNv3 [57], which combines the group convolution with DCNv2 for the sake of imitating multi-head mechanism of Transformer [55]. DCNv4 [65] aims to speedup DCNv3, achieving a more efficient model. Additionally, Trans4PASS [82,83] proposes Deformable Patch Embedding (DPE) and Deformable MLP (DMLP) to handle the distortion that occurs in panoramas. Incorporating the principles of deformable convolution into vision transformers [14], Xia et al. propose DeformableViT [62,63], extending the applicability of deformable neural networks to the domain of transformer architectures. After rethinking a series of DCN designs, we propose DAO to improve the model's ability to handle image distortion in the OPS task. + +# 3 Methodology + +# 3.1 Open Panoramic Segmentation + +Open Panoramic Segmentation (OPS) task is proposed to deal with three challenging problems: ① the narrow FoV, ② the restricted range of classes, and ③ the scarcity of panoramic labels. OPS gives three answers to the aforementioned issues: ① Open the FoV, ② Open the Vocabulary, and ③ Open the Domain. The OPS task paradigm is shown in Fig. 1c. Models are trained in the narrow-FoV pinhole source domain in an open-vocabulary setting while evaluated in the wide-FoV panoramic target domain. + +# 3.2 Model Architecture + +A foundation model [49] can be efficiently transferred to downstream tasks by using an adapter [68]. To improve the modeling ability for panoramas, we design the OOOPS model. As shown in Fig. 2, it comprises a frozen CLIP model and a proposed Deformable Adapter Network (DAN) that incorporates multiple Transformer Layers and the novel DAO, which will be detailed in Sec. 3.3. Feature fusion takes place between the intermediate layers of CLIP and DAN. One of the two outputs from DAN consists of mask proposals, while the other serves as deep supervised guidance for CLIP to generate proposal logits. In the training phase, pinhole images are forwarded into OOOPS, generating mask proposals and proposal logits for the loss calculation. In the inference phase, panoramas are forwarded into OOOPS to generate segmentation predictions via the multiplication of mask proposals and the corresponding proposal logits. The frozen CLIP is necessary for the zero-shot learning ability of OOOPS. + +![](images/96240dd66b057b5eea2695ecf8ab59c7653cff69998751351a5c6e81494ba866.jpg) +Fig. 2: Overview of the OOOPS model architecture. It consists of a frozen CLIP model and a Deformable Adapter Network (DAN) which includes Transformer Layers and the proposed DAO. + +# 3.3 Deformable Adapter Network + +The Deformable Adapter Network is a combination of multiple Transformer [55] Layers and the proposed DAO. Since distortion exists in panoramic images, which is a big challenge when utilizing informative panoramas, we delve deeper into the deformable design [57] and sampling methods like APES [60], MateRobot [85], proposing DAO to tackle the image distortions and object deformations of panoramas. To provide a detailed rethinking process of the deformable convolution design, we introduce the preliminary of the DCN series as follows. + +Revisiting DCN Series. The seminal work DCN [10] can enable the traditional CNN with spatial deformation-aware ability. Given a convolutional kernel of $K$ sampling locations, let $\mathbf{w}_k$ and $\mathbf{p}_k$ denote the weight and pre-specified offset for the $k$ -th location, respectively. For example, $K = 9$ and $\mathbf{p}_k \in \{(1,1), \dots, (-1, -1)\}$ defines a $3 \times 3$ convolutional kernel of dilation 1. Let $\mathbf{x}(\mathbf{p})$ and $\mathbf{y}(\mathbf{p})$ denote the features at location $\mathbf{p}$ from the input feature maps $\mathbf{x}$ and output feature maps $\mathbf{y}$ , respectively. DCN is formulated as: + +$$ +\mathbf {y} (\mathbf {p}) = \sum_ {k = 1} ^ {K} \mathbf {w} _ {k} \mathbf {x} \left(\mathbf {p} + \mathbf {p} _ {k} + \Delta \mathbf {p} _ {k}\right), \tag {1} +$$ + +where $\Delta \mathbf{p}_k$ is the learnable offset for the $k$ -th location. Although DCN is capable of capturing spatial deformation, every sampling location is treated equally when calculating the local features. DCNv2 [90] is proposed with an additional term called modulation scalar. Specifically, DCNv2 can be formulated as: + +$$ +\mathbf {y} (\mathbf {p}) = \sum_ {k = 1} ^ {K} \mathbf {w} _ {k} \mathbf {m} _ {k} \mathbf {x} \left(\mathbf {p} + \mathbf {p} _ {k} + \Delta \mathbf {p} _ {k}\right), \tag {2} +$$ + +where $\mathbf{m}_k$ is the learnable modulation scalar for the $k$ -th location. Inspired by Transformer, DCNv3 [57] is proposed with a grouping operation, further boosting the deformation-aware ability of DCNv2. DCNv3 can be addressed with the following formula: + +$$ +\mathbf {y} (\mathbf {p}) = \sum_ {g = 1} ^ {G} \sum_ {k = 1} ^ {K} \mathbf {w} _ {g} \mathbf {m} _ {g k} \mathbf {x} _ {g} \left(\mathbf {p} + \mathbf {p} _ {k} + \Delta \mathbf {p} _ {g k}\right), \tag {3} +$$ + +where $G$ denotes the total number of aggregation groups. DCNv4 [65] is similar to DCNv3, achieving similar performance while significantly reducing the runtime. + +![](images/f06c2a86499b210dcd1eacf22acc69164e03fbc1dd56f82f5f5913404a1e637b.jpg) +Fig. 3: Salient map generation in DAO. + +Deformable Adapter Operator (DAO). When dealing with distortions in panoramas, DCNv3 and DCNv4 fall short of meeting the distortion-aware requirements. This is illustrated by the visualization of Fig. 6 in Sec. 4.5. Therefore, DAO is proposed to tackle the distortion problem in panoramic images using the following expression: + +$$ +\mathbf {y} (\mathbf {p}) = \mathbf {s} (\mathbf {p}) \sum_ {g = 1} ^ {G} \sum_ {k = 1} ^ {K} \mathbf {w} _ {g} \mathbf {m} _ {g k} \mathbf {x} _ {g} \left(\mathbf {p} + \mathbf {p} _ {k} + \Delta \mathbf {p} _ {g k}\right), \tag {4} +$$ + +where $\mathbf{s}(\mathbf{p})$ is the learnable salient scalar at location $\mathbf{p}$ . Inheriting from DCNv3, DAO comes up with an additional term called salient scalar indicating the importance of each pixel within the whole panorama. It's worth noting that DCNv3 and DCNv4 share the same mathematical expression, but DCNv3 is more robust according to our experiments before designing DAO. Therefore, we adopt DCNv3 as a part of DAO rather than DCNv4. Illustrated in Fig. 2, the feature map outputted by DCNv3 passes through the Patch Similarity Layer, Normalization Layer, and Standard Deviation Layer sequentially to form a salient map. The intuition for such a design is straightforward. The salient pixels in an image are those that differ significantly from their neighboring pixels, e.g., edge pixels. If all pixels within an image patch are different, the standard deviation of the + +![](images/8357fdae817e0482fcaef067b2002045730139dec0591a682b911a14d7cbaa62.jpg) + +![](images/a89652183d1d5894817bbf5376ab43d72310c9add4d9018828b1b24a11131f62.jpg) + +![](images/34f27cedfb3013359b313cdd70477ac0779ffcfc05bbdeb66082d67e9348efbb.jpg) +(a) Equirectangular Projection (ERP) maps a globe into a panorama. It's clear that strong distortion occurs in the panorama, especially on object edges in the horizontal direction. + +![](images/9df6f7374ef486d09cff92e1638f00aae16e2a5daab186c82eef4380a1420a81.jpg) + +![](images/32aeedbff43f2742c95dc8624148e698b2877a25b42a6214db3d2f7f1cfeff71.jpg) + +![](images/e378c1784ab7463c61bbff1685d866a257219ea04f85e8641d388a8382f2d5a6.jpg) +Fig. 4: (a) Visualization of ERP on a panoramic image and (b) Our proposed RERP on pinhole images. + +![](images/0eaf5c01de9eeed8815723df3ff32c6c0732a6171a46e45f37ba7a7b821e454f.jpg) +(b) Visualization of ERP and RERP of pinhole images. First column: pinhole images. Second column: ERP on pinhole images. Third column: RERP on pinhole images. + +![](images/4b814a83bf59069de5a5b863c893dfb42b267022716d7216b0ba6491ee1b7c17.jpg) + +![](images/12937113d2c42facd50178b70956f57af3e93762ce4e218f71ec156d35af2ff5.jpg) + +pixel similarity for this patch is higher than the one containing similar pixels, resulting in a higher salient scalar. Fig. 3 gives a more detailed explanation of the salient map generation. Given a feature map, DAO first calculates the cosine similarity between the center pixel and all pixels within a kernel, e.g., 9 pixels within a $3 \times 3$ kernel in Fig. 3, resulting in a 9-dimensional cosine similarity vector. Softmax normalization is then applied to the vector. Afterward, DAO calculates the standard deviation of this normalized cosine similarity vector, indicating the importance of the center pixel. By traversing every pixel across the entire feature map, a salient map is generated for the sake of enhancing the salient pixels, which are usually the edge pixels of an image where strong panoramic distortion occurs. + +# 3.4 Random Equirectangular Projection + +The Equirectangular Projection [50] (ERP) is one of the most common methods for mapping a globe into a panoramic plane, which transforms spherical coordinates into planar coordinates expressed as follows: + +$$ +x = R \left(\lambda - \lambda_ {0}\right) \cos \left(\varphi_ {1}\right), \tag {5} +$$ + +$$ +y = R \left(\varphi - \varphi_ {0}\right), \tag {6} +$$ + +where $\lambda$ and $\varphi$ are the longitude and latitude of the location to project, respectively. $\varphi_{1}$ are the standard parallels. $\lambda_0$ and $\varphi_0$ are the central meridian and central parallel of the map, respectively. $R$ is the radius of the globe. $x$ denotes the horizontal coordinate, while $y$ represents the vertical coordinate of the projected location on the map. Fig. 4a visualizes the ERP on a panoramic plane. It can be observed that after equirectangular projection strong distortion occurs in the panorama, e.g., a straight line is transformed into a curve. To further boost the performance, we propose Random Equirectangular Projection (RERP) on pinhole images since the OPS task requires the model to train with pinhole images instead of panoramas. + +We divide the pinhole image into four parts and randomly shuffle the image patches. Afterward, we apply equirectangular projection to the distortion-free pinhole image. Fig. 4b visualizes pinhole images after RERP. The first column is the pinhole images without any data augmentation. The second column is ERP on pinhole images without random shuffling. The last column is our proposed RERP. It can be observed that the panorama-like distortion also occurs in the pinhole images after RERP. Random shuffling is employed to enhance robustness and promote generalization. + +# 4 Experiments + +# 4.1 Datasets + +We train our model using the pinhole general-object COCOStuff-164k [5] while evaluating the model performance on three different panoramic datasets: panoramic street-view WildPASS [76], panoramic indoor-scene Stanford2D3D [2] and panoramic indoor-scene Matterport3D [6]. + +COCOStuff-164k. The pinhole dataset comprises 164k images across 171 annotated classes, partitioned into training, validation, and test sets, with 118k, 5k, and 41k images, respectively. In our experiments, we utilize the entire 118k-image training set as the training data by default. + +WildPASS. The panoramic street-view dataset contains 2,500 panoramas in 65 cities from all continents Asia, Europe, Africa, Oceania, and North and South America (excluding only Antarctica) with 8 dense-annotated classes. All images have a $70^{\circ} \times 360^{\circ}$ FoV in the size of $400 \times 2048$ . + +Stanford2D3D. The panoramic indoor dataset consists of 1,413 images in the size of $512 \times 1024$ and 13 object classes. All results are averaged over 3 cross-validation folds. + +Matterport3D. The panoramic indoor dataset includes 10,800 panoramic views from 194,400 RGB-D images. According to the data preprocessing pipeline of 360BEV [54], a subset of 772 images in the size of $512 \times 1024$ with 20 classes is used to evaluate the segmentation performance. + +# 4.2 Metrics + +Close-vocabulary Metric. Mean Intersection over Union (mIoU) is used as the evaluation metric in close-vocabulary semantic segmentation tasks. Specifically, IoU is calculated as the area of overlap between the predicted segmentation and the ground truth, divided by the area of union between them. mIoU is a hard metric designed for close-vocabulary segmentation ignoring the semantic meanings of similar classes referring to the open-vocabulary setting. + +Open-vocabulary Metric. Open mean Intersection over Union [88], open mIoU for short, is a soft metric compared to mIoU since the similarity scores between different classes are also taken into account, which are calculated by the WordNet approach [43]. In this work, we will report both close and open metrics for a comprehensive understanding of all models. + +Table 1: Results on WildPASS. The mIoU, open mIoU, and per-class IoU are in percentage (%). #PARAMs is the number of learnable parameters in millions. + +
Network#PARAMSmIoUOpen mIoUCarRoadSidewalkCrosswalkCurbpersonTruckBus
Close-Vocabulary
Fast-SCNN [48]1.124.8-45.960.031.79.717.16.014.213.8
PSPNet18 [84]17.528.2-58.266.828.413.019.26.218.215.6
SwiftNet [47]11.830.0-55.764.129.216.222.98.521.122.2
DenseASPP [77]8.333.2-51.169.138.116.426.38.727.428.4
ERF-PSPNet [73]2.534.0-66.370.536.56.424.19.426.532.0
PSPNet50 [84]53.346.1-80.074.951.723.931.419.838.948.1
DANet [15]47.447.2-74.872.249.928.923.825.451.950.6
OOSS (ERF-PSPNet) [74]2.556.1-87.279.360.828.038.154.548.852.2
PASS (ERF-PSPNet) [73]2.564.7-87.380.061.471.149.972.237.557.9
ERF-PSPNet (omni-sup') [76]2.566.8-90.582.765.670.551.558.262.053.1
ECANet (with attention) [76]2.667.7-90.483.768.467.852.161.454.563.3
ECANet (with fusion) [76]2.669.0-90.685.768.067.952.166.059.362.3
Open-Vocabulary
SAN [68]8.455.656.984.381.969.610.12.872.254.969.1
CAT-seg [9]59.555.757.184.181.570.010.23.072.554.869.3
OpenSeeD [80]65.455.857.284.381.669.710.33.073.654.969.1
OOOPS (ours, w/o RERP)8.757.058.584.981.668.715.18.471.355.370.4
OOOPS (ours, w/ RERP)8.758.059.784.981.668.719.48.373.854.972.2
+ +# 4.3 Implementation Details + +We utilize CLIP ViT-B/16 as the frozen CLIP model. All models are trained with the pinhole general-object COCOStuff-164k training set with $4 \times \mathrm{A}40$ GPUs while evaluated on three panoramic datasets which are in totally different domains compared with the pinhole dataset. The AdamW optimizer is applied with an initial learning rate of 0.0001, weight decay of 0.0001, batch size of 32, and a total of 60,000 training iterations. Throughout the training, the learning rate follows a polynomial schedule with a power of 0.9. Data augmentations, including random image resizing within the short-side range of [320, 1024] and a crop size of $640 \times 640$ are applied. Additionally, when verifying the effectiveness of the proposed RERP, it is applied as an extra data augmentation. Following the practice of SAN [68], apart from the supervision of mask recognition with cross-entropy loss, we utilize the dice loss and binary cross-entropy loss for mask proposal generation. + +# 4.4 Comparison with the State of the Art + +Results on WildPASS. Comparisons with both close- and open-vocabulary state-of-the-art methods on WildPASS are listed in Table 1. It is not surprising that ECANet (with fusion) [76] achieves the best performance among all methods since it is a close-vocabulary method, meaning that ECANet is only able to predict the classes predefined by the dataset. However, our OOOPS model is qualified for predicting classes that are not annotated in the dataset. The OOOPS method (with RERP) surpasses other state-of-the-art open-vocabulary panoramic segmentation methods with over $+2.2\%$ mIoU while introducing + +Fig. 5: (a) Comparison on the WildPASS dataset and (b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings. +![](images/d5d995f006c1007f0e6a53716257988c635e6044cfc30b8070d4c849eec7e24f.jpg) +(a) Comparison with baseline method SAN [68] on the WildPASS dataset with predefined 8 classes. First row: ground truth. Second row: prediction from SAN. Third row: prediction from OOOPS with RERP. + +![](images/6d1a70ea15719e6c11daca8a1148376fe15be37ed7e72f6e7e884c63a3a9c1c0.jpg) +(b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings. First row: RGB. Second row: prediction with close vocabularies. Third row: prediction with open vocabularies. Zoom in for a better view. + +only a minor increment of 0.3 million learnable parameters compared to the lightweight baseline SAN [68], offering a promising step towards bridging the gap between close- and open-vocabulary settings. Moreover, our OOOPS method enhanced by RERP demonstrates significant advancements in challenging categories such as crosswalk and curb, achieving $+9.3\%$ and $+5.5\%$ mIoU compared with the baseline SAN, respectively. It is worth noting that in some categories, e.g., sidewalk, person, and bus, our proposed OOOPS (both with and without RERP) even outperforms the models trained in a close-vocabulary setting. All open-vocabulary methods achieve better performance with open mIoU metric compared to the close mIoU since similar semantic meanings are also considered in the calculation. Fig. 1a showcases the performance degradation on both metrics. Even though all methods have degraded performance when FoV increases, our OOOPS achieves a smaller drop $(-10.0\%)$ while others have a larger performance decrease $(-12.0\%)$ , indicating the effectiveness of the OOOPS model. + +Fig. 5a visualizes the prediction of the challenging curb category from baseline SAN and our proposed OOOPS with RERP. From top to bottom are the ground truth, prediction from SAN, and prediction from OOOPS with RERP, respectively. The baseline SAN cannot recognize the challenging curb category while our proposed OOOPS with RERP is capable of detecting the curb in the middle of the panorama highlighted in gray. Fig. 5b showcases the visualization of the prediction from OOOPS on WildPASS in close- and open-vocabulary settings. From top to bottom are input RGB, prediction from OOOPS with 8 predefined close vocabularies from WildPASS, and prediction from OOOPS with unlimited open vocabularies. It is evident that our OOOPS with RERP is capable of tackling unseen categories in the open-vocabulary setting. Every pixel on the panorama is assigned a meaningful semantic label. For example, "mailbox" is not a predefined category from WildPASS. While existing models can not recognize the mailbox under a close-vocabulary setting, the rare classes can be easily detected by our model in an open-vocabulary one, indicating the great potential of open-vocabulary panoramic segmentation for parsing unconstrained traffic surroundings. + +Table 2: Results on Stanford2D3D. The mIoU, open mIoU, and per-class IoU are in percentage (\%). #PARAMs is the number of learnable parameters in millions. + +
Network#PARAMsmIoUOpen mIoUBeamBoardBookcaseCeilingChairClutterColumnDoorFloorSofaTableWallWindow
Close-Vocabulary
Trans4PASS [82]28.252.1-11.460.651.181.859.737.923.554.688.834.653.968.151.0
360BEV [54]28.454.3-06.063.655.582.762.739.916.458.890.845.257.171.157.6
Open-Vocabulary
SAN [68]8.438.539.50.251.539.666.552.110.70.045.049.331.849.250.053.4
CAT-seg [9]59.538.639.60.051.642.366.851.48.00.047.450.731.247.551.453.5
OpenSeeD [80]65.438.740.00.051.842.067.051.68.40.247.550.031.447.151.753.6
OOOPS (w/o RERP)8.739.541.00.052.344.167.153.79.70.050.152.830.547.252.253.8
OOOPS (w/ RERP)8.741.142.60.050.744.868.851.88.10.048.271.133.647.255.454.5
+ +Table 3: Results on Matterport3D. The mIoU, open mIoU, and per-class IoU are in percentage (\%). #PARAMs is the number of learnable parameters in millions. + +
Network≈PARAMsinotOpen mIoUWallFloorChairDoorTablePictureFurnitureObjectsWindowSofbBedSinkStarsCeilingToiletMirrorshowerBathtubCounterShelving
Close-Vocabulary
Trans4PASS+ [83]28.542.6-63.479.139.140.332.836.031.031.537.544.063.220.641.877.640.724.323.758.334.332.9
360BEV [54]28.446.4-64.183.145.845.038.041.132.335.140.648.769.825.147.880.246.028.722.360.138.634.8
Open-Vocabulary
SAN [68]8.430.531.454.476.118.034.122.321.113.01.732.734.251.011.019.974.223.412.031.631.919.028.4
CAT-seg [9]59.530.231.154.374.817.034.121.920.812.61.932.934.050.110.718.873.522.411.631.932.919.228.5
OpenSecD [80]65.430.631.654.276.318.134.322.521.013.02.132.834.150.911.220.274.123.512.231.131.819.128.6
OOOPS (w/o RERP)8.731.132.455.277.017.135.725.017.513.52.333.032.451.812.620.474.522.610.834.136.819.429.7
OOOPS (w/ RERP)8.731.232.555.476.418.434.823.921.213.22.633.334.552.912.120.374.628.212.531.929.319.428.6
+ +Results on Stanford2D3D. In Table 2, the comparison is conducted between state-of-the-art methods on the Stanford2D3D dataset. All reported results are averaged over 3 cross-validation folds. Without RERP, OOOPS achieves $39.5\%$ mIoU, outperforming other open-vocabulary methods like SAN [68], CAT-seg [9], and OpenSeeD [80], which score between $38.5\%$ and $38.7\%$ . With RERP, OOOPS's mIoU increases to $41.1\%$ with a $+2.4\%$ gain, narrowing the gap with advanced close-vocabulary methods. + +Results on Matterport3D. In Table 3, the comparison is conducted between state-of-the-art methods on the Matterport3D dataset. Our OOOPS method showcases consistent performance. Without augmentation, OOOPS achieves mIoU of $31.1\%$ , slightly outperforming other open-vocabulary methods like SAN [68] $(30.5\%)$ , CAT-seg [9] $(30.2\%)$ , and OpenSeeD [80] $(30.6\%)$ . With augmentation, OOOPS's mIoU marginally increases to $31.2\%$ , further reinforcing its lead over these open-vocabulary methods. Both variants of OOOPS still lag behind close-vocabulary methods like Trans4PASS+ [83] and 360BEV [54], which however, are trained in a supervised fashion in a close-vocabulary setting on panoramas and have mIoUs ranging from $42.6\%$ to $46.4\%$ . + +Table 4: Task difference between Domain Adaptation (DA) and the proposed OPS. + +
TaskDomain#ClassesmIoU (%)
SourceTargetWildPASSMatterport3DStanford2D3D
DA✓✓limited72.247.255.4
OPS✓✗unlimited58.031.241.1
+ +(a) Ablation study of OOOPS components on the WildPASS Dataset. The mIoU is reported in percentage $(\%)$ . #PARAMs stands for the number of learnable parameters in millions. + +(b) Ablation study of different numbers of shuffling patches on WildPASS, Matterport3D and Stanford2D3D datasets. The mIoU is reported in percentage $(\%)$ . + +Table 5: Ablation study of OOOPS components and the number of shuffling patches. + +
MethodERPRERP#PARAMsmIoU
SAN [68]✘✘8.455.6
SAN [68]✓✘8.456.0
+ DCN [68]✘✓8.456.5 (+0.9)
+ DCN✘✘8.655.6
+ DCN✓✘8.656.0
+ DCN✘✓8.656.6 (+1.0)
+ DCNv2✘✘8.755.7
+ DCNv2✓✘8.756.4
+ DCNv2✘✓8.756.6 (+0.9)
+ DCNv3✘✘8.755.8
+ DCNv3✓✘8.756.4
+ DCNv3✘✓8.756.8 (+1.0)
+ DCNv4✘✘8.755.8
+ DCNv4✓✘8.756.4
+ DCNv4✘✓8.756.8 (+1.0)
+ DAO (ours)✘✘8.757.0
+ DAO (ours)✓✘8.757.6
+ DAO (ours)✘✓8.758.0 (+1.0)
+ +
mIoU (%) #Patches Dataset4 9 16 25
WildPASS58.0 58.5 58.2 58.0
Matterport3D31.2 31.6 31.4 31.0
Stanford2D3D41.1 41.9 41.5 41.1
+ +Results in DA Subtask. The proposed OPS task differs from Domain Adaptation since no data from the target domain is available to models in OPS while models have access to the data in DA during training. Moreover, models are required to recognize an unlimited amount of classes in OPS while only a limited number of classes can be recognized in DA. DA is a special case of OPS, meaning that models in OPS can be easily adapted to DA and achieve better performance. The converse doesn't hold. Table 4 showcases the task difference between OPS and DA. It can be observed that over $+14\%$ gain in mIoU is achieved when adapting our OOOPS model in the DA subtask. + +# 4.5 Ablation Study + +We replace DAO in OOOPS, as illustrated in Fig. 2, with DCN [10], DCNv2 [90], DCNv3 [57] and DCNv4 [65]. When referring to the component substitution without data augmentation in Table 5a, it is clear that our DAO surpasses other counterparts with a maximum performance gain of $+1.4\%$ mIoU. As for the component substitution with ERP, all models achieve a boost in the range of $+0.3\%$ to $+0.7\%$ compared with the one without data augmentation. The gain is even bigger considering the component substitution with RERP, achieving $+1.0\%$ mIoU. It is worth noting that RERP is able to boost the model performance stably among all models. The best model performance is achieved by the combination of DAO and RERP, yielding $58.0\%$ mIoU with only a slight increment in parameters. Table 5b showcases the influence of different numbers of shuffling patches. The performance is not proportional to the number of patches, indicating the importance of moderate content continuity during training. + +![](images/7d5096a80cb57a68af065098d6f774bc40acd6312e3c2f1dd0f08756f7b19cbb.jpg) +(a) Deformable offsets without RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO. + +![](images/03505442d6f0e8b1e5367307847b1b80fcb74392fefd5aaf81cd3c8d0c69a639.jpg) +(b) Deformable offsets with RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO. +Fig. 6: Visualization of deformable offsets. The green points $\bullet$ are sampling locations. The red points $\bullet$ are deformable offsets in 2 levels, indicating a deformable receptive field (e.g., each level has a $3 \times 3$ kernel size, resulting in $(3 \times 3)^{2} = 81$ red points). + +To further illustrate the deformable ability of DAO and the superiority of RERP when tackling the distortion of panoramas, we visualize the deformable offset in Fig. 6. Fig. 6a presents the deformable offsets without RERP. It can be observed that DAO is better aware of the panoramic distortion compared with the rest in the DCN series. Since the strongest distortion occurs on the edge of objects, DAO focuses more on edge pixels for better distortion-aware feature extraction. The standard deviation of patch similarity in DAO implies the salience of the image patch center. By multiplicatively combining salient scalars and features, significant pixels, such as those representing object edges, are enhanced in their representation. In Fig. 6b, it is evident that the inclusion of RERP improves the distortion information capturing ability of all models in the DCN series, as opposed to their counterparts without RERP. This observation is consistent with the results in Table 5a. The incorporation of both DAO and RERP in OOOPS showcases a remarkable distortion-aware modeling capability. + +# 5 Conclusion + +In this work, we formulate the novel task of Open Panoramic Segmentation (OPS) for unfolding holistic scene understanding, which reaches Open FoV, Open Vocabulary, and Open Domain in surrounding scene segmentation. We have designed the Deformable Adapter Network (DAN) and Random Equirectangular Projection (RERP) for enhancing distortion-aware segmentation capacities of the novel proposed OOOPS model. Extensive experiments on WildPASS, Stanford2D3D, and Matterport3D benchmarks show that the proposed method significantly elevates the performance of state-of-the-art panoramic segmenters. + +# Acknowledgements + +This work was supported in part by the Ministry of Science, Research and the Arts of Baden-Württemberg (MWK) through the Cooperative Graduate School Accessibility through AI-based Assistive Technology (KATE) under Grant BW6-03, in part by BMBF through a fellowship within the IFI programme of DAAD, in part by the Helmholtz Association Initiative and Networking Fund on the HAICORE@KIT and HOREKA@KIT partition, in part by the National Key RD Program under Grant 2022YFB4701400, and in part by Hangzhou SurImage Technology Company Ltd. + +# References + +1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022) +2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017) +3. Athwale, A., Afrasiyabi, A., Lagüie, J., Shili, I., Ahmad, O., Lalonde, J.F.: DarSwin: Distortion aware radial swim transformer. In: ICCV (2023) +4. Berenguel-Baeta, B., Bermudez-Cameo, J., Guerrero, J.J.: FreDSNet: Joint monocular depth and semantic segmentation with fast fourier convolutions from single panoramas. In: ICRA (2023) +5. Caesar, H., Uijlings, J.R.R., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. CVPR (2018) +6. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. 3DV (2017) +7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023) +8. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. arXiv preprint arXiv:2303.11324 (2023) +9. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.: CAT-Seg: Cost aggregation for open-vocabulary semantic segmentation. In: CVPR (2023) +0. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: ICCV (2017) +1. Dao, S.D., Shi, H., Phung, D., Cai, J.: Class enhancement losses with pseudo labels for open-vocabulary semantic segmentation. TMM (2023) +2. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023) +3. Dong, B., Gu, J., Han, J., Xu, H., Zuo, W.: Towards universal vision-language omni-supervised segmentation. arXiv preprint arXiv:2303.06547 (2023) +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +5. Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., Lu, H.: Dual attention network for scene segmentation. In: CVPR (2019) + +16. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhou, X., Geiger, A., Liao, Y.: PanopticNeRF-360: Panoramic 3D-to-2D label transfer in urban scenes. arXiv preprint arXiv:2309.10815 (2023) +17. Gao, S., Yang, K., Shi, H., Wang, K., Bai, J.: Review on panoramic imaging and its applications in scene understanding. TIM (2022) +18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022) +19. Guo, J., Wang, Q., Gao, Y., Jiang, X., Tang, X., Hu, Y., Zhang, B.: MVP-SEG: Multi-view prompt learning for open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.06957 (2023) +20. Guttikonda, S., Rambach, J.: Single frame semantic segmentation using multimodal spherical images. In: WACV (2024) +21. Han, K., Liu, Y., Liew, J.H., Ding, H., Wei, Y., Liu, J., Wang, Y., Tang, Y., Yang, Y., Feng, J., Zhao, Y.: Global knowledge calibration for fast open-vocabulary segmentation. arXiv preprint arXiv:2303.09181 (2023) +22. Hu, X., An, Y., Shao, C., Hu, H.: Distortion convolution module for semantic segmentation of panoramic images based on the image-forming principle. TIM (2022) +23. Jang, S., Na, J., Oh, D.: DaDA: Distortion-aware domain adaptation for unsupervised semantic segmentation. In: NeurIPS (2022) +24. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Towards complete surrounding understanding via unsupervised contrastive learning. In: IV (2021) +25. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Insights into surrounding parsing for mobile agents via unsupervised contrastive learning. T-ITS (2023) +26. Jiang, Q., Gao, S., Gao, Y., Yang, K., Yi, Z., Shi, H., Sun, L., Wang, K.: Minimalist and high-quality panoramic imaging with PSF-aware transformers. TIP (2024) +27. Jiang, Q., Shi, H., Sun, L., Gao, S., Yang, K., Wang, K.: Annular computational imaging: Capture clear panoramic images through simple lens. TCI (2022) +28. Jiang, W., Wu, Y.: DFNet: Semantic segmentation on panoramic images with dynamic loss weights and residual fusion block. In: ICRA (2018) +29. Jiayun, L., Khandelwal, S., Sigal, L., Li, B.: Plug-and-play, dense-label-free extraction of open-vocabulary semantic segmentation from vision-language models. arXiv preprint arXiv:2311.17095 (2023) +30. Kim, J., Jeong, S., Sohn, K.: PASTS: Toward effective distilling transformer for panoramic semantic segmentation. In: ICIP (2022) +31. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) +32. Li, J., Chen, P., Qian, S., Jia, J.: TagCLIP: Improving discrimination ability of open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.07547 (2023) +33. Li, X., Wu, T., Qi, Z., Wang, G., Shan, Y., Li, X.: SGAT4PASS: Spherical geometry-aware transformer for panoramic semantic segmentation. In: IJCAI (2023) +34. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023) +35. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Mar-culescu, D.: Open-vocabulary semantic segmentation with mask-adapted CLIP. In: CVPR (2023) +36. Ling, Z., Xing, Z., Zhou, X., Cao, M., Zhou, G.: PanoSwin: A pano-style swim transformer for panorama understanding. In: CVPR (2023) + +37. Liu, Y., Ge, P., Liu, Q., Huang, D.: Multi-grained cross-modal alignment for learning open-vocabulary semantic segmentation from text supervision. arXiv preprint arXiv:2403.03707 (2024) +38. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023) +39. Ma, C., Yang, Y.H., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: BMVC (2022) +40. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096 (2023) +41. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: DensePASS: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: ITSC (2021) +42. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: ECCV (2022) +43. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995) +44. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H.S., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023) +45. Oin, J., Wu, J., Yan, P., Li, M., Ren, Y., Xiao, X., Wang, Y., Wang, R., Wen, S., Pan, X., Wang, X.: FreeSeg: Unified, universal and open-vocabulary image segmentation. In: CVPR (2023) +46. Orhan, S., Bastanlar, Y.: Semantic segmentation of outdoor panoramic images. SIVP (2021) +47. Orsic, M., Kreso, I., Bevandic, P., Segvic, S.: In defense of pre-trained ImageNet architectures for real-time semantic segmentation of road-driving images. In: CVPR (2019) +48. Poudel, R.P.K., Liwicki, S., Cipolla, R.: Fast-SCNN: Fast semantic segmentation network. In: BMVC (2019) +49. Radford, A., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +50. Ray, B., Jung, J., Larabi, M.: A low-complexity video encoder for equirectangular projected 360 video content. In: ICASSP (2018) +51. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: PanoFormer: Panorama transformer for indoor $360^{\circ}$ depth estimation. In: ECCV (2022) +52. Shi, H., Zhou, Y., Yang, K., Yin, X., Wang, Z., Ye, Y., Yin, Z., Meng, S., Li, P., Wang, K.: PanoFlow: learning $360^{\circ}$ optical flow for surrounding temporal understanding. T-ITS (2023) +53. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. arXiv preprint arXiv:2306.13631 (2023) +54. Teng, Z., Zhang, J., Yang, K., Peng, K., Shi, H., Reifs, S., Cao, K., Stiefelhagen, R.: 360BEV: Panoramic semantic mapping for indoor bird's-eye view. In: WACV (2024) +55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +56. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023) + +57. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., Wang, X., Qiao, Y.: InternImage: Exploring large-scale vision foundation models with deformable convolutions. In: CVPR (2023) +58. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. arXiv preprint arXiv:2307.00764 (2023) +59. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: OV-PARTS: Towards open-vocabulary part segmentation. arXiv preprint arXiv:2310.05107 (2023) +60. Wu, C., Zheng, J., Pfrommer, J., Beyerer, J.: Attention-based point cloud edge sampling. In: CVPR (2023) +61. Wysocza'nska, M., Ramamonjisoa, M., Trzci'nski, T., Simeoni, O.: CLIP-DIY: CLIP dense inference yields open-vocabulary semantic segmentation for-free. arXiv preprint arXiv:2309.14289 (2023) +62. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022) +63. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: DAT++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430 (2023) +64. Xie, B., Cao, J., Xie, J., Khan, F.S., Pang, Y.: SED: A simple encoder-decoder for open-vocabulary semantic segmentation. arXiv preprint arXiv:2311.15537 (2023) +65. Xiong, Y., Li, Z., Chen, Y., Wang, F., Zhu, X., Luo, J., Wang, W., Lu, T., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications. arXiv preprint arXiv:2401.06197 (2024) +66. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) +67. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023) +68. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: SAN: Side adapter network for open-vocabulary semantic segmentation. TPAMI (2023) +69. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023) +70. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022) +71. Xu, X., Xiong, T., Ding, Z., Tu, Z.: MasQCLIP for open-vocabulary universal image segmentation. In: ICCV (2023) +72. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Huang, X., Sun, D., Wang, K.: Can we PASS beyond the field of view? Panoramic annular semantic segmentation for real-world surrounding perception. In: IV (2019) +73. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Wang, K.: PASS: Panoramic annular semantic segmentation. T-ITS (2020) +74. Yang, K., Hu, X., Fang, Y., Wang, K., Stiefelhagen, R.: Omnisupervised omnidirectional semantic segmentation. T-ITS (2022) +75. Yang, K., Hu, X., Stiefelhagen, R.: Is context-aware CNN ready for the surroundings? Panoramic semantic segmentation in the wild. TIP (2021) +76. Yang, K., Zhang, J., Reif, S., Hu, X., Stiefelhagen, R.: Capturing omni-range context for omnidirectional segmentation. In: CVPR (2021) +77. Yang, M., Yu, K., Zhang, C., Li, Z., Yang, K.: DenseASPP for semantic segmentation in street scenes. In: CVPR (2018) + +78. Yin, W., Liu, Y., Shen, C., Hengel, A.v.d., Sun, B.: The devil is in the labels: Semantic segmentation from sentences. arXiv preprint arXiv:2202.02002 (2022) +79. Yu, H., He, L., Jian, B., Feng, W., Liu, S.: PanelNet: Understanding 360 indoor environment via panel representation. In: CVPR (2023) +80. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Gao, J., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV (2023) +81. Zhang, J., Ma, C., Yang, K., Roitberg, A., Peng, K., Stiefelhagen, R.: Transfer beyond the field of view: Dense panoramic semantic segmentation via unsupervised domain adaptation. T-ITS (2022) +82. Zhang, J., Yang, K., Ma, C., Reif, S., Peng, K., Stiefelhagen, R.: Bending reality: Distortion-aware transformers for adapting to panoramic semantic segmentation. In: CVPR (2022) +83. Zhang, J., Yang, K., Shi, H., Reif, S., Peng, K., Ma, C., Fu, H., Torr, P.H.S., Wang, K., Stiefelhagen, R.: Behind every domain there is a shift: Adapting distortion-aware vision transformers for panoramic semantic segmentation. TPAMI (2024) +84. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017) +85. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Materobot: Material recognition in wearable robotics for people with visual impairments. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024) +86. Zheng, X., Pan, T., Luo, Y., Wang, L.: Look at the neighbor: Distortion-aware unsupervised domain adaptation for panoramic semantic segmentation. In: ICCV (2023) +87. Zheng, X., Zhu, J., Liu, Y., Cao, Z., Fu, C., Wang, L.: Both style and distortion matter: Dual-path unsupervised domain adaptation for panoramic semantic segmentation. In: CVPR (2023) +88. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentaion. arXiv preprint arXiv:2311.03352 (2023) +89. Zhou, Q., Liu, Y., Yu, C., Li, J., Wang, Z., Wang, F.: LMSeg: Language-guided multi-dataset segmentation. In: ICLR (2023) +90. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable ConvNets V2: More deformable, better results. In: CVPR (2019) \ No newline at end of file diff --git a/2024/Open Panoramic Segmentation/images.zip b/2024/Open Panoramic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8576e131d6d048526f37fd631060938fbcdc9e68 --- /dev/null +++ b/2024/Open Panoramic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a8d56dceacfe62e498168b179623fd977dd5c78ca70dd91ccb83070bde36077 +size 574898 diff --git a/2024/Open Panoramic Segmentation/layout.json b/2024/Open Panoramic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0105e5fa2ee6685d06106d6b25db8cbb57196a70 --- /dev/null +++ b/2024/Open Panoramic Segmentation/layout.json @@ -0,0 +1,10450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 198, + 112, + 416, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 112, + 416, + 129 + ], + "spans": [ + { + "bbox": [ + 198, + 112, + 416, + 129 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "spans": [ + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": "Junwei Zheng" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": ", Ruiping Liu" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": ", Yufan Chen" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": ", Kunyu Peng" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": ", Chengzhi Wu" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": ", Kailun Yang" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": ", Jiaming Zhang" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "text", + "content": ", and Rainer Stiefelhagen" + }, + { + "bbox": [ + 197, + 149, + 416, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 233, + 195, + 380, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 195, + 380, + 218 + ], + "spans": [ + { + "bbox": [ + 233, + 195, + 380, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 233, + 195, + 380, + 218 + ], + "type": "text", + "content": " Karlsruhe Institute of Technology \n" + }, + { + "bbox": [ + 233, + 195, + 380, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 233, + 195, + 380, + 218 + ], + "type": "text", + "content": " Hunan University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "text", + "content": "Abstract. Panoramic images, capturing a " + }, + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "text", + "content": " field of view (FoV), encompass omnidirectional spatial information crucial for scene understanding. However, it is not only costly to obtain training-sufficient dense-annotated panoramas but also application-restricted when training models in a close-vocabulary setting. To tackle this problem, in this work, we define a new task termed Open Panoramic Segmentation (OPS), where models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain, enabling the zero-shot open panoramic semantic segmentation ability of models. Moreover, we propose a model named OOOPS with a Deformable Adapter Network (DAN), which significantly improves zero-shot panoramic semantic segmentation performance. To further enhance the distortion-aware modeling ability from the pinhole source domain, we propose a novel data augmentation method called Random Equirectangular Projection (RERP) which is specifically designed to address object deformations in advance. Surpassing other state-of-the-art open-vocabulary semantic segmentation approaches, a remarkable performance boost on three panoramic datasets, WildPASS, Stanford2D3D, and Matterport3D, proves the effectiveness of our proposed OOOPS model with RERP on the OPS task, especially " + }, + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "text", + "content": " on outdoor WildPASS and " + }, + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 160, + 243, + 455, + 485 + ], + "type": "text", + "content": " mIoU on indoor Stanford2D3D. The source code is publicly available at OPS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 504, + 230, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 230, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 230, + 516 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 527, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 648 + ], + "type": "text", + "content": "Panoramic imaging systems [26, 27] have advanced significantly in recent years, which has fostered a wide variety of panoramic vision applications [1, 17]. Due to the comprehensive " + }, + { + "bbox": [ + 130, + 527, + 482, + 648 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 527, + 482, + 648 + ], + "type": "text", + "content": " Field of View (FoV), omnidirectional panoramas provide more informative visual cues when perceiving surroundings in a broad spectrum of scene understanding tasks [3, 36, 51, 52, 54, 79], enabling a more complete and immersive capture of environmental data, which is crucial for in-depth scene understanding. This wide perspective surpasses the limited scope of pinhole images, significantly enhancing the capability of computer vision systems to perceive and interpret surroundings in a variety of applications. While the benefits of utilizing panoramic images in computer vision applications are apparent compared" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 313, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 313, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 313, + 665 + ], + "type": "text", + "content": "† Correspondence: jiaming.zhang@kit.edu" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 305, + 170 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 305, + 170 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 305, + 170 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 305, + 170 + ], + "type": "image", + "image_path": "9944832e53237798be69d5ca51307d5e129be6a789df41bbc1f109e8b19903ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 171, + 306, + 187 + ], + "lines": [ + { + "bbox": [ + 134, + 171, + 306, + 187 + ], + "spans": [ + { + "bbox": [ + 134, + 171, + 306, + 187 + ], + "type": "text", + "content": "(a) The challenge of wider FoV. Performance (mIoU / open mIoU) drops when FoV increases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 188, + 306, + 250 + ], + "blocks": [ + { + "bbox": [ + 134, + 188, + 306, + 250 + ], + "lines": [ + { + "bbox": [ + 134, + 188, + 306, + 250 + ], + "spans": [ + { + "bbox": [ + 134, + 188, + 306, + 250 + ], + "type": "image", + "image_path": "e5ee89e04e53003570a9fdc65f0a3771b2b85b1be7edfbcc37974a025792c7b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 251, + 306, + 292 + ], + "lines": [ + { + "bbox": [ + 133, + 251, + 306, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 306, + 292 + ], + "type": "text", + "content": "(b) Close-vocabulary vs. Open-vocabulary segmentation on panoramas. While limited categories are recognized in a close-vocabulary setting, all objects in " + }, + { + "bbox": [ + 133, + 251, + 306, + 292 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 133, + 251, + 306, + 292 + ], + "type": "text", + "content": " can be covered in an open-vocabulary setting." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 118, + 479, + 267 + ], + "blocks": [ + { + "bbox": [ + 309, + 118, + 479, + 267 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 479, + 267 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 479, + 267 + ], + "type": "image", + "image_path": "1dbbd406fe1a711b5b77fb9494ea9d1c15795e07be5c8e8e54e96fef582feb1f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 270, + 479, + 288 + ], + "lines": [ + { + "bbox": [ + 307, + 270, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 479, + 288 + ], + "type": "text", + "content": "(c) The paradigm of Open Panoramic Segmentation (OPS) task." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 300, + 482, + 378 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 378 + ], + "type": "text", + "content": "Fig. 1: (a) The challenge of existing state-of-the-art segmentation models. (b) The limitation of categories in traditional close-vocabulary panoramic segmentation tasks. (c) Our newly defined Open Panoramic Segmentation (OPS) task aims at tackling the above challenges. OPS consists of three important elements: Open the FoV targeted at the challenge of " + }, + { + "bbox": [ + 130, + 300, + 482, + 378 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 300, + 482, + 378 + ], + "type": "text", + "content": " FoV, Open the Vocabulary targeted at the drawback of close-vocabulary panoramic segmentation and Open the Domain targeted at the challenge of scarcity of panoramic labels." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "text", + "content": "with pinhole images, it is imperative to give increasing consideration to some noteworthy challenges as follows: (1) The challenge of wider FoV. For example, Fig. 1a illustrates the degraded performance of state-of-the-art open-vocabulary semantic segmentation approaches [9,68,80] with an increasing FoV, i.e., from pinhole images to " + }, + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "text", + "content": " panoramic images. Over " + }, + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 130, + 402, + 482, + 605 + ], + "type": "text", + "content": " mIoU performance degradation can be observed, indicating the challenges introduced by the large divergence of semantic and structural information across narrow-wide imagery. (2) The limitation of categories. The traditional close-vocabulary segmentation task paradigm [74,82] only provides a limited amount of labeled categories, which cannot handle the inestimable amount of categories in real-world applications. Fig. 1b illustrates the difference between close- and open-vocabulary segmentation. Compared to the close-vocabulary setting (first row) the open-vocabulary setting (second row) is not limited by the number of categories of the datasets. Only four predefined categories (highlighted in different colors) are recognized in the close-vocabulary setting in contrast to the open one, where every pixel of the panorama has its own semantic meaning even though the categories are not annotated in the dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": "To further release the great potential of panoramic images three critical problems are eager to be solved: How to obtain a holistic perception via a single image? How to break the barrier of limited recognizable categories in existing panoramic datasets so that downstream vision applications can benefit from unrestricted informative visual hints? How to deal with the scarcity of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "panoramic labels? Based on the aforementioned three questions, we propose a new task named Open Panoramic Segmentation (OPS) that comprehensively addresses these challenges, aiming to better leverage the advantages brought by panoramic images. The new task paradigm is shown in Fig. 1c. The open panoramic segmentation task consists of three important elements considering the three problems. The concept of \"Open\" is three-fold: omnidirectional panoramic images (Open the FoV), an unrestricted range of recognizable classes (Open the Vocabulary), and additionally, models are trained using pinhole images in the source domain while evaluated with panoramic images in the target domain (Open the Domain). Given that densely annotated pinhole segmentation labels incur lower costs compared to panoramic ones [74], it is cost-efficient to open different domains. Note that OPS is different from Domain Adaptation (DA). In the context of training in DA, the utilization of data encompasses both the source and target domains while in OPS, exclusive reliance is placed on data originating solely from the source domain during the whole training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "text", + "content": "Apart from the new task, we propose a new model named OOOPS to address the aforementioned three openness-related challenges in the OPS task. It consists of a frozen CLIP model [49] and a key component named Deformable Adapter Network (DAN), which serves two critical purposes: (1) efficiently adapting the frozen CLIP model to the panoramic segmentation task and (2) addressing object deformation and image distortion inherent in panoramas. More specifically, a novel Deformable Adapter Operator (DAO), the key component of DAN, is designed to tackle panoramic distortion. To advance the modeling capabilities for distortion awareness in the pinhole source domain, we further introduce Random Equirectangular Projection (RERP) explicitly crafted to tackle object deformations and image distortions. A pinhole image is divided into four image patches that are randomly shuffled. Equirectangular Projection, one of the most common methods for mapping a globe into a panoramic plane, is then introduced in the shuffled image. Our OOOPS model with RERP outperforms other state-of-the-art open-vocabulary segmentation methods with " + }, + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "inline_equation", + "content": "+0.6\\%" + }, + { + "bbox": [ + 130, + 315, + 482, + 507 + ], + "type": "text", + "content": " mIoU on WildPASS [76], Stanford2D3D [2], and Matterport3D [6], respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 514, + 386, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 514, + 386, + 526 + ], + "spans": [ + { + "bbox": [ + 146, + 514, + 386, + 526 + ], + "type": "text", + "content": "To summarize, we present the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 533, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 533, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 593 + ], + "type": "text", + "content": "- We introduce a new task termed open panoramic segmentation, OPS for short, including Open FoV, Open Vocabulary and Open Domain. Models are trained with FoV-restricted pinhole images in the source domain in an open-vocabulary setting while evaluated with FoV-open panoramic images in the target domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 594, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 481, + 640 + ], + "type": "text", + "content": "- We put forward a model called OOOPS with the aim to address three openness-related challenges at once. A Deformable Adapter Network (DAN) is proposed to transfer the zero-shot learning ability of a frozen CLIP model from the pinhole domain to different panoramic domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 642, + 481, + 665 + ], + "type": "text", + "content": "- A novel data augmentation strategy named Random Equirectangular Projection (RERP), which is specifically designed for the proposed OPS" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "content": "task, further boosts the accuracy of the OOOPS model, achieving state-of-the-art performance in the open panoramic segmentation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 175 + ], + "type": "text", + "content": "- For benchmarking OPS, we conduct comprehensive evaluations on both indoor and outdoor datasets (WildPASS, Stanford2D3D and Matterport3D), involving more than 10 close- and open-vocabulary segmentation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 237, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 237, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 237, + 205 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 217, + 367, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 367, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 367, + 229 + ], + "type": "text", + "content": "2.1 Open-vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 485, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 485, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 485, + 439 + ], + "type": "text", + "content": "Motivated by the success of vision-language pre-training models, Open-Vocabulary Semantic Segmentation (OVSS) has been recently explored to achieve open-set pixel-wise understanding [7, 11, 12, 18, 19, 29, 31, 32, 34, 34, 37, 39, 40, 44, 53, 56, 58, 59, 61, 64, 70, 71]. Most works leverage the open-vocabulary capabilities of CLIP [49] to reason about relations between dense pixel semantics and arbitrary class labels [9, 12, 32, 38, 61, 71]. In [78, 89], multiple datasets are unified by transferring heterogeneous labels into text embeddings for training to achieve OVSS. In [35, 67], image-caption datasets are employed for learning alignment between visual regions and text entities. To reduce the computational costs of using a frozen pre-trained CLIP model for identifying novel classes, latest studies [21, 69] have also proposed efficient architectures for fast OVSS. SAN [68, 69] has designed an asymmetric input resolution strategy to decouple mask proposal prediction and attention bias prediction applied to CLIP for mask class recognition. Furthermore, open-vocabulary panoptic segmentation has been tackled in [8, 12, 13, 34, 45, 66, 66, 71, 80] to render universal image understanding. In this work, we focus on the open-vocabulary zero-shot learning on panoramas instead of pinhole images discussed in the aforementioned works." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 456, + 335, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 335, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 335, + 467 + ], + "type": "text", + "content": "2.2 Panoramic Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Panoramic semantic segmentation enables more informative decision-making and interaction within complex and dynamic environments, advancing the capabilities of intelligent systems for navigation and perception [4, 16, 24, 25, 28, 33, 42, 46, 75]. PASS [72, 73] achieves panoramic segmentation on annular images, whereas 360BEV [54] tackles panoramic semantic mapping. ECANets [76] address omnidirectional semantic segmentation with multi-source omni-supervised learning to enhance generalization. Zhang et al. [81] propose to use attention-augmented designs for panoramic representation learning. Guttikonda et al. [20] propose a single-frame panoramic semantic segmentation approach using multimodal spherical images. Hu et al. [22] propose a distortion convolution module for panoramic semantic segmentation. An important cluster of works revisits panoramic semantic segmentation from an unsupervised domain adaptation perspective [23,30,41,81,82,86,87], by adapting from label-rich pinhole images to the panoramic images. Unlike these works, we propose a method to tackle open-vocabulary panoramic semantic segmentation for unconstrained surrounding understanding." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "type": "text", + "content": "2.3 Deformable Neural Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 327 + ], + "type": "text", + "content": "In the realm of deformable neural networks, notable advancements have been made through various iterations. The pioneering work [10] introduces the concept of deformable convolution, laying the foundation for subsequent developments. Building upon this foundation, DCNv2 [90] further refines deformable convolution by introducing a modulation mechanism that expands the scope of deformation modeling. The evolution of deformable neural networks continues with the introduction of DCNv3 [57], which combines the group convolution with DCNv2 for the sake of imitating multi-head mechanism of Transformer [55]. DCNv4 [65] aims to speedup DCNv3, achieving a more efficient model. Additionally, Trans4PASS [82,83] proposes Deformable Patch Embedding (DPE) and Deformable MLP (DMLP) to handle the distortion that occurs in panoramas. Incorporating the principles of deformable convolution into vision transformers [14], Xia et al. propose DeformableViT [62,63], extending the applicability of deformable neural networks to the domain of transformer architectures. After rethinking a series of DCN designs, we propose DAO to improve the model's ability to handle image distortion in the OPS task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 345, + 233, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 233, + 359 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 233, + 359 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 369, + 317, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 317, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 317, + 382 + ], + "type": "text", + "content": "3.1 Open Panoramic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 473 + ], + "type": "text", + "content": "Open Panoramic Segmentation (OPS) task is proposed to deal with three challenging problems: ① the narrow FoV, ② the restricted range of classes, and ③ the scarcity of panoramic labels. OPS gives three answers to the aforementioned issues: ① Open the FoV, ② Open the Vocabulary, and ③ Open the Domain. The OPS task paradigm is shown in Fig. 1c. Models are trained in the narrow-FoV pinhole source domain in an open-vocabulary setting while evaluated in the wide-FoV panoramic target domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 490, + 261, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 261, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 261, + 502 + ], + "type": "text", + "content": "3.2 Model Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "A foundation model [49] can be efficiently transferred to downstream tasks by using an adapter [68]. To improve the modeling ability for panoramas, we design the OOOPS model. As shown in Fig. 2, it comprises a frozen CLIP model and a proposed Deformable Adapter Network (DAN) that incorporates multiple Transformer Layers and the novel DAO, which will be detailed in Sec. 3.3. Feature fusion takes place between the intermediate layers of CLIP and DAN. One of the two outputs from DAN consists of mask proposals, while the other serves as deep supervised guidance for CLIP to generate proposal logits. In the training phase, pinhole images are forwarded into OOOPS, generating mask proposals and proposal logits for the loss calculation. In the inference phase, panoramas are forwarded into OOOPS to generate segmentation predictions via the multiplication of mask proposals and the corresponding proposal logits. The frozen CLIP is necessary for the zero-shot learning ability of OOOPS." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 286 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 286 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 286 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 286 + ], + "type": "image", + "image_path": "96240dd66b057b5eea2695ecf8ab59c7653cff69998751351a5c6e81494ba866.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 328 + ], + "type": "text", + "content": "Fig. 2: Overview of the OOOPS model architecture. It consists of a frozen CLIP model and a Deformable Adapter Network (DAN) which includes Transformer Layers and the proposed DAO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 350, + 311, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 311, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 311, + 363 + ], + "type": "text", + "content": "3.3 Deformable Adapter Network" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": "The Deformable Adapter Network is a combination of multiple Transformer [55] Layers and the proposed DAO. Since distortion exists in panoramic images, which is a big challenge when utilizing informative panoramas, we delve deeper into the deformable design [57] and sampling methods like APES [60], MateRobot [85], proposing DAO to tackle the image distortions and object deformations of panoramas. To provide a detailed rethinking process of the deformable convolution design, we introduce the preliminary of the DCN series as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": "Revisiting DCN Series. The seminal work DCN [10] can enable the traditional CNN with spatial deformation-aware ability. Given a convolutional kernel of " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " sampling locations, let " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_k" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " denote the weight and pre-specified offset for the " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": "-th location, respectively. For example, " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "K = 9" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k \\in \\{(1,1), \\dots, (-1, -1)\\}" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " defines a " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " convolutional kernel of dilation 1. Let " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{y}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " denote the features at location " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " from the input feature maps " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": " and output feature maps " + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 451, + 482, + 536 + ], + "type": "text", + "content": ", respectively. DCN is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 542, + 481, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 542, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 233, + 542, + 481, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {y} (\\mathbf {p}) = \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {k} \\mathbf {x} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {k}\\right), \\tag {1}", + "image_path": "55cee41059c310d54cc0ccf12d9c099e0b0a098dcde7ca80ef9111ce32347830.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 580, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 580, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_k" + }, + { + "bbox": [ + 130, + 580, + 482, + 629 + ], + "type": "text", + "content": " is the learnable offset for the " + }, + { + "bbox": [ + 130, + 580, + 482, + 629 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 580, + 482, + 629 + ], + "type": "text", + "content": "-th location. Although DCN is capable of capturing spatial deformation, every sampling location is treated equally when calculating the local features. DCNv2 [90] is proposed with an additional term called modulation scalar. Specifically, DCNv2 can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 226, + 635, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 635, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 226, + 635, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\mathbf {y} (\\mathbf {p}) = \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {k} \\mathbf {m} _ {k} \\mathbf {x} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {k}\\right), \\tag {2}", + "image_path": "11e803181590708b7541f4726b48c55f47fadec5eeba60f076289730cb45736d.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_k" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is the learnable modulation scalar for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "-th location. Inspired by Transformer, DCNv3 [57] is proposed with a grouping operation, further boosting the deformation-aware ability of DCNv2. DCNv3 can be addressed with the following formula:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 172, + 481, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 172, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 212, + 172, + 481, + 207 + ], + "type": "interline_equation", + "content": "\\mathbf {y} (\\mathbf {p}) = \\sum_ {g = 1} ^ {G} \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {g} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {g k}\\right), \\tag {3}", + "image_path": "5730d2a3bb7a293f09c144df7e30a5f3fc7ca215b0cd9813ce5e88f4c34b183d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "text", + "content": " denotes the total number of aggregation groups. DCNv4 [65] is similar to DCNv3, achieving similar performance while significantly reducing the runtime." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 262, + 396, + 376 + ], + "blocks": [ + { + "bbox": [ + 219, + 262, + 396, + 376 + ], + "lines": [ + { + "bbox": [ + 219, + 262, + 396, + 376 + ], + "spans": [ + { + "bbox": [ + 219, + 262, + 396, + 376 + ], + "type": "image", + "image_path": "f06c2a86499b210dcd1eacf22acc69164e03fbc1dd56f82f5f5913404a1e637b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 388, + 389, + 400 + ], + "lines": [ + { + "bbox": [ + 224, + 388, + 389, + 400 + ], + "spans": [ + { + "bbox": [ + 224, + 388, + 389, + 400 + ], + "type": "text", + "content": "Fig. 3: Salient map generation in DAO." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 483 + ], + "type": "text", + "content": "Deformable Adapter Operator (DAO). When dealing with distortions in panoramas, DCNv3 and DCNv4 fall short of meeting the distortion-aware requirements. This is illustrated by the visualization of Fig. 6 in Sec. 4.5. Therefore, DAO is proposed to tackle the distortion problem in panoramic images using the following expression:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 201, + 491, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 491, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 201, + 491, + 481, + 526 + ], + "type": "interline_equation", + "content": "\\mathbf {y} (\\mathbf {p}) = \\mathbf {s} (\\mathbf {p}) \\sum_ {g = 1} ^ {G} \\sum_ {k = 1} ^ {K} \\mathbf {w} _ {g} \\mathbf {m} _ {g k} \\mathbf {x} _ {g} \\left(\\mathbf {p} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {g k}\\right), \\tag {4}", + "image_path": "62a445df312d69d90f74d15235c4dbabc88a0c82a9b0d1b2d7a9dd44da3d79db.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{s}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " is the learnable salient scalar at location " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". Inheriting from DCNv3, DAO comes up with an additional term called salient scalar indicating the importance of each pixel within the whole panorama. It's worth noting that DCNv3 and DCNv4 share the same mathematical expression, but DCNv3 is more robust according to our experiments before designing DAO. Therefore, we adopt DCNv3 as a part of DAO rather than DCNv4. Illustrated in Fig. 2, the feature map outputted by DCNv3 passes through the Patch Similarity Layer, Normalization Layer, and Standard Deviation Layer sequentially to form a salient map. The intuition for such a design is straightforward. The salient pixels in an image are those that differ significantly from their neighboring pixels, e.g., edge pixels. If all pixels within an image patch are different, the standard deviation of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 114, + 253, + 160 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 253, + 160 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 253, + 160 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 253, + 160 + ], + "type": "image", + "image_path": "8357fdae817e0482fcaef067b2002045730139dec0591a682b911a14d7cbaa62.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 218, + 163, + 248, + 183 + ], + "blocks": [ + { + "bbox": [ + 218, + 163, + 248, + 183 + ], + "lines": [ + { + "bbox": [ + 218, + 163, + 248, + 183 + ], + "spans": [ + { + "bbox": [ + 218, + 163, + 248, + 183 + ], + "type": "image", + "image_path": "a89652183d1d5894817bbf5376ab43d72310c9add4d9018828b1b24a11131f62.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 146, + 184, + 301, + 214 + ], + "blocks": [ + { + "bbox": [ + 146, + 184, + 301, + 214 + ], + "lines": [ + { + "bbox": [ + 146, + 184, + 301, + 214 + ], + "spans": [ + { + "bbox": [ + 146, + 184, + 301, + 214 + ], + "type": "image", + "image_path": "34f27cedfb3013359b313cdd70477ac0779ffcfc05bbdeb66082d67e9348efbb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 216, + 303, + 250 + ], + "lines": [ + { + "bbox": [ + 145, + 216, + 303, + 250 + ], + "spans": [ + { + "bbox": [ + 145, + 216, + 303, + 250 + ], + "type": "text", + "content": "(a) Equirectangular Projection (ERP) maps a globe into a panorama. It's clear that strong distortion occurs in the panorama, especially on object edges in the horizontal direction." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 361, + 163 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 361, + 163 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 361, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 361, + 163 + ], + "type": "image", + "image_path": "9df6f7374ef486d09cff92e1638f00aae16e2a5daab186c82eef4380a1420a81.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 413, + 163 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 413, + 163 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 413, + 163 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 413, + 163 + ], + "type": "image", + "image_path": "32aeedbff43f2742c95dc8624148e698b2877a25b42a6214db3d2f7f1cfeff71.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 165, + 361, + 214 + ], + "blocks": [ + { + "bbox": [ + 312, + 165, + 361, + 214 + ], + "lines": [ + { + "bbox": [ + 312, + 165, + 361, + 214 + ], + "spans": [ + { + "bbox": [ + 312, + 165, + 361, + 214 + ], + "type": "image", + "image_path": "e378c1784ab7463c61bbff1685d866a257219ea04f85e8641d388a8382f2d5a6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 258, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 131, + 258, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 480, + 281 + ], + "type": "text", + "content": "Fig. 4: (a) Visualization of ERP on a panoramic image and (b) Our proposed RERP on pinhole images." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 363, + 164, + 413, + 214 + ], + "blocks": [ + { + "bbox": [ + 363, + 164, + 413, + 214 + ], + "lines": [ + { + "bbox": [ + 363, + 164, + 413, + 214 + ], + "spans": [ + { + "bbox": [ + 363, + 164, + 413, + 214 + ], + "type": "image", + "image_path": "0eaf5c01de9eeed8815723df3ff32c6c0732a6171a46e45f37ba7a7b821e454f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 216, + 468, + 250 + ], + "lines": [ + { + "bbox": [ + 310, + 216, + 468, + 250 + ], + "spans": [ + { + "bbox": [ + 310, + 216, + 468, + 250 + ], + "type": "text", + "content": "(b) Visualization of ERP and RERP of pinhole images. First column: pinhole images. Second column: ERP on pinhole images. Third column: RERP on pinhole images." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 416, + 114, + 466, + 163 + ], + "blocks": [ + { + "bbox": [ + 416, + 114, + 466, + 163 + ], + "lines": [ + { + "bbox": [ + 416, + 114, + 466, + 163 + ], + "spans": [ + { + "bbox": [ + 416, + 114, + 466, + 163 + ], + "type": "image", + "image_path": "4b814a83bf59069de5a5b863c893dfb42b267022716d7216b0ba6491ee1b7c17.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 416, + 164, + 466, + 214 + ], + "blocks": [ + { + "bbox": [ + 416, + 164, + 466, + 214 + ], + "lines": [ + { + "bbox": [ + 416, + 164, + 466, + 214 + ], + "spans": [ + { + "bbox": [ + 416, + 164, + 466, + 214 + ], + "type": "image", + "image_path": "12937113d2c42facd50178b70956f57af3e93762ce4e218f71ec156d35af2ff5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 302, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 434 + ], + "type": "text", + "content": "pixel similarity for this patch is higher than the one containing similar pixels, resulting in a higher salient scalar. Fig. 3 gives a more detailed explanation of the salient map generation. Given a feature map, DAO first calculates the cosine similarity between the center pixel and all pixels within a kernel, e.g., 9 pixels within a " + }, + { + "bbox": [ + 130, + 302, + 482, + 434 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 302, + 482, + 434 + ], + "type": "text", + "content": " kernel in Fig. 3, resulting in a 9-dimensional cosine similarity vector. Softmax normalization is then applied to the vector. Afterward, DAO calculates the standard deviation of this normalized cosine similarity vector, indicating the importance of the center pixel. By traversing every pixel across the entire feature map, a salient map is generated for the sake of enhancing the salient pixels, which are usually the edge pixels of an image where strong panoramic distortion occurs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 449, + 343, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 343, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 343, + 462 + ], + "type": "text", + "content": "3.4 Random Equirectangular Projection" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 468, + 481, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 504 + ], + "type": "text", + "content": "The Equirectangular Projection [50] (ERP) is one of the most common methods for mapping a globe into a panoramic plane, which transforms spherical coordinates into planar coordinates expressed as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 255, + 511, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 511, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 255, + 511, + 480, + 525 + ], + "type": "interline_equation", + "content": "x = R \\left(\\lambda - \\lambda_ {0}\\right) \\cos \\left(\\varphi_ {1}\\right), \\tag {5}", + "image_path": "1e5aa588102e835747dab84dfdd6d58fad245ee930d565224964400a405e1654.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 256, + 526, + 480, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 526, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 256, + 526, + 480, + 539 + ], + "type": "interline_equation", + "content": "y = R \\left(\\varphi - \\varphi_ {0}\\right), \\tag {6}", + "image_path": "2acb3f855f7c8e34982f9bfdecde5484b99d681d52ee07b469b4fc9db9730588.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " are the longitude and latitude of the location to project, respectively. " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varphi_{1}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " are the standard parallels. " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_0" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varphi_0" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " are the central meridian and central parallel of the map, respectively. " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is the radius of the globe. " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " denotes the horizontal coordinate, while " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " represents the vertical coordinate of the projected location on the map. Fig. 4a visualizes the ERP on a panoramic plane. It can be observed that after equirectangular projection strong distortion occurs in the panorama, e.g., a straight line is transformed into a curve. To further boost the performance, we propose Random Equirectangular Projection (RERP) on pinhole images since the OPS task requires the model to train with pinhole images instead of panoramas." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 212 + ], + "type": "text", + "content": "We divide the pinhole image into four parts and randomly shuffle the image patches. Afterward, we apply equirectangular projection to the distortion-free pinhole image. Fig. 4b visualizes pinhole images after RERP. The first column is the pinhole images without any data augmentation. The second column is ERP on pinhole images without random shuffling. The last column is our proposed RERP. It can be observed that the panorama-like distortion also occurs in the pinhole images after RERP. Random shuffling is employed to enhance robustness and promote generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 228, + 229, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 228, + 229, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 228, + 229, + 243 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 253, + 204, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 253, + 204, + 264 + ], + "spans": [ + { + "bbox": [ + 134, + 253, + 204, + 264 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 271, + 497, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 497, + 319 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 497, + 319 + ], + "type": "text", + "content": "We train our model using the pinhole general-object COCOStuff-164k [5] while evaluating the model performance on three different panoramic datasets: panoramic street-view WildPASS [76], panoramic indoor-scene Stanford2D3D [2] and panoramic indoor-scene Matterport3D [6]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 319, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 319, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 134, + 319, + 481, + 366 + ], + "type": "text", + "content": "COCOStuff-164k. The pinhole dataset comprises 164k images across 171 annotated classes, partitioned into training, validation, and test sets, with 118k, 5k, and 41k images, respectively. In our experiments, we utilize the entire 118k-image training set as the training data by default." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 367, + 481, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 367, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 134, + 367, + 481, + 414 + ], + "type": "text", + "content": "WildPASS. The panoramic street-view dataset contains 2,500 panoramas in 65 cities from all continents Asia, Europe, Africa, Oceania, and North and South America (excluding only Antarctica) with 8 dense-annotated classes. All images have a " + }, + { + "bbox": [ + 134, + 367, + 481, + 414 + ], + "type": "inline_equation", + "content": "70^{\\circ} \\times 360^{\\circ}" + }, + { + "bbox": [ + 134, + 367, + 481, + 414 + ], + "type": "text", + "content": " FoV in the size of " + }, + { + "bbox": [ + 134, + 367, + 481, + 414 + ], + "type": "inline_equation", + "content": "400 \\times 2048" + }, + { + "bbox": [ + 134, + 367, + 481, + 414 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 415, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 415, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 134, + 415, + 481, + 449 + ], + "type": "text", + "content": "Stanford2D3D. The panoramic indoor dataset consists of 1,413 images in the size of " + }, + { + "bbox": [ + 134, + 415, + 481, + 449 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 134, + 415, + 481, + 449 + ], + "type": "text", + "content": " and 13 object classes. All results are averaged over 3 cross-validation folds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 450, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 450, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 134, + 450, + 481, + 498 + ], + "type": "text", + "content": "Matterport3D. The panoramic indoor dataset includes 10,800 panoramic views from 194,400 RGB-D images. According to the data preprocessing pipeline of 360BEV [54], a subset of 772 images in the size of " + }, + { + "bbox": [ + 134, + 450, + 481, + 498 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 134, + 450, + 481, + 498 + ], + "type": "text", + "content": " with 20 classes is used to evaluate the segmentation performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 515, + 198, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 515, + 198, + 526 + ], + "spans": [ + { + "bbox": [ + 134, + 515, + 198, + 526 + ], + "type": "text", + "content": "4.2 Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 533, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 481, + 605 + ], + "type": "text", + "content": "Close-vocabulary Metric. Mean Intersection over Union (mIoU) is used as the evaluation metric in close-vocabulary semantic segmentation tasks. Specifically, IoU is calculated as the area of overlap between the predicted segmentation and the ground truth, divided by the area of union between them. mIoU is a hard metric designed for close-vocabulary segmentation ignoring the semantic meanings of similar classes referring to the open-vocabulary setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 605, + 481, + 665 + ], + "type": "text", + "content": "Open-vocabulary Metric. Open mean Intersection over Union [88], open mIoU for short, is a soft metric compared to mIoU since the similarity scores between different classes are also taken into account, which are calculated by the WordNet approach [43]. In this work, we will report both close and open metrics for a comprehensive understanding of all models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 149, + 479, + 328 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 1: Results on WildPASS. The mIoU, open mIoU, and per-class IoU are in percentage (%). #PARAMs is the number of learnable parameters in millions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 149, + 479, + 328 + ], + "lines": [ + { + "bbox": [ + 140, + 149, + 479, + 328 + ], + "spans": [ + { + "bbox": [ + 140, + 149, + 479, + 328 + ], + "type": "table", + "html": "
Network#PARAMSmIoUOpen mIoUCarRoadSidewalkCrosswalkCurbpersonTruckBus
Close-Vocabulary
Fast-SCNN [48]1.124.8-45.960.031.79.717.16.014.213.8
PSPNet18 [84]17.528.2-58.266.828.413.019.26.218.215.6
SwiftNet [47]11.830.0-55.764.129.216.222.98.521.122.2
DenseASPP [77]8.333.2-51.169.138.116.426.38.727.428.4
ERF-PSPNet [73]2.534.0-66.370.536.56.424.19.426.532.0
PSPNet50 [84]53.346.1-80.074.951.723.931.419.838.948.1
DANet [15]47.447.2-74.872.249.928.923.825.451.950.6
OOSS (ERF-PSPNet) [74]2.556.1-87.279.360.828.038.154.548.852.2
PASS (ERF-PSPNet) [73]2.564.7-87.380.061.471.149.972.237.557.9
ERF-PSPNet (omni-sup') [76]2.566.8-90.582.765.670.551.558.262.053.1
ECANet (with attention) [76]2.667.7-90.483.768.467.852.161.454.563.3
ECANet (with fusion) [76]2.669.0-90.685.768.067.952.166.059.362.3
Open-Vocabulary
SAN [68]8.455.656.984.381.969.610.12.872.254.969.1
CAT-seg [9]59.555.757.184.181.570.010.23.072.554.869.3
OpenSeeD [80]65.455.857.284.381.669.710.33.073.654.969.1
OOOPS (ours, w/o RERP)8.757.058.584.981.668.715.18.471.355.370.4
OOOPS (ours, w/ RERP)8.758.059.784.981.668.719.48.373.854.972.2
", + "image_path": "dda5503e70337909b4f5c8276767944cd43d39a150eaea0cf615464abf7c3313.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 360, + 279, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 279, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 279, + 372 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 536 + ], + "type": "text", + "content": "We utilize CLIP ViT-B/16 as the frozen CLIP model. All models are trained with the pinhole general-object COCOStuff-164k training set with " + }, + { + "bbox": [ + 130, + 378, + 482, + 536 + ], + "type": "inline_equation", + "content": "4 \\times \\mathrm{A}40" + }, + { + "bbox": [ + 130, + 378, + 482, + 536 + ], + "type": "text", + "content": " GPUs while evaluated on three panoramic datasets which are in totally different domains compared with the pinhole dataset. The AdamW optimizer is applied with an initial learning rate of 0.0001, weight decay of 0.0001, batch size of 32, and a total of 60,000 training iterations. Throughout the training, the learning rate follows a polynomial schedule with a power of 0.9. Data augmentations, including random image resizing within the short-side range of [320, 1024] and a crop size of " + }, + { + "bbox": [ + 130, + 378, + 482, + 536 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 378, + 482, + 536 + ], + "type": "text", + "content": " are applied. Additionally, when verifying the effectiveness of the proposed RERP, it is applied as an extra data augmentation. Following the practice of SAN [68], apart from the supervision of mask recognition with cross-entropy loss, we utilize the dice loss and binary cross-entropy loss for mask proposal generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 551, + 353, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 353, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 353, + 563 + ], + "type": "text", + "content": "4.4 Comparison with the State of the Art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Results on WildPASS. Comparisons with both close- and open-vocabulary state-of-the-art methods on WildPASS are listed in Table 1. It is not surprising that ECANet (with fusion) [76] achieves the best performance among all methods since it is a close-vocabulary method, meaning that ECANet is only able to predict the classes predefined by the dataset. However, our OOOPS model is qualified for predicting classes that are not annotated in the dataset. The OOOPS method (with RERP) surpasses other state-of-the-art open-vocabulary panoramic segmentation methods with over " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " mIoU while introducing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 304, + 185 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 304, + 185 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 304, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 304, + 185 + ], + "type": "image", + "image_path": "d5d995f006c1007f0e6a53716257988c635e6044cfc30b8070d4c849eec7e24f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 187, + 304, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 187, + 304, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 304, + 228 + ], + "type": "text", + "content": "(a) Comparison with baseline method SAN [68] on the WildPASS dataset with predefined 8 classes. First row: ground truth. Second row: prediction from SAN. Third row: prediction from OOOPS with RERP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 132, + 238, + 480, + 260 + ], + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 260 + ], + "type": "text", + "content": "Fig. 5: (a) Comparison on the WildPASS dataset and (b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 480, + 185 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 480, + 185 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 480, + 185 + ], + "type": "image", + "image_path": "6d1a70ea15719e6c11daca8a1148376fe15be37ed7e72f6e7e884c63a3a9c1c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 187, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 309, + 187, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 187, + 481, + 228 + ], + "type": "text", + "content": "(b) Visualization of the prediction from OOOPS in close- and open-vocabulary settings. First row: RGB. Second row: prediction with close vocabularies. Third row: prediction with open vocabularies. Zoom in for a better view." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "text", + "content": "only a minor increment of 0.3 million learnable parameters compared to the lightweight baseline SAN [68], offering a promising step towards bridging the gap between close- and open-vocabulary settings. Moreover, our OOOPS method enhanced by RERP demonstrates significant advancements in challenging categories such as crosswalk and curb, achieving " + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "inline_equation", + "content": "+9.3\\%" + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "inline_equation", + "content": "+5.5\\%" + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "text", + "content": " mIoU compared with the baseline SAN, respectively. It is worth noting that in some categories, e.g., sidewalk, person, and bus, our proposed OOOPS (both with and without RERP) even outperforms the models trained in a close-vocabulary setting. All open-vocabulary methods achieve better performance with open mIoU metric compared to the close mIoU since similar semantic meanings are also considered in the calculation. Fig. 1a showcases the performance degradation on both metrics. Even though all methods have degraded performance when FoV increases, our OOOPS achieves a smaller drop " + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "inline_equation", + "content": "(-10.0\\%)" + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "text", + "content": " while others have a larger performance decrease " + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "inline_equation", + "content": "(-12.0\\%)" + }, + { + "bbox": [ + 130, + 291, + 482, + 459 + ], + "type": "text", + "content": ", indicating the effectiveness of the OOOPS model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Fig. 5a visualizes the prediction of the challenging curb category from baseline SAN and our proposed OOOPS with RERP. From top to bottom are the ground truth, prediction from SAN, and prediction from OOOPS with RERP, respectively. The baseline SAN cannot recognize the challenging curb category while our proposed OOOPS with RERP is capable of detecting the curb in the middle of the panorama highlighted in gray. Fig. 5b showcases the visualization of the prediction from OOOPS on WildPASS in close- and open-vocabulary settings. From top to bottom are input RGB, prediction from OOOPS with 8 predefined close vocabularies from WildPASS, and prediction from OOOPS with unlimited open vocabularies. It is evident that our OOOPS with RERP is capable of tackling unseen categories in the open-vocabulary setting. Every pixel on the panorama is assigned a meaningful semantic label. For example, \"mailbox\" is not a predefined category from WildPASS. While existing models can not recognize the mailbox under a close-vocabulary setting, the rare classes can be easily detected by our model in an open-vocabulary one, indicating the great potential of open-vocabulary panoramic segmentation for parsing unconstrained traffic surroundings." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 152, + 479, + 276 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Results on Stanford2D3D. The mIoU, open mIoU, and per-class IoU are in percentage (\\%). #PARAMs is the number of learnable parameters in millions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 152, + 479, + 276 + ], + "lines": [ + { + "bbox": [ + 137, + 152, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 137, + 152, + 479, + 276 + ], + "type": "table", + "html": "
Network#PARAMsmIoUOpen mIoUBeamBoardBookcaseCeilingChairClutterColumnDoorFloorSofaTableWallWindow
Close-Vocabulary
Trans4PASS [82]28.252.1-11.460.651.181.859.737.923.554.688.834.653.968.151.0
360BEV [54]28.454.3-06.063.655.582.762.739.916.458.890.845.257.171.157.6
Open-Vocabulary
SAN [68]8.438.539.50.251.539.666.552.110.70.045.049.331.849.250.053.4
CAT-seg [9]59.538.639.60.051.642.366.851.48.00.047.450.731.247.551.453.5
OpenSeeD [80]65.438.740.00.051.842.067.051.68.40.247.550.031.447.151.753.6
OOOPS (w/o RERP)8.739.541.00.052.344.167.153.79.70.050.152.830.547.252.253.8
OOOPS (w/ RERP)8.741.142.60.050.744.868.851.88.10.048.271.133.647.255.454.5
", + "image_path": "97019023d5e34c99681f23af5ec363fd8e5f88aed1656823fc13cffb6634646c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 334, + 479, + 427 + ], + "blocks": [ + { + "bbox": [ + 132, + 297, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 132, + 297, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 482, + 321 + ], + "type": "text", + "content": "Table 3: Results on Matterport3D. The mIoU, open mIoU, and per-class IoU are in percentage (\\%). #PARAMs is the number of learnable parameters in millions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 334, + 479, + 427 + ], + "lines": [ + { + "bbox": [ + 136, + 334, + 479, + 427 + ], + "spans": [ + { + "bbox": [ + 136, + 334, + 479, + 427 + ], + "type": "table", + "html": "
Network≈PARAMsinotOpen mIoUWallFloorChairDoorTablePictureFurnitureObjectsWindowSofbBedSinkStarsCeilingToiletMirrorshowerBathtubCounterShelving
Close-Vocabulary
Trans4PASS+ [83]28.542.6-63.479.139.140.332.836.031.031.537.544.063.220.641.877.640.724.323.758.334.332.9
360BEV [54]28.446.4-64.183.145.845.038.041.132.335.140.648.769.825.147.880.246.028.722.360.138.634.8
Open-Vocabulary
SAN [68]8.430.531.454.476.118.034.122.321.113.01.732.734.251.011.019.974.223.412.031.631.919.028.4
CAT-seg [9]59.530.231.154.374.817.034.121.920.812.61.932.934.050.110.718.873.522.411.631.932.919.228.5
OpenSecD [80]65.430.631.654.276.318.134.322.521.013.02.132.834.150.911.220.274.123.512.231.131.819.128.6
OOOPS (w/o RERP)8.731.132.455.277.017.135.725.017.513.52.333.032.451.812.620.474.522.610.834.136.819.429.7
OOOPS (w/ RERP)8.731.232.555.476.418.434.823.921.213.22.633.334.552.912.120.374.628.212.531.929.319.428.6
", + "image_path": "5ae218e7f7d60f156392cf66d9eb8c6eef53217a440c1e317c5460a2dc5afca5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "content": "Results on Stanford2D3D. In Table 2, the comparison is conducted between state-of-the-art methods on the Stanford2D3D dataset. All reported results are averaged over 3 cross-validation folds. Without RERP, OOOPS achieves " + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "inline_equation", + "content": "39.5\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "content": " mIoU, outperforming other open-vocabulary methods like SAN [68], CAT-seg [9], and OpenSeeD [80], which score between " + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "inline_equation", + "content": "38.5\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "content": ". With RERP, OOOPS's mIoU increases to " + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 544 + ], + "type": "text", + "content": " gain, narrowing the gap with advanced close-vocabulary methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": "Results on Matterport3D. In Table 3, the comparison is conducted between state-of-the-art methods on the Matterport3D dataset. Our OOOPS method showcases consistent performance. Without augmentation, OOOPS achieves mIoU of " + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "inline_equation", + "content": "31.1\\%" + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": ", slightly outperforming other open-vocabulary methods like SAN [68] " + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "inline_equation", + "content": "(30.5\\%)" + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": ", CAT-seg [9] " + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "inline_equation", + "content": "(30.2\\%)" + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": ", and OpenSeeD [80] " + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "inline_equation", + "content": "(30.6\\%)" + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": ". With augmentation, OOOPS's mIoU marginally increases to " + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "inline_equation", + "content": "31.2\\%" + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": ", further reinforcing its lead over these open-vocabulary methods. Both variants of OOOPS still lag behind close-vocabulary methods like Trans4PASS+ [83] and 360BEV [54], which however, are trained in a supervised fashion in a close-vocabulary setting on panoramas and have mIoUs ranging from " + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "inline_equation", + "content": "42.6\\%" + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "inline_equation", + "content": "46.4\\%" + }, + { + "bbox": [ + 130, + 545, + 487, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 135, + 480, + 178 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "type": "text", + "content": "Table 4: Task difference between Domain Adaptation (DA) and the proposed OPS." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 135, + 480, + 178 + ], + "lines": [ + { + "bbox": [ + 136, + 135, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 136, + 135, + 480, + 178 + ], + "type": "table", + "html": "
TaskDomain#ClassesmIoU (%)
SourceTargetWildPASSMatterport3DStanford2D3D
DA✓✓limited72.247.255.4
OPS✓✗unlimited58.031.241.1
", + "image_path": "3a9acf26bcb2c2345a11bd001df00da8e9eb83811bdd7991a5cbb61720317c48.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 211, + 305, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 211, + 305, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 305, + 245 + ], + "type": "text", + "content": "(a) Ablation study of OOOPS components on the WildPASS Dataset. The mIoU is reported in percentage " + }, + { + "bbox": [ + 132, + 211, + 305, + 245 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 211, + 305, + 245 + ], + "type": "text", + "content": ". #PARAMs stands for the number of learnable parameters in millions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 212, + 481, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 212, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 481, + 246 + ], + "type": "text", + "content": "(b) Ablation study of different numbers of shuffling patches on WildPASS, Matterport3D and Stanford2D3D datasets. The mIoU is reported in percentage " + }, + { + "bbox": [ + 308, + 212, + 481, + 246 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 308, + 212, + 481, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 142, + 254, + 297, + 346 + ], + "blocks": [ + { + "bbox": [ + 132, + 187, + 479, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 187, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 479, + 198 + ], + "type": "text", + "content": "Table 5: Ablation study of OOOPS components and the number of shuffling patches." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 254, + 297, + 346 + ], + "lines": [ + { + "bbox": [ + 142, + 254, + 297, + 346 + ], + "spans": [ + { + "bbox": [ + 142, + 254, + 297, + 346 + ], + "type": "table", + "html": "
MethodERPRERP#PARAMsmIoU
SAN [68]✘✘8.455.6
SAN [68]✓✘8.456.0
+ DCN [68]✘✓8.456.5 (+0.9)
+ DCN✘✘8.655.6
+ DCN✓✘8.656.0
+ DCN✘✓8.656.6 (+1.0)
+ DCNv2✘✘8.755.7
+ DCNv2✓✘8.756.4
+ DCNv2✘✓8.756.6 (+0.9)
+ DCNv3✘✘8.755.8
+ DCNv3✓✘8.756.4
+ DCNv3✘✓8.756.8 (+1.0)
+ DCNv4✘✘8.755.8
+ DCNv4✓✘8.756.4
+ DCNv4✘✓8.756.8 (+1.0)
+ DAO (ours)✘✘8.757.0
+ DAO (ours)✓✘8.757.6
+ DAO (ours)✘✓8.758.0 (+1.0)
", + "image_path": "b897021ddfb07d27b2e5315df1f6ce4de1ace7d6e344acb532b5bfa9f16120a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 312, + 255, + 480, + 346 + ], + "blocks": [ + { + "bbox": [ + 312, + 255, + 480, + 346 + ], + "lines": [ + { + "bbox": [ + 312, + 255, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 312, + 255, + 480, + 346 + ], + "type": "table", + "html": "
mIoU (%) #Patches Dataset4 9 16 25
WildPASS58.0 58.5 58.2 58.0
Matterport3D31.2 31.6 31.4 31.0
Stanford2D3D41.1 41.9 41.5 41.1
", + "image_path": "88cf2b1e700bc10e5bbc325f1b50cfce194aa3c604195de4b692a84b5df0565f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 366, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 475 + ], + "type": "text", + "content": "Results in DA Subtask. The proposed OPS task differs from Domain Adaptation since no data from the target domain is available to models in OPS while models have access to the data in DA during training. Moreover, models are required to recognize an unlimited amount of classes in OPS while only a limited number of classes can be recognized in DA. DA is a special case of OPS, meaning that models in OPS can be easily adapted to DA and achieve better performance. The converse doesn't hold. Table 4 showcases the task difference between OPS and DA. It can be observed that over " + }, + { + "bbox": [ + 130, + 366, + 482, + 475 + ], + "type": "inline_equation", + "content": "+14\\%" + }, + { + "bbox": [ + 130, + 366, + 482, + 475 + ], + "type": "text", + "content": " gain in mIoU is achieved when adapting our OOOPS model in the DA subtask." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 491, + 239, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 239, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 239, + 503 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "We replace DAO in OOOPS, as illustrated in Fig. 2, with DCN [10], DCNv2 [90], DCNv3 [57] and DCNv4 [65]. When referring to the component substitution without data augmentation in Table 5a, it is clear that our DAO surpasses other counterparts with a maximum performance gain of " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "+1.4\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " mIoU. As for the component substitution with ERP, all models achieve a boost in the range of " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " compared with the one without data augmentation. The gain is even bigger considering the component substitution with RERP, achieving " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "+1.0\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " mIoU. It is worth noting that RERP is able to boost the model performance stably among all models. The best model performance is achieved by the combination of DAO and RERP, yielding " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "58.0\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " mIoU with only a slight increment in parameters. Table 5b showcases the influence of different numbers of shuffling patches. The performance is not proportional to the number of patches, indicating the importance of moderate content continuity during training." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 305, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 305, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 305, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 305, + 228 + ], + "type": "image", + "image_path": "7d5096a80cb57a68af065098d6f774bc40acd6312e3c2f1dd0f08756f7b19cbb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 230, + 305, + 257 + ], + "lines": [ + { + "bbox": [ + 132, + 230, + 305, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 305, + 257 + ], + "type": "text", + "content": "(a) Deformable offsets without RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 481, + 228 + ], + "type": "image", + "image_path": "03505442d6f0e8b1e5367307847b1b80fcb74392fefd5aaf81cd3c8d0c69a639.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 230, + 481, + 257 + ], + "lines": [ + { + "bbox": [ + 309, + 230, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 309, + 230, + 481, + 257 + ], + "type": "text", + "content": "(b) Deformable offsets with RERP. From top to bottom are deformable offsets of DCN, DCNv2, DCNv3, DCNv4, and proposed DAO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": "Fig. 6: Visualization of deformable offsets. The green points " + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": " are sampling locations. The red points " + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": " are deformable offsets in 2 levels, indicating a deformable receptive field (e.g., each level has a " + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": " kernel size, resulting in " + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "inline_equation", + "content": "(3 \\times 3)^{2} = 81" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": " red points)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 483, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 483, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 483, + 494 + ], + "type": "text", + "content": "To further illustrate the deformable ability of DAO and the superiority of RERP when tackling the distortion of panoramas, we visualize the deformable offset in Fig. 6. Fig. 6a presents the deformable offsets without RERP. It can be observed that DAO is better aware of the panoramic distortion compared with the rest in the DCN series. Since the strongest distortion occurs on the edge of objects, DAO focuses more on edge pixels for better distortion-aware feature extraction. The standard deviation of patch similarity in DAO implies the salience of the image patch center. By multiplicatively combining salient scalars and features, significant pixels, such as those representing object edges, are enhanced in their representation. In Fig. 6b, it is evident that the inclusion of RERP improves the distortion information capturing ability of all models in the DCN series, as opposed to their counterparts without RERP. This observation is consistent with the results in Table 5a. The incorporation of both DAO and RERP in OOOPS showcases a remarkable distortion-aware modeling capability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "In this work, we formulate the novel task of Open Panoramic Segmentation (OPS) for unfolding holistic scene understanding, which reaches Open FoV, Open Vocabulary, and Open Domain in surrounding scene segmentation. We have designed the Deformable Adapter Network (DAN) and Random Equirectangular Projection (RERP) for enhancing distortion-aware segmentation capacities of the novel proposed OOOPS model. Extensive experiments on WildPASS, Stanford2D3D, and Matterport3D benchmarks show that the proposed method significantly elevates the performance of state-of-the-art panoramic segmenters." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "type": "text", + "content": "This work was supported in part by the Ministry of Science, Research and the Arts of Baden-Württemberg (MWK) through the Cooperative Graduate School Accessibility through AI-based Assistive Technology (KATE) under Grant BW6-03, in part by BMBF through a fellowship within the IFI programme of DAAD, in part by the Helmholtz Association Initiative and Networking Fund on the HAICORE@KIT and HOREKA@KIT partition, in part by the National Key RD Program under Grant 2022YFB4701400, and in part by Hangzhou SurImage Technology Company Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 275, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 275, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 275, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 275, + 480, + 308 + ], + "type": "text", + "content": "1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 329 + ], + "type": "text", + "content": "2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 330, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 330, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 330, + 480, + 350 + ], + "type": "text", + "content": "3. Athwale, A., Afrasiyabi, A., Lagüie, J., Shili, I., Ahmad, O., Lalonde, J.F.: DarSwin: Distortion aware radial swim transformer. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 352, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 480, + 384 + ], + "type": "text", + "content": "4. Berenguel-Baeta, B., Bermudez-Cameo, J., Guerrero, J.J.: FreDSNet: Joint monocular depth and semantic segmentation with fast fourier convolutions from single panoramas. In: ICRA (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 384, + 480, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 405 + ], + "type": "text", + "content": "5. Caesar, H., Uijlings, J.R.R., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. CVPR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 406, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 480, + 437 + ], + "type": "text", + "content": "6. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. 3DV (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 438, + 480, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 480, + 470 + ], + "type": "text", + "content": "7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 471, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 492 + ], + "type": "text", + "content": "8. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. arXiv preprint arXiv:2303.11324 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 514 + ], + "type": "text", + "content": "9. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.: CAT-Seg: Cost aggregation for open-vocabulary semantic segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 535 + ], + "type": "text", + "content": "0. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: ICCV (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 556 + ], + "type": "text", + "content": "1. Dao, S.D., Shi, H., Phung, D., Cai, J.: Class enhancement losses with pseudo labels for open-vocabulary semantic segmentation. TMM (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "type": "text", + "content": "2. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "text", + "content": "3. Dong, B., Gu, J., Han, J., Xu, H., Zuo, W.: Towards universal vision-language omni-supervised segmentation. arXiv preprint arXiv:2303.06547 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 643 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "5. Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., Lu, H.: Dual attention network for scene segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Fu, X., Zhang, S., Chen, T., Lu, Y., Zhou, X., Geiger, A., Liao, Y.: PanopticNeRF-360: Panoramic 3D-to-2D label transfer in urban scenes. arXiv preprint arXiv:2309.10815 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "17. Gao, S., Yang, K., Shi, H., Wang, K., Bai, J.: Review on panoramic imaging and its applications in scene understanding. TIM (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "text", + "content": "19. Guo, J., Wang, Q., Gao, Y., Jiang, X., Tang, X., Hu, Y., Zhang, B.: MVP-SEG: Multi-view prompt learning for open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.06957 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "20. Guttikonda, S., Rambach, J.: Single frame semantic segmentation using multimodal spherical images. In: WACV (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "text", + "content": "21. Han, K., Liu, Y., Liew, J.H., Ding, H., Wei, Y., Liu, J., Wang, Y., Tang, Y., Yang, Y., Feng, J., Zhao, Y.: Global knowledge calibration for fast open-vocabulary segmentation. arXiv preprint arXiv:2303.09181 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "text", + "content": "22. Hu, X., An, Y., Shao, C., Hu, H.: Distortion convolution module for semantic segmentation of panoramic images based on the image-forming principle. TIM (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "23. Jang, S., Na, J., Oh, D.: DaDA: Distortion-aware domain adaptation for unsupervised semantic segmentation. In: NeurIPS (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "24. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Towards complete surrounding understanding via unsupervised contrastive learning. In: IV (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "text", + "content": "25. Jaus, A., Yang, K., Stiefelhagen, R.: Panoramic panoptic segmentation: Insights into surrounding parsing for mobile agents via unsupervised contrastive learning. T-ITS (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "26. Jiang, Q., Gao, S., Gao, Y., Yang, K., Yi, Z., Shi, H., Sun, L., Wang, K.: Minimalist and high-quality panoramic imaging with PSF-aware transformers. TIP (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 440 + ], + "type": "text", + "content": "27. Jiang, Q., Shi, H., Sun, L., Gao, S., Yang, K., Wang, K.: Annular computational imaging: Capture clear panoramic images through simple lens. TCI (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "text", + "content": "28. Jiang, W., Wu, Y.: DFNet: Semantic segmentation on panoramic images with dynamic loss weights and residual fusion block. In: ICRA (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "type": "text", + "content": "29. Jiayun, L., Khandelwal, S., Sigal, L., Li, B.: Plug-and-play, dense-label-free extraction of open-vocabulary semantic segmentation from vision-language models. arXiv preprint arXiv:2311.17095 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 494, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 515 + ], + "type": "text", + "content": "30. Kim, J., Jeong, S., Sohn, K.: PASTS: Toward effective distilling transformer for panoramic semantic segmentation. In: ICIP (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "text", + "content": "31. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "32. Li, J., Chen, P., Qian, S., Jia, J.: TagCLIP: Improving discrimination ability of open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.07547 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "33. Li, X., Wu, T., Qi, Z., Wang, G., Shan, Y., Li, X.: SGAT4PASS: Spherical geometry-aware transformer for panoramic semantic segmentation. In: IJCAI (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "34. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "35. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Mar-culescu, D.: Open-vocabulary semantic segmentation with mask-adapted CLIP. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "36. Ling, Z., Xing, Z., Zhou, X., Cao, M., Zhou, G.: PanoSwin: A pano-style swim transformer for panorama understanding. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Liu, Y., Ge, P., Liu, Q., Huang, D.: Multi-grained cross-modal alignment for learning open-vocabulary semantic segmentation from text supervision. arXiv preprint arXiv:2403.03707 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "38. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "39. Ma, C., Yang, Y.H., Wang, Y., Zhang, Y., Xie, W.: Open-vocabulary semantic segmentation with frozen vision-language models. In: BMVC (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "40. Ma, C., Yang, Y., Ju, C., Zhang, F., Zhang, Y., Wang, Y.: Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "type": "text", + "content": "41. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: DensePASS: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: ITSC (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "type": "text", + "content": "42. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: ECCV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "43. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "44. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H.S., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "45. Oin, J., Wu, J., Yan, P., Li, M., Ren, Y., Xiao, X., Wang, Y., Wang, R., Wen, S., Pan, X., Wang, X.: FreeSeg: Unified, universal and open-vocabulary image segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "text", + "content": "46. Orhan, S., Bastanlar, Y.: Semantic segmentation of outdoor panoramic images. SIVP (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "47. Orsic, M., Kreso, I., Bevandic, P., Segvic, S.: In defense of pre-trained ImageNet architectures for real-time semantic segmentation of road-driving images. In: CVPR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "text", + "content": "48. Poudel, R.P.K., Liwicki, S., Cipolla, R.: Fast-SCNN: Fast semantic segmentation network. In: BMVC (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "49. Radford, A., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "50. Ray, B., Jung, J., Larabi, M.: A low-complexity video encoder for equirectangular projected 360 video content. In: ICASSP (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "text", + "content": "51. Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: PanoFormer: Panorama transformer for indoor " + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "text", + "content": " depth estimation. In: ECCV (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "52. Shi, H., Zhou, Y., Yang, K., Yin, X., Wang, Z., Ye, Y., Yin, Z., Meng, S., Li, P., Wang, K.: PanoFlow: learning " + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": " optical flow for surrounding temporal understanding. T-ITS (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 577 + ], + "type": "text", + "content": "53. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. arXiv preprint arXiv:2306.13631 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "54. Teng, Z., Zhang, J., Yang, K., Peng, K., Shi, H., Reifs, S., Cao, K., Stiefelhagen, R.: 360BEV: Panoramic semantic mapping for indoor bird's-eye view. In: WACV (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "56. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "57. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., Wang, X., Qiao, Y.: InternImage: Exploring large-scale vision foundation models with deformable convolutions. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "58. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. arXiv preprint arXiv:2307.00764 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "59. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: OV-PARTS: Towards open-vocabulary part segmentation. arXiv preprint arXiv:2310.05107 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "60. Wu, C., Zheng, J., Pfrommer, J., Beyerer, J.: Attention-based point cloud edge sampling. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "61. Wysocza'nska, M., Ramamonjisoa, M., Trzci'nski, T., Simeoni, O.: CLIP-DIY: CLIP dense inference yields open-vocabulary semantic segmentation for-free. arXiv preprint arXiv:2309.14289 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "62. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "text", + "content": "63. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: DAT++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "64. Xie, B., Cao, J., Xie, J., Khan, F.S., Pang, Y.: SED: A simple encoder-decoder for open-vocabulary semantic segmentation. arXiv preprint arXiv:2311.15537 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "text", + "content": "65. Xiong, Y., Li, Z., Chen, Y., Wang, F., Zhu, X., Luo, J., Wang, W., Lu, T., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications. arXiv preprint arXiv:2401.06197 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "66. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "67. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "68. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: SAN: Side adapter network for open-vocabulary semantic segmentation. TPAMI (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "69. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "70. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "71. Xu, X., Xiong, T., Ding, Z., Tu, Z.: MasQCLIP for open-vocabulary universal image segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "72. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Huang, X., Sun, D., Wang, K.: Can we PASS beyond the field of view? Panoramic annular semantic segmentation for real-world surrounding perception. In: IV (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "73. Yang, K., Hu, X., Bergasa, L.M., Romera, E., Wang, K.: PASS: Panoramic annular semantic segmentation. T-ITS (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "74. Yang, K., Hu, X., Fang, Y., Wang, K., Stiefelhagen, R.: Omnisupervised omnidirectional semantic segmentation. T-ITS (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "75. Yang, K., Hu, X., Stiefelhagen, R.: Is context-aware CNN ready for the surroundings? Panoramic semantic segmentation in the wild. TIP (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "76. Yang, K., Zhang, J., Reif, S., Hu, X., Stiefelhagen, R.: Capturing omni-range context for omnidirectional segmentation. In: CVPR (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "77. Yang, M., Yu, K., Zhang, C., Li, Z., Yang, K.: DenseASPP for semantic segmentation in street scenes. In: CVPR (2018)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "J. Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 479 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "78. Yin, W., Liu, Y., Shen, C., Hengel, A.v.d., Sun, B.: The devil is in the labels: Semantic segmentation from sentences. arXiv preprint arXiv:2202.02002 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "text", + "content": "79. Yu, H., He, L., Jian, B., Feng, W., Liu, S.: PanelNet: Understanding 360 indoor environment via panel representation. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 480, + 182 + ], + "type": "text", + "content": "80. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Gao, J., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 215 + ], + "type": "text", + "content": "81. Zhang, J., Ma, C., Yang, K., Roitberg, A., Peng, K., Stiefelhagen, R.: Transfer beyond the field of view: Dense panoramic semantic segmentation via unsupervised domain adaptation. T-ITS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 215, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 480, + 247 + ], + "type": "text", + "content": "82. Zhang, J., Yang, K., Ma, C., Reif, S., Peng, K., Stiefelhagen, R.: Bending reality: Distortion-aware transformers for adapting to panoramic semantic segmentation. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 281 + ], + "type": "text", + "content": "83. Zhang, J., Yang, K., Shi, H., Reif, S., Peng, K., Ma, C., Fu, H., Torr, P.H.S., Wang, K., Stiefelhagen, R.: Behind every domain there is a shift: Adapting distortion-aware vision transformers for panoramic semantic segmentation. TPAMI (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 303 + ], + "type": "text", + "content": "84. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 303, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 480, + 335 + ], + "type": "text", + "content": "85. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: Materobot: Material recognition in wearable robotics for people with visual impairments. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 335, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 369 + ], + "type": "text", + "content": "86. Zheng, X., Pan, T., Luo, Y., Wang, L.: Look at the neighbor: Distortion-aware unsupervised domain adaptation for panoramic semantic segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 402 + ], + "type": "text", + "content": "87. Zheng, X., Zhu, J., Liu, Y., Cao, Z., Fu, C., Wang, L.: Both style and distortion matter: Dual-path unsupervised domain adaptation for panoramic semantic segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 435 + ], + "type": "text", + "content": "88. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentaion. arXiv preprint arXiv:2311.03352 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 480, + 456 + ], + "type": "text", + "content": "89. Zhou, Q., Liu, Y., Yu, C., Li, J., Wang, Z., Wang, F.: LMSeg: Language-guided multi-dataset segmentation. In: ICLR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 456, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 480, + 479 + ], + "type": "text", + "content": "90. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable ConvNets V2: More deformable, better results. In: CVPR (2019)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Panoramic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_content_list.json b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c801e268760bf280d3bd3ca10a1c1ee5865506a2 --- /dev/null +++ b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_content_list.json @@ -0,0 +1,1619 @@ +[ + { + "type": "text", + "text": "Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation", + "text_level": 1, + "bbox": [ + 235, + 140, + 769, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengfei Wang $^{1,2}$ , Yuxi Wang $^{2}$ , Shuai Li $^{1}$ , Zhaoxiang Zhang $^{1,2,3,4(\\boxtimes)}$ , Zhen Lei $^{1,2,3,4}$ , and Lei Zhang $^{1(\\boxtimes)}$", + "bbox": [ + 217, + 210, + 782, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong Polytechnic University", + "bbox": [ + 362, + 253, + 640, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Centre for Artificial Intelligence and Robotics, HKISI, CAS", + "bbox": [ + 294, + 268, + 705, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA", + "bbox": [ + 241, + 282, + 759, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)", + "bbox": [ + 215, + 296, + 784, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pengfei.wang@connect.polyu.hk, zhaoxiang.zhang@ia.ac.cn,", + "bbox": [ + 282, + 311, + 715, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cslzhang@comp.polyu.edu.hk", + "bbox": [ + 398, + 325, + 602, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The scarcity of large-scale 3D-text paired data poses a great challenge on open vocabulary 3D scene understanding, and hence it is popular to leverage internet-scale 2D data and transfer their open vocabulary capabilities to 3D models through knowledge distillation. However, the existing distillation-based 3D scene understanding approaches rely on the representation capacity of 2D models, disregarding the exploration of geometric priors and inherent representational advantages offered by 3D data. In this paper, we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained models. Specifically, we first design a geometry guided distillation module to distill knowledge from 2D models, and then leverage the 3D geometric priors to alleviate the inherent noise in 2D models and enhance the representation learning process. Due to the advantages of 3D representation, the performance of the distilled 3D student model can significantly surpass that of the 2D teacher model. This motivates us to further leverage the representation advantages of 3D data through self-distillation. As a result, our proposed GGSD approach outperforms the existing open vocabulary 3D scene understanding methods by a large margin, as demonstrated by our experiments on both indoor and outdoor benchmark datasets. Codes are available at https://github.com/Wang-pengfei/GGSD.", + "bbox": [ + 259, + 375, + 743, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D scene understanding $\\cdot$ Open vocabulary $\\cdot$ Self-distillation", + "bbox": [ + 261, + 678, + 740, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 718, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding plays a vital role in various real-world applications, including robot manipulation, virtual reality, and autonomous driving, etc. Traditional 3D scene understanding methods mostly employ supervised training on labeled 3D datasets with a limited number of classes [18,42,56]. These approaches, however, face difficulties when they encounter unknown categories in open environments. Such a limitation in practical applications motivates researchers to", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a94125fd6d8516321dc0fe7eff58ddf59c6524eb05cd7782aaa5dbadc76ec7e7.jpg", + "image_caption": [ + "Fig. 1: Example of the problems when utilizing 2D models, e.g., LSeg [34], for pixel-to-point distillation. We can see the noise caused by the factors such as occlusion, lighting variations and view angles in 2D images. Therefore, only employing the 2D model in distillation brings errors to the 3D model, limiting its 3D representation capacity." + ], + "image_footnote": [], + "bbox": [ + 349, + 148, + 656, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "explore new 3D scene understanding models with open vocabulary capability. Inspired by the success of vision-language foundation models like CLIP [48], which aligns language features to image features, in various image understanding tasks [19,28,34,50,61,67,69], an intuitive strategy is to leverage a substantial amount of paired 3D point cloud and text data for training. Unfortunately, the high cost of annotating 3D datasets (e.g., annotating a scene with 20 categories takes approximately 22.3 minutes [12]) makes it impractical to manually label enough data with numerous real-world categories.", + "bbox": [ + 212, + 474, + 787, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An alternative yet more practical approach is to leverage internet-scale 2D data to acquire open vocabulary capability, and then transfer this capability to 3D scene understanding models via distillation [6, 12, 45, 62, 66]. For instance, PLA [12] utilizes a 2D caption model [58] to generate textual descriptions for images, which are then refined using GPT2 [1]. Subsequently, these descriptions are projected onto the 3D point cloud, and the vision-language knowledge can be passed to the 3D model via 3D-text contrastive learning. The recent state-of-the-art OpenScene [45] utilizes 2D open-vocabulary segmentors such as LSeg [34] and OpenSeg [17] to extract pixel-level 2D embeddings for the 3D point cloud. The ability of pixel-text alignment is then transferred to align 3D point to text via feature distillation, enabling 3D semantic-level scene understanding.", + "bbox": [ + 212, + 597, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though the above mentioned distillation-based methods have shown promising results in 3D scene understanding, they basically imitate the 2D models and inherit their limitations, restricting the upper bound of representation power. As shown in Fig. 1, the 2D model may suffer from errors due to issues such as occlusion, lighting variations, and viewing angles. The misclassification of ex", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tracted 2D features leads to erroneous supervisory signals during the distillation process of 3D point cloud learning. Instead of solely imitating 2D models, our objective is to harness both the geometric priors in 3D point clouds and the representational advantages of 3D data to guide the knowledge distillation process from pre-trained 2D models.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve the above mentioned objectives, in this paper we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained open vocabulary segmentors. Given a sparse point cloud composed of multiple semantic categories, geometrically homogeneous neighborhoods typically correspond to the same semantic information. Based on this observation, we first design a geometry guided distillation module to learn knowledge from 2D pre-trained models. Specifically, we generate geometrically homogeneous neighborhoods as superpoints and utilize their semantic consistency to guide the distillation process. The utilization of 3D geometric priors could mitigate the inherent noise of the 2D model and enhance the representation learning capability. Furthermore, due to the representation advantages of 3D data, which are less affected by factors such as lighting and viewing angles, the distilled 3D student model can exhibit superior performance compared to the 2D teacher model. This motivates us to further distill knowledge within the 3D network itself through self-distillation. We employ an exponential moving average (EMA) model to predict each 3D point cloud and perform voting within the superpoint to provide supervision signals for the online 3D model. This approach ensures that the online model could receive reliable and accurate guidance by leveraging the consensus among multiple-point predictions. Finally, our proposed GGSD method could learn robust 3D representations from the 3D distillation model itself, achieving significantly superior performance to the 2D teacher model.", + "bbox": [ + 212, + 222, + 785, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this work are summarized as follows. First, we analyze the limitations of pixel-point distillation methods and propose leveraging the advantages of 3D data representation through self-distillation, instead of relying solely on 2D representations. Secondly, we utilize 3D geometric priors to mitigate the noise in pixel-point distillation process, thereby enhancing the capability of representation learning. Third, we conduct experiments on both indoor and outdoor datasets, including ScanNet [10], Matterport [3], and nuScenes [2]. The results demonstrate that GGSD sets a new state-of-the-art for 3D open vocabulary scene understanding in both indoor and outdoor datasets.", + "bbox": [ + 212, + 555, + 787, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 715, + 387, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Closed-set 3D Scene Understanding. Previous studies in closed-set 3D scene understanding [9,21,24,26,27,36,43,46,52] have led to significant advancements across diverse 3D scene understanding benchmarks, encompassing tasks such as 3D object classification, 3D object detection, 3D semantic and instance segmentation, etc. Some researches [4] have explored the use of 2D pre-trained models, such as MAE [22], for pre-training point encoders. However, they mainly em", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "phasize on leveraging pre-training to enhance supervised 3D downstream tasks, rather than addressing open-vocabulary query scenarios. Existing closed-set 3D scene understanding methods excel in specific categories but often struggle to recognize new categories in open-set scenarios. Our objective is to develop a solution to effectively recognize and comprehend new categories in open-set environments, thus overcoming this limitation.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Vocabulary 2D Scene Understanding. Traditional closed-set scene understanding has seen significant performance improvements [8,14,57,59,60,65], while open-vocabulary scene understanding is also beginning to flourish. Prior studies have explored knowledge distillation from visual-language models, such as CLIP, to transfer their open vocabulary capabilities in 2D vision tasks [13,15, 16,20,35,37,40,51,64,68]. Recent research endeavors have aimed to establish correlations between dense image features and embeddings derived from language models [17, 19, 32, 34, 38, 49]. Due to their pixel-level text alignment capabilities, 2D open vocabulary scene understanding models, such as OpenSeg [17] and LSeg [34], can serve as teacher models for distillation of 3D models compared to CLIP. In this study, we rely on pre-trained 2D models with open vocabulary capabilities, specifically OpenSeg and LSeg, and propose a novel geometry-guided self-distillation strategy to conduct 3D scene-level understanding tasks without the need for any ground-truth 2D or 3D training data.", + "bbox": [ + 212, + 238, + 789, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Vocabulary 3D Scene Understanding. Recent works on open-vocabulary 3D scene understanding can be classified into two categories. The first category [25,30,39,47], represented by methods like LERF [30], focuses on optimization-based approaches that utilize techniques such as NeRF [41] or 3DGS [29] to distill 2D features into 3D scenes. However, these methods lack generalization capability when confronted with new scenes. The second category [6,12,45,54], represented by methods like OpenScene [45], is based on learning-based approaches. This paper falls in the second category. CLIP2Scene [6] leverages CLIP image model to extract 2D image features and utilizes textual category features to classify pixels, which are then projected onto 3D point clouds. OpenScene [45] aggregates pixel features from multiple view images on every 3D point, achieving state-of-the-art open vocabulary 3D scene understanding performance. Despite the promising results of these methods, their simple imitation of 2D models limits the representational capacity of 3D models.", + "bbox": [ + 212, + 452, + 810, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 694, + 330, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An overview of our approach is illustrated in Fig. 2. Firstly, we follow OpenScene [45] to create pixel-point feature pairs for the 3D point cloud (see Section 3.1). Subsequently, we analyze the limitations of pixel-point distillation methods (see Section 3.2). Then, we propose a geometry guided distillation module to enhance the learning of 3D representations from the 2D model (see Section 3.3). Finally, we introduce a geometry guided self-distillation module to further distill the knowledge within the 3D network itself (see Section 3.4).", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d58c3b1a390d9f2b02b655de569ad1d524f644ccdb63d53b087e25c10197cbd7.jpg", + "image_caption": [ + "Fig. 2: Overview of our Geometry Guided Self-Distillation (GGSD) framework, which consists of two main components: Geometry Guided Distillation and Self-Distillation. The first module leverages 3D geometric priors to mitigate the inherent noise in 2D models. In the second module, the 3D network learns from its own predictions. This self-distillation step allows the 3D network to continuously enhance its understanding and representation capabilities. Overall, the GGSD framework combines geometry guidance and self-distillation to fully exploit the valuable information hidden in 3D data, rather than solely relying on 2D pre-trained models." + ], + "image_footnote": [], + "bbox": [ + 232, + 146, + 772, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Creating Pixel-Point Feature Pairs", + "text_level": 1, + "bbox": [ + 215, + 547, + 549, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following OpenScene [45], our approach begins by extracting dense per-pixel embeddings from RGB images using a pre-trained 2D vision language segmentation model, such as LSeg [34] or OpenSeg [17]. After obtaining the pixel features, we can establish correspondences between each point in the point cloud and the pixels in different RGB images by utilizing the intrinsic matrix and the world-to-camera extrinsic matrix. Next, we employ a multi-view fusion strategy to combine the features from multiple views into a single feature vector, denoted as $\\mathbf{f}^{2\\mathrm{D}} = \\phi (\\mathbf{f}_1,\\dots ,\\mathbf{f}_K)$ , for each point cloud $\\mathbf{p}$ . This fusion is achieved using an average pooling operator $\\phi$ over the $K$ -view features. After obtaining the fused 2D features for each of the M points in the scene, we obtain a 2D pixel feature set $\\mathbf{F}^{2\\mathrm{D}} = \\mathbf{f}_1^{2\\mathrm{D}},\\dots ,\\mathbf{f}_M^{2\\mathrm{D}}\\in \\mathbb{R}^{M\\times C}$ corresponding to the point cloud scene. Finally, these embedded features are back-projected into the 3D point cloud scene. For more details, please refer to OpenScene [45].", + "bbox": [ + 212, + 570, + 787, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Limitations of Pixel-Point 3D Distillation", + "text_level": 1, + "bbox": [ + 215, + 786, + 604, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the text-image alignment capability of CLIP, the 2D features $\\mathbf{F}^{2\\mathrm{D}}$ can be directly utilized for language-driven open vocabulary scene understanding.", + "bbox": [ + 212, + 808, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To transfer 2D vision-language knowledge to a 3D point cloud network, the distillation approach has been popularly used. One up-to-date pixel-point 3D distillation work is OpenScene [45]. Specifically, OpenScene aims to learn a 3D encoder that extracts 3D features from each point cloud, which are expected to be similar to their corresponding 2D pixel features. This allows the 3D encoder to distill the open vocabulary capability of the 2D model. To this end, OpenScene utilizes cosine distance to perform pixel-point distillation:", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = 1 - \\cos (\\mathbf {F} ^ {\\mathrm {2 D}}, \\mathbf {F} ^ {\\mathrm {3 D}}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 263, + 785, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, the current pixel-to-point distillation methods are mostly a simplistic imitation of the 2D models, which are limited in extracting useful information from 3D point data. Firstly, relying solely on the projection relationship between point clouds and pixels, these approaches fail to fully exploit the geometric priors inherent in 3D data, leading to suboptimal performance. In this paper, we propose to incorporate 3D geometric priors throughout the entire distillation process to address the issue of noise in 2D models by constraining semantic consistency. Secondly, the current pixel-point distillation methods fail to fully exploit the representational advantage offered by 3D point clouds. It has been observed that the distilled 3D model can outperform the directly projected 2D models. This can be attributed to the inherent capability of 3D data to handle challenges like lighting variations, occlusion, and viewpoint changes, which are prevalent in 2D representations. Therefore, the accuracy of 3D model segmentation results, particularly for foreground classes, can be significantly improved. This observation provides a strong motivation for us to further leverage the advantages of 3D representation within the 3D network itself via self-distillation, rather than relying solely on information from the 2D model.", + "bbox": [ + 212, + 292, + 787, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Geometry Guided Distillation", + "text_level": 1, + "bbox": [ + 215, + 571, + 508, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first design a geometry guided distillation method to learn knowledge from 2D models. Given a sparse point cloud composed of multiple semantic categories, we can decompose it into many simple parts, i.e., superpoints, according to their geometric position in the space. As shown in Fig. 3, the superpoints for an indoor room are constructed by VCCS [44]. The details for creating superpoints can be found in the supplementary file. For a specific input point cloud $\\mathbf{P} \\in \\mathbb{R}^{M \\times 3}$ , its superpoints are denoted as $\\{\\tilde{\\mathbf{p}}_1 \\cdots \\tilde{\\mathbf{p}}_n \\cdots \\tilde{\\mathbf{p}}_N\\}$ , where each superpoint $\\tilde{\\mathbf{p}}_n$ consists of a small subset of original point cloud $\\mathbf{p}$ . These superpoints are geometrically simple and usually cannot cover objects of different classes. Ideally, each superpoint as a whole should belong to the same semantic category. The simplicity and coherence of superpoints naturally facilitate the implementation of semantic consistency constraints.", + "bbox": [ + 212, + 598, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To mitigate the noise issue in 2D models, we can leverage the semantic consistency provided by 3D geometric priors. In particular, for a specific input point cloud $\\mathbf{P}$ , we have its neural features $\\mathbf{F}^{3\\mathrm{D}} = \\{\\mathbf{f}_1^{3\\mathrm{D}}, \\dots, \\mathbf{f}_M^{3\\mathrm{D}}\\}$ , the corresponding pixel fusion features $\\mathbf{F}^{2\\mathrm{D}} = \\{\\mathbf{f}_1^{2\\mathrm{D}}, \\dots, \\mathbf{f}_M^{2\\mathrm{D}}\\} \\in \\mathbb{R}^{M \\times C}$ and superpoints", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3341e6b5f098d0e334bb0f17ea4fd7d4fe1fb622bda8ad9b1889ea59436c3811.jpg", + "image_caption": [ + "Input Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 294, + 146, + 480, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a132b87accc65d2d1bcca8c038c6fe1456a91835910cc43bfaef98d3296ae4d.jpg", + "image_caption": [ + "Superpoints", + "Fig. 3: Example of the superpoints generated using VCCS [44] on ScanNet scene point clouds. Each colored patch corresponds to a distinct superpoint, showcasing a simple geometric structure. Additionally, the points within each superpoint consistently exhibit a shared semantic category." + ], + "image_footnote": [], + "bbox": [ + 513, + 146, + 704, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\{\\tilde{\\mathbf{p}}_1 \\cdots \\tilde{\\mathbf{p}}_n \\cdots \\tilde{\\mathbf{p}}_N\\}$ . Firstly, we compute the mean 2D pixel features for superpoints, denoted as $\\mathbf{F}_{sp}^{2D} = \\{\\tilde{\\mathbf{f}}_1^{2D}, \\cdots, \\tilde{\\mathbf{f}}_N^{2D}\\}$ :", + "bbox": [ + 214, + 390, + 782, + 422 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {f}} _ {n} ^ {2 \\mathrm {D}} = \\frac {1}{Q} \\sum_ {q = 1} ^ {Q} \\mathbf {f} _ {q} ^ {2 \\mathrm {D}}, \\quad \\mathbf {f} _ {q} ^ {2 \\mathrm {D}} \\in \\mathbb {R} ^ {1 \\times C}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 434, + 784, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $Q$ is the total number of 3D points within a superpoint $\\tilde{\\mathbf{p}}_n$ , and $\\mathbf{f}_q^{2\\mathrm{D}}$ is the feature vector for the $q^{th}$ pixel that corresponds to the 3D point in the superpoint. Similarly, we can employ the same approach to calculate the mean 3D point features $\\mathbf{F}_{sp}^{3\\mathrm{D}} = \\{\\tilde{\\mathbf{f}}_1^{3\\mathrm{D}}, \\dots, \\tilde{\\mathbf{f}}_N^{3\\mathrm{D}}\\}$ .", + "bbox": [ + 212, + 489, + 782, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To enforce the superpoint feature $\\mathbf{F}_{sp}^{3D}$ to be consistent with the 2D fused features $\\mathbf{F}_{sp}^{2D}$ , we use a cosine similarity loss as that in pixel-point distillation:", + "bbox": [ + 212, + 554, + 782, + 588 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s p} = 1 - \\cos \\left(\\mathbf {F} _ {s p} ^ {\\mathrm {2 D}}, \\mathbf {F} _ {s p} ^ {\\mathrm {3 D}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 598, + 784, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The final optimization objective for geometry guided distillation is as follows:", + "bbox": [ + 214, + 628, + 771, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\mathcal {L} _ {p} + \\mathcal {L} _ {s p}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 655, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Geometry Guided Self-Distillation", + "text_level": 1, + "bbox": [ + 214, + 693, + 542, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With geometry guided distillation, the performance of the 3D model can surpass that of the directly projected 2D model. This inspires us to further design a self-distillation module to fully leverage the inherent advantages of 3D data in representation. Recent methods such as CNS [5] and CLIP2Scene [6] typically transfer the pixel-level pseudo-labels from the 2D image to the 3D point cloud using a 2D-3D calibration matrix. In contrast to existing methods that rely on projecting from 2D, we introduce a geometry guided self-distillation approach to distill knowledge from the 3D model itself. Specifically, we generate predictions", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "directly on 3D point clouds instead of projecting 2D predictions back to 3D, as our 3D model possesses enhanced feature representation capabilities. Specifically, given the 3D point features $\\mathbf{F}^{3\\mathrm{D}} = \\{\\mathbf{f}_1^{3\\mathrm{D}},\\dots ,\\mathbf{f}_M^{3\\mathrm{D}}\\} \\in \\mathbb{R}^{M\\times C}$ , we assign pseudolabels to each point $\\mathbf{f}_n^{3\\mathrm{D}}$ as follows:", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {\\hat {t}} = \\underset {l} {\\arg \\max } \\psi \\left(\\mathbf {f} _ {n} ^ {\\mathrm {3 D}}, \\mathbf {f} _ {l} ^ {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 220, + 785, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_l^t$ is CLIP's text embedding of the $l$ -th class name with a pre-defined template. $\\psi$ is the dot product operation. Furthermore, in the self-distillation stage, we also incorporate 3D geometric priors to enhance the semantic consistency of the predictions.", + "bbox": [ + 212, + 258, + 787, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, the research findings from self-supervised methods [7,23] have highlighted that directly using the network's predictions as supervision signals can lead to model collapse due to overconfidence. To address this issue, we leverage the predictions from the exponential moving average (EMA) model as supervision signals to train the 3D point cloud network, aiming to enhance stability and mitigate the impact of overconfidence. Specifically, we use the features extracted by the EMA model to calculate the similarity with the text feature according to Equation 5, resulting in generating a point-text pair for each point cloud.", + "bbox": [ + 212, + 320, + 787, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although using EMA model can prevent mode collapse, it is still inevitable to have noise in the prediction results. To mitigate this issue, we incorporate the semantic consistency of 3D geometric priors during the self-distillation process. Specifically, we implement a voting mechanism within each superpoint to refine the prediction results. After assigning the most similar textual feature to each point cloud according to Equation 5, we calculate the proportion of textual features within each superpoint. We then modify the text corresponding to all point clouds within that superpoint to match the textual feature with the highest proportion. This correction process effectively mitigates the impact of noise and improves the accuracy of prediction. The voting outcome generates point-language pairs for each point cloud, then we can pull paired point and language features together while pushing unmatched features away through contrastive loss. The geometry guided self-distillation loss is as follows:", + "bbox": [ + 212, + 441, + 789, + 637 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s d} = - \\log \\frac {\\exp \\left(\\mathbf {f} ^ {\\mathrm {3 D}} \\cdot \\mathbf {f} ^ {\\hat {t}} / \\tau\\right)}{\\sum_ {i = 1} ^ {n _ {t}} \\exp \\left(\\mathbf {f} ^ {\\mathrm {3 D}} \\cdot \\mathbf {f} _ {i} ^ {t} / \\tau\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 641, + 785, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{f}^t$ represents the textual features corresponding to the point cloud $\\mathbf{f}^{3\\mathrm{D}}$ ; $n_t$ is the number of point-language pairs and $\\tau$ is a temperature factor.", + "bbox": [ + 212, + 679, + 787, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Training and Inference", + "text_level": 1, + "bbox": [ + 214, + 736, + 450, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training. The training process comprises two stages. In the initial stage, we employ geometry guided distillation to train the 3D network for 70 epochs, enabling it to acquire knowledge from the 2D model. Then, we incorporate the GGSD module into an additional 30 training epochs. The total number of training epochs is consistent with OpenScene.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference. Using any per-point feature $\\mathbf{f}^{3D}$ and CLIP text features obtained from a set of text prompts, we can estimate their similarities by calculating the cosine similarity score between them. This similarity score is utilized in all of our scene understanding tasks. In all experiments, unless otherwise stated, we exclusively use pure 3D point clouds for predictions during the inference stage, without employing the 2D-3D ensemble strategy.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 258, + 375, + 275 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 214, + 289, + 413, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. To evaluate the effectiveness of our method, we conduct experiments on three widely used public benchmarks: ScanNet [10], Matterport [3], and nuScenes [2]. The first two datasets provide RGBD images and 3D meshes of indoor scenes, while the last one offers LIDAR scans of outdoor scenes. Additionally, Matterport3D is a challenging dataset with intricately detailed scenes, which allows us to thoroughly examine the performance of open-vocabulary queries across different datasets.", + "bbox": [ + 212, + 311, + 787, + 416 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the settings outlined in OpenScene [45], utilize MinkowskiNet18A [9] as the 3D backbone network, and employ a simple prompt engineering technique to extract CLIP text features. Specifically, we modify the text prompts for each object class \"XX\" to \"a XX in a scene\", such as \"a chair in a scene\". On the ScanNet and Matterport3D datasets, we employ a voxel size of $2\\mathrm{cm}$ with MinkowskiNet while utilizing a voxel size of $5\\mathrm{cm}$ on the nuScenes dataset. Based on the findings of OpenScene, we employ LSeg [34] for indoor datasets and OpenSeg [17] for outdoor dataset as the 2D teacher model. The temperature $\\tau$ in Eq. 6 is empirically set as 0.01. We utilize the Adam optimizer with an initial learning rate of $1e-4$ for the entire training process. We use a batch size of 8 on ScanNet and Matterport3D with a single NVIDIA A100 GPU (80G). On nuScenes, we use a batch size of 16 with 4 A100 GPUs. It takes around 24 hours to train and 0.1 second for inference. More details can be found in the supplementary file.", + "bbox": [ + 212, + 417, + 787, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 650, + 372, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted a series of experiments, including traditional closed-set scenarios, in-domain open vocabulary scenarios, and out-of-domain open vocabulary scenarios to test the effectiveness of the proposed methods in various 3D scene understanding tasks.", + "bbox": [ + 212, + 672, + 785, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison on Standard Benchmarks. We compare our approach with both fully-supervised and open vocabulary methods on all classes of the ScanNet [10] validation set and nuScenes [2] validation set. As shown in the bottom panel of Table 1, our approach, GGSD, consistently outperforms recent open vocabulary methods PLA [12], CLIP2Scene [6], CNS [5] and CLIP-FO3D [66] by a large margin on both datasets. Additionally, our method exhibits significant performance improvements compared to the baseline model OpenScene [45].", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/481394f58f8c278c0317a58ba0bc117b25c63242b4596442b094b528a889f02a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScanNet [10]nuScenes [2]
mIoUmAccmIoUmAcc
Fully-supervised methods
TangentConv [55]40.9---
TextureNet [27]54.8---
ScanComplete [11]56.6---
DCM-Net [53]65.8---
Mix3D [43]73.6---
VMNet [26]73.2---
LidarMultiNet [63]--82.0-
MinkowskiNet [9]69.077.578.083.7
Open vocabulary methods
MSeg [33] Voting45.654.431.036.9
PLA [12]17.733.5--
CLIP2Scene [6]25.1-20.8-
CNS [5]26.8-33.5-
CLIP-FO3D [66]30.249.1--
OpenScene†[45]54.266.642.161.8
OpenScene‡[45]52.963.242.957.1
GGSD (Ours)56.568.646.159.2
", + "bbox": [ + 333, + 143, + 674, + 414 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Comparisons on Standard Benchmarks. The numbers in gray are obtained from previous papers. The symbol $\\dagger$ indicates the use of the 2D-3D ensemble strategy, while $\\ddagger$ denotes the pure 3D results obtained from the official open-source model. Our method surpasses the current state-of-the-art by a large margin in both indoor and outdoor scenarios without using the 2D-3D ensemble strategy.", + "bbox": [ + 212, + 424, + 787, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Specifically, on the ScanNet dataset, GGSD achieves a $3.6\\%$ increase in mIoU and a $5.4\\%$ increase in mACC. On the nuScenes dataset, GGSD demonstrates a $3.2\\%$ increase in mIoU and a $2.1\\%$ increase in mAcc. Except for OpenScene, the other methods in Table 1 do not employ the 2D-3D ensemble strategy due to the significant computational and storage costs associated with maintaining a 2D segmentation network and processing a large number of images per scene during testing. It is worth noting that our method using only point clouds still outperforms OpenScene with the ensemble strategy. Given that our method is evaluated on both indoor and outdoor scenes, and it even surpasses OpenScene with ensemble strategy, these results clearly demonstrate the effectiveness and superiority of GGSD.", + "bbox": [ + 212, + 546, + 787, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison on Open-Vocabulary Scenarios. The standard ScanNet benchmark only contains a small vocabulary of 20 classes. To evaluate the open vocabulary capability of GGSD, we expand the original vocabulary size by using the NYU-40 label set. We remove NYU-40 labels that lack specific semantics, such as \"other structure\", \"other furniture\", and \"other prop\". The remaining categories are evenly divided into Head, Common, and Tail based on the number of samples in each category (see supplementary file for specific division). We present the quantitative results on these three categories in Table 2.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/062a5c34ec4c7443834c3bb5e5e97864a629c85f59a39e9c7ae37d919b1bf577.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HeadCommonTail
mIoUmAccmIoUmAccmIoUmAcc
Inference by feature projection
MaskCLIP-3D19.833.113.326.87.515.3
OpenScene‡[45]64.476.827.835.114.921.4
GGSD (Ours)66.878.830.140.916.024.2
", + "bbox": [ + 313, + 143, + 692, + 241 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/62a6e1b1428e20312e3bb60da88e1d51f95b6e1cb6403e287162532565196370.jpg", + "table_caption": [ + "Table 2: Comparison on open-vocabulary scenarios. Open-vocabulary semantic segmentation on ScanNet with extended labels from the NYU label set. Categories are divided into Head, Common and Tail by point numbers. The numbers in gray are obtained from previous papers. The symbol $\\ddagger$ denotes the pure 3D point results obtained from the official open-source model." + ], + "table_footnote": [], + "table_body": "
Matterport21Matterport40Matterport80Matterport160
mIoUmAccmIoUmAccmIoUmAccmIoUmAcc
OpenScene‡[45]36.048.021.127.510.813.96.08.1
GGSD (Ours)40.154.422.831.611.916.26.39.6
", + "bbox": [ + 263, + 348, + 743, + 414 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Comparison on cross-domain scenarios. Both GGSD and OpenScene are trained on ScanNet, and zero-shot tested on the Matterport3D dataset. The symbol $\\ddagger$ denotes the pure 3D results obtained from the official open-source model. $K = 21$ is derived from the original Matterport3D benchmark. For $K = 40,80,160$ , we use the $K$ most common categories from the NYU label set provided in the benchmark.", + "bbox": [ + 212, + 424, + 787, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "GGSD achieves superior results over OpenScene in all categories, with a specific improvement of $2.4\\%$ and $2.3\\%$ in mIOU for the Head and Common categories, respectively. Since the categories in the Common set typically consist of objects with smaller sizes (e.g., night stand and floor mat), the $2.3\\%$ mIOU improvement demonstrates the stronger open vocabulary capability of our method. However, for the Tail category, although GGSD surpasses OpenScene by $2.8\\%$ in mACC, the performance of both OpenScene and GGSD is not satisfactory. This can be attributed to the small sizes and the limited number of samples in these categories (e.g., lamp, box, bag, and book).", + "bbox": [ + 212, + 537, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on Cross-Domain Scenarios. Our GGSD method has demonstrated excellent potential in solving open vocabulary scene understanding tasks within a specific domain. However, the exploration of open vocabulary learners that can transfer across datasets is also worth investigation, as they face the dual challenges of category and data distribution. We conduct zero-shot domain transfer experiments by training the model on ScanNet [10] and testing it on the Matterport3D [3] dataset without fine-tuning. It should be noted that as $K$ increases, the test set will contain more small-sized objects. As shown in Table 3, our GGSD method consistently outperforms OpenScene in Matterport21, Matterport40, Matterport80, and Matterport160, demonstrating its strong open vocabulary capability and generalization ability.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/72fc6aa754e0326cfae86d715bc71cc7176d069e957960012d6facd3e3f2c581.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
2D Fusion Projection50.062.732.340.0
Pixel-Point Distillation52.963.236.148.0
Geometry Guided Distillation53.565.036.749.3
Self-Distillation56.168.239.053.3
Geometry Guided Self-distillation56.568.640.154.4
", + "bbox": [ + 290, + 143, + 715, + 243 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e41c89647602306f1436d22fdd367e587988c65d78723aa22801614a484d6b5a.jpg", + "table_caption": [ + "Table 4: Ablation study of GGSD. Comparison of semantic segmentation performance of each component of our GGSD method." + ], + "table_footnote": [], + "table_body": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Geometry Guided Distillation54.065.036.749.3
Fixed 2D model51.964.634.245.7
Fixed 3D model55.065.837.349.8
EMA model56.168.239.053.3
", + "bbox": [ + 290, + 306, + 715, + 400 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5: Comparisons of the 3D supervision signals provided by different models.", + "bbox": [ + 225, + 410, + 774, + 425 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies and Analysis", + "text_level": 1, + "bbox": [ + 215, + 467, + 511, + 483 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we investigate our key components through an in-depth ablation study. The experiments are conducted on the ScanNet dataset and subsequently tested on both ScanNet and Matterport3D datasets.", + "bbox": [ + 212, + 492, + 784, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Component Analysis. As shown in Table 4, even without utilizing 3D geometric priors, pixel-point distillation can outperform 2D direct projection, demonstrating the advantage of 3D representation. By incorporating 3D geometric priors, the geometry guided distillation method further enhances performance, resulting in a $0.6\\%$ increase in mIoU and a $1.8\\%$ increase in mAcc on the ScanNet dataset. Through self-distillation, we achieve a notable enhancement in performance. Specifically, on the ScanNet dataset, there is a substantial increase of $2.6\\%$ in mIoU and $3.2\\%$ in mAcc. On the Matterport3D dataset, the mIoU is increased by $2.3\\%$ , and the mAcc is improved by $4.0\\%$ . Finally, the utilization of the voting mechanism within the superpoints leads to improved performance by mitigating the impact of noise through the refinement model's predictions. It is noteworthy that our GGSD method demonstrates substantial improvements compared to the 2D teacher model, i.e., 2D Fusion Projection, achieving a $+6.5\\%$ mIoU and $+5.9\\%$ mAcc on the ScanNet dataset, as well as a $+7.8\\%$ mIoU and $+14.4\\%$ mAcc on the Matterport3D dataset. These results provide strong evidence for the effectiveness of our method.", + "bbox": [ + 212, + 537, + 785, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The Effectiveness of the EMA Model. In the second stage, we employ the EMA model to provide supervised signals for self-distillation. In Table 5, we investigate the performance of different models in providing supervised signals. When using the fixed 2D model for pixel prediction and projecting it onto", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0a8f265d939766f5628b723b90e255b0fab830d5a68741a1efac1615b16708b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Pixel-Point Distillation52.963.236.148.0
SAM [31] Refine53.063.335.948.4
Geometry Guided Distillation53.565.036.749.3
", + "bbox": [ + 277, + 143, + 730, + 227 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Comparisons between the utilization of geometric priors and SAM [31].", + "bbox": [ + 228, + 238, + 771, + 253 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "points, we observe lower performance compared to geometry-guided distillation. This can be attributed to the fact that the 2D model provides more erroneous supervised signals. Furthermore, using the fixed first-stage distilled 3D model only results in marginal performance improvement. However, when utilizing the EMA model, we achieve the best performance, primarily because the EMA model gradually obtains more accurate prediction results as training progresses.", + "bbox": [ + 212, + 296, + 782, + 387 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Effectiveness of 3D Geometric Priors. The recent work CNS [5] utilized the foundation model SAM [31] to refine pseudo-labels from the 2D model, achieving performance improvements on a lower baseline (27.3% mIoU in ScanNet). We also apply this approach to LSeg. Specifically, we follow the official implementation of SAM to generate masks for all images and perform average pooling within each mask to refine them. Finally, we utilize the extracted refined features for distillation. However, as shown in Table 6, this approach yields minimal improvements. Conversely, leveraging 3D geometric priors results in a 0.6% mIoU improvement in ScanNet.", + "bbox": [ + 212, + 387, + 784, + 521 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Results. As shown in Fig. 4, our method successfully segments all categories that exist in closed-set ScanNet benchmarks, such as wall and table. Additionally, our method successfully segments some categories that are not annotated in traditional benchmarks, such as box, lamp, and clothes, demonstrating our model's excellent open vocabulary capability. On the Matterport3D, the GGSD method successfully segments unlabeled television and also identifies mislabeled pillow and coffee table in the ground truth. Furthermore, there have been instances of segmentation failures in Matterport3D. Due to linguistic ambiguity, sofa chairs are sometimes segmented as separate entities, such as sofas and chairs, while coffee tables can be segmented as coffee tables and tables. The misclassification of certain pillow points as sofas may be attributed to their strong color resemblance to sofas. In the case of nuScenes, GGSD also successfully segments unannotated barrier. These qualitative results obtained from GGSD demonstrate its excellent open vocabulary capability. More visualization results can be found in the supplementary file.", + "bbox": [ + 212, + 523, + 785, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 775, + 359, + 791 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this study, we proposed an effective approach called Geometry Guided Self-Distillation (GGSD) to learn superior 3D representations from 2D pre-trained", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/825d679fcf2d5232836ede9f7225d7928ba17b191d3edbd0012a27aad8530cfa.jpg", + "image_caption": [ + "Fig. 4: Qualitative results. We present the qualitative results of 3D semantic segmentation on public indoor and outdoor benchmarks." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 787, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "models. We first introduced a geometry guided distillation module that leveraged 3D geometric priors to alleviate noise in 2D models. Moreover, we harnessed the representation advantages of 3D data through geometry guided self-distillation, resulting in significant improvement in the model performance for 3D scene understanding. Our GGSD method significantly outperformed existing open-vocabulary 3D scene understanding methods on both indoor and outdoor datasets. This highlighted the crucial role of geometric priors and of 3D data representation in distillation-based approaches for 3D scene understanding.", + "bbox": [ + 212, + 565, + 787, + 688 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Though GGSD has achieved promising results in 3D open vocabulary scene understanding tasks, it has two main limitations. Firstly, while geometric priors can significantly improve the model performance on objects with pronounced geometric structures, the improvement can be limited on objects with fewer prominent geometric characteristics. Secondly, similar to other open vocabulary methods, our approach is susceptible to the influence of language ambiguity, as illustrated by the qualitative results in Fig. 4.", + "bbox": [ + 212, + 702, + 787, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the InnoHK program.", + "bbox": [ + 215, + 824, + 741, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Vit-gpt2 image captioning. https://huggingface.co/nlpconnect/vit-gpt2-image-captioning/discussions", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)", + "4. Chen, A., Zhang, K., Zhang, R., Wang, Z., Lu, Y., Guo, Y., Zhang, S.: Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5291-5301 (2023)", + "5. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023)", + "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)", + "9. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)", + "10. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)", + "1. Dai, A., Ritchie, D., Bokeloh, M., Reed, S., Sturm, J., Nießner, M.: Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2018)", + "2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7010-7019 (2023)", + "3. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.C.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022)", + "4. Fan, J., Wang, Y., Guan, H., Song, C., Zhang, Z.: Toward few-shot domain adaptation with perturbation-invariant representation and transferable prototypes. Frontiers of Computer Science 16(3), 163347 (2022)", + "5. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV (2022)" + ], + "bbox": [ + 225, + 179, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. In: ECCV (2022)", + "17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: European Conference on Computer Vision. pp. 540-557. Springer (2022)", + "18. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)", + "19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)", + "20. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)", + "21. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2940-2949 (2020)", + "22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "24. Hu, W., Zhao, H., Jiang, L., Jia, J., Wong, T.T.: Bidirectional projection network for cross dimension scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14373-14382 (2021)", + "25. Hu, X., Wang, Y., Fan, L., Fan, J., Peng, J., Lei, Z., Li, Q., Zhang, Z.: Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857 (2024)", + "26. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15488-15498 (2021)", + "27. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4440-4449 (2019)", + "28. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. arXiv preprint arXiv:2210.01055 (2022)", + "29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "30. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)", + "31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "32. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639 (2022)", + "33. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: Mseg: A composite dataset for multi-domain semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2879-2888 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN", + "35. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)", + "36. Li, J., He, X., Wen, Y., Gao, Y., Cheng, X., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11809-11818 (2022)", + "37. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv abs/2210.04150 (2022)", + "38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)", + "39. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems 36 (2024)", + "40. Lüddecke, T., Ecker, A.S.: Image segmentation using text and image prompts. In: CVPR (2021)", + "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "42. Misra, I., Girdhar, R., Joulin, A.: An End-to-End Transformer Model for 3D Object Detection. In: ICCV (2021)", + "43. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 International Conference on 3D Vision (3DV). pp. 116-125. IEEE (2021)", + "44. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013)", + "45. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023)", + "46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "47. Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: Langsplat: 3d language gaussian splatting. arXiv preprint arXiv:2312.16084 (2023)", + "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "49. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Dense-clip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: 36th Conference on Neural Information Processing Systems (NIPS) (2022)", + "51. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022)", + "52. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022)", + "53. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: Dualconvmesh-net: Joint geodesic and euclidean convolutions on 3d meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8612-8622 (2020)", + "54. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023)", + "55. Tatarchenko, M., Park, J., Koltun, V., Zhou, Q.Y.: Tangent convolutions for dense prediction in 3d. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3887-3896 (2018)", + "56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)", + "57. Wang, H., Shen, Y., Fei, J., Li, W., Wu, L., Wang, Y., Zhang, Z.: Pulling target to source: A new perspective on domain adaptive semantic segmentation. arXiv preprint arXiv:2305.13752 (2023)", + "58. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR abs/2202.03052 (2022)", + "59. Wang, Y., Liang, J., Xiao, J., Mei, S., Yang, Y., Zhang, Z.: Informative data mining for one-shot cross-domain semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1064-1074 (2023)", + "60. Wang, Y., Peng, J., Zhang, Z.: Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9092-9101 (2021)", + "61. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757 (2021)", + "62. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023)", + "63. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidarmulti-net: Towards a unified multi-task network for lidar perception. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3231-3240 (2023)", + "64. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV (2022)", + "65. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023)", + "66. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)", + "68. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N.C.F., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2021)", + "69. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 258 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "GGSD", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_model.json b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3da852fd94b6e59293c8a8ef5ed84b08a773157f --- /dev/null +++ b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_model.json @@ -0,0 +1,2361 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.141, + 0.77, + 0.187 + ], + "angle": 0, + "content": "Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.211, + 0.784, + 0.245 + ], + "angle": 0, + "content": "Pengfei Wang\\(^{1,2}\\), Yuxi Wang\\(^{2}\\), Shuai Li\\(^{1}\\), Zhaoxiang Zhang\\(^{1,2,3,4(\\boxtimes)}\\), Zhen Lei\\(^{1,2,3,4}\\), and Lei Zhang\\(^{1(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.255, + 0.641, + 0.269 + ], + "angle": 0, + "content": "1 The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.269, + 0.707, + 0.282 + ], + "angle": 0, + "content": "2 Centre for Artificial Intelligence and Robotics, HKISI, CAS" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.283, + 0.761, + 0.297 + ], + "angle": 0, + "content": "3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.311 + ], + "angle": 0, + "content": "\\(^{4}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.312, + 0.717, + 0.326 + ], + "angle": 0, + "content": "pengfei.wang@connect.polyu.hk, zhaoxiang.zhang@ia.ac.cn," + }, + { + "type": "text", + "bbox": [ + 0.4, + 0.327, + 0.603, + 0.339 + ], + "angle": 0, + "content": "cslzhang@comp.polyu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.375, + 0.744, + 0.668 + ], + "angle": 0, + "content": "Abstract. The scarcity of large-scale 3D-text paired data poses a great challenge on open vocabulary 3D scene understanding, and hence it is popular to leverage internet-scale 2D data and transfer their open vocabulary capabilities to 3D models through knowledge distillation. However, the existing distillation-based 3D scene understanding approaches rely on the representation capacity of 2D models, disregarding the exploration of geometric priors and inherent representational advantages offered by 3D data. In this paper, we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained models. Specifically, we first design a geometry guided distillation module to distill knowledge from 2D models, and then leverage the 3D geometric priors to alleviate the inherent noise in 2D models and enhance the representation learning process. Due to the advantages of 3D representation, the performance of the distilled 3D student model can significantly surpass that of the 2D teacher model. This motivates us to further leverage the representation advantages of 3D data through self-distillation. As a result, our proposed GGSD approach outperforms the existing open vocabulary 3D scene understanding methods by a large margin, as demonstrated by our experiments on both indoor and outdoor benchmark datasets. Codes are available at https://github.com/Wang-pengfei/GGSD." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.679, + 0.741, + 0.694 + ], + "angle": 0, + "content": "Keywords: 3D scene understanding \\(\\cdot\\) Open vocabulary \\(\\cdot\\) Self-distillation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.377, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3D scene understanding plays a vital role in various real-world applications, including robot manipulation, virtual reality, and autonomous driving, etc. Traditional 3D scene understanding methods mostly employ supervised training on labeled 3D datasets with a limited number of classes [18,42,56]. These approaches, however, face difficulties when they encounter unknown categories in open environments. Such a limitation in practical applications motivates researchers to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.149, + 0.658, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.387, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 1: Example of the problems when utilizing 2D models, e.g., LSeg [34], for pixel-to-point distillation. We can see the noise caused by the factors such as occlusion, lighting variations and view angles in 2D images. Therefore, only employing the 2D model in distillation brings errors to the 3D model, limiting its 3D representation capacity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.596 + ], + "angle": 0, + "content": "explore new 3D scene understanding models with open vocabulary capability. Inspired by the success of vision-language foundation models like CLIP [48], which aligns language features to image features, in various image understanding tasks [19,28,34,50,61,67,69], an intuitive strategy is to leverage a substantial amount of paired 3D point cloud and text data for training. Unfortunately, the high cost of annotating 3D datasets (e.g., annotating a scene with 20 categories takes approximately 22.3 minutes [12]) makes it impractical to manually label enough data with numerous real-world categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.764 + ], + "angle": 0, + "content": "An alternative yet more practical approach is to leverage internet-scale 2D data to acquire open vocabulary capability, and then transfer this capability to 3D scene understanding models via distillation [6, 12, 45, 62, 66]. For instance, PLA [12] utilizes a 2D caption model [58] to generate textual descriptions for images, which are then refined using GPT2 [1]. Subsequently, these descriptions are projected onto the 3D point cloud, and the vision-language knowledge can be passed to the 3D model via 3D-text contrastive learning. The recent state-of-the-art OpenScene [45] utilizes 2D open-vocabulary segmentors such as LSeg [34] and OpenSeg [17] to extract pixel-level 2D embeddings for the 3D point cloud. The ability of pixel-text alignment is then transferred to align 3D point to text via feature distillation, enabling 3D semantic-level scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Though the above mentioned distillation-based methods have shown promising results in 3D scene understanding, they basically imitate the 2D models and inherit their limitations, restricting the upper bound of representation power. As shown in Fig. 1, the 2D model may suffer from errors due to issues such as occlusion, lighting variations, and viewing angles. The misclassification of ex" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "tracted 2D features leads to erroneous supervisory signals during the distillation process of 3D point cloud learning. Instead of solely imitating 2D models, our objective is to harness both the geometric priors in 3D point clouds and the representational advantages of 3D data to guide the knowledge distillation process from pre-trained 2D models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.786, + 0.553 + ], + "angle": 0, + "content": "To achieve the above mentioned objectives, in this paper we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained open vocabulary segmentors. Given a sparse point cloud composed of multiple semantic categories, geometrically homogeneous neighborhoods typically correspond to the same semantic information. Based on this observation, we first design a geometry guided distillation module to learn knowledge from 2D pre-trained models. Specifically, we generate geometrically homogeneous neighborhoods as superpoints and utilize their semantic consistency to guide the distillation process. The utilization of 3D geometric priors could mitigate the inherent noise of the 2D model and enhance the representation learning capability. Furthermore, due to the representation advantages of 3D data, which are less affected by factors such as lighting and viewing angles, the distilled 3D student model can exhibit superior performance compared to the 2D teacher model. This motivates us to further distill knowledge within the 3D network itself through self-distillation. We employ an exponential moving average (EMA) model to predict each 3D point cloud and perform voting within the superpoint to provide supervision signals for the online 3D model. This approach ensures that the online model could receive reliable and accurate guidance by leveraging the consensus among multiple-point predictions. Finally, our proposed GGSD method could learn robust 3D representations from the 3D distillation model itself, achieving significantly superior performance to the 2D teacher model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.693 + ], + "angle": 0, + "content": "The contributions of this work are summarized as follows. First, we analyze the limitations of pixel-point distillation methods and propose leveraging the advantages of 3D data representation through self-distillation, instead of relying solely on 2D representations. Secondly, we utilize 3D geometric priors to mitigate the noise in pixel-point distillation process, thereby enhancing the capability of representation learning. Third, we conduct experiments on both indoor and outdoor datasets, including ScanNet [10], Matterport [3], and nuScenes [2]. The results demonstrate that GGSD sets a new state-of-the-art for 3D open vocabulary scene understanding in both indoor and outdoor datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.716, + 0.388, + 0.733 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Closed-set 3D Scene Understanding. Previous studies in closed-set 3D scene understanding [9,21,24,26,27,36,43,46,52] have led to significant advancements across diverse 3D scene understanding benchmarks, encompassing tasks such as 3D object classification, 3D object detection, 3D semantic and instance segmentation, etc. Some researches [4] have explored the use of 2D pre-trained models, such as MAE [22], for pre-training point encoders. However, they mainly em" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "phasize on leveraging pre-training to enhance supervised 3D downstream tasks, rather than addressing open-vocabulary query scenarios. Existing closed-set 3D scene understanding methods excel in specific categories but often struggle to recognize new categories in open-set scenarios. Our objective is to develop a solution to effectively recognize and comprehend new categories in open-set environments, thus overcoming this limitation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.79, + 0.451 + ], + "angle": 0, + "content": "Open-Vocabulary 2D Scene Understanding. Traditional closed-set scene understanding has seen significant performance improvements [8,14,57,59,60,65], while open-vocabulary scene understanding is also beginning to flourish. Prior studies have explored knowledge distillation from visual-language models, such as CLIP, to transfer their open vocabulary capabilities in 2D vision tasks [13,15, 16,20,35,37,40,51,64,68]. Recent research endeavors have aimed to establish correlations between dense image features and embeddings derived from language models [17, 19, 32, 34, 38, 49]. Due to their pixel-level text alignment capabilities, 2D open vocabulary scene understanding models, such as OpenSeg [17] and LSeg [34], can serve as teacher models for distillation of 3D models compared to CLIP. In this study, we rely on pre-trained 2D models with open vocabulary capabilities, specifically OpenSeg and LSeg, and propose a novel geometry-guided self-distillation strategy to conduct 3D scene-level understanding tasks without the need for any ground-truth 2D or 3D training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.811, + 0.664 + ], + "angle": 0, + "content": "Open-Vocabulary 3D Scene Understanding. Recent works on open-vocabulary 3D scene understanding can be classified into two categories. The first category [25,30,39,47], represented by methods like LERF [30], focuses on optimization-based approaches that utilize techniques such as NeRF [41] or 3DGS [29] to distill 2D features into 3D scenes. However, these methods lack generalization capability when confronted with new scenes. The second category [6,12,45,54], represented by methods like OpenScene [45], is based on learning-based approaches. This paper falls in the second category. CLIP2Scene [6] leverages CLIP image model to extract 2D image features and utilizes textual category features to classify pixels, which are then projected onto 3D point clouds. OpenScene [45] aggregates pixel features from multiple view images on every 3D point, achieving state-of-the-art open vocabulary 3D scene understanding performance. Despite the promising results of these methods, their simple imitation of 2D models limits the representational capacity of 3D models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.331, + 0.71 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "An overview of our approach is illustrated in Fig. 2. Firstly, we follow OpenScene [45] to create pixel-point feature pairs for the 3D point cloud (see Section 3.1). Subsequently, we analyze the limitations of pixel-point distillation methods (see Section 3.2). Then, we propose a geometry guided distillation module to enhance the learning of 3D representations from the 2D model (see Section 3.3). Finally, we introduce a geometry guided self-distillation module to further distill the knowledge within the 3D network itself (see Section 3.4)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.147, + 0.773, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 2: Overview of our Geometry Guided Self-Distillation (GGSD) framework, which consists of two main components: Geometry Guided Distillation and Self-Distillation. The first module leverages 3D geometric priors to mitigate the inherent noise in 2D models. In the second module, the 3D network learns from its own predictions. This self-distillation step allows the 3D network to continuously enhance its understanding and representation capabilities. Overall, the GGSD framework combines geometry guidance and self-distillation to fully exploit the valuable information hidden in 3D data, rather than solely relying on 2D pre-trained models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.549, + 0.55, + 0.563 + ], + "angle": 0, + "content": "3.1 Creating Pixel-Point Feature Pairs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.769 + ], + "angle": 0, + "content": "Following OpenScene [45], our approach begins by extracting dense per-pixel embeddings from RGB images using a pre-trained 2D vision language segmentation model, such as LSeg [34] or OpenSeg [17]. After obtaining the pixel features, we can establish correspondences between each point in the point cloud and the pixels in different RGB images by utilizing the intrinsic matrix and the world-to-camera extrinsic matrix. Next, we employ a multi-view fusion strategy to combine the features from multiple views into a single feature vector, denoted as \\(\\mathbf{f}^{2\\mathrm{D}} = \\phi (\\mathbf{f}_1,\\dots ,\\mathbf{f}_K)\\), for each point cloud \\(\\mathbf{p}\\). This fusion is achieved using an average pooling operator \\(\\phi\\) over the \\(K\\)-view features. After obtaining the fused 2D features for each of the M points in the scene, we obtain a 2D pixel feature set \\(\\mathbf{F}^{2\\mathrm{D}} = \\mathbf{f}_1^{2\\mathrm{D}},\\dots ,\\mathbf{f}_M^{2\\mathrm{D}}\\in \\mathbb{R}^{M\\times C}\\) corresponding to the point cloud scene. Finally, these embedded features are back-projected into the 3D point cloud scene. For more details, please refer to OpenScene [45]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.605, + 0.801 + ], + "angle": 0, + "content": "3.2 Limitations of Pixel-Point 3D Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Due to the text-image alignment capability of CLIP, the 2D features \\(\\mathbf{F}^{2\\mathrm{D}}\\) can be directly utilized for language-driven open vocabulary scene understanding." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "To transfer 2D vision-language knowledge to a 3D point cloud network, the distillation approach has been popularly used. One up-to-date pixel-point 3D distillation work is OpenScene [45]. Specifically, OpenScene aims to learn a 3D encoder that extracts 3D features from each point cloud, which are expected to be similar to their corresponding 2D pixel features. This allows the 3D encoder to distill the open vocabulary capability of the 2D model. To this end, OpenScene utilizes cosine distance to perform pixel-point distillation:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.265, + 0.786, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = 1 - \\cos (\\mathbf {F} ^ {\\mathrm {2 D}}, \\mathbf {F} ^ {\\mathrm {3 D}}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.789, + 0.552 + ], + "angle": 0, + "content": "However, the current pixel-to-point distillation methods are mostly a simplistic imitation of the 2D models, which are limited in extracting useful information from 3D point data. Firstly, relying solely on the projection relationship between point clouds and pixels, these approaches fail to fully exploit the geometric priors inherent in 3D data, leading to suboptimal performance. In this paper, we propose to incorporate 3D geometric priors throughout the entire distillation process to address the issue of noise in 2D models by constraining semantic consistency. Secondly, the current pixel-point distillation methods fail to fully exploit the representational advantage offered by 3D point clouds. It has been observed that the distilled 3D model can outperform the directly projected 2D models. This can be attributed to the inherent capability of 3D data to handle challenges like lighting variations, occlusion, and viewpoint changes, which are prevalent in 2D representations. Therefore, the accuracy of 3D model segmentation results, particularly for foreground classes, can be significantly improved. This observation provides a strong motivation for us to further leverage the advantages of 3D representation within the 3D network itself via self-distillation, rather than relying solely on information from the 2D model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.509, + 0.589 + ], + "angle": 0, + "content": "3.3 Geometry Guided Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.779 + ], + "angle": 0, + "content": "We first design a geometry guided distillation method to learn knowledge from 2D models. Given a sparse point cloud composed of multiple semantic categories, we can decompose it into many simple parts, i.e., superpoints, according to their geometric position in the space. As shown in Fig. 3, the superpoints for an indoor room are constructed by VCCS [44]. The details for creating superpoints can be found in the supplementary file. For a specific input point cloud \\(\\mathbf{P} \\in \\mathbb{R}^{M \\times 3}\\), its superpoints are denoted as \\(\\{\\tilde{\\mathbf{p}}_1 \\cdots \\tilde{\\mathbf{p}}_n \\cdots \\tilde{\\mathbf{p}}_N\\}\\), where each superpoint \\(\\tilde{\\mathbf{p}}_n\\) consists of a small subset of original point cloud \\(\\mathbf{p}\\). These superpoints are geometrically simple and usually cannot cover objects of different classes. Ideally, each superpoint as a whole should belong to the same semantic category. The simplicity and coherence of superpoints naturally facilitate the implementation of semantic consistency constraints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To mitigate the noise issue in 2D models, we can leverage the semantic consistency provided by 3D geometric priors. In particular, for a specific input point cloud \\(\\mathbf{P}\\), we have its neural features \\(\\mathbf{F}^{3\\mathrm{D}} = \\{\\mathbf{f}_1^{3\\mathrm{D}}, \\dots, \\mathbf{f}_M^{3\\mathrm{D}}\\}\\), the corresponding pixel fusion features \\(\\mathbf{F}^{2\\mathrm{D}} = \\{\\mathbf{f}_1^{2\\mathrm{D}}, \\dots, \\mathbf{f}_M^{2\\mathrm{D}}\\} \\in \\mathbb{R}^{M \\times C}\\) and superpoints" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.147, + 0.482, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.301, + 0.45, + 0.314 + ], + "angle": 0, + "content": "Input Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.147, + 0.705, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.302, + 0.655, + 0.314 + ], + "angle": 0, + "content": "Superpoints" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig. 3: Example of the superpoints generated using VCCS [44] on ScanNet scene point clouds. Each colored patch corresponds to a distinct superpoint, showcasing a simple geometric structure. Additionally, the points within each superpoint consistently exhibit a shared semantic category." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.391, + 0.784, + 0.424 + ], + "angle": 0, + "content": "\\(\\{\\tilde{\\mathbf{p}}_1 \\cdots \\tilde{\\mathbf{p}}_n \\cdots \\tilde{\\mathbf{p}}_N\\}\\). Firstly, we compute the mean 2D pixel features for superpoints, denoted as \\(\\mathbf{F}_{sp}^{2D} = \\{\\tilde{\\mathbf{f}}_1^{2D}, \\cdots, \\tilde{\\mathbf{f}}_N^{2D}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.435, + 0.785, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {f}} _ {n} ^ {2 \\mathrm {D}} = \\frac {1}{Q} \\sum_ {q = 1} ^ {Q} \\mathbf {f} _ {q} ^ {2 \\mathrm {D}}, \\quad \\mathbf {f} _ {q} ^ {2 \\mathrm {D}} \\in \\mathbb {R} ^ {1 \\times C}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.784, + 0.554 + ], + "angle": 0, + "content": "where \\(Q\\) is the total number of 3D points within a superpoint \\(\\tilde{\\mathbf{p}}_n\\), and \\(\\mathbf{f}_q^{2\\mathrm{D}}\\) is the feature vector for the \\(q^{th}\\) pixel that corresponds to the 3D point in the superpoint. Similarly, we can employ the same approach to calculate the mean 3D point features \\(\\mathbf{F}_{sp}^{3\\mathrm{D}} = \\{\\tilde{\\mathbf{f}}_1^{3\\mathrm{D}}, \\dots, \\tilde{\\mathbf{f}}_N^{3\\mathrm{D}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.784, + 0.589 + ], + "angle": 0, + "content": "To enforce the superpoint feature \\(\\mathbf{F}_{sp}^{3D}\\) to be consistent with the 2D fused features \\(\\mathbf{F}_{sp}^{2D}\\), we use a cosine similarity loss as that in pixel-point distillation:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.599, + 0.785, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s p} = 1 - \\cos \\left(\\mathbf {F} _ {s p} ^ {\\mathrm {2 D}}, \\mathbf {F} _ {s p} ^ {\\mathrm {3 D}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.629, + 0.772, + 0.644 + ], + "angle": 0, + "content": "The final optimization objective for geometry guided distillation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.656, + 0.785, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\mathcal {L} _ {p} + \\mathcal {L} _ {s p}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.544, + 0.71 + ], + "angle": 0, + "content": "3.4 Geometry Guided Self-Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "With geometry guided distillation, the performance of the 3D model can surpass that of the directly projected 2D model. This inspires us to further design a self-distillation module to fully leverage the inherent advantages of 3D data in representation. Recent methods such as CNS [5] and CLIP2Scene [6] typically transfer the pixel-level pseudo-labels from the 2D image to the 3D point cloud using a 2D-3D calibration matrix. In contrast to existing methods that rely on projecting from 2D, we introduce a geometry guided self-distillation approach to distill knowledge from the 3D model itself. Specifically, we generate predictions" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "directly on 3D point clouds instead of projecting 2D predictions back to 3D, as our 3D model possesses enhanced feature representation capabilities. Specifically, given the 3D point features \\(\\mathbf{F}^{3\\mathrm{D}} = \\{\\mathbf{f}_1^{3\\mathrm{D}},\\dots ,\\mathbf{f}_M^{3\\mathrm{D}}\\} \\in \\mathbb{R}^{M\\times C}\\), we assign pseudolabels to each point \\(\\mathbf{f}_n^{3\\mathrm{D}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.222, + 0.786, + 0.247 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {\\hat {t}} = \\underset {l} {\\arg \\max } \\psi \\left(\\mathbf {f} _ {n} ^ {\\mathrm {3 D}}, \\mathbf {f} _ {l} ^ {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.789, + 0.319 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_l^t\\) is CLIP's text embedding of the \\(l\\)-th class name with a pre-defined template. \\(\\psi\\) is the dot product operation. Furthermore, in the self-distillation stage, we also incorporate 3D geometric priors to enhance the semantic consistency of the predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.789, + 0.441 + ], + "angle": 0, + "content": "However, the research findings from self-supervised methods [7,23] have highlighted that directly using the network's predictions as supervision signals can lead to model collapse due to overconfidence. To address this issue, we leverage the predictions from the exponential moving average (EMA) model as supervision signals to train the 3D point cloud network, aiming to enhance stability and mitigate the impact of overconfidence. Specifically, we use the features extracted by the EMA model to calculate the similarity with the text feature according to Equation 5, resulting in generating a point-text pair for each point cloud." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.79, + 0.638 + ], + "angle": 0, + "content": "Although using EMA model can prevent mode collapse, it is still inevitable to have noise in the prediction results. To mitigate this issue, we incorporate the semantic consistency of 3D geometric priors during the self-distillation process. Specifically, we implement a voting mechanism within each superpoint to refine the prediction results. After assigning the most similar textual feature to each point cloud according to Equation 5, we calculate the proportion of textual features within each superpoint. We then modify the text corresponding to all point clouds within that superpoint to match the textual feature with the highest proportion. This correction process effectively mitigates the impact of noise and improves the accuracy of prediction. The voting outcome generates point-language pairs for each point cloud, then we can pull paired point and language features together while pushing unmatched features away through contrastive loss. The geometry guided self-distillation loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.642, + 0.787, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s d} = - \\log \\frac {\\exp \\left(\\mathbf {f} ^ {\\mathrm {3 D}} \\cdot \\mathbf {f} ^ {\\hat {t}} / \\tau\\right)}{\\sum_ {i = 1} ^ {n _ {t}} \\exp \\left(\\mathbf {f} ^ {\\mathrm {3 D}} \\cdot \\mathbf {f} _ {i} ^ {t} / \\tau\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.713 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}^t\\) represents the textual features corresponding to the point cloud \\(\\mathbf{f}^{3\\mathrm{D}}\\); \\(n_t\\) is the number of point-language pairs and \\(\\tau\\) is a temperature factor." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.737, + 0.452, + 0.753 + ], + "angle": 0, + "content": "3.5 Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training. The training process comprises two stages. In the initial stage, we employ geometry guided distillation to train the 3D network for 70 epochs, enabling it to acquire knowledge from the 2D model. Then, we incorporate the GGSD module into an additional 30 training epochs. The total number of training epochs is consistent with OpenScene." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Inference. Using any per-point feature \\(\\mathbf{f}^{3D}\\) and CLIP text features obtained from a set of text prompts, we can estimate their similarities by calculating the cosine similarity score between them. This similarity score is utilized in all of our scene understanding tasks. In all experiments, unless otherwise stated, we exclusively use pure 3D point clouds for predictions during the inference stage, without employing the 2D-3D ensemble strategy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.26, + 0.376, + 0.276 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.414, + 0.306 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Datasets. To evaluate the effectiveness of our method, we conduct experiments on three widely used public benchmarks: ScanNet [10], Matterport [3], and nuScenes [2]. The first two datasets provide RGBD images and 3D meshes of indoor scenes, while the last one offers LIDAR scans of outdoor scenes. Additionally, Matterport3D is a challenging dataset with intricately detailed scenes, which allows us to thoroughly examine the performance of open-vocabulary queries across different datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Implementation Details. We adopt the settings outlined in OpenScene [45], utilize MinkowskiNet18A [9] as the 3D backbone network, and employ a simple prompt engineering technique to extract CLIP text features. Specifically, we modify the text prompts for each object class \"XX\" to \"a XX in a scene\", such as \"a chair in a scene\". On the ScanNet and Matterport3D datasets, we employ a voxel size of \\(2\\mathrm{cm}\\) with MinkowskiNet while utilizing a voxel size of \\(5\\mathrm{cm}\\) on the nuScenes dataset. Based on the findings of OpenScene, we employ LSeg [34] for indoor datasets and OpenSeg [17] for outdoor dataset as the 2D teacher model. The temperature \\(\\tau\\) in Eq. 6 is empirically set as 0.01. We utilize the Adam optimizer with an initial learning rate of \\(1e-4\\) for the entire training process. We use a batch size of 8 on ScanNet and Matterport3D with a single NVIDIA A100 GPU (80G). On nuScenes, we use a batch size of 16 with 4 A100 GPUs. It takes around 24 hours to train and 0.1 second for inference. More details can be found in the supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.373, + 0.665 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.735 + ], + "angle": 0, + "content": "We conducted a series of experiments, including traditional closed-set scenarios, in-domain open vocabulary scenarios, and out-of-domain open vocabulary scenarios to test the effectiveness of the proposed methods in various 3D scene understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Comparison on Standard Benchmarks. We compare our approach with both fully-supervised and open vocabulary methods on all classes of the ScanNet [10] validation set and nuScenes [2] validation set. As shown in the bottom panel of Table 1, our approach, GGSD, consistently outperforms recent open vocabulary methods PLA [12], CLIP2Scene [6], CNS [5] and CLIP-FO3D [66] by a large margin on both datasets. Additionally, our method exhibits significant performance improvements compared to the baseline model OpenScene [45]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.144, + 0.675, + 0.415 + ], + "angle": 0, + "content": "
ScanNet [10]nuScenes [2]
mIoUmAccmIoUmAcc
Fully-supervised methods
TangentConv [55]40.9---
TextureNet [27]54.8---
ScanComplete [11]56.6---
DCM-Net [53]65.8---
Mix3D [43]73.6---
VMNet [26]73.2---
LidarMultiNet [63]--82.0-
MinkowskiNet [9]69.077.578.083.7
Open vocabulary methods
MSeg [33] Voting45.654.431.036.9
PLA [12]17.733.5--
CLIP2Scene [6]25.1-20.8-
CNS [5]26.8-33.5-
CLIP-FO3D [66]30.249.1--
OpenScene†[45]54.266.642.161.8
OpenScene‡[45]52.963.242.957.1
GGSD (Ours)56.568.646.159.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Table 1: Comparisons on Standard Benchmarks. The numbers in gray are obtained from previous papers. The symbol \\(\\dagger\\) indicates the use of the 2D-3D ensemble strategy, while \\(\\ddagger\\) denotes the pure 3D results obtained from the official open-source model. Our method surpasses the current state-of-the-art by a large margin in both indoor and outdoor scenarios without using the 2D-3D ensemble strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Specifically, on the ScanNet dataset, GGSD achieves a \\(3.6\\%\\) increase in mIoU and a \\(5.4\\%\\) increase in mACC. On the nuScenes dataset, GGSD demonstrates a \\(3.2\\%\\) increase in mIoU and a \\(2.1\\%\\) increase in mAcc. Except for OpenScene, the other methods in Table 1 do not employ the 2D-3D ensemble strategy due to the significant computational and storage costs associated with maintaining a 2D segmentation network and processing a large number of images per scene during testing. It is worth noting that our method using only point clouds still outperforms OpenScene with the ensemble strategy. Given that our method is evaluated on both indoor and outdoor scenes, and it even surpasses OpenScene with ensemble strategy, these results clearly demonstrate the effectiveness and superiority of GGSD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Comparison on Open-Vocabulary Scenarios. The standard ScanNet benchmark only contains a small vocabulary of 20 classes. To evaluate the open vocabulary capability of GGSD, we expand the original vocabulary size by using the NYU-40 label set. We remove NYU-40 labels that lack specific semantics, such as \"other structure\", \"other furniture\", and \"other prop\". The remaining categories are evenly divided into Head, Common, and Tail based on the number of samples in each category (see supplementary file for specific division). We present the quantitative results on these three categories in Table 2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.144, + 0.694, + 0.242 + ], + "angle": 0, + "content": "
HeadCommonTail
mIoUmAccmIoUmAccmIoUmAcc
Inference by feature projection
MaskCLIP-3D19.833.113.326.87.515.3
OpenScene‡[45]64.476.827.835.114.921.4
GGSD (Ours)66.878.830.140.916.024.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.252, + 0.788, + 0.322 + ], + "angle": 0, + "content": "Table 2: Comparison on open-vocabulary scenarios. Open-vocabulary semantic segmentation on ScanNet with extended labels from the NYU label set. Categories are divided into Head, Common and Tail by point numbers. The numbers in gray are obtained from previous papers. The symbol \\(\\ddagger\\) denotes the pure 3D point results obtained from the official open-source model." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.349, + 0.744, + 0.415 + ], + "angle": 0, + "content": "
Matterport21Matterport40Matterport80Matterport160
mIoUmAccmIoUmAccmIoUmAccmIoUmAcc
OpenScene‡[45]36.048.021.127.510.813.96.08.1
GGSD (Ours)40.154.422.831.611.916.26.39.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Table 3: Comparison on cross-domain scenarios. Both GGSD and OpenScene are trained on ScanNet, and zero-shot tested on the Matterport3D dataset. The symbol \\(\\ddagger\\) denotes the pure 3D results obtained from the official open-source model. \\(K = 21\\) is derived from the original Matterport3D benchmark. For \\(K = 40,80,160\\), we use the \\(K\\) most common categories from the NYU label set provided in the benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.674 + ], + "angle": 0, + "content": "GGSD achieves superior results over OpenScene in all categories, with a specific improvement of \\(2.4\\%\\) and \\(2.3\\%\\) in mIOU for the Head and Common categories, respectively. Since the categories in the Common set typically consist of objects with smaller sizes (e.g., night stand and floor mat), the \\(2.3\\%\\) mIOU improvement demonstrates the stronger open vocabulary capability of our method. However, for the Tail category, although GGSD surpasses OpenScene by \\(2.8\\%\\) in mACC, the performance of both OpenScene and GGSD is not satisfactory. This can be attributed to the small sizes and the limited number of samples in these categories (e.g., lamp, box, bag, and book)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison on Cross-Domain Scenarios. Our GGSD method has demonstrated excellent potential in solving open vocabulary scene understanding tasks within a specific domain. However, the exploration of open vocabulary learners that can transfer across datasets is also worth investigation, as they face the dual challenges of category and data distribution. We conduct zero-shot domain transfer experiments by training the model on ScanNet [10] and testing it on the Matterport3D [3] dataset without fine-tuning. It should be noted that as \\( K \\) increases, the test set will contain more small-sized objects. As shown in Table 3, our GGSD method consistently outperforms OpenScene in Matterport21, Matterport40, Matterport80, and Matterport160, demonstrating its strong open vocabulary capability and generalization ability." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "table", + "bbox": [ + 0.292, + 0.144, + 0.716, + 0.244 + ], + "angle": 0, + "content": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
2D Fusion Projection50.062.732.340.0
Pixel-Point Distillation52.963.236.148.0
Geometry Guided Distillation53.565.036.749.3
Self-Distillation56.168.239.053.3
Geometry Guided Self-distillation56.568.640.154.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.254, + 0.785, + 0.281 + ], + "angle": 0, + "content": "Table 4: Ablation study of GGSD. Comparison of semantic segmentation performance of each component of our GGSD method." + }, + { + "type": "table", + "bbox": [ + 0.292, + 0.308, + 0.716, + 0.401 + ], + "angle": 0, + "content": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Geometry Guided Distillation54.065.036.749.3
Fixed 2D model51.964.634.245.7
Fixed 3D model55.065.837.349.8
EMA model56.168.239.053.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.411, + 0.775, + 0.426 + ], + "angle": 0, + "content": "Table 5: Comparisons of the 3D supervision signals provided by different models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.513, + 0.484 + ], + "angle": 0, + "content": "4.3 Ablation Studies and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.538 + ], + "angle": 0, + "content": "In this section, we investigate our key components through an in-depth ablation study. The experiments are conducted on the ScanNet dataset and subsequently tested on both ScanNet and Matterport3D datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Component Analysis. As shown in Table 4, even without utilizing 3D geometric priors, pixel-point distillation can outperform 2D direct projection, demonstrating the advantage of 3D representation. By incorporating 3D geometric priors, the geometry guided distillation method further enhances performance, resulting in a \\(0.6\\%\\) increase in mIoU and a \\(1.8\\%\\) increase in mAcc on the ScanNet dataset. Through self-distillation, we achieve a notable enhancement in performance. Specifically, on the ScanNet dataset, there is a substantial increase of \\(2.6\\%\\) in mIoU and \\(3.2\\%\\) in mAcc. On the Matterport3D dataset, the mIoU is increased by \\(2.3\\%\\), and the mAcc is improved by \\(4.0\\%\\). Finally, the utilization of the voting mechanism within the superpoints leads to improved performance by mitigating the impact of noise through the refinement model's predictions. It is noteworthy that our GGSD method demonstrates substantial improvements compared to the 2D teacher model, i.e., 2D Fusion Projection, achieving a \\(+6.5\\%\\) mIoU and \\(+5.9\\%\\) mAcc on the ScanNet dataset, as well as a \\(+7.8\\%\\) mIoU and \\(+14.4\\%\\) mAcc on the Matterport3D dataset. These results provide strong evidence for the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The Effectiveness of the EMA Model. In the second stage, we employ the EMA model to provide supervised signals for self-distillation. In Table 5, we investigate the performance of different models in providing supervised signals. When using the fixed 2D model for pixel prediction and projecting it onto" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.279, + 0.144, + 0.731, + 0.228 + ], + "angle": 0, + "content": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Pixel-Point Distillation52.963.236.148.0
SAM [31] Refine53.063.335.948.4
Geometry Guided Distillation53.565.036.749.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.239, + 0.772, + 0.254 + ], + "angle": 0, + "content": "Table 6: Comparisons between the utilization of geometric priors and SAM [31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.784, + 0.388 + ], + "angle": 0, + "content": "points, we observe lower performance compared to geometry-guided distillation. This can be attributed to the fact that the 2D model provides more erroneous supervised signals. Furthermore, using the fixed first-stage distilled 3D model only results in marginal performance improvement. However, when utilizing the EMA model, we achieve the best performance, primarily because the EMA model gradually obtains more accurate prediction results as training progresses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.522 + ], + "angle": 0, + "content": "The Effectiveness of 3D Geometric Priors. The recent work CNS [5] utilized the foundation model SAM [31] to refine pseudo-labels from the 2D model, achieving performance improvements on a lower baseline (27.3% mIoU in ScanNet). We also apply this approach to LSeg. Specifically, we follow the official implementation of SAM to generate masks for all images and perform average pooling within each mask to refine them. Finally, we utilize the extracted refined features for distillation. However, as shown in Table 6, this approach yields minimal improvements. Conversely, leveraging 3D geometric priors results in a 0.6% mIoU improvement in ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.787, + 0.751 + ], + "angle": 0, + "content": "Qualitative Results. As shown in Fig. 4, our method successfully segments all categories that exist in closed-set ScanNet benchmarks, such as wall and table. Additionally, our method successfully segments some categories that are not annotated in traditional benchmarks, such as box, lamp, and clothes, demonstrating our model's excellent open vocabulary capability. On the Matterport3D, the GGSD method successfully segments unlabeled television and also identifies mislabeled pillow and coffee table in the ground truth. Furthermore, there have been instances of segmentation failures in Matterport3D. Due to linguistic ambiguity, sofa chairs are sometimes segmented as separate entities, such as sofas and chairs, while coffee tables can be segmented as coffee tables and tables. The misclassification of certain pillow points as sofas may be attributed to their strong color resemblance to sofas. In the case of nuScenes, GGSD also successfully segments unannotated barrier. These qualitative results obtained from GGSD demonstrate its excellent open vocabulary capability. More visualization results can be found in the supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.36, + 0.792 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this study, we proposed an effective approach called Geometry Guided Self-Distillation (GGSD) to learn superior 3D representations from 2D pre-trained" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.789, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.475, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results. We present the qualitative results of 3D semantic segmentation on public indoor and outdoor benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.689 + ], + "angle": 0, + "content": "models. We first introduced a geometry guided distillation module that leveraged 3D geometric priors to alleviate noise in 2D models. Moreover, we harnessed the representation advantages of 3D data through geometry guided self-distillation, resulting in significant improvement in the model performance for 3D scene understanding. Our GGSD method significantly outperformed existing open-vocabulary 3D scene understanding methods on both indoor and outdoor datasets. This highlighted the crucial role of geometric priors and of 3D data representation in distillation-based approaches for 3D scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Limitations. Though GGSD has achieved promising results in 3D open vocabulary scene understanding tasks, it has two main limitations. Firstly, while geometric priors can significantly improve the model performance on objects with pronounced geometric structures, the improvement can be limited on objects with fewer prominent geometric characteristics. Secondly, similar to other open vocabulary methods, our approach is susceptible to the influence of language ambiguity, as illustrated by the qualitative results in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.742, + 0.842 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the InnoHK program." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.784, + 0.208 + ], + "angle": 0, + "content": "1. Vit-gpt2 image captioning. https://huggingface.co/nlpconnect/vit-gpt2-image-captioning/discussions" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.264 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.307 + ], + "angle": 0, + "content": "3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.362 + ], + "angle": 0, + "content": "4. Chen, A., Zhang, K., Zhang, R., Wang, Z., Lu, Y., Guo, Y., Zhang, S.: Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5291-5301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.785, + 0.405 + ], + "angle": 0, + "content": "5. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.46 + ], + "angle": 0, + "content": "6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.503 + ], + "angle": 0, + "content": "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.558 + ], + "angle": 0, + "content": "8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.601 + ], + "angle": 0, + "content": "9. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.643 + ], + "angle": 0, + "content": "10. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.699 + ], + "angle": 0, + "content": "1. Dai, A., Ritchie, D., Bokeloh, M., Reed, S., Sturm, J., Nießner, M.: Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7010-7019 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "3. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.C.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "4. Fan, J., Wang, Y., Guan, H., Song, C., Zhang, Z.: Toward few-shot domain adaptation with perturbation-invariant representation and transferable prototypes. Frontiers of Computer Science 16(3), 163347 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.215 + ], + "angle": 0, + "content": "17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: European Conference on Computer Vision. pp. 540-557. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.215, + 0.785, + 0.257 + ], + "angle": 0, + "content": "18. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.31 + ], + "angle": 0, + "content": "20. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.31, + 0.785, + 0.351 + ], + "angle": 0, + "content": "21. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2940-2949 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.351, + 0.785, + 0.392 + ], + "angle": 0, + "content": "22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.433 + ], + "angle": 0, + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "24. Hu, W., Zhao, H., Jiang, L., Jia, J., Wong, T.T.: Bidirectional projection network for cross dimension scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14373-14382 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.5 + ], + "angle": 0, + "content": "25. Hu, X., Wang, Y., Fan, L., Fan, J., Peng, J., Lei, Z., Li, Q., Zhang, Z.: Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.5, + 0.785, + 0.554 + ], + "angle": 0, + "content": "26. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15488-15498 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.785, + 0.609 + ], + "angle": 0, + "content": "27. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4440-4449 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "28. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. arXiv preprint arXiv:2210.01055 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.676 + ], + "angle": 0, + "content": "29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.676, + 0.785, + 0.717 + ], + "angle": 0, + "content": "30. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: Mseg: A composite dataset for multi-domain semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2879-2888 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "35. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.219, + 0.787, + 0.275 + ], + "angle": 0, + "content": "36. Li, J., He, X., Wen, Y., Gao, Y., Cheng, X., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11809-11818 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "37. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv abs/2210.04150 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.318, + 0.787, + 0.374 + ], + "angle": 0, + "content": "38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "39. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.417, + 0.787, + 0.444 + ], + "angle": 0, + "content": "40. Lüddecke, T., Ecker, A.S.: Image segmentation using text and image prompts. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.488, + 0.787, + 0.515 + ], + "angle": 0, + "content": "42. Misra, I., Girdhar, R., Joulin, A.: An End-to-End Transformer Model for 3D Object Detection. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.516, + 0.787, + 0.558 + ], + "angle": 0, + "content": "43. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 International Conference on 3D Vision (3DV). pp. 116-125. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.559, + 0.787, + 0.601 + ], + "angle": 0, + "content": "44. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.601, + 0.787, + 0.657 + ], + "angle": 0, + "content": "45. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.7, + 0.787, + 0.728 + ], + "angle": 0, + "content": "47. Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: Langsplat: 3d language gaussian splatting. arXiv preprint arXiv:2312.16084 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.787, + 0.785 + ], + "angle": 0, + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.787, + 0.841 + ], + "angle": 0, + "content": "49. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Dense-clip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "P. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "50. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: 36th Conference on Neural Information Processing Systems (NIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.785, + 0.232 + ], + "angle": 0, + "content": "51. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.274 + ], + "angle": 0, + "content": "52. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.33 + ], + "angle": 0, + "content": "53. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: Dualconvmesh-net: Joint geodesic and euclidean convolutions on 3d meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8612-8622 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.373 + ], + "angle": 0, + "content": "54. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.785, + 0.415 + ], + "angle": 0, + "content": "55. Tatarchenko, M., Park, J., Koltun, V., Zhou, Q.Y.: Tangent convolutions for dense prediction in 3d. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3887-3896 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.444 + ], + "angle": 0, + "content": "56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.487 + ], + "angle": 0, + "content": "57. Wang, H., Shen, Y., Fei, J., Li, W., Wu, L., Wang, Y., Zhang, Z.: Pulling target to source: A new perspective on domain adaptive semantic segmentation. arXiv preprint arXiv:2305.13752 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "58. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR abs/2202.03052 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.572 + ], + "angle": 0, + "content": "59. Wang, Y., Liang, J., Xiao, J., Mei, S., Yang, Y., Zhang, Z.: Informative data mining for one-shot cross-domain semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1064-1074 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.613 + ], + "angle": 0, + "content": "60. Wang, Y., Peng, J., Zhang, Z.: Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9092-9101 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.655 + ], + "angle": 0, + "content": "61. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.698 + ], + "angle": 0, + "content": "62. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.74 + ], + "angle": 0, + "content": "63. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidarmulti-net: Towards a unified multi-task network for lidar perception. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3231-3240 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.769 + ], + "angle": 0, + "content": "64. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "65. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "66. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "GGSD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "67. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "68. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N.C.F., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.259 + ], + "angle": 0, + "content": "69. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.259 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_origin.pdf b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9ed140f4dde3e698566868ac629162913b366a49 --- /dev/null +++ b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/c70a8b2d-3925-490e-bfc0-2945ff5cecb6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e645e7b66e3e8c67af12965ec0084acde10219a4247c7261821bdfb1707a1b3 +size 3393761 diff --git a/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/full.md b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3c5883b25f212c0f6abb939436b3611e02233293 --- /dev/null +++ b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/full.md @@ -0,0 +1,281 @@ +# Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation + +Pengfei Wang $^{1,2}$ , Yuxi Wang $^{2}$ , Shuai Li $^{1}$ , Zhaoxiang Zhang $^{1,2,3,4(\boxtimes)}$ , Zhen Lei $^{1,2,3,4}$ , and Lei Zhang $^{1(\boxtimes)}$ + +1 The Hong Kong Polytechnic University + +2 Centre for Artificial Intelligence and Robotics, HKISI, CAS + +3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA + +$^{4}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) + +pengfei.wang@connect.polyu.hk, zhaoxiang.zhang@ia.ac.cn, + +cslzhang@comp.polyu.edu.hk + +Abstract. The scarcity of large-scale 3D-text paired data poses a great challenge on open vocabulary 3D scene understanding, and hence it is popular to leverage internet-scale 2D data and transfer their open vocabulary capabilities to 3D models through knowledge distillation. However, the existing distillation-based 3D scene understanding approaches rely on the representation capacity of 2D models, disregarding the exploration of geometric priors and inherent representational advantages offered by 3D data. In this paper, we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained models. Specifically, we first design a geometry guided distillation module to distill knowledge from 2D models, and then leverage the 3D geometric priors to alleviate the inherent noise in 2D models and enhance the representation learning process. Due to the advantages of 3D representation, the performance of the distilled 3D student model can significantly surpass that of the 2D teacher model. This motivates us to further leverage the representation advantages of 3D data through self-distillation. As a result, our proposed GGSD approach outperforms the existing open vocabulary 3D scene understanding methods by a large margin, as demonstrated by our experiments on both indoor and outdoor benchmark datasets. Codes are available at https://github.com/Wang-pengfei/GGSD. + +Keywords: 3D scene understanding $\cdot$ Open vocabulary $\cdot$ Self-distillation + +# 1 Introduction + +3D scene understanding plays a vital role in various real-world applications, including robot manipulation, virtual reality, and autonomous driving, etc. Traditional 3D scene understanding methods mostly employ supervised training on labeled 3D datasets with a limited number of classes [18,42,56]. These approaches, however, face difficulties when they encounter unknown categories in open environments. Such a limitation in practical applications motivates researchers to + +![](images/a94125fd6d8516321dc0fe7eff58ddf59c6524eb05cd7782aaa5dbadc76ec7e7.jpg) +Fig. 1: Example of the problems when utilizing 2D models, e.g., LSeg [34], for pixel-to-point distillation. We can see the noise caused by the factors such as occlusion, lighting variations and view angles in 2D images. Therefore, only employing the 2D model in distillation brings errors to the 3D model, limiting its 3D representation capacity. + +explore new 3D scene understanding models with open vocabulary capability. Inspired by the success of vision-language foundation models like CLIP [48], which aligns language features to image features, in various image understanding tasks [19,28,34,50,61,67,69], an intuitive strategy is to leverage a substantial amount of paired 3D point cloud and text data for training. Unfortunately, the high cost of annotating 3D datasets (e.g., annotating a scene with 20 categories takes approximately 22.3 minutes [12]) makes it impractical to manually label enough data with numerous real-world categories. + +An alternative yet more practical approach is to leverage internet-scale 2D data to acquire open vocabulary capability, and then transfer this capability to 3D scene understanding models via distillation [6, 12, 45, 62, 66]. For instance, PLA [12] utilizes a 2D caption model [58] to generate textual descriptions for images, which are then refined using GPT2 [1]. Subsequently, these descriptions are projected onto the 3D point cloud, and the vision-language knowledge can be passed to the 3D model via 3D-text contrastive learning. The recent state-of-the-art OpenScene [45] utilizes 2D open-vocabulary segmentors such as LSeg [34] and OpenSeg [17] to extract pixel-level 2D embeddings for the 3D point cloud. The ability of pixel-text alignment is then transferred to align 3D point to text via feature distillation, enabling 3D semantic-level scene understanding. + +Though the above mentioned distillation-based methods have shown promising results in 3D scene understanding, they basically imitate the 2D models and inherit their limitations, restricting the upper bound of representation power. As shown in Fig. 1, the 2D model may suffer from errors due to issues such as occlusion, lighting variations, and viewing angles. The misclassification of ex + +tracted 2D features leads to erroneous supervisory signals during the distillation process of 3D point cloud learning. Instead of solely imitating 2D models, our objective is to harness both the geometric priors in 3D point clouds and the representational advantages of 3D data to guide the knowledge distillation process from pre-trained 2D models. + +To achieve the above mentioned objectives, in this paper we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained open vocabulary segmentors. Given a sparse point cloud composed of multiple semantic categories, geometrically homogeneous neighborhoods typically correspond to the same semantic information. Based on this observation, we first design a geometry guided distillation module to learn knowledge from 2D pre-trained models. Specifically, we generate geometrically homogeneous neighborhoods as superpoints and utilize their semantic consistency to guide the distillation process. The utilization of 3D geometric priors could mitigate the inherent noise of the 2D model and enhance the representation learning capability. Furthermore, due to the representation advantages of 3D data, which are less affected by factors such as lighting and viewing angles, the distilled 3D student model can exhibit superior performance compared to the 2D teacher model. This motivates us to further distill knowledge within the 3D network itself through self-distillation. We employ an exponential moving average (EMA) model to predict each 3D point cloud and perform voting within the superpoint to provide supervision signals for the online 3D model. This approach ensures that the online model could receive reliable and accurate guidance by leveraging the consensus among multiple-point predictions. Finally, our proposed GGSD method could learn robust 3D representations from the 3D distillation model itself, achieving significantly superior performance to the 2D teacher model. + +The contributions of this work are summarized as follows. First, we analyze the limitations of pixel-point distillation methods and propose leveraging the advantages of 3D data representation through self-distillation, instead of relying solely on 2D representations. Secondly, we utilize 3D geometric priors to mitigate the noise in pixel-point distillation process, thereby enhancing the capability of representation learning. Third, we conduct experiments on both indoor and outdoor datasets, including ScanNet [10], Matterport [3], and nuScenes [2]. The results demonstrate that GGSD sets a new state-of-the-art for 3D open vocabulary scene understanding in both indoor and outdoor datasets. + +# 2 Related Work + +Closed-set 3D Scene Understanding. Previous studies in closed-set 3D scene understanding [9,21,24,26,27,36,43,46,52] have led to significant advancements across diverse 3D scene understanding benchmarks, encompassing tasks such as 3D object classification, 3D object detection, 3D semantic and instance segmentation, etc. Some researches [4] have explored the use of 2D pre-trained models, such as MAE [22], for pre-training point encoders. However, they mainly em + +phasize on leveraging pre-training to enhance supervised 3D downstream tasks, rather than addressing open-vocabulary query scenarios. Existing closed-set 3D scene understanding methods excel in specific categories but often struggle to recognize new categories in open-set scenarios. Our objective is to develop a solution to effectively recognize and comprehend new categories in open-set environments, thus overcoming this limitation. + +Open-Vocabulary 2D Scene Understanding. Traditional closed-set scene understanding has seen significant performance improvements [8,14,57,59,60,65], while open-vocabulary scene understanding is also beginning to flourish. Prior studies have explored knowledge distillation from visual-language models, such as CLIP, to transfer their open vocabulary capabilities in 2D vision tasks [13,15, 16,20,35,37,40,51,64,68]. Recent research endeavors have aimed to establish correlations between dense image features and embeddings derived from language models [17, 19, 32, 34, 38, 49]. Due to their pixel-level text alignment capabilities, 2D open vocabulary scene understanding models, such as OpenSeg [17] and LSeg [34], can serve as teacher models for distillation of 3D models compared to CLIP. In this study, we rely on pre-trained 2D models with open vocabulary capabilities, specifically OpenSeg and LSeg, and propose a novel geometry-guided self-distillation strategy to conduct 3D scene-level understanding tasks without the need for any ground-truth 2D or 3D training data. + +Open-Vocabulary 3D Scene Understanding. Recent works on open-vocabulary 3D scene understanding can be classified into two categories. The first category [25,30,39,47], represented by methods like LERF [30], focuses on optimization-based approaches that utilize techniques such as NeRF [41] or 3DGS [29] to distill 2D features into 3D scenes. However, these methods lack generalization capability when confronted with new scenes. The second category [6,12,45,54], represented by methods like OpenScene [45], is based on learning-based approaches. This paper falls in the second category. CLIP2Scene [6] leverages CLIP image model to extract 2D image features and utilizes textual category features to classify pixels, which are then projected onto 3D point clouds. OpenScene [45] aggregates pixel features from multiple view images on every 3D point, achieving state-of-the-art open vocabulary 3D scene understanding performance. Despite the promising results of these methods, their simple imitation of 2D models limits the representational capacity of 3D models. + +# 3 Method + +An overview of our approach is illustrated in Fig. 2. Firstly, we follow OpenScene [45] to create pixel-point feature pairs for the 3D point cloud (see Section 3.1). Subsequently, we analyze the limitations of pixel-point distillation methods (see Section 3.2). Then, we propose a geometry guided distillation module to enhance the learning of 3D representations from the 2D model (see Section 3.3). Finally, we introduce a geometry guided self-distillation module to further distill the knowledge within the 3D network itself (see Section 3.4). + +![](images/d58c3b1a390d9f2b02b655de569ad1d524f644ccdb63d53b087e25c10197cbd7.jpg) +Fig. 2: Overview of our Geometry Guided Self-Distillation (GGSD) framework, which consists of two main components: Geometry Guided Distillation and Self-Distillation. The first module leverages 3D geometric priors to mitigate the inherent noise in 2D models. In the second module, the 3D network learns from its own predictions. This self-distillation step allows the 3D network to continuously enhance its understanding and representation capabilities. Overall, the GGSD framework combines geometry guidance and self-distillation to fully exploit the valuable information hidden in 3D data, rather than solely relying on 2D pre-trained models. + +# 3.1 Creating Pixel-Point Feature Pairs + +Following OpenScene [45], our approach begins by extracting dense per-pixel embeddings from RGB images using a pre-trained 2D vision language segmentation model, such as LSeg [34] or OpenSeg [17]. After obtaining the pixel features, we can establish correspondences between each point in the point cloud and the pixels in different RGB images by utilizing the intrinsic matrix and the world-to-camera extrinsic matrix. Next, we employ a multi-view fusion strategy to combine the features from multiple views into a single feature vector, denoted as $\mathbf{f}^{2\mathrm{D}} = \phi (\mathbf{f}_1,\dots ,\mathbf{f}_K)$ , for each point cloud $\mathbf{p}$ . This fusion is achieved using an average pooling operator $\phi$ over the $K$ -view features. After obtaining the fused 2D features for each of the M points in the scene, we obtain a 2D pixel feature set $\mathbf{F}^{2\mathrm{D}} = \mathbf{f}_1^{2\mathrm{D}},\dots ,\mathbf{f}_M^{2\mathrm{D}}\in \mathbb{R}^{M\times C}$ corresponding to the point cloud scene. Finally, these embedded features are back-projected into the 3D point cloud scene. For more details, please refer to OpenScene [45]. + +# 3.2 Limitations of Pixel-Point 3D Distillation + +Due to the text-image alignment capability of CLIP, the 2D features $\mathbf{F}^{2\mathrm{D}}$ can be directly utilized for language-driven open vocabulary scene understanding. + +To transfer 2D vision-language knowledge to a 3D point cloud network, the distillation approach has been popularly used. One up-to-date pixel-point 3D distillation work is OpenScene [45]. Specifically, OpenScene aims to learn a 3D encoder that extracts 3D features from each point cloud, which are expected to be similar to their corresponding 2D pixel features. This allows the 3D encoder to distill the open vocabulary capability of the 2D model. To this end, OpenScene utilizes cosine distance to perform pixel-point distillation: + +$$ +\mathcal {L} _ {p} = 1 - \cos (\mathbf {F} ^ {\mathrm {2 D}}, \mathbf {F} ^ {\mathrm {3 D}}). \tag {1} +$$ + +However, the current pixel-to-point distillation methods are mostly a simplistic imitation of the 2D models, which are limited in extracting useful information from 3D point data. Firstly, relying solely on the projection relationship between point clouds and pixels, these approaches fail to fully exploit the geometric priors inherent in 3D data, leading to suboptimal performance. In this paper, we propose to incorporate 3D geometric priors throughout the entire distillation process to address the issue of noise in 2D models by constraining semantic consistency. Secondly, the current pixel-point distillation methods fail to fully exploit the representational advantage offered by 3D point clouds. It has been observed that the distilled 3D model can outperform the directly projected 2D models. This can be attributed to the inherent capability of 3D data to handle challenges like lighting variations, occlusion, and viewpoint changes, which are prevalent in 2D representations. Therefore, the accuracy of 3D model segmentation results, particularly for foreground classes, can be significantly improved. This observation provides a strong motivation for us to further leverage the advantages of 3D representation within the 3D network itself via self-distillation, rather than relying solely on information from the 2D model. + +# 3.3 Geometry Guided Distillation + +We first design a geometry guided distillation method to learn knowledge from 2D models. Given a sparse point cloud composed of multiple semantic categories, we can decompose it into many simple parts, i.e., superpoints, according to their geometric position in the space. As shown in Fig. 3, the superpoints for an indoor room are constructed by VCCS [44]. The details for creating superpoints can be found in the supplementary file. For a specific input point cloud $\mathbf{P} \in \mathbb{R}^{M \times 3}$ , its superpoints are denoted as $\{\tilde{\mathbf{p}}_1 \cdots \tilde{\mathbf{p}}_n \cdots \tilde{\mathbf{p}}_N\}$ , where each superpoint $\tilde{\mathbf{p}}_n$ consists of a small subset of original point cloud $\mathbf{p}$ . These superpoints are geometrically simple and usually cannot cover objects of different classes. Ideally, each superpoint as a whole should belong to the same semantic category. The simplicity and coherence of superpoints naturally facilitate the implementation of semantic consistency constraints. + +To mitigate the noise issue in 2D models, we can leverage the semantic consistency provided by 3D geometric priors. In particular, for a specific input point cloud $\mathbf{P}$ , we have its neural features $\mathbf{F}^{3\mathrm{D}} = \{\mathbf{f}_1^{3\mathrm{D}}, \dots, \mathbf{f}_M^{3\mathrm{D}}\}$ , the corresponding pixel fusion features $\mathbf{F}^{2\mathrm{D}} = \{\mathbf{f}_1^{2\mathrm{D}}, \dots, \mathbf{f}_M^{2\mathrm{D}}\} \in \mathbb{R}^{M \times C}$ and superpoints + +![](images/3341e6b5f098d0e334bb0f17ea4fd7d4fe1fb622bda8ad9b1889ea59436c3811.jpg) +Input Point Cloud + +![](images/6a132b87accc65d2d1bcca8c038c6fe1456a91835910cc43bfaef98d3296ae4d.jpg) +Superpoints +Fig. 3: Example of the superpoints generated using VCCS [44] on ScanNet scene point clouds. Each colored patch corresponds to a distinct superpoint, showcasing a simple geometric structure. Additionally, the points within each superpoint consistently exhibit a shared semantic category. + +$\{\tilde{\mathbf{p}}_1 \cdots \tilde{\mathbf{p}}_n \cdots \tilde{\mathbf{p}}_N\}$ . Firstly, we compute the mean 2D pixel features for superpoints, denoted as $\mathbf{F}_{sp}^{2D} = \{\tilde{\mathbf{f}}_1^{2D}, \cdots, \tilde{\mathbf{f}}_N^{2D}\}$ : + +$$ +\tilde {\mathbf {f}} _ {n} ^ {2 \mathrm {D}} = \frac {1}{Q} \sum_ {q = 1} ^ {Q} \mathbf {f} _ {q} ^ {2 \mathrm {D}}, \quad \mathbf {f} _ {q} ^ {2 \mathrm {D}} \in \mathbb {R} ^ {1 \times C}, \tag {2} +$$ + +where $Q$ is the total number of 3D points within a superpoint $\tilde{\mathbf{p}}_n$ , and $\mathbf{f}_q^{2\mathrm{D}}$ is the feature vector for the $q^{th}$ pixel that corresponds to the 3D point in the superpoint. Similarly, we can employ the same approach to calculate the mean 3D point features $\mathbf{F}_{sp}^{3\mathrm{D}} = \{\tilde{\mathbf{f}}_1^{3\mathrm{D}}, \dots, \tilde{\mathbf{f}}_N^{3\mathrm{D}}\}$ . + +To enforce the superpoint feature $\mathbf{F}_{sp}^{3D}$ to be consistent with the 2D fused features $\mathbf{F}_{sp}^{2D}$ , we use a cosine similarity loss as that in pixel-point distillation: + +$$ +\mathcal {L} _ {s p} = 1 - \cos \left(\mathbf {F} _ {s p} ^ {\mathrm {2 D}}, \mathbf {F} _ {s p} ^ {\mathrm {3 D}}\right). \tag {3} +$$ + +The final optimization objective for geometry guided distillation is as follows: + +$$ +\mathcal {L} _ {d} = \mathcal {L} _ {p} + \mathcal {L} _ {s p}. \tag {4} +$$ + +# 3.4 Geometry Guided Self-Distillation + +With geometry guided distillation, the performance of the 3D model can surpass that of the directly projected 2D model. This inspires us to further design a self-distillation module to fully leverage the inherent advantages of 3D data in representation. Recent methods such as CNS [5] and CLIP2Scene [6] typically transfer the pixel-level pseudo-labels from the 2D image to the 3D point cloud using a 2D-3D calibration matrix. In contrast to existing methods that rely on projecting from 2D, we introduce a geometry guided self-distillation approach to distill knowledge from the 3D model itself. Specifically, we generate predictions + +directly on 3D point clouds instead of projecting 2D predictions back to 3D, as our 3D model possesses enhanced feature representation capabilities. Specifically, given the 3D point features $\mathbf{F}^{3\mathrm{D}} = \{\mathbf{f}_1^{3\mathrm{D}},\dots ,\mathbf{f}_M^{3\mathrm{D}}\} \in \mathbb{R}^{M\times C}$ , we assign pseudolabels to each point $\mathbf{f}_n^{3\mathrm{D}}$ as follows: + +$$ +\mathbf {f} ^ {\hat {t}} = \underset {l} {\arg \max } \psi \left(\mathbf {f} _ {n} ^ {\mathrm {3 D}}, \mathbf {f} _ {l} ^ {t}\right), \tag {5} +$$ + +where $\mathbf{f}_l^t$ is CLIP's text embedding of the $l$ -th class name with a pre-defined template. $\psi$ is the dot product operation. Furthermore, in the self-distillation stage, we also incorporate 3D geometric priors to enhance the semantic consistency of the predictions. + +However, the research findings from self-supervised methods [7,23] have highlighted that directly using the network's predictions as supervision signals can lead to model collapse due to overconfidence. To address this issue, we leverage the predictions from the exponential moving average (EMA) model as supervision signals to train the 3D point cloud network, aiming to enhance stability and mitigate the impact of overconfidence. Specifically, we use the features extracted by the EMA model to calculate the similarity with the text feature according to Equation 5, resulting in generating a point-text pair for each point cloud. + +Although using EMA model can prevent mode collapse, it is still inevitable to have noise in the prediction results. To mitigate this issue, we incorporate the semantic consistency of 3D geometric priors during the self-distillation process. Specifically, we implement a voting mechanism within each superpoint to refine the prediction results. After assigning the most similar textual feature to each point cloud according to Equation 5, we calculate the proportion of textual features within each superpoint. We then modify the text corresponding to all point clouds within that superpoint to match the textual feature with the highest proportion. This correction process effectively mitigates the impact of noise and improves the accuracy of prediction. The voting outcome generates point-language pairs for each point cloud, then we can pull paired point and language features together while pushing unmatched features away through contrastive loss. The geometry guided self-distillation loss is as follows: + +$$ +\mathcal {L} _ {s d} = - \log \frac {\exp \left(\mathbf {f} ^ {\mathrm {3 D}} \cdot \mathbf {f} ^ {\hat {t}} / \tau\right)}{\sum_ {i = 1} ^ {n _ {t}} \exp \left(\mathbf {f} ^ {\mathrm {3 D}} \cdot \mathbf {f} _ {i} ^ {t} / \tau\right)}, \tag {6} +$$ + +where $\mathbf{f}^t$ represents the textual features corresponding to the point cloud $\mathbf{f}^{3\mathrm{D}}$ ; $n_t$ is the number of point-language pairs and $\tau$ is a temperature factor. + +# 3.5 Training and Inference + +Training. The training process comprises two stages. In the initial stage, we employ geometry guided distillation to train the 3D network for 70 epochs, enabling it to acquire knowledge from the 2D model. Then, we incorporate the GGSD module into an additional 30 training epochs. The total number of training epochs is consistent with OpenScene. + +Inference. Using any per-point feature $\mathbf{f}^{3D}$ and CLIP text features obtained from a set of text prompts, we can estimate their similarities by calculating the cosine similarity score between them. This similarity score is utilized in all of our scene understanding tasks. In all experiments, unless otherwise stated, we exclusively use pure 3D point clouds for predictions during the inference stage, without employing the 2D-3D ensemble strategy. + +# 4 Experiments + +# 4.1 Experiment Setup + +Datasets. To evaluate the effectiveness of our method, we conduct experiments on three widely used public benchmarks: ScanNet [10], Matterport [3], and nuScenes [2]. The first two datasets provide RGBD images and 3D meshes of indoor scenes, while the last one offers LIDAR scans of outdoor scenes. Additionally, Matterport3D is a challenging dataset with intricately detailed scenes, which allows us to thoroughly examine the performance of open-vocabulary queries across different datasets. + +Implementation Details. We adopt the settings outlined in OpenScene [45], utilize MinkowskiNet18A [9] as the 3D backbone network, and employ a simple prompt engineering technique to extract CLIP text features. Specifically, we modify the text prompts for each object class "XX" to "a XX in a scene", such as "a chair in a scene". On the ScanNet and Matterport3D datasets, we employ a voxel size of $2\mathrm{cm}$ with MinkowskiNet while utilizing a voxel size of $5\mathrm{cm}$ on the nuScenes dataset. Based on the findings of OpenScene, we employ LSeg [34] for indoor datasets and OpenSeg [17] for outdoor dataset as the 2D teacher model. The temperature $\tau$ in Eq. 6 is empirically set as 0.01. We utilize the Adam optimizer with an initial learning rate of $1e-4$ for the entire training process. We use a batch size of 8 on ScanNet and Matterport3D with a single NVIDIA A100 GPU (80G). On nuScenes, we use a batch size of 16 with 4 A100 GPUs. It takes around 24 hours to train and 0.1 second for inference. More details can be found in the supplementary file. + +# 4.2 Main Results + +We conducted a series of experiments, including traditional closed-set scenarios, in-domain open vocabulary scenarios, and out-of-domain open vocabulary scenarios to test the effectiveness of the proposed methods in various 3D scene understanding tasks. + +Comparison on Standard Benchmarks. We compare our approach with both fully-supervised and open vocabulary methods on all classes of the ScanNet [10] validation set and nuScenes [2] validation set. As shown in the bottom panel of Table 1, our approach, GGSD, consistently outperforms recent open vocabulary methods PLA [12], CLIP2Scene [6], CNS [5] and CLIP-FO3D [66] by a large margin on both datasets. Additionally, our method exhibits significant performance improvements compared to the baseline model OpenScene [45]. + +
ScanNet [10]nuScenes [2]
mIoUmAccmIoUmAcc
Fully-supervised methods
TangentConv [55]40.9---
TextureNet [27]54.8---
ScanComplete [11]56.6---
DCM-Net [53]65.8---
Mix3D [43]73.6---
VMNet [26]73.2---
LidarMultiNet [63]--82.0-
MinkowskiNet [9]69.077.578.083.7
Open vocabulary methods
MSeg [33] Voting45.654.431.036.9
PLA [12]17.733.5--
CLIP2Scene [6]25.1-20.8-
CNS [5]26.8-33.5-
CLIP-FO3D [66]30.249.1--
OpenScene†[45]54.266.642.161.8
OpenScene‡[45]52.963.242.957.1
GGSD (Ours)56.568.646.159.2
+ +Table 1: Comparisons on Standard Benchmarks. The numbers in gray are obtained from previous papers. The symbol $\dagger$ indicates the use of the 2D-3D ensemble strategy, while $\ddagger$ denotes the pure 3D results obtained from the official open-source model. Our method surpasses the current state-of-the-art by a large margin in both indoor and outdoor scenarios without using the 2D-3D ensemble strategy. + +Specifically, on the ScanNet dataset, GGSD achieves a $3.6\%$ increase in mIoU and a $5.4\%$ increase in mACC. On the nuScenes dataset, GGSD demonstrates a $3.2\%$ increase in mIoU and a $2.1\%$ increase in mAcc. Except for OpenScene, the other methods in Table 1 do not employ the 2D-3D ensemble strategy due to the significant computational and storage costs associated with maintaining a 2D segmentation network and processing a large number of images per scene during testing. It is worth noting that our method using only point clouds still outperforms OpenScene with the ensemble strategy. Given that our method is evaluated on both indoor and outdoor scenes, and it even surpasses OpenScene with ensemble strategy, these results clearly demonstrate the effectiveness and superiority of GGSD. + +Comparison on Open-Vocabulary Scenarios. The standard ScanNet benchmark only contains a small vocabulary of 20 classes. To evaluate the open vocabulary capability of GGSD, we expand the original vocabulary size by using the NYU-40 label set. We remove NYU-40 labels that lack specific semantics, such as "other structure", "other furniture", and "other prop". The remaining categories are evenly divided into Head, Common, and Tail based on the number of samples in each category (see supplementary file for specific division). We present the quantitative results on these three categories in Table 2. + +
HeadCommonTail
mIoUmAccmIoUmAccmIoUmAcc
Inference by feature projection
MaskCLIP-3D19.833.113.326.87.515.3
OpenScene‡[45]64.476.827.835.114.921.4
GGSD (Ours)66.878.830.140.916.024.2
+ +Table 2: Comparison on open-vocabulary scenarios. Open-vocabulary semantic segmentation on ScanNet with extended labels from the NYU label set. Categories are divided into Head, Common and Tail by point numbers. The numbers in gray are obtained from previous papers. The symbol $\ddagger$ denotes the pure 3D point results obtained from the official open-source model. + +
Matterport21Matterport40Matterport80Matterport160
mIoUmAccmIoUmAccmIoUmAccmIoUmAcc
OpenScene‡[45]36.048.021.127.510.813.96.08.1
GGSD (Ours)40.154.422.831.611.916.26.39.6
+ +Table 3: Comparison on cross-domain scenarios. Both GGSD and OpenScene are trained on ScanNet, and zero-shot tested on the Matterport3D dataset. The symbol $\ddagger$ denotes the pure 3D results obtained from the official open-source model. $K = 21$ is derived from the original Matterport3D benchmark. For $K = 40,80,160$ , we use the $K$ most common categories from the NYU label set provided in the benchmark. + +GGSD achieves superior results over OpenScene in all categories, with a specific improvement of $2.4\%$ and $2.3\%$ in mIOU for the Head and Common categories, respectively. Since the categories in the Common set typically consist of objects with smaller sizes (e.g., night stand and floor mat), the $2.3\%$ mIOU improvement demonstrates the stronger open vocabulary capability of our method. However, for the Tail category, although GGSD surpasses OpenScene by $2.8\%$ in mACC, the performance of both OpenScene and GGSD is not satisfactory. This can be attributed to the small sizes and the limited number of samples in these categories (e.g., lamp, box, bag, and book). + +Comparison on Cross-Domain Scenarios. Our GGSD method has demonstrated excellent potential in solving open vocabulary scene understanding tasks within a specific domain. However, the exploration of open vocabulary learners that can transfer across datasets is also worth investigation, as they face the dual challenges of category and data distribution. We conduct zero-shot domain transfer experiments by training the model on ScanNet [10] and testing it on the Matterport3D [3] dataset without fine-tuning. It should be noted that as $K$ increases, the test set will contain more small-sized objects. As shown in Table 3, our GGSD method consistently outperforms OpenScene in Matterport21, Matterport40, Matterport80, and Matterport160, demonstrating its strong open vocabulary capability and generalization ability. + +
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
2D Fusion Projection50.062.732.340.0
Pixel-Point Distillation52.963.236.148.0
Geometry Guided Distillation53.565.036.749.3
Self-Distillation56.168.239.053.3
Geometry Guided Self-distillation56.568.640.154.4
+ +Table 4: Ablation study of GGSD. Comparison of semantic segmentation performance of each component of our GGSD method. + +
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Geometry Guided Distillation54.065.036.749.3
Fixed 2D model51.964.634.245.7
Fixed 3D model55.065.837.349.8
EMA model56.168.239.053.3
+ +Table 5: Comparisons of the 3D supervision signals provided by different models. + +# 4.3 Ablation Studies and Analysis + +In this section, we investigate our key components through an in-depth ablation study. The experiments are conducted on the ScanNet dataset and subsequently tested on both ScanNet and Matterport3D datasets. + +Component Analysis. As shown in Table 4, even without utilizing 3D geometric priors, pixel-point distillation can outperform 2D direct projection, demonstrating the advantage of 3D representation. By incorporating 3D geometric priors, the geometry guided distillation method further enhances performance, resulting in a $0.6\%$ increase in mIoU and a $1.8\%$ increase in mAcc on the ScanNet dataset. Through self-distillation, we achieve a notable enhancement in performance. Specifically, on the ScanNet dataset, there is a substantial increase of $2.6\%$ in mIoU and $3.2\%$ in mAcc. On the Matterport3D dataset, the mIoU is increased by $2.3\%$ , and the mAcc is improved by $4.0\%$ . Finally, the utilization of the voting mechanism within the superpoints leads to improved performance by mitigating the impact of noise through the refinement model's predictions. It is noteworthy that our GGSD method demonstrates substantial improvements compared to the 2D teacher model, i.e., 2D Fusion Projection, achieving a $+6.5\%$ mIoU and $+5.9\%$ mAcc on the ScanNet dataset, as well as a $+7.8\%$ mIoU and $+14.4\%$ mAcc on the Matterport3D dataset. These results provide strong evidence for the effectiveness of our method. + +The Effectiveness of the EMA Model. In the second stage, we employ the EMA model to provide supervised signals for self-distillation. In Table 5, we investigate the performance of different models in providing supervised signals. When using the fixed 2D model for pixel prediction and projecting it onto + +
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Pixel-Point Distillation52.963.236.148.0
SAM [31] Refine53.063.335.948.4
Geometry Guided Distillation53.565.036.749.3
+ +Table 6: Comparisons between the utilization of geometric priors and SAM [31]. + +points, we observe lower performance compared to geometry-guided distillation. This can be attributed to the fact that the 2D model provides more erroneous supervised signals. Furthermore, using the fixed first-stage distilled 3D model only results in marginal performance improvement. However, when utilizing the EMA model, we achieve the best performance, primarily because the EMA model gradually obtains more accurate prediction results as training progresses. + +The Effectiveness of 3D Geometric Priors. The recent work CNS [5] utilized the foundation model SAM [31] to refine pseudo-labels from the 2D model, achieving performance improvements on a lower baseline (27.3% mIoU in ScanNet). We also apply this approach to LSeg. Specifically, we follow the official implementation of SAM to generate masks for all images and perform average pooling within each mask to refine them. Finally, we utilize the extracted refined features for distillation. However, as shown in Table 6, this approach yields minimal improvements. Conversely, leveraging 3D geometric priors results in a 0.6% mIoU improvement in ScanNet. + +Qualitative Results. As shown in Fig. 4, our method successfully segments all categories that exist in closed-set ScanNet benchmarks, such as wall and table. Additionally, our method successfully segments some categories that are not annotated in traditional benchmarks, such as box, lamp, and clothes, demonstrating our model's excellent open vocabulary capability. On the Matterport3D, the GGSD method successfully segments unlabeled television and also identifies mislabeled pillow and coffee table in the ground truth. Furthermore, there have been instances of segmentation failures in Matterport3D. Due to linguistic ambiguity, sofa chairs are sometimes segmented as separate entities, such as sofas and chairs, while coffee tables can be segmented as coffee tables and tables. The misclassification of certain pillow points as sofas may be attributed to their strong color resemblance to sofas. In the case of nuScenes, GGSD also successfully segments unannotated barrier. These qualitative results obtained from GGSD demonstrate its excellent open vocabulary capability. More visualization results can be found in the supplementary file. + +# 5 Conclusion + +In this study, we proposed an effective approach called Geometry Guided Self-Distillation (GGSD) to learn superior 3D representations from 2D pre-trained + +![](images/825d679fcf2d5232836ede9f7225d7928ba17b191d3edbd0012a27aad8530cfa.jpg) +Fig. 4: Qualitative results. We present the qualitative results of 3D semantic segmentation on public indoor and outdoor benchmarks. + +models. We first introduced a geometry guided distillation module that leveraged 3D geometric priors to alleviate noise in 2D models. Moreover, we harnessed the representation advantages of 3D data through geometry guided self-distillation, resulting in significant improvement in the model performance for 3D scene understanding. Our GGSD method significantly outperformed existing open-vocabulary 3D scene understanding methods on both indoor and outdoor datasets. This highlighted the crucial role of geometric priors and of 3D data representation in distillation-based approaches for 3D scene understanding. + +Limitations. Though GGSD has achieved promising results in 3D open vocabulary scene understanding tasks, it has two main limitations. Firstly, while geometric priors can significantly improve the model performance on objects with pronounced geometric structures, the improvement can be limited on objects with fewer prominent geometric characteristics. Secondly, similar to other open vocabulary methods, our approach is susceptible to the influence of language ambiguity, as illustrated by the qualitative results in Fig. 4. + +Acknowledgement. This work was supported by the InnoHK program. + +# References + +1. Vit-gpt2 image captioning. https://huggingface.co/nlpconnect/vit-gpt2-image-captioning/discussions +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017) +4. Chen, A., Zhang, K., Zhang, R., Wang, Z., Lu, Y., Guo, Y., Zhang, S.: Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5291-5301 (2023) +5. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023) +7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022) +9. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019) +10. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) +1. Dai, A., Ritchie, D., Bokeloh, M., Reed, S., Sturm, J., Nießner, M.: Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2018) +2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7010-7019 (2023) +3. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.C.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022) +4. Fan, J., Wang, Y., Guan, H., Song, C., Zhang, Z.: Toward few-shot domain adaptation with perturbation-invariant representation and transferable prototypes. Frontiers of Computer Science 16(3), 163347 (2022) +5. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV (2022) + +16. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. In: ECCV (2022) +17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: European Conference on Computer Vision. pp. 540-557. Springer (2022) +18. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018) +19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021) +20. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021) +21. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2940-2949 (2020) +22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +24. Hu, W., Zhao, H., Jiang, L., Jia, J., Wong, T.T.: Bidirectional projection network for cross dimension scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14373-14382 (2021) +25. Hu, X., Wang, Y., Fan, L., Fan, J., Peng, J., Lei, Z., Li, Q., Zhang, Z.: Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857 (2024) +26. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15488-15498 (2021) +27. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4440-4449 (2019) +28. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. arXiv preprint arXiv:2210.01055 (2022) +29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +30. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023) +31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +32. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639 (2022) +33. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: Mseg: A composite dataset for multi-domain semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2879-2888 (2020) + +34. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN +35. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) +36. Li, J., He, X., Wen, Y., Gao, Y., Cheng, X., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11809-11818 (2022) +37. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv abs/2210.04150 (2022) +38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023) +39. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems 36 (2024) +40. Lüddecke, T., Ecker, A.S.: Image segmentation using text and image prompts. In: CVPR (2021) +41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +42. Misra, I., Girdhar, R., Joulin, A.: An End-to-End Transformer Model for 3D Object Detection. In: ICCV (2021) +43. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 International Conference on 3D Vision (3DV). pp. 116-125. IEEE (2021) +44. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013) +45. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023) +46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +47. Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: Langsplat: 3d language gaussian splatting. arXiv preprint arXiv:2312.16084 (2023) +48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +49. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Dense-clip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) + +50. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: 36th Conference on Neural Information Processing Systems (NIPS) (2022) +51. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022) +52. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022) +53. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: Dualconvmesh-net: Joint geodesic and euclidean convolutions on 3d meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8612-8622 (2020) +54. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023) +55. Tatarchenko, M., Park, J., Koltun, V., Zhou, Q.Y.: Tangent convolutions for dense prediction in 3d. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3887-3896 (2018) +56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022) +57. Wang, H., Shen, Y., Fei, J., Li, W., Wu, L., Wang, Y., Zhang, Z.: Pulling target to source: A new perspective on domain adaptive semantic segmentation. arXiv preprint arXiv:2305.13752 (2023) +58. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR abs/2202.03052 (2022) +59. Wang, Y., Liang, J., Xiao, J., Mei, S., Yang, Y., Zhang, Z.: Informative data mining for one-shot cross-domain semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1064-1074 (2023) +60. Wang, Y., Peng, J., Zhang, Z.: Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9092-9101 (2021) +61. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757 (2021) +62. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023) +63. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidarmulti-net: Towards a unified multi-task network for lidar perception. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3231-3240 (2023) +64. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV (2022) +65. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023) +66. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023) + +67. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022) +68. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N.C.F., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2021) +69. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022) \ No newline at end of file diff --git a/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/images.zip b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d05cd553e622e48e91f1c1da532322732f30abff --- /dev/null +++ b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d0b61a68ae7a6036b5722fae0184f6f3a91e1cd6c3568e9fa9f87a19f1ab4e1 +size 459229 diff --git a/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/layout.json b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7222509462e006e6527fa63b18baf8f4a5ad2c43 --- /dev/null +++ b/2024/Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation/layout.json @@ -0,0 +1,8616 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 111, + 471, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 111, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 144, + 111, + 471, + 148 + ], + "type": "text", + "content": "Open Vocabulary 3D Scene Understanding via Geometry Guided Self-Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "text", + "content": "Pengfei Wang" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Yuxi Wang" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Shuai Li" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Zhaoxiang Zhang" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3,4(\\boxtimes)}" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Zhen Lei" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3,4}" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "text", + "content": ", and Lei Zhang" + }, + { + "bbox": [ + 133, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1(\\boxtimes)}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 222, + 201, + 392, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 201, + 392, + 213 + ], + "spans": [ + { + "bbox": [ + 222, + 201, + 392, + 213 + ], + "type": "text", + "content": "1 The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 180, + 213, + 432, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 213, + 432, + 223 + ], + "spans": [ + { + "bbox": [ + 180, + 213, + 432, + 223 + ], + "type": "text", + "content": "2 Centre for Artificial Intelligence and Robotics, HKISI, CAS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 224, + 465, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 224, + 465, + 235 + ], + "spans": [ + { + "bbox": [ + 148, + 224, + 465, + 235 + ], + "type": "text", + "content": "3 State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 235, + 480, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 235, + 480, + 246 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 173, + 247, + 438, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 247, + 438, + 258 + ], + "spans": [ + { + "bbox": [ + 173, + 247, + 438, + 258 + ], + "type": "text", + "content": "pengfei.wang@connect.polyu.hk, zhaoxiang.zhang@ia.ac.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 258, + 369, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 258, + 369, + 268 + ], + "spans": [ + { + "bbox": [ + 244, + 258, + 369, + 268 + ], + "type": "text", + "content": "cslzhang@comp.polyu.edu.hk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 297, + 455, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 297, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 297, + 455, + 529 + ], + "type": "text", + "content": "Abstract. The scarcity of large-scale 3D-text paired data poses a great challenge on open vocabulary 3D scene understanding, and hence it is popular to leverage internet-scale 2D data and transfer their open vocabulary capabilities to 3D models through knowledge distillation. However, the existing distillation-based 3D scene understanding approaches rely on the representation capacity of 2D models, disregarding the exploration of geometric priors and inherent representational advantages offered by 3D data. In this paper, we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained models. Specifically, we first design a geometry guided distillation module to distill knowledge from 2D models, and then leverage the 3D geometric priors to alleviate the inherent noise in 2D models and enhance the representation learning process. Due to the advantages of 3D representation, the performance of the distilled 3D student model can significantly surpass that of the 2D teacher model. This motivates us to further leverage the representation advantages of 3D data through self-distillation. As a result, our proposed GGSD approach outperforms the existing open vocabulary 3D scene understanding methods by a large margin, as demonstrated by our experiments on both indoor and outdoor benchmark datasets. Codes are available at https://github.com/Wang-pengfei/GGSD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 537, + 453, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 537, + 453, + 549 + ], + "spans": [ + { + "bbox": [ + 160, + 537, + 453, + 549 + ], + "type": "text", + "content": "Keywords: 3D scene understanding " + }, + { + "bbox": [ + 160, + 537, + 453, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 537, + 453, + 549 + ], + "type": "text", + "content": " Open vocabulary " + }, + { + "bbox": [ + 160, + 537, + 453, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 537, + 453, + 549 + ], + "type": "text", + "content": " Self-distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "3D scene understanding plays a vital role in various real-world applications, including robot manipulation, virtual reality, and autonomous driving, etc. Traditional 3D scene understanding methods mostly employ supervised training on labeled 3D datasets with a limited number of classes [18,42,56]. These approaches, however, face difficulties when they encounter unknown categories in open environments. Such a limitation in practical applications motivates researchers to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 214, + 118, + 402, + 294 + ], + "blocks": [ + { + "bbox": [ + 214, + 118, + 402, + 294 + ], + "lines": [ + { + "bbox": [ + 214, + 118, + 402, + 294 + ], + "spans": [ + { + "bbox": [ + 214, + 118, + 402, + 294 + ], + "type": "image", + "image_path": "a94125fd6d8516321dc0fe7eff58ddf59c6524eb05cd7782aaa5dbadc76ec7e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 306, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 131, + 306, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 306, + 482, + 352 + ], + "type": "text", + "content": "Fig. 1: Example of the problems when utilizing 2D models, e.g., LSeg [34], for pixel-to-point distillation. We can see the noise caused by the factors such as occlusion, lighting variations and view angles in 2D images. Therefore, only employing the 2D model in distillation brings errors to the 3D model, limiting its 3D representation capacity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "explore new 3D scene understanding models with open vocabulary capability. Inspired by the success of vision-language foundation models like CLIP [48], which aligns language features to image features, in various image understanding tasks [19,28,34,50,61,67,69], an intuitive strategy is to leverage a substantial amount of paired 3D point cloud and text data for training. Unfortunately, the high cost of annotating 3D datasets (e.g., annotating a scene with 20 categories takes approximately 22.3 minutes [12]) makes it impractical to manually label enough data with numerous real-world categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "content": "An alternative yet more practical approach is to leverage internet-scale 2D data to acquire open vocabulary capability, and then transfer this capability to 3D scene understanding models via distillation [6, 12, 45, 62, 66]. For instance, PLA [12] utilizes a 2D caption model [58] to generate textual descriptions for images, which are then refined using GPT2 [1]. Subsequently, these descriptions are projected onto the 3D point cloud, and the vision-language knowledge can be passed to the 3D model via 3D-text contrastive learning. The recent state-of-the-art OpenScene [45] utilizes 2D open-vocabulary segmentors such as LSeg [34] and OpenSeg [17] to extract pixel-level 2D embeddings for the 3D point cloud. The ability of pixel-text alignment is then transferred to align 3D point to text via feature distillation, enabling 3D semantic-level scene understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Though the above mentioned distillation-based methods have shown promising results in 3D scene understanding, they basically imitate the 2D models and inherit their limitations, restricting the upper bound of representation power. As shown in Fig. 1, the 2D model may suffer from errors due to issues such as occlusion, lighting variations, and viewing angles. The misclassification of ex" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "tracted 2D features leads to erroneous supervisory signals during the distillation process of 3D point cloud learning. Instead of solely imitating 2D models, our objective is to harness both the geometric priors in 3D point clouds and the representational advantages of 3D data to guide the knowledge distillation process from pre-trained 2D models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 481, + 437 + ], + "type": "text", + "content": "To achieve the above mentioned objectives, in this paper we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained open vocabulary segmentors. Given a sparse point cloud composed of multiple semantic categories, geometrically homogeneous neighborhoods typically correspond to the same semantic information. Based on this observation, we first design a geometry guided distillation module to learn knowledge from 2D pre-trained models. Specifically, we generate geometrically homogeneous neighborhoods as superpoints and utilize their semantic consistency to guide the distillation process. The utilization of 3D geometric priors could mitigate the inherent noise of the 2D model and enhance the representation learning capability. Furthermore, due to the representation advantages of 3D data, which are less affected by factors such as lighting and viewing angles, the distilled 3D student model can exhibit superior performance compared to the 2D teacher model. This motivates us to further distill knowledge within the 3D network itself through self-distillation. We employ an exponential moving average (EMA) model to predict each 3D point cloud and perform voting within the superpoint to provide supervision signals for the online 3D model. This approach ensures that the online model could receive reliable and accurate guidance by leveraging the consensus among multiple-point predictions. Finally, our proposed GGSD method could learn robust 3D representations from the 3D distillation model itself, achieving significantly superior performance to the 2D teacher model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "type": "text", + "content": "The contributions of this work are summarized as follows. First, we analyze the limitations of pixel-point distillation methods and propose leveraging the advantages of 3D data representation through self-distillation, instead of relying solely on 2D representations. Secondly, we utilize 3D geometric priors to mitigate the noise in pixel-point distillation process, thereby enhancing the capability of representation learning. Third, we conduct experiments on both indoor and outdoor datasets, including ScanNet [10], Matterport [3], and nuScenes [2]. The results demonstrate that GGSD sets a new state-of-the-art for 3D open vocabulary scene understanding in both indoor and outdoor datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 567, + 237, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 237, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 237, + 580 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Closed-set 3D Scene Understanding. Previous studies in closed-set 3D scene understanding [9,21,24,26,27,36,43,46,52] have led to significant advancements across diverse 3D scene understanding benchmarks, encompassing tasks such as 3D object classification, 3D object detection, 3D semantic and instance segmentation, etc. Some researches [4] have explored the use of 2D pre-trained models, such as MAE [22], for pre-training point encoders. However, they mainly em" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "phasize on leveraging pre-training to enhance supervised 3D downstream tasks, rather than addressing open-vocabulary query scenarios. Existing closed-set 3D scene understanding methods excel in specific categories but often struggle to recognize new categories in open-set scenarios. Our objective is to develop a solution to effectively recognize and comprehend new categories in open-set environments, thus overcoming this limitation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 483, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 483, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 483, + 357 + ], + "type": "text", + "content": "Open-Vocabulary 2D Scene Understanding. Traditional closed-set scene understanding has seen significant performance improvements [8,14,57,59,60,65], while open-vocabulary scene understanding is also beginning to flourish. Prior studies have explored knowledge distillation from visual-language models, such as CLIP, to transfer their open vocabulary capabilities in 2D vision tasks [13,15, 16,20,35,37,40,51,64,68]. Recent research endeavors have aimed to establish correlations between dense image features and embeddings derived from language models [17, 19, 32, 34, 38, 49]. Due to their pixel-level text alignment capabilities, 2D open vocabulary scene understanding models, such as OpenSeg [17] and LSeg [34], can serve as teacher models for distillation of 3D models compared to CLIP. In this study, we rely on pre-trained 2D models with open vocabulary capabilities, specifically OpenSeg and LSeg, and propose a novel geometry-guided self-distillation strategy to conduct 3D scene-level understanding tasks without the need for any ground-truth 2D or 3D training data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 358, + 496, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 496, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 496, + 525 + ], + "type": "text", + "content": "Open-Vocabulary 3D Scene Understanding. Recent works on open-vocabulary 3D scene understanding can be classified into two categories. The first category [25,30,39,47], represented by methods like LERF [30], focuses on optimization-based approaches that utilize techniques such as NeRF [41] or 3DGS [29] to distill 2D features into 3D scenes. However, these methods lack generalization capability when confronted with new scenes. The second category [6,12,45,54], represented by methods like OpenScene [45], is based on learning-based approaches. This paper falls in the second category. CLIP2Scene [6] leverages CLIP image model to extract 2D image features and utilizes textual category features to classify pixels, which are then projected onto 3D point clouds. OpenScene [45] aggregates pixel features from multiple view images on every 3D point, achieving state-of-the-art open vocabulary 3D scene understanding performance. Despite the promising results of these methods, their simple imitation of 2D models limits the representational capacity of 3D models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 550, + 202, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 202, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 202, + 562 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "An overview of our approach is illustrated in Fig. 2. Firstly, we follow OpenScene [45] to create pixel-point feature pairs for the 3D point cloud (see Section 3.1). Subsequently, we analyze the limitations of pixel-point distillation methods (see Section 3.2). Then, we propose a geometry guided distillation module to enhance the learning of 3D representations from the 2D model (see Section 3.3). Finally, we introduce a geometry guided self-distillation module to further distill the knowledge within the 3D network itself (see Section 3.4)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 116, + 473, + 306 + ], + "blocks": [ + { + "bbox": [ + 142, + 116, + 473, + 306 + ], + "lines": [ + { + "bbox": [ + 142, + 116, + 473, + 306 + ], + "spans": [ + { + "bbox": [ + 142, + 116, + 473, + 306 + ], + "type": "image", + "image_path": "d58c3b1a390d9f2b02b655de569ad1d524f644ccdb63d53b087e25c10197cbd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 413 + ], + "type": "text", + "content": "Fig. 2: Overview of our Geometry Guided Self-Distillation (GGSD) framework, which consists of two main components: Geometry Guided Distillation and Self-Distillation. The first module leverages 3D geometric priors to mitigate the inherent noise in 2D models. In the second module, the 3D network learns from its own predictions. This self-distillation step allows the 3D network to continuously enhance its understanding and representation capabilities. Overall, the GGSD framework combines geometry guidance and self-distillation to fully exploit the valuable information hidden in 3D data, rather than solely relying on 2D pre-trained models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 434, + 336, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 336, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 336, + 445 + ], + "type": "text", + "content": "3.1 Creating Pixel-Point Feature Pairs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "text", + "content": "Following OpenScene [45], our approach begins by extracting dense per-pixel embeddings from RGB images using a pre-trained 2D vision language segmentation model, such as LSeg [34] or OpenSeg [17]. After obtaining the pixel features, we can establish correspondences between each point in the point cloud and the pixels in different RGB images by utilizing the intrinsic matrix and the world-to-camera extrinsic matrix. Next, we employ a multi-view fusion strategy to combine the features from multiple views into a single feature vector, denoted as " + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{2\\mathrm{D}} = \\phi (\\mathbf{f}_1,\\dots ,\\mathbf{f}_K)" + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "text", + "content": ", for each point cloud " + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "text", + "content": ". This fusion is achieved using an average pooling operator " + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "text", + "content": " over the " + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "text", + "content": "-view features. After obtaining the fused 2D features for each of the M points in the scene, we obtain a 2D pixel feature set " + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{2\\mathrm{D}} = \\mathbf{f}_1^{2\\mathrm{D}},\\dots ,\\mathbf{f}_M^{2\\mathrm{D}}\\in \\mathbb{R}^{M\\times C}" + }, + { + "bbox": [ + 130, + 452, + 482, + 609 + ], + "type": "text", + "content": " corresponding to the point cloud scene. Finally, these embedded features are back-projected into the 3D point cloud scene. For more details, please refer to OpenScene [45]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 623, + 370, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 370, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 370, + 634 + ], + "type": "text", + "content": "3.2 Limitations of Pixel-Point 3D Distillation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": "Due to the text-image alignment capability of CLIP, the 2D features " + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": " can be directly utilized for language-driven open vocabulary scene understanding." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "To transfer 2D vision-language knowledge to a 3D point cloud network, the distillation approach has been popularly used. One up-to-date pixel-point 3D distillation work is OpenScene [45]. Specifically, OpenScene aims to learn a 3D encoder that extracts 3D features from each point cloud, which are expected to be similar to their corresponding 2D pixel features. This allows the 3D encoder to distill the open vocabulary capability of the 2D model. To this end, OpenScene utilizes cosine distance to perform pixel-point distillation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 209, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 209, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 255, + 209, + 481, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = 1 - \\cos (\\mathbf {F} ^ {\\mathrm {2 D}}, \\mathbf {F} ^ {\\mathrm {3 D}}). \\tag {1}", + "image_path": "297e37e8f42aca214b19d9394f3c77436012f4afc71f0cc7b48a84cca51f44ec.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 437 + ], + "type": "text", + "content": "However, the current pixel-to-point distillation methods are mostly a simplistic imitation of the 2D models, which are limited in extracting useful information from 3D point data. Firstly, relying solely on the projection relationship between point clouds and pixels, these approaches fail to fully exploit the geometric priors inherent in 3D data, leading to suboptimal performance. In this paper, we propose to incorporate 3D geometric priors throughout the entire distillation process to address the issue of noise in 2D models by constraining semantic consistency. Secondly, the current pixel-point distillation methods fail to fully exploit the representational advantage offered by 3D point clouds. It has been observed that the distilled 3D model can outperform the directly projected 2D models. This can be attributed to the inherent capability of 3D data to handle challenges like lighting variations, occlusion, and viewpoint changes, which are prevalent in 2D representations. Therefore, the accuracy of 3D model segmentation results, particularly for foreground classes, can be significantly improved. This observation provides a strong motivation for us to further leverage the advantages of 3D representation within the 3D network itself via self-distillation, rather than relying solely on information from the 2D model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 453, + 311, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 311, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 311, + 466 + ], + "type": "text", + "content": "3.3 Geometry Guided Distillation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": "We first design a geometry guided distillation method to learn knowledge from 2D models. Given a sparse point cloud composed of multiple semantic categories, we can decompose it into many simple parts, i.e., superpoints, according to their geometric position in the space. As shown in Fig. 3, the superpoints for an indoor room are constructed by VCCS [44]. The details for creating superpoints can be found in the supplementary file. For a specific input point cloud " + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{M \\times 3}" + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": ", its superpoints are denoted as " + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{p}}_1 \\cdots \\tilde{\\mathbf{p}}_n \\cdots \\tilde{\\mathbf{p}}_N\\}" + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": ", where each superpoint " + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{p}}_n" + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": " consists of a small subset of original point cloud " + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 474, + 482, + 616 + ], + "type": "text", + "content": ". These superpoints are geometrically simple and usually cannot cover objects of different classes. Ideally, each superpoint as a whole should belong to the same semantic category. The simplicity and coherence of superpoints naturally facilitate the implementation of semantic consistency constraints." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To mitigate the noise issue in 2D models, we can leverage the semantic consistency provided by 3D geometric priors. In particular, for a specific input point cloud " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", we have its neural features " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{3\\mathrm{D}} = \\{\\mathbf{f}_1^{3\\mathrm{D}}, \\dots, \\mathbf{f}_M^{3\\mathrm{D}}\\}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", the corresponding pixel fusion features " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{2\\mathrm{D}} = \\{\\mathbf{f}_1^{2\\mathrm{D}}, \\dots, \\mathbf{f}_M^{2\\mathrm{D}}\\} \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and superpoints" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 180, + 116, + 294, + 232 + ], + "blocks": [ + { + "bbox": [ + 180, + 116, + 294, + 232 + ], + "lines": [ + { + "bbox": [ + 180, + 116, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 180, + 116, + 294, + 232 + ], + "type": "image", + "image_path": "3341e6b5f098d0e334bb0f17ea4fd7d4fe1fb622bda8ad9b1889ea59436c3811.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 238, + 275, + 248 + ], + "lines": [ + { + "bbox": [ + 206, + 238, + 275, + 248 + ], + "spans": [ + { + "bbox": [ + 206, + 238, + 275, + 248 + ], + "type": "text", + "content": "Input Point Cloud" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 116, + 431, + 232 + ], + "blocks": [ + { + "bbox": [ + 314, + 116, + 431, + 232 + ], + "lines": [ + { + "bbox": [ + 314, + 116, + 431, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 116, + 431, + 232 + ], + "type": "image", + "image_path": "6a132b87accc65d2d1bcca8c038c6fe1456a91835910cc43bfaef98d3296ae4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 239, + 400, + 248 + ], + "lines": [ + { + "bbox": [ + 354, + 239, + 400, + 248 + ], + "spans": [ + { + "bbox": [ + 354, + 239, + 400, + 248 + ], + "type": "text", + "content": "Superpoints" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 251, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 480, + 297 + ], + "type": "text", + "content": "Fig. 3: Example of the superpoints generated using VCCS [44] on ScanNet scene point clouds. Each colored patch corresponds to a distinct superpoint, showcasing a simple geometric structure. Additionally, the points within each superpoint consistently exhibit a shared semantic category." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 309, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 309, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 309, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{p}}_1 \\cdots \\tilde{\\mathbf{p}}_n \\cdots \\tilde{\\mathbf{p}}_N\\}" + }, + { + "bbox": [ + 131, + 309, + 479, + 335 + ], + "type": "text", + "content": ". Firstly, we compute the mean 2D pixel features for superpoints, denoted as " + }, + { + "bbox": [ + 131, + 309, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{sp}^{2D} = \\{\\tilde{\\mathbf{f}}_1^{2D}, \\cdots, \\tilde{\\mathbf{f}}_N^{2D}\\}" + }, + { + "bbox": [ + 131, + 309, + 479, + 335 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 344, + 480, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 344, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 233, + 344, + 480, + 378 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {f}} _ {n} ^ {2 \\mathrm {D}} = \\frac {1}{Q} \\sum_ {q = 1} ^ {Q} \\mathbf {f} _ {q} ^ {2 \\mathrm {D}}, \\quad \\mathbf {f} _ {q} ^ {2 \\mathrm {D}} \\in \\mathbb {R} ^ {1 \\times C}, \\tag {2}", + "image_path": "d2324e1e8f305c7e0027cd5f46ad2c253bcf177db2b43e776ea8883f1e90677d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "text", + "content": " is the total number of 3D points within a superpoint " + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{p}}_n" + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_q^{2\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "text", + "content": " is the feature vector for the " + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "inline_equation", + "content": "q^{th}" + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "text", + "content": " pixel that corresponds to the 3D point in the superpoint. Similarly, we can employ the same approach to calculate the mean 3D point features " + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{sp}^{3\\mathrm{D}} = \\{\\tilde{\\mathbf{f}}_1^{3\\mathrm{D}}, \\dots, \\tilde{\\mathbf{f}}_N^{3\\mathrm{D}}\\}" + }, + { + "bbox": [ + 130, + 388, + 479, + 438 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 439, + 479, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 479, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 479, + 466 + ], + "type": "text", + "content": "To enforce the superpoint feature " + }, + { + "bbox": [ + 130, + 439, + 479, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{sp}^{3D}" + }, + { + "bbox": [ + 130, + 439, + 479, + 466 + ], + "type": "text", + "content": " to be consistent with the 2D fused features " + }, + { + "bbox": [ + 130, + 439, + 479, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{sp}^{2D}" + }, + { + "bbox": [ + 130, + 439, + 479, + 466 + ], + "type": "text", + "content": ", we use a cosine similarity loss as that in pixel-point distillation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 474, + 480, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 474, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 249, + 474, + 480, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s p} = 1 - \\cos \\left(\\mathbf {F} _ {s p} ^ {\\mathrm {2 D}}, \\mathbf {F} _ {s p} ^ {\\mathrm {3 D}}\\right). \\tag {3}", + "image_path": "341fb3d23bda89d1cc5e967bff352ae11018bec9b63f57efa9a20d5ceed1f818.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 498, + 472, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 472, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 472, + 510 + ], + "type": "text", + "content": "The final optimization objective for geometry guided distillation is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 271, + 519, + 480, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 519, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 271, + 519, + 480, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\mathcal {L} _ {p} + \\mathcal {L} _ {s p}. \\tag {4}", + "image_path": "6ed85cc10fe5c8fb1e5cdb5dc0e17e3d1a697d34e895a04c26dc90883f92c8b6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 549, + 332, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 332, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 332, + 562 + ], + "type": "text", + "content": "3.4 Geometry Guided Self-Distillation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "With geometry guided distillation, the performance of the 3D model can surpass that of the directly projected 2D model. This inspires us to further design a self-distillation module to fully leverage the inherent advantages of 3D data in representation. Recent methods such as CNS [5] and CLIP2Scene [6] typically transfer the pixel-level pseudo-labels from the 2D image to the 3D point cloud using a 2D-3D calibration matrix. In contrast to existing methods that rely on projecting from 2D, we introduce a geometry guided self-distillation approach to distill knowledge from the 3D model itself. Specifically, we generate predictions" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "directly on 3D point clouds instead of projecting 2D predictions back to 3D, as our 3D model possesses enhanced feature representation capabilities. Specifically, given the 3D point features " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{3\\mathrm{D}} = \\{\\mathbf{f}_1^{3\\mathrm{D}},\\dots ,\\mathbf{f}_M^{3\\mathrm{D}}\\} \\in \\mathbb{R}^{M\\times C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", we assign pseudolabels to each point " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 175, + 481, + 195 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 175, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 255, + 175, + 481, + 195 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {\\hat {t}} = \\underset {l} {\\arg \\max } \\psi \\left(\\mathbf {f} _ {n} ^ {\\mathrm {3 D}}, \\mathbf {f} _ {l} ^ {t}\\right), \\tag {5}", + "image_path": "c663d48b35380a6e43599cd52b0b16229acf4c6c72e86663ea6e8a3347e26ced.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l^t" + }, + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "text", + "content": " is CLIP's text embedding of the " + }, + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "text", + "content": "-th class name with a pre-defined template. " + }, + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 205, + 482, + 252 + ], + "type": "text", + "content": " is the dot product operation. Furthermore, in the self-distillation stage, we also incorporate 3D geometric priors to enhance the semantic consistency of the predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 254, + 482, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 349 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 349 + ], + "type": "text", + "content": "However, the research findings from self-supervised methods [7,23] have highlighted that directly using the network's predictions as supervision signals can lead to model collapse due to overconfidence. To address this issue, we leverage the predictions from the exponential moving average (EMA) model as supervision signals to train the 3D point cloud network, aiming to enhance stability and mitigate the impact of overconfidence. Specifically, we use the features extracted by the EMA model to calculate the similarity with the text feature according to Equation 5, resulting in generating a point-text pair for each point cloud." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 483, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 483, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 483, + 505 + ], + "type": "text", + "content": "Although using EMA model can prevent mode collapse, it is still inevitable to have noise in the prediction results. To mitigate this issue, we incorporate the semantic consistency of 3D geometric priors during the self-distillation process. Specifically, we implement a voting mechanism within each superpoint to refine the prediction results. After assigning the most similar textual feature to each point cloud according to Equation 5, we calculate the proportion of textual features within each superpoint. We then modify the text corresponding to all point clouds within that superpoint to match the textual feature with the highest proportion. This correction process effectively mitigates the impact of noise and improves the accuracy of prediction. The voting outcome generates point-language pairs for each point cloud, then we can pull paired point and language features together while pushing unmatched features away through contrastive loss. The geometry guided self-distillation loss is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 508, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 508, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 233, + 508, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s d} = - \\log \\frac {\\exp \\left(\\mathbf {f} ^ {\\mathrm {3 D}} \\cdot \\mathbf {f} ^ {\\hat {t}} / \\tau\\right)}{\\sum_ {i = 1} ^ {n _ {t}} \\exp \\left(\\mathbf {f} ^ {\\mathrm {3 D}} \\cdot \\mathbf {f} _ {i} ^ {t} / \\tau\\right)}, \\tag {6}", + "image_path": "2c5c022a6c403a99933cc4ed4201b974bc4b8e7d8c0a96f86e4234be5ab5f027.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^t" + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "text", + "content": " represents the textual features corresponding to the point cloud " + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "text", + "content": " is the number of point-language pairs and " + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 538, + 482, + 564 + ], + "type": "text", + "content": " is a temperature factor." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 583, + 276, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 583, + 276, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 583, + 276, + 596 + ], + "type": "text", + "content": "3.5 Training and Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Training. The training process comprises two stages. In the initial stage, we employ geometry guided distillation to train the 3D network for 70 epochs, enabling it to acquire knowledge from the 2D model. Then, we incorporate the GGSD module into an additional 30 training epochs. The total number of training epochs is consistent with OpenScene." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Inference. Using any per-point feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{3D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " and CLIP text features obtained from a set of text prompts, we can estimate their similarities by calculating the cosine similarity score between them. This similarity score is utilized in all of our scene understanding tasks. In all experiments, unless otherwise stated, we exclusively use pure 3D point clouds for predictions during the inference stage, without employing the 2D-3D ensemble strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 230, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 230, + 218 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 230, + 218 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 229, + 253, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 253, + 242 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 253, + 242 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 330 + ], + "type": "text", + "content": "Datasets. To evaluate the effectiveness of our method, we conduct experiments on three widely used public benchmarks: ScanNet [10], Matterport [3], and nuScenes [2]. The first two datasets provide RGBD images and 3D meshes of indoor scenes, while the last one offers LIDAR scans of outdoor scenes. Additionally, Matterport3D is a challenging dataset with intricately detailed scenes, which allows us to thoroughly examine the performance of open-vocabulary queries across different datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "content": "Implementation Details. We adopt the settings outlined in OpenScene [45], utilize MinkowskiNet18A [9] as the 3D backbone network, and employ a simple prompt engineering technique to extract CLIP text features. Specifically, we modify the text prompts for each object class \"XX\" to \"a XX in a scene\", such as \"a chair in a scene\". On the ScanNet and Matterport3D datasets, we employ a voxel size of " + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "content": " with MinkowskiNet while utilizing a voxel size of " + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "content": " on the nuScenes dataset. Based on the findings of OpenScene, we employ LSeg [34] for indoor datasets and OpenSeg [17] for outdoor dataset as the 2D teacher model. The temperature " + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "content": " in Eq. 6 is empirically set as 0.01. We utilize the Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "inline_equation", + "content": "1e-4" + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "content": " for the entire training process. We use a batch size of 8 on ScanNet and Matterport3D with a single NVIDIA A100 GPU (80G). On nuScenes, we use a batch size of 16 with 4 A100 GPUs. It takes around 24 hours to train and 0.1 second for inference. More details can be found in the supplementary file." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 515, + 228, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 228, + 526 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 228, + 526 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 582 + ], + "type": "text", + "content": "We conducted a series of experiments, including traditional closed-set scenarios, in-domain open vocabulary scenarios, and out-of-domain open vocabulary scenarios to test the effectiveness of the proposed methods in various 3D scene understanding tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Comparison on Standard Benchmarks. We compare our approach with both fully-supervised and open vocabulary methods on all classes of the ScanNet [10] validation set and nuScenes [2] validation set. As shown in the bottom panel of Table 1, our approach, GGSD, consistently outperforms recent open vocabulary methods PLA [12], CLIP2Scene [6], CNS [5] and CLIP-FO3D [66] by a large margin on both datasets. Additionally, our method exhibits significant performance improvements compared to the baseline model OpenScene [45]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 204, + 114, + 413, + 328 + ], + "blocks": [ + { + "bbox": [ + 204, + 114, + 413, + 328 + ], + "lines": [ + { + "bbox": [ + 204, + 114, + 413, + 328 + ], + "spans": [ + { + "bbox": [ + 204, + 114, + 413, + 328 + ], + "type": "table", + "html": "
ScanNet [10]nuScenes [2]
mIoUmAccmIoUmAcc
Fully-supervised methods
TangentConv [55]40.9---
TextureNet [27]54.8---
ScanComplete [11]56.6---
DCM-Net [53]65.8---
Mix3D [43]73.6---
VMNet [26]73.2---
LidarMultiNet [63]--82.0-
MinkowskiNet [9]69.077.578.083.7
Open vocabulary methods
MSeg [33] Voting45.654.431.036.9
PLA [12]17.733.5--
CLIP2Scene [6]25.1-20.8-
CNS [5]26.8-33.5-
CLIP-FO3D [66]30.249.1--
OpenScene†[45]54.266.642.161.8
OpenScene‡[45]52.963.242.957.1
GGSD (Ours)56.568.646.159.2
", + "image_path": "481394f58f8c278c0317a58ba0bc117b25c63242b4596442b094b528a889f02a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": "Table 1: Comparisons on Standard Benchmarks. The numbers in gray are obtained from previous papers. The symbol " + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": " indicates the use of the 2D-3D ensemble strategy, while " + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": " denotes the pure 3D results obtained from the official open-source model. Our method surpasses the current state-of-the-art by a large margin in both indoor and outdoor scenarios without using the 2D-3D ensemble strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "text", + "content": "Specifically, on the ScanNet dataset, GGSD achieves a " + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "text", + "content": " increase in mIoU and a " + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "text", + "content": " increase in mACC. On the nuScenes dataset, GGSD demonstrates a " + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "text", + "content": " increase in mIoU and a " + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 433, + 482, + 567 + ], + "type": "text", + "content": " increase in mAcc. Except for OpenScene, the other methods in Table 1 do not employ the 2D-3D ensemble strategy due to the significant computational and storage costs associated with maintaining a 2D segmentation network and processing a large number of images per scene during testing. It is worth noting that our method using only point clouds still outperforms OpenScene with the ensemble strategy. Given that our method is evaluated on both indoor and outdoor scenes, and it even surpasses OpenScene with ensemble strategy, these results clearly demonstrate the effectiveness and superiority of GGSD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Comparison on Open-Vocabulary Scenarios. The standard ScanNet benchmark only contains a small vocabulary of 20 classes. To evaluate the open vocabulary capability of GGSD, we expand the original vocabulary size by using the NYU-40 label set. We remove NYU-40 labels that lack specific semantics, such as \"other structure\", \"other furniture\", and \"other prop\". The remaining categories are evenly divided into Head, Common, and Tail based on the number of samples in each category (see supplementary file for specific division). We present the quantitative results on these three categories in Table 2." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 114, + 424, + 191 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 424, + 191 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 424, + 191 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 424, + 191 + ], + "type": "table", + "html": "
HeadCommonTail
mIoUmAccmIoUmAccmIoUmAcc
Inference by feature projection
MaskCLIP-3D19.833.113.326.87.515.3
OpenScene‡[45]64.476.827.835.114.921.4
GGSD (Ours)66.878.830.140.916.024.2
", + "image_path": "062a5c34ec4c7443834c3bb5e5e97864a629c85f59a39e9c7ae37d919b1bf577.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 161, + 276, + 455, + 328 + ], + "blocks": [ + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "type": "text", + "content": "Table 2: Comparison on open-vocabulary scenarios. Open-vocabulary semantic segmentation on ScanNet with extended labels from the NYU label set. Categories are divided into Head, Common and Tail by point numbers. The numbers in gray are obtained from previous papers. The symbol " + }, + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "type": "text", + "content": " denotes the pure 3D point results obtained from the official open-source model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 276, + 455, + 328 + ], + "lines": [ + { + "bbox": [ + 161, + 276, + 455, + 328 + ], + "spans": [ + { + "bbox": [ + 161, + 276, + 455, + 328 + ], + "type": "table", + "html": "
Matterport21Matterport40Matterport80Matterport160
mIoUmAccmIoUmAccmIoUmAccmIoUmAcc
OpenScene‡[45]36.048.021.127.510.813.96.08.1
GGSD (Ours)40.154.422.831.611.916.26.39.6
", + "image_path": "62a6e1b1428e20312e3bb60da88e1d51f95b6e1cb6403e287162532565196370.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": "Table 3: Comparison on cross-domain scenarios. Both GGSD and OpenScene are trained on ScanNet, and zero-shot tested on the Matterport3D dataset. The symbol " + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": " denotes the pure 3D results obtained from the official open-source model. " + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "inline_equation", + "content": "K = 21" + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": " is derived from the original Matterport3D benchmark. For " + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "inline_equation", + "content": "K = 40,80,160" + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": ", we use the " + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 336, + 482, + 392 + ], + "type": "text", + "content": " most common categories from the NYU label set provided in the benchmark." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": "GGSD achieves superior results over OpenScene in all categories, with a specific improvement of " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": " in mIOU for the Head and Common categories, respectively. Since the categories in the Common set typically consist of objects with smaller sizes (e.g., night stand and floor mat), the " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": " mIOU improvement demonstrates the stronger open vocabulary capability of our method. However, for the Tail category, although GGSD surpasses OpenScene by " + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": " in mACC, the performance of both OpenScene and GGSD is not satisfactory. This can be attributed to the small sizes and the limited number of samples in these categories (e.g., lamp, box, bag, and book)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Comparison on Cross-Domain Scenarios. Our GGSD method has demonstrated excellent potential in solving open vocabulary scene understanding tasks within a specific domain. However, the exploration of open vocabulary learners that can transfer across datasets is also worth investigation, as they face the dual challenges of category and data distribution. We conduct zero-shot domain transfer experiments by training the model on ScanNet [10] and testing it on the Matterport3D [3] dataset without fine-tuning. It should be noted that as " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " increases, the test set will contain more small-sized objects. As shown in Table 3, our GGSD method consistently outperforms OpenScene in Matterport21, Matterport40, Matterport80, and Matterport160, demonstrating its strong open vocabulary capability and generalization ability." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 178, + 114, + 438, + 193 + ], + "blocks": [ + { + "bbox": [ + 178, + 114, + 438, + 193 + ], + "lines": [ + { + "bbox": [ + 178, + 114, + 438, + 193 + ], + "spans": [ + { + "bbox": [ + 178, + 114, + 438, + 193 + ], + "type": "table", + "html": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
2D Fusion Projection50.062.732.340.0
Pixel-Point Distillation52.963.236.148.0
Geometry Guided Distillation53.565.036.749.3
Self-Distillation56.168.239.053.3
Geometry Guided Self-distillation56.568.640.154.4
", + "image_path": "72fc6aa754e0326cfae86d715bc71cc7176d069e957960012d6facd3e3f2c581.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 178, + 243, + 438, + 317 + ], + "blocks": [ + { + "bbox": [ + 132, + 201, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 480, + 222 + ], + "type": "text", + "content": "Table 4: Ablation study of GGSD. Comparison of semantic segmentation performance of each component of our GGSD method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 243, + 438, + 317 + ], + "lines": [ + { + "bbox": [ + 178, + 243, + 438, + 317 + ], + "spans": [ + { + "bbox": [ + 178, + 243, + 438, + 317 + ], + "type": "table", + "html": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Geometry Guided Distillation54.065.036.749.3
Fixed 2D model51.964.634.245.7
Fixed 3D model55.065.837.349.8
EMA model56.168.239.053.3
", + "image_path": "e41c89647602306f1436d22fdd367e587988c65d78723aa22801614a484d6b5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 325, + 474, + 337 + ], + "lines": [ + { + "bbox": [ + 138, + 325, + 474, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 474, + 337 + ], + "type": "text", + "content": "Table 5: Comparisons of the 3D supervision signals provided by different models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 370, + 313, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 313, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 313, + 383 + ], + "type": "text", + "content": "4.3 Ablation Studies and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 390, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 426 + ], + "type": "text", + "content": "In this section, we investigate our key components through an in-depth ablation study. The experiments are conducted on the ScanNet dataset and subsequently tested on both ScanNet and Matterport3D datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": "Component Analysis. As shown in Table 4, even without utilizing 3D geometric priors, pixel-point distillation can outperform 2D direct projection, demonstrating the advantage of 3D representation. By incorporating 3D geometric priors, the geometry guided distillation method further enhances performance, resulting in a " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " increase in mIoU and a " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " increase in mAcc on the ScanNet dataset. Through self-distillation, we achieve a notable enhancement in performance. Specifically, on the ScanNet dataset, there is a substantial increase of " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " in mAcc. On the Matterport3D dataset, the mIoU is increased by " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": ", and the mAcc is improved by " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": ". Finally, the utilization of the voting mechanism within the superpoints leads to improved performance by mitigating the impact of noise through the refinement model's predictions. It is noteworthy that our GGSD method demonstrates substantial improvements compared to the 2D teacher model, i.e., 2D Fusion Projection, achieving a " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "+6.5\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "+5.9\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " mAcc on the ScanNet dataset, as well as a " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "+7.8\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "inline_equation", + "content": "+14.4\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 616 + ], + "type": "text", + "content": " mAcc on the Matterport3D dataset. These results provide strong evidence for the effectiveness of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "The Effectiveness of the EMA Model. In the second stage, we employ the EMA model to provide supervised signals for self-distillation. In Table 5, we investigate the performance of different models in providing supervised signals. When using the fixed 2D model for pixel prediction and projecting it onto" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 447, + 180 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 447, + 180 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 447, + 180 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 447, + 180 + ], + "type": "table", + "html": "
ScanNet [10]Matterport3D [3]
mIoUmAccmIoUmAcc
Pixel-Point Distillation52.963.236.148.0
SAM [31] Refine53.063.335.948.4
Geometry Guided Distillation53.565.036.749.3
", + "image_path": "0a8f265d939766f5628b723b90e255b0fab830d5a68741a1efac1615b16708b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 189, + 472, + 201 + ], + "lines": [ + { + "bbox": [ + 140, + 189, + 472, + 201 + ], + "spans": [ + { + "bbox": [ + 140, + 189, + 472, + 201 + ], + "type": "text", + "content": "Table 6: Comparisons between the utilization of geometric priors and SAM [31]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 235, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 479, + 307 + ], + "type": "text", + "content": "points, we observe lower performance compared to geometry-guided distillation. This can be attributed to the fact that the 2D model provides more erroneous supervised signals. Furthermore, using the fixed first-stage distilled 3D model only results in marginal performance improvement. However, when utilizing the EMA model, we achieve the best performance, primarily because the EMA model gradually obtains more accurate prediction results as training progresses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 413 + ], + "type": "text", + "content": "The Effectiveness of 3D Geometric Priors. The recent work CNS [5] utilized the foundation model SAM [31] to refine pseudo-labels from the 2D model, achieving performance improvements on a lower baseline (27.3% mIoU in ScanNet). We also apply this approach to LSeg. Specifically, we follow the official implementation of SAM to generate masks for all images and perform average pooling within each mask to refine them. Finally, we utilize the extracted refined features for distillation. However, as shown in Table 6, this approach yields minimal improvements. Conversely, leveraging 3D geometric priors results in a 0.6% mIoU improvement in ScanNet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 594 + ], + "type": "text", + "content": "Qualitative Results. As shown in Fig. 4, our method successfully segments all categories that exist in closed-set ScanNet benchmarks, such as wall and table. Additionally, our method successfully segments some categories that are not annotated in traditional benchmarks, such as box, lamp, and clothes, demonstrating our model's excellent open vocabulary capability. On the Matterport3D, the GGSD method successfully segments unlabeled television and also identifies mislabeled pillow and coffee table in the ground truth. Furthermore, there have been instances of segmentation failures in Matterport3D. Due to linguistic ambiguity, sofa chairs are sometimes segmented as separate entities, such as sofas and chairs, while coffee tables can be segmented as coffee tables and tables. The misclassification of certain pillow points as sofas may be attributed to their strong color resemblance to sofas. In the case of nuScenes, GGSD also successfully segments unannotated barrier. These qualitative results obtained from GGSD demonstrate its excellent open vocabulary capability. More visualization results can be found in the supplementary file." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 614, + 220, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 220, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 220, + 627 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "In this study, we proposed an effective approach called Geometry Guided Self-Distillation (GGSD) to learn superior 3D representations from 2D pre-trained" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 482, + 361 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 482, + 361 + ], + "type": "image", + "image_path": "825d679fcf2d5232836ede9f7225d7928ba17b191d3edbd0012a27aad8530cfa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 376, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 131, + 376, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 482, + 399 + ], + "type": "text", + "content": "Fig. 4: Qualitative results. We present the qualitative results of 3D semantic segmentation on public indoor and outdoor benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 448, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 545 + ], + "type": "text", + "content": "models. We first introduced a geometry guided distillation module that leveraged 3D geometric priors to alleviate noise in 2D models. Moreover, we harnessed the representation advantages of 3D data through geometry guided self-distillation, resulting in significant improvement in the model performance for 3D scene understanding. Our GGSD method significantly outperformed existing open-vocabulary 3D scene understanding methods on both indoor and outdoor datasets. This highlighted the crucial role of geometric priors and of 3D data representation in distillation-based approaches for 3D scene understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "text", + "content": "Limitations. Though GGSD has achieved promising results in 3D open vocabulary scene understanding tasks, it has two main limitations. Firstly, while geometric priors can significantly improve the model performance on objects with pronounced geometric structures, the improvement can be limited on objects with fewer prominent geometric characteristics. Secondly, similar to other open vocabulary methods, our approach is susceptible to the influence of language ambiguity, as illustrated by the qualitative results in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 653, + 454, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 454, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 454, + 666 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the InnoHK program." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 142, + 479, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 479, + 164 + ], + "type": "text", + "content": "1. Vit-gpt2 image captioning. https://huggingface.co/nlpconnect/vit-gpt2-image-captioning/discussions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 209 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "text", + "content": "3. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 480, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 286 + ], + "type": "text", + "content": "4. Chen, A., Zhang, K., Zhang, R., Wang, Z., Lu, Y., Guo, Y., Zhang, S.: Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5291-5301 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 480, + 320 + ], + "type": "text", + "content": "5. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 364 + ], + "type": "text", + "content": "6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "text", + "content": "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 399, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 441 + ], + "type": "text", + "content": "8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "type": "text", + "content": "9. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "type": "text", + "content": "10. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 510, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 553 + ], + "type": "text", + "content": "1. Dai, A., Ritchie, D., Bokeloh, M., Reed, S., Sturm, J., Nießner, M.: Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4578-4587 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "text", + "content": "2. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7010-7019 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "type": "text", + "content": "3. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.C.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "4. Fan, J., Wang, Y., Guan, H., Song, C., Zhang, Z.: Toward few-shot domain adaptation with perturbation-invariant representation and transferable prototypes. Frontiers of Computer Science 16(3), 163347 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "5. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. In: ECCV (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "type": "text", + "content": "17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: European Conference on Computer Vision. pp. 540-557. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 170, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 480, + 203 + ], + "type": "text", + "content": "18. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "19. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "text", + "content": "20. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 245, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 480, + 277 + ], + "type": "text", + "content": "21. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2940-2949 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "type": "text", + "content": "22. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "text", + "content": "23. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "24. Hu, W., Zhao, H., Jiang, L., Jia, J., Wong, T.T.: Bidirectional projection network for cross dimension scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14373-14382 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "25. Hu, X., Wang, Y., Fan, L., Fan, J., Peng, J., Lei, Z., Li, Q., Zhang, Z.: Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 438 + ], + "type": "text", + "content": "26. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15488-15498 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 438, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 480, + 482 + ], + "type": "text", + "content": "27. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4440-4449 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "28. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. arXiv preprint arXiv:2210.01055 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "text", + "content": "29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 480, + 567 + ], + "type": "text", + "content": "30. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "32. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "33. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: Mseg: A composite dataset for multi-domain semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2879-2888 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "35. Li, B., Weinberger, K.Q., Belongie, S.J., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "text", + "content": "36. Li, J., He, X., Wen, Y., Gao, Y., Cheng, X., Zhang, D.: Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11809-11818 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "37. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv abs/2210.04150 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "text", + "content": "38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "39. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "text", + "content": "40. Lüddecke, T., Ecker, A.S.: Image segmentation using text and image prompts. In: CVPR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "42. Misra, I., Girdhar, R., Joulin, A.: An End-to-End Transformer Model for 3D Object Detection. In: ICCV (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "43. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 International Conference on 3D Vision (3DV). pp. 116-125. IEEE (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "44. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "text", + "content": "45. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "46. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "type": "text", + "content": "47. Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: Langsplat: 3d language gaussian splatting. arXiv preprint arXiv:2312.16084 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "type": "text", + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "text", + "content": "49. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Dense-clip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "50. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: 36th Conference on Neural Information Processing Systems (NIPS) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 480, + 183 + ], + "type": "text", + "content": "51. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "text", + "content": "52. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 261 + ], + "type": "text", + "content": "53. Schult, J., Engelmann, F., Kontogianni, T., Leibe, B.: Dualconvmesh-net: Joint geodesic and euclidean convolutions on 3d meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8612-8622 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "text", + "content": "54. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 480, + 328 + ], + "type": "text", + "content": "55. Tatarchenko, M., Park, J., Koltun, V., Zhou, Q.Y.: Tangent convolutions for dense prediction in 3d. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3887-3896 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 351 + ], + "type": "text", + "content": "56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "text", + "content": "57. Wang, H., Shen, Y., Fei, J., Li, W., Wu, L., Wang, Y., Zhang, Z.: Pulling target to source: A new perspective on domain adaptive semantic segmentation. arXiv preprint arXiv:2305.13752 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "58. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR abs/2202.03052 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "text", + "content": "59. Wang, Y., Liang, J., Xiao, J., Mei, S., Yang, Y., Zhang, Z.: Informative data mining for one-shot cross-domain semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1064-1074 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "type": "text", + "content": "60. Wang, Y., Peng, J., Zhang, Z.: Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9092-9101 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "type": "text", + "content": "61. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "type": "text", + "content": "62. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. arXiv preprint arXiv:2304.00962 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "text", + "content": "63. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidarmulti-net: Towards a unified multi-task network for lidar perception. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3231-3240 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "64. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "text", + "content": "65. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "66. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "P. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 205 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "67. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "68. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N.C.F., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "69. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "GGSD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_content_list.json b/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5d478bcc12357560d7f9a6e6adc7739b85c7abcb --- /dev/null +++ b/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Open Vocabulary Multi-Label Video Classification", + "text_level": 1, + "bbox": [ + 292, + 140, + 709, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohit Gupta1, Mamshad Nayeem Rizve2, Jayakrishnan Unnikrishnan2,", + "bbox": [ + 238, + 210, + 761, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ashish Tawari $^{2}$ , Son Tran $^{2}$ , Mubarak Shah $^{1,2}$", + "bbox": [ + 334, + 227, + 666, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benjamin Yao $^{2}$ , and Trishul Chilimbi $^{2}$", + "bbox": [ + 362, + 243, + 638, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Center for Research in Computer Vision, University of Central Florida", + "bbox": [ + 256, + 268, + 743, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rohit.gupta@ucf.edu, shah@crv.ucf.edu", + "bbox": [ + 352, + 285, + 648, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Amazon", + "bbox": [ + 465, + 297, + 535, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jayunn,atakari,sontran}@amazon.com", + "bbox": [ + 366, + 311, + 635, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Pre-trained vision-language models (VLMs) have enabled significant progress in open vocabulary computer vision tasks such as image classification, object detection and image segmentation. Some recent works have focused on extending VLMs to open vocabulary single label action classification in videos. However, previous methods fall short in holistic video understanding which requires the ability to simultaneously recognize multiple actions and entities e.g., objects in the video in an open vocabulary setting. We formulate this problem as open vocabulary multi-label video classification and propose a method to adapt a pre-trained VLM such as CLIP to solve this task. We leverage large language models (LLMs) to provide semantic guidance to the VLM about class labels to improve its open vocabulary performance with two key contributions. First, we propose an end-to-end trainable architecture that learns to prompt an LLM to generate soft attributes for the CLIP text-encoder to enable it to recognize novel classes. Second, we integrate a temporal modeling module into CLIP's vision encoder to effectively model the spatio-temporal dynamics of video concepts as well as propose a novel regularized finetuning technique to ensure strong open vocabulary classification performance in the video domain. Our extensive experimentation showcases the efficacy of our approach on multiple benchmark datasets.", + "bbox": [ + 261, + 359, + 740, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open Vocabulary $\\cdot$ Multi-Modal $\\cdot$ Video Understanding", + "bbox": [ + 261, + 648, + 720, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 688, + 374, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video classification is a critical challenge in computer vision. Its goal involves recognizing a diverse array of concepts depicted in a video, which may include primarily static entities such as objects and scenes, as well as dynamic actions. In the classic setting, the vocabulary of all possible classes of interest is known in advance, and the model is trained in a supervised manner using a labeled dataset. The labor-intensive process of manual annotation often results in video datasets that are narrowly focused, such as those limited to specific sports or simple activities, which restricts the breadth of concepts models can learn. As video", + "bbox": [ + 212, + 718, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa6d656ede06a3b9bd52a63c9f82f23519323f4871da1b926beb4fb11ac58ace.jpg", + "image_caption": [ + "Fig. 1: Our task: recognize multiple open-vocabulary classes in videos at inference from an open vocabulary, including entities (in blue) such as objects and scenes, and actions (in red). Prior zero-shot image classification approaches (e.g. CLIP) can label the salient entity in each frame (left), while zero-shot action-recognition approaches (e.g. ViFi-CLIP) (middle) can classify video-level action. Our method (right) can recognize all action or entity classes present in the video." + ], + "image_footnote": [], + "bbox": [ + 264, + 143, + 743, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "applications are becoming widespread, there is an increasing need for developing video models that can recognize a broad range of concepts. This necessitates the development of open-vocabulary approaches for video classification that can recognize a diverse range of video concepts, including those that were not part of the class vocabulary present in the labeled training dataset. In this work, to solve this problem, we aim to leverage recent advances in vision-language models (VLM), which are trained to align visual and textual data and large language models (LLM), which have been demonstrated to possess a rich understanding of the world due to large scale pre-training.", + "bbox": [ + 212, + 474, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLMs that are trained for image-language alignment on image-text pairs at large-scale exhibit remarkable zero-shot visual recognition performance across a diverse set of tasks [11, 19, 29, 32]. For classification tasks, the pretrained text encoder acts as a \"label encoder\" for the class labels, mapping them into the same representation space as the visual embedding of the image, which allows for the image to be classified by ranking the labels in order of representation similarity. Since VLMs are primarily trained to rank, picking the top match usually results in excellent performance on single label zero-shot classification. However, merely ranking the labels is insufficient to achieve open-vocabulary multi-label classification, and as a result, specialized methods [40] have been developed for multi-label image classification using VLMs. VLM based zero-shot image classification has been further improved by LLM-based prompting, in which LLMs are used to generate descriptive class definitions which are then used as prompts to the VLM text encoder during inference [27, 31]. Other works have adapted VLMs for zero-shot single-label action classification in videos [42, 43]. A typical video,", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "however, contains multiple concepts including objects, actions, scenes, events, etc. Our goal is to develop a video classifier that can simultaneously recognize the presence of these different concepts from any vocabulary specified at inference. We refer to this task as open-vocabulary multi-label video classification. An illustration of the subtle differences between the prior tasks and our proposed task is presented in Fig. 1.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-vocabulary multi-label video classification presents two unique challenges. First, unlike the single label open-vocabulary setting, in the multi-label setting, identifying relevant concepts in a video based on VLM similarity score performs poorly, as VLM similarity scores for different types of concepts (e.g., actions, objects) often fall within different ranges, even with LLM-guided prompting (see Section 4.5). Addressing this issue requires end-to-end finetuning of the VLM with LLM guidance on datasets with diverse video concepts. However, performing end-to-end finetuning while simultaneously incorporating the benefits of LLM-guided prompting, presents additional technical challenges, particularly in backpropagating gradients through the VLMs text encoder's tokenizer. Second, adapting pretrained image-language models to recognize video concepts while retaining strong zero-shot performance is challenging, as the image-language pretraining datasets are orders of magnitude larger than the largest labeled video datasets. VLM adaptation tends to overfit to the video data easily, thus losing their generalization capability [42].", + "bbox": [ + 212, + 238, + 787, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we attempt to address these challenges for open vocabulary multi-label video classification. First, to simultaneously recognize multiple video concepts in open-domain, we finetune the VLM in an end-to-end manner with LLM guidance to ensure proper ranking between different types of video concepts. Particularly, we extend the LLM-guided prompting approach for open-vocabulary classification by introducing learnable prefixes for prompting the LLM. To directly utilize the LLM output representations in the VLM's text encoder for end-to-end training, we propose a prompt transformer. This prompt transformer, in conjunction with the learnable prefixes, not only facilitates the integration of the LLM's world knowledge into the VLM's text representation but also aids in mitigating the discrepancy between VLM scores for various types of concepts. Second, for the effective recognition of temporally varying video concepts, we integrate an auxiliary temporal modeling branch with the VLM's vision encoder [24] and introduce a novel regularization penalty to retain the VLM's zero-shot performance during the adaptation to videos. Through extensive experimentation and analysis we demonstrate that our proposed approach significantly outperforms baseline solutions obtained by extending prior work on single label open vocabulary image and video classification.", + "bbox": [ + 212, + 465, + 787, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, in this work we make three major technical contributions:", + "bbox": [ + 214, + 737, + 725, + 751 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel end-to-end trainable approach to learn a strong label encoder for open vocabulary multi-label video classification by adapting an LLM to learn to prompt a VLM's text encoder.", + "- An approach for enhancing pre-trained VLM image encoders with temporal modeling capability while also retaining strong open-vocabulary performance." + ], + "bbox": [ + 214, + 763, + 782, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Defining a new open vocabulary multi-label video classification benchmark using 5 datasets (2 closed, 3 open vocabulary) and benchmarking our approach against 6 strong baselines, significantly outperforming them.", + "bbox": [ + 215, + 146, + 787, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 215, + 397, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Language Representation Learning: Recently, image-language models [12, 18, 32, 39, 49, 50] have drawn huge attention because of their effectiveness in learning generic visual representations that are transferable to several downstream tasks like classification, retrieval, etc. This success can partly be attributed to the large-scale image-text datasets [4, 35-37, 41]. However, this is not the case for the video data. Therefore, to perform video-language pretraining, most recent works [6, 7, 26, 30, 47] bootstrap from a pretrained image-language model and adapt them to video representation tasks.", + "bbox": [ + 212, + 250, + 787, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Vocabulary Classification: Vision-language models do not perform optimally in visual recognition tasks out of the box. To improve the zero-shot open-vocabulary visual recognition capability of the vision-language models, the recent works either resort to regularized finetuning [43, 45] or prompt learning [40, 42, 52]. However, these works either focus on solving single label tasks or only perform adaptation for image recognition tasks. In contrast to these works, we focus on open-vocabulary video classification task.", + "bbox": [ + 212, + 371, + 787, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large Language Models in Vision: Recently, Large Language Models (LLM) have been utilized in vision tasks primarily because of their excellent in-context and zero-shot learning performance especially in commonsense reasoning tasks [15]. To be particular, LLMs have been used to rewrite noisy ASR text [8,38], to generate classification labels by mapping textual description to a predefined task list [22], extracting verbs from textual descriptions to improve action understanding [21]. LLMs have also been used to initialize textual encoders for vision language encoder [51]. In some recent works, vision representations have been aligned with LLM input space to solve multitude of vision-language tasks [17,23]. These multi-modal models have also been used to generate captions for images which are then utilized to train better vision-language models [5].", + "bbox": [ + 212, + 477, + 795, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompting Vision-Language Models with LLMs: Some prior works have leveraged LLMs to augment the open vocabulary understanding capabilities of Vision-Language models. Classification By Description [27], CuPL [31], LLM guided concept bottleneck [48] all utilize LLM generated class descriptors to prompt CLIP to improve image classification, and make it more interpretable. WaffleCLIP [34] demonstrated that LLM generated higher level concepts to describe datasets can also improve classification. LLM based approaches have also been extended to object detection [13] and point cloud understanding [53]", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameter efficient fine-tuning of LLMs: As full parameter finetuning of LLMs is not practical, quite a few parameter efficient finetuning methods have been developed. The lightweight Learnable Soft prompts [16] is the approach we utilize to guide the LLM in our method. Other approaches such as LoRA [10], Prefix-Tuning [20] and P-Tuning [25] could also be adopted in principle.", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposal to boost the open vocabulary multi-label video classification capabilities of CLIP consists of two key parts: first, an end-to-end trainable label", + "bbox": [ + 212, + 204, + 782, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encoder leveraging an LLM for strong open vocabulary capabilities, and the CLIP text encoder for visual alignment. The label encoder consists of a frozen LLM guided with learnable prefixes, whose outputs are mapped to the frozen CLIP text encoder using a learnable prompting transformer. Second, we enhance the CLIP image encoder's capabilities for understanding temporal dynamics. This is achieved by using a lightweight temporal modeling branch to enhance the CLIP image encoder. The details of our approach are illustrated in Figure 3. Each part of our method is discussed in detail in this section.", + "bbox": [ + 212, + 239, + 460, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of our approach is provided in Fig. 2. Our method has three broad stages. Firstly, during the training phase, both the label encoder and video encoder are trained simultaneously, as shown in (a).", + "bbox": [ + 212, + 532, + 459, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secondly, during the classifier vocabulary expansion stage, embeddings for class labels are calculated and saved into a label embedding database, as shown in (b). This vocabulary can be extended at any point after training, thus allowing our method to be used in the open vocabulary setting. Finally, during the inference", + "bbox": [ + 212, + 643, + 460, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "stage, video features are computed and compared against label embeddings from the database. As the label embeddings are pre-computed, the computational overhead during inference over standard CLIP models is minimal.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37fd6f002647e864d9f3e79b261b57ffc0482b5da92c96721421e1e836e18748.jpg", + "image_caption": [ + "(a) Training Stage" + ], + "image_footnote": [], + "bbox": [ + 472, + 273, + 781, + 391 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ef2898a2fb44cfde50cab01708b4430ea91f11e42e7457ca4dadaa2a06612c4.jpg", + "image_caption": [ + "(b) Classifier Vocabulary Expansion Stage" + ], + "image_footnote": [], + "bbox": [ + 472, + 409, + 777, + 484 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/329d6aad43e3bd6950d5b340653c41a44f31103ba06fac1de88a87961f4b2783.jpg", + "image_caption": [ + "Fig.2: Our open vocabulary classification method includes three stages of operation. During the (a) training stage, we train our label and video encoders on closed set training labels. New class labels can be added to the vocabulary after training by employing the (b) classifier vocabulary expansion stage. During the (c) inference stage video embeddings are computed and matched with the label embeddings database to get the classification scores." + ], + "image_footnote": [], + "bbox": [ + 483, + 496, + 772, + 588 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 LLMs for Semantic Enrichment of Label Embeddings", + "text_level": 1, + "bbox": [ + 215, + 146, + 699, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretrained VLMs have strong open-vocabulary image classification performance owing to their large scale image-text pretraining. LLMs on the other hand are trained on even larger scale of text data with training objectives that enable a rich semantic understanding. As a result, compared to LLMs, VLMs have a more limited semantic understanding of natural language and have difficulty understanding concepts such as the relationship between different class labels. For instance, prior works have found that while VLMs are able to classify ImageNet images at a fine-grained level, their accuracy at higher levels of the class hierarchy is poor [46]. E.g., while they may recognize an image of a lion, they are unable to understand that a lion is also a feline, a mammal and an animal. However, an LLM is very effective in comprehending the hierarchical relationship between these labels, among many other capabilities, due to the large-scale training. Hence, in this work, we utilize an LLM to generate complementary information about the class labels that can be utilized to improve the vision-language alignment of VLM leading to better open-vocabulary classification performance.", + "bbox": [ + 212, + 174, + 785, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fixed LLM prompting.", + "text_level": 1, + "bbox": [ + 215, + 401, + 410, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a baseline, we first develop an LLM-based prompting method to adapt CLIP for open vocabulary video classification, extending the approach of Menon & Vondrick [27] for open vocabulary image classification. We design a prompt template (see implementation details in Section K of Supplementary material) for an encoder-decoder based LLM with the class name and a question asking it to generate useful features for visually distinguishing that class. The LLM output is then parsed into a list of textual descriptions, hereafter referred to as attributes. These attributes along with the class name is used to prompt the CLIP text encoder and the resultant text-embeddings are mean-pooled to obtain an attribute enriched text-embedding for the class. We mean-pool CLIP vision embedding of video frames to generate video embeddings and then perform open vocabulary classification by matching video embeddings with the text-embeddings of different classes. However, this simple baseline has a crucial weakness: it is not trainable, as there is a need to de-tokenize and tokenize the LLM text output to CLIP, which doesn't allow the flow of gradients. As a result, the process of generating class attributes by prompting an LLM cannot be improved by training on a labeled video dataset. To remedy this, we propose an end-to-end trainable architecture that integrates the LLM with the CLIP text encoder.", + "bbox": [ + 212, + 417, + 785, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "End-to-end learnable LLM prompting. Our proposed architecture, summarized in Figure 3, incorporates a learnable prompting framework with the frozen LLM to generate the inputs to a frozen CLIP text-encoder. The learnable components on the text side are limited to $N$ learnable prefixes/vectors to the prompt template used for querying the LLM and a prompt transformer that transforms the sequence of tokens from the LLM to input soft prompts for the CLIP text-encoder. In our implementation we avoid the discrete and nondifferentiable operations like detokenization at the LLM decoder output and re-tokenization at the CLIP text-encoder input. This allows the prompt transformer to directly connect the LLM output semantic space to the CLIP input", + "bbox": [ + 212, + 689, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/608e33bb894c29f04f60afc3af13fe110c4bb80a9a29785624ce708da6e654c3.jpg", + "image_caption": [ + "Fig. 3: Our end-to-end trainable system for open vocabulary video classification. The class labels are used by the LLM to generate useful class attributes for the CLIP text encoder which provides a visually aligned label embedding. The learnable input prompts to the LLM guide it to generate soft-attributes useful for video classification. Prompting transformer learns to map from the LLM output space to the CLIP input space. To add video understanding to CLIP's vision encoder we add additional spatiotemporal modeling layers. Details about each component are in Section 3." + ], + "image_footnote": [], + "bbox": [ + 277, + 150, + 733, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "semantic space, ensuring that the entire text model is differentiable and therefore end-to-end learnable. We describe the details below.", + "bbox": [ + 212, + 641, + 787, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each of the $N$ LLM prefixes, we first construct an input sequence to the LLM comprising of the prefix followed by a fixed prompt template that contains the label name $\\ell$ from the input data sample (see Figure 3). The text input to the LLM is mapped using a tokenizer and the LLM's embedding layer into a sequence of tokens $I\\in \\mathbb{R}^{M\\times d}$ , where $M$ and $d$ represent number of input tokens and the dimension of the embedding space respectively. All our $N$ learnable prefixes are $d$ -dimensional vectors. We use $P_{i}\\in \\mathbb{R}^{d}$ to represent the $i$ -th prefix. Each prefix is concatenated with the tokens of the prompt template, yielding a unified sequence of tokens $[P_i;I]\\in \\mathbb{R}^{(1 + M)\\times d}$ . This combined sequence is then processed through the frozen encoder-decoder layers of the LLM. For each prefix, we run $KL$ decoding iterations of the LLM decoder to generate a sequence of", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$KL$ decoded tokens. These tokens represent useful class features in the LLM's semantic space. As the prompt template specifically asks LLM to output features as a list, we split the sequence of tokens evenly into $K$ subsequences of $L$ tokens each. Due to the nature of LLM prompt template, which prompts the LLM to output useful visual features, we refer to the subsequences of tokens as soft attributes, as they are sets of continuous vectors instead of discrete attributes in natural language. The $K$ soft attributes are then individually processed by the prompt transformer to generate $K$ soft prompts to the CLIP text encoder. Repeating this operation for each of the $N$ LLM prefixes, we get $NK$ soft prompts. Each soft-prompt is then concatenated with the tokenized label embedding and then processed by the frozen CLIP text-encoder to generate an attribute feature. All $NK$ resulting features are mean-pooled and normalized to obtain the final label embedding $f_{t}(\\ell)$ , where $\\ell$ is the label name.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Regularized Parallel Temporal Modeling", + "text_level": 1, + "bbox": [ + 214, + 363, + 598, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The simple image transformer of the CLIP vision encoder does not model the temporal dynamics of the video, which is essential to enhance the ability to recognize entities in videos. We enhance our vision encoder by adding a parallel temporal modeling branch to the last $T$ layers of the CLIP vision encoder as illustrated in Figure 3. We freeze the CLIP vision branch and train only the newly added temporal layers. Each block in the newly added branch consists of a spatial attention layer initialized from their corresponding CLIP weights, and a temporal attention layer that is randomly initialized. All frames of the video are processed independently by the CLIP vision backbone. Then, at the first temporal modeling block, the temporal token TMP is created by averaging the CLS tokens across frames, meanwhile learnable spatial and temporal positional embeddings are added to each patch token. The $t$ th temporal modeling block takes in as input a", + "bbox": [ + 212, + 386, + 488, + 779 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/069d3ff05a91fc9e23e28564d2b1bfc60ef56395b8b489f0990ad159afb84780.jpg", + "image_caption": [ + "Fig. 4: Our Temporal Block projects frame patch tokens from the CLIP image encoder, fusing them with previous block's temporal branch tokens. The temporal token (TMP) incorporates all frames' CLS tokens. Divided Space-Time attention layers form the core of the block. Spatial attention layers are initialized from CLIP weights and regularized using stochastic weight averaging." + ], + "image_footnote": [], + "bbox": [ + 504, + 417, + 754, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "weighted combination of the previous temporal modeling block and the corresponding layer in the CLIP backbone. Symbolically if $V_{t}^{\\top}$ represents the patch tokens from the temporal block at the $t$ th block and $V_{t}$ represents the patch tokens for the corresponding CLIP layers, the patch tokens for the $t$ th block", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$V_{t}^{\\mathsf{T}} = V_{t - 1}^{\\mathsf{T}} + Proj_{\\text{spatial}}(V_{s})$ , where $V_{s}$ is the corresponding token from the CLIP backbone. After that the tokens are processed by the divided space-time attention layer and spatial layer. The overall video embedding is generated at the end by mean pooling the final TMP token and the CLS tokens for each frame from the CLIP backbone.", + "bbox": [ + 212, + 145, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A key drawback of adding a deep parallel branch is that while it improves performance in the finetuned setting, it diminishes CLIP's zero-shot performance. In order to overcome this limitation in the zero-shot setting, we propose to use weight regularization on the spatial attention layers. The weight regularization operation tries to match predictions made by the current set of finetuned weights and a randomly weighted average of the current weights and the original CLIP weights. Symbolically, at each iteration we use weights $\\theta$ which are a stochastic weighted average of the finetuned and the original frozen weights as given by:", + "bbox": [ + 212, + 222, + 784, + 343 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\alpha \\theta_ {\\mathrm {f t}} + (1 - \\alpha) \\theta_ {\\mathrm {f r o z e n}}, \\quad \\mathrm {w h e r e} \\quad \\alpha \\sim U (0, \\lambda)\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 359, + 671, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We observe that an empirical value of $\\lambda = 0.5$ works well across datasets.", + "bbox": [ + 214, + 382, + 741, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Relationship to prior works: Prior temporal modeling approaches split into two camps: improving finetuned performance or preserving CLIP's zero-shot abilities. STAN [24] is a state-of-the-art finetuning focused approach which adds a parallel temporal modeling branch which is able to leverage both high-level and low-level features from the CLIP vision encoder. The divided space-time attention in STAN's temporal block is inspired by the TimeSformer [2] architecture. We differ from STAN in order to achieve stronger open vocabulary performance in a number of ways. Firstly, unlike STAN we do not finetune the CLIP image encoder, only the added parallel branch. Additionally, we propose stochastic weight regularization to prevent the temporal branch from drifting too far from the original CLIP feature space. Our ablation experiments (Table 3) demonstrate that both these choices significantly improve our performance. Some form of weight regularization has been explored by methods such as Open-VCLIP [44]. However, Open V-CLIP takes in only 3 frames at a time, and its temporal modeling range is limited. Our approach is able to achieve the best of both approaches.", + "bbox": [ + 212, + 412, + 787, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Training objective", + "text_level": 1, + "bbox": [ + 214, + 662, + 413, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train our model on multi-label video datasets. To construct a batch, we first sample a set $\\mathcal{B}$ of $B$ videos from the dataset. Each video $v$ is associated with a set of positive class labels which we denote $\\mathcal{P}(v)$ . In addition to the positive labels, we augment the batch with random negative class labels for each video. To obtain the negatives, we first identify the set $\\mathcal{P}_{\\mathcal{B}} := \\cup_{v \\in \\mathcal{B}} \\mathcal{P}(v)$ of distinct classes among all the positive labels from all the videos in the batch. We then sample a random set $\\mathcal{N}_{\\mathcal{B}}$ of $4B - |\\mathcal{P}_{\\mathcal{B}}|$ classes from the rest of the class vocabulary of the dataset. For each video $v$ , we then choose all non-positive classes from $\\mathcal{P}_{\\mathcal{B}} \\cup \\mathcal{N}_{\\mathcal{B}}$ as negative labels. We use $\\mathcal{N}(v) := (\\mathcal{P}_{\\mathcal{B}} \\cup \\mathcal{N}_{\\mathcal{B}}) \\setminus \\mathcal{P}(v)$ to denote the set of all negative labels for $v$ . The resulting positive and negative video-label pairs are", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "then treated as training samples for binary classification. Each training sample is a label, video pair $(\\ell, v)$ where $\\ell$ represents the label name and $v$ the video. With this batch construction, the total number of training samples in a batch could be variable, but the total number of videos is fixed at $B$ and the total number of classes present among the samples is fixed at $4B$ .", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For each training data sample $(\\ell, v)$ , the text and video encoders respectively generate a unit-norm text embedding $f_{t}(\\ell) \\in \\mathbb{R}^{D}$ and a unit-norm video embedding $f_{v}(v) \\in \\mathbb{R}^{D}$ where $D$ is the common embedding dimension. The score $s(\\ell, v)$ for this data sample is given by the inner product:", + "bbox": [ + 212, + 222, + 517, + 344 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\ns (\\ell , v) = \\left(f _ {t} (\\ell)\\right) ^ {\\mathsf {T}} f _ {v} (v). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 353, + 514, + 371 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The model is trained with a weighted binary cross entropy loss:", + "bbox": [ + 212, + 381, + 514, + 411 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9da06156b01cb9b87c553dd63e811e03d91b0fa9ea419ee607798789c29d7da3.jpg", + "image_caption": [ + "Fig.5: Our training objective applies binary cross-entropy to predicted Video-Label feature similarities sharpened by a temperature-scaled sigmoid." + ], + "image_footnote": [], + "bbox": [ + 526, + 253, + 769, + 367 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathcal {B}) = - \\sum_ {v \\in \\mathcal {B}} \\left[ \\sum_ {\\ell \\in \\mathcal {P} (v)} \\log p (\\ell , v) + w \\sum_ {\\ell \\in \\mathcal {N} (v)} \\log \\left(1 - p (\\ell , v)\\right) \\right] ^ {\\text {e n e d b y a t}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 421, + 602, + 486 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $p(\\ell, v) \\coloneqq \\sigma\\left(\\frac{s(\\ell, v)}{\\tau}\\right)$ , $\\tau$ is a temperature parameter, $\\sigma(.)$ is the sigmoid function and $w > 0$ is a weight hyperparameter.", + "bbox": [ + 212, + 498, + 785, + 535 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments and Results", + "text_level": 1, + "bbox": [ + 214, + 556, + 496, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 588, + 334, + 602 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In order to train our model on a wide range of concepts, we train it on a mix of YouTube8M [1] and Kinetics-400 [14]. YouTube-8M (YT-8M) is primarily labeled with entities, whereas Kinetics-400 (K400) is labeled only with actions. As YT-8M is composed of a random sample of the whole of YouTube, a significant portion of video its samples are from video game streams. Many of these videos are just tagged with the video game title as the label irrespective of the actual entities and actions present in the video. A significant portion of its label vocabulary is dedicated to video game titles (771 out of 3862 total) are dedicated to video game titles. We observe training instabilities due to this, and to fix this issue, we remove video game titles from the training vocabulary. We also remove some of the least frequently occurring labels during training to reach a training vocabulary of 2429 classes. This leads to stabilized training without further issues. Note that for closed set evaluation on YT-8M, we use the YouTube-8M Segments validation set, which contains human verified labels for 1000 classes. K400 on the other hand is widely used, and has a clean vocabulary and high", + "bbox": [ + 212, + 613, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "quality human verified action labels. From the training perspective, the only limitation is the absence of entity (object, scenes etc.) labels.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For evaluation, we use 3 open-vocabulary test datasets: TAO (Tracking Any Object) [3], ActivityNet [9] and RareAct [28]. TAO and ActivityNet are exclusively object and action focused datasets respectively. RareAct has object, action labels, and focuses on unusual object-action combination. TAO dataset was developed for evaluating object trackers, however analogous to how object detection datasets like MS-COCO are used for evaluating multi-label image classification methods such as DualCoOp [40], we transform TAO into a multi-label video classification dataset by ignoring localization annotations. ActivityNet often has multiple actions in a given video. For our evaluation each annotation from RareAct provides three labels: object, action and the (unusual) object-action combination, this provides an interesting test of the open vocabulary capabilities. Altogether, these five datasets provide a comprehensive evaluation of a model's closed and open vocabulary video classification performance.", + "bbox": [ + 212, + 176, + 787, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2Baselines", + "text_level": 1, + "bbox": [ + 215, + 397, + 339, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As there are no prior works that explicitly address multilabel open vocabulary video classification, we extend the following methods from the zero shot image and action classification literature to our multi-label video classification setting for comparison.", + "bbox": [ + 212, + 424, + 782, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CoOp [52]: Short for Context Optimization, CoOp learns prompts for the CLIP text encoder as a lightweight adaptation technique for classification.", + "bbox": [ + 212, + 484, + 782, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "DualCoOp [40]: Is an extension of CoOP to the multi-label setting where both positive and negative prompts are learnt for the CLIP text encoder. For a given label, prediction is based on whether the positive or negative prompted version has higher similarity with the image feature.", + "bbox": [ + 212, + 515, + 782, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "LLM + CLIP (Frozen): This baseline was discussed in Section 3.1 and is illustrated in Figure 8 of Supplementary.", + "bbox": [ + 212, + 575, + 782, + 603 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ViFi-CLIP [33]: In this baseline, there is no temporal modeling, and both the CLIP image and text encoder are finetuned on the training dataset.", + "bbox": [ + 212, + 607, + 782, + 633 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As CoOp and ViFi-CLIP were developed for single label classification, to utilize them in our setting we replace their contrastive loss function with our multi-label classification loss. We also drop the region aggregation aspect of DualCoOp and only test the dual prompting architecture. We refer to these baselines as CoOp*, DualCoOp* and ViFi-CLIP* to reflect these differences.", + "bbox": [ + 212, + 638, + 782, + 710 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Metrics", + "text_level": 1, + "bbox": [ + 215, + 737, + 326, + 750 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We report two sets of metrics to evaluate the performance of the models in the open vocabulary setting. Area Under Precision-Recall curve (AUPR) summarizes the overall classification performance across the entire precision-recall trade-off. Secondly, we report the Peak F1-Score for the model on each dataset, which is the F1-Score achieved by that model at the optimal threshold chosen by an oracle", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b950e4f5d8abb9678f6387e26166fb93acf10b1ba098ee931073dcfe488a8a4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities+Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt6.326.243.844.29.5
CLIP with Prompt Templates6.830.546.045.911.4
CLIP with Fixed LLM Prompts6.930.650.246.811.5
Trainable Baseline Methods
CoOp2.717.835.028.83.5
DualCoOp8.323.947.133.07.6
ViFi-CLIP3.410.958.317.24.1
Ours
CLIP + Learnable LLM Prompts9.432.851.447.111.9
+ Temporal Modeling14.842.063.847.112.4
+ Synthetic Labels16.743.265.550.213.2
", + "bbox": [ + 228, + 143, + 776, + 329 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1: AUPR scores for all methods on all datasets.", + "bbox": [ + 316, + 329, + 683, + 340 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "for that dataset. This metric captures the classification performance obtained by an open vocabulary classification method on a dataset, if the threshold alone could be tuned, e.g., by using a labeled validation set.", + "bbox": [ + 212, + 386, + 784, + 433 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Results", + "text_level": 1, + "bbox": [ + 215, + 455, + 325, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "AUPR scores for each method are presented in Table 1 and Peak F1 scores are presented in Table 2. Our learnable LLM prompting approach outperforms both frozen and trainable baselines across both closed-vocabulary and open-vocabulary classification tasks. Also notable is the enhancement in performance due to our temporal modeling approach (average gain of $\\approx 5\\%$ in AUPR). As previously mentioned in our discussion of the dataset, the training datasets YT-8M and K400 have certain limitations (in particular, YT-8M has few action labels and K400 has no object/entity labels). We present further improved results obtained by augmenting our training dataset with additional class labels obtained using a multimodal LLM (pipeline is detailed in Section I of the Supplementary Material). As seen from the last rows of Tables 1 and 2, this approach gives a further $\\approx 2\\%$ improvement, and is specially effective for actions.", + "bbox": [ + 212, + 484, + 787, + 667 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Improved score calibration across datasets", + "text_level": 1, + "bbox": [ + 215, + 691, + 609, + 707 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In order to use a classifier in a truly open vocabulary setting in practice, the classification score needs to be well calibrated across different types of concepts. This would allow us to pre-select an optimal threshold to generate binary classification for all concepts. In Figure 6 we demonstrate the robustness offered by our method in setting a threshold that works for a wide range of concepts, and the advantage offered by our method over the frozen CLIP + LLM baseline. These figures plot the F1 scores for multi-label classification across thresholds provided by both methods with Figure 6a showing the numbers on the validation splits", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a20171eded079ba94b2510e1958b0f166b4c3a4806de8ec6248c66ebbae71a88.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities + Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt14.934.244.647.117.6
CLIP with Prompt Templates20.836.849.250.120.5
CLIP with Fixed LLM Prompts21.637.350.251.419.8
Trainable Baseline Methods
CoOp5.825.543.935.510.5
DualCoOp16.233.249.040.515.0
ViFi-CLIP5.219.354.424.79.6
Ours
CLIP + Learnable LLM Prompts23.642.452.851.122.6
+ Temporal Modeling31.546.259.652.624.3
+ Synthetic Labels32.746.656.653.825.1
", + "bbox": [ + 223, + 143, + 779, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2: Peak F1 scores for all methods on all datasets.", + "bbox": [ + 310, + 329, + 689, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dafc88c6d70e2d588a4642d629d68729f841aad4b170287f80a1391bc5f6fb25.jpg", + "image_caption": [ + "(a) Closed Vocabulary Evaluation." + ], + "image_footnote": [], + "bbox": [ + 223, + 376, + 464, + 476 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dd25dc6544a7dc4be501e6c89cee168e4206348622993b4b60c00812bbc95075.jpg", + "image_caption": [ + "(b) Open Vocabulary Evaluation.", + "Fig. 6: F1-Scores at different thresholds for closed and open vocabulary evaluation datasets. Our end-to-end trained model can achieve better performance across datasets with a single threshold (labeled by gray vertical line) chosen on the supervised datasets." + ], + "image_footnote": [], + "bbox": [ + 535, + 376, + 777, + 476 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of the training datasets, and Figure 6b showing the numbers in the open vocabulary setting on two diverse datasets. A reasonable choice for a classification threshold would be to pick the threshold that maximizes the minimum validation F1-score for the datasets used in training. The resulting choice of thresholds for our method and for frozen CLIP are shown in Figure 6a. We see that for the baseline method the performance on TAO is poor, while that on ActivityNet is modest. However with our proposed method the F1 scores for both evaluation datasets are simultaneously high. This result empirically validates the versatility that is achieved by our learnable LLM-based prompting scheme over the baseline, which makes it possible to pre-select a threshold for using the solution as a black box classifier for all unseen concepts.", + "bbox": [ + 212, + 580, + 787, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Ablations", + "text_level": 1, + "bbox": [ + 215, + 777, + 341, + 791 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We scientifically ablate each part of our framework, reporting ablation evaluations (AUPR) on TAO (Objects) and ActivityNet (Actions).", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/17fcecd83deb8830b4465785bbed4729c4335d63feca9b8051da43ed2e42d1d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Open Vocabulary
# BlocksObjectsActions
565.448.1
465.550.2
264.649.1
162.347.7
", + "bbox": [ + 222, + 143, + 410, + 227 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/435dfd0060c91a52ca54709f4591c1bfa1285e3f1d779e444f9e2527412fdb50.jpg", + "table_caption": [ + "(a) Effect of Temporal Modeling" + ], + "table_footnote": [], + "table_body": "
Open Vocabulary
Spatial Reg.BackboneObjectsActionsGeo. Mean
✓56.653.855.2
X62.132.544.9
X42.728.334.8
", + "bbox": [ + 439, + 143, + 779, + 229 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/25b93de6d33b88c17175ba0b35af8af51e0c5effdb04e8e371b8c0ca9c2f8abd.jpg", + "table_caption": [ + "(b) Effect of our Regularization", + "Table 3: Ablations of Temporal Modeling Branch Architecture and Regularization" + ], + "table_footnote": [], + "table_body": "
Label EncoderOpen Vocabulary
LLMPTCLIPObjects Actions
Learnable Prompt✓✗25.2 8.5
Fixed Prompt✗✓62.5 41.3
Fixed Prompt✓✓64.9 48.1
Learnable Prompt✓✓65.5 50.2
✗✗✓60.1 34.5
", + "bbox": [ + 352, + 295, + 648, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 4: Ablating the Label Encoder. PT $\\rightarrow$ Prompt Transformer", + "bbox": [ + 277, + 388, + 723, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "- Temporal Modeling Blocks: We find that using more Blocks improves performance, until around 4. (Table 3a)", + "bbox": [ + 223, + 445, + 784, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "- Temporal Modeling Weight Regularization: We find that using our weight regularization strategy for the spatial layers of the temporal branch prevents the model from overfitting to certain concepts. We also validate our choice of keeping the CLIP vision backbone frozen and find that unfreezing the backbone leads to severe over-fitting. (Table 3b)", + "bbox": [ + 225, + 474, + 785, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "- Label Encoder: We demonstrate the effectiveness of combining an LLM with learnable prompts and CLIP's text encoder as the label encoder. First, we show that removing CLIP text encoder leads to a significant drop in performance. This demonstrates that the visual-text alignment learned by CLIP is essential. Next we demonstrate the benefits of learnable prompts over fixed prompts. We also provide results with the LLM removed (Table 4).", + "bbox": [ + 225, + 550, + 785, + 641 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 223, + 660, + 359, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced the problem of open vocabulary multi-label video classification and proposed a solution leveraging LLMs and pre-trained VLMs. Through our extensive experiments we demonstrated strong performance on both actions and entities with a single model. Our proposed approach benefits from two key innovations. First, we proposed a method to prompt VLMs more effectively by utilizing the LLM output representations through a trainable prompt transformer and learnable LLM prompts. Second, we introduced a temporal modeling architecture and a regularized finetuning approach to improve the video understanding capability of the vision encoder while retaining strong open-vocabulary performance. Our ablations validate the efficacy of each of these contributions.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark (2016)", + "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: Proceedings of the International Conference on Machine Learning (ICML) (July 2021)", + "3. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 436-454. Springer (2020)", + "4. Desai, K., Kaul, G., Aysola, Z., Johnson, J.: Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431 (2021)", + "5. Fan, L., Krishnan, D., Isola, P., Katabi, D., Tian, Y.: Improving clip training with language rewrites. In: NeurIPS (2023)", + "6. Fang, H., Xiong, P., Xu, L., Chen, Y.: Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097 (2021)", + "7. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)", + "8. Gupta, R., Roy, A., Christensen, C., Kim, S., Gerard, S., Cincebeaux, M., Divakaran, A., Grindal, T., Shah, M.: Class prototypes based contrastive learning for classifying multi-label and fine-grained educational videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19923-19933 (June 2023)", + "9. Heilbron, F.C., Niebles, J.C.: Collecting and annotating human activities in web videos. In: Proceedings of International Conference on Multimedia Retrieval. p. 377-384. ICMR '14, Association for Computing Machinery, New York, NY, USA (2014). https://doi.org/10.1145/2578726.2578775, https://doi.org/10.1145/2578726.2578775", + "0. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9", + "1. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html", + "2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "3. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. In: International Conference on Machine Learning (2023)", + "4. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., Suleyman, M., Zisserman, A.: The kinetics human action video dataset (2017)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=e2TBb5y0yFf", + "16. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243", + "17. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 12888-12900. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/1i22n.html", + "18. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)", + "19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10965-10975 (June 2022)", + "20. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353", + "21. Lin, W., Karlinsky, L., Shvetsova, N., Possegger, H., Kozinski, M., Panda, R., Feris, R., Kuehne, H., Bischof, H.: Match, expand and improve: Unsupervised finetuning for zero-shot action recognition with language knowledge. In: ICCV (2023)", + "22. Lin, X., Petroni, F., Bertasius, G., Rohrbach, M., Chang, S.F., Torresani, L.: Learning to recognize procedural activities with distant supervision. arXiv preprint arXiv:2201.10990 (2022)", + "23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems. Curran Associates, Inc. (2023)", + "24. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023)", + "25. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. AI Open (2023). https://doi.org/https://doi.org/10.1016/j.aiopen.2023.08.012, https://www.sciencedirect.com/science/article/pii/S2666651023000141", + "26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Menon, S., Vondrick, C.: Visual classification via description from large language models. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=j1AjNL8z5cs", + "28. Miech, A., Alayrac, J.B., Laptev, I., Sivic, J., Zisserman, A.: Rareact: A video dataset of unusual interactions. arxiv:2008.01018 (2020)", + "29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 728-755. Springer Nature Switzerland, Cham (2022)", + "30. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)", + "31. Pratt, S., Covert, I., Liu, R., Farhadi, A.: What does a platypus look like? generating customized prompts for zero-shot image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15691-15701 (2023)", + "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "33. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "34. Roth, K., Kim, J.M., Koepke, A.S., Vinyals, O., Schmid, C., Akata, Z.: Waffling around for performance: Visual classification with random words and broad concepts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15746-15757 (October 2023)", + "35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)", + "38. Shvetsova, N., Kukleva, A., Hong, X., Rupprecht, C., Schiele, B., Kuehne, H.: HowtocapTION: Prompting llms to transform video annotations at scale (2023)", + "39. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022)", + "40. Sun, X., Hu, P., Saenko, K.: Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems 35, 30569-30582 (2022)", + "41. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Commun. ACM" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Open Vocabulary Multi-Label Video Classification", + "bbox": [ + 393, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "59(2), 64-73 (jan 2016). https://doi.org/10.1145/2812802, https://doi.org/10.1145/2812802", + "42. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23034–23044 (2023)", + "43. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-VCLIP: Transforming CLIP to an open-vocabulary video model via interpolated weight optimization. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 36978-36989. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/weng23b.htm1", + "44. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In: ICML (2023)", + "45. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022)", + "46. Xu, Z., Zhu, Y., Deng, T., Mittal, A., Chen, Y., Wang, M., Favaro, P., Tighe, J., Modolo, D.: Challenges of zero-shot recognition with vision-language models: Granularity and correctness (2023)", + "47. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022)", + "48. Yang, Y., Panagopoulou, A., Zhou, S., Jin, D., Callison-Burch, C., Yatskar, M.: Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19187-19197 (June 2023)", + "49. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)", + "50. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)", + "51. Zhao, Y., Misra, I., Krahenbuhl, P., Girdhar, R.: Learning video representations from large language models. In: CVPR (2023)", + "52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2639-2650 (October 2023)" + ], + "bbox": [ + 217, + 147, + 785, + 728 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Gupta et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_model.json b/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ffa71d9cd75ac6e1a2d3ef372f7e6ca400a6503a --- /dev/null +++ b/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_model.json @@ -0,0 +1,2326 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.293, + 0.141, + 0.71, + 0.184 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.212, + 0.763, + 0.227 + ], + "angle": 0, + "content": "Rohit Gupta1, Mamshad Nayeem Rizve2, Jayakrishnan Unnikrishnan2," + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.228, + 0.668, + 0.243 + ], + "angle": 0, + "content": "Ashish Tawari\\(^{2}\\), Son Tran\\(^{2}\\), Mubarak Shah\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.244, + 0.64, + 0.258 + ], + "angle": 0, + "content": "Benjamin Yao\\(^{2}\\), and Trishul Chilimbi\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.269, + 0.744, + 0.284 + ], + "angle": 0, + "content": "1 Center for Research in Computer Vision, University of Central Florida" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.286, + 0.649, + 0.298 + ], + "angle": 0, + "content": "rohit.gupta@ucf.edu, shah@crv.ucf.edu" + }, + { + "type": "text", + "bbox": [ + 0.466, + 0.299, + 0.537, + 0.311 + ], + "angle": 0, + "content": "2 Amazon" + }, + { + "type": "text", + "bbox": [ + 0.367, + 0.313, + 0.637, + 0.326 + ], + "angle": 0, + "content": "{jayunn,atakari,sontran}@amazon.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.36, + 0.741, + 0.637 + ], + "angle": 0, + "content": "Abstract. Pre-trained vision-language models (VLMs) have enabled significant progress in open vocabulary computer vision tasks such as image classification, object detection and image segmentation. Some recent works have focused on extending VLMs to open vocabulary single label action classification in videos. However, previous methods fall short in holistic video understanding which requires the ability to simultaneously recognize multiple actions and entities e.g., objects in the video in an open vocabulary setting. We formulate this problem as open vocabulary multi-label video classification and propose a method to adapt a pre-trained VLM such as CLIP to solve this task. We leverage large language models (LLMs) to provide semantic guidance to the VLM about class labels to improve its open vocabulary performance with two key contributions. First, we propose an end-to-end trainable architecture that learns to prompt an LLM to generate soft attributes for the CLIP text-encoder to enable it to recognize novel classes. Second, we integrate a temporal modeling module into CLIP's vision encoder to effectively model the spatio-temporal dynamics of video concepts as well as propose a novel regularized finetuning technique to ensure strong open vocabulary classification performance in the video domain. Our extensive experimentation showcases the efficacy of our approach on multiple benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.65, + 0.721, + 0.665 + ], + "angle": 0, + "content": "Keywords: Open Vocabulary \\(\\cdot\\) Multi-Modal \\(\\cdot\\) Video Understanding" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.689, + 0.375, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Video classification is a critical challenge in computer vision. Its goal involves recognizing a diverse array of concepts depicted in a video, which may include primarily static entities such as objects and scenes, as well as dynamic actions. In the classic setting, the vocabulary of all possible classes of interest is known in advance, and the model is trained in a supervised manner using a labeled dataset. The labor-intensive process of manual annotation often results in video datasets that are narrowly focused, such as those limited to specific sports or simple activities, which restricts the breadth of concepts models can learn. As video" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.144, + 0.744, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Fig. 1: Our task: recognize multiple open-vocabulary classes in videos at inference from an open vocabulary, including entities (in blue) such as objects and scenes, and actions (in red). Prior zero-shot image classification approaches (e.g. CLIP) can label the salient entity in each frame (left), while zero-shot action-recognition approaches (e.g. ViFi-CLIP) (middle) can classify video-level action. Our method (right) can recognize all action or entity classes present in the video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.612 + ], + "angle": 0, + "content": "applications are becoming widespread, there is an increasing need for developing video models that can recognize a broad range of concepts. This necessitates the development of open-vocabulary approaches for video classification that can recognize a diverse range of video concepts, including those that were not part of the class vocabulary present in the labeled training dataset. In this work, to solve this problem, we aim to leverage recent advances in vision-language models (VLM), which are trained to align visual and textual data and large language models (LLM), which have been demonstrated to possess a rich understanding of the world due to large scale pre-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "VLMs that are trained for image-language alignment on image-text pairs at large-scale exhibit remarkable zero-shot visual recognition performance across a diverse set of tasks [11, 19, 29, 32]. For classification tasks, the pretrained text encoder acts as a \"label encoder\" for the class labels, mapping them into the same representation space as the visual embedding of the image, which allows for the image to be classified by ranking the labels in order of representation similarity. Since VLMs are primarily trained to rank, picking the top match usually results in excellent performance on single label zero-shot classification. However, merely ranking the labels is insufficient to achieve open-vocabulary multi-label classification, and as a result, specialized methods [40] have been developed for multi-label image classification using VLMs. VLM based zero-shot image classification has been further improved by LLM-based prompting, in which LLMs are used to generate descriptive class definitions which are then used as prompts to the VLM text encoder during inference [27, 31]. Other works have adapted VLMs for zero-shot single-label action classification in videos [42, 43]. A typical video," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "however, contains multiple concepts including objects, actions, scenes, events, etc. Our goal is to develop a video classifier that can simultaneously recognize the presence of these different concepts from any vocabulary specified at inference. We refer to this task as open-vocabulary multi-label video classification. An illustration of the subtle differences between the prior tasks and our proposed task is presented in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Open-vocabulary multi-label video classification presents two unique challenges. First, unlike the single label open-vocabulary setting, in the multi-label setting, identifying relevant concepts in a video based on VLM similarity score performs poorly, as VLM similarity scores for different types of concepts (e.g., actions, objects) often fall within different ranges, even with LLM-guided prompting (see Section 4.5). Addressing this issue requires end-to-end finetuning of the VLM with LLM guidance on datasets with diverse video concepts. However, performing end-to-end finetuning while simultaneously incorporating the benefits of LLM-guided prompting, presents additional technical challenges, particularly in backpropagating gradients through the VLMs text encoder's tokenizer. Second, adapting pretrained image-language models to recognize video concepts while retaining strong zero-shot performance is challenging, as the image-language pretraining datasets are orders of magnitude larger than the largest labeled video datasets. VLM adaptation tends to overfit to the video data easily, thus losing their generalization capability [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.737 + ], + "angle": 0, + "content": "In this work, we attempt to address these challenges for open vocabulary multi-label video classification. First, to simultaneously recognize multiple video concepts in open-domain, we finetune the VLM in an end-to-end manner with LLM guidance to ensure proper ranking between different types of video concepts. Particularly, we extend the LLM-guided prompting approach for open-vocabulary classification by introducing learnable prefixes for prompting the LLM. To directly utilize the LLM output representations in the VLM's text encoder for end-to-end training, we propose a prompt transformer. This prompt transformer, in conjunction with the learnable prefixes, not only facilitates the integration of the LLM's world knowledge into the VLM's text representation but also aids in mitigating the discrepancy between VLM scores for various types of concepts. Second, for the effective recognition of temporally varying video concepts, we integrate an auxiliary temporal modeling branch with the VLM's vision encoder [24] and introduce a novel regularization penalty to retain the VLM's zero-shot performance during the adaptation to videos. Through extensive experimentation and analysis we demonstrate that our proposed approach significantly outperforms baseline solutions obtained by extending prior work on single label open vocabulary image and video classification." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.738, + 0.727, + 0.752 + ], + "angle": 0, + "content": "In summary, in this work we make three major technical contributions:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.784, + 0.808 + ], + "angle": 0, + "content": "- A novel end-to-end trainable approach to learn a strong label encoder for open vocabulary multi-label video classification by adapting an LLM to learn to prompt a VLM's text encoder." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- An approach for enhancing pre-trained VLM image encoders with temporal modeling capability while also retaining strong open-vocabulary performance." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.765, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.13 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "- Defining a new open vocabulary multi-label video classification benchmark using 5 datasets (2 closed, 3 open vocabulary) and benchmarking our approach against 6 strong baselines, significantly outperforming them." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.217, + 0.398, + 0.234 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.37 + ], + "angle": 0, + "content": "Vision-Language Representation Learning: Recently, image-language models [12, 18, 32, 39, 49, 50] have drawn huge attention because of their effectiveness in learning generic visual representations that are transferable to several downstream tasks like classification, retrieval, etc. This success can partly be attributed to the large-scale image-text datasets [4, 35-37, 41]. However, this is not the case for the video data. Therefore, to perform video-language pretraining, most recent works [6, 7, 26, 30, 47] bootstrap from a pretrained image-language model and adapt them to video representation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Open-Vocabulary Classification: Vision-language models do not perform optimally in visual recognition tasks out of the box. To improve the zero-shot open-vocabulary visual recognition capability of the vision-language models, the recent works either resort to regularized finetuning [43, 45] or prompt learning [40, 42, 52]. However, these works either focus on solving single label tasks or only perform adaptation for image recognition tasks. In contrast to these works, we focus on open-vocabulary video classification task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.796, + 0.643 + ], + "angle": 0, + "content": "Large Language Models in Vision: Recently, Large Language Models (LLM) have been utilized in vision tasks primarily because of their excellent in-context and zero-shot learning performance especially in commonsense reasoning tasks [15]. To be particular, LLMs have been used to rewrite noisy ASR text [8,38], to generate classification labels by mapping textual description to a predefined task list [22], extracting verbs from textual descriptions to improve action understanding [21]. LLMs have also been used to initialize textual encoders for vision language encoder [51]. In some recent works, vision representations have been aligned with LLM input space to solve multitude of vision-language tasks [17,23]. These multi-modal models have also been used to generate captions for images which are then utilized to train better vision-language models [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Prompting Vision-Language Models with LLMs: Some prior works have leveraged LLMs to augment the open vocabulary understanding capabilities of Vision-Language models. Classification By Description [27], CuPL [31], LLM guided concept bottleneck [48] all utilize LLM generated class descriptors to prompt CLIP to improve image classification, and make it more interpretable. WaffleCLIP [34] demonstrated that LLM generated higher level concepts to describe datasets can also improve classification. LLM based approaches have also been extended to object detection [13] and point cloud understanding [53]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Parameter efficient fine-tuning of LLMs: As full parameter finetuning of LLMs is not practical, quite a few parameter efficient finetuning methods have been developed. The lightweight Learnable Soft prompts [16] is the approach we utilize to guide the LLM in our method. Other approaches such as LoRA [10], Prefix-Tuning [20] and P-Tuning [25] could also be adopted in principle." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.205, + 0.784, + 0.235 + ], + "angle": 0, + "content": "Our proposal to boost the open vocabulary multi-label video classification capabilities of CLIP consists of two key parts: first, an end-to-end trainable label" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.461, + 0.527 + ], + "angle": 0, + "content": "encoder leveraging an LLM for strong open vocabulary capabilities, and the CLIP text encoder for visual alignment. The label encoder consists of a frozen LLM guided with learnable prefixes, whose outputs are mapped to the frozen CLIP text encoder using a learnable prompting transformer. Second, we enhance the CLIP image encoder's capabilities for understanding temporal dynamics. This is achieved by using a lightweight temporal modeling branch to enhance the CLIP image encoder. The details of our approach are illustrated in Figure 3. Each part of our method is discussed in detail in this section." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.46, + 0.639 + ], + "angle": 0, + "content": "An overview of our approach is provided in Fig. 2. Our method has three broad stages. Firstly, during the training phase, both the label encoder and video encoder are trained simultaneously, as shown in (a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.462, + 0.795 + ], + "angle": 0, + "content": "Secondly, during the classifier vocabulary expansion stage, embeddings for class labels are calculated and saved into a label embedding database, as shown in (b). This vocabulary can be extended at any point after training, thus allowing our method to be used in the open vocabulary setting. Finally, during the inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "stage, video features are computed and compared against label embeddings from the database. As the label embeddings are pre-computed, the computational overhead during inference over standard CLIP models is minimal." + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.274, + 0.782, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.394, + 0.642, + 0.404 + ], + "angle": 0, + "content": "(a) Training Stage" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.41, + 0.779, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.486, + 0.713, + 0.495 + ], + "angle": 0, + "content": "(b) Classifier Vocabulary Expansion Stage" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.497, + 0.774, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.604, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Fig.2: Our open vocabulary classification method includes three stages of operation. During the (a) training stage, we train our label and video encoders on closed set training labels. New class labels can be added to the vocabulary after training by employing the (b) classifier vocabulary expansion stage. During the (c) inference stage video embeddings are computed and matched with the label embeddings database to get the classification scores." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.7, + 0.162 + ], + "angle": 0, + "content": "3.1 LLMs for Semantic Enrichment of Label Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.401 + ], + "angle": 0, + "content": "Pretrained VLMs have strong open-vocabulary image classification performance owing to their large scale image-text pretraining. LLMs on the other hand are trained on even larger scale of text data with training objectives that enable a rich semantic understanding. As a result, compared to LLMs, VLMs have a more limited semantic understanding of natural language and have difficulty understanding concepts such as the relationship between different class labels. For instance, prior works have found that while VLMs are able to classify ImageNet images at a fine-grained level, their accuracy at higher levels of the class hierarchy is poor [46]. E.g., while they may recognize an image of a lion, they are unable to understand that a lion is also a feline, a mammal and an animal. However, an LLM is very effective in comprehending the hierarchical relationship between these labels, among many other capabilities, due to the large-scale training. Hence, in this work, we utilize an LLM to generate complementary information about the class labels that can be utilized to improve the vision-language alignment of VLM leading to better open-vocabulary classification performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.402, + 0.411, + 0.417 + ], + "angle": 0, + "content": "Fixed LLM prompting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.787, + 0.688 + ], + "angle": 0, + "content": "As a baseline, we first develop an LLM-based prompting method to adapt CLIP for open vocabulary video classification, extending the approach of Menon & Vondrick [27] for open vocabulary image classification. We design a prompt template (see implementation details in Section K of Supplementary material) for an encoder-decoder based LLM with the class name and a question asking it to generate useful features for visually distinguishing that class. The LLM output is then parsed into a list of textual descriptions, hereafter referred to as attributes. These attributes along with the class name is used to prompt the CLIP text encoder and the resultant text-embeddings are mean-pooled to obtain an attribute enriched text-embedding for the class. We mean-pool CLIP vision embedding of video frames to generate video embeddings and then perform open vocabulary classification by matching video embeddings with the text-embeddings of different classes. However, this simple baseline has a crucial weakness: it is not trainable, as there is a need to de-tokenize and tokenize the LLM text output to CLIP, which doesn't allow the flow of gradients. As a result, the process of generating class attributes by prompting an LLM cannot be improved by training on a labeled video dataset. To remedy this, we propose an end-to-end trainable architecture that integrates the LLM with the CLIP text encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.84 + ], + "angle": 0, + "content": "End-to-end learnable LLM prompting. Our proposed architecture, summarized in Figure 3, incorporates a learnable prompting framework with the frozen LLM to generate the inputs to a frozen CLIP text-encoder. The learnable components on the text side are limited to \\( N \\) learnable prefixes/vectors to the prompt template used for querying the LLM and a prompt transformer that transforms the sequence of tokens from the LLM to input soft prompts for the CLIP text-encoder. In our implementation we avoid the discrete and nondifferentiable operations like detokenization at the LLM decoder output and re-tokenization at the CLIP text-encoder input. This allows the prompt transformer to directly connect the LLM output semantic space to the CLIP input" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.151, + 0.734, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Fig. 3: Our end-to-end trainable system for open vocabulary video classification. The class labels are used by the LLM to generate useful class attributes for the CLIP text encoder which provides a visually aligned label embedding. The learnable input prompts to the LLM guide it to generate soft-attributes useful for video classification. Prompting transformer learns to map from the LLM output space to the CLIP input space. To add video understanding to CLIP's vision encoder we add additional spatiotemporal modeling layers. Details about each component are in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.672 + ], + "angle": 0, + "content": "semantic space, ensuring that the entire text model is differentiable and therefore end-to-end learnable. We describe the details below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "For each of the \\(N\\) LLM prefixes, we first construct an input sequence to the LLM comprising of the prefix followed by a fixed prompt template that contains the label name \\(\\ell\\) from the input data sample (see Figure 3). The text input to the LLM is mapped using a tokenizer and the LLM's embedding layer into a sequence of tokens \\(I\\in \\mathbb{R}^{M\\times d}\\), where \\(M\\) and \\(d\\) represent number of input tokens and the dimension of the embedding space respectively. All our \\(N\\) learnable prefixes are \\(d\\)-dimensional vectors. We use \\(P_{i}\\in \\mathbb{R}^{d}\\) to represent the \\(i\\)-th prefix. Each prefix is concatenated with the tokens of the prompt template, yielding a unified sequence of tokens \\([P_i;I]\\in \\mathbb{R}^{(1 + M)\\times d}\\). This combined sequence is then processed through the frozen encoder-decoder layers of the LLM. For each prefix, we run \\(KL\\) decoding iterations of the LLM decoder to generate a sequence of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "\\(KL\\) decoded tokens. These tokens represent useful class features in the LLM's semantic space. As the prompt template specifically asks LLM to output features as a list, we split the sequence of tokens evenly into \\(K\\) subsequences of \\(L\\) tokens each. Due to the nature of LLM prompt template, which prompts the LLM to output useful visual features, we refer to the subsequences of tokens as soft attributes, as they are sets of continuous vectors instead of discrete attributes in natural language. The \\(K\\) soft attributes are then individually processed by the prompt transformer to generate \\(K\\) soft prompts to the CLIP text encoder. Repeating this operation for each of the \\(N\\) LLM prefixes, we get \\(NK\\) soft prompts. Each soft-prompt is then concatenated with the tokenized label embedding and then processed by the frozen CLIP text-encoder to generate an attribute feature. All \\(NK\\) resulting features are mean-pooled and normalized to obtain the final label embedding \\(f_{t}(\\ell)\\), where \\(\\ell\\) is the label name." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.364, + 0.599, + 0.382 + ], + "angle": 0, + "content": "3.2 Regularized Parallel Temporal Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.49, + 0.78 + ], + "angle": 0, + "content": "The simple image transformer of the CLIP vision encoder does not model the temporal dynamics of the video, which is essential to enhance the ability to recognize entities in videos. We enhance our vision encoder by adding a parallel temporal modeling branch to the last \\(T\\) layers of the CLIP vision encoder as illustrated in Figure 3. We freeze the CLIP vision branch and train only the newly added temporal layers. Each block in the newly added branch consists of a spatial attention layer initialized from their corresponding CLIP weights, and a temporal attention layer that is randomly initialized. All frames of the video are processed independently by the CLIP vision backbone. Then, at the first temporal modeling block, the temporal token TMP is created by averaging the CLS tokens across frames, meanwhile learnable spatial and temporal positional embeddings are added to each patch token. The \\(t\\)th temporal modeling block takes in as input a" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.418, + 0.755, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.619, + 0.789, + 0.746 + ], + "angle": 0, + "content": "Fig. 4: Our Temporal Block projects frame patch tokens from the CLIP image encoder, fusing them with previous block's temporal branch tokens. The temporal token (TMP) incorporates all frames' CLS tokens. Divided Space-Time attention layers form the core of the block. Spatial attention layers are initialized from CLIP weights and regularized using stochastic weight averaging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "weighted combination of the previous temporal modeling block and the corresponding layer in the CLIP backbone. Symbolically if \\( V_{t}^{\\top} \\) represents the patch tokens from the temporal block at the \\( t \\)th block and \\( V_{t} \\) represents the patch tokens for the corresponding CLIP layers, the patch tokens for the \\( t \\)th block" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.221 + ], + "angle": 0, + "content": "\\(V_{t}^{\\mathsf{T}} = V_{t - 1}^{\\mathsf{T}} + Proj_{\\text{spatial}}(V_{s})\\), where \\(V_{s}\\) is the corresponding token from the CLIP backbone. After that the tokens are processed by the divided space-time attention layer and spatial layer. The overall video embedding is generated at the end by mean pooling the final TMP token and the CLS tokens for each frame from the CLIP backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.344 + ], + "angle": 0, + "content": "A key drawback of adding a deep parallel branch is that while it improves performance in the finetuned setting, it diminishes CLIP's zero-shot performance. In order to overcome this limitation in the zero-shot setting, we propose to use weight regularization on the spatial attention layers. The weight regularization operation tries to match predictions made by the current set of finetuned weights and a randomly weighted average of the current weights and the original CLIP weights. Symbolically, at each iteration we use weights \\(\\theta\\) which are a stochastic weighted average of the finetuned and the original frozen weights as given by:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.361, + 0.672, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\theta = \\alpha \\theta_ {\\mathrm {f t}} + (1 - \\alpha) \\theta_ {\\mathrm {f r o z e n}}, \\quad \\mathrm {w h e r e} \\quad \\alpha \\sim U (0, \\lambda)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.383, + 0.743, + 0.399 + ], + "angle": 0, + "content": "We observe that an empirical value of \\(\\lambda = 0.5\\) works well across datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Relationship to prior works: Prior temporal modeling approaches split into two camps: improving finetuned performance or preserving CLIP's zero-shot abilities. STAN [24] is a state-of-the-art finetuning focused approach which adds a parallel temporal modeling branch which is able to leverage both high-level and low-level features from the CLIP vision encoder. The divided space-time attention in STAN's temporal block is inspired by the TimeSformer [2] architecture. We differ from STAN in order to achieve stronger open vocabulary performance in a number of ways. Firstly, unlike STAN we do not finetune the CLIP image encoder, only the added parallel branch. Additionally, we propose stochastic weight regularization to prevent the temporal branch from drifting too far from the original CLIP feature space. Our ablation experiments (Table 3) demonstrate that both these choices significantly improve our performance. Some form of weight regularization has been explored by methods such as Open-VCLIP [44]. However, Open V-CLIP takes in only 3 frames at a time, and its temporal modeling range is limited. Our approach is able to achieve the best of both approaches." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.663, + 0.414, + 0.679 + ], + "angle": 0, + "content": "3.3 Training objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We train our model on multi-label video datasets. To construct a batch, we first sample a set \\(\\mathcal{B}\\) of \\(B\\) videos from the dataset. Each video \\(v\\) is associated with a set of positive class labels which we denote \\(\\mathcal{P}(v)\\). In addition to the positive labels, we augment the batch with random negative class labels for each video. To obtain the negatives, we first identify the set \\(\\mathcal{P}_{\\mathcal{B}} := \\cup_{v \\in \\mathcal{B}} \\mathcal{P}(v)\\) of distinct classes among all the positive labels from all the videos in the batch. We then sample a random set \\(\\mathcal{N}_{\\mathcal{B}}\\) of \\(4B - |\\mathcal{P}_{\\mathcal{B}}|\\) classes from the rest of the class vocabulary of the dataset. For each video \\(v\\), we then choose all non-positive classes from \\(\\mathcal{P}_{\\mathcal{B}} \\cup \\mathcal{N}_{\\mathcal{B}}\\) as negative labels. We use \\(\\mathcal{N}(v) := (\\mathcal{P}_{\\mathcal{B}} \\cup \\mathcal{N}_{\\mathcal{B}}) \\setminus \\mathcal{P}(v)\\) to denote the set of all negative labels for \\(v\\). The resulting positive and negative video-label pairs are" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "then treated as training samples for binary classification. Each training sample is a label, video pair \\((\\ell, v)\\) where \\(\\ell\\) represents the label name and \\(v\\) the video. With this batch construction, the total number of training samples in a batch could be variable, but the total number of videos is fixed at \\(B\\) and the total number of classes present among the samples is fixed at \\(4B\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.518, + 0.345 + ], + "angle": 0, + "content": "For each training data sample \\((\\ell, v)\\), the text and video encoders respectively generate a unit-norm text embedding \\(f_{t}(\\ell) \\in \\mathbb{R}^{D}\\) and a unit-norm video embedding \\(f_{v}(v) \\in \\mathbb{R}^{D}\\) where \\(D\\) is the common embedding dimension. The score \\(s(\\ell, v)\\) for this data sample is given by the inner product:" + }, + { + "type": "equation", + "bbox": [ + 0.283, + 0.354, + 0.516, + 0.372 + ], + "angle": 0, + "content": "\\[\ns (\\ell , v) = \\left(f _ {t} (\\ell)\\right) ^ {\\mathsf {T}} f _ {v} (v). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.516, + 0.412 + ], + "angle": 0, + "content": "The model is trained with a weighted binary cross entropy loss:" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.255, + 0.77, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.379, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig.5: Our training objective applies binary cross-entropy to predicted Video-Label feature similarities sharpened by a temperature-scaled sigmoid." + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.422, + 0.603, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathcal {B}) = - \\sum_ {v \\in \\mathcal {B}} \\left[ \\sum_ {\\ell \\in \\mathcal {P} (v)} \\log p (\\ell , v) + w \\sum_ {\\ell \\in \\mathcal {N} (v)} \\log \\left(1 - p (\\ell , v)\\right) \\right] ^ {\\text {e n e d b y a t}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.787, + 0.536 + ], + "angle": 0, + "content": "where \\( p(\\ell, v) \\coloneqq \\sigma\\left(\\frac{s(\\ell, v)}{\\tau}\\right) \\), \\( \\tau \\) is a temperature parameter, \\( \\sigma(.) \\) is the sigmoid function and \\( w > 0 \\) is a weight hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.558, + 0.497, + 0.575 + ], + "angle": 0, + "content": "4 Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.336, + 0.603 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In order to train our model on a wide range of concepts, we train it on a mix of YouTube8M [1] and Kinetics-400 [14]. YouTube-8M (YT-8M) is primarily labeled with entities, whereas Kinetics-400 (K400) is labeled only with actions. As YT-8M is composed of a random sample of the whole of YouTube, a significant portion of video its samples are from video game streams. Many of these videos are just tagged with the video game title as the label irrespective of the actual entities and actions present in the video. A significant portion of its label vocabulary is dedicated to video game titles (771 out of 3862 total) are dedicated to video game titles. We observe training instabilities due to this, and to fix this issue, we remove video game titles from the training vocabulary. We also remove some of the least frequently occurring labels during training to reach a training vocabulary of 2429 classes. This leads to stabilized training without further issues. Note that for closed set evaluation on YT-8M, we use the YouTube-8M Segments validation set, which contains human verified labels for 1000 classes. K400 on the other hand is widely used, and has a clean vocabulary and high" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "quality human verified action labels. From the training perspective, the only limitation is the absence of entity (object, scenes etc.) labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.375 + ], + "angle": 0, + "content": "For evaluation, we use 3 open-vocabulary test datasets: TAO (Tracking Any Object) [3], ActivityNet [9] and RareAct [28]. TAO and ActivityNet are exclusively object and action focused datasets respectively. RareAct has object, action labels, and focuses on unusual object-action combination. TAO dataset was developed for evaluating object trackers, however analogous to how object detection datasets like MS-COCO are used for evaluating multi-label image classification methods such as DualCoOp [40], we transform TAO into a multi-label video classification dataset by ignoring localization annotations. ActivityNet often has multiple actions in a given video. For our evaluation each annotation from RareAct provides three labels: object, action and the (unusual) object-action combination, this provides an interesting test of the open vocabulary capabilities. Altogether, these five datasets provide a comprehensive evaluation of a model's closed and open vocabulary video classification performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.398, + 0.341, + 0.413 + ], + "angle": 0, + "content": "4.2Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.784, + 0.485 + ], + "angle": 0, + "content": "As there are no prior works that explicitly address multilabel open vocabulary video classification, we extend the following methods from the zero shot image and action classification literature to our multi-label video classification setting for comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.784, + 0.514 + ], + "angle": 0, + "content": "CoOp [52]: Short for Context Optimization, CoOp learns prompts for the CLIP text encoder as a lightweight adaptation technique for classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.784, + 0.573 + ], + "angle": 0, + "content": "DualCoOp [40]: Is an extension of CoOP to the multi-label setting where both positive and negative prompts are learnt for the CLIP text encoder. For a given label, prediction is based on whether the positive or negative prompted version has higher similarity with the image feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.784, + 0.604 + ], + "angle": 0, + "content": "LLM + CLIP (Frozen): This baseline was discussed in Section 3.1 and is illustrated in Figure 8 of Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.784, + 0.635 + ], + "angle": 0, + "content": "ViFi-CLIP [33]: In this baseline, there is no temporal modeling, and both the CLIP image and text encoder are finetuned on the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.784, + 0.711 + ], + "angle": 0, + "content": "As CoOp and ViFi-CLIP were developed for single label classification, to utilize them in our setting we replace their contrastive loss function with our multi-label classification loss. We also drop the region aggregation aspect of DualCoOp and only test the dual prompting architecture. We refer to these baselines as CoOp*, DualCoOp* and ViFi-CLIP* to reflect these differences." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.327, + 0.751 + ], + "angle": 0, + "content": "4.3 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We report two sets of metrics to evaluate the performance of the models in the open vocabulary setting. Area Under Precision-Recall curve (AUPR) summarizes the overall classification performance across the entire precision-recall trade-off. Secondly, we report the Peak F1-Score for the model on each dataset, which is the F1-Score achieved by that model at the optimal threshold chosen by an oracle" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.144, + 0.777, + 0.33 + ], + "angle": 0, + "content": "
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities+Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt6.326.243.844.29.5
CLIP with Prompt Templates6.830.546.045.911.4
CLIP with Fixed LLM Prompts6.930.650.246.811.5
Trainable Baseline Methods
CoOp2.717.835.028.83.5
DualCoOp8.323.947.133.07.6
ViFi-CLIP3.410.958.317.24.1
Ours
CLIP + Learnable LLM Prompts9.432.851.447.111.9
+ Temporal Modeling14.842.063.847.112.4
+ Synthetic Labels16.743.265.550.213.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.318, + 0.33, + 0.684, + 0.342 + ], + "angle": 0, + "content": "Table 1: AUPR scores for all methods on all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.434 + ], + "angle": 0, + "content": "for that dataset. This metric captures the classification performance obtained by an open vocabulary classification method on a dataset, if the threshold alone could be tuned, e.g., by using a labeled validation set." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.326, + 0.471 + ], + "angle": 0, + "content": "4.4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.668 + ], + "angle": 0, + "content": "AUPR scores for each method are presented in Table 1 and Peak F1 scores are presented in Table 2. Our learnable LLM prompting approach outperforms both frozen and trainable baselines across both closed-vocabulary and open-vocabulary classification tasks. Also notable is the enhancement in performance due to our temporal modeling approach (average gain of \\(\\approx 5\\%\\) in AUPR). As previously mentioned in our discussion of the dataset, the training datasets YT-8M and K400 have certain limitations (in particular, YT-8M has few action labels and K400 has no object/entity labels). We present further improved results obtained by augmenting our training dataset with additional class labels obtained using a multimodal LLM (pipeline is detailed in Section I of the Supplementary Material). As seen from the last rows of Tables 1 and 2, this approach gives a further \\(\\approx 2\\%\\) improvement, and is specially effective for actions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.611, + 0.708 + ], + "angle": 0, + "content": "4.5 Improved score calibration across datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In order to use a classifier in a truly open vocabulary setting in practice, the classification score needs to be well calibrated across different types of concepts. This would allow us to pre-select an optimal threshold to generate binary classification for all concepts. In Figure 6 we demonstrate the robustness offered by our method in setting a threshold that works for a wide range of concepts, and the advantage offered by our method over the frozen CLIP + LLM baseline. These figures plot the F1 scores for multi-label classification across thresholds provided by both methods with Figure 6a showing the numbers on the validation splits" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.144, + 0.781, + 0.329 + ], + "angle": 0, + "content": "
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities + Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt14.934.244.647.117.6
CLIP with Prompt Templates20.836.849.250.120.5
CLIP with Fixed LLM Prompts21.637.350.251.419.8
Trainable Baseline Methods
CoOp5.825.543.935.510.5
DualCoOp16.233.249.040.515.0
ViFi-CLIP5.219.354.424.79.6
Ours
CLIP + Learnable LLM Prompts23.642.452.851.122.6
+ Temporal Modeling31.546.259.652.624.3
+ Synthetic Labels32.746.656.653.825.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.33, + 0.691, + 0.341 + ], + "angle": 0, + "content": "Table 2: Peak F1 scores for all methods on all datasets." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.377, + 0.465, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.484, + 0.446, + 0.496 + ], + "angle": 0, + "content": "(a) Closed Vocabulary Evaluation." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.377, + 0.779, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.484, + 0.754, + 0.496 + ], + "angle": 0, + "content": "(b) Open Vocabulary Evaluation." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Fig. 6: F1-Scores at different thresholds for closed and open vocabulary evaluation datasets. Our end-to-end trained model can achieve better performance across datasets with a single threshold (labeled by gray vertical line) chosen on the supervised datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.75 + ], + "angle": 0, + "content": "of the training datasets, and Figure 6b showing the numbers in the open vocabulary setting on two diverse datasets. A reasonable choice for a classification threshold would be to pick the threshold that maximizes the minimum validation F1-score for the datasets used in training. The resulting choice of thresholds for our method and for frozen CLIP are shown in Figure 6a. We see that for the baseline method the performance on TAO is poor, while that on ActivityNet is modest. However with our proposed method the F1 scores for both evaluation datasets are simultaneously high. This result empirically validates the versatility that is achieved by our learnable LLM-based prompting scheme over the baseline, which makes it possible to pre-select a threshold for using the solution as a black box classifier for all unseen concepts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.779, + 0.342, + 0.792 + ], + "angle": 0, + "content": "4.6 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.786, + 0.842 + ], + "angle": 0, + "content": "We scientifically ablate each part of our framework, reporting ablation evaluations (AUPR) on TAO (Objects) and ActivityNet (Actions)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.145, + 0.411, + 0.228 + ], + "angle": 0, + "content": "
Open Vocabulary
# BlocksObjectsActions
565.448.1
465.550.2
264.649.1
162.347.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.23, + 0.411, + 0.242 + ], + "angle": 0, + "content": "(a) Effect of Temporal Modeling" + }, + { + "type": "table", + "bbox": [ + 0.441, + 0.144, + 0.78, + 0.23 + ], + "angle": 0, + "content": "
Open Vocabulary
Spatial Reg.BackboneObjectsActionsGeo. Mean
✓56.653.855.2
X62.132.544.9
X42.728.334.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.231, + 0.702, + 0.243 + ], + "angle": 0, + "content": "(b) Effect of our Regularization" + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.255, + 0.779, + 0.27 + ], + "angle": 0, + "content": "Table 3: Ablations of Temporal Modeling Branch Architecture and Regularization" + }, + { + "type": "table", + "bbox": [ + 0.354, + 0.296, + 0.65, + 0.389 + ], + "angle": 0, + "content": "
Label EncoderOpen Vocabulary
LLMPTCLIPObjects Actions
Learnable Prompt✓✗25.2 8.5
Fixed Prompt✗✓62.5 41.3
Fixed Prompt✓✓64.9 48.1
Learnable Prompt✓✓65.5 50.2
✗✗✓60.1 34.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.278, + 0.389, + 0.725, + 0.403 + ], + "angle": 0, + "content": "Table 4: Ablating the Label Encoder. PT \\(\\rightarrow\\) Prompt Transformer" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.446, + 0.785, + 0.475 + ], + "angle": 0, + "content": "- Temporal Modeling Blocks: We find that using more Blocks improves performance, until around 4. (Table 3a)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.55 + ], + "angle": 0, + "content": "- Temporal Modeling Weight Regularization: We find that using our weight regularization strategy for the spatial layers of the temporal branch prevents the model from overfitting to certain concepts. We also validate our choice of keeping the CLIP vision backbone frozen and find that unfreezing the backbone leads to severe over-fitting. (Table 3b)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.551, + 0.787, + 0.642 + ], + "angle": 0, + "content": "- Label Encoder: We demonstrate the effectiveness of combining an LLM with learnable prompts and CLIP's text encoder as the label encoder. First, we show that removing CLIP text encoder leads to a significant drop in performance. This demonstrates that the visual-text alignment learned by CLIP is essential. Next we demonstrate the benefits of learnable prompts over fixed prompts. We also provide results with the LLM removed (Table 4)." + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.661, + 0.36, + 0.677 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduced the problem of open vocabulary multi-label video classification and proposed a solution leveraging LLMs and pre-trained VLMs. Through our extensive experiments we demonstrated strong performance on both actions and entities with a single model. Our proposed approach benefits from two key innovations. First, we proposed a method to prompt VLMs more effectively by utilizing the LLM output representations through a trainable prompt transformer and learnable LLM prompts. Second, we introduced a temporal modeling architecture and a regularized finetuning approach to improve the video understanding capability of the vision encoder while retaining strong open-vocabulary performance. Our ablations validate the efficacy of each of these contributions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.255 + ], + "angle": 0, + "content": "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: Proceedings of the International Conference on Machine Learning (ICML) (July 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.255, + 0.785, + 0.31 + ], + "angle": 0, + "content": "3. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 436-454. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Desai, K., Kaul, G., Aysola, Z., Johnson, J.: Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.363 + ], + "angle": 0, + "content": "5. Fan, L., Krishnan, D., Isola, P., Katabi, D., Tian, Y.: Improving clip training with language rewrites. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.785, + 0.39 + ], + "angle": 0, + "content": "6. Fang, H., Xiong, P., Xu, L., Chen, Y.: Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.785, + 0.444 + ], + "angle": 0, + "content": "7. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.514 + ], + "angle": 0, + "content": "8. Gupta, R., Roy, A., Christensen, C., Kim, S., Gerard, S., Cincebeaux, M., Divakaran, A., Grindal, T., Shah, M.: Class prototypes based contrastive learning for classifying multi-label and fine-grained educational videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19923-19933 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.58 + ], + "angle": 0, + "content": "9. Heilbron, F.C., Niebles, J.C.: Collecting and annotating human activities in web videos. In: Proceedings of International Conference on Multimedia Retrieval. p. 377-384. ICMR '14, Association for Computing Machinery, New York, NY, USA (2014). https://doi.org/10.1145/2578726.2578775, https://doi.org/10.1145/2578726.2578775" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.635 + ], + "angle": 0, + "content": "0. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.718 + ], + "angle": 0, + "content": "1. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.772 + ], + "angle": 0, + "content": "2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "3. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. In: International Conference on Machine Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., Suleyman, M., Zisserman, A.: The kinetics human action video dataset (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.205 + ], + "angle": 0, + "content": "15. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=e2TBb5y0yFf" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.206, + 0.787, + 0.288 + ], + "angle": 0, + "content": "16. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.289, + 0.787, + 0.373 + ], + "angle": 0, + "content": "17. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 12888-12900. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/1i22n.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.415 + ], + "angle": 0, + "content": "18. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.416, + 0.787, + 0.471 + ], + "angle": 0, + "content": "19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10965-10975 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.57 + ], + "angle": 0, + "content": "20. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.571, + 0.787, + 0.613 + ], + "angle": 0, + "content": "21. Lin, W., Karlinsky, L., Shvetsova, N., Possegger, H., Kozinski, M., Panda, R., Feris, R., Kuehne, H., Bischof, H.: Match, expand and improve: Unsupervised finetuning for zero-shot action recognition with language knowledge. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.655 + ], + "angle": 0, + "content": "22. Lin, X., Petroni, F., Bertasius, G., Rohrbach, M., Chang, S.F., Torresani, L.: Learning to recognize procedural activities with distant supervision. arXiv preprint arXiv:2201.10990 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.684 + ], + "angle": 0, + "content": "23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems. Curran Associates, Inc. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.787, + 0.741 + ], + "angle": 0, + "content": "24. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.797 + ], + "angle": 0, + "content": "25. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. AI Open (2023). https://doi.org/https://doi.org/10.1016/j.aiopen.2023.08.012, https://www.sciencedirect.com/science/article/pii/S2666651023000141" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.798, + 0.787, + 0.841 + ], + "angle": 0, + "content": "26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.395, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary Multi-Label Video Classification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "27. Menon, S., Vondrick, C.: Visual classification via description from large language models. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=j1AjNL8z5cs" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "28. Miech, A., Alayrac, J.B., Laptev, I., Sivic, J., Zisserman, A.: Rareact: A video dataset of unusual interactions. arxiv:2008.01018 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.286 + ], + "angle": 0, + "content": "29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 728-755. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "30. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "31. Pratt, S., Covert, I., Liu, R., Farhadi, A.: What does a platypus look like? generating customized prompts for zero-shot image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15691-15701 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.438 + ], + "angle": 0, + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "33. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.536 + ], + "angle": 0, + "content": "34. Roth, K., Kim, J.M., Koepke, A.S., Vinyals, O., Schmid, C., Akata, Z.: Waffling around for performance: Visual classification with random words and broad concepts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15746-15757 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.591 + ], + "angle": 0, + "content": "35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.633 + ], + "angle": 0, + "content": "36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.688 + ], + "angle": 0, + "content": "37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.716 + ], + "angle": 0, + "content": "38. Shvetsova, N., Kukleva, A., Hong, X., Rupprecht, C., Schiele, B., Kuehne, H.: HowtocapTION: Prompting llms to transform video annotations at scale (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.77 + ], + "angle": 0, + "content": "39. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "40. Sun, X., Hu, P., Saenko, K.: Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems 35, 30569-30582 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Commun. ACM" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Gupta et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.174 + ], + "angle": 0, + "content": "59(2), 64-73 (jan 2016). https://doi.org/10.1145/2812802, https://doi.org/10.1145/2812802" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "42. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23034–23044 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.3 + ], + "angle": 0, + "content": "43. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-VCLIP: Transforming CLIP to an open-vocabulary video model via interpolated weight optimization. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 36978-36989. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/weng23b.htm1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.786, + 0.341 + ], + "angle": 0, + "content": "44. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.342, + 0.786, + 0.397 + ], + "angle": 0, + "content": "45. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.786, + 0.438 + ], + "angle": 0, + "content": "46. Xu, Z., Zhu, Y., Deng, T., Mittal, A., Chen, Y., Wang, M., Favaro, P., Tighe, J., Modolo, D.: Challenges of zero-shot recognition with vision-language models: Granularity and correctness (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "47. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.786, + 0.535 + ], + "angle": 0, + "content": "48. Yang, Y., Panagopoulou, A., Zhou, S., Jin, D., Callison-Burch, C., Yatskar, M.: Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19187-19197 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.536, + 0.786, + 0.577 + ], + "angle": 0, + "content": "49. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.786, + 0.618 + ], + "angle": 0, + "content": "50. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.619, + 0.786, + 0.646 + ], + "angle": 0, + "content": "51. Zhao, Y., Misra, I., Krahenbuhl, P., Girdhar, R.: Learning video representations from large language models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.786, + 0.729 + ], + "angle": 0, + "content": "53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2639-2650 (October 2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.786, + 0.729 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_origin.pdf b/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d5d501c6d6b2e628cd65af4075f8dc419b2164dc --- /dev/null +++ b/2024/Open Vocabulary Multi-Label Video Classification/6bdc56f6-e0f6-48fd-ba36-782933480233_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffe4d2aaa00bcf9c023f40860adcddf41ac47dda59ab93d93795c9de1cff78ab +size 1136872 diff --git a/2024/Open Vocabulary Multi-Label Video Classification/full.md b/2024/Open Vocabulary Multi-Label Video Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7bc3e96a93c3cdbcef7f54e256cd82dcc0782e55 --- /dev/null +++ b/2024/Open Vocabulary Multi-Label Video Classification/full.md @@ -0,0 +1,281 @@ +# Open Vocabulary Multi-Label Video Classification + +Rohit Gupta1, Mamshad Nayeem Rizve2, Jayakrishnan Unnikrishnan2, + +Ashish Tawari $^{2}$ , Son Tran $^{2}$ , Mubarak Shah $^{1,2}$ + +Benjamin Yao $^{2}$ , and Trishul Chilimbi $^{2}$ + +1 Center for Research in Computer Vision, University of Central Florida + +rohit.gupta@ucf.edu, shah@crv.ucf.edu + +2 Amazon + +{jayunn,atakari,sontran}@amazon.com + +Abstract. Pre-trained vision-language models (VLMs) have enabled significant progress in open vocabulary computer vision tasks such as image classification, object detection and image segmentation. Some recent works have focused on extending VLMs to open vocabulary single label action classification in videos. However, previous methods fall short in holistic video understanding which requires the ability to simultaneously recognize multiple actions and entities e.g., objects in the video in an open vocabulary setting. We formulate this problem as open vocabulary multi-label video classification and propose a method to adapt a pre-trained VLM such as CLIP to solve this task. We leverage large language models (LLMs) to provide semantic guidance to the VLM about class labels to improve its open vocabulary performance with two key contributions. First, we propose an end-to-end trainable architecture that learns to prompt an LLM to generate soft attributes for the CLIP text-encoder to enable it to recognize novel classes. Second, we integrate a temporal modeling module into CLIP's vision encoder to effectively model the spatio-temporal dynamics of video concepts as well as propose a novel regularized finetuning technique to ensure strong open vocabulary classification performance in the video domain. Our extensive experimentation showcases the efficacy of our approach on multiple benchmark datasets. + +Keywords: Open Vocabulary $\cdot$ Multi-Modal $\cdot$ Video Understanding + +# 1 Introduction + +Video classification is a critical challenge in computer vision. Its goal involves recognizing a diverse array of concepts depicted in a video, which may include primarily static entities such as objects and scenes, as well as dynamic actions. In the classic setting, the vocabulary of all possible classes of interest is known in advance, and the model is trained in a supervised manner using a labeled dataset. The labor-intensive process of manual annotation often results in video datasets that are narrowly focused, such as those limited to specific sports or simple activities, which restricts the breadth of concepts models can learn. As video + +![](images/fa6d656ede06a3b9bd52a63c9f82f23519323f4871da1b926beb4fb11ac58ace.jpg) +Fig. 1: Our task: recognize multiple open-vocabulary classes in videos at inference from an open vocabulary, including entities (in blue) such as objects and scenes, and actions (in red). Prior zero-shot image classification approaches (e.g. CLIP) can label the salient entity in each frame (left), while zero-shot action-recognition approaches (e.g. ViFi-CLIP) (middle) can classify video-level action. Our method (right) can recognize all action or entity classes present in the video. + +applications are becoming widespread, there is an increasing need for developing video models that can recognize a broad range of concepts. This necessitates the development of open-vocabulary approaches for video classification that can recognize a diverse range of video concepts, including those that were not part of the class vocabulary present in the labeled training dataset. In this work, to solve this problem, we aim to leverage recent advances in vision-language models (VLM), which are trained to align visual and textual data and large language models (LLM), which have been demonstrated to possess a rich understanding of the world due to large scale pre-training. + +VLMs that are trained for image-language alignment on image-text pairs at large-scale exhibit remarkable zero-shot visual recognition performance across a diverse set of tasks [11, 19, 29, 32]. For classification tasks, the pretrained text encoder acts as a "label encoder" for the class labels, mapping them into the same representation space as the visual embedding of the image, which allows for the image to be classified by ranking the labels in order of representation similarity. Since VLMs are primarily trained to rank, picking the top match usually results in excellent performance on single label zero-shot classification. However, merely ranking the labels is insufficient to achieve open-vocabulary multi-label classification, and as a result, specialized methods [40] have been developed for multi-label image classification using VLMs. VLM based zero-shot image classification has been further improved by LLM-based prompting, in which LLMs are used to generate descriptive class definitions which are then used as prompts to the VLM text encoder during inference [27, 31]. Other works have adapted VLMs for zero-shot single-label action classification in videos [42, 43]. A typical video, + +however, contains multiple concepts including objects, actions, scenes, events, etc. Our goal is to develop a video classifier that can simultaneously recognize the presence of these different concepts from any vocabulary specified at inference. We refer to this task as open-vocabulary multi-label video classification. An illustration of the subtle differences between the prior tasks and our proposed task is presented in Fig. 1. + +Open-vocabulary multi-label video classification presents two unique challenges. First, unlike the single label open-vocabulary setting, in the multi-label setting, identifying relevant concepts in a video based on VLM similarity score performs poorly, as VLM similarity scores for different types of concepts (e.g., actions, objects) often fall within different ranges, even with LLM-guided prompting (see Section 4.5). Addressing this issue requires end-to-end finetuning of the VLM with LLM guidance on datasets with diverse video concepts. However, performing end-to-end finetuning while simultaneously incorporating the benefits of LLM-guided prompting, presents additional technical challenges, particularly in backpropagating gradients through the VLMs text encoder's tokenizer. Second, adapting pretrained image-language models to recognize video concepts while retaining strong zero-shot performance is challenging, as the image-language pretraining datasets are orders of magnitude larger than the largest labeled video datasets. VLM adaptation tends to overfit to the video data easily, thus losing their generalization capability [42]. + +In this work, we attempt to address these challenges for open vocabulary multi-label video classification. First, to simultaneously recognize multiple video concepts in open-domain, we finetune the VLM in an end-to-end manner with LLM guidance to ensure proper ranking between different types of video concepts. Particularly, we extend the LLM-guided prompting approach for open-vocabulary classification by introducing learnable prefixes for prompting the LLM. To directly utilize the LLM output representations in the VLM's text encoder for end-to-end training, we propose a prompt transformer. This prompt transformer, in conjunction with the learnable prefixes, not only facilitates the integration of the LLM's world knowledge into the VLM's text representation but also aids in mitigating the discrepancy between VLM scores for various types of concepts. Second, for the effective recognition of temporally varying video concepts, we integrate an auxiliary temporal modeling branch with the VLM's vision encoder [24] and introduce a novel regularization penalty to retain the VLM's zero-shot performance during the adaptation to videos. Through extensive experimentation and analysis we demonstrate that our proposed approach significantly outperforms baseline solutions obtained by extending prior work on single label open vocabulary image and video classification. + +In summary, in this work we make three major technical contributions: + +- A novel end-to-end trainable approach to learn a strong label encoder for open vocabulary multi-label video classification by adapting an LLM to learn to prompt a VLM's text encoder. +- An approach for enhancing pre-trained VLM image encoders with temporal modeling capability while also retaining strong open-vocabulary performance. + +- Defining a new open vocabulary multi-label video classification benchmark using 5 datasets (2 closed, 3 open vocabulary) and benchmarking our approach against 6 strong baselines, significantly outperforming them. + +# 2 Related Works + +Vision-Language Representation Learning: Recently, image-language models [12, 18, 32, 39, 49, 50] have drawn huge attention because of their effectiveness in learning generic visual representations that are transferable to several downstream tasks like classification, retrieval, etc. This success can partly be attributed to the large-scale image-text datasets [4, 35-37, 41]. However, this is not the case for the video data. Therefore, to perform video-language pretraining, most recent works [6, 7, 26, 30, 47] bootstrap from a pretrained image-language model and adapt them to video representation tasks. + +Open-Vocabulary Classification: Vision-language models do not perform optimally in visual recognition tasks out of the box. To improve the zero-shot open-vocabulary visual recognition capability of the vision-language models, the recent works either resort to regularized finetuning [43, 45] or prompt learning [40, 42, 52]. However, these works either focus on solving single label tasks or only perform adaptation for image recognition tasks. In contrast to these works, we focus on open-vocabulary video classification task. + +Large Language Models in Vision: Recently, Large Language Models (LLM) have been utilized in vision tasks primarily because of their excellent in-context and zero-shot learning performance especially in commonsense reasoning tasks [15]. To be particular, LLMs have been used to rewrite noisy ASR text [8,38], to generate classification labels by mapping textual description to a predefined task list [22], extracting verbs from textual descriptions to improve action understanding [21]. LLMs have also been used to initialize textual encoders for vision language encoder [51]. In some recent works, vision representations have been aligned with LLM input space to solve multitude of vision-language tasks [17,23]. These multi-modal models have also been used to generate captions for images which are then utilized to train better vision-language models [5]. + +Prompting Vision-Language Models with LLMs: Some prior works have leveraged LLMs to augment the open vocabulary understanding capabilities of Vision-Language models. Classification By Description [27], CuPL [31], LLM guided concept bottleneck [48] all utilize LLM generated class descriptors to prompt CLIP to improve image classification, and make it more interpretable. WaffleCLIP [34] demonstrated that LLM generated higher level concepts to describe datasets can also improve classification. LLM based approaches have also been extended to object detection [13] and point cloud understanding [53] + +Parameter efficient fine-tuning of LLMs: As full parameter finetuning of LLMs is not practical, quite a few parameter efficient finetuning methods have been developed. The lightweight Learnable Soft prompts [16] is the approach we utilize to guide the LLM in our method. Other approaches such as LoRA [10], Prefix-Tuning [20] and P-Tuning [25] could also be adopted in principle. + +# 3 Method + +Our proposal to boost the open vocabulary multi-label video classification capabilities of CLIP consists of two key parts: first, an end-to-end trainable label + +encoder leveraging an LLM for strong open vocabulary capabilities, and the CLIP text encoder for visual alignment. The label encoder consists of a frozen LLM guided with learnable prefixes, whose outputs are mapped to the frozen CLIP text encoder using a learnable prompting transformer. Second, we enhance the CLIP image encoder's capabilities for understanding temporal dynamics. This is achieved by using a lightweight temporal modeling branch to enhance the CLIP image encoder. The details of our approach are illustrated in Figure 3. Each part of our method is discussed in detail in this section. + +An overview of our approach is provided in Fig. 2. Our method has three broad stages. Firstly, during the training phase, both the label encoder and video encoder are trained simultaneously, as shown in (a). + +Secondly, during the classifier vocabulary expansion stage, embeddings for class labels are calculated and saved into a label embedding database, as shown in (b). This vocabulary can be extended at any point after training, thus allowing our method to be used in the open vocabulary setting. Finally, during the inference + +stage, video features are computed and compared against label embeddings from the database. As the label embeddings are pre-computed, the computational overhead during inference over standard CLIP models is minimal. + +![](images/37fd6f002647e864d9f3e79b261b57ffc0482b5da92c96721421e1e836e18748.jpg) +(a) Training Stage + +![](images/4ef2898a2fb44cfde50cab01708b4430ea91f11e42e7457ca4dadaa2a06612c4.jpg) +(b) Classifier Vocabulary Expansion Stage + +![](images/329d6aad43e3bd6950d5b340653c41a44f31103ba06fac1de88a87961f4b2783.jpg) +Fig.2: Our open vocabulary classification method includes three stages of operation. During the (a) training stage, we train our label and video encoders on closed set training labels. New class labels can be added to the vocabulary after training by employing the (b) classifier vocabulary expansion stage. During the (c) inference stage video embeddings are computed and matched with the label embeddings database to get the classification scores. + +# 3.1 LLMs for Semantic Enrichment of Label Embeddings + +Pretrained VLMs have strong open-vocabulary image classification performance owing to their large scale image-text pretraining. LLMs on the other hand are trained on even larger scale of text data with training objectives that enable a rich semantic understanding. As a result, compared to LLMs, VLMs have a more limited semantic understanding of natural language and have difficulty understanding concepts such as the relationship between different class labels. For instance, prior works have found that while VLMs are able to classify ImageNet images at a fine-grained level, their accuracy at higher levels of the class hierarchy is poor [46]. E.g., while they may recognize an image of a lion, they are unable to understand that a lion is also a feline, a mammal and an animal. However, an LLM is very effective in comprehending the hierarchical relationship between these labels, among many other capabilities, due to the large-scale training. Hence, in this work, we utilize an LLM to generate complementary information about the class labels that can be utilized to improve the vision-language alignment of VLM leading to better open-vocabulary classification performance. + +# Fixed LLM prompting. + +As a baseline, we first develop an LLM-based prompting method to adapt CLIP for open vocabulary video classification, extending the approach of Menon & Vondrick [27] for open vocabulary image classification. We design a prompt template (see implementation details in Section K of Supplementary material) for an encoder-decoder based LLM with the class name and a question asking it to generate useful features for visually distinguishing that class. The LLM output is then parsed into a list of textual descriptions, hereafter referred to as attributes. These attributes along with the class name is used to prompt the CLIP text encoder and the resultant text-embeddings are mean-pooled to obtain an attribute enriched text-embedding for the class. We mean-pool CLIP vision embedding of video frames to generate video embeddings and then perform open vocabulary classification by matching video embeddings with the text-embeddings of different classes. However, this simple baseline has a crucial weakness: it is not trainable, as there is a need to de-tokenize and tokenize the LLM text output to CLIP, which doesn't allow the flow of gradients. As a result, the process of generating class attributes by prompting an LLM cannot be improved by training on a labeled video dataset. To remedy this, we propose an end-to-end trainable architecture that integrates the LLM with the CLIP text encoder. + +End-to-end learnable LLM prompting. Our proposed architecture, summarized in Figure 3, incorporates a learnable prompting framework with the frozen LLM to generate the inputs to a frozen CLIP text-encoder. The learnable components on the text side are limited to $N$ learnable prefixes/vectors to the prompt template used for querying the LLM and a prompt transformer that transforms the sequence of tokens from the LLM to input soft prompts for the CLIP text-encoder. In our implementation we avoid the discrete and nondifferentiable operations like detokenization at the LLM decoder output and re-tokenization at the CLIP text-encoder input. This allows the prompt transformer to directly connect the LLM output semantic space to the CLIP input + +![](images/608e33bb894c29f04f60afc3af13fe110c4bb80a9a29785624ce708da6e654c3.jpg) +Fig. 3: Our end-to-end trainable system for open vocabulary video classification. The class labels are used by the LLM to generate useful class attributes for the CLIP text encoder which provides a visually aligned label embedding. The learnable input prompts to the LLM guide it to generate soft-attributes useful for video classification. Prompting transformer learns to map from the LLM output space to the CLIP input space. To add video understanding to CLIP's vision encoder we add additional spatiotemporal modeling layers. Details about each component are in Section 3. + +semantic space, ensuring that the entire text model is differentiable and therefore end-to-end learnable. We describe the details below. + +For each of the $N$ LLM prefixes, we first construct an input sequence to the LLM comprising of the prefix followed by a fixed prompt template that contains the label name $\ell$ from the input data sample (see Figure 3). The text input to the LLM is mapped using a tokenizer and the LLM's embedding layer into a sequence of tokens $I\in \mathbb{R}^{M\times d}$ , where $M$ and $d$ represent number of input tokens and the dimension of the embedding space respectively. All our $N$ learnable prefixes are $d$ -dimensional vectors. We use $P_{i}\in \mathbb{R}^{d}$ to represent the $i$ -th prefix. Each prefix is concatenated with the tokens of the prompt template, yielding a unified sequence of tokens $[P_i;I]\in \mathbb{R}^{(1 + M)\times d}$ . This combined sequence is then processed through the frozen encoder-decoder layers of the LLM. For each prefix, we run $KL$ decoding iterations of the LLM decoder to generate a sequence of + +$KL$ decoded tokens. These tokens represent useful class features in the LLM's semantic space. As the prompt template specifically asks LLM to output features as a list, we split the sequence of tokens evenly into $K$ subsequences of $L$ tokens each. Due to the nature of LLM prompt template, which prompts the LLM to output useful visual features, we refer to the subsequences of tokens as soft attributes, as they are sets of continuous vectors instead of discrete attributes in natural language. The $K$ soft attributes are then individually processed by the prompt transformer to generate $K$ soft prompts to the CLIP text encoder. Repeating this operation for each of the $N$ LLM prefixes, we get $NK$ soft prompts. Each soft-prompt is then concatenated with the tokenized label embedding and then processed by the frozen CLIP text-encoder to generate an attribute feature. All $NK$ resulting features are mean-pooled and normalized to obtain the final label embedding $f_{t}(\ell)$ , where $\ell$ is the label name. + +# 3.2 Regularized Parallel Temporal Modeling + +The simple image transformer of the CLIP vision encoder does not model the temporal dynamics of the video, which is essential to enhance the ability to recognize entities in videos. We enhance our vision encoder by adding a parallel temporal modeling branch to the last $T$ layers of the CLIP vision encoder as illustrated in Figure 3. We freeze the CLIP vision branch and train only the newly added temporal layers. Each block in the newly added branch consists of a spatial attention layer initialized from their corresponding CLIP weights, and a temporal attention layer that is randomly initialized. All frames of the video are processed independently by the CLIP vision backbone. Then, at the first temporal modeling block, the temporal token TMP is created by averaging the CLS tokens across frames, meanwhile learnable spatial and temporal positional embeddings are added to each patch token. The $t$ th temporal modeling block takes in as input a + +![](images/069d3ff05a91fc9e23e28564d2b1bfc60ef56395b8b489f0990ad159afb84780.jpg) +Fig. 4: Our Temporal Block projects frame patch tokens from the CLIP image encoder, fusing them with previous block's temporal branch tokens. The temporal token (TMP) incorporates all frames' CLS tokens. Divided Space-Time attention layers form the core of the block. Spatial attention layers are initialized from CLIP weights and regularized using stochastic weight averaging. + +weighted combination of the previous temporal modeling block and the corresponding layer in the CLIP backbone. Symbolically if $V_{t}^{\top}$ represents the patch tokens from the temporal block at the $t$ th block and $V_{t}$ represents the patch tokens for the corresponding CLIP layers, the patch tokens for the $t$ th block + +$V_{t}^{\mathsf{T}} = V_{t - 1}^{\mathsf{T}} + Proj_{\text{spatial}}(V_{s})$ , where $V_{s}$ is the corresponding token from the CLIP backbone. After that the tokens are processed by the divided space-time attention layer and spatial layer. The overall video embedding is generated at the end by mean pooling the final TMP token and the CLS tokens for each frame from the CLIP backbone. + +A key drawback of adding a deep parallel branch is that while it improves performance in the finetuned setting, it diminishes CLIP's zero-shot performance. In order to overcome this limitation in the zero-shot setting, we propose to use weight regularization on the spatial attention layers. The weight regularization operation tries to match predictions made by the current set of finetuned weights and a randomly weighted average of the current weights and the original CLIP weights. Symbolically, at each iteration we use weights $\theta$ which are a stochastic weighted average of the finetuned and the original frozen weights as given by: + +$$ +\theta = \alpha \theta_ {\mathrm {f t}} + (1 - \alpha) \theta_ {\mathrm {f r o z e n}}, \quad \mathrm {w h e r e} \quad \alpha \sim U (0, \lambda) +$$ + +We observe that an empirical value of $\lambda = 0.5$ works well across datasets. + +Relationship to prior works: Prior temporal modeling approaches split into two camps: improving finetuned performance or preserving CLIP's zero-shot abilities. STAN [24] is a state-of-the-art finetuning focused approach which adds a parallel temporal modeling branch which is able to leverage both high-level and low-level features from the CLIP vision encoder. The divided space-time attention in STAN's temporal block is inspired by the TimeSformer [2] architecture. We differ from STAN in order to achieve stronger open vocabulary performance in a number of ways. Firstly, unlike STAN we do not finetune the CLIP image encoder, only the added parallel branch. Additionally, we propose stochastic weight regularization to prevent the temporal branch from drifting too far from the original CLIP feature space. Our ablation experiments (Table 3) demonstrate that both these choices significantly improve our performance. Some form of weight regularization has been explored by methods such as Open-VCLIP [44]. However, Open V-CLIP takes in only 3 frames at a time, and its temporal modeling range is limited. Our approach is able to achieve the best of both approaches. + +# 3.3 Training objective + +We train our model on multi-label video datasets. To construct a batch, we first sample a set $\mathcal{B}$ of $B$ videos from the dataset. Each video $v$ is associated with a set of positive class labels which we denote $\mathcal{P}(v)$ . In addition to the positive labels, we augment the batch with random negative class labels for each video. To obtain the negatives, we first identify the set $\mathcal{P}_{\mathcal{B}} := \cup_{v \in \mathcal{B}} \mathcal{P}(v)$ of distinct classes among all the positive labels from all the videos in the batch. We then sample a random set $\mathcal{N}_{\mathcal{B}}$ of $4B - |\mathcal{P}_{\mathcal{B}}|$ classes from the rest of the class vocabulary of the dataset. For each video $v$ , we then choose all non-positive classes from $\mathcal{P}_{\mathcal{B}} \cup \mathcal{N}_{\mathcal{B}}$ as negative labels. We use $\mathcal{N}(v) := (\mathcal{P}_{\mathcal{B}} \cup \mathcal{N}_{\mathcal{B}}) \setminus \mathcal{P}(v)$ to denote the set of all negative labels for $v$ . The resulting positive and negative video-label pairs are + +then treated as training samples for binary classification. Each training sample is a label, video pair $(\ell, v)$ where $\ell$ represents the label name and $v$ the video. With this batch construction, the total number of training samples in a batch could be variable, but the total number of videos is fixed at $B$ and the total number of classes present among the samples is fixed at $4B$ . + +For each training data sample $(\ell, v)$ , the text and video encoders respectively generate a unit-norm text embedding $f_{t}(\ell) \in \mathbb{R}^{D}$ and a unit-norm video embedding $f_{v}(v) \in \mathbb{R}^{D}$ where $D$ is the common embedding dimension. The score $s(\ell, v)$ for this data sample is given by the inner product: + +$$ +s (\ell , v) = \left(f _ {t} (\ell)\right) ^ {\mathsf {T}} f _ {v} (v). \tag {1} +$$ + +The model is trained with a weighted binary cross entropy loss: + +![](images/9da06156b01cb9b87c553dd63e811e03d91b0fa9ea419ee607798789c29d7da3.jpg) +Fig.5: Our training objective applies binary cross-entropy to predicted Video-Label feature similarities sharpened by a temperature-scaled sigmoid. + +$$ +\mathcal {L} (\mathcal {B}) = - \sum_ {v \in \mathcal {B}} \left[ \sum_ {\ell \in \mathcal {P} (v)} \log p (\ell , v) + w \sum_ {\ell \in \mathcal {N} (v)} \log \left(1 - p (\ell , v)\right) \right] ^ {\text {e n e d b y a t}} \tag {2} +$$ + +where $p(\ell, v) \coloneqq \sigma\left(\frac{s(\ell, v)}{\tau}\right)$ , $\tau$ is a temperature parameter, $\sigma(.)$ is the sigmoid function and $w > 0$ is a weight hyperparameter. + +# 4 Experiments and Results + +# 4.1 Datasets + +In order to train our model on a wide range of concepts, we train it on a mix of YouTube8M [1] and Kinetics-400 [14]. YouTube-8M (YT-8M) is primarily labeled with entities, whereas Kinetics-400 (K400) is labeled only with actions. As YT-8M is composed of a random sample of the whole of YouTube, a significant portion of video its samples are from video game streams. Many of these videos are just tagged with the video game title as the label irrespective of the actual entities and actions present in the video. A significant portion of its label vocabulary is dedicated to video game titles (771 out of 3862 total) are dedicated to video game titles. We observe training instabilities due to this, and to fix this issue, we remove video game titles from the training vocabulary. We also remove some of the least frequently occurring labels during training to reach a training vocabulary of 2429 classes. This leads to stabilized training without further issues. Note that for closed set evaluation on YT-8M, we use the YouTube-8M Segments validation set, which contains human verified labels for 1000 classes. K400 on the other hand is widely used, and has a clean vocabulary and high + +quality human verified action labels. From the training perspective, the only limitation is the absence of entity (object, scenes etc.) labels. + +For evaluation, we use 3 open-vocabulary test datasets: TAO (Tracking Any Object) [3], ActivityNet [9] and RareAct [28]. TAO and ActivityNet are exclusively object and action focused datasets respectively. RareAct has object, action labels, and focuses on unusual object-action combination. TAO dataset was developed for evaluating object trackers, however analogous to how object detection datasets like MS-COCO are used for evaluating multi-label image classification methods such as DualCoOp [40], we transform TAO into a multi-label video classification dataset by ignoring localization annotations. ActivityNet often has multiple actions in a given video. For our evaluation each annotation from RareAct provides three labels: object, action and the (unusual) object-action combination, this provides an interesting test of the open vocabulary capabilities. Altogether, these five datasets provide a comprehensive evaluation of a model's closed and open vocabulary video classification performance. + +# 4.2Baselines + +As there are no prior works that explicitly address multilabel open vocabulary video classification, we extend the following methods from the zero shot image and action classification literature to our multi-label video classification setting for comparison. + +CoOp [52]: Short for Context Optimization, CoOp learns prompts for the CLIP text encoder as a lightweight adaptation technique for classification. + +DualCoOp [40]: Is an extension of CoOP to the multi-label setting where both positive and negative prompts are learnt for the CLIP text encoder. For a given label, prediction is based on whether the positive or negative prompted version has higher similarity with the image feature. + +LLM + CLIP (Frozen): This baseline was discussed in Section 3.1 and is illustrated in Figure 8 of Supplementary. + +ViFi-CLIP [33]: In this baseline, there is no temporal modeling, and both the CLIP image and text encoder are finetuned on the training dataset. + +As CoOp and ViFi-CLIP were developed for single label classification, to utilize them in our setting we replace their contrastive loss function with our multi-label classification loss. We also drop the region aggregation aspect of DualCoOp and only test the dual prompting architecture. We refer to these baselines as CoOp*, DualCoOp* and ViFi-CLIP* to reflect these differences. + +# 4.3 Metrics + +We report two sets of metrics to evaluate the performance of the models in the open vocabulary setting. Area Under Precision-Recall curve (AUPR) summarizes the overall classification performance across the entire precision-recall trade-off. Secondly, we report the Peak F1-Score for the model on each dataset, which is the F1-Score achieved by that model at the optimal threshold chosen by an oracle + +
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities+Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt6.326.243.844.29.5
CLIP with Prompt Templates6.830.546.045.911.4
CLIP with Fixed LLM Prompts6.930.650.246.811.5
Trainable Baseline Methods
CoOp2.717.835.028.83.5
DualCoOp8.323.947.133.07.6
ViFi-CLIP3.410.958.317.24.1
Ours
CLIP + Learnable LLM Prompts9.432.851.447.111.9
+ Temporal Modeling14.842.063.847.112.4
+ Synthetic Labels16.743.265.550.213.2
+ +Table 1: AUPR scores for all methods on all datasets. + +for that dataset. This metric captures the classification performance obtained by an open vocabulary classification method on a dataset, if the threshold alone could be tuned, e.g., by using a labeled validation set. + +# 4.4 Results + +AUPR scores for each method are presented in Table 1 and Peak F1 scores are presented in Table 2. Our learnable LLM prompting approach outperforms both frozen and trainable baselines across both closed-vocabulary and open-vocabulary classification tasks. Also notable is the enhancement in performance due to our temporal modeling approach (average gain of $\approx 5\%$ in AUPR). As previously mentioned in our discussion of the dataset, the training datasets YT-8M and K400 have certain limitations (in particular, YT-8M has few action labels and K400 has no object/entity labels). We present further improved results obtained by augmenting our training dataset with additional class labels obtained using a multimodal LLM (pipeline is detailed in Section I of the Supplementary Material). As seen from the last rows of Tables 1 and 2, this approach gives a further $\approx 2\%$ improvement, and is specially effective for actions. + +# 4.5 Improved score calibration across datasets + +In order to use a classifier in a truly open vocabulary setting in practice, the classification score needs to be well calibrated across different types of concepts. This would allow us to pre-select an optimal threshold to generate binary classification for all concepts. In Figure 6 we demonstrate the robustness offered by our method in setting a threshold that works for a wide range of concepts, and the advantage offered by our method over the frozen CLIP + LLM baseline. These figures plot the F1 scores for multi-label classification across thresholds provided by both methods with Figure 6a showing the numbers on the validation splits + +
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities + Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt14.934.244.647.117.6
CLIP with Prompt Templates20.836.849.250.120.5
CLIP with Fixed LLM Prompts21.637.350.251.419.8
Trainable Baseline Methods
CoOp5.825.543.935.510.5
DualCoOp16.233.249.040.515.0
ViFi-CLIP5.219.354.424.79.6
Ours
CLIP + Learnable LLM Prompts23.642.452.851.122.6
+ Temporal Modeling31.546.259.652.624.3
+ Synthetic Labels32.746.656.653.825.1
+ +Table 2: Peak F1 scores for all methods on all datasets. + +![](images/dafc88c6d70e2d588a4642d629d68729f841aad4b170287f80a1391bc5f6fb25.jpg) +(a) Closed Vocabulary Evaluation. + +![](images/dd25dc6544a7dc4be501e6c89cee168e4206348622993b4b60c00812bbc95075.jpg) +(b) Open Vocabulary Evaluation. +Fig. 6: F1-Scores at different thresholds for closed and open vocabulary evaluation datasets. Our end-to-end trained model can achieve better performance across datasets with a single threshold (labeled by gray vertical line) chosen on the supervised datasets. + +of the training datasets, and Figure 6b showing the numbers in the open vocabulary setting on two diverse datasets. A reasonable choice for a classification threshold would be to pick the threshold that maximizes the minimum validation F1-score for the datasets used in training. The resulting choice of thresholds for our method and for frozen CLIP are shown in Figure 6a. We see that for the baseline method the performance on TAO is poor, while that on ActivityNet is modest. However with our proposed method the F1 scores for both evaluation datasets are simultaneously high. This result empirically validates the versatility that is achieved by our learnable LLM-based prompting scheme over the baseline, which makes it possible to pre-select a threshold for using the solution as a black box classifier for all unseen concepts. + +# 4.6 Ablations + +We scientifically ablate each part of our framework, reporting ablation evaluations (AUPR) on TAO (Objects) and ActivityNet (Actions). + +
Open Vocabulary
# BlocksObjectsActions
565.448.1
465.550.2
264.649.1
162.347.7
+ +(a) Effect of Temporal Modeling + +
Open Vocabulary
Spatial Reg.BackboneObjectsActionsGeo. Mean
✓56.653.855.2
X62.132.544.9
X42.728.334.8
+ +(b) Effect of our Regularization +Table 3: Ablations of Temporal Modeling Branch Architecture and Regularization + +
Label EncoderOpen Vocabulary
LLMPTCLIPObjects Actions
Learnable Prompt✓✗25.2 8.5
Fixed Prompt✗✓62.5 41.3
Fixed Prompt✓✓64.9 48.1
Learnable Prompt✓✓65.5 50.2
✗✗✓60.1 34.5
+ +Table 4: Ablating the Label Encoder. PT $\rightarrow$ Prompt Transformer + +- Temporal Modeling Blocks: We find that using more Blocks improves performance, until around 4. (Table 3a) + +- Temporal Modeling Weight Regularization: We find that using our weight regularization strategy for the spatial layers of the temporal branch prevents the model from overfitting to certain concepts. We also validate our choice of keeping the CLIP vision backbone frozen and find that unfreezing the backbone leads to severe over-fitting. (Table 3b) + +- Label Encoder: We demonstrate the effectiveness of combining an LLM with learnable prompts and CLIP's text encoder as the label encoder. First, we show that removing CLIP text encoder leads to a significant drop in performance. This demonstrates that the visual-text alignment learned by CLIP is essential. Next we demonstrate the benefits of learnable prompts over fixed prompts. We also provide results with the LLM removed (Table 4). + +# 5 Conclusion + +We introduced the problem of open vocabulary multi-label video classification and proposed a solution leveraging LLMs and pre-trained VLMs. Through our extensive experiments we demonstrated strong performance on both actions and entities with a single model. Our proposed approach benefits from two key innovations. First, we proposed a method to prompt VLMs more effectively by utilizing the LLM output representations through a trainable prompt transformer and learnable LLM prompts. Second, we introduced a temporal modeling architecture and a regularized finetuning approach to improve the video understanding capability of the vision encoder while retaining strong open-vocabulary performance. Our ablations validate the efficacy of each of these contributions. + +# References + +1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark (2016) +2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: Proceedings of the International Conference on Machine Learning (ICML) (July 2021) +3. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 436-454. Springer (2020) +4. Desai, K., Kaul, G., Aysola, Z., Johnson, J.: Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431 (2021) +5. Fan, L., Krishnan, D., Isola, P., Katabi, D., Tian, Y.: Improving clip training with language rewrites. In: NeurIPS (2023) +6. Fang, H., Xiong, P., Xu, L., Chen, Y.: Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097 (2021) +7. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022) +8. Gupta, R., Roy, A., Christensen, C., Kim, S., Gerard, S., Cincebeaux, M., Divakaran, A., Grindal, T., Shah, M.: Class prototypes based contrastive learning for classifying multi-label and fine-grained educational videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19923-19933 (June 2023) +9. Heilbron, F.C., Niebles, J.C.: Collecting and annotating human activities in web videos. In: Proceedings of International Conference on Multimedia Retrieval. p. 377-384. ICMR '14, Association for Computing Machinery, New York, NY, USA (2014). https://doi.org/10.1145/2578726.2578775, https://doi.org/10.1145/2578726.2578775 +0. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9 +1. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html +2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +3. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. In: International Conference on Machine Learning (2023) +4. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., Suleyman, M., Zisserman, A.: The kinetics human action video dataset (2017) + +15. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=e2TBb5y0yFf +16. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243 +17. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 12888-12900. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/1i22n.html +18. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021) +19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10965-10975 (June 2022) +20. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353 +21. Lin, W., Karlinsky, L., Shvetsova, N., Possegger, H., Kozinski, M., Panda, R., Feris, R., Kuehne, H., Bischof, H.: Match, expand and improve: Unsupervised finetuning for zero-shot action recognition with language knowledge. In: ICCV (2023) +22. Lin, X., Petroni, F., Bertasius, G., Rohrbach, M., Chang, S.F., Torresani, L.: Learning to recognize procedural activities with distant supervision. arXiv preprint arXiv:2201.10990 (2022) +23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems. Curran Associates, Inc. (2023) +24. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023) +25. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. AI Open (2023). https://doi.org/https://doi.org/10.1016/j.aiopen.2023.08.012, https://www.sciencedirect.com/science/article/pii/S2666651023000141 +26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022) + +27. Menon, S., Vondrick, C.: Visual classification via description from large language models. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=j1AjNL8z5cs +28. Miech, A., Alayrac, J.B., Laptev, I., Sivic, J., Zisserman, A.: Rareact: A video dataset of unusual interactions. arxiv:2008.01018 (2020) +29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 728-755. Springer Nature Switzerland, Cham (2022) +30. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: European Conference on Computer Vision. pp. 1-18. Springer (2022) +31. Pratt, S., Covert, I., Liu, R., Farhadi, A.: What does a platypus look like? generating customized prompts for zero-shot image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15691-15701 (2023) +32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +33. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +34. Roth, K., Kim, J.M., Koepke, A.S., Vinyals, O., Schmid, C., Akata, Z.: Waffling around for performance: Visual classification with random words and broad concepts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15746-15757 (October 2023) +35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018) +38. Shvetsova, N., Kukleva, A., Hong, X., Rupprecht, C., Schiele, B., Kuehne, H.: HowtocapTION: Prompting llms to transform video annotations at scale (2023) +39. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022) +40. Sun, X., Hu, P., Saenko, K.: Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems 35, 30569-30582 (2022) +41. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Commun. ACM + +59(2), 64-73 (jan 2016). https://doi.org/10.1145/2812802, https://doi.org/10.1145/2812802 +42. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23034–23044 (2023) +43. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-VCLIP: Transforming CLIP to an open-vocabulary video model via interpolated weight optimization. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 36978-36989. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/weng23b.htm1 +44. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In: ICML (2023) +45. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022) +46. Xu, Z., Zhu, Y., Deng, T., Mittal, A., Chen, Y., Wang, M., Favaro, P., Tighe, J., Modolo, D.: Challenges of zero-shot recognition with vision-language models: Granularity and correctness (2023) +47. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022) +48. Yang, Y., Panagopoulou, A., Zhou, S., Jin, D., Callison-Burch, C., Yatskar, M.: Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19187-19197 (June 2023) +49. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021) +50. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022) +51. Zhao, Y., Misra, I., Krahenbuhl, P., Girdhar, R.: Learning video representations from large language models. In: CVPR (2023) +52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2639-2650 (October 2023) \ No newline at end of file diff --git a/2024/Open Vocabulary Multi-Label Video Classification/images.zip b/2024/Open Vocabulary Multi-Label Video Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c805557643ef25cc04a95d5a1b5d157b0bab2780 --- /dev/null +++ b/2024/Open Vocabulary Multi-Label Video Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc2f10f67760a822e5889a41b2111d68e10ccb13710306ecae9a8ed0985778e8 +size 394335 diff --git a/2024/Open Vocabulary Multi-Label Video Classification/layout.json b/2024/Open Vocabulary Multi-Label Video Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d20d7ebd11dce291be5db1ed8ec87c515df988f4 --- /dev/null +++ b/2024/Open Vocabulary Multi-Label Video Classification/layout.json @@ -0,0 +1,8448 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 179, + 111, + 434, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 111, + 434, + 145 + ], + "spans": [ + { + "bbox": [ + 179, + 111, + 434, + 145 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 466, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 466, + 179 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 466, + 179 + ], + "type": "text", + "content": "Rohit Gupta1, Mamshad Nayeem Rizve2, Jayakrishnan Unnikrishnan2," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "spans": [ + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "type": "text", + "content": "Ashish Tawari" + }, + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "type": "text", + "content": ", Son Tran" + }, + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "type": "text", + "content": ", Mubarak Shah" + }, + { + "bbox": [ + 205, + 180, + 408, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 193, + 391, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 193, + 391, + 204 + ], + "spans": [ + { + "bbox": [ + 222, + 193, + 391, + 204 + ], + "type": "text", + "content": "Benjamin Yao" + }, + { + "bbox": [ + 222, + 193, + 391, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 222, + 193, + 391, + 204 + ], + "type": "text", + "content": ", and Trishul Chilimbi" + }, + { + "bbox": [ + 222, + 193, + 391, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 213, + 455, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 213, + 455, + 224 + ], + "spans": [ + { + "bbox": [ + 157, + 213, + 455, + 224 + ], + "type": "text", + "content": "1 Center for Research in Computer Vision, University of Central Florida" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 226, + 397, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 226, + 397, + 236 + ], + "spans": [ + { + "bbox": [ + 216, + 226, + 397, + 236 + ], + "type": "text", + "content": "rohit.gupta@ucf.edu, shah@crv.ucf.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 285, + 236, + 328, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 236, + 328, + 246 + ], + "spans": [ + { + "bbox": [ + 285, + 236, + 328, + 246 + ], + "type": "text", + "content": "2 Amazon" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 247, + 389, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 247, + 389, + 258 + ], + "spans": [ + { + "bbox": [ + 224, + 247, + 389, + 258 + ], + "type": "text", + "content": "{jayunn,atakari,sontran}@amazon.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 285, + 453, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 285, + 453, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 285, + 453, + 504 + ], + "type": "text", + "content": "Abstract. Pre-trained vision-language models (VLMs) have enabled significant progress in open vocabulary computer vision tasks such as image classification, object detection and image segmentation. Some recent works have focused on extending VLMs to open vocabulary single label action classification in videos. However, previous methods fall short in holistic video understanding which requires the ability to simultaneously recognize multiple actions and entities e.g., objects in the video in an open vocabulary setting. We formulate this problem as open vocabulary multi-label video classification and propose a method to adapt a pre-trained VLM such as CLIP to solve this task. We leverage large language models (LLMs) to provide semantic guidance to the VLM about class labels to improve its open vocabulary performance with two key contributions. First, we propose an end-to-end trainable architecture that learns to prompt an LLM to generate soft attributes for the CLIP text-encoder to enable it to recognize novel classes. Second, we integrate a temporal modeling module into CLIP's vision encoder to effectively model the spatio-temporal dynamics of video concepts as well as propose a novel regularized finetuning technique to ensure strong open vocabulary classification performance in the video domain. Our extensive experimentation showcases the efficacy of our approach on multiple benchmark datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 514, + 441, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 514, + 441, + 526 + ], + "spans": [ + { + "bbox": [ + 160, + 514, + 441, + 526 + ], + "type": "text", + "content": "Keywords: Open Vocabulary " + }, + { + "bbox": [ + 160, + 514, + 441, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 441, + 526 + ], + "type": "text", + "content": " Multi-Modal " + }, + { + "bbox": [ + 160, + 514, + 441, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 441, + 526 + ], + "type": "text", + "content": " Video Understanding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 545, + 229, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 229, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 229, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "Video classification is a critical challenge in computer vision. Its goal involves recognizing a diverse array of concepts depicted in a video, which may include primarily static entities such as objects and scenes, as well as dynamic actions. In the classic setting, the vocabulary of all possible classes of interest is known in advance, and the model is trained in a supervised manner using a labeled dataset. The labor-intensive process of manual annotation often results in video datasets that are narrowly focused, such as those limited to specific sports or simple activities, which restricts the breadth of concepts models can learn. As video" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 114, + 455, + 276 + ], + "blocks": [ + { + "bbox": [ + 162, + 114, + 455, + 276 + ], + "lines": [ + { + "bbox": [ + 162, + 114, + 455, + 276 + ], + "spans": [ + { + "bbox": [ + 162, + 114, + 455, + 276 + ], + "type": "image", + "image_path": "fa6d656ede06a3b9bd52a63c9f82f23519323f4871da1b926beb4fb11ac58ace.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 283, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 350 + ], + "type": "text", + "content": "Fig. 1: Our task: recognize multiple open-vocabulary classes in videos at inference from an open vocabulary, including entities (in blue) such as objects and scenes, and actions (in red). Prior zero-shot image classification approaches (e.g. CLIP) can label the salient entity in each frame (left), while zero-shot action-recognition approaches (e.g. ViFi-CLIP) (middle) can classify video-level action. Our method (right) can recognize all action or entity classes present in the video." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "type": "text", + "content": "applications are becoming widespread, there is an increasing need for developing video models that can recognize a broad range of concepts. This necessitates the development of open-vocabulary approaches for video classification that can recognize a diverse range of video concepts, including those that were not part of the class vocabulary present in the labeled training dataset. In this work, to solve this problem, we aim to leverage recent advances in vision-language models (VLM), which are trained to align visual and textual data and large language models (LLM), which have been demonstrated to possess a rich understanding of the world due to large scale pre-training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "VLMs that are trained for image-language alignment on image-text pairs at large-scale exhibit remarkable zero-shot visual recognition performance across a diverse set of tasks [11, 19, 29, 32]. For classification tasks, the pretrained text encoder acts as a \"label encoder\" for the class labels, mapping them into the same representation space as the visual embedding of the image, which allows for the image to be classified by ranking the labels in order of representation similarity. Since VLMs are primarily trained to rank, picking the top match usually results in excellent performance on single label zero-shot classification. However, merely ranking the labels is insufficient to achieve open-vocabulary multi-label classification, and as a result, specialized methods [40] have been developed for multi-label image classification using VLMs. VLM based zero-shot image classification has been further improved by LLM-based prompting, in which LLMs are used to generate descriptive class definitions which are then used as prompts to the VLM text encoder during inference [27, 31]. Other works have adapted VLMs for zero-shot single-label action classification in videos [42, 43]. A typical video," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "however, contains multiple concepts including objects, actions, scenes, events, etc. Our goal is to develop a video classifier that can simultaneously recognize the presence of these different concepts from any vocabulary specified at inference. We refer to this task as open-vocabulary multi-label video classification. An illustration of the subtle differences between the prior tasks and our proposed task is presented in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 368 + ], + "type": "text", + "content": "Open-vocabulary multi-label video classification presents two unique challenges. First, unlike the single label open-vocabulary setting, in the multi-label setting, identifying relevant concepts in a video based on VLM similarity score performs poorly, as VLM similarity scores for different types of concepts (e.g., actions, objects) often fall within different ranges, even with LLM-guided prompting (see Section 4.5). Addressing this issue requires end-to-end finetuning of the VLM with LLM guidance on datasets with diverse video concepts. However, performing end-to-end finetuning while simultaneously incorporating the benefits of LLM-guided prompting, presents additional technical challenges, particularly in backpropagating gradients through the VLMs text encoder's tokenizer. Second, adapting pretrained image-language models to recognize video concepts while retaining strong zero-shot performance is challenging, as the image-language pretraining datasets are orders of magnitude larger than the largest labeled video datasets. VLM adaptation tends to overfit to the video data easily, thus losing their generalization capability [42]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 369, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 583 + ], + "type": "text", + "content": "In this work, we attempt to address these challenges for open vocabulary multi-label video classification. First, to simultaneously recognize multiple video concepts in open-domain, we finetune the VLM in an end-to-end manner with LLM guidance to ensure proper ranking between different types of video concepts. Particularly, we extend the LLM-guided prompting approach for open-vocabulary classification by introducing learnable prefixes for prompting the LLM. To directly utilize the LLM output representations in the VLM's text encoder for end-to-end training, we propose a prompt transformer. This prompt transformer, in conjunction with the learnable prefixes, not only facilitates the integration of the LLM's world knowledge into the VLM's text representation but also aids in mitigating the discrepancy between VLM scores for various types of concepts. Second, for the effective recognition of temporally varying video concepts, we integrate an auxiliary temporal modeling branch with the VLM's vision encoder [24] and introduce a novel regularization penalty to retain the VLM's zero-shot performance during the adaptation to videos. Through extensive experimentation and analysis we demonstrate that our proposed approach significantly outperforms baseline solutions obtained by extending prior work on single label open vocabulary image and video classification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 584, + 444, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 444, + 595 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 444, + 595 + ], + "type": "text", + "content": "In summary, in this work we make three major technical contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 605, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 131, + 605, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 639 + ], + "type": "text", + "content": "- A novel end-to-end trainable approach to learn a strong label encoder for open vocabulary multi-label video classification by adapting an LLM to learn to prompt a VLM's text encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 664 + ], + "type": "text", + "content": "- An approach for enhancing pre-trained VLM image encoders with temporal modeling capability while also retaining strong open-vocabulary performance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": "- Defining a new open vocabulary multi-label video classification benchmark using 5 datasets (2 closed, 3 open vocabulary) and benchmarking our approach against 6 strong baselines, significantly outperforming them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 243, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 243, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 243, + 185 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 293 + ], + "type": "text", + "content": "Vision-Language Representation Learning: Recently, image-language models [12, 18, 32, 39, 49, 50] have drawn huge attention because of their effectiveness in learning generic visual representations that are transferable to several downstream tasks like classification, retrieval, etc. This success can partly be attributed to the large-scale image-text datasets [4, 35-37, 41]. However, this is not the case for the video data. Therefore, to perform video-language pretraining, most recent works [6, 7, 26, 30, 47] bootstrap from a pretrained image-language model and adapt them to video representation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 376 + ], + "type": "text", + "content": "Open-Vocabulary Classification: Vision-language models do not perform optimally in visual recognition tasks out of the box. To improve the zero-shot open-vocabulary visual recognition capability of the vision-language models, the recent works either resort to regularized finetuning [43, 45] or prompt learning [40, 42, 52]. However, these works either focus on solving single label tasks or only perform adaptation for image recognition tasks. In contrast to these works, we focus on open-vocabulary video classification task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 487, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 487, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 487, + 509 + ], + "type": "text", + "content": "Large Language Models in Vision: Recently, Large Language Models (LLM) have been utilized in vision tasks primarily because of their excellent in-context and zero-shot learning performance especially in commonsense reasoning tasks [15]. To be particular, LLMs have been used to rewrite noisy ASR text [8,38], to generate classification labels by mapping textual description to a predefined task list [22], extracting verbs from textual descriptions to improve action understanding [21]. LLMs have also been used to initialize textual encoders for vision language encoder [51]. In some recent works, vision representations have been aligned with LLM input space to solve multitude of vision-language tasks [17,23]. These multi-modal models have also been used to generate captions for images which are then utilized to train better vision-language models [5]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "Prompting Vision-Language Models with LLMs: Some prior works have leveraged LLMs to augment the open vocabulary understanding capabilities of Vision-Language models. Classification By Description [27], CuPL [31], LLM guided concept bottleneck [48] all utilize LLM generated class descriptors to prompt CLIP to improve image classification, and make it more interpretable. WaffleCLIP [34] demonstrated that LLM generated higher level concepts to describe datasets can also improve classification. LLM based approaches have also been extended to object detection [13] and point cloud understanding [53]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Parameter efficient fine-tuning of LLMs: As full parameter finetuning of LLMs is not practical, quite a few parameter efficient finetuning methods have been developed. The lightweight Learnable Soft prompts [16] is the approach we utilize to guide the LLM in our method. Other approaches such as LoRA [10], Prefix-Tuning [20] and P-Tuning [25] could also be adopted in principle." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 479, + 186 + ], + "type": "text", + "content": "Our proposal to boost the open vocabulary multi-label video classification capabilities of CLIP consists of two key parts: first, an end-to-end trainable label" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 282, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 282, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 282, + 417 + ], + "type": "text", + "content": "encoder leveraging an LLM for strong open vocabulary capabilities, and the CLIP text encoder for visual alignment. The label encoder consists of a frozen LLM guided with learnable prefixes, whose outputs are mapped to the frozen CLIP text encoder using a learnable prompting transformer. Second, we enhance the CLIP image encoder's capabilities for understanding temporal dynamics. This is achieved by using a lightweight temporal modeling branch to enhance the CLIP image encoder. The details of our approach are illustrated in Figure 3. Each part of our method is discussed in detail in this section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 281, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 281, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 281, + 506 + ], + "type": "text", + "content": "An overview of our approach is provided in Fig. 2. Our method has three broad stages. Firstly, during the training phase, both the label encoder and video encoder are trained simultaneously, as shown in (a)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 282, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 282, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 282, + 629 + ], + "type": "text", + "content": "Secondly, during the classifier vocabulary expansion stage, embeddings for class labels are calculated and saved into a label embedding database, as shown in (b). This vocabulary can be extended at any point after training, thus allowing our method to be used in the open vocabulary setting. Finally, during the inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "stage, video features are computed and compared against label embeddings from the database. As the label embeddings are pre-computed, the computational overhead during inference over standard CLIP models is minimal." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 289, + 217, + 478, + 310 + ], + "blocks": [ + { + "bbox": [ + 289, + 217, + 478, + 310 + ], + "lines": [ + { + "bbox": [ + 289, + 217, + 478, + 310 + ], + "spans": [ + { + "bbox": [ + 289, + 217, + 478, + 310 + ], + "type": "image", + "image_path": "37fd6f002647e864d9f3e79b261b57ffc0482b5da92c96721421e1e836e18748.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 312, + 392, + 319 + ], + "lines": [ + { + "bbox": [ + 342, + 312, + 392, + 319 + ], + "spans": [ + { + "bbox": [ + 342, + 312, + 392, + 319 + ], + "type": "text", + "content": "(a) Training Stage" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 289, + 324, + 476, + 384 + ], + "blocks": [ + { + "bbox": [ + 289, + 324, + 476, + 384 + ], + "lines": [ + { + "bbox": [ + 289, + 324, + 476, + 384 + ], + "spans": [ + { + "bbox": [ + 289, + 324, + 476, + 384 + ], + "type": "image", + "image_path": "4ef2898a2fb44cfde50cab01708b4430ea91f11e42e7457ca4dadaa2a06612c4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 384, + 436, + 392 + ], + "lines": [ + { + "bbox": [ + 323, + 384, + 436, + 392 + ], + "spans": [ + { + "bbox": [ + 323, + 384, + 436, + 392 + ], + "type": "text", + "content": "(b) Classifier Vocabulary Expansion Stage" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 296, + 393, + 473, + 466 + ], + "blocks": [ + { + "bbox": [ + 296, + 393, + 473, + 466 + ], + "lines": [ + { + "bbox": [ + 296, + 393, + 473, + 466 + ], + "spans": [ + { + "bbox": [ + 296, + 393, + 473, + 466 + ], + "type": "image", + "image_path": "329d6aad43e3bd6950d5b340653c41a44f31103ba06fac1de88a87961f4b2783.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 478, + 482, + 588 + ], + "lines": [ + { + "bbox": [ + 286, + 478, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 286, + 478, + 482, + 588 + ], + "type": "text", + "content": "Fig.2: Our open vocabulary classification method includes three stages of operation. During the (a) training stage, we train our label and video encoders on closed set training labels. New class labels can be added to the vocabulary after training by employing the (b) classifier vocabulary expansion stage. During the (c) inference stage video embeddings are computed and matched with the label embeddings database to get the classification scores." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 428, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 428, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 428, + 128 + ], + "type": "text", + "content": "3.1 LLMs for Semantic Enrichment of Label Embeddings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 317 + ], + "type": "text", + "content": "Pretrained VLMs have strong open-vocabulary image classification performance owing to their large scale image-text pretraining. LLMs on the other hand are trained on even larger scale of text data with training objectives that enable a rich semantic understanding. As a result, compared to LLMs, VLMs have a more limited semantic understanding of natural language and have difficulty understanding concepts such as the relationship between different class labels. For instance, prior works have found that while VLMs are able to classify ImageNet images at a fine-grained level, their accuracy at higher levels of the class hierarchy is poor [46]. E.g., while they may recognize an image of a lion, they are unable to understand that a lion is also a feline, a mammal and an animal. However, an LLM is very effective in comprehending the hierarchical relationship between these labels, among many other capabilities, due to the large-scale training. Hence, in this work, we utilize an LLM to generate complementary information about the class labels that can be utilized to improve the vision-language alignment of VLM leading to better open-vocabulary classification performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 318, + 251, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 251, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 251, + 330 + ], + "type": "text", + "content": "Fixed LLM prompting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 544 + ], + "type": "text", + "content": "As a baseline, we first develop an LLM-based prompting method to adapt CLIP for open vocabulary video classification, extending the approach of Menon & Vondrick [27] for open vocabulary image classification. We design a prompt template (see implementation details in Section K of Supplementary material) for an encoder-decoder based LLM with the class name and a question asking it to generate useful features for visually distinguishing that class. The LLM output is then parsed into a list of textual descriptions, hereafter referred to as attributes. These attributes along with the class name is used to prompt the CLIP text encoder and the resultant text-embeddings are mean-pooled to obtain an attribute enriched text-embedding for the class. We mean-pool CLIP vision embedding of video frames to generate video embeddings and then perform open vocabulary classification by matching video embeddings with the text-embeddings of different classes. However, this simple baseline has a crucial weakness: it is not trainable, as there is a need to de-tokenize and tokenize the LLM text output to CLIP, which doesn't allow the flow of gradients. As a result, the process of generating class attributes by prompting an LLM cannot be improved by training on a labeled video dataset. To remedy this, we propose an end-to-end trainable architecture that integrates the LLM with the CLIP text encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": "End-to-end learnable LLM prompting. Our proposed architecture, summarized in Figure 3, incorporates a learnable prompting framework with the frozen LLM to generate the inputs to a frozen CLIP text-encoder. The learnable components on the text side are limited to " + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 546, + 481, + 665 + ], + "type": "text", + "content": " learnable prefixes/vectors to the prompt template used for querying the LLM and a prompt transformer that transforms the sequence of tokens from the LLM to input soft prompts for the CLIP text-encoder. In our implementation we avoid the discrete and nondifferentiable operations like detokenization at the LLM decoder output and re-tokenization at the CLIP text-encoder input. This allows the prompt transformer to directly connect the LLM output semantic space to the CLIP input" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 119, + 449, + 386 + ], + "blocks": [ + { + "bbox": [ + 170, + 119, + 449, + 386 + ], + "lines": [ + { + "bbox": [ + 170, + 119, + 449, + 386 + ], + "spans": [ + { + "bbox": [ + 170, + 119, + 449, + 386 + ], + "type": "image", + "image_path": "608e33bb894c29f04f60afc3af13fe110c4bb80a9a29785624ce708da6e654c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 405, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 483 + ], + "type": "text", + "content": "Fig. 3: Our end-to-end trainable system for open vocabulary video classification. The class labels are used by the LLM to generate useful class attributes for the CLIP text encoder which provides a visually aligned label embedding. The learnable input prompts to the LLM guide it to generate soft-attributes useful for video classification. Prompting transformer learns to map from the LLM output space to the CLIP input space. To add video understanding to CLIP's vision encoder we add additional spatiotemporal modeling layers. Details about each component are in Section 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 508, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 532 + ], + "type": "text", + "content": "semantic space, ensuring that the entire text model is differentiable and therefore end-to-end learnable. We describe the details below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "For each of the " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " LLM prefixes, we first construct an input sequence to the LLM comprising of the prefix followed by a fixed prompt template that contains the label name " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " from the input data sample (see Figure 3). The text input to the LLM is mapped using a tokenizer and the LLM's embedding layer into a sequence of tokens " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "I\\in \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " represent number of input tokens and the dimension of the embedding space respectively. All our " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " learnable prefixes are " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "-dimensional vectors. We use " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{i}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " to represent the " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "-th prefix. Each prefix is concatenated with the tokens of the prompt template, yielding a unified sequence of tokens " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "[P_i;I]\\in \\mathbb{R}^{(1 + M)\\times d}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ". This combined sequence is then processed through the frozen encoder-decoder layers of the LLM. For each prefix, we run " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " decoding iterations of the LLM decoder to generate a sequence of" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " decoded tokens. These tokens represent useful class features in the LLM's semantic space. As the prompt template specifically asks LLM to output features as a list, we split the sequence of tokens evenly into " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " subsequences of " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " tokens each. Due to the nature of LLM prompt template, which prompts the LLM to output useful visual features, we refer to the subsequences of tokens as soft attributes, as they are sets of continuous vectors instead of discrete attributes in natural language. The " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " soft attributes are then individually processed by the prompt transformer to generate " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " soft prompts to the CLIP text encoder. Repeating this operation for each of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " LLM prefixes, we get " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "NK" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " soft prompts. Each soft-prompt is then concatenated with the tokenized label embedding and then processed by the frozen CLIP text-encoder to generate an attribute feature. All " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "NK" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " resulting features are mean-pooled and normalized to obtain the final label embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "f_{t}(\\ell)" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " is the label name." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 288, + 366, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 366, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 366, + 302 + ], + "type": "text", + "content": "3.2 Regularized Parallel Temporal Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 306, + 299, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 299, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 299, + 617 + ], + "type": "text", + "content": "The simple image transformer of the CLIP vision encoder does not model the temporal dynamics of the video, which is essential to enhance the ability to recognize entities in videos. We enhance our vision encoder by adding a parallel temporal modeling branch to the last " + }, + { + "bbox": [ + 130, + 306, + 299, + 617 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 306, + 299, + 617 + ], + "type": "text", + "content": " layers of the CLIP vision encoder as illustrated in Figure 3. We freeze the CLIP vision branch and train only the newly added temporal layers. Each block in the newly added branch consists of a spatial attention layer initialized from their corresponding CLIP weights, and a temporal attention layer that is randomly initialized. All frames of the video are processed independently by the CLIP vision backbone. Then, at the first temporal modeling block, the temporal token TMP is created by averaging the CLS tokens across frames, meanwhile learnable spatial and temporal positional embeddings are added to each patch token. The " + }, + { + "bbox": [ + 130, + 306, + 299, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 306, + 299, + 617 + ], + "type": "text", + "content": "th temporal modeling block takes in as input a" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 331, + 462, + 482 + ], + "blocks": [ + { + "bbox": [ + 309, + 331, + 462, + 482 + ], + "lines": [ + { + "bbox": [ + 309, + 331, + 462, + 482 + ], + "spans": [ + { + "bbox": [ + 309, + 331, + 462, + 482 + ], + "type": "image", + "image_path": "069d3ff05a91fc9e23e28564d2b1bfc60ef56395b8b489f0990ad159afb84780.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 490, + 482, + 590 + ], + "lines": [ + { + "bbox": [ + 304, + 490, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 482, + 590 + ], + "type": "text", + "content": "Fig. 4: Our Temporal Block projects frame patch tokens from the CLIP image encoder, fusing them with previous block's temporal branch tokens. The temporal token (TMP) incorporates all frames' CLS tokens. Divided Space-Time attention layers form the core of the block. Spatial attention layers are initialized from CLIP weights and regularized using stochastic weight averaging." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "weighted combination of the previous temporal modeling block and the corresponding layer in the CLIP backbone. Symbolically if " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "V_{t}^{\\top}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": " represents the patch tokens from the temporal block at the " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "th block and " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": " represents the patch tokens for the corresponding CLIP layers, the patch tokens for the " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "th block" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "V_{t}^{\\mathsf{T}} = V_{t - 1}^{\\mathsf{T}} + Proj_{\\text{spatial}}(V_{s})" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "V_{s}" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": " is the corresponding token from the CLIP backbone. After that the tokens are processed by the divided space-time attention layer and spatial layer. The overall video embedding is generated at the end by mean pooling the final TMP token and the CLS tokens for each frame from the CLIP backbone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 272 + ], + "type": "text", + "content": "A key drawback of adding a deep parallel branch is that while it improves performance in the finetuned setting, it diminishes CLIP's zero-shot performance. In order to overcome this limitation in the zero-shot setting, we propose to use weight regularization on the spatial attention layers. The weight regularization operation tries to match predictions made by the current set of finetuned weights and a randomly weighted average of the current weights and the original CLIP weights. Symbolically, at each iteration we use weights " + }, + { + "bbox": [ + 130, + 176, + 480, + 272 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 176, + 480, + 272 + ], + "type": "text", + "content": " which are a stochastic weighted average of the finetuned and the original frozen weights as given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 285, + 411, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 285, + 411, + 298 + ], + "spans": [ + { + "bbox": [ + 203, + 285, + 411, + 298 + ], + "type": "interline_equation", + "content": "\\theta = \\alpha \\theta_ {\\mathrm {f t}} + (1 - \\alpha) \\theta_ {\\mathrm {f r o z e n}}, \\quad \\mathrm {w h e r e} \\quad \\alpha \\sim U (0, \\lambda)", + "image_path": "b97c3de8c28adccf5ae034ca55c7521735f4cddb0c2c61f247553fb857ef4036.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 303, + 454, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 303, + 454, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 454, + 316 + ], + "type": "text", + "content": "We observe that an empirical value of " + }, + { + "bbox": [ + 131, + 303, + 454, + 316 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 131, + 303, + 454, + 316 + ], + "type": "text", + "content": " works well across datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 327, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 507 + ], + "type": "text", + "content": "Relationship to prior works: Prior temporal modeling approaches split into two camps: improving finetuned performance or preserving CLIP's zero-shot abilities. STAN [24] is a state-of-the-art finetuning focused approach which adds a parallel temporal modeling branch which is able to leverage both high-level and low-level features from the CLIP vision encoder. The divided space-time attention in STAN's temporal block is inspired by the TimeSformer [2] architecture. We differ from STAN in order to achieve stronger open vocabulary performance in a number of ways. Firstly, unlike STAN we do not finetune the CLIP image encoder, only the added parallel branch. Additionally, we propose stochastic weight regularization to prevent the temporal branch from drifting too far from the original CLIP feature space. Our ablation experiments (Table 3) demonstrate that both these choices significantly improve our performance. Some form of weight regularization has been explored by methods such as Open-VCLIP [44]. However, Open V-CLIP takes in only 3 frames at a time, and its temporal modeling range is limited. Our approach is able to achieve the best of both approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 525, + 253, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 253, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 253, + 537 + ], + "type": "text", + "content": "3.3 Training objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We train our model on multi-label video datasets. To construct a batch, we first sample a set " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " videos from the dataset. Each video " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is associated with a set of positive class labels which we denote " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(v)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". In addition to the positive labels, we augment the batch with random negative class labels for each video. To obtain the negatives, we first identify the set " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}} := \\cup_{v \\in \\mathcal{B}} \\mathcal{P}(v)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " of distinct classes among all the positive labels from all the videos in the batch. We then sample a random set " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\mathcal{B}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "4B - |\\mathcal{P}_{\\mathcal{B}}|" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " classes from the rest of the class vocabulary of the dataset. For each video " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", we then choose all non-positive classes from " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}} \\cup \\mathcal{N}_{\\mathcal{B}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " as negative labels. We use " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(v) := (\\mathcal{P}_{\\mathcal{B}} \\cup \\mathcal{N}_{\\mathcal{B}}) \\setminus \\mathcal{P}(v)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " to denote the set of all negative labels for " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". The resulting positive and negative video-label pairs are" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "then treated as training samples for binary classification. Each training sample is a label, video pair " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "(\\ell, v)" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " represents the label name and " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " the video. With this batch construction, the total number of training samples in a batch could be variable, but the total number of videos is fixed at " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " and the total number of classes present among the samples is fixed at " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "4B" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "text", + "content": "For each training data sample " + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "inline_equation", + "content": "(\\ell, v)" + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "text", + "content": ", the text and video encoders respectively generate a unit-norm text embedding " + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "inline_equation", + "content": "f_{t}(\\ell) \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "text", + "content": " and a unit-norm video embedding " + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "inline_equation", + "content": "f_{v}(v) \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "text", + "content": " is the common embedding dimension. The score " + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "inline_equation", + "content": "s(\\ell, v)" + }, + { + "bbox": [ + 130, + 176, + 317, + 273 + ], + "type": "text", + "content": " for this data sample is given by the inner product:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 280, + 315, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 280, + 315, + 294 + ], + "spans": [ + { + "bbox": [ + 173, + 280, + 315, + 294 + ], + "type": "interline_equation", + "content": "s (\\ell , v) = \\left(f _ {t} (\\ell)\\right) ^ {\\mathsf {T}} f _ {v} (v). \\tag {1}", + "image_path": "96e75401798ffc0cf4a0ee72e0c2ae3f55985be67fac4d1ac5377d4c96c1b4cf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 315, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 315, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 315, + 326 + ], + "type": "text", + "content": "The model is trained with a weighted binary cross entropy loss:" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 201, + 471, + 291 + ], + "blocks": [ + { + "bbox": [ + 322, + 201, + 471, + 291 + ], + "lines": [ + { + "bbox": [ + 322, + 201, + 471, + 291 + ], + "spans": [ + { + "bbox": [ + 322, + 201, + 471, + 291 + ], + "type": "image", + "image_path": "9da06156b01cb9b87c553dd63e811e03d91b0fa9ea419ee607798789c29d7da3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 300, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 321, + 300, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 321, + 300, + 482, + 344 + ], + "type": "text", + "content": "Fig.5: Our training objective applies binary cross-entropy to predicted Video-Label feature similarities sharpened by a temperature-scaled sigmoid." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 334, + 369, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 334, + 369, + 385 + ], + "spans": [ + { + "bbox": [ + 125, + 334, + 369, + 385 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathcal {B}) = - \\sum_ {v \\in \\mathcal {B}} \\left[ \\sum_ {\\ell \\in \\mathcal {P} (v)} \\log p (\\ell , v) + w \\sum_ {\\ell \\in \\mathcal {N} (v)} \\log \\left(1 - p (\\ell , v)\\right) \\right] ^ {\\text {e n e d b y a t}} \\tag {2}", + "image_path": "3cffb16c9e016d8c82a9093634daf71ffc9dc9f56e44dacc3791052cadb3545f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "inline_equation", + "content": "p(\\ell, v) \\coloneqq \\sigma\\left(\\frac{s(\\ell, v)}{\\tau}\\right)" + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "text", + "content": " is a temperature parameter, " + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "inline_equation", + "content": "\\sigma(.)" + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "text", + "content": " is the sigmoid function and " + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "inline_equation", + "content": "w > 0" + }, + { + "bbox": [ + 130, + 395, + 481, + 424 + ], + "type": "text", + "content": " is a weight hyperparameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 441, + 304, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 304, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 304, + 455 + ], + "type": "text", + "content": "4 Experiments and Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 466, + 205, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 205, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 205, + 477 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "In order to train our model on a wide range of concepts, we train it on a mix of YouTube8M [1] and Kinetics-400 [14]. YouTube-8M (YT-8M) is primarily labeled with entities, whereas Kinetics-400 (K400) is labeled only with actions. As YT-8M is composed of a random sample of the whole of YouTube, a significant portion of video its samples are from video game streams. Many of these videos are just tagged with the video game title as the label irrespective of the actual entities and actions present in the video. A significant portion of its label vocabulary is dedicated to video game titles (771 out of 3862 total) are dedicated to video game titles. We observe training instabilities due to this, and to fix this issue, we remove video game titles from the training vocabulary. We also remove some of the least frequently occurring labels during training to reach a training vocabulary of 2429 classes. This leads to stabilized training without further issues. Note that for closed set evaluation on YT-8M, we use the YouTube-8M Segments validation set, which contains human verified labels for 1000 classes. K400 on the other hand is widely used, and has a clean vocabulary and high" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "quality human verified action labels. From the training perspective, the only limitation is the absence of entity (object, scenes etc.) labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 297 + ], + "type": "text", + "content": "For evaluation, we use 3 open-vocabulary test datasets: TAO (Tracking Any Object) [3], ActivityNet [9] and RareAct [28]. TAO and ActivityNet are exclusively object and action focused datasets respectively. RareAct has object, action labels, and focuses on unusual object-action combination. TAO dataset was developed for evaluating object trackers, however analogous to how object detection datasets like MS-COCO are used for evaluating multi-label image classification methods such as DualCoOp [40], we transform TAO into a multi-label video classification dataset by ignoring localization annotations. ActivityNet often has multiple actions in a given video. For our evaluation each annotation from RareAct provides three labels: object, action and the (unusual) object-action combination, this provides an interesting test of the open vocabulary capabilities. Altogether, these five datasets provide a comprehensive evaluation of a model's closed and open vocabulary video classification performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 315, + 208, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 208, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 208, + 327 + ], + "type": "text", + "content": "4.2Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 479, + 384 + ], + "type": "text", + "content": "As there are no prior works that explicitly address multilabel open vocabulary video classification, we extend the following methods from the zero shot image and action classification literature to our multi-label video classification setting for comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 479, + 407 + ], + "type": "text", + "content": "CoOp [52]: Short for Context Optimization, CoOp learns prompts for the CLIP text encoder as a lightweight adaptation technique for classification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 408, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 479, + 453 + ], + "type": "text", + "content": "DualCoOp [40]: Is an extension of CoOP to the multi-label setting where both positive and negative prompts are learnt for the CLIP text encoder. For a given label, prediction is based on whether the positive or negative prompted version has higher similarity with the image feature." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 456, + 479, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 479, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 479, + 478 + ], + "type": "text", + "content": "LLM + CLIP (Frozen): This baseline was discussed in Section 3.1 and is illustrated in Figure 8 of Supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 481, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 479, + 502 + ], + "type": "text", + "content": "ViFi-CLIP [33]: In this baseline, there is no temporal modeling, and both the CLIP image and text encoder are finetuned on the training dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 506, + 479, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 479, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 479, + 563 + ], + "type": "text", + "content": "As CoOp and ViFi-CLIP were developed for single label classification, to utilize them in our setting we replace their contrastive loss function with our multi-label classification loss. We also drop the region aggregation aspect of DualCoOp and only test the dual prompting architecture. We refer to these baselines as CoOp*, DualCoOp* and ViFi-CLIP* to reflect these differences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 584, + 200, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 200, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 200, + 594 + ], + "type": "text", + "content": "4.3 Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "We report two sets of metrics to evaluate the performance of the models in the open vocabulary setting. Area Under Precision-Recall curve (AUPR) summarizes the overall classification performance across the entire precision-recall trade-off. Secondly, we report the Peak F1-Score for the model on each dataset, which is the F1-Score achieved by that model at the optimal threshold chosen by an oracle" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 475, + 261 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 475, + 261 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 475, + 261 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 475, + 261 + ], + "type": "table", + "html": "
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities+Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt6.326.243.844.29.5
CLIP with Prompt Templates6.830.546.045.911.4
CLIP with Fixed LLM Prompts6.930.650.246.811.5
Trainable Baseline Methods
CoOp2.717.835.028.83.5
DualCoOp8.323.947.133.07.6
ViFi-CLIP3.410.958.317.24.1
Ours
CLIP + Learnable LLM Prompts9.432.851.447.111.9
+ Temporal Modeling14.842.063.847.112.4
+ Synthetic Labels16.743.265.550.213.2
", + "image_path": "b950e4f5d8abb9678f6387e26166fb93acf10b1ba098ee931073dcfe488a8a4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 261, + 418, + 270 + ], + "lines": [ + { + "bbox": [ + 194, + 261, + 418, + 270 + ], + "spans": [ + { + "bbox": [ + 194, + 261, + 418, + 270 + ], + "type": "text", + "content": "Table 1: AUPR scores for all methods on all datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 306, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 343 + ], + "type": "text", + "content": "for that dataset. This metric captures the classification performance obtained by an open vocabulary classification method on a dataset, if the threshold alone could be tuned, e.g., by using a labeled validation set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 199, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 199, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 199, + 373 + ], + "type": "text", + "content": "4.4 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": "AUPR scores for each method are presented in Table 1 and Peak F1 scores are presented in Table 2. Our learnable LLM prompting approach outperforms both frozen and trainable baselines across both closed-vocabulary and open-vocabulary classification tasks. Also notable is the enhancement in performance due to our temporal modeling approach (average gain of " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\approx 5\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " in AUPR). As previously mentioned in our discussion of the dataset, the training datasets YT-8M and K400 have certain limitations (in particular, YT-8M has few action labels and K400 has no object/entity labels). We present further improved results obtained by augmenting our training dataset with additional class labels obtained using a multimodal LLM (pipeline is detailed in Section I of the Supplementary Material). As seen from the last rows of Tables 1 and 2, this approach gives a further " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\approx 2\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " improvement, and is specially effective for actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 548, + 373, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 373, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 373, + 560 + ], + "type": "text", + "content": "4.5 Improved score calibration across datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In order to use a classifier in a truly open vocabulary setting in practice, the classification score needs to be well calibrated across different types of concepts. This would allow us to pre-select an optimal threshold to generate binary classification for all concepts. In Figure 6 we demonstrate the robustness offered by our method in setting a threshold that works for a wide range of concepts, and the advantage offered by our method over the frozen CLIP + LLM baseline. These figures plot the F1 scores for multi-label classification across thresholds provided by both methods with Figure 6a showing the numbers on the validation splits" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 114, + 477, + 260 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 477, + 260 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 477, + 260 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 477, + 260 + ], + "type": "table", + "html": "
MethodClosed-VocabularyOpen-Vocabulary
YouTube-8MKineticsTAO(Entities)ActivityNet(Actions)RareAct(Entities + Actions)
Frozen CLIP-based Methods
CLIP with Class name Prompt14.934.244.647.117.6
CLIP with Prompt Templates20.836.849.250.120.5
CLIP with Fixed LLM Prompts21.637.350.251.419.8
Trainable Baseline Methods
CoOp5.825.543.935.510.5
DualCoOp16.233.249.040.515.0
ViFi-CLIP5.219.354.424.79.6
Ours
CLIP + Learnable LLM Prompts23.642.452.851.122.6
+ Temporal Modeling31.546.259.652.624.3
+ Synthetic Labels32.746.656.653.825.1
", + "image_path": "a20171eded079ba94b2510e1958b0f166b4c3a4806de8ec6248c66ebbae71a88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 190, + 261, + 422, + 270 + ], + "lines": [ + { + "bbox": [ + 190, + 261, + 422, + 270 + ], + "spans": [ + { + "bbox": [ + 190, + 261, + 422, + 270 + ], + "type": "text", + "content": "Table 2: Peak F1 scores for all methods on all datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 137, + 298, + 284, + 377 + ], + "blocks": [ + { + "bbox": [ + 137, + 298, + 284, + 377 + ], + "lines": [ + { + "bbox": [ + 137, + 298, + 284, + 377 + ], + "spans": [ + { + "bbox": [ + 137, + 298, + 284, + 377 + ], + "type": "image", + "image_path": "dafc88c6d70e2d588a4642d629d68729f841aad4b170287f80a1391bc5f6fb25.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 383, + 272, + 392 + ], + "lines": [ + { + "bbox": [ + 149, + 383, + 272, + 392 + ], + "spans": [ + { + "bbox": [ + 149, + 383, + 272, + 392 + ], + "type": "text", + "content": "(a) Closed Vocabulary Evaluation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 328, + 298, + 476, + 377 + ], + "blocks": [ + { + "bbox": [ + 328, + 298, + 476, + 377 + ], + "lines": [ + { + "bbox": [ + 328, + 298, + 476, + 377 + ], + "spans": [ + { + "bbox": [ + 328, + 298, + 476, + 377 + ], + "type": "image", + "image_path": "dd25dc6544a7dc4be501e6c89cee168e4206348622993b4b60c00812bbc95075.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 383, + 461, + 392 + ], + "lines": [ + { + "bbox": [ + 342, + 383, + 461, + 392 + ], + "spans": [ + { + "bbox": [ + 342, + 383, + 461, + 392 + ], + "type": "text", + "content": "(b) Open Vocabulary Evaluation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "text", + "content": "Fig. 6: F1-Scores at different thresholds for closed and open vocabulary evaluation datasets. Our end-to-end trained model can achieve better performance across datasets with a single threshold (labeled by gray vertical line) chosen on the supervised datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 594 + ], + "type": "text", + "content": "of the training datasets, and Figure 6b showing the numbers in the open vocabulary setting on two diverse datasets. A reasonable choice for a classification threshold would be to pick the threshold that maximizes the minimum validation F1-score for the datasets used in training. The resulting choice of thresholds for our method and for frozen CLIP are shown in Figure 6a. We see that for the baseline method the performance on TAO is poor, while that on ActivityNet is modest. However with our proposed method the F1 scores for both evaluation datasets are simultaneously high. This result empirically validates the versatility that is achieved by our learnable LLM-based prompting scheme over the baseline, which makes it possible to pre-select a threshold for using the solution as a black box classifier for all unseen concepts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 616, + 209, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 209, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 209, + 627 + ], + "type": "text", + "content": "4.6 Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We scientifically ablate each part of our framework, reporting ablation evaluations (AUPR) on TAO (Objects) and ActivityNet (Actions)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 251, + 180 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 251, + 180 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 251, + 180 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 251, + 180 + ], + "type": "table", + "html": "
Open Vocabulary
# BlocksObjectsActions
565.448.1
465.550.2
264.649.1
162.347.7
", + "image_path": "17fcecd83deb8830b4465785bbed4729c4335d63feca9b8051da43ed2e42d1d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 269, + 114, + 477, + 182 + ], + "blocks": [ + { + "bbox": [ + 135, + 182, + 251, + 191 + ], + "lines": [ + { + "bbox": [ + 135, + 182, + 251, + 191 + ], + "spans": [ + { + "bbox": [ + 135, + 182, + 251, + 191 + ], + "type": "text", + "content": "(a) Effect of Temporal Modeling" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 269, + 114, + 477, + 182 + ], + "lines": [ + { + "bbox": [ + 269, + 114, + 477, + 182 + ], + "spans": [ + { + "bbox": [ + 269, + 114, + 477, + 182 + ], + "type": "table", + "html": "
Open Vocabulary
Spatial Reg.BackboneObjectsActionsGeo. Mean
✓56.653.855.2
X62.132.544.9
X42.728.334.8
", + "image_path": "435dfd0060c91a52ca54709f4591c1bfa1285e3f1d779e444f9e2527412fdb50.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 216, + 234, + 397, + 308 + ], + "blocks": [ + { + "bbox": [ + 316, + 182, + 429, + 192 + ], + "lines": [ + { + "bbox": [ + 316, + 182, + 429, + 192 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 429, + 192 + ], + "type": "text", + "content": "(b) Effect of our Regularization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 201, + 476, + 213 + ], + "lines": [ + { + "bbox": [ + 136, + 201, + 476, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 476, + 213 + ], + "type": "text", + "content": "Table 3: Ablations of Temporal Modeling Branch Architecture and Regularization" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 216, + 234, + 397, + 308 + ], + "lines": [ + { + "bbox": [ + 216, + 234, + 397, + 308 + ], + "spans": [ + { + "bbox": [ + 216, + 234, + 397, + 308 + ], + "type": "table", + "html": "
Label EncoderOpen Vocabulary
LLMPTCLIPObjects Actions
Learnable Prompt✓✗25.2 8.5
Fixed Prompt✗✓62.5 41.3
Fixed Prompt✓✓64.9 48.1
Learnable Prompt✓✓65.5 50.2
✗✗✓60.1 34.5
", + "image_path": "25b93de6d33b88c17175ba0b35af8af51e0c5effdb04e8e371b8c0ca9c2f8abd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 170, + 308, + 443, + 319 + ], + "lines": [ + { + "bbox": [ + 170, + 308, + 443, + 319 + ], + "spans": [ + { + "bbox": [ + 170, + 308, + 443, + 319 + ], + "type": "text", + "content": "Table 4: Ablating the Label Encoder. PT " + }, + { + "bbox": [ + 170, + 308, + 443, + 319 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 170, + 308, + 443, + 319 + ], + "type": "text", + "content": " Prompt Transformer" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 137, + 353, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 353, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 137, + 353, + 480, + 376 + ], + "type": "text", + "content": "- Temporal Modeling Blocks: We find that using more Blocks improves performance, until around 4. (Table 3a)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 481, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 435 + ], + "type": "text", + "content": "- Temporal Modeling Weight Regularization: We find that using our weight regularization strategy for the spatial layers of the temporal branch prevents the model from overfitting to certain concepts. We also validate our choice of keeping the CLIP vision backbone frozen and find that unfreezing the backbone leads to severe over-fitting. (Table 3b)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 436, + 481, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 508 + ], + "type": "text", + "content": "- Label Encoder: We demonstrate the effectiveness of combining an LLM with learnable prompts and CLIP's text encoder as the label encoder. First, we show that removing CLIP text encoder leads to a significant drop in performance. This demonstrates that the visual-text alignment learned by CLIP is essential. Next we demonstrate the benefits of learnable prompts over fixed prompts. We also provide results with the LLM removed (Table 4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 523, + 220, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 523, + 220, + 536 + ], + "spans": [ + { + "bbox": [ + 137, + 523, + 220, + 536 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We introduced the problem of open vocabulary multi-label video classification and proposed a solution leveraging LLMs and pre-trained VLMs. Through our extensive experiments we demonstrated strong performance on both actions and entities with a single model. Our proposed approach benefits from two key innovations. First, we proposed a method to prompt VLMs more effectively by utilizing the LLM output representations through a trainable prompt transformer and learnable LLM prompts. Second, we introduced a temporal modeling architecture and a regularized finetuning approach to improve the video understanding capability of the vision encoder while retaining strong open-vocabulary performance. Our ablations validate the efficacy of each of these contributions." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "text", + "content": "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: Proceedings of the International Conference on Machine Learning (ICML) (July 2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 201, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 480, + 245 + ], + "type": "text", + "content": "3. Dave, A., Khurana, T., Tokmakov, P., Schmid, C., Ramanan, D.: Tao: A large-scale benchmark for tracking any object. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 436-454. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "4. Desai, K., Kaul, G., Aysola, Z., Johnson, J.: Redcaps: Web-curated image-text data created by the people, for the people. arXiv preprint arXiv:2111.11431 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 287 + ], + "type": "text", + "content": "5. Fan, L., Krishnan, D., Isola, P., Katabi, D., Tian, Y.: Improving clip training with language rewrites. In: NeurIPS (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 287, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 480, + 308 + ], + "type": "text", + "content": "6. Fang, H., Xiong, P., Xu, L., Chen, Y.: Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 308, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 351 + ], + "type": "text", + "content": "7. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 351, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 480, + 407 + ], + "type": "text", + "content": "8. Gupta, R., Roy, A., Christensen, C., Kim, S., Gerard, S., Cincebeaux, M., Divakaran, A., Grindal, T., Shah, M.: Class prototypes based contrastive learning for classifying multi-label and fine-grained educational videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19923-19933 (June 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 459 + ], + "type": "text", + "content": "9. Heilbron, F.C., Niebles, J.C.: Collecting and annotating human activities in web videos. In: Proceedings of International Conference on Multimedia Retrieval. p. 377-384. ICMR '14, Association for Computing Machinery, New York, NY, USA (2014). https://doi.org/10.1145/2578726.2578775, https://doi.org/10.1145/2578726.2578775" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "text", + "content": "0. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 568 + ], + "type": "text", + "content": "1. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "type": "text", + "content": "2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "text", + "content": "3. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. In: International Conference on Machine Learning (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., Suleyman, M., Zisserman, A.: The kinetics human action video dataset (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 162 + ], + "type": "text", + "content": "15. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=e2TBb5y0yFf" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 163, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 163, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 481, + 228 + ], + "type": "text", + "content": "16. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 228, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 481, + 295 + ], + "type": "text", + "content": "17. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 12888-12900. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/1i22n.html" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 296, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 328 + ], + "type": "text", + "content": "18. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 329, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 481, + 373 + ], + "type": "text", + "content": "19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10965-10975 (June 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 374, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 451 + ], + "type": "text", + "content": "20. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 452, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 481, + 485 + ], + "type": "text", + "content": "21. Lin, W., Karlinsky, L., Shvetsova, N., Possegger, H., Kozinski, M., Panda, R., Feris, R., Kuehne, H., Bischof, H.: Match, expand and improve: Unsupervised finetuning for zero-shot action recognition with language knowledge. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "text", + "content": "22. Lin, X., Petroni, F., Bertasius, G., Rohrbach, M., Chang, S.F., Torresani, L.: Learning to recognize procedural activities with distant supervision. arXiv preprint arXiv:2201.10990 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 520, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 541 + ], + "type": "text", + "content": "23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Advances in Neural Information Processing Systems. Curran Associates, Inc. (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 542, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 481, + 586 + ], + "type": "text", + "content": "24. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6555-6564 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 587, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 631 + ], + "type": "text", + "content": "25. Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., Tang, J.: Gpt understands, too. AI Open (2023). https://doi.org/https://doi.org/10.1016/j.aiopen.2023.08.012, https://www.sciencedirect.com/science/article/pii/S2666651023000141" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "27. Menon, S., Vondrick, C.: Visual classification via description from large language models. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=j1AjNL8z5cs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "28. Miech, A., Alayrac, J.B., Laptev, I., Sivic, J., Zisserman, A.: Rareact: A video dataset of unusual interactions. arxiv:2008.01018 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 226 + ], + "type": "text", + "content": "29. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 728-755. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "30. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "text", + "content": "31. Pratt, S., Covert, I., Liu, R., Farhadi, A.: What does a platypus look like? generating customized prompts for zero-shot image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15691-15701 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 346 + ], + "type": "text", + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "33. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 424 + ], + "type": "text", + "content": "34. Roth, K., Kim, J.M., Koepke, A.S., Vinyals, O., Schmid, C., Akata, Z.: Waffling around for performance: Visual classification with random words and broad concepts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15746-15757 (October 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 468 + ], + "type": "text", + "content": "35. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "36. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "text", + "content": "37. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "text", + "content": "38. Shvetsova, N., Kukleva, A., Hong, X., Rupprecht, C., Schiele, B., Kuehne, H.: HowtocapTION: Prompting llms to transform video annotations at scale (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "text", + "content": "39. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "text", + "content": "40. Sun, X., Hu, P., Saenko, K.: Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems 35, 30569-30582 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "41. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Commun. ACM" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 241, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary Multi-Label Video Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 577 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "type": "text", + "content": "59(2), 64-73 (jan 2016). https://doi.org/10.1145/2812802, https://doi.org/10.1145/2812802" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "42. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23034–23044 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 237 + ], + "type": "text", + "content": "43. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-VCLIP: Transforming CLIP to an open-vocabulary video model via interpolated weight optimization. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 36978-36989. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/weng23b.htm1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "44. Weng, Z., Yang, X., Li, A., Wu, Z., Jiang, Y.G.: Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In: ICML (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 270, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 270, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 270, + 481, + 314 + ], + "type": "text", + "content": "45. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "type": "text", + "content": "46. Xu, Z., Zhu, Y., Deng, T., Mittal, A., Chen, Y., Wang, M., Favaro, P., Tighe, J., Modolo, D.: Challenges of zero-shot recognition with vision-language models: Granularity and correctness (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "type": "text", + "content": "47. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 423 + ], + "type": "text", + "content": "48. Yang, Y., Panagopoulou, A., Zhou, S., Jin, D., Callison-Burch, C., Yatskar, M.: Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19187-19197 (June 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "type": "text", + "content": "49. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "type": "text", + "content": "50. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 490, + 481, + 511 + ], + "type": "text", + "content": "51. Zhao, Y., Misra, I., Krahenbuhl, P., Girdhar, R.: Learning video representations from large language models. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 481, + 533 + ], + "type": "text", + "content": "52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "type": "text", + "content": "53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2639-2650 (October 2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_content_list.json b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..583b7b86d8f389284b0c923e84b37164166ab691 --- /dev/null +++ b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_content_list.json @@ -0,0 +1,2257 @@ +[ + { + "type": "text", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiyang Su, Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu", + "bbox": [ + 243, + 212, + 759, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michigan State University, East Lansing, MI 48824, USA {suyiyan1, kimminc2, liufeng6, jain, liuxm}@msu.edu", + "bbox": [ + 302, + 239, + 702, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Biometric recognition has primarily addressed closed-set identification, assuming all probe subjects are in the gallery. However, most practical applications involve open-set biometrics, where probe subjects may or may not be present in the gallery. This poses distinct challenges in effectively distinguishing individuals in the gallery while minimizing false detections. While it is commonly believed that powerful biometric models can excel in both closed- and open-set scenarios, existing loss functions are inconsistent with open-set evaluation. They treat genuine (mated) and imposter (non-mated) similarity scores symmetrically and neglect the relative magnitudes of imposter scores. To address these issues, we simulate open-set evaluation using minibatches during training and introduce novel loss functions: (1) the identification-detection loss optimized for open-set performance under selective thresholds and (2) relative threshold minimization to reduce the maximum negative score for each probe. Across diverse biometric tasks, including face recognition, gait recognition, and person re-identification, our experiments demonstrate the effectiveness of the proposed loss functions, significantly enhancing open-set performance while positively impacting closed-set performance. Our code and models are available here.", + "bbox": [ + 261, + 308, + 743, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open-set biometrics $\\cdot$ Face recognition $\\cdot$ Gait recognition $\\cdot$ Person reID", + "bbox": [ + 261, + 585, + 740, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 640, + 375, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Biometrics, the study of unique physiological or behavioral characteristics for human identification, plays a pivotal role in applications such as access control, surveillance, and forensic analysis [52]. Each biometric task, from face recognition, gait recognition to person re-identification (reID), uses features tailored to its specific requirements [39, 55, 71].", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world applications, after test subjects are enrolled in the gallery, biometric systems inevitably encounter non-mated subjects, those not present in the gallery. Thus, it becomes imperative for such systems to demonstrate robust performance in the open-set framework with non-mated probes [16, 18, 28, 58]. The closed-set framework assumes all probe subjects are mated and predominantly focuses on identification, associating probes with the closest match in the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1894c19c93dc6f73e2dc162e99b0289153a08461e1bce4793f1754b8e568e40c.jpg", + "image_caption": [ + "Fig. 1: A toy example for FNIR@FPIR calculation. The score matrix and distribution between non-mated probes and all gallery subjects are shown in blue. The maximum score for each non-mated probe is shown in red. A percentile of their distribution determines the threshold. A user-defined percentile (e.g., $33\\%$ ) sets the threshold to be 0.75 as 1 out of 3 (the $2^{\\mathrm{nd}}$ ) non-mated probes is incorrectly classified as \"in the gallery\". The genuine scores of mated probes are shown in green. A mated probe is an FN if either of the conditions is met: (1) in detection (Det.), it is incorrectly classified as \"not in the gallery\" ( $\\mathcal{G}$ ) as the genuine score is below the threshold (the $1^{\\mathrm{st}}$ mated probe); (2) its genuine score is not within a specified rank $R$ , e.g., 1, in the gallery (the $2^{\\mathrm{nd}}$ mated probe). Hence, FNIR@33%FPIR is $75\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 212, + 143, + 787, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gallery. However, the open-set framework introduces an additional challenge: detection, determining whether a probe subject is mated. For example, in airport surveillance, faces in CCTV videos are compared against a criminal watch list. The goal is to reduce the misidentification (false positives) of non-mated subjects such as airport staff and innocent passengers (improving detection), while simultaneously ensuring accurate recognition of known criminals (enhancing identification). For this reason, open-set biometrics is a harder and more general task than closed-set.", + "bbox": [ + 212, + 460, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The biometric community has predominantly centered on the closed-set framework. While some face recognition benchmarks have open-set evaluation protocols [18,19,28], they are primarily used for evaluation rather than for developing tailored strategies to address the open-set challenges. In gait recognition and person reID, most benchmarks only provide official closed-set protocols, lacking open-set ones. To address this gap, we take the initiative to introduce open-set benchmarks and propose strategies to train models specifically for improving open-set performances while maintaining closed-set performance.", + "bbox": [ + 212, + 582, + 787, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The false negative identification rate at a given false positive identification rate (FNIR@FPIR) is a de facto open-set metric [16]. We first compute the similarity score matrix between non-mated probes and the gallery, from which we extract the maximum score for each non-mated probe across the gallery, forming the maximum-per-probe non-mated score distribution. This distribution represents the likelihood that a non-mated probe has a match in the gallery. A threshold is set based on a user-defined percentile (e.g., $1\\%$ ) of this distribution. Mated probes with scores below this threshold are counted as false negatives (FNs), as they are incorrectly considered not to be in the gallery. Similar to the", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "closed-set scenario, a mated probe is also an FN if its gallery match doesn't rank among the gallery set within a user-specified rank $R$ . FNIR is the proportion of mated probes classified as FNs. Fig. 1 depicts its calculation pipeline.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike closed-set or verification metrics, FNIR@FPIR introduces unique challenges. It employs a threshold based on maximum scores for each non-mated probe, creating an asymmetry between genuine and imposter scores. SOTA biometric recognition methods achieve strong performance in closed-set scenarios leveraging triplet loss [20] and/or softmax cross-entropy based losses [9, 25, 29, 46, 60], both of which excel in reducing intra-person variance and increasing inter-person variance. However, they do not take into account the fact that the non-mated score distribution determines the threshold for the false negative.", + "bbox": [ + 212, + 193, + 784, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In other words, conventional loss functions, while effective in closed-set scenarios, exhibit limitations in addressing the nuances of the FNIR@FPIR metric, which necessitate a handling of the maximum of imposter scores. These high-scoring non-mated probes disproportionately influence the threshold setting for false negatives, yet conventional loss functions often treat them equivalently to other scores. As a result, they face challenges in open-set scenarios. GaitBase [13], a SOTA gait recognition model, can only achieve a $61\\%$ FNIR@1%FPIR on the training set and $86\\%$ on the test set. This challenge extends to face recognition and person reID.", + "bbox": [ + 212, + 316, + 785, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this challenge, we propose to create simulated open-set evaluation episodes during training, whose score distributions serve as proxies to those of the test set. Specifically, we introduce an identification-detection loss tailored to optimize the two core elements in open-set biometrics, i.e., detection and identification. For detection, the model is trained to accurately pinpoint the exact probes whose mate is present in the gallery set; for identification, the model learns to establish associations between the gallery and probe samples of mated subjects. By integrating the aspects of detection and identification within training batches, we improve the model's efficacy in both domains, thus yielding enhanced testing performance. Further, we minimize similarity scores between non-mated subjects and the gallery, enabling a more comprehensive detection of mated subjects (see Fig. 3).", + "bbox": [ + 212, + 454, + 787, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirically, we apply our proposed loss function to three biometric modalities, face, gait, and person reID, and show performance improvement in all tasks. While our focus is open-set biometrics, we also observe performance gain in closed-set search due to improved feature representations.", + "bbox": [ + 212, + 637, + 784, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions include:", + "bbox": [ + 238, + 700, + 537, + 714 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\checkmark$ We show that the existing loss functions, e.g., triplet loss and softmax loss, do not treat genuine and imposter training samples in accordance with open-set evaluation. This inconsistency sheds light on why the open-set challenges have not been adequately addressed.", + "$\\checkmark$ We propose a novel loss function that leverages open-set mini-batches during training to optimize identification and detection performance and minimize open-set detection thresholds." + ], + "bbox": [ + 220, + 732, + 782, + 835 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8229de76d49520d971e21e38a586d6917b82ff56eb1fcba2dd37204d1abf2606.jpg", + "image_caption": [ + "Fig. 2: Comparison between open-set recognition and biometrics. Closed-set recognition has the same set of classes for training and testing. Open-set recognition involves testing on unseen classes during training. On the other hand, in biometrics, the sets of training and testing subjects are always disjoint. In closed-set biometrics, all probe subjects are in the gallery, and in open-set biometrics, some probe subjects are not enrolled in the gallery. [Legends: ▪ gallery subjects, ▪ probe subjects]" + ], + "image_footnote": [], + "bbox": [ + 263, + 142, + 339, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Testing classes", + "bbox": [ + 254, + 181, + 349, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Closed-Set Recognition", + "bbox": [ + 267, + 205, + 336, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3287ea4dc50c33faa3b3076a0d310355d1537a07da695711a9af3e152d35debc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 143, + 472, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Testing classes", + "bbox": [ + 374, + 181, + 470, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Set Recognition", + "bbox": [ + 388, + 205, + 459, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5b732f0aacb12364654e5d6f26e60e7327007c202f5d086397307626c56fa04b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 146, + 527, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training subjects", + "bbox": [ + 493, + 181, + 540, + 205 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff77196acab4689c6cb553d048134e6766f969b6c2b2770e143e63d8afc3e0c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 143, + 609, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Testing subjects", + "bbox": [ + 560, + 181, + 602, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Closed-Set Biometrics", + "bbox": [ + 516, + 205, + 581, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d382f7589b5738bf0e795179809598bc523c3e350e4f6f81bf39439f8df7c02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 148, + 666, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training subjects", + "bbox": [ + 632, + 181, + 679, + 205 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3feb03aabdf216d3310d60c1dc57b76fd707d2f7c537c2fd8ac7b7130935b009.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 148, + 754, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Testing subjects", + "bbox": [ + 697, + 181, + 743, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Set Biometrics", + "bbox": [ + 656, + 205, + 720, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\checkmark$ We experimentally demonstrate that our proposed loss function can effectively improve both closed- and open-set performance across diverse biometric modalities, including face, gait, and person reID.", + "bbox": [ + 220, + 363, + 784, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 428, + 387, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Set Biometrics. While various biometric recognition tasks, such as face recognition [9, 24, 25, 29-31, 41, 43, 46, 50, 60], gait recognition [3, 4, 7, 10-14, 23, 26, 36-38, 44, 54, 61, 62, 68, 74-77], and person reID [5, 17, 21, 35, 36, 40, 42, 45, 48, 56, 63, 66], have witnessed significant progress, the majority of their focus remains on closed-set scenarios. Although EVM has been applied to open-set face recognition [18], EVM takes extracted features, instead of images/videos, as input. Our approach improves feature representation and EVM improves score calculation. While some face recognition datasets provide opens-set evaluation protocols [28], open-set performance is often perceived as a consequence rather than a direct objective in advancing recognition models. Further, even the most recent gait recognition [68] and person reID [42] models are exclusively evaluated under closed-set settings due to the absence of open-set evaluation protocols.", + "bbox": [ + 212, + 457, + 787, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-Set Recognition. While open-set biometrics is relatively underexplored, open-set recognition [1,2,6,22,32-34,51,59,64,67,69,72,73,78] has attracted attention in the literature. Open-set recognition distinguishes between known and unknown classes in image classification. In both open-set image recognition and open-set biometrics, the primary concern is how the models deal with instances that belong to unknown classes or non-mated probes. However, as illustrated in Fig. 2, in open-set recognition known classes come from the training set and the unknown classes refer to classes not used for training [15]. On the other hand, even in closed-set biometrics, all test subjects are typically not used during training and mated probes belong to test subjects enrolled in the gallery [27,47]. Open-set biometrics differs from closed-set biometrics in that there are non-mated subjects are neither in the training set nor in the gallery [16,18,58]. In", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6898f0d2ae795b6537516797e76054f9c31695e15459e8316eb56ca4572e2159.jpg", + "image_caption": [ + "Fig. 3: Overview of our approach. We introduce an additional modality-agnostic loss function specifically designed for open-set biometrics, which simultaneously optimizes identification (Id.) and detection (Det.) and Relatie Threshold Minimization (RTM) for open-set thresholds." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 743, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "evaluation, open-set recognition models typically use closed-set recognition accuracy and AUROC for detecting classes not seen during training [15] and open-set biometrics uses FNIR@FPIR [16].", + "bbox": [ + 212, + 321, + 782, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 392, + 330, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 426, + 437, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $S_{\\mathrm{test}}$ be the test set, $\\mathcal{G} \\subset S_{\\mathrm{test}}$ represent the gallery set containing mated subjects, and $\\mathcal{P} \\subset S_{\\mathrm{test}}$ denote the probe set consisting of test samples. For each probe sample $p_i \\in \\mathcal{P}$ , there are two possible outcomes: (1) If $p_i$ is mated, i.e., its corresponding subject $g_i$ is in $\\mathcal{G}$ , the biometric model must identify the correct subject. (2) If $p_i$ is non-mated, the model must reject it. The open-set scenario poses a more challenging problem than the traditional closed-set one, as the model needs to recognize mated subjects while rejecting non-mated ones.", + "bbox": [ + 212, + 452, + 787, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Revisiting Loss Functions for Biometrics", + "text_level": 1, + "bbox": [ + 215, + 582, + 596, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compute gradients of two common loss functions in biometrics: softmax and triplet loss. Many loss functions are derived from them and share similar characteristics. For instance, center loss [65] can be analyzed similarly to softmax loss, and the logic of triplet loss can also be applied to circle loss [57].", + "bbox": [ + 212, + 608, + 785, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Softmax loss. The softmax loss, often employed in closed-set classification, operates by computing the cross-entropy between the normalized logit of a sample and its corresponding ground truth class. Mathematically, for a sample $i$ in a batch, the softmax loss $L_{\\mathrm{softmax}}(i)$ is:", + "bbox": [ + 212, + 691, + 787, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s o f t m a x}} (i) = - \\log \\left(\\frac {e ^ {z _ {i , s}}}{\\sum_ {t} e ^ {z _ {i , t}}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 765, + 785, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $z_{i,s}$ denotes the logit for sample $i$ belonging to subject $s$ . This loss aims to minimize the negative log-likelihood of the true subject $s$ for sample $i$ . For one", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sample, the gradient of the softmax loss $w.r.t.$ the logit $z_{i,s}$ is given by:", + "bbox": [ + 215, + 146, + 728, + 162 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L _ {\\mathrm {s o f t m a x}} (i)}{\\partial z _ {i , s}} = \\frac {e ^ {z _ {i , s}}}{\\sum_ {t} e ^ {z _ {i , t}}} - \\delta_ {i, s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 174, + 785, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\frac{e^{z_{i,s}}}{\\sum_{j=1}^{N} e^{z_{j,s}}}$ represents the probability that sample $i$ belongs to subject $s$ and $\\delta_{i,c}$ is the Kronecker delta, which is 1 when $i = c$ (the true subject) and 0 otherwise. We visualize gradients of the softmax loss in Fig. 4(a).", + "bbox": [ + 214, + 217, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, it overlooks the relative magnitude of each imposter score within the batch. An imposter score may be deemed small by, e.g., AdaFace [29] standards but still rank high among maximum-per-probe non-mated scores, raising the open-set threshold. Further, the softmax loss assumes that each sample can be assigned to one of the training subjects, which is not the case in open-set settings. Consequently, the softmax loss is not inherently designed to handle open-set biometrics.", + "bbox": [ + 214, + 270, + 787, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Triplet loss. The triplet loss [53] is defined as:", + "bbox": [ + 215, + 397, + 563, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {t r i}} (a, p, n) = \\max (0, d (a, p) - d (a, n) + m). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 425, + 785, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that $d(a, p)$ is the distance between the anchor $a$ and positive sample $p$ , $d(a, n)$ is between $a$ and the negative sample $n$ , and $m$ is a constant margin. The partial derivatives of $L(a, p, n)$ w.r.t. $d(a, p)$ and $d(a, n)$ are piecewise functions:", + "bbox": [ + 214, + 452, + 787, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L _ {\\operatorname {t r i}} (a , p , n)}{\\partial d (a , p)} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} d (a, p) + m > d (a, n) \\\\ 0, & \\text {i f} d (a, p) + m \\leq d (a, n) \\end{array} ; \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 522, + 785, + 560 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L _ {\\operatorname {t r i}} (a , p , n)}{\\partial d (a , n)} = \\left\\{ \\begin{array}{l l} - 1, & \\text {i f} d (a, p) + m > d (a, n) \\\\ 0, & \\text {i f} d (a, p) + m \\leq d (a, n) \\end{array} \\right.. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 565, + 785, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The magnitude of the gradient is constant regardless of the distance between two samples. This inability to differentiate samples in the gradient implies that all negative samples in the triplet are contributing equally, which is not reflective of the open-set scenario. One remedy often used with triplet loss is hard negative mining. However, it comes with a critical drawback of being susceptible to label noise [70]. Noisy and mislabeled samples are often considered hard negatives, which can make the loss function sensitive to noisy data when focusing on the hardest samples. Thus, we believe that a loss function that considers all negative samples while being able to adjust the gradient magnitude based on the anchor-negative distances would be effective in open-set biometrics.", + "bbox": [ + 214, + 613, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, we propose a novel loss function, shown in Fig. 3. It optimizes identification and detection at various thresholds and minimizes non-mated scores to reduce open-set thresholds. It can be applied to any biometric modality. Unlike existing loss functions, it is consistent with FNIR@FPIR in that it treats genuine and imposter scores asymmetrically and tackles large non-mated scores.", + "bbox": [ + 214, + 765, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52c76ac312b534708c2225e80f66d0cc7ce2254ab9a3e9cde2a79db71cede52d.jpg", + "image_caption": [ + "(a) SoftMax" + ], + "image_footnote": [ + "$\\oplus$ Genuine probe $(p)$" + ], + "bbox": [ + 223, + 159, + 305, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e4793ee91db32ed3139b83a6b44fa7ca4588bbf0788886e7406ed3bbc271270.jpg", + "image_caption": [ + "Gradient Magnitude" + ], + "image_footnote": [ + "Threshold", + "$\\rightarrow$ Gradient" + ], + "bbox": [ + 225, + 265, + 316, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a830a71b69130fadc524c4716deb1d21977307162096105aa84fd8873d5218c.jpg", + "image_caption": [ + "(b) Triplet Loss" + ], + "image_footnote": [ + "$\\ominus$ Imposter probe $(n)$" + ], + "bbox": [ + 326, + 159, + 405, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5577a7bd635537aa144436de87afdb42dc6b0919ea9794c11b1cd967daf7b703.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 220, + 403, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a1469d03aa844543f7c8ac099f771de10b7f5c1d72b440e31d5bf457079ae5b.jpg", + "image_caption": [ + "(c) Ours", + "Identification Detection", + "$p_i$ Most Emphasized" + ], + "image_footnote": [ + "different $p_i$ gradient magnitude based on threshold" + ], + "bbox": [ + 540, + 154, + 620, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28ca7ef5ed26341493e44c4f99627cdc714a0fb9ad7e2a21015cc390d5fb56e3.jpg", + "image_caption": [ + "Relative Threshold Minimization", + "$n_2$ Emphasized", + "Fig. 4: Visualization of the gradient of different loss functions. The colors represent gradient magnitude and the arrows represent gradient directions. a) SoftMax loss pushes the sample toward the class center and away from the other classes, with gradual shifts in gradient magnitude. b) On the other hand, triplet loss has a hard cut gradient based on triplet samples. c) For $S_{\\tau}^{det}$ , as shown in the top row, the thresholds are determined by non-mated sample, $n_j$ . $p_i$ 's gradient has the greatest magnitude when it has a similar distance from the gallery $g_i$ to $n_j$ . For Relative Threshold Minimization, as shown in the bottom row, as $n_2$ moves away from the gallery, its gradient decreases. While $n_1$ remains at the same location, its gradient increases because it becomes closer to $g_i$ than $n_2$ . The gradients w.r.t. genuine scores adapt to non-mated scores and the gradients w.r.t. non-mated scores are adaptive to other non-mated scores." + ], + "image_footnote": [ + "different $n_j$ gradient magnitude based on relative distance" + ], + "bbox": [ + 540, + 226, + 620, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3667c27e033dc36e5f678e648abe413ccbb9658585d27bde538392110b58c8a0.jpg", + "image_caption": [ + "$p_i$ Emphasized" + ], + "image_footnote": [], + "bbox": [ + 622, + 154, + 700, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29ddcc80faf53ffbfafc1a5378346db052fd4bc681ce2845a5690b34d8e805fe.jpg", + "image_caption": [ + "$n_1,n_2$ Emphasized" + ], + "image_footnote": [], + "bbox": [ + 622, + 226, + 700, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d75486625275e10b438d32d877414329e2dda47c910815db9dc4e69a12346b4.jpg", + "image_caption": [ + "$p_i$ Less Emphasized" + ], + "image_footnote": [], + "bbox": [ + 704, + 155, + 782, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bed9bd62c7db3fceb67f87596eebbd49389e55ec4dda998918f834a5e5221c12.jpg", + "image_caption": [ + "$n_1$ Emphasized" + ], + "image_footnote": [], + "bbox": [ + 704, + 226, + 782, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Our Approach", + "text_level": 1, + "bbox": [ + 215, + 482, + 380, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To design loss functions tailored for open-set biometrics, we create a training scenario mirroring the testing environment. During training, we partition each batch into gallery and probe sets to simulate open-set evaluation conditions. In this process, we randomly designate a certain percentage, denoted as $p\\%$ , of the subjects as mated subjects, while the rest are categorized as non-mated. For the mated subjects, we further randomly allocate their exemplars to the gallery and probe sets. Conversely, all exemplars of the non-mated subjects are placed in the probe set. This approach yields a more realistic and representative training scenario, preparing the biometric features to excel under open-set conditions. To avoid any potential notation ambiguity, we refer to the mated probe set created within each training batch as $\\mathcal{P}_{\\mathrm{K}}^{\\prime}$ , the non-mated probe set as $\\mathcal{P}_{\\mathrm{U}}^{\\prime}$ , and the gallery sets within each training batch as $\\mathcal{G}^{\\prime}$ .", + "bbox": [ + 212, + 506, + 787, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure robustness, when the metric space inhabited by the biometric features is defined under the Euclidean distance, we calculate the similarity scores between two features as:", + "bbox": [ + 212, + 686, + 785, + 731 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {s} \\left(p _ {i}, g _ {j}\\right) = \\frac {1}{1 + \\mathrm {d} \\left(p _ {i} , g _ {j}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 738, + 785, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{d}(\\cdot, \\cdot)$ represents the Euclidean distance. Alternatively, if the metric space is defined under the cosine distance, we simply adopt the cosine similarity.", + "bbox": [ + 212, + 779, + 784, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure that the model can generalize to open-set scenarios during testing, we propose loss functions that penalize the model for three types of errors during", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training: (1) failing to detect a mated probe with a threshold $\\tau$ , (2) failing to identify a mated probe within the top rank- $r$ positions and (3) setting the threshold $\\tau$ too high, leading to poor differentiation between mated and non-mated sample distributions.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(1) Detection. Detection is assessing whether the computed similarity score exceeds a predefined threshold $\\tau$ . That is, for a given probe $p_i \\in \\mathcal{P}_{\\mathrm{K}}'$ and its corresponding gallery subject $g_i \\in \\mathcal{G}'$ , we calculate a detection score as:", + "bbox": [ + 212, + 227, + 787, + 272 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\tau} ^ {d e t} \\left(p _ {i}, g _ {i}\\right) = \\sigma_ {\\alpha} \\left(\\mathrm {s} \\left(p _ {i}, g _ {i}\\right) - \\tau\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 281, + 785, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sigma_{\\alpha}(x) = 1 / (1 + \\exp (-\\alpha x))$ is the Sigmoid function with temperature hyperparameter $\\alpha$ . The $S_{\\tau}^{det}(p_i, g_i)$ score quantifies the model's ability to detect if $p_i$ and $g_i$ belong to the same subject, given the threshold $\\tau$ . A high similarity score, well above $\\tau$ , indicates successful detection and $S_{\\tau}^{det}(p_i, g_i)$ approaches 1. In contrast, if the model fails to detect that $p_i$ and $g_i$ belong to the same subject, $s(p_i, g_i)$ remains much smaller than $\\tau$ , causing $S_{\\tau}^{det}(p_i, g_i)$ to approach 0. Since the Sigmoid function has the highest gradient near 0, the term $\\mathrm{s}(p_i, g_i) - \\tau$ makes the loss function focus more on the samples near the threshold $\\tau$ .", + "bbox": [ + 212, + 306, + 787, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During testing, the open-set detection threshold is determined by the non-mated score distribution, which we do not know during training. Thus, we sample $\\tau$ from the non-mated scores in a batch. Specifically, we let", + "bbox": [ + 212, + 429, + 787, + 473 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {d e t} = \\frac {1}{| \\mathcal {T} |} \\sum_ {\\tau \\in \\mathcal {T}} S _ {\\tau} ^ {d e t} \\left(p _ {i}, g _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 482, + 785, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{T} = \\{s(n_j,g_i)|n_j\\in \\mathcal{P}_{\\mathrm{U}}'\\}$ is the set of all non-mated scores. Fig. 4(c) illustrates the gradient of Det w.r.t. mated probes under different thresholds induced by various non-mated probes. The gradient is adaptable to different thresholds, increasing as $\\tau$ decreases, even when $s(g,p)$ is constant. This adaptability is desirable in open-set settings, as it promotes mated probes to be more similar to their ground-truth gallery than non-mated probes. Note that the thresholds in Eq. (8) are only used during training. The detection thresholds during testing are automatically determined by the FNIR@FPIR metric as discussed in Sec. 1.", + "bbox": [ + 212, + 525, + 787, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(2) Identification. It is not sufficient for a biometric model to distinguish mated subjects from non-mated ones based on their scores. Equally important is the model's ability to accurately identify the correct subject among the gallery subjects. Therefore, we introduce an identification score $S^{id}(p_i, g_i)$ as:", + "bbox": [ + 212, + 667, + 787, + 728 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {i d} \\left(p _ {i}, g _ {i}\\right) = \\sigma_ {\\beta} \\left(1 - \\operatorname {s o f t r a n k} \\left(p _ {i}, g _ {i}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 737, + 785, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sigma_{\\beta}$ is a Sigmoid function with temperature $\\beta$ and sofrank is a differentiable function with a Sigmoid function $\\sigma_{\\gamma}$ defined as:", + "bbox": [ + 212, + 762, + 787, + 792 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s o f t r a n k} \\left(p _ {i}, g _ {i}\\right) = \\sum_ {g _ {j} \\in \\mathcal {G} ^ {\\prime}} \\sigma_ {\\gamma} \\left(\\mathrm {s} \\left(p _ {i}, g _ {j}\\right) - \\mathrm {s} \\left(p _ {i}, g _ {i}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 804, + 785, + 835 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The softrank function accumulates values close to 1 for gallery subjects $g_{j}$ more similar to the probe $p_{i}$ than the correct match $g_{i}$ and close to 0 otherwise. It essentially reflects the rank of the correct match $g_{i}$ among all gallery subjects for probe $p_{i}$ . The $S^{id}(p_i, g_i)$ is an indicator, approaching 1 if $p_{i}$ has a softrank value less than 1, indicating successful identification.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To optimize open-set detection and identification simultaneously, the overall identification-detection loss $\\mathcal{L}_{r,\\tau}$ is formulated as the product of two components,", + "bbox": [ + 212, + 222, + 782, + 252 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {I D L}} = - \\frac {1}{| \\mathcal {P} _ {\\mathrm {K}} ^ {\\prime} |} \\sum_ {p _ {i} \\in \\mathcal {P} _ {\\mathrm {K}} ^ {\\prime}} S ^ {d e t} \\left(p _ {i}, g _ {i}\\right) \\cdot S ^ {i d} \\left(p _ {i}, g _ {i}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 270, + 784, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In other words, $\\mathcal{L}^{\\mathrm{IDL}}$ is only close to $-1$ if both $S^{id}(p_i,g_i)$ and $S^{det}(p_i,g_i)$ approach 1. Conversely, should either or both not approximate 1, $\\mathcal{L}^{\\mathrm{IDL}}$ converges to 0.", + "bbox": [ + 212, + 313, + 782, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(3) Relative Threshold Minimization $\\mathcal{L}^{\\mathrm{IDL}}$ optimizes detection and identification under fixed thresholds. We also introduce a strategy to minimize the maximum scores for non-mated probes, discouraging the model from assigning high scores to non-mated probes. This helps mitigate the risk of false positive detections in open-set scenarios, enhancing overall robustness.", + "bbox": [ + 212, + 377, + 782, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To achieve this, we minimize the weighted average of the similarities between the probe and each gallery subject, where the weights are obtained by applying the softmax function to each score. Mathematically, given a probe sample $p_i$ and a set of gallery subjects $\\mathcal{G}' = \\{g_1, g_2, \\dots, g_n\\}$ with corresponding similarity scores $s_1, s_2, \\dots, s_n$ , we minimize the weighted average as follows:", + "bbox": [ + 212, + 454, + 784, + 529 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {R T M}} = \\frac {1}{\\sum_ {j = 1} ^ {n} e ^ {s _ {j}}} \\sum_ {j = 1} ^ {n} e ^ {s _ {j}} \\cdot s _ {j}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 535, + 784, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $e^{s_j}$ is the softmax value for each similarity score.", + "bbox": [ + 212, + 579, + 612, + 593 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 4 (c) shows gradients of $\\mathcal{L}_{\\mathrm{RTM}}$ w.r.t. non-mated probes. This gradient adapts to the relative position of non-mated probes, assigning higher gradients to samples closer to the gallery. Unlike softmax-based losses and the triplet loss, RTM consistently emphasizes imposter scores that rank higher compared to others within the batch to learn better feature representations for open-set biometrics. This property is desirable as it effectively minimizes the maximum score of each non-mated probe.", + "bbox": [ + 212, + 595, + 782, + 698 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The max function is an alternative to the weighted average. Yet, simply minimizing the maximum score in training would not minimize other scores even if they are close to the maximum. Hence, the weighted average would make it easier for the trained model to generalize during testing.", + "bbox": [ + 212, + 700, + 782, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Overall Loss", + "text_level": 1, + "bbox": [ + 214, + 777, + 362, + 791 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall objective is to minimize the combination of $\\mathcal{L}^{\\mathrm{IDL}}$ and $\\mathcal{L}^{\\mathrm{RTM}}$ ,", + "bbox": [ + 214, + 797, + 736, + 813 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL = \\mathcal {L} ^ {\\mathrm {I D L}} + \\lambda \\cdot \\mathcal {L} ^ {\\mathrm {R T M}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 819, + 782, + 835 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda$ is the loss weight. Compared to existing loss functions, our loss function treats genuine and imposter scores differently. It aims to reduce open-set thresholds by pushing non-mated probes away from gallery while pulling genuine scores toward their corresponding galleries, similar to open-set evaluation. Moreover, the relative magnitude of non-mated scores is utilized by RTM to assign larger gradients to harder non-mated probes.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 258, + 375, + 276 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 289, + 428, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocols. In face recognition experiments, we use the official closed- and open-set evaluation protocols of IJB-S [28]. However, gait recognition and person reID datasets typically lack official open-set evaluation protocols, where some probe subjects are excluded from the gallery. Naïvely removing gallery subjects often results in highly unstable outcomes. This instability arises because threshold calculation heavily relies on a small fraction (1%) of non-mated probes due to the limited number of probes in these datasets. For instance, in a dataset with 500 non-mated probes, the threshold is computed from the highest scores of only 5 probes.", + "bbox": [ + 212, + 313, + 787, + 449 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To ensure robust open-set evaluation, we create $N (= 50)$ random sets of non-mated probes and compute FNIR@1%FPIR for each set. In each set, $q\\% = 21.5\\%$ subjects are designated as non-mated. We report the median value and the standard deviation from $N$ trials. The choice of median over mean is deliberate to avoid undue influence from extreme values and provide a more stable metric.", + "bbox": [ + 212, + 450, + 787, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate in both open- and closed-set settings. In closed-set settings, we use the Cumulative Matching Curve (CMC) at rank-1; for open-set, we use FNIR@FPIR at $1\\%$ and $0.1\\%$ FPIR and rank $R = 20$ .", + "bbox": [ + 214, + 546, + 460, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. For face recognition, we train on WebFace4M [80] and test on IJB-S [28]. Our experiments in gait recognition involve the Gait3D [77] and GREW [79] datasets, while for person reID, we employ the MEVID [8] and", + "bbox": [ + 214, + 672, + 460, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Summary of datasets used in our experiments. [Keys: Mod.=Modality; Img.=Open-set image recognition; Subj.=the number of subjects ; Media=the number of images for face/image recognition datasets and the number of sequences/tracklets for gait and video-based person reID datasets]", + "bbox": [ + 468, + 561, + 787, + 660 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/52784991c13ebb13c4b71009d3ce5c01e52099f124cb2d2eb9bf0ccaec739f99.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetYear Mod.TrainTest
Subj.MediaSubj.Media
IJB-S [28]2018Face-202>10M
Gait3D [77]2022Gait3,00018,9406,369
GREW [79]2021Gait20,000102,88724,000
MEVID [8]2023ReID1046,33855
CCVID [17]2022ReID75948151
", + "bbox": [ + 473, + 661, + 782, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CCVID [17] datasets. Further details are provided in Tab. 1.", + "bbox": [ + 214, + 779, + 651, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "One technical caveat is that although FNIR@1%FPIR implicitly assumes that each subject only appears once in the gallery, a typical gait recognition or person re-id dataset contains more than one gallery template per subject. To", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "address this issue, we take the mean feature over all gallery templates for each subject. This benefits the models compared to randomly sampling one gallery template (details are in the Supp.).", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We implement our proposed approach in PyTorch [49]. For fair comparisons, we follow the experiment setup of our baselines. For details on the hyperparameters, please refer to the Supp.", + "bbox": [ + 212, + 214, + 802, + 261 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 282, + 401, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct our ablation studies on the Gait3D [77] dataset using GaitBase [13] as the baseline. In the ablation studies, we hold out $20\\%$ of subjects in the training set as the validation set.", + "bbox": [ + 212, + 306, + 785, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of Det, Id and RTM.", + "text_level": 1, + "bbox": [ + 212, + 376, + 457, + 390 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We ablate the effect of Det, Id and RTM in Tab. 2. RTM consistently enhances both closed- and open-set performance. Id further improves open-set performance but has a minor impact on rank-1 accuracy. Meanwhile, Det significantly enhances both closed- and open-set performance. In summary, each component contributes to its effectiveness.", + "bbox": [ + 212, + 391, + 460, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of Hyperparameters.", + "text_level": 1, + "bbox": [ + 214, + 580, + 457, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We ablate the effect of hyperparameters $\\alpha, \\beta, \\gamma,$ and $\\lambda$ in Tab. 2. approach can consistently improve line, GaitBase [13]. Also, while the hyperparameters, our approach with the best set of tested hyperp", + "bbox": [ + 212, + 595, + 459, + 686 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/43a3dabde20b093603ab03588afeadecabf303b03214e392839aaf4cef512bab.jpg", + "table_caption": [ + "Table 2: Ablation study results. [Keys: $\\checkmark =$ Kept; $\\checkmark =$ Removed; R@1=rank-1 accuracy]" + ], + "table_footnote": [], + "table_body": "
IdDetRTMα,γβλR@1FNIR@1%FPIR
XXX———64.686.43±5.44
XX✓6.00.24.065.185.81±5.78
X✓✓64.485.16±5.13
✓✓✓3.00.24.063.985.22±4.73
9.064.284.08±6.28
✓✓✓6.00.14.063.185.10±5.67
0.463.885.92±5.45
✓✓✓6.00.22.064.883.12±5.94
8.061.685.67±5.15
✓✓✓6.00.24.064.882.99±4.50
", + "bbox": [ + 470, + 417, + 784, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison with State of the Art", + "text_level": 1, + "bbox": [ + 214, + 708, + 542, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Across biometric modalities, existing methods exhibit competitive closed-set recognition performance, as shown in Tab. 3 for face, Tab. 4 for gait, and Tab. 5 for person reID. In particular, in gait recognition, GaitBase [13] and SwinGait [12] showcase strong closed-set performance with a notable rank-1 accuracies of greater than $60\\%$ and $70\\%$ , respectively, for in-the-wild gait datasets Gait3D [77] and GREW [79]. However, in open-set scenarios, they face challenges, as evidenced by GaitBase's high FNIR of $86.43\\%$ at $1\\%$ FPIR in Gait3D.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 113, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fd8ff5b58ab5bcb981f700956a7345d17163847089f59cd0b354b6d03920e3ff.jpg", + "table_caption": [ + "Table 3: Closed- and open-set face recognition on two official protocols of IJB-S [28]. [Keys: R@1=rank-1 accuracy; 0.1%=FNIR@0.1%FPIR; 1%: FNIR@1%FPIR]" + ], + "table_footnote": [], + "table_body": "
ModelSurveillance-SingleSurveillance-Booking
R@1↑0.1%↓1%↓R@1↑0.1%↓1%↓
AdaFace [29]68.4445.1838.3768.8445.9837.80
AdaFace+Ours69.9244.3036.7370.0043.4836.17
", + "bbox": [ + 333, + 184, + 671, + 250 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "It is evident that the inclusion of our loss functions consistently yields improvements in both closed- and open-set performance across three modalities. In face recognition, we boost the rank-1 accuracy of AdaFace [29] to $69.92\\%$ $(+1.48\\%)$ in Surveillance-to-Single. In the challenging open-set scenario, our losses further demonstrate their effectiveness by reducing FNIR@0.1%FPIR from $45.98\\%$ to $43.48\\%$ in Surveillance-Booking.", + "bbox": [ + 212, + 279, + 784, + 369 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In gait recognition, when applied to Swin-Gait [12] backbone, our losses improve the rank-1 accuracy to $75.3\\%$ $(+2.9\\%)$ on Gait3D, and FNIR@1%FPIR decreases to $79.30\\%$ $(-4.20\\%)$ on Gait3D. In person reID, integrating our loss with CAL [17] reaches a rank-1 accuracy of $84.8\\%$ $(+2.2\\%)$ and a FNIR of $30.81\\%$ $(-3.08\\%)$ on MEVID [8]. We conduct the independent two-sample t-test to assess the statistical significance of perfor", + "bbox": [ + 212, + 372, + 431, + 611 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ac624f147c1e1c0c087629b8ea5d85ea9b34196a7b1718fd9ad266ce5613aaf6.jpg", + "table_caption": [ + "Table 4: Closed- and open-set gait recognition on Gait3D [77] and GREW [79]. SMPLGait [77] takes SMPL parameters as additional input. SwinGait-3D [12] uses our own implementation. [Keys: Rank@1=rank-1 accuracy]" + ], + "table_footnote": [], + "table_body": "
ModelRank@1 ↑FNIR@1%FPIR ↓
Gait3DGREWGait3DGREW
GaitSet [4]36.748.496.05±2.3081.17±1.45
GaitPart [14]28.247.696.88±1.8088.43±1.26
GaitGL [38]29.741.594.97±1.2480.55±1.11
SMPLGait [77]46.3—93.18±2.58—
GaitBase [13]64.660.186.43±5.4477.67±1.60
+Ours64.860.182.99±4.5076.55±1.54
SwinGait-3D [12]72.479.383.50±6.2058.32±1.79
+Ours75.380.479.30±6.7457.29±1.83
", + "bbox": [ + 446, + 458, + 784, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "mance improvements in open-set scenarios. For instance, our approach outperforms CAL on the CCVID [17] dataset, with a p-value of $0.052\\%$ .", + "bbox": [ + 212, + 614, + 782, + 643 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Collectively, these findings emphasize the consistent effectiveness of our proposed loss functions in enhancing both closed- and open-set performance across a spectrum of biometric tasks, ultimately improving the models' ability to handle open-set challenges, while maintaining strong closed-set performance.", + "bbox": [ + 212, + 645, + 784, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Efficiency. We present the floating point operations (FLOPs) of training baselines with and without our loss functions in Tab. 6. It is expected that incorporating our loss into baselines will increase the computational load in training, but the increase is very minor. For example, when SwinGait [12] is used, our loss functions only incur $46.9\\%$ of the FLOPs of the triplet loss, which is negligible compared to the model. Moreover, our approach does not incur any computation overhead during testing while improving closed- and open-set performance.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b4d4641ac96c79efb2ba14c16aac6284e943a1b55b15f80b9d09393519f81513.jpg", + "table_caption": [ + "Table 5: Closed- and open-set person reID on MEVID [8] and CCVID [17]." + ], + "table_footnote": [], + "table_body": "
ModelDatasetRank-1Accuracy ↑FNIR@1%FPIR ↓
AGRL [66]MEVID [8]48.466.67±9.57
TCLNet [21]MEVID [8]48.170.34±11.68
PSTA [63]MEVID [8]48.277.11±11.72
CAL [17]MEVID [8]52.566.95±11.59
+OursMEVID [8]55.164.80±10.71
CAL [17]CCVID [17]82.633.89±7.65
+OursCCVID [17]84.830.81±8.02
", + "bbox": [ + 313, + 170, + 687, + 282 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/686ba18ec30bace317a506ac35b29f802f6d35d02ef4fff028e3eb3901ced444.jpg", + "table_caption": [ + "Table 6: FLOPs in training different baseline models with and without our approach. We report the mean FLOPs across a minibatch. Our approach incurs no additional FLOP during testing." + ], + "table_footnote": [], + "table_body": "
ModalityModelFLOPS+Triplet+Ours
FaceAdaFace [29]24.19G—503K
GaitGaitBase [13]35.43G524.3K245.8K
GaitSwinGait [12]120G491.5K230.4K
BodyCAL [17]32.75G4.09K15.38K
", + "bbox": [ + 348, + 354, + 655, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Visualizations", + "text_level": 1, + "bbox": [ + 215, + 462, + 375, + 476 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Score Distributions. Fig. 5 visualizes the non-mated, maximum non-mated score for each probe, and genuine scores of SwinGait [12], with and without our loss. Our introduced RTM effectively reduces the distribution of maximum-per-probe scores, shifting it towards lower values. After linearly normalizing all scores between 0 and 1, the threshold at $1\\%$ FPIR is 0.61 with our loss (Fig. 5 (b)) compared to 0.69 without ours (Fig. 5 (a)). Moreover, our loss does not significantly alter the rightmost portion of the genuine score distribution, enhancing detec", + "bbox": [ + 212, + 505, + 433, + 806 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tion. This analysis offers insights into how our loss changes score distributions, and through which improves open-set performance.", + "bbox": [ + 212, + 806, + 784, + 838 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/88bea38f4110d2aaa25187068b72f57ff1d1f84bae263255a5d3755e50b27da5.jpg", + "image_caption": [ + "Fig.5: Score distributions on the test set of Gait3D [77] dataset. Ours (b) effectively reduces the threshold at $1\\%$ FPIR while only shifting the genuine score distribution marginally to the left compared to SwinGait [12] (a). Both distributions are shifted and scaled horizontally so that the min/max scores are within $[0,1]$ ." + ], + "image_footnote": [], + "bbox": [ + 444, + 537, + 782, + 681 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e40c633b2b5acffa1a07a4b2234ff56dc6b8a0048e7f6c1dcba3742e20b58b85.jpg", + "image_caption": [ + "(a) Face" + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 395, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/139e2e548bdfcefcdf8f42006380bd9c91aa26073c4d9770523905ddb18e25cb.jpg", + "image_caption": [ + "(b) Gait", + "Fig. 6: Open-set FNs due to detection, identification, and either. The average across the 50 open-set evaluation protocols is plotted. Blue indicates FNs that are shared by the baseline, SwinGait [12], and our approach. Orange indicates FNs of the baseline only. Green indicates FNs of our approach." + ], + "image_footnote": [], + "bbox": [ + 408, + 147, + 586, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3bb64effdd26f3bdd71d47e99941ed7a6c51b0e2f8ec97ec859cd1522e65ba8c.jpg", + "image_caption": [ + "(c) Person ReID" + ], + "image_footnote": [], + "bbox": [ + 602, + 148, + 779, + 231 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "False Negatives Breakdown. As illustrated in Fig. 6, across different biometric and generic tasks, almost all of the false negatives are failed detection, e.g. the $1^{\\mathrm{st}}$ mated probe in Fig. 1. A small portion of false negatives have both failed identification and detection, e.g. the $3^{\\mathrm{rd}}$ mated probe in Fig. 1. When we apply our loss functions, the number of FNs due to detection is reduced and the number of FNs due to identification is comparable to that of the baseline. These results highlight that our approach can improve open-set detection without sacrificing identification in various settings.", + "bbox": [ + 212, + 344, + 787, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 488, + 367, + 505 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our work addresses the critical gap in open-set biometrics, a practical generalization of the well-studied closed-set biometrics. We introduce novel I-DL and RTM losses aiming to improve biometric models' open-set performance. Through extensive experiments across diverse modalities, including face, gait, and person reID, we demonstrate the efficacy of our approach. We not only advance the state of the art in the challenging open-set domain but also enhance closed-set recognition capabilities. We advocate for increased attention to the open-set problem within the biometric community.", + "bbox": [ + 212, + 521, + 787, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations While our paper makes significant strides in addressing open-set challenges in biometrics through loss functions, our exploration does not extensively explore dedicated architectures for this specific context. Future research endeavors may benefit from investigating tailored architectures that complement and enhance the effectiveness of our proposed loss functions.", + "bbox": [ + 212, + 665, + 787, + 742 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Potential Societal Impacts. The use of biometric datasets in our work containing personally identifiable images prompts a crucial discussion on privacy and ethical implications. As biometric technologies become more prevalent, it is imperative to ensure that the collection of such datasets is aligned with ethical standards and privacy regulations.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via 2022-21102100004. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 291, + 321, + 306 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: CVPR (2016)", + "2. Cen, J., Luan, D., Zhang, S., Pei, Y., Zhang, Y., Zhao, D., Shen, S., Chen, Q.: The devil is in the wrongly-classified samples: Towards unified open-set recognition. In: ICLR (2023)", + "3. Chai, T., Li, A., Zhang, S., Li, Z., Wang, Y.: Lagrange motion analysis and view embeddings for improved gait recognition. In: CVPR (2022)", + "4. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: AAAI (2019)", + "5. Chen, C., Ye, M., Jiang, D.: Towards modality-agnostic person re-identification with descriptive query. In: CVPR (2023)", + "6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE TPAMI (2021)", + "7. Cui, Y., Kang, Y.: Multi-modal gait recognition via effective spatial-temporal feature fusion. In: CVPR (2023)", + "8. Davila, D., Du, D., Lewis, B., Funk, C., Van Pelt, J., Collins, R., Corona, K., Brown, M., McCloskey, S., Hoogs, A., et al.: MEVID: multi-view extended videos with identities for video person re-identification. In: WACV (2023)", + "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR (2019)", + "0. Dou, H., Zhang, P., Su, W., Yu, Y., Li, X.: Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In: ECCV (2022)", + "1. Dou, H., Zhang, P., Su, W., Yu, Y., Lin, Y., Li, X.: Gaitgci: Generative counterfactual intervention for gait recognition. In: CVPR (2023)", + "2. Fan, C., Hou, S., Huang, Y., Yu, S.: Exploring deep models for practical gait recognition. arXiv preprint arXiv:2303.03301 (2023)", + "3. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: CVPR (2023)", + "4. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: CVPR (2020)", + "5. Geng, C., Huang, S., Chen, S.: Recent advances in open set recognition: A survey. IEEE TPAMI (2020)", + "6. Grother, P., Ngan, M., Hanaoka, K.: Face recognition vendor test (FRVT) part 2: Identification (2019), https://www.nist.gov/document/nistir827120190911pdf", + "7. Gu, X., Chang, H., Ma, B., Bai, S., Shan, S., Chen, X.: Clothes-changing person re-identification with rgb modality only. In: CVPR (2022)", + "8. Gunther, M., Cruz, S., Rudd, E.M., Boult, T.E.: Toward open-set face recognition. In: CVPR Workshops (2017)" + ], + "bbox": [ + 225, + 323, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Günther, M., Hu, P., Herrmann, C., Chan, C.H., Jiang, M., Yang, S., Dhamija, A.R., Ramanan, D., Beyerer, J., Kittler, J., et al.: Unconstrained face detection and open-set face recognition challenge. In: IJCB. IEEE (2017)", + "20. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)", + "21. Hou, R., Chang, H., Ma, B., Shan, S., Chen, X.: Temporal complementary learning for video person re-identification. In: ECCV (2020)", + "22. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE TPAMI (2022)", + "23. Huang, X., Zhu, D., Wang, H., Wang, X., Yang, B., He, B., Liu, W., Feng, B.: Context-sensitive temporal feature learning for gait recognition. In: ICCV (2021)", + "24. Huang, Y., Shen, P., Tai, Y., Li, S., Liu, X., Li, J., Huang, F., Ji, R.: Improving face recognition from hard samples via distribution distillation loss. In: ECCV (2020)", + "25. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: CVPR (2020)", + "26. Huang, Z., Xue, D., Shen, X., Tian, X., Li, H., Huang, J., Hua, X.S.: 3d local convolutional neural networks for gait recognition. In: ICCV (2021)", + "27. Jain, A.K., Deb, D., Engelsma, J.J.: Biometrics: Trust, but verify. IEEE Transactions on Biometrics, Behavior, and Identity Science (2021)", + "28. Kalka, N.D., Maze, B., Duncan, J.A., O'Connor, K., Elliott, S., Hebert, K., Bryan, J., Jain, A.K.: IJB-S: IARPA janus surveillance video benchmark. In: BTAS (2018)", + "29. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. In: CVPR (2022)", + "30. Kim, M., Liu, F., Jain, A.K., Liu, X.: Cluster and aggregate: Face recognition with large probe set. In: NeurIPS (2022)", + "31. Kim, M., Su, Y., Liu, F., Jain, A., Liu, X.: Keypoint relative position encoding for face recognition. In: CVPR (2024)", + "32. Koch, T., Riess, C., Kohler, T.: Lord: Leveraging open-set recognition with unknown data. In: ICCV (2023)", + "33. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: ICCV (2021)", + "34. Kuchibhotla, H.C., Malagi, S.S., Chandhok, S., Balasubramanian, V.N.: Unseen classes at a later time? no problem. In: CVPR (2022)", + "35. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: ICCV (2023)", + "36. Li, W., Hou, S., Zhang, C., Cao, C., Liu, X., Huang, Y., Zhao, Y.: An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In: CVPR (2023)", + "37. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: ECCV (2022)", + "38. Lin, B., Zhang, S., Yu, X.: Gait recognition via effective global-local feature representation and local temporal aggregation. In: ICCV (2021)", + "39. Liu, F., Ashbaugh, R., Chimitt, N., Hassan, N., Hassani, A., Jaiswal, A., Kim, M., Mao, Z., Perry, C., Ren, Z., et al.: Farsight: A physics-driven whole-body biometric system at large distance and altitude. In: WACV (2024)", + "40. Liu, F., Kim, M., Gu, Z., Jain, A., Liu, X.: Learning clothing and pose invariant 3d shape representation for long-term person re-identification. In: ICCV (2023)", + "41. Liu, F., Kim, M., Jain, A., Liu, X.: Controllable and guided face synthesis for unconstrained face recognition. In: ECCV (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Liu, F., Kim, M., Ren, Z., Liu, X.: Distilling CLIP with dual guidance for learning discriminative human body shape representation. In: CVPR (2024)", + "43. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: CVPR (2017)", + "44. Ma, K., Fu, Y., Zheng, D., Cao, C., Hu, X., Huang, Y.: Dynamic aggregated network for gait recognition. In: CVPR (2023)", + "45. Mekhazni, D., Dufau, M., Desrosiers, C., Pedersoli, M., Granger, E.: Camera alignment and weighted contrastive learning for domain adaptation in video person reid. In: WACV (2023)", + "46. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: CVPR (2021)", + "47. Minaee, S., Abdelrashidi, A., Su, H., Bennamoun, M., Zhang, D.: Biometrics recognition using deep learning: A survey. Artificial Intelligence Review (2023)", + "48. Ni, H., Li, Y., Gao, L., Shen, H.T., Song, J.: Part-aware transformer for generalizable person re-identification. In: ICCV (2023)", + "49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)", + "50. Qin, L., Wang, M., Deng, C., Wang, K., Chen, X., Hu, J., Deng, W.: Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. TCSVT (2023)", + "51. Ren, Z., Su, Y., Liu, X.: Chatgpt-powered hierarchical comparisons for image classification. In: NeurIPS (2023)", + "52. Ross, A., Banerjee, S., Chen, C., Chowdhury, A., Mirjalili, V., Sharma, R., Swearinen, T., Yadav, S.: Some research problems in biometrics: The future beckons. In: ICB (2019)", + "53. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: CVPR (2015)", + "54. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3d gait recognition with point clouds. In: CVPR (2023)", + "55. Shen, C., Yu, S., Wang, J., Huang, G.Q., Wang, L.: A comprehensive survey on deep gait recognition: algorithms, datasets and challenges. arXiv preprint arXiv:2206.13732 (2022)", + "56. Somers, V., De Vleeschouwer, C., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: CVPR (2023)", + "57. Sun, Y., Cheng, C., Zhang, Y., Zhang, C., Zheng, L., Wang, Z., Wei, Y.: Circle loss: A unified perspective of pair similarity optimization. In: CVPR (2020)", + "58. Tabassi, E., Watson, C., Fiumara, G., Salamon, W., Flanagan, P., Cheng, S.L.: Performance evaluation of fingerprint open-set identification algorithms. In: IJCB (2014)", + "59. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need. In: ICLR (2022)", + "60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: CVPR (2018)", + "61. Wang, L., Liu, B., Liang, F., Wang, B.: Hierarchical spatio-temporal representation learning for gait recognition. In: ICCV (2023)", + "62. Wang, M., Guo, X., Lin, B., Yang, T., Zhu, Z., Li, L., Zhang, S., Yu, X.: Dygait: Exploiting dynamic representations for high-performance gait recognition. In: ICCV (2023)", + "63. Wang, Y., Zhang, P., Gao, S., Geng, X., Lu, H., Wang, D.: Pyramid spatial-temporal aggregation for video-based person re-identification. In: ICCV (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Open-Set Biometrics: Beyond Good Closed-Set Models", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards AUC-oriented open-set recognition. In: NeurIPS (2022)", + "65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: ECCV (2016)", + "66. Wu, Y., Bourahla, O.E.F., Li, X., Wu, F., Tian, Q., Zhou, X.: Adaptive graph representation learning for video person re-identification. TIP (2020)", + "67. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. In: NeurIPS (2022)", + "68. Ye, D., Fan, C., Ma, J., Liu, X., Yu, S.: Biggait: Learning gait representation you want by large vision models. In: CVPR (2024)", + "69. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: CVPR (2019)", + "70. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: CVPR (2020)", + "71. Zahra, A., Perwaiz, N., Shahzad, M., Fraz, M.M.: Person re-identification: A retrospective on domain specific open challenges and future trends. Pattern Recognition (2023)", + "72. Zhai, Y., Liu, Z., Wu, Z., Wu, Y., Zhou, C., Doermann, D., Yuan, J., Hua, G.: Soar: Scene-debiasing open-set action recognition. In: ICCV (2023)", + "73. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: ECCV (2020)", + "74. Zhang, Z., Tran, L., Liu, F., Liu, X.: On learning disentangled representations for gait recognition. IEEE TPAMI (2020)", + "75. Zhang, Z., Tran, L., Yin, X., Atoum, Y., Liu, X., Wan, J., Wang, N.: Gait recognition via disentangled representation learning. In: CVPR (2019)", + "76. Zheng, J., Liu, X., Gu, X., Sun, Y., Gan, C., Zhang, J., Liu, W., Yan, C.: Gait recognition in the wild with multi-hop temporal switch. In: ACMMM (2022)", + "77. Zheng, J., Liu, X., Liu, W., He, L., Yan, C., Mei, T.: Gait recognition in the wild with dense 3d representations and a benchmark. In: CVPR (2022)", + "78. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: CVPR (2021)", + "79. Zhu, Z., Guo, X., Yang, T., Huang, J., Deng, J., Huang, G., Du, D., Lu, J., Zhou, J.: Gait recognition in the wild: A benchmark. In: ICCV (2021)", + "80. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., et al.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: CVPR (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 674 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Su et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_model.json b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6ca5654561330f6f2c9271a49c9f7523d54c1cc0 --- /dev/null +++ b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_model.json @@ -0,0 +1,3316 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.773, + 0.186 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.213, + 0.761, + 0.229 + ], + "angle": 0, + "content": "Yiyang Su, Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.24, + 0.703, + 0.27 + ], + "angle": 0, + "content": "Michigan State University, East Lansing, MI 48824, USA {suyiyan1, kimminc2, liufeng6, jain, liuxm}@msu.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.309, + 0.745, + 0.571 + ], + "angle": 0, + "content": "Abstract. Biometric recognition has primarily addressed closed-set identification, assuming all probe subjects are in the gallery. However, most practical applications involve open-set biometrics, where probe subjects may or may not be present in the gallery. This poses distinct challenges in effectively distinguishing individuals in the gallery while minimizing false detections. While it is commonly believed that powerful biometric models can excel in both closed- and open-set scenarios, existing loss functions are inconsistent with open-set evaluation. They treat genuine (mated) and imposter (non-mated) similarity scores symmetrically and neglect the relative magnitudes of imposter scores. To address these issues, we simulate open-set evaluation using minibatches during training and introduce novel loss functions: (1) the identification-detection loss optimized for open-set performance under selective thresholds and (2) relative threshold minimization to reduce the maximum negative score for each probe. Across diverse biometric tasks, including face recognition, gait recognition, and person re-identification, our experiments demonstrate the effectiveness of the proposed loss functions, significantly enhancing open-set performance while positively impacting closed-set performance. Our code and models are available here." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.586, + 0.741, + 0.613 + ], + "angle": 0, + "content": "Keywords: Open-set biometrics \\(\\cdot\\) Face recognition \\(\\cdot\\) Gait recognition \\(\\cdot\\) Person reID" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.641, + 0.377, + 0.657 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Biometrics, the study of unique physiological or behavioral characteristics for human identification, plays a pivotal role in applications such as access control, surveillance, and forensic analysis [52]. Each biometric task, from face recognition, gait recognition to person re-identification (reID), uses features tailored to its specific requirements [39, 55, 71]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In real-world applications, after test subjects are enrolled in the gallery, biometric systems inevitably encounter non-mated subjects, those not present in the gallery. Thus, it becomes imperative for such systems to demonstrate robust performance in the open-set framework with non-mated probes [16, 18, 28, 58]. The closed-set framework assumes all probe subjects are mated and predominantly focuses on identification, associating probes with the closest match in the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.79, + 0.432 + ], + "angle": 0, + "content": "Fig. 1: A toy example for FNIR@FPIR calculation. The score matrix and distribution between non-mated probes and all gallery subjects are shown in blue. The maximum score for each non-mated probe is shown in red. A percentile of their distribution determines the threshold. A user-defined percentile (e.g., \\(33\\%\\)) sets the threshold to be 0.75 as 1 out of 3 (the \\(2^{\\mathrm{nd}}\\)) non-mated probes is incorrectly classified as \"in the gallery\". The genuine scores of mated probes are shown in green. A mated probe is an FN if either of the conditions is met: (1) in detection (Det.), it is incorrectly classified as \"not in the gallery\" (\\(\\mathcal{G}\\)) as the genuine score is below the threshold (the \\(1^{\\mathrm{st}}\\) mated probe); (2) its genuine score is not within a specified rank \\(R\\), e.g., 1, in the gallery (the \\(2^{\\mathrm{nd}}\\) mated probe). Hence, FNIR@33%FPIR is \\(75\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.583 + ], + "angle": 0, + "content": "gallery. However, the open-set framework introduces an additional challenge: detection, determining whether a probe subject is mated. For example, in airport surveillance, faces in CCTV videos are compared against a criminal watch list. The goal is to reduce the misidentification (false positives) of non-mated subjects such as airport staff and innocent passengers (improving detection), while simultaneously ensuring accurate recognition of known criminals (enhancing identification). For this reason, open-set biometrics is a harder and more general task than closed-set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.704 + ], + "angle": 0, + "content": "The biometric community has predominantly centered on the closed-set framework. While some face recognition benchmarks have open-set evaluation protocols [18,19,28], they are primarily used for evaluation rather than for developing tailored strategies to address the open-set challenges. In gait recognition and person reID, most benchmarks only provide official closed-set protocols, lacking open-set ones. To address this gap, we take the initiative to introduce open-set benchmarks and propose strategies to train models specifically for improving open-set performances while maintaining closed-set performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The false negative identification rate at a given false positive identification rate (FNIR@FPIR) is a de facto open-set metric [16]. We first compute the similarity score matrix between non-mated probes and the gallery, from which we extract the maximum score for each non-mated probe across the gallery, forming the maximum-per-probe non-mated score distribution. This distribution represents the likelihood that a non-mated probe has a match in the gallery. A threshold is set based on a user-defined percentile (e.g., \\(1\\%\\)) of this distribution. Mated probes with scores below this threshold are counted as false negatives (FNs), as they are incorrectly considered not to be in the gallery. Similar to the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "closed-set scenario, a mated probe is also an FN if its gallery match doesn't rank among the gallery set within a user-specified rank \\( R \\). FNIR is the proportion of mated probes classified as FNs. Fig. 1 depicts its calculation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.785, + 0.315 + ], + "angle": 0, + "content": "Unlike closed-set or verification metrics, FNIR@FPIR introduces unique challenges. It employs a threshold based on maximum scores for each non-mated probe, creating an asymmetry between genuine and imposter scores. SOTA biometric recognition methods achieve strong performance in closed-set scenarios leveraging triplet loss [20] and/or softmax cross-entropy based losses [9, 25, 29, 46, 60], both of which excel in reducing intra-person variance and increasing inter-person variance. However, they do not take into account the fact that the non-mated score distribution determines the threshold for the false negative." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.787, + 0.452 + ], + "angle": 0, + "content": "In other words, conventional loss functions, while effective in closed-set scenarios, exhibit limitations in addressing the nuances of the FNIR@FPIR metric, which necessitate a handling of the maximum of imposter scores. These high-scoring non-mated probes disproportionately influence the threshold setting for false negatives, yet conventional loss functions often treat them equivalently to other scores. As a result, they face challenges in open-set scenarios. GaitBase [13], a SOTA gait recognition model, can only achieve a \\(61\\%\\) FNIR@1%FPIR on the training set and \\(86\\%\\) on the test set. This challenge extends to face recognition and person reID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.636 + ], + "angle": 0, + "content": "To address this challenge, we propose to create simulated open-set evaluation episodes during training, whose score distributions serve as proxies to those of the test set. Specifically, we introduce an identification-detection loss tailored to optimize the two core elements in open-set biometrics, i.e., detection and identification. For detection, the model is trained to accurately pinpoint the exact probes whose mate is present in the gallery set; for identification, the model learns to establish associations between the gallery and probe samples of mated subjects. By integrating the aspects of detection and identification within training batches, we improve the model's efficacy in both domains, thus yielding enhanced testing performance. Further, we minimize similarity scores between non-mated subjects and the gallery, enabling a more comprehensive detection of mated subjects (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.785, + 0.699 + ], + "angle": 0, + "content": "Empirically, we apply our proposed loss function to three biometric modalities, face, gait, and person reID, and show performance improvement in all tasks. While our focus is open-set biometrics, we also observe performance gain in closed-set search due to improved feature representations." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.701, + 0.538, + 0.715 + ], + "angle": 0, + "content": "To summarize, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.733, + 0.784, + 0.791 + ], + "angle": 0, + "content": "\\(\\checkmark\\) We show that the existing loss functions, e.g., triplet loss and softmax loss, do not treat genuine and imposter training samples in accordance with open-set evaluation. This inconsistency sheds light on why the open-set challenges have not been adequately addressed." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.795, + 0.784, + 0.837 + ], + "angle": 0, + "content": "\\(\\checkmark\\) We propose a novel loss function that leverages open-set mini-batches during training to optimize identification and detection performance and minimize open-set detection thresholds." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.733, + 0.784, + 0.837 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.143, + 0.341, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.183, + 0.35, + 0.206 + ], + "angle": 0, + "content": "Training Testing classes" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.207, + 0.338, + 0.232 + ], + "angle": 0, + "content": "Closed-Set Recognition" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.144, + 0.473, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.183, + 0.471, + 0.206 + ], + "angle": 0, + "content": "Training Testing classes" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.207, + 0.46, + 0.232 + ], + "angle": 0, + "content": "Open-Set Recognition" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.147, + 0.528, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.495, + 0.183, + 0.542, + 0.206 + ], + "angle": 0, + "content": "Training subjects" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.144, + 0.611, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.183, + 0.604, + 0.206 + ], + "angle": 0, + "content": "Testing subjects" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.207, + 0.582, + 0.231 + ], + "angle": 0, + "content": "Closed-Set Biometrics" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.149, + 0.667, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.183, + 0.68, + 0.206 + ], + "angle": 0, + "content": "Training subjects" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.149, + 0.755, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.183, + 0.744, + 0.206 + ], + "angle": 0, + "content": "Testing subjects" + }, + { + "type": "text", + "bbox": [ + 0.658, + 0.207, + 0.722, + 0.231 + ], + "angle": 0, + "content": "Open-Set Biometrics" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 2: Comparison between open-set recognition and biometrics. Closed-set recognition has the same set of classes for training and testing. Open-set recognition involves testing on unseen classes during training. On the other hand, in biometrics, the sets of training and testing subjects are always disjoint. In closed-set biometrics, all probe subjects are in the gallery, and in open-set biometrics, some probe subjects are not enrolled in the gallery. [Legends: ▪ gallery subjects, ▪ probe subjects]" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.364, + 0.785, + 0.41 + ], + "angle": 0, + "content": "\\(\\checkmark\\) We experimentally demonstrate that our proposed loss function can effectively improve both closed- and open-set performance across diverse biometric modalities, including face, gait, and person reID." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.429, + 0.388, + 0.445 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Open-Set Biometrics. While various biometric recognition tasks, such as face recognition [9, 24, 25, 29-31, 41, 43, 46, 50, 60], gait recognition [3, 4, 7, 10-14, 23, 26, 36-38, 44, 54, 61, 62, 68, 74-77], and person reID [5, 17, 21, 35, 36, 40, 42, 45, 48, 56, 63, 66], have witnessed significant progress, the majority of their focus remains on closed-set scenarios. Although EVM has been applied to open-set face recognition [18], EVM takes extracted features, instead of images/videos, as input. Our approach improves feature representation and EVM improves score calculation. While some face recognition datasets provide opens-set evaluation protocols [28], open-set performance is often perceived as a consequence rather than a direct objective in advancing recognition models. Further, even the most recent gait recognition [68] and person reID [42] models are exclusively evaluated under closed-set settings due to the absence of open-set evaluation protocols." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Open-Set Recognition. While open-set biometrics is relatively underexplored, open-set recognition [1,2,6,22,32-34,51,59,64,67,69,72,73,78] has attracted attention in the literature. Open-set recognition distinguishes between known and unknown classes in image classification. In both open-set image recognition and open-set biometrics, the primary concern is how the models deal with instances that belong to unknown classes or non-mated probes. However, as illustrated in Fig. 2, in open-set recognition known classes come from the training set and the unknown classes refer to classes not used for training [15]. On the other hand, even in closed-set biometrics, all test subjects are typically not used during training and mated probes belong to test subjects enrolled in the gallery [27,47]. Open-set biometrics differs from closed-set biometrics in that there are non-mated subjects are neither in the training set nor in the gallery [16,18,58]. In" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.145, + 0.744, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.237, + 0.788, + 0.293 + ], + "angle": 0, + "content": "Fig. 3: Overview of our approach. We introduce an additional modality-agnostic loss function specifically designed for open-set biometrics, which simultaneously optimizes identification (Id.) and detection (Det.) and Relatie Threshold Minimization (RTM) for open-set thresholds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.784, + 0.37 + ], + "angle": 0, + "content": "evaluation, open-set recognition models typically use closed-set recognition accuracy and AUROC for detecting classes not seen during training [15] and open-set biometrics uses FNIR@FPIR [16]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.393, + 0.331, + 0.409 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.439, + 0.441 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Let \\( S_{\\mathrm{test}} \\) be the test set, \\( \\mathcal{G} \\subset S_{\\mathrm{test}} \\) represent the gallery set containing mated subjects, and \\( \\mathcal{P} \\subset S_{\\mathrm{test}} \\) denote the probe set consisting of test samples. For each probe sample \\( p_i \\in \\mathcal{P} \\), there are two possible outcomes: (1) If \\( p_i \\) is mated, i.e., its corresponding subject \\( g_i \\) is in \\( \\mathcal{G} \\), the biometric model must identify the correct subject. (2) If \\( p_i \\) is non-mated, the model must reject it. The open-set scenario poses a more challenging problem than the traditional closed-set one, as the model needs to recognize mated subjects while rejecting non-mated ones." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.583, + 0.598, + 0.598 + ], + "angle": 0, + "content": "3.2 Revisiting Loss Functions for Biometrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.787, + 0.673 + ], + "angle": 0, + "content": "In this section, we compute gradients of two common loss functions in biometrics: softmax and triplet loss. Many loss functions are derived from them and share similar characteristics. For instance, center loss [65] can be analyzed similarly to softmax loss, and the logic of triplet loss can also be applied to circle loss [57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.755 + ], + "angle": 0, + "content": "Softmax loss. The softmax loss, often employed in closed-set classification, operates by computing the cross-entropy between the normalized logit of a sample and its corresponding ground truth class. Mathematically, for a sample \\( i \\) in a batch, the softmax loss \\( L_{\\mathrm{softmax}}(i) \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.766, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s o f t m a x}} (i) = - \\log \\left(\\frac {e ^ {z _ {i , s}}}{\\sum_ {t} e ^ {z _ {i , t}}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Here, \\( z_{i,s} \\) denotes the logit for sample \\( i \\) belonging to subject \\( s \\). This loss aims to minimize the negative log-likelihood of the true subject \\( s \\) for sample \\( i \\). For one" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.729, + 0.163 + ], + "angle": 0, + "content": "sample, the gradient of the softmax loss \\( w.r.t. \\) the logit \\( z_{i,s} \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.175, + 0.786, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L _ {\\mathrm {s o f t m a x}} (i)}{\\partial z _ {i , s}} = \\frac {e ^ {z _ {i , s}}}{\\sum_ {t} e ^ {z _ {i , t}}} - \\delta_ {i, s}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.218, + 0.787, + 0.269 + ], + "angle": 0, + "content": "where \\(\\frac{e^{z_{i,s}}}{\\sum_{j=1}^{N} e^{z_{j,s}}}\\) represents the probability that sample \\(i\\) belongs to subject \\(s\\) and \\(\\delta_{i,c}\\) is the Kronecker delta, which is 1 when \\(i = c\\) (the true subject) and 0 otherwise. We visualize gradients of the softmax loss in Fig. 4(a)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.271, + 0.789, + 0.377 + ], + "angle": 0, + "content": "However, it overlooks the relative magnitude of each imposter score within the batch. An imposter score may be deemed small by, e.g., AdaFace [29] standards but still rank high among maximum-per-probe non-mated scores, raising the open-set threshold. Further, the softmax loss assumes that each sample can be assigned to one of the training subjects, which is not the case in open-set settings. Consequently, the softmax loss is not inherently designed to handle open-set biometrics." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.398, + 0.564, + 0.414 + ], + "angle": 0, + "content": "Triplet loss. The triplet loss [53] is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.426, + 0.786, + 0.442 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {t r i}} (a, p, n) = \\max (0, d (a, p) - d (a, n) + m). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.453, + 0.788, + 0.5 + ], + "angle": 0, + "content": "Note that \\( d(a, p) \\) is the distance between the anchor \\( a \\) and positive sample \\( p \\), \\( d(a, n) \\) is between \\( a \\) and the negative sample \\( n \\), and \\( m \\) is a constant margin. The partial derivatives of \\( L(a, p, n) \\) w.r.t. \\( d(a, p) \\) and \\( d(a, n) \\) are piecewise functions:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.523, + 0.786, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L _ {\\operatorname {t r i}} (a , p , n)}{\\partial d (a , p)} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} d (a, p) + m > d (a, n) \\\\ 0, & \\text {i f} d (a, p) + m \\leq d (a, n) \\end{array} ; \\right. \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.566, + 0.787, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L _ {\\operatorname {t r i}} (a , p , n)}{\\partial d (a , n)} = \\left\\{ \\begin{array}{l l} - 1, & \\text {i f} d (a, p) + m > d (a, n) \\\\ 0, & \\text {i f} d (a, p) + m \\leq d (a, n) \\end{array} \\right.. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.614, + 0.788, + 0.764 + ], + "angle": 0, + "content": "The magnitude of the gradient is constant regardless of the distance between two samples. This inability to differentiate samples in the gradient implies that all negative samples in the triplet are contributing equally, which is not reflective of the open-set scenario. One remedy often used with triplet loss is hard negative mining. However, it comes with a critical drawback of being susceptible to label noise [70]. Noisy and mislabeled samples are often considered hard negatives, which can make the loss function sensitive to noisy data when focusing on the hardest samples. Thus, we believe that a loss function that considers all negative samples while being able to adjust the gradient magnitude based on the anchor-negative distances would be effective in open-set biometrics." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Hence, we propose a novel loss function, shown in Fig. 3. It optimizes identification and detection at various thresholds and minimizes non-mated scores to reduce open-set thresholds. It can be applied to any biometric modality. Unlike existing loss functions, it is consistent with FNIR@FPIR in that it treats genuine and imposter scores asymmetrically and tackles large non-mated scores." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.148, + 0.292, + 0.157 + ], + "angle": 0, + "content": "(a) SoftMax" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.16, + 0.307, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.226, + 0.226, + 0.317, + 0.235 + ], + "angle": 0, + "content": "\\(\\oplus\\) Genuine probe \\((p)\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.226, + 0.235, + 0.317, + 0.244 + ], + "angle": 0, + "content": "\\(\\ominus\\) Imposter probe \\((n)\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.226, + 0.244, + 0.283, + 0.254 + ], + "angle": 0, + "content": "Threshold" + }, + { + "type": "image_footnote", + "bbox": [ + 0.226, + 0.254, + 0.278, + 0.263 + ], + "angle": 0, + "content": "\\(\\rightarrow\\) Gradient" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.226, + 0.317, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.266, + 0.317, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.279, + 0.317, + 0.288 + ], + "angle": 0, + "content": "Gradient Magnitude" + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.149, + 0.401, + 0.157 + ], + "angle": 0, + "content": "(b) Triplet Loss" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.16, + 0.406, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.222, + 0.405, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.149, + 0.496, + 0.157 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.161, + 0.532, + 0.17 + ], + "angle": 0, + "content": "Identification Detection" + }, + { + "type": "image_footnote", + "bbox": [ + 0.439, + 0.176, + 0.524, + 0.202 + ], + "angle": 0, + "content": "different \\(p_i\\) gradient magnitude based on threshold" + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.23, + 0.523, + 0.248 + ], + "angle": 0, + "content": "Relative Threshold Minimization" + }, + { + "type": "image_footnote", + "bbox": [ + 0.439, + 0.253, + 0.523, + 0.281 + ], + "angle": 0, + "content": "different \\(n_j\\) gradient magnitude based on relative distance" + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.146, + 0.622, + 0.155 + ], + "angle": 0, + "content": "\\(p_i\\) Most Emphasized" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.155, + 0.621, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.219, + 0.611, + 0.227 + ], + "angle": 0, + "content": "\\(n_2\\) Emphasized" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.227, + 0.621, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.147, + 0.692, + 0.155 + ], + "angle": 0, + "content": "\\(p_i\\) Emphasized" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.155, + 0.702, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.219, + 0.696, + 0.227 + ], + "angle": 0, + "content": "\\(n_1,n_2\\) Emphasized" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.227, + 0.702, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.147, + 0.782, + 0.155 + ], + "angle": 0, + "content": "\\(p_i\\) Less Emphasized" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.156, + 0.783, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.219, + 0.772, + 0.227 + ], + "angle": 0, + "content": "\\(n_1\\) Emphasized" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.227, + 0.784, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.453 + ], + "angle": 0, + "content": "Fig. 4: Visualization of the gradient of different loss functions. The colors represent gradient magnitude and the arrows represent gradient directions. a) SoftMax loss pushes the sample toward the class center and away from the other classes, with gradual shifts in gradient magnitude. b) On the other hand, triplet loss has a hard cut gradient based on triplet samples. c) For \\( S_{\\tau}^{det} \\), as shown in the top row, the thresholds are determined by non-mated sample, \\( n_j \\). \\( p_i \\)'s gradient has the greatest magnitude when it has a similar distance from the gallery \\( g_i \\) to \\( n_j \\). For Relative Threshold Minimization, as shown in the bottom row, as \\( n_2 \\) moves away from the gallery, its gradient decreases. While \\( n_1 \\) remains at the same location, its gradient increases because it becomes closer to \\( g_i \\) than \\( n_2 \\). The gradients w.r.t. genuine scores adapt to non-mated scores and the gradients w.r.t. non-mated scores are adaptive to other non-mated scores." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.382, + 0.498 + ], + "angle": 0, + "content": "3.3 Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.687 + ], + "angle": 0, + "content": "To design loss functions tailored for open-set biometrics, we create a training scenario mirroring the testing environment. During training, we partition each batch into gallery and probe sets to simulate open-set evaluation conditions. In this process, we randomly designate a certain percentage, denoted as \\( p\\% \\), of the subjects as mated subjects, while the rest are categorized as non-mated. For the mated subjects, we further randomly allocate their exemplars to the gallery and probe sets. Conversely, all exemplars of the non-mated subjects are placed in the probe set. This approach yields a more realistic and representative training scenario, preparing the biometric features to excel under open-set conditions. To avoid any potential notation ambiguity, we refer to the mated probe set created within each training batch as \\( \\mathcal{P}_{\\mathrm{K}}^{\\prime} \\), the non-mated probe set as \\( \\mathcal{P}_{\\mathrm{U}}^{\\prime} \\), and the gallery sets within each training batch as \\( \\mathcal{G}^{\\prime} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.787, + 0.732 + ], + "angle": 0, + "content": "To ensure robustness, when the metric space inhabited by the biometric features is defined under the Euclidean distance, we calculate the similarity scores between two features as:" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.739, + 0.786, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathrm {s} \\left(p _ {i}, g _ {j}\\right) = \\frac {1}{1 + \\mathrm {d} \\left(p _ {i} , g _ {j}\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.81 + ], + "angle": 0, + "content": "where \\( \\mathrm{d}(\\cdot, \\cdot) \\) represents the Euclidean distance. Alternatively, if the metric space is defined under the cosine distance, we simply adopt the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To ensure that the model can generalize to open-set scenarios during testing, we propose loss functions that penalize the model for three types of errors during" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "training: (1) failing to detect a mated probe with a threshold \\(\\tau\\), (2) failing to identify a mated probe within the top rank-\\(r\\) positions and (3) setting the threshold \\(\\tau\\) too high, leading to poor differentiation between mated and non-mated sample distributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.788, + 0.273 + ], + "angle": 0, + "content": "(1) Detection. Detection is assessing whether the computed similarity score exceeds a predefined threshold \\(\\tau\\). That is, for a given probe \\(p_i \\in \\mathcal{P}_{\\mathrm{K}}'\\) and its corresponding gallery subject \\(g_i \\in \\mathcal{G}'\\), we calculate a detection score as:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.282, + 0.786, + 0.3 + ], + "angle": 0, + "content": "\\[\nS _ {\\tau} ^ {d e t} \\left(p _ {i}, g _ {i}\\right) = \\sigma_ {\\alpha} \\left(\\mathrm {s} \\left(p _ {i}, g _ {i}\\right) - \\tau\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.429 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\alpha}(x) = 1 / (1 + \\exp (-\\alpha x))\\) is the Sigmoid function with temperature hyperparameter \\(\\alpha\\). The \\(S_{\\tau}^{det}(p_i, g_i)\\) score quantifies the model's ability to detect if \\(p_i\\) and \\(g_i\\) belong to the same subject, given the threshold \\(\\tau\\). A high similarity score, well above \\(\\tau\\), indicates successful detection and \\(S_{\\tau}^{det}(p_i, g_i)\\) approaches 1. In contrast, if the model fails to detect that \\(p_i\\) and \\(g_i\\) belong to the same subject, \\(s(p_i, g_i)\\) remains much smaller than \\(\\tau\\), causing \\(S_{\\tau}^{det}(p_i, g_i)\\) to approach 0. Since the Sigmoid function has the highest gradient near 0, the term \\(\\mathrm{s}(p_i, g_i) - \\tau\\) makes the loss function focus more on the samples near the threshold \\(\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.474 + ], + "angle": 0, + "content": "During testing, the open-set detection threshold is determined by the non-mated score distribution, which we do not know during training. Thus, we sample \\(\\tau\\) from the non-mated scores in a batch. Specifically, we let" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.483, + 0.786, + 0.518 + ], + "angle": 0, + "content": "\\[\nS ^ {d e t} = \\frac {1}{| \\mathcal {T} |} \\sum_ {\\tau \\in \\mathcal {T}} S _ {\\tau} ^ {d e t} \\left(p _ {i}, g _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.648 + ], + "angle": 0, + "content": "where \\(\\mathcal{T} = \\{s(n_j,g_i)|n_j\\in \\mathcal{P}_{\\mathrm{U}}'\\}\\) is the set of all non-mated scores. Fig. 4(c) illustrates the gradient of Det w.r.t. mated probes under different thresholds induced by various non-mated probes. The gradient is adaptable to different thresholds, increasing as \\(\\tau\\) decreases, even when \\(s(g,p)\\) is constant. This adaptability is desirable in open-set settings, as it promotes mated probes to be more similar to their ground-truth gallery than non-mated probes. Note that the thresholds in Eq. (8) are only used during training. The detection thresholds during testing are automatically determined by the FNIR@FPIR metric as discussed in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.729 + ], + "angle": 0, + "content": "(2) Identification. It is not sufficient for a biometric model to distinguish mated subjects from non-mated ones based on their scores. Equally important is the model's ability to accurately identify the correct subject among the gallery subjects. Therefore, we introduce an identification score \\( S^{id}(p_i, g_i) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.738, + 0.786, + 0.755 + ], + "angle": 0, + "content": "\\[\nS ^ {i d} \\left(p _ {i}, g _ {i}\\right) = \\sigma_ {\\beta} \\left(1 - \\operatorname {s o f t r a n k} \\left(p _ {i}, g _ {i}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.794 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\beta}\\) is a Sigmoid function with temperature \\(\\beta\\) and sofrank is a differentiable function with a Sigmoid function \\(\\sigma_{\\gamma}\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.805, + 0.786, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s o f t r a n k} \\left(p _ {i}, g _ {i}\\right) = \\sum_ {g _ {j} \\in \\mathcal {G} ^ {\\prime}} \\sigma_ {\\gamma} \\left(\\mathrm {s} \\left(p _ {i}, g _ {j}\\right) - \\mathrm {s} \\left(p _ {i}, g _ {i}\\right)\\right), \\tag {10}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "The softrank function accumulates values close to 1 for gallery subjects \\( g_{j} \\) more similar to the probe \\( p_{i} \\) than the correct match \\( g_{i} \\) and close to 0 otherwise. It essentially reflects the rank of the correct match \\( g_{i} \\) among all gallery subjects for probe \\( p_{i} \\). The \\( S^{id}(p_i, g_i) \\) is an indicator, approaching 1 if \\( p_{i} \\) has a softrank value less than 1, indicating successful identification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.253 + ], + "angle": 0, + "content": "To optimize open-set detection and identification simultaneously, the overall identification-detection loss \\(\\mathcal{L}_{r,\\tau}\\) is formulated as the product of two components," + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.271, + 0.785, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {I D L}} = - \\frac {1}{| \\mathcal {P} _ {\\mathrm {K}} ^ {\\prime} |} \\sum_ {p _ {i} \\in \\mathcal {P} _ {\\mathrm {K}} ^ {\\prime}} S ^ {d e t} \\left(p _ {i}, g _ {i}\\right) \\cdot S ^ {i d} \\left(p _ {i}, g _ {i}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.359 + ], + "angle": 0, + "content": "In other words, \\(\\mathcal{L}^{\\mathrm{IDL}}\\) is only close to \\(-1\\) if both \\(S^{id}(p_i,g_i)\\) and \\(S^{det}(p_i,g_i)\\) approach 1. Conversely, should either or both not approximate 1, \\(\\mathcal{L}^{\\mathrm{IDL}}\\) converges to 0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.784, + 0.453 + ], + "angle": 0, + "content": "(3) Relative Threshold Minimization \\(\\mathcal{L}^{\\mathrm{IDL}}\\) optimizes detection and identification under fixed thresholds. We also introduce a strategy to minimize the maximum scores for non-mated probes, discouraging the model from assigning high scores to non-mated probes. This helps mitigate the risk of false positive detections in open-set scenarios, enhancing overall robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.785, + 0.53 + ], + "angle": 0, + "content": "To achieve this, we minimize the weighted average of the similarities between the probe and each gallery subject, where the weights are obtained by applying the softmax function to each score. Mathematically, given a probe sample \\( p_i \\) and a set of gallery subjects \\( \\mathcal{G}' = \\{g_1, g_2, \\dots, g_n\\} \\) with corresponding similarity scores \\( s_1, s_2, \\dots, s_n \\), we minimize the weighted average as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.536, + 0.785, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {R T M}} = \\frac {1}{\\sum_ {j = 1} ^ {n} e ^ {s _ {j}}} \\sum_ {j = 1} ^ {n} e ^ {s _ {j}} \\cdot s _ {j}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.614, + 0.594 + ], + "angle": 0, + "content": "where \\( e^{s_j} \\) is the softmax value for each similarity score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.784, + 0.699 + ], + "angle": 0, + "content": "Fig. 4 (c) shows gradients of \\(\\mathcal{L}_{\\mathrm{RTM}}\\) w.r.t. non-mated probes. This gradient adapts to the relative position of non-mated probes, assigning higher gradients to samples closer to the gallery. Unlike softmax-based losses and the triplet loss, RTM consistently emphasizes imposter scores that rank higher compared to others within the batch to learn better feature representations for open-set biometrics. This property is desirable as it effectively minimizes the maximum score of each non-mated probe." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.784, + 0.761 + ], + "angle": 0, + "content": "The max function is an alternative to the weighted average. Yet, simply minimizing the maximum score in training would not minimize other scores even if they are close to the maximum. Hence, the weighted average would make it easier for the trained model to generalize during testing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.779, + 0.364, + 0.792 + ], + "angle": 0, + "content": "3.4 Overall Loss" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.798, + 0.738, + 0.814 + ], + "angle": 0, + "content": "The overall objective is to minimize the combination of \\(\\mathcal{L}^{\\mathrm{IDL}}\\) and \\(\\mathcal{L}^{\\mathrm{RTM}}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.82, + 0.784, + 0.836 + ], + "angle": 0, + "content": "\\[\nL = \\mathcal {L} ^ {\\mathrm {I D L}} + \\lambda \\cdot \\mathcal {L} ^ {\\mathrm {R T M}}. \\tag {13}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the loss weight. Compared to existing loss functions, our loss function treats genuine and imposter scores differently. It aims to reduce open-set thresholds by pushing non-mated probes away from gallery while pulling genuine scores toward their corresponding galleries, similar to open-set evaluation. Moreover, the relative magnitude of non-mated scores is utilized by RTM to assign larger gradients to harder non-mated probes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.26, + 0.376, + 0.277 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.429, + 0.307 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Evaluation Protocols. In face recognition experiments, we use the official closed- and open-set evaluation protocols of IJB-S [28]. However, gait recognition and person reID datasets typically lack official open-set evaluation protocols, where some probe subjects are excluded from the gallery. Naïvely removing gallery subjects often results in highly unstable outcomes. This instability arises because threshold calculation heavily relies on a small fraction (1%) of non-mated probes due to the limited number of probes in these datasets. For instance, in a dataset with 500 non-mated probes, the threshold is computed from the highest scores of only 5 probes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.527 + ], + "angle": 0, + "content": "To ensure robust open-set evaluation, we create \\( N (= 50) \\) random sets of non-mated probes and compute FNIR@1%FPIR for each set. In each set, \\( q\\% = 21.5\\% \\) subjects are designated as non-mated. We report the median value and the standard deviation from \\( N \\) trials. The choice of median over mean is deliberate to avoid undue influence from extreme values and provide a more stable metric." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.547, + 0.461, + 0.653 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate in both open- and closed-set settings. In closed-set settings, we use the Cumulative Matching Curve (CMC) at rank-1; for open-set, we use FNIR@FPIR at \\(1\\%\\) and \\(0.1\\%\\) FPIR and rank \\(R = 20\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.461, + 0.78 + ], + "angle": 0, + "content": "Datasets. For face recognition, we train on WebFace4M [80] and test on IJB-S [28]. Our experiments in gait recognition involve the Gait3D [77] and GREW [79] datasets, while for person reID, we employ the MEVID [8] and" + }, + { + "type": "text", + "bbox": [ + 0.469, + 0.563, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Table 1: Summary of datasets used in our experiments. [Keys: Mod.=Modality; Img.=Open-set image recognition; Subj.=the number of subjects ; Media=the number of images for face/image recognition datasets and the number of sequences/tracklets for gait and video-based person reID datasets]" + }, + { + "type": "table", + "bbox": [ + 0.474, + 0.662, + 0.784, + 0.768 + ], + "angle": 0, + "content": "
DatasetYear Mod.TrainTest
Subj.MediaSubj.Media
IJB-S [28]2018Face-202>10M
Gait3D [77]2022Gait3,00018,9406,369
GREW [79]2021Gait20,000102,88724,000
MEVID [8]2023ReID1046,33855
CCVID [17]2022ReID75948151
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.652, + 0.795 + ], + "angle": 0, + "content": "CCVID [17] datasets. Further details are provided in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "One technical caveat is that although FNIR@1%FPIR implicitly assumes that each subject only appears once in the gallery, a typical gait recognition or person re-id dataset contains more than one gallery template per subject. To" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "address this issue, we take the mean feature over all gallery templates for each subject. This benefits the models compared to randomly sampling one gallery template (details are in the Supp.)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.215, + 0.803, + 0.262 + ], + "angle": 0, + "content": "Implementation Details. We implement our proposed approach in PyTorch [49]. For fair comparisons, we follow the experiment setup of our baselines. For details on the hyperparameters, please refer to the Supp." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.283, + 0.403, + 0.297 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.787, + 0.355 + ], + "angle": 0, + "content": "We conduct our ablation studies on the Gait3D [77] dataset using GaitBase [13] as the baseline. In the ablation studies, we hold out \\(20\\%\\) of subjects in the training set as the validation set." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.377, + 0.459, + 0.391 + ], + "angle": 0, + "content": "Effect of Det, Id and RTM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.461, + 0.557 + ], + "angle": 0, + "content": "We ablate the effect of Det, Id and RTM in Tab. 2. RTM consistently enhances both closed- and open-set performance. Id further improves open-set performance but has a minor impact on rank-1 accuracy. Meanwhile, Det significantly enhances both closed- and open-set performance. In summary, each component contributes to its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.458, + 0.596 + ], + "angle": 0, + "content": "Effect of Hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.46, + 0.687 + ], + "angle": 0, + "content": "We ablate the effect of hyperparameters \\(\\alpha, \\beta, \\gamma,\\) and \\(\\lambda\\) in Tab. 2. approach can consistently improve line, GaitBase [13]. Also, while the hyperparameters, our approach with the best set of tested hyperp" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.389, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Table 2: Ablation study results. [Keys: \\(\\checkmark =\\) Kept; \\(\\checkmark =\\) Removed; R@1=rank-1 accuracy]" + }, + { + "type": "table", + "bbox": [ + 0.472, + 0.419, + 0.785, + 0.595 + ], + "angle": 0, + "content": "
IdDetRTMα,γβλR@1FNIR@1%FPIR
XXX———64.686.43±5.44
XX✓6.00.24.065.185.81±5.78
X✓✓64.485.16±5.13
✓✓✓3.00.24.063.985.22±4.73
9.064.284.08±6.28
✓✓✓6.00.14.063.185.10±5.67
0.463.885.92±5.45
✓✓✓6.00.22.064.883.12±5.94
8.061.685.67±5.15
✓✓✓6.00.24.064.882.99±4.50
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.709, + 0.544, + 0.725 + ], + "angle": 0, + "content": "4.3 Comparison with State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Across biometric modalities, existing methods exhibit competitive closed-set recognition performance, as shown in Tab. 3 for face, Tab. 4 for gait, and Tab. 5 for person reID. In particular, in gait recognition, GaitBase [13] and SwinGait [12] showcase strong closed-set performance with a notable rank-1 accuracies of greater than \\(60\\%\\) and \\(70\\%\\), respectively, for in-the-wild gait datasets Gait3D [77] and GREW [79]. However, in open-set scenarios, they face challenges, as evidenced by GaitBase's high FNIR of \\(86.43\\%\\) at \\(1\\%\\) FPIR in Gait3D." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 3: Closed- and open-set face recognition on two official protocols of IJB-S [28]. [Keys: R@1=rank-1 accuracy; 0.1%=FNIR@0.1%FPIR; 1%: FNIR@1%FPIR]" + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.185, + 0.673, + 0.251 + ], + "angle": 0, + "content": "
ModelSurveillance-SingleSurveillance-Booking
R@1↑0.1%↓1%↓R@1↑0.1%↓1%↓
AdaFace [29]68.4445.1838.3768.8445.9837.80
AdaFace+Ours69.9244.3036.7370.0043.4836.17
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.785, + 0.37 + ], + "angle": 0, + "content": "It is evident that the inclusion of our loss functions consistently yields improvements in both closed- and open-set performance across three modalities. In face recognition, we boost the rank-1 accuracy of AdaFace [29] to \\(69.92\\%\\) \\((+1.48\\%)\\) in Surveillance-to-Single. In the challenging open-set scenario, our losses further demonstrate their effectiveness by reducing FNIR@0.1%FPIR from \\(45.98\\%\\) to \\(43.48\\%\\) in Surveillance-Booking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.432, + 0.612 + ], + "angle": 0, + "content": "In gait recognition, when applied to Swin-Gait [12] backbone, our losses improve the rank-1 accuracy to \\(75.3\\%\\) \\((+2.9\\%)\\) on Gait3D, and FNIR@1%FPIR decreases to \\(79.30\\%\\) \\((-4.20\\%)\\) on Gait3D. In person reID, integrating our loss with CAL [17] reaches a rank-1 accuracy of \\(84.8\\%\\) \\((+2.2\\%)\\) and a FNIR of \\(30.81\\%\\) \\((-3.08\\%)\\) on MEVID [8]. We conduct the independent two-sample t-test to assess the statistical significance of perfor" + }, + { + "type": "table_caption", + "bbox": [ + 0.442, + 0.389, + 0.788, + 0.458 + ], + "angle": 0, + "content": "Table 4: Closed- and open-set gait recognition on Gait3D [77] and GREW [79]. SMPLGait [77] takes SMPL parameters as additional input. SwinGait-3D [12] uses our own implementation. [Keys: Rank@1=rank-1 accuracy]" + }, + { + "type": "table", + "bbox": [ + 0.447, + 0.459, + 0.785, + 0.597 + ], + "angle": 0, + "content": "
ModelRank@1 ↑FNIR@1%FPIR ↓
Gait3DGREWGait3DGREW
GaitSet [4]36.748.496.05±2.3081.17±1.45
GaitPart [14]28.247.696.88±1.8088.43±1.26
GaitGL [38]29.741.594.97±1.2480.55±1.11
SMPLGait [77]46.3—93.18±2.58—
GaitBase [13]64.660.186.43±5.4477.67±1.60
+Ours64.860.182.99±4.5076.55±1.54
SwinGait-3D [12]72.479.383.50±6.2058.32±1.79
+Ours75.380.479.30±6.7457.29±1.83
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.784, + 0.645 + ], + "angle": 0, + "content": "mance improvements in open-set scenarios. For instance, our approach outperforms CAL on the CCVID [17] dataset, with a p-value of \\(0.052\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.785, + 0.706 + ], + "angle": 0, + "content": "Collectively, these findings emphasize the consistent effectiveness of our proposed loss functions in enhancing both closed- and open-set performance across a spectrum of biometric tasks, ultimately improving the models' ability to handle open-set challenges, while maintaining strong closed-set performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Efficiency. We present the floating point operations (FLOPs) of training baselines with and without our loss functions in Tab. 6. It is expected that incorporating our loss into baselines will increase the computational load in training, but the increase is very minor. For example, when SwinGait [12] is used, our loss functions only incur \\(46.9\\%\\) of the FLOPs of the triplet loss, which is negligible compared to the model. Moreover, our approach does not incur any computation overhead during testing while improving closed- and open-set performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.247, + 0.144, + 0.756, + 0.16 + ], + "angle": 0, + "content": "Table 5: Closed- and open-set person reID on MEVID [8] and CCVID [17]." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.171, + 0.689, + 0.283 + ], + "angle": 0, + "content": "
ModelDatasetRank-1Accuracy ↑FNIR@1%FPIR ↓
AGRL [66]MEVID [8]48.466.67±9.57
TCLNet [21]MEVID [8]48.170.34±11.68
PSTA [63]MEVID [8]48.277.11±11.72
CAL [17]MEVID [8]52.566.95±11.59
+OursMEVID [8]55.164.80±10.71
CAL [17]CCVID [17]82.633.89±7.65
+OursCCVID [17]84.830.81±8.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.301, + 0.787, + 0.345 + ], + "angle": 0, + "content": "Table 6: FLOPs in training different baseline models with and without our approach. We report the mean FLOPs across a minibatch. Our approach incurs no additional FLOP during testing." + }, + { + "type": "table", + "bbox": [ + 0.349, + 0.356, + 0.656, + 0.429 + ], + "angle": 0, + "content": "
ModalityModelFLOPS+Triplet+Ours
FaceAdaFace [29]24.19G—503K
GaitGaitBase [13]35.43G524.3K245.8K
GaitSwinGait [12]120G491.5K230.4K
BodyCAL [17]32.75G4.09K15.38K
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.377, + 0.477 + ], + "angle": 0, + "content": "4.4 Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.434, + 0.808 + ], + "angle": 0, + "content": "Score Distributions. Fig. 5 visualizes the non-mated, maximum non-mated score for each probe, and genuine scores of SwinGait [12], with and without our loss. Our introduced RTM effectively reduces the distribution of maximum-per-probe scores, shifting it towards lower values. After linearly normalizing all scores between 0 and 1, the threshold at \\(1\\%\\) FPIR is 0.61 with our loss (Fig. 5 (b)) compared to 0.69 without ours (Fig. 5 (a)). Moreover, our loss does not significantly alter the rightmost portion of the genuine score distribution, enhancing detec" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.785, + 0.839 + ], + "angle": 0, + "content": "tion. This analysis offers insights into how our loss changes score distributions, and through which improves open-set performance." + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.538, + 0.784, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.695, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Fig.5: Score distributions on the test set of Gait3D [77] dataset. Ours (b) effectively reduces the threshold at \\(1\\%\\) FPIR while only shifting the genuine score distribution marginally to the left compared to SwinGait [12] (a). Both distributions are shifted and scaled horizontally so that the min/max scores are within \\([0,1]\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.396, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.235, + 0.336, + 0.248 + ], + "angle": 0, + "content": "(a) Face" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.148, + 0.588, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.235, + 0.528, + 0.247 + ], + "angle": 0, + "content": "(b) Gait" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.149, + 0.781, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.235, + 0.743, + 0.248 + ], + "angle": 0, + "content": "(c) Person ReID" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.316 + ], + "angle": 0, + "content": "Fig. 6: Open-set FNs due to detection, identification, and either. The average across the 50 open-set evaluation protocols is plotted. Blue indicates FNs that are shared by the baseline, SwinGait [12], and our approach. Orange indicates FNs of the baseline only. Green indicates FNs of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.789, + 0.466 + ], + "angle": 0, + "content": "False Negatives Breakdown. As illustrated in Fig. 6, across different biometric and generic tasks, almost all of the false negatives are failed detection, e.g. the \\(1^{\\mathrm{st}}\\) mated probe in Fig. 1. A small portion of false negatives have both failed identification and detection, e.g. the \\(3^{\\mathrm{rd}}\\) mated probe in Fig. 1. When we apply our loss functions, the number of FNs due to detection is reduced and the number of FNs due to identification is comparable to that of the baseline. These results highlight that our approach can improve open-set detection without sacrificing identification in various settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.369, + 0.506 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Our work addresses the critical gap in open-set biometrics, a practical generalization of the well-studied closed-set biometrics. We introduce novel I-DL and RTM losses aiming to improve biometric models' open-set performance. Through extensive experiments across diverse modalities, including face, gait, and person reID, we demonstrate the efficacy of our approach. We not only advance the state of the art in the challenging open-set domain but also enhance closed-set recognition capabilities. We advocate for increased attention to the open-set problem within the biometric community." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Limitations While our paper makes significant strides in addressing open-set challenges in biometrics through loss functions, our exploration does not extensively explore dedicated architectures for this specific context. Future research endeavors may benefit from investigating tailored architectures that complement and enhance the effectiveness of our proposed loss functions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Potential Societal Impacts. The use of biometric datasets in our work containing personally identifiable images prompts a crucial discussion on privacy and ethical implications. As biometric technologies become more prevalent, it is imperative to ensure that the collection of such datasets is aligned with ethical standards and privacy regulations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Acknowledgements. This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via 2022-21102100004. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.292, + 0.323, + 0.308 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.756, + 0.339 + ], + "angle": 0, + "content": "1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.786, + 0.38 + ], + "angle": 0, + "content": "2. Cen, J., Luan, D., Zhang, S., Pei, Y., Zhang, Y., Zhao, D., Shen, S., Chen, Q.: The devil is in the wrongly-classified samples: Towards unified open-set recognition. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.786, + 0.408 + ], + "angle": 0, + "content": "3. Chai, T., Li, A., Zhang, S., Li, Z., Wang, Y.: Lagrange motion analysis and view embeddings for improved gait recognition. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.786, + 0.436 + ], + "angle": 0, + "content": "4. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: AAAI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.786, + 0.464 + ], + "angle": 0, + "content": "5. Chen, C., Ye, M., Jiang, D.: Towards modality-agnostic person re-identification with descriptive query. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.786, + 0.492 + ], + "angle": 0, + "content": "6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.786, + 0.519 + ], + "angle": 0, + "content": "7. Cui, Y., Kang, Y.: Multi-modal gait recognition via effective spatial-temporal feature fusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.52, + 0.786, + 0.561 + ], + "angle": 0, + "content": "8. Davila, D., Du, D., Lewis, B., Funk, C., Van Pelt, J., Collins, R., Corona, K., Brown, M., McCloskey, S., Hoogs, A., et al.: MEVID: multi-view extended videos with identities for video person re-identification. In: WACV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.786, + 0.589 + ], + "angle": 0, + "content": "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.786, + 0.618 + ], + "angle": 0, + "content": "0. Dou, H., Zhang, P., Su, W., Yu, Y., Li, X.: Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.786, + 0.645 + ], + "angle": 0, + "content": "1. Dou, H., Zhang, P., Su, W., Yu, Y., Lin, Y., Li, X.: Gaitgci: Generative counterfactual intervention for gait recognition. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.786, + 0.673 + ], + "angle": 0, + "content": "2. Fan, C., Hou, S., Huang, Y., Yu, S.: Exploring deep models for practical gait recognition. arXiv preprint arXiv:2303.03301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.786, + 0.701 + ], + "angle": 0, + "content": "3. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.729 + ], + "angle": 0, + "content": "4. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.786, + 0.757 + ], + "angle": 0, + "content": "5. Geng, C., Huang, S., Chen, S.: Recent advances in open set recognition: A survey. IEEE TPAMI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.784 + ], + "angle": 0, + "content": "6. Grother, P., Ngan, M., Hanaoka, K.: Face recognition vendor test (FRVT) part 2: Identification (2019), https://www.nist.gov/document/nistir827120190911pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.812 + ], + "angle": 0, + "content": "7. Gu, X., Chang, H., Ma, B., Bai, S., Shan, S., Chen, X.: Clothes-changing person re-identification with rgb modality only. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "8. Gunther, M., Cruz, S., Rudd, E.M., Boult, T.E.: Toward open-set face recognition. In: CVPR Workshops (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.324, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Günther, M., Hu, P., Herrmann, C., Chan, C.H., Jiang, M., Yang, S., Dhamija, A.R., Ramanan, D., Beyerer, J., Kittler, J., et al.: Unconstrained face detection and open-set face recognition challenge. In: IJCB. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "20. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "21. Hou, R., Chang, H., Ma, B., Shan, S., Chen, X.: Temporal complementary learning for video person re-identification. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.271 + ], + "angle": 0, + "content": "22. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "23. Huang, X., Zhu, D., Wang, H., Wang, X., Yang, B., He, B., Liu, W., Feng, B.: Context-sensitive temporal feature learning for gait recognition. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "24. Huang, Y., Shen, P., Tai, Y., Li, S., Liu, X., Li, J., Huang, F., Ji, R.: Improving face recognition from hard samples via distribution distillation loss. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.366 + ], + "angle": 0, + "content": "25. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.393 + ], + "angle": 0, + "content": "26. Huang, Z., Xue, D., Shen, X., Tian, X., Li, H., Huang, J., Hua, X.S.: 3d local convolutional neural networks for gait recognition. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.42 + ], + "angle": 0, + "content": "27. Jain, A.K., Deb, D., Engelsma, J.J.: Biometrics: Trust, but verify. IEEE Transactions on Biometrics, Behavior, and Identity Science (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "28. Kalka, N.D., Maze, B., Duncan, J.A., O'Connor, K., Elliott, S., Hebert, K., Bryan, J., Jain, A.K.: IJB-S: IARPA janus surveillance video benchmark. In: BTAS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "29. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "30. Kim, M., Liu, F., Jain, A.K., Liu, X.: Cluster and aggregate: Face recognition with large probe set. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "31. Kim, M., Su, Y., Liu, F., Jain, A., Liu, X.: Keypoint relative position encoding for face recognition. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.555 + ], + "angle": 0, + "content": "32. Koch, T., Riess, C., Kohler, T.: Lord: Leveraging open-set recognition with unknown data. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.582 + ], + "angle": 0, + "content": "33. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "34. Kuchibhotla, H.C., Malagi, S.S., Chandhok, S., Balasubramanian, V.N.: Unseen classes at a later time? no problem. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "35. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "36. Li, W., Hou, S., Zhang, C., Cao, C., Liu, X., Huang, Y., Zhao, Y.: An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "37. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.746 + ], + "angle": 0, + "content": "38. Lin, B., Zhang, S., Yu, X.: Gait recognition via effective global-local feature representation and local temporal aggregation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "39. Liu, F., Ashbaugh, R., Chimitt, N., Hassan, N., Hassani, A., Jaiswal, A., Kim, M., Mao, Z., Perry, C., Ren, Z., et al.: Farsight: A physics-driven whole-body biometric system at large distance and altitude. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "40. Liu, F., Kim, M., Gu, Z., Jain, A., Liu, X.: Learning clothing and pose invariant 3d shape representation for long-term person re-identification. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Liu, F., Kim, M., Jain, A., Liu, X.: Controllable and guided face synthesis for unconstrained face recognition. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "42. Liu, F., Kim, M., Ren, Z., Liu, X.: Distilling CLIP with dual guidance for learning discriminative human body shape representation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "43. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "44. Ma, K., Fu, Y., Zheng, D., Cao, C., Hu, X., Huang, Y.: Dynamic aggregated network for gait recognition. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.27 + ], + "angle": 0, + "content": "45. Mekhazni, D., Dufau, M., Desrosiers, C., Pedersoli, M., Granger, E.: Camera alignment and weighted contrastive learning for domain adaptation in video person reid. In: WACV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "46. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "47. Minaee, S., Abdelrashidi, A., Su, H., Bennamoun, M., Zhang, D.: Biometrics recognition using deep learning: A survey. Artificial Intelligence Review (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.785, + 0.352 + ], + "angle": 0, + "content": "48. Ni, H., Li, Y., Gao, L., Shen, H.T., Song, J.: Part-aware transformer for generalizable person re-identification. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.392 + ], + "angle": 0, + "content": "49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "50. Qin, L., Wang, M., Deng, C., Wang, K., Chen, X., Hu, J., Deng, W.: Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. TCSVT (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "51. Ren, Z., Su, Y., Liu, X.: Chatgpt-powered hierarchical comparisons for image classification. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "52. Ross, A., Banerjee, S., Chen, C., Chowdhury, A., Mirjalili, V., Sharma, R., Swearinen, T., Yadav, S.: Some research problems in biometrics: The future beckons. In: ICB (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "53. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.555 + ], + "angle": 0, + "content": "54. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3d gait recognition with point clouds. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.595 + ], + "angle": 0, + "content": "55. Shen, C., Yu, S., Wang, J., Huang, G.Q., Wang, L.: A comprehensive survey on deep gait recognition: algorithms, datasets and challenges. arXiv preprint arXiv:2206.13732 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.622 + ], + "angle": 0, + "content": "56. Somers, V., De Vleeschouwer, C., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.65 + ], + "angle": 0, + "content": "57. Sun, Y., Cheng, C., Zhang, Y., Zhang, C., Zheng, L., Wang, Z., Wei, Y.: Circle loss: A unified perspective of pair similarity optimization. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.69 + ], + "angle": 0, + "content": "58. Tabassi, E., Watson, C., Fiumara, G., Salamon, W., Flanagan, P., Cheng, S.L.: Performance evaluation of fingerprint open-set identification algorithms. In: IJCB (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "59. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.746 + ], + "angle": 0, + "content": "60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "61. Wang, L., Liu, B., Liang, F., Wang, B.: Hierarchical spatio-temporal representation learning for gait recognition. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.812 + ], + "angle": 0, + "content": "62. Wang, M., Guo, X., Lin, B., Yang, T., Zhu, Z., Li, L., Zhang, S., Yu, X.: Dygait: Exploiting dynamic representations for high-performance gait recognition. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Wang, Y., Zhang, P., Gao, S., Geng, X., Lu, H., Wang, D.: Pyramid spatial-temporal aggregation for video-based person re-identification. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Su et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "64. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards AUC-oriented open-set recognition. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "66. Wu, Y., Bourahla, O.E.F., Li, X., Wu, F., Tian, Q., Zhou, X.: Adaptive graph representation learning for video person re-identification. TIP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "67. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "68. Ye, D., Fan, C., Ma, J., Liu, X., Yu, S.: Biggait: Learning gait representation you want by large vision models. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "69. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "70. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.41 + ], + "angle": 0, + "content": "71. Zahra, A., Perwaiz, N., Shahzad, M., Fraz, M.M.: Person re-identification: A retrospective on domain specific open challenges and future trends. Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "72. Zhai, Y., Liu, Z., Wu, Z., Wu, Y., Zhou, C., Doermann, D., Yuan, J., Hua, G.: Soar: Scene-debiasing open-set action recognition. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "73. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "74. Zhang, Z., Tran, L., Liu, F., Liu, X.: On learning disentangled representations for gait recognition. IEEE TPAMI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "75. Zhang, Z., Tran, L., Yin, X., Atoum, Y., Liu, X., Wan, J., Wang, N.: Gait recognition via disentangled representation learning. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "76. Zheng, J., Liu, X., Gu, X., Sun, Y., Gan, C., Zhang, J., Liu, W., Yan, C.: Gait recognition in the wild with multi-hop temporal switch. In: ACMMM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "77. Zheng, J., Liu, X., Liu, W., He, L., Yan, C., Mei, T.: Gait recognition in the wild with dense 3d representations and a benchmark. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "78. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "79. Zhu, Z., Guo, X., Yang, T., Huang, J., Deng, J., Huang, G., Du, D., Lu, J., Zhou, J.: Gait recognition in the wild: A benchmark. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "80. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., et al.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_origin.pdf b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db42a7965f3dc71518e32ba8d00bd177d5ce98b1 --- /dev/null +++ b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/5d8636e8-0041-40e7-8842-7b710a232989_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb37c8adce19fd8f5d8384e3e8d470ba6c0e7eb824e735d920f6039352b9b024 +size 1034716 diff --git a/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/full.md b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d36bfa216e7a77ca5ec8801899ae7e62ef30ba1 --- /dev/null +++ b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/full.md @@ -0,0 +1,437 @@ +# Open-Set Biometrics: Beyond Good Closed-Set Models + +Yiyang Su, Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu + +Michigan State University, East Lansing, MI 48824, USA {suyiyan1, kimminc2, liufeng6, jain, liuxm}@msu.edu + +Abstract. Biometric recognition has primarily addressed closed-set identification, assuming all probe subjects are in the gallery. However, most practical applications involve open-set biometrics, where probe subjects may or may not be present in the gallery. This poses distinct challenges in effectively distinguishing individuals in the gallery while minimizing false detections. While it is commonly believed that powerful biometric models can excel in both closed- and open-set scenarios, existing loss functions are inconsistent with open-set evaluation. They treat genuine (mated) and imposter (non-mated) similarity scores symmetrically and neglect the relative magnitudes of imposter scores. To address these issues, we simulate open-set evaluation using minibatches during training and introduce novel loss functions: (1) the identification-detection loss optimized for open-set performance under selective thresholds and (2) relative threshold minimization to reduce the maximum negative score for each probe. Across diverse biometric tasks, including face recognition, gait recognition, and person re-identification, our experiments demonstrate the effectiveness of the proposed loss functions, significantly enhancing open-set performance while positively impacting closed-set performance. Our code and models are available here. + +Keywords: Open-set biometrics $\cdot$ Face recognition $\cdot$ Gait recognition $\cdot$ Person reID + +# 1 Introduction + +Biometrics, the study of unique physiological or behavioral characteristics for human identification, plays a pivotal role in applications such as access control, surveillance, and forensic analysis [52]. Each biometric task, from face recognition, gait recognition to person re-identification (reID), uses features tailored to its specific requirements [39, 55, 71]. + +In real-world applications, after test subjects are enrolled in the gallery, biometric systems inevitably encounter non-mated subjects, those not present in the gallery. Thus, it becomes imperative for such systems to demonstrate robust performance in the open-set framework with non-mated probes [16, 18, 28, 58]. The closed-set framework assumes all probe subjects are mated and predominantly focuses on identification, associating probes with the closest match in the + +![](images/1894c19c93dc6f73e2dc162e99b0289153a08461e1bce4793f1754b8e568e40c.jpg) +Fig. 1: A toy example for FNIR@FPIR calculation. The score matrix and distribution between non-mated probes and all gallery subjects are shown in blue. The maximum score for each non-mated probe is shown in red. A percentile of their distribution determines the threshold. A user-defined percentile (e.g., $33\%$ ) sets the threshold to be 0.75 as 1 out of 3 (the $2^{\mathrm{nd}}$ ) non-mated probes is incorrectly classified as "in the gallery". The genuine scores of mated probes are shown in green. A mated probe is an FN if either of the conditions is met: (1) in detection (Det.), it is incorrectly classified as "not in the gallery" ( $\mathcal{G}$ ) as the genuine score is below the threshold (the $1^{\mathrm{st}}$ mated probe); (2) its genuine score is not within a specified rank $R$ , e.g., 1, in the gallery (the $2^{\mathrm{nd}}$ mated probe). Hence, FNIR@33%FPIR is $75\%$ . + +gallery. However, the open-set framework introduces an additional challenge: detection, determining whether a probe subject is mated. For example, in airport surveillance, faces in CCTV videos are compared against a criminal watch list. The goal is to reduce the misidentification (false positives) of non-mated subjects such as airport staff and innocent passengers (improving detection), while simultaneously ensuring accurate recognition of known criminals (enhancing identification). For this reason, open-set biometrics is a harder and more general task than closed-set. + +The biometric community has predominantly centered on the closed-set framework. While some face recognition benchmarks have open-set evaluation protocols [18,19,28], they are primarily used for evaluation rather than for developing tailored strategies to address the open-set challenges. In gait recognition and person reID, most benchmarks only provide official closed-set protocols, lacking open-set ones. To address this gap, we take the initiative to introduce open-set benchmarks and propose strategies to train models specifically for improving open-set performances while maintaining closed-set performance. + +The false negative identification rate at a given false positive identification rate (FNIR@FPIR) is a de facto open-set metric [16]. We first compute the similarity score matrix between non-mated probes and the gallery, from which we extract the maximum score for each non-mated probe across the gallery, forming the maximum-per-probe non-mated score distribution. This distribution represents the likelihood that a non-mated probe has a match in the gallery. A threshold is set based on a user-defined percentile (e.g., $1\%$ ) of this distribution. Mated probes with scores below this threshold are counted as false negatives (FNs), as they are incorrectly considered not to be in the gallery. Similar to the + +closed-set scenario, a mated probe is also an FN if its gallery match doesn't rank among the gallery set within a user-specified rank $R$ . FNIR is the proportion of mated probes classified as FNs. Fig. 1 depicts its calculation pipeline. + +Unlike closed-set or verification metrics, FNIR@FPIR introduces unique challenges. It employs a threshold based on maximum scores for each non-mated probe, creating an asymmetry between genuine and imposter scores. SOTA biometric recognition methods achieve strong performance in closed-set scenarios leveraging triplet loss [20] and/or softmax cross-entropy based losses [9, 25, 29, 46, 60], both of which excel in reducing intra-person variance and increasing inter-person variance. However, they do not take into account the fact that the non-mated score distribution determines the threshold for the false negative. + +In other words, conventional loss functions, while effective in closed-set scenarios, exhibit limitations in addressing the nuances of the FNIR@FPIR metric, which necessitate a handling of the maximum of imposter scores. These high-scoring non-mated probes disproportionately influence the threshold setting for false negatives, yet conventional loss functions often treat them equivalently to other scores. As a result, they face challenges in open-set scenarios. GaitBase [13], a SOTA gait recognition model, can only achieve a $61\%$ FNIR@1%FPIR on the training set and $86\%$ on the test set. This challenge extends to face recognition and person reID. + +To address this challenge, we propose to create simulated open-set evaluation episodes during training, whose score distributions serve as proxies to those of the test set. Specifically, we introduce an identification-detection loss tailored to optimize the two core elements in open-set biometrics, i.e., detection and identification. For detection, the model is trained to accurately pinpoint the exact probes whose mate is present in the gallery set; for identification, the model learns to establish associations between the gallery and probe samples of mated subjects. By integrating the aspects of detection and identification within training batches, we improve the model's efficacy in both domains, thus yielding enhanced testing performance. Further, we minimize similarity scores between non-mated subjects and the gallery, enabling a more comprehensive detection of mated subjects (see Fig. 3). + +Empirically, we apply our proposed loss function to three biometric modalities, face, gait, and person reID, and show performance improvement in all tasks. While our focus is open-set biometrics, we also observe performance gain in closed-set search due to improved feature representations. + +To summarize, our contributions include: + +$\checkmark$ We show that the existing loss functions, e.g., triplet loss and softmax loss, do not treat genuine and imposter training samples in accordance with open-set evaluation. This inconsistency sheds light on why the open-set challenges have not been adequately addressed. +$\checkmark$ We propose a novel loss function that leverages open-set mini-batches during training to optimize identification and detection performance and minimize open-set detection thresholds. + +![](images/8229de76d49520d971e21e38a586d6917b82ff56eb1fcba2dd37204d1abf2606.jpg) +Fig. 2: Comparison between open-set recognition and biometrics. Closed-set recognition has the same set of classes for training and testing. Open-set recognition involves testing on unseen classes during training. On the other hand, in biometrics, the sets of training and testing subjects are always disjoint. In closed-set biometrics, all probe subjects are in the gallery, and in open-set biometrics, some probe subjects are not enrolled in the gallery. [Legends: ▪ gallery subjects, ▪ probe subjects] + +Training Testing classes + +Closed-Set Recognition + +![](images/3287ea4dc50c33faa3b3076a0d310355d1537a07da695711a9af3e152d35debc.jpg) + +Training Testing classes + +Open-Set Recognition + +![](images/5b732f0aacb12364654e5d6f26e60e7327007c202f5d086397307626c56fa04b.jpg) + +Training subjects + +![](images/ff77196acab4689c6cb553d048134e6766f969b6c2b2770e143e63d8afc3e0c8.jpg) + +Testing subjects + +Closed-Set Biometrics + +![](images/4d382f7589b5738bf0e795179809598bc523c3e350e4f6f81bf39439f8df7c02.jpg) + +Training subjects + +![](images/3feb03aabdf216d3310d60c1dc57b76fd707d2f7c537c2fd8ac7b7130935b009.jpg) + +Testing subjects + +Open-Set Biometrics + +$\checkmark$ We experimentally demonstrate that our proposed loss function can effectively improve both closed- and open-set performance across diverse biometric modalities, including face, gait, and person reID. + +# 2 Related Work + +Open-Set Biometrics. While various biometric recognition tasks, such as face recognition [9, 24, 25, 29-31, 41, 43, 46, 50, 60], gait recognition [3, 4, 7, 10-14, 23, 26, 36-38, 44, 54, 61, 62, 68, 74-77], and person reID [5, 17, 21, 35, 36, 40, 42, 45, 48, 56, 63, 66], have witnessed significant progress, the majority of their focus remains on closed-set scenarios. Although EVM has been applied to open-set face recognition [18], EVM takes extracted features, instead of images/videos, as input. Our approach improves feature representation and EVM improves score calculation. While some face recognition datasets provide opens-set evaluation protocols [28], open-set performance is often perceived as a consequence rather than a direct objective in advancing recognition models. Further, even the most recent gait recognition [68] and person reID [42] models are exclusively evaluated under closed-set settings due to the absence of open-set evaluation protocols. + +Open-Set Recognition. While open-set biometrics is relatively underexplored, open-set recognition [1,2,6,22,32-34,51,59,64,67,69,72,73,78] has attracted attention in the literature. Open-set recognition distinguishes between known and unknown classes in image classification. In both open-set image recognition and open-set biometrics, the primary concern is how the models deal with instances that belong to unknown classes or non-mated probes. However, as illustrated in Fig. 2, in open-set recognition known classes come from the training set and the unknown classes refer to classes not used for training [15]. On the other hand, even in closed-set biometrics, all test subjects are typically not used during training and mated probes belong to test subjects enrolled in the gallery [27,47]. Open-set biometrics differs from closed-set biometrics in that there are non-mated subjects are neither in the training set nor in the gallery [16,18,58]. In + +![](images/6898f0d2ae795b6537516797e76054f9c31695e15459e8316eb56ca4572e2159.jpg) +Fig. 3: Overview of our approach. We introduce an additional modality-agnostic loss function specifically designed for open-set biometrics, which simultaneously optimizes identification (Id.) and detection (Det.) and Relatie Threshold Minimization (RTM) for open-set thresholds. + +evaluation, open-set recognition models typically use closed-set recognition accuracy and AUROC for detecting classes not seen during training [15] and open-set biometrics uses FNIR@FPIR [16]. + +# 3 Method + +# 3.1 Problem Formulation + +Let $S_{\mathrm{test}}$ be the test set, $\mathcal{G} \subset S_{\mathrm{test}}$ represent the gallery set containing mated subjects, and $\mathcal{P} \subset S_{\mathrm{test}}$ denote the probe set consisting of test samples. For each probe sample $p_i \in \mathcal{P}$ , there are two possible outcomes: (1) If $p_i$ is mated, i.e., its corresponding subject $g_i$ is in $\mathcal{G}$ , the biometric model must identify the correct subject. (2) If $p_i$ is non-mated, the model must reject it. The open-set scenario poses a more challenging problem than the traditional closed-set one, as the model needs to recognize mated subjects while rejecting non-mated ones. + +# 3.2 Revisiting Loss Functions for Biometrics + +In this section, we compute gradients of two common loss functions in biometrics: softmax and triplet loss. Many loss functions are derived from them and share similar characteristics. For instance, center loss [65] can be analyzed similarly to softmax loss, and the logic of triplet loss can also be applied to circle loss [57]. + +Softmax loss. The softmax loss, often employed in closed-set classification, operates by computing the cross-entropy between the normalized logit of a sample and its corresponding ground truth class. Mathematically, for a sample $i$ in a batch, the softmax loss $L_{\mathrm{softmax}}(i)$ is: + +$$ +L _ {\text {s o f t m a x}} (i) = - \log \left(\frac {e ^ {z _ {i , s}}}{\sum_ {t} e ^ {z _ {i , t}}}\right). \tag {1} +$$ + +Here, $z_{i,s}$ denotes the logit for sample $i$ belonging to subject $s$ . This loss aims to minimize the negative log-likelihood of the true subject $s$ for sample $i$ . For one + +sample, the gradient of the softmax loss $w.r.t.$ the logit $z_{i,s}$ is given by: + +$$ +\frac {\partial L _ {\mathrm {s o f t m a x}} (i)}{\partial z _ {i , s}} = \frac {e ^ {z _ {i , s}}}{\sum_ {t} e ^ {z _ {i , t}}} - \delta_ {i, s}, \tag {2} +$$ + +where $\frac{e^{z_{i,s}}}{\sum_{j=1}^{N} e^{z_{j,s}}}$ represents the probability that sample $i$ belongs to subject $s$ and $\delta_{i,c}$ is the Kronecker delta, which is 1 when $i = c$ (the true subject) and 0 otherwise. We visualize gradients of the softmax loss in Fig. 4(a). + +However, it overlooks the relative magnitude of each imposter score within the batch. An imposter score may be deemed small by, e.g., AdaFace [29] standards but still rank high among maximum-per-probe non-mated scores, raising the open-set threshold. Further, the softmax loss assumes that each sample can be assigned to one of the training subjects, which is not the case in open-set settings. Consequently, the softmax loss is not inherently designed to handle open-set biometrics. + +Triplet loss. The triplet loss [53] is defined as: + +$$ +L _ {\mathrm {t r i}} (a, p, n) = \max (0, d (a, p) - d (a, n) + m). \tag {3} +$$ + +Note that $d(a, p)$ is the distance between the anchor $a$ and positive sample $p$ , $d(a, n)$ is between $a$ and the negative sample $n$ , and $m$ is a constant margin. The partial derivatives of $L(a, p, n)$ w.r.t. $d(a, p)$ and $d(a, n)$ are piecewise functions: + +$$ +\frac {\partial L _ {\operatorname {t r i}} (a , p , n)}{\partial d (a , p)} = \left\{ \begin{array}{l l} 1, & \text {i f} d (a, p) + m > d (a, n) \\ 0, & \text {i f} d (a, p) + m \leq d (a, n) \end{array} ; \right. \tag {4} +$$ + +$$ +\frac {\partial L _ {\operatorname {t r i}} (a , p , n)}{\partial d (a , n)} = \left\{ \begin{array}{l l} - 1, & \text {i f} d (a, p) + m > d (a, n) \\ 0, & \text {i f} d (a, p) + m \leq d (a, n) \end{array} \right.. \tag {5} +$$ + +The magnitude of the gradient is constant regardless of the distance between two samples. This inability to differentiate samples in the gradient implies that all negative samples in the triplet are contributing equally, which is not reflective of the open-set scenario. One remedy often used with triplet loss is hard negative mining. However, it comes with a critical drawback of being susceptible to label noise [70]. Noisy and mislabeled samples are often considered hard negatives, which can make the loss function sensitive to noisy data when focusing on the hardest samples. Thus, we believe that a loss function that considers all negative samples while being able to adjust the gradient magnitude based on the anchor-negative distances would be effective in open-set biometrics. + +Hence, we propose a novel loss function, shown in Fig. 3. It optimizes identification and detection at various thresholds and minimizes non-mated scores to reduce open-set thresholds. It can be applied to any biometric modality. Unlike existing loss functions, it is consistent with FNIR@FPIR in that it treats genuine and imposter scores asymmetrically and tackles large non-mated scores. + +(a) SoftMax +![](images/52c76ac312b534708c2225e80f66d0cc7ce2254ab9a3e9cde2a79db71cede52d.jpg) +$\oplus$ Genuine probe $(p)$ + +Gradient Magnitude +![](images/9e4793ee91db32ed3139b83a6b44fa7ca4588bbf0788886e7406ed3bbc271270.jpg) +Threshold +$\rightarrow$ Gradient + +(b) Triplet Loss +![](images/9a830a71b69130fadc524c4716deb1d21977307162096105aa84fd8873d5218c.jpg) +$\ominus$ Imposter probe $(n)$ + +![](images/5577a7bd635537aa144436de87afdb42dc6b0919ea9794c11b1cd967daf7b703.jpg) + +(c) Ours +Identification Detection +$p_i$ Most Emphasized +![](images/4a1469d03aa844543f7c8ac099f771de10b7f5c1d72b440e31d5bf457079ae5b.jpg) +different $p_i$ gradient magnitude based on threshold + +Relative Threshold Minimization +$n_2$ Emphasized +Fig. 4: Visualization of the gradient of different loss functions. The colors represent gradient magnitude and the arrows represent gradient directions. a) SoftMax loss pushes the sample toward the class center and away from the other classes, with gradual shifts in gradient magnitude. b) On the other hand, triplet loss has a hard cut gradient based on triplet samples. c) For $S_{\tau}^{det}$ , as shown in the top row, the thresholds are determined by non-mated sample, $n_j$ . $p_i$ 's gradient has the greatest magnitude when it has a similar distance from the gallery $g_i$ to $n_j$ . For Relative Threshold Minimization, as shown in the bottom row, as $n_2$ moves away from the gallery, its gradient decreases. While $n_1$ remains at the same location, its gradient increases because it becomes closer to $g_i$ than $n_2$ . The gradients w.r.t. genuine scores adapt to non-mated scores and the gradients w.r.t. non-mated scores are adaptive to other non-mated scores. +![](images/28ca7ef5ed26341493e44c4f99627cdc714a0fb9ad7e2a21015cc390d5fb56e3.jpg) +different $n_j$ gradient magnitude based on relative distance + +![](images/3667c27e033dc36e5f678e648abe413ccbb9658585d27bde538392110b58c8a0.jpg) +$p_i$ Emphasized + +![](images/29ddcc80faf53ffbfafc1a5378346db052fd4bc681ce2845a5690b34d8e805fe.jpg) +$n_1,n_2$ Emphasized + +![](images/7d75486625275e10b438d32d877414329e2dda47c910815db9dc4e69a12346b4.jpg) +$p_i$ Less Emphasized + +![](images/bed9bd62c7db3fceb67f87596eebbd49389e55ec4dda998918f834a5e5221c12.jpg) +$n_1$ Emphasized + +# 3.3 Our Approach + +To design loss functions tailored for open-set biometrics, we create a training scenario mirroring the testing environment. During training, we partition each batch into gallery and probe sets to simulate open-set evaluation conditions. In this process, we randomly designate a certain percentage, denoted as $p\%$ , of the subjects as mated subjects, while the rest are categorized as non-mated. For the mated subjects, we further randomly allocate their exemplars to the gallery and probe sets. Conversely, all exemplars of the non-mated subjects are placed in the probe set. This approach yields a more realistic and representative training scenario, preparing the biometric features to excel under open-set conditions. To avoid any potential notation ambiguity, we refer to the mated probe set created within each training batch as $\mathcal{P}_{\mathrm{K}}^{\prime}$ , the non-mated probe set as $\mathcal{P}_{\mathrm{U}}^{\prime}$ , and the gallery sets within each training batch as $\mathcal{G}^{\prime}$ . + +To ensure robustness, when the metric space inhabited by the biometric features is defined under the Euclidean distance, we calculate the similarity scores between two features as: + +$$ +\mathrm {s} \left(p _ {i}, g _ {j}\right) = \frac {1}{1 + \mathrm {d} \left(p _ {i} , g _ {j}\right)}, \tag {6} +$$ + +where $\mathrm{d}(\cdot, \cdot)$ represents the Euclidean distance. Alternatively, if the metric space is defined under the cosine distance, we simply adopt the cosine similarity. + +To ensure that the model can generalize to open-set scenarios during testing, we propose loss functions that penalize the model for three types of errors during + +training: (1) failing to detect a mated probe with a threshold $\tau$ , (2) failing to identify a mated probe within the top rank- $r$ positions and (3) setting the threshold $\tau$ too high, leading to poor differentiation between mated and non-mated sample distributions. + +(1) Detection. Detection is assessing whether the computed similarity score exceeds a predefined threshold $\tau$ . That is, for a given probe $p_i \in \mathcal{P}_{\mathrm{K}}'$ and its corresponding gallery subject $g_i \in \mathcal{G}'$ , we calculate a detection score as: + +$$ +S _ {\tau} ^ {d e t} \left(p _ {i}, g _ {i}\right) = \sigma_ {\alpha} \left(\mathrm {s} \left(p _ {i}, g _ {i}\right) - \tau\right), \tag {7} +$$ + +where $\sigma_{\alpha}(x) = 1 / (1 + \exp (-\alpha x))$ is the Sigmoid function with temperature hyperparameter $\alpha$ . The $S_{\tau}^{det}(p_i, g_i)$ score quantifies the model's ability to detect if $p_i$ and $g_i$ belong to the same subject, given the threshold $\tau$ . A high similarity score, well above $\tau$ , indicates successful detection and $S_{\tau}^{det}(p_i, g_i)$ approaches 1. In contrast, if the model fails to detect that $p_i$ and $g_i$ belong to the same subject, $s(p_i, g_i)$ remains much smaller than $\tau$ , causing $S_{\tau}^{det}(p_i, g_i)$ to approach 0. Since the Sigmoid function has the highest gradient near 0, the term $\mathrm{s}(p_i, g_i) - \tau$ makes the loss function focus more on the samples near the threshold $\tau$ . + +During testing, the open-set detection threshold is determined by the non-mated score distribution, which we do not know during training. Thus, we sample $\tau$ from the non-mated scores in a batch. Specifically, we let + +$$ +S ^ {d e t} = \frac {1}{| \mathcal {T} |} \sum_ {\tau \in \mathcal {T}} S _ {\tau} ^ {d e t} \left(p _ {i}, g _ {i}\right), \tag {8} +$$ + +where $\mathcal{T} = \{s(n_j,g_i)|n_j\in \mathcal{P}_{\mathrm{U}}'\}$ is the set of all non-mated scores. Fig. 4(c) illustrates the gradient of Det w.r.t. mated probes under different thresholds induced by various non-mated probes. The gradient is adaptable to different thresholds, increasing as $\tau$ decreases, even when $s(g,p)$ is constant. This adaptability is desirable in open-set settings, as it promotes mated probes to be more similar to their ground-truth gallery than non-mated probes. Note that the thresholds in Eq. (8) are only used during training. The detection thresholds during testing are automatically determined by the FNIR@FPIR metric as discussed in Sec. 1. + +(2) Identification. It is not sufficient for a biometric model to distinguish mated subjects from non-mated ones based on their scores. Equally important is the model's ability to accurately identify the correct subject among the gallery subjects. Therefore, we introduce an identification score $S^{id}(p_i, g_i)$ as: + +$$ +S ^ {i d} \left(p _ {i}, g _ {i}\right) = \sigma_ {\beta} \left(1 - \operatorname {s o f t r a n k} \left(p _ {i}, g _ {i}\right)\right), \tag {9} +$$ + +where $\sigma_{\beta}$ is a Sigmoid function with temperature $\beta$ and sofrank is a differentiable function with a Sigmoid function $\sigma_{\gamma}$ defined as: + +$$ +\operatorname {s o f t r a n k} \left(p _ {i}, g _ {i}\right) = \sum_ {g _ {j} \in \mathcal {G} ^ {\prime}} \sigma_ {\gamma} \left(\mathrm {s} \left(p _ {i}, g _ {j}\right) - \mathrm {s} \left(p _ {i}, g _ {i}\right)\right), \tag {10} +$$ + +The softrank function accumulates values close to 1 for gallery subjects $g_{j}$ more similar to the probe $p_{i}$ than the correct match $g_{i}$ and close to 0 otherwise. It essentially reflects the rank of the correct match $g_{i}$ among all gallery subjects for probe $p_{i}$ . The $S^{id}(p_i, g_i)$ is an indicator, approaching 1 if $p_{i}$ has a softrank value less than 1, indicating successful identification. + +To optimize open-set detection and identification simultaneously, the overall identification-detection loss $\mathcal{L}_{r,\tau}$ is formulated as the product of two components, + +$$ +\mathcal {L} ^ {\mathrm {I D L}} = - \frac {1}{| \mathcal {P} _ {\mathrm {K}} ^ {\prime} |} \sum_ {p _ {i} \in \mathcal {P} _ {\mathrm {K}} ^ {\prime}} S ^ {d e t} \left(p _ {i}, g _ {i}\right) \cdot S ^ {i d} \left(p _ {i}, g _ {i}\right). \tag {11} +$$ + +In other words, $\mathcal{L}^{\mathrm{IDL}}$ is only close to $-1$ if both $S^{id}(p_i,g_i)$ and $S^{det}(p_i,g_i)$ approach 1. Conversely, should either or both not approximate 1, $\mathcal{L}^{\mathrm{IDL}}$ converges to 0. + +(3) Relative Threshold Minimization $\mathcal{L}^{\mathrm{IDL}}$ optimizes detection and identification under fixed thresholds. We also introduce a strategy to minimize the maximum scores for non-mated probes, discouraging the model from assigning high scores to non-mated probes. This helps mitigate the risk of false positive detections in open-set scenarios, enhancing overall robustness. + +To achieve this, we minimize the weighted average of the similarities between the probe and each gallery subject, where the weights are obtained by applying the softmax function to each score. Mathematically, given a probe sample $p_i$ and a set of gallery subjects $\mathcal{G}' = \{g_1, g_2, \dots, g_n\}$ with corresponding similarity scores $s_1, s_2, \dots, s_n$ , we minimize the weighted average as follows: + +$$ +\mathcal {L} ^ {\mathrm {R T M}} = \frac {1}{\sum_ {j = 1} ^ {n} e ^ {s _ {j}}} \sum_ {j = 1} ^ {n} e ^ {s _ {j}} \cdot s _ {j}, \tag {12} +$$ + +where $e^{s_j}$ is the softmax value for each similarity score. + +Fig. 4 (c) shows gradients of $\mathcal{L}_{\mathrm{RTM}}$ w.r.t. non-mated probes. This gradient adapts to the relative position of non-mated probes, assigning higher gradients to samples closer to the gallery. Unlike softmax-based losses and the triplet loss, RTM consistently emphasizes imposter scores that rank higher compared to others within the batch to learn better feature representations for open-set biometrics. This property is desirable as it effectively minimizes the maximum score of each non-mated probe. + +The max function is an alternative to the weighted average. Yet, simply minimizing the maximum score in training would not minimize other scores even if they are close to the maximum. Hence, the weighted average would make it easier for the trained model to generalize during testing. + +# 3.4 Overall Loss + +The overall objective is to minimize the combination of $\mathcal{L}^{\mathrm{IDL}}$ and $\mathcal{L}^{\mathrm{RTM}}$ , + +$$ +L = \mathcal {L} ^ {\mathrm {I D L}} + \lambda \cdot \mathcal {L} ^ {\mathrm {R T M}}. \tag {13} +$$ + +where $\lambda$ is the loss weight. Compared to existing loss functions, our loss function treats genuine and imposter scores differently. It aims to reduce open-set thresholds by pushing non-mated probes away from gallery while pulling genuine scores toward their corresponding galleries, similar to open-set evaluation. Moreover, the relative magnitude of non-mated scores is utilized by RTM to assign larger gradients to harder non-mated probes. + +# 4 Experiments + +# 4.1 Experimental Setup + +Evaluation Protocols. In face recognition experiments, we use the official closed- and open-set evaluation protocols of IJB-S [28]. However, gait recognition and person reID datasets typically lack official open-set evaluation protocols, where some probe subjects are excluded from the gallery. Naïvely removing gallery subjects often results in highly unstable outcomes. This instability arises because threshold calculation heavily relies on a small fraction (1%) of non-mated probes due to the limited number of probes in these datasets. For instance, in a dataset with 500 non-mated probes, the threshold is computed from the highest scores of only 5 probes. + +To ensure robust open-set evaluation, we create $N (= 50)$ random sets of non-mated probes and compute FNIR@1%FPIR for each set. In each set, $q\% = 21.5\%$ subjects are designated as non-mated. We report the median value and the standard deviation from $N$ trials. The choice of median over mean is deliberate to avoid undue influence from extreme values and provide a more stable metric. + +Evaluation Metrics. We evaluate in both open- and closed-set settings. In closed-set settings, we use the Cumulative Matching Curve (CMC) at rank-1; for open-set, we use FNIR@FPIR at $1\%$ and $0.1\%$ FPIR and rank $R = 20$ . + +Datasets. For face recognition, we train on WebFace4M [80] and test on IJB-S [28]. Our experiments in gait recognition involve the Gait3D [77] and GREW [79] datasets, while for person reID, we employ the MEVID [8] and + +Table 1: Summary of datasets used in our experiments. [Keys: Mod.=Modality; Img.=Open-set image recognition; Subj.=the number of subjects ; Media=the number of images for face/image recognition datasets and the number of sequences/tracklets for gait and video-based person reID datasets] + +
DatasetYear Mod.TrainTest
Subj.MediaSubj.Media
IJB-S [28]2018Face-202>10M
Gait3D [77]2022Gait3,00018,9406,369
GREW [79]2021Gait20,000102,88724,000
MEVID [8]2023ReID1046,33855
CCVID [17]2022ReID75948151
+ +CCVID [17] datasets. Further details are provided in Tab. 1. + +One technical caveat is that although FNIR@1%FPIR implicitly assumes that each subject only appears once in the gallery, a typical gait recognition or person re-id dataset contains more than one gallery template per subject. To + +address this issue, we take the mean feature over all gallery templates for each subject. This benefits the models compared to randomly sampling one gallery template (details are in the Supp.). + +Implementation Details. We implement our proposed approach in PyTorch [49]. For fair comparisons, we follow the experiment setup of our baselines. For details on the hyperparameters, please refer to the Supp. + +# 4.2 Ablation Studies + +We conduct our ablation studies on the Gait3D [77] dataset using GaitBase [13] as the baseline. In the ablation studies, we hold out $20\%$ of subjects in the training set as the validation set. + +# Effect of Det, Id and RTM. + +We ablate the effect of Det, Id and RTM in Tab. 2. RTM consistently enhances both closed- and open-set performance. Id further improves open-set performance but has a minor impact on rank-1 accuracy. Meanwhile, Det significantly enhances both closed- and open-set performance. In summary, each component contributes to its effectiveness. + +# Effect of Hyperparameters. + +We ablate the effect of hyperparameters $\alpha, \beta, \gamma,$ and $\lambda$ in Tab. 2. approach can consistently improve line, GaitBase [13]. Also, while the hyperparameters, our approach with the best set of tested hyperp + +Table 2: Ablation study results. [Keys: $\checkmark =$ Kept; $\checkmark =$ Removed; R@1=rank-1 accuracy] + +
IdDetRTMα,γβλR@1FNIR@1%FPIR
XXX———64.686.43±5.44
XX✓6.00.24.065.185.81±5.78
X✓✓64.485.16±5.13
✓✓✓3.00.24.063.985.22±4.73
9.064.284.08±6.28
✓✓✓6.00.14.063.185.10±5.67
0.463.885.92±5.45
✓✓✓6.00.22.064.883.12±5.94
8.061.685.67±5.15
✓✓✓6.00.24.064.882.99±4.50
+ +# 4.3 Comparison with State of the Art + +Across biometric modalities, existing methods exhibit competitive closed-set recognition performance, as shown in Tab. 3 for face, Tab. 4 for gait, and Tab. 5 for person reID. In particular, in gait recognition, GaitBase [13] and SwinGait [12] showcase strong closed-set performance with a notable rank-1 accuracies of greater than $60\%$ and $70\%$ , respectively, for in-the-wild gait datasets Gait3D [77] and GREW [79]. However, in open-set scenarios, they face challenges, as evidenced by GaitBase's high FNIR of $86.43\%$ at $1\%$ FPIR in Gait3D. + +Table 3: Closed- and open-set face recognition on two official protocols of IJB-S [28]. [Keys: R@1=rank-1 accuracy; 0.1%=FNIR@0.1%FPIR; 1%: FNIR@1%FPIR] + +
ModelSurveillance-SingleSurveillance-Booking
R@1↑0.1%↓1%↓R@1↑0.1%↓1%↓
AdaFace [29]68.4445.1838.3768.8445.9837.80
AdaFace+Ours69.9244.3036.7370.0043.4836.17
+ +It is evident that the inclusion of our loss functions consistently yields improvements in both closed- and open-set performance across three modalities. In face recognition, we boost the rank-1 accuracy of AdaFace [29] to $69.92\%$ $(+1.48\%)$ in Surveillance-to-Single. In the challenging open-set scenario, our losses further demonstrate their effectiveness by reducing FNIR@0.1%FPIR from $45.98\%$ to $43.48\%$ in Surveillance-Booking. + +In gait recognition, when applied to Swin-Gait [12] backbone, our losses improve the rank-1 accuracy to $75.3\%$ $(+2.9\%)$ on Gait3D, and FNIR@1%FPIR decreases to $79.30\%$ $(-4.20\%)$ on Gait3D. In person reID, integrating our loss with CAL [17] reaches a rank-1 accuracy of $84.8\%$ $(+2.2\%)$ and a FNIR of $30.81\%$ $(-3.08\%)$ on MEVID [8]. We conduct the independent two-sample t-test to assess the statistical significance of perfor + +Table 4: Closed- and open-set gait recognition on Gait3D [77] and GREW [79]. SMPLGait [77] takes SMPL parameters as additional input. SwinGait-3D [12] uses our own implementation. [Keys: Rank@1=rank-1 accuracy] + +
ModelRank@1 ↑FNIR@1%FPIR ↓
Gait3DGREWGait3DGREW
GaitSet [4]36.748.496.05±2.3081.17±1.45
GaitPart [14]28.247.696.88±1.8088.43±1.26
GaitGL [38]29.741.594.97±1.2480.55±1.11
SMPLGait [77]46.3—93.18±2.58—
GaitBase [13]64.660.186.43±5.4477.67±1.60
+Ours64.860.182.99±4.5076.55±1.54
SwinGait-3D [12]72.479.383.50±6.2058.32±1.79
+Ours75.380.479.30±6.7457.29±1.83
+ +mance improvements in open-set scenarios. For instance, our approach outperforms CAL on the CCVID [17] dataset, with a p-value of $0.052\%$ . + +Collectively, these findings emphasize the consistent effectiveness of our proposed loss functions in enhancing both closed- and open-set performance across a spectrum of biometric tasks, ultimately improving the models' ability to handle open-set challenges, while maintaining strong closed-set performance. + +Efficiency. We present the floating point operations (FLOPs) of training baselines with and without our loss functions in Tab. 6. It is expected that incorporating our loss into baselines will increase the computational load in training, but the increase is very minor. For example, when SwinGait [12] is used, our loss functions only incur $46.9\%$ of the FLOPs of the triplet loss, which is negligible compared to the model. Moreover, our approach does not incur any computation overhead during testing while improving closed- and open-set performance. + +Table 5: Closed- and open-set person reID on MEVID [8] and CCVID [17]. + +
ModelDatasetRank-1Accuracy ↑FNIR@1%FPIR ↓
AGRL [66]MEVID [8]48.466.67±9.57
TCLNet [21]MEVID [8]48.170.34±11.68
PSTA [63]MEVID [8]48.277.11±11.72
CAL [17]MEVID [8]52.566.95±11.59
+OursMEVID [8]55.164.80±10.71
CAL [17]CCVID [17]82.633.89±7.65
+OursCCVID [17]84.830.81±8.02
+ +Table 6: FLOPs in training different baseline models with and without our approach. We report the mean FLOPs across a minibatch. Our approach incurs no additional FLOP during testing. + +
ModalityModelFLOPS+Triplet+Ours
FaceAdaFace [29]24.19G—503K
GaitGaitBase [13]35.43G524.3K245.8K
GaitSwinGait [12]120G491.5K230.4K
BodyCAL [17]32.75G4.09K15.38K
+ +# 4.4 Visualizations + +Score Distributions. Fig. 5 visualizes the non-mated, maximum non-mated score for each probe, and genuine scores of SwinGait [12], with and without our loss. Our introduced RTM effectively reduces the distribution of maximum-per-probe scores, shifting it towards lower values. After linearly normalizing all scores between 0 and 1, the threshold at $1\%$ FPIR is 0.61 with our loss (Fig. 5 (b)) compared to 0.69 without ours (Fig. 5 (a)). Moreover, our loss does not significantly alter the rightmost portion of the genuine score distribution, enhancing detec + +tion. This analysis offers insights into how our loss changes score distributions, and through which improves open-set performance. + +![](images/88bea38f4110d2aaa25187068b72f57ff1d1f84bae263255a5d3755e50b27da5.jpg) +Fig.5: Score distributions on the test set of Gait3D [77] dataset. Ours (b) effectively reduces the threshold at $1\%$ FPIR while only shifting the genuine score distribution marginally to the left compared to SwinGait [12] (a). Both distributions are shifted and scaled horizontally so that the min/max scores are within $[0,1]$ . + +![](images/e40c633b2b5acffa1a07a4b2234ff56dc6b8a0048e7f6c1dcba3742e20b58b85.jpg) +(a) Face + +![](images/139e2e548bdfcefcdf8f42006380bd9c91aa26073c4d9770523905ddb18e25cb.jpg) +(b) Gait +Fig. 6: Open-set FNs due to detection, identification, and either. The average across the 50 open-set evaluation protocols is plotted. Blue indicates FNs that are shared by the baseline, SwinGait [12], and our approach. Orange indicates FNs of the baseline only. Green indicates FNs of our approach. + +![](images/3bb64effdd26f3bdd71d47e99941ed7a6c51b0e2f8ec97ec859cd1522e65ba8c.jpg) +(c) Person ReID + +False Negatives Breakdown. As illustrated in Fig. 6, across different biometric and generic tasks, almost all of the false negatives are failed detection, e.g. the $1^{\mathrm{st}}$ mated probe in Fig. 1. A small portion of false negatives have both failed identification and detection, e.g. the $3^{\mathrm{rd}}$ mated probe in Fig. 1. When we apply our loss functions, the number of FNs due to detection is reduced and the number of FNs due to identification is comparable to that of the baseline. These results highlight that our approach can improve open-set detection without sacrificing identification in various settings. + +# 5 Conclusions + +Our work addresses the critical gap in open-set biometrics, a practical generalization of the well-studied closed-set biometrics. We introduce novel I-DL and RTM losses aiming to improve biometric models' open-set performance. Through extensive experiments across diverse modalities, including face, gait, and person reID, we demonstrate the efficacy of our approach. We not only advance the state of the art in the challenging open-set domain but also enhance closed-set recognition capabilities. We advocate for increased attention to the open-set problem within the biometric community. + +Limitations While our paper makes significant strides in addressing open-set challenges in biometrics through loss functions, our exploration does not extensively explore dedicated architectures for this specific context. Future research endeavors may benefit from investigating tailored architectures that complement and enhance the effectiveness of our proposed loss functions. + +Potential Societal Impacts. The use of biometric datasets in our work containing personally identifiable images prompts a crucial discussion on privacy and ethical implications. As biometric technologies become more prevalent, it is imperative to ensure that the collection of such datasets is aligned with ethical standards and privacy regulations. + +Acknowledgements. This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via 2022-21102100004. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein. + +# References + +1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: CVPR (2016) +2. Cen, J., Luan, D., Zhang, S., Pei, Y., Zhang, Y., Zhao, D., Shen, S., Chen, Q.: The devil is in the wrongly-classified samples: Towards unified open-set recognition. In: ICLR (2023) +3. Chai, T., Li, A., Zhang, S., Li, Z., Wang, Y.: Lagrange motion analysis and view embeddings for improved gait recognition. In: CVPR (2022) +4. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: AAAI (2019) +5. Chen, C., Ye, M., Jiang, D.: Towards modality-agnostic person re-identification with descriptive query. In: CVPR (2023) +6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE TPAMI (2021) +7. Cui, Y., Kang, Y.: Multi-modal gait recognition via effective spatial-temporal feature fusion. In: CVPR (2023) +8. Davila, D., Du, D., Lewis, B., Funk, C., Van Pelt, J., Collins, R., Corona, K., Brown, M., McCloskey, S., Hoogs, A., et al.: MEVID: multi-view extended videos with identities for video person re-identification. In: WACV (2023) +9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR (2019) +0. Dou, H., Zhang, P., Su, W., Yu, Y., Li, X.: Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In: ECCV (2022) +1. Dou, H., Zhang, P., Su, W., Yu, Y., Lin, Y., Li, X.: Gaitgci: Generative counterfactual intervention for gait recognition. In: CVPR (2023) +2. Fan, C., Hou, S., Huang, Y., Yu, S.: Exploring deep models for practical gait recognition. arXiv preprint arXiv:2303.03301 (2023) +3. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: CVPR (2023) +4. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: CVPR (2020) +5. Geng, C., Huang, S., Chen, S.: Recent advances in open set recognition: A survey. IEEE TPAMI (2020) +6. Grother, P., Ngan, M., Hanaoka, K.: Face recognition vendor test (FRVT) part 2: Identification (2019), https://www.nist.gov/document/nistir827120190911pdf +7. Gu, X., Chang, H., Ma, B., Bai, S., Shan, S., Chen, X.: Clothes-changing person re-identification with rgb modality only. In: CVPR (2022) +8. Gunther, M., Cruz, S., Rudd, E.M., Boult, T.E.: Toward open-set face recognition. In: CVPR Workshops (2017) + +19. Günther, M., Hu, P., Herrmann, C., Chan, C.H., Jiang, M., Yang, S., Dhamija, A.R., Ramanan, D., Beyerer, J., Kittler, J., et al.: Unconstrained face detection and open-set face recognition challenge. In: IJCB. IEEE (2017) +20. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017) +21. Hou, R., Chang, H., Ma, B., Shan, S., Chen, X.: Temporal complementary learning for video person re-identification. In: ECCV (2020) +22. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE TPAMI (2022) +23. Huang, X., Zhu, D., Wang, H., Wang, X., Yang, B., He, B., Liu, W., Feng, B.: Context-sensitive temporal feature learning for gait recognition. In: ICCV (2021) +24. Huang, Y., Shen, P., Tai, Y., Li, S., Liu, X., Li, J., Huang, F., Ji, R.: Improving face recognition from hard samples via distribution distillation loss. In: ECCV (2020) +25. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: CVPR (2020) +26. Huang, Z., Xue, D., Shen, X., Tian, X., Li, H., Huang, J., Hua, X.S.: 3d local convolutional neural networks for gait recognition. In: ICCV (2021) +27. Jain, A.K., Deb, D., Engelsma, J.J.: Biometrics: Trust, but verify. IEEE Transactions on Biometrics, Behavior, and Identity Science (2021) +28. Kalka, N.D., Maze, B., Duncan, J.A., O'Connor, K., Elliott, S., Hebert, K., Bryan, J., Jain, A.K.: IJB-S: IARPA janus surveillance video benchmark. In: BTAS (2018) +29. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. In: CVPR (2022) +30. Kim, M., Liu, F., Jain, A.K., Liu, X.: Cluster and aggregate: Face recognition with large probe set. In: NeurIPS (2022) +31. Kim, M., Su, Y., Liu, F., Jain, A., Liu, X.: Keypoint relative position encoding for face recognition. In: CVPR (2024) +32. Koch, T., Riess, C., Kohler, T.: Lord: Leveraging open-set recognition with unknown data. In: ICCV (2023) +33. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: ICCV (2021) +34. Kuchibhotla, H.C., Malagi, S.S., Chandhok, S., Balasubramanian, V.N.: Unseen classes at a later time? no problem. In: CVPR (2022) +35. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: ICCV (2023) +36. Li, W., Hou, S., Zhang, C., Cao, C., Liu, X., Huang, Y., Zhao, Y.: An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In: CVPR (2023) +37. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: ECCV (2022) +38. Lin, B., Zhang, S., Yu, X.: Gait recognition via effective global-local feature representation and local temporal aggregation. In: ICCV (2021) +39. Liu, F., Ashbaugh, R., Chimitt, N., Hassan, N., Hassani, A., Jaiswal, A., Kim, M., Mao, Z., Perry, C., Ren, Z., et al.: Farsight: A physics-driven whole-body biometric system at large distance and altitude. In: WACV (2024) +40. Liu, F., Kim, M., Gu, Z., Jain, A., Liu, X.: Learning clothing and pose invariant 3d shape representation for long-term person re-identification. In: ICCV (2023) +41. Liu, F., Kim, M., Jain, A., Liu, X.: Controllable and guided face synthesis for unconstrained face recognition. In: ECCV (2022) + +42. Liu, F., Kim, M., Ren, Z., Liu, X.: Distilling CLIP with dual guidance for learning discriminative human body shape representation. In: CVPR (2024) +43. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: CVPR (2017) +44. Ma, K., Fu, Y., Zheng, D., Cao, C., Hu, X., Huang, Y.: Dynamic aggregated network for gait recognition. In: CVPR (2023) +45. Mekhazni, D., Dufau, M., Desrosiers, C., Pedersoli, M., Granger, E.: Camera alignment and weighted contrastive learning for domain adaptation in video person reid. In: WACV (2023) +46. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: CVPR (2021) +47. Minaee, S., Abdelrashidi, A., Su, H., Bennamoun, M., Zhang, D.: Biometrics recognition using deep learning: A survey. Artificial Intelligence Review (2023) +48. Ni, H., Li, Y., Gao, L., Shen, H.T., Song, J.: Part-aware transformer for generalizable person re-identification. In: ICCV (2023) +49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019) +50. Qin, L., Wang, M., Deng, C., Wang, K., Chen, X., Hu, J., Deng, W.: Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. TCSVT (2023) +51. Ren, Z., Su, Y., Liu, X.: Chatgpt-powered hierarchical comparisons for image classification. In: NeurIPS (2023) +52. Ross, A., Banerjee, S., Chen, C., Chowdhury, A., Mirjalili, V., Sharma, R., Swearinen, T., Yadav, S.: Some research problems in biometrics: The future beckons. In: ICB (2019) +53. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: CVPR (2015) +54. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3d gait recognition with point clouds. In: CVPR (2023) +55. Shen, C., Yu, S., Wang, J., Huang, G.Q., Wang, L.: A comprehensive survey on deep gait recognition: algorithms, datasets and challenges. arXiv preprint arXiv:2206.13732 (2022) +56. Somers, V., De Vleeschouwer, C., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: CVPR (2023) +57. Sun, Y., Cheng, C., Zhang, Y., Zhang, C., Zheng, L., Wang, Z., Wei, Y.: Circle loss: A unified perspective of pair similarity optimization. In: CVPR (2020) +58. Tabassi, E., Watson, C., Fiumara, G., Salamon, W., Flanagan, P., Cheng, S.L.: Performance evaluation of fingerprint open-set identification algorithms. In: IJCB (2014) +59. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need. In: ICLR (2022) +60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: CVPR (2018) +61. Wang, L., Liu, B., Liang, F., Wang, B.: Hierarchical spatio-temporal representation learning for gait recognition. In: ICCV (2023) +62. Wang, M., Guo, X., Lin, B., Yang, T., Zhu, Z., Li, L., Zhang, S., Yu, X.: Dygait: Exploiting dynamic representations for high-performance gait recognition. In: ICCV (2023) +63. Wang, Y., Zhang, P., Gao, S., Geng, X., Lu, H., Wang, D.: Pyramid spatial-temporal aggregation for video-based person re-identification. In: ICCV (2021) + +64. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards AUC-oriented open-set recognition. In: NeurIPS (2022) +65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: ECCV (2016) +66. Wu, Y., Bourahla, O.E.F., Li, X., Wu, F., Tian, Q., Zhou, X.: Adaptive graph representation learning for video person re-identification. TIP (2020) +67. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. In: NeurIPS (2022) +68. Ye, D., Fan, C., Ma, J., Liu, X., Yu, S.: Biggait: Learning gait representation you want by large vision models. In: CVPR (2024) +69. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: CVPR (2019) +70. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: CVPR (2020) +71. Zahra, A., Perwaiz, N., Shahzad, M., Fraz, M.M.: Person re-identification: A retrospective on domain specific open challenges and future trends. Pattern Recognition (2023) +72. Zhai, Y., Liu, Z., Wu, Z., Wu, Y., Zhou, C., Doermann, D., Yuan, J., Hua, G.: Soar: Scene-debiasing open-set action recognition. In: ICCV (2023) +73. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: ECCV (2020) +74. Zhang, Z., Tran, L., Liu, F., Liu, X.: On learning disentangled representations for gait recognition. IEEE TPAMI (2020) +75. Zhang, Z., Tran, L., Yin, X., Atoum, Y., Liu, X., Wan, J., Wang, N.: Gait recognition via disentangled representation learning. In: CVPR (2019) +76. Zheng, J., Liu, X., Gu, X., Sun, Y., Gan, C., Zhang, J., Liu, W., Yan, C.: Gait recognition in the wild with multi-hop temporal switch. In: ACMMM (2022) +77. Zheng, J., Liu, X., Liu, W., He, L., Yan, C., Mei, T.: Gait recognition in the wild with dense 3d representations and a benchmark. In: CVPR (2022) +78. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: CVPR (2021) +79. Zhu, Z., Guo, X., Yang, T., Huang, J., Deng, J., Huang, G., Du, D., Lu, J., Zhou, J.: Gait recognition in the wild: A benchmark. In: ICCV (2021) +80. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., et al.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: CVPR (2021) \ No newline at end of file diff --git a/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/images.zip b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fbae30a664ea8d313816c91f0e5ccb0fa57d6ce3 --- /dev/null +++ b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31abf3237c4a862352aa38ad09a5088db71f0d3ea94d105ae6232b3d768b1e17 +size 394548 diff --git a/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/layout.json b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f91d08343fcf3ab8c5f3e1a0dc4ab4e1817dd3a --- /dev/null +++ b/2024/Open-Set Biometrics_ Beyond Good Closed-Set Models/layout.json @@ -0,0 +1,13294 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 147 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 147 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 168, + 465, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 168, + 465, + 181 + ], + "spans": [ + { + "bbox": [ + 149, + 168, + 465, + 181 + ], + "type": "text", + "content": "Yiyang Su, Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 185, + 190, + 430, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 190, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 185, + 190, + 430, + 213 + ], + "type": "text", + "content": "Michigan State University, East Lansing, MI 48824, USA {suyiyan1, kimminc2, liufeng6, jain, liuxm}@msu.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 244, + 455, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 244, + 455, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 244, + 455, + 452 + ], + "type": "text", + "content": "Abstract. Biometric recognition has primarily addressed closed-set identification, assuming all probe subjects are in the gallery. However, most practical applications involve open-set biometrics, where probe subjects may or may not be present in the gallery. This poses distinct challenges in effectively distinguishing individuals in the gallery while minimizing false detections. While it is commonly believed that powerful biometric models can excel in both closed- and open-set scenarios, existing loss functions are inconsistent with open-set evaluation. They treat genuine (mated) and imposter (non-mated) similarity scores symmetrically and neglect the relative magnitudes of imposter scores. To address these issues, we simulate open-set evaluation using minibatches during training and introduce novel loss functions: (1) the identification-detection loss optimized for open-set performance under selective thresholds and (2) relative threshold minimization to reduce the maximum negative score for each probe. Across diverse biometric tasks, including face recognition, gait recognition, and person re-identification, our experiments demonstrate the effectiveness of the proposed loss functions, significantly enhancing open-set performance while positively impacting closed-set performance. Our code and models are available here." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "spans": [ + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "text", + "content": "Keywords: Open-set biometrics " + }, + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "text", + "content": " Face recognition " + }, + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "text", + "content": " Gait recognition " + }, + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 464, + 453, + 485 + ], + "type": "text", + "content": " Person reID" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 507, + 230, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 507, + 230, + 520 + ], + "spans": [ + { + "bbox": [ + 133, + 507, + 230, + 520 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "Biometrics, the study of unique physiological or behavioral characteristics for human identification, plays a pivotal role in applications such as access control, surveillance, and forensic analysis [52]. Each biometric task, from face recognition, gait recognition to person re-identification (reID), uses features tailored to its specific requirements [39, 55, 71]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In real-world applications, after test subjects are enrolled in the gallery, biometric systems inevitably encounter non-mated subjects, those not present in the gallery. Thus, it becomes imperative for such systems to demonstrate robust performance in the open-set framework with non-mated probes [16, 18, 28, 58]. The closed-set framework assumes all probe subjects are mated and predominantly focuses on identification, associating probes with the closest match in the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 114, + 482, + 223 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 223 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 223 + ], + "type": "image", + "image_path": "1894c19c93dc6f73e2dc162e99b0289153a08461e1bce4793f1754b8e568e40c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": "Fig. 1: A toy example for FNIR@FPIR calculation. The score matrix and distribution between non-mated probes and all gallery subjects are shown in blue. The maximum score for each non-mated probe is shown in red. A percentile of their distribution determines the threshold. A user-defined percentile (e.g., " + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": ") sets the threshold to be 0.75 as 1 out of 3 (the " + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": ") non-mated probes is incorrectly classified as \"in the gallery\". The genuine scores of mated probes are shown in green. A mated probe is an FN if either of the conditions is met: (1) in detection (Det.), it is incorrectly classified as \"not in the gallery\" (" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": ") as the genuine score is below the threshold (the " + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": " mated probe); (2) its genuine score is not within a specified rank " + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": ", e.g., 1, in the gallery (the " + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": " mated probe). Hence, FNIR@33%FPIR is " + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 231, + 483, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": "gallery. However, the open-set framework introduces an additional challenge: detection, determining whether a probe subject is mated. For example, in airport surveillance, faces in CCTV videos are compared against a criminal watch list. The goal is to reduce the misidentification (false positives) of non-mated subjects such as airport staff and innocent passengers (improving detection), while simultaneously ensuring accurate recognition of known criminals (enhancing identification). For this reason, open-set biometrics is a harder and more general task than closed-set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "text", + "content": "The biometric community has predominantly centered on the closed-set framework. While some face recognition benchmarks have open-set evaluation protocols [18,19,28], they are primarily used for evaluation rather than for developing tailored strategies to address the open-set challenges. In gait recognition and person reID, most benchmarks only provide official closed-set protocols, lacking open-set ones. To address this gap, we take the initiative to introduce open-set benchmarks and propose strategies to train models specifically for improving open-set performances while maintaining closed-set performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The false negative identification rate at a given false positive identification rate (FNIR@FPIR) is a de facto open-set metric [16]. We first compute the similarity score matrix between non-mated probes and the gallery, from which we extract the maximum score for each non-mated probe across the gallery, forming the maximum-per-probe non-mated score distribution. This distribution represents the likelihood that a non-mated probe has a match in the gallery. A threshold is set based on a user-defined percentile (e.g., " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ") of this distribution. Mated probes with scores below this threshold are counted as false negatives (FNs), as they are incorrectly considered not to be in the gallery. Similar to the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "closed-set scenario, a mated probe is also an FN if its gallery match doesn't rank among the gallery set within a user-specified rank " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ". FNIR is the proportion of mated probes classified as FNs. Fig. 1 depicts its calculation pipeline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 480, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 480, + 249 + ], + "type": "text", + "content": "Unlike closed-set or verification metrics, FNIR@FPIR introduces unique challenges. It employs a threshold based on maximum scores for each non-mated probe, creating an asymmetry between genuine and imposter scores. SOTA biometric recognition methods achieve strong performance in closed-set scenarios leveraging triplet loss [20] and/or softmax cross-entropy based losses [9, 25, 29, 46, 60], both of which excel in reducing intra-person variance and increasing inter-person variance. However, they do not take into account the fact that the non-mated score distribution determines the threshold for the false negative." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 357 + ], + "type": "text", + "content": "In other words, conventional loss functions, while effective in closed-set scenarios, exhibit limitations in addressing the nuances of the FNIR@FPIR metric, which necessitate a handling of the maximum of imposter scores. These high-scoring non-mated probes disproportionately influence the threshold setting for false negatives, yet conventional loss functions often treat them equivalently to other scores. As a result, they face challenges in open-set scenarios. GaitBase [13], a SOTA gait recognition model, can only achieve a " + }, + { + "bbox": [ + 130, + 251, + 481, + 357 + ], + "type": "inline_equation", + "content": "61\\%" + }, + { + "bbox": [ + 130, + 251, + 481, + 357 + ], + "type": "text", + "content": " FNIR@1%FPIR on the training set and " + }, + { + "bbox": [ + 130, + 251, + 481, + 357 + ], + "type": "inline_equation", + "content": "86\\%" + }, + { + "bbox": [ + 130, + 251, + 481, + 357 + ], + "type": "text", + "content": " on the test set. This challenge extends to face recognition and person reID." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": "To address this challenge, we propose to create simulated open-set evaluation episodes during training, whose score distributions serve as proxies to those of the test set. Specifically, we introduce an identification-detection loss tailored to optimize the two core elements in open-set biometrics, i.e., detection and identification. For detection, the model is trained to accurately pinpoint the exact probes whose mate is present in the gallery set; for identification, the model learns to establish associations between the gallery and probe samples of mated subjects. By integrating the aspects of detection and identification within training batches, we improve the model's efficacy in both domains, thus yielding enhanced testing performance. Further, we minimize similarity scores between non-mated subjects and the gallery, enabling a more comprehensive detection of mated subjects (see Fig. 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 505, + 480, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 480, + 553 + ], + "type": "text", + "content": "Empirically, we apply our proposed loss function to three biometric modalities, face, gait, and person reID, and show performance improvement in all tasks. While our focus is open-set biometrics, we also observe performance gain in closed-set search due to improved feature representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 555, + 329, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 555, + 329, + 566 + ], + "spans": [ + { + "bbox": [ + 146, + 555, + 329, + 566 + ], + "type": "text", + "content": "To summarize, our contributions include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 580, + 479, + 662 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 135, + 580, + 479, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 580, + 479, + 626 + ], + "spans": [ + { + "bbox": [ + 135, + 580, + 479, + 626 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 135, + 580, + 479, + 626 + ], + "type": "text", + "content": " We show that the existing loss functions, e.g., triplet loss and softmax loss, do not treat genuine and imposter training samples in accordance with open-set evaluation. This inconsistency sheds light on why the open-set challenges have not been adequately addressed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 629, + 479, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 629, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 135, + 629, + 479, + 662 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 135, + 629, + 479, + 662 + ], + "type": "text", + "content": " We propose a novel loss function that leverages open-set mini-batches during training to optimize identification and detection performance and minimize open-set detection thresholds." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 113, + 208, + 144 + ], + "blocks": [ + { + "bbox": [ + 161, + 113, + 208, + 144 + ], + "lines": [ + { + "bbox": [ + 161, + 113, + 208, + 144 + ], + "spans": [ + { + "bbox": [ + 161, + 113, + 208, + 144 + ], + "type": "image", + "image_path": "8229de76d49520d971e21e38a586d6917b82ff56eb1fcba2dd37204d1abf2606.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 200, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 266 + ], + "type": "text", + "content": "Fig. 2: Comparison between open-set recognition and biometrics. Closed-set recognition has the same set of classes for training and testing. Open-set recognition involves testing on unseen classes during training. On the other hand, in biometrics, the sets of training and testing subjects are always disjoint. In closed-set biometrics, all probe subjects are in the gallery, and in open-set biometrics, some probe subjects are not enrolled in the gallery. [Legends: ▪ gallery subjects, ▪ probe subjects]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 144, + 214, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 144, + 214, + 163 + ], + "spans": [ + { + "bbox": [ + 156, + 144, + 214, + 163 + ], + "type": "text", + "content": "Training Testing classes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 163, + 206, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 163, + 206, + 183 + ], + "spans": [ + { + "bbox": [ + 164, + 163, + 206, + 183 + ], + "type": "text", + "content": "Closed-Set Recognition" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 289, + 144 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 289, + 144 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 289, + 144 + ], + "type": "image", + "image_path": "3287ea4dc50c33faa3b3076a0d310355d1537a07da695711a9af3e152d35debc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 229, + 144, + 288, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 144, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 229, + 144, + 288, + 163 + ], + "type": "text", + "content": "Training Testing classes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 238, + 163, + 281, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 163, + 281, + 183 + ], + "spans": [ + { + "bbox": [ + 238, + 163, + 281, + 183 + ], + "type": "text", + "content": "Open-Set Recognition" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 299, + 116, + 323, + 142 + ], + "blocks": [ + { + "bbox": [ + 299, + 116, + 323, + 142 + ], + "lines": [ + { + "bbox": [ + 299, + 116, + 323, + 142 + ], + "spans": [ + { + "bbox": [ + 299, + 116, + 323, + 142 + ], + "type": "image", + "image_path": "5b732f0aacb12364654e5d6f26e60e7327007c202f5d086397307626c56fa04b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 144, + 331, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 144, + 331, + 163 + ], + "spans": [ + { + "bbox": [ + 302, + 144, + 331, + 163 + ], + "type": "text", + "content": "Training subjects" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 114, + 373, + 143 + ], + "blocks": [ + { + "bbox": [ + 343, + 114, + 373, + 143 + ], + "lines": [ + { + "bbox": [ + 343, + 114, + 373, + 143 + ], + "spans": [ + { + "bbox": [ + 343, + 114, + 373, + 143 + ], + "type": "image", + "image_path": "ff77196acab4689c6cb553d048134e6766f969b6c2b2770e143e63d8afc3e0c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 144, + 369, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 144, + 369, + 163 + ], + "spans": [ + { + "bbox": [ + 343, + 144, + 369, + 163 + ], + "type": "text", + "content": "Testing subjects" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 163, + 356, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 356, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 356, + 182 + ], + "type": "text", + "content": "Closed-Set Biometrics" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 384, + 118, + 408, + 142 + ], + "blocks": [ + { + "bbox": [ + 384, + 118, + 408, + 142 + ], + "lines": [ + { + "bbox": [ + 384, + 118, + 408, + 142 + ], + "spans": [ + { + "bbox": [ + 384, + 118, + 408, + 142 + ], + "type": "image", + "image_path": "4d382f7589b5738bf0e795179809598bc523c3e350e4f6f81bf39439f8df7c02.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 387, + 144, + 416, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 144, + 416, + 163 + ], + "spans": [ + { + "bbox": [ + 387, + 144, + 416, + 163 + ], + "type": "text", + "content": "Training subjects" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 118, + 462, + 141 + ], + "blocks": [ + { + "bbox": [ + 425, + 118, + 462, + 141 + ], + "lines": [ + { + "bbox": [ + 425, + 118, + 462, + 141 + ], + "spans": [ + { + "bbox": [ + 425, + 118, + 462, + 141 + ], + "type": "image", + "image_path": "3feb03aabdf216d3310d60c1dc57b76fd707d2f7c537c2fd8ac7b7130935b009.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 427, + 144, + 455, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 144, + 455, + 163 + ], + "spans": [ + { + "bbox": [ + 427, + 144, + 455, + 163 + ], + "type": "text", + "content": "Testing subjects" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 402, + 163, + 441, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 163, + 441, + 182 + ], + "spans": [ + { + "bbox": [ + 402, + 163, + 441, + 182 + ], + "type": "text", + "content": "Open-Set Biometrics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 288, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 288, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 135, + 288, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 135, + 288, + 480, + 324 + ], + "type": "text", + "content": " We experimentally demonstrate that our proposed loss function can effectively improve both closed- and open-set performance across diverse biometric modalities, including face, gait, and person reID." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 339, + 237, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 237, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 237, + 352 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 362, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 507 + ], + "type": "text", + "content": "Open-Set Biometrics. While various biometric recognition tasks, such as face recognition [9, 24, 25, 29-31, 41, 43, 46, 50, 60], gait recognition [3, 4, 7, 10-14, 23, 26, 36-38, 44, 54, 61, 62, 68, 74-77], and person reID [5, 17, 21, 35, 36, 40, 42, 45, 48, 56, 63, 66], have witnessed significant progress, the majority of their focus remains on closed-set scenarios. Although EVM has been applied to open-set face recognition [18], EVM takes extracted features, instead of images/videos, as input. Our approach improves feature representation and EVM improves score calculation. While some face recognition datasets provide opens-set evaluation protocols [28], open-set performance is often perceived as a consequence rather than a direct objective in advancing recognition models. Further, even the most recent gait recognition [68] and person reID [42] models are exclusively evaluated under closed-set settings due to the absence of open-set evaluation protocols." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Open-Set Recognition. While open-set biometrics is relatively underexplored, open-set recognition [1,2,6,22,32-34,51,59,64,67,69,72,73,78] has attracted attention in the literature. Open-set recognition distinguishes between known and unknown classes in image classification. In both open-set image recognition and open-set biometrics, the primary concern is how the models deal with instances that belong to unknown classes or non-mated probes. However, as illustrated in Fig. 2, in open-set recognition known classes come from the training set and the unknown classes refer to classes not used for training [15]. On the other hand, even in closed-set biometrics, all test subjects are typically not used during training and mated probes belong to test subjects enrolled in the gallery [27,47]. Open-set biometrics differs from closed-set biometrics in that there are non-mated subjects are neither in the training set nor in the gallery [16,18,58]. In" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 455, + 178 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 455, + 178 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 455, + 178 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 455, + 178 + ], + "type": "image", + "image_path": "6898f0d2ae795b6537516797e76054f9c31695e15459e8316eb56ca4572e2159.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 187, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 130, + 187, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 482, + 232 + ], + "type": "text", + "content": "Fig. 3: Overview of our approach. We introduce an additional modality-agnostic loss function specifically designed for open-set biometrics, which simultaneously optimizes identification (Id.) and detection (Det.) and Relatie Threshold Minimization (RTM) for open-set thresholds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 255, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 293 + ], + "type": "text", + "content": "evaluation, open-set recognition models typically use closed-set recognition accuracy and AUROC for detecting classes not seen during training [15] and open-set biometrics uses FNIR@FPIR [16]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 311, + 202, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 202, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 202, + 323 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 338, + 268, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 268, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 268, + 349 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{test}}" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": " be the test set, " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{G} \\subset S_{\\mathrm{test}}" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": " represent the gallery set containing mated subjects, and " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\subset S_{\\mathrm{test}}" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": " denote the probe set consisting of test samples. For each probe sample " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathcal{P}" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": ", there are two possible outcomes: (1) If " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": " is mated, i.e., its corresponding subject " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": " is in " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": ", the biometric model must identify the correct subject. (2) If " + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 358, + 482, + 443 + ], + "type": "text", + "content": " is non-mated, the model must reject it. The open-set scenario poses a more challenging problem than the traditional closed-set one, as the model needs to recognize mated subjects while rejecting non-mated ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 461, + 365, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 365, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 365, + 473 + ], + "type": "text", + "content": "3.2 Revisiting Loss Functions for Biometrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 482, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 481, + 533 + ], + "type": "text", + "content": "In this section, we compute gradients of two common loss functions in biometrics: softmax and triplet loss. Many loss functions are derived from them and share similar characteristics. For instance, center loss [65] can be analyzed similarly to softmax loss, and the logic of triplet loss can also be applied to circle loss [57]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "content": "Softmax loss. The softmax loss, often employed in closed-set classification, operates by computing the cross-entropy between the normalized logit of a sample and its corresponding ground truth class. Mathematically, for a sample " + }, + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "content": " in a batch, the softmax loss " + }, + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{softmax}}(i)" + }, + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 235, + 606, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 606, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 235, + 606, + 481, + 634 + ], + "type": "interline_equation", + "content": "L _ {\\text {s o f t m a x}} (i) = - \\log \\left(\\frac {e ^ {z _ {i , s}}}{\\sum_ {t} e ^ {z _ {i , t}}}\\right). \\tag {1}", + "image_path": "dc8f44e03f25b89275fd09038c577a2f50107186b3ae6737b24059387243fde0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "z_{i,s}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " denotes the logit for sample " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " belonging to subject " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ". This loss aims to minimize the negative log-likelihood of the true subject " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " for sample " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ". For one" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 446, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 446, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 446, + 129 + ], + "type": "text", + "content": "sample, the gradient of the softmax loss " + }, + { + "bbox": [ + 132, + 116, + 446, + 129 + ], + "type": "inline_equation", + "content": "w.r.t." + }, + { + "bbox": [ + 132, + 116, + 446, + 129 + ], + "type": "text", + "content": " the logit " + }, + { + "bbox": [ + 132, + 116, + 446, + 129 + ], + "type": "inline_equation", + "content": "z_{i,s}" + }, + { + "bbox": [ + 132, + 116, + 446, + 129 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 242, + 138, + 481, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 138, + 481, + 165 + ], + "spans": [ + { + "bbox": [ + 242, + 138, + 481, + 165 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L _ {\\mathrm {s o f t m a x}} (i)}{\\partial z _ {i , s}} = \\frac {e ^ {z _ {i , s}}}{\\sum_ {t} e ^ {z _ {i , t}}} - \\delta_ {i, s}, \\tag {2}", + "image_path": "1d7b06931c3a249df36ee29e230eb44dc14a07d165bc7d88a9bd8412bb61f88f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\frac{e^{z_{i,s}}}{\\sum_{j=1}^{N} e^{z_{j,s}}}" + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "text", + "content": " represents the probability that sample " + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "text", + "content": " belongs to subject " + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\delta_{i,c}" + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "text", + "content": " is the Kronecker delta, which is 1 when " + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "inline_equation", + "content": "i = c" + }, + { + "bbox": [ + 131, + 172, + 481, + 213 + ], + "type": "text", + "content": " (the true subject) and 0 otherwise. We visualize gradients of the softmax loss in Fig. 4(a)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 214, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 214, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 482, + 298 + ], + "type": "text", + "content": "However, it overlooks the relative magnitude of each imposter score within the batch. An imposter score may be deemed small by, e.g., AdaFace [29] standards but still rank high among maximum-per-probe non-mated scores, raising the open-set threshold. Further, the softmax loss assumes that each sample can be assigned to one of the training subjects, which is not the case in open-set settings. Consequently, the softmax loss is not inherently designed to handle open-set biometrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 315, + 345, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 345, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 345, + 327 + ], + "type": "text", + "content": "Triplet loss. The triplet loss [53] is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 212, + 337, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 337, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 212, + 337, + 481, + 350 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {t r i}} (a, p, n) = \\max (0, d (a, p) - d (a, n) + m). \\tag {3}", + "image_path": "2cd7a451d13bfe9ae0f5ba95ddf9a068953e8645b03f3be702b27e8b7c5b3608.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "d(a, p)" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " is the distance between the anchor " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " and positive sample " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "d(a, n)" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " is between " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " and the negative sample " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " is a constant margin. The partial derivatives of " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "L(a, p, n)" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "d(a, p)" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "inline_equation", + "content": "d(a, n)" + }, + { + "bbox": [ + 131, + 358, + 482, + 396 + ], + "type": "text", + "content": " are piecewise functions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 203, + 414, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 414, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 203, + 414, + 481, + 444 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L _ {\\operatorname {t r i}} (a , p , n)}{\\partial d (a , p)} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} d (a, p) + m > d (a, n) \\\\ 0, & \\text {i f} d (a, p) + m \\leq d (a, n) \\end{array} ; \\right. \\tag {4}", + "image_path": "792689d578ed392a25155f96148b5eb8a60abb40827c11412d18785776b9d0e4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 448, + 481, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 448, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 203, + 448, + 481, + 478 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L _ {\\operatorname {t r i}} (a , p , n)}{\\partial d (a , n)} = \\left\\{ \\begin{array}{l l} - 1, & \\text {i f} d (a, p) + m > d (a, n) \\\\ 0, & \\text {i f} d (a, p) + m \\leq d (a, n) \\end{array} \\right.. \\tag {5}", + "image_path": "3a6a8e79ffc1896f1c71b6dc73efd7f78a3966fad63eda842f9e489f229b52c8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 486, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 482, + 605 + ], + "type": "text", + "content": "The magnitude of the gradient is constant regardless of the distance between two samples. This inability to differentiate samples in the gradient implies that all negative samples in the triplet are contributing equally, which is not reflective of the open-set scenario. One remedy often used with triplet loss is hard negative mining. However, it comes with a critical drawback of being susceptible to label noise [70]. Noisy and mislabeled samples are often considered hard negatives, which can make the loss function sensitive to noisy data when focusing on the hardest samples. Thus, we believe that a loss function that considers all negative samples while being able to adjust the gradient magnitude based on the anchor-negative distances would be effective in open-set biometrics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 482, + 666 + ], + "type": "text", + "content": "Hence, we propose a novel loss function, shown in Fig. 3. It optimizes identification and detection at various thresholds and minimizes non-mated scores to reduce open-set thresholds. It can be applied to any biometric modality. Unlike existing loss functions, it is consistent with FNIR@FPIR in that it treats genuine and imposter scores asymmetrically and tackles large non-mated scores." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 126, + 187, + 175 + ], + "blocks": [ + { + "bbox": [ + 144, + 117, + 178, + 124 + ], + "lines": [ + { + "bbox": [ + 144, + 117, + 178, + 124 + ], + "spans": [ + { + "bbox": [ + 144, + 117, + 178, + 124 + ], + "type": "text", + "content": "(a) SoftMax" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 126, + 187, + 175 + ], + "lines": [ + { + "bbox": [ + 137, + 126, + 187, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 126, + 187, + 175 + ], + "type": "image", + "image_path": "52c76ac312b534708c2225e80f66d0cc7ce2254ab9a3e9cde2a79db71cede52d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 178, + 194, + 186 + ], + "lines": [ + { + "bbox": [ + 138, + 178, + 194, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 178, + 194, + 186 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 138, + 178, + 194, + 186 + ], + "type": "text", + "content": " Genuine probe " + }, + { + "bbox": [ + 138, + 178, + 194, + 186 + ], + "type": "inline_equation", + "content": "(p)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 210, + 194, + 220 + ], + "blocks": [ + { + "bbox": [ + 138, + 193, + 173, + 201 + ], + "lines": [ + { + "bbox": [ + 138, + 193, + 173, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 193, + 173, + 201 + ], + "type": "text", + "content": "Threshold" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 138, + 201, + 170, + 208 + ], + "lines": [ + { + "bbox": [ + 138, + 201, + 170, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 170, + 208 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 138, + 201, + 170, + 208 + ], + "type": "text", + "content": " Gradient" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 138, + 210, + 194, + 220 + ], + "lines": [ + { + "bbox": [ + 138, + 210, + 194, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 194, + 220 + ], + "type": "image", + "image_path": "9e4793ee91db32ed3139b83a6b44fa7ca4588bbf0788886e7406ed3bbc271270.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 220, + 194, + 228 + ], + "lines": [ + { + "bbox": [ + 138, + 220, + 194, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 194, + 228 + ], + "type": "text", + "content": "Gradient Magnitude" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 200, + 126, + 248, + 175 + ], + "blocks": [ + { + "bbox": [ + 138, + 186, + 194, + 193 + ], + "lines": [ + { + "bbox": [ + 138, + 186, + 194, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 194, + 193 + ], + "type": "inline_equation", + "content": "\\ominus" + }, + { + "bbox": [ + 138, + 186, + 194, + 193 + ], + "type": "text", + "content": " Imposter probe " + }, + { + "bbox": [ + 138, + 186, + 194, + 193 + ], + "type": "inline_equation", + "content": "(n)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 201, + 118, + 245, + 124 + ], + "lines": [ + { + "bbox": [ + 201, + 118, + 245, + 124 + ], + "spans": [ + { + "bbox": [ + 201, + 118, + 245, + 124 + ], + "type": "text", + "content": "(b) Triplet Loss" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 126, + 248, + 175 + ], + "lines": [ + { + "bbox": [ + 200, + 126, + 248, + 175 + ], + "spans": [ + { + "bbox": [ + 200, + 126, + 248, + 175 + ], + "type": "image", + "image_path": "9a830a71b69130fadc524c4716deb1d21977307162096105aa84fd8873d5218c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 200, + 175, + 247, + 223 + ], + "blocks": [ + { + "bbox": [ + 200, + 175, + 247, + 223 + ], + "lines": [ + { + "bbox": [ + 200, + 175, + 247, + 223 + ], + "spans": [ + { + "bbox": [ + 200, + 175, + 247, + 223 + ], + "type": "image", + "image_path": "5577a7bd635537aa144436de87afdb42dc6b0919ea9794c11b1cd967daf7b703.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 331, + 122, + 380, + 171 + ], + "blocks": [ + { + "bbox": [ + 280, + 118, + 303, + 124 + ], + "lines": [ + { + "bbox": [ + 280, + 118, + 303, + 124 + ], + "spans": [ + { + "bbox": [ + 280, + 118, + 303, + 124 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 127, + 325, + 134 + ], + "lines": [ + { + "bbox": [ + 261, + 127, + 325, + 134 + ], + "spans": [ + { + "bbox": [ + 261, + 127, + 325, + 134 + ], + "type": "text", + "content": "Identification Detection" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 139, + 320, + 159 + ], + "lines": [ + { + "bbox": [ + 268, + 139, + 320, + 159 + ], + "spans": [ + { + "bbox": [ + 268, + 139, + 320, + 159 + ], + "type": "text", + "content": "different " + }, + { + "bbox": [ + 268, + 139, + 320, + 159 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 268, + 139, + 320, + 159 + ], + "type": "text", + "content": " gradient magnitude based on threshold" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 332, + 115, + 380, + 122 + ], + "lines": [ + { + "bbox": [ + 332, + 115, + 380, + 122 + ], + "spans": [ + { + "bbox": [ + 332, + 115, + 380, + 122 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 332, + 115, + 380, + 122 + ], + "type": "text", + "content": " Most Emphasized" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 122, + 380, + 171 + ], + "lines": [ + { + "bbox": [ + 331, + 122, + 380, + 171 + ], + "spans": [ + { + "bbox": [ + 331, + 122, + 380, + 171 + ], + "type": "image", + "image_path": "4a1469d03aa844543f7c8ac099f771de10b7f5c1d72b440e31d5bf457079ae5b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 331, + 179, + 380, + 228 + ], + "blocks": [ + { + "bbox": [ + 268, + 182, + 320, + 196 + ], + "lines": [ + { + "bbox": [ + 268, + 182, + 320, + 196 + ], + "spans": [ + { + "bbox": [ + 268, + 182, + 320, + 196 + ], + "type": "text", + "content": "Relative Threshold Minimization" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 200, + 320, + 222 + ], + "lines": [ + { + "bbox": [ + 268, + 200, + 320, + 222 + ], + "spans": [ + { + "bbox": [ + 268, + 200, + 320, + 222 + ], + "type": "text", + "content": "different " + }, + { + "bbox": [ + 268, + 200, + 320, + 222 + ], + "type": "inline_equation", + "content": "n_j" + }, + { + "bbox": [ + 268, + 200, + 320, + 222 + ], + "type": "text", + "content": " gradient magnitude based on relative distance" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 336, + 173, + 373, + 179 + ], + "lines": [ + { + "bbox": [ + 336, + 173, + 373, + 179 + ], + "spans": [ + { + "bbox": [ + 336, + 173, + 373, + 179 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 336, + 173, + 373, + 179 + ], + "type": "text", + "content": " Emphasized" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 179, + 380, + 228 + ], + "lines": [ + { + "bbox": [ + 331, + 179, + 380, + 228 + ], + "spans": [ + { + "bbox": [ + 331, + 179, + 380, + 228 + ], + "type": "image", + "image_path": "28ca7ef5ed26341493e44c4f99627cdc714a0fb9ad7e2a21015cc390d5fb56e3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": "Fig. 4: Visualization of the gradient of different loss functions. The colors represent gradient magnitude and the arrows represent gradient directions. a) SoftMax loss pushes the sample toward the class center and away from the other classes, with gradual shifts in gradient magnitude. b) On the other hand, triplet loss has a hard cut gradient based on triplet samples. c) For " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "S_{\\tau}^{det}" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": ", as shown in the top row, the thresholds are determined by non-mated sample, " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "n_j" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": "'s gradient has the greatest magnitude when it has a similar distance from the gallery " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "n_j" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": ". For Relative Threshold Minimization, as shown in the bottom row, as " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": " moves away from the gallery, its gradient decreases. While " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": " remains at the same location, its gradient increases because it becomes closer to " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 132, + 238, + 481, + 358 + ], + "type": "text", + "content": ". The gradients w.r.t. genuine scores adapt to non-mated scores and the gradients w.r.t. non-mated scores are adaptive to other non-mated scores." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 381, + 122, + 429, + 171 + ], + "blocks": [ + { + "bbox": [ + 388, + 116, + 423, + 122 + ], + "lines": [ + { + "bbox": [ + 388, + 116, + 423, + 122 + ], + "spans": [ + { + "bbox": [ + 388, + 116, + 423, + 122 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 388, + 116, + 423, + 122 + ], + "type": "text", + "content": " Emphasized" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 122, + 429, + 171 + ], + "lines": [ + { + "bbox": [ + 381, + 122, + 429, + 171 + ], + "spans": [ + { + "bbox": [ + 381, + 122, + 429, + 171 + ], + "type": "image", + "image_path": "3667c27e033dc36e5f678e648abe413ccbb9658585d27bde538392110b58c8a0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 381, + 179, + 429, + 228 + ], + "blocks": [ + { + "bbox": [ + 383, + 173, + 425, + 179 + ], + "lines": [ + { + "bbox": [ + 383, + 173, + 425, + 179 + ], + "spans": [ + { + "bbox": [ + 383, + 173, + 425, + 179 + ], + "type": "inline_equation", + "content": "n_1,n_2" + }, + { + "bbox": [ + 383, + 173, + 425, + 179 + ], + "type": "text", + "content": " Emphasized" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 179, + 429, + 228 + ], + "lines": [ + { + "bbox": [ + 381, + 179, + 429, + 228 + ], + "spans": [ + { + "bbox": [ + 381, + 179, + 429, + 228 + ], + "type": "image", + "image_path": "29ddcc80faf53ffbfafc1a5378346db052fd4bc681ce2845a5690b34d8e805fe.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 431, + 123, + 479, + 171 + ], + "blocks": [ + { + "bbox": [ + 432, + 116, + 478, + 122 + ], + "lines": [ + { + "bbox": [ + 432, + 116, + 478, + 122 + ], + "spans": [ + { + "bbox": [ + 432, + 116, + 478, + 122 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 432, + 116, + 478, + 122 + ], + "type": "text", + "content": " Less Emphasized" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 123, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 431, + 123, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 431, + 123, + 479, + 171 + ], + "type": "image", + "image_path": "7d75486625275e10b438d32d877414329e2dda47c910815db9dc4e69a12346b4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 431, + 179, + 479, + 228 + ], + "blocks": [ + { + "bbox": [ + 436, + 173, + 472, + 179 + ], + "lines": [ + { + "bbox": [ + 436, + 173, + 472, + 179 + ], + "spans": [ + { + "bbox": [ + 436, + 173, + 472, + 179 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 436, + 173, + 472, + 179 + ], + "type": "text", + "content": " Emphasized" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 179, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 431, + 179, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 431, + 179, + 479, + 228 + ], + "type": "image", + "image_path": "bed9bd62c7db3fceb67f87596eebbd49389e55ec4dda998918f834a5e5221c12.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 382, + 233, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 233, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 233, + 394 + ], + "type": "text", + "content": "3.3 Our Approach" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "text", + "content": "To design loss functions tailored for open-set biometrics, we create a training scenario mirroring the testing environment. During training, we partition each batch into gallery and probe sets to simulate open-set evaluation conditions. In this process, we randomly designate a certain percentage, denoted as " + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "text", + "content": ", of the subjects as mated subjects, while the rest are categorized as non-mated. For the mated subjects, we further randomly allocate their exemplars to the gallery and probe sets. Conversely, all exemplars of the non-mated subjects are placed in the probe set. This approach yields a more realistic and representative training scenario, preparing the biometric features to excel under open-set conditions. To avoid any potential notation ambiguity, we refer to the mated probe set created within each training batch as " + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{K}}^{\\prime}" + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "text", + "content": ", the non-mated probe set as " + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{U}}^{\\prime}" + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "text", + "content": ", and the gallery sets within each training batch as " + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\prime}" + }, + { + "bbox": [ + 130, + 401, + 482, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 544, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 579 + ], + "type": "text", + "content": "To ensure robustness, when the metric space inhabited by the biometric features is defined under the Euclidean distance, we calculate the similarity scores between two features as:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 252, + 585, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 585, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 252, + 585, + 481, + 611 + ], + "type": "interline_equation", + "content": "\\mathrm {s} \\left(p _ {i}, g _ {j}\\right) = \\frac {1}{1 + \\mathrm {d} \\left(p _ {i} , g _ {j}\\right)}, \\tag {6}", + "image_path": "f5cde5b9de09890b808b5a9d59df58b702d39bea5073ba1b505379c960837860.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 617, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 480, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{d}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 617, + 480, + 641 + ], + "type": "text", + "content": " represents the Euclidean distance. Alternatively, if the metric space is defined under the cosine distance, we simply adopt the cosine similarity." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "To ensure that the model can generalize to open-set scenarios during testing, we propose loss functions that penalize the model for three types of errors during" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "training: (1) failing to detect a mated probe with a threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", (2) failing to identify a mated probe within the top rank-" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " positions and (3) setting the threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " too high, leading to poor differentiation between mated and non-mated sample distributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "text", + "content": "(1) Detection. Detection is assessing whether the computed similarity score exceeds a predefined threshold " + }, + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "text", + "content": ". That is, for a given probe " + }, + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathcal{P}_{\\mathrm{K}}'" + }, + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "text", + "content": " and its corresponding gallery subject " + }, + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "inline_equation", + "content": "g_i \\in \\mathcal{G}'" + }, + { + "bbox": [ + 130, + 180, + 482, + 216 + ], + "type": "text", + "content": ", we calculate a detection score as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 223, + 481, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 223, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 239, + 223, + 481, + 237 + ], + "type": "interline_equation", + "content": "S _ {\\tau} ^ {d e t} \\left(p _ {i}, g _ {i}\\right) = \\sigma_ {\\alpha} \\left(\\mathrm {s} \\left(p _ {i}, g _ {i}\\right) - \\tau\\right), \\tag {7}", + "image_path": "fc09a210b0f38e0f5bf378b4a0612da5d254e153d92b35bafa5d6dd56220f824.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\sigma_{\\alpha}(x) = 1 / (1 + \\exp (-\\alpha x))" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " is the Sigmoid function with temperature hyperparameter " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "S_{\\tau}^{det}(p_i, g_i)" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " score quantifies the model's ability to detect if " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " belong to the same subject, given the threshold " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": ". A high similarity score, well above " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": ", indicates successful detection and " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "S_{\\tau}^{det}(p_i, g_i)" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " approaches 1. In contrast, if the model fails to detect that " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " belong to the same subject, " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "s(p_i, g_i)" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " remains much smaller than " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": ", causing " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "S_{\\tau}^{det}(p_i, g_i)" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " to approach 0. Since the Sigmoid function has the highest gradient near 0, the term " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{s}(p_i, g_i) - \\tau" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": " makes the loss function focus more on the samples near the threshold " + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 243, + 482, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 340, + 482, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 375 + ], + "type": "text", + "content": "During testing, the open-set detection threshold is determined by the non-mated score distribution, which we do not know during training. Thus, we sample " + }, + { + "bbox": [ + 130, + 340, + 482, + 375 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 340, + 482, + 375 + ], + "type": "text", + "content": " from the non-mated scores in a batch. Specifically, we let" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 382, + 481, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 382, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 249, + 382, + 481, + 410 + ], + "type": "interline_equation", + "content": "S ^ {d e t} = \\frac {1}{| \\mathcal {T} |} \\sum_ {\\tau \\in \\mathcal {T}} S _ {\\tau} ^ {d e t} \\left(p _ {i}, g _ {i}\\right), \\tag {8}", + "image_path": "714ed92b30d850ea654ebe38d0e6f82ea1e2eb4023b0bd4ac5926a175214a89f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{s(n_j,g_i)|n_j\\in \\mathcal{P}_{\\mathrm{U}}'\\}" + }, + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "text", + "content": " is the set of all non-mated scores. Fig. 4(c) illustrates the gradient of Det w.r.t. mated probes under different thresholds induced by various non-mated probes. The gradient is adaptable to different thresholds, increasing as " + }, + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "text", + "content": " decreases, even when " + }, + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "inline_equation", + "content": "s(g,p)" + }, + { + "bbox": [ + 130, + 416, + 482, + 513 + ], + "type": "text", + "content": " is constant. This adaptability is desirable in open-set settings, as it promotes mated probes to be more similar to their ground-truth gallery than non-mated probes. Note that the thresholds in Eq. (8) are only used during training. The detection thresholds during testing are automatically determined by the FNIR@FPIR metric as discussed in Sec. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 529, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 577 + ], + "type": "text", + "content": "(2) Identification. It is not sufficient for a biometric model to distinguish mated subjects from non-mated ones based on their scores. Equally important is the model's ability to accurately identify the correct subject among the gallery subjects. Therefore, we introduce an identification score " + }, + { + "bbox": [ + 130, + 529, + 482, + 577 + ], + "type": "inline_equation", + "content": "S^{id}(p_i, g_i)" + }, + { + "bbox": [ + 130, + 529, + 482, + 577 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 584, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 584, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 228, + 584, + 481, + 597 + ], + "type": "interline_equation", + "content": "S ^ {i d} \\left(p _ {i}, g _ {i}\\right) = \\sigma_ {\\beta} \\left(1 - \\operatorname {s o f t r a n k} \\left(p _ {i}, g _ {i}\\right)\\right), \\tag {9}", + "image_path": "ddf77c7e245470c9637cff4d3cff29d657a3c51d533fe7a5b517f180bf5e7c04.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\beta}" + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": " is a Sigmoid function with temperature " + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": " and sofrank is a differentiable function with a Sigmoid function " + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_{\\gamma}" + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 205, + 637, + 481, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 637, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 205, + 637, + 481, + 662 + ], + "type": "interline_equation", + "content": "\\operatorname {s o f t r a n k} \\left(p _ {i}, g _ {i}\\right) = \\sum_ {g _ {j} \\in \\mathcal {G} ^ {\\prime}} \\sigma_ {\\gamma} \\left(\\mathrm {s} \\left(p _ {i}, g _ {j}\\right) - \\mathrm {s} \\left(p _ {i}, g _ {i}\\right)\\right), \\tag {10}", + "image_path": "0060a189797134884746d21ce4c8b5afc5a205b5e04612bf6048c487c4abb1f1.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "The softrank function accumulates values close to 1 for gallery subjects " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "g_{j}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " more similar to the probe " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " than the correct match " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " and close to 0 otherwise. It essentially reflects the rank of the correct match " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " among all gallery subjects for probe " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "S^{id}(p_i, g_i)" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " is an indicator, approaching 1 if " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " has a softrank value less than 1, indicating successful identification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "content": "To optimize open-set detection and identification simultaneously, the overall identification-detection loss " + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{r,\\tau}" + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "content": " is formulated as the product of two components," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 214, + 480, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 214, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 212, + 214, + 480, + 243 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {I D L}} = - \\frac {1}{| \\mathcal {P} _ {\\mathrm {K}} ^ {\\prime} |} \\sum_ {p _ {i} \\in \\mathcal {P} _ {\\mathrm {K}} ^ {\\prime}} S ^ {d e t} \\left(p _ {i}, g _ {i}\\right) \\cdot S ^ {i d} \\left(p _ {i}, g _ {i}\\right). \\tag {11}", + "image_path": "edf274985a76d9bdd2a4df8ce8cbbd0b157fc588bdf4de434561aebfa01500a7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": "In other words, " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{IDL}}" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " is only close to " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " if both " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "S^{id}(p_i,g_i)" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "S^{det}(p_i,g_i)" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " approach 1. Conversely, should either or both not approximate 1, " + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{IDL}}" + }, + { + "bbox": [ + 130, + 248, + 479, + 284 + ], + "type": "text", + "content": " converges to 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 299, + 479, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 479, + 358 + ], + "type": "text", + "content": "(3) Relative Threshold Minimization " + }, + { + "bbox": [ + 130, + 299, + 479, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{IDL}}" + }, + { + "bbox": [ + 130, + 299, + 479, + 358 + ], + "type": "text", + "content": " optimizes detection and identification under fixed thresholds. We also introduce a strategy to minimize the maximum scores for non-mated probes, discouraging the model from assigning high scores to non-mated probes. This helps mitigate the risk of false positive detections in open-set scenarios, enhancing overall robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "text", + "content": "To achieve this, we minimize the weighted average of the similarities between the probe and each gallery subject, where the weights are obtained by applying the softmax function to each score. Mathematically, given a probe sample " + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "text", + "content": " and a set of gallery subjects " + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{G}' = \\{g_1, g_2, \\dots, g_n\\}" + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "text", + "content": " with corresponding similarity scores " + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "inline_equation", + "content": "s_1, s_2, \\dots, s_n" + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "text", + "content": ", we minimize the weighted average as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 424, + 480, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 424, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 242, + 424, + 480, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {R T M}} = \\frac {1}{\\sum_ {j = 1} ^ {n} e ^ {s _ {j}}} \\sum_ {j = 1} ^ {n} e ^ {s _ {j}} \\cdot s _ {j}, \\tag {12}", + "image_path": "a394b88ebb6cb8b2d2dd4ac2cf69d3022f1b89161c381989dbd6f6ab97f931d4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 459, + 375, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 375, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 375, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 459, + 375, + 470 + ], + "type": "inline_equation", + "content": "e^{s_j}" + }, + { + "bbox": [ + 130, + 459, + 375, + 470 + ], + "type": "text", + "content": " is the softmax value for each similarity score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 472, + 479, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 479, + 553 + ], + "type": "text", + "content": "Fig. 4 (c) shows gradients of " + }, + { + "bbox": [ + 130, + 472, + 479, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RTM}}" + }, + { + "bbox": [ + 130, + 472, + 479, + 553 + ], + "type": "text", + "content": " w.r.t. non-mated probes. This gradient adapts to the relative position of non-mated probes, assigning higher gradients to samples closer to the gallery. Unlike softmax-based losses and the triplet loss, RTM consistently emphasizes imposter scores that rank higher compared to others within the batch to learn better feature representations for open-set biometrics. This property is desirable as it effectively minimizes the maximum score of each non-mated probe." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 555, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 479, + 602 + ], + "type": "text", + "content": "The max function is an alternative to the weighted average. Yet, simply minimizing the maximum score in training would not minimize other scores even if they are close to the maximum. Hence, the weighted average would make it easier for the trained model to generalize during testing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 616, + 222, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 616, + 222, + 627 + ], + "spans": [ + { + "bbox": [ + 131, + 616, + 222, + 627 + ], + "type": "text", + "content": "3.4 Overall Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 632, + 451, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 632, + 451, + 644 + ], + "spans": [ + { + "bbox": [ + 131, + 632, + 451, + 644 + ], + "type": "text", + "content": "The overall objective is to minimize the combination of " + }, + { + "bbox": [ + 131, + 632, + 451, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{IDL}}" + }, + { + "bbox": [ + 131, + 632, + 451, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 632, + 451, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{RTM}}" + }, + { + "bbox": [ + 131, + 632, + 451, + 644 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 260, + 649, + 479, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 649, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 260, + 649, + 479, + 662 + ], + "type": "interline_equation", + "content": "L = \\mathcal {L} ^ {\\mathrm {I D L}} + \\lambda \\cdot \\mathcal {L} ^ {\\mathrm {R T M}}. \\tag {13}", + "image_path": "c3ecc45b57d19240fff233040b3b043fd0e772b13640549db03d045fbdf245f0.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is the loss weight. Compared to existing loss functions, our loss function treats genuine and imposter scores differently. It aims to reduce open-set thresholds by pushing non-mated probes away from gallery while pulling genuine scores toward their corresponding galleries, similar to open-set evaluation. Moreover, the relative magnitude of non-mated scores is utilized by RTM to assign larger gradients to harder non-mated probes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 230, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 230, + 219 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 230, + 219 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 229, + 262, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 262, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 262, + 243 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 356 + ], + "type": "text", + "content": "Evaluation Protocols. In face recognition experiments, we use the official closed- and open-set evaluation protocols of IJB-S [28]. However, gait recognition and person reID datasets typically lack official open-set evaluation protocols, where some probe subjects are excluded from the gallery. Naïvely removing gallery subjects often results in highly unstable outcomes. This instability arises because threshold calculation heavily relies on a small fraction (1%) of non-mated probes due to the limited number of probes in these datasets. For instance, in a dataset with 500 non-mated probes, the threshold is computed from the highest scores of only 5 probes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "text", + "content": "To ensure robust open-set evaluation, we create " + }, + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "inline_equation", + "content": "N (= 50)" + }, + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "text", + "content": " random sets of non-mated probes and compute FNIR@1%FPIR for each set. In each set, " + }, + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "inline_equation", + "content": "q\\% = 21.5\\%" + }, + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "text", + "content": " subjects are designated as non-mated. We report the median value and the standard deviation from " + }, + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 357, + 482, + 417 + ], + "type": "text", + "content": " trials. The choice of median over mean is deliberate to avoid undue influence from extreme values and provide a more stable metric." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate in both open- and closed-set settings. In closed-set settings, we use the Cumulative Matching Curve (CMC) at rank-1; for open-set, we use FNIR@FPIR at " + }, + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "text", + "content": " FPIR and rank " + }, + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "inline_equation", + "content": "R = 20" + }, + { + "bbox": [ + 131, + 433, + 282, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 533, + 282, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 282, + 617 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 282, + 617 + ], + "type": "text", + "content": "Datasets. For face recognition, we train on WebFace4M [80] and test on IJB-S [28]. Our experiments in gait recognition involve the Gait3D [77] and GREW [79] datasets, while for person reID, we employ the MEVID [8] and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 287, + 445, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 445, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 287, + 445, + 482, + 523 + ], + "type": "text", + "content": "Table 1: Summary of datasets used in our experiments. [Keys: Mod.=Modality; Img.=Open-set image recognition; Subj.=the number of subjects ; Media=the number of images for face/image recognition datasets and the number of sequences/tracklets for gait and video-based person reID datasets]" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 290, + 524, + 479, + 608 + ], + "blocks": [ + { + "bbox": [ + 290, + 524, + 479, + 608 + ], + "lines": [ + { + "bbox": [ + 290, + 524, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 290, + 524, + 479, + 608 + ], + "type": "table", + "html": "
DatasetYear Mod.TrainTest
Subj.MediaSubj.Media
IJB-S [28]2018Face-202>10M
Gait3D [77]2022Gait3,00018,9406,369
GREW [79]2021Gait20,000102,88724,000
MEVID [8]2023ReID1046,33855
CCVID [17]2022ReID75948151
", + "image_path": "52784991c13ebb13c4b71009d3ce5c01e52099f124cb2d2eb9bf0ccaec739f99.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 399, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 399, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 399, + 629 + ], + "type": "text", + "content": "CCVID [17] datasets. Further details are provided in Tab. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "One technical caveat is that although FNIR@1%FPIR implicitly assumes that each subject only appears once in the gallery, a typical gait recognition or person re-id dataset contains more than one gallery template per subject. To" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "address this issue, we take the mean feature over all gallery templates for each subject. This benefits the models compared to randomly sampling one gallery template (details are in the Supp.)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 170, + 491, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 491, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 491, + 207 + ], + "type": "text", + "content": "Implementation Details. We implement our proposed approach in PyTorch [49]. For fair comparisons, we follow the experiment setup of our baselines. For details on the hyperparameters, please refer to the Supp." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 224, + 246, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 246, + 235 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 246, + 235 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 281 + ], + "type": "text", + "content": "We conduct our ablation studies on the Gait3D [77] dataset using GaitBase [13] as the baseline. In the ablation studies, we hold out " + }, + { + "bbox": [ + 130, + 243, + 481, + 281 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 243, + 481, + 281 + ], + "type": "text", + "content": " of subjects in the training set as the validation set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 298, + 280, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 280, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 280, + 309 + ], + "type": "text", + "content": "Effect of Det, Id and RTM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 310, + 282, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 282, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 282, + 441 + ], + "type": "text", + "content": "We ablate the effect of Det, Id and RTM in Tab. 2. RTM consistently enhances both closed- and open-set performance. Id further improves open-set performance but has a minor impact on rank-1 accuracy. Meanwhile, Det significantly enhances both closed- and open-set performance. In summary, each component contributes to its effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 460, + 280, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 280, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 280, + 472 + ], + "type": "text", + "content": "Effect of Hyperparameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 472, + 281, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 281, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 281, + 544 + ], + "type": "text", + "content": "We ablate the effect of hyperparameters " + }, + { + "bbox": [ + 130, + 472, + 281, + 544 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma," + }, + { + "bbox": [ + 130, + 472, + 281, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 281, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 472, + 281, + 544 + ], + "type": "text", + "content": " in Tab. 2. approach can consistently improve line, GaitBase [13]. Also, while the hyperparameters, our approach with the best set of tested hyperp" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 288, + 331, + 480, + 471 + ], + "blocks": [ + { + "bbox": [ + 287, + 308, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 287, + 308, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 287, + 308, + 482, + 331 + ], + "type": "text", + "content": "Table 2: Ablation study results. [Keys: " + }, + { + "bbox": [ + 287, + 308, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\checkmark =" + }, + { + "bbox": [ + 287, + 308, + 482, + 331 + ], + "type": "text", + "content": " Kept; " + }, + { + "bbox": [ + 287, + 308, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\checkmark =" + }, + { + "bbox": [ + 287, + 308, + 482, + 331 + ], + "type": "text", + "content": " Removed; R@1=rank-1 accuracy]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 288, + 331, + 480, + 471 + ], + "lines": [ + { + "bbox": [ + 288, + 331, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 288, + 331, + 480, + 471 + ], + "type": "table", + "html": "
IdDetRTMα,γβλR@1FNIR@1%FPIR
XXX———64.686.43±5.44
XX✓6.00.24.065.185.81±5.78
X✓✓64.485.16±5.13
✓✓✓3.00.24.063.985.22±4.73
9.064.284.08±6.28
✓✓✓6.00.14.063.185.10±5.67
0.463.885.92±5.45
✓✓✓6.00.22.064.883.12±5.94
8.061.685.67±5.15
✓✓✓6.00.24.064.882.99±4.50
", + "image_path": "43a3dabde20b093603ab03588afeadecabf303b03214e392839aaf4cef512bab.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 561, + 332, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 561, + 332, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 561, + 332, + 574 + ], + "type": "text", + "content": "4.3 Comparison with State of the Art" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Across biometric modalities, existing methods exhibit competitive closed-set recognition performance, as shown in Tab. 3 for face, Tab. 4 for gait, and Tab. 5 for person reID. In particular, in gait recognition, GaitBase [13] and SwinGait [12] showcase strong closed-set performance with a notable rank-1 accuracies of greater than " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", respectively, for in-the-wild gait datasets Gait3D [77] and GREW [79]. However, in open-set scenarios, they face challenges, as evidenced by GaitBase's high FNIR of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "86.43\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " FPIR in Gait3D." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 90, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 204, + 146, + 411, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 3: Closed- and open-set face recognition on two official protocols of IJB-S [28]. [Keys: R@1=rank-1 accuracy; 0.1%=FNIR@0.1%FPIR; 1%: FNIR@1%FPIR]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 146, + 411, + 198 + ], + "lines": [ + { + "bbox": [ + 204, + 146, + 411, + 198 + ], + "spans": [ + { + "bbox": [ + 204, + 146, + 411, + 198 + ], + "type": "table", + "html": "
ModelSurveillance-SingleSurveillance-Booking
R@1↑0.1%↓1%↓R@1↑0.1%↓1%↓
AdaFace [29]68.4445.1838.3768.8445.9837.80
AdaFace+Ours69.9244.3036.7370.0043.4836.17
", + "image_path": "fd8ff5b58ab5bcb981f700956a7345d17163847089f59cd0b354b6d03920e3ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "text", + "content": "It is evident that the inclusion of our loss functions consistently yields improvements in both closed- and open-set performance across three modalities. In face recognition, we boost the rank-1 accuracy of AdaFace [29] to " + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "inline_equation", + "content": "69.92\\%" + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "inline_equation", + "content": "(+1.48\\%)" + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "text", + "content": " in Surveillance-to-Single. In the challenging open-set scenario, our losses further demonstrate their effectiveness by reducing FNIR@0.1%FPIR from " + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "inline_equation", + "content": "45.98\\%" + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "inline_equation", + "content": "43.48\\%" + }, + { + "bbox": [ + 130, + 221, + 480, + 293 + ], + "type": "text", + "content": " in Surveillance-Booking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "text", + "content": "In gait recognition, when applied to Swin-Gait [12] backbone, our losses improve the rank-1 accuracy to " + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "75.3\\%" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "(+2.9\\%)" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "text", + "content": " on Gait3D, and FNIR@1%FPIR decreases to " + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "79.30\\%" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "(-4.20\\%)" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "text", + "content": " on Gait3D. In person reID, integrating our loss with CAL [17] reaches a rank-1 accuracy of " + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "84.8\\%" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "(+2.2\\%)" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "text", + "content": " and a FNIR of " + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "30.81\\%" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "inline_equation", + "content": "(-3.08\\%)" + }, + { + "bbox": [ + 130, + 295, + 264, + 484 + ], + "type": "text", + "content": " on MEVID [8]. We conduct the independent two-sample t-test to assess the statistical significance of perfor" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 273, + 363, + 480, + 472 + ], + "blocks": [ + { + "bbox": [ + 270, + 308, + 482, + 362 + ], + "lines": [ + { + "bbox": [ + 270, + 308, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 270, + 308, + 482, + 362 + ], + "type": "text", + "content": "Table 4: Closed- and open-set gait recognition on Gait3D [77] and GREW [79]. SMPLGait [77] takes SMPL parameters as additional input. SwinGait-3D [12] uses our own implementation. [Keys: Rank@1=rank-1 accuracy]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 273, + 363, + 480, + 472 + ], + "lines": [ + { + "bbox": [ + 273, + 363, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 273, + 363, + 480, + 472 + ], + "type": "table", + "html": "
ModelRank@1 ↑FNIR@1%FPIR ↓
Gait3DGREWGait3DGREW
GaitSet [4]36.748.496.05±2.3081.17±1.45
GaitPart [14]28.247.696.88±1.8088.43±1.26
GaitGL [38]29.741.594.97±1.2480.55±1.11
SMPLGait [77]46.3—93.18±2.58—
GaitBase [13]64.660.186.43±5.4477.67±1.60
+Ours64.860.182.99±4.5076.55±1.54
SwinGait-3D [12]72.479.383.50±6.2058.32±1.79
+Ours75.380.479.30±6.7457.29±1.83
", + "image_path": "ac624f147c1e1c0c087629b8ea5d85ea9b34196a7b1718fd9ad266ce5613aaf6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 487, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 479, + 510 + ], + "type": "text", + "content": "mance improvements in open-set scenarios. For instance, our approach outperforms CAL on the CCVID [17] dataset, with a p-value of " + }, + { + "bbox": [ + 130, + 487, + 479, + 510 + ], + "type": "inline_equation", + "content": "0.052\\%" + }, + { + "bbox": [ + 130, + 487, + 479, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 511, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 480, + 559 + ], + "type": "text", + "content": "Collectively, these findings emphasize the consistent effectiveness of our proposed loss functions in enhancing both closed- and open-set performance across a spectrum of biometric tasks, ultimately improving the models' ability to handle open-set challenges, while maintaining strong closed-set performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "Efficiency. We present the floating point operations (FLOPs) of training baselines with and without our loss functions in Tab. 6. It is expected that incorporating our loss into baselines will increase the computational load in training, but the increase is very minor. For example, when SwinGait [12] is used, our loss functions only incur " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "46.9\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": " of the FLOPs of the triplet loss, which is negligible compared to the model. Moreover, our approach does not incur any computation overhead during testing while improving closed- and open-set performance." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 135, + 421, + 224 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 126 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 126 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 126 + ], + "type": "text", + "content": "Table 5: Closed- and open-set person reID on MEVID [8] and CCVID [17]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 135, + 421, + 224 + ], + "lines": [ + { + "bbox": [ + 192, + 135, + 421, + 224 + ], + "spans": [ + { + "bbox": [ + 192, + 135, + 421, + 224 + ], + "type": "table", + "html": "
ModelDatasetRank-1Accuracy ↑FNIR@1%FPIR ↓
AGRL [66]MEVID [8]48.466.67±9.57
TCLNet [21]MEVID [8]48.170.34±11.68
PSTA [63]MEVID [8]48.277.11±11.72
CAL [17]MEVID [8]52.566.95±11.59
+OursMEVID [8]55.164.80±10.71
CAL [17]CCVID [17]82.633.89±7.65
+OursCCVID [17]84.830.81±8.02
", + "image_path": "b4d4641ac96c79efb2ba14c16aac6284e943a1b55b15f80b9d09393519f81513.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 213, + 281, + 401, + 339 + ], + "blocks": [ + { + "bbox": [ + 131, + 238, + 481, + 273 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 481, + 273 + ], + "type": "text", + "content": "Table 6: FLOPs in training different baseline models with and without our approach. We report the mean FLOPs across a minibatch. Our approach incurs no additional FLOP during testing." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 213, + 281, + 401, + 339 + ], + "lines": [ + { + "bbox": [ + 213, + 281, + 401, + 339 + ], + "spans": [ + { + "bbox": [ + 213, + 281, + 401, + 339 + ], + "type": "table", + "html": "
ModalityModelFLOPS+Triplet+Ours
FaceAdaFace [29]24.19G—503K
GaitGaitBase [13]35.43G524.3K245.8K
GaitSwinGait [12]120G491.5K230.4K
BodyCAL [17]32.75G4.09K15.38K
", + "image_path": "686ba18ec30bace317a506ac35b29f802f6d35d02ef4fff028e3eb3901ced444.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 366, + 230, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 230, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 230, + 377 + ], + "type": "text", + "content": "4.4 Visualizations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 400, + 265, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 265, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 265, + 639 + ], + "type": "text", + "content": "Score Distributions. Fig. 5 visualizes the non-mated, maximum non-mated score for each probe, and genuine scores of SwinGait [12], with and without our loss. Our introduced RTM effectively reduces the distribution of maximum-per-probe scores, shifting it towards lower values. After linearly normalizing all scores between 0 and 1, the threshold at " + }, + { + "bbox": [ + 130, + 400, + 265, + 639 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 400, + 265, + 639 + ], + "type": "text", + "content": " FPIR is 0.61 with our loss (Fig. 5 (b)) compared to 0.69 without ours (Fig. 5 (a)). Moreover, our loss does not significantly alter the rightmost portion of the genuine score distribution, enhancing detec" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 639, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 480, + 664 + ], + "type": "text", + "content": "tion. This analysis offers insights into how our loss changes score distributions, and through which improves open-set performance." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 272, + 426, + 479, + 540 + ], + "blocks": [ + { + "bbox": [ + 272, + 426, + 479, + 540 + ], + "lines": [ + { + "bbox": [ + 272, + 426, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 272, + 426, + 479, + 540 + ], + "type": "image", + "image_path": "88bea38f4110d2aaa25187068b72f57ff1d1f84bae263255a5d3755e50b27da5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 550, + 482, + 628 + ], + "lines": [ + { + "bbox": [ + 269, + 550, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 269, + 550, + 482, + 628 + ], + "type": "text", + "content": "Fig.5: Score distributions on the test set of Gait3D [77] dataset. Ours (b) effectively reduces the threshold at " + }, + { + "bbox": [ + 269, + 550, + 482, + 628 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 269, + 550, + 482, + 628 + ], + "type": "text", + "content": " FPIR while only shifting the genuine score distribution marginally to the left compared to SwinGait [12] (a). Both distributions are shifted and scaled horizontally so that the min/max scores are within " + }, + { + "bbox": [ + 269, + 550, + 482, + 628 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 269, + 550, + 482, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 242, + 183 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 242, + 183 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 242, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 242, + 183 + ], + "type": "image", + "image_path": "e40c633b2b5acffa1a07a4b2234ff56dc6b8a0048e7f6c1dcba3742e20b58b85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 186, + 205, + 196 + ], + "lines": [ + { + "bbox": [ + 174, + 186, + 205, + 196 + ], + "spans": [ + { + "bbox": [ + 174, + 186, + 205, + 196 + ], + "type": "text", + "content": "(a) Face" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 117, + 359, + 183 + ], + "blocks": [ + { + "bbox": [ + 250, + 117, + 359, + 183 + ], + "lines": [ + { + "bbox": [ + 250, + 117, + 359, + 183 + ], + "spans": [ + { + "bbox": [ + 250, + 117, + 359, + 183 + ], + "type": "image", + "image_path": "139e2e548bdfcefcdf8f42006380bd9c91aa26073c4d9770523905ddb18e25cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 186, + 323, + 195 + ], + "lines": [ + { + "bbox": [ + 291, + 186, + 323, + 195 + ], + "spans": [ + { + "bbox": [ + 291, + 186, + 323, + 195 + ], + "type": "text", + "content": "(b) Gait" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 204, + 482, + 250 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 250 + ], + "type": "text", + "content": "Fig. 6: Open-set FNs due to detection, identification, and either. The average across the 50 open-set evaluation protocols is plotted. Blue indicates FNs that are shared by the baseline, SwinGait [12], and our approach. Orange indicates FNs of the baseline only. Green indicates FNs of our approach." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 118, + 477, + 183 + ], + "blocks": [ + { + "bbox": [ + 369, + 118, + 477, + 183 + ], + "lines": [ + { + "bbox": [ + 369, + 118, + 477, + 183 + ], + "spans": [ + { + "bbox": [ + 369, + 118, + 477, + 183 + ], + "type": "image", + "image_path": "3bb64effdd26f3bdd71d47e99941ed7a6c51b0e2f8ec97ec859cd1522e65ba8c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 186, + 454, + 196 + ], + "lines": [ + { + "bbox": [ + 395, + 186, + 454, + 196 + ], + "spans": [ + { + "bbox": [ + 395, + 186, + 454, + 196 + ], + "type": "text", + "content": "(c) Person ReID" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 273, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 369 + ], + "type": "text", + "content": "False Negatives Breakdown. As illustrated in Fig. 6, across different biometric and generic tasks, almost all of the false negatives are failed detection, e.g. the " + }, + { + "bbox": [ + 130, + 273, + 482, + 369 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 273, + 482, + 369 + ], + "type": "text", + "content": " mated probe in Fig. 1. A small portion of false negatives have both failed identification and detection, e.g. the " + }, + { + "bbox": [ + 130, + 273, + 482, + 369 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 130, + 273, + 482, + 369 + ], + "type": "text", + "content": " mated probe in Fig. 1. When we apply our loss functions, the number of FNs due to detection is reduced and the number of FNs due to identification is comparable to that of the baseline. These results highlight that our approach can improve open-set detection without sacrificing identification in various settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 387, + 225, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 225, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 225, + 400 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 510 + ], + "type": "text", + "content": "Our work addresses the critical gap in open-set biometrics, a practical generalization of the well-studied closed-set biometrics. We introduce novel I-DL and RTM losses aiming to improve biometric models' open-set performance. Through extensive experiments across diverse modalities, including face, gait, and person reID, we demonstrate the efficacy of our approach. We not only advance the state of the art in the challenging open-set domain but also enhance closed-set recognition capabilities. We advocate for increased attention to the open-set problem within the biometric community." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 527, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 588 + ], + "type": "text", + "content": "Limitations While our paper makes significant strides in addressing open-set challenges in biometrics through loss functions, our exploration does not extensively explore dedicated architectures for this specific context. Future research endeavors may benefit from investigating tailored architectures that complement and enhance the effectiveness of our proposed loss functions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Potential Societal Impacts. The use of biometric datasets in our work containing personally identifiable images prompts a crucial discussion on privacy and ethical implications. As biometric technologies become more prevalent, it is imperative to ensure that the collection of such datasets is aligned with ethical standards and privacy regulations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Acknowledgements. This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via 2022-21102100004. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 231, + 197, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 231, + 197, + 243 + ], + "spans": [ + { + "bbox": [ + 133, + 231, + 197, + 243 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 256, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 256, + 462, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 462, + 268 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 462, + 268 + ], + "type": "text", + "content": "1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: CVPR (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "text", + "content": "2. Cen, J., Luan, D., Zhang, S., Pei, Y., Zhang, Y., Zhao, D., Shen, S., Chen, Q.: The devil is in the wrongly-classified samples: Towards unified open-set recognition. In: ICLR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "text", + "content": "3. Chai, T., Li, A., Zhang, S., Li, Z., Wang, Y.: Lagrange motion analysis and view embeddings for improved gait recognition. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 323, + 481, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 345 + ], + "type": "text", + "content": "4. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: AAAI (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "type": "text", + "content": "5. Chen, C., Ye, M., Jiang, D.: Towards modality-agnostic person re-identification with descriptive query. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "type": "text", + "content": "6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE TPAMI (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 411 + ], + "type": "text", + "content": "7. Cui, Y., Kang, Y.: Multi-modal gait recognition via effective spatial-temporal feature fusion. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 481, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 444 + ], + "type": "text", + "content": "8. Davila, D., Du, D., Lewis, B., Funk, C., Van Pelt, J., Collins, R., Corona, K., Brown, M., McCloskey, S., Hoogs, A., et al.: MEVID: multi-view extended videos with identities for video person re-identification. In: WACV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "text", + "content": "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 489 + ], + "type": "text", + "content": "0. Dou, H., Zhang, P., Su, W., Yu, Y., Li, X.: Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In: ECCV (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "type": "text", + "content": "1. Dou, H., Zhang, P., Su, W., Yu, Y., Lin, Y., Li, X.: Gaitgci: Generative counterfactual intervention for gait recognition. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "type": "text", + "content": "2. Fan, C., Hou, S., Huang, Y., Yu, S.: Exploring deep models for practical gait recognition. arXiv preprint arXiv:2303.03301 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "text", + "content": "3. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 577 + ], + "type": "text", + "content": "4. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: CVPR (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 599 + ], + "type": "text", + "content": "5. Geng, C., Huang, S., Chen, S.: Recent advances in open set recognition: A survey. IEEE TPAMI (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "text", + "content": "6. Grother, P., Ngan, M., Hanaoka, K.: Face recognition vendor test (FRVT) part 2: Identification (2019), https://www.nist.gov/document/nistir827120190911pdf" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "text", + "content": "7. Gu, X., Chang, H., Ma, B., Bai, S., Shan, S., Chen, X.: Clothes-changing person re-identification with rgb modality only. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "8. Gunther, M., Cruz, S., Rudd, E.M., Boult, T.E.: Toward open-set face recognition. In: CVPR Workshops (2017)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "19. Günther, M., Hu, P., Herrmann, C., Chan, C.H., Jiang, M., Yang, S., Dhamija, A.R., Ramanan, D., Beyerer, J., Kittler, J., et al.: Unconstrained face detection and open-set face recognition challenge. In: IJCB. IEEE (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "20. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "21. Hou, R., Chang, H., Ma, B., Shan, S., Chen, X.: Temporal complementary learning for video person re-identification. In: ECCV (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "text", + "content": "22. Huang, H., Wang, Y., Hu, Q., Cheng, M.M.: Class-specific semantic reconstruction for open set recognition. IEEE TPAMI (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 236 + ], + "type": "text", + "content": "23. Huang, X., Zhu, D., Wang, H., Wang, X., Yang, B., He, B., Liu, W., Feng, B.: Context-sensitive temporal feature learning for gait recognition. In: ICCV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "text", + "content": "24. Huang, Y., Shen, P., Tai, Y., Li, S., Liu, X., Li, J., Huang, F., Ji, R.: Improving face recognition from hard samples via distribution distillation loss. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "25. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "26. Huang, Z., Xue, D., Shen, X., Tian, X., Li, H., Huang, J., Hua, X.S.: 3d local convolutional neural networks for gait recognition. In: ICCV (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "text", + "content": "27. Jain, A.K., Deb, D., Engelsma, J.J.: Biometrics: Trust, but verify. IEEE Transactions on Biometrics, Behavior, and Identity Science (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "text", + "content": "28. Kalka, N.D., Maze, B., Duncan, J.A., O'Connor, K., Elliott, S., Hebert, K., Bryan, J., Jain, A.K.: IJB-S: IARPA janus surveillance video benchmark. In: BTAS (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "29. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "30. Kim, M., Liu, F., Jain, A.K., Liu, X.: Cluster and aggregate: Face recognition with large probe set. In: NeurIPS (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "31. Kim, M., Su, Y., Liu, F., Jain, A., Liu, X.: Keypoint relative position encoding for face recognition. In: CVPR (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "32. Koch, T., Riess, C., Kohler, T.: Lord: Leveraging open-set recognition with unknown data. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "text", + "content": "33. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "text", + "content": "34. Kuchibhotla, H.C., Malagi, S.S., Chandhok, S., Balasubramanian, V.N.: Unseen classes at a later time? no problem. In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "35. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "36. Li, W., Hou, S., Zhang, C., Cao, C., Liu, X., Huang, Y., Zhao, Y.: An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "37. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: ECCV (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "text", + "content": "38. Lin, B., Zhang, S., Yu, X.: Gait recognition via effective global-local feature representation and local temporal aggregation. In: ICCV (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "39. Liu, F., Ashbaugh, R., Chimitt, N., Hassan, N., Hassani, A., Jaiswal, A., Kim, M., Mao, Z., Perry, C., Ren, Z., et al.: Farsight: A physics-driven whole-body biometric system at large distance and altitude. In: WACV (2024)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "40. Liu, F., Kim, M., Gu, Z., Jain, A., Liu, X.: Learning clothing and pose invariant 3d shape representation for long-term person re-identification. In: ICCV (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "41. Liu, F., Kim, M., Jain, A., Liu, X.: Controllable and guided face synthesis for unconstrained face recognition. In: ECCV (2022)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "42. Liu, F., Kim, M., Ren, Z., Liu, X.: Distilling CLIP with dual guidance for learning discriminative human body shape representation. In: CVPR (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "text", + "content": "43. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: CVPR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "44. Ma, K., Fu, Y., Zheng, D., Cao, C., Hu, X., Huang, Y.: Dynamic aggregated network for gait recognition. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "text", + "content": "45. Mekhazni, D., Dufau, M., Desrosiers, C., Pedersoli, M., Granger, E.: Camera alignment and weighted contrastive learning for domain adaptation in video person reid. In: WACV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "text", + "content": "46. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "text", + "content": "47. Minaee, S., Abdelrashidi, A., Su, H., Bennamoun, M., Zhang, D.: Biometrics recognition using deep learning: A survey. Artificial Intelligence Review (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 258, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 278 + ], + "type": "text", + "content": "48. Ni, H., Li, Y., Gao, L., Shen, H.T., Song, J.: Part-aware transformer for generalizable person re-identification. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "type": "text", + "content": "49. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "text", + "content": "50. Qin, L., Wang, M., Deng, C., Wang, K., Chen, X., Hu, J., Deng, W.: Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. TCSVT (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "text", + "content": "51. Ren, Z., Su, Y., Liu, X.: Chatgpt-powered hierarchical comparisons for image classification. In: NeurIPS (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "52. Ross, A., Banerjee, S., Chen, C., Chowdhury, A., Mirjalili, V., Sharma, R., Swearinen, T., Yadav, S.: Some research problems in biometrics: The future beckons. In: ICB (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "53. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: CVPR (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "54. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3d gait recognition with point clouds. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "type": "text", + "content": "55. Shen, C., Yu, S., Wang, J., Huang, G.Q., Wang, L.: A comprehensive survey on deep gait recognition: algorithms, datasets and challenges. arXiv preprint arXiv:2206.13732 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 472, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 492 + ], + "type": "text", + "content": "56. Somers, V., De Vleeschouwer, C., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "text", + "content": "57. Sun, Y., Cheng, C., Zhang, Y., Zhang, C., Zheng, L., Wang, Z., Wei, Y.: Circle loss: A unified perspective of pair similarity optimization. In: CVPR (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "type": "text", + "content": "58. Tabassi, E., Watson, C., Fiumara, G., Salamon, W., Flanagan, P., Cheng, S.L.: Performance evaluation of fingerprint open-set identification algorithms. In: IJCB (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "59. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need. In: ICLR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "60. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: CVPR (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "61. Wang, L., Liu, B., Liang, F., Wang, B.: Hierarchical spatio-temporal representation learning for gait recognition. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "text", + "content": "62. Wang, M., Guo, X., Lin, B., Yang, T., Zhu, Z., Li, L., Zhang, S., Yu, X.: Dygait: Exploiting dynamic representations for high-performance gait recognition. In: ICCV (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "63. Wang, Y., Zhang, P., Gao, S., Geng, X., Lu, H., Wang, D.: Pyramid spatial-temporal aggregation for video-based person re-identification. In: ICCV (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Set Biometrics: Beyond Good Closed-Set Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 534 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "64. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards AUC-oriented open-set recognition. In: NeurIPS (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: ECCV (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "66. Wu, Y., Bourahla, O.E.F., Li, X., Wu, F., Tian, Q., Zhou, X.: Adaptive graph representation learning for video person re-identification. TIP (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "67. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. In: NeurIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "68. Ye, D., Fan, C., Ma, J., Liu, X., Yu, S.: Biggait: Learning gait representation you want by large vision models. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "69. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: CVPR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "70. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "71. Zahra, A., Perwaiz, N., Shahzad, M., Fraz, M.M.: Person re-identification: A retrospective on domain specific open challenges and future trends. Pattern Recognition (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "72. Zhai, Y., Liu, Z., Wu, Z., Wu, Y., Zhou, C., Doermann, D., Yuan, J., Hua, G.: Soar: Scene-debiasing open-set action recognition. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "73. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "74. Zhang, Z., Tran, L., Liu, F., Liu, X.: On learning disentangled representations for gait recognition. IEEE TPAMI (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "75. Zhang, Z., Tran, L., Yin, X., Atoum, Y., Liu, X., Wan, J., Wang, N.: Gait recognition via disentangled representation learning. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "76. Zheng, J., Liu, X., Gu, X., Sun, Y., Gan, C., Zhang, J., Liu, W., Yan, C.: Gait recognition in the wild with multi-hop temporal switch. In: ACMMM (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "77. Zheng, J., Liu, X., Liu, W., He, L., Yan, C., Mei, T.: Gait recognition in the wild with dense 3d representations and a benchmark. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "78. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: CVPR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "79. Zhu, Z., Guo, X., Yang, T., Huang, J., Deng, J., Huang, G., Du, D., Lu, J., Zhou, J.: Gait recognition in the wild: A benchmark. In: ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "80. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., et al.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: CVPR (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_content_list.json b/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dfaca726346c83c3211311b85f2622e586c054f8 --- /dev/null +++ b/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "text_level": 1, + "bbox": [ + 297, + 140, + 705, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dimity Miller, Niko Sünderhauf, Alex Kenna, and Keita Mason", + "bbox": [ + 271, + 212, + 727, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Queensland University of Technology, Brisbane, Australia\\*\\* {d24.miller, nikoo.suenderhauf}@qut.edu.au", + "bbox": [ + 300, + 239, + 697, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Are vision-language models (VLMs) for open-vocabulary perception inherently open-set models because they are trained on internet-scale datasets? We answer this question with a clear no - VLMs introduce closed-set assumptions via their finite query set, making them vulnerable to open-set conditions. We systematically evaluate VLMs for open-set recognition and find they frequently misclassify objects not contained in their query set, leading to alarmingly low precision when tuned for high recall and vice versa. We show that naively increasing the size of the query set to contain more and more classes does not mitigate this problem, but instead causes diminishing task performance and open-set performance. We establish a revised definition of the open-set problem for the age of VLMs, define a new benchmark and evaluation protocol to facilitate standardised evaluation and research in this important area, and evaluate promising baseline approaches based on predictive uncertainty and dedicated negative embeddings on a range of open-vocabulary VLM classifiers and object detectors.", + "bbox": [ + 261, + 306, + 738, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 556, + 374, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In 2013, Scheirer et al. [37] described the closed-set assumption ingrained in almost all vision models at the time: all test classes are known during training and the model is never exposed to novel or unknown classes during testing. Real-world applications challenge this assumption and test in open-set conditions - encountering unexpected objects not in the training dataset - and causing potentially dangerous overconfident misclassifications [4,10,26,38]. Open-set recognition [37] was introduced to address this issue by evaluating the ability of vision models to identify and reject open-set inputs as unknown.", + "bbox": [ + 212, + 588, + 785, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have recently revolutionised the fields of image classification [14, 19, 33] and object detection [15, 48, 50]. Trained on vast and diverse internet-scale datasets, VLMs recognise an extensive variety of classes and seemingly are open-set by default [42]. Our paper challenges this view.", + "bbox": [ + 212, + 709, + 785, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the age of VLMs, we introduce an updated definition of the open-set problem - while previous definitions focused on a finite training set [37], we show that", + "bbox": [ + 212, + 771, + 785, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "** The authors acknowledge ongoing support from the QUT Centre for Robotics.", + "bbox": [ + 214, + 811, + 777, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Evaluation code is publicly available at github.com/dimitymiller/openset_vlms", + "bbox": [ + 232, + 824, + 776, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2d187fdd875a836e0a2ce850aefe17c3593dabaa0f57cb9ec0655ef931791ba.jpg", + "image_caption": [ + "Fig. 1: Left: Traditional open-set recognition arises when testing classes outside a model's finite training class set. While VLMs are trained on internet-scale datasets containing most conceivable object classes, they use a finite query set for classification. Open-set recognition arises when test classes are not included in the query set. Right: When testing an object present in the predefined query set, VLMs often correctly classify the object. When testing an object that is not present in the query set (i.e. an open-set object), VLMs often misclassify the object as a query class with high confidence (i.e. an open-set error)." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 416, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3acd71d845b28255b00d1b39ed65607a18bfa03b3ed61cb8b2ec5be10371755f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 142, + 787, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLMs impose closed-set assumptions through a finite query set. As illustrated in Fig. 1, VLMs for open-vocabulary image classification or object detection are still evaluated under closed-set conditions: they compare image embeddings with text embeddings from a predefined, finite query set of class labels. This introduces a closed-set assumption that all classes encountered during testing are included in this query set. Open-set conditions emerge when VLMs encounter objects that are not included in the query set (unknown objects). As we show in Section 5, even state-of-the-art VLMs heavily degrade in performance and misclassify unknown objects as belonging to the query set with high confidence.", + "bbox": [ + 212, + 506, + 787, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In theory, this problem disappears if a VLM's query set contains every possible class label in a vocabulary. However, we show that naively increasing the query set to contain more and more class labels results in worse task performance, where increasing numbers of misclassifications occur (see Sec. 5.4). The serious concerns about deploying vision models in safety-critical applications voiced in the pre-VLM era [2,38] continues to be relevant and underscore the urgent need for further research into the open-set problem for VLMs.", + "bbox": [ + 212, + 643, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our paper is the first to systematically investigate and quantify the open-set vulnerabilities of VLM-based open-vocabulary object detectors and zero-shot image classifiers. We re-define the open-set problem for VLMs (Section 3) and compare promising baseline strategies, leveraging predictive uncertainty (Section 5.1) and different approaches to inserting dedicated negative (unknown) class representations into the VLM pipeline in Section 5.2. We propose a new", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "benchmark and evaluation protocol in Section 4 to foster further research and evaluation of VLMs for open-set recognition.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 214, + 200, + 369, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Vision-Language Models for Zero-shot Classification", + "text_level": 1, + "bbox": [ + 214, + 233, + 691, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A recently emerged foundation model is the vision-language model (VLM), which learns a multi-modal feature space that can jointly represent visual features and text [1, 19, 33, 46, 51] (and in some cases other modalities as well [14, 47]). Large-scale contrastive, self-supervised learning on internet-scale datasets provides VLMs with a rich feature representation that can be adapted for many downstream tasks, one of which is zero-shot or open vocabulary image classification [1, 14, 19, 33, 46, 47, 51]. Despite not being trained for classification tasks or datasets, VLMs generalise surprisingly well and achieve near state-of-the-art supervised learning performance [46]. Alongside this, a distinguishing advantage of VLMs is their adaptability to different datasets with different classes - the set of classes can be changed at will depending on the dataset [19, 33].", + "bbox": [ + 212, + 258, + 787, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Vision-Language Models for Open Vocabulary Object Detection", + "text_level": 1, + "bbox": [ + 214, + 448, + 785, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlocked by the advances in VLMs for zero-shot classification, [50] proposed open vocabulary object detection, where detectors must generalise to an arbitrary set of object classes at test time, even object classes that were not seen during training [50]. This is related to our proposed open-set recognition, but assumes that novel object classes are added to the query set, whereas we test against object classes that are not present in the query set.", + "bbox": [ + 212, + 473, + 787, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing open vocabulary detectors leverage VLM backbones to enable generalisation to new object classes, embedded within Faster R-CNN-style architectures [11, 12, 15, 20, 22, 31, 41, 43, 50, 54, 55] or DETR-style architectures [44, 48]. CLIP [33] is frequently utilised by these open vocabulary detectors, directly included in the architecture via its text encoder [11, 12, 15, 22, 31, 41, 43, 48, 54, 55] or in its entirety [20, 44], or used as a supervision signal during training [11, 12, 15, 31, 41, 43, 48, 54]. We refer the reader to [42] for a comprehensive survey on open vocabulary learning.", + "bbox": [ + 212, + 565, + 787, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Open-set Recognition", + "text_level": 1, + "bbox": [ + 214, + 709, + 441, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scheirer et al. [37] introduced open-set recognition to challenge the prevailing closed-set assumptions of image recognition systems. Closed-set models only test on samples from a predefined set of classes that were known during model training [37]. In open-set recognition, samples can instead come from an expansive range of classes that cannot be predefined, and therefore are unknown during model training [37]. Open-set recognition tests the ability of models to identify and reject unknown samples that do not belong to one of the training classes.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The open-set concept has been explored widely in computer vision, including in image classification [4, 7, 28, 30, 40, 45, 53], object detection [10, 16, 26, 27, 56], image segmentation [6, 17, 32] and other vision tasks [3, 24, 25, 29, 35].", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our paper is the first to systematically investigate the open-set problem for VLMs, and our findings challenge the assumption that the open-set problem has been solved for VLMs due to their exposure to internet-scale training data. The following sections define the closed-set assumptions existing in current VLMs, demonstrate that VLMs continue to be vulnerable to open-set conditions and evaluate a number of baseline strategies for open-set recognition with VLMs.", + "bbox": [ + 212, + 191, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem Definition", + "text_level": 1, + "bbox": [ + 215, + 308, + 439, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core concept behind VLMs, mapping images and text into a joint embedding space, makes them powerful foundations for image classification and object detection. Classification is performed by comparing an image embedding $\\mathbf{e}^I$ with text embeddings $\\mathcal{E}^q = \\{\\mathbf{e}_1^q,\\dots ,\\mathbf{e}_N^q\\}$ obtained from a set of query labels $\\mathcal{Q} = \\{q_1,\\ldots ,q_N\\}$ which are class names such as \"dog\". The similarities between the image embedding and all text embeddings from $\\mathcal{Q}$ are measured by a similarity function $\\mathbf{s}^q = \\mathrm{sim}(\\mathbf{e}^I,\\mathcal{E}^q)$ , e.g. using the cosine similarity. The text embedding with the highest similarity to the image embedding, i.e. $\\arg \\max \\mathbf{s}^q$ , determines the predicted class label. The process is the same for object detection, with $\\mathbf{e}^I$ now representing the embedding of an image region corresponding to an object proposal.", + "bbox": [ + 212, + 342, + 787, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-set and Closed-set: While the pre-VLM definition of the closed-set and open-set concepts focused on the challenges arising from a finite training set [37], our updated definition emphasises the limitations of the finite query set: The closed-set assumption in VLMs is that all classes seen during testing are predefined prior to testing and included in the query set. An open-set situation arises when the VLM encounters an object or image class that is not part of this predefined query set $\\mathcal{Q}$ .", + "bbox": [ + 212, + 508, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Baseline Approaches to Open-Set Recognition for VLMs", + "text_level": 1, + "bbox": [ + 214, + 638, + 728, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In an open-set situation, the VLM should be able to identify an input not represented by the query set and reject it as unknown, instead of misclassifying it as one of the query classes from $\\mathcal{Q}$ . Similar to the pre-VLM open-set literature, there are multiple ways of recognising an input as unknown [13]. We investigate baseline methods using predictive uncertainty and different approaches of negative embeddings.", + "bbox": [ + 212, + 665, + 787, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using Uncertainty: Different measures of predictive uncertainty $\\psi$ can be used to reject inputs as unknown, for example based on the class similarities, e.g.", + "bbox": [ + 214, + 756, + 787, + 787 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 Most methods insert an additional step here and generate a set of prompts from each query label $q_{i}$ , such as {\"A photo of a dog\", \"A picture of a dog\", ...} and compare the embeddings of these prompts with $\\mathbf{e}^{I}$ . For clarity of notation, we omit this step.", + "bbox": [ + 217, + 796, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\psi^{\\mathrm{Cosine}} = \\max (\\mathbf{s}^q)$ or $\\psi^{\\mathrm{Softmax}} = \\max (\\mathrm{softmax}(\\mathbf{s}^q))$ , or the entropy of the soft-max similarities $\\psi^{\\mathrm{Entropy}} = -H(\\mathrm{softmax}(\\mathbf{s}^q))$ . Only inputs that pass a threshold test $\\psi \\geq \\theta$ are classified according to $\\arg \\max \\mathbf{s}^q$ . Otherwise, the input is declared to be open-set and unknown.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dedicated Negative Embeddings: Another approach is to create a set of augmented query embeddings $\\tilde{\\mathcal{E}}^q = \\mathcal{E}^q \\cup \\mathcal{E}^n$ by adding $M$ dedicated negative class embeddings $\\mathcal{E}^n = \\{\\mathbf{e}_1^n, \\ldots, \\mathbf{e}_M^n\\}$ . The input image or detection can then be rejected as unknown if the image embedding is most similar to one of those negative embeddings $\\mathbf{e}_i^n$ rather than a query class embedding $\\mathbf{e}_j^q$ . These negative embeddings can be created in one of two ways: (1) by creating negative words, which are then passed through the VLM text encoder to create the negative embeddings, or (2) by directly creating negative embeddings.", + "bbox": [ + 212, + 207, + 784, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We investigate two baseline options for negative embeddings: a random words method, where we generate random strings with random length between 2-8 characters (e.g. \"braxl\"); and a random embeddings method, where we randomly draw $M$ embedding vectors from a Gaussian distribution, i.e. $\\mathbf{e}_i^{\\mathrm{n}}\\sim \\mathcal{N}(\\boldsymbol {\\mu},\\boldsymbol {\\Sigma})$ , using the mean and standard deviation from the query embedding vectors $\\mathcal{E}^q$ . We compare the effectiveness of these approaches, including for different numbers of $M$ random negative queries, in our experiments. Interestingly, many open vocabulary object detectors already use a negative query approach, though typically with a only single embedding vector of all zeroes [22,44,50,54,55].", + "bbox": [ + 212, + 329, + 785, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Evaluation Protocol", + "text_level": 1, + "bbox": [ + 215, + 505, + 444, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Creating an Open-set Recognition Dataset for VLMs", + "text_level": 1, + "bbox": [ + 215, + 539, + 697, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now introduce how to use existing image classification or object detection datasets to evaluate open-set recognition. In short, we test all images twice: once in a standard closed-set manner with all dataset classes included in the query set (see Fig. 1 top-right), and once in an open-set manner with a query set containing all dataset classes not present in the image (see Fig. 1 bottom-right).", + "bbox": [ + 212, + 566, + 784, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classification and object detection datasets have a pre-defined set of $K$ labelled target classes $\\mathcal{L}$ . In image classification, there is a single corresponding class label $y \\in \\mathcal{L}$ per image. In object detection, there are multiple ground truth objects in each image and each has a corresponding class label. For both tasks, we can define a set $\\mathcal{V}$ that contains the ground truth class labels for each image: $\\mathcal{V}$ contains a single label in classification and multiple labels in object detection.", + "bbox": [ + 212, + 643, + 785, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing Closed-set Recognition: We first test images in a standard manner to establish the closed-set performance of the VLM and identify all true positive $(TP)$ predictions. This involves testing all images with a query set identical to the dataset class list, $\\mathcal{Q} = \\mathcal{L}$ . To identify the TP predictions, we follow standard protocol for classification and object detection datasets - in classification, a TP has a predicted class matching the ground-truth label; in object detection, this involves a object-detection assignment process that requires correct object", + "bbox": [ + 214, + 734, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "classification and localisation (see Supplementary Sec. S1). TPs represent correct closed-set predictions that should not be rejected as unknown, and we collect the uncertainty measures $\\psi$ associated with all TPs for later evaluation.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For object detection datasets, unlabelled objects are present in the background of the image [10]. These unlabelled objects are not in the query set $\\mathcal{Q}$ , and therefore meet our definition for open-set objects. Detections of these open-set objects will present as false positives (FPs) in a closed-set test, however not all FPs arise from open-set objects - FPs in object detection can also arise due to poor object localisation, misclassification between query classes, or duplicate detections of a query class (see [5]). Rather than attempting to categorise different FP types to distinguish open-set errors, our approach definitively identifies open-set errors for both object detection and image classification.", + "bbox": [ + 212, + 191, + 787, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Testing Open-set Recognition: To evaluate open-set recognition, we test each image in the dataset with a modified query set, $\\tilde{\\mathcal{Q}}$ that contains all the class labels not in the image ground-truth class list, $\\tilde{\\mathcal{Q}} = \\mathcal{L} - \\mathcal{Y}$ . This ensures the image only contains object classes that are not in the modified query set - every single prediction must be an open-set error $(OSE)^2$ . In image classification, this modifies the query set to have $K - 1$ labels. In object detection, the modified query set size depends on the number of unique object classes in the image. We collect the uncertainty measures $\\psi$ of all OSE predictions for evaluation.", + "bbox": [ + 212, + 330, + 787, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Metrics", + "text_level": 1, + "bbox": [ + 215, + 479, + 326, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use performance metrics that measure the ability of the VLM to reject OSEs (with high uncertainty or with a negative class), while maintaining the closed-set task performance.", + "bbox": [ + 212, + 511, + 785, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Precision-Recall Curves: Open-set recognition can be formulated as a binary classification task that uses uncertainty as the decision threshold. Correct closed-set predictions (introduced as TPs in Sec. 4.1) and open-set errors (OSEs) are considered as the positive and negative class respectively. We construct precision-recall curves, which are well-suited to this problem as they are affected by the size of the negative class (OSEs) $[36]^3$ as well as prediction uncertainty. High precision indicates that little-to-none negatives (OSEs) remain after the uncertainty thresholding. High recall indicates that most positives (TPs) remain after the uncertainty thresholding. We summarise the performance of the precision-recall curves with 3 core metrics: (1) Area under the PR curve (AuPR), (2) Precision at $95\\%$ Recall, and (3) Recall at $95\\%$ Precision.", + "bbox": [ + 212, + 556, + 787, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task Metrics: We use the established task metrics used to evaluate classification and detection - top-1 accuracy [34] and mean Average Precision (mAP) [23]", + "bbox": [ + 212, + 724, + 785, + 756 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2 This assumes accurate labelling of ground-truth classes, though most datasets will contain some label error. This approach is unsuitable for weakly-labelled datasets.", + "bbox": [ + 217, + 768, + 785, + 797 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{3}$ ROC curves do not have this property - we discuss this further in Sec. S3 of the Supplementary Material where we report AuROC for the object detection experiments.", + "bbox": [ + 218, + 797, + 785, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "respectively. We evaluate mAP at an IoU threshold of 0.5 - we are primarily interested in the classification ability of the detector (rather than localisation) and this is standard protocol in open vocabulary detection [15,22,44,48,50,54,55]. Absolute Counts: We also report the absolute counts of TPs and OSEs. This can be particularly useful for object detection, where increased task performance does not necessarily indicate increased TPs and detectors predict variable numbers of OSEs. We stress that absolute counts should not be considered alone as a performance indicator, e.g. an increased number of OSEs can be acceptable if all OSEs have very high uncertainty.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 VLM Baselines", + "text_level": 1, + "bbox": [ + 214, + 305, + 388, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We test with six VLM classifiers - CLIP [33], ALIGN [19], CoCa [46], Image-Bind [14], SigLIP [52] and LanguageBind [58] - and seven open vocabulary object detectors - OVR-CNN [50], ViLD [15], OV-DETR [48], RegionCLIP [54], Detric [55], VLDet [22] and CORA [44] (see Supplementary Sec. S2 for implementation details). These VLMs cover a range of architectures and training paradigms and have publicly-available implementations. Our goal with testing this range of VLMs is not to find the \"best\" VLM for open-set recognition, but instead to exhibit the general vulnerability of VLMs to open-set conditions. For each VLM, we extract the prediction Softmax score, Cosine similarity, and Entropy of the softmax distribution as baseline uncertainty measures. Some VLMs use Sigmoid activation rather than Softmax [22,44,48,52,55] - for these methods, we only report the Sigmoid score as its results are identical to the cosine similarity and the entropy is unsuitable as the scores are not normalised to a distribution. We do not impose a minimum score threshold on any predictions, as the AuPR metric captures performance over all possible thresholds.", + "bbox": [ + 212, + 330, + 787, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Datasets", + "text_level": 1, + "bbox": [ + 214, + 580, + 333, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We focus on general object recognition and test with the large-scale and prevailing datasets used to benchmark image classification and object detection: For image classification, we test with the ImageNet1k validation dataset [9, 34]. It contains 50,000 images from 1000 classes and is publicly available for research. For object detection, we test with the MS COCO 2017 validation dataset [23]. It contains 4,952 images with 36,781 annotated objects from 80 classes and is publicly available for research. We show results for domain-specific classification datasets in the Supplementary Sec. S4.", + "bbox": [ + 212, + 606, + 787, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Experiments and Results", + "text_level": 1, + "bbox": [ + 214, + 751, + 496, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 State-of-the-art VLMs Perform Poorly in Open-set Conditions.", + "text_level": 1, + "bbox": [ + 214, + 782, + 781, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classification: Tab. 1 shows that all six VLM classifiers perform poorly for open-set recognition. For scenarios that require high recall of true positive", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a106caac5e6f9cf0cd64ba36f6d5e13bf393841a22e436c913696df0b67b7e9.jpg", + "table_caption": [ + "Table 1: The open-set performance of state-of-the-art VLM classifiers when tested on ImageNet1k. Arrows indicate direction of optimal performance and - indicates operating point could not be achieved." + ], + "table_footnote": [], + "table_body": "
ClassifierUncertaintyAuROCAuPRP@95RR@95PTPOSEAccuracy
↑↑↑↑↑↓↑
CLIP [33]Softmax79.771.546.23.4
Cosine72.261.342.50.131,02350,00062.1
Entropy80.272.345.42.8
ALIGN [19]Softmax81.074.748.18.3
Cosine72.462.144.8-32,61850,00065.2
Entropy80.975.046.88.6
CoCa [46]Softmax79.071.746.03.4
Cosine73.263.043.30.131,72550,00063.4
Entropy80.573.046.83.5
ImageBind [14]Softmax82.879.152.85.0
Cosine79.273.950.70.238,40550,00076.8
Entropy84.380.156.35.6
SigLIP [52]Sigmoid81.076.252.12.237,85150,00075.7
LanguageBind [58]Softmax83.980.754.710.5
Cosine81.677.452.80.839,24350,00078.5
Entropy85.481.758.411.1
", + "bbox": [ + 215, + 198, + 795, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "predictions (i.e. $95\\%$ recall), approximately every second prediction returned is an open-set error (i.e. precision between $46.2\\%$ and $58.4\\%$ ). For scenarios that require high precision of predictions (i.e. $95\\%$ precision), less than $12\\%$ of the true positive predictions are retained (i.e. recall between $3.4\\%$ and $11.1\\%$ ). Comparing across uncertainty types, the VLM classifiers show better uncertainty performance when using the predicted Softmax score or Entropy rather than the predicted Cosine similarity. This suggests raw similarity in VLM feature space is not the best indicator of open-set error, with measures of relative similarity comparative to other classes offering better performance.", + "bbox": [ + 212, + 530, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Detection: All seven of the tested VLM object detectors are vulnerable to open-set errors - see Tab. 2. Unlike VLM classifiers, most VLM detectors already contain a negative query to capture objects not in the query set (often referred to as the background class). Despite this, the VLM detectors produce open-set error counts in the range of 100,000 to 1,500,000 when tested on only 4,952 images, i.e. between 20 and 300 open-set error per image on average.", + "bbox": [ + 212, + 670, + 787, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "High open-set error counts are not inherently a problem if they are produced with high uncertainty. Yet for all tested detectors, the baseline uncertainty methods cannot adequately distinguish true positive detections from open-set errors. When thresholding for high recall of true positives, at best only every fifth detection is a true positive (i.e. precision ranges between $5.9\\%$ and $21.9\\%$ ). When", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/62576fbc24da883ef1704c6d1c72a828ab3dceb047558b2b17f1b0ce111946df.jpg", + "table_caption": [ + "Table 2: The open-set performance of state-of-the-art open-vocabulary object detectors (with ResNet backbone indicated), tested on COCO. Arrows indicate the direction of optimal performance." + ], + "table_footnote": [], + "table_body": "
DetectorUncertaintyNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]SoftmaxZero-Emb75.413.042.9
CosineZero-Emb77.215.343.613,544190,14634.7
EntropyZero-Emb63.27.543.4
ViLD (R152) [15]Softmax“background”60.17.515.4
Cosine“background”33.16.90.116,0111,485,60046.4
Entropy“background”62.28.416.3
OV-DETR (R50) [48]SigmoidNone75.85.947.315,8181,485,60043.0
RegionCLIP (R50) [54]SoftmaxZero-Emb73.19.939.4
CosineZero-Emb74.07.140.915,741493,94044.6
EntropyZero-Emb35.43.917.0
Detic (R50) [55]SigmoidZero-Emb72.68.142.214,670495,20039.1
VLDet (R50) [22]SigmoidZero-Emb79.821.948.514,751112,92340.6
CORA (R50) [44]SigmoidZero-Emb67.28.124.713,763495,20031.6
", + "bbox": [ + 215, + 198, + 782, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "thresholding for high precision, at best less than half of true positives are recalled (i.e. recall ranges between $16.3\\%$ and $48.5\\%$ ). We include some qualitative examples of highly confident open-set errors in the Supplementary Sec. S6.", + "bbox": [ + 215, + 460, + 784, + 505 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Notably, the precision-recall characteristics for VLM classification and object detection differ significantly – VLM classifiers achieve better performance at high recall, whereas VLM detectors achieve better performance at high precision. This suggests that while the VLM classifiers struggle with over-confident open-set errors, the VLM detectors suffer from under-confident true positives. We include histograms that visualise this trend in the Supplementary Sec. S7.", + "bbox": [ + 215, + 506, + 784, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also consider how the performance of uncertainty differs for identifying open-set error versus closed-set misclassifications, with results and a detailed discussion in the Supplementary Sec. S5. We find that the performance of softmax uncertainty is lower for open-set error than for closed-set error in both classification and detection tasks. In classification, open-set performance is particularly decremented when thresholding for high precision - likely due to a greater proportion of overconfident open-set error - whereas detection primarily degrades with increased numbers of errors, with ViLD for example producing $11.7 \\times$ more open-set than closed-set misclassifications.", + "bbox": [ + 215, + 597, + 784, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Can Negative Queries Improve Open-set Performance?", + "text_level": 1, + "bbox": [ + 217, + 753, + 712, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Classification: Negative class queries provide the VLM classifiers with the ability to select \"unknown\" when testing an input. When used in larger quantities, negative queries can capture and remove significant portions of open-set error, yet this happens in trade-off with a reduction in closed-set performance -", + "bbox": [ + 215, + 779, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1a28376a8efa8409199a5438b1595099f4eb1671116037bf086cfda79f8d2bcf.jpg", + "table_caption": [ + "Table 3: The open-set performance of state-of-the-art VLM classifiers on ImageNet1k when used with negative random words (R-W) or random embeddings (R-E)." + ], + "table_footnote": [], + "table_body": "
ClassifierNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓Accuracy ↑
CLIP [33] (Softmax)None71.546.23.431,02350,00062.1
500 R-W73.448.13.329,96141,42859.9
2500 R-W72.849.81.628,65735,09257.3
500 R-E71.947.22.530,16044,67360.3
2500 R-E71.747.10.929,29140,49758.6
ALIGN [19] (Softmax)None74.748.18.332,61850,00065.2
500 R-W75.348.910.431,73143,19863.5
2500 R-W74.649.58.930,58837,73861.2
500 R-E74.948.78.632,34247,22864.7
2500 R-E75.149.77.231,46241,21062.9
CoCa [46] (Softmax)None71.746.03.431,72550,00063.4
500 R-W75.449.60.130,27639,52860.6
2500 R-W75.950.96.029,10734,36058.2
500 R-E71.746.03.431,72349,96663.4
2500 R-E71.746.13.431,71949,86663.4
ImageBind [14] (Softmax)None79.152.85.038,40550,00076.8
500 R-W83.863.17.735,85929,17271.7
2500 R-W84.666.18.334,17922,87268.4
500 R-E81.057.39.036,96638,63673.9
2500 R-E81.759.011.235,95433,34971.9
SigLIP [52] (Sigmoid)None76.252.15.237,85150,00075.7
500 R-W77.656.72.336,11133,28772.2
2500 R-W78.860.22.434,50726,11069.0
500 R-E84.773.03.622,7729,57045.5
2500 R-E90.183.97.910,5612,25821.1
LanguageBind [58] (Softmax)None80.754.710.539,24350,00078.5
500 R-W84.764.913.136,59628,33773.2
2500 R-W85.468.211.734,54021,23069.1
500 R-E80.854.910.839,22649,64678.5
2500 R-E80.955.110.939,21049,03278.4
", + "bbox": [ + 272, + 184, + 728, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "see Tab. 3. For example, ImageBing using 2,500 random word negative queries effectively halves the open-set error, but this occurs in trade-off with a closed-set accuracy reduction of $8.4\\%$ (approximately 4,000 of the closed-set images are incorrectly classified as one of the random words).", + "bbox": [ + 212, + 642, + 785, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "While negative queries noticeably reduce the raw open-set error count, this is not always mirrored with increased performance in the open-set uncertainty metrics. We present the change in closed-set and open-set performance against the number of negative query points in Fig. 2. Generally the use of negative random words appears the most effective, with CoCa, ImageBind, SigLIP and Language-Bind showing improved open-set performance with minimal loss of closed-set performance. For some of the classifiers, there is negligible change in open-set performance despite the reduction in open-set error. We explore this further in the Supplementary Sec. S8, where we show that in some VLMs the negative", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1d6d2a49c30f0b30b0ddb8cc7229d817202099d78c738a68d062201a52532894.jpg", + "image_caption": [ + "Fig. 2: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM classifiers." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 774, + 316 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b25352584581fd230f2ec6b9cb5cb48e7df50ae17f800868cba8f9bf92df34e3.jpg", + "image_caption": [ + "Fig. 3: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM object detectors." + ], + "image_footnote": [], + "bbox": [ + 220, + 383, + 774, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "queries only capture and remove \"easy\" open-set errors, i.e. open-set errors that already have high uncertainty.", + "bbox": [ + 212, + 638, + 782, + 670 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Object Detection: The VLM detectors do not show a consistent response to the use of negative queries. Different detectors show different sensitivities to negative query counts, different trade-off rates between closed-set and open-set performance, and differences in effectiveness of embedding versus word negative query types - see Tab. 4 and Fig. 3. Looking at Fig. 3, we can consider the sensitivity, trade-off rate and response of the detectors to negative queries: greater performance changes with increasing queries indicates sensitivity; large decreases in mAP with more queries indicates increased trade-off rates between open-set and closed-set performance; and a positive response to the negative queries presents as an increase in AuPR when increasing queries (negative response decreases AuPR).", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 325, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fae1a9361dc708aa2202875fbf145f1737b9f0822f9d66c95bd953f23da7f19f.jpg", + "table_caption": [ + "Table 4: The open-set performance of state-of-the-art open vocabulary object detectors (with ResNet backbone indicated) when additional negative embeddings are included, tested on COCO. Arrows indicate the direction of optimal performance. Negative embeddings introduced as words are highlighted gray." + ], + "table_footnote": [], + "table_body": "
DetectorNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]Zero-Emb75.413.042.913,544190,14634.7
+5 Rand-Embs75.613.143.113,543186,12034.7
+100 Rand-Embs78.318.245.613,129109,26734.0
ViLD (R152) [15]“background”60.17.515.416,0111,485,60046.4
+5 Rand-Words57.56.013.415,9261,400,61246.6
+100 Rand-Words55.02.710.315,7721,088,72547.3
+5 Rand-Embs60.17.515.216,0111,485,23146.4
+100 Rand-Embs60.27.615.216,0191,478,80646.4
OV-DETR (R50) [48]None75.85.947.315,8181,485,60043.0
+5 Rand-Words77.28.148.315,510923,39843.0
+100 Rand-Words84.727.055.713,433239,11840.1
+5 Rand-Embs76.66.647.815,6521,178,82743.0
+100 Rand-Embs81.018.250.914,699356,38042.6
RegionCLIP (R50) [54]Zero-Emb73.19.939.315,752493,94944.6
+5 Rand-Words73.811.139.715,662322,59644.6
+100 Rand-Words77.017.543.115,178154,26244.0
+5 Rand-Embs91.877.244.68,8285,06227.0
+100 Rand-Embs99.999.383.628721.6
Detic (R50) [55]Zero-Emb72.68.142.214,670495,20039.1
+5 Rand-Words72.88.542.414,637451,63339.1
+100 Rand-Words74.511.543.514,287265,37338.8
+5 Rand-Embs93.474.561.810,0706,34630.6
+100 Rand-Embs100.0100.0100.037100.7
VLDet (R50) [22]Zero-Emb79.821.948.414,751112,95440.6
+5 Rand-Words80.122.348.714,672111,89340.5
+100 Rand-Words86.536.758.411,13250,12233.2
+5 Rand-Embs97.391.384.98,4641,38225.8
+100 Rand-Embs99.099.33.72,531225.9
CORA (R50) [44]Zero-Emb67.28.124.913,745495,20031.5
+5 Rand-Words66.37.627.013,700413,63631.6
+100 Rand-Words69.110.128.613,722329,39330.1
+5 Rand-Embs63.16.221.813,689495,20031.5
+100 Rand-Embs59.96.016.713,680495,17430.6
", + "bbox": [ + 228, + 212, + 769, + 694 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "OV-DETR is the only detector to show a clear positive response, low trade-off rate, and high sensitivity to the negative query approach, particularly with negative embeddings. While OVR-CNN (embeddings), RegionCLIP (words), Detric (words) also show a positive response and low trade-off rate, this is with very little sensitivity to increasing numbers of negative queries. In contrast, the open-set performance of Detric, RegionCLIP, and VLDet are very sensitive to negative", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3f1faff813e5f05cedd03c8050f31cdf1a9a20886cba024b9b3d2175f2460624.jpg", + "image_caption": [ + "(a) Classification task." + ], + "image_footnote": [], + "bbox": [ + 308, + 146, + 496, + 289 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a1aa75051a50cd3509a999c13d057de3e83f58c8d57e9d70281e624101b0a004.jpg", + "image_caption": [ + "(b) Object Detection task.", + "Fig. 4: The relationship between closed-set and open-set performance in VLMs." + ], + "image_footnote": [], + "bbox": [ + 500, + 146, + 694, + 289 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "embeddings, but at a rapid trade-off with closed-set performance. Both ViLD and CORA show a negative response or no sensitivity to negative queries.", + "bbox": [ + 212, + 358, + 782, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is tempting to relate these trends to the architecture or training paradigm of the VLM detectors, however we could not identify many common characteristics. RegionCLIP, Detic and VLDet all learn from auxiliary datasets that have been pseudo-labelled with regions [22, 54, 55], although this does not clearly indicate a reason for high sensitivity and trade-off rate for negative embeddings. Vild [15] and CORA [44] both heavily rely on CLIP's visual feature encoding (either by using CLIP directly [44] or knowledge distillation [15]) - their poor response is consistent with CLIP in the classification experiments on negative queries. Yet OV-DETR also uses CLIP's visual features during training to condition the object proposals [48], but exhibits the best response to negative queries.", + "bbox": [ + 212, + 390, + 784, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Is a good classifier all you need for VLM Open-set Recognition?", + "text_level": 1, + "bbox": [ + 212, + 561, + 782, + 575 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In traditional open-set recognition, Vaze et. al [40] identified a strong positive correlation between closed-set and open-set performance. Testing this theory for VLM open-set recognition, we compare classifier and object detector closed-set and open-set performance in Fig. 4.", + "bbox": [ + 212, + 584, + 784, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Classification: The correlation observed by [40] appears to hold for VLM classifiers, with greater accuracy (closed-set performance) relating to greater AuPR scores (open-set performance). It is worth noting that AuPR is a summary metric and this behaviour may not hold when examining performance at different operating points on the precision-recall curves.", + "bbox": [ + 212, + 645, + 784, + 719 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Object Detection: There is not a clear correlation between a VLM object detector's closed-set and open-set performance. ViLD in particular challenges this correlation, with the greatest mAP yet worst AuPR of all the tested detectors.", + "bbox": [ + 212, + 720, + 784, + 765 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 How Does Query Set Size Impact Performance?", + "text_level": 1, + "bbox": [ + 212, + 786, + 656, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Fig. 5, we test performance of the VLM classifiers on ImageNet while progressively increasing the query set size. We select the additional classes from the", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 325, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/767a2cee1c481a09e25d58855181a77c904d0bd4bf3f18a8db1b1a4624d5ab31.jpg", + "image_caption": [ + "Fig. 5: Influence of the query set size on closed-set and open-set performance on ImageNet1k classification, showing mean and standard deviation from 10 random seeds." + ], + "image_footnote": [], + "bbox": [ + 305, + 146, + 455, + 248 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/337bb0c7993599552ced58d7bfa22c1571e1cebb85b3304efa8ca74b15080da8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 146, + 697, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "WordNet hierarchy by choosing leaf nodes that are related, but not too closely related, to existing ImageNet classes (i.e. share a grandparent in the hierarchy but do not share a direct parent). Increasing the number of target classes and thus the query set size decrements both the closed-set and open-set performance of a VLM classifier. This indicates that testing VLMs with an entire vocabulary of nouns may not be a feasible approach to avoid open-set recognition.", + "bbox": [ + 212, + 318, + 784, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 431, + 354, + 448 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Open Challenges: The tested baseline uncertainty measures were ineffective at identifying open-set errors, indicating a need to research better alternatives. While uncertainty representations for VLMs have been proposed [18, 39], their utility for open-set recognition remains unexplored.", + "bbox": [ + 212, + 460, + 784, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While the use of negative random embeddings was able to reduce open-set error, large numbers were typically required to have any effect and this was often as the cost of falsely captured true positive predictions. Interestingly, some negative embeddings were much more effective at capturing open-set errors than others (e.g. the random string \"awy\" tested with ImageBind, see the Supplementary Sec. S9). Identifying the characteristics of these embeddings, learning them through prompt-tuning [8, 11, 12, 21, 57] or introducing auto-labelling techniques [49] are promising directions.", + "bbox": [ + 212, + 523, + 784, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusions: VLMs are increasingly applied in diverse real-world applications, from robotics to medical image analysis. We hope that by highlighting their weaknesses, we encourage further research to improve these models and enhance their positive societal impact. We have clearly demonstrated that VLMs for open-vocabulary image classification and object detection introduce closed-set assumptions by relying on a finite query set. We thus refuted the assumption that vision-language models (VLMs) are inherently open-set models due to their training on extensive internet-scale datasets. While the investigated baseline approaches to identify open-set errors showed promise, we found none of them sufficiently robust and reliable for safety-critical applications. Our new benchmark and evaluation protocol is designed to foster and support this research. Understanding VLMs' limitations ultimately reduces risks, improves transparency, and ensures safer deployment in critical applications.", + "bbox": [ + 212, + 643, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "2. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., Mané, D.: Concrete problems in ai safety. arXiv preprint arXiv:1606.06565 (2016)", + "3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021)", + "4. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016)", + "5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 558-573. Springer (2020)", + "6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021)", + "7. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021)", + "8. Cho, E., Kim, J., Kim, H.J.: Distribution-aware prompt tuning for vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22004-22013 (2023)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "0. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: The IEEE Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020)", + "1. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14084-14093 (2022)", + "2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022)", + "3. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020)", + "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023)", + "5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022)" + ], + "bbox": [ + 225, + 178, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Han, J., Ren, Y., Ding, J., Pan, X., Yan, K., Xia, G.S.: Expanding low-density latent regions for open-set object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9591-9600 (2022)", + "17. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021)", + "18. Ji, Y., Wang, J., Gong, Y., Zhang, L., Zhu, Y., Wang, H., Zhang, J., Sakai, T., Yang, Y.: Map: Multimodal uncertainty-aware vision-language pre-training model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23262-23271 (2023)", + "19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html", + "20. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. In: International Conference on Learning Representations (2023)", + "21. Li, H., Zhang, R., Yao, H., Song, X., Hao, Y., Zhao, Y., Li, L., Chen, Y.: Learning domain-aware detection head with prompt tuning. Advances in Neural Information Processing Systems 36 (2024)", + "22. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: International Conference on Learning Representations (2023)", + "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "24. Liu, Y.C., Ma, C.Y., Dai, X., Tian, J., Vajda, P., He, Z., Kira, Z.: Open-set semi-supervised object detection. In: European Conference on Computer Vision. pp. 143-159. Springer (2022)", + "25. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024)", + "26. Miller, D., Nicholson, L., Dayoub, F., Sunderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249. IEEE (2018)", + "27. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. IEEE Robotics and Automation Letters 7(1), 215-222 (2022). https://doi.org/10.1109/LRA.2021.3123374", + "28. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019)", + "29. Panareda Busto, P., Gall, J.: Open set domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 754-763 (2017)", + "30. Perera, P., Morariu, V.I., Jain, R., Manjunatha, V., Wigington, C., Ordonez, V., Patel, V.M.: Generative-discriminative feature representations for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11814-11823 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Pham, C., Vu, T., Nguyen, K.: Lp-ovod: Open-vocabulary object detection by linear probing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 779–788 (2024)", + "32. Pham, T., Do, T.T., Carneiro, G., Reid, I., et al.: Bayesian semantic instance segmentation in open set world. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 3-18 (2018)", + "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html", + "34. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)", + "35. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-168 (2018)", + "36. Saito, T., Rehmsmeier, M.: The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PloS one 10(3), e0118432 (2015)", + "37. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(7), 1757-1772 (2013). https://doi.org/10.1109/TPAMI.2012.256", + "38. Sünderhauf, N., Brock, O., Scheirer, W., Hadsell, R., Fox, D., Leitner, J., Upcroft, B., Abbeel, P., Burgard, W., Milford, M., et al.: The limits and potentials of deep learning for robotics. The International Journal of Robotics Research 37(4-5), 405-420 (2018)", + "39. Upadhyay, U., Karthik, S., Mancini, M., Akata, Z.: Probvlm: Probabilistic adapter for frozen vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1899-1910 (2023)", + "40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022)", + "41. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6759-6769 (2023)", + "42. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards Open Vocabulary Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)", + "43. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023)", + "44. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023)", + "45. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Open-Set Recognition in the Age of Vision-Language Models", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022), https://openreview.net/forum?id=Ee277P3AYC", + "47. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)", + "48. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: European Conference on Computer Vision. pp. 106-122. Springer (2022)", + "49. Zara, G., Roy, S., Rota, P., Ricci, E.: Autolabel: Clip-based framework for open-set video domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11504-11513 (2023)", + "50. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)", + "51. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022)", + "52. Zhai, X., et al.: Sigmoid loss for language image pre-training. In: Intl. Conf. on Computer Vision. pp. 11975-11986 (2023)", + "53. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 102-117. Springer (2020)", + "54. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)", + "55. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)", + "56. Zhou, Z., Yang, Y., Wang, Y., Xiong, R.: Open-set object detection using classification-free object proposal and instance-level contrastive learning. IEEE Robotics and Automation Letters 8(3), 1691-1698 (2023)", + "57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)", + "58. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: Intl. Conf. on Learning Representations (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 715 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "D. Miller et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_model.json b/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1f578c30b432a09ecfe9b13fe8b568cd9a3a2ea5 --- /dev/null +++ b/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_model.json @@ -0,0 +1,2271 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.298, + 0.141, + 0.707, + 0.186 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.213, + 0.728, + 0.228 + ], + "angle": 0, + "content": "Dimity Miller, Niko Sünderhauf, Alex Kenna, and Keita Mason" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.24, + 0.698, + 0.268 + ], + "angle": 0, + "content": "Queensland University of Technology, Brisbane, Australia\\*\\* {d24.miller, nikoo.suenderhauf}@qut.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.308, + 0.74, + 0.529 + ], + "angle": 0, + "content": "Abstract. Are vision-language models (VLMs) for open-vocabulary perception inherently open-set models because they are trained on internet-scale datasets? We answer this question with a clear no - VLMs introduce closed-set assumptions via their finite query set, making them vulnerable to open-set conditions. We systematically evaluate VLMs for open-set recognition and find they frequently misclassify objects not contained in their query set, leading to alarmingly low precision when tuned for high recall and vice versa. We show that naively increasing the size of the query set to contain more and more classes does not mitigate this problem, but instead causes diminishing task performance and open-set performance. We establish a revised definition of the open-set problem for the age of VLMs, define a new benchmark and evaluation protocol to facilitate standardised evaluation and research in this important area, and evaluate promising baseline approaches based on predictive uncertainty and dedicated negative embeddings on a range of open-vocabulary VLM classifiers and object detectors." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.557, + 0.375, + 0.572 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.787, + 0.71 + ], + "angle": 0, + "content": "In 2013, Scheirer et al. [37] described the closed-set assumption ingrained in almost all vision models at the time: all test classes are known during training and the model is never exposed to novel or unknown classes during testing. Real-world applications challenge this assumption and test in open-set conditions - encountering unexpected objects not in the training dataset - and causing potentially dangerous overconfident misclassifications [4,10,26,38]. Open-set recognition [37] was introduced to address this issue by evaluating the ability of vision models to identify and reject open-set inputs as unknown." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.787, + 0.771 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have recently revolutionised the fields of image classification [14, 19, 33] and object detection [15, 48, 50]. Trained on vast and diverse internet-scale datasets, VLMs recognise an extensive variety of classes and seemingly are open-set by default [42]. Our paper challenges this view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.803 + ], + "angle": 0, + "content": "In the age of VLMs, we introduce an updated definition of the open-set problem - while previous definitions focused on a finite training set [37], we show that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.812, + 0.778, + 0.825 + ], + "angle": 0, + "content": "** The authors acknowledge ongoing support from the QUT Centre for Robotics." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.777, + 0.84 + ], + "angle": 0, + "content": "Evaluation code is publicly available at github.com/dimitymiller/openset_vlms" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.144, + 0.418, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.143, + 0.788, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig. 1: Left: Traditional open-set recognition arises when testing classes outside a model's finite training class set. While VLMs are trained on internet-scale datasets containing most conceivable object classes, they use a finite query set for classification. Open-set recognition arises when test classes are not included in the query set. Right: When testing an object present in the predefined query set, VLMs often correctly classify the object. When testing an object that is not present in the query set (i.e. an open-set object), VLMs often misclassify the object as a query class with high confidence (i.e. an open-set error)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.643 + ], + "angle": 0, + "content": "VLMs impose closed-set assumptions through a finite query set. As illustrated in Fig. 1, VLMs for open-vocabulary image classification or object detection are still evaluated under closed-set conditions: they compare image embeddings with text embeddings from a predefined, finite query set of class labels. This introduces a closed-set assumption that all classes encountered during testing are included in this query set. Open-set conditions emerge when VLMs encounter objects that are not included in the query set (unknown objects). As we show in Section 5, even state-of-the-art VLMs heavily degrade in performance and misclassify unknown objects as belonging to the query set with high confidence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In theory, this problem disappears if a VLM's query set contains every possible class label in a vocabulary. However, we show that naively increasing the query set to contain more and more class labels results in worse task performance, where increasing numbers of misclassifications occur (see Sec. 5.4). The serious concerns about deploying vision models in safety-critical applications voiced in the pre-VLM era [2,38] continues to be relevant and underscore the urgent need for further research into the open-set problem for VLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our paper is the first to systematically investigate and quantify the open-set vulnerabilities of VLM-based open-vocabulary object detectors and zero-shot image classifiers. We re-define the open-set problem for VLMs (Section 3) and compare promising baseline strategies, leveraging predictive uncertainty (Section 5.1) and different approaches to inserting dedicated negative (unknown) class representations into the VLM pipeline in Section 5.2. We propose a new" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "benchmark and evaluation protocol in Section 4 to foster further research and evaluation of VLMs for open-set recognition." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.201, + 0.37, + 0.218 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.234, + 0.692, + 0.249 + ], + "angle": 0, + "content": "2.1 Vision-Language Models for Zero-shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.427 + ], + "angle": 0, + "content": "A recently emerged foundation model is the vision-language model (VLM), which learns a multi-modal feature space that can jointly represent visual features and text [1, 19, 33, 46, 51] (and in some cases other modalities as well [14, 47]). Large-scale contrastive, self-supervised learning on internet-scale datasets provides VLMs with a rich feature representation that can be adapted for many downstream tasks, one of which is zero-shot or open vocabulary image classification [1, 14, 19, 33, 46, 47, 51]. Despite not being trained for classification tasks or datasets, VLMs generalise surprisingly well and achieve near state-of-the-art supervised learning performance [46]. Alongside this, a distinguishing advantage of VLMs is their adaptability to different datasets with different classes - the set of classes can be changed at will depending on the dataset [19, 33]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.449, + 0.787, + 0.466 + ], + "angle": 0, + "content": "2.2 Vision-Language Models for Open Vocabulary Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.565 + ], + "angle": 0, + "content": "Unlocked by the advances in VLMs for zero-shot classification, [50] proposed open vocabulary object detection, where detectors must generalise to an arbitrary set of object classes at test time, even object classes that were not seen during training [50]. This is related to our proposed open-set recognition, but assumes that novel object classes are added to the query set, whereas we test against object classes that are not present in the query set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Existing open vocabulary detectors leverage VLM backbones to enable generalisation to new object classes, embedded within Faster R-CNN-style architectures [11, 12, 15, 20, 22, 31, 41, 43, 50, 54, 55] or DETR-style architectures [44, 48]. CLIP [33] is frequently utilised by these open vocabulary detectors, directly included in the architecture via its text encoder [11, 12, 15, 22, 31, 41, 43, 48, 54, 55] or in its entirety [20, 44], or used as a supervision signal during training [11, 12, 15, 31, 41, 43, 48, 54]. We refer the reader to [42] for a comprehensive survey on open vocabulary learning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.442, + 0.725 + ], + "angle": 0, + "content": "2.3 Open-set Recognition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Scheirer et al. [37] introduced open-set recognition to challenge the prevailing closed-set assumptions of image recognition systems. Closed-set models only test on samples from a predefined set of classes that were known during model training [37]. In open-set recognition, samples can instead come from an expansive range of classes that cannot be predefined, and therefore are unknown during model training [37]. Open-set recognition tests the ability of models to identify and reject unknown samples that do not belong to one of the training classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "The open-set concept has been explored widely in computer vision, including in image classification [4, 7, 28, 30, 40, 45, 53], object detection [10, 16, 26, 27, 56], image segmentation [6, 17, 32] and other vision tasks [3, 24, 25, 29, 35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Our paper is the first to systematically investigate the open-set problem for VLMs, and our findings challenge the assumption that the open-set problem has been solved for VLMs due to their exposure to internet-scale training data. The following sections define the closed-set assumptions existing in current VLMs, demonstrate that VLMs continue to be vulnerable to open-set conditions and evaluate a number of baseline strategies for open-set recognition with VLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.309, + 0.441, + 0.325 + ], + "angle": 0, + "content": "3 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.508 + ], + "angle": 0, + "content": "The core concept behind VLMs, mapping images and text into a joint embedding space, makes them powerful foundations for image classification and object detection. Classification is performed by comparing an image embedding \\(\\mathbf{e}^I\\) with text embeddings \\(\\mathcal{E}^q = \\{\\mathbf{e}_1^q,\\dots ,\\mathbf{e}_N^q\\}\\) obtained from a set of query labels \\(\\mathcal{Q} = \\{q_1,\\ldots ,q_N\\}\\) which are class names such as \"dog\". The similarities between the image embedding and all text embeddings from \\(\\mathcal{Q}\\) are measured by a similarity function \\(\\mathbf{s}^q = \\mathrm{sim}(\\mathbf{e}^I,\\mathcal{E}^q)\\), e.g. using the cosine similarity. The text embedding with the highest similarity to the image embedding, i.e. \\(\\arg \\max \\mathbf{s}^q\\), determines the predicted class label. The process is the same for object detection, with \\(\\mathbf{e}^I\\) now representing the embedding of an image region corresponding to an object proposal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Open-set and Closed-set: While the pre-VLM definition of the closed-set and open-set concepts focused on the challenges arising from a finite training set [37], our updated definition emphasises the limitations of the finite query set: The closed-set assumption in VLMs is that all classes seen during testing are predefined prior to testing and included in the query set. An open-set situation arises when the VLM encounters an object or image class that is not part of this predefined query set \\(\\mathcal{Q}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.639, + 0.729, + 0.654 + ], + "angle": 0, + "content": "3.1 Baseline Approaches to Open-Set Recognition for VLMs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.756 + ], + "angle": 0, + "content": "In an open-set situation, the VLM should be able to identify an input not represented by the query set and reject it as unknown, instead of misclassifying it as one of the query classes from \\(\\mathcal{Q}\\). Similar to the pre-VLM open-set literature, there are multiple ways of recognising an input as unknown [13]. We investigate baseline methods using predictive uncertainty and different approaches of negative embeddings." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.757, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Using Uncertainty: Different measures of predictive uncertainty \\(\\psi\\) can be used to reject inputs as unknown, for example based on the class similarities, e.g." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.842 + ], + "angle": 0, + "content": "1 Most methods insert an additional step here and generate a set of prompts from each query label \\( q_{i} \\), such as {\"A photo of a dog\", \"A picture of a dog\", ...} and compare the embeddings of these prompts with \\( \\mathbf{e}^{I} \\). For clarity of notation, we omit this step." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "\\(\\psi^{\\mathrm{Cosine}} = \\max (\\mathbf{s}^q)\\) or \\(\\psi^{\\mathrm{Softmax}} = \\max (\\mathrm{softmax}(\\mathbf{s}^q))\\), or the entropy of the soft-max similarities \\(\\psi^{\\mathrm{Entropy}} = -H(\\mathrm{softmax}(\\mathbf{s}^q))\\). Only inputs that pass a threshold test \\(\\psi \\geq \\theta\\) are classified according to \\(\\arg \\max \\mathbf{s}^q\\). Otherwise, the input is declared to be open-set and unknown." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Dedicated Negative Embeddings: Another approach is to create a set of augmented query embeddings \\(\\tilde{\\mathcal{E}}^q = \\mathcal{E}^q \\cup \\mathcal{E}^n\\) by adding \\(M\\) dedicated negative class embeddings \\(\\mathcal{E}^n = \\{\\mathbf{e}_1^n, \\ldots, \\mathbf{e}_M^n\\}\\). The input image or detection can then be rejected as unknown if the image embedding is most similar to one of those negative embeddings \\(\\mathbf{e}_i^n\\) rather than a query class embedding \\(\\mathbf{e}_j^q\\). These negative embeddings can be created in one of two ways: (1) by creating negative words, which are then passed through the VLM text encoder to create the negative embeddings, or (2) by directly creating negative embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.787, + 0.482 + ], + "angle": 0, + "content": "We investigate two baseline options for negative embeddings: a random words method, where we generate random strings with random length between 2-8 characters (e.g. \"braxl\"); and a random embeddings method, where we randomly draw \\(M\\) embedding vectors from a Gaussian distribution, i.e. \\(\\mathbf{e}_i^{\\mathrm{n}}\\sim \\mathcal{N}(\\boldsymbol {\\mu},\\boldsymbol {\\Sigma})\\), using the mean and standard deviation from the query embedding vectors \\(\\mathcal{E}^q\\). We compare the effectiveness of these approaches, including for different numbers of \\(M\\) random negative queries, in our experiments. Interestingly, many open vocabulary object detectors already use a negative query approach, though typically with a only single embedding vector of all zeroes [22,44,50,54,55]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.506, + 0.446, + 0.523 + ], + "angle": 0, + "content": "4 Evaluation Protocol" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.54, + 0.699, + 0.556 + ], + "angle": 0, + "content": "4.1 Creating an Open-set Recognition Dataset for VLMs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.644 + ], + "angle": 0, + "content": "We now introduce how to use existing image classification or object detection datasets to evaluate open-set recognition. In short, we test all images twice: once in a standard closed-set manner with all dataset classes included in the query set (see Fig. 1 top-right), and once in an open-set manner with a query set containing all dataset classes not present in the image (see Fig. 1 bottom-right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.735 + ], + "angle": 0, + "content": "Classification and object detection datasets have a pre-defined set of \\( K \\) labelled target classes \\( \\mathcal{L} \\). In image classification, there is a single corresponding class label \\( y \\in \\mathcal{L} \\) per image. In object detection, there are multiple ground truth objects in each image and each has a corresponding class label. For both tasks, we can define a set \\( \\mathcal{V} \\) that contains the ground truth class labels for each image: \\( \\mathcal{V} \\) contains a single label in classification and multiple labels in object detection." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Testing Closed-set Recognition: We first test images in a standard manner to establish the closed-set performance of the VLM and identify all true positive \\((TP)\\) predictions. This involves testing all images with a query set identical to the dataset class list, \\(\\mathcal{Q} = \\mathcal{L}\\). To identify the TP predictions, we follow standard protocol for classification and object detection datasets - in classification, a TP has a predicted class matching the ground-truth label; in object detection, this involves a object-detection assignment process that requires correct object" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.192 + ], + "angle": 0, + "content": "classification and localisation (see Supplementary Sec. S1). TPs represent correct closed-set predictions that should not be rejected as unknown, and we collect the uncertainty measures \\(\\psi\\) associated with all TPs for later evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.329 + ], + "angle": 0, + "content": "For object detection datasets, unlabelled objects are present in the background of the image [10]. These unlabelled objects are not in the query set \\(\\mathcal{Q}\\), and therefore meet our definition for open-set objects. Detections of these open-set objects will present as false positives (FPs) in a closed-set test, however not all FPs arise from open-set objects - FPs in object detection can also arise due to poor object localisation, misclassification between query classes, or duplicate detections of a query class (see [5]). Rather than attempting to categorise different FP types to distinguish open-set errors, our approach definitively identifies open-set errors for both object detection and image classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Testing Open-set Recognition: To evaluate open-set recognition, we test each image in the dataset with a modified query set, \\(\\tilde{\\mathcal{Q}}\\) that contains all the class labels not in the image ground-truth class list, \\(\\tilde{\\mathcal{Q}} = \\mathcal{L} - \\mathcal{Y}\\). This ensures the image only contains object classes that are not in the modified query set - every single prediction must be an open-set error \\((OSE)^2\\). In image classification, this modifies the query set to have \\(K - 1\\) labels. In object detection, the modified query set size depends on the number of unique object classes in the image. We collect the uncertainty measures \\(\\psi\\) of all OSE predictions for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.327, + 0.495 + ], + "angle": 0, + "content": "4.2 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.787, + 0.557 + ], + "angle": 0, + "content": "We use performance metrics that measure the ability of the VLM to reject OSEs (with high uncertainty or with a negative class), while maintaining the closed-set task performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Precision-Recall Curves: Open-set recognition can be formulated as a binary classification task that uses uncertainty as the decision threshold. Correct closed-set predictions (introduced as TPs in Sec. 4.1) and open-set errors (OSEs) are considered as the positive and negative class respectively. We construct precision-recall curves, which are well-suited to this problem as they are affected by the size of the negative class (OSEs) \\([36]^3\\) as well as prediction uncertainty. High precision indicates that little-to-none negatives (OSEs) remain after the uncertainty thresholding. High recall indicates that most positives (TPs) remain after the uncertainty thresholding. We summarise the performance of the precision-recall curves with 3 core metrics: (1) Area under the PR curve (AuPR), (2) Precision at \\(95\\%\\) Recall, and (3) Recall at \\(95\\%\\) Precision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.787, + 0.757 + ], + "angle": 0, + "content": "Task Metrics: We use the established task metrics used to evaluate classification and detection - top-1 accuracy [34] and mean Average Precision (mAP) [23]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.77, + 0.786, + 0.798 + ], + "angle": 0, + "content": "2 This assumes accurate labelling of ground-truth classes, though most datasets will contain some label error. This approach is unsuitable for weakly-labelled datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.799, + 0.787, + 0.839 + ], + "angle": 0, + "content": "\\(^{3}\\) ROC curves do not have this property - we discuss this further in Sec. S3 of the Supplementary Material where we report AuROC for the object detection experiments." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.77, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "respectively. We evaluate mAP at an IoU threshold of 0.5 - we are primarily interested in the classification ability of the detector (rather than localisation) and this is standard protocol in open vocabulary detection [15,22,44,48,50,54,55]. Absolute Counts: We also report the absolute counts of TPs and OSEs. This can be particularly useful for object detection, where increased task performance does not necessarily indicate increased TPs and detectors predict variable numbers of OSEs. We stress that absolute counts should not be considered alone as a performance indicator, e.g. an increased number of OSEs can be acceptable if all OSEs have very high uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.306, + 0.389, + 0.321 + ], + "angle": 0, + "content": "4.3 VLM Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.789, + 0.56 + ], + "angle": 0, + "content": "We test with six VLM classifiers - CLIP [33], ALIGN [19], CoCa [46], Image-Bind [14], SigLIP [52] and LanguageBind [58] - and seven open vocabulary object detectors - OVR-CNN [50], ViLD [15], OV-DETR [48], RegionCLIP [54], Detric [55], VLDet [22] and CORA [44] (see Supplementary Sec. S2 for implementation details). These VLMs cover a range of architectures and training paradigms and have publicly-available implementations. Our goal with testing this range of VLMs is not to find the \"best\" VLM for open-set recognition, but instead to exhibit the general vulnerability of VLMs to open-set conditions. For each VLM, we extract the prediction Softmax score, Cosine similarity, and Entropy of the softmax distribution as baseline uncertainty measures. Some VLMs use Sigmoid activation rather than Softmax [22,44,48,52,55] - for these methods, we only report the Sigmoid score as its results are identical to the cosine similarity and the entropy is unsuitable as the scores are not normalised to a distribution. We do not impose a minimum score threshold on any predictions, as the AuPR metric captures performance over all possible thresholds." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.334, + 0.595 + ], + "angle": 0, + "content": "4.4 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.729 + ], + "angle": 0, + "content": "We focus on general object recognition and test with the large-scale and prevailing datasets used to benchmark image classification and object detection: For image classification, we test with the ImageNet1k validation dataset [9, 34]. It contains 50,000 images from 1000 classes and is publicly available for research. For object detection, we test with the MS COCO 2017 validation dataset [23]. It contains 4,952 images with 36,781 annotated objects from 80 classes and is publicly available for research. We show results for domain-specific classification datasets in the Supplementary Sec. S4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.752, + 0.498, + 0.769 + ], + "angle": 0, + "content": "5 Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.782, + 0.8 + ], + "angle": 0, + "content": "5.1 State-of-the-art VLMs Perform Poorly in Open-set Conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Classification: Tab. 1 shows that all six VLM classifiers perform poorly for open-set recognition. For scenarios that require high recall of true positive" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: The open-set performance of state-of-the-art VLM classifiers when tested on ImageNet1k. Arrows indicate direction of optimal performance and - indicates operating point could not be achieved." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.199, + 0.797, + 0.498 + ], + "angle": 0, + "content": "
ClassifierUncertaintyAuROCAuPRP@95RR@95PTPOSEAccuracy
↑↑↑↑↑↓↑
CLIP [33]Softmax79.771.546.23.4
Cosine72.261.342.50.131,02350,00062.1
Entropy80.272.345.42.8
ALIGN [19]Softmax81.074.748.18.3
Cosine72.462.144.8-32,61850,00065.2
Entropy80.975.046.88.6
CoCa [46]Softmax79.071.746.03.4
Cosine73.263.043.30.131,72550,00063.4
Entropy80.573.046.83.5
ImageBind [14]Softmax82.879.152.85.0
Cosine79.273.950.70.238,40550,00076.8
Entropy84.380.156.35.6
SigLIP [52]Sigmoid81.076.252.12.237,85150,00075.7
LanguageBind [58]Softmax83.980.754.710.5
Cosine81.677.452.80.839,24350,00078.5
Entropy85.481.758.411.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.669 + ], + "angle": 0, + "content": "predictions (i.e. \\(95\\%\\) recall), approximately every second prediction returned is an open-set error (i.e. precision between \\(46.2\\%\\) and \\(58.4\\%\\)). For scenarios that require high precision of predictions (i.e. \\(95\\%\\) precision), less than \\(12\\%\\) of the true positive predictions are retained (i.e. recall between \\(3.4\\%\\) and \\(11.1\\%\\)). Comparing across uncertainty types, the VLM classifiers show better uncertainty performance when using the predicted Softmax score or Entropy rather than the predicted Cosine similarity. This suggests raw similarity in VLM feature space is not the best indicator of open-set error, with measures of relative similarity comparative to other classes offering better performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Object Detection: All seven of the tested VLM object detectors are vulnerable to open-set errors - see Tab. 2. Unlike VLM classifiers, most VLM detectors already contain a negative query to capture objects not in the query set (often referred to as the background class). Despite this, the VLM detectors produce open-set error counts in the range of 100,000 to 1,500,000 when tested on only 4,952 images, i.e. between 20 and 300 open-set error per image on average." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "High open-set error counts are not inherently a problem if they are produced with high uncertainty. Yet for all tested detectors, the baseline uncertainty methods cannot adequately distinguish true positive detections from open-set errors. When thresholding for high recall of true positives, at best only every fifth detection is a true positive (i.e. precision ranges between \\(5.9\\%\\) and \\(21.9\\%\\)). When" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 2: The open-set performance of state-of-the-art open-vocabulary object detectors (with ResNet backbone indicated), tested on COCO. Arrows indicate the direction of optimal performance." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.199, + 0.784, + 0.433 + ], + "angle": 0, + "content": "
DetectorUncertaintyNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]SoftmaxZero-Emb75.413.042.9
CosineZero-Emb77.215.343.613,544190,14634.7
EntropyZero-Emb63.27.543.4
ViLD (R152) [15]Softmax“background”60.17.515.4
Cosine“background”33.16.90.116,0111,485,60046.4
Entropy“background”62.28.416.3
OV-DETR (R50) [48]SigmoidNone75.85.947.315,8181,485,60043.0
RegionCLIP (R50) [54]SoftmaxZero-Emb73.19.939.4
CosineZero-Emb74.07.140.915,741493,94044.6
EntropyZero-Emb35.43.917.0
Detic (R50) [55]SigmoidZero-Emb72.68.142.214,670495,20039.1
VLDet (R50) [22]SigmoidZero-Emb79.821.948.514,751112,92340.6
CORA (R50) [44]SigmoidZero-Emb67.28.124.713,763495,20031.6
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.506 + ], + "angle": 0, + "content": "thresholding for high precision, at best less than half of true positives are recalled (i.e. recall ranges between \\(16.3\\%\\) and \\(48.5\\%\\)). We include some qualitative examples of highly confident open-set errors in the Supplementary Sec. S6." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.507, + 0.785, + 0.597 + ], + "angle": 0, + "content": "Notably, the precision-recall characteristics for VLM classification and object detection differ significantly – VLM classifiers achieve better performance at high recall, whereas VLM detectors achieve better performance at high precision. This suggests that while the VLM classifiers struggle with over-confident open-set errors, the VLM detectors suffer from under-confident true positives. We include histograms that visualise this trend in the Supplementary Sec. S7." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.598, + 0.785, + 0.732 + ], + "angle": 0, + "content": "We also consider how the performance of uncertainty differs for identifying open-set error versus closed-set misclassifications, with results and a detailed discussion in the Supplementary Sec. S5. We find that the performance of softmax uncertainty is lower for open-set error than for closed-set error in both classification and detection tasks. In classification, open-set performance is particularly decremented when thresholding for high precision - likely due to a greater proportion of overconfident open-set error - whereas detection primarily degrades with increased numbers of errors, with ViLD for example producing \\(11.7 \\times\\) more open-set than closed-set misclassifications." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.755, + 0.713, + 0.77 + ], + "angle": 0, + "content": "5.2 Can Negative Queries Improve Open-set Performance?" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Classification: Negative class queries provide the VLM classifiers with the ability to select \"unknown\" when testing an input. When used in larger quantities, negative queries can capture and remove significant portions of open-set error, yet this happens in trade-off with a reduction in closed-set performance -" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 3: The open-set performance of state-of-the-art VLM classifiers on ImageNet1k when used with negative random words (R-W) or random embeddings (R-E)." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.185, + 0.729, + 0.615 + ], + "angle": 0, + "content": "
ClassifierNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓Accuracy ↑
CLIP [33] (Softmax)None71.546.23.431,02350,00062.1
500 R-W73.448.13.329,96141,42859.9
2500 R-W72.849.81.628,65735,09257.3
500 R-E71.947.22.530,16044,67360.3
2500 R-E71.747.10.929,29140,49758.6
ALIGN [19] (Softmax)None74.748.18.332,61850,00065.2
500 R-W75.348.910.431,73143,19863.5
2500 R-W74.649.58.930,58837,73861.2
500 R-E74.948.78.632,34247,22864.7
2500 R-E75.149.77.231,46241,21062.9
CoCa [46] (Softmax)None71.746.03.431,72550,00063.4
500 R-W75.449.60.130,27639,52860.6
2500 R-W75.950.96.029,10734,36058.2
500 R-E71.746.03.431,72349,96663.4
2500 R-E71.746.13.431,71949,86663.4
ImageBind [14] (Softmax)None79.152.85.038,40550,00076.8
500 R-W83.863.17.735,85929,17271.7
2500 R-W84.666.18.334,17922,87268.4
500 R-E81.057.39.036,96638,63673.9
2500 R-E81.759.011.235,95433,34971.9
SigLIP [52] (Sigmoid)None76.252.15.237,85150,00075.7
500 R-W77.656.72.336,11133,28772.2
2500 R-W78.860.22.434,50726,11069.0
500 R-E84.773.03.622,7729,57045.5
2500 R-E90.183.97.910,5612,25821.1
LanguageBind [58] (Softmax)None80.754.710.539,24350,00078.5
500 R-W84.764.913.136,59628,33773.2
2500 R-W85.468.211.734,54021,23069.1
500 R-E80.854.910.839,22649,64678.5
2500 R-E80.955.110.939,21049,03278.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.703 + ], + "angle": 0, + "content": "see Tab. 3. For example, ImageBing using 2,500 random word negative queries effectively halves the open-set error, but this occurs in trade-off with a closed-set accuracy reduction of \\(8.4\\%\\) (approximately 4,000 of the closed-set images are incorrectly classified as one of the random words)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While negative queries noticeably reduce the raw open-set error count, this is not always mirrored with increased performance in the open-set uncertainty metrics. We present the change in closed-set and open-set performance against the number of negative query points in Fig. 2. Generally the use of negative random words appears the most effective, with CoCa, ImageBind, SigLIP and Language-Bind showing improved open-set performance with minimal loss of closed-set performance. For some of the classifiers, there is negligible change in open-set performance despite the reduction in open-set error. We explore this further in the Supplementary Sec. S8, where we show that in some VLMs the negative" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.775, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "Fig. 2: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM classifiers." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.385, + 0.775, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.571, + 0.785, + 0.599 + ], + "angle": 0, + "content": "Fig. 3: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM object detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.784, + 0.671 + ], + "angle": 0, + "content": "queries only capture and remove \"easy\" open-set errors, i.e. open-set errors that already have high uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Object Detection: The VLM detectors do not show a consistent response to the use of negative queries. Different detectors show different sensitivities to negative query counts, different trade-off rates between closed-set and open-set performance, and differences in effectiveness of embedding versus word negative query types - see Tab. 4 and Fig. 3. Looking at Fig. 3, we can consider the sensitivity, trade-off rate and response of the detectors to negative queries: greater performance changes with increasing queries indicates sensitivity; large decreases in mAP with more queries indicates increased trade-off rates between open-set and closed-set performance; and a positive response to the negative queries presents as an increase in AuPR when increasing queries (negative response decreases AuPR)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: The open-set performance of state-of-the-art open vocabulary object detectors (with ResNet backbone indicated) when additional negative embeddings are included, tested on COCO. Arrows indicate the direction of optimal performance. Negative embeddings introduced as words are highlighted gray." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.213, + 0.77, + 0.695 + ], + "angle": 0, + "content": "
DetectorNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]Zero-Emb75.413.042.913,544190,14634.7
+5 Rand-Embs75.613.143.113,543186,12034.7
+100 Rand-Embs78.318.245.613,129109,26734.0
ViLD (R152) [15]“background”60.17.515.416,0111,485,60046.4
+5 Rand-Words57.56.013.415,9261,400,61246.6
+100 Rand-Words55.02.710.315,7721,088,72547.3
+5 Rand-Embs60.17.515.216,0111,485,23146.4
+100 Rand-Embs60.27.615.216,0191,478,80646.4
OV-DETR (R50) [48]None75.85.947.315,8181,485,60043.0
+5 Rand-Words77.28.148.315,510923,39843.0
+100 Rand-Words84.727.055.713,433239,11840.1
+5 Rand-Embs76.66.647.815,6521,178,82743.0
+100 Rand-Embs81.018.250.914,699356,38042.6
RegionCLIP (R50) [54]Zero-Emb73.19.939.315,752493,94944.6
+5 Rand-Words73.811.139.715,662322,59644.6
+100 Rand-Words77.017.543.115,178154,26244.0
+5 Rand-Embs91.877.244.68,8285,06227.0
+100 Rand-Embs99.999.383.628721.6
Detic (R50) [55]Zero-Emb72.68.142.214,670495,20039.1
+5 Rand-Words72.88.542.414,637451,63339.1
+100 Rand-Words74.511.543.514,287265,37338.8
+5 Rand-Embs93.474.561.810,0706,34630.6
+100 Rand-Embs100.0100.0100.037100.7
VLDet (R50) [22]Zero-Emb79.821.948.414,751112,95440.6
+5 Rand-Words80.122.348.714,672111,89340.5
+100 Rand-Words86.536.758.411,13250,12233.2
+5 Rand-Embs97.391.384.98,4641,38225.8
+100 Rand-Embs99.099.33.72,531225.9
CORA (R50) [44]Zero-Emb67.28.124.913,745495,20031.5
+5 Rand-Words66.37.627.013,700413,63631.6
+100 Rand-Words69.110.128.613,722329,39330.1
+5 Rand-Embs63.16.221.813,689495,20031.5
+100 Rand-Embs59.96.016.713,680495,17430.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.842 + ], + "angle": 0, + "content": "OV-DETR is the only detector to show a clear positive response, low trade-off rate, and high sensitivity to the negative query approach, particularly with negative embeddings. While OVR-CNN (embeddings), RegionCLIP (words), Detric (words) also show a positive response and low trade-off rate, this is with very little sensitivity to increasing numbers of negative queries. In contrast, the open-set performance of Detric, RegionCLIP, and VLDet are very sensitive to negative" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.147, + 0.497, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.293, + 0.469, + 0.306 + ], + "angle": 0, + "content": "(a) Classification task." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.147, + 0.695, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.293, + 0.675, + 0.306 + ], + "angle": 0, + "content": "(b) Object Detection task." + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.317, + 0.766, + 0.332 + ], + "angle": 0, + "content": "Fig. 4: The relationship between closed-set and open-set performance in VLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.784, + 0.39 + ], + "angle": 0, + "content": "embeddings, but at a rapid trade-off with closed-set performance. Both ViLD and CORA show a negative response or no sensitivity to negative queries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.542 + ], + "angle": 0, + "content": "It is tempting to relate these trends to the architecture or training paradigm of the VLM detectors, however we could not identify many common characteristics. RegionCLIP, Detic and VLDet all learn from auxiliary datasets that have been pseudo-labelled with regions [22, 54, 55], although this does not clearly indicate a reason for high sensitivity and trade-off rate for negative embeddings. Vild [15] and CORA [44] both heavily rely on CLIP's visual feature encoding (either by using CLIP directly [44] or knowledge distillation [15]) - their poor response is consistent with CLIP in the classification experiments on negative queries. Yet OV-DETR also uses CLIP's visual features during training to condition the object proposals [48], but exhibits the best response to negative queries." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.562, + 0.784, + 0.577 + ], + "angle": 0, + "content": "5.3 Is a good classifier all you need for VLM Open-set Recognition?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.785, + 0.645 + ], + "angle": 0, + "content": "In traditional open-set recognition, Vaze et. al [40] identified a strong positive correlation between closed-set and open-set performance. Testing this theory for VLM open-set recognition, we compare classifier and object detector closed-set and open-set performance in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.785, + 0.72 + ], + "angle": 0, + "content": "Classification: The correlation observed by [40] appears to hold for VLM classifiers, with greater accuracy (closed-set performance) relating to greater AuPR scores (open-set performance). It is worth noting that AuPR is a summary metric and this behaviour may not hold when examining performance at different operating points on the precision-recall curves." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.721, + 0.785, + 0.766 + ], + "angle": 0, + "content": "Object Detection: There is not a clear correlation between a VLM object detector's closed-set and open-set performance. ViLD in particular challenges this correlation, with the greatest mAP yet worst AuPR of all the tested detectors." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.787, + 0.657, + 0.803 + ], + "angle": 0, + "content": "5.4 How Does Query Set Size Impact Performance?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In Fig. 5, we test performance of the VLM classifiers on ImageNet while progressively increasing the query set size. We select the additional classes from the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.127 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.147, + 0.457, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.147, + 0.698, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.784, + 0.292 + ], + "angle": 0, + "content": "Fig. 5: Influence of the query set size on closed-set and open-set performance on ImageNet1k classification, showing mean and standard deviation from 10 random seeds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.785, + 0.412 + ], + "angle": 0, + "content": "WordNet hierarchy by choosing leaf nodes that are related, but not too closely related, to existing ImageNet classes (i.e. share a grandparent in the hierarchy but do not share a direct parent). Increasing the number of target classes and thus the query set size decrements both the closed-set and open-set performance of a VLM classifier. This indicates that testing VLMs with an entire vocabulary of nouns may not be a feasible approach to avoid open-set recognition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.355, + 0.449 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.524 + ], + "angle": 0, + "content": "Open Challenges: The tested baseline uncertainty measures were ineffective at identifying open-set errors, indicating a need to research better alternatives. While uncertainty representations for VLMs have been proposed [18, 39], their utility for open-set recognition remains unexplored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.785, + 0.644 + ], + "angle": 0, + "content": "While the use of negative random embeddings was able to reduce open-set error, large numbers were typically required to have any effect and this was often as the cost of falsely captured true positive predictions. Interestingly, some negative embeddings were much more effective at capturing open-set errors than others (e.g. the random string \"awy\" tested with ImageBind, see the Supplementary Sec. S9). Identifying the characteristics of these embeddings, learning them through prompt-tuning [8, 11, 12, 21, 57] or introducing auto-labelling techniques [49] are promising directions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Conclusions: VLMs are increasingly applied in diverse real-world applications, from robotics to medical image analysis. We hope that by highlighting their weaknesses, we encourage further research to improve these models and enhance their positive societal impact. We have clearly demonstrated that VLMs for open-vocabulary image classification and object detection introduce closed-set assumptions by relying on a finite query set. We thus refuted the assumption that vision-language models (VLMs) are inherently open-set models due to their training on extensive internet-scale datasets. While the investigated baseline approaches to identify open-set errors showed promise, we found none of them sufficiently robust and reliable for safety-critical applications. Our new benchmark and evaluation protocol is designed to foster and support this research. Understanding VLMs' limitations ultimately reduces risks, improves transparency, and ensures safer deployment in critical applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.235 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.786, + 0.264 + ], + "angle": 0, + "content": "2. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., Mané, D.: Concrete problems in ai safety. arXiv preprint arXiv:1606.06565 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.786, + 0.307 + ], + "angle": 0, + "content": "3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.786, + 0.335 + ], + "angle": 0, + "content": "4. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.786, + 0.39 + ], + "angle": 0, + "content": "5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 558-573. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.786, + 0.433 + ], + "angle": 0, + "content": "6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.786, + 0.475 + ], + "angle": 0, + "content": "7. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.786, + 0.517 + ], + "angle": 0, + "content": "8. Cho, E., Kim, J., Kim, H.J.: Distribution-aware prompt tuning for vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22004-22013 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.786, + 0.56 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.786, + 0.602 + ], + "angle": 0, + "content": "0. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: The IEEE Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.786, + 0.657 + ], + "angle": 0, + "content": "1. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14084-14093 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.786, + 0.701 + ], + "angle": 0, + "content": "2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.742 + ], + "angle": 0, + "content": "3. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Han, J., Ren, Y., Ding, J., Pan, X., Yan, K., Xia, G.S.: Expanding low-density latent regions for open-set object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9591-9600 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.229 + ], + "angle": 0, + "content": "17. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.229, + 0.787, + 0.284 + ], + "angle": 0, + "content": "18. Ji, Y., Wang, J., Gong, Y., Zhang, L., Zhu, Y., Wang, H., Zhang, J., Sakai, T., Yang, Y.: Map: Multimodal uncertainty-aware vision-language pre-training model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23262-23271 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.284, + 0.787, + 0.366 + ], + "angle": 0, + "content": "19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.407 + ], + "angle": 0, + "content": "20. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Li, H., Zhang, R., Yao, H., Song, X., Hao, Y., Zhao, Y., Li, L., Chen, Y.: Learning domain-aware detection head with prompt tuning. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "22. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.543 + ], + "angle": 0, + "content": "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.787, + 0.583 + ], + "angle": 0, + "content": "24. Liu, Y.C., Ma, C.Y., Dai, X., Tian, J., Vajda, P., He, Z., Kira, Z.: Open-set semi-supervised object detection. In: European Conference on Computer Vision. pp. 143-159. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.638 + ], + "angle": 0, + "content": "25. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.787, + 0.678 + ], + "angle": 0, + "content": "26. Miller, D., Nicholson, L., Dayoub, F., Sunderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "27. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. IEEE Robotics and Automation Letters 7(1), 215-222 (2022). https://doi.org/10.1109/LRA.2021.3123374" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "28. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.786 + ], + "angle": 0, + "content": "29. Panareda Busto, P., Gall, J.: Open set domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 754-763 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Perera, P., Morariu, V.I., Jain, R., Manjunatha, V., Wigington, C., Ordonez, V., Patel, V.M.: Generative-discriminative feature representations for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11814-11823 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Set Recognition in the Age of Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Pham, C., Vu, T., Nguyen, K.: Lp-ovod: Open-vocabulary object detection by linear probing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 779–788 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "32. Pham, T., Do, T.T., Carneiro, G., Reid, I., et al.: Bayesian semantic instance segmentation in open set world. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 3-18 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.311 + ], + "angle": 0, + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.353 + ], + "angle": 0, + "content": "34. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.393 + ], + "angle": 0, + "content": "35. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-168 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.434 + ], + "angle": 0, + "content": "36. Saito, T., Rehmsmeier, M.: The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PloS one 10(3), e0118432 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.473 + ], + "angle": 0, + "content": "37. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(7), 1757-1772 (2013). https://doi.org/10.1109/TPAMI.2012.256" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.528 + ], + "angle": 0, + "content": "38. Sünderhauf, N., Brock, O., Scheirer, W., Hadsell, R., Fox, D., Leitner, J., Upcroft, B., Abbeel, P., Burgard, W., Milford, M., et al.: The limits and potentials of deep learning for robotics. The International Journal of Robotics Research 37(4-5), 405-420 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.57 + ], + "angle": 0, + "content": "39. Upadhyay, U., Karthik, S., Mancini, M., Akata, Z.: Probvlm: Probabilistic adapter for frozen vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1899-1910 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.609 + ], + "angle": 0, + "content": "40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.651 + ], + "angle": 0, + "content": "41. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6759-6769 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.691 + ], + "angle": 0, + "content": "42. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards Open Vocabulary Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.732 + ], + "angle": 0, + "content": "43. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.785 + ], + "angle": 0, + "content": "44. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.84 + ], + "angle": 0, + "content": "45. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "D. Miller et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "46. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022), https://openreview.net/forum?id=Ee277P3AYC" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "47. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "48. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: European Conference on Computer Vision. pp. 106-122. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "49. Zara, G., Roy, S., Rota, P., Ricci, E.: Autolabel: Clip-based framework for open-set video domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11504-11513 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "50. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.411 + ], + "angle": 0, + "content": "51. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.439 + ], + "angle": 0, + "content": "52. Zhai, X., et al.: Sigmoid loss for language image pre-training. In: Intl. Conf. on Computer Vision. pp. 11975-11986 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "53. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 102-117. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.536 + ], + "angle": 0, + "content": "54. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.578 + ], + "angle": 0, + "content": "55. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "56. Zhou, Z., Yang, Y., Wang, Y., Xiong, R.: Open-set object detection using classification-free object proposal and instance-level contrastive learning. IEEE Robotics and Automation Letters 8(3), 1691-1698 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.716 + ], + "angle": 0, + "content": "58. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: Intl. Conf. on Learning Representations (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.716 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_origin.pdf b/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a8f5e76878a9218eb800c5d83fc04b5731b62c82 --- /dev/null +++ b/2024/Open-Set Recognition in the Age of Vision-Language Models/7fef71b5-35c7-46f2-bce9-ab91b5f839b4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c92adacf4e6f50e80125ed87bbdaef862cea4cd47c2da4fd51bd20d6e995e208 +size 705279 diff --git a/2024/Open-Set Recognition in the Age of Vision-Language Models/full.md b/2024/Open-Set Recognition in the Age of Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7b6523595ff8e31353be88e06d895d4f3f0ed34 --- /dev/null +++ b/2024/Open-Set Recognition in the Age of Vision-Language Models/full.md @@ -0,0 +1,256 @@ +# Open-Set Recognition in the Age of Vision-Language Models + +Dimity Miller, Niko Sünderhauf, Alex Kenna, and Keita Mason + +Queensland University of Technology, Brisbane, Australia\*\* {d24.miller, nikoo.suenderhauf}@qut.edu.au + +Abstract. Are vision-language models (VLMs) for open-vocabulary perception inherently open-set models because they are trained on internet-scale datasets? We answer this question with a clear no - VLMs introduce closed-set assumptions via their finite query set, making them vulnerable to open-set conditions. We systematically evaluate VLMs for open-set recognition and find they frequently misclassify objects not contained in their query set, leading to alarmingly low precision when tuned for high recall and vice versa. We show that naively increasing the size of the query set to contain more and more classes does not mitigate this problem, but instead causes diminishing task performance and open-set performance. We establish a revised definition of the open-set problem for the age of VLMs, define a new benchmark and evaluation protocol to facilitate standardised evaluation and research in this important area, and evaluate promising baseline approaches based on predictive uncertainty and dedicated negative embeddings on a range of open-vocabulary VLM classifiers and object detectors. + +# 1 Introduction + +In 2013, Scheirer et al. [37] described the closed-set assumption ingrained in almost all vision models at the time: all test classes are known during training and the model is never exposed to novel or unknown classes during testing. Real-world applications challenge this assumption and test in open-set conditions - encountering unexpected objects not in the training dataset - and causing potentially dangerous overconfident misclassifications [4,10,26,38]. Open-set recognition [37] was introduced to address this issue by evaluating the ability of vision models to identify and reject open-set inputs as unknown. + +Vision-language models (VLMs) have recently revolutionised the fields of image classification [14, 19, 33] and object detection [15, 48, 50]. Trained on vast and diverse internet-scale datasets, VLMs recognise an extensive variety of classes and seemingly are open-set by default [42]. Our paper challenges this view. + +In the age of VLMs, we introduce an updated definition of the open-set problem - while previous definitions focused on a finite training set [37], we show that + +![](images/c2d187fdd875a836e0a2ce850aefe17c3593dabaa0f57cb9ec0655ef931791ba.jpg) +Fig. 1: Left: Traditional open-set recognition arises when testing classes outside a model's finite training class set. While VLMs are trained on internet-scale datasets containing most conceivable object classes, they use a finite query set for classification. Open-set recognition arises when test classes are not included in the query set. Right: When testing an object present in the predefined query set, VLMs often correctly classify the object. When testing an object that is not present in the query set (i.e. an open-set object), VLMs often misclassify the object as a query class with high confidence (i.e. an open-set error). + +![](images/3acd71d845b28255b00d1b39ed65607a18bfa03b3ed61cb8b2ec5be10371755f.jpg) + +VLMs impose closed-set assumptions through a finite query set. As illustrated in Fig. 1, VLMs for open-vocabulary image classification or object detection are still evaluated under closed-set conditions: they compare image embeddings with text embeddings from a predefined, finite query set of class labels. This introduces a closed-set assumption that all classes encountered during testing are included in this query set. Open-set conditions emerge when VLMs encounter objects that are not included in the query set (unknown objects). As we show in Section 5, even state-of-the-art VLMs heavily degrade in performance and misclassify unknown objects as belonging to the query set with high confidence. + +In theory, this problem disappears if a VLM's query set contains every possible class label in a vocabulary. However, we show that naively increasing the query set to contain more and more class labels results in worse task performance, where increasing numbers of misclassifications occur (see Sec. 5.4). The serious concerns about deploying vision models in safety-critical applications voiced in the pre-VLM era [2,38] continues to be relevant and underscore the urgent need for further research into the open-set problem for VLMs. + +Our paper is the first to systematically investigate and quantify the open-set vulnerabilities of VLM-based open-vocabulary object detectors and zero-shot image classifiers. We re-define the open-set problem for VLMs (Section 3) and compare promising baseline strategies, leveraging predictive uncertainty (Section 5.1) and different approaches to inserting dedicated negative (unknown) class representations into the VLM pipeline in Section 5.2. We propose a new + +benchmark and evaluation protocol in Section 4 to foster further research and evaluation of VLMs for open-set recognition. + +# 2 Background + +# 2.1 Vision-Language Models for Zero-shot Classification + +A recently emerged foundation model is the vision-language model (VLM), which learns a multi-modal feature space that can jointly represent visual features and text [1, 19, 33, 46, 51] (and in some cases other modalities as well [14, 47]). Large-scale contrastive, self-supervised learning on internet-scale datasets provides VLMs with a rich feature representation that can be adapted for many downstream tasks, one of which is zero-shot or open vocabulary image classification [1, 14, 19, 33, 46, 47, 51]. Despite not being trained for classification tasks or datasets, VLMs generalise surprisingly well and achieve near state-of-the-art supervised learning performance [46]. Alongside this, a distinguishing advantage of VLMs is their adaptability to different datasets with different classes - the set of classes can be changed at will depending on the dataset [19, 33]. + +# 2.2 Vision-Language Models for Open Vocabulary Object Detection + +Unlocked by the advances in VLMs for zero-shot classification, [50] proposed open vocabulary object detection, where detectors must generalise to an arbitrary set of object classes at test time, even object classes that were not seen during training [50]. This is related to our proposed open-set recognition, but assumes that novel object classes are added to the query set, whereas we test against object classes that are not present in the query set. + +Existing open vocabulary detectors leverage VLM backbones to enable generalisation to new object classes, embedded within Faster R-CNN-style architectures [11, 12, 15, 20, 22, 31, 41, 43, 50, 54, 55] or DETR-style architectures [44, 48]. CLIP [33] is frequently utilised by these open vocabulary detectors, directly included in the architecture via its text encoder [11, 12, 15, 22, 31, 41, 43, 48, 54, 55] or in its entirety [20, 44], or used as a supervision signal during training [11, 12, 15, 31, 41, 43, 48, 54]. We refer the reader to [42] for a comprehensive survey on open vocabulary learning. + +# 2.3 Open-set Recognition + +Scheirer et al. [37] introduced open-set recognition to challenge the prevailing closed-set assumptions of image recognition systems. Closed-set models only test on samples from a predefined set of classes that were known during model training [37]. In open-set recognition, samples can instead come from an expansive range of classes that cannot be predefined, and therefore are unknown during model training [37]. Open-set recognition tests the ability of models to identify and reject unknown samples that do not belong to one of the training classes. + +The open-set concept has been explored widely in computer vision, including in image classification [4, 7, 28, 30, 40, 45, 53], object detection [10, 16, 26, 27, 56], image segmentation [6, 17, 32] and other vision tasks [3, 24, 25, 29, 35]. + +Our paper is the first to systematically investigate the open-set problem for VLMs, and our findings challenge the assumption that the open-set problem has been solved for VLMs due to their exposure to internet-scale training data. The following sections define the closed-set assumptions existing in current VLMs, demonstrate that VLMs continue to be vulnerable to open-set conditions and evaluate a number of baseline strategies for open-set recognition with VLMs. + +# 3 Problem Definition + +The core concept behind VLMs, mapping images and text into a joint embedding space, makes them powerful foundations for image classification and object detection. Classification is performed by comparing an image embedding $\mathbf{e}^I$ with text embeddings $\mathcal{E}^q = \{\mathbf{e}_1^q,\dots ,\mathbf{e}_N^q\}$ obtained from a set of query labels $\mathcal{Q} = \{q_1,\ldots ,q_N\}$ which are class names such as "dog". The similarities between the image embedding and all text embeddings from $\mathcal{Q}$ are measured by a similarity function $\mathbf{s}^q = \mathrm{sim}(\mathbf{e}^I,\mathcal{E}^q)$ , e.g. using the cosine similarity. The text embedding with the highest similarity to the image embedding, i.e. $\arg \max \mathbf{s}^q$ , determines the predicted class label. The process is the same for object detection, with $\mathbf{e}^I$ now representing the embedding of an image region corresponding to an object proposal. + +Open-set and Closed-set: While the pre-VLM definition of the closed-set and open-set concepts focused on the challenges arising from a finite training set [37], our updated definition emphasises the limitations of the finite query set: The closed-set assumption in VLMs is that all classes seen during testing are predefined prior to testing and included in the query set. An open-set situation arises when the VLM encounters an object or image class that is not part of this predefined query set $\mathcal{Q}$ . + +# 3.1 Baseline Approaches to Open-Set Recognition for VLMs + +In an open-set situation, the VLM should be able to identify an input not represented by the query set and reject it as unknown, instead of misclassifying it as one of the query classes from $\mathcal{Q}$ . Similar to the pre-VLM open-set literature, there are multiple ways of recognising an input as unknown [13]. We investigate baseline methods using predictive uncertainty and different approaches of negative embeddings. + +Using Uncertainty: Different measures of predictive uncertainty $\psi$ can be used to reject inputs as unknown, for example based on the class similarities, e.g. + +$\psi^{\mathrm{Cosine}} = \max (\mathbf{s}^q)$ or $\psi^{\mathrm{Softmax}} = \max (\mathrm{softmax}(\mathbf{s}^q))$ , or the entropy of the soft-max similarities $\psi^{\mathrm{Entropy}} = -H(\mathrm{softmax}(\mathbf{s}^q))$ . Only inputs that pass a threshold test $\psi \geq \theta$ are classified according to $\arg \max \mathbf{s}^q$ . Otherwise, the input is declared to be open-set and unknown. + +Dedicated Negative Embeddings: Another approach is to create a set of augmented query embeddings $\tilde{\mathcal{E}}^q = \mathcal{E}^q \cup \mathcal{E}^n$ by adding $M$ dedicated negative class embeddings $\mathcal{E}^n = \{\mathbf{e}_1^n, \ldots, \mathbf{e}_M^n\}$ . The input image or detection can then be rejected as unknown if the image embedding is most similar to one of those negative embeddings $\mathbf{e}_i^n$ rather than a query class embedding $\mathbf{e}_j^q$ . These negative embeddings can be created in one of two ways: (1) by creating negative words, which are then passed through the VLM text encoder to create the negative embeddings, or (2) by directly creating negative embeddings. + +We investigate two baseline options for negative embeddings: a random words method, where we generate random strings with random length between 2-8 characters (e.g. "braxl"); and a random embeddings method, where we randomly draw $M$ embedding vectors from a Gaussian distribution, i.e. $\mathbf{e}_i^{\mathrm{n}}\sim \mathcal{N}(\boldsymbol {\mu},\boldsymbol {\Sigma})$ , using the mean and standard deviation from the query embedding vectors $\mathcal{E}^q$ . We compare the effectiveness of these approaches, including for different numbers of $M$ random negative queries, in our experiments. Interestingly, many open vocabulary object detectors already use a negative query approach, though typically with a only single embedding vector of all zeroes [22,44,50,54,55]. + +# 4 Evaluation Protocol + +# 4.1 Creating an Open-set Recognition Dataset for VLMs + +We now introduce how to use existing image classification or object detection datasets to evaluate open-set recognition. In short, we test all images twice: once in a standard closed-set manner with all dataset classes included in the query set (see Fig. 1 top-right), and once in an open-set manner with a query set containing all dataset classes not present in the image (see Fig. 1 bottom-right). + +Classification and object detection datasets have a pre-defined set of $K$ labelled target classes $\mathcal{L}$ . In image classification, there is a single corresponding class label $y \in \mathcal{L}$ per image. In object detection, there are multiple ground truth objects in each image and each has a corresponding class label. For both tasks, we can define a set $\mathcal{V}$ that contains the ground truth class labels for each image: $\mathcal{V}$ contains a single label in classification and multiple labels in object detection. + +Testing Closed-set Recognition: We first test images in a standard manner to establish the closed-set performance of the VLM and identify all true positive $(TP)$ predictions. This involves testing all images with a query set identical to the dataset class list, $\mathcal{Q} = \mathcal{L}$ . To identify the TP predictions, we follow standard protocol for classification and object detection datasets - in classification, a TP has a predicted class matching the ground-truth label; in object detection, this involves a object-detection assignment process that requires correct object + +classification and localisation (see Supplementary Sec. S1). TPs represent correct closed-set predictions that should not be rejected as unknown, and we collect the uncertainty measures $\psi$ associated with all TPs for later evaluation. + +For object detection datasets, unlabelled objects are present in the background of the image [10]. These unlabelled objects are not in the query set $\mathcal{Q}$ , and therefore meet our definition for open-set objects. Detections of these open-set objects will present as false positives (FPs) in a closed-set test, however not all FPs arise from open-set objects - FPs in object detection can also arise due to poor object localisation, misclassification between query classes, or duplicate detections of a query class (see [5]). Rather than attempting to categorise different FP types to distinguish open-set errors, our approach definitively identifies open-set errors for both object detection and image classification. + +Testing Open-set Recognition: To evaluate open-set recognition, we test each image in the dataset with a modified query set, $\tilde{\mathcal{Q}}$ that contains all the class labels not in the image ground-truth class list, $\tilde{\mathcal{Q}} = \mathcal{L} - \mathcal{Y}$ . This ensures the image only contains object classes that are not in the modified query set - every single prediction must be an open-set error $(OSE)^2$ . In image classification, this modifies the query set to have $K - 1$ labels. In object detection, the modified query set size depends on the number of unique object classes in the image. We collect the uncertainty measures $\psi$ of all OSE predictions for evaluation. + +# 4.2 Metrics + +We use performance metrics that measure the ability of the VLM to reject OSEs (with high uncertainty or with a negative class), while maintaining the closed-set task performance. + +Precision-Recall Curves: Open-set recognition can be formulated as a binary classification task that uses uncertainty as the decision threshold. Correct closed-set predictions (introduced as TPs in Sec. 4.1) and open-set errors (OSEs) are considered as the positive and negative class respectively. We construct precision-recall curves, which are well-suited to this problem as they are affected by the size of the negative class (OSEs) $[36]^3$ as well as prediction uncertainty. High precision indicates that little-to-none negatives (OSEs) remain after the uncertainty thresholding. High recall indicates that most positives (TPs) remain after the uncertainty thresholding. We summarise the performance of the precision-recall curves with 3 core metrics: (1) Area under the PR curve (AuPR), (2) Precision at $95\%$ Recall, and (3) Recall at $95\%$ Precision. + +Task Metrics: We use the established task metrics used to evaluate classification and detection - top-1 accuracy [34] and mean Average Precision (mAP) [23] + +respectively. We evaluate mAP at an IoU threshold of 0.5 - we are primarily interested in the classification ability of the detector (rather than localisation) and this is standard protocol in open vocabulary detection [15,22,44,48,50,54,55]. Absolute Counts: We also report the absolute counts of TPs and OSEs. This can be particularly useful for object detection, where increased task performance does not necessarily indicate increased TPs and detectors predict variable numbers of OSEs. We stress that absolute counts should not be considered alone as a performance indicator, e.g. an increased number of OSEs can be acceptable if all OSEs have very high uncertainty. + +# 4.3 VLM Baselines + +We test with six VLM classifiers - CLIP [33], ALIGN [19], CoCa [46], Image-Bind [14], SigLIP [52] and LanguageBind [58] - and seven open vocabulary object detectors - OVR-CNN [50], ViLD [15], OV-DETR [48], RegionCLIP [54], Detric [55], VLDet [22] and CORA [44] (see Supplementary Sec. S2 for implementation details). These VLMs cover a range of architectures and training paradigms and have publicly-available implementations. Our goal with testing this range of VLMs is not to find the "best" VLM for open-set recognition, but instead to exhibit the general vulnerability of VLMs to open-set conditions. For each VLM, we extract the prediction Softmax score, Cosine similarity, and Entropy of the softmax distribution as baseline uncertainty measures. Some VLMs use Sigmoid activation rather than Softmax [22,44,48,52,55] - for these methods, we only report the Sigmoid score as its results are identical to the cosine similarity and the entropy is unsuitable as the scores are not normalised to a distribution. We do not impose a minimum score threshold on any predictions, as the AuPR metric captures performance over all possible thresholds. + +# 4.4 Datasets + +We focus on general object recognition and test with the large-scale and prevailing datasets used to benchmark image classification and object detection: For image classification, we test with the ImageNet1k validation dataset [9, 34]. It contains 50,000 images from 1000 classes and is publicly available for research. For object detection, we test with the MS COCO 2017 validation dataset [23]. It contains 4,952 images with 36,781 annotated objects from 80 classes and is publicly available for research. We show results for domain-specific classification datasets in the Supplementary Sec. S4. + +# 5 Experiments and Results + +# 5.1 State-of-the-art VLMs Perform Poorly in Open-set Conditions. + +Classification: Tab. 1 shows that all six VLM classifiers perform poorly for open-set recognition. For scenarios that require high recall of true positive + +Table 1: The open-set performance of state-of-the-art VLM classifiers when tested on ImageNet1k. Arrows indicate direction of optimal performance and - indicates operating point could not be achieved. + +
ClassifierUncertaintyAuROCAuPRP@95RR@95PTPOSEAccuracy
↑↑↑↑↑↓↑
CLIP [33]Softmax79.771.546.23.4
Cosine72.261.342.50.131,02350,00062.1
Entropy80.272.345.42.8
ALIGN [19]Softmax81.074.748.18.3
Cosine72.462.144.8-32,61850,00065.2
Entropy80.975.046.88.6
CoCa [46]Softmax79.071.746.03.4
Cosine73.263.043.30.131,72550,00063.4
Entropy80.573.046.83.5
ImageBind [14]Softmax82.879.152.85.0
Cosine79.273.950.70.238,40550,00076.8
Entropy84.380.156.35.6
SigLIP [52]Sigmoid81.076.252.12.237,85150,00075.7
LanguageBind [58]Softmax83.980.754.710.5
Cosine81.677.452.80.839,24350,00078.5
Entropy85.481.758.411.1
+ +predictions (i.e. $95\%$ recall), approximately every second prediction returned is an open-set error (i.e. precision between $46.2\%$ and $58.4\%$ ). For scenarios that require high precision of predictions (i.e. $95\%$ precision), less than $12\%$ of the true positive predictions are retained (i.e. recall between $3.4\%$ and $11.1\%$ ). Comparing across uncertainty types, the VLM classifiers show better uncertainty performance when using the predicted Softmax score or Entropy rather than the predicted Cosine similarity. This suggests raw similarity in VLM feature space is not the best indicator of open-set error, with measures of relative similarity comparative to other classes offering better performance. + +Object Detection: All seven of the tested VLM object detectors are vulnerable to open-set errors - see Tab. 2. Unlike VLM classifiers, most VLM detectors already contain a negative query to capture objects not in the query set (often referred to as the background class). Despite this, the VLM detectors produce open-set error counts in the range of 100,000 to 1,500,000 when tested on only 4,952 images, i.e. between 20 and 300 open-set error per image on average. + +High open-set error counts are not inherently a problem if they are produced with high uncertainty. Yet for all tested detectors, the baseline uncertainty methods cannot adequately distinguish true positive detections from open-set errors. When thresholding for high recall of true positives, at best only every fifth detection is a true positive (i.e. precision ranges between $5.9\%$ and $21.9\%$ ). When + +Table 2: The open-set performance of state-of-the-art open-vocabulary object detectors (with ResNet backbone indicated), tested on COCO. Arrows indicate the direction of optimal performance. + +
DetectorUncertaintyNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]SoftmaxZero-Emb75.413.042.9
CosineZero-Emb77.215.343.613,544190,14634.7
EntropyZero-Emb63.27.543.4
ViLD (R152) [15]Softmax“background”60.17.515.4
Cosine“background”33.16.90.116,0111,485,60046.4
Entropy“background”62.28.416.3
OV-DETR (R50) [48]SigmoidNone75.85.947.315,8181,485,60043.0
RegionCLIP (R50) [54]SoftmaxZero-Emb73.19.939.4
CosineZero-Emb74.07.140.915,741493,94044.6
EntropyZero-Emb35.43.917.0
Detic (R50) [55]SigmoidZero-Emb72.68.142.214,670495,20039.1
VLDet (R50) [22]SigmoidZero-Emb79.821.948.514,751112,92340.6
CORA (R50) [44]SigmoidZero-Emb67.28.124.713,763495,20031.6
+ +thresholding for high precision, at best less than half of true positives are recalled (i.e. recall ranges between $16.3\%$ and $48.5\%$ ). We include some qualitative examples of highly confident open-set errors in the Supplementary Sec. S6. + +Notably, the precision-recall characteristics for VLM classification and object detection differ significantly – VLM classifiers achieve better performance at high recall, whereas VLM detectors achieve better performance at high precision. This suggests that while the VLM classifiers struggle with over-confident open-set errors, the VLM detectors suffer from under-confident true positives. We include histograms that visualise this trend in the Supplementary Sec. S7. + +We also consider how the performance of uncertainty differs for identifying open-set error versus closed-set misclassifications, with results and a detailed discussion in the Supplementary Sec. S5. We find that the performance of softmax uncertainty is lower for open-set error than for closed-set error in both classification and detection tasks. In classification, open-set performance is particularly decremented when thresholding for high precision - likely due to a greater proportion of overconfident open-set error - whereas detection primarily degrades with increased numbers of errors, with ViLD for example producing $11.7 \times$ more open-set than closed-set misclassifications. + +# 5.2 Can Negative Queries Improve Open-set Performance? + +Classification: Negative class queries provide the VLM classifiers with the ability to select "unknown" when testing an input. When used in larger quantities, negative queries can capture and remove significant portions of open-set error, yet this happens in trade-off with a reduction in closed-set performance - + +Table 3: The open-set performance of state-of-the-art VLM classifiers on ImageNet1k when used with negative random words (R-W) or random embeddings (R-E). + +
ClassifierNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓Accuracy ↑
CLIP [33] (Softmax)None71.546.23.431,02350,00062.1
500 R-W73.448.13.329,96141,42859.9
2500 R-W72.849.81.628,65735,09257.3
500 R-E71.947.22.530,16044,67360.3
2500 R-E71.747.10.929,29140,49758.6
ALIGN [19] (Softmax)None74.748.18.332,61850,00065.2
500 R-W75.348.910.431,73143,19863.5
2500 R-W74.649.58.930,58837,73861.2
500 R-E74.948.78.632,34247,22864.7
2500 R-E75.149.77.231,46241,21062.9
CoCa [46] (Softmax)None71.746.03.431,72550,00063.4
500 R-W75.449.60.130,27639,52860.6
2500 R-W75.950.96.029,10734,36058.2
500 R-E71.746.03.431,72349,96663.4
2500 R-E71.746.13.431,71949,86663.4
ImageBind [14] (Softmax)None79.152.85.038,40550,00076.8
500 R-W83.863.17.735,85929,17271.7
2500 R-W84.666.18.334,17922,87268.4
500 R-E81.057.39.036,96638,63673.9
2500 R-E81.759.011.235,95433,34971.9
SigLIP [52] (Sigmoid)None76.252.15.237,85150,00075.7
500 R-W77.656.72.336,11133,28772.2
2500 R-W78.860.22.434,50726,11069.0
500 R-E84.773.03.622,7729,57045.5
2500 R-E90.183.97.910,5612,25821.1
LanguageBind [58] (Softmax)None80.754.710.539,24350,00078.5
500 R-W84.764.913.136,59628,33773.2
2500 R-W85.468.211.734,54021,23069.1
500 R-E80.854.910.839,22649,64678.5
2500 R-E80.955.110.939,21049,03278.4
+ +see Tab. 3. For example, ImageBing using 2,500 random word negative queries effectively halves the open-set error, but this occurs in trade-off with a closed-set accuracy reduction of $8.4\%$ (approximately 4,000 of the closed-set images are incorrectly classified as one of the random words). + +While negative queries noticeably reduce the raw open-set error count, this is not always mirrored with increased performance in the open-set uncertainty metrics. We present the change in closed-set and open-set performance against the number of negative query points in Fig. 2. Generally the use of negative random words appears the most effective, with CoCa, ImageBind, SigLIP and Language-Bind showing improved open-set performance with minimal loss of closed-set performance. For some of the classifiers, there is negligible change in open-set performance despite the reduction in open-set error. We explore this further in the Supplementary Sec. S8, where we show that in some VLMs the negative + +![](images/1d6d2a49c30f0b30b0ddb8cc7229d817202099d78c738a68d062201a52532894.jpg) +Fig. 2: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM classifiers. + +![](images/b25352584581fd230f2ec6b9cb5cb48e7df50ae17f800868cba8f9bf92df34e3.jpg) +Fig. 3: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM object detectors. + +queries only capture and remove "easy" open-set errors, i.e. open-set errors that already have high uncertainty. + +Object Detection: The VLM detectors do not show a consistent response to the use of negative queries. Different detectors show different sensitivities to negative query counts, different trade-off rates between closed-set and open-set performance, and differences in effectiveness of embedding versus word negative query types - see Tab. 4 and Fig. 3. Looking at Fig. 3, we can consider the sensitivity, trade-off rate and response of the detectors to negative queries: greater performance changes with increasing queries indicates sensitivity; large decreases in mAP with more queries indicates increased trade-off rates between open-set and closed-set performance; and a positive response to the negative queries presents as an increase in AuPR when increasing queries (negative response decreases AuPR). + +Table 4: The open-set performance of state-of-the-art open vocabulary object detectors (with ResNet backbone indicated) when additional negative embeddings are included, tested on COCO. Arrows indicate the direction of optimal performance. Negative embeddings introduced as words are highlighted gray. + +
DetectorNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]Zero-Emb75.413.042.913,544190,14634.7
+5 Rand-Embs75.613.143.113,543186,12034.7
+100 Rand-Embs78.318.245.613,129109,26734.0
ViLD (R152) [15]“background”60.17.515.416,0111,485,60046.4
+5 Rand-Words57.56.013.415,9261,400,61246.6
+100 Rand-Words55.02.710.315,7721,088,72547.3
+5 Rand-Embs60.17.515.216,0111,485,23146.4
+100 Rand-Embs60.27.615.216,0191,478,80646.4
OV-DETR (R50) [48]None75.85.947.315,8181,485,60043.0
+5 Rand-Words77.28.148.315,510923,39843.0
+100 Rand-Words84.727.055.713,433239,11840.1
+5 Rand-Embs76.66.647.815,6521,178,82743.0
+100 Rand-Embs81.018.250.914,699356,38042.6
RegionCLIP (R50) [54]Zero-Emb73.19.939.315,752493,94944.6
+5 Rand-Words73.811.139.715,662322,59644.6
+100 Rand-Words77.017.543.115,178154,26244.0
+5 Rand-Embs91.877.244.68,8285,06227.0
+100 Rand-Embs99.999.383.628721.6
Detic (R50) [55]Zero-Emb72.68.142.214,670495,20039.1
+5 Rand-Words72.88.542.414,637451,63339.1
+100 Rand-Words74.511.543.514,287265,37338.8
+5 Rand-Embs93.474.561.810,0706,34630.6
+100 Rand-Embs100.0100.0100.037100.7
VLDet (R50) [22]Zero-Emb79.821.948.414,751112,95440.6
+5 Rand-Words80.122.348.714,672111,89340.5
+100 Rand-Words86.536.758.411,13250,12233.2
+5 Rand-Embs97.391.384.98,4641,38225.8
+100 Rand-Embs99.099.33.72,531225.9
CORA (R50) [44]Zero-Emb67.28.124.913,745495,20031.5
+5 Rand-Words66.37.627.013,700413,63631.6
+100 Rand-Words69.110.128.613,722329,39330.1
+5 Rand-Embs63.16.221.813,689495,20031.5
+100 Rand-Embs59.96.016.713,680495,17430.6
+ +OV-DETR is the only detector to show a clear positive response, low trade-off rate, and high sensitivity to the negative query approach, particularly with negative embeddings. While OVR-CNN (embeddings), RegionCLIP (words), Detric (words) also show a positive response and low trade-off rate, this is with very little sensitivity to increasing numbers of negative queries. In contrast, the open-set performance of Detric, RegionCLIP, and VLDet are very sensitive to negative + +![](images/3f1faff813e5f05cedd03c8050f31cdf1a9a20886cba024b9b3d2175f2460624.jpg) +(a) Classification task. + +![](images/a1aa75051a50cd3509a999c13d057de3e83f58c8d57e9d70281e624101b0a004.jpg) +(b) Object Detection task. +Fig. 4: The relationship between closed-set and open-set performance in VLMs. + +embeddings, but at a rapid trade-off with closed-set performance. Both ViLD and CORA show a negative response or no sensitivity to negative queries. + +It is tempting to relate these trends to the architecture or training paradigm of the VLM detectors, however we could not identify many common characteristics. RegionCLIP, Detic and VLDet all learn from auxiliary datasets that have been pseudo-labelled with regions [22, 54, 55], although this does not clearly indicate a reason for high sensitivity and trade-off rate for negative embeddings. Vild [15] and CORA [44] both heavily rely on CLIP's visual feature encoding (either by using CLIP directly [44] or knowledge distillation [15]) - their poor response is consistent with CLIP in the classification experiments on negative queries. Yet OV-DETR also uses CLIP's visual features during training to condition the object proposals [48], but exhibits the best response to negative queries. + +# 5.3 Is a good classifier all you need for VLM Open-set Recognition? + +In traditional open-set recognition, Vaze et. al [40] identified a strong positive correlation between closed-set and open-set performance. Testing this theory for VLM open-set recognition, we compare classifier and object detector closed-set and open-set performance in Fig. 4. + +Classification: The correlation observed by [40] appears to hold for VLM classifiers, with greater accuracy (closed-set performance) relating to greater AuPR scores (open-set performance). It is worth noting that AuPR is a summary metric and this behaviour may not hold when examining performance at different operating points on the precision-recall curves. + +Object Detection: There is not a clear correlation between a VLM object detector's closed-set and open-set performance. ViLD in particular challenges this correlation, with the greatest mAP yet worst AuPR of all the tested detectors. + +# 5.4 How Does Query Set Size Impact Performance? + +In Fig. 5, we test performance of the VLM classifiers on ImageNet while progressively increasing the query set size. We select the additional classes from the + +![](images/767a2cee1c481a09e25d58855181a77c904d0bd4bf3f18a8db1b1a4624d5ab31.jpg) +Fig. 5: Influence of the query set size on closed-set and open-set performance on ImageNet1k classification, showing mean and standard deviation from 10 random seeds. + +![](images/337bb0c7993599552ced58d7bfa22c1571e1cebb85b3304efa8ca74b15080da8.jpg) + +WordNet hierarchy by choosing leaf nodes that are related, but not too closely related, to existing ImageNet classes (i.e. share a grandparent in the hierarchy but do not share a direct parent). Increasing the number of target classes and thus the query set size decrements both the closed-set and open-set performance of a VLM classifier. This indicates that testing VLMs with an entire vocabulary of nouns may not be a feasible approach to avoid open-set recognition. + +# 6 Discussion + +Open Challenges: The tested baseline uncertainty measures were ineffective at identifying open-set errors, indicating a need to research better alternatives. While uncertainty representations for VLMs have been proposed [18, 39], their utility for open-set recognition remains unexplored. + +While the use of negative random embeddings was able to reduce open-set error, large numbers were typically required to have any effect and this was often as the cost of falsely captured true positive predictions. Interestingly, some negative embeddings were much more effective at capturing open-set errors than others (e.g. the random string "awy" tested with ImageBind, see the Supplementary Sec. S9). Identifying the characteristics of these embeddings, learning them through prompt-tuning [8, 11, 12, 21, 57] or introducing auto-labelling techniques [49] are promising directions. + +Conclusions: VLMs are increasingly applied in diverse real-world applications, from robotics to medical image analysis. We hope that by highlighting their weaknesses, we encourage further research to improve these models and enhance their positive societal impact. We have clearly demonstrated that VLMs for open-vocabulary image classification and object detection introduce closed-set assumptions by relying on a finite query set. We thus refuted the assumption that vision-language models (VLMs) are inherently open-set models due to their training on extensive internet-scale datasets. While the investigated baseline approaches to identify open-set errors showed promise, we found none of them sufficiently robust and reliable for safety-critical applications. Our new benchmark and evaluation protocol is designed to foster and support this research. Understanding VLMs' limitations ultimately reduces risks, improves transparency, and ensures safer deployment in critical applications. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +2. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., Mané, D.: Concrete problems in ai safety. arXiv preprint arXiv:1606.06565 (2016) +3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021) +4. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016) +5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 558-573. Springer (2020) +6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021) +7. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021) +8. Cho, E., Kim, J., Kim, H.J.: Distribution-aware prompt tuning for vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22004-22013 (2023) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +0. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: The IEEE Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020) +1. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14084-14093 (2022) +2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) +3. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020) +4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023) +5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022) + +16. Han, J., Ren, Y., Ding, J., Pan, X., Yan, K., Xia, G.S.: Expanding low-density latent regions for open-set object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9591-9600 (2022) +17. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021) +18. Ji, Y., Wang, J., Gong, Y., Zhang, L., Zhu, Y., Wang, H., Zhang, J., Sakai, T., Yang, Y.: Map: Multimodal uncertainty-aware vision-language pre-training model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23262-23271 (2023) +19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html +20. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. In: International Conference on Learning Representations (2023) +21. Li, H., Zhang, R., Yao, H., Song, X., Hao, Y., Zhao, Y., Li, L., Chen, Y.: Learning domain-aware detection head with prompt tuning. Advances in Neural Information Processing Systems 36 (2024) +22. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: International Conference on Learning Representations (2023) +23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +24. Liu, Y.C., Ma, C.Y., Dai, X., Tian, J., Vajda, P., He, Z., Kira, Z.: Open-set semi-supervised object detection. In: European Conference on Computer Vision. pp. 143-159. Springer (2022) +25. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024) +26. Miller, D., Nicholson, L., Dayoub, F., Sunderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249. IEEE (2018) +27. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. IEEE Robotics and Automation Letters 7(1), 215-222 (2022). https://doi.org/10.1109/LRA.2021.3123374 +28. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019) +29. Panareda Busto, P., Gall, J.: Open set domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 754-763 (2017) +30. Perera, P., Morariu, V.I., Jain, R., Manjunatha, V., Wigington, C., Ordonez, V., Patel, V.M.: Generative-discriminative feature representations for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11814-11823 (2020) + +31. Pham, C., Vu, T., Nguyen, K.: Lp-ovod: Open-vocabulary object detection by linear probing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 779–788 (2024) +32. Pham, T., Do, T.T., Carneiro, G., Reid, I., et al.: Bayesian semantic instance segmentation in open set world. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 3-18 (2018) +33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html +34. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) +35. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-168 (2018) +36. Saito, T., Rehmsmeier, M.: The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PloS one 10(3), e0118432 (2015) +37. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(7), 1757-1772 (2013). https://doi.org/10.1109/TPAMI.2012.256 +38. Sünderhauf, N., Brock, O., Scheirer, W., Hadsell, R., Fox, D., Leitner, J., Upcroft, B., Abbeel, P., Burgard, W., Milford, M., et al.: The limits and potentials of deep learning for robotics. The International Journal of Robotics Research 37(4-5), 405-420 (2018) +39. Upadhyay, U., Karthik, S., Mancini, M., Akata, Z.: Probvlm: Probabilistic adapter for frozen vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1899-1910 (2023) +40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022) +41. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6759-6769 (2023) +42. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards Open Vocabulary Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) +43. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023) +44. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023) +45. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019) + +46. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022), https://openreview.net/forum?id=Ee277P3AYC +47. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) +48. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: European Conference on Computer Vision. pp. 106-122. Springer (2022) +49. Zara, G., Roy, S., Rota, P., Ricci, E.: Autolabel: Clip-based framework for open-set video domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11504-11513 (2023) +50. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021) +51. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022) +52. Zhai, X., et al.: Sigmoid loss for language image pre-training. In: Intl. Conf. on Computer Vision. pp. 11975-11986 (2023) +53. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 102-117. Springer (2020) +54. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022) +55. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022) +56. Zhou, Z., Yang, Y., Wang, Y., Xiong, R.: Open-set object detection using classification-free object proposal and instance-level contrastive learning. IEEE Robotics and Automation Letters 8(3), 1691-1698 (2023) +57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023) +58. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: Intl. Conf. on Learning Representations (2023) \ No newline at end of file diff --git a/2024/Open-Set Recognition in the Age of Vision-Language Models/images.zip b/2024/Open-Set Recognition in the Age of Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a6bc61559c3531a67c211d92d6cc2632814680b4 --- /dev/null +++ b/2024/Open-Set Recognition in the Age of Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:562298f3ba847a2038d0a509e0e32eef6297bd542f35013234f934e8c16daa18 +size 633384 diff --git a/2024/Open-Set Recognition in the Age of Vision-Language Models/layout.json b/2024/Open-Set Recognition in the Age of Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e973c7ddf454ab27560eb21601aab77cf5622a2f --- /dev/null +++ b/2024/Open-Set Recognition in the Age of Vision-Language Models/layout.json @@ -0,0 +1,8062 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 182, + 111, + 432, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 111, + 432, + 147 + ], + "spans": [ + { + "bbox": [ + 182, + 111, + 432, + 147 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 168, + 445, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 168, + 445, + 180 + ], + "spans": [ + { + "bbox": [ + 166, + 168, + 445, + 180 + ], + "type": "text", + "content": "Dimity Miller, Niko Sünderhauf, Alex Kenna, and Keita Mason" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 190, + 427, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 190, + 427, + 212 + ], + "spans": [ + { + "bbox": [ + 184, + 190, + 427, + 212 + ], + "type": "text", + "content": "Queensland University of Technology, Brisbane, Australia\\*\\* {d24.miller, nikoo.suenderhauf}@qut.edu.au" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 452, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 452, + 418 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 452, + 418 + ], + "type": "text", + "content": "Abstract. Are vision-language models (VLMs) for open-vocabulary perception inherently open-set models because they are trained on internet-scale datasets? We answer this question with a clear no - VLMs introduce closed-set assumptions via their finite query set, making them vulnerable to open-set conditions. We systematically evaluate VLMs for open-set recognition and find they frequently misclassify objects not contained in their query set, leading to alarmingly low precision when tuned for high recall and vice versa. We show that naively increasing the size of the query set to contain more and more classes does not mitigate this problem, but instead causes diminishing task performance and open-set performance. We establish a revised definition of the open-set problem for the age of VLMs, define a new benchmark and evaluation protocol to facilitate standardised evaluation and research in this important area, and evaluate promising baseline approaches based on predictive uncertainty and dedicated negative embeddings on a range of open-vocabulary VLM classifiers and object detectors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 441, + 229, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 441, + 229, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 441, + 229, + 453 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 466, + 481, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 481, + 562 + ], + "type": "text", + "content": "In 2013, Scheirer et al. [37] described the closed-set assumption ingrained in almost all vision models at the time: all test classes are known during training and the model is never exposed to novel or unknown classes during testing. Real-world applications challenge this assumption and test in open-set conditions - encountering unexpected objects not in the training dataset - and causing potentially dangerous overconfident misclassifications [4,10,26,38]. Open-set recognition [37] was introduced to address this issue by evaluating the ability of vision models to identify and reject open-set inputs as unknown." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 562, + 481, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 481, + 610 + ], + "type": "text", + "content": "Vision-language models (VLMs) have recently revolutionised the fields of image classification [14, 19, 33] and object detection [15, 48, 50]. Trained on vast and diverse internet-scale datasets, VLMs recognise an extensive variety of classes and seemingly are open-set by default [42]. Our paper challenges this view." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "type": "text", + "content": "In the age of VLMs, we introduce an updated definition of the open-set problem - while previous definitions focused on a finite training set [37], we show that" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 131, + 643, + 476, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 643, + 476, + 653 + ], + "spans": [ + { + "bbox": [ + 131, + 643, + 476, + 653 + ], + "type": "text", + "content": "** The authors acknowledge ongoing support from the QUT Centre for Robotics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 653, + 475, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 475, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 475, + 665 + ], + "type": "text", + "content": "Evaluation code is publicly available at github.com/dimitymiller/openset_vlms" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 255, + 281 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 255, + 281 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 255, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 255, + 281 + ], + "type": "image", + "image_path": "c2d187fdd875a836e0a2ce850aefe17c3593dabaa0f57cb9ec0655ef931791ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 288, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 376 + ], + "type": "text", + "content": "Fig. 1: Left: Traditional open-set recognition arises when testing classes outside a model's finite training class set. While VLMs are trained on internet-scale datasets containing most conceivable object classes, they use a finite query set for classification. Open-set recognition arises when test classes are not included in the query set. Right: When testing an object present in the predefined query set, VLMs often correctly classify the object. When testing an object that is not present in the query set (i.e. an open-set object), VLMs often misclassify the object as a query class with high confidence (i.e. an open-set error)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 113, + 482, + 281 + ], + "blocks": [ + { + "bbox": [ + 258, + 113, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 258, + 113, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 258, + 113, + 482, + 281 + ], + "type": "image", + "image_path": "3acd71d845b28255b00d1b39ed65607a18bfa03b3ed61cb8b2ec5be10371755f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 509 + ], + "type": "text", + "content": "VLMs impose closed-set assumptions through a finite query set. As illustrated in Fig. 1, VLMs for open-vocabulary image classification or object detection are still evaluated under closed-set conditions: they compare image embeddings with text embeddings from a predefined, finite query set of class labels. This introduces a closed-set assumption that all classes encountered during testing are included in this query set. Open-set conditions emerge when VLMs encounter objects that are not included in the query set (unknown objects). As we show in Section 5, even state-of-the-art VLMs heavily degrade in performance and misclassify unknown objects as belonging to the query set with high confidence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": "In theory, this problem disappears if a VLM's query set contains every possible class label in a vocabulary. However, we show that naively increasing the query set to contain more and more class labels results in worse task performance, where increasing numbers of misclassifications occur (see Sec. 5.4). The serious concerns about deploying vision models in safety-critical applications voiced in the pre-VLM era [2,38] continues to be relevant and underscore the urgent need for further research into the open-set problem for VLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Our paper is the first to systematically investigate and quantify the open-set vulnerabilities of VLM-based open-vocabulary object detectors and zero-shot image classifiers. We re-define the open-set problem for VLMs (Section 3) and compare promising baseline strategies, leveraging predictive uncertainty (Section 5.1) and different approaches to inserting dedicated negative (unknown) class representations into the VLM pipeline in Section 5.2. We propose a new" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "benchmark and evaluation protocol in Section 4 to foster further research and evaluation of VLMs for open-set recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 159, + 226, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 159, + 226, + 172 + ], + "spans": [ + { + "bbox": [ + 131, + 159, + 226, + 172 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 185, + 423, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 185, + 423, + 197 + ], + "spans": [ + { + "bbox": [ + 131, + 185, + 423, + 197 + ], + "type": "text", + "content": "2.1 Vision-Language Models for Zero-shot Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 338 + ], + "type": "text", + "content": "A recently emerged foundation model is the vision-language model (VLM), which learns a multi-modal feature space that can jointly represent visual features and text [1, 19, 33, 46, 51] (and in some cases other modalities as well [14, 47]). Large-scale contrastive, self-supervised learning on internet-scale datasets provides VLMs with a rich feature representation that can be adapted for many downstream tasks, one of which is zero-shot or open vocabulary image classification [1, 14, 19, 33, 46, 47, 51]. Despite not being trained for classification tasks or datasets, VLMs generalise surprisingly well and achieve near state-of-the-art supervised learning performance [46]. Alongside this, a distinguishing advantage of VLMs is their adaptability to different datasets with different classes - the set of classes can be changed at will depending on the dataset [19, 33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 355, + 481, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 355, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 355, + 481, + 369 + ], + "type": "text", + "content": "2.2 Vision-Language Models for Open Vocabulary Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 375, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 447 + ], + "type": "text", + "content": "Unlocked by the advances in VLMs for zero-shot classification, [50] proposed open vocabulary object detection, where detectors must generalise to an arbitrary set of object classes at test time, even object classes that were not seen during training [50]. This is related to our proposed open-set recognition, but assumes that novel object classes are added to the query set, whereas we test against object classes that are not present in the query set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": "Existing open vocabulary detectors leverage VLM backbones to enable generalisation to new object classes, embedded within Faster R-CNN-style architectures [11, 12, 15, 20, 22, 31, 41, 43, 50, 54, 55] or DETR-style architectures [44, 48]. CLIP [33] is frequently utilised by these open vocabulary detectors, directly included in the architecture via its text encoder [11, 12, 15, 22, 31, 41, 43, 48, 54, 55] or in its entirety [20, 44], or used as a supervision signal during training [11, 12, 15, 31, 41, 43, 48, 54]. We refer the reader to [42] for a comprehensive survey on open vocabulary learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 562, + 270, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 270, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 270, + 574 + ], + "type": "text", + "content": "2.3 Open-set Recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Scheirer et al. [37] introduced open-set recognition to challenge the prevailing closed-set assumptions of image recognition systems. Closed-set models only test on samples from a predefined set of classes that were known during model training [37]. In open-set recognition, samples can instead come from an expansive range of classes that cannot be predefined, and therefore are unknown during model training [37]. Open-set recognition tests the ability of models to identify and reject unknown samples that do not belong to one of the training classes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "The open-set concept has been explored widely in computer vision, including in image classification [4, 7, 28, 30, 40, 45, 53], object detection [10, 16, 26, 27, 56], image segmentation [6, 17, 32] and other vision tasks [3, 24, 25, 29, 35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 224 + ], + "type": "text", + "content": "Our paper is the first to systematically investigate the open-set problem for VLMs, and our findings challenge the assumption that the open-set problem has been solved for VLMs due to their exposure to internet-scale training data. The following sections define the closed-set assumptions existing in current VLMs, demonstrate that VLMs continue to be vulnerable to open-set conditions and evaluate a number of baseline strategies for open-set recognition with VLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 244, + 269, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 244, + 269, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 244, + 269, + 257 + ], + "type": "text", + "content": "3 Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": "The core concept behind VLMs, mapping images and text into a joint embedding space, makes them powerful foundations for image classification and object detection. Classification is performed by comparing an image embedding " + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^I" + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": " with text embeddings " + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^q = \\{\\mathbf{e}_1^q,\\dots ,\\mathbf{e}_N^q\\}" + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": " obtained from a set of query labels " + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{q_1,\\ldots ,q_N\\}" + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": " which are class names such as \"dog\". The similarities between the image embedding and all text embeddings from " + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": " are measured by a similarity function " + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^q = \\mathrm{sim}(\\mathbf{e}^I,\\mathcal{E}^q)" + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": ", e.g. using the cosine similarity. The text embedding with the highest similarity to the image embedding, i.e. " + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\arg \\max \\mathbf{s}^q" + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": ", determines the predicted class label. The process is the same for object detection, with " + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^I" + }, + { + "bbox": [ + 130, + 271, + 482, + 402 + ], + "type": "text", + "content": " now representing the embedding of an image region corresponding to an object proposal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 486 + ], + "type": "text", + "content": "Open-set and Closed-set: While the pre-VLM definition of the closed-set and open-set concepts focused on the challenges arising from a finite training set [37], our updated definition emphasises the limitations of the finite query set: The closed-set assumption in VLMs is that all classes seen during testing are predefined prior to testing and included in the query set. An open-set situation arises when the VLM encounters an object or image class that is not part of this predefined query set " + }, + { + "bbox": [ + 130, + 403, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 403, + 482, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 506, + 446, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 446, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 446, + 517 + ], + "type": "text", + "content": "3.1 Baseline Approaches to Open-Set Recognition for VLMs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "type": "text", + "content": "In an open-set situation, the VLM should be able to identify an input not represented by the query set and reject it as unknown, instead of misclassifying it as one of the query classes from " + }, + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "type": "text", + "content": ". Similar to the pre-VLM open-set literature, there are multiple ways of recognising an input as unknown [13]. We investigate baseline methods using predictive uncertainty and different approaches of negative embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 599, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 482, + 624 + ], + "type": "text", + "content": "Using Uncertainty: Different measures of predictive uncertainty " + }, + { + "bbox": [ + 131, + 599, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 131, + 599, + 482, + 624 + ], + "type": "text", + "content": " can be used to reject inputs as unknown, for example based on the class similarities, e.g." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": "1 Most methods insert an additional step here and generate a set of prompts from each query label " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": ", such as {\"A photo of a dog\", \"A picture of a dog\", ...} and compare the embeddings of these prompts with " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{I}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": ". For clarity of notation, we omit this step." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\psi^{\\mathrm{Cosine}} = \\max (\\mathbf{s}^q)" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\psi^{\\mathrm{Softmax}} = \\max (\\mathrm{softmax}(\\mathbf{s}^q))" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ", or the entropy of the soft-max similarities " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\psi^{\\mathrm{Entropy}} = -H(\\mathrm{softmax}(\\mathbf{s}^q))" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ". Only inputs that pass a threshold test " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\psi \\geq \\theta" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " are classified according to " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\arg \\max \\mathbf{s}^q" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ". Otherwise, the input is declared to be open-set and unknown." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "content": "Dedicated Negative Embeddings: Another approach is to create a set of augmented query embeddings " + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^q = \\mathcal{E}^q \\cup \\mathcal{E}^n" + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "content": " by adding " + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "content": " dedicated negative class embeddings " + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^n = \\{\\mathbf{e}_1^n, \\ldots, \\mathbf{e}_M^n\\}" + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "content": ". The input image or detection can then be rejected as unknown if the image embedding is most similar to one of those negative embeddings " + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i^n" + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "content": " rather than a query class embedding " + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_j^q" + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "content": ". These negative embeddings can be created in one of two ways: (1) by creating negative words, which are then passed through the VLM text encoder to create the negative embeddings, or (2) by directly creating negative embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "text", + "content": "We investigate two baseline options for negative embeddings: a random words method, where we generate random strings with random length between 2-8 characters (e.g. \"braxl\"); and a random embeddings method, where we randomly draw " + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "text", + "content": " embedding vectors from a Gaussian distribution, i.e. " + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i^{\\mathrm{n}}\\sim \\mathcal{N}(\\boldsymbol {\\mu},\\boldsymbol {\\Sigma})" + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "text", + "content": ", using the mean and standard deviation from the query embedding vectors " + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^q" + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "text", + "content": ". We compare the effectiveness of these approaches, including for different numbers of " + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 261, + 481, + 381 + ], + "type": "text", + "content": " random negative queries, in our experiments. Interestingly, many open vocabulary object detectors already use a negative query approach, though typically with a only single embedding vector of all zeroes [22,44,50,54,55]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 400, + 272, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 272, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 272, + 414 + ], + "type": "text", + "content": "4 Evaluation Protocol" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 427, + 427, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 427, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 427, + 440 + ], + "type": "text", + "content": "4.1 Creating an Open-set Recognition Dataset for VLMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 510 + ], + "type": "text", + "content": "We now introduce how to use existing image classification or object detection datasets to evaluate open-set recognition. In short, we test all images twice: once in a standard closed-set manner with all dataset classes included in the query set (see Fig. 1 top-right), and once in an open-set manner with a query set containing all dataset classes not present in the image (see Fig. 1 bottom-right)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "content": "Classification and object detection datasets have a pre-defined set of " + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "content": " labelled target classes " + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "content": ". In image classification, there is a single corresponding class label " + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{L}" + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "content": " per image. In object detection, there are multiple ground truth objects in each image and each has a corresponding class label. For both tasks, we can define a set " + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "content": " that contains the ground truth class labels for each image: " + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 510, + 481, + 582 + ], + "type": "text", + "content": " contains a single label in classification and multiple labels in object detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 582, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 481, + 666 + ], + "type": "text", + "content": "Testing Closed-set Recognition: We first test images in a standard manner to establish the closed-set performance of the VLM and identify all true positive " + }, + { + "bbox": [ + 131, + 582, + 481, + 666 + ], + "type": "inline_equation", + "content": "(TP)" + }, + { + "bbox": [ + 131, + 582, + 481, + 666 + ], + "type": "text", + "content": " predictions. This involves testing all images with a query set identical to the dataset class list, " + }, + { + "bbox": [ + 131, + 582, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\mathcal{L}" + }, + { + "bbox": [ + 131, + 582, + 481, + 666 + ], + "type": "text", + "content": ". To identify the TP predictions, we follow standard protocol for classification and object detection datasets - in classification, a TP has a predicted class matching the ground-truth label; in object detection, this involves a object-detection assignment process that requires correct object" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "classification and localisation (see Supplementary Sec. S1). TPs represent correct closed-set predictions that should not be rejected as unknown, and we collect the uncertainty measures " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " associated with all TPs for later evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": "For object detection datasets, unlabelled objects are present in the background of the image [10]. These unlabelled objects are not in the query set " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": ", and therefore meet our definition for open-set objects. Detections of these open-set objects will present as false positives (FPs) in a closed-set test, however not all FPs arise from open-set objects - FPs in object detection can also arise due to poor object localisation, misclassification between query classes, or duplicate detections of a query class (see [5]). Rather than attempting to categorise different FP types to distinguish open-set errors, our approach definitively identifies open-set errors for both object detection and image classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "text", + "content": "Testing Open-set Recognition: To evaluate open-set recognition, we test each image in the dataset with a modified query set, " + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "text", + "content": " that contains all the class labels not in the image ground-truth class list, " + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Q}} = \\mathcal{L} - \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "text", + "content": ". This ensures the image only contains object classes that are not in the modified query set - every single prediction must be an open-set error " + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "inline_equation", + "content": "(OSE)^2" + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "text", + "content": ". In image classification, this modifies the query set to have " + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "text", + "content": " labels. In object detection, the modified query set size depends on the number of unique object classes in the image. We collect the uncertainty measures " + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 262, + 482, + 357 + ], + "type": "text", + "content": " of all OSE predictions for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 380, + 200, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 200, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 200, + 392 + ], + "type": "text", + "content": "4.2 Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 405, + 481, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 481, + 441 + ], + "type": "text", + "content": "We use performance metrics that measure the ability of the VLM to reject OSEs (with high uncertainty or with a negative class), while maintaining the closed-set task performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "content": "Precision-Recall Curves: Open-set recognition can be formulated as a binary classification task that uses uncertainty as the decision threshold. Correct closed-set predictions (introduced as TPs in Sec. 4.1) and open-set errors (OSEs) are considered as the positive and negative class respectively. We construct precision-recall curves, which are well-suited to this problem as they are affected by the size of the negative class (OSEs) " + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "inline_equation", + "content": "[36]^3" + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "content": " as well as prediction uncertainty. High precision indicates that little-to-none negatives (OSEs) remain after the uncertainty thresholding. High recall indicates that most positives (TPs) remain after the uncertainty thresholding. We summarise the performance of the precision-recall curves with 3 core metrics: (1) Area under the PR curve (AuPR), (2) Precision at " + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "content": " Recall, and (3) Recall at " + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "content": " Precision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 574, + 481, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 481, + 599 + ], + "type": "text", + "content": "Task Metrics: We use the established task metrics used to evaluate classification and detection - top-1 accuracy [34] and mean Average Precision (mAP) [23]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 609, + 481, + 632 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 609, + 481, + 632 + ], + "type": "text", + "content": "2 This assumes accurate labelling of ground-truth classes, though most datasets will contain some label error. This approach is unsuitable for weakly-labelled datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 632, + 481, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 664 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 632, + 481, + 664 + ], + "type": "text", + "content": " ROC curves do not have this property - we discuss this further in Sec. S3 of the Supplementary Material where we report AuROC for the object detection experiments." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "respectively. We evaluate mAP at an IoU threshold of 0.5 - we are primarily interested in the classification ability of the detector (rather than localisation) and this is standard protocol in open vocabulary detection [15,22,44,48,50,54,55]. Absolute Counts: We also report the absolute counts of TPs and OSEs. This can be particularly useful for object detection, where increased task performance does not necessarily indicate increased TPs and detectors predict variable numbers of OSEs. We stress that absolute counts should not be considered alone as a performance indicator, e.g. an increased number of OSEs can be acceptable if all OSEs have very high uncertainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 242, + 238, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 238, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 238, + 254 + ], + "type": "text", + "content": "4.3 VLM Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 443 + ], + "type": "text", + "content": "We test with six VLM classifiers - CLIP [33], ALIGN [19], CoCa [46], Image-Bind [14], SigLIP [52] and LanguageBind [58] - and seven open vocabulary object detectors - OVR-CNN [50], ViLD [15], OV-DETR [48], RegionCLIP [54], Detric [55], VLDet [22] and CORA [44] (see Supplementary Sec. S2 for implementation details). These VLMs cover a range of architectures and training paradigms and have publicly-available implementations. Our goal with testing this range of VLMs is not to find the \"best\" VLM for open-set recognition, but instead to exhibit the general vulnerability of VLMs to open-set conditions. For each VLM, we extract the prediction Softmax score, Cosine similarity, and Entropy of the softmax distribution as baseline uncertainty measures. Some VLMs use Sigmoid activation rather than Softmax [22,44,48,52,55] - for these methods, we only report the Sigmoid score as its results are identical to the cosine similarity and the entropy is unsuitable as the scores are not normalised to a distribution. We do not impose a minimum score threshold on any predictions, as the AuPR metric captures performance over all possible thresholds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 460, + 204, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 204, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 204, + 471 + ], + "type": "text", + "content": "4.4 Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 480, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 577 + ], + "type": "text", + "content": "We focus on general object recognition and test with the large-scale and prevailing datasets used to benchmark image classification and object detection: For image classification, we test with the ImageNet1k validation dataset [9, 34]. It contains 50,000 images from 1000 classes and is publicly available for research. For object detection, we test with the MS COCO 2017 validation dataset [23]. It contains 4,952 images with 36,781 annotated objects from 80 classes and is publicly available for research. We show results for domain-specific classification datasets in the Supplementary Sec. S4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 595, + 304, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 595, + 304, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 595, + 304, + 609 + ], + "type": "text", + "content": "5 Experiments and Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 620, + 478, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 478, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 478, + 633 + ], + "type": "text", + "content": "5.1 State-of-the-art VLMs Perform Poorly in Open-set Conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Classification: Tab. 1 shows that all six VLM classifiers perform poorly for open-set recognition. For scenarios that require high recall of true positive" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 487, + 394 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: The open-set performance of state-of-the-art VLM classifiers when tested on ImageNet1k. Arrows indicate direction of optimal performance and - indicates operating point could not be achieved." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 487, + 394 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 487, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 487, + 394 + ], + "type": "table", + "html": "
ClassifierUncertaintyAuROCAuPRP@95RR@95PTPOSEAccuracy
↑↑↑↑↑↓↑
CLIP [33]Softmax79.771.546.23.4
Cosine72.261.342.50.131,02350,00062.1
Entropy80.272.345.42.8
ALIGN [19]Softmax81.074.748.18.3
Cosine72.462.144.8-32,61850,00065.2
Entropy80.975.046.88.6
CoCa [46]Softmax79.071.746.03.4
Cosine73.263.043.30.131,72550,00063.4
Entropy80.573.046.83.5
ImageBind [14]Softmax82.879.152.85.0
Cosine79.273.950.70.238,40550,00076.8
Entropy84.380.156.35.6
SigLIP [52]Sigmoid81.076.252.12.237,85150,00075.7
LanguageBind [58]Softmax83.980.754.710.5
Cosine81.677.452.80.839,24350,00078.5
Entropy85.481.758.411.1
", + "image_path": "0a106caac5e6f9cf0cd64ba36f6d5e13bf393841a22e436c913696df0b67b7e9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": "predictions (i.e. " + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": " recall), approximately every second prediction returned is an open-set error (i.e. precision between " + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "inline_equation", + "content": "46.2\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "inline_equation", + "content": "58.4\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": "). For scenarios that require high precision of predictions (i.e. " + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": " precision), less than " + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": " of the true positive predictions are retained (i.e. recall between " + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": "). Comparing across uncertainty types, the VLM classifiers show better uncertainty performance when using the predicted Softmax score or Entropy rather than the predicted Cosine similarity. This suggests raw similarity in VLM feature space is not the best indicator of open-set error, with measures of relative similarity comparative to other classes offering better performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 531, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 604 + ], + "type": "text", + "content": "Object Detection: All seven of the tested VLM object detectors are vulnerable to open-set errors - see Tab. 2. Unlike VLM classifiers, most VLM detectors already contain a negative query to capture objects not in the query set (often referred to as the background class). Despite this, the VLM detectors produce open-set error counts in the range of 100,000 to 1,500,000 when tested on only 4,952 images, i.e. between 20 and 300 open-set error per image on average." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "High open-set error counts are not inherently a problem if they are produced with high uncertainty. Yet for all tested detectors, the baseline uncertainty methods cannot adequately distinguish true positive detections from open-set errors. When thresholding for high recall of true positives, at best only every fifth detection is a true positive (i.e. precision ranges between " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "21.9\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "). When" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 479, + 342 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 2: The open-set performance of state-of-the-art open-vocabulary object detectors (with ResNet backbone indicated), tested on COCO. Arrows indicate the direction of optimal performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 479, + 342 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 479, + 342 + ], + "type": "table", + "html": "
DetectorUncertaintyNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]SoftmaxZero-Emb75.413.042.9
CosineZero-Emb77.215.343.613,544190,14634.7
EntropyZero-Emb63.27.543.4
ViLD (R152) [15]Softmax“background”60.17.515.4
Cosine“background”33.16.90.116,0111,485,60046.4
Entropy“background”62.28.416.3
OV-DETR (R50) [48]SigmoidNone75.85.947.315,8181,485,60043.0
RegionCLIP (R50) [54]SoftmaxZero-Emb73.19.939.4
CosineZero-Emb74.07.140.915,741493,94044.6
EntropyZero-Emb35.43.917.0
Detic (R50) [55]SigmoidZero-Emb72.68.142.214,670495,20039.1
VLDet (R50) [22]SigmoidZero-Emb79.821.948.514,751112,92340.6
CORA (R50) [44]SigmoidZero-Emb67.28.124.713,763495,20031.6
", + "image_path": "62576fbc24da883ef1704c6d1c72a828ab3dceb047558b2b17f1b0ce111946df.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 365, + 480, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 400 + ], + "type": "text", + "content": "thresholding for high precision, at best less than half of true positives are recalled (i.e. recall ranges between " + }, + { + "bbox": [ + 132, + 365, + 480, + 400 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 132, + 365, + 480, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 365, + 480, + 400 + ], + "type": "inline_equation", + "content": "48.5\\%" + }, + { + "bbox": [ + 132, + 365, + 480, + 400 + ], + "type": "text", + "content": "). We include some qualitative examples of highly confident open-set errors in the Supplementary Sec. S6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 401, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 401, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 480, + 472 + ], + "type": "text", + "content": "Notably, the precision-recall characteristics for VLM classification and object detection differ significantly – VLM classifiers achieve better performance at high recall, whereas VLM detectors achieve better performance at high precision. This suggests that while the VLM classifiers struggle with over-confident open-set errors, the VLM detectors suffer from under-confident true positives. We include histograms that visualise this trend in the Supplementary Sec. S7." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 473, + 480, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 480, + 579 + ], + "type": "text", + "content": "We also consider how the performance of uncertainty differs for identifying open-set error versus closed-set misclassifications, with results and a detailed discussion in the Supplementary Sec. S5. We find that the performance of softmax uncertainty is lower for open-set error than for closed-set error in both classification and detection tasks. In classification, open-set performance is particularly decremented when thresholding for high precision - likely due to a greater proportion of overconfident open-set error - whereas detection primarily degrades with increased numbers of errors, with ViLD for example producing " + }, + { + "bbox": [ + 132, + 473, + 480, + 579 + ], + "type": "inline_equation", + "content": "11.7 \\times" + }, + { + "bbox": [ + 132, + 473, + 480, + 579 + ], + "type": "text", + "content": " more open-set than closed-set misclassifications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 597, + 436, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 597, + 436, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 597, + 436, + 609 + ], + "type": "text", + "content": "5.2 Can Negative Queries Improve Open-set Performance?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 480, + 665 + ], + "type": "text", + "content": "Classification: Negative class queries provide the VLM classifiers with the ability to select \"unknown\" when testing an input. When used in larger quantities, negative queries can capture and remove significant portions of open-set error, yet this happens in trade-off with a reduction in closed-set performance -" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 146, + 446, + 487 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 3: The open-set performance of state-of-the-art VLM classifiers on ImageNet1k when used with negative random words (R-W) or random embeddings (R-E)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 146, + 446, + 487 + ], + "lines": [ + { + "bbox": [ + 167, + 146, + 446, + 487 + ], + "spans": [ + { + "bbox": [ + 167, + 146, + 446, + 487 + ], + "type": "table", + "html": "
ClassifierNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓Accuracy ↑
CLIP [33] (Softmax)None71.546.23.431,02350,00062.1
500 R-W73.448.13.329,96141,42859.9
2500 R-W72.849.81.628,65735,09257.3
500 R-E71.947.22.530,16044,67360.3
2500 R-E71.747.10.929,29140,49758.6
ALIGN [19] (Softmax)None74.748.18.332,61850,00065.2
500 R-W75.348.910.431,73143,19863.5
2500 R-W74.649.58.930,58837,73861.2
500 R-E74.948.78.632,34247,22864.7
2500 R-E75.149.77.231,46241,21062.9
CoCa [46] (Softmax)None71.746.03.431,72550,00063.4
500 R-W75.449.60.130,27639,52860.6
2500 R-W75.950.96.029,10734,36058.2
500 R-E71.746.03.431,72349,96663.4
2500 R-E71.746.13.431,71949,86663.4
ImageBind [14] (Softmax)None79.152.85.038,40550,00076.8
500 R-W83.863.17.735,85929,17271.7
2500 R-W84.666.18.334,17922,87268.4
500 R-E81.057.39.036,96638,63673.9
2500 R-E81.759.011.235,95433,34971.9
SigLIP [52] (Sigmoid)None76.252.15.237,85150,00075.7
500 R-W77.656.72.336,11133,28772.2
2500 R-W78.860.22.434,50726,11069.0
500 R-E84.773.03.622,7729,57045.5
2500 R-E90.183.97.910,5612,25821.1
LanguageBind [58] (Softmax)None80.754.710.539,24350,00078.5
500 R-W84.764.913.136,59628,33773.2
2500 R-W85.468.211.734,54021,23069.1
500 R-E80.854.910.839,22649,64678.5
2500 R-E80.955.110.939,21049,03278.4
", + "image_path": "1a28376a8efa8409199a5438b1595099f4eb1671116037bf086cfda79f8d2bcf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "type": "text", + "content": "see Tab. 3. For example, ImageBing using 2,500 random word negative queries effectively halves the open-set error, but this occurs in trade-off with a closed-set accuracy reduction of " + }, + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 130, + 509, + 481, + 556 + ], + "type": "text", + "content": " (approximately 4,000 of the closed-set images are incorrectly classified as one of the random words)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "While negative queries noticeably reduce the raw open-set error count, this is not always mirrored with increased performance in the open-set uncertainty metrics. We present the change in closed-set and open-set performance against the number of negative query points in Fig. 2. Generally the use of negative random words appears the most effective, with CoCa, ImageBind, SigLIP and Language-Bind showing improved open-set performance with minimal loss of closed-set performance. For some of the classifiers, there is negligible change in open-set performance despite the reduction in open-set error. We explore this further in the Supplementary Sec. S8, where we show that in some VLMs the negative" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 474, + 251 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 474, + 251 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 474, + 251 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 474, + 251 + ], + "type": "image", + "image_path": "1d6d2a49c30f0b30b0ddb8cc7229d817202099d78c738a68d062201a52532894.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "text", + "content": "Fig. 2: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM classifiers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 304, + 474, + 441 + ], + "blocks": [ + { + "bbox": [ + 135, + 304, + 474, + 441 + ], + "lines": [ + { + "bbox": [ + 135, + 304, + 474, + 441 + ], + "spans": [ + { + "bbox": [ + 135, + 304, + 474, + 441 + ], + "type": "image", + "image_path": "b25352584581fd230f2ec6b9cb5cb48e7df50ae17f800868cba8f9bf92df34e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 452, + 480, + 474 + ], + "lines": [ + { + "bbox": [ + 132, + 452, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 480, + 474 + ], + "type": "text", + "content": "Fig. 3: The trade-off between closed-set and open-set performance with increasing number of negative queries for different VLM object detectors." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 506, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 479, + 531 + ], + "type": "text", + "content": "queries only capture and remove \"easy\" open-set errors, i.e. open-set errors that already have high uncertainty." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Object Detection: The VLM detectors do not show a consistent response to the use of negative queries. Different detectors show different sensitivities to negative query counts, different trade-off rates between closed-set and open-set performance, and differences in effectiveness of embedding versus word negative query types - see Tab. 4 and Fig. 3. Looking at Fig. 3, we can consider the sensitivity, trade-off rate and response of the detectors to negative queries: greater performance changes with increasing queries indicates sensitivity; large decreases in mAP with more queries indicates increased trade-off rates between open-set and closed-set performance; and a positive response to the negative queries presents as an increase in AuPR when increasing queries (negative response decreases AuPR)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 168, + 471, + 550 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: The open-set performance of state-of-the-art open vocabulary object detectors (with ResNet backbone indicated) when additional negative embeddings are included, tested on COCO. Arrows indicate the direction of optimal performance. Negative embeddings introduced as words are highlighted gray." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 168, + 471, + 550 + ], + "lines": [ + { + "bbox": [ + 140, + 168, + 471, + 550 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 471, + 550 + ], + "type": "table", + "html": "
DetectorNegative EmbeddingAuPR ↑P@95R ↑R@95P ↑TP ↑OSE ↓mAP ↑
OVR-CNN (R50) [50]Zero-Emb75.413.042.913,544190,14634.7
+5 Rand-Embs75.613.143.113,543186,12034.7
+100 Rand-Embs78.318.245.613,129109,26734.0
ViLD (R152) [15]“background”60.17.515.416,0111,485,60046.4
+5 Rand-Words57.56.013.415,9261,400,61246.6
+100 Rand-Words55.02.710.315,7721,088,72547.3
+5 Rand-Embs60.17.515.216,0111,485,23146.4
+100 Rand-Embs60.27.615.216,0191,478,80646.4
OV-DETR (R50) [48]None75.85.947.315,8181,485,60043.0
+5 Rand-Words77.28.148.315,510923,39843.0
+100 Rand-Words84.727.055.713,433239,11840.1
+5 Rand-Embs76.66.647.815,6521,178,82743.0
+100 Rand-Embs81.018.250.914,699356,38042.6
RegionCLIP (R50) [54]Zero-Emb73.19.939.315,752493,94944.6
+5 Rand-Words73.811.139.715,662322,59644.6
+100 Rand-Words77.017.543.115,178154,26244.0
+5 Rand-Embs91.877.244.68,8285,06227.0
+100 Rand-Embs99.999.383.628721.6
Detic (R50) [55]Zero-Emb72.68.142.214,670495,20039.1
+5 Rand-Words72.88.542.414,637451,63339.1
+100 Rand-Words74.511.543.514,287265,37338.8
+5 Rand-Embs93.474.561.810,0706,34630.6
+100 Rand-Embs100.0100.0100.037100.7
VLDet (R50) [22]Zero-Emb79.821.948.414,751112,95440.6
+5 Rand-Words80.122.348.714,672111,89340.5
+100 Rand-Words86.536.758.411,13250,12233.2
+5 Rand-Embs97.391.384.98,4641,38225.8
+100 Rand-Embs99.099.33.72,531225.9
CORA (R50) [44]Zero-Emb67.28.124.913,745495,20031.5
+5 Rand-Words66.37.627.013,700413,63631.6
+100 Rand-Words69.110.128.613,722329,39330.1
+5 Rand-Embs63.16.221.813,689495,20031.5
+100 Rand-Embs59.96.016.713,680495,17430.6
", + "image_path": "fae1a9361dc708aa2202875fbf145f1737b9f0822f9d66c95bd953f23da7f19f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "OV-DETR is the only detector to show a clear positive response, low trade-off rate, and high sensitivity to the negative query approach, particularly with negative embeddings. While OVR-CNN (embeddings), RegionCLIP (words), Detric (words) also show a positive response and low trade-off rate, this is with very little sensitivity to increasing numbers of negative queries. In contrast, the open-set performance of Detric, RegionCLIP, and VLDet are very sensitive to negative" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 116, + 304, + 229 + ], + "blocks": [ + { + "bbox": [ + 189, + 116, + 304, + 229 + ], + "lines": [ + { + "bbox": [ + 189, + 116, + 304, + 229 + ], + "spans": [ + { + "bbox": [ + 189, + 116, + 304, + 229 + ], + "type": "image", + "image_path": "3f1faff813e5f05cedd03c8050f31cdf1a9a20886cba024b9b3d2175f2460624.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 232, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 206, + 232, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 206, + 232, + 287, + 242 + ], + "type": "text", + "content": "(a) Classification task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 116, + 425, + 229 + ], + "blocks": [ + { + "bbox": [ + 306, + 116, + 425, + 229 + ], + "lines": [ + { + "bbox": [ + 306, + 116, + 425, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 425, + 229 + ], + "type": "image", + "image_path": "a1aa75051a50cd3509a999c13d057de3e83f58c8d57e9d70281e624101b0a004.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 232, + 413, + 242 + ], + "lines": [ + { + "bbox": [ + 318, + 232, + 413, + 242 + ], + "spans": [ + { + "bbox": [ + 318, + 232, + 413, + 242 + ], + "type": "text", + "content": "(b) Object Detection task." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 251, + 468, + 262 + ], + "lines": [ + { + "bbox": [ + 143, + 251, + 468, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 468, + 262 + ], + "type": "text", + "content": "Fig. 4: The relationship between closed-set and open-set performance in VLMs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 479, + 308 + ], + "type": "text", + "content": "embeddings, but at a rapid trade-off with closed-set performance. Both ViLD and CORA show a negative response or no sensitivity to negative queries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 309, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 429 + ], + "type": "text", + "content": "It is tempting to relate these trends to the architecture or training paradigm of the VLM detectors, however we could not identify many common characteristics. RegionCLIP, Detic and VLDet all learn from auxiliary datasets that have been pseudo-labelled with regions [22, 54, 55], although this does not clearly indicate a reason for high sensitivity and trade-off rate for negative embeddings. Vild [15] and CORA [44] both heavily rely on CLIP's visual feature encoding (either by using CLIP directly [44] or knowledge distillation [15]) - their poor response is consistent with CLIP in the classification experiments on negative queries. Yet OV-DETR also uses CLIP's visual features during training to condition the object proposals [48], but exhibits the best response to negative queries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 445, + 479, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 479, + 456 + ], + "type": "text", + "content": "5.3 Is a good classifier all you need for VLM Open-set Recognition?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 463, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 480, + 510 + ], + "type": "text", + "content": "In traditional open-set recognition, Vaze et. al [40] identified a strong positive correlation between closed-set and open-set performance. Testing this theory for VLM open-set recognition, we compare classifier and object detector closed-set and open-set performance in Fig. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 511, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 480, + 570 + ], + "type": "text", + "content": "Classification: The correlation observed by [40] appears to hold for VLM classifiers, with greater accuracy (closed-set performance) relating to greater AuPR scores (open-set performance). It is worth noting that AuPR is a summary metric and this behaviour may not hold when examining performance at different operating points on the precision-recall curves." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 571, + 480, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 480, + 606 + ], + "type": "text", + "content": "Object Detection: There is not a clear correlation between a VLM object detector's closed-set and open-set performance. ViLD in particular challenges this correlation, with the greatest mAP yet worst AuPR of all the tested detectors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 623, + 402, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 402, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 402, + 635 + ], + "type": "text", + "content": "5.4 How Does Query Set Size Impact Performance?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "In Fig. 5, we test performance of the VLM classifiers on ImageNet while progressively increasing the query set size. We select the additional classes from the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 116, + 279, + 197 + ], + "blocks": [ + { + "bbox": [ + 187, + 116, + 279, + 197 + ], + "lines": [ + { + "bbox": [ + 187, + 116, + 279, + 197 + ], + "spans": [ + { + "bbox": [ + 187, + 116, + 279, + 197 + ], + "type": "image", + "image_path": "767a2cee1c481a09e25d58855181a77c904d0bd4bf3f18a8db1b1a4624d5ab31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 479, + 231 + ], + "type": "text", + "content": "Fig. 5: Influence of the query set size on closed-set and open-set performance on ImageNet1k classification, showing mean and standard deviation from 10 random seeds." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 284, + 116, + 427, + 198 + ], + "blocks": [ + { + "bbox": [ + 284, + 116, + 427, + 198 + ], + "lines": [ + { + "bbox": [ + 284, + 116, + 427, + 198 + ], + "spans": [ + { + "bbox": [ + 284, + 116, + 427, + 198 + ], + "type": "image", + "image_path": "337bb0c7993599552ced58d7bfa22c1571e1cebb85b3304efa8ca74b15080da8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 252, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 480, + 326 + ], + "type": "text", + "content": "WordNet hierarchy by choosing leaf nodes that are related, but not too closely related, to existing ImageNet classes (i.e. share a grandparent in the hierarchy but do not share a direct parent). Increasing the number of target classes and thus the query set size decrements both the closed-set and open-set performance of a VLM classifier. This indicates that testing VLMs with an entire vocabulary of nouns may not be a feasible approach to avoid open-set recognition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 342, + 217, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 217, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 217, + 355 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 415 + ], + "type": "text", + "content": "Open Challenges: The tested baseline uncertainty measures were ineffective at identifying open-set errors, indicating a need to research better alternatives. While uncertainty representations for VLMs have been proposed [18, 39], their utility for open-set recognition remains unexplored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 510 + ], + "type": "text", + "content": "While the use of negative random embeddings was able to reduce open-set error, large numbers were typically required to have any effect and this was often as the cost of falsely captured true positive predictions. Interestingly, some negative embeddings were much more effective at capturing open-set errors than others (e.g. the random string \"awy\" tested with ImageBind, see the Supplementary Sec. S9). Identifying the characteristics of these embeddings, learning them through prompt-tuning [8, 11, 12, 21, 57] or introducing auto-labelling techniques [49] are promising directions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "content": "Conclusions: VLMs are increasingly applied in diverse real-world applications, from robotics to medical image analysis. We hope that by highlighting their weaknesses, we encourage further research to improve these models and enhance their positive societal impact. We have clearly demonstrated that VLMs for open-vocabulary image classification and object detection introduce closed-set assumptions by relying on a finite query set. We thus refuted the assumption that vision-language models (VLMs) are inherently open-set models due to their training on extensive internet-scale datasets. While the investigated baseline approaches to identify open-set errors showed promise, we found none of them sufficiently robust and reliable for safety-critical applications. Our new benchmark and evaluation protocol is designed to foster and support this research. Understanding VLMs' limitations ultimately reduces risks, improves transparency, and ensures safer deployment in critical applications." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 100 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 186, + 481, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 481, + 209 + ], + "type": "text", + "content": "2. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., Mané, D.: Concrete problems in ai safety. arXiv preprint arXiv:1606.06565 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 243 + ], + "type": "text", + "content": "3. Bao, W., Yu, Q., Kong, Y.: Evidential deep learning for open set action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13349-13358 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 481, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 265 + ], + "type": "text", + "content": "4. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1563-1572 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 308 + ], + "type": "text", + "content": "5. Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying object detection errors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 558-573. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "type": "text", + "content": "6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Deep metric learning for open world semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15333-15342 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 376 + ], + "type": "text", + "content": "7. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 8065-8081 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "text", + "content": "8. Cho, E., Kim, J., Kim, H.J.: Distribution-aware prompt tuning for vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22004-22013 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 410, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 443 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 444, + 481, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 481, + 476 + ], + "type": "text", + "content": "0. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: The IEEE Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 477, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 481, + 520 + ], + "type": "text", + "content": "1. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14084-14093 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 555 + ], + "type": "text", + "content": "2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "3. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "type": "text", + "content": "4. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Han, J., Ren, Y., Ding, J., Pan, X., Yan, K., Xia, G.S.: Expanding low-density latent regions for open-set object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9591-9600 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "type": "text", + "content": "17. Hwang, J., Oh, S.W., Lee, J.Y., Han, B.: Exemplar-based open-set panoptic segmentation network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1184 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 181, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 181, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 481, + 224 + ], + "type": "text", + "content": "18. Ji, Y., Wang, J., Gong, Y., Zhang, L., Zhu, Y., Wang, H., Zhang, J., Sakai, T., Yang, Y.: Map: Multimodal uncertainty-aware vision-language pre-training model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23262-23271 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 224, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 481, + 289 + ], + "type": "text", + "content": "19. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "text", + "content": "20. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "text", + "content": "21. Li, H., Zhang, R., Yao, H., Song, X., Hao, Y., Zhao, Y., Li, L., Chen, Y.: Learning domain-aware detection head with prompt tuning. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "22. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "text", + "content": "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 461 + ], + "type": "text", + "content": "24. Liu, Y.C., Ma, C.Y., Dai, X., Tian, J., Vajda, P., He, Z., Kira, Z.: Open-set semi-supervised object detection. In: European Conference on Computer Vision. pp. 143-159. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 461, + 481, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 505 + ], + "type": "text", + "content": "25. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 505, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 481, + 536 + ], + "type": "text", + "content": "26. Miller, D., Nicholson, L., Dayoub, F., Sunderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: 2018 IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249. IEEE (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "27. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. IEEE Robotics and Automation Letters 7(1), 215-222 (2022). https://doi.org/10.1109/LRA.2021.3123374" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "28. Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2307-2316 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "29. Panareda Busto, P., Gall, J.: Open set domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 754-763 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "30. Perera, P., Morariu, V.I., Jain, R., Manjunatha, V., Wigington, C., Ordonez, V., Patel, V.M.: Generative-discriminative feature representations for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11814-11823 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Pham, C., Vu, T., Nguyen, K.: Lp-ovod: Open-vocabulary object detection by linear probing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 779–788 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "32. Pham, T., Do, T.T., Carneiro, G., Reid, I., et al.: Bayesian semantic instance segmentation in open set world. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 3-18 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 246 + ], + "type": "text", + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 246, + 482, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 279 + ], + "type": "text", + "content": "34. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 279, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 311 + ], + "type": "text", + "content": "35. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-168 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 343 + ], + "type": "text", + "content": "36. Saito, T., Rehmsmeier, M.: The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PloS one 10(3), e0118432 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 343, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 374 + ], + "type": "text", + "content": "37. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(7), 1757-1772 (2013). https://doi.org/10.1109/TPAMI.2012.256" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "type": "text", + "content": "38. Sünderhauf, N., Brock, O., Scheirer, W., Hadsell, R., Fox, D., Leitner, J., Upcroft, B., Abbeel, P., Burgard, W., Milford, M., et al.: The limits and potentials of deep learning for robotics. The International Journal of Robotics Research 37(4-5), 405-420 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 418, + 482, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 451 + ], + "type": "text", + "content": "39. Upadhyay, U., Karthik, S., Mancini, M., Akata, Z.: Probvlm: Probabilistic adapter for frozen vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1899-1910 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 450, + 482, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 482 + ], + "type": "text", + "content": "40. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 482, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 515 + ], + "type": "text", + "content": "41. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6759-6769 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "type": "text", + "content": "42. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards Open Vocabulary Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 547, + 482, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 579 + ], + "type": "text", + "content": "43. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 579, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 621 + ], + "type": "text", + "content": "44. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "text", + "content": "45. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4016-4025 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Set Recognition in the Age of Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 567 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "46. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (2022), https://openreview.net/forum?id=Ee277P3AYC" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "47. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "48. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: European Conference on Computer Vision. pp. 106-122. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "49. Zara, G., Roy, S., Rota, P., Ricci, E.: Autolabel: Clip-based framework for open-set video domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11504-11513 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "50. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "text", + "content": "51. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18123-18133 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "text", + "content": "52. Zhai, X., et al.: Sigmoid loss for language image pre-training. In: Intl. Conf. on Computer Vision. pp. 11975-11986 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "53. Zhang, H., Li, A., Guo, J., Guo, Y.: Hybrid models for open set recognition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 102-117. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "text", + "content": "54. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "text", + "content": "55. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "56. Zhou, Z., Yang, Y., Wang, Y., Xiong, R.: Open-set object detection using classification-free object proposal and instance-level contrastive learning. IEEE Robotics and Automation Letters 8(3), 1691-1698 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 567 + ], + "type": "text", + "content": "58. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: Intl. Conf. on Learning Representations (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "D. Miller et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_content_list.json b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e8bbab1cddd09f5df4b06a63413ba2781c7dc26a --- /dev/null +++ b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_content_list.json @@ -0,0 +1,1641 @@ +[ + { + "type": "text", + "text": "Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyu Zhu $^{1\\dagger}$ , Hao Zhou $^{2}$ , Pengfei Xing $^{2}$ , Long Zhao $^{2}$ , Hao Xu $^{3}$ , Junwei Liang $^{4}$ , Alexander Hauptmann $^{1}$ , Ting Liu $^{2}$ , and Andrew Gallagher $^{2}$", + "bbox": [ + 240, + 212, + 763, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Carnegie Mellon University", + "bbox": [ + 401, + 253, + 602, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Google DeepMind", + "bbox": [ + 433, + 268, + 568, + 282 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3 Google", + "HKUST" + ], + "bbox": [ + 467, + 268, + 568, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c1649a6e2e03548aeb19c4a1910e53519f3b2317cd11c94fccc95c6cf5789f8.jpg", + "image_caption": [ + "Fig. 1: Illustration of open-vocabulary 3D semantic scene understanding. We propose Diff2Scene, a 3D model that performs open-vocabulary semantic segmentation and visual grounding tasks given novel text prompts, without relying on any annotated 3D data. By leveraging discriminative-based and generative-based 2D foundation models, Diff2Scene can handle a wide variety of novel text queries for both common and rare classes, like \"desk\" and \"soap dispenser\". It can also handle compositional queries, such as \"find the white sneakers that are closer to the desk chair.\"" + ], + "image_footnote": [], + "bbox": [ + 223, + 354, + 354, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af3c50172caea8bbfc4b814c0a98bdfa0afc31dda57ca2d5ab7f13de774727ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 354, + 501, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a60d0c114989817290c1260c459e45eb10c59e66c07554867844ef8b833786c8.jpg", + "image_caption": [ + "\"This backpack is alongside a square green chair.\"" + ], + "image_footnote": [], + "bbox": [ + 504, + 354, + 637, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d1a5dc0d29cf6bef148b985b981427b14a55caede524a272704da0556eb99cf.jpg", + "image_caption": [ + "Find the white sneakers that are closer to the desk chair." + ], + "image_footnote": [], + "bbox": [ + 647, + 354, + 779, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we investigate the use of diffusion models which are pre-trained on large-scale image-caption pairs for open-vocabulary 3D semantic understanding. We propose a novel method, namely Diff2Scene, which leverages frozen representations from text-image generative models, along with salient-aware and geometric-aware masks, for open-vocabulary 3D semantic segmentation and visual grounding tasks. Diff2Scene gets rid of any labeled 3D data and effectively identifies objects, appearances, materials, locations and their compositions in 3D scenes. We show that it outperforms competitive baselines and achieves significant improvements over state-of-the-art methods. In particular, Diff2Scene improves the state-of-the-art method on ScanNet200 by $12\\%$ .", + "bbox": [ + 259, + 590, + 766, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Semantic Understanding $\\cdot$ Open-Vocabulary Perception $\\cdot$ Diffusion Model", + "bbox": [ + 261, + 758, + 740, + 785 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was partially done while the author was a student researcher at Google.", + "bbox": [ + 230, + 825, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D semantic scene understanding, with the task of assigning semantics to every 3D point, plays a fundamental role in many computer vision applications, such as robotics [88], autonomous driving [36], human-computer interaction [22], and augmented reality [27]. Traditional studies in this field usually target solving this problem in a closed-set fashion [16,73], resulting in models that can only be used to make predictions within the predefined label space.", + "bbox": [ + 212, + 175, + 785, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent progress in computer vision have witnessed the emerging interests in solving semantic understanding tasks in open-vocabulary settings [35, 62, 67, 78, 94]. In contrast to closed-set setting, models targeting open-vocabulary tasks must make predictions for any semantics described in text, including object category and fine-grained attributes (e.g., shape, color, material, property) as well as their complicated compositions. However, this is a challenging task due to the wide diversity and complexity of possible queries. Motivated by the advance of aligning text and image embeddings with large-scale foundation models [2, 39, 48, 65], existing methods mitigate this challenge by lifting the image features from foundation models such as CLIP [65] or their descendants [25, 47] to 3D. These lifted feature representations for 3D points can then be used to query with open-vocabulary descriptions, achieving semantic understanding in 3D. Despite these achievements, contrastively trained CLIP-based models exhibit limitations in handling fine-grained classes [66] and novel compositional text queries [58], restricting their performance in open-vocabulary 3D semantic understanding.", + "bbox": [ + 212, + 267, + 785, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recently developed text-to-image diffusion models have shown outstanding abilities for image generation even with challenging text prompts [55,68,72], such as combinational descriptions with multiple attributes (e.g., A bucket bag made of blue suede with intricate golden paisley patterns.) The internal visual representation of these models, entangled with text embedding through cross-attention, have proven correlate well with semantic concepts described by language [43, 60, 61, 89]. On the other hand, CLIP-based foundation models have been shown to struggle with compositionality [58]. Moreover, compared with the CLIP model which is optimized for global representation, diffusion models have proven to be superior at local representation [80], which is a key for dense prediction tasks. Specifically, ODISE [87] applied the internal representations of Stable Diffusion [68] to open-vocabulary 2D semantic understanding tasks and achieved promising results.", + "bbox": [ + 212, + 492, + 785, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One of the key challenges in 3D perception is the severe scarcity of point clouds and their dense labels. Several existing methods have been proposed to solve the lack of data issue in a zero-shot fashion by leveraging the CLIP model pre-trained on large-scale text-image data [37,62,79]. The prior art [62] extracts dense CLIP features from 2D images and distill the knowledge of their lifted 3D counterpart into a 3D mask predictor. However, CLIP features, as discussed above, struggle to handle fine-grained classes [66] and show worse localization capability compared with diffusion features. We leverage diffusion model as feature backbone along with a mask-based segmentation head (e.g., Mask2Fromer [10]) for its intrinsic nature that decouples mask and its semantic representations.", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models and further learning geometrically-accurate masks from the 3D branch. However, performing multi-modal distillation with mask-based segmentation head is a non-trivial task. The frozen features extracted from the decoder of the U-Net in the diffusion model are trained with generative objectives, and cannot be directly used for the perception task. Therefore, directly distilling knowledge from these features as normally done in prior art [54,56,62] is infeasible. Another intuitive way is to leverage a supervised 3D mask proposal network and pool the feature representations from 2D CLIP features for each mask [79]. However, the training of 3D mask proposal network requires labeled 3D data, which may not be feasible in practice.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate these issues, we propose a novel mask distillation method tailored to distill knowledge from the Mask2Former style 2D branch [10, 87] to the 3D branch, which is shown in Fig. 2. Specifically, we design our 3D branch to take a 3D point cloud as input and to predict their 3D features. The semantically meaningful mask embeddings produced from our 2D branch are used as linear classifiers to assign class probability to these 3D features. Their corresponding 2D masks are lifted to 3D based on pixel-point correspondence and used to force the consistency learning of the 2D and 3D branch.", + "bbox": [ + 212, + 313, + 787, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate Diff2Scene quantitatively on ScanNet [17], ScanNet200 [69], Matterport 3D [7] and Replica [77] for open-vocabulary 3D semantic segmentation and qualitatively on $\\mathrm{Nr3D}$ [1] for visual grounding tasks. Our experimental results show that Diff2Scene outperforms state-of-the-art models [62] on all the four semantic segmentation datasets and achieves promising results on visual grounding tasks. In summary, we make the following contributions:", + "bbox": [ + 212, + 431, + 787, + 525 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to leverage text-image diffusion to perform open-vocabulary 3D semantic segmentation.", + "- We propose a novel mask distillation method to train a 3D mask prediction model by distilling knowledge from the Mask2Former style 2D segmentation model.", + "- The proposed method achieves state-of-the-art performance on several open-vocabulary 3D semantic segmentation and visual grounding benchmarks." + ], + "bbox": [ + 223, + 532, + 782, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 672, + 387, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Closed-vocabulary 3D semantic segmentation. In 3D semantic segmentation, a semantic category is assigned for each 3D point. It has been long studied [3-5, 20, 21, 26, 31, 32, 42, 45, 49, 57, 63, 64, 82, 86, 95] due to its importance in computer vision and robotics applications. One challenge of this task is that 3D point clouds are not in a regular structured format; network architectures that work well for 2D tasks cannot handle 3D point clouds effectively. As a result, most of the early studies focus on designing effective and efficient network architectures that are suitable for 3D point clouds [16, 20, 21, 26, 31, 32, 45, 63, 64, 83]. This line of work achieved great success and significantly improves the results of", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b81f04d59bb4554ff4d29a7be3fd55452524173e13d56186929e41e9a90639c.jpg", + "image_caption": [ + "(a) Feature-based Distillation w/o labeled 3D data" + ], + "image_footnote": [], + "bbox": [ + 218, + 156, + 390, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9532b963500d02d5ef4f4e6800e7a96b3012ccf79b64bfe9218b71441c59e085.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 157, + 557, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4e832f01c47e6a31fe445588ff7d58114dc077557186fe8d22f11f8c2b978263.jpg", + "image_caption": [ + "(b) Mask-based Feature Pooling w/ labeled 3D data", + "(c) Mask-based Distillation w/o labeled 3D data (Ours)", + "Fig. 2: Illustration of open-vocabulary 3D perception methods. $L_{PD}$ and $L_{MD}$ denote point-based distillation loss and mask-based distillation loss. $M_{3D}$ denote a set of predicted 3D masks; $M_{2D}$ and $Z_{mf}$ denote a set of predicted 2D masks and their semantic embeddings; $Z_{gf}$ denote the high-resolution 3D feature map. (a) Directly minimizing the per-point feature distance between the CLIP-based model and the tuned 3D model [62]. (b) Directly using a 3D mask proposal network trained on labeled 3D data to produce class-agnostic masks, and then pool corresponding representations from the CLIP feature map [79]. (c) The proposed mask distillation approach, namely Diff2Scene, that uses Stable Diffusion and performs mask-based distillation. Diff2Scene leverages the semantically-rich mask embeddings from 2D foundation models and geometrically accurate masks from the tuned 3D model, and thus achieves superior performance compared to previous methods." + ], + "image_footnote": [], + "bbox": [ + 562, + 156, + 812, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D semantic segmentation. Another challenge is the lack of large scale data with ground truth annotations. Due to the intensive labeling effort and high cost of data annotation [69], the available datasets for 3D semantic segmentation are usually small in scale. In the absence of large scale data, early studies usually target solving this problem in a closed-vocabulary setting, where the trained model can only predict categories that appear during training. To mitigate the scale limitation of existing datasets, a handful of works [12-15, 52, 60, 91] have applied zero-shot learning in 3D scene understanding tasks. [12-15, 91] focused on 3D point classification task and [52, 60] tried to address the 3D semantic segmentation problem. However, these zero-short methods still require ground truth annotations for a certain amount of 3D point clouds.", + "bbox": [ + 212, + 501, + 787, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-vocabulary 3D segmentation. The recent progress of large-scale vision and language representation learning [2,28,39,48,65,90] has advanced the study of semantic and instance segmentation in an open-vocabulary setting. [25, 47, 51] first explored open-vocabulary 2D semantic segmentation. They proposed aligning per-pixel features [47] or features from mask regions [25, 51] with the corresponding text embedding. Following these works, [8,19,24,33,37,59,62,74, 75] focus on 3D semantic segmentation in an open-vocabulary setting. Among them, [8,24,33,37,59,71,74] project 3D points to 2D images and solve the 3D problem in the 2D space, instead of targeting the 3D open-vocabulary semantic segmentation directly. As [19] pointed out, the projection from 3D to 2D has information loss and the solution is suboptimal.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make better use of information from the 3D point cloud, [19], [18] and [62] proposed to directly applying semantic segmentation on the 3D point cloud. [19] and its extension [18] proposed associating captions generated for 2D images to corresponding 3D point clouds to build the pseudo-ground truth captions for 3D point clouds. A neural network is trained to associate the 3D point cloud with these pseudo labels through contrastive loss. Similar to the zero-shot setting, [19] evaluated their model in a leave-one-out fashion, which still requires annotations for 3D point cloud. Inspired by the strong open-vocabulary ability of large-scale vision and language models, Peng et al. [62] proposed distilling knowledge to a 3D point cloud model. They trained 3D semantic segmentation model by only distilling the knowledge from a CLIP-style [65] 2D open-vocabulary semantic segmentation model [25, 47]. They demonstrated that without training with any ground truth labels, the model can achieve great performance on many open-vocabulary tasks. However, we observe that [62] is strongly limited by the 2D open-vocabulary semantic segmentation models used as the teacher. Its performance on rare classes that are not used in training these models are not satisfactory. Our method follows this idea by distilling the knowledge of 2D open-vocabulary semantic segmentation model to a 3D model. In constraint to the approach in [62], we use a diffusion-based 2D open vocabulary semantic segmentation model [87] as the teacher model.", + "bbox": [ + 217, + 145, + 785, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion models for scene understanding. The last few years have witnessed the success of diffusion models in image generation [68, 70]. Recent studies also observed the diffusion models are strong representation learners [43, 60, 61, 89]. As a result, researchers have applied it to many understanding tasks such as image classification [46], object detection [9], image semantic segmentation [6, 38, 87], instance segmentation [50], human pose estimation [23, 29, 76], action segmentation [53], camera pose estimation [84], to name a few, and achieved great success. Especially, [87] and [50] showed that Stable Diffusion [68], whose internal representation being well correlated with text embedding, has strong open-vocabulary abilities for understanding tasks. Inspired by this, we are the first to apply text-to-image diffusion models to open vocabulary 3D semantic segmentation task.", + "bbox": [ + 217, + 449, + 785, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Diff2Scene", + "text_level": 1, + "bbox": [ + 217, + 657, + 354, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce Diff2Scene, an open-vocabulary 3D semantic understanding method. Similar to [62], our proposed model operates in a zero-shot fashion, where no ground truth 3D annotations are needed during training.", + "bbox": [ + 217, + 693, + 792, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 217, + 765, + 339, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of Diff2Scene is shown in Fig. 3. It takes posed RGB images and the reconstructed 3D point cloud as model inputs. The model predicts the semantic label for each 3D point. Diff2Scene has two branches. The 2D branch is designed", + "bbox": [ + 217, + 794, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3e326a7ce26f33488dd8a67bc7b5dc5e9f7724c15ea98183a0869e98a8559f45.jpg", + "image_caption": [ + "Fig. 3: Overview of our method. We propose Diff2Scene, an open-vocabulary 3D semantic understanding model. Diff2Scene contains two branches. The 2D branch is designed to be a diffusion-based 2D semantic segmentation model. It accepts a 2D image as input and predicts a set of 2D probabilistic masks with corresponding semantically-rich mask embeddings. The 3D branch utilizes the point cloud and 2D mask embeddings as input. The 2D mask embeddings are used as \"semantic queries\" to generate corresponding 3D probabilistic masks. The model learns salient patterns from the RGB images and geometric information from the point clouds." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 785, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to be an open-vocabulary 2D semantic segmentation model. It leverages text-to-image generative model [68] which is pre-trained on massive text-image pairs. The model takes a 2D image as input to predict a set of 2D probabilistic masks with their corresponding 2D mask embeddings. Thanks to the generative pretraining process with large-scale text-image pairs, the 2D mask embeddings are semantically rich. The model leverages the salient patterns in RGB images to produce the 2D salient masks. The 3D branch takes the point cloud and the 2D mask embeddings as inputs. The 2D mask embeddings are used as linear classifiers to assign class probabilities to each of 3D features output from the 3D branch, resulting in a 3D probabilistic mask termed as geometric masks. To predict the per-point semantic class, the model first computes the per-mask category logits for both salient and geometric masks. Then we ensemble the per-mask logits for those two types of masks. In the way, the model can learn salient patterns from the RGB images and geometric information from the point clouds.", + "bbox": [ + 212, + 454, + 789, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 2D Semantic Understanding Model", + "text_level": 1, + "bbox": [ + 215, + 690, + 553, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One challenge of 3D semantic understanding is the severe scarcity of 3D point clouds with groundtruth labels. To tackle the challenge brought by limited training data, vision-language foundation models have been used to transfer semantically-rich 2D features into the 3D space [37,62,79]. [79] used on a model trained on labeled 3D data to produce class-agnostic masks, and then pooled the corresponding 2D representations as the mask embeddings. On the other hand, [62] proposed to leverage a pre-trained 2D semantic segmentation model as feature extractor to perform open-vocabulary 3D segmentation, and no ground", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "truth 3D annotations are needed during training. In this work, we follow the setting in [62] to reduce the 3D annotation efforts.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The 2D segmentation model consists of an image backbone $\\phi$ which is a foundation model pretrained on large-scale text-image pairs; and a segmentation head $\\sigma$ to predict the semantic embedding. There are multiple design options for the 2D backbone $\\phi$ and segmentation head $\\sigma$ . (1) The 2D backbone could either be contrastively pretrained or generatively pre-trianed. The popular frameworks for contrastive representation learning include CLIP [65] and ALIGN [40]. On the other hand, a few works [85,87,92] have demonstrated promising performance by using generatively pre-trained representations for perception task. The feature representations from Diffusion U-Net blocks are extracted for different downstream tasks.", + "bbox": [ + 212, + 178, + 785, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once feature representations from text-image foundation models are extracted, a segmentation head is added upon those features to predict the perpoint semantic classes. The segmentation problem could be formulated as pixel-based classification or mask-based classification. For pixel-based classification [25,47], the intermediate output of segmentation head is of shape $H \\times W \\times C$ , where $H$ and $W$ is image height and width, and $C$ is the dimension of feature embedding. For mask-based classification [10, 11, 87], the segmentation head takes the 2D feature map $\\mathbf{F}^{2d}$ and $N$ fixed mask queries $\\{q_i\\}_{i=1}^N$ as input. The intermediate output is $N$ 2D probabilistic masks $\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N$ and their corresponding mask embeddings $\\{f_i^{2d}\\}_{i=1}^N$ .", + "bbox": [ + 212, + 332, + 787, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we choose diffusion model as the feature backbone $\\phi$ , considering its strong localization ability brought by generative pre-training. Besides, we leverage mask-based segmentation head for its intrinsic nature that decouples mask and its semantic representations. This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models, and further learn geometrically-accurate masks from the 3D branch.", + "bbox": [ + 212, + 484, + 787, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Geometry-Aware 3D Mask Model", + "text_level": 1, + "bbox": [ + 214, + 608, + 540, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While mask-based segmentation has achieved promising performance in fully-supervised setting [10, 11, 73], it has been rarely explored to transfer the learned mask-level representations into another domain. On the other hand, the point-based feature representations from 2D foundation model can be naively distilled by minimizing the per-point feature distance. For example, [62] proposed to train a 3D model to predicts 3D semantic meaningful features by distilling pixel aligned 2D features. However, similar methods are not applicable in our proposed method. First of all, our 2D semantic understanding model uses a mask-based segmentation head which does not provide semantically-rich features in the pixel level. Secondly, the backbone of our 2D semantic understanding model is a frozen stable diffusion model [68] which is designed to generate realistic images with rich details and not tuned for semantic segmentation tasks. The per-pixel features extracted from it are not feasible to supervise the training of our 3D mask", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model†. In the following, we introduce our proposed mask distillation which is tailored to distill knowledge from the mask-based 2D foundation model to the geometry-aware 3D mask model.", + "bbox": [ + 212, + 145, + 782, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The mask-based 2D foundation model predicts $N$ 2D probabilistic masks $\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N$ and their corresponding mask embeddings $\\{f_i^{2d}\\}_{i=1}^N$ . Specifically, $\\mathcal{B}_i^{2d}$ represents a probabilistic map whose elements represent the probability of the corresponding pixel being foreground. We first compute the pixel-point correspondence following [62]. Subsequently, a set of 3D probabilistic masks $\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N$ can be generated by lifting the 2D masks $\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N$ to 3D space based on the pixel-point correspondence. We proposed a novel mask distillation which distills information from both 3D probabilistic masks $\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N$ and the corresponding semantic rich mask embeddings $\\{f_i^{2d}\\}_{i=1}^N$ generated from the 2D branch. Specifically, we train a Minkowski network [16] as the 3D mask prediction model to generate geometry-aware 3D masks. The 3D point cloud is quantized into voxels by averaging the pixels within each voxel to save memory and reduce computes. The 3D mask prediction model generates a 3D feature to represent each voxel and this feature is assigned to all points within the voxel. This produces a full feature map $\\mathbf{F}^{3d} \\in \\mathbb{R}^{M \\times D}$ for the point cloud, where $D$ is the dimension of the 3D feature. The semantic rich 2D mask embeddings $\\{f_i^{2d}\\}_{i=1}^N$ are used as linear classifiers to compute the logits $\\mathcal{S}_i \\in \\mathbb{R}^M$ of a 3D feature belonging to the corresponding class:", + "bbox": [ + 212, + 191, + 787, + 462 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i} = \\left\\langle \\mathbf {F} ^ {3 d}, f _ {i} ^ {2 d} \\right\\rangle , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 462, + 785, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\langle \\cdot, \\cdot \\rangle$ denotes inner product. The 3D probabilistic mask $\\mathcal{B}_i^{3d}$ is then generated by applying the sigmoid function on $S_i$ . We propose a multimodal mask distillation loss to train our 3D mask generator:", + "bbox": [ + 212, + 487, + 787, + 532 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} 1 - \\cos \\left(\\mathcal {B} _ {i} ^ {\\prime 3 d}, \\mathcal {B} _ {i} ^ {3 d}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 546, + 785, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The distillation loss aims at forcing the 2D and 3D branch to make consistent predictions. It serves as an implicit distillation objective to make the 3D model learn high-resolution, semantically-rich feature representations.", + "bbox": [ + 212, + 597, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Open-Vocabulary Inference", + "text_level": 1, + "bbox": [ + 214, + 666, + 488, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During inference, $Diff2Scene$ takes a 3D point cloud and its multiview 2D images as inputs. The 2D semantic understanding model consumes the 2D images and generates a set of 2D probabilistic masks $\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N$ with their corresponding mask embeddings $\\{f_i^{2d}\\}_{i=1}^N$ , where $f_i^{2d} \\in \\mathbb{R}^D$ . The 3D mask model takes the 3D point cloud and the mask embeddings $\\{f_i^{2d}\\}_{i=1}^N$ as inputs to predict the 3D probabilistic mask $\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N$ . To ground a semantic label $c$ to the 3D point cloud, we first apply the same idea from [87] to compute the geometric mean (denoted as $p_i^c$ ) of label probabilities from diffusion and discriminative models for each 2D", + "bbox": [ + 212, + 691, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "†The 3D mask model trained to distill these features does not converge.", + "bbox": [ + 230, + 825, + 710, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mask $\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N$ . Next, the label probabilities $\\mathbf{p}^c$ are assigned to 3D points via the following equation:", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {c} = \\lambda \\sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \\mathcal {B} _ {i} ^ {3 d} + (1 - \\lambda) \\sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \\mathcal {B} _ {i} ^ {\\prime 3 d}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 188, + 784, + 228 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda = 0.5$ . When multiple labels can be assigned to a 3D point, the label with the highest probability from Eq. 3 is taken.", + "bbox": [ + 214, + 239, + 782, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 292, + 366, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct a series of experiments to demonstrate the effectiveness of Diff2Scene on a variety of zero-shot 3D scene understanding benchmarks. We first evaluate the proposed model on zero-shot open-vocabulary semantic segmentation tasks following the evaluation protocol of [62]. We then perform comprehensive ablation studies to validate our designs. Finally, we qualitatively demonstrate the strong ability of the proposed model for open-vocabulary 3D segmentation and grounding complicated compositional text queries.", + "bbox": [ + 212, + 324, + 782, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 452, + 334, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use ScanNet [17], Matterport3D [7], ScanNet200 [69] and Replica [77] for the open-vocabulary 3D semantic segmentation task. We provide qualitative analysis of the visual grounding task on Nr3D [1]. Except for Replica, point clouds and multi-view images in the training split without ground truth annotations are used for model training. As Replica does not provide the training data, we perform training on ScanNet and perform evaluation on Replica, following the setting in [79].", + "bbox": [ + 212, + 477, + 782, + 584 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ScanNet is one of the largest 3D semantic segmentation dataset. It provides 80,554 images from 1201 scans for training and 21,300 images from 312 scans for testing with 20 semantic labels.", + "bbox": [ + 214, + 598, + 782, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matterport3D is a large scale RGB-D dataset containing 10,800 panoramic views from 194,000 RGB-D images of 90 building-scale scenes. It splits 61 scenes for training, 11 scenes for validation and 18 for testing. We train our 3D branch using the images in the training splits and report the results on test split.", + "bbox": [ + 214, + 657, + 782, + 719 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ScanNet200 has 200 semantic labels with long-tailed classes. It also provides a grouping of the 200 categories based on the number of labeled surface points in the training set, resulting in 3 subsets: head, common, and tail. This enables us to evaluate the performance of our method on the long-tail distribution, making ScanNet200 a natural choice as an evaluation dataset. We report the mean intersection over union (mIoU) metric on the validation set consisting of 312 scenes following the split in [62, 69, 78].", + "bbox": [ + 212, + 733, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 243, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/bb3d87afa40798d3f1c2230ffbbbb4a161f40ccd3d71020af9636ea48cb16bad.jpg", + "table_caption": [ + "Table 1: Comparison to state-of-the-art models. We report mIoU for all benchmarks. Best results in zero-shot, open-vocabulary setting are shown in bold." + ], + "table_footnote": [], + "table_body": "
ScanNetMatterport3DScanNet200Replica
AllAllHeadCommonTailAllHeadTailAll
Fully-supervised
TangentConv [81]40.9--------
TextureNet [34]54.8--------
SFSS-MMSI [16]-35.9-------
CSC-Pretrain [30]--45.517.17.924.9---
SupCon [93]--48.619.210.326.0---
LGround [69]--48.518.410.627.2---
MinkowskiNet [16]69.054.246.315.410.225.3---
Zero-shot, open-vocabulary
MSeg Voting [44]31.033.4-------
ConceptFusion [37]33.3-17.56.32.88.811.63.54.6
OpenMask3D [79]34.0-19.67.54.510.513.23.44.8
OpenScene (2D) [62]41.432.421.910.85.512.733.411.514.5
OpenScene (3D) [62]46.041.317.60.00.06.332.67.711.1
OpenScene (2D/3D) [62]47.542.620.09.75.111.634.211.914.9
Diff2Scene (Ours)48.645.525.611.56.914.246.212.917.5
", + "bbox": [ + 215, + 184, + 785, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Replica contains 51 categories, and we further split those categories into head and tail sets based on their appearance frequency. We report the mIoU on the office0, office1, office2, office3, office4, room0, room1, and room2.", + "bbox": [ + 215, + 465, + 784, + 512 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$Nr3D$ is a 3D visual grounding dataset which contains diverse text prompts. To further evaluate the ability of our model to distinguish between objects in the same class but with different attributes, we perform qualitative evaluation on the visual grounding dataset $Nr3D$ [1]. We perform zero-shot evaluation on the validation set without training on any labeled data for the visual grounding task.", + "bbox": [ + 212, + 529, + 784, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Baseline Methods", + "text_level": 1, + "bbox": [ + 215, + 645, + 408, + 659 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare Diff2Scene with the current state-of-the-art fully-supervised 3D semantic segmentation models including TangentConv [81], TextureNet [34], SFSS-MMSI [16], CSC-Pretrain [30], SupCon [93], LGround [69] and MinkowskiNet [16] on the 3D semantic segmentation benchmark. We also compare our model against OpenScene [62] and ConceptFusion [37], the recently proposed open-vocabulary 3D semantic understanding model. For OpenScene [62], we compare with its OpenSeg [25] variant which has the same feature and pretrained datasets for a fair comparison. We also compare our model with its three different variants (2D Fusion, 3D Distill, and 2D/3D Ensemble). Besides, we adapt the state-of-the-art 3D instance segmentation model OpenMask3D [79] for comparison on the 3D semantic segmentation benchmark.", + "bbox": [ + 212, + 672, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 146, + 455, + 161 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We use posed multi-view RGB images and 3D point clouds for all the datasets. ODISE [87], which consists of a diffusion backbone and mask-based segmentation head, is used as the model in our 2D branch. It uses a stable diffusion model [68] pre-trained on Laion-5B [72] as the feature backbone. The dimensions for diffusion and CLIP features are 256 and 768 respectively. The number of queries of Mask2Former [10] is 100. Similar to OpenScene [62], we use MinkowskiNet18A [16] as the model in our 3D branch to extract 3D features from the 3D point clouds. Our 3D model is trained for 200 epochs with a batch size of 8. Adam optimizer [41] is used with a learning rate of 0.0001 and polynomial learning rate policy is used as the learning rate scheduler with power 0.9. During inference, text-embeddings are computed by the ViT-L/14 CLIP model [65] for each of the semantic categories and grounding queries. We use the same pre-processing step and pre-trained dataset as OpenScene [62] (OpenSeg [25]) for a fair comparison.", + "bbox": [ + 212, + 172, + 787, + 369 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 391, + 431, + 405 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation on zero-shot 3D semantic segmentation. We first compare our method with the state-of-the-art open-vocabulary scene understanding models and fully-supervised 3D segmentation models. We report the mIoU for Scannet, Matterport3D, Scannet200, and Replica in Table 1. We find that our method achieves better results than the state-of-the-art open-vocabulary models and their variants on all the benchmarks. Besides, although our zero-shot model has noticeable performance drop compared with fully-supervised model, the gap of tail categories between the proposed method and those methods are relatively small (e.g. 6.9 vs. 7.9 from CSC-Pretrain) on Scannet200. This demonstrates the strong potential of the proposed method for long-trailed 3D semantic segmentation tasks.", + "bbox": [ + 212, + 416, + 787, + 582 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generalization to unseen dataset. To test the generalization ability of our proposed model, we evaluate it on an unseen dataset Replica [77] and report the results in Table 1. The results shown that our proposed method significantly outperforms the state-of-the-art models on head, tail and all categories in Replica. This demonstrates the strong generalization ability of the proposed model on novel datasets.", + "bbox": [ + 212, + 583, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of Different Distillation Settings. We compare our mask-based distillation method with point-based ones under different settings and report the performance of the 3D branch on Replica [77] in Table 2. The supervisions for point-based method include: (1) Fine-tuned CLIP feature, which follows the same setting as OpenScene [62]; (2) Frozen diffusion feature extracted from the last layer of diffusion U-Net block. We observe that distilling frozen diffusion features does not converge. Our proposed method, by introducing the semantic meaningful mask embedding output from the 2D branch as a fixed classifier, significantly boost the performance of the 3D branch.", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation studies. We conduct ablation studies using the Replica dataset [77] and show the results in Table 3. We first analyze the effectiveness of combining", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dee1d4f789cb5ea660d6a875954805b99953c20b16edbc575d5ee334339b957b.jpg", + "table_caption": [ + "Table 2: Effectiveness of Different Distillation Settings. We report mIoU of different methods on the Replica [77] dataset." + ], + "table_footnote": [], + "table_body": "
SettingDistillation TypeHead Tail All
fine-tuned CLIP feature [62]Point-based32.6 7.7 11.1
frozen diffusion featurePoint-basedDivergence
multimodal mask distillation (ours)Mask-based43.3 8.0 12.8
", + "bbox": [ + 264, + 184, + 733, + 258 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4c9796ec18757c702295c8c1f61fcb8a6f206f0e8c73600c79cfc8d631a802ae.jpg", + "table_caption": [ + "Table 3: Performance of different model ablations. We observe that each component of our model gains consistent improvements." + ], + "table_footnote": [], + "table_body": "
MethodmIoU
Our full model17.5
Without 2D (salient) mask12.8
Without 3D (geometric) mask16.5
Without discriminative (CLIP) features15.5
Without generative (Stable Diffusion) features15.3
", + "bbox": [ + 320, + 311, + 678, + 412 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "2D and 3D masks using equation 3. We observe that compared with using salient or geometric mask only, using both types of masks achieves the best performance. This is intuitive as both salient patterns and geometric information are helpful to segment accurate class boundaries. We then analyze the effectiveness of different semantic features. We find that discriminative and diffusion features serve as strong complementary to each other. We also observe that using those two types of semantic features jointly can significantly outperform using any of them alone.", + "bbox": [ + 212, + 441, + 787, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Qualitative analysis", + "text_level": 1, + "bbox": [ + 215, + 585, + 426, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualizations of zero-shot semantic segmentation. In Fig. 4, we provide qualitative analysis of our approach and OpenScene for the zero-shot 3D semantic segmentation task. Compared with OpenScene, our model generates coherent and consistent masks (e.g., the table mask in first column and the bed mask in third column) thanks to the mask-instance representations. It predicts accurate semantic labels for both head and tail categories by leveraging both CLIP and diffusion features.", + "bbox": [ + 212, + 612, + 787, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualizations of visual grounding results. We provide qualitative analysis of our approach and OpenScene for the zero-shot visual grounding task in Fig. 5. We observe that our model can accurately identify objects given complicated text queries. It demonstrates that the proposed method, Diff2Scene, has good capability at the following types of queries. Fig. 5 (a) describes object shape and color, and even in comparative degree (It's the shorter, red box); Fig. 5 (b) describes a rare object (rack) and its surrounded object with surface appearance descriptions (wrinkled towel); Fig. 5 (c) describes the relative location of the", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e4d887f69930dad3a32fee80548a629c70118e6b949e5dac3bd27c4ef47ec20f.jpg", + "image_caption": [ + "Fig. 4: Qualitative results from our model and OpenScene on zero-shot semantic segmentation. We visualize the segmentation results on the validation set of ScanNet200 [69]. We observe that our model can predict coherent masks with accurate semantic labels compared to OpenScene for both head and tail categories." + ], + "image_footnote": [], + "bbox": [ + 240, + 142, + 754, + 448 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "object (next to the desk); Fig. 5 (d) describes the usage of the object (recycling). In addition, we can see that given vague usage descriptions without common category names like trash bin in the text prompts, the model can still accurately identify the object.", + "bbox": [ + 212, + 546, + 787, + 607 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 628, + 359, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we investigate the problem of leveraging frozen representations from large text-to-image diffusion models for open-vocabulary 3D semantic understanding. Diff2Scene sets a new state-of-the-art in the zero-shot 3D semantic segmentation task and shows promising performance in the visual grounding task. Our method also shows outstanding generalization ability towards unseen datasets and novel text queries. It provides a new way to effectively leverage generative text-to-image foundation models for 3D semantic scene understanding tasks.", + "bbox": [ + 212, + 657, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "There are several limitations of the proposed model. First, while our model achieves better performance compared to existing methods in small objects, it still misclassified some small and rare categories (e.g. rail). Second, we observe that the model can be easily confused by fine-grained categories that with similar", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f1777dd22d3990c8e5f12ffe403ef29e25d42cfbba22c91322224bfdfbb4e05b.jpg", + "image_caption": [ + "Fig. 5: Qualitative results from our model and OpenScene on zero-shot visual grounding. Our open-vocabulary semantic understanding model is capable of handling different types of novel and compositional queries. Novel object classes as well as objects described by colors, shapes, appearances, locations, and usages are successfully retrieved by our method. Note that the located points are colored in yellow." + ], + "image_footnote": [], + "bbox": [ + 225, + 141, + 781, + 529 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "semantic meaning. For example, the model sometimes wrongly classifies points of windowsill to the window class. In future work, it will be interesting to design models that can accurately distinguish between fine-grained categories in the open-vocabulary setting.", + "bbox": [ + 212, + 650, + 787, + 712 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 756, + 403, + 773 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We sincerely thank Amir Hertz, Weicheng Kuo, and Lijun Yu for the helpful discussions.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.J.: ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020) 3, 9, 10", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 2, 4", + "3. Anand, A., Koppula, H.S., Joachims, T., Saxena, A.: Contextually guided semantic labeling and search for 3d point clouds. In: IJRR (2011) 3", + "4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3", + "5. Atzmon, M., Maron, H., Lipman, Y.: Point convolutional neural networks by extension operators. ACM Transactions on Graphics (2018) 3", + "6. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022) 5", + "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017) 3, 9", + "8. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. arXiv preprint arXiv:2209.09874 (2022) 4", + "9. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. In: ICCV (2023) 5", + "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022) 2, 3, 7, 11", + "1. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) 7", + "2. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: WACV (2020) 4", + "3. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. In: BMVC (2019) 4", + "4. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. IJCV (2022) 4", + "5. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: MVA (2019) 4", + "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019) 2, 3, 8, 10, 11", + "7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9", + "8. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. arXiv preprint arXiv:2308.00353 (2023) 5", + "9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 4, 5", + "20. Engelmann, F., Kontogianni, T., Hermans, A., Leibe, B.: Exploring spatial context for 3d semantic segmentation of point clouds. In: ICCV workshop (2017) 3" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Engelmann, F., Kontogianni, T., Schult, J., Leibe, B.: Know what your neighbors do: 3d semantic segmentation of point clouds. In: ECCV workshop (2019) 3", + "22. Fan, J., Zheng, P., Li, S.: Vision-based holistic scene understanding towards proactive human-robot collaboration. Robotics and Computer-Integrated Manufacturing 75, 102304 (2022) 2", + "23. Feng, R., Gao, Y., Tse, T.H.E., Ma, X., Chang, H.J.: Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In: ICCV (2023) 5", + "24. Gadre, S.Y., Wortsman, M., Ilharco, G., Schmidt, L., Song, S.: Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In: CVPR (2023) 4", + "25. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022) 2, 4, 5, 7, 10, 11", + "26. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018) 3", + "27. Han, L., Zheng, T., Zhu, Y., Xu, L., Fang, L.: Live semantic 3d perception for immersive augmented reality. IEEE Trans. visualization and computer graphics 26(5), 2012-2022 (2020) 2", + "28. He, Q., Peng, J., Jiang, Z., Wu, K., Ji, X., Zhang, J., Wang, Y., Wang, C., Chen, M., Wu, Y.: Unim-ov3d: Uni-modality open-vocabulary 3d scene understanding with fine-grained feature representation. IJCAI (2024) 4", + "29. Holmquist, K., Wandt, B.: Diffpose: Multi-hypothesis human pose estimation using diffusion models. In: ICCV (2023) 5", + "30. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: CVPR (2021) 10", + "31. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: ICCV (2021) 3", + "32. Hua, B.S., Tran, M.K., Yeung, S.K.: Pointwise convolutional neural networks. In: CVPR (2018) 3", + "33. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: ICRA (2023) 4", + "34. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: CVPR (2019) 10", + "35. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: CVPR (2022) 2", + "36. Huang, Z., Lv, C., Xing, Y., Wu, J.: Multi-modal sensor fusion-based deep neural network for end-to-end autonomous driving with scene understanding. IEEE Sensors Journal 21(10), 11781-11790 (2020) 2", + "37. Jatavallabhula, K., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., Tenenbaum, J., de Melo, C., Krishna, M., Paull, L., Shkurti, F., Torralba, A.: Conceptfusion: Open-set multimodal 3d mapping. In: Robotics science and systems (2023) 2, 4, 6, 10", + "38. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. In: ICCV (2023) 5", + "39. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 2, 4", + "40. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 7" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 11", + "42. Koppula, H., Anand, A., Joachims, T., Saxena, A.: Semantic labeling of 3d point clouds for indoor scenes. In: NeurIPS (2011) 3", + "43. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023) 2, 5", + "44. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: MSeg: A composite dataset for multi-domain semantic segmentation. In: CVPR (2020) 10", + "45. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs. In: CVPR (2018) 3", + "46. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: ICCV (2023) 5", + "47. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) 2, 4, 5, 7", + "48. Li, J., Selvaraju, R.R., Gotmare, A.D., Joty, S.R., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021) 2, 4", + "49. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018) 3", + "50. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023) 5", + "51. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023) 4", + "52. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2022) 4", + "53. Liu, D., Li, Q., Dinh, A.D., Jiang, T., Shah, M., Xu, C.: Diffusion action segmentation. In: ICCV (2023) 5", + "54. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., Saddik, A.E., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: NeurIPS (2023) 3", + "55. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022) 2", + "56. Liu, Z., Qi, X., Fu, C.W.: 3d-to-2d distillation for indoor scene parsing. In: CVPR (2021) 3", + "57. Lu, Y., Rasmussen, C.: Simplified markov random fields for efficient semantic labeling of 3d point clouds. In: ICIRS (2012) 3", + "58. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2023) 2", + "59. Mazur, K., Sucar, E., Davison, A.: Feature-realistic neural fusion for real-time, open set scene understanding. In: ICRA (2023) 4", + "60. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3D point cloud. In: 3DV (2021) 2, 4, 5", + "61. Mittal, S., Abstreiter, K., Bauer, S., Scholkopf, B., Mehrjou, A.: Diffusion based representation learning. In: ICML (2023) 2, 5", + "62. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12", + "63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) 3" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 3", + "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 5, 7, 11", + "66. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2", + "67. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Linguagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022) 2", + "68. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 5, 6, 7, 11", + "69. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 3, 4, 9, 10, 13", + "70. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 5", + "71. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 4", + "72. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022) 2, 11", + "73. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 2, 7", + "74. Shafiullah, N.M.M., Paxton, C., Pinto, L., Chintala, S., Szlam, A.: CLIP-fields: Weakly supervised semantic fields for robotic memory. In: CoRL Workshop on Language and Robotics (2022) 4", + "75. Shah, D., Osinski, B., Ichter, B., Levine, S.: LM-nav: Robotic navigation with large pre-trained models of language, vision, and action. In: CoRL (2022) 4", + "76. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV (2023) 5", + "77. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) 3, 9, 11, 12", + "78. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 9", + "79. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023) 2, 3, 4, 6, 9, 10", + "80. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023) 2", + "81. Tatarchenko, M., Park, J., Koltun, V., Zhou., Q.Y.: Tangent convolutions for dense prediction in 3D. CVPR (2018) 10" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "82. Tchapmi, L.P., Choy, C.B., Armeni, I., Gwak, J., Savarese, S.: Segcloud: Semantic segmentation of 3d point clouds. In: 3DV (2017) 3", + "83. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019) 3", + "84. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: ICCV (2023) 5", + "85. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023) 7", + "86. Wang, T., Li, J., An, X.: An efficient scene semantic labeling approach for 3d point cloud. In: ITSC (2015) 3", + "87. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) 2, 3, 5, 7, 8, 11", + "88. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020) 2", + "89. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023) 2, 5", + "90. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023) 4", + "91. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR (2022) 4", + "92. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023) 7", + "93. Zheng, M., Wang, F., You, S., Qian, C., Zhang, C., Wang, X., Xu, C.: Weakly supervised contrastive learning. In: ICCV (2021) 10", + "94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022) 2", + "95. Zhu, X., Chen, J., Zeng, X., Liang, J., Li, C., Liu, S., Behpour, S., Xu, M.: Weakly supervised 3d semantic segmentation using cross-image consensus and inter-voxel affinity relations. In: ICCV (2021) 3" + ], + "bbox": [ + 215, + 146, + 784, + 575 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_model.json b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed50120efded73b6ccc20ed2b17d8f7ac124003c --- /dev/null +++ b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_model.json @@ -0,0 +1,2735 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.213, + 0.764, + 0.246 + ], + "angle": 0, + "content": "Xiaoyu Zhu\\(^{1\\dagger}\\), Hao Zhou\\(^{2}\\), Pengfei Xing\\(^{2}\\), Long Zhao\\(^{2}\\), Hao Xu\\(^{3}\\), Junwei Liang\\(^{4}\\), Alexander Hauptmann\\(^{1}\\), Ting Liu\\(^{2}\\), and Andrew Gallagher\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.255, + 0.603, + 0.27 + ], + "angle": 0, + "content": "1 Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.27, + 0.57, + 0.284 + ], + "angle": 0, + "content": "2 Google DeepMind" + }, + { + "type": "text", + "bbox": [ + 0.471, + 0.284, + 0.534, + 0.298 + ], + "angle": 0, + "content": "3 Google" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.298, + 0.537, + 0.31 + ], + "angle": 0, + "content": "HKUST" + }, + { + "type": "list", + "bbox": [ + 0.468, + 0.27, + 0.57, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.355, + 0.356, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.355, + 0.503, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.355, + 0.638, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.418, + 0.626, + 0.434 + ], + "angle": 0, + "content": "\"This backpack is alongside a square green chair.\"" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.355, + 0.78, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.419, + 0.766, + 0.434 + ], + "angle": 0, + "content": "Find the white sneakers that are closer to the desk chair." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Fig. 1: Illustration of open-vocabulary 3D semantic scene understanding. We propose Diff2Scene, a 3D model that performs open-vocabulary semantic segmentation and visual grounding tasks given novel text prompts, without relying on any annotated 3D data. By leveraging discriminative-based and generative-based 2D foundation models, Diff2Scene can handle a wide variety of novel text queries for both common and rare classes, like \"desk\" and \"soap dispenser\". It can also handle compositional queries, such as \"find the white sneakers that are closer to the desk chair.\"" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.592, + 0.767, + 0.745 + ], + "angle": 0, + "content": "Abstract. In this paper, we investigate the use of diffusion models which are pre-trained on large-scale image-caption pairs for open-vocabulary 3D semantic understanding. We propose a novel method, namely Diff2Scene, which leverages frozen representations from text-image generative models, along with salient-aware and geometric-aware masks, for open-vocabulary 3D semantic segmentation and visual grounding tasks. Diff2Scene gets rid of any labeled 3D data and effectively identifies objects, appearances, materials, locations and their compositions in 3D scenes. We show that it outperforms competitive baselines and achieves significant improvements over state-of-the-art methods. In particular, Diff2Scene improves the state-of-the-art method on ScanNet200 by \\(12\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.759, + 0.741, + 0.786 + ], + "angle": 0, + "content": "Keywords: 3D Semantic Understanding \\(\\cdot\\) Open-Vocabulary Perception \\(\\cdot\\) Diffusion Model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.786, + 0.841 + ], + "angle": 0, + "content": "This work was partially done while the author was a student researcher at Google." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.266 + ], + "angle": 0, + "content": "3D semantic scene understanding, with the task of assigning semantics to every 3D point, plays a fundamental role in many computer vision applications, such as robotics [88], autonomous driving [36], human-computer interaction [22], and augmented reality [27]. Traditional studies in this field usually target solving this problem in a closed-set fashion [16,73], resulting in models that can only be used to make predictions within the predefined label space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.787, + 0.493 + ], + "angle": 0, + "content": "Recent progress in computer vision have witnessed the emerging interests in solving semantic understanding tasks in open-vocabulary settings [35, 62, 67, 78, 94]. In contrast to closed-set setting, models targeting open-vocabulary tasks must make predictions for any semantics described in text, including object category and fine-grained attributes (e.g., shape, color, material, property) as well as their complicated compositions. However, this is a challenging task due to the wide diversity and complexity of possible queries. Motivated by the advance of aligning text and image embeddings with large-scale foundation models [2, 39, 48, 65], existing methods mitigate this challenge by lifting the image features from foundation models such as CLIP [65] or their descendants [25, 47] to 3D. These lifted feature representations for 3D points can then be used to query with open-vocabulary descriptions, achieving semantic understanding in 3D. Despite these achievements, contrastively trained CLIP-based models exhibit limitations in handling fine-grained classes [66] and novel compositional text queries [58], restricting their performance in open-vocabulary 3D semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.689 + ], + "angle": 0, + "content": "The recently developed text-to-image diffusion models have shown outstanding abilities for image generation even with challenging text prompts [55,68,72], such as combinational descriptions with multiple attributes (e.g., A bucket bag made of blue suede with intricate golden paisley patterns.) The internal visual representation of these models, entangled with text embedding through cross-attention, have proven correlate well with semantic concepts described by language [43, 60, 61, 89]. On the other hand, CLIP-based foundation models have been shown to struggle with compositionality [58]. Moreover, compared with the CLIP model which is optimized for global representation, diffusion models have proven to be superior at local representation [80], which is a key for dense prediction tasks. Specifically, ODISE [87] applied the internal representations of Stable Diffusion [68] to open-vocabulary 2D semantic understanding tasks and achieved promising results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "One of the key challenges in 3D perception is the severe scarcity of point clouds and their dense labels. Several existing methods have been proposed to solve the lack of data issue in a zero-shot fashion by leveraging the CLIP model pre-trained on large-scale text-image data [37,62,79]. The prior art [62] extracts dense CLIP features from 2D images and distill the knowledge of their lifted 3D counterpart into a 3D mask predictor. However, CLIP features, as discussed above, struggle to handle fine-grained classes [66] and show worse localization capability compared with diffusion features. We leverage diffusion model as feature backbone along with a mask-based segmentation head (e.g., Mask2Fromer [10]) for its intrinsic nature that decouples mask and its semantic representations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models and further learning geometrically-accurate masks from the 3D branch. However, performing multi-modal distillation with mask-based segmentation head is a non-trivial task. The frozen features extracted from the decoder of the U-Net in the diffusion model are trained with generative objectives, and cannot be directly used for the perception task. Therefore, directly distilling knowledge from these features as normally done in prior art [54,56,62] is infeasible. Another intuitive way is to leverage a supervised 3D mask proposal network and pool the feature representations from 2D CLIP features for each mask [79]. However, the training of 3D mask proposal network requires labeled 3D data, which may not be feasible in practice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.433 + ], + "angle": 0, + "content": "To mitigate these issues, we propose a novel mask distillation method tailored to distill knowledge from the Mask2Former style 2D branch [10, 87] to the 3D branch, which is shown in Fig. 2. Specifically, we design our 3D branch to take a 3D point cloud as input and to predict their 3D features. The semantically meaningful mask embeddings produced from our 2D branch are used as linear classifiers to assign class probability to these 3D features. Their corresponding 2D masks are lifted to 3D based on pixel-point correspondence and used to force the consistency learning of the 2D and 3D branch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.526 + ], + "angle": 0, + "content": "We evaluate Diff2Scene quantitatively on ScanNet [17], ScanNet200 [69], Matterport 3D [7] and Replica [77] for open-vocabulary 3D semantic segmentation and qualitatively on \\(\\mathrm{Nr3D}\\) [1] for visual grounding tasks. Our experimental results show that Diff2Scene outperforms state-of-the-art models [62] on all the four semantic segmentation datasets and achieves promising results on visual grounding tasks. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.534, + 0.784, + 0.561 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to leverage text-image diffusion to perform open-vocabulary 3D semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.564, + 0.784, + 0.605 + ], + "angle": 0, + "content": "- We propose a novel mask distillation method to train a 3D mask prediction model by distilling knowledge from the Mask2Former style 2D segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.609, + 0.784, + 0.637 + ], + "angle": 0, + "content": "- The proposed method achieves state-of-the-art performance on several open-vocabulary 3D semantic segmentation and visual grounding benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.534, + 0.784, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.673, + 0.388, + 0.689 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Closed-vocabulary 3D semantic segmentation. In 3D semantic segmentation, a semantic category is assigned for each 3D point. It has been long studied [3-5, 20, 21, 26, 31, 32, 42, 45, 49, 57, 63, 64, 82, 86, 95] due to its importance in computer vision and robotics applications. One challenge of this task is that 3D point clouds are not in a regular structured format; network architectures that work well for 2D tasks cannot handle 3D point clouds effectively. As a result, most of the early studies focus on designing effective and efficient network architectures that are suitable for 3D point clouds [16, 20, 21, 26, 31, 32, 45, 63, 64, 83]. This line of work achieved great success and significantly improves the results of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.157, + 0.391, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.244, + 0.365, + 0.263 + ], + "angle": 0, + "content": "(a) Feature-based Distillation w/o labeled 3D data" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.158, + 0.558, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.244, + 0.563, + 0.264 + ], + "angle": 0, + "content": "(b) Mask-based Feature Pooling w/ labeled 3D data" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.157, + 0.813, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.245, + 0.776, + 0.266 + ], + "angle": 0, + "content": "(c) Mask-based Distillation w/o labeled 3D data (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.294, + 0.789, + 0.461 + ], + "angle": 0, + "content": "Fig. 2: Illustration of open-vocabulary 3D perception methods. \\(L_{PD}\\) and \\(L_{MD}\\) denote point-based distillation loss and mask-based distillation loss. \\(M_{3D}\\) denote a set of predicted 3D masks; \\(M_{2D}\\) and \\(Z_{mf}\\) denote a set of predicted 2D masks and their semantic embeddings; \\(Z_{gf}\\) denote the high-resolution 3D feature map. (a) Directly minimizing the per-point feature distance between the CLIP-based model and the tuned 3D model [62]. (b) Directly using a 3D mask proposal network trained on labeled 3D data to produce class-agnostic masks, and then pool corresponding representations from the CLIP feature map [79]. (c) The proposed mask distillation approach, namely Diff2Scene, that uses Stable Diffusion and performs mask-based distillation. Diff2Scene leverages the semantically-rich mask embeddings from 2D foundation models and geometrically accurate masks from the tuned 3D model, and thus achieves superior performance compared to previous methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.669 + ], + "angle": 0, + "content": "3D semantic segmentation. Another challenge is the lack of large scale data with ground truth annotations. Due to the intensive labeling effort and high cost of data annotation [69], the available datasets for 3D semantic segmentation are usually small in scale. In the absence of large scale data, early studies usually target solving this problem in a closed-vocabulary setting, where the trained model can only predict categories that appear during training. To mitigate the scale limitation of existing datasets, a handful of works [12-15, 52, 60, 91] have applied zero-shot learning in 3D scene understanding tasks. [12-15, 91] focused on 3D point classification task and [52, 60] tried to address the 3D semantic segmentation problem. However, these zero-short methods still require ground truth annotations for a certain amount of 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Open-vocabulary 3D segmentation. The recent progress of large-scale vision and language representation learning [2,28,39,48,65,90] has advanced the study of semantic and instance segmentation in an open-vocabulary setting. [25, 47, 51] first explored open-vocabulary 2D semantic segmentation. They proposed aligning per-pixel features [47] or features from mask regions [25, 51] with the corresponding text embedding. Following these works, [8,19,24,33,37,59,62,74, 75] focus on 3D semantic segmentation in an open-vocabulary setting. Among them, [8,24,33,37,59,71,74] project 3D points to 2D images and solve the 3D problem in the 2D space, instead of targeting the 3D open-vocabulary semantic segmentation directly. As [19] pointed out, the projection from 3D to 2D has information loss and the solution is suboptimal." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.786, + 0.448 + ], + "angle": 0, + "content": "To make better use of information from the 3D point cloud, [19], [18] and [62] proposed to directly applying semantic segmentation on the 3D point cloud. [19] and its extension [18] proposed associating captions generated for 2D images to corresponding 3D point clouds to build the pseudo-ground truth captions for 3D point clouds. A neural network is trained to associate the 3D point cloud with these pseudo labels through contrastive loss. Similar to the zero-shot setting, [19] evaluated their model in a leave-one-out fashion, which still requires annotations for 3D point cloud. Inspired by the strong open-vocabulary ability of large-scale vision and language models, Peng et al. [62] proposed distilling knowledge to a 3D point cloud model. They trained 3D semantic segmentation model by only distilling the knowledge from a CLIP-style [65] 2D open-vocabulary semantic segmentation model [25, 47]. They demonstrated that without training with any ground truth labels, the model can achieve great performance on many open-vocabulary tasks. However, we observe that [62] is strongly limited by the 2D open-vocabulary semantic segmentation models used as the teacher. Its performance on rare classes that are not used in training these models are not satisfactory. Our method follows this idea by distilling the knowledge of 2D open-vocabulary semantic segmentation model to a 3D model. In constraint to the approach in [62], we use a diffusion-based 2D open vocabulary semantic segmentation model [87] as the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.45, + 0.787, + 0.631 + ], + "angle": 0, + "content": "Diffusion models for scene understanding. The last few years have witnessed the success of diffusion models in image generation [68, 70]. Recent studies also observed the diffusion models are strong representation learners [43, 60, 61, 89]. As a result, researchers have applied it to many understanding tasks such as image classification [46], object detection [9], image semantic segmentation [6, 38, 87], instance segmentation [50], human pose estimation [23, 29, 76], action segmentation [53], camera pose estimation [84], to name a few, and achieved great success. Especially, [87] and [50] showed that Stable Diffusion [68], whose internal representation being well correlated with text embedding, has strong open-vocabulary abilities for understanding tasks. Inspired by this, we are the first to apply text-to-image diffusion models to open vocabulary 3D semantic segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.658, + 0.355, + 0.673 + ], + "angle": 0, + "content": "3 Diff2Scene" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.694, + 0.793, + 0.74 + ], + "angle": 0, + "content": "We introduce Diff2Scene, an open-vocabulary 3D semantic understanding method. Similar to [62], our proposed model operates in a zero-shot fashion, where no ground truth 3D annotations are needed during training." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.766, + 0.34, + 0.78 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.795, + 0.786, + 0.84 + ], + "angle": 0, + "content": "An overview of Diff2Scene is shown in Fig. 3. It takes posed RGB images and the reconstructed 3D point cloud as model inputs. The model predicts the semantic label for each 3D point. Diff2Scene has two branches. The 2D branch is designed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.787, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.427 + ], + "angle": 0, + "content": "Fig. 3: Overview of our method. We propose Diff2Scene, an open-vocabulary 3D semantic understanding model. Diff2Scene contains two branches. The 2D branch is designed to be a diffusion-based 2D semantic segmentation model. It accepts a 2D image as input and predicts a set of 2D probabilistic masks with corresponding semantically-rich mask embeddings. The 3D branch utilizes the point cloud and 2D mask embeddings as input. The 2D mask embeddings are used as \"semantic queries\" to generate corresponding 3D probabilistic masks. The model learns salient patterns from the RGB images and geometric information from the point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.79, + 0.668 + ], + "angle": 0, + "content": "to be an open-vocabulary 2D semantic segmentation model. It leverages text-to-image generative model [68] which is pre-trained on massive text-image pairs. The model takes a 2D image as input to predict a set of 2D probabilistic masks with their corresponding 2D mask embeddings. Thanks to the generative pretraining process with large-scale text-image pairs, the 2D mask embeddings are semantically rich. The model leverages the salient patterns in RGB images to produce the 2D salient masks. The 3D branch takes the point cloud and the 2D mask embeddings as inputs. The 2D mask embeddings are used as linear classifiers to assign class probabilities to each of 3D features output from the 3D branch, resulting in a 3D probabilistic mask termed as geometric masks. To predict the per-point semantic class, the model first computes the per-mask category logits for both salient and geometric masks. Then we ensemble the per-mask logits for those two types of masks. In the way, the model can learn salient patterns from the RGB images and geometric information from the point clouds." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.555, + 0.708 + ], + "angle": 0, + "content": "3.2 2D Semantic Understanding Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "One challenge of 3D semantic understanding is the severe scarcity of 3D point clouds with groundtruth labels. To tackle the challenge brought by limited training data, vision-language foundation models have been used to transfer semantically-rich 2D features into the 3D space [37,62,79]. [79] used on a model trained on labeled 3D data to produce class-agnostic masks, and then pooled the corresponding 2D representations as the mask embeddings. On the other hand, [62] proposed to leverage a pre-trained 2D semantic segmentation model as feature extractor to perform open-vocabulary 3D segmentation, and no ground" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "truth 3D annotations are needed during training. In this work, we follow the setting in [62] to reduce the 3D annotation efforts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.787, + 0.329 + ], + "angle": 0, + "content": "The 2D segmentation model consists of an image backbone \\(\\phi\\) which is a foundation model pretrained on large-scale text-image pairs; and a segmentation head \\(\\sigma\\) to predict the semantic embedding. There are multiple design options for the 2D backbone \\(\\phi\\) and segmentation head \\(\\sigma\\). (1) The 2D backbone could either be contrastively pretrained or generatively pre-trianed. The popular frameworks for contrastive representation learning include CLIP [65] and ALIGN [40]. On the other hand, a few works [85,87,92] have demonstrated promising performance by using generatively pre-trained representations for perception task. The feature representations from Diffusion U-Net blocks are extracted for different downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.485 + ], + "angle": 0, + "content": "Once feature representations from text-image foundation models are extracted, a segmentation head is added upon those features to predict the perpoint semantic classes. The segmentation problem could be formulated as pixel-based classification or mask-based classification. For pixel-based classification [25,47], the intermediate output of segmentation head is of shape \\( H \\times W \\times C \\), where \\( H \\) and \\( W \\) is image height and width, and \\( C \\) is the dimension of feature embedding. For mask-based classification [10, 11, 87], the segmentation head takes the 2D feature map \\( \\mathbf{F}^{2d} \\) and \\( N \\) fixed mask queries \\( \\{q_i\\}_{i=1}^N \\) as input. The intermediate output is \\( N \\) 2D probabilistic masks \\( \\{\\mathcal{B}_i^{2d}\\}_{i=1}^N \\) and their corresponding mask embeddings \\( \\{f_i^{2d}\\}_{i=1}^N \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.577 + ], + "angle": 0, + "content": "In this work, we choose diffusion model as the feature backbone \\(\\phi\\), considering its strong localization ability brought by generative pre-training. Besides, we leverage mask-based segmentation head for its intrinsic nature that decouples mask and its semantic representations. This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models, and further learn geometrically-accurate masks from the 3D branch." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.609, + 0.542, + 0.624 + ], + "angle": 0, + "content": "3.3 Geometry-Aware 3D Mask Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While mask-based segmentation has achieved promising performance in fully-supervised setting [10, 11, 73], it has been rarely explored to transfer the learned mask-level representations into another domain. On the other hand, the point-based feature representations from 2D foundation model can be naively distilled by minimizing the per-point feature distance. For example, [62] proposed to train a 3D model to predicts 3D semantic meaningful features by distilling pixel aligned 2D features. However, similar methods are not applicable in our proposed method. First of all, our 2D semantic understanding model uses a mask-based segmentation head which does not provide semantically-rich features in the pixel level. Secondly, the backbone of our 2D semantic understanding model is a frozen stable diffusion model [68] which is designed to generate realistic images with rich details and not tuned for semantic segmentation tasks. The per-pixel features extracted from it are not feasible to supervise the training of our 3D mask" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.191 + ], + "angle": 0, + "content": "model†. In the following, we introduce our proposed mask distillation which is tailored to distill knowledge from the mask-based 2D foundation model to the geometry-aware 3D mask model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.463 + ], + "angle": 0, + "content": "The mask-based 2D foundation model predicts \\(N\\) 2D probabilistic masks \\(\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N\\) and their corresponding mask embeddings \\(\\{f_i^{2d}\\}_{i=1}^N\\). Specifically, \\(\\mathcal{B}_i^{2d}\\) represents a probabilistic map whose elements represent the probability of the corresponding pixel being foreground. We first compute the pixel-point correspondence following [62]. Subsequently, a set of 3D probabilistic masks \\(\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N\\) can be generated by lifting the 2D masks \\(\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N\\) to 3D space based on the pixel-point correspondence. We proposed a novel mask distillation which distills information from both 3D probabilistic masks \\(\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N\\) and the corresponding semantic rich mask embeddings \\(\\{f_i^{2d}\\}_{i=1}^N\\) generated from the 2D branch. Specifically, we train a Minkowski network [16] as the 3D mask prediction model to generate geometry-aware 3D masks. The 3D point cloud is quantized into voxels by averaging the pixels within each voxel to save memory and reduce computes. The 3D mask prediction model generates a 3D feature to represent each voxel and this feature is assigned to all points within the voxel. This produces a full feature map \\(\\mathbf{F}^{3d} \\in \\mathbb{R}^{M \\times D}\\) for the point cloud, where \\(D\\) is the dimension of the 3D feature. The semantic rich 2D mask embeddings \\(\\{f_i^{2d}\\}_{i=1}^N\\) are used as linear classifiers to compute the logits \\(\\mathcal{S}_i \\in \\mathbb{R}^M\\) of a 3D feature belonging to the corresponding class:" + }, + { + "type": "equation", + "bbox": [ + 0.443, + 0.463, + 0.786, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i} = \\left\\langle \\mathbf {F} ^ {3 d}, f _ {i} ^ {2 d} \\right\\rangle , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.534 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot, \\cdot \\rangle\\) denotes inner product. The 3D probabilistic mask \\(\\mathcal{B}_i^{3d}\\) is then generated by applying the sigmoid function on \\(S_i\\). We propose a multimodal mask distillation loss to train our 3D mask generator:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.547, + 0.786, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} 1 - \\cos \\left(\\mathcal {B} _ {i} ^ {\\prime 3 d}, \\mathcal {B} _ {i} ^ {3 d}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.644 + ], + "angle": 0, + "content": "The distillation loss aims at forcing the 2D and 3D branch to make consistent predictions. It serves as an implicit distillation objective to make the 3D model learn high-resolution, semantically-rich feature representations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.667, + 0.49, + 0.683 + ], + "angle": 0, + "content": "3.4 Open-Vocabulary Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.817 + ], + "angle": 0, + "content": "During inference, \\(Diff2Scene\\) takes a 3D point cloud and its multiview 2D images as inputs. The 2D semantic understanding model consumes the 2D images and generates a set of 2D probabilistic masks \\(\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N\\) with their corresponding mask embeddings \\(\\{f_i^{2d}\\}_{i=1}^N\\), where \\(f_i^{2d} \\in \\mathbb{R}^D\\). The 3D mask model takes the 3D point cloud and the mask embeddings \\(\\{f_i^{2d}\\}_{i=1}^N\\) as inputs to predict the 3D probabilistic mask \\(\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N\\). To ground a semantic label \\(c\\) to the 3D point cloud, we first apply the same idea from [87] to compute the geometric mean (denoted as \\(p_i^c\\)) of label probabilities from diffusion and discriminative models for each 2D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.712, + 0.841 + ], + "angle": 0, + "content": "†The 3D mask model trained to distill these features does not converge." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "mask \\(\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N\\). Next, the label probabilities \\(\\mathbf{p}^c\\) are assigned to 3D points via the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {c} = \\lambda \\sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \\mathcal {B} _ {i} ^ {3 d} + (1 - \\lambda) \\sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \\mathcal {B} _ {i} ^ {\\prime 3 d}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.24, + 0.784, + 0.27 + ], + "angle": 0, + "content": "where \\(\\lambda = 0.5\\). When multiple labels can be assigned to a 3D point, the label with the highest probability from Eq. 3 is taken." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.293, + 0.367, + 0.312 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.431 + ], + "angle": 0, + "content": "We conduct a series of experiments to demonstrate the effectiveness of Diff2Scene on a variety of zero-shot 3D scene understanding benchmarks. We first evaluate the proposed model on zero-shot open-vocabulary semantic segmentation tasks following the evaluation protocol of [62]. We then perform comprehensive ablation studies to validate our designs. Finally, we qualitatively demonstrate the strong ability of the proposed model for open-vocabulary 3D segmentation and grounding complicated compositional text queries." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.453, + 0.335, + 0.468 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.784, + 0.585 + ], + "angle": 0, + "content": "We use ScanNet [17], Matterport3D [7], ScanNet200 [69] and Replica [77] for the open-vocabulary 3D semantic segmentation task. We provide qualitative analysis of the visual grounding task on Nr3D [1]. Except for Replica, point clouds and multi-view images in the training split without ground truth annotations are used for model training. As Replica does not provide the training data, we perform training on ScanNet and perform evaluation on Replica, following the setting in [79]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.784, + 0.644 + ], + "angle": 0, + "content": "ScanNet is one of the largest 3D semantic segmentation dataset. It provides 80,554 images from 1201 scans for training and 21,300 images from 312 scans for testing with 20 semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.784, + 0.72 + ], + "angle": 0, + "content": "Matterport3D is a large scale RGB-D dataset containing 10,800 panoramic views from 194,000 RGB-D images of 90 building-scale scenes. It splits 61 scenes for training, 11 scenes for validation and 18 for testing. We train our 3D branch using the images in the training splits and report the results on test split." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.841 + ], + "angle": 0, + "content": "ScanNet200 has 200 semantic labels with long-tailed classes. It also provides a grouping of the 200 categories based on the number of labeled surface points in the training set, resulting in 3 subsets: head, common, and tail. This enables us to evaluate the performance of our method on the long-tail distribution, making ScanNet200 a natural choice as an evaluation dataset. We report the mean intersection over union (mIoU) metric on the validation set consisting of 312 scenes following the split in [62, 69, 78]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Comparison to state-of-the-art models. We report mIoU for all benchmarks. Best results in zero-shot, open-vocabulary setting are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.787, + 0.436 + ], + "angle": 0, + "content": "
ScanNetMatterport3DScanNet200Replica
AllAllHeadCommonTailAllHeadTailAll
Fully-supervised
TangentConv [81]40.9--------
TextureNet [34]54.8--------
SFSS-MMSI [16]-35.9-------
CSC-Pretrain [30]--45.517.17.924.9---
SupCon [93]--48.619.210.326.0---
LGround [69]--48.518.410.627.2---
MinkowskiNet [16]69.054.246.315.410.225.3---
Zero-shot, open-vocabulary
MSeg Voting [44]31.033.4-------
ConceptFusion [37]33.3-17.56.32.88.811.63.54.6
OpenMask3D [79]34.0-19.67.54.510.513.23.44.8
OpenScene (2D) [62]41.432.421.910.85.512.733.411.514.5
OpenScene (3D) [62]46.041.317.60.00.06.332.67.711.1
OpenScene (2D/3D) [62]47.542.620.09.75.111.634.211.914.9
Diff2Scene (Ours)48.645.525.611.56.914.246.212.917.5
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.467, + 0.785, + 0.513 + ], + "angle": 0, + "content": "Replica contains 51 categories, and we further split those categories into head and tail sets based on their appearance frequency. We report the mIoU on the office0, office1, office2, office3, office4, room0, room1, and room2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.785, + 0.62 + ], + "angle": 0, + "content": "\\(Nr3D\\) is a 3D visual grounding dataset which contains diverse text prompts. To further evaluate the ability of our model to distinguish between objects in the same class but with different attributes, we perform qualitative evaluation on the visual grounding dataset \\(Nr3D\\) [1]. We perform zero-shot evaluation on the validation set without training on any labeled data for the visual grounding task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.409, + 0.66 + ], + "angle": 0, + "content": "4.2 Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We compare Diff2Scene with the current state-of-the-art fully-supervised 3D semantic segmentation models including TangentConv [81], TextureNet [34], SFSS-MMSI [16], CSC-Pretrain [30], SupCon [93], LGround [69] and MinkowskiNet [16] on the 3D semantic segmentation benchmark. We also compare our model against OpenScene [62] and ConceptFusion [37], the recently proposed open-vocabulary 3D semantic understanding model. For OpenScene [62], we compare with its OpenSeg [25] variant which has the same feature and pretrained datasets for a fair comparison. We also compare our model with its three different variants (2D Fusion, 3D Distill, and 2D/3D Ensemble). Besides, we adapt the state-of-the-art 3D instance segmentation model OpenMask3D [79] for comparison on the 3D semantic segmentation benchmark." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.457, + 0.162 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.37 + ], + "angle": 0, + "content": "We use posed multi-view RGB images and 3D point clouds for all the datasets. ODISE [87], which consists of a diffusion backbone and mask-based segmentation head, is used as the model in our 2D branch. It uses a stable diffusion model [68] pre-trained on Laion-5B [72] as the feature backbone. The dimensions for diffusion and CLIP features are 256 and 768 respectively. The number of queries of Mask2Former [10] is 100. Similar to OpenScene [62], we use MinkowskiNet18A [16] as the model in our 3D branch to extract 3D features from the 3D point clouds. Our 3D model is trained for 200 epochs with a batch size of 8. Adam optimizer [41] is used with a learning rate of 0.0001 and polynomial learning rate policy is used as the learning rate scheduler with power 0.9. During inference, text-embeddings are computed by the ViT-L/14 CLIP model [65] for each of the semantic categories and grounding queries. We use the same pre-processing step and pre-trained dataset as OpenScene [62] (OpenSeg [25]) for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.392, + 0.433, + 0.406 + ], + "angle": 0, + "content": "4.4 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Evaluation on zero-shot 3D semantic segmentation. We first compare our method with the state-of-the-art open-vocabulary scene understanding models and fully-supervised 3D segmentation models. We report the mIoU for Scannet, Matterport3D, Scannet200, and Replica in Table 1. We find that our method achieves better results than the state-of-the-art open-vocabulary models and their variants on all the benchmarks. Besides, although our zero-shot model has noticeable performance drop compared with fully-supervised model, the gap of tail categories between the proposed method and those methods are relatively small (e.g. 6.9 vs. 7.9 from CSC-Pretrain) on Scannet200. This demonstrates the strong potential of the proposed method for long-trailed 3D semantic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Generalization to unseen dataset. To test the generalization ability of our proposed model, we evaluate it on an unseen dataset Replica [77] and report the results in Table 1. The results shown that our proposed method significantly outperforms the state-of-the-art models on head, tail and all categories in Replica. This demonstrates the strong generalization ability of the proposed model on novel datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Effectiveness of Different Distillation Settings. We compare our mask-based distillation method with point-based ones under different settings and report the performance of the 3D branch on Replica [77] in Table 2. The supervisions for point-based method include: (1) Fine-tuned CLIP feature, which follows the same setting as OpenScene [62]; (2) Frozen diffusion feature extracted from the last layer of diffusion U-Net block. We observe that distilling frozen diffusion features does not converge. Our proposed method, by introducing the semantic meaningful mask embedding output from the 2D branch as a fixed classifier, significantly boost the performance of the 3D branch." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Ablation studies. We conduct ablation studies using the Replica dataset [77] and show the results in Table 3. We first analyze the effectiveness of combining" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Effectiveness of Different Distillation Settings. We report mIoU of different methods on the Replica [77] dataset." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.185, + 0.734, + 0.259 + ], + "angle": 0, + "content": "
SettingDistillation TypeHead Tail All
fine-tuned CLIP feature [62]Point-based32.6 7.7 11.1
frozen diffusion featurePoint-basedDivergence
multimodal mask distillation (ours)Mask-based43.3 8.0 12.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.272, + 0.785, + 0.3 + ], + "angle": 0, + "content": "Table 3: Performance of different model ablations. We observe that each component of our model gains consistent improvements." + }, + { + "type": "table", + "bbox": [ + 0.321, + 0.312, + 0.679, + 0.414 + ], + "angle": 0, + "content": "
MethodmIoU
Our full model17.5
Without 2D (salient) mask12.8
Without 3D (geometric) mask16.5
Without discriminative (CLIP) features15.5
Without generative (Stable Diffusion) features15.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.549 + ], + "angle": 0, + "content": "2D and 3D masks using equation 3. We observe that compared with using salient or geometric mask only, using both types of masks achieves the best performance. This is intuitive as both salient patterns and geometric information are helpful to segment accurate class boundaries. We then analyze the effectiveness of different semantic features. We find that discriminative and diffusion features serve as strong complementary to each other. We also observe that using those two types of semantic features jointly can significantly outperform using any of them alone." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.427, + 0.603 + ], + "angle": 0, + "content": "4.5 Qualitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Visualizations of zero-shot semantic segmentation. In Fig. 4, we provide qualitative analysis of our approach and OpenScene for the zero-shot 3D semantic segmentation task. Compared with OpenScene, our model generates coherent and consistent masks (e.g., the table mask in first column and the bed mask in third column) thanks to the mask-instance representations. It predicts accurate semantic labels for both head and tail categories by leveraging both CLIP and diffusion features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Visualizations of visual grounding results. We provide qualitative analysis of our approach and OpenScene for the zero-shot visual grounding task in Fig. 5. We observe that our model can accurately identify objects given complicated text queries. It demonstrates that the proposed method, Diff2Scene, has good capability at the following types of queries. Fig. 5 (a) describes object shape and color, and even in comparative degree (It's the shorter, red box); Fig. 5 (b) describes a rare object (rack) and its surrounded object with surface appearance descriptions (wrinkled towel); Fig. 5 (c) describes the relative location of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.143, + 0.756, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.463, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results from our model and OpenScene on zero-shot semantic segmentation. We visualize the segmentation results on the validation set of ScanNet200 [69]. We observe that our model can predict coherent masks with accurate semantic labels compared to OpenScene for both head and tail categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.608 + ], + "angle": 0, + "content": "object (next to the desk); Fig. 5 (d) describes the usage of the object (recycling). In addition, we can see that given vague usage descriptions without common category names like trash bin in the text prompts, the model can still accurately identify the object." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.36, + 0.645 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In this paper, we investigate the problem of leveraging frozen representations from large text-to-image diffusion models for open-vocabulary 3D semantic understanding. Diff2Scene sets a new state-of-the-art in the zero-shot 3D semantic segmentation task and shows promising performance in the visual grounding task. Our method also shows outstanding generalization ability towards unseen datasets and novel text queries. It provides a new way to effectively leverage generative text-to-image foundation models for 3D semantic scene understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "There are several limitations of the proposed model. First, while our model achieves better performance compared to existing methods in small objects, it still misclassified some small and rare categories (e.g. rail). Second, we observe that the model can be easily confused by fine-grained categories that with similar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.142, + 0.782, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results from our model and OpenScene on zero-shot visual grounding. Our open-vocabulary semantic understanding model is capable of handling different types of novel and compositional queries. Novel object classes as well as objects described by colors, shapes, appearances, locations, and usages are successfully retrieved by our method. Note that the located points are colored in yellow." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.713 + ], + "angle": 0, + "content": "semantic meaning. For example, the model sometimes wrongly classifies points of windowsill to the window class. In future work, it will be interesting to design models that can accurately distinguish between fine-grained categories in the open-vocabulary setting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.405, + 0.775 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We sincerely thank Amir Hertz, Weicheng Kuo, and Lijun Yu for the helpful discussions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.J.: ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020) 3, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.3 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "3. Anand, A., Koppula, H.S., Joachims, T., Saxena, A.: Contextually guided semantic labeling and search for 3d point clouds. In: IJRR (2011) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.357 + ], + "angle": 0, + "content": "4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.384 + ], + "angle": 0, + "content": "5. Atzmon, M., Maron, H., Lipman, Y.: Point convolutional neural networks by extension operators. ACM Transactions on Graphics (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "6. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.495 + ], + "angle": 0, + "content": "8. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. arXiv preprint arXiv:2209.09874 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "9. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022) 2, 3, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "1. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "2. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: WACV (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "3. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. In: BMVC (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "4. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. IJCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "5. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: MVA (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019) 2, 3, 8, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.784 + ], + "angle": 0, + "content": "8. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. arXiv preprint arXiv:2308.00353 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Engelmann, F., Kontogianni, T., Hermans, A., Leibe, B.: Exploring spatial context for 3d semantic segmentation of point clouds. In: ICCV workshop (2017) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Engelmann, F., Kontogianni, T., Schult, J., Leibe, B.: Know what your neighbors do: 3d semantic segmentation of point clouds. In: ECCV workshop (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.215 + ], + "angle": 0, + "content": "22. Fan, J., Zheng, P., Li, S.: Vision-based holistic scene understanding towards proactive human-robot collaboration. Robotics and Computer-Integrated Manufacturing 75, 102304 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.242 + ], + "angle": 0, + "content": "23. Feng, R., Gao, Y., Tse, T.H.E., Ma, X., Chang, H.J.: Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.785, + 0.284 + ], + "angle": 0, + "content": "24. Gadre, S.Y., Wortsman, M., Ilharco, G., Schmidt, L., Song, S.: Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.311 + ], + "angle": 0, + "content": "25. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022) 2, 4, 5, 7, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "26. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.378 + ], + "angle": 0, + "content": "27. Han, L., Zheng, T., Zhu, Y., Xu, L., Fang, L.: Live semantic 3d perception for immersive augmented reality. IEEE Trans. visualization and computer graphics 26(5), 2012-2022 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "28. He, Q., Peng, J., Jiang, Z., Wu, K., Ji, X., Zhang, J., Wang, Y., Wang, C., Chen, M., Wu, Y.: Unim-ov3d: Uni-modality open-vocabulary 3d scene understanding with fine-grained feature representation. IJCAI (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "29. Holmquist, K., Wandt, B.: Diffpose: Multi-hypothesis human pose estimation using diffusion models. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "30. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: CVPR (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "31. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.54 + ], + "angle": 0, + "content": "32. Hua, B.S., Tran, M.K., Yeung, S.K.: Pointwise convolutional neural networks. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.568 + ], + "angle": 0, + "content": "33. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: ICRA (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "34. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: CVPR (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.636 + ], + "angle": 0, + "content": "35. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: CVPR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.676 + ], + "angle": 0, + "content": "36. Huang, Z., Lv, C., Xing, Y., Wu, J.: Multi-modal sensor fusion-based deep neural network for end-to-end autonomous driving with scene understanding. IEEE Sensors Journal 21(10), 11781-11790 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.732 + ], + "angle": 0, + "content": "37. Jatavallabhula, K., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., Tenenbaum, J., de Melo, C., Krishna, M., Paull, L., Shkurti, F., Torralba, A.: Conceptfusion: Open-set multimodal 3d mapping. In: Robotics science and systems (2023) 2, 4, 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.758 + ], + "angle": 0, + "content": "38. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "39. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 7" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "41. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "42. Koppula, H., Anand, A., Joachims, T., Saxena, A.: Semantic labeling of 3d point clouds for indoor scenes. In: NeurIPS (2011) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "43. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "44. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: MSeg: A composite dataset for multi-domain semantic segmentation. In: CVPR (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "45. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.31 + ], + "angle": 0, + "content": "46. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "47. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "48. Li, J., Selvaraju, R.R., Gotmare, A.D., Joty, S.R., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "49. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "50. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "51. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.5 + ], + "angle": 0, + "content": "52. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "53. Liu, D., Li, Q., Dinh, A.D., Jiang, T., Shah, M., Xu, C.: Diffusion action segmentation. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.567 + ], + "angle": 0, + "content": "54. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., Saddik, A.E., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: NeurIPS (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.596 + ], + "angle": 0, + "content": "55. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "56. Liu, Z., Qi, X., Fu, C.W.: 3d-to-2d distillation for indoor scene parsing. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.65 + ], + "angle": 0, + "content": "57. Lu, Y., Rasmussen, C.: Simplified markov random fields for efficient semantic labeling of 3d point clouds. In: ICIRS (2012) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.69 + ], + "angle": 0, + "content": "58. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "59. Mazur, K., Sucar, E., Davison, A.: Feature-realistic neural fusion for real-time, open set scene understanding. In: ICRA (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "60. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3D point cloud. In: 3DV (2021) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "61. Mittal, S., Abstreiter, K., Bauer, S., Scholkopf, B., Mehrjou, A.: Diffusion based representation learning. In: ICML (2023) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.812 + ], + "angle": 0, + "content": "62. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "X. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "66. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "67. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Linguagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "68. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 5, 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.356 + ], + "angle": 0, + "content": "69. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 3, 4, 9, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.412 + ], + "angle": 0, + "content": "70. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.464 + ], + "angle": 0, + "content": "71. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.521 + ], + "angle": 0, + "content": "72. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "73. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "74. Shafiullah, N.M.M., Paxton, C., Pinto, L., Chintala, S., Szlam, A.: CLIP-fields: Weakly supervised semantic fields for robotic memory. In: CoRL Workshop on Language and Robotics (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "75. Shah, D., Osinski, B., Ichter, B., Levine, S.: LM-nav: Robotic navigation with large pre-trained models of language, vision, and action. In: CoRL (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "76. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "77. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) 3, 9, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "78. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "79. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023) 2, 3, 4, 6, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "80. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.841 + ], + "angle": 0, + "content": "81. Tatarchenko, M., Park, J., Koltun, V., Zhou., Q.Y.: Tangent convolutions for dense prediction in 3D. CVPR (2018) 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "82. Tchapmi, L.P., Choy, C.B., Armeni, I., Gwak, J., Savarese, S.: Segcloud: Semantic segmentation of 3d point clouds. In: 3DV (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "83. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "84. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: ICCV (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "85. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.299 + ], + "angle": 0, + "content": "86. Wang, T., Li, J., An, X.: An efficient scene semantic labeling approach for 3d point cloud. In: ITSC (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.34 + ], + "angle": 0, + "content": "87. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) 2, 3, 5, 7, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.369 + ], + "angle": 0, + "content": "88. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.395 + ], + "angle": 0, + "content": "89. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.424 + ], + "angle": 0, + "content": "90. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "91. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "92. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "93. Zheng, M., Wang, F., You, S., Qian, C., Zhang, C., Wang, X., Xu, C.: Weakly supervised contrastive learning. In: ICCV (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "95. Zhu, X., Chen, J., Zeng, X., Liang, J., Li, C., Liu, S., Behpour, S., Xu, M.: Weakly supervised 3d semantic segmentation using cross-image consensus and inter-voxel affinity relations. In: ICCV (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_origin.pdf b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01a12f81521ca2ff910b531255525b65c3f9820f --- /dev/null +++ b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/76d1ce31-5c37-42e9-9b04-6a8f345c9c3a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20f5aa1b482f93888844ef2e58a636856e5a9c80f4c95f420012ffc6306288e9 +size 9184639 diff --git a/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/full.md b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..27f482ac6c8404f465fbeabab4c02a1985336095 --- /dev/null +++ b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/full.md @@ -0,0 +1,302 @@ +# Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models + +Xiaoyu Zhu $^{1\dagger}$ , Hao Zhou $^{2}$ , Pengfei Xing $^{2}$ , Long Zhao $^{2}$ , Hao Xu $^{3}$ , Junwei Liang $^{4}$ , Alexander Hauptmann $^{1}$ , Ting Liu $^{2}$ , and Andrew Gallagher $^{2}$ + +1 Carnegie Mellon University + +2 Google DeepMind + +3 Google +HKUST + +![](images/4c1649a6e2e03548aeb19c4a1910e53519f3b2317cd11c94fccc95c6cf5789f8.jpg) +Fig. 1: Illustration of open-vocabulary 3D semantic scene understanding. We propose Diff2Scene, a 3D model that performs open-vocabulary semantic segmentation and visual grounding tasks given novel text prompts, without relying on any annotated 3D data. By leveraging discriminative-based and generative-based 2D foundation models, Diff2Scene can handle a wide variety of novel text queries for both common and rare classes, like "desk" and "soap dispenser". It can also handle compositional queries, such as "find the white sneakers that are closer to the desk chair." + +![](images/af3c50172caea8bbfc4b814c0a98bdfa0afc31dda57ca2d5ab7f13de774727ec.jpg) + +![](images/a60d0c114989817290c1260c459e45eb10c59e66c07554867844ef8b833786c8.jpg) +"This backpack is alongside a square green chair." + +![](images/0d1a5dc0d29cf6bef148b985b981427b14a55caede524a272704da0556eb99cf.jpg) +Find the white sneakers that are closer to the desk chair. + +Abstract. In this paper, we investigate the use of diffusion models which are pre-trained on large-scale image-caption pairs for open-vocabulary 3D semantic understanding. We propose a novel method, namely Diff2Scene, which leverages frozen representations from text-image generative models, along with salient-aware and geometric-aware masks, for open-vocabulary 3D semantic segmentation and visual grounding tasks. Diff2Scene gets rid of any labeled 3D data and effectively identifies objects, appearances, materials, locations and their compositions in 3D scenes. We show that it outperforms competitive baselines and achieves significant improvements over state-of-the-art methods. In particular, Diff2Scene improves the state-of-the-art method on ScanNet200 by $12\%$ . + +Keywords: 3D Semantic Understanding $\cdot$ Open-Vocabulary Perception $\cdot$ Diffusion Model + +# 1 Introduction + +3D semantic scene understanding, with the task of assigning semantics to every 3D point, plays a fundamental role in many computer vision applications, such as robotics [88], autonomous driving [36], human-computer interaction [22], and augmented reality [27]. Traditional studies in this field usually target solving this problem in a closed-set fashion [16,73], resulting in models that can only be used to make predictions within the predefined label space. + +Recent progress in computer vision have witnessed the emerging interests in solving semantic understanding tasks in open-vocabulary settings [35, 62, 67, 78, 94]. In contrast to closed-set setting, models targeting open-vocabulary tasks must make predictions for any semantics described in text, including object category and fine-grained attributes (e.g., shape, color, material, property) as well as their complicated compositions. However, this is a challenging task due to the wide diversity and complexity of possible queries. Motivated by the advance of aligning text and image embeddings with large-scale foundation models [2, 39, 48, 65], existing methods mitigate this challenge by lifting the image features from foundation models such as CLIP [65] or their descendants [25, 47] to 3D. These lifted feature representations for 3D points can then be used to query with open-vocabulary descriptions, achieving semantic understanding in 3D. Despite these achievements, contrastively trained CLIP-based models exhibit limitations in handling fine-grained classes [66] and novel compositional text queries [58], restricting their performance in open-vocabulary 3D semantic understanding. + +The recently developed text-to-image diffusion models have shown outstanding abilities for image generation even with challenging text prompts [55,68,72], such as combinational descriptions with multiple attributes (e.g., A bucket bag made of blue suede with intricate golden paisley patterns.) The internal visual representation of these models, entangled with text embedding through cross-attention, have proven correlate well with semantic concepts described by language [43, 60, 61, 89]. On the other hand, CLIP-based foundation models have been shown to struggle with compositionality [58]. Moreover, compared with the CLIP model which is optimized for global representation, diffusion models have proven to be superior at local representation [80], which is a key for dense prediction tasks. Specifically, ODISE [87] applied the internal representations of Stable Diffusion [68] to open-vocabulary 2D semantic understanding tasks and achieved promising results. + +One of the key challenges in 3D perception is the severe scarcity of point clouds and their dense labels. Several existing methods have been proposed to solve the lack of data issue in a zero-shot fashion by leveraging the CLIP model pre-trained on large-scale text-image data [37,62,79]. The prior art [62] extracts dense CLIP features from 2D images and distill the knowledge of their lifted 3D counterpart into a 3D mask predictor. However, CLIP features, as discussed above, struggle to handle fine-grained classes [66] and show worse localization capability compared with diffusion features. We leverage diffusion model as feature backbone along with a mask-based segmentation head (e.g., Mask2Fromer [10]) for its intrinsic nature that decouples mask and its semantic representations. + +This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models and further learning geometrically-accurate masks from the 3D branch. However, performing multi-modal distillation with mask-based segmentation head is a non-trivial task. The frozen features extracted from the decoder of the U-Net in the diffusion model are trained with generative objectives, and cannot be directly used for the perception task. Therefore, directly distilling knowledge from these features as normally done in prior art [54,56,62] is infeasible. Another intuitive way is to leverage a supervised 3D mask proposal network and pool the feature representations from 2D CLIP features for each mask [79]. However, the training of 3D mask proposal network requires labeled 3D data, which may not be feasible in practice. + +To mitigate these issues, we propose a novel mask distillation method tailored to distill knowledge from the Mask2Former style 2D branch [10, 87] to the 3D branch, which is shown in Fig. 2. Specifically, we design our 3D branch to take a 3D point cloud as input and to predict their 3D features. The semantically meaningful mask embeddings produced from our 2D branch are used as linear classifiers to assign class probability to these 3D features. Their corresponding 2D masks are lifted to 3D based on pixel-point correspondence and used to force the consistency learning of the 2D and 3D branch. + +We evaluate Diff2Scene quantitatively on ScanNet [17], ScanNet200 [69], Matterport 3D [7] and Replica [77] for open-vocabulary 3D semantic segmentation and qualitatively on $\mathrm{Nr3D}$ [1] for visual grounding tasks. Our experimental results show that Diff2Scene outperforms state-of-the-art models [62] on all the four semantic segmentation datasets and achieves promising results on visual grounding tasks. In summary, we make the following contributions: + +- To the best of our knowledge, we are the first to leverage text-image diffusion to perform open-vocabulary 3D semantic segmentation. +- We propose a novel mask distillation method to train a 3D mask prediction model by distilling knowledge from the Mask2Former style 2D segmentation model. +- The proposed method achieves state-of-the-art performance on several open-vocabulary 3D semantic segmentation and visual grounding benchmarks. + +# 2 Related Work + +Closed-vocabulary 3D semantic segmentation. In 3D semantic segmentation, a semantic category is assigned for each 3D point. It has been long studied [3-5, 20, 21, 26, 31, 32, 42, 45, 49, 57, 63, 64, 82, 86, 95] due to its importance in computer vision and robotics applications. One challenge of this task is that 3D point clouds are not in a regular structured format; network architectures that work well for 2D tasks cannot handle 3D point clouds effectively. As a result, most of the early studies focus on designing effective and efficient network architectures that are suitable for 3D point clouds [16, 20, 21, 26, 31, 32, 45, 63, 64, 83]. This line of work achieved great success and significantly improves the results of + +![](images/5b81f04d59bb4554ff4d29a7be3fd55452524173e13d56186929e41e9a90639c.jpg) +(a) Feature-based Distillation w/o labeled 3D data + +![](images/9532b963500d02d5ef4f4e6800e7a96b3012ccf79b64bfe9218b71441c59e085.jpg) + +![](images/4e832f01c47e6a31fe445588ff7d58114dc077557186fe8d22f11f8c2b978263.jpg) +(b) Mask-based Feature Pooling w/ labeled 3D data +(c) Mask-based Distillation w/o labeled 3D data (Ours) +Fig. 2: Illustration of open-vocabulary 3D perception methods. $L_{PD}$ and $L_{MD}$ denote point-based distillation loss and mask-based distillation loss. $M_{3D}$ denote a set of predicted 3D masks; $M_{2D}$ and $Z_{mf}$ denote a set of predicted 2D masks and their semantic embeddings; $Z_{gf}$ denote the high-resolution 3D feature map. (a) Directly minimizing the per-point feature distance between the CLIP-based model and the tuned 3D model [62]. (b) Directly using a 3D mask proposal network trained on labeled 3D data to produce class-agnostic masks, and then pool corresponding representations from the CLIP feature map [79]. (c) The proposed mask distillation approach, namely Diff2Scene, that uses Stable Diffusion and performs mask-based distillation. Diff2Scene leverages the semantically-rich mask embeddings from 2D foundation models and geometrically accurate masks from the tuned 3D model, and thus achieves superior performance compared to previous methods. + +3D semantic segmentation. Another challenge is the lack of large scale data with ground truth annotations. Due to the intensive labeling effort and high cost of data annotation [69], the available datasets for 3D semantic segmentation are usually small in scale. In the absence of large scale data, early studies usually target solving this problem in a closed-vocabulary setting, where the trained model can only predict categories that appear during training. To mitigate the scale limitation of existing datasets, a handful of works [12-15, 52, 60, 91] have applied zero-shot learning in 3D scene understanding tasks. [12-15, 91] focused on 3D point classification task and [52, 60] tried to address the 3D semantic segmentation problem. However, these zero-short methods still require ground truth annotations for a certain amount of 3D point clouds. + +Open-vocabulary 3D segmentation. The recent progress of large-scale vision and language representation learning [2,28,39,48,65,90] has advanced the study of semantic and instance segmentation in an open-vocabulary setting. [25, 47, 51] first explored open-vocabulary 2D semantic segmentation. They proposed aligning per-pixel features [47] or features from mask regions [25, 51] with the corresponding text embedding. Following these works, [8,19,24,33,37,59,62,74, 75] focus on 3D semantic segmentation in an open-vocabulary setting. Among them, [8,24,33,37,59,71,74] project 3D points to 2D images and solve the 3D problem in the 2D space, instead of targeting the 3D open-vocabulary semantic segmentation directly. As [19] pointed out, the projection from 3D to 2D has information loss and the solution is suboptimal. + +To make better use of information from the 3D point cloud, [19], [18] and [62] proposed to directly applying semantic segmentation on the 3D point cloud. [19] and its extension [18] proposed associating captions generated for 2D images to corresponding 3D point clouds to build the pseudo-ground truth captions for 3D point clouds. A neural network is trained to associate the 3D point cloud with these pseudo labels through contrastive loss. Similar to the zero-shot setting, [19] evaluated their model in a leave-one-out fashion, which still requires annotations for 3D point cloud. Inspired by the strong open-vocabulary ability of large-scale vision and language models, Peng et al. [62] proposed distilling knowledge to a 3D point cloud model. They trained 3D semantic segmentation model by only distilling the knowledge from a CLIP-style [65] 2D open-vocabulary semantic segmentation model [25, 47]. They demonstrated that without training with any ground truth labels, the model can achieve great performance on many open-vocabulary tasks. However, we observe that [62] is strongly limited by the 2D open-vocabulary semantic segmentation models used as the teacher. Its performance on rare classes that are not used in training these models are not satisfactory. Our method follows this idea by distilling the knowledge of 2D open-vocabulary semantic segmentation model to a 3D model. In constraint to the approach in [62], we use a diffusion-based 2D open vocabulary semantic segmentation model [87] as the teacher model. + +Diffusion models for scene understanding. The last few years have witnessed the success of diffusion models in image generation [68, 70]. Recent studies also observed the diffusion models are strong representation learners [43, 60, 61, 89]. As a result, researchers have applied it to many understanding tasks such as image classification [46], object detection [9], image semantic segmentation [6, 38, 87], instance segmentation [50], human pose estimation [23, 29, 76], action segmentation [53], camera pose estimation [84], to name a few, and achieved great success. Especially, [87] and [50] showed that Stable Diffusion [68], whose internal representation being well correlated with text embedding, has strong open-vocabulary abilities for understanding tasks. Inspired by this, we are the first to apply text-to-image diffusion models to open vocabulary 3D semantic segmentation task. + +# 3 Diff2Scene + +We introduce Diff2Scene, an open-vocabulary 3D semantic understanding method. Similar to [62], our proposed model operates in a zero-shot fashion, where no ground truth 3D annotations are needed during training. + +# 3.1 Overview + +An overview of Diff2Scene is shown in Fig. 3. It takes posed RGB images and the reconstructed 3D point cloud as model inputs. The model predicts the semantic label for each 3D point. Diff2Scene has two branches. The 2D branch is designed + +![](images/3e326a7ce26f33488dd8a67bc7b5dc5e9f7724c15ea98183a0869e98a8559f45.jpg) +Fig. 3: Overview of our method. We propose Diff2Scene, an open-vocabulary 3D semantic understanding model. Diff2Scene contains two branches. The 2D branch is designed to be a diffusion-based 2D semantic segmentation model. It accepts a 2D image as input and predicts a set of 2D probabilistic masks with corresponding semantically-rich mask embeddings. The 3D branch utilizes the point cloud and 2D mask embeddings as input. The 2D mask embeddings are used as "semantic queries" to generate corresponding 3D probabilistic masks. The model learns salient patterns from the RGB images and geometric information from the point clouds. + +to be an open-vocabulary 2D semantic segmentation model. It leverages text-to-image generative model [68] which is pre-trained on massive text-image pairs. The model takes a 2D image as input to predict a set of 2D probabilistic masks with their corresponding 2D mask embeddings. Thanks to the generative pretraining process with large-scale text-image pairs, the 2D mask embeddings are semantically rich. The model leverages the salient patterns in RGB images to produce the 2D salient masks. The 3D branch takes the point cloud and the 2D mask embeddings as inputs. The 2D mask embeddings are used as linear classifiers to assign class probabilities to each of 3D features output from the 3D branch, resulting in a 3D probabilistic mask termed as geometric masks. To predict the per-point semantic class, the model first computes the per-mask category logits for both salient and geometric masks. Then we ensemble the per-mask logits for those two types of masks. In the way, the model can learn salient patterns from the RGB images and geometric information from the point clouds. + +# 3.2 2D Semantic Understanding Model + +One challenge of 3D semantic understanding is the severe scarcity of 3D point clouds with groundtruth labels. To tackle the challenge brought by limited training data, vision-language foundation models have been used to transfer semantically-rich 2D features into the 3D space [37,62,79]. [79] used on a model trained on labeled 3D data to produce class-agnostic masks, and then pooled the corresponding 2D representations as the mask embeddings. On the other hand, [62] proposed to leverage a pre-trained 2D semantic segmentation model as feature extractor to perform open-vocabulary 3D segmentation, and no ground + +truth 3D annotations are needed during training. In this work, we follow the setting in [62] to reduce the 3D annotation efforts. + +The 2D segmentation model consists of an image backbone $\phi$ which is a foundation model pretrained on large-scale text-image pairs; and a segmentation head $\sigma$ to predict the semantic embedding. There are multiple design options for the 2D backbone $\phi$ and segmentation head $\sigma$ . (1) The 2D backbone could either be contrastively pretrained or generatively pre-trianed. The popular frameworks for contrastive representation learning include CLIP [65] and ALIGN [40]. On the other hand, a few works [85,87,92] have demonstrated promising performance by using generatively pre-trained representations for perception task. The feature representations from Diffusion U-Net blocks are extracted for different downstream tasks. + +Once feature representations from text-image foundation models are extracted, a segmentation head is added upon those features to predict the perpoint semantic classes. The segmentation problem could be formulated as pixel-based classification or mask-based classification. For pixel-based classification [25,47], the intermediate output of segmentation head is of shape $H \times W \times C$ , where $H$ and $W$ is image height and width, and $C$ is the dimension of feature embedding. For mask-based classification [10, 11, 87], the segmentation head takes the 2D feature map $\mathbf{F}^{2d}$ and $N$ fixed mask queries $\{q_i\}_{i=1}^N$ as input. The intermediate output is $N$ 2D probabilistic masks $\{\mathcal{B}_i^{2d}\}_{i=1}^N$ and their corresponding mask embeddings $\{f_i^{2d}\}_{i=1}^N$ . + +In this work, we choose diffusion model as the feature backbone $\phi$ , considering its strong localization ability brought by generative pre-training. Besides, we leverage mask-based segmentation head for its intrinsic nature that decouples mask and its semantic representations. This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models, and further learn geometrically-accurate masks from the 3D branch. + +# 3.3 Geometry-Aware 3D Mask Model + +While mask-based segmentation has achieved promising performance in fully-supervised setting [10, 11, 73], it has been rarely explored to transfer the learned mask-level representations into another domain. On the other hand, the point-based feature representations from 2D foundation model can be naively distilled by minimizing the per-point feature distance. For example, [62] proposed to train a 3D model to predicts 3D semantic meaningful features by distilling pixel aligned 2D features. However, similar methods are not applicable in our proposed method. First of all, our 2D semantic understanding model uses a mask-based segmentation head which does not provide semantically-rich features in the pixel level. Secondly, the backbone of our 2D semantic understanding model is a frozen stable diffusion model [68] which is designed to generate realistic images with rich details and not tuned for semantic segmentation tasks. The per-pixel features extracted from it are not feasible to supervise the training of our 3D mask + +model†. In the following, we introduce our proposed mask distillation which is tailored to distill knowledge from the mask-based 2D foundation model to the geometry-aware 3D mask model. + +The mask-based 2D foundation model predicts $N$ 2D probabilistic masks $\{\mathcal{B}_i^{2d}\}_{i=1}^N$ and their corresponding mask embeddings $\{f_i^{2d}\}_{i=1}^N$ . Specifically, $\mathcal{B}_i^{2d}$ represents a probabilistic map whose elements represent the probability of the corresponding pixel being foreground. We first compute the pixel-point correspondence following [62]. Subsequently, a set of 3D probabilistic masks $\{\mathcal{B}_i^{3d}\}_{i=1}^N$ can be generated by lifting the 2D masks $\{\mathcal{B}_i^{2d}\}_{i=1}^N$ to 3D space based on the pixel-point correspondence. We proposed a novel mask distillation which distills information from both 3D probabilistic masks $\{\mathcal{B}_i^{3d}\}_{i=1}^N$ and the corresponding semantic rich mask embeddings $\{f_i^{2d}\}_{i=1}^N$ generated from the 2D branch. Specifically, we train a Minkowski network [16] as the 3D mask prediction model to generate geometry-aware 3D masks. The 3D point cloud is quantized into voxels by averaging the pixels within each voxel to save memory and reduce computes. The 3D mask prediction model generates a 3D feature to represent each voxel and this feature is assigned to all points within the voxel. This produces a full feature map $\mathbf{F}^{3d} \in \mathbb{R}^{M \times D}$ for the point cloud, where $D$ is the dimension of the 3D feature. The semantic rich 2D mask embeddings $\{f_i^{2d}\}_{i=1}^N$ are used as linear classifiers to compute the logits $\mathcal{S}_i \in \mathbb{R}^M$ of a 3D feature belonging to the corresponding class: + +$$ +\mathcal {S} _ {i} = \left\langle \mathbf {F} ^ {3 d}, f _ {i} ^ {2 d} \right\rangle , \tag {1} +$$ + +where $\langle \cdot, \cdot \rangle$ denotes inner product. The 3D probabilistic mask $\mathcal{B}_i^{3d}$ is then generated by applying the sigmoid function on $S_i$ . We propose a multimodal mask distillation loss to train our 3D mask generator: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {N} 1 - \cos \left(\mathcal {B} _ {i} ^ {\prime 3 d}, \mathcal {B} _ {i} ^ {3 d}\right). \tag {2} +$$ + +The distillation loss aims at forcing the 2D and 3D branch to make consistent predictions. It serves as an implicit distillation objective to make the 3D model learn high-resolution, semantically-rich feature representations. + +# 3.4 Open-Vocabulary Inference + +During inference, $Diff2Scene$ takes a 3D point cloud and its multiview 2D images as inputs. The 2D semantic understanding model consumes the 2D images and generates a set of 2D probabilistic masks $\{\mathcal{B}_i^{2d}\}_{i=1}^N$ with their corresponding mask embeddings $\{f_i^{2d}\}_{i=1}^N$ , where $f_i^{2d} \in \mathbb{R}^D$ . The 3D mask model takes the 3D point cloud and the mask embeddings $\{f_i^{2d}\}_{i=1}^N$ as inputs to predict the 3D probabilistic mask $\{\mathcal{B}_i^{3d}\}_{i=1}^N$ . To ground a semantic label $c$ to the 3D point cloud, we first apply the same idea from [87] to compute the geometric mean (denoted as $p_i^c$ ) of label probabilities from diffusion and discriminative models for each 2D + +mask $\{\mathcal{B}_i^{2d}\}_{i=1}^N$ . Next, the label probabilities $\mathbf{p}^c$ are assigned to 3D points via the following equation: + +$$ +\mathbf {p} ^ {c} = \lambda \sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \mathcal {B} _ {i} ^ {3 d} + (1 - \lambda) \sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \mathcal {B} _ {i} ^ {\prime 3 d}, \tag {3} +$$ + +where $\lambda = 0.5$ . When multiple labels can be assigned to a 3D point, the label with the highest probability from Eq. 3 is taken. + +# 4 Experiment + +We conduct a series of experiments to demonstrate the effectiveness of Diff2Scene on a variety of zero-shot 3D scene understanding benchmarks. We first evaluate the proposed model on zero-shot open-vocabulary semantic segmentation tasks following the evaluation protocol of [62]. We then perform comprehensive ablation studies to validate our designs. Finally, we qualitatively demonstrate the strong ability of the proposed model for open-vocabulary 3D segmentation and grounding complicated compositional text queries. + +# 4.1 Datasets + +We use ScanNet [17], Matterport3D [7], ScanNet200 [69] and Replica [77] for the open-vocabulary 3D semantic segmentation task. We provide qualitative analysis of the visual grounding task on Nr3D [1]. Except for Replica, point clouds and multi-view images in the training split without ground truth annotations are used for model training. As Replica does not provide the training data, we perform training on ScanNet and perform evaluation on Replica, following the setting in [79]. + +ScanNet is one of the largest 3D semantic segmentation dataset. It provides 80,554 images from 1201 scans for training and 21,300 images from 312 scans for testing with 20 semantic labels. + +Matterport3D is a large scale RGB-D dataset containing 10,800 panoramic views from 194,000 RGB-D images of 90 building-scale scenes. It splits 61 scenes for training, 11 scenes for validation and 18 for testing. We train our 3D branch using the images in the training splits and report the results on test split. + +ScanNet200 has 200 semantic labels with long-tailed classes. It also provides a grouping of the 200 categories based on the number of labeled surface points in the training set, resulting in 3 subsets: head, common, and tail. This enables us to evaluate the performance of our method on the long-tail distribution, making ScanNet200 a natural choice as an evaluation dataset. We report the mean intersection over union (mIoU) metric on the validation set consisting of 312 scenes following the split in [62, 69, 78]. + +Table 1: Comparison to state-of-the-art models. We report mIoU for all benchmarks. Best results in zero-shot, open-vocabulary setting are shown in bold. + +
ScanNetMatterport3DScanNet200Replica
AllAllHeadCommonTailAllHeadTailAll
Fully-supervised
TangentConv [81]40.9--------
TextureNet [34]54.8--------
SFSS-MMSI [16]-35.9-------
CSC-Pretrain [30]--45.517.17.924.9---
SupCon [93]--48.619.210.326.0---
LGround [69]--48.518.410.627.2---
MinkowskiNet [16]69.054.246.315.410.225.3---
Zero-shot, open-vocabulary
MSeg Voting [44]31.033.4-------
ConceptFusion [37]33.3-17.56.32.88.811.63.54.6
OpenMask3D [79]34.0-19.67.54.510.513.23.44.8
OpenScene (2D) [62]41.432.421.910.85.512.733.411.514.5
OpenScene (3D) [62]46.041.317.60.00.06.332.67.711.1
OpenScene (2D/3D) [62]47.542.620.09.75.111.634.211.914.9
Diff2Scene (Ours)48.645.525.611.56.914.246.212.917.5
+ +Replica contains 51 categories, and we further split those categories into head and tail sets based on their appearance frequency. We report the mIoU on the office0, office1, office2, office3, office4, room0, room1, and room2. + +$Nr3D$ is a 3D visual grounding dataset which contains diverse text prompts. To further evaluate the ability of our model to distinguish between objects in the same class but with different attributes, we perform qualitative evaluation on the visual grounding dataset $Nr3D$ [1]. We perform zero-shot evaluation on the validation set without training on any labeled data for the visual grounding task. + +# 4.2 Baseline Methods + +We compare Diff2Scene with the current state-of-the-art fully-supervised 3D semantic segmentation models including TangentConv [81], TextureNet [34], SFSS-MMSI [16], CSC-Pretrain [30], SupCon [93], LGround [69] and MinkowskiNet [16] on the 3D semantic segmentation benchmark. We also compare our model against OpenScene [62] and ConceptFusion [37], the recently proposed open-vocabulary 3D semantic understanding model. For OpenScene [62], we compare with its OpenSeg [25] variant which has the same feature and pretrained datasets for a fair comparison. We also compare our model with its three different variants (2D Fusion, 3D Distill, and 2D/3D Ensemble). Besides, we adapt the state-of-the-art 3D instance segmentation model OpenMask3D [79] for comparison on the 3D semantic segmentation benchmark. + +# 4.3 Implementation Details + +We use posed multi-view RGB images and 3D point clouds for all the datasets. ODISE [87], which consists of a diffusion backbone and mask-based segmentation head, is used as the model in our 2D branch. It uses a stable diffusion model [68] pre-trained on Laion-5B [72] as the feature backbone. The dimensions for diffusion and CLIP features are 256 and 768 respectively. The number of queries of Mask2Former [10] is 100. Similar to OpenScene [62], we use MinkowskiNet18A [16] as the model in our 3D branch to extract 3D features from the 3D point clouds. Our 3D model is trained for 200 epochs with a batch size of 8. Adam optimizer [41] is used with a learning rate of 0.0001 and polynomial learning rate policy is used as the learning rate scheduler with power 0.9. During inference, text-embeddings are computed by the ViT-L/14 CLIP model [65] for each of the semantic categories and grounding queries. We use the same pre-processing step and pre-trained dataset as OpenScene [62] (OpenSeg [25]) for a fair comparison. + +# 4.4 Quantitative Results + +Evaluation on zero-shot 3D semantic segmentation. We first compare our method with the state-of-the-art open-vocabulary scene understanding models and fully-supervised 3D segmentation models. We report the mIoU for Scannet, Matterport3D, Scannet200, and Replica in Table 1. We find that our method achieves better results than the state-of-the-art open-vocabulary models and their variants on all the benchmarks. Besides, although our zero-shot model has noticeable performance drop compared with fully-supervised model, the gap of tail categories between the proposed method and those methods are relatively small (e.g. 6.9 vs. 7.9 from CSC-Pretrain) on Scannet200. This demonstrates the strong potential of the proposed method for long-trailed 3D semantic segmentation tasks. + +Generalization to unseen dataset. To test the generalization ability of our proposed model, we evaluate it on an unseen dataset Replica [77] and report the results in Table 1. The results shown that our proposed method significantly outperforms the state-of-the-art models on head, tail and all categories in Replica. This demonstrates the strong generalization ability of the proposed model on novel datasets. + +Effectiveness of Different Distillation Settings. We compare our mask-based distillation method with point-based ones under different settings and report the performance of the 3D branch on Replica [77] in Table 2. The supervisions for point-based method include: (1) Fine-tuned CLIP feature, which follows the same setting as OpenScene [62]; (2) Frozen diffusion feature extracted from the last layer of diffusion U-Net block. We observe that distilling frozen diffusion features does not converge. Our proposed method, by introducing the semantic meaningful mask embedding output from the 2D branch as a fixed classifier, significantly boost the performance of the 3D branch. + +Ablation studies. We conduct ablation studies using the Replica dataset [77] and show the results in Table 3. We first analyze the effectiveness of combining + +Table 2: Effectiveness of Different Distillation Settings. We report mIoU of different methods on the Replica [77] dataset. + +
SettingDistillation TypeHead Tail All
fine-tuned CLIP feature [62]Point-based32.6 7.7 11.1
frozen diffusion featurePoint-basedDivergence
multimodal mask distillation (ours)Mask-based43.3 8.0 12.8
+ +Table 3: Performance of different model ablations. We observe that each component of our model gains consistent improvements. + +
MethodmIoU
Our full model17.5
Without 2D (salient) mask12.8
Without 3D (geometric) mask16.5
Without discriminative (CLIP) features15.5
Without generative (Stable Diffusion) features15.3
+ +2D and 3D masks using equation 3. We observe that compared with using salient or geometric mask only, using both types of masks achieves the best performance. This is intuitive as both salient patterns and geometric information are helpful to segment accurate class boundaries. We then analyze the effectiveness of different semantic features. We find that discriminative and diffusion features serve as strong complementary to each other. We also observe that using those two types of semantic features jointly can significantly outperform using any of them alone. + +# 4.5 Qualitative analysis + +Visualizations of zero-shot semantic segmentation. In Fig. 4, we provide qualitative analysis of our approach and OpenScene for the zero-shot 3D semantic segmentation task. Compared with OpenScene, our model generates coherent and consistent masks (e.g., the table mask in first column and the bed mask in third column) thanks to the mask-instance representations. It predicts accurate semantic labels for both head and tail categories by leveraging both CLIP and diffusion features. + +Visualizations of visual grounding results. We provide qualitative analysis of our approach and OpenScene for the zero-shot visual grounding task in Fig. 5. We observe that our model can accurately identify objects given complicated text queries. It demonstrates that the proposed method, Diff2Scene, has good capability at the following types of queries. Fig. 5 (a) describes object shape and color, and even in comparative degree (It's the shorter, red box); Fig. 5 (b) describes a rare object (rack) and its surrounded object with surface appearance descriptions (wrinkled towel); Fig. 5 (c) describes the relative location of the + +![](images/e4d887f69930dad3a32fee80548a629c70118e6b949e5dac3bd27c4ef47ec20f.jpg) +Fig. 4: Qualitative results from our model and OpenScene on zero-shot semantic segmentation. We visualize the segmentation results on the validation set of ScanNet200 [69]. We observe that our model can predict coherent masks with accurate semantic labels compared to OpenScene for both head and tail categories. + +object (next to the desk); Fig. 5 (d) describes the usage of the object (recycling). In addition, we can see that given vague usage descriptions without common category names like trash bin in the text prompts, the model can still accurately identify the object. + +# 5 Conclusion + +In this paper, we investigate the problem of leveraging frozen representations from large text-to-image diffusion models for open-vocabulary 3D semantic understanding. Diff2Scene sets a new state-of-the-art in the zero-shot 3D semantic segmentation task and shows promising performance in the visual grounding task. Our method also shows outstanding generalization ability towards unseen datasets and novel text queries. It provides a new way to effectively leverage generative text-to-image foundation models for 3D semantic scene understanding tasks. + +There are several limitations of the proposed model. First, while our model achieves better performance compared to existing methods in small objects, it still misclassified some small and rare categories (e.g. rail). Second, we observe that the model can be easily confused by fine-grained categories that with similar + +![](images/f1777dd22d3990c8e5f12ffe403ef29e25d42cfbba22c91322224bfdfbb4e05b.jpg) +Fig. 5: Qualitative results from our model and OpenScene on zero-shot visual grounding. Our open-vocabulary semantic understanding model is capable of handling different types of novel and compositional queries. Novel object classes as well as objects described by colors, shapes, appearances, locations, and usages are successfully retrieved by our method. Note that the located points are colored in yellow. + +semantic meaning. For example, the model sometimes wrongly classifies points of windowsill to the window class. In future work, it will be interesting to design models that can accurately distinguish between fine-grained categories in the open-vocabulary setting. + +# Acknowledgements + +We sincerely thank Amir Hertz, Weicheng Kuo, and Lijun Yu for the helpful discussions. + +# References + +1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.J.: ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020) 3, 9, 10 +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 2, 4 +3. Anand, A., Koppula, H.S., Joachims, T., Saxena, A.: Contextually guided semantic labeling and search for 3d point clouds. In: IJRR (2011) 3 +4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3 +5. Atzmon, M., Maron, H., Lipman, Y.: Point convolutional neural networks by extension operators. ACM Transactions on Graphics (2018) 3 +6. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022) 5 +7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017) 3, 9 +8. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. arXiv preprint arXiv:2209.09874 (2022) 4 +9. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. In: ICCV (2023) 5 +0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022) 2, 3, 7, 11 +1. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) 7 +2. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: WACV (2020) 4 +3. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. In: BMVC (2019) 4 +4. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. IJCV (2022) 4 +5. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: MVA (2019) 4 +6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019) 2, 3, 8, 10, 11 +7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9 +8. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. arXiv preprint arXiv:2308.00353 (2023) 5 +9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 4, 5 +20. Engelmann, F., Kontogianni, T., Hermans, A., Leibe, B.: Exploring spatial context for 3d semantic segmentation of point clouds. In: ICCV workshop (2017) 3 + +21. Engelmann, F., Kontogianni, T., Schult, J., Leibe, B.: Know what your neighbors do: 3d semantic segmentation of point clouds. In: ECCV workshop (2019) 3 +22. Fan, J., Zheng, P., Li, S.: Vision-based holistic scene understanding towards proactive human-robot collaboration. Robotics and Computer-Integrated Manufacturing 75, 102304 (2022) 2 +23. Feng, R., Gao, Y., Tse, T.H.E., Ma, X., Chang, H.J.: Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In: ICCV (2023) 5 +24. Gadre, S.Y., Wortsman, M., Ilharco, G., Schmidt, L., Song, S.: Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In: CVPR (2023) 4 +25. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022) 2, 4, 5, 7, 10, 11 +26. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018) 3 +27. Han, L., Zheng, T., Zhu, Y., Xu, L., Fang, L.: Live semantic 3d perception for immersive augmented reality. IEEE Trans. visualization and computer graphics 26(5), 2012-2022 (2020) 2 +28. He, Q., Peng, J., Jiang, Z., Wu, K., Ji, X., Zhang, J., Wang, Y., Wang, C., Chen, M., Wu, Y.: Unim-ov3d: Uni-modality open-vocabulary 3d scene understanding with fine-grained feature representation. IJCAI (2024) 4 +29. Holmquist, K., Wandt, B.: Diffpose: Multi-hypothesis human pose estimation using diffusion models. In: ICCV (2023) 5 +30. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: CVPR (2021) 10 +31. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: ICCV (2021) 3 +32. Hua, B.S., Tran, M.K., Yeung, S.K.: Pointwise convolutional neural networks. In: CVPR (2018) 3 +33. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: ICRA (2023) 4 +34. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: CVPR (2019) 10 +35. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: CVPR (2022) 2 +36. Huang, Z., Lv, C., Xing, Y., Wu, J.: Multi-modal sensor fusion-based deep neural network for end-to-end autonomous driving with scene understanding. IEEE Sensors Journal 21(10), 11781-11790 (2020) 2 +37. Jatavallabhula, K., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., Tenenbaum, J., de Melo, C., Krishna, M., Paull, L., Shkurti, F., Torralba, A.: Conceptfusion: Open-set multimodal 3d mapping. In: Robotics science and systems (2023) 2, 4, 6, 10 +38. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. In: ICCV (2023) 5 +39. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 2, 4 +40. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 7 + +41. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 11 +42. Koppula, H., Anand, A., Joachims, T., Saxena, A.: Semantic labeling of 3d point clouds for indoor scenes. In: NeurIPS (2011) 3 +43. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023) 2, 5 +44. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: MSeg: A composite dataset for multi-domain semantic segmentation. In: CVPR (2020) 10 +45. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs. In: CVPR (2018) 3 +46. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: ICCV (2023) 5 +47. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) 2, 4, 5, 7 +48. Li, J., Selvaraju, R.R., Gotmare, A.D., Joty, S.R., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021) 2, 4 +49. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018) 3 +50. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023) 5 +51. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023) 4 +52. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2022) 4 +53. Liu, D., Li, Q., Dinh, A.D., Jiang, T., Shah, M., Xu, C.: Diffusion action segmentation. In: ICCV (2023) 5 +54. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., Saddik, A.E., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: NeurIPS (2023) 3 +55. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022) 2 +56. Liu, Z., Qi, X., Fu, C.W.: 3d-to-2d distillation for indoor scene parsing. In: CVPR (2021) 3 +57. Lu, Y., Rasmussen, C.: Simplified markov random fields for efficient semantic labeling of 3d point clouds. In: ICIRS (2012) 3 +58. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2023) 2 +59. Mazur, K., Sucar, E., Davison, A.: Feature-realistic neural fusion for real-time, open set scene understanding. In: ICRA (2023) 4 +60. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3D point cloud. In: 3DV (2021) 2, 4, 5 +61. Mittal, S., Abstreiter, K., Bauer, S., Scholkopf, B., Mehrjou, A.: Diffusion based representation learning. In: ICML (2023) 2, 5 +62. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 +63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) 3 + +64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 3 +65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 5, 7, 11 +66. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2 +67. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Linguagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022) 2 +68. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 5, 6, 7, 11 +69. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 3, 4, 9, 10, 13 +70. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 5 +71. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 4 +72. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022) 2, 11 +73. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 2, 7 +74. Shafiullah, N.M.M., Paxton, C., Pinto, L., Chintala, S., Szlam, A.: CLIP-fields: Weakly supervised semantic fields for robotic memory. In: CoRL Workshop on Language and Robotics (2022) 4 +75. Shah, D., Osinski, B., Ichter, B., Levine, S.: LM-nav: Robotic navigation with large pre-trained models of language, vision, and action. In: CoRL (2022) 4 +76. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV (2023) 5 +77. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) 3, 9, 11, 12 +78. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 9 +79. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023) 2, 3, 4, 6, 9, 10 +80. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023) 2 +81. Tatarchenko, M., Park, J., Koltun, V., Zhou., Q.Y.: Tangent convolutions for dense prediction in 3D. CVPR (2018) 10 + +82. Tchapmi, L.P., Choy, C.B., Armeni, I., Gwak, J., Savarese, S.: Segcloud: Semantic segmentation of 3d point clouds. In: 3DV (2017) 3 +83. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019) 3 +84. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: ICCV (2023) 5 +85. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023) 7 +86. Wang, T., Li, J., An, X.: An efficient scene semantic labeling approach for 3d point cloud. In: ITSC (2015) 3 +87. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) 2, 3, 5, 7, 8, 11 +88. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020) 2 +89. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023) 2, 5 +90. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023) 4 +91. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR (2022) 4 +92. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023) 7 +93. Zheng, M., Wang, F., You, S., Qian, C., Zhang, C., Wang, X., Xu, C.: Weakly supervised contrastive learning. In: ICCV (2021) 10 +94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022) 2 +95. Zhu, X., Chen, J., Zeng, X., Liang, J., Li, C., Liu, S., Behpour, S., Xu, M.: Weakly supervised 3d semantic segmentation using cross-image consensus and inter-voxel affinity relations. In: ICCV (2021) 3 \ No newline at end of file diff --git a/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/images.zip b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b186c4f5f125a208dcce221978fa9baa3bc92f3e --- /dev/null +++ b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d51c94b1b6848ae36391f3a92f841187d65432dd7a76f40bb049d3a710d9e0b4 +size 508793 diff --git a/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/layout.json b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2655d213263b70665d509ea181f5a1ba0baad6ee --- /dev/null +++ b/2024/Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,9549 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "spans": [ + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": "Xiaoyu Zhu" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", Hao Zhou" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", Pengfei Xing" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", Long Zhao" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", Hao Xu" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", Junwei Liang" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", Alexander Hauptmann" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", Ting Liu" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "text", + "content": ", and Andrew Gallagher" + }, + { + "bbox": [ + 147, + 168, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 246, + 201, + 369, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 201, + 369, + 213 + ], + "spans": [ + { + "bbox": [ + 246, + 201, + 369, + 213 + ], + "type": "text", + "content": "1 Carnegie Mellon University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 265, + 213, + 348, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 213, + 348, + 224 + ], + "spans": [ + { + "bbox": [ + 265, + 213, + 348, + 224 + ], + "type": "text", + "content": "2 Google DeepMind" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 286, + 213, + 348, + 245 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 288, + 224, + 326, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 224, + 326, + 236 + ], + "spans": [ + { + "bbox": [ + 288, + 224, + 326, + 236 + ], + "type": "text", + "content": "3 Google" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 286, + 236, + 328, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 236, + 328, + 245 + ], + "spans": [ + { + "bbox": [ + 286, + 236, + 328, + 245 + ], + "type": "text", + "content": "HKUST" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 137, + 281, + 217, + 350 + ], + "blocks": [ + { + "bbox": [ + 137, + 281, + 217, + 350 + ], + "lines": [ + { + "bbox": [ + 137, + 281, + 217, + 350 + ], + "spans": [ + { + "bbox": [ + 137, + 281, + 217, + 350 + ], + "type": "image", + "image_path": "4c1649a6e2e03548aeb19c4a1910e53519f3b2317cd11c94fccc95c6cf5789f8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 368, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 445 + ], + "type": "text", + "content": "Fig. 1: Illustration of open-vocabulary 3D semantic scene understanding. We propose Diff2Scene, a 3D model that performs open-vocabulary semantic segmentation and visual grounding tasks given novel text prompts, without relying on any annotated 3D data. By leveraging discriminative-based and generative-based 2D foundation models, Diff2Scene can handle a wide variety of novel text queries for both common and rare classes, like \"desk\" and \"soap dispenser\". It can also handle compositional queries, such as \"find the white sneakers that are closer to the desk chair.\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 223, + 281, + 307, + 349 + ], + "blocks": [ + { + "bbox": [ + 223, + 281, + 307, + 349 + ], + "lines": [ + { + "bbox": [ + 223, + 281, + 307, + 349 + ], + "spans": [ + { + "bbox": [ + 223, + 281, + 307, + 349 + ], + "type": "image", + "image_path": "af3c50172caea8bbfc4b814c0a98bdfa0afc31dda57ca2d5ab7f13de774727ec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 281, + 390, + 328 + ], + "blocks": [ + { + "bbox": [ + 309, + 281, + 390, + 328 + ], + "lines": [ + { + "bbox": [ + 309, + 281, + 390, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 281, + 390, + 328 + ], + "type": "image", + "image_path": "a60d0c114989817290c1260c459e45eb10c59e66c07554867844ef8b833786c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 331, + 383, + 343 + ], + "lines": [ + { + "bbox": [ + 317, + 331, + 383, + 343 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 383, + 343 + ], + "type": "text", + "content": "\"This backpack is alongside a square green chair.\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 396, + 281, + 477, + 329 + ], + "blocks": [ + { + "bbox": [ + 396, + 281, + 477, + 329 + ], + "lines": [ + { + "bbox": [ + 396, + 281, + 477, + 329 + ], + "spans": [ + { + "bbox": [ + 396, + 281, + 477, + 329 + ], + "type": "image", + "image_path": "0d1a5dc0d29cf6bef148b985b981427b14a55caede524a272704da0556eb99cf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 331, + 468, + 343 + ], + "lines": [ + { + "bbox": [ + 405, + 331, + 468, + 343 + ], + "spans": [ + { + "bbox": [ + 405, + 331, + 468, + 343 + ], + "type": "text", + "content": "Find the white sneakers that are closer to the desk chair." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 468, + 469, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 468, + 469, + 590 + ], + "spans": [ + { + "bbox": [ + 159, + 468, + 469, + 590 + ], + "type": "text", + "content": "Abstract. In this paper, we investigate the use of diffusion models which are pre-trained on large-scale image-caption pairs for open-vocabulary 3D semantic understanding. We propose a novel method, namely Diff2Scene, which leverages frozen representations from text-image generative models, along with salient-aware and geometric-aware masks, for open-vocabulary 3D semantic segmentation and visual grounding tasks. Diff2Scene gets rid of any labeled 3D data and effectively identifies objects, appearances, materials, locations and their compositions in 3D scenes. We show that it outperforms competitive baselines and achieves significant improvements over state-of-the-art methods. In particular, Diff2Scene improves the state-of-the-art method on ScanNet200 by " + }, + { + "bbox": [ + 159, + 468, + 469, + 590 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 159, + 468, + 469, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 160, + 601, + 453, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 601, + 453, + 622 + ], + "spans": [ + { + "bbox": [ + 160, + 601, + 453, + 622 + ], + "type": "text", + "content": "Keywords: 3D Semantic Understanding " + }, + { + "bbox": [ + 160, + 601, + 453, + 622 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 601, + 453, + 622 + ], + "type": "text", + "content": " Open-Vocabulary Perception " + }, + { + "bbox": [ + 160, + 601, + 453, + 622 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 601, + 453, + 622 + ], + "type": "text", + "content": " Diffusion Model" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 481, + 666 + ], + "type": "text", + "content": "This work was partially done while the author was a student researcher at Google." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 210 + ], + "type": "text", + "content": "3D semantic scene understanding, with the task of assigning semantics to every 3D point, plays a fundamental role in many computer vision applications, such as robotics [88], autonomous driving [36], human-computer interaction [22], and augmented reality [27]. Traditional studies in this field usually target solving this problem in a closed-set fashion [16,73], resulting in models that can only be used to make predictions within the predefined label space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 481, + 390 + ], + "type": "text", + "content": "Recent progress in computer vision have witnessed the emerging interests in solving semantic understanding tasks in open-vocabulary settings [35, 62, 67, 78, 94]. In contrast to closed-set setting, models targeting open-vocabulary tasks must make predictions for any semantics described in text, including object category and fine-grained attributes (e.g., shape, color, material, property) as well as their complicated compositions. However, this is a challenging task due to the wide diversity and complexity of possible queries. Motivated by the advance of aligning text and image embeddings with large-scale foundation models [2, 39, 48, 65], existing methods mitigate this challenge by lifting the image features from foundation models such as CLIP [65] or their descendants [25, 47] to 3D. These lifted feature representations for 3D points can then be used to query with open-vocabulary descriptions, achieving semantic understanding in 3D. Despite these achievements, contrastively trained CLIP-based models exhibit limitations in handling fine-grained classes [66] and novel compositional text queries [58], restricting their performance in open-vocabulary 3D semantic understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "text", + "content": "The recently developed text-to-image diffusion models have shown outstanding abilities for image generation even with challenging text prompts [55,68,72], such as combinational descriptions with multiple attributes (e.g., A bucket bag made of blue suede with intricate golden paisley patterns.) The internal visual representation of these models, entangled with text embedding through cross-attention, have proven correlate well with semantic concepts described by language [43, 60, 61, 89]. On the other hand, CLIP-based foundation models have been shown to struggle with compositionality [58]. Moreover, compared with the CLIP model which is optimized for global representation, diffusion models have proven to be superior at local representation [80], which is a key for dense prediction tasks. Specifically, ODISE [87] applied the internal representations of Stable Diffusion [68] to open-vocabulary 2D semantic understanding tasks and achieved promising results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "One of the key challenges in 3D perception is the severe scarcity of point clouds and their dense labels. Several existing methods have been proposed to solve the lack of data issue in a zero-shot fashion by leveraging the CLIP model pre-trained on large-scale text-image data [37,62,79]. The prior art [62] extracts dense CLIP features from 2D images and distill the knowledge of their lifted 3D counterpart into a 3D mask predictor. However, CLIP features, as discussed above, struggle to handle fine-grained classes [66] and show worse localization capability compared with diffusion features. We leverage diffusion model as feature backbone along with a mask-based segmentation head (e.g., Mask2Fromer [10]) for its intrinsic nature that decouples mask and its semantic representations." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models and further learning geometrically-accurate masks from the 3D branch. However, performing multi-modal distillation with mask-based segmentation head is a non-trivial task. The frozen features extracted from the decoder of the U-Net in the diffusion model are trained with generative objectives, and cannot be directly used for the perception task. Therefore, directly distilling knowledge from these features as normally done in prior art [54,56,62] is infeasible. Another intuitive way is to leverage a supervised 3D mask proposal network and pool the feature representations from 2D CLIP features for each mask [79]. However, the training of 3D mask proposal network requires labeled 3D data, which may not be feasible in practice." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 342 + ], + "type": "text", + "content": "To mitigate these issues, we propose a novel mask distillation method tailored to distill knowledge from the Mask2Former style 2D branch [10, 87] to the 3D branch, which is shown in Fig. 2. Specifically, we design our 3D branch to take a 3D point cloud as input and to predict their 3D features. The semantically meaningful mask embeddings produced from our 2D branch are used as linear classifiers to assign class probability to these 3D features. Their corresponding 2D masks are lifted to 3D based on pixel-point correspondence and used to force the consistency learning of the 2D and 3D branch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "content": "We evaluate Diff2Scene quantitatively on ScanNet [17], ScanNet200 [69], Matterport 3D [7] and Replica [77] for open-vocabulary 3D semantic segmentation and qualitatively on " + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "content": " [1] for visual grounding tasks. Our experimental results show that Diff2Scene outperforms state-of-the-art models [62] on all the four semantic segmentation datasets and achieves promising results on visual grounding tasks. In summary, we make the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 422, + 479, + 504 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 422, + 479, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 479, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 479, + 444 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to leverage text-image diffusion to perform open-vocabulary 3D semantic segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 446, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 446, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 137, + 446, + 479, + 479 + ], + "type": "text", + "content": "- We propose a novel mask distillation method to train a 3D mask prediction model by distilling knowledge from the Mask2Former style 2D segmentation model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 482, + 479, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 482, + 479, + 504 + ], + "spans": [ + { + "bbox": [ + 137, + 482, + 479, + 504 + ], + "type": "text", + "content": "- The proposed method achieves state-of-the-art performance on several open-vocabulary 3D semantic segmentation and visual grounding benchmarks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 533, + 237, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 237, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 237, + 545 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Closed-vocabulary 3D semantic segmentation. In 3D semantic segmentation, a semantic category is assigned for each 3D point. It has been long studied [3-5, 20, 21, 26, 31, 32, 42, 45, 49, 57, 63, 64, 82, 86, 95] due to its importance in computer vision and robotics applications. One challenge of this task is that 3D point clouds are not in a regular structured format; network architectures that work well for 2D tasks cannot handle 3D point clouds effectively. As a result, most of the early studies focus on designing effective and efficient network architectures that are suitable for 3D point clouds [16, 20, 21, 26, 31, 32, 45, 63, 64, 83]. This line of work achieved great success and significantly improves the results of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 124, + 239, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 124, + 239, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 124, + 239, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 124, + 239, + 190 + ], + "type": "image", + "image_path": "5b81f04d59bb4554ff4d29a7be3fd55452524173e13d56186929e41e9a90639c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 193, + 223, + 208 + ], + "lines": [ + { + "bbox": [ + 138, + 193, + 223, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 193, + 223, + 208 + ], + "type": "text", + "content": "(a) Feature-based Distillation w/o labeled 3D data" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 125, + 341, + 190 + ], + "blocks": [ + { + "bbox": [ + 244, + 125, + 341, + 190 + ], + "lines": [ + { + "bbox": [ + 244, + 125, + 341, + 190 + ], + "spans": [ + { + "bbox": [ + 244, + 125, + 341, + 190 + ], + "type": "image", + "image_path": "9532b963500d02d5ef4f4e6800e7a96b3012ccf79b64bfe9218b71441c59e085.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 344, + 124, + 497, + 193 + ], + "blocks": [ + { + "bbox": [ + 251, + 193, + 344, + 209 + ], + "lines": [ + { + "bbox": [ + 251, + 193, + 344, + 209 + ], + "spans": [ + { + "bbox": [ + 251, + 193, + 344, + 209 + ], + "type": "text", + "content": "(b) Mask-based Feature Pooling w/ labeled 3D data" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 344, + 124, + 497, + 193 + ], + "lines": [ + { + "bbox": [ + 344, + 124, + 497, + 193 + ], + "spans": [ + { + "bbox": [ + 344, + 124, + 497, + 193 + ], + "type": "image", + "image_path": "4e832f01c47e6a31fe445588ff7d58114dc077557186fe8d22f11f8c2b978263.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 194, + 474, + 210 + ], + "lines": [ + { + "bbox": [ + 395, + 194, + 474, + 210 + ], + "spans": [ + { + "bbox": [ + 395, + 194, + 474, + 210 + ], + "type": "text", + "content": "(c) Mask-based Distillation w/o labeled 3D data (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "text", + "content": "Fig. 2: Illustration of open-vocabulary 3D perception methods. " + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "inline_equation", + "content": "L_{PD}" + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "inline_equation", + "content": "L_{MD}" + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "text", + "content": " denote point-based distillation loss and mask-based distillation loss. " + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "inline_equation", + "content": "M_{3D}" + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "text", + "content": " denote a set of predicted 3D masks; " + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "inline_equation", + "content": "M_{2D}" + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "inline_equation", + "content": "Z_{mf}" + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "text", + "content": " denote a set of predicted 2D masks and their semantic embeddings; " + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "inline_equation", + "content": "Z_{gf}" + }, + { + "bbox": [ + 131, + 232, + 482, + 365 + ], + "type": "text", + "content": " denote the high-resolution 3D feature map. (a) Directly minimizing the per-point feature distance between the CLIP-based model and the tuned 3D model [62]. (b) Directly using a 3D mask proposal network trained on labeled 3D data to produce class-agnostic masks, and then pool corresponding representations from the CLIP feature map [79]. (c) The proposed mask distillation approach, namely Diff2Scene, that uses Stable Diffusion and performs mask-based distillation. Diff2Scene leverages the semantically-rich mask embeddings from 2D foundation models and geometrically accurate masks from the tuned 3D model, and thus achieves superior performance compared to previous methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "text", + "content": "3D semantic segmentation. Another challenge is the lack of large scale data with ground truth annotations. Due to the intensive labeling effort and high cost of data annotation [69], the available datasets for 3D semantic segmentation are usually small in scale. In the absence of large scale data, early studies usually target solving this problem in a closed-vocabulary setting, where the trained model can only predict categories that appear during training. To mitigate the scale limitation of existing datasets, a handful of works [12-15, 52, 60, 91] have applied zero-shot learning in 3D scene understanding tasks. [12-15, 91] focused on 3D point classification task and [52, 60] tried to address the 3D semantic segmentation problem. However, these zero-short methods still require ground truth annotations for a certain amount of 3D point clouds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "Open-vocabulary 3D segmentation. The recent progress of large-scale vision and language representation learning [2,28,39,48,65,90] has advanced the study of semantic and instance segmentation in an open-vocabulary setting. [25, 47, 51] first explored open-vocabulary 2D semantic segmentation. They proposed aligning per-pixel features [47] or features from mask regions [25, 51] with the corresponding text embedding. Following these works, [8,19,24,33,37,59,62,74, 75] focus on 3D semantic segmentation in an open-vocabulary setting. Among them, [8,24,33,37,59,71,74] project 3D points to 2D images and solve the 3D problem in the 2D space, instead of targeting the 3D open-vocabulary semantic segmentation directly. As [19] pointed out, the projection from 3D to 2D has information loss and the solution is suboptimal." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 354 + ], + "type": "text", + "content": "To make better use of information from the 3D point cloud, [19], [18] and [62] proposed to directly applying semantic segmentation on the 3D point cloud. [19] and its extension [18] proposed associating captions generated for 2D images to corresponding 3D point clouds to build the pseudo-ground truth captions for 3D point clouds. A neural network is trained to associate the 3D point cloud with these pseudo labels through contrastive loss. Similar to the zero-shot setting, [19] evaluated their model in a leave-one-out fashion, which still requires annotations for 3D point cloud. Inspired by the strong open-vocabulary ability of large-scale vision and language models, Peng et al. [62] proposed distilling knowledge to a 3D point cloud model. They trained 3D semantic segmentation model by only distilling the knowledge from a CLIP-style [65] 2D open-vocabulary semantic segmentation model [25, 47]. They demonstrated that without training with any ground truth labels, the model can achieve great performance on many open-vocabulary tasks. However, we observe that [62] is strongly limited by the 2D open-vocabulary semantic segmentation models used as the teacher. Its performance on rare classes that are not used in training these models are not satisfactory. Our method follows this idea by distilling the knowledge of 2D open-vocabulary semantic segmentation model to a 3D model. In constraint to the approach in [62], we use a diffusion-based 2D open vocabulary semantic segmentation model [87] as the teacher model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 356, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 356, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 133, + 356, + 481, + 499 + ], + "type": "text", + "content": "Diffusion models for scene understanding. The last few years have witnessed the success of diffusion models in image generation [68, 70]. Recent studies also observed the diffusion models are strong representation learners [43, 60, 61, 89]. As a result, researchers have applied it to many understanding tasks such as image classification [46], object detection [9], image semantic segmentation [6, 38, 87], instance segmentation [50], human pose estimation [23, 29, 76], action segmentation [53], camera pose estimation [84], to name a few, and achieved great success. Especially, [87] and [50] showed that Stable Diffusion [68], whose internal representation being well correlated with text embedding, has strong open-vocabulary abilities for understanding tasks. Inspired by this, we are the first to apply text-to-image diffusion models to open vocabulary 3D semantic segmentation task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 521, + 217, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 217, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 217, + 533 + ], + "type": "text", + "content": "3 Diff2Scene" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 549, + 485, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 549, + 485, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 549, + 485, + 586 + ], + "type": "text", + "content": "We introduce Diff2Scene, an open-vocabulary 3D semantic understanding method. Similar to [62], our proposed model operates in a zero-shot fashion, where no ground truth 3D annotations are needed during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 606, + 208, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 606, + 208, + 617 + ], + "spans": [ + { + "bbox": [ + 133, + 606, + 208, + 617 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "type": "text", + "content": "An overview of Diff2Scene is shown in Fig. 3. It takes posed RGB images and the reconstructed 3D point cloud as model inputs. The model predicts the semantic label for each 3D point. Diff2Scene has two branches. The 2D branch is designed" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 481, + 239 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 481, + 239 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 481, + 239 + ], + "type": "image", + "image_path": "3e326a7ce26f33488dd8a67bc7b5dc5e9f7724c15ea98183a0869e98a8559f45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 338 + ], + "type": "text", + "content": "Fig. 3: Overview of our method. We propose Diff2Scene, an open-vocabulary 3D semantic understanding model. Diff2Scene contains two branches. The 2D branch is designed to be a diffusion-based 2D semantic segmentation model. It accepts a 2D image as input and predicts a set of 2D probabilistic masks with corresponding semantically-rich mask embeddings. The 3D branch utilizes the point cloud and 2D mask embeddings as input. The 2D mask embeddings are used as \"semantic queries\" to generate corresponding 3D probabilistic masks. The model learns salient patterns from the RGB images and geometric information from the point clouds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "text", + "content": "to be an open-vocabulary 2D semantic segmentation model. It leverages text-to-image generative model [68] which is pre-trained on massive text-image pairs. The model takes a 2D image as input to predict a set of 2D probabilistic masks with their corresponding 2D mask embeddings. Thanks to the generative pretraining process with large-scale text-image pairs, the 2D mask embeddings are semantically rich. The model leverages the salient patterns in RGB images to produce the 2D salient masks. The 3D branch takes the point cloud and the 2D mask embeddings as inputs. The 2D mask embeddings are used as linear classifiers to assign class probabilities to each of 3D features output from the 3D branch, resulting in a 3D probabilistic mask termed as geometric masks. To predict the per-point semantic class, the model first computes the per-mask category logits for both salient and geometric masks. Then we ensemble the per-mask logits for those two types of masks. In the way, the model can learn salient patterns from the RGB images and geometric information from the point clouds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 547, + 339, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 339, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 339, + 560 + ], + "type": "text", + "content": "3.2 2D Semantic Understanding Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "One challenge of 3D semantic understanding is the severe scarcity of 3D point clouds with groundtruth labels. To tackle the challenge brought by limited training data, vision-language foundation models have been used to transfer semantically-rich 2D features into the 3D space [37,62,79]. [79] used on a model trained on labeled 3D data to produce class-agnostic masks, and then pooled the corresponding 2D representations as the mask embeddings. On the other hand, [62] proposed to leverage a pre-trained 2D semantic segmentation model as feature extractor to perform open-vocabulary 3D segmentation, and no ground" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "truth 3D annotations are needed during training. In this work, we follow the setting in [62] to reduce the 3D annotation efforts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "text", + "content": "The 2D segmentation model consists of an image backbone " + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "text", + "content": " which is a foundation model pretrained on large-scale text-image pairs; and a segmentation head " + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "text", + "content": " to predict the semantic embedding. There are multiple design options for the 2D backbone " + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "text", + "content": " and segmentation head " + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 141, + 481, + 260 + ], + "type": "text", + "content": ". (1) The 2D backbone could either be contrastively pretrained or generatively pre-trianed. The popular frameworks for contrastive representation learning include CLIP [65] and ALIGN [40]. On the other hand, a few works [85,87,92] have demonstrated promising performance by using generatively pre-trained representations for perception task. The feature representations from Diffusion U-Net blocks are extracted for different downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": "Once feature representations from text-image foundation models are extracted, a segmentation head is added upon those features to predict the perpoint semantic classes. The segmentation problem could be formulated as pixel-based classification or mask-based classification. For pixel-based classification [25,47], the intermediate output of segmentation head is of shape " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "H \\times W \\times C" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " is image height and width, and " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " is the dimension of feature embedding. For mask-based classification [10, 11, 87], the segmentation head takes the 2D feature map " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{2d}" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " fixed mask queries " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\{q_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " as input. The intermediate output is " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " 2D probabilistic masks " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": " and their corresponding mask embeddings " + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\{f_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 263, + 482, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 456 + ], + "type": "text", + "content": "In this work, we choose diffusion model as the feature backbone " + }, + { + "bbox": [ + 130, + 384, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 384, + 482, + 456 + ], + "type": "text", + "content": ", considering its strong localization ability brought by generative pre-training. Besides, we leverage mask-based segmentation head for its intrinsic nature that decouples mask and its semantic representations. This is intuitively suitable for leveraging semantically-rich embeddings from 2D foundation models, and further learn geometrically-accurate masks from the 3D branch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 482, + 331, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 482, + 331, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 482, + 331, + 494 + ], + "type": "text", + "content": "3.3 Geometry-Aware 3D Mask Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "While mask-based segmentation has achieved promising performance in fully-supervised setting [10, 11, 73], it has been rarely explored to transfer the learned mask-level representations into another domain. On the other hand, the point-based feature representations from 2D foundation model can be naively distilled by minimizing the per-point feature distance. For example, [62] proposed to train a 3D model to predicts 3D semantic meaningful features by distilling pixel aligned 2D features. However, similar methods are not applicable in our proposed method. First of all, our 2D semantic understanding model uses a mask-based segmentation head which does not provide semantically-rich features in the pixel level. Secondly, the backbone of our 2D semantic understanding model is a frozen stable diffusion model [68] which is designed to generate realistic images with rich details and not tuned for semantic segmentation tasks. The per-pixel features extracted from it are not feasible to supervise the training of our 3D mask" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": "model†. In the following, we introduce our proposed mask distillation which is tailored to distill knowledge from the mask-based 2D foundation model to the geometry-aware 3D mask model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": "The mask-based 2D foundation model predicts " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " 2D probabilistic masks " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " and their corresponding mask embeddings " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\{f_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i^{2d}" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " represents a probabilistic map whose elements represent the probability of the corresponding pixel being foreground. We first compute the pixel-point correspondence following [62]. Subsequently, a set of 3D probabilistic masks " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " can be generated by lifting the 2D masks " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " to 3D space based on the pixel-point correspondence. We proposed a novel mask distillation which distills information from both 3D probabilistic masks " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " and the corresponding semantic rich mask embeddings " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\{f_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " generated from the 2D branch. Specifically, we train a Minkowski network [16] as the 3D mask prediction model to generate geometry-aware 3D masks. The 3D point cloud is quantized into voxels by averaging the pixels within each voxel to save memory and reduce computes. The 3D mask prediction model generates a 3D feature to represent each voxel and this feature is assigned to all points within the voxel. This produces a full feature map " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{3d} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " for the point cloud, where " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " is the dimension of the 3D feature. The semantic rich 2D mask embeddings " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\{f_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " are used as linear classifiers to compute the logits " + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_i \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": " of a 3D feature belonging to the corresponding class:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 271, + 366, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 366, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 271, + 366, + 481, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i} = \\left\\langle \\mathbf {F} ^ {3 d}, f _ {i} ^ {2 d} \\right\\rangle , \\tag {1}", + "image_path": "6843b0fd66cc709c189f48aebe4c5865a420c847ddfc09464a323571fad73668.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "text", + "content": " denotes inner product. The 3D probabilistic mask " + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i^{3d}" + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "text", + "content": " is then generated by applying the sigmoid function on " + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "text", + "content": ". We propose a multimodal mask distillation loss to train our 3D mask generator:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 433, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 433, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 247, + 433, + 481, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {N} 1 - \\cos \\left(\\mathcal {B} _ {i} ^ {\\prime 3 d}, \\mathcal {B} _ {i} ^ {3 d}\\right). \\tag {2}", + "image_path": "1f23259e70776418f861fa65ec88501a096e6574ccea0fe66f98a1d59ee427f8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 510 + ], + "type": "text", + "content": "The distillation loss aims at forcing the 2D and 3D branch to make consistent predictions. It serves as an implicit distillation objective to make the 3D model learn high-resolution, semantically-rich feature representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 528, + 299, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 299, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 299, + 540 + ], + "type": "text", + "content": "3.4 Open-Vocabulary Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": "During inference, " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "Diff2Scene" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": " takes a 3D point cloud and its multiview 2D images as inputs. The 2D semantic understanding model consumes the 2D images and generates a set of 2D probabilistic masks " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": " with their corresponding mask embeddings " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\{f_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "f_i^{2d} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": ". The 3D mask model takes the 3D point cloud and the mask embeddings " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\{f_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": " as inputs to predict the 3D probabilistic mask " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{3d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": ". To ground a semantic label " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": " to the 3D point cloud, we first apply the same idea from [87] to compute the geometric mean (denoted as " + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "inline_equation", + "content": "p_i^c" + }, + { + "bbox": [ + 130, + 548, + 482, + 647 + ], + "type": "text", + "content": ") of label probabilities from diffusion and discriminative models for each 2D" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 141, + 654, + 435, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 435, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 435, + 666 + ], + "type": "text", + "content": "†The 3D mask model trained to distill these features does not converge." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "mask " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i^{2d}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ". Next, the label probabilities " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^c" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " are assigned to 3D points via the following equation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 149, + 480, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 216, + 149, + 480, + 181 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {c} = \\lambda \\sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \\mathcal {B} _ {i} ^ {3 d} + (1 - \\lambda) \\sum_ {i = 1} ^ {N} p _ {i} ^ {c} * \\mathcal {B} _ {i} ^ {\\prime 3 d}, \\tag {3}", + "image_path": "ecacba3ed9fbe0b5a91cb7e7bc622c5ea8907bcb566dfb437a6cdbce530e2a0f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 190, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 479, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 190, + 479, + 213 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 131, + 190, + 479, + 213 + ], + "type": "text", + "content": ". When multiple labels can be assigned to a 3D point, the label with the highest probability from Eq. 3 is taken." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 232, + 224, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 224, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 224, + 247 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 479, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 341 + ], + "type": "text", + "content": "We conduct a series of experiments to demonstrate the effectiveness of Diff2Scene on a variety of zero-shot 3D scene understanding benchmarks. We first evaluate the proposed model on zero-shot open-vocabulary semantic segmentation tasks following the evaluation protocol of [62]. We then perform comprehensive ablation studies to validate our designs. Finally, we qualitatively demonstrate the strong ability of the proposed model for open-vocabulary 3D segmentation and grounding complicated compositional text queries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 358, + 205, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 205, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 205, + 370 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 378, + 479, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 479, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 479, + 463 + ], + "type": "text", + "content": "We use ScanNet [17], Matterport3D [7], ScanNet200 [69] and Replica [77] for the open-vocabulary 3D semantic segmentation task. We provide qualitative analysis of the visual grounding task on Nr3D [1]. Except for Replica, point clouds and multi-view images in the training split without ground truth annotations are used for model training. As Replica does not provide the training data, we perform training on ScanNet and perform evaluation on Replica, following the setting in [79]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 474, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 479, + 510 + ], + "type": "text", + "content": "ScanNet is one of the largest 3D semantic segmentation dataset. It provides 80,554 images from 1201 scans for training and 21,300 images from 312 scans for testing with 20 semantic labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 521, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 479, + 570 + ], + "type": "text", + "content": "Matterport3D is a large scale RGB-D dataset containing 10,800 panoramic views from 194,000 RGB-D images of 90 building-scale scenes. It splits 61 scenes for training, 11 scenes for validation and 18 for testing. We train our 3D branch using the images in the training splits and report the results on test split." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 666 + ], + "type": "text", + "content": "ScanNet200 has 200 semantic labels with long-tailed classes. It also provides a grouping of the 200 categories based on the number of labeled surface points in the training set, resulting in 3 subsets: head, common, and tail. This enables us to evaluate the performance of our method on the long-tail distribution, making ScanNet200 a natural choice as an evaluation dataset. We report the mean intersection over union (mIoU) metric on the validation set consisting of 312 scenes following the split in [62, 69, 78]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 345 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Comparison to state-of-the-art models. We report mIoU for all benchmarks. Best results in zero-shot, open-vocabulary setting are shown in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 345 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 345 + ], + "type": "table", + "html": "
ScanNetMatterport3DScanNet200Replica
AllAllHeadCommonTailAllHeadTailAll
Fully-supervised
TangentConv [81]40.9--------
TextureNet [34]54.8--------
SFSS-MMSI [16]-35.9-------
CSC-Pretrain [30]--45.517.17.924.9---
SupCon [93]--48.619.210.326.0---
LGround [69]--48.518.410.627.2---
MinkowskiNet [16]69.054.246.315.410.225.3---
Zero-shot, open-vocabulary
MSeg Voting [44]31.033.4-------
ConceptFusion [37]33.3-17.56.32.88.811.63.54.6
OpenMask3D [79]34.0-19.67.54.510.513.23.44.8
OpenScene (2D) [62]41.432.421.910.85.512.733.411.514.5
OpenScene (3D) [62]46.041.317.60.00.06.332.67.711.1
OpenScene (2D/3D) [62]47.542.620.09.75.111.634.211.914.9
Diff2Scene (Ours)48.645.525.611.56.914.246.212.917.5
", + "image_path": "bb3d87afa40798d3f1c2230ffbbbb4a161f40ccd3d71020af9636ea48cb16bad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 369, + 480, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 406 + ], + "type": "text", + "content": "Replica contains 51 categories, and we further split those categories into head and tail sets based on their appearance frequency. We report the mIoU on the office0, office1, office2, office3, office4, room0, room1, and room2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 419, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 480, + 491 + ], + "type": "inline_equation", + "content": "Nr3D" + }, + { + "bbox": [ + 130, + 419, + 480, + 491 + ], + "type": "text", + "content": " is a 3D visual grounding dataset which contains diverse text prompts. To further evaluate the ability of our model to distinguish between objects in the same class but with different attributes, we perform qualitative evaluation on the visual grounding dataset " + }, + { + "bbox": [ + 130, + 419, + 480, + 491 + ], + "type": "inline_equation", + "content": "Nr3D" + }, + { + "bbox": [ + 130, + 419, + 480, + 491 + ], + "type": "text", + "content": " [1]. We perform zero-shot evaluation on the validation set without training on any labeled data for the visual grounding task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 511, + 250, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 250, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 250, + 522 + ], + "type": "text", + "content": "4.2 Baseline Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 665 + ], + "type": "text", + "content": "We compare Diff2Scene with the current state-of-the-art fully-supervised 3D semantic segmentation models including TangentConv [81], TextureNet [34], SFSS-MMSI [16], CSC-Pretrain [30], SupCon [93], LGround [69] and MinkowskiNet [16] on the 3D semantic segmentation benchmark. We also compare our model against OpenScene [62] and ConceptFusion [37], the recently proposed open-vocabulary 3D semantic understanding model. For OpenScene [62], we compare with its OpenSeg [25] variant which has the same feature and pretrained datasets for a fair comparison. We also compare our model with its three different variants (2D Fusion, 3D Distill, and 2D/3D Ensemble). Besides, we adapt the state-of-the-art 3D instance segmentation model OpenMask3D [79] for comparison on the 3D semantic segmentation benchmark." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 293 + ], + "type": "text", + "content": "We use posed multi-view RGB images and 3D point clouds for all the datasets. ODISE [87], which consists of a diffusion backbone and mask-based segmentation head, is used as the model in our 2D branch. It uses a stable diffusion model [68] pre-trained on Laion-5B [72] as the feature backbone. The dimensions for diffusion and CLIP features are 256 and 768 respectively. The number of queries of Mask2Former [10] is 100. Similar to OpenScene [62], we use MinkowskiNet18A [16] as the model in our 3D branch to extract 3D features from the 3D point clouds. Our 3D model is trained for 200 epochs with a batch size of 8. Adam optimizer [41] is used with a learning rate of 0.0001 and polynomial learning rate policy is used as the learning rate scheduler with power 0.9. During inference, text-embeddings are computed by the ViT-L/14 CLIP model [65] for each of the semantic categories and grounding queries. We use the same pre-processing step and pre-trained dataset as OpenScene [62] (OpenSeg [25]) for a fair comparison." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 310, + 264, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 264, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 264, + 321 + ], + "type": "text", + "content": "4.4 Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 461 + ], + "type": "text", + "content": "Evaluation on zero-shot 3D semantic segmentation. We first compare our method with the state-of-the-art open-vocabulary scene understanding models and fully-supervised 3D segmentation models. We report the mIoU for Scannet, Matterport3D, Scannet200, and Replica in Table 1. We find that our method achieves better results than the state-of-the-art open-vocabulary models and their variants on all the benchmarks. Besides, although our zero-shot model has noticeable performance drop compared with fully-supervised model, the gap of tail categories between the proposed method and those methods are relatively small (e.g. 6.9 vs. 7.9 from CSC-Pretrain) on Scannet200. This demonstrates the strong potential of the proposed method for long-trailed 3D semantic segmentation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 533 + ], + "type": "text", + "content": "Generalization to unseen dataset. To test the generalization ability of our proposed model, we evaluate it on an unseen dataset Replica [77] and report the results in Table 1. The results shown that our proposed method significantly outperforms the state-of-the-art models on head, tail and all categories in Replica. This demonstrates the strong generalization ability of the proposed model on novel datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "Effectiveness of Different Distillation Settings. We compare our mask-based distillation method with point-based ones under different settings and report the performance of the 3D branch on Replica [77] in Table 2. The supervisions for point-based method include: (1) Fine-tuned CLIP feature, which follows the same setting as OpenScene [62]; (2) Frozen diffusion feature extracted from the last layer of diffusion U-Net block. We observe that distilling frozen diffusion features does not converge. Our proposed method, by introducing the semantic meaningful mask embedding output from the 2D branch as a fixed classifier, significantly boost the performance of the 3D branch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Ablation studies. We conduct ablation studies using the Replica dataset [77] and show the results in Table 3. We first analyze the effectiveness of combining" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 162, + 146, + 449, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Effectiveness of Different Distillation Settings. We report mIoU of different methods on the Replica [77] dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 146, + 449, + 205 + ], + "lines": [ + { + "bbox": [ + 162, + 146, + 449, + 205 + ], + "spans": [ + { + "bbox": [ + 162, + 146, + 449, + 205 + ], + "type": "table", + "html": "
SettingDistillation TypeHead Tail All
fine-tuned CLIP feature [62]Point-based32.6 7.7 11.1
frozen diffusion featurePoint-basedDivergence
multimodal mask distillation (ours)Mask-based43.3 8.0 12.8
", + "image_path": "dee1d4f789cb5ea660d6a875954805b99953c20b16edbc575d5ee334339b957b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 196, + 247, + 415, + 327 + ], + "blocks": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "text", + "content": "Table 3: Performance of different model ablations. We observe that each component of our model gains consistent improvements." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 196, + 247, + 415, + 327 + ], + "lines": [ + { + "bbox": [ + 196, + 247, + 415, + 327 + ], + "spans": [ + { + "bbox": [ + 196, + 247, + 415, + 327 + ], + "type": "table", + "html": "
MethodmIoU
Our full model17.5
Without 2D (salient) mask12.8
Without 3D (geometric) mask16.5
Without discriminative (CLIP) features15.5
Without generative (Stable Diffusion) features15.3
", + "image_path": "4c9796ec18757c702295c8c1f61fcb8a6f206f0e8c73600c79cfc8d631a802ae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "text", + "content": "2D and 3D masks using equation 3. We observe that compared with using salient or geometric mask only, using both types of masks achieves the best performance. This is intuitive as both salient patterns and geometric information are helpful to segment accurate class boundaries. We then analyze the effectiveness of different semantic features. We find that discriminative and diffusion features serve as strong complementary to each other. We also observe that using those two types of semantic features jointly can significantly outperform using any of them alone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 464, + 261, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 261, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 261, + 477 + ], + "type": "text", + "content": "4.5 Qualitative analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": "Visualizations of zero-shot semantic segmentation. In Fig. 4, we provide qualitative analysis of our approach and OpenScene for the zero-shot 3D semantic segmentation task. Compared with OpenScene, our model generates coherent and consistent masks (e.g., the table mask in first column and the bed mask in third column) thanks to the mask-instance representations. It predicts accurate semantic labels for both head and tail categories by leveraging both CLIP and diffusion features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Visualizations of visual grounding results. We provide qualitative analysis of our approach and OpenScene for the zero-shot visual grounding task in Fig. 5. We observe that our model can accurately identify objects given complicated text queries. It demonstrates that the proposed method, Diff2Scene, has good capability at the following types of queries. Fig. 5 (a) describes object shape and color, and even in comparative degree (It's the shorter, red box); Fig. 5 (b) describes a rare object (rack) and its surrounded object with surface appearance descriptions (wrinkled towel); Fig. 5 (c) describes the relative location of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 113, + 462, + 355 + ], + "blocks": [ + { + "bbox": [ + 147, + 113, + 462, + 355 + ], + "lines": [ + { + "bbox": [ + 147, + 113, + 462, + 355 + ], + "spans": [ + { + "bbox": [ + 147, + 113, + 462, + 355 + ], + "type": "image", + "image_path": "e4d887f69930dad3a32fee80548a629c70118e6b949e5dac3bd27c4ef47ec20f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 366, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 131, + 366, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 482, + 411 + ], + "type": "text", + "content": "Fig. 4: Qualitative results from our model and OpenScene on zero-shot semantic segmentation. We visualize the segmentation results on the validation set of ScanNet200 [69]. We observe that our model can predict coherent masks with accurate semantic labels compared to OpenScene for both head and tail categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 433, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 481 + ], + "type": "text", + "content": "object (next to the desk); Fig. 5 (d) describes the usage of the object (recycling). In addition, we can see that given vague usage descriptions without common category names like trash bin in the text prompts, the model can still accurately identify the object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 498, + 220, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 220, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 220, + 510 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": "In this paper, we investigate the problem of leveraging frozen representations from large text-to-image diffusion models for open-vocabulary 3D semantic understanding. Diff2Scene sets a new state-of-the-art in the zero-shot 3D semantic segmentation task and shows promising performance in the visual grounding task. Our method also shows outstanding generalization ability towards unseen datasets and novel text queries. It provides a new way to effectively leverage generative text-to-image foundation models for 3D semantic scene understanding tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "There are several limitations of the proposed model. First, while our model achieves better performance compared to existing methods in small objects, it still misclassified some small and rare categories (e.g. rail). Second, we observe that the model can be easily confused by fine-grained categories that with similar" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 112, + 478, + 419 + ], + "blocks": [ + { + "bbox": [ + 138, + 112, + 478, + 419 + ], + "lines": [ + { + "bbox": [ + 138, + 112, + 478, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 478, + 419 + ], + "type": "image", + "image_path": "f1777dd22d3990c8e5f12ffe403ef29e25d42cfbba22c91322224bfdfbb4e05b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 485 + ], + "type": "text", + "content": "Fig. 5: Qualitative results from our model and OpenScene on zero-shot visual grounding. Our open-vocabulary semantic understanding model is capable of handling different types of novel and compositional queries. Novel object classes as well as objects described by colors, shapes, appearances, locations, and usages are successfully retrieved by our method. Note that the located points are colored in yellow." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 515, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 564 + ], + "type": "text", + "content": "semantic meaning. For example, the model sometimes wrongly classifies points of windowsill to the window class. In future work, it will be interesting to design models that can accurately distinguish between fine-grained categories in the open-vocabulary setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 599, + 247, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 247, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 247, + 613 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We sincerely thank Amir Hertz, Weicheng Kuo, and Lijun Yu for the helpful discussions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Achlioptas, P., Abdelreheem, A., Xia, F., Elhoseiny, M., Guibas, L.J.: ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In: ECCV (2020) 3, 9, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 237 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "type": "text", + "content": "3. Anand, A., Koppula, H.S., Joachims, T., Saxena, A.: Contextually guided semantic labeling and search for 3d point clouds. In: IJRR (2011) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "text", + "content": "4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "text", + "content": "5. Atzmon, M., Maron, H., Lipman, Y.: Point convolutional neural networks by extension operators. ACM Transactions on Graphics (2018) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "text", + "content": "6. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from rgb-d data in indoor environments. In: 3DV (2017) 3, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "type": "text", + "content": "8. Chen, B., Xia, F., Ichter, B., Rao, K., Gopalakrishnan, K., Ryoo, M.S., Stone, A., Kappler, D.: Open-vocabulary queryable scene representations for real world planning. arXiv preprint arXiv:2209.09874 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "9. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. In: ICCV (2023) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "text", + "content": "0. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022) 2, 3, 7, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "text", + "content": "1. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "text", + "content": "2. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: WACV (2020) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "text", + "content": "3. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. In: BMVC (2019) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "text", + "content": "4. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. IJCV (2022) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "5. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: MVA (2019) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: CVPR (2019) 2, 3, 8, 10, 11" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "7. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 620 + ], + "type": "text", + "content": "8. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. arXiv preprint arXiv:2308.00353 (2023) 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "20. Engelmann, F., Kontogianni, T., Hermans, A., Leibe, B.: Exploring spatial context for 3d semantic segmentation of point clouds. In: ICCV workshop (2017) 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "21. Engelmann, F., Kontogianni, T., Schult, J., Leibe, B.: Know what your neighbors do: 3d semantic segmentation of point clouds. In: ECCV workshop (2019) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 170 + ], + "type": "text", + "content": "22. Fan, J., Zheng, P., Li, S.: Vision-based holistic scene understanding towards proactive human-robot collaboration. Robotics and Computer-Integrated Manufacturing 75, 102304 (2022) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "text", + "content": "23. Feng, R., Gao, Y., Tse, T.H.E., Ma, X., Chang, H.J.: Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In: ICCV (2023) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 193, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 480, + 224 + ], + "type": "text", + "content": "24. Gadre, S.Y., Wortsman, M., Ilharco, G., Schmidt, L., Song, S.: Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In: CVPR (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "text", + "content": "25. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022) 2, 4, 5, 7, 10, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "text", + "content": "26. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: CVPR (2018) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "type": "text", + "content": "27. Han, L., Zheng, T., Zhu, Y., Xu, L., Fang, L.: Live semantic 3d perception for immersive augmented reality. IEEE Trans. visualization and computer graphics 26(5), 2012-2022 (2020) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "28. He, Q., Peng, J., Jiang, Z., Wu, K., Ji, X., Zhang, J., Wang, Y., Wang, C., Chen, M., Wu, Y.: Unim-ov3d: Uni-modality open-vocabulary 3d scene understanding with fine-grained feature representation. IJCAI (2024) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "text", + "content": "29. Holmquist, K., Wandt, B.: Diffpose: Multi-hypothesis human pose estimation using diffusion models. In: ICCV (2023) 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "30. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: CVPR (2021) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "31. Hu, Z., Bai, X., Shang, J., Zhang, R., Dong, J., Wang, X., Sun, G., Fu, H., Tai, C.L.: Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In: ICCV (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 427 + ], + "type": "text", + "content": "32. Hua, B.S., Tran, M.K., Yeung, S.K.: Pointwise convolutional neural networks. In: CVPR (2018) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 449 + ], + "type": "text", + "content": "33. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: ICRA (2023) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "text", + "content": "34. Huang, J., Zhang, H., Yi, L., Funkhouser, T., Nießner, M., Guibas, L.J.: Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In: CVPR (2019) 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "text", + "content": "35. Huang, S., Chen, Y., Jia, J., Wang, L.: Multi-view transformer for 3d visual grounding. In: CVPR (2022) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "text", + "content": "36. Huang, Z., Lv, C., Xing, Y., Wu, J.: Multi-modal sensor fusion-based deep neural network for end-to-end autonomous driving with scene understanding. IEEE Sensors Journal 21(10), 11781-11790 (2020) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "type": "text", + "content": "37. Jatavallabhula, K., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., Tenenbaum, J., de Melo, C., Krishna, M., Paull, L., Shkurti, F., Torralba, A.: Conceptfusion: Open-set multimodal 3d mapping. In: Robotics science and systems (2023) 2, 4, 6, 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 580, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 600 + ], + "type": "text", + "content": "38. Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. In: ICCV (2023) 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "39. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "40. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "41. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "text", + "content": "42. Koppula, H., Anand, A., Joachims, T., Saxena, A.: Semantic labeling of 3d point clouds for indoor scenes. In: NeurIPS (2011) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "43. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023) 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "44. Lambert, J., Liu, Z., Sener, O., Hays, J., Koltun, V.: MSeg: A composite dataset for multi-domain semantic segmentation. In: CVPR (2020) 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "45. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs. In: CVPR (2018) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "text", + "content": "46. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: ICCV (2023) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "47. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) 2, 4, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "48. Li, J., Selvaraju, R.R., Gotmare, A.D., Joty, S.R., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "text", + "content": "49. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "50. Li, Z., Zhou, Q., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Open-vocabulary object segmentation with diffusion models. In: ICCV (2023) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "text", + "content": "51. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "52. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2022) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "53. Liu, D., Li, Q., Dinh, A.D., Jiang, T., Shah, M., Xu, C.: Diffusion action segmentation. In: ICCV (2023) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "54. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., Saddik, A.E., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: NeurIPS (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "type": "text", + "content": "55. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "56. Liu, Z., Qi, X., Fu, C.W.: 3d-to-2d distillation for indoor scene parsing. In: CVPR (2021) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "text", + "content": "57. Lu, Y., Rasmussen, C.: Simplified markov random fields for efficient semantic labeling of 3d point clouds. In: ICIRS (2012) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "type": "text", + "content": "58. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2023) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "59. Mazur, K., Sucar, E., Davison, A.: Feature-realistic neural fusion for real-time, open set scene understanding. In: ICRA (2023) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "60. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3D point cloud. In: 3DV (2021) 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "61. Mittal, S., Abstreiter, K., Bauer, S., Scholkopf, B., Mehrjou, A.: Diffusion based representation learning. In: ICML (2023) 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "text", + "content": "62. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "63. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "64. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "65. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 5, 7, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "66. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "67. Roh, J., Desingh, K., Farhadi, A., Fox, D.: Linguagerefer: Spatial-language model for 3d visual grounding. In: Conference on Robot Learning. pp. 1046-1056. PMLR (2022) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "68. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 5, 6, 7, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "69. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 3, 4, 9, 10, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 326 + ], + "type": "text", + "content": "70. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Gontijo-Lopes, R., Ayan, B.K., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 367 + ], + "type": "text", + "content": "71. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 412 + ], + "type": "text", + "content": "72. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022) 2, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "73. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "74. Shafiullah, N.M.M., Paxton, C., Pinto, L., Chintala, S., Szlam, A.: CLIP-fields: Weakly supervised semantic fields for robotic memory. In: CoRL Workshop on Language and Robotics (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "75. Shah, D., Osinski, B., Ichter, B., Levine, S.: LM-nav: Robotic navigation with large pre-trained models of language, vision, and action. In: CoRL (2022) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "76. Shan, W., Liu, Z., Zhang, X., Wang, Z., Han, K., Wang, S., Ma, S., Gao, W.: Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In: ICCV (2023) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "77. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) 3, 9, 11, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "78. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: NeurIPS (2023) 2, 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "79. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631 (2023) 2, 3, 4, 6, 9, 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "80. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. arXiv preprint arXiv:2306.03881 (2023) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 666 + ], + "type": "text", + "content": "81. Tatarchenko, M., Park, J., Koltun, V., Zhou., Q.Y.: Tangent convolutions for dense prediction in 3D. CVPR (2018) 10" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 456 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "82. Tchapmi, L.P., Choy, C.B., Armeni, I., Gwak, J., Savarese, S.: Segcloud: Semantic segmentation of 3d point clouds. In: 3DV (2017) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "83. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "84. Wang, J., Rupprecht, C., Novotny, D.: Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In: ICCV (2023) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "85. Wang, J., Li, X., Zhang, J., Xu, Q., Zhou, Q., Yu, Q., Sheng, L., Xu, D.: Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773 (2023) 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "text", + "content": "86. Wang, T., Li, J., An, X.: An efficient scene semantic labeling approach for 3d point cloud. In: ITSC (2015) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 269 + ], + "type": "text", + "content": "87. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) 2, 3, 5, 7, 8, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 292 + ], + "type": "text", + "content": "88. Xu, Z., He, Z., Wu, J., Song, S.: Learning 3d dynamic scene representations for robot manipulation. arXiv preprint arXiv:2011.01968 (2020) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 312 + ], + "type": "text", + "content": "89. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023) 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "text", + "content": "90. Zhang, J., Dong, R., Ma, K.: Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. arXiv preprint arXiv:2303.04748 (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "91. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "92. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. In: ICCV (2023) 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "93. Zheng, M., Wang, F., You, S., Qian, C., Zhang, C., Wang, X., Xu, C.: Weakly supervised contrastive learning. In: ICCV (2021) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "94. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision (ECCV) (2022) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "95. Zhu, X., Chen, J., Zeng, X., Liang, J., Li, C., Liu, S., Behpour, S., Xu, M.: Weakly supervised 3d semantic segmentation using cross-image consensus and inter-voxel affinity relations. In: ICCV (2021) 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open Vocabulary 3D Segmentation with Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_content_list.json b/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c38a878d870ea5495a635d7d27e2383f6c25d03b --- /dev/null +++ b/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_content_list.json @@ -0,0 +1,1607 @@ +[ + { + "type": "text", + "text": "Open-Vocabulary Camouflaged Object Segmentation", + "text_level": 1, + "bbox": [ + 282, + 140, + 722, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youwei Pang $^{1,2\\star}$ , Xiaqi Zhao $^{1,2\\star}$ , Jiaming Zuo $^{2}$ , Lihe Zhang $^{1\\star\\star}$ , and Huchuan Lu $^{1}$", + "bbox": [ + 326, + 212, + 673, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dalian University of Technology \n $^{2}$ X3000 Inspection Co., Ltd \n{lartpang, zxq} $@$ mail.dlut.edu.cn \nklaus@3000gy.com, {zhanglihe, lhchuan} $@$ dlut.edu.cn", + "bbox": [ + 308, + 253, + 691, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recently, the emergence of the large-scale vision-language model (VLM), such as CLIP, has opened the way towards open-world object perception. Many works have explored the utilization of pre-trained VLM for the challenging open-vocabulary dense prediction task that requires perceiving diverse objects with novel classes at inference time. Existing methods construct experiments based on the public datasets of related tasks, which are not tailored for open vocabulary and rarely involve imperceptible objects camouflaged in complex scenes due to data collection bias and annotation costs. To fill in the gaps, we introduce a new task, open-vocabulary camouflaged object segmentation (OVCOS), and construct a large-scale complex scene dataset (OVCampo) containing 11,483 hand-selected images with fine annotations and corresponding object classes. Further, we build a strong single-stage open-vocabulary camouflaged object segmentation transformer baseline OVCoser attached to the parameter-fixed CLIP with iterative semantic guidance and structure enhancement. By integrating the guidance of class semantic knowledge and the supplement of visual structure cues from the edge and depth information, the proposed method can efficiently capture camouflaged objects. Moreover, this effective framework also surpasses previous state-of-the-arts of open-vocabulary semantic image segmentation by a large margin on our OVCampo dataset. With the proposed dataset and baseline, we hope that this new task with more practical value can further expand the research on open-vocabulary dense prediction tasks. Our code and data can be found in the link.", + "bbox": [ + 261, + 351, + 738, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open Vocabulary $\\cdot$ Camouflaged Object Segmentation", + "bbox": [ + 261, + 696, + 712, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 738, + 374, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to identify and reason about object regions in a visual scene is essential for a wide range of human activities. As a complex and fundamental task in", + "bbox": [ + 215, + 771, + 785, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution.", + "bbox": [ + 217, + 811, + 367, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding Author.", + "bbox": [ + 215, + 825, + 387, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "computer vision, detecting and segmenting objects with diverse appearances in complex scenes is also an important challenge, which is crucial for applications across vision and robotics fields, including autonomous driving [3], medical image analysis [34], and intelligent robotics [84], to name a few. In the past few years, numerous typical methods [6, 7, 19, 69] have emerged with the help of a mass of labeled data, which have greatly promoted the development of related fields such as semantic image segmentation (SIS) [11,54]. However, existing works have mainly focused on predefined closed-set scenarios, where all semantic concepts are seen during both the inference and training phases. Such a scenario setting oversimplifies the real-world complexity. To this end, many explorations have been contributed to open vocabulary settings [56, 83]. In existing work, open vocabulary learning utilizes visual-related language data (e.g., object class text or descriptions) to align images and language features, enabling models to perceive novel classes without extensive labeled data [56]. Recently, large-scale pretrained visual language models (VLMs) such as CLIP [9,47] have been gaining attention. Image-text matching-based learning mechanism provides them with the ability to align textual and visual signals well, and many works demonstrate their potential for open vocabulary tasks. Besides, data plays an important role in open-vocabulary tasks. Existing open-vocabulary semantic image segmentation (OVSIS) tasks rely on related public datasets [4,12,40,41,79], while they are not designed for the open-vocabulary setting, and there is high semantic similarity between their class definitions as revealed in [59]. Due to the data collection bias and annotation cost constraints, the existing open-vocabulary benchmarks lack special attention to finely perceive the objects of interest in concealed scenes. And the widely used VLMs are pre-trained on image-text pairs with inherent object concept bias, thus, their ability to segment objects in complex scenes remains to be verified.", + "bbox": [ + 212, + 146, + 787, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a new open-vocabulary segmentation task OVCOS dedicated to analyzing camouflaged object perception in diverse natural scenes. And a large-scale data benchmark, named OVCamo, is carefully constructed. Besides, we also design a strong baseline OVCoser for the proposed OVCOS, based on the VLM-driven single-stage paradigm. The camouflage arises from several sources, including similar patterns to the environment (e.g., color and texture) and imperceptible attributes (e.g., small size and heavy occlusion) as statistically illustrated in Fig. 3.", + "bbox": [ + 212, + 556, + 486, + 797 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7051fe0903c93b57e760b3d3fc06af83e0bdb08e8028394555ff2fc233ebb1ef.jpg", + "image_caption": [ + "Fig.1: The perception and recognition of camouflaged objects require the collaboration of information from multiple sources such as appearance clues, spatial structure, contour details, and object semantics." + ], + "image_footnote": [], + "bbox": [ + 500, + 585, + 785, + 704 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3 To ensure the reliability of the definition, we follow existing published literature and collect data from the class-agnostic camouflage perception field, i.e., CSU [15].", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/15933d076926daf2632908164e92c8e2ffdefb4c006b632973dc4b56a762a330.jpg", + "image_caption": [ + "Fig. 2: Class distribution in the proposed OVCamo dataset and some visual examples." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the imperceptible appearance of camouflaged objects, accurate recognition and capture actually depend more on the cooperation of multi-source knowledge. As shown in Fig. 1, in addition to visual appearance cues, we introduce the depth for the spatial structure of the scene, the edge for the regional changes about objects, and the text for the context-aware class semantics. Considering the cooperative relationship between class recognition and object perception, the iterative learning strategy is introduced to feed back the optimized semantic relationship, resulting in more accurate object semantic guidance. This top-down conceptual reinforcement can further optimize open-vocabulary segmentation performance. With the help of the iterative multi-source information joint learning strategy, our method OVCoser shows good performance in the proposed OVCOS task.", + "bbox": [ + 212, + 368, + 787, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold as follows:", + "bbox": [ + 238, + 549, + 640, + 563 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- New Challenge. In view of the limitations of the existing OVSIS, we introduce a more challenging OVCOS task for open-vocabulary segmentation of camouflaged objects.", + "- New Benchmark. A new large-scale benchmark OVCamo with diverse samples carefully collected from existing publicly available data is proposed to better evaluate and analyze the generalization of algorithms on the proposed task.", + "- **Strong Baseline.** We build a robust single-stage baseline based on CLIP, in which the proposed iterative semantic guidance and structure enhancement are embedded. Under the joint optimization of multi-source information, our approach OVCoser outperforms existing OVSIS algorithms on the new benchmark." + ], + "bbox": [ + 225, + 571, + 784, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 780, + 395, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Pre-training. The core goal of visual-language pre-training is to learn generic representations of vision and language, and connect visual and", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "language concepts. Early approaches [5, 29, 32, 52] are based on some relatively small and clean public datasets, which limits their achievable performance, and their fine-tuning on specific downstream tasks hinders application flexibility. In [39], image-text pairs collected from the Internet bring clear performance improvement for the retrieval task. This also indirectly encourages subsequent further exploration, such as CLIP [9,47]. It benefits from the larger scale of noisy data from web pages covering diverse and rich concepts and exhibits impressive open-vocabulary capabilities. This work introduces CLIP as the bedrock of open-vocabulary capability and borrows its strong image-text matching capability to build an effective baseline for the OVCOS task.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open Vocabulary Semantic Image Segmentation (OVSIS). Although a variety of pipelines have emerged as summarized in [83], from an overall perspective, their efforts are similar, namely, how to align class name/description semantic embedding with visual features to anchor relevant object cues in the representation space. The early pioneer work [70] attempts to connect word concepts and semantic relations, and encodes word concept hierarchy to parse images. Due to the leading performance of the VLM on the image-text joint modeling, it has been gradually applied in the OVSIS field. In terms of structure, existing schemes can be roughly categorized into two types: two-stage [33,58,60,65] and single-stage [10,64,80]. In [65], a rough segmentation map for each class is created from a VLM and then refined by the test-time augmentation. These rough maps are utilized as pseudo-labels for subsequent fine segmentation by stochastic pixel sampling. SimSeg [60] adopts a cascaded design including class-agnostic proposal generation by MaskFormer [7] and class assignment by CLIP [47]. Furthermore, OVSeg [33] fine-tunes CLIP on the noisy but diverse data to improve its generalization to masked images. In lieu of using existing SIS models, a text-to-image diffusion model is introduced to generate mask features with implicit image captions in [58]. The single-stage design is more flexible and simpler. MaskCLIP [80] directly modifies CLIP for semantic segmentation without training, while SAN [59] achieves better performance with the help of adapters. CAT-Seg [10] highlights the importance of the cost aggregation between image and text embeddings for the OVSIS decoding. Recent FC-CLIP [64] investigates the hierarchical CLIP image encoder. Although these methods show impressive OVSIS performance, they still follow the generalized segmentation paradigm [6,7] in SIS, ignoring valuable auxiliary cues for object perception. This also causes them to struggle with objects camouflaged in complex scenes. Unlike them, our method, which is tailored for OVCOS, can effectively tap into the camouflaged object by integrating task-specific multi-source knowledge from visual appearance, spatial structure, object contour, and class semantics.", + "bbox": [ + 212, + 303, + 787, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camouflaged Scene Understanding (CSU). This is a research hotspot in the computer vision community, aiming to perceive objects with camouflage [15]. Different from traditional object perception like salient object detection [22, 23, 30, 31, 42, 45, 46, 63, 66–68, 72–77] and semantic object segmentation [21, 71], CSU is obviously a more challenging problem. It can be applied in some specific fields, such as medical analysis [13, 16] and agricultural man", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "agement [35, 49]. This topic is currently defined as the class-agnostic form, focusing on the area of camouflaged objects in the visual scene. The available work [8, 13, 14, 24, 28, 43, 44, 62] to date has demonstrated promising performance on existing data benchmarks [2, 8, 14, 27, 38, 51]. Unlike previous settings, the proposed OVCOS task requires further perception of object classes. Admittedly, the publicly available data provides critical support for this new task. It helps us take the first step.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 OVCamo Dataset", + "text_level": 1, + "bbox": [ + 214, + 277, + 421, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work focuses on open-vocabulary segmentation in the camouflaged scene. It enriches the connotation of OVSIS and provides a more challenging benchmark. Note that our aim is not to claim novelty through class annotations alone, but rather to focus on how such annotations can pave the way for new tasks and methods. Our dataset provides additional features such as clean data, less ambiguity, and sensible class hierarchies, which provide the necessary data for solving the new and complex task. It not only adds complexity and challenge, but also promotes innovation and advances the field through the need for advanced methods. We believe that this dataset can lead to new opportunities in areas such as wildlife conservation and intelligence sensing.", + "bbox": [ + 212, + 311, + 787, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Collection. Our data is collected from existing CSU datasets that have finely annotated segmentation maps. Specifically, the OV-Camo integrates 11,483 hand-selected images covering 75 object classes reconstructed from several public datasets [2, 8, 14, 61, 78]. The distribution of the number of samples in different classes is shown in Fig. 2. Meanwhile, we consider attributes of objects when selecting images, such as object concentration, average color ratio, object-image area ratio, number", + "bbox": [ + 212, + 462, + 486, + 672 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/87288ba2cf10ba746f92188663604bc7473ab6c60629936fe6d656cf21caeeab.jpg", + "table_caption": [ + "Table 1: Attributes involved in the dataset analysis." + ], + "table_footnote": [], + "table_body": "
AttributeDescription
Object ConcentrationObject pixel concentration, which calculates the area ratio between the object region and its minimum rotatable bounding box.
Average Color RatioColor ratio of the object to the background, which calculates the average of the three color channels in their respective regions.
Object-Image Area RatioArea ratio of the object relative to the image.
Number of Object PartsNumber of separate areas in the image that belong to the object.
Normalized CentroidObject centroid coordinates, which are normalized using the image shape.
", + "bbox": [ + 496, + 534, + 782, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of object parts, and normalized centroid. Tab. 1 gives their definitions. And Fig. 3 visualizes the attribute distribution of the proposed dataset. The camouflaged objects of interest usually have complex shape Fig. 3a, high similarity to the background Fig. 3b, and small size Fig. 3c. And the image often contains multiple camouflaged objects or sub-regions with a central bias as shown in Figs. 3d and 3e.", + "bbox": [ + 212, + 674, + 787, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotation. The original annotation cannot be used directly in the open-vocabulary setting, due to the following semantic ambiguities caused by different annotation standards, i.e., 1) Broad concepts, such as \"fish\" and \"bug\"; 2) Vague definitions, such as \"small fish\" and \"black cat\"; 3) Inconsistent granularity, such as the coexistence of \"orchid mantis\" and \"mantis\"; 4) Non-", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c782ee5c2f964548f52a19423457a3aaf7e420ac8f8c0afc6769d410db394d68.jpg", + "image_caption": [ + "(a) Object Concentration." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 401, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c97fb15807e94d5f10c05f70741d4a3bf8e6fd0b429f63f7f7ec2b508af86437.jpg", + "image_caption": [ + "(b) Average Color Ratio." + ], + "image_footnote": [], + "bbox": [ + 406, + 143, + 584, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62e7170139bc7d1baf091e121a780fb85c2bedfba0578a1cc3cae7741dea61a0.jpg", + "image_caption": [ + "(c) Object-Image Area Ratio.", + "Fig. 3: Attribute visualization of objects in our OVCampo dataset, including object concentration, object-background color ratio, object-image area ratio, number of object parts, and normalized centroid. Please refer to Tab. 1 for details about attributes." + ], + "image_footnote": [], + "bbox": [ + 222, + 227, + 401, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d70c181511b1804039f21147df1f7722d96498ca630affdf21e8fa183371c299.jpg", + "image_caption": [ + "(d) Number of Object Parts." + ], + "image_footnote": [], + "bbox": [ + 406, + 228, + 581, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5347468efba2f7697113810c0c2056e4576654a558a32bc3c64a69445569da9.jpg", + "image_caption": [ + "(e) Normalized Centroid." + ], + "image_footnote": [], + "bbox": [ + 588, + 146, + 781, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "entity concepts, such as \"other\". These issues can lead to unreasonable and unreliable results for the open-vocabulary prediction as discussed in [81]. To this end, we relabel the classes of all camouflaged objects and take the generality of the concept as the criterion for class definition, which also ensures lower semantic similarity. The similarity analysis and rough hierarchy of the class set $\\mathcal{C}$ can be found in the supplementary material.", + "bbox": [ + 212, + 393, + 787, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Division. To objectively evaluate the open-vocabulary segmentation algorithm on unseen classes, we assign as many classes as possible to the test set and control the sample ratio of the training set to the testing set to be 7:3. Specifically, 14 classes in the dataset are taken as the training set and all the remaining 61 classes are used for testing. Such a setting follows the existing practice in OVSIS where the number of seen classes (e.g., 171 [4]) is usually fewer than unseen classes (e.g., 847 [79]) which is also closer to the real-world setting. Such a setup can ensure the quantity of training samples while reinforcing the complexity of the test set. Finally, the overall ratio of samples is 7713:3770.", + "bbox": [ + 212, + 484, + 787, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 642, + 380, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce a strong baseline OVCoser. The overall framework is first described, followed by the details of the key components.", + "bbox": [ + 212, + 672, + 785, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Architecture. We follow the common encoder-decoder paradigm and the pipeline is shown in Fig. 4. Specifically, we first leverage the textual encoder $\\mathbf{E}_{\\mathbf{t}}$ of the frozen CLIP to extract semantic embedding $f_{t}$ from the class label set $\\mathcal{C}$ , and the visual encoder $\\mathbf{E}_{\\mathbf{v}}$ to extract multi-scale image features $\\{f^{i}\\}_{i = 1}^{5}$ . Both of them are fed into the decoder as the information bedrock of object segmentation, as shown in Fig. 5. Structural cues such as depth and edge are also introduced to assist the iterative refinement process. Finally, the class-related segmentation $P_{s}$ which is the segmentation logits $M_{s}$ after the sigmoid function processing, is used to remove the interference from the background in the high-level image", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98d7bc5ca58c4f9b33fea5e9602818bdff4276a6ebd2c44fadc098d84aef0019.jpg", + "image_caption": [ + "Fig. 4: Overview of our single-stage open-vocabulary camouflaged object segmentation framework, OVCoser. It is based on the frozen CLIP model which includes feature encoder $\\mathbf{E}_v$ and embedding layer $\\mathbf{T}_v$ for visual appearance cues and textual encoder $\\mathbf{E}_t$ for class semantic information. The well-constructed prompt template set Camo-Prompts $\\mathcal{P}$ enables textual embedding $f_t$ to be more appropriate to OVCOS. In the iterative refinement decoder (Fig. 5), a more accurate class-related segmentation $P_s$ , can be obtained with the assistance of the semantic guidance $\\mathbf{C}$ (Fig. 6a) from the class semantic and the structure enhancement (Fig. 6b) from the auxiliary depth and edge supervisions, i.e., $\\mathbf{D}$ and $\\mathbf{E}$ . After being masked average pooled (MAP) by $P_s$ , the high-level feature $f^5$ is used to assign a class $P_c$ to the object from the class set $\\mathcal{C}$ . Note that $\\mathbf{D}$ and $\\mathbf{E}$ are only used as auxiliary supervision during training and are not required during inference." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "feature and guide the generation of object-oriented visual representation $f_{v}$ . And the class label $P_{c}$ is determined by the similarity matching between $f_{v}$ and $f_{t}$ . More details about the proposed architecture are listed in the supplementary material.", + "bbox": [ + 212, + 546, + 787, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Guidance (SG). The class label definition usually is independent of the image scene. It is very important to fully utilize the class prior for object recognition in complex scenes. In the proposed decoder, normalized textual embedding is introduced into each stage to highlight semantically relevant cues. As shown in Fig. 6a, we design a semantic guidance component SG to inject concept cues into the self-enhancement of the image feature. Specifically, in our semantic guidance attention (SGA), the normalized image feature $X$ is linearly mapped to $Q$ , $K$ , and $V$ , while the textual embedding $f_{t}$ is transformed to the class guidance vector $G_{t}$ . The similarity between $Q$ and $G_{t}$ reflects the activation of different classes in spatial locations. In Agg of Fig. 6a, the base weight $W_{b}$ for the spatial guidance is obtained after highlighting the most relevant class information by softmax operation. And then $V$ is modulated and fed into MHSA, i.e., multi-head self-attention [55].", + "bbox": [ + 212, + 609, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Structure Enhancement (SE). Existing methods demonstrate that low-level structural information, such as the edge [53] and the depth [57], plays an important role in the design of neural networks.", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14a8ad6ee324e9607f42bbf7fea10bb3c82c438b707e7854408c3893b4ced91a.jpg", + "image_caption": [ + "Fig. 5: Proposed pipeline of the iterative refinement decoder with semantic guidance (SG) and structure enhancement (SE) components denoted as \"SG- $\\star$ \" and \"SE- $\\star$ \". $t$ and $T$ denote the current step and total number of iterations, respectively." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 727, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/688eea1ab6c8901f7cc64598553b5ac3d07abb0b4da8eb5ddc8f59b7749019bb.jpg", + "image_caption": [ + "(a) SG" + ], + "image_footnote": [], + "bbox": [ + 232, + 345, + 485, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cade2dedc198785643b56f9429a0e2194f2b5d60d85332e0c2aa122919bb6c40.jpg", + "image_caption": [ + "(b) SE", + "Fig. 6: Proposed semantic guidance (SG) and structure enhancement (SE) components in which semantic guidance attention (SGA) and structure enhancement attention (SEA) are embedded. The dashed blocks represent the feedback paths that come into play during iteration." + ], + "image_footnote": [], + "bbox": [ + 532, + 345, + 756, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tant role in CSU, which is closely related to the mechanism of the human visual system. So the SE component attached to the low-level SG is proposed to integrate the edge-aware and depth-aware cues and improve the structural details. Specifically, the output of the SG is fed into two separate branches containing the convolutional stem and head for the edge and depth estimation as shown in Fig. 5. The edge and depth logits maps, i.e., $M_e^i$ and $M_d^i$ , from the head in the branch of the layer $i \\in \\{1,2,3\\}$ are directly supervised. And the outputs $f_e^i$ and $f_d^i$ of the stem are fed into the SE. In the structure enhancement attention (SEA), they independently update the normalized visual features $X$ using MHSA [55], and the corresponding outputs are combined with the learnable weight $\\alpha$ as in Fig. 6b.", + "bbox": [ + 212, + 566, + 787, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iterative Refinement. In the SG component, the aggregation process between image features and class semantics is not aligned and requires data-driven optimization. Considering the aligned embedding space of the pre-trained CLIP, we introduce the correlation matrix $M_{cor}$ between the visual and textual embeddings into the SG as shown in Fig. 6a. Meanwhile, due to the emphasis of the decoder output on the object region, the object-aware representation $f_{obj}$ is also inputted, which comes from the image features pooled by the coarse", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "segmentation prediction in the last iteration. By combining the two, we obtain task-oriented object cues, which is actually inspired by the top-down attention mechanism in the human cognitive system [1,25]. The spatial activation map $W_{r}$ of such object cues over image features is used to re-modulate $W_{b}$ . Besides, the SE in the iteration also helps the model to further optimize the texture details. To benefit as much as possible from the assistance from structure enhancement while avoiding over-computation, we set the iteration entry to the third decoding layer as shown in Fig. 5.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CamoPrompts. As mentioned in [47], prompt engineering and ensembling are important for the transfer performance of CLIP on downstream tasks, and the prompt template should be more relevant to the data type. Because additional task-related cues are generally able to impose the necessary contextual constraint to the flexible CLIP. Hence, instead of common practices [18, 47, 82], we design a simpler yet more effective template set CamoPrompts tailored for OVCOS to decorate the class name, and average their textual embeddings as the final semantic embedding for each class. Its full form is depicted in Tab. 4, while it also achieves better classification performance in comparison with other forms.", + "bbox": [ + 212, + 268, + 787, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Supervision. In each iteration, in addition to semantic object segmentation, we also need to perform depth estimation and edge estimation as auxiliary tasks. For the segmentation prediction, we follow the commonly used weighted segmentation loss function $l_{s}^{t} = l_{s}(P_{s}^{t},G_{s})$ [53,62]. For the edge estimation, considering the imbalance problem of positive and negative samples, we introduce the dice loss function as $l_{e}^{i,t} = l_{e}(P_{e}^{i,t},G_{e})$ . The summation of L1 and SSIM losses, i.e., $l_{d}^{i,t} = l_{d}(P_{d}^{i,t},G_{d})$ , is used for the depth estimation. The total loss $L$ of our method can be formulated as follows: $L = \\sum_{t=1}^{T}\\left[l_{s}^{t} + \\sum_{i=1}^{3}\\left(l_{e}^{i,t} + l_{d}^{i,t}\\right)\\right]$ , where $t$ and $i$ are used to index iterations and layers, respectively. And the total number $T$ of iterations is set to 2 as mentioned in Sec. 5.3.", + "bbox": [ + 214, + 405, + 787, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 592, + 375, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 628, + 455, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset Settings. As mentioned in Sec. 3, the proposed dataset is divided according to classes into two disjoint subsets, i.e., $\\mathcal{C}_{seen}$ and $\\mathcal{C}_{unseen}$ . The former contains 14 classes for training and the latter contains the remaining 61 classes for testing. The proposed dataset itself is provided with only images and masks. We use the typical monocular depth estimation methods [26,48] to obtain the depth map $G_{d}$ for training, while the edge map $G_{e}$ is generated by dilating and eroding operations. We introduce depth and edge maps only in the training phase and they are not required during inference, which also avoids information leakage during testing. So, our depth and edge maps are labor-free and training-only, and directly generated by existing algorithms. Such a design can help our model get rid of the dependence on them in real-world applications. Our structure enhancement based on depth and edge information belongs to a deeper exploration", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/90754cec10618a0faa07e5dd09531883e702474a4cbd1e7fd5c5b512665c2d20.jpg", + "image_caption": [ + "Fig. 7: Visual results on OVCamo. Existing methods are either disrupted by chaotic backgrounds, imperceptible appearances, blurry details, or severe occlusion, while our algorithm can effectively capture and remain well-exposed object details. And three different colors are used to represent human annotations, correct predictions, and incorrect predictions." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of this task. These generated depth maps will be made publicly available with the dataset.", + "bbox": [ + 212, + 474, + 784, + 503 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Details. Following previous settings in [59, 60, 64], the pre-trained CLIP is frozen during training, and the remaining parameters are learnable and are randomly initialized. The AdamW [37] optimizer with the learning rate of 3e-6, weight decay of 5e-4, batch size of 4, and training epoch of 30 is used to optimize model parameters. Some basic data augmentations including random flipping, rotating, and color jittering, are introduced to preprocess training data. The input data and the output logits are bilinearly interpolated to $384 \\times 384$ during training and inference.", + "bbox": [ + 212, + 505, + 785, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol. To reasonably evaluate the performance of OVCOS, we modify the metrics in the original CSU task [8,13] to $\\mathrm{cS}_m$ , $\\mathrm{cF}_{\\beta}^{\\omega}$ , cMAE, $\\mathrm{cF}_{\\beta}$ , $\\mathrm{cE}_m$ , and cIoU which follows the common settings in the OVSIS field [10,33,59, 60,64,83] and takes into account both classification and segmentation.", + "bbox": [ + 212, + 626, + 785, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 System-level Comparison", + "text_level": 1, + "bbox": [ + 214, + 709, + 470, + 724 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To show the complexity of the proposed OVCOS task and also to verify the effectiveness of the proposed method, we compare OVCoser with several recent state-of-the-art methods in OVSIS, including [10,33,58-60,64]. Since this is a new task, existing methods need to be re-evaluated to understand their generalization ability. Based on the public code and weights trained on COCO-Stuff [4] provided by the authors, we show the performance of these methods under three different testing schemes, including S.I) testing directly with their trained weights; S.II)", + "bbox": [ + 212, + 733, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1c13bee2fcd4e7bf16c8ebfcafd0f9e9ca7115f6825adbfac77b3d54af693415.jpg", + "table_caption": [ + "Table 2: Comparison with recent state-of-the-art CLIP-based open-vocabulary semantic image segmentation methods with different training settings on the proposed OVCamo dataset. The best three results are highlighted in red, green and blue." + ], + "table_footnote": [], + "table_body": "
ModelVLMFeature BackboneText PromptcSm↑cFωβ↑cMAE↓cFβ↑cEm↑cloU↑
Test on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.1280.1050.8380.1120.1430.094
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.3410.3060.5840.3250.3840.273
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.4090.3390.5000.3410.4210.302
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.4140.3430.4890.3570.4560.319
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.4300.3440.4480.3660.4590.310
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.3740.3060.5390.3200.4090.285
Finetune on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0980.0710.8520.0810.1280.066
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.1640.1310.7630.1470.2080.123
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1820.1250.6910.2190.3090.189
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.3210.2160.5500.2360.3310.204
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1850.0940.7020.1100.1850.088
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.1240.0740.7980.0880.1620.072
Train on OVCamo.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0530.0490.9210.0560.0980.047
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.0240.0460.9540.0560.1300.046
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1870.1190.7000.2110.2980.167
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.2750.2020.6120.2200.3180.189
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1810.1060.7190.1230.1960.094
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.0800.0760.8720.0900.1910.072
OursCLIP-ConvNeXt-L [9]—CamoPrompts0.5790.4900.3360.5200.6160.443
", + "bbox": [ + 215, + 199, + 782, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "further fine-tuning based on trained weights before testing; and S.III) testing after re-training directly on our training set.", + "bbox": [ + 212, + 488, + 784, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Evaluation. For the sake of fairness in comparisons, we report the performance of the \"Large\" versions for these methods, except for SimSeg [60] where the authors only provide the \"Base\" version. All results are summarized in Tab. 2 and our approach consistently outperforms these competitors. It is worth noting that existing methods perform better at S.I. This may be attributed to the training process on the larger-scale COCO-Stuff dataset, which provides a more general understanding of the concepts. However, direct fine-tuning (i.e., S.II) may destroy this knowledge and even cause oblivion to some extent, resulting in performance degradation. If we follow S.III to re-train them, the relatively small-scale training data may not be enough to train these models with more complex structures. At the same time, the existing methods also lack targeted optimization for the OVCOS task. These problems can lead to further deterioration of performance. However, our approach tailored for OVCOS achieves leading performance by the iterative refinement strategy of multi-source information, which comprehensively considers different characteristics of the task.", + "bbox": [ + 212, + 521, + 784, + 746 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Evaluation. We also visualize the results of some recent methods on a variety of data in Fig. 7. It can be seen that the proposed method shows better performance and adaptability to diverse objects, including large objects (Col. 1-2), middle objects (Col. 3-5), small objects (Col. 6-9), multiple objects (Col. 8), complex shapes (Col. 3-5), blurred edges (Col. 1-5), severe occlusion (Col. 6), and background interference (Col. 2-6).", + "bbox": [ + 214, + 748, + 782, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/53c314442fd23bf85afe9affbc4394d0d326aa93002cfa4afc67846b41526484.jpg", + "table_caption": [ + "Table 3: Ablation comparison of proposed components. $\\Delta$ represents the average relative gain in performance of the corresponding model over the baseline for OVCOS. $\\mathcal{P}$ : CamoPrompts. $C$ : Semantic guidance. $D$ : Depth estimation auxiliary task. $E$ : Edge estimation auxiliary task. $T$ : Number of iterations which is set to 2 by default due to the best performance. \"lim $\\sup_{P_s \\to G_s}$ Perf.\" The ideal performance for our framework." + ], + "table_footnote": [], + "table_body": "
ModelcSm↑ cFβ↑ cMAE ↓ cFβ ↑ cEm ↑ cIoU ↑Δ
Comparison of the proposed modules.
Baseline0.5170.4080.3740.4510.5490.3590.0%
+P0.5430.4350.3460.4800.5810.3836.3%
+P,C0.5500.4530.3410.4910.5970.3979.1%
+P,C,D0.5650.4730.3360.5070.6060.42212.6%
+P,C,E0.5670.4810.3390.5110.6070.43213.5%
+P,C,D,E (i.e., T=1)0.5700.4880.3380.5180.6100.43614.5%
SE (Fig. 6b) → Addition Fusion0.5520.4570.3400.4970.5990.4029.9%
Comparison of the proposed iterative refinement.
T=10.5700.4880.3380.5180.6100.43614.5%
T=20.5790.4900.3360.5200.6160.44315.5%
w/o Mcor as in Fig. 6a0.5750.4870.3370.5150.6110.44114.8%
w/o obj as in Fig. 6a0.5710.4760.3390.5060.6080.43413.4%
T=30.5760.4840.3330.5140.6140.43714.8%
lim supPseg→GsegPerf.0.7030.7030.2970.7010.7010.70151.2%
", + "bbox": [ + 267, + 226, + 720, + 438 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Analysis and Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 469, + 501, + 486 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Better Baseline Model. Interestingly, our baseline model in Tab. 3 outperforms existing methods in Tab. 2. We attribute this to the fact that the overly complex feature pipeline of the conventional OVSIS architecture is not well adapted to this task. And our framework, from hyperparameter setting to structural evolution, is explored for the proposed task and data. Thus, we actually provide a more effective and robust baseline setup for OVCOS.", + "bbox": [ + 212, + 505, + 787, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Importance of Modules. To specifically analyze the effect of different components, we evaluate their performance in Tab. 3. As can be seen, the proposed modules all show positive gains. Both the semantic guidance in the class-aware decoding and the structure enhancement from auxiliary tasks of depth and edge estimation consistently boost the final performance. The ablation comparison also demonstrates that explicit guidance of the spatial and contour information is important for the detection of camouflaged objects and their positive gains are subject to the effective fusion design. We replace the proposed structure enhancement (SE) component with the vanilla addition fusion with a similar number of parameters, which results in detrimental effects. Besides, in Tab. 3, we also give the ideal results of our CLIP-driven framework where $G_{s}$ is treated as $P_{s}$ . Although our method exhibits leading performance compared to existing methods as shown in Tab. 2, there is still a long way to go to solve this problem. At the same time, the current ideal performance is still far from the limit, suggesting that future breakthroughs in this field may require more powerful paradigms as discussed in the supplementary material.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Importance of Iterative Refinement. The top-down iterative refinement strategy significantly improves the OVCOS performance as shown in Tab. 3. When the number $T$ of iterations is 2, our algorithm obtains the best OVCOS performance. As $T$ increases, there is no further improvement in performance, so it is set to 2 by default. In addition, the object-aware representation $f_{obj}$ from the output space plays an important role, and introducing the correlation guidance $M_{cor}$ also has positive gains.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Importance of CamoPrompts.", + "text_level": 1, + "bbox": [ + 212, + 252, + 485, + 267 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our prompt template set CamoPrompts, which takes task attributes into account, shows better performance in Tab. 4. To further understand the influence of different templates on the semantic embedding, we calculate the Hausdorff distance between training and testing class labels in the embedding space, as shown in Fig. 8. The figure presents an interesting phenomenon that those templates with better classification performance tend to reduce the distance, which inspires further explorations for more effective prompt engineering.", + "bbox": [ + 212, + 268, + 486, + 493 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Importance Between Edge and Depth. In the proposed SE as shown in Fig. 6b, the interaction components for the edge and depth are combined by a coefficient vector $\\alpha$ , which can also reflect the relative importance of", + "bbox": [ + 212, + 494, + 486, + 584 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the two kinds of information. In Fig. 9, we plot $\\alpha$ from different decoding layers. It can be seen that the values are usually greater than 0.5, which indicates a preference for the edge information flow.", + "bbox": [ + 212, + 584, + 785, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Model Complexity. We compare the proposed method with the CLIP-based competitors [10, 33, 58-60, 64] in terms of the number of trainable and total parameters, and FLOPs. To be fair, we test all methods following the setting of their original inference settings. As can be seen from Tab. 5, our method has fewer trainable parameters ( $< 2\\%$ ) and less computational complexity ( $0.2T$ ), which is superior to these competitors.", + "bbox": [ + 212, + 630, + 785, + 720 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 746, + 359, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this work, we propose a new challenging task, OVCOS, to explore open-vocabulary semantic image segmentation (OVSIS) for the camouflaged objects in more complex natural scenes, and carefully collect and construct a large-scale data benchmark OVCamo. Meanwhile, by considering the characteristics", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1b3d99441641a29ae4526167b1b4f06898f27b8d62d06c59dc36418b7ebeca55.jpg", + "image_caption": [ + "Fig.8: Illustration of classification accuracy $\\mathcal{A}$ and Hausdorff distance $\\mathcal{H}$ between the textual embeddings of training and testing class splits by using the plain CLIP on the OVCamo dataset. The approximate negative correlation between them may reveal that the templates that bring training and testing class embeddings closer usually also improve classification accuracy. Tab. 4 shows the details of these templates." + ], + "image_footnote": [], + "bbox": [ + 498, + 282, + 782, + 419 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 113, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3adf9f569bd1219b761fb570cb4a05905910d60f3ac4da1640f3b74d4e345a1b.jpg", + "table_caption": [ + "Table 4: Classification accuracy $\\mathcal{A}$ of the plain CLIP using different prompt templates on OVCamo, which is based on the masked average pooling with the ground truth mask." + ], + "table_footnote": [], + "table_body": "
IDPrompt TemplateA
0".class" w/o MAP based on ground truth0.538
Task-generic templates.
1".class"0.671
2.1"The.class"0.648
2.2"the.class"0.632
2.3"A.class"0.677
2.4"a.class"0.675
3.1"A photo of a.class"0.684
3.2"A photo of the.class"0.691
3.3"The photo of a.class"0.672
3.4"The photo of the.class"0.680
3.5"a photo of a.class"0.684
3.6"a photo of the.class"0.689
3.7"the photo of a.class"0.675
3.8"the photo of the.class"0.674
4templates from [47]0.686
5templates from [18]0.682
Task-related templates.
6.1"A photo of the camouflaged.class"0.681
6.2"A photo of the concealed.class"0.674
6.36.1, 6.20.685
6.4"A photo of the.classcamouflagedin the background."0.684
6.5"A photo of the.classconcealedin the background."0.694
6.66.4, 6.50.694
6.7"A photo of the.classcamouflagedto blend in with its surroundings."0.700
6.8"A photo of the.classconcealedto blend in with its surroundings."0.697
6.96.7, 6.80.703
6.106.1 - 6.9 i.e., CamoPrompts0.704
", + "bbox": [ + 215, + 186, + 777, + 393 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/af83dab3bfb4be02ccb6f28297540cc4a50cde561178f275ceb21931f0cee4d1.jpg", + "table_caption": [ + "Table 5: Efficiency comparison with other methods. \"Trainable Param.\" and \"Total Param.\" stand for the number of trainable and total parameters. In OVSeg [33], the CLIP [47] model is fine-tuned by the authors, resulting in more trainable parameters." + ], + "table_footnote": [], + "table_body": "
ModelTrainable Param.Total Param.FLOPs
SimSeg21 [60]61M (28.91%)211M1.9T
OVSeg22 [33]531M (100.00%)531M8.0T
ODISE23 [58]28M (1.80%)1522M5.5T
SAN23 [59]9M (2.06%)437M0.4T
CAT-Seg23 [10]104M (21.22%)490M0.3T
FC-CLIP23 [64]20M (5.38%)372M0.8T
Ours7M (1.95%)359M0.2T
", + "bbox": [ + 217, + 521, + 557, + 628 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3c67441ce3f068c9b28af38ea7fa0dc43c7ead4639ceff3461a87dbc0b8e12d2.jpg", + "image_caption": [ + "Fig. 9: $\\alpha$ corresponding to different heads in Fig. 6b from different decoding layers." + ], + "image_footnote": [], + "bbox": [ + 576, + 439, + 774, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "of the task and data, we propose a strong single-stage baseline OVCoser with the advanced pre-trained vision-language model. Specifically, the well-designed prompt templates are introduced to reinforce the task-relevant semantic context. We introduce additional multi-source information including class semantic cues, depth spatial structure, object edge details, and top-down iterative guidance from the output space. With the help of these components, OVCoser can perceive and segment camouflaged objects in complex environments. Extensive experiments demonstrate the effectiveness of the proposed method and its superior performance compared with the existing state-of-the-art OVSIS algorithms on OVCamo.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant 62276046 and by Dalian Science and Technology Innovation Foundation under Grant 2023JJ12GX015. We sincerely thank these authors of the public datasets [2,8,14,61,78] used in this work: Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao for COD10K [14], Yunfei Zheng, Xiongwei Zhang, Feng Wang, Tieyong Cao, Meng Sun, and Xiaobing Wang for CPD1K [78], Jinyu Yang for PlantCamo [61], Pia Bideau and Erik Learned-Miller for CAD [2], and Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge for MoCA-Mask [8]. It is their efforts that underpin this new exploration of camouflaged object segmentation.", + "bbox": [ + 212, + 172, + 787, + 339 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 359, + 321, + 376 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018)", + "2. Bideau, P., Learned-Miller, E.: It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In: Proceedings of European Conference on Computer Vision (2016)", + "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)", + "4. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018)", + "5. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: Proceedings of European Conference on Computer Vision (2020)", + "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. International Conference on Neural Information Processing Systems (2021)", + "8. Cheng, X., Xiong, H., Fan, D.P., Zhong, Y., Harandi, M., Drummond, T., Ge, Z.: Implicit motion handling for video camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "9. Cherti, M., Beaumont, R., Wightman, R., Wortzman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "0. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.W.: Catseg: Cost aggregation for open-vocabulary semantic segmentation. arXiv preprint (2023)" + ], + "bbox": [ + 225, + 388, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Csurka, G., Volpi, R., Chidlovskii, B.: Semantic image segmentation: Two decades of research. Foundations and Trends in Computer Graphics and Vision (2022)", + "12. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (2010)", + "13. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "14. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)", + "15. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence (2023)", + "16. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging (2020)", + "17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (2022)", + "18. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2021)", + "19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. IEEE Transactions on Pattern Analysis and Machine Intelligence (Feb 2020)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2016)", + "21. Ji, W., Li, J., Bian, C., Zhou, Z., Zhao, J., Yuille, A.L., Cheng, L.: Multispectral video semantic segmentation: A benchmark dataset and baseline. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)", + "22. Ji, W., Li, J., Zhang, M., Piao, Y., Lu, H.: Accurate rgb-d salient object detection via collaborative learning. In: Proceedings of European Conference on Computer Vision (2020)", + "23. Ji, W., Yan, G., Li, J., Piao, Y., Yao, S., Zhang, M., Cheng, L., Lu, H.: DMra: Depth-induced multi-scale recurrent attention network for rgb-d saliency detection. IEEE Transactions on Image Processing (2022)", + "24. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "25. Katsuki, F., Constantinidis, C.: Bottom-up and top-down attention: different processes and overlapping neural systems. The Neuroscientist (2014)", + "26. Kim, S.Y., Zhang, J., Niklaus, S., Fan, Y., Chen, S., Lin, Z., Kim, M.: Layered depth refinement with mask guidance. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "27. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding (2019)", + "28. Li, A., Zhang, J., Lyu, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Li, G., Duan, N., Fang, Y., Gong, M., Jiang, D.: Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In: AAAI Conference on Artificial Intelligence (2020)", + "30. Li, J., Ji, W., Wang, S., Li, W., Cheng, L.: Dvsod: RGB-d video salient object detection. In: International Conference on Neural Information Processing Systems (2023)", + "31. Li, J., Ji, W., Zhang, M., Piao, Y., Lu, H., Cheng, L.: Delving into calibrated depth for accurate rgb-d salient object detection. International Journal of Computer Vision (2023)", + "32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Proceedings of European Conference on Computer Vision (2020)", + "33. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "34. Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., van der Laak, J.A., van Ginneken, B., Sánchez, C.I.: A survey on deep learning in medical image analysis. Medical Image Analysis (2017)", + "35. Liu, L., Wang, R., Xie, C., Yang, P., Wang, F., Sudirman, S., Liu, W.: Pestnet: An end-to-end deep learning approach for large-scale multi-class pest detection and classification. IEEE Access (2019)", + "36. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. Proceedings of the IEEE International Conference on Computer Vision (2021)", + "37. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)", + "38. Lyu, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021)", + "39. Mithun, N.C., Panda, R., Papalexakis, E.E., Roy-Chowdhury, A.K.: Webly supervised joint embedding for cross-modal image-text retrieval. In: Proceedings of the ACM International Conference on Multimedia (2018)", + "40. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2014)", + "41. Neuhold, G., Ollmann, T., Bulo, S.R., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: Proceedings of the IEEE International Conference on Computer Vision (2017)", + "42. Pang, Y., Zhang, L., Zhao, X., Lu, H.: Hierarchical dynamic filtering network for rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020)", + "43. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "44. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)", + "46. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE Transactions on Image Processing (2023)", + "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (2021)", + "48. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE International Conference on Computer Vision (2021)", + "49. Rizzo, M., Marcuzzo, M., Zangari, A., Gasparetto, A., Albarelli, A.: Fruit ripeness classification: A survey (2023)", + "50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "51. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database (2017), http://kgwisc.aei.polsl.pl/index.php/pl/dataset/63-animal-camouflage-analysis", + "52. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. In: International Conference on Learning Representations (2019)", + "53. Sun, Y., Wang, S., Chen, C., Xiang, T.Z.: Boundary-guided camouflaged object detection. In: International Joint Conference on Artificial Intelligence (2022)", + "54. Thisanke, H., Deshan, C., Chamith, K., Seneviratne, S., Vidanaarachchi, R., Herath, D.: Semantic segmentation using vision transformers: A survey. arXiv preprint (2023)", + "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, u., Polosukhin, I.: Attention is all you need. In: International Conference on Neural Information Processing Systems (2017)", + "56. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. arXiv preprint (2024)", + "57. Xiang, M., Zhang, J., Lv, Y., Li, A., Zhong, Y., Dai, Y.: Exploring depth contribution for camouflaged object detection (2022)", + "58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)", + "59. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)", + "60. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: Proceedings of European Conference on Computer Vision (2021)", + "61. Yang, J.: Plantcamo dataset (2023), https://github.com/yjybuaa/PlantCamo", + "62. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection (2022)", + "63. Yu, Q., Zhao, X., Pang, Y., Zhang, L., Lu, H.: Multi-view aggregation network for dichotomous image segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In: International Conference on Neural Information Processing Systems (2023)", + "65. Zabari, N., Hoshen, Y.: Open-vocabulary semantic segmentation using test-time distillation. In: European Conference on Computer Vision Workshops (2023)", + "66. Zhang, M., Ji, W., Piao, Y., Li, J., Zhang, Y., Xu, S., Lu, H.: Lfnet: Light field fusion network for salient object detection. IEEE Transactions on Image Processing (2020)", + "67. Zhang, M., Liu, J., Wang, Y., Piao, Y., Yao, S., Ji, W., Li, J., Lu, H., Luo, Z.: Dynamic context-sensitive filtering network for video salient object detection. In: Proceedings of the IEEE International Conference on Computer Vision (2021)", + "68. Zhang, M., Yao, S., Hu, B., Piao, Y., Ji, W.: C2dfnet: Criss-cross dynamic filter network for rgb-d salient object detection. IEEE Transactions on Multimedia (2023)", + "69. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. International Conference on Neural Information Processing Systems (2021)", + "70. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: Proceedings of the IEEE International Conference on Computer Vision (2017)", + "71. Zhao, X., Chang, S., Pang, Y., Yang, J., Zhang, L., Lu, H.: Multi-source fusion and automatic predictor selection for zero-shot video object segmentation. International Journal of Computer Vision (2024)", + "72. Zhao, X., Pang, Y., Ji, W., Sheng, B., Zuo, J., Zhang, L., Lu, H.: Spider: A unified framework for context-dependent concept understanding. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024)", + "73. Zhao, X., Pang, Y., Zhang, L., Lu, H.: Joint learning of salient object detection, depth estimation and contour extraction. IEEE Transactions on Image Processing (2022)", + "74. Zhao, X., Pang, Y., Zhang, L., Lu, H., Ruan, X.: Self-supervised pretraining for rgb-d salient object detection. In: AAAI Conference on Artificial Intelligence (2022)", + "75. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Suppress and balance: A simple gated network for salient object detection. In: Proceedings of European Conference on Computer Vision (2020)", + "76. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Towards diverse binary segmentation via a simple yet general gated network. International Journal of Computer Vision (2024)", + "77. Zhao, X., Zhang, L., Pang, Y., Lu, H., Zhang, L.: A single stream network for robust and real-time rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020)", + "78. Zheng, Y., Zhang, X., Wang, F., Cao, T., Sun, M., Wang, X.: Detection of people with camouflage pattern via dense deconvolution network. IEEE Signal Processing Letters (Jan 2019)", + "79. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2017)", + "80. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Proceedings of European Conference on Computer Vision (2022)", + "81. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. ArXiv abs/2311.03352 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "OVCOS", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "82. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision (2021)", + "83. Zhu, C., Chen, L.: A survey on open-vocabulary detection and segmentation: Past, present, and future. arXiv preprint (2023)", + "84. Zhu, F., Zhu, Y., Lee, V., Liang, X., Chang, X.: Deep learning for embodied vision navigation: A survey. arXiv preprint (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 229 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Y. Pang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_model.json b/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2fd19722d41c937a1d5679f1a1c1cbbdceb3dfb4 --- /dev/null +++ b/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_model.json @@ -0,0 +1,2638 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.283, + 0.141, + 0.723, + 0.187 + ], + "angle": 0, + "content": "Open-Vocabulary Camouflaged Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.213, + 0.674, + 0.244 + ], + "angle": 0, + "content": "Youwei Pang\\(^{1,2\\star}\\), Xiaqi Zhao\\(^{1,2\\star}\\), Jiaming Zuo\\(^{2}\\), Lihe Zhang\\(^{1\\star\\star}\\), and Huchuan Lu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.255, + 0.692, + 0.313 + ], + "angle": 0, + "content": "\\(^{1}\\) Dalian University of Technology \n\\(^{2}\\) X3000 Inspection Co., Ltd \n{lartpang, zxq} \\(@\\) mail.dlut.edu.cn \nklaus@3000gy.com, {zhanglihe, lhchuan} \\(@\\) dlut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.352, + 0.74, + 0.683 + ], + "angle": 0, + "content": "Abstract. Recently, the emergence of the large-scale vision-language model (VLM), such as CLIP, has opened the way towards open-world object perception. Many works have explored the utilization of pre-trained VLM for the challenging open-vocabulary dense prediction task that requires perceiving diverse objects with novel classes at inference time. Existing methods construct experiments based on the public datasets of related tasks, which are not tailored for open vocabulary and rarely involve imperceptible objects camouflaged in complex scenes due to data collection bias and annotation costs. To fill in the gaps, we introduce a new task, open-vocabulary camouflaged object segmentation (OVCOS), and construct a large-scale complex scene dataset (OVCampo) containing 11,483 hand-selected images with fine annotations and corresponding object classes. Further, we build a strong single-stage open-vocabulary camouflaged object segmentation transformer baseline OVCoser attached to the parameter-fixed CLIP with iterative semantic guidance and structure enhancement. By integrating the guidance of class semantic knowledge and the supplement of visual structure cues from the edge and depth information, the proposed method can efficiently capture camouflaged objects. Moreover, this effective framework also surpasses previous state-of-the-arts of open-vocabulary semantic image segmentation by a large margin on our OVCampo dataset. With the proposed dataset and baseline, we hope that this new task with more practical value can further expand the research on open-vocabulary dense prediction tasks. Our code and data can be found in the link." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.697, + 0.714, + 0.712 + ], + "angle": 0, + "content": "Keywords: Open Vocabulary \\(\\cdot\\) Camouflaged Object Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.739, + 0.375, + 0.755 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.802 + ], + "angle": 0, + "content": "The ability to identify and reason about object regions in a visual scene is essential for a wide range of human activities. As a complex and fundamental task in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.369, + 0.825 + ], + "angle": 0, + "content": "* Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.217, + 0.826, + 0.388, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.555 + ], + "angle": 0, + "content": "computer vision, detecting and segmenting objects with diverse appearances in complex scenes is also an important challenge, which is crucial for applications across vision and robotics fields, including autonomous driving [3], medical image analysis [34], and intelligent robotics [84], to name a few. In the past few years, numerous typical methods [6, 7, 19, 69] have emerged with the help of a mass of labeled data, which have greatly promoted the development of related fields such as semantic image segmentation (SIS) [11,54]. However, existing works have mainly focused on predefined closed-set scenarios, where all semantic concepts are seen during both the inference and training phases. Such a scenario setting oversimplifies the real-world complexity. To this end, many explorations have been contributed to open vocabulary settings [56, 83]. In existing work, open vocabulary learning utilizes visual-related language data (e.g., object class text or descriptions) to align images and language features, enabling models to perceive novel classes without extensive labeled data [56]. Recently, large-scale pretrained visual language models (VLMs) such as CLIP [9,47] have been gaining attention. Image-text matching-based learning mechanism provides them with the ability to align textual and visual signals well, and many works demonstrate their potential for open vocabulary tasks. Besides, data plays an important role in open-vocabulary tasks. Existing open-vocabulary semantic image segmentation (OVSIS) tasks rely on related public datasets [4,12,40,41,79], while they are not designed for the open-vocabulary setting, and there is high semantic similarity between their class definitions as revealed in [59]. Due to the data collection bias and annotation cost constraints, the existing open-vocabulary benchmarks lack special attention to finely perceive the objects of interest in concealed scenes. And the widely used VLMs are pre-trained on image-text pairs with inherent object concept bias, thus, their ability to segment objects in complex scenes remains to be verified." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.487, + 0.798 + ], + "angle": 0, + "content": "In this paper, we introduce a new open-vocabulary segmentation task OVCOS dedicated to analyzing camouflaged object perception in diverse natural scenes. And a large-scale data benchmark, named OVCamo, is carefully constructed. Besides, we also design a strong baseline OVCoser for the proposed OVCOS, based on the VLM-driven single-stage paradigm. The camouflage arises from several sources, including similar patterns to the environment (e.g., color and texture) and imperceptible attributes (e.g., small size and heavy occlusion) as statistically illustrated in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.587, + 0.787, + 0.705 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.715, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Fig.1: The perception and recognition of camouflaged objects require the collaboration of information from multiple sources such as appearance clues, spatial structure, contour details, and object semantics." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "3 To ensure the reliability of the definition, we follow existing published literature and collect data from the class-agnostic camouflage perception field, i.e., CSU [15]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.326, + 0.784, + 0.34 + ], + "angle": 0, + "content": "Fig. 2: Class distribution in the proposed OVCamo dataset and some visual examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Considering the imperceptible appearance of camouflaged objects, accurate recognition and capture actually depend more on the cooperation of multi-source knowledge. As shown in Fig. 1, in addition to visual appearance cues, we introduce the depth for the spatial structure of the scene, the edge for the regional changes about objects, and the text for the context-aware class semantics. Considering the cooperative relationship between class recognition and object perception, the iterative learning strategy is introduced to feed back the optimized semantic relationship, resulting in more accurate object semantic guidance. This top-down conceptual reinforcement can further optimize open-vocabulary segmentation performance. With the help of the iterative multi-source information joint learning strategy, our method OVCoser shows good performance in the proposed OVCOS task." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.55, + 0.641, + 0.564 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.572, + 0.785, + 0.614 + ], + "angle": 0, + "content": "- New Challenge. In view of the limitations of the existing OVSIS, we introduce a more challenging OVCOS task for open-vocabulary segmentation of camouflaged objects." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.674 + ], + "angle": 0, + "content": "- New Benchmark. A new large-scale benchmark OVCamo with diverse samples carefully collected from existing publicly available data is proposed to better evaluate and analyze the generalization of algorithms on the proposed task." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.749 + ], + "angle": 0, + "content": "- **Strong Baseline.** We build a robust single-stage baseline based on CLIP, in which the proposed iterative semantic guidance and structure enhancement are embedded. Under the joint optimization of multi-source information, our approach OVCoser outperforms existing OVSIS algorithms on the new benchmark." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.572, + 0.785, + 0.749 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.781, + 0.396, + 0.796 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Vision-Language Pre-training. The core goal of visual-language pre-training is to learn generic representations of vision and language, and connect visual and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "language concepts. Early approaches [5, 29, 32, 52] are based on some relatively small and clean public datasets, which limits their achievable performance, and their fine-tuning on specific downstream tasks hinders application flexibility. In [39], image-text pairs collected from the Internet bring clear performance improvement for the retrieval task. This also indirectly encourages subsequent further exploration, such as CLIP [9,47]. It benefits from the larger scale of noisy data from web pages covering diverse and rich concepts and exhibits impressive open-vocabulary capabilities. This work introduces CLIP as the bedrock of open-vocabulary capability and borrows its strong image-text matching capability to build an effective baseline for the OVCOS task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Open Vocabulary Semantic Image Segmentation (OVSIS). Although a variety of pipelines have emerged as summarized in [83], from an overall perspective, their efforts are similar, namely, how to align class name/description semantic embedding with visual features to anchor relevant object cues in the representation space. The early pioneer work [70] attempts to connect word concepts and semantic relations, and encodes word concept hierarchy to parse images. Due to the leading performance of the VLM on the image-text joint modeling, it has been gradually applied in the OVSIS field. In terms of structure, existing schemes can be roughly categorized into two types: two-stage [33,58,60,65] and single-stage [10,64,80]. In [65], a rough segmentation map for each class is created from a VLM and then refined by the test-time augmentation. These rough maps are utilized as pseudo-labels for subsequent fine segmentation by stochastic pixel sampling. SimSeg [60] adopts a cascaded design including class-agnostic proposal generation by MaskFormer [7] and class assignment by CLIP [47]. Furthermore, OVSeg [33] fine-tunes CLIP on the noisy but diverse data to improve its generalization to masked images. In lieu of using existing SIS models, a text-to-image diffusion model is introduced to generate mask features with implicit image captions in [58]. The single-stage design is more flexible and simpler. MaskCLIP [80] directly modifies CLIP for semantic segmentation without training, while SAN [59] achieves better performance with the help of adapters. CAT-Seg [10] highlights the importance of the cost aggregation between image and text embeddings for the OVSIS decoding. Recent FC-CLIP [64] investigates the hierarchical CLIP image encoder. Although these methods show impressive OVSIS performance, they still follow the generalized segmentation paradigm [6,7] in SIS, ignoring valuable auxiliary cues for object perception. This also causes them to struggle with objects camouflaged in complex scenes. Unlike them, our method, which is tailored for OVCOS, can effectively tap into the camouflaged object by integrating task-specific multi-source knowledge from visual appearance, spatial structure, object contour, and class semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Camouflaged Scene Understanding (CSU). This is a research hotspot in the computer vision community, aiming to perceive objects with camouflage [15]. Different from traditional object perception like salient object detection [22, 23, 30, 31, 42, 45, 46, 63, 66–68, 72–77] and semantic object segmentation [21, 71], CSU is obviously a more challenging problem. It can be applied in some specific fields, such as medical analysis [13, 16] and agricultural man" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "agement [35, 49]. This topic is currently defined as the class-agnostic form, focusing on the area of camouflaged objects in the visual scene. The available work [8, 13, 14, 24, 28, 43, 44, 62] to date has demonstrated promising performance on existing data benchmarks [2, 8, 14, 27, 38, 51]. Unlike previous settings, the proposed OVCOS task requires further perception of object classes. Admittedly, the publicly available data provides critical support for this new task. It helps us take the first step." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.278, + 0.423, + 0.294 + ], + "angle": 0, + "content": "3 OVCamo Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.463 + ], + "angle": 0, + "content": "This work focuses on open-vocabulary segmentation in the camouflaged scene. It enriches the connotation of OVSIS and provides a more challenging benchmark. Note that our aim is not to claim novelty through class annotations alone, but rather to focus on how such annotations can pave the way for new tasks and methods. Our dataset provides additional features such as clean data, less ambiguity, and sensible class hierarchies, which provide the necessary data for solving the new and complex task. It not only adds complexity and challenge, but also promotes innovation and advances the field through the need for advanced methods. We believe that this dataset can lead to new opportunities in areas such as wildlife conservation and intelligence sensing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.488, + 0.674 + ], + "angle": 0, + "content": "Image Collection. Our data is collected from existing CSU datasets that have finely annotated segmentation maps. Specifically, the OV-Camo integrates 11,483 hand-selected images covering 75 object classes reconstructed from several public datasets [2, 8, 14, 61, 78]. The distribution of the number of samples in different classes is shown in Fig. 2. Meanwhile, we consider attributes of objects when selecting images, such as object concentration, average color ratio, object-image area ratio, number" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.507, + 0.787, + 0.534 + ], + "angle": 0, + "content": "Table 1: Attributes involved in the dataset analysis." + }, + { + "type": "table", + "bbox": [ + 0.498, + 0.535, + 0.784, + 0.659 + ], + "angle": 0, + "content": "
AttributeDescription
Object ConcentrationObject pixel concentration, which calculates the area ratio between the object region and its minimum rotatable bounding box.
Average Color RatioColor ratio of the object to the background, which calculates the average of the three color channels in their respective regions.
Object-Image Area RatioArea ratio of the object relative to the image.
Number of Object PartsNumber of separate areas in the image that belong to the object.
Normalized CentroidObject centroid coordinates, which are normalized using the image shape.
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.763 + ], + "angle": 0, + "content": "of object parts, and normalized centroid. Tab. 1 gives their definitions. And Fig. 3 visualizes the attribute distribution of the proposed dataset. The camouflaged objects of interest usually have complex shape Fig. 3a, high similarity to the background Fig. 3b, and small size Fig. 3c. And the image often contains multiple camouflaged objects or sub-regions with a central bias as shown in Figs. 3d and 3e." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Annotation. The original annotation cannot be used directly in the open-vocabulary setting, due to the following semantic ambiguities caused by different annotation standards, i.e., 1) Broad concepts, such as \"fish\" and \"bug\"; 2) Vague definitions, such as \"small fish\" and \"black cat\"; 3) Inconsistent granularity, such as the coexistence of \"orchid mantis\" and \"mantis\"; 4) Non-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.144, + 0.402, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.217, + 0.388, + 0.228 + ], + "angle": 0, + "content": "(a) Object Concentration." + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.144, + 0.585, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.217, + 0.567, + 0.228 + ], + "angle": 0, + "content": "(b) Average Color Ratio." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.228, + 0.403, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.3, + 0.398, + 0.312 + ], + "angle": 0, + "content": "(c) Object-Image Area Ratio." + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.229, + 0.583, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.3, + 0.578, + 0.312 + ], + "angle": 0, + "content": "(d) Number of Object Parts." + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.147, + 0.782, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.298, + 0.758, + 0.31 + ], + "angle": 0, + "content": "(e) Normalized Centroid." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.323, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: Attribute visualization of objects in our OVCampo dataset, including object concentration, object-background color ratio, object-image area ratio, number of object parts, and normalized centroid. Please refer to Tab. 1 for details about attributes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.485 + ], + "angle": 0, + "content": "entity concepts, such as \"other\". These issues can lead to unreasonable and unreliable results for the open-vocabulary prediction as discussed in [81]. To this end, we relabel the classes of all camouflaged objects and take the generality of the concept as the criterion for class definition, which also ensures lower semantic similarity. The similarity analysis and rough hierarchy of the class set \\(\\mathcal{C}\\) can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.622 + ], + "angle": 0, + "content": "Data Division. To objectively evaluate the open-vocabulary segmentation algorithm on unseen classes, we assign as many classes as possible to the test set and control the sample ratio of the training set to the testing set to be 7:3. Specifically, 14 classes in the dataset are taken as the training set and all the remaining 61 classes are used for testing. Such a setting follows the existing practice in OVSIS where the number of seen classes (e.g., 171 [4]) is usually fewer than unseen classes (e.g., 847 [79]) which is also closer to the real-world setting. Such a setup can ensure the quantity of training samples while reinforcing the complexity of the test set. Finally, the overall ratio of samples is 7713:3770." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.381, + 0.661 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.704 + ], + "angle": 0, + "content": "In this section, we introduce a strong baseline OVCoser. The overall framework is first described, followed by the details of the key components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Overall Architecture. We follow the common encoder-decoder paradigm and the pipeline is shown in Fig. 4. Specifically, we first leverage the textual encoder \\(\\mathbf{E}_{\\mathbf{t}}\\) of the frozen CLIP to extract semantic embedding \\(f_{t}\\) from the class label set \\(\\mathcal{C}\\), and the visual encoder \\(\\mathbf{E}_{\\mathbf{v}}\\) to extract multi-scale image features \\(\\{f^{i}\\}_{i = 1}^{5}\\). Both of them are fed into the decoder as the information bedrock of object segmentation, as shown in Fig. 5. Structural cues such as depth and edge are also introduced to assist the iterative refinement process. Finally, the class-related segmentation \\(P_{s}\\) which is the segmentation logits \\(M_{s}\\) after the sigmoid function processing, is used to remove the interference from the background in the high-level image" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.789, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.344, + 0.789, + 0.512 + ], + "angle": 0, + "content": "Fig. 4: Overview of our single-stage open-vocabulary camouflaged object segmentation framework, OVCoser. It is based on the frozen CLIP model which includes feature encoder \\(\\mathbf{E}_v\\) and embedding layer \\(\\mathbf{T}_v\\) for visual appearance cues and textual encoder \\(\\mathbf{E}_t\\) for class semantic information. The well-constructed prompt template set Camo-Prompts \\(\\mathcal{P}\\) enables textual embedding \\(f_t\\) to be more appropriate to OVCOS. In the iterative refinement decoder (Fig. 5), a more accurate class-related segmentation \\(P_s\\), can be obtained with the assistance of the semantic guidance \\(\\mathbf{C}\\) (Fig. 6a) from the class semantic and the structure enhancement (Fig. 6b) from the auxiliary depth and edge supervisions, i.e., \\(\\mathbf{D}\\) and \\(\\mathbf{E}\\). After being masked average pooled (MAP) by \\(P_s\\), the high-level feature \\(f^5\\) is used to assign a class \\(P_c\\) to the object from the class set \\(\\mathcal{C}\\). Note that \\(\\mathbf{D}\\) and \\(\\mathbf{E}\\) are only used as auxiliary supervision during training and are not required during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.789, + 0.606 + ], + "angle": 0, + "content": "feature and guide the generation of object-oriented visual representation \\( f_{v} \\). And the class label \\( P_{c} \\) is determined by the similarity matching between \\( f_{v} \\) and \\( f_{t} \\). More details about the proposed architecture are listed in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.789, + 0.807 + ], + "angle": 0, + "content": "Semantic Guidance (SG). The class label definition usually is independent of the image scene. It is very important to fully utilize the class prior for object recognition in complex scenes. In the proposed decoder, normalized textual embedding is introduced into each stage to highlight semantically relevant cues. As shown in Fig. 6a, we design a semantic guidance component SG to inject concept cues into the self-enhancement of the image feature. Specifically, in our semantic guidance attention (SGA), the normalized image feature \\( X \\) is linearly mapped to \\( Q \\), \\( K \\), and \\( V \\), while the textual embedding \\( f_{t} \\) is transformed to the class guidance vector \\( G_{t} \\). The similarity between \\( Q \\) and \\( G_{t} \\) reflects the activation of different classes in spatial locations. In Agg of Fig. 6a, the base weight \\( W_{b} \\) for the spatial guidance is obtained after highlighting the most relevant class information by softmax operation. And then \\( V \\) is modulated and fed into MHSA, i.e., multi-head self-attention [55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Structure Enhancement (SE). Existing methods demonstrate that low-level structural information, such as the edge [53] and the depth [57], plays an important role in the design of neural networks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.145, + 0.728, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.289, + 0.789, + 0.332 + ], + "angle": 0, + "content": "Fig. 5: Proposed pipeline of the iterative refinement decoder with semantic guidance (SG) and structure enhancement (SE) components denoted as \"SG- \\( \\star \\)\" and \"SE- \\( \\star \\)\". \\( t \\) and \\( T \\) denote the current step and total number of iterations, respectively." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.347, + 0.486, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.456, + 0.379, + 0.468 + ], + "angle": 0, + "content": "(a) SG" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.347, + 0.758, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.456, + 0.668, + 0.468 + ], + "angle": 0, + "content": "(b) SE" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.479, + 0.789, + 0.536 + ], + "angle": 0, + "content": "Fig. 6: Proposed semantic guidance (SG) and structure enhancement (SE) components in which semantic guidance attention (SGA) and structure enhancement attention (SEA) are embedded. The dashed blocks represent the feedback paths that come into play during iteration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.734 + ], + "angle": 0, + "content": "tant role in CSU, which is closely related to the mechanism of the human visual system. So the SE component attached to the low-level SG is proposed to integrate the edge-aware and depth-aware cues and improve the structural details. Specifically, the output of the SG is fed into two separate branches containing the convolutional stem and head for the edge and depth estimation as shown in Fig. 5. The edge and depth logits maps, i.e., \\( M_e^i \\) and \\( M_d^i \\), from the head in the branch of the layer \\( i \\in \\{1,2,3\\} \\) are directly supervised. And the outputs \\( f_e^i \\) and \\( f_d^i \\) of the stem are fed into the SE. In the structure enhancement attention (SEA), they independently update the normalized visual features \\( X \\) using MHSA [55], and the corresponding outputs are combined with the learnable weight \\( \\alpha \\) as in Fig. 6b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Iterative Refinement. In the SG component, the aggregation process between image features and class semantics is not aligned and requires data-driven optimization. Considering the aligned embedding space of the pre-trained CLIP, we introduce the correlation matrix \\( M_{cor} \\) between the visual and textual embeddings into the SG as shown in Fig. 6a. Meanwhile, due to the emphasis of the decoder output on the object region, the object-aware representation \\( f_{obj} \\) is also inputted, which comes from the image features pooled by the coarse" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "segmentation prediction in the last iteration. By combining the two, we obtain task-oriented object cues, which is actually inspired by the top-down attention mechanism in the human cognitive system [1,25]. The spatial activation map \\( W_{r} \\) of such object cues over image features is used to re-modulate \\( W_{b} \\). Besides, the SE in the iteration also helps the model to further optimize the texture details. To benefit as much as possible from the assistance from structure enhancement while avoiding over-computation, we set the iteration entry to the third decoding layer as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.405 + ], + "angle": 0, + "content": "CamoPrompts. As mentioned in [47], prompt engineering and ensembling are important for the transfer performance of CLIP on downstream tasks, and the prompt template should be more relevant to the data type. Because additional task-related cues are generally able to impose the necessary contextual constraint to the flexible CLIP. Hence, instead of common practices [18, 47, 82], we design a simpler yet more effective template set CamoPrompts tailored for OVCOS to decorate the class name, and average their textual embeddings as the final semantic embedding for each class. Its full form is depicted in Tab. 4, while it also achieves better classification performance in comparison with other forms." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.406, + 0.789, + 0.565 + ], + "angle": 0, + "content": "Supervision. In each iteration, in addition to semantic object segmentation, we also need to perform depth estimation and edge estimation as auxiliary tasks. For the segmentation prediction, we follow the commonly used weighted segmentation loss function \\( l_{s}^{t} = l_{s}(P_{s}^{t},G_{s}) \\) [53,62]. For the edge estimation, considering the imbalance problem of positive and negative samples, we introduce the dice loss function as \\( l_{e}^{i,t} = l_{e}(P_{e}^{i,t},G_{e}) \\). The summation of L1 and SSIM losses, i.e., \\( l_{d}^{i,t} = l_{d}(P_{d}^{i,t},G_{d}) \\), is used for the depth estimation. The total loss \\( L \\) of our method can be formulated as follows: \\( L = \\sum_{t=1}^{T}\\left[l_{s}^{t} + \\sum_{i=1}^{3}\\left(l_{e}^{i,t} + l_{d}^{i,t}\\right)\\right] \\), where \\( t \\) and \\( i \\) are used to index iterations and layers, respectively. And the total number \\( T \\) of iterations is set to 2 as mentioned in Sec. 5.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.593, + 0.377, + 0.61 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.457, + 0.645 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dataset Settings. As mentioned in Sec. 3, the proposed dataset is divided according to classes into two disjoint subsets, i.e., \\(\\mathcal{C}_{seen}\\) and \\(\\mathcal{C}_{unseen}\\). The former contains 14 classes for training and the latter contains the remaining 61 classes for testing. The proposed dataset itself is provided with only images and masks. We use the typical monocular depth estimation methods [26,48] to obtain the depth map \\(G_{d}\\) for training, while the edge map \\(G_{e}\\) is generated by dilating and eroding operations. We introduce depth and edge maps only in the training phase and they are not required during inference, which also avoids information leakage during testing. So, our depth and edge maps are labor-free and training-only, and directly generated by existing algorithms. Such a design can help our model get rid of the dependence on them in real-world applications. Our structure enhancement based on depth and edge information belongs to a deeper exploration" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 7: Visual results on OVCamo. Existing methods are either disrupted by chaotic backgrounds, imperceptible appearances, blurry details, or severe occlusion, while our algorithm can effectively capture and remain well-exposed object details. And three different colors are used to represent human annotations, correct predictions, and incorrect predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.785, + 0.505 + ], + "angle": 0, + "content": "of this task. These generated depth maps will be made publicly available with the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.787, + 0.626 + ], + "angle": 0, + "content": "Training Details. Following previous settings in [59, 60, 64], the pre-trained CLIP is frozen during training, and the remaining parameters are learnable and are randomly initialized. The AdamW [37] optimizer with the learning rate of 3e-6, weight decay of 5e-4, batch size of 4, and training epoch of 30 is used to optimize model parameters. Some basic data augmentations including random flipping, rotating, and color jittering, are introduced to preprocess training data. The input data and the output logits are bilinearly interpolated to \\(384 \\times 384\\) during training and inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.787, + 0.687 + ], + "angle": 0, + "content": "Evaluation Protocol. To reasonably evaluate the performance of OVCOS, we modify the metrics in the original CSU task [8,13] to \\(\\mathrm{cS}_m\\), \\(\\mathrm{cF}_{\\beta}^{\\omega}\\), cMAE, \\(\\mathrm{cF}_{\\beta}\\), \\(\\mathrm{cE}_m\\), and cIoU which follows the common settings in the OVSIS field [10,33,59, 60,64,83] and takes into account both classification and segmentation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.471, + 0.725 + ], + "angle": 0, + "content": "5.2 System-level Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To show the complexity of the proposed OVCOS task and also to verify the effectiveness of the proposed method, we compare OVCoser with several recent state-of-the-art methods in OVSIS, including [10,33,58-60,64]. Since this is a new task, existing methods need to be re-evaluated to understand their generalization ability. Based on the public code and weights trained on COCO-Stuff [4] provided by the authors, we show the performance of these methods under three different testing schemes, including S.I) testing directly with their trained weights; S.II)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison with recent state-of-the-art CLIP-based open-vocabulary semantic image segmentation methods with different training settings on the proposed OVCamo dataset. The best three results are highlighted in red, green and blue." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.2, + 0.784, + 0.457 + ], + "angle": 0, + "content": "
ModelVLMFeature BackboneText PromptcSm↑cFωβ↑cMAE↓cFβ↑cEm↑cloU↑
Test on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.1280.1050.8380.1120.1430.094
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.3410.3060.5840.3250.3840.273
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.4090.3390.5000.3410.4210.302
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.4140.3430.4890.3570.4560.319
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.4300.3440.4480.3660.4590.310
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.3740.3060.5390.3200.4090.285
Finetune on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0980.0710.8520.0810.1280.066
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.1640.1310.7630.1470.2080.123
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1820.1250.6910.2190.3090.189
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.3210.2160.5500.2360.3310.204
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1850.0940.7020.1100.1850.088
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.1240.0740.7980.0880.1620.072
Train on OVCamo.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0530.0490.9210.0560.0980.047
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.0240.0460.9540.0560.1300.046
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1870.1190.7000.2110.2980.167
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.2750.2020.6120.2200.3180.189
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1810.1060.7190.1230.1960.094
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.0800.0760.8720.0900.1910.072
OursCLIP-ConvNeXt-L [9]—CamoPrompts0.5790.4900.3360.5200.6160.443
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.52 + ], + "angle": 0, + "content": "further fine-tuning based on trained weights before testing; and S.III) testing after re-training directly on our training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.747 + ], + "angle": 0, + "content": "Quantitative Evaluation. For the sake of fairness in comparisons, we report the performance of the \"Large\" versions for these methods, except for SimSeg [60] where the authors only provide the \"Base\" version. All results are summarized in Tab. 2 and our approach consistently outperforms these competitors. It is worth noting that existing methods perform better at S.I. This may be attributed to the training process on the larger-scale COCO-Stuff dataset, which provides a more general understanding of the concepts. However, direct fine-tuning (i.e., S.II) may destroy this knowledge and even cause oblivion to some extent, resulting in performance degradation. If we follow S.III to re-train them, the relatively small-scale training data may not be enough to train these models with more complex structures. At the same time, the existing methods also lack targeted optimization for the OVCOS task. These problems can lead to further deterioration of performance. However, our approach tailored for OVCOS achieves leading performance by the iterative refinement strategy of multi-source information, which comprehensively considers different characteristics of the task." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.749, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Qualitative Evaluation. We also visualize the results of some recent methods on a variety of data in Fig. 7. It can be seen that the proposed method shows better performance and adaptability to diverse objects, including large objects (Col. 1-2), middle objects (Col. 3-5), small objects (Col. 6-9), multiple objects (Col. 8), complex shapes (Col. 3-5), blurred edges (Col. 1-5), severe occlusion (Col. 6), and background interference (Col. 2-6)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.219 + ], + "angle": 0, + "content": "Table 3: Ablation comparison of proposed components. \\(\\Delta\\) represents the average relative gain in performance of the corresponding model over the baseline for OVCOS. \\(\\mathcal{P}\\): CamoPrompts. \\(C\\): Semantic guidance. \\(D\\): Depth estimation auxiliary task. \\(E\\): Edge estimation auxiliary task. \\(T\\): Number of iterations which is set to 2 by default due to the best performance. \"lim \\(\\sup_{P_s \\to G_s}\\) Perf.\" The ideal performance for our framework." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.227, + 0.722, + 0.439 + ], + "angle": 0, + "content": "
ModelcSm↑ cFβ↑ cMAE ↓ cFβ ↑ cEm ↑ cIoU ↑Δ
Comparison of the proposed modules.
Baseline0.5170.4080.3740.4510.5490.3590.0%
+P0.5430.4350.3460.4800.5810.3836.3%
+P,C0.5500.4530.3410.4910.5970.3979.1%
+P,C,D0.5650.4730.3360.5070.6060.42212.6%
+P,C,E0.5670.4810.3390.5110.6070.43213.5%
+P,C,D,E (i.e., T=1)0.5700.4880.3380.5180.6100.43614.5%
SE (Fig. 6b) → Addition Fusion0.5520.4570.3400.4970.5990.4029.9%
Comparison of the proposed iterative refinement.
T=10.5700.4880.3380.5180.6100.43614.5%
T=20.5790.4900.3360.5200.6160.44315.5%
w/o Mcor as in Fig. 6a0.5750.4870.3370.5150.6110.44114.8%
w/o obj as in Fig. 6a0.5710.4760.3390.5060.6080.43413.4%
T=30.5760.4840.3330.5140.6140.43714.8%
lim supPseg→GsegPerf.0.7030.7030.2970.7010.7010.70151.2%
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.502, + 0.487 + ], + "angle": 0, + "content": "5.3 Analysis and Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Better Baseline Model. Interestingly, our baseline model in Tab. 3 outperforms existing methods in Tab. 2. We attribute this to the fact that the overly complex feature pipeline of the conventional OVSIS architecture is not well adapted to this task. And our framework, from hyperparameter setting to structural evolution, is explored for the proposed task and data. Thus, we actually provide a more effective and robust baseline setup for OVCOS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Importance of Modules. To specifically analyze the effect of different components, we evaluate their performance in Tab. 3. As can be seen, the proposed modules all show positive gains. Both the semantic guidance in the class-aware decoding and the structure enhancement from auxiliary tasks of depth and edge estimation consistently boost the final performance. The ablation comparison also demonstrates that explicit guidance of the spatial and contour information is important for the detection of camouflaged objects and their positive gains are subject to the effective fusion design. We replace the proposed structure enhancement (SE) component with the vanilla addition fusion with a similar number of parameters, which results in detrimental effects. Besides, in Tab. 3, we also give the ideal results of our CLIP-driven framework where \\( G_{s} \\) is treated as \\( P_{s} \\). Although our method exhibits leading performance compared to existing methods as shown in Tab. 2, there is still a long way to go to solve this problem. At the same time, the current ideal performance is still far from the limit, suggesting that future breakthroughs in this field may require more powerful paradigms as discussed in the supplementary material." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Importance of Iterative Refinement. The top-down iterative refinement strategy significantly improves the OVCOS performance as shown in Tab. 3. When the number \\( T \\) of iterations is 2, our algorithm obtains the best OVCOS performance. As \\( T \\) increases, there is no further improvement in performance, so it is set to 2 by default. In addition, the object-aware representation \\( f_{obj} \\) from the output space plays an important role, and introducing the correlation guidance \\( M_{cor} \\) also has positive gains." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.253, + 0.486, + 0.268 + ], + "angle": 0, + "content": "Importance of CamoPrompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.487, + 0.494 + ], + "angle": 0, + "content": "Our prompt template set CamoPrompts, which takes task attributes into account, shows better performance in Tab. 4. To further understand the influence of different templates on the semantic embedding, we calculate the Hausdorff distance between training and testing class labels in the embedding space, as shown in Fig. 8. The figure presents an interesting phenomenon that those templates with better classification performance tend to reduce the distance, which inspires further explorations for more effective prompt engineering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.488, + 0.585 + ], + "angle": 0, + "content": "Importance Between Edge and Depth. In the proposed SE as shown in Fig. 6b, the interaction components for the edge and depth are combined by a coefficient vector \\(\\alpha\\), which can also reflect the relative importance of" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.631 + ], + "angle": 0, + "content": "the two kinds of information. In Fig. 9, we plot \\(\\alpha\\) from different decoding layers. It can be seen that the values are usually greater than 0.5, which indicates a preference for the edge information flow." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.787, + 0.722 + ], + "angle": 0, + "content": "Model Complexity. We compare the proposed method with the CLIP-based competitors [10, 33, 58-60, 64] in terms of the number of trainable and total parameters, and FLOPs. To be fair, we test all methods following the setting of their original inference settings. As can be seen from Tab. 5, our method has fewer trainable parameters (\\(< 2\\%\\)) and less computational complexity (\\(0.2T\\)), which is superior to these competitors." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.747, + 0.36, + 0.763 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this work, we propose a new challenging task, OVCOS, to explore open-vocabulary semantic image segmentation (OVSIS) for the camouflaged objects in more complex natural scenes, and carefully collect and construct a large-scale data benchmark OVCamo. Meanwhile, by considering the characteristics" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.283, + 0.784, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.431, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Fig.8: Illustration of classification accuracy \\(\\mathcal{A}\\) and Hausdorff distance \\(\\mathcal{H}\\) between the textual embeddings of training and testing class splits by using the plain CLIP on the OVCamo dataset. The approximate negative correlation between them may reveal that the templates that bring training and testing class embeddings closer usually also improve classification accuracy. Tab. 4 shows the details of these templates." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.158, + 0.787, + 0.186 + ], + "angle": 0, + "content": "Table 4: Classification accuracy \\(\\mathcal{A}\\) of the plain CLIP using different prompt templates on OVCamo, which is based on the masked average pooling with the ground truth mask." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.187, + 0.778, + 0.395 + ], + "angle": 0, + "content": "
IDPrompt TemplateA
0".class" w/o MAP based on ground truth0.538
Task-generic templates.
1".class"0.671
2.1"The.class"0.648
2.2"the.class"0.632
2.3"A.class"0.677
2.4"a.class"0.675
3.1"A photo of a.class"0.684
3.2"A photo of the.class"0.691
3.3"The photo of a.class"0.672
3.4"The photo of the.class"0.680
3.5"a photo of a.class"0.684
3.6"a photo of the.class"0.689
3.7"the photo of a.class"0.675
3.8"the photo of the.class"0.674
4templates from [47]0.686
5templates from [18]0.682
Task-related templates.
6.1"A photo of the camouflaged.class"0.681
6.2"A photo of the concealed.class"0.674
6.36.1, 6.20.685
6.4"A photo of the.classcamouflagedin the background."0.684
6.5"A photo of the.classconcealedin the background."0.694
6.66.4, 6.50.694
6.7"A photo of the.classcamouflagedto blend in with its surroundings."0.700
6.8"A photo of the.classconcealedto blend in with its surroundings."0.697
6.96.7, 6.80.703
6.106.1 - 6.9 i.e., CamoPrompts0.704
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.45, + 0.566, + 0.52 + ], + "angle": 0, + "content": "Table 5: Efficiency comparison with other methods. \"Trainable Param.\" and \"Total Param.\" stand for the number of trainable and total parameters. In OVSeg [33], the CLIP [47] model is fine-tuned by the authors, resulting in more trainable parameters." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.522, + 0.558, + 0.63 + ], + "angle": 0, + "content": "
ModelTrainable Param.Total Param.FLOPs
SimSeg21 [60]61M (28.91%)211M1.9T
OVSeg22 [33]531M (100.00%)531M8.0T
ODISE23 [58]28M (1.80%)1522M5.5T
SAN23 [59]9M (2.06%)437M0.4T
CAT-Seg23 [10]104M (21.22%)490M0.3T
FC-CLIP23 [64]20M (5.38%)372M0.8T
Ours7M (1.95%)359M0.2T
" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.44, + 0.776, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.583, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Fig. 9: \\(\\alpha\\) corresponding to different heads in Fig. 6b from different decoding layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.84 + ], + "angle": 0, + "content": "of the task and data, we propose a strong single-stage baseline OVCoser with the advanced pre-trained vision-language model. Specifically, the well-designed prompt templates are introduced to reinforce the task-relevant semantic context. We introduce additional multi-source information including class semantic cues, depth spatial structure, object edge details, and top-down iterative guidance from the output space. With the help of these components, OVCoser can perceive and segment camouflaged objects in complex environments. Extensive experiments demonstrate the effectiveness of the proposed method and its superior performance compared with the existing state-of-the-art OVSIS algorithms on OVCamo." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.34 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62276046 and by Dalian Science and Technology Innovation Foundation under Grant 2023JJ12GX015. We sincerely thank these authors of the public datasets [2,8,14,61,78] used in this work: Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao for COD10K [14], Yunfei Zheng, Xiongwei Zhang, Feng Wang, Tieyong Cao, Meng Sun, and Xiaobing Wang for CPD1K [78], Jinyu Yang for PlantCamo [61], Pia Bideau and Erik Learned-Miller for CAD [2], and Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge for MoCA-Mask [8]. It is their efforts that underpin this new exploration of camouflaged object segmentation." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.361, + 0.323, + 0.377 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.785, + 0.445 + ], + "angle": 0, + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.785, + 0.486 + ], + "angle": 0, + "content": "2. Bideau, P., Learned-Miller, E.: It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In: Proceedings of European Conference on Computer Vision (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.485, + 0.785, + 0.54 + ], + "angle": 0, + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.785, + 0.58 + ], + "angle": 0, + "content": "4. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.622 + ], + "angle": 0, + "content": "5. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: Proceedings of European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.663 + ], + "angle": 0, + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.703 + ], + "angle": 0, + "content": "7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. International Conference on Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.744 + ], + "angle": 0, + "content": "8. Cheng, X., Xiong, H., Fan, D.P., Zhong, Y., Harandi, M., Drummond, T., Ge, Z.: Implicit motion handling for video camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "9. Cherti, M., Beaumont, R., Wightman, R., Wortzman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "0. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.W.: Catseg: Cost aggregation for open-vocabulary semantic segmentation. arXiv preprint (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.389, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "11. Csurka, G., Volpi, R., Chidlovskii, B.: Semantic image segmentation: Two decades of research. Foundations and Trends in Computer Graphics and Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "12. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.247 + ], + "angle": 0, + "content": "13. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.248, + 0.785, + 0.289 + ], + "angle": 0, + "content": "14. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.29, + 0.785, + 0.318 + ], + "angle": 0, + "content": "15. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.319, + 0.785, + 0.36 + ], + "angle": 0, + "content": "16. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.403, + 0.785, + 0.445 + ], + "angle": 0, + "content": "18. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.446, + 0.785, + 0.473 + ], + "angle": 0, + "content": "19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. IEEE Transactions on Pattern Analysis and Machine Intelligence (Feb 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.474, + 0.785, + 0.516 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "21. Ji, W., Li, J., Bian, C., Zhou, Z., Zhao, J., Yuille, A.L., Cheng, L.: Multispectral video semantic segmentation: A benchmark dataset and baseline. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.559, + 0.785, + 0.6 + ], + "angle": 0, + "content": "22. Ji, W., Li, J., Zhang, M., Piao, Y., Lu, H.: Accurate rgb-d salient object detection via collaborative learning. In: Proceedings of European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "23. Ji, W., Yan, G., Li, J., Piao, Y., Yao, S., Zhang, M., Cheng, L., Lu, H.: DMra: Depth-induced multi-scale recurrent attention network for rgb-d saliency detection. IEEE Transactions on Image Processing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "24. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "25. Katsuki, F., Constantinidis, C.: Bottom-up and top-down attention: different processes and overlapping neural systems. The Neuroscientist (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "26. Kim, S.Y., Zhang, J., Niklaus, S., Fan, Y., Chen, S., Lin, Z., Kim, M.: Layered depth refinement with mask guidance. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "27. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Li, A., Zhang, J., Lyu, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Li, G., Duan, N., Fang, Y., Gong, M., Jiang, D.: Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In: AAAI Conference on Artificial Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "30. Li, J., Ji, W., Wang, S., Li, W., Cheng, L.: Dvsod: RGB-d video salient object detection. In: International Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.234, + 0.786, + 0.274 + ], + "angle": 0, + "content": "31. Li, J., Ji, W., Zhang, M., Piao, Y., Lu, H., Cheng, L.: Delving into calibrated depth for accurate rgb-d salient object detection. International Journal of Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Proceedings of European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.318, + 0.787, + 0.373 + ], + "angle": 0, + "content": "33. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "34. Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., van der Laak, J.A., van Ginneken, B., Sánchez, C.I.: A survey on deep learning in medical image analysis. Medical Image Analysis (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "35. Liu, L., Wang, R., Xie, C., Yang, P., Wang, F., Sudirman, S., Liu, W.: Pestnet: An end-to-end deep learning approach for large-scale multi-class pest detection and classification. IEEE Access (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "36. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. Proceedings of the IEEE International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.502, + 0.787, + 0.529 + ], + "angle": 0, + "content": "37. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.531, + 0.787, + 0.572 + ], + "angle": 0, + "content": "38. Lyu, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.573, + 0.787, + 0.614 + ], + "angle": 0, + "content": "39. Mithun, N.C., Panda, R., Papalexakis, E.E., Roy-Chowdhury, A.K.: Webly supervised joint embedding for cross-modal image-text retrieval. In: Proceedings of the ACM International Conference on Multimedia (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.615, + 0.787, + 0.671 + ], + "angle": 0, + "content": "40. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.672, + 0.787, + 0.712 + ], + "angle": 0, + "content": "41. Neuhold, G., Ollmann, T., Bulo, S.R., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: Proceedings of the IEEE International Conference on Computer Vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "42. Pang, Y., Zhang, L., Zhao, X., Lu, H.: Hierarchical dynamic filtering network for rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "43. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "44. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "45. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "46. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE Transactions on Image Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.285 + ], + "angle": 0, + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.312 + ], + "angle": 0, + "content": "48. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.787, + 0.338 + ], + "angle": 0, + "content": "49. Rizzo, M., Marcuzzo, M., Zangari, A., Gasparetto, A., Albarelli, A.: Fruit ripeness classification: A survey (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "51. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database (2017), http://kgwisc.aei.polsl.pl/index.php/pl/dataset/63-animal-camouflage-analysis" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.461 + ], + "angle": 0, + "content": "52. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.488 + ], + "angle": 0, + "content": "53. Sun, Y., Wang, S., Chen, C., Xiang, T.Z.: Boundary-guided camouflaged object detection. In: International Joint Conference on Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "54. Thisanke, H., Deshan, C., Chamith, K., Seneviratne, S., Vidanaarachchi, R., Herath, D.: Semantic segmentation using vision transformers: A survey. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.57 + ], + "angle": 0, + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, u., Polosukhin, I.: Attention is all you need. In: International Conference on Neural Information Processing Systems (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.57, + 0.787, + 0.611 + ], + "angle": 0, + "content": "56. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. arXiv preprint (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.611, + 0.787, + 0.637 + ], + "angle": 0, + "content": "57. Xiang, M., Zhang, J., Lv, Y., Li, A., Zhong, Y., Dai, Y.: Exploring depth contribution for camouflaged object detection (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "59. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.76 + ], + "angle": 0, + "content": "60. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: Proceedings of European Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.767, + 0.772 + ], + "angle": 0, + "content": "61. Yang, J.: Plantcamo dataset (2023), https://github.com/yjybuaa/PlantCamo" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.8 + ], + "angle": 0, + "content": "62. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "63. Yu, Q., Zhao, X., Pang, Y., Zhang, L., Lu, H.: Multi-view aggregation network for dichotomous image segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "OVCOS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "64. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In: International Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "65. Zabari, N., Hoshen, Y.: Open-vocabulary semantic segmentation using test-time distillation. In: European Conference on Computer Vision Workshops (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.258 + ], + "angle": 0, + "content": "66. Zhang, M., Ji, W., Piao, Y., Li, J., Zhang, Y., Xu, S., Lu, H.: Lfnet: Light field fusion network for salient object detection. IEEE Transactions on Image Processing (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.3 + ], + "angle": 0, + "content": "67. Zhang, M., Liu, J., Wang, Y., Piao, Y., Yao, S., Ji, W., Li, J., Lu, H., Luo, Z.: Dynamic context-sensitive filtering network for video salient object detection. In: Proceedings of the IEEE International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "68. Zhang, M., Yao, S., Hu, B., Piao, Y., Ji, W.: C2dfnet: Criss-cross dynamic filter network for rgb-d salient object detection. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "69. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. International Conference on Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "70. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: Proceedings of the IEEE International Conference on Computer Vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.452 + ], + "angle": 0, + "content": "71. Zhao, X., Chang, S., Pang, Y., Yang, J., Zhang, L., Lu, H.: Multi-source fusion and automatic predictor selection for zero-shot video object segmentation. International Journal of Computer Vision (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.494 + ], + "angle": 0, + "content": "72. Zhao, X., Pang, Y., Ji, W., Sheng, B., Zuo, J., Zhang, L., Lu, H.: Spider: A unified framework for context-dependent concept understanding. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "73. Zhao, X., Pang, Y., Zhang, L., Lu, H.: Joint learning of salient object detection, depth estimation and contour extraction. IEEE Transactions on Image Processing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.564 + ], + "angle": 0, + "content": "74. Zhao, X., Pang, Y., Zhang, L., Lu, H., Ruan, X.: Self-supervised pretraining for rgb-d salient object detection. In: AAAI Conference on Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "75. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Suppress and balance: A simple gated network for salient object detection. In: Proceedings of European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "76. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Towards diverse binary segmentation via a simple yet general gated network. International Journal of Computer Vision (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "77. Zhao, X., Zhang, L., Pang, Y., Lu, H., Zhang, L.: A single stream network for robust and real-time rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "78. Zheng, Y., Zhang, X., Wang, F., Cao, T., Sun, M., Wang, X.: Detection of people with camouflage pattern via dense deconvolution network. IEEE Signal Processing Letters (Jan 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "79. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "80. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Proceedings of European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "81. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. ArXiv abs/2311.03352 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "Y. Pang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "82. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "83. Zhu, C., Chen, L.: A survey on open-vocabulary detection and segmentation: Past, present, and future. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "84. Zhu, F., Zhu, Y., Lee, V., Liang, X., Chang, X.: Deep learning for embodied vision navigation: A survey. arXiv preprint (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.231 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_origin.pdf b/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3c972eec7302c3573a014f5457c8cedefec1bdb9 --- /dev/null +++ b/2024/Open-Vocabulary Camouflaged Object Segmentation/3f202b98-48c3-4934-b87c-d93d6216aebc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34439b5be28a78b873e934c73cec4dc5d6c8e328972f447ae7985c8377e346dc +size 2829514 diff --git a/2024/Open-Vocabulary Camouflaged Object Segmentation/full.md b/2024/Open-Vocabulary Camouflaged Object Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e3918c13c4c10c6379c236dd7e992d98f4a60501 --- /dev/null +++ b/2024/Open-Vocabulary Camouflaged Object Segmentation/full.md @@ -0,0 +1,282 @@ +# Open-Vocabulary Camouflaged Object Segmentation + +Youwei Pang $^{1,2\star}$ , Xiaqi Zhao $^{1,2\star}$ , Jiaming Zuo $^{2}$ , Lihe Zhang $^{1\star\star}$ , and Huchuan Lu $^{1}$ + +$^{1}$ Dalian University of Technology + $^{2}$ X3000 Inspection Co., Ltd +{lartpang, zxq} $@$ mail.dlut.edu.cn +klaus@3000gy.com, {zhanglihe, lhchuan} $@$ dlut.edu.cn + +Abstract. Recently, the emergence of the large-scale vision-language model (VLM), such as CLIP, has opened the way towards open-world object perception. Many works have explored the utilization of pre-trained VLM for the challenging open-vocabulary dense prediction task that requires perceiving diverse objects with novel classes at inference time. Existing methods construct experiments based on the public datasets of related tasks, which are not tailored for open vocabulary and rarely involve imperceptible objects camouflaged in complex scenes due to data collection bias and annotation costs. To fill in the gaps, we introduce a new task, open-vocabulary camouflaged object segmentation (OVCOS), and construct a large-scale complex scene dataset (OVCampo) containing 11,483 hand-selected images with fine annotations and corresponding object classes. Further, we build a strong single-stage open-vocabulary camouflaged object segmentation transformer baseline OVCoser attached to the parameter-fixed CLIP with iterative semantic guidance and structure enhancement. By integrating the guidance of class semantic knowledge and the supplement of visual structure cues from the edge and depth information, the proposed method can efficiently capture camouflaged objects. Moreover, this effective framework also surpasses previous state-of-the-arts of open-vocabulary semantic image segmentation by a large margin on our OVCampo dataset. With the proposed dataset and baseline, we hope that this new task with more practical value can further expand the research on open-vocabulary dense prediction tasks. Our code and data can be found in the link. + +Keywords: Open Vocabulary $\cdot$ Camouflaged Object Segmentation + +# 1 Introduction + +The ability to identify and reason about object regions in a visual scene is essential for a wide range of human activities. As a complex and fundamental task in + +computer vision, detecting and segmenting objects with diverse appearances in complex scenes is also an important challenge, which is crucial for applications across vision and robotics fields, including autonomous driving [3], medical image analysis [34], and intelligent robotics [84], to name a few. In the past few years, numerous typical methods [6, 7, 19, 69] have emerged with the help of a mass of labeled data, which have greatly promoted the development of related fields such as semantic image segmentation (SIS) [11,54]. However, existing works have mainly focused on predefined closed-set scenarios, where all semantic concepts are seen during both the inference and training phases. Such a scenario setting oversimplifies the real-world complexity. To this end, many explorations have been contributed to open vocabulary settings [56, 83]. In existing work, open vocabulary learning utilizes visual-related language data (e.g., object class text or descriptions) to align images and language features, enabling models to perceive novel classes without extensive labeled data [56]. Recently, large-scale pretrained visual language models (VLMs) such as CLIP [9,47] have been gaining attention. Image-text matching-based learning mechanism provides them with the ability to align textual and visual signals well, and many works demonstrate their potential for open vocabulary tasks. Besides, data plays an important role in open-vocabulary tasks. Existing open-vocabulary semantic image segmentation (OVSIS) tasks rely on related public datasets [4,12,40,41,79], while they are not designed for the open-vocabulary setting, and there is high semantic similarity between their class definitions as revealed in [59]. Due to the data collection bias and annotation cost constraints, the existing open-vocabulary benchmarks lack special attention to finely perceive the objects of interest in concealed scenes. And the widely used VLMs are pre-trained on image-text pairs with inherent object concept bias, thus, their ability to segment objects in complex scenes remains to be verified. + +In this paper, we introduce a new open-vocabulary segmentation task OVCOS dedicated to analyzing camouflaged object perception in diverse natural scenes. And a large-scale data benchmark, named OVCamo, is carefully constructed. Besides, we also design a strong baseline OVCoser for the proposed OVCOS, based on the VLM-driven single-stage paradigm. The camouflage arises from several sources, including similar patterns to the environment (e.g., color and texture) and imperceptible attributes (e.g., small size and heavy occlusion) as statistically illustrated in Fig. 3. + +![](images/7051fe0903c93b57e760b3d3fc06af83e0bdb08e8028394555ff2fc233ebb1ef.jpg) +Fig.1: The perception and recognition of camouflaged objects require the collaboration of information from multiple sources such as appearance clues, spatial structure, contour details, and object semantics. + +![](images/15933d076926daf2632908164e92c8e2ffdefb4c006b632973dc4b56a762a330.jpg) +Fig. 2: Class distribution in the proposed OVCamo dataset and some visual examples. + +Considering the imperceptible appearance of camouflaged objects, accurate recognition and capture actually depend more on the cooperation of multi-source knowledge. As shown in Fig. 1, in addition to visual appearance cues, we introduce the depth for the spatial structure of the scene, the edge for the regional changes about objects, and the text for the context-aware class semantics. Considering the cooperative relationship between class recognition and object perception, the iterative learning strategy is introduced to feed back the optimized semantic relationship, resulting in more accurate object semantic guidance. This top-down conceptual reinforcement can further optimize open-vocabulary segmentation performance. With the help of the iterative multi-source information joint learning strategy, our method OVCoser shows good performance in the proposed OVCOS task. + +In summary, our contributions are three-fold as follows: + +- New Challenge. In view of the limitations of the existing OVSIS, we introduce a more challenging OVCOS task for open-vocabulary segmentation of camouflaged objects. +- New Benchmark. A new large-scale benchmark OVCamo with diverse samples carefully collected from existing publicly available data is proposed to better evaluate and analyze the generalization of algorithms on the proposed task. +- **Strong Baseline.** We build a robust single-stage baseline based on CLIP, in which the proposed iterative semantic guidance and structure enhancement are embedded. Under the joint optimization of multi-source information, our approach OVCoser outperforms existing OVSIS algorithms on the new benchmark. + +# 2 Related Works + +Vision-Language Pre-training. The core goal of visual-language pre-training is to learn generic representations of vision and language, and connect visual and + +language concepts. Early approaches [5, 29, 32, 52] are based on some relatively small and clean public datasets, which limits their achievable performance, and their fine-tuning on specific downstream tasks hinders application flexibility. In [39], image-text pairs collected from the Internet bring clear performance improvement for the retrieval task. This also indirectly encourages subsequent further exploration, such as CLIP [9,47]. It benefits from the larger scale of noisy data from web pages covering diverse and rich concepts and exhibits impressive open-vocabulary capabilities. This work introduces CLIP as the bedrock of open-vocabulary capability and borrows its strong image-text matching capability to build an effective baseline for the OVCOS task. + +Open Vocabulary Semantic Image Segmentation (OVSIS). Although a variety of pipelines have emerged as summarized in [83], from an overall perspective, their efforts are similar, namely, how to align class name/description semantic embedding with visual features to anchor relevant object cues in the representation space. The early pioneer work [70] attempts to connect word concepts and semantic relations, and encodes word concept hierarchy to parse images. Due to the leading performance of the VLM on the image-text joint modeling, it has been gradually applied in the OVSIS field. In terms of structure, existing schemes can be roughly categorized into two types: two-stage [33,58,60,65] and single-stage [10,64,80]. In [65], a rough segmentation map for each class is created from a VLM and then refined by the test-time augmentation. These rough maps are utilized as pseudo-labels for subsequent fine segmentation by stochastic pixel sampling. SimSeg [60] adopts a cascaded design including class-agnostic proposal generation by MaskFormer [7] and class assignment by CLIP [47]. Furthermore, OVSeg [33] fine-tunes CLIP on the noisy but diverse data to improve its generalization to masked images. In lieu of using existing SIS models, a text-to-image diffusion model is introduced to generate mask features with implicit image captions in [58]. The single-stage design is more flexible and simpler. MaskCLIP [80] directly modifies CLIP for semantic segmentation without training, while SAN [59] achieves better performance with the help of adapters. CAT-Seg [10] highlights the importance of the cost aggregation between image and text embeddings for the OVSIS decoding. Recent FC-CLIP [64] investigates the hierarchical CLIP image encoder. Although these methods show impressive OVSIS performance, they still follow the generalized segmentation paradigm [6,7] in SIS, ignoring valuable auxiliary cues for object perception. This also causes them to struggle with objects camouflaged in complex scenes. Unlike them, our method, which is tailored for OVCOS, can effectively tap into the camouflaged object by integrating task-specific multi-source knowledge from visual appearance, spatial structure, object contour, and class semantics. + +Camouflaged Scene Understanding (CSU). This is a research hotspot in the computer vision community, aiming to perceive objects with camouflage [15]. Different from traditional object perception like salient object detection [22, 23, 30, 31, 42, 45, 46, 63, 66–68, 72–77] and semantic object segmentation [21, 71], CSU is obviously a more challenging problem. It can be applied in some specific fields, such as medical analysis [13, 16] and agricultural man + +agement [35, 49]. This topic is currently defined as the class-agnostic form, focusing on the area of camouflaged objects in the visual scene. The available work [8, 13, 14, 24, 28, 43, 44, 62] to date has demonstrated promising performance on existing data benchmarks [2, 8, 14, 27, 38, 51]. Unlike previous settings, the proposed OVCOS task requires further perception of object classes. Admittedly, the publicly available data provides critical support for this new task. It helps us take the first step. + +# 3 OVCamo Dataset + +This work focuses on open-vocabulary segmentation in the camouflaged scene. It enriches the connotation of OVSIS and provides a more challenging benchmark. Note that our aim is not to claim novelty through class annotations alone, but rather to focus on how such annotations can pave the way for new tasks and methods. Our dataset provides additional features such as clean data, less ambiguity, and sensible class hierarchies, which provide the necessary data for solving the new and complex task. It not only adds complexity and challenge, but also promotes innovation and advances the field through the need for advanced methods. We believe that this dataset can lead to new opportunities in areas such as wildlife conservation and intelligence sensing. + +Image Collection. Our data is collected from existing CSU datasets that have finely annotated segmentation maps. Specifically, the OV-Camo integrates 11,483 hand-selected images covering 75 object classes reconstructed from several public datasets [2, 8, 14, 61, 78]. The distribution of the number of samples in different classes is shown in Fig. 2. Meanwhile, we consider attributes of objects when selecting images, such as object concentration, average color ratio, object-image area ratio, number + +Table 1: Attributes involved in the dataset analysis. + +
AttributeDescription
Object ConcentrationObject pixel concentration, which calculates the area ratio between the object region and its minimum rotatable bounding box.
Average Color RatioColor ratio of the object to the background, which calculates the average of the three color channels in their respective regions.
Object-Image Area RatioArea ratio of the object relative to the image.
Number of Object PartsNumber of separate areas in the image that belong to the object.
Normalized CentroidObject centroid coordinates, which are normalized using the image shape.
+ +of object parts, and normalized centroid. Tab. 1 gives their definitions. And Fig. 3 visualizes the attribute distribution of the proposed dataset. The camouflaged objects of interest usually have complex shape Fig. 3a, high similarity to the background Fig. 3b, and small size Fig. 3c. And the image often contains multiple camouflaged objects or sub-regions with a central bias as shown in Figs. 3d and 3e. + +Annotation. The original annotation cannot be used directly in the open-vocabulary setting, due to the following semantic ambiguities caused by different annotation standards, i.e., 1) Broad concepts, such as "fish" and "bug"; 2) Vague definitions, such as "small fish" and "black cat"; 3) Inconsistent granularity, such as the coexistence of "orchid mantis" and "mantis"; 4) Non- + +![](images/c782ee5c2f964548f52a19423457a3aaf7e420ac8f8c0afc6769d410db394d68.jpg) +(a) Object Concentration. + +![](images/c97fb15807e94d5f10c05f70741d4a3bf8e6fd0b429f63f7f7ec2b508af86437.jpg) +(b) Average Color Ratio. + +![](images/62e7170139bc7d1baf091e121a780fb85c2bedfba0578a1cc3cae7741dea61a0.jpg) +(c) Object-Image Area Ratio. +Fig. 3: Attribute visualization of objects in our OVCampo dataset, including object concentration, object-background color ratio, object-image area ratio, number of object parts, and normalized centroid. Please refer to Tab. 1 for details about attributes. + +![](images/d70c181511b1804039f21147df1f7722d96498ca630affdf21e8fa183371c299.jpg) +(d) Number of Object Parts. + +![](images/a5347468efba2f7697113810c0c2056e4576654a558a32bc3c64a69445569da9.jpg) +(e) Normalized Centroid. + +entity concepts, such as "other". These issues can lead to unreasonable and unreliable results for the open-vocabulary prediction as discussed in [81]. To this end, we relabel the classes of all camouflaged objects and take the generality of the concept as the criterion for class definition, which also ensures lower semantic similarity. The similarity analysis and rough hierarchy of the class set $\mathcal{C}$ can be found in the supplementary material. + +Data Division. To objectively evaluate the open-vocabulary segmentation algorithm on unseen classes, we assign as many classes as possible to the test set and control the sample ratio of the training set to the testing set to be 7:3. Specifically, 14 classes in the dataset are taken as the training set and all the remaining 61 classes are used for testing. Such a setting follows the existing practice in OVSIS where the number of seen classes (e.g., 171 [4]) is usually fewer than unseen classes (e.g., 847 [79]) which is also closer to the real-world setting. Such a setup can ensure the quantity of training samples while reinforcing the complexity of the test set. Finally, the overall ratio of samples is 7713:3770. + +# 4 Methodology + +In this section, we introduce a strong baseline OVCoser. The overall framework is first described, followed by the details of the key components. + +Overall Architecture. We follow the common encoder-decoder paradigm and the pipeline is shown in Fig. 4. Specifically, we first leverage the textual encoder $\mathbf{E}_{\mathbf{t}}$ of the frozen CLIP to extract semantic embedding $f_{t}$ from the class label set $\mathcal{C}$ , and the visual encoder $\mathbf{E}_{\mathbf{v}}$ to extract multi-scale image features $\{f^{i}\}_{i = 1}^{5}$ . Both of them are fed into the decoder as the information bedrock of object segmentation, as shown in Fig. 5. Structural cues such as depth and edge are also introduced to assist the iterative refinement process. Finally, the class-related segmentation $P_{s}$ which is the segmentation logits $M_{s}$ after the sigmoid function processing, is used to remove the interference from the background in the high-level image + +![](images/98d7bc5ca58c4f9b33fea5e9602818bdff4276a6ebd2c44fadc098d84aef0019.jpg) +Fig. 4: Overview of our single-stage open-vocabulary camouflaged object segmentation framework, OVCoser. It is based on the frozen CLIP model which includes feature encoder $\mathbf{E}_v$ and embedding layer $\mathbf{T}_v$ for visual appearance cues and textual encoder $\mathbf{E}_t$ for class semantic information. The well-constructed prompt template set Camo-Prompts $\mathcal{P}$ enables textual embedding $f_t$ to be more appropriate to OVCOS. In the iterative refinement decoder (Fig. 5), a more accurate class-related segmentation $P_s$ , can be obtained with the assistance of the semantic guidance $\mathbf{C}$ (Fig. 6a) from the class semantic and the structure enhancement (Fig. 6b) from the auxiliary depth and edge supervisions, i.e., $\mathbf{D}$ and $\mathbf{E}$ . After being masked average pooled (MAP) by $P_s$ , the high-level feature $f^5$ is used to assign a class $P_c$ to the object from the class set $\mathcal{C}$ . Note that $\mathbf{D}$ and $\mathbf{E}$ are only used as auxiliary supervision during training and are not required during inference. + +feature and guide the generation of object-oriented visual representation $f_{v}$ . And the class label $P_{c}$ is determined by the similarity matching between $f_{v}$ and $f_{t}$ . More details about the proposed architecture are listed in the supplementary material. + +Semantic Guidance (SG). The class label definition usually is independent of the image scene. It is very important to fully utilize the class prior for object recognition in complex scenes. In the proposed decoder, normalized textual embedding is introduced into each stage to highlight semantically relevant cues. As shown in Fig. 6a, we design a semantic guidance component SG to inject concept cues into the self-enhancement of the image feature. Specifically, in our semantic guidance attention (SGA), the normalized image feature $X$ is linearly mapped to $Q$ , $K$ , and $V$ , while the textual embedding $f_{t}$ is transformed to the class guidance vector $G_{t}$ . The similarity between $Q$ and $G_{t}$ reflects the activation of different classes in spatial locations. In Agg of Fig. 6a, the base weight $W_{b}$ for the spatial guidance is obtained after highlighting the most relevant class information by softmax operation. And then $V$ is modulated and fed into MHSA, i.e., multi-head self-attention [55]. + +Structure Enhancement (SE). Existing methods demonstrate that low-level structural information, such as the edge [53] and the depth [57], plays an important role in the design of neural networks. + +![](images/14a8ad6ee324e9607f42bbf7fea10bb3c82c438b707e7854408c3893b4ced91a.jpg) +Fig. 5: Proposed pipeline of the iterative refinement decoder with semantic guidance (SG) and structure enhancement (SE) components denoted as "SG- $\star$ " and "SE- $\star$ ". $t$ and $T$ denote the current step and total number of iterations, respectively. + +![](images/688eea1ab6c8901f7cc64598553b5ac3d07abb0b4da8eb5ddc8f59b7749019bb.jpg) +(a) SG + +![](images/cade2dedc198785643b56f9429a0e2194f2b5d60d85332e0c2aa122919bb6c40.jpg) +(b) SE +Fig. 6: Proposed semantic guidance (SG) and structure enhancement (SE) components in which semantic guidance attention (SGA) and structure enhancement attention (SEA) are embedded. The dashed blocks represent the feedback paths that come into play during iteration. + +tant role in CSU, which is closely related to the mechanism of the human visual system. So the SE component attached to the low-level SG is proposed to integrate the edge-aware and depth-aware cues and improve the structural details. Specifically, the output of the SG is fed into two separate branches containing the convolutional stem and head for the edge and depth estimation as shown in Fig. 5. The edge and depth logits maps, i.e., $M_e^i$ and $M_d^i$ , from the head in the branch of the layer $i \in \{1,2,3\}$ are directly supervised. And the outputs $f_e^i$ and $f_d^i$ of the stem are fed into the SE. In the structure enhancement attention (SEA), they independently update the normalized visual features $X$ using MHSA [55], and the corresponding outputs are combined with the learnable weight $\alpha$ as in Fig. 6b. + +Iterative Refinement. In the SG component, the aggregation process between image features and class semantics is not aligned and requires data-driven optimization. Considering the aligned embedding space of the pre-trained CLIP, we introduce the correlation matrix $M_{cor}$ between the visual and textual embeddings into the SG as shown in Fig. 6a. Meanwhile, due to the emphasis of the decoder output on the object region, the object-aware representation $f_{obj}$ is also inputted, which comes from the image features pooled by the coarse + +segmentation prediction in the last iteration. By combining the two, we obtain task-oriented object cues, which is actually inspired by the top-down attention mechanism in the human cognitive system [1,25]. The spatial activation map $W_{r}$ of such object cues over image features is used to re-modulate $W_{b}$ . Besides, the SE in the iteration also helps the model to further optimize the texture details. To benefit as much as possible from the assistance from structure enhancement while avoiding over-computation, we set the iteration entry to the third decoding layer as shown in Fig. 5. + +CamoPrompts. As mentioned in [47], prompt engineering and ensembling are important for the transfer performance of CLIP on downstream tasks, and the prompt template should be more relevant to the data type. Because additional task-related cues are generally able to impose the necessary contextual constraint to the flexible CLIP. Hence, instead of common practices [18, 47, 82], we design a simpler yet more effective template set CamoPrompts tailored for OVCOS to decorate the class name, and average their textual embeddings as the final semantic embedding for each class. Its full form is depicted in Tab. 4, while it also achieves better classification performance in comparison with other forms. + +Supervision. In each iteration, in addition to semantic object segmentation, we also need to perform depth estimation and edge estimation as auxiliary tasks. For the segmentation prediction, we follow the commonly used weighted segmentation loss function $l_{s}^{t} = l_{s}(P_{s}^{t},G_{s})$ [53,62]. For the edge estimation, considering the imbalance problem of positive and negative samples, we introduce the dice loss function as $l_{e}^{i,t} = l_{e}(P_{e}^{i,t},G_{e})$ . The summation of L1 and SSIM losses, i.e., $l_{d}^{i,t} = l_{d}(P_{d}^{i,t},G_{d})$ , is used for the depth estimation. The total loss $L$ of our method can be formulated as follows: $L = \sum_{t=1}^{T}\left[l_{s}^{t} + \sum_{i=1}^{3}\left(l_{e}^{i,t} + l_{d}^{i,t}\right)\right]$ , where $t$ and $i$ are used to index iterations and layers, respectively. And the total number $T$ of iterations is set to 2 as mentioned in Sec. 5.3. + +# 5 Experiments + +# 5.1 Implementation Details + +Dataset Settings. As mentioned in Sec. 3, the proposed dataset is divided according to classes into two disjoint subsets, i.e., $\mathcal{C}_{seen}$ and $\mathcal{C}_{unseen}$ . The former contains 14 classes for training and the latter contains the remaining 61 classes for testing. The proposed dataset itself is provided with only images and masks. We use the typical monocular depth estimation methods [26,48] to obtain the depth map $G_{d}$ for training, while the edge map $G_{e}$ is generated by dilating and eroding operations. We introduce depth and edge maps only in the training phase and they are not required during inference, which also avoids information leakage during testing. So, our depth and edge maps are labor-free and training-only, and directly generated by existing algorithms. Such a design can help our model get rid of the dependence on them in real-world applications. Our structure enhancement based on depth and edge information belongs to a deeper exploration + +![](images/90754cec10618a0faa07e5dd09531883e702474a4cbd1e7fd5c5b512665c2d20.jpg) +Fig. 7: Visual results on OVCamo. Existing methods are either disrupted by chaotic backgrounds, imperceptible appearances, blurry details, or severe occlusion, while our algorithm can effectively capture and remain well-exposed object details. And three different colors are used to represent human annotations, correct predictions, and incorrect predictions. + +of this task. These generated depth maps will be made publicly available with the dataset. + +Training Details. Following previous settings in [59, 60, 64], the pre-trained CLIP is frozen during training, and the remaining parameters are learnable and are randomly initialized. The AdamW [37] optimizer with the learning rate of 3e-6, weight decay of 5e-4, batch size of 4, and training epoch of 30 is used to optimize model parameters. Some basic data augmentations including random flipping, rotating, and color jittering, are introduced to preprocess training data. The input data and the output logits are bilinearly interpolated to $384 \times 384$ during training and inference. + +Evaluation Protocol. To reasonably evaluate the performance of OVCOS, we modify the metrics in the original CSU task [8,13] to $\mathrm{cS}_m$ , $\mathrm{cF}_{\beta}^{\omega}$ , cMAE, $\mathrm{cF}_{\beta}$ , $\mathrm{cE}_m$ , and cIoU which follows the common settings in the OVSIS field [10,33,59, 60,64,83] and takes into account both classification and segmentation. + +# 5.2 System-level Comparison + +To show the complexity of the proposed OVCOS task and also to verify the effectiveness of the proposed method, we compare OVCoser with several recent state-of-the-art methods in OVSIS, including [10,33,58-60,64]. Since this is a new task, existing methods need to be re-evaluated to understand their generalization ability. Based on the public code and weights trained on COCO-Stuff [4] provided by the authors, we show the performance of these methods under three different testing schemes, including S.I) testing directly with their trained weights; S.II) + +Table 2: Comparison with recent state-of-the-art CLIP-based open-vocabulary semantic image segmentation methods with different training settings on the proposed OVCamo dataset. The best three results are highlighted in red, green and blue. + +
ModelVLMFeature BackboneText PromptcSm↑cFωβ↑cMAE↓cFβ↑cEm↑cloU↑
Test on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.1280.1050.8380.1120.1430.094
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.3410.3060.5840.3250.3840.273
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.4090.3390.5000.3410.4210.302
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.4140.3430.4890.3570.4560.319
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.4300.3440.4480.3660.4590.310
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.3740.3060.5390.3200.4090.285
Finetune on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0980.0710.8520.0810.1280.066
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.1640.1310.7630.1470.2080.123
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1820.1250.6910.2190.3090.189
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.3210.2160.5500.2360.3310.204
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1850.0940.7020.1100.1850.088
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.1240.0740.7980.0880.1620.072
Train on OVCamo.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0530.0490.9210.0560.0980.047
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.0240.0460.9540.0560.1300.046
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1870.1190.7000.2110.2980.167
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.2750.2020.6120.2200.3180.189
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1810.1060.7190.1230.1960.094
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.0800.0760.8720.0900.1910.072
OursCLIP-ConvNeXt-L [9]—CamoPrompts0.5790.4900.3360.5200.6160.443
+ +further fine-tuning based on trained weights before testing; and S.III) testing after re-training directly on our training set. + +Quantitative Evaluation. For the sake of fairness in comparisons, we report the performance of the "Large" versions for these methods, except for SimSeg [60] where the authors only provide the "Base" version. All results are summarized in Tab. 2 and our approach consistently outperforms these competitors. It is worth noting that existing methods perform better at S.I. This may be attributed to the training process on the larger-scale COCO-Stuff dataset, which provides a more general understanding of the concepts. However, direct fine-tuning (i.e., S.II) may destroy this knowledge and even cause oblivion to some extent, resulting in performance degradation. If we follow S.III to re-train them, the relatively small-scale training data may not be enough to train these models with more complex structures. At the same time, the existing methods also lack targeted optimization for the OVCOS task. These problems can lead to further deterioration of performance. However, our approach tailored for OVCOS achieves leading performance by the iterative refinement strategy of multi-source information, which comprehensively considers different characteristics of the task. + +Qualitative Evaluation. We also visualize the results of some recent methods on a variety of data in Fig. 7. It can be seen that the proposed method shows better performance and adaptability to diverse objects, including large objects (Col. 1-2), middle objects (Col. 3-5), small objects (Col. 6-9), multiple objects (Col. 8), complex shapes (Col. 3-5), blurred edges (Col. 1-5), severe occlusion (Col. 6), and background interference (Col. 2-6). + +Table 3: Ablation comparison of proposed components. $\Delta$ represents the average relative gain in performance of the corresponding model over the baseline for OVCOS. $\mathcal{P}$ : CamoPrompts. $C$ : Semantic guidance. $D$ : Depth estimation auxiliary task. $E$ : Edge estimation auxiliary task. $T$ : Number of iterations which is set to 2 by default due to the best performance. "lim $\sup_{P_s \to G_s}$ Perf." The ideal performance for our framework. + +
ModelcSm↑ cFβ↑ cMAE ↓ cFβ ↑ cEm ↑ cIoU ↑Δ
Comparison of the proposed modules.
Baseline0.5170.4080.3740.4510.5490.3590.0%
+P0.5430.4350.3460.4800.5810.3836.3%
+P,C0.5500.4530.3410.4910.5970.3979.1%
+P,C,D0.5650.4730.3360.5070.6060.42212.6%
+P,C,E0.5670.4810.3390.5110.6070.43213.5%
+P,C,D,E (i.e., T=1)0.5700.4880.3380.5180.6100.43614.5%
SE (Fig. 6b) → Addition Fusion0.5520.4570.3400.4970.5990.4029.9%
Comparison of the proposed iterative refinement.
T=10.5700.4880.3380.5180.6100.43614.5%
T=20.5790.4900.3360.5200.6160.44315.5%
w/o Mcor as in Fig. 6a0.5750.4870.3370.5150.6110.44114.8%
w/o obj as in Fig. 6a0.5710.4760.3390.5060.6080.43413.4%
T=30.5760.4840.3330.5140.6140.43714.8%
lim supPseg→GsegPerf.0.7030.7030.2970.7010.7010.70151.2%
+ +# 5.3 Analysis and Ablation Study + +Better Baseline Model. Interestingly, our baseline model in Tab. 3 outperforms existing methods in Tab. 2. We attribute this to the fact that the overly complex feature pipeline of the conventional OVSIS architecture is not well adapted to this task. And our framework, from hyperparameter setting to structural evolution, is explored for the proposed task and data. Thus, we actually provide a more effective and robust baseline setup for OVCOS. + +Importance of Modules. To specifically analyze the effect of different components, we evaluate their performance in Tab. 3. As can be seen, the proposed modules all show positive gains. Both the semantic guidance in the class-aware decoding and the structure enhancement from auxiliary tasks of depth and edge estimation consistently boost the final performance. The ablation comparison also demonstrates that explicit guidance of the spatial and contour information is important for the detection of camouflaged objects and their positive gains are subject to the effective fusion design. We replace the proposed structure enhancement (SE) component with the vanilla addition fusion with a similar number of parameters, which results in detrimental effects. Besides, in Tab. 3, we also give the ideal results of our CLIP-driven framework where $G_{s}$ is treated as $P_{s}$ . Although our method exhibits leading performance compared to existing methods as shown in Tab. 2, there is still a long way to go to solve this problem. At the same time, the current ideal performance is still far from the limit, suggesting that future breakthroughs in this field may require more powerful paradigms as discussed in the supplementary material. + +Importance of Iterative Refinement. The top-down iterative refinement strategy significantly improves the OVCOS performance as shown in Tab. 3. When the number $T$ of iterations is 2, our algorithm obtains the best OVCOS performance. As $T$ increases, there is no further improvement in performance, so it is set to 2 by default. In addition, the object-aware representation $f_{obj}$ from the output space plays an important role, and introducing the correlation guidance $M_{cor}$ also has positive gains. + +# Importance of CamoPrompts. + +Our prompt template set CamoPrompts, which takes task attributes into account, shows better performance in Tab. 4. To further understand the influence of different templates on the semantic embedding, we calculate the Hausdorff distance between training and testing class labels in the embedding space, as shown in Fig. 8. The figure presents an interesting phenomenon that those templates with better classification performance tend to reduce the distance, which inspires further explorations for more effective prompt engineering. + +Importance Between Edge and Depth. In the proposed SE as shown in Fig. 6b, the interaction components for the edge and depth are combined by a coefficient vector $\alpha$ , which can also reflect the relative importance of + +the two kinds of information. In Fig. 9, we plot $\alpha$ from different decoding layers. It can be seen that the values are usually greater than 0.5, which indicates a preference for the edge information flow. + +Model Complexity. We compare the proposed method with the CLIP-based competitors [10, 33, 58-60, 64] in terms of the number of trainable and total parameters, and FLOPs. To be fair, we test all methods following the setting of their original inference settings. As can be seen from Tab. 5, our method has fewer trainable parameters ( $< 2\%$ ) and less computational complexity ( $0.2T$ ), which is superior to these competitors. + +# 6 Conclusion + +In this work, we propose a new challenging task, OVCOS, to explore open-vocabulary semantic image segmentation (OVSIS) for the camouflaged objects in more complex natural scenes, and carefully collect and construct a large-scale data benchmark OVCamo. Meanwhile, by considering the characteristics + +![](images/1b3d99441641a29ae4526167b1b4f06898f27b8d62d06c59dc36418b7ebeca55.jpg) +Fig.8: Illustration of classification accuracy $\mathcal{A}$ and Hausdorff distance $\mathcal{H}$ between the textual embeddings of training and testing class splits by using the plain CLIP on the OVCamo dataset. The approximate negative correlation between them may reveal that the templates that bring training and testing class embeddings closer usually also improve classification accuracy. Tab. 4 shows the details of these templates. + +Table 4: Classification accuracy $\mathcal{A}$ of the plain CLIP using different prompt templates on OVCamo, which is based on the masked average pooling with the ground truth mask. + +
IDPrompt TemplateA
0".class" w/o MAP based on ground truth0.538
Task-generic templates.
1".class"0.671
2.1"The.class"0.648
2.2"the.class"0.632
2.3"A.class"0.677
2.4"a.class"0.675
3.1"A photo of a.class"0.684
3.2"A photo of the.class"0.691
3.3"The photo of a.class"0.672
3.4"The photo of the.class"0.680
3.5"a photo of a.class"0.684
3.6"a photo of the.class"0.689
3.7"the photo of a.class"0.675
3.8"the photo of the.class"0.674
4templates from [47]0.686
5templates from [18]0.682
Task-related templates.
6.1"A photo of the camouflaged.class"0.681
6.2"A photo of the concealed.class"0.674
6.36.1, 6.20.685
6.4"A photo of the.classcamouflagedin the background."0.684
6.5"A photo of the.classconcealedin the background."0.694
6.66.4, 6.50.694
6.7"A photo of the.classcamouflagedto blend in with its surroundings."0.700
6.8"A photo of the.classconcealedto blend in with its surroundings."0.697
6.96.7, 6.80.703
6.106.1 - 6.9 i.e., CamoPrompts0.704
+ +Table 5: Efficiency comparison with other methods. "Trainable Param." and "Total Param." stand for the number of trainable and total parameters. In OVSeg [33], the CLIP [47] model is fine-tuned by the authors, resulting in more trainable parameters. + +
ModelTrainable Param.Total Param.FLOPs
SimSeg21 [60]61M (28.91%)211M1.9T
OVSeg22 [33]531M (100.00%)531M8.0T
ODISE23 [58]28M (1.80%)1522M5.5T
SAN23 [59]9M (2.06%)437M0.4T
CAT-Seg23 [10]104M (21.22%)490M0.3T
FC-CLIP23 [64]20M (5.38%)372M0.8T
Ours7M (1.95%)359M0.2T
+ +![](images/3c67441ce3f068c9b28af38ea7fa0dc43c7ead4639ceff3461a87dbc0b8e12d2.jpg) +Fig. 9: $\alpha$ corresponding to different heads in Fig. 6b from different decoding layers. + +of the task and data, we propose a strong single-stage baseline OVCoser with the advanced pre-trained vision-language model. Specifically, the well-designed prompt templates are introduced to reinforce the task-relevant semantic context. We introduce additional multi-source information including class semantic cues, depth spatial structure, object edge details, and top-down iterative guidance from the output space. With the help of these components, OVCoser can perceive and segment camouflaged objects in complex environments. Extensive experiments demonstrate the effectiveness of the proposed method and its superior performance compared with the existing state-of-the-art OVSIS algorithms on OVCamo. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China under Grant 62276046 and by Dalian Science and Technology Innovation Foundation under Grant 2023JJ12GX015. We sincerely thank these authors of the public datasets [2,8,14,61,78] used in this work: Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao for COD10K [14], Yunfei Zheng, Xiongwei Zhang, Feng Wang, Tieyong Cao, Meng Sun, and Xiaobing Wang for CPD1K [78], Jinyu Yang for PlantCamo [61], Pia Bideau and Erik Learned-Miller for CAD [2], and Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge for MoCA-Mask [8]. It is their efforts that underpin this new exploration of camouflaged object segmentation. + +# References + +1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018) +2. Bideau, P., Learned-Miller, E.: It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In: Proceedings of European Conference on Computer Vision (2016) +3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020) +4. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018) +5. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: Proceedings of European Conference on Computer Vision (2020) +6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. International Conference on Neural Information Processing Systems (2021) +8. Cheng, X., Xiong, H., Fan, D.P., Zhong, Y., Harandi, M., Drummond, T., Ge, Z.: Implicit motion handling for video camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +9. Cherti, M., Beaumont, R., Wightman, R., Wortzman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +0. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.W.: Catseg: Cost aggregation for open-vocabulary semantic segmentation. arXiv preprint (2023) + +11. Csurka, G., Volpi, R., Chidlovskii, B.: Semantic image segmentation: Two decades of research. Foundations and Trends in Computer Graphics and Vision (2022) +12. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (2010) +13. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +14. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020) +15. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence (2023) +16. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging (2020) +17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (2022) +18. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2021) +19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. IEEE Transactions on Pattern Analysis and Machine Intelligence (Feb 2020) +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2016) +21. Ji, W., Li, J., Bian, C., Zhou, Z., Zhao, J., Yuille, A.L., Cheng, L.: Multispectral video semantic segmentation: A benchmark dataset and baseline. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023) +22. Ji, W., Li, J., Zhang, M., Piao, Y., Lu, H.: Accurate rgb-d salient object detection via collaborative learning. In: Proceedings of European Conference on Computer Vision (2020) +23. Ji, W., Yan, G., Li, J., Piao, Y., Yao, S., Zhang, M., Cheng, L., Lu, H.: DMra: Depth-induced multi-scale recurrent attention network for rgb-d saliency detection. IEEE Transactions on Image Processing (2022) +24. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +25. Katsuki, F., Constantinidis, C.: Bottom-up and top-down attention: different processes and overlapping neural systems. The Neuroscientist (2014) +26. Kim, S.Y., Zhang, J., Niklaus, S., Fan, Y., Chen, S., Lin, Z., Kim, M.: Layered depth refinement with mask guidance. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +27. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding (2019) +28. Li, A., Zhang, J., Lyu, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021) + +29. Li, G., Duan, N., Fang, Y., Gong, M., Jiang, D.: Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In: AAAI Conference on Artificial Intelligence (2020) +30. Li, J., Ji, W., Wang, S., Li, W., Cheng, L.: Dvsod: RGB-d video salient object detection. In: International Conference on Neural Information Processing Systems (2023) +31. Li, J., Ji, W., Zhang, M., Piao, Y., Lu, H., Cheng, L.: Delving into calibrated depth for accurate rgb-d salient object detection. International Journal of Computer Vision (2023) +32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Proceedings of European Conference on Computer Vision (2020) +33. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +34. Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., van der Laak, J.A., van Ginneken, B., Sánchez, C.I.: A survey on deep learning in medical image analysis. Medical Image Analysis (2017) +35. Liu, L., Wang, R., Xie, C., Yang, P., Wang, F., Sudirman, S., Liu, W.: Pestnet: An end-to-end deep learning approach for large-scale multi-class pest detection and classification. IEEE Access (2019) +36. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. Proceedings of the IEEE International Conference on Computer Vision (2021) +37. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019) +38. Lyu, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021) +39. Mithun, N.C., Panda, R., Papalexakis, E.E., Roy-Chowdhury, A.K.: Webly supervised joint embedding for cross-modal image-text retrieval. In: Proceedings of the ACM International Conference on Multimedia (2018) +40. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2014) +41. Neuhold, G., Ollmann, T., Bulo, S.R., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: Proceedings of the IEEE International Conference on Computer Vision (2017) +42. Pang, Y., Zhang, L., Zhao, X., Lu, H.: Hierarchical dynamic filtering network for rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020) +43. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +44. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) + +45. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020) +46. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE Transactions on Image Processing (2023) +47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (2021) +48. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE International Conference on Computer Vision (2021) +49. Rizzo, M., Marcuzzo, M., Zangari, A., Gasparetto, A., Albarelli, A.: Fruit ripeness classification: A survey (2023) +50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022) +51. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database (2017), http://kgwisc.aei.polsl.pl/index.php/pl/dataset/63-animal-camouflage-analysis +52. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. In: International Conference on Learning Representations (2019) +53. Sun, Y., Wang, S., Chen, C., Xiang, T.Z.: Boundary-guided camouflaged object detection. In: International Joint Conference on Artificial Intelligence (2022) +54. Thisanke, H., Deshan, C., Chamith, K., Seneviratne, S., Vidanaarachchi, R., Herath, D.: Semantic segmentation using vision transformers: A survey. arXiv preprint (2023) +55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, u., Polosukhin, I.: Attention is all you need. In: International Conference on Neural Information Processing Systems (2017) +56. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. arXiv preprint (2024) +57. Xiang, M., Zhang, J., Lv, Y., Li, A., Zhong, Y., Dai, Y.: Exploring depth contribution for camouflaged object detection (2022) +58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023) +59. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023) +60. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: Proceedings of European Conference on Computer Vision (2021) +61. Yang, J.: Plantcamo dataset (2023), https://github.com/yjybuaa/PlantCamo +62. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection (2022) +63. Yu, Q., Zhao, X., Pang, Y., Zhang, L., Lu, H.: Multi-view aggregation network for dichotomous image segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024) + +64. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In: International Conference on Neural Information Processing Systems (2023) +65. Zabari, N., Hoshen, Y.: Open-vocabulary semantic segmentation using test-time distillation. In: European Conference on Computer Vision Workshops (2023) +66. Zhang, M., Ji, W., Piao, Y., Li, J., Zhang, Y., Xu, S., Lu, H.: Lfnet: Light field fusion network for salient object detection. IEEE Transactions on Image Processing (2020) +67. Zhang, M., Liu, J., Wang, Y., Piao, Y., Yao, S., Ji, W., Li, J., Lu, H., Luo, Z.: Dynamic context-sensitive filtering network for video salient object detection. In: Proceedings of the IEEE International Conference on Computer Vision (2021) +68. Zhang, M., Yao, S., Hu, B., Piao, Y., Ji, W.: C2dfnet: Criss-cross dynamic filter network for rgb-d salient object detection. IEEE Transactions on Multimedia (2023) +69. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. International Conference on Neural Information Processing Systems (2021) +70. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: Proceedings of the IEEE International Conference on Computer Vision (2017) +71. Zhao, X., Chang, S., Pang, Y., Yang, J., Zhang, L., Lu, H.: Multi-source fusion and automatic predictor selection for zero-shot video object segmentation. International Journal of Computer Vision (2024) +72. Zhao, X., Pang, Y., Ji, W., Sheng, B., Zuo, J., Zhang, L., Lu, H.: Spider: A unified framework for context-dependent concept understanding. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024) +73. Zhao, X., Pang, Y., Zhang, L., Lu, H.: Joint learning of salient object detection, depth estimation and contour extraction. IEEE Transactions on Image Processing (2022) +74. Zhao, X., Pang, Y., Zhang, L., Lu, H., Ruan, X.: Self-supervised pretraining for rgb-d salient object detection. In: AAAI Conference on Artificial Intelligence (2022) +75. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Suppress and balance: A simple gated network for salient object detection. In: Proceedings of European Conference on Computer Vision (2020) +76. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Towards diverse binary segmentation via a simple yet general gated network. International Journal of Computer Vision (2024) +77. Zhao, X., Zhang, L., Pang, Y., Lu, H., Zhang, L.: A single stream network for robust and real-time rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020) +78. Zheng, Y., Zhang, X., Wang, F., Cao, T., Sun, M., Wang, X.: Detection of people with camouflage pattern via dense deconvolution network. IEEE Signal Processing Letters (Jan 2019) +79. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2017) +80. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Proceedings of European Conference on Computer Vision (2022) +81. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. ArXiv abs/2311.03352 (2023) + +82. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision (2021) +83. Zhu, C., Chen, L.: A survey on open-vocabulary detection and segmentation: Past, present, and future. arXiv preprint (2023) +84. Zhu, F., Zhu, Y., Lee, V., Liang, X., Chang, X.: Deep learning for embodied vision navigation: A survey. arXiv preprint (2021) \ No newline at end of file diff --git a/2024/Open-Vocabulary Camouflaged Object Segmentation/images.zip b/2024/Open-Vocabulary Camouflaged Object Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2717bd8eedd35754ea3dc441a2be5a54406a73f2 --- /dev/null +++ b/2024/Open-Vocabulary Camouflaged Object Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6815c4c5b4996c09baacd66c26c313885ac40a508d3215f9480699912961fbb1 +size 747006 diff --git a/2024/Open-Vocabulary Camouflaged Object Segmentation/layout.json b/2024/Open-Vocabulary Camouflaged Object Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..36d2206faa999492a4fed5d6e5ea69e474aa7b28 --- /dev/null +++ b/2024/Open-Vocabulary Camouflaged Object Segmentation/layout.json @@ -0,0 +1,10094 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "text", + "content": "Open-Vocabulary Camouflaged Object Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "spans": [ + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "content": "Youwei Pang" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "content": ", Xiaqi Zhao" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "content": ", Jiaming Zuo" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "content": ", Lihe Zhang" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "inline_equation", + "content": "^{1\\star\\star}" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "content": ", and Huchuan Lu" + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "spans": [ + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "text", + "content": " Dalian University of Technology \n" + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "text", + "content": " X3000 Inspection Co., Ltd \n{lartpang, zxq} " + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "text", + "content": " mail.dlut.edu.cn \nklaus@3000gy.com, {zhanglihe, lhchuan} " + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 189, + 201, + 423, + 247 + ], + "type": "text", + "content": " dlut.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 278, + 452, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 278, + 452, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 278, + 452, + 540 + ], + "type": "text", + "content": "Abstract. Recently, the emergence of the large-scale vision-language model (VLM), such as CLIP, has opened the way towards open-world object perception. Many works have explored the utilization of pre-trained VLM for the challenging open-vocabulary dense prediction task that requires perceiving diverse objects with novel classes at inference time. Existing methods construct experiments based on the public datasets of related tasks, which are not tailored for open vocabulary and rarely involve imperceptible objects camouflaged in complex scenes due to data collection bias and annotation costs. To fill in the gaps, we introduce a new task, open-vocabulary camouflaged object segmentation (OVCOS), and construct a large-scale complex scene dataset (OVCampo) containing 11,483 hand-selected images with fine annotations and corresponding object classes. Further, we build a strong single-stage open-vocabulary camouflaged object segmentation transformer baseline OVCoser attached to the parameter-fixed CLIP with iterative semantic guidance and structure enhancement. By integrating the guidance of class semantic knowledge and the supplement of visual structure cues from the edge and depth information, the proposed method can efficiently capture camouflaged objects. Moreover, this effective framework also surpasses previous state-of-the-arts of open-vocabulary semantic image segmentation by a large margin on our OVCampo dataset. With the proposed dataset and baseline, we hope that this new task with more practical value can further expand the research on open-vocabulary dense prediction tasks. Our code and data can be found in the link." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 552, + 436, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 552, + 436, + 563 + ], + "spans": [ + { + "bbox": [ + 160, + 552, + 436, + 563 + ], + "type": "text", + "content": "Keywords: Open Vocabulary " + }, + { + "bbox": [ + 160, + 552, + 436, + 563 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 552, + 436, + 563 + ], + "type": "text", + "content": " Camouflaged Object Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 585, + 229, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 585, + 229, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 585, + 229, + 597 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 611, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 635 + ], + "type": "text", + "content": "The ability to identify and reason about object regions in a visual scene is essential for a wide range of human activities. As a complex and fundamental task in" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 225, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 225, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 225, + 653 + ], + "type": "text", + "content": "* Equal Contribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 654, + 237, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 237, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 237, + 665 + ], + "type": "text", + "content": "\\*\\* Corresponding Author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 439 + ], + "type": "text", + "content": "computer vision, detecting and segmenting objects with diverse appearances in complex scenes is also an important challenge, which is crucial for applications across vision and robotics fields, including autonomous driving [3], medical image analysis [34], and intelligent robotics [84], to name a few. In the past few years, numerous typical methods [6, 7, 19, 69] have emerged with the help of a mass of labeled data, which have greatly promoted the development of related fields such as semantic image segmentation (SIS) [11,54]. However, existing works have mainly focused on predefined closed-set scenarios, where all semantic concepts are seen during both the inference and training phases. Such a scenario setting oversimplifies the real-world complexity. To this end, many explorations have been contributed to open vocabulary settings [56, 83]. In existing work, open vocabulary learning utilizes visual-related language data (e.g., object class text or descriptions) to align images and language features, enabling models to perceive novel classes without extensive labeled data [56]. Recently, large-scale pretrained visual language models (VLMs) such as CLIP [9,47] have been gaining attention. Image-text matching-based learning mechanism provides them with the ability to align textual and visual signals well, and many works demonstrate their potential for open vocabulary tasks. Besides, data plays an important role in open-vocabulary tasks. Existing open-vocabulary semantic image segmentation (OVSIS) tasks rely on related public datasets [4,12,40,41,79], while they are not designed for the open-vocabulary setting, and there is high semantic similarity between their class definitions as revealed in [59]. Due to the data collection bias and annotation cost constraints, the existing open-vocabulary benchmarks lack special attention to finely perceive the objects of interest in concealed scenes. And the widely used VLMs are pre-trained on image-text pairs with inherent object concept bias, thus, their ability to segment objects in complex scenes remains to be verified." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 441, + 298, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 298, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 298, + 632 + ], + "type": "text", + "content": "In this paper, we introduce a new open-vocabulary segmentation task OVCOS dedicated to analyzing camouflaged object perception in diverse natural scenes. And a large-scale data benchmark, named OVCamo, is carefully constructed. Besides, we also design a strong baseline OVCoser for the proposed OVCOS, based on the VLM-driven single-stage paradigm. The camouflage arises from several sources, including similar patterns to the environment (e.g., color and texture) and imperceptible attributes (e.g., small size and heavy occlusion) as statistically illustrated in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 464, + 481, + 558 + ], + "blocks": [ + { + "bbox": [ + 306, + 464, + 481, + 558 + ], + "lines": [ + { + "bbox": [ + 306, + 464, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 481, + 558 + ], + "type": "image", + "image_path": "7051fe0903c93b57e760b3d3fc06af83e0bdb08e8028394555ff2fc233ebb1ef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 566, + 482, + 622 + ], + "lines": [ + { + "bbox": [ + 304, + 566, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 482, + 622 + ], + "type": "text", + "content": "Fig.1: The perception and recognition of camouflaged objects require the collaboration of information from multiple sources such as appearance clues, spatial structure, contour details, and object semantics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "3 To ensure the reliability of the definition, we follow existing published literature and collect data from the class-agnostic camouflage perception field, i.e., CSU [15]." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 249 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 249 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 249 + ], + "type": "image", + "image_path": "15933d076926daf2632908164e92c8e2ffdefb4c006b632973dc4b56a762a330.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 258, + 479, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 258, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 479, + 269 + ], + "type": "text", + "content": "Fig. 2: Class distribution in the proposed OVCamo dataset and some visual examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 434 + ], + "type": "text", + "content": "Considering the imperceptible appearance of camouflaged objects, accurate recognition and capture actually depend more on the cooperation of multi-source knowledge. As shown in Fig. 1, in addition to visual appearance cues, we introduce the depth for the spatial structure of the scene, the edge for the regional changes about objects, and the text for the context-aware class semantics. Considering the cooperative relationship between class recognition and object perception, the iterative learning strategy is introduced to feed back the optimized semantic relationship, resulting in more accurate object semantic guidance. This top-down conceptual reinforcement can further optimize open-vocabulary segmentation performance. With the help of the iterative multi-source information joint learning strategy, our method OVCoser shows good performance in the proposed OVCOS task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 435, + 392, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 435, + 392, + 446 + ], + "spans": [ + { + "bbox": [ + 146, + 435, + 392, + 446 + ], + "type": "text", + "content": "In summary, our contributions are three-fold as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 453, + 480, + 593 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 453, + 480, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 480, + 486 + ], + "type": "text", + "content": "- New Challenge. In view of the limitations of the existing OVSIS, we introduce a more challenging OVCOS task for open-vocabulary segmentation of camouflaged objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 488, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 533 + ], + "type": "text", + "content": "- New Benchmark. A new large-scale benchmark OVCamo with diverse samples carefully collected from existing publicly available data is proposed to better evaluate and analyze the generalization of algorithms on the proposed task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 536, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 593 + ], + "type": "text", + "content": "- **Strong Baseline.** We build a robust single-stage baseline based on CLIP, in which the proposed iterative semantic guidance and structure enhancement are embedded. Under the joint optimization of multi-source information, our approach OVCoser outperforms existing OVSIS algorithms on the new benchmark." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 618, + 242, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 242, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 242, + 630 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Vision-Language Pre-training. The core goal of visual-language pre-training is to learn generic representations of vision and language, and connect visual and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "language concepts. Early approaches [5, 29, 32, 52] are based on some relatively small and clean public datasets, which limits their achievable performance, and their fine-tuning on specific downstream tasks hinders application flexibility. In [39], image-text pairs collected from the Internet bring clear performance improvement for the retrieval task. This also indirectly encourages subsequent further exploration, such as CLIP [9,47]. It benefits from the larger scale of noisy data from web pages covering diverse and rich concepts and exhibits impressive open-vocabulary capabilities. This work introduces CLIP as the bedrock of open-vocabulary capability and borrows its strong image-text matching capability to build an effective baseline for the OVCOS task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 240, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 589 + ], + "type": "text", + "content": "Open Vocabulary Semantic Image Segmentation (OVSIS). Although a variety of pipelines have emerged as summarized in [83], from an overall perspective, their efforts are similar, namely, how to align class name/description semantic embedding with visual features to anchor relevant object cues in the representation space. The early pioneer work [70] attempts to connect word concepts and semantic relations, and encodes word concept hierarchy to parse images. Due to the leading performance of the VLM on the image-text joint modeling, it has been gradually applied in the OVSIS field. In terms of structure, existing schemes can be roughly categorized into two types: two-stage [33,58,60,65] and single-stage [10,64,80]. In [65], a rough segmentation map for each class is created from a VLM and then refined by the test-time augmentation. These rough maps are utilized as pseudo-labels for subsequent fine segmentation by stochastic pixel sampling. SimSeg [60] adopts a cascaded design including class-agnostic proposal generation by MaskFormer [7] and class assignment by CLIP [47]. Furthermore, OVSeg [33] fine-tunes CLIP on the noisy but diverse data to improve its generalization to masked images. In lieu of using existing SIS models, a text-to-image diffusion model is introduced to generate mask features with implicit image captions in [58]. The single-stage design is more flexible and simpler. MaskCLIP [80] directly modifies CLIP for semantic segmentation without training, while SAN [59] achieves better performance with the help of adapters. CAT-Seg [10] highlights the importance of the cost aggregation between image and text embeddings for the OVSIS decoding. Recent FC-CLIP [64] investigates the hierarchical CLIP image encoder. Although these methods show impressive OVSIS performance, they still follow the generalized segmentation paradigm [6,7] in SIS, ignoring valuable auxiliary cues for object perception. This also causes them to struggle with objects camouflaged in complex scenes. Unlike them, our method, which is tailored for OVCOS, can effectively tap into the camouflaged object by integrating task-specific multi-source knowledge from visual appearance, spatial structure, object contour, and class semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Camouflaged Scene Understanding (CSU). This is a research hotspot in the computer vision community, aiming to perceive objects with camouflage [15]. Different from traditional object perception like salient object detection [22, 23, 30, 31, 42, 45, 46, 63, 66–68, 72–77] and semantic object segmentation [21, 71], CSU is obviously a more challenging problem. It can be applied in some specific fields, such as medical analysis [13, 16] and agricultural man" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "agement [35, 49]. This topic is currently defined as the class-agnostic form, focusing on the area of camouflaged objects in the visual scene. The available work [8, 13, 14, 24, 28, 43, 44, 62] to date has demonstrated promising performance on existing data benchmarks [2, 8, 14, 27, 38, 51]. Unlike previous settings, the proposed OVCOS task requires further perception of object classes. Admittedly, the publicly available data provides critical support for this new task. It helps us take the first step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 220, + 258, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 258, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 258, + 232 + ], + "type": "text", + "content": "3 OVCamo Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 366 + ], + "type": "text", + "content": "This work focuses on open-vocabulary segmentation in the camouflaged scene. It enriches the connotation of OVSIS and provides a more challenging benchmark. Note that our aim is not to claim novelty through class annotations alone, but rather to focus on how such annotations can pave the way for new tasks and methods. Our dataset provides additional features such as clean data, less ambiguity, and sensible class hierarchies, which provide the necessary data for solving the new and complex task. It not only adds complexity and challenge, but also promotes innovation and advances the field through the need for advanced methods. We believe that this dataset can lead to new opportunities in areas such as wildlife conservation and intelligence sensing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 298, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 298, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 298, + 533 + ], + "type": "text", + "content": "Image Collection. Our data is collected from existing CSU datasets that have finely annotated segmentation maps. Specifically, the OV-Camo integrates 11,483 hand-selected images covering 75 object classes reconstructed from several public datasets [2, 8, 14, 61, 78]. The distribution of the number of samples in different classes is shown in Fig. 2. Meanwhile, we consider attributes of objects when selecting images, such as object concentration, average color ratio, object-image area ratio, number" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 304, + 423, + 479, + 521 + ], + "blocks": [ + { + "bbox": [ + 304, + 401, + 481, + 422 + ], + "lines": [ + { + "bbox": [ + 304, + 401, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 481, + 422 + ], + "type": "text", + "content": "Table 1: Attributes involved in the dataset analysis." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 423, + 479, + 521 + ], + "lines": [ + { + "bbox": [ + 304, + 423, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 479, + 521 + ], + "type": "table", + "html": "
AttributeDescription
Object ConcentrationObject pixel concentration, which calculates the area ratio between the object region and its minimum rotatable bounding box.
Average Color RatioColor ratio of the object to the background, which calculates the average of the three color channels in their respective regions.
Object-Image Area RatioArea ratio of the object relative to the image.
Number of Object PartsNumber of separate areas in the image that belong to the object.
Normalized CentroidObject centroid coordinates, which are normalized using the image shape.
", + "image_path": "87288ba2cf10ba746f92188663604bc7473ab6c60629936fe6d656cf21caeeab.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 604 + ], + "type": "text", + "content": "of object parts, and normalized centroid. Tab. 1 gives their definitions. And Fig. 3 visualizes the attribute distribution of the proposed dataset. The camouflaged objects of interest usually have complex shape Fig. 3a, high similarity to the background Fig. 3b, and small size Fig. 3c. And the image often contains multiple camouflaged objects or sub-regions with a central bias as shown in Figs. 3d and 3e." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Annotation. The original annotation cannot be used directly in the open-vocabulary setting, due to the following semantic ambiguities caused by different annotation standards, i.e., 1) Broad concepts, such as \"fish\" and \"bug\"; 2) Vague definitions, such as \"small fish\" and \"black cat\"; 3) Inconsistent granularity, such as the coexistence of \"orchid mantis\" and \"mantis\"; 4) Non-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 246, + 171 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 246, + 171 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 246, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 246, + 171 + ], + "type": "image", + "image_path": "c782ee5c2f964548f52a19423457a3aaf7e420ac8f8c0afc6769d410db394d68.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 171, + 237, + 180 + ], + "lines": [ + { + "bbox": [ + 144, + 171, + 237, + 180 + ], + "spans": [ + { + "bbox": [ + 144, + 171, + 237, + 180 + ], + "type": "text", + "content": "(a) Object Concentration." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 358, + 171 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 358, + 171 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 358, + 171 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 358, + 171 + ], + "type": "image", + "image_path": "c97fb15807e94d5f10c05f70741d4a3bf8e6fd0b429f63f7f7ec2b508af86437.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 171, + 347, + 180 + ], + "lines": [ + { + "bbox": [ + 258, + 171, + 347, + 180 + ], + "spans": [ + { + "bbox": [ + 258, + 171, + 347, + 180 + ], + "type": "text", + "content": "(b) Average Color Ratio." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 180, + 246, + 236 + ], + "blocks": [ + { + "bbox": [ + 136, + 180, + 246, + 236 + ], + "lines": [ + { + "bbox": [ + 136, + 180, + 246, + 236 + ], + "spans": [ + { + "bbox": [ + 136, + 180, + 246, + 236 + ], + "type": "image", + "image_path": "62e7170139bc7d1baf091e121a780fb85c2bedfba0578a1cc3cae7741dea61a0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 237, + 243, + 247 + ], + "lines": [ + { + "bbox": [ + 138, + 237, + 243, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 237, + 243, + 247 + ], + "type": "text", + "content": "(c) Object-Image Area Ratio." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "text", + "content": "Fig. 3: Attribute visualization of objects in our OVCampo dataset, including object concentration, object-background color ratio, object-image area ratio, number of object parts, and normalized centroid. Please refer to Tab. 1 for details about attributes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 181, + 356, + 236 + ], + "blocks": [ + { + "bbox": [ + 249, + 181, + 356, + 236 + ], + "lines": [ + { + "bbox": [ + 249, + 181, + 356, + 236 + ], + "spans": [ + { + "bbox": [ + 249, + 181, + 356, + 236 + ], + "type": "image", + "image_path": "d70c181511b1804039f21147df1f7722d96498ca630affdf21e8fa183371c299.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 237, + 353, + 247 + ], + "lines": [ + { + "bbox": [ + 251, + 237, + 353, + 247 + ], + "spans": [ + { + "bbox": [ + 251, + 237, + 353, + 247 + ], + "type": "text", + "content": "(d) Number of Object Parts." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 360, + 116, + 478, + 235 + ], + "blocks": [ + { + "bbox": [ + 360, + 116, + 478, + 235 + ], + "lines": [ + { + "bbox": [ + 360, + 116, + 478, + 235 + ], + "spans": [ + { + "bbox": [ + 360, + 116, + 478, + 235 + ], + "type": "image", + "image_path": "a5347468efba2f7697113810c0c2056e4576654a558a32bc3c64a69445569da9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 236, + 463, + 245 + ], + "lines": [ + { + "bbox": [ + 374, + 236, + 463, + 245 + ], + "spans": [ + { + "bbox": [ + 374, + 236, + 463, + 245 + ], + "type": "text", + "content": "(e) Normalized Centroid." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": "entity concepts, such as \"other\". These issues can lead to unreasonable and unreliable results for the open-vocabulary prediction as discussed in [81]. To this end, we relabel the classes of all camouflaged objects and take the generality of the concept as the criterion for class definition, which also ensures lower semantic similarity. The similarity analysis and rough hierarchy of the class set " + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 312, + 482, + 384 + ], + "type": "text", + "content": " can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "type": "text", + "content": "Data Division. To objectively evaluate the open-vocabulary segmentation algorithm on unseen classes, we assign as many classes as possible to the test set and control the sample ratio of the training set to the testing set to be 7:3. Specifically, 14 classes in the dataset are taken as the training set and all the remaining 61 classes are used for testing. Such a setting follows the existing practice in OVSIS where the number of seen classes (e.g., 171 [4]) is usually fewer than unseen classes (e.g., 847 [79]) which is also closer to the real-world setting. Such a setup can ensure the quantity of training samples while reinforcing the complexity of the test set. Finally, the overall ratio of samples is 7713:3770." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 233, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 233, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 233, + 523 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 533, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 557 + ], + "type": "text", + "content": "In this section, we introduce a strong baseline OVCoser. The overall framework is first described, followed by the details of the key components." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Overall Architecture. We follow the common encoder-decoder paradigm and the pipeline is shown in Fig. 4. Specifically, we first leverage the textual encoder " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{t}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " of the frozen CLIP to extract semantic embedding " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " from the class label set " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", and the visual encoder " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " to extract multi-scale image features " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{f^{i}\\}_{i = 1}^{5}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". Both of them are fed into the decoder as the information bedrock of object segmentation, as shown in Fig. 5. Structural cues such as depth and edge are also introduced to assist the iterative refinement process. Finally, the class-related segmentation " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " which is the segmentation logits " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " after the sigmoid function processing, is used to remove the interference from the background in the high-level image" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 264 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 264 + ], + "type": "image", + "image_path": "98d7bc5ca58c4f9b33fea5e9602818bdff4276a6ebd2c44fadc098d84aef0019.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": "Fig. 4: Overview of our single-stage open-vocabulary camouflaged object segmentation framework, OVCoser. It is based on the frozen CLIP model which includes feature encoder " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_v" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " and embedding layer " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_v" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " for visual appearance cues and textual encoder " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " for class semantic information. The well-constructed prompt template set Camo-Prompts " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " enables textual embedding " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " to be more appropriate to OVCOS. In the iterative refinement decoder (Fig. 5), a more accurate class-related segmentation " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "P_s" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": ", can be obtained with the assistance of the semantic guidance " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " (Fig. 6a) from the class semantic and the structure enhancement (Fig. 6b) from the auxiliary depth and edge supervisions, i.e., " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": ". After being masked average pooled (MAP) by " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "P_s" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": ", the high-level feature " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "f^5" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " is used to assign a class " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "P_c" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " to the object from the class set " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 130, + 272, + 482, + 405 + ], + "type": "text", + "content": " are only used as auxiliary supervision during training and are not required during inference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "text", + "content": "feature and guide the generation of object-oriented visual representation " + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "text", + "content": ". And the class label " + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "text", + "content": " is determined by the similarity matching between " + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 433, + 482, + 479 + ], + "type": "text", + "content": ". More details about the proposed architecture are listed in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": "Semantic Guidance (SG). The class label definition usually is independent of the image scene. It is very important to fully utilize the class prior for object recognition in complex scenes. In the proposed decoder, normalized textual embedding is introduced into each stage to highlight semantically relevant cues. As shown in Fig. 6a, we design a semantic guidance component SG to inject concept cues into the self-enhancement of the image feature. Specifically, in our semantic guidance attention (SGA), the normalized image feature " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": " is linearly mapped to " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": ", while the textual embedding " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": " is transformed to the class guidance vector " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": ". The similarity between " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": " reflects the activation of different classes in spatial locations. In Agg of Fig. 6a, the base weight " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "W_{b}" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": " for the spatial guidance is obtained after highlighting the most relevant class information by softmax operation. And then " + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 483, + 482, + 639 + ], + "type": "text", + "content": " is modulated and fed into MHSA, i.e., multi-head self-attention [55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Structure Enhancement (SE). Existing methods demonstrate that low-level structural information, such as the edge [53] and the depth [57], plays an important role in the design of neural networks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 445, + 220 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 445, + 220 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 445, + 220 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 445, + 220 + ], + "type": "image", + "image_path": "14a8ad6ee324e9607f42bbf7fea10bb3c82c438b707e7854408c3893b4ced91a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "text", + "content": "Fig. 5: Proposed pipeline of the iterative refinement decoder with semantic guidance (SG) and structure enhancement (SE) components denoted as \"SG- " + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "text", + "content": "\" and \"SE- " + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "text", + "content": "\". " + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 228, + 482, + 262 + ], + "type": "text", + "content": " denote the current step and total number of iterations, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 274, + 297, + 358 + ], + "blocks": [ + { + "bbox": [ + 142, + 274, + 297, + 358 + ], + "lines": [ + { + "bbox": [ + 142, + 274, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 142, + 274, + 297, + 358 + ], + "type": "image", + "image_path": "688eea1ab6c8901f7cc64598553b5ac3d07abb0b4da8eb5ddc8f59b7749019bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 361, + 231, + 370 + ], + "lines": [ + { + "bbox": [ + 206, + 361, + 231, + 370 + ], + "spans": [ + { + "bbox": [ + 206, + 361, + 231, + 370 + ], + "type": "text", + "content": "(a) SG" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 274, + 463, + 358 + ], + "blocks": [ + { + "bbox": [ + 326, + 274, + 463, + 358 + ], + "lines": [ + { + "bbox": [ + 326, + 274, + 463, + 358 + ], + "spans": [ + { + "bbox": [ + 326, + 274, + 463, + 358 + ], + "type": "image", + "image_path": "cade2dedc198785643b56f9429a0e2194f2b5d60d85332e0c2aa122919bb6c40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 361, + 408, + 370 + ], + "lines": [ + { + "bbox": [ + 383, + 361, + 408, + 370 + ], + "spans": [ + { + "bbox": [ + 383, + 361, + 408, + 370 + ], + "type": "text", + "content": "(b) SE" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 379, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 131, + 379, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 482, + 424 + ], + "type": "text", + "content": "Fig. 6: Proposed semantic guidance (SG) and structure enhancement (SE) components in which semantic guidance attention (SGA) and structure enhancement attention (SEA) are embedded. The dashed blocks represent the feedback paths that come into play during iteration." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": "tant role in CSU, which is closely related to the mechanism of the human visual system. So the SE component attached to the low-level SG is proposed to integrate the edge-aware and depth-aware cues and improve the structural details. Specifically, the output of the SG is fed into two separate branches containing the convolutional stem and head for the edge and depth estimation as shown in Fig. 5. The edge and depth logits maps, i.e., " + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "inline_equation", + "content": "M_e^i" + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "inline_equation", + "content": "M_d^i" + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": ", from the head in the branch of the layer " + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,3\\}" + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": " are directly supervised. And the outputs " + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "inline_equation", + "content": "f_e^i" + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "inline_equation", + "content": "f_d^i" + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": " of the stem are fed into the SE. In the structure enhancement attention (SEA), they independently update the normalized visual features " + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": " using MHSA [55], and the corresponding outputs are combined with the learnable weight " + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 449, + 482, + 581 + ], + "type": "text", + "content": " as in Fig. 6b." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Iterative Refinement. In the SG component, the aggregation process between image features and class semantics is not aligned and requires data-driven optimization. Considering the aligned embedding space of the pre-trained CLIP, we introduce the correlation matrix " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{cor}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " between the visual and textual embeddings into the SG as shown in Fig. 6a. Meanwhile, due to the emphasis of the decoder output on the object region, the object-aware representation " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{obj}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is also inputted, which comes from the image features pooled by the coarse" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "segmentation prediction in the last iteration. By combining the two, we obtain task-oriented object cues, which is actually inspired by the top-down attention mechanism in the human cognitive system [1,25]. The spatial activation map " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "W_{r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " of such object cues over image features is used to re-modulate " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "W_{b}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ". Besides, the SE in the iteration also helps the model to further optimize the texture details. To benefit as much as possible from the assistance from structure enhancement while avoiding over-computation, we set the iteration entry to the third decoding layer as shown in Fig. 5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 320 + ], + "type": "text", + "content": "CamoPrompts. As mentioned in [47], prompt engineering and ensembling are important for the transfer performance of CLIP on downstream tasks, and the prompt template should be more relevant to the data type. Because additional task-related cues are generally able to impose the necessary contextual constraint to the flexible CLIP. Hence, instead of common practices [18, 47, 82], we design a simpler yet more effective template set CamoPrompts tailored for OVCOS to decorate the class name, and average their textual embeddings as the final semantic embedding for each class. Its full form is depicted in Tab. 4, while it also achieves better classification performance in comparison with other forms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": "Supervision. In each iteration, in addition to semantic object segmentation, we also need to perform depth estimation and edge estimation as auxiliary tasks. For the segmentation prediction, we follow the commonly used weighted segmentation loss function " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "l_{s}^{t} = l_{s}(P_{s}^{t},G_{s})" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": " [53,62]. For the edge estimation, considering the imbalance problem of positive and negative samples, we introduce the dice loss function as " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "l_{e}^{i,t} = l_{e}(P_{e}^{i,t},G_{e})" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": ". The summation of L1 and SSIM losses, i.e., " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "l_{d}^{i,t} = l_{d}(P_{d}^{i,t},G_{d})" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": ", is used for the depth estimation. The total loss " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": " of our method can be formulated as follows: " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "L = \\sum_{t=1}^{T}\\left[l_{s}^{t} + \\sum_{i=1}^{3}\\left(l_{e}^{i,t} + l_{d}^{i,t}\\right)\\right]" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": " are used to index iterations and layers, respectively. And the total number " + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 321, + 482, + 447 + ], + "type": "text", + "content": " of iterations is set to 2 as mentioned in Sec. 5.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 469, + 230, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 230, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 230, + 483 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Dataset Settings. As mentioned in Sec. 3, the proposed dataset is divided according to classes into two disjoint subsets, i.e., " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{seen}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{unseen}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". The former contains 14 classes for training and the latter contains the remaining 61 classes for testing. The proposed dataset itself is provided with only images and masks. We use the typical monocular depth estimation methods [26,48] to obtain the depth map " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " for training, while the edge map " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{e}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is generated by dilating and eroding operations. We introduce depth and edge maps only in the training phase and they are not required during inference, which also avoids information leakage during testing. So, our depth and edge maps are labor-free and training-only, and directly generated by existing algorithms. Such a design can help our model get rid of the dependence on them in real-world applications. Our structure enhancement based on depth and edge information belongs to a deeper exploration" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 289 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 289 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 289 + ], + "type": "image", + "image_path": "90754cec10618a0faa07e5dd09531883e702474a4cbd1e7fd5c5b512665c2d20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 352 + ], + "type": "text", + "content": "Fig. 7: Visual results on OVCamo. Existing methods are either disrupted by chaotic backgrounds, imperceptible appearances, blurry details, or severe occlusion, while our algorithm can effectively capture and remain well-exposed object details. And three different colors are used to represent human annotations, correct predictions, and incorrect predictions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "content": "of this task. These generated depth maps will be made publicly available with the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 481, + 495 + ], + "type": "text", + "content": "Training Details. Following previous settings in [59, 60, 64], the pre-trained CLIP is frozen during training, and the remaining parameters are learnable and are randomly initialized. The AdamW [37] optimizer with the learning rate of 3e-6, weight decay of 5e-4, batch size of 4, and training epoch of 30 is used to optimize model parameters. Some basic data augmentations including random flipping, rotating, and color jittering, are introduced to preprocess training data. The input data and the output logits are bilinearly interpolated to " + }, + { + "bbox": [ + 130, + 400, + 481, + 495 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 130, + 400, + 481, + 495 + ], + "type": "text", + "content": " during training and inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "text", + "content": "Evaluation Protocol. To reasonably evaluate the performance of OVCOS, we modify the metrics in the original CSU task [8,13] to " + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{cS}_m" + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{cF}_{\\beta}^{\\omega}" + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "text", + "content": ", cMAE, " + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{cF}_{\\beta}" + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{cE}_m" + }, + { + "bbox": [ + 130, + 496, + 481, + 544 + ], + "type": "text", + "content": ", and cIoU which follows the common settings in the OVSIS field [10,33,59, 60,64,83] and takes into account both classification and segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 562, + 288, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 288, + 574 + ], + "type": "text", + "content": "5.2 System-level Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "To show the complexity of the proposed OVCOS task and also to verify the effectiveness of the proposed method, we compare OVCoser with several recent state-of-the-art methods in OVSIS, including [10,33,58-60,64]. Since this is a new task, existing methods need to be re-evaluated to understand their generalization ability. Based on the public code and weights trained on COCO-Stuff [4] provided by the authors, we show the performance of these methods under three different testing schemes, including S.I) testing directly with their trained weights; S.II)" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 158, + 479, + 361 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 2: Comparison with recent state-of-the-art CLIP-based open-vocabulary semantic image segmentation methods with different training settings on the proposed OVCamo dataset. The best three results are highlighted in red, green and blue." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 158, + 479, + 361 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 479, + 361 + ], + "type": "table", + "html": "
ModelVLMFeature BackboneText PromptcSm↑cFωβ↑cMAE↓cFβ↑cEm↑cloU↑
Test on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.1280.1050.8380.1120.1430.094
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.3410.3060.5840.3250.3840.273
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.4090.3390.5000.3410.4210.302
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.4140.3430.4890.3570.4560.319
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.4300.3440.4480.3660.4590.310
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.3740.3060.5390.3200.4090.285
Finetune on OVCamo with the weight trained on COCO.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0980.0710.8520.0810.1280.066
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.1640.1310.7630.1470.2080.123
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1820.1250.6910.2190.3090.189
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.3210.2160.5500.2360.3310.204
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1850.0940.7020.1100.1850.088
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.1240.0740.7980.0880.1620.072
Train on OVCamo.
SimSeg21[60]CLIP-ViT-B/16 [47]ResNet-101 [20]Learnable [82]0.0530.0490.9210.0560.0980.047
OVSeg22[33]CLIP-ViT-L/14 [47]Swin-B [36][18]0.0240.0460.9540.0560.1300.046
ODISE23[58]CLIP-ViT-L/14 [47]StableDiffusionv1.3 [50][17]0.1870.1190.7000.2110.2980.167
SAN23[59]CLIP-ViT-L/14 [47]ViT Adapter[18]0.2750.2020.6120.2200.3180.189
CAT-Seg23[10]CLIP-ViT-L/14 [47]Swin-B [36][47]0.1810.1060.7190.1230.1960.094
FC-CLIP23[64]CLIP-ConvNeXt-L [9]—[18]0.0800.0760.8720.0900.1910.072
OursCLIP-ConvNeXt-L [9]—CamoPrompts0.5790.4900.3360.5200.6160.443
", + "image_path": "1c13bee2fcd4e7bf16c8ebfcafd0f9e9ca7115f6825adbfac77b3d54af693415.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 387, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 411 + ], + "type": "text", + "content": "further fine-tuning based on trained weights before testing; and S.III) testing after re-training directly on our training set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 591 + ], + "type": "text", + "content": "Quantitative Evaluation. For the sake of fairness in comparisons, we report the performance of the \"Large\" versions for these methods, except for SimSeg [60] where the authors only provide the \"Base\" version. All results are summarized in Tab. 2 and our approach consistently outperforms these competitors. It is worth noting that existing methods perform better at S.I. This may be attributed to the training process on the larger-scale COCO-Stuff dataset, which provides a more general understanding of the concepts. However, direct fine-tuning (i.e., S.II) may destroy this knowledge and even cause oblivion to some extent, resulting in performance degradation. If we follow S.III to re-train them, the relatively small-scale training data may not be enough to train these models with more complex structures. At the same time, the existing methods also lack targeted optimization for the OVCOS task. These problems can lead to further deterioration of performance. However, our approach tailored for OVCOS achieves leading performance by the iterative refinement strategy of multi-source information, which comprehensively considers different characteristics of the task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 593, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 479, + 666 + ], + "type": "text", + "content": "Qualitative Evaluation. We also visualize the results of some recent methods on a variety of data in Fig. 7. It can be seen that the proposed method shows better performance and adaptability to diverse objects, including large objects (Col. 1-2), middle objects (Col. 3-5), small objects (Col. 6-9), multiple objects (Col. 8), complex shapes (Col. 3-5), blurred edges (Col. 1-5), severe occlusion (Col. 6), and background interference (Col. 2-6)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 179, + 441, + 347 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": "Table 3: Ablation comparison of proposed components. " + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": " represents the average relative gain in performance of the corresponding model over the baseline for OVCOS. " + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": ": CamoPrompts. " + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": ": Semantic guidance. " + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": ": Depth estimation auxiliary task. " + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": ": Edge estimation auxiliary task. " + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": ": Number of iterations which is set to 2 by default due to the best performance. \"lim " + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "inline_equation", + "content": "\\sup_{P_s \\to G_s}" + }, + { + "bbox": [ + 130, + 114, + 482, + 173 + ], + "type": "text", + "content": " Perf.\" The ideal performance for our framework." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 179, + 441, + 347 + ], + "lines": [ + { + "bbox": [ + 164, + 179, + 441, + 347 + ], + "spans": [ + { + "bbox": [ + 164, + 179, + 441, + 347 + ], + "type": "table", + "html": "
ModelcSm↑ cFβ↑ cMAE ↓ cFβ ↑ cEm ↑ cIoU ↑Δ
Comparison of the proposed modules.
Baseline0.5170.4080.3740.4510.5490.3590.0%
+P0.5430.4350.3460.4800.5810.3836.3%
+P,C0.5500.4530.3410.4910.5970.3979.1%
+P,C,D0.5650.4730.3360.5070.6060.42212.6%
+P,C,E0.5670.4810.3390.5110.6070.43213.5%
+P,C,D,E (i.e., T=1)0.5700.4880.3380.5180.6100.43614.5%
SE (Fig. 6b) → Addition Fusion0.5520.4570.3400.4970.5990.4029.9%
Comparison of the proposed iterative refinement.
T=10.5700.4880.3380.5180.6100.43614.5%
T=20.5790.4900.3360.5200.6160.44315.5%
w/o Mcor as in Fig. 6a0.5750.4870.3370.5150.6110.44114.8%
w/o obj as in Fig. 6a0.5710.4760.3390.5060.6080.43413.4%
T=30.5760.4840.3330.5140.6140.43714.8%
lim supPseg→GsegPerf.0.7030.7030.2970.7010.7010.70151.2%
", + "image_path": "53c314442fd23bf85afe9affbc4394d0d326aa93002cfa4afc67846b41526484.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 372, + 307, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 307, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 307, + 385 + ], + "type": "text", + "content": "5.3 Analysis and Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": "Better Baseline Model. Interestingly, our baseline model in Tab. 3 outperforms existing methods in Tab. 2. We attribute this to the fact that the overly complex feature pipeline of the conventional OVSIS architecture is not well adapted to this task. And our framework, from hyperparameter setting to structural evolution, is explored for the proposed task and data. Thus, we actually provide a more effective and robust baseline setup for OVCOS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Importance of Modules. To specifically analyze the effect of different components, we evaluate their performance in Tab. 3. As can be seen, the proposed modules all show positive gains. Both the semantic guidance in the class-aware decoding and the structure enhancement from auxiliary tasks of depth and edge estimation consistently boost the final performance. The ablation comparison also demonstrates that explicit guidance of the spatial and contour information is important for the detection of camouflaged objects and their positive gains are subject to the effective fusion design. We replace the proposed structure enhancement (SE) component with the vanilla addition fusion with a similar number of parameters, which results in detrimental effects. Besides, in Tab. 3, we also give the ideal results of our CLIP-driven framework where " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " is treated as " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": ". Although our method exhibits leading performance compared to existing methods as shown in Tab. 2, there is still a long way to go to solve this problem. At the same time, the current ideal performance is still far from the limit, suggesting that future breakthroughs in this field may require more powerful paradigms as discussed in the supplementary material." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Importance of Iterative Refinement. The top-down iterative refinement strategy significantly improves the OVCOS performance as shown in Tab. 3. When the number " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " of iterations is 2, our algorithm obtains the best OVCOS performance. As " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " increases, there is no further improvement in performance, so it is set to 2 by default. In addition, the object-aware representation " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f_{obj}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " from the output space plays an important role, and introducing the correlation guidance " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "M_{cor}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " also has positive gains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 297, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 297, + 212 + ], + "type": "text", + "content": "Importance of CamoPrompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 298, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 298, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 298, + 391 + ], + "type": "text", + "content": "Our prompt template set CamoPrompts, which takes task attributes into account, shows better performance in Tab. 4. To further understand the influence of different templates on the semantic embedding, we calculate the Hausdorff distance between training and testing class labels in the embedding space, as shown in Fig. 8. The figure presents an interesting phenomenon that those templates with better classification performance tend to reduce the distance, which inspires further explorations for more effective prompt engineering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 298, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 298, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 298, + 463 + ], + "type": "text", + "content": "Importance Between Edge and Depth. In the proposed SE as shown in Fig. 6b, the interaction components for the edge and depth are combined by a coefficient vector " + }, + { + "bbox": [ + 130, + 392, + 298, + 463 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 392, + 298, + 463 + ], + "type": "text", + "content": ", which can also reflect the relative importance of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "type": "text", + "content": "the two kinds of information. In Fig. 9, we plot " + }, + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "type": "text", + "content": " from different decoding layers. It can be seen that the values are usually greater than 0.5, which indicates a preference for the edge information flow." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 499, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 481, + 571 + ], + "type": "text", + "content": "Model Complexity. We compare the proposed method with the CLIP-based competitors [10, 33, 58-60, 64] in terms of the number of trainable and total parameters, and FLOPs. To be fair, we test all methods following the setting of their original inference settings. As can be seen from Tab. 5, our method has fewer trainable parameters (" + }, + { + "bbox": [ + 130, + 499, + 481, + 571 + ], + "type": "inline_equation", + "content": "< 2\\%" + }, + { + "bbox": [ + 130, + 499, + 481, + 571 + ], + "type": "text", + "content": ") and less computational complexity (" + }, + { + "bbox": [ + 130, + 499, + 481, + 571 + ], + "type": "inline_equation", + "content": "0.2T" + }, + { + "bbox": [ + 130, + 499, + 481, + 571 + ], + "type": "text", + "content": "), which is superior to these competitors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 591, + 220, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 591, + 220, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 591, + 220, + 604 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "In this work, we propose a new challenging task, OVCOS, to explore open-vocabulary semantic image segmentation (OVSIS) for the camouflaged objects in more complex natural scenes, and carefully collect and construct a large-scale data benchmark OVCamo. Meanwhile, by considering the characteristics" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 305, + 224, + 479, + 332 + ], + "blocks": [ + { + "bbox": [ + 305, + 224, + 479, + 332 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 479, + 332 + ], + "type": "image", + "image_path": "1b3d99441641a29ae4526167b1b4f06898f27b8d62d06c59dc36418b7ebeca55.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 341, + 482, + 452 + ], + "lines": [ + { + "bbox": [ + 304, + 341, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 482, + 452 + ], + "type": "text", + "content": "Fig.8: Illustration of classification accuracy " + }, + { + "bbox": [ + 304, + 341, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 341, + 482, + 452 + ], + "type": "text", + "content": " and Hausdorff distance " + }, + { + "bbox": [ + 304, + 341, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 341, + 482, + 452 + ], + "type": "text", + "content": " between the textual embeddings of training and testing class splits by using the plain CLIP on the OVCamo dataset. The approximate negative correlation between them may reveal that the templates that bring training and testing class embeddings closer usually also improve classification accuracy. Tab. 4 shows the details of these templates." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 90, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 148, + 476, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 481, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 481, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 481, + 147 + ], + "type": "text", + "content": "Table 4: Classification accuracy " + }, + { + "bbox": [ + 132, + 125, + 481, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 132, + 125, + 481, + 147 + ], + "type": "text", + "content": " of the plain CLIP using different prompt templates on OVCamo, which is based on the masked average pooling with the ground truth mask." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 148, + 476, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 148, + 476, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 476, + 312 + ], + "type": "table", + "html": "
IDPrompt TemplateA
0".class" w/o MAP based on ground truth0.538
Task-generic templates.
1".class"0.671
2.1"The.class"0.648
2.2"the.class"0.632
2.3"A.class"0.677
2.4"a.class"0.675
3.1"A photo of a.class"0.684
3.2"A photo of the.class"0.691
3.3"The photo of a.class"0.672
3.4"The photo of the.class"0.680
3.5"a photo of a.class"0.684
3.6"a photo of the.class"0.689
3.7"the photo of a.class"0.675
3.8"the photo of the.class"0.674
4templates from [47]0.686
5templates from [18]0.682
Task-related templates.
6.1"A photo of the camouflaged.class"0.681
6.2"A photo of the concealed.class"0.674
6.36.1, 6.20.685
6.4"A photo of the.classcamouflagedin the background."0.684
6.5"A photo of the.classconcealedin the background."0.694
6.66.4, 6.50.694
6.7"A photo of the.classcamouflagedto blend in with its surroundings."0.700
6.8"A photo of the.classconcealedto blend in with its surroundings."0.697
6.96.7, 6.80.703
6.106.1 - 6.9 i.e., CamoPrompts0.704
", + "image_path": "3adf9f569bd1219b761fb570cb4a05905910d60f3ac4da1640f3b74d4e345a1b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 413, + 341, + 498 + ], + "blocks": [ + { + "bbox": [ + 132, + 356, + 346, + 411 + ], + "lines": [ + { + "bbox": [ + 132, + 356, + 346, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 346, + 411 + ], + "type": "text", + "content": "Table 5: Efficiency comparison with other methods. \"Trainable Param.\" and \"Total Param.\" stand for the number of trainable and total parameters. In OVSeg [33], the CLIP [47] model is fine-tuned by the authors, resulting in more trainable parameters." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 413, + 341, + 498 + ], + "lines": [ + { + "bbox": [ + 133, + 413, + 341, + 498 + ], + "spans": [ + { + "bbox": [ + 133, + 413, + 341, + 498 + ], + "type": "table", + "html": "
ModelTrainable Param.Total Param.FLOPs
SimSeg21 [60]61M (28.91%)211M1.9T
OVSeg22 [33]531M (100.00%)531M8.0T
ODISE23 [58]28M (1.80%)1522M5.5T
SAN23 [59]9M (2.06%)437M0.4T
CAT-Seg23 [10]104M (21.22%)490M0.3T
FC-CLIP23 [64]20M (5.38%)372M0.8T
Ours7M (1.95%)359M0.2T
", + "image_path": "af83dab3bfb4be02ccb6f28297540cc4a50cde561178f275ceb21931f0cee4d1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 353, + 348, + 474, + 453 + ], + "blocks": [ + { + "bbox": [ + 353, + 348, + 474, + 453 + ], + "lines": [ + { + "bbox": [ + 353, + 348, + 474, + 453 + ], + "spans": [ + { + "bbox": [ + 353, + 348, + 474, + 453 + ], + "type": "image", + "image_path": "3c67441ce3f068c9b28af38ea7fa0dc43c7ead4639ceff3461a87dbc0b8e12d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 461, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 346, + 461, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 346, + 461, + 482, + 495 + ], + "type": "text", + "content": "Fig. 9: " + }, + { + "bbox": [ + 346, + 461, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 346, + 461, + 482, + 495 + ], + "type": "text", + "content": " corresponding to different heads in Fig. 6b from different decoding layers." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "of the task and data, we propose a strong single-stage baseline OVCoser with the advanced pre-trained vision-language model. Specifically, the well-designed prompt templates are introduced to reinforce the task-relevant semantic context. We introduce additional multi-source information including class semantic cues, depth spatial structure, object edge details, and top-down iterative guidance from the output space. With the help of these components, OVCoser can perceive and segment camouflaged objects in complex environments. Extensive experiments demonstrate the effectiveness of the proposed method and its superior performance compared with the existing state-of-the-art OVSIS algorithms on OVCamo." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 269 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62276046 and by Dalian Science and Technology Innovation Foundation under Grant 2023JJ12GX015. We sincerely thank these authors of the public datasets [2,8,14,61,78] used in this work: Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao for COD10K [14], Yunfei Zheng, Xiongwei Zhang, Feng Wang, Tieyong Cao, Meng Sun, and Xiaobing Wang for CPD1K [78], Jinyu Yang for PlantCamo [61], Pia Bideau and Erik Learned-Miller for CAD [2], and Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge for MoCA-Mask [8]. It is their efforts that underpin this new exploration of camouflaged object segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 285, + 197, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 197, + 298 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 197, + 298 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 308, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 308, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 352 + ], + "type": "text", + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 352, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 480, + 384 + ], + "type": "text", + "content": "2. Bideau, P., Learned-Miller, E.: It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In: Proceedings of European Conference on Computer Vision (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "type": "text", + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 427, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 480, + 459 + ], + "type": "text", + "content": "4. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 459, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 492 + ], + "type": "text", + "content": "5. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: Proceedings of European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 492, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 525 + ], + "type": "text", + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "text", + "content": "7. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. International Conference on Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "type": "text", + "content": "8. Cheng, X., Xiong, H., Fan, D.P., Zhong, Y., Harandi, M., Drummond, T., Ge, Z.: Implicit motion handling for video camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "text", + "content": "9. Cherti, M., Beaumont, R., Wightman, R., Wortzman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "0. Cho, S., Shin, H., Hong, S., An, S., Lee, S., Arnab, A., Seo, P.H., Kim, S.W.: Catseg: Cost aggregation for open-vocabulary semantic segmentation. arXiv preprint (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "11. Csurka, G., Volpi, R., Chidlovskii, B.: Semantic image segmentation: Two decades of research. Foundations and Trends in Computer Graphics and Vision (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "12. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 195 + ], + "type": "text", + "content": "13. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 196, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 196, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 480, + 228 + ], + "type": "text", + "content": "14. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 229, + 480, + 251 + ], + "type": "text", + "content": "15. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 252, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 252, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 133, + 252, + 480, + 285 + ], + "type": "text", + "content": "16. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 480, + 318 + ], + "type": "text", + "content": "17. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Proceedings of European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 480, + 352 + ], + "type": "text", + "content": "18. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 353, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 353, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 353, + 480, + 374 + ], + "type": "text", + "content": "19. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. IEEE Transactions on Pattern Analysis and Machine Intelligence (Feb 2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 375, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 375, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 133, + 375, + 480, + 408 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "text", + "content": "21. Ji, W., Li, J., Bian, C., Zhou, Z., Zhao, J., Yuille, A.L., Cheng, L.: Multispectral video semantic segmentation: A benchmark dataset and baseline. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 442, + 480, + 475 + ], + "type": "text", + "content": "22. Ji, W., Li, J., Zhang, M., Piao, Y., Lu, H.: Accurate rgb-d salient object detection via collaborative learning. In: Proceedings of European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 480, + 508 + ], + "type": "text", + "content": "23. Ji, W., Yan, G., Li, J., Piao, Y., Yao, S., Zhang, M., Cheng, L., Lu, H.: DMra: Depth-induced multi-scale recurrent attention network for rgb-d saliency detection. IEEE Transactions on Image Processing (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 509, + 480, + 542 + ], + "type": "text", + "content": "24. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "type": "text", + "content": "25. Katsuki, F., Constantinidis, C.: Bottom-up and top-down attention: different processes and overlapping neural systems. The Neuroscientist (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "type": "text", + "content": "26. Kim, S.Y., Zhang, J., Niklaus, S., Fan, Y., Chen, S., Lin, Z., Kim, M.: Layered depth refinement with mask guidance. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "text", + "content": "27. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "28. Li, A., Zhang, J., Lyu, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Li, G., Duan, N., Fang, Y., Gong, M., Jiang, D.: Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In: AAAI Conference on Artificial Intelligence (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "text", + "content": "30. Li, J., Ji, W., Wang, S., Li, W., Cheng, L.: Dvsod: RGB-d video salient object detection. In: International Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "31. Li, J., Ji, W., Zhang, M., Piao, Y., Lu, H., Cheng, L.: Delving into calibrated depth for accurate rgb-d salient object detection. International Journal of Computer Vision (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Proceedings of European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 295 + ], + "type": "text", + "content": "33. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "34. Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., van der Laak, J.A., van Ginneken, B., Sánchez, C.I.: A survey on deep learning in medical image analysis. Medical Image Analysis (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "35. Liu, L., Wang, R., Xie, C., Yang, P., Wang, F., Sudirman, S., Liu, W.: Pestnet: An end-to-end deep learning approach for large-scale multi-class pest detection and classification. IEEE Access (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "36. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. Proceedings of the IEEE International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "37. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "text", + "content": "38. Lyu, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "39. Mithun, N.C., Panda, R., Papalexakis, E.E., Roy-Chowdhury, A.K.: Webly supervised joint embedding for cross-modal image-text retrieval. In: Proceedings of the ACM International Conference on Multimedia (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "text", + "content": "40. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "text", + "content": "41. Neuhold, G., Ollmann, T., Bulo, S.R., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: Proceedings of the IEEE International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "42. Pang, Y., Zhang, L., Zhao, X., Lu, H.: Hierarchical dynamic filtering network for rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "43. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "44. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "45. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "text", + "content": "46. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE Transactions on Image Processing (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "text", + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "type": "text", + "content": "48. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "type": "text", + "content": "49. Rizzo, M., Marcuzzo, M., Zangari, A., Gasparetto, A., Albarelli, A.: Fruit ripeness classification: A survey (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "50. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "51. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database (2017), http://kgwisc.aei.polsl.pl/index.php/pl/dataset/63-animal-camouflage-analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "text", + "content": "52. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "text", + "content": "53. Sun, Y., Wang, S., Chen, C., Xiang, T.Z.: Boundary-guided camouflaged object detection. In: International Joint Conference on Artificial Intelligence (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "54. Thisanke, H., Deshan, C., Chamith, K., Seneviratne, S., Vidanaarachchi, R., Herath, D.: Semantic segmentation using vision transformers: A survey. arXiv preprint (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "text", + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, u., Polosukhin, I.: Attention is all you need. In: International Conference on Neural Information Processing Systems (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 451, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 481, + 483 + ], + "type": "text", + "content": "56. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. arXiv preprint (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "text", + "content": "57. Xiang, M., Zhang, J., Lv, Y., Li, A., Zhong, Y., Dai, Y.: Exploring depth contribution for camouflaged object detection (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "59. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "60. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: Proceedings of European Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 469, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 469, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 469, + 611 + ], + "type": "text", + "content": "61. Yang, J.: Plantcamo dataset (2023), https://github.com/yjybuaa/PlantCamo" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "type": "text", + "content": "62. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "63. Yu, Q., Zhao, X., Pang, Y., Zhang, L., Lu, H.: Multi-view aggregation network for dichotomous image segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "64. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In: International Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "65. Zabari, N., Hoshen, Y.: Open-vocabulary semantic segmentation using test-time distillation. In: European Conference on Computer Vision Workshops (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "text", + "content": "66. Zhang, M., Ji, W., Piao, Y., Li, J., Zhang, Y., Xu, S., Lu, H.: Lfnet: Light field fusion network for salient object detection. IEEE Transactions on Image Processing (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "67. Zhang, M., Liu, J., Wang, Y., Piao, Y., Yao, S., Ji, W., Li, J., Lu, H., Luo, Z.: Dynamic context-sensitive filtering network for video salient object detection. In: Proceedings of the IEEE International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "68. Zhang, M., Yao, S., Hu, B., Piao, Y., Ji, W.: C2dfnet: Criss-cross dynamic filter network for rgb-d salient object detection. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "69. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. International Conference on Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "70. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: Proceedings of the IEEE International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "71. Zhao, X., Chang, S., Pang, Y., Yang, J., Zhang, L., Lu, H.: Multi-source fusion and automatic predictor selection for zero-shot video object segmentation. International Journal of Computer Vision (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 391 + ], + "type": "text", + "content": "72. Zhao, X., Pang, Y., Ji, W., Sheng, B., Zuo, J., Zhang, L., Lu, H.: Spider: A unified framework for context-dependent concept understanding. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "73. Zhao, X., Pang, Y., Zhang, L., Lu, H.: Joint learning of salient object detection, depth estimation and contour extraction. IEEE Transactions on Image Processing (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "text", + "content": "74. Zhao, X., Pang, Y., Zhang, L., Lu, H., Ruan, X.: Self-supervised pretraining for rgb-d salient object detection. In: AAAI Conference on Artificial Intelligence (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "75. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Suppress and balance: A simple gated network for salient object detection. In: Proceedings of European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "76. Zhao, X., Pang, Y., Zhang, L., Lu, H., Zhang, L.: Towards diverse binary segmentation via a simple yet general gated network. International Journal of Computer Vision (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "77. Zhao, X., Zhang, L., Pang, Y., Lu, H., Zhang, L.: A single stream network for robust and real-time rgb-d salient object detection. In: Proceedings of European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "78. Zheng, Y., Zhang, X., Wang, F., Cao, T., Sun, M., Wang, X.: Detection of people with camouflage pattern via dense deconvolution network. IEEE Signal Processing Letters (Jan 2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "79. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "80. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: Proceedings of European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "81. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. ArXiv abs/2311.03352 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "OVCOS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 182 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "82. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "83. Zhu, C., Chen, L.: A survey on open-vocabulary detection and segmentation: Past, present, and future. arXiv preprint (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "84. Zhu, F., Zhu, Y., Lee, V., Liang, X., Chang, X.: Deep learning for embodied vision navigation: A survey. arXiv preprint (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "Y. Pang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_content_list.json b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..572f5c50cf9b846c725b8ad88e72958f56102bb3 --- /dev/null +++ b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_content_list.json @@ -0,0 +1,1618 @@ +[ + { + "type": "text", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "text_level": 1, + "bbox": [ + 259, + 140, + 741, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guoqiang Zhao $^{1}$ , Junjie Huang $^{1}$ , Xiaoyun Yan $^{1,2\\dagger}$ , Zhaojing Wang $^{1,2}$ , Junwei Tang $^{1}$ , Yangjun Ou $^{1}$ , Xinrong Hu $^{1,2}$ , and Tao Peng $^{1}$", + "bbox": [ + 223, + 210, + 779, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Computer Science and Artificial Intelligence, Wuhan Textile University zgq_lyy@163.com, xyyan@wtu.edu.cn", + "bbox": [ + 225, + 253, + 776, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Hubei Garment Informatization Engineering Technology Research Center", + "bbox": [ + 250, + 282, + 750, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. RGB-Thermal (RGB-T) semantic segmentation is an important research branch of multi-modal image segmentation. The current RGB-T semantic segmentation methods generally have two unsolved and typical shortcomings. First, they do not have the open-vocabulary recognition ability, which significantly limits their application scenarios. Second, when fusing RGB and thermal images, they often need to design complex fusion network structures, which usually results in low network training efficiency. We present OpenRSS, the Open-vocabulary RGB-T Semantic Segmentation method, to solve these two disadvantages. To our knowledge, OpenRSS is the first RGB-T semantic segmentation method with open-vocabulary segmentation capability. OpenRSS modifies the basic segmentation model SAM for RGB-T semantic segmentation by adding the proposed thermal information prompt module and dynamic low-rank adaptation strategy to SAM. These designs effectively fuse the RGB and thermal information, but with much fewer trainable parameters than other methods. OpenRSS achieves the open-vocabulary capability by jointly utilizing the vision-language model CLIP and the modified SAM. Through extensive experiments, OpenRSS demonstrates its effective open-vocabulary semantic segmentation ability on RGB-T images. It outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: $+12.1\\%$ mIoU on the MFNet dataset, $+18.4\\%$ mIoU on the MCubeS dataset, and $+21.4\\%$ mIoU on the Freiburg Thermal dataset. Code will be released at https://github.com/SXDR/OpenRSS.", + "bbox": [ + 261, + 335, + 738, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open-vocabulary $\\cdot$ RGB-T semantic segmentation $\\cdot$ Prompt learning", + "bbox": [ + 264, + 681, + 738, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 737, + 374, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing deep learning-based semantic segmentation methods [22, 25, 41] mainly rely on single-modal RGB images, which are effective only under well-lit conditions during the daytime and perform poorly in challenging lighting conditions", + "bbox": [ + 217, + 770, + 785, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 220, + 824, + 380, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(e.g., nighttime, foggy weather). With the popularity of thermal cameras, researchers have recently started incorporating thermal images as a supplementary modality into RGB semantic segmentation [8], known as RGB-T semantic segmentation.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advancement of deep learning, current RGB-T semantic segmentation models mainly rely on supervised training with a predefined set of fixed categories, which cannot generalize to categories beyond the training set. This shortcoming significantly limits the application scenarios of current RGB-T semantic segmentation methods. In contrast, humans understand scenes in an open-vocabulary manner and recognize around 30,000 basic categories [1, 7]. Moreover, humans can further identify new categories through language descriptions.", + "bbox": [ + 212, + 210, + 784, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, large-scale vision-language models, represented by CLIP [26], achieve image-level classification across arbitrary categories, i.e., open-vocabulary classification. However, since the CLIP model is trained using image-level contrastive learning, how to effectively utilize CLIP in semantic segmentation is still an open problem. Some researchers adopt a two-stage training paradigm [5,20,37]: first, using a model to generate category-agnostic mask proposals, and then a pre-trained CLIP model for open-vocabulary classification on each mask. While others choose to train a side network [36] to integrate pre-trained CLIP model features into the side network for mask recognition. The promising performance of CLIP prompts us to explore its applicability in RGB-T semantic segmentation. But unlike the methods mentioned above, in this paper, we jointly utilize the basic segmentation model SAM [16] and CLIP through an elegant way to achieve open-vocabulary semantic segmentation, which effectively takes advantage of well-learned visual features and high segmentation performance of SAM and powerful open-vocabulary classification ability of CLIP.", + "bbox": [ + 212, + 335, + 787, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While SAM has learned a powerful and general visual segmentation model using the largest segmentation dataset to date, SA-1B, it is still unclear how to effectively transfer its knowledge from RGB images to RGB-T images. The most direct strategy is to train the SAM from scratch, specifically for RGB-T images. However, on the one hand, full model fine-tuning is time-consuming, inefficient, and burdensome regarding parameter storage. On the other hand, with the limited quantity of RGB-T semantic segmentation datasets, it is easy to produce over-fitting results. Thus, a natural question arises: is there a more efficient and effective method to enable SAM models trained on RGB images to be utilized for RGB-T images? Our answer to this question is affirmative, and we provide a solution with the paradigm of prompt learning. Prompt learning, also known as prompt tuning, originated from natural language processing (NLP) [17, 19], aims to inject text prompts into language models to transfer the knowledge of foundation models to various downstream tasks. Inspired by this, researchers apply prompt tuning to visual tasks [3,6,13,33,42,43,45] by freezing the entire foundation model and adding some learnable parameters only to the input side to learn effective visual prompts. Intuitively, there is a significant correlation between multimodal and RGB modalities, and they should share most of the", + "bbox": [ + 212, + 568, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prior knowledge about feature extraction. Following this, we propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation method.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike other RGB-T semantic segmentation methods that always need complex fusion networks to fuse RGB and thermal information, which inevitably adds considerable computation and training burden to their approaches. In OpenRSS, only a small amount of parameters are trainable, which makes the training of OpenRSS quite efficient and also decreases the risk of over-fitting. During the training, OpenRSS freezes the entire image encoder of SAM and only learns visual prompts based on RGB and thermal modality. Specifically, OpenRSS freezes the image encoder of SAM and adds several simple and efficient Thermal Information Prompt (TIP) modules to SAM to produce visual prompts to learn the complementarity between RGB and thermal modality effectively. Then, in the mask decoder of SAM, we align pixel embeddings with the text embeddings corresponding to the semantic classes generated by CLIP. Due to the flexible label representations provided by the text encoder, OpenRSS can perform open-vocabulary segmentation without requiring additional training samples. Additionally, we employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to fine-tune the image encoder of SAM. For fairness and reproducibility, our research is based on the officially released SAM and CLIP models. The main contributions of this paper are as follows:", + "bbox": [ + 212, + 183, + 787, + 455 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation model, where, with the aid of prompt learning and dynamic low-rank adaptation strategy, existing foundation models can effectively adapt to open-vocabulary RGB-T semantic segmentation.", + "- We introduce a Thermal Information Prompt (TIP) module that generates simple yet efficient visual prompts to fuse RGB and thermal information, avoiding the complex and redundant additional branches designed in previous methods.", + "- We employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to optimize a series of parameters in the image encoder of SAM.", + "- Extensive experiments demonstrate that OpenRSS outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: $+12.1\\%$ mIoU on the MFNet dataset, $+18.4\\%$ mIoU on the MCubeS dataset, and $+21.4\\%$ mIoU on the Freiburg Thermal dataset, while maintaining parameter efficiency (only $1\\%$ trainable parameters)." + ], + "bbox": [ + 225, + 489, + 784, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The remainder of this article is organized as follows. We review related work in Sec. 2. In Sec. 3, we provide a detailed description of our network architecture. Sec. 4 present experimental results and discussions. Finally, we summarize the work in Sec. 5.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 RGB-T Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 176, + 522, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For complex scenes with challenging lighting conditions, thermal images are highly complementary to RGB images. Existing RGB-T semantic segmentation methods focus on designing complex cross-modal fusion paradigms. We briefly review RGB-T semantic segmentation from three paradigms: encoder fusion, decoder fusion, and feature fusion. For the encoder fusion paradigm, RTFNet [28] and FuseSeg [29] integrate thermal features into RGB features by element-wise summation. FEANet [4] introduces a two-stage feature-enhanced attention network, which utilizes channel and spatial attention to fuse RGB and thermal information in a complementary way. For the decoder fusion paradigm, MFNet [8] uses two identical encoders to extract RGB and thermal features and feed them into the decoder for integration. For the feature fusion paradigm, EGFNet [44] embeds prior edge information into feature maps for cross-modal fusion. ABMDRNet [40] employs a bi-directional image-to-image translation based method to reduce the modality differences, followed by channel-wise weighted fusion for final prediction. CMX [38] first combines features from other modalities in both spatial and channel directions to calibrate the current modality's features, then fuses them for cross-modal fusion.", + "bbox": [ + 212, + 200, + 785, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike the methods mentioned above, which always have to train complex additional network branches to extract thermal-related features, OpenRSS only needs to learn a small amount of trainable parameters to inject thermal-related information into the image encoder of SAM, whose parameters are frozen during training. This merit gives OpenRSS high parameters and training efficiency.", + "bbox": [ + 212, + 458, + 785, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Large-scale Models", + "text_level": 1, + "bbox": [ + 215, + 556, + 419, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Transformer's robust scalability allows large-scale models to have hundreds of billions of parameters. Large-scale models, such as BERT [14], Instruct-GPT [24], GPT-4 [23], etc, achieve groundbreaking progress in natural language processing tasks. Large-scale models like SAM [16] and SegGPT [34] demonstrate outstanding zero-shot generalization capabilities in computer vision. Recently, large-scale vision-language models represented by CLIP [26] and ALIGN [12] can recognize arbitrary categories described in the text, known as open-vocabulary classification.", + "bbox": [ + 212, + 580, + 785, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work further explores how to effectively utilize pre-trained visual and visual-language large-scale models to perform open-vocabulary semantic segmentation on RGB-T images, and accordingly proposes our solution: the OpenRSS model.", + "bbox": [ + 212, + 700, + 785, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Visual Prompt Learning", + "text_level": 1, + "bbox": [ + 215, + 784, + 464, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In computer vision, downstream tasks are always executed by fine-tuning large pre-trained models comprehensively for a long time. Specifically, different down-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stream data are used to retrain downstream tasks to update all parameters. Although this approach yields satisfactory performance, it is inefficient regarding parameter usage and ultimately loses the universal knowledge learned by the pretrained models. Prompt learning (also known as prompt tuning) is recognized as a trade-off between the training burden and the final performance in the downstream tasks and has recently garnered widespread attention. It initially shone in natural language processing before being introduced to various computer vision tasks. For example, VPT [13] presents a set of learnable parameters for Transformer encoders and significantly outperforms full fine-tuning on multiple downstream tasks. ViPT [45] introduces prompts for pre-trained single-modal trackers, making them applicable to multimodal tracking tasks. UniSOD [33] designs a switchable prompt generation block to interact with single-modal or multimodal inputs.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In previous works, they often used simple $1 \\times 1$ or $3 \\times 3$ convolutions to generate prompts. In contrast, we specifically design a Thermal Information Prompt (TIP) module with a Transformer block inside to generate prompts for dense prediction, enabling the pre-trained SAM model to adapt to RGB-T images.", + "bbox": [ + 212, + 343, + 787, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.4 Open-vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 426, + 602, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Open-vocabulary semantic segmentation refers to identifying pixel-level categories in images through arbitrary text descriptions. In initial works like ZS3Net [2], generative models are used to synthesize pixel-level features through word embeddings of unseen classes. GroupVit [35] introduces a grouping mechanism to cluster images automatically under text supervision. With the remarkable performance of pre-trained large-scale vision-language models in open-vocabulary classification, researchers begin exploring their performance in open-vocabulary semantic segmentation. These approaches can be broadly categorized into the following three types: 1)LSeg [18] aligns pixel-level embeddings with corresponding semantic class text embeddings for pixel-level classification. 2) [5, 20, 37] adopt a two-stage segmentation framework, where MaskFormer first generates a set of category-agnostic mask proposals, followed by CLIP performing open-vocabulary classification on each mask. 3)SAN [36] designs a side network to enable CLIP to perceive the predictions and recognition of masks.", + "bbox": [ + 212, + 452, + 787, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike these methods, OpenRSS is an end-to-end network that elegantly utilizes SAM and CLIP to achieve open-vocabulary semantic segmentation.", + "bbox": [ + 214, + 662, + 785, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 715, + 331, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we present OpenRSS, an end-to-end framework that combines pretrained large-scale models SAM and CLIP to perform open-vocabulary RGB-T semantic segmentation. Instead of fully fine-tuning all the parameters of the models, OpenRSS only learns a small amount of prompt information. Furthermore, we introduce trainable dynamic LoRA bypasses to adjust the frozen image encoder.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4f45d9c8f5e3fc70233e6904cf555900af7a336c1cfbfbf4da333bdcad0341a.jpg", + "image_caption": [ + "Fig. 1: Overview architecture of our OpenRSS. It inputs a pair of RGB and thermal images and generates corresponding RGB and thermal tokens through patch embeddings. Then, it utilizes the frozen Transformer encoder blocks of SAM for feature extraction. The TIP modules are inserted into the image encoder of SAM to learn effective visual prompts. DLoRA further adjusts the frozen image encoder. Finally, the mask decoder aligns image and text embeddings to generate the final segmentation map." + ], + "image_footnote": [], + "bbox": [ + 223, + 148, + 782, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 558, + 339, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall architecture of our OpenRSS is shown in Fig. 1. Formally, OpenRSS incorporates an additional thermal image input stream instead of a thermal feature extraction branch, which is fed into the network simultaneously with the RGB stream. Given a pair of RGB and thermal images, OpenRSS first projects the two information streams $(T_{RGB}$ and $T_{TH})$ into patch embedding layers separately. After projection, we obtain two C-dimensional embeddings, which we denote as RGB tokens $\\pmb{I}_{rgb}$ and thermal tokens $\\pmb{I}_{th}$ , respectively. Then, $\\pmb{I}_{rgb}$ and $\\pmb{I}_{th}$ are input into the TIP module to generate thermal information prompts. The learned prompt information is added to the original RGB information stream $\\pmb{I}_{rgb}$ in a residual manner before being fed into the fully frozen image encoder of SAM:", + "bbox": [ + 212, + 585, + 787, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} ^ {i} = \\boldsymbol {I} _ {r g b} ^ {i} + \\boldsymbol {T} ^ {i + 1}, \\quad i = 0, 1, \\dots , L - 1, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 752, + 785, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{I}^i$ is the final sequence of image tokens fed into the Transformer layers of the encoder, and $\\pmb{T}^{i + 1}$ is the prompt tokens sequence from the $(i + 1)$ th TIP module. By directly adding the thermal information prompts to the intermediate features generated by SAM, our OpenRSS can effortlessly perform downstream", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tasks quickly. Furthermore, we also design a trainable dynamic LoRA bypass for each Transformer layer. For each bypass, the features generated by the Transformer layers are first mapped to a low-rank space and then projected back to the initial dimension, aligning it with the original output dimension of the frozen Transformer layers. We use a points grid prompt for the prompt encoder, significantly enhancing SAM's attention to every detail in the image compared to the default embeddings. For the text encoder, we use the frozen text encoder of CLIP. During training, we employ manually crafted prompts for the text encoder, consisting of a natural sentence and category names. However, any sentence can be used during inference to achieve the goal of open-vocabulary. We modify the original mask decoder of SAM to align its dimensionality with the feature embeddings of the text encoder. It simultaneously takes the embeddings from the prompt encoder and the text encoder as input, generating the final segmentation map.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Thermal Information Prompt", + "text_level": 1, + "bbox": [ + 215, + 378, + 504, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In recent years, following the success of text prompts in NLP, some research has begun to explore the introduction of visual prompts into frozen pre-trained models. Prompt tuning is achieved through some learnable parameters and yields promising results on many downstream visual tasks. A key to our OpenRSS is to generate appropriate thermal information prompts to guide the network in effective open-vocabulary RGB-T semantic segmentation. As shown in Fig. 1, the TIP module is inserted before each Transformer layer. Formally, the thermal information prompts are generated from the token sequences $\\{I_{rgb}^{0}, I_{th}^{0}\\}$ of both modalities. The TIP module learns the complementarity between these two modalities, which can be represented as follows:", + "bbox": [ + 212, + 401, + 787, + 551 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} ^ {i} = T I P ^ {i} \\left(\\boldsymbol {I} ^ {i - 1}, \\boldsymbol {T} ^ {i - 1}\\right), \\quad i = 1, 2, \\dots , L, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 560, + 784, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $TIP^i$ denotes the $i$ th TIP module. Specifically, $I^0 = I_{rqb}^0$ and $T^0 = I_{th}^0$ .", + "bbox": [ + 212, + 587, + 766, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The detailed design of the TIP is shown in Fig. 2. TIP has two input branches, each receiving token sequences from two input streams. Specifically, considering the redundancy of single-modal features and the lightness of the prompt block, we use two $1 \\times 1$ convolution operations to project each input stream to the size of $[ \\alpha \\times H \\times W ]$ by:", + "bbox": [ + 212, + 604, + 787, + 679 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} _ {R G B} = c _ {1} \\left(\\boldsymbol {I} ^ {i - 1}\\right), \\quad \\boldsymbol {Z} _ {T H} = c _ {2} \\left(\\boldsymbol {T} ^ {i - 1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 688, + 784, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ is set to 8 in all TIP modules. $c_{1}(\\cdot)$ and $c_{2}(\\cdot)$ are $1\\times 1$ convolution operations. Then, inspired by [32], we use a Transformer block with correlative self-attention for dense prediction to generate enhanced embeddings for $\\mathbf{Z}_{RGB}^{e}$ , as shown in Fig. 2. The formula for Correlative_attn is expressed as:", + "bbox": [ + 212, + 713, + 787, + 773 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {C o r r e l a t i v e} \\_ \\text {a t t n} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {X} \\boldsymbol {W} _ {q} \\boldsymbol {W} _ {q} ^ {T} \\boldsymbol {X} ^ {T} / \\tau\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 782, + 784, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{X}$ denotes the input and $\\mathbf{W}_q$ is the projection matrix. The temperature coefficient $\\tau$ is by default set to $\\sqrt{d}$ following traditional self-attention. Finally,", + "bbox": [ + 212, + 806, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 352, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00ed787fe22808959742094ec4063222c64a618e11002e7c4a6bc6edd89762d5.jpg", + "image_caption": [ + "Fig. 2: Overview architecture of our TIP module. Multi-modal input flows are sent separately to the $1 \\times 1$ convolutions for distilling to the lower-dimension latent space. After passing through the Correlative Self-Attention block, the RGB embeddings are fused with thermal embeddings through element-wise addition. Finally, the multi-modal hybrid embeddings are mapped to the original dimension through a $1 \\times 1$ convolutional layer." + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 774, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we obtain the mixed-modal embeddings by element-wise addition and project them back to the original dimension by:", + "bbox": [ + 212, + 482, + 787, + 513 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} ^ {i} = c _ {3} \\left(\\boldsymbol {Z} _ {R G B} ^ {e} + \\boldsymbol {Z} _ {T H}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 521, + 785, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $c_{3}(\\cdot)$ is the same as $c_{1}(\\cdot)$ and $c_{2}(\\cdot)$ .", + "bbox": [ + 214, + 546, + 516, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 DLoRA in Image Encoder", + "text_level": 1, + "bbox": [ + 214, + 580, + 478, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In OpenRSS, the thermal information prompts generated by TIP are fed into the frozen pre-trained image encoder of SAM to extract features. However, this may not be optimal due to the differences between the RGB and thermal modalities. Inspired by the idea from [10, 30], we design a Dynamic Low-Rank Adaptation (DLoRA) strategy to adjust the frozen image encoder of SAM to address this issue.", + "bbox": [ + 212, + 604, + 787, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given an input token sequence $\\mathbf{X} \\in \\mathbb{R}^{B \\times N \\times C_{in}}$ , the output token sequence $\\mathbf{X}' \\in \\mathbb{R}^{B \\times N \\times C_{out}}$ is obtained through a projection layer $\\mathbf{W} \\in \\mathbb{R}^{C_{in} \\times C_{out}}$ . LoRA assumes the updates to the weights have a low \"intrinsic rank\" during adaptation, thus proposing a low-rank decomposition to constrain its updates. Following this strategy, the Transformer layer is first frozen to fix $\\mathbf{W}$ , and then a bypass is added to achieve low-rank decomposition. This bypass consists of two linear layers $\\mathbf{A} \\in \\mathbb{R}^{C_{in} \\times r}$ and $\\mathbf{B} \\in \\mathbb{R}^{r \\times C_{out}}$ , where $r \\ll \\min(C_{in}, C_{out})$ . Therefore, the update of $\\mathbf{W}$ can be described as:", + "bbox": [ + 212, + 693, + 789, + 815 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {W}} = \\boldsymbol {W} + \\Delta \\boldsymbol {W} = \\boldsymbol {W} + \\boldsymbol {B} \\boldsymbol {A}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 821, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fd25287071bda43007e633aad1ab66e0bbaf4f3f3db226bcd14f508a0793673.jpg", + "image_caption": [ + "Fig. 3: Overview architecture of our DLoRA. We apply the DLoRA layer to the q and v projection layers of each of the Transformer layers in the image encoder of SAM." + ], + "image_footnote": [], + "bbox": [ + 292, + 148, + 707, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We maintain consistency with [39] and apply LoRA to the query and value projection layers of the multi-head self-attention. Therefore, the modified query and value projection layers will become:", + "bbox": [ + 214, + 392, + 782, + 438 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Q} = \\hat {\\boldsymbol {W}} _ {q} \\boldsymbol {X} = \\boldsymbol {W} _ {q} \\boldsymbol {X} + \\boldsymbol {B} _ {q} \\boldsymbol {A} _ {q} \\boldsymbol {X}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 448, + 785, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_q$ and $\\mathbf{W}_v$ are frozen from image encoder of SAM, and $\\mathbf{A}_q, \\mathbf{B}_q, \\mathbf{A}_v$ and $\\mathbf{B}_v$ are trainable parameters.", + "bbox": [ + 214, + 492, + 782, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While LoRA achieves remarkable results, it heavily relies on the choice of rank $r$ , requiring a manual and iterative search for the optimal rank. Therefore, we propose a Dynamic LoRA (DLoRA). Building upon the original LoRA, we increase the number of bypasses (as shown in Fig. 3) from 1 to 4, each with different ranks $\\gamma$ and weights $\\beta$ . Through experimental observation, we find that our DLoRA can surpass the optimal rank found through manual iterative training in just one training.", + "bbox": [ + 212, + 522, + 784, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Generate Segmentation Map", + "text_level": 1, + "bbox": [ + 214, + 650, + 498, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Prompt Encoder and Mask Decoder. SAM passes the image embeddings generated by its image encoder and the prompts generated by its prompt encoder to its mask decoder for dense prediction. The types of prompts include point, box, and mask. SAM represents the first two sparse prompts as positional encodings and feeds them into the Transformer layers of the mask decoder. SAM performs convolution operations for dense prompts such as masks and then adds them element-wise with the image embeddings.", + "bbox": [ + 212, + 672, + 784, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Due to SAM's prompt mechanism, it is susceptible to the position and quantity of prompt points. Therefore, in OpenRSS, we use a points grid as prompts to ensure that SAM does not overlook any detail in the image. SAM's mask decoder comprises a lightweight Transformer layer and a segmentation head. OpenRSS", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "modifies this segmentation head to align the dimension of the feature map with the dimension of CLIP's text features, enabling pixel-level classification.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pixel-level Classification via CLIP. Existing CLIP-based open-vocabulary segmentation methods mainly adopt a two-stage region recognition approach. In the first stage, an independent model is trained to generate a set of category-agnostic mask proposals. In the second stage, visual language pre-trained models such as CLIP are used to identify the category of each mask individually. However, this two-stage approach heavily relies on the quality of the masks generated in the first stage. Moreover, the cropped masks, due to missing backgrounds and varying sizes, are not suitable for recognition by CLIP.", + "bbox": [ + 212, + 204, + 782, + 323 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To address the problem of region recognition, our OpenRSS adopts pixel-level recognition. We define the feature map outputted by the mask decoder as $\\pmb{F} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times D}$ , and the class embeddings generated by the CLIP text encoder as $\\pmb{E} \\in \\mathbb{R}^{N \\times D}$ . Since they have the same embedding dimension, we can directly link them together through the inner product to obtain the final segmentation map:", + "bbox": [ + 212, + 325, + 782, + 412 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} = \\boldsymbol {F} \\times \\boldsymbol {E} ^ {T}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 454, + 419, + 782, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\pmb{S} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times N}$ . This is the standard output of semantic segmentation, hence compatible with mainstream semantic segmentation evaluations.", + "bbox": [ + 212, + 445, + 782, + 478 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 506, + 375, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Dataset and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 542, + 522, + 556 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct experiments on 4 datasets: MF dataset [9], KP dataset [11], MCubeS dataset [21] and Freiburg Thermal dataset [31]. All models are trained on the training set of the KP dataset and evaluated on other datasets.", + "bbox": [ + 212, + 571, + 782, + 617 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MF Dataset. The dataset consists of urban street scenes captured using an InfRec R500 camera. It includes 1569 pairs of RGB and thermal images, each with a resolution of $480 \\times 640$ pixels. Among these are 820 pairs of daytime images and 749 pairs of nighttime images. The dataset consists of eight manually labeled classes, as well as a background class.", + "bbox": [ + 212, + 645, + 782, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "KP Dataset. The dataset is an RGB-T paired urban driving scene dataset, providing 95328 pairs of RGB and thermal images with a resolution of $512 \\times 640$ . Initially, the KP dataset only provides bounding box labels for object detection tasks. However, fortunately, [15] provides semantic segmentation task labels for 503 daytime and 447 nighttime images. Furthermore, [27] partitions these 950 images into training, validation, and test sets.", + "bbox": [ + 212, + 748, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MCubeS Dataset. MCubeS is a dataset with pairs of RGB, Near-Infrared (NIR), Degree of Linear Polarization (DoLP), and Angle of Linear Polarization (AoLP) to study semantic material segmentation of 20 classes with a resolution of $1024 \\times 1224$ .", + "bbox": [ + 214, + 146, + 782, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Freiburg Thermal Dataset. The dataset includes 12051 daytime and 8596 nighttime time-synchronized RGB-T image pairs captured in rural and urban environments with a resolution of $650 \\times 1920$ . However, only the test set includes 32 daytime and 32 nighttime images, comprising 12 object classes and one background class.", + "bbox": [ + 214, + 226, + 782, + 301 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use mean Accuracy (mAcc) and mean Intersection over Union (mIoU) as the evaluation metrics for our model. The formulas for calculation are as follows:", + "bbox": [ + 214, + 321, + 782, + 364 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nm A c c = \\frac {1}{n _ {c l s}} \\cdot \\sum_ {i} \\frac {n _ {i i}}{t _ {i}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 371, + 784, + 405 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nm I o U = \\frac {1}{n _ {c l s}} \\cdot \\sum_ {i} \\frac {n _ {i i}}{t _ {i} + \\sum_ {j} n _ {j i} - n _ {i i}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 415, + 784, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $n_{ii}$ is the number of pixels where class i is predicted as class j, $n_{cls}$ is the total number of target classes (including the background), and $t_i = \\sum_j n_{ij}$ is the total number of pixels for target class i (ground truth labels).", + "bbox": [ + 214, + 459, + 782, + 505 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 523, + 455, + 539 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct all experiments on NVIDIA A100 GPUs. If not specified, we default to using SAM with the ViT-H backbone and CLIP with the ViT-L/16 backbone. We apply DLoRA fine-tuning to the frozen encoder Transformer layers, using four bypasses with ranks selected as 1, 2, 4, and 8, and weights as 0.1, 0.2, 0.4, and 0.3, respectively. All models are trained on the KP dataset with image flipping as a data augmentation technique. We use the AdamW optimizer with a weight decay of 0.01 and an initial learning rate of 5e-4. For the scheduler, we use the WarmupPolyLR scheduler. The loss function consists of DiceLoss and SoftCrossEntropy with weighting coefficients of 0.5 each:", + "bbox": [ + 214, + 546, + 782, + 681 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nD i c e L o s s = 1 - \\frac {2 \\sum_ {i = 1} ^ {N} y _ {i} \\widehat {y} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} \\widehat {y} _ {i} ^ {2}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 691, + 784, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $y_{i}$ and $\\widehat{y_i}$ represent the true and predicted label values of pixel I, respectively, and $N$ represents the total number of pixels.", + "bbox": [ + 214, + 765, + 782, + 795 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S o f t C r o s s E n t r o p y l o s s} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {c} \\hat {y} _ {i j} \\log \\left(y _ {i j} ^ {d}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 804, + 782, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3a9aa0ef160e807d432e243c5a4e85f9716dbd4415352a052a4da27dd48bd2da.jpg", + "table_caption": [ + "Table 1: Performance comparison with state-of-the-art methods. For a fair comparison, we reproduce their method with their officially released code. The bold font highlights the best result in each column." + ], + "table_footnote": [], + "table_body": "
MethodBackboneMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
ZegFormer [5]R-5011.1110.3110.003.437.690.11
OVSeg [20]R-10110.869.4210.092.738.320.98
SAN [36]ViT-B14.644.1317.2410.107.131.31
ZegFormer [5]R-10111.5510.7110.464.117.850.79
OVSeg [20]Swin-B10.206.7816.496.5810.203.74
SAN [36]ViT-L3.810.386.904.2710.606.10
OpenRSS(ours)SAM-B22.2118.9631.7620.2716.018.14
SAM-L23.9120.0833.9524.9125.3316.65
SAM-H30.7322.8137.8828.5837.2527.55
", + "bbox": [ + 243, + 198, + 761, + 435 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where the $n$ is the number of the batch, $\\hat{y}_{ij}$ is a binary indicator if class label $c$ is the correct classification for pixel $i$ , and $y_{ij}^{d}$ is the corresponding predicted probability be normalized to a probability distribution. These two losses are summed to obtain the overall loss function:", + "bbox": [ + 212, + 469, + 787, + 530 + ], + "page_idx": 11 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = 0. 5 L _ {\\text {D i c e L o s s}} + 0. 5 L _ {\\text {S o f t C r o s s E n t r o p y l o s s}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 549, + 785, + 566 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Results and Analysis", + "text_level": 1, + "bbox": [ + 215, + 602, + 434, + 618 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 1 presents the quantitative comparison results of our method with other state-of-the-art open-vocabulary semantic segmentation methods. We only use RGB images as input, rather than pairs of RGB and thermal images, allowing the general open vocabulary semantic segmentation in Tab. 1 to perform RGB-T semantic segmentation. From the table, it can be observed that our method achieves superior performance on both metrics across all three datasets, indicating the effectiveness of our approach. Specifically, our method surpasses other methods with an average of $+9.17\\%$ mAcc and $+6.82\\%$ mIoU for SAM-B, $+13.57\\%$ mAcc and $+11.58\\%$ mIoU for SAM-L, and $+20.46\\%$ mAcc and $+17.34\\%$ mIoU for SAM-H. This further validates that open-vocabulary semantic segmentation methods trained on RGB images are unsuitable for RGB-T images. Furthermore, we present the qualitative experimental results of OpenRSS on the Freiburg Thermal dataset in Fig. 4.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b5a36f6a248b47730193dacdee5984d2b242f73af4118f603193569a831f7ebc.jpg", + "image_caption": [ + "Fig. 4: Qualitative results of OpenRSS on Freiburg Thermal dataset. From left to right: RGB images, Thermal images, ground-truths, OpenRSS results." + ], + "image_footnote": [], + "bbox": [ + 243, + 148, + 763, + 429 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/167ba4482994e1f231b93d2c4bd86e5b7439ea8989cd0646e4bcabcc3ac3629f.jpg", + "table_caption": [ + "Table 2: Ablation studies of TIP module Table 3: Ablation studies of attn type in OpenRSS. in TIP module." + ], + "table_footnote": [], + "table_body": "
MethodMFMSFT
mIoU↑
TIP-empty10.2016.213.90
TIP-shallow19.9826.8325.44
TIP-deep(ours)22.8128.5827.55
attn-fovea18.4523.0118.58
attn-msa22.5421.7221.56
attn-cor(ours)22.8128.5827.55
", + "bbox": [ + 240, + 554, + 782, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 686, + 388, + 704 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Importance of TIP. Tab. 2 shows the ablation experiments of the proposed TIP module. To validate the effectiveness of the TIP module, we replace it with additive fusion, resulting in an average mIoU performance drop of $16.21\\%$ . Furthermore, we discuss the specific placement of the TIP module. We only position the TIP module before the first Transformer layer, called TIP-shallow. The results indicate that placing the TIP module at each Transformer layer yields the best performance. Finally, we conduct ablation experiments on the correlative self-attention within the TIP module. As shown in Tab. 3, the results demon", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "strate the effectiveness of correlative self-attention for dense tasks by comparing it to spatial fovea attention and the original multi-head self-attention.", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Importance of DLoRA. Tab. 4 shows the ablation experiments of the proposed DLoRA. By manually setting the ranks to 1, 2, 4, and 8, the performance degradation of average mIoU is observed to be $4.69\\%$ , $2.55\\%$ , $3.03\\%$ , and $1.36\\%$ , respectively. This demonstrates the significance of DLoRA while also avoiding spending considerable time selecting an optimal low-rank space. More detailed results and analysis, including failure cases, are discussed in the supplementary material.", + "bbox": [ + 212, + 200, + 787, + 305 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/56c1b9f56524ee7da6515364734567104cf2010d8024d2b0111e09c26ddc2e76.jpg", + "table_caption": [ + "Table 4: Ablation of rank in DLoRA. Manually set rank vs. dynamically adaptive rank. The bold font highlights the best result in each column." + ], + "table_footnote": [], + "table_body": "
MethodsMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
OpenRSS-rank-129.2020.6137.4324.6928.4319.57
OpenRSS-rank-226.6121.7835.7725.9832.6023.51
OpenRSS-rank-425.3821.6336.5327.7329.6420.49
OpenRSS-rank-827.2622.5137.0726.2835.2326.06
OpenRSS-DLoRA(ours)30.7322.8137.8828.5837.2527.55
", + "bbox": [ + 251, + 371, + 750, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 580, + 359, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work proposes the first open-vocabulary RGB-T semantic segmentation framework: OpenRSS. Our framework achieves the open-vocabulary capability by jointly utilizing the CLIP and the modified SAM model. In addition, when fusing RGB and thermal modality information, OpenRSS only needs to add a few learnable parameters to SAM to achieve effective multi-modal information fusion. Extensive experiments show that the proposed framework significantly outperforms previous methods on three RGB-T semantic segmentation benchmarks, validating the effectiveness and generalization of OpenRSS. Our work provides a feasible solution to open-vocabulary semantic segmentation with RGB-T images and a strong baseline for the community to advance future research.", + "bbox": [ + 212, + 612, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grant 62203337; in part by the Hubei Province Natural Science Foundation, China, under Grant 2022CFC074; and in part by the Open Fund of the Hubei Garment Informatization Engineering Technology Research Center, Wuhan Textile University, under Grant 2022HBCI04.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987)", + "2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Advances in Neural Information Processing Systems 32 (2019)", + "3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)", + "4. Deng, F., Feng, H., Liang, M., Wang, H., Yang, Y., Gao, Y., Chen, J., Hu, J., Guo, X., Lam, T.L.: Fanet: Feature-enhanced attention network for rgb-thermal real-time semantic segmentation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4467-4473. IEEE (2021)", + "5. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022)", + "6. Dong, S., Feng, Y., Yang, Q., Huang, Y., Liu, D., Fan, H.: Efficient multimodal semantic segmentation via dual-prompt learning. arXiv preprint arXiv:2312.00360 (2023)", + "7. Fu, Y., Xiang, T., Jiang, Y.G., Xue, X., Sigal, L., Gong, S.: Recent advances in zero-shot recognition. arXiv preprint arXiv:1710.04837 (2017)", + "8. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)", + "9. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)", + "10. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "1. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1037-1045 (2015)", + "2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "3. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)", + "4. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)", + "5. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497-6504 (2021)", + "6. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)", + "18. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022)", + "19. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)", + "20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)", + "21. Liang, Y., Wakaki, R., Nobuhara, S., Nishino, K.: Multimodal material segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19800-19808 (2022)", + "22. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)", + "23. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article 2, 13 (2023)", + "24. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)", + "25. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016)", + "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "27. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. arXiv preprint arXiv:2303.17386 (2023)", + "28. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019)", + "29. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2020)", + "30. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 (2022)", + "31. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020)", + "32. Wang, F., Mei, J., Yuille, A.: Sclip: Rethinking self-attention for dense vision-language inference. arXiv preprint arXiv:2312.01597 (2023)", + "33. Wang, K., Li, C., Tu, Z., Luo, B.: Unified-modal salient object detection via adaptive prompt learning. arXiv preprint arXiv:2311.16835 (2023)", + "34. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Zhao, J. Huang, X. Yan, et al.", + "bbox": [ + 271, + 114, + 495, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)", + "36. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)", + "37. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022)", + "38. Zhang, J., Liu, H., Yang, K., Hu, X., Liu, R., Stiefelhagen, R.: Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems (2023)", + "39. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)", + "40. Zhang, Q., Zhao, S., Luo, Y., Zhang, D., Huang, N., Han, J.: Abmdrnet: Adaptive-weighted bi-directional modality difference reduction network for rgb-t semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2633-2642 (2021)", + "41. Zhao, H., Qi, X., Shen, X., Shi, J., Jia, J.: Icnet for real-time semantic segmentation on high-resolution images. In: Proceedings of the European conference on computer vision (ECCV). pp. 405-420 (2018)", + "42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)", + "43. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "44. Zhou, W., Dong, S., Xu, C., Qian, Y.: Edge-aware guidance fusion network for rgb-thermal scene parsing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3571-3579 (2022)", + "45. Zhu, J., Lai, S., Chen, X., Wang, D., Lu, H.: Visual prompt multi-modal tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9516-9526 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 604 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Open-Vocabulary RGB-Thermal Semantic Segmentation", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_model.json b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f272abebe3f1d8394536d11368cb33ea5757e145 --- /dev/null +++ b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_model.json @@ -0,0 +1,2148 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.261, + 0.141, + 0.743, + 0.186 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.212, + 0.78, + 0.244 + ], + "angle": 0, + "content": "Guoqiang Zhao\\(^{1}\\), Junjie Huang\\(^{1}\\), Xiaoyun Yan\\(^{1,2\\dagger}\\), Zhaojing Wang\\(^{1,2}\\), Junwei Tang\\(^{1}\\), Yangjun Ou\\(^{1}\\), Xinrong Hu\\(^{1,2}\\), and Tao Peng\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.255, + 0.777, + 0.282 + ], + "angle": 0, + "content": "1 School of Computer Science and Artificial Intelligence, Wuhan Textile University zgq_lyy@163.com, xyyan@wtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.283, + 0.751, + 0.296 + ], + "angle": 0, + "content": "\\(^{2}\\) Hubei Garment Informatization Engineering Technology Research Center" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.337, + 0.739, + 0.668 + ], + "angle": 0, + "content": "Abstract. RGB-Thermal (RGB-T) semantic segmentation is an important research branch of multi-modal image segmentation. The current RGB-T semantic segmentation methods generally have two unsolved and typical shortcomings. First, they do not have the open-vocabulary recognition ability, which significantly limits their application scenarios. Second, when fusing RGB and thermal images, they often need to design complex fusion network structures, which usually results in low network training efficiency. We present OpenRSS, the Open-vocabulary RGB-T Semantic Segmentation method, to solve these two disadvantages. To our knowledge, OpenRSS is the first RGB-T semantic segmentation method with open-vocabulary segmentation capability. OpenRSS modifies the basic segmentation model SAM for RGB-T semantic segmentation by adding the proposed thermal information prompt module and dynamic low-rank adaptation strategy to SAM. These designs effectively fuse the RGB and thermal information, but with much fewer trainable parameters than other methods. OpenRSS achieves the open-vocabulary capability by jointly utilizing the vision-language model CLIP and the modified SAM. Through extensive experiments, OpenRSS demonstrates its effective open-vocabulary semantic segmentation ability on RGB-T images. It outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: \\(+12.1\\%\\) mIoU on the MFNet dataset, \\(+18.4\\%\\) mIoU on the MCubeS dataset, and \\(+21.4\\%\\) mIoU on the Freiburg Thermal dataset. Code will be released at https://github.com/SXDR/OpenRSS." + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.683, + 0.739, + 0.71 + ], + "angle": 0, + "content": "Keywords: Open-vocabulary \\(\\cdot\\) RGB-T semantic segmentation \\(\\cdot\\) Prompt learning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.738, + 0.375, + 0.753 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.771, + 0.786, + 0.815 + ], + "angle": 0, + "content": "Existing deep learning-based semantic segmentation methods [22, 25, 41] mainly rely on single-modal RGB images, which are effective only under well-lit conditions during the daytime and perform poorly in challenging lighting conditions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.382, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "(e.g., nighttime, foggy weather). With the popularity of thermal cameras, researchers have recently started incorporating thermal images as a supplementary modality into RGB semantic segmentation [8], known as RGB-T semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.785, + 0.332 + ], + "angle": 0, + "content": "With the advancement of deep learning, current RGB-T semantic segmentation models mainly rely on supervised training with a predefined set of fixed categories, which cannot generalize to categories beyond the training set. This shortcoming significantly limits the application scenarios of current RGB-T semantic segmentation methods. In contrast, humans understand scenes in an open-vocabulary manner and recognize around 30,000 basic categories [1, 7]. Moreover, humans can further identify new categories through language descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.789, + 0.563 + ], + "angle": 0, + "content": "Recently, large-scale vision-language models, represented by CLIP [26], achieve image-level classification across arbitrary categories, i.e., open-vocabulary classification. However, since the CLIP model is trained using image-level contrastive learning, how to effectively utilize CLIP in semantic segmentation is still an open problem. Some researchers adopt a two-stage training paradigm [5,20,37]: first, using a model to generate category-agnostic mask proposals, and then a pre-trained CLIP model for open-vocabulary classification on each mask. While others choose to train a side network [36] to integrate pre-trained CLIP model features into the side network for mask recognition. The promising performance of CLIP prompts us to explore its applicability in RGB-T semantic segmentation. But unlike the methods mentioned above, in this paper, we jointly utilize the basic segmentation model SAM [16] and CLIP through an elegant way to achieve open-vocabulary semantic segmentation, which effectively takes advantage of well-learned visual features and high segmentation performance of SAM and powerful open-vocabulary classification ability of CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.84 + ], + "angle": 0, + "content": "While SAM has learned a powerful and general visual segmentation model using the largest segmentation dataset to date, SA-1B, it is still unclear how to effectively transfer its knowledge from RGB images to RGB-T images. The most direct strategy is to train the SAM from scratch, specifically for RGB-T images. However, on the one hand, full model fine-tuning is time-consuming, inefficient, and burdensome regarding parameter storage. On the other hand, with the limited quantity of RGB-T semantic segmentation datasets, it is easy to produce over-fitting results. Thus, a natural question arises: is there a more efficient and effective method to enable SAM models trained on RGB images to be utilized for RGB-T images? Our answer to this question is affirmative, and we provide a solution with the paradigm of prompt learning. Prompt learning, also known as prompt tuning, originated from natural language processing (NLP) [17, 19], aims to inject text prompts into language models to transfer the knowledge of foundation models to various downstream tasks. Inspired by this, researchers apply prompt tuning to visual tasks [3,6,13,33,42,43,45] by freezing the entire foundation model and adding some learnable parameters only to the input side to learn effective visual prompts. Intuitively, there is a significant correlation between multimodal and RGB modalities, and they should share most of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "prior knowledge about feature extraction. Following this, we propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Unlike other RGB-T semantic segmentation methods that always need complex fusion networks to fuse RGB and thermal information, which inevitably adds considerable computation and training burden to their approaches. In OpenRSS, only a small amount of parameters are trainable, which makes the training of OpenRSS quite efficient and also decreases the risk of over-fitting. During the training, OpenRSS freezes the entire image encoder of SAM and only learns visual prompts based on RGB and thermal modality. Specifically, OpenRSS freezes the image encoder of SAM and adds several simple and efficient Thermal Information Prompt (TIP) modules to SAM to produce visual prompts to learn the complementarity between RGB and thermal modality effectively. Then, in the mask decoder of SAM, we align pixel embeddings with the text embeddings corresponding to the semantic classes generated by CLIP. Due to the flexible label representations provided by the text encoder, OpenRSS can perform open-vocabulary segmentation without requiring additional training samples. Additionally, we employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to fine-tune the image encoder of SAM. For fairness and reproducibility, our research is based on the officially released SAM and CLIP models. The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.49, + 0.784, + 0.55 + ], + "angle": 0, + "content": "- We propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation model, where, with the aid of prompt learning and dynamic low-rank adaptation strategy, existing foundation models can effectively adapt to open-vocabulary RGB-T semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.557, + 0.785, + 0.616 + ], + "angle": 0, + "content": "- We introduce a Thermal Information Prompt (TIP) module that generates simple yet efficient visual prompts to fuse RGB and thermal information, avoiding the complex and redundant additional branches designed in previous methods." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.624, + 0.784, + 0.653 + ], + "angle": 0, + "content": "- We employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to optimize a series of parameters in the image encoder of SAM." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.75 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that OpenRSS outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: \\(+12.1\\%\\) mIoU on the MFNet dataset, \\(+18.4\\%\\) mIoU on the MCubeS dataset, and \\(+21.4\\%\\) mIoU on the Freiburg Thermal dataset, while maintaining parameter efficiency (only \\(1\\%\\) trainable parameters)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The remainder of this article is organized as follows. We review related work in Sec. 2. In Sec. 3, we provide a detailed description of our network architecture. Sec. 4 present experimental results and discussions. Finally, we summarize the work in Sec. 5." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.523, + 0.192 + ], + "angle": 0, + "content": "2.1 RGB-T Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.787, + 0.457 + ], + "angle": 0, + "content": "For complex scenes with challenging lighting conditions, thermal images are highly complementary to RGB images. Existing RGB-T semantic segmentation methods focus on designing complex cross-modal fusion paradigms. We briefly review RGB-T semantic segmentation from three paradigms: encoder fusion, decoder fusion, and feature fusion. For the encoder fusion paradigm, RTFNet [28] and FuseSeg [29] integrate thermal features into RGB features by element-wise summation. FEANet [4] introduces a two-stage feature-enhanced attention network, which utilizes channel and spatial attention to fuse RGB and thermal information in a complementary way. For the decoder fusion paradigm, MFNet [8] uses two identical encoders to extract RGB and thermal features and feed them into the decoder for integration. For the feature fusion paradigm, EGFNet [44] embeds prior edge information into feature maps for cross-modal fusion. ABMDRNet [40] employs a bi-directional image-to-image translation based method to reduce the modality differences, followed by channel-wise weighted fusion for final prediction. CMX [38] first combines features from other modalities in both spatial and channel directions to calibrate the current modality's features, then fuses them for cross-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.787, + 0.535 + ], + "angle": 0, + "content": "Unlike the methods mentioned above, which always have to train complex additional network branches to extract thermal-related features, OpenRSS only needs to learn a small amount of trainable parameters to inject thermal-related information into the image encoder of SAM, whose parameters are frozen during training. This merit gives OpenRSS high parameters and training efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.421, + 0.571 + ], + "angle": 0, + "content": "2.2 Large-scale Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.701 + ], + "angle": 0, + "content": "The Transformer's robust scalability allows large-scale models to have hundreds of billions of parameters. Large-scale models, such as BERT [14], Instruct-GPT [24], GPT-4 [23], etc, achieve groundbreaking progress in natural language processing tasks. Large-scale models like SAM [16] and SegGPT [34] demonstrate outstanding zero-shot generalization capabilities in computer vision. Recently, large-scale vision-language models represented by CLIP [26] and ALIGN [12] can recognize arbitrary categories described in the text, known as open-vocabulary classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Our work further explores how to effectively utilize pre-trained visual and visual-language large-scale models to perform open-vocabulary semantic segmentation on RGB-T images, and accordingly proposes our solution: the OpenRSS model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.465, + 0.801 + ], + "angle": 0, + "content": "2.3 Visual Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In computer vision, downstream tasks are always executed by fine-tuning large pre-trained models comprehensively for a long time. Specifically, different down-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "stream data are used to retrain downstream tasks to update all parameters. Although this approach yields satisfactory performance, it is inefficient regarding parameter usage and ultimately loses the universal knowledge learned by the pretrained models. Prompt learning (also known as prompt tuning) is recognized as a trade-off between the training burden and the final performance in the downstream tasks and has recently garnered widespread attention. It initially shone in natural language processing before being introduced to various computer vision tasks. For example, VPT [13] presents a set of learnable parameters for Transformer encoders and significantly outperforms full fine-tuning on multiple downstream tasks. ViPT [45] introduces prompts for pre-trained single-modal trackers, making them applicable to multimodal tracking tasks. UniSOD [33] designs a switchable prompt generation block to interact with single-modal or multimodal inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.406 + ], + "angle": 0, + "content": "In previous works, they often used simple \\(1 \\times 1\\) or \\(3 \\times 3\\) convolutions to generate prompts. In contrast, we specifically design a Thermal Information Prompt (TIP) module with a Transformer block inside to generate prompts for dense prediction, enabling the pre-trained SAM model to adapt to RGB-T images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.427, + 0.603, + 0.442 + ], + "angle": 0, + "content": "2.4 Open-vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Open-vocabulary semantic segmentation refers to identifying pixel-level categories in images through arbitrary text descriptions. In initial works like ZS3Net [2], generative models are used to synthesize pixel-level features through word embeddings of unseen classes. GroupVit [35] introduces a grouping mechanism to cluster images automatically under text supervision. With the remarkable performance of pre-trained large-scale vision-language models in open-vocabulary classification, researchers begin exploring their performance in open-vocabulary semantic segmentation. These approaches can be broadly categorized into the following three types: 1)LSeg [18] aligns pixel-level embeddings with corresponding semantic class text embeddings for pixel-level classification. 2) [5, 20, 37] adopt a two-stage segmentation framework, where MaskFormer first generates a set of category-agnostic mask proposals, followed by CLIP performing open-vocabulary classification on each mask. 3)SAN [36] designs a side network to enable CLIP to perceive the predictions and recognition of masks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.664, + 0.787, + 0.695 + ], + "angle": 0, + "content": "Unlike these methods, OpenRSS is an end-to-end network that elegantly utilizes SAM and CLIP to achieve open-vocabulary semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.717, + 0.332, + 0.733 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we present OpenRSS, an end-to-end framework that combines pretrained large-scale models SAM and CLIP to perform open-vocabulary RGB-T semantic segmentation. Instead of fully fine-tuning all the parameters of the models, OpenRSS only learns a small amount of prompt information. Furthermore, we introduce trainable dynamic LoRA bypasses to adjust the frozen image encoder." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.149, + 0.784, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Fig. 1: Overview architecture of our OpenRSS. It inputs a pair of RGB and thermal images and generates corresponding RGB and thermal tokens through patch embeddings. Then, it utilizes the frozen Transformer encoder blocks of SAM for feature extraction. The TIP modules are inserted into the image encoder of SAM to learn effective visual prompts. DLoRA further adjusts the frozen image encoder. Finally, the mask decoder aligns image and text embeddings to generate the final segmentation map." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.559, + 0.341, + 0.573 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.789, + 0.75 + ], + "angle": 0, + "content": "The overall architecture of our OpenRSS is shown in Fig. 1. Formally, OpenRSS incorporates an additional thermal image input stream instead of a thermal feature extraction branch, which is fed into the network simultaneously with the RGB stream. Given a pair of RGB and thermal images, OpenRSS first projects the two information streams \\((T_{RGB}\\) and \\(T_{TH})\\) into patch embedding layers separately. After projection, we obtain two C-dimensional embeddings, which we denote as RGB tokens \\(\\pmb{I}_{rgb}\\) and thermal tokens \\(\\pmb{I}_{th}\\), respectively. Then, \\(\\pmb{I}_{rgb}\\) and \\(\\pmb{I}_{th}\\) are input into the TIP module to generate thermal information prompts. The learned prompt information is added to the original RGB information stream \\(\\pmb{I}_{rgb}\\) in a residual manner before being fed into the fully frozen image encoder of SAM:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.753, + 0.787, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} ^ {i} = \\boldsymbol {I} _ {r g b} ^ {i} + \\boldsymbol {T} ^ {i + 1}, \\quad i = 0, 1, \\dots , L - 1, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(\\pmb{I}^i\\) is the final sequence of image tokens fed into the Transformer layers of the encoder, and \\(\\pmb{T}^{i + 1}\\) is the prompt tokens sequence from the \\((i + 1)\\)th TIP module. By directly adding the thermal information prompts to the intermediate features generated by SAM, our OpenRSS can effortlessly perform downstream" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "tasks quickly. Furthermore, we also design a trainable dynamic LoRA bypass for each Transformer layer. For each bypass, the features generated by the Transformer layers are first mapped to a low-rank space and then projected back to the initial dimension, aligning it with the original output dimension of the frozen Transformer layers. We use a points grid prompt for the prompt encoder, significantly enhancing SAM's attention to every detail in the image compared to the default embeddings. For the text encoder, we use the frozen text encoder of CLIP. During training, we employ manually crafted prompts for the text encoder, consisting of a natural sentence and category names. However, any sentence can be used during inference to achieve the goal of open-vocabulary. We modify the original mask decoder of SAM to align its dimensionality with the feature embeddings of the text encoder. It simultaneously takes the embeddings from the prompt encoder and the text encoder as input, generating the final segmentation map." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.506, + 0.394 + ], + "angle": 0, + "content": "3.2 Thermal Information Prompt" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.552 + ], + "angle": 0, + "content": "In recent years, following the success of text prompts in NLP, some research has begun to explore the introduction of visual prompts into frozen pre-trained models. Prompt tuning is achieved through some learnable parameters and yields promising results on many downstream visual tasks. A key to our OpenRSS is to generate appropriate thermal information prompts to guide the network in effective open-vocabulary RGB-T semantic segmentation. As shown in Fig. 1, the TIP module is inserted before each Transformer layer. Formally, the thermal information prompts are generated from the token sequences \\(\\{I_{rgb}^{0}, I_{th}^{0}\\}\\) of both modalities. The TIP module learns the complementarity between these two modalities, which can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.561, + 0.785, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} ^ {i} = T I P ^ {i} \\left(\\boldsymbol {I} ^ {i - 1}, \\boldsymbol {T} ^ {i - 1}\\right), \\quad i = 1, 2, \\dots , L, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.767, + 0.605 + ], + "angle": 0, + "content": "where \\( TIP^i \\) denotes the \\( i \\)th TIP module. Specifically, \\( I^0 = I_{rqb}^0 \\) and \\( T^0 = I_{th}^0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.68 + ], + "angle": 0, + "content": "The detailed design of the TIP is shown in Fig. 2. TIP has two input branches, each receiving token sequences from two input streams. Specifically, considering the redundancy of single-modal features and the lightness of the prompt block, we use two \\(1 \\times 1\\) convolution operations to project each input stream to the size of \\([ \\alpha \\times H \\times W ]\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.689, + 0.785, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} _ {R G B} = c _ {1} \\left(\\boldsymbol {I} ^ {i - 1}\\right), \\quad \\boldsymbol {Z} _ {T H} = c _ {2} \\left(\\boldsymbol {T} ^ {i - 1}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.775 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is set to 8 in all TIP modules. \\(c_{1}(\\cdot)\\) and \\(c_{2}(\\cdot)\\) are \\(1\\times 1\\) convolution operations. Then, inspired by [32], we use a Transformer block with correlative self-attention for dense prediction to generate enhanced embeddings for \\(\\mathbf{Z}_{RGB}^{e}\\), as shown in Fig. 2. The formula for Correlative_attn is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.783, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\text {C o r r e l a t i v e} \\_ \\text {a t t n} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {X} \\boldsymbol {W} _ {q} \\boldsymbol {W} _ {q} ^ {T} \\boldsymbol {X} ^ {T} / \\tau\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}\\) denotes the input and \\(\\mathbf{W}_q\\) is the projection matrix. The temperature coefficient \\(\\tau\\) is by default set to \\(\\sqrt{d}\\) following traditional self-attention. Finally," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.149, + 0.776, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.789, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Overview architecture of our TIP module. Multi-modal input flows are sent separately to the \\(1 \\times 1\\) convolutions for distilling to the lower-dimension latent space. After passing through the Correlative Self-Attention block, the RGB embeddings are fused with thermal embeddings through element-wise addition. Finally, the multi-modal hybrid embeddings are mapped to the original dimension through a \\(1 \\times 1\\) convolutional layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.789, + 0.514 + ], + "angle": 0, + "content": "we obtain the mixed-modal embeddings by element-wise addition and project them back to the original dimension by:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.522, + 0.787, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} ^ {i} = c _ {3} \\left(\\boldsymbol {Z} _ {R G B} ^ {e} + \\boldsymbol {Z} _ {T H}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.547, + 0.517, + 0.564 + ], + "angle": 0, + "content": "where \\(c_{3}(\\cdot)\\) is the same as \\(c_{1}(\\cdot)\\) and \\(c_{2}(\\cdot)\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.582, + 0.48, + 0.598 + ], + "angle": 0, + "content": "3.3 DLoRA in Image Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.789, + 0.693 + ], + "angle": 0, + "content": "In OpenRSS, the thermal information prompts generated by TIP are fed into the frozen pre-trained image encoder of SAM to extract features. However, this may not be optimal due to the differences between the RGB and thermal modalities. Inspired by the idea from [10, 30], we design a Dynamic Low-Rank Adaptation (DLoRA) strategy to adjust the frozen image encoder of SAM to address this issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.79, + 0.816 + ], + "angle": 0, + "content": "Given an input token sequence \\( \\mathbf{X} \\in \\mathbb{R}^{B \\times N \\times C_{in}} \\), the output token sequence \\( \\mathbf{X}' \\in \\mathbb{R}^{B \\times N \\times C_{out}} \\) is obtained through a projection layer \\( \\mathbf{W} \\in \\mathbb{R}^{C_{in} \\times C_{out}} \\). LoRA assumes the updates to the weights have a low \"intrinsic rank\" during adaptation, thus proposing a low-rank decomposition to constrain its updates. Following this strategy, the Transformer layer is first frozen to fix \\( \\mathbf{W} \\), and then a bypass is added to achieve low-rank decomposition. This bypass consists of two linear layers \\( \\mathbf{A} \\in \\mathbb{R}^{C_{in} \\times r} \\) and \\( \\mathbf{B} \\in \\mathbb{R}^{r \\times C_{out}} \\), where \\( r \\ll \\min(C_{in}, C_{out}) \\). Therefore, the update of \\( \\mathbf{W} \\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.823, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {W}} = \\boldsymbol {W} + \\Delta \\boldsymbol {W} = \\boldsymbol {W} + \\boldsymbol {B} \\boldsymbol {A}. \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.15, + 0.709, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.336, + 0.785, + 0.365 + ], + "angle": 0, + "content": "Fig. 3: Overview architecture of our DLoRA. We apply the DLoRA layer to the q and v projection layers of each of the Transformer layers in the image encoder of SAM." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.393, + 0.784, + 0.439 + ], + "angle": 0, + "content": "We maintain consistency with [39] and apply LoRA to the query and value projection layers of the multi-head self-attention. Therefore, the modified query and value projection layers will become:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.449, + 0.786, + 0.474 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Q} = \\hat {\\boldsymbol {W}} _ {q} \\boldsymbol {X} = \\boldsymbol {W} _ {q} \\boldsymbol {X} + \\boldsymbol {B} _ {q} \\boldsymbol {A} _ {q} \\boldsymbol {X}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.493, + 0.784, + 0.522 + ], + "angle": 0, + "content": "where \\( \\mathbf{W}_q \\) and \\( \\mathbf{W}_v \\) are frozen from image encoder of SAM, and \\( \\mathbf{A}_q, \\mathbf{B}_q, \\mathbf{A}_v \\) and \\( \\mathbf{B}_v \\) are trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.785, + 0.629 + ], + "angle": 0, + "content": "While LoRA achieves remarkable results, it heavily relies on the choice of rank \\( r \\), requiring a manual and iterative search for the optimal rank. Therefore, we propose a Dynamic LoRA (DLoRA). Building upon the original LoRA, we increase the number of bypasses (as shown in Fig. 3) from 1 to 4, each with different ranks \\( \\gamma \\) and weights \\( \\beta \\). Through experimental observation, we find that our DLoRA can surpass the optimal rank found through manual iterative training in just one training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.499, + 0.667 + ], + "angle": 0, + "content": "3.4 Generate Segmentation Map" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Prompt Encoder and Mask Decoder. SAM passes the image embeddings generated by its image encoder and the prompts generated by its prompt encoder to its mask decoder for dense prediction. The types of prompts include point, box, and mask. SAM represents the first two sparse prompts as positional encodings and feeds them into the Transformer layers of the mask decoder. SAM performs convolution operations for dense prompts such as masks and then adds them element-wise with the image embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Due to SAM's prompt mechanism, it is susceptible to the position and quantity of prompt points. Therefore, in OpenRSS, we use a points grid as prompts to ensure that SAM does not overlook any detail in the image. SAM's mask decoder comprises a lightweight Transformer layer and a segmentation head. OpenRSS" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "modifies this segmentation head to align the dimension of the feature map with the dimension of CLIP's text features, enabling pixel-level classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.205, + 0.784, + 0.324 + ], + "angle": 0, + "content": "Pixel-level Classification via CLIP. Existing CLIP-based open-vocabulary segmentation methods mainly adopt a two-stage region recognition approach. In the first stage, an independent model is trained to generate a set of category-agnostic mask proposals. In the second stage, visual language pre-trained models such as CLIP are used to identify the category of each mask individually. However, this two-stage approach heavily relies on the quality of the masks generated in the first stage. Moreover, the cropped masks, due to missing backgrounds and varying sizes, are not suitable for recognition by CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.414 + ], + "angle": 0, + "content": "To address the problem of region recognition, our OpenRSS adopts pixel-level recognition. We define the feature map outputted by the mask decoder as \\(\\pmb{F} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times D}\\), and the class embeddings generated by the CLIP text encoder as \\(\\pmb{E} \\in \\mathbb{R}^{N \\times D}\\). Since they have the same embedding dimension, we can directly link them together through the inner product to obtain the final segmentation map:" + }, + { + "type": "equation", + "bbox": [ + 0.455, + 0.42, + 0.784, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} = \\boldsymbol {F} \\times \\boldsymbol {E} ^ {T}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.783, + 0.479 + ], + "angle": 0, + "content": "where \\(\\pmb{S} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times N}\\). This is the standard output of semantic segmentation, hence compatible with mainstream semantic segmentation evaluations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.507, + 0.376, + 0.525 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.524, + 0.558 + ], + "angle": 0, + "content": "4.1 Dataset and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.784, + 0.618 + ], + "angle": 0, + "content": "We conduct experiments on 4 datasets: MF dataset [9], KP dataset [11], MCubeS dataset [21] and Freiburg Thermal dataset [31]. All models are trained on the training set of the KP dataset and evaluated on other datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.784, + 0.722 + ], + "angle": 0, + "content": "MF Dataset. The dataset consists of urban street scenes captured using an InfRec R500 camera. It includes 1569 pairs of RGB and thermal images, each with a resolution of \\(480 \\times 640\\) pixels. Among these are 820 pairs of daytime images and 749 pairs of nighttime images. The dataset consists of eight manually labeled classes, as well as a background class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.784, + 0.84 + ], + "angle": 0, + "content": "KP Dataset. The dataset is an RGB-T paired urban driving scene dataset, providing 95328 pairs of RGB and thermal images with a resolution of \\(512 \\times 640\\). Initially, the KP dataset only provides bounding box labels for object detection tasks. However, fortunately, [15] provides semantic segmentation task labels for 503 daytime and 447 nighttime images. Furthermore, [27] partitions these 950 images into training, validation, and test sets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "MCubeS Dataset. MCubeS is a dataset with pairs of RGB, Near-Infrared (NIR), Degree of Linear Polarization (DoLP), and Angle of Linear Polarization (AoLP) to study semantic material segmentation of 20 classes with a resolution of \\(1024 \\times 1224\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.227, + 0.784, + 0.303 + ], + "angle": 0, + "content": "Freiburg Thermal Dataset. The dataset includes 12051 daytime and 8596 nighttime time-synchronized RGB-T image pairs captured in rural and urban environments with a resolution of \\(650 \\times 1920\\). However, only the test set includes 32 daytime and 32 nighttime images, comprising 12 object classes and one background class." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.322, + 0.784, + 0.366 + ], + "angle": 0, + "content": "Evaluation Metrics. We use mean Accuracy (mAcc) and mean Intersection over Union (mIoU) as the evaluation metrics for our model. The formulas for calculation are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.372, + 0.785, + 0.406 + ], + "angle": 0, + "content": "\\[\nm A c c = \\frac {1}{n _ {c l s}} \\cdot \\sum_ {i} \\frac {n _ {i i}}{t _ {i}}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.416, + 0.785, + 0.456 + ], + "angle": 0, + "content": "\\[\nm I o U = \\frac {1}{n _ {c l s}} \\cdot \\sum_ {i} \\frac {n _ {i i}}{t _ {i} + \\sum_ {j} n _ {j i} - n _ {i i}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.46, + 0.784, + 0.506 + ], + "angle": 0, + "content": "where \\( n_{ii} \\) is the number of pixels where class i is predicted as class j, \\( n_{cls} \\) is the total number of target classes (including the background), and \\( t_i = \\sum_j n_{ij} \\) is the total number of pixels for target class i (ground truth labels)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.525, + 0.456, + 0.54 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.547, + 0.784, + 0.683 + ], + "angle": 0, + "content": "We conduct all experiments on NVIDIA A100 GPUs. If not specified, we default to using SAM with the ViT-H backbone and CLIP with the ViT-L/16 backbone. We apply DLoRA fine-tuning to the frozen encoder Transformer layers, using four bypasses with ranks selected as 1, 2, 4, and 8, and weights as 0.1, 0.2, 0.4, and 0.3, respectively. All models are trained on the KP dataset with image flipping as a data augmentation technique. We use the AdamW optimizer with a weight decay of 0.01 and an initial learning rate of 5e-4. For the scheduler, we use the WarmupPolyLR scheduler. The loss function consists of DiceLoss and SoftCrossEntropy with weighting coefficients of 0.5 each:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.692, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\nD i c e L o s s = 1 - \\frac {2 \\sum_ {i = 1} ^ {N} y _ {i} \\widehat {y} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} \\widehat {y} _ {i} ^ {2}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.766, + 0.784, + 0.796 + ], + "angle": 0, + "content": "where \\( y_{i} \\) and \\( \\widehat{y_i} \\) represent the true and predicted label values of pixel I, respectively, and \\( N \\) represents the total number of pixels." + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.805, + 0.784, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\text {S o f t C r o s s E n t r o p y l o s s} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {c} \\hat {y} _ {i j} \\log \\left(y _ {i j} ^ {d}\\right), \\tag {12}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Performance comparison with state-of-the-art methods. For a fair comparison, we reproduce their method with their officially released code. The bold font highlights the best result in each column." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.199, + 0.763, + 0.436 + ], + "angle": 0, + "content": "
MethodBackboneMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
ZegFormer [5]R-5011.1110.3110.003.437.690.11
OVSeg [20]R-10110.869.4210.092.738.320.98
SAN [36]ViT-B14.644.1317.2410.107.131.31
ZegFormer [5]R-10111.5510.7110.464.117.850.79
OVSeg [20]Swin-B10.206.7816.496.5810.203.74
SAN [36]ViT-L3.810.386.904.2710.606.10
OpenRSS(ours)SAM-B22.2118.9631.7620.2716.018.14
SAM-L23.9120.0833.9524.9125.3316.65
SAM-H30.7322.8137.8828.5837.2527.55
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.531 + ], + "angle": 0, + "content": "where the \\( n \\) is the number of the batch, \\( \\hat{y}_{ij} \\) is a binary indicator if class label \\( c \\) is the correct classification for pixel \\( i \\), and \\( y_{ij}^{d} \\) is the corresponding predicted probability be normalized to a probability distribution. These two losses are summed to obtain the overall loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.55, + 0.786, + 0.567 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = 0. 5 L _ {\\text {D i c e L o s s}} + 0. 5 L _ {\\text {S o f t C r o s s E n t r o p y l o s s}}. \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.603, + 0.436, + 0.619 + ], + "angle": 0, + "content": "4.3 Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Tab. 1 presents the quantitative comparison results of our method with other state-of-the-art open-vocabulary semantic segmentation methods. We only use RGB images as input, rather than pairs of RGB and thermal images, allowing the general open vocabulary semantic segmentation in Tab. 1 to perform RGB-T semantic segmentation. From the table, it can be observed that our method achieves superior performance on both metrics across all three datasets, indicating the effectiveness of our approach. Specifically, our method surpasses other methods with an average of \\(+9.17\\%\\) mAcc and \\(+6.82\\%\\) mIoU for SAM-B, \\(+13.57\\%\\) mAcc and \\(+11.58\\%\\) mIoU for SAM-L, and \\(+20.46\\%\\) mAcc and \\(+17.34\\%\\) mIoU for SAM-H. This further validates that open-vocabulary semantic segmentation methods trained on RGB images are unsuitable for RGB-T images. Furthermore, we present the qualitative experimental results of OpenRSS on the Freiburg Thermal dataset in Fig. 4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.149, + 0.764, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.444, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results of OpenRSS on Freiburg Thermal dataset. From left to right: RGB images, Thermal images, ground-truths, OpenRSS results." + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.514, + 0.789, + 0.544 + ], + "angle": 0, + "content": "Table 2: Ablation studies of TIP module Table 3: Ablation studies of attn type in OpenRSS. in TIP module." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.555, + 0.784, + 0.664 + ], + "angle": 0, + "content": "
MethodMFMSFT
mIoU↑
TIP-empty10.2016.213.90
TIP-shallow19.9826.8325.44
TIP-deep(ours)22.8128.5827.55
attn-fovea18.4523.0118.58
attn-msa22.5421.7221.56
attn-cor(ours)22.8128.5827.55
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.39, + 0.705 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Importance of TIP. Tab. 2 shows the ablation experiments of the proposed TIP module. To validate the effectiveness of the TIP module, we replace it with additive fusion, resulting in an average mIoU performance drop of \\(16.21\\%\\). Furthermore, we discuss the specific placement of the TIP module. We only position the TIP module before the first Transformer layer, called TIP-shallow. The results indicate that placing the TIP module at each Transformer layer yields the best performance. Finally, we conduct ablation experiments on the correlative self-attention within the TIP module. As shown in Tab. 3, the results demon" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "strate the effectiveness of correlative self-attention for dense tasks by comparing it to spatial fovea attention and the original multi-head self-attention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.201, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Importance of DLoRA. Tab. 4 shows the ablation experiments of the proposed DLoRA. By manually setting the ranks to 1, 2, 4, and 8, the performance degradation of average mIoU is observed to be \\(4.69\\%\\), \\(2.55\\%\\), \\(3.03\\%\\), and \\(1.36\\%\\), respectively. This demonstrates the significance of DLoRA while also avoiding spending considerable time selecting an optimal low-rank space. More detailed results and analysis, including failure cases, are discussed in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.331, + 0.787, + 0.36 + ], + "angle": 0, + "content": "Table 4: Ablation of rank in DLoRA. Manually set rank vs. dynamically adaptive rank. The bold font highlights the best result in each column." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.372, + 0.751, + 0.525 + ], + "angle": 0, + "content": "
MethodsMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
OpenRSS-rank-129.2020.6137.4324.6928.4319.57
OpenRSS-rank-226.6121.7835.7725.9832.6023.51
OpenRSS-rank-425.3821.6336.5327.7329.6420.49
OpenRSS-rank-827.2622.5137.0726.2835.2326.06
OpenRSS-DLoRA(ours)30.7322.8137.8828.5837.2527.55
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.36, + 0.597 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.764 + ], + "angle": 0, + "content": "This work proposes the first open-vocabulary RGB-T semantic segmentation framework: OpenRSS. Our framework achieves the open-vocabulary capability by jointly utilizing the CLIP and the modified SAM model. In addition, when fusing RGB and thermal modality information, OpenRSS only needs to add a few learnable parameters to SAM to achieve effective multi-modal information fusion. Extensive experiments show that the proposed framework significantly outperforms previous methods on three RGB-T semantic segmentation benchmarks, validating the effectiveness and generalization of OpenRSS. Our work provides a feasible solution to open-vocabulary semantic segmentation with RGB-T images and a strong baseline for the community to advance future research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grant 62203337; in part by the Hubei Province Natural Science Foundation, China, under Grant 2022CFC074; and in part by the Open Fund of the Hubei Garment Informatization Engineering Technology Research Center, Wuhan Textile University, under Grant 2022HBCI04." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.227 + ], + "angle": 0, + "content": "2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.785, + 0.268 + ], + "angle": 0, + "content": "3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.269, + 0.785, + 0.323 + ], + "angle": 0, + "content": "4. Deng, F., Feng, H., Liang, M., Wang, H., Yang, Y., Gao, Y., Chen, J., Hu, J., Guo, X., Lam, T.L.: Fanet: Feature-enhanced attention network for rgb-thermal real-time semantic segmentation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4467-4473. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.363 + ], + "angle": 0, + "content": "5. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Dong, S., Feng, Y., Yang, Q., Huang, Y., Liu, D., Fan, H.: Efficient multimodal semantic segmentation via dual-prompt learning. arXiv preprint arXiv:2312.00360 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.431 + ], + "angle": 0, + "content": "7. Fu, Y., Xiang, T., Jiang, Y.G., Xue, X., Sigal, L., Gong, S.: Recent advances in zero-shot recognition. arXiv preprint arXiv:1710.04837 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.486 + ], + "angle": 0, + "content": "8. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.54 + ], + "angle": 0, + "content": "9. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "10. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "1. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1037-1045 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.676 + ], + "angle": 0, + "content": "2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "3. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "4. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497-6504 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.496, + 0.129 + ], + "angle": 0, + "content": "G. Zhao, J. Huang, X. Yan, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "18. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.247 + ], + "angle": 0, + "content": "19. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.303 + ], + "angle": 0, + "content": "20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "21. Liang, Y., Wakaki, R., Nobuhara, S., Nishino, K.: Multimodal material segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19800-19808 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.389 + ], + "angle": 0, + "content": "22. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.403 + ], + "angle": 0, + "content": "23. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article 2, 13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.459 + ], + "angle": 0, + "content": "24. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.502 + ], + "angle": 0, + "content": "25. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.557 + ], + "angle": 0, + "content": "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.586 + ], + "angle": 0, + "content": "27. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. arXiv preprint arXiv:2303.17386 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "28. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.671 + ], + "angle": 0, + "content": "29. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.785, + 0.712 + ], + "angle": 0, + "content": "30. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.785, + 0.755 + ], + "angle": 0, + "content": "31. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.784 + ], + "angle": 0, + "content": "32. Wang, F., Mei, J., Yuille, A.: Sclip: Rethinking self-attention for dense vision-language inference. arXiv preprint arXiv:2312.01597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.784, + 0.785, + 0.812 + ], + "angle": 0, + "content": "33. Wang, K., Li, C., Tu, Z., Luo, B.: Unified-modal salient object detection via adaptive prompt learning. arXiv preprint arXiv:2311.16835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.841 + ], + "angle": 0, + "content": "34. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "35. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.245 + ], + "angle": 0, + "content": "36. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "37. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "38. Zhang, J., Liu, H., Yang, K., Hu, X., Liu, R., Stiefelhagen, R.: Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.355 + ], + "angle": 0, + "content": "39. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.412 + ], + "angle": 0, + "content": "40. Zhang, Q., Zhao, S., Luo, Y., Zhang, D., Huang, N., Han, J.: Abmdrnet: Adaptive-weighted bi-directional modality difference reduction network for rgb-t semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2633-2642 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.453 + ], + "angle": 0, + "content": "41. Zhao, H., Qi, X., Shen, X., Shi, J., Jia, J.: Icnet for real-time semantic segmentation on high-resolution images. In: Proceedings of the European conference on computer vision (ECCV). pp. 405-420 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.495 + ], + "angle": 0, + "content": "42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.522 + ], + "angle": 0, + "content": "43. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "44. Zhou, W., Dong, S., Xu, C., Qian, Y.: Edge-aware guidance fusion network for rgb-thermal scene parsing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3571-3579 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "45. Zhu, J., Lai, S., Chen, X., Wang, D., Lu, H.: Visual prompt multi-modal tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9516-9526 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.605 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_origin.pdf b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b40c74711ca3caf899ad83cdf49356af31b0dfa7 --- /dev/null +++ b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/8ecf13d6-1b91-4616-aea0-71a3f9b67335_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5797e939c4cbca0e028cd22c882939319366ffc4da629c7e1ae730284320328 +size 845288 diff --git a/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/full.md b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b139d88e97eb43a39bd35ca4748caf6f7652139b --- /dev/null +++ b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/full.md @@ -0,0 +1,286 @@ +# Open-Vocabulary RGB-Thermal Semantic Segmentation + +Guoqiang Zhao $^{1}$ , Junjie Huang $^{1}$ , Xiaoyun Yan $^{1,2\dagger}$ , Zhaojing Wang $^{1,2}$ , Junwei Tang $^{1}$ , Yangjun Ou $^{1}$ , Xinrong Hu $^{1,2}$ , and Tao Peng $^{1}$ + +1 School of Computer Science and Artificial Intelligence, Wuhan Textile University zgq_lyy@163.com, xyyan@wtu.edu.cn + +$^{2}$ Hubei Garment Informatization Engineering Technology Research Center + +Abstract. RGB-Thermal (RGB-T) semantic segmentation is an important research branch of multi-modal image segmentation. The current RGB-T semantic segmentation methods generally have two unsolved and typical shortcomings. First, they do not have the open-vocabulary recognition ability, which significantly limits their application scenarios. Second, when fusing RGB and thermal images, they often need to design complex fusion network structures, which usually results in low network training efficiency. We present OpenRSS, the Open-vocabulary RGB-T Semantic Segmentation method, to solve these two disadvantages. To our knowledge, OpenRSS is the first RGB-T semantic segmentation method with open-vocabulary segmentation capability. OpenRSS modifies the basic segmentation model SAM for RGB-T semantic segmentation by adding the proposed thermal information prompt module and dynamic low-rank adaptation strategy to SAM. These designs effectively fuse the RGB and thermal information, but with much fewer trainable parameters than other methods. OpenRSS achieves the open-vocabulary capability by jointly utilizing the vision-language model CLIP and the modified SAM. Through extensive experiments, OpenRSS demonstrates its effective open-vocabulary semantic segmentation ability on RGB-T images. It outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: $+12.1\%$ mIoU on the MFNet dataset, $+18.4\%$ mIoU on the MCubeS dataset, and $+21.4\%$ mIoU on the Freiburg Thermal dataset. Code will be released at https://github.com/SXDR/OpenRSS. + +Keywords: Open-vocabulary $\cdot$ RGB-T semantic segmentation $\cdot$ Prompt learning + +# 1 Introduction + +Existing deep learning-based semantic segmentation methods [22, 25, 41] mainly rely on single-modal RGB images, which are effective only under well-lit conditions during the daytime and perform poorly in challenging lighting conditions + +(e.g., nighttime, foggy weather). With the popularity of thermal cameras, researchers have recently started incorporating thermal images as a supplementary modality into RGB semantic segmentation [8], known as RGB-T semantic segmentation. + +With the advancement of deep learning, current RGB-T semantic segmentation models mainly rely on supervised training with a predefined set of fixed categories, which cannot generalize to categories beyond the training set. This shortcoming significantly limits the application scenarios of current RGB-T semantic segmentation methods. In contrast, humans understand scenes in an open-vocabulary manner and recognize around 30,000 basic categories [1, 7]. Moreover, humans can further identify new categories through language descriptions. + +Recently, large-scale vision-language models, represented by CLIP [26], achieve image-level classification across arbitrary categories, i.e., open-vocabulary classification. However, since the CLIP model is trained using image-level contrastive learning, how to effectively utilize CLIP in semantic segmentation is still an open problem. Some researchers adopt a two-stage training paradigm [5,20,37]: first, using a model to generate category-agnostic mask proposals, and then a pre-trained CLIP model for open-vocabulary classification on each mask. While others choose to train a side network [36] to integrate pre-trained CLIP model features into the side network for mask recognition. The promising performance of CLIP prompts us to explore its applicability in RGB-T semantic segmentation. But unlike the methods mentioned above, in this paper, we jointly utilize the basic segmentation model SAM [16] and CLIP through an elegant way to achieve open-vocabulary semantic segmentation, which effectively takes advantage of well-learned visual features and high segmentation performance of SAM and powerful open-vocabulary classification ability of CLIP. + +While SAM has learned a powerful and general visual segmentation model using the largest segmentation dataset to date, SA-1B, it is still unclear how to effectively transfer its knowledge from RGB images to RGB-T images. The most direct strategy is to train the SAM from scratch, specifically for RGB-T images. However, on the one hand, full model fine-tuning is time-consuming, inefficient, and burdensome regarding parameter storage. On the other hand, with the limited quantity of RGB-T semantic segmentation datasets, it is easy to produce over-fitting results. Thus, a natural question arises: is there a more efficient and effective method to enable SAM models trained on RGB images to be utilized for RGB-T images? Our answer to this question is affirmative, and we provide a solution with the paradigm of prompt learning. Prompt learning, also known as prompt tuning, originated from natural language processing (NLP) [17, 19], aims to inject text prompts into language models to transfer the knowledge of foundation models to various downstream tasks. Inspired by this, researchers apply prompt tuning to visual tasks [3,6,13,33,42,43,45] by freezing the entire foundation model and adding some learnable parameters only to the input side to learn effective visual prompts. Intuitively, there is a significant correlation between multimodal and RGB modalities, and they should share most of the + +prior knowledge about feature extraction. Following this, we propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation method. + +Unlike other RGB-T semantic segmentation methods that always need complex fusion networks to fuse RGB and thermal information, which inevitably adds considerable computation and training burden to their approaches. In OpenRSS, only a small amount of parameters are trainable, which makes the training of OpenRSS quite efficient and also decreases the risk of over-fitting. During the training, OpenRSS freezes the entire image encoder of SAM and only learns visual prompts based on RGB and thermal modality. Specifically, OpenRSS freezes the image encoder of SAM and adds several simple and efficient Thermal Information Prompt (TIP) modules to SAM to produce visual prompts to learn the complementarity between RGB and thermal modality effectively. Then, in the mask decoder of SAM, we align pixel embeddings with the text embeddings corresponding to the semantic classes generated by CLIP. Due to the flexible label representations provided by the text encoder, OpenRSS can perform open-vocabulary segmentation without requiring additional training samples. Additionally, we employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to fine-tune the image encoder of SAM. For fairness and reproducibility, our research is based on the officially released SAM and CLIP models. The main contributions of this paper are as follows: + +- We propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation model, where, with the aid of prompt learning and dynamic low-rank adaptation strategy, existing foundation models can effectively adapt to open-vocabulary RGB-T semantic segmentation. +- We introduce a Thermal Information Prompt (TIP) module that generates simple yet efficient visual prompts to fuse RGB and thermal information, avoiding the complex and redundant additional branches designed in previous methods. +- We employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to optimize a series of parameters in the image encoder of SAM. +- Extensive experiments demonstrate that OpenRSS outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: $+12.1\%$ mIoU on the MFNet dataset, $+18.4\%$ mIoU on the MCubeS dataset, and $+21.4\%$ mIoU on the Freiburg Thermal dataset, while maintaining parameter efficiency (only $1\%$ trainable parameters). + +The remainder of this article is organized as follows. We review related work in Sec. 2. In Sec. 3, we provide a detailed description of our network architecture. Sec. 4 present experimental results and discussions. Finally, we summarize the work in Sec. 5. + +# 2 Related Work + +# 2.1 RGB-T Semantic Segmentation + +For complex scenes with challenging lighting conditions, thermal images are highly complementary to RGB images. Existing RGB-T semantic segmentation methods focus on designing complex cross-modal fusion paradigms. We briefly review RGB-T semantic segmentation from three paradigms: encoder fusion, decoder fusion, and feature fusion. For the encoder fusion paradigm, RTFNet [28] and FuseSeg [29] integrate thermal features into RGB features by element-wise summation. FEANet [4] introduces a two-stage feature-enhanced attention network, which utilizes channel and spatial attention to fuse RGB and thermal information in a complementary way. For the decoder fusion paradigm, MFNet [8] uses two identical encoders to extract RGB and thermal features and feed them into the decoder for integration. For the feature fusion paradigm, EGFNet [44] embeds prior edge information into feature maps for cross-modal fusion. ABMDRNet [40] employs a bi-directional image-to-image translation based method to reduce the modality differences, followed by channel-wise weighted fusion for final prediction. CMX [38] first combines features from other modalities in both spatial and channel directions to calibrate the current modality's features, then fuses them for cross-modal fusion. + +Unlike the methods mentioned above, which always have to train complex additional network branches to extract thermal-related features, OpenRSS only needs to learn a small amount of trainable parameters to inject thermal-related information into the image encoder of SAM, whose parameters are frozen during training. This merit gives OpenRSS high parameters and training efficiency. + +# 2.2 Large-scale Models + +The Transformer's robust scalability allows large-scale models to have hundreds of billions of parameters. Large-scale models, such as BERT [14], Instruct-GPT [24], GPT-4 [23], etc, achieve groundbreaking progress in natural language processing tasks. Large-scale models like SAM [16] and SegGPT [34] demonstrate outstanding zero-shot generalization capabilities in computer vision. Recently, large-scale vision-language models represented by CLIP [26] and ALIGN [12] can recognize arbitrary categories described in the text, known as open-vocabulary classification. + +Our work further explores how to effectively utilize pre-trained visual and visual-language large-scale models to perform open-vocabulary semantic segmentation on RGB-T images, and accordingly proposes our solution: the OpenRSS model. + +# 2.3 Visual Prompt Learning + +In computer vision, downstream tasks are always executed by fine-tuning large pre-trained models comprehensively for a long time. Specifically, different down- + +stream data are used to retrain downstream tasks to update all parameters. Although this approach yields satisfactory performance, it is inefficient regarding parameter usage and ultimately loses the universal knowledge learned by the pretrained models. Prompt learning (also known as prompt tuning) is recognized as a trade-off between the training burden and the final performance in the downstream tasks and has recently garnered widespread attention. It initially shone in natural language processing before being introduced to various computer vision tasks. For example, VPT [13] presents a set of learnable parameters for Transformer encoders and significantly outperforms full fine-tuning on multiple downstream tasks. ViPT [45] introduces prompts for pre-trained single-modal trackers, making them applicable to multimodal tracking tasks. UniSOD [33] designs a switchable prompt generation block to interact with single-modal or multimodal inputs. + +In previous works, they often used simple $1 \times 1$ or $3 \times 3$ convolutions to generate prompts. In contrast, we specifically design a Thermal Information Prompt (TIP) module with a Transformer block inside to generate prompts for dense prediction, enabling the pre-trained SAM model to adapt to RGB-T images. + +# 2.4 Open-vocabulary Semantic Segmentation + +Open-vocabulary semantic segmentation refers to identifying pixel-level categories in images through arbitrary text descriptions. In initial works like ZS3Net [2], generative models are used to synthesize pixel-level features through word embeddings of unseen classes. GroupVit [35] introduces a grouping mechanism to cluster images automatically under text supervision. With the remarkable performance of pre-trained large-scale vision-language models in open-vocabulary classification, researchers begin exploring their performance in open-vocabulary semantic segmentation. These approaches can be broadly categorized into the following three types: 1)LSeg [18] aligns pixel-level embeddings with corresponding semantic class text embeddings for pixel-level classification. 2) [5, 20, 37] adopt a two-stage segmentation framework, where MaskFormer first generates a set of category-agnostic mask proposals, followed by CLIP performing open-vocabulary classification on each mask. 3)SAN [36] designs a side network to enable CLIP to perceive the predictions and recognition of masks. + +Unlike these methods, OpenRSS is an end-to-end network that elegantly utilizes SAM and CLIP to achieve open-vocabulary semantic segmentation. + +# 3 Method + +In this paper, we present OpenRSS, an end-to-end framework that combines pretrained large-scale models SAM and CLIP to perform open-vocabulary RGB-T semantic segmentation. Instead of fully fine-tuning all the parameters of the models, OpenRSS only learns a small amount of prompt information. Furthermore, we introduce trainable dynamic LoRA bypasses to adjust the frozen image encoder. + +![](images/b4f45d9c8f5e3fc70233e6904cf555900af7a336c1cfbfbf4da333bdcad0341a.jpg) +Fig. 1: Overview architecture of our OpenRSS. It inputs a pair of RGB and thermal images and generates corresponding RGB and thermal tokens through patch embeddings. Then, it utilizes the frozen Transformer encoder blocks of SAM for feature extraction. The TIP modules are inserted into the image encoder of SAM to learn effective visual prompts. DLoRA further adjusts the frozen image encoder. Finally, the mask decoder aligns image and text embeddings to generate the final segmentation map. + +# 3.1 Overview + +The overall architecture of our OpenRSS is shown in Fig. 1. Formally, OpenRSS incorporates an additional thermal image input stream instead of a thermal feature extraction branch, which is fed into the network simultaneously with the RGB stream. Given a pair of RGB and thermal images, OpenRSS first projects the two information streams $(T_{RGB}$ and $T_{TH})$ into patch embedding layers separately. After projection, we obtain two C-dimensional embeddings, which we denote as RGB tokens $\pmb{I}_{rgb}$ and thermal tokens $\pmb{I}_{th}$ , respectively. Then, $\pmb{I}_{rgb}$ and $\pmb{I}_{th}$ are input into the TIP module to generate thermal information prompts. The learned prompt information is added to the original RGB information stream $\pmb{I}_{rgb}$ in a residual manner before being fed into the fully frozen image encoder of SAM: + +$$ +\boldsymbol {I} ^ {i} = \boldsymbol {I} _ {r g b} ^ {i} + \boldsymbol {T} ^ {i + 1}, \quad i = 0, 1, \dots , L - 1, \tag {1} +$$ + +where $\pmb{I}^i$ is the final sequence of image tokens fed into the Transformer layers of the encoder, and $\pmb{T}^{i + 1}$ is the prompt tokens sequence from the $(i + 1)$ th TIP module. By directly adding the thermal information prompts to the intermediate features generated by SAM, our OpenRSS can effortlessly perform downstream + +tasks quickly. Furthermore, we also design a trainable dynamic LoRA bypass for each Transformer layer. For each bypass, the features generated by the Transformer layers are first mapped to a low-rank space and then projected back to the initial dimension, aligning it with the original output dimension of the frozen Transformer layers. We use a points grid prompt for the prompt encoder, significantly enhancing SAM's attention to every detail in the image compared to the default embeddings. For the text encoder, we use the frozen text encoder of CLIP. During training, we employ manually crafted prompts for the text encoder, consisting of a natural sentence and category names. However, any sentence can be used during inference to achieve the goal of open-vocabulary. We modify the original mask decoder of SAM to align its dimensionality with the feature embeddings of the text encoder. It simultaneously takes the embeddings from the prompt encoder and the text encoder as input, generating the final segmentation map. + +# 3.2 Thermal Information Prompt + +In recent years, following the success of text prompts in NLP, some research has begun to explore the introduction of visual prompts into frozen pre-trained models. Prompt tuning is achieved through some learnable parameters and yields promising results on many downstream visual tasks. A key to our OpenRSS is to generate appropriate thermal information prompts to guide the network in effective open-vocabulary RGB-T semantic segmentation. As shown in Fig. 1, the TIP module is inserted before each Transformer layer. Formally, the thermal information prompts are generated from the token sequences $\{I_{rgb}^{0}, I_{th}^{0}\}$ of both modalities. The TIP module learns the complementarity between these two modalities, which can be represented as follows: + +$$ +\boldsymbol {T} ^ {i} = T I P ^ {i} \left(\boldsymbol {I} ^ {i - 1}, \boldsymbol {T} ^ {i - 1}\right), \quad i = 1, 2, \dots , L, \tag {2} +$$ + +where $TIP^i$ denotes the $i$ th TIP module. Specifically, $I^0 = I_{rqb}^0$ and $T^0 = I_{th}^0$ . + +The detailed design of the TIP is shown in Fig. 2. TIP has two input branches, each receiving token sequences from two input streams. Specifically, considering the redundancy of single-modal features and the lightness of the prompt block, we use two $1 \times 1$ convolution operations to project each input stream to the size of $[ \alpha \times H \times W ]$ by: + +$$ +\boldsymbol {Z} _ {R G B} = c _ {1} \left(\boldsymbol {I} ^ {i - 1}\right), \quad \boldsymbol {Z} _ {T H} = c _ {2} \left(\boldsymbol {T} ^ {i - 1}\right), \tag {3} +$$ + +where $\alpha$ is set to 8 in all TIP modules. $c_{1}(\cdot)$ and $c_{2}(\cdot)$ are $1\times 1$ convolution operations. Then, inspired by [32], we use a Transformer block with correlative self-attention for dense prediction to generate enhanced embeddings for $\mathbf{Z}_{RGB}^{e}$ , as shown in Fig. 2. The formula for Correlative_attn is expressed as: + +$$ +\text {C o r r e l a t i v e} \_ \text {a t t n} = \operatorname {S o f t m a x} \left(\boldsymbol {X} \boldsymbol {W} _ {q} \boldsymbol {W} _ {q} ^ {T} \boldsymbol {X} ^ {T} / \tau\right), \tag {4} +$$ + +where $\mathbf{X}$ denotes the input and $\mathbf{W}_q$ is the projection matrix. The temperature coefficient $\tau$ is by default set to $\sqrt{d}$ following traditional self-attention. Finally, + +![](images/00ed787fe22808959742094ec4063222c64a618e11002e7c4a6bc6edd89762d5.jpg) +Fig. 2: Overview architecture of our TIP module. Multi-modal input flows are sent separately to the $1 \times 1$ convolutions for distilling to the lower-dimension latent space. After passing through the Correlative Self-Attention block, the RGB embeddings are fused with thermal embeddings through element-wise addition. Finally, the multi-modal hybrid embeddings are mapped to the original dimension through a $1 \times 1$ convolutional layer. + +we obtain the mixed-modal embeddings by element-wise addition and project them back to the original dimension by: + +$$ +\boldsymbol {T} ^ {i} = c _ {3} \left(\boldsymbol {Z} _ {R G B} ^ {e} + \boldsymbol {Z} _ {T H}\right), \tag {5} +$$ + +where $c_{3}(\cdot)$ is the same as $c_{1}(\cdot)$ and $c_{2}(\cdot)$ . + +# 3.3 DLoRA in Image Encoder + +In OpenRSS, the thermal information prompts generated by TIP are fed into the frozen pre-trained image encoder of SAM to extract features. However, this may not be optimal due to the differences between the RGB and thermal modalities. Inspired by the idea from [10, 30], we design a Dynamic Low-Rank Adaptation (DLoRA) strategy to adjust the frozen image encoder of SAM to address this issue. + +Given an input token sequence $\mathbf{X} \in \mathbb{R}^{B \times N \times C_{in}}$ , the output token sequence $\mathbf{X}' \in \mathbb{R}^{B \times N \times C_{out}}$ is obtained through a projection layer $\mathbf{W} \in \mathbb{R}^{C_{in} \times C_{out}}$ . LoRA assumes the updates to the weights have a low "intrinsic rank" during adaptation, thus proposing a low-rank decomposition to constrain its updates. Following this strategy, the Transformer layer is first frozen to fix $\mathbf{W}$ , and then a bypass is added to achieve low-rank decomposition. This bypass consists of two linear layers $\mathbf{A} \in \mathbb{R}^{C_{in} \times r}$ and $\mathbf{B} \in \mathbb{R}^{r \times C_{out}}$ , where $r \ll \min(C_{in}, C_{out})$ . Therefore, the update of $\mathbf{W}$ can be described as: + +$$ +\hat {\boldsymbol {W}} = \boldsymbol {W} + \Delta \boldsymbol {W} = \boldsymbol {W} + \boldsymbol {B} \boldsymbol {A}. \tag {6} +$$ + +![](images/3fd25287071bda43007e633aad1ab66e0bbaf4f3f3db226bcd14f508a0793673.jpg) +Fig. 3: Overview architecture of our DLoRA. We apply the DLoRA layer to the q and v projection layers of each of the Transformer layers in the image encoder of SAM. + +We maintain consistency with [39] and apply LoRA to the query and value projection layers of the multi-head self-attention. Therefore, the modified query and value projection layers will become: + +$$ +\boldsymbol {Q} = \hat {\boldsymbol {W}} _ {q} \boldsymbol {X} = \boldsymbol {W} _ {q} \boldsymbol {X} + \boldsymbol {B} _ {q} \boldsymbol {A} _ {q} \boldsymbol {X}, \tag {7} +$$ + +where $\mathbf{W}_q$ and $\mathbf{W}_v$ are frozen from image encoder of SAM, and $\mathbf{A}_q, \mathbf{B}_q, \mathbf{A}_v$ and $\mathbf{B}_v$ are trainable parameters. + +While LoRA achieves remarkable results, it heavily relies on the choice of rank $r$ , requiring a manual and iterative search for the optimal rank. Therefore, we propose a Dynamic LoRA (DLoRA). Building upon the original LoRA, we increase the number of bypasses (as shown in Fig. 3) from 1 to 4, each with different ranks $\gamma$ and weights $\beta$ . Through experimental observation, we find that our DLoRA can surpass the optimal rank found through manual iterative training in just one training. + +# 3.4 Generate Segmentation Map + +Prompt Encoder and Mask Decoder. SAM passes the image embeddings generated by its image encoder and the prompts generated by its prompt encoder to its mask decoder for dense prediction. The types of prompts include point, box, and mask. SAM represents the first two sparse prompts as positional encodings and feeds them into the Transformer layers of the mask decoder. SAM performs convolution operations for dense prompts such as masks and then adds them element-wise with the image embeddings. + +Due to SAM's prompt mechanism, it is susceptible to the position and quantity of prompt points. Therefore, in OpenRSS, we use a points grid as prompts to ensure that SAM does not overlook any detail in the image. SAM's mask decoder comprises a lightweight Transformer layer and a segmentation head. OpenRSS + +modifies this segmentation head to align the dimension of the feature map with the dimension of CLIP's text features, enabling pixel-level classification. + +Pixel-level Classification via CLIP. Existing CLIP-based open-vocabulary segmentation methods mainly adopt a two-stage region recognition approach. In the first stage, an independent model is trained to generate a set of category-agnostic mask proposals. In the second stage, visual language pre-trained models such as CLIP are used to identify the category of each mask individually. However, this two-stage approach heavily relies on the quality of the masks generated in the first stage. Moreover, the cropped masks, due to missing backgrounds and varying sizes, are not suitable for recognition by CLIP. + +To address the problem of region recognition, our OpenRSS adopts pixel-level recognition. We define the feature map outputted by the mask decoder as $\pmb{F} \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times D}$ , and the class embeddings generated by the CLIP text encoder as $\pmb{E} \in \mathbb{R}^{N \times D}$ . Since they have the same embedding dimension, we can directly link them together through the inner product to obtain the final segmentation map: + +$$ +\boldsymbol {S} = \boldsymbol {F} \times \boldsymbol {E} ^ {T}, \tag {8} +$$ + +where $\pmb{S} \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times N}$ . This is the standard output of semantic segmentation, hence compatible with mainstream semantic segmentation evaluations. + +# 4 Experiments + +# 4.1 Dataset and Evaluation Metrics + +We conduct experiments on 4 datasets: MF dataset [9], KP dataset [11], MCubeS dataset [21] and Freiburg Thermal dataset [31]. All models are trained on the training set of the KP dataset and evaluated on other datasets. + +MF Dataset. The dataset consists of urban street scenes captured using an InfRec R500 camera. It includes 1569 pairs of RGB and thermal images, each with a resolution of $480 \times 640$ pixels. Among these are 820 pairs of daytime images and 749 pairs of nighttime images. The dataset consists of eight manually labeled classes, as well as a background class. + +KP Dataset. The dataset is an RGB-T paired urban driving scene dataset, providing 95328 pairs of RGB and thermal images with a resolution of $512 \times 640$ . Initially, the KP dataset only provides bounding box labels for object detection tasks. However, fortunately, [15] provides semantic segmentation task labels for 503 daytime and 447 nighttime images. Furthermore, [27] partitions these 950 images into training, validation, and test sets. + +MCubeS Dataset. MCubeS is a dataset with pairs of RGB, Near-Infrared (NIR), Degree of Linear Polarization (DoLP), and Angle of Linear Polarization (AoLP) to study semantic material segmentation of 20 classes with a resolution of $1024 \times 1224$ . + +Freiburg Thermal Dataset. The dataset includes 12051 daytime and 8596 nighttime time-synchronized RGB-T image pairs captured in rural and urban environments with a resolution of $650 \times 1920$ . However, only the test set includes 32 daytime and 32 nighttime images, comprising 12 object classes and one background class. + +Evaluation Metrics. We use mean Accuracy (mAcc) and mean Intersection over Union (mIoU) as the evaluation metrics for our model. The formulas for calculation are as follows: + +$$ +m A c c = \frac {1}{n _ {c l s}} \cdot \sum_ {i} \frac {n _ {i i}}{t _ {i}}, \tag {9} +$$ + +$$ +m I o U = \frac {1}{n _ {c l s}} \cdot \sum_ {i} \frac {n _ {i i}}{t _ {i} + \sum_ {j} n _ {j i} - n _ {i i}}, \tag {10} +$$ + +where $n_{ii}$ is the number of pixels where class i is predicted as class j, $n_{cls}$ is the total number of target classes (including the background), and $t_i = \sum_j n_{ij}$ is the total number of pixels for target class i (ground truth labels). + +# 4.2 Implementation Details + +We conduct all experiments on NVIDIA A100 GPUs. If not specified, we default to using SAM with the ViT-H backbone and CLIP with the ViT-L/16 backbone. We apply DLoRA fine-tuning to the frozen encoder Transformer layers, using four bypasses with ranks selected as 1, 2, 4, and 8, and weights as 0.1, 0.2, 0.4, and 0.3, respectively. All models are trained on the KP dataset with image flipping as a data augmentation technique. We use the AdamW optimizer with a weight decay of 0.01 and an initial learning rate of 5e-4. For the scheduler, we use the WarmupPolyLR scheduler. The loss function consists of DiceLoss and SoftCrossEntropy with weighting coefficients of 0.5 each: + +$$ +D i c e L o s s = 1 - \frac {2 \sum_ {i = 1} ^ {N} y _ {i} \widehat {y} _ {i}}{\sum_ {i = 1} ^ {N} y _ {i} ^ {2} + \sum_ {i = 1} ^ {N} \widehat {y} _ {i} ^ {2}}, \tag {11} +$$ + +where $y_{i}$ and $\widehat{y_i}$ represent the true and predicted label values of pixel I, respectively, and $N$ represents the total number of pixels. + +$$ +\text {S o f t C r o s s E n t r o p y l o s s} = - \frac {1}{n} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {c} \hat {y} _ {i j} \log \left(y _ {i j} ^ {d}\right), \tag {12} +$$ + +Table 1: Performance comparison with state-of-the-art methods. For a fair comparison, we reproduce their method with their officially released code. The bold font highlights the best result in each column. + +
MethodBackboneMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
ZegFormer [5]R-5011.1110.3110.003.437.690.11
OVSeg [20]R-10110.869.4210.092.738.320.98
SAN [36]ViT-B14.644.1317.2410.107.131.31
ZegFormer [5]R-10111.5510.7110.464.117.850.79
OVSeg [20]Swin-B10.206.7816.496.5810.203.74
SAN [36]ViT-L3.810.386.904.2710.606.10
OpenRSS(ours)SAM-B22.2118.9631.7620.2716.018.14
SAM-L23.9120.0833.9524.9125.3316.65
SAM-H30.7322.8137.8828.5837.2527.55
+ +where the $n$ is the number of the batch, $\hat{y}_{ij}$ is a binary indicator if class label $c$ is the correct classification for pixel $i$ , and $y_{ij}^{d}$ is the corresponding predicted probability be normalized to a probability distribution. These two losses are summed to obtain the overall loss function: + +$$ +L _ {t o t a l} = 0. 5 L _ {\text {D i c e L o s s}} + 0. 5 L _ {\text {S o f t C r o s s E n t r o p y l o s s}}. \tag {13} +$$ + +# 4.3 Results and Analysis + +Tab. 1 presents the quantitative comparison results of our method with other state-of-the-art open-vocabulary semantic segmentation methods. We only use RGB images as input, rather than pairs of RGB and thermal images, allowing the general open vocabulary semantic segmentation in Tab. 1 to perform RGB-T semantic segmentation. From the table, it can be observed that our method achieves superior performance on both metrics across all three datasets, indicating the effectiveness of our approach. Specifically, our method surpasses other methods with an average of $+9.17\%$ mAcc and $+6.82\%$ mIoU for SAM-B, $+13.57\%$ mAcc and $+11.58\%$ mIoU for SAM-L, and $+20.46\%$ mAcc and $+17.34\%$ mIoU for SAM-H. This further validates that open-vocabulary semantic segmentation methods trained on RGB images are unsuitable for RGB-T images. Furthermore, we present the qualitative experimental results of OpenRSS on the Freiburg Thermal dataset in Fig. 4. + +![](images/b5a36f6a248b47730193dacdee5984d2b242f73af4118f603193569a831f7ebc.jpg) +Fig. 4: Qualitative results of OpenRSS on Freiburg Thermal dataset. From left to right: RGB images, Thermal images, ground-truths, OpenRSS results. + +Table 2: Ablation studies of TIP module Table 3: Ablation studies of attn type in OpenRSS. in TIP module. + +
MethodMFMSFT
mIoU↑
TIP-empty10.2016.213.90
TIP-shallow19.9826.8325.44
TIP-deep(ours)22.8128.5827.55
attn-fovea18.4523.0118.58
attn-msa22.5421.7221.56
attn-cor(ours)22.8128.5827.55
+ +# 4.4 Ablation Study + +Importance of TIP. Tab. 2 shows the ablation experiments of the proposed TIP module. To validate the effectiveness of the TIP module, we replace it with additive fusion, resulting in an average mIoU performance drop of $16.21\%$ . Furthermore, we discuss the specific placement of the TIP module. We only position the TIP module before the first Transformer layer, called TIP-shallow. The results indicate that placing the TIP module at each Transformer layer yields the best performance. Finally, we conduct ablation experiments on the correlative self-attention within the TIP module. As shown in Tab. 3, the results demon + +strate the effectiveness of correlative self-attention for dense tasks by comparing it to spatial fovea attention and the original multi-head self-attention. + +Importance of DLoRA. Tab. 4 shows the ablation experiments of the proposed DLoRA. By manually setting the ranks to 1, 2, 4, and 8, the performance degradation of average mIoU is observed to be $4.69\%$ , $2.55\%$ , $3.03\%$ , and $1.36\%$ , respectively. This demonstrates the significance of DLoRA while also avoiding spending considerable time selecting an optimal low-rank space. More detailed results and analysis, including failure cases, are discussed in the supplementary material. + +Table 4: Ablation of rank in DLoRA. Manually set rank vs. dynamically adaptive rank. The bold font highlights the best result in each column. + +
MethodsMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
OpenRSS-rank-129.2020.6137.4324.6928.4319.57
OpenRSS-rank-226.6121.7835.7725.9832.6023.51
OpenRSS-rank-425.3821.6336.5327.7329.6420.49
OpenRSS-rank-827.2622.5137.0726.2835.2326.06
OpenRSS-DLoRA(ours)30.7322.8137.8828.5837.2527.55
+ +# 5 Conclusion + +This work proposes the first open-vocabulary RGB-T semantic segmentation framework: OpenRSS. Our framework achieves the open-vocabulary capability by jointly utilizing the CLIP and the modified SAM model. In addition, when fusing RGB and thermal modality information, OpenRSS only needs to add a few learnable parameters to SAM to achieve effective multi-modal information fusion. Extensive experiments show that the proposed framework significantly outperforms previous methods on three RGB-T semantic segmentation benchmarks, validating the effectiveness and generalization of OpenRSS. Our work provides a feasible solution to open-vocabulary semantic segmentation with RGB-T images and a strong baseline for the community to advance future research. + +Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grant 62203337; in part by the Hubei Province Natural Science Foundation, China, under Grant 2022CFC074; and in part by the Open Fund of the Hubei Garment Informatization Engineering Technology Research Center, Wuhan Textile University, under Grant 2022HBCI04. + +# References + +1. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987) +2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Advances in Neural Information Processing Systems 32 (2019) +3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) +4. Deng, F., Feng, H., Liang, M., Wang, H., Yang, Y., Gao, Y., Chen, J., Hu, J., Guo, X., Lam, T.L.: Fanet: Feature-enhanced attention network for rgb-thermal real-time semantic segmentation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4467-4473. IEEE (2021) +5. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022) +6. Dong, S., Feng, Y., Yang, Q., Huang, Y., Liu, D., Fan, H.: Efficient multimodal semantic segmentation via dual-prompt learning. arXiv preprint arXiv:2312.00360 (2023) +7. Fu, Y., Xiang, T., Jiang, Y.G., Xue, X., Sigal, L., Gong, S.: Recent advances in zero-shot recognition. arXiv preprint arXiv:1710.04837 (2017) +8. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017) +9. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017) +10. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +1. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1037-1045 (2015) +2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +3. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) +4. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019) +5. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497-6504 (2021) +6. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) + +17. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) +18. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022) +19. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) +20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023) +21. Liang, Y., Wakaki, R., Nobuhara, S., Nishino, K.: Multimodal material segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19800-19808 (2022) +22. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015) +23. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article 2, 13 (2023) +24. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) +25. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016) +26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +27. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. arXiv preprint arXiv:2303.17386 (2023) +28. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019) +29. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2020) +30. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 (2022) +31. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020) +32. Wang, F., Mei, J., Yuille, A.: Sclip: Rethinking self-attention for dense vision-language inference. arXiv preprint arXiv:2312.01597 (2023) +33. Wang, K., Li, C., Tu, Z., Luo, B.: Unified-modal salient object detection via adaptive prompt learning. arXiv preprint arXiv:2311.16835 (2023) +34. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023) + +35. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022) +36. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023) +37. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022) +38. Zhang, J., Liu, H., Yang, K., Hu, X., Liu, R., Stiefelhagen, R.: Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems (2023) +39. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023) +40. Zhang, Q., Zhao, S., Luo, Y., Zhang, D., Huang, N., Han, J.: Abmdrnet: Adaptive-weighted bi-directional modality difference reduction network for rgb-t semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2633-2642 (2021) +41. Zhao, H., Qi, X., Shen, X., Shi, J., Jia, J.: Icnet for real-time semantic segmentation on high-resolution images. In: Proceedings of the European conference on computer vision (ECCV). pp. 405-420 (2018) +42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) +43. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +44. Zhou, W., Dong, S., Xu, C., Qian, Y.: Edge-aware guidance fusion network for rgb-thermal scene parsing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3571-3579 (2022) +45. Zhu, J., Lai, S., Chen, X., Wang, D., Lu, H.: Visual prompt multi-modal tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9516-9526 (2023) \ No newline at end of file diff --git a/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/images.zip b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..749155dad5536908f99b65d245c72c8b973a860d --- /dev/null +++ b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47d847dc7611f248245a5a4d81c01778a8c90942a9ef30df7ef501ef0440420b +size 420803 diff --git a/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/layout.json b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e8623f05194b8aa280c7fd04cf89740ee1204da8 --- /dev/null +++ b/2024/Open-Vocabulary RGB-Thermal Semantic Segmentation/layout.json @@ -0,0 +1,8433 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 454, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 454, + 147 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 454, + 147 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": "Guoqiang Zhao" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Junjie Huang" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Xiaoyun Yan" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Zhaojing Wang" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Junwei Tang" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Yangjun Ou" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Xinrong Hu" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", and Tao Peng" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 201, + 475, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 475, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 475, + 223 + ], + "type": "text", + "content": "1 School of Computer Science and Artificial Intelligence, Wuhan Textile University zgq_lyy@163.com, xyyan@wtu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 224, + 459, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 224, + 459, + 234 + ], + "spans": [ + { + "bbox": [ + 153, + 224, + 459, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 224, + 459, + 234 + ], + "type": "text", + "content": " Hubei Garment Informatization Engineering Technology Research Center" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "spans": [ + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "text", + "content": "Abstract. RGB-Thermal (RGB-T) semantic segmentation is an important research branch of multi-modal image segmentation. The current RGB-T semantic segmentation methods generally have two unsolved and typical shortcomings. First, they do not have the open-vocabulary recognition ability, which significantly limits their application scenarios. Second, when fusing RGB and thermal images, they often need to design complex fusion network structures, which usually results in low network training efficiency. We present OpenRSS, the Open-vocabulary RGB-T Semantic Segmentation method, to solve these two disadvantages. To our knowledge, OpenRSS is the first RGB-T semantic segmentation method with open-vocabulary segmentation capability. OpenRSS modifies the basic segmentation model SAM for RGB-T semantic segmentation by adding the proposed thermal information prompt module and dynamic low-rank adaptation strategy to SAM. These designs effectively fuse the RGB and thermal information, but with much fewer trainable parameters than other methods. OpenRSS achieves the open-vocabulary capability by jointly utilizing the vision-language model CLIP and the modified SAM. Through extensive experiments, OpenRSS demonstrates its effective open-vocabulary semantic segmentation ability on RGB-T images. It outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: " + }, + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "inline_equation", + "content": "+12.1\\%" + }, + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "text", + "content": " mIoU on the MFNet dataset, " + }, + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "inline_equation", + "content": "+18.4\\%" + }, + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "text", + "content": " mIoU on the MCubeS dataset, and " + }, + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "inline_equation", + "content": "+21.4\\%" + }, + { + "bbox": [ + 160, + 266, + 452, + 529 + ], + "type": "text", + "content": " mIoU on the Freiburg Thermal dataset. Code will be released at https://github.com/SXDR/OpenRSS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 540, + 452, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 540, + 452, + 562 + ], + "spans": [ + { + "bbox": [ + 162, + 540, + 452, + 562 + ], + "type": "text", + "content": "Keywords: Open-vocabulary " + }, + { + "bbox": [ + 162, + 540, + 452, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 162, + 540, + 452, + 562 + ], + "type": "text", + "content": " RGB-T semantic segmentation " + }, + { + "bbox": [ + 162, + 540, + 452, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 162, + 540, + 452, + 562 + ], + "type": "text", + "content": " Prompt learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 584, + 229, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 584, + 229, + 596 + ], + "spans": [ + { + "bbox": [ + 133, + 584, + 229, + 596 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 610, + 481, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 610, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 133, + 610, + 481, + 645 + ], + "type": "text", + "content": "Existing deep learning-based semantic segmentation methods [22, 25, 41] mainly rely on single-modal RGB images, which are effective only under well-lit conditions during the daytime and perform poorly in challenging lighting conditions" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 653, + 233, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 233, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 233, + 665 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "(e.g., nighttime, foggy weather). With the popularity of thermal cameras, researchers have recently started incorporating thermal images as a supplementary modality into RGB semantic segmentation [8], known as RGB-T semantic segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 480, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 480, + 262 + ], + "type": "text", + "content": "With the advancement of deep learning, current RGB-T semantic segmentation models mainly rely on supervised training with a predefined set of fixed categories, which cannot generalize to categories beyond the training set. This shortcoming significantly limits the application scenarios of current RGB-T semantic segmentation methods. In contrast, humans understand scenes in an open-vocabulary manner and recognize around 30,000 basic categories [1, 7]. Moreover, humans can further identify new categories through language descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 445 + ], + "type": "text", + "content": "Recently, large-scale vision-language models, represented by CLIP [26], achieve image-level classification across arbitrary categories, i.e., open-vocabulary classification. However, since the CLIP model is trained using image-level contrastive learning, how to effectively utilize CLIP in semantic segmentation is still an open problem. Some researchers adopt a two-stage training paradigm [5,20,37]: first, using a model to generate category-agnostic mask proposals, and then a pre-trained CLIP model for open-vocabulary classification on each mask. While others choose to train a side network [36] to integrate pre-trained CLIP model features into the side network for mask recognition. The promising performance of CLIP prompts us to explore its applicability in RGB-T semantic segmentation. But unlike the methods mentioned above, in this paper, we jointly utilize the basic segmentation model SAM [16] and CLIP through an elegant way to achieve open-vocabulary semantic segmentation, which effectively takes advantage of well-learned visual features and high segmentation performance of SAM and powerful open-vocabulary classification ability of CLIP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 665 + ], + "type": "text", + "content": "While SAM has learned a powerful and general visual segmentation model using the largest segmentation dataset to date, SA-1B, it is still unclear how to effectively transfer its knowledge from RGB images to RGB-T images. The most direct strategy is to train the SAM from scratch, specifically for RGB-T images. However, on the one hand, full model fine-tuning is time-consuming, inefficient, and burdensome regarding parameter storage. On the other hand, with the limited quantity of RGB-T semantic segmentation datasets, it is easy to produce over-fitting results. Thus, a natural question arises: is there a more efficient and effective method to enable SAM models trained on RGB images to be utilized for RGB-T images? Our answer to this question is affirmative, and we provide a solution with the paradigm of prompt learning. Prompt learning, also known as prompt tuning, originated from natural language processing (NLP) [17, 19], aims to inject text prompts into language models to transfer the knowledge of foundation models to various downstream tasks. Inspired by this, researchers apply prompt tuning to visual tasks [3,6,13,33,42,43,45] by freezing the entire foundation model and adding some learnable parameters only to the input side to learn effective visual prompts. Intuitively, there is a significant correlation between multimodal and RGB modalities, and they should share most of the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "prior knowledge about feature extraction. Following this, we propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 361 + ], + "type": "text", + "content": "Unlike other RGB-T semantic segmentation methods that always need complex fusion networks to fuse RGB and thermal information, which inevitably adds considerable computation and training burden to their approaches. In OpenRSS, only a small amount of parameters are trainable, which makes the training of OpenRSS quite efficient and also decreases the risk of over-fitting. During the training, OpenRSS freezes the entire image encoder of SAM and only learns visual prompts based on RGB and thermal modality. Specifically, OpenRSS freezes the image encoder of SAM and adds several simple and efficient Thermal Information Prompt (TIP) modules to SAM to produce visual prompts to learn the complementarity between RGB and thermal modality effectively. Then, in the mask decoder of SAM, we align pixel embeddings with the text embeddings corresponding to the semantic classes generated by CLIP. Due to the flexible label representations provided by the text encoder, OpenRSS can perform open-vocabulary segmentation without requiring additional training samples. Additionally, we employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to fine-tune the image encoder of SAM. For fairness and reproducibility, our research is based on the officially released SAM and CLIP models. The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 388, + 480, + 594 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 388, + 479, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 479, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 479, + 435 + ], + "type": "text", + "content": "- We propose OpenRSS, the first open-vocabulary RGB-T semantic segmentation model, where, with the aid of prompt learning and dynamic low-rank adaptation strategy, existing foundation models can effectively adapt to open-vocabulary RGB-T semantic segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 441, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 441, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 441, + 480, + 487 + ], + "type": "text", + "content": "- We introduce a Thermal Information Prompt (TIP) module that generates simple yet efficient visual prompts to fuse RGB and thermal information, avoiding the complex and redundant additional branches designed in previous methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 494, + 479, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 479, + 517 + ], + "type": "text", + "content": "- We employ a Dynamic Low-Rank Adaptation (DLoRA) strategy to optimize a series of parameters in the image encoder of SAM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that OpenRSS outperforms other state-of-the-art RGB open-vocabulary semantic segmentation methods on multiple RGB-T semantic segmentation benchmarks: " + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "inline_equation", + "content": "+12.1\\%" + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "text", + "content": " mIoU on the MFNet dataset, " + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "inline_equation", + "content": "+18.4\\%" + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "text", + "content": " mIoU on the MCubeS dataset, and " + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "inline_equation", + "content": "+21.4\\%" + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "text", + "content": " mIoU on the Freiburg Thermal dataset, while maintaining parameter efficiency (only " + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 138, + 523, + 480, + 594 + ], + "type": "text", + "content": " trainable parameters)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "The remainder of this article is organized as follows. We review related work in Sec. 2. In Sec. 3, we provide a detailed description of our network architecture. Sec. 4 present experimental results and discussions. Finally, we summarize the work in Sec. 5." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 320, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 320, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 320, + 152 + ], + "type": "text", + "content": "2.1 RGB-T Semantic Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 481, + 361 + ], + "type": "text", + "content": "For complex scenes with challenging lighting conditions, thermal images are highly complementary to RGB images. Existing RGB-T semantic segmentation methods focus on designing complex cross-modal fusion paradigms. We briefly review RGB-T semantic segmentation from three paradigms: encoder fusion, decoder fusion, and feature fusion. For the encoder fusion paradigm, RTFNet [28] and FuseSeg [29] integrate thermal features into RGB features by element-wise summation. FEANet [4] introduces a two-stage feature-enhanced attention network, which utilizes channel and spatial attention to fuse RGB and thermal information in a complementary way. For the decoder fusion paradigm, MFNet [8] uses two identical encoders to extract RGB and thermal features and feed them into the decoder for integration. For the feature fusion paradigm, EGFNet [44] embeds prior edge information into feature maps for cross-modal fusion. ABMDRNet [40] employs a bi-directional image-to-image translation based method to reduce the modality differences, followed by channel-wise weighted fusion for final prediction. CMX [38] first combines features from other modalities in both spatial and channel directions to calibrate the current modality's features, then fuses them for cross-modal fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 481, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 481, + 423 + ], + "type": "text", + "content": "Unlike the methods mentioned above, which always have to train complex additional network branches to extract thermal-related features, OpenRSS only needs to learn a small amount of trainable parameters to inject thermal-related information into the image encoder of SAM, whose parameters are frozen during training. This merit gives OpenRSS high parameters and training efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 441, + 257, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 257, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 257, + 452 + ], + "type": "text", + "content": "2.2 Large-scale Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 555 + ], + "type": "text", + "content": "The Transformer's robust scalability allows large-scale models to have hundreds of billions of parameters. Large-scale models, such as BERT [14], Instruct-GPT [24], GPT-4 [23], etc, achieve groundbreaking progress in natural language processing tasks. Large-scale models like SAM [16] and SegGPT [34] demonstrate outstanding zero-shot generalization capabilities in computer vision. Recently, large-scale vision-language models represented by CLIP [26] and ALIGN [12] can recognize arbitrary categories described in the text, known as open-vocabulary classification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "content": "Our work further explores how to effectively utilize pre-trained visual and visual-language large-scale models to perform open-vocabulary semantic segmentation on RGB-T images, and accordingly proposes our solution: the OpenRSS model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 621, + 284, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 284, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 284, + 634 + ], + "type": "text", + "content": "2.3 Visual Prompt Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In computer vision, downstream tasks are always executed by fine-tuning large pre-trained models comprehensively for a long time. Specifically, different down-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "stream data are used to retrain downstream tasks to update all parameters. Although this approach yields satisfactory performance, it is inefficient regarding parameter usage and ultimately loses the universal knowledge learned by the pretrained models. Prompt learning (also known as prompt tuning) is recognized as a trade-off between the training burden and the final performance in the downstream tasks and has recently garnered widespread attention. It initially shone in natural language processing before being introduced to various computer vision tasks. For example, VPT [13] presents a set of learnable parameters for Transformer encoders and significantly outperforms full fine-tuning on multiple downstream tasks. ViPT [45] introduces prompts for pre-trained single-modal trackers, making them applicable to multimodal tracking tasks. UniSOD [33] designs a switchable prompt generation block to interact with single-modal or multimodal inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 321 + ], + "type": "text", + "content": "In previous works, they often used simple " + }, + { + "bbox": [ + 130, + 272, + 482, + 321 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 272, + 482, + 321 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 272, + 482, + 321 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 272, + 482, + 321 + ], + "type": "text", + "content": " convolutions to generate prompts. In contrast, we specifically design a Thermal Information Prompt (TIP) module with a Transformer block inside to generate prompts for dense prediction, enabling the pre-trained SAM model to adapt to RGB-T images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 338, + 369, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 369, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 369, + 350 + ], + "type": "text", + "content": "2.4 Open-vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 525 + ], + "type": "text", + "content": "Open-vocabulary semantic segmentation refers to identifying pixel-level categories in images through arbitrary text descriptions. In initial works like ZS3Net [2], generative models are used to synthesize pixel-level features through word embeddings of unseen classes. GroupVit [35] introduces a grouping mechanism to cluster images automatically under text supervision. With the remarkable performance of pre-trained large-scale vision-language models in open-vocabulary classification, researchers begin exploring their performance in open-vocabulary semantic segmentation. These approaches can be broadly categorized into the following three types: 1)LSeg [18] aligns pixel-level embeddings with corresponding semantic class text embeddings for pixel-level classification. 2) [5, 20, 37] adopt a two-stage segmentation framework, where MaskFormer first generates a set of category-agnostic mask proposals, followed by CLIP performing open-vocabulary classification on each mask. 3)SAN [36] designs a side network to enable CLIP to perceive the predictions and recognition of masks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 525, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 481, + 550 + ], + "type": "text", + "content": "Unlike these methods, OpenRSS is an end-to-end network that elegantly utilizes SAM and CLIP to achieve open-vocabulary semantic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 567, + 203, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 567, + 203, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 567, + 203, + 580 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we present OpenRSS, an end-to-end framework that combines pretrained large-scale models SAM and CLIP to perform open-vocabulary RGB-T semantic segmentation. Instead of fully fine-tuning all the parameters of the models, OpenRSS only learns a small amount of prompt information. Furthermore, we introduce trainable dynamic LoRA bypasses to adjust the frozen image encoder." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 479, + 340 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 479, + 340 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 479, + 340 + ], + "type": "image", + "image_path": "b4f45d9c8f5e3fc70233e6904cf555900af7a336c1cfbfbf4da333bdcad0341a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 351, + 482, + 418 + ], + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 418 + ], + "type": "text", + "content": "Fig. 1: Overview architecture of our OpenRSS. It inputs a pair of RGB and thermal images and generates corresponding RGB and thermal tokens through patch embeddings. Then, it utilizes the frozen Transformer encoder blocks of SAM for feature extraction. The TIP modules are inserted into the image encoder of SAM to learn effective visual prompts. DLoRA further adjusts the frozen image encoder. Finally, the mask decoder aligns image and text embeddings to generate the final segmentation map." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 442, + 208, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 208, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 208, + 453 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": "The overall architecture of our OpenRSS is shown in Fig. 1. Formally, OpenRSS incorporates an additional thermal image input stream instead of a thermal feature extraction branch, which is fed into the network simultaneously with the RGB stream. Given a pair of RGB and thermal images, OpenRSS first projects the two information streams " + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "inline_equation", + "content": "(T_{RGB}" + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "inline_equation", + "content": "T_{TH})" + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": " into patch embedding layers separately. After projection, we obtain two C-dimensional embeddings, which we denote as RGB tokens " + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{rgb}" + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": " and thermal tokens " + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{th}" + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": ", respectively. Then, " + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{rgb}" + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{th}" + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": " are input into the TIP module to generate thermal information prompts. The learned prompt information is added to the original RGB information stream " + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{rgb}" + }, + { + "bbox": [ + 130, + 464, + 482, + 594 + ], + "type": "text", + "content": " in a residual manner before being fed into the fully frozen image encoder of SAM:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 596, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 596, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 221, + 596, + 481, + 611 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} ^ {i} = \\boldsymbol {I} _ {r g b} ^ {i} + \\boldsymbol {T} ^ {i + 1}, \\quad i = 0, 1, \\dots , L - 1, \\tag {1}", + "image_path": "1fad65a592e30db873f3453ba7cb6bb9572bab0d26437b0c6f0164f33c6d41c4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{I}^i" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the final sequence of image tokens fed into the Transformer layers of the encoder, and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{T}^{i + 1}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the prompt tokens sequence from the " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "th TIP module. By directly adding the thermal information prompts to the intermediate features generated by SAM, our OpenRSS can effortlessly perform downstream" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "tasks quickly. Furthermore, we also design a trainable dynamic LoRA bypass for each Transformer layer. For each bypass, the features generated by the Transformer layers are first mapped to a low-rank space and then projected back to the initial dimension, aligning it with the original output dimension of the frozen Transformer layers. We use a points grid prompt for the prompt encoder, significantly enhancing SAM's attention to every detail in the image compared to the default embeddings. For the text encoder, we use the frozen text encoder of CLIP. During training, we employ manually crafted prompts for the text encoder, consisting of a natural sentence and category names. However, any sentence can be used during inference to achieve the goal of open-vocabulary. We modify the original mask decoder of SAM to align its dimensionality with the feature embeddings of the text encoder. It simultaneously takes the embeddings from the prompt encoder and the text encoder as input, generating the final segmentation map." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 300, + 309, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 309, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 309, + 312 + ], + "type": "text", + "content": "3.2 Thermal Information Prompt" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "type": "text", + "content": "In recent years, following the success of text prompts in NLP, some research has begun to explore the introduction of visual prompts into frozen pre-trained models. Prompt tuning is achieved through some learnable parameters and yields promising results on many downstream visual tasks. A key to our OpenRSS is to generate appropriate thermal information prompts to guide the network in effective open-vocabulary RGB-T semantic segmentation. As shown in Fig. 1, the TIP module is inserted before each Transformer layer. Formally, the thermal information prompts are generated from the token sequences " + }, + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\{I_{rgb}^{0}, I_{th}^{0}\\}" + }, + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "type": "text", + "content": " of both modalities. The TIP module learns the complementarity between these two modalities, which can be represented as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 444, + 480, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 444, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 216, + 444, + 480, + 458 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} ^ {i} = T I P ^ {i} \\left(\\boldsymbol {I} ^ {i - 1}, \\boldsymbol {T} ^ {i - 1}\\right), \\quad i = 1, 2, \\dots , L, \\tag {2}", + "image_path": "d573aac670437144a3a21ee65243cd7aef52a3a96cce5f6c0ec7648b83ced4b8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "inline_equation", + "content": "TIP^i" + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "text", + "content": "th TIP module. Specifically, " + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "inline_equation", + "content": "I^0 = I_{rqb}^0" + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "inline_equation", + "content": "T^0 = I_{th}^0" + }, + { + "bbox": [ + 130, + 465, + 469, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 538 + ], + "type": "text", + "content": "The detailed design of the TIP is shown in Fig. 2. TIP has two input branches, each receiving token sequences from two input streams. Specifically, considering the redundancy of single-modal features and the lightness of the prompt block, we use two " + }, + { + "bbox": [ + 130, + 479, + 482, + 538 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 479, + 482, + 538 + ], + "type": "text", + "content": " convolution operations to project each input stream to the size of " + }, + { + "bbox": [ + 130, + 479, + 482, + 538 + ], + "type": "inline_equation", + "content": "[ \\alpha \\times H \\times W ]" + }, + { + "bbox": [ + 130, + 479, + 482, + 538 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 545, + 480, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 545, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 220, + 545, + 480, + 559 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} _ {R G B} = c _ {1} \\left(\\boldsymbol {I} ^ {i - 1}\\right), \\quad \\boldsymbol {Z} _ {T H} = c _ {2} \\left(\\boldsymbol {T} ^ {i - 1}\\right), \\tag {3}", + "image_path": "ba9685ae0b8ee517f9bdf497aff156a5e334afc767899a356cd4ba18fa09c1ac.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "text", + "content": " is set to 8 in all TIP modules. " + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "inline_equation", + "content": "c_{1}(\\cdot)" + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "inline_equation", + "content": "c_{2}(\\cdot)" + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "text", + "content": " convolution operations. Then, inspired by [32], we use a Transformer block with correlative self-attention for dense prediction to generate enhanced embeddings for " + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{RGB}^{e}" + }, + { + "bbox": [ + 130, + 565, + 482, + 613 + ], + "type": "text", + "content": ", as shown in Fig. 2. The formula for Correlative_attn is expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 201, + 620, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 620, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 201, + 620, + 480, + 634 + ], + "type": "interline_equation", + "content": "\\text {C o r r e l a t i v e} \\_ \\text {a t t n} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {X} \\boldsymbol {W} _ {q} \\boldsymbol {W} _ {q} ^ {T} \\boldsymbol {X} ^ {T} / \\tau\\right), \\tag {4}", + "image_path": "36a02c99181a4a7e46ab0396a42997af16fd76b9538ee416fe353af7afd21862.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "content": " denotes the input and " + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "content": " is the projection matrix. The temperature coefficient " + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "content": " is by default set to " + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\sqrt{d}" + }, + { + "bbox": [ + 130, + 639, + 482, + 665 + ], + "type": "text", + "content": " following traditional self-attention. Finally," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 448, + 103 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 474, + 283 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 474, + 283 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 474, + 283 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 474, + 283 + ], + "type": "image", + "image_path": "00ed787fe22808959742094ec4063222c64a618e11002e7c4a6bc6edd89762d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Overview architecture of our TIP module. Multi-modal input flows are sent separately to the " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": " convolutions for distilling to the lower-dimension latent space. After passing through the Correlative Self-Attention block, the RGB embeddings are fused with thermal embeddings through element-wise addition. Finally, the multi-modal hybrid embeddings are mapped to the original dimension through a " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": " convolutional layer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 382, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 407 + ], + "type": "text", + "content": "we obtain the mixed-modal embeddings by element-wise addition and project them back to the original dimension by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 413, + 481, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 413, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 250, + 413, + 481, + 427 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} ^ {i} = c _ {3} \\left(\\boldsymbol {Z} _ {R G B} ^ {e} + \\boldsymbol {Z} _ {T H}\\right), \\tag {5}", + "image_path": "6f1bf9022e8a7cc53a1a7d529164261c1f14d215528faa8e1199057d7c0e2091.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "inline_equation", + "content": "c_{3}(\\cdot)" + }, + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "text", + "content": " is the same as " + }, + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "inline_equation", + "content": "c_{1}(\\cdot)" + }, + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "inline_equation", + "content": "c_{2}(\\cdot)" + }, + { + "bbox": [ + 131, + 433, + 316, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 460, + 293, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 293, + 473 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 293, + 473 + ], + "type": "text", + "content": "3.3 DLoRA in Image Encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 479, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 548 + ], + "type": "text", + "content": "In OpenRSS, the thermal information prompts generated by TIP are fed into the frozen pre-trained image encoder of SAM to extract features. However, this may not be optimal due to the differences between the RGB and thermal modalities. Inspired by the idea from [10, 30], we design a Dynamic Low-Rank Adaptation (DLoRA) strategy to adjust the frozen image encoder of SAM to address this issue." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": "Given an input token sequence " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{B \\times N \\times C_{in}}" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": ", the output token sequence " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{X}' \\in \\mathbb{R}^{B \\times N \\times C_{out}}" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": " is obtained through a projection layer " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{C_{in} \\times C_{out}}" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": ". LoRA assumes the updates to the weights have a low \"intrinsic rank\" during adaptation, thus proposing a low-rank decomposition to constrain its updates. Following this strategy, the Transformer layer is first frozen to fix " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": ", and then a bypass is added to achieve low-rank decomposition. This bypass consists of two linear layers " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{C_{in} \\times r}" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{r \\times C_{out}}" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "r \\ll \\min(C_{in}, C_{out})" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": ". Therefore, the update of " + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 549, + 483, + 646 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 244, + 651, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 651, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 244, + 651, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {W}} = \\boldsymbol {W} + \\Delta \\boldsymbol {W} = \\boldsymbol {W} + \\boldsymbol {B} \\boldsymbol {A}. \\tag {6}", + "image_path": "269ffff4a8e8ab385cce447273c4b8173fa74c04e375c19aabf46dc2b9fe5d97.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 118, + 433, + 253 + ], + "blocks": [ + { + "bbox": [ + 179, + 118, + 433, + 253 + ], + "lines": [ + { + "bbox": [ + 179, + 118, + 433, + 253 + ], + "spans": [ + { + "bbox": [ + 179, + 118, + 433, + 253 + ], + "type": "image", + "image_path": "3fd25287071bda43007e633aad1ab66e0bbaf4f3f3db226bcd14f508a0793673.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 266, + 480, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 266, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 480, + 289 + ], + "type": "text", + "content": "Fig. 3: Overview architecture of our DLoRA. We apply the DLoRA layer to the q and v projection layers of each of the Transformer layers in the image encoder of SAM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 311, + 479, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 311, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 131, + 311, + 479, + 347 + ], + "type": "text", + "content": "We maintain consistency with [39] and apply LoRA to the query and value projection layers of the multi-head self-attention. Therefore, the modified query and value projection layers will become:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 237, + 355, + 481, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 355, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 237, + 355, + 481, + 375 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Q} = \\hat {\\boldsymbol {W}} _ {q} \\boldsymbol {X} = \\boldsymbol {W} _ {q} \\boldsymbol {X} + \\boldsymbol {B} _ {q} \\boldsymbol {A} _ {q} \\boldsymbol {X}, \\tag {7}", + "image_path": "4a68fb89b1e87506e86f102b66d886d3fdde987bd1642482ff2b30dffe3b73f1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "text", + "content": " are frozen from image encoder of SAM, and " + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_q, \\mathbf{B}_q, \\mathbf{A}_v" + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_v" + }, + { + "bbox": [ + 131, + 390, + 479, + 413 + ], + "type": "text", + "content": " are trainable parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "text", + "content": "While LoRA achieves remarkable results, it heavily relies on the choice of rank " + }, + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "text", + "content": ", requiring a manual and iterative search for the optimal rank. Therefore, we propose a Dynamic LoRA (DLoRA). Building upon the original LoRA, we increase the number of bypasses (as shown in Fig. 3) from 1 to 4, each with different ranks " + }, + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "text", + "content": " and weights " + }, + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 414, + 480, + 498 + ], + "type": "text", + "content": ". Through experimental observation, we find that our DLoRA can surpass the optimal rank found through manual iterative training in just one training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 515, + 305, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 305, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 305, + 528 + ], + "type": "text", + "content": "3.4 Generate Segmentation Map" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "type": "text", + "content": "Prompt Encoder and Mask Decoder. SAM passes the image embeddings generated by its image encoder and the prompts generated by its prompt encoder to its mask decoder for dense prediction. The types of prompts include point, box, and mask. SAM represents the first two sparse prompts as positional encodings and feeds them into the Transformer layers of the mask decoder. SAM performs convolution operations for dense prompts such as masks and then adds them element-wise with the image embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Due to SAM's prompt mechanism, it is susceptible to the position and quantity of prompt points. Therefore, in OpenRSS, we use a points grid as prompts to ensure that SAM does not overlook any detail in the image. SAM's mask decoder comprises a lightweight Transformer layer and a segmentation head. OpenRSS" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "modifies this segmentation head to align the dimension of the feature map with the dimension of CLIP's text features, enabling pixel-level classification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 479, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 479, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 479, + 256 + ], + "type": "text", + "content": "Pixel-level Classification via CLIP. Existing CLIP-based open-vocabulary segmentation methods mainly adopt a two-stage region recognition approach. In the first stage, an independent model is trained to generate a set of category-agnostic mask proposals. In the second stage, visual language pre-trained models such as CLIP are used to identify the category of each mask individually. However, this two-stage approach heavily relies on the quality of the masks generated in the first stage. Moreover, the cropped masks, due to missing backgrounds and varying sizes, are not suitable for recognition by CLIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 327 + ], + "type": "text", + "content": "To address the problem of region recognition, our OpenRSS adopts pixel-level recognition. We define the feature map outputted by the mask decoder as " + }, + { + "bbox": [ + 130, + 258, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times D}" + }, + { + "bbox": [ + 130, + 258, + 479, + 327 + ], + "type": "text", + "content": ", and the class embeddings generated by the CLIP text encoder as " + }, + { + "bbox": [ + 130, + 258, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{E} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 258, + 479, + 327 + ], + "type": "text", + "content": ". Since they have the same embedding dimension, we can directly link them together through the inner product to obtain the final segmentation map:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 278, + 332, + 479, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 332, + 479, + 345 + ], + "spans": [ + { + "bbox": [ + 278, + 332, + 479, + 345 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} = \\boldsymbol {F} \\times \\boldsymbol {E} ^ {T}, \\tag {8}", + "image_path": "4b75306b1ddaae1f493dde1109ea7bd70f03f06c4e5bd23b06759a582b3cbd24.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 479, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 479, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 479, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 353, + 479, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{S} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times N}" + }, + { + "bbox": [ + 130, + 353, + 479, + 379 + ], + "type": "text", + "content": ". This is the standard output of semantic segmentation, hence compatible with mainstream semantic segmentation evaluations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 401, + 230, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 401, + 230, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 230, + 415 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 430, + 320, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 320, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 320, + 441 + ], + "type": "text", + "content": "4.1 Dataset and Evaluation Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 453, + 479, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 479, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 479, + 489 + ], + "type": "text", + "content": "We conduct experiments on 4 datasets: MF dataset [9], KP dataset [11], MCubeS dataset [21] and Freiburg Thermal dataset [31]. All models are trained on the training set of the KP dataset and evaluated on other datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 511, + 479, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 479, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 479, + 571 + ], + "type": "text", + "content": "MF Dataset. The dataset consists of urban street scenes captured using an InfRec R500 camera. It includes 1569 pairs of RGB and thermal images, each with a resolution of " + }, + { + "bbox": [ + 130, + 511, + 479, + 571 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 130, + 511, + 479, + 571 + ], + "type": "text", + "content": " pixels. Among these are 820 pairs of daytime images and 749 pairs of nighttime images. The dataset consists of eight manually labeled classes, as well as a background class." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "text", + "content": "KP Dataset. The dataset is an RGB-T paired urban driving scene dataset, providing 95328 pairs of RGB and thermal images with a resolution of " + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 640" + }, + { + "bbox": [ + 130, + 593, + 479, + 665 + ], + "type": "text", + "content": ". Initially, the KP dataset only provides bounding box labels for object detection tasks. However, fortunately, [15] provides semantic segmentation task labels for 503 daytime and 447 nighttime images. Furthermore, [27] partitions these 950 images into training, validation, and test sets." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "text", + "content": "MCubeS Dataset. MCubeS is a dataset with pairs of RGB, Near-Infrared (NIR), Degree of Linear Polarization (DoLP), and Angle of Linear Polarization (AoLP) to study semantic material segmentation of 20 classes with a resolution of " + }, + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "1024 \\times 1224" + }, + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 179, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 179, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 179, + 479, + 239 + ], + "type": "text", + "content": "Freiburg Thermal Dataset. The dataset includes 12051 daytime and 8596 nighttime time-synchronized RGB-T image pairs captured in rural and urban environments with a resolution of " + }, + { + "bbox": [ + 131, + 179, + 479, + 239 + ], + "type": "inline_equation", + "content": "650 \\times 1920" + }, + { + "bbox": [ + 131, + 179, + 479, + 239 + ], + "type": "text", + "content": ". However, only the test set includes 32 daytime and 32 nighttime images, comprising 12 object classes and one background class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 255, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 479, + 289 + ], + "type": "text", + "content": "Evaluation Metrics. We use mean Accuracy (mAcc) and mean Intersection over Union (mIoU) as the evaluation metrics for our model. The formulas for calculation are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 294, + 480, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 294, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 258, + 294, + 480, + 321 + ], + "type": "interline_equation", + "content": "m A c c = \\frac {1}{n _ {c l s}} \\cdot \\sum_ {i} \\frac {n _ {i i}}{t _ {i}}, \\tag {9}", + "image_path": "4457a30ba5a02e3d6ad928280696b91695053093b70e1db59de6f9ae6f11220e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 329, + 480, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 329, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 228, + 329, + 480, + 361 + ], + "type": "interline_equation", + "content": "m I o U = \\frac {1}{n _ {c l s}} \\cdot \\sum_ {i} \\frac {n _ {i i}}{t _ {i} + \\sum_ {j} n _ {j i} - n _ {i i}}, \\tag {10}", + "image_path": "863e48f2f16a07bb091bb9b9b55d8605c2e306b8609d4a16e47f51b399dc4405.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "spans": [ + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "inline_equation", + "content": "n_{ii}" + }, + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "text", + "content": " is the number of pixels where class i is predicted as class j, " + }, + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "inline_equation", + "content": "n_{cls}" + }, + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "text", + "content": " is the total number of target classes (including the background), and " + }, + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "inline_equation", + "content": "t_i = \\sum_j n_{ij}" + }, + { + "bbox": [ + 131, + 364, + 479, + 400 + ], + "type": "text", + "content": " is the total number of pixels for target class i (ground truth labels)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 415, + 279, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 279, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 279, + 427 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 433, + 479, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 479, + 540 + ], + "type": "text", + "content": "We conduct all experiments on NVIDIA A100 GPUs. If not specified, we default to using SAM with the ViT-H backbone and CLIP with the ViT-L/16 backbone. We apply DLoRA fine-tuning to the frozen encoder Transformer layers, using four bypasses with ranks selected as 1, 2, 4, and 8, and weights as 0.1, 0.2, 0.4, and 0.3, respectively. All models are trained on the KP dataset with image flipping as a data augmentation technique. We use the AdamW optimizer with a weight decay of 0.01 and an initial learning rate of 5e-4. For the scheduler, we use the WarmupPolyLR scheduler. The loss function consists of DiceLoss and SoftCrossEntropy with weighting coefficients of 0.5 each:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 238, + 548, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 548, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 238, + 548, + 480, + 601 + ], + "type": "interline_equation", + "content": "D i c e L o s s = 1 - \\frac {2 \\sum_ {i = 1} ^ {N} y _ {i} \\widehat {y} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} \\widehat {y} _ {i} ^ {2}}, \\tag {11}", + "image_path": "a17f75dcb8f2d17e7aeb1abc9bf9a06eeb55516c35008e87165f0a118f34fddd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "inline_equation", + "content": "\\widehat{y_i}" + }, + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "text", + "content": " represent the true and predicted label values of pixel I, respectively, and " + }, + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 606, + 479, + 630 + ], + "type": "text", + "content": " represents the total number of pixels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 197, + 637, + 479, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 637, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 197, + 637, + 479, + 667 + ], + "type": "interline_equation", + "content": "\\text {S o f t C r o s s E n t r o p y l o s s} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {c} \\hat {y} _ {i j} \\log \\left(y _ {i j} ^ {d}\\right), \\tag {12}", + "image_path": "a9d674efd848b8a795149b4364b5d631ffd950e423d596da12ba1860f0ec86df.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 157, + 466, + 345 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Performance comparison with state-of-the-art methods. For a fair comparison, we reproduce their method with their officially released code. The bold font highlights the best result in each column." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 157, + 466, + 345 + ], + "lines": [ + { + "bbox": [ + 149, + 157, + 466, + 345 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 466, + 345 + ], + "type": "table", + "html": "
MethodBackboneMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
ZegFormer [5]R-5011.1110.3110.003.437.690.11
OVSeg [20]R-10110.869.4210.092.738.320.98
SAN [36]ViT-B14.644.1317.2410.107.131.31
ZegFormer [5]R-10111.5510.7110.464.117.850.79
OVSeg [20]Swin-B10.206.7816.496.5810.203.74
SAN [36]ViT-L3.810.386.904.2710.606.10
OpenRSS(ours)SAM-B22.2118.9631.7620.2716.018.14
SAM-L23.9120.0833.9524.9125.3316.65
SAM-H30.7322.8137.8828.5837.2527.55
", + "image_path": "3a9aa0ef160e807d432e243c5a4e85f9716dbd4415352a052a4da27dd48bd2da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "text", + "content": " is the number of the batch, " + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "inline_equation", + "content": "\\hat{y}_{ij}" + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "text", + "content": " is a binary indicator if class label " + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "text", + "content": " is the correct classification for pixel " + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "inline_equation", + "content": "y_{ij}^{d}" + }, + { + "bbox": [ + 130, + 372, + 482, + 420 + ], + "type": "text", + "content": " is the corresponding predicted probability be normalized to a probability distribution. These two losses are summed to obtain the overall loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 210, + 435, + 481, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 435, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 210, + 435, + 481, + 449 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = 0. 5 L _ {\\text {D i c e L o s s}} + 0. 5 L _ {\\text {S o f t C r o s s E n t r o p y l o s s}}. \\tag {13}", + "image_path": "0f43a51afb4b98165a442e488e2745ef754aa5ca81bf80a5ed8a2c3bf1d4bd1a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 477, + 266, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 477, + 266, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 266, + 490 + ], + "type": "text", + "content": "4.3 Results and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Tab. 1 presents the quantitative comparison results of our method with other state-of-the-art open-vocabulary semantic segmentation methods. We only use RGB images as input, rather than pairs of RGB and thermal images, allowing the general open vocabulary semantic segmentation in Tab. 1 to perform RGB-T semantic segmentation. From the table, it can be observed that our method achieves superior performance on both metrics across all three datasets, indicating the effectiveness of our approach. Specifically, our method surpasses other methods with an average of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "+9.17\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mAcc and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "+6.82\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mIoU for SAM-B, " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "+13.57\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mAcc and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "+11.58\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mIoU for SAM-L, and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "+20.46\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mAcc and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "+17.34\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mIoU for SAM-H. This further validates that open-vocabulary semantic segmentation methods trained on RGB images are unsuitable for RGB-T images. Furthermore, we present the qualitative experimental results of OpenRSS on the Freiburg Thermal dataset in Fig. 4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 118, + 467, + 340 + ], + "blocks": [ + { + "bbox": [ + 149, + 118, + 467, + 340 + ], + "lines": [ + { + "bbox": [ + 149, + 118, + 467, + 340 + ], + "spans": [ + { + "bbox": [ + 149, + 118, + 467, + 340 + ], + "type": "image", + "image_path": "b5a36f6a248b47730193dacdee5984d2b242f73af4118f603193569a831f7ebc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 351, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 131, + 351, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 351, + 482, + 374 + ], + "type": "text", + "content": "Fig. 4: Qualitative results of OpenRSS on Freiburg Thermal dataset. From left to right: RGB images, Thermal images, ground-truths, OpenRSS results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 147, + 439, + 479, + 525 + ], + "blocks": [ + { + "bbox": [ + 146, + 407, + 482, + 430 + ], + "lines": [ + { + "bbox": [ + 146, + 407, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 146, + 407, + 482, + 430 + ], + "type": "text", + "content": "Table 2: Ablation studies of TIP module Table 3: Ablation studies of attn type in OpenRSS. in TIP module." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 439, + 479, + 525 + ], + "lines": [ + { + "bbox": [ + 147, + 439, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 147, + 439, + 479, + 525 + ], + "type": "table", + "html": "
MethodMFMSFT
mIoU↑
TIP-empty10.2016.213.90
TIP-shallow19.9826.8325.44
TIP-deep(ours)22.8128.5827.55
attn-fovea18.4523.0118.58
attn-msa22.5421.7221.56
attn-cor(ours)22.8128.5827.55
", + "image_path": "167ba4482994e1f231b93d2c4bd86e5b7439ea8989cd0646e4bcabcc3ac3629f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 544, + 238, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 238, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 238, + 558 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Importance of TIP. Tab. 2 shows the ablation experiments of the proposed TIP module. To validate the effectiveness of the TIP module, we replace it with additive fusion, resulting in an average mIoU performance drop of " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "16.21\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ". Furthermore, we discuss the specific placement of the TIP module. We only position the TIP module before the first Transformer layer, called TIP-shallow. The results indicate that placing the TIP module at each Transformer layer yields the best performance. Finally, we conduct ablation experiments on the correlative self-attention within the TIP module. As shown in Tab. 3, the results demon" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "strate the effectiveness of correlative self-attention for dense tasks by comparing it to spatial fovea attention and the original multi-head self-attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "text", + "content": "Importance of DLoRA. Tab. 4 shows the ablation experiments of the proposed DLoRA. By manually setting the ranks to 1, 2, 4, and 8, the performance degradation of average mIoU is observed to be " + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "inline_equation", + "content": "4.69\\%" + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "inline_equation", + "content": "2.55\\%" + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "inline_equation", + "content": "3.03\\%" + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "inline_equation", + "content": "1.36\\%" + }, + { + "bbox": [ + 130, + 159, + 482, + 242 + ], + "type": "text", + "content": ", respectively. This demonstrates the significance of DLoRA while also avoiding spending considerable time selecting an optimal low-rank space. More detailed results and analysis, including failure cases, are discussed in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 154, + 294, + 459, + 415 + ], + "blocks": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "type": "text", + "content": "Table 4: Ablation of rank in DLoRA. Manually set rank vs. dynamically adaptive rank. The bold font highlights the best result in each column." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 294, + 459, + 415 + ], + "lines": [ + { + "bbox": [ + 154, + 294, + 459, + 415 + ], + "spans": [ + { + "bbox": [ + 154, + 294, + 459, + 415 + ], + "type": "table", + "html": "
MethodsMFMCubeSFT
mAcc↑mIoU↑mAcc↑mIoU↑mAcc↑mIoU↑
OpenRSS-rank-129.2020.6137.4324.6928.4319.57
OpenRSS-rank-226.6121.7835.7725.9832.6023.51
OpenRSS-rank-425.3821.6336.5327.7329.6420.49
OpenRSS-rank-827.2622.5137.0726.2835.2326.06
OpenRSS-DLoRA(ours)30.7322.8137.8828.5837.2527.55
", + "image_path": "56c1b9f56524ee7da6515364734567104cf2010d8024d2b0111e09c26ddc2e76.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": "This work proposes the first open-vocabulary RGB-T semantic segmentation framework: OpenRSS. Our framework achieves the open-vocabulary capability by jointly utilizing the CLIP and the modified SAM model. In addition, when fusing RGB and thermal modality information, OpenRSS only needs to add a few learnable parameters to SAM to achieve effective multi-modal information fusion. Extensive experiments show that the proposed framework significantly outperforms previous methods on three RGB-T semantic segmentation benchmarks, validating the effectiveness and generalization of OpenRSS. Our work provides a feasible solution to open-vocabulary semantic segmentation with RGB-T images and a strong baseline for the community to advance future research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grant 62203337; in part by the Hubei Province Natural Science Foundation, China, under Grant 2022CFC074; and in part by the Open Fund of the Hubei Garment Informatization Engineering Technology Research Center, Wuhan Textile University, under Grant 2022HBCI04." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review 94(2), 115 (1987)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "text", + "content": "2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 480, + 212 + ], + "type": "text", + "content": "3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 213, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 255 + ], + "type": "text", + "content": "4. Deng, F., Feng, H., Liang, M., Wang, H., Yang, Y., Gao, Y., Chen, J., Hu, J., Guo, X., Lam, T.L.: Fanet: Feature-enhanced attention network for rgb-thermal real-time semantic segmentation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4467-4473. IEEE (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "type": "text", + "content": "5. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11583-11592 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "text", + "content": "6. Dong, S., Feng, Y., Yang, Q., Huang, Y., Liu, D., Fan, H.: Efficient multimodal semantic segmentation via dual-prompt learning. arXiv preprint arXiv:2312.00360 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 341 + ], + "type": "text", + "content": "7. Fu, Y., Xiang, T., Jiang, Y.G., Xue, X., Sigal, L., Gong, S.: Recent advances in zero-shot recognition. arXiv preprint arXiv:1710.04837 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 384 + ], + "type": "text", + "content": "8. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "type": "text", + "content": "9. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "10. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "text", + "content": "1. Hwang, S., Park, J., Kim, N., Choi, Y., So Kweon, I.: Multispectral pedestrian detection: Benchmark dataset and baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1037-1045 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 492, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 535 + ], + "type": "text", + "content": "2. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "3. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "4. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "5. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497-6504 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "6. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "18. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 195 + ], + "type": "text", + "content": "19. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 239 + ], + "type": "text", + "content": "20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "21. Liang, Y., Wakaki, R., Nobuhara, S., Nishino, K.: Multimodal material segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19800-19808 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 308 + ], + "type": "text", + "content": "22. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 319 + ], + "type": "text", + "content": "23. OpenAI, R.: Gpt-4 technical report. arxiv 2303.08774. View in Article 2, 13 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 363 + ], + "type": "text", + "content": "24. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 397 + ], + "type": "text", + "content": "25. Paszke, A., Chaurasia, A., Kim, S., Culurciello, E.: Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "text", + "content": "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "type": "text", + "content": "27. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. arXiv preprint arXiv:2303.17386 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "text", + "content": "28. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "text", + "content": "29. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "type": "text", + "content": "30. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 564, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 480, + 597 + ], + "type": "text", + "content": "31. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "text", + "content": "32. Wang, F., Mei, J., Yuille, A.: Sclip: Rethinking self-attention for dense vision-language inference. arXiv preprint arXiv:2312.01597 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "type": "text", + "content": "33. Wang, K., Li, C., Tu, Z., Luo, B.: Unified-modal salient object detection via adaptive prompt learning. arXiv preprint arXiv:2311.16835 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "34. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 303, + 102 + ], + "type": "text", + "content": "G. Zhao, J. Huang, X. Yan, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 479 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "35. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 194 + ], + "type": "text", + "content": "36. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "37. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "38. Zhang, J., Liu, H., Yang, K., Hu, X., Liu, R., Stiefelhagen, R.: Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "text", + "content": "39. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": "40. Zhang, Q., Zhao, S., Luo, Y., Zhang, D., Huang, N., Han, J.: Abmdrnet: Adaptive-weighted bi-directional modality difference reduction network for rgb-t semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2633-2642 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "type": "text", + "content": "41. Zhao, H., Qi, X., Shen, X., Shi, J., Jia, J.: Icnet for real-time semantic segmentation on high-resolution images. In: Proceedings of the European conference on computer vision (ECCV). pp. 405-420 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "type": "text", + "content": "42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "text", + "content": "43. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "text", + "content": "44. Zhou, W., Dong, S., Xu, C., Qian, Y.: Edge-aware guidance fusion network for rgb-thermal scene parsing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3571-3579 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "45. Zhu, J., Lai, S., Chen, X., Wang, D., Lu, H.: Visual prompt multi-modal tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9516-9526 (2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary RGB-Thermal Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_content_list.json b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db7e784d32ca425d685d2de9c0d0ea61c05693a5 --- /dev/null +++ b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_content_list.json @@ -0,0 +1,1774 @@ +[ + { + "type": "text", + "text": "Open-Vocabulary SAM: Segment and Recognize Twenty-thousand Classes Interactively", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haobo Yuan $^{1}$ , Xiangtai Li $^{1}$ , Chong Zhou $^{1}$ , Yining Li $^{2}$ , Kai Chen $^{2}$ , and Chen Change Loy $^{1}$", + "bbox": [ + 220, + 210, + 782, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University", + "bbox": [ + 354, + 253, + 645, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shanghai AI Laboratory", + "bbox": [ + 411, + 268, + 589, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: https://www.mmlab-ntu.com/project/ovsam", + "bbox": [ + 300, + 282, + 700, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "E-mail: yuanhaobo@whu.edu.cn, xiangtai94@gmail.com", + "bbox": [ + 313, + 297, + 687, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The CLIP and Segment Anything Model (SAM) are remarkable vision foundation models (VFMs). SAM excels in segmentation tasks across diverse domains, whereas CLIP is renowned for its zero-shot recognition capabilities. This paper presents an in-depth exploration of integrating these two models into a unified framework. Specifically, we introduce the Open-Vocabulary SAM, a SAM-inspired model designed for simultaneous interactive segmentation and recognition, leveraging two unique knowledge transfer modules: SAM2CLIP and CLIP2SAM. The former adapts SAM's knowledge into the CLIP via distillation and learnable transformer adapters, while the latter transfers CLIP knowledge into SAM, enhancing its recognition capabilities. Extensive experiments on various datasets and detectors show the effectiveness of Open-Vocabulary SAM in both segmentation and recognition tasks, significantly outperforming the naive baselines of simply combining SAM and CLIP. Furthermore, aided with image classification data training, our method can segment and recognize approximately 22,000 classes.", + "bbox": [ + 259, + 349, + 743, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Scene Understanding $\\cdot$ Promptable Segmentation", + "bbox": [ + 261, + 584, + 679, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 626, + 375, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Segment Anything Model (SAM) [30] and CLIP [53] have made significant strides in various vision tasks, showcasing remarkable generalization capabilities in segmentation and recognition, respectively. SAM, in particular, has been trained with a massive dataset of mask labels, making it highly adaptable to a wide range of downstream tasks through interactive prompts. On the other hand, CLIP's training with billions of text-image pairs has given it an unprecedented ability in zero-shot visual recognition. This has led to numerous studies [18,64,72,78] exploring the extension of CLIP to open vocabulary tasks, such as detection and segmentation.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While SAM and CLIP offer considerable advantages, they also have inherent limitations in their original designs. SAM, for instance, lacks the capability to recognize the segments it identifies. Efforts to overcome this by integrating a", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d08e96c2b61e9aa040e1b9d3bd0377d9bb614b4f3b9d2cd64f68dc34bde8d6c.jpg", + "image_caption": [ + "Fig. 1: Open-Vocabulary SAM not only can segment anything with prompts just like SAM but also has the capability of recognition in the real world, like CLIP. With drastically lower computational cost, Open-Vocabulary SAM has a higher recognition performance than directly combining SAM and CLIP with image or feature cropping (measured on the COCO open vocabulary benchmark)." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 782, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "classification head have been made [32,83], but these solutions are constrained to specific datasets or closed-set settings. On the other hand, CLIP, which is trained using image-level contrastive losses, faces challenges in adapting its representations for dense prediction tasks. To address this, several studies [18,57,66,67,76] have investigated ways to align CLIP's representation for dense predictions. However, these approaches tend to be dataset-specific and not universally applicable. For example, some research has focused on open vocabulary segmentation on the ADE-20k [84] dataset, using the COCO [43] dataset for pre-training. Merging SAM and CLIP in a naive manner, as illustrated in Fig. 2 (a) and (b), proves to be inefficient. This approach incurs substantial computational expenses and yields subpar results, including recognition of small-scale objects, as evidenced by our experimental results.", + "bbox": [ + 212, + 398, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we address these challenges with a unified encoder-decoder framework that integrates a CLIP encoder and a SAM decoder, as depicted in Fig. 2 (c). To bridge these two distinct components effectively, we introduce two novel modules, $SAM2CLIP$ and $CLIP2SAM$ , facilitating dual knowledge transfer. First, we distill knowledge from the SAM encoder to the CLIP encoder using $SAM2CLIP$ . This distillation process is uniquely executed not directly on the CLIP encoder, which is kept frozen to maintain its existing knowledge, but rather on a lightweight transformer-like adapter using a pixel-wise distillation loss. The adapter takes multi-scale features as input, with the goal of aligning CLIP features with SAM representation. On the decoding side, the $CLIP2SAM$ module transfers knowledge from the frozen CLIP encoder to the SAM decoder. In particular, we design a feature pyramid adapter with a RoIAlign operator to be jointly trained with the SAM decoder. Both modules are lightweight and naturally combine the strengths of CLIP and SAM.", + "bbox": [ + 212, + 582, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the spirit of SAM, we enhance our model's recognition capabilities by harnessing the power of established semantic datasets, including COCO [43], LVIS [19], and ImageNet-22k [11]. This strategy elevates our model to the ver", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/02fb8c05999b7fa2ffcd113ee85a8e4675525b43f314fb2d20cc0715817dd86c.jpg", + "image_caption": [ + "(a) Image Cropping Baseline" + ], + "image_footnote": [], + "bbox": [ + 230, + 148, + 401, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/036dd7cd4526dbb8dea5ba2796fbe8704ec17769b5b5cedcbfa6c7888ca8bfc0.jpg", + "image_caption": [ + "(b) Feature Cropping Baseline" + ], + "image_footnote": [], + "bbox": [ + 403, + 148, + 575, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd97026e7811052495b98f88b11f2c801374954c3d1434e44db7e51e597e0f68.jpg", + "image_caption": [ + "(c) Open-Vocabulary SAM", + "Fig. 2: Comparison of two simple SAM-CLIP combination baselines (a) and (b), and our proposed single encoder architecture (c). The adapters for (a) and (b) are optional and can be replaced with various designs (please refer to Sec. 4.1 for details). Note that, in our method, the SAM encoder will be discarded during inference." + ], + "image_footnote": [], + "bbox": [ + 578, + 148, + 777, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "satility of SAM, endowing it with enhanced capability to segment and recognize any objects, as shown in Fig. 1. As our approach is an adaptation of SAM, it is flexible enough to be integrated with various detectors, making it suitable for both closed-set and open-set environments.", + "bbox": [ + 212, + 363, + 784, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct extensive experiments across a range of datasets and scenarios, encompassing closed-set and open-vocabulary interactive segmentation. Notably, when compared to basic combined baselines, our approach demonstrates superior performance, achieving over $2\\%$ improvement in IoU and $3\\%$ in mAP with various detectors on the COCO dataset. In particular, in the case of recognition on LVIS, our approach achieves over $20\\%$ improvements over previous adapters. Furthermore, by expanding our approach with a more diverse array of datasets, we have developed a versatile, interactive tool suitable for practical applications. For detailed results, we direct the reader to Sec. 4 and the appendix.", + "bbox": [ + 212, + 424, + 784, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 582, + 387, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Language Models (VLMs). Vision-language pre-training has given rise to models with aligned image and text representations [25, 26, 28, 41, 53]. Recent studies on contrastive vision-language pre-training [26, 53, 58, 80] have significantly improved the generalization ability of recognition models. Meanwhile, several works [28, 33-35] aim to design better optimization goals for downstream multi-modal tasks, including caption and visual question answering. Among these works, CLIP models [53] that are pre-trained on billion-scale image-text pairs have shown impressive zero-shot classification performance on a wide range of datasets. Our goal is to enable SAM to perform recognition tasks with the help of pre-trained VLMs.", + "bbox": [ + 212, + 613, + 787, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open Vocabulary Dense Prediction. This direction aims to recognize region visual concepts of arbitrary categories described by texts, which includes object detection [18,63,64,69,79], semantic segmentation [37,38,40,71,86,87], and panoptic segmentation [70,75,76]. This necessitates the alignment between region and text representations with the help of VLMs [26,53,58]. For open-vocabulary", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "detection, a series of works [18, 57, 66, 78] distill knowledge from the CLIP models to recognize novel objects. In contrast to distillation-based methods, several works [31, 68] directly build object detectors upon frozen CLIP CNNs. For open-vocabulary segmentation, the typical works [12, 70, 72, 76] first generate class-agnostic mask proposals and then classify the proposals with CLIP. Recently, several works [70, 76] build the mask generator upon the frozen diffusion model [54] and CLIP model [53]. Meanwhile, several studies [27, 47, 51, 52] focus on class-agnostic segmentation and detection to enrich generalization ability in various domains. However, most approaches are trained and tested on specific datasets. Our approach is based on SAM, which provides a general, interactive tool to support different open vocabulary detectors.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompting in Computer Vision. Prompting, originating from in-context learning in natural language processing (NLP) as seen in works like Brown et al. [4] and Rubin et al. [55], leverages a large language model to infer unseen tasks through context-specific input-output pairs. Recent studies [1,3,16,45,61,62,88] have explored in-context learning for visual tasks. Common techniques involve mask image modeling [2, 21, 77] for cross-task visual prompting, as employed by approaches like Painter [61] and Bar et al. [3]. SAM [30] demonstrates in-context learning through interactive segmentation, using diverse visual prompts like points, boxes, and masks, although it is limited to class-agnostic mask prediction. Meanwhile, other studies [8,17,23,39,42] have concentrated on efficient parameter tuning of visual foundation models, typically focusing on a single model. Our work uniquely bridges two models, CLIP and SAM, exploring their combined potential for enhanced general segmentation and recognition capabilities. In particular, we adopt visual prompts (box, point) as the model's inputs. Segmentation Anything Model. SAM [30] presents a new data engine and portable model for segmentation. Subsequent research has employed SAM as an interactive segmentation tool for various vision tasks, including grounding [44], tracking [10], distillation [81, 85], medical analysis [5, 65], and generation [82]. While some studies use SAM and CLIP for segmentation [6, 20, 36, 49, 59], none have yet integrated VLMs and SAM into a unified model capable of both segmentation and recognition of novel classes. Our work makes the first attempt to merge the capabilities of VLMs with SAM for enhanced task versatility.", + "bbox": [ + 212, + 313, + 787, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 667, + 380, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first review the SAM, CLIP, and combined baselines in Sec. 3.1. Then, we detail our Open Vocabulary SAM in Sec. 3.2. Last, we present our model's training details and application in Sec. 3.3.", + "bbox": [ + 212, + 700, + 785, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries and Baselines", + "text_level": 1, + "bbox": [ + 215, + 768, + 490, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAM. SAM is a prompt-driven segmentor. It contains an image encoder, a prompt encoder, and a light-weight mask decoder. Here, we use box prompts as an example. We denote an input image as $X \\in \\mathbb{R}^{H \\times W \\times 3}$ and input visual", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/41f0f796fe80f45ef8f18e6bcd77ea34d6644844e567929ea8b544f32d1b3e8b.jpg", + "image_caption": [ + "Fig. 3: Illustration of Open-Vocabulary SAM. For training, the SAM encoder is as a teacher network, while SAM2CLIP plays the role of a student network and aligns the knowledge of SAM into CLIP. The CLIP2SAM transfers the CLIP knowledge to the SAM decoder and performs joint segmentation and classification for close-set and open vocabulary settings." + ], + "image_footnote": [], + "bbox": [ + 217, + 147, + 782, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prompts as $P \\in \\mathbb{R}^{N \\times 4}$ , where $H \\times W$ are the spatial size, $N$ is the number of box prompts. The image encoder is a modified vision transformer (ViT). It encodes an image into dense feature $F_{SAM} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times d}$ . The prompt encoder encodes $P$ into sparse prompts $Q_{sp}$ . Meanwhile, mask tokens $Q_{mask}$ and an IoU token $Q_{IoU}$ are initialized for the mask decoder.", + "bbox": [ + 212, + 375, + 782, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mask decoder takes the image feature $F$ , sparse prompts $Q_{sp}$ , mask tokens $Q_{mask}$ , and the IoU token $Q_{IoU}$ as input. All the inputs will be concatenated and encoded with a lightweight two-way transformer. Consequently, each mask token is transformed into a dynamic linear classifier, capable of calculating the foreground mask probability for every sparse prompt. Simultaneously, the IoU token is tasked with predicting the confidence score for each mask. Considering the multi-granular nature of SAM's data annotations, encompassing both instance and part level, $Q_{mask}$ naturally encodes multi-granularity. Our study concentrates exclusively on the object level, which aligns more closely with prevalent real-world applications and datasets such as COCO [43] and LVIS [19]. CLIP. Given an input image $X$ and a corresponding caption $C$ , the CLIP framework processes these modalities to produce respective embeddings: the image embedding $E_I$ , derived from its image encoder, and the text embedding $\\mathbf{t}$ , obtained from its text encoder. In the context of open-vocabulary object detection and segmentation, CLIP's capability to generalize beyond fixed class labels is leveraged to replace traditional classifiers. For instance, in open-vocabulary detection scenarios, the text embedding $\\mathbf{t_c}$ for the $c$ -th object category is generated by inputting the category name into the CLIP text encoder. This process can employ a single template prompt, such as \"a photo of {category}\", or multiple prompt templates. Subsequently, for a given region embedding $r$ , that is produced by the RoI-Align [22], the classification score for the $c$ -th category is computed as: $p_c = \\frac{\\exp(\\tau \\cdot <\\mathbf{r}, \\mathbf{t_c} >)}{\\sum_{i=0}^{C} \\exp(\\tau \\cdot <\\mathbf{r}, \\mathbf{t_i} >)}$ , where $<\\cdot, \\cdot>$ denotes the cosine similarity, and $\\tau$ is a learnable or fixed temperature to re-scale the value.", + "bbox": [ + 212, + 454, + 787, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combined Baselines. We introduce two different baselines for combining CLIP and SAM, as depicted in Fig. 2 (a) and (b). The first approach, termed the", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "'cropped image baseline', employs the SAM mask decoder's output to segment and resize the original input image. This processed image then serves as the input for the CLIP image encoder, and, in conjunction with the CLIP text embedding, the mask is classified using the similarities between visual and text embeddings. The second approach referred to as the 'cropped CLIP image feature baseline', employs the same initial CLIP feature extraction step. However, in this method, masks predicted by the SAM decoder are used to crop the CLIP image features. Subsequent pooling of these masked features yields the final label, akin to baseline (a).", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While both baselines enable zero-shot inference of images, they exhibit a noticeable knowledge gap on specific datasets. To address this, we draw inspiration from recent advancements in visual prompting or adapters [8,88]. Specifically, we propose incorporating additional learnable tokens as an adapter to fine-tune the model for enhanced performance on downstream datasets. These zero-shot inference capabilities and the fine-tuned models constitute our primary comparison baselines under various experimental conditions, detailed in Sec. 4.1.", + "bbox": [ + 212, + 282, + 787, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Open Vocabulary SAM", + "text_level": 1, + "bbox": [ + 215, + 409, + 455, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While both baseline models can be enhanced through visual prompting or adapters, as we will discuss in Sec. 4, they face several challenges in real-world applications. First, the requirement for two independent backbones in the combined model increases computational costs (Prob.1). Second, SAM and CLIP are trained with distinct objectives – SAM through supervised learning and CLIP via contrastive learning – and there is limited research on knowledge transfer between such diverse architectures (Prob.2). Third, despite adapter integration, significant performance gaps remain in recognizing small objects (Prob.3). Fourth, there is a lack of exploration into integrating open-vocabulary capabilities for SAM and CLIP, particularly in the context of feature fusion and data scaling (Prob.4). Our work aims to solve these problems in a unified yet effective framework.", + "bbox": [ + 212, + 431, + 797, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unified Architecture. We design a unified architecture for both segmentation and recognition to address Prob.1. Specifically, we adopt the frozen CLIP visual encoder as our feature extractor. Then, both SAM's mask decoder and prompt encoder are appended behind the CLIP encoder. The meta-architecture of open-vocabulary SAM is shown in Fig. 2 (c), with the more detailed version shown in Fig. 3. This unified architecture is made possible via the SAM2CLIP, which transfers knowledge of SAM to CLIP with distillation, and CLIP2SAM, which employs CLIP knowledge and combines the SAM mask decoder for recognition. We have chosen convolution-based visual backbones for the frozen CLIP backbone, aligning with previous studies that have highlighted their superiority in capturing spatial structures [31, 73]. The efficacy of different CLIP backbones is further explored in Sec. 4.2.", + "bbox": [ + 212, + 598, + 787, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SAM2CLIP. To resolve Prob.2, we design the SAM2CLIP module that bridges the gap in feature representations learned by SAM and CLIP, using adaptation and distillation methods. Through comprehensive experiments, we discovered that employing distillation loss $L_{distill}$ along with transformer-based", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adapters [14], yields effective results. Specifically, the distillation process involves a simple pixel-wise approach, where SAM-Huge serves as the teacher, and the frozen CLIP equipped with an adapter assumes the student's role. We then implement a per-pixel mean squared error (MSE) loss to align the SAM feature $F_{sam}$ with the CLIP feature $E_I$ , as detailed below:", + "bbox": [ + 214, + 146, + 787, + 223 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d i s t i l l} = \\operatorname {M S E} \\left(F _ {s a m}, E _ {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 237, + 785, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We design a multi-scale adapter $A_{sam2clip}$ to align the features from CLIP and SAM. In particular, we take pyramid CLIP features $E_I^i$ , $i = 1, 2, 3$ as the inputs. Such pyramid features contain both high-resolution and semantic information, which is proven crucial for semantic segmentation [29]. The MSE loss is revised as follows:", + "bbox": [ + 214, + 265, + 787, + 340 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d i s t i l l} = \\operatorname {M S E} \\left(F _ {s a m}, A _ {s a m 2 c l i p} \\left(\\operatorname {F u s i o n} \\left(E _ {I} ^ {i}\\right)\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 356, + 785, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $A_{sam2clip}$ comprises several transformer layers, and Fusion is achieved by bilinear upsampling and addition.", + "bbox": [ + 212, + 383, + 785, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With SAM2CLIP, we can even achieve comparable segmentation results with the SAM-Huge with much lower computational costs. As detailed in Sec. 4.2, we observe that employing convolution-based methods specifically designed for backbone adaptation [8, 15] results in sub-optimal outcomes. The reason for this might be in the inherent architecture of the SAM encoder, which is purely based on ViT. A symmetrical structure is crucial for effective knowledge transfer. With the implementation of SAM2CLIP, we can achieve segmentation results comparable to those of SAM-Huge across various detectors, while significantly reducing computational costs.", + "bbox": [ + 212, + 415, + 787, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CLIP2SAM. This module aims to leverage CLIP's knowledge to enhance the recognition capabilities of the SAM decoder. A straightforward approach involves appending a label token $Q_{label}$ to the existing mask token $Q_{mask}$ and IoU token $Q_{IoU}$ . Using $Q_{label}$ , we introduce a specialized adapter to facilitate the transfer of knowledge from the frozen CLIP to the SAM decoder. Subsequently, the enhanced $Q_{label}$ , combined with the output of the prompt encoder and adapted CLIP features, is fed into a two-way transformer. Following the cross-attention process, the improved $Q_{label}$ undergoes further refinement through a multilayer perceptron (MLP), ensuring better alignment with CLIP's text embedding. The final labels are derived by calculating the distance between the refined label token and the CLIP text embedding.", + "bbox": [ + 212, + 551, + 787, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This design, however, falls short of recognizing small objects (Prob.3) since the adaptation only involves the single-scale feature, which is mainly focused on segmentation. We present a simple yet effective solution to handle this issue, introducing a lightweight feature pyramid network (FPN) for CLIP2SAM adaption. As shown in Fig. 3, the pyramid network extracts multi-scale CLIP features as the inputs. Then, we apply the RoI-Align [22] operation to extract region features. Like the R-CNN framework [22], we apply one convolution layer and a MLP to learn the feature embedding without introducing cross-attention", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in the mask decoder. In particular, for point prompts, we first obtain the corresponding masks via the SAM decoder and obtain the box via the corresponding masks. For box prompts, we can directly send it to the FPN for region feature extraction. Given that our method incorporates only a few convolution layers, it does not significantly increase computational costs compared to the original SAM.", + "bbox": [ + 212, + 146, + 787, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Open Vocabulary. To tackle Prob.4, the open-vocabulary challenge, we leverage the knowledge embedded in the frozen CLIP backbone, which aids in recognizing novel and unseen objects during inference. In line with previous studies [31,67], we fuse the learned class scores with those from the frozen CLIP via a geometric mean to leverage information from both the CLIP and CLIP2SAM. Additionally, we investigate various strategies to expand the vocabulary size, such as joint training with multiple datasets, as detailed in Sec. 4.2. Our experimental results show that the model scales effectively with large datasets.", + "bbox": [ + 212, + 237, + 787, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our approach can address the open vocabulary challenge, it is important to distinguish the setting of Open-Vocabulary SAM from that of previous open vocabulary segmentation methods [6,13,24,74]. Unlike previous techniques, which typically depend on a separate segmentor to produce mask proposals, our method only uses visual prompts, such as boxes and points, to generate masks and labels. Furthermore, our method specifically targets foreground objects, aligning with the ImageNet [11] datasets and the purpose of box prompts. To show the effectiveness of our method, we compare the performance of our method with previous open vocabulary segmentation methods in Sec. 4.1.", + "bbox": [ + 212, + 359, + 787, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Training and Application", + "text_level": 1, + "bbox": [ + 215, + 521, + 472, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training and Loss Function. We first use the SA-1B (1%) dataset [30] for training the SAM2CLIP module to transfer SAM's knowledge into open-vocabulary SAM, with the loss $L_{distill}$ (Equ. (2)). Then, we joint train the CLIP2SAM and mask decoder using segmentation mask and label annotations from COCO or LVIS. The final loss function is given as $L = \\lambda_{cls}L_{t\\_cls} + \\lambda_{ce}L_{t\\_ce} + \\lambda_{dice}L_{t\\_dice}$ . Here, $L_{t\\_ce}$ is the Cross-Entropy (CE) loss for mask classification, and $\\bar{L}_{t\\_ce}$ and $L_{t\\_dice}$ are mask Cross Entropy (CE) loss and Dice loss [48] for segmentation, respectively. In addition, we adopt joint training with the ImageNet dataset for classifying over 22,000 categories.", + "bbox": [ + 212, + 549, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference and Demo Tools. Our model performs inference like SAM, with points and boxes as visual prompts. Specifically, we test boxes and points as visual prompts for the encoder in Sec. 4. In the appendix, we show a demo of our model, which can segment and recognize with prompts.", + "bbox": [ + 212, + 686, + 787, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 773, + 375, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets and Metrics. We mainly use COCO [43] and LVIS [19] datasets for the experiments. Moreover, we also use part of SA-1B [30] data (1%) for", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a5d7b5de3e38996c1944b760761c477848bd60e4afde51eec7092eb73a5c71d.jpg", + "table_caption": [ + "Table 1: Comparison of combined baselines and Open-Vocabulary SAM using visual prompts. \\*\\* indicates using mask center point as prompts, while others indicate using ground truth boxes prompts. $I o U_{b}$ and $I o U_{n}$ refer to the average IoU for each mask of base classes and novel classes, respectively." + ], + "table_footnote": [], + "table_body": "
MethodCOCO\nIoUbIoUn AccLVIS\nIoUbIoUn AccFLOPs #Param
Image-Crop baseline78.181.446.278.381.69.6
Feature-Crop baseline78.181.455.178.381.626.5
Image-Crop baseline + CoOp [88]79.682.162.080.182.032.1
Feature-Crop baseline + CoOp [88]79.682.170.980.182.048.2
Open-Vocabulary SAM81.584.084.380.483.166.6
Image-Crop baseline*60.766.724.553.062.36.2
Feature-Crop baseline*60.766.732.153.062.311.0
Image-Crop baseline + CoOp [88]*64.766.728.258.964.28.3
Feature-Crop baseline + CoOp [88]*64.766.735.158.964.213.2
Open-Vocabulary SAM*68.465.276.763.667.960.4
", + "bbox": [ + 220, + 212, + 784, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SAM2CLIP knowledge transfer. For COCO, we report the results of both close-set and open-vocabulary settings for the instance segmentation task. In particular, following Zareian et al. [79], we split 48 base classes with annotations and 17 target classes without annotations. We use the base class annotations for training. For LVIS datasets, we adopt the open-vocabulary setting and report the results of $AP_{\\text{rare}}$ for novel classes. For evaluation metrics, we report the accuracy of recognition for reference to evaluate the recognition capability. Meanwhile, each prompt's intersection-over-union (IoU) with its ground truth mask is also adopted to evaluate the segmentation ability of our method. As mentioned in Sec. 3.2, different from previous open vocabulary segmentation tasks, boxes or points serve as visual prompts in our method.", + "bbox": [ + 212, + 428, + 785, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. As shown in Fig. 2 (a) and (b), based on different adapter designs, we append these adapters to the different locations of the combined models. For example, when using CoOp [88], we append the learnable tokens by combining them with CLIP features. For several convolution-based adapters [8], we add the extra convolution layers along with SAM or CLIP backbone for fair comparison. By default, we adopt SAM-huge and CLIP R50x16.", + "bbox": [ + 212, + 595, + 785, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We implement our models in PyTorch [50] with MMDetection [7]. We use 8 A100 GPUs for distributed training. Each minibatch has two images per GPU. The optimizer is AdamW [46] with a weight decay of 0.0001. We adopt full image size for a random crop in the pre-training and training process following Cheng et al. [9]. All the class names are transferred into CLIP text embedding, following previous works [18]. We train each model for 12 epochs for fair comparison. Due to the limitation of computation costs, we do not adopt joint SAM data and COCO data training. We first perform training the SAM2CLIP on SA-1B (1%), and then we finetune the model on COCO or LVIS data. Please refer to the supplementary material for more details.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2cab27e2b8242f30abc6e10876ff8048c6ca35b54f3aa1c8be322e87bd632158.jpg", + "table_caption": [ + "Table 2: Comparison of combined baselines and Open-Vocabulary SAM on prompts generated by the open vocabulary detector. For the LVIS dataset, only 'normal' and 'frequent' classes are in the training set. The labels are generated by each baseline or our method. We adopt Detric [89] as the OV-Detector to provide box prompts." + ], + "table_footnote": [], + "table_body": "
MethodCOCOLVISFLOPs#Params
APbaseAPnovelAPAPrareAPnormAPfreqAP
Image-Crop baseline + CoOp [88]26.231.227.319.818.316.317.23,748G808M
Feature-Crop baseline + CoOp [88]28.033.829.524.221.418.620.83,545G808M
Open-Vocabulary SAM31.136.032.424.021.322.922.41,180G304M
", + "bbox": [ + 220, + 212, + 782, + 287 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/116c88fffdc74fa4b24ac5daeab0a75a1351854ae7a7ac8b62505390c9c0caad.jpg", + "table_caption": [ + "Table 3: Comparison of Open-Vocabulary SAM with previous methods on open vocabulary instance segmentation. The results presented in the table all use the Mask-RCNN [22] with ResNet-50 backbone. Different from previous works, Open-Vocabulary SAM uses the bounding box as the prompt to generate the mask. We report mask AP50 following base-novel setting as [24]." + ], + "table_footnote": [], + "table_body": "
MethodVenueConstrainedGeneralized
BaseNovelBaseNovelAll
XPM [24]CVPR'2242.424.041.521.636.3
MaskCLIP [13]ICML'2342.823.242.621.737.2
MasQCLIP [74]ICCV'2340.930.140.728.437.5
Open-Vocabulary SAM(Ours)41.737.539.339.839.4
", + "bbox": [ + 316, + 383, + 679, + 474 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bff0a36dea899c88451f3ba151355b5220cca8e3085e2a421a1de441806bb2ea.jpg", + "table_caption": [ + "Table 4: Comparison of mask quality with various detectors on COCO dataset. We report the mask mean AP for comparison. The masks are generated by each method, while the labels are from the corresponding detectors." + ], + "table_footnote": [], + "table_body": "
MethodDetectorsmAPAP50AP75APSAPMAPL#ParamsFLOPs
SAM-HugeFaster-RCNN (R50)35.654.938.417.239.151.4641M3,001G
SAM-Huge (finetuned)Faster-RCNN (R50)35.855.038.416.538.653.0641M3,001G
Open-Vocabulary SAMFaster-RCNN (R50)35.855.638.316.038.953.1304M1,180G
SAM-HugeDetic (swin-base)36.457.139.421.440.854.6641M3,001G
SAM-Huge (finetuned)Detic (swin-base)36.857.439.820.840.655.1641M3,001G
Open-Vocabulary SAMDetic (swin-base)36.757.239.720.740.854.9304M1,180G
SAM-HugeViTDet (Huge)46.372.049.825.245.559.6641M3,001G
SAM-Huge (finetuned)ViTDet (Huge)46.572.350.325.245.860.1641M3,001G
Open-Vocabulary SAMViTDet (Huge)48.873.852.924.846.364.2304M1,180G
", + "bbox": [ + 232, + 542, + 767, + 674 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmark Setup. Open Vocabulary SAM is different from both SAM [30] and open vocabulary segmentation methods [13,24,74]. Compared with Open-Vocabulary SAM, SAM [30] cannot recognize segmented objects while open vocabulary segmentation methods [13,24,74] usually rely on a separate segmentor (e.g., Mask-RCNN [22]) instead of prompts like points or boxes. For a complete evaluation of Open-Vocabulary SAM, we set the benchmark as three parts, including 1). comparison with combined baselines to verify the effectiveness; 2). comparison with open-vocabulary segmentation methods to evaluate the recognition; 3). comparison with SAM models to evaluate segmentation.", + "bbox": [ + 215, + 703, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/50fd20f70167d55efa7bc7b0a554428da02ceb3cdc434517c3f2a8b0e62a2e51.jpg", + "table_caption": [ + "Table 5: Scaling up with large-scale datasets." + ], + "table_footnote": [], + "table_body": "
DatasetsAccuracy#vocabulary#images
LVIS83.11,20399K
V3Det78.713,204183K
I-21k44.519,16713M
V3Det + LVIS82.713,844282K
V3Det + LVIS + I-21k83.325,89813M
V3Det + LVIS + I-21k + Object36583.025,97015M
", + "bbox": [ + 289, + 171, + 709, + 277 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e63293e85b30ef8f9e11888eca78b86d4f67714449443d4c646213cbd519809d.jpg", + "table_caption": [ + "Table 6: The effectiveness of each component. We use Detic [89] as the detector. The labels are generated by the corresponding model. S2C and C2S denote SAM2CLIP and CLIP2SAM respectively. The baseline refers to the image-crop variant." + ], + "table_footnote": [], + "table_body": "
SettingAPAPbaseAPnovelFLOPs (G)#Params (M)
Baseline + CoOp [88]29.528.033.83,545808
Our + S2C28.727.333.31,127291
Our + S2C + C2S34.433.138.01,180304
", + "bbox": [ + 289, + 344, + 709, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Main Results", + "text_level": 1, + "bbox": [ + 215, + 438, + 372, + 452 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Combined Baselines Using Ground Truth. To avoid the influence of other modules, we first demonstrate the recognition ability of our model in Tab. 1. Compared to the simple combined approaches, adding adapters with joint co-training leads to better results. However, the recognition ability is still limited on both COCO and LVIS. Our Open-Vocabulary SAM achieves the best results on both boxes and points as visual prompts. We observe more significant gains on LVIS datasets. We argue that LVIS contains more small objects, which is more challenging than COCO. Our method can solve Prob.2 and lead to over $20\\%$ accuracy improvement. Although the segmentation quality is pretty good (about 80 IoU on COCO and LVIS with box prompt), our method still achieves $2\\%$ IoU improvements. This indicates the effectiveness of our joint co-training on mask prediction and classification. Compared with boxes as prompts, using points as prompts is more challenging since the location clues of points are much weaker than boxes. However, our approach is still better than combined baselines or them with adapters.", + "bbox": [ + 212, + 460, + 787, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Combined Baselines on OV-Detector. In Tab. 2, we adopt a more challenging setting by using the box prediction from the existing open-vocabulary detector to simulate the interactive segmentation process with deviation. We choose the representative Detic [89] as the open-vocabulary detector. Again, our method also achieves the best performance on both COCO and LVIS datasets. In particular, on COCO, compared with previous works [88], our method achieves 3.0 mask mAP improvements with much lower costs.", + "bbox": [ + 212, + 688, + 787, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Open-vocabulary Segmentation. In Tab. 3, we compare our method with previous open-vocabulary instance segmentation methods. Open-Vocabulary SAM uses the boxes generated by Mask- RCNN [22] and", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9918da2f62c9985e09ff0d4c1cc3e2c16155fd07add3a9dd9b99be80d1a10104.jpg", + "table_caption": [ + "Table 7: Ablation studies on COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (left): Ablation on SAM2CLIP design. (right): Ablation on CLIP2SAM design." + ], + "table_footnote": [], + "table_body": "
SettingIoU FLOPs (G) #Params (M)SettingIoU Acc APbase APnovel
SAM78.73,001641SAM2CLIP (baseline)78.154.227.6
Conv-L + MultiScale Neck78.31,313321+ Cls Token81.379.329.8
Conv-L + SingleScale Neck73.61,280307+ Cls Token & CLIP MLP fusion80.978.929.0
R50x16 + MultiScale Neck78.11,180304+ light FPN (Ours)81.584.331.1
R50 + MultiScale Neck77.3728165
", + "bbox": [ + 218, + 199, + 784, + 267 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c24eb787b2849193b6da45e3f018f3397a26050fc3519360d8a04dc8bce67beb.jpg", + "table_caption": [ + "Table 8: (left): Ablation study on different CLIP backbone. We test results on the COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (right): Comparison with other SAM models." + ], + "table_footnote": [], + "table_body": "
BackboneIoU Acc #FLOPs(G) #Params (M)
RN5077.3 50.8728165
RN50x1678.1 55.11,180304
RN50x6478.1 54.12,098568
ConvNeXt-L78.3 59.11,313321
ViT-L-1438.6 14.32,294441
", + "bbox": [ + 220, + 334, + 483, + 412 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d039bd88ff2ea66dfe77b2dd7532086769393aa82695eca1ce59ecd40a853c29.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method1-IoU (COCO)cls. open.
SAM [30] (H)78.2--
SEEM [90] (T)73.7✓-
Semantic-SAM [32] (T)76.1✓-
OV-SAM (ours)81.7✓✓
", + "bbox": [ + 519, + 338, + 782, + 409 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "previous works use the masks of Mask-RCNN. The results show that our Open-Vocabulary SAM performs strongly, especially in the novel classes.", + "bbox": [ + 212, + 441, + 782, + 472 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with SAM on Various Detectors. In Tab. 4, we test the mask prediction quality of our model and original SAM on different detectors. Our method performs better than the original SAM and performs comparably with fine-tuned SAM. It is worth noting that our Open-Vocabulary SAM has much lower computational costs and parameters than SAM.", + "bbox": [ + 212, + 473, + 784, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on Interactive Segmentation. In Tab. 8 (right), we compare interactive segmentation performance. Notably, our approach excels beyond interactive segmentation and can recognize classes in an open-vocabulary setting. Visualization Comparison. In Fig. 4, we compare our approach with the feature-crop baseline. Our model shows better performance in classifying small and rare object classification and handling occlusion scenarios.", + "bbox": [ + 212, + 547, + 784, + 638 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Model as a Zero Shot Annotation Tool. In addition to COCO and LVIS standard datasets training, following the spirit of SAM, we also scale up our model by training it with more data (Tab. 5). In particular, we adopt more detection data (V3Det [60], Object365 [56]) and classification data (ImageNet22k [11]). Owing to significant costs, we have not conducted comparisons with other baselines for this setting. Rather, we have adapted our method into an interactive annotation tool capable of segmenting and recognizing over 22,000 classes.", + "bbox": [ + 212, + 638, + 787, + 747 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Ablation Studies and Analysis", + "text_level": 1, + "bbox": [ + 214, + 768, + 511, + 784 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of SAM2CLIP and CLIP2SAM. We first verify the effectiveness of our proposed two modules in Tab. 6. We adopt image-crop variant of the baseline for comparison. In particular, by sharing a single backbone, we observe", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d6313c36a159f09236bb855814d903f7f107015db0866a095e03c2955d7f718f.jpg", + "image_caption": [ + "Fig. 4: Visualization Comparison. We compare the mask and classification results of the image-crop baseline (SAM + CLIP) and Open-Vocabulary SAM (OV SAM). The predicted labels are presented on the mask." + ], + "image_footnote": [], + "bbox": [ + 289, + 148, + 715, + 392 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "a significant drop in the number of parameters and FLOPs, with a little drop in segmentation performance. The slight drop is caused by the domain gap between SAM data and COCO data during the training of the SAM2CLIP module. However, after adding our CLIP2SAM module and joint co-training with mask classification and prediction, a significant improvement in both segmentation and classification is observed, with just a negligible increase in compute cost.", + "bbox": [ + 212, + 482, + 787, + 573 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Detailed Design on SAM2CLIP. In Tab. 7, we explore the detailed design of SAM2CLIP in the first stage of open-vocabulary SAM training. The results show that distillation benefits most when multi-scale features are adopted, suggesting that both high-resolution features and high-level semantics are important to align CLIP's feature with the SAM's feature.", + "bbox": [ + 212, + 575, + 787, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Detailed Design on CLIP2SAM. In Tab. 7, we present extensive design for the CLIP2SAM module. We compare two designs: a simple classification token with cross attention (Cls Token) and a combination of this token with mask pooled CLIP feature (CLS Token & CLIP MLP fusion). These designs work better than the combined baseline shown in the first row. Nonetheless, due to resolution constraints, these variants cannot handle small objects well, as shown in Fig. 4. In contrast, our design improves the performance considerably.", + "bbox": [ + 212, + 655, + 787, + 761 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on Different CLIP Backbones. In Tab. 8, we explore the effect of frozen CLIP visual backbone. We do not add the CLIP2SAM. Motivated by recent works [31,67,73,76], CNN-based CLIPs encapsulate more structural information, which is good for our goal since we have location-sensitive visual prompts. Thus, adopt CNN-based CLIPs, aligned with previous works. As shown", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/86073dd589b87850741a8711cbd5818db5351f08272fe52c5b27f7081571749e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 354, + 224 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/40c3b8fc36b2611ec703be7c0697015b1721136679dd41c15a7d98e9fcc94efa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 143, + 506, + 224 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/272da928ca6b914a200c89de556f6229024baf88ca0c863ac0b1db449c613a3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 143, + 645, + 224 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f4334d26a305ab5ff9ca5c37b2ab0548afcfa594cd5b7792c7c4d9c54f41f92a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 143, + 712, + 224 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/18b84b886e35bdd4419970d61e0c59756f35d2f22a6e91395d9b94e00972133c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 143, + 784, + 224 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8ee2ece8e7a8041242d99d76926ea13c7d1fafa9c227801ce597acf9e3aeffb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 231, + 354, + 308 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/35b3153c1e006fbca89453d14eb1201b03473d9a0da96ab4516a83d2ea97b9ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 234, + 508, + 308 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/132df602141ec273e4a06ea2eb3288e0daa0cbfba59a2ca3690716b1b1d0e8da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 234, + 645, + 308 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3c37d9c18656873ff7c01e4a9314c867cd3c8b8ae20e96d3e68e03961cb8a0f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 234, + 712, + 308 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/602e18d6d337cb3b4dc3003acacca4f334ab6fddbadf5caa9e835d8380375a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 234, + 784, + 308 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/97e4442e4ea0a5ba3e9aa030b6741888660ab91e74f5bebeeefa6337ef0b18e8.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of Open-Vocabulary SAM. In the visualization, boxes refer to box prompts (up), and the red stars refer to the point prompts (middle). The masks can be merged into a segmentation map (bottom). We show the mask and labels generated by the proposed Open-Vocabulary SAM. Our method can segment and recognize open vocabulary objects in diverse scenes identified by prompts." + ], + "image_footnote": [], + "bbox": [ + 217, + 320, + 401, + 445 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6d8e6627bdbf8bfe7ab8d9af8261374ff569ec69d675c86b541c6fa985134cc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 320, + 576, + 445 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4ceaf8098f57f5b9958e3bc7c0b047e8531aa48981fd85898a20d212e91a5cf7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 321, + 784, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "in the table, we find ConvNext large achieves the best performance, but we choose RN50x16 since it has comparable performance and better efficiency.", + "bbox": [ + 212, + 534, + 787, + 565 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 592, + 359, + 608 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present Open Vocabulary SAM, a SAM-inspired method for interactive segmentation and recognition. Unlike previous open-vocabulary detection and segmentation methods, our method explores interactive open-vocabulary segmentation for the first time. Given the user's inputs, such as boxes or points, the proposed approach can interactively segment and label each visual prompt. Compared with the combined baselines and various visual adapters, our proposed CLIP2SAM and SAM2CLIP are both efficient and effective in various settings. Our open vocabulary segmentation is compatible with different detectors, including open-vocabulary detectors and close-set detectors. With more data, our model plays a similar role as SAM, offering an effective annotation tool for both segmentation and instance labeling. In particular, our method can perform large vocabulary segmentation and recognition over 22K classes. We hope our Open-Vocabulary SAM can provide a solid baseline for combining the strengths of different forms of vision foundation models and inspire further research.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). The project is also supported by Singapore MOE AcRF Tier 1 (RG16/21) and the National Key R&D Program of China (No. 2022ZD0161600).", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 246, + 321, + 262 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Balăzevic, I., Steiner, D., Parthasarathy, N., Arandjelović, R., Henaff, O.J.: Towards in-context scene understanding. In: NeurIPS (2023)", + "2. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2022)", + "3. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. In: NeurIPS (2022)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)", + "5. Chen, C., Miao, J., Wu, D., Yan, Z., Kim, S., Hu, J., Zhong, A., Liu, Z., Sun, L., Li, X., Liu, T., Heng, P.A., Li, Q.: MA-SAM: Modality-agnostic SAM adaptation for 3d medical image segmentation. arXiv preprint arXiv:2309.08842 (2023)", + "6. Chen, J., Yang, Z., Zhang, L.: Semantic segment anything (2023)", + "7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)", + "8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023)", + "9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022)", + "0. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023)", + "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009)", + "2. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022)", + "3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023)", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "5. Fang, Y., Yang, S., Wang, S., Ge, Y., Shan, Y., Wang, X.: Unleashing vanilla vision transformer with masked image modeling for object detection. In: ICCV (2023)", + "6. Fang, Z., Li, X., Li, X., Buhmann, J.M., Loy, C.C., Liu, M.: Explore in-context learning for 3d point cloud understanding. In: NeurIPS (2023)", + "7. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: CLIP-Adapter: Better vision-language models with feature adapters. IJCV (2024)", + "8. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)" + ], + "bbox": [ + 225, + 280, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)", + "20. Han, X., Wei, L., Yu, X., Dou, Z., He, X., Wang, K., Han, Z., Tian, Q.: Boosting segment anything model towards open-vocabulary learning. arXiv preprint arXiv:2312.03628 (2023)", + "21. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)", + "22. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV (2017)", + "23. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)", + "24. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: CVPR (2022)", + "25. Jayaraman, D., Grauman, K.: Zero-shot recognition with unreliable attributes. In: NeurIPS (2014)", + "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)", + "27. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. RA-L (2022)", + "28. Kim, W., Son, B., Kim, I.: ViLT: Vision-and-language transformer without convolution or region supervision. In: ICML (2021)", + "29. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: CVPR (2019)", + "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)", + "31. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A.J., Angelova, A.: F-VLM: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023)", + "32. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-SAM: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)", + "33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)", + "34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)", + "35. Li, J., Selvaraju, R.R., Gotmare, A., Joty, S.R., Xiong, C., Hoi, S.C.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021)", + "36. Li, S., Cao, J., Ye, P., Ding, Y., Tu, C., Chen, T.: ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665 (2024)", + "37. Li, X., Ding, H., Zhang, W., Yuan, H., Cheng, G., Jiangmiao, P., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv pre-print (2023)", + "38. Li, X., You, A., Zhu, Z., Zhao, H., Yang, M., Yang, K., Tong, Y.: Semantic flow for fast and accurate scene parsing. In: ECCV (2020)", + "39. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: OMG-Seg: Is one model good enough for all segmentation? In: CVPR (2024)", + "40. Li, X., Zhang, J., Yang, Y., Cheng, G., Yang, K., Tong, Y., Tao, D.: Sfnet: Faster and accurate domain agnostic semantic segmentation via semantic flow. IJCV (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: CVPR (2023)", + "42. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: NeurIPS (2022)", + "43. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)", + "44. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "45. Liu, Y., Qi, L., Tsai, Y.J., Li, X., Chan, K.C.K., Yang, M.H.: Effective adapter for face recognition in the wild. arXiv preprint (2023)", + "46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)", + "47. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022)", + "48. Miletari, F., Navab, N., Ahmadi, S.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. In: 3DV (2016)", + "49. Pan, T., Tang, L., Wang, X., Shan, S.: Tokenize anything via prompting. arXiv preprint arXiv:2312.09128 (2023)", + "50. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)", + "51. Qi, L., Kuen, J., Shen, T., Gu, J., Guo, W., Jia, J., Lin, Z., Yang, M.H.: High-quality entity segmentation. In: ICCV (2023)", + "52. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE TPAMI (2022)", + "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "55. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv:2112.08633 (2021)", + "56. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: ICCV (2019)", + "57. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV (2023)", + "58. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: EVA-CLIP: Improved training techniques for CLIP at scale. arXiv preprint arXiv:2303.15389 (2023)", + "59. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: SAM-CLIP: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023)", + "60. Wang, J., Zhang, P., Chu, T., Cao, Y., Zhou, Y., Wu, T., Wang, B., He, C., Lin, D.: V3det: Vast vocabulary visual detection dataset. In: ICCV (2023)", + "61. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR (2023)", + "62. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. In: ICCV (2023)", + "63. Wu, J., Li, X., Ding, H., Li, X., Cheng, G., Tong, Y., Loy, C.C.: Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. In: ICCV (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024)", + "65. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)", + "66. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: CVPR (2023)", + "67. Wu, S., Zhang, W., Xu, L., Jin, S., Li, X., Liu, W., Loy, C.C.: CLIPSelf: Vision transformer distills itself for open-vocabulary dense prediction. In: ICLR (2024)", + "68. Wu, X., Zhu, F., Zhao, R., Li, H.: CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In: CVPR (2023)", + "69. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023)", + "70. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)", + "71. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023)", + "72. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022)", + "73. Xu, S., Li, X., Wu, S., Zhang, W., Li, Y., Cheng, G., Tong, Y., Chen, K., Loy, C.C.: Dst-det: Simple dynamic self-training for open-vocabulary object detection. arXiv pre-print (2023)", + "74. Xu, X., Xiong, T., Ding, Z., Tu, Z.: Masqclip for open-vocabulary universal image segmentation. In: ICCV (2023)", + "75. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., et al.: Panoptic video scene graph generation. In: CVPR (2023)", + "76. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional CLIP. In: NeurIPS (2023)", + "77. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022)", + "78. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV (2022)", + "79. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021)", + "80. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: LiT: Zero-shot transfer with locked-image text tuning. In: CVPR (2022)", + "81. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight SAM for mobile applications. arXiv preprint arXiv:2306.14289 (2023)", + "82. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)", + "83. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)", + "84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)", + "85. Zhou, C., Li, X., Loy, C.C., Dai, B.: EdgeSAM: Prompt-in-the-loop distillation for on-device deployment of SAM. arXiv preprint arXiv:2312.06660 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "86. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from CLIP. In: ECCV (2022)", + "87. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. arXiv preprint arXiv:2311.03352 (2023)", + "88. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV (2022)", + "89. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022)", + "90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: NeurIPS (2023)" + ], + "bbox": [ + 212, + 146, + 785, + 285 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Open-Vocabulary SAM", + "bbox": [ + 573, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_model.json b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b133d7951087ec24a2c99c4e5b72741ebda3c1f8 --- /dev/null +++ b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_model.json @@ -0,0 +1,2768 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Open-Vocabulary SAM: Segment and Recognize Twenty-thousand Classes Interactively" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.212, + 0.783, + 0.245 + ], + "angle": 0, + "content": "Haobo Yuan\\(^{1}\\), Xiangtai Li\\(^{1}\\), Chong Zhou\\(^{1}\\), Yining Li\\(^{2}\\), Kai Chen\\(^{2}\\), and Chen Change Loy\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.255, + 0.646, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.269, + 0.59, + 0.284 + ], + "angle": 0, + "content": "2 Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.284, + 0.702, + 0.298 + ], + "angle": 0, + "content": "Project Page: https://www.mmlab-ntu.com/project/ovsam" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.299, + 0.689, + 0.311 + ], + "angle": 0, + "content": "E-mail: yuanhaobo@whu.edu.cn, xiangtai94@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.35, + 0.744, + 0.573 + ], + "angle": 0, + "content": "Abstract. The CLIP and Segment Anything Model (SAM) are remarkable vision foundation models (VFMs). SAM excels in segmentation tasks across diverse domains, whereas CLIP is renowned for its zero-shot recognition capabilities. This paper presents an in-depth exploration of integrating these two models into a unified framework. Specifically, we introduce the Open-Vocabulary SAM, a SAM-inspired model designed for simultaneous interactive segmentation and recognition, leveraging two unique knowledge transfer modules: SAM2CLIP and CLIP2SAM. The former adapts SAM's knowledge into the CLIP via distillation and learnable transformer adapters, while the latter transfers CLIP knowledge into SAM, enhancing its recognition capabilities. Extensive experiments on various datasets and detectors show the effectiveness of Open-Vocabulary SAM in both segmentation and recognition tasks, significantly outperforming the naive baselines of simply combining SAM and CLIP. Furthermore, aided with image classification data training, our method can segment and recognize approximately 22,000 classes." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.585, + 0.68, + 0.6 + ], + "angle": 0, + "content": "Keywords: Scene Understanding \\(\\cdot\\) Promptable Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.627, + 0.377, + 0.643 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "The Segment Anything Model (SAM) [30] and CLIP [53] have made significant strides in various vision tasks, showcasing remarkable generalization capabilities in segmentation and recognition, respectively. SAM, in particular, has been trained with a massive dataset of mask labels, making it highly adaptable to a wide range of downstream tasks through interactive prompts. On the other hand, CLIP's training with billions of text-image pairs has given it an unprecedented ability in zero-shot visual recognition. This has led to numerous studies [18,64,72,78] exploring the extension of CLIP to open vocabulary tasks, such as detection and segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "While SAM and CLIP offer considerable advantages, they also have inherent limitations in their original designs. SAM, for instance, lacks the capability to recognize the segments it identifies. Efforts to overcome this by integrating a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 1: Open-Vocabulary SAM not only can segment anything with prompts just like SAM but also has the capability of recognition in the real world, like CLIP. With drastically lower computational cost, Open-Vocabulary SAM has a higher recognition performance than directly combining SAM and CLIP with image or feature cropping (measured on the COCO open vocabulary benchmark)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.581 + ], + "angle": 0, + "content": "classification head have been made [32,83], but these solutions are constrained to specific datasets or closed-set settings. On the other hand, CLIP, which is trained using image-level contrastive losses, faces challenges in adapting its representations for dense prediction tasks. To address this, several studies [18,57,66,67,76] have investigated ways to align CLIP's representation for dense predictions. However, these approaches tend to be dataset-specific and not universally applicable. For example, some research has focused on open vocabulary segmentation on the ADE-20k [84] dataset, using the COCO [43] dataset for pre-training. Merging SAM and CLIP in a naive manner, as illustrated in Fig. 2 (a) and (b), proves to be inefficient. This approach incurs substantial computational expenses and yields subpar results, including recognition of small-scale objects, as evidenced by our experimental results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.793 + ], + "angle": 0, + "content": "In this study, we address these challenges with a unified encoder-decoder framework that integrates a CLIP encoder and a SAM decoder, as depicted in Fig. 2 (c). To bridge these two distinct components effectively, we introduce two novel modules, \\(SAM2CLIP\\) and \\(CLIP2SAM\\), facilitating dual knowledge transfer. First, we distill knowledge from the SAM encoder to the CLIP encoder using \\(SAM2CLIP\\). This distillation process is uniquely executed not directly on the CLIP encoder, which is kept frozen to maintain its existing knowledge, but rather on a lightweight transformer-like adapter using a pixel-wise distillation loss. The adapter takes multi-scale features as input, with the goal of aligning CLIP features with SAM representation. On the decoding side, the \\(CLIP2SAM\\) module transfers knowledge from the frozen CLIP encoder to the SAM decoder. In particular, we design a feature pyramid adapter with a RoIAlign operator to be jointly trained with the SAM decoder. Both modules are lightweight and naturally combine the strengths of CLIP and SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Following the spirit of SAM, we enhance our model's recognition capabilities by harnessing the power of established semantic datasets, including COCO [43], LVIS [19], and ImageNet-22k [11]. This strategy elevates our model to the ver" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.575, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.149, + 0.402, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.262, + 0.378, + 0.274 + ], + "angle": 0, + "content": "(a) Image Cropping Baseline" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.149, + 0.576, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.262, + 0.557, + 0.274 + ], + "angle": 0, + "content": "(b) Feature Cropping Baseline" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.149, + 0.779, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.262, + 0.722, + 0.274 + ], + "angle": 0, + "content": "(c) Open-Vocabulary SAM" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.289, + 0.787, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: Comparison of two simple SAM-CLIP combination baselines (a) and (b), and our proposed single encoder architecture (c). The adapters for (a) and (b) are optional and can be replaced with various designs (please refer to Sec. 4.1 for details). Note that, in our method, the SAM encoder will be discarded during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.785, + 0.424 + ], + "angle": 0, + "content": "satility of SAM, endowing it with enhanced capability to segment and recognize any objects, as shown in Fig. 1. As our approach is an adaptation of SAM, it is flexible enough to be integrated with various detectors, making it suitable for both closed-set and open-set environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.56 + ], + "angle": 0, + "content": "We conduct extensive experiments across a range of datasets and scenarios, encompassing closed-set and open-vocabulary interactive segmentation. Notably, when compared to basic combined baselines, our approach demonstrates superior performance, achieving over \\(2\\%\\) improvement in IoU and \\(3\\%\\) in mAP with various detectors on the COCO dataset. In particular, in the case of recognition on LVIS, our approach achieves over \\(20\\%\\) improvements over previous adapters. Furthermore, by expanding our approach with a more diverse array of datasets, we have developed a versatile, interactive tool suitable for practical applications. For detailed results, we direct the reader to Sec. 4 and the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.583, + 0.388, + 0.599 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Vision Language Models (VLMs). Vision-language pre-training has given rise to models with aligned image and text representations [25, 26, 28, 41, 53]. Recent studies on contrastive vision-language pre-training [26, 53, 58, 80] have significantly improved the generalization ability of recognition models. Meanwhile, several works [28, 33-35] aim to design better optimization goals for downstream multi-modal tasks, including caption and visual question answering. Among these works, CLIP models [53] that are pre-trained on billion-scale image-text pairs have shown impressive zero-shot classification performance on a wide range of datasets. Our goal is to enable SAM to perform recognition tasks with the help of pre-trained VLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Open Vocabulary Dense Prediction. This direction aims to recognize region visual concepts of arbitrary categories described by texts, which includes object detection [18,63,64,69,79], semantic segmentation [37,38,40,71,86,87], and panoptic segmentation [70,75,76]. This necessitates the alignment between region and text representations with the help of VLMs [26,53,58]. For open-vocabulary" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "detection, a series of works [18, 57, 66, 78] distill knowledge from the CLIP models to recognize novel objects. In contrast to distillation-based methods, several works [31, 68] directly build object detectors upon frozen CLIP CNNs. For open-vocabulary segmentation, the typical works [12, 70, 72, 76] first generate class-agnostic mask proposals and then classify the proposals with CLIP. Recently, several works [70, 76] build the mask generator upon the frozen diffusion model [54] and CLIP model [53]. Meanwhile, several studies [27, 47, 51, 52] focus on class-agnostic segmentation and detection to enrich generalization ability in various domains. However, most approaches are trained and tested on specific datasets. Our approach is based on SAM, which provides a general, interactive tool to support different open vocabulary detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Prompting in Computer Vision. Prompting, originating from in-context learning in natural language processing (NLP) as seen in works like Brown et al. [4] and Rubin et al. [55], leverages a large language model to infer unseen tasks through context-specific input-output pairs. Recent studies [1,3,16,45,61,62,88] have explored in-context learning for visual tasks. Common techniques involve mask image modeling [2, 21, 77] for cross-task visual prompting, as employed by approaches like Painter [61] and Bar et al. [3]. SAM [30] demonstrates in-context learning through interactive segmentation, using diverse visual prompts like points, boxes, and masks, although it is limited to class-agnostic mask prediction. Meanwhile, other studies [8,17,23,39,42] have concentrated on efficient parameter tuning of visual foundation models, typically focusing on a single model. Our work uniquely bridges two models, CLIP and SAM, exploring their combined potential for enhanced general segmentation and recognition capabilities. In particular, we adopt visual prompts (box, point) as the model's inputs. Segmentation Anything Model. SAM [30] presents a new data engine and portable model for segmentation. Subsequent research has employed SAM as an interactive segmentation tool for various vision tasks, including grounding [44], tracking [10], distillation [81, 85], medical analysis [5, 65], and generation [82]. While some studies use SAM and CLIP for segmentation [6, 20, 36, 49, 59], none have yet integrated VLMs and SAM into a unified model capable of both segmentation and recognition of novel classes. Our work makes the first attempt to merge the capabilities of VLMs with SAM for enhanced task versatility." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.669, + 0.381, + 0.687 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.787, + 0.748 + ], + "angle": 0, + "content": "We first review the SAM, CLIP, and combined baselines in Sec. 3.1. Then, we detail our Open Vocabulary SAM in Sec. 3.2. Last, we present our model's training details and application in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.491, + 0.785 + ], + "angle": 0, + "content": "3.1 Preliminaries and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "SAM. SAM is a prompt-driven segmentor. It contains an image encoder, a prompt encoder, and a light-weight mask decoder. Here, we use box prompts as an example. We denote an input image as \\( X \\in \\mathbb{R}^{H \\times W \\times 3} \\) and input visual" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.575, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.148, + 0.784, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 3: Illustration of Open-Vocabulary SAM. For training, the SAM encoder is as a teacher network, while SAM2CLIP plays the role of a student network and aligns the knowledge of SAM into CLIP. The CLIP2SAM transfers the CLIP knowledge to the SAM decoder and performs joint segmentation and classification for close-set and open vocabulary settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.784, + 0.452 + ], + "angle": 0, + "content": "prompts as \\( P \\in \\mathbb{R}^{N \\times 4} \\), where \\( H \\times W \\) are the spatial size, \\( N \\) is the number of box prompts. The image encoder is a modified vision transformer (ViT). It encodes an image into dense feature \\( F_{SAM} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times d} \\). The prompt encoder encodes \\( P \\) into sparse prompts \\( Q_{sp} \\). Meanwhile, mask tokens \\( Q_{mask} \\) and an IoU token \\( Q_{IoU} \\) are initialized for the mask decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.789, + 0.808 + ], + "angle": 0, + "content": "The mask decoder takes the image feature \\( F \\), sparse prompts \\( Q_{sp} \\), mask tokens \\( Q_{mask} \\), and the IoU token \\( Q_{IoU} \\) as input. All the inputs will be concatenated and encoded with a lightweight two-way transformer. Consequently, each mask token is transformed into a dynamic linear classifier, capable of calculating the foreground mask probability for every sparse prompt. Simultaneously, the IoU token is tasked with predicting the confidence score for each mask. Considering the multi-granular nature of SAM's data annotations, encompassing both instance and part level, \\( Q_{mask} \\) naturally encodes multi-granularity. Our study concentrates exclusively on the object level, which aligns more closely with prevalent real-world applications and datasets such as COCO [43] and LVIS [19]. CLIP. Given an input image \\( X \\) and a corresponding caption \\( C \\), the CLIP framework processes these modalities to produce respective embeddings: the image embedding \\( E_I \\), derived from its image encoder, and the text embedding \\( \\mathbf{t} \\), obtained from its text encoder. In the context of open-vocabulary object detection and segmentation, CLIP's capability to generalize beyond fixed class labels is leveraged to replace traditional classifiers. For instance, in open-vocabulary detection scenarios, the text embedding \\( \\mathbf{t_c} \\) for the \\( c \\)-th object category is generated by inputting the category name into the CLIP text encoder. This process can employ a single template prompt, such as \"a photo of {category}\", or multiple prompt templates. Subsequently, for a given region embedding \\( r \\), that is produced by the RoI-Align [22], the classification score for the \\( c \\)-th category is computed as: \\( p_c = \\frac{\\exp(\\tau \\cdot <\\mathbf{r}, \\mathbf{t_c} >)}{\\sum_{i=0}^{C} \\exp(\\tau \\cdot <\\mathbf{r}, \\mathbf{t_i} >)} \\), where \\( <\\cdot, \\cdot> \\) denotes the cosine similarity, and \\( \\tau \\) is a learnable or fixed temperature to re-scale the value." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Combined Baselines. We introduce two different baselines for combining CLIP and SAM, as depicted in Fig. 2 (a) and (b). The first approach, termed the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "'cropped image baseline', employs the SAM mask decoder's output to segment and resize the original input image. This processed image then serves as the input for the CLIP image encoder, and, in conjunction with the CLIP text embedding, the mask is classified using the similarities between visual and text embeddings. The second approach referred to as the 'cropped CLIP image feature baseline', employs the same initial CLIP feature extraction step. However, in this method, masks predicted by the SAM decoder are used to crop the CLIP image features. Subsequent pooling of these masked features yields the final label, akin to baseline (a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.39 + ], + "angle": 0, + "content": "While both baselines enable zero-shot inference of images, they exhibit a noticeable knowledge gap on specific datasets. To address this, we draw inspiration from recent advancements in visual prompting or adapters [8,88]. Specifically, we propose incorporating additional learnable tokens as an adapter to fine-tune the model for enhanced performance on downstream datasets. These zero-shot inference capabilities and the fine-tuned models constitute our primary comparison baselines under various experimental conditions, detailed in Sec. 4.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.456, + 0.426 + ], + "angle": 0, + "content": "3.2 Open Vocabulary SAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.799, + 0.599 + ], + "angle": 0, + "content": "While both baseline models can be enhanced through visual prompting or adapters, as we will discuss in Sec. 4, they face several challenges in real-world applications. First, the requirement for two independent backbones in the combined model increases computational costs (Prob.1). Second, SAM and CLIP are trained with distinct objectives – SAM through supervised learning and CLIP via contrastive learning – and there is limited research on knowledge transfer between such diverse architectures (Prob.2). Third, despite adapter integration, significant performance gaps remain in recognizing small objects (Prob.3). Fourth, there is a lack of exploration into integrating open-vocabulary capabilities for SAM and CLIP, particularly in the context of feature fusion and data scaling (Prob.4). Our work aims to solve these problems in a unified yet effective framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Unified Architecture. We design a unified architecture for both segmentation and recognition to address Prob.1. Specifically, we adopt the frozen CLIP visual encoder as our feature extractor. Then, both SAM's mask decoder and prompt encoder are appended behind the CLIP encoder. The meta-architecture of open-vocabulary SAM is shown in Fig. 2 (c), with the more detailed version shown in Fig. 3. This unified architecture is made possible via the SAM2CLIP, which transfers knowledge of SAM to CLIP with distillation, and CLIP2SAM, which employs CLIP knowledge and combines the SAM mask decoder for recognition. We have chosen convolution-based visual backbones for the frozen CLIP backbone, aligning with previous studies that have highlighted their superiority in capturing spatial structures [31, 73]. The efficacy of different CLIP backbones is further explored in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "SAM2CLIP. To resolve Prob.2, we design the SAM2CLIP module that bridges the gap in feature representations learned by SAM and CLIP, using adaptation and distillation methods. Through comprehensive experiments, we discovered that employing distillation loss \\( L_{distill} \\) along with transformer-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.574, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "adapters [14], yields effective results. Specifically, the distillation process involves a simple pixel-wise approach, where SAM-Huge serves as the teacher, and the frozen CLIP equipped with an adapter assumes the student's role. We then implement a per-pixel mean squared error (MSE) loss to align the SAM feature \\( F_{sam} \\) with the CLIP feature \\( E_I \\), as detailed below:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.238, + 0.786, + 0.253 + ], + "angle": 0, + "content": "\\[\nL _ {d i s t i l l} = \\operatorname {M S E} \\left(F _ {s a m}, E _ {I}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.266, + 0.788, + 0.342 + ], + "angle": 0, + "content": "We design a multi-scale adapter \\( A_{sam2clip} \\) to align the features from CLIP and SAM. In particular, we take pyramid CLIP features \\( E_I^i \\), \\( i = 1, 2, 3 \\) as the inputs. Such pyramid features contain both high-resolution and semantic information, which is proven crucial for semantic segmentation [29]. The MSE loss is revised as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.357, + 0.786, + 0.373 + ], + "angle": 0, + "content": "\\[\nL _ {d i s t i l l} = \\operatorname {M S E} \\left(F _ {s a m}, A _ {s a m 2 c l i p} \\left(\\operatorname {F u s i o n} \\left(E _ {I} ^ {i}\\right)\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.786, + 0.414 + ], + "angle": 0, + "content": "where \\( A_{sam2clip} \\) comprises several transformer layers, and Fusion is achieved by bilinear upsampling and addition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.552 + ], + "angle": 0, + "content": "With SAM2CLIP, we can even achieve comparable segmentation results with the SAM-Huge with much lower computational costs. As detailed in Sec. 4.2, we observe that employing convolution-based methods specifically designed for backbone adaptation [8, 15] results in sub-optimal outcomes. The reason for this might be in the inherent architecture of the SAM encoder, which is purely based on ViT. A symmetrical structure is crucial for effective knowledge transfer. With the implementation of SAM2CLIP, we can achieve segmentation results comparable to those of SAM-Huge across various detectors, while significantly reducing computational costs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.718 + ], + "angle": 0, + "content": "CLIP2SAM. This module aims to leverage CLIP's knowledge to enhance the recognition capabilities of the SAM decoder. A straightforward approach involves appending a label token \\( Q_{label} \\) to the existing mask token \\( Q_{mask} \\) and IoU token \\( Q_{IoU} \\). Using \\( Q_{label} \\), we introduce a specialized adapter to facilitate the transfer of knowledge from the frozen CLIP to the SAM decoder. Subsequently, the enhanced \\( Q_{label} \\), combined with the output of the prompt encoder and adapted CLIP features, is fed into a two-way transformer. Following the cross-attention process, the improved \\( Q_{label} \\) undergoes further refinement through a multilayer perceptron (MLP), ensuring better alignment with CLIP's text embedding. The final labels are derived by calculating the distance between the refined label token and the CLIP text embedding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This design, however, falls short of recognizing small objects (Prob.3) since the adaptation only involves the single-scale feature, which is mainly focused on segmentation. We present a simple yet effective solution to handle this issue, introducing a lightweight feature pyramid network (FPN) for CLIP2SAM adaption. As shown in Fig. 3, the pyramid network extracts multi-scale CLIP features as the inputs. Then, we apply the RoI-Align [22] operation to extract region features. Like the R-CNN framework [22], we apply one convolution layer and a MLP to learn the feature embedding without introducing cross-attention" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.235 + ], + "angle": 0, + "content": "in the mask decoder. In particular, for point prompts, we first obtain the corresponding masks via the SAM decoder and obtain the box via the corresponding masks. For box prompts, we can directly send it to the FPN for region feature extraction. Given that our method incorporates only a few convolution layers, it does not significantly increase computational costs compared to the original SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Open Vocabulary. To tackle Prob.4, the open-vocabulary challenge, we leverage the knowledge embedded in the frozen CLIP backbone, which aids in recognizing novel and unseen objects during inference. In line with previous studies [31,67], we fuse the learned class scores with those from the frozen CLIP via a geometric mean to leverage information from both the CLIP and CLIP2SAM. Additionally, we investigate various strategies to expand the vocabulary size, such as joint training with multiple datasets, as detailed in Sec. 4.2. Our experimental results show that the model scales effectively with large datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.496 + ], + "angle": 0, + "content": "While our approach can address the open vocabulary challenge, it is important to distinguish the setting of Open-Vocabulary SAM from that of previous open vocabulary segmentation methods [6,13,24,74]. Unlike previous techniques, which typically depend on a separate segmentor to produce mask proposals, our method only uses visual prompts, such as boxes and points, to generate masks and labels. Furthermore, our method specifically targets foreground objects, aligning with the ImageNet [11] datasets and the purpose of box prompts. To show the effectiveness of our method, we compare the performance of our method with previous open vocabulary segmentation methods in Sec. 4.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.473, + 0.538 + ], + "angle": 0, + "content": "3.3 Training and Application" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Training and Loss Function. We first use the SA-1B (1%) dataset [30] for training the SAM2CLIP module to transfer SAM's knowledge into open-vocabulary SAM, with the loss \\( L_{distill} \\) (Equ. (2)). Then, we joint train the CLIP2SAM and mask decoder using segmentation mask and label annotations from COCO or LVIS. The final loss function is given as \\( L = \\lambda_{cls}L_{t\\_cls} + \\lambda_{ce}L_{t\\_ce} + \\lambda_{dice}L_{t\\_dice} \\). Here, \\( L_{t\\_ce} \\) is the Cross-Entropy (CE) loss for mask classification, and \\( \\bar{L}_{t\\_ce} \\) and \\( L_{t\\_dice} \\) are mask Cross Entropy (CE) loss and Dice loss [48] for segmentation, respectively. In addition, we adopt joint training with the ImageNet dataset for classifying over 22,000 categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Inference and Demo Tools. Our model performs inference like SAM, with points and boxes as visual prompts. Specifically, we test boxes and points as visual prompts for the encoder in Sec. 4. In the appendix, we show a demo of our model, which can segment and recognize with prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.775, + 0.377, + 0.792 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Datasets and Metrics. We mainly use COCO [43] and LVIS [19] datasets for the experiments. Moreover, we also use part of SA-1B [30] data (1%) for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.575, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Comparison of combined baselines and Open-Vocabulary SAM using visual prompts. \\*\\* indicates using mask center point as prompts, while others indicate using ground truth boxes prompts. \\(I o U_{b}\\) and \\(I o U_{n}\\) refer to the average IoU for each mask of base classes and novel classes, respectively." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.785, + 0.398 + ], + "angle": 0, + "content": "
MethodCOCO\nIoUbIoUn AccLVIS\nIoUbIoUn AccFLOPs #Param
Image-Crop baseline78.181.446.278.381.69.6
Feature-Crop baseline78.181.455.178.381.626.5
Image-Crop baseline + CoOp [88]79.682.162.080.182.032.1
Feature-Crop baseline + CoOp [88]79.682.170.980.182.048.2
Open-Vocabulary SAM81.584.084.380.483.166.6
Image-Crop baseline*60.766.724.553.062.36.2
Feature-Crop baseline*60.766.732.153.062.311.0
Image-Crop baseline + CoOp [88]*64.766.728.258.964.28.3
Feature-Crop baseline + CoOp [88]*64.766.735.158.964.213.2
Open-Vocabulary SAM*68.465.276.763.667.960.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.786, + 0.595 + ], + "angle": 0, + "content": "SAM2CLIP knowledge transfer. For COCO, we report the results of both close-set and open-vocabulary settings for the instance segmentation task. In particular, following Zareian et al. [79], we split 48 base classes with annotations and 17 target classes without annotations. We use the base class annotations for training. For LVIS datasets, we adopt the open-vocabulary setting and report the results of \\(AP_{\\text{rare}}\\) for novel classes. For evaluation metrics, we report the accuracy of recognition for reference to evaluate the recognition capability. Meanwhile, each prompt's intersection-over-union (IoU) with its ground truth mask is also adopted to evaluate the segmentation ability of our method. As mentioned in Sec. 3.2, different from previous open vocabulary segmentation tasks, boxes or points serve as visual prompts in our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.687 + ], + "angle": 0, + "content": "Baselines. As shown in Fig. 2 (a) and (b), based on different adapter designs, we append these adapters to the different locations of the combined models. For example, when using CoOp [88], we append the learnable tokens by combining them with CLIP features. For several convolution-based adapters [8], we add the extra convolution layers along with SAM or CLIP backbone for fair comparison. By default, we adopt SAM-huge and CLIP R50x16." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. We implement our models in PyTorch [50] with MMDetection [7]. We use 8 A100 GPUs for distributed training. Each minibatch has two images per GPU. The optimizer is AdamW [46] with a weight decay of 0.0001. We adopt full image size for a random crop in the pre-training and training process following Cheng et al. [9]. All the class names are transferred into CLIP text embedding, following previous works [18]. We train each model for 12 epochs for fair comparison. Due to the limitation of computation costs, we do not adopt joint SAM data and COCO data training. We first perform training the SAM2CLIP on SA-1B (1%), and then we finetune the model on COCO or LVIS data. Please refer to the supplementary material for more details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Comparison of combined baselines and Open-Vocabulary SAM on prompts generated by the open vocabulary detector. For the LVIS dataset, only 'normal' and 'frequent' classes are in the training set. The labels are generated by each baseline or our method. We adopt Detric [89] as the OV-Detector to provide box prompts." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.784, + 0.288 + ], + "angle": 0, + "content": "
MethodCOCOLVISFLOPs#Params
APbaseAPnovelAPAPrareAPnormAPfreqAP
Image-Crop baseline + CoOp [88]26.231.227.319.818.316.317.23,748G808M
Feature-Crop baseline + CoOp [88]28.033.829.524.221.418.620.83,545G808M
Open-Vocabulary SAM31.136.032.424.021.322.922.41,180G304M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.303, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Table 3: Comparison of Open-Vocabulary SAM with previous methods on open vocabulary instance segmentation. The results presented in the table all use the Mask-RCNN [22] with ResNet-50 backbone. Different from previous works, Open-Vocabulary SAM uses the bounding box as the prompt to generate the mask. We report mask AP50 following base-novel setting as [24]." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.385, + 0.681, + 0.475 + ], + "angle": 0, + "content": "
MethodVenueConstrainedGeneralized
BaseNovelBaseNovelAll
XPM [24]CVPR'2242.424.041.521.636.3
MaskCLIP [13]ICML'2342.823.242.621.737.2
MasQCLIP [74]ICCV'2340.930.140.728.437.5
Open-Vocabulary SAM(Ours)41.737.539.339.839.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.488, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Table 4: Comparison of mask quality with various detectors on COCO dataset. We report the mask mean AP for comparison. The masks are generated by each method, while the labels are from the corresponding detectors." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.543, + 0.768, + 0.675 + ], + "angle": 0, + "content": "
MethodDetectorsmAPAP50AP75APSAPMAPL#ParamsFLOPs
SAM-HugeFaster-RCNN (R50)35.654.938.417.239.151.4641M3,001G
SAM-Huge (finetuned)Faster-RCNN (R50)35.855.038.416.538.653.0641M3,001G
Open-Vocabulary SAMFaster-RCNN (R50)35.855.638.316.038.953.1304M1,180G
SAM-HugeDetic (swin-base)36.457.139.421.440.854.6641M3,001G
SAM-Huge (finetuned)Detic (swin-base)36.857.439.820.840.655.1641M3,001G
Open-Vocabulary SAMDetic (swin-base)36.757.239.720.740.854.9304M1,180G
SAM-HugeViTDet (Huge)46.372.049.825.245.559.6641M3,001G
SAM-Huge (finetuned)ViTDet (Huge)46.572.350.325.245.860.1641M3,001G
Open-Vocabulary SAMViTDet (Huge)48.873.852.924.846.364.2304M1,180G
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Benchmark Setup. Open Vocabulary SAM is different from both SAM [30] and open vocabulary segmentation methods [13,24,74]. Compared with Open-Vocabulary SAM, SAM [30] cannot recognize segmented objects while open vocabulary segmentation methods [13,24,74] usually rely on a separate segmentor (e.g., Mask-RCNN [22]) instead of prompts like points or boxes. For a complete evaluation of Open-Vocabulary SAM, we set the benchmark as three parts, including 1). comparison with combined baselines to verify the effectiveness; 2). comparison with open-vocabulary segmentation methods to evaluate the recognition; 3). comparison with SAM models to evaluate segmentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.574, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.347, + 0.145, + 0.655, + 0.159 + ], + "angle": 0, + "content": "Table 5: Scaling up with large-scale datasets." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.172, + 0.71, + 0.279 + ], + "angle": 0, + "content": "
DatasetsAccuracy#vocabulary#images
LVIS83.11,20399K
V3Det78.713,204183K
I-21k44.519,16713M
V3Det + LVIS82.713,844282K
V3Det + LVIS + I-21k83.325,89813M
V3Det + LVIS + I-21k + Object36583.025,97015M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.29, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Table 6: The effectiveness of each component. We use Detic [89] as the detector. The labels are generated by the corresponding model. S2C and C2S denote SAM2CLIP and CLIP2SAM respectively. The baseline refers to the image-crop variant." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.345, + 0.71, + 0.414 + ], + "angle": 0, + "content": "
SettingAPAPbaseAPnovelFLOPs (G)#Params (M)
Baseline + CoOp [88]29.528.033.83,545808
Our + S2C28.727.333.31,127291
Our + S2C + C2S34.433.138.01,180304
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.439, + 0.373, + 0.453 + ], + "angle": 0, + "content": "4.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Comparison with Combined Baselines Using Ground Truth. To avoid the influence of other modules, we first demonstrate the recognition ability of our model in Tab. 1. Compared to the simple combined approaches, adding adapters with joint co-training leads to better results. However, the recognition ability is still limited on both COCO and LVIS. Our Open-Vocabulary SAM achieves the best results on both boxes and points as visual prompts. We observe more significant gains on LVIS datasets. We argue that LVIS contains more small objects, which is more challenging than COCO. Our method can solve Prob.2 and lead to over \\(20\\%\\) accuracy improvement. Although the segmentation quality is pretty good (about 80 IoU on COCO and LVIS with box prompt), our method still achieves \\(2\\%\\) IoU improvements. This indicates the effectiveness of our joint co-training on mask prediction and classification. Compared with boxes as prompts, using points as prompts is more challenging since the location clues of points are much weaker than boxes. However, our approach is still better than combined baselines or them with adapters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Comparison with Combined Baselines on OV-Detector. In Tab. 2, we adopt a more challenging setting by using the box prediction from the existing open-vocabulary detector to simulate the interactive segmentation process with deviation. We choose the representative Detic [89] as the open-vocabulary detector. Again, our method also achieves the best performance on both COCO and LVIS datasets. In particular, on COCO, compared with previous works [88], our method achieves 3.0 mask mAP improvements with much lower costs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Comparison with Open-vocabulary Segmentation. In Tab. 3, we compare our method with previous open-vocabulary instance segmentation methods. Open-Vocabulary SAM uses the boxes generated by Mask- RCNN [22] and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 7: Ablation studies on COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (left): Ablation on SAM2CLIP design. (right): Ablation on CLIP2SAM design." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.2, + 0.785, + 0.268 + ], + "angle": 0, + "content": "
SettingIoU FLOPs (G) #Params (M)SettingIoU Acc APbase APnovel
SAM78.73,001641SAM2CLIP (baseline)78.154.227.6
Conv-L + MultiScale Neck78.31,313321+ Cls Token81.379.329.8
Conv-L + SingleScale Neck73.61,280307+ Cls Token & CLIP MLP fusion80.978.929.0
R50x16 + MultiScale Neck78.11,180304+ light FPN (Ours)81.584.331.1
R50 + MultiScale Neck77.3728165
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.324 + ], + "angle": 0, + "content": "Table 8: (left): Ablation study on different CLIP backbone. We test results on the COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (right): Comparison with other SAM models." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.335, + 0.484, + 0.414 + ], + "angle": 0, + "content": "
BackboneIoU Acc #FLOPs(G) #Params (M)
RN5077.3 50.8728165
RN50x1678.1 55.11,180304
RN50x6478.1 54.12,098568
ConvNeXt-L78.3 59.11,313321
ViT-L-1438.6 14.32,294441
" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.339, + 0.784, + 0.41 + ], + "angle": 0, + "content": "
Method1-IoU (COCO)cls. open.
SAM [30] (H)78.2--
SEEM [90] (T)73.7✓-
Semantic-SAM [32] (T)76.1✓-
OV-SAM (ours)81.7✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.784, + 0.473 + ], + "angle": 0, + "content": "previous works use the masks of Mask-RCNN. The results show that our Open-Vocabulary SAM performs strongly, especially in the novel classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.548 + ], + "angle": 0, + "content": "Comparison with SAM on Various Detectors. In Tab. 4, we test the mask prediction quality of our model and original SAM on different detectors. Our method performs better than the original SAM and performs comparably with fine-tuned SAM. It is worth noting that our Open-Vocabulary SAM has much lower computational costs and parameters than SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.785, + 0.64 + ], + "angle": 0, + "content": "Comparison on Interactive Segmentation. In Tab. 8 (right), we compare interactive segmentation performance. Notably, our approach excels beyond interactive segmentation and can recognize classes in an open-vocabulary setting. Visualization Comparison. In Fig. 4, we compare our approach with the feature-crop baseline. Our model shows better performance in classifying small and rare object classification and handling occlusion scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Model as a Zero Shot Annotation Tool. In addition to COCO and LVIS standard datasets training, following the spirit of SAM, we also scale up our model by training it with more data (Tab. 5). In particular, we adopt more detection data (V3Det [60], Object365 [56]) and classification data (ImageNet22k [11]). Owing to significant costs, we have not conducted comparisons with other baselines for this setting. Rather, we have adapted our method into an interactive annotation tool capable of segmenting and recognizing over 22,000 classes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.513, + 0.785 + ], + "angle": 0, + "content": "4.2 Ablation Studies and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Effectiveness of SAM2CLIP and CLIP2SAM. We first verify the effectiveness of our proposed two modules in Tab. 6. We adopt image-crop variant of the baseline for comparison. In particular, by sharing a single backbone, we observe" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.575, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.149, + 0.716, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.404, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Fig. 4: Visualization Comparison. We compare the mask and classification results of the image-crop baseline (SAM + CLIP) and Open-Vocabulary SAM (OV SAM). The predicted labels are presented on the mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.574 + ], + "angle": 0, + "content": "a significant drop in the number of parameters and FLOPs, with a little drop in segmentation performance. The slight drop is caused by the domain gap between SAM data and COCO data during the training of the SAM2CLIP module. However, after adding our CLIP2SAM module and joint co-training with mask classification and prediction, a significant improvement in both segmentation and classification is observed, with just a negligible increase in compute cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.652 + ], + "angle": 0, + "content": "Detailed Design on SAM2CLIP. In Tab. 7, we explore the detailed design of SAM2CLIP in the first stage of open-vocabulary SAM training. The results show that distillation benefits most when multi-scale features are adopted, suggesting that both high-resolution features and high-level semantics are important to align CLIP's feature with the SAM's feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.762 + ], + "angle": 0, + "content": "Detailed Design on CLIP2SAM. In Tab. 7, we present extensive design for the CLIP2SAM module. We compare two designs: a simple classification token with cross attention (Cls Token) and a combination of this token with mask pooled CLIP feature (CLS Token & CLIP MLP fusion). These designs work better than the combined baseline shown in the first row. Nonetheless, due to resolution constraints, these variants cannot handle small objects well, as shown in Fig. 4. In contrast, our design improves the performance considerably." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Ablation on Different CLIP Backbones. In Tab. 8, we explore the effect of frozen CLIP visual backbone. We do not add the CLIP2SAM. Motivated by recent works [31,67,73,76], CNN-based CLIPs encapsulate more structural information, which is good for our goal since we have location-sensitive visual prompts. Thus, adopt CNN-based CLIPs, aligned with previous works. As shown" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.355, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.144, + 0.508, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.145, + 0.647, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.145, + 0.713, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.145, + 0.785, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.232, + 0.355, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.236, + 0.509, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.236, + 0.647, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.236, + 0.713, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.236, + 0.785, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.321, + 0.403, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.321, + 0.577, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.322, + 0.785, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.527 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of Open-Vocabulary SAM. In the visualization, boxes refer to box prompts (up), and the red stars refer to the point prompts (middle). The masks can be merged into a segmentation map (bottom). We show the mask and labels generated by the proposed Open-Vocabulary SAM. Our method can segment and recognize open vocabulary objects in diverse scenes identified by prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.566 + ], + "angle": 0, + "content": "in the table, we find ConvNext large achieves the best performance, but we choose RN50x16 since it has comparable performance and better efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.593, + 0.36, + 0.609 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We present Open Vocabulary SAM, a SAM-inspired method for interactive segmentation and recognition. Unlike previous open-vocabulary detection and segmentation methods, our method explores interactive open-vocabulary segmentation for the first time. Given the user's inputs, such as boxes or points, the proposed approach can interactively segment and label each visual prompt. Compared with the combined baselines and various visual adapters, our proposed CLIP2SAM and SAM2CLIP are both efficient and effective in various settings. Our open vocabulary segmentation is compatible with different detectors, including open-vocabulary detectors and close-set detectors. With more data, our model plays a similar role as SAM, offering an effective annotation tool for both segmentation and instance labeling. In particular, our method can perform large vocabulary segmentation and recognition over 22K classes. We hope our Open-Vocabulary SAM can provide a solid baseline for combining the strengths of different forms of vision foundation models and inspire further research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.574, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). The project is also supported by Singapore MOE AcRF Tier 1 (RG16/21) and the National Key R&D Program of China (No. 2022ZD0161600)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.247, + 0.323, + 0.263 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.308 + ], + "angle": 0, + "content": "1. Balăzevic, I., Steiner, D., Parthasarathy, N., Arandjelović, R., Henaff, O.J.: Towards in-context scene understanding. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.336 + ], + "angle": 0, + "content": "2. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.364 + ], + "angle": 0, + "content": "3. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.785, + 0.449 + ], + "angle": 0, + "content": "5. Chen, C., Miao, J., Wu, D., Yan, Z., Kim, S., Hu, J., Zhong, A., Liu, Z., Sun, L., Li, X., Liu, T., Heng, P.A., Li, Q.: MA-SAM: Modality-agnostic SAM adaptation for 3d medical image segmentation. arXiv preprint arXiv:2309.08842 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.678, + 0.463 + ], + "angle": 0, + "content": "6. Chen, J., Yang, Z., Zhang, L.: Semantic segment anything (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.785, + 0.505 + ], + "angle": 0, + "content": "7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.785, + 0.533 + ], + "angle": 0, + "content": "8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.56 + ], + "angle": 0, + "content": "9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.785, + 0.589 + ], + "angle": 0, + "content": "0. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.616 + ], + "angle": 0, + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.644 + ], + "angle": 0, + "content": "2. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.672 + ], + "angle": 0, + "content": "3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.728 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "5. Fang, Y., Yang, S., Wang, S., Ge, Y., Shan, Y., Wang, X.: Unleashing vanilla vision transformer with masked image modeling for object detection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "6. Fang, Z., Li, X., Li, X., Buhmann, J.M., Loy, C.C., Liu, M.: Explore in-context learning for 3d point cloud understanding. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "7. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: CLIP-Adapter: Better vision-language models with feature adapters. IJCV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "20. Han, X., Wei, L., Yu, X., Dou, Z., He, X., Wang, K., Han, Z., Tian, Q.: Boosting segment anything model towards open-vocabulary learning. arXiv preprint arXiv:2312.03628 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "21. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.76, + 0.257 + ], + "angle": 0, + "content": "22. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "23. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "24. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.337 + ], + "angle": 0, + "content": "25. Jayaraman, D., Grauman, K.: Zero-shot recognition with unreliable attributes. In: NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.337, + 0.785, + 0.378 + ], + "angle": 0, + "content": "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.406 + ], + "angle": 0, + "content": "27. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. RA-L (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "28. Kim, W., Son, B., Kim, I.: ViLT: Vision-and-language transformer without convolution or region supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.46 + ], + "angle": 0, + "content": "29. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "31. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A.J., Angelova, A.: F-VLM: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "32. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-SAM: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.596 + ], + "angle": 0, + "content": "33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.623 + ], + "angle": 0, + "content": "34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "35. Li, J., Selvaraju, R.R., Gotmare, A., Joty, S.R., Xiong, C., Hoi, S.C.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.704 + ], + "angle": 0, + "content": "36. Li, S., Cao, J., Ye, P., Ding, Y., Tu, C., Chen, T.: ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.746 + ], + "angle": 0, + "content": "37. Li, X., Ding, H., Zhang, W., Yuan, H., Cheng, G., Jiangmiao, P., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv pre-print (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "38. Li, X., You, A., Zhu, Z., Zhao, H., Yang, M., Yang, K., Tong, Y.: Semantic flow for fast and accurate scene parsing. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "39. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: OMG-Seg: Is one model good enough for all segmentation? In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "40. Li, X., Zhang, J., Yang, Y., Cheng, G., Yang, K., Tong, Y., Tao, D.: Sfnet: Faster and accurate domain agnostic semantic segmentation via semantic flow. IJCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.575, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "41. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.202 + ], + "angle": 0, + "content": "42. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.23 + ], + "angle": 0, + "content": "43. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.787, + 0.27 + ], + "angle": 0, + "content": "44. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "45. Liu, Y., Qi, L., Tsai, Y.J., Li, X., Chan, K.C.K., Yang, M.H.: Effective adapter for face recognition in the wild. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.311 + ], + "angle": 0, + "content": "46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.338 + ], + "angle": 0, + "content": "47. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.787, + 0.365 + ], + "angle": 0, + "content": "48. Miletari, F., Navab, N., Ahmadi, S.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. In: 3DV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.787, + 0.392 + ], + "angle": 0, + "content": "49. Pan, T., Tang, L., Wang, X., Shan, S.: Tokenize anything via prompting. arXiv preprint arXiv:2312.09128 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "50. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.46 + ], + "angle": 0, + "content": "51. Qi, L., Kuen, J., Shen, T., Gu, J., Guo, W., Jia, J., Lin, Z., Yang, M.H.: High-quality entity segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.787, + 0.487 + ], + "angle": 0, + "content": "52. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.528 + ], + "angle": 0, + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.529, + 0.787, + 0.555 + ], + "angle": 0, + "content": "54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.556, + 0.787, + 0.582 + ], + "angle": 0, + "content": "55. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv:2112.08633 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.61 + ], + "angle": 0, + "content": "56. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.787, + 0.636 + ], + "angle": 0, + "content": "57. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.636, + 0.787, + 0.663 + ], + "angle": 0, + "content": "58. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: EVA-CLIP: Improved training techniques for CLIP at scale. arXiv preprint arXiv:2303.15389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.663, + 0.787, + 0.718 + ], + "angle": 0, + "content": "59. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: SAM-CLIP: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.787, + 0.746 + ], + "angle": 0, + "content": "60. Wang, J., Zhang, P., Chu, T., Cao, Y., Zhou, Y., Wu, T., Wang, B., He, C., Lin, D.: V3det: Vast vocabulary visual detection dataset. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.746, + 0.787, + 0.772 + ], + "angle": 0, + "content": "61. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "62. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "63. Wu, J., Li, X., Ding, H., Li, X., Cheng, G., Tong, Y., Loy, C.C.: Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "64. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "65. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "66. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "67. Wu, S., Zhang, W., Xu, L., Jin, S., Li, X., Liu, W., Loy, C.C.: CLIPSelf: Vision transformer distills itself for open-vocabulary dense prediction. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "68. Wu, X., Zhu, F., Zhao, R., Li, H.: CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "69. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.38 + ], + "angle": 0, + "content": "70. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.785, + 0.406 + ], + "angle": 0, + "content": "71. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.447 + ], + "angle": 0, + "content": "72. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "73. Xu, S., Li, X., Wu, S., Zhang, W., Li, Y., Cheng, G., Tong, Y., Chen, K., Loy, C.C.: Dst-det: Simple dynamic self-training for open-vocabulary object detection. arXiv pre-print (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.515 + ], + "angle": 0, + "content": "74. Xu, X., Xiong, T., Ding, Z., Tu, Z.: Masqclip for open-vocabulary universal image segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.542 + ], + "angle": 0, + "content": "75. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., et al.: Panoptic video scene graph generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.583 + ], + "angle": 0, + "content": "76. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional CLIP. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.61 + ], + "angle": 0, + "content": "77. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "78. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "79. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.691 + ], + "angle": 0, + "content": "80. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: LiT: Zero-shot transfer with locked-image text tuning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "81. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight SAM for mobile applications. arXiv preprint arXiv:2306.14289 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "82. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "83. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "85. Zhou, C., Li, X., Loy, C.C., Dai, B.: EdgeSAM: Prompt-in-the-loop distillation for on-device deployment of SAM. arXiv preprint arXiv:2312.06660 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.575, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-Vocabulary SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.161 + ], + "angle": 0, + "content": "86. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from CLIP. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.162, + 0.786, + 0.203 + ], + "angle": 0, + "content": "87. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. arXiv preprint arXiv:2311.03352 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.786, + 0.231 + ], + "angle": 0, + "content": "88. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "89. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.286 + ], + "angle": 0, + "content": "90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: NeurIPS (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.286 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_origin.pdf b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e9ee3b35634fe32f07a4775ebdb58672144f9a16 --- /dev/null +++ b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/b48c8638-747a-47cc-91e7-420dcc5582d8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:770d38e396dcd3cabee91df0574d931dcd4e1ad345571191520fcfe0958572cd +size 3609845 diff --git a/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/full.md b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93fb7e609bb8729cc7406bdc4dee729c3eb16fe2 --- /dev/null +++ b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/full.md @@ -0,0 +1,326 @@ +# Open-Vocabulary SAM: Segment and Recognize Twenty-thousand Classes Interactively + +Haobo Yuan $^{1}$ , Xiangtai Li $^{1}$ , Chong Zhou $^{1}$ , Yining Li $^{2}$ , Kai Chen $^{2}$ , and Chen Change Loy $^{1}$ + +$^{1}$ S-Lab, Nanyang Technological University + +2 Shanghai AI Laboratory + +Project Page: https://www.mmlab-ntu.com/project/ovsam + +E-mail: yuanhaobo@whu.edu.cn, xiangtai94@gmail.com + +Abstract. The CLIP and Segment Anything Model (SAM) are remarkable vision foundation models (VFMs). SAM excels in segmentation tasks across diverse domains, whereas CLIP is renowned for its zero-shot recognition capabilities. This paper presents an in-depth exploration of integrating these two models into a unified framework. Specifically, we introduce the Open-Vocabulary SAM, a SAM-inspired model designed for simultaneous interactive segmentation and recognition, leveraging two unique knowledge transfer modules: SAM2CLIP and CLIP2SAM. The former adapts SAM's knowledge into the CLIP via distillation and learnable transformer adapters, while the latter transfers CLIP knowledge into SAM, enhancing its recognition capabilities. Extensive experiments on various datasets and detectors show the effectiveness of Open-Vocabulary SAM in both segmentation and recognition tasks, significantly outperforming the naive baselines of simply combining SAM and CLIP. Furthermore, aided with image classification data training, our method can segment and recognize approximately 22,000 classes. + +Keywords: Scene Understanding $\cdot$ Promptable Segmentation + +# 1 Introduction + +The Segment Anything Model (SAM) [30] and CLIP [53] have made significant strides in various vision tasks, showcasing remarkable generalization capabilities in segmentation and recognition, respectively. SAM, in particular, has been trained with a massive dataset of mask labels, making it highly adaptable to a wide range of downstream tasks through interactive prompts. On the other hand, CLIP's training with billions of text-image pairs has given it an unprecedented ability in zero-shot visual recognition. This has led to numerous studies [18,64,72,78] exploring the extension of CLIP to open vocabulary tasks, such as detection and segmentation. + +While SAM and CLIP offer considerable advantages, they also have inherent limitations in their original designs. SAM, for instance, lacks the capability to recognize the segments it identifies. Efforts to overcome this by integrating a + +![](images/6d08e96c2b61e9aa040e1b9d3bd0377d9bb614b4f3b9d2cd64f68dc34bde8d6c.jpg) +Fig. 1: Open-Vocabulary SAM not only can segment anything with prompts just like SAM but also has the capability of recognition in the real world, like CLIP. With drastically lower computational cost, Open-Vocabulary SAM has a higher recognition performance than directly combining SAM and CLIP with image or feature cropping (measured on the COCO open vocabulary benchmark). + +classification head have been made [32,83], but these solutions are constrained to specific datasets or closed-set settings. On the other hand, CLIP, which is trained using image-level contrastive losses, faces challenges in adapting its representations for dense prediction tasks. To address this, several studies [18,57,66,67,76] have investigated ways to align CLIP's representation for dense predictions. However, these approaches tend to be dataset-specific and not universally applicable. For example, some research has focused on open vocabulary segmentation on the ADE-20k [84] dataset, using the COCO [43] dataset for pre-training. Merging SAM and CLIP in a naive manner, as illustrated in Fig. 2 (a) and (b), proves to be inefficient. This approach incurs substantial computational expenses and yields subpar results, including recognition of small-scale objects, as evidenced by our experimental results. + +In this study, we address these challenges with a unified encoder-decoder framework that integrates a CLIP encoder and a SAM decoder, as depicted in Fig. 2 (c). To bridge these two distinct components effectively, we introduce two novel modules, $SAM2CLIP$ and $CLIP2SAM$ , facilitating dual knowledge transfer. First, we distill knowledge from the SAM encoder to the CLIP encoder using $SAM2CLIP$ . This distillation process is uniquely executed not directly on the CLIP encoder, which is kept frozen to maintain its existing knowledge, but rather on a lightweight transformer-like adapter using a pixel-wise distillation loss. The adapter takes multi-scale features as input, with the goal of aligning CLIP features with SAM representation. On the decoding side, the $CLIP2SAM$ module transfers knowledge from the frozen CLIP encoder to the SAM decoder. In particular, we design a feature pyramid adapter with a RoIAlign operator to be jointly trained with the SAM decoder. Both modules are lightweight and naturally combine the strengths of CLIP and SAM. + +Following the spirit of SAM, we enhance our model's recognition capabilities by harnessing the power of established semantic datasets, including COCO [43], LVIS [19], and ImageNet-22k [11]. This strategy elevates our model to the ver + +![](images/02fb8c05999b7fa2ffcd113ee85a8e4675525b43f314fb2d20cc0715817dd86c.jpg) +(a) Image Cropping Baseline + +![](images/036dd7cd4526dbb8dea5ba2796fbe8704ec17769b5b5cedcbfa6c7888ca8bfc0.jpg) +(b) Feature Cropping Baseline + +![](images/dd97026e7811052495b98f88b11f2c801374954c3d1434e44db7e51e597e0f68.jpg) +(c) Open-Vocabulary SAM +Fig. 2: Comparison of two simple SAM-CLIP combination baselines (a) and (b), and our proposed single encoder architecture (c). The adapters for (a) and (b) are optional and can be replaced with various designs (please refer to Sec. 4.1 for details). Note that, in our method, the SAM encoder will be discarded during inference. + +satility of SAM, endowing it with enhanced capability to segment and recognize any objects, as shown in Fig. 1. As our approach is an adaptation of SAM, it is flexible enough to be integrated with various detectors, making it suitable for both closed-set and open-set environments. + +We conduct extensive experiments across a range of datasets and scenarios, encompassing closed-set and open-vocabulary interactive segmentation. Notably, when compared to basic combined baselines, our approach demonstrates superior performance, achieving over $2\%$ improvement in IoU and $3\%$ in mAP with various detectors on the COCO dataset. In particular, in the case of recognition on LVIS, our approach achieves over $20\%$ improvements over previous adapters. Furthermore, by expanding our approach with a more diverse array of datasets, we have developed a versatile, interactive tool suitable for practical applications. For detailed results, we direct the reader to Sec. 4 and the appendix. + +# 2 Related Work + +Vision Language Models (VLMs). Vision-language pre-training has given rise to models with aligned image and text representations [25, 26, 28, 41, 53]. Recent studies on contrastive vision-language pre-training [26, 53, 58, 80] have significantly improved the generalization ability of recognition models. Meanwhile, several works [28, 33-35] aim to design better optimization goals for downstream multi-modal tasks, including caption and visual question answering. Among these works, CLIP models [53] that are pre-trained on billion-scale image-text pairs have shown impressive zero-shot classification performance on a wide range of datasets. Our goal is to enable SAM to perform recognition tasks with the help of pre-trained VLMs. + +Open Vocabulary Dense Prediction. This direction aims to recognize region visual concepts of arbitrary categories described by texts, which includes object detection [18,63,64,69,79], semantic segmentation [37,38,40,71,86,87], and panoptic segmentation [70,75,76]. This necessitates the alignment between region and text representations with the help of VLMs [26,53,58]. For open-vocabulary + +detection, a series of works [18, 57, 66, 78] distill knowledge from the CLIP models to recognize novel objects. In contrast to distillation-based methods, several works [31, 68] directly build object detectors upon frozen CLIP CNNs. For open-vocabulary segmentation, the typical works [12, 70, 72, 76] first generate class-agnostic mask proposals and then classify the proposals with CLIP. Recently, several works [70, 76] build the mask generator upon the frozen diffusion model [54] and CLIP model [53]. Meanwhile, several studies [27, 47, 51, 52] focus on class-agnostic segmentation and detection to enrich generalization ability in various domains. However, most approaches are trained and tested on specific datasets. Our approach is based on SAM, which provides a general, interactive tool to support different open vocabulary detectors. + +Prompting in Computer Vision. Prompting, originating from in-context learning in natural language processing (NLP) as seen in works like Brown et al. [4] and Rubin et al. [55], leverages a large language model to infer unseen tasks through context-specific input-output pairs. Recent studies [1,3,16,45,61,62,88] have explored in-context learning for visual tasks. Common techniques involve mask image modeling [2, 21, 77] for cross-task visual prompting, as employed by approaches like Painter [61] and Bar et al. [3]. SAM [30] demonstrates in-context learning through interactive segmentation, using diverse visual prompts like points, boxes, and masks, although it is limited to class-agnostic mask prediction. Meanwhile, other studies [8,17,23,39,42] have concentrated on efficient parameter tuning of visual foundation models, typically focusing on a single model. Our work uniquely bridges two models, CLIP and SAM, exploring their combined potential for enhanced general segmentation and recognition capabilities. In particular, we adopt visual prompts (box, point) as the model's inputs. Segmentation Anything Model. SAM [30] presents a new data engine and portable model for segmentation. Subsequent research has employed SAM as an interactive segmentation tool for various vision tasks, including grounding [44], tracking [10], distillation [81, 85], medical analysis [5, 65], and generation [82]. While some studies use SAM and CLIP for segmentation [6, 20, 36, 49, 59], none have yet integrated VLMs and SAM into a unified model capable of both segmentation and recognition of novel classes. Our work makes the first attempt to merge the capabilities of VLMs with SAM for enhanced task versatility. + +# 3 Methodology + +We first review the SAM, CLIP, and combined baselines in Sec. 3.1. Then, we detail our Open Vocabulary SAM in Sec. 3.2. Last, we present our model's training details and application in Sec. 3.3. + +# 3.1 Preliminaries and Baselines + +SAM. SAM is a prompt-driven segmentor. It contains an image encoder, a prompt encoder, and a light-weight mask decoder. Here, we use box prompts as an example. We denote an input image as $X \in \mathbb{R}^{H \times W \times 3}$ and input visual + +![](images/41f0f796fe80f45ef8f18e6bcd77ea34d6644844e567929ea8b544f32d1b3e8b.jpg) +Fig. 3: Illustration of Open-Vocabulary SAM. For training, the SAM encoder is as a teacher network, while SAM2CLIP plays the role of a student network and aligns the knowledge of SAM into CLIP. The CLIP2SAM transfers the CLIP knowledge to the SAM decoder and performs joint segmentation and classification for close-set and open vocabulary settings. + +prompts as $P \in \mathbb{R}^{N \times 4}$ , where $H \times W$ are the spatial size, $N$ is the number of box prompts. The image encoder is a modified vision transformer (ViT). It encodes an image into dense feature $F_{SAM} \in \mathbb{R}^{\frac{H}{16} \times \frac{W}{16} \times d}$ . The prompt encoder encodes $P$ into sparse prompts $Q_{sp}$ . Meanwhile, mask tokens $Q_{mask}$ and an IoU token $Q_{IoU}$ are initialized for the mask decoder. + +The mask decoder takes the image feature $F$ , sparse prompts $Q_{sp}$ , mask tokens $Q_{mask}$ , and the IoU token $Q_{IoU}$ as input. All the inputs will be concatenated and encoded with a lightweight two-way transformer. Consequently, each mask token is transformed into a dynamic linear classifier, capable of calculating the foreground mask probability for every sparse prompt. Simultaneously, the IoU token is tasked with predicting the confidence score for each mask. Considering the multi-granular nature of SAM's data annotations, encompassing both instance and part level, $Q_{mask}$ naturally encodes multi-granularity. Our study concentrates exclusively on the object level, which aligns more closely with prevalent real-world applications and datasets such as COCO [43] and LVIS [19]. CLIP. Given an input image $X$ and a corresponding caption $C$ , the CLIP framework processes these modalities to produce respective embeddings: the image embedding $E_I$ , derived from its image encoder, and the text embedding $\mathbf{t}$ , obtained from its text encoder. In the context of open-vocabulary object detection and segmentation, CLIP's capability to generalize beyond fixed class labels is leveraged to replace traditional classifiers. For instance, in open-vocabulary detection scenarios, the text embedding $\mathbf{t_c}$ for the $c$ -th object category is generated by inputting the category name into the CLIP text encoder. This process can employ a single template prompt, such as "a photo of {category}", or multiple prompt templates. Subsequently, for a given region embedding $r$ , that is produced by the RoI-Align [22], the classification score for the $c$ -th category is computed as: $p_c = \frac{\exp(\tau \cdot <\mathbf{r}, \mathbf{t_c} >)}{\sum_{i=0}^{C} \exp(\tau \cdot <\mathbf{r}, \mathbf{t_i} >)}$ , where $<\cdot, \cdot>$ denotes the cosine similarity, and $\tau$ is a learnable or fixed temperature to re-scale the value. + +Combined Baselines. We introduce two different baselines for combining CLIP and SAM, as depicted in Fig. 2 (a) and (b). The first approach, termed the + +'cropped image baseline', employs the SAM mask decoder's output to segment and resize the original input image. This processed image then serves as the input for the CLIP image encoder, and, in conjunction with the CLIP text embedding, the mask is classified using the similarities between visual and text embeddings. The second approach referred to as the 'cropped CLIP image feature baseline', employs the same initial CLIP feature extraction step. However, in this method, masks predicted by the SAM decoder are used to crop the CLIP image features. Subsequent pooling of these masked features yields the final label, akin to baseline (a). + +While both baselines enable zero-shot inference of images, they exhibit a noticeable knowledge gap on specific datasets. To address this, we draw inspiration from recent advancements in visual prompting or adapters [8,88]. Specifically, we propose incorporating additional learnable tokens as an adapter to fine-tune the model for enhanced performance on downstream datasets. These zero-shot inference capabilities and the fine-tuned models constitute our primary comparison baselines under various experimental conditions, detailed in Sec. 4.1. + +# 3.2 Open Vocabulary SAM + +While both baseline models can be enhanced through visual prompting or adapters, as we will discuss in Sec. 4, they face several challenges in real-world applications. First, the requirement for two independent backbones in the combined model increases computational costs (Prob.1). Second, SAM and CLIP are trained with distinct objectives – SAM through supervised learning and CLIP via contrastive learning – and there is limited research on knowledge transfer between such diverse architectures (Prob.2). Third, despite adapter integration, significant performance gaps remain in recognizing small objects (Prob.3). Fourth, there is a lack of exploration into integrating open-vocabulary capabilities for SAM and CLIP, particularly in the context of feature fusion and data scaling (Prob.4). Our work aims to solve these problems in a unified yet effective framework. + +Unified Architecture. We design a unified architecture for both segmentation and recognition to address Prob.1. Specifically, we adopt the frozen CLIP visual encoder as our feature extractor. Then, both SAM's mask decoder and prompt encoder are appended behind the CLIP encoder. The meta-architecture of open-vocabulary SAM is shown in Fig. 2 (c), with the more detailed version shown in Fig. 3. This unified architecture is made possible via the SAM2CLIP, which transfers knowledge of SAM to CLIP with distillation, and CLIP2SAM, which employs CLIP knowledge and combines the SAM mask decoder for recognition. We have chosen convolution-based visual backbones for the frozen CLIP backbone, aligning with previous studies that have highlighted their superiority in capturing spatial structures [31, 73]. The efficacy of different CLIP backbones is further explored in Sec. 4.2. + +SAM2CLIP. To resolve Prob.2, we design the SAM2CLIP module that bridges the gap in feature representations learned by SAM and CLIP, using adaptation and distillation methods. Through comprehensive experiments, we discovered that employing distillation loss $L_{distill}$ along with transformer-based + +adapters [14], yields effective results. Specifically, the distillation process involves a simple pixel-wise approach, where SAM-Huge serves as the teacher, and the frozen CLIP equipped with an adapter assumes the student's role. We then implement a per-pixel mean squared error (MSE) loss to align the SAM feature $F_{sam}$ with the CLIP feature $E_I$ , as detailed below: + +$$ +L _ {d i s t i l l} = \operatorname {M S E} \left(F _ {s a m}, E _ {I}\right). \tag {1} +$$ + +We design a multi-scale adapter $A_{sam2clip}$ to align the features from CLIP and SAM. In particular, we take pyramid CLIP features $E_I^i$ , $i = 1, 2, 3$ as the inputs. Such pyramid features contain both high-resolution and semantic information, which is proven crucial for semantic segmentation [29]. The MSE loss is revised as follows: + +$$ +L _ {d i s t i l l} = \operatorname {M S E} \left(F _ {s a m}, A _ {s a m 2 c l i p} \left(\operatorname {F u s i o n} \left(E _ {I} ^ {i}\right)\right)\right), \tag {2} +$$ + +where $A_{sam2clip}$ comprises several transformer layers, and Fusion is achieved by bilinear upsampling and addition. + +With SAM2CLIP, we can even achieve comparable segmentation results with the SAM-Huge with much lower computational costs. As detailed in Sec. 4.2, we observe that employing convolution-based methods specifically designed for backbone adaptation [8, 15] results in sub-optimal outcomes. The reason for this might be in the inherent architecture of the SAM encoder, which is purely based on ViT. A symmetrical structure is crucial for effective knowledge transfer. With the implementation of SAM2CLIP, we can achieve segmentation results comparable to those of SAM-Huge across various detectors, while significantly reducing computational costs. + +CLIP2SAM. This module aims to leverage CLIP's knowledge to enhance the recognition capabilities of the SAM decoder. A straightforward approach involves appending a label token $Q_{label}$ to the existing mask token $Q_{mask}$ and IoU token $Q_{IoU}$ . Using $Q_{label}$ , we introduce a specialized adapter to facilitate the transfer of knowledge from the frozen CLIP to the SAM decoder. Subsequently, the enhanced $Q_{label}$ , combined with the output of the prompt encoder and adapted CLIP features, is fed into a two-way transformer. Following the cross-attention process, the improved $Q_{label}$ undergoes further refinement through a multilayer perceptron (MLP), ensuring better alignment with CLIP's text embedding. The final labels are derived by calculating the distance between the refined label token and the CLIP text embedding. + +This design, however, falls short of recognizing small objects (Prob.3) since the adaptation only involves the single-scale feature, which is mainly focused on segmentation. We present a simple yet effective solution to handle this issue, introducing a lightweight feature pyramid network (FPN) for CLIP2SAM adaption. As shown in Fig. 3, the pyramid network extracts multi-scale CLIP features as the inputs. Then, we apply the RoI-Align [22] operation to extract region features. Like the R-CNN framework [22], we apply one convolution layer and a MLP to learn the feature embedding without introducing cross-attention + +in the mask decoder. In particular, for point prompts, we first obtain the corresponding masks via the SAM decoder and obtain the box via the corresponding masks. For box prompts, we can directly send it to the FPN for region feature extraction. Given that our method incorporates only a few convolution layers, it does not significantly increase computational costs compared to the original SAM. + +Open Vocabulary. To tackle Prob.4, the open-vocabulary challenge, we leverage the knowledge embedded in the frozen CLIP backbone, which aids in recognizing novel and unseen objects during inference. In line with previous studies [31,67], we fuse the learned class scores with those from the frozen CLIP via a geometric mean to leverage information from both the CLIP and CLIP2SAM. Additionally, we investigate various strategies to expand the vocabulary size, such as joint training with multiple datasets, as detailed in Sec. 4.2. Our experimental results show that the model scales effectively with large datasets. + +While our approach can address the open vocabulary challenge, it is important to distinguish the setting of Open-Vocabulary SAM from that of previous open vocabulary segmentation methods [6,13,24,74]. Unlike previous techniques, which typically depend on a separate segmentor to produce mask proposals, our method only uses visual prompts, such as boxes and points, to generate masks and labels. Furthermore, our method specifically targets foreground objects, aligning with the ImageNet [11] datasets and the purpose of box prompts. To show the effectiveness of our method, we compare the performance of our method with previous open vocabulary segmentation methods in Sec. 4.1. + +# 3.3 Training and Application + +Training and Loss Function. We first use the SA-1B (1%) dataset [30] for training the SAM2CLIP module to transfer SAM's knowledge into open-vocabulary SAM, with the loss $L_{distill}$ (Equ. (2)). Then, we joint train the CLIP2SAM and mask decoder using segmentation mask and label annotations from COCO or LVIS. The final loss function is given as $L = \lambda_{cls}L_{t\_cls} + \lambda_{ce}L_{t\_ce} + \lambda_{dice}L_{t\_dice}$ . Here, $L_{t\_ce}$ is the Cross-Entropy (CE) loss for mask classification, and $\bar{L}_{t\_ce}$ and $L_{t\_dice}$ are mask Cross Entropy (CE) loss and Dice loss [48] for segmentation, respectively. In addition, we adopt joint training with the ImageNet dataset for classifying over 22,000 categories. + +Inference and Demo Tools. Our model performs inference like SAM, with points and boxes as visual prompts. Specifically, we test boxes and points as visual prompts for the encoder in Sec. 4. In the appendix, we show a demo of our model, which can segment and recognize with prompts. + +# 4 Experiments + +Datasets and Metrics. We mainly use COCO [43] and LVIS [19] datasets for the experiments. Moreover, we also use part of SA-1B [30] data (1%) for + +Table 1: Comparison of combined baselines and Open-Vocabulary SAM using visual prompts. \*\* indicates using mask center point as prompts, while others indicate using ground truth boxes prompts. $I o U_{b}$ and $I o U_{n}$ refer to the average IoU for each mask of base classes and novel classes, respectively. + +
MethodCOCO +IoUbIoUn AccLVIS +IoUbIoUn AccFLOPs #Param
Image-Crop baseline78.181.446.278.381.69.6
Feature-Crop baseline78.181.455.178.381.626.5
Image-Crop baseline + CoOp [88]79.682.162.080.182.032.1
Feature-Crop baseline + CoOp [88]79.682.170.980.182.048.2
Open-Vocabulary SAM81.584.084.380.483.166.6
Image-Crop baseline*60.766.724.553.062.36.2
Feature-Crop baseline*60.766.732.153.062.311.0
Image-Crop baseline + CoOp [88]*64.766.728.258.964.28.3
Feature-Crop baseline + CoOp [88]*64.766.735.158.964.213.2
Open-Vocabulary SAM*68.465.276.763.667.960.4
+ +SAM2CLIP knowledge transfer. For COCO, we report the results of both close-set and open-vocabulary settings for the instance segmentation task. In particular, following Zareian et al. [79], we split 48 base classes with annotations and 17 target classes without annotations. We use the base class annotations for training. For LVIS datasets, we adopt the open-vocabulary setting and report the results of $AP_{\text{rare}}$ for novel classes. For evaluation metrics, we report the accuracy of recognition for reference to evaluate the recognition capability. Meanwhile, each prompt's intersection-over-union (IoU) with its ground truth mask is also adopted to evaluate the segmentation ability of our method. As mentioned in Sec. 3.2, different from previous open vocabulary segmentation tasks, boxes or points serve as visual prompts in our method. + +Baselines. As shown in Fig. 2 (a) and (b), based on different adapter designs, we append these adapters to the different locations of the combined models. For example, when using CoOp [88], we append the learnable tokens by combining them with CLIP features. For several convolution-based adapters [8], we add the extra convolution layers along with SAM or CLIP backbone for fair comparison. By default, we adopt SAM-huge and CLIP R50x16. + +Implementation Details. We implement our models in PyTorch [50] with MMDetection [7]. We use 8 A100 GPUs for distributed training. Each minibatch has two images per GPU. The optimizer is AdamW [46] with a weight decay of 0.0001. We adopt full image size for a random crop in the pre-training and training process following Cheng et al. [9]. All the class names are transferred into CLIP text embedding, following previous works [18]. We train each model for 12 epochs for fair comparison. Due to the limitation of computation costs, we do not adopt joint SAM data and COCO data training. We first perform training the SAM2CLIP on SA-1B (1%), and then we finetune the model on COCO or LVIS data. Please refer to the supplementary material for more details. + +Table 2: Comparison of combined baselines and Open-Vocabulary SAM on prompts generated by the open vocabulary detector. For the LVIS dataset, only 'normal' and 'frequent' classes are in the training set. The labels are generated by each baseline or our method. We adopt Detric [89] as the OV-Detector to provide box prompts. + +
MethodCOCOLVISFLOPs#Params
APbaseAPnovelAPAPrareAPnormAPfreqAP
Image-Crop baseline + CoOp [88]26.231.227.319.818.316.317.23,748G808M
Feature-Crop baseline + CoOp [88]28.033.829.524.221.418.620.83,545G808M
Open-Vocabulary SAM31.136.032.424.021.322.922.41,180G304M
+ +Table 3: Comparison of Open-Vocabulary SAM with previous methods on open vocabulary instance segmentation. The results presented in the table all use the Mask-RCNN [22] with ResNet-50 backbone. Different from previous works, Open-Vocabulary SAM uses the bounding box as the prompt to generate the mask. We report mask AP50 following base-novel setting as [24]. + +
MethodVenueConstrainedGeneralized
BaseNovelBaseNovelAll
XPM [24]CVPR'2242.424.041.521.636.3
MaskCLIP [13]ICML'2342.823.242.621.737.2
MasQCLIP [74]ICCV'2340.930.140.728.437.5
Open-Vocabulary SAM(Ours)41.737.539.339.839.4
+ +Table 4: Comparison of mask quality with various detectors on COCO dataset. We report the mask mean AP for comparison. The masks are generated by each method, while the labels are from the corresponding detectors. + +
MethodDetectorsmAPAP50AP75APSAPMAPL#ParamsFLOPs
SAM-HugeFaster-RCNN (R50)35.654.938.417.239.151.4641M3,001G
SAM-Huge (finetuned)Faster-RCNN (R50)35.855.038.416.538.653.0641M3,001G
Open-Vocabulary SAMFaster-RCNN (R50)35.855.638.316.038.953.1304M1,180G
SAM-HugeDetic (swin-base)36.457.139.421.440.854.6641M3,001G
SAM-Huge (finetuned)Detic (swin-base)36.857.439.820.840.655.1641M3,001G
Open-Vocabulary SAMDetic (swin-base)36.757.239.720.740.854.9304M1,180G
SAM-HugeViTDet (Huge)46.372.049.825.245.559.6641M3,001G
SAM-Huge (finetuned)ViTDet (Huge)46.572.350.325.245.860.1641M3,001G
Open-Vocabulary SAMViTDet (Huge)48.873.852.924.846.364.2304M1,180G
+ +Benchmark Setup. Open Vocabulary SAM is different from both SAM [30] and open vocabulary segmentation methods [13,24,74]. Compared with Open-Vocabulary SAM, SAM [30] cannot recognize segmented objects while open vocabulary segmentation methods [13,24,74] usually rely on a separate segmentor (e.g., Mask-RCNN [22]) instead of prompts like points or boxes. For a complete evaluation of Open-Vocabulary SAM, we set the benchmark as three parts, including 1). comparison with combined baselines to verify the effectiveness; 2). comparison with open-vocabulary segmentation methods to evaluate the recognition; 3). comparison with SAM models to evaluate segmentation. + +Table 5: Scaling up with large-scale datasets. + +
DatasetsAccuracy#vocabulary#images
LVIS83.11,20399K
V3Det78.713,204183K
I-21k44.519,16713M
V3Det + LVIS82.713,844282K
V3Det + LVIS + I-21k83.325,89813M
V3Det + LVIS + I-21k + Object36583.025,97015M
+ +Table 6: The effectiveness of each component. We use Detic [89] as the detector. The labels are generated by the corresponding model. S2C and C2S denote SAM2CLIP and CLIP2SAM respectively. The baseline refers to the image-crop variant. + +
SettingAPAPbaseAPnovelFLOPs (G)#Params (M)
Baseline + CoOp [88]29.528.033.83,545808
Our + S2C28.727.333.31,127291
Our + S2C + C2S34.433.138.01,180304
+ +# 4.1 Main Results + +Comparison with Combined Baselines Using Ground Truth. To avoid the influence of other modules, we first demonstrate the recognition ability of our model in Tab. 1. Compared to the simple combined approaches, adding adapters with joint co-training leads to better results. However, the recognition ability is still limited on both COCO and LVIS. Our Open-Vocabulary SAM achieves the best results on both boxes and points as visual prompts. We observe more significant gains on LVIS datasets. We argue that LVIS contains more small objects, which is more challenging than COCO. Our method can solve Prob.2 and lead to over $20\%$ accuracy improvement. Although the segmentation quality is pretty good (about 80 IoU on COCO and LVIS with box prompt), our method still achieves $2\%$ IoU improvements. This indicates the effectiveness of our joint co-training on mask prediction and classification. Compared with boxes as prompts, using points as prompts is more challenging since the location clues of points are much weaker than boxes. However, our approach is still better than combined baselines or them with adapters. + +Comparison with Combined Baselines on OV-Detector. In Tab. 2, we adopt a more challenging setting by using the box prediction from the existing open-vocabulary detector to simulate the interactive segmentation process with deviation. We choose the representative Detic [89] as the open-vocabulary detector. Again, our method also achieves the best performance on both COCO and LVIS datasets. In particular, on COCO, compared with previous works [88], our method achieves 3.0 mask mAP improvements with much lower costs. + +Comparison with Open-vocabulary Segmentation. In Tab. 3, we compare our method with previous open-vocabulary instance segmentation methods. Open-Vocabulary SAM uses the boxes generated by Mask- RCNN [22] and + +Table 7: Ablation studies on COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (left): Ablation on SAM2CLIP design. (right): Ablation on CLIP2SAM design. + +
SettingIoU FLOPs (G) #Params (M)SettingIoU Acc APbase APnovel
SAM78.73,001641SAM2CLIP (baseline)78.154.227.6
Conv-L + MultiScale Neck78.31,313321+ Cls Token81.379.329.8
Conv-L + SingleScale Neck73.61,280307+ Cls Token & CLIP MLP fusion80.978.929.0
R50x16 + MultiScale Neck78.11,180304+ light FPN (Ours)81.584.331.1
R50 + MultiScale Neck77.3728165
+ +Table 8: (left): Ablation study on different CLIP backbone. We test results on the COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (right): Comparison with other SAM models. + +
BackboneIoU Acc #FLOPs(G) #Params (M)
RN5077.3 50.8728165
RN50x1678.1 55.11,180304
RN50x6478.1 54.12,098568
ConvNeXt-L78.3 59.11,313321
ViT-L-1438.6 14.32,294441
+ +
Method1-IoU (COCO)cls. open.
SAM [30] (H)78.2--
SEEM [90] (T)73.7✓-
Semantic-SAM [32] (T)76.1✓-
OV-SAM (ours)81.7✓✓
+ +previous works use the masks of Mask-RCNN. The results show that our Open-Vocabulary SAM performs strongly, especially in the novel classes. + +Comparison with SAM on Various Detectors. In Tab. 4, we test the mask prediction quality of our model and original SAM on different detectors. Our method performs better than the original SAM and performs comparably with fine-tuned SAM. It is worth noting that our Open-Vocabulary SAM has much lower computational costs and parameters than SAM. + +Comparison on Interactive Segmentation. In Tab. 8 (right), we compare interactive segmentation performance. Notably, our approach excels beyond interactive segmentation and can recognize classes in an open-vocabulary setting. Visualization Comparison. In Fig. 4, we compare our approach with the feature-crop baseline. Our model shows better performance in classifying small and rare object classification and handling occlusion scenarios. + +Model as a Zero Shot Annotation Tool. In addition to COCO and LVIS standard datasets training, following the spirit of SAM, we also scale up our model by training it with more data (Tab. 5). In particular, we adopt more detection data (V3Det [60], Object365 [56]) and classification data (ImageNet22k [11]). Owing to significant costs, we have not conducted comparisons with other baselines for this setting. Rather, we have adapted our method into an interactive annotation tool capable of segmenting and recognizing over 22,000 classes. + +# 4.2 Ablation Studies and Analysis + +Effectiveness of SAM2CLIP and CLIP2SAM. We first verify the effectiveness of our proposed two modules in Tab. 6. We adopt image-crop variant of the baseline for comparison. In particular, by sharing a single backbone, we observe + +![](images/d6313c36a159f09236bb855814d903f7f107015db0866a095e03c2955d7f718f.jpg) +Fig. 4: Visualization Comparison. We compare the mask and classification results of the image-crop baseline (SAM + CLIP) and Open-Vocabulary SAM (OV SAM). The predicted labels are presented on the mask. + +a significant drop in the number of parameters and FLOPs, with a little drop in segmentation performance. The slight drop is caused by the domain gap between SAM data and COCO data during the training of the SAM2CLIP module. However, after adding our CLIP2SAM module and joint co-training with mask classification and prediction, a significant improvement in both segmentation and classification is observed, with just a negligible increase in compute cost. + +Detailed Design on SAM2CLIP. In Tab. 7, we explore the detailed design of SAM2CLIP in the first stage of open-vocabulary SAM training. The results show that distillation benefits most when multi-scale features are adopted, suggesting that both high-resolution features and high-level semantics are important to align CLIP's feature with the SAM's feature. + +Detailed Design on CLIP2SAM. In Tab. 7, we present extensive design for the CLIP2SAM module. We compare two designs: a simple classification token with cross attention (Cls Token) and a combination of this token with mask pooled CLIP feature (CLS Token & CLIP MLP fusion). These designs work better than the combined baseline shown in the first row. Nonetheless, due to resolution constraints, these variants cannot handle small objects well, as shown in Fig. 4. In contrast, our design improves the performance considerably. + +Ablation on Different CLIP Backbones. In Tab. 8, we explore the effect of frozen CLIP visual backbone. We do not add the CLIP2SAM. Motivated by recent works [31,67,73,76], CNN-based CLIPs encapsulate more structural information, which is good for our goal since we have location-sensitive visual prompts. Thus, adopt CNN-based CLIPs, aligned with previous works. As shown + +![](images/86073dd589b87850741a8711cbd5818db5351f08272fe52c5b27f7081571749e.jpg) + +![](images/40c3b8fc36b2611ec703be7c0697015b1721136679dd41c15a7d98e9fcc94efa.jpg) + +![](images/272da928ca6b914a200c89de556f6229024baf88ca0c863ac0b1db449c613a3f.jpg) + +![](images/f4334d26a305ab5ff9ca5c37b2ab0548afcfa594cd5b7792c7c4d9c54f41f92a.jpg) + +![](images/18b84b886e35bdd4419970d61e0c59756f35d2f22a6e91395d9b94e00972133c.jpg) + +![](images/8ee2ece8e7a8041242d99d76926ea13c7d1fafa9c227801ce597acf9e3aeffb8.jpg) + +![](images/35b3153c1e006fbca89453d14eb1201b03473d9a0da96ab4516a83d2ea97b9ae.jpg) + +![](images/132df602141ec273e4a06ea2eb3288e0daa0cbfba59a2ca3690716b1b1d0e8da.jpg) + +![](images/3c37d9c18656873ff7c01e4a9314c867cd3c8b8ae20e96d3e68e03961cb8a0f7.jpg) + +![](images/602e18d6d337cb3b4dc3003acacca4f334ab6fddbadf5caa9e835d8380375a26.jpg) + +![](images/97e4442e4ea0a5ba3e9aa030b6741888660ab91e74f5bebeeefa6337ef0b18e8.jpg) +Fig. 5: Qualitative results of Open-Vocabulary SAM. In the visualization, boxes refer to box prompts (up), and the red stars refer to the point prompts (middle). The masks can be merged into a segmentation map (bottom). We show the mask and labels generated by the proposed Open-Vocabulary SAM. Our method can segment and recognize open vocabulary objects in diverse scenes identified by prompts. + +![](images/6d8e6627bdbf8bfe7ab8d9af8261374ff569ec69d675c86b541c6fa985134cc1.jpg) + +![](images/4ceaf8098f57f5b9958e3bc7c0b047e8531aa48981fd85898a20d212e91a5cf7.jpg) + +in the table, we find ConvNext large achieves the best performance, but we choose RN50x16 since it has comparable performance and better efficiency. + +# 5 Conclusion + +We present Open Vocabulary SAM, a SAM-inspired method for interactive segmentation and recognition. Unlike previous open-vocabulary detection and segmentation methods, our method explores interactive open-vocabulary segmentation for the first time. Given the user's inputs, such as boxes or points, the proposed approach can interactively segment and label each visual prompt. Compared with the combined baselines and various visual adapters, our proposed CLIP2SAM and SAM2CLIP are both efficient and effective in various settings. Our open vocabulary segmentation is compatible with different detectors, including open-vocabulary detectors and close-set detectors. With more data, our model plays a similar role as SAM, offering an effective annotation tool for both segmentation and instance labeling. In particular, our method can perform large vocabulary segmentation and recognition over 22K classes. We hope our Open-Vocabulary SAM can provide a solid baseline for combining the strengths of different forms of vision foundation models and inspire further research. + +Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). The project is also supported by Singapore MOE AcRF Tier 1 (RG16/21) and the National Key R&D Program of China (No. 2022ZD0161600). + +# References + +1. Balăzevic, I., Steiner, D., Parthasarathy, N., Arandjelović, R., Henaff, O.J.: Towards in-context scene understanding. In: NeurIPS (2023) +2. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2022) +3. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. In: NeurIPS (2022) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) +5. Chen, C., Miao, J., Wu, D., Yan, Z., Kim, S., Hu, J., Zhong, A., Liu, Z., Sun, L., Li, X., Liu, T., Heng, P.A., Li, Q.: MA-SAM: Modality-agnostic SAM adaptation for 3d medical image segmentation. arXiv preprint arXiv:2309.08842 (2023) +6. Chen, J., Yang, Z., Zhang, L.: Semantic segment anything (2023) +7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019) +8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023) +9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022) +0. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023) +1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009) +2. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022) +3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023) +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +5. Fang, Y., Yang, S., Wang, S., Ge, Y., Shan, Y., Wang, X.: Unleashing vanilla vision transformer with masked image modeling for object detection. In: ICCV (2023) +6. Fang, Z., Li, X., Li, X., Buhmann, J.M., Loy, C.C., Liu, M.: Explore in-context learning for 3d point cloud understanding. In: NeurIPS (2023) +7. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: CLIP-Adapter: Better vision-language models with feature adapters. IJCV (2024) +8. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021) + +19. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019) +20. Han, X., Wei, L., Yu, X., Dou, Z., He, X., Wang, K., Han, Z., Tian, Q.: Boosting segment anything model towards open-vocabulary learning. arXiv preprint arXiv:2312.03628 (2023) +21. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) +22. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV (2017) +23. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022) +24. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: CVPR (2022) +25. Jayaraman, D., Grauman, K.: Zero-shot recognition with unreliable attributes. In: NeurIPS (2014) +26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) +27. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. RA-L (2022) +28. Kim, W., Son, B., Kim, I.: ViLT: Vision-and-language transformer without convolution or region supervision. In: ICML (2021) +29. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: CVPR (2019) +30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023) +31. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A.J., Angelova, A.: F-VLM: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023) +32. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-SAM: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023) +33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023) +34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022) +35. Li, J., Selvaraju, R.R., Gotmare, A., Joty, S.R., Xiong, C., Hoi, S.C.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021) +36. Li, S., Cao, J., Ye, P., Ding, Y., Tu, C., Chen, T.: ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665 (2024) +37. Li, X., Ding, H., Zhang, W., Yuan, H., Cheng, G., Jiangmiao, P., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv pre-print (2023) +38. Li, X., You, A., Zhu, Z., Zhao, H., Yang, M., Yang, K., Tong, Y.: Semantic flow for fast and accurate scene parsing. In: ECCV (2020) +39. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: OMG-Seg: Is one model good enough for all segmentation? In: CVPR (2024) +40. Li, X., Zhang, J., Yang, Y., Cheng, G., Yang, K., Tong, Y., Tao, D.: Sfnet: Faster and accurate domain agnostic semantic segmentation via semantic flow. IJCV (2023) + +41. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: CVPR (2023) +42. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: NeurIPS (2022) +43. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014) +44. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +45. Liu, Y., Qi, L., Tsai, Y.J., Li, X., Chan, K.C.K., Yang, M.H.: Effective adapter for face recognition in the wild. arXiv preprint (2023) +46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) +47. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022) +48. Miletari, F., Navab, N., Ahmadi, S.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. In: 3DV (2016) +49. Pan, T., Tang, L., Wang, X., Shan, S.: Tokenize anything via prompting. arXiv preprint arXiv:2312.09128 (2023) +50. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019) +51. Qi, L., Kuen, J., Shen, T., Gu, J., Guo, W., Jia, J., Lin, Z., Yang, M.H.: High-quality entity segmentation. In: ICCV (2023) +52. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE TPAMI (2022) +53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +55. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv:2112.08633 (2021) +56. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: ICCV (2019) +57. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV (2023) +58. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: EVA-CLIP: Improved training techniques for CLIP at scale. arXiv preprint arXiv:2303.15389 (2023) +59. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: SAM-CLIP: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023) +60. Wang, J., Zhang, P., Chu, T., Cao, Y., Zhou, Y., Wu, T., Wang, B., He, C., Lin, D.: V3det: Vast vocabulary visual detection dataset. In: ICCV (2023) +61. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR (2023) +62. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. In: ICCV (2023) +63. Wu, J., Li, X., Ding, H., Li, X., Cheng, G., Tong, Y., Loy, C.C.: Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. In: ICCV (2023) + +64. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024) +65. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023) +66. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: CVPR (2023) +67. Wu, S., Zhang, W., Xu, L., Jin, S., Li, X., Liu, W., Loy, C.C.: CLIPSelf: Vision transformer distills itself for open-vocabulary dense prediction. In: ICLR (2024) +68. Wu, X., Zhu, F., Zhao, R., Li, H.: CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In: CVPR (2023) +69. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023) +70. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) +71. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023) +72. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022) +73. Xu, S., Li, X., Wu, S., Zhang, W., Li, Y., Cheng, G., Tong, Y., Chen, K., Loy, C.C.: Dst-det: Simple dynamic self-training for open-vocabulary object detection. arXiv pre-print (2023) +74. Xu, X., Xiong, T., Ding, Z., Tu, Z.: Masqclip for open-vocabulary universal image segmentation. In: ICCV (2023) +75. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., et al.: Panoptic video scene graph generation. In: CVPR (2023) +76. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional CLIP. In: NeurIPS (2023) +77. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022) +78. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV (2022) +79. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021) +80. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: LiT: Zero-shot transfer with locked-image text tuning. In: CVPR (2022) +81. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight SAM for mobile applications. arXiv preprint arXiv:2306.14289 (2023) +82. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023) +83. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023) +84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019) +85. Zhou, C., Li, X., Loy, C.C., Dai, B.: EdgeSAM: Prompt-in-the-loop distillation for on-device deployment of SAM. arXiv preprint arXiv:2312.06660 (2023) + +86. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from CLIP. In: ECCV (2022) +87. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. arXiv preprint arXiv:2311.03352 (2023) +88. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV (2022) +89. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) +90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: NeurIPS (2023) \ No newline at end of file diff --git a/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/images.zip b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dcc6ecb1a5aef81d42bb7e0baf1ef1c82e604252 --- /dev/null +++ b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2d6246b40ca5453debd18f362769c37c8aef7f4369442053e42afc35779d8ac +size 674987 diff --git a/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/layout.json b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b607abe06925a47f96180d19b35d29ab0de35fb5 --- /dev/null +++ b/2024/Open-Vocabulary SAM_ Segment and Recognize Twenty-thousand Classes Interactively/layout.json @@ -0,0 +1,9890 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "Open-Vocabulary SAM: Segment and Recognize Twenty-thousand Classes Interactively" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "text", + "content": "Haobo Yuan" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Xiangtai Li" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Chong Zhou" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Yining Li" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "text", + "content": ", Kai Chen" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "text", + "content": ", and Chen Change Loy" + }, + { + "bbox": [ + 135, + 167, + 479, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 217, + 201, + 395, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 201, + 395, + 213 + ], + "spans": [ + { + "bbox": [ + 217, + 201, + 395, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 217, + 201, + 395, + 213 + ], + "type": "text", + "content": " S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 252, + 213, + 361, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 213, + 361, + 224 + ], + "spans": [ + { + "bbox": [ + 252, + 213, + 361, + 224 + ], + "type": "text", + "content": "2 Shanghai AI Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 224, + 429, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 224, + 429, + 236 + ], + "spans": [ + { + "bbox": [ + 184, + 224, + 429, + 236 + ], + "type": "text", + "content": "Project Page: https://www.mmlab-ntu.com/project/ovsam" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 236, + 421, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 236, + 421, + 246 + ], + "spans": [ + { + "bbox": [ + 192, + 236, + 421, + 246 + ], + "type": "text", + "content": "E-mail: yuanhaobo@whu.edu.cn, xiangtai94@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 277, + 455, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 277, + 455, + 453 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 455, + 453 + ], + "type": "text", + "content": "Abstract. The CLIP and Segment Anything Model (SAM) are remarkable vision foundation models (VFMs). SAM excels in segmentation tasks across diverse domains, whereas CLIP is renowned for its zero-shot recognition capabilities. This paper presents an in-depth exploration of integrating these two models into a unified framework. Specifically, we introduce the Open-Vocabulary SAM, a SAM-inspired model designed for simultaneous interactive segmentation and recognition, leveraging two unique knowledge transfer modules: SAM2CLIP and CLIP2SAM. The former adapts SAM's knowledge into the CLIP via distillation and learnable transformer adapters, while the latter transfers CLIP knowledge into SAM, enhancing its recognition capabilities. Extensive experiments on various datasets and detectors show the effectiveness of Open-Vocabulary SAM in both segmentation and recognition tasks, significantly outperforming the naive baselines of simply combining SAM and CLIP. Furthermore, aided with image classification data training, our method can segment and recognize approximately 22,000 classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 463, + 416, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 463, + 416, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 463, + 416, + 475 + ], + "type": "text", + "content": "Keywords: Scene Understanding " + }, + { + "bbox": [ + 160, + 463, + 416, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 463, + 416, + 475 + ], + "type": "text", + "content": " Promptable Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 496, + 230, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 230, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 230, + 509 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "The Segment Anything Model (SAM) [30] and CLIP [53] have made significant strides in various vision tasks, showcasing remarkable generalization capabilities in segmentation and recognition, respectively. SAM, in particular, has been trained with a massive dataset of mask labels, making it highly adaptable to a wide range of downstream tasks through interactive prompts. On the other hand, CLIP's training with billions of text-image pairs has given it an unprecedented ability in zero-shot visual recognition. This has led to numerous studies [18,64,72,78] exploring the extension of CLIP to open vocabulary tasks, such as detection and segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "While SAM and CLIP offer considerable advantages, they also have inherent limitations in their original designs. SAM, for instance, lacks the capability to recognize the segments it identifies. Efforts to overcome this by integrating a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 479, + 235 + ], + "type": "image", + "image_path": "6d08e96c2b61e9aa040e1b9d3bd0377d9bb614b4f3b9d2cd64f68dc34bde8d6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": "Fig. 1: Open-Vocabulary SAM not only can segment anything with prompts just like SAM but also has the capability of recognition in the real world, like CLIP. With drastically lower computational cost, Open-Vocabulary SAM has a higher recognition performance than directly combining SAM and CLIP with image or feature cropping (measured on the COCO open vocabulary benchmark)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 460 + ], + "type": "text", + "content": "classification head have been made [32,83], but these solutions are constrained to specific datasets or closed-set settings. On the other hand, CLIP, which is trained using image-level contrastive losses, faces challenges in adapting its representations for dense prediction tasks. To address this, several studies [18,57,66,67,76] have investigated ways to align CLIP's representation for dense predictions. However, these approaches tend to be dataset-specific and not universally applicable. For example, some research has focused on open vocabulary segmentation on the ADE-20k [84] dataset, using the COCO [43] dataset for pre-training. Merging SAM and CLIP in a naive manner, as illustrated in Fig. 2 (a) and (b), proves to be inefficient. This approach incurs substantial computational expenses and yields subpar results, including recognition of small-scale objects, as evidenced by our experimental results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "text", + "content": "In this study, we address these challenges with a unified encoder-decoder framework that integrates a CLIP encoder and a SAM decoder, as depicted in Fig. 2 (c). To bridge these two distinct components effectively, we introduce two novel modules, " + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "inline_equation", + "content": "SAM2CLIP" + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "inline_equation", + "content": "CLIP2SAM" + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "text", + "content": ", facilitating dual knowledge transfer. First, we distill knowledge from the SAM encoder to the CLIP encoder using " + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "inline_equation", + "content": "SAM2CLIP" + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "text", + "content": ". This distillation process is uniquely executed not directly on the CLIP encoder, which is kept frozen to maintain its existing knowledge, but rather on a lightweight transformer-like adapter using a pixel-wise distillation loss. The adapter takes multi-scale features as input, with the goal of aligning CLIP features with SAM representation. On the decoding side, the " + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "inline_equation", + "content": "CLIP2SAM" + }, + { + "bbox": [ + 130, + 461, + 482, + 628 + ], + "type": "text", + "content": " module transfers knowledge from the frozen CLIP encoder to the SAM decoder. In particular, we design a feature pyramid adapter with a RoIAlign operator to be jointly trained with the SAM decoder. Both modules are lightweight and naturally combine the strengths of CLIP and SAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Following the spirit of SAM, we enhance our model's recognition capabilities by harnessing the power of established semantic datasets, including COCO [43], LVIS [19], and ImageNet-22k [11]. This strategy elevates our model to the ver" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 118, + 246, + 202 + ], + "blocks": [ + { + "bbox": [ + 141, + 118, + 246, + 202 + ], + "lines": [ + { + "bbox": [ + 141, + 118, + 246, + 202 + ], + "spans": [ + { + "bbox": [ + 141, + 118, + 246, + 202 + ], + "type": "image", + "image_path": "02fb8c05999b7fa2ffcd113ee85a8e4675525b43f314fb2d20cc0715817dd86c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 207, + 231, + 217 + ], + "lines": [ + { + "bbox": [ + 148, + 207, + 231, + 217 + ], + "spans": [ + { + "bbox": [ + 148, + 207, + 231, + 217 + ], + "type": "text", + "content": "(a) Image Cropping Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 118, + 352, + 202 + ], + "blocks": [ + { + "bbox": [ + 247, + 118, + 352, + 202 + ], + "lines": [ + { + "bbox": [ + 247, + 118, + 352, + 202 + ], + "spans": [ + { + "bbox": [ + 247, + 118, + 352, + 202 + ], + "type": "image", + "image_path": "036dd7cd4526dbb8dea5ba2796fbe8704ec17769b5b5cedcbfa6c7888ca8bfc0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 207, + 340, + 217 + ], + "lines": [ + { + "bbox": [ + 254, + 207, + 340, + 217 + ], + "spans": [ + { + "bbox": [ + 254, + 207, + 340, + 217 + ], + "type": "text", + "content": "(b) Feature Cropping Baseline" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 354, + 118, + 476, + 205 + ], + "blocks": [ + { + "bbox": [ + 354, + 118, + 476, + 205 + ], + "lines": [ + { + "bbox": [ + 354, + 118, + 476, + 205 + ], + "spans": [ + { + "bbox": [ + 354, + 118, + 476, + 205 + ], + "type": "image", + "image_path": "dd97026e7811052495b98f88b11f2c801374954c3d1434e44db7e51e597e0f68.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 207, + 441, + 217 + ], + "lines": [ + { + "bbox": [ + 364, + 207, + 441, + 217 + ], + "spans": [ + { + "bbox": [ + 364, + 207, + 441, + 217 + ], + "type": "text", + "content": "(c) Open-Vocabulary SAM" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 228, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 481, + 275 + ], + "type": "text", + "content": "Fig. 2: Comparison of two simple SAM-CLIP combination baselines (a) and (b), and our proposed single encoder architecture (c). The adapters for (a) and (b) are optional and can be replaced with various designs (please refer to Sec. 4.1 for details). Note that, in our method, the SAM encoder will be discarded during inference." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 288, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 480, + 335 + ], + "type": "text", + "content": "satility of SAM, endowing it with enhanced capability to segment and recognize any objects, as shown in Fig. 1. As our approach is an adaptation of SAM, it is flexible enough to be integrated with various detectors, making it suitable for both closed-set and open-set environments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "text", + "content": "We conduct extensive experiments across a range of datasets and scenarios, encompassing closed-set and open-vocabulary interactive segmentation. Notably, when compared to basic combined baselines, our approach demonstrates superior performance, achieving over " + }, + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "text", + "content": " improvement in IoU and " + }, + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "text", + "content": " in mAP with various detectors on the COCO dataset. In particular, in the case of recognition on LVIS, our approach achieves over " + }, + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 336, + 480, + 443 + ], + "type": "text", + "content": " improvements over previous adapters. Furthermore, by expanding our approach with a more diverse array of datasets, we have developed a versatile, interactive tool suitable for practical applications. For detailed results, we direct the reader to Sec. 4 and the appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 461, + 237, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 237, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 237, + 474 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "content": "Vision Language Models (VLMs). Vision-language pre-training has given rise to models with aligned image and text representations [25, 26, 28, 41, 53]. Recent studies on contrastive vision-language pre-training [26, 53, 58, 80] have significantly improved the generalization ability of recognition models. Meanwhile, several works [28, 33-35] aim to design better optimization goals for downstream multi-modal tasks, including caption and visual question answering. Among these works, CLIP models [53] that are pre-trained on billion-scale image-text pairs have shown impressive zero-shot classification performance on a wide range of datasets. Our goal is to enable SAM to perform recognition tasks with the help of pre-trained VLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Open Vocabulary Dense Prediction. This direction aims to recognize region visual concepts of arbitrary categories described by texts, which includes object detection [18,63,64,69,79], semantic segmentation [37,38,40,71,86,87], and panoptic segmentation [70,75,76]. This necessitates the alignment between region and text representations with the help of VLMs [26,53,58]. For open-vocabulary" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "detection, a series of works [18, 57, 66, 78] distill knowledge from the CLIP models to recognize novel objects. In contrast to distillation-based methods, several works [31, 68] directly build object detectors upon frozen CLIP CNNs. For open-vocabulary segmentation, the typical works [12, 70, 72, 76] first generate class-agnostic mask proposals and then classify the proposals with CLIP. Recently, several works [70, 76] build the mask generator upon the frozen diffusion model [54] and CLIP model [53]. Meanwhile, several studies [27, 47, 51, 52] focus on class-agnostic segmentation and detection to enrich generalization ability in various domains. However, most approaches are trained and tested on specific datasets. Our approach is based on SAM, which provides a general, interactive tool to support different open vocabulary detectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 511 + ], + "type": "text", + "content": "Prompting in Computer Vision. Prompting, originating from in-context learning in natural language processing (NLP) as seen in works like Brown et al. [4] and Rubin et al. [55], leverages a large language model to infer unseen tasks through context-specific input-output pairs. Recent studies [1,3,16,45,61,62,88] have explored in-context learning for visual tasks. Common techniques involve mask image modeling [2, 21, 77] for cross-task visual prompting, as employed by approaches like Painter [61] and Bar et al. [3]. SAM [30] demonstrates in-context learning through interactive segmentation, using diverse visual prompts like points, boxes, and masks, although it is limited to class-agnostic mask prediction. Meanwhile, other studies [8,17,23,39,42] have concentrated on efficient parameter tuning of visual foundation models, typically focusing on a single model. Our work uniquely bridges two models, CLIP and SAM, exploring their combined potential for enhanced general segmentation and recognition capabilities. In particular, we adopt visual prompts (box, point) as the model's inputs. Segmentation Anything Model. SAM [30] presents a new data engine and portable model for segmentation. Subsequent research has employed SAM as an interactive segmentation tool for various vision tasks, including grounding [44], tracking [10], distillation [81, 85], medical analysis [5, 65], and generation [82]. While some studies use SAM and CLIP for segmentation [6, 20, 36, 49, 59], none have yet integrated VLMs and SAM into a unified model capable of both segmentation and recognition of novel classes. Our work makes the first attempt to merge the capabilities of VLMs with SAM for enhanced task versatility." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 529, + 233, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 233, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 233, + 544 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 555, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 592 + ], + "type": "text", + "content": "We first review the SAM, CLIP, and combined baselines in Sec. 3.1. Then, we detail our Open Vocabulary SAM in Sec. 3.2. Last, we present our model's training details and application in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 609, + 300, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 300, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 300, + 621 + ], + "type": "text", + "content": "3.1 Preliminaries and Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "SAM. SAM is a prompt-driven segmentor. It contains an image encoder, a prompt encoder, and a light-weight mask decoder. Here, we use box prompts as an example. We denote an input image as " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " and input visual" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 117, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 479, + 205 + ], + "type": "image", + "image_path": "41f0f796fe80f45ef8f18e6bcd77ea34d6644844e567929ea8b544f32d1b3e8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 214, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 270 + ], + "type": "text", + "content": "Fig. 3: Illustration of Open-Vocabulary SAM. For training, the SAM encoder is as a teacher network, while SAM2CLIP plays the role of a student network and aligns the knowledge of SAM into CLIP. The CLIP2SAM transfers the CLIP knowledge to the SAM decoder and performs joint segmentation and classification for close-set and open vocabulary settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": "prompts as " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 4}" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": " are the spatial size, " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": " is the number of box prompts. The image encoder is a modified vision transformer (ViT). It encodes an image into dense feature " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "F_{SAM} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times d}" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": ". The prompt encoder encodes " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": " into sparse prompts " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "Q_{sp}" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": ". Meanwhile, mask tokens " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "Q_{mask}" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": " and an IoU token " + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "inline_equation", + "content": "Q_{IoU}" + }, + { + "bbox": [ + 130, + 297, + 479, + 357 + ], + "type": "text", + "content": " are initialized for the mask decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": "The mask decoder takes the image feature " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", sparse prompts " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q_{sp}" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", mask tokens " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q_{mask}" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", and the IoU token " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q_{IoU}" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": " as input. All the inputs will be concatenated and encoded with a lightweight two-way transformer. Consequently, each mask token is transformed into a dynamic linear classifier, capable of calculating the foreground mask probability for every sparse prompt. Simultaneously, the IoU token is tasked with predicting the confidence score for each mask. Considering the multi-granular nature of SAM's data annotations, encompassing both instance and part level, " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "Q_{mask}" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": " naturally encodes multi-granularity. Our study concentrates exclusively on the object level, which aligns more closely with prevalent real-world applications and datasets such as COCO [43] and LVIS [19]. CLIP. Given an input image " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": " and a corresponding caption " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", the CLIP framework processes these modalities to produce respective embeddings: the image embedding " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", derived from its image encoder, and the text embedding " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", obtained from its text encoder. In the context of open-vocabulary object detection and segmentation, CLIP's capability to generalize beyond fixed class labels is leveraged to replace traditional classifiers. For instance, in open-vocabulary detection scenarios, the text embedding " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{t_c}" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": "-th object category is generated by inputting the category name into the CLIP text encoder. This process can employ a single template prompt, such as \"a photo of {category}\", or multiple prompt templates. Subsequently, for a given region embedding " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", that is produced by the RoI-Align [22], the classification score for the " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": "-th category is computed as: " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "p_c = \\frac{\\exp(\\tau \\cdot <\\mathbf{r}, \\mathbf{t_c} >)}{\\sum_{i=0}^{C} \\exp(\\tau \\cdot <\\mathbf{r}, \\mathbf{t_i} >)}" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "<\\cdot, \\cdot>" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": " denotes the cosine similarity, and " + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 360, + 482, + 639 + ], + "type": "text", + "content": " is a learnable or fixed temperature to re-scale the value." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Combined Baselines. We introduce two different baselines for combining CLIP and SAM, as depicted in Fig. 2 (a) and (b). The first approach, termed the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "'cropped image baseline', employs the SAM mask decoder's output to segment and resize the original input image. This processed image then serves as the input for the CLIP image encoder, and, in conjunction with the CLIP text embedding, the mask is classified using the similarities between visual and text embeddings. The second approach referred to as the 'cropped CLIP image feature baseline', employs the same initial CLIP feature extraction step. However, in this method, masks predicted by the SAM decoder are used to crop the CLIP image features. Subsequent pooling of these masked features yields the final label, akin to baseline (a)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": "While both baselines enable zero-shot inference of images, they exhibit a noticeable knowledge gap on specific datasets. To address this, we draw inspiration from recent advancements in visual prompting or adapters [8,88]. Specifically, we propose incorporating additional learnable tokens as an adapter to fine-tune the model for enhanced performance on downstream datasets. These zero-shot inference capabilities and the fine-tuned models constitute our primary comparison baselines under various experimental conditions, detailed in Sec. 4.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 324, + 279, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 279, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 279, + 337 + ], + "type": "text", + "content": "3.2 Open Vocabulary SAM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 488, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 488, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 488, + 474 + ], + "type": "text", + "content": "While both baseline models can be enhanced through visual prompting or adapters, as we will discuss in Sec. 4, they face several challenges in real-world applications. First, the requirement for two independent backbones in the combined model increases computational costs (Prob.1). Second, SAM and CLIP are trained with distinct objectives – SAM through supervised learning and CLIP via contrastive learning – and there is limited research on knowledge transfer between such diverse architectures (Prob.2). Third, despite adapter integration, significant performance gaps remain in recognizing small objects (Prob.3). Fourth, there is a lack of exploration into integrating open-vocabulary capabilities for SAM and CLIP, particularly in the context of feature fusion and data scaling (Prob.4). Our work aims to solve these problems in a unified yet effective framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 617 + ], + "type": "text", + "content": "Unified Architecture. We design a unified architecture for both segmentation and recognition to address Prob.1. Specifically, we adopt the frozen CLIP visual encoder as our feature extractor. Then, both SAM's mask decoder and prompt encoder are appended behind the CLIP encoder. The meta-architecture of open-vocabulary SAM is shown in Fig. 2 (c), with the more detailed version shown in Fig. 3. This unified architecture is made possible via the SAM2CLIP, which transfers knowledge of SAM to CLIP with distillation, and CLIP2SAM, which employs CLIP knowledge and combines the SAM mask decoder for recognition. We have chosen convolution-based visual backbones for the frozen CLIP backbone, aligning with previous studies that have highlighted their superiority in capturing spatial structures [31, 73]. The efficacy of different CLIP backbones is further explored in Sec. 4.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "SAM2CLIP. To resolve Prob.2, we design the SAM2CLIP module that bridges the gap in feature representations learned by SAM and CLIP, using adaptation and distillation methods. Through comprehensive experiments, we discovered that employing distillation loss " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{distill}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " along with transformer-based" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": "adapters [14], yields effective results. Specifically, the distillation process involves a simple pixel-wise approach, where SAM-Huge serves as the teacher, and the frozen CLIP equipped with an adapter assumes the student's role. We then implement a per-pixel mean squared error (MSE) loss to align the SAM feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "F_{sam}" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " with the CLIP feature " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": ", as detailed below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 250, + 188, + 481, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 188, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 250, + 188, + 481, + 200 + ], + "type": "interline_equation", + "content": "L _ {d i s t i l l} = \\operatorname {M S E} \\left(F _ {s a m}, E _ {I}\\right). \\tag {1}", + "image_path": "80a578d690b6f12bacdc9b7010f75865a2a205cb8b3d58a1c987bd2b7e09d562.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "text", + "content": "We design a multi-scale adapter " + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "inline_equation", + "content": "A_{sam2clip}" + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "text", + "content": " to align the features from CLIP and SAM. In particular, we take pyramid CLIP features " + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "inline_equation", + "content": "E_I^i" + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "inline_equation", + "content": "i = 1, 2, 3" + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "text", + "content": " as the inputs. Such pyramid features contain both high-resolution and semantic information, which is proven crucial for semantic segmentation [29]. The MSE loss is revised as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 282, + 481, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 282, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 209, + 282, + 481, + 295 + ], + "type": "interline_equation", + "content": "L _ {d i s t i l l} = \\operatorname {M S E} \\left(F _ {s a m}, A _ {s a m 2 c l i p} \\left(\\operatorname {F u s i o n} \\left(E _ {I} ^ {i}\\right)\\right)\\right), \\tag {2}", + "image_path": "431fbd54ce712521fd2003910656edada301230f3b75bdc4e7c6bbfd7f4fc185.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 304, + 481, + 327 + ], + "type": "inline_equation", + "content": "A_{sam2clip}" + }, + { + "bbox": [ + 130, + 304, + 481, + 327 + ], + "type": "text", + "content": " comprises several transformer layers, and Fusion is achieved by bilinear upsampling and addition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 437 + ], + "type": "text", + "content": "With SAM2CLIP, we can even achieve comparable segmentation results with the SAM-Huge with much lower computational costs. As detailed in Sec. 4.2, we observe that employing convolution-based methods specifically designed for backbone adaptation [8, 15] results in sub-optimal outcomes. The reason for this might be in the inherent architecture of the SAM encoder, which is purely based on ViT. A symmetrical structure is crucial for effective knowledge transfer. With the implementation of SAM2CLIP, we can achieve segmentation results comparable to those of SAM-Huge across various detectors, while significantly reducing computational costs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": "CLIP2SAM. This module aims to leverage CLIP's knowledge to enhance the recognition capabilities of the SAM decoder. A straightforward approach involves appending a label token " + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q_{label}" + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": " to the existing mask token " + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q_{mask}" + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": " and IoU token " + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q_{IoU}" + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q_{label}" + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": ", we introduce a specialized adapter to facilitate the transfer of knowledge from the frozen CLIP to the SAM decoder. Subsequently, the enhanced " + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q_{label}" + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": ", combined with the output of the prompt encoder and adapted CLIP features, is fed into a two-way transformer. Following the cross-attention process, the improved " + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q_{label}" + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": " undergoes further refinement through a multilayer perceptron (MLP), ensuring better alignment with CLIP's text embedding. The final labels are derived by calculating the distance between the refined label token and the CLIP text embedding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "This design, however, falls short of recognizing small objects (Prob.3) since the adaptation only involves the single-scale feature, which is mainly focused on segmentation. We present a simple yet effective solution to handle this issue, introducing a lightweight feature pyramid network (FPN) for CLIP2SAM adaption. As shown in Fig. 3, the pyramid network extracts multi-scale CLIP features as the inputs. Then, we apply the RoI-Align [22] operation to extract region features. Like the R-CNN framework [22], we apply one convolution layer and a MLP to learn the feature embedding without introducing cross-attention" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": "in the mask decoder. In particular, for point prompts, we first obtain the corresponding masks via the SAM decoder and obtain the box via the corresponding masks. For box prompts, we can directly send it to the FPN for region feature extraction. Given that our method incorporates only a few convolution layers, it does not significantly increase computational costs compared to the original SAM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 284 + ], + "type": "text", + "content": "Open Vocabulary. To tackle Prob.4, the open-vocabulary challenge, we leverage the knowledge embedded in the frozen CLIP backbone, which aids in recognizing novel and unseen objects during inference. In line with previous studies [31,67], we fuse the learned class scores with those from the frozen CLIP via a geometric mean to leverage information from both the CLIP and CLIP2SAM. Additionally, we investigate various strategies to expand the vocabulary size, such as joint training with multiple datasets, as detailed in Sec. 4.2. Our experimental results show that the model scales effectively with large datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 392 + ], + "type": "text", + "content": "While our approach can address the open vocabulary challenge, it is important to distinguish the setting of Open-Vocabulary SAM from that of previous open vocabulary segmentation methods [6,13,24,74]. Unlike previous techniques, which typically depend on a separate segmentor to produce mask proposals, our method only uses visual prompts, such as boxes and points, to generate masks and labels. Furthermore, our method specifically targets foreground objects, aligning with the ImageNet [11] datasets and the purpose of box prompts. To show the effectiveness of our method, we compare the performance of our method with previous open vocabulary segmentation methods in Sec. 4.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 413, + 289, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 289, + 426 + ], + "type": "text", + "content": "3.3 Training and Application" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "content": "Training and Loss Function. We first use the SA-1B (1%) dataset [30] for training the SAM2CLIP module to transfer SAM's knowledge into open-vocabulary SAM, with the loss " + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "inline_equation", + "content": "L_{distill}" + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "content": " (Equ. (2)). Then, we joint train the CLIP2SAM and mask decoder using segmentation mask and label annotations from COCO or LVIS. The final loss function is given as " + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "inline_equation", + "content": "L = \\lambda_{cls}L_{t\\_cls} + \\lambda_{ce}L_{t\\_ce} + \\lambda_{dice}L_{t\\_dice}" + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "inline_equation", + "content": "L_{t\\_ce}" + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "content": " is the Cross-Entropy (CE) loss for mask classification, and " + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\bar{L}_{t\\_ce}" + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "inline_equation", + "content": "L_{t\\_dice}" + }, + { + "bbox": [ + 130, + 435, + 482, + 544 + ], + "type": "text", + "content": " are mask Cross Entropy (CE) loss and Dice loss [48] for segmentation, respectively. In addition, we adopt joint training with the ImageNet dataset for classifying over 22,000 categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "content": "Inference and Demo Tools. Our model performs inference like SAM, with points and boxes as visual prompts. Specifically, we test boxes and points as visual prompts for the encoder in Sec. 4. In the appendix, we show a demo of our model, which can segment and recognize with prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 613, + 230, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 613, + 230, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 613, + 230, + 627 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Datasets and Metrics. We mainly use COCO [43] and LVIS [19] datasets for the experiments. Moreover, we also use part of SA-1B [30] data (1%) for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 480, + 315 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Comparison of combined baselines and Open-Vocabulary SAM using visual prompts. \\*\\* indicates using mask center point as prompts, while others indicate using ground truth boxes prompts. " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "I o U_{b}" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "I o U_{n}" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " refer to the average IoU for each mask of base classes and novel classes, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 480, + 315 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 480, + 315 + ], + "type": "table", + "html": "
MethodCOCO\nIoUbIoUn AccLVIS\nIoUbIoUn AccFLOPs #Param
Image-Crop baseline78.181.446.278.381.69.6
Feature-Crop baseline78.181.455.178.381.626.5
Image-Crop baseline + CoOp [88]79.682.162.080.182.032.1
Feature-Crop baseline + CoOp [88]79.682.170.980.182.048.2
Open-Vocabulary SAM81.584.084.380.483.166.6
Image-Crop baseline*60.766.724.553.062.36.2
Feature-Crop baseline*60.766.732.153.062.311.0
Image-Crop baseline + CoOp [88]*64.766.728.258.964.28.3
Feature-Crop baseline + CoOp [88]*64.766.735.158.964.213.2
Open-Vocabulary SAM*68.465.276.763.667.960.4
", + "image_path": "7a5d7b5de3e38996c1944b760761c477848bd60e4afde51eec7092eb73a5c71d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 339, + 481, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 481, + 471 + ], + "type": "text", + "content": "SAM2CLIP knowledge transfer. For COCO, we report the results of both close-set and open-vocabulary settings for the instance segmentation task. In particular, following Zareian et al. [79], we split 48 base classes with annotations and 17 target classes without annotations. We use the base class annotations for training. For LVIS datasets, we adopt the open-vocabulary setting and report the results of " + }, + { + "bbox": [ + 130, + 339, + 481, + 471 + ], + "type": "inline_equation", + "content": "AP_{\\text{rare}}" + }, + { + "bbox": [ + 130, + 339, + 481, + 471 + ], + "type": "text", + "content": " for novel classes. For evaluation metrics, we report the accuracy of recognition for reference to evaluate the recognition capability. Meanwhile, each prompt's intersection-over-union (IoU) with its ground truth mask is also adopted to evaluate the segmentation ability of our method. As mentioned in Sec. 3.2, different from previous open vocabulary segmentation tasks, boxes or points serve as visual prompts in our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 544 + ], + "type": "text", + "content": "Baselines. As shown in Fig. 2 (a) and (b), based on different adapter designs, we append these adapters to the different locations of the combined models. For example, when using CoOp [88], we append the learnable tokens by combining them with CLIP features. For several convolution-based adapters [8], we add the extra convolution layers along with SAM or CLIP backbone for fair comparison. By default, we adopt SAM-huge and CLIP R50x16." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Implementation Details. We implement our models in PyTorch [50] with MMDetection [7]. We use 8 A100 GPUs for distributed training. Each minibatch has two images per GPU. The optimizer is AdamW [46] with a weight decay of 0.0001. We adopt full image size for a random crop in the pre-training and training process following Cheng et al. [9]. All the class names are transferred into CLIP text embedding, following previous works [18]. We train each model for 12 epochs for fair comparison. Due to the limitation of computation costs, we do not adopt joint SAM data and COCO data training. We first perform training the SAM2CLIP on SA-1B (1%), and then we finetune the model on COCO or LVIS data. Please refer to the supplementary material for more details." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 479, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Comparison of combined baselines and Open-Vocabulary SAM on prompts generated by the open vocabulary detector. For the LVIS dataset, only 'normal' and 'frequent' classes are in the training set. The labels are generated by each baseline or our method. We adopt Detric [89] as the OV-Detector to provide box prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 479, + 228 + ], + "type": "table", + "html": "
MethodCOCOLVISFLOPs#Params
APbaseAPnovelAPAPrareAPnormAPfreqAP
Image-Crop baseline + CoOp [88]26.231.227.319.818.316.317.23,748G808M
Feature-Crop baseline + CoOp [88]28.033.829.524.221.418.620.83,545G808M
Open-Vocabulary SAM31.136.032.424.021.322.922.41,180G304M
", + "image_path": "2cab27e2b8242f30abc6e10876ff8048c6ca35b54f3aa1c8be322e87bd632158.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 194, + 304, + 416, + 376 + ], + "blocks": [ + { + "bbox": [ + 132, + 239, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 482, + 295 + ], + "type": "text", + "content": "Table 3: Comparison of Open-Vocabulary SAM with previous methods on open vocabulary instance segmentation. The results presented in the table all use the Mask-RCNN [22] with ResNet-50 backbone. Different from previous works, Open-Vocabulary SAM uses the bounding box as the prompt to generate the mask. We report mask AP50 following base-novel setting as [24]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 304, + 416, + 376 + ], + "lines": [ + { + "bbox": [ + 194, + 304, + 416, + 376 + ], + "spans": [ + { + "bbox": [ + 194, + 304, + 416, + 376 + ], + "type": "table", + "html": "
MethodVenueConstrainedGeneralized
BaseNovelBaseNovelAll
XPM [24]CVPR'2242.424.041.521.636.3
MaskCLIP [13]ICML'2342.823.242.621.737.2
MasQCLIP [74]ICCV'2340.930.140.728.437.5
Open-Vocabulary SAM(Ours)41.737.539.339.839.4
", + "image_path": "116c88fffdc74fa4b24ac5daeab0a75a1351854ae7a7ac8b62505390c9c0caad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 142, + 430, + 470, + 534 + ], + "blocks": [ + { + "bbox": [ + 132, + 386, + 482, + 420 + ], + "lines": [ + { + "bbox": [ + 132, + 386, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 482, + 420 + ], + "type": "text", + "content": "Table 4: Comparison of mask quality with various detectors on COCO dataset. We report the mask mean AP for comparison. The masks are generated by each method, while the labels are from the corresponding detectors." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 430, + 470, + 534 + ], + "lines": [ + { + "bbox": [ + 142, + 430, + 470, + 534 + ], + "spans": [ + { + "bbox": [ + 142, + 430, + 470, + 534 + ], + "type": "table", + "html": "
MethodDetectorsmAPAP50AP75APSAPMAPL#ParamsFLOPs
SAM-HugeFaster-RCNN (R50)35.654.938.417.239.151.4641M3,001G
SAM-Huge (finetuned)Faster-RCNN (R50)35.855.038.416.538.653.0641M3,001G
Open-Vocabulary SAMFaster-RCNN (R50)35.855.638.316.038.953.1304M1,180G
SAM-HugeDetic (swin-base)36.457.139.421.440.854.6641M3,001G
SAM-Huge (finetuned)Detic (swin-base)36.857.439.820.840.655.1641M3,001G
Open-Vocabulary SAMDetic (swin-base)36.757.239.720.740.854.9304M1,180G
SAM-HugeViTDet (Huge)46.372.049.825.245.559.6641M3,001G
SAM-Huge (finetuned)ViTDet (Huge)46.572.350.325.245.860.1641M3,001G
Open-Vocabulary SAMViTDet (Huge)48.873.852.924.846.364.2304M1,180G
", + "image_path": "bff0a36dea899c88451f3ba151355b5220cca8e3085e2a421a1de441806bb2ea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 482, + 666 + ], + "type": "text", + "content": "Benchmark Setup. Open Vocabulary SAM is different from both SAM [30] and open vocabulary segmentation methods [13,24,74]. Compared with Open-Vocabulary SAM, SAM [30] cannot recognize segmented objects while open vocabulary segmentation methods [13,24,74] usually rely on a separate segmentor (e.g., Mask-RCNN [22]) instead of prompts like points or boxes. For a complete evaluation of Open-Vocabulary SAM, we set the benchmark as three parts, including 1). comparison with combined baselines to verify the effectiveness; 2). comparison with open-vocabulary segmentation methods to evaluate the recognition; 3). comparison with SAM models to evaluate segmentation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 136, + 434, + 220 + ], + "blocks": [ + { + "bbox": [ + 212, + 114, + 400, + 125 + ], + "lines": [ + { + "bbox": [ + 212, + 114, + 400, + 125 + ], + "spans": [ + { + "bbox": [ + 212, + 114, + 400, + 125 + ], + "type": "text", + "content": "Table 5: Scaling up with large-scale datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 136, + 434, + 220 + ], + "lines": [ + { + "bbox": [ + 177, + 136, + 434, + 220 + ], + "spans": [ + { + "bbox": [ + 177, + 136, + 434, + 220 + ], + "type": "table", + "html": "
DatasetsAccuracy#vocabulary#images
LVIS83.11,20399K
V3Det78.713,204183K
I-21k44.519,16713M
V3Det + LVIS82.713,844282K
V3Det + LVIS + I-21k83.325,89813M
V3Det + LVIS + I-21k + Object36583.025,97015M
", + "image_path": "50fd20f70167d55efa7bc7b0a554428da02ceb3cdc434517c3f2a8b0e62a2e51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 177, + 273, + 434, + 327 + ], + "blocks": [ + { + "bbox": [ + 130, + 229, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 229, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 482, + 264 + ], + "type": "text", + "content": "Table 6: The effectiveness of each component. We use Detic [89] as the detector. The labels are generated by the corresponding model. S2C and C2S denote SAM2CLIP and CLIP2SAM respectively. The baseline refers to the image-crop variant." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 273, + 434, + 327 + ], + "lines": [ + { + "bbox": [ + 177, + 273, + 434, + 327 + ], + "spans": [ + { + "bbox": [ + 177, + 273, + 434, + 327 + ], + "type": "table", + "html": "
SettingAPAPbaseAPnovelFLOPs (G)#Params (M)
Baseline + CoOp [88]29.528.033.83,545808
Our + S2C28.727.333.31,127291
Our + S2C + C2S34.433.138.01,180304
", + "image_path": "e63293e85b30ef8f9e11888eca78b86d4f67714449443d4c646213cbd519809d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 347, + 228, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 228, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 228, + 358 + ], + "type": "text", + "content": "4.1 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 545 + ], + "type": "text", + "content": "Comparison with Combined Baselines Using Ground Truth. To avoid the influence of other modules, we first demonstrate the recognition ability of our model in Tab. 1. Compared to the simple combined approaches, adding adapters with joint co-training leads to better results. However, the recognition ability is still limited on both COCO and LVIS. Our Open-Vocabulary SAM achieves the best results on both boxes and points as visual prompts. We observe more significant gains on LVIS datasets. We argue that LVIS contains more small objects, which is more challenging than COCO. Our method can solve Prob.2 and lead to over " + }, + { + "bbox": [ + 130, + 365, + 482, + 545 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 545 + ], + "type": "text", + "content": " accuracy improvement. Although the segmentation quality is pretty good (about 80 IoU on COCO and LVIS with box prompt), our method still achieves " + }, + { + "bbox": [ + 130, + 365, + 482, + 545 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 545 + ], + "type": "text", + "content": " IoU improvements. This indicates the effectiveness of our joint co-training on mask prediction and classification. Compared with boxes as prompts, using points as prompts is more challenging since the location clues of points are much weaker than boxes. However, our approach is still better than combined baselines or them with adapters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": "Comparison with Combined Baselines on OV-Detector. In Tab. 2, we adopt a more challenging setting by using the box prediction from the existing open-vocabulary detector to simulate the interactive segmentation process with deviation. We choose the representative Detic [89] as the open-vocabulary detector. Again, our method also achieves the best performance on both COCO and LVIS datasets. In particular, on COCO, compared with previous works [88], our method achieves 3.0 mask mAP improvements with much lower costs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Comparison with Open-vocabulary Segmentation. In Tab. 3, we compare our method with previous open-vocabulary instance segmentation methods. Open-Vocabulary SAM uses the boxes generated by Mask- RCNN [22] and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 158, + 480, + 212 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 7: Ablation studies on COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (left): Ablation on SAM2CLIP design. (right): Ablation on CLIP2SAM design." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 158, + 480, + 212 + ], + "lines": [ + { + "bbox": [ + 134, + 158, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 134, + 158, + 480, + 212 + ], + "type": "table", + "html": "
SettingIoU FLOPs (G) #Params (M)SettingIoU Acc APbase APnovel
SAM78.73,001641SAM2CLIP (baseline)78.154.227.6
Conv-L + MultiScale Neck78.31,313321+ Cls Token81.379.329.8
Conv-L + SingleScale Neck73.61,280307+ Cls Token & CLIP MLP fusion80.978.929.0
R50x16 + MultiScale Neck78.11,180304+ light FPN (Ours)81.584.331.1
R50 + MultiScale Neck77.3728165
", + "image_path": "9918da2f62c9985e09ff0d4c1cc3e2c16155fd07add3a9dd9b99be80d1a10104.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 265, + 296, + 327 + ], + "blocks": [ + { + "bbox": [ + 130, + 221, + 482, + 256 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 256 + ], + "type": "text", + "content": "Table 8: (left): Ablation study on different CLIP backbone. We test results on the COCO open-vocabulary dataset. We use boxes of the ground truth as a prompt to generate masks and labels. (right): Comparison with other SAM models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 265, + 296, + 327 + ], + "lines": [ + { + "bbox": [ + 135, + 265, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 135, + 265, + 296, + 327 + ], + "type": "table", + "html": "
BackboneIoU Acc #FLOPs(G) #Params (M)
RN5077.3 50.8728165
RN50x1678.1 55.11,180304
RN50x6478.1 54.12,098568
ConvNeXt-L78.3 59.11,313321
ViT-L-1438.6 14.32,294441
", + "image_path": "c24eb787b2849193b6da45e3f018f3397a26050fc3519360d8a04dc8bce67beb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 268, + 479, + 324 + ], + "blocks": [ + { + "bbox": [ + 318, + 268, + 479, + 324 + ], + "lines": [ + { + "bbox": [ + 318, + 268, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 318, + 268, + 479, + 324 + ], + "type": "table", + "html": "
Method1-IoU (COCO)cls. open.
SAM [30] (H)78.2--
SEEM [90] (T)73.7✓-
Semantic-SAM [32] (T)76.1✓-
OV-SAM (ours)81.7✓✓
", + "image_path": "d039bd88ff2ea66dfe77b2dd7532086769393aa82695eca1ce59ecd40a853c29.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 350, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 374 + ], + "type": "text", + "content": "previous works use the masks of Mask-RCNN. The results show that our Open-Vocabulary SAM performs strongly, especially in the novel classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "content": "Comparison with SAM on Various Detectors. In Tab. 4, we test the mask prediction quality of our model and original SAM on different detectors. Our method performs better than the original SAM and performs comparably with fine-tuned SAM. It is worth noting that our Open-Vocabulary SAM has much lower computational costs and parameters than SAM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 434, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 506 + ], + "type": "text", + "content": "Comparison on Interactive Segmentation. In Tab. 8 (right), we compare interactive segmentation performance. Notably, our approach excels beyond interactive segmentation and can recognize classes in an open-vocabulary setting. Visualization Comparison. In Fig. 4, we compare our approach with the feature-crop baseline. Our model shows better performance in classifying small and rare object classification and handling occlusion scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 506, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 592 + ], + "type": "text", + "content": "Model as a Zero Shot Annotation Tool. In addition to COCO and LVIS standard datasets training, following the spirit of SAM, we also scale up our model by training it with more data (Tab. 5). In particular, we adopt more detection data (V3Det [60], Object365 [56]) and classification data (ImageNet22k [11]). Owing to significant costs, we have not conducted comparisons with other baselines for this setting. Rather, we have adapted our method into an interactive annotation tool capable of segmenting and recognizing over 22,000 classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 609, + 313, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 313, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 313, + 621 + ], + "type": "text", + "content": "4.2 Ablation Studies and Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Effectiveness of SAM2CLIP and CLIP2SAM. We first verify the effectiveness of our proposed two modules in Tab. 6. We adopt image-crop variant of the baseline for comparison. In particular, by sharing a single backbone, we observe" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 118, + 438, + 311 + ], + "blocks": [ + { + "bbox": [ + 177, + 118, + 438, + 311 + ], + "lines": [ + { + "bbox": [ + 177, + 118, + 438, + 311 + ], + "spans": [ + { + "bbox": [ + 177, + 118, + 438, + 311 + ], + "type": "image", + "image_path": "d6313c36a159f09236bb855814d903f7f107015db0866a095e03c2955d7f718f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 319, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 131, + 319, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 482, + 353 + ], + "type": "text", + "content": "Fig. 4: Visualization Comparison. We compare the mask and classification results of the image-crop baseline (SAM + CLIP) and Open-Vocabulary SAM (OV SAM). The predicted labels are presented on the mask." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 454 + ], + "type": "text", + "content": "a significant drop in the number of parameters and FLOPs, with a little drop in segmentation performance. The slight drop is caused by the domain gap between SAM data and COCO data during the training of the SAM2CLIP module. However, after adding our CLIP2SAM module and joint co-training with mask classification and prediction, a significant improvement in both segmentation and classification is observed, with just a negligible increase in compute cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 456, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 516 + ], + "type": "text", + "content": "Detailed Design on SAM2CLIP. In Tab. 7, we explore the detailed design of SAM2CLIP in the first stage of open-vocabulary SAM training. The results show that distillation benefits most when multi-scale features are adopted, suggesting that both high-resolution features and high-level semantics are important to align CLIP's feature with the SAM's feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 519, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 603 + ], + "type": "text", + "content": "Detailed Design on CLIP2SAM. In Tab. 7, we present extensive design for the CLIP2SAM module. We compare two designs: a simple classification token with cross attention (Cls Token) and a combination of this token with mask pooled CLIP feature (CLS Token & CLIP MLP fusion). These designs work better than the combined baseline shown in the first row. Nonetheless, due to resolution constraints, these variants cannot handle small objects well, as shown in Fig. 4. In contrast, our design improves the performance considerably." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Ablation on Different CLIP Backbones. In Tab. 8, we explore the effect of frozen CLIP visual backbone. We do not add the CLIP2SAM. Motivated by recent works [31,67,73,76], CNN-based CLIPs encapsulate more structural information, which is good for our goal since we have location-sensitive visual prompts. Thus, adopt CNN-based CLIPs, aligned with previous works. As shown" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 217, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 217, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 217, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 217, + 178 + ], + "type": "image", + "image_path": "86073dd589b87850741a8711cbd5818db5351f08272fe52c5b27f7081571749e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 114, + 310, + 178 + ], + "blocks": [ + { + "bbox": [ + 217, + 114, + 310, + 178 + ], + "lines": [ + { + "bbox": [ + 217, + 114, + 310, + 178 + ], + "spans": [ + { + "bbox": [ + 217, + 114, + 310, + 178 + ], + "type": "image", + "image_path": "40c3b8fc36b2611ec703be7c0697015b1721136679dd41c15a7d98e9fcc94efa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 395, + 178 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 395, + 178 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 395, + 178 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 395, + 178 + ], + "type": "image", + "image_path": "272da928ca6b914a200c89de556f6229024baf88ca0c863ac0b1db449c613a3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 396, + 114, + 436, + 178 + ], + "blocks": [ + { + "bbox": [ + 396, + 114, + 436, + 178 + ], + "lines": [ + { + "bbox": [ + 396, + 114, + 436, + 178 + ], + "spans": [ + { + "bbox": [ + 396, + 114, + 436, + 178 + ], + "type": "image", + "image_path": "f4334d26a305ab5ff9ca5c37b2ab0548afcfa594cd5b7792c7c4d9c54f41f92a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 437, + 114, + 480, + 178 + ], + "blocks": [ + { + "bbox": [ + 437, + 114, + 480, + 178 + ], + "lines": [ + { + "bbox": [ + 437, + 114, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 437, + 114, + 480, + 178 + ], + "type": "image", + "image_path": "18b84b886e35bdd4419970d61e0c59756f35d2f22a6e91395d9b94e00972133c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 183, + 217, + 244 + ], + "blocks": [ + { + "bbox": [ + 133, + 183, + 217, + 244 + ], + "lines": [ + { + "bbox": [ + 133, + 183, + 217, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 217, + 244 + ], + "type": "image", + "image_path": "8ee2ece8e7a8041242d99d76926ea13c7d1fafa9c227801ce597acf9e3aeffb8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 217, + 186, + 311, + 244 + ], + "blocks": [ + { + "bbox": [ + 217, + 186, + 311, + 244 + ], + "lines": [ + { + "bbox": [ + 217, + 186, + 311, + 244 + ], + "spans": [ + { + "bbox": [ + 217, + 186, + 311, + 244 + ], + "type": "image", + "image_path": "35b3153c1e006fbca89453d14eb1201b03473d9a0da96ab4516a83d2ea97b9ae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 186, + 395, + 244 + ], + "blocks": [ + { + "bbox": [ + 312, + 186, + 395, + 244 + ], + "lines": [ + { + "bbox": [ + 312, + 186, + 395, + 244 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 395, + 244 + ], + "type": "image", + "image_path": "132df602141ec273e4a06ea2eb3288e0daa0cbfba59a2ca3690716b1b1d0e8da.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 396, + 186, + 436, + 244 + ], + "blocks": [ + { + "bbox": [ + 396, + 186, + 436, + 244 + ], + "lines": [ + { + "bbox": [ + 396, + 186, + 436, + 244 + ], + "spans": [ + { + "bbox": [ + 396, + 186, + 436, + 244 + ], + "type": "image", + "image_path": "3c37d9c18656873ff7c01e4a9314c867cd3c8b8ae20e96d3e68e03961cb8a0f7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 437, + 186, + 480, + 244 + ], + "blocks": [ + { + "bbox": [ + 437, + 186, + 480, + 244 + ], + "lines": [ + { + "bbox": [ + 437, + 186, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 437, + 186, + 480, + 244 + ], + "type": "image", + "image_path": "602e18d6d337cb3b4dc3003acacca4f334ab6fddbadf5caa9e835d8380375a26.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 133, + 254, + 246, + 353 + ], + "blocks": [ + { + "bbox": [ + 133, + 254, + 246, + 353 + ], + "lines": [ + { + "bbox": [ + 133, + 254, + 246, + 353 + ], + "spans": [ + { + "bbox": [ + 133, + 254, + 246, + 353 + ], + "type": "image", + "image_path": "97e4442e4ea0a5ba3e9aa030b6741888660ab91e74f5bebeeefa6337ef0b18e8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 361, + 482, + 417 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 417 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of Open-Vocabulary SAM. In the visualization, boxes refer to box prompts (up), and the red stars refer to the point prompts (middle). The masks can be merged into a segmentation map (bottom). We show the mask and labels generated by the proposed Open-Vocabulary SAM. Our method can segment and recognize open vocabulary objects in diverse scenes identified by prompts." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 248, + 254, + 353, + 353 + ], + "blocks": [ + { + "bbox": [ + 248, + 254, + 353, + 353 + ], + "lines": [ + { + "bbox": [ + 248, + 254, + 353, + 353 + ], + "spans": [ + { + "bbox": [ + 248, + 254, + 353, + 353 + ], + "type": "image", + "image_path": "6d8e6627bdbf8bfe7ab8d9af8261374ff569ec69d675c86b541c6fa985134cc1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 353, + 255, + 480, + 353 + ], + "blocks": [ + { + "bbox": [ + 353, + 255, + 480, + 353 + ], + "lines": [ + { + "bbox": [ + 353, + 255, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 353, + 255, + 480, + 353 + ], + "type": "image", + "image_path": "4ceaf8098f57f5b9958e3bc7c0b047e8531aa48981fd85898a20d212e91a5cf7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "text", + "content": "in the table, we find ConvNext large achieves the best performance, but we choose RN50x16 since it has comparable performance and better efficiency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 469, + 220, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 220, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 220, + 482 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "We present Open Vocabulary SAM, a SAM-inspired method for interactive segmentation and recognition. Unlike previous open-vocabulary detection and segmentation methods, our method explores interactive open-vocabulary segmentation for the first time. Given the user's inputs, such as boxes or points, the proposed approach can interactively segment and label each visual prompt. Compared with the combined baselines and various visual adapters, our proposed CLIP2SAM and SAM2CLIP are both efficient and effective in various settings. Our open vocabulary segmentation is compatible with different detectors, including open-vocabulary detectors and close-set detectors. With more data, our model plays a similar role as SAM, offering an effective annotation tool for both segmentation and instance labeling. In particular, our method can perform large vocabulary segmentation and recognition over 22K classes. We hope our Open-Vocabulary SAM can provide a solid baseline for combining the strengths of different forms of vision foundation models and inspire further research." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). The project is also supported by Singapore MOE AcRF Tier 1 (RG16/21) and the National Key R&D Program of China (No. 2022ZD0161600)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 195, + 197, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 195, + 197, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 195, + 197, + 208 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 222, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "text", + "content": "1. Balăzevic, I., Steiner, D., Parthasarathy, N., Arandjelović, R., Henaff, O.J.: Towards in-context scene understanding. In: NeurIPS (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "2. Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image transformers. In: ICLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 267, + 480, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 288 + ], + "type": "text", + "content": "3. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. In: NeurIPS (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 322 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "text", + "content": "5. Chen, C., Miao, J., Wu, D., Yan, Z., Kim, S., Hu, J., Zhong, A., Liu, Z., Sun, L., Li, X., Liu, T., Heng, P.A., Li, Q.: MA-SAM: Modality-agnostic SAM adaptation for 3d medical image segmentation. arXiv preprint arXiv:2309.08842 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 356, + 414, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 414, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 414, + 366 + ], + "type": "text", + "content": "6. Chen, J., Yang, Z., Zhang, L.: Semantic segment anything (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "type": "text", + "content": "7. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "type": "text", + "content": "8. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. In: ICLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 422, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 443 + ], + "type": "text", + "content": "9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 466 + ], + "type": "text", + "content": "0. Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 467, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 487 + ], + "type": "text", + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "text", + "content": "2. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "text", + "content": "3. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with MaskCLIP. In: ICML (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 533, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 576 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "text", + "content": "5. Fang, Y., Yang, S., Wang, S., Ge, Y., Shan, Y., Wang, X.: Unleashing vanilla vision transformer with masked image modeling for object detection. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "6. Fang, Z., Li, X., Li, X., Buhmann, J.M., Loy, C.C., Liu, M.: Explore in-context learning for 3d point cloud understanding. In: NeurIPS (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "7. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: CLIP-Adapter: Better vision-language models with feature adapters. IJCV (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "8. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "20. Han, X., Wei, L., Yu, X., Dou, Z., He, X., Wang, K., Han, Z., Tian, Q.: Boosting segment anything model towards open-vocabulary learning. arXiv preprint arXiv:2312.03628 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "21. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 465, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 465, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 465, + 203 + ], + "type": "text", + "content": "22. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "23. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "24. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 266 + ], + "type": "text", + "content": "25. Jayaraman, D., Grauman, K.: Zero-shot recognition with unreliable attributes. In: NeurIPS (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 266, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 480, + 299 + ], + "type": "text", + "content": "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "text", + "content": "27. Kim, D., Lin, T.Y., Angelova, A., Kweon, I.S., Kuo, W.: Learning open-world object proposals without learning to classify. RA-L (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "28. Kim, W., Son, B., Kim, I.: ViLT: Vision-and-language transformer without convolution or region supervision. In: ICML (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 342, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 364 + ], + "type": "text", + "content": "29. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "text", + "content": "31. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A.J., Angelova, A.: F-VLM: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "32. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-SAM: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "type": "text", + "content": "33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "35. Li, J., Selvaraju, R.R., Gotmare, A., Joty, S.R., Xiong, C., Hoi, S.C.: Align before fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "36. Li, S., Cao, J., Ye, P., Ding, Y., Tu, C., Chen, T.: ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "37. Li, X., Ding, H., Zhang, W., Yuan, H., Cheng, G., Jiangmiao, P., Chen, K., Liu, Z., Loy, C.C.: Transformer-based visual segmentation: A survey. arXiv pre-print (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "38. Li, X., You, A., Zhu, Z., Zhao, H., Yang, M., Yang, K., Tong, Y.: Semantic flow for fast and accurate scene parsing. In: ECCV (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "39. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: OMG-Seg: Is one model good enough for all segmentation? In: CVPR (2024)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "text", + "content": "40. Li, X., Zhang, J., Yang, Y., Cheng, G., Yang, K., Tong, Y., Tao, D.: Sfnet: Faster and accurate domain agnostic semantic segmentation via semantic flow. IJCV (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "41. Li, Y., Fan, H., Hu, R., Feichtenhofer, C., He, K.: Scaling language-image pretraining via masking. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 159 + ], + "type": "text", + "content": "42. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "43. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 213 + ], + "type": "text", + "content": "44. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 481, + 236 + ], + "type": "text", + "content": "45. Liu, Y., Qi, L., Tsai, Y.J., Li, X., Chan, K.C.K., Yang, M.H.: Effective adapter for face recognition in the wild. arXiv preprint (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 236, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 246 + ], + "type": "text", + "content": "46. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 247, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 267 + ], + "type": "text", + "content": "47. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 267, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 481, + 289 + ], + "type": "text", + "content": "48. Miletari, F., Navab, N., Ahmadi, S.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. In: 3DV (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 289, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 310 + ], + "type": "text", + "content": "49. Pan, T., Tang, L., Wang, X., Shan, S.: Tokenize anything via prompting. arXiv preprint arXiv:2312.09128 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 481, + 342 + ], + "type": "text", + "content": "50. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 364 + ], + "type": "text", + "content": "51. Qi, L., Kuen, J., Shen, T., Gu, J., Guo, W., Jia, J., Lin, Z., Yang, M.H.: High-quality entity segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 364, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 481, + 385 + ], + "type": "text", + "content": "52. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE TPAMI (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 418 + ], + "type": "text", + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 439 + ], + "type": "text", + "content": "54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 440, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 481, + 460 + ], + "type": "text", + "content": "55. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv:2112.08633 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 461, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 483 + ], + "type": "text", + "content": "56. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: ICCV (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 483, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 503 + ], + "type": "text", + "content": "57. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 503, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 481, + 525 + ], + "type": "text", + "content": "58. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: EVA-CLIP: Improved training techniques for CLIP at scale. arXiv preprint arXiv:2303.15389 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 481, + 568 + ], + "type": "text", + "content": "59. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: SAM-CLIP: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 568, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 481, + 590 + ], + "type": "text", + "content": "60. Wang, J., Zhang, P., Chu, T., Cao, Y., Zhou, Y., Wu, T., Wang, B., He, C., Lin, D.: V3det: Vast vocabulary visual detection dataset. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 481, + 611 + ], + "type": "text", + "content": "61. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 632 + ], + "type": "text", + "content": "62. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. In: ICCV (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "63. Wu, J., Li, X., Ding, H., Li, X., Cheng, G., Tong, Y., Loy, C.C.: Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "64. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., Ghanem, B., Tao, D.: Towards open vocabulary learning: A survey. IEEE TPAMI (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "65. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical SAM Adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "66. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "67. Wu, S., Zhang, W., Xu, L., Jin, S., Li, X., Liu, W., Loy, C.C.: CLIPSelf: Vision transformer distills itself for open-vocabulary dense prediction. In: ICLR (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "68. Wu, X., Zhu, F., Zhao, R., Li, H.: CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "text", + "content": "69. Xie, J., Li, W., Li, X., Liu, Z., Ong, Y.S., Loy, C.C.: Mosaic fusion: Diffusion models as data augmenters for large vocabulary instance segmentation. arXiv preprint arXiv:2309.13042 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "70. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "text", + "content": "71. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "type": "text", + "content": "72. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "text", + "content": "73. Xu, S., Li, X., Wu, S., Zhang, W., Li, Y., Cheng, G., Tong, Y., Chen, K., Loy, C.C.: Dst-det: Simple dynamic self-training for open-vocabulary object detection. arXiv pre-print (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "74. Xu, X., Xiong, T., Ding, Z., Tu, Z.: Masqclip for open-vocabulary universal image segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "75. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., et al.: Panoptic video scene graph generation. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "type": "text", + "content": "76. Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional CLIP. In: NeurIPS (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 483 + ], + "type": "text", + "content": "77. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "78. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "79. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "text", + "content": "80. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: LiT: Zero-shot transfer with locked-image text tuning. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "81. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight SAM for mobile applications. arXiv preprint arXiv:2306.14289 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "82. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "83. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "85. Zhou, C., Li, X., Loy, C.C., Dai, B.: EdgeSAM: Prompt-in-the-loop distillation for on-device deployment of SAM. arXiv preprint arXiv:2312.06660 (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 226 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 127 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 127 + ], + "type": "text", + "content": "86. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from CLIP. In: ECCV (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 128, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 128, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 481, + 160 + ], + "type": "text", + "content": "87. Zhou, H., Shen, T., Yang, X., Huang, H., Li, X., Qi, L., Yang, M.H.: Rethinking evaluation metrics of open-vocabulary segmentation. arXiv preprint arXiv:2311.03352 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "text", + "content": "88. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "89. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: NeurIPS (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-Vocabulary SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_content_list.json b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a572a5a181386c3201b9a7b421ddbac54c30b8a3 --- /dev/null +++ b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_content_list.json @@ -0,0 +1,1683 @@ +[ + { + "type": "text", + "text": "Open-World Dynamic Prompt and Continual Visual Representation Learning", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngeun Kim $^{1,*†}$ , Jun Fang $^{2,*‡}$ , Qin Zhang $^{2}$ , Zhaowei Cai $^{3}$ , Yantao Shen $^{2}$ , Rahul Duggal $^{2}$ , Dripta S. Raychaudhuri $^{2}$ , Zhuowen Tu $^{2}$ , Yifan Xing $^{2}$ , Onkar Dabeer $^{2}$", + "bbox": [ + 264, + 210, + 736, + 258 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Yale University", + "AWS AI Labs", + "3 Amazon AGI" + ], + "bbox": [ + 442, + 268, + 560, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "youngun.kim@yale.edu {junfa, qzaamz, zhaoweic, yantaos, dugrahul, driptarc, ztu, yifax, onkardab}@amazon.com", + "bbox": [ + 240, + 311, + 756, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The open world is inherently dynamic, characterized by ever-evolving concepts and distributions. Continual learning (CL) in this dynamic open-world environment presents a significant challenge in effectively generalizing to unseen test-time classes. To address this challenge, we introduce a new practical CL setting tailored for open-world visual representation learning. In this setting, subsequent data streams systematically introduce novel classes that are disjoint from those seen in previous training phases, while also remaining distinct from the unseen test classes. In response, we present Dynamic Prompt and Representation Learner (DPaRL), a simple yet effective Prompt-based CL (PCL) method. Our DPaRL learns to generate dynamic prompts for inference, as opposed to relying on a static prompt pool in previous PCL methods. In addition, DPaRL jointly learns dynamic prompt generation and discriminative representation at each training stage whereas prior PCL methods only refine the prompt learning throughout the process. Our experimental results demonstrate the superiority of our approach, surpassing state-of-the-art methods on well-established open-world image retrieval benchmarks by an average of $4.7\\%$ improvement in Recall@1 performance.", + "bbox": [ + 259, + 377, + 751, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Dynamic Prompt Generation $\\cdot$ Continual Learning $\\cdot$ Open-World Visual Representation Learning", + "bbox": [ + 259, + 638, + 740, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 694, + 375, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continual learning (CL) without catastrophic forgetting is a challenging task in practice as retraining with old data is prohibitively expensive when data volume grows [10,14,23]. It is particularly challenging for real-world applications in the open-world [2,5], where new concepts keep emerging and distribution shifts are", + "bbox": [ + 212, + 726, + 787, + 787 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 217, + 796, + 364, + 810 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work conducted during an internship at Amazon.", + "bbox": [ + 220, + 811, + 562, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "‡ Corresponding author.", + "bbox": [ + 222, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4956c950113a6c80b7d90ac9556c93ec61c1c08e9318344612d80e0fab2e977.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 233, + 148, + 352, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe91abc385f9ade7dd745fc2cc5a86198522244945daaa10c6dc39d75cb5caa7.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 357, + 148, + 475, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3f30019b810bc0377ab9000dc55444fadbc6d010816e1508fdb87ac41a36b015.jpg", + "image_caption": [ + "(c)", + "Fig. 1: Illustration of the distinctions between our problem setting and traditional settings. In our work, we aim to address the problem where training splits have no class overlaps. (a) Closed-world setting: both training and testing classes are identical. (b) Continual learning (CL) in closed-world setting: training classes are split into multiple divisions and are introduced through various CL stages. (c) Open-world setting: training and testing classes remain separate. The aim is to learn a robust representation from training classes that generalize to unseen classes. (d) Open-World Continual Representation Learning (Our Problem Setting): continual learning tailored for an open-world scenario. We sequentially introduce training classes over multiple CL stages, ensuring they remain distinct from test-time unseen classes." + ], + "image_footnote": [], + "bbox": [ + 482, + 148, + 612, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/92d55fa6a85a963a9e0bd211454df4ae89c774157d98a3df799e3e34df60b55a.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 627, + 148, + 767, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inevitable. Recent studies on Prompt-based Continual Learning (PCL) [32,38, 39] demonstrated success in mitigating catastrophic forgetting without access to any past data. However, they heavily focus on the closed-world setting where test-time categories come from those encountered in various training stages.", + "bbox": [ + 212, + 443, + 787, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Whereas in practice, the world is dynamic and concepts evolve over time and new categories are often introduced after a model is trained. Consider, for instance, an image retrieval system for clothing brands in online e-commerce. During training, one would gather images of all existing clothing brands available to train a model. However, after deployment, the system encounters new brands. The new data collected for system upgrade, constituting a new training stage, will likely include brands that are disjoint to those in older stages. Furthermore, the system will constantly face new brands at test-time that are disjoint to any of those encountered in the previous training stages and is expected to generalize well to those new categories.", + "bbox": [ + 212, + 506, + 787, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, to comprehensively evaluate the ability of a model in retaining semantic information from continuous streams of data for generalization to new concepts at test-time for practical applications, we propose a new continual open-world representation learning setting, as illustrated in Figure 1(d). In this setting, different data streams comprise distinct classes, and test-time classes are disjoint to those encountered in all training phases. Our setting is notably more challenging than the previous closed-world CL scenario. It closely mirrors real-world demands on recognition systems, where, during any evaluation stage, the test set consists of unseen entities to the training process up to that point. This setup reflects the need for practical recognition systems to encode general semantic information from training-time classes and apply this knowledge to new, unseen concepts without access to older training data during system updates.", + "bbox": [ + 212, + 657, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start with evaluating state-of-the-art PCL methods [32, 38, 39] in our proposed setting and find that they fail to generalize with the absence of past data. For example, in our experiments, the gap between the paragon of standard training with all data in one phase and the best PCL baseline (Coda) shows up to $20\\%$ gap in Recall@1 (Table 1), significantly larger than the $5\\%$ gap in the closed-world setting, which suggests that prior PCL methods encounter difficulties and are less effective in the open-world setting.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle this challenging yet practical open-world CL setting, we introduce a simple yet effective method called Dynamic Prompt and Representation Learner (DPaRL). Unlike previous prompt pool based PCL methods [32, 38, 39] that depend on a static prompt pool (Figure 2(a)) for test-time predictions. Notably, DPaRL trains an innovative dynamic prompt generation network jointly with the discriminative representation backbone model (Figure 2(b)), departing from the prompt-only learning paradigm in prior PCL methods.", + "bbox": [ + 212, + 253, + 784, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key insight lies in that simply combining the fixed prompts learnt over training classes does not generalize well to unseen test-time classes in the practical open-world setting. On the other hand, the on-the-fly dynamically generated prompts in our method exhibit a stronger capability of capturing the diverse semantics spread across the numerous divisions of classes observed during the different training stages of continual learning, improving generalization to test-time unseen classes. This is achieved through the proposed dynamic prompt generation network, which possesses stronger representation power with a learnable mapping function compared to the traditional approach of combining multiple prompts in a static prompt pool. Additionally, unlike prior PCL methods, DPaRL jointly updates the dynamic prompt generation process and the discriminative representation backbone weights, facilitating maximized interaction and comparison between old concepts and newer ones to encapsulate diverse semantics that aid generalization to unseen test-time classes.", + "bbox": [ + 212, + 361, + 784, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consequently, over the four open-world image retrieval benchmarks on cars, online catalogs, products, and species images where the task is to search for the most similar images belonging to the same class as the query, via a distance measurement in an embedding space, DPaRL achieves significant accuracy boosts. It improves the average Recall@1 performance by absolute $4.7\\%$ and $9.1\\%$ compared to state-of-the-art continual learning methods, from $76.1\\%$ to $80.8\\%$ over 10 CL stages and from $68.0\\%$ to $77.1\\%$ over 100 CL stages, respectively.", + "bbox": [ + 212, + 575, + 782, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are three folds:", + "bbox": [ + 238, + 683, + 573, + 696 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We establish a new practical setting on continual visual representation learning in the open-world.", + "2. We introduce a simple yet powerful method, Dynamic Prompt and Representation Learner (DPaRL), which dynamically generates prompts while effectively updating the discriminative representation backbone. This enhancement improves generalization for unseen open-world classes at test time.", + "3. We outperform state-of-the-art continual learning methods in the proposed practical setting, both rehearsal-free and rehearsal-based." + ], + "bbox": [ + 222, + 715, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/102be1890616f633366a02d43040179e7334815d274287997908dc1476d130bb.jpg", + "image_caption": [ + "(a) Previous method" + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 493, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/740775948f3f6a33e91d0b5fd35cc9ddb232c9a1a8d5101dfb2e98a836469bb1.jpg", + "image_caption": [ + "(b) Our work", + "Fig. 2: Prior PCL methods combine prompts from a static prompt pool trained on the training class distribution, leading to a loss of generalization capability when facing unseen classes during test time. Our work introduces a Dynamic Prompt Generation network that generates dynamic prompts on the fly by integrating a given image with stage tokens, followed by a specialized mapping function and adjustable discriminative representation backbone weights, providing generalizable prompts for unseen testing classes, distinguishing it from prior PCL methods." + ], + "image_footnote": [], + "bbox": [ + 511, + 147, + 771, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 402, + 387, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continual learning. Continual learning (CL) has been recognized for decades as a crucial research domain, focusing on the ongoing adaptation of models over time. Early works primarily centered around regularization-based methods in an effort to address catastrophic forgetting [1,21,43] by introducing regularization between its current and previously learned parameters. However, they often underperform when faced with complex datasets.", + "bbox": [ + 212, + 431, + 787, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A separate line of work, architecture-based approach, involves expanding the network architecture in response to new learning phases [20,26,29,30,42]. While these methods generally outperform regularization-based counterparts, they do so at the cost of introducing additional parameters. Recent advancements include rehearsal-based methods that incorporate a memory component for storing past data [4,6,7,15,28]. This archived data aids in training newer stages, typically yielding superior performance compared to other CL techniques. Yet, this data retention can introduce privacy concerns and significant memory overhead.", + "bbox": [ + 212, + 523, + 787, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given these constraints, rehearsal-free strategies are gaining attention in the community. These aim to mitigate catastrophic forgetting without relying on historical data. A prominent method within this category uses model inversion to produce rehearsal images [9,13,31,41]. However, model inversion is computationally intensive and time-consuming. Additionally, these methods lag significantly in performance when compared to their rehearsal-based counterparts.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent developments have witnessed the rise of prompt-based approaches [27, 32, 34, 37-39, 45]. These strategies offer robust protection against catastrophic forgetting by learning a small set of parameters, or prompts, instead of directly training all parameters. For instance, L2P [39] employs a pool of prompts, selecting suitable ones for insertion based on input data clustering. Building on this concept, DualPrompt [38] presents general prompts designed to encapsulate shared knowledge across various learning stages. Advancing this further,", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the state-of-the-art method CodaPrompt [32] facilitates end-to-end training by incorporating an attention mechanism for prompt combination within a pool. However, during inference, the above works rely on the combination of prompts from a static pool learned from training to feed into a frozen backbone model, which has limitations in cases where testing data has distribution shifts or disjoint classes from the training data. This motivates us to design a Dynamic Prompt and Representation Learner (DPaRL) to address these drawbacks.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Open-World Visual Representation Learning. In open-world visual representation learning, the model learns discriminative representations that align distances between representations with their semantic similarities, and uses these knowledge for testing on unseen classes. Previous works in this domain have primarily focused on designing loss functions [11, 17, 24] for enhanced accuracy. Recently, several studies [2, 12] have demonstrated the effectiveness of utilizing pre-trained visual foundation models and fine-tuning them to achieve significant performance improvements. While open-world visual representation learning has been under investigation for an extended period, there has been limited exploration within the continual learning (CL) framework for this task. [44] investigated biometric datasets using a regularization-based CL approach. However, the investigation into state-of-the-art prompt-based CL methods remains missing. Furthermore, as their studies primarily centered on biometric data, there is a pressing need to explore established continual learning benchmarks specific to open-world visual representation learning in more natural settings.", + "bbox": [ + 212, + 273, + 787, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 522, + 380, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our open-world Continual Learning (CL) setting and motivate our Dynamic Prompt and Representation Learner (DPaRL).", + "bbox": [ + 212, + 553, + 787, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Setting", + "text_level": 1, + "bbox": [ + 215, + 604, + 398, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two fundamental visual representation learning settings are the closed-world and open-world paradigms. In a closed-world setting (as depicted in Figure 1(a)), the training and testing data classes are entirely identical. Conversely, the open-world setting (illustrated in Figure 1(c)) presents a more practical scenario. Here, training and testing classes are entirely distinct, thus requiring the model to learn representations that generalize to unseen concepts.", + "bbox": [ + 212, + 628, + 787, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When faced with continuous streams of data for visual recognition system updates, built upon the closed-world paradigm, Figure 1(b) showcases the well-known continual learning setting for image recognition. In this work, we introduce a new practical setting: Open-World Continual Representation Learning as shown in Figure 1(d). Instead of testing on seen classes of the various training stages thus far, the testing classes are entirely disjoint to those faced in different training stages, necessitating the model to not only learn continually but also generalize to new concepts, making it a more dynamic and challenging setting.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9bfe3eb4e63ee6f78904d8649b61ea55486f17d7086fdc13b3687209d646347c.jpg", + "image_caption": [ + "Fig. 3: Illustration of our proposed method: Dynamic Prompt and Representation Leaner (DPaRL). We dynamically generate prompt tokens on the fly from Dynamic Prompt Generation (DPG) networks by integrating information from stage tokens and image tokens. The [CLS] token from DPG is converted to prompt tokens via a linear mapping function. The generated prompt tokens are added to the backbone ViT and trained with a loss function. The learnable parameters include the current stage token, weights in mapping function, weights in backbone, and weights in loss function." + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 754, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Prompt-based Continual Learning Paradigm", + "text_level": 1, + "bbox": [ + 215, + 412, + 627, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt-based Continual Learning (PCL) methods [32,38,39] use a pre-trained Vision Transformer (ViT) as a discriminative backbone for closed-world image classification, as shown in the Figure 2(a). These methods create a prompt pool with multiple prompt tokens, updating only the learnable parameters in this pool during training. In inference, the learned prompt pool is static, and PCL methods select tokens from this pool to feed into multiple ViT backbone layers for predictions. Our method adopts this PCL paradigm but adapts to the open-world representation learning setting with a prototypical metric learning loss. Additionally, we introduce a dynamic prompt generation (DPG) network to replace the static prompt pool for effective joint learning with the discriminative representation backbone model, as detailed below.", + "bbox": [ + 212, + 435, + 787, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Dynamic Prompt and Representation Leaner (DPaRL)", + "text_level": 1, + "bbox": [ + 215, + 621, + 714, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Pipeline. In the open-world setting where training and testing classes are disjoint, existing prompt pool designs in PCL [32,38,39] exhibit limitations, i.e., limited separation between intra- and inter- testing class distance distributions as shown in Figure 4 and large performance gap from paragon in Table 1. To mitigate, we first propose dynamically generating prompts by blending the given image information and continual learning stage's information through a dynamic prompt generation (DPG) network. It aims to provide better representation power compared to the prompt pool design, which combines a fixed prompt pool trained from the training class distribution. Moreover, we enhance the capability of the discriminative representation backbone model to more effectively generalize to the open-world concepts through a joint dynamic prompt and representation learning paradigm. We refer to Figure 2, 3 for detailed design differences between our approach and prior prompt pool-based methods.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Prompt Generation Network. In order to retain information from each CL stage of our Dynamic Prompt Generation (DPG) process, we introduce stage tokens, denoted as $S$ . During training stage $t$ , we train a stage token $S_{t}$ while freezing the previous stage tokens $S_{t - (q - 1)} \\sim S_{t - 1}$ in a first-in-first-out (FIFO) queue manner, where $q$ is the max queue size. Guided by the ablation study detailed in Section 4.4, we typically set $q = 5$ using a FIFO order to aim for both high accuracy and scalability. This ensures that the knowledge from the previous stages remains unchanged, and the total number of stage tokens limits by the queue size without linearly scaling with stage numbers. Within the deep neural network such as ViT, the self-attention block facilitates the integration of information between the stage-wise tokens and the instance-wise image tokens.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, the DPG produces a [CLS] token to dynamically acquire task-specific high-level information for prior knowledge of class-level features. However, it differs in size from the prompt tokens required by the discriminative backbone model (right side of Figure 3). To address this, we introduce a mapping function between the [CLS] token and the prompt tokens.", + "bbox": [ + 212, + 311, + 787, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A straightforward approach is to employ a single linear layer for dimensionality transformation. However, such a method introduces an excessive number of additional parameters, and the resulting over-parameterization can lead to overfitting of the highly compressed [CLS] token information. To tackle this issue, we impose a constraint on the weight parameter $W = AB^{T} \\in \\mathbb{R}^{C_{in} \\times C_{out}}$ , where $A \\in \\mathbb{R}^{C_{in} \\times R}$ , $B \\in \\mathbb{R}^{C_{out} \\times R}$ , ensuring a maximum rank of $R < \\min(C_{in}, C_{out})$ . This constraint is inspired by the success of the LoRA technique [16]. The experimental results in Section 4.4 illustrate the critical importance of the low-rank mapping function design for our DPG module. Moreover, we apply Dropout [33] and LayerNorm [3] to this low-rank linear mapping function", + "bbox": [ + 212, + 388, + 787, + 541 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM a p p i n g (x) = L a y e r N o r m \\left(L o w R a n k L i n e a r (D r o p O u t (x))\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 547, + 785, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which further helps to avoid overfitting and stabilize training.", + "bbox": [ + 214, + 571, + 660, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the help of this specialized mapping function along with a pre-trained neural network, the prompt $P$ is dynamically obtained from the input image $I$ and stage tokens $S_{t,q} \\coloneqq S_{t - (q - 1)} \\sim S_t$ :", + "bbox": [ + 212, + 585, + 787, + 633 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP = \\text {M a p p i n g} (\\text {N e t w o r k} ([ S _ {t, q}; I ]). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 641, + 784, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We reshape $P$ to a size of $N_{p} \\times C \\times L$ , where $N_{p}$ is the number of prompts, $C$ is the channel dimension, and $L$ indicates the number of layers where prompts are applied to the backbone model. The generated prompts are inserted across the first $L$ layers in the ViT backbone. In our DPG approach to prompting, we follow the technical foundations of prompting from the prior state-of-the-art PCL methods [32,38]. This is because we want to make fair comparison with the previous work, and our primary focus is on the formation of prompt. The prompt for the $l$ -th layer, denoted as $P_{l} \\in \\mathbb{R}^{N \\times C}$ , is partitioned into $\\{P_{l,k}, P_{l,v}\\} \\in \\mathbb{R}^{\\frac{N}{2} \\times C}$ and is prefixed to the input token embedding for both key and value in the attention mechanism [36]. This can be formulated as:", + "bbox": [ + 212, + 664, + 787, + 816 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\text {A t t e n t i o n} \\left(X _ {l} W _ {q} ^ {i}, [ P _ {k}; X ] W _ {k} ^ {i}, [ P _ {v}; X ] W _ {v} ^ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 824, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 436, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we omit layer index $l$ for brevity, and $i$ is head index in multi-head self-attention of transformer architecture.", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Joint Dynamic Prompt and Representation Learning. Differing from prior PCL methods [32,38,39] that freeze the backbone model (on the right side of the pipeline in Figure 2(a)) during training, we propose our Dynamic Prompt and Representation Learner (DPaRL), to jointly learn the dynamic prompt generation with the discriminative representation learning backbone (on the right side of Figure 2(b) and Figure 3). This approach aims to maximize the capability of entire pipeline to integrate old-stage concepts and new ones to encapsulate diverse semantics that aid generalization to open-world unseen classes and unknown domain shifts. Leveraging parameter-efficient fine-tuning techniques, our DPaRL successfully maximizes accuracy performance and minimizes catastrophic forgetting simultaneously. The overall framework is depicted in Figure 3. Note that the DPG encoder weights (on the left side of the pipeline) for prompt generation are frozen in our method, while the learnable parameters include stage tokens, weights in the mapping function and representation loss function, as well as the discriminative representation backbone weights.", + "bbox": [ + 212, + 200, + 789, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 453, + 375, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we empirically evaluate the effectiveness of our Dynamic Prompt and Representation Learner (DPaRL) in the established challenging continual open-world visual representation learning.", + "bbox": [ + 212, + 486, + 787, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets & Evaluation Metric", + "text_level": 1, + "bbox": [ + 215, + 556, + 511, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We conducted our training and testing across 4 prominent open-world image retrieval benchmarks: Cars [19], In-Shop [22], SOP [25] and iNat2018 [35]. We evaluate baselines and our proposed method following our practical continual open-world visual representation learning setting. In this dynamic scenario, classes are introduced incrementally at each learning stage, and the testing phase encompasses classes that have not been encountered during training. Our experiments consist of 10 or 100 stages, with the total training classes approximately distributed uniformly across these stages. The detailed data information including number of classes and samples are provided in the Supplementary Material. For inference, we use the entire testing dataset across all CL stages since there is no overlap between training and testing classes. Thus, our objective is to learn class relations while maintaining generalization across CL stages.", + "bbox": [ + 212, + 582, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metric. To evaluate model performance, we compute the Recall@1 metric using the [CLS] token at the last layer of the discriminative representation backbone network. Given our goal to optimize the model's adaptability throughout the CL phases, we report average Recall@1 across all CL stage. Let $R_{n}$ be the Recall@1 performance at stage $n$ , then we report $R_{N} = \\frac{1}{N_{s}}\\sum_{n = 1}^{N_{s}}R_{n}$ where $N_{s}$", + "bbox": [ + 212, + 765, + 787, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is the total number of stages. We also report the forgetting score $F_{N}$ by following [32,38] to measure the performance drop on the same unseen testing dataset across different training stages, that is, $F_{N} = \\frac{1}{N_{s} - 1}\\sum_{n = 2}^{N_{s}}\\max_{m\\in \\{1,\\dots,n - 1\\}}(0,R_{m} - R_{n})$ . We stress that $R_{N}$ holds greater importance as it captures both learning capacity and forgetting, while $F_{N}$ merely offers supplementary context.", + "bbox": [ + 212, + 146, + 787, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Baselines and Implementations", + "text_level": 1, + "bbox": [ + 214, + 253, + 517, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We validate 3 types of representative baselines:", + "bbox": [ + 214, + 277, + 642, + 292 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "-Rehearsal-based:Experience Replay (ER) [7],", + "- Regularization-based: Learning without Forgetting (LwF) [21]", + "- Rehearsal-free prompt-based: Learning to Prompt (L2P) [39], DualPrompt [38], and CodaPrompt [32]." + ], + "bbox": [ + 225, + 301, + 782, + 362 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "**Implementations Details.** Consistently across all methods, we adhere to the training setups by following [32]. Specifically, we utilize the Adam optimizer [18] with a batch size of 128 to train for 20 epochs to ensure comprehensive model convergence. The learning rate is cosine decaying with an initial value of 0.001.", + "bbox": [ + 212, + 369, + 785, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We utilize ArcFace [11] as the training loss function as it is a prevalent choice in open-world representation learning to achieve state-of-the-art performance [2]. For new-stage training, the prototype weights in the loss function of old training stages are appended to the new training stage as initialization. All prototype weights in the loss function, including those corresponding to the classes for old and new stages, are trainable during each training stage.", + "bbox": [ + 212, + 431, + 785, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For established PCL techniques, we maintain the prompt lengths and locations for L2P [39], DualPrompt [38], and CodaPrompt [32] as suggested in these works. For our DPaRL method, in each CL stage, we incorporate a learnable stage token with the first-in-first-out (FIFO) order and a max size of 5 tokens based on insights from Section 4.4. We apply our specialized mapping function in the DPG module to avoid over-fitting and stabilize training. We adopt LoRA [16] as the PEFT method to update backbone weights as guided by Section 4.5.", + "bbox": [ + 212, + 521, + 785, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Performance Comparison with Prior Arts", + "text_level": 1, + "bbox": [ + 214, + 648, + 606, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1 shows a comprehensive performance comparison with various CL methods across all four open-world image retrieval tasks. We adopt the official codebase of CodaPrompt [32] for the implementation of PCL baselines including L2P [39], DualPrompt [38], and CodaPrompt [32]. Additionally, we report the performance of representative rehearsal-based and regularization-based approaches, i.e., ER [7] and LWF [21], respectively, for a more comprehensive comparison. Note that, for a fair comparison, all methods use the pre-trained ViT-Base on ImageNet-21k as the backbone model, as shown on the right side of Figure 3.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "From the results in the Table 1, we observe that the performance trend of PCL-based methods (L2P, DualPrompt, CodaPrompt, our DPaRL) is similar to what was observed in the closed-world image recognition task [32], where PCL", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 436, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7c38699a069e9141d1d55a261eaaf4fee9b491c8b34d4e724dc2ac8f5edd990e.jpg", + "table_caption": [ + "Table 1: Performance on 4 open-world image retrieval benchmarks with 10 CL stages. We report the averaged $Recall@1$ ( $R_N$ ) and $Forgetting\\ Score$ ( $F_N$ ) across all continual learning stages, the results are averaged over 3 runs. We present the LowerBound with the zero-shot performance of ImageNet-21k pre-trained ViT-B/16 backbone. UpperBound denotes the upper bound performance when trained on the entire dataset at a single stage. All PCL methods, including L2P, DualPrompt, CodaPrompt, and our DPaRL, utilize the same pre-trained encoder, ViT-B/16 on ImageNet-21k, to generate the prompt tokens. The numbers in brackets show the $Recall@1$ improvement of our DPaRL over the previous best baselines, with the previous best baseline underlined." + ], + "table_footnote": [], + "table_body": "
DatasetCars [19]In-Shop [22]SOP [25]iNat2018 [35]Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
LowerBound53.86-47.16-62.85-71.74-58.90-
UpperBound85.96-89.28-85.21-82.68-85.78-
ER [7]65.380.4982.992.4865.380.5036.9418.5862.675.51
LwF [21]56.532.9181.523.3966.1813.2039.6933.8160.9813.33
L2P [39]56.270.0067.850.0076.880.4078.950.0069.990.10
DualPrompt [38]55.540.0065.560.0077.040.6678.710.0369.210.17
CodaPrompt [32]65.230.3678.610.0081.620.0178.970.0176.110.10
DPaRL (Ours)73.22(↑7.8)0.0386.28(↑3.3)0.0283.69(↑2.1)0.0780.02(↑1.0)0.0380.80(↑4.7)0.09
", + "bbox": [ + 230, + 282, + 772, + 397 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "methods outperform other continual learning approaches by a significant margin on both recall@1 metric $R_N$ and forgetting score $F_N^4$ , aligning with findings from previous research [32, 38, 39]. This suggests that PCL methods are strong baselines for continual learning on both closed-world image recognition tasks and open-world visual representation learning tasks. Notably, our DPaRL approach consistently outperforms other methods across all datasets, resulting in an average Recall@1 boost of $4.7\\%$ over the leading previous method, CodaPrompt [32]. This supports our hypothesis about the limitations of the conventional PCL design in open-world scenarios.", + "bbox": [ + 212, + 428, + 787, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the following, we provide step-by-step guidance on the design choices of our DPaRL. Additional analysis and results are presented in the supplementary material, covering the superior performance of DPaRL in few-shot settings, effectiveness of DPaRL in closed-world benchmarks, dynamic changes in the accuracy curve across CL stages, and robustness of DPaRL against various DPG encoders. Due to space limitations, we refer to the supplementary material for a more comprehensive exploration of these aspects.", + "bbox": [ + 212, + 565, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Effectiveness of Dynamic Prompt Generation", + "text_level": 1, + "bbox": [ + 214, + 696, + 635, + 713 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To solely investigate the effectiveness of our Dynamic Prompt Generation design, we freeze the backbone weights similar to prior PCL methods [32, 38, 39] in our DPaRL and denoted as the DPG method.", + "bbox": [ + 212, + 728, + 785, + 772 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "$^{4}$ PCL methods consistently achieve forgetting scores close to zero, indicating that performance continually improves. This trend is also evident in the dynamic change of accuracy curve in the supplementary material, highlighting that PCL methods serve as robust baselines for addressing catastrophic forgetting in open-world tasks.", + "bbox": [ + 217, + 782, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/97169a63cef55964afa7bb192180a1118f496ee640cc082bebb2f7b9dadb904c.jpg", + "image_caption": [ + "Fig. 4: Histogram of L2 distance between a pair of samples with embedding features extracted by Coda [32] (SOTA PCL method) and our DPG (the naive version of DPaRL with freezing backbone weights). Left and right figures are distributions of the seen and unseen classes in training and testing Cars [19] dataset, respectively. Our DPG exhibits enhanced separation between inter- and intra-classes, particularly on open-world unseen test classes." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 357, + 236 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/139fdfb8ac68d17f6424db357b5d5f8908e68f7b5e8fa91a82ce609ca6625ee3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 145, + 493, + 237 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/55c375d5732f61354c1ea85b0adfd0a16d298eb8bb0acffb4d1e3a1cb3ff2ca3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 143, + 640, + 236 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/593ec7899c6673d945afc746e8519d66b3670489bdee5179b831a66130a51f81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 145, + 774, + 236 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We visualize the distributions of L2 distance between pairwise samples within intra- and inter-classes of CodaPrompt (Coda) [32] and our DPG in Figure 4 to assess generalizability. A good model should exhibit distinguishable distributions between intra- and inter-classes for both closed-world seen classes and open-world unseen classes. Coda, employing a static prompt pool design, achieves a notable separation on closed-world intra- and inter-classes but exhibits a close distribution for open-world intra- and inter-classes, impeding effective generalization to unseen test classes. In contrast, our DPG shows a more distinct separation, with an average gap of 0.14 (twice as large as Coda's gap) between inter- and intra-class distances in testing data. This indicates better generalizability with a dynamic prompt generation mechanism for unseen classes.", + "bbox": [ + 212, + 380, + 787, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, accuracy results with Recall@1 in Table 2 and Table 4 reveal that our DPG significantly outperforms the best baseline (Coda), with an average of $9.1\\%$ boost in 100 CL stages and $3.2\\%$ boost in 10 CL stages, respectively. This highlights the superiority of Dynamic Prompting Generation design, particularly in practical scenarios involving a large number of stages. In the next Section 4.4.1 and Section 4.4.2, we investigate the importance of our design components of stage tokens and mapping function in the DPG network.", + "bbox": [ + 212, + 547, + 789, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4.1 Effectiveness of Stage Tokens", + "text_level": 1, + "bbox": [ + 215, + 681, + 504, + 698 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In our DPaRL design with Dynamic Prompt Generation (DPG), we introduce stage tokens to retain information on old training stages, and we impose a max limit on the number of stage tokens ( $M_{stg\\_tok} = 5$ ) to control storage and computation costs as the number of stages scales up, e.g., $N_{stg} = 100$ . Hence, when the stage number goes beyond the max limit $M_{stg\\_tok}$ , we replace some of previous stage tokens. Our hypothesis is that the earlier stage information learned from previous stage tokens can propagate into the later ones. In this section, we conduct ablations in a practical CL setting with a large number of stages (100 stages) to study the impact of stage tokens.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5cd6cd1a14bc290a63c620957ca939b8535d08595153b6db9f9fb11dcb081fb5.jpg", + "table_caption": [ + "Table 2: Impact of stage tokens with a large number of CL stages, $N_{stg} = 100$ . Recall@1 results on In-Shop, SOP, and iNat2018 datasets are reported. It's worth noting that the Cars dataset is relatively small, preventing its division into 100 stages. DPG is a naive version of our DPaRL method by freezing the backbone weights." + ], + "table_footnote": [], + "table_body": "
Method# Stage TokenOrderIn-ShopSOPiNatAverage
L2P [39]N/AN/A48.3970.3976.8265.20
Dual [38]N/AN/A52.5073.7676.9167.72
Coda [32]N/AN/A52.9673.9276.9867.95
DPGNoneN/A71.1880.0477.5776.26 (↑8.31)
DPG5FIFO (default)72.0580.8178.2977.05 (↑9.10)
DPG5FILO71.7680.6078.1276.83 (↑8.88)
DPG5Random72.0080.8378.2077.01 (↑9.06)
DPG10FIFO72.1580.8178.2877.08 (↑9.13)
DPG100full-size baseline72.2080.8478.3777.14 (↑9.19)
", + "bbox": [ + 269, + 212, + 733, + 329 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "With and without stage token. From Table 2, DPG with 5 stage tokens achieves a noticeable accuracy improvement over DPG without stage tokens, 77.05 vs. 76.26 on average, indicating the importance of stage token for retaining stage-wise class information in old training stages.", + "bbox": [ + 212, + 357, + 782, + 416 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Impact of the order. We conduct ablations on three different orders: first-in-first-out (FIFO), first-in-last-out (FILO), and randomly replacing one stage token (Random). Table 2 reveals that FIFO performs the best on average. FILO achieves slightly lower accuracy. Nevertheless, the Random method achieves performance very close to the best, indicating the robustness of our DPG design to different orders of stage tokens. We hypothesize this due to stage information learned from previous ones effectively propagating to the new stage tokens, regardless of the order or the location where the new stage token is inserted.", + "bbox": [ + 212, + 417, + 784, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Selection of maximum limit $M_{stg\\_tok}$ . When the stage number is large ( $N_{stg} = 100$ ), keeping a maximum token number of 5 achieves comparable performances (77.05 vs. 77.14 on average) to that of the full token number baseline with $M_{stg\\_tok} = 100$ . This signifies a highly effective selection with a constant number of stage tokens to achieve high accuracy, as well as good scalability with respect to the stage number. Hence, we use $M_{stg\\_tok} = 5$ throughout this paper.", + "bbox": [ + 215, + 537, + 787, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4.2 Effectiveness of Mapping Function", + "text_level": 1, + "bbox": [ + 215, + 651, + 547, + 666 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In our DPaRL framework (Figure 3), as elaborated in Section 3.3, we introduce a mapping function to match the size of [CLS] token and prompt tokens. We apply rank- $R$ constraint to avoid overfitting and reduce the memory consumption from the fully-connected layer inside the mapping function. To measure the sensitivity of parameter $R$ , we conduct experiments with varying $R$ from 64 to 768.", + "bbox": [ + 212, + 672, + 784, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results for both In-Shop and iNat in Table 3 indicate the existence of an optimal rank with $R_{opt} = 256$ that strikes a balance between accuracy and the number of parameters, saving $2/3$ of the total parameters compared to the full rank case. This default rank $R_{opt} = 256$ is subsequently used for all experiments.", + "bbox": [ + 212, + 750, + 784, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Note that when the rank reaches 768, representing the full rank of the mapping function essentially serving as a fully-connected layer, a significant per", + "bbox": [ + 212, + 810, + 784, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5fcb6b44d46be5391df05335ebcb05ab7b5da7ebfc039b971be1e9348c959346.jpg", + "table_caption": [ + "Table 3: Impact of the rank of weight matrix inside the mapping function on our DPG design. Recall@1 results on In-Shop and iNat2018 with 10 CL stages are reported." + ], + "table_footnote": [], + "table_body": "
Rank64128256384512768
In-Shop82.3383.1784.0983.8682.9871.40
iNat201879.2179.8580.0279.7479.6378.98
#Parameters2.0M3.9M7.9M12.0M16.1M23.5M
", + "bbox": [ + 313, + 184, + 687, + 243 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "formance drop of $12.69\\%$ is observed for the In-Shop dataset in Table 3. This suggests that an overparameterized expanding layer results in a severe accuracy degradation for a highly compressed [CLS] token, indicating the importance of our design. Moreover, our design has a robust rank range from 128 to 384, which closely matches the accuracy performance of the peak result.", + "bbox": [ + 212, + 268, + 787, + 345 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Effectiveness of Dynamic Prompt and Representation Learner", + "text_level": 1, + "bbox": [ + 214, + 364, + 772, + 381 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Previous PCL methods [32,38,39] freeze backbone weights of the discriminative model to reduce catastrophic forgetting. However, this approach could limit accuracy performance as it doesn't provide the degree of freedom to update the backbone, particularly in open-world setting where data distribution and domain shifts naturally exist. In this section, we explore effective ways to update backbone weights jointly with our dynamic prompt learning. We explore 3 approaches here: 1) fine-tuning $(FT)$ the entire backbone weights, 2) parameter-efficient fine-tuning (PEFT) with AdaptFromer [8] (Adapter) or Low-Rank Adaptation [16] weights on the backbone $(LoRA)$ , and 3) joint prompt generation and representation learning on the backbone weights (Coda and $DPaRL$ , $w / X$ ).", + "bbox": [ + 212, + 388, + 787, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4 shows that vanilla $FT$ or fine-tuning with PCL methods not only leads to lower accuracy but also suffers from higher catastrophic forgetting compared to DPG. While, LoRA gets higher accuracy $(+10.4\\%)$ on average) and reduces forgetting $(-2.3\\%)$ on average) compared to vanilla $FT$ , as well as outperforms Adapter, indicating a more effective way to update the backbone. Still, LoRA is less effective than DPG, with an average accuracy gap of $1.7\\%$ and a forgetting gap of $0.6\\%$ , indicating the superiority of our DPG design for open-world CL problems. Lastly, our joint Dynamic Prompt and Representation Learner (DPaRL) with LoRA improves average accuracy by $3.2\\%$ and reduces forgetting by $0.5\\%$ compared to LoRA-only baseline. It also boosts accuracy by $1.4\\%$ on average compared to DPG method without updating backbone weights at a negligible cost of 0.09 forgetting raises, indicating the effectiveness of our DPaRL design to maximizing accuracy without causing catastrophic forgetting issues. Therefore, unless stated otherwise, we adopt joint learning with dynamic prompt generation with LoRA as our default DPaRL method in this paper.", + "bbox": [ + 212, + 539, + 787, + 767 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Performance on Closed-World Setting in Other Domains", + "text_level": 1, + "bbox": [ + 214, + 786, + 730, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we assess the effectiveness of our DPaRL method in domains other than the established image retrieval benchmarks such as medical image", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 436, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/db99ec69f63a930675a3c5cf9511515f941994f5f79dbca9856fc05623082c91.jpg", + "table_caption": [ + "Table 4: Accuracy performance $(R_N, \\uparrow)$ and forgetting scores $(F_N, \\downarrow)$ in 10 CL stages with different fine-tuning methods: fine-tuning (FT) the entire backbone weights, PEFT with prompt tuning (Coda and our DPG), Adapter [8], and LoRA [16]." + ], + "table_footnote": [], + "table_body": "
DatasetCarsIn-ShopSOPiNat2018Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
Coda [32], freeze(frz) backbone65.230.3678.610.0081.620.0178.970.0176.110.10
DPG (Ours), frz backbone70.620.0084.090.0082.690.0080.020.0179.360.00
FineTune [32]57.575.0482.230.5473.452.2355.693.8467.242.91
Adapter [8]64.592.4783.940.0082.180.0677.270.3176.990.71
LoRA [16]67.770.9983.290.4081.970.2777.470.7977.630.61
Coda, w/ FT50.0312.7982.290.4574.033.2957.424.3965.945.23
Coda, w/ Adapter64.390.8683.560.4581.980.0578.560.0677.120.36
Coda, w/ LoRA69.061.2382.790.3382.510.2579.120.3478.370.54
DPaRL, w/ FT65.004.6182.360.3374.242.7156.734.1269.582.94
DPaRL, w/ Adapter68.471.0085.720.0182.570.0178.040.0378.700.26
DPaRL, w/ LoRA (default)73.220.0386.280.0283.690.0780.020.2280.800.09
", + "bbox": [ + 222, + 198, + 779, + 325 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "domain. However, there are no available open-world benchmarks so we follow the standard closed-world setting in Figure 1 (b). We conduct experiments on MedMNIST [40], which consists of approximately 700K biomedical images. It includes 7 datasets, each with $7\\sim 11$ classes. We evenly split the classes into 3 CL stages and report the average top-1 accuracy over each stage. Results in Table 5 demonstrate that DPaRL generalizes well to the medical domain, outperforming other methods with an average improvement of $6.50\\%$ . For additional tasks in other domains, please refer to the supplementary material.", + "bbox": [ + 215, + 351, + 784, + 470 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/06ada58c40519e3df5205419138826cf469aa58e39b9d87e92dda5fbe2dd9803.jpg", + "table_caption": [ + "Table 5: Top-1 (%) on medical data MedMNIST [40]. Our DPaRL surpasses other PCL methods by $6.50\\%$ on average." + ], + "table_footnote": [], + "table_body": "
MethodsBloodDermaOrganAOrganCOrganSPathTissueAverage
L2P56.0334.0061.1252.8746.2556.3941.2949.71
Dual66.2833.1665.4257.7948.2260.6940.3453.13
Coda57.3733.7265.4165.9846.8279.8532.3054.49
DPaRL (Ours)70.4938.7170.3971.2951.2381.4443.3860.99(↑6.50)
", + "bbox": [ + 245, + 537, + 753, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 643, + 357, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a practical setting for continual visual representation learning in the open-world. Existing continual learning methods fail to generalize well in this proposed setting. To mitigate this challenge, we propose Dynamic Prompt and Representation Learner (DPaRL), to enhance the adaptability and performance in continual learning for open-world visual recognition. Our method consistently demonstrates superior performance across various open-world image retrieval datasets and settings, by dynamically generating prompts from a deep neural network coupled with a dedicated mapping function, and effectively updating the discriminative representation backbone. Notably, it overcomes the intrinsic constraints found in static prompt pool designs, thereby expanding the limits of model generalization in open-world scenarios.", + "bbox": [ + 215, + 672, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018)", + "2. An, X., Deng, J., Yang, K., Li, J., Feng, Z., Guo, J., Yang, J., Liu, T.: Unicom: Universal and compact representation learning for image retrieval. arXiv preprint arXiv:2304.05884 (2023)", + "3. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "4. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020)", + "5. Cao, T., Wang, Y., Xing, Y., Xiao, T., He, T., Zhang, Z., Zhou, H., Tighe, J.: Pss: Progressive sample selection for open-world visual representation learning. In: European Conference on Computer Vision. pp. 278-294. Springer (2022)", + "6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018)", + "7. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019)", + "8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535 (2022)", + "9. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021)", + "10. De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence 44(7), 3366-3385 (2021)", + "11. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)", + "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "13. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)", + "14. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020)", + "15. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9769-9776. IEEE (2019)", + "16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + ], + "bbox": [ + 218, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 436, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Kim, S., Kim, D., Cho, M., Kwak, S.: Proxy anchor loss for deep metric learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3238-3247 (2020)", + "18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)", + "20. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: International Conference on Machine Learning. pp. 3925-3934. PMLR (2019)", + "21. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)", + "22. Liu, Z., Luo, P., Qiu, S., Wang, X., Tang, X.: Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1096-1104 (2016)", + "23. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022)", + "24. Movshovitz-Attias, Y., Toshev, A., Leung, T.K., Ioffe, S., Singh, S.: No fuss distance metric learning using proxies. In: Proceedings of the IEEE international conference on computer vision. pp. 360-368 (2017)", + "25. Oh Song, H., Xiang, Y., Jegelka, S., Savarese, S.: Deep metric learning via lifted structured feature embedding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4004-4012 (2016)", + "26. Rao, D., Visin, F., Rusu, A., Pascanu, R., Teh, Y.W., Hadsell, R.: Continual unsupervised representation learning. Advances in neural information processing systems 32 (2019)", + "27. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023)", + "28. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)", + "29. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)", + "30. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: International conference on machine learning. pp. 4548-4557. PMLR (2018)", + "31. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021)", + "32. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)", + "33. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV. pp. 1706-1716 (2023)", + "35. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)", + "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "37. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2023)", + "38. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)", + "39. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)", + "40. Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., Ni, B.: Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data 10(1), 41 (2023)", + "41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020)", + "42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017)", + "43. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017)", + "44. Zhao, B., Tang, S., Chen, D., Bilen, H., Zhao, R.: Continual representation learning for biometric identification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 1198-1208 (2021)", + "45. Zhou, D.W., Ye, H.J., Zhan, D.C., Liu, Z.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 659 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "DPaRL for Open-World Continual Learning", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_model.json b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9ba5bc4269c358420f45ec68afb2720f5f92e73f --- /dev/null +++ b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_model.json @@ -0,0 +1,2335 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.261, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "Open-World Dynamic Prompt and Continual Visual Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.212, + 0.737, + 0.259 + ], + "angle": 0, + "content": "Youngeun Kim\\(^{1,*†}\\), Jun Fang\\(^{2,*‡}\\), Qin Zhang\\(^{2}\\), Zhaowei Cai\\(^{3}\\), Yantao Shen\\(^{2}\\), Rahul Duggal\\(^{2}\\), Dripta S. Raychaudhuri\\(^{2}\\), Zhuowen Tu\\(^{2}\\), Yifan Xing\\(^{2}\\), Onkar Dabeer\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.27, + 0.561, + 0.284 + ], + "angle": 0, + "content": "Yale University" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.284, + 0.558, + 0.297 + ], + "angle": 0, + "content": "AWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.297, + 0.555, + 0.311 + ], + "angle": 0, + "content": "3 Amazon AGI" + }, + { + "type": "list", + "bbox": [ + 0.443, + 0.27, + 0.561, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.313, + 0.758, + 0.34 + ], + "angle": 0, + "content": "youngun.kim@yale.edu {junfa, qzaamz, zhaoweic, yantaos, dugrahul, driptarc, ztu, yifax, onkardab}@amazon.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.378, + 0.752, + 0.628 + ], + "angle": 0, + "content": "Abstract. The open world is inherently dynamic, characterized by ever-evolving concepts and distributions. Continual learning (CL) in this dynamic open-world environment presents a significant challenge in effectively generalizing to unseen test-time classes. To address this challenge, we introduce a new practical CL setting tailored for open-world visual representation learning. In this setting, subsequent data streams systematically introduce novel classes that are disjoint from those seen in previous training phases, while also remaining distinct from the unseen test classes. In response, we present Dynamic Prompt and Representation Learner (DPaRL), a simple yet effective Prompt-based CL (PCL) method. Our DPaRL learns to generate dynamic prompts for inference, as opposed to relying on a static prompt pool in previous PCL methods. In addition, DPaRL jointly learns dynamic prompt generation and discriminative representation at each training stage whereas prior PCL methods only refine the prompt learning throughout the process. Our experimental results demonstrate the superiority of our approach, surpassing state-of-the-art methods on well-established open-world image retrieval benchmarks by an average of \\(4.7\\%\\) improvement in Recall@1 performance." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.64, + 0.741, + 0.671 + ], + "angle": 0, + "content": "Keywords: Dynamic Prompt Generation \\(\\cdot\\) Continual Learning \\(\\cdot\\) Open-World Visual Representation Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.377, + 0.712 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Continual learning (CL) without catastrophic forgetting is a challenging task in practice as retraining with old data is prohibitively expensive when data volume grows [10,14,23]. It is particularly challenging for real-world applications in the open-world [2,5], where new concepts keep emerging and distribution shifts are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.365, + 0.811 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.812, + 0.563, + 0.825 + ], + "angle": 0, + "content": "† Work conducted during an internship at Amazon." + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.826, + 0.384, + 0.84 + ], + "angle": 0, + "content": "‡ Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.563, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.15, + 0.353, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.234, + 0.314, + 0.248 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.15, + 0.477, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.234, + 0.436, + 0.248 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.15, + 0.613, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.234, + 0.562, + 0.248 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.149, + 0.768, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.234, + 0.71, + 0.248 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.272, + 0.789, + 0.412 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the distinctions between our problem setting and traditional settings. In our work, we aim to address the problem where training splits have no class overlaps. (a) Closed-world setting: both training and testing classes are identical. (b) Continual learning (CL) in closed-world setting: training classes are split into multiple divisions and are introduced through various CL stages. (c) Open-world setting: training and testing classes remain separate. The aim is to learn a robust representation from training classes that generalize to unseen classes. (d) Open-World Continual Representation Learning (Our Problem Setting): continual learning tailored for an open-world scenario. We sequentially introduce training classes over multiple CL stages, ensuring they remain distinct from test-time unseen classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.506 + ], + "angle": 0, + "content": "inevitable. Recent studies on Prompt-based Continual Learning (PCL) [32,38, 39] demonstrated success in mitigating catastrophic forgetting without access to any past data. However, they heavily focus on the closed-world setting where test-time categories come from those encountered in various training stages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.657 + ], + "angle": 0, + "content": "Whereas in practice, the world is dynamic and concepts evolve over time and new categories are often introduced after a model is trained. Consider, for instance, an image retrieval system for clothing brands in online e-commerce. During training, one would gather images of all existing clothing brands available to train a model. However, after deployment, the system encounters new brands. The new data collected for system upgrade, constituting a new training stage, will likely include brands that are disjoint to those in older stages. Furthermore, the system will constantly face new brands at test-time that are disjoint to any of those encountered in the previous training stages and is expected to generalize well to those new categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Thus, to comprehensively evaluate the ability of a model in retaining semantic information from continuous streams of data for generalization to new concepts at test-time for practical applications, we propose a new continual open-world representation learning setting, as illustrated in Figure 1(d). In this setting, different data streams comprise distinct classes, and test-time classes are disjoint to those encountered in all training phases. Our setting is notably more challenging than the previous closed-world CL scenario. It closely mirrors real-world demands on recognition systems, where, during any evaluation stage, the test set consists of unseen entities to the training process up to that point. This setup reflects the need for practical recognition systems to encode general semantic information from training-time classes and apply this knowledge to new, unseen concepts without access to older training data during system updates." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.438, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "We start with evaluating state-of-the-art PCL methods [32, 38, 39] in our proposed setting and find that they fail to generalize with the absence of past data. For example, in our experiments, the gap between the paragon of standard training with all data in one phase and the best PCL baseline (Coda) shows up to \\(20\\%\\) gap in Recall@1 (Table 1), significantly larger than the \\(5\\%\\) gap in the closed-world setting, which suggests that prior PCL methods encounter difficulties and are less effective in the open-world setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.785, + 0.36 + ], + "angle": 0, + "content": "To tackle this challenging yet practical open-world CL setting, we introduce a simple yet effective method called Dynamic Prompt and Representation Learner (DPaRL). Unlike previous prompt pool based PCL methods [32, 38, 39] that depend on a static prompt pool (Figure 2(a)) for test-time predictions. Notably, DPaRL trains an innovative dynamic prompt generation network jointly with the discriminative representation backbone model (Figure 2(b)), departing from the prompt-only learning paradigm in prior PCL methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.572 + ], + "angle": 0, + "content": "Our key insight lies in that simply combining the fixed prompts learnt over training classes does not generalize well to unseen test-time classes in the practical open-world setting. On the other hand, the on-the-fly dynamically generated prompts in our method exhibit a stronger capability of capturing the diverse semantics spread across the numerous divisions of classes observed during the different training stages of continual learning, improving generalization to test-time unseen classes. This is achieved through the proposed dynamic prompt generation network, which possesses stronger representation power with a learnable mapping function compared to the traditional approach of combining multiple prompts in a static prompt pool. Additionally, unlike prior PCL methods, DPaRL jointly updates the dynamic prompt generation process and the discriminative representation backbone weights, facilitating maximized interaction and comparison between old concepts and newer ones to encapsulate diverse semantics that aid generalization to unseen test-time classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.784, + 0.681 + ], + "angle": 0, + "content": "Consequently, over the four open-world image retrieval benchmarks on cars, online catalogs, products, and species images where the task is to search for the most similar images belonging to the same class as the query, via a distance measurement in an embedding space, DPaRL achieves significant accuracy boosts. It improves the average Recall@1 performance by absolute \\(4.7\\%\\) and \\(9.1\\%\\) compared to state-of-the-art continual learning methods, from \\(76.1\\%\\) to \\(80.8\\%\\) over 10 CL stages and from \\(68.0\\%\\) to \\(77.1\\%\\) over 100 CL stages, respectively." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.684, + 0.574, + 0.698 + ], + "angle": 0, + "content": "In summary, our contributions are three folds:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.716, + 0.784, + 0.745 + ], + "angle": 0, + "content": "1. We establish a new practical setting on continual visual representation learning in the open-world." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.748, + 0.785, + 0.808 + ], + "angle": 0, + "content": "2. We introduce a simple yet powerful method, Dynamic Prompt and Representation Learner (DPaRL), which dynamically generates prompts while effectively updating the discriminative representation backbone. This enhancement improves generalization for unseen open-world classes at test time." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "3. We outperform state-of-the-art continual learning methods in the proposed practical setting, both rehearsal-free and rehearsal-based." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.716, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.147, + 0.495, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.241, + 0.436, + 0.255 + ], + "angle": 0, + "content": "(a) Previous method" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.148, + 0.772, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.241, + 0.69, + 0.255 + ], + "angle": 0, + "content": "(b) Our work" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 2: Prior PCL methods combine prompts from a static prompt pool trained on the training class distribution, leading to a loss of generalization capability when facing unseen classes during test time. Our work introduces a Dynamic Prompt Generation network that generates dynamic prompts on the fly by integrating a given image with stage tokens, followed by a specialized mapping function and adjustable discriminative representation backbone weights, providing generalizable prompts for unseen testing classes, distinguishing it from prior PCL methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.403, + 0.388, + 0.419 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Continual learning. Continual learning (CL) has been recognized for decades as a crucial research domain, focusing on the ongoing adaptation of models over time. Early works primarily centered around regularization-based methods in an effort to address catastrophic forgetting [1,21,43] by introducing regularization between its current and previously learned parameters. However, they often underperform when faced with complex datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.644 + ], + "angle": 0, + "content": "A separate line of work, architecture-based approach, involves expanding the network architecture in response to new learning phases [20,26,29,30,42]. While these methods generally outperform regularization-based counterparts, they do so at the cost of introducing additional parameters. Recent advancements include rehearsal-based methods that incorporate a memory component for storing past data [4,6,7,15,28]. This archived data aids in training newer stages, typically yielding superior performance compared to other CL techniques. Yet, this data retention can introduce privacy concerns and significant memory overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Given these constraints, rehearsal-free strategies are gaining attention in the community. These aim to mitigate catastrophic forgetting without relying on historical data. A prominent method within this category uses model inversion to produce rehearsal images [9,13,31,41]. However, model inversion is computationally intensive and time-consuming. Additionally, these methods lag significantly in performance when compared to their rehearsal-based counterparts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Recent developments have witnessed the rise of prompt-based approaches [27, 32, 34, 37-39, 45]. These strategies offer robust protection against catastrophic forgetting by learning a small set of parameters, or prompts, instead of directly training all parameters. For instance, L2P [39] employs a pool of prompts, selecting suitable ones for insertion based on input data clustering. Building on this concept, DualPrompt [38] presents general prompts designed to encapsulate shared knowledge across various learning stages. Advancing this further," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.438, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "the state-of-the-art method CodaPrompt [32] facilitates end-to-end training by incorporating an attention mechanism for prompt combination within a pool. However, during inference, the above works rely on the combination of prompts from a static pool learned from training to feed into a frozen backbone model, which has limitations in cases where testing data has distribution shifts or disjoint classes from the training data. This motivates us to design a Dynamic Prompt and Representation Learner (DPaRL) to address these drawbacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.789, + 0.502 + ], + "angle": 0, + "content": "Open-World Visual Representation Learning. In open-world visual representation learning, the model learns discriminative representations that align distances between representations with their semantic similarities, and uses these knowledge for testing on unseen classes. Previous works in this domain have primarily focused on designing loss functions [11, 17, 24] for enhanced accuracy. Recently, several studies [2, 12] have demonstrated the effectiveness of utilizing pre-trained visual foundation models and fine-tuning them to achieve significant performance improvements. While open-world visual representation learning has been under investigation for an extended period, there has been limited exploration within the continual learning (CL) framework for this task. [44] investigated biometric datasets using a regularization-based CL approach. However, the investigation into state-of-the-art prompt-based CL methods remains missing. Furthermore, as their studies primarily centered on biometric data, there is a pressing need to explore established continual learning benchmarks specific to open-world visual representation learning in more natural settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.381, + 0.541 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.585 + ], + "angle": 0, + "content": "In this section, we introduce our open-world Continual Learning (CL) setting and motivate our Dynamic Prompt and Representation Learner (DPaRL)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.4, + 0.621 + ], + "angle": 0, + "content": "3.1 Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Two fundamental visual representation learning settings are the closed-world and open-world paradigms. In a closed-world setting (as depicted in Figure 1(a)), the training and testing data classes are entirely identical. Conversely, the open-world setting (illustrated in Figure 1(c)) presents a more practical scenario. Here, training and testing classes are entirely distinct, thus requiring the model to learn representations that generalize to unseen concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "When faced with continuous streams of data for visual recognition system updates, built upon the closed-world paradigm, Figure 1(b) showcases the well-known continual learning setting for image recognition. In this work, we introduce a new practical setting: Open-World Continual Representation Learning as shown in Figure 1(d). Instead of testing on seen classes of the various training stages thus far, the testing classes are entirely disjoint to those faced in different training stages, necessitating the model to not only learn continually but also generalize to new concepts, making it a more dynamic and challenging setting." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.148, + 0.756, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.788, + 0.387 + ], + "angle": 0, + "content": "Fig. 3: Illustration of our proposed method: Dynamic Prompt and Representation Leaner (DPaRL). We dynamically generate prompt tokens on the fly from Dynamic Prompt Generation (DPG) networks by integrating information from stage tokens and image tokens. The [CLS] token from DPG is converted to prompt tokens via a linear mapping function. The generated prompt tokens are added to the backbone ViT and trained with a loss function. The learnable parameters include the current stage token, weights in mapping function, weights in backbone, and weights in loss function." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.628, + 0.429 + ], + "angle": 0, + "content": "3.2 Prompt-based Continual Learning Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.789, + 0.603 + ], + "angle": 0, + "content": "Prompt-based Continual Learning (PCL) methods [32,38,39] use a pre-trained Vision Transformer (ViT) as a discriminative backbone for closed-world image classification, as shown in the Figure 2(a). These methods create a prompt pool with multiple prompt tokens, updating only the learnable parameters in this pool during training. In inference, the learned prompt pool is static, and PCL methods select tokens from this pool to feed into multiple ViT backbone layers for predictions. Our method adopts this PCL paradigm but adapts to the open-world representation learning setting with a prototypical metric learning loss. Additionally, we introduce a dynamic prompt generation (DPG) network to replace the static prompt pool for effective joint learning with the discriminative representation backbone model, as detailed below." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.715, + 0.639 + ], + "angle": 0, + "content": "3.3 Dynamic Prompt and Representation Leaner (DPaRL)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Overall Pipeline. In the open-world setting where training and testing classes are disjoint, existing prompt pool designs in PCL [32,38,39] exhibit limitations, i.e., limited separation between intra- and inter- testing class distance distributions as shown in Figure 4 and large performance gap from paragon in Table 1. To mitigate, we first propose dynamically generating prompts by blending the given image information and continual learning stage's information through a dynamic prompt generation (DPG) network. It aims to provide better representation power compared to the prompt pool design, which combines a fixed prompt pool trained from the training class distribution. Moreover, we enhance the capability of the discriminative representation backbone model to more effectively generalize to the open-world concepts through a joint dynamic prompt and representation learning paradigm. We refer to Figure 2, 3 for detailed design differences between our approach and prior prompt pool-based methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Dynamic Prompt Generation Network. In order to retain information from each CL stage of our Dynamic Prompt Generation (DPG) process, we introduce stage tokens, denoted as \\( S \\). During training stage \\( t \\), we train a stage token \\( S_{t} \\) while freezing the previous stage tokens \\( S_{t - (q - 1)} \\sim S_{t - 1} \\) in a first-in-first-out (FIFO) queue manner, where \\( q \\) is the max queue size. Guided by the ablation study detailed in Section 4.4, we typically set \\( q = 5 \\) using a FIFO order to aim for both high accuracy and scalability. This ensures that the knowledge from the previous stages remains unchanged, and the total number of stage tokens limits by the queue size without linearly scaling with stage numbers. Within the deep neural network such as ViT, the self-attention block facilitates the integration of information between the stage-wise tokens and the instance-wise image tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.388 + ], + "angle": 0, + "content": "As illustrated in Figure 3, the DPG produces a [CLS] token to dynamically acquire task-specific high-level information for prior knowledge of class-level features. However, it differs in size from the prompt tokens required by the discriminative backbone model (right side of Figure 3). To address this, we introduce a mapping function between the [CLS] token and the prompt tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.542 + ], + "angle": 0, + "content": "A straightforward approach is to employ a single linear layer for dimensionality transformation. However, such a method introduces an excessive number of additional parameters, and the resulting over-parameterization can lead to overfitting of the highly compressed [CLS] token information. To tackle this issue, we impose a constraint on the weight parameter \\( W = AB^{T} \\in \\mathbb{R}^{C_{in} \\times C_{out}} \\), where \\( A \\in \\mathbb{R}^{C_{in} \\times R} \\), \\( B \\in \\mathbb{R}^{C_{out} \\times R} \\), ensuring a maximum rank of \\( R < \\min(C_{in}, C_{out}) \\). This constraint is inspired by the success of the LoRA technique [16]. The experimental results in Section 4.4 illustrate the critical importance of the low-rank mapping function design for our DPG module. Moreover, we apply Dropout [33] and LayerNorm [3] to this low-rank linear mapping function" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.549, + 0.786, + 0.565 + ], + "angle": 0, + "content": "\\[\nM a p p i n g (x) = L a y e r N o r m \\left(L o w R a n k L i n e a r (D r o p O u t (x))\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.572, + 0.661, + 0.587 + ], + "angle": 0, + "content": "which further helps to avoid overfitting and stabilize training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.634 + ], + "angle": 0, + "content": "With the help of this specialized mapping function along with a pre-trained neural network, the prompt \\( P \\) is dynamically obtained from the input image \\( I \\) and stage tokens \\( S_{t,q} \\coloneqq S_{t - (q - 1)} \\sim S_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.642, + 0.785, + 0.657 + ], + "angle": 0, + "content": "\\[\nP = \\text {M a p p i n g} (\\text {N e t w o r k} ([ S _ {t, q}; I ]). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.817 + ], + "angle": 0, + "content": "We reshape \\( P \\) to a size of \\( N_{p} \\times C \\times L \\), where \\( N_{p} \\) is the number of prompts, \\( C \\) is the channel dimension, and \\( L \\) indicates the number of layers where prompts are applied to the backbone model. The generated prompts are inserted across the first \\( L \\) layers in the ViT backbone. In our DPG approach to prompting, we follow the technical foundations of prompting from the prior state-of-the-art PCL methods [32,38]. This is because we want to make fair comparison with the previous work, and our primary focus is on the formation of prompt. The prompt for the \\( l \\)-th layer, denoted as \\( P_{l} \\in \\mathbb{R}^{N \\times C} \\), is partitioned into \\( \\{P_{l,k}, P_{l,v}\\} \\in \\mathbb{R}^{\\frac{N}{2} \\times C} \\) and is prefixed to the input token embedding for both key and value in the attention mechanism [36]. This can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.825, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\text {A t t e n t i o n} \\left(X _ {l} W _ {q} ^ {i}, [ P _ {k}; X ] W _ {k} ^ {i}, [ P _ {v}; X ] W _ {v} ^ {i}\\right). \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.177 + ], + "angle": 0, + "content": "Here, we omit layer index \\( l \\) for brevity, and \\( i \\) is head index in multi-head self-attention of transformer architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.201, + 0.79, + 0.43 + ], + "angle": 0, + "content": "Joint Dynamic Prompt and Representation Learning. Differing from prior PCL methods [32,38,39] that freeze the backbone model (on the right side of the pipeline in Figure 2(a)) during training, we propose our Dynamic Prompt and Representation Learner (DPaRL), to jointly learn the dynamic prompt generation with the discriminative representation learning backbone (on the right side of Figure 2(b) and Figure 3). This approach aims to maximize the capability of entire pipeline to integrate old-stage concepts and new ones to encapsulate diverse semantics that aid generalization to open-world unseen classes and unknown domain shifts. Leveraging parameter-efficient fine-tuning techniques, our DPaRL successfully maximizes accuracy performance and minimizes catastrophic forgetting simultaneously. The overall framework is depicted in Figure 3. Note that the DPG encoder weights (on the left side of the pipeline) for prompt generation are frozen in our method, while the learnable parameters include stage tokens, weights in the mapping function and representation loss function, as well as the discriminative representation backbone weights." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.377, + 0.471 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.789, + 0.534 + ], + "angle": 0, + "content": "In this section, we empirically evaluate the effectiveness of our Dynamic Prompt and Representation Learner (DPaRL) in the established challenging continual open-world visual representation learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.512, + 0.571 + ], + "angle": 0, + "content": "4.1 Datasets & Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Datasets. We conducted our training and testing across 4 prominent open-world image retrieval benchmarks: Cars [19], In-Shop [22], SOP [25] and iNat2018 [35]. We evaluate baselines and our proposed method following our practical continual open-world visual representation learning setting. In this dynamic scenario, classes are introduced incrementally at each learning stage, and the testing phase encompasses classes that have not been encountered during training. Our experiments consist of 10 or 100 stages, with the total training classes approximately distributed uniformly across these stages. The detailed data information including number of classes and samples are provided in the Supplementary Material. For inference, we use the entire testing dataset across all CL stages since there is no overlap between training and testing classes. Thus, our objective is to learn class relations while maintaining generalization across CL stages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.789, + 0.845 + ], + "angle": 0, + "content": "Evaluation Metric. To evaluate model performance, we compute the Recall@1 metric using the [CLS] token at the last layer of the discriminative representation backbone network. Given our goal to optimize the model's adaptability throughout the CL phases, we report average Recall@1 across all CL stage. Let \\( R_{n} \\) be the Recall@1 performance at stage \\( n \\), then we report \\( R_{N} = \\frac{1}{N_{s}}\\sum_{n = 1}^{N_{s}}R_{n} \\) where \\( N_{s} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.234 + ], + "angle": 0, + "content": "is the total number of stages. We also report the forgetting score \\( F_{N} \\) by following [32,38] to measure the performance drop on the same unseen testing dataset across different training stages, that is, \\( F_{N} = \\frac{1}{N_{s} - 1}\\sum_{n = 2}^{N_{s}}\\max_{m\\in \\{1,\\dots,n - 1\\}}(0,R_{m} - R_{n}) \\). We stress that \\( R_{N} \\) holds greater importance as it captures both learning capacity and forgetting, while \\( F_{N} \\) merely offers supplementary context." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.254, + 0.519, + 0.269 + ], + "angle": 0, + "content": "4.2 Baselines and Implementations" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.643, + 0.294 + ], + "angle": 0, + "content": "Baselines. We validate 3 types of representative baselines:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.302, + 0.574, + 0.317 + ], + "angle": 0, + "content": "-Rehearsal-based:Experience Replay (ER) [7]," + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.317, + 0.69, + 0.332 + ], + "angle": 0, + "content": "- Regularization-based: Learning without Forgetting (LwF) [21]" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.332, + 0.784, + 0.363 + ], + "angle": 0, + "content": "- Rehearsal-free prompt-based: Learning to Prompt (L2P) [39], DualPrompt [38], and CodaPrompt [32]." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.302, + 0.784, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.786, + 0.432 + ], + "angle": 0, + "content": "**Implementations Details.** Consistently across all methods, we adhere to the training setups by following [32]. Specifically, we utilize the Adam optimizer [18] with a batch size of 128 to train for 20 epochs to ensure comprehensive model convergence. The learning rate is cosine decaying with an initial value of 0.001." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.787, + 0.522 + ], + "angle": 0, + "content": "We utilize ArcFace [11] as the training loss function as it is a prevalent choice in open-world representation learning to achieve state-of-the-art performance [2]. For new-stage training, the prototype weights in the loss function of old training stages are appended to the new training stage as initialization. All prototype weights in the loss function, including those corresponding to the classes for old and new stages, are trainable during each training stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.629 + ], + "angle": 0, + "content": "For established PCL techniques, we maintain the prompt lengths and locations for L2P [39], DualPrompt [38], and CodaPrompt [32] as suggested in these works. For our DPaRL method, in each CL stage, we incorporate a learnable stage token with the first-in-first-out (FIFO) order and a max size of 5 tokens based on insights from Section 4.4. We apply our specialized mapping function in the DPG module to avoid over-fitting and stabilize training. We adopt LoRA [16] as the PEFT method to update backbone weights as guided by Section 4.5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.65, + 0.607, + 0.667 + ], + "angle": 0, + "content": "4.3 Performance Comparison with Prior Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Table 1 shows a comprehensive performance comparison with various CL methods across all four open-world image retrieval tasks. We adopt the official codebase of CodaPrompt [32] for the implementation of PCL baselines including L2P [39], DualPrompt [38], and CodaPrompt [32]. Additionally, we report the performance of representative rehearsal-based and regularization-based approaches, i.e., ER [7] and LWF [21], respectively, for a more comprehensive comparison. Note that, for a fair comparison, all methods use the pre-trained ViT-Base on ImageNet-21k as the backbone model, as shown on the right side of Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "From the results in the Table 1, we observe that the performance trend of PCL-based methods (L2P, DualPrompt, CodaPrompt, our DPaRL) is similar to what was observed in the closed-world image recognition task [32], where PCL" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Table 1: Performance on 4 open-world image retrieval benchmarks with 10 CL stages. We report the averaged \\( Recall@1 \\) (\\( R_N \\)) and \\( Forgetting\\ Score \\) (\\( F_N \\)) across all continual learning stages, the results are averaged over 3 runs. We present the LowerBound with the zero-shot performance of ImageNet-21k pre-trained ViT-B/16 backbone. UpperBound denotes the upper bound performance when trained on the entire dataset at a single stage. All PCL methods, including L2P, DualPrompt, CodaPrompt, and our DPaRL, utilize the same pre-trained encoder, ViT-B/16 on ImageNet-21k, to generate the prompt tokens. The numbers in brackets show the \\( Recall@1 \\) improvement of our DPaRL over the previous best baselines, with the previous best baseline underlined." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.283, + 0.773, + 0.398 + ], + "angle": 0, + "content": "
DatasetCars [19]In-Shop [22]SOP [25]iNat2018 [35]Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
LowerBound53.86-47.16-62.85-71.74-58.90-
UpperBound85.96-89.28-85.21-82.68-85.78-
ER [7]65.380.4982.992.4865.380.5036.9418.5862.675.51
LwF [21]56.532.9181.523.3966.1813.2039.6933.8160.9813.33
L2P [39]56.270.0067.850.0076.880.4078.950.0069.990.10
DualPrompt [38]55.540.0065.560.0077.040.6678.710.0369.210.17
CodaPrompt [32]65.230.3678.610.0081.620.0178.970.0176.110.10
DPaRL (Ours)73.22(↑7.8)0.0386.28(↑3.3)0.0283.69(↑2.1)0.0780.02(↑1.0)0.0380.80(↑4.7)0.09
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.563 + ], + "angle": 0, + "content": "methods outperform other continual learning approaches by a significant margin on both recall@1 metric \\( R_N \\) and forgetting score \\( F_N^4 \\), aligning with findings from previous research [32, 38, 39]. This suggests that PCL methods are strong baselines for continual learning on both closed-world image recognition tasks and open-world visual representation learning tasks. Notably, our DPaRL approach consistently outperforms other methods across all datasets, resulting in an average Recall@1 boost of \\( 4.7\\% \\) over the leading previous method, CodaPrompt [32]. This supports our hypothesis about the limitations of the conventional PCL design in open-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.673 + ], + "angle": 0, + "content": "In the following, we provide step-by-step guidance on the design choices of our DPaRL. Additional analysis and results are presented in the supplementary material, covering the superior performance of DPaRL in few-shot settings, effectiveness of DPaRL in closed-world benchmarks, dynamic changes in the accuracy curve across CL stages, and robustness of DPaRL against various DPG encoders. Due to space limitations, we refer to the supplementary material for a more comprehensive exploration of these aspects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.637, + 0.714 + ], + "angle": 0, + "content": "4.4 Effectiveness of Dynamic Prompt Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.787, + 0.773 + ], + "angle": 0, + "content": "To solely investigate the effectiveness of our Dynamic Prompt Generation design, we freeze the backbone weights similar to prior PCL methods [32, 38, 39] in our DPaRL and denoted as the DPG method." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.784, + 0.788, + 0.842 + ], + "angle": 0, + "content": "\\(^{4}\\) PCL methods consistently achieve forgetting scores close to zero, indicating that performance continually improves. This trend is also evident in the dynamic change of accuracy curve in the supplementary material, highlighting that PCL methods serve as robust baselines for addressing catastrophic forgetting in open-world tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.438, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.145, + 0.359, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.146, + 0.495, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.145, + 0.642, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.146, + 0.776, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.264, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig. 4: Histogram of L2 distance between a pair of samples with embedding features extracted by Coda [32] (SOTA PCL method) and our DPG (the naive version of DPaRL with freezing backbone weights). Left and right figures are distributions of the seen and unseen classes in training and testing Cars [19] dataset, respectively. Our DPG exhibits enhanced separation between inter- and intra-classes, particularly on open-world unseen test classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.547 + ], + "angle": 0, + "content": "We visualize the distributions of L2 distance between pairwise samples within intra- and inter-classes of CodaPrompt (Coda) [32] and our DPG in Figure 4 to assess generalizability. A good model should exhibit distinguishable distributions between intra- and inter-classes for both closed-world seen classes and open-world unseen classes. Coda, employing a static prompt pool design, achieves a notable separation on closed-world intra- and inter-classes but exhibits a close distribution for open-world intra- and inter-classes, impeding effective generalization to unseen test classes. In contrast, our DPG shows a more distinct separation, with an average gap of 0.14 (twice as large as Coda's gap) between inter- and intra-class distances in testing data. This indicates better generalizability with a dynamic prompt generation mechanism for unseen classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.79, + 0.655 + ], + "angle": 0, + "content": "Moreover, accuracy results with Recall@1 in Table 2 and Table 4 reveal that our DPG significantly outperforms the best baseline (Coda), with an average of \\(9.1\\%\\) boost in 100 CL stages and \\(3.2\\%\\) boost in 10 CL stages, respectively. This highlights the superiority of Dynamic Prompting Generation design, particularly in practical scenarios involving a large number of stages. In the next Section 4.4.1 and Section 4.4.2, we investigate the importance of our design components of stage tokens and mapping function in the DPG network." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.682, + 0.506, + 0.699 + ], + "angle": 0, + "content": "4.4.1 Effectiveness of Stage Tokens" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In our DPaRL design with Dynamic Prompt Generation (DPG), we introduce stage tokens to retain information on old training stages, and we impose a max limit on the number of stage tokens (\\( M_{stg\\_tok} = 5 \\)) to control storage and computation costs as the number of stages scales up, e.g., \\( N_{stg} = 100 \\). Hence, when the stage number goes beyond the max limit \\( M_{stg\\_tok} \\), we replace some of previous stage tokens. Our hypothesis is that the earlier stage information learned from previous stage tokens can propagate into the later ones. In this section, we conduct ablations in a practical CL setting with a large number of stages (100 stages) to study the impact of stage tokens." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Impact of stage tokens with a large number of CL stages, \\( N_{stg} = 100 \\). Recall@1 results on In-Shop, SOP, and iNat2018 datasets are reported. It's worth noting that the Cars dataset is relatively small, preventing its division into 100 stages. DPG is a naive version of our DPaRL method by freezing the backbone weights." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.213, + 0.734, + 0.33 + ], + "angle": 0, + "content": "
Method# Stage TokenOrderIn-ShopSOPiNatAverage
L2P [39]N/AN/A48.3970.3976.8265.20
Dual [38]N/AN/A52.5073.7676.9167.72
Coda [32]N/AN/A52.9673.9276.9867.95
DPGNoneN/A71.1880.0477.5776.26 (↑8.31)
DPG5FIFO (default)72.0580.8178.2977.05 (↑9.10)
DPG5FILO71.7680.6078.1276.83 (↑8.88)
DPG5Random72.0080.8378.2077.01 (↑9.06)
DPG10FIFO72.1580.8178.2877.08 (↑9.13)
DPG100full-size baseline72.2080.8478.3777.14 (↑9.19)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.784, + 0.417 + ], + "angle": 0, + "content": "With and without stage token. From Table 2, DPG with 5 stage tokens achieves a noticeable accuracy improvement over DPG without stage tokens, 77.05 vs. 76.26 on average, indicating the importance of stage token for retaining stage-wise class information in old training stages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Impact of the order. We conduct ablations on three different orders: first-in-first-out (FIFO), first-in-last-out (FILO), and randomly replacing one stage token (Random). Table 2 reveals that FIFO performs the best on average. FILO achieves slightly lower accuracy. Nevertheless, the Random method achieves performance very close to the best, indicating the robustness of our DPG design to different orders of stage tokens. We hypothesize this due to stage information learned from previous ones effectively propagating to the new stage tokens, regardless of the order or the location where the new stage token is inserted." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.539, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Selection of maximum limit \\( M_{stg\\_tok} \\). When the stage number is large (\\( N_{stg} = 100 \\)), keeping a maximum token number of 5 achieves comparable performances (77.05 vs. 77.14 on average) to that of the full token number baseline with \\( M_{stg\\_tok} = 100 \\). This signifies a highly effective selection with a constant number of stage tokens to achieve high accuracy, as well as good scalability with respect to the stage number. Hence, we use \\( M_{stg\\_tok} = 5 \\) throughout this paper." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.549, + 0.667 + ], + "angle": 0, + "content": "4.4.2 Effectiveness of Mapping Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.749 + ], + "angle": 0, + "content": "In our DPaRL framework (Figure 3), as elaborated in Section 3.3, we introduce a mapping function to match the size of [CLS] token and prompt tokens. We apply rank-\\(R\\) constraint to avoid overfitting and reduce the memory consumption from the fully-connected layer inside the mapping function. To measure the sensitivity of parameter \\(R\\), we conduct experiments with varying \\(R\\) from 64 to 768." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.81 + ], + "angle": 0, + "content": "The results for both In-Shop and iNat in Table 3 indicate the existence of an optimal rank with \\( R_{opt} = 256 \\) that strikes a balance between accuracy and the number of parameters, saving \\( 2/3 \\) of the total parameters compared to the full rank case. This default rank \\( R_{opt} = 256 \\) is subsequently used for all experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Note that when the rank reaches 768, representing the full rank of the mapping function essentially serving as a fully-connected layer, a significant per" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Impact of the rank of weight matrix inside the mapping function on our DPG design. Recall@1 results on In-Shop and iNat2018 with 10 CL stages are reported." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.185, + 0.689, + 0.244 + ], + "angle": 0, + "content": "
Rank64128256384512768
In-Shop82.3383.1784.0983.8682.9871.40
iNat201879.2179.8580.0279.7479.6378.98
#Parameters2.0M3.9M7.9M12.0M16.1M23.5M
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.346 + ], + "angle": 0, + "content": "formance drop of \\(12.69\\%\\) is observed for the In-Shop dataset in Table 3. This suggests that an overparameterized expanding layer results in a severe accuracy degradation for a highly compressed [CLS] token, indicating the importance of our design. Moreover, our design has a robust rank range from 128 to 384, which closely matches the accuracy performance of the peak result." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.366, + 0.774, + 0.382 + ], + "angle": 0, + "content": "4.5 Effectiveness of Dynamic Prompt and Representation Learner" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.539 + ], + "angle": 0, + "content": "Previous PCL methods [32,38,39] freeze backbone weights of the discriminative model to reduce catastrophic forgetting. However, this approach could limit accuracy performance as it doesn't provide the degree of freedom to update the backbone, particularly in open-world setting where data distribution and domain shifts naturally exist. In this section, we explore effective ways to update backbone weights jointly with our dynamic prompt learning. We explore 3 approaches here: 1) fine-tuning \\((FT)\\) the entire backbone weights, 2) parameter-efficient fine-tuning (PEFT) with AdaptFromer [8] (Adapter) or Low-Rank Adaptation [16] weights on the backbone \\((LoRA)\\), and 3) joint prompt generation and representation learning on the backbone weights (Coda and \\(DPaRL\\), \\(w / X\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Table 4 shows that vanilla \\(FT\\) or fine-tuning with PCL methods not only leads to lower accuracy but also suffers from higher catastrophic forgetting compared to DPG. While, LoRA gets higher accuracy \\((+10.4\\%)\\) on average) and reduces forgetting \\((-2.3\\%)\\) on average) compared to vanilla \\(FT\\), as well as outperforms Adapter, indicating a more effective way to update the backbone. Still, LoRA is less effective than DPG, with an average accuracy gap of \\(1.7\\%\\) and a forgetting gap of \\(0.6\\%\\), indicating the superiority of our DPG design for open-world CL problems. Lastly, our joint Dynamic Prompt and Representation Learner (DPaRL) with LoRA improves average accuracy by \\(3.2\\%\\) and reduces forgetting by \\(0.5\\%\\) compared to LoRA-only baseline. It also boosts accuracy by \\(1.4\\%\\) on average compared to DPG method without updating backbone weights at a negligible cost of 0.09 forgetting raises, indicating the effectiveness of our DPaRL design to maximizing accuracy without causing catastrophic forgetting issues. Therefore, unless stated otherwise, we adopt joint learning with dynamic prompt generation with LoRA as our default DPaRL method in this paper." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.731, + 0.803 + ], + "angle": 0, + "content": "4.6 Performance on Closed-World Setting in Other Domains" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we assess the effectiveness of our DPaRL method in domains other than the established image retrieval benchmarks such as medical image" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 4: Accuracy performance \\((R_N, \\uparrow)\\) and forgetting scores \\((F_N, \\downarrow)\\) in 10 CL stages with different fine-tuning methods: fine-tuning (FT) the entire backbone weights, PEFT with prompt tuning (Coda and our DPG), Adapter [8], and LoRA [16]." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.199, + 0.78, + 0.326 + ], + "angle": 0, + "content": "
DatasetCarsIn-ShopSOPiNat2018Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
Coda [32], freeze(frz) backbone65.230.3678.610.0081.620.0178.970.0176.110.10
DPG (Ours), frz backbone70.620.0084.090.0082.690.0080.020.0179.360.00
FineTune [32]57.575.0482.230.5473.452.2355.693.8467.242.91
Adapter [8]64.592.4783.940.0082.180.0677.270.3176.990.71
LoRA [16]67.770.9983.290.4081.970.2777.470.7977.630.61
Coda, w/ FT50.0312.7982.290.4574.033.2957.424.3965.945.23
Coda, w/ Adapter64.390.8683.560.4581.980.0578.560.0677.120.36
Coda, w/ LoRA69.061.2382.790.3382.510.2579.120.3478.370.54
DPaRL, w/ FT65.004.6182.360.3374.242.7156.734.1269.582.94
DPaRL, w/ Adapter68.471.0085.720.0182.570.0178.040.0378.700.26
DPaRL, w/ LoRA (default)73.220.0386.280.0283.690.0780.020.2280.800.09
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.472 + ], + "angle": 0, + "content": "domain. However, there are no available open-world benchmarks so we follow the standard closed-world setting in Figure 1 (b). We conduct experiments on MedMNIST [40], which consists of approximately 700K biomedical images. It includes 7 datasets, each with \\(7\\sim 11\\) classes. We evenly split the classes into 3 CL stages and report the average top-1 accuracy over each stage. Results in Table 5 demonstrate that DPaRL generalizes well to the medical domain, outperforming other methods with an average improvement of \\(6.50\\%\\). For additional tasks in other domains, please refer to the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.497, + 0.785, + 0.527 + ], + "angle": 0, + "content": "Table 5: Top-1 (%) on medical data MedMNIST [40]. Our DPaRL surpasses other PCL methods by \\(6.50\\%\\) on average." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.538, + 0.754, + 0.596 + ], + "angle": 0, + "content": "
MethodsBloodDermaOrganAOrganCOrganSPathTissueAverage
L2P56.0334.0061.1252.8746.2556.3941.2949.71
Dual66.2833.1665.4257.7948.2260.6940.3453.13
Coda57.3733.7265.4165.9846.8279.8532.3054.49
DPaRL (Ours)70.4938.7170.3971.2951.2381.4443.3860.99(↑6.50)
" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.644, + 0.358, + 0.66 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.674, + 0.786, + 0.84 + ], + "angle": 0, + "content": "In this paper, we introduce a practical setting for continual visual representation learning in the open-world. Existing continual learning methods fail to generalize well in this proposed setting. To mitigate this challenge, we propose Dynamic Prompt and Representation Learner (DPaRL), to enhance the adaptability and performance in continual learning for open-world visual recognition. Our method consistently demonstrates superior performance across various open-world image retrieval datasets and settings, by dynamically generating prompts from a deep neural network coupled with a dedicated mapping function, and effectively updating the discriminative representation backbone. Notably, it overcomes the intrinsic constraints found in static prompt pool designs, thereby expanding the limits of model generalization in open-world scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.223 + ], + "angle": 0, + "content": "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.264 + ], + "angle": 0, + "content": "2. An, X., Deng, J., Yang, K., Li, J., Feng, Z., Guo, J., Yang, J., Liu, T.: Unicom: Universal and compact representation learning for image retrieval. arXiv preprint arXiv:2304.05884 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.334 + ], + "angle": 0, + "content": "4. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Cao, T., Wang, Y., Xing, Y., Xiao, T., He, T., Zhang, Z., Zhou, H., Tighe, J.: Pss: Progressive sample selection for open-world visual representation learning. In: European Conference on Computer Vision. pp. 278-294. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.447 + ], + "angle": 0, + "content": "7. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.489 + ], + "angle": 0, + "content": "8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.532 + ], + "angle": 0, + "content": "9. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.533, + 0.785, + 0.587 + ], + "angle": 0, + "content": "10. De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence 44(7), 3366-3385 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.588, + 0.785, + 0.629 + ], + "angle": 0, + "content": "11. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.63, + 0.785, + 0.685 + ], + "angle": 0, + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.687, + 0.785, + 0.728 + ], + "angle": 0, + "content": "13. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "14. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.785, + 0.799 + ], + "angle": 0, + "content": "15. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9769-9776. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "Y. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Kim, S., Kim, D., Cho, M., Kwak, S.: Proxy anchor loss for deep metric learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3238-3247 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "20. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: International Conference on Machine Learning. pp. 3925-3934. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "21. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "22. Liu, Z., Luo, P., Qiu, S., Wang, X., Tang, X.: Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1096-1104 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "23. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "24. Movshovitz-Attias, Y., Toshev, A., Leung, T.K., Ioffe, S., Singh, S.: No fuss distance metric learning using proxies. In: Proceedings of the IEEE international conference on computer vision. pp. 360-368 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "25. Oh Song, H., Xiang, Y., Jegelka, S., Savarese, S.: Deep metric learning via lifted structured feature embedding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4004-4012 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "26. Rao, D., Visin, F., Rusu, A., Pascanu, R., Teh, Y.W., Hadsell, R.: Continual unsupervised representation learning. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "27. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "28. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.647 + ], + "angle": 0, + "content": "29. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "30. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: International conference on machine learning. pp. 4548-4557. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "31. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "33. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.438, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "DPaRL for Open-World Continual Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "34. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV. pp. 1706-1716 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.232 + ], + "angle": 0, + "content": "35. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.313 + ], + "angle": 0, + "content": "37. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.37 + ], + "angle": 0, + "content": "38. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.424 + ], + "angle": 0, + "content": "39. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.467 + ], + "angle": 0, + "content": "40. Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., Ni, B.: Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data 10(1), 41 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.522 + ], + "angle": 0, + "content": "41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.55 + ], + "angle": 0, + "content": "42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.578 + ], + "angle": 0, + "content": "43. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.619 + ], + "angle": 0, + "content": "44. Zhao, B., Tang, S., Chen, D., Bilen, H., Zhao, R.: Continual representation learning for biometric identification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 1198-1208 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "45. Zhou, D.W., Ye, H.J., Zhan, D.C., Liu, Z.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.66 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_origin.pdf b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..707a41db8786ec43ea1ff2f1e4cce5b924d1f078 --- /dev/null +++ b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/5cd14218-379f-4209-a524-2be87a9eba2f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4456668a2c8c11a3745040bc1623696e37063e074b513d65a4c45eeec9616cfe +size 661476 diff --git a/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/full.md b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..94e34866cceb547957358af40d177d69b81faafc --- /dev/null +++ b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/full.md @@ -0,0 +1,279 @@ +# Open-World Dynamic Prompt and Continual Visual Representation Learning + +Youngeun Kim $^{1,*†}$ , Jun Fang $^{2,*‡}$ , Qin Zhang $^{2}$ , Zhaowei Cai $^{3}$ , Yantao Shen $^{2}$ , Rahul Duggal $^{2}$ , Dripta S. Raychaudhuri $^{2}$ , Zhuowen Tu $^{2}$ , Yifan Xing $^{2}$ , Onkar Dabeer $^{2}$ + +Yale University +AWS AI Labs +3 Amazon AGI + +youngun.kim@yale.edu {junfa, qzaamz, zhaoweic, yantaos, dugrahul, driptarc, ztu, yifax, onkardab}@amazon.com + +Abstract. The open world is inherently dynamic, characterized by ever-evolving concepts and distributions. Continual learning (CL) in this dynamic open-world environment presents a significant challenge in effectively generalizing to unseen test-time classes. To address this challenge, we introduce a new practical CL setting tailored for open-world visual representation learning. In this setting, subsequent data streams systematically introduce novel classes that are disjoint from those seen in previous training phases, while also remaining distinct from the unseen test classes. In response, we present Dynamic Prompt and Representation Learner (DPaRL), a simple yet effective Prompt-based CL (PCL) method. Our DPaRL learns to generate dynamic prompts for inference, as opposed to relying on a static prompt pool in previous PCL methods. In addition, DPaRL jointly learns dynamic prompt generation and discriminative representation at each training stage whereas prior PCL methods only refine the prompt learning throughout the process. Our experimental results demonstrate the superiority of our approach, surpassing state-of-the-art methods on well-established open-world image retrieval benchmarks by an average of $4.7\%$ improvement in Recall@1 performance. + +Keywords: Dynamic Prompt Generation $\cdot$ Continual Learning $\cdot$ Open-World Visual Representation Learning + +# 1 Introduction + +Continual learning (CL) without catastrophic forgetting is a challenging task in practice as retraining with old data is prohibitively expensive when data volume grows [10,14,23]. It is particularly challenging for real-world applications in the open-world [2,5], where new concepts keep emerging and distribution shifts are + +![](images/c4956c950113a6c80b7d90ac9556c93ec61c1c08e9318344612d80e0fab2e977.jpg) +(a) + +![](images/fe91abc385f9ade7dd745fc2cc5a86198522244945daaa10c6dc39d75cb5caa7.jpg) +(b) + +![](images/3f30019b810bc0377ab9000dc55444fadbc6d010816e1508fdb87ac41a36b015.jpg) +(c) +Fig. 1: Illustration of the distinctions between our problem setting and traditional settings. In our work, we aim to address the problem where training splits have no class overlaps. (a) Closed-world setting: both training and testing classes are identical. (b) Continual learning (CL) in closed-world setting: training classes are split into multiple divisions and are introduced through various CL stages. (c) Open-world setting: training and testing classes remain separate. The aim is to learn a robust representation from training classes that generalize to unseen classes. (d) Open-World Continual Representation Learning (Our Problem Setting): continual learning tailored for an open-world scenario. We sequentially introduce training classes over multiple CL stages, ensuring they remain distinct from test-time unseen classes. + +![](images/92d55fa6a85a963a9e0bd211454df4ae89c774157d98a3df799e3e34df60b55a.jpg) +(d) + +inevitable. Recent studies on Prompt-based Continual Learning (PCL) [32,38, 39] demonstrated success in mitigating catastrophic forgetting without access to any past data. However, they heavily focus on the closed-world setting where test-time categories come from those encountered in various training stages. + +Whereas in practice, the world is dynamic and concepts evolve over time and new categories are often introduced after a model is trained. Consider, for instance, an image retrieval system for clothing brands in online e-commerce. During training, one would gather images of all existing clothing brands available to train a model. However, after deployment, the system encounters new brands. The new data collected for system upgrade, constituting a new training stage, will likely include brands that are disjoint to those in older stages. Furthermore, the system will constantly face new brands at test-time that are disjoint to any of those encountered in the previous training stages and is expected to generalize well to those new categories. + +Thus, to comprehensively evaluate the ability of a model in retaining semantic information from continuous streams of data for generalization to new concepts at test-time for practical applications, we propose a new continual open-world representation learning setting, as illustrated in Figure 1(d). In this setting, different data streams comprise distinct classes, and test-time classes are disjoint to those encountered in all training phases. Our setting is notably more challenging than the previous closed-world CL scenario. It closely mirrors real-world demands on recognition systems, where, during any evaluation stage, the test set consists of unseen entities to the training process up to that point. This setup reflects the need for practical recognition systems to encode general semantic information from training-time classes and apply this knowledge to new, unseen concepts without access to older training data during system updates. + +We start with evaluating state-of-the-art PCL methods [32, 38, 39] in our proposed setting and find that they fail to generalize with the absence of past data. For example, in our experiments, the gap between the paragon of standard training with all data in one phase and the best PCL baseline (Coda) shows up to $20\%$ gap in Recall@1 (Table 1), significantly larger than the $5\%$ gap in the closed-world setting, which suggests that prior PCL methods encounter difficulties and are less effective in the open-world setting. + +To tackle this challenging yet practical open-world CL setting, we introduce a simple yet effective method called Dynamic Prompt and Representation Learner (DPaRL). Unlike previous prompt pool based PCL methods [32, 38, 39] that depend on a static prompt pool (Figure 2(a)) for test-time predictions. Notably, DPaRL trains an innovative dynamic prompt generation network jointly with the discriminative representation backbone model (Figure 2(b)), departing from the prompt-only learning paradigm in prior PCL methods. + +Our key insight lies in that simply combining the fixed prompts learnt over training classes does not generalize well to unseen test-time classes in the practical open-world setting. On the other hand, the on-the-fly dynamically generated prompts in our method exhibit a stronger capability of capturing the diverse semantics spread across the numerous divisions of classes observed during the different training stages of continual learning, improving generalization to test-time unseen classes. This is achieved through the proposed dynamic prompt generation network, which possesses stronger representation power with a learnable mapping function compared to the traditional approach of combining multiple prompts in a static prompt pool. Additionally, unlike prior PCL methods, DPaRL jointly updates the dynamic prompt generation process and the discriminative representation backbone weights, facilitating maximized interaction and comparison between old concepts and newer ones to encapsulate diverse semantics that aid generalization to unseen test-time classes. + +Consequently, over the four open-world image retrieval benchmarks on cars, online catalogs, products, and species images where the task is to search for the most similar images belonging to the same class as the query, via a distance measurement in an embedding space, DPaRL achieves significant accuracy boosts. It improves the average Recall@1 performance by absolute $4.7\%$ and $9.1\%$ compared to state-of-the-art continual learning methods, from $76.1\%$ to $80.8\%$ over 10 CL stages and from $68.0\%$ to $77.1\%$ over 100 CL stages, respectively. + +In summary, our contributions are three folds: + +1. We establish a new practical setting on continual visual representation learning in the open-world. +2. We introduce a simple yet powerful method, Dynamic Prompt and Representation Learner (DPaRL), which dynamically generates prompts while effectively updating the discriminative representation backbone. This enhancement improves generalization for unseen open-world classes at test time. +3. We outperform state-of-the-art continual learning methods in the proposed practical setting, both rehearsal-free and rehearsal-based. + +![](images/102be1890616f633366a02d43040179e7334815d274287997908dc1476d130bb.jpg) +(a) Previous method + +![](images/740775948f3f6a33e91d0b5fd35cc9ddb232c9a1a8d5101dfb2e98a836469bb1.jpg) +(b) Our work +Fig. 2: Prior PCL methods combine prompts from a static prompt pool trained on the training class distribution, leading to a loss of generalization capability when facing unseen classes during test time. Our work introduces a Dynamic Prompt Generation network that generates dynamic prompts on the fly by integrating a given image with stage tokens, followed by a specialized mapping function and adjustable discriminative representation backbone weights, providing generalizable prompts for unseen testing classes, distinguishing it from prior PCL methods. + +# 2 Related Work + +Continual learning. Continual learning (CL) has been recognized for decades as a crucial research domain, focusing on the ongoing adaptation of models over time. Early works primarily centered around regularization-based methods in an effort to address catastrophic forgetting [1,21,43] by introducing regularization between its current and previously learned parameters. However, they often underperform when faced with complex datasets. + +A separate line of work, architecture-based approach, involves expanding the network architecture in response to new learning phases [20,26,29,30,42]. While these methods generally outperform regularization-based counterparts, they do so at the cost of introducing additional parameters. Recent advancements include rehearsal-based methods that incorporate a memory component for storing past data [4,6,7,15,28]. This archived data aids in training newer stages, typically yielding superior performance compared to other CL techniques. Yet, this data retention can introduce privacy concerns and significant memory overhead. + +Given these constraints, rehearsal-free strategies are gaining attention in the community. These aim to mitigate catastrophic forgetting without relying on historical data. A prominent method within this category uses model inversion to produce rehearsal images [9,13,31,41]. However, model inversion is computationally intensive and time-consuming. Additionally, these methods lag significantly in performance when compared to their rehearsal-based counterparts. + +Recent developments have witnessed the rise of prompt-based approaches [27, 32, 34, 37-39, 45]. These strategies offer robust protection against catastrophic forgetting by learning a small set of parameters, or prompts, instead of directly training all parameters. For instance, L2P [39] employs a pool of prompts, selecting suitable ones for insertion based on input data clustering. Building on this concept, DualPrompt [38] presents general prompts designed to encapsulate shared knowledge across various learning stages. Advancing this further, + +the state-of-the-art method CodaPrompt [32] facilitates end-to-end training by incorporating an attention mechanism for prompt combination within a pool. However, during inference, the above works rely on the combination of prompts from a static pool learned from training to feed into a frozen backbone model, which has limitations in cases where testing data has distribution shifts or disjoint classes from the training data. This motivates us to design a Dynamic Prompt and Representation Learner (DPaRL) to address these drawbacks. + +Open-World Visual Representation Learning. In open-world visual representation learning, the model learns discriminative representations that align distances between representations with their semantic similarities, and uses these knowledge for testing on unseen classes. Previous works in this domain have primarily focused on designing loss functions [11, 17, 24] for enhanced accuracy. Recently, several studies [2, 12] have demonstrated the effectiveness of utilizing pre-trained visual foundation models and fine-tuning them to achieve significant performance improvements. While open-world visual representation learning has been under investigation for an extended period, there has been limited exploration within the continual learning (CL) framework for this task. [44] investigated biometric datasets using a regularization-based CL approach. However, the investigation into state-of-the-art prompt-based CL methods remains missing. Furthermore, as their studies primarily centered on biometric data, there is a pressing need to explore established continual learning benchmarks specific to open-world visual representation learning in more natural settings. + +# 3 Methodology + +In this section, we introduce our open-world Continual Learning (CL) setting and motivate our Dynamic Prompt and Representation Learner (DPaRL). + +# 3.1 Problem Setting + +Two fundamental visual representation learning settings are the closed-world and open-world paradigms. In a closed-world setting (as depicted in Figure 1(a)), the training and testing data classes are entirely identical. Conversely, the open-world setting (illustrated in Figure 1(c)) presents a more practical scenario. Here, training and testing classes are entirely distinct, thus requiring the model to learn representations that generalize to unseen concepts. + +When faced with continuous streams of data for visual recognition system updates, built upon the closed-world paradigm, Figure 1(b) showcases the well-known continual learning setting for image recognition. In this work, we introduce a new practical setting: Open-World Continual Representation Learning as shown in Figure 1(d). Instead of testing on seen classes of the various training stages thus far, the testing classes are entirely disjoint to those faced in different training stages, necessitating the model to not only learn continually but also generalize to new concepts, making it a more dynamic and challenging setting. + +![](images/9bfe3eb4e63ee6f78904d8649b61ea55486f17d7086fdc13b3687209d646347c.jpg) +Fig. 3: Illustration of our proposed method: Dynamic Prompt and Representation Leaner (DPaRL). We dynamically generate prompt tokens on the fly from Dynamic Prompt Generation (DPG) networks by integrating information from stage tokens and image tokens. The [CLS] token from DPG is converted to prompt tokens via a linear mapping function. The generated prompt tokens are added to the backbone ViT and trained with a loss function. The learnable parameters include the current stage token, weights in mapping function, weights in backbone, and weights in loss function. + +# 3.2 Prompt-based Continual Learning Paradigm + +Prompt-based Continual Learning (PCL) methods [32,38,39] use a pre-trained Vision Transformer (ViT) as a discriminative backbone for closed-world image classification, as shown in the Figure 2(a). These methods create a prompt pool with multiple prompt tokens, updating only the learnable parameters in this pool during training. In inference, the learned prompt pool is static, and PCL methods select tokens from this pool to feed into multiple ViT backbone layers for predictions. Our method adopts this PCL paradigm but adapts to the open-world representation learning setting with a prototypical metric learning loss. Additionally, we introduce a dynamic prompt generation (DPG) network to replace the static prompt pool for effective joint learning with the discriminative representation backbone model, as detailed below. + +# 3.3 Dynamic Prompt and Representation Leaner (DPaRL) + +Overall Pipeline. In the open-world setting where training and testing classes are disjoint, existing prompt pool designs in PCL [32,38,39] exhibit limitations, i.e., limited separation between intra- and inter- testing class distance distributions as shown in Figure 4 and large performance gap from paragon in Table 1. To mitigate, we first propose dynamically generating prompts by blending the given image information and continual learning stage's information through a dynamic prompt generation (DPG) network. It aims to provide better representation power compared to the prompt pool design, which combines a fixed prompt pool trained from the training class distribution. Moreover, we enhance the capability of the discriminative representation backbone model to more effectively generalize to the open-world concepts through a joint dynamic prompt and representation learning paradigm. We refer to Figure 2, 3 for detailed design differences between our approach and prior prompt pool-based methods. + +Dynamic Prompt Generation Network. In order to retain information from each CL stage of our Dynamic Prompt Generation (DPG) process, we introduce stage tokens, denoted as $S$ . During training stage $t$ , we train a stage token $S_{t}$ while freezing the previous stage tokens $S_{t - (q - 1)} \sim S_{t - 1}$ in a first-in-first-out (FIFO) queue manner, where $q$ is the max queue size. Guided by the ablation study detailed in Section 4.4, we typically set $q = 5$ using a FIFO order to aim for both high accuracy and scalability. This ensures that the knowledge from the previous stages remains unchanged, and the total number of stage tokens limits by the queue size without linearly scaling with stage numbers. Within the deep neural network such as ViT, the self-attention block facilitates the integration of information between the stage-wise tokens and the instance-wise image tokens. + +As illustrated in Figure 3, the DPG produces a [CLS] token to dynamically acquire task-specific high-level information for prior knowledge of class-level features. However, it differs in size from the prompt tokens required by the discriminative backbone model (right side of Figure 3). To address this, we introduce a mapping function between the [CLS] token and the prompt tokens. + +A straightforward approach is to employ a single linear layer for dimensionality transformation. However, such a method introduces an excessive number of additional parameters, and the resulting over-parameterization can lead to overfitting of the highly compressed [CLS] token information. To tackle this issue, we impose a constraint on the weight parameter $W = AB^{T} \in \mathbb{R}^{C_{in} \times C_{out}}$ , where $A \in \mathbb{R}^{C_{in} \times R}$ , $B \in \mathbb{R}^{C_{out} \times R}$ , ensuring a maximum rank of $R < \min(C_{in}, C_{out})$ . This constraint is inspired by the success of the LoRA technique [16]. The experimental results in Section 4.4 illustrate the critical importance of the low-rank mapping function design for our DPG module. Moreover, we apply Dropout [33] and LayerNorm [3] to this low-rank linear mapping function + +$$ +M a p p i n g (x) = L a y e r N o r m \left(L o w R a n k L i n e a r (D r o p O u t (x))\right), \tag {1} +$$ + +which further helps to avoid overfitting and stabilize training. + +With the help of this specialized mapping function along with a pre-trained neural network, the prompt $P$ is dynamically obtained from the input image $I$ and stage tokens $S_{t,q} \coloneqq S_{t - (q - 1)} \sim S_t$ : + +$$ +P = \text {M a p p i n g} (\text {N e t w o r k} ([ S _ {t, q}; I ]). \tag {2} +$$ + +We reshape $P$ to a size of $N_{p} \times C \times L$ , where $N_{p}$ is the number of prompts, $C$ is the channel dimension, and $L$ indicates the number of layers where prompts are applied to the backbone model. The generated prompts are inserted across the first $L$ layers in the ViT backbone. In our DPG approach to prompting, we follow the technical foundations of prompting from the prior state-of-the-art PCL methods [32,38]. This is because we want to make fair comparison with the previous work, and our primary focus is on the formation of prompt. The prompt for the $l$ -th layer, denoted as $P_{l} \in \mathbb{R}^{N \times C}$ , is partitioned into $\{P_{l,k}, P_{l,v}\} \in \mathbb{R}^{\frac{N}{2} \times C}$ and is prefixed to the input token embedding for both key and value in the attention mechanism [36]. This can be formulated as: + +$$ +h _ {i} = \text {A t t e n t i o n} \left(X _ {l} W _ {q} ^ {i}, [ P _ {k}; X ] W _ {k} ^ {i}, [ P _ {v}; X ] W _ {v} ^ {i}\right). \tag {3} +$$ + +Here, we omit layer index $l$ for brevity, and $i$ is head index in multi-head self-attention of transformer architecture. + +Joint Dynamic Prompt and Representation Learning. Differing from prior PCL methods [32,38,39] that freeze the backbone model (on the right side of the pipeline in Figure 2(a)) during training, we propose our Dynamic Prompt and Representation Learner (DPaRL), to jointly learn the dynamic prompt generation with the discriminative representation learning backbone (on the right side of Figure 2(b) and Figure 3). This approach aims to maximize the capability of entire pipeline to integrate old-stage concepts and new ones to encapsulate diverse semantics that aid generalization to open-world unseen classes and unknown domain shifts. Leveraging parameter-efficient fine-tuning techniques, our DPaRL successfully maximizes accuracy performance and minimizes catastrophic forgetting simultaneously. The overall framework is depicted in Figure 3. Note that the DPG encoder weights (on the left side of the pipeline) for prompt generation are frozen in our method, while the learnable parameters include stage tokens, weights in the mapping function and representation loss function, as well as the discriminative representation backbone weights. + +# 4 Experiments + +In this section, we empirically evaluate the effectiveness of our Dynamic Prompt and Representation Learner (DPaRL) in the established challenging continual open-world visual representation learning. + +# 4.1 Datasets & Evaluation Metric + +Datasets. We conducted our training and testing across 4 prominent open-world image retrieval benchmarks: Cars [19], In-Shop [22], SOP [25] and iNat2018 [35]. We evaluate baselines and our proposed method following our practical continual open-world visual representation learning setting. In this dynamic scenario, classes are introduced incrementally at each learning stage, and the testing phase encompasses classes that have not been encountered during training. Our experiments consist of 10 or 100 stages, with the total training classes approximately distributed uniformly across these stages. The detailed data information including number of classes and samples are provided in the Supplementary Material. For inference, we use the entire testing dataset across all CL stages since there is no overlap between training and testing classes. Thus, our objective is to learn class relations while maintaining generalization across CL stages. + +Evaluation Metric. To evaluate model performance, we compute the Recall@1 metric using the [CLS] token at the last layer of the discriminative representation backbone network. Given our goal to optimize the model's adaptability throughout the CL phases, we report average Recall@1 across all CL stage. Let $R_{n}$ be the Recall@1 performance at stage $n$ , then we report $R_{N} = \frac{1}{N_{s}}\sum_{n = 1}^{N_{s}}R_{n}$ where $N_{s}$ + +is the total number of stages. We also report the forgetting score $F_{N}$ by following [32,38] to measure the performance drop on the same unseen testing dataset across different training stages, that is, $F_{N} = \frac{1}{N_{s} - 1}\sum_{n = 2}^{N_{s}}\max_{m\in \{1,\dots,n - 1\}}(0,R_{m} - R_{n})$ . We stress that $R_{N}$ holds greater importance as it captures both learning capacity and forgetting, while $F_{N}$ merely offers supplementary context. + +# 4.2 Baselines and Implementations + +Baselines. We validate 3 types of representative baselines: + +-Rehearsal-based:Experience Replay (ER) [7], +- Regularization-based: Learning without Forgetting (LwF) [21] +- Rehearsal-free prompt-based: Learning to Prompt (L2P) [39], DualPrompt [38], and CodaPrompt [32]. + +**Implementations Details.** Consistently across all methods, we adhere to the training setups by following [32]. Specifically, we utilize the Adam optimizer [18] with a batch size of 128 to train for 20 epochs to ensure comprehensive model convergence. The learning rate is cosine decaying with an initial value of 0.001. + +We utilize ArcFace [11] as the training loss function as it is a prevalent choice in open-world representation learning to achieve state-of-the-art performance [2]. For new-stage training, the prototype weights in the loss function of old training stages are appended to the new training stage as initialization. All prototype weights in the loss function, including those corresponding to the classes for old and new stages, are trainable during each training stage. + +For established PCL techniques, we maintain the prompt lengths and locations for L2P [39], DualPrompt [38], and CodaPrompt [32] as suggested in these works. For our DPaRL method, in each CL stage, we incorporate a learnable stage token with the first-in-first-out (FIFO) order and a max size of 5 tokens based on insights from Section 4.4. We apply our specialized mapping function in the DPG module to avoid over-fitting and stabilize training. We adopt LoRA [16] as the PEFT method to update backbone weights as guided by Section 4.5. + +# 4.3 Performance Comparison with Prior Arts + +Table 1 shows a comprehensive performance comparison with various CL methods across all four open-world image retrieval tasks. We adopt the official codebase of CodaPrompt [32] for the implementation of PCL baselines including L2P [39], DualPrompt [38], and CodaPrompt [32]. Additionally, we report the performance of representative rehearsal-based and regularization-based approaches, i.e., ER [7] and LWF [21], respectively, for a more comprehensive comparison. Note that, for a fair comparison, all methods use the pre-trained ViT-Base on ImageNet-21k as the backbone model, as shown on the right side of Figure 3. + +From the results in the Table 1, we observe that the performance trend of PCL-based methods (L2P, DualPrompt, CodaPrompt, our DPaRL) is similar to what was observed in the closed-world image recognition task [32], where PCL + +Table 1: Performance on 4 open-world image retrieval benchmarks with 10 CL stages. We report the averaged $Recall@1$ ( $R_N$ ) and $Forgetting\ Score$ ( $F_N$ ) across all continual learning stages, the results are averaged over 3 runs. We present the LowerBound with the zero-shot performance of ImageNet-21k pre-trained ViT-B/16 backbone. UpperBound denotes the upper bound performance when trained on the entire dataset at a single stage. All PCL methods, including L2P, DualPrompt, CodaPrompt, and our DPaRL, utilize the same pre-trained encoder, ViT-B/16 on ImageNet-21k, to generate the prompt tokens. The numbers in brackets show the $Recall@1$ improvement of our DPaRL over the previous best baselines, with the previous best baseline underlined. + +
DatasetCars [19]In-Shop [22]SOP [25]iNat2018 [35]Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
LowerBound53.86-47.16-62.85-71.74-58.90-
UpperBound85.96-89.28-85.21-82.68-85.78-
ER [7]65.380.4982.992.4865.380.5036.9418.5862.675.51
LwF [21]56.532.9181.523.3966.1813.2039.6933.8160.9813.33
L2P [39]56.270.0067.850.0076.880.4078.950.0069.990.10
DualPrompt [38]55.540.0065.560.0077.040.6678.710.0369.210.17
CodaPrompt [32]65.230.3678.610.0081.620.0178.970.0176.110.10
DPaRL (Ours)73.22(↑7.8)0.0386.28(↑3.3)0.0283.69(↑2.1)0.0780.02(↑1.0)0.0380.80(↑4.7)0.09
+ +methods outperform other continual learning approaches by a significant margin on both recall@1 metric $R_N$ and forgetting score $F_N^4$ , aligning with findings from previous research [32, 38, 39]. This suggests that PCL methods are strong baselines for continual learning on both closed-world image recognition tasks and open-world visual representation learning tasks. Notably, our DPaRL approach consistently outperforms other methods across all datasets, resulting in an average Recall@1 boost of $4.7\%$ over the leading previous method, CodaPrompt [32]. This supports our hypothesis about the limitations of the conventional PCL design in open-world scenarios. + +In the following, we provide step-by-step guidance on the design choices of our DPaRL. Additional analysis and results are presented in the supplementary material, covering the superior performance of DPaRL in few-shot settings, effectiveness of DPaRL in closed-world benchmarks, dynamic changes in the accuracy curve across CL stages, and robustness of DPaRL against various DPG encoders. Due to space limitations, we refer to the supplementary material for a more comprehensive exploration of these aspects. + +# 4.4 Effectiveness of Dynamic Prompt Generation + +To solely investigate the effectiveness of our Dynamic Prompt Generation design, we freeze the backbone weights similar to prior PCL methods [32, 38, 39] in our DPaRL and denoted as the DPG method. + +![](images/97169a63cef55964afa7bb192180a1118f496ee640cc082bebb2f7b9dadb904c.jpg) +Fig. 4: Histogram of L2 distance between a pair of samples with embedding features extracted by Coda [32] (SOTA PCL method) and our DPG (the naive version of DPaRL with freezing backbone weights). Left and right figures are distributions of the seen and unseen classes in training and testing Cars [19] dataset, respectively. Our DPG exhibits enhanced separation between inter- and intra-classes, particularly on open-world unseen test classes. + +![](images/139fdfb8ac68d17f6424db357b5d5f8908e68f7b5e8fa91a82ce609ca6625ee3.jpg) + +![](images/55c375d5732f61354c1ea85b0adfd0a16d298eb8bb0acffb4d1e3a1cb3ff2ca3.jpg) + +![](images/593ec7899c6673d945afc746e8519d66b3670489bdee5179b831a66130a51f81.jpg) + +We visualize the distributions of L2 distance between pairwise samples within intra- and inter-classes of CodaPrompt (Coda) [32] and our DPG in Figure 4 to assess generalizability. A good model should exhibit distinguishable distributions between intra- and inter-classes for both closed-world seen classes and open-world unseen classes. Coda, employing a static prompt pool design, achieves a notable separation on closed-world intra- and inter-classes but exhibits a close distribution for open-world intra- and inter-classes, impeding effective generalization to unseen test classes. In contrast, our DPG shows a more distinct separation, with an average gap of 0.14 (twice as large as Coda's gap) between inter- and intra-class distances in testing data. This indicates better generalizability with a dynamic prompt generation mechanism for unseen classes. + +Moreover, accuracy results with Recall@1 in Table 2 and Table 4 reveal that our DPG significantly outperforms the best baseline (Coda), with an average of $9.1\%$ boost in 100 CL stages and $3.2\%$ boost in 10 CL stages, respectively. This highlights the superiority of Dynamic Prompting Generation design, particularly in practical scenarios involving a large number of stages. In the next Section 4.4.1 and Section 4.4.2, we investigate the importance of our design components of stage tokens and mapping function in the DPG network. + +# 4.4.1 Effectiveness of Stage Tokens + +In our DPaRL design with Dynamic Prompt Generation (DPG), we introduce stage tokens to retain information on old training stages, and we impose a max limit on the number of stage tokens ( $M_{stg\_tok} = 5$ ) to control storage and computation costs as the number of stages scales up, e.g., $N_{stg} = 100$ . Hence, when the stage number goes beyond the max limit $M_{stg\_tok}$ , we replace some of previous stage tokens. Our hypothesis is that the earlier stage information learned from previous stage tokens can propagate into the later ones. In this section, we conduct ablations in a practical CL setting with a large number of stages (100 stages) to study the impact of stage tokens. + +Table 2: Impact of stage tokens with a large number of CL stages, $N_{stg} = 100$ . Recall@1 results on In-Shop, SOP, and iNat2018 datasets are reported. It's worth noting that the Cars dataset is relatively small, preventing its division into 100 stages. DPG is a naive version of our DPaRL method by freezing the backbone weights. + +
Method# Stage TokenOrderIn-ShopSOPiNatAverage
L2P [39]N/AN/A48.3970.3976.8265.20
Dual [38]N/AN/A52.5073.7676.9167.72
Coda [32]N/AN/A52.9673.9276.9867.95
DPGNoneN/A71.1880.0477.5776.26 (↑8.31)
DPG5FIFO (default)72.0580.8178.2977.05 (↑9.10)
DPG5FILO71.7680.6078.1276.83 (↑8.88)
DPG5Random72.0080.8378.2077.01 (↑9.06)
DPG10FIFO72.1580.8178.2877.08 (↑9.13)
DPG100full-size baseline72.2080.8478.3777.14 (↑9.19)
+ +With and without stage token. From Table 2, DPG with 5 stage tokens achieves a noticeable accuracy improvement over DPG without stage tokens, 77.05 vs. 76.26 on average, indicating the importance of stage token for retaining stage-wise class information in old training stages. + +Impact of the order. We conduct ablations on three different orders: first-in-first-out (FIFO), first-in-last-out (FILO), and randomly replacing one stage token (Random). Table 2 reveals that FIFO performs the best on average. FILO achieves slightly lower accuracy. Nevertheless, the Random method achieves performance very close to the best, indicating the robustness of our DPG design to different orders of stage tokens. We hypothesize this due to stage information learned from previous ones effectively propagating to the new stage tokens, regardless of the order or the location where the new stage token is inserted. + +Selection of maximum limit $M_{stg\_tok}$ . When the stage number is large ( $N_{stg} = 100$ ), keeping a maximum token number of 5 achieves comparable performances (77.05 vs. 77.14 on average) to that of the full token number baseline with $M_{stg\_tok} = 100$ . This signifies a highly effective selection with a constant number of stage tokens to achieve high accuracy, as well as good scalability with respect to the stage number. Hence, we use $M_{stg\_tok} = 5$ throughout this paper. + +# 4.4.2 Effectiveness of Mapping Function + +In our DPaRL framework (Figure 3), as elaborated in Section 3.3, we introduce a mapping function to match the size of [CLS] token and prompt tokens. We apply rank- $R$ constraint to avoid overfitting and reduce the memory consumption from the fully-connected layer inside the mapping function. To measure the sensitivity of parameter $R$ , we conduct experiments with varying $R$ from 64 to 768. + +The results for both In-Shop and iNat in Table 3 indicate the existence of an optimal rank with $R_{opt} = 256$ that strikes a balance between accuracy and the number of parameters, saving $2/3$ of the total parameters compared to the full rank case. This default rank $R_{opt} = 256$ is subsequently used for all experiments. + +Note that when the rank reaches 768, representing the full rank of the mapping function essentially serving as a fully-connected layer, a significant per + +Table 3: Impact of the rank of weight matrix inside the mapping function on our DPG design. Recall@1 results on In-Shop and iNat2018 with 10 CL stages are reported. + +
Rank64128256384512768
In-Shop82.3383.1784.0983.8682.9871.40
iNat201879.2179.8580.0279.7479.6378.98
#Parameters2.0M3.9M7.9M12.0M16.1M23.5M
+ +formance drop of $12.69\%$ is observed for the In-Shop dataset in Table 3. This suggests that an overparameterized expanding layer results in a severe accuracy degradation for a highly compressed [CLS] token, indicating the importance of our design. Moreover, our design has a robust rank range from 128 to 384, which closely matches the accuracy performance of the peak result. + +# 4.5 Effectiveness of Dynamic Prompt and Representation Learner + +Previous PCL methods [32,38,39] freeze backbone weights of the discriminative model to reduce catastrophic forgetting. However, this approach could limit accuracy performance as it doesn't provide the degree of freedom to update the backbone, particularly in open-world setting where data distribution and domain shifts naturally exist. In this section, we explore effective ways to update backbone weights jointly with our dynamic prompt learning. We explore 3 approaches here: 1) fine-tuning $(FT)$ the entire backbone weights, 2) parameter-efficient fine-tuning (PEFT) with AdaptFromer [8] (Adapter) or Low-Rank Adaptation [16] weights on the backbone $(LoRA)$ , and 3) joint prompt generation and representation learning on the backbone weights (Coda and $DPaRL$ , $w / X$ ). + +Table 4 shows that vanilla $FT$ or fine-tuning with PCL methods not only leads to lower accuracy but also suffers from higher catastrophic forgetting compared to DPG. While, LoRA gets higher accuracy $(+10.4\%)$ on average) and reduces forgetting $(-2.3\%)$ on average) compared to vanilla $FT$ , as well as outperforms Adapter, indicating a more effective way to update the backbone. Still, LoRA is less effective than DPG, with an average accuracy gap of $1.7\%$ and a forgetting gap of $0.6\%$ , indicating the superiority of our DPG design for open-world CL problems. Lastly, our joint Dynamic Prompt and Representation Learner (DPaRL) with LoRA improves average accuracy by $3.2\%$ and reduces forgetting by $0.5\%$ compared to LoRA-only baseline. It also boosts accuracy by $1.4\%$ on average compared to DPG method without updating backbone weights at a negligible cost of 0.09 forgetting raises, indicating the effectiveness of our DPaRL design to maximizing accuracy without causing catastrophic forgetting issues. Therefore, unless stated otherwise, we adopt joint learning with dynamic prompt generation with LoRA as our default DPaRL method in this paper. + +# 4.6 Performance on Closed-World Setting in Other Domains + +In this section, we assess the effectiveness of our DPaRL method in domains other than the established image retrieval benchmarks such as medical image + +Table 4: Accuracy performance $(R_N, \uparrow)$ and forgetting scores $(F_N, \downarrow)$ in 10 CL stages with different fine-tuning methods: fine-tuning (FT) the entire backbone weights, PEFT with prompt tuning (Coda and our DPG), Adapter [8], and LoRA [16]. + +
DatasetCarsIn-ShopSOPiNat2018Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
Coda [32], freeze(frz) backbone65.230.3678.610.0081.620.0178.970.0176.110.10
DPG (Ours), frz backbone70.620.0084.090.0082.690.0080.020.0179.360.00
FineTune [32]57.575.0482.230.5473.452.2355.693.8467.242.91
Adapter [8]64.592.4783.940.0082.180.0677.270.3176.990.71
LoRA [16]67.770.9983.290.4081.970.2777.470.7977.630.61
Coda, w/ FT50.0312.7982.290.4574.033.2957.424.3965.945.23
Coda, w/ Adapter64.390.8683.560.4581.980.0578.560.0677.120.36
Coda, w/ LoRA69.061.2382.790.3382.510.2579.120.3478.370.54
DPaRL, w/ FT65.004.6182.360.3374.242.7156.734.1269.582.94
DPaRL, w/ Adapter68.471.0085.720.0182.570.0178.040.0378.700.26
DPaRL, w/ LoRA (default)73.220.0386.280.0283.690.0780.020.2280.800.09
+ +domain. However, there are no available open-world benchmarks so we follow the standard closed-world setting in Figure 1 (b). We conduct experiments on MedMNIST [40], which consists of approximately 700K biomedical images. It includes 7 datasets, each with $7\sim 11$ classes. We evenly split the classes into 3 CL stages and report the average top-1 accuracy over each stage. Results in Table 5 demonstrate that DPaRL generalizes well to the medical domain, outperforming other methods with an average improvement of $6.50\%$ . For additional tasks in other domains, please refer to the supplementary material. + +Table 5: Top-1 (%) on medical data MedMNIST [40]. Our DPaRL surpasses other PCL methods by $6.50\%$ on average. + +
MethodsBloodDermaOrganAOrganCOrganSPathTissueAverage
L2P56.0334.0061.1252.8746.2556.3941.2949.71
Dual66.2833.1665.4257.7948.2260.6940.3453.13
Coda57.3733.7265.4165.9846.8279.8532.3054.49
DPaRL (Ours)70.4938.7170.3971.2951.2381.4443.3860.99(↑6.50)
+ +# 5 Conclusion + +In this paper, we introduce a practical setting for continual visual representation learning in the open-world. Existing continual learning methods fail to generalize well in this proposed setting. To mitigate this challenge, we propose Dynamic Prompt and Representation Learner (DPaRL), to enhance the adaptability and performance in continual learning for open-world visual recognition. Our method consistently demonstrates superior performance across various open-world image retrieval datasets and settings, by dynamically generating prompts from a deep neural network coupled with a dedicated mapping function, and effectively updating the discriminative representation backbone. Notably, it overcomes the intrinsic constraints found in static prompt pool designs, thereby expanding the limits of model generalization in open-world scenarios. + +# References + +1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018) +2. An, X., Deng, J., Yang, K., Li, J., Feng, Z., Guo, J., Yang, J., Liu, T.: Unicom: Universal and compact representation learning for image retrieval. arXiv preprint arXiv:2304.05884 (2023) +3. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +4. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020) +5. Cao, T., Wang, Y., Xing, Y., Xiao, T., He, T., Zhang, Z., Zhou, H., Tighe, J.: Pss: Progressive sample selection for open-world visual representation learning. In: European Conference on Computer Vision. pp. 278-294. Springer (2022) +6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018) +7. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019) +8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535 (2022) +9. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021) +10. De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence 44(7), 3366-3385 (2021) +11. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019) +12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +13. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) +14. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020) +15. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9769-9776. IEEE (2019) +16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) + +17. Kim, S., Kim, D., Cho, M., Kwak, S.: Proxy anchor loss for deep metric learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3238-3247 (2020) +18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013) +20. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: International Conference on Machine Learning. pp. 3925-3934. PMLR (2019) +21. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017) +22. Liu, Z., Luo, P., Qiu, S., Wang, X., Tang, X.: Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1096-1104 (2016) +23. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022) +24. Movshovitz-Attias, Y., Toshev, A., Leung, T.K., Ioffe, S., Singh, S.: No fuss distance metric learning using proxies. In: Proceedings of the IEEE international conference on computer vision. pp. 360-368 (2017) +25. Oh Song, H., Xiang, Y., Jegelka, S., Savarese, S.: Deep metric learning via lifted structured feature embedding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4004-4012 (2016) +26. Rao, D., Visin, F., Rusu, A., Pascanu, R., Teh, Y.W., Hadsell, R.: Continual unsupervised representation learning. Advances in neural information processing systems 32 (2019) +27. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023) +28. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) +29. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) +30. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: International conference on machine learning. pp. 4548-4557. PMLR (2018) +31. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021) +32. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) +33. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) + +34. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV. pp. 1706-1716 (2023) +35. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018) +36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +37. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2023) +38. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022) +39. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) +40. Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., Ni, B.: Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data 10(1), 41 (2023) +41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020) +42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017) +43. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017) +44. Zhao, B., Tang, S., Chen, D., Bilen, H., Zhao, R.: Continual representation learning for biometric identification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 1198-1208 (2021) +45. Zhou, D.W., Ye, H.J., Zhan, D.C., Liu, Z.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338 (2023) \ No newline at end of file diff --git a/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/images.zip b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d249782e18487c64bcee42573670468b5f253d8b --- /dev/null +++ b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed8d4e891fce9b9b1894f9988c51ca03556911c1af1463c9780bac06ee7b9bba +size 367146 diff --git a/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/layout.json b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f07fd2731f190dcc9a22d40db671e26180b245a --- /dev/null +++ b/2024/Open-World Dynamic Prompt and Continual Visual Representation Learning/layout.json @@ -0,0 +1,9053 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "Open-World Dynamic Prompt and Continual Visual Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "spans": [ + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": "Youngeun Kim" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{1,*†}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Jun Fang" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2,*‡}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Qin Zhang" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Zhaowei Cai" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Yantao Shen" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Rahul Duggal" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Dripta S. Raychaudhuri" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Zhuowen Tu" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Yifan Xing" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "text", + "content": ", Onkar Dabeer" + }, + { + "bbox": [ + 162, + 167, + 451, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 271, + 213, + 343, + 246 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 271, + 213, + 343, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 213, + 343, + 224 + ], + "spans": [ + { + "bbox": [ + 271, + 213, + 343, + 224 + ], + "type": "text", + "content": "Yale University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 271, + 224, + 341, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 224, + 341, + 235 + ], + "spans": [ + { + "bbox": [ + 271, + 224, + 341, + 235 + ], + "type": "text", + "content": "AWS AI Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 271, + 235, + 339, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 235, + 339, + 246 + ], + "spans": [ + { + "bbox": [ + 271, + 235, + 339, + 246 + ], + "type": "text", + "content": "3 Amazon AGI" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 147, + 247, + 463, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 247, + 463, + 269 + ], + "spans": [ + { + "bbox": [ + 147, + 247, + 463, + 269 + ], + "type": "text", + "content": "youngun.kim@yale.edu {junfa, qzaamz, zhaoweic, yantaos, dugrahul, driptarc, ztu, yifax, onkardab}@amazon.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 299, + 460, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 299, + 460, + 497 + ], + "spans": [ + { + "bbox": [ + 159, + 299, + 460, + 497 + ], + "type": "text", + "content": "Abstract. The open world is inherently dynamic, characterized by ever-evolving concepts and distributions. Continual learning (CL) in this dynamic open-world environment presents a significant challenge in effectively generalizing to unseen test-time classes. To address this challenge, we introduce a new practical CL setting tailored for open-world visual representation learning. In this setting, subsequent data streams systematically introduce novel classes that are disjoint from those seen in previous training phases, while also remaining distinct from the unseen test classes. In response, we present Dynamic Prompt and Representation Learner (DPaRL), a simple yet effective Prompt-based CL (PCL) method. Our DPaRL learns to generate dynamic prompts for inference, as opposed to relying on a static prompt pool in previous PCL methods. In addition, DPaRL jointly learns dynamic prompt generation and discriminative representation at each training stage whereas prior PCL methods only refine the prompt learning throughout the process. Our experimental results demonstrate the superiority of our approach, surpassing state-of-the-art methods on well-established open-world image retrieval benchmarks by an average of " + }, + { + "bbox": [ + 159, + 299, + 460, + 497 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 159, + 299, + 460, + 497 + ], + "type": "text", + "content": " improvement in Recall@1 performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 506, + 453, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 506, + 453, + 531 + ], + "spans": [ + { + "bbox": [ + 159, + 506, + 453, + 531 + ], + "type": "text", + "content": "Keywords: Dynamic Prompt Generation " + }, + { + "bbox": [ + 159, + 506, + 453, + 531 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 531 + ], + "type": "text", + "content": " Continual Learning " + }, + { + "bbox": [ + 159, + 506, + 453, + 531 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 531 + ], + "type": "text", + "content": " Open-World Visual Representation Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 550, + 230, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 230, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 230, + 563 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 575, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 624 + ], + "type": "text", + "content": "Continual learning (CL) without catastrophic forgetting is a challenging task in practice as retraining with old data is prohibitively expensive when data volume grows [10,14,23]. It is particularly challenging for real-world applications in the open-world [2,5], where new concepts keep emerging and distribution shifts are" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 631, + 223, + 642 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 223, + 642 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 223, + 642 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 643, + 344, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 344, + 653 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 344, + 653 + ], + "type": "text", + "content": "† Work conducted during an internship at Amazon." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 136, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 235, + 665 + ], + "type": "text", + "content": "‡ Corresponding author." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 118, + 216, + 183 + ], + "blocks": [ + { + "bbox": [ + 143, + 118, + 216, + 183 + ], + "lines": [ + { + "bbox": [ + 143, + 118, + 216, + 183 + ], + "spans": [ + { + "bbox": [ + 143, + 118, + 216, + 183 + ], + "type": "image", + "image_path": "c4956c950113a6c80b7d90ac9556c93ec61c1c08e9318344612d80e0fab2e977.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 185, + 192, + 196 + ], + "lines": [ + { + "bbox": [ + 179, + 185, + 192, + 196 + ], + "spans": [ + { + "bbox": [ + 179, + 185, + 192, + 196 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 118, + 291, + 183 + ], + "blocks": [ + { + "bbox": [ + 219, + 118, + 291, + 183 + ], + "lines": [ + { + "bbox": [ + 219, + 118, + 291, + 183 + ], + "spans": [ + { + "bbox": [ + 219, + 118, + 291, + 183 + ], + "type": "image", + "image_path": "fe91abc385f9ade7dd745fc2cc5a86198522244945daaa10c6dc39d75cb5caa7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 185, + 266, + 196 + ], + "lines": [ + { + "bbox": [ + 253, + 185, + 266, + 196 + ], + "spans": [ + { + "bbox": [ + 253, + 185, + 266, + 196 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 295, + 118, + 375, + 183 + ], + "blocks": [ + { + "bbox": [ + 295, + 118, + 375, + 183 + ], + "lines": [ + { + "bbox": [ + 295, + 118, + 375, + 183 + ], + "spans": [ + { + "bbox": [ + 295, + 118, + 375, + 183 + ], + "type": "image", + "image_path": "3f30019b810bc0377ab9000dc55444fadbc6d010816e1508fdb87ac41a36b015.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 185, + 343, + 196 + ], + "lines": [ + { + "bbox": [ + 331, + 185, + 343, + 196 + ], + "spans": [ + { + "bbox": [ + 331, + 185, + 343, + 196 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 215, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 131, + 215, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 482, + 326 + ], + "type": "text", + "content": "Fig. 1: Illustration of the distinctions between our problem setting and traditional settings. In our work, we aim to address the problem where training splits have no class overlaps. (a) Closed-world setting: both training and testing classes are identical. (b) Continual learning (CL) in closed-world setting: training classes are split into multiple divisions and are introduced through various CL stages. (c) Open-world setting: training and testing classes remain separate. The aim is to learn a robust representation from training classes that generalize to unseen classes. (d) Open-World Continual Representation Learning (Our Problem Setting): continual learning tailored for an open-world scenario. We sequentially introduce training classes over multiple CL stages, ensuring they remain distinct from test-time unseen classes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 384, + 118, + 470, + 183 + ], + "blocks": [ + { + "bbox": [ + 384, + 118, + 470, + 183 + ], + "lines": [ + { + "bbox": [ + 384, + 118, + 470, + 183 + ], + "spans": [ + { + "bbox": [ + 384, + 118, + 470, + 183 + ], + "type": "image", + "image_path": "92d55fa6a85a963a9e0bd211454df4ae89c774157d98a3df799e3e34df60b55a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 185, + 434, + 196 + ], + "lines": [ + { + "bbox": [ + 421, + 185, + 434, + 196 + ], + "spans": [ + { + "bbox": [ + 421, + 185, + 434, + 196 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 351, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 400 + ], + "type": "text", + "content": "inevitable. Recent studies on Prompt-based Continual Learning (PCL) [32,38, 39] demonstrated success in mitigating catastrophic forgetting without access to any past data. However, they heavily focus on the closed-world setting where test-time categories come from those encountered in various training stages." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 401, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 520 + ], + "type": "text", + "content": "Whereas in practice, the world is dynamic and concepts evolve over time and new categories are often introduced after a model is trained. Consider, for instance, an image retrieval system for clothing brands in online e-commerce. During training, one would gather images of all existing clothing brands available to train a model. However, after deployment, the system encounters new brands. The new data collected for system upgrade, constituting a new training stage, will likely include brands that are disjoint to those in older stages. Furthermore, the system will constantly face new brands at test-time that are disjoint to any of those encountered in the previous training stages and is expected to generalize well to those new categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": "Thus, to comprehensively evaluate the ability of a model in retaining semantic information from continuous streams of data for generalization to new concepts at test-time for practical applications, we propose a new continual open-world representation learning setting, as illustrated in Figure 1(d). In this setting, different data streams comprise distinct classes, and test-time classes are disjoint to those encountered in all training phases. Our setting is notably more challenging than the previous closed-world CL scenario. It closely mirrors real-world demands on recognition systems, where, during any evaluation stage, the test set consists of unseen entities to the training process up to that point. This setup reflects the need for practical recognition systems to encode general semantic information from training-time classes and apply this knowledge to new, unseen concepts without access to older training data during system updates." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "We start with evaluating state-of-the-art PCL methods [32, 38, 39] in our proposed setting and find that they fail to generalize with the absence of past data. For example, in our experiments, the gap between the paragon of standard training with all data in one phase and the best PCL baseline (Coda) shows up to " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " gap in Recall@1 (Table 1), significantly larger than the " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " gap in the closed-world setting, which suggests that prior PCL methods encounter difficulties and are less effective in the open-world setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 480, + 285 + ], + "type": "text", + "content": "To tackle this challenging yet practical open-world CL setting, we introduce a simple yet effective method called Dynamic Prompt and Representation Learner (DPaRL). Unlike previous prompt pool based PCL methods [32, 38, 39] that depend on a static prompt pool (Figure 2(a)) for test-time predictions. Notably, DPaRL trains an innovative dynamic prompt generation network jointly with the discriminative representation backbone model (Figure 2(b)), departing from the prompt-only learning paradigm in prior PCL methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 286, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 453 + ], + "type": "text", + "content": "Our key insight lies in that simply combining the fixed prompts learnt over training classes does not generalize well to unseen test-time classes in the practical open-world setting. On the other hand, the on-the-fly dynamically generated prompts in our method exhibit a stronger capability of capturing the diverse semantics spread across the numerous divisions of classes observed during the different training stages of continual learning, improving generalization to test-time unseen classes. This is achieved through the proposed dynamic prompt generation network, which possesses stronger representation power with a learnable mapping function compared to the traditional approach of combining multiple prompts in a static prompt pool. Additionally, unlike prior PCL methods, DPaRL jointly updates the dynamic prompt generation process and the discriminative representation backbone weights, facilitating maximized interaction and comparison between old concepts and newer ones to encapsulate diverse semantics that aid generalization to unseen test-time classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "content": "Consequently, over the four open-world image retrieval benchmarks on cars, online catalogs, products, and species images where the task is to search for the most similar images belonging to the same class as the query, via a distance measurement in an embedding space, DPaRL achieves significant accuracy boosts. It improves the average Recall@1 performance by absolute " + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "inline_equation", + "content": "9.1\\%" + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "content": " compared to state-of-the-art continual learning methods, from " + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "inline_equation", + "content": "76.1\\%" + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "inline_equation", + "content": "80.8\\%" + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "content": " over 10 CL stages and from " + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "inline_equation", + "content": "68.0\\%" + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "inline_equation", + "content": "77.1\\%" + }, + { + "bbox": [ + 130, + 456, + 479, + 539 + ], + "type": "text", + "content": " over 100 CL stages, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 541, + 351, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 541, + 351, + 552 + ], + "spans": [ + { + "bbox": [ + 146, + 541, + 351, + 552 + ], + "type": "text", + "content": "In summary, our contributions are three folds:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 567, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 567, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 567, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 136, + 567, + 479, + 590 + ], + "type": "text", + "content": "1. We establish a new practical setting on continual visual representation learning in the open-world." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 592, + 480, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 592, + 480, + 639 + ], + "spans": [ + { + "bbox": [ + 136, + 592, + 480, + 639 + ], + "type": "text", + "content": "2. We introduce a simple yet powerful method, Dynamic Prompt and Representation Learner (DPaRL), which dynamically generates prompts while effectively updating the discriminative representation backbone. This enhancement improves generalization for unseen open-world classes at test time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 479, + 665 + ], + "type": "text", + "content": "3. We outperform state-of-the-art continual learning methods in the proposed practical setting, both rehearsal-free and rehearsal-based." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 302, + 186 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 302, + 186 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 302, + 186 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 302, + 186 + ], + "type": "image", + "image_path": "102be1890616f633366a02d43040179e7334815d274287997908dc1476d130bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 190, + 266, + 201 + ], + "lines": [ + { + "bbox": [ + 181, + 190, + 266, + 201 + ], + "spans": [ + { + "bbox": [ + 181, + 190, + 266, + 201 + ], + "type": "text", + "content": "(a) Previous method" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 117, + 472, + 186 + ], + "blocks": [ + { + "bbox": [ + 313, + 117, + 472, + 186 + ], + "lines": [ + { + "bbox": [ + 313, + 117, + 472, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 472, + 186 + ], + "type": "image", + "image_path": "740775948f3f6a33e91d0b5fd35cc9ddb232c9a1a8d5101dfb2e98a836469bb1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 190, + 422, + 201 + ], + "lines": [ + { + "bbox": [ + 365, + 190, + 422, + 201 + ], + "spans": [ + { + "bbox": [ + 365, + 190, + 422, + 201 + ], + "type": "text", + "content": "(b) Our work" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 221, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 299 + ], + "type": "text", + "content": "Fig. 2: Prior PCL methods combine prompts from a static prompt pool trained on the training class distribution, leading to a loss of generalization capability when facing unseen classes during test time. Our work introduces a Dynamic Prompt Generation network that generates dynamic prompts on the fly by integrating a given image with stage tokens, followed by a specialized mapping function and adjustable discriminative representation backbone weights, providing generalizable prompts for unseen testing classes, distinguishing it from prior PCL methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 319, + 237, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 237, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 237, + 331 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "type": "text", + "content": "Continual learning. Continual learning (CL) has been recognized for decades as a crucial research domain, focusing on the ongoing adaptation of models over time. Early works primarily centered around regularization-based methods in an effort to address catastrophic forgetting [1,21,43] by introducing regularization between its current and previously learned parameters. However, they often underperform when faced with complex datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "content": "A separate line of work, architecture-based approach, involves expanding the network architecture in response to new learning phases [20,26,29,30,42]. While these methods generally outperform regularization-based counterparts, they do so at the cost of introducing additional parameters. Recent advancements include rehearsal-based methods that incorporate a memory component for storing past data [4,6,7,15,28]. This archived data aids in training newer stages, typically yielding superior performance compared to other CL techniques. Yet, this data retention can introduce privacy concerns and significant memory overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "Given these constraints, rehearsal-free strategies are gaining attention in the community. These aim to mitigate catastrophic forgetting without relying on historical data. A prominent method within this category uses model inversion to produce rehearsal images [9,13,31,41]. However, model inversion is computationally intensive and time-consuming. Additionally, these methods lag significantly in performance when compared to their rehearsal-based counterparts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Recent developments have witnessed the rise of prompt-based approaches [27, 32, 34, 37-39, 45]. These strategies offer robust protection against catastrophic forgetting by learning a small set of parameters, or prompts, instead of directly training all parameters. For instance, L2P [39] employs a pool of prompts, selecting suitable ones for insertion based on input data clustering. Building on this concept, DualPrompt [38] presents general prompts designed to encapsulate shared knowledge across various learning stages. Advancing this further," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "the state-of-the-art method CodaPrompt [32] facilitates end-to-end training by incorporating an attention mechanism for prompt combination within a pool. However, during inference, the above works rely on the combination of prompts from a static pool learned from training to feed into a frozen backbone model, which has limitations in cases where testing data has distribution shifts or disjoint classes from the training data. This motivates us to design a Dynamic Prompt and Representation Learner (DPaRL) to address these drawbacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 217, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 397 + ], + "type": "text", + "content": "Open-World Visual Representation Learning. In open-world visual representation learning, the model learns discriminative representations that align distances between representations with their semantic similarities, and uses these knowledge for testing on unseen classes. Previous works in this domain have primarily focused on designing loss functions [11, 17, 24] for enhanced accuracy. Recently, several studies [2, 12] have demonstrated the effectiveness of utilizing pre-trained visual foundation models and fine-tuning them to achieve significant performance improvements. While open-world visual representation learning has been under investigation for an extended period, there has been limited exploration within the continual learning (CL) framework for this task. [44] investigated biometric datasets using a regularization-based CL approach. However, the investigation into state-of-the-art prompt-based CL methods remains missing. Furthermore, as their studies primarily centered on biometric data, there is a pressing need to explore established continual learning benchmarks specific to open-world visual representation learning in more natural settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 414, + 233, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 233, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 233, + 428 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 463 + ], + "type": "text", + "content": "In this section, we introduce our open-world Continual Learning (CL) setting and motivate our Dynamic Prompt and Representation Learner (DPaRL)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 479, + 244, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 244, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 244, + 491 + ], + "type": "text", + "content": "3.1 Problem Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": "Two fundamental visual representation learning settings are the closed-world and open-world paradigms. In a closed-world setting (as depicted in Figure 1(a)), the training and testing data classes are entirely identical. Conversely, the open-world setting (illustrated in Figure 1(c)) presents a more practical scenario. Here, training and testing classes are entirely distinct, thus requiring the model to learn representations that generalize to unseen concepts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "When faced with continuous streams of data for visual recognition system updates, built upon the closed-world paradigm, Figure 1(b) showcases the well-known continual learning setting for image recognition. In this work, we introduce a new practical setting: Open-World Continual Representation Learning as shown in Figure 1(d). Instead of testing on seen classes of the various training stages thus far, the testing classes are entirely disjoint to those faced in different training stages, necessitating the model to not only learn continually but also generalize to new concepts, making it a more dynamic and challenging setting." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 462, + 209 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 462, + 209 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 462, + 209 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 462, + 209 + ], + "type": "image", + "image_path": "9bfe3eb4e63ee6f78904d8649b61ea55486f17d7086fdc13b3687209d646347c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "type": "text", + "content": "Fig. 3: Illustration of our proposed method: Dynamic Prompt and Representation Leaner (DPaRL). We dynamically generate prompt tokens on the fly from Dynamic Prompt Generation (DPG) networks by integrating information from stage tokens and image tokens. The [CLS] token from DPG is converted to prompt tokens via a linear mapping function. The generated prompt tokens are added to the backbone ViT and trained with a loss function. The learnable parameters include the current stage token, weights in mapping function, weights in backbone, and weights in loss function." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 327, + 384, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 384, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 384, + 339 + ], + "type": "text", + "content": "3.2 Prompt-based Continual Learning Paradigm" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 477 + ], + "type": "text", + "content": "Prompt-based Continual Learning (PCL) methods [32,38,39] use a pre-trained Vision Transformer (ViT) as a discriminative backbone for closed-world image classification, as shown in the Figure 2(a). These methods create a prompt pool with multiple prompt tokens, updating only the learnable parameters in this pool during training. In inference, the learned prompt pool is static, and PCL methods select tokens from this pool to feed into multiple ViT backbone layers for predictions. Our method adopts this PCL paradigm but adapts to the open-world representation learning setting with a prototypical metric learning loss. Additionally, we introduce a dynamic prompt generation (DPG) network to replace the static prompt pool for effective joint learning with the discriminative representation backbone model, as detailed below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 492, + 437, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 437, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 437, + 506 + ], + "type": "text", + "content": "3.3 Dynamic Prompt and Representation Leaner (DPaRL)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Overall Pipeline. In the open-world setting where training and testing classes are disjoint, existing prompt pool designs in PCL [32,38,39] exhibit limitations, i.e., limited separation between intra- and inter- testing class distance distributions as shown in Figure 4 and large performance gap from paragon in Table 1. To mitigate, we first propose dynamically generating prompts by blending the given image information and continual learning stage's information through a dynamic prompt generation (DPG) network. It aims to provide better representation power compared to the prompt pool design, which combines a fixed prompt pool trained from the training class distribution. Moreover, we enhance the capability of the discriminative representation backbone model to more effectively generalize to the open-world concepts through a joint dynamic prompt and representation learning paradigm. We refer to Figure 2, 3 for detailed design differences between our approach and prior prompt pool-based methods." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "Dynamic Prompt Generation Network. In order to retain information from each CL stage of our Dynamic Prompt Generation (DPG) process, we introduce stage tokens, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ". During training stage " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ", we train a stage token " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " while freezing the previous stage tokens " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "S_{t - (q - 1)} \\sim S_{t - 1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " in a first-in-first-out (FIFO) queue manner, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " is the max queue size. Guided by the ablation study detailed in Section 4.4, we typically set " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "q = 5" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " using a FIFO order to aim for both high accuracy and scalability. This ensures that the knowledge from the previous stages remains unchanged, and the total number of stage tokens limits by the queue size without linearly scaling with stage numbers. Within the deep neural network such as ViT, the self-attention block facilitates the integration of information between the stage-wise tokens and the instance-wise image tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 307 + ], + "type": "text", + "content": "As illustrated in Figure 3, the DPG produces a [CLS] token to dynamically acquire task-specific high-level information for prior knowledge of class-level features. However, it differs in size from the prompt tokens required by the discriminative backbone model (right side of Figure 3). To address this, we introduce a mapping function between the [CLS] token and the prompt tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": "A straightforward approach is to employ a single linear layer for dimensionality transformation. However, such a method introduces an excessive number of additional parameters, and the resulting over-parameterization can lead to overfitting of the highly compressed [CLS] token information. To tackle this issue, we impose a constraint on the weight parameter " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "W = AB^{T} \\in \\mathbb{R}^{C_{in} \\times C_{out}}" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{C_{in} \\times R}" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{C_{out} \\times R}" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": ", ensuring a maximum rank of " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "R < \\min(C_{in}, C_{out})" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": ". This constraint is inspired by the success of the LoRA technique [16]. The experimental results in Section 4.4 illustrate the critical importance of the low-rank mapping function design for our DPG module. Moreover, we apply Dropout [33] and LayerNorm [3] to this low-rank linear mapping function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 434, + 481, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 434, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 174, + 434, + 481, + 447 + ], + "type": "interline_equation", + "content": "M a p p i n g (x) = L a y e r N o r m \\left(L o w R a n k L i n e a r (D r o p O u t (x))\\right), \\tag {1}", + "image_path": "765b2c40498e790dc5996abba5540ff7ddc09dc1f03ce6080a3c4a829c6577db.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 453, + 404, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 404, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 404, + 464 + ], + "type": "text", + "content": "which further helps to avoid overfitting and stabilize training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "text", + "content": "With the help of this specialized mapping function along with a pre-trained neural network, the prompt " + }, + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "text", + "content": " is dynamically obtained from the input image " + }, + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "text", + "content": " and stage tokens " + }, + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "inline_equation", + "content": "S_{t,q} \\coloneqq S_{t - (q - 1)} \\sim S_t" + }, + { + "bbox": [ + 130, + 464, + 482, + 502 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 508, + 480, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 508, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 231, + 508, + 480, + 520 + ], + "type": "interline_equation", + "content": "P = \\text {M a p p i n g} (\\text {N e t w o r k} ([ S _ {t, q}; I ]). \\tag {2}", + "image_path": "dc91ea29664ce82c0deecdd29969fc35ff0d814dc0d36362a6b20ca484251824.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": "We reshape " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": " to a size of " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "N_{p} \\times C \\times L" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": " is the number of prompts, " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": " is the channel dimension, and " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": " indicates the number of layers where prompts are applied to the backbone model. The generated prompts are inserted across the first " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": " layers in the ViT backbone. In our DPG approach to prompting, we follow the technical foundations of prompting from the prior state-of-the-art PCL methods [32,38]. This is because we want to make fair comparison with the previous work, and our primary focus is on the formation of prompt. The prompt for the " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": "-th layer, denoted as " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "P_{l} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": ", is partitioned into " + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\{P_{l,k}, P_{l,v}\\} \\in \\mathbb{R}^{\\frac{N}{2} \\times C}" + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": " and is prefixed to the input token embedding for both key and value in the attention mechanism [36]. This can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 653, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 653, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 205, + 653, + 480, + 667 + ], + "type": "interline_equation", + "content": "h _ {i} = \\text {A t t e n t i o n} \\left(X _ {l} W _ {q} ^ {i}, [ P _ {k}; X ] W _ {k} ^ {i}, [ P _ {v}; X ] W _ {v} ^ {i}\\right). \\tag {3}", + "image_path": "df122f642a13635a1685b2e29dd0a8f9bfc1eacc3b34be129bcb875139fdf53b.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Here, we omit layer index " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " for brevity, and " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": " is head index in multi-head self-attention of transformer architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 159, + 483, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 483, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 483, + 340 + ], + "type": "text", + "content": "Joint Dynamic Prompt and Representation Learning. Differing from prior PCL methods [32,38,39] that freeze the backbone model (on the right side of the pipeline in Figure 2(a)) during training, we propose our Dynamic Prompt and Representation Learner (DPaRL), to jointly learn the dynamic prompt generation with the discriminative representation learning backbone (on the right side of Figure 2(b) and Figure 3). This approach aims to maximize the capability of entire pipeline to integrate old-stage concepts and new ones to encapsulate diverse semantics that aid generalization to open-world unseen classes and unknown domain shifts. Leveraging parameter-efficient fine-tuning techniques, our DPaRL successfully maximizes accuracy performance and minimizes catastrophic forgetting simultaneously. The overall framework is depicted in Figure 3. Note that the DPG encoder weights (on the left side of the pipeline) for prompt generation are frozen in our method, while the learnable parameters include stage tokens, weights in the mapping function and representation loss function, as well as the discriminative representation backbone weights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 359, + 230, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 230, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 230, + 373 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 422 + ], + "type": "text", + "content": "In this section, we empirically evaluate the effectiveness of our Dynamic Prompt and Representation Learner (DPaRL) in the established challenging continual open-world visual representation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 441, + 313, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 313, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 313, + 452 + ], + "type": "text", + "content": "4.1 Datasets & Evaluation Metric" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "content": "Datasets. We conducted our training and testing across 4 prominent open-world image retrieval benchmarks: Cars [19], In-Shop [22], SOP [25] and iNat2018 [35]. We evaluate baselines and our proposed method following our practical continual open-world visual representation learning setting. In this dynamic scenario, classes are introduced incrementally at each learning stage, and the testing phase encompasses classes that have not been encountered during training. Our experiments consist of 10 or 100 stages, with the total training classes approximately distributed uniformly across these stages. The detailed data information including number of classes and samples are provided in the Supplementary Material. For inference, we use the entire testing dataset across all CL stages since there is no overlap between training and testing classes. Thus, our objective is to learn class relations while maintaining generalization across CL stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "text", + "content": "Evaluation Metric. To evaluate model performance, we compute the Recall@1 metric using the [CLS] token at the last layer of the discriminative representation backbone network. Given our goal to optimize the model's adaptability throughout the CL phases, we report average Recall@1 across all CL stage. Let " + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "inline_equation", + "content": "R_{n}" + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "text", + "content": " be the Recall@1 performance at stage " + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "text", + "content": ", then we report " + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "inline_equation", + "content": "R_{N} = \\frac{1}{N_{s}}\\sum_{n = 1}^{N_{s}}R_{n}" + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 606, + 482, + 669 + ], + "type": "inline_equation", + "content": "N_{s}" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": "is the total number of stages. We also report the forgetting score " + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "inline_equation", + "content": "F_{N}" + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": " by following [32,38] to measure the performance drop on the same unseen testing dataset across different training stages, that is, " + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "inline_equation", + "content": "F_{N} = \\frac{1}{N_{s} - 1}\\sum_{n = 2}^{N_{s}}\\max_{m\\in \\{1,\\dots,n - 1\\}}(0,R_{m} - R_{n})" + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": ". We stress that " + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "inline_equation", + "content": "R_{N}" + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": " holds greater importance as it captures both learning capacity and forgetting, while " + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "inline_equation", + "content": "F_{N}" + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": " merely offers supplementary context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 201, + 317, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 317, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 317, + 213 + ], + "type": "text", + "content": "4.2 Baselines and Implementations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 393, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 393, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 393, + 232 + ], + "type": "text", + "content": "Baselines. We validate 3 types of representative baselines:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 239, + 479, + 287 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 239, + 351, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 351, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 351, + 251 + ], + "type": "text", + "content": "-Rehearsal-based:Experience Replay (ER) [7]," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 422, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 422, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 422, + 262 + ], + "type": "text", + "content": "- Regularization-based: Learning without Forgetting (LwF) [21]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 262, + 479, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 479, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 479, + 287 + ], + "type": "text", + "content": "- Rehearsal-free prompt-based: Learning to Prompt (L2P) [39], DualPrompt [38], and CodaPrompt [32]." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 293, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 342 + ], + "type": "text", + "content": "**Implementations Details.** Consistently across all methods, we adhere to the training setups by following [32]. Specifically, we utilize the Adam optimizer [18] with a batch size of 128 to train for 20 epochs to ensure comprehensive model convergence. The learning rate is cosine decaying with an initial value of 0.001." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 342, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 413 + ], + "type": "text", + "content": "We utilize ArcFace [11] as the training loss function as it is a prevalent choice in open-world representation learning to achieve state-of-the-art performance [2]. For new-stage training, the prototype weights in the loss function of old training stages are appended to the new training stage as initialization. All prototype weights in the loss function, including those corresponding to the classes for old and new stages, are trainable during each training stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 498 + ], + "type": "text", + "content": "For established PCL techniques, we maintain the prompt lengths and locations for L2P [39], DualPrompt [38], and CodaPrompt [32] as suggested in these works. For our DPaRL method, in each CL stage, we incorporate a learnable stage token with the first-in-first-out (FIFO) order and a max size of 5 tokens based on insights from Section 4.4. We apply our specialized mapping function in the DPG module to avoid over-fitting and stabilize training. We adopt LoRA [16] as the PEFT method to update backbone weights as guided by Section 4.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 514, + 371, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 371, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 371, + 528 + ], + "type": "text", + "content": "4.3 Performance Comparison with Prior Arts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "Table 1 shows a comprehensive performance comparison with various CL methods across all four open-world image retrieval tasks. We adopt the official codebase of CodaPrompt [32] for the implementation of PCL baselines including L2P [39], DualPrompt [38], and CodaPrompt [32]. Additionally, we report the performance of representative rehearsal-based and regularization-based approaches, i.e., ER [7] and LWF [21], respectively, for a more comprehensive comparison. Note that, for a fair comparison, all methods use the pre-trained ViT-Base on ImageNet-21k as the backbone model, as shown on the right side of Figure 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "From the results in the Table 1, we observe that the performance trend of PCL-based methods (L2P, DualPrompt, CodaPrompt, our DPaRL) is similar to what was observed in the closed-world image recognition task [32], where PCL" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 224, + 473, + 315 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": "Table 1: Performance on 4 open-world image retrieval benchmarks with 10 CL stages. We report the averaged " + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "Recall@1" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "R_N" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "Forgetting\\ Score" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": ") across all continual learning stages, the results are averaged over 3 runs. We present the LowerBound with the zero-shot performance of ImageNet-21k pre-trained ViT-B/16 backbone. UpperBound denotes the upper bound performance when trained on the entire dataset at a single stage. All PCL methods, including L2P, DualPrompt, CodaPrompt, and our DPaRL, utilize the same pre-trained encoder, ViT-B/16 on ImageNet-21k, to generate the prompt tokens. The numbers in brackets show the " + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "Recall@1" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": " improvement of our DPaRL over the previous best baselines, with the previous best baseline underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 224, + 473, + 315 + ], + "lines": [ + { + "bbox": [ + 141, + 224, + 473, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 473, + 315 + ], + "type": "table", + "html": "
DatasetCars [19]In-Shop [22]SOP [25]iNat2018 [35]Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
LowerBound53.86-47.16-62.85-71.74-58.90-
UpperBound85.96-89.28-85.21-82.68-85.78-
ER [7]65.380.4982.992.4865.380.5036.9418.5862.675.51
LwF [21]56.532.9181.523.3966.1813.2039.6933.8160.9813.33
L2P [39]56.270.0067.850.0076.880.4078.950.0069.990.10
DualPrompt [38]55.540.0065.560.0077.040.6678.710.0369.210.17
CodaPrompt [32]65.230.3678.610.0081.620.0178.970.0176.110.10
DPaRL (Ours)73.22(↑7.8)0.0386.28(↑3.3)0.0283.69(↑2.1)0.0780.02(↑1.0)0.0380.80(↑4.7)0.09
", + "image_path": "7c38699a069e9141d1d55a261eaaf4fee9b491c8b34d4e724dc2ac8f5edd990e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "text", + "content": "methods outperform other continual learning approaches by a significant margin on both recall@1 metric " + }, + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "inline_equation", + "content": "R_N" + }, + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "text", + "content": " and forgetting score " + }, + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "inline_equation", + "content": "F_N^4" + }, + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "text", + "content": ", aligning with findings from previous research [32, 38, 39]. This suggests that PCL methods are strong baselines for continual learning on both closed-world image recognition tasks and open-world visual representation learning tasks. Notably, our DPaRL approach consistently outperforms other methods across all datasets, resulting in an average Recall@1 boost of " + }, + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 339, + 482, + 445 + ], + "type": "text", + "content": " over the leading previous method, CodaPrompt [32]. This supports our hypothesis about the limitations of the conventional PCL design in open-world scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "type": "text", + "content": "In the following, we provide step-by-step guidance on the design choices of our DPaRL. Additional analysis and results are presented in the supplementary material, covering the superior performance of DPaRL in few-shot settings, effectiveness of DPaRL in closed-world benchmarks, dynamic changes in the accuracy curve across CL stages, and robustness of DPaRL against various DPG encoders. Due to space limitations, we refer to the supplementary material for a more comprehensive exploration of these aspects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 552, + 389, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 389, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 389, + 565 + ], + "type": "text", + "content": "4.4 Effectiveness of Dynamic Prompt Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 577, + 481, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 481, + 612 + ], + "type": "text", + "content": "To solely investigate the effectiveness of our Dynamic Prompt Generation design, we freeze the backbone weights similar to prior PCL methods [32, 38, 39] in our DPaRL and denoted as the DPG method." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " PCL methods consistently achieve forgetting scores close to zero, indicating that performance continually improves. This trend is also evident in the dynamic change of accuracy curve in the supplementary material, highlighting that PCL methods serve as robust baselines for addressing catastrophic forgetting in open-world tasks." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 219, + 187 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 219, + 187 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 219, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 219, + 187 + ], + "type": "image", + "image_path": "97169a63cef55964afa7bb192180a1118f496ee640cc082bebb2f7b9dadb904c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 209, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 276 + ], + "type": "text", + "content": "Fig. 4: Histogram of L2 distance between a pair of samples with embedding features extracted by Coda [32] (SOTA PCL method) and our DPG (the naive version of DPaRL with freezing backbone weights). Left and right figures are distributions of the seen and unseen classes in training and testing Cars [19] dataset, respectively. Our DPG exhibits enhanced separation between inter- and intra-classes, particularly on open-world unseen test classes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 115, + 302, + 188 + ], + "blocks": [ + { + "bbox": [ + 221, + 115, + 302, + 188 + ], + "lines": [ + { + "bbox": [ + 221, + 115, + 302, + 188 + ], + "spans": [ + { + "bbox": [ + 221, + 115, + 302, + 188 + ], + "type": "image", + "image_path": "139fdfb8ac68d17f6424db357b5d5f8908e68f7b5e8fa91a82ce609ca6625ee3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 392, + 187 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 392, + 187 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 392, + 187 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 392, + 187 + ], + "type": "image", + "image_path": "55c375d5732f61354c1ea85b0adfd0a16d298eb8bb0acffb4d1e3a1cb3ff2ca3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 115, + 474, + 187 + ], + "blocks": [ + { + "bbox": [ + 394, + 115, + 474, + 187 + ], + "lines": [ + { + "bbox": [ + 394, + 115, + 474, + 187 + ], + "spans": [ + { + "bbox": [ + 394, + 115, + 474, + 187 + ], + "type": "image", + "image_path": "593ec7899c6673d945afc746e8519d66b3670489bdee5179b831a66130a51f81.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 301, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 433 + ], + "type": "text", + "content": "We visualize the distributions of L2 distance between pairwise samples within intra- and inter-classes of CodaPrompt (Coda) [32] and our DPG in Figure 4 to assess generalizability. A good model should exhibit distinguishable distributions between intra- and inter-classes for both closed-world seen classes and open-world unseen classes. Coda, employing a static prompt pool design, achieves a notable separation on closed-world intra- and inter-classes but exhibits a close distribution for open-world intra- and inter-classes, impeding effective generalization to unseen test classes. In contrast, our DPG shows a more distinct separation, with an average gap of 0.14 (twice as large as Coda's gap) between inter- and intra-class distances in testing data. This indicates better generalizability with a dynamic prompt generation mechanism for unseen classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "text", + "content": "Moreover, accuracy results with Recall@1 in Table 2 and Table 4 reveal that our DPG significantly outperforms the best baseline (Coda), with an average of " + }, + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "inline_equation", + "content": "9.1\\%" + }, + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "text", + "content": " boost in 100 CL stages and " + }, + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "text", + "content": " boost in 10 CL stages, respectively. This highlights the superiority of Dynamic Prompting Generation design, particularly in practical scenarios involving a large number of stages. In the next Section 4.4.1 and Section 4.4.2, we investigate the importance of our design components of stage tokens and mapping function in the DPG network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 540, + 309, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 309, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 309, + 553 + ], + "type": "text", + "content": "4.4.1 Effectiveness of Stage Tokens" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In our DPaRL design with Dynamic Prompt Generation (DPG), we introduce stage tokens to retain information on old training stages, and we impose a max limit on the number of stage tokens (" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{stg\\_tok} = 5" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ") to control storage and computation costs as the number of stages scales up, e.g., " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{stg} = 100" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". Hence, when the stage number goes beyond the max limit " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{stg\\_tok}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", we replace some of previous stage tokens. Our hypothesis is that the earlier stage information learned from previous stage tokens can propagate into the later ones. In this section, we conduct ablations in a practical CL setting with a large number of stages (100 stages) to study the impact of stage tokens." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 168, + 449, + 261 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Impact of stage tokens with a large number of CL stages, " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "N_{stg} = 100" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": ". Recall@1 results on In-Shop, SOP, and iNat2018 datasets are reported. It's worth noting that the Cars dataset is relatively small, preventing its division into 100 stages. DPG is a naive version of our DPaRL method by freezing the backbone weights." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 168, + 449, + 261 + ], + "lines": [ + { + "bbox": [ + 165, + 168, + 449, + 261 + ], + "spans": [ + { + "bbox": [ + 165, + 168, + 449, + 261 + ], + "type": "table", + "html": "
Method# Stage TokenOrderIn-ShopSOPiNatAverage
L2P [39]N/AN/A48.3970.3976.8265.20
Dual [38]N/AN/A52.5073.7676.9167.72
Coda [32]N/AN/A52.9673.9276.9867.95
DPGNoneN/A71.1880.0477.5776.26 (↑8.31)
DPG5FIFO (default)72.0580.8178.2977.05 (↑9.10)
DPG5FILO71.7680.6078.1276.83 (↑8.88)
DPG5Random72.0080.8378.2077.01 (↑9.06)
DPG10FIFO72.1580.8178.2877.08 (↑9.13)
DPG100full-size baseline72.2080.8478.3777.14 (↑9.19)
", + "image_path": "5cd6cd1a14bc290a63c620957ca939b8535d08595153b6db9f9fb11dcb081fb5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 479, + 330 + ], + "type": "text", + "content": "With and without stage token. From Table 2, DPG with 5 stage tokens achieves a noticeable accuracy improvement over DPG without stage tokens, 77.05 vs. 76.26 on average, indicating the importance of stage token for retaining stage-wise class information in old training stages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 426 + ], + "type": "text", + "content": "Impact of the order. We conduct ablations on three different orders: first-in-first-out (FIFO), first-in-last-out (FILO), and randomly replacing one stage token (Random). Table 2 reveals that FIFO performs the best on average. FILO achieves slightly lower accuracy. Nevertheless, the Random method achieves performance very close to the best, indicating the robustness of our DPG design to different orders of stage tokens. We hypothesize this due to stage information learned from previous ones effectively propagating to the new stage tokens, regardless of the order or the location where the new stage token is inserted." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "text", + "content": "Selection of maximum limit " + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "M_{stg\\_tok}" + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "text", + "content": ". When the stage number is large (" + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "N_{stg} = 100" + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "text", + "content": "), keeping a maximum token number of 5 achieves comparable performances (77.05 vs. 77.14 on average) to that of the full token number baseline with " + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "M_{stg\\_tok} = 100" + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "text", + "content": ". This signifies a highly effective selection with a constant number of stage tokens to achieve high accuracy, as well as good scalability with respect to the stage number. Hence, we use " + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "M_{stg\\_tok} = 5" + }, + { + "bbox": [ + 132, + 426, + 482, + 499 + ], + "type": "text", + "content": " throughout this paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 516, + 335, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 335, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 335, + 528 + ], + "type": "text", + "content": "4.4.2 Effectiveness of Mapping Function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "text", + "content": "In our DPaRL framework (Figure 3), as elaborated in Section 3.3, we introduce a mapping function to match the size of [CLS] token and prompt tokens. We apply rank-" + }, + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "text", + "content": " constraint to avoid overfitting and reduce the memory consumption from the fully-connected layer inside the mapping function. To measure the sensitivity of parameter " + }, + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "text", + "content": ", we conduct experiments with varying " + }, + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 533, + 480, + 593 + ], + "type": "text", + "content": " from 64 to 768." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": "The results for both In-Shop and iNat in Table 3 indicate the existence of an optimal rank with " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "R_{opt} = 256" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": " that strikes a balance between accuracy and the number of parameters, saving " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "2/3" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": " of the total parameters compared to the full rank case. This default rank " + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "inline_equation", + "content": "R_{opt} = 256" + }, + { + "bbox": [ + 130, + 594, + 480, + 641 + ], + "type": "text", + "content": " is subsequently used for all experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "content": "Note that when the rank reaches 768, representing the full rank of the mapping function essentially serving as a fully-connected layer, a significant per" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 146, + 421, + 193 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Impact of the rank of weight matrix inside the mapping function on our DPG design. Recall@1 results on In-Shop and iNat2018 with 10 CL stages are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 146, + 421, + 193 + ], + "lines": [ + { + "bbox": [ + 192, + 146, + 421, + 193 + ], + "spans": [ + { + "bbox": [ + 192, + 146, + 421, + 193 + ], + "type": "table", + "html": "
Rank64128256384512768
In-Shop82.3383.1784.0983.8682.9871.40
iNat201879.2179.8580.0279.7479.6378.98
#Parameters2.0M3.9M7.9M12.0M16.1M23.5M
", + "image_path": "5fcb6b44d46be5391df05335ebcb05ab7b5da7ebfc039b971be1e9348c959346.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 274 + ], + "type": "text", + "content": "formance drop of " + }, + { + "bbox": [ + 130, + 213, + 482, + 274 + ], + "type": "inline_equation", + "content": "12.69\\%" + }, + { + "bbox": [ + 130, + 213, + 482, + 274 + ], + "type": "text", + "content": " is observed for the In-Shop dataset in Table 3. This suggests that an overparameterized expanding layer results in a severe accuracy degradation for a highly compressed [CLS] token, indicating the importance of our design. Moreover, our design has a robust rank range from 128 to 384, which closely matches the accuracy performance of the peak result." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 473, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 473, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 473, + 302 + ], + "type": "text", + "content": "4.5 Effectiveness of Dynamic Prompt and Representation Learner" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "text", + "content": "Previous PCL methods [32,38,39] freeze backbone weights of the discriminative model to reduce catastrophic forgetting. However, this approach could limit accuracy performance as it doesn't provide the degree of freedom to update the backbone, particularly in open-world setting where data distribution and domain shifts naturally exist. In this section, we explore effective ways to update backbone weights jointly with our dynamic prompt learning. We explore 3 approaches here: 1) fine-tuning " + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "inline_equation", + "content": "(FT)" + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "text", + "content": " the entire backbone weights, 2) parameter-efficient fine-tuning (PEFT) with AdaptFromer [8] (Adapter) or Low-Rank Adaptation [16] weights on the backbone " + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "inline_equation", + "content": "(LoRA)" + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "text", + "content": ", and 3) joint prompt generation and representation learning on the backbone weights (Coda and " + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "inline_equation", + "content": "DPaRL" + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "inline_equation", + "content": "w / X" + }, + { + "bbox": [ + 130, + 308, + 482, + 426 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": "Table 4 shows that vanilla " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": " or fine-tuning with PCL methods not only leads to lower accuracy but also suffers from higher catastrophic forgetting compared to DPG. While, LoRA gets higher accuracy " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "(+10.4\\%)" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": " on average) and reduces forgetting " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "(-2.3\\%)" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": " on average) compared to vanilla " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": ", as well as outperforms Adapter, indicating a more effective way to update the backbone. Still, LoRA is less effective than DPG, with an average accuracy gap of " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": " and a forgetting gap of " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": ", indicating the superiority of our DPG design for open-world CL problems. Lastly, our joint Dynamic Prompt and Representation Learner (DPaRL) with LoRA improves average accuracy by " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": " and reduces forgetting by " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": " compared to LoRA-only baseline. It also boosts accuracy by " + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 427, + 482, + 608 + ], + "type": "text", + "content": " on average compared to DPG method without updating backbone weights at a negligible cost of 0.09 forgetting raises, indicating the effectiveness of our DPaRL design to maximizing accuracy without causing catastrophic forgetting issues. Therefore, unless stated otherwise, we adopt joint learning with dynamic prompt generation with LoRA as our default DPaRL method in this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 623, + 447, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 447, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 447, + 635 + ], + "type": "text", + "content": "4.6 Performance on Closed-World Setting in Other Domains" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we assess the effectiveness of our DPaRL method in domains other than the established image retrieval benchmarks such as medical image" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 157, + 477, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 4: Accuracy performance " + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "(R_N, \\uparrow)" + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": " and forgetting scores " + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "(F_N, \\downarrow)" + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": " in 10 CL stages with different fine-tuning methods: fine-tuning (FT) the entire backbone weights, PEFT with prompt tuning (Coda and our DPG), Adapter [8], and LoRA [16]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 157, + 477, + 258 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 477, + 258 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 477, + 258 + ], + "type": "table", + "html": "
DatasetCarsIn-ShopSOPiNat2018Average
MetricRN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)RN(↑)FN(↓)
Coda [32], freeze(frz) backbone65.230.3678.610.0081.620.0178.970.0176.110.10
DPG (Ours), frz backbone70.620.0084.090.0082.690.0080.020.0179.360.00
FineTune [32]57.575.0482.230.5473.452.2355.693.8467.242.91
Adapter [8]64.592.4783.940.0082.180.0677.270.3176.990.71
LoRA [16]67.770.9983.290.4081.970.2777.470.7977.630.61
Coda, w/ FT50.0312.7982.290.4574.033.2957.424.3965.945.23
Coda, w/ Adapter64.390.8683.560.4581.980.0578.560.0677.120.36
Coda, w/ LoRA69.061.2382.790.3382.510.2579.120.3478.370.54
DPaRL, w/ FT65.004.6182.360.3374.242.7156.734.1269.582.94
DPaRL, w/ Adapter68.471.0085.720.0182.570.0178.040.0378.700.26
DPaRL, w/ LoRA (default)73.220.0386.280.0283.690.0780.020.2280.800.09
", + "image_path": "db99ec69f63a930675a3c5cf9511515f941994f5f79dbca9856fc05623082c91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 278, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 373 + ], + "type": "text", + "content": "domain. However, there are no available open-world benchmarks so we follow the standard closed-world setting in Figure 1 (b). We conduct experiments on MedMNIST [40], which consists of approximately 700K biomedical images. It includes 7 datasets, each with " + }, + { + "bbox": [ + 132, + 278, + 480, + 373 + ], + "type": "inline_equation", + "content": "7\\sim 11" + }, + { + "bbox": [ + 132, + 278, + 480, + 373 + ], + "type": "text", + "content": " classes. We evenly split the classes into 3 CL stages and report the average top-1 accuracy over each stage. Results in Table 5 demonstrate that DPaRL generalizes well to the medical domain, outperforming other methods with an average improvement of " + }, + { + "bbox": [ + 132, + 278, + 480, + 373 + ], + "type": "inline_equation", + "content": "6.50\\%" + }, + { + "bbox": [ + 132, + 278, + 480, + 373 + ], + "type": "text", + "content": ". For additional tasks in other domains, please refer to the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 150, + 426, + 461, + 472 + ], + "blocks": [ + { + "bbox": [ + 132, + 393, + 480, + 417 + ], + "lines": [ + { + "bbox": [ + 132, + 393, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 480, + 417 + ], + "type": "text", + "content": "Table 5: Top-1 (%) on medical data MedMNIST [40]. Our DPaRL surpasses other PCL methods by " + }, + { + "bbox": [ + 132, + 393, + 480, + 417 + ], + "type": "inline_equation", + "content": "6.50\\%" + }, + { + "bbox": [ + 132, + 393, + 480, + 417 + ], + "type": "text", + "content": " on average." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 426, + 461, + 472 + ], + "lines": [ + { + "bbox": [ + 150, + 426, + 461, + 472 + ], + "spans": [ + { + "bbox": [ + 150, + 426, + 461, + 472 + ], + "type": "table", + "html": "
MethodsBloodDermaOrganAOrganCOrganSPathTissueAverage
L2P56.0334.0061.1252.8746.2556.3941.2949.71
Dual66.2833.1665.4257.7948.2260.6940.3453.13
Coda57.3733.7265.4165.9846.8279.8532.3054.49
DPaRL (Ours)70.4938.7170.3971.2951.2381.4443.3860.99(↑6.50)
", + "image_path": "06ada58c40519e3df5205419138826cf469aa58e39b9d87e92dda5fbe2dd9803.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 510, + 219, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 219, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 219, + 522 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we introduce a practical setting for continual visual representation learning in the open-world. Existing continual learning methods fail to generalize well in this proposed setting. To mitigate this challenge, we propose Dynamic Prompt and Representation Learner (DPaRL), to enhance the adaptability and performance in continual learning for open-world visual recognition. Our method consistently demonstrates superior performance across various open-world image retrieval datasets and settings, by dynamically generating prompts from a deep neural network coupled with a dedicated mapping function, and effectively updating the discriminative representation backbone. Notably, it overcomes the intrinsic constraints found in static prompt pool designs, thereby expanding the limits of model generalization in open-world scenarios." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "text", + "content": "1. Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV). pp. 139-154 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 209 + ], + "type": "text", + "content": "2. An, X., Deng, J., Yang, K., Li, J., Feng, Z., Guo, J., Yang, J., Liu, T.: Unicom: Universal and compact representation learning for image retrieval. arXiv preprint arXiv:2304.05884 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "text", + "content": "3. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 264 + ], + "type": "text", + "content": "4. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33, 15920-15930 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 264, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 298 + ], + "type": "text", + "content": "5. Cao, T., Wang, Y., Xing, Y., Xiao, T., He, T., Zhang, Z., Zhou, H., Tighe, J.: Pss: Progressive sample selection for open-world visual representation learning. In: European Conference on Computer Vision. pp. 278-294. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "text", + "content": "6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "type": "text", + "content": "7. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 387 + ], + "type": "text", + "content": "8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "text", + "content": "9. Choi, Y., El-Khamy, M., Lee, J.: Dual-teacher class-incremental learning with data-free generative replay. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3543–3552 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 422, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 422, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 134, + 422, + 480, + 464 + ], + "type": "text", + "content": "10. De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence 44(7), 3366-3385 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 465, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 465, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 134, + 465, + 480, + 498 + ], + "type": "text", + "content": "11. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 498, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 498, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 498, + 480, + 542 + ], + "type": "text", + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 480, + 576 + ], + "type": "text", + "content": "13. Gao, Q., Zhao, C., Ghanem, B., Zhang, J.: R-dfcil: Relation-guided representation learning for data-free class incremental learning. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "type": "text", + "content": "14. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "text", + "content": "15. Hayes, T.L., Cahill, N.D., Kanan, C.: Memory efficient experience replay for streaming learning. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9769-9776. IEEE (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Kim, S., Kim, D., Cho, M., Kwak, S.: Proxy anchor loss for deep metric learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3238-3247 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "19. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "20. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: International Conference on Machine Learning. pp. 3925-3934. PMLR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "21. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "22. Liu, Z., Luo, P., Qiu, S., Wang, X., Tang, X.: Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1096-1104 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "23. Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing 469, 28-51 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "text", + "content": "24. Movshovitz-Attias, Y., Toshev, A., Leung, T.K., Ioffe, S., Singh, S.: No fuss distance metric learning using proxies. In: Proceedings of the IEEE international conference on computer vision. pp. 360-368 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "25. Oh Song, H., Xiang, Y., Jegelka, S., Savarese, S.: Deep metric learning via lifted structured feature embedding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4004-4012 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "26. Rao, D., Visin, F., Rusu, A., Pascanu, R., Teh, Y.W., Hadsell, R.: Continual unsupervised representation learning. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "27. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "28. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "29. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "30. Serra, J., Suris, D., Miron, M., Karatzoglou, A.: Overcoming catastrophic forgetting with hard attention to the task. In: International conference on machine learning. pp. 4548-4557. PMLR (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "31. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9374-9384 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 632 + ], + "type": "text", + "content": "32. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "33. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "Y. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 522 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "34. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV. pp. 1706-1716 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 183 + ], + "type": "text", + "content": "35. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "text", + "content": "37. Wang, L., Xie, J., Zhang, X., Huang, M., Su, H., Zhu, J.: Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. NeurIPS (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "text", + "content": "38. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "text", + "content": "39. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": "40. Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., Ni, B.: Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data 10(1), 41 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "text", + "content": "41. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "text", + "content": "42. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 435, + 482, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 457 + ], + "type": "text", + "content": "43. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International conference on machine learning. pp. 3987-3995. PMLR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "type": "text", + "content": "44. Zhao, B., Tang, S., Chen, D., Bilen, H., Zhao, R.: Continual representation learning for biometric identification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 1198-1208 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "text", + "content": "45. Zhou, D.W., Ye, H.J., Zhan, D.C., Liu, Z.: Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "DPaRL for Open-World Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_content_list.json b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..38042fa8a0c38f0da007f33034cabebea9d2ce07 --- /dev/null +++ b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_content_list.json @@ -0,0 +1,1977 @@ +[ + { + "type": "text", + "text": "Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning", + "text_level": 1, + "bbox": [ + 215, + 140, + 787, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dexuan Zhang $^{1}$ , Thomas Westfechtel $^{1}$ , and Tatsuya Harada $^{1,2}$", + "bbox": [ + 274, + 212, + 725, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Tokyo \n2 RIKEN", + "bbox": [ + 410, + 239, + 591, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Open-set domain adaptation aims to improve the generalization performance of a learning algorithm on a more realistic problem of open-set domain shift where the target data contains an additional unknown class that is not present in the source data. Most existing algorithms include two phases that can be described as closed-set domain adaptation given heuristic unknown class separation. Therefore, the generalization error cannot be strictly bounded due to the gap between the true distribution and samples inferred from heuristics. In this paper, we propose an end-to-end algorithm that tightly bound the risk of the entire target task by positive-unlabeled (PU) learning theory and the joint error from domain adaptation. Extensive experiments on various data sets demonstrate the effectiveness and efficiency of our proposed algorithm over open-set domain adaptation baselines.", + "bbox": [ + 259, + 310, + 743, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 518, + 375, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given enough annotated training data, deep learning models can significantly improve the performance across a wide variety of machine-learning tasks but usually cannot generalize well to new domains, as it is commonly assumed that the training and test data are drawn from the same distribution. In practice, however, this assumption can be violated by several factors, such as the change in light, noise, the angle at which the image is captured, and different types of sensors, which is referred to as the domain shift that can harm the performance when predicting the test data. As a solution, domain adaptation (DA) aims to transfer the knowledge learned from a source domain, which is typically fully labeled, into a different (although related) target domain.", + "bbox": [ + 212, + 551, + 787, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a more realistic setting [28], open-set domain adaptation allows the target data to contain an additional unknown category, covering all irrelevant classes not present in the source domain. The main idea of open-set unsupervised domain adaptation (OUDA) approaches [1, 4, 8, 17, 28] is to jointly learn a classifier from a hypothesis space for known and unknown classes in the source and target domains. According to [2], the target error can be bounded by the source error, the discrepancy distance between domains, the joint error coming from the conditional shift [38], and the open-set risk that indicates the precision of unknown class recognition. Open-set risk contributes substantially to the error", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bound when a large amount of target data is from the unknown class. However, the latter two terms cannot be explicitly computed.", + "bbox": [ + 212, + 146, + 784, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised domain adaptation (UDA) methods [9, 19-21, 27, 32] show effectiveness in adapting unlabeled data to new domains by distribution alignment, but can fail to learn discriminative class boundaries, especially when the domain shift is large. Most existing methods tend to ignore the joint error and only focus on minimizing the discrepancy between domains, where samples from different classes can be grouped if the domain shift is large enough. In that case, the joint error becomes non-negligible, and the target error cannot be strictly bounded [9, 38]. [37] provided a solution to this problem by incorporating the joint error into the target error upper bound in the unsupervised setting, and we generalize this idea towards the open-set domain adaptation setting.", + "bbox": [ + 212, + 180, + 787, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the promising performance, existing OUDA approaches [1, 3, 4, 8, 15, 17, 28, 34] lack an essential theoretical analysis of the generalization error for the target risk. Some works claim to derive a rigorous target error bound, while the minimization of joint error and the open-set risk are not theoretically guaranteed [7, 22]. Moreover, most existing algorithms can be described as closed-set UDA after separation of the unknown [3, 22, 34, 35]. In that case, the open-set risk cannot be controlled since the generalization error is also affected by the gap between empirical known target distribution and inferred known target samples, which becomes hard to analyze.", + "bbox": [ + 212, + 334, + 787, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1a illustrates the potential problems of existing methods. Generally, it is impossible to perfectly distinguish the unknown such that the following marginal distribution alignment for UDA part is conducted under a large label shit. In that case, the joint error would be crucial as there exists a trade-off between marginal discrepancy and joint error [38]. Even if somehow we manage to obtain a perfect separation, some target data class will be dragged towards the outside of the corresponding decision boundary. In addition, the unknown class cannot be grouped as a single cluster. Fig. 1b intuitively explains the core concept of our algorithm PUJE Sec. 3, where we introduce two additional task classifiers and minimize the predictive discrepancy between the classifiers with same color, such that the unknown class will be pushed away from source data and grouped as a single cluster. Furthermore, those target data outside the corresponding decision boundary will be aligned back to the correct places.", + "bbox": [ + 212, + 474, + 787, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OUDA problems can be considered as multi-class positive-unlabeled learning problems (MPU) [36] with domain shift, where all the shared classes appearing in both source and target domains are positive and the unknown class is negative. Therefore, to minimize the aforementioned joint error and open-set risk to achieve a tighter error bound for open-set adaptation, instead of unknown separation and closed-set UDA, we combine PU learning theory and joint error-based target error bound into an end-to-end learning framework. The following methods also introduced PU learning to tackle OUDA problems. [18] required the unsupervised source-like reconstruction for target data via encoder-decoder models whose performance cannot be guaranteed especially for large domain shift and complex dataset; [10] only dealt with the open-set label shift and assumed an identical", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17b323858c0514ade4d8d4ed8fcfad8dfee3f213db051ac5107f6b8b35596180.jpg", + "image_caption": [ + "Fig. 1: Intuitive explanation of the difference between our algorithm PUJE and existing methods. (a) existing methods do not explicitly minimize joint error and cannot group unknown class as a single cluster; (b) our proposal is an upper bound of joint error which can address large domain shift and group unknown class into a single cluster." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 743, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "class conditional distribution of each instance across domains; [35] tackled a slightly different task where the target data is changing over time and it did not cover joint error and the gap between the empirical and inferred known target distribution. Compared with previous works, the main contributions of this paper can be summarized as follows:", + "bbox": [ + 212, + 334, + 782, + 409 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an end-to-end algorithm with theoretical analysis for open-set problem setting by PU learning-induced joint error based target error bound;", + "- We define a novel discrepancy measurement, namely open-set margin discrepancy (OMD), which allows us to smoothly apply the strategy in PU learning to the target error bound in domain adaptation;", + "- Extensive experiments confirm the effectiveness and efficiency of the proposal." + ], + "bbox": [ + 230, + 419, + 785, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 530, + 380, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce notations, problem settings, and theoretical definitions for the tasks of closed-set unsupervised domain adaptation, open-set unsupervised domain adaptation, and multi-class positive and unlabeled learning.", + "bbox": [ + 212, + 560, + 787, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Closed-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of $n$ labeled points $\\hat{S} = \\{(x_s^i,y_s^i)\\in (\\mathcal{X}\\subseteq \\mathbb{R}^D\\times \\mathcal{Y} = \\{1,\\dots,K\\})\\}_{i = 1}^n$ from the source domain $S$ , and a set of $m$ unlabeled points $\\hat{T} = \\{(x_t^i)\\in \\mathcal{X}\\}_{i = 1}^m$ from a different target domain $T$ . With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier $f:\\mathcal{X}\\to \\mathcal{Y}$ .", + "bbox": [ + 212, + 614, + 787, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Multi-class Positive and Unlabeled Learning). For a classification task, the learning algorithm has access to a set of $n$ labeled points $\\hat{P} = \\{(x_p^i, y_p^i) \\in (\\mathcal{X} \\times \\mathcal{Y}' = \\{1, ..., K - 1\\})\\}_{i=1}^n$ from the positive domain $P$ , and a set of $m$ unlabeled points $\\hat{U} = \\{(x_u^i) \\in \\mathcal{X}\\}_{i=1}^m$ from the unlabeled domain $U$ , where $y_u^i \\in \\mathcal{Y}$ . With training samples drawn i.i.d from positive and unlabeled domains, the goal is to learn an optimal classifier $f: \\mathcal{X} \\to \\mathcal{Y}$ for the unlabeled domain. Here, $U$ includes samples from unknown class $K$ , but the class conditional distributions for known classes remain invariant between $P, U$ .", + "bbox": [ + 212, + 715, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Open-set DA via Joint Error based PU Learning", + "bbox": [ + 406, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3 (Open-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of $n$ labeled points $\\hat{S}' = \\{(x_{s'}^i, y_{s'}^i) \\in (\\mathcal{X} \\times \\mathcal{Y}')\\}_{i=1}^n$ from the incomplete source domain $S'$ , and a set of $m$ unlabeled points $\\hat{T} = \\{(x_t^i) \\in \\mathcal{X}\\}_{i=1}^m$ from a different target domain $T$ , where $y_t^i \\in \\mathcal{Y}$ . With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier $f: \\mathcal{X} \\to \\mathcal{Y}$ .", + "bbox": [ + 215, + 146, + 787, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show the relation between open-set domain adaptation and multi-class positive and unlabeled learning as follows, which includes the hits on how to apply the domain adaptation algorithm in the open-set scenario.", + "bbox": [ + 212, + 248, + 785, + 294 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In Definition 1, there is a domain shift between the distributions of the samples from the source and target domains such that a classifier trained on $S$ cannot generalize on $T$ . A typical solution is to align the feature distributions of both domains by a feature extractor [9,32].", + "- In Definition 2, the conditional distributions of the samples from known class are identical for positive and unlabeled domains. As proved in [36], the expected error on the unlabeled domain can be approximated by the expected error on the positive domain and the probability that the unlabeled sample has not been classified as unknown.", + "- In Definition 3, the incomplete source domain $S'$ can be regarded as the positive domain $P$ in MPU. The target domain $T$ can be treated as $U$ with a domain shift. Assume that we can bridge the feature distributions with UDA techniques, it becomes possible to address open-set problems with well-established positive and unlabeled (PU) learning theories." + ], + "bbox": [ + 230, + 301, + 782, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 PU Learning induced Joint Error based OUDA (PUJE)", + "text_level": 1, + "bbox": [ + 215, + 534, + 785, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, inspired by [37], we first rigorously reformulate the joint error-based target upper bound to facilitate the further derivation of a practical objective function for OUDA. To estimate the expectation over the complete source domain $S$ , we deploy the theory from multi-class positive and unlabeled learning to address the open-set problems where the source domain $S'$ is incomplete. We lastly show the entire loss of PUJE by combining the aforementioned joint error and PU learning.", + "bbox": [ + 212, + 565, + 785, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 (Approximated Joint Error based Target Upper Bound $^3$ ). Given the output space $\\mathcal{K} = \\{k|k\\in \\mathbb{R}^K:\\sum_{y\\in \\mathcal{Y}}k[y] = 1,k[y]\\in [0,1]\\}$ , let $f_{S},f_{T}:\\mathcal{X}\\to \\mathcal{K}$ be the true labeling functions for the source and target domains respectively, whose outputs are one-hot vectors denoting the corresponding labels of inputs. Let $\\epsilon :\\mathcal{K}\\times \\mathcal{K}\\rightarrow \\mathbb{R}$ denote a distance metric and $\\epsilon_D(f,f')\\coloneqq \\mathbb{E}_{x\\sim D}\\epsilon (f(x),f'(x))$ measure the expected disagreement between the outputs of $f,f':\\mathcal{X}\\to \\mathcal{K}$ over a distribution $D$ on $\\mathcal{X}$ . For $\\forall f_S^*\\in \\mathcal{H}_S\\subseteq \\mathcal{H},\\forall f_T^*\\in \\mathcal{H}_T\\subseteq \\mathcal{H},\\forall h\\in \\mathcal{H}:\\mathcal{X}\\to \\mathcal{K}$ where $h(x)[y]$ indicates the probability of $x\\in \\mathcal{X}$ labeled as $y\\in \\mathcal{V}$ , the expected target error is bounded by", + "bbox": [ + 215, + 680, + 787, + 816 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3 see proofs in the supplemental material", + "bbox": [ + 217, + 823, + 496, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {T} (h) \\leq \\epsilon_ {S} (h) + \\epsilon_ {T} \\left(f _ {S} ^ {*}, f _ {T} ^ {*}\\right) + \\epsilon_ {T} \\left(h, f _ {S} ^ {*}\\right) - \\left| \\epsilon_ {S} \\left(f _ {S} ^ {*}, f _ {T} ^ {*}\\right) - \\epsilon_ {S} \\left(h, f _ {T} ^ {*}\\right) \\right| + \\theta \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 171, + 787, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = 2 \\epsilon_ {T} \\left(f _ {S}, f _ {S} ^ {*}\\right) + \\epsilon_ {S} \\left(f _ {S}, f _ {S} ^ {*}\\right) + 2 \\epsilon_ {S} \\left(f _ {T} ^ {*}, f _ {T}\\right) + \\epsilon_ {T} \\left(f _ {T} ^ {*}, f _ {T}\\right) = \\theta_ {f _ {S}} + \\theta_ {f _ {T}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 190, + 785, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to Theorem 1, we show that the expected target error $\\epsilon_T(h) \\coloneqq \\epsilon_T(h, f_T)$ is bounded by the expected source error $\\epsilon_S(h) \\coloneqq \\epsilon_S(h, f_S)$ , the discrepancy between domains $[\\epsilon_T(f_S^*, f_T^*) + \\epsilon_T(h, f_S^*) - |\\epsilon_S(f_S^*, f_T^*) - \\epsilon_S(h, f_T^*)|]$ , and the deviation from true labeling functions $\\theta$ . The following Assumption 1 allows us to derive the generalization error of Theorem 1 such that we can ignore the residual term $\\theta$ and relate target error bound with the empirical estimation of source error and the discrepancy between domains.", + "bbox": [ + 212, + 215, + 787, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assumption 1. Let $\\hat{S},\\hat{T}$ denote empirical samples with finite size from source and target domains. Assume that there exists approximated labeling functions $f_{S}^{*},f_{T}^{*}$ such that the empirical deviation $\\hat{\\theta}_{f_S},\\hat{\\theta}_{f_T}$ are close enough to zero.", + "bbox": [ + 214, + 329, + 787, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 4 (Empirical Rademacher Complexity). Let $\\mathcal{F}$ be a class of real-valued functions: $\\mathcal{X} \\to \\mathbb{R}$ and $\\hat{D} = \\{x_1, \\ldots, x_m\\}$ a finite sample drawn i.i.d. from a distribution $D$ , the empirical Rademacher Complexity of $\\mathcal{F}$ is defined by", + "bbox": [ + 214, + 383, + 789, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathfrak {R}} _ {\\hat {D}} (\\mathcal {F}) = \\frac {1}{m} \\mathbb {E} _ {\\sigma} [ \\sup _ {f \\in \\mathcal {F}} \\sum_ {i = 1} ^ {m} \\sigma_ {i} f (x _ {i}) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 441, + 787, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_{i}$ is an independent uniform random variable taking values in $\\{-1, + 1\\}$ .", + "bbox": [ + 214, + 484, + 787, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2 (Generalization Error $^3$ ). Given Theorem 1 and Definition 4, Assumption 1 and function space $\\mathcal{F}_{f,f'}^{\\epsilon}:\\mathcal{X}\\to \\epsilon (f(X),f'(X))$ bounded by $M > 0$ , for any $\\delta >0$ , with probability at least $1 - 2\\delta$ , for $\\forall h\\in \\mathcal{H}$ :", + "bbox": [ + 214, + 508, + 789, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {T} (h) \\leq \\epsilon_ {\\hat {S}} (h) + \\epsilon_ {\\hat {T}} (f _ {S} ^ {*}, f _ {T} ^ {*}) + \\epsilon_ {\\hat {T}} (h, f _ {S} ^ {*}) - | \\epsilon_ {\\hat {S}} (f _ {S} ^ {*}, f _ {T} ^ {*}) - \\epsilon_ {\\hat {S}} (h, f _ {T} ^ {*}) | + 2 [ \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {h, f _ {S}} ^ {\\epsilon}) ] \\\\ + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {h, f _ {S} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {h, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\\epsilon}) (4) \\\\ + 2 \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\\epsilon}) + 2 \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\\epsilon}) ] + 3 M (5 \\sqrt {\\frac {\\log \\frac {9}{\\delta}}{2 m}} + 6 \\sqrt {\\frac {\\log \\frac {9}{\\delta}}{2 n}}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 577, + 787, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following part, we introduce the techniques from PU learning to estimate the expectation over source domain $S$ by the incomplete source domain $S'$ and target domain $T$ for the open-set scenario.", + "bbox": [ + 214, + 680, + 787, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 5 (Unknown Predictive Discrepancy). Let $v: \\mathcal{K} \\times \\mathcal{K} \\to \\mathbb{R}$ denote the Unknown Predictive Discrepancy as a distance metric and $v_{D}(f, f') := \\mathbb{E}_{x \\sim D} v(f(x), f'(x))$ measure the expected disagreement between the $K$ -th outputs of $f, f': \\mathcal{X} \\to \\mathcal{K}$ over a distribution $D$ on $\\mathcal{X}$ . Let $e^{K}: \\mathcal{X} \\to [0, \\dots, 1] \\in \\mathcal{K}$ denote a function that can predict any input as unknown. The deviation from $e^{K}$ for a hypothesis $h \\in \\mathcal{H}$ is referred to as the shorthand $v_{D}(h) := v_{D}(h, e^{K})$ that measures the probability that samples from $D$ have not been classified as unknown.", + "bbox": [ + 214, + 733, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Open-set DA via Joint Error based PU Learning", + "bbox": [ + 406, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assumption 2. Let $S^i = P_S(x|y = i), T^i = P_T(x|y = i)$ denote class conditional distributions, $S' = P_S(x|y \\neq K), T' = P_T(x|y \\neq K)$ indicate incomplete domains that do not contain unknown class $S^K, T^K$ . Given a feature extractor $g: \\mathcal{X} \\subseteq \\mathbb{R}^D \\to \\mathcal{Z} \\subseteq \\mathbb{R}^F$ , assume that the feature space can be aligned by UDA techniques such that $Z^K = P_{S^K}(z) = P_{T^K}(z), Z' = P_{S'}(z) = P_{T'}(z)$ .", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 1. Let $h \\in \\mathcal{H}_{F_g} : \\mathcal{Z} \\to \\mathcal{K}$ denote the decomposed hypothesis where $h \\circ g \\in \\mathcal{H}$ . Given Definition 5, Assumption 2, $v_{S^K}(h \\circ g) = v_{T^K}(h \\circ g)$ that represents probability that samples from $S^K$ have not been classified as unknown class, is equivalent to $\\epsilon_{S^K}(h \\circ g)$ that measures the classification error on unknown class according to the definition.", + "bbox": [ + 212, + 234, + 787, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 1 (Estimated Source Error $^3$ ). Let $\\sum_{i=1}^{K} \\pi_S^i = 1$ , $\\sum_{i=1}^{K} \\pi_T^i = 1$ denote the label distribution of $S$ and $T$ respectively. Given Proposition 1, the expected error on $S$ can be estimated by the error on $S'$ and Unknown Predictive Discrepancy (Definition 5) on $T$ with a mild condition that $\\pi_S^K = \\pi_T^K = 1 - \\alpha$ :", + "bbox": [ + 212, + 323, + 787, + 387 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {S} (h \\circ g) = \\alpha [ \\epsilon_ {S ^ {\\prime}} (h \\circ g) - v _ {S ^ {\\prime}} (h \\circ g) ] + v _ {T} (h \\circ g) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 414, + 787, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 1. According to Definition 5, minimizing $v_{T}(h \\circ g)$ means classifying target samples as the unknown class. In practice, a multiplier $\\beta < 1$ is applied on $v_{T}(h \\circ g)$ to prevent all target samples from being recognized as unknown.", + "bbox": [ + 212, + 444, + 787, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To reformulate the intractable discrepancy $\\epsilon_S(f_S^*,f_T^*)$ , $\\epsilon_S(h,f_T^*)$ in Theorem 1, we decompose the approximated labeling functions given feature extractor $g$ such that $f_{S}^{\\star}\\circ g = f_{S}^{\\star},f_{T}^{\\star}\\circ g = f_{T}^{\\star}$ where $f_{S}^{\\star}\\in \\mathcal{H}_{S_{g}}\\subseteq \\mathcal{H}_{F_{g}},f_{T}^{\\star}\\in \\mathcal{H}_{T_{g}}\\subseteq \\mathcal{H}_{F_{g}}$ . We further define $\\epsilon$ in a way such that, under some circumstances, it is equivalent to the Unknown Predictive Discrepancy for $\\forall f\\in \\mathcal{H}_{F_g}$", + "bbox": [ + 212, + 503, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {T ^ {K}} \\left(f \\circ g, f _ {T} ^ {\\star} \\circ g\\right) = v _ {T ^ {k}} \\left(f \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 595, + 787, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Corollary 1. Given Assumption 2, the feature extractor $g$ can align the feature distributions of source and target domain such that the expectation over $S^K$ is equal to that over $T^K$ , which leads Eq. (6) to: $\\epsilon_{S^K}(f \\circ g, f_T^\\star \\circ g) = \\epsilon_{T^K}(f \\circ g, f_T^\\star \\circ g) = v_{T^K}(f \\circ g, f_T^\\star \\circ g) = v_{S^K}(f \\circ g, f_T^\\star \\circ g)$ .", + "bbox": [ + 212, + 625, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 2 (Estimated Discrepancy $^3$ ). Given Corollary 1, analogous to Lemma 1, the discrepancy measured on source domain $\\epsilon_S(f_S^*, f_T^*)$ , $\\epsilon_S(h, f_T^*)$ can be estimated by the discrepancy on $S'$ and Unknown Predictive Discrepancy (Definition 5) with a mild condition that $\\pi_S^K = \\pi_T^K = 1 - \\alpha$ :", + "bbox": [ + 212, + 700, + 787, + 762 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\epsilon_ {S} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) & = \\alpha \\left[ \\epsilon_ {S ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {S ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] + v _ {T} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\\\ \\epsilon_ {S} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) & = \\alpha \\left[ \\epsilon_ {S ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {S ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] + v _ {T} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 786, + 815, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 6 (Overall Loss of PUJE). Let $\\hat{S}',\\hat{T}$ denote empirical data from $S',T$ . Given a feature extractor $g:\\mathcal{X}\\subseteq R^{D}\\to \\mathcal{Z}\\subseteq \\mathbb{R}^{F}$ , decomposed hypothesis and approximated labeling functions $h,f_{S}^{\\star},f_{T}^{\\star}\\in \\mathcal{H}_{F_{g}}:\\mathcal{Z}\\to \\mathcal{K}$ , replacing the intractable terms in Theorem 2 with Eqs. (5) and (7) and ignoring the Rademacher Complexity terms, the overall loss of PUJE can be regrouped as the sum of classification loss $L_{cls}$ and the discrepancy between domains $L_{dis}$ :", + "bbox": [ + 214, + 145, + 789, + 238 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {p u j e} = \\alpha \\left[ \\epsilon_ {\\hat {S} ^ {\\prime}} (h \\circ g) - v _ {\\hat {S} ^ {\\prime}} (h \\circ g) \\right] + \\beta v _ {\\hat {T}} (h \\circ g) + \\epsilon_ {\\hat {T}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) + \\epsilon_ {\\hat {T}} (h \\circ g, f _ {S} ^ {\\star} \\circ g) \\\\ - \\left| v _ {\\hat {T}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {\\hat {T}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) + \\alpha \\left[ \\epsilon_ {\\hat {S} ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right. \\right. \\\\ \\left. + v _ {\\hat {S} ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - \\epsilon_ {\\hat {S} ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {\\hat {S} ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] | \\\\ = L _ {c l s} (h; g) + L _ {d i s} \\left(f _ {S} ^ {\\star}, f _ {T} ^ {\\star}, h; g\\right) \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 260, + 821, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 354, + 380, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we first define Approximated Labeling Function Space $\\mathcal{H}_{S_g},\\mathcal{H}_{T_g}$ such that the intractable $L_{dis}(f_S^\\star ,f_T^\\star ,h;g)$ due to unknown $f_S^\\star ,f_T^\\star$ can be upper bounded by $\\sup_{f_S'\\in \\mathcal{H}_{S_g},f_T'\\in \\mathcal{H}_{T_g}}L_{dis}(f_S',f_T',h;g)$ . To fulfill Eq. (6), we propose Open-set Margin Discrepancy to quantify $\\epsilon$ that measures the disagreement between classifiers. We lastly show the entire training algorithm.", + "bbox": [ + 214, + 382, + 787, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Approximated Labeling Function Space", + "text_level": 1, + "bbox": [ + 214, + 479, + 588, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proposition 2. Let $\\mathcal{H}_S, \\mathcal{H}_T : \\mathcal{X} \\to \\mathcal{K}$ be two sets of functions that can minimize a part of the empirical residual $\\hat{\\theta}_S, \\hat{\\theta}_T$ respectively, $f_S^* \\in \\mathcal{H}_S$ , $f_T^* \\in \\mathcal{H}_T$ must hold as $f_S^*, f_T^*$ can minimize the entire $\\hat{\\theta}$ given Assumption 1. Accordingly, decomposed functions $f_S^*, f_T^*$ must lie in $\\mathcal{H}_{S_g} = \\{f|\\forall f \\circ g \\in \\mathcal{H}_S\\}$ , $\\mathcal{H}_{T_g} = \\{f|\\forall f \\circ g \\in \\mathcal{H}_T\\}$ such that a sufficient condition of the following inequality: $L_{dis}(f_S^*, f_T^*, h; g) \\leq \\sup_{f_S' \\in \\mathcal{H}_{S_g}, f_T' \\in \\mathcal{H}_{T_g}} L_{dis}(f_S', f_T', h; g)$ is fulfilled.", + "bbox": [ + 212, + 501, + 787, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 7 (Approximated Labeling Function Space). Let $L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T}$ denote a part of the empirical residual $\\hat{\\theta}_{fs}, \\hat{\\theta}_{f_T}$ respectively. Let $f_S^\\star \\circ g = f_S^* \\in \\mathcal{H}_S, f_T^\\star \\circ g = f_T^* \\in \\mathcal{H}_T$ denote the decomposed approximated labeling functions where $f_S^\\star \\in \\mathcal{H}_{S_g} \\subseteq \\mathcal{H}_{F_g}, f_T^\\star \\in \\mathcal{H}_{T_g} \\subseteq \\mathcal{H}_{F_g}$ . Given Proposition 2, Approximated Labeling Function Space $\\mathcal{H}_{S_g}, \\mathcal{H}_{T_g}$ can be defined as the sets whose members $f_S', f_T' \\in \\mathcal{H}_{F_g}$ can minimize $L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T}$ :", + "bbox": [ + 214, + 604, + 787, + 699 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathcal {H} _ {S _ {g}} = \\left\\{f _ {S} ^ {\\prime} \\right| \\arg \\min _ {g, f _ {S} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) := L _ {c l s} \\left(f _ {S} ^ {\\prime}; g\\right) \\} \\\\ \\mathcal {H} _ {T _ {g}} = \\left\\{f _ {T} ^ {\\prime} \\right| \\arg \\min _ {g, f _ {T} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) := (1 - \\gamma) L _ {c l s} \\left(f _ {T} ^ {\\prime}; g\\right) + \\gamma L _ {r e g} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 715, + 787, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{H}_{S_g}$ consists of functions minimizing the expected error on $S$ , which can be estimated by Lemma 1. To build a reliable function space $\\mathcal{H}_{T_g}$ without target labels, we approximate the target error with the weighted average $(\\gamma \\in [0,1])$ of error rate on labeled samples and a semi-supervised regularization term $L_{reg} = L_{ent} + L_{pse} + \\omega L_{con}$ .", + "bbox": [ + 214, + 763, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Open-set DA via Joint Error based PU Learning", + "bbox": [ + 405, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regularized Entropy Minimization As introduced in [11,27,31], we impose a class balance prior that can penalize classifiers with complex decision boundaries on entropy minimization [12] to yield a more sensible solution:", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {e n t} = - \\mathbb {E} _ {x \\in \\hat {T}} \\sum_ {y \\in \\mathcal {Y}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\log f _ {T} ^ {\\prime} (g (x)) [ y ] \\\\ + \\sum_ {y \\in \\mathcal {Y}} \\mathbb {E} _ {x \\in \\hat {T}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\log \\mathbb {E} _ {x \\in \\hat {T}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 218, + 787, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pseudo Labeling As introduced in [30, 31], for input $x \\in \\hat{T}$ and its random augmentation $x'$ [5], we minimize cross entropy for $x$ with pseudo labels of $x'$ :", + "bbox": [ + 215, + 286, + 785, + 316 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p s e} = - \\mathbb {E} _ {x \\in \\hat {T}} \\log f _ {T} ^ {\\prime} (g (x)) [ \\arg \\max _ {y \\in \\mathcal {Y}} h (g (x ^ {\\prime})) [ y ] ] \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 333, + 787, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consistency Regularization As introduced in [14, 29], we penalize the difference of the outputs for input $x \\in \\hat{T}$ and its random augmentation $x'$ :", + "bbox": [ + 215, + 364, + 787, + 396 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o n} = \\mathbb {E} _ {x \\in \\hat {T}} | f _ {T} ^ {\\prime} (g (x)) - f _ {T} ^ {\\prime} (g (x ^ {\\prime})) | \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 412, + 787, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Open-set Margin Discrepancy", + "text_level": 1, + "bbox": [ + 215, + 450, + 509, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Definition 8 (Induced Labeling Function). Let $f: \\mathcal{X} \\to \\mathcal{K}$ denote a multiclass labeling function, where $f(x)[y]$ indicating $y$ -th element of the output $f(x)$ for the probability of $x$ classified as $y$ . Thus an induced labeling function $l \\circ f: \\mathcal{X} \\to \\mathcal{Y}$ is given by: $l(f(x)) = \\arg \\max_{y \\in \\mathcal{Y}} f(x)[y]$ .", + "bbox": [ + 214, + 478, + 789, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Definition 9 (Open-set Margin Discrepancy (OMD)). Let $y = l(f(x)), y' = l(f'(x))$ denote the predictions on input $x$ from $f, f': \\mathcal{X} \\to \\mathcal{K}$ given Definition 8. Open-set Margin Discrepancy between $f, f'$ over a distribution $D$ is given by", + "bbox": [ + 215, + 551, + 792, + 598 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {D} (f, f ^ {\\prime}) = \\mathbb {E} _ {x \\sim D} [ \\operatorname {o m d} (f (x), f ^ {\\prime} (x)) ] (13) \\\\ \\operatorname {o m d} (f (x), f ^ {\\prime} (x)) = \\max \\left(\\left| \\log (1 - f (x) [ y ]) - \\log (1 - f ^ {\\prime} (x) [ y ]) \\right|\\right), \\\\ \\left. \\left| \\log \\left(1 - f (x) \\left[ y ^ {\\prime} \\right]\\right) - \\log \\left(1 - f ^ {\\prime} (x) \\left[ y ^ {\\prime} \\right]\\right) \\right|\\right) (14) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 625, + 787, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For OUDA, generally the classifier is not discriminative for unknown class $K$ due to the lack of labeled samples to reduce the misclassification rate. To address this, we model $\\log (1 - f(x)[K])$ that has a more stable gradient.", + "bbox": [ + 214, + 691, + 785, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark 2. According to Definition 5, $v_{D}(f,f^{\\prime})$ measures the expectation of disagreement on the K-th output between two functions $f,f^{\\prime}:\\mathcal{X}\\to \\mathcal{K}$ over a distribution $D$ , which can also indicate the difference between probabilities of categorizing inputs as known classes. To fulfill Eq. (6), we quantify $v_{D}(f,f^{\\prime})$ as:", + "bbox": [ + 214, + 747, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nv _ {D} (f, f ^ {\\prime}) = \\mathbb {E} _ {x \\sim D} | \\log (1 - f (x) [ K ]) - \\log (1 - f ^ {\\prime} (x) [ K ]) | \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 824, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Algorithm 1 PUJE", + "text_level": 1, + "bbox": [ + 217, + 146, + 367, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Input: incomplete source data $\\hat{S}^{\\prime}$ ; unlabeled target data $\\hat{T}$", + "bbox": [ + 230, + 167, + 625, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Output: labeling functions $f_S', f_T'$ ; feature extractor $g$ ; hypothesis $h$", + "bbox": [ + 230, + 181, + 686, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Parameter: trade-off parameter $\\lambda$ ; learning rate $\\eta$ ; known class ratio estimator $\\alpha$ for epoch = 1, 2, ... do", + "bbox": [ + 230, + 196, + 776, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Step 1: estimate known class ratio $\\alpha$ on $T$ with prediction of $g, h$", + "bbox": [ + 254, + 223, + 692, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Step 2: optimize $g, f_S', f_T'$ to satisfy the approximated labeling function space", + "bbox": [ + 254, + 237, + 772, + 251 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n(g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) \\leftarrow (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) + \\eta \\Delta (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime})\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 251, + 531, + 265 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\varDelta (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) = - \\frac {\\partial (L _ {\\mathcal {H} _ {S}} (f _ {S} ^ {\\prime} ; g) + L _ {\\mathcal {H} _ {T}} (f _ {T} ^ {\\prime} ; g))}{\\partial (g , f _ {S} ^ {\\prime} , f _ {T} ^ {\\prime})}\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 265, + 532, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Step 3: maximize the discrepancy w.r.t. $f_{S}^{\\prime}, f_{T}^{\\prime}$ within the function space", + "bbox": [ + 254, + 287, + 741, + 300 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right) \\leftarrow \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right) + \\eta \\Delta \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 300, + 486, + 314 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\varDelta(f_{S}^{\\prime},f_{T}^{\\prime})=-\\frac{\\partial(L_{\\mathcal{H}_{S}}(f_{S}^{\\prime};g)+L_{\\mathcal{H}_{T}}(f_{T}^{\\prime};g)-\\lambda L_{dis}(f_{S}^{\\prime},f_{T}^{\\prime},h;g))}{\\partial(f_{S}^{\\prime},f_{T}^{\\prime})}\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 314, + 622, + 337 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Step 4: minimize the entire target error bound w.r.t. $g, h$ for fixed $f_S', f_T'$ .", + "bbox": [ + 254, + 335, + 741, + 349 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n(g, h) \\leftarrow (g, h) + \\eta \\Delta (g, h)\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 349, + 437, + 364 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta (g, h) = - \\frac {\\partial \\left(L _ {c l s} (h ; g) + \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime} , f _ {T} ^ {\\prime} , h ; g\\right)\\right)}{\\partial (g , h)}\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 364, + 524, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 230, + 382, + 287, + 392 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Training Algorithm", + "text_level": 1, + "bbox": [ + 215, + 425, + 421, + 440 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given the feature extractor $g: \\mathcal{X} \\subseteq \\mathbb{R}^D \\to \\mathcal{Z} \\subseteq \\mathbb{R}^F$ and hypotheses $h, f_S', f_T' \\in \\mathcal{H}_{F_g}: \\mathcal{Z} \\to \\mathcal{K}$ , we introduce a trade-off parameter $\\lambda$ to balance the classification loss and discrepancy as Eq. (16). During each epoch, we first estimate the known class ratio $\\alpha$ (initially set to 1) in target data by the predictions of $h$ . Then we train the hypothesis $h$ , approximated labeling functions $f_S', f_T'$ , feature extractor $g$ to minimize the error on labeled data $L_{cls}$ with hypothesis space constraints $L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T}$ , while optimizing the discrepancy $L_{dis}$ adversarially by the min-max game over $g$ and $f_S', f_T'$ (Algorithm 1):", + "bbox": [ + 212, + 449, + 787, + 571 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\min _ {f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) + L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) - \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}, h; g\\right) \\right. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 593, + 785, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\lfloor \\min _ {h \\in \\mathcal {H} _ {F _ {g}}, g} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) + L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) + L _ {c l s} (h; g) + \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}, h; g\\right) \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 618, + 732, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Intuition", + "text_level": 1, + "bbox": [ + 215, + 664, + 334, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we intuitively explain how our method can align the source and target domains while separating the unknown in Fig. 2. According to [24], unlike the entropy minimization that tends to cut through unlabeled data, consistency regularization helps to draw a manifold-aware decision boundary. Based on this conclusion, the decision boundary of $f_{T}^{\\prime}$ induced by consistency regularization $L_{con}$ can preserve the cluster structure (Fig. 2a). According to Algorithm 1, $f_{S}^{\\prime}$ tries to maximize $L_{dis}$ while classify target samples as unknown class to minimize $L_{\\mathcal{H}_S}(f_S';g)$ (Definition 7 and Lemma 1), which can lead to a new decision boundary shown in Fig. 2b. Meanwhile, feature extractor $g$ tries to minimize $L_{dis}$ by pushing target samples close to $f_{S}^{\\prime}$ towards source clusters,", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Open-set DA via Joint Error based PU Learning", + "bbox": [ + 406, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/27899cb4ba862fd93fb884d91e7c91c655ce01785873ac7abe2c8609ed4c46ae.jpg", + "image_caption": [ + "(a) $\\min_{f_T'} L_{\\mathcal{H}_T}$" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 377, + 257 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/67564d6910b9b0116b35c85a28e1451426b0f91251ff90ed0a78e0bb78b10f4d.jpg", + "image_caption": [ + "(b) $\\min_{f_S'}(L_{\\mathcal{H}_S} - L_{dis})$" + ], + "image_footnote": [], + "bbox": [ + 380, + 142, + 537, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2dd6be1cf92e80d3475bcda7814410c86f3dfef6fd68d894034898d11410acb7.jpg", + "image_caption": [ + "(c) $\\min_g L_{dis}$", + "Fig. 2: Intuitive explanation of the mechanism of the proposed PUJE in Algorithm 1. (a) $L_{\\mathcal{H}_T}$ includes consistency regularization $L_{con}$ that helps to draw a manifold-aware decision boundary for $f_T'$ (Step 2); (b) $f_S'$ can increase $L_{dis}$ by classifying most of the unlabeled target samples as unknown, which also decreases $L_{\\mathcal{H}_S}$ (Step 3); (c) feature extractor $g$ pushes unlabeled target samples close to the decision boundary of $f_S'$ towards source cluster, while those close to $f_T'$ towards outside to reduce $L_{dis}$ , which can lead to a separated unknown cluster (Step 4)." + ], + "image_footnote": [ + "$L_{dis}$ $f_T^{\\prime}$ $f_S^{\\prime}$", + "- Positive Source\n- Negative Source\n- Unknown Target\n- Pseudo Positive\n- Pseudo Negative" + ], + "bbox": [ + 537, + 145, + 663, + 257 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "while those close to $f_{T}^{\\prime}$ towards outside such that unlabeled target samples far from source clusters can be eventually separated and grouped into a new cluster for the unknown progressively as illustrated in Fig. 2c.", + "bbox": [ + 212, + 414, + 784, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 484, + 356, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluated our proposal on two benchmarks, Office-Home and Syn2Real-O. In the implementation, the $L_{reg}$ weight $\\gamma$ and the trade-off parameter $\\lambda$ were set to 0.1 and 0.01 according to [37]. In addition, we empirically set PU coefficient $\\beta$ to 0.03 and consistency coefficient $\\omega$ to 20 for Office-Home and 5 for Syn2Real-O. All experiments were conducted with the ImageNet [6] pre-trained ResNet-50 [13] as the feature extractor $g$ and 2-layer linear networks for classifiers $f_S', f_T', h$ . We trained the model by Stochastic Gradient Descent optimizer with a 0.001 learning rate annealed according to [9], 24 batch-size, a momentum of 0.9. We quantitatively compare our results against various OUDA baselines, including OSBP [28], STA [17], DAOD [8], PGL [22], ROS [3], OSLPP [34] and ANNA [15].", + "bbox": [ + 212, + 518, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics To evaluate the proposed method and the baselines, we utilize the widely used measures [22, 28], i.e., normalized accuracy for all classes (OS), normalized accuracy for the known classes only $(\\mathrm{OS}^{\\star})$ and harmonic mean $\\mathrm{HOS} = 2(\\mathrm{OS}^{\\star}\\times \\mathrm{UNK}) / (\\mathrm{OS}^{\\star} + \\mathrm{UNK})$ [3,15,18,34].", + "bbox": [ + 212, + 686, + 787, + 750 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Office-Home [33] is a widely-used domain adaptation benchmark, which consists of 15,500 images from 65 categories and four domains: Art (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw). Following the same splits used in previous methods [22], we select the first 25 classes in alphabetical order as the known classes and group the rest as the unknown.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c93574545a28fb067e293ef80ae8f8d455aadc4a915fe544c3c2fd88b10564eb.jpg", + "table_caption": [ + "Table 1: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (OS)" + ], + "table_footnote": [], + "table_body": "
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP47.165.364.638.348.748.327.081.679.537.173.572.129.355.354.337.781.980.2
STA77.048.449.595.440.842.959.177.376.671.268.668.761.045.446.058.974.573.9
DAOD44.367.766.844.760.359.740.885.083.349.873.272.347.460.459.956.882.881.8
PGL34.773.772.238.459.258.427.684.882.66.181.578.625.166.865.038.084.883.0
PUJE38.474.873.442.264.463.637.987.785.823.481.078.839.271.270.039.188.286.3
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP28.657.256.125.877.875.823.085.483.033.077.275.516.771.369.232.165.964.6
STA64.145.946.662.067.267.066.276.676.257.465.264.960.257.657.772.749.350.2
DAOD71.155.556.166.669.269.163.779.378.754.878.277.354.670.269.655.162.962.6
PGL19.163.361.632.178.977.140.987.785.95.385.982.824.573.972.033.870.268.8
PUJE57.166.666.352.182.881.645.288.987.255.981.880.851.481.580.342.571.069.9
", + "bbox": [ + 217, + 180, + 785, + 327 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ad919c5ac303e819cbc74c22204d829156213168786e51ba5d0d776311935742.jpg", + "table_caption": [ + "Table 2: Accuracy of ResNet-50 model fine-tuned on Syn2Real-O dataset. * indicates our re-implementation with the officially released code." + ], + "table_footnote": [], + "table_body": "
METHODplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckUNKOS*OSHOS
OSBP73.657.958.265.267.429.584.147.067.85.789.10.566.653.854.859.5
STA64.170.353.759.480.820.890.012.563.230.278.22.759.152.252.755.4
PGL81.568.374.260.691.945.492.241.087.967.579.26.449.666.865.556.9
PGL*81.377.566.571.790.635.792.441.282.646.982.34.959.664.564.162.0
ANNA41.254.639.759.451.225.682.654.878.910.576.62.067.448.149.656.1
PUJE90.773.875.172.477.833.891.258.879.546.584.95.169.065.866.067.3
", + "bbox": [ + 217, + 407, + 785, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Syn2Real-O [26] is a more challenging synthetic-to-real benchmark, which is constructed from the VisDA-17 [25]. The Syn2Real-O dataset significantly increases the ratio of unknown data in the target domain to 0.9 by introducing additional unknown samples. In this dataset, the source domain contains training data from the VisDA-17 as the known set, and the target domain includes the test data from the VisDA-17 (known set) plus 50k images from irrelevant categories of MSCOCO [16] dataset (unknown set).", + "bbox": [ + 217, + 535, + 785, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As reported in Tabs. 1 and 2, we observe that our method PUJE consistently outperforms the state-of-the-art results, improving OS and HOS by $3.0\\%$ and $5.3\\%$ on the benchmark datasets of Office-Home and Syn2Real-O respectively. Note that our proposed approach provides significant performance gains for the more challenging dataset of Syn2Real-O that requires knowledge transfer across different modalities. In the sub-tasks with a larger domain shift, e.g., $\\mathrm{Rw} \\rightarrow \\mathrm{Cl}$ and $\\mathrm{Pr} \\rightarrow \\mathrm{Cl}$ in Office-Home, we can observe similar results demonstrating the strong adaptation ability of the proposed framework. PGL reported a higher OS score in Syn2Real-O, but we cannot reproduce the results with the officially released code. During the re-implementation based on their early stop strategy, we found the UNK was around $30\\%$ , which is far below the reported score. A potential reason to explain this is that maybe they did not remove the unknown category from the source data.", + "bbox": [ + 217, + 643, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Open-set DA via Joint Error based PU Learning", + "bbox": [ + 406, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/22fc7f5120fbf46d1e5d2e6bb70e4645304e492da73720d3c40dad39c3cbb96f.jpg", + "table_caption": [ + "Table 3: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (HOS)" + ], + "table_footnote": [], + "table_body": "
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOS
ROS64.357.360.671.246.556.378.470.874.470.867.068.8
OSLPP76.254.663.667.153.159.371.277.074.075.060.867.2
ANNA70.363.066.574.854.663.178.974.376.677.366.171.3
PUJE77.856.965.776.851.261.575.280.177.567.371.169.1
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*
ROS74.150.660.170.368.469.377.275.876.572.265.368.6
OSLPP67.155.961.073.172.572.869.480.174.373.967.266.9
ANNA78.761.469.079.968.373.779.774.176.880.266.973.0
PUJE83.558.568.878.979.679.376.382.179.180.671.275.6
", + "bbox": [ + 217, + 180, + 785, + 308 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d8cb973b1306fe0394dbed14ea61becd07d57074a5569d35668bd143d07413e0.jpg", + "table_caption": [ + "Table 4: Ablation studies of semi-supervised regularization losses. We report the Accuracy $(\\%)$ on Office-Home of $\\mathrm{A}\\rightarrow \\mathrm{R}$ and $\\mathrm{C}\\to \\mathrm{P}$ using a ResNet50 backbone." + ], + "table_footnote": [], + "table_body": "
Config.Ar→PrCl→Rw
UNK OS* OSUNK OS* OS
default52.182.881.655.981.880.8
w/o Lpse69.678.177.864.877.877.3
w/o Lcon26.381.779.633.380.979.1
w/o Lent50.281.179.953.980.579.5
", + "bbox": [ + 361, + 373, + 635, + 462 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "HOS Metric Instead of OS score, [18] proposed the harmonic mean of $\\mathrm{OS}^{\\star}$ and unk, $\\mathrm{HOS} = 2(\\mathrm{OS}^{\\star} \\times \\mathrm{UNK}) / (\\mathrm{OS}^{\\star} + \\mathrm{UNK})$ to penalize large gaps between $\\mathrm{OS}^{\\star}$ and unk. We agree that the OS score does not reflect UNK when the number of known classes grows. However, the HOS score is still unfair. $\\mathrm{OS}^{\\star}$ and UNK are equally valued in HOS, but UNK is much easier to improve than $\\mathrm{OS}^{\\star}$ which depends on the power of the adaptation algorithm. [15] utilizes a weak adaptation algorithm based on [28] that leads to a relatively loose alignment (Fig. 4) on the feature space of source and target domains such that unlabeled data is more frequently assigned to the unknown class. We believe that if a method can achieve a high $\\mathrm{OS}^{\\star}$ score, it is possible to obtain a decent HOS score by tuning hyper-parameters or additional losses. For our proposal, we can increase $\\beta = 0.25$ to improve UNK at the sacrifice of $\\mathrm{OS}^{\\star}$ (Tab. 3).", + "bbox": [ + 212, + 489, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Study To investigate the impact of the semi-supervised regularization, we compare three variants of the PUJE model on the Office-Home dataset shown in Tab. 4. Without $L_{pse}$ , $\\mathrm{OS}^{\\star}$ decreases because it can no longer leverage the pseudo labels that can help to build reliable hypothesis space $\\mathcal{H}_{T_g}$ (Definition 7). Without $L_{con}$ , UNK drops as it is not likely to draw a manifold-aware decision boundary that is beneficial to the separation of unknown data (Fig. 2a).", + "bbox": [ + 212, + 686, + 787, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Robustness against Varying Openness To verify the robustness of the proposed PUJE, we conducted experiments on the Syn2Real-O with the openness varying in $\\{0.25, 0.5, 0.75, 0.9\\}$ . Here, openness is defined by the ratio of unknown", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9f623d8a8102881900f455a6d93b300329f2c188562ea187d1995eb93e2f1a24.jpg", + "image_caption": [ + "(a) PU Coefficient" + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 403, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4a0b854f463874efefecc0c4a3413973e7bfba59c6490d606dc1dc83bac6687a.jpg", + "image_caption": [ + "(b) SSL Coefficient" + ], + "image_footnote": [], + "bbox": [ + 413, + 148, + 589, + 252 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/70c99c257b1de897960d3c2589029d98b9bfb0dd150cea78234988c9b4aa9322.jpg", + "image_caption": [ + "(c) Openness" + ], + "image_footnote": [], + "bbox": [ + 599, + 148, + 776, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/91fdbbd4da8cbf009efa2f4814fa03746412ba4ff43bd324322679c5de6f2ac8.jpg", + "image_caption": [ + "(d) Trade-off in $\\mathrm{Ar}\\rightarrow \\mathrm{Cl}$" + ], + "image_footnote": [], + "bbox": [ + 228, + 276, + 403, + 380 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a2ed7e612b215ec82d0d1fd5bd51401cad5048f151cfdf52cfebd24bacc35099.jpg", + "image_caption": [ + "(e) Convergence in $\\mathrm{Ar}\\rightarrow \\mathrm{Cl}$" + ], + "image_footnote": [], + "bbox": [ + 413, + 275, + 589, + 380 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5441e85fb22f0abe9a4774e6021834e77cc845b6e0979bbd0bc3acc26c3b8b2c.jpg", + "image_caption": [ + "(f) Convergence in $\\mathrm{Rw}\\rightarrow \\mathrm{Cl}$", + "Fig. 3: (a)-(b): sensitivity to varying loss coefficient $\\beta, \\omega$ verified in $\\mathrm{Ar} \\rightarrow \\mathrm{Pr}$ and $\\mathrm{Cl} \\rightarrow \\mathrm{Rw}$ tasks; (c): performance comparisons w.r.t. varying openness of the Syn2Real-O task; (d): training cure of the proposed algorithm that characterizes a trade-off between the classification accuracy on known classes and the unknown class; (e)-(f): convergence analysis of the Office-Home task compared to other baselines with confidence intervals." + ], + "image_footnote": [], + "bbox": [ + 599, + 276, + 776, + 380 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "samples in the entire target data. We show the results of PGL and PUJE in Fig. 3c. Note that the PGL approach heuristically sets the hyper-parameter according to the true unknown ratio to control the openness, while PUJE automatically estimates the weight $\\alpha$ during the training procedure. We observe that PUJE consistently outperforms the baseline by a large margin, especially on unknown recognition, which confirms its robustness to the change in openness.", + "bbox": [ + 212, + 508, + 787, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training Curve In Fig. 3d, we illustrate the recognition performance of PUJE over training steps in the $\\mathrm{Ar} \\rightarrow \\mathrm{Cl}$ task. All metrics show a performance gain over the first several steps. Then, the normalized accuracy $\\mathrm{OS}^{\\star}$ experiences a downward while the accuracy for the unknown class keeps improving, which characterizes a trade-off between $\\mathrm{OS}^{\\star}$ and UNK. In addition, the behavior of $f_{S}^{\\prime}, f_{T}^{\\prime}$ meets the expectation of our alignment mechanism where $f_{S}^{\\prime}$ starts to detect unknown data, which is followed by $f_{T}^{\\prime}$ (Fig. 2b). From Figs. 3e and 3f, we observe that some previous works [15,22] do not converge at the reported optimum, and they tackle this problem by ceasing the model updates before the score starts to decline. However, this early stop strategy can fail without enough labeled target data to validate the current performance. In contrast, our method can reach a reliable convergence.", + "bbox": [ + 212, + 613, + 789, + 796 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sensitivity to PU & SSL Coefficient We show the sensitivity of our approach to various PU and SSL loss coefficient $\\beta, \\omega$ . We tested the value of $\\beta$ from [0.01, 0.5]", + "bbox": [ + 214, + 808, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Open-set DA via Joint Error based PU Learning", + "bbox": [ + 406, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9ff734422a3e51cc04995add15f161b0c758ecc1876ef43b88b339e080dfeb75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 143, + 362, + 220 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1dc159318049829926bfd8c878ddb1b3c91617cf4b36aff3e011932ecac04b31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 145, + 496, + 220 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4e1db5e706e0f59f9854bce07b19bda165d78dfccb2ed755691039c74c9939e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 633, + 220 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e769a3fb4f133ecd3154fde9819ce3329846d8a94ee3522faa23d8ade5a02d41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 146, + 767, + 220 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/42b22b296ec64a08cf44f844076cc89671fc3e60cdc868bf1447f3765c44ef86.jpg", + "image_caption": [ + "(a) OSBP", + "(e) OSBP", + "Fig. 4: Feature space visualized by t-SNE in (a)-(d): Office-Home; (e)-(h): Syn2Real-O." + ], + "image_footnote": [], + "bbox": [ + 232, + 239, + 361, + 315 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/48ab1a9f47aa3e015200d4c8e50f058504a6cd167c7369bdcfec0c79066ec9bf.jpg", + "image_caption": [ + "(b) PGL", + "(f) PGL" + ], + "image_footnote": [], + "bbox": [ + 370, + 239, + 496, + 315 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2763721c66614cb069558809ebe9e454e7ef87ea3fe187ab5e56815d0a9e14ac.jpg", + "image_caption": [ + "(c) ANNA", + "(g) ANNA" + ], + "image_footnote": [], + "bbox": [ + 504, + 239, + 632, + 315 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2a2f759faff32a31f774512916fb1891f8a590265e81e170c81277d7181e42c4.jpg", + "image_caption": [ + "(d) PUJE", + "(h) PUJE" + ], + "image_footnote": [], + "bbox": [ + 640, + 239, + 767, + 315 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "in $\\mathrm{Ar}\\rightarrow \\mathrm{Pr}$ task and $\\omega$ from [5, 25] in $\\mathrm{Cl}\\rightarrow \\mathrm{Rw}$ task while fixing other parameters to the default setting. We can draw two observations from Figs. 3a and 3b: the OS score is relatively stable, and the unknown recognition achieves a reliable performance when coefficient $\\beta$ is within the interval of [0.02, 0.05]; Generally, a larger $\\omega$ , which is set for lower openness, will lead to a better performance on recognizing unknown data while hurt the accuracy for known classes.", + "bbox": [ + 212, + 396, + 787, + 488 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Feature Space Visualization To intuitively visualize the effectiveness of OUDA approaches, we extracted features from the baseline models (OSBP, PGL, ANNA) and our proposed model on the $\\mathrm{Ar}\\rightarrow \\mathrm{Cl}$ task (Office-Home) and Syn2Real-O task with the ResNet-50 backbone [13]. The feature distributions were processed with t-SNE [23] afterward. As shown in Fig. 4, compared with baselines, our method PUJE achieves a better alignment between source and target distributions, especially when the domain shift is large. Benefiting from our joint error-based adversarial alignment mechanism, the extracted feature space, including the cluster of unknown target data, has a more discriminative class-wise decision boundary.", + "bbox": [ + 212, + 520, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 713, + 359, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have addressed the open-set domain shift problem by introducing a novel learning theory based on multi-class positive-unlabeled learning that can reduce the open-set risk and the joint error. Experiments show that our proposed learning theory performs consistently better on challenging object recognition benchmarks for open-set adaptation with significant domain gaps and label shifts.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo.", + "bbox": [ + 212, + 174, + 787, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 256 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baktashmotlagh, M., Faraki, M., Drummond, T., Salzmann, M.: Learning factorized representations for open-set domain adaptation. In: ICLR (2019)", + "2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.: A theory of learning from different domains. Machine Learning 79, 151-175 (2010)", + "3. Bucci, S., Loghmani, M.R., Tommasi, T.: On the effectiveness of image rotation for open set domain adaptation. In: ECCV (2020)", + "4. Busto, P.P., Gall, J.: Open set domain adaptation. In: ICCV (2017)", + "5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: CVPRW (2020)", + "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009)", + "7. Fang, Z., Lu, J., Liu, F., Xuan, J., Zhang, G.: Open set domain adaptation: Theoretical bound and algorithm. arXiv abs/1907.08375 (2019)", + "8. Feng, Q., Kang, G., Fan, H., Yang, Y.: Attract or distract: Exploit the margin of open set. In: ICCV (2019)", + "9. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of Machine Learning Research 17(1), 2096-2030 (2016)", + "0. Garg, S., Balakrishnan, S., Lipton, Z.C.: Domain adaptation under open set label shift. In: NeurIPS (2022)", + "1. Gomes, R., Krause, A., Perona, P.: Discriminative clustering by regularized information maximization. In: NIPS (2010)", + "2. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NIPS (2005)", + "3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. IEEE Conference on Computer Vision and Pattern Recognition pp. 770-778 (2015)", + "4. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)", + "5. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR (2023)", + "6. Lin, T.Y., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014)", + "7. Liu, H., Cao, Z., Long, M., Wang, J., Yang, Q.: Separate to adapt: Open set domain adaptation via progressive separation. In: CVPR (2019)", + "8. Loghmania, M.R., Vinczea, M., Tommasi, T.: Positive-unlabeled learning for open set domain adaptation. Pattern Recognition Letters 136, 198-204 (2020)", + "9. Long, M., Cao, Y., Wang, J., Jordan, M.I.: Learning transferable features with deep adaptation networks. In: Proceedings of the 32nd International Conference on International Conference on Machine Learning. vol. 37, pp. 97-105. JMLR.org (2015)" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Open-set DA via Joint Error based PU Learning", + "bbox": [ + 406, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Long, M., CAO, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. In: Advances in Neural Information Processing Systems, pp. 1640-1650. Curran Associates, Inc. (2018)", + "21. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 2208-2217. JMLR.org (2017)", + "22. Luo, Y., Wang, Z., Huang, Z., Baktashmotlagh, M.: Progressive graph learning for open-set domain adaptation. In: ICML (2020)", + "23. van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008)", + "24. Oliver, A., Odena, A., Raffel, C., Cubuk, E.D., Goodfellow, I.J.: Realistic evaluation of deep semi-supervised learning algorithms. In: NeurIPS (2018)", + "25. Peng, X., Usman, B., Kaushik, N., Hoffman, J., Wang, D., Saenko, K.: Visda: The visual domain adaptation challenge. arXiv abs/1710.06924 (2017)", + "26. Peng, X., Usmana, B., Saito, K., Kaushik, N., Hoffman, J., Saenko, K.: Syn2real: A new benchmark for synthetic-to-real visual domain adaptation. arXiv abs/1806.09755 (2018)", + "27. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3723-3732 (2017)", + "28. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: ECCV (2018)", + "29. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NIPS (2016)", + "30. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NIPS (2020)", + "31. Tang, H., Chen, K.C., Jia, K.: Unsupervised domain adaptation via structurally regularized deep clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (2020)", + "32. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 2962-2971 (2017)", + "33. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 5385-5394 (2017)", + "34. Wang, Q., Meng, F., Breckon, T.P.: Progressively select and reject pseudo-labelled samples for open-set domain adaptation. In: AAAI (2022)", + "35. Wu, J., He, J.: Domain adaptation with dynamic open-set targets. In: KDD (2022)", + "36. Xu, Y., Xu, C., Xu, C., Tao, D.: Multi-positive and unlabeled learning. In: IJCAI (2017)", + "37. Zhang, D., Westfechtel, T., Harada, T.: Unsupervised domain adaptation via minimized joint error. In: TMLR (2023)", + "38. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Proceedings of Machine Learning Research. vol. 97, pp. 7523-7532. PMLR (2019)" + ], + "bbox": [ + 215, + 146, + 787, + 784 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D.Zhang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_model.json b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89941627367314ffd4dbefd35a015df4a2a89a43 --- /dev/null +++ b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_model.json @@ -0,0 +1,2619 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.217, + 0.141, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.213, + 0.726, + 0.229 + ], + "angle": 0, + "content": "Dexuan Zhang\\(^{1}\\), Thomas Westfechtel\\(^{1}\\), and Tatsuya Harada\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.24, + 0.593, + 0.268 + ], + "angle": 0, + "content": "1 The University of Tokyo \n2 RIKEN" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.311, + 0.744, + 0.492 + ], + "angle": 0, + "content": "Abstract. Open-set domain adaptation aims to improve the generalization performance of a learning algorithm on a more realistic problem of open-set domain shift where the target data contains an additional unknown class that is not present in the source data. Most existing algorithms include two phases that can be described as closed-set domain adaptation given heuristic unknown class separation. Therefore, the generalization error cannot be strictly bounded due to the gap between the true distribution and samples inferred from heuristics. In this paper, we propose an end-to-end algorithm that tightly bound the risk of the entire target task by positive-unlabeled (PU) learning theory and the joint error from domain adaptation. Extensive experiments on various data sets demonstrate the effectiveness and efficiency of our proposed algorithm over open-set domain adaptation baselines." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.519, + 0.377, + 0.535 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Given enough annotated training data, deep learning models can significantly improve the performance across a wide variety of machine-learning tasks but usually cannot generalize well to new domains, as it is commonly assumed that the training and test data are drawn from the same distribution. In practice, however, this assumption can be violated by several factors, such as the change in light, noise, the angle at which the image is captured, and different types of sensors, which is referred to as the domain shift that can harm the performance when predicting the test data. As a solution, domain adaptation (DA) aims to transfer the knowledge learned from a source domain, which is typically fully labeled, into a different (although related) target domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As a more realistic setting [28], open-set domain adaptation allows the target data to contain an additional unknown category, covering all irrelevant classes not present in the source domain. The main idea of open-set unsupervised domain adaptation (OUDA) approaches [1, 4, 8, 17, 28] is to jointly learn a classifier from a hypothesis space for known and unknown classes in the source and target domains. According to [2], the target error can be bounded by the source error, the discrepancy distance between domains, the joint error coming from the conditional shift [38], and the open-set risk that indicates the precision of unknown class recognition. Open-set risk contributes substantially to the error" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "bound when a large amount of target data is from the unknown class. However, the latter two terms cannot be explicitly computed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Unsupervised domain adaptation (UDA) methods [9, 19-21, 27, 32] show effectiveness in adapting unlabeled data to new domains by distribution alignment, but can fail to learn discriminative class boundaries, especially when the domain shift is large. Most existing methods tend to ignore the joint error and only focus on minimizing the discrepancy between domains, where samples from different classes can be grouped if the domain shift is large enough. In that case, the joint error becomes non-negligible, and the target error cannot be strictly bounded [9, 38]. [37] provided a solution to this problem by incorporating the joint error into the target error upper bound in the unsupervised setting, and we generalize this idea towards the open-set domain adaptation setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Despite the promising performance, existing OUDA approaches [1, 3, 4, 8, 15, 17, 28, 34] lack an essential theoretical analysis of the generalization error for the target risk. Some works claim to derive a rigorous target error bound, while the minimization of joint error and the open-set risk are not theoretically guaranteed [7, 22]. Moreover, most existing algorithms can be described as closed-set UDA after separation of the unknown [3, 22, 34, 35]. In that case, the open-set risk cannot be controlled since the generalization error is also affected by the gap between empirical known target distribution and inferred known target samples, which becomes hard to analyze." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Fig. 1a illustrates the potential problems of existing methods. Generally, it is impossible to perfectly distinguish the unknown such that the following marginal distribution alignment for UDA part is conducted under a large label shit. In that case, the joint error would be crucial as there exists a trade-off between marginal discrepancy and joint error [38]. Even if somehow we manage to obtain a perfect separation, some target data class will be dragged towards the outside of the corresponding decision boundary. In addition, the unknown class cannot be grouped as a single cluster. Fig. 1b intuitively explains the core concept of our algorithm PUJE Sec. 3, where we introduce two additional task classifiers and minimize the predictive discrepancy between the classifiers with same color, such that the unknown class will be pushed away from source data and grouped as a single cluster. Furthermore, those target data outside the corresponding decision boundary will be aligned back to the correct places." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "OUDA problems can be considered as multi-class positive-unlabeled learning problems (MPU) [36] with domain shift, where all the shared classes appearing in both source and target domains are positive and the unknown class is negative. Therefore, to minimize the aforementioned joint error and open-set risk to achieve a tighter error bound for open-set adaptation, instead of unknown separation and closed-set UDA, we combine PU learning theory and joint error-based target error bound into an end-to-end learning framework. The following methods also introduced PU learning to tackle OUDA problems. [18] required the unsupervised source-like reconstruction for target data via encoder-decoder models whose performance cannot be guaranteed especially for large domain shift and complex dataset; [10] only dealt with the open-set label shift and assumed an identical" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Open-set DA via Joint Error based PU Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.744, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.251, + 0.788, + 0.308 + ], + "angle": 0, + "content": "Fig. 1: Intuitive explanation of the difference between our algorithm PUJE and existing methods. (a) existing methods do not explicitly minimize joint error and cannot group unknown class as a single cluster; (b) our proposal is an upper bound of joint error which can address large domain shift and group unknown class into a single cluster." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.784, + 0.41 + ], + "angle": 0, + "content": "class conditional distribution of each instance across domains; [35] tackled a slightly different task where the target data is changing over time and it did not cover joint error and the gap between the empirical and inferred known target distribution. Compared with previous works, the main contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.42, + 0.786, + 0.449 + ], + "angle": 0, + "content": "- We propose an end-to-end algorithm with theoretical analysis for open-set problem setting by PU learning-induced joint error based target error bound;" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.449, + 0.786, + 0.493 + ], + "angle": 0, + "content": "- We define a novel discrepancy measurement, namely open-set margin discrepancy (OMD), which allows us to smoothly apply the strategy in PU learning to the target error bound in domain adaptation;" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.493, + 0.787, + 0.509 + ], + "angle": 0, + "content": "- Extensive experiments confirm the effectiveness and efficiency of the proposal." + }, + { + "type": "list", + "bbox": [ + 0.231, + 0.42, + 0.787, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.382, + 0.546 + ], + "angle": 0, + "content": "2 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.607 + ], + "angle": 0, + "content": "In this section, we introduce notations, problem settings, and theoretical definitions for the tasks of closed-set unsupervised domain adaptation, open-set unsupervised domain adaptation, and multi-class positive and unlabeled learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.708 + ], + "angle": 0, + "content": "Definition 1 (Closed-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of \\( n \\) labeled points \\( \\hat{S} = \\{(x_s^i,y_s^i)\\in (\\mathcal{X}\\subseteq \\mathbb{R}^D\\times \\mathcal{Y} = \\{1,\\dots,K\\})\\}_{i = 1}^n \\) from the source domain \\( S \\), and a set of \\( m \\) unlabeled points \\( \\hat{T} = \\{(x_t^i)\\in \\mathcal{X}\\}_{i = 1}^m \\) from a different target domain \\( T \\). With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier \\( f:\\mathcal{X}\\to \\mathcal{Y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Definition 2 (Multi-class Positive and Unlabeled Learning). For a classification task, the learning algorithm has access to a set of \\( n \\) labeled points \\( \\hat{P} = \\{(x_p^i, y_p^i) \\in (\\mathcal{X} \\times \\mathcal{Y}' = \\{1, ..., K - 1\\})\\}_{i=1}^n \\) from the positive domain \\( P \\), and a set of \\( m \\) unlabeled points \\( \\hat{U} = \\{(x_u^i) \\in \\mathcal{X}\\}_{i=1}^m \\) from the unlabeled domain \\( U \\), where \\( y_u^i \\in \\mathcal{Y} \\). With training samples drawn i.i.d from positive and unlabeled domains, the goal is to learn an optimal classifier \\( f: \\mathcal{X} \\to \\mathcal{Y} \\) for the unlabeled domain. Here, \\( U \\) includes samples from unknown class \\( K \\), but the class conditional distributions for known classes remain invariant between \\( P, U \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.24 + ], + "angle": 0, + "content": "Definition 3 (Open-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of \\( n \\) labeled points \\( \\hat{S}' = \\{(x_{s'}^i, y_{s'}^i) \\in (\\mathcal{X} \\times \\mathcal{Y}')\\}_{i=1}^n \\) from the incomplete source domain \\( S' \\), and a set of \\( m \\) unlabeled points \\( \\hat{T} = \\{(x_t^i) \\in \\mathcal{X}\\}_{i=1}^m \\) from a different target domain \\( T \\), where \\( y_t^i \\in \\mathcal{Y} \\). With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier \\( f: \\mathcal{X} \\to \\mathcal{Y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.787, + 0.295 + ], + "angle": 0, + "content": "We show the relation between open-set domain adaptation and multi-class positive and unlabeled learning as follows, which includes the hits on how to apply the domain adaptation algorithm in the open-set scenario." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.303, + 0.784, + 0.36 + ], + "angle": 0, + "content": "- In Definition 1, there is a domain shift between the distributions of the samples from the source and target domains such that a classifier trained on \\( S \\) cannot generalize on \\( T \\). A typical solution is to align the feature distributions of both domains by a feature extractor [9,32]." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.363, + 0.784, + 0.434 + ], + "angle": 0, + "content": "- In Definition 2, the conditional distributions of the samples from known class are identical for positive and unlabeled domains. As proved in [36], the expected error on the unlabeled domain can be approximated by the expected error on the positive domain and the probability that the unlabeled sample has not been classified as unknown." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.438, + 0.784, + 0.51 + ], + "angle": 0, + "content": "- In Definition 3, the incomplete source domain \\( S' \\) can be regarded as the positive domain \\( P \\) in MPU. The target domain \\( T \\) can be treated as \\( U \\) with a domain shift. Assume that we can bridge the feature distributions with UDA techniques, it becomes possible to address open-set problems with well-established positive and unlabeled (PU) learning theories." + }, + { + "type": "list", + "bbox": [ + 0.231, + 0.303, + 0.784, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.535, + 0.787, + 0.554 + ], + "angle": 0, + "content": "3 PU Learning induced Joint Error based OUDA (PUJE)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.787, + 0.672 + ], + "angle": 0, + "content": "In this section, inspired by [37], we first rigorously reformulate the joint error-based target upper bound to facilitate the further derivation of a practical objective function for OUDA. To estimate the expectation over the complete source domain \\( S \\), we deploy the theory from multi-class positive and unlabeled learning to address the open-set problems where the source domain \\( S' \\) is incomplete. We lastly show the entire loss of PUJE by combining the aforementioned joint error and PU learning." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.681, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Theorem 1 (Approximated Joint Error based Target Upper Bound\\(^3\\)). Given the output space \\(\\mathcal{K} = \\{k|k\\in \\mathbb{R}^K:\\sum_{y\\in \\mathcal{Y}}k[y] = 1,k[y]\\in [0,1]\\}\\), let \\(f_{S},f_{T}:\\mathcal{X}\\to \\mathcal{K}\\) be the true labeling functions for the source and target domains respectively, whose outputs are one-hot vectors denoting the corresponding labels of inputs. Let \\(\\epsilon :\\mathcal{K}\\times \\mathcal{K}\\rightarrow \\mathbb{R}\\) denote a distance metric and \\(\\epsilon_D(f,f')\\coloneqq \\mathbb{E}_{x\\sim D}\\epsilon (f(x),f'(x))\\) measure the expected disagreement between the outputs of \\(f,f':\\mathcal{X}\\to \\mathcal{K}\\) over a distribution \\(D\\) on \\(\\mathcal{X}\\). For \\(\\forall f_S^*\\in \\mathcal{H}_S\\subseteq \\mathcal{H},\\forall f_T^*\\in \\mathcal{H}_T\\subseteq \\mathcal{H},\\forall h\\in \\mathcal{H}:\\mathcal{X}\\to \\mathcal{K}\\) where \\(h(x)[y]\\) indicates the probability of \\(x\\in \\mathcal{X}\\) labeled as \\(y\\in \\mathcal{V}\\), the expected target error is bounded by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.498, + 0.84 + ], + "angle": 0, + "content": "3 see proofs in the supplemental material" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Open-set DA via Joint Error based PU Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.172, + 0.788, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {T} (h) \\leq \\epsilon_ {S} (h) + \\epsilon_ {T} \\left(f _ {S} ^ {*}, f _ {T} ^ {*}\\right) + \\epsilon_ {T} \\left(h, f _ {S} ^ {*}\\right) - \\left| \\epsilon_ {S} \\left(f _ {S} ^ {*}, f _ {T} ^ {*}\\right) - \\epsilon_ {S} \\left(h, f _ {T} ^ {*}\\right) \\right| + \\theta \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.191, + 0.787, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\theta = 2 \\epsilon_ {T} \\left(f _ {S}, f _ {S} ^ {*}\\right) + \\epsilon_ {S} \\left(f _ {S}, f _ {S} ^ {*}\\right) + 2 \\epsilon_ {S} \\left(f _ {T} ^ {*}, f _ {T}\\right) + \\epsilon_ {T} \\left(f _ {T} ^ {*}, f _ {T}\\right) = \\theta_ {f _ {S}} + \\theta_ {f _ {T}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.216, + 0.789, + 0.323 + ], + "angle": 0, + "content": "According to Theorem 1, we show that the expected target error \\(\\epsilon_T(h) \\coloneqq \\epsilon_T(h, f_T)\\) is bounded by the expected source error \\(\\epsilon_S(h) \\coloneqq \\epsilon_S(h, f_S)\\), the discrepancy between domains \\([\\epsilon_T(f_S^*, f_T^*) + \\epsilon_T(h, f_S^*) - |\\epsilon_S(f_S^*, f_T^*) - \\epsilon_S(h, f_T^*)|]\\), and the deviation from true labeling functions \\(\\theta\\). The following Assumption 1 allows us to derive the generalization error of Theorem 1 such that we can ignore the residual term \\(\\theta\\) and relate target error bound with the empirical estimation of source error and the discrepancy between domains." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.33, + 0.789, + 0.379 + ], + "angle": 0, + "content": "Assumption 1. Let \\(\\hat{S},\\hat{T}\\) denote empirical samples with finite size from source and target domains. Assume that there exists approximated labeling functions \\(f_{S}^{*},f_{T}^{*}\\) such that the empirical deviation \\(\\hat{\\theta}_{f_S},\\hat{\\theta}_{f_T}\\) are close enough to zero." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.385, + 0.79, + 0.432 + ], + "angle": 0, + "content": "Definition 4 (Empirical Rademacher Complexity). Let \\(\\mathcal{F}\\) be a class of real-valued functions: \\(\\mathcal{X} \\to \\mathbb{R}\\) and \\(\\hat{D} = \\{x_1, \\ldots, x_m\\}\\) a finite sample drawn i.i.d. from a distribution \\(D\\), the empirical Rademacher Complexity of \\(\\mathcal{F}\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.442, + 0.788, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathfrak {R}} _ {\\hat {D}} (\\mathcal {F}) = \\frac {1}{m} \\mathbb {E} _ {\\sigma} [ \\sup _ {f \\in \\mathcal {F}} \\sum_ {i = 1} ^ {m} \\sigma_ {i} f (x _ {i}) ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.485, + 0.788, + 0.503 + ], + "angle": 0, + "content": "where \\(\\sigma_{i}\\) is an independent uniform random variable taking values in \\(\\{-1, + 1\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.51, + 0.79, + 0.556 + ], + "angle": 0, + "content": "Theorem 2 (Generalization Error\\(^3\\)). Given Theorem 1 and Definition 4, Assumption 1 and function space \\(\\mathcal{F}_{f,f'}^{\\epsilon}:\\mathcal{X}\\to \\epsilon (f(X),f'(X))\\) bounded by \\(M > 0\\), for any \\(\\delta >0\\), with probability at least \\(1 - 2\\delta\\), for \\(\\forall h\\in \\mathcal{H}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.578, + 0.788, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {T} (h) \\leq \\epsilon_ {\\hat {S}} (h) + \\epsilon_ {\\hat {T}} (f _ {S} ^ {*}, f _ {T} ^ {*}) + \\epsilon_ {\\hat {T}} (h, f _ {S} ^ {*}) - | \\epsilon_ {\\hat {S}} (f _ {S} ^ {*}, f _ {T} ^ {*}) - \\epsilon_ {\\hat {S}} (h, f _ {T} ^ {*}) | + 2 [ \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {h, f _ {S}} ^ {\\epsilon}) ] \\\\ + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {h, f _ {S} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {h, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\\epsilon}) (4) \\\\ + 2 \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\\epsilon}) + 2 \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\\epsilon}) ] + 3 M (5 \\sqrt {\\frac {\\log \\frac {9}{\\delta}}{2 m}} + 6 \\sqrt {\\frac {\\log \\frac {9}{\\delta}}{2 n}}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.681, + 0.788, + 0.726 + ], + "angle": 0, + "content": "In the following part, we introduce the techniques from PU learning to estimate the expectation over source domain \\( S \\) by the incomplete source domain \\( S' \\) and target domain \\( T \\) for the open-set scenario." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Definition 5 (Unknown Predictive Discrepancy). Let \\( v: \\mathcal{K} \\times \\mathcal{K} \\to \\mathbb{R} \\) denote the Unknown Predictive Discrepancy as a distance metric and \\( v_{D}(f, f') := \\mathbb{E}_{x \\sim D} v(f(x), f'(x)) \\) measure the expected disagreement between the \\( K \\)-th outputs of \\( f, f': \\mathcal{X} \\to \\mathcal{K} \\) over a distribution \\( D \\) on \\( \\mathcal{X} \\). Let \\( e^{K}: \\mathcal{X} \\to [0, \\dots, 1] \\in \\mathcal{K} \\) denote a function that can predict any input as unknown. The deviation from \\( e^{K} \\) for a hypothesis \\( h \\in \\mathcal{H} \\) is referred to as the shorthand \\( v_{D}(h) := v_{D}(h, e^{K}) \\) that measures the probability that samples from \\( D \\) have not been classified as unknown." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Assumption 2. Let \\( S^i = P_S(x|y = i), T^i = P_T(x|y = i) \\) denote class conditional distributions, \\( S' = P_S(x|y \\neq K), T' = P_T(x|y \\neq K) \\) indicate incomplete domains that do not contain unknown class \\( S^K, T^K \\). Given a feature extractor \\( g: \\mathcal{X} \\subseteq \\mathbb{R}^D \\to \\mathcal{Z} \\subseteq \\mathbb{R}^F \\), assume that the feature space can be aligned by UDA techniques such that \\( Z^K = P_{S^K}(z) = P_{T^K}(z), Z' = P_{S'}(z) = P_{T'}(z) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Proposition 1. Let \\( h \\in \\mathcal{H}_{F_g} : \\mathcal{Z} \\to \\mathcal{K} \\) denote the decomposed hypothesis where \\( h \\circ g \\in \\mathcal{H} \\). Given Definition 5, Assumption 2, \\( v_{S^K}(h \\circ g) = v_{T^K}(h \\circ g) \\) that represents probability that samples from \\( S^K \\) have not been classified as unknown class, is equivalent to \\( \\epsilon_{S^K}(h \\circ g) \\) that measures the classification error on unknown class according to the definition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Lemma 1 (Estimated Source Error\\(^3\\)). Let \\(\\sum_{i=1}^{K} \\pi_S^i = 1\\), \\(\\sum_{i=1}^{K} \\pi_T^i = 1\\) denote the label distribution of \\(S\\) and \\(T\\) respectively. Given Proposition 1, the expected error on \\(S\\) can be estimated by the error on \\(S'\\) and Unknown Predictive Discrepancy (Definition 5) on \\(T\\) with a mild condition that \\(\\pi_S^K = \\pi_T^K = 1 - \\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.415, + 0.788, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {S} (h \\circ g) = \\alpha [ \\epsilon_ {S ^ {\\prime}} (h \\circ g) - v _ {S ^ {\\prime}} (h \\circ g) ] + v _ {T} (h \\circ g) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Remark 1. According to Definition 5, minimizing \\( v_{T}(h \\circ g) \\) means classifying target samples as the unknown class. In practice, a multiplier \\( \\beta < 1 \\) is applied on \\( v_{T}(h \\circ g) \\) to prevent all target samples from being recognized as unknown." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.582 + ], + "angle": 0, + "content": "To reformulate the intractable discrepancy \\(\\epsilon_S(f_S^*,f_T^*)\\), \\(\\epsilon_S(h,f_T^*)\\) in Theorem 1, we decompose the approximated labeling functions given feature extractor \\(g\\) such that \\(f_{S}^{\\star}\\circ g = f_{S}^{\\star},f_{T}^{\\star}\\circ g = f_{T}^{\\star}\\) where \\(f_{S}^{\\star}\\in \\mathcal{H}_{S_{g}}\\subseteq \\mathcal{H}_{F_{g}},f_{T}^{\\star}\\in \\mathcal{H}_{T_{g}}\\subseteq \\mathcal{H}_{F_{g}}\\). We further define \\(\\epsilon\\) in a way such that, under some circumstances, it is equivalent to the Unknown Predictive Discrepancy for \\(\\forall f\\in \\mathcal{H}_{F_g}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.597, + 0.788, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {T ^ {K}} \\left(f \\circ g, f _ {T} ^ {\\star} \\circ g\\right) = v _ {T ^ {k}} \\left(f \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Corollary 1. Given Assumption 2, the feature extractor \\( g \\) can align the feature distributions of source and target domain such that the expectation over \\( S^K \\) is equal to that over \\( T^K \\), which leads Eq. (6) to: \\( \\epsilon_{S^K}(f \\circ g, f_T^\\star \\circ g) = \\epsilon_{T^K}(f \\circ g, f_T^\\star \\circ g) = v_{T^K}(f \\circ g, f_T^\\star \\circ g) = v_{S^K}(f \\circ g, f_T^\\star \\circ g) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Lemma 2 (Estimated Discrepancy\\(^3\\)). Given Corollary 1, analogous to Lemma 1, the discrepancy measured on source domain \\(\\epsilon_S(f_S^*, f_T^*)\\), \\(\\epsilon_S(h, f_T^*)\\) can be estimated by the discrepancy on \\(S'\\) and Unknown Predictive Discrepancy (Definition 5) with a mild condition that \\(\\pi_S^K = \\pi_T^K = 1 - \\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.787, + 0.816, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\epsilon_ {S} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) & = \\alpha \\left[ \\epsilon_ {S ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {S ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] + v _ {T} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\\\ \\epsilon_ {S} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) & = \\alpha \\left[ \\epsilon_ {S ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {S ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] + v _ {T} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\end{array} \\right. \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Open-set DA via Joint Error based PU Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.79, + 0.239 + ], + "angle": 0, + "content": "Definition 6 (Overall Loss of PUJE). Let \\(\\hat{S}',\\hat{T}\\) denote empirical data from \\(S',T\\). Given a feature extractor \\(g:\\mathcal{X}\\subseteq R^{D}\\to \\mathcal{Z}\\subseteq \\mathbb{R}^{F}\\), decomposed hypothesis and approximated labeling functions \\(h,f_{S}^{\\star},f_{T}^{\\star}\\in \\mathcal{H}_{F_{g}}:\\mathcal{Z}\\to \\mathcal{K}\\), replacing the intractable terms in Theorem 2 with Eqs. (5) and (7) and ignoring the Rademacher Complexity terms, the overall loss of PUJE can be regrouped as the sum of classification loss \\(L_{cls}\\) and the discrepancy between domains \\(L_{dis}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.261, + 0.822, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {p u j e} = \\alpha \\left[ \\epsilon_ {\\hat {S} ^ {\\prime}} (h \\circ g) - v _ {\\hat {S} ^ {\\prime}} (h \\circ g) \\right] + \\beta v _ {\\hat {T}} (h \\circ g) + \\epsilon_ {\\hat {T}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) + \\epsilon_ {\\hat {T}} (h \\circ g, f _ {S} ^ {\\star} \\circ g) \\\\ - \\left| v _ {\\hat {T}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {\\hat {T}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) + \\alpha \\left[ \\epsilon_ {\\hat {S} ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right. \\right. \\\\ \\left. + v _ {\\hat {S} ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - \\epsilon_ {\\hat {S} ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {\\hat {S} ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] | \\\\ = L _ {c l s} (h; g) + L _ {d i s} \\left(f _ {S} ^ {\\star}, f _ {T} ^ {\\star}, h; g\\right) \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.355, + 0.381, + 0.372 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.383, + 0.789, + 0.462 + ], + "angle": 0, + "content": "In this section, we first define Approximated Labeling Function Space \\(\\mathcal{H}_{S_g},\\mathcal{H}_{T_g}\\) such that the intractable \\(L_{dis}(f_S^\\star ,f_T^\\star ,h;g)\\) due to unknown \\(f_S^\\star ,f_T^\\star\\) can be upper bounded by \\(\\sup_{f_S'\\in \\mathcal{H}_{S_g},f_T'\\in \\mathcal{H}_{T_g}}L_{dis}(f_S',f_T',h;g)\\). To fulfill Eq. (6), we propose Open-set Margin Discrepancy to quantify \\(\\epsilon\\) that measures the disagreement between classifiers. We lastly show the entire training algorithm." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.48, + 0.589, + 0.496 + ], + "angle": 0, + "content": "4.1 Approximated Labeling Function Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.599 + ], + "angle": 0, + "content": "Proposition 2. Let \\(\\mathcal{H}_S, \\mathcal{H}_T : \\mathcal{X} \\to \\mathcal{K}\\) be two sets of functions that can minimize a part of the empirical residual \\(\\hat{\\theta}_S, \\hat{\\theta}_T\\) respectively, \\(f_S^* \\in \\mathcal{H}_S\\), \\(f_T^* \\in \\mathcal{H}_T\\) must hold as \\(f_S^*, f_T^*\\) can minimize the entire \\(\\hat{\\theta}\\) given Assumption 1. Accordingly, decomposed functions \\(f_S^*, f_T^*\\) must lie in \\(\\mathcal{H}_{S_g} = \\{f|\\forall f \\circ g \\in \\mathcal{H}_S\\}\\), \\(\\mathcal{H}_{T_g} = \\{f|\\forall f \\circ g \\in \\mathcal{H}_T\\}\\) such that a sufficient condition of the following inequality: \\(L_{dis}(f_S^*, f_T^*, h; g) \\leq \\sup_{f_S' \\in \\mathcal{H}_{S_g}, f_T' \\in \\mathcal{H}_{T_g}} L_{dis}(f_S', f_T', h; g)\\) is fulfilled." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.605, + 0.789, + 0.7 + ], + "angle": 0, + "content": "Definition 7 (Approximated Labeling Function Space). Let \\( L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T} \\) denote a part of the empirical residual \\( \\hat{\\theta}_{fs}, \\hat{\\theta}_{f_T} \\) respectively. Let \\( f_S^\\star \\circ g = f_S^* \\in \\mathcal{H}_S, f_T^\\star \\circ g = f_T^* \\in \\mathcal{H}_T \\) denote the decomposed approximated labeling functions where \\( f_S^\\star \\in \\mathcal{H}_{S_g} \\subseteq \\mathcal{H}_{F_g}, f_T^\\star \\in \\mathcal{H}_{T_g} \\subseteq \\mathcal{H}_{F_g} \\). Given Proposition 2, Approximated Labeling Function Space \\( \\mathcal{H}_{S_g}, \\mathcal{H}_{T_g} \\) can be defined as the sets whose members \\( f_S', f_T' \\in \\mathcal{H}_{F_g} \\) can minimize \\( L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.717, + 0.788, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathcal {H} _ {S _ {g}} = \\left\\{f _ {S} ^ {\\prime} \\right| \\arg \\min _ {g, f _ {S} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) := L _ {c l s} \\left(f _ {S} ^ {\\prime}; g\\right) \\} \\\\ \\mathcal {H} _ {T _ {g}} = \\left\\{f _ {T} ^ {\\prime} \\right| \\arg \\min _ {g, f _ {T} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) := (1 - \\gamma) L _ {c l s} \\left(f _ {T} ^ {\\prime}; g\\right) + \\gamma L _ {r e g} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "\\(\\mathcal{H}_{S_g}\\) consists of functions minimizing the expected error on \\(S\\), which can be estimated by Lemma 1. To build a reliable function space \\(\\mathcal{H}_{T_g}\\) without target labels, we approximate the target error with the weighted average \\((\\gamma \\in [0,1])\\) of error rate on labeled samples and a semi-supervised regularization term \\(L_{reg} = L_{ent} + L_{pse} + \\omega L_{con}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Regularized Entropy Minimization As introduced in [11,27,31], we impose a class balance prior that can penalize classifiers with complex decision boundaries on entropy minimization [12] to yield a more sensible solution:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.219, + 0.788, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {e n t} = - \\mathbb {E} _ {x \\in \\hat {T}} \\sum_ {y \\in \\mathcal {Y}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\log f _ {T} ^ {\\prime} (g (x)) [ y ] \\\\ + \\sum_ {y \\in \\mathcal {Y}} \\mathbb {E} _ {x \\in \\hat {T}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\log \\mathbb {E} _ {x \\in \\hat {T}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.318 + ], + "angle": 0, + "content": "Pseudo Labeling As introduced in [30, 31], for input \\( x \\in \\hat{T} \\) and its random augmentation \\( x' \\) [5], we minimize cross entropy for \\( x \\) with pseudo labels of \\( x' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.334, + 0.788, + 0.353 + ], + "angle": 0, + "content": "\\[\nL _ {p s e} = - \\mathbb {E} _ {x \\in \\hat {T}} \\log f _ {T} ^ {\\prime} (g (x)) [ \\arg \\max _ {y \\in \\mathcal {Y}} h (g (x ^ {\\prime})) [ y ] ] \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.366, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Consistency Regularization As introduced in [14, 29], we penalize the difference of the outputs for input \\( x \\in \\hat{T} \\) and its random augmentation \\( x' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.413, + 0.788, + 0.431 + ], + "angle": 0, + "content": "\\[\nL _ {c o n} = \\mathbb {E} _ {x \\in \\hat {T}} | f _ {T} ^ {\\prime} (g (x)) - f _ {T} ^ {\\prime} (g (x ^ {\\prime})) | \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.452, + 0.51, + 0.468 + ], + "angle": 0, + "content": "4.2 Open-set Margin Discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.479, + 0.79, + 0.543 + ], + "angle": 0, + "content": "Definition 8 (Induced Labeling Function). Let \\( f: \\mathcal{X} \\to \\mathcal{K} \\) denote a multiclass labeling function, where \\( f(x)[y] \\) indicating \\( y \\)-th element of the output \\( f(x) \\) for the probability of \\( x \\) classified as \\( y \\). Thus an induced labeling function \\( l \\circ f: \\mathcal{X} \\to \\mathcal{Y} \\) is given by: \\( l(f(x)) = \\arg \\max_{y \\in \\mathcal{Y}} f(x)[y] \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.552, + 0.794, + 0.599 + ], + "angle": 0, + "content": "Definition 9 (Open-set Margin Discrepancy (OMD)). Let \\( y = l(f(x)), y' = l(f'(x)) \\) denote the predictions on input \\( x \\) from \\( f, f': \\mathcal{X} \\to \\mathcal{K} \\) given Definition 8. Open-set Margin Discrepancy between \\( f, f' \\) over a distribution \\( D \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.626, + 0.788, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {D} (f, f ^ {\\prime}) = \\mathbb {E} _ {x \\sim D} [ \\operatorname {o m d} (f (x), f ^ {\\prime} (x)) ] (13) \\\\ \\operatorname {o m d} (f (x), f ^ {\\prime} (x)) = \\max \\left(\\left| \\log (1 - f (x) [ y ]) - \\log (1 - f ^ {\\prime} (x) [ y ]) \\right|\\right), \\\\ \\left. \\left| \\log \\left(1 - f (x) \\left[ y ^ {\\prime} \\right]\\right) - \\log \\left(1 - f ^ {\\prime} (x) \\left[ y ^ {\\prime} \\right]\\right) \\right|\\right) (14) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.693, + 0.787, + 0.74 + ], + "angle": 0, + "content": "For OUDA, generally the classifier is not discriminative for unknown class \\( K \\) due to the lack of labeled samples to reduce the misclassification rate. To address this, we model \\( \\log (1 - f(x)[K]) \\) that has a more stable gradient." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.748, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Remark 2. According to Definition 5, \\( v_{D}(f,f^{\\prime}) \\) measures the expectation of disagreement on the K-th output between two functions \\( f,f^{\\prime}:\\mathcal{X}\\to \\mathcal{K} \\) over a distribution \\( D \\), which can also indicate the difference between probabilities of categorizing inputs as known classes. To fulfill Eq. (6), we quantify \\( v_{D}(f,f^{\\prime}) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.825, + 0.788, + 0.843 + ], + "angle": 0, + "content": "\\[\nv _ {D} (f, f ^ {\\prime}) = \\mathbb {E} _ {x \\sim D} | \\log (1 - f (x) [ K ]) - \\log (1 - f ^ {\\prime} (x) [ K ]) | \\tag {15}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Open-set DA via Joint Error based PU Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.147, + 0.369, + 0.161 + ], + "angle": 0, + "content": "Algorithm 1 PUJE" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.168, + 0.626, + 0.182 + ], + "angle": 0, + "content": "Input: incomplete source data \\(\\hat{S}^{\\prime}\\); unlabeled target data \\(\\hat{T}\\)" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.183, + 0.687, + 0.197 + ], + "angle": 0, + "content": "Output: labeling functions \\( f_S', f_T' \\); feature extractor \\( g \\); hypothesis \\( h \\)" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.197, + 0.777, + 0.223 + ], + "angle": 0, + "content": "Parameter: trade-off parameter \\(\\lambda\\); learning rate \\(\\eta\\); known class ratio estimator \\(\\alpha\\) for epoch = 1, 2, ... do" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.224, + 0.693, + 0.238 + ], + "angle": 0, + "content": "Step 1: estimate known class ratio \\(\\alpha\\) on \\(T\\) with prediction of \\(g, h\\)" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.238, + 0.773, + 0.252 + ], + "angle": 0, + "content": "Step 2: optimize \\( g, f_S', f_T' \\) to satisfy the approximated labeling function space" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.252, + 0.532, + 0.266 + ], + "angle": 0, + "content": "\\[\n(g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) \\leftarrow (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) + \\eta \\Delta (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.266, + 0.534, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\varDelta (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) = - \\frac {\\partial (L _ {\\mathcal {H} _ {S}} (f _ {S} ^ {\\prime} ; g) + L _ {\\mathcal {H} _ {T}} (f _ {T} ^ {\\prime} ; g))}{\\partial (g , f _ {S} ^ {\\prime} , f _ {T} ^ {\\prime})}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.288, + 0.742, + 0.301 + ], + "angle": 0, + "content": "Step 3: maximize the discrepancy w.r.t. \\( f_{S}^{\\prime}, f_{T}^{\\prime} \\) within the function space" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.301, + 0.488, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right) \\leftarrow \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right) + \\eta \\Delta \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.315, + 0.623, + 0.338 + ], + "angle": 0, + "content": "\\[\n\\varDelta(f_{S}^{\\prime},f_{T}^{\\prime})=-\\frac{\\partial(L_{\\mathcal{H}_{S}}(f_{S}^{\\prime};g)+L_{\\mathcal{H}_{T}}(f_{T}^{\\prime};g)-\\lambda L_{dis}(f_{S}^{\\prime},f_{T}^{\\prime},h;g))}{\\partial(f_{S}^{\\prime},f_{T}^{\\prime})}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.337, + 0.742, + 0.351 + ], + "angle": 0, + "content": "Step 4: minimize the entire target error bound w.r.t. \\(g, h\\) for fixed \\(f_S', f_T'\\)." + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.351, + 0.439, + 0.365 + ], + "angle": 0, + "content": "\\[\n(g, h) \\leftarrow (g, h) + \\eta \\Delta (g, h)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.365, + 0.526, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\Delta (g, h) = - \\frac {\\partial \\left(L _ {c l s} (h ; g) + \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime} , f _ {T} ^ {\\prime} , h ; g\\right)\\right)}{\\partial (g , h)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.383, + 0.289, + 0.393 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.426, + 0.423, + 0.441 + ], + "angle": 0, + "content": "4.3 Training Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Given the feature extractor \\( g: \\mathcal{X} \\subseteq \\mathbb{R}^D \\to \\mathcal{Z} \\subseteq \\mathbb{R}^F \\) and hypotheses \\( h, f_S', f_T' \\in \\mathcal{H}_{F_g}: \\mathcal{Z} \\to \\mathcal{K} \\), we introduce a trade-off parameter \\( \\lambda \\) to balance the classification loss and discrepancy as Eq. (16). During each epoch, we first estimate the known class ratio \\( \\alpha \\) (initially set to 1) in target data by the predictions of \\( h \\). Then we train the hypothesis \\( h \\), approximated labeling functions \\( f_S', f_T' \\), feature extractor \\( g \\) to minimize the error on labeled data \\( L_{cls} \\) with hypothesis space constraints \\( L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T} \\), while optimizing the discrepancy \\( L_{dis} \\) adversarially by the min-max game over \\( g \\) and \\( f_S', f_T' \\) (Algorithm 1):" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.594, + 0.787, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\min _ {f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) + L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) - \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}, h; g\\right) \\right. \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.619, + 0.733, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\left\\lfloor \\min _ {h \\in \\mathcal {H} _ {F _ {g}}, g} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) + L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) + L _ {c l s} (h; g) + \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}, h; g\\right) \\right.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.335, + 0.679 + ], + "angle": 0, + "content": "4.4 Intuition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this section, we intuitively explain how our method can align the source and target domains while separating the unknown in Fig. 2. According to [24], unlike the entropy minimization that tends to cut through unlabeled data, consistency regularization helps to draw a manifold-aware decision boundary. Based on this conclusion, the decision boundary of \\( f_{T}^{\\prime} \\) induced by consistency regularization \\( L_{con} \\) can preserve the cluster structure (Fig. 2a). According to Algorithm 1, \\( f_{S}^{\\prime} \\) tries to maximize \\( L_{dis} \\) while classify target samples as unknown class to minimize \\( L_{\\mathcal{H}_S}(f_S';g) \\) (Definition 7 and Lemma 1), which can lead to a new decision boundary shown in Fig. 2b. Meanwhile, feature extractor \\( g \\) tries to minimize \\( L_{dis} \\) by pushing target samples close to \\( f_{S}^{\\prime} \\) towards source clusters," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.379, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.261, + 0.345, + 0.277 + ], + "angle": 0, + "content": "(a) \\(\\min_{f_T'} L_{\\mathcal{H}_T}\\)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.143, + 0.538, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.261, + 0.528, + 0.277 + ], + "angle": 0, + "content": "(b) \\(\\min_{f_S'}(L_{\\mathcal{H}_S} - L_{dis})\\)" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.146, + 0.665, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.675, + 0.147, + 0.714, + 0.179 + ], + "angle": 0, + "content": "\\(L_{dis}\\) \\(f_T^{\\prime}\\) \\(f_S^{\\prime}\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.685, + 0.188, + 0.782, + 0.236 + ], + "angle": 0, + "content": "- Positive Source\n- Negative Source\n- Unknown Target\n- Pseudo Positive\n- Pseudo Negative" + }, + { + "type": "list", + "bbox": [ + 0.685, + 0.188, + 0.782, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.261, + 0.701, + 0.274 + ], + "angle": 0, + "content": "(c) \\(\\min_g L_{dis}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Fig. 2: Intuitive explanation of the mechanism of the proposed PUJE in Algorithm 1. (a) \\( L_{\\mathcal{H}_T} \\) includes consistency regularization \\( L_{con} \\) that helps to draw a manifold-aware decision boundary for \\( f_T' \\) (Step 2); (b) \\( f_S' \\) can increase \\( L_{dis} \\) by classifying most of the unlabeled target samples as unknown, which also decreases \\( L_{\\mathcal{H}_S} \\) (Step 3); (c) feature extractor \\( g \\) pushes unlabeled target samples close to the decision boundary of \\( f_S' \\) towards source cluster, while those close to \\( f_T' \\) towards outside to reduce \\( L_{dis} \\), which can lead to a separated unknown cluster (Step 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.462 + ], + "angle": 0, + "content": "while those close to \\( f_{T}^{\\prime} \\) towards outside such that unlabeled target samples far from source clusters can be eventually separated and grouped into a new cluster for the unknown progressively as illustrated in Fig. 2c." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.357, + 0.501 + ], + "angle": 0, + "content": "5 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.789, + 0.673 + ], + "angle": 0, + "content": "We evaluated our proposal on two benchmarks, Office-Home and Syn2Real-O. In the implementation, the \\( L_{reg} \\) weight \\( \\gamma \\) and the trade-off parameter \\( \\lambda \\) were set to 0.1 and 0.01 according to [37]. In addition, we empirically set PU coefficient \\( \\beta \\) to 0.03 and consistency coefficient \\( \\omega \\) to 20 for Office-Home and 5 for Syn2Real-O. All experiments were conducted with the ImageNet [6] pre-trained ResNet-50 [13] as the feature extractor \\( g \\) and 2-layer linear networks for classifiers \\( f_S', f_T', h \\). We trained the model by Stochastic Gradient Descent optimizer with a 0.001 learning rate annealed according to [9], 24 batch-size, a momentum of 0.9. We quantitatively compare our results against various OUDA baselines, including OSBP [28], STA [17], DAOD [8], PGL [22], ROS [3], OSLPP [34] and ANNA [15]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Evaluation Metrics To evaluate the proposed method and the baselines, we utilize the widely used measures [22, 28], i.e., normalized accuracy for all classes (OS), normalized accuracy for the known classes only \\((\\mathrm{OS}^{\\star})\\) and harmonic mean \\(\\mathrm{HOS} = 2(\\mathrm{OS}^{\\star}\\times \\mathrm{UNK}) / (\\mathrm{OS}^{\\star} + \\mathrm{UNK})\\) [3,15,18,34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Office-Home [33] is a widely-used domain adaptation benchmark, which consists of 15,500 images from 65 categories and four domains: Art (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw). Following the same splits used in previous methods [22], we select the first 25 classes in alphabetical order as the known classes and group the rest as the unknown." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-set DA via Joint Error based PU Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.145, + 0.767, + 0.159 + ], + "angle": 0, + "content": "Table 1: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (OS)" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.181, + 0.787, + 0.328 + ], + "angle": 0, + "content": "
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP47.165.364.638.348.748.327.081.679.537.173.572.129.355.354.337.781.980.2
STA77.048.449.595.440.842.959.177.376.671.268.668.761.045.446.058.974.573.9
DAOD44.367.766.844.760.359.740.885.083.349.873.272.347.460.459.956.882.881.8
PGL34.773.772.238.459.258.427.684.882.66.181.578.625.166.865.038.084.883.0
PUJE38.474.873.442.264.463.637.987.785.823.481.078.839.271.270.039.188.286.3
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP28.657.256.125.877.875.823.085.483.033.077.275.516.771.369.232.165.964.6
STA64.145.946.662.067.267.066.276.676.257.465.264.960.257.657.772.749.350.2
DAOD71.155.556.166.669.269.163.779.378.754.878.277.354.670.269.655.162.962.6
PGL19.163.361.632.178.977.140.987.785.95.385.982.824.573.972.033.870.268.8
PUJE57.166.666.352.182.881.645.288.987.255.981.880.851.481.580.342.571.069.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.358, + 0.785, + 0.385 + ], + "angle": 0, + "content": "Table 2: Accuracy of ResNet-50 model fine-tuned on Syn2Real-O dataset. * indicates our re-implementation with the officially released code." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.409, + 0.787, + 0.492 + ], + "angle": 0, + "content": "
METHODplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckUNKOS*OSHOS
OSBP73.657.958.265.267.429.584.147.067.85.789.10.566.653.854.859.5
STA64.170.353.759.480.820.890.012.563.230.278.22.759.152.252.755.4
PGL81.568.374.260.691.945.492.241.087.967.579.26.449.666.865.556.9
PGL*81.377.566.571.790.635.792.441.282.646.982.34.959.664.564.162.0
ANNA41.254.639.759.451.225.682.654.878.910.576.62.067.448.149.656.1
PUJE90.773.875.172.477.833.891.258.879.546.584.95.169.065.866.067.3
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.536, + 0.787, + 0.643 + ], + "angle": 0, + "content": "Syn2Real-O [26] is a more challenging synthetic-to-real benchmark, which is constructed from the VisDA-17 [25]. The Syn2Real-O dataset significantly increases the ratio of unknown data in the target domain to 0.9 by introducing additional unknown samples. In this dataset, the source domain contains training data from the VisDA-17 as the known set, and the target domain includes the test data from the VisDA-17 (known set) plus 50k images from irrelevant categories of MSCOCO [16] dataset (unknown set)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.645, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As reported in Tabs. 1 and 2, we observe that our method PUJE consistently outperforms the state-of-the-art results, improving OS and HOS by \\(3.0\\%\\) and \\(5.3\\%\\) on the benchmark datasets of Office-Home and Syn2Real-O respectively. Note that our proposed approach provides significant performance gains for the more challenging dataset of Syn2Real-O that requires knowledge transfer across different modalities. In the sub-tasks with a larger domain shift, e.g., \\(\\mathrm{Rw} \\rightarrow \\mathrm{Cl}\\) and \\(\\mathrm{Pr} \\rightarrow \\mathrm{Cl}\\) in Office-Home, we can observe similar results demonstrating the strong adaptation ability of the proposed framework. PGL reported a higher OS score in Syn2Real-O, but we cannot reproduce the results with the officially released code. During the re-implementation based on their early stop strategy, we found the UNK was around \\(30\\%\\), which is far below the reported score. A potential reason to explain this is that maybe they did not remove the unknown category from the source data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.145, + 0.774, + 0.16 + ], + "angle": 0, + "content": "Table 3: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (HOS)" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.181, + 0.787, + 0.309 + ], + "angle": 0, + "content": "
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOS
ROS64.357.360.671.246.556.378.470.874.470.867.068.8
OSLPP76.254.663.667.153.159.371.277.074.075.060.867.2
ANNA70.363.066.574.854.663.178.974.376.677.366.171.3
PUJE77.856.965.776.851.261.575.280.177.567.371.169.1
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*
ROS74.150.660.170.368.469.377.275.876.572.265.368.6
OSLPP67.155.961.073.172.572.869.480.174.373.967.266.9
ANNA78.761.469.079.968.373.779.774.176.880.266.973.0
PUJE83.558.568.878.979.679.376.382.179.180.671.275.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.334, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Table 4: Ablation studies of semi-supervised regularization losses. We report the Accuracy \\((\\%)\\) on Office-Home of \\(\\mathrm{A}\\rightarrow \\mathrm{R}\\) and \\(\\mathrm{C}\\to \\mathrm{P}\\) using a ResNet50 backbone." + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.374, + 0.637, + 0.463 + ], + "angle": 0, + "content": "
Config.Ar→PrCl→Rw
UNK OS* OSUNK OS* OS
default52.182.881.655.981.880.8
w/o Lpse69.678.177.864.877.877.3
w/o Lcon26.381.779.633.380.979.1
w/o Lent50.281.179.953.980.579.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.673 + ], + "angle": 0, + "content": "HOS Metric Instead of OS score, [18] proposed the harmonic mean of \\(\\mathrm{OS}^{\\star}\\) and unk, \\(\\mathrm{HOS} = 2(\\mathrm{OS}^{\\star} \\times \\mathrm{UNK}) / (\\mathrm{OS}^{\\star} + \\mathrm{UNK})\\) to penalize large gaps between \\(\\mathrm{OS}^{\\star}\\) and unk. We agree that the OS score does not reflect UNK when the number of known classes grows. However, the HOS score is still unfair. \\(\\mathrm{OS}^{\\star}\\) and UNK are equally valued in HOS, but UNK is much easier to improve than \\(\\mathrm{OS}^{\\star}\\) which depends on the power of the adaptation algorithm. [15] utilizes a weak adaptation algorithm based on [28] that leads to a relatively loose alignment (Fig. 4) on the feature space of source and target domains such that unlabeled data is more frequently assigned to the unknown class. We believe that if a method can achieve a high \\(\\mathrm{OS}^{\\star}\\) score, it is possible to obtain a decent HOS score by tuning hyper-parameters or additional losses. For our proposal, we can increase \\(\\beta = 0.25\\) to improve UNK at the sacrifice of \\(\\mathrm{OS}^{\\star}\\) (Tab. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.789, + 0.78 + ], + "angle": 0, + "content": "Ablation Study To investigate the impact of the semi-supervised regularization, we compare three variants of the PUJE model on the Office-Home dataset shown in Tab. 4. Without \\( L_{pse} \\), \\( \\mathrm{OS}^{\\star} \\) decreases because it can no longer leverage the pseudo labels that can help to build reliable hypothesis space \\( \\mathcal{H}_{T_g} \\) (Definition 7). Without \\( L_{con} \\), UNK drops as it is not likely to draw a manifold-aware decision boundary that is beneficial to the separation of unknown data (Fig. 2a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Robustness against Varying Openness To verify the robustness of the proposed PUJE, we conducted experiments on the Syn2Real-O with the openness varying in \\(\\{0.25, 0.5, 0.75, 0.9\\}\\). Here, openness is defined by the ratio of unknown" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Open-set DA via Joint Error based PU Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.149, + 0.404, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.26, + 0.367, + 0.272 + ], + "angle": 0, + "content": "(a) PU Coefficient" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.149, + 0.591, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.26, + 0.555, + 0.272 + ], + "angle": 0, + "content": "(b) SSL Coefficient" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.149, + 0.777, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.26, + 0.726, + 0.272 + ], + "angle": 0, + "content": "(c) Openness" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.277, + 0.404, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.387, + 0.384, + 0.399 + ], + "angle": 0, + "content": "(d) Trade-off in \\(\\mathrm{Ar}\\rightarrow \\mathrm{Cl}\\)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.276, + 0.591, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.387, + 0.58, + 0.4 + ], + "angle": 0, + "content": "(e) Convergence in \\(\\mathrm{Ar}\\rightarrow \\mathrm{Cl}\\)" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.277, + 0.777, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.387, + 0.768, + 0.4 + ], + "angle": 0, + "content": "(f) Convergence in \\(\\mathrm{Rw}\\rightarrow \\mathrm{Cl}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.482 + ], + "angle": 0, + "content": "Fig. 3: (a)-(b): sensitivity to varying loss coefficient \\(\\beta, \\omega\\) verified in \\(\\mathrm{Ar} \\rightarrow \\mathrm{Pr}\\) and \\(\\mathrm{Cl} \\rightarrow \\mathrm{Rw}\\) tasks; (c): performance comparisons w.r.t. varying openness of the Syn2Real-O task; (d): training cure of the proposed algorithm that characterizes a trade-off between the classification accuracy on known classes and the unknown class; (e)-(f): convergence analysis of the Office-Home task compared to other baselines with confidence intervals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.789, + 0.603 + ], + "angle": 0, + "content": "samples in the entire target data. We show the results of PGL and PUJE in Fig. 3c. Note that the PGL approach heuristically sets the hyper-parameter according to the true unknown ratio to control the openness, while PUJE automatically estimates the weight \\(\\alpha\\) during the training procedure. We observe that PUJE consistently outperforms the baseline by a large margin, especially on unknown recognition, which confirms its robustness to the change in openness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.79, + 0.797 + ], + "angle": 0, + "content": "Training Curve In Fig. 3d, we illustrate the recognition performance of PUJE over training steps in the \\(\\mathrm{Ar} \\rightarrow \\mathrm{Cl}\\) task. All metrics show a performance gain over the first several steps. Then, the normalized accuracy \\(\\mathrm{OS}^{\\star}\\) experiences a downward while the accuracy for the unknown class keeps improving, which characterizes a trade-off between \\(\\mathrm{OS}^{\\star}\\) and UNK. In addition, the behavior of \\(f_{S}^{\\prime}, f_{T}^{\\prime}\\) meets the expectation of our alignment mechanism where \\(f_{S}^{\\prime}\\) starts to detect unknown data, which is followed by \\(f_{T}^{\\prime}\\) (Fig. 2b). From Figs. 3e and 3f, we observe that some previous works [15,22] do not converge at the reported optimum, and they tackle this problem by ceasing the model updates before the score starts to decline. However, this early stop strategy can fail without enough labeled target data to validate the current performance. In contrast, our method can reach a reliable convergence." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.809, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Sensitivity to PU & SSL Coefficient We show the sensitivity of our approach to various PU and SSL loss coefficient \\(\\beta, \\omega\\). We tested the value of \\(\\beta\\) from [0.01, 0.5]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.363, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.226, + 0.329, + 0.239 + ], + "angle": 0, + "content": "(a) OSBP" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.146, + 0.498, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.226, + 0.461, + 0.239 + ], + "angle": 0, + "content": "(b) PGL" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.634, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.226, + 0.601, + 0.239 + ], + "angle": 0, + "content": "(c) ANNA" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.147, + 0.769, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.226, + 0.736, + 0.239 + ], + "angle": 0, + "content": "(d) PUJE" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.241, + 0.362, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.321, + 0.328, + 0.333 + ], + "angle": 0, + "content": "(e) OSBP" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.241, + 0.497, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.321, + 0.46, + 0.334 + ], + "angle": 0, + "content": "(f) PGL" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.241, + 0.633, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.321, + 0.602, + 0.334 + ], + "angle": 0, + "content": "(g) ANNA" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.241, + 0.768, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.321, + 0.736, + 0.334 + ], + "angle": 0, + "content": "(h) PUJE" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.344, + 0.785, + 0.36 + ], + "angle": 0, + "content": "Fig. 4: Feature space visualized by t-SNE in (a)-(d): Office-Home; (e)-(h): Syn2Real-O." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.489 + ], + "angle": 0, + "content": "in \\(\\mathrm{Ar}\\rightarrow \\mathrm{Pr}\\) task and \\(\\omega\\) from [5, 25] in \\(\\mathrm{Cl}\\rightarrow \\mathrm{Rw}\\) task while fixing other parameters to the default setting. We can draw two observations from Figs. 3a and 3b: the OS score is relatively stable, and the unknown recognition achieves a reliable performance when coefficient \\(\\beta\\) is within the interval of [0.02, 0.05]; Generally, a larger \\(\\omega\\), which is set for lower openness, will lead to a better performance on recognizing unknown data while hurt the accuracy for known classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.789, + 0.673 + ], + "angle": 0, + "content": "Feature Space Visualization To intuitively visualize the effectiveness of OUDA approaches, we extracted features from the baseline models (OSBP, PGL, ANNA) and our proposed model on the \\(\\mathrm{Ar}\\rightarrow \\mathrm{Cl}\\) task (Office-Home) and Syn2Real-O task with the ResNet-50 backbone [13]. The feature distributions were processed with t-SNE [23] afterward. As shown in Fig. 4, compared with baselines, our method PUJE achieves a better alignment between source and target distributions, especially when the domain shift is large. Benefiting from our joint error-based adversarial alignment mechanism, the extracted feature space, including the cluster of unknown target data, has a more discriminative class-wise decision boundary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.714, + 0.36, + 0.73 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We have addressed the open-set domain shift problem by introducing a novel learning theory based on multi-class positive-unlabeled learning that can reduce the open-set risk and the joint error. Experiments show that our proposed learning theory performs consistently better on challenging object recognition benchmarks for open-set adaptation with significant domain gaps and label shifts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Open-set DA via Joint Error based PU Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.221 + ], + "angle": 0, + "content": "This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.257 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.298 + ], + "angle": 0, + "content": "1. Baktashmotlagh, M., Faraki, M., Drummond, T., Salzmann, M.: Learning factorized representations for open-set domain adaptation. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.299, + 0.786, + 0.325 + ], + "angle": 0, + "content": "2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.: A theory of learning from different domains. Machine Learning 79, 151-175 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.786, + 0.352 + ], + "angle": 0, + "content": "3. Bucci, S., Loghmani, M.R., Tommasi, T.: On the effectiveness of image rotation for open set domain adaptation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.694, + 0.365 + ], + "angle": 0, + "content": "4. Busto, P.P., Gall, J.: Open set domain adaptation. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.786, + 0.392 + ], + "angle": 0, + "content": "5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: CVPRW (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.786, + 0.433 + ], + "angle": 0, + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.786, + 0.46 + ], + "angle": 0, + "content": "7. Fang, Z., Lu, J., Liu, F., Xuan, J., Zhang, G.: Open set domain adaptation: Theoretical bound and algorithm. arXiv abs/1907.08375 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.786, + 0.487 + ], + "angle": 0, + "content": "8. Feng, Q., Kang, G., Fan, H., Yang, Y.: Attract or distract: Exploit the margin of open set. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.786, + 0.528 + ], + "angle": 0, + "content": "9. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of Machine Learning Research 17(1), 2096-2030 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.786, + 0.555 + ], + "angle": 0, + "content": "0. Garg, S., Balakrishnan, S., Lipton, Z.C.: Domain adaptation under open set label shift. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.786, + 0.582 + ], + "angle": 0, + "content": "1. Gomes, R., Krause, A., Perona, P.: Discriminative clustering by regularized information maximization. In: NIPS (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.786, + 0.609 + ], + "angle": 0, + "content": "2. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NIPS (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.786, + 0.637 + ], + "angle": 0, + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. IEEE Conference on Computer Vision and Pattern Recognition pp. 770-778 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.786, + 0.663 + ], + "angle": 0, + "content": "4. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.786, + 0.69 + ], + "angle": 0, + "content": "5. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.786, + 0.732 + ], + "angle": 0, + "content": "6. Lin, T.Y., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.733, + 0.786, + 0.758 + ], + "angle": 0, + "content": "7. Liu, H., Cao, Z., Long, M., Wang, J., Yang, Q.: Separate to adapt: Open set domain adaptation via progressive separation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.785 + ], + "angle": 0, + "content": "8. Loghmania, M.R., Vinczea, M., Tommasi, T.: Positive-unlabeled learning for open set domain adaptation. Pattern Recognition Letters 136, 198-204 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "9. Long, M., Cao, Y., Wang, J., Jordan, M.I.: Learning transferable features with deep adaptation networks. In: Proceedings of the 32nd International Conference on International Conference on Machine Learning. vol. 37, pp. 97-105. JMLR.org (2015)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "D.Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "20. Long, M., CAO, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. In: Advances in Neural Information Processing Systems, pp. 1640-1650. Curran Associates, Inc. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "21. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 2208-2217. JMLR.org (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "22. Luo, Y., Wang, Z., Huang, Z., Baktashmotlagh, M.: Progressive graph learning for open-set domain adaptation. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.286 + ], + "angle": 0, + "content": "23. van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "24. Oliver, A., Odena, A., Raffel, C., Cubuk, E.D., Goodfellow, I.J.: Realistic evaluation of deep semi-supervised learning algorithms. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "25. Peng, X., Usman, B., Kaushik, N., Hoffman, J., Wang, D., Saenko, K.: Visda: The visual domain adaptation challenge. arXiv abs/1710.06924 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "26. Peng, X., Usmana, B., Saito, K., Kaushik, N., Hoffman, J., Saenko, K.: Syn2real: A new benchmark for synthetic-to-real visual domain adaptation. arXiv abs/1806.09755 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "27. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3723-3732 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.787, + 0.452 + ], + "angle": 0, + "content": "28. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "29. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NIPS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "30. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "31. Tang, H., Chen, K.C., Jia, K.: Unsupervised domain adaptation via structurally regularized deep clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "32. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 2962-2971 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "33. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 5385-5394 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "34. Wang, Q., Meng, F., Breckon, T.P.: Progressively select and reject pseudo-labelled samples for open-set domain adaptation. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.688 + ], + "angle": 0, + "content": "35. Wu, J., He, J.: Domain adaptation with dynamic open-set targets. In: KDD (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "36. Xu, Y., Xu, C., Xu, C., Tao, D.: Multi-positive and unlabeled learning. In: IJCAI (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.743 + ], + "angle": 0, + "content": "37. Zhang, D., Westfechtel, T., Harada, T.: Unsupervised domain adaptation via minimized joint error. In: TMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "38. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Proceedings of Machine Learning Research. vol. 97, pp. 7523-7532. PMLR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_origin.pdf b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed6c27ac88e89338924b253fa1707a65bfe60719 --- /dev/null +++ b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/15ee46ff-3473-42f3-b21c-e6a7457772f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bb94a424f1428f0b1e47c3cf1565539eca4197d447db5b1e75eeab8f8ae312c +size 2934220 diff --git a/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/full.md b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2440f07253c83acc1166193fec11a17b3276b71d --- /dev/null +++ b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/full.md @@ -0,0 +1,383 @@ +# Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning + +Dexuan Zhang $^{1}$ , Thomas Westfechtel $^{1}$ , and Tatsuya Harada $^{1,2}$ + +1 The University of Tokyo +2 RIKEN + +Abstract. Open-set domain adaptation aims to improve the generalization performance of a learning algorithm on a more realistic problem of open-set domain shift where the target data contains an additional unknown class that is not present in the source data. Most existing algorithms include two phases that can be described as closed-set domain adaptation given heuristic unknown class separation. Therefore, the generalization error cannot be strictly bounded due to the gap between the true distribution and samples inferred from heuristics. In this paper, we propose an end-to-end algorithm that tightly bound the risk of the entire target task by positive-unlabeled (PU) learning theory and the joint error from domain adaptation. Extensive experiments on various data sets demonstrate the effectiveness and efficiency of our proposed algorithm over open-set domain adaptation baselines. + +# 1 Introduction + +Given enough annotated training data, deep learning models can significantly improve the performance across a wide variety of machine-learning tasks but usually cannot generalize well to new domains, as it is commonly assumed that the training and test data are drawn from the same distribution. In practice, however, this assumption can be violated by several factors, such as the change in light, noise, the angle at which the image is captured, and different types of sensors, which is referred to as the domain shift that can harm the performance when predicting the test data. As a solution, domain adaptation (DA) aims to transfer the knowledge learned from a source domain, which is typically fully labeled, into a different (although related) target domain. + +As a more realistic setting [28], open-set domain adaptation allows the target data to contain an additional unknown category, covering all irrelevant classes not present in the source domain. The main idea of open-set unsupervised domain adaptation (OUDA) approaches [1, 4, 8, 17, 28] is to jointly learn a classifier from a hypothesis space for known and unknown classes in the source and target domains. According to [2], the target error can be bounded by the source error, the discrepancy distance between domains, the joint error coming from the conditional shift [38], and the open-set risk that indicates the precision of unknown class recognition. Open-set risk contributes substantially to the error + +bound when a large amount of target data is from the unknown class. However, the latter two terms cannot be explicitly computed. + +Unsupervised domain adaptation (UDA) methods [9, 19-21, 27, 32] show effectiveness in adapting unlabeled data to new domains by distribution alignment, but can fail to learn discriminative class boundaries, especially when the domain shift is large. Most existing methods tend to ignore the joint error and only focus on minimizing the discrepancy between domains, where samples from different classes can be grouped if the domain shift is large enough. In that case, the joint error becomes non-negligible, and the target error cannot be strictly bounded [9, 38]. [37] provided a solution to this problem by incorporating the joint error into the target error upper bound in the unsupervised setting, and we generalize this idea towards the open-set domain adaptation setting. + +Despite the promising performance, existing OUDA approaches [1, 3, 4, 8, 15, 17, 28, 34] lack an essential theoretical analysis of the generalization error for the target risk. Some works claim to derive a rigorous target error bound, while the minimization of joint error and the open-set risk are not theoretically guaranteed [7, 22]. Moreover, most existing algorithms can be described as closed-set UDA after separation of the unknown [3, 22, 34, 35]. In that case, the open-set risk cannot be controlled since the generalization error is also affected by the gap between empirical known target distribution and inferred known target samples, which becomes hard to analyze. + +Fig. 1a illustrates the potential problems of existing methods. Generally, it is impossible to perfectly distinguish the unknown such that the following marginal distribution alignment for UDA part is conducted under a large label shit. In that case, the joint error would be crucial as there exists a trade-off between marginal discrepancy and joint error [38]. Even if somehow we manage to obtain a perfect separation, some target data class will be dragged towards the outside of the corresponding decision boundary. In addition, the unknown class cannot be grouped as a single cluster. Fig. 1b intuitively explains the core concept of our algorithm PUJE Sec. 3, where we introduce two additional task classifiers and minimize the predictive discrepancy between the classifiers with same color, such that the unknown class will be pushed away from source data and grouped as a single cluster. Furthermore, those target data outside the corresponding decision boundary will be aligned back to the correct places. + +OUDA problems can be considered as multi-class positive-unlabeled learning problems (MPU) [36] with domain shift, where all the shared classes appearing in both source and target domains are positive and the unknown class is negative. Therefore, to minimize the aforementioned joint error and open-set risk to achieve a tighter error bound for open-set adaptation, instead of unknown separation and closed-set UDA, we combine PU learning theory and joint error-based target error bound into an end-to-end learning framework. The following methods also introduced PU learning to tackle OUDA problems. [18] required the unsupervised source-like reconstruction for target data via encoder-decoder models whose performance cannot be guaranteed especially for large domain shift and complex dataset; [10] only dealt with the open-set label shift and assumed an identical + +![](images/17b323858c0514ade4d8d4ed8fcfad8dfee3f213db051ac5107f6b8b35596180.jpg) +Fig. 1: Intuitive explanation of the difference between our algorithm PUJE and existing methods. (a) existing methods do not explicitly minimize joint error and cannot group unknown class as a single cluster; (b) our proposal is an upper bound of joint error which can address large domain shift and group unknown class into a single cluster. + +class conditional distribution of each instance across domains; [35] tackled a slightly different task where the target data is changing over time and it did not cover joint error and the gap between the empirical and inferred known target distribution. Compared with previous works, the main contributions of this paper can be summarized as follows: + +- We propose an end-to-end algorithm with theoretical analysis for open-set problem setting by PU learning-induced joint error based target error bound; +- We define a novel discrepancy measurement, namely open-set margin discrepancy (OMD), which allows us to smoothly apply the strategy in PU learning to the target error bound in domain adaptation; +- Extensive experiments confirm the effectiveness and efficiency of the proposal. + +# 2 Preliminaries + +In this section, we introduce notations, problem settings, and theoretical definitions for the tasks of closed-set unsupervised domain adaptation, open-set unsupervised domain adaptation, and multi-class positive and unlabeled learning. + +Definition 1 (Closed-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of $n$ labeled points $\hat{S} = \{(x_s^i,y_s^i)\in (\mathcal{X}\subseteq \mathbb{R}^D\times \mathcal{Y} = \{1,\dots,K\})\}_{i = 1}^n$ from the source domain $S$ , and a set of $m$ unlabeled points $\hat{T} = \{(x_t^i)\in \mathcal{X}\}_{i = 1}^m$ from a different target domain $T$ . With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier $f:\mathcal{X}\to \mathcal{Y}$ . + +Definition 2 (Multi-class Positive and Unlabeled Learning). For a classification task, the learning algorithm has access to a set of $n$ labeled points $\hat{P} = \{(x_p^i, y_p^i) \in (\mathcal{X} \times \mathcal{Y}' = \{1, ..., K - 1\})\}_{i=1}^n$ from the positive domain $P$ , and a set of $m$ unlabeled points $\hat{U} = \{(x_u^i) \in \mathcal{X}\}_{i=1}^m$ from the unlabeled domain $U$ , where $y_u^i \in \mathcal{Y}$ . With training samples drawn i.i.d from positive and unlabeled domains, the goal is to learn an optimal classifier $f: \mathcal{X} \to \mathcal{Y}$ for the unlabeled domain. Here, $U$ includes samples from unknown class $K$ , but the class conditional distributions for known classes remain invariant between $P, U$ . + +Definition 3 (Open-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of $n$ labeled points $\hat{S}' = \{(x_{s'}^i, y_{s'}^i) \in (\mathcal{X} \times \mathcal{Y}')\}_{i=1}^n$ from the incomplete source domain $S'$ , and a set of $m$ unlabeled points $\hat{T} = \{(x_t^i) \in \mathcal{X}\}_{i=1}^m$ from a different target domain $T$ , where $y_t^i \in \mathcal{Y}$ . With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier $f: \mathcal{X} \to \mathcal{Y}$ . + +We show the relation between open-set domain adaptation and multi-class positive and unlabeled learning as follows, which includes the hits on how to apply the domain adaptation algorithm in the open-set scenario. + +- In Definition 1, there is a domain shift between the distributions of the samples from the source and target domains such that a classifier trained on $S$ cannot generalize on $T$ . A typical solution is to align the feature distributions of both domains by a feature extractor [9,32]. +- In Definition 2, the conditional distributions of the samples from known class are identical for positive and unlabeled domains. As proved in [36], the expected error on the unlabeled domain can be approximated by the expected error on the positive domain and the probability that the unlabeled sample has not been classified as unknown. +- In Definition 3, the incomplete source domain $S'$ can be regarded as the positive domain $P$ in MPU. The target domain $T$ can be treated as $U$ with a domain shift. Assume that we can bridge the feature distributions with UDA techniques, it becomes possible to address open-set problems with well-established positive and unlabeled (PU) learning theories. + +# 3 PU Learning induced Joint Error based OUDA (PUJE) + +In this section, inspired by [37], we first rigorously reformulate the joint error-based target upper bound to facilitate the further derivation of a practical objective function for OUDA. To estimate the expectation over the complete source domain $S$ , we deploy the theory from multi-class positive and unlabeled learning to address the open-set problems where the source domain $S'$ is incomplete. We lastly show the entire loss of PUJE by combining the aforementioned joint error and PU learning. + +Theorem 1 (Approximated Joint Error based Target Upper Bound $^3$ ). Given the output space $\mathcal{K} = \{k|k\in \mathbb{R}^K:\sum_{y\in \mathcal{Y}}k[y] = 1,k[y]\in [0,1]\}$ , let $f_{S},f_{T}:\mathcal{X}\to \mathcal{K}$ be the true labeling functions for the source and target domains respectively, whose outputs are one-hot vectors denoting the corresponding labels of inputs. Let $\epsilon :\mathcal{K}\times \mathcal{K}\rightarrow \mathbb{R}$ denote a distance metric and $\epsilon_D(f,f')\coloneqq \mathbb{E}_{x\sim D}\epsilon (f(x),f'(x))$ measure the expected disagreement between the outputs of $f,f':\mathcal{X}\to \mathcal{K}$ over a distribution $D$ on $\mathcal{X}$ . For $\forall f_S^*\in \mathcal{H}_S\subseteq \mathcal{H},\forall f_T^*\in \mathcal{H}_T\subseteq \mathcal{H},\forall h\in \mathcal{H}:\mathcal{X}\to \mathcal{K}$ where $h(x)[y]$ indicates the probability of $x\in \mathcal{X}$ labeled as $y\in \mathcal{V}$ , the expected target error is bounded by + +$$ +\epsilon_ {T} (h) \leq \epsilon_ {S} (h) + \epsilon_ {T} \left(f _ {S} ^ {*}, f _ {T} ^ {*}\right) + \epsilon_ {T} \left(h, f _ {S} ^ {*}\right) - \left| \epsilon_ {S} \left(f _ {S} ^ {*}, f _ {T} ^ {*}\right) - \epsilon_ {S} \left(h, f _ {T} ^ {*}\right) \right| + \theta \tag {1} +$$ + +$$ +\theta = 2 \epsilon_ {T} \left(f _ {S}, f _ {S} ^ {*}\right) + \epsilon_ {S} \left(f _ {S}, f _ {S} ^ {*}\right) + 2 \epsilon_ {S} \left(f _ {T} ^ {*}, f _ {T}\right) + \epsilon_ {T} \left(f _ {T} ^ {*}, f _ {T}\right) = \theta_ {f _ {S}} + \theta_ {f _ {T}} \tag {2} +$$ + +According to Theorem 1, we show that the expected target error $\epsilon_T(h) \coloneqq \epsilon_T(h, f_T)$ is bounded by the expected source error $\epsilon_S(h) \coloneqq \epsilon_S(h, f_S)$ , the discrepancy between domains $[\epsilon_T(f_S^*, f_T^*) + \epsilon_T(h, f_S^*) - |\epsilon_S(f_S^*, f_T^*) - \epsilon_S(h, f_T^*)|]$ , and the deviation from true labeling functions $\theta$ . The following Assumption 1 allows us to derive the generalization error of Theorem 1 such that we can ignore the residual term $\theta$ and relate target error bound with the empirical estimation of source error and the discrepancy between domains. + +Assumption 1. Let $\hat{S},\hat{T}$ denote empirical samples with finite size from source and target domains. Assume that there exists approximated labeling functions $f_{S}^{*},f_{T}^{*}$ such that the empirical deviation $\hat{\theta}_{f_S},\hat{\theta}_{f_T}$ are close enough to zero. + +Definition 4 (Empirical Rademacher Complexity). Let $\mathcal{F}$ be a class of real-valued functions: $\mathcal{X} \to \mathbb{R}$ and $\hat{D} = \{x_1, \ldots, x_m\}$ a finite sample drawn i.i.d. from a distribution $D$ , the empirical Rademacher Complexity of $\mathcal{F}$ is defined by + +$$ +\hat {\mathfrak {R}} _ {\hat {D}} (\mathcal {F}) = \frac {1}{m} \mathbb {E} _ {\sigma} [ \sup _ {f \in \mathcal {F}} \sum_ {i = 1} ^ {m} \sigma_ {i} f (x _ {i}) ], \tag {3} +$$ + +where $\sigma_{i}$ is an independent uniform random variable taking values in $\{-1, + 1\}$ . + +Theorem 2 (Generalization Error $^3$ ). Given Theorem 1 and Definition 4, Assumption 1 and function space $\mathcal{F}_{f,f'}^{\epsilon}:\mathcal{X}\to \epsilon (f(X),f'(X))$ bounded by $M > 0$ , for any $\delta >0$ , with probability at least $1 - 2\delta$ , for $\forall h\in \mathcal{H}$ : + +$$ +\begin{array}{l} \epsilon_ {T} (h) \leq \epsilon_ {\hat {S}} (h) + \epsilon_ {\hat {T}} (f _ {S} ^ {*}, f _ {T} ^ {*}) + \epsilon_ {\hat {T}} (h, f _ {S} ^ {*}) - | \epsilon_ {\hat {S}} (f _ {S} ^ {*}, f _ {T} ^ {*}) - \epsilon_ {\hat {S}} (h, f _ {T} ^ {*}) | + 2 [ \hat {\Re} _ {\hat {S}} (\mathcal {F} _ {h, f _ {S}} ^ {\epsilon}) ] \\ + \hat {\Re} _ {\hat {T}} (\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\epsilon}) + \hat {\Re} _ {\hat {T}} (\mathcal {F} _ {h, f _ {S} ^ {*}} ^ {\epsilon}) + \hat {\Re} _ {\hat {S}} (\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\epsilon}) + \hat {\Re} _ {\hat {S}} (\mathcal {F} _ {h, f _ {T} ^ {*}} ^ {\epsilon}) + \hat {\Re} _ {\hat {S}} (\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\epsilon}) (4) \\ + 2 \hat {\Re} _ {\hat {S}} (\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\epsilon}) + \hat {\Re} _ {\hat {T}} (\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\epsilon}) + 2 \hat {\Re} _ {\hat {T}} (\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\epsilon}) ] + 3 M (5 \sqrt {\frac {\log \frac {9}{\delta}}{2 m}} + 6 \sqrt {\frac {\log \frac {9}{\delta}}{2 n}}) \\ \end{array} +$$ + +In the following part, we introduce the techniques from PU learning to estimate the expectation over source domain $S$ by the incomplete source domain $S'$ and target domain $T$ for the open-set scenario. + +Definition 5 (Unknown Predictive Discrepancy). Let $v: \mathcal{K} \times \mathcal{K} \to \mathbb{R}$ denote the Unknown Predictive Discrepancy as a distance metric and $v_{D}(f, f') := \mathbb{E}_{x \sim D} v(f(x), f'(x))$ measure the expected disagreement between the $K$ -th outputs of $f, f': \mathcal{X} \to \mathcal{K}$ over a distribution $D$ on $\mathcal{X}$ . Let $e^{K}: \mathcal{X} \to [0, \dots, 1] \in \mathcal{K}$ denote a function that can predict any input as unknown. The deviation from $e^{K}$ for a hypothesis $h \in \mathcal{H}$ is referred to as the shorthand $v_{D}(h) := v_{D}(h, e^{K})$ that measures the probability that samples from $D$ have not been classified as unknown. + +Assumption 2. Let $S^i = P_S(x|y = i), T^i = P_T(x|y = i)$ denote class conditional distributions, $S' = P_S(x|y \neq K), T' = P_T(x|y \neq K)$ indicate incomplete domains that do not contain unknown class $S^K, T^K$ . Given a feature extractor $g: \mathcal{X} \subseteq \mathbb{R}^D \to \mathcal{Z} \subseteq \mathbb{R}^F$ , assume that the feature space can be aligned by UDA techniques such that $Z^K = P_{S^K}(z) = P_{T^K}(z), Z' = P_{S'}(z) = P_{T'}(z)$ . + +Proposition 1. Let $h \in \mathcal{H}_{F_g} : \mathcal{Z} \to \mathcal{K}$ denote the decomposed hypothesis where $h \circ g \in \mathcal{H}$ . Given Definition 5, Assumption 2, $v_{S^K}(h \circ g) = v_{T^K}(h \circ g)$ that represents probability that samples from $S^K$ have not been classified as unknown class, is equivalent to $\epsilon_{S^K}(h \circ g)$ that measures the classification error on unknown class according to the definition. + +Lemma 1 (Estimated Source Error $^3$ ). Let $\sum_{i=1}^{K} \pi_S^i = 1$ , $\sum_{i=1}^{K} \pi_T^i = 1$ denote the label distribution of $S$ and $T$ respectively. Given Proposition 1, the expected error on $S$ can be estimated by the error on $S'$ and Unknown Predictive Discrepancy (Definition 5) on $T$ with a mild condition that $\pi_S^K = \pi_T^K = 1 - \alpha$ : + +$$ +\epsilon_ {S} (h \circ g) = \alpha [ \epsilon_ {S ^ {\prime}} (h \circ g) - v _ {S ^ {\prime}} (h \circ g) ] + v _ {T} (h \circ g) \tag {5} +$$ + +Remark 1. According to Definition 5, minimizing $v_{T}(h \circ g)$ means classifying target samples as the unknown class. In practice, a multiplier $\beta < 1$ is applied on $v_{T}(h \circ g)$ to prevent all target samples from being recognized as unknown. + +To reformulate the intractable discrepancy $\epsilon_S(f_S^*,f_T^*)$ , $\epsilon_S(h,f_T^*)$ in Theorem 1, we decompose the approximated labeling functions given feature extractor $g$ such that $f_{S}^{\star}\circ g = f_{S}^{\star},f_{T}^{\star}\circ g = f_{T}^{\star}$ where $f_{S}^{\star}\in \mathcal{H}_{S_{g}}\subseteq \mathcal{H}_{F_{g}},f_{T}^{\star}\in \mathcal{H}_{T_{g}}\subseteq \mathcal{H}_{F_{g}}$ . We further define $\epsilon$ in a way such that, under some circumstances, it is equivalent to the Unknown Predictive Discrepancy for $\forall f\in \mathcal{H}_{F_g}$ + +$$ +\epsilon_ {T ^ {K}} \left(f \circ g, f _ {T} ^ {\star} \circ g\right) = v _ {T ^ {k}} \left(f \circ g, f _ {T} ^ {\star} \circ g\right) \tag {6} +$$ + +Corollary 1. Given Assumption 2, the feature extractor $g$ can align the feature distributions of source and target domain such that the expectation over $S^K$ is equal to that over $T^K$ , which leads Eq. (6) to: $\epsilon_{S^K}(f \circ g, f_T^\star \circ g) = \epsilon_{T^K}(f \circ g, f_T^\star \circ g) = v_{T^K}(f \circ g, f_T^\star \circ g) = v_{S^K}(f \circ g, f_T^\star \circ g)$ . + +Lemma 2 (Estimated Discrepancy $^3$ ). Given Corollary 1, analogous to Lemma 1, the discrepancy measured on source domain $\epsilon_S(f_S^*, f_T^*)$ , $\epsilon_S(h, f_T^*)$ can be estimated by the discrepancy on $S'$ and Unknown Predictive Discrepancy (Definition 5) with a mild condition that $\pi_S^K = \pi_T^K = 1 - \alpha$ : + +$$ +\left\{ \begin{array}{l l} \epsilon_ {S} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) & = \alpha \left[ \epsilon_ {S ^ {\prime}} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) - v _ {S ^ {\prime}} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) \right] + v _ {T} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) \\ \epsilon_ {S} \left(h \circ g, f _ {T} ^ {\star} \circ g\right) & = \alpha \left[ \epsilon_ {S ^ {\prime}} \left(h \circ g, f _ {T} ^ {\star} \circ g\right) - v _ {S ^ {\prime}} \left(h \circ g, f _ {T} ^ {\star} \circ g\right) \right] + v _ {T} \left(h \circ g, f _ {T} ^ {\star} \circ g\right) \end{array} \right. \tag {7} +$$ + +Definition 6 (Overall Loss of PUJE). Let $\hat{S}',\hat{T}$ denote empirical data from $S',T$ . Given a feature extractor $g:\mathcal{X}\subseteq R^{D}\to \mathcal{Z}\subseteq \mathbb{R}^{F}$ , decomposed hypothesis and approximated labeling functions $h,f_{S}^{\star},f_{T}^{\star}\in \mathcal{H}_{F_{g}}:\mathcal{Z}\to \mathcal{K}$ , replacing the intractable terms in Theorem 2 with Eqs. (5) and (7) and ignoring the Rademacher Complexity terms, the overall loss of PUJE can be regrouped as the sum of classification loss $L_{cls}$ and the discrepancy between domains $L_{dis}$ : + +$$ +\begin{array}{l} L _ {p u j e} = \alpha \left[ \epsilon_ {\hat {S} ^ {\prime}} (h \circ g) - v _ {\hat {S} ^ {\prime}} (h \circ g) \right] + \beta v _ {\hat {T}} (h \circ g) + \epsilon_ {\hat {T}} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) + \epsilon_ {\hat {T}} (h \circ g, f _ {S} ^ {\star} \circ g) \\ - \left| v _ {\hat {T}} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) - v _ {\hat {T}} \left(h \circ g, f _ {T} ^ {\star} \circ g\right) + \alpha \left[ \epsilon_ {\hat {S} ^ {\prime}} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) \right. \right. \\ \left. + v _ {\hat {S} ^ {\prime}} \left(h \circ g, f _ {T} ^ {\star} \circ g\right) - \epsilon_ {\hat {S} ^ {\prime}} \left(h \circ g, f _ {T} ^ {\star} \circ g\right) - v _ {\hat {S} ^ {\prime}} \left(f _ {S} ^ {\star} \circ g, f _ {T} ^ {\star} \circ g\right) \right] | \\ = L _ {c l s} (h; g) + L _ {d i s} \left(f _ {S} ^ {\star}, f _ {T} ^ {\star}, h; g\right) \tag {8} \\ \end{array} +$$ + +# 4 Methodology + +In this section, we first define Approximated Labeling Function Space $\mathcal{H}_{S_g},\mathcal{H}_{T_g}$ such that the intractable $L_{dis}(f_S^\star ,f_T^\star ,h;g)$ due to unknown $f_S^\star ,f_T^\star$ can be upper bounded by $\sup_{f_S'\in \mathcal{H}_{S_g},f_T'\in \mathcal{H}_{T_g}}L_{dis}(f_S',f_T',h;g)$ . To fulfill Eq. (6), we propose Open-set Margin Discrepancy to quantify $\epsilon$ that measures the disagreement between classifiers. We lastly show the entire training algorithm. + +# 4.1 Approximated Labeling Function Space + +Proposition 2. Let $\mathcal{H}_S, \mathcal{H}_T : \mathcal{X} \to \mathcal{K}$ be two sets of functions that can minimize a part of the empirical residual $\hat{\theta}_S, \hat{\theta}_T$ respectively, $f_S^* \in \mathcal{H}_S$ , $f_T^* \in \mathcal{H}_T$ must hold as $f_S^*, f_T^*$ can minimize the entire $\hat{\theta}$ given Assumption 1. Accordingly, decomposed functions $f_S^*, f_T^*$ must lie in $\mathcal{H}_{S_g} = \{f|\forall f \circ g \in \mathcal{H}_S\}$ , $\mathcal{H}_{T_g} = \{f|\forall f \circ g \in \mathcal{H}_T\}$ such that a sufficient condition of the following inequality: $L_{dis}(f_S^*, f_T^*, h; g) \leq \sup_{f_S' \in \mathcal{H}_{S_g}, f_T' \in \mathcal{H}_{T_g}} L_{dis}(f_S', f_T', h; g)$ is fulfilled. + +Definition 7 (Approximated Labeling Function Space). Let $L_{\mathcal{H}_S}, L_{\mathcal{H}_T}$ denote a part of the empirical residual $\hat{\theta}_{fs}, \hat{\theta}_{f_T}$ respectively. Let $f_S^\star \circ g = f_S^* \in \mathcal{H}_S, f_T^\star \circ g = f_T^* \in \mathcal{H}_T$ denote the decomposed approximated labeling functions where $f_S^\star \in \mathcal{H}_{S_g} \subseteq \mathcal{H}_{F_g}, f_T^\star \in \mathcal{H}_{T_g} \subseteq \mathcal{H}_{F_g}$ . Given Proposition 2, Approximated Labeling Function Space $\mathcal{H}_{S_g}, \mathcal{H}_{T_g}$ can be defined as the sets whose members $f_S', f_T' \in \mathcal{H}_{F_g}$ can minimize $L_{\mathcal{H}_S}, L_{\mathcal{H}_T}$ : + +$$ +\left\{ \begin{array}{l} \mathcal {H} _ {S _ {g}} = \left\{f _ {S} ^ {\prime} \right| \arg \min _ {g, f _ {S} ^ {\prime} \in \mathcal {H} _ {F _ {g}}} L _ {\mathcal {H} _ {S}} \left(f _ {S} ^ {\prime}; g\right) := L _ {c l s} \left(f _ {S} ^ {\prime}; g\right) \} \\ \mathcal {H} _ {T _ {g}} = \left\{f _ {T} ^ {\prime} \right| \arg \min _ {g, f _ {T} ^ {\prime} \in \mathcal {H} _ {F _ {g}}} L _ {\mathcal {H} _ {T}} \left(f _ {T} ^ {\prime}; g\right) := (1 - \gamma) L _ {c l s} \left(f _ {T} ^ {\prime}; g\right) + \gamma L _ {r e g} \end{array} \right. \tag {9} +$$ + +$\mathcal{H}_{S_g}$ consists of functions minimizing the expected error on $S$ , which can be estimated by Lemma 1. To build a reliable function space $\mathcal{H}_{T_g}$ without target labels, we approximate the target error with the weighted average $(\gamma \in [0,1])$ of error rate on labeled samples and a semi-supervised regularization term $L_{reg} = L_{ent} + L_{pse} + \omega L_{con}$ . + +Regularized Entropy Minimization As introduced in [11,27,31], we impose a class balance prior that can penalize classifiers with complex decision boundaries on entropy minimization [12] to yield a more sensible solution: + +$$ +\begin{array}{l} L _ {e n t} = - \mathbb {E} _ {x \in \hat {T}} \sum_ {y \in \mathcal {Y}} f _ {T} ^ {\prime} (g (x)) [ y ] \log f _ {T} ^ {\prime} (g (x)) [ y ] \\ + \sum_ {y \in \mathcal {Y}} \mathbb {E} _ {x \in \hat {T}} f _ {T} ^ {\prime} (g (x)) [ y ] \log \mathbb {E} _ {x \in \hat {T}} f _ {T} ^ {\prime} (g (x)) [ y ] \tag {10} \\ \end{array} +$$ + +Pseudo Labeling As introduced in [30, 31], for input $x \in \hat{T}$ and its random augmentation $x'$ [5], we minimize cross entropy for $x$ with pseudo labels of $x'$ : + +$$ +L _ {p s e} = - \mathbb {E} _ {x \in \hat {T}} \log f _ {T} ^ {\prime} (g (x)) [ \arg \max _ {y \in \mathcal {Y}} h (g (x ^ {\prime})) [ y ] ] \tag {11} +$$ + +Consistency Regularization As introduced in [14, 29], we penalize the difference of the outputs for input $x \in \hat{T}$ and its random augmentation $x'$ : + +$$ +L _ {c o n} = \mathbb {E} _ {x \in \hat {T}} | f _ {T} ^ {\prime} (g (x)) - f _ {T} ^ {\prime} (g (x ^ {\prime})) | \tag {12} +$$ + +# 4.2 Open-set Margin Discrepancy + +Definition 8 (Induced Labeling Function). Let $f: \mathcal{X} \to \mathcal{K}$ denote a multiclass labeling function, where $f(x)[y]$ indicating $y$ -th element of the output $f(x)$ for the probability of $x$ classified as $y$ . Thus an induced labeling function $l \circ f: \mathcal{X} \to \mathcal{Y}$ is given by: $l(f(x)) = \arg \max_{y \in \mathcal{Y}} f(x)[y]$ . + +Definition 9 (Open-set Margin Discrepancy (OMD)). Let $y = l(f(x)), y' = l(f'(x))$ denote the predictions on input $x$ from $f, f': \mathcal{X} \to \mathcal{K}$ given Definition 8. Open-set Margin Discrepancy between $f, f'$ over a distribution $D$ is given by + +$$ +\begin{array}{l} \epsilon_ {D} (f, f ^ {\prime}) = \mathbb {E} _ {x \sim D} [ \operatorname {o m d} (f (x), f ^ {\prime} (x)) ] (13) \\ \operatorname {o m d} (f (x), f ^ {\prime} (x)) = \max \left(\left| \log (1 - f (x) [ y ]) - \log (1 - f ^ {\prime} (x) [ y ]) \right|\right), \\ \left. \left| \log \left(1 - f (x) \left[ y ^ {\prime} \right]\right) - \log \left(1 - f ^ {\prime} (x) \left[ y ^ {\prime} \right]\right) \right|\right) (14) \\ \end{array} +$$ + +For OUDA, generally the classifier is not discriminative for unknown class $K$ due to the lack of labeled samples to reduce the misclassification rate. To address this, we model $\log (1 - f(x)[K])$ that has a more stable gradient. + +Remark 2. According to Definition 5, $v_{D}(f,f^{\prime})$ measures the expectation of disagreement on the K-th output between two functions $f,f^{\prime}:\mathcal{X}\to \mathcal{K}$ over a distribution $D$ , which can also indicate the difference between probabilities of categorizing inputs as known classes. To fulfill Eq. (6), we quantify $v_{D}(f,f^{\prime})$ as: + +$$ +v _ {D} (f, f ^ {\prime}) = \mathbb {E} _ {x \sim D} | \log (1 - f (x) [ K ]) - \log (1 - f ^ {\prime} (x) [ K ]) | \tag {15} +$$ + +# Algorithm 1 PUJE + +Input: incomplete source data $\hat{S}^{\prime}$ ; unlabeled target data $\hat{T}$ + +Output: labeling functions $f_S', f_T'$ ; feature extractor $g$ ; hypothesis $h$ + +Parameter: trade-off parameter $\lambda$ ; learning rate $\eta$ ; known class ratio estimator $\alpha$ for epoch = 1, 2, ... do + +Step 1: estimate known class ratio $\alpha$ on $T$ with prediction of $g, h$ + +Step 2: optimize $g, f_S', f_T'$ to satisfy the approximated labeling function space + +$$ +(g, f _ {S} ^ {\prime}, f _ {T} ^ {\prime}) \leftarrow (g, f _ {S} ^ {\prime}, f _ {T} ^ {\prime}) + \eta \Delta (g, f _ {S} ^ {\prime}, f _ {T} ^ {\prime}) +$$ + +$$ +\varDelta (g, f _ {S} ^ {\prime}, f _ {T} ^ {\prime}) = - \frac {\partial (L _ {\mathcal {H} _ {S}} (f _ {S} ^ {\prime} ; g) + L _ {\mathcal {H} _ {T}} (f _ {T} ^ {\prime} ; g))}{\partial (g , f _ {S} ^ {\prime} , f _ {T} ^ {\prime})} +$$ + +Step 3: maximize the discrepancy w.r.t. $f_{S}^{\prime}, f_{T}^{\prime}$ within the function space + +$$ +\left(f _ {S} ^ {\prime}, f _ {T} ^ {\prime}\right) \leftarrow \left(f _ {S} ^ {\prime}, f _ {T} ^ {\prime}\right) + \eta \Delta \left(f _ {S} ^ {\prime}, f _ {T} ^ {\prime}\right) +$$ + +$$ +\varDelta(f_{S}^{\prime},f_{T}^{\prime})=-\frac{\partial(L_{\mathcal{H}_{S}}(f_{S}^{\prime};g)+L_{\mathcal{H}_{T}}(f_{T}^{\prime};g)-\lambda L_{dis}(f_{S}^{\prime},f_{T}^{\prime},h;g))}{\partial(f_{S}^{\prime},f_{T}^{\prime})} +$$ + +Step 4: minimize the entire target error bound w.r.t. $g, h$ for fixed $f_S', f_T'$ . + +$$ +(g, h) \leftarrow (g, h) + \eta \Delta (g, h) +$$ + +$$ +\Delta (g, h) = - \frac {\partial \left(L _ {c l s} (h ; g) + \lambda L _ {d i s} \left(f _ {S} ^ {\prime} , f _ {T} ^ {\prime} , h ; g\right)\right)}{\partial (g , h)} +$$ + +end for + +# 4.3 Training Algorithm + +Given the feature extractor $g: \mathcal{X} \subseteq \mathbb{R}^D \to \mathcal{Z} \subseteq \mathbb{R}^F$ and hypotheses $h, f_S', f_T' \in \mathcal{H}_{F_g}: \mathcal{Z} \to \mathcal{K}$ , we introduce a trade-off parameter $\lambda$ to balance the classification loss and discrepancy as Eq. (16). During each epoch, we first estimate the known class ratio $\alpha$ (initially set to 1) in target data by the predictions of $h$ . Then we train the hypothesis $h$ , approximated labeling functions $f_S', f_T'$ , feature extractor $g$ to minimize the error on labeled data $L_{cls}$ with hypothesis space constraints $L_{\mathcal{H}_S}, L_{\mathcal{H}_T}$ , while optimizing the discrepancy $L_{dis}$ adversarially by the min-max game over $g$ and $f_S', f_T'$ (Algorithm 1): + +$$ +\left\{\min _ {f _ {S} ^ {\prime}, f _ {T} ^ {\prime} \in \mathcal {H} _ {F _ {g}}} L _ {\mathcal {H} _ {S}} \left(f _ {S} ^ {\prime}; g\right) + L _ {\mathcal {H} _ {T}} \left(f _ {T} ^ {\prime}; g\right) - \lambda L _ {d i s} \left(f _ {S} ^ {\prime}, f _ {T} ^ {\prime}, h; g\right) \right. \tag {16} +$$ + +$$ +\left\lfloor \min _ {h \in \mathcal {H} _ {F _ {g}}, g} L _ {\mathcal {H} _ {S}} \left(f _ {S} ^ {\prime}; g\right) + L _ {\mathcal {H} _ {T}} \left(f _ {T} ^ {\prime}; g\right) + L _ {c l s} (h; g) + \lambda L _ {d i s} \left(f _ {S} ^ {\prime}, f _ {T} ^ {\prime}, h; g\right) \right. +$$ + +# 4.4 Intuition + +In this section, we intuitively explain how our method can align the source and target domains while separating the unknown in Fig. 2. According to [24], unlike the entropy minimization that tends to cut through unlabeled data, consistency regularization helps to draw a manifold-aware decision boundary. Based on this conclusion, the decision boundary of $f_{T}^{\prime}$ induced by consistency regularization $L_{con}$ can preserve the cluster structure (Fig. 2a). According to Algorithm 1, $f_{S}^{\prime}$ tries to maximize $L_{dis}$ while classify target samples as unknown class to minimize $L_{\mathcal{H}_S}(f_S';g)$ (Definition 7 and Lemma 1), which can lead to a new decision boundary shown in Fig. 2b. Meanwhile, feature extractor $g$ tries to minimize $L_{dis}$ by pushing target samples close to $f_{S}^{\prime}$ towards source clusters, + +![](images/27899cb4ba862fd93fb884d91e7c91c655ce01785873ac7abe2c8609ed4c46ae.jpg) +(a) $\min_{f_T'} L_{\mathcal{H}_T}$ + +![](images/67564d6910b9b0116b35c85a28e1451426b0f91251ff90ed0a78e0bb78b10f4d.jpg) +(b) $\min_{f_S'}(L_{\mathcal{H}_S} - L_{dis})$ + +(c) $\min_g L_{dis}$ +Fig. 2: Intuitive explanation of the mechanism of the proposed PUJE in Algorithm 1. (a) $L_{\mathcal{H}_T}$ includes consistency regularization $L_{con}$ that helps to draw a manifold-aware decision boundary for $f_T'$ (Step 2); (b) $f_S'$ can increase $L_{dis}$ by classifying most of the unlabeled target samples as unknown, which also decreases $L_{\mathcal{H}_S}$ (Step 3); (c) feature extractor $g$ pushes unlabeled target samples close to the decision boundary of $f_S'$ towards source cluster, while those close to $f_T'$ towards outside to reduce $L_{dis}$ , which can lead to a separated unknown cluster (Step 4). +![](images/2dd6be1cf92e80d3475bcda7814410c86f3dfef6fd68d894034898d11410acb7.jpg) +$L_{dis}$ $f_T^{\prime}$ $f_S^{\prime}$ +- Positive Source +- Negative Source +- Unknown Target +- Pseudo Positive +- Pseudo Negative + +while those close to $f_{T}^{\prime}$ towards outside such that unlabeled target samples far from source clusters can be eventually separated and grouped into a new cluster for the unknown progressively as illustrated in Fig. 2c. + +# 5 Evaluation + +We evaluated our proposal on two benchmarks, Office-Home and Syn2Real-O. In the implementation, the $L_{reg}$ weight $\gamma$ and the trade-off parameter $\lambda$ were set to 0.1 and 0.01 according to [37]. In addition, we empirically set PU coefficient $\beta$ to 0.03 and consistency coefficient $\omega$ to 20 for Office-Home and 5 for Syn2Real-O. All experiments were conducted with the ImageNet [6] pre-trained ResNet-50 [13] as the feature extractor $g$ and 2-layer linear networks for classifiers $f_S', f_T', h$ . We trained the model by Stochastic Gradient Descent optimizer with a 0.001 learning rate annealed according to [9], 24 batch-size, a momentum of 0.9. We quantitatively compare our results against various OUDA baselines, including OSBP [28], STA [17], DAOD [8], PGL [22], ROS [3], OSLPP [34] and ANNA [15]. + +Evaluation Metrics To evaluate the proposed method and the baselines, we utilize the widely used measures [22, 28], i.e., normalized accuracy for all classes (OS), normalized accuracy for the known classes only $(\mathrm{OS}^{\star})$ and harmonic mean $\mathrm{HOS} = 2(\mathrm{OS}^{\star}\times \mathrm{UNK}) / (\mathrm{OS}^{\star} + \mathrm{UNK})$ [3,15,18,34]. + +Office-Home [33] is a widely-used domain adaptation benchmark, which consists of 15,500 images from 65 categories and four domains: Art (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw). Following the same splits used in previous methods [22], we select the first 25 classes in alphabetical order as the known classes and group the rest as the unknown. + +Table 1: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (OS) + +
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP47.165.364.638.348.748.327.081.679.537.173.572.129.355.354.337.781.980.2
STA77.048.449.595.440.842.959.177.376.671.268.668.761.045.446.058.974.573.9
DAOD44.367.766.844.760.359.740.885.083.349.873.272.347.460.459.956.882.881.8
PGL34.773.772.238.459.258.427.684.882.66.181.578.625.166.865.038.084.883.0
PUJE38.474.873.442.264.463.637.987.785.823.481.078.839.271.270.039.188.286.3
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP28.657.256.125.877.875.823.085.483.033.077.275.516.771.369.232.165.964.6
STA64.145.946.662.067.267.066.276.676.257.465.264.960.257.657.772.749.350.2
DAOD71.155.556.166.669.269.163.779.378.754.878.277.354.670.269.655.162.962.6
PGL19.163.361.632.178.977.140.987.785.95.385.982.824.573.972.033.870.268.8
PUJE57.166.666.352.182.881.645.288.987.255.981.880.851.481.580.342.571.069.9
+ +Table 2: Accuracy of ResNet-50 model fine-tuned on Syn2Real-O dataset. * indicates our re-implementation with the officially released code. + +
METHODplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckUNKOS*OSHOS
OSBP73.657.958.265.267.429.584.147.067.85.789.10.566.653.854.859.5
STA64.170.353.759.480.820.890.012.563.230.278.22.759.152.252.755.4
PGL81.568.374.260.691.945.492.241.087.967.579.26.449.666.865.556.9
PGL*81.377.566.571.790.635.792.441.282.646.982.34.959.664.564.162.0
ANNA41.254.639.759.451.225.682.654.878.910.576.62.067.448.149.656.1
PUJE90.773.875.172.477.833.891.258.879.546.584.95.169.065.866.067.3
+ +Syn2Real-O [26] is a more challenging synthetic-to-real benchmark, which is constructed from the VisDA-17 [25]. The Syn2Real-O dataset significantly increases the ratio of unknown data in the target domain to 0.9 by introducing additional unknown samples. In this dataset, the source domain contains training data from the VisDA-17 as the known set, and the target domain includes the test data from the VisDA-17 (known set) plus 50k images from irrelevant categories of MSCOCO [16] dataset (unknown set). + +As reported in Tabs. 1 and 2, we observe that our method PUJE consistently outperforms the state-of-the-art results, improving OS and HOS by $3.0\%$ and $5.3\%$ on the benchmark datasets of Office-Home and Syn2Real-O respectively. Note that our proposed approach provides significant performance gains for the more challenging dataset of Syn2Real-O that requires knowledge transfer across different modalities. In the sub-tasks with a larger domain shift, e.g., $\mathrm{Rw} \rightarrow \mathrm{Cl}$ and $\mathrm{Pr} \rightarrow \mathrm{Cl}$ in Office-Home, we can observe similar results demonstrating the strong adaptation ability of the proposed framework. PGL reported a higher OS score in Syn2Real-O, but we cannot reproduce the results with the officially released code. During the re-implementation based on their early stop strategy, we found the UNK was around $30\%$ , which is far below the reported score. A potential reason to explain this is that maybe they did not remove the unknown category from the source data. + +Table 3: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (HOS) + +
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOS
ROS64.357.360.671.246.556.378.470.874.470.867.068.8
OSLPP76.254.663.667.153.159.371.277.074.075.060.867.2
ANNA70.363.066.574.854.663.178.974.376.677.366.171.3
PUJE77.856.965.776.851.261.575.280.177.567.371.169.1
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*
ROS74.150.660.170.368.469.377.275.876.572.265.368.6
OSLPP67.155.961.073.172.572.869.480.174.373.967.266.9
ANNA78.761.469.079.968.373.779.774.176.880.266.973.0
PUJE83.558.568.878.979.679.376.382.179.180.671.275.6
+ +Table 4: Ablation studies of semi-supervised regularization losses. We report the Accuracy $(\%)$ on Office-Home of $\mathrm{A}\rightarrow \mathrm{R}$ and $\mathrm{C}\to \mathrm{P}$ using a ResNet50 backbone. + +
Config.Ar→PrCl→Rw
UNK OS* OSUNK OS* OS
default52.182.881.655.981.880.8
w/o Lpse69.678.177.864.877.877.3
w/o Lcon26.381.779.633.380.979.1
w/o Lent50.281.179.953.980.579.5
+ +HOS Metric Instead of OS score, [18] proposed the harmonic mean of $\mathrm{OS}^{\star}$ and unk, $\mathrm{HOS} = 2(\mathrm{OS}^{\star} \times \mathrm{UNK}) / (\mathrm{OS}^{\star} + \mathrm{UNK})$ to penalize large gaps between $\mathrm{OS}^{\star}$ and unk. We agree that the OS score does not reflect UNK when the number of known classes grows. However, the HOS score is still unfair. $\mathrm{OS}^{\star}$ and UNK are equally valued in HOS, but UNK is much easier to improve than $\mathrm{OS}^{\star}$ which depends on the power of the adaptation algorithm. [15] utilizes a weak adaptation algorithm based on [28] that leads to a relatively loose alignment (Fig. 4) on the feature space of source and target domains such that unlabeled data is more frequently assigned to the unknown class. We believe that if a method can achieve a high $\mathrm{OS}^{\star}$ score, it is possible to obtain a decent HOS score by tuning hyper-parameters or additional losses. For our proposal, we can increase $\beta = 0.25$ to improve UNK at the sacrifice of $\mathrm{OS}^{\star}$ (Tab. 3). + +Ablation Study To investigate the impact of the semi-supervised regularization, we compare three variants of the PUJE model on the Office-Home dataset shown in Tab. 4. Without $L_{pse}$ , $\mathrm{OS}^{\star}$ decreases because it can no longer leverage the pseudo labels that can help to build reliable hypothesis space $\mathcal{H}_{T_g}$ (Definition 7). Without $L_{con}$ , UNK drops as it is not likely to draw a manifold-aware decision boundary that is beneficial to the separation of unknown data (Fig. 2a). + +Robustness against Varying Openness To verify the robustness of the proposed PUJE, we conducted experiments on the Syn2Real-O with the openness varying in $\{0.25, 0.5, 0.75, 0.9\}$ . Here, openness is defined by the ratio of unknown + +![](images/9f623d8a8102881900f455a6d93b300329f2c188562ea187d1995eb93e2f1a24.jpg) +(a) PU Coefficient + +![](images/4a0b854f463874efefecc0c4a3413973e7bfba59c6490d606dc1dc83bac6687a.jpg) +(b) SSL Coefficient + +![](images/70c99c257b1de897960d3c2589029d98b9bfb0dd150cea78234988c9b4aa9322.jpg) +(c) Openness + +![](images/91fdbbd4da8cbf009efa2f4814fa03746412ba4ff43bd324322679c5de6f2ac8.jpg) +(d) Trade-off in $\mathrm{Ar}\rightarrow \mathrm{Cl}$ + +![](images/a2ed7e612b215ec82d0d1fd5bd51401cad5048f151cfdf52cfebd24bacc35099.jpg) +(e) Convergence in $\mathrm{Ar}\rightarrow \mathrm{Cl}$ + +![](images/5441e85fb22f0abe9a4774e6021834e77cc845b6e0979bbd0bc3acc26c3b8b2c.jpg) +(f) Convergence in $\mathrm{Rw}\rightarrow \mathrm{Cl}$ +Fig. 3: (a)-(b): sensitivity to varying loss coefficient $\beta, \omega$ verified in $\mathrm{Ar} \rightarrow \mathrm{Pr}$ and $\mathrm{Cl} \rightarrow \mathrm{Rw}$ tasks; (c): performance comparisons w.r.t. varying openness of the Syn2Real-O task; (d): training cure of the proposed algorithm that characterizes a trade-off between the classification accuracy on known classes and the unknown class; (e)-(f): convergence analysis of the Office-Home task compared to other baselines with confidence intervals. + +samples in the entire target data. We show the results of PGL and PUJE in Fig. 3c. Note that the PGL approach heuristically sets the hyper-parameter according to the true unknown ratio to control the openness, while PUJE automatically estimates the weight $\alpha$ during the training procedure. We observe that PUJE consistently outperforms the baseline by a large margin, especially on unknown recognition, which confirms its robustness to the change in openness. + +Training Curve In Fig. 3d, we illustrate the recognition performance of PUJE over training steps in the $\mathrm{Ar} \rightarrow \mathrm{Cl}$ task. All metrics show a performance gain over the first several steps. Then, the normalized accuracy $\mathrm{OS}^{\star}$ experiences a downward while the accuracy for the unknown class keeps improving, which characterizes a trade-off between $\mathrm{OS}^{\star}$ and UNK. In addition, the behavior of $f_{S}^{\prime}, f_{T}^{\prime}$ meets the expectation of our alignment mechanism where $f_{S}^{\prime}$ starts to detect unknown data, which is followed by $f_{T}^{\prime}$ (Fig. 2b). From Figs. 3e and 3f, we observe that some previous works [15,22] do not converge at the reported optimum, and they tackle this problem by ceasing the model updates before the score starts to decline. However, this early stop strategy can fail without enough labeled target data to validate the current performance. In contrast, our method can reach a reliable convergence. + +Sensitivity to PU & SSL Coefficient We show the sensitivity of our approach to various PU and SSL loss coefficient $\beta, \omega$ . We tested the value of $\beta$ from [0.01, 0.5] + +![](images/9ff734422a3e51cc04995add15f161b0c758ecc1876ef43b88b339e080dfeb75.jpg) + +![](images/1dc159318049829926bfd8c878ddb1b3c91617cf4b36aff3e011932ecac04b31.jpg) + +![](images/4e1db5e706e0f59f9854bce07b19bda165d78dfccb2ed755691039c74c9939e7.jpg) + +![](images/e769a3fb4f133ecd3154fde9819ce3329846d8a94ee3522faa23d8ade5a02d41.jpg) + +![](images/42b22b296ec64a08cf44f844076cc89671fc3e60cdc868bf1447f3765c44ef86.jpg) +(a) OSBP +(e) OSBP +Fig. 4: Feature space visualized by t-SNE in (a)-(d): Office-Home; (e)-(h): Syn2Real-O. + +![](images/48ab1a9f47aa3e015200d4c8e50f058504a6cd167c7369bdcfec0c79066ec9bf.jpg) +(b) PGL +(f) PGL + +![](images/2763721c66614cb069558809ebe9e454e7ef87ea3fe187ab5e56815d0a9e14ac.jpg) +(c) ANNA +(g) ANNA + +![](images/2a2f759faff32a31f774512916fb1891f8a590265e81e170c81277d7181e42c4.jpg) +(d) PUJE +(h) PUJE + +in $\mathrm{Ar}\rightarrow \mathrm{Pr}$ task and $\omega$ from [5, 25] in $\mathrm{Cl}\rightarrow \mathrm{Rw}$ task while fixing other parameters to the default setting. We can draw two observations from Figs. 3a and 3b: the OS score is relatively stable, and the unknown recognition achieves a reliable performance when coefficient $\beta$ is within the interval of [0.02, 0.05]; Generally, a larger $\omega$ , which is set for lower openness, will lead to a better performance on recognizing unknown data while hurt the accuracy for known classes. + +Feature Space Visualization To intuitively visualize the effectiveness of OUDA approaches, we extracted features from the baseline models (OSBP, PGL, ANNA) and our proposed model on the $\mathrm{Ar}\rightarrow \mathrm{Cl}$ task (Office-Home) and Syn2Real-O task with the ResNet-50 backbone [13]. The feature distributions were processed with t-SNE [23] afterward. As shown in Fig. 4, compared with baselines, our method PUJE achieves a better alignment between source and target distributions, especially when the domain shift is large. Benefiting from our joint error-based adversarial alignment mechanism, the extracted feature space, including the cluster of unknown target data, has a more discriminative class-wise decision boundary. + +# 6 Conclusion + +We have addressed the open-set domain shift problem by introducing a novel learning theory based on multi-class positive-unlabeled learning that can reduce the open-set risk and the joint error. Experiments show that our proposed learning theory performs consistently better on challenging object recognition benchmarks for open-set adaptation with significant domain gaps and label shifts. + +# Acknowledgements + +This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo. + +# References + +1. Baktashmotlagh, M., Faraki, M., Drummond, T., Salzmann, M.: Learning factorized representations for open-set domain adaptation. In: ICLR (2019) +2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.: A theory of learning from different domains. Machine Learning 79, 151-175 (2010) +3. Bucci, S., Loghmani, M.R., Tommasi, T.: On the effectiveness of image rotation for open set domain adaptation. In: ECCV (2020) +4. Busto, P.P., Gall, J.: Open set domain adaptation. In: ICCV (2017) +5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: CVPRW (2020) +6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009) +7. Fang, Z., Lu, J., Liu, F., Xuan, J., Zhang, G.: Open set domain adaptation: Theoretical bound and algorithm. arXiv abs/1907.08375 (2019) +8. Feng, Q., Kang, G., Fan, H., Yang, Y.: Attract or distract: Exploit the margin of open set. In: ICCV (2019) +9. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of Machine Learning Research 17(1), 2096-2030 (2016) +0. Garg, S., Balakrishnan, S., Lipton, Z.C.: Domain adaptation under open set label shift. In: NeurIPS (2022) +1. Gomes, R., Krause, A., Perona, P.: Discriminative clustering by regularized information maximization. In: NIPS (2010) +2. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NIPS (2005) +3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. IEEE Conference on Computer Vision and Pattern Recognition pp. 770-778 (2015) +4. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017) +5. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR (2023) +6. Lin, T.Y., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014) +7. Liu, H., Cao, Z., Long, M., Wang, J., Yang, Q.: Separate to adapt: Open set domain adaptation via progressive separation. In: CVPR (2019) +8. Loghmania, M.R., Vinczea, M., Tommasi, T.: Positive-unlabeled learning for open set domain adaptation. Pattern Recognition Letters 136, 198-204 (2020) +9. Long, M., Cao, Y., Wang, J., Jordan, M.I.: Learning transferable features with deep adaptation networks. In: Proceedings of the 32nd International Conference on International Conference on Machine Learning. vol. 37, pp. 97-105. JMLR.org (2015) + +20. Long, M., CAO, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. In: Advances in Neural Information Processing Systems, pp. 1640-1650. Curran Associates, Inc. (2018) +21. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 2208-2217. JMLR.org (2017) +22. Luo, Y., Wang, Z., Huang, Z., Baktashmotlagh, M.: Progressive graph learning for open-set domain adaptation. In: ICML (2020) +23. van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008) +24. Oliver, A., Odena, A., Raffel, C., Cubuk, E.D., Goodfellow, I.J.: Realistic evaluation of deep semi-supervised learning algorithms. In: NeurIPS (2018) +25. Peng, X., Usman, B., Kaushik, N., Hoffman, J., Wang, D., Saenko, K.: Visda: The visual domain adaptation challenge. arXiv abs/1710.06924 (2017) +26. Peng, X., Usmana, B., Saito, K., Kaushik, N., Hoffman, J., Saenko, K.: Syn2real: A new benchmark for synthetic-to-real visual domain adaptation. arXiv abs/1806.09755 (2018) +27. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3723-3732 (2017) +28. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: ECCV (2018) +29. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NIPS (2016) +30. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NIPS (2020) +31. Tang, H., Chen, K.C., Jia, K.: Unsupervised domain adaptation via structurally regularized deep clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (2020) +32. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 2962-2971 (2017) +33. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 5385-5394 (2017) +34. Wang, Q., Meng, F., Breckon, T.P.: Progressively select and reject pseudo-labelled samples for open-set domain adaptation. In: AAAI (2022) +35. Wu, J., He, J.: Domain adaptation with dynamic open-set targets. In: KDD (2022) +36. Xu, Y., Xu, C., Xu, C., Tao, D.: Multi-positive and unlabeled learning. In: IJCAI (2017) +37. Zhang, D., Westfechtel, T., Harada, T.: Unsupervised domain adaptation via minimized joint error. In: TMLR (2023) +38. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Proceedings of Machine Learning Research. vol. 97, pp. 7523-7532. PMLR (2019) \ No newline at end of file diff --git a/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/images.zip b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..77627ff4ddebd5df49f0de9dad2295f0d8957f6b --- /dev/null +++ b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6c7a3777c4d9ff7323f52a5c2d15343a0d98c7e213ca80a0158a965e2c1bb3a +size 659183 diff --git a/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/layout.json b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..921c7b3fd23899eb07231c7bf1a64f3b91463ec0 --- /dev/null +++ b/2024/Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning/layout.json @@ -0,0 +1,13637 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 111, + 482, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 111, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 111, + 482, + 148 + ], + "type": "text", + "content": "Open-set Domain Adaptation via Joint Error based Multi-class Positive and Unlabeled Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "spans": [ + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "type": "text", + "content": "Dexuan Zhang" + }, + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "type": "text", + "content": ", Thomas Westfechtel" + }, + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "type": "text", + "content": ", and Tatsuya Harada" + }, + { + "bbox": [ + 168, + 168, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 251, + 190, + 362, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 190, + 362, + 212 + ], + "spans": [ + { + "bbox": [ + 251, + 190, + 362, + 212 + ], + "type": "text", + "content": "1 The University of Tokyo \n2 RIKEN" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 246, + 455, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 246, + 455, + 389 + ], + "spans": [ + { + "bbox": [ + 159, + 246, + 455, + 389 + ], + "type": "text", + "content": "Abstract. Open-set domain adaptation aims to improve the generalization performance of a learning algorithm on a more realistic problem of open-set domain shift where the target data contains an additional unknown class that is not present in the source data. Most existing algorithms include two phases that can be described as closed-set domain adaptation given heuristic unknown class separation. Therefore, the generalization error cannot be strictly bounded due to the gap between the true distribution and samples inferred from heuristics. In this paper, we propose an end-to-end algorithm that tightly bound the risk of the entire target task by positive-unlabeled (PU) learning theory and the joint error from domain adaptation. Extensive experiments on various data sets demonstrate the effectiveness and efficiency of our proposed algorithm over open-set domain adaptation baselines." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 411, + 230, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 230, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 230, + 423 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "Given enough annotated training data, deep learning models can significantly improve the performance across a wide variety of machine-learning tasks but usually cannot generalize well to new domains, as it is commonly assumed that the training and test data are drawn from the same distribution. In practice, however, this assumption can be violated by several factors, such as the change in light, noise, the angle at which the image is captured, and different types of sensors, which is referred to as the domain shift that can harm the performance when predicting the test data. As a solution, domain adaptation (DA) aims to transfer the knowledge learned from a source domain, which is typically fully labeled, into a different (although related) target domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "As a more realistic setting [28], open-set domain adaptation allows the target data to contain an additional unknown category, covering all irrelevant classes not present in the source domain. The main idea of open-set unsupervised domain adaptation (OUDA) approaches [1, 4, 8, 17, 28] is to jointly learn a classifier from a hypothesis space for known and unknown classes in the source and target domains. According to [2], the target error can be bounded by the source error, the discrepancy distance between domains, the joint error coming from the conditional shift [38], and the open-set risk that indicates the precision of unknown class recognition. Open-set risk contributes substantially to the error" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "content": "bound when a large amount of target data is from the unknown class. However, the latter two terms cannot be explicitly computed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 262 + ], + "type": "text", + "content": "Unsupervised domain adaptation (UDA) methods [9, 19-21, 27, 32] show effectiveness in adapting unlabeled data to new domains by distribution alignment, but can fail to learn discriminative class boundaries, especially when the domain shift is large. Most existing methods tend to ignore the joint error and only focus on minimizing the discrepancy between domains, where samples from different classes can be grouped if the domain shift is large enough. In that case, the joint error becomes non-negligible, and the target error cannot be strictly bounded [9, 38]. [37] provided a solution to this problem by incorporating the joint error into the target error upper bound in the unsupervised setting, and we generalize this idea towards the open-set domain adaptation setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 373 + ], + "type": "text", + "content": "Despite the promising performance, existing OUDA approaches [1, 3, 4, 8, 15, 17, 28, 34] lack an essential theoretical analysis of the generalization error for the target risk. Some works claim to derive a rigorous target error bound, while the minimization of joint error and the open-set risk are not theoretically guaranteed [7, 22]. Moreover, most existing algorithms can be described as closed-set UDA after separation of the unknown [3, 22, 34, 35]. In that case, the open-set risk cannot be controlled since the generalization error is also affected by the gap between empirical known target distribution and inferred known target samples, which becomes hard to analyze." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 532 + ], + "type": "text", + "content": "Fig. 1a illustrates the potential problems of existing methods. Generally, it is impossible to perfectly distinguish the unknown such that the following marginal distribution alignment for UDA part is conducted under a large label shit. In that case, the joint error would be crucial as there exists a trade-off between marginal discrepancy and joint error [38]. Even if somehow we manage to obtain a perfect separation, some target data class will be dragged towards the outside of the corresponding decision boundary. In addition, the unknown class cannot be grouped as a single cluster. Fig. 1b intuitively explains the core concept of our algorithm PUJE Sec. 3, where we introduce two additional task classifiers and minimize the predictive discrepancy between the classifiers with same color, such that the unknown class will be pushed away from source data and grouped as a single cluster. Furthermore, those target data outside the corresponding decision boundary will be aligned back to the correct places." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "OUDA problems can be considered as multi-class positive-unlabeled learning problems (MPU) [36] with domain shift, where all the shared classes appearing in both source and target domains are positive and the unknown class is negative. Therefore, to minimize the aforementioned joint error and open-set risk to achieve a tighter error bound for open-set adaptation, instead of unknown separation and closed-set UDA, we combine PU learning theory and joint error-based target error bound into an end-to-end learning framework. The following methods also introduced PU learning to tackle OUDA problems. [18] required the unsupervised source-like reconstruction for target data via encoder-decoder models whose performance cannot be guaranteed especially for large domain shift and complex dataset; [10] only dealt with the open-set label shift and assumed an identical" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 455, + 190 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 455, + 190 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 455, + 190 + ], + "type": "image", + "image_path": "17b323858c0514ade4d8d4ed8fcfad8dfee3f213db051ac5107f6b8b35596180.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 198, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 243 + ], + "type": "text", + "content": "Fig. 1: Intuitive explanation of the difference between our algorithm PUJE and existing methods. (a) existing methods do not explicitly minimize joint error and cannot group unknown class as a single cluster; (b) our proposal is an upper bound of joint error which can address large domain shift and group unknown class into a single cluster." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 265, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 479, + 324 + ], + "type": "text", + "content": "class conditional distribution of each instance across domains; [35] tackled a slightly different task where the target data is changing over time and it did not cover joint error and the gap between the empirical and inferred known target distribution. Compared with previous works, the main contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 332, + 481, + 403 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 141, + 332, + 481, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 332, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 481, + 355 + ], + "type": "text", + "content": "- We propose an end-to-end algorithm with theoretical analysis for open-set problem setting by PU learning-induced joint error based target error bound;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 355, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 355, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 481, + 390 + ], + "type": "text", + "content": "- We define a novel discrepancy measurement, namely open-set margin discrepancy (OMD), which allows us to smoothly apply the strategy in PU learning to the target error bound in domain adaptation;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 390, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 390, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 390, + 481, + 403 + ], + "type": "text", + "content": "- Extensive experiments confirm the effectiveness and efficiency of the proposal." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 420, + 233, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 233, + 432 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 233, + 432 + ], + "type": "text", + "content": "2 Preliminaries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": "In this section, we introduce notations, problem settings, and theoretical definitions for the tasks of closed-set unsupervised domain adaptation, open-set unsupervised domain adaptation, and multi-class positive and unlabeled learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": "Definition 1 (Closed-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of " + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": " labeled points " + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "inline_equation", + "content": "\\hat{S} = \\{(x_s^i,y_s^i)\\in (\\mathcal{X}\\subseteq \\mathbb{R}^D\\times \\mathcal{Y} = \\{1,\\dots,K\\})\\}_{i = 1}^n" + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": " from the source domain " + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": ", and a set of " + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": " unlabeled points " + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "inline_equation", + "content": "\\hat{T} = \\{(x_t^i)\\in \\mathcal{X}\\}_{i = 1}^m" + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": " from a different target domain " + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": ". With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier " + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "inline_equation", + "content": "f:\\mathcal{X}\\to \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 487, + 482, + 560 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": "Definition 2 (Multi-class Positive and Unlabeled Learning). For a classification task, the learning algorithm has access to a set of " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": " labeled points " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{P} = \\{(x_p^i, y_p^i) \\in (\\mathcal{X} \\times \\mathcal{Y}' = \\{1, ..., K - 1\\})\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": " from the positive domain " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": ", and a set of " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": " unlabeled points " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{U} = \\{(x_u^i) \\in \\mathcal{X}\\}_{i=1}^m" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": " from the unlabeled domain " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "y_u^i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": ". With training samples drawn i.i.d from positive and unlabeled domains, the goal is to learn an optimal classifier " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": " for the unlabeled domain. Here, " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": " includes samples from unknown class " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": ", but the class conditional distributions for known classes remain invariant between " + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "inline_equation", + "content": "P, U" + }, + { + "bbox": [ + 130, + 567, + 482, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-set DA via Joint Error based PU Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": "Definition 3 (Open-set Unsupervised Domain Adaptation). For a classification task, the learning algorithm has access to a set of " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " labeled points " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\hat{S}' = \\{(x_{s'}^i, y_{s'}^i) \\in (\\mathcal{X} \\times \\mathcal{Y}')\\}_{i=1}^n" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " from the incomplete source domain " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ", and a set of " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " unlabeled points " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "\\hat{T} = \\{(x_t^i) \\in \\mathcal{X}\\}_{i=1}^m" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": " from a different target domain " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "y_t^i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": ". With training samples drawn i.i.d from both domains, the goal is to learn an optimal target classifier " + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 132, + 116, + 482, + 190 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 197, + 481, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 481, + 233 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 481, + 233 + ], + "type": "text", + "content": "We show the relation between open-set domain adaptation and multi-class positive and unlabeled learning as follows, which includes the hits on how to apply the domain adaptation algorithm in the open-set scenario." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 239, + 479, + 403 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 141, + 239, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 239, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 239, + 479, + 285 + ], + "type": "text", + "content": "- In Definition 1, there is a domain shift between the distributions of the samples from the source and target domains such that a classifier trained on " + }, + { + "bbox": [ + 141, + 239, + 479, + 285 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 141, + 239, + 479, + 285 + ], + "type": "text", + "content": " cannot generalize on " + }, + { + "bbox": [ + 141, + 239, + 479, + 285 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 141, + 239, + 479, + 285 + ], + "type": "text", + "content": ". A typical solution is to align the feature distributions of both domains by a feature extractor [9,32]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 287, + 479, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 287, + 479, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 287, + 479, + 343 + ], + "type": "text", + "content": "- In Definition 2, the conditional distributions of the samples from known class are identical for positive and unlabeled domains. As proved in [36], the expected error on the unlabeled domain can be approximated by the expected error on the positive domain and the probability that the unlabeled sample has not been classified as unknown." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "text", + "content": "- In Definition 3, the incomplete source domain " + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "text", + "content": " can be regarded as the positive domain " + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "text", + "content": " in MPU. The target domain " + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "text", + "content": " can be treated as " + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 141, + 346, + 479, + 403 + ], + "type": "text", + "content": " with a domain shift. Assume that we can bridge the feature distributions with UDA techniques, it becomes possible to address open-set problems with well-established positive and unlabeled (PU) learning theories." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 423, + 481, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 481, + 438 + ], + "type": "text", + "content": "3 PU Learning induced Joint Error based OUDA (PUJE)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 481, + 532 + ], + "type": "text", + "content": "In this section, inspired by [37], we first rigorously reformulate the joint error-based target upper bound to facilitate the further derivation of a practical objective function for OUDA. To estimate the expectation over the complete source domain " + }, + { + "bbox": [ + 130, + 448, + 481, + 532 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 448, + 481, + 532 + ], + "type": "text", + "content": ", we deploy the theory from multi-class positive and unlabeled learning to address the open-set problems where the source domain " + }, + { + "bbox": [ + 130, + 448, + 481, + 532 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 130, + 448, + 481, + 532 + ], + "type": "text", + "content": " is incomplete. We lastly show the entire loss of PUJE by combining the aforementioned joint error and PU learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": "Theorem 1 (Approximated Joint Error based Target Upper Bound" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": "). Given the output space " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{K} = \\{k|k\\in \\mathbb{R}^K:\\sum_{y\\in \\mathcal{Y}}k[y] = 1,k[y]\\in [0,1]\\}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "f_{S},f_{T}:\\mathcal{X}\\to \\mathcal{K}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " be the true labeling functions for the source and target domains respectively, whose outputs are one-hot vectors denoting the corresponding labels of inputs. Let " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\epsilon :\\mathcal{K}\\times \\mathcal{K}\\rightarrow \\mathbb{R}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " denote a distance metric and " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\epsilon_D(f,f')\\coloneqq \\mathbb{E}_{x\\sim D}\\epsilon (f(x),f'(x))" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " measure the expected disagreement between the outputs of " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "f,f':\\mathcal{X}\\to \\mathcal{K}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " over a distribution " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\forall f_S^*\\in \\mathcal{H}_S\\subseteq \\mathcal{H},\\forall f_T^*\\in \\mathcal{H}_T\\subseteq \\mathcal{H},\\forall h\\in \\mathcal{H}:\\mathcal{X}\\to \\mathcal{K}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "h(x)[y]" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " indicates the probability of " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{X}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": " labeled as " + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{V}" + }, + { + "bbox": [ + 132, + 539, + 482, + 647 + ], + "type": "text", + "content": ", the expected target error is bounded by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 304, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 304, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 304, + 665 + ], + "type": "text", + "content": "3 see proofs in the supplemental material" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 142, + 136, + 482, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 136, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 142, + 136, + 482, + 149 + ], + "type": "interline_equation", + "content": "\\epsilon_ {T} (h) \\leq \\epsilon_ {S} (h) + \\epsilon_ {T} \\left(f _ {S} ^ {*}, f _ {T} ^ {*}\\right) + \\epsilon_ {T} \\left(h, f _ {S} ^ {*}\\right) - \\left| \\epsilon_ {S} \\left(f _ {S} ^ {*}, f _ {T} ^ {*}\\right) - \\epsilon_ {S} \\left(h, f _ {T} ^ {*}\\right) \\right| + \\theta \\tag {1}", + "image_path": "bd8c3d2b2491eb3c2c3050bd3b35db10a8aa0d120baad672b820bb7bed03ec00.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 151, + 481, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 151, + 481, + 165 + ], + "spans": [ + { + "bbox": [ + 161, + 151, + 481, + 165 + ], + "type": "interline_equation", + "content": "\\theta = 2 \\epsilon_ {T} \\left(f _ {S}, f _ {S} ^ {*}\\right) + \\epsilon_ {S} \\left(f _ {S}, f _ {S} ^ {*}\\right) + 2 \\epsilon_ {S} \\left(f _ {T} ^ {*}, f _ {T}\\right) + \\epsilon_ {T} \\left(f _ {T} ^ {*}, f _ {T}\\right) = \\theta_ {f _ {S}} + \\theta_ {f _ {T}} \\tag {2}", + "image_path": "6dcd10870621ed043890af463826e2a88e94702510927a2f5045a7ea8800bbbf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "text", + "content": "According to Theorem 1, we show that the expected target error " + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\epsilon_T(h) \\coloneqq \\epsilon_T(h, f_T)" + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "text", + "content": " is bounded by the expected source error " + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\epsilon_S(h) \\coloneqq \\epsilon_S(h, f_S)" + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "text", + "content": ", the discrepancy between domains " + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "inline_equation", + "content": "[\\epsilon_T(f_S^*, f_T^*) + \\epsilon_T(h, f_S^*) - |\\epsilon_S(f_S^*, f_T^*) - \\epsilon_S(h, f_T^*)|]" + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "text", + "content": ", and the deviation from true labeling functions " + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "text", + "content": ". The following Assumption 1 allows us to derive the generalization error of Theorem 1 such that we can ignore the residual term " + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 171, + 482, + 255 + ], + "type": "text", + "content": " and relate target error bound with the empirical estimation of source error and the discrepancy between domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "text", + "content": "Assumption 1. Let " + }, + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\hat{S},\\hat{T}" + }, + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "text", + "content": " denote empirical samples with finite size from source and target domains. Assume that there exists approximated labeling functions " + }, + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "inline_equation", + "content": "f_{S}^{*},f_{T}^{*}" + }, + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "text", + "content": " such that the empirical deviation " + }, + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{f_S},\\hat{\\theta}_{f_T}" + }, + { + "bbox": [ + 131, + 261, + 482, + 300 + ], + "type": "text", + "content": " are close enough to zero." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "text", + "content": "Definition 4 (Empirical Rademacher Complexity). Let " + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "text", + "content": " be a class of real-valued functions: " + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\to \\mathbb{R}" + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "inline_equation", + "content": "\\hat{D} = \\{x_1, \\ldots, x_m\\}" + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "text", + "content": " a finite sample drawn i.i.d. from a distribution " + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "text", + "content": ", the empirical Rademacher Complexity of " + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 131, + 304, + 483, + 342 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 350, + 482, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 350, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 233, + 350, + 482, + 380 + ], + "type": "interline_equation", + "content": "\\hat {\\mathfrak {R}} _ {\\hat {D}} (\\mathcal {F}) = \\frac {1}{m} \\mathbb {E} _ {\\sigma} [ \\sup _ {f \\in \\mathcal {F}} \\sum_ {i = 1} ^ {m} \\sigma_ {i} f (x _ {i}) ], \\tag {3}", + "image_path": "83b336454bcc7f5f5f4b5585c56f47b6b10ab585974d0a9fd1b6b17f5cf2d834.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 384, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 482, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 384, + 482, + 398 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 131, + 384, + 482, + 398 + ], + "type": "text", + "content": " is an independent uniform random variable taking values in " + }, + { + "bbox": [ + 131, + 384, + 482, + 398 + ], + "type": "inline_equation", + "content": "\\{-1, + 1\\}" + }, + { + "bbox": [ + 131, + 384, + 482, + 398 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "content": "Theorem 2 (Generalization Error" + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "content": "). Given Theorem 1 and Definition 4, Assumption 1 and function space " + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{f,f'}^{\\epsilon}:\\mathcal{X}\\to \\epsilon (f(X),f'(X))" + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "content": " bounded by " + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "inline_equation", + "content": "M > 0" + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "content": ", for any " + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "inline_equation", + "content": "\\delta >0" + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "content": ", with probability at least " + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "inline_equation", + "content": "1 - 2\\delta" + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "inline_equation", + "content": "\\forall h\\in \\mathcal{H}" + }, + { + "bbox": [ + 131, + 403, + 483, + 440 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 457, + 482, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 482, + 526 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {T} (h) \\leq \\epsilon_ {\\hat {S}} (h) + \\epsilon_ {\\hat {T}} (f _ {S} ^ {*}, f _ {T} ^ {*}) + \\epsilon_ {\\hat {T}} (h, f _ {S} ^ {*}) - | \\epsilon_ {\\hat {S}} (f _ {S} ^ {*}, f _ {T} ^ {*}) - \\epsilon_ {\\hat {S}} (h, f _ {T} ^ {*}) | + 2 [ \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {h, f _ {S}} ^ {\\epsilon}) ] \\\\ + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {h, f _ {S} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {S} ^ {*}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {h, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\\epsilon}) (4) \\\\ + 2 \\hat {\\Re} _ {\\hat {S}} (\\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\\epsilon}) + \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {T}, f _ {T} ^ {*}} ^ {\\epsilon}) + 2 \\hat {\\Re} _ {\\hat {T}} (\\mathcal {F} _ {f _ {S}, f _ {S} ^ {*}} ^ {\\epsilon}) ] + 3 M (5 \\sqrt {\\frac {\\log \\frac {9}{\\delta}}{2 m}} + 6 \\sqrt {\\frac {\\log \\frac {9}{\\delta}}{2 n}}) \\\\ \\end{array}", + "image_path": "8c4cb0e401a58ff28639b6a0aa95baa5b305bfa958e66247fb7dee7f9b667e59.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "text", + "content": "In the following part, we introduce the techniques from PU learning to estimate the expectation over source domain " + }, + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "text", + "content": " by the incomplete source domain " + }, + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "text", + "content": " and target domain " + }, + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 539, + 482, + 574 + ], + "type": "text", + "content": " for the open-set scenario." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": "Definition 5 (Unknown Predictive Discrepancy). Let " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "v: \\mathcal{K} \\times \\mathcal{K} \\to \\mathbb{R}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " denote the Unknown Predictive Discrepancy as a distance metric and " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{D}(f, f') := \\mathbb{E}_{x \\sim D} v(f(x), f'(x))" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " measure the expected disagreement between the " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": "-th outputs of " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "f, f': \\mathcal{X} \\to \\mathcal{K}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " over a distribution " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "e^{K}: \\mathcal{X} \\to [0, \\dots, 1] \\in \\mathcal{K}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " denote a function that can predict any input as unknown. The deviation from " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "e^{K}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " for a hypothesis " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "h \\in \\mathcal{H}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " is referred to as the shorthand " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{D}(h) := v_{D}(h, e^{K})" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " that measures the probability that samples from " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " have not been classified as unknown." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 448, + 103 + ], + "type": "text", + "content": "Open-set DA via Joint Error based PU Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Assumption 2. Let " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "S^i = P_S(x|y = i), T^i = P_T(x|y = i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " denote class conditional distributions, " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "S' = P_S(x|y \\neq K), T' = P_T(x|y \\neq K)" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " indicate incomplete domains that do not contain unknown class " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "S^K, T^K" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". Given a feature extractor " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "g: \\mathcal{X} \\subseteq \\mathbb{R}^D \\to \\mathcal{Z} \\subseteq \\mathbb{R}^F" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", assume that the feature space can be aligned by UDA techniques such that " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "Z^K = P_{S^K}(z) = P_{T^K}(z), Z' = P_{S'}(z) = P_{T'}(z)" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "text", + "content": "Proposition 1. Let " + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "inline_equation", + "content": "h \\in \\mathcal{H}_{F_g} : \\mathcal{Z} \\to \\mathcal{K}" + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "text", + "content": " denote the decomposed hypothesis where " + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "inline_equation", + "content": "h \\circ g \\in \\mathcal{H}" + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "text", + "content": ". Given Definition 5, Assumption 2, " + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "inline_equation", + "content": "v_{S^K}(h \\circ g) = v_{T^K}(h \\circ g)" + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "text", + "content": " that represents probability that samples from " + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "inline_equation", + "content": "S^K" + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "text", + "content": " have not been classified as unknown class, is equivalent to " + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon_{S^K}(h \\circ g)" + }, + { + "bbox": [ + 130, + 186, + 482, + 247 + ], + "type": "text", + "content": " that measures the classification error on unknown class according to the definition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": "Lemma 1 (Estimated Source Error" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": "). Let " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{K} \\pi_S^i = 1" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{K} \\pi_T^i = 1" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": " denote the label distribution of " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": " respectively. Given Proposition 1, the expected error on " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": " can be estimated by the error on " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": " and Unknown Predictive Discrepancy (Definition 5) on " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": " with a mild condition that " + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\pi_S^K = \\pi_T^K = 1 - \\alpha" + }, + { + "bbox": [ + 130, + 256, + 482, + 307 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 328, + 482, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 328, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 200, + 328, + 482, + 342 + ], + "type": "interline_equation", + "content": "\\epsilon_ {S} (h \\circ g) = \\alpha [ \\epsilon_ {S ^ {\\prime}} (h \\circ g) - v _ {S ^ {\\prime}} (h \\circ g) ] + v _ {T} (h \\circ g) \\tag {5}", + "image_path": "ecdabdb60be53502a341cd17824c1b00ac788f5c05c452970b10a6cfc3db1dac.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": "Remark 1. According to Definition 5, minimizing " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "v_{T}(h \\circ g)" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": " means classifying target samples as the unknown class. In practice, a multiplier " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\beta < 1" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": " is applied on " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "v_{T}(h \\circ g)" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": " to prevent all target samples from being recognized as unknown." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "content": "To reformulate the intractable discrepancy " + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon_S(f_S^*,f_T^*)" + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon_S(h,f_T^*)" + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "content": " in Theorem 1, we decompose the approximated labeling functions given feature extractor " + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "inline_equation", + "content": "f_{S}^{\\star}\\circ g = f_{S}^{\\star},f_{T}^{\\star}\\circ g = f_{T}^{\\star}" + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "inline_equation", + "content": "f_{S}^{\\star}\\in \\mathcal{H}_{S_{g}}\\subseteq \\mathcal{H}_{F_{g}},f_{T}^{\\star}\\in \\mathcal{H}_{T_{g}}\\subseteq \\mathcal{H}_{F_{g}}" + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "content": ". We further define " + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "text", + "content": " in a way such that, under some circumstances, it is equivalent to the Unknown Predictive Discrepancy for " + }, + { + "bbox": [ + 130, + 399, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\forall f\\in \\mathcal{H}_{F_g}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 223, + 472, + 482, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 472, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 223, + 472, + 482, + 487 + ], + "type": "interline_equation", + "content": "\\epsilon_ {T ^ {K}} \\left(f \\circ g, f _ {T} ^ {\\star} \\circ g\\right) = v _ {T ^ {k}} \\left(f \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\tag {6}", + "image_path": "0c8bb0206c26af39af242f3a609f65980c2bd82959a854b3830d16905ea8036d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "text", + "content": "Corollary 1. Given Assumption 2, the feature extractor " + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "text", + "content": " can align the feature distributions of source and target domain such that the expectation over " + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "inline_equation", + "content": "S^K" + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "text", + "content": " is equal to that over " + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "inline_equation", + "content": "T^K" + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "text", + "content": ", which leads Eq. (6) to: " + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_{S^K}(f \\circ g, f_T^\\star \\circ g) = \\epsilon_{T^K}(f \\circ g, f_T^\\star \\circ g) = v_{T^K}(f \\circ g, f_T^\\star \\circ g) = v_{S^K}(f \\circ g, f_T^\\star \\circ g)" + }, + { + "bbox": [ + 130, + 495, + 482, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": "Lemma 2 (Estimated Discrepancy" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": "). Given Corollary 1, analogous to Lemma 1, the discrepancy measured on source domain " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon_S(f_S^*, f_T^*)" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon_S(h, f_T^*)" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " can be estimated by the discrepancy on " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " and Unknown Predictive Discrepancy (Definition 5) with a mild condition that " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\pi_S^K = \\pi_T^K = 1 - \\alpha" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 623, + 499, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 499, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 499, + 666 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\epsilon_ {S} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) & = \\alpha \\left[ \\epsilon_ {S ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {S ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] + v _ {T} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\\\ \\epsilon_ {S} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) & = \\alpha \\left[ \\epsilon_ {S ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {S ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] + v _ {T} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\end{array} \\right. \\tag {7}", + "image_path": "8183b19435151e9975d84ac30fed7e1bd32fd043e4d45d98b07834c12eb475fd.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "content": "Definition 6 (Overall Loss of PUJE). Let " + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "inline_equation", + "content": "\\hat{S}',\\hat{T}" + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "content": " denote empirical data from " + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "inline_equation", + "content": "S',T" + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "content": ". Given a feature extractor " + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "inline_equation", + "content": "g:\\mathcal{X}\\subseteq R^{D}\\to \\mathcal{Z}\\subseteq \\mathbb{R}^{F}" + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "content": ", decomposed hypothesis and approximated labeling functions " + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "inline_equation", + "content": "h,f_{S}^{\\star},f_{T}^{\\star}\\in \\mathcal{H}_{F_{g}}:\\mathcal{Z}\\to \\mathcal{K}" + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "content": ", replacing the intractable terms in Theorem 2 with Eqs. (5) and (7) and ignoring the Rademacher Complexity terms, the overall loss of PUJE can be regrouped as the sum of classification loss " + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "content": " and the discrepancy between domains " + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 131, + 115, + 483, + 189 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 206, + 503, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 503, + 265 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 503, + 265 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {p u j e} = \\alpha \\left[ \\epsilon_ {\\hat {S} ^ {\\prime}} (h \\circ g) - v _ {\\hat {S} ^ {\\prime}} (h \\circ g) \\right] + \\beta v _ {\\hat {T}} (h \\circ g) + \\epsilon_ {\\hat {T}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) + \\epsilon_ {\\hat {T}} (h \\circ g, f _ {S} ^ {\\star} \\circ g) \\\\ - \\left| v _ {\\hat {T}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {\\hat {T}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) + \\alpha \\left[ \\epsilon_ {\\hat {S} ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right. \\right. \\\\ \\left. + v _ {\\hat {S} ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - \\epsilon_ {\\hat {S} ^ {\\prime}} \\left(h \\circ g, f _ {T} ^ {\\star} \\circ g\\right) - v _ {\\hat {S} ^ {\\prime}} \\left(f _ {S} ^ {\\star} \\circ g, f _ {T} ^ {\\star} \\circ g\\right) \\right] | \\\\ = L _ {c l s} (h; g) + L _ {d i s} \\left(f _ {S} ^ {\\star}, f _ {T} ^ {\\star}, h; g\\right) \\tag {8} \\\\ \\end{array}", + "image_path": "e18ac70609cba8814cadf0f8cd0bb2297e862e83978cb1786c95ed61ada9bb33.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 233, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 233, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 233, + 294 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "text", + "content": "In this section, we first define Approximated Labeling Function Space " + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{S_g},\\mathcal{H}_{T_g}" + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "text", + "content": " such that the intractable " + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "inline_equation", + "content": "L_{dis}(f_S^\\star ,f_T^\\star ,h;g)" + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "text", + "content": " due to unknown " + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "inline_equation", + "content": "f_S^\\star ,f_T^\\star" + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "text", + "content": " can be upper bounded by " + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\sup_{f_S'\\in \\mathcal{H}_{S_g},f_T'\\in \\mathcal{H}_{T_g}}L_{dis}(f_S',f_T',h;g)" + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "text", + "content": ". To fulfill Eq. (6), we propose Open-set Margin Discrepancy to quantify " + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 303, + 482, + 365 + ], + "type": "text", + "content": " that measures the disagreement between classifiers. We lastly show the entire training algorithm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 380, + 360, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 360, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 360, + 392 + ], + "type": "text", + "content": "4.1 Approximated Labeling Function Space" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": "Proposition 2. Let " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_S, \\mathcal{H}_T : \\mathcal{X} \\to \\mathcal{K}" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " be two sets of functions that can minimize a part of the empirical residual " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_S, \\hat{\\theta}_T" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "f_S^* \\in \\mathcal{H}_S" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "f_T^* \\in \\mathcal{H}_T" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " must hold as " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "f_S^*, f_T^*" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " can minimize the entire " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " given Assumption 1. Accordingly, decomposed functions " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "f_S^*, f_T^*" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " must lie in " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{S_g} = \\{f|\\forall f \\circ g \\in \\mathcal{H}_S\\}" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{T_g} = \\{f|\\forall f \\circ g \\in \\mathcal{H}_T\\}" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " such that a sufficient condition of the following inequality: " + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "inline_equation", + "content": "L_{dis}(f_S^*, f_T^*, h; g) \\leq \\sup_{f_S' \\in \\mathcal{H}_{S_g}, f_T' \\in \\mathcal{H}_{T_g}} L_{dis}(f_S', f_T', h; g)" + }, + { + "bbox": [ + 130, + 397, + 482, + 474 + ], + "type": "text", + "content": " is fulfilled." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": "Definition 7 (Approximated Labeling Function Space). Let " + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T}" + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": " denote a part of the empirical residual " + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{fs}, \\hat{\\theta}_{f_T}" + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": " respectively. Let " + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "inline_equation", + "content": "f_S^\\star \\circ g = f_S^* \\in \\mathcal{H}_S, f_T^\\star \\circ g = f_T^* \\in \\mathcal{H}_T" + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": " denote the decomposed approximated labeling functions where " + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "inline_equation", + "content": "f_S^\\star \\in \\mathcal{H}_{S_g} \\subseteq \\mathcal{H}_{F_g}, f_T^\\star \\in \\mathcal{H}_{T_g} \\subseteq \\mathcal{H}_{F_g}" + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": ". Given Proposition 2, Approximated Labeling Function Space " + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{S_g}, \\mathcal{H}_{T_g}" + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": " can be defined as the sets whose members " + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "inline_equation", + "content": "f_S', f_T' \\in \\mathcal{H}_{F_g}" + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": " can minimize " + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T}" + }, + { + "bbox": [ + 131, + 479, + 482, + 554 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 567, + 482, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 482, + 602 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathcal {H} _ {S _ {g}} = \\left\\{f _ {S} ^ {\\prime} \\right| \\arg \\min _ {g, f _ {S} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) := L _ {c l s} \\left(f _ {S} ^ {\\prime}; g\\right) \\} \\\\ \\mathcal {H} _ {T _ {g}} = \\left\\{f _ {T} ^ {\\prime} \\right| \\arg \\min _ {g, f _ {T} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) := (1 - \\gamma) L _ {c l s} \\left(f _ {T} ^ {\\prime}; g\\right) + \\gamma L _ {r e g} \\end{array} \\right. \\tag {9}", + "image_path": "65da16ef49076ab343624d8b0f83368c9321657bf13c00942534bb3c033c39de.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{S_g}" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " consists of functions minimizing the expected error on " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": ", which can be estimated by Lemma 1. To build a reliable function space " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{T_g}" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " without target labels, we approximate the target error with the weighted average " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\gamma \\in [0,1])" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " of error rate on labeled samples and a semi-supervised regularization term " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{reg} = L_{ent} + L_{pse} + \\omega L_{con}" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-set DA via Joint Error based PU Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Regularized Entropy Minimization As introduced in [11,27,31], we impose a class balance prior that can penalize classifiers with complex decision boundaries on entropy minimization [12] to yield a more sensible solution:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 196, + 173, + 482, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 173, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 196, + 173, + 482, + 215 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {e n t} = - \\mathbb {E} _ {x \\in \\hat {T}} \\sum_ {y \\in \\mathcal {Y}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\log f _ {T} ^ {\\prime} (g (x)) [ y ] \\\\ + \\sum_ {y \\in \\mathcal {Y}} \\mathbb {E} _ {x \\in \\hat {T}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\log \\mathbb {E} _ {x \\in \\hat {T}} f _ {T} ^ {\\prime} (g (x)) [ y ] \\tag {10} \\\\ \\end{array}", + "image_path": "ccc62c4a5c2f1795e9434c1df14e4b240930d34126a91a88e707a0e55622f551.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": "Pseudo Labeling As introduced in [30, 31], for input " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "x \\in \\hat{T}" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " and its random augmentation " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " [5], we minimize cross entropy for " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " with pseudo labels of " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 264, + 482, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 264, + 482, + 279 + ], + "spans": [ + { + "bbox": [ + 197, + 264, + 482, + 279 + ], + "type": "interline_equation", + "content": "L _ {p s e} = - \\mathbb {E} _ {x \\in \\hat {T}} \\log f _ {T} ^ {\\prime} (g (x)) [ \\arg \\max _ {y \\in \\mathcal {Y}} h (g (x ^ {\\prime})) [ y ] ] \\tag {11}", + "image_path": "a304055007ae35d05a9d0e5a1ea730d9a48ff5b73b5e14655b587a1e9e14bcf7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 289, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 482, + 314 + ], + "type": "text", + "content": "Consistency Regularization As introduced in [14, 29], we penalize the difference of the outputs for input " + }, + { + "bbox": [ + 132, + 289, + 482, + 314 + ], + "type": "inline_equation", + "content": "x \\in \\hat{T}" + }, + { + "bbox": [ + 132, + 289, + 482, + 314 + ], + "type": "text", + "content": " and its random augmentation " + }, + { + "bbox": [ + 132, + 289, + 482, + 314 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 132, + 289, + 482, + 314 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 327, + 482, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 327, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 231, + 327, + 482, + 341 + ], + "type": "interline_equation", + "content": "L _ {c o n} = \\mathbb {E} _ {x \\in \\hat {T}} | f _ {T} ^ {\\prime} (g (x)) - f _ {T} ^ {\\prime} (g (x ^ {\\prime})) | \\tag {12}", + "image_path": "2084bf810d6961d46daba213b06e1a8d54e87798ef7a347c349e929e34a54cef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 357, + 312, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 312, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 312, + 370 + ], + "type": "text", + "content": "4.2 Open-set Margin Discrepancy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": "Definition 8 (Induced Labeling Function). Let " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{K}" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": " denote a multiclass labeling function, where " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "f(x)[y]" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": " indicating " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": "-th element of the output " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": " for the probability of " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": " classified as " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": ". Thus an induced labeling function " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "l \\circ f: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": " is given by: " + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "inline_equation", + "content": "l(f(x)) = \\arg \\max_{y \\in \\mathcal{Y}} f(x)[y]" + }, + { + "bbox": [ + 131, + 379, + 483, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "text", + "content": "Definition 9 (Open-set Margin Discrepancy (OMD)). Let " + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "inline_equation", + "content": "y = l(f(x)), y' = l(f'(x))" + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "text", + "content": " denote the predictions on input " + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "inline_equation", + "content": "f, f': \\mathcal{X} \\to \\mathcal{K}" + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "text", + "content": " given Definition 8. Open-set Margin Discrepancy between " + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "inline_equation", + "content": "f, f'" + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "text", + "content": " over a distribution " + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 437, + 485, + 474 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 171, + 495, + 482, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 495, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 171, + 495, + 482, + 539 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {D} (f, f ^ {\\prime}) = \\mathbb {E} _ {x \\sim D} [ \\operatorname {o m d} (f (x), f ^ {\\prime} (x)) ] (13) \\\\ \\operatorname {o m d} (f (x), f ^ {\\prime} (x)) = \\max \\left(\\left| \\log (1 - f (x) [ y ]) - \\log (1 - f ^ {\\prime} (x) [ y ]) \\right|\\right), \\\\ \\left. \\left| \\log \\left(1 - f (x) \\left[ y ^ {\\prime} \\right]\\right) - \\log \\left(1 - f ^ {\\prime} (x) \\left[ y ^ {\\prime} \\right]\\right) \\right|\\right) (14) \\\\ \\end{array}", + "image_path": "766e5499316d6b711c80c4a9faae6d1a24c2a11f1d704c35b44e06c034809654.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 548, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 481, + 586 + ], + "type": "text", + "content": "For OUDA, generally the classifier is not discriminative for unknown class " + }, + { + "bbox": [ + 131, + 548, + 481, + 586 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 548, + 481, + 586 + ], + "type": "text", + "content": " due to the lack of labeled samples to reduce the misclassification rate. To address this, we model " + }, + { + "bbox": [ + 131, + 548, + 481, + 586 + ], + "type": "inline_equation", + "content": "\\log (1 - f(x)[K])" + }, + { + "bbox": [ + 131, + 548, + 481, + 586 + ], + "type": "text", + "content": " that has a more stable gradient." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "text", + "content": "Remark 2. According to Definition 5, " + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "inline_equation", + "content": "v_{D}(f,f^{\\prime})" + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "text", + "content": " measures the expectation of disagreement on the K-th output between two functions " + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "inline_equation", + "content": "f,f^{\\prime}:\\mathcal{X}\\to \\mathcal{K}" + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "text", + "content": " over a distribution " + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "text", + "content": ", which can also indicate the difference between probabilities of categorizing inputs as known classes. To fulfill Eq. (6), we quantify " + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "inline_equation", + "content": "v_{D}(f,f^{\\prime})" + }, + { + "bbox": [ + 131, + 592, + 482, + 641 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 185, + 653, + 482, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 653, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 185, + 653, + 482, + 667 + ], + "type": "interline_equation", + "content": "v _ {D} (f, f ^ {\\prime}) = \\mathbb {E} _ {x \\sim D} | \\log (1 - f (x) [ K ]) - \\log (1 - f ^ {\\prime} (x) [ K ]) | \\tag {15}", + "image_path": "79f5ce806eff0c1e1734348c77b5630e3a2194e7223447edcb65cc666744c512.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 225, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 225, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 225, + 127 + ], + "type": "text", + "content": "Algorithm 1 PUJE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 133, + 383, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 133, + 383, + 144 + ], + "spans": [ + { + "bbox": [ + 141, + 133, + 383, + 144 + ], + "type": "text", + "content": "Input: incomplete source data " + }, + { + "bbox": [ + 141, + 133, + 383, + 144 + ], + "type": "inline_equation", + "content": "\\hat{S}^{\\prime}" + }, + { + "bbox": [ + 141, + 133, + 383, + 144 + ], + "type": "text", + "content": "; unlabeled target data " + }, + { + "bbox": [ + 141, + 133, + 383, + 144 + ], + "type": "inline_equation", + "content": "\\hat{T}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "spans": [ + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "type": "text", + "content": "Output: labeling functions " + }, + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "type": "inline_equation", + "content": "f_S', f_T'" + }, + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "type": "text", + "content": "; feature extractor " + }, + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "type": "text", + "content": "; hypothesis " + }, + { + "bbox": [ + 141, + 144, + 420, + 156 + ], + "type": "inline_equation", + "content": "h" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "spans": [ + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "text", + "content": "Parameter: trade-off parameter " + }, + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "text", + "content": "; learning rate " + }, + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "text", + "content": "; known class ratio estimator " + }, + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 141, + 156, + 475, + 176 + ], + "type": "text", + "content": " for epoch = 1, 2, ... do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "spans": [ + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "type": "text", + "content": "Step 1: estimate known class ratio " + }, + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "type": "text", + "content": " with prediction of " + }, + { + "bbox": [ + 156, + 177, + 424, + 188 + ], + "type": "inline_equation", + "content": "g, h" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 156, + 188, + 473, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 188, + 473, + 199 + ], + "spans": [ + { + "bbox": [ + 156, + 188, + 473, + 199 + ], + "type": "text", + "content": "Step 2: optimize " + }, + { + "bbox": [ + 156, + 188, + 473, + 199 + ], + "type": "inline_equation", + "content": "g, f_S', f_T'" + }, + { + "bbox": [ + 156, + 188, + 473, + 199 + ], + "type": "text", + "content": " to satisfy the approximated labeling function space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 166, + 199, + 325, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 199, + 325, + 210 + ], + "spans": [ + { + "bbox": [ + 166, + 199, + 325, + 210 + ], + "type": "interline_equation", + "content": "(g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) \\leftarrow (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) + \\eta \\Delta (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime})", + "image_path": "123230d3f20b4c9be4ebbd14f8da6319d6b3749dbeb2314202f091b763e6de1e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 166, + 210, + 326, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 210, + 326, + 228 + ], + "spans": [ + { + "bbox": [ + 166, + 210, + 326, + 228 + ], + "type": "interline_equation", + "content": "\\varDelta (g, f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}) = - \\frac {\\partial (L _ {\\mathcal {H} _ {S}} (f _ {S} ^ {\\prime} ; g) + L _ {\\mathcal {H} _ {T}} (f _ {T} ^ {\\prime} ; g))}{\\partial (g , f _ {S} ^ {\\prime} , f _ {T} ^ {\\prime})}", + "image_path": "69369ba11a23302a5aae708b43d32979dd160525003201cad5ef231cb42287d0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 156, + 228, + 454, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 228, + 454, + 238 + ], + "spans": [ + { + "bbox": [ + 156, + 228, + 454, + 238 + ], + "type": "text", + "content": "Step 3: maximize the discrepancy w.r.t. " + }, + { + "bbox": [ + 156, + 228, + 454, + 238 + ], + "type": "inline_equation", + "content": "f_{S}^{\\prime}, f_{T}^{\\prime}" + }, + { + "bbox": [ + 156, + 228, + 454, + 238 + ], + "type": "text", + "content": " within the function space" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 166, + 238, + 298, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 238, + 298, + 249 + ], + "spans": [ + { + "bbox": [ + 166, + 238, + 298, + 249 + ], + "type": "interline_equation", + "content": "\\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right) \\leftarrow \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right) + \\eta \\Delta \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}\\right)", + "image_path": "33a6f0b3cbb4cdac1416eb057c77cbe527b801e4c8e341623144c02f445eb5a6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 166, + 249, + 381, + 267 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 249, + 381, + 267 + ], + "spans": [ + { + "bbox": [ + 166, + 249, + 381, + 267 + ], + "type": "interline_equation", + "content": "\\varDelta(f_{S}^{\\prime},f_{T}^{\\prime})=-\\frac{\\partial(L_{\\mathcal{H}_{S}}(f_{S}^{\\prime};g)+L_{\\mathcal{H}_{T}}(f_{T}^{\\prime};g)-\\lambda L_{dis}(f_{S}^{\\prime},f_{T}^{\\prime},h;g))}{\\partial(f_{S}^{\\prime},f_{T}^{\\prime})}", + "image_path": "b4024027fbbdac6a1e3735222069f373a09a27f69d35bf410a400cacb02ac011.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 156, + 266, + 454, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 266, + 454, + 277 + ], + "spans": [ + { + "bbox": [ + 156, + 266, + 454, + 277 + ], + "type": "text", + "content": "Step 4: minimize the entire target error bound w.r.t. " + }, + { + "bbox": [ + 156, + 266, + 454, + 277 + ], + "type": "inline_equation", + "content": "g, h" + }, + { + "bbox": [ + 156, + 266, + 454, + 277 + ], + "type": "text", + "content": " for fixed " + }, + { + "bbox": [ + 156, + 266, + 454, + 277 + ], + "type": "inline_equation", + "content": "f_S', f_T'" + }, + { + "bbox": [ + 156, + 266, + 454, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 166, + 277, + 268, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 277, + 268, + 289 + ], + "spans": [ + { + "bbox": [ + 166, + 277, + 268, + 289 + ], + "type": "interline_equation", + "content": "(g, h) \\leftarrow (g, h) + \\eta \\Delta (g, h)", + "image_path": "06bd66b1d8655c22263914effe95cf3f87cb1fb983a047eca23b51e65f2da47f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 166, + 289, + 321, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 289, + 321, + 304 + ], + "spans": [ + { + "bbox": [ + 166, + 289, + 321, + 304 + ], + "type": "interline_equation", + "content": "\\Delta (g, h) = - \\frac {\\partial \\left(L _ {c l s} (h ; g) + \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime} , f _ {T} ^ {\\prime} , h ; g\\right)\\right)}{\\partial (g , h)}", + "image_path": "932dc93ce112116c929b3dd9fa2355a73ac9b3035ddce2e60230e6dc7ee3a59b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 303, + 176, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 303, + 176, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 176, + 311 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 337, + 258, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 258, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 258, + 349 + ], + "type": "text", + "content": "4.3 Training Algorithm" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": "Given the feature extractor " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "g: \\mathcal{X} \\subseteq \\mathbb{R}^D \\to \\mathcal{Z} \\subseteq \\mathbb{R}^F" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " and hypotheses " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "h, f_S', f_T' \\in \\mathcal{H}_{F_g}: \\mathcal{Z} \\to \\mathcal{K}" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ", we introduce a trade-off parameter " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " to balance the classification loss and discrepancy as Eq. (16). During each epoch, we first estimate the known class ratio " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " (initially set to 1) in target data by the predictions of " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ". Then we train the hypothesis " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ", approximated labeling functions " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "f_S', f_T'" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ", feature extractor " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " to minimize the error on labeled data " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " with hypothesis space constraints " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{H}_S}, L_{\\mathcal{H}_T}" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": ", while optimizing the discrepancy " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " adversarially by the min-max game over " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "inline_equation", + "content": "f_S', f_T'" + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": " (Algorithm 1):" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "type": "interline_equation", + "content": "\\left\\{\\min _ {f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime} \\in \\mathcal {H} _ {F _ {g}}} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) + L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) - \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}, h; g\\right) \\right. \\tag {16}", + "image_path": "9cdcfb302781c045a92b4ca8a029ed0b13f261214d2466956f4b8835d42b7d8e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 490, + 448, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 448, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 448, + 502 + ], + "type": "interline_equation", + "content": "\\left\\lfloor \\min _ {h \\in \\mathcal {H} _ {F _ {g}}, g} L _ {\\mathcal {H} _ {S}} \\left(f _ {S} ^ {\\prime}; g\\right) + L _ {\\mathcal {H} _ {T}} \\left(f _ {T} ^ {\\prime}; g\\right) + L _ {c l s} (h; g) + \\lambda L _ {d i s} \\left(f _ {S} ^ {\\prime}, f _ {T} ^ {\\prime}, h; g\\right) \\right.", + "image_path": "ba29b4197b05e91e9710287f4293eee4b7ce3b60c971dd9727c84b66bd199210.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 526, + 205, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 205, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 205, + 537 + ], + "type": "text", + "content": "4.4 Intuition" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this section, we intuitively explain how our method can align the source and target domains while separating the unknown in Fig. 2. According to [24], unlike the entropy minimization that tends to cut through unlabeled data, consistency regularization helps to draw a manifold-aware decision boundary. Based on this conclusion, the decision boundary of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{T}^{\\prime}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " induced by consistency regularization " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " can preserve the cluster structure (Fig. 2a). According to Algorithm 1, " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{S}^{\\prime}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " tries to maximize " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " while classify target samples as unknown class to minimize " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{H}_S}(f_S';g)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " (Definition 7 and Lemma 1), which can lead to a new decision boundary shown in Fig. 2b. Meanwhile, feature extractor " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " tries to minimize " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " by pushing target samples close to " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{S}^{\\prime}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " towards source clusters," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-set DA via Joint Error based PU Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 231, + 204 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 231, + 204 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 231, + 204 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 231, + 204 + ], + "type": "image", + "image_path": "27899cb4ba862fd93fb884d91e7c91c655ce01785873ac7abe2c8609ed4c46ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 206, + 211, + 219 + ], + "lines": [ + { + "bbox": [ + 155, + 206, + 211, + 219 + ], + "spans": [ + { + "bbox": [ + 155, + 206, + 211, + 219 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 155, + 206, + 211, + 219 + ], + "type": "inline_equation", + "content": "\\min_{f_T'} L_{\\mathcal{H}_T}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 113, + 329, + 205 + ], + "blocks": [ + { + "bbox": [ + 233, + 113, + 329, + 205 + ], + "lines": [ + { + "bbox": [ + 233, + 113, + 329, + 205 + ], + "spans": [ + { + "bbox": [ + 233, + 113, + 329, + 205 + ], + "type": "image", + "image_path": "67564d6910b9b0116b35c85a28e1451426b0f91251ff90ed0a78e0bb78b10f4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 206, + 323, + 219 + ], + "lines": [ + { + "bbox": [ + 236, + 206, + 323, + 219 + ], + "spans": [ + { + "bbox": [ + 236, + 206, + 323, + 219 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 236, + 206, + 323, + 219 + ], + "type": "inline_equation", + "content": "\\min_{f_S'}(L_{\\mathcal{H}_S} - L_{dis})" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 115, + 406, + 204 + ], + "blocks": [ + { + "bbox": [ + 329, + 115, + 406, + 204 + ], + "lines": [ + { + "bbox": [ + 329, + 115, + 406, + 204 + ], + "spans": [ + { + "bbox": [ + 329, + 115, + 406, + 204 + ], + "type": "image", + "image_path": "2dd6be1cf92e80d3475bcda7814410c86f3dfef6fd68d894034898d11410acb7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 116, + 436, + 141 + ], + "lines": [ + { + "bbox": [ + 413, + 116, + 436, + 141 + ], + "spans": [ + { + "bbox": [ + 413, + 116, + 436, + 141 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 413, + 116, + 436, + 141 + ], + "type": "inline_equation", + "content": "f_T^{\\prime}" + }, + { + "bbox": [ + 413, + 116, + 436, + 141 + ], + "type": "inline_equation", + "content": "f_S^{\\prime}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 419, + 148, + 478, + 186 + ], + "lines": [ + { + "bbox": [ + 419, + 148, + 478, + 186 + ], + "spans": [ + { + "bbox": [ + 419, + 148, + 478, + 186 + ], + "type": "text", + "content": "- Positive Source\n- Negative Source\n- Unknown Target\n- Pseudo Positive\n- Pseudo Negative" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 380, + 206, + 429, + 217 + ], + "lines": [ + { + "bbox": [ + 380, + 206, + 429, + 217 + ], + "spans": [ + { + "bbox": [ + 380, + 206, + 429, + 217 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 380, + 206, + 429, + 217 + ], + "type": "inline_equation", + "content": "\\min_g L_{dis}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": "Fig. 2: Intuitive explanation of the mechanism of the proposed PUJE in Algorithm 1. (a) " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{H}_T}" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " includes consistency regularization " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " that helps to draw a manifold-aware decision boundary for " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "f_T'" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " (Step 2); (b) " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "f_S'" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " can increase " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " by classifying most of the unlabeled target samples as unknown, which also decreases " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{H}_S}" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " (Step 3); (c) feature extractor " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " pushes unlabeled target samples close to the decision boundary of " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "f_S'" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " towards source cluster, while those close to " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "f_T'" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": " towards outside to reduce " + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 130, + 227, + 482, + 304 + ], + "type": "text", + "content": ", which can lead to a separated unknown cluster (Step 4)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "type": "text", + "content": "while those close to " + }, + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "type": "inline_equation", + "content": "f_{T}^{\\prime}" + }, + { + "bbox": [ + 130, + 328, + 480, + 365 + ], + "type": "text", + "content": " towards outside such that unlabeled target samples far from source clusters can be eventually separated and grouped into a new cluster for the unknown progressively as illustrated in Fig. 2c." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 384, + 218, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 218, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 218, + 396 + ], + "type": "text", + "content": "5 Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": "We evaluated our proposal on two benchmarks, Office-Home and Syn2Real-O. In the implementation, the " + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": " weight " + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": " and the trade-off parameter " + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": " were set to 0.1 and 0.01 according to [37]. In addition, we empirically set PU coefficient " + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": " to 0.03 and consistency coefficient " + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": " to 20 for Office-Home and 5 for Syn2Real-O. All experiments were conducted with the ImageNet [6] pre-trained ResNet-50 [13] as the feature extractor " + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": " and 2-layer linear networks for classifiers " + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "inline_equation", + "content": "f_S', f_T', h" + }, + { + "bbox": [ + 130, + 411, + 482, + 533 + ], + "type": "text", + "content": ". We trained the model by Stochastic Gradient Descent optimizer with a 0.001 learning rate annealed according to [9], 24 batch-size, a momentum of 0.9. We quantitatively compare our results against various OUDA baselines, including OSBP [28], STA [17], DAOD [8], PGL [22], ROS [3], OSLPP [34] and ANNA [15]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "text", + "content": "Evaluation Metrics To evaluate the proposed method and the baselines, we utilize the widely used measures [22, 28], i.e., normalized accuracy for all classes (OS), normalized accuracy for the known classes only " + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "inline_equation", + "content": "(\\mathrm{OS}^{\\star})" + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "text", + "content": " and harmonic mean " + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{HOS} = 2(\\mathrm{OS}^{\\star}\\times \\mathrm{UNK}) / (\\mathrm{OS}^{\\star} + \\mathrm{UNK})" + }, + { + "bbox": [ + 130, + 544, + 482, + 594 + ], + "type": "text", + "content": " [3,15,18,34]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Office-Home [33] is a widely-used domain adaptation benchmark, which consists of 15,500 images from 65 categories and four domains: Art (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw). Following the same splits used in previous methods [22], we select the first 25 classes in alphabetical order as the known classes and group the rest as the unknown." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 143, + 481, + 259 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 469, + 125 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 469, + 125 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 469, + 125 + ], + "type": "text", + "content": "Table 1: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (OS)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 143, + 481, + 259 + ], + "lines": [ + { + "bbox": [ + 133, + 143, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 143, + 481, + 259 + ], + "type": "table", + "html": "
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP47.165.364.638.348.748.327.081.679.537.173.572.129.355.354.337.781.980.2
STA77.048.449.595.440.842.959.177.376.671.268.668.761.045.446.058.974.573.9
DAOD44.367.766.844.760.359.740.885.083.349.873.272.347.460.459.956.882.881.8
PGL34.773.772.238.459.258.427.684.882.66.181.578.625.166.865.038.084.883.0
PUJE38.474.873.442.264.463.637.987.785.823.481.078.839.271.270.039.188.286.3
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OSUNK OS*OS
OSBP28.657.256.125.877.875.823.085.483.033.077.275.516.771.369.232.165.964.6
STA64.145.946.662.067.267.066.276.676.257.465.264.960.257.657.772.749.350.2
DAOD71.155.556.166.669.269.163.779.378.754.878.277.354.670.269.655.162.962.6
PGL19.163.361.632.178.977.140.987.785.95.385.982.824.573.972.033.870.268.8
PUJE57.166.666.352.182.881.645.288.987.255.981.880.851.481.580.342.571.069.9
", + "image_path": "c93574545a28fb067e293ef80ae8f8d455aadc4a915fe544c3c2fd88b10564eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 323, + 481, + 389 + ], + "blocks": [ + { + "bbox": [ + 132, + 283, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 283, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 480, + 304 + ], + "type": "text", + "content": "Table 2: Accuracy of ResNet-50 model fine-tuned on Syn2Real-O dataset. * indicates our re-implementation with the officially released code." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 323, + 481, + 389 + ], + "lines": [ + { + "bbox": [ + 133, + 323, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 133, + 323, + 481, + 389 + ], + "type": "table", + "html": "
METHODplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckUNKOS*OSHOS
OSBP73.657.958.265.267.429.584.147.067.85.789.10.566.653.854.859.5
STA64.170.353.759.480.820.890.012.563.230.278.22.759.152.252.755.4
PGL81.568.374.260.691.945.492.241.087.967.579.26.449.666.865.556.9
PGL*81.377.566.571.790.635.792.441.282.646.982.34.959.664.564.162.0
ANNA41.254.639.759.451.225.682.654.878.910.576.62.067.448.149.656.1
PUJE90.773.875.172.477.833.891.258.879.546.584.95.169.065.866.067.3
", + "image_path": "ad919c5ac303e819cbc74c22204d829156213168786e51ba5d0d776311935742.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 424, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 424, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 133, + 424, + 481, + 509 + ], + "type": "text", + "content": "Syn2Real-O [26] is a more challenging synthetic-to-real benchmark, which is constructed from the VisDA-17 [25]. The Syn2Real-O dataset significantly increases the ratio of unknown data in the target domain to 0.9 by introducing additional unknown samples. In this dataset, the source domain contains training data from the VisDA-17 as the known set, and the target domain includes the test data from the VisDA-17 (known set) plus 50k images from irrelevant categories of MSCOCO [16] dataset (unknown set)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "content": "As reported in Tabs. 1 and 2, we observe that our method PUJE consistently outperforms the state-of-the-art results, improving OS and HOS by " + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "content": " on the benchmark datasets of Office-Home and Syn2Real-O respectively. Note that our proposed approach provides significant performance gains for the more challenging dataset of Syn2Real-O that requires knowledge transfer across different modalities. In the sub-tasks with a larger domain shift, e.g., " + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{Rw} \\rightarrow \\mathrm{Cl}" + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr} \\rightarrow \\mathrm{Cl}" + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "content": " in Office-Home, we can observe similar results demonstrating the strong adaptation ability of the proposed framework. PGL reported a higher OS score in Syn2Real-O, but we cannot reproduce the results with the officially released code. During the re-implementation based on their early stop strategy, we found the UNK was around " + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 133, + 510, + 481, + 665 + ], + "type": "text", + "content": ", which is far below the reported score. A potential reason to explain this is that maybe they did not remove the unknown category from the source data." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-set DA via Joint Error based PU Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 143, + 481, + 244 + ], + "blocks": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "lines": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "spans": [ + { + "bbox": [ + 139, + 114, + 473, + 126 + ], + "type": "text", + "content": "Table 3: Accuracy of ResNet-50 model fine-tuned on Office-Home dataset (HOS)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 143, + 481, + 244 + ], + "lines": [ + { + "bbox": [ + 133, + 143, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 143, + 481, + 244 + ], + "type": "table", + "html": "
METHODPr→ArPr→ClPr→RwRw→ArRW→ClRw→Pr
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOS
ROS64.357.360.671.246.556.378.470.874.470.867.068.8
OSLPP76.254.663.667.153.159.371.277.074.075.060.867.2
ANNA70.363.066.574.854.663.178.974.376.677.366.171.3
PUJE77.856.965.776.851.261.575.280.177.567.371.169.1
METHODAr→ClAr→PrAr→RwCl→RwCl→PrCl→ArMEAN
UNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*HOSUNK OS*
ROS74.150.660.170.368.469.377.275.876.572.265.368.6
OSLPP67.155.961.073.172.572.869.480.174.373.967.266.9
ANNA78.761.469.079.968.373.779.774.176.880.266.973.0
PUJE83.558.568.878.979.679.376.382.179.180.671.275.6
", + "image_path": "22fc7f5120fbf46d1e5d2e6bb70e4645304e492da73720d3c40dad39c3cbb96f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 221, + 296, + 389, + 366 + ], + "blocks": [ + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "text", + "content": "Table 4: Ablation studies of semi-supervised regularization losses. We report the Accuracy " + }, + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "text", + "content": " on Office-Home of " + }, + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\rightarrow \\mathrm{R}" + }, + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "inline_equation", + "content": "\\mathrm{C}\\to \\mathrm{P}" + }, + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "text", + "content": " using a ResNet50 backbone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 221, + 296, + 389, + 366 + ], + "lines": [ + { + "bbox": [ + 221, + 296, + 389, + 366 + ], + "spans": [ + { + "bbox": [ + 221, + 296, + 389, + 366 + ], + "type": "table", + "html": "
Config.Ar→PrCl→Rw
UNK OS* OSUNK OS* OS
default52.182.881.655.981.880.8
w/o Lpse69.678.177.864.877.877.3
w/o Lcon26.381.779.633.380.979.1
w/o Lent50.281.179.953.980.579.5
", + "image_path": "d8cb973b1306fe0394dbed14ea61becd07d57074a5569d35668bd143d07413e0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": "HOS Metric Instead of OS score, [18] proposed the harmonic mean of " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " and unk, " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{HOS} = 2(\\mathrm{OS}^{\\star} \\times \\mathrm{UNK}) / (\\mathrm{OS}^{\\star} + \\mathrm{UNK})" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " to penalize large gaps between " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " and unk. We agree that the OS score does not reflect UNK when the number of known classes grows. However, the HOS score is still unfair. " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " and UNK are equally valued in HOS, but UNK is much easier to improve than " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " which depends on the power of the adaptation algorithm. [15] utilizes a weak adaptation algorithm based on [28] that leads to a relatively loose alignment (Fig. 4) on the feature space of source and target domains such that unlabeled data is more frequently assigned to the unknown class. We believe that if a method can achieve a high " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " score, it is possible to obtain a decent HOS score by tuning hyper-parameters or additional losses. For our proposal, we can increase " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\beta = 0.25" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " to improve UNK at the sacrifice of " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " (Tab. 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": "Ablation Study To investigate the impact of the semi-supervised regularization, we compare three variants of the PUJE model on the Office-Home dataset shown in Tab. 4. Without " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "L_{pse}" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": " decreases because it can no longer leverage the pseudo labels that can help to build reliable hypothesis space " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{T_g}" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": " (Definition 7). Without " + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 544, + 482, + 617 + ], + "type": "text", + "content": ", UNK drops as it is not likely to draw a manifold-aware decision boundary that is beneficial to the separation of unknown data (Fig. 2a)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Robustness against Varying Openness To verify the robustness of the proposed PUJE, we conducted experiments on the Syn2Real-O with the openness varying in " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{0.25, 0.5, 0.75, 0.9\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ". Here, openness is defined by the ratio of unknown" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 247, + 201 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 247, + 201 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 247, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 247, + 201 + ], + "type": "image", + "image_path": "9f623d8a8102881900f455a6d93b300329f2c188562ea187d1995eb93e2f1a24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 205, + 224, + 215 + ], + "lines": [ + { + "bbox": [ + 157, + 205, + 224, + 215 + ], + "spans": [ + { + "bbox": [ + 157, + 205, + 224, + 215 + ], + "type": "text", + "content": "(a) PU Coefficient" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 118, + 361, + 200 + ], + "blocks": [ + { + "bbox": [ + 253, + 118, + 361, + 200 + ], + "lines": [ + { + "bbox": [ + 253, + 118, + 361, + 200 + ], + "spans": [ + { + "bbox": [ + 253, + 118, + 361, + 200 + ], + "type": "image", + "image_path": "4a0b854f463874efefecc0c4a3413973e7bfba59c6490d606dc1dc83bac6687a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 205, + 339, + 215 + ], + "lines": [ + { + "bbox": [ + 270, + 205, + 339, + 215 + ], + "spans": [ + { + "bbox": [ + 270, + 205, + 339, + 215 + ], + "type": "text", + "content": "(b) SSL Coefficient" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 367, + 118, + 475, + 201 + ], + "blocks": [ + { + "bbox": [ + 367, + 118, + 475, + 201 + ], + "lines": [ + { + "bbox": [ + 367, + 118, + 475, + 201 + ], + "spans": [ + { + "bbox": [ + 367, + 118, + 475, + 201 + ], + "type": "image", + "image_path": "70c99c257b1de897960d3c2589029d98b9bfb0dd150cea78234988c9b4aa9322.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 205, + 444, + 215 + ], + "lines": [ + { + "bbox": [ + 396, + 205, + 444, + 215 + ], + "spans": [ + { + "bbox": [ + 396, + 205, + 444, + 215 + ], + "type": "text", + "content": "(c) Openness" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 140, + 219, + 247, + 301 + ], + "blocks": [ + { + "bbox": [ + 140, + 219, + 247, + 301 + ], + "lines": [ + { + "bbox": [ + 140, + 219, + 247, + 301 + ], + "spans": [ + { + "bbox": [ + 140, + 219, + 247, + 301 + ], + "type": "image", + "image_path": "91fdbbd4da8cbf009efa2f4814fa03746412ba4ff43bd324322679c5de6f2ac8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 306, + 235, + 316 + ], + "lines": [ + { + "bbox": [ + 149, + 306, + 235, + 316 + ], + "spans": [ + { + "bbox": [ + 149, + 306, + 235, + 316 + ], + "type": "text", + "content": "(d) Trade-off in " + }, + { + "bbox": [ + 149, + 306, + 235, + 316 + ], + "type": "inline_equation", + "content": "\\mathrm{Ar}\\rightarrow \\mathrm{Cl}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 253, + 218, + 361, + 301 + ], + "blocks": [ + { + "bbox": [ + 253, + 218, + 361, + 301 + ], + "lines": [ + { + "bbox": [ + 253, + 218, + 361, + 301 + ], + "spans": [ + { + "bbox": [ + 253, + 218, + 361, + 301 + ], + "type": "image", + "image_path": "a2ed7e612b215ec82d0d1fd5bd51401cad5048f151cfdf52cfebd24bacc35099.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 306, + 354, + 316 + ], + "lines": [ + { + "bbox": [ + 258, + 306, + 354, + 316 + ], + "spans": [ + { + "bbox": [ + 258, + 306, + 354, + 316 + ], + "type": "text", + "content": "(e) Convergence in " + }, + { + "bbox": [ + 258, + 306, + 354, + 316 + ], + "type": "inline_equation", + "content": "\\mathrm{Ar}\\rightarrow \\mathrm{Cl}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 367, + 219, + 475, + 301 + ], + "blocks": [ + { + "bbox": [ + 367, + 219, + 475, + 301 + ], + "lines": [ + { + "bbox": [ + 367, + 219, + 475, + 301 + ], + "spans": [ + { + "bbox": [ + 367, + 219, + 475, + 301 + ], + "type": "image", + "image_path": "5441e85fb22f0abe9a4774e6021834e77cc845b6e0979bbd0bc3acc26c3b8b2c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 306, + 470, + 316 + ], + "lines": [ + { + "bbox": [ + 372, + 306, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 372, + 306, + 470, + 316 + ], + "type": "text", + "content": "(f) Convergence in " + }, + { + "bbox": [ + 372, + 306, + 470, + 316 + ], + "type": "inline_equation", + "content": "\\mathrm{Rw}\\rightarrow \\mathrm{Cl}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "type": "text", + "content": "Fig. 3: (a)-(b): sensitivity to varying loss coefficient " + }, + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\beta, \\omega" + }, + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "type": "text", + "content": " verified in " + }, + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{Ar} \\rightarrow \\mathrm{Pr}" + }, + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{Cl} \\rightarrow \\mathrm{Rw}" + }, + { + "bbox": [ + 130, + 324, + 482, + 381 + ], + "type": "text", + "content": " tasks; (c): performance comparisons w.r.t. varying openness of the Syn2Real-O task; (d): training cure of the proposed algorithm that characterizes a trade-off between the classification accuracy on known classes and the unknown class; (e)-(f): convergence analysis of the Office-Home task compared to other baselines with confidence intervals." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "text", + "content": "samples in the entire target data. We show the results of PGL and PUJE in Fig. 3c. Note that the PGL approach heuristically sets the hyper-parameter according to the true unknown ratio to control the openness, while PUJE automatically estimates the weight " + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "text", + "content": " during the training procedure. We observe that PUJE consistently outperforms the baseline by a large margin, especially on unknown recognition, which confirms its robustness to the change in openness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "content": "Training Curve In Fig. 3d, we illustrate the recognition performance of PUJE over training steps in the " + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "inline_equation", + "content": "\\mathrm{Ar} \\rightarrow \\mathrm{Cl}" + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "content": " task. All metrics show a performance gain over the first several steps. Then, the normalized accuracy " + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "content": " experiences a downward while the accuracy for the unknown class keeps improving, which characterizes a trade-off between " + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "inline_equation", + "content": "\\mathrm{OS}^{\\star}" + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "content": " and UNK. In addition, the behavior of " + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "inline_equation", + "content": "f_{S}^{\\prime}, f_{T}^{\\prime}" + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "content": " meets the expectation of our alignment mechanism where " + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "inline_equation", + "content": "f_{S}^{\\prime}" + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "content": " starts to detect unknown data, which is followed by " + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "inline_equation", + "content": "f_{T}^{\\prime}" + }, + { + "bbox": [ + 130, + 486, + 483, + 631 + ], + "type": "text", + "content": " (Fig. 2b). From Figs. 3e and 3f, we observe that some previous works [15,22] do not converge at the reported optimum, and they tackle this problem by ceasing the model updates before the score starts to decline. However, this early stop strategy can fail without enough labeled target data to validate the current performance. In contrast, our method can reach a reliable convergence." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 640, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 640, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 640, + 482, + 666 + ], + "type": "text", + "content": "Sensitivity to PU & SSL Coefficient We show the sensitivity of our approach to various PU and SSL loss coefficient " + }, + { + "bbox": [ + 131, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta, \\omega" + }, + { + "bbox": [ + 131, + 640, + 482, + 666 + ], + "type": "text", + "content": ". We tested the value of " + }, + { + "bbox": [ + 131, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 640, + 482, + 666 + ], + "type": "text", + "content": " from [0.01, 0.5]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 448, + 102 + ], + "type": "text", + "content": "Open-set DA via Joint Error based PU Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 222, + 175 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 222, + 175 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 222, + 175 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 222, + 175 + ], + "type": "image", + "image_path": "9ff734422a3e51cc04995add15f161b0c758ecc1876ef43b88b339e080dfeb75.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 226, + 115, + 304, + 175 + ], + "blocks": [ + { + "bbox": [ + 226, + 115, + 304, + 175 + ], + "lines": [ + { + "bbox": [ + 226, + 115, + 304, + 175 + ], + "spans": [ + { + "bbox": [ + 226, + 115, + 304, + 175 + ], + "type": "image", + "image_path": "1dc159318049829926bfd8c878ddb1b3c91617cf4b36aff3e011932ecac04b31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 388, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 388, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 388, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 388, + 175 + ], + "type": "image", + "image_path": "4e1db5e706e0f59f9854bce07b19bda165d78dfccb2ed755691039c74c9939e7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 116, + 470, + 175 + ], + "blocks": [ + { + "bbox": [ + 392, + 116, + 470, + 175 + ], + "lines": [ + { + "bbox": [ + 392, + 116, + 470, + 175 + ], + "spans": [ + { + "bbox": [ + 392, + 116, + 470, + 175 + ], + "type": "image", + "image_path": "e769a3fb4f133ecd3154fde9819ce3329846d8a94ee3522faa23d8ade5a02d41.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 142, + 190, + 221, + 250 + ], + "blocks": [ + { + "bbox": [ + 163, + 178, + 201, + 189 + ], + "lines": [ + { + "bbox": [ + 163, + 178, + 201, + 189 + ], + "spans": [ + { + "bbox": [ + 163, + 178, + 201, + 189 + ], + "type": "text", + "content": "(a) OSBP" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 142, + 190, + 221, + 250 + ], + "lines": [ + { + "bbox": [ + 142, + 190, + 221, + 250 + ], + "spans": [ + { + "bbox": [ + 142, + 190, + 221, + 250 + ], + "type": "image", + "image_path": "42b22b296ec64a08cf44f844076cc89671fc3e60cdc868bf1447f3765c44ef86.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 254, + 200, + 263 + ], + "lines": [ + { + "bbox": [ + 163, + 254, + 200, + 263 + ], + "spans": [ + { + "bbox": [ + 163, + 254, + 200, + 263 + ], + "type": "text", + "content": "(e) OSBP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 272, + 480, + 285 + ], + "lines": [ + { + "bbox": [ + 132, + 272, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 480, + 285 + ], + "type": "text", + "content": "Fig. 4: Feature space visualized by t-SNE in (a)-(d): Office-Home; (e)-(h): Syn2Real-O." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 227, + 190, + 304, + 250 + ], + "blocks": [ + { + "bbox": [ + 248, + 178, + 282, + 189 + ], + "lines": [ + { + "bbox": [ + 248, + 178, + 282, + 189 + ], + "spans": [ + { + "bbox": [ + 248, + 178, + 282, + 189 + ], + "type": "text", + "content": "(b) PGL" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 190, + 304, + 250 + ], + "lines": [ + { + "bbox": [ + 227, + 190, + 304, + 250 + ], + "spans": [ + { + "bbox": [ + 227, + 190, + 304, + 250 + ], + "type": "image", + "image_path": "48ab1a9f47aa3e015200d4c8e50f058504a6cd167c7369bdcfec0c79066ec9bf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 254, + 281, + 264 + ], + "lines": [ + { + "bbox": [ + 249, + 254, + 281, + 264 + ], + "spans": [ + { + "bbox": [ + 249, + 254, + 281, + 264 + ], + "type": "text", + "content": "(f) PGL" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 190, + 387, + 250 + ], + "blocks": [ + { + "bbox": [ + 329, + 178, + 367, + 189 + ], + "lines": [ + { + "bbox": [ + 329, + 178, + 367, + 189 + ], + "spans": [ + { + "bbox": [ + 329, + 178, + 367, + 189 + ], + "type": "text", + "content": "(c) ANNA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 190, + 387, + 250 + ], + "lines": [ + { + "bbox": [ + 309, + 190, + 387, + 250 + ], + "spans": [ + { + "bbox": [ + 309, + 190, + 387, + 250 + ], + "type": "image", + "image_path": "2763721c66614cb069558809ebe9e454e7ef87ea3fe187ab5e56815d0a9e14ac.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 254, + 368, + 264 + ], + "lines": [ + { + "bbox": [ + 328, + 254, + 368, + 264 + ], + "spans": [ + { + "bbox": [ + 328, + 254, + 368, + 264 + ], + "type": "text", + "content": "(g) ANNA" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 392, + 190, + 470, + 250 + ], + "blocks": [ + { + "bbox": [ + 413, + 178, + 450, + 189 + ], + "lines": [ + { + "bbox": [ + 413, + 178, + 450, + 189 + ], + "spans": [ + { + "bbox": [ + 413, + 178, + 450, + 189 + ], + "type": "text", + "content": "(d) PUJE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 190, + 470, + 250 + ], + "lines": [ + { + "bbox": [ + 392, + 190, + 470, + 250 + ], + "spans": [ + { + "bbox": [ + 392, + 190, + 470, + 250 + ], + "type": "image", + "image_path": "2a2f759faff32a31f774512916fb1891f8a590265e81e170c81277d7181e42c4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 254, + 450, + 264 + ], + "lines": [ + { + "bbox": [ + 413, + 254, + 450, + 264 + ], + "spans": [ + { + "bbox": [ + 413, + 254, + 450, + 264 + ], + "type": "text", + "content": "(h) PUJE" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "content": "in " + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{Ar}\\rightarrow \\mathrm{Pr}" + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "content": " task and " + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "content": " from [5, 25] in " + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{Cl}\\rightarrow \\mathrm{Rw}" + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "content": " task while fixing other parameters to the default setting. We can draw two observations from Figs. 3a and 3b: the OS score is relatively stable, and the unknown recognition achieves a reliable performance when coefficient " + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "content": " is within the interval of [0.02, 0.05]; Generally, a larger " + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "content": ", which is set for lower openness, will lead to a better performance on recognizing unknown data while hurt the accuracy for known classes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": "Feature Space Visualization To intuitively visualize the effectiveness of OUDA approaches, we extracted features from the baseline models (OSBP, PGL, ANNA) and our proposed model on the " + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{Ar}\\rightarrow \\mathrm{Cl}" + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": " task (Office-Home) and Syn2Real-O task with the ResNet-50 backbone [13]. The feature distributions were processed with t-SNE [23] afterward. As shown in Fig. 4, compared with baselines, our method PUJE achieves a better alignment between source and target distributions, especially when the domain shift is large. Benefiting from our joint error-based adversarial alignment mechanism, the extracted feature space, including the cluster of unknown target data, has a more discriminative class-wise decision boundary." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 565, + 220, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 220, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 220, + 578 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We have addressed the open-set domain shift problem by introducing a novel learning theory based on multi-class positive-unlabeled learning that can reduce the open-set risk and the joint error. Experiments show that our proposed learning theory performs consistently better on challenging object recognition benchmarks for open-set adaptation with significant domain gaps and label shifts." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 175 + ], + "type": "text", + "content": "This research is partially supported by JST Moonshot R&D Grant Number JPMJPS2011, CREST Grant Number JPMJCR2015 and Basic Research Grant (Super AI) of Institute for AI and Beyond of the University of Tokyo." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 203 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 203 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 236 + ], + "type": "text", + "content": "1. Baktashmotlagh, M., Faraki, M., Drummond, T., Salzmann, M.: Learning factorized representations for open-set domain adaptation. In: ICLR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 481, + 257 + ], + "type": "text", + "content": "2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.: A theory of learning from different domains. Machine Learning 79, 151-175 (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 258, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 481, + 278 + ], + "type": "text", + "content": "3. Bucci, S., Loghmani, M.R., Tommasi, T.: On the effectiveness of image rotation for open set domain adaptation. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 279, + 424, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 424, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 424, + 289 + ], + "type": "text", + "content": "4. Busto, P.P., Gall, J.: Open set domain adaptation. In: ICCV (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "text", + "content": "5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: CVPRW (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 311, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 481, + 342 + ], + "type": "text", + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. IEEE Conference on Computer Vision and Pattern Recognition pp. 248-255 (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "type": "text", + "content": "7. Fang, Z., Lu, J., Liu, F., Xuan, J., Zhang, G.: Open set domain adaptation: Theoretical bound and algorithm. arXiv abs/1907.08375 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 365, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 385 + ], + "type": "text", + "content": "8. Feng, Q., Kang, G., Fan, H., Yang, Y.: Attract or distract: Exploit the margin of open set. In: ICCV (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "text", + "content": "9. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of Machine Learning Research 17(1), 2096-2030 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 439 + ], + "type": "text", + "content": "0. Garg, S., Balakrishnan, S., Lipton, Z.C.: Domain adaptation under open set label shift. In: NeurIPS (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 440, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 481, + 460 + ], + "type": "text", + "content": "1. Gomes, R., Krause, A., Perona, P.: Discriminative clustering by regularized information maximization. In: NIPS (2010)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 461, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 481, + 482 + ], + "type": "text", + "content": "2. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. In: NIPS (2005)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 483, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 481, + 504 + ], + "type": "text", + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. IEEE Conference on Computer Vision and Pattern Recognition pp. 770-778 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "text", + "content": "4. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 546 + ], + "type": "text", + "content": "5. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "text", + "content": "6. Lin, T.Y., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 580, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 481, + 600 + ], + "type": "text", + "content": "7. Liu, H., Cao, Z., Long, M., Wang, J., Yang, Q.: Separate to adapt: Open set domain adaptation via progressive separation. In: CVPR (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "type": "text", + "content": "8. Loghmania, M.R., Vinczea, M., Tommasi, T.: Positive-unlabeled learning for open set domain adaptation. Pattern Recognition Letters 136, 198-204 (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "9. Long, M., Cao, Y., Wang, J., Jordan, M.I.: Learning transferable features with deep adaptation networks. In: Proceedings of the 32nd International Conference on International Conference on Machine Learning. vol. 37, pp. 97-105. JMLR.org (2015)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "text", + "content": "Open-set DA via Joint Error based PU Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 621 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "20. Long, M., CAO, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. In: Advances in Neural Information Processing Systems, pp. 1640-1650. Curran Associates, Inc. (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "21. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Proceedings of the 34th International Conference on Machine Learning. vol. 70, pp. 2208-2217. JMLR.org (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "22. Luo, Y., Wang, Z., Huang, Z., Baktashmotlagh, M.: Progressive graph learning for open-set domain adaptation. In: ICML (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "23. van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "24. Oliver, A., Odena, A., Raffel, C., Cubuk, E.D., Goodfellow, I.J.: Realistic evaluation of deep semi-supervised learning algorithms. In: NeurIPS (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "25. Peng, X., Usman, B., Kaushik, N., Hoffman, J., Wang, D., Saenko, K.: Visda: The visual domain adaptation challenge. arXiv abs/1710.06924 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "26. Peng, X., Usmana, B., Saito, K., Kaushik, N., Hoffman, J., Saenko, K.: Syn2real: A new benchmark for synthetic-to-real visual domain adaptation. arXiv abs/1806.09755 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "27. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3723-3732 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "text", + "content": "28. Saito, K., Yamamoto, S., Ushiku, Y., Harada, T.: Open set domain adaptation by backpropagation. In: ECCV (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "29. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NIPS (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "30. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NIPS (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "31. Tang, H., Chen, K.C., Jia, K.: Unsupervised domain adaptation via structurally regularized deep clustering. In: IEEE Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "32. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 2962-2971 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "33. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. IEEE Conference on Computer Vision and Pattern Recognition pp. 5385-5394 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "34. Wang, Q., Meng, F., Breckon, T.P.: Progressively select and reject pseudo-labelled samples for open-set domain adaptation. In: AAAI (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 544 + ], + "type": "text", + "content": "35. Wu, J., He, J.: Domain adaptation with dynamic open-set targets. In: KDD (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "36. Xu, Y., Xu, C., Xu, C., Tao, D.: Multi-positive and unlabeled learning. In: IJCAI (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "37. Zhang, D., Westfechtel, T., Harada, T.: Unsupervised domain adaptation via minimized joint error. In: TMLR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "38. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Proceedings of Machine Learning Research. vol. 97, pp. 7523-7532. PMLR (2019)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "D.Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_content_list.json b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fa554fcec03126205ec34e67a66a2dd3102c76f5 --- /dev/null +++ b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_content_list.json @@ -0,0 +1,1692 @@ +[ + { + "type": "text", + "text": "Opens3D: Snap and Lookup for 3D Open-vocabulary Instance Segmentation", + "text_level": 1, + "bbox": [ + 250, + 141, + 751, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhening Huang1", + "bbox": [ + 316, + 212, + 449, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyang $\\mathbf{W u}^{2}$", + "bbox": [ + 467, + 212, + 584, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xi Chen $^{2}$", + "bbox": [ + 604, + 212, + 684, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hengshuang Zhao\\*", + "bbox": [ + 302, + 228, + 459, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Zhu $^{3,4}$", + "bbox": [ + 473, + 228, + 563, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joan Lasenby", + "bbox": [ + 581, + 228, + 700, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Cambridge 2The University of Hong Kong", + "bbox": [ + 305, + 253, + 697, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Hong Kong University of Science and Technology", + "bbox": [ + 333, + 268, + 668, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4The Hong Kong University of Science and Technology (Guangzhou)", + "bbox": [ + 272, + 282, + 728, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://zheninghuang.github.io/OpenIns3D", + "bbox": [ + 344, + 300, + 656, + 313 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc2489a3e5f401235117350154fd6d85d7b0ac99b092789ef21aad86adc82d8b.jpg", + "image_caption": [ + "\"furniture that is capable of producing music.\"" + ], + "image_footnote": [], + "bbox": [ + 227, + 320, + 406, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0fe4c32c28f73c06a571d0f37b7b6051fc3628edb9df459fe3726dd327c647a.jpg", + "image_caption": [ + "device to watch BBC news", + "Fig. 1: Complex Queries 3D Instance Segmentation with OpenIns3D." + ], + "image_footnote": [], + "bbox": [ + 411, + 320, + 591, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd1a3f2ed6ae89ed11399adf6ff3197706b7ba264a727977b35d7b7b595e7ef1.jpg", + "image_caption": [ + "\"Ma Long's domain of excellence" + ], + "image_footnote": [], + "bbox": [ + 594, + 320, + 774, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this work, we introduce OpenIns3D, a new 3D-input-only framework for 3D open-vocabulary scene understanding. The OpenIns3D framework employs a \"Mask-Snap-Lookup\" scheme. The \"Mask\" module learns class-agnostic mask proposals in 3D point clouds, the \"Snap\" module generates synthetic scene-level images at multiple scales and leverages 2D vision-language models to extract interesting objects, and the \"Lookup\" module searches through the outcomes of \"Snap\" to assign category names to the proposed masks. This approach yet simple, achieves state-of-the-art performance across a wide range of 3D open-vocabulary tasks, including recognition, object detection, and instance segmentation, on both indoor and outdoor datasets. Moreover, OpenIns3D facilitates effortless switching between different 2D detectors without requiring retraining. When integrated with powerful 2D open-world models, it achieves excellent results in scene understanding tasks. Furthermore, when combined with LLM-powered 2D models, OpenIns3D exhibits an impressive capability to comprehend and process highly complex text queries that demand intricate reasoning and real-world knowledge.", + "bbox": [ + 259, + 462, + 740, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open-Vocabulary Understanding $\\cdot$ 3D Scene Understanding $\\cdot$ Vision-Language Model", + "bbox": [ + 259, + 710, + 740, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 753, + 375, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding plays a critical role in various domains, such as autonomous driving, robotic sensing, AR/VR, and manufacturing, among others.", + "bbox": [ + 212, + 785, + 785, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57840a1969993fc1483db598a8b507bc99a38ca0dc2030bd1f620f9f98a0388b.jpg", + "image_caption": [ + "Fig. 2: High-level Illustrations of OpenIns3D and Quantitative Results. (a) OpenIns3D follows the \"Mask-Snap-Lookup\" steps for open-vocabulary scene understanding. (b) A list of SOTA results has been achieved on both indoor and outdoor datasets. OV-Rec: open-vocabulary object recognition. OVOD: open-vocabulary object detection. OVIS: open-vocabulary instance segmentation. PointCLIPV2 [43]; Uni3D [40]; Open3DIS [22]; FM-OV3D [36]" + ], + "image_footnote": [], + "bbox": [ + 223, + 152, + 566, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8b2f93d3e12882c7e597be4a8a9716fc2eae35b6904cd515dba5b847a408e329.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 152, + 781, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the development of 3D closed-set understanding is relatively mature, scene understanding in an open-vocabulary setting is still in its infancy. Closed-set understanding can only handle a predefined set of concepts and scenarios but fails to provide valid responses when faced with unfamiliar concepts or variations in language usage. This limitation impacts its performance in dynamic and ever-changing contexts.", + "bbox": [ + 212, + 430, + 787, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to internet-scale image-text datasets, significant progress has been made in 2D image open-vocabulary understanding [4, 7, 11, 16, 24, 33, 38, 39]. However, unlike 2D data that can be easily collected from the internet, constructing a large-scale 3D-text dataset poses a challenge. As a result, the most viable approach to achieving 3D open-vocabulary understanding involves leveraging 2D images to bridge language and 3D data. In this direction, there have been several notable works, such as OpenScene [23], PLA-family [9, 10, 34], and CLIP2Scene [6]. These works leverage well-aligned 2D images and 3D point clouds to conduct feature distillation or employ 2D caption models to construct 3D-text pairs. One prerequisite of these methods, however, is the availability of well-aligned 2D images and 3D point clouds. This means that posed 2D images, associated depth maps and camera models need to be accessible as inputs to the network. In real-life scenarios, there are numerous cases where 2D images or the information required to align 2D and 3D data are unavailable. For instance, to save storage space, point clouds generated with LiDAR are often stored without accompanying 2D images (example datasets include [13, 25, 31]). In cases where point clouds are obtained from the registration of multiple scans from different sensors or are converted from 3D simulations/CAD models [12, 21], 2D images are often non-existent.", + "bbox": [ + 212, + 521, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We believe that developing a 3D open-vocabulary framework without relying on well-aligned 2D images is meaningful, as this will simplify deployment pre", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "requisites and enhance its applicability across a wide range of scenarios. To this end, we introduce OpenIns3D, a framework designed to effectively perform 3D open-vocabulary scene understanding tasks without relying on 2D aligned images. Overall, OpenIns3D comprises three core steps: Mask, Snap, and Lookup. An overall illustration of OpenIns3D is presented in Figure 2a.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask: Given a 3D point cloud, the first part of OpenIns3D learns class-agnostic mask proposals with a Mask Proposal Module (MPM). This process is trained without any classification labels. To control the quality of the mask, MPM proposes a learnable Mask Scoring module to predict the quality of each mask output and implements a list of Mask Filtering techniques to discard invalid, low-quality masks. MPM outputs a list of class-agnostic masks in the scene.", + "bbox": [ + 212, + 224, + 784, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Snap: Multiple synthetic scene-level images are generated with calibrated and optimized camera poses and intrinsic parameters. These images are specifically designed to encompass part or all of the relevant masks, aiming to minimize the need for multiple renderings. Instead of individually predicting the category of each mask proposal [20, 37, 43], the scene-level images are input into 2D open-vocabulary models for the simultaneous understanding of all interesting objects present in the scene. A Class Lookup Table (CLT) is then constructed to store all the detected object categories alongside their respective pixel locations.", + "bbox": [ + 212, + 319, + 785, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lookup: To precisely determine the positions of mask proposals in each image, Mask2Pixel maps are constructed. These maps project all 3D mask proposals onto 2D images with identical camera parameters used in Snap. In the Lookup phase, OpenIns3D searches through the CLT with the help of Mask2Pixel maps to precisely assign category names to 3D mask proposals. Results from multiple views are combined to establish initial mask classification outcomes. For remaining masks, a similar Lookup procedure is carried out on a local scale to facilitate classification. Lastly, the 3D mask proposals are refined by removing masks lacking class assignments after both global and local Lookup.", + "bbox": [ + 212, + 444, + 785, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OpenIns3D demonstrates promising performance in extensive comparisons with other methods, as summarized in Figure 2b. This simple and flexible approach achieves a range of state-of-the-art (SOTA) results on various 3D open-vocabulary tasks. Specifically, SOTA results are achieved on open-vocabulary instance segmentation for the indoor S3DIS [1] dataset and the outdoor STPLS3D [5] dataset. OpenIns3D even outperforms OpenMask3D [30], a concurrent work that heavily utilizes 2D images, for instance, segmentation on the challenging Replica [28] dataset. We also find that the \"Snap and Lookup\" serves as a powerful open-vocabulary object recognition engine, which achieved SOTA object recognition tasks on ScanNet [8], outperforming the previous SOTA, a 1-billion parameter 3D foundation model [40]. Lastly, when converting mask proposals into 3D bounding boxes, OpenIns3D also achieved state-of-the-art results in open-vocabulary object detection (OVOD) on ScanNet, outperforming previous image-dependent methods [36]. The design of OpenIns3D also allows 2D detectors to be changed without the need for retraining. This provides the model with the capability to evolve alongside the latest development of the 2D open-vocabulary models. Moreover, when 2D detectors are coupled with Large Lan-", + "bbox": [ + 212, + 583, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/56e04e05cf354c1035c4332cae16288847e2d3ee8f0c2b3747bd7bf4f8ab3c8e.jpg", + "image_caption": [ + "(a) 3D Feature Distillation" + ], + "image_footnote": [], + "bbox": [ + 223, + 152, + 346, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5700f25ca1d75f620c5e91366bded4bc61d16fca6ad8c22ebb93857c706f28a5.jpg", + "image_caption": [ + "(b) Building 3D-text pair" + ], + "image_footnote": [], + "bbox": [ + 367, + 152, + 491, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ea3b54d16ccd3fb51b8ca573a330619e299eeda3b24ec51d7d245f1cfdae1ed.jpg", + "image_caption": [ + "(c) CLIP and Projection", + "Fig. 3: Four Categories of Open-Vocabulary 3D Scene Understanding Models. a) 3D feature distillation frameworks, where 2D images are used as a bridge to distil language-aligned features into 3D, with typical works including OpenScene [23] and Clip2Scene [6]. b) Building 3D-text pairs, where 2D captioning models are used to build 3D-text pairs for feature learning, with typical works including the PLA-family [9,10,34] c) CLIP and Projection, where objects are cropped out of 2D images before being processed by CLIP, and the results are directly projected into 3D, including OpenMask3D [30], OV-3DET [20], $\\mathrm{CLIP}^2$ [35] and Open3DIS [22]. d) OpenIns3D" + ], + "image_footnote": [], + "bbox": [ + 511, + 152, + 635, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f02f5869c5d4c0c810eba53bf727f6dedfc9597e07fb45b5b818b7436e848ba4.jpg", + "image_caption": [ + "(d) Mask-Snap-Lookup (ours)" + ], + "image_footnote": [], + "bbox": [ + 656, + 152, + 776, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "guage Models (LLMs), OpenIns3D can enable complex query understanding capability. When integrated with LISA [18], an LLM-powered reasoning segmentation model, OpenIns3D exhibits a strong ability to comprehend highly intricate language queries and performs reasoning segmentation in 3D, as illustrated in Figure 1. In summary, our contributions are:", + "bbox": [ + 212, + 426, + 787, + 502 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- OpenIns3D employs a distinct pipeline that operates without the need for well-aligned images. This approach achieves state-of-the-art results across a range of benchmarks and possesses the ability to comprehend highly complex input queries.", + "- The proposed \"Snap and Lookup\" combination can serve as a powerful 3D object recognition engine, especially for noisy 3D objects extracted from scene-level scans. This capability is demonstrated by achieving state-of-the-art results with a large margin." + ], + "bbox": [ + 218, + 503, + 787, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 643, + 382, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D open-vocabulary understanding. Progress in 3D open-vocabulary understanding has been relatively slow compared to that of images. In the domain of 3D object classification tasks, methods like PointCLIP [37], PointCLIPV2 [43], and CLIP2Point [14] project 3D point clouds into depth maps and link them with 2D models for classification. However, these methods lack performance in scene-level understanding, where points are often overlapped and incomplete. For scene-level understanding, most work has primarily focused on leveraging well-aligned 2D posed images, depth maps, and point clouds [6,9,10,23,26,34,35]. One notable example is OpenScene [23], which takes posed 2D images, depth maps, and 3D data as input, and feature distillation is performed to transfer 2D language-aligned features from images to 3D point clouds. Similarly,", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clip2Scene [6] builds dense pixel-point pairs by calibrating the LiDAR point cloud with corresponding images captured by six cameras. However, achieving instance-level understanding is challenging with these methods as they focus solely on semantic-level understanding. In contrast, PLA [9] and its follow-up work RegionPLC [34] and Lowis3D [10] utilize a 2D caption model to construct 3D-text pairs to learn features. However, the PLA-family works rely on a binary head to classify the input object into base categories or novel categories, and the transferability of this binary head to different base-novel splits is very limited, posing a challenge for flexible applications. One current work, OpenMask3D [30], utilizes well-aligned 2D images to learn features for mask proposals, leading to impressive results in open-vocabulary instance segmentation. Open3DIS [22] follows the same procedure, enhancing both mask proposals and mask classification with images, yielding better performance. A common issue with these methods is their reliance on well-aligned 3D and 2D pairs in the input, which may not always be available in real applications. We summarise the key difference between these methods in Figure 3. Simplifying input requirements improves method flexibility and compatibility. In this work, we explore how to conduct 3D open-vocabulary understanding without relying on 2D images.", + "bbox": [ + 217, + 145, + 785, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image generation from 3D. Projection-based methods have been extensively explored in the past for 3D understanding and have proven to be beneficial for obtaining complementary features. For instance, MVCNN [29] projects 3D objects to different views to aid in feature learning, while LAR [3] introduces object centre projection methods to generate images for 3D objects from various angles, assisting visual grounding tasks. Additionally, Virtual View Fusion [17] employs the original camera pose but enlarges the field of view, resulting in enhanced 2D feature transfer. However, these methods encounter challenges like best view selection, object occlusion, information loss during projection, and long rendering times. In the context of open-vocabulary settings, the quality of the projected image plays a crucial role in model performance. In our work, we evaluate different projection methods, along with their compatibility with 2D open-vocabulary models, to identify an optimal solution that achieves good results and is efficient to implement.", + "bbox": [ + 217, + 420, + 785, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 OpenIns3D", + "text_level": 1, + "bbox": [ + 217, + 661, + 359, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Baseline and Challenges", + "text_level": 1, + "bbox": [ + 217, + 700, + 459, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline. We build a naive baseline by adopting the recent 3D instance segmentation backbone Mask3D [27] to generate mask proposals. To make the Mask Proposal Module (MPM) fit for the open-vocabulary setting, we remove all components in Mask3D that use the classification labels. Later, PointCLIP [43] is adopted for mask understanding. This naive approach, although satisfying the requirement of 3D inputs only, has long rendering times and unsatisfactory performance (See in Table 7). There are several problems in this baseline model.", + "bbox": [ + 217, + 734, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d1b9b92f820f6790837997a5d5d46ee43c3e3a0409ae66129bfb0c616b33ab2.jpg", + "image_caption": [ + "Fig. 4: General Pipeline of OpenIns3D OpenIns3D first processes point clouds with MPM to generate 3D mask proposals and mask scores. The Snap module (detailed in Figure 5) then renders $N$ synthetic scene-level images, which are later passed into the 2D open-world model along with the input text queries. The detection results from the 2D model are stored in the Class Lookup Table (CLT). Finally, both the mask proposals and CLT are fed into the Lookup module, where Mask2Pixel Guided Lookup (detailed in Figure 6) is performed at the global level, followed by a Local Enforced Lookup at the local level to unlock the semantic meaning of mask proposals. The final mask filtering refines the mask proposals and obtains the final results." + ], + "image_footnote": [], + "bbox": [ + 222, + 130, + 782, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Challenge 1: excessive mask proposals. in Mask3D [27], mask proposals are filtered by the mask classification logit, which is removed in the class-agnostic setting. Therefore, an effective mask filtering scheme is needed.", + "bbox": [ + 212, + 441, + 787, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Challenge 2: low quality of 3D instances. 3D instances extracted from scene scans are typically broken, uncompleted, and sparse. Therefore, generated images by simple projection are not easily understood by 2D VL models.", + "bbox": [ + 212, + 487, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Challenge 3: lack of context information. Humans recognize imperfect 3D point cloud objects through scene comprehension. However, isolated instance point cloud projections lack this context. A solution is to project both the mask and background onto images, yet this introduces distracting elements, potentially confusing classification-level models like CLIP.", + "bbox": [ + 212, + 532, + 787, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Challenge 4: domain gap between projected images and natural images. Rendered images often differ significantly from natural images used in training, posing a challenge for 2D visual language models to comprehend.", + "bbox": [ + 212, + 607, + 787, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Overall Framework", + "text_level": 1, + "bbox": [ + 214, + 674, + 419, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present our design of OpenIns3D, which targets the aforementioned four problems. The pipeline of OpenIns3D is shown in Figure 4.", + "bbox": [ + 212, + 698, + 787, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mask: class-agnostic mask proposal", + "text_level": 1, + "bbox": [ + 214, + 750, + 511, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Within the Mask Proposal Module (MPM), we introduce two straightforward designs aimed at filtering low-quality masks generated in the baseline model. Mask scoring. Inspired by [7, 15, 16], We feed the instance queries generated from the Mask Module into a shallow MLP module to predict the quality, i.e", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "IoU of the mask. The predicted IoU $(IoU_{m})$ is supervised by the ground truth IoU $(IoU_{gt})$ value during the training stages, which is calculated between the predicted mask and its matched ground truth mask in the Bipartite Matching. For unmatched prediction masks $(IoU_{u})$ , we label the ground truth IoU value as zero. The loss is computed using $L2$ . To avoid overly low IoU predictions, a hyper-parameter $\\gamma$ is introduced to reduce the weight of loss for unmatched masks. Therefore, the total loss function for the MPM is:", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma \\sum \\left(I o U _ {u}\\right) ^ {2} + \\sum \\left(I o U _ {m} - I o U _ {\\mathrm {g t}}\\right) ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 266, + 784, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask filtering. To enhance mask quality, three filters are applied. Firstly, we retain masks with a model-predicted IoU score above a threshold of $\\beta$ , ensuring that only high-quality masks are kept. Secondly, drawing inspiration from SAM [16], we focus on stable masks by comparing two binary masks derived from the same underlying soft mask using different threshold values. Specifically, we introduce an offset value $\\alpha$ and select masks where the IoU between the pair of thresholded masks (one with $-\\alpha$ and the other with $+\\alpha$ ) exceeds $80\\%$ . Lastly, small objects in the scene often lead to invalid proposals, so we filter out mask proposals that have a point number lower than $N_{\\mathrm{min}}$ . By employing these techniques, the quantity of mask proposals will decrease, resulting in cleaner and higher-quality masks for subsequent mask understanding tasks (Challenge 1).", + "bbox": [ + 212, + 297, + 784, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Snap: Synthetic Scene-level View Generation", + "text_level": 1, + "bbox": [ + 214, + 489, + 589, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rendering images from points can be a time-consuming task, especially when the number of rendering jobs is high. We propose a multi-scale synthetic scene-level image scheme.", + "bbox": [ + 212, + 521, + 782, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera pose selection. The Snap module captures scene-level images at three scales: global, corner, and wide-angle, as shown in Figure 5. For global-level images, cameras are positioned above the scene and point directly toward the centre of the scene. For corner images, cameras are positioned above the centre and point toward the corner, while for the wide-angle images, cameras are positioned at the 3 by 3 grid interaction points, pointing toward the furthest corner. Using the camera position coordinates $P_{\\mathrm{cam}}$ , target coordinates $P_{\\mathrm{target}}$ , and the up axis of the scene $U$ , the Lookat function can be employed to determine the pose matrix Pose. A more detailed mathematical formulation of this is presented in the supplementary materials.", + "bbox": [ + 212, + 566, + 784, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera intrinsic calibration. Once the camera extrinsic matrix is established, the fields of view for the captured images are adjusted by modifying the camera's intrinsic parameters. The goal is to ensure that the entire scene or a specific part of it is encompassed within the captured images. To achieve this goal, we initialize an arbitrary camera intrinsic matrix and then adjust the focal lengths $(f_x$ and $f_y)$ and the principal point coordinates $(c_x$ and $c_y)$ through scaling. The scaling is performed by readjusting the projected areas in the image coordinate space. For example: if the projected points of the pre-defined area", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af8221a65488d1b74b92faed1c87ef1836ad9ca7513063a556a1967303d0672f.jpg", + "image_caption": [ + "Fig.5: Snap and Mask2Pixel Maps. Multiscale snaps are conducted to render images with different levels of detail for scene understanding, including wide-corner snaps, wide-angle snaps, and global snaps. Cameras are positioned on the top of the scene and point towards the centre or corners, and the field of view is determined with the calibrated intrinsic matrix. With the defined camera models, Mask2Pixel maps are built to store the location of each 3D mask in the 2D image (using the same colour to represent 2D-3D correspondences) to guide the search for category names." + ], + "image_footnote": [], + "bbox": [ + 250, + 145, + 787, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "were located in image coordinates within the range of $[-1000, -192]$ in the $x$ -domain, our calibrated intrinsic parameters transform this range to $[0, 1000]$ in $x$ . Importantly, we preserve the aspect ratio between the $x$ and $y$ coordinates to maintain the proportions of the final image without any distortion. This procedure ensures that each captured image is fully utilized and encompasses all regions of interest within the scene (Challenge 2 & 3).", + "bbox": [ + 212, + 454, + 787, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class lookup table. Upon obtaining $N$ synthetic scene-level images, we input them into a 2D open-vocabulary detector. With text queries provided for interested classes, a list of detected objects in synthetic images can be obtained. Subsequently, information about detected objects, including their location and class, are stored in a designated Class Lookup Table (CLT). This table will later be retrieved to allocate class categories to 3D mask proposals (Challenge 4).", + "bbox": [ + 212, + 545, + 787, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lookup: Mask Classification through Searching", + "text_level": 1, + "bbox": [ + 214, + 657, + 609, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a multi-level search to assign category labels to the mask proposals generated from Mask step.", + "bbox": [ + 212, + 688, + 785, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask2Pixel guided lookup. We introduce a Mask2Pixel Guided Lookup (MGL) to search within CLT. The concept involves projecting each 3D mask proposal onto a 2D plane using the same camera extrinsic and intrinsic matrices that are utilized to generate the 2D image, as depicted in Figure 5. With knowledge of the precise pixel locations of each mask in images, we can conduct an accurate search through the CLT to identify the most likely class for each mask. The development of MPM takes into account occlusion by integrating depth information. To accomplish the matching, we follow a three-step approach: 1. based", + "bbox": [ + 212, + 719, + 792, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on the mask's projection onto the 2D plane, we select the best-matched class categories in terms of IoU values; 2. if the IoU value of the best-matched object on the 2D plane is below $20\\%$ , the match is disregarded. 3. we aggregate results from multiple views to formulate the final prediction, calculating probability scores using their normalized average IoU values, as illustrated in Figure 6.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Local enforced lookup. While the Mask2Pixel Guided Lookup assigns class categories to mask proposals, some masks may not correspond to objects in the CLT. To address this, we introduce a Local Enforced Lookup (LEL) approach. We crop out the remaining masks from 2D scene-level images using enlarged bounding boxes and process them with the 2D detector to encourage detection. To select the best views, we introduce an Occlusion Report method to assess occlusion conditions for each mask in each projection, and then choose the top $K$ views for LEL. More details for Occlusion Report can be found in the supplementary materials.", + "bbox": [ + 212, + 222, + 486, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Final mask refinement. With the previous lookup approaches, a large proportion of mask proposals obtain a category prediction. All masks that have no category predictions after the MGL and LEL stages are eliminated.", + "bbox": [ + 212, + 493, + 486, + 583 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8398733af888cae47acd68594093de78c76a573e84a9b273f7fcbcfd6487bfa8.jpg", + "image_caption": [ + "Fig. 6: Mask2Pixel Guided Lookup Illustration. IoUs between the 2D detection results and the projected masks are the guidance to assign class names to 3D masks. Multiple image results are ensembled." + ], + "image_footnote": [], + "bbox": [ + 504, + 234, + 781, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 599, + 375, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets and evaluation scheme. We tested OpenIns3D on five datasets, including four indoor datasets, namely S3DIS [2], ScanNetv2 [8], ScanNet200 [26], Replica [28], and one outdoor dataset, STPLS3D [5]. Among them, S3DIS, ScanNetv2, and ScanNet200 are indoor point cloud datasets generated from RGB-D images, Replica is a photo-realistic 3D indoor scene reconstruction, while STPLS3D is an aerial photogrammetry-constructed outdoor dataset. We exclusively used the 3D data with colour from these datasets and did not employ any 2D images, poses, or depth maps. Following the settings of prior work [9], we excluded the \"other furniture\" class in ScanNetv2 and the \"clutter\" class in S3DIS due to their vague meanings. Replica and ScanNet200 are evaluated by following the settings of OpenMask3D [30]. For STPLS3D, we merged the low, medium, and high vegetation classes into one \"vegetation\" class and kept all the rest.", + "bbox": [ + 212, + 628, + 787, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. For the S3DIS, ScanNetv2, Scannet200, and STPLS datasets, the MPM module is trained without utilizing any category labels, and", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5a58ddd1d6957f4a3ae0db4080c24410a4be6b0f74446b141dff21944c46f9f7.jpg", + "table_caption": [ + "Table 1: Zero-shot object classification on ScanNetv2. OpenIns3D's Snap and Lookup approach for mask classification, surpasses all previous methods, including the latest language-aligned large-scale 3D foundation model [40]." + ], + "table_footnote": [], + "table_body": "
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointCLIP [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
PointCLIP V2 [37]11.00.00.023.80.00.00.00.07.80.090.70.00.00.00.064.40.00.0
CLIP2Point [14]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP2[35]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
Uni3D [40]45.858.53.778.883.754.931.339.470.135.11.927.394.213.838.710.788.147.6
OpenIns3D60.885.227.487.677.346.954.864.271.49.980.882.771.661.438.70.087.985.7
", + "bbox": [ + 225, + 208, + 784, + 309 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$\\lambda$ is set to 0.1 to reduce the weight of zero-IoU. The mask proposal of Replica is trained on ScanNet200, followed by OpenMask3D [30]. The Snap module captures images with a size of $1000 \\times 1000$ including 16 global snaps, 4 corner snaps, and 4 wide-angle snaps. The top $0.5\\mathrm{m}$ of the scene is removed for S3DIS, as the rooms are enclosed. For STPLIS3D, we followed Mask3D to split the large outdoor scene into patches of $50m \\times 50m$ and lifted the camera up to $10m$ . More implementation details are presented in the supplementary materials.", + "bbox": [ + 212, + 337, + 782, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Open-vocabulary point cloud recognition. We first evaluate OpenIns3D performance among all existing 3D open-vocabulary models with the same setting, i.e. using only 3D inputs. These models are most commonly tested on recognition tasks, including PointCLIP(v1&v2) [43] [37], Clip2Point [14], CLIP² [35]. We also compare with large-scale 3D foundation models, such as Uni3D [40], which has 1-billion parameters, and was trained with large-scale 3D shapes and image-text pairs. We followed their evaluation scheme and reported the Top-1 accuracy of instance classification on ScanNetv2. The results are presented in Table 1.", + "bbox": [ + 212, + 443, + 784, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Open-vocabulary instance segmentation. We adopted various comparison schemes to align with existing methods. For 3D Open-vocabulary Instance Segmentation, we compared with PLA [9], and its follow-up works RegionPLC [34] and Lowis3D [10]. For a fair comparison, we followed their category splits and compared our results on novel classes, as demonstrated in Table 2. For STPLS3D, we compared OpenIns3D with baseline models whose classification module is PointCLIP and PointCLIPV2 [43] (Table 5). We also explored the performance of OpenIns3D on a more challenging dataset with more class categories. Specifically, we compared the performance of OpenIns3D with OpenMask3D, OpenScene, on Replica (Table 4) as well as ScanNet200 (Table 6). Following OpenMask3D, we used Mask3D to generate mask proposals for OpenScene for evaluation.", + "bbox": [ + 212, + 582, + 784, + 747 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Open-vocabulary object detection. Since there were limited works on 3D open-vocabulary instance segmentation at the time of conducting this work, we also selected some of the latest methods in the 3D open-world object detection domain for a more comprehensive evaluation. The evaluation is performed by converting generated masks into axis-aligned bounding boxes, and the results are shown in Table 3.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/959332eff066c6b5aa442a74270fab5510dce4d12128d6508a9eb0f00559ca1a.jpg", + "table_caption": [ + "Table 2: 3D Open-vocabulary Instance Segmentation Results on S3DIS and ScanNetv2. We compare our zero-shot performance on the novel categories defined in the PLA-family work. Significant improvements are achieved on the S3DIS dataset, and competitive results are observed on ScanNetv2 (B/N: Base/Novel)." + ], + "table_footnote": [], + "table_body": "
OVISS3DISScanNetv2
MethodB/NAP50AP25B/NAP50AP25require 2D
PLA [9]8/408.6-10/721.9-✓
RegionPLC [34]8/4--10/732.3-✓
Lowis3D [10]8/413.8-10/731.2-✓
Open3DIS [22]8/426.3-10/7--✓
Mask3d+PointClip [43]- / 405.410.3- / 704.507.8X
OpenIns3D- / 437.039.3- / 727.942.6X
improvement- / 4(+10.7)(+29.0)- / 7-(+34.8)X
PLA [9]6/609.8-8/925.1-✓
RegionPLC [34]6/6--8/932.2-✓
Lowis3D [10]6/615.8-8/938.1-✓
Open3DIS [22]6/629.0-8/9--✓
Mask3d+PointClip [43]- / 608.510.6- / 905.606.7X
OpenIns3D- / 633.038.9- / 919.527.9X
improvement- / 6(+4.0)(+28.3)- / 9-(+21.2)X
Mask3d+PointClip [43]- / 1208.609.3- / 1704.514.4X
OpenIns3D- / 1228.329.5- / 1728.738.9X
improvement- / 12(+19.7)(+20.2)- / 17(+24.2)(+24.5)X
", + "bbox": [ + 222, + 212, + 777, + 439 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/033b301a8e4acfcf4fbc0b6e193c49b948d200570093cb609681b73eee187b6a.jpg", + "table_caption": [ + "Table 3: Open-vocabulary Object Detection $(AP_{25})$ on unseen classes in ScanNet." + ], + "table_footnote": [], + "table_body": "
Methodsmeantoiletbedchairsofadressertablecabinetbookshelfpillowsink
OV-PointCLIP [37]3.16.62.36.33.90.77.20.72.10.60.8
OV-Image2Point [32]0.80.20.81.01.40.22.81.00.90.00.1
Detic-ModelNet [41]1.74.21.04.61.20.23.20.61.20.00.7
Detic-ImageNet [41]0.40.00.00.20.00.51.80.50.30.00.7
OV-3DETIC [20]12.749.02.67.318.62.814.32.44.53.921.1
L3Det [42]24.656.336.216.123.08.123.114.717.323.427.9
FM-OV3D [36]21.555.038.819.241.923.83.50.46.017.48.8
OpenIns3D43.779.570.576.915.80.053.140.141.27.153.1
improvement(+19.1)(+23.2)(+31.7)(+57.7)--(+30)(+25.4)(+23.9)(+25.2)
", + "bbox": [ + 217, + 458, + 795, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Results and Discussion", + "text_level": 1, + "bbox": [ + 215, + 609, + 473, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Comparison with SOTA", + "text_level": 1, + "bbox": [ + 215, + 645, + 459, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "First of all, OpenIns3D demonstrates impressive performance in open-vocabulary point cloud recognition, surpassing all previous methods, including the large-scale 3D foundation model by $15\\%$ . This proves the effectiveness of the zero-shot \"Snap\" and \"lookup\" scheme. With the enhanced recognition capability, the performance of 3D open-vocabulary Object Detection among the ScanNet dataset has also achieved state-of-the-art results by a large margin. For 3D instance segmentation, compared to works in the PLA family [9, 10, 34] and the latest work Open3DIS [22], OpenIns3D does not require aligned images as input, (still need RGB information for points), achieves higher results on the S3DIS dataset, both in the 4 novel categories split and the 6 novel categories split. In SPTLS3D, OpenIns3D outperforms the baseline model PointCLIPV2 by $9.3\\%$ in AP. On", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3cc6073981b81418449e38e2c47aa7f0196dcec7b513fd7cbff6561248a81270.jpg", + "table_caption": [ + "Table 4: OVIS on the indoor Replica dataset." + ], + "table_footnote": [], + "table_body": "
Model2DAP\\( AP_{50} \\)\\( AP_{25} \\)
OpenScene [23] (2D Fusion)✓10.915.617.3
OpenScene [23] (2D/3D Ens.)✓8.210.413.3
OpenMask3D [30]✓13.118.424.2
OpenScene [23] (3D Distill)✗8.210.512.6
OpenIns3D✗13.618.019.7
improvement✗(+5.4)(+7.5)(+7.1)
", + "bbox": [ + 215, + 184, + 537, + 301 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ab849c100e8f51877e20188bf17d56c86210d04251060644384940c679ecfe6e.jpg", + "table_caption": [ + "Table 5: OVIS on outdoor STPLS3D dataset" + ], + "table_footnote": [], + "table_body": "
ModelAPAP50AP25
PointCLIP [43]02.002.604.0
PointCLIPv2 [37]02.103.105.2
OpenIns3D11.414.217.2
improvement(+9.3)(+11.9)(+12.0)
", + "bbox": [ + 539, + 184, + 776, + 301 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a4fac431bf628534861606e8f8675e334c58de44b67fb8f612e77846f5dd4028.jpg", + "table_caption": [ + "Table 6: 3D instance segmentation results on the ScanNet200 validation set. OpenIns3D demonstrates robust performance when compared to 2D-input-free models. However, notable limitations emerge when dealing with small objects in the common and tail classes." + ], + "table_footnote": [], + "table_body": "
Modeluse 2DAPheadAPcommonAPtailAPAP50AP25
OpenScene (2D Fusion) [23]✓13.411.69.911.715.217.8
OpenScene (2D/3D Ens.) [23]✓11.03.21.15.36.78.1
OpenMask3D✓17.114.114.915.419.923.1
OpenScene (3D Distill)✗10.62.60.74.86.27.2
OpenIns3D✗16.06.54.28.810.314.4
improvement✗(+5.4)(+3.9)(+3.6)(+4.0)(+4.1)(+7.2)
", + "bbox": [ + 225, + 386, + 779, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the Replica dataset, OpenIns3D even outperformed OpenMask3D, which relies on well-aligned images for mask understanding. In the case of ScanNet200, OpenIns3D attains the highest performance compared to all other 3D input baselines. However, we have observed a decline in performance on tail and common classes within ScanNet200. The decrease in performance can be attributed to the low-quality and unclear reconstruction of smaller objects in the ScanNet200 scene. This is a noticeable limitation of OpenIns3D, but for better reconstructions, such as in Replica, and for Head classes within ScanNet200, OpenIns3D still demonstrates decent performance.", + "bbox": [ + 212, + 517, + 787, + 652 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In summary, OpenIns3D demonstrates the best performance among all existing methods if only 3D data is used as input and outperforms many existing state-of-the-art methods that require 2D images. It also shows certain limitations on small objects that are not well-reconstructed in 3D scenes.", + "bbox": [ + 212, + 652, + 787, + 714 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 737, + 387, + 753 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mask quality ablation. Following the evaluation on ScanNetv2, we assessed the class-agnostic mask quality using the average precision (AP) score. We treated all classes as universal since the predictions are class-agnostic. The evaluation was conducted on the ScanNetv2 validation set. Table 8 demonstrates the effectiveness of the Mask Scoring and Mask Filtering designs.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b64e81d3a19f420713b956449ae944df67733b6291ee320283f98d3234a2f426.jpg", + "table_caption": [ + "Table 7: Rendering and Inference Time Ablations. Results tested on typical ScanNet scenes with 50 masks. OpenIns3D requires less rendering and inference time." + ], + "table_footnote": [], + "table_body": "
RenderingNum of Img needed2D backboneImg size (w × h)\\( T_{render} \\) (s/scene)\\( T_{infer} \\) (s/scene)\\( T_{total} \\) (s/scene)AP25 (%)
PointCLIP250CLIP\\( 128^2 \\)5.215.320.59.3
LAR250CLIP\\( 128^2 \\)14.318.733.010.5
Mask rendering250CLIP\\( 128^2 \\)42.619.562.17.3
OpenIns3D (ours)8G-DINO\\( 1000^2 \\)2.36.28.529.8
OpenIns3D (ours)8ODISE\\( 1000^2 \\)2.38.210.535.1
", + "bbox": [ + 217, + 184, + 785, + 303 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/71b5cb7a727fac59e27441d98f19b587e37cc1cfaa4533163211aa2c4a5ea9ae.jpg", + "table_caption": [ + "Table 8: MPM Ablation. MS: Mask scoring. MF: Mask filtering." + ], + "table_footnote": [], + "table_body": "
MethodAP50AP25
Mask3d-Supervised74.780.9
CA-Mask3d47.549.2
CA-Mask3d + MS50.2 (+02.7)53.3 (+04.1)
CA-Mask3d + MF61.6 (+14.1)71.0 (+21.8)
CA-Mask3d + MS + MF64.6 (+17.0)73.4 (+24.2)
", + "bbox": [ + 217, + 354, + 524, + 446 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c8ec054766ec7a02501d42b4dd8099de73847bd240a059147df6a64f749452d1.jpg", + "table_caption": [ + "Table 9: Number of Views Abla-tion. LEL: Local Enforced Lookup" + ], + "table_footnote": [], + "table_body": "
IDX4816LELAP50AP25
i✓18.327.1
ii✓22.735.1
iii✓24.837.5
iv✓✓28.738.9
", + "bbox": [ + 545, + 356, + 790, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multi-view ablation. We also studied the effects of using different numbers of views (Table 9). Increasing the number of views used in the Lookup module leads to better results. Additionally, Look Enforced Lookup provided a final improvement to the results.", + "bbox": [ + 212, + 453, + 784, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Projection and 2D backbone ablation. We conducted a comprehensive study on various rendering methods and their interaction with the 2D backbone to identify a suitable approach. We report the rendering time and inference performance for each method (Table 7), and more details can be found in the supplementary materials. The key observation is that the scene-level rendering and understanding approach excels in speed while also demonstrating strong performance. Switching from Grounding Dino [19] to the latest ODISE [33] also brings gains in performance, indicating that the OpenIns3D framework can easily benefit from the rapid development of 2D open-world detectors.", + "bbox": [ + 212, + 516, + 784, + 652 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Cross-domain analysis. To evaluate the generalization capability of MPM across different domains, we trained and tested OpenIns3D on two different datasets, as shown in Table 10. The cross-domain models also demonstrate impressive performance on both datasets when compared with the baseline. Notably, within the 17 classes of ScanNetv2, 11 classes do not exist in S3DIS. OpenIns3D, trained on S3DIS, still achieves decent performance among these unseen classes.", + "bbox": [ + 212, + 655, + 784, + 760 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Free-flow language capability. The snap and lookup scheme outsources the mask understanding tasks to a 2D vision-language model. Therefore when integrated with a 2D model powered by large language models (LLMs) like LISA [18], OpenIns3D can perform reasoning-based segmentation tasks (as depicted in Figure 7). For instance, when given the query \"pencilling down ideas during brain-", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "storming,\" OpenIns3D accurately segments the whiteboard, while for \"furniture offers recreational enjoyment with friends,\" it precisely identifies and segments the pool table.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1df998dcb8085692438928bd7bbade94c4fe596135cb01b58cdbdadfc6a06a98.jpg", + "image_caption": [ + "Fig. 7: Qualitative Results from OpenIns3D. OpenIns3D (ODISE) demonstrates the ability to manage a versatile vocabulary. OpenIns3D (LISA) can conduct 3D reasoning segmentation." + ], + "image_footnote": [], + "bbox": [ + 223, + 208, + 776, + 406 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/4562aa1fd1f7de6fd3a1427677f75da2244989a0bf75c74df202f0d551aa2819.jpg", + "table_caption": [ + "Table 10: Cross-domain Ablation. We trained and tested OpenIns3D on two different datasets to examine its cross-domain capability. While S3DIS and ScanNetV2 have non-overlapping classes, OpenIns3D demonstrates decent generalization capability." + ], + "table_footnote": [], + "table_body": "
TestModelTraining DataAP50AP25
ScanNetv2 [8]Mask3D-Pointclip [43]ScanNetv204.514.4
OpenIns3DScanNetv228.738.9
OpenIns3DS3DIS21.533.6
S3DIS [1]Mask3D-Pointclip [43]S3DIS03.506.8
OpenIns3DS3DIS28.329.5
OpenIns3DScanNetv214.219.8
", + "bbox": [ + 240, + 518, + 759, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 631, + 359, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Achieving 3D open-vocabulary scene understanding is a challenging task, primarily due to the lack of extensive 3D-text data. Currently, most work in this domain focuses on using 2D images to bridge the gap between 3D and language. This, however, not only requires a good alignment between 2D and 3D but also evolves slowly due to the significant effort needed for retraining when changing the 2D backbone. OpenIns3D introduces a new pipeline, i.e. Mask-Snap-Lookup, for this task. The Mask module generates authentic masks in the 3D domain, while Snap renders scene-level images in 2D domains, and the Lookup module links the results from 2D to 3D precisely. This pipeline requires no image input, achieves better performance, and can evolve seamlessly with 2D models without training. We hope our work will provide a fresh perspective for researchers working towards open-world 3D scene understanding.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the Girton College Graduate Research Awards at the University of Cambridge, School of Technology, National Highways sponsored through EPSRC Centre for Doctoral Training, the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR, the National Natural Science Foundation of China No. 62201484, HKU Startup Fund, and HKU Seed Fund for Basic Research. We would like to express our gratitude to Yunhan Yang for his assistance in exploring rendering techniques and to Chengyao Wang for sharing his implementation of Mask3D.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 321, + 321, + 337 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3, 14", + "2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 9", + "3. Bakr, E.M., Alsaedy, Y.Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: NeurIPS (2022) 5", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 2", + "5. Chen, M., Hu, Q., Yu, Z., THOMAS, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. In: BMVA (2022) 3, 9", + "6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023) 2, 4, 5", + "7. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. In: ICCV (2023) 2, 6", + "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9, 14", + "9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 2, 4, 5, 9, 10, 11", + "0. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. In: TPAMI (2024) 2, 4, 5, 10, 11", + "1. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: ICML (2023) 2", + "2. Griffiths, D., Boehm, J.: SynthCity: A large-scale synthetic point cloud. In: arXiv (2019) 2", + "3. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: Semantic3D.net: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017) 2", + "4. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV (2023) 4, 10" + ], + "bbox": [ + 225, + 353, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Huang, Z., Huang, L., Gong, Y., Huang, C., Wang, X.: Mask scoring R-CNN. In: CVPR (2019) 6", + "16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 2, 6, 7", + "17. Kundu, A., Yin, X., Fathi, A., Ross, D.A., Brewington, B., Funkhouser, T.A., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: ECCV (2020) 5", + "18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024) 4, 13", + "19. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv (2023) 13", + "20. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR (2023) 3, 4, 11", + "21. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019) 2", + "22. Nguyen, P.D.A., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024) 2, 4, 5, 11", + "23. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 4, 12", + "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2", + "25. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-lille-3d: A large and high-quality ground-truth urban point cloud dataset for automatic segmentation and classification. In: The International booktitle of Robotics Research (2018) 2", + "26. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 4, 9", + "27. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 5, 6", + "28. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces. In: arXiv (2019) 3, 9", + "29. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.G.: Multi-view convolutional neural networks for 3d shape recognition. In: ICCV (2015) 5", + "30. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023) 3, 4, 5, 9, 10, 12", + "31. Tan, W., Qin, N., Ma, L., Li, Y., Du, J., Cai, G., Yang, K., Li, J.: Toronto-3D: A large-scale mobile lidar dataset for semantic segmentation of urban roadways. In: CVPRW (2020) 2" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z.Huang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Xu, C., Yang, S., Galanti, T., Wu, B., Yue, X., Zhai, B., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: Image2point: 3d point-cloud understanding with 2d image pretrained models. In: ECCV (2022) 11", + "33. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In: CVPR (2023) 2, 13", + "34. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024) 2, 4, 5, 10, 11", + "35. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: $\\mathrm{Clip}^2$ : Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR (2023) 4, 10", + "36. Zhang, D., Li, C., Zhang, R., Xie, S., Xue, W., Xie, X., Zhang, S.: Fm-ov3d: Foundation model-based cross-modal knowledge blending for open-vocabulary 3d detection. In: AAAI (2024) 2, 3, 11", + "37. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by CLIP. In: CVPR (2022) 3, 4, 10, 11, 12", + "38. Zhou, C., Loy, C.C., Dai, B.: Denseclip: Extract free dense labels from clip. In: CVPR (2022) 2", + "39. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022) 2", + "40. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024) 2, 3, 10", + "41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 11", + "42. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. In: arXiv (2023) 11", + "43. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV (2023) 2, 3, 4, 5, 10, 11, 12, 14" + ], + "bbox": [ + 212, + 146, + 787, + 575 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OpenIns3D: 3D Open-vocabulary Instance Segmentation", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_model.json b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6c6eb88b9ba481744bc2a686edbc967488649213 --- /dev/null +++ b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_model.json @@ -0,0 +1,2313 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.142, + 0.753, + 0.187 + ], + "angle": 0, + "content": "Opens3D: Snap and Lookup for 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.213, + 0.45, + 0.228 + ], + "angle": 0, + "content": "Zhening Huang1" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.213, + 0.585, + 0.228 + ], + "angle": 0, + "content": "Xiaoyang \\(\\mathbf{W u}^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.605, + 0.213, + 0.685, + 0.227 + ], + "angle": 0, + "content": "Xi Chen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.229, + 0.46, + 0.243 + ], + "angle": 0, + "content": "Hengshuang Zhao\\*" + }, + { + "type": "text", + "bbox": [ + 0.475, + 0.229, + 0.565, + 0.243 + ], + "angle": 0, + "content": "Lei Zhu\\(^{3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.229, + 0.701, + 0.243 + ], + "angle": 0, + "content": "Joan Lasenby" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.255, + 0.699, + 0.27 + ], + "angle": 0, + "content": "1University of Cambridge 2The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.269, + 0.669, + 0.284 + ], + "angle": 0, + "content": "\\(^{3}\\)Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.283, + 0.73, + 0.298 + ], + "angle": 0, + "content": "4The Hong Kong University of Science and Technology (Guangzhou)" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.301, + 0.657, + 0.314 + ], + "angle": 0, + "content": "https://zheninghuang.github.io/OpenIns3D" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.321, + 0.408, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.406, + 0.407, + 0.417 + ], + "angle": 0, + "content": "\"furniture that is capable of producing music.\"" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.321, + 0.592, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.406, + 0.559, + 0.415 + ], + "angle": 0, + "content": "device to watch BBC news" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.321, + 0.776, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.406, + 0.753, + 0.415 + ], + "angle": 0, + "content": "\"Ma Long's domain of excellence" + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.422, + 0.763, + 0.436 + ], + "angle": 0, + "content": "Fig. 1: Complex Queries 3D Instance Segmentation with OpenIns3D." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.463, + 0.741, + 0.699 + ], + "angle": 0, + "content": "Abstract. In this work, we introduce OpenIns3D, a new 3D-input-only framework for 3D open-vocabulary scene understanding. The OpenIns3D framework employs a \"Mask-Snap-Lookup\" scheme. The \"Mask\" module learns class-agnostic mask proposals in 3D point clouds, the \"Snap\" module generates synthetic scene-level images at multiple scales and leverages 2D vision-language models to extract interesting objects, and the \"Lookup\" module searches through the outcomes of \"Snap\" to assign category names to the proposed masks. This approach yet simple, achieves state-of-the-art performance across a wide range of 3D open-vocabulary tasks, including recognition, object detection, and instance segmentation, on both indoor and outdoor datasets. Moreover, OpenIns3D facilitates effortless switching between different 2D detectors without requiring retraining. When integrated with powerful 2D open-world models, it achieves excellent results in scene understanding tasks. Furthermore, when combined with LLM-powered 2D models, OpenIns3D exhibits an impressive capability to comprehend and process highly complex text queries that demand intricate reasoning and real-world knowledge." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.712, + 0.741, + 0.74 + ], + "angle": 0, + "content": "Keywords: Open-Vocabulary Understanding \\(\\cdot\\) 3D Scene Understanding \\(\\cdot\\) Vision-Language Model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.376, + 0.769 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.787, + 0.817 + ], + "angle": 0, + "content": "3D scene understanding plays a critical role in various domains, such as autonomous driving, robotic sensing, AR/VR, and manufacturing, among others." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.154, + 0.568, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.154, + 0.782, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 2: High-level Illustrations of OpenIns3D and Quantitative Results. (a) OpenIns3D follows the \"Mask-Snap-Lookup\" steps for open-vocabulary scene understanding. (b) A list of SOTA results has been achieved on both indoor and outdoor datasets. OV-Rec: open-vocabulary object recognition. OVOD: open-vocabulary object detection. OVIS: open-vocabulary instance segmentation. PointCLIPV2 [43]; Uni3D [40]; Open3DIS [22]; FM-OV3D [36]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.522 + ], + "angle": 0, + "content": "While the development of 3D closed-set understanding is relatively mature, scene understanding in an open-vocabulary setting is still in its infancy. Closed-set understanding can only handle a predefined set of concepts and scenarios but fails to provide valid responses when faced with unfamiliar concepts or variations in language usage. This limitation impacts its performance in dynamic and ever-changing contexts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Thanks to internet-scale image-text datasets, significant progress has been made in 2D image open-vocabulary understanding [4, 7, 11, 16, 24, 33, 38, 39]. However, unlike 2D data that can be easily collected from the internet, constructing a large-scale 3D-text dataset poses a challenge. As a result, the most viable approach to achieving 3D open-vocabulary understanding involves leveraging 2D images to bridge language and 3D data. In this direction, there have been several notable works, such as OpenScene [23], PLA-family [9, 10, 34], and CLIP2Scene [6]. These works leverage well-aligned 2D images and 3D point clouds to conduct feature distillation or employ 2D caption models to construct 3D-text pairs. One prerequisite of these methods, however, is the availability of well-aligned 2D images and 3D point clouds. This means that posed 2D images, associated depth maps and camera models need to be accessible as inputs to the network. In real-life scenarios, there are numerous cases where 2D images or the information required to align 2D and 3D data are unavailable. For instance, to save storage space, point clouds generated with LiDAR are often stored without accompanying 2D images (example datasets include [13, 25, 31]). In cases where point clouds are obtained from the registration of multiple scans from different sensors or are converted from 3D simulations/CAD models [12, 21], 2D images are often non-existent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We believe that developing a 3D open-vocabulary framework without relying on well-aligned 2D images is meaningful, as this will simplify deployment pre" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "requisites and enhance its applicability across a wide range of scenarios. To this end, we introduce OpenIns3D, a framework designed to effectively perform 3D open-vocabulary scene understanding tasks without relying on 2D aligned images. Overall, OpenIns3D comprises three core steps: Mask, Snap, and Lookup. An overall illustration of OpenIns3D is presented in Figure 2a." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.785, + 0.317 + ], + "angle": 0, + "content": "Mask: Given a 3D point cloud, the first part of OpenIns3D learns class-agnostic mask proposals with a Mask Proposal Module (MPM). This process is trained without any classification labels. To control the quality of the mask, MPM proposes a learnable Mask Scoring module to predict the quality of each mask output and implements a list of Mask Filtering techniques to discard invalid, low-quality masks. MPM outputs a list of class-agnostic masks in the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.787, + 0.44 + ], + "angle": 0, + "content": "Snap: Multiple synthetic scene-level images are generated with calibrated and optimized camera poses and intrinsic parameters. These images are specifically designed to encompass part or all of the relevant masks, aiming to minimize the need for multiple renderings. Instead of individually predicting the category of each mask proposal [20, 37, 43], the scene-level images are input into 2D open-vocabulary models for the simultaneous understanding of all interesting objects present in the scene. A Class Lookup Table (CLT) is then constructed to store all the detected object categories alongside their respective pixel locations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.787, + 0.581 + ], + "angle": 0, + "content": "Lookup: To precisely determine the positions of mask proposals in each image, Mask2Pixel maps are constructed. These maps project all 3D mask proposals onto 2D images with identical camera parameters used in Snap. In the Lookup phase, OpenIns3D searches through the CLT with the help of Mask2Pixel maps to precisely assign category names to 3D mask proposals. Results from multiple views are combined to establish initial mask classification outcomes. For remaining masks, a similar Lookup procedure is carried out on a local scale to facilitate classification. Lastly, the 3D mask proposals are refined by removing masks lacking class assignments after both global and local Lookup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.84 + ], + "angle": 0, + "content": "OpenIns3D demonstrates promising performance in extensive comparisons with other methods, as summarized in Figure 2b. This simple and flexible approach achieves a range of state-of-the-art (SOTA) results on various 3D open-vocabulary tasks. Specifically, SOTA results are achieved on open-vocabulary instance segmentation for the indoor S3DIS [1] dataset and the outdoor STPLS3D [5] dataset. OpenIns3D even outperforms OpenMask3D [30], a concurrent work that heavily utilizes 2D images, for instance, segmentation on the challenging Replica [28] dataset. We also find that the \"Snap and Lookup\" serves as a powerful open-vocabulary object recognition engine, which achieved SOTA object recognition tasks on ScanNet [8], outperforming the previous SOTA, a 1-billion parameter 3D foundation model [40]. Lastly, when converting mask proposals into 3D bounding boxes, OpenIns3D also achieved state-of-the-art results in open-vocabulary object detection (OVOD) on ScanNet, outperforming previous image-dependent methods [36]. The design of OpenIns3D also allows 2D detectors to be changed without the need for retraining. This provides the model with the capability to evolve alongside the latest development of the 2D open-vocabulary models. Moreover, when 2D detectors are coupled with Large Lan-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.153, + 0.348, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.287, + 0.328, + 0.298 + ], + "angle": 0, + "content": "(a) 3D Feature Distillation" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.153, + 0.493, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.287, + 0.469, + 0.299 + ], + "angle": 0, + "content": "(b) Building 3D-text pair" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.153, + 0.636, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.287, + 0.607, + 0.298 + ], + "angle": 0, + "content": "(c) CLIP and Projection" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.153, + 0.777, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.287, + 0.771, + 0.299 + ], + "angle": 0, + "content": "(d) Mask-Snap-Lookup (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.306, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 3: Four Categories of Open-Vocabulary 3D Scene Understanding Models. a) 3D feature distillation frameworks, where 2D images are used as a bridge to distil language-aligned features into 3D, with typical works including OpenScene [23] and Clip2Scene [6]. b) Building 3D-text pairs, where 2D captioning models are used to build 3D-text pairs for feature learning, with typical works including the PLA-family [9,10,34] c) CLIP and Projection, where objects are cropped out of 2D images before being processed by CLIP, and the results are directly projected into 3D, including OpenMask3D [30], OV-3DET [20], \\(\\mathrm{CLIP}^2\\) [35] and Open3DIS [22]. d) OpenIns3D" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.503 + ], + "angle": 0, + "content": "guage Models (LLMs), OpenIns3D can enable complex query understanding capability. When integrated with LISA [18], an LLM-powered reasoning segmentation model, OpenIns3D exhibits a strong ability to comprehend highly intricate language queries and performs reasoning segmentation in 3D, as illustrated in Figure 1. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.504, + 0.788, + 0.564 + ], + "angle": 0, + "content": "- OpenIns3D employs a distinct pipeline that operates without the need for well-aligned images. This approach achieves state-of-the-art results across a range of benchmarks and possesses the ability to comprehend highly complex input queries." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.564, + 0.788, + 0.624 + ], + "angle": 0, + "content": "- The proposed \"Snap and Lookup\" combination can serve as a powerful 3D object recognition engine, especially for noisy 3D objects extracted from scene-level scans. This capability is demonstrated by achieving state-of-the-art results with a large margin." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.504, + 0.788, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.383, + 0.66 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.844 + ], + "angle": 0, + "content": "3D open-vocabulary understanding. Progress in 3D open-vocabulary understanding has been relatively slow compared to that of images. In the domain of 3D object classification tasks, methods like PointCLIP [37], PointCLIPV2 [43], and CLIP2Point [14] project 3D point clouds into depth maps and link them with 2D models for classification. However, these methods lack performance in scene-level understanding, where points are often overlapped and incomplete. For scene-level understanding, most work has primarily focused on leveraging well-aligned 2D posed images, depth maps, and point clouds [6,9,10,23,26,34,35]. One notable example is OpenScene [23], which takes posed 2D images, depth maps, and 3D data as input, and feature distillation is performed to transfer 2D language-aligned features from images to 3D point clouds. Similarly," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.786, + 0.419 + ], + "angle": 0, + "content": "Clip2Scene [6] builds dense pixel-point pairs by calibrating the LiDAR point cloud with corresponding images captured by six cameras. However, achieving instance-level understanding is challenging with these methods as they focus solely on semantic-level understanding. In contrast, PLA [9] and its follow-up work RegionPLC [34] and Lowis3D [10] utilize a 2D caption model to construct 3D-text pairs to learn features. However, the PLA-family works rely on a binary head to classify the input object into base categories or novel categories, and the transferability of this binary head to different base-novel splits is very limited, posing a challenge for flexible applications. One current work, OpenMask3D [30], utilizes well-aligned 2D images to learn features for mask proposals, leading to impressive results in open-vocabulary instance segmentation. Open3DIS [22] follows the same procedure, enhancing both mask proposals and mask classification with images, yielding better performance. A common issue with these methods is their reliance on well-aligned 3D and 2D pairs in the input, which may not always be available in real applications. We summarise the key difference between these methods in Figure 3. Simplifying input requirements improves method flexibility and compatibility. In this work, we explore how to conduct 3D open-vocabulary understanding without relying on 2D images." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.421, + 0.786, + 0.631 + ], + "angle": 0, + "content": "Image generation from 3D. Projection-based methods have been extensively explored in the past for 3D understanding and have proven to be beneficial for obtaining complementary features. For instance, MVCNN [29] projects 3D objects to different views to aid in feature learning, while LAR [3] introduces object centre projection methods to generate images for 3D objects from various angles, assisting visual grounding tasks. Additionally, Virtual View Fusion [17] employs the original camera pose but enlarges the field of view, resulting in enhanced 2D feature transfer. However, these methods encounter challenges like best view selection, object occlusion, information loss during projection, and long rendering times. In the context of open-vocabulary settings, the quality of the projected image plays a crucial role in model performance. In our work, we evaluate different projection methods, along with their compatibility with 2D open-vocabulary models, to identify an optimal solution that achieves good results and is efficient to implement." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.662, + 0.36, + 0.68 + ], + "angle": 0, + "content": "3 OpenIns3D" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.702, + 0.46, + 0.717 + ], + "angle": 0, + "content": "3.1 Baseline and Challenges" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.735, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Baseline. We build a naive baseline by adopting the recent 3D instance segmentation backbone Mask3D [27] to generate mask proposals. To make the Mask Proposal Module (MPM) fit for the open-vocabulary setting, we remove all components in Mask3D that use the classification labels. Later, PointCLIP [43] is adopted for mask understanding. This naive approach, although satisfying the requirement of 3D inputs only, has long rendering times and unsatisfactory performance (See in Table 7). There are several problems in this baseline model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.131, + 0.784, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.789, + 0.43 + ], + "angle": 0, + "content": "Fig. 4: General Pipeline of OpenIns3D OpenIns3D first processes point clouds with MPM to generate 3D mask proposals and mask scores. The Snap module (detailed in Figure 5) then renders \\(N\\) synthetic scene-level images, which are later passed into the 2D open-world model along with the input text queries. The detection results from the 2D model are stored in the Class Lookup Table (CLT). Finally, both the mask proposals and CLT are fed into the Lookup module, where Mask2Pixel Guided Lookup (detailed in Figure 6) is performed at the global level, followed by a Local Enforced Lookup at the local level to unlock the semantic meaning of mask proposals. The final mask filtering refines the mask proposals and obtains the final results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.789, + 0.488 + ], + "angle": 0, + "content": "Challenge 1: excessive mask proposals. in Mask3D [27], mask proposals are filtered by the mask classification logit, which is removed in the class-agnostic setting. Therefore, an effective mask filtering scheme is needed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.789, + 0.533 + ], + "angle": 0, + "content": "Challenge 2: low quality of 3D instances. 3D instances extracted from scene scans are typically broken, uncompleted, and sparse. Therefore, generated images by simple projection are not easily understood by 2D VL models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.789, + 0.608 + ], + "angle": 0, + "content": "Challenge 3: lack of context information. Humans recognize imperfect 3D point cloud objects through scene comprehension. However, isolated instance point cloud projections lack this context. A solution is to project both the mask and background onto images, yet this introduces distracting elements, potentially confusing classification-level models like CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.789, + 0.656 + ], + "angle": 0, + "content": "Challenge 4: domain gap between projected images and natural images. Rendered images often differ significantly from natural images used in training, posing a challenge for 2D visual language models to comprehend." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.675, + 0.421, + 0.689 + ], + "angle": 0, + "content": "3.2 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.789, + 0.73 + ], + "angle": 0, + "content": "In this section, we present our design of OpenIns3D, which targets the aforementioned four problems. The pipeline of OpenIns3D is shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.75, + 0.513, + 0.765 + ], + "angle": 0, + "content": "Mask: class-agnostic mask proposal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Within the Mask Proposal Module (MPM), we introduce two straightforward designs aimed at filtering low-quality masks generated in the baseline model. Mask scoring. Inspired by [7, 15, 16], We feed the instance queries generated from the Mask Module into a shallow MLP module to predict the quality, i.e" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "IoU of the mask. The predicted IoU \\((IoU_{m})\\) is supervised by the ground truth IoU \\((IoU_{gt})\\) value during the training stages, which is calculated between the predicted mask and its matched ground truth mask in the Bipartite Matching. For unmatched prediction masks \\((IoU_{u})\\), we label the ground truth IoU value as zero. The loss is computed using \\(L2\\). To avoid overly low IoU predictions, a hyper-parameter \\(\\gamma\\) is introduced to reduce the weight of loss for unmatched masks. Therefore, the total loss function for the MPM is:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.267, + 0.785, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma \\sum \\left(I o U _ {u}\\right) ^ {2} + \\sum \\left(I o U _ {m} - I o U _ {\\mathrm {g t}}\\right) ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.785, + 0.466 + ], + "angle": 0, + "content": "Mask filtering. To enhance mask quality, three filters are applied. Firstly, we retain masks with a model-predicted IoU score above a threshold of \\(\\beta\\), ensuring that only high-quality masks are kept. Secondly, drawing inspiration from SAM [16], we focus on stable masks by comparing two binary masks derived from the same underlying soft mask using different threshold values. Specifically, we introduce an offset value \\(\\alpha\\) and select masks where the IoU between the pair of thresholded masks (one with \\(-\\alpha\\) and the other with \\(+\\alpha\\)) exceeds \\(80\\%\\). Lastly, small objects in the scene often lead to invalid proposals, so we filter out mask proposals that have a point number lower than \\(N_{\\mathrm{min}}\\). By employing these techniques, the quantity of mask proposals will decrease, resulting in cleaner and higher-quality masks for subsequent mask understanding tasks (Challenge 1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.491, + 0.591, + 0.507 + ], + "angle": 0, + "content": "Snap: Synthetic Scene-level View Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.784, + 0.566 + ], + "angle": 0, + "content": "Rendering images from points can be a time-consuming task, especially when the number of rendering jobs is high. We propose a multi-scale synthetic scene-level image scheme." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Camera pose selection. The Snap module captures scene-level images at three scales: global, corner, and wide-angle, as shown in Figure 5. For global-level images, cameras are positioned above the scene and point directly toward the centre of the scene. For corner images, cameras are positioned above the centre and point toward the corner, while for the wide-angle images, cameras are positioned at the 3 by 3 grid interaction points, pointing toward the furthest corner. Using the camera position coordinates \\( P_{\\mathrm{cam}} \\), target coordinates \\( P_{\\mathrm{target}} \\), and the up axis of the scene \\( U \\), the Lookat function can be employed to determine the pose matrix Pose. A more detailed mathematical formulation of this is presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Camera intrinsic calibration. Once the camera extrinsic matrix is established, the fields of view for the captured images are adjusted by modifying the camera's intrinsic parameters. The goal is to ensure that the entire scene or a specific part of it is encompassed within the captured images. To achieve this goal, we initialize an arbitrary camera intrinsic matrix and then adjust the focal lengths \\((f_x\\) and \\(f_y)\\) and the principal point coordinates \\((c_x\\) and \\(c_y)\\) through scaling. The scaling is performed by readjusting the projected areas in the image coordinate space. For example: if the projected points of the pre-defined area" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.146, + 0.788, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.44 + ], + "angle": 0, + "content": "Fig.5: Snap and Mask2Pixel Maps. Multiscale snaps are conducted to render images with different levels of detail for scene understanding, including wide-corner snaps, wide-angle snaps, and global snaps. Cameras are positioned on the top of the scene and point towards the centre or corners, and the field of view is determined with the calibrated intrinsic matrix. With the defined camera models, Mask2Pixel maps are built to store the location of each 3D mask in the 2D image (using the same colour to represent 2D-3D correspondences) to guide the search for category names." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.546 + ], + "angle": 0, + "content": "were located in image coordinates within the range of \\([-1000, -192]\\) in the \\(x\\)-domain, our calibrated intrinsic parameters transform this range to \\([0, 1000]\\) in \\(x\\). Importantly, we preserve the aspect ratio between the \\(x\\) and \\(y\\) coordinates to maintain the proportions of the final image without any distortion. This procedure ensures that each captured image is fully utilized and encompasses all regions of interest within the scene (Challenge 2 & 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.637 + ], + "angle": 0, + "content": "Class lookup table. Upon obtaining \\( N \\) synthetic scene-level images, we input them into a 2D open-vocabulary detector. With text queries provided for interested classes, a list of detected objects in synthetic images can be obtained. Subsequently, information about detected objects, including their location and class, are stored in a designated Class Lookup Table (CLT). This table will later be retrieved to allocate class categories to 3D mask proposals (Challenge 4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.659, + 0.61, + 0.675 + ], + "angle": 0, + "content": "Lookup: Mask Classification through Searching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.719 + ], + "angle": 0, + "content": "We conduct a multi-level search to assign category labels to the mask proposals generated from Mask step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.794, + 0.841 + ], + "angle": 0, + "content": "Mask2Pixel guided lookup. We introduce a Mask2Pixel Guided Lookup (MGL) to search within CLT. The concept involves projecting each 3D mask proposal onto a 2D plane using the same camera extrinsic and intrinsic matrices that are utilized to generate the 2D image, as depicted in Figure 5. With knowledge of the precise pixel locations of each mask in images, we can conduct an accurate search through the CLT to identify the most likely class for each mask. The development of MPM takes into account occlusion by integrating depth information. To accomplish the matching, we follow a three-step approach: 1. based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "on the mask's projection onto the 2D plane, we select the best-matched class categories in terms of IoU values; 2. if the IoU value of the best-matched object on the 2D plane is below \\(20\\%\\), the match is disregarded. 3. we aggregate results from multiple views to formulate the final prediction, calculating probability scores using their normalized average IoU values, as illustrated in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.488, + 0.493 + ], + "angle": 0, + "content": "Local enforced lookup. While the Mask2Pixel Guided Lookup assigns class categories to mask proposals, some masks may not correspond to objects in the CLT. To address this, we introduce a Local Enforced Lookup (LEL) approach. We crop out the remaining masks from 2D scene-level images using enlarged bounding boxes and process them with the 2D detector to encourage detection. To select the best views, we introduce an Occlusion Report method to assess occlusion conditions for each mask in each projection, and then choose the top \\(K\\) views for LEL. More details for Occlusion Report can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.487, + 0.584 + ], + "angle": 0, + "content": "Final mask refinement. With the previous lookup approaches, a large proportion of mask proposals obtain a category prediction. All masks that have no category predictions after the MGL and LEL stages are eliminated." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.236, + 0.782, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.489, + 0.788, + 0.559 + ], + "angle": 0, + "content": "Fig. 6: Mask2Pixel Guided Lookup Illustration. IoUs between the 2D detection results and the projected masks are the guidance to assign class names to 3D masks. Multiple image results are ensembled." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.376, + 0.617 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Datasets and evaluation scheme. We tested OpenIns3D on five datasets, including four indoor datasets, namely S3DIS [2], ScanNetv2 [8], ScanNet200 [26], Replica [28], and one outdoor dataset, STPLS3D [5]. Among them, S3DIS, ScanNetv2, and ScanNet200 are indoor point cloud datasets generated from RGB-D images, Replica is a photo-realistic 3D indoor scene reconstruction, while STPLS3D is an aerial photogrammetry-constructed outdoor dataset. We exclusively used the 3D data with colour from these datasets and did not employ any 2D images, poses, or depth maps. Following the settings of prior work [9], we excluded the \"other furniture\" class in ScanNetv2 and the \"clutter\" class in S3DIS due to their vague meanings. Replica and ScanNet200 are evaluated by following the settings of OpenMask3D [30]. For STPLS3D, we merged the low, medium, and high vegetation classes into one \"vegetation\" class and kept all the rest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Implementation details. For the S3DIS, ScanNetv2, Scannet200, and STPLS datasets, the MPM module is trained without utilizing any category labels, and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Zero-shot object classification on ScanNetv2. OpenIns3D's Snap and Lookup approach for mask classification, surpasses all previous methods, including the latest language-aligned large-scale 3D foundation model [40]." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.209, + 0.785, + 0.31 + ], + "angle": 0, + "content": "
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointCLIP [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
PointCLIP V2 [37]11.00.00.023.80.00.00.00.07.80.090.70.00.00.00.064.40.00.0
CLIP2Point [14]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP2[35]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
Uni3D [40]45.858.53.778.883.754.931.339.470.135.11.927.394.213.838.710.788.147.6
OpenIns3D60.885.227.487.677.346.954.864.271.49.980.882.771.661.438.70.087.985.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.784, + 0.443 + ], + "angle": 0, + "content": "\\(\\lambda\\) is set to 0.1 to reduce the weight of zero-IoU. The mask proposal of Replica is trained on ScanNet200, followed by OpenMask3D [30]. The Snap module captures images with a size of \\(1000 \\times 1000\\) including 16 global snaps, 4 corner snaps, and 4 wide-angle snaps. The top \\(0.5\\mathrm{m}\\) of the scene is removed for S3DIS, as the rooms are enclosed. For STPLIS3D, we followed Mask3D to split the large outdoor scene into patches of \\(50m \\times 50m\\) and lifted the camera up to \\(10m\\). More implementation details are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Open-vocabulary point cloud recognition. We first evaluate OpenIns3D performance among all existing 3D open-vocabulary models with the same setting, i.e. using only 3D inputs. These models are most commonly tested on recognition tasks, including PointCLIP(v1&v2) [43] [37], Clip2Point [14], CLIP² [35]. We also compare with large-scale 3D foundation models, such as Uni3D [40], which has 1-billion parameters, and was trained with large-scale 3D shapes and image-text pairs. We followed their evaluation scheme and reported the Top-1 accuracy of instance classification on ScanNetv2. The results are presented in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.785, + 0.748 + ], + "angle": 0, + "content": "Open-vocabulary instance segmentation. We adopted various comparison schemes to align with existing methods. For 3D Open-vocabulary Instance Segmentation, we compared with PLA [9], and its follow-up works RegionPLC [34] and Lowis3D [10]. For a fair comparison, we followed their category splits and compared our results on novel classes, as demonstrated in Table 2. For STPLS3D, we compared OpenIns3D with baseline models whose classification module is PointCLIP and PointCLIPV2 [43] (Table 5). We also explored the performance of OpenIns3D on a more challenging dataset with more class categories. Specifically, we compared the performance of OpenIns3D with OpenMask3D, OpenScene, on Replica (Table 4) as well as ScanNet200 (Table 6). Following OpenMask3D, we used Mask3D to generate mask proposals for OpenScene for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Open-vocabulary object detection. Since there were limited works on 3D open-vocabulary instance segmentation at the time of conducting this work, we also selected some of the latest methods in the 3D open-world object detection domain for a more comprehensive evaluation. The evaluation is performed by converting generated masks into axis-aligned bounding boxes, and the results are shown in Table 3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: 3D Open-vocabulary Instance Segmentation Results on S3DIS and ScanNetv2. We compare our zero-shot performance on the novel categories defined in the PLA-family work. Significant improvements are achieved on the S3DIS dataset, and competitive results are observed on ScanNetv2 (B/N: Base/Novel)." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.213, + 0.778, + 0.44 + ], + "angle": 0, + "content": "
OVISS3DISScanNetv2
MethodB/NAP50AP25B/NAP50AP25require 2D
PLA [9]8/408.6-10/721.9-✓
RegionPLC [34]8/4--10/732.3-✓
Lowis3D [10]8/413.8-10/731.2-✓
Open3DIS [22]8/426.3-10/7--✓
Mask3d+PointClip [43]- / 405.410.3- / 704.507.8X
OpenIns3D- / 437.039.3- / 727.942.6X
improvement- / 4(+10.7)(+29.0)- / 7-(+34.8)X
PLA [9]6/609.8-8/925.1-✓
RegionPLC [34]6/6--8/932.2-✓
Lowis3D [10]6/615.8-8/938.1-✓
Open3DIS [22]6/629.0-8/9--✓
Mask3d+PointClip [43]- / 608.510.6- / 905.606.7X
OpenIns3D- / 633.038.9- / 919.527.9X
improvement- / 6(+4.0)(+28.3)- / 9-(+21.2)X
Mask3d+PointClip [43]- / 1208.609.3- / 1704.514.4X
OpenIns3D- / 1228.329.5- / 1728.738.9X
improvement- / 12(+19.7)(+20.2)- / 17(+24.2)(+24.5)X
" + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.444, + 0.776, + 0.458 + ], + "angle": 0, + "content": "Table 3: Open-vocabulary Object Detection \\((AP_{25})\\) on unseen classes in ScanNet." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.459, + 0.796, + 0.581 + ], + "angle": 0, + "content": "
Methodsmeantoiletbedchairsofadressertablecabinetbookshelfpillowsink
OV-PointCLIP [37]3.16.62.36.33.90.77.20.72.10.60.8
OV-Image2Point [32]0.80.20.81.01.40.22.81.00.90.00.1
Detic-ModelNet [41]1.74.21.04.61.20.23.20.61.20.00.7
Detic-ImageNet [41]0.40.00.00.20.00.51.80.50.30.00.7
OV-3DETIC [20]12.749.02.67.318.62.814.32.44.53.921.1
L3Det [42]24.656.336.216.123.08.123.114.717.323.427.9
FM-OV3D [36]21.555.038.819.241.923.83.50.46.017.48.8
OpenIns3D43.779.570.576.915.80.053.140.141.27.153.1
improvement(+19.1)(+23.2)(+31.7)(+57.7)--(+30)(+25.4)(+23.9)(+25.2)
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.475, + 0.626 + ], + "angle": 0, + "content": "5 Results and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.46, + 0.66 + ], + "angle": 0, + "content": "5.1 Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "First of all, OpenIns3D demonstrates impressive performance in open-vocabulary point cloud recognition, surpassing all previous methods, including the large-scale 3D foundation model by \\(15\\%\\). This proves the effectiveness of the zero-shot \"Snap\" and \"lookup\" scheme. With the enhanced recognition capability, the performance of 3D open-vocabulary Object Detection among the ScanNet dataset has also achieved state-of-the-art results by a large margin. For 3D instance segmentation, compared to works in the PLA family [9, 10, 34] and the latest work Open3DIS [22], OpenIns3D does not require aligned images as input, (still need RGB information for points), achieves higher results on the S3DIS dataset, both in the 4 novel categories split and the 6 novel categories split. In SPTLS3D, OpenIns3D outperforms the baseline model PointCLIPV2 by \\(9.3\\%\\) in AP. On" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.532, + 0.172 + ], + "angle": 0, + "content": "Table 4: OVIS on the indoor Replica dataset." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.538, + 0.303 + ], + "angle": 0, + "content": "
Model2DAP\\( AP_{50} \\)\\( AP_{25} \\)
OpenScene [23] (2D Fusion)✓10.915.617.3
OpenScene [23] (2D/3D Ens.)✓8.210.413.3
OpenMask3D [30]✓13.118.424.2
OpenScene [23] (3D Distill)✗8.210.512.6
OpenIns3D✗13.618.019.7
improvement✗(+5.4)(+7.5)(+7.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.145, + 0.771, + 0.172 + ], + "angle": 0, + "content": "Table 5: OVIS on outdoor STPLS3D dataset" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.185, + 0.777, + 0.303 + ], + "angle": 0, + "content": "
ModelAPAP50AP25
PointCLIP [43]02.002.604.0
PointCLIPv2 [37]02.103.105.2
OpenIns3D11.414.217.2
improvement(+9.3)(+11.9)(+12.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.319, + 0.787, + 0.374 + ], + "angle": 0, + "content": "Table 6: 3D instance segmentation results on the ScanNet200 validation set. OpenIns3D demonstrates robust performance when compared to 2D-input-free models. However, notable limitations emerge when dealing with small objects in the common and tail classes." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.387, + 0.78, + 0.503 + ], + "angle": 0, + "content": "
Modeluse 2DAPheadAPcommonAPtailAPAP50AP25
OpenScene (2D Fusion) [23]✓13.411.69.911.715.217.8
OpenScene (2D/3D Ens.) [23]✓11.03.21.15.36.78.1
OpenMask3D✓17.114.114.915.419.923.1
OpenScene (3D Distill)✗10.62.60.74.86.27.2
OpenIns3D✗16.06.54.28.810.314.4
improvement✗(+5.4)(+3.9)(+3.6)(+4.0)(+4.1)(+7.2)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.653 + ], + "angle": 0, + "content": "the Replica dataset, OpenIns3D even outperformed OpenMask3D, which relies on well-aligned images for mask understanding. In the case of ScanNet200, OpenIns3D attains the highest performance compared to all other 3D input baselines. However, we have observed a decline in performance on tail and common classes within ScanNet200. The decrease in performance can be attributed to the low-quality and unclear reconstruction of smaller objects in the ScanNet200 scene. This is a noticeable limitation of OpenIns3D, but for better reconstructions, such as in Replica, and for Head classes within ScanNet200, OpenIns3D still demonstrates decent performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.715 + ], + "angle": 0, + "content": "In summary, OpenIns3D demonstrates the best performance among all existing methods if only 3D data is used as input and outperforms many existing state-of-the-art methods that require 2D images. It also shows certain limitations on small objects that are not well-reconstructed in 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.388, + 0.754 + ], + "angle": 0, + "content": "5.2 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Mask quality ablation. Following the evaluation on ScanNetv2, we assessed the class-agnostic mask quality using the average precision (AP) score. We treated all classes as universal since the predictions are class-agnostic. The evaluation was conducted on the ScanNetv2 validation set. Table 8 demonstrates the effectiveness of the Mask Scoring and Mask Filtering designs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 7: Rendering and Inference Time Ablations. Results tested on typical ScanNet scenes with 50 masks. OpenIns3D requires less rendering and inference time." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.787, + 0.304 + ], + "angle": 0, + "content": "
RenderingNum of Img needed2D backboneImg size (w × h)\\( T_{render} \\) (s/scene)\\( T_{infer} \\) (s/scene)\\( T_{total} \\) (s/scene)AP25 (%)
PointCLIP250CLIP\\( 128^2 \\)5.215.320.59.3
LAR250CLIP\\( 128^2 \\)14.318.733.010.5
Mask rendering250CLIP\\( 128^2 \\)42.619.562.17.3
OpenIns3D (ours)8G-DINO\\( 1000^2 \\)2.36.28.529.8
OpenIns3D (ours)8ODISE\\( 1000^2 \\)2.38.210.535.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.316, + 0.53, + 0.344 + ], + "angle": 0, + "content": "Table 8: MPM Ablation. MS: Mask scoring. MF: Mask filtering." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.356, + 0.526, + 0.448 + ], + "angle": 0, + "content": "
MethodAP50AP25
Mask3d-Supervised74.780.9
CA-Mask3d47.549.2
CA-Mask3d + MS50.2 (+02.7)53.3 (+04.1)
CA-Mask3d + MF61.6 (+14.1)71.0 (+21.8)
CA-Mask3d + MS + MF64.6 (+17.0)73.4 (+24.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.316, + 0.799, + 0.345 + ], + "angle": 0, + "content": "Table 9: Number of Views Abla-tion. LEL: Local Enforced Lookup" + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.357, + 0.791, + 0.446 + ], + "angle": 0, + "content": "
IDX4816LELAP50AP25
i✓18.327.1
ii✓22.735.1
iii✓24.837.5
iv✓✓28.738.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.785, + 0.514 + ], + "angle": 0, + "content": "Multi-view ablation. We also studied the effects of using different numbers of views (Table 9). Increasing the number of views used in the Lookup module leads to better results. Additionally, Look Enforced Lookup provided a final improvement to the results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.785, + 0.653 + ], + "angle": 0, + "content": "Projection and 2D backbone ablation. We conducted a comprehensive study on various rendering methods and their interaction with the 2D backbone to identify a suitable approach. We report the rendering time and inference performance for each method (Table 7), and more details can be found in the supplementary materials. The key observation is that the scene-level rendering and understanding approach excels in speed while also demonstrating strong performance. Switching from Grounding Dino [19] to the latest ODISE [33] also brings gains in performance, indicating that the OpenIns3D framework can easily benefit from the rapid development of 2D open-world detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.785, + 0.761 + ], + "angle": 0, + "content": "Cross-domain analysis. To evaluate the generalization capability of MPM across different domains, we trained and tested OpenIns3D on two different datasets, as shown in Table 10. The cross-domain models also demonstrate impressive performance on both datasets when compared with the baseline. Notably, within the 17 classes of ScanNetv2, 11 classes do not exist in S3DIS. OpenIns3D, trained on S3DIS, still achieves decent performance among these unseen classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Free-flow language capability. The snap and lookup scheme outsources the mask understanding tasks to a 2D vision-language model. Therefore when integrated with a 2D model powered by large language models (LLMs) like LISA [18], OpenIns3D can perform reasoning-based segmentation tasks (as depicted in Figure 7). For instance, when given the query \"pencilling down ideas during brain-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "storming,\" OpenIns3D accurately segments the whiteboard, while for \"furniture offers recreational enjoyment with friends,\" it precisely identifies and segments the pool table." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.209, + 0.777, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 7: Qualitative Results from OpenIns3D. OpenIns3D (ODISE) demonstrates the ability to manage a versatile vocabulary. OpenIns3D (LISA) can conduct 3D reasoning segmentation." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Table 10: Cross-domain Ablation. We trained and tested OpenIns3D on two different datasets to examine its cross-domain capability. While S3DIS and ScanNetV2 have non-overlapping classes, OpenIns3D demonstrates decent generalization capability." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.52, + 0.76, + 0.61 + ], + "angle": 0, + "content": "
TestModelTraining DataAP50AP25
ScanNetv2 [8]Mask3D-Pointclip [43]ScanNetv204.514.4
OpenIns3DScanNetv228.738.9
OpenIns3DS3DIS21.533.6
S3DIS [1]Mask3D-Pointclip [43]S3DIS03.506.8
OpenIns3DS3DIS28.329.5
OpenIns3DScanNetv214.219.8
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.36, + 0.648 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Achieving 3D open-vocabulary scene understanding is a challenging task, primarily due to the lack of extensive 3D-text data. Currently, most work in this domain focuses on using 2D images to bridge the gap between 3D and language. This, however, not only requires a good alignment between 2D and 3D but also evolves slowly due to the significant effort needed for retraining when changing the 2D backbone. OpenIns3D introduces a new pipeline, i.e. Mask-Snap-Lookup, for this task. The Mask module generates authentic masks in the 3D domain, while Snap renders scene-level images in 2D domains, and the Lookup module links the results from 2D to 3D precisely. This pipeline requires no image input, achieves better performance, and can evolve seamlessly with 2D models without training. We hope our work will provide a fresh perspective for researchers working towards open-world 3D scene understanding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.298 + ], + "angle": 0, + "content": "This work is supported by the Girton College Graduate Research Awards at the University of Cambridge, School of Technology, National Highways sponsored through EPSRC Centre for Doctoral Training, the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR, the National Natural Science Foundation of China No. 62201484, HKU Startup Fund, and HKU Seed Fund for Basic Research. We would like to express our gratitude to Yunhan Yang for his assistance in exploring rendering techniques and to Chengyao Wang for sharing his implementation of Mask3D." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.323, + 0.323, + 0.338 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.381 + ], + "angle": 0, + "content": "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.41 + ], + "angle": 0, + "content": "2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.785, + 0.437 + ], + "angle": 0, + "content": "3. Bakr, E.M., Alsaedy, Y.Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: NeurIPS (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.52 + ], + "angle": 0, + "content": "5. Chen, M., Hu, Q., Yu, Z., THOMAS, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. In: BMVA (2022) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.562 + ], + "angle": 0, + "content": "6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.59 + ], + "angle": 0, + "content": "7. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. In: ICCV (2023) 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.618 + ], + "angle": 0, + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 2, 4, 5, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.686 + ], + "angle": 0, + "content": "0. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. In: TPAMI (2024) 2, 4, 5, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.715 + ], + "angle": 0, + "content": "1. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: ICML (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "2. Griffiths, D., Boehm, J.: SynthCity: A large-scale synthetic point cloud. In: arXiv (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.798 + ], + "angle": 0, + "content": "3. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: Semantic3D.net: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV (2023) 4, 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z.Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Huang, Z., Huang, L., Gong, Y., Huang, C., Wang, X.: Mask scoring R-CNN. In: CVPR (2019) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "17. Kundu, A., Yin, X., Fathi, A., Ross, D.A., Brewington, B., Funkhouser, T.A., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: ECCV (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.289 + ], + "angle": 0, + "content": "18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024) 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.331 + ], + "angle": 0, + "content": "19. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.373 + ], + "angle": 0, + "content": "20. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR (2023) 3, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.785, + 0.416 + ], + "angle": 0, + "content": "21. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.459 + ], + "angle": 0, + "content": "22. Nguyen, P.D.A., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024) 2, 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "23. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.543 + ], + "angle": 0, + "content": "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.586 + ], + "angle": 0, + "content": "25. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-lille-3d: A large and high-quality ground-truth urban point cloud dataset for automatic segmentation and classification. In: The International booktitle of Robotics Research (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.614 + ], + "angle": 0, + "content": "26. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.643 + ], + "angle": 0, + "content": "27. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.727 + ], + "angle": 0, + "content": "28. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces. In: arXiv (2019) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "29. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.G.: Multi-view convolutional neural networks for 3d shape recognition. In: ICCV (2015) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.797 + ], + "angle": 0, + "content": "30. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023) 3, 4, 5, 9, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.798, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Tan, W., Qin, N., Ma, L., Li, Y., Du, J., Cai, G., Yang, K., Li, J.: Toronto-3D: A large-scale mobile lidar dataset for semantic segmentation of urban roadways. In: CVPRW (2020) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Xu, C., Yang, S., Galanti, T., Wu, B., Yue, X., Zhai, B., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: Image2point: 3d point-cloud understanding with 2d image pretrained models. In: ECCV (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.23 + ], + "angle": 0, + "content": "33. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In: CVPR (2023) 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.271 + ], + "angle": 0, + "content": "34. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024) 2, 4, 5, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.313 + ], + "angle": 0, + "content": "35. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: \\(\\mathrm{Clip}^2\\): Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR (2023) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.355 + ], + "angle": 0, + "content": "36. Zhang, D., Li, C., Zhang, R., Xie, S., Xue, W., Xie, X., Zhang, S.: Fm-ov3d: Foundation model-based cross-modal knowledge blending for open-vocabulary 3d detection. In: AAAI (2024) 2, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.395 + ], + "angle": 0, + "content": "37. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by CLIP. In: CVPR (2022) 3, 4, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.424 + ], + "angle": 0, + "content": "38. Zhou, C., Loy, C.C., Dai, B.: Denseclip: Extract free dense labels from clip. In: CVPR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.452 + ], + "angle": 0, + "content": "39. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.479 + ], + "angle": 0, + "content": "40. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024) 2, 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.508 + ], + "angle": 0, + "content": "41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.536 + ], + "angle": 0, + "content": "42. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. In: arXiv (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.577 + ], + "angle": 0, + "content": "43. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV (2023) 2, 3, 4, 5, 10, 11, 12, 14" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_origin.pdf b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d8a0a21d3d7dcaa56e082cdb7d17490e6f121914 --- /dev/null +++ b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/f206041e-1c46-4539-919a-ce8570607515_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a899a4fba4679b7f91074cbb63cc3531da4c3b35ea414da1436e8f201276cc2 +size 7406385 diff --git a/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/full.md b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5506ab19e7fbfcb002c9f8ffd2120520a9f7e092 --- /dev/null +++ b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/full.md @@ -0,0 +1,286 @@ +# Opens3D: Snap and Lookup for 3D Open-vocabulary Instance Segmentation + +Zhening Huang1 + +Xiaoyang $\mathbf{W u}^{2}$ + +Xi Chen $^{2}$ + +Hengshuang Zhao\* + +Lei Zhu $^{3,4}$ + +Joan Lasenby + +1University of Cambridge 2The University of Hong Kong + +$^{3}$ Hong Kong University of Science and Technology + +4The Hong Kong University of Science and Technology (Guangzhou) + +https://zheninghuang.github.io/OpenIns3D + +![](images/fc2489a3e5f401235117350154fd6d85d7b0ac99b092789ef21aad86adc82d8b.jpg) +"furniture that is capable of producing music." + +![](images/b0fe4c32c28f73c06a571d0f37b7b6051fc3628edb9df459fe3726dd327c647a.jpg) +device to watch BBC news +Fig. 1: Complex Queries 3D Instance Segmentation with OpenIns3D. + +![](images/bd1a3f2ed6ae89ed11399adf6ff3197706b7ba264a727977b35d7b7b595e7ef1.jpg) +"Ma Long's domain of excellence + +Abstract. In this work, we introduce OpenIns3D, a new 3D-input-only framework for 3D open-vocabulary scene understanding. The OpenIns3D framework employs a "Mask-Snap-Lookup" scheme. The "Mask" module learns class-agnostic mask proposals in 3D point clouds, the "Snap" module generates synthetic scene-level images at multiple scales and leverages 2D vision-language models to extract interesting objects, and the "Lookup" module searches through the outcomes of "Snap" to assign category names to the proposed masks. This approach yet simple, achieves state-of-the-art performance across a wide range of 3D open-vocabulary tasks, including recognition, object detection, and instance segmentation, on both indoor and outdoor datasets. Moreover, OpenIns3D facilitates effortless switching between different 2D detectors without requiring retraining. When integrated with powerful 2D open-world models, it achieves excellent results in scene understanding tasks. Furthermore, when combined with LLM-powered 2D models, OpenIns3D exhibits an impressive capability to comprehend and process highly complex text queries that demand intricate reasoning and real-world knowledge. + +Keywords: Open-Vocabulary Understanding $\cdot$ 3D Scene Understanding $\cdot$ Vision-Language Model + +# 1 Introduction + +3D scene understanding plays a critical role in various domains, such as autonomous driving, robotic sensing, AR/VR, and manufacturing, among others. + +![](images/57840a1969993fc1483db598a8b507bc99a38ca0dc2030bd1f620f9f98a0388b.jpg) +Fig. 2: High-level Illustrations of OpenIns3D and Quantitative Results. (a) OpenIns3D follows the "Mask-Snap-Lookup" steps for open-vocabulary scene understanding. (b) A list of SOTA results has been achieved on both indoor and outdoor datasets. OV-Rec: open-vocabulary object recognition. OVOD: open-vocabulary object detection. OVIS: open-vocabulary instance segmentation. PointCLIPV2 [43]; Uni3D [40]; Open3DIS [22]; FM-OV3D [36] + +![](images/8b2f93d3e12882c7e597be4a8a9716fc2eae35b6904cd515dba5b847a408e329.jpg) + +While the development of 3D closed-set understanding is relatively mature, scene understanding in an open-vocabulary setting is still in its infancy. Closed-set understanding can only handle a predefined set of concepts and scenarios but fails to provide valid responses when faced with unfamiliar concepts or variations in language usage. This limitation impacts its performance in dynamic and ever-changing contexts. + +Thanks to internet-scale image-text datasets, significant progress has been made in 2D image open-vocabulary understanding [4, 7, 11, 16, 24, 33, 38, 39]. However, unlike 2D data that can be easily collected from the internet, constructing a large-scale 3D-text dataset poses a challenge. As a result, the most viable approach to achieving 3D open-vocabulary understanding involves leveraging 2D images to bridge language and 3D data. In this direction, there have been several notable works, such as OpenScene [23], PLA-family [9, 10, 34], and CLIP2Scene [6]. These works leverage well-aligned 2D images and 3D point clouds to conduct feature distillation or employ 2D caption models to construct 3D-text pairs. One prerequisite of these methods, however, is the availability of well-aligned 2D images and 3D point clouds. This means that posed 2D images, associated depth maps and camera models need to be accessible as inputs to the network. In real-life scenarios, there are numerous cases where 2D images or the information required to align 2D and 3D data are unavailable. For instance, to save storage space, point clouds generated with LiDAR are often stored without accompanying 2D images (example datasets include [13, 25, 31]). In cases where point clouds are obtained from the registration of multiple scans from different sensors or are converted from 3D simulations/CAD models [12, 21], 2D images are often non-existent. + +We believe that developing a 3D open-vocabulary framework without relying on well-aligned 2D images is meaningful, as this will simplify deployment pre + +requisites and enhance its applicability across a wide range of scenarios. To this end, we introduce OpenIns3D, a framework designed to effectively perform 3D open-vocabulary scene understanding tasks without relying on 2D aligned images. Overall, OpenIns3D comprises three core steps: Mask, Snap, and Lookup. An overall illustration of OpenIns3D is presented in Figure 2a. + +Mask: Given a 3D point cloud, the first part of OpenIns3D learns class-agnostic mask proposals with a Mask Proposal Module (MPM). This process is trained without any classification labels. To control the quality of the mask, MPM proposes a learnable Mask Scoring module to predict the quality of each mask output and implements a list of Mask Filtering techniques to discard invalid, low-quality masks. MPM outputs a list of class-agnostic masks in the scene. + +Snap: Multiple synthetic scene-level images are generated with calibrated and optimized camera poses and intrinsic parameters. These images are specifically designed to encompass part or all of the relevant masks, aiming to minimize the need for multiple renderings. Instead of individually predicting the category of each mask proposal [20, 37, 43], the scene-level images are input into 2D open-vocabulary models for the simultaneous understanding of all interesting objects present in the scene. A Class Lookup Table (CLT) is then constructed to store all the detected object categories alongside their respective pixel locations. + +Lookup: To precisely determine the positions of mask proposals in each image, Mask2Pixel maps are constructed. These maps project all 3D mask proposals onto 2D images with identical camera parameters used in Snap. In the Lookup phase, OpenIns3D searches through the CLT with the help of Mask2Pixel maps to precisely assign category names to 3D mask proposals. Results from multiple views are combined to establish initial mask classification outcomes. For remaining masks, a similar Lookup procedure is carried out on a local scale to facilitate classification. Lastly, the 3D mask proposals are refined by removing masks lacking class assignments after both global and local Lookup. + +OpenIns3D demonstrates promising performance in extensive comparisons with other methods, as summarized in Figure 2b. This simple and flexible approach achieves a range of state-of-the-art (SOTA) results on various 3D open-vocabulary tasks. Specifically, SOTA results are achieved on open-vocabulary instance segmentation for the indoor S3DIS [1] dataset and the outdoor STPLS3D [5] dataset. OpenIns3D even outperforms OpenMask3D [30], a concurrent work that heavily utilizes 2D images, for instance, segmentation on the challenging Replica [28] dataset. We also find that the "Snap and Lookup" serves as a powerful open-vocabulary object recognition engine, which achieved SOTA object recognition tasks on ScanNet [8], outperforming the previous SOTA, a 1-billion parameter 3D foundation model [40]. Lastly, when converting mask proposals into 3D bounding boxes, OpenIns3D also achieved state-of-the-art results in open-vocabulary object detection (OVOD) on ScanNet, outperforming previous image-dependent methods [36]. The design of OpenIns3D also allows 2D detectors to be changed without the need for retraining. This provides the model with the capability to evolve alongside the latest development of the 2D open-vocabulary models. Moreover, when 2D detectors are coupled with Large Lan- + +![](images/56e04e05cf354c1035c4332cae16288847e2d3ee8f0c2b3747bd7bf4f8ab3c8e.jpg) +(a) 3D Feature Distillation + +![](images/5700f25ca1d75f620c5e91366bded4bc61d16fca6ad8c22ebb93857c706f28a5.jpg) +(b) Building 3D-text pair + +![](images/0ea3b54d16ccd3fb51b8ca573a330619e299eeda3b24ec51d7d245f1cfdae1ed.jpg) +(c) CLIP and Projection +Fig. 3: Four Categories of Open-Vocabulary 3D Scene Understanding Models. a) 3D feature distillation frameworks, where 2D images are used as a bridge to distil language-aligned features into 3D, with typical works including OpenScene [23] and Clip2Scene [6]. b) Building 3D-text pairs, where 2D captioning models are used to build 3D-text pairs for feature learning, with typical works including the PLA-family [9,10,34] c) CLIP and Projection, where objects are cropped out of 2D images before being processed by CLIP, and the results are directly projected into 3D, including OpenMask3D [30], OV-3DET [20], $\mathrm{CLIP}^2$ [35] and Open3DIS [22]. d) OpenIns3D + +![](images/f02f5869c5d4c0c810eba53bf727f6dedfc9597e07fb45b5b818b7436e848ba4.jpg) +(d) Mask-Snap-Lookup (ours) + +guage Models (LLMs), OpenIns3D can enable complex query understanding capability. When integrated with LISA [18], an LLM-powered reasoning segmentation model, OpenIns3D exhibits a strong ability to comprehend highly intricate language queries and performs reasoning segmentation in 3D, as illustrated in Figure 1. In summary, our contributions are: + +- OpenIns3D employs a distinct pipeline that operates without the need for well-aligned images. This approach achieves state-of-the-art results across a range of benchmarks and possesses the ability to comprehend highly complex input queries. +- The proposed "Snap and Lookup" combination can serve as a powerful 3D object recognition engine, especially for noisy 3D objects extracted from scene-level scans. This capability is demonstrated by achieving state-of-the-art results with a large margin. + +# 2 Related work + +3D open-vocabulary understanding. Progress in 3D open-vocabulary understanding has been relatively slow compared to that of images. In the domain of 3D object classification tasks, methods like PointCLIP [37], PointCLIPV2 [43], and CLIP2Point [14] project 3D point clouds into depth maps and link them with 2D models for classification. However, these methods lack performance in scene-level understanding, where points are often overlapped and incomplete. For scene-level understanding, most work has primarily focused on leveraging well-aligned 2D posed images, depth maps, and point clouds [6,9,10,23,26,34,35]. One notable example is OpenScene [23], which takes posed 2D images, depth maps, and 3D data as input, and feature distillation is performed to transfer 2D language-aligned features from images to 3D point clouds. Similarly, + +Clip2Scene [6] builds dense pixel-point pairs by calibrating the LiDAR point cloud with corresponding images captured by six cameras. However, achieving instance-level understanding is challenging with these methods as they focus solely on semantic-level understanding. In contrast, PLA [9] and its follow-up work RegionPLC [34] and Lowis3D [10] utilize a 2D caption model to construct 3D-text pairs to learn features. However, the PLA-family works rely on a binary head to classify the input object into base categories or novel categories, and the transferability of this binary head to different base-novel splits is very limited, posing a challenge for flexible applications. One current work, OpenMask3D [30], utilizes well-aligned 2D images to learn features for mask proposals, leading to impressive results in open-vocabulary instance segmentation. Open3DIS [22] follows the same procedure, enhancing both mask proposals and mask classification with images, yielding better performance. A common issue with these methods is their reliance on well-aligned 3D and 2D pairs in the input, which may not always be available in real applications. We summarise the key difference between these methods in Figure 3. Simplifying input requirements improves method flexibility and compatibility. In this work, we explore how to conduct 3D open-vocabulary understanding without relying on 2D images. + +Image generation from 3D. Projection-based methods have been extensively explored in the past for 3D understanding and have proven to be beneficial for obtaining complementary features. For instance, MVCNN [29] projects 3D objects to different views to aid in feature learning, while LAR [3] introduces object centre projection methods to generate images for 3D objects from various angles, assisting visual grounding tasks. Additionally, Virtual View Fusion [17] employs the original camera pose but enlarges the field of view, resulting in enhanced 2D feature transfer. However, these methods encounter challenges like best view selection, object occlusion, information loss during projection, and long rendering times. In the context of open-vocabulary settings, the quality of the projected image plays a crucial role in model performance. In our work, we evaluate different projection methods, along with their compatibility with 2D open-vocabulary models, to identify an optimal solution that achieves good results and is efficient to implement. + +# 3 OpenIns3D + +# 3.1 Baseline and Challenges + +Baseline. We build a naive baseline by adopting the recent 3D instance segmentation backbone Mask3D [27] to generate mask proposals. To make the Mask Proposal Module (MPM) fit for the open-vocabulary setting, we remove all components in Mask3D that use the classification labels. Later, PointCLIP [43] is adopted for mask understanding. This naive approach, although satisfying the requirement of 3D inputs only, has long rendering times and unsatisfactory performance (See in Table 7). There are several problems in this baseline model. + +![](images/8d1b9b92f820f6790837997a5d5d46ee43c3e3a0409ae66129bfb0c616b33ab2.jpg) +Fig. 4: General Pipeline of OpenIns3D OpenIns3D first processes point clouds with MPM to generate 3D mask proposals and mask scores. The Snap module (detailed in Figure 5) then renders $N$ synthetic scene-level images, which are later passed into the 2D open-world model along with the input text queries. The detection results from the 2D model are stored in the Class Lookup Table (CLT). Finally, both the mask proposals and CLT are fed into the Lookup module, where Mask2Pixel Guided Lookup (detailed in Figure 6) is performed at the global level, followed by a Local Enforced Lookup at the local level to unlock the semantic meaning of mask proposals. The final mask filtering refines the mask proposals and obtains the final results. + +Challenge 1: excessive mask proposals. in Mask3D [27], mask proposals are filtered by the mask classification logit, which is removed in the class-agnostic setting. Therefore, an effective mask filtering scheme is needed. + +Challenge 2: low quality of 3D instances. 3D instances extracted from scene scans are typically broken, uncompleted, and sparse. Therefore, generated images by simple projection are not easily understood by 2D VL models. + +Challenge 3: lack of context information. Humans recognize imperfect 3D point cloud objects through scene comprehension. However, isolated instance point cloud projections lack this context. A solution is to project both the mask and background onto images, yet this introduces distracting elements, potentially confusing classification-level models like CLIP. + +Challenge 4: domain gap between projected images and natural images. Rendered images often differ significantly from natural images used in training, posing a challenge for 2D visual language models to comprehend. + +# 3.2 Overall Framework + +In this section, we present our design of OpenIns3D, which targets the aforementioned four problems. The pipeline of OpenIns3D is shown in Figure 4. + +# Mask: class-agnostic mask proposal + +Within the Mask Proposal Module (MPM), we introduce two straightforward designs aimed at filtering low-quality masks generated in the baseline model. Mask scoring. Inspired by [7, 15, 16], We feed the instance queries generated from the Mask Module into a shallow MLP module to predict the quality, i.e + +IoU of the mask. The predicted IoU $(IoU_{m})$ is supervised by the ground truth IoU $(IoU_{gt})$ value during the training stages, which is calculated between the predicted mask and its matched ground truth mask in the Bipartite Matching. For unmatched prediction masks $(IoU_{u})$ , we label the ground truth IoU value as zero. The loss is computed using $L2$ . To avoid overly low IoU predictions, a hyper-parameter $\gamma$ is introduced to reduce the weight of loss for unmatched masks. Therefore, the total loss function for the MPM is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \gamma \sum \left(I o U _ {u}\right) ^ {2} + \sum \left(I o U _ {m} - I o U _ {\mathrm {g t}}\right) ^ {2} \tag {1} +$$ + +Mask filtering. To enhance mask quality, three filters are applied. Firstly, we retain masks with a model-predicted IoU score above a threshold of $\beta$ , ensuring that only high-quality masks are kept. Secondly, drawing inspiration from SAM [16], we focus on stable masks by comparing two binary masks derived from the same underlying soft mask using different threshold values. Specifically, we introduce an offset value $\alpha$ and select masks where the IoU between the pair of thresholded masks (one with $-\alpha$ and the other with $+\alpha$ ) exceeds $80\%$ . Lastly, small objects in the scene often lead to invalid proposals, so we filter out mask proposals that have a point number lower than $N_{\mathrm{min}}$ . By employing these techniques, the quantity of mask proposals will decrease, resulting in cleaner and higher-quality masks for subsequent mask understanding tasks (Challenge 1). + +# Snap: Synthetic Scene-level View Generation + +Rendering images from points can be a time-consuming task, especially when the number of rendering jobs is high. We propose a multi-scale synthetic scene-level image scheme. + +Camera pose selection. The Snap module captures scene-level images at three scales: global, corner, and wide-angle, as shown in Figure 5. For global-level images, cameras are positioned above the scene and point directly toward the centre of the scene. For corner images, cameras are positioned above the centre and point toward the corner, while for the wide-angle images, cameras are positioned at the 3 by 3 grid interaction points, pointing toward the furthest corner. Using the camera position coordinates $P_{\mathrm{cam}}$ , target coordinates $P_{\mathrm{target}}$ , and the up axis of the scene $U$ , the Lookat function can be employed to determine the pose matrix Pose. A more detailed mathematical formulation of this is presented in the supplementary materials. + +Camera intrinsic calibration. Once the camera extrinsic matrix is established, the fields of view for the captured images are adjusted by modifying the camera's intrinsic parameters. The goal is to ensure that the entire scene or a specific part of it is encompassed within the captured images. To achieve this goal, we initialize an arbitrary camera intrinsic matrix and then adjust the focal lengths $(f_x$ and $f_y)$ and the principal point coordinates $(c_x$ and $c_y)$ through scaling. The scaling is performed by readjusting the projected areas in the image coordinate space. For example: if the projected points of the pre-defined area + +![](images/af8221a65488d1b74b92faed1c87ef1836ad9ca7513063a556a1967303d0672f.jpg) +Fig.5: Snap and Mask2Pixel Maps. Multiscale snaps are conducted to render images with different levels of detail for scene understanding, including wide-corner snaps, wide-angle snaps, and global snaps. Cameras are positioned on the top of the scene and point towards the centre or corners, and the field of view is determined with the calibrated intrinsic matrix. With the defined camera models, Mask2Pixel maps are built to store the location of each 3D mask in the 2D image (using the same colour to represent 2D-3D correspondences) to guide the search for category names. + +were located in image coordinates within the range of $[-1000, -192]$ in the $x$ -domain, our calibrated intrinsic parameters transform this range to $[0, 1000]$ in $x$ . Importantly, we preserve the aspect ratio between the $x$ and $y$ coordinates to maintain the proportions of the final image without any distortion. This procedure ensures that each captured image is fully utilized and encompasses all regions of interest within the scene (Challenge 2 & 3). + +Class lookup table. Upon obtaining $N$ synthetic scene-level images, we input them into a 2D open-vocabulary detector. With text queries provided for interested classes, a list of detected objects in synthetic images can be obtained. Subsequently, information about detected objects, including their location and class, are stored in a designated Class Lookup Table (CLT). This table will later be retrieved to allocate class categories to 3D mask proposals (Challenge 4). + +# Lookup: Mask Classification through Searching + +We conduct a multi-level search to assign category labels to the mask proposals generated from Mask step. + +Mask2Pixel guided lookup. We introduce a Mask2Pixel Guided Lookup (MGL) to search within CLT. The concept involves projecting each 3D mask proposal onto a 2D plane using the same camera extrinsic and intrinsic matrices that are utilized to generate the 2D image, as depicted in Figure 5. With knowledge of the precise pixel locations of each mask in images, we can conduct an accurate search through the CLT to identify the most likely class for each mask. The development of MPM takes into account occlusion by integrating depth information. To accomplish the matching, we follow a three-step approach: 1. based + +on the mask's projection onto the 2D plane, we select the best-matched class categories in terms of IoU values; 2. if the IoU value of the best-matched object on the 2D plane is below $20\%$ , the match is disregarded. 3. we aggregate results from multiple views to formulate the final prediction, calculating probability scores using their normalized average IoU values, as illustrated in Figure 6. + +Local enforced lookup. While the Mask2Pixel Guided Lookup assigns class categories to mask proposals, some masks may not correspond to objects in the CLT. To address this, we introduce a Local Enforced Lookup (LEL) approach. We crop out the remaining masks from 2D scene-level images using enlarged bounding boxes and process them with the 2D detector to encourage detection. To select the best views, we introduce an Occlusion Report method to assess occlusion conditions for each mask in each projection, and then choose the top $K$ views for LEL. More details for Occlusion Report can be found in the supplementary materials. + +Final mask refinement. With the previous lookup approaches, a large proportion of mask proposals obtain a category prediction. All masks that have no category predictions after the MGL and LEL stages are eliminated. + +![](images/8398733af888cae47acd68594093de78c76a573e84a9b273f7fcbcfd6487bfa8.jpg) +Fig. 6: Mask2Pixel Guided Lookup Illustration. IoUs between the 2D detection results and the projected masks are the guidance to assign class names to 3D masks. Multiple image results are ensembled. + +# 4 Experiments + +Datasets and evaluation scheme. We tested OpenIns3D on five datasets, including four indoor datasets, namely S3DIS [2], ScanNetv2 [8], ScanNet200 [26], Replica [28], and one outdoor dataset, STPLS3D [5]. Among them, S3DIS, ScanNetv2, and ScanNet200 are indoor point cloud datasets generated from RGB-D images, Replica is a photo-realistic 3D indoor scene reconstruction, while STPLS3D is an aerial photogrammetry-constructed outdoor dataset. We exclusively used the 3D data with colour from these datasets and did not employ any 2D images, poses, or depth maps. Following the settings of prior work [9], we excluded the "other furniture" class in ScanNetv2 and the "clutter" class in S3DIS due to their vague meanings. Replica and ScanNet200 are evaluated by following the settings of OpenMask3D [30]. For STPLS3D, we merged the low, medium, and high vegetation classes into one "vegetation" class and kept all the rest. + +Implementation details. For the S3DIS, ScanNetv2, Scannet200, and STPLS datasets, the MPM module is trained without utilizing any category labels, and + +Table 1: Zero-shot object classification on ScanNetv2. OpenIns3D's Snap and Lookup approach for mask classification, surpasses all previous methods, including the latest language-aligned large-scale 3D foundation model [40]. + +
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointCLIP [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
PointCLIP V2 [37]11.00.00.023.80.00.00.00.07.80.090.70.00.00.00.064.40.00.0
CLIP2Point [14]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP2[35]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
Uni3D [40]45.858.53.778.883.754.931.339.470.135.11.927.394.213.838.710.788.147.6
OpenIns3D60.885.227.487.677.346.954.864.271.49.980.882.771.661.438.70.087.985.7
+ +$\lambda$ is set to 0.1 to reduce the weight of zero-IoU. The mask proposal of Replica is trained on ScanNet200, followed by OpenMask3D [30]. The Snap module captures images with a size of $1000 \times 1000$ including 16 global snaps, 4 corner snaps, and 4 wide-angle snaps. The top $0.5\mathrm{m}$ of the scene is removed for S3DIS, as the rooms are enclosed. For STPLIS3D, we followed Mask3D to split the large outdoor scene into patches of $50m \times 50m$ and lifted the camera up to $10m$ . More implementation details are presented in the supplementary materials. + +Open-vocabulary point cloud recognition. We first evaluate OpenIns3D performance among all existing 3D open-vocabulary models with the same setting, i.e. using only 3D inputs. These models are most commonly tested on recognition tasks, including PointCLIP(v1&v2) [43] [37], Clip2Point [14], CLIP² [35]. We also compare with large-scale 3D foundation models, such as Uni3D [40], which has 1-billion parameters, and was trained with large-scale 3D shapes and image-text pairs. We followed their evaluation scheme and reported the Top-1 accuracy of instance classification on ScanNetv2. The results are presented in Table 1. + +Open-vocabulary instance segmentation. We adopted various comparison schemes to align with existing methods. For 3D Open-vocabulary Instance Segmentation, we compared with PLA [9], and its follow-up works RegionPLC [34] and Lowis3D [10]. For a fair comparison, we followed their category splits and compared our results on novel classes, as demonstrated in Table 2. For STPLS3D, we compared OpenIns3D with baseline models whose classification module is PointCLIP and PointCLIPV2 [43] (Table 5). We also explored the performance of OpenIns3D on a more challenging dataset with more class categories. Specifically, we compared the performance of OpenIns3D with OpenMask3D, OpenScene, on Replica (Table 4) as well as ScanNet200 (Table 6). Following OpenMask3D, we used Mask3D to generate mask proposals for OpenScene for evaluation. + +Open-vocabulary object detection. Since there were limited works on 3D open-vocabulary instance segmentation at the time of conducting this work, we also selected some of the latest methods in the 3D open-world object detection domain for a more comprehensive evaluation. The evaluation is performed by converting generated masks into axis-aligned bounding boxes, and the results are shown in Table 3. + +Table 2: 3D Open-vocabulary Instance Segmentation Results on S3DIS and ScanNetv2. We compare our zero-shot performance on the novel categories defined in the PLA-family work. Significant improvements are achieved on the S3DIS dataset, and competitive results are observed on ScanNetv2 (B/N: Base/Novel). + +
OVISS3DISScanNetv2
MethodB/NAP50AP25B/NAP50AP25require 2D
PLA [9]8/408.6-10/721.9-✓
RegionPLC [34]8/4--10/732.3-✓
Lowis3D [10]8/413.8-10/731.2-✓
Open3DIS [22]8/426.3-10/7--✓
Mask3d+PointClip [43]- / 405.410.3- / 704.507.8X
OpenIns3D- / 437.039.3- / 727.942.6X
improvement- / 4(+10.7)(+29.0)- / 7-(+34.8)X
PLA [9]6/609.8-8/925.1-✓
RegionPLC [34]6/6--8/932.2-✓
Lowis3D [10]6/615.8-8/938.1-✓
Open3DIS [22]6/629.0-8/9--✓
Mask3d+PointClip [43]- / 608.510.6- / 905.606.7X
OpenIns3D- / 633.038.9- / 919.527.9X
improvement- / 6(+4.0)(+28.3)- / 9-(+21.2)X
Mask3d+PointClip [43]- / 1208.609.3- / 1704.514.4X
OpenIns3D- / 1228.329.5- / 1728.738.9X
improvement- / 12(+19.7)(+20.2)- / 17(+24.2)(+24.5)X
+ +Table 3: Open-vocabulary Object Detection $(AP_{25})$ on unseen classes in ScanNet. + +
Methodsmeantoiletbedchairsofadressertablecabinetbookshelfpillowsink
OV-PointCLIP [37]3.16.62.36.33.90.77.20.72.10.60.8
OV-Image2Point [32]0.80.20.81.01.40.22.81.00.90.00.1
Detic-ModelNet [41]1.74.21.04.61.20.23.20.61.20.00.7
Detic-ImageNet [41]0.40.00.00.20.00.51.80.50.30.00.7
OV-3DETIC [20]12.749.02.67.318.62.814.32.44.53.921.1
L3Det [42]24.656.336.216.123.08.123.114.717.323.427.9
FM-OV3D [36]21.555.038.819.241.923.83.50.46.017.48.8
OpenIns3D43.779.570.576.915.80.053.140.141.27.153.1
improvement(+19.1)(+23.2)(+31.7)(+57.7)--(+30)(+25.4)(+23.9)(+25.2)
+ +# 5 Results and Discussion + +# 5.1 Comparison with SOTA + +First of all, OpenIns3D demonstrates impressive performance in open-vocabulary point cloud recognition, surpassing all previous methods, including the large-scale 3D foundation model by $15\%$ . This proves the effectiveness of the zero-shot "Snap" and "lookup" scheme. With the enhanced recognition capability, the performance of 3D open-vocabulary Object Detection among the ScanNet dataset has also achieved state-of-the-art results by a large margin. For 3D instance segmentation, compared to works in the PLA family [9, 10, 34] and the latest work Open3DIS [22], OpenIns3D does not require aligned images as input, (still need RGB information for points), achieves higher results on the S3DIS dataset, both in the 4 novel categories split and the 6 novel categories split. In SPTLS3D, OpenIns3D outperforms the baseline model PointCLIPV2 by $9.3\%$ in AP. On + +Table 4: OVIS on the indoor Replica dataset. + +
Model2DAP\( AP_{50} \)\( AP_{25} \)
OpenScene [23] (2D Fusion)✓10.915.617.3
OpenScene [23] (2D/3D Ens.)✓8.210.413.3
OpenMask3D [30]✓13.118.424.2
OpenScene [23] (3D Distill)✗8.210.512.6
OpenIns3D✗13.618.019.7
improvement✗(+5.4)(+7.5)(+7.1)
+ +Table 5: OVIS on outdoor STPLS3D dataset + +
ModelAPAP50AP25
PointCLIP [43]02.002.604.0
PointCLIPv2 [37]02.103.105.2
OpenIns3D11.414.217.2
improvement(+9.3)(+11.9)(+12.0)
+ +Table 6: 3D instance segmentation results on the ScanNet200 validation set. OpenIns3D demonstrates robust performance when compared to 2D-input-free models. However, notable limitations emerge when dealing with small objects in the common and tail classes. + +
Modeluse 2DAPheadAPcommonAPtailAPAP50AP25
OpenScene (2D Fusion) [23]✓13.411.69.911.715.217.8
OpenScene (2D/3D Ens.) [23]✓11.03.21.15.36.78.1
OpenMask3D✓17.114.114.915.419.923.1
OpenScene (3D Distill)✗10.62.60.74.86.27.2
OpenIns3D✗16.06.54.28.810.314.4
improvement✗(+5.4)(+3.9)(+3.6)(+4.0)(+4.1)(+7.2)
+ +the Replica dataset, OpenIns3D even outperformed OpenMask3D, which relies on well-aligned images for mask understanding. In the case of ScanNet200, OpenIns3D attains the highest performance compared to all other 3D input baselines. However, we have observed a decline in performance on tail and common classes within ScanNet200. The decrease in performance can be attributed to the low-quality and unclear reconstruction of smaller objects in the ScanNet200 scene. This is a noticeable limitation of OpenIns3D, but for better reconstructions, such as in Replica, and for Head classes within ScanNet200, OpenIns3D still demonstrates decent performance. + +In summary, OpenIns3D demonstrates the best performance among all existing methods if only 3D data is used as input and outperforms many existing state-of-the-art methods that require 2D images. It also shows certain limitations on small objects that are not well-reconstructed in 3D scenes. + +# 5.2 Ablation study + +Mask quality ablation. Following the evaluation on ScanNetv2, we assessed the class-agnostic mask quality using the average precision (AP) score. We treated all classes as universal since the predictions are class-agnostic. The evaluation was conducted on the ScanNetv2 validation set. Table 8 demonstrates the effectiveness of the Mask Scoring and Mask Filtering designs. + +Table 7: Rendering and Inference Time Ablations. Results tested on typical ScanNet scenes with 50 masks. OpenIns3D requires less rendering and inference time. + +
RenderingNum of Img needed2D backboneImg size (w × h)\( T_{render} \) (s/scene)\( T_{infer} \) (s/scene)\( T_{total} \) (s/scene)AP25 (%)
PointCLIP250CLIP\( 128^2 \)5.215.320.59.3
LAR250CLIP\( 128^2 \)14.318.733.010.5
Mask rendering250CLIP\( 128^2 \)42.619.562.17.3
OpenIns3D (ours)8G-DINO\( 1000^2 \)2.36.28.529.8
OpenIns3D (ours)8ODISE\( 1000^2 \)2.38.210.535.1
+ +Table 8: MPM Ablation. MS: Mask scoring. MF: Mask filtering. + +
MethodAP50AP25
Mask3d-Supervised74.780.9
CA-Mask3d47.549.2
CA-Mask3d + MS50.2 (+02.7)53.3 (+04.1)
CA-Mask3d + MF61.6 (+14.1)71.0 (+21.8)
CA-Mask3d + MS + MF64.6 (+17.0)73.4 (+24.2)
+ +Table 9: Number of Views Abla-tion. LEL: Local Enforced Lookup + +
IDX4816LELAP50AP25
i✓18.327.1
ii✓22.735.1
iii✓24.837.5
iv✓✓28.738.9
+ +Multi-view ablation. We also studied the effects of using different numbers of views (Table 9). Increasing the number of views used in the Lookup module leads to better results. Additionally, Look Enforced Lookup provided a final improvement to the results. + +Projection and 2D backbone ablation. We conducted a comprehensive study on various rendering methods and their interaction with the 2D backbone to identify a suitable approach. We report the rendering time and inference performance for each method (Table 7), and more details can be found in the supplementary materials. The key observation is that the scene-level rendering and understanding approach excels in speed while also demonstrating strong performance. Switching from Grounding Dino [19] to the latest ODISE [33] also brings gains in performance, indicating that the OpenIns3D framework can easily benefit from the rapid development of 2D open-world detectors. + +Cross-domain analysis. To evaluate the generalization capability of MPM across different domains, we trained and tested OpenIns3D on two different datasets, as shown in Table 10. The cross-domain models also demonstrate impressive performance on both datasets when compared with the baseline. Notably, within the 17 classes of ScanNetv2, 11 classes do not exist in S3DIS. OpenIns3D, trained on S3DIS, still achieves decent performance among these unseen classes. + +Free-flow language capability. The snap and lookup scheme outsources the mask understanding tasks to a 2D vision-language model. Therefore when integrated with a 2D model powered by large language models (LLMs) like LISA [18], OpenIns3D can perform reasoning-based segmentation tasks (as depicted in Figure 7). For instance, when given the query "pencilling down ideas during brain- + +storming," OpenIns3D accurately segments the whiteboard, while for "furniture offers recreational enjoyment with friends," it precisely identifies and segments the pool table. + +![](images/1df998dcb8085692438928bd7bbade94c4fe596135cb01b58cdbdadfc6a06a98.jpg) +Fig. 7: Qualitative Results from OpenIns3D. OpenIns3D (ODISE) demonstrates the ability to manage a versatile vocabulary. OpenIns3D (LISA) can conduct 3D reasoning segmentation. + +Table 10: Cross-domain Ablation. We trained and tested OpenIns3D on two different datasets to examine its cross-domain capability. While S3DIS and ScanNetV2 have non-overlapping classes, OpenIns3D demonstrates decent generalization capability. + +
TestModelTraining DataAP50AP25
ScanNetv2 [8]Mask3D-Pointclip [43]ScanNetv204.514.4
OpenIns3DScanNetv228.738.9
OpenIns3DS3DIS21.533.6
S3DIS [1]Mask3D-Pointclip [43]S3DIS03.506.8
OpenIns3DS3DIS28.329.5
OpenIns3DScanNetv214.219.8
+ +# 6 Conclusion + +Achieving 3D open-vocabulary scene understanding is a challenging task, primarily due to the lack of extensive 3D-text data. Currently, most work in this domain focuses on using 2D images to bridge the gap between 3D and language. This, however, not only requires a good alignment between 2D and 3D but also evolves slowly due to the significant effort needed for retraining when changing the 2D backbone. OpenIns3D introduces a new pipeline, i.e. Mask-Snap-Lookup, for this task. The Mask module generates authentic masks in the 3D domain, while Snap renders scene-level images in 2D domains, and the Lookup module links the results from 2D to 3D precisely. This pipeline requires no image input, achieves better performance, and can evolve seamlessly with 2D models without training. We hope our work will provide a fresh perspective for researchers working towards open-world 3D scene understanding. + +# Acknowledgments + +This work is supported by the Girton College Graduate Research Awards at the University of Cambridge, School of Technology, National Highways sponsored through EPSRC Centre for Doctoral Training, the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR, the National Natural Science Foundation of China No. 62201484, HKU Startup Fund, and HKU Seed Fund for Basic Research. We would like to express our gratitude to Yunhan Yang for his assistance in exploring rendering techniques and to Chengyao Wang for sharing his implementation of Mask3D. + +# References + +1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3, 14 +2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 9 +3. Bakr, E.M., Alsaedy, Y.Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: NeurIPS (2022) 5 +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 2 +5. Chen, M., Hu, Q., Yu, Z., THOMAS, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. In: BMVA (2022) 3, 9 +6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023) 2, 4, 5 +7. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. In: ICCV (2023) 2, 6 +8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9, 14 +9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 2, 4, 5, 9, 10, 11 +0. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. In: TPAMI (2024) 2, 4, 5, 10, 11 +1. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: ICML (2023) 2 +2. Griffiths, D., Boehm, J.: SynthCity: A large-scale synthetic point cloud. In: arXiv (2019) 2 +3. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: Semantic3D.net: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017) 2 +4. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV (2023) 4, 10 + +15. Huang, Z., Huang, L., Gong, Y., Huang, C., Wang, X.: Mask scoring R-CNN. In: CVPR (2019) 6 +16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 2, 6, 7 +17. Kundu, A., Yin, X., Fathi, A., Ross, D.A., Brewington, B., Funkhouser, T.A., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: ECCV (2020) 5 +18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024) 4, 13 +19. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv (2023) 13 +20. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR (2023) 3, 4, 11 +21. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019) 2 +22. Nguyen, P.D.A., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024) 2, 4, 5, 11 +23. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 4, 12 +24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2 +25. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-lille-3d: A large and high-quality ground-truth urban point cloud dataset for automatic segmentation and classification. In: The International booktitle of Robotics Research (2018) 2 +26. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 4, 9 +27. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 5, 6 +28. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces. In: arXiv (2019) 3, 9 +29. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.G.: Multi-view convolutional neural networks for 3d shape recognition. In: ICCV (2015) 5 +30. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023) 3, 4, 5, 9, 10, 12 +31. Tan, W., Qin, N., Ma, L., Li, Y., Du, J., Cai, G., Yang, K., Li, J.: Toronto-3D: A large-scale mobile lidar dataset for semantic segmentation of urban roadways. In: CVPRW (2020) 2 + +32. Xu, C., Yang, S., Galanti, T., Wu, B., Yue, X., Zhai, B., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: Image2point: 3d point-cloud understanding with 2d image pretrained models. In: ECCV (2022) 11 +33. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In: CVPR (2023) 2, 13 +34. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024) 2, 4, 5, 10, 11 +35. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: $\mathrm{Clip}^2$ : Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR (2023) 4, 10 +36. Zhang, D., Li, C., Zhang, R., Xie, S., Xue, W., Xie, X., Zhang, S.: Fm-ov3d: Foundation model-based cross-modal knowledge blending for open-vocabulary 3d detection. In: AAAI (2024) 2, 3, 11 +37. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by CLIP. In: CVPR (2022) 3, 4, 10, 11, 12 +38. Zhou, C., Loy, C.C., Dai, B.: Denseclip: Extract free dense labels from clip. In: CVPR (2022) 2 +39. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022) 2 +40. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024) 2, 3, 10 +41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 11 +42. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. In: arXiv (2023) 11 +43. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV (2023) 2, 3, 4, 5, 10, 11, 12, 14 \ No newline at end of file diff --git a/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/images.zip b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3bd51349d9e3a21424c2c6c2b19869e68d2f9799 --- /dev/null +++ b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b007b6f136a8ed501120308b60fd6c48db8cc5507df304b4b7328ac0378e66ed +size 695519 diff --git a/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/layout.json b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..534cbff45cd927ad8adf13accc9c83bcd0871889 --- /dev/null +++ b/2024/OpenIns3D_ Snap and Lookup for 3D Open-vocabulary Instance Segmentation/layout.json @@ -0,0 +1,8065 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 112, + 460, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 112, + 460, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 112, + 460, + 148 + ], + "type": "text", + "content": "Opens3D: Snap and Lookup for 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 194, + 168, + 275, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 168, + 275, + 180 + ], + "spans": [ + { + "bbox": [ + 194, + 168, + 275, + 180 + ], + "type": "text", + "content": "Zhening Huang1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 286, + 168, + 358, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 168, + 358, + 180 + ], + "spans": [ + { + "bbox": [ + 286, + 168, + 358, + 180 + ], + "type": "text", + "content": "Xiaoyang " + }, + { + "bbox": [ + 286, + 168, + 358, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{W u}^{2}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 370, + 168, + 419, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 168, + 419, + 179 + ], + "spans": [ + { + "bbox": [ + 370, + 168, + 419, + 179 + ], + "type": "text", + "content": "Xi Chen" + }, + { + "bbox": [ + 370, + 168, + 419, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 185, + 181, + 281, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 181, + 281, + 192 + ], + "spans": [ + { + "bbox": [ + 185, + 181, + 281, + 192 + ], + "type": "text", + "content": "Hengshuang Zhao\\*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 290, + 181, + 345, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 181, + 345, + 192 + ], + "spans": [ + { + "bbox": [ + 290, + 181, + 345, + 192 + ], + "type": "text", + "content": "Lei Zhu" + }, + { + "bbox": [ + 290, + 181, + 345, + 192 + ], + "type": "inline_equation", + "content": "^{3,4}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 356, + 181, + 429, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 181, + 429, + 192 + ], + "spans": [ + { + "bbox": [ + 356, + 181, + 429, + 192 + ], + "type": "text", + "content": "Joan Lasenby" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 187, + 201, + 427, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 201, + 427, + 213 + ], + "spans": [ + { + "bbox": [ + 187, + 201, + 427, + 213 + ], + "type": "text", + "content": "1University of Cambridge 2The University of Hong Kong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 204, + 213, + 409, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 213, + 409, + 224 + ], + "spans": [ + { + "bbox": [ + 204, + 213, + 409, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 204, + 213, + 409, + 224 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 167, + 224, + 446, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 224, + 446, + 236 + ], + "spans": [ + { + "bbox": [ + 167, + 224, + 446, + 236 + ], + "type": "text", + "content": "4The Hong Kong University of Science and Technology (Guangzhou)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 211, + 238, + 402, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 238, + 402, + 248 + ], + "spans": [ + { + "bbox": [ + 211, + 238, + 402, + 248 + ], + "type": "text", + "content": "https://zheninghuang.github.io/OpenIns3D" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 139, + 254, + 249, + 320 + ], + "blocks": [ + { + "bbox": [ + 139, + 254, + 249, + 320 + ], + "lines": [ + { + "bbox": [ + 139, + 254, + 249, + 320 + ], + "spans": [ + { + "bbox": [ + 139, + 254, + 249, + 320 + ], + "type": "image", + "image_path": "fc2489a3e5f401235117350154fd6d85d7b0ac99b092789ef21aad86adc82d8b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 321, + 249, + 330 + ], + "lines": [ + { + "bbox": [ + 140, + 321, + 249, + 330 + ], + "spans": [ + { + "bbox": [ + 140, + 321, + 249, + 330 + ], + "type": "text", + "content": "\"furniture that is capable of producing music.\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 252, + 254, + 362, + 320 + ], + "blocks": [ + { + "bbox": [ + 252, + 254, + 362, + 320 + ], + "lines": [ + { + "bbox": [ + 252, + 254, + 362, + 320 + ], + "spans": [ + { + "bbox": [ + 252, + 254, + 362, + 320 + ], + "type": "image", + "image_path": "b0fe4c32c28f73c06a571d0f37b7b6051fc3628edb9df459fe3726dd327c647a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 321, + 342, + 328 + ], + "lines": [ + { + "bbox": [ + 273, + 321, + 342, + 328 + ], + "spans": [ + { + "bbox": [ + 273, + 321, + 342, + 328 + ], + "type": "text", + "content": "device to watch BBC news" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 334, + 466, + 345 + ], + "lines": [ + { + "bbox": [ + 145, + 334, + 466, + 345 + ], + "spans": [ + { + "bbox": [ + 145, + 334, + 466, + 345 + ], + "type": "text", + "content": "Fig. 1: Complex Queries 3D Instance Segmentation with OpenIns3D." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 364, + 254, + 474, + 320 + ], + "blocks": [ + { + "bbox": [ + 364, + 254, + 474, + 320 + ], + "lines": [ + { + "bbox": [ + 364, + 254, + 474, + 320 + ], + "spans": [ + { + "bbox": [ + 364, + 254, + 474, + 320 + ], + "type": "image", + "image_path": "bd1a3f2ed6ae89ed11399adf6ff3197706b7ba264a727977b35d7b7b595e7ef1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 321, + 460, + 328 + ], + "lines": [ + { + "bbox": [ + 378, + 321, + 460, + 328 + ], + "spans": [ + { + "bbox": [ + 378, + 321, + 460, + 328 + ], + "type": "text", + "content": "\"Ma Long's domain of excellence" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 159, + 366, + 453, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 366, + 453, + 553 + ], + "spans": [ + { + "bbox": [ + 159, + 366, + 453, + 553 + ], + "type": "text", + "content": "Abstract. In this work, we introduce OpenIns3D, a new 3D-input-only framework for 3D open-vocabulary scene understanding. The OpenIns3D framework employs a \"Mask-Snap-Lookup\" scheme. The \"Mask\" module learns class-agnostic mask proposals in 3D point clouds, the \"Snap\" module generates synthetic scene-level images at multiple scales and leverages 2D vision-language models to extract interesting objects, and the \"Lookup\" module searches through the outcomes of \"Snap\" to assign category names to the proposed masks. This approach yet simple, achieves state-of-the-art performance across a wide range of 3D open-vocabulary tasks, including recognition, object detection, and instance segmentation, on both indoor and outdoor datasets. Moreover, OpenIns3D facilitates effortless switching between different 2D detectors without requiring retraining. When integrated with powerful 2D open-world models, it achieves excellent results in scene understanding tasks. Furthermore, when combined with LLM-powered 2D models, OpenIns3D exhibits an impressive capability to comprehend and process highly complex text queries that demand intricate reasoning and real-world knowledge." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "spans": [ + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "content": "Keywords: Open-Vocabulary Understanding " + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "content": " 3D Scene Understanding " + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "content": " Vision-Language Model" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 597, + 230, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 230, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 230, + 609 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "content": "3D scene understanding plays a critical role in various domains, such as autonomous driving, robotic sensing, AR/VR, and manufacturing, among others." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 121, + 347, + 254 + ], + "blocks": [ + { + "bbox": [ + 137, + 121, + 347, + 254 + ], + "lines": [ + { + "bbox": [ + 137, + 121, + 347, + 254 + ], + "spans": [ + { + "bbox": [ + 137, + 121, + 347, + 254 + ], + "type": "image", + "image_path": "57840a1969993fc1483db598a8b507bc99a38ca0dc2030bd1f620f9f98a0388b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "type": "text", + "content": "Fig. 2: High-level Illustrations of OpenIns3D and Quantitative Results. (a) OpenIns3D follows the \"Mask-Snap-Lookup\" steps for open-vocabulary scene understanding. (b) A list of SOTA results has been achieved on both indoor and outdoor datasets. OV-Rec: open-vocabulary object recognition. OVOD: open-vocabulary object detection. OVIS: open-vocabulary instance segmentation. PointCLIPV2 [43]; Uni3D [40]; Open3DIS [22]; FM-OV3D [36]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 350, + 121, + 478, + 254 + ], + "blocks": [ + { + "bbox": [ + 350, + 121, + 478, + 254 + ], + "lines": [ + { + "bbox": [ + 350, + 121, + 478, + 254 + ], + "spans": [ + { + "bbox": [ + 350, + 121, + 478, + 254 + ], + "type": "image", + "image_path": "8b2f93d3e12882c7e597be4a8a9716fc2eae35b6904cd515dba5b847a408e329.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "type": "text", + "content": "While the development of 3D closed-set understanding is relatively mature, scene understanding in an open-vocabulary setting is still in its infancy. Closed-set understanding can only handle a predefined set of concepts and scenarios but fails to provide valid responses when faced with unfamiliar concepts or variations in language usage. This limitation impacts its performance in dynamic and ever-changing contexts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 640 + ], + "type": "text", + "content": "Thanks to internet-scale image-text datasets, significant progress has been made in 2D image open-vocabulary understanding [4, 7, 11, 16, 24, 33, 38, 39]. However, unlike 2D data that can be easily collected from the internet, constructing a large-scale 3D-text dataset poses a challenge. As a result, the most viable approach to achieving 3D open-vocabulary understanding involves leveraging 2D images to bridge language and 3D data. In this direction, there have been several notable works, such as OpenScene [23], PLA-family [9, 10, 34], and CLIP2Scene [6]. These works leverage well-aligned 2D images and 3D point clouds to conduct feature distillation or employ 2D caption models to construct 3D-text pairs. One prerequisite of these methods, however, is the availability of well-aligned 2D images and 3D point clouds. This means that posed 2D images, associated depth maps and camera models need to be accessible as inputs to the network. In real-life scenarios, there are numerous cases where 2D images or the information required to align 2D and 3D data are unavailable. For instance, to save storage space, point clouds generated with LiDAR are often stored without accompanying 2D images (example datasets include [13, 25, 31]). In cases where point clouds are obtained from the registration of multiple scans from different sensors or are converted from 3D simulations/CAD models [12, 21], 2D images are often non-existent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "We believe that developing a 3D open-vocabulary framework without relying on well-aligned 2D images is meaningful, as this will simplify deployment pre" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "requisites and enhance its applicability across a wide range of scenarios. To this end, we introduce OpenIns3D, a framework designed to effectively perform 3D open-vocabulary scene understanding tasks without relying on 2D aligned images. Overall, OpenIns3D comprises three core steps: Mask, Snap, and Lookup. An overall illustration of OpenIns3D is presented in Figure 2a." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 178, + 480, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 251 + ], + "type": "text", + "content": "Mask: Given a 3D point cloud, the first part of OpenIns3D learns class-agnostic mask proposals with a Mask Proposal Module (MPM). This process is trained without any classification labels. To control the quality of the mask, MPM proposes a learnable Mask Scoring module to predict the quality of each mask output and implements a list of Mask Filtering techniques to discard invalid, low-quality masks. MPM outputs a list of class-agnostic masks in the scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 481, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 481, + 348 + ], + "type": "text", + "content": "Snap: Multiple synthetic scene-level images are generated with calibrated and optimized camera poses and intrinsic parameters. These images are specifically designed to encompass part or all of the relevant masks, aiming to minimize the need for multiple renderings. Instead of individually predicting the category of each mask proposal [20, 37, 43], the scene-level images are input into 2D open-vocabulary models for the simultaneous understanding of all interesting objects present in the scene. A Class Lookup Table (CLT) is then constructed to store all the detected object categories alongside their respective pixel locations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 460 + ], + "type": "text", + "content": "Lookup: To precisely determine the positions of mask proposals in each image, Mask2Pixel maps are constructed. These maps project all 3D mask proposals onto 2D images with identical camera parameters used in Snap. In the Lookup phase, OpenIns3D searches through the CLT with the help of Mask2Pixel maps to precisely assign category names to 3D mask proposals. Results from multiple views are combined to establish initial mask classification outcomes. For remaining masks, a similar Lookup procedure is carried out on a local scale to facilitate classification. Lastly, the 3D mask proposals are refined by removing masks lacking class assignments after both global and local Lookup." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 665 + ], + "type": "text", + "content": "OpenIns3D demonstrates promising performance in extensive comparisons with other methods, as summarized in Figure 2b. This simple and flexible approach achieves a range of state-of-the-art (SOTA) results on various 3D open-vocabulary tasks. Specifically, SOTA results are achieved on open-vocabulary instance segmentation for the indoor S3DIS [1] dataset and the outdoor STPLS3D [5] dataset. OpenIns3D even outperforms OpenMask3D [30], a concurrent work that heavily utilizes 2D images, for instance, segmentation on the challenging Replica [28] dataset. We also find that the \"Snap and Lookup\" serves as a powerful open-vocabulary object recognition engine, which achieved SOTA object recognition tasks on ScanNet [8], outperforming the previous SOTA, a 1-billion parameter 3D foundation model [40]. Lastly, when converting mask proposals into 3D bounding boxes, OpenIns3D also achieved state-of-the-art results in open-vocabulary object detection (OVOD) on ScanNet, outperforming previous image-dependent methods [36]. The design of OpenIns3D also allows 2D detectors to be changed without the need for retraining. This provides the model with the capability to evolve alongside the latest development of the 2D open-vocabulary models. Moreover, when 2D detectors are coupled with Large Lan-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 121, + 212, + 223 + ], + "blocks": [ + { + "bbox": [ + 137, + 121, + 212, + 223 + ], + "lines": [ + { + "bbox": [ + 137, + 121, + 212, + 223 + ], + "spans": [ + { + "bbox": [ + 137, + 121, + 212, + 223 + ], + "type": "image", + "image_path": "56e04e05cf354c1035c4332cae16288847e2d3ee8f0c2b3747bd7bf4f8ab3c8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 227, + 200, + 236 + ], + "lines": [ + { + "bbox": [ + 137, + 227, + 200, + 236 + ], + "spans": [ + { + "bbox": [ + 137, + 227, + 200, + 236 + ], + "type": "text", + "content": "(a) 3D Feature Distillation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 121, + 301, + 223 + ], + "blocks": [ + { + "bbox": [ + 225, + 121, + 301, + 223 + ], + "lines": [ + { + "bbox": [ + 225, + 121, + 301, + 223 + ], + "spans": [ + { + "bbox": [ + 225, + 121, + 301, + 223 + ], + "type": "image", + "image_path": "5700f25ca1d75f620c5e91366bded4bc61d16fca6ad8c22ebb93857c706f28a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 227, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 226, + 227, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 226, + 227, + 287, + 236 + ], + "type": "text", + "content": "(b) Building 3D-text pair" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 121, + 389, + 223 + ], + "blocks": [ + { + "bbox": [ + 313, + 121, + 389, + 223 + ], + "lines": [ + { + "bbox": [ + 313, + 121, + 389, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 389, + 223 + ], + "type": "image", + "image_path": "0ea3b54d16ccd3fb51b8ca573a330619e299eeda3b24ec51d7d245f1cfdae1ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 227, + 371, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 371, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 371, + 236 + ], + "type": "text", + "content": "(c) CLIP and Projection" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 242, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 331 + ], + "type": "text", + "content": "Fig. 3: Four Categories of Open-Vocabulary 3D Scene Understanding Models. a) 3D feature distillation frameworks, where 2D images are used as a bridge to distil language-aligned features into 3D, with typical works including OpenScene [23] and Clip2Scene [6]. b) Building 3D-text pairs, where 2D captioning models are used to build 3D-text pairs for feature learning, with typical works including the PLA-family [9,10,34] c) CLIP and Projection, where objects are cropped out of 2D images before being processed by CLIP, and the results are directly projected into 3D, including OpenMask3D [30], OV-3DET [20], " + }, + { + "bbox": [ + 130, + 242, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 130, + 242, + 482, + 331 + ], + "type": "text", + "content": " [35] and Open3DIS [22]. d) OpenIns3D" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 402, + 121, + 475, + 221 + ], + "blocks": [ + { + "bbox": [ + 402, + 121, + 475, + 221 + ], + "lines": [ + { + "bbox": [ + 402, + 121, + 475, + 221 + ], + "spans": [ + { + "bbox": [ + 402, + 121, + 475, + 221 + ], + "type": "image", + "image_path": "f02f5869c5d4c0c810eba53bf727f6dedfc9597e07fb45b5b818b7436e848ba4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 227, + 471, + 236 + ], + "lines": [ + { + "bbox": [ + 399, + 227, + 471, + 236 + ], + "spans": [ + { + "bbox": [ + 399, + 227, + 471, + 236 + ], + "type": "text", + "content": "(d) Mask-Snap-Lookup (ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 338, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 398 + ], + "type": "text", + "content": "guage Models (LLMs), OpenIns3D can enable complex query understanding capability. When integrated with LISA [18], an LLM-powered reasoning segmentation model, OpenIns3D exhibits a strong ability to comprehend highly intricate language queries and performs reasoning segmentation in 3D, as illustrated in Figure 1. In summary, our contributions are:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 399, + 482, + 494 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 134, + 399, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 399, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 134, + 399, + 482, + 446 + ], + "type": "text", + "content": "- OpenIns3D employs a distinct pipeline that operates without the need for well-aligned images. This approach achieves state-of-the-art results across a range of benchmarks and possesses the ability to comprehend highly complex input queries." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 446, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 446, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 135, + 446, + 482, + 494 + ], + "type": "text", + "content": "- The proposed \"Snap and Lookup\" combination can serve as a powerful 3D object recognition engine, especially for noisy 3D objects extracted from scene-level scans. This capability is demonstrated by achieving state-of-the-art results with a large margin." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 510, + 234, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 234, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 234, + 522 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "3D open-vocabulary understanding. Progress in 3D open-vocabulary understanding has been relatively slow compared to that of images. In the domain of 3D object classification tasks, methods like PointCLIP [37], PointCLIPV2 [43], and CLIP2Point [14] project 3D point clouds into depth maps and link them with 2D models for classification. However, these methods lack performance in scene-level understanding, where points are often overlapped and incomplete. For scene-level understanding, most work has primarily focused on leveraging well-aligned 2D posed images, depth maps, and point clouds [6,9,10,23,26,34,35]. One notable example is OpenScene [23], which takes posed 2D images, depth maps, and 3D data as input, and feature distillation is performed to transfer 2D language-aligned features from images to 3D point clouds. Similarly," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 331 + ], + "type": "text", + "content": "Clip2Scene [6] builds dense pixel-point pairs by calibrating the LiDAR point cloud with corresponding images captured by six cameras. However, achieving instance-level understanding is challenging with these methods as they focus solely on semantic-level understanding. In contrast, PLA [9] and its follow-up work RegionPLC [34] and Lowis3D [10] utilize a 2D caption model to construct 3D-text pairs to learn features. However, the PLA-family works rely on a binary head to classify the input object into base categories or novel categories, and the transferability of this binary head to different base-novel splits is very limited, posing a challenge for flexible applications. One current work, OpenMask3D [30], utilizes well-aligned 2D images to learn features for mask proposals, leading to impressive results in open-vocabulary instance segmentation. Open3DIS [22] follows the same procedure, enhancing both mask proposals and mask classification with images, yielding better performance. A common issue with these methods is their reliance on well-aligned 3D and 2D pairs in the input, which may not always be available in real applications. We summarise the key difference between these methods in Figure 3. Simplifying input requirements improves method flexibility and compatibility. In this work, we explore how to conduct 3D open-vocabulary understanding without relying on 2D images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 333, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 333, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 133, + 333, + 481, + 499 + ], + "type": "text", + "content": "Image generation from 3D. Projection-based methods have been extensively explored in the past for 3D understanding and have proven to be beneficial for obtaining complementary features. For instance, MVCNN [29] projects 3D objects to different views to aid in feature learning, while LAR [3] introduces object centre projection methods to generate images for 3D objects from various angles, assisting visual grounding tasks. Additionally, Virtual View Fusion [17] employs the original camera pose but enlarges the field of view, resulting in enhanced 2D feature transfer. However, these methods encounter challenges like best view selection, object occlusion, information loss during projection, and long rendering times. In the context of open-vocabulary settings, the quality of the projected image plays a crucial role in model performance. In our work, we evaluate different projection methods, along with their compatibility with 2D open-vocabulary models, to identify an optimal solution that achieves good results and is efficient to implement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 524, + 220, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 524, + 220, + 538 + ], + "spans": [ + { + "bbox": [ + 133, + 524, + 220, + 538 + ], + "type": "text", + "content": "3 OpenIns3D" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 555, + 281, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 281, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 281, + 567 + ], + "type": "text", + "content": "3.1 Baseline and Challenges" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": "Baseline. We build a naive baseline by adopting the recent 3D instance segmentation backbone Mask3D [27] to generate mask proposals. To make the Mask Proposal Module (MPM) fit for the open-vocabulary setting, we remove all components in Mask3D that use the classification labels. Later, PointCLIP [43] is adopted for mask understanding. This naive approach, although satisfying the requirement of 3D inputs only, has long rendering times and unsatisfactory performance (See in Table 7). There are several problems in this baseline model." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 103, + 479, + 230 + ], + "blocks": [ + { + "bbox": [ + 136, + 103, + 479, + 230 + ], + "lines": [ + { + "bbox": [ + 136, + 103, + 479, + 230 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 479, + 230 + ], + "type": "image", + "image_path": "8d1b9b92f820f6790837997a5d5d46ee43c3e3a0409ae66129bfb0c616b33ab2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "type": "text", + "content": "Fig. 4: General Pipeline of OpenIns3D OpenIns3D first processes point clouds with MPM to generate 3D mask proposals and mask scores. The Snap module (detailed in Figure 5) then renders " + }, + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "type": "text", + "content": " synthetic scene-level images, which are later passed into the 2D open-world model along with the input text queries. The detection results from the 2D model are stored in the Class Lookup Table (CLT). Finally, both the mask proposals and CLT are fed into the Lookup module, where Mask2Pixel Guided Lookup (detailed in Figure 6) is performed at the global level, followed by a Local Enforced Lookup at the local level to unlock the semantic meaning of mask proposals. The final mask filtering refines the mask proposals and obtains the final results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 350, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 386 + ], + "type": "text", + "content": "Challenge 1: excessive mask proposals. in Mask3D [27], mask proposals are filtered by the mask classification logit, which is removed in the class-agnostic setting. Therefore, an effective mask filtering scheme is needed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 422 + ], + "type": "text", + "content": "Challenge 2: low quality of 3D instances. 3D instances extracted from scene scans are typically broken, uncompleted, and sparse. Therefore, generated images by simple projection are not easily understood by 2D VL models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "type": "text", + "content": "Challenge 3: lack of context information. Humans recognize imperfect 3D point cloud objects through scene comprehension. However, isolated instance point cloud projections lack this context. A solution is to project both the mask and background onto images, yet this introduces distracting elements, potentially confusing classification-level models like CLIP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 519 + ], + "type": "text", + "content": "Challenge 4: domain gap between projected images and natural images. Rendered images often differ significantly from natural images used in training, posing a challenge for 2D visual language models to comprehend." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 534, + 257, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 257, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 257, + 545 + ], + "type": "text", + "content": "3.2 Overall Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 553, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 578 + ], + "type": "text", + "content": "In this section, we present our design of OpenIns3D, which targets the aforementioned four problems. The pipeline of OpenIns3D is shown in Figure 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 594, + 313, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 313, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 313, + 605 + ], + "type": "text", + "content": "Mask: class-agnostic mask proposal" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Within the Mask Proposal Module (MPM), we introduce two straightforward designs aimed at filtering low-quality masks generated in the baseline model. Mask scoring. Inspired by [7, 15, 16], We feed the instance queries generated from the Mask Module into a shallow MLP module to predict the quality, i.e" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "IoU of the mask. The predicted IoU " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "(IoU_{m})" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " is supervised by the ground truth IoU " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "(IoU_{gt})" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " value during the training stages, which is calculated between the predicted mask and its matched ground truth mask in the Bipartite Matching. For unmatched prediction masks " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "(IoU_{u})" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": ", we label the ground truth IoU value as zero. The loss is computed using " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": ". To avoid overly low IoU predictions, a hyper-parameter " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " is introduced to reduce the weight of loss for unmatched masks. Therefore, the total loss function for the MPM is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 211, + 480, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 211, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 209, + 211, + 480, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\gamma \\sum \\left(I o U _ {u}\\right) ^ {2} + \\sum \\left(I o U _ {m} - I o U _ {\\mathrm {g t}}\\right) ^ {2} \\tag {1}", + "image_path": "f34a94f9a9884d8ff0b82eadf121aed94c818d572c1192ff31863c6d2db5ef4b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "content": "Mask filtering. To enhance mask quality, three filters are applied. Firstly, we retain masks with a model-predicted IoU score above a threshold of " + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "content": ", ensuring that only high-quality masks are kept. Secondly, drawing inspiration from SAM [16], we focus on stable masks by comparing two binary masks derived from the same underlying soft mask using different threshold values. Specifically, we introduce an offset value " + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "content": " and select masks where the IoU between the pair of thresholded masks (one with " + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "inline_equation", + "content": "-\\alpha" + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "content": " and the other with " + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "inline_equation", + "content": "+\\alpha" + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "content": ") exceeds " + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "content": ". Lastly, small objects in the scene often lead to invalid proposals, so we filter out mask proposals that have a point number lower than " + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{min}}" + }, + { + "bbox": [ + 130, + 236, + 480, + 369 + ], + "type": "text", + "content": ". By employing these techniques, the quantity of mask proposals will decrease, resulting in cleaner and higher-quality masks for subsequent mask understanding tasks (Challenge 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 388, + 361, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 361, + 401 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 361, + 401 + ], + "type": "text", + "content": "Snap: Synthetic Scene-level View Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "text", + "content": "Rendering images from points can be a time-consuming task, especially when the number of rendering jobs is high. We propose a multi-scale synthetic scene-level image scheme." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "text", + "content": "Camera pose selection. The Snap module captures scene-level images at three scales: global, corner, and wide-angle, as shown in Figure 5. For global-level images, cameras are positioned above the scene and point directly toward the centre of the scene. For corner images, cameras are positioned above the centre and point toward the corner, while for the wide-angle images, cameras are positioned at the 3 by 3 grid interaction points, pointing toward the furthest corner. Using the camera position coordinates " + }, + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "text", + "content": ", target coordinates " + }, + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{target}}" + }, + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "text", + "content": ", and the up axis of the scene " + }, + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 449, + 480, + 568 + ], + "type": "text", + "content": ", the Lookat function can be employed to determine the pose matrix Pose. A more detailed mathematical formulation of this is presented in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "Camera intrinsic calibration. Once the camera extrinsic matrix is established, the fields of view for the captured images are adjusted by modifying the camera's intrinsic parameters. The goal is to ensure that the entire scene or a specific part of it is encompassed within the captured images. To achieve this goal, we initialize an arbitrary camera intrinsic matrix and then adjust the focal lengths " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "(f_x" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "f_y)" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " and the principal point coordinates " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "(c_x" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "c_y)" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " through scaling. The scaling is performed by readjusting the projected areas in the image coordinate space. For example: if the projected points of the pre-defined area" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 115, + 482, + 261 + ], + "blocks": [ + { + "bbox": [ + 153, + 115, + 482, + 261 + ], + "lines": [ + { + "bbox": [ + 153, + 115, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 153, + 115, + 482, + 261 + ], + "type": "image", + "image_path": "af8221a65488d1b74b92faed1c87ef1836ad9ca7513063a556a1967303d0672f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 348 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 348 + ], + "type": "text", + "content": "Fig.5: Snap and Mask2Pixel Maps. Multiscale snaps are conducted to render images with different levels of detail for scene understanding, including wide-corner snaps, wide-angle snaps, and global snaps. Cameras are positioned on the top of the scene and point towards the centre or corners, and the field of view is determined with the calibrated intrinsic matrix. With the defined camera models, Mask2Pixel maps are built to store the location of each 3D mask in the 2D image (using the same colour to represent 2D-3D correspondences) to guide the search for category names." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "content": "were located in image coordinates within the range of " + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "inline_equation", + "content": "[-1000, -192]" + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "content": "-domain, our calibrated intrinsic parameters transform this range to " + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "inline_equation", + "content": "[0, 1000]" + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "content": ". Importantly, we preserve the aspect ratio between the " + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 360, + 482, + 432 + ], + "type": "text", + "content": " coordinates to maintain the proportions of the final image without any distortion. This procedure ensures that each captured image is fully utilized and encompasses all regions of interest within the scene (Challenge 2 & 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 432, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 504 + ], + "type": "text", + "content": "Class lookup table. Upon obtaining " + }, + { + "bbox": [ + 130, + 432, + 482, + 504 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 432, + 482, + 504 + ], + "type": "text", + "content": " synthetic scene-level images, we input them into a 2D open-vocabulary detector. With text queries provided for interested classes, a list of detected objects in synthetic images can be obtained. Subsequently, information about detected objects, including their location and class, are stored in a designated Class Lookup Table (CLT). This table will later be retrieved to allocate class categories to 3D mask proposals (Challenge 4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 521, + 373, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 373, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 373, + 534 + ], + "type": "text", + "content": "Lookup: Mask Classification through Searching" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 569 + ], + "type": "text", + "content": "We conduct a multi-level search to assign category labels to the mask proposals generated from Mask step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 485, + 666 + ], + "type": "text", + "content": "Mask2Pixel guided lookup. We introduce a Mask2Pixel Guided Lookup (MGL) to search within CLT. The concept involves projecting each 3D mask proposal onto a 2D plane using the same camera extrinsic and intrinsic matrices that are utilized to generate the 2D image, as depicted in Figure 5. With knowledge of the precise pixel locations of each mask in images, we can conduct an accurate search through the CLT to identify the most likely class for each mask. The development of MPM takes into account occlusion by integrating depth information. To accomplish the matching, we follow a three-step approach: 1. based" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "on the mask's projection onto the 2D plane, we select the best-matched class categories in terms of IoU values; 2. if the IoU value of the best-matched object on the 2D plane is below " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ", the match is disregarded. 3. we aggregate results from multiple views to formulate the final prediction, calculating probability scores using their normalized average IoU values, as illustrated in Figure 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 298, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 298, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 298, + 390 + ], + "type": "text", + "content": "Local enforced lookup. While the Mask2Pixel Guided Lookup assigns class categories to mask proposals, some masks may not correspond to objects in the CLT. To address this, we introduce a Local Enforced Lookup (LEL) approach. We crop out the remaining masks from 2D scene-level images using enlarged bounding boxes and process them with the 2D detector to encourage detection. To select the best views, we introduce an Occlusion Report method to assess occlusion conditions for each mask in each projection, and then choose the top " + }, + { + "bbox": [ + 130, + 176, + 298, + 390 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 176, + 298, + 390 + ], + "type": "text", + "content": " views for LEL. More details for Occlusion Report can be found in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 391, + 298, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 298, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 298, + 462 + ], + "type": "text", + "content": "Final mask refinement. With the previous lookup approaches, a large proportion of mask proposals obtain a category prediction. All masks that have no category predictions after the MGL and LEL stages are eliminated." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 186, + 478, + 375 + ], + "blocks": [ + { + "bbox": [ + 309, + 186, + 478, + 375 + ], + "lines": [ + { + "bbox": [ + 309, + 186, + 478, + 375 + ], + "spans": [ + { + "bbox": [ + 309, + 186, + 478, + 375 + ], + "type": "image", + "image_path": "8398733af888cae47acd68594093de78c76a573e84a9b273f7fcbcfd6487bfa8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 387, + 482, + 442 + ], + "lines": [ + { + "bbox": [ + 304, + 387, + 482, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 482, + 442 + ], + "type": "text", + "content": "Fig. 6: Mask2Pixel Guided Lookup Illustration. IoUs between the 2D detection results and the projected masks are the guidance to assign class names to 3D masks. Multiple image results are ensembled." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": "Datasets and evaluation scheme. We tested OpenIns3D on five datasets, including four indoor datasets, namely S3DIS [2], ScanNetv2 [8], ScanNet200 [26], Replica [28], and one outdoor dataset, STPLS3D [5]. Among them, S3DIS, ScanNetv2, and ScanNet200 are indoor point cloud datasets generated from RGB-D images, Replica is a photo-realistic 3D indoor scene reconstruction, while STPLS3D is an aerial photogrammetry-constructed outdoor dataset. We exclusively used the 3D data with colour from these datasets and did not employ any 2D images, poses, or depth maps. Following the settings of prior work [9], we excluded the \"other furniture\" class in ScanNetv2 and the \"clutter\" class in S3DIS due to their vague meanings. Replica and ScanNet200 are evaluated by following the settings of OpenMask3D [30]. For STPLS3D, we merged the low, medium, and high vegetation classes into one \"vegetation\" class and kept all the rest." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Implementation details. For the S3DIS, ScanNetv2, Scannet200, and STPLS datasets, the MPM module is trained without utilizing any category labels, and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 165, + 480, + 245 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Zero-shot object classification on ScanNetv2. OpenIns3D's Snap and Lookup approach for mask classification, surpasses all previous methods, including the latest language-aligned large-scale 3D foundation model [40]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 165, + 480, + 245 + ], + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 245 + ], + "type": "table", + "html": "
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointCLIP [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
PointCLIP V2 [37]11.00.00.023.80.00.00.00.07.80.090.70.00.00.00.064.40.00.0
CLIP2Point [14]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP2[35]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
Uni3D [40]45.858.53.778.883.754.931.339.470.135.11.927.394.213.838.710.788.147.6
OpenIns3D60.885.227.487.677.346.954.864.271.49.980.882.771.661.438.70.087.985.7
", + "image_path": "5a58ddd1d6957f4a3ae0db4080c24410a4be6b0f74446b141dff21944c46f9f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "text", + "content": " is set to 0.1 to reduce the weight of zero-IoU. The mask proposal of Replica is trained on ScanNet200, followed by OpenMask3D [30]. The Snap module captures images with a size of " + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "inline_equation", + "content": "1000 \\times 1000" + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "text", + "content": " including 16 global snaps, 4 corner snaps, and 4 wide-angle snaps. The top " + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "text", + "content": " of the scene is removed for S3DIS, as the rooms are enclosed. For STPLIS3D, we followed Mask3D to split the large outdoor scene into patches of " + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "inline_equation", + "content": "50m \\times 50m" + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "text", + "content": " and lifted the camera up to " + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "inline_equation", + "content": "10m" + }, + { + "bbox": [ + 130, + 267, + 479, + 350 + ], + "type": "text", + "content": ". More implementation details are presented in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 480, + 460 + ], + "type": "text", + "content": "Open-vocabulary point cloud recognition. We first evaluate OpenIns3D performance among all existing 3D open-vocabulary models with the same setting, i.e. using only 3D inputs. These models are most commonly tested on recognition tasks, including PointCLIP(v1&v2) [43] [37], Clip2Point [14], CLIP² [35]. We also compare with large-scale 3D foundation models, such as Uni3D [40], which has 1-billion parameters, and was trained with large-scale 3D shapes and image-text pairs. We followed their evaluation scheme and reported the Top-1 accuracy of instance classification on ScanNetv2. The results are presented in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 480, + 592 + ], + "type": "text", + "content": "Open-vocabulary instance segmentation. We adopted various comparison schemes to align with existing methods. For 3D Open-vocabulary Instance Segmentation, we compared with PLA [9], and its follow-up works RegionPLC [34] and Lowis3D [10]. For a fair comparison, we followed their category splits and compared our results on novel classes, as demonstrated in Table 2. For STPLS3D, we compared OpenIns3D with baseline models whose classification module is PointCLIP and PointCLIPV2 [43] (Table 5). We also explored the performance of OpenIns3D on a more challenging dataset with more class categories. Specifically, we compared the performance of OpenIns3D with OpenMask3D, OpenScene, on Replica (Table 4) as well as ScanNet200 (Table 6). Following OpenMask3D, we used Mask3D to generate mask proposals for OpenScene for evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Open-vocabulary object detection. Since there were limited works on 3D open-vocabulary instance segmentation at the time of conducting this work, we also selected some of the latest methods in the 3D open-world object detection domain for a more comprehensive evaluation. The evaluation is performed by converting generated masks into axis-aligned bounding boxes, and the results are shown in Table 3." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 168, + 476, + 348 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: 3D Open-vocabulary Instance Segmentation Results on S3DIS and ScanNetv2. We compare our zero-shot performance on the novel categories defined in the PLA-family work. Significant improvements are achieved on the S3DIS dataset, and competitive results are observed on ScanNetv2 (B/N: Base/Novel)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 168, + 476, + 348 + ], + "lines": [ + { + "bbox": [ + 136, + 168, + 476, + 348 + ], + "spans": [ + { + "bbox": [ + 136, + 168, + 476, + 348 + ], + "type": "table", + "html": "
OVISS3DISScanNetv2
MethodB/NAP50AP25B/NAP50AP25require 2D
PLA [9]8/408.6-10/721.9-✓
RegionPLC [34]8/4--10/732.3-✓
Lowis3D [10]8/413.8-10/731.2-✓
Open3DIS [22]8/426.3-10/7--✓
Mask3d+PointClip [43]- / 405.410.3- / 704.507.8X
OpenIns3D- / 437.039.3- / 727.942.6X
improvement- / 4(+10.7)(+29.0)- / 7-(+34.8)X
PLA [9]6/609.8-8/925.1-✓
RegionPLC [34]6/6--8/932.2-✓
Lowis3D [10]6/615.8-8/938.1-✓
Open3DIS [22]6/629.0-8/9--✓
Mask3d+PointClip [43]- / 608.510.6- / 905.606.7X
OpenIns3D- / 633.038.9- / 919.527.9X
improvement- / 6(+4.0)(+28.3)- / 9-(+21.2)X
Mask3d+PointClip [43]- / 1208.609.3- / 1704.514.4X
OpenIns3D- / 1228.329.5- / 1728.738.9X
improvement- / 12(+19.7)(+20.2)- / 17(+24.2)(+24.5)X
", + "image_path": "959332eff066c6b5aa442a74270fab5510dce4d12128d6508a9eb0f00559ca1a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 363, + 487, + 460 + ], + "blocks": [ + { + "bbox": [ + 138, + 351, + 474, + 362 + ], + "lines": [ + { + "bbox": [ + 138, + 351, + 474, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 474, + 362 + ], + "type": "text", + "content": "Table 3: Open-vocabulary Object Detection " + }, + { + "bbox": [ + 138, + 351, + 474, + 362 + ], + "type": "inline_equation", + "content": "(AP_{25})" + }, + { + "bbox": [ + 138, + 351, + 474, + 362 + ], + "type": "text", + "content": " on unseen classes in ScanNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 363, + 487, + 460 + ], + "lines": [ + { + "bbox": [ + 133, + 363, + 487, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 363, + 487, + 460 + ], + "type": "table", + "html": "
Methodsmeantoiletbedchairsofadressertablecabinetbookshelfpillowsink
OV-PointCLIP [37]3.16.62.36.33.90.77.20.72.10.60.8
OV-Image2Point [32]0.80.20.81.01.40.22.81.00.90.00.1
Detic-ModelNet [41]1.74.21.04.61.20.23.20.61.20.00.7
Detic-ImageNet [41]0.40.00.00.20.00.51.80.50.30.00.7
OV-3DETIC [20]12.749.02.67.318.62.814.32.44.53.921.1
L3Det [42]24.656.336.216.123.08.123.114.717.323.427.9
FM-OV3D [36]21.555.038.819.241.923.83.50.46.017.48.8
OpenIns3D43.779.570.576.915.80.053.140.141.27.153.1
improvement(+19.1)(+23.2)(+31.7)(+57.7)--(+30)(+25.4)(+23.9)(+25.2)
", + "image_path": "033b301a8e4acfcf4fbc0b6e193c49b948d200570093cb609681b73eee187b6a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 483, + 290, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 290, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 290, + 495 + ], + "type": "text", + "content": "5 Results and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 511, + 281, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 281, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 281, + 522 + ], + "type": "text", + "content": "5.1 Comparison with SOTA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "First of all, OpenIns3D demonstrates impressive performance in open-vocabulary point cloud recognition, surpassing all previous methods, including the large-scale 3D foundation model by " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". This proves the effectiveness of the zero-shot \"Snap\" and \"lookup\" scheme. With the enhanced recognition capability, the performance of 3D open-vocabulary Object Detection among the ScanNet dataset has also achieved state-of-the-art results by a large margin. For 3D instance segmentation, compared to works in the PLA family [9, 10, 34] and the latest work Open3DIS [22], OpenIns3D does not require aligned images as input, (still need RGB information for points), achieves higher results on the S3DIS dataset, both in the 4 novel categories split and the 6 novel categories split. In SPTLS3D, OpenIns3D outperforms the baseline model PointCLIPV2 by " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "9.3\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " in AP. On" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 329, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 325, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 325, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 325, + 136 + ], + "type": "text", + "content": "Table 4: OVIS on the indoor Replica dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 329, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 329, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 329, + 239 + ], + "type": "table", + "html": "
Model2DAP\\( AP_{50} \\)\\( AP_{25} \\)
OpenScene [23] (2D Fusion)✓10.915.617.3
OpenScene [23] (2D/3D Ens.)✓8.210.413.3
OpenMask3D [30]✓13.118.424.2
OpenScene [23] (3D Distill)✗8.210.512.6
OpenIns3D✗13.618.019.7
improvement✗(+5.4)(+7.5)(+7.1)
", + "image_path": "3cc6073981b81418449e38e2c47aa7f0196dcec7b513fd7cbff6561248a81270.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 330, + 146, + 475, + 239 + ], + "blocks": [ + { + "bbox": [ + 330, + 114, + 471, + 136 + ], + "lines": [ + { + "bbox": [ + 330, + 114, + 471, + 136 + ], + "spans": [ + { + "bbox": [ + 330, + 114, + 471, + 136 + ], + "type": "text", + "content": "Table 5: OVIS on outdoor STPLS3D dataset" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 146, + 475, + 239 + ], + "lines": [ + { + "bbox": [ + 330, + 146, + 475, + 239 + ], + "spans": [ + { + "bbox": [ + 330, + 146, + 475, + 239 + ], + "type": "table", + "html": "
ModelAPAP50AP25
PointCLIP [43]02.002.604.0
PointCLIPv2 [37]02.103.105.2
OpenIns3D11.414.217.2
improvement(+9.3)(+11.9)(+12.0)
", + "image_path": "ab849c100e8f51877e20188bf17d56c86210d04251060644384940c679ecfe6e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 138, + 306, + 477, + 398 + ], + "blocks": [ + { + "bbox": [ + 131, + 252, + 481, + 296 + ], + "lines": [ + { + "bbox": [ + 131, + 252, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 481, + 296 + ], + "type": "text", + "content": "Table 6: 3D instance segmentation results on the ScanNet200 validation set. OpenIns3D demonstrates robust performance when compared to 2D-input-free models. However, notable limitations emerge when dealing with small objects in the common and tail classes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 306, + 477, + 398 + ], + "lines": [ + { + "bbox": [ + 138, + 306, + 477, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 477, + 398 + ], + "type": "table", + "html": "
Modeluse 2DAPheadAPcommonAPtailAPAP50AP25
OpenScene (2D Fusion) [23]✓13.411.69.911.715.217.8
OpenScene (2D/3D Ens.) [23]✓11.03.21.15.36.78.1
OpenMask3D✓17.114.114.915.419.923.1
OpenScene (3D Distill)✗10.62.60.74.86.27.2
OpenIns3D✗16.06.54.28.810.314.4
improvement✗(+5.4)(+3.9)(+3.6)(+4.0)(+4.1)(+7.2)
", + "image_path": "a4fac431bf628534861606e8f8675e334c58de44b67fb8f612e77846f5dd4028.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 517 + ], + "type": "text", + "content": "the Replica dataset, OpenIns3D even outperformed OpenMask3D, which relies on well-aligned images for mask understanding. In the case of ScanNet200, OpenIns3D attains the highest performance compared to all other 3D input baselines. However, we have observed a decline in performance on tail and common classes within ScanNet200. The decrease in performance can be attributed to the low-quality and unclear reconstruction of smaller objects in the ScanNet200 scene. This is a noticeable limitation of OpenIns3D, but for better reconstructions, such as in Replica, and for Head classes within ScanNet200, OpenIns3D still demonstrates decent performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 517, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 566 + ], + "type": "text", + "content": "In summary, OpenIns3D demonstrates the best performance among all existing methods if only 3D data is used as input and outperforms many existing state-of-the-art methods that require 2D images. It also shows certain limitations on small objects that are not well-reconstructed in 3D scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 584, + 237, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 237, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 237, + 597 + ], + "type": "text", + "content": "5.2 Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Mask quality ablation. Following the evaluation on ScanNetv2, we assessed the class-agnostic mask quality using the average precision (AP) score. We treated all classes as universal since the predictions are class-agnostic. The evaluation was conducted on the ScanNetv2 validation set. Table 8 demonstrates the effectiveness of the Mask Scoring and Mask Filtering designs." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 481, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 7: Rendering and Inference Time Ablations. Results tested on typical ScanNet scenes with 50 masks. OpenIns3D requires less rendering and inference time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 481, + 240 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 481, + 240 + ], + "type": "table", + "html": "
RenderingNum of Img needed2D backboneImg size (w × h)\\( T_{render} \\) (s/scene)\\( T_{infer} \\) (s/scene)\\( T_{total} \\) (s/scene)AP25 (%)
PointCLIP250CLIP\\( 128^2 \\)5.215.320.59.3
LAR250CLIP\\( 128^2 \\)14.318.733.010.5
Mask rendering250CLIP\\( 128^2 \\)42.619.562.17.3
OpenIns3D (ours)8G-DINO\\( 1000^2 \\)2.36.28.529.8
OpenIns3D (ours)8ODISE\\( 1000^2 \\)2.38.210.535.1
", + "image_path": "b64e81d3a19f420713b956449ae944df67733b6291ee320283f98d3234a2f426.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 281, + 321, + 354 + ], + "blocks": [ + { + "bbox": [ + 132, + 250, + 324, + 272 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 324, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 324, + 272 + ], + "type": "text", + "content": "Table 8: MPM Ablation. MS: Mask scoring. MF: Mask filtering." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 281, + 321, + 354 + ], + "lines": [ + { + "bbox": [ + 133, + 281, + 321, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 321, + 354 + ], + "type": "table", + "html": "
MethodAP50AP25
Mask3d-Supervised74.780.9
CA-Mask3d47.549.2
CA-Mask3d + MS50.2 (+02.7)53.3 (+04.1)
CA-Mask3d + MF61.6 (+14.1)71.0 (+21.8)
CA-Mask3d + MS + MF64.6 (+17.0)73.4 (+24.2)
", + "image_path": "71b5cb7a727fac59e27441d98f19b587e37cc1cfaa4533163211aa2c4a5ea9ae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 334, + 282, + 484, + 353 + ], + "blocks": [ + { + "bbox": [ + 330, + 250, + 488, + 273 + ], + "lines": [ + { + "bbox": [ + 330, + 250, + 488, + 273 + ], + "spans": [ + { + "bbox": [ + 330, + 250, + 488, + 273 + ], + "type": "text", + "content": "Table 9: Number of Views Abla-tion. LEL: Local Enforced Lookup" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 282, + 484, + 353 + ], + "lines": [ + { + "bbox": [ + 334, + 282, + 484, + 353 + ], + "spans": [ + { + "bbox": [ + 334, + 282, + 484, + 353 + ], + "type": "table", + "html": "
IDX4816LELAP50AP25
i✓18.327.1
ii✓22.735.1
iii✓24.837.5
iv✓✓28.738.9
", + "image_path": "c8ec054766ec7a02501d42b4dd8099de73847bd240a059147df6a64f749452d1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": "Multi-view ablation. We also studied the effects of using different numbers of views (Table 9). Increasing the number of views used in the Lookup module leads to better results. Additionally, Look Enforced Lookup provided a final improvement to the results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 409, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 480, + 517 + ], + "type": "text", + "content": "Projection and 2D backbone ablation. We conducted a comprehensive study on various rendering methods and their interaction with the 2D backbone to identify a suitable approach. We report the rendering time and inference performance for each method (Table 7), and more details can be found in the supplementary materials. The key observation is that the scene-level rendering and understanding approach excels in speed while also demonstrating strong performance. Switching from Grounding Dino [19] to the latest ODISE [33] also brings gains in performance, indicating that the OpenIns3D framework can easily benefit from the rapid development of 2D open-world detectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 519, + 480, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 480, + 602 + ], + "type": "text", + "content": "Cross-domain analysis. To evaluate the generalization capability of MPM across different domains, we trained and tested OpenIns3D on two different datasets, as shown in Table 10. The cross-domain models also demonstrate impressive performance on both datasets when compared with the baseline. Notably, within the 17 classes of ScanNetv2, 11 classes do not exist in S3DIS. OpenIns3D, trained on S3DIS, still achieves decent performance among these unseen classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Free-flow language capability. The snap and lookup scheme outsources the mask understanding tasks to a 2D vision-language model. Therefore when integrated with a 2D model powered by large language models (LLMs) like LISA [18], OpenIns3D can perform reasoning-based segmentation tasks (as depicted in Figure 7). For instance, when given the query \"pencilling down ideas during brain-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "storming,\" OpenIns3D accurately segments the whiteboard, while for \"furniture offers recreational enjoyment with friends,\" it precisely identifies and segments the pool table." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 165, + 475, + 322 + ], + "blocks": [ + { + "bbox": [ + 137, + 165, + 475, + 322 + ], + "lines": [ + { + "bbox": [ + 137, + 165, + 475, + 322 + ], + "spans": [ + { + "bbox": [ + 137, + 165, + 475, + 322 + ], + "type": "image", + "image_path": "1df998dcb8085692438928bd7bbade94c4fe596135cb01b58cdbdadfc6a06a98.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 327, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 361 + ], + "type": "text", + "content": "Fig. 7: Qualitative Results from OpenIns3D. OpenIns3D (ODISE) demonstrates the ability to manage a versatile vocabulary. OpenIns3D (LISA) can conduct 3D reasoning segmentation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 147, + 411, + 465, + 483 + ], + "blocks": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "type": "text", + "content": "Table 10: Cross-domain Ablation. We trained and tested OpenIns3D on two different datasets to examine its cross-domain capability. While S3DIS and ScanNetV2 have non-overlapping classes, OpenIns3D demonstrates decent generalization capability." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 411, + 465, + 483 + ], + "lines": [ + { + "bbox": [ + 147, + 411, + 465, + 483 + ], + "spans": [ + { + "bbox": [ + 147, + 411, + 465, + 483 + ], + "type": "table", + "html": "
TestModelTraining DataAP50AP25
ScanNetv2 [8]Mask3D-Pointclip [43]ScanNetv204.514.4
OpenIns3DScanNetv228.738.9
OpenIns3DS3DIS21.533.6
S3DIS [1]Mask3D-Pointclip [43]S3DIS03.506.8
OpenIns3DS3DIS28.329.5
OpenIns3DScanNetv214.219.8
", + "image_path": "4562aa1fd1f7de6fd3a1427677f75da2244989a0bf75c74df202f0d551aa2819.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Achieving 3D open-vocabulary scene understanding is a challenging task, primarily due to the lack of extensive 3D-text data. Currently, most work in this domain focuses on using 2D images to bridge the gap between 3D and language. This, however, not only requires a good alignment between 2D and 3D but also evolves slowly due to the significant effort needed for retraining when changing the 2D backbone. OpenIns3D introduces a new pipeline, i.e. Mask-Snap-Lookup, for this task. The Mask module generates authentic masks in the 3D domain, while Snap renders scene-level images in 2D domains, and the Lookup module links the results from 2D to 3D precisely. This pipeline requires no image input, achieves better performance, and can evolve seamlessly with 2D models without training. We hope our work will provide a fresh perspective for researchers working towards open-world 3D scene understanding." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "This work is supported by the Girton College Graduate Research Awards at the University of Cambridge, School of Technology, National Highways sponsored through EPSRC Centre for Doctoral Training, the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR, the National Natural Science Foundation of China No. 62201484, HKU Startup Fund, and HKU Seed Fund for Basic Research. We would like to express our gratitude to Yunhan Yang for his assistance in exploring rendering techniques and to Chengyao Wang for sharing his implementation of Mask3D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 255, + 197, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 255, + 197, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 255, + 197, + 267 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 280, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "text", + "content": "1. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 3, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 324 + ], + "type": "text", + "content": "2. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR (2016) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 480, + 346 + ], + "type": "text", + "content": "3. Bakr, E.M., Alsaedy, Y.Y., Elhoseiny, M.: Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. In: NeurIPS (2022) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "text", + "content": "5. Chen, M., Hu, Q., Yu, Z., THOMAS, H., Feng, A., Hou, Y., McCullough, K., Ren, F., Soibelman, L.: Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. In: BMVA (2022) 3, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 445 + ], + "type": "text", + "content": "6. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023) 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 446, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 467 + ], + "type": "text", + "content": "7. Chen, X., Li, S., Lim, S.N., Torralba, A., Zhao, H.: Open-vocabulary panoptic segmentation with embedding modulation. In: ICCV (2023) 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 489 + ], + "type": "text", + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 3, 9, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "9. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR (2023) 2, 4, 5, 9, 10, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 543 + ], + "type": "text", + "content": "0. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3d: Language-driven open-world instance-level 3d scene understanding. In: TPAMI (2024) 2, 4, 5, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "text", + "content": "1. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip. In: ICML (2023) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "2. Griffiths, D., Boehm, J.: SynthCity: A large-scale synthetic point cloud. In: arXiv (2019) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "3. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: Semantic3D.net: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "4. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV (2023) 4, 10" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "15. Huang, Z., Huang, L., Gong, Y., Huang, C., Wang, X.: Mask scoring R-CNN. In: CVPR (2019) 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "16. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "text", + "content": "17. Kundu, A., Yin, X., Fathi, A., Ross, D.A., Brewington, B., Funkhouser, T.A., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: ECCV (2020) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "text", + "content": "18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024) 4, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "text", + "content": "19. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: arXiv (2023) 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "text", + "content": "20. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR (2023) 3, 4, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 296, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 480, + 329 + ], + "type": "text", + "content": "21. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "text", + "content": "22. Nguyen, P.D.A., Ngo, T.D., Gan, C., Kalogerakis, E., Tran, A., Pham, C., Nguyen, K.: Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In: CVPR (2024) 2, 4, 5, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "23. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) 2, 4, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 430 + ], + "type": "text", + "content": "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 464 + ], + "type": "text", + "content": "25. Roynard, X., Deschaud, J.E., Goulette, F.: Paris-lille-3d: A large and high-quality ground-truth urban point cloud dataset for automatic segmentation and classification. In: The International booktitle of Robotics Research (2018) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 486 + ], + "type": "text", + "content": "26. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: ECCV (2022) 4, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "text", + "content": "27. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In: ICRA (2023) 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 510, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 575 + ], + "type": "text", + "content": "28. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces. In: arXiv (2019) 3, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "text", + "content": "29. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.G.: Multi-view convolutional neural networks for 3d shape recognition. In: ICCV (2015) 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 598, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 631 + ], + "type": "text", + "content": "30. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: Openmask3d: Open-vocabulary 3d instance segmentation. In: NeurIPS (2023) 3, 4, 5, 9, 10, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Tan, W., Qin, N., Ma, L., Li, Y., Du, J., Cai, G., Yang, K., Li, J.: Toronto-3D: A large-scale mobile lidar dataset for semantic segmentation of urban roadways. In: CVPRW (2020) 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z.Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 456 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Xu, C., Yang, S., Galanti, T., Wu, B., Yue, X., Zhai, B., Zhan, W., Vajda, P., Keutzer, K., Tomizuka, M.: Image2point: 3d point-cloud understanding with 2d image pretrained models. In: ECCV (2022) 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "33. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In: CVPR (2023) 2, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "text", + "content": "34. Yang, J., Ding, R., Wang, Z., Qi, X.: Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In: CVPR (2024) 2, 4, 5, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "text", + "content": "35. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: " + }, + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathrm{Clip}^2" + }, + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "text", + "content": ": Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR (2023) 4, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "36. Zhang, D., Li, C., Zhang, R., Xie, S., Xue, W., Xie, X., Zhang, S.: Fm-ov3d: Foundation model-based cross-modal knowledge blending for open-vocabulary 3d detection. In: AAAI (2024) 2, 3, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 312 + ], + "type": "text", + "content": "37. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by CLIP. In: CVPR (2022) 3, 4, 10, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "text", + "content": "38. Zhou, C., Loy, C.C., Dai, B.: Denseclip: Extract free dense labels from clip. In: CVPR (2022) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "text", + "content": "39. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "type": "text", + "content": "40. Zhou, J., Wang, J., Ma, B., Liu, Y.S., Huang, T., Wang, X.: Uni3d: Exploring unified 3d representation at scale. In: ICLR (2024) 2, 3, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "text", + "content": "41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "type": "text", + "content": "42. Zhu, C., Zhang, W., Wang, T., Liu, X., Chen, K.: Object2scene: Putting objects in context for open-vocabulary 3d detection. In: arXiv (2023) 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "text", + "content": "43. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV (2023) 2, 3, 4, 5, 10, 11, 12, 14" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenIns3D: 3D Open-vocabulary Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_content_list.json b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e05ad0c908a46ee508cc4586f6f54b38cb863411 --- /dev/null +++ b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_content_list.json @@ -0,0 +1,1837 @@ +[ + { + "type": "text", + "text": "OpenKD: Opening Prompt Diversity for Zero-and Few-shot Keypoint Detection", + "text_level": 1, + "bbox": [ + 236, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changsheng Lu1, Zheyuan Liu1, and Piotr Koniusz*,2,1", + "bbox": [ + 285, + 212, + 718, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Australian National University 2Data61/CSIRO changshengluu@gmail.com, zheyuan.david.liu@outlook.com, piotr.koniusz@data61.csiro.au", + "bbox": [ + 292, + 239, + 710, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Exploiting foundation models (e.g., CLIP) to build a versatile keypoint detector has gained increasing attention. Most existing models accept either the text prompt (e.g., \"the nose of a cat\"), or the visual prompt (e.g., support image with keypoint annotations), to detect the corresponding keypoints in query image, thereby, exhibiting either zero-shot or few-shot detection ability. However, the research on multimodal prompting is still underexplored, and the prompt diversity in semantics and language is far from opened. For example, how to handle unseen text prompts for novel keypoint detection and the diverse text prompts like \"Can you detect the nose and ears of a cat?\" In this work, we open the prompt diversity in three aspects: modality, semantics (seen vs. unseen), and language, to enable a more general zero- and few-shot keypoint detection (Z-FSKD). We propose a novel OpenKD model which leverages a multimodal prototype set to support both visual and textual prompting. Further, to infer the keypoint location of unseen texts, we add the auxiliary keypoints and texts interpolated in visual and textual domains into training, which improves the spatial reasoning of our model and significantly enhances zero-shot novel keypoint detection. We also find large language model (LLM) is a good parser, which achieves over $96\\%$ accuracy when parsing keypoints from texts. With LLM, OpenKD can handle diverse text prompts. Experimental results show that our method achieves state-of-the-art performance on Z-FSKD and initiates new ways of dealing with unseen text and diverse texts. The source code and data are available at https://github.com/AlanLuSun/OpenKD.", + "bbox": [ + 261, + 319, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 678, + 375, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keypoint detection is a fundamental research problem in computer vision and has numerous applications such as pose estimation for humans [6,8,12,33,37, 52] and animals [2,5,23,34], action recognition [30,47], and fine-grained image classification [39,55]. Over the past decade, significant advancements have been made in deep keypoint detection. However, most existing methods are tailored to close-set detection and struggle to predict novel body parts and species with limited data. Thus, it necessitates such a zero- or few-shot keypoint detection:", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ad9eb483430f672d537314cbf53fe745ba8da0ec13c63a2ff5371291e28867f.jpg", + "image_caption": [ + "(a)", + "Visual prompt", + "Query image" + ], + "image_footnote": [], + "bbox": [ + 248, + 146, + 305, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eeddedc453b57159ecbeca957bfdd1fecf49d36147cb73b3c82fe732cb6ba0dc.jpg", + "image_caption": [ + "(b)", + "Text prompt", + "Query image" + ], + "image_footnote": [], + "bbox": [ + 380, + 146, + 436, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4b3173369d2239a793bbb82c3b6e78217ec0d99280db7dfb7c9937b1d9e3e06b.jpg", + "image_caption": [ + "(c)", + "Visual + Text prompt" + ], + "image_footnote": [], + "bbox": [ + 550, + 146, + 604, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/191f43a2b19a7913c29425d4d0be443d6f81f9c025d80fffe649a8e9b925114f.jpg", + "image_caption": [ + "(d)", + "Results on AwA dataset One shot Zero shot One shot w/ text", + "Fig. 1: Illustration of multimodal prompting for keypoint detection. Our model can successfully detect keypoints given visual prompts formed by support images and keypoints (a), text prompts (b), or both (c). Graph (d) shows our model well combines the advantages of different modalities, mitigating the weakness induced by either modality." + ], + "image_footnote": [], + "bbox": [ + 651, + 185, + 776, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "after being trained on a diverse dataset, the model can quickly recognize novel or base keypoints in unseen species given only zero, one or a few labelled samples.", + "bbox": [ + 212, + 369, + 784, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the one hand, researchers have recently explored few-shot keypoint detection (FSKD) [27-29, 53]. The FSKD inspired by few-shot learning [13, 20, 36, 38, 44] is general and offers higher flexibility of detecting a varying number of keypoints in a query image, given the prompts formed by support image with keypoint annotations. In this regard, the support set is namely visual prompt, which is required when evaluating new classes. On the other hand, in the era of foundation model, the vision-language model (e.g., CLIP [35]) injects new life to detecting keypoints, enabling zero-shot keypoint detection (ZSKD) [54, 56]. Compared to FSKD, ZSKD does not need support image with annotations, instead, using the text prompt to instruct the model to detect the specified keypoints in query image. In this case, the support set is text prompt. Though impressive for the pioneering FSKD and ZSKD models, we identify an important issue limiting the progress: the prompt diversity for current keypoint detection methods is far from opened, especially in the aspects of modality, semantics, and language.", + "bbox": [ + 212, + 401, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modality diversity. Most existing keypoint detection models cannot support multimodal prompts, e.g., image, text, or both (see Fig. 1). Multimodal prompting is more friendly in real-world interaction, and coherent with the human concept recognition. We not only see the objects, but also describe objects with language. Our work extends existing FSKD and ZSKD, building a more general zero- and few-shot keypoint detection by leveraging a multimodal prototype set and aligning the visual keypoint features towards the textual features. While straightforward, our method enables one to study the advantages of respective modality data and exploit them for better model training and testing.", + "bbox": [ + 212, + 613, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic diversity. Considering keypoints between seen and unseen species, there exist large similarity yet difference in semantics. A strong advantage of text prompt is that the keypoints with the same semantics share high similarity in language across species, which enables excellent ZSKD on base keypoints. However, if the text has different semantics, significant domain gap arises. For instance, \"the eye of a cat\" vs. \"the eye of a dog\" has cosine similarity of 0.93 using CLIP", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/edc17f53fcff8c7a4630b46334b70b5498c5751eafb02e71876a3c5942cd8f39.jpg", + "image_caption": [ + "Can you make out the back-right knee, front-left knee, body-middle-left on cow?", + "(a)", + "Fig. 2: Examples of keypoint detection under diverse text prompting. With LLM, our method can deal with diverse texts, showing potential for real-world applications. The circles and crosses refer to predictions and GT, respectively. A keypoint is regarded as a correct detection if falling in the white shadow area that signifies PCK@0.1." + ], + "image_footnote": [], + "bbox": [ + 218, + 170, + 328, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4dfedab7274c9b8fb5b3e53d2855bc6555fa4f09036a74dbb7c777aea709bf2.jpg", + "image_caption": [ + "Can you detect the body-middle-right, nose, left earbase, front-right paw, left eye of grizzly bear?", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 333, + 170, + 442, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a2dfed2b56fc4d8c6f9153bb7fe7486675ef237eac93d87e8fe76088f34c307a.jpg", + "image_caption": [ + "Please find the throat, belly, breast, back, left leg for the evening grosbeak.", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 447, + 170, + 557, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb53182c4349e9bc99cd7d84c6f4d5c337d6351eebb881579c0292134217ab41.jpg", + "image_caption": [ + "Where is the breast, throat for pine warbler?", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 560, + 170, + 669, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4487459394e3567a13976a47814d8e0d171b141de22a36cbfaa6759b6a8988da.jpg", + "image_caption": [ + "Can you recognize the back, crown, throat, tail, belly?", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 674, + 170, + 784, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text embeddings, whereas for \"the eye of a cat\" vs. \"the knee of a dog\" it drops to 0.77. We observe that the keypoint detectors cannot perform well if text is unseen during training. To bridge the gap, we propose to open the semantic diversity by adding auxiliary keypoints and texts into training, where the auxiliary texts are reasoned by LLM given the base keypoints. To ease the reasoning error, we ask LLM a good question using chain of thought [48]. We also propose a novel sampling strategy with false text control to improve the matching between auxiliary keypoints and texts. With them together, our model significantly improves ZSKD on novel keypoints.", + "bbox": [ + 212, + 366, + 787, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language diversity. Following zero-shot image classification by CLIP [35], existing ZSKD also constructs simple text prompts based on templates, e.g., \"〈keypoint〉\" or \"the〈keypoint〉 of a〈category〉 in the photo\", where $\\langle \\cdot \\rangle$ is replaced by class names. In real-world interaction, humans tend to question in diverse styles, which results in diverse text prompts. For example, \"Please locate the right-back leg on cat\", \"Can you find the left eye, right ear and nose in image?\", etc. A natural question is how to handle the diverse text prompting? To address this, we propose a simple yet effective approach, whose key idea is to borrow the large language model (LLM) such as GPT3.5 [4] or Vicuna [9, 58] to parse the diverse texts via prompt engineering. After extracting the keypoint type and object category, they will be synthesized into simple text prompts to instruct the detection model. As such, the diverse text prompting can be transformed into simple text prompting, which sheds light on opening language diversity (Fig. 2).", + "bbox": [ + 212, + 505, + 787, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, in this paper, we propose an OpenKD model with several intriguing features: 1) supporting both visual and textual prompts, 2) having the potential to handle unseen texts and diverse texts, and 3) maintaining strong generality and performance on ZSKD and FSKD. We report that LLM is capable of being a reasoner for text interpolation, and a good language parser for parsing diverse texts. We also contribute four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42] for fair evaluations. To our best knowledge, we are the first to open semantics and language diversity of text prompts for keypoint detection.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keypoint Detection has been widely studied ranging from the traditional interest points [10, 25] to deep corner detection [57], semi-supervised [18, 31, 46] and fully-supervised keypoint estimation [6, 8, 12, 33, 37, 40, 51, 52]. There are two major classes of methods for deep keypoint localization: i) direct coordinates based regression [7, 41] and ii) heatmap based regression. In contrast to existing heatmap based models dedicated to specific body parts, e.g., top-down [12, 16, 37] and bottom-up pose estimators [6, 8, 33], our OpenKD model offers more flexible keypoint detection, breaking the limitation of keypoint types to be detected.", + "bbox": [ + 212, + 175, + 785, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Few-shot Keypoint Detection is more versatile and data-efficient compared to the supervised paradigms. Similar to standard few-shot learning (FSL) [44], FSKD also takes episodes for training and evaluation. Many well-known FSL methods such as ProtoNet [36], RelationNet [38], LwoF [15] and MAML [13] have been extended to the field of keypoint detection [3,14,27] and serve as baselines. Recently, Lu and Koniusz [27] formalize the comprehensive FSKD settings and model the localization uncertainty of keypoints. Further, they propose a lightweight FSKD model [28] and explore FSKD under transductive setting and occlusions [29]. Bohdal et al. [3] propose a dataset-of-datasets, benchmarking the FSL algorithms universal to various vision tasks. FSKD also inspires class-agnostic animal pose estimation [27, 28, 50]. Compared to existing FSKD, our work pushes further by opening the text prompting, which enables more versatile keypoint detection. Moreover, text prompts can complement visual prompts, delivering better performance of detecting base keypoints on unseen species.", + "bbox": [ + 212, + 296, + 785, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero-shot Keypoint Detection becomes feasible thanks to the great progress of vision-language pre-training (e.g., CLIP [35] and BLIP [22]). Since CLIP was pre-trained on internet-scale image-text pairs that contain more semantic concepts than other datasets, CLIP can provide semantically rich features for various downstream tasks [32, 56, 59], showing strong zero-shot transfer ability.", + "bbox": [ + 212, + 508, + 785, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, CLAMP [56] adapts CLIP for animal pose estimation in feature-aware and spatial-aware aspects; Approach [54] also employs CLIP as backbone and proposes language-driven keypoint detection. Compared to FSKD which still requires one or more annotations, ZSKD merely demands cheap language descriptions, which make it intriguing for keypoint detection. Nevertheless, most existing ZSKD works cannot support multimodal prompts and their prompt diversity is limited. To address these issues, we propose an OpenKD model which further opens modality, semantics, and language. To handle unseen texts, we propose to generate auxiliary keypoint-text pairs which significantly help reason novel keypoints on unseen species. Moreover, we contribute a simple yet effective language parsing module to handle diverse text prompts.", + "bbox": [ + 212, + 583, + 785, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Foundation Models include a broad set of models pre-trained over large-scale datasets, including LLMs (e.g., GPT3.5 [4]), VLMs (e.g., CLIP [35]), and Multimodal models (e.g., GPT4 [1]). Transferring the knowledge from these foundation models to specific tasks in a cost-effective way is popular in computer vision [19,24,26,54,56]. Thus, we follow this trend and leverage LLM like GPT3.5 (relatively cheap) to perform reasoning and parsing for keypoints.", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Model Architecture", + "text_level": 1, + "bbox": [ + 215, + 174, + 426, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the general few-shot learning [36, 38, 44], Z-FSKD model is evaluated on episodes, each of which includes a support set and a query set. The query set is comprised of query images, while the support set gives the prompts. If the visual prompts, i.e., $K$ support images with keypoint annotations ( $K \\geq 1$ ), are given, then the problem is defined as $K$ -shot detection. If only the text prompt (i.e., the language description) is given, then $K = 0$ thus the problem becomes zero-shot detection. The goal of Z-FSKD is to detect the corresponding keypoints in query image given the prompts, whether visual prompt, text prompt, or both. Such an approach allows the model to effectively respond to diverse inputs.", + "bbox": [ + 212, + 196, + 787, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overview of our model inference is shown in Fig. 3, which mainly includes four stages: i) image/text feature extraction, ii) feature adaptation, iii) keypoint prototype set building, iv) correlation, decoding and heatamp fusion. The overview of our model training is shown in Fig. 4, which includes the novel approaches for performance improvement.", + "bbox": [ + 212, + 333, + 787, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image/Text Feature Extraction Since the input of our model involves vision and language two different modalities, we wish the features extracted from image and text have smaller modality gap, so that the text can somewhat correlate with image regions of the same semantics. Such a property helps the model find keypoint locations in query image. To this end, we resort to CLIP [35] pre-trained on large-scale image-text pairs. The CLIP includes image and text encoders. The text encoder is generally a transformer [43] while the image encoder can be CNN [21] or ViT [11]. We empirically found that the CLIP image encoder based on RN50 [17, 35], a CNN model, gives high efficiency in consideration of both performance and cost. Thus, we choose it as the default backbone. When extracting the image features, original CLIP image encoder only retains the classification token $\\mathbf{x}_{\\mathrm{class}}$ and discards image tokens $\\mathbf{X}$ by attention pooling as", + "bbox": [ + 212, + 409, + 787, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\text {c l a s s}} = \\operatorname {A v g P o o l} (\\mathbf {X}), \\quad \\mathbf {x} _ {\\text {c l a s s}} ^ {\\prime} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {x} _ {\\text {c l a s s}} \\mathbf {W} _ {q}, \\mathbf {X} \\mathbf {W} _ {k}, \\mathbf {X} \\mathbf {W} _ {v}\\right) \\mathbf {W} _ {o}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 599, + 784, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While it is natural for image-text matching, we require the image tokens to recover the spatial locations of keypoints. Thus, we propose to obtain the image tokens via a projection composed by $\\mathbf{W}_v$ and $\\mathbf{W}_o$ as", + "bbox": [ + 212, + 623, + 787, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {\\prime} = \\mathbf {X} \\mathbf {W} _ {v} \\mathbf {W} _ {o}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 679, + 784, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we can not only handle the channel inconsistency between raw image tokens $\\mathbf{X}$ and text features, but also reuse the projection layers $\\mathbf{W}_v$ and $\\mathbf{W}_o$ .", + "bbox": [ + 212, + 704, + 785, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For brevity, we assume that per input episode contains one query image $\\mathbf{I}^{\\mathrm{q}}$ , one support image $\\mathbf{I}^{\\mathrm{s}}$ with $N$ annotated keypoints, and $N$ keypoint texts. With Eq. 2, both the support and query images can be encoded and projected as the support and query feature maps as $\\mathbf{X}^{\\mathrm{s}} = \\mathcal{F}_{\\mathrm{v}}(\\mathbf{I}^{\\mathrm{s}})$ and $\\mathbf{X}^{\\mathrm{q}} = \\mathcal{F}_{\\mathrm{v}}(\\mathbf{I}^{\\mathrm{q}})$ in deep feature space $\\mathbb{R}^{l\\times l\\times d}$ via a shared CLIP image encoder $\\mathcal{F}_{\\mathrm{v}}$ . Further, with the CLIP text encoder $\\mathcal{F}_{\\mathrm{t}}$ , the $N$ texts are firstly tokenized, then encoded as the text features $\\mathbf{t}_n\\in \\mathbb{R}^{m\\times d}$ , where $n = 1,2,\\dots ,N$ ; and $m$ is sequence length.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e71fb7a801ead380a8c60b5890eccc813b55be1ecd8c053e1b174e17a4e4a28d.jpg", + "image_caption": [ + "Fig. 3: The sketch of model inference. Our OpenKD allows testing under visual prompt, text prompt, or both. For clarification, we show the \"both\" case (i.e., 1-shot with text testing). We firstly extract the deep features of texts, support and query images via CLIP, and then adapt both modalities of features via residual refinement. After extracting the visual keypoint prototype (VKP) and textual counterpart, we build the prototype set to perform class-agnostic correlation and heatmap decoding. Finally, we fuse the heatmaps induced by two modalities (i.e., M1 & M2) to obtain predictions." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature Adaptation As CLIP provides general yet transferrable features, we adapt it to a new multimodal feature space suitable for keypoint detection. Inspired by residual learning [17], we propose two adaptation nets $\\mathcal{A}_{\\mathrm{v}}$ and $\\mathcal{A}_{\\mathrm{t}}$ to respectively refine the obtained image and text features in a residual way:", + "bbox": [ + 212, + 433, + 787, + 494 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {\\mathrm {s}} := \\mathbf {X} ^ {\\mathrm {s}} + \\mathcal {A} _ {\\mathrm {v}} \\left(\\mathbf {X} ^ {\\mathrm {s}}\\right), \\quad \\mathbf {X} ^ {\\mathrm {q}} := \\mathbf {X} ^ {\\mathrm {q}} + \\mathcal {A} _ {\\mathrm {v}} \\left(\\mathbf {X} ^ {\\mathrm {q}}\\right), \\quad \\mathbf {t} _ {n} := \\mathbf {t} _ {n} + \\mathcal {A} _ {\\mathrm {t}} \\left(\\mathbf {t} _ {n}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 506, + 785, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The adaptation net $\\mathcal{A}$ is highly scalable and can be based on transformer [43] or bottleneck [17] as refinement blocks. We found that $\\mathcal{A}$ can adapt features well.", + "bbox": [ + 212, + 534, + 787, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Keypoint Prototype Set Once we obtain the adapted features of support image and text prompts, i.e., $\\mathbf{X}^{\\mathrm{s}}$ and $\\mathbf{t}_n$ , we propose to convert them into keypoint prototypes, which are unified in a keypoint prototype set.", + "bbox": [ + 212, + 568, + 787, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For visual prompt, recall that keypoint labels are provided along with the support image. Thus, one can aggregate the local features for each support keypoint $\\mathbf{p}_n$ by using pixelwise weighted summation between feature map $\\mathbf{X}^{\\mathrm{s}}$ and Gaussian heatmap $\\mathbf{H}(\\mathbf{p}_n;\\sigma)$ , yielding the visual keypoint representation (VKR) as $\\Phi_n\\in \\mathbb{R}^d$ . The $\\sigma$ is the standard deviation that controls the Gaussian spread. If $K$ support images are given, namely in few-shot case, the VKRs of the same type of keypoints $\\Phi_{k,n}$ will be averaged to build the visual keypoint prototype (VKP) $\\Psi_{n}^{\\mathrm{v}} = \\frac{1}{K}\\sum_{k}\\Phi_{k,n}$ . Analogously, one can develop textual keypoint prototype (TKP) $\\Psi_{n}^{\\mathrm{t}}$ by text features if provided with multiple texts per keypoint. Finally, one can build the keypoint prototype set as $\\mathcal{T} = \\mathcal{T}^{\\mathrm{v}}\\cup \\mathcal{T}^{\\mathrm{t}}$ , where $\\mathcal{T}^{\\mathrm{v}} = \\{\\Psi_n^{\\mathrm{v}}\\}$ is the VKPs and $\\mathcal{T}^{\\mathrm{t}} = \\{\\Psi_n^{\\mathrm{t}}\\}$ is the TKPs. In this way, prompts of different modalities can be summarized as keypoint prototypes in a shared $d$ dimensional feature space, which enables our model to flexibly handle various modalities. Moreover, each prototype in set $\\mathcal{T}$ can guide the model to induce a keypoint heatmap, establishing the zero- and few-shot keypoint detection.", + "bbox": [ + 212, + 614, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81028be5d25f5ed9afb43f7501320e54413ea1abb266a4d015de338622d577d6.jpg", + "image_caption": [ + "(a) Model training" + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 485, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99d668236a6c35de77c09fab149453a4ae2da4e27dd33cb61065baac7cfa5e11.jpg", + "image_caption": [ + "(b) Auxiliary text interpolation", + "Fig. 4: Model training and text interpolation. (a) In addition to multi-group heatmap regression, we improve model performance by introducing intra- and inter-modality contrastive learning and the novel auxiliary keypoint and text learning. (b) We exploit LLM for auxiliary texts interpolation and explore incorporating visual keypoint features for text selection in order to mitigate the noisy texts or false texts." + ], + "image_footnote": [], + "bbox": [ + 488, + 150, + 785, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Correlation, Decoding and Fusion To discover the corresponding keypoints in query image $\\mathbf{I}^{\\mathrm{q}}$ , each keypoint prototype $\\Psi_{n}$ , either VKP $\\Psi_{n}^{\\mathrm{v}}$ or TKP $\\Psi_{n}^{\\mathrm{t}}$ , is required to be correlated with the query feature map $\\mathbf{X}^{\\mathrm{q}}$ . To this end, we leverage a correlation module $\\mathcal{C}$ which takes both $\\Psi_{n}$ and $\\mathbf{X}^{\\mathrm{q}}$ as input to produce the attentive feature maps $\\mathbf{A}_{n} = \\mathcal{C}(\\Psi_{n},\\mathbf{X}^{\\mathrm{q}})$ , where $\\mathbf{A}_{n}\\in \\mathbb{R}^{l\\times l\\times d}$ and $n = 1,2,\\dots ,N$ . For the correlation module $\\mathcal{C}$ , we explored multiple variants and found that the simple cross correlation already yields high performance, i.e., $\\mathbf{A}_n = \\mathbf{X}^{\\mathrm{q}}\\odot \\Psi_n$ , where $\\odot$ denotes the channel-wise multiplication.", + "bbox": [ + 212, + 393, + 787, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subsequently, a class-agnostic decoder $\\mathcal{D}$ is devised to convert each attentive feature map into a keypoint heatmap, i.e., $\\mathbf{H}_n = \\mathcal{D}(\\mathbf{A}_n)$ , where $\\mathbf{H}_n \\in \\mathbb{R}^{l \\times l}$ . For heatmap regression based keypoint localization, the higher resolution of heatmap can greatly reduce the coordinate decoding error. Thus, a lightweight upsampling module $\\mathcal{U}$ is adopted to further refine heatmaps, i.e., $\\mathbf{H}_n := \\mathcal{U}(\\mathbf{H}_n)$ .", + "bbox": [ + 212, + 517, + 787, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since each modality of prototypes can induce one group of heatmaps, considering visual and textual modalities, we have two groups of heatmaps $\\mathbf{H}^{\\mathrm{v}} = \\{\\mathbf{H}_n^{\\mathrm{v}}\\}_{n=1}^N$ and $\\mathbf{H}^{\\mathrm{t}} = \\{\\mathbf{H}_n^{\\mathrm{t}}\\}_{n=1}^N$ . In testing phase, we fuse the upsampled multi-group heatmaps as the final output $\\mathbf{H} \\in \\mathbb{R}^{N \\times ul \\times ul}$ ( $u$ is upsampling factor):", + "bbox": [ + 212, + 592, + 787, + 654 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\left(\\mathbf {H} ^ {\\mathrm {v}} + \\mathbf {H} ^ {\\mathrm {t}}\\right) / 2. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 666, + 784, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training phase, inspired by HigherHRNet [8], we perform group-specific supervision for heatmaps as", + "bbox": [ + 212, + 689, + 787, + 720 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k p}} = \\left(\\| \\mathbf {H} ^ {\\mathrm {v}} - \\mathbf {H} ^ {*} \\| _ {2} ^ {2} + \\| \\mathbf {H} ^ {\\mathrm {t}} - \\mathbf {H} ^ {*} \\| _ {2} ^ {2}\\right) / 2, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 731, + 784, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{H}^*$ denotes the groundtruth heatmap that encodes the query keypoint.", + "bbox": [ + 212, + 760, + 772, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Improving Model by Contrastive Learning Since the keypoint prototype directly impacts performance, we aim for its representativeness to be as good as possible. The same type of keypoints should have higher invariance across species, e.g., the ear of cat, dog, and cow, while different types are required to", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be sufficiently discriminative such that our model can distinguish the ambiguous texts or fine-grained texts such as \"the left ear\" vs. \"the right ear\". To this end, we propose to introduce a contrative loss over TKPs. If randomly sampling two species $(s,s^{\\prime})$ at a time and each species pertaining to an episode, we have pairwise sets of TKPs, i.e., $\\mathcal{T}_s^{\\mathrm{t}} = \\{\\Psi_n\\}$ and $\\mathcal{T}_{s^{\\prime}}^{\\mathrm{t}} = \\{\\Psi_{n}^{\\prime}\\}$ , which can form similarity matrix as", + "bbox": [ + 212, + 146, + 787, + 233 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} \\left(\\mathcal {T} _ {s} ^ {\\mathrm {t}}, \\mathcal {T} _ {s ^ {\\prime}} ^ {\\mathrm {t}}\\right) = \\left( \\begin{array}{c c} \\cos \\left(\\Psi_ {1}, \\Psi_ {1} ^ {\\prime}\\right) & \\dots \\cos \\left(\\Psi_ {1}, \\Psi_ {N} ^ {\\prime}\\right) \\\\ \\vdots & \\vdots \\\\ \\cos \\left(\\Psi_ {N}, \\Psi_ {1} ^ {\\prime}\\right) & \\dots \\cos \\left(\\Psi_ {N}, \\Psi_ {N} ^ {\\prime}\\right) \\end{array} \\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 233, + 785, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ denotes the cosine similarity. Then, the contrastive loss within textual modality $\\mathcal{L}_{\\mathrm{tt}}$ becomes", + "bbox": [ + 214, + 294, + 787, + 325 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t t}} ^ {s \\rightarrow s ^ {\\prime}} = - \\langle \\mathbb {I}, \\log (\\operatorname {s o f t m a x} (\\mathbf {J} (\\mathcal {T} _ {s} ^ {\\mathrm {t}}, \\mathcal {T} _ {s ^ {\\prime}} ^ {\\mathrm {t}}) / \\tau)) \\rangle , \\quad \\mathcal {L} _ {\\mathrm {t t}} = \\frac {1}{2} (\\mathcal {L} _ {\\mathrm {t t}} ^ {s \\rightarrow s ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {t t}} ^ {s ^ {\\prime} \\rightarrow s}), \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 335, + 785, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbb{I}$ is the identity matrix and $\\langle \\cdot ,\\cdot \\rangle$ denotes inner product. Since CLIP is trained via image-level alignment instead of keypoint-level alignment with text, we wish VKPs and TKPs were aligned so that we can exploit their respective advantages to help each other. This motivates us to propose the second contrastive loss between VKPs and TKPs. Similar to Eq. 7, we build $\\mathcal{L}_{\\mathrm{vt}}$ . We find that the visual-textual alignment improves 1-shot testing, but unintentionally degrades 0-shot performance slightly. We discover the principle behind this is better clustering effects of textual keypoint representations than visual ones. To address this issue, we perform stop gradient on TKPs in $\\mathcal{L}_{\\mathrm{vt}}$ , thus enforcing the VKPs to align towards TKPs. We also find adding $\\mathcal{L}_{\\mathrm{vv}}$ does not further improve scores. We conjecture that after aligning VKPs to TKPs via $\\mathcal{L}_{\\mathrm{vt}}$ , the visual features will also improve discrimination along with textual features as $\\mathcal{L}_{\\mathrm{tt}}$ is applied.", + "bbox": [ + 212, + 373, + 787, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization By incorporating the heatmap regression loss $\\mathcal{L}_{\\mathrm{kp}}$ , intra- and inter-modality contrastive loss $\\mathcal{L}_{\\mathrm{tt}}$ and $\\mathcal{L}_{\\mathrm{vt}}$ , we have the overall loss as $\\mathcal{L} = \\lambda_1\\mathcal{L}_{\\mathrm{kp}} + \\lambda_2\\mathcal{L}_{\\mathrm{tt}} + \\lambda_3\\mathcal{L}_{\\mathrm{vt}}$ . We set the loss weight $\\lambda_{1} = 1$ , $\\lambda_{2} = \\lambda_{3} = 0.002$ .", + "bbox": [ + 214, + 559, + 789, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Training with Auxiliary Keypoints and Texts", + "text_level": 1, + "bbox": [ + 214, + 626, + 635, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Localizing the seen keypoints and texts is relatively easy to handle, but it becomes difficult if the unseen ones are encountered. Thus, we propose to add auxiliary keypoints and texts interpolated from visual and textual domains into training, which significantly improves the ability to detect novel keypoints.", + "bbox": [ + 212, + 651, + 787, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Interpolation Following [27], we generate the auxiliary keypoints in visual domain via interpolation $\\mathcal{I}^{\\mathrm{v}}(z;\\mathbf{p}_1,\\mathbf{p}_2)$ , where $z\\in (0,1)$ is the interpolation node and $\\mathbf{p}_i\\in \\mathbb{R}^2$ are the end points that form an interpolation path. An off-the-shelf saliency detector [49] is adopted to filter the auxiliary keypoints outside of the foreground [27]. Via interpolation $\\mathcal{I}^{\\mathrm{v}}$ in the visual domain, the auxiliary keypoints $\\hat{\\mathbf{p}}$ can boost the visual diversity of training keypoints.", + "bbox": [ + 212, + 715, + 787, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Interpolation and Selection The $\\mathcal{L}^{\\nu}$ can assist the model in handling unseen keypoints in visual prompt, but it helps little if prompting unseen texts.", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consequently, we propose using text interpolation $\\mathcal{I}^{\\mathrm{t}}(z;t(\\mathbf{p}_1),t(\\mathbf{p}_2),c)$ to mitigate this issue, where $t(\\mathbf{p}_i)$ are the names of end points $\\mathbf{p}_i$ and $c$ is the object category. Thanks to great language processing ability of LLM (e.g., GPT3.5 [4]), we propose to ask LLM a good question to reason the auxiliary keypoint texts given base keypoints (Fig. 4(b)), e.g., to infer the keypoint text between \"high\" and \"paw\" of a \"cat\", we can design a vanilla prompt for the LLM as:", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "\"Please give me three most common body parts/ keypoints at $1 / 2$ between thigh and paw of a cat. Please answer in concise words.\"", + "bbox": [ + 212, + 237, + 782, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Apparently, we expect that the answers returned by LLM should include \"knee\" or \"elbow\". In practice, the keypoint types are usually more complex thus the reasoning is challenging. To enhance reasoning, we improve the above vanilla prompt into a more advanced one by using chain of thought (CoT) [48], whose key idea is to firstly provide an example with analysis and then ask the question. We provide the detailed prompts and examples in §A of Suppl.", + "bbox": [ + 212, + 267, + 784, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Considering the LLM suffers from some randomness and might produce erroneous answers, we let LLM give three most possible answers for each interpolation path and repeat this step $R$ times, yielding a text pool with $3R$ texts, i.e., $\\{\\hat{t}_i\\}_{i=1}^{3R}$ . As such, the GT text has higher chance to be included. Afterwards, we need to select one interpolated text from the pool to form a pair with a visually interpolated keypoint. To this end, a simple yet effective approach is to sample from top-1 results of all repetitions. However, it does not use the visual features for selection. An interesting approach is to use the correlation between visual and text features for selection as follows:", + "bbox": [ + 212, + 358, + 784, + 492 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nj = \\arg \\max _ {i} \\cos (\\hat {\\boldsymbol {\\Phi}}, \\hat {\\mathbf {t}} _ {i}), \\quad \\hat {t} ^ {*} = \\hat {t} _ {j}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 503, + 784, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{\\Phi}$ is the auxiliary keypoint feature of $\\hat{\\mathbf{p}}$ , $\\hat{\\mathbf{t}}_i$ is text feature of $\\hat{t}_i$ , and $\\hat{t}^*$ is the selected text. However, we found that sometimes it leads to performance drop. We suspect it is because CLIP's keypoint-text features are not fine-grained aligned enough on those regions beyond annotations. Opposed to picking, we observe such a correlation could be used to reject low-quality text if keypoint-text similarity $\\cos (\\hat{\\Phi},\\hat{\\mathbf{t}}_i)$ is low. Consequently, we propose a more advanced sampling with false text control (FTC). Namely, we sample the text within the range top- $\\eta$ but the text will be rejected if the similarity $\\cos (\\hat{\\Phi},\\hat{\\mathbf{t}}_i)$ is below the threshold $\\alpha$ . In this way, we can form auxiliary keypoint-text pairs with higher quality to enhance model training. We apply $\\mathcal{L}^{\\mathrm{v}}$ to both support and query images to generate support and query auxiliary keypoints $\\hat{\\mathbf{p}}^{\\mathrm{s}}$ and $\\hat{\\mathbf{p}}^{\\mathrm{q}}$ , respectively. As in Eq. 5, $\\hat{\\mathbf{p}}^{\\mathrm{q}}$ encodes GT heatmap supervising the heatmaps induced by $\\hat{\\mathbf{p}}^{\\mathrm{s}}$ and $\\hat{t}^*$ .", + "bbox": [ + 212, + 534, + 785, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 LLM is a Good Language Parser", + "text_level": 1, + "bbox": [ + 212, + 739, + 532, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To handle diverse text prompts in keypoint detection, unlike current fashion of utilizing LLMs for text generation, we use LLM for text decomposition. For example, if our model receives a diverse text prompt like \"Can you localize the left eye and nose of cat?\", we firstly leverage an LLM (e.g., GPT3.5 [4]) to perform text parsing by providing it with a simple yet effective prompt as follows:", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Please extract the animal and keypoint keywords from the below text: \"Can you localize the left eye and nose of cat?\"", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Then, LLM returns the parsed texts regarding ⟨keypoint⟩ and ⟨object⟩, i.e., the \"left eye\", \"nose\", and \"cat\" in the example. We directly utilize parsed results to synthesize simple text prompts to instruct our model and predict all the keypoints corresponding to the text. The problem of diverse text prompting becomes simple text prompting, thus opening the language diversity for keypoint detection.", + "bbox": [ + 212, + 176, + 782, + 250 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To fairly evaluate our approach, we manually collect 100 diverse text templates which cover most scenarios for questions ( $\\S \\mathbf{B}$ of Suppl.). Take one as an example: \"Where is the for . Afterwards, we develop four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42]. Specifically, we accompany each object instance with a diverse text prompt synthesized by randomly sampling one template and one to $N$ valid keypoints. Compared to traditional text prompts which only have $N$ variants, our constructed text prompt sets are more diverse, whose space could be as large as $100 \\cdot (C_N^1 + C_N^2 + \\dots + C_N^N) = 100 \\cdot 2^N$ variants. For example, as CUB has 11 keypoints, its space is larger than $10^5$ . We randomly sample 1000 diverse texts from each dataset and evaluate the LLM's parsing performance. Table 5b reports that GPT3.5 has over $96\\%$ accuracy in parsing keypoints from text, which shows that LLM can be a good language parser.", + "bbox": [ + 212, + 252, + 784, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 469, + 375, + 486 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 500, + 444, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets and Splits. 1) Animal pose dataset [5] has five mammal species cat, dog, cow, horse, and sheep. Each animal species is alternately chosen as unseen species for testing while the remaining four as seen species for training, which yields five subproblems; 2) AwA [2] has 35 diverse animal species with 10064 images. For AwA, 25 species are for training and 10 for testing; 3) CUB [45] consists of 200 bird species with 15 keypoint annotations. We use 100 species for training, 50 for validation, and 50 for testing; 4) NABird [42] is larger than CUB and has 555 categories. The species splits are 333, 111, and 111 for training, validation, and testing. To test model generalization, we follow [27] to split keypoints into base and novel sets, and report the performance of both base and novel keypoint detection on unseen species.", + "bbox": [ + 212, + 522, + 782, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metric. The percentage of correct keypoints (PCK) is used. A predicted keypoint is correct if its distance to GT $d \\leq \\rho \\cdot \\max(w_{\\mathrm{bbx}}, h_{\\mathrm{bbx}})$ , where $w_{\\mathrm{bbx}}$ and $h_{\\mathrm{bbx}}$ are the edges of object bounding box. Following [27], we set $\\rho$ to 0.1.", + "bbox": [ + 212, + 688, + 782, + 732 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. The input image size for all models is $384 \\times 384$ . We freeze the CLIP text encoder but finetune last two layers of image encoder. The temperature $\\tau$ used in contrastive loss is 0.05. Same to [27], the visual auxiliary keypoints are generated on pre-defined interpolation path, and we set $z = 0.5$ . For the text interpolation, to reduce randomness, we use GPT3.5 to reason $R = 3$ times for all datasets except CUB with $R = 10$ . Our model is trained with 40k episodes. We report results using 1000 test episodes.", + "bbox": [ + 212, + 734, + 782, + 837 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c28d6d353b412283ce4bb914ec76643db4e4d35a68b5e0f6299f3da800681cd0.jpg", + "table_caption": [ + "Table 1: Main results on the Animal pose dataset. Each row shows the results tested on the model trained in one setting. Each PCK score is the average of five subproblems." + ], + "table_footnote": [], + "table_body": "
#Training settings1-shot testing0-shot testing1-shot w/ text
KpAux. kpTextAux. textNovelBaseNovelBaseNovelBase
1✓21.3650.841.261.9316.1632.81
2✓✓47.5449.452.002.1039.2434.11
3✓16.1835.4625.6061.6427.0261.21
4✓✓15.8731.5763.1465.3163.3065.52
5✓✓21.4652.1522.4261.0723.8860.49
6✓✓✓✓50.3254.3963.3765.5963.1964.93
", + "bbox": [ + 220, + 178, + 782, + 319 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Compared Methods. For few-shot keypoint detection (FSKD), as in previous works, we compare the few-shot learning models such as ProtoNet [36], RelationNet [38], and LwoF [15], and FSKD-dedicated works $FSKD-R/-D$ [27]. For zero-shot keypoint detection (ZSKD), we adopt the source code of CLAMP [56] and compare it in our experiments for fairness. Moreover, we build a Baseline sharing same backbone with our model but we do not use the specifically proposed auxiliary texts, contrastive loss, etc. We denote our method as OpenKD.", + "bbox": [ + 212, + 344, + 784, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 469, + 372, + 483 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Firstly, we explore single-modal training and the impacts of whether or not to add the auxiliary keypoints/texts (1st-4th row, Table 1). When only using main keypoints (1st row) or main texts (3rd row), one can observe that the models perform well on base keypoints in 1-shot $(50.84\\%)$ or 0-shot $(61.64\\%)$ , but fail to detect novel keypoints. However, if the model is trained by adding the auxiliary keypoints (2nd row) or auxiliary texts (4th row), one can observe the remarkable improvements on novel keypoint detection, e.g., $47.54\\%$ vs. $21.36\\%$ in 1-shot testing for the model trained by visual prompts (2nd row vs. 1st row); and also $63.14\\%$ vs. $25.60\\%$ in 0-shot testing for the model trained by text prompts (4th row vs. 3rd row). The significant gains confirm the benefits of visual interpolation [27] and also highlight the effectiveness of our proposed textual interpolation. Secondly, we investigate multimodal training (5th-6th row, Table 1). As one can see, our model performs well under multimodal training (6th row), and greatly outperforms the model trained by multimodal prompts without auxiliary keypoints and texts (5th row). The higher performance is due to the fact that auxiliary keypoints and texts could boost the visual and textual semantic diversity, thus enabling novel keypoint detection under 1-shot and 0-shot setting. Thirdly, we observe that 0-shot testing on base keypoints is much higher than 1-shot, which shows that the texts can be an excellent representation to guide keypoint detection. Lastly, we examine multimodal testing, i.e., performing 1-shot with text testing, our model obtains $63.19\\%$ and $64.93\\%$ on novel and base keypoints, which highlights that our model strongly combines the advantages of both modalities, mitigating the weakness of individual modality.", + "bbox": [ + 212, + 492, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5131b6f9966a66cc7c33467c6f8dcd4cdec40b569a3afe3f4eddc6a304d22994.jpg", + "table_caption": [ + "Table 2: Results on 1-shot keypoint detection for unseen species. The PCK scores on novel and base keypoint detection are reported. $+ \\mathbf{T}$ means adding texts during testing." + ], + "table_footnote": [], + "table_body": "
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelProtoNet19.6816.1814.3912.0515.0615.4729.5751.3236.65
RelationNet22.1517.1915.4713.5816.5516.9920.9156.5934.02
LwoF22.4719.3916.8216.4016.9418.4028.5454.7534.19
FSKD-R46.0540.6637.5538.0931.5038.7751.8177.9054.01
FSKD-D52.3647.9444.0742.7736.6044.7564.7677.8956.04
OpenKD60.3653.5847.5949.0141.0550.3266.7178.3953.35
OpenKD+T69.2666.8162.4063.2154.2763.1979.0273.2953.40
BaseProtoNet45.8039.8334.8835.8032.3337.7357.1780.3673.18
RelationNet51.0345.8539.8641.9737.1943.1857.3179.4078.85
LwoF50.0544.6443.4743.3537.8443.8763.8781.9681.39
FSKD-R57.1251.1247.8349.7143.7149.9065.2687.9487.84
FSKD-D56.3851.2948.2449.7743.9549.9366.3987.7186.99
OpenKD63.6155.4351.1853.8747.8654.3974.2287.4585.11
OpenKD+T70.4765.0962.8466.4659.8164.9384.5091.8191.23
", + "bbox": [ + 222, + 178, + 781, + 429 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1606a2e6a8576e7f72e179c5beddb8a255372473fd82c690e22f03b3f847913f.jpg", + "table_caption": [ + "Table 3: Results on the 0-shot keypoint detection for unseen species. $\\mathrm{CLAMP}^{\\dagger}$ is the variant trained by adding our interpolated auxiliary texts." + ], + "table_footnote": [], + "table_body": "
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelBaseline25.6424.2925.2918.1425.5123.7728.9534.9430.00
CLAMP20.9024.0627.0716.8620.7321.9238.9637.0918.18
\\( CLAMP^† \\)61.7058.0961.4264.8653.1359.8477.6669.3050.81
OpenKD71.0766.4960.3065.5353.4563.3778.6470.1652.29
BaseBaseline58.6956.6256.5860.0051.3556.6573.4587.1271.25
CLAMP60.7154.0462.5862.7357.3059.4784.1690.9779.10
\\( CLAMP^† \\)60.3056.4962.6960.9057.1659.5183.7690.6583.65
OpenKD71.3466.6064.5367.2658.2365.5984.3291.7290.63
", + "bbox": [ + 223, + 491, + 776, + 660 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparisons on FSKD and ZSKD: We comprehensively conduct the few-shot keypoint detection across four datasets. Table 2 shows our OpenKD model significantly outperforms compared methods in 15 out of 18 tasks in detecting novel or base keypoints. Moreover, if adding the texts during testing, the performance of our model further improves, yielding $63.19\\%$ on the Animal pose dataset (63.19% vs. 44.75% of prior-art FSKD-D) and $79.02\\%$ on AwA. We also notice the improvements on CUB and NAB are modest, which might be due to the auxiliary texts being relatively harder to reason, thus bringing lesser benefit to few-shot model. For ZSKD (Table 3), again, our model outperforms other", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dcde36d53e7d806056951f660123c52190278beee66b2b9b4856c02bb25c78e5.jpg", + "image_caption": [ + "Fig. 5: Different clustering effects of two modalities on base keypoints in unseen species (a) and statistical feature variance per keypoint (b)." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 370, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/55dbc83131d445db9e8286c416cb2ef640096370e80f1f1dad30ea45ca477827.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 372, + 145, + 527, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/46f8128bb1ff7c6c139ae5dac6acf88a1ff74087fb9a1b4a164fb39ba1efe543.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 146, + 591, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a9b85bc8269ff46d1daa6dfa3506600b87d635c5837f8e441dbcb02e4e1d8fe.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 614, + 145, + 777, + 255 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ad36af4f4fcc3497856be9290efb448927b4e4c1ee32c7ed0afbf94ec8ae01df.jpg", + "table_caption": [ + "Table 4: Ablation study. (a) Intra- and inter-modality contrastive loss. $\\mathcal{L}_{\\mathrm{vt}}^*$ denotes no stop gradient on text features; (b) Text interpolation. Baseline†: w/o using auxiliary texts; CoT: chain of thought; Corr: Eq. 8; FTC: sampling with false text control (ours)." + ], + "table_footnote": [], + "table_body": "
AwA1-shot0-shot
NovelBaseNovelBase
1: w/o CL65.5672.1776.7181.67
2: Lt66.0572.3578.2884.00
3: Ltt+L*vt66.7073.9177.1083.85
4: Ltt+Lvt66.7174.2278.6484.32
5: Ltt+Lvt+Lvv66.2774.0878.6083.42
", + "bbox": [ + 225, + 358, + 493, + 463 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/cf4458a4c6364c25f16ab4f8133204d1881ab0f5b520580805c03d4b31f80128.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
0-shot testingAwACUB
NovelBaseNovelBase
1: Baselined† (▲)34.3183.5336.8988.85
2: ▲+w/o CoT73.8082.7968.0388.93
3: ▲+CoT78.3084.3268.6888.98
4: ▲+CoT+Corr62.1283.1371.8094.27
5: ▲+CoT+FTC78.6484.3270.1691.72
", + "bbox": [ + 503, + 358, + 777, + 463 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 629, + 464, + 648, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "methods (in 17 out of 18 tasks). The improvement of $\\mathrm{CLAMP}^{\\dagger}$ on novel keypoint detection strongly shows the benefits of adding auxiliary texts into training.", + "bbox": [ + 212, + 507, + 782, + 539 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Model and Performance Analysis", + "text_level": 1, + "bbox": [ + 214, + 559, + 535, + 574 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Why 0-shot can outperform 1-shot? One may observe the 0-shot testing scores on base keypoints are always higher than 1-shot testing. In this case, the training texts are accurate (cf. auxiliary text is noisy) and have high semantic similarity to testing texts. We dig more to visualize the corresponding text and keypoint features. Fig. 5 shows that textual keypoint features enjoy better clustering effects and lower variance than visual ones, which suggests textual representations of base keypoints are better thus 0-shot yields higher scores.", + "bbox": [ + 212, + 583, + 785, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Contrastive Learning (CL). After applying $\\mathcal{L}_{\\mathrm{tt}}$ , we observe the 0-shot performance significantly improves, e.g. $84.00\\%$ vs. $81.67\\%$ (2nd row, Table 4a), as CL improves the discrimination between text features. However, when adding visual-textual alignment (3rd row), 1-shot testing scores boost but slightly decreasing 0-shot scores. We suspect it is due to the imbalance of representation quality between multimodal features. After adding stop gradient, it mitigates negative impacts from \"weak\" modality, thus achieving higher scores (4th row). We also observe $\\mathcal{L}_{\\mathrm{vv}}$ does not further improve scores in our losses (5th row).", + "bbox": [ + 212, + 688, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Text Interpolation. Table 4b shows that using chain of thought (CoT) significantly improves performance, especially on AwA (3rd row), as CoT improves the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/81fa18534bbd4b4cfa7ae4c25f5282652bfefcdaea24e52d2338eafd3809b0d4.jpg", + "table_caption": [ + "Table 5: (a) Study on repetition $R$ ; (b) Parsing results with GPT3.5 and Vicuna." + ], + "table_footnote": [], + "table_body": "
RAwACUB
NovelBaseNovelBase
167.3483.9864.2191.40
378.64(+11.3)84.3269.81(+5.60)92.38
576.36(+9.02)83.7369.81(+5.60)92.02
1076.61(+9.27)83.3670.16(+5.95)91.72
", + "bbox": [ + 220, + 164, + 485, + 268 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/afa97f50a108196d6280af317aa8787f75ed5166cbbdf2415600d0d94b7618c9.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
DatasetGPT3.5Vicuna
Acc. kpAcc. objAcc. kpAcc. obj
Animal0.970.990.940.98
AwA0.961.000.930.98
CUB0.970.960.941.00
NABird0.970.990.951.00
", + "bbox": [ + 500, + 164, + 787, + 268 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/4d4ff6ead49932460c50bfb0dcbfadcb76d3e1f2b1ae31d5074017d710407108.jpg", + "table_caption": [ + "(b)", + "Table 6: Diverse text prompting evaluation. Results are the average of three runs." + ], + "table_footnote": [], + "table_body": "
MethodAnimal PoseAwACUBNABAvg Drop
NovelBaseNovelBaseNovelBaseNovelBaseNovelBase
OpenKD63.3765.5978.6484.3270.1691.7252.2990.63--
No parsing13.8915.1813.4421.7022.0917.4814.5517.6250.12↓65.07↓
Vicuna57.5459.8473.9778.1867.8987.3950.4586.593.65↓5.06↓
GPT3.561.4562.8276.2380.0769.3990.6552.2189.141.29↓2.39↓
", + "bbox": [ + 220, + 319, + 776, + 426 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "quality of reasoned texts. After applying FTC, the scores strike a high balance in all datasets (5-th row), as our FTC rejects low-quality texts and further reduces noise after including visual features into decision. We also explore the repetition $R$ for text reasoning. Table 5a shows the importance of setting $R > 1$ .", + "bbox": [ + 212, + 453, + 784, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Diverse Text Prompting. To evaluate our model under diverse text prompting, we randomly sample 1000 diverse texts from each dataset to conduct zero-shot testing. As shown in Table 6, after coupling with GPT3.5, our model retains the high performance even under diverse texts (with less than $3\\%$ drop). The strong results compared to no parsing show that leveraging LLM as a parser to handle language diversity is a possible solution. Moreover, we also found coupling with GPT3.5 has higher performance compared to Vicuna, as GPT3.5 enjoys a stronger keypoint text parsing ability, e.g., $96\\%$ vs. $93\\%$ in AwA (Table 5b).", + "bbox": [ + 212, + 513, + 785, + 635 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 657, + 357, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose to open the prompt diversity from the aspects of modality, semantics and language, to enable a more general zero- and few-shot keypoint detection. To this end, we build a versatile OpenKD model which supports both visual and textual prompting. Moreover, to bridge the semantics gap between seen and unseen texts, we propose a novel text interpolation and the selection strategy with false text control, which significantly improves zero-shot novel keypoint detection. We also discover that LLM is a good language parser. After coupling with LLM, our model can handle diverse texts well. We hope the proposed model, text interpolation and the parsing approach will provide useful insights for versatile keypoint detection, thus we highly recommend it to vision community.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Changsheng Lu is supported by Australian Government Research Training Program (AGRTP) International Scholarship. Piotr Koniusz is supported by CSIRO's Science Digital.", + "bbox": [ + 212, + 178, + 792, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 248, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "2. Banik, P., Li, L., Dong, X.: A novel dataset for keypoint detection of quadruped animals from images. arXiv preprint arXiv:2108.13958 (2021)", + "3. Bohdal, O., Tian, Y., Zong, Y., Chavhan, R., Li, D., Gouk, H., Guo, L., Hospedales, T.: Meta omnium: A benchmark for general-purpose learning-to-learn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7693-7703 (2023)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "5. Cao, J., Tang, H., Fang, H.S., Shen, X., Lu, C., Tai, Y.W.: Cross-domain adaptation for animal pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9498-9507 (2019)", + "6. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: realtime multiperson 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence 43(1), 172-186 (2019)", + "7. Carreira, J., Agrawal, P., Fragkiadaki, K., Malik, J.: Human pose estimation with iterative error feedback. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4733-4742 (2016)", + "8. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020)", + "9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)", + "10. Derpanis, K.G.: The harris corner detector. York University pp. 2-3 (2004)", + "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)", + "12. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017)", + "13. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: International Conference on Machine Learning. pp. 1126-1135. PMLR (2017)" + ], + "bbox": [ + 217, + 281, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Ge, Y., Zhang, R., Luo, P.: Metacloth: Learning unseen tasks of dense fashion landmark detection from a few samples. IEEE Transactions on Image Processing 31, 1120-1133 (2021)", + "15. Gidaris, S., Komodakis, N.: Dynamic few-shot visual learning without forgetting. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4367-4375 (2018)", + "16. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)", + "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "18. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1546-1555 (2018)", + "19. Jiao, B., Liu, L., Gao, L., Wu, R., Lin, G., Wang, P., Zhang, Y.: Toward re-identifying any animal. Advances in Neural Information Processing Systems 36 (2024)", + "20. Koch, G., Zemel, R., Salakhutdinov, R., et al.: Siamese neural networks for one-shot image recognition. In: ICML deep learning workshop. vol. 2. Lille (2015)", + "21. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)", + "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "23. Li, S., Gunel, S., Ostrek, M., Ramdya, P., Fua, P., Rhodin, H.: Deformation-aware unpaired image translation for pose estimation on laboratory animals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13158-13168 (2020)", + "24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "25. Lowe, G.: Sift-the scale invariant feature transform. Int. J 2, 91-110 (2004)", + "26. Lu, C., Gu, C., Wu, K., Xia, S., Wang, H., Guan, X.: Deep transfer neural network using hybrid representations of domain discrepancy. Neurocomputing 409, 60-73 (2020)", + "27. Lu, C., Koniusz, P.: Few-shot keypoint detection with uncertainty learning for unseen species. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19416-19426 (2022)", + "28. Lu, C., Koniusz, P.: Detect any keypoints: An efficient light-weight few-shot keypoint detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3882-3890 (2024)", + "29. Lu, C., Zhu, H., Koniusz, P.: From saliency to dino: Saliency-guided vision transformer for few-shot keypoint detection. arXiv preprint arXiv:2304.03140 (2023)", + "30. Mathis, A., Mamidanna, P., Cury, K.M., Abe, T., Murthy, V.N., Mathis, M.W., Bethge, M.: Deeplabcut: markerless pose estimation of user-defined body parts with deep learning. Nature neuroscience 21(9), 1281-1289 (2018)", + "31. Moskvyak, O., Maire, F., Dayoub, F., Baktashmotlagh, M.: Semi-supervised keypoint localization. arXiv preprint arXiv:2101.07988 (2021)", + "32. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning." + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19413-19423 (2023)", + "33. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 483-499. Springer (2016)", + "34. Pereira, T.D., Aldarondo, D.E., Willmore, L., Kislin, M., Wang, S.S.H., Murthy, M., Shaevitz, J.W.: Fast animal pose estimation using deep neural networks. Nature methods 16(1), 117-125 (2019)", + "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "36. Snell, J., Swersky, K., Zemel, R.S.: Prototypical networks for few-shot learning. arXiv preprint arXiv:1703.05175 (2017)", + "37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)", + "38. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1199-1208 (2018)", + "39. Tang, L., Wertheimer, D., Hariharan, B.: Revisiting pose-normalization for fine-grained few-shot recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14352-14361 (2020)", + "40. Thompson, J.J., Jain, A., LeCun, Y., Bregler, C.: Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems 27 (2014)", + "41. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014)", + "42. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 595-604 (2015)", + "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "44. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. Advances in neural information processing systems 29, 3630-3638 (2016)", + "45. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)", + "46. Wang, C., Jin, S., Guan, Y., Liu, W., Qian, C., Luo, P., Ouyang, W.: Pseudolabeled auto-curriculum learning for semi-supervised keypoint localization. arXiv preprint arXiv:2201.08613 (2022)", + "47. Wang, L., Huynh, D.Q., Koniusz, P.: A comparative review of recent kinect-based action recognition algorithms. IEEE Transactions on Image Processing 29, 15-28 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "49. Wu, Z., Su, L., Huang, Q.: Stacked cross refinement network for edge-aware salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7264-7273 (2019)", + "50. Xu, L., Jin, S., Zeng, W., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Pose for everything: Towards category-agnostic pose estimation. In: European Conference on Computer Vision. pp. 398-416. Springer (2022)", + "51. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022)", + "52. Yang, S., Quan, Z., Nie, M., Yang, W.: Transpose: Keypoint localization via transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11802-11812 (2021)", + "53. Yao, Q., Quan, Q., Xiao, L., Kevin Zhou, S.: One-shot medical landmark detection. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 177-188. Springer (2021)", + "54. Zhang, H., Zhang, K., Xu, L., Lai, S., Shao, W., Zheng, N., Luo, P., Qiao, Y.: Language-driven open-vocabulary keypoint detection for animal body and face. arXiv preprint arXiv:2310.05056 (2023)", + "55. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014)", + "56. Zhang, X., Wang, W., Chen, Z., Xu, Y., Zhang, J., Tao, D.: Clamp: Prompt-based contrastive learning for connecting language and animal pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23272-23281 (2023)", + "57. Zhao, S., Gong, M., Zhao, H., Zhang, J., Tao, D.: Deep corner. International Journal of Computer Vision pp. 1-25 (2023)", + "58. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena (2023)", + "59. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 674 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Lu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_model.json b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3548548d0ec4ddb9dddae780ed78ab381d3edd49 --- /dev/null +++ b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_model.json @@ -0,0 +1,2744 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.237, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "OpenKD: Opening Prompt Diversity for Zero-and Few-shot Keypoint Detection" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.213, + 0.72, + 0.229 + ], + "angle": 0, + "content": "Changsheng Lu1, Zheyuan Liu1, and Piotr Koniusz*,2,1" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.24, + 0.711, + 0.282 + ], + "angle": 0, + "content": "1The Australian National University 2Data61/CSIRO changshengluu@gmail.com, zheyuan.david.liu@outlook.com, piotr.koniusz@data61.csiro.au" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.32, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Exploiting foundation models (e.g., CLIP) to build a versatile keypoint detector has gained increasing attention. Most existing models accept either the text prompt (e.g., \"the nose of a cat\"), or the visual prompt (e.g., support image with keypoint annotations), to detect the corresponding keypoints in query image, thereby, exhibiting either zero-shot or few-shot detection ability. However, the research on multimodal prompting is still underexplored, and the prompt diversity in semantics and language is far from opened. For example, how to handle unseen text prompts for novel keypoint detection and the diverse text prompts like \"Can you detect the nose and ears of a cat?\" In this work, we open the prompt diversity in three aspects: modality, semantics (seen vs. unseen), and language, to enable a more general zero- and few-shot keypoint detection (Z-FSKD). We propose a novel OpenKD model which leverages a multimodal prototype set to support both visual and textual prompting. Further, to infer the keypoint location of unseen texts, we add the auxiliary keypoints and texts interpolated in visual and textual domains into training, which improves the spatial reasoning of our model and significantly enhances zero-shot novel keypoint detection. We also find large language model (LLM) is a good parser, which achieves over \\(96\\%\\) accuracy when parsing keypoints from texts. With LLM, OpenKD can handle diverse text prompts. Experimental results show that our method achieves state-of-the-art performance on Z-FSKD and initiates new ways of dealing with unseen text and diverse texts. The source code and data are available at https://github.com/AlanLuSun/OpenKD." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.679, + 0.377, + 0.695 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Keypoint detection is a fundamental research problem in computer vision and has numerous applications such as pose estimation for humans [6,8,12,33,37, 52] and animals [2,5,23,34], action recognition [30,47], and fine-grained image classification [39,55]. Over the past decade, significant advancements have been made in deep keypoint detection. However, most existing methods are tailored to close-set detection and struggle to predict novel body parts and species with limited data. Thus, it necessitates such a zero- or few-shot keypoint detection:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.387, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.147, + 0.237, + 0.159 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.147, + 0.307, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.261, + 0.275, + 0.269 + ], + "angle": 0, + "content": "Visual prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.261, + 0.333, + 0.269 + ], + "angle": 0, + "content": "Query image" + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.147, + 0.365, + 0.159 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.147, + 0.437, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.261, + 0.398, + 0.269 + ], + "angle": 0, + "content": "Text prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.261, + 0.461, + 0.269 + ], + "angle": 0, + "content": "Query image" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.149, + 0.511, + 0.159 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.147, + 0.605, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.261, + 0.568, + 0.269 + ], + "angle": 0, + "content": "Visual + Text prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.149, + 0.662, + 0.159 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.149, + 0.764, + 0.183 + ], + "angle": 0, + "content": "Results on AwA dataset One shot Zero shot One shot w/ text" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.186, + 0.777, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 1: Illustration of multimodal prompting for keypoint detection. Our model can successfully detect keypoints given visual prompts formed by support images and keypoints (a), text prompts (b), or both (c). Graph (d) shows our model well combines the advantages of different modalities, mitigating the weakness induced by either modality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.4 + ], + "angle": 0, + "content": "after being trained on a diverse dataset, the model can quickly recognize novel or base keypoints in unseen species given only zero, one or a few labelled samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.612 + ], + "angle": 0, + "content": "On the one hand, researchers have recently explored few-shot keypoint detection (FSKD) [27-29, 53]. The FSKD inspired by few-shot learning [13, 20, 36, 38, 44] is general and offers higher flexibility of detecting a varying number of keypoints in a query image, given the prompts formed by support image with keypoint annotations. In this regard, the support set is namely visual prompt, which is required when evaluating new classes. On the other hand, in the era of foundation model, the vision-language model (e.g., CLIP [35]) injects new life to detecting keypoints, enabling zero-shot keypoint detection (ZSKD) [54, 56]. Compared to FSKD, ZSKD does not need support image with annotations, instead, using the text prompt to instruct the model to detect the specified keypoints in query image. In this case, the support set is text prompt. Though impressive for the pioneering FSKD and ZSKD models, we identify an important issue limiting the progress: the prompt diversity for current keypoint detection methods is far from opened, especially in the aspects of modality, semantics, and language." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Modality diversity. Most existing keypoint detection models cannot support multimodal prompts, e.g., image, text, or both (see Fig. 1). Multimodal prompting is more friendly in real-world interaction, and coherent with the human concept recognition. We not only see the objects, but also describe objects with language. Our work extends existing FSKD and ZSKD, building a more general zero- and few-shot keypoint detection by leveraging a multimodal prototype set and aligning the visual keypoint features towards the textual features. While straightforward, our method enables one to study the advantages of respective modality data and exploit them for better model training and testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Semantic diversity. Considering keypoints between seen and unseen species, there exist large similarity yet difference in semantics. A strong advantage of text prompt is that the keypoints with the same semantics share high similarity in language across species, which enables excellent ZSKD on base keypoints. However, if the text has different semantics, significant domain gap arises. For instance, \"the eye of a cat\" vs. \"the eye of a dog\" has cosine similarity of 0.93 using CLIP" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.147, + 0.329, + 0.169 + ], + "angle": 0, + "content": "Can you make out the back-right knee, front-left knee, body-middle-left on cow?" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.171, + 0.33, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.257, + 0.282, + 0.266 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.147, + 0.444, + 0.169 + ], + "angle": 0, + "content": "Can you detect the body-middle-right, nose, left earbase, front-right paw, left eye of grizzly bear?" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.171, + 0.444, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.257, + 0.396, + 0.266 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.147, + 0.547, + 0.169 + ], + "angle": 0, + "content": "Please find the throat, belly, breast, back, left leg for the evening grosbeak." + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.171, + 0.558, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.257, + 0.511, + 0.266 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.149, + 0.671, + 0.166 + ], + "angle": 0, + "content": "Where is the breast, throat for pine warbler?" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.171, + 0.671, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.257, + 0.624, + 0.266 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.15, + 0.774, + 0.166 + ], + "angle": 0, + "content": "Can you recognize the back, crown, throat, tail, belly?" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.171, + 0.785, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.257, + 0.737, + 0.266 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 2: Examples of keypoint detection under diverse text prompting. With LLM, our method can deal with diverse texts, showing potential for real-world applications. The circles and crosses refer to predictions and GT, respectively. A keypoint is regarded as a correct detection if falling in the white shadow area that signifies PCK@0.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.504 + ], + "angle": 0, + "content": "text embeddings, whereas for \"the eye of a cat\" vs. \"the knee of a dog\" it drops to 0.77. We observe that the keypoint detectors cannot perform well if text is unseen during training. To bridge the gap, we propose to open the semantic diversity by adding auxiliary keypoints and texts into training, where the auxiliary texts are reasoned by LLM given the base keypoints. To ease the reasoning error, we ask LLM a good question using chain of thought [48]. We also propose a novel sampling strategy with false text control to improve the matching between auxiliary keypoints and texts. With them together, our model significantly improves ZSKD on novel keypoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Language diversity. Following zero-shot image classification by CLIP [35], existing ZSKD also constructs simple text prompts based on templates, e.g., \"〈keypoint〉\" or \"the〈keypoint〉 of a〈category〉 in the photo\", where \\( \\langle \\cdot \\rangle \\) is replaced by class names. In real-world interaction, humans tend to question in diverse styles, which results in diverse text prompts. For example, \"Please locate the right-back leg on cat\", \"Can you find the left eye, right ear and nose in image?\", etc. A natural question is how to handle the diverse text prompting? To address this, we propose a simple yet effective approach, whose key idea is to borrow the large language model (LLM) such as GPT3.5 [4] or Vicuna [9, 58] to parse the diverse texts via prompt engineering. After extracting the keypoint type and object category, they will be synthesized into simple text prompts to instruct the detection model. As such, the diverse text prompting can be transformed into simple text prompting, which sheds light on opening language diversity (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To summarize, in this paper, we propose an OpenKD model with several intriguing features: 1) supporting both visual and textual prompts, 2) having the potential to handle unseen texts and diverse texts, and 3) maintaining strong generality and performance on ZSKD and FSKD. We report that LLM is capable of being a reasoner for text interpolation, and a good language parser for parsing diverse texts. We also contribute four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42] for fair evaluations. To our best knowledge, we are the first to open semantics and language diversity of text prompts for keypoint detection." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.297 + ], + "angle": 0, + "content": "Keypoint Detection has been widely studied ranging from the traditional interest points [10, 25] to deep corner detection [57], semi-supervised [18, 31, 46] and fully-supervised keypoint estimation [6, 8, 12, 33, 37, 40, 51, 52]. There are two major classes of methods for deep keypoint localization: i) direct coordinates based regression [7, 41] and ii) heatmap based regression. In contrast to existing heatmap based models dedicated to specific body parts, e.g., top-down [12, 16, 37] and bottom-up pose estimators [6, 8, 33], our OpenKD model offers more flexible keypoint detection, breaking the limitation of keypoint types to be detected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.787, + 0.508 + ], + "angle": 0, + "content": "Few-shot Keypoint Detection is more versatile and data-efficient compared to the supervised paradigms. Similar to standard few-shot learning (FSL) [44], FSKD also takes episodes for training and evaluation. Many well-known FSL methods such as ProtoNet [36], RelationNet [38], LwoF [15] and MAML [13] have been extended to the field of keypoint detection [3,14,27] and serve as baselines. Recently, Lu and Koniusz [27] formalize the comprehensive FSKD settings and model the localization uncertainty of keypoints. Further, they propose a lightweight FSKD model [28] and explore FSKD under transductive setting and occlusions [29]. Bohdal et al. [3] propose a dataset-of-datasets, benchmarking the FSL algorithms universal to various vision tasks. FSKD also inspires class-agnostic animal pose estimation [27, 28, 50]. Compared to existing FSKD, our work pushes further by opening the text prompting, which enables more versatile keypoint detection. Moreover, text prompts can complement visual prompts, delivering better performance of detecting base keypoints on unseen species." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.583 + ], + "angle": 0, + "content": "Zero-shot Keypoint Detection becomes feasible thanks to the great progress of vision-language pre-training (e.g., CLIP [35] and BLIP [22]). Since CLIP was pre-trained on internet-scale image-text pairs that contain more semantic concepts than other datasets, CLIP can provide semantically rich features for various downstream tasks [32, 56, 59], showing strong zero-shot transfer ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.75 + ], + "angle": 0, + "content": "Recently, CLAMP [56] adapts CLIP for animal pose estimation in feature-aware and spatial-aware aspects; Approach [54] also employs CLIP as backbone and proposes language-driven keypoint detection. Compared to FSKD which still requires one or more annotations, ZSKD merely demands cheap language descriptions, which make it intriguing for keypoint detection. Nevertheless, most existing ZSKD works cannot support multimodal prompts and their prompt diversity is limited. To address these issues, we propose an OpenKD model which further opens modality, semantics, and language. To handle unseen texts, we propose to generate auxiliary keypoint-text pairs which significantly help reason novel keypoints on unseen species. Moreover, we contribute a simple yet effective language parsing module to handle diverse text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Foundation Models include a broad set of models pre-trained over large-scale datasets, including LLMs (e.g., GPT3.5 [4]), VLMs (e.g., CLIP [35]), and Multimodal models (e.g., GPT4 [1]). Transferring the knowledge from these foundation models to specific tasks in a cost-effective way is popular in computer vision [19,24,26,54,56]. Thus, we follow this trend and leverage LLM like GPT3.5 (relatively cheap) to perform reasoning and parsing for keypoints." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.262, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.175, + 0.427, + 0.189 + ], + "angle": 0, + "content": "3.1 Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.333 + ], + "angle": 0, + "content": "Following the general few-shot learning [36, 38, 44], Z-FSKD model is evaluated on episodes, each of which includes a support set and a query set. The query set is comprised of query images, while the support set gives the prompts. If the visual prompts, i.e., \\( K \\) support images with keypoint annotations (\\( K \\geq 1 \\)), are given, then the problem is defined as \\( K \\)-shot detection. If only the text prompt (i.e., the language description) is given, then \\( K = 0 \\) thus the problem becomes zero-shot detection. The goal of Z-FSKD is to detect the corresponding keypoints in query image given the prompts, whether visual prompt, text prompt, or both. Such an approach allows the model to effectively respond to diverse inputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.409 + ], + "angle": 0, + "content": "The overview of our model inference is shown in Fig. 3, which mainly includes four stages: i) image/text feature extraction, ii) feature adaptation, iii) keypoint prototype set building, iv) correlation, decoding and heatamp fusion. The overview of our model training is shown in Fig. 4, which includes the novel approaches for performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.592 + ], + "angle": 0, + "content": "Image/Text Feature Extraction Since the input of our model involves vision and language two different modalities, we wish the features extracted from image and text have smaller modality gap, so that the text can somewhat correlate with image regions of the same semantics. Such a property helps the model find keypoint locations in query image. To this end, we resort to CLIP [35] pre-trained on large-scale image-text pairs. The CLIP includes image and text encoders. The text encoder is generally a transformer [43] while the image encoder can be CNN [21] or ViT [11]. We empirically found that the CLIP image encoder based on RN50 [17, 35], a CNN model, gives high efficiency in consideration of both performance and cost. Thus, we choose it as the default backbone. When extracting the image features, original CLIP image encoder only retains the classification token \\(\\mathbf{x}_{\\mathrm{class}}\\) and discards image tokens \\(\\mathbf{X}\\) by attention pooling as" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.6, + 0.785, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\text {c l a s s}} = \\operatorname {A v g P o o l} (\\mathbf {X}), \\quad \\mathbf {x} _ {\\text {c l a s s}} ^ {\\prime} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {x} _ {\\text {c l a s s}} \\mathbf {W} _ {q}, \\mathbf {X} \\mathbf {W} _ {k}, \\mathbf {X} \\mathbf {W} _ {v}\\right) \\mathbf {W} _ {o}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.672 + ], + "angle": 0, + "content": "While it is natural for image-text matching, we require the image tokens to recover the spatial locations of keypoints. Thus, we propose to obtain the image tokens via a projection composed by \\(\\mathbf{W}_v\\) and \\(\\mathbf{W}_o\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.68, + 0.785, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {\\prime} = \\mathbf {X} \\mathbf {W} _ {v} \\mathbf {W} _ {o}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.735 + ], + "angle": 0, + "content": "In this way, we can not only handle the channel inconsistency between raw image tokens \\(\\mathbf{X}\\) and text features, but also reuse the projection layers \\(\\mathbf{W}_v\\) and \\(\\mathbf{W}_o\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For brevity, we assume that per input episode contains one query image \\(\\mathbf{I}^{\\mathrm{q}}\\), one support image \\(\\mathbf{I}^{\\mathrm{s}}\\) with \\(N\\) annotated keypoints, and \\(N\\) keypoint texts. With Eq. 2, both the support and query images can be encoded and projected as the support and query feature maps as \\(\\mathbf{X}^{\\mathrm{s}} = \\mathcal{F}_{\\mathrm{v}}(\\mathbf{I}^{\\mathrm{s}})\\) and \\(\\mathbf{X}^{\\mathrm{q}} = \\mathcal{F}_{\\mathrm{v}}(\\mathbf{I}^{\\mathrm{q}})\\) in deep feature space \\(\\mathbb{R}^{l\\times l\\times d}\\) via a shared CLIP image encoder \\(\\mathcal{F}_{\\mathrm{v}}\\). Further, with the CLIP text encoder \\(\\mathcal{F}_{\\mathrm{t}}\\), the \\(N\\) texts are firstly tokenized, then encoded as the text features \\(\\mathbf{t}_n\\in \\mathbb{R}^{m\\times d}\\), where \\(n = 1,2,\\dots ,N\\); and \\(m\\) is sequence length." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.229, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 3: The sketch of model inference. Our OpenKD allows testing under visual prompt, text prompt, or both. For clarification, we show the \"both\" case (i.e., 1-shot with text testing). We firstly extract the deep features of texts, support and query images via CLIP, and then adapt both modalities of features via residual refinement. After extracting the visual keypoint prototype (VKP) and textual counterpart, we build the prototype set to perform class-agnostic correlation and heatmap decoding. Finally, we fuse the heatmaps induced by two modalities (i.e., M1 & M2) to obtain predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Feature Adaptation As CLIP provides general yet transferrable features, we adapt it to a new multimodal feature space suitable for keypoint detection. Inspired by residual learning [17], we propose two adaptation nets \\(\\mathcal{A}_{\\mathrm{v}}\\) and \\(\\mathcal{A}_{\\mathrm{t}}\\) to respectively refine the obtained image and text features in a residual way:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.507, + 0.786, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {\\mathrm {s}} := \\mathbf {X} ^ {\\mathrm {s}} + \\mathcal {A} _ {\\mathrm {v}} \\left(\\mathbf {X} ^ {\\mathrm {s}}\\right), \\quad \\mathbf {X} ^ {\\mathrm {q}} := \\mathbf {X} ^ {\\mathrm {q}} + \\mathcal {A} _ {\\mathrm {v}} \\left(\\mathbf {X} ^ {\\mathrm {q}}\\right), \\quad \\mathbf {t} _ {n} := \\mathbf {t} _ {n} + \\mathcal {A} _ {\\mathrm {t}} \\left(\\mathbf {t} _ {n}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.566 + ], + "angle": 0, + "content": "The adaptation net \\(\\mathcal{A}\\) is highly scalable and can be based on transformer [43] or bottleneck [17] as refinement blocks. We found that \\(\\mathcal{A}\\) can adapt features well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Keypoint Prototype Set Once we obtain the adapted features of support image and text prompts, i.e., \\(\\mathbf{X}^{\\mathrm{s}}\\) and \\(\\mathbf{t}_n\\), we propose to convert them into keypoint prototypes, which are unified in a keypoint prototype set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.789, + 0.841 + ], + "angle": 0, + "content": "For visual prompt, recall that keypoint labels are provided along with the support image. Thus, one can aggregate the local features for each support keypoint \\(\\mathbf{p}_n\\) by using pixelwise weighted summation between feature map \\(\\mathbf{X}^{\\mathrm{s}}\\) and Gaussian heatmap \\(\\mathbf{H}(\\mathbf{p}_n;\\sigma)\\), yielding the visual keypoint representation (VKR) as \\(\\Phi_n\\in \\mathbb{R}^d\\). The \\(\\sigma\\) is the standard deviation that controls the Gaussian spread. If \\(K\\) support images are given, namely in few-shot case, the VKRs of the same type of keypoints \\(\\Phi_{k,n}\\) will be averaged to build the visual keypoint prototype (VKP) \\(\\Psi_{n}^{\\mathrm{v}} = \\frac{1}{K}\\sum_{k}\\Phi_{k,n}\\). Analogously, one can develop textual keypoint prototype (TKP) \\(\\Psi_{n}^{\\mathrm{t}}\\) by text features if provided with multiple texts per keypoint. Finally, one can build the keypoint prototype set as \\(\\mathcal{T} = \\mathcal{T}^{\\mathrm{v}}\\cup \\mathcal{T}^{\\mathrm{t}}\\), where \\(\\mathcal{T}^{\\mathrm{v}} = \\{\\Psi_n^{\\mathrm{v}}\\}\\) is the VKPs and \\(\\mathcal{T}^{\\mathrm{t}} = \\{\\Psi_n^{\\mathrm{t}}\\}\\) is the TKPs. In this way, prompts of different modalities can be summarized as keypoint prototypes in a shared \\(d\\) dimensional feature space, which enables our model to flexibly handle various modalities. Moreover, each prototype in set \\(\\mathcal{T}\\) can guide the model to induce a keypoint heatmap, establishing the zero- and few-shot keypoint detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.146, + 0.486, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.275, + 0.395, + 0.285 + ], + "angle": 0, + "content": "(a) Model training" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.151, + 0.787, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.274, + 0.715, + 0.286 + ], + "angle": 0, + "content": "(b) Auxiliary text interpolation" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 4: Model training and text interpolation. (a) In addition to multi-group heatmap regression, we improve model performance by introducing intra- and inter-modality contrastive learning and the novel auxiliary keypoint and text learning. (b) We exploit LLM for auxiliary texts interpolation and explore incorporating visual keypoint features for text selection in order to mitigate the noisy texts or false texts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Correlation, Decoding and Fusion To discover the corresponding keypoints in query image \\(\\mathbf{I}^{\\mathrm{q}}\\), each keypoint prototype \\(\\Psi_{n}\\), either VKP \\(\\Psi_{n}^{\\mathrm{v}}\\) or TKP \\(\\Psi_{n}^{\\mathrm{t}}\\), is required to be correlated with the query feature map \\(\\mathbf{X}^{\\mathrm{q}}\\). To this end, we leverage a correlation module \\(\\mathcal{C}\\) which takes both \\(\\Psi_{n}\\) and \\(\\mathbf{X}^{\\mathrm{q}}\\) as input to produce the attentive feature maps \\(\\mathbf{A}_{n} = \\mathcal{C}(\\Psi_{n},\\mathbf{X}^{\\mathrm{q}})\\), where \\(\\mathbf{A}_{n}\\in \\mathbb{R}^{l\\times l\\times d}\\) and \\(n = 1,2,\\dots ,N\\). For the correlation module \\(\\mathcal{C}\\), we explored multiple variants and found that the simple cross correlation already yields high performance, i.e., \\(\\mathbf{A}_n = \\mathbf{X}^{\\mathrm{q}}\\odot \\Psi_n\\), where \\(\\odot\\) denotes the channel-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Subsequently, a class-agnostic decoder \\(\\mathcal{D}\\) is devised to convert each attentive feature map into a keypoint heatmap, i.e., \\(\\mathbf{H}_n = \\mathcal{D}(\\mathbf{A}_n)\\), where \\(\\mathbf{H}_n \\in \\mathbb{R}^{l \\times l}\\). For heatmap regression based keypoint localization, the higher resolution of heatmap can greatly reduce the coordinate decoding error. Thus, a lightweight upsampling module \\(\\mathcal{U}\\) is adopted to further refine heatmaps, i.e., \\(\\mathbf{H}_n := \\mathcal{U}(\\mathbf{H}_n)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Since each modality of prototypes can induce one group of heatmaps, considering visual and textual modalities, we have two groups of heatmaps \\(\\mathbf{H}^{\\mathrm{v}} = \\{\\mathbf{H}_n^{\\mathrm{v}}\\}_{n=1}^N\\) and \\(\\mathbf{H}^{\\mathrm{t}} = \\{\\mathbf{H}_n^{\\mathrm{t}}\\}_{n=1}^N\\). In testing phase, we fuse the upsampled multi-group heatmaps as the final output \\(\\mathbf{H} \\in \\mathbb{R}^{N \\times ul \\times ul}\\) (\\(u\\) is upsampling factor):" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.667, + 0.785, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\left(\\mathbf {H} ^ {\\mathrm {v}} + \\mathbf {H} ^ {\\mathrm {t}}\\right) / 2. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.722 + ], + "angle": 0, + "content": "During training phase, inspired by HigherHRNet [8], we perform group-specific supervision for heatmaps as" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.732, + 0.785, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k p}} = \\left(\\| \\mathbf {H} ^ {\\mathrm {v}} - \\mathbf {H} ^ {*} \\| _ {2} ^ {2} + \\| \\mathbf {H} ^ {\\mathrm {t}} - \\mathbf {H} ^ {*} \\| _ {2} ^ {2}\\right) / 2, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.774, + 0.777 + ], + "angle": 0, + "content": "where \\(\\mathbf{H}^*\\) denotes the groundtruth heatmap that encodes the query keypoint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Improving Model by Contrastive Learning Since the keypoint prototype directly impacts performance, we aim for its representativeness to be as good as possible. The same type of keypoints should have higher invariance across species, e.g., the ear of cat, dog, and cow, while different types are required to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.234 + ], + "angle": 0, + "content": "be sufficiently discriminative such that our model can distinguish the ambiguous texts or fine-grained texts such as \"the left ear\" vs. \"the right ear\". To this end, we propose to introduce a contrative loss over TKPs. If randomly sampling two species \\((s,s^{\\prime})\\) at a time and each species pertaining to an episode, we have pairwise sets of TKPs, i.e., \\(\\mathcal{T}_s^{\\mathrm{t}} = \\{\\Psi_n\\}\\) and \\(\\mathcal{T}_{s^{\\prime}}^{\\mathrm{t}} = \\{\\Psi_{n}^{\\prime}\\}\\), which can form similarity matrix as" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.234, + 0.786, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} \\left(\\mathcal {T} _ {s} ^ {\\mathrm {t}}, \\mathcal {T} _ {s ^ {\\prime}} ^ {\\mathrm {t}}\\right) = \\left( \\begin{array}{c c} \\cos \\left(\\Psi_ {1}, \\Psi_ {1} ^ {\\prime}\\right) & \\dots \\cos \\left(\\Psi_ {1}, \\Psi_ {N} ^ {\\prime}\\right) \\\\ \\vdots & \\vdots \\\\ \\cos \\left(\\Psi_ {N}, \\Psi_ {1} ^ {\\prime}\\right) & \\dots \\cos \\left(\\Psi_ {N}, \\Psi_ {N} ^ {\\prime}\\right) \\end{array} \\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.295, + 0.789, + 0.327 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) denotes the cosine similarity. Then, the contrastive loss within textual modality \\(\\mathcal{L}_{\\mathrm{tt}}\\) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.337, + 0.786, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t t}} ^ {s \\rightarrow s ^ {\\prime}} = - \\langle \\mathbb {I}, \\log (\\operatorname {s o f t m a x} (\\mathbf {J} (\\mathcal {T} _ {s} ^ {\\mathrm {t}}, \\mathcal {T} _ {s ^ {\\prime}} ^ {\\mathrm {t}}) / \\tau)) \\rangle , \\quad \\mathcal {L} _ {\\mathrm {t t}} = \\frac {1}{2} (\\mathcal {L} _ {\\mathrm {t t}} ^ {s \\rightarrow s ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {t t}} ^ {s ^ {\\prime} \\rightarrow s}), \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.789, + 0.557 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}\\) is the identity matrix and \\(\\langle \\cdot ,\\cdot \\rangle\\) denotes inner product. Since CLIP is trained via image-level alignment instead of keypoint-level alignment with text, we wish VKPs and TKPs were aligned so that we can exploit their respective advantages to help each other. This motivates us to propose the second contrastive loss between VKPs and TKPs. Similar to Eq. 7, we build \\(\\mathcal{L}_{\\mathrm{vt}}\\). We find that the visual-textual alignment improves 1-shot testing, but unintentionally degrades 0-shot performance slightly. We discover the principle behind this is better clustering effects of textual keypoint representations than visual ones. To address this issue, we perform stop gradient on TKPs in \\(\\mathcal{L}_{\\mathrm{vt}}\\), thus enforcing the VKPs to align towards TKPs. We also find adding \\(\\mathcal{L}_{\\mathrm{vv}}\\) does not further improve scores. We conjecture that after aligning VKPs to TKPs via \\(\\mathcal{L}_{\\mathrm{vt}}\\), the visual features will also improve discrimination along with textual features as \\(\\mathcal{L}_{\\mathrm{tt}}\\) is applied." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.56, + 0.79, + 0.606 + ], + "angle": 0, + "content": "Optimization By incorporating the heatmap regression loss \\(\\mathcal{L}_{\\mathrm{kp}}\\), intra- and inter-modality contrastive loss \\(\\mathcal{L}_{\\mathrm{tt}}\\) and \\(\\mathcal{L}_{\\mathrm{vt}}\\), we have the overall loss as \\(\\mathcal{L} = \\lambda_1\\mathcal{L}_{\\mathrm{kp}} + \\lambda_2\\mathcal{L}_{\\mathrm{tt}} + \\lambda_3\\mathcal{L}_{\\mathrm{vt}}\\). We set the loss weight \\(\\lambda_{1} = 1\\), \\(\\lambda_{2} = \\lambda_{3} = 0.002\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.637, + 0.643 + ], + "angle": 0, + "content": "3.2 Training with Auxiliary Keypoints and Texts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Localizing the seen keypoints and texts is relatively easy to handle, but it becomes difficult if the unseen ones are encountered. Thus, we propose to add auxiliary keypoints and texts interpolated from visual and textual domains into training, which significantly improves the ability to detect novel keypoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Visual Interpolation Following [27], we generate the auxiliary keypoints in visual domain via interpolation \\(\\mathcal{I}^{\\mathrm{v}}(z;\\mathbf{p}_1,\\mathbf{p}_2)\\), where \\(z\\in (0,1)\\) is the interpolation node and \\(\\mathbf{p}_i\\in \\mathbb{R}^2\\) are the end points that form an interpolation path. An off-the-shelf saliency detector [49] is adopted to filter the auxiliary keypoints outside of the foreground [27]. Via interpolation \\(\\mathcal{I}^{\\mathrm{v}}\\) in the visual domain, the auxiliary keypoints \\(\\hat{\\mathbf{p}}\\) can boost the visual diversity of training keypoints." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Text Interpolation and Selection The \\(\\mathcal{L}^{\\nu}\\) can assist the model in handling unseen keypoints in visual prompt, but it helps little if prompting unseen texts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.262, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "Consequently, we propose using text interpolation \\(\\mathcal{I}^{\\mathrm{t}}(z;t(\\mathbf{p}_1),t(\\mathbf{p}_2),c)\\) to mitigate this issue, where \\(t(\\mathbf{p}_i)\\) are the names of end points \\(\\mathbf{p}_i\\) and \\(c\\) is the object category. Thanks to great language processing ability of LLM (e.g., GPT3.5 [4]), we propose to ask LLM a good question to reason the auxiliary keypoint texts given base keypoints (Fig. 4(b)), e.g., to infer the keypoint text between \"high\" and \"paw\" of a \"cat\", we can design a vanilla prompt for the LLM as:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.265 + ], + "angle": 0, + "content": "\"Please give me three most common body parts/ keypoints at \\(1 / 2\\) between thigh and paw of a cat. Please answer in concise words.\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Apparently, we expect that the answers returned by LLM should include \"knee\" or \"elbow\". In practice, the keypoint types are usually more complex thus the reasoning is challenging. To enhance reasoning, we improve the above vanilla prompt into a more advanced one by using chain of thought (CoT) [48], whose key idea is to firstly provide an example with analysis and then ask the question. We provide the detailed prompts and examples in §A of Suppl." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.785, + 0.493 + ], + "angle": 0, + "content": "Considering the LLM suffers from some randomness and might produce erroneous answers, we let LLM give three most possible answers for each interpolation path and repeat this step \\( R \\) times, yielding a text pool with \\( 3R \\) texts, i.e., \\( \\{\\hat{t}_i\\}_{i=1}^{3R} \\). As such, the GT text has higher chance to be included. Afterwards, we need to select one interpolated text from the pool to form a pair with a visually interpolated keypoint. To this end, a simple yet effective approach is to sample from top-1 results of all repetitions. However, it does not use the visual features for selection. An interesting approach is to use the correlation between visual and text features for selection as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.505, + 0.785, + 0.524 + ], + "angle": 0, + "content": "\\[\nj = \\arg \\max _ {i} \\cos (\\hat {\\boldsymbol {\\Phi}}, \\hat {\\mathbf {t}} _ {i}), \\quad \\hat {t} ^ {*} = \\hat {t} _ {j}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.786, + 0.718 + ], + "angle": 0, + "content": "where \\(\\hat{\\Phi}\\) is the auxiliary keypoint feature of \\(\\hat{\\mathbf{p}}\\), \\(\\hat{\\mathbf{t}}_i\\) is text feature of \\(\\hat{t}_i\\), and \\(\\hat{t}^*\\) is the selected text. However, we found that sometimes it leads to performance drop. We suspect it is because CLIP's keypoint-text features are not fine-grained aligned enough on those regions beyond annotations. Opposed to picking, we observe such a correlation could be used to reject low-quality text if keypoint-text similarity \\(\\cos (\\hat{\\Phi},\\hat{\\mathbf{t}}_i)\\) is low. Consequently, we propose a more advanced sampling with false text control (FTC). Namely, we sample the text within the range top- \\(\\eta\\) but the text will be rejected if the similarity \\(\\cos (\\hat{\\Phi},\\hat{\\mathbf{t}}_i)\\) is below the threshold \\(\\alpha\\). In this way, we can form auxiliary keypoint-text pairs with higher quality to enhance model training. We apply \\(\\mathcal{L}^{\\mathrm{v}}\\) to both support and query images to generate support and query auxiliary keypoints \\(\\hat{\\mathbf{p}}^{\\mathrm{s}}\\) and \\(\\hat{\\mathbf{p}}^{\\mathrm{q}}\\), respectively. As in Eq. 5, \\(\\hat{\\mathbf{p}}^{\\mathrm{q}}\\) encodes GT heatmap supervising the heatmaps induced by \\(\\hat{\\mathbf{p}}^{\\mathrm{s}}\\) and \\(\\hat{t}^*\\)." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.74, + 0.533, + 0.756 + ], + "angle": 0, + "content": "3.3 LLM is a Good Language Parser" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To handle diverse text prompts in keypoint detection, unlike current fashion of utilizing LLMs for text generation, we use LLM for text decomposition. For example, if our model receives a diverse text prompt like \"Can you localize the left eye and nose of cat?\", we firstly leverage an LLM (e.g., GPT3.5 [4]) to perform text parsing by providing it with a simple yet effective prompt as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "Please extract the animal and keypoint keywords from the below text: \"Can you localize the left eye and nose of cat?\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.251 + ], + "angle": 0, + "content": "Then, LLM returns the parsed texts regarding ⟨keypoint⟩ and ⟨object⟩, i.e., the \"left eye\", \"nose\", and \"cat\" in the example. We directly utilize parsed results to synthesize simple text prompts to instruct our model and predict all the keypoints corresponding to the text. The problem of diverse text prompting becomes simple text prompting, thus opening the language diversity for keypoint detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.447 + ], + "angle": 0, + "content": "To fairly evaluate our approach, we manually collect 100 diverse text templates which cover most scenarios for questions (\\(\\S \\mathbf{B}\\) of Suppl.). Take one as an example: \"Where is the for . Afterwards, we develop four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42]. Specifically, we accompany each object instance with a diverse text prompt synthesized by randomly sampling one template and one to \\(N\\) valid keypoints. Compared to traditional text prompts which only have \\(N\\) variants, our constructed text prompt sets are more diverse, whose space could be as large as \\(100 \\cdot (C_N^1 + C_N^2 + \\dots + C_N^N) = 100 \\cdot 2^N\\) variants. For example, as CUB has 11 keypoints, its space is larger than \\(10^5\\). We randomly sample 1000 diverse texts from each dataset and evaluate the LLM's parsing performance. Table 5b reports that GPT3.5 has over \\(96\\%\\) accuracy in parsing keypoints from text, which shows that LLM can be a good language parser." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.377, + 0.487 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.5, + 0.446, + 0.515 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.687 + ], + "angle": 0, + "content": "Datasets and Splits. 1) Animal pose dataset [5] has five mammal species cat, dog, cow, horse, and sheep. Each animal species is alternately chosen as unseen species for testing while the remaining four as seen species for training, which yields five subproblems; 2) AwA [2] has 35 diverse animal species with 10064 images. For AwA, 25 species are for training and 10 for testing; 3) CUB [45] consists of 200 bird species with 15 keypoint annotations. We use 100 species for training, 50 for validation, and 50 for testing; 4) NABird [42] is larger than CUB and has 555 categories. The species splits are 333, 111, and 111 for training, validation, and testing. To test model generalization, we follow [27] to split keypoints into base and novel sets, and report the performance of both base and novel keypoint detection on unseen species." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.784, + 0.733 + ], + "angle": 0, + "content": "Metric. The percentage of correct keypoints (PCK) is used. A predicted keypoint is correct if its distance to GT \\( d \\leq \\rho \\cdot \\max(w_{\\mathrm{bbx}}, h_{\\mathrm{bbx}}) \\), where \\( w_{\\mathrm{bbx}} \\) and \\( h_{\\mathrm{bbx}} \\) are the edges of object bounding box. Following [27], we set \\( \\rho \\) to 0.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.838 + ], + "angle": 0, + "content": "Implementation Details. The input image size for all models is \\(384 \\times 384\\). We freeze the CLIP text encoder but finetune last two layers of image encoder. The temperature \\(\\tau\\) used in contrastive loss is 0.05. Same to [27], the visual auxiliary keypoints are generated on pre-defined interpolation path, and we set \\(z = 0.5\\). For the text interpolation, to reduce randomness, we use GPT3.5 to reason \\(R = 3\\) times for all datasets except CUB with \\(R = 10\\). Our model is trained with 40k episodes. We report results using 1000 test episodes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.148, + 0.784, + 0.175 + ], + "angle": 0, + "content": "Table 1: Main results on the Animal pose dataset. Each row shows the results tested on the model trained in one setting. Each PCK score is the average of five subproblems." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.179, + 0.784, + 0.32 + ], + "angle": 0, + "content": "
#Training settings1-shot testing0-shot testing1-shot w/ text
KpAux. kpTextAux. textNovelBaseNovelBaseNovelBase
1✓21.3650.841.261.9316.1632.81
2✓✓47.5449.452.002.1039.2434.11
3✓16.1835.4625.6061.6427.0261.21
4✓✓15.8731.5763.1465.3163.3065.52
5✓✓21.4652.1522.4261.0723.8860.49
6✓✓✓✓50.3254.3963.3765.5963.1964.93
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.785, + 0.451 + ], + "angle": 0, + "content": "Compared Methods. For few-shot keypoint detection (FSKD), as in previous works, we compare the few-shot learning models such as ProtoNet [36], RelationNet [38], and LwoF [15], and FSKD-dedicated works \\( FSKD-R/-D \\) [27]. For zero-shot keypoint detection (ZSKD), we adopt the source code of CLAMP [56] and compare it in our experiments for fairness. Moreover, we build a Baseline sharing same backbone with our model but we do not use the specifically proposed auxiliary texts, contrastive loss, etc. We denote our method as OpenKD." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.373, + 0.484 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Firstly, we explore single-modal training and the impacts of whether or not to add the auxiliary keypoints/texts (1st-4th row, Table 1). When only using main keypoints (1st row) or main texts (3rd row), one can observe that the models perform well on base keypoints in 1-shot \\((50.84\\%)\\) or 0-shot \\((61.64\\%)\\), but fail to detect novel keypoints. However, if the model is trained by adding the auxiliary keypoints (2nd row) or auxiliary texts (4th row), one can observe the remarkable improvements on novel keypoint detection, e.g., \\(47.54\\%\\) vs. \\(21.36\\%\\) in 1-shot testing for the model trained by visual prompts (2nd row vs. 1st row); and also \\(63.14\\%\\) vs. \\(25.60\\%\\) in 0-shot testing for the model trained by text prompts (4th row vs. 3rd row). The significant gains confirm the benefits of visual interpolation [27] and also highlight the effectiveness of our proposed textual interpolation. Secondly, we investigate multimodal training (5th-6th row, Table 1). As one can see, our model performs well under multimodal training (6th row), and greatly outperforms the model trained by multimodal prompts without auxiliary keypoints and texts (5th row). The higher performance is due to the fact that auxiliary keypoints and texts could boost the visual and textual semantic diversity, thus enabling novel keypoint detection under 1-shot and 0-shot setting. Thirdly, we observe that 0-shot testing on base keypoints is much higher than 1-shot, which shows that the texts can be an excellent representation to guide keypoint detection. Lastly, we examine multimodal testing, i.e., performing 1-shot with text testing, our model obtains \\(63.19\\%\\) and \\(64.93\\%\\) on novel and base keypoints, which highlights that our model strongly combines the advantages of both modalities, mitigating the weakness of individual modality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.148, + 0.787, + 0.176 + ], + "angle": 0, + "content": "Table 2: Results on 1-shot keypoint detection for unseen species. The PCK scores on novel and base keypoint detection are reported. \\( + \\mathbf{T} \\) means adding texts during testing." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.179, + 0.782, + 0.43 + ], + "angle": 0, + "content": "
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelProtoNet19.6816.1814.3912.0515.0615.4729.5751.3236.65
RelationNet22.1517.1915.4713.5816.5516.9920.9156.5934.02
LwoF22.4719.3916.8216.4016.9418.4028.5454.7534.19
FSKD-R46.0540.6637.5538.0931.5038.7751.8177.9054.01
FSKD-D52.3647.9444.0742.7736.6044.7564.7677.8956.04
OpenKD60.3653.5847.5949.0141.0550.3266.7178.3953.35
OpenKD+T69.2666.8162.4063.2154.2763.1979.0273.2953.40
BaseProtoNet45.8039.8334.8835.8032.3337.7357.1780.3673.18
RelationNet51.0345.8539.8641.9737.1943.1857.3179.4078.85
LwoF50.0544.6443.4743.3537.8443.8763.8781.9681.39
FSKD-R57.1251.1247.8349.7143.7149.9065.2687.9487.84
FSKD-D56.3851.2948.2449.7743.9549.9366.3987.7186.99
OpenKD63.6155.4351.1853.8747.8654.3974.2287.4585.11
OpenKD+T70.4765.0962.8466.4659.8164.9384.5091.8191.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.46, + 0.787, + 0.489 + ], + "angle": 0, + "content": "Table 3: Results on the 0-shot keypoint detection for unseen species. \\(\\mathrm{CLAMP}^{\\dagger}\\) is the variant trained by adding our interpolated auxiliary texts." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.492, + 0.777, + 0.661 + ], + "angle": 0, + "content": "
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelBaseline25.6424.2925.2918.1425.5123.7728.9534.9430.00
CLAMP20.9024.0627.0716.8620.7321.9238.9637.0918.18
\\( CLAMP^† \\)61.7058.0961.4264.8653.1359.8477.6669.3050.81
OpenKD71.0766.4960.3065.5353.4563.3778.6470.1652.29
BaseBaseline58.6956.6256.5860.0051.3556.6573.4587.1271.25
CLAMP60.7154.0462.5862.7357.3059.4784.1690.9779.10
\\( CLAMP^† \\)60.3056.4962.6960.9057.1659.5183.7690.6583.65
OpenKD71.3466.6064.5367.2658.2365.5984.3291.7290.63
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparisons on FSKD and ZSKD: We comprehensively conduct the few-shot keypoint detection across four datasets. Table 2 shows our OpenKD model significantly outperforms compared methods in 15 out of 18 tasks in detecting novel or base keypoints. Moreover, if adding the texts during testing, the performance of our model further improves, yielding \\(63.19\\%\\) on the Animal pose dataset (63.19% vs. 44.75% of prior-art FSKD-D) and \\(79.02\\%\\) on AwA. We also notice the improvements on CUB and NAB are modest, which might be due to the auxiliary texts being relatively harder to reason, thus bringing lesser benefit to few-shot model. For ZSKD (Table 3), again, our model outperforms other" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.371, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.146, + 0.528, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.147, + 0.592, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.259, + 0.416, + 0.27 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.146, + 0.779, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.259, + 0.71, + 0.27 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Fig. 5: Different clustering effects of two modalities on base keypoints in unseen species (a) and statistical feature variance per keypoint (b)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Table 4: Ablation study. (a) Intra- and inter-modality contrastive loss. \\(\\mathcal{L}_{\\mathrm{vt}}^*\\) denotes no stop gradient on text features; (b) Text interpolation. Baseline†: w/o using auxiliary texts; CoT: chain of thought; Corr: Eq. 8; FTC: sampling with false text control (ours)." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.359, + 0.495, + 0.464 + ], + "angle": 0, + "content": "
AwA1-shot0-shot
NovelBaseNovelBase
1: w/o CL65.5672.1776.7181.67
2: Lt66.0572.3578.2884.00
3: Ltt+L*vt66.7073.9177.1083.85
4: Ltt+Lvt66.7174.2278.6484.32
5: Ltt+Lvt+Lvv66.2774.0878.6083.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.351, + 0.465, + 0.371, + 0.478 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.359, + 0.778, + 0.464 + ], + "angle": 0, + "content": "
0-shot testingAwACUB
NovelBaseNovelBase
1: Baselined† (▲)34.3183.5336.8988.85
2: ▲+w/o CoT73.8082.7968.0388.93
3: ▲+CoT78.3084.3268.6888.98
4: ▲+CoT+Corr62.1283.1371.8094.27
5: ▲+CoT+FTC78.6484.3270.1691.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.63, + 0.465, + 0.65, + 0.478 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.54 + ], + "angle": 0, + "content": "methods (in 17 out of 18 tasks). The improvement of \\(\\mathrm{CLAMP}^{\\dagger}\\) on novel keypoint detection strongly shows the benefits of adding auxiliary texts into training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.536, + 0.575 + ], + "angle": 0, + "content": "4.3 Model and Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.689 + ], + "angle": 0, + "content": "Why 0-shot can outperform 1-shot? One may observe the 0-shot testing scores on base keypoints are always higher than 1-shot testing. In this case, the training texts are accurate (cf. auxiliary text is noisy) and have high semantic similarity to testing texts. We dig more to visualize the corresponding text and keypoint features. Fig. 5 shows that textual keypoint features enjoy better clustering effects and lower variance than visual ones, which suggests textual representations of base keypoints are better thus 0-shot yields higher scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Contrastive Learning (CL). After applying \\(\\mathcal{L}_{\\mathrm{tt}}\\), we observe the 0-shot performance significantly improves, e.g. \\(84.00\\%\\) vs. \\(81.67\\%\\) (2nd row, Table 4a), as CL improves the discrimination between text features. However, when adding visual-textual alignment (3rd row), 1-shot testing scores boost but slightly decreasing 0-shot scores. We suspect it is due to the imbalance of representation quality between multimodal features. After adding stop gradient, it mitigates negative impacts from \"weak\" modality, thus achieving higher scores (4th row). We also observe \\(\\mathcal{L}_{\\mathrm{vv}}\\) does not further improve scores in our losses (5th row)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Text Interpolation. Table 4b shows that using chain of thought (CoT) significantly improves performance, especially on AwA (3rd row), as CoT improves the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.147, + 0.774, + 0.162 + ], + "angle": 0, + "content": "Table 5: (a) Study on repetition \\( R \\) ; (b) Parsing results with GPT3.5 and Vicuna." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.165, + 0.486, + 0.27 + ], + "angle": 0, + "content": "
RAwACUB
NovelBaseNovelBase
167.3483.9864.2191.40
378.64(+11.3)84.3269.81(+5.60)92.38
576.36(+9.02)83.7369.81(+5.60)92.02
1076.61(+9.27)83.3670.16(+5.95)91.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.355, + 0.272, + 0.376, + 0.285 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.165, + 0.788, + 0.27 + ], + "angle": 0, + "content": "
DatasetGPT3.5Vicuna
Acc. kpAcc. objAcc. kpAcc. obj
Animal0.970.990.940.98
AwA0.961.000.930.98
CUB0.970.960.941.00
NABird0.970.990.951.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.631, + 0.272, + 0.651, + 0.285 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.303, + 0.776, + 0.318 + ], + "angle": 0, + "content": "Table 6: Diverse text prompting evaluation. Results are the average of three runs." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.32, + 0.777, + 0.428 + ], + "angle": 0, + "content": "
MethodAnimal PoseAwACUBNABAvg Drop
NovelBaseNovelBaseNovelBaseNovelBaseNovelBase
OpenKD63.3765.5978.6484.3270.1691.7252.2990.63--
No parsing13.8915.1813.4421.7022.0917.4814.5517.6250.12↓65.07↓
Vicuna57.5459.8473.9778.1867.8987.3950.4586.593.65↓5.06↓
GPT3.561.4562.8276.2380.0769.3990.6552.2189.141.29↓2.39↓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.785, + 0.515 + ], + "angle": 0, + "content": "quality of reasoned texts. After applying FTC, the scores strike a high balance in all datasets (5-th row), as our FTC rejects low-quality texts and further reduces noise after including visual features into decision. We also explore the repetition \\( R \\) for text reasoning. Table 5a shows the importance of setting \\( R > 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.787, + 0.636 + ], + "angle": 0, + "content": "Diverse Text Prompting. To evaluate our model under diverse text prompting, we randomly sample 1000 diverse texts from each dataset to conduct zero-shot testing. As shown in Table 6, after coupling with GPT3.5, our model retains the high performance even under diverse texts (with less than \\(3\\%\\) drop). The strong results compared to no parsing show that leveraging LLM as a parser to handle language diversity is a possible solution. Moreover, we also found coupling with GPT3.5 has higher performance compared to Vicuna, as GPT3.5 enjoys a stronger keypoint text parsing ability, e.g., \\(96\\%\\) vs. \\(93\\%\\) in AwA (Table 5b)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.658, + 0.359, + 0.674 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose to open the prompt diversity from the aspects of modality, semantics and language, to enable a more general zero- and few-shot keypoint detection. To this end, we build a versatile OpenKD model which supports both visual and textual prompting. Moreover, to bridge the semantics gap between seen and unseen texts, we propose a novel text interpolation and the selection strategy with false text control, which significantly improves zero-shot novel keypoint detection. We also discover that LLM is a good language parser. After coupling with LLM, our model can handle diverse texts well. We hope the proposed model, text interpolation and the parsing approach will provide useful insights for versatile keypoint detection, thus we highly recommend it to vision community." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.262, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.794, + 0.224 + ], + "angle": 0, + "content": "Changsheng Lu is supported by Australian Government Research Training Program (AGRTP) International Scholarship. Piotr Koniusz is supported by CSIRO's Science Digital." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.249, + 0.323, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.786, + 0.323 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.786, + 0.351 + ], + "angle": 0, + "content": "2. Banik, P., Li, L., Dong, X.: A novel dataset for keypoint detection of quadruped animals from images. arXiv preprint arXiv:2108.13958 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.786, + 0.407 + ], + "angle": 0, + "content": "3. Bohdal, O., Tian, Y., Zong, Y., Chavhan, R., Li, D., Gouk, H., Guo, L., Hospedales, T.: Meta omnium: A benchmark for general-purpose learning-to-learn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7693-7703 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.786, + 0.449 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.786, + 0.491 + ], + "angle": 0, + "content": "5. Cao, J., Tang, H., Fang, H.S., Shen, X., Lu, C., Tai, Y.W.: Cross-domain adaptation for animal pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9498-9507 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.786, + 0.533 + ], + "angle": 0, + "content": "6. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: realtime multiperson 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence 43(1), 172-186 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.786, + 0.575 + ], + "angle": 0, + "content": "7. Carreira, J., Agrawal, P., Fragkiadaki, K., Malik, J.: Human pose estimation with iterative error feedback. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4733-4742 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.786, + 0.63 + ], + "angle": 0, + "content": "8. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.786, + 0.686 + ], + "angle": 0, + "content": "9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.688, + 0.743, + 0.7 + ], + "angle": 0, + "content": "10. Derpanis, K.G.: The harris corner detector. York University pp. 2-3 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.702, + 0.786, + 0.756 + ], + "angle": 0, + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.786, + 0.797 + ], + "angle": 0, + "content": "12. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "13. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: International Conference on Machine Learning. pp. 1126-1135. PMLR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.282, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "14. Ge, Y., Zhang, R., Luo, P.: Metacloth: Learning unseen tasks of dense fashion landmark detection from a few samples. IEEE Transactions on Image Processing 31, 1120-1133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "15. Gidaris, S., Komodakis, N.: Dynamic few-shot visual learning without forgetting. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4367-4375 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "16. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.786, + 0.3 + ], + "angle": 0, + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "18. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1546-1555 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "19. Jiao, B., Liu, L., Gao, L., Wu, R., Lin, G., Wang, P., Zhang, Y.: Toward re-identifying any animal. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.412 + ], + "angle": 0, + "content": "20. Koch, G., Zemel, R., Salakhutdinov, R., et al.: Siamese neural networks for one-shot image recognition. In: ICML deep learning workshop. vol. 2. Lille (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "21. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.786, + 0.481 + ], + "angle": 0, + "content": "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.536 + ], + "angle": 0, + "content": "23. Li, S., Gunel, S., Ostrek, M., Ramdya, P., Fua, P., Rhodin, H.: Deformation-aware unpaired image translation for pose estimation on laboratory animals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13158-13168 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.577 + ], + "angle": 0, + "content": "24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.744, + 0.591 + ], + "angle": 0, + "content": "25. Lowe, G.: Sift-the scale invariant feature transform. Int. J 2, 91-110 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.632 + ], + "angle": 0, + "content": "26. Lu, C., Gu, C., Wu, K., Xia, S., Wang, H., Guan, X.: Deep transfer neural network using hybrid representations of domain discrepancy. Neurocomputing 409, 60-73 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.674 + ], + "angle": 0, + "content": "27. Lu, C., Koniusz, P.: Few-shot keypoint detection with uncertainty learning for unseen species. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19416-19426 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.715 + ], + "angle": 0, + "content": "28. Lu, C., Koniusz, P.: Detect any keypoints: An efficient light-weight few-shot keypoint detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3882-3890 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.743 + ], + "angle": 0, + "content": "29. Lu, C., Zhu, H., Koniusz, P.: From saliency to dino: Saliency-guided vision transformer for few-shot keypoint detection. arXiv preprint arXiv:2304.03140 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.785 + ], + "angle": 0, + "content": "30. Mathis, A., Mamidanna, P., Cury, K.M., Abe, T., Murthy, V.N., Mathis, M.W., Bethge, M.: Deeplabcut: markerless pose estimation of user-defined body parts with deep learning. Nature neuroscience 21(9), 1281-1289 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.813 + ], + "angle": 0, + "content": "31. Moskvyak, O., Maire, F., Dayoub, F., Baktashmotlagh, M.: Semi-supervised keypoint localization. arXiv preprint arXiv:2101.07988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.786, + 0.841 + ], + "angle": 0, + "content": "32. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.176 + ], + "angle": 0, + "content": "In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19413-19423 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.232 + ], + "angle": 0, + "content": "33. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 483-499. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.274 + ], + "angle": 0, + "content": "34. Pereira, T.D., Aldarondo, D.E., Willmore, L., Kislin, M., Wang, S.S.H., Murthy, M., Shaevitz, J.W.: Fast animal pose estimation using deep neural networks. Nature methods 16(1), 117-125 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.331 + ], + "angle": 0, + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.786, + 0.36 + ], + "angle": 0, + "content": "36. Snell, J., Swersky, K., Zemel, R.S.: Prototypical networks for few-shot learning. arXiv preprint arXiv:1703.05175 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.786, + 0.402 + ], + "angle": 0, + "content": "37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "38. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1199-1208 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "39. Tang, L., Wertheimer, D., Hariharan, B.: Revisiting pose-normalization for fine-grained few-shot recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14352-14361 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.53 + ], + "angle": 0, + "content": "40. Thompson, J.J., Jain, A., LeCun, Y., Bregler, C.: Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.572 + ], + "angle": 0, + "content": "41. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.573, + 0.787, + 0.628 + ], + "angle": 0, + "content": "42. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 595-604 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.787, + 0.67 + ], + "angle": 0, + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.712 + ], + "angle": 0, + "content": "44. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. Advances in neural information processing systems 29, 3630-3638 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "45. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "46. Wang, C., Jin, S., Guan, Y., Liu, W., Qian, C., Luo, P., Ouyang, W.: Pseudolabeled auto-curriculum learning for semi-supervised keypoint localization. arXiv preprint arXiv:2201.08613 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Wang, L., Huynh, D.Q., Koniusz, P.: A comparative review of recent kinect-based action recognition algorithms. IEEE Transactions on Image Processing 29, 15-28 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "C. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "49. Wu, Z., Su, L., Huang, Q.: Stacked cross refinement network for edge-aware salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7264-7273 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "50. Xu, L., Jin, S., Zeng, W., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Pose for everything: Towards category-agnostic pose estimation. In: European Conference on Computer Vision. pp. 398-416. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "51. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "52. Yang, S., Quan, Z., Nie, M., Yang, W.: Transpose: Keypoint localization via transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11802-11812 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.397 + ], + "angle": 0, + "content": "53. Yao, Q., Quan, Q., Xiao, L., Kevin Zhou, S.: One-shot medical landmark detection. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 177-188. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.438 + ], + "angle": 0, + "content": "54. Zhang, H., Zhang, K., Xu, L., Lai, S., Shao, W., Zheng, N., Luo, P., Qiao, Y.: Language-driven open-vocabulary keypoint detection for animal body and face. arXiv preprint arXiv:2310.05056 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.495 + ], + "angle": 0, + "content": "55. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.55 + ], + "angle": 0, + "content": "56. Zhang, X., Wang, W., Chen, Z., Xu, Y., Zhang, J., Tao, D.: Clamp: Prompt-based contrastive learning for connecting language and animal pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23272-23281 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "57. Zhao, S., Gong, M., Zhao, H., Zhang, J., Tao, D.: Deep corner. International Journal of Computer Vision pp. 1-25 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "58. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.675 + ], + "angle": 0, + "content": "59. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_origin.pdf b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e626f297bbfd28099ffc1fdc7712317bc420917f --- /dev/null +++ b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/dd93a63c-860b-4bc5-8581-032ed835a163_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a424344b76a0232b3d0088e37db97af7d853fa902698673aa3f41a60a1e533c +size 1718966 diff --git a/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/full.md b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a0d12b7d8c8e184be9dd34be8b98bb07d07fdf5e --- /dev/null +++ b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/full.md @@ -0,0 +1,343 @@ +# OpenKD: Opening Prompt Diversity for Zero-and Few-shot Keypoint Detection + +Changsheng Lu1, Zheyuan Liu1, and Piotr Koniusz*,2,1 + +1The Australian National University 2Data61/CSIRO changshengluu@gmail.com, zheyuan.david.liu@outlook.com, piotr.koniusz@data61.csiro.au + +Abstract. Exploiting foundation models (e.g., CLIP) to build a versatile keypoint detector has gained increasing attention. Most existing models accept either the text prompt (e.g., "the nose of a cat"), or the visual prompt (e.g., support image with keypoint annotations), to detect the corresponding keypoints in query image, thereby, exhibiting either zero-shot or few-shot detection ability. However, the research on multimodal prompting is still underexplored, and the prompt diversity in semantics and language is far from opened. For example, how to handle unseen text prompts for novel keypoint detection and the diverse text prompts like "Can you detect the nose and ears of a cat?" In this work, we open the prompt diversity in three aspects: modality, semantics (seen vs. unseen), and language, to enable a more general zero- and few-shot keypoint detection (Z-FSKD). We propose a novel OpenKD model which leverages a multimodal prototype set to support both visual and textual prompting. Further, to infer the keypoint location of unseen texts, we add the auxiliary keypoints and texts interpolated in visual and textual domains into training, which improves the spatial reasoning of our model and significantly enhances zero-shot novel keypoint detection. We also find large language model (LLM) is a good parser, which achieves over $96\%$ accuracy when parsing keypoints from texts. With LLM, OpenKD can handle diverse text prompts. Experimental results show that our method achieves state-of-the-art performance on Z-FSKD and initiates new ways of dealing with unseen text and diverse texts. The source code and data are available at https://github.com/AlanLuSun/OpenKD. + +# 1 Introduction + +Keypoint detection is a fundamental research problem in computer vision and has numerous applications such as pose estimation for humans [6,8,12,33,37, 52] and animals [2,5,23,34], action recognition [30,47], and fine-grained image classification [39,55]. Over the past decade, significant advancements have been made in deep keypoint detection. However, most existing methods are tailored to close-set detection and struggle to predict novel body parts and species with limited data. Thus, it necessitates such a zero- or few-shot keypoint detection: + +![](images/6ad9eb483430f672d537314cbf53fe745ba8da0ec13c63a2ff5371291e28867f.jpg) +(a) +Visual prompt +Query image + +![](images/eeddedc453b57159ecbeca957bfdd1fecf49d36147cb73b3c82fe732cb6ba0dc.jpg) +(b) +Text prompt +Query image + +![](images/4b3173369d2239a793bbb82c3b6e78217ec0d99280db7dfb7c9937b1d9e3e06b.jpg) +(c) +Visual + Text prompt + +![](images/191f43a2b19a7913c29425d4d0be443d6f81f9c025d80fffe649a8e9b925114f.jpg) +(d) +Results on AwA dataset One shot Zero shot One shot w/ text +Fig. 1: Illustration of multimodal prompting for keypoint detection. Our model can successfully detect keypoints given visual prompts formed by support images and keypoints (a), text prompts (b), or both (c). Graph (d) shows our model well combines the advantages of different modalities, mitigating the weakness induced by either modality. + +after being trained on a diverse dataset, the model can quickly recognize novel or base keypoints in unseen species given only zero, one or a few labelled samples. + +On the one hand, researchers have recently explored few-shot keypoint detection (FSKD) [27-29, 53]. The FSKD inspired by few-shot learning [13, 20, 36, 38, 44] is general and offers higher flexibility of detecting a varying number of keypoints in a query image, given the prompts formed by support image with keypoint annotations. In this regard, the support set is namely visual prompt, which is required when evaluating new classes. On the other hand, in the era of foundation model, the vision-language model (e.g., CLIP [35]) injects new life to detecting keypoints, enabling zero-shot keypoint detection (ZSKD) [54, 56]. Compared to FSKD, ZSKD does not need support image with annotations, instead, using the text prompt to instruct the model to detect the specified keypoints in query image. In this case, the support set is text prompt. Though impressive for the pioneering FSKD and ZSKD models, we identify an important issue limiting the progress: the prompt diversity for current keypoint detection methods is far from opened, especially in the aspects of modality, semantics, and language. + +Modality diversity. Most existing keypoint detection models cannot support multimodal prompts, e.g., image, text, or both (see Fig. 1). Multimodal prompting is more friendly in real-world interaction, and coherent with the human concept recognition. We not only see the objects, but also describe objects with language. Our work extends existing FSKD and ZSKD, building a more general zero- and few-shot keypoint detection by leveraging a multimodal prototype set and aligning the visual keypoint features towards the textual features. While straightforward, our method enables one to study the advantages of respective modality data and exploit them for better model training and testing. + +Semantic diversity. Considering keypoints between seen and unseen species, there exist large similarity yet difference in semantics. A strong advantage of text prompt is that the keypoints with the same semantics share high similarity in language across species, which enables excellent ZSKD on base keypoints. However, if the text has different semantics, significant domain gap arises. For instance, "the eye of a cat" vs. "the eye of a dog" has cosine similarity of 0.93 using CLIP + +![](images/edc17f53fcff8c7a4630b46334b70b5498c5751eafb02e71876a3c5942cd8f39.jpg) +Can you make out the back-right knee, front-left knee, body-middle-left on cow? +(a) +Fig. 2: Examples of keypoint detection under diverse text prompting. With LLM, our method can deal with diverse texts, showing potential for real-world applications. The circles and crosses refer to predictions and GT, respectively. A keypoint is regarded as a correct detection if falling in the white shadow area that signifies PCK@0.1. + +![](images/d4dfedab7274c9b8fb5b3e53d2855bc6555fa4f09036a74dbb7c777aea709bf2.jpg) +Can you detect the body-middle-right, nose, left earbase, front-right paw, left eye of grizzly bear? +(b) + +![](images/a2dfed2b56fc4d8c6f9153bb7fe7486675ef237eac93d87e8fe76088f34c307a.jpg) +Please find the throat, belly, breast, back, left leg for the evening grosbeak. +(c) + +![](images/bb53182c4349e9bc99cd7d84c6f4d5c337d6351eebb881579c0292134217ab41.jpg) +Where is the breast, throat for pine warbler? +(d) + +![](images/4487459394e3567a13976a47814d8e0d171b141de22a36cbfaa6759b6a8988da.jpg) +Can you recognize the back, crown, throat, tail, belly? +(e) + +text embeddings, whereas for "the eye of a cat" vs. "the knee of a dog" it drops to 0.77. We observe that the keypoint detectors cannot perform well if text is unseen during training. To bridge the gap, we propose to open the semantic diversity by adding auxiliary keypoints and texts into training, where the auxiliary texts are reasoned by LLM given the base keypoints. To ease the reasoning error, we ask LLM a good question using chain of thought [48]. We also propose a novel sampling strategy with false text control to improve the matching between auxiliary keypoints and texts. With them together, our model significantly improves ZSKD on novel keypoints. + +Language diversity. Following zero-shot image classification by CLIP [35], existing ZSKD also constructs simple text prompts based on templates, e.g., "〈keypoint〉" or "the〈keypoint〉 of a〈category〉 in the photo", where $\langle \cdot \rangle$ is replaced by class names. In real-world interaction, humans tend to question in diverse styles, which results in diverse text prompts. For example, "Please locate the right-back leg on cat", "Can you find the left eye, right ear and nose in image?", etc. A natural question is how to handle the diverse text prompting? To address this, we propose a simple yet effective approach, whose key idea is to borrow the large language model (LLM) such as GPT3.5 [4] or Vicuna [9, 58] to parse the diverse texts via prompt engineering. After extracting the keypoint type and object category, they will be synthesized into simple text prompts to instruct the detection model. As such, the diverse text prompting can be transformed into simple text prompting, which sheds light on opening language diversity (Fig. 2). + +To summarize, in this paper, we propose an OpenKD model with several intriguing features: 1) supporting both visual and textual prompts, 2) having the potential to handle unseen texts and diverse texts, and 3) maintaining strong generality and performance on ZSKD and FSKD. We report that LLM is capable of being a reasoner for text interpolation, and a good language parser for parsing diverse texts. We also contribute four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42] for fair evaluations. To our best knowledge, we are the first to open semantics and language diversity of text prompts for keypoint detection. + +# 2 Related Work + +Keypoint Detection has been widely studied ranging from the traditional interest points [10, 25] to deep corner detection [57], semi-supervised [18, 31, 46] and fully-supervised keypoint estimation [6, 8, 12, 33, 37, 40, 51, 52]. There are two major classes of methods for deep keypoint localization: i) direct coordinates based regression [7, 41] and ii) heatmap based regression. In contrast to existing heatmap based models dedicated to specific body parts, e.g., top-down [12, 16, 37] and bottom-up pose estimators [6, 8, 33], our OpenKD model offers more flexible keypoint detection, breaking the limitation of keypoint types to be detected. + +Few-shot Keypoint Detection is more versatile and data-efficient compared to the supervised paradigms. Similar to standard few-shot learning (FSL) [44], FSKD also takes episodes for training and evaluation. Many well-known FSL methods such as ProtoNet [36], RelationNet [38], LwoF [15] and MAML [13] have been extended to the field of keypoint detection [3,14,27] and serve as baselines. Recently, Lu and Koniusz [27] formalize the comprehensive FSKD settings and model the localization uncertainty of keypoints. Further, they propose a lightweight FSKD model [28] and explore FSKD under transductive setting and occlusions [29]. Bohdal et al. [3] propose a dataset-of-datasets, benchmarking the FSL algorithms universal to various vision tasks. FSKD also inspires class-agnostic animal pose estimation [27, 28, 50]. Compared to existing FSKD, our work pushes further by opening the text prompting, which enables more versatile keypoint detection. Moreover, text prompts can complement visual prompts, delivering better performance of detecting base keypoints on unseen species. + +Zero-shot Keypoint Detection becomes feasible thanks to the great progress of vision-language pre-training (e.g., CLIP [35] and BLIP [22]). Since CLIP was pre-trained on internet-scale image-text pairs that contain more semantic concepts than other datasets, CLIP can provide semantically rich features for various downstream tasks [32, 56, 59], showing strong zero-shot transfer ability. + +Recently, CLAMP [56] adapts CLIP for animal pose estimation in feature-aware and spatial-aware aspects; Approach [54] also employs CLIP as backbone and proposes language-driven keypoint detection. Compared to FSKD which still requires one or more annotations, ZSKD merely demands cheap language descriptions, which make it intriguing for keypoint detection. Nevertheless, most existing ZSKD works cannot support multimodal prompts and their prompt diversity is limited. To address these issues, we propose an OpenKD model which further opens modality, semantics, and language. To handle unseen texts, we propose to generate auxiliary keypoint-text pairs which significantly help reason novel keypoints on unseen species. Moreover, we contribute a simple yet effective language parsing module to handle diverse text prompts. + +Foundation Models include a broad set of models pre-trained over large-scale datasets, including LLMs (e.g., GPT3.5 [4]), VLMs (e.g., CLIP [35]), and Multimodal models (e.g., GPT4 [1]). Transferring the knowledge from these foundation models to specific tasks in a cost-effective way is popular in computer vision [19,24,26,54,56]. Thus, we follow this trend and leverage LLM like GPT3.5 (relatively cheap) to perform reasoning and parsing for keypoints. + +# 3 Method + +# 3.1 Model Architecture + +Following the general few-shot learning [36, 38, 44], Z-FSKD model is evaluated on episodes, each of which includes a support set and a query set. The query set is comprised of query images, while the support set gives the prompts. If the visual prompts, i.e., $K$ support images with keypoint annotations ( $K \geq 1$ ), are given, then the problem is defined as $K$ -shot detection. If only the text prompt (i.e., the language description) is given, then $K = 0$ thus the problem becomes zero-shot detection. The goal of Z-FSKD is to detect the corresponding keypoints in query image given the prompts, whether visual prompt, text prompt, or both. Such an approach allows the model to effectively respond to diverse inputs. + +The overview of our model inference is shown in Fig. 3, which mainly includes four stages: i) image/text feature extraction, ii) feature adaptation, iii) keypoint prototype set building, iv) correlation, decoding and heatamp fusion. The overview of our model training is shown in Fig. 4, which includes the novel approaches for performance improvement. + +Image/Text Feature Extraction Since the input of our model involves vision and language two different modalities, we wish the features extracted from image and text have smaller modality gap, so that the text can somewhat correlate with image regions of the same semantics. Such a property helps the model find keypoint locations in query image. To this end, we resort to CLIP [35] pre-trained on large-scale image-text pairs. The CLIP includes image and text encoders. The text encoder is generally a transformer [43] while the image encoder can be CNN [21] or ViT [11]. We empirically found that the CLIP image encoder based on RN50 [17, 35], a CNN model, gives high efficiency in consideration of both performance and cost. Thus, we choose it as the default backbone. When extracting the image features, original CLIP image encoder only retains the classification token $\mathbf{x}_{\mathrm{class}}$ and discards image tokens $\mathbf{X}$ by attention pooling as + +$$ +\mathbf {x} _ {\text {c l a s s}} = \operatorname {A v g P o o l} (\mathbf {X}), \quad \mathbf {x} _ {\text {c l a s s}} ^ {\prime} = \operatorname {A t t e n t i o n} \left(\mathbf {x} _ {\text {c l a s s}} \mathbf {W} _ {q}, \mathbf {X} \mathbf {W} _ {k}, \mathbf {X} \mathbf {W} _ {v}\right) \mathbf {W} _ {o}. \tag {1} +$$ + +While it is natural for image-text matching, we require the image tokens to recover the spatial locations of keypoints. Thus, we propose to obtain the image tokens via a projection composed by $\mathbf{W}_v$ and $\mathbf{W}_o$ as + +$$ +\mathbf {X} ^ {\prime} = \mathbf {X} \mathbf {W} _ {v} \mathbf {W} _ {o}. \tag {2} +$$ + +In this way, we can not only handle the channel inconsistency between raw image tokens $\mathbf{X}$ and text features, but also reuse the projection layers $\mathbf{W}_v$ and $\mathbf{W}_o$ . + +For brevity, we assume that per input episode contains one query image $\mathbf{I}^{\mathrm{q}}$ , one support image $\mathbf{I}^{\mathrm{s}}$ with $N$ annotated keypoints, and $N$ keypoint texts. With Eq. 2, both the support and query images can be encoded and projected as the support and query feature maps as $\mathbf{X}^{\mathrm{s}} = \mathcal{F}_{\mathrm{v}}(\mathbf{I}^{\mathrm{s}})$ and $\mathbf{X}^{\mathrm{q}} = \mathcal{F}_{\mathrm{v}}(\mathbf{I}^{\mathrm{q}})$ in deep feature space $\mathbb{R}^{l\times l\times d}$ via a shared CLIP image encoder $\mathcal{F}_{\mathrm{v}}$ . Further, with the CLIP text encoder $\mathcal{F}_{\mathrm{t}}$ , the $N$ texts are firstly tokenized, then encoded as the text features $\mathbf{t}_n\in \mathbb{R}^{m\times d}$ , where $n = 1,2,\dots ,N$ ; and $m$ is sequence length. + +![](images/e71fb7a801ead380a8c60b5890eccc813b55be1ecd8c053e1b174e17a4e4a28d.jpg) +Fig. 3: The sketch of model inference. Our OpenKD allows testing under visual prompt, text prompt, or both. For clarification, we show the "both" case (i.e., 1-shot with text testing). We firstly extract the deep features of texts, support and query images via CLIP, and then adapt both modalities of features via residual refinement. After extracting the visual keypoint prototype (VKP) and textual counterpart, we build the prototype set to perform class-agnostic correlation and heatmap decoding. Finally, we fuse the heatmaps induced by two modalities (i.e., M1 & M2) to obtain predictions. + +Feature Adaptation As CLIP provides general yet transferrable features, we adapt it to a new multimodal feature space suitable for keypoint detection. Inspired by residual learning [17], we propose two adaptation nets $\mathcal{A}_{\mathrm{v}}$ and $\mathcal{A}_{\mathrm{t}}$ to respectively refine the obtained image and text features in a residual way: + +$$ +\mathbf {X} ^ {\mathrm {s}} := \mathbf {X} ^ {\mathrm {s}} + \mathcal {A} _ {\mathrm {v}} \left(\mathbf {X} ^ {\mathrm {s}}\right), \quad \mathbf {X} ^ {\mathrm {q}} := \mathbf {X} ^ {\mathrm {q}} + \mathcal {A} _ {\mathrm {v}} \left(\mathbf {X} ^ {\mathrm {q}}\right), \quad \mathbf {t} _ {n} := \mathbf {t} _ {n} + \mathcal {A} _ {\mathrm {t}} \left(\mathbf {t} _ {n}\right). \tag {3} +$$ + +The adaptation net $\mathcal{A}$ is highly scalable and can be based on transformer [43] or bottleneck [17] as refinement blocks. We found that $\mathcal{A}$ can adapt features well. + +Keypoint Prototype Set Once we obtain the adapted features of support image and text prompts, i.e., $\mathbf{X}^{\mathrm{s}}$ and $\mathbf{t}_n$ , we propose to convert them into keypoint prototypes, which are unified in a keypoint prototype set. + +For visual prompt, recall that keypoint labels are provided along with the support image. Thus, one can aggregate the local features for each support keypoint $\mathbf{p}_n$ by using pixelwise weighted summation between feature map $\mathbf{X}^{\mathrm{s}}$ and Gaussian heatmap $\mathbf{H}(\mathbf{p}_n;\sigma)$ , yielding the visual keypoint representation (VKR) as $\Phi_n\in \mathbb{R}^d$ . The $\sigma$ is the standard deviation that controls the Gaussian spread. If $K$ support images are given, namely in few-shot case, the VKRs of the same type of keypoints $\Phi_{k,n}$ will be averaged to build the visual keypoint prototype (VKP) $\Psi_{n}^{\mathrm{v}} = \frac{1}{K}\sum_{k}\Phi_{k,n}$ . Analogously, one can develop textual keypoint prototype (TKP) $\Psi_{n}^{\mathrm{t}}$ by text features if provided with multiple texts per keypoint. Finally, one can build the keypoint prototype set as $\mathcal{T} = \mathcal{T}^{\mathrm{v}}\cup \mathcal{T}^{\mathrm{t}}$ , where $\mathcal{T}^{\mathrm{v}} = \{\Psi_n^{\mathrm{v}}\}$ is the VKPs and $\mathcal{T}^{\mathrm{t}} = \{\Psi_n^{\mathrm{t}}\}$ is the TKPs. In this way, prompts of different modalities can be summarized as keypoint prototypes in a shared $d$ dimensional feature space, which enables our model to flexibly handle various modalities. Moreover, each prototype in set $\mathcal{T}$ can guide the model to induce a keypoint heatmap, establishing the zero- and few-shot keypoint detection. + +![](images/81028be5d25f5ed9afb43f7501320e54413ea1abb266a4d015de338622d577d6.jpg) +(a) Model training + +![](images/99d668236a6c35de77c09fab149453a4ae2da4e27dd33cb61065baac7cfa5e11.jpg) +(b) Auxiliary text interpolation +Fig. 4: Model training and text interpolation. (a) In addition to multi-group heatmap regression, we improve model performance by introducing intra- and inter-modality contrastive learning and the novel auxiliary keypoint and text learning. (b) We exploit LLM for auxiliary texts interpolation and explore incorporating visual keypoint features for text selection in order to mitigate the noisy texts or false texts. + +Correlation, Decoding and Fusion To discover the corresponding keypoints in query image $\mathbf{I}^{\mathrm{q}}$ , each keypoint prototype $\Psi_{n}$ , either VKP $\Psi_{n}^{\mathrm{v}}$ or TKP $\Psi_{n}^{\mathrm{t}}$ , is required to be correlated with the query feature map $\mathbf{X}^{\mathrm{q}}$ . To this end, we leverage a correlation module $\mathcal{C}$ which takes both $\Psi_{n}$ and $\mathbf{X}^{\mathrm{q}}$ as input to produce the attentive feature maps $\mathbf{A}_{n} = \mathcal{C}(\Psi_{n},\mathbf{X}^{\mathrm{q}})$ , where $\mathbf{A}_{n}\in \mathbb{R}^{l\times l\times d}$ and $n = 1,2,\dots ,N$ . For the correlation module $\mathcal{C}$ , we explored multiple variants and found that the simple cross correlation already yields high performance, i.e., $\mathbf{A}_n = \mathbf{X}^{\mathrm{q}}\odot \Psi_n$ , where $\odot$ denotes the channel-wise multiplication. + +Subsequently, a class-agnostic decoder $\mathcal{D}$ is devised to convert each attentive feature map into a keypoint heatmap, i.e., $\mathbf{H}_n = \mathcal{D}(\mathbf{A}_n)$ , where $\mathbf{H}_n \in \mathbb{R}^{l \times l}$ . For heatmap regression based keypoint localization, the higher resolution of heatmap can greatly reduce the coordinate decoding error. Thus, a lightweight upsampling module $\mathcal{U}$ is adopted to further refine heatmaps, i.e., $\mathbf{H}_n := \mathcal{U}(\mathbf{H}_n)$ . + +Since each modality of prototypes can induce one group of heatmaps, considering visual and textual modalities, we have two groups of heatmaps $\mathbf{H}^{\mathrm{v}} = \{\mathbf{H}_n^{\mathrm{v}}\}_{n=1}^N$ and $\mathbf{H}^{\mathrm{t}} = \{\mathbf{H}_n^{\mathrm{t}}\}_{n=1}^N$ . In testing phase, we fuse the upsampled multi-group heatmaps as the final output $\mathbf{H} \in \mathbb{R}^{N \times ul \times ul}$ ( $u$ is upsampling factor): + +$$ +\mathbf {H} = \left(\mathbf {H} ^ {\mathrm {v}} + \mathbf {H} ^ {\mathrm {t}}\right) / 2. \tag {4} +$$ + +During training phase, inspired by HigherHRNet [8], we perform group-specific supervision for heatmaps as + +$$ +\mathcal {L} _ {\mathrm {k p}} = \left(\| \mathbf {H} ^ {\mathrm {v}} - \mathbf {H} ^ {*} \| _ {2} ^ {2} + \| \mathbf {H} ^ {\mathrm {t}} - \mathbf {H} ^ {*} \| _ {2} ^ {2}\right) / 2, \tag {5} +$$ + +where $\mathbf{H}^*$ denotes the groundtruth heatmap that encodes the query keypoint. + +Improving Model by Contrastive Learning Since the keypoint prototype directly impacts performance, we aim for its representativeness to be as good as possible. The same type of keypoints should have higher invariance across species, e.g., the ear of cat, dog, and cow, while different types are required to + +be sufficiently discriminative such that our model can distinguish the ambiguous texts or fine-grained texts such as "the left ear" vs. "the right ear". To this end, we propose to introduce a contrative loss over TKPs. If randomly sampling two species $(s,s^{\prime})$ at a time and each species pertaining to an episode, we have pairwise sets of TKPs, i.e., $\mathcal{T}_s^{\mathrm{t}} = \{\Psi_n\}$ and $\mathcal{T}_{s^{\prime}}^{\mathrm{t}} = \{\Psi_{n}^{\prime}\}$ , which can form similarity matrix as + +$$ +\mathbf {J} \left(\mathcal {T} _ {s} ^ {\mathrm {t}}, \mathcal {T} _ {s ^ {\prime}} ^ {\mathrm {t}}\right) = \left( \begin{array}{c c} \cos \left(\Psi_ {1}, \Psi_ {1} ^ {\prime}\right) & \dots \cos \left(\Psi_ {1}, \Psi_ {N} ^ {\prime}\right) \\ \vdots & \vdots \\ \cos \left(\Psi_ {N}, \Psi_ {1} ^ {\prime}\right) & \dots \cos \left(\Psi_ {N}, \Psi_ {N} ^ {\prime}\right) \end{array} \right), \tag {6} +$$ + +where $\cos (\cdot ,\cdot)$ denotes the cosine similarity. Then, the contrastive loss within textual modality $\mathcal{L}_{\mathrm{tt}}$ becomes + +$$ +\mathcal {L} _ {\mathrm {t t}} ^ {s \rightarrow s ^ {\prime}} = - \langle \mathbb {I}, \log (\operatorname {s o f t m a x} (\mathbf {J} (\mathcal {T} _ {s} ^ {\mathrm {t}}, \mathcal {T} _ {s ^ {\prime}} ^ {\mathrm {t}}) / \tau)) \rangle , \quad \mathcal {L} _ {\mathrm {t t}} = \frac {1}{2} (\mathcal {L} _ {\mathrm {t t}} ^ {s \rightarrow s ^ {\prime}} + \mathcal {L} _ {\mathrm {t t}} ^ {s ^ {\prime} \rightarrow s}), \qquad (7) +$$ + +where $\mathbb{I}$ is the identity matrix and $\langle \cdot ,\cdot \rangle$ denotes inner product. Since CLIP is trained via image-level alignment instead of keypoint-level alignment with text, we wish VKPs and TKPs were aligned so that we can exploit their respective advantages to help each other. This motivates us to propose the second contrastive loss between VKPs and TKPs. Similar to Eq. 7, we build $\mathcal{L}_{\mathrm{vt}}$ . We find that the visual-textual alignment improves 1-shot testing, but unintentionally degrades 0-shot performance slightly. We discover the principle behind this is better clustering effects of textual keypoint representations than visual ones. To address this issue, we perform stop gradient on TKPs in $\mathcal{L}_{\mathrm{vt}}$ , thus enforcing the VKPs to align towards TKPs. We also find adding $\mathcal{L}_{\mathrm{vv}}$ does not further improve scores. We conjecture that after aligning VKPs to TKPs via $\mathcal{L}_{\mathrm{vt}}$ , the visual features will also improve discrimination along with textual features as $\mathcal{L}_{\mathrm{tt}}$ is applied. + +Optimization By incorporating the heatmap regression loss $\mathcal{L}_{\mathrm{kp}}$ , intra- and inter-modality contrastive loss $\mathcal{L}_{\mathrm{tt}}$ and $\mathcal{L}_{\mathrm{vt}}$ , we have the overall loss as $\mathcal{L} = \lambda_1\mathcal{L}_{\mathrm{kp}} + \lambda_2\mathcal{L}_{\mathrm{tt}} + \lambda_3\mathcal{L}_{\mathrm{vt}}$ . We set the loss weight $\lambda_{1} = 1$ , $\lambda_{2} = \lambda_{3} = 0.002$ . + +# 3.2 Training with Auxiliary Keypoints and Texts + +Localizing the seen keypoints and texts is relatively easy to handle, but it becomes difficult if the unseen ones are encountered. Thus, we propose to add auxiliary keypoints and texts interpolated from visual and textual domains into training, which significantly improves the ability to detect novel keypoints. + +Visual Interpolation Following [27], we generate the auxiliary keypoints in visual domain via interpolation $\mathcal{I}^{\mathrm{v}}(z;\mathbf{p}_1,\mathbf{p}_2)$ , where $z\in (0,1)$ is the interpolation node and $\mathbf{p}_i\in \mathbb{R}^2$ are the end points that form an interpolation path. An off-the-shelf saliency detector [49] is adopted to filter the auxiliary keypoints outside of the foreground [27]. Via interpolation $\mathcal{I}^{\mathrm{v}}$ in the visual domain, the auxiliary keypoints $\hat{\mathbf{p}}$ can boost the visual diversity of training keypoints. + +Text Interpolation and Selection The $\mathcal{L}^{\nu}$ can assist the model in handling unseen keypoints in visual prompt, but it helps little if prompting unseen texts. + +Consequently, we propose using text interpolation $\mathcal{I}^{\mathrm{t}}(z;t(\mathbf{p}_1),t(\mathbf{p}_2),c)$ to mitigate this issue, where $t(\mathbf{p}_i)$ are the names of end points $\mathbf{p}_i$ and $c$ is the object category. Thanks to great language processing ability of LLM (e.g., GPT3.5 [4]), we propose to ask LLM a good question to reason the auxiliary keypoint texts given base keypoints (Fig. 4(b)), e.g., to infer the keypoint text between "high" and "paw" of a "cat", we can design a vanilla prompt for the LLM as: + +"Please give me three most common body parts/ keypoints at $1 / 2$ between thigh and paw of a cat. Please answer in concise words." + +Apparently, we expect that the answers returned by LLM should include "knee" or "elbow". In practice, the keypoint types are usually more complex thus the reasoning is challenging. To enhance reasoning, we improve the above vanilla prompt into a more advanced one by using chain of thought (CoT) [48], whose key idea is to firstly provide an example with analysis and then ask the question. We provide the detailed prompts and examples in §A of Suppl. + +Considering the LLM suffers from some randomness and might produce erroneous answers, we let LLM give three most possible answers for each interpolation path and repeat this step $R$ times, yielding a text pool with $3R$ texts, i.e., $\{\hat{t}_i\}_{i=1}^{3R}$ . As such, the GT text has higher chance to be included. Afterwards, we need to select one interpolated text from the pool to form a pair with a visually interpolated keypoint. To this end, a simple yet effective approach is to sample from top-1 results of all repetitions. However, it does not use the visual features for selection. An interesting approach is to use the correlation between visual and text features for selection as follows: + +$$ +j = \arg \max _ {i} \cos (\hat {\boldsymbol {\Phi}}, \hat {\mathbf {t}} _ {i}), \quad \hat {t} ^ {*} = \hat {t} _ {j}, \tag {8} +$$ + +where $\hat{\Phi}$ is the auxiliary keypoint feature of $\hat{\mathbf{p}}$ , $\hat{\mathbf{t}}_i$ is text feature of $\hat{t}_i$ , and $\hat{t}^*$ is the selected text. However, we found that sometimes it leads to performance drop. We suspect it is because CLIP's keypoint-text features are not fine-grained aligned enough on those regions beyond annotations. Opposed to picking, we observe such a correlation could be used to reject low-quality text if keypoint-text similarity $\cos (\hat{\Phi},\hat{\mathbf{t}}_i)$ is low. Consequently, we propose a more advanced sampling with false text control (FTC). Namely, we sample the text within the range top- $\eta$ but the text will be rejected if the similarity $\cos (\hat{\Phi},\hat{\mathbf{t}}_i)$ is below the threshold $\alpha$ . In this way, we can form auxiliary keypoint-text pairs with higher quality to enhance model training. We apply $\mathcal{L}^{\mathrm{v}}$ to both support and query images to generate support and query auxiliary keypoints $\hat{\mathbf{p}}^{\mathrm{s}}$ and $\hat{\mathbf{p}}^{\mathrm{q}}$ , respectively. As in Eq. 5, $\hat{\mathbf{p}}^{\mathrm{q}}$ encodes GT heatmap supervising the heatmaps induced by $\hat{\mathbf{p}}^{\mathrm{s}}$ and $\hat{t}^*$ . + +# 3.3 LLM is a Good Language Parser + +To handle diverse text prompts in keypoint detection, unlike current fashion of utilizing LLMs for text generation, we use LLM for text decomposition. For example, if our model receives a diverse text prompt like "Can you localize the left eye and nose of cat?", we firstly leverage an LLM (e.g., GPT3.5 [4]) to perform text parsing by providing it with a simple yet effective prompt as follows: + +Please extract the animal and keypoint keywords from the below text: "Can you localize the left eye and nose of cat?" + +Then, LLM returns the parsed texts regarding ⟨keypoint⟩ and ⟨object⟩, i.e., the "left eye", "nose", and "cat" in the example. We directly utilize parsed results to synthesize simple text prompts to instruct our model and predict all the keypoints corresponding to the text. The problem of diverse text prompting becomes simple text prompting, thus opening the language diversity for keypoint detection. + +To fairly evaluate our approach, we manually collect 100 diverse text templates which cover most scenarios for questions ( $\S \mathbf{B}$ of Suppl.). Take one as an example: "Where is the for . Afterwards, we develop four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42]. Specifically, we accompany each object instance with a diverse text prompt synthesized by randomly sampling one template and one to $N$ valid keypoints. Compared to traditional text prompts which only have $N$ variants, our constructed text prompt sets are more diverse, whose space could be as large as $100 \cdot (C_N^1 + C_N^2 + \dots + C_N^N) = 100 \cdot 2^N$ variants. For example, as CUB has 11 keypoints, its space is larger than $10^5$ . We randomly sample 1000 diverse texts from each dataset and evaluate the LLM's parsing performance. Table 5b reports that GPT3.5 has over $96\%$ accuracy in parsing keypoints from text, which shows that LLM can be a good language parser. + +# 4 Experiments + +# 4.1 Experimental Settings + +Datasets and Splits. 1) Animal pose dataset [5] has five mammal species cat, dog, cow, horse, and sheep. Each animal species is alternately chosen as unseen species for testing while the remaining four as seen species for training, which yields five subproblems; 2) AwA [2] has 35 diverse animal species with 10064 images. For AwA, 25 species are for training and 10 for testing; 3) CUB [45] consists of 200 bird species with 15 keypoint annotations. We use 100 species for training, 50 for validation, and 50 for testing; 4) NABird [42] is larger than CUB and has 555 categories. The species splits are 333, 111, and 111 for training, validation, and testing. To test model generalization, we follow [27] to split keypoints into base and novel sets, and report the performance of both base and novel keypoint detection on unseen species. + +Metric. The percentage of correct keypoints (PCK) is used. A predicted keypoint is correct if its distance to GT $d \leq \rho \cdot \max(w_{\mathrm{bbx}}, h_{\mathrm{bbx}})$ , where $w_{\mathrm{bbx}}$ and $h_{\mathrm{bbx}}$ are the edges of object bounding box. Following [27], we set $\rho$ to 0.1. + +Implementation Details. The input image size for all models is $384 \times 384$ . We freeze the CLIP text encoder but finetune last two layers of image encoder. The temperature $\tau$ used in contrastive loss is 0.05. Same to [27], the visual auxiliary keypoints are generated on pre-defined interpolation path, and we set $z = 0.5$ . For the text interpolation, to reduce randomness, we use GPT3.5 to reason $R = 3$ times for all datasets except CUB with $R = 10$ . Our model is trained with 40k episodes. We report results using 1000 test episodes. + +Table 1: Main results on the Animal pose dataset. Each row shows the results tested on the model trained in one setting. Each PCK score is the average of five subproblems. + +
#Training settings1-shot testing0-shot testing1-shot w/ text
KpAux. kpTextAux. textNovelBaseNovelBaseNovelBase
1✓21.3650.841.261.9316.1632.81
2✓✓47.5449.452.002.1039.2434.11
3✓16.1835.4625.6061.6427.0261.21
4✓✓15.8731.5763.1465.3163.3065.52
5✓✓21.4652.1522.4261.0723.8860.49
6✓✓✓✓50.3254.3963.3765.5963.1964.93
+ +Compared Methods. For few-shot keypoint detection (FSKD), as in previous works, we compare the few-shot learning models such as ProtoNet [36], RelationNet [38], and LwoF [15], and FSKD-dedicated works $FSKD-R/-D$ [27]. For zero-shot keypoint detection (ZSKD), we adopt the source code of CLAMP [56] and compare it in our experiments for fairness. Moreover, we build a Baseline sharing same backbone with our model but we do not use the specifically proposed auxiliary texts, contrastive loss, etc. We denote our method as OpenKD. + +# 4.2 Main Results + +Firstly, we explore single-modal training and the impacts of whether or not to add the auxiliary keypoints/texts (1st-4th row, Table 1). When only using main keypoints (1st row) or main texts (3rd row), one can observe that the models perform well on base keypoints in 1-shot $(50.84\%)$ or 0-shot $(61.64\%)$ , but fail to detect novel keypoints. However, if the model is trained by adding the auxiliary keypoints (2nd row) or auxiliary texts (4th row), one can observe the remarkable improvements on novel keypoint detection, e.g., $47.54\%$ vs. $21.36\%$ in 1-shot testing for the model trained by visual prompts (2nd row vs. 1st row); and also $63.14\%$ vs. $25.60\%$ in 0-shot testing for the model trained by text prompts (4th row vs. 3rd row). The significant gains confirm the benefits of visual interpolation [27] and also highlight the effectiveness of our proposed textual interpolation. Secondly, we investigate multimodal training (5th-6th row, Table 1). As one can see, our model performs well under multimodal training (6th row), and greatly outperforms the model trained by multimodal prompts without auxiliary keypoints and texts (5th row). The higher performance is due to the fact that auxiliary keypoints and texts could boost the visual and textual semantic diversity, thus enabling novel keypoint detection under 1-shot and 0-shot setting. Thirdly, we observe that 0-shot testing on base keypoints is much higher than 1-shot, which shows that the texts can be an excellent representation to guide keypoint detection. Lastly, we examine multimodal testing, i.e., performing 1-shot with text testing, our model obtains $63.19\%$ and $64.93\%$ on novel and base keypoints, which highlights that our model strongly combines the advantages of both modalities, mitigating the weakness of individual modality. + +Table 2: Results on 1-shot keypoint detection for unseen species. The PCK scores on novel and base keypoint detection are reported. $+ \mathbf{T}$ means adding texts during testing. + +
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelProtoNet19.6816.1814.3912.0515.0615.4729.5751.3236.65
RelationNet22.1517.1915.4713.5816.5516.9920.9156.5934.02
LwoF22.4719.3916.8216.4016.9418.4028.5454.7534.19
FSKD-R46.0540.6637.5538.0931.5038.7751.8177.9054.01
FSKD-D52.3647.9444.0742.7736.6044.7564.7677.8956.04
OpenKD60.3653.5847.5949.0141.0550.3266.7178.3953.35
OpenKD+T69.2666.8162.4063.2154.2763.1979.0273.2953.40
BaseProtoNet45.8039.8334.8835.8032.3337.7357.1780.3673.18
RelationNet51.0345.8539.8641.9737.1943.1857.3179.4078.85
LwoF50.0544.6443.4743.3537.8443.8763.8781.9681.39
FSKD-R57.1251.1247.8349.7143.7149.9065.2687.9487.84
FSKD-D56.3851.2948.2449.7743.9549.9366.3987.7186.99
OpenKD63.6155.4351.1853.8747.8654.3974.2287.4585.11
OpenKD+T70.4765.0962.8466.4659.8164.9384.5091.8191.23
+ +Table 3: Results on the 0-shot keypoint detection for unseen species. $\mathrm{CLAMP}^{\dagger}$ is the variant trained by adding our interpolated auxiliary texts. + +
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelBaseline25.6424.2925.2918.1425.5123.7728.9534.9430.00
CLAMP20.9024.0627.0716.8620.7321.9238.9637.0918.18
\( CLAMP^† \)61.7058.0961.4264.8653.1359.8477.6669.3050.81
OpenKD71.0766.4960.3065.5353.4563.3778.6470.1652.29
BaseBaseline58.6956.6256.5860.0051.3556.6573.4587.1271.25
CLAMP60.7154.0462.5862.7357.3059.4784.1690.9779.10
\( CLAMP^† \)60.3056.4962.6960.9057.1659.5183.7690.6583.65
OpenKD71.3466.6064.5367.2658.2365.5984.3291.7290.63
+ +Comparisons on FSKD and ZSKD: We comprehensively conduct the few-shot keypoint detection across four datasets. Table 2 shows our OpenKD model significantly outperforms compared methods in 15 out of 18 tasks in detecting novel or base keypoints. Moreover, if adding the texts during testing, the performance of our model further improves, yielding $63.19\%$ on the Animal pose dataset (63.19% vs. 44.75% of prior-art FSKD-D) and $79.02\%$ on AwA. We also notice the improvements on CUB and NAB are modest, which might be due to the auxiliary texts being relatively harder to reason, thus bringing lesser benefit to few-shot model. For ZSKD (Table 3), again, our model outperforms other + +![](images/dcde36d53e7d806056951f660123c52190278beee66b2b9b4856c02bb25c78e5.jpg) +Fig. 5: Different clustering effects of two modalities on base keypoints in unseen species (a) and statistical feature variance per keypoint (b). + +![](images/55dbc83131d445db9e8286c416cb2ef640096370e80f1f1dad30ea45ca477827.jpg) +(a) + +![](images/46f8128bb1ff7c6c139ae5dac6acf88a1ff74087fb9a1b4a164fb39ba1efe543.jpg) + +![](images/9a9b85bc8269ff46d1daa6dfa3506600b87d635c5837f8e441dbcb02e4e1d8fe.jpg) +(b) + +Table 4: Ablation study. (a) Intra- and inter-modality contrastive loss. $\mathcal{L}_{\mathrm{vt}}^*$ denotes no stop gradient on text features; (b) Text interpolation. Baseline†: w/o using auxiliary texts; CoT: chain of thought; Corr: Eq. 8; FTC: sampling with false text control (ours). + +
AwA1-shot0-shot
NovelBaseNovelBase
1: w/o CL65.5672.1776.7181.67
2: Lt66.0572.3578.2884.00
3: Ltt+L*vt66.7073.9177.1083.85
4: Ltt+Lvt66.7174.2278.6484.32
5: Ltt+Lvt+Lvv66.2774.0878.6083.42
+ +(a) + +
0-shot testingAwACUB
NovelBaseNovelBase
1: Baselined† (▲)34.3183.5336.8988.85
2: ▲+w/o CoT73.8082.7968.0388.93
3: ▲+CoT78.3084.3268.6888.98
4: ▲+CoT+Corr62.1283.1371.8094.27
5: ▲+CoT+FTC78.6484.3270.1691.72
+ +(b) + +methods (in 17 out of 18 tasks). The improvement of $\mathrm{CLAMP}^{\dagger}$ on novel keypoint detection strongly shows the benefits of adding auxiliary texts into training. + +# 4.3 Model and Performance Analysis + +Why 0-shot can outperform 1-shot? One may observe the 0-shot testing scores on base keypoints are always higher than 1-shot testing. In this case, the training texts are accurate (cf. auxiliary text is noisy) and have high semantic similarity to testing texts. We dig more to visualize the corresponding text and keypoint features. Fig. 5 shows that textual keypoint features enjoy better clustering effects and lower variance than visual ones, which suggests textual representations of base keypoints are better thus 0-shot yields higher scores. + +Contrastive Learning (CL). After applying $\mathcal{L}_{\mathrm{tt}}$ , we observe the 0-shot performance significantly improves, e.g. $84.00\%$ vs. $81.67\%$ (2nd row, Table 4a), as CL improves the discrimination between text features. However, when adding visual-textual alignment (3rd row), 1-shot testing scores boost but slightly decreasing 0-shot scores. We suspect it is due to the imbalance of representation quality between multimodal features. After adding stop gradient, it mitigates negative impacts from "weak" modality, thus achieving higher scores (4th row). We also observe $\mathcal{L}_{\mathrm{vv}}$ does not further improve scores in our losses (5th row). + +Text Interpolation. Table 4b shows that using chain of thought (CoT) significantly improves performance, especially on AwA (3rd row), as CoT improves the + +Table 5: (a) Study on repetition $R$ ; (b) Parsing results with GPT3.5 and Vicuna. + +
RAwACUB
NovelBaseNovelBase
167.3483.9864.2191.40
378.64(+11.3)84.3269.81(+5.60)92.38
576.36(+9.02)83.7369.81(+5.60)92.02
1076.61(+9.27)83.3670.16(+5.95)91.72
+ +(a) + +
DatasetGPT3.5Vicuna
Acc. kpAcc. objAcc. kpAcc. obj
Animal0.970.990.940.98
AwA0.961.000.930.98
CUB0.970.960.941.00
NABird0.970.990.951.00
+ +(b) +Table 6: Diverse text prompting evaluation. Results are the average of three runs. + +
MethodAnimal PoseAwACUBNABAvg Drop
NovelBaseNovelBaseNovelBaseNovelBaseNovelBase
OpenKD63.3765.5978.6484.3270.1691.7252.2990.63--
No parsing13.8915.1813.4421.7022.0917.4814.5517.6250.12↓65.07↓
Vicuna57.5459.8473.9778.1867.8987.3950.4586.593.65↓5.06↓
GPT3.561.4562.8276.2380.0769.3990.6552.2189.141.29↓2.39↓
+ +quality of reasoned texts. After applying FTC, the scores strike a high balance in all datasets (5-th row), as our FTC rejects low-quality texts and further reduces noise after including visual features into decision. We also explore the repetition $R$ for text reasoning. Table 5a shows the importance of setting $R > 1$ . + +Diverse Text Prompting. To evaluate our model under diverse text prompting, we randomly sample 1000 diverse texts from each dataset to conduct zero-shot testing. As shown in Table 6, after coupling with GPT3.5, our model retains the high performance even under diverse texts (with less than $3\%$ drop). The strong results compared to no parsing show that leveraging LLM as a parser to handle language diversity is a possible solution. Moreover, we also found coupling with GPT3.5 has higher performance compared to Vicuna, as GPT3.5 enjoys a stronger keypoint text parsing ability, e.g., $96\%$ vs. $93\%$ in AwA (Table 5b). + +# 5 Conclusion + +We propose to open the prompt diversity from the aspects of modality, semantics and language, to enable a more general zero- and few-shot keypoint detection. To this end, we build a versatile OpenKD model which supports both visual and textual prompting. Moreover, to bridge the semantics gap between seen and unseen texts, we propose a novel text interpolation and the selection strategy with false text control, which significantly improves zero-shot novel keypoint detection. We also discover that LLM is a good language parser. After coupling with LLM, our model can handle diverse texts well. We hope the proposed model, text interpolation and the parsing approach will provide useful insights for versatile keypoint detection, thus we highly recommend it to vision community. + +# Acknowledgment + +Changsheng Lu is supported by Australian Government Research Training Program (AGRTP) International Scholarship. Piotr Koniusz is supported by CSIRO's Science Digital. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +2. Banik, P., Li, L., Dong, X.: A novel dataset for keypoint detection of quadruped animals from images. arXiv preprint arXiv:2108.13958 (2021) +3. Bohdal, O., Tian, Y., Zong, Y., Chavhan, R., Li, D., Gouk, H., Guo, L., Hospedales, T.: Meta omnium: A benchmark for general-purpose learning-to-learn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7693-7703 (2023) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +5. Cao, J., Tang, H., Fang, H.S., Shen, X., Lu, C., Tai, Y.W.: Cross-domain adaptation for animal pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9498-9507 (2019) +6. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: realtime multiperson 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence 43(1), 172-186 (2019) +7. Carreira, J., Agrawal, P., Fragkiadaki, K., Malik, J.: Human pose estimation with iterative error feedback. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4733-4742 (2016) +8. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020) +9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) +10. Derpanis, K.G.: The harris corner detector. York University pp. 2-3 (2004) +11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021) +12. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017) +13. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: International Conference on Machine Learning. pp. 1126-1135. PMLR (2017) + +14. Ge, Y., Zhang, R., Luo, P.: Metacloth: Learning unseen tasks of dense fashion landmark detection from a few samples. IEEE Transactions on Image Processing 31, 1120-1133 (2021) +15. Gidaris, S., Komodakis, N.: Dynamic few-shot visual learning without forgetting. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4367-4375 (2018) +16. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) +17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +18. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1546-1555 (2018) +19. Jiao, B., Liu, L., Gao, L., Wu, R., Lin, G., Wang, P., Zhang, Y.: Toward re-identifying any animal. Advances in Neural Information Processing Systems 36 (2024) +20. Koch, G., Zemel, R., Salakhutdinov, R., et al.: Siamese neural networks for one-shot image recognition. In: ICML deep learning workshop. vol. 2. Lille (2015) +21. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015) +22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +23. Li, S., Gunel, S., Ostrek, M., Ramdya, P., Fua, P., Rhodin, H.: Deformation-aware unpaired image translation for pose estimation on laboratory animals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13158-13168 (2020) +24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +25. Lowe, G.: Sift-the scale invariant feature transform. Int. J 2, 91-110 (2004) +26. Lu, C., Gu, C., Wu, K., Xia, S., Wang, H., Guan, X.: Deep transfer neural network using hybrid representations of domain discrepancy. Neurocomputing 409, 60-73 (2020) +27. Lu, C., Koniusz, P.: Few-shot keypoint detection with uncertainty learning for unseen species. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19416-19426 (2022) +28. Lu, C., Koniusz, P.: Detect any keypoints: An efficient light-weight few-shot keypoint detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3882-3890 (2024) +29. Lu, C., Zhu, H., Koniusz, P.: From saliency to dino: Saliency-guided vision transformer for few-shot keypoint detection. arXiv preprint arXiv:2304.03140 (2023) +30. Mathis, A., Mamidanna, P., Cury, K.M., Abe, T., Murthy, V.N., Mathis, M.W., Bethge, M.: Deeplabcut: markerless pose estimation of user-defined body parts with deep learning. Nature neuroscience 21(9), 1281-1289 (2018) +31. Moskvyak, O., Maire, F., Dayoub, F., Baktashmotlagh, M.: Semi-supervised keypoint localization. arXiv preprint arXiv:2101.07988 (2021) +32. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. + +In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19413-19423 (2023) +33. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 483-499. Springer (2016) +34. Pereira, T.D., Aldarondo, D.E., Willmore, L., Kislin, M., Wang, S.S.H., Murthy, M., Shaevitz, J.W.: Fast animal pose estimation using deep neural networks. Nature methods 16(1), 117-125 (2019) +35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +36. Snell, J., Swersky, K., Zemel, R.S.: Prototypical networks for few-shot learning. arXiv preprint arXiv:1703.05175 (2017) +37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019) +38. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1199-1208 (2018) +39. Tang, L., Wertheimer, D., Hariharan, B.: Revisiting pose-normalization for fine-grained few-shot recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14352-14361 (2020) +40. Thompson, J.J., Jain, A., LeCun, Y., Bregler, C.: Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems 27 (2014) +41. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014) +42. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 595-604 (2015) +43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +44. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. Advances in neural information processing systems 29, 3630-3638 (2016) +45. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011) +46. Wang, C., Jin, S., Guan, Y., Liu, W., Qian, C., Luo, P., Ouyang, W.: Pseudolabeled auto-curriculum learning for semi-supervised keypoint localization. arXiv preprint arXiv:2201.08613 (2022) +47. Wang, L., Huynh, D.Q., Koniusz, P.: A comparative review of recent kinect-based action recognition algorithms. IEEE Transactions on Image Processing 29, 15-28 (2019) + +48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +49. Wu, Z., Su, L., Huang, Q.: Stacked cross refinement network for edge-aware salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7264-7273 (2019) +50. Xu, L., Jin, S., Zeng, W., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Pose for everything: Towards category-agnostic pose estimation. In: European Conference on Computer Vision. pp. 398-416. Springer (2022) +51. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022) +52. Yang, S., Quan, Z., Nie, M., Yang, W.: Transpose: Keypoint localization via transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11802-11812 (2021) +53. Yao, Q., Quan, Q., Xiao, L., Kevin Zhou, S.: One-shot medical landmark detection. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 177-188. Springer (2021) +54. Zhang, H., Zhang, K., Xu, L., Lai, S., Shao, W., Zheng, N., Luo, P., Qiao, Y.: Language-driven open-vocabulary keypoint detection for animal body and face. arXiv preprint arXiv:2310.05056 (2023) +55. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014) +56. Zhang, X., Wang, W., Chen, Z., Xu, Y., Zhang, J., Tao, D.: Clamp: Prompt-based contrastive learning for connecting language and animal pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23272-23281 (2023) +57. Zhao, S., Gong, M., Zhao, H., Zhang, J., Tao, D.: Deep corner. International Journal of Computer Vision pp. 1-25 (2023) +58. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena (2023) +59. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022) \ No newline at end of file diff --git a/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/images.zip b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..63c37ef83d0c74b3ec9bad81f8b6524d4f3ddb25 --- /dev/null +++ b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:423609bf6acdbc16fe484f597677e212aaee2023a6d0eb0840b4c796dfd324f0 +size 634775 diff --git a/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/layout.json b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b79e2c500b9def3b3d4b39a7216195b2c22b29aa --- /dev/null +++ b/2024/OpenKD_ Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection/layout.json @@ -0,0 +1,11710 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "type": "text", + "content": "OpenKD: Opening Prompt Diversity for Zero-and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 175, + 168, + 440, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 168, + 440, + 181 + ], + "spans": [ + { + "bbox": [ + 175, + 168, + 440, + 181 + ], + "type": "text", + "content": "Changsheng Lu1, Zheyuan Liu1, and Piotr Koniusz*,2,1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 190, + 435, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 190, + 435, + 223 + ], + "spans": [ + { + "bbox": [ + 179, + 190, + 435, + 223 + ], + "type": "text", + "content": "1The Australian National University 2Data61/CSIRO changshengluu@gmail.com, zheyuan.david.liu@outlook.com, piotr.koniusz@data61.csiro.au" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 253, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 253, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 253, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Exploiting foundation models (e.g., CLIP) to build a versatile keypoint detector has gained increasing attention. Most existing models accept either the text prompt (e.g., \"the nose of a cat\"), or the visual prompt (e.g., support image with keypoint annotations), to detect the corresponding keypoints in query image, thereby, exhibiting either zero-shot or few-shot detection ability. However, the research on multimodal prompting is still underexplored, and the prompt diversity in semantics and language is far from opened. For example, how to handle unseen text prompts for novel keypoint detection and the diverse text prompts like \"Can you detect the nose and ears of a cat?\" In this work, we open the prompt diversity in three aspects: modality, semantics (seen vs. unseen), and language, to enable a more general zero- and few-shot keypoint detection (Z-FSKD). We propose a novel OpenKD model which leverages a multimodal prototype set to support both visual and textual prompting. Further, to infer the keypoint location of unseen texts, we add the auxiliary keypoints and texts interpolated in visual and textual domains into training, which improves the spatial reasoning of our model and significantly enhances zero-shot novel keypoint detection. We also find large language model (LLM) is a good parser, which achieves over " + }, + { + "bbox": [ + 160, + 253, + 455, + 517 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 160, + 253, + 455, + 517 + ], + "type": "text", + "content": " accuracy when parsing keypoints from texts. With LLM, OpenKD can handle diverse text prompts. Experimental results show that our method achieves state-of-the-art performance on Z-FSKD and initiates new ways of dealing with unseen text and diverse texts. The source code and data are available at https://github.com/AlanLuSun/OpenKD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 537, + 230, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 230, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 230, + 550 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "Keypoint detection is a fundamental research problem in computer vision and has numerous applications such as pose estimation for humans [6,8,12,33,37, 52] and animals [2,5,23,34], action recognition [30,47], and fine-grained image classification [39,55]. Over the past decade, significant advancements have been made in deep keypoint detection. However, most existing methods are tailored to close-set detection and struggle to predict novel body parts and species with limited data. Thus, it necessitates such a zero- or few-shot keypoint detection:" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 116, + 187, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 145, + 125 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 145, + 125 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 145, + 125 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 152, + 116, + 187, + 205 + ], + "lines": [ + { + "bbox": [ + 152, + 116, + 187, + 205 + ], + "spans": [ + { + "bbox": [ + 152, + 116, + 187, + 205 + ], + "type": "image", + "image_path": "6ad9eb483430f672d537314cbf53fe745ba8da0ec13c63a2ff5371291e28867f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 206, + 168, + 213 + ], + "lines": [ + { + "bbox": [ + 136, + 206, + 168, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 206, + 168, + 213 + ], + "type": "text", + "content": "Visual prompt" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 206, + 203, + 213 + ], + "lines": [ + { + "bbox": [ + 171, + 206, + 203, + 213 + ], + "spans": [ + { + "bbox": [ + 171, + 206, + 203, + 213 + ], + "type": "text", + "content": "Query image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 233, + 116, + 267, + 205 + ], + "blocks": [ + { + "bbox": [ + 212, + 116, + 223, + 125 + ], + "lines": [ + { + "bbox": [ + 212, + 116, + 223, + 125 + ], + "spans": [ + { + "bbox": [ + 212, + 116, + 223, + 125 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 116, + 267, + 205 + ], + "lines": [ + { + "bbox": [ + 233, + 116, + 267, + 205 + ], + "spans": [ + { + "bbox": [ + 233, + 116, + 267, + 205 + ], + "type": "image", + "image_path": "eeddedc453b57159ecbeca957bfdd1fecf49d36147cb73b3c82fe732cb6ba0dc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 206, + 243, + 213 + ], + "lines": [ + { + "bbox": [ + 214, + 206, + 243, + 213 + ], + "spans": [ + { + "bbox": [ + 214, + 206, + 243, + 213 + ], + "type": "text", + "content": "Text prompt" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 206, + 282, + 213 + ], + "lines": [ + { + "bbox": [ + 247, + 206, + 282, + 213 + ], + "spans": [ + { + "bbox": [ + 247, + 206, + 282, + 213 + ], + "type": "text", + "content": "Query image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 337, + 116, + 370, + 205 + ], + "blocks": [ + { + "bbox": [ + 302, + 118, + 312, + 125 + ], + "lines": [ + { + "bbox": [ + 302, + 118, + 312, + 125 + ], + "spans": [ + { + "bbox": [ + 302, + 118, + 312, + 125 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 116, + 370, + 205 + ], + "lines": [ + { + "bbox": [ + 337, + 116, + 370, + 205 + ], + "spans": [ + { + "bbox": [ + 337, + 116, + 370, + 205 + ], + "type": "image", + "image_path": "4b3173369d2239a793bbb82c3b6e78217ec0d99280db7dfb7c9937b1d9e3e06b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 206, + 347, + 213 + ], + "lines": [ + { + "bbox": [ + 302, + 206, + 347, + 213 + ], + "spans": [ + { + "bbox": [ + 302, + 206, + 347, + 213 + ], + "type": "text", + "content": "Visual + Text prompt" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 399, + 147, + 475, + 213 + ], + "blocks": [ + { + "bbox": [ + 394, + 118, + 405, + 125 + ], + "lines": [ + { + "bbox": [ + 394, + 118, + 405, + 125 + ], + "spans": [ + { + "bbox": [ + 394, + 118, + 405, + 125 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 118, + 467, + 144 + ], + "lines": [ + { + "bbox": [ + 414, + 118, + 467, + 144 + ], + "spans": [ + { + "bbox": [ + 414, + 118, + 467, + 144 + ], + "type": "text", + "content": "Results on AwA dataset One shot Zero shot One shot w/ text" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 147, + 475, + 213 + ], + "lines": [ + { + "bbox": [ + 399, + 147, + 475, + 213 + ], + "spans": [ + { + "bbox": [ + 399, + 147, + 475, + 213 + ], + "type": "image", + "image_path": "191f43a2b19a7913c29425d4d0be443d6f81f9c025d80fffe649a8e9b925114f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 269 + ], + "type": "text", + "content": "Fig. 1: Illustration of multimodal prompting for keypoint detection. Our model can successfully detect keypoints given visual prompts formed by support images and keypoints (a), text prompts (b), or both (c). Graph (d) shows our model well combines the advantages of different modalities, mitigating the weakness induced by either modality." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 316 + ], + "type": "text", + "content": "after being trained on a diverse dataset, the model can quickly recognize novel or base keypoints in unseen species given only zero, one or a few labelled samples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 318, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 484 + ], + "type": "text", + "content": "On the one hand, researchers have recently explored few-shot keypoint detection (FSKD) [27-29, 53]. The FSKD inspired by few-shot learning [13, 20, 36, 38, 44] is general and offers higher flexibility of detecting a varying number of keypoints in a query image, given the prompts formed by support image with keypoint annotations. In this regard, the support set is namely visual prompt, which is required when evaluating new classes. On the other hand, in the era of foundation model, the vision-language model (e.g., CLIP [35]) injects new life to detecting keypoints, enabling zero-shot keypoint detection (ZSKD) [54, 56]. Compared to FSKD, ZSKD does not need support image with annotations, instead, using the text prompt to instruct the model to detect the specified keypoints in query image. In this case, the support set is text prompt. Though impressive for the pioneering FSKD and ZSKD models, we identify an important issue limiting the progress: the prompt diversity for current keypoint detection methods is far from opened, especially in the aspects of modality, semantics, and language." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "type": "text", + "content": "Modality diversity. Most existing keypoint detection models cannot support multimodal prompts, e.g., image, text, or both (see Fig. 1). Multimodal prompting is more friendly in real-world interaction, and coherent with the human concept recognition. We not only see the objects, but also describe objects with language. Our work extends existing FSKD and ZSKD, building a more general zero- and few-shot keypoint detection by leveraging a multimodal prototype set and aligning the visual keypoint features towards the textual features. While straightforward, our method enables one to study the advantages of respective modality data and exploit them for better model training and testing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Semantic diversity. Considering keypoints between seen and unseen species, there exist large similarity yet difference in semantics. A strong advantage of text prompt is that the keypoints with the same semantics share high similarity in language across species, which enables excellent ZSKD on base keypoints. However, if the text has different semantics, significant domain gap arises. For instance, \"the eye of a cat\" vs. \"the eye of a dog\" has cosine similarity of 0.93 using CLIP" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 135, + 201, + 202 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 201, + 133 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 201, + 133 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 201, + 133 + ], + "type": "text", + "content": "Can you make out the back-right knee, front-left knee, body-middle-left on cow?" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 135, + 201, + 202 + ], + "lines": [ + { + "bbox": [ + 134, + 135, + 201, + 202 + ], + "spans": [ + { + "bbox": [ + 134, + 135, + 201, + 202 + ], + "type": "image", + "image_path": "edc17f53fcff8c7a4630b46334b70b5498c5751eafb02e71876a3c5942cd8f39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 203, + 172, + 210 + ], + "lines": [ + { + "bbox": [ + 164, + 203, + 172, + 210 + ], + "spans": [ + { + "bbox": [ + 164, + 203, + 172, + 210 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 220, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 264 + ], + "type": "text", + "content": "Fig. 2: Examples of keypoint detection under diverse text prompting. With LLM, our method can deal with diverse texts, showing potential for real-world applications. The circles and crosses refer to predictions and GT, respectively. A keypoint is regarded as a correct detection if falling in the white shadow area that signifies PCK@0.1." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 204, + 135, + 271, + 202 + ], + "blocks": [ + { + "bbox": [ + 205, + 116, + 271, + 133 + ], + "lines": [ + { + "bbox": [ + 205, + 116, + 271, + 133 + ], + "spans": [ + { + "bbox": [ + 205, + 116, + 271, + 133 + ], + "type": "text", + "content": "Can you detect the body-middle-right, nose, left earbase, front-right paw, left eye of grizzly bear?" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 204, + 135, + 271, + 202 + ], + "lines": [ + { + "bbox": [ + 204, + 135, + 271, + 202 + ], + "spans": [ + { + "bbox": [ + 204, + 135, + 271, + 202 + ], + "type": "image", + "image_path": "d4dfedab7274c9b8fb5b3e53d2855bc6555fa4f09036a74dbb7c777aea709bf2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 203, + 242, + 210 + ], + "lines": [ + { + "bbox": [ + 233, + 203, + 242, + 210 + ], + "spans": [ + { + "bbox": [ + 233, + 203, + 242, + 210 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 274, + 135, + 341, + 202 + ], + "blocks": [ + { + "bbox": [ + 280, + 116, + 334, + 133 + ], + "lines": [ + { + "bbox": [ + 280, + 116, + 334, + 133 + ], + "spans": [ + { + "bbox": [ + 280, + 116, + 334, + 133 + ], + "type": "text", + "content": "Please find the throat, belly, breast, back, left leg for the evening grosbeak." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 274, + 135, + 341, + 202 + ], + "lines": [ + { + "bbox": [ + 274, + 135, + 341, + 202 + ], + "spans": [ + { + "bbox": [ + 274, + 135, + 341, + 202 + ], + "type": "image", + "image_path": "a2dfed2b56fc4d8c6f9153bb7fe7486675ef237eac93d87e8fe76088f34c307a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 203, + 312, + 210 + ], + "lines": [ + { + "bbox": [ + 303, + 203, + 312, + 210 + ], + "spans": [ + { + "bbox": [ + 303, + 203, + 312, + 210 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 135, + 410, + 202 + ], + "blocks": [ + { + "bbox": [ + 343, + 118, + 410, + 131 + ], + "lines": [ + { + "bbox": [ + 343, + 118, + 410, + 131 + ], + "spans": [ + { + "bbox": [ + 343, + 118, + 410, + 131 + ], + "type": "text", + "content": "Where is the breast, throat for pine warbler?" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 135, + 410, + 202 + ], + "lines": [ + { + "bbox": [ + 343, + 135, + 410, + 202 + ], + "spans": [ + { + "bbox": [ + 343, + 135, + 410, + 202 + ], + "type": "image", + "image_path": "bb53182c4349e9bc99cd7d84c6f4d5c337d6351eebb881579c0292134217ab41.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 203, + 381, + 210 + ], + "lines": [ + { + "bbox": [ + 372, + 203, + 381, + 210 + ], + "spans": [ + { + "bbox": [ + 372, + 203, + 381, + 210 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 413, + 135, + 480, + 202 + ], + "blocks": [ + { + "bbox": [ + 419, + 118, + 473, + 131 + ], + "lines": [ + { + "bbox": [ + 419, + 118, + 473, + 131 + ], + "spans": [ + { + "bbox": [ + 419, + 118, + 473, + 131 + ], + "type": "text", + "content": "Can you recognize the back, crown, throat, tail, belly?" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 135, + 480, + 202 + ], + "lines": [ + { + "bbox": [ + 413, + 135, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 413, + 135, + 480, + 202 + ], + "type": "image", + "image_path": "4487459394e3567a13976a47814d8e0d171b141de22a36cbfaa6759b6a8988da.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 203, + 451, + 210 + ], + "lines": [ + { + "bbox": [ + 442, + 203, + 451, + 210 + ], + "spans": [ + { + "bbox": [ + 442, + 203, + 451, + 210 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 290, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 399 + ], + "type": "text", + "content": "text embeddings, whereas for \"the eye of a cat\" vs. \"the knee of a dog\" it drops to 0.77. We observe that the keypoint detectors cannot perform well if text is unseen during training. To bridge the gap, we propose to open the semantic diversity by adding auxiliary keypoints and texts into training, where the auxiliary texts are reasoned by LLM given the base keypoints. To ease the reasoning error, we ask LLM a good question using chain of thought [48]. We also propose a novel sampling strategy with false text control to improve the matching between auxiliary keypoints and texts. With them together, our model significantly improves ZSKD on novel keypoints." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 400, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 556 + ], + "type": "text", + "content": "Language diversity. Following zero-shot image classification by CLIP [35], existing ZSKD also constructs simple text prompts based on templates, e.g., \"〈keypoint〉\" or \"the〈keypoint〉 of a〈category〉 in the photo\", where " + }, + { + "bbox": [ + 130, + 400, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 130, + 400, + 482, + 556 + ], + "type": "text", + "content": " is replaced by class names. In real-world interaction, humans tend to question in diverse styles, which results in diverse text prompts. For example, \"Please locate the right-back leg on cat\", \"Can you find the left eye, right ear and nose in image?\", etc. A natural question is how to handle the diverse text prompting? To address this, we propose a simple yet effective approach, whose key idea is to borrow the large language model (LLM) such as GPT3.5 [4] or Vicuna [9, 58] to parse the diverse texts via prompt engineering. After extracting the keypoint type and object category, they will be synthesized into simple text prompts to instruct the detection model. As such, the diverse text prompting can be transformed into simple text prompting, which sheds light on opening language diversity (Fig. 2)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "To summarize, in this paper, we propose an OpenKD model with several intriguing features: 1) supporting both visual and textual prompts, 2) having the potential to handle unseen texts and diverse texts, and 3) maintaining strong generality and performance on ZSKD and FSKD. We report that LLM is capable of being a reasoner for text interpolation, and a good language parser for parsing diverse texts. We also contribute four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42] for fair evaluations. To our best knowledge, we are the first to open semantics and language diversity of text prompts for keypoint detection." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 235 + ], + "type": "text", + "content": "Keypoint Detection has been widely studied ranging from the traditional interest points [10, 25] to deep corner detection [57], semi-supervised [18, 31, 46] and fully-supervised keypoint estimation [6, 8, 12, 33, 37, 40, 51, 52]. There are two major classes of methods for deep keypoint localization: i) direct coordinates based regression [7, 41] and ii) heatmap based regression. In contrast to existing heatmap based models dedicated to specific body parts, e.g., top-down [12, 16, 37] and bottom-up pose estimators [6, 8, 33], our OpenKD model offers more flexible keypoint detection, breaking the limitation of keypoint types to be detected." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 481, + 402 + ], + "type": "text", + "content": "Few-shot Keypoint Detection is more versatile and data-efficient compared to the supervised paradigms. Similar to standard few-shot learning (FSL) [44], FSKD also takes episodes for training and evaluation. Many well-known FSL methods such as ProtoNet [36], RelationNet [38], LwoF [15] and MAML [13] have been extended to the field of keypoint detection [3,14,27] and serve as baselines. Recently, Lu and Koniusz [27] formalize the comprehensive FSKD settings and model the localization uncertainty of keypoints. Further, they propose a lightweight FSKD model [28] and explore FSKD under transductive setting and occlusions [29]. Bohdal et al. [3] propose a dataset-of-datasets, benchmarking the FSL algorithms universal to various vision tasks. FSKD also inspires class-agnostic animal pose estimation [27, 28, 50]. Compared to existing FSKD, our work pushes further by opening the text prompting, which enables more versatile keypoint detection. Moreover, text prompts can complement visual prompts, delivering better performance of detecting base keypoints on unseen species." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "text", + "content": "Zero-shot Keypoint Detection becomes feasible thanks to the great progress of vision-language pre-training (e.g., CLIP [35] and BLIP [22]). Since CLIP was pre-trained on internet-scale image-text pairs that contain more semantic concepts than other datasets, CLIP can provide semantically rich features for various downstream tasks [32, 56, 59], showing strong zero-shot transfer ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 594 + ], + "type": "text", + "content": "Recently, CLAMP [56] adapts CLIP for animal pose estimation in feature-aware and spatial-aware aspects; Approach [54] also employs CLIP as backbone and proposes language-driven keypoint detection. Compared to FSKD which still requires one or more annotations, ZSKD merely demands cheap language descriptions, which make it intriguing for keypoint detection. Nevertheless, most existing ZSKD works cannot support multimodal prompts and their prompt diversity is limited. To address these issues, we propose an OpenKD model which further opens modality, semantics, and language. To handle unseen texts, we propose to generate auxiliary keypoint-text pairs which significantly help reason novel keypoints on unseen species. Moreover, we contribute a simple yet effective language parsing module to handle diverse text prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "Foundation Models include a broad set of models pre-trained over large-scale datasets, including LLMs (e.g., GPT3.5 [4]), VLMs (e.g., CLIP [35]), and Multimodal models (e.g., GPT4 [1]). Transferring the knowledge from these foundation models to specific tasks in a cost-effective way is popular in computer vision [19,24,26,54,56]. Thus, we follow this trend and leverage LLM like GPT3.5 (relatively cheap) to perform reasoning and parsing for keypoints." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 261, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 261, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 261, + 149 + ], + "type": "text", + "content": "3.1 Model Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "content": "Following the general few-shot learning [36, 38, 44], Z-FSKD model is evaluated on episodes, each of which includes a support set and a query set. The query set is comprised of query images, while the support set gives the prompts. If the visual prompts, i.e., " + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "content": " support images with keypoint annotations (" + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "inline_equation", + "content": "K \\geq 1" + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "content": "), are given, then the problem is defined as " + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "content": "-shot detection. If only the text prompt (i.e., the language description) is given, then " + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 130, + 156, + 482, + 263 + ], + "type": "text", + "content": " thus the problem becomes zero-shot detection. The goal of Z-FSKD is to detect the corresponding keypoints in query image given the prompts, whether visual prompt, text prompt, or both. Such an approach allows the model to effectively respond to diverse inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 264, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 323 + ], + "type": "text", + "content": "The overview of our model inference is shown in Fig. 3, which mainly includes four stages: i) image/text feature extraction, ii) feature adaptation, iii) keypoint prototype set building, iv) correlation, decoding and heatamp fusion. The overview of our model training is shown in Fig. 4, which includes the novel approaches for performance improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "text", + "content": "Image/Text Feature Extraction Since the input of our model involves vision and language two different modalities, we wish the features extracted from image and text have smaller modality gap, so that the text can somewhat correlate with image regions of the same semantics. Such a property helps the model find keypoint locations in query image. To this end, we resort to CLIP [35] pre-trained on large-scale image-text pairs. The CLIP includes image and text encoders. The text encoder is generally a transformer [43] while the image encoder can be CNN [21] or ViT [11]. We empirically found that the CLIP image encoder based on RN50 [17, 35], a CNN model, gives high efficiency in consideration of both performance and cost. Thus, we choose it as the default backbone. When extracting the image features, original CLIP image encoder only retains the classification token " + }, + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{class}}" + }, + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "text", + "content": " and discards image tokens " + }, + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "text", + "content": " by attention pooling as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 475, + 480, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 475, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 146, + 475, + 480, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\text {c l a s s}} = \\operatorname {A v g P o o l} (\\mathbf {X}), \\quad \\mathbf {x} _ {\\text {c l a s s}} ^ {\\prime} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {x} _ {\\text {c l a s s}} \\mathbf {W} _ {q}, \\mathbf {X} \\mathbf {W} _ {k}, \\mathbf {X} \\mathbf {W} _ {v}\\right) \\mathbf {W} _ {o}. \\tag {1}", + "image_path": "49f88c92fee02e41aa8974ea5c02bcead236ad7c858806e6945effbf4cfe90a2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 494, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 532 + ], + "type": "text", + "content": "While it is natural for image-text matching, we require the image tokens to recover the spatial locations of keypoints. Thus, we propose to obtain the image tokens via a projection composed by " + }, + { + "bbox": [ + 130, + 494, + 482, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 130, + 494, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 494, + 482, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o" + }, + { + "bbox": [ + 130, + 494, + 482, + 532 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 272, + 538, + 480, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 538, + 480, + 551 + ], + "spans": [ + { + "bbox": [ + 272, + 538, + 480, + 551 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {\\prime} = \\mathbf {X} \\mathbf {W} _ {v} \\mathbf {W} _ {o}. \\tag {2}", + "image_path": "92c7702b3b4a0aadb4ca0249a168afcc5c1e93caab26f0624145654ecf6c4fd3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "text", + "content": "In this way, we can not only handle the channel inconsistency between raw image tokens " + }, + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "text", + "content": " and text features, but also reuse the projection layers " + }, + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o" + }, + { + "bbox": [ + 130, + 558, + 481, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "For brevity, we assume that per input episode contains one query image " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{q}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", one support image " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " annotated keypoints, and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " keypoint texts. With Eq. 2, both the support and query images can be encoded and projected as the support and query feature maps as " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{s}} = \\mathcal{F}_{\\mathrm{v}}(\\mathbf{I}^{\\mathrm{s}})" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{q}} = \\mathcal{F}_{\\mathrm{v}}(\\mathbf{I}^{\\mathrm{q}})" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " in deep feature space " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{l\\times l\\times d}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " via a shared CLIP image encoder " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{v}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Further, with the CLIP text encoder " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " texts are firstly tokenized, then encoded as the text features " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_n\\in \\mathbb{R}^{m\\times d}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "n = 1,2,\\dots ,N" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "; and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is sequence length." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 103 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "type": "image", + "image_path": "e71fb7a801ead380a8c60b5890eccc813b55be1ecd8c053e1b174e17a4e4a28d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "text", + "content": "Fig. 3: The sketch of model inference. Our OpenKD allows testing under visual prompt, text prompt, or both. For clarification, we show the \"both\" case (i.e., 1-shot with text testing). We firstly extract the deep features of texts, support and query images via CLIP, and then adapt both modalities of features via residual refinement. After extracting the visual keypoint prototype (VKP) and textual counterpart, we build the prototype set to perform class-agnostic correlation and heatmap decoding. Finally, we fuse the heatmaps induced by two modalities (i.e., M1 & M2) to obtain predictions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "content": "Feature Adaptation As CLIP provides general yet transferrable features, we adapt it to a new multimodal feature space suitable for keypoint detection. Inspired by residual learning [17], we propose two adaptation nets " + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{v}}" + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "content": " to respectively refine the obtained image and text features in a residual way:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 401, + 481, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 401, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 162, + 401, + 481, + 415 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {\\mathrm {s}} := \\mathbf {X} ^ {\\mathrm {s}} + \\mathcal {A} _ {\\mathrm {v}} \\left(\\mathbf {X} ^ {\\mathrm {s}}\\right), \\quad \\mathbf {X} ^ {\\mathrm {q}} := \\mathbf {X} ^ {\\mathrm {q}} + \\mathcal {A} _ {\\mathrm {v}} \\left(\\mathbf {X} ^ {\\mathrm {q}}\\right), \\quad \\mathbf {t} _ {n} := \\mathbf {t} _ {n} + \\mathcal {A} _ {\\mathrm {t}} \\left(\\mathbf {t} _ {n}\\right). \\tag {3}", + "image_path": "2fe6759924a8cd14b0df81e56fcfa1cacb789071a7c4ed86f54be669c8dd9043.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "text", + "content": "The adaptation net " + }, + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "text", + "content": " is highly scalable and can be based on transformer [43] or bottleneck [17] as refinement blocks. We found that " + }, + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 423, + 482, + 448 + ], + "type": "text", + "content": " can adapt features well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 486 + ], + "type": "text", + "content": "Keypoint Prototype Set Once we obtain the adapted features of support image and text prompts, i.e., " + }, + { + "bbox": [ + 130, + 450, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 450, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_n" + }, + { + "bbox": [ + 130, + 450, + 482, + 486 + ], + "type": "text", + "content": ", we propose to convert them into keypoint prototypes, which are unified in a keypoint prototype set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": "For visual prompt, recall that keypoint labels are provided along with the support image. Thus, one can aggregate the local features for each support keypoint " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_n" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " by using pixelwise weighted summation between feature map " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " and Gaussian heatmap " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{H}(\\mathbf{p}_n;\\sigma)" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": ", yielding the visual keypoint representation (VKR) as " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Phi_n\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " is the standard deviation that controls the Gaussian spread. If " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " support images are given, namely in few-shot case, the VKRs of the same type of keypoints " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Phi_{k,n}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " will be averaged to build the visual keypoint prototype (VKP) " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Psi_{n}^{\\mathrm{v}} = \\frac{1}{K}\\sum_{k}\\Phi_{k,n}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": ". Analogously, one can develop textual keypoint prototype (TKP) " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Psi_{n}^{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " by text features if provided with multiple texts per keypoint. Finally, one can build the keypoint prototype set as " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\mathcal{T}^{\\mathrm{v}}\\cup \\mathcal{T}^{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{v}} = \\{\\Psi_n^{\\mathrm{v}}\\}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " is the VKPs and " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{t}} = \\{\\Psi_n^{\\mathrm{t}}\\}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " is the TKPs. In this way, prompts of different modalities can be summarized as keypoint prototypes in a shared " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " dimensional feature space, which enables our model to flexibly handle various modalities. Moreover, each prototype in set " + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 487, + 482, + 666 + ], + "type": "text", + "content": " can guide the model to induce a keypoint heatmap, establishing the zero- and few-shot keypoint detection." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 297, + 215 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 297, + 215 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 297, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 297, + 215 + ], + "type": "image", + "image_path": "81028be5d25f5ed9afb43f7501320e54413ea1abb266a4d015de338622d577d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 217, + 241, + 225 + ], + "lines": [ + { + "bbox": [ + 187, + 217, + 241, + 225 + ], + "spans": [ + { + "bbox": [ + 187, + 217, + 241, + 225 + ], + "type": "text", + "content": "(a) Model training" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 119, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 299, + 119, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 299, + 119, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 299, + 119, + 481, + 213 + ], + "type": "image", + "image_path": "99d668236a6c35de77c09fab149453a4ae2da4e27dd33cb61065baac7cfa5e11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 217, + 437, + 226 + ], + "lines": [ + { + "bbox": [ + 348, + 217, + 437, + 226 + ], + "spans": [ + { + "bbox": [ + 348, + 217, + 437, + 226 + ], + "type": "text", + "content": "(b) Auxiliary text interpolation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 234, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 289 + ], + "type": "text", + "content": "Fig. 4: Model training and text interpolation. (a) In addition to multi-group heatmap regression, we improve model performance by introducing intra- and inter-modality contrastive learning and the novel auxiliary keypoint and text learning. (b) We exploit LLM for auxiliary texts interpolation and explore incorporating visual keypoint features for text selection in order to mitigate the noisy texts or false texts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": "Correlation, Decoding and Fusion To discover the corresponding keypoints in query image " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{q}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ", each keypoint prototype " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\Psi_{n}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ", either VKP " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\Psi_{n}^{\\mathrm{v}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": " or TKP " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\Psi_{n}^{\\mathrm{t}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ", is required to be correlated with the query feature map " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{q}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ". To this end, we leverage a correlation module " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": " which takes both " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\Psi_{n}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{q}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": " as input to produce the attentive feature maps " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{n} = \\mathcal{C}(\\Psi_{n},\\mathbf{X}^{\\mathrm{q}})" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{n}\\in \\mathbb{R}^{l\\times l\\times d}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "n = 1,2,\\dots ,N" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ". For the correlation module " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ", we explored multiple variants and found that the simple cross correlation already yields high performance, i.e., " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_n = \\mathbf{X}^{\\mathrm{q}}\\odot \\Psi_n" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 312, + 482, + 409 + ], + "type": "text", + "content": " denotes the channel-wise multiplication." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "text", + "content": "Subsequently, a class-agnostic decoder " + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "text", + "content": " is devised to convert each attentive feature map into a keypoint heatmap, i.e., " + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_n = \\mathcal{D}(\\mathbf{A}_n)" + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_n \\in \\mathbb{R}^{l \\times l}" + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "text", + "content": ". For heatmap regression based keypoint localization, the higher resolution of heatmap can greatly reduce the coordinate decoding error. Thus, a lightweight upsampling module " + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "text", + "content": " is adopted to further refine heatmaps, i.e., " + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_n := \\mathcal{U}(\\mathbf{H}_n)" + }, + { + "bbox": [ + 130, + 410, + 482, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "content": "Since each modality of prototypes can induce one group of heatmaps, considering visual and textual modalities, we have two groups of heatmaps " + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{\\mathrm{v}} = \\{\\mathbf{H}_n^{\\mathrm{v}}\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{\\mathrm{t}} = \\{\\mathbf{H}_n^{\\mathrm{t}}\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "content": ". In testing phase, we fuse the upsampled multi-group heatmaps as the final output " + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{H} \\in \\mathbb{R}^{N \\times ul \\times ul}" + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "content": " is upsampling factor):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 263, + 528, + 480, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 528, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 263, + 528, + 480, + 541 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\left(\\mathbf {H} ^ {\\mathrm {v}} + \\mathbf {H} ^ {\\mathrm {t}}\\right) / 2. \\tag {4}", + "image_path": "07ffca3ea041864200b567bce32958226b6a44cc2a1ba9801d74cfa0449a660a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 546, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 571 + ], + "type": "text", + "content": "During training phase, inspired by HigherHRNet [8], we perform group-specific supervision for heatmaps as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 220, + 579, + 480, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 579, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 220, + 579, + 480, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k p}} = \\left(\\| \\mathbf {H} ^ {\\mathrm {v}} - \\mathbf {H} ^ {*} \\| _ {2} ^ {2} + \\| \\mathbf {H} ^ {\\mathrm {t}} - \\mathbf {H} ^ {*} \\| _ {2} ^ {2}\\right) / 2, \\tag {5}", + "image_path": "5c613ca11af6d3c65c274e52061fe64b961fe8c1b69656eb700ba80191482ebc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 602, + 473, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 473, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 473, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 602, + 473, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^*" + }, + { + "bbox": [ + 130, + 602, + 473, + 615 + ], + "type": "text", + "content": " denotes the groundtruth heatmap that encodes the query keypoint." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Improving Model by Contrastive Learning Since the keypoint prototype directly impacts performance, we aim for its representativeness to be as good as possible. The same type of keypoints should have higher invariance across species, e.g., the ear of cat, dog, and cow, while different types are required to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 448, + 102 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": "be sufficiently discriminative such that our model can distinguish the ambiguous texts or fine-grained texts such as \"the left ear\" vs. \"the right ear\". To this end, we propose to introduce a contrative loss over TKPs. If randomly sampling two species " + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "inline_equation", + "content": "(s,s^{\\prime})" + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": " at a time and each species pertaining to an episode, we have pairwise sets of TKPs, i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s^{\\mathrm{t}} = \\{\\Psi_n\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{s^{\\prime}}^{\\mathrm{t}} = \\{\\Psi_{n}^{\\prime}\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 185 + ], + "type": "text", + "content": ", which can form similarity matrix as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 185, + 481, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 185, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 201, + 185, + 481, + 229 + ], + "type": "interline_equation", + "content": "\\mathbf {J} \\left(\\mathcal {T} _ {s} ^ {\\mathrm {t}}, \\mathcal {T} _ {s ^ {\\prime}} ^ {\\mathrm {t}}\\right) = \\left( \\begin{array}{c c} \\cos \\left(\\Psi_ {1}, \\Psi_ {1} ^ {\\prime}\\right) & \\dots \\cos \\left(\\Psi_ {1}, \\Psi_ {N} ^ {\\prime}\\right) \\\\ \\vdots & \\vdots \\\\ \\cos \\left(\\Psi_ {N}, \\Psi_ {1} ^ {\\prime}\\right) & \\dots \\cos \\left(\\Psi_ {N}, \\Psi_ {N} ^ {\\prime}\\right) \\end{array} \\right), \\tag {6}", + "image_path": "f0476bb7d7aafddca2f85c6c9d61d312588e0357dac1ed3219d05548ff10a1ba.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 233, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 233, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 482, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 233, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 131, + 233, + 482, + 258 + ], + "type": "text", + "content": " denotes the cosine similarity. Then, the contrastive loss within textual modality " + }, + { + "bbox": [ + 131, + 233, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tt}}" + }, + { + "bbox": [ + 131, + 233, + 482, + 258 + ], + "type": "text", + "content": " becomes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 150, + 266, + 481, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 266, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 150, + 266, + 481, + 289 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t t}} ^ {s \\rightarrow s ^ {\\prime}} = - \\langle \\mathbb {I}, \\log (\\operatorname {s o f t m a x} (\\mathbf {J} (\\mathcal {T} _ {s} ^ {\\mathrm {t}}, \\mathcal {T} _ {s ^ {\\prime}} ^ {\\mathrm {t}}) / \\tau)) \\rangle , \\quad \\mathcal {L} _ {\\mathrm {t t}} = \\frac {1}{2} (\\mathcal {L} _ {\\mathrm {t t}} ^ {s \\rightarrow s ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {t t}} ^ {s ^ {\\prime} \\rightarrow s}), \\qquad (7)", + "image_path": "b5c337b4565f8ab9c631ceaa70ee297373d89d6cf7594df1775a409f166549bc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": " is the identity matrix and " + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": " denotes inner product. Since CLIP is trained via image-level alignment instead of keypoint-level alignment with text, we wish VKPs and TKPs were aligned so that we can exploit their respective advantages to help each other. This motivates us to propose the second contrastive loss between VKPs and TKPs. Similar to Eq. 7, we build " + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vt}}" + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": ". We find that the visual-textual alignment improves 1-shot testing, but unintentionally degrades 0-shot performance slightly. We discover the principle behind this is better clustering effects of textual keypoint representations than visual ones. To address this issue, we perform stop gradient on TKPs in " + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vt}}" + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": ", thus enforcing the VKPs to align towards TKPs. We also find adding " + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vv}}" + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": " does not further improve scores. We conjecture that after aligning VKPs to TKPs via " + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vt}}" + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": ", the visual features will also improve discrimination along with textual features as " + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tt}}" + }, + { + "bbox": [ + 130, + 296, + 482, + 441 + ], + "type": "text", + "content": " is applied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": "Optimization By incorporating the heatmap regression loss " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp}}" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": ", intra- and inter-modality contrastive loss " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tt}}" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vt}}" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": ", we have the overall loss as " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\lambda_1\\mathcal{L}_{\\mathrm{kp}} + \\lambda_2\\mathcal{L}_{\\mathrm{tt}} + \\lambda_3\\mathcal{L}_{\\mathrm{vt}}" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": ". We set the loss weight " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 1" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = \\lambda_{3} = 0.002" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 496, + 389, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 389, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 389, + 509 + ], + "type": "text", + "content": "3.2 Training with Auxiliary Keypoints and Texts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "type": "text", + "content": "Localizing the seen keypoints and texts is relatively easy to handle, but it becomes difficult if the unseen ones are encountered. Thus, we propose to add auxiliary keypoints and texts interpolated from visual and textual domains into training, which significantly improves the ability to detect novel keypoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": "Visual Interpolation Following [27], we generate the auxiliary keypoints in visual domain via interpolation " + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{\\mathrm{v}}(z;\\mathbf{p}_1,\\mathbf{p}_2)" + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "inline_equation", + "content": "z\\in (0,1)" + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": " is the interpolation node and " + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": " are the end points that form an interpolation path. An off-the-shelf saliency detector [49] is adopted to filter the auxiliary keypoints outside of the foreground [27]. Via interpolation " + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{\\mathrm{v}}" + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": " in the visual domain, the auxiliary keypoints " + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 567, + 482, + 639 + ], + "type": "text", + "content": " can boost the visual diversity of training keypoints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Text Interpolation and Selection The " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\nu}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " can assist the model in handling unseen keypoints in visual prompt, but it helps little if prompting unseen texts." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "Consequently, we propose using text interpolation " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{\\mathrm{t}}(z;t(\\mathbf{p}_1),t(\\mathbf{p}_2),c)" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " to mitigate this issue, where " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "t(\\mathbf{p}_i)" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " are the names of end points " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " is the object category. Thanks to great language processing ability of LLM (e.g., GPT3.5 [4]), we propose to ask LLM a good question to reason the auxiliary keypoint texts given base keypoints (Fig. 4(b)), e.g., to infer the keypoint text between \"high\" and \"paw\" of a \"cat\", we can design a vanilla prompt for the LLM as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 209 + ], + "type": "text", + "content": "\"Please give me three most common body parts/ keypoints at " + }, + { + "bbox": [ + 130, + 188, + 479, + 209 + ], + "type": "inline_equation", + "content": "1 / 2" + }, + { + "bbox": [ + 130, + 188, + 479, + 209 + ], + "type": "text", + "content": " between thigh and paw of a cat. Please answer in concise words.\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 281 + ], + "type": "text", + "content": "Apparently, we expect that the answers returned by LLM should include \"knee\" or \"elbow\". In practice, the keypoint types are usually more complex thus the reasoning is challenging. To enhance reasoning, we improve the above vanilla prompt into a more advanced one by using chain of thought (CoT) [48], whose key idea is to firstly provide an example with analysis and then ask the question. We provide the detailed prompts and examples in §A of Suppl." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "text", + "content": "Considering the LLM suffers from some randomness and might produce erroneous answers, we let LLM give three most possible answers for each interpolation path and repeat this step " + }, + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "text", + "content": " times, yielding a text pool with " + }, + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "inline_equation", + "content": "3R" + }, + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "text", + "content": " texts, i.e., " + }, + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\{\\hat{t}_i\\}_{i=1}^{3R}" + }, + { + "bbox": [ + 130, + 284, + 480, + 390 + ], + "type": "text", + "content": ". As such, the GT text has higher chance to be included. Afterwards, we need to select one interpolated text from the pool to form a pair with a visually interpolated keypoint. To this end, a simple yet effective approach is to sample from top-1 results of all repetitions. However, it does not use the visual features for selection. An interesting approach is to use the correlation between visual and text features for selection as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 399, + 480, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 399, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 233, + 399, + 480, + 415 + ], + "type": "interline_equation", + "content": "j = \\arg \\max _ {i} \\cos (\\hat {\\boldsymbol {\\Phi}}, \\hat {\\mathbf {t}} _ {i}), \\quad \\hat {t} ^ {*} = \\hat {t} _ {j}, \\tag {8}", + "image_path": "5bcdb82cf2f10d17388f78379b58d17cfef9b7f8f010ceb455d769858f481aea.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " is the auxiliary keypoint feature of " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_i" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " is text feature of " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{t}_i" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{t}^*" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " is the selected text. However, we found that sometimes it leads to performance drop. We suspect it is because CLIP's keypoint-text features are not fine-grained aligned enough on those regions beyond annotations. Opposed to picking, we observe such a correlation could be used to reject low-quality text if keypoint-text similarity " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\cos (\\hat{\\Phi},\\hat{\\mathbf{t}}_i)" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " is low. Consequently, we propose a more advanced sampling with false text control (FTC). Namely, we sample the text within the range top- " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " but the text will be rejected if the similarity " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\cos (\\hat{\\Phi},\\hat{\\mathbf{t}}_i)" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " is below the threshold " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": ". In this way, we can form auxiliary keypoint-text pairs with higher quality to enhance model training. We apply " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{v}}" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " to both support and query images to generate support and query auxiliary keypoints " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}^{\\mathrm{q}}" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": ", respectively. As in Eq. 5, " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}^{\\mathrm{q}}" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " encodes GT heatmap supervising the heatmaps induced by " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\hat{t}^*" + }, + { + "bbox": [ + 130, + 423, + 481, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 586, + 326, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 326, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 326, + 598 + ], + "type": "text", + "content": "3.3 LLM is a Good Language Parser" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "To handle diverse text prompts in keypoint detection, unlike current fashion of utilizing LLMs for text generation, we use LLM for text decomposition. For example, if our model receives a diverse text prompt like \"Can you localize the left eye and nose of cat?\", we firstly leverage an LLM (e.g., GPT3.5 [4]) to perform text parsing by providing it with a simple yet effective prompt as follows:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 103 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "Please extract the animal and keypoint keywords from the below text: \"Can you localize the left eye and nose of cat?\"" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 198 + ], + "type": "text", + "content": "Then, LLM returns the parsed texts regarding ⟨keypoint⟩ and ⟨object⟩, i.e., the \"left eye\", \"nose\", and \"cat\" in the example. We directly utilize parsed results to synthesize simple text prompts to instruct our model and predict all the keypoints corresponding to the text. The problem of diverse text prompting becomes simple text prompting, thus opening the language diversity for keypoint detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "content": "To fairly evaluate our approach, we manually collect 100 diverse text templates which cover most scenarios for questions (" + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "inline_equation", + "content": "\\S \\mathbf{B}" + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "content": " of Suppl.). Take one as an example: \"Where is the for . Afterwards, we develop four diverse text prompt sets for the popular Animal pose dataset [5], AwA [2], CUB [45], and NABird [42]. Specifically, we accompany each object instance with a diverse text prompt synthesized by randomly sampling one template and one to " + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "content": " valid keypoints. Compared to traditional text prompts which only have " + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "content": " variants, our constructed text prompt sets are more diverse, whose space could be as large as " + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "inline_equation", + "content": "100 \\cdot (C_N^1 + C_N^2 + \\dots + C_N^N) = 100 \\cdot 2^N" + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "content": " variants. For example, as CUB has 11 keypoints, its space is larger than " + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "inline_equation", + "content": "10^5" + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "content": ". We randomly sample 1000 diverse texts from each dataset and evaluate the LLM's parsing performance. Table 5b reports that GPT3.5 has over " + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 130, + 200, + 480, + 354 + ], + "type": "text", + "content": " accuracy in parsing keypoints from text, which shows that LLM can be a good language parser." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 372, + 230, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 230, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 230, + 385 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 396, + 272, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 272, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 272, + 407 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 544 + ], + "type": "text", + "content": "Datasets and Splits. 1) Animal pose dataset [5] has five mammal species cat, dog, cow, horse, and sheep. Each animal species is alternately chosen as unseen species for testing while the remaining four as seen species for training, which yields five subproblems; 2) AwA [2] has 35 diverse animal species with 10064 images. For AwA, 25 species are for training and 10 for testing; 3) CUB [45] consists of 200 bird species with 15 keypoint annotations. We use 100 species for training, 50 for validation, and 50 for testing; 4) NABird [42] is larger than CUB and has 555 categories. The species splits are 333, 111, and 111 for training, validation, and testing. To test model generalization, we follow [27] to split keypoints into base and novel sets, and report the performance of both base and novel keypoint detection on unseen species." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "text", + "content": "Metric. The percentage of correct keypoints (PCK) is used. A predicted keypoint is correct if its distance to GT " + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "inline_equation", + "content": "d \\leq \\rho \\cdot \\max(w_{\\mathrm{bbx}}, h_{\\mathrm{bbx}})" + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{bbx}}" + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{bbx}}" + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "text", + "content": " are the edges of object bounding box. Following [27], we set " + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 545, + 479, + 580 + ], + "type": "text", + "content": " to 0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "content": "Implementation Details. The input image size for all models is " + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "content": ". We freeze the CLIP text encoder but finetune last two layers of image encoder. The temperature " + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "content": " used in contrastive loss is 0.05. Same to [27], the visual auxiliary keypoints are generated on pre-defined interpolation path, and we set " + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "inline_equation", + "content": "z = 0.5" + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "content": ". For the text interpolation, to reduce randomness, we use GPT3.5 to reason " + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "inline_equation", + "content": "R = 3" + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "content": " times for all datasets except CUB with " + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "inline_equation", + "content": "R = 10" + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "content": ". Our model is trained with 40k episodes. We report results using 1000 test episodes." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 141, + 479, + 253 + ], + "blocks": [ + { + "bbox": [ + 131, + 117, + 479, + 138 + ], + "lines": [ + { + "bbox": [ + 131, + 117, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 131, + 117, + 479, + 138 + ], + "type": "text", + "content": "Table 1: Main results on the Animal pose dataset. Each row shows the results tested on the model trained in one setting. Each PCK score is the average of five subproblems." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 141, + 479, + 253 + ], + "lines": [ + { + "bbox": [ + 135, + 141, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 135, + 141, + 479, + 253 + ], + "type": "table", + "html": "
#Training settings1-shot testing0-shot testing1-shot w/ text
KpAux. kpTextAux. textNovelBaseNovelBaseNovelBase
1✓21.3650.841.261.9316.1632.81
2✓✓47.5449.452.002.1039.2434.11
3✓16.1835.4625.6061.6427.0261.21
4✓✓15.8731.5763.1465.3163.3065.52
5✓✓21.4652.1522.4261.0723.8860.49
6✓✓✓✓50.3254.3963.3765.5963.1964.93
", + "image_path": "c28d6d353b412283ce4bb914ec76643db4e4d35a68b5e0f6299f3da800681cd0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 273, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 480, + 357 + ], + "type": "text", + "content": "Compared Methods. For few-shot keypoint detection (FSKD), as in previous works, we compare the few-shot learning models such as ProtoNet [36], RelationNet [38], and LwoF [15], and FSKD-dedicated works " + }, + { + "bbox": [ + 130, + 273, + 480, + 357 + ], + "type": "inline_equation", + "content": "FSKD-R/-D" + }, + { + "bbox": [ + 130, + 273, + 480, + 357 + ], + "type": "text", + "content": " [27]. For zero-shot keypoint detection (ZSKD), we adopt the source code of CLAMP [56] and compare it in our experiments for fairness. Moreover, we build a Baseline sharing same backbone with our model but we do not use the specifically proposed auxiliary texts, contrastive loss, etc. We denote our method as OpenKD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 372, + 228, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 228, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 228, + 383 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": "Firstly, we explore single-modal training and the impacts of whether or not to add the auxiliary keypoints/texts (1st-4th row, Table 1). When only using main keypoints (1st row) or main texts (3rd row), one can observe that the models perform well on base keypoints in 1-shot " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "(50.84\\%)" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": " or 0-shot " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "(61.64\\%)" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": ", but fail to detect novel keypoints. However, if the model is trained by adding the auxiliary keypoints (2nd row) or auxiliary texts (4th row), one can observe the remarkable improvements on novel keypoint detection, e.g., " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "47.54\\%" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "21.36\\%" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": " in 1-shot testing for the model trained by visual prompts (2nd row vs. 1st row); and also " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "63.14\\%" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "25.60\\%" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": " in 0-shot testing for the model trained by text prompts (4th row vs. 3rd row). The significant gains confirm the benefits of visual interpolation [27] and also highlight the effectiveness of our proposed textual interpolation. Secondly, we investigate multimodal training (5th-6th row, Table 1). As one can see, our model performs well under multimodal training (6th row), and greatly outperforms the model trained by multimodal prompts without auxiliary keypoints and texts (5th row). The higher performance is due to the fact that auxiliary keypoints and texts could boost the visual and textual semantic diversity, thus enabling novel keypoint detection under 1-shot and 0-shot setting. Thirdly, we observe that 0-shot testing on base keypoints is much higher than 1-shot, which shows that the texts can be an excellent representation to guide keypoint detection. Lastly, we examine multimodal testing, i.e., performing 1-shot with text testing, our model obtains " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "63.19\\%" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "inline_equation", + "content": "64.93\\%" + }, + { + "bbox": [ + 130, + 390, + 481, + 666 + ], + "type": "text", + "content": " on novel and base keypoints, which highlights that our model strongly combines the advantages of both modalities, mitigating the weakness of individual modality." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 141, + 478, + 340 + ], + "blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 139 + ], + "lines": [ + { + "bbox": [ + 132, + 117, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 481, + 139 + ], + "type": "text", + "content": "Table 2: Results on 1-shot keypoint detection for unseen species. The PCK scores on novel and base keypoint detection are reported. " + }, + { + "bbox": [ + 132, + 117, + 481, + 139 + ], + "type": "inline_equation", + "content": "+ \\mathbf{T}" + }, + { + "bbox": [ + 132, + 117, + 481, + 139 + ], + "type": "text", + "content": " means adding texts during testing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 141, + 478, + 340 + ], + "lines": [ + { + "bbox": [ + 136, + 141, + 478, + 340 + ], + "spans": [ + { + "bbox": [ + 136, + 141, + 478, + 340 + ], + "type": "table", + "html": "
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelProtoNet19.6816.1814.3912.0515.0615.4729.5751.3236.65
RelationNet22.1517.1915.4713.5816.5516.9920.9156.5934.02
LwoF22.4719.3916.8216.4016.9418.4028.5454.7534.19
FSKD-R46.0540.6637.5538.0931.5038.7751.8177.9054.01
FSKD-D52.3647.9444.0742.7736.6044.7564.7677.8956.04
OpenKD60.3653.5847.5949.0141.0550.3266.7178.3953.35
OpenKD+T69.2666.8162.4063.2154.2763.1979.0273.2953.40
BaseProtoNet45.8039.8334.8835.8032.3337.7357.1780.3673.18
RelationNet51.0345.8539.8641.9737.1943.1857.3179.4078.85
LwoF50.0544.6443.4743.3537.8443.8763.8781.9681.39
FSKD-R57.1251.1247.8349.7143.7149.9065.2687.9487.84
FSKD-D56.3851.2948.2449.7743.9549.9366.3987.7186.99
OpenKD63.6155.4351.1853.8747.8654.3974.2287.4585.11
OpenKD+T70.4765.0962.8466.4659.8164.9384.5091.8191.23
", + "image_path": "5131b6f9966a66cc7c33467c6f8dcd4cdec40b569a3afe3f4eddc6a304d22994.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 137, + 389, + 475, + 523 + ], + "blocks": [ + { + "bbox": [ + 132, + 364, + 481, + 387 + ], + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 387 + ], + "type": "text", + "content": "Table 3: Results on the 0-shot keypoint detection for unseen species. " + }, + { + "bbox": [ + 132, + 364, + 481, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{CLAMP}^{\\dagger}" + }, + { + "bbox": [ + 132, + 364, + 481, + 387 + ], + "type": "text", + "content": " is the variant trained by adding our interpolated auxiliary texts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 389, + 475, + 523 + ], + "lines": [ + { + "bbox": [ + 137, + 389, + 475, + 523 + ], + "spans": [ + { + "bbox": [ + 137, + 389, + 475, + 523 + ], + "type": "table", + "html": "
SettingModelAnimal Pose DatasetAwACUBNAB
CatDogCowHorseSheepAvg
NovelBaseline25.6424.2925.2918.1425.5123.7728.9534.9430.00
CLAMP20.9024.0627.0716.8620.7321.9238.9637.0918.18
\\( CLAMP^† \\)61.7058.0961.4264.8653.1359.8477.6669.3050.81
OpenKD71.0766.4960.3065.5353.4563.3778.6470.1652.29
BaseBaseline58.6956.6256.5860.0051.3556.6573.4587.1271.25
CLAMP60.7154.0462.5862.7357.3059.4784.1690.9779.10
\\( CLAMP^† \\)60.3056.4962.6960.9057.1659.5183.7690.6583.65
OpenKD71.3466.6064.5367.2658.2365.5984.3291.7290.63
", + "image_path": "1606a2e6a8576e7f72e179c5beddb8a255372473fd82c690e22f03b3f847913f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Comparisons on FSKD and ZSKD: We comprehensively conduct the few-shot keypoint detection across four datasets. Table 2 shows our OpenKD model significantly outperforms compared methods in 15 out of 18 tasks in detecting novel or base keypoints. Moreover, if adding the texts during testing, the performance of our model further improves, yielding " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "63.19\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " on the Animal pose dataset (63.19% vs. 44.75% of prior-art FSKD-D) and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "79.02\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " on AwA. We also notice the improvements on CUB and NAB are modest, which might be due to the auxiliary texts being relatively harder to reason, thus bringing lesser benefit to few-shot model. For ZSKD (Table 3), again, our model outperforms other" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 227, + 202 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 227, + 202 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 227, + 202 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 227, + 202 + ], + "type": "image", + "image_path": "dcde36d53e7d806056951f660123c52190278beee66b2b9b4856c02bb25c78e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 480, + 236 + ], + "type": "text", + "content": "Fig. 5: Different clustering effects of two modalities on base keypoints in unseen species (a) and statistical feature variance per keypoint (b)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 115, + 323, + 202 + ], + "blocks": [ + { + "bbox": [ + 228, + 115, + 323, + 202 + ], + "lines": [ + { + "bbox": [ + 228, + 115, + 323, + 202 + ], + "spans": [ + { + "bbox": [ + 228, + 115, + 323, + 202 + ], + "type": "image", + "image_path": "55dbc83131d445db9e8286c416cb2ef640096370e80f1f1dad30ea45ca477827.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 205, + 254, + 213 + ], + "lines": [ + { + "bbox": [ + 242, + 205, + 254, + 213 + ], + "spans": [ + { + "bbox": [ + 242, + 205, + 254, + 213 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 116, + 362, + 192 + ], + "blocks": [ + { + "bbox": [ + 328, + 116, + 362, + 192 + ], + "lines": [ + { + "bbox": [ + 328, + 116, + 362, + 192 + ], + "spans": [ + { + "bbox": [ + 328, + 116, + 362, + 192 + ], + "type": "image", + "image_path": "46f8128bb1ff7c6c139ae5dac6acf88a1ff74087fb9a1b4a164fb39ba1efe543.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 376, + 115, + 476, + 202 + ], + "blocks": [ + { + "bbox": [ + 376, + 115, + 476, + 202 + ], + "lines": [ + { + "bbox": [ + 376, + 115, + 476, + 202 + ], + "spans": [ + { + "bbox": [ + 376, + 115, + 476, + 202 + ], + "type": "image", + "image_path": "9a9b85bc8269ff46d1daa6dfa3506600b87d635c5837f8e441dbcb02e4e1d8fe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 205, + 434, + 213 + ], + "lines": [ + { + "bbox": [ + 421, + 205, + 434, + 213 + ], + "spans": [ + { + "bbox": [ + 421, + 205, + 434, + 213 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 138, + 284, + 302, + 367 + ], + "blocks": [ + { + "bbox": [ + 130, + 248, + 480, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 282 + ], + "type": "text", + "content": "Table 4: Ablation study. (a) Intra- and inter-modality contrastive loss. " + }, + { + "bbox": [ + 130, + 248, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vt}}^*" + }, + { + "bbox": [ + 130, + 248, + 480, + 282 + ], + "type": "text", + "content": " denotes no stop gradient on text features; (b) Text interpolation. Baseline†: w/o using auxiliary texts; CoT: chain of thought; Corr: Eq. 8; FTC: sampling with false text control (ours)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 284, + 302, + 367 + ], + "lines": [ + { + "bbox": [ + 138, + 284, + 302, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 302, + 367 + ], + "type": "table", + "html": "
AwA1-shot0-shot
NovelBaseNovelBase
1: w/o CL65.5672.1776.7181.67
2: Lt66.0572.3578.2884.00
3: Ltt+L*vt66.7073.9177.1083.85
4: Ltt+Lvt66.7174.2278.6484.32
5: Ltt+Lvt+Lvv66.2774.0878.6083.42
", + "image_path": "ad36af4f4fcc3497856be9290efb448927b4e4c1ee32c7ed0afbf94ec8ae01df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 284, + 476, + 367 + ], + "blocks": [ + { + "bbox": [ + 214, + 368, + 227, + 378 + ], + "lines": [ + { + "bbox": [ + 214, + 368, + 227, + 378 + ], + "spans": [ + { + "bbox": [ + 214, + 368, + 227, + 378 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 284, + 476, + 367 + ], + "lines": [ + { + "bbox": [ + 308, + 284, + 476, + 367 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 476, + 367 + ], + "type": "table", + "html": "
0-shot testingAwACUB
NovelBaseNovelBase
1: Baselined† (▲)34.3183.5336.8988.85
2: ▲+w/o CoT73.8082.7968.0388.93
3: ▲+CoT78.3084.3268.6888.98
4: ▲+CoT+Corr62.1283.1371.8094.27
5: ▲+CoT+FTC78.6484.3270.1691.72
", + "image_path": "cf4458a4c6364c25f16ab4f8133204d1881ab0f5b520580805c03d4b31f80128.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 385, + 368, + 397, + 378 + ], + "lines": [ + { + "bbox": [ + 385, + 368, + 397, + 378 + ], + "spans": [ + { + "bbox": [ + 385, + 368, + 397, + 378 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 402, + 479, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 427 + ], + "type": "text", + "content": "methods (in 17 out of 18 tasks). The improvement of " + }, + { + "bbox": [ + 130, + 402, + 479, + 427 + ], + "type": "inline_equation", + "content": "\\mathrm{CLAMP}^{\\dagger}" + }, + { + "bbox": [ + 130, + 402, + 479, + 427 + ], + "type": "text", + "content": " on novel keypoint detection strongly shows the benefits of adding auxiliary texts into training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 443, + 328, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 328, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 328, + 455 + ], + "type": "text", + "content": "4.3 Model and Performance Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 462, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 545 + ], + "type": "text", + "content": "Why 0-shot can outperform 1-shot? One may observe the 0-shot testing scores on base keypoints are always higher than 1-shot testing. In this case, the training texts are accurate (cf. auxiliary text is noisy) and have high semantic similarity to testing texts. We dig more to visualize the corresponding text and keypoint features. Fig. 5 shows that textual keypoint features enjoy better clustering effects and lower variance than visual ones, which suggests textual representations of base keypoints are better thus 0-shot yields higher scores." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": "Contrastive Learning (CL). After applying " + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tt}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": ", we observe the 0-shot performance significantly improves, e.g. " + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "84.00\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "81.67\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": " (2nd row, Table 4a), as CL improves the discrimination between text features. However, when adding visual-textual alignment (3rd row), 1-shot testing scores boost but slightly decreasing 0-shot scores. We suspect it is due to the imbalance of representation quality between multimodal features. After adding stop gradient, it mitigates negative impacts from \"weak\" modality, thus achieving higher scores (4th row). We also observe " + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vv}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": " does not further improve scores in our losses (5th row)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Text Interpolation. Table 4b shows that using chain of thought (CoT) significantly improves performance, especially on AwA (3rd row), as CoT improves the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 130, + 297, + 213 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 473, + 128 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 473, + 128 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 473, + 128 + ], + "type": "text", + "content": "Table 5: (a) Study on repetition " + }, + { + "bbox": [ + 138, + 116, + 473, + 128 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 138, + 116, + 473, + 128 + ], + "type": "text", + "content": " ; (b) Parsing results with GPT3.5 and Vicuna." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 130, + 297, + 213 + ], + "lines": [ + { + "bbox": [ + 135, + 130, + 297, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 130, + 297, + 213 + ], + "type": "table", + "html": "
RAwACUB
NovelBaseNovelBase
167.3483.9864.2191.40
378.64(+11.3)84.3269.81(+5.60)92.38
576.36(+9.02)83.7369.81(+5.60)92.02
1076.61(+9.27)83.3670.16(+5.95)91.72
", + "image_path": "81fa18534bbd4b4cfa7ae4c25f5282652bfefcdaea24e52d2338eafd3809b0d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 306, + 130, + 482, + 213 + ], + "blocks": [ + { + "bbox": [ + 217, + 215, + 230, + 225 + ], + "lines": [ + { + "bbox": [ + 217, + 215, + 230, + 225 + ], + "spans": [ + { + "bbox": [ + 217, + 215, + 230, + 225 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 130, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 306, + 130, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 130, + 482, + 213 + ], + "type": "table", + "html": "
DatasetGPT3.5Vicuna
Acc. kpAcc. objAcc. kpAcc. obj
Animal0.970.990.940.98
AwA0.961.000.930.98
CUB0.970.960.941.00
NABird0.970.990.951.00
", + "image_path": "afa97f50a108196d6280af317aa8787f75ed5166cbbdf2415600d0d94b7618c9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 135, + 253, + 475, + 338 + ], + "blocks": [ + { + "bbox": [ + 386, + 215, + 398, + 225 + ], + "lines": [ + { + "bbox": [ + 386, + 215, + 398, + 225 + ], + "spans": [ + { + "bbox": [ + 386, + 215, + 398, + 225 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 239, + 474, + 251 + ], + "lines": [ + { + "bbox": [ + 137, + 239, + 474, + 251 + ], + "spans": [ + { + "bbox": [ + 137, + 239, + 474, + 251 + ], + "type": "text", + "content": "Table 6: Diverse text prompting evaluation. Results are the average of three runs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 253, + 475, + 338 + ], + "lines": [ + { + "bbox": [ + 135, + 253, + 475, + 338 + ], + "spans": [ + { + "bbox": [ + 135, + 253, + 475, + 338 + ], + "type": "table", + "html": "
MethodAnimal PoseAwACUBNABAvg Drop
NovelBaseNovelBaseNovelBaseNovelBaseNovelBase
OpenKD63.3765.5978.6484.3270.1691.7252.2990.63--
No parsing13.8915.1813.4421.7022.0917.4814.5517.6250.12↓65.07↓
Vicuna57.5459.8473.9778.1867.8987.3950.4586.593.65↓5.06↓
GPT3.561.4562.8276.2380.0769.3990.6552.2189.141.29↓2.39↓
", + "image_path": "4d4ff6ead49932460c50bfb0dcbfadcb76d3e1f2b1ae31d5074017d710407108.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": "quality of reasoned texts. After applying FTC, the scores strike a high balance in all datasets (5-th row), as our FTC rejects low-quality texts and further reduces noise after including visual features into decision. We also explore the repetition " + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": " for text reasoning. Table 5a shows the importance of setting " + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "inline_equation", + "content": "R > 1" + }, + { + "bbox": [ + 130, + 359, + 480, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "text", + "content": "Diverse Text Prompting. To evaluate our model under diverse text prompting, we randomly sample 1000 diverse texts from each dataset to conduct zero-shot testing. As shown in Table 6, after coupling with GPT3.5, our model retains the high performance even under diverse texts (with less than " + }, + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "text", + "content": " drop). The strong results compared to no parsing show that leveraging LLM as a parser to handle language diversity is a possible solution. Moreover, we also found coupling with GPT3.5 has higher performance compared to Vicuna, as GPT3.5 enjoys a stronger keypoint text parsing ability, e.g., " + }, + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 130, + 407, + 481, + 503 + ], + "type": "text", + "content": " in AwA (Table 5b)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 521, + 219, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 219, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 219, + 533 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We propose to open the prompt diversity from the aspects of modality, semantics and language, to enable a more general zero- and few-shot keypoint detection. To this end, we build a versatile OpenKD model which supports both visual and textual prompting. Moreover, to bridge the semantics gap between seen and unseen texts, we propose a novel text interpolation and the selection strategy with false text control, which significantly improves zero-shot novel keypoint detection. We also discover that LLM is a good language parser. After coupling with LLM, our model can handle diverse texts well. We hope the proposed model, text interpolation and the parsing approach will provide useful insights for versatile keypoint detection, thus we highly recommend it to vision community." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 485, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 485, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 485, + 177 + ], + "type": "text", + "content": "Changsheng Lu is supported by Australian Government Research Training Program (AGRTP) International Scholarship. Piotr Koniusz is supported by CSIRO's Science Digital." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 197, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 223, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 223, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 481, + 255 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 481, + 277 + ], + "type": "text", + "content": "2. Banik, P., Li, L., Dong, X.: A novel dataset for keypoint detection of quadruped animals from images. arXiv preprint arXiv:2108.13958 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 481, + 322 + ], + "type": "text", + "content": "3. Bohdal, O., Tian, Y., Zong, Y., Chavhan, R., Li, D., Gouk, H., Guo, L., Hospedales, T.: Meta omnium: A benchmark for general-purpose learning-to-learn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7693-7703 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "text", + "content": "5. Cao, J., Tang, H., Fang, H.S., Shen, X., Lu, C., Tai, Y.W.: Cross-domain adaptation for animal pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9498-9507 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 390, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 422 + ], + "type": "text", + "content": "6. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: Openpose: realtime multiperson 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence 43(1), 172-186 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "type": "text", + "content": "7. Carreira, J., Agrawal, P., Fragkiadaki, K., Malik, J.: Human pose estimation with iterative error feedback. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4733-4742 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 498 + ], + "type": "text", + "content": "8. Cheng, B., Xiao, B., Wang, J., Shi, H., Huang, T.S., Zhang, L.: Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5386-5395 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 500, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 543 + ], + "type": "text", + "content": "9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 500, + 481, + 543 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 500, + 481, + 543 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 544, + 454, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 454, + 554 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 454, + 554 + ], + "type": "text", + "content": "10. Derpanis, K.G.: The harris corner detector. York University pp. 2-3 (2004)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 598 + ], + "type": "text", + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 600, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 481, + 631 + ], + "type": "text", + "content": "12. Fang, H.S., Xie, S., Tai, Y.W., Lu, C.: Rmpe: Regional multi-person pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 2334-2343 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "13. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: International Conference on Machine Learning. pp. 1126-1135. PMLR (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Ge, Y., Zhang, R., Luo, P.: Metacloth: Learning unseen tasks of dense fashion landmark detection from a few samples. IEEE Transactions on Image Processing 31, 1120-1133 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "15. Gidaris, S., Komodakis, N.: Dynamic few-shot visual learning without forgetting. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4367-4375 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "16. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "17. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "18. Honari, S., Molchanov, P., Tyree, S., Vincent, P., Pal, C., Kautz, J.: Improving landmark localization with semi-supervised learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1546-1555 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "19. Jiao, B., Liu, L., Gao, L., Wu, R., Lin, G., Wang, P., Zhang, Y.: Toward re-identifying any animal. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "text", + "content": "20. Koch, G., Zemel, R., Salakhutdinov, R., et al.: Siamese neural networks for one-shot image recognition. In: ICML deep learning workshop. vol. 2. Lille (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "21. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "text", + "content": "23. Li, S., Gunel, S., Ostrek, M., Ramdya, P., Fua, P., Rhodin, H.: Deformation-aware unpaired image translation for pose estimation on laboratory animals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13158-13168 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "24. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 455, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 455, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 455, + 468 + ], + "type": "text", + "content": "25. Lowe, G.: Sift-the scale invariant feature transform. Int. J 2, 91-110 (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "26. Lu, C., Gu, C., Wu, K., Xia, S., Wang, H., Guan, X.: Deep transfer neural network using hybrid representations of domain discrepancy. Neurocomputing 409, 60-73 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "27. Lu, C., Koniusz, P.: Few-shot keypoint detection with uncertainty learning for unseen species. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19416-19426 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "28. Lu, C., Koniusz, P.: Detect any keypoints: An efficient light-weight few-shot keypoint detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3882-3890 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "29. Lu, C., Zhu, H., Koniusz, P.: From saliency to dino: Saliency-guided vision transformer for few-shot keypoint detection. arXiv preprint arXiv:2304.03140 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "30. Mathis, A., Mamidanna, P., Cury, K.M., Abe, T., Murthy, V.N., Mathis, M.W., Bethge, M.: Deeplabcut: markerless pose estimation of user-defined body parts with deep learning. Nature neuroscience 21(9), 1281-1289 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "31. Moskvyak, O., Maire, F., Dayoub, F., Baktashmotlagh, M.: Semi-supervised keypoint localization. arXiv preprint arXiv:2101.07988 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "text", + "content": "32. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 139 + ], + "type": "text", + "content": "In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19413-19423 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "text", + "content": "33. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for human pose estimation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14. pp. 483-499. Springer (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "34. Pereira, T.D., Aldarondo, D.E., Willmore, L., Kislin, M., Wang, S.S.H., Murthy, M., Shaevitz, J.W.: Fast animal pose estimation using deep neural networks. Nature methods 16(1), 117-125 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "text", + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "type": "text", + "content": "36. Snell, J., Swersky, K., Zemel, R.S.: Prototypical networks for few-shot learning. arXiv preprint arXiv:1703.05175 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "37. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5693-5703 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "38. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1199-1208 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "text", + "content": "39. Tang, L., Wertheimer, D., Hariharan, B.: Revisiting pose-normalization for fine-grained few-shot recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14352-14361 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "text", + "content": "40. Thompson, J.J., Jain, A., LeCun, Y., Bregler, C.: Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "text", + "content": "41. Toshev, A., Szegedy, C.: Deepest: Human pose estimation via deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1653-1660 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 453, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 453, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 453, + 481, + 497 + ], + "type": "text", + "content": "42. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 595-604 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "text", + "content": "44. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. Advances in neural information processing systems 29, 3630-3638 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "45. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "46. Wang, C., Jin, S., Guan, Y., Liu, W., Qian, C., Luo, P., Ouyang, W.: Pseudolabeled auto-curriculum learning for semi-supervised keypoint localization. arXiv preprint arXiv:2201.08613 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "47. Wang, L., Huynh, D.Q., Koniusz, P.: A comparative review of recent kinect-based action recognition algorithms. IEEE Transactions on Image Processing 29, 15-28 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 534 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "49. Wu, Z., Su, L., Huang, Q.: Stacked cross refinement network for edge-aware salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7264-7273 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "text", + "content": "50. Xu, L., Jin, S., Zeng, W., Liu, W., Qian, C., Ouyang, W., Luo, P., Wang, X.: Pose for everything: Towards category-agnostic pose estimation. In: European Conference on Computer Vision. pp. 398-416. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "51. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: Vitpose+: Vision transformer foundation model for generic body pose estimation. arXiv preprint arXiv:2212.04246 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "52. Yang, S., Quan, Z., Nie, M., Yang, W.: Transpose: Keypoint localization via transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11802-11812 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "text", + "content": "53. Yao, Q., Quan, Q., Xiao, L., Kevin Zhou, S.: One-shot medical landmark detection. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24. pp. 177-188. Springer (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "text", + "content": "54. Zhang, H., Zhang, K., Xu, L., Lai, S., Shao, W., Zheng, N., Luo, P., Qiao, Y.: Language-driven open-vocabulary keypoint detection for animal body and face. arXiv preprint arXiv:2310.05056 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 392 + ], + "type": "text", + "content": "55. Zhang, N., Donahue, J., Girshick, R., Darrell, T.: Part-based r-cnns for fine-grained category detection. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 834-849. Springer (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 435 + ], + "type": "text", + "content": "56. Zhang, X., Wang, W., Chen, Z., Xu, Y., Zhang, J., Tao, D.: Clamp: Prompt-based contrastive learning for connecting language and animal pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23272-23281 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "57. Zhao, S., Gong, M., Zhao, H., Zhang, J., Tao, D.: Deep corner. International Journal of Computer Vision pp. 1-25 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "text", + "content": "58. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging llm-as-a-judge with mt-bench and chatbot arena (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 534 + ], + "type": "text", + "content": "59. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "C. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_content_list.json b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..717ae1033538263a85c6b8d2d0e00e6948075080 --- /dev/null +++ b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_content_list.json @@ -0,0 +1,1678 @@ +[ + { + "type": "text", + "text": "OpenPSG: Open-set Panoptic Scene Graph Generation via Large Multimodal Models", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijian Zhou $^{1}$ Zheng Zhu $^{2}$ Holger Caesar $^{3}$ Miaojing Shi $^{4,5\\boxtimes}$", + "bbox": [ + 264, + 210, + 736, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Informatics, King's College London $^{2}$ GigaAI", + "bbox": [ + 289, + 238, + 712, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Intelligent Vehicles Lab, Delft University of Technology", + "bbox": [ + 310, + 253, + 691, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4College of Electronic and Information Engineering, Tongji University \n5Shanghai Institute of Intelligent Science and Technology, Tongji University", + "bbox": [ + 246, + 267, + 754, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Panoptic Scene Graph Generation (PSG) aims to segment objects and recognize their relations, enabling the structured understanding of an image. Previous methods focus on predicting predefined object and relation categories, hence limiting their applications in the open world scenarios. With the rapid development of large multimodal models (LMMs), significant progress has been made in open-set object detection and segmentation, yet open-set relation prediction in PSG remains unexplored. In this paper, we focus on the task of open-set relation prediction integrated with a pretrained open-set panoptic segmentation model to achieve true open-set panoptic scene graph generation (OpenPSG). Our OpenPSG leverages LMMs to achieve open-set relation prediction in an autoregressive manner. We introduce a relation query transformer to efficiently extract visual features of object pairs and estimate the existence of relations between them. The latter can enhance the prediction efficiency by filtering irrelevant pairs. Finally, we design the generation and judgement instructions to perform open-set relation prediction in PSG autoregressively. To our knowledge, we are the first to propose the open-set PSG task. Extensive experiments demonstrate that our method achieves state-of-the-art performance in open-set relation prediction and panoptic scene graph generation. Code is available at https://github.com/franciszzj/OpenPSG.", + "bbox": [ + 259, + 335, + 743, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Panoptic Scene Graph Generation $\\cdot$ Open-set $\\cdot$ Large Multimodal Models", + "bbox": [ + 259, + 633, + 740, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 689, + 375, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panoptic scene graph generation (PSG) [36] aims to segment objects within an image and recognize the relations among them, thereby constructing a panoptic scene graph for a structured understanding of the image. Given its significant potential in applications such as visual question answering [13], image captioning [4, 10], and embodied navigation [29], PSG has attracted considerable attention from researchers ever since its emerging [19, 33, 35, 46-48].", + "bbox": [ + 212, + 723, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxdot$ Corresponding author.", + "bbox": [ + 230, + 824, + 400, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39f4576820ef7e83836b9ae10496dfac3e2e398fbd2c3948dd76d7f41caf0287.jpg", + "image_caption": [ + "Image", + "Fig. 1: The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects. Our method can predict both known (close-set) relations, e.g., (0_person, playing, 1_skateboard), (2_person, looking at, 0_person), and unknown (open-set) relations, e.g., (0_person, pop shove-it, 1_skateboard), (2_person, recording, 0_person)." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 403, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af3d444f853e1b60198c8c5d2a37ab60ad3336fc0bb50681d4adabee86c03426.jpg", + "image_caption": [ + "Panoptic Segmentation" + ], + "image_footnote": [], + "bbox": [ + 408, + 146, + 594, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1f18b16b500d07235addb206562c29206da4aebacedb346cbbceb2399ca831fe.jpg", + "image_caption": [ + "Relation Prediction" + ], + "image_footnote": [], + "bbox": [ + 598, + 146, + 785, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous PSG methods [33, 47, 48] are only capable of predicting closed-set object and relation categories while failing to recognize objects/relations beyond predefined categories. Recently, with the advent of large multimodal models (LMMs) such as CLIP [27], BLIP-2 [18] etc., a significant number of open-set prediction methods for object detection [9, 21, 37, 42, 44] and segmentation [11, 20, 39, 44] are introduced, attributing to LMMs' rich understanding of language and strong connections between vision and language. Nevertheless, open-set prediction of relations has been largely unexplored so far.", + "bbox": [ + 212, + 367, + 787, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to open-set object detection and segmentation, open-set relation prediction is more complex: the model is required to both understand different objects and recognize relations of object pairs based on their interactions; especially, the computation of the latter can be exponentially increased. To bridge the gap, in this paper, we focus on the open-set relation prediction.", + "bbox": [ + 212, + 489, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs [18, 25, 49] have demonstrated exceptional semantic analysis and understanding abilities across various multimodal tasks. In particular with the text processing, LMMs are not only good at interpreting on nouns (i.e. representing objects) but also pay considerable attention on predicates (i.e. representing relations between objects), ensuring their generated contents to be sufficiently coherent [1]. Inspired by this, we propose the Open-set Panoptic Scene Graph Generation architecture, OpenPSG, leveraging the capabilities of LMMs for open-set relation prediction.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we utilize a large multimodal model (e.g., BLIP-2 [18]) to achieve open-set relation prediction. Specifically, our model comprises three parts. First, the open-set panoptic segmenter, we adapt an existing model (e.g., OpenSeeD [44]) which is capable of extracting open-set object categories, masks, and visual features from the whole image, forming object pairs and pair masks. Second, the relation query transformer, which has two functions: extracting visual features of object pairs based on pair masks and with a special focus on pair interactions; judging the potential relations between object pairs. They are realized by two sets of queries, pair feature extraction query and relation existence estimation query. Only those object pairs that are judged to likely have", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 271, + 114, + 506, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relations are fed into the third part, the multimodal relation decoder. This decoder directly inherits from the LMM to predict the open-set relations given an object pair in an auto-regressive manner, on condition of specifically-designed text instructions and pre-extracted pair visual features.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, we are the first to propose the task of open-set panoptic scene graph generation, enabling the open-set prediction of both object masks and relations. Extensive experiments demonstrate that our OpenPSG achieves state-of-the-art results in the closed-set setting and exhibits outstanding performance in the open-set setting.", + "bbox": [ + 212, + 207, + 785, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 304, + 387, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Panoptic Scene Graph Generation", + "text_level": 1, + "bbox": [ + 214, + 334, + 542, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Panoptic scene graph generation stems from scene graph generation (SGG) by replacing bounding boxes with panoptic segmentation masks to represent the objects, so as to achieve a more comprehensive understanding of scenes. Following the introduction of PSG [36], a series of related works [19, 33, 36, 46-48] emerge, significantly advancing its performance. For example, Yang et al. [36], based on DETR [3], introduce an end-to-end framework with learnable queries to generate panoptic scene graphs. Wang et al. [33] design a pair proposal network to filter irrespective subject-object pairs, achieving performance improvement of the PSG. Subsequently, Zhou et al. [47] build a HiLo architecture based on Mask2Former [6] and devise separate branches for high- and low-frequency relations respectively, hence achieving an unbiased relation prediction method. Li et al. [19] re-balance the relation prediction by adaptively transferring information from high-frequency to low-frequency relations. Additionally, Zhao et al. [46] implement a weakly-supervised PSG method given only image-text pairs as annotations, allowing for the learning of panoptic scene graphs from image-level ground truth. Recently, Zhou et al. [48] leverage the rich language information inherent in large language models [1] and design effective image-text interaction modules to assist unbiased PSG. All these works are learned in the closed-set setting, in this paper, we study the open-set PSG, which has been unexplored.", + "bbox": [ + 212, + 357, + 787, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Open-set Scene Graph Generation", + "text_level": 1, + "bbox": [ + 214, + 665, + 545, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-set SGG has been studied in recent years. Earlier works [15, 40], often termed as zero-shot SGG, focus on transferring the knowledge from known relations to unknown relations given their prior connections; for example, Kan et al. [15] leverage the external commonsense knowledge while Yu et al. [40] use knowledge graphs for the knowledge transfer. Subsequently, with the advancement of large multimodal models, various open-set SGG works [5, 12, 38, 45] have emerged; for example, He et al. [12] and Zhang et al. [45] focus on predicting relations between unknown objects in the SGG using multimodal models. Yu et al. [38] and Chen et al. [5] on the other hand focus on the open-set relation prediction in PSG, sharing the same aim with us. Specifically, Yu et al. [38] utilize", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the CLIP model to match visual features with textual relation features for open-set relation prediction; Chen et al. [5] employ a student-teacher network to align visual concepts in multimodal models for predicting open-set relations. In this paper, different from previous works, we introduce an auto-regressive method based on the LMM to achieve open-set relation prediction.", + "bbox": [ + 215, + 146, + 784, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Large Multimodal Models", + "text_level": 1, + "bbox": [ + 215, + 244, + 475, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the introduction of large models like the GPT series [1], recent years have witnessed rapid development in large multimodal models [18, 25, 27]. Benefiting from its nature of connecting vision and language, LMMs have significantly advanced various downstream tasks, ranging from computer vision [9, 39] to natural language processing [14, 28, 32]. Early multimodal models, such as CLIP [27], trained on image-text paired datasets through contrastive learning to align the visual and textual information. Subsequently, owing to enlightenment of the autoregressive prediction in large language models [1, 31], there has been an explosive growth of LMMs [18, 24, 25]; by introducing mechanisms that can transform visual information into large language models, they facilitate the communication between visual and textual information. Furthermore, this has endowed LMMs with the capability to generate free text, leading to substantial improvements in numerous multimodal tasks. In this paper, we leverage LMMs to design a multimodal relation decoder to predict relations in an open-set scenarios.", + "bbox": [ + 215, + 268, + 784, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Task Definition", + "text_level": 1, + "bbox": [ + 215, + 505, + 398, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the task, open-set panoptic scene graph generation. Given an image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , the objective of this task is to extract an open-set panoptic scene graph $G = \\{O, R\\}$ from the image $I$ , where $H$ and $W$ are the height and width of the image. Here:", + "bbox": [ + 215, + 537, + 784, + 597 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $O = \\{o_i\\}_{i=1}^N$ represents $N$ objects segmented from the image, each defined as $o_i = \\{c, m\\}$ , where $c$ is the object category that can belong to either predefined base object categories $C_{base}$ or undefined novel object categories $C_{novel}$ . $m$ represents the binary mask in $\\{0, 1\\}^{H \\times W}$ of the object.", + "- $R = \\{r_{i,j} \\mid i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}$ represents the relations between objects, where $r_{i,j}$ denotes the relation between $o_i$ and $o_j$ , with $o_i$ as the subject and $o_j$ as the object. Each relation $r$ can belong to either predefined base relation categories $K_{base}$ or undefined novel relation categories $K_{novel}$ ." + ], + "bbox": [ + 222, + 607, + 782, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 762, + 328, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, our OpenPSG comprises three components: object segmenter, relation query transformer (RelQ-Former), and multimodal relation decoder (RelDecoder). For the object segmenter (Sec. 4.1), we utilize a pretrained", + "bbox": [ + 215, + 794, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 271, + 114, + 506, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/faeaf50c5ce4cd0274b1cf8fd14822d16869518baec9b9d0b087f96837661315.jpg", + "image_caption": [ + "Fig. 2: The overall framework of our OpenPSG, which comprises three components: object segmenter, relation query transformer and multimodal relation decoder." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "open-set panoptic segmentation model to transform the input image into object categories and masks, as well as visual feature representing the whole image. Subsequently, we input the object categories, masks, and visual feature into the RelQ-Former (Sec. 4.2). Through two sets of learnable queries and complemented by designed instructions, we obtain visual features of object pairs compatible with LMMs' input format as well as judgements on potential relation existence. Finally, only those object pairs being judged to likely have a relation are sent into the RelDecoder (Sec. 4.3) for open-set relation prediction, ultimately yielding an open-set panoptic scene graph.", + "bbox": [ + 212, + 386, + 784, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Object Segmenter", + "text_level": 1, + "bbox": [ + 215, + 546, + 410, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an image $I$ , we utilize the pretrained open-set object segmenter (e.g., OpenSeeD [44]) to predict the objects $O$ within the image and the whole-image visual feature $F_{I} \\in \\mathbb{R}^{h \\times w \\times D}$ . Here, $h$ and $w$ represent the height and width of $F_{I}$ , and $D$ denotes the feature dimension. The segmenter has a similar architecture to Mask2Former [6] including a pixel decoder. The whole-image visual feature $F_{I}$ refers to the visual feature output by the pixel decoder. Below, we develop the patchify module and pairwise module to process the output of the segmenter, generating the input for RelQ-Former.", + "bbox": [ + 212, + 573, + 784, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patchify Module. Patchify module aims to serialize visual feature $F_{I}$ and object masks $m$ , enabling them to be processed as inputs by the RelQ-Former (Sec. 4.2). Similar to the input patchify layer of vision transformer (ViT) [8], we utilize a single convolution layer to transform the extracted $F_{I}$ into a sequence of visual tokens $F_{Iseq} \\in \\mathbb{R}^{L \\times D}$ , where $L$ is the number of patches and $D$ is the feature dimension. When the kernel size and stride of the convolution layer are both $p$ , $L$ is calculated as $L = \\frac{h}{p} \\times \\frac{w}{p}$ . Simultaneously, we employ nearest neighbor interpolation to each extracted object's mask $m_{i}$ , where the size of $m_{i}$", + "bbox": [ + 212, + 718, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is of height $\\frac{h}{p}$ and width $\\frac{w}{p}$ , and then reshape it to a one-dimensional vector with the length $L$ . After processing all masks in the same way, we obtain the mask sequence $m_{seq} \\in \\{0,1\\}^{N \\times L}$ for all objects.", + "bbox": [ + 215, + 145, + 784, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pairwise Module. Pairwise module aims to construct subject-object pairs. Given $N$ objects in the image $I$ , we pairwise all objects $O$ into subject-object pairs $P = \\{(o_i, o_j) | i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}$ . The number of subject-object pairs in $P$ is $N \\times (N - 1)$ , which exhibits exponential growth as $N$ increases. Consequently, we also obtain the combined subject-object pair category set $c^{pair} \\in \\{(c_i, c_j) | i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}$ . We construct the mask sequences for the two objects corresponding to the indices $i$ and $j$ from $m_{seq}$ for each subject-object pair by using the logical $OR$ operation. This operation is performed for all subject-object pairs, resulting in the pair mask sequence $m_{seq}^{pair} \\in \\{0, 1\\}^{N \\times (N - 1) \\times L}$ for the subject-object pairs, where $L$ is the number of patches.", + "bbox": [ + 215, + 215, + 785, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Relation Query Transformer", + "text_level": 1, + "bbox": [ + 215, + 393, + 495, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relation query transformer, leveraging the obtained $F_{Iseq}$ , $c^{pair}$ , and $m_{seq}^{pair}$ , employs two distinct types of queries, pair feature extraction query and relation existence estimation query, along with customized instructions. This approach facilitates the extraction of subject-object pair features and assesses which subject-object pairs likely have relations.", + "bbox": [ + 215, + 420, + 784, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pair Feature Extraction Query. The objective of the pair feature extraction query is to extract corresponding subject-object pair features from the whole image visual feature based on the subject-object pair masks. A common extraction method involves mask pooling [7], which extracts features for the target subject-object pair, treating each area on the subject-object pair equally. However, for features used in relation prediction, they should focus more on areas where interactions between objects occur. By leveraging attention mechanisms, we facilitate interactions among visual tokens representing different areas within the visual feature sequence $F_{Iseq}$ of a subject-object pair. This way can enhance areas that are crucial for relation predictions. Furthermore, inspired by [24], we design an instruction to assist the this learnable query in understanding its purpose for extracting subject-object pair features.", + "bbox": [ + 215, + 520, + 785, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, for each subject-object pair $(o_i,o_j)$ , we first input the pair feature extraction query $Q^{feat}\\in \\mathbb{R}^{E\\times D}$ into a self-attention layer $(SA(\\cdot))$ , along with the pair instruction designed specifically for the pair feature extraction query. This pair instruction is processed through a tokenizer layer to obtain $F_{Inst}^{feat}\\in \\mathbb{R}^{X^{feat}\\times D}$ , which specifies the function of the pair feature extraction query, namely \"Extracting subject-object $(c_i,c_j)$ features from visual features according to the mask\". Here $E$ is the token number of the pair feature extraction query, and $X^{feat}$ is the token number of the pair instruction. Note that we also incorporate the category names of the subject and object $(c_i,c_j)$ into this", + "bbox": [ + 215, + 702, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pair instruction. This operation is formulated as", + "bbox": [ + 215, + 146, + 565, + 161 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {S A} ^ {\\text {f e a t}} = \\operatorname {T r u n c} \\left(S A \\left(\\operatorname {C o n c a t} \\left(Q ^ {\\text {f e a t}}, F _ {\\text {I n s t}} ^ {\\text {f e a t}}\\right)\\right), E\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 176, + 785, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\text{Concat}(\\cdot)$ denotes the concatenation operation, $\\text{Trunc}(\\cdot)$ represents the truncation operation, and $E$ in this truncation operation indicates that we only extract the first $E$ features, namely the features corresponding to the pair feature extraction query. Next, we use a mask cross-attention layer ( $\\text{MaskCA}(\\cdot)$ ), with $F_{SA}^{\\text{feat}}$ as the query, $F_{Iseq}$ as key and value, and $m_{seq}$ as the mask, to extract features corresponding to the subject-object pair, formulated as", + "bbox": [ + 215, + 210, + 787, + 303 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {C A} ^ {f e a t} = \\operatorname {M a s k C A} \\left(F _ {S A} ^ {f e a t}, F _ {I s e q}, m _ {s e q}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 319, + 785, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The features $F_{CA}^{feat}$ are further refined through a feed-forward network $(FFN(\\cdot))$ , formulated as $F_{FFN}^{feat} = FFN(F_{CA}^{feat})$ .", + "bbox": [ + 215, + 353, + 784, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By repeating this process twice, we obtain the visual features for the subject-object pair to be input into the multimodal relation decoder, $F_{I}^{pair(i,j)} \\in \\mathbb{R}^{E \\times D}$ . We perform these operations in parallel for all subject-object pairs to obtain the corresponding features for all pairs.", + "bbox": [ + 215, + 388, + 785, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relation Existence Estimation Query. In addition to the pair feature extraction query, we also design a relation existence estimation query to determine whether a relation likely exists between the subject $o_i$ and object $o_j$ , without predicting the specific relation category. The objective is to filter out irrelevant subject-object pairs to save the computation for subsequent LMM decoding.", + "bbox": [ + 215, + 484, + 785, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, for each subject-object pair $(o_i,o_j)$ , the relation existence estimation query $Q^{exist}\\in \\mathbb{R}^{1\\times D}$ , similar to the pair feature extraction query, is input into the self-attention, mask cross-attention, and feed-forward network layers, interacting respectively with $F_{Iseq}$ , $m_{seq}$ and the specially designed relation instruction. The purpose of the relation instruction is to direct the relation existence estimation query towards determining whether a relation likely exists in the subject-object pair, e.g. \"Is there a relation between $o_i$ and $o_j$ ?\". The relation instruction, after being processed by the tokenizer, results in $F_{Inst}^{exist}\\in \\mathbb{R}^{X^{exist}\\times D}$ , where $X^{exist}$ represents the number of tokens. Eventually, the extracted features are input into a relation existence prediction layer, which includes a 2-layer MLP, and the predicted scores are normalized to [0, 1] using the sigmoid function. It is worth noting that we train it using binary labels indicating whether a relation exists between objects, and a selector module specified below is utilized to perform filtering during inference.", + "bbox": [ + 215, + 564, + 787, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selector. The selector module implemented by 2-layer MLP is set to filter irrelevant subject-object pairs. Only those with a score higher than the threshold $\\theta$ can be input into the multimodal relation decoder. Compared to predicting for all subject-object pairs, this can enables a $20\\times$ speedup in our experiment.", + "bbox": [ + 215, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Multimodal Relation Decoder", + "text_level": 1, + "bbox": [ + 215, + 146, + 509, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal relation decoder aims to utilize the subject-object pair feature $F_{I}^{i,j}$ extracted by the aforementioned modules, combined with an instruction guiding it to achieve open-set relation prediction. Inspired by [39, 41], we first design a generation instruction to perform open-set relation prediction in an autoregressive manner. This works well, yet we find that it tends to favor common relations more or less. Therefore, we further design a judgement instruction, leveraging the LMMs' strong analytical and judgment capabilities. The judgement instruction also utilizes an autoregressive manner but to judge whether a specific relation exists between objects, thereby simplifying the complexity of open-set relation prediction. Next, we specify the two instructions, respectively.", + "bbox": [ + 212, + 167, + 787, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generation Instruction. For the generation instruction, we follow the instruction design used in open-set object recognition [39], utilizing \"What are the relations between $c_i$ and $c_j$ ?\" Here, $c_i$ and $c_j$ respectively refers to the name of the subject and object. We convert this instruction into features $F_{inst}^{gen} \\in \\mathbb{R}^{X^{gen} \\times D}$ using the tokenizer, where $X^{gen}$ is the token number of the generation instruction. We input the features of this generation instruction $F_{inst}^{gen}$ together with the subject-object pair features $F_I^{pair(i,j)}$ into the multimodal relation decoder $Dec(\\cdot)$ , predicting all possible relations in an autoregressive way, formulated as", + "bbox": [ + 214, + 339, + 787, + 464 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i, j} = \\operatorname {D e c} \\left(\\operatorname {C o n c a t} \\left(F _ {I} ^ {\\text {p a i r} (i, j)}, F _ {\\text {i n s t}} ^ {\\text {g e n}}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 470, + 785, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "If multiple relations are predicted, they are separated by the delimiter \"[SEP]\".", + "bbox": [ + 214, + 497, + 781, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Judgement Instruction. Unlike generation instruction, the judgement instruction guides relation decoder to judge, based on a given relation name, whether this relation exists between the subject and object. For example, \"Please judge between $c_i$ and $c_j$ whether there is a relation $r_k$ \". In this case, we only need the multimodal relation decoder to answer \"Yes\" or \"No\" to determine the existence of this relation. Note that inputting the complete judgement instruction for each relation into the decoder can be costly. Therefore, we place the relation name at the end of the instruction. During inference we divide the judgement instruction into two parts: the section before the relation name, transformed into $F_{inst}^{judge}$ through the tokenizer, and the relation name itself, processed into $F_{inst}^{rel}$ . Benefiting from the autoregressive manner for open-set relation prediction, we initially input the subject-object pair feature $F_I^{pair(i,j)}$ and $F_{inst}^{judge}$ into the multimodal relation decoder, formulated as", + "bbox": [ + 214, + 532, + 787, + 731 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF _ {p r e f i x} ^ {(i, j)} = D e c (C o n c a t \\left(F _ {I} ^ {p a i r (i, j)}, F _ {i n s t} ^ {j u d g e}\\right)), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 737, + 785, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which is then cached for subsequent calculations for each relation. For each relation $r_k$ , the multimodal relation decoder only needs to process $F_{prefix}$ and $F_{inst}^{rel(k)}$ to achieve relation prediction, formulated as:", + "bbox": [ + 214, + 765, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {i, j, k} = \\operatorname {D e c} \\left(\\operatorname {C o n c a t} \\left(F _ {\\text {p r e f i x}} ^ {(i, j)}, F _ {\\text {i n s t}} ^ {\\text {r e l} (k)}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 821, + 785, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 269, + 114, + 506, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $J_{i,j,k}$ represents the judgement for $(o_i, r_k, o_j)$ triplet. When $J_{i,j,k}$ is \"Yes\", it indicates that the relation $r_k$ exists between $o_i$ and $o_j$ ; otherwise, it does not exist. Through this approach, we can maintain the same prediction time as with the generation instruction.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We perform the aforementioned process for all subject-object pairs that may have a relation, ultimately achieving open-set relation prediction. For method using generation instruction, we denote it as OpenPSG-G, and for those using judgement instruction, as OpenPSG-J. In next section, OpenPSG by default refers to the latter.", + "bbox": [ + 212, + 207, + 782, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Loss Function", + "text_level": 1, + "bbox": [ + 214, + 306, + 377, + 320 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During the model training, there are two losses involved: the binary cross-entropy loss $\\mathcal{L}_{\\text{exist}}$ for estimating the existence of a relation using the relation existence estimation query in the relation query transformer, and the cross-entropy loss $\\mathcal{L}_{LM}$ consistent with language model training used by the multimodal relation decoder. The total loss is: $\\mathcal{L} = \\lambda \\mathcal{L}_{\\text{exist}} + \\mathcal{L}_{LM}$ , where $\\lambda$ is a weight factor.", + "bbox": [ + 212, + 330, + 782, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 431, + 375, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 463, + 334, + 477 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Panoptic Scene Graph (PSG) dataset [36] is constructed based on the COCO dataset [2, 22], consisting of 48,749 annotated images: 46,563 for training and 2,186 for testing. It encompasses 80 \"thing\" object categories [22] and 53 \"stuff\" object categories [2], as well as 56 relation categories.", + "bbox": [ + 212, + 488, + 782, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visual Genome (VG) dataset [17] is a widely used dataset in the SGG task. To further validate our method, we follow previous works [5, 38] and test our method on the VG-150 variant, which contains 150 object categories and 50 relation categories.", + "bbox": [ + 212, + 549, + 782, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Tasks and metrics", + "text_level": 1, + "bbox": [ + 214, + 633, + 411, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tasks. In both PSG and SGG, there are three distinct subtasks: Predicate Classification (PredCls), Scene Graph Classification (SGCls), and Scene Graph Detection (SGDet) [34]. In PredCls, the categories and locations of objects within the image are given, and only the relation categories between the objects need to be predicted. SGCls requires predicting both the categories of objects and the relations between them, given the locations of objects within the image. SGDet requires the simultaneous prediction of object categories, locations, and relations between objects, based on the given image. In this paper, we focus on the PredCls and SGDet subtasks. The PredCls excludes the influence of segmentation performance and only compares the relation prediction performance, while the SGDet considers the combined results for both object segmentation and relation prediction.", + "bbox": [ + 212, + 657, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, to validate our method's capability in open-set relation prediction, we divide the dataset into base relations and novel relations at a ratio of 7:3. For the PSG dataset, please refer to the supplementary material for the specific division method. For the division of the VG dataset, we follow the practice in previous works [5, 38]. In the open-set scenarios, our model is trained with data only from base relations and tested on both base and novel relations. It is noteworthy that the test sets for open-set and closed-set are same.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. Following previous works [36, 47], we use Recall@K (R@K) and mean Recall@K (mR@K) as our evaluation metrics. Additionally, in open-set scenarios, we also report the R@K and mR@K metrics for base and novel relations.", + "bbox": [ + 212, + 252, + 782, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Implementation details", + "text_level": 1, + "bbox": [ + 214, + 318, + 452, + 333 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our experiments, we utilize the pretrained OpenSeeD [44] as the open-set object segmenter. The patch size $p$ of the patchify module is set to 8. Within the relation query transformer, the length $E$ of the pair feature extraction query is 32, and the threshold $\\theta$ used to filter subject-object pairs is set to 0.35. In the multimodal relation decoder, we employ the decoder of BLIP-2 [18]. During model training, the weight factor $\\lambda$ for the loss is set to 10. We adopt the same data augmentation strategies as in previous methods [36,47]. To train our model, we use the AdamW [26] optimizer with a learning rate of $1e^{-4}$ and a weight decay of $5e^{-2}$ . Our model is trained for a total of 12 epochs, reducing the learning rate to $1e^{-5}$ at the $8_{th}$ epoch. The experimental platform uses four A100 GPUs. Note that during training we freeze the parameters of the object segmenter and multimodal relation decoder but only train the proposed RelQ-Former.", + "bbox": [ + 212, + 342, + 782, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.4 Comparison to the state of the art", + "text_level": 1, + "bbox": [ + 214, + 544, + 545, + 559 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "PSG dataset. Tab. 1 consists of two parts, comparing the performance of our method against previous methods in closed-set and open-set scenarios under the subtasks of predicate classification and scene graph detection. For the first part, in the closed-set scenario, our method significantly surpasses previous methods. For the predicate classification subtask, only methods that predict segmentation and relation separately are applicable [36], and our method achieves a substantial improvement compared to previous best results, for instance, a $26.6\\%$ increase in R@100 and a $25.0\\%$ increase in mR@100. For scene graph detection, our method also achieves a significant increase of $9.0\\%$ over the best previous method [47] in R@100 and a $17.0\\%$ increase in mR@100. This indicates that in the closed-set scenario, our method demonstrates significant performance improvements. For the second part, in the open-set scenario, we train the model only on base relations and test it on all relations. Notably, for predicate classification, our method even outperforms previous methods trained on all relations, which demonstrates the superiority of our method. For example, compared with previous best results, we achieve a $39\\%$ improvement in R@100 and a $7.2\\%$ improvement in mR@100. For the scene graph detection subtask, our method with judgement instruction remains very competitive compared to [36] trained on all relations.", + "bbox": [ + 212, + 568, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5f2e79004ffbe3f759de3159ab268baf6d7d40b0590c50f199ad4b73ec8d56dc.jpg", + "table_caption": [ + "Table 1: Comparison between our OpenPSG and other methods on the PSG dataset in both closed-set and open-set scenarios. Our method shows superior performance compared to all previous methods." + ], + "table_footnote": [], + "table_body": "
MethodPredicate ClassificationScene Graph Detection
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Train on all relations (closed-set)
IMP [34]31.9/9.636.8/10.938.9/11.616.5/6.518.2/7.118.6/7.2
Motifs [43]44.9/20.250.4/22.152.4/22.920.0/9.121.7/9.622.0/9.7
VCTree [30]45.3/20.550.8/22.652.7/23.320.6/9.722.1/10.222.5/10.2
GPSNet [23]31.5/13.239.9/16.444.7/18.317.8/7.019.6/7.520.1/7.7
PSGTR [36]- / -- / -- / -28.4/16.634.4/20.836.3/22.1
PSGFormer [36]- / -- / -- / -18.0/14.819.6/17.020.1/17.6
ADTrans [19]- / 29.0- / 36.2- / 38.826.0/26.429.6/29.730.0/30.0
PairNet [33]- / -- / -- / -29.6/24.735.6/28.539.6/30.6
HiLo [47]- / -- / -- / -34.1/23.740.7/30.343.0/33.1
OpenPSG55.1/39.270.6/53.879.3/63.838.1/32.346.8/40.952.0/50.1
Train on base relations (open-set)
OpenPSG45.1/29.155.5/38.761.5/46.025.9/20.931.6/24.036.7/25.4
", + "bbox": [ + 220, + 198, + 782, + 445 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/be488ea68c5071fc26d3dd32e8760236668cdfedc2e7c63f4e3ff2ec2b998686.jpg", + "table_caption": [ + "Table 2: Comparison between our OpenPSG and other methods on the VG dataset in both closed-set and open-set scenarios with predicate classification subtask." + ], + "table_footnote": [], + "table_body": "
MethodTrain on all relations (closed-set)Train on base relations (open-set)
R/mR@50R/mR@100R/mR@50R/mR@100
Motifs [43]65.2/15.967.1/17.2- / -- / -
VCTree [30]66.4/16.868.1/19.4- / -- / -
Cacao+Epic [38]- /39.0- /40.8- /16.5- /21.8
OvSGTR [5]36.4/ -42.4/ -22.9/ -26.7/ -
OpenPSG60.2/45.871.4/50.325.7/21.530.6/27.2
", + "bbox": [ + 246, + 503, + 751, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "VG dataset. To further validate our method on the VG dataset, we compare to two closed-set SGG methods [30, 43] and two recent open-set SGG works [5, 38]. Since our method relies on an object segmentation model while [5, 30, 38, 43] rely on an object detector, for a fair comparison, we only present the results for predicate classification subtask. Tab. 2 shows the results in both closed-set and open-set scenarios. In the closed-set scenario, compared with previous closed-set methods [30, 43], our method only performs a few points lower on R@50, yet yields significantly better results on the other metrics. In addition, compared with previous best open-set methods [5, 38], our method improves by $29.0\\%$ in R@100 and by $9.5\\%$ in mR@100. In open-set scenario, our method improves upon [5, 38] by $3.9\\%$ in R@100 and by $5.4\\%$ in mR@100. These results demonstrate the effectiveness of our method in both closed-set and open-set relation prediction.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9b6443fb734edc22a9575f772882d6705261c7e9c5b9d8aac36e9d1855dbc18c.jpg", + "table_caption": [ + "Table 3: Ablation study of comparison of different segmenters." + ], + "table_footnote": [], + "table_body": "
SegmenterPredicate ClassificationScene Graph Detection
PQR/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
OpenSeeD [44]55.155.1/39.270.6/53.879.3/63.838.1/32.346.8/40.951.9/50.1
Mask2Former [6]51.754.8/39.170.4/52.978.9/63.736.1/30.344.8/38.448.4/47.8
", + "bbox": [ + 220, + 171, + 782, + 246 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 273, + 385, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To validate the effectiveness of each module in our method, we conduct ablation studies, using the model trained on all relations with judgement instruction for relation prediction as the baseline unless otherwise specified. To eliminate interference, we validate the object segmenter and modules in RelQ-Former under closed-set settings, and test two types of instructions in the multimodal relation decoder under open-set settings.", + "bbox": [ + 212, + 296, + 782, + 385 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Different segmenters. Our method is based on a pretrained segmenter. To verify the impact of different segmenters on model performance, we experiment with two options: the closed-set Mask2Former [6] and the open-set OpenSeeD [44]. As shown in Tab. 3, OpenSeeD outperforms Mask2Former on the Panoptic Quality (PQ) [16] metric (55.1 vs. 51.7) in PSG test set in closed-set scenario. For the predicate classification subtask, the results show that using OpenSeeD is slightly better than using Mask2Former, with R@100 higher by $0.4\\%$ and mR@100 by $0.1\\%$ . For the scene graph detection subtask, OpenSeeD is a better segmenter than Mask2Former, with an increase of $3.5\\%$ in R@100 and $2.3\\%$ in mR@100, due to its superior object segmentation capability (see PQ in Tab. 3).", + "bbox": [ + 212, + 387, + 784, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Subject-object pair features extraction. To validate the effectiveness of the RelQ-Former in extracting subject-object pair features via the attention mechanism, we design an experiment by extracting subject-object pair features using mask pooling. Specifically, for mask pooling, we derive object features by applying mask pooling to their respective positions in the visual features and then concatenate these to form the subject-object pair features. As shown in Tab. 4, it demonstrates that our attention mechanism outperforms mask pooling based pair feature extraction method by $5.2\\%$ in R@100 and by $4.7\\%$ in mR@100. This indicates that our method using RelQ-Former can better focus the extracted features on the interactions between objects, thereby enhancing the performance for relation prediction.", + "bbox": [ + 212, + 537, + 785, + 703 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Selector in RelQ-Former. To validate the efficacy of the selector, we set $\\theta$ to 0, meaning that during inference, we predict relations for all subject-object pairs. As shown in Tab. 4, we find that when we set $\\theta$ to 0, the model performance is approximately the same. However, under these conditions, the model takes 20 times longer to run on the whole PSG test set. This indicates that our selector allows for a significant improvement in computational efficiency with a small sacrifice in performance. For more details, please refer to supplementary material.", + "bbox": [ + 212, + 704, + 785, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Relation Existence Loss. To evaluate the impact of the relation existence estimation loss (Sec. 4.4) on the model, we set the training $\\lambda$ to 0, thereby", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 271, + 114, + 506, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cf4e5fae65723027f38a5b0c5c72f15a22df0b31536858288efabe3c65248c2b.jpg", + "table_caption": [ + "Table 4: Ablation study of comparison of different designs in RelQ-Former." + ], + "table_footnote": [], + "table_body": "
MethodPredicate Classification
R@20mR@20R@50mR@50R@100mR@100
RelQ-Former55.139.270.653.879.363.8
Change to Mask Pooling51.536.365.949.374.159.1
Set θ = 0 in Selector56.840.770.954.079.663.9
Set λ = 0 in Loss54.938.769.853.678.563.2
", + "bbox": [ + 274, + 171, + 723, + 272 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/732c9dc4181813aa11d579d5683653b3441d3576a1c154a174a3d616f2c155cc.jpg", + "table_caption": [ + "Table 5: Ablation study of different instruction types in multimodal relation decoder under various base-to-novel relation ratios. OpenPSG-G: using generation instruction. OpenPSG-J: using judgement instruction." + ], + "table_footnote": [], + "table_body": "
base:novelOpenPSG-GOpenPSG-J
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
7:340.1/24.549.4/32.457.1/36.845.1/29.155.5/38.761.5/46.0
6:434.9/21.346.3/29.252.9/32.040.6/25.153.9/35.357.0/43.5
5:525.4/14.038.6/23.240.1/24.236.4/19.248.9/29.350.3/40.8
4:619.8/11.232.9/14.333.2/18.829.8/17.340.9/21.644.3/34.7
3:713.0/10.717.6/13.018.5/14.422.8/15.728.3/19.735.1/23.7
", + "bbox": [ + 246, + 361, + 750, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "removing this loss for model training. We discover that this loss not only aids in training a relation existence classifier, but also has beneficial effect on the model. As shown in Tab. 4, setting $\\lambda$ to 0 leads to a decrease in R@100 by $0.8\\%$ and in mR@100 by $0.6\\%$ .", + "bbox": [ + 212, + 527, + 784, + 587 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis of instruction types in multimodal relation decoder. To further analyze the two types of instructions, generation instruction and judgement instruction of open-set relation prediction in our multimodal relation decoder (Sec. 4.3), we conduct a series of experiments by adjusting the proportion of novel relation categories, conducting tests with base:novel ratios of 7:3, 6:4, 5:5, 4:6, and 3:7, to validate the performance, and results shown in Tab. 5. First, under the same base:novel ratio, the method using judgement instruction (OpenPSG-J) consistently outperforms the one using generation instruction (OpenPSG-G). For example, at a base:novel ratio of 7:3, OpenPSG-J is $4.4\\%$ higher than OpenPSG-G in R@100, and is $9.2\\%$ higher in mR@100. Second, the results indicate that as the proportion of novel relations increases, the performance of both OpenPSG-G and OpenPSG-G gradually decreases. For OpenPSG-G, R@100 decreases from $57.1\\%$ at a base:novel ratio of 7:3 to $18.5\\%$ at a base:novel ratio of 3:7, a drop of $38.6\\%$ . OpenPSG-J's R@100 decreases by $26.4\\%$ , which is less than the decrease for OpenPSG-G, indicating that OpenPSG-J has a stronger capability for relation prediction in an open world.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/945e0dec144194d49b19bc00ffea3edede98fc965df40486901bd91fbade8f3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 147, + 405, + 243 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/43f82896c0cf6d296dee8e0ed5f3a47527d2c0f25367b9a2d389b575251be0fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 147, + 593, + 244 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ea5ff6e6ef660d009ed346d9ad8071d523a51e9916ae6d1a33c7e9fdc6169e85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 148, + 782, + 243 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/71276d8d5b425ea14af5a23887aee4543350e8b7b6b6e5e1b23e5ff279cabd5c.jpg", + "image_caption": [ + "Image", + "Fig. 3: Visualization results produced by our OpenPSG. The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects." + ], + "image_footnote": [], + "bbox": [ + 220, + 247, + 405, + 349 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/67a9e9e758a42e6dd13f367a4c4a22f0368a4062381af332bd2bdf374f9913e0.jpg", + "image_caption": [ + "Panoptic Segmentation" + ], + "image_footnote": [], + "bbox": [ + 408, + 247, + 593, + 349 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/440cd1caf0ba872cd3fa1b9e07d3a7bc114939b763e4cccb263ae2cff9f0193b.jpg", + "image_caption": [ + "Relation Prediction" + ], + "image_footnote": [], + "bbox": [ + 596, + 247, + 782, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.6 Visualization", + "text_level": 1, + "bbox": [ + 214, + 446, + 370, + 462 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our OpenPSG method can predict relations not defined in the dataset, such as \"flying\", \"observing\", and \"fighting\", which qualitatively demonstrates the excellent open-set relation prediction capability of our method.", + "bbox": [ + 212, + 472, + 784, + 518 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 541, + 359, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose the open-set PSG task and introduce the OpenPSG to accomplish open-set relation prediction. With the help of large multimodal models, our method employs an autoregressive approach to predict open-set relations. Additionally, we have developed a relation query transformer which contains pair feature extraction and relation existence estimation queries, one for extracting features of subject-object pairs, the other for predicting the existence of relations between them to filter out irrelevant pairs. Furthermore, we design generation and judgement instructions to enable the multimodal relation decoder to predict open-set relations. Extensive experiments demonstrate that our method achieves excellent performance in open-set relation prediction. In the future, we plan to employ model distillation to reduce the model size, thus enhancing the practicality in various real-world scenarios while ensuring its ability to predict open-set relations.", + "bbox": [ + 212, + 573, + 787, + 770 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 271, + 114, + 506, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 212, + 175, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2, 3, 4", + "2. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018) 9", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020) 3", + "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: CVPR (2020) 1", + "5. Chen, Z., Wu, J., Lei, Z., Zhang, Z., Chen, C.: Expanding scene graph boundaries: Fully open-vocabulary scene graph generation via visual-concept alignment and retention. arXiv preprint arXiv:2311.10988 (2023) 3, 4, 9, 10, 11", + "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 3, 5, 12", + "7. Dai, J., He, K., Sun, J.: Convolutional feature masking for joint object and stuff segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3992-4000 (2015) 6", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 5", + "9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR. pp. 14084-14093 (2022) 2, 4", + "0. Gao, L., Wang, B., Wang, W.: Image captioning with scene-graph based semantic concepts. In: ICMLC (2018) 1", + "1. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV. pp. 540-557. Springer (2022) 2", + "2. He, T., Gao, L., Song, J., Li, Y.F.: Towards open-vocabulary scene graph generation with prompt-based finetuning. In: ECCV (2022) 3", + "3. Hildebrandt, M., Li, H., Koner, R., Tresp, V., Gunnemann, S.: Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072 (2020) 1", + "4. Joshi, M., Levy, O., Weld, D.S., Zettlemoyer, L.: Bert for coreference resolution: Baselines and analysis. arXiv preprint arXiv:1908.09091 (2019) 4", + "5. Kan, X., Cui, H., Yang, C.: Zero-shot scene graph relation prediction through commonsense knowledge integration. In: Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part II 21. pp. 466-482. Springer (2021) 3" + ], + "bbox": [ + 225, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019) 12", + "17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision pp. 32-73 (2017) 9", + "18. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 2, 4, 10", + "19. Li, L., Ji, W., Wu, Y., Li, M., Qin, Y., Wei, L., Zimmermann, R.: Panoptic scene graph generation with semantics-prototype learning. arXiv preprint arXiv:2307.15567 (2023) 1, 3, 11", + "20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR. pp. 7061-7070 (2023) 2", + "21. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843 (2022) 2", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 9", + "23. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: CVPR (2020) 11", + "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) 4, 6", + "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS 36 (2024) 2, 4", + "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 10", + "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) 2, 4", + "28. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 4", + "29. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: ICCV (2023) 1", + "30. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: CVPR (2019) 11", + "31. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4", + "32. Wang, C., Liu, X., Song, D.: Language models are open knowledge graphs. arXiv preprint arXiv:2010.11967 (2020) 4", + "33. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023) 1, 2, 3, 11", + "34. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: CVPR (2017) 9, 11" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi", + "bbox": [ + 271, + 114, + 506, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Yang, J., Wang, C., Liu, Z., Wu, J., Wang, D., Yang, L., Cao, X.: Focusing on flexible masks: A novel framework for panoptic scene graph generation with relation constraints. In: ACM MM. pp. 4209-4218 (2023) 1", + "36. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation. In: ECCV. pp. 178-196 (2022) 1, 3, 9, 10, 11", + "37. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: CVPR. pp. 23497-23506 (2023) 2", + "38. Yu, Q., Li, J., Wu, Y., Tang, S., Ji, W., Zhuang, Y.: Visually-prompted language model for fine-grained scene graph generation in an open world. arXiv preprint arXiv:2303.13233 (2023) 3, 9, 10, 11", + "39. Yu, Q., Shen, X., Chen, L.C.: Towards open-ended visual recognition with large language model. arXiv preprint arXiv:2311.08400 (2023) 2, 4, 8", + "40. Yu, X., Chen, R., Li, J., Sun, J., Yuan, S., Ji, H., Lu, X., Wu, C.: Zero-shot scene graph generation with knowledge graph completion. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022) 3", + "41. Yue, K., Chen, B.C., Geiping, J., Li, H., Goldstein, T., Lim, S.N.: Object recognition as next token prediction. arXiv preprint arXiv:2312.02142 (2023) 8", + "42. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 2", + "43. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR (2018) 11", + "44. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV. pp. 1020-1031 (2023) 2, 5, 10, 12", + "45. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Learning to generate language-supervised and open-vocabulary scene graph using pre-trained visual-semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2915–2924 (2023) 3", + "46. Zhao, C., Shen, Y., Chen, Z., Ding, M., Gan, C.: Textpsg: Panoptic scene graph generation from textual descriptions. In: ICCV. pp. 2839-2850 (2023) 1, 3", + "47. Zhou, Z., Shi, M., Caesar, H.: Hilo: Exploiting high low frequency relations for unbiased panoptic scene graph generation. arXiv preprint arXiv:2303.15994 (2023) 1, 2, 3, 10, 11", + "48. Zhou, Z., Shi, M., Caesar, H.: Vlprompt: Vision-language prompting for panoptic scene graph generation. arXiv preprint arXiv:2311.16492 (2023) 1, 2, 3", + "49. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 2" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OpenPSG", + "bbox": [ + 661, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_model.json b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_model.json new file mode 100644 index 0000000000000000000000000000000000000000..538f9c02e6fe0015a86292484f1d40aec4f649e3 --- /dev/null +++ b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_model.json @@ -0,0 +1,2280 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "OpenPSG: Open-set Panoptic Scene Graph Generation via Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.212, + 0.737, + 0.229 + ], + "angle": 0, + "content": "Zijian Zhou\\(^{1}\\) Zheng Zhu\\(^{2}\\) Holger Caesar\\(^{3}\\) Miaojing Shi\\(^{4,5\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.239, + 0.714, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Informatics, King's College London \\(^{2}\\)GigaAI" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.254, + 0.692, + 0.268 + ], + "angle": 0, + "content": "3Intelligent Vehicles Lab, Delft University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.268, + 0.756, + 0.297 + ], + "angle": 0, + "content": "4College of Electronic and Information Engineering, Tongji University \n5Shanghai Institute of Intelligent Science and Technology, Tongji University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.336, + 0.744, + 0.629 + ], + "angle": 0, + "content": "Abstract. Panoptic Scene Graph Generation (PSG) aims to segment objects and recognize their relations, enabling the structured understanding of an image. Previous methods focus on predicting predefined object and relation categories, hence limiting their applications in the open world scenarios. With the rapid development of large multimodal models (LMMs), significant progress has been made in open-set object detection and segmentation, yet open-set relation prediction in PSG remains unexplored. In this paper, we focus on the task of open-set relation prediction integrated with a pretrained open-set panoptic segmentation model to achieve true open-set panoptic scene graph generation (OpenPSG). Our OpenPSG leverages LMMs to achieve open-set relation prediction in an autoregressive manner. We introduce a relation query transformer to efficiently extract visual features of object pairs and estimate the existence of relations between them. The latter can enhance the prediction efficiency by filtering irrelevant pairs. Finally, we design the generation and judgement instructions to perform open-set relation prediction in PSG autoregressively. To our knowledge, we are the first to propose the open-set PSG task. Extensive experiments demonstrate that our method achieves state-of-the-art performance in open-set relation prediction and panoptic scene graph generation. Code is available at https://github.com/franciszzj/OpenPSG." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.634, + 0.741, + 0.661 + ], + "angle": 0, + "content": "Keywords: Panoptic Scene Graph Generation \\(\\cdot\\) Open-set \\(\\cdot\\) Large Multimodal Models" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.377, + 0.706 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Panoptic scene graph generation (PSG) [36] aims to segment objects within an image and recognize the relations among them, thereby constructing a panoptic scene graph for a structured understanding of the image. Given its significant potential in applications such as visual question answering [13], image captioning [4, 10], and embodied navigation [29], PSG has attracted considerable attention from researchers ever since its emerging [19, 33, 35, 46-48]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.401, + 0.84 + ], + "angle": 0, + "content": "\\(\\boxdot\\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.405, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.244, + 0.326, + 0.254 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.147, + 0.596, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.244, + 0.547, + 0.254 + ], + "angle": 0, + "content": "Panoptic Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.147, + 0.787, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.244, + 0.731, + 0.254 + ], + "angle": 0, + "content": "Relation Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.267, + 0.788, + 0.339 + ], + "angle": 0, + "content": "Fig. 1: The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects. Our method can predict both known (close-set) relations, e.g., (0_person, playing, 1_skateboard), (2_person, looking at, 0_person), and unknown (open-set) relations, e.g., (0_person, pop shove-it, 1_skateboard), (2_person, recording, 0_person)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.49 + ], + "angle": 0, + "content": "Previous PSG methods [33, 47, 48] are only capable of predicting closed-set object and relation categories while failing to recognize objects/relations beyond predefined categories. Recently, with the advent of large multimodal models (LMMs) such as CLIP [27], BLIP-2 [18] etc., a significant number of open-set prediction methods for object detection [9, 21, 37, 42, 44] and segmentation [11, 20, 39, 44] are introduced, attributing to LMMs' rich understanding of language and strong connections between vision and language. Nevertheless, open-set prediction of relations has been largely unexplored so far." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Compared to open-set object detection and segmentation, open-set relation prediction is more complex: the model is required to both understand different objects and recognize relations of object pairs based on their interactions; especially, the computation of the latter can be exponentially increased. To bridge the gap, in this paper, we focus on the open-set relation prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.688 + ], + "angle": 0, + "content": "LLMs [18, 25, 49] have demonstrated exceptional semantic analysis and understanding abilities across various multimodal tasks. In particular with the text processing, LMMs are not only good at interpreting on nouns (i.e. representing objects) but also pay considerable attention on predicates (i.e. representing relations between objects), ensuring their generated contents to be sufficiently coherent [1]. Inspired by this, we propose the Open-set Panoptic Scene Graph Generation architecture, OpenPSG, leveraging the capabilities of LMMs for open-set relation prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To this end, we utilize a large multimodal model (e.g., BLIP-2 [18]) to achieve open-set relation prediction. Specifically, our model comprises three parts. First, the open-set panoptic segmenter, we adapt an existing model (e.g., OpenSeeD [44]) which is capable of extracting open-set object categories, masks, and visual features from the whole image, forming object pairs and pair masks. Second, the relation query transformer, which has two functions: extracting visual features of object pairs based on pair masks and with a special focus on pair interactions; judging the potential relations between object pairs. They are realized by two sets of queries, pair feature extraction query and relation existence estimation query. Only those object pairs that are judged to likely have" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "relations are fed into the third part, the multimodal relation decoder. This decoder directly inherits from the LMM to predict the open-set relations given an object pair in an auto-regressive manner, on condition of specifically-designed text instructions and pre-extracted pair visual features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.283 + ], + "angle": 0, + "content": "To the best of our knowledge, we are the first to propose the task of open-set panoptic scene graph generation, enabling the open-set prediction of both object masks and relations. Extensive experiments demonstrate that our OpenPSG achieves state-of-the-art results in the closed-set setting and exhibits outstanding performance in the open-set setting." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.305, + 0.388, + 0.32 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.335, + 0.544, + 0.35 + ], + "angle": 0, + "content": "2.1 Panoptic Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Panoptic scene graph generation stems from scene graph generation (SGG) by replacing bounding boxes with panoptic segmentation masks to represent the objects, so as to achieve a more comprehensive understanding of scenes. Following the introduction of PSG [36], a series of related works [19, 33, 36, 46-48] emerge, significantly advancing its performance. For example, Yang et al. [36], based on DETR [3], introduce an end-to-end framework with learnable queries to generate panoptic scene graphs. Wang et al. [33] design a pair proposal network to filter irrespective subject-object pairs, achieving performance improvement of the PSG. Subsequently, Zhou et al. [47] build a HiLo architecture based on Mask2Former [6] and devise separate branches for high- and low-frequency relations respectively, hence achieving an unbiased relation prediction method. Li et al. [19] re-balance the relation prediction by adaptively transferring information from high-frequency to low-frequency relations. Additionally, Zhao et al. [46] implement a weakly-supervised PSG method given only image-text pairs as annotations, allowing for the learning of panoptic scene graphs from image-level ground truth. Recently, Zhou et al. [48] leverage the rich language information inherent in large language models [1] and design effective image-text interaction modules to assist unbiased PSG. All these works are learned in the closed-set setting, in this paper, we study the open-set PSG, which has been unexplored." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.546, + 0.681 + ], + "angle": 0, + "content": "2.2 Open-set Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Open-set SGG has been studied in recent years. Earlier works [15, 40], often termed as zero-shot SGG, focus on transferring the knowledge from known relations to unknown relations given their prior connections; for example, Kan et al. [15] leverage the external commonsense knowledge while Yu et al. [40] use knowledge graphs for the knowledge transfer. Subsequently, with the advancement of large multimodal models, various open-set SGG works [5, 12, 38, 45] have emerged; for example, He et al. [12] and Zhang et al. [45] focus on predicting relations between unknown objects in the SGG using multimodal models. Yu et al. [38] and Chen et al. [5] on the other hand focus on the open-set relation prediction in PSG, sharing the same aim with us. Specifically, Yu et al. [38] utilize" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.222 + ], + "angle": 0, + "content": "the CLIP model to match visual features with textual relation features for open-set relation prediction; Chen et al. [5] employ a student-teacher network to align visual concepts in multimodal models for predicting open-set relations. In this paper, different from previous works, we introduce an auto-regressive method based on the LMM to achieve open-set relation prediction." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.246, + 0.476, + 0.26 + ], + "angle": 0, + "content": "2.3 Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.481 + ], + "angle": 0, + "content": "Since the introduction of large models like the GPT series [1], recent years have witnessed rapid development in large multimodal models [18, 25, 27]. Benefiting from its nature of connecting vision and language, LMMs have significantly advanced various downstream tasks, ranging from computer vision [9, 39] to natural language processing [14, 28, 32]. Early multimodal models, such as CLIP [27], trained on image-text paired datasets through contrastive learning to align the visual and textual information. Subsequently, owing to enlightenment of the autoregressive prediction in large language models [1, 31], there has been an explosive growth of LMMs [18, 24, 25]; by introducing mechanisms that can transform visual information into large language models, they facilitate the communication between visual and textual information. Furthermore, this has endowed LMMs with the capability to generate free text, leading to substantial improvements in numerous multimodal tasks. In this paper, we leverage LMMs to design a multimodal relation decoder to predict relations in an open-set scenarios." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.506, + 0.4, + 0.521 + ], + "angle": 0, + "content": "3 Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.538, + 0.785, + 0.598 + ], + "angle": 0, + "content": "We define the task, open-set panoptic scene graph generation. Given an image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), the objective of this task is to extract an open-set panoptic scene graph \\( G = \\{O, R\\} \\) from the image \\( I \\), where \\( H \\) and \\( W \\) are the height and width of the image. Here:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.608, + 0.784, + 0.667 + ], + "angle": 0, + "content": "- \\( O = \\{o_i\\}_{i=1}^N \\) represents \\( N \\) objects segmented from the image, each defined as \\( o_i = \\{c, m\\} \\), where \\( c \\) is the object category that can belong to either predefined base object categories \\( C_{base} \\) or undefined novel object categories \\( C_{novel} \\). \\( m \\) represents the binary mask in \\( \\{0, 1\\}^{H \\times W} \\) of the object." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.669, + 0.784, + 0.727 + ], + "angle": 0, + "content": "- \\( R = \\{r_{i,j} \\mid i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\} \\) represents the relations between objects, where \\( r_{i,j} \\) denotes the relation between \\( o_i \\) and \\( o_j \\), with \\( o_i \\) as the subject and \\( o_j \\) as the object. Each relation \\( r \\) can belong to either predefined base relation categories \\( K_{base} \\) or undefined novel relation categories \\( K_{novel} \\)." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.608, + 0.784, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.763, + 0.329, + 0.779 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, our OpenPSG comprises three components: object segmenter, relation query transformer (RelQ-Former), and multimodal relation decoder (RelDecoder). For the object segmenter (Sec. 4.1), we utilize a pretrained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.356 + ], + "angle": 0, + "content": "Fig. 2: The overall framework of our OpenPSG, which comprises three components: object segmenter, relation query transformer and multimodal relation decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.523 + ], + "angle": 0, + "content": "open-set panoptic segmentation model to transform the input image into object categories and masks, as well as visual feature representing the whole image. Subsequently, we input the object categories, masks, and visual feature into the RelQ-Former (Sec. 4.2). Through two sets of learnable queries and complemented by designed instructions, we obtain visual features of object pairs compatible with LMMs' input format as well as judgements on potential relation existence. Finally, only those object pairs being judged to likely have a relation are sent into the RelDecoder (Sec. 4.3) for open-set relation prediction, ultimately yielding an open-set panoptic scene graph." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.547, + 0.411, + 0.563 + ], + "angle": 0, + "content": "4.1 Object Segmenter" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.785, + 0.695 + ], + "angle": 0, + "content": "Given an image \\(I\\), we utilize the pretrained open-set object segmenter (e.g., OpenSeeD [44]) to predict the objects \\(O\\) within the image and the whole-image visual feature \\(F_{I} \\in \\mathbb{R}^{h \\times w \\times D}\\). Here, \\(h\\) and \\(w\\) represent the height and width of \\(F_{I}\\), and \\(D\\) denotes the feature dimension. The segmenter has a similar architecture to Mask2Former [6] including a pixel decoder. The whole-image visual feature \\(F_{I}\\) refers to the visual feature output by the pixel decoder. Below, we develop the patchify module and pairwise module to process the output of the segmenter, generating the input for RelQ-Former." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Patchify Module. Patchify module aims to serialize visual feature \\( F_{I} \\) and object masks \\( m \\), enabling them to be processed as inputs by the RelQ-Former (Sec. 4.2). Similar to the input patchify layer of vision transformer (ViT) [8], we utilize a single convolution layer to transform the extracted \\( F_{I} \\) into a sequence of visual tokens \\( F_{Iseq} \\in \\mathbb{R}^{L \\times D} \\), where \\( L \\) is the number of patches and \\( D \\) is the feature dimension. When the kernel size and stride of the convolution layer are both \\( p \\), \\( L \\) is calculated as \\( L = \\frac{h}{p} \\times \\frac{w}{p} \\). Simultaneously, we employ nearest neighbor interpolation to each extracted object's mask \\( m_{i} \\), where the size of \\( m_{i} \\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.506, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.194 + ], + "angle": 0, + "content": "is of height \\( \\frac{h}{p} \\) and width \\( \\frac{w}{p} \\), and then reshape it to a one-dimensional vector with the length \\( L \\). After processing all masks in the same way, we obtain the mask sequence \\( m_{seq} \\in \\{0,1\\}^{N \\times L} \\) for all objects." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.216, + 0.787, + 0.371 + ], + "angle": 0, + "content": "Pairwise Module. Pairwise module aims to construct subject-object pairs. Given \\(N\\) objects in the image \\(I\\), we pairwise all objects \\(O\\) into subject-object pairs \\(P = \\{(o_i, o_j) | i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}\\). The number of subject-object pairs in \\(P\\) is \\(N \\times (N - 1)\\), which exhibits exponential growth as \\(N\\) increases. Consequently, we also obtain the combined subject-object pair category set \\(c^{pair} \\in \\{(c_i, c_j) | i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}\\). We construct the mask sequences for the two objects corresponding to the indices \\(i\\) and \\(j\\) from \\(m_{seq}\\) for each subject-object pair by using the logical \\(OR\\) operation. This operation is performed for all subject-object pairs, resulting in the pair mask sequence \\(m_{seq}^{pair} \\in \\{0, 1\\}^{N \\times (N - 1) \\times L}\\) for the subject-object pairs, where \\(L\\) is the number of patches." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.394, + 0.496, + 0.41 + ], + "angle": 0, + "content": "4.2 Relation Query Transformer" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.497 + ], + "angle": 0, + "content": "Relation query transformer, leveraging the obtained \\( F_{Iseq} \\), \\( c^{pair} \\), and \\( m_{seq}^{pair} \\), employs two distinct types of queries, pair feature extraction query and relation existence estimation query, along with customized instructions. This approach facilitates the extraction of subject-object pair features and assesses which subject-object pairs likely have relations." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.521, + 0.786, + 0.702 + ], + "angle": 0, + "content": "Pair Feature Extraction Query. The objective of the pair feature extraction query is to extract corresponding subject-object pair features from the whole image visual feature based on the subject-object pair masks. A common extraction method involves mask pooling [7], which extracts features for the target subject-object pair, treating each area on the subject-object pair equally. However, for features used in relation prediction, they should focus more on areas where interactions between objects occur. By leveraging attention mechanisms, we facilitate interactions among visual tokens representing different areas within the visual feature sequence \\( F_{Iseq} \\) of a subject-object pair. This way can enhance areas that are crucial for relation predictions. Furthermore, inspired by [24], we design an instruction to assist the this learnable query in understanding its purpose for extracting subject-object pair features." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.703, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Specifically, for each subject-object pair \\((o_i,o_j)\\), we first input the pair feature extraction query \\(Q^{feat}\\in \\mathbb{R}^{E\\times D}\\) into a self-attention layer \\((SA(\\cdot))\\), along with the pair instruction designed specifically for the pair feature extraction query. This pair instruction is processed through a tokenizer layer to obtain \\(F_{Inst}^{feat}\\in \\mathbb{R}^{X^{feat}\\times D}\\), which specifies the function of the pair feature extraction query, namely \"Extracting subject-object \\((c_i,c_j)\\) features from visual features according to the mask\". Here \\(E\\) is the token number of the pair feature extraction query, and \\(X^{feat}\\) is the token number of the pair instruction. Note that we also incorporate the category names of the subject and object \\((c_i,c_j)\\) into this" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.566, + 0.162 + ], + "angle": 0, + "content": "pair instruction. This operation is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.178, + 0.786, + 0.197 + ], + "angle": 0, + "content": "\\[\nF _ {S A} ^ {\\text {f e a t}} = \\operatorname {T r u n c} \\left(S A \\left(\\operatorname {C o n c a t} \\left(Q ^ {\\text {f e a t}}, F _ {\\text {I n s t}} ^ {\\text {f e a t}}\\right)\\right), E\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.788, + 0.304 + ], + "angle": 0, + "content": "where \\( \\text{Concat}(\\cdot) \\) denotes the concatenation operation, \\( \\text{Trunc}(\\cdot) \\) represents the truncation operation, and \\( E \\) in this truncation operation indicates that we only extract the first \\( E \\) features, namely the features corresponding to the pair feature extraction query. Next, we use a mask cross-attention layer (\\( \\text{MaskCA}(\\cdot) \\)), with \\( F_{SA}^{\\text{feat}} \\) as the query, \\( F_{Iseq} \\) as key and value, and \\( m_{seq} \\) as the mask, to extract features corresponding to the subject-object pair, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.32, + 0.786, + 0.339 + ], + "angle": 0, + "content": "\\[\nF _ {C A} ^ {f e a t} = \\operatorname {M a s k C A} \\left(F _ {S A} ^ {f e a t}, F _ {I s e q}, m _ {s e q}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.354, + 0.785, + 0.389 + ], + "angle": 0, + "content": "The features \\( F_{CA}^{feat} \\) are further refined through a feed-forward network \\( (FFN(\\cdot)) \\), formulated as \\( F_{FFN}^{feat} = FFN(F_{CA}^{feat}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.39, + 0.787, + 0.452 + ], + "angle": 0, + "content": "By repeating this process twice, we obtain the visual features for the subject-object pair to be input into the multimodal relation decoder, \\( F_{I}^{pair(i,j)} \\in \\mathbb{R}^{E \\times D} \\). We perform these operations in parallel for all subject-object pairs to obtain the corresponding features for all pairs." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.486, + 0.787, + 0.561 + ], + "angle": 0, + "content": "Relation Existence Estimation Query. In addition to the pair feature extraction query, we also design a relation existence estimation query to determine whether a relation likely exists between the subject \\( o_i \\) and object \\( o_j \\), without predicting the specific relation category. The objective is to filter out irrelevant subject-object pairs to save the computation for subsequent LMM decoding." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.565, + 0.788, + 0.778 + ], + "angle": 0, + "content": "Specifically, for each subject-object pair \\((o_i,o_j)\\), the relation existence estimation query \\(Q^{exist}\\in \\mathbb{R}^{1\\times D}\\), similar to the pair feature extraction query, is input into the self-attention, mask cross-attention, and feed-forward network layers, interacting respectively with \\(F_{Iseq}\\), \\(m_{seq}\\) and the specially designed relation instruction. The purpose of the relation instruction is to direct the relation existence estimation query towards determining whether a relation likely exists in the subject-object pair, e.g. \"Is there a relation between \\(o_i\\) and \\(o_j\\)?\". The relation instruction, after being processed by the tokenizer, results in \\(F_{Inst}^{exist}\\in \\mathbb{R}^{X^{exist}\\times D}\\), where \\(X^{exist}\\) represents the number of tokens. Eventually, the extracted features are input into a relation existence prediction layer, which includes a 2-layer MLP, and the predicted scores are normalized to [0, 1] using the sigmoid function. It is worth noting that we train it using binary labels indicating whether a relation exists between objects, and a selector module specified below is utilized to perform filtering during inference." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Selector. The selector module implemented by 2-layer MLP is set to filter irrelevant subject-object pairs. Only those with a score higher than the threshold \\(\\theta\\) can be input into the multimodal relation decoder. Compared to predicting for all subject-object pairs, this can enables a \\(20\\times\\) speedup in our experiment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.508, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.51, + 0.161 + ], + "angle": 0, + "content": "4.3 Multimodal Relation Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.321 + ], + "angle": 0, + "content": "Multimodal relation decoder aims to utilize the subject-object pair feature \\( F_{I}^{i,j} \\) extracted by the aforementioned modules, combined with an instruction guiding it to achieve open-set relation prediction. Inspired by [39, 41], we first design a generation instruction to perform open-set relation prediction in an autoregressive manner. This works well, yet we find that it tends to favor common relations more or less. Therefore, we further design a judgement instruction, leveraging the LMMs' strong analytical and judgment capabilities. The judgement instruction also utilizes an autoregressive manner but to judge whether a specific relation exists between objects, thereby simplifying the complexity of open-set relation prediction. Next, we specify the two instructions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.34, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Generation Instruction. For the generation instruction, we follow the instruction design used in open-set object recognition [39], utilizing \"What are the relations between \\( c_i \\) and \\( c_j \\)?\" Here, \\( c_i \\) and \\( c_j \\) respectively refers to the name of the subject and object. We convert this instruction into features \\( F_{inst}^{gen} \\in \\mathbb{R}^{X^{gen} \\times D} \\) using the tokenizer, where \\( X^{gen} \\) is the token number of the generation instruction. We input the features of this generation instruction \\( F_{inst}^{gen} \\) together with the subject-object pair features \\( F_I^{pair(i,j)} \\) into the multimodal relation decoder \\( Dec(\\cdot) \\), predicting all possible relations in an autoregressive way, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.471, + 0.786, + 0.492 + ], + "angle": 0, + "content": "\\[\nr _ {i, j} = \\operatorname {D e c} \\left(\\operatorname {C o n c a t} \\left(F _ {I} ^ {\\text {p a i r} (i, j)}, F _ {\\text {i n s t}} ^ {\\text {g e n}}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.782, + 0.514 + ], + "angle": 0, + "content": "If multiple relations are predicted, they are separated by the delimiter \"[SEP]\"." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.533, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Judgement Instruction. Unlike generation instruction, the judgement instruction guides relation decoder to judge, based on a given relation name, whether this relation exists between the subject and object. For example, \"Please judge between \\( c_i \\) and \\( c_j \\) whether there is a relation \\( r_k \\)\". In this case, we only need the multimodal relation decoder to answer \"Yes\" or \"No\" to determine the existence of this relation. Note that inputting the complete judgement instruction for each relation into the decoder can be costly. Therefore, we place the relation name at the end of the instruction. During inference we divide the judgement instruction into two parts: the section before the relation name, transformed into \\( F_{inst}^{judge} \\) through the tokenizer, and the relation name itself, processed into \\( F_{inst}^{rel} \\). Benefiting from the autoregressive manner for open-set relation prediction, we initially input the subject-object pair feature \\( F_I^{pair(i,j)} \\) and \\( F_{inst}^{judge} \\) into the multimodal relation decoder, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.738, + 0.786, + 0.761 + ], + "angle": 0, + "content": "\\[\nF _ {p r e f i x} ^ {(i, j)} = D e c (C o n c a t \\left(F _ {I} ^ {p a i r (i, j)}, F _ {i n s t} ^ {j u d g e}\\right)), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.766, + 0.788, + 0.817 + ], + "angle": 0, + "content": "which is then cached for subsequent calculations for each relation. For each relation \\( r_k \\), the multimodal relation decoder only needs to process \\( F_{prefix} \\) and \\( F_{inst}^{rel(k)} \\) to achieve relation prediction, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.822, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\nJ _ {i, j, k} = \\operatorname {D e c} \\left(\\operatorname {C o n c a t} \\left(F _ {\\text {p r e f i x}} ^ {(i, j)}, F _ {\\text {i n s t}} ^ {\\text {r e l} (k)}\\right)\\right), \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "where \\( J_{i,j,k} \\) represents the judgement for \\( (o_i, r_k, o_j) \\) triplet. When \\( J_{i,j,k} \\) is \"Yes\", it indicates that the relation \\( r_k \\) exists between \\( o_i \\) and \\( o_j \\); otherwise, it does not exist. Through this approach, we can maintain the same prediction time as with the generation instruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.784, + 0.283 + ], + "angle": 0, + "content": "We perform the aforementioned process for all subject-object pairs that may have a relation, ultimately achieving open-set relation prediction. For method using generation instruction, we denote it as OpenPSG-G, and for those using judgement instruction, as OpenPSG-J. In next section, OpenPSG by default refers to the latter." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.378, + 0.321 + ], + "angle": 0, + "content": "4.4 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.784, + 0.409 + ], + "angle": 0, + "content": "During the model training, there are two losses involved: the binary cross-entropy loss \\(\\mathcal{L}_{\\text{exist}}\\) for estimating the existence of a relation using the relation existence estimation query in the relation query transformer, and the cross-entropy loss \\(\\mathcal{L}_{LM}\\) consistent with language model training used by the multimodal relation decoder. The total loss is: \\(\\mathcal{L} = \\lambda \\mathcal{L}_{\\text{exist}} + \\mathcal{L}_{LM}\\), where \\(\\lambda\\) is a weight factor." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.432, + 0.377, + 0.449 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.464, + 0.335, + 0.478 + ], + "angle": 0, + "content": "5.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.784, + 0.549 + ], + "angle": 0, + "content": "Panoptic Scene Graph (PSG) dataset [36] is constructed based on the COCO dataset [2, 22], consisting of 48,749 annotated images: 46,563 for training and 2,186 for testing. It encompasses 80 \"thing\" object categories [22] and 53 \"stuff\" object categories [2], as well as 56 relation categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.784, + 0.61 + ], + "angle": 0, + "content": "Visual Genome (VG) dataset [17] is a widely used dataset in the SGG task. To further validate our method, we follow previous works [5, 38] and test our method on the VG-150 variant, which contains 150 object categories and 50 relation categories." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.412, + 0.648 + ], + "angle": 0, + "content": "5.2 Tasks and metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Tasks. In both PSG and SGG, there are three distinct subtasks: Predicate Classification (PredCls), Scene Graph Classification (SGCls), and Scene Graph Detection (SGDet) [34]. In PredCls, the categories and locations of objects within the image are given, and only the relation categories between the objects need to be predicted. SGCls requires predicting both the categories of objects and the relations between them, given the locations of objects within the image. SGDet requires the simultaneous prediction of object categories, locations, and relations between objects, based on the given image. In this paper, we focus on the PredCls and SGDet subtasks. The PredCls excludes the influence of segmentation performance and only compares the relation prediction performance, while the SGDet considers the combined results for both object segmentation and relation prediction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.506, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.252 + ], + "angle": 0, + "content": "Furthermore, to validate our method's capability in open-set relation prediction, we divide the dataset into base relations and novel relations at a ratio of 7:3. For the PSG dataset, please refer to the supplementary material for the specific division method. For the division of the VG dataset, we follow the practice in previous works [5, 38]. In the open-set scenarios, our model is trained with data only from base relations and tested on both base and novel relations. It is noteworthy that the test sets for open-set and closed-set are same." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.783, + 0.297 + ], + "angle": 0, + "content": "Metrics. Following previous works [36, 47], we use Recall@K (R@K) and mean Recall@K (mR@K) as our evaluation metrics. Additionally, in open-set scenarios, we also report the R@K and mR@K metrics for base and novel relations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.319, + 0.453, + 0.334 + ], + "angle": 0, + "content": "5.3 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.525 + ], + "angle": 0, + "content": "In our experiments, we utilize the pretrained OpenSeeD [44] as the open-set object segmenter. The patch size \\( p \\) of the patchify module is set to 8. Within the relation query transformer, the length \\( E \\) of the pair feature extraction query is 32, and the threshold \\( \\theta \\) used to filter subject-object pairs is set to 0.35. In the multimodal relation decoder, we employ the decoder of BLIP-2 [18]. During model training, the weight factor \\( \\lambda \\) for the loss is set to 10. We adopt the same data augmentation strategies as in previous methods [36,47]. To train our model, we use the AdamW [26] optimizer with a learning rate of \\( 1e^{-4} \\) and a weight decay of \\( 5e^{-2} \\). Our model is trained for a total of 12 epochs, reducing the learning rate to \\( 1e^{-5} \\) at the \\( 8_{th} \\) epoch. The experimental platform uses four A100 GPUs. Note that during training we freeze the parameters of the object segmenter and multimodal relation decoder but only train the proposed RelQ-Former." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.547, + 0.56 + ], + "angle": 0, + "content": "5.4 Comparison to the state of the art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.784, + 0.84 + ], + "angle": 0, + "content": "PSG dataset. Tab. 1 consists of two parts, comparing the performance of our method against previous methods in closed-set and open-set scenarios under the subtasks of predicate classification and scene graph detection. For the first part, in the closed-set scenario, our method significantly surpasses previous methods. For the predicate classification subtask, only methods that predict segmentation and relation separately are applicable [36], and our method achieves a substantial improvement compared to previous best results, for instance, a \\(26.6\\%\\) increase in R@100 and a \\(25.0\\%\\) increase in mR@100. For scene graph detection, our method also achieves a significant increase of \\(9.0\\%\\) over the best previous method [47] in R@100 and a \\(17.0\\%\\) increase in mR@100. This indicates that in the closed-set scenario, our method demonstrates significant performance improvements. For the second part, in the open-set scenario, we train the model only on base relations and test it on all relations. Notably, for predicate classification, our method even outperforms previous methods trained on all relations, which demonstrates the superiority of our method. For example, compared with previous best results, we achieve a \\(39\\%\\) improvement in R@100 and a \\(7.2\\%\\) improvement in mR@100. For the scene graph detection subtask, our method with judgement instruction remains very competitive compared to [36] trained on all relations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison between our OpenPSG and other methods on the PSG dataset in both closed-set and open-set scenarios. Our method shows superior performance compared to all previous methods." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.199, + 0.784, + 0.446 + ], + "angle": 0, + "content": "
MethodPredicate ClassificationScene Graph Detection
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Train on all relations (closed-set)
IMP [34]31.9/9.636.8/10.938.9/11.616.5/6.518.2/7.118.6/7.2
Motifs [43]44.9/20.250.4/22.152.4/22.920.0/9.121.7/9.622.0/9.7
VCTree [30]45.3/20.550.8/22.652.7/23.320.6/9.722.1/10.222.5/10.2
GPSNet [23]31.5/13.239.9/16.444.7/18.317.8/7.019.6/7.520.1/7.7
PSGTR [36]- / -- / -- / -28.4/16.634.4/20.836.3/22.1
PSGFormer [36]- / -- / -- / -18.0/14.819.6/17.020.1/17.6
ADTrans [19]- / 29.0- / 36.2- / 38.826.0/26.429.6/29.730.0/30.0
PairNet [33]- / -- / -- / -29.6/24.735.6/28.539.6/30.6
HiLo [47]- / -- / -- / -34.1/23.740.7/30.343.0/33.1
OpenPSG55.1/39.270.6/53.879.3/63.838.1/32.346.8/40.952.0/50.1
Train on base relations (open-set)
OpenPSG45.1/29.155.5/38.761.5/46.025.9/20.931.6/24.036.7/25.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.463, + 0.784, + 0.491 + ], + "angle": 0, + "content": "Table 2: Comparison between our OpenPSG and other methods on the VG dataset in both closed-set and open-set scenarios with predicate classification subtask." + }, + { + "type": "table", + "bbox": [ + 0.247, + 0.504, + 0.752, + 0.613 + ], + "angle": 0, + "content": "
MethodTrain on all relations (closed-set)Train on base relations (open-set)
R/mR@50R/mR@100R/mR@50R/mR@100
Motifs [43]65.2/15.967.1/17.2- / -- / -
VCTree [30]66.4/16.868.1/19.4- / -- / -
Cacao+Epic [38]- /39.0- /40.8- /16.5- /21.8
OvSGTR [5]36.4/ -42.4/ -22.9/ -26.7/ -
OpenPSG60.2/45.871.4/50.325.7/21.530.6/27.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "VG dataset. To further validate our method on the VG dataset, we compare to two closed-set SGG methods [30, 43] and two recent open-set SGG works [5, 38]. Since our method relies on an object segmentation model while [5, 30, 38, 43] rely on an object detector, for a fair comparison, we only present the results for predicate classification subtask. Tab. 2 shows the results in both closed-set and open-set scenarios. In the closed-set scenario, compared with previous closed-set methods [30, 43], our method only performs a few points lower on R@50, yet yields significantly better results on the other metrics. In addition, compared with previous best open-set methods [5, 38], our method improves by \\(29.0\\%\\) in R@100 and by \\(9.5\\%\\) in mR@100. In open-set scenario, our method improves upon [5, 38] by \\(3.9\\%\\) in R@100 and by \\(5.4\\%\\) in mR@100. These results demonstrate the effectiveness of our method in both closed-set and open-set relation prediction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.145, + 0.712, + 0.159 + ], + "angle": 0, + "content": "Table 3: Ablation study of comparison of different segmenters." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.172, + 0.784, + 0.247 + ], + "angle": 0, + "content": "
SegmenterPredicate ClassificationScene Graph Detection
PQR/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
OpenSeeD [44]55.155.1/39.270.6/53.879.3/63.838.1/32.346.8/40.951.9/50.1
Mask2Former [6]51.754.8/39.170.4/52.978.9/63.736.1/30.344.8/38.448.4/47.8
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.386, + 0.289 + ], + "angle": 0, + "content": "5.5 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.784, + 0.386 + ], + "angle": 0, + "content": "To validate the effectiveness of each module in our method, we conduct ablation studies, using the model trained on all relations with judgement instruction for relation prediction as the baseline unless otherwise specified. To eliminate interference, we validate the object segmenter and modules in RelQ-Former under closed-set settings, and test two types of instructions in the multimodal relation decoder under open-set settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Different segmenters. Our method is based on a pretrained segmenter. To verify the impact of different segmenters on model performance, we experiment with two options: the closed-set Mask2Former [6] and the open-set OpenSeeD [44]. As shown in Tab. 3, OpenSeeD outperforms Mask2Former on the Panoptic Quality (PQ) [16] metric (55.1 vs. 51.7) in PSG test set in closed-set scenario. For the predicate classification subtask, the results show that using OpenSeeD is slightly better than using Mask2Former, with R@100 higher by \\(0.4\\%\\) and mR@100 by \\(0.1\\%\\). For the scene graph detection subtask, OpenSeeD is a better segmenter than Mask2Former, with an increase of \\(3.5\\%\\) in R@100 and \\(2.3\\%\\) in mR@100, due to its superior object segmentation capability (see PQ in Tab. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.704 + ], + "angle": 0, + "content": "Subject-object pair features extraction. To validate the effectiveness of the RelQ-Former in extracting subject-object pair features via the attention mechanism, we design an experiment by extracting subject-object pair features using mask pooling. Specifically, for mask pooling, we derive object features by applying mask pooling to their respective positions in the visual features and then concatenate these to form the subject-object pair features. As shown in Tab. 4, it demonstrates that our attention mechanism outperforms mask pooling based pair feature extraction method by \\(5.2\\%\\) in R@100 and by \\(4.7\\%\\) in mR@100. This indicates that our method using RelQ-Former can better focus the extracted features on the interactions between objects, thereby enhancing the performance for relation prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Selector in RelQ-Former. To validate the efficacy of the selector, we set \\(\\theta\\) to 0, meaning that during inference, we predict relations for all subject-object pairs. As shown in Tab. 4, we find that when we set \\(\\theta\\) to 0, the model performance is approximately the same. However, under these conditions, the model takes 20 times longer to run on the whole PSG test set. This indicates that our selector allows for a significant improvement in computational efficiency with a small sacrifice in performance. For more details, please refer to supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Relation Existence Loss. To evaluate the impact of the relation existence estimation loss (Sec. 4.4) on the model, we set the training \\(\\lambda\\) to 0, thereby" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.145, + 0.754, + 0.16 + ], + "angle": 0, + "content": "Table 4: Ablation study of comparison of different designs in RelQ-Former." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.172, + 0.725, + 0.273 + ], + "angle": 0, + "content": "
MethodPredicate Classification
R@20mR@20R@50mR@50R@100mR@100
RelQ-Former55.139.270.653.879.363.8
Change to Mask Pooling51.536.365.949.374.159.1
Set θ = 0 in Selector56.840.770.954.079.663.9
Set λ = 0 in Loss54.938.769.853.678.563.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.308, + 0.785, + 0.351 + ], + "angle": 0, + "content": "Table 5: Ablation study of different instruction types in multimodal relation decoder under various base-to-novel relation ratios. OpenPSG-G: using generation instruction. OpenPSG-J: using judgement instruction." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.362, + 0.751, + 0.478 + ], + "angle": 0, + "content": "
base:novelOpenPSG-GOpenPSG-J
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
7:340.1/24.549.4/32.457.1/36.845.1/29.155.5/38.761.5/46.0
6:434.9/21.346.3/29.252.9/32.040.6/25.153.9/35.357.0/43.5
5:525.4/14.038.6/23.240.1/24.236.4/19.248.9/29.350.3/40.8
4:619.8/11.232.9/14.333.2/18.829.8/17.340.9/21.644.3/34.7
3:713.0/10.717.6/13.018.5/14.422.8/15.728.3/19.735.1/23.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.785, + 0.588 + ], + "angle": 0, + "content": "removing this loss for model training. We discover that this loss not only aids in training a relation existence classifier, but also has beneficial effect on the model. As shown in Tab. 4, setting \\(\\lambda\\) to 0 leads to a decrease in R@100 by \\(0.8\\%\\) and in mR@100 by \\(0.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Analysis of instruction types in multimodal relation decoder. To further analyze the two types of instructions, generation instruction and judgement instruction of open-set relation prediction in our multimodal relation decoder (Sec. 4.3), we conduct a series of experiments by adjusting the proportion of novel relation categories, conducting tests with base:novel ratios of 7:3, 6:4, 5:5, 4:6, and 3:7, to validate the performance, and results shown in Tab. 5. First, under the same base:novel ratio, the method using judgement instruction (OpenPSG-J) consistently outperforms the one using generation instruction (OpenPSG-G). For example, at a base:novel ratio of 7:3, OpenPSG-J is \\(4.4\\%\\) higher than OpenPSG-G in R@100, and is \\(9.2\\%\\) higher in mR@100. Second, the results indicate that as the proportion of novel relations increases, the performance of both OpenPSG-G and OpenPSG-G gradually decreases. For OpenPSG-G, R@100 decreases from \\(57.1\\%\\) at a base:novel ratio of 7:3 to \\(18.5\\%\\) at a base:novel ratio of 3:7, a drop of \\(38.6\\%\\). OpenPSG-J's R@100 decreases by \\(26.4\\%\\), which is less than the decrease for OpenPSG-G, indicating that OpenPSG-J has a stronger capability for relation prediction in an open world." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.148, + 0.406, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.148, + 0.594, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.149, + 0.783, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.248, + 0.406, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.351, + 0.328, + 0.361 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.248, + 0.594, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.352, + 0.55, + 0.361 + ], + "angle": 0, + "content": "Panoptic Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.248, + 0.783, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.351, + 0.732, + 0.36 + ], + "angle": 0, + "content": "Relation Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.376, + 0.785, + 0.418 + ], + "angle": 0, + "content": "Fig. 3: Visualization results produced by our OpenPSG. The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.448, + 0.371, + 0.463 + ], + "angle": 0, + "content": "5.6 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.785, + 0.519 + ], + "angle": 0, + "content": "As shown in Fig. 3, our OpenPSG method can predict relations not defined in the dataset, such as \"flying\", \"observing\", and \"fighting\", which qualitatively demonstrates the excellent open-set relation prediction capability of our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.542, + 0.36, + 0.558 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.771 + ], + "angle": 0, + "content": "In this paper, we propose the open-set PSG task and introduce the OpenPSG to accomplish open-set relation prediction. With the help of large multimodal models, our method employs an autoregressive approach to predict open-set relations. Additionally, we have developed a relation query transformer which contains pair feature extraction and relation existence estimation queries, one for extracting features of subject-object pairs, the other for predicting the existence of relations between them to filter out irrelevant pairs. Furthermore, we design generation and judgement instructions to enable the multimodal relation decoder to predict open-set relations. Extensive experiments demonstrate that our method achieves excellent performance in open-set relation prediction. In the future, we plan to employ model distillation to reduce the model size, thus enhancing the practicality in various real-world scenarios while ensuring its ability to predict open-set relations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.237 + ], + "angle": 0, + "content": "The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the Fundamental Research Funds for the Central Universities." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.358 + ], + "angle": 0, + "content": "2. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.785, + 0.386 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.785, + 0.413 + ], + "angle": 0, + "content": "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: CVPR (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.414, + 0.785, + 0.455 + ], + "angle": 0, + "content": "5. Chen, Z., Wu, J., Lei, Z., Zhang, Z., Chen, C.: Expanding scene graph boundaries: Fully open-vocabulary scene graph generation via visual-concept alignment and retention. arXiv preprint arXiv:2311.10988 (2023) 3, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.785, + 0.496 + ], + "angle": 0, + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 3, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.537 + ], + "angle": 0, + "content": "7. Dai, J., He, K., Sun, J.: Convolutional feature masking for joint object and stuff segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3992-4000 (2015) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.592 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.634 + ], + "angle": 0, + "content": "9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR. pp. 14084-14093 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.661 + ], + "angle": 0, + "content": "0. Gao, L., Wang, B., Wang, W.: Image captioning with scene-graph based semantic concepts. In: ICMLC (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.785, + 0.689 + ], + "angle": 0, + "content": "1. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV. pp. 540-557. Springer (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.716 + ], + "angle": 0, + "content": "2. He, T., Gao, L., Song, J., Li, Y.F.: Towards open-vocabulary scene graph generation with prompt-based finetuning. In: ECCV (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "3. Hildebrandt, M., Li, H., Koner, R., Tresp, V., Gunnemann, S.: Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "4. Joshi, M., Levy, O., Weld, D.S., Zettlemoyer, L.: Bert for coreference resolution: Baselines and analysis. arXiv preprint arXiv:1908.09091 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Kan, X., Cui, H., Yang, C.: Zero-shot scene graph relation prediction through commonsense knowledge integration. In: Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part II 21. pp. 466-482. Springer (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.128 + ], + "angle": 0, + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision pp. 32-73 (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.285 + ], + "angle": 0, + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "19. Li, L., Ji, W., Wu, Y., Li, M., Qin, Y., Wei, L., Zimmermann, R.: Panoptic scene graph generation with semantics-prototype learning. arXiv preprint arXiv:2307.15567 (2023) 1, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR. pp. 7061-7070 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.41 + ], + "angle": 0, + "content": "21. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "23. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: CVPR (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.521 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS 36 (2024) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.548 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.645 + ], + "angle": 0, + "content": "28. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "29. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: ICCV (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "30. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: CVPR (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "31. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "32. Wang, C., Liu, X., Song, D.: Language models are open knowledge graphs. arXiv preprint arXiv:2010.11967 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.811 + ], + "angle": 0, + "content": "33. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023) 1, 2, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: CVPR (2017) 9, 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OpenPSG" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Yang, J., Wang, C., Liu, Z., Wu, J., Wang, D., Yang, L., Cao, X.: Focusing on flexible masks: A novel framework for panoptic scene graph generation with relation constraints. In: ACM MM. pp. 4209-4218 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "36. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation. In: ECCV. pp. 178-196 (2022) 1, 3, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "37. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: CVPR. pp. 23497-23506 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "38. Yu, Q., Li, J., Wu, Y., Tang, S., Ji, W., Zhuang, Y.: Visually-prompted language model for fine-grained scene graph generation in an open world. arXiv preprint arXiv:2303.13233 (2023) 3, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "39. Yu, Q., Shen, X., Chen, L.C.: Towards open-ended visual recognition with large language model. arXiv preprint arXiv:2311.08400 (2023) 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "40. Yu, X., Chen, R., Li, J., Sun, J., Yuan, S., Ji, H., Lu, X., Wu, C.: Zero-shot scene graph generation with knowledge graph completion. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.398 + ], + "angle": 0, + "content": "41. Yue, K., Chen, B.C., Geiping, J., Li, H., Goldstein, T., Lim, S.N.: Object recognition as next token prediction. arXiv preprint arXiv:2312.02142 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.425 + ], + "angle": 0, + "content": "42. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "43. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR (2018) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "44. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV. pp. 1020-1031 (2023) 2, 5, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.55 + ], + "angle": 0, + "content": "45. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Learning to generate language-supervised and open-vocabulary scene graph using pre-trained visual-semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2915–2924 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.578 + ], + "angle": 0, + "content": "46. Zhao, C., Shen, Y., Chen, Z., Ding, M., Gan, C.: Textpsg: Panoptic scene graph generation from textual descriptions. In: ICCV. pp. 2839-2850 (2023) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "47. Zhou, Z., Shi, M., Caesar, H.: Hilo: Exploiting high low frequency relations for unbiased panoptic scene graph generation. arXiv preprint arXiv:2303.15994 (2023) 1, 2, 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.647 + ], + "angle": 0, + "content": "48. Zhou, Z., Shi, M., Caesar, H.: Vlprompt: Vision-language prompting for panoptic scene graph generation. arXiv preprint arXiv:2311.16492 (2023) 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "49. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_origin.pdf b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..426296ad36166f9333c1f9da0c89884654ee4ded --- /dev/null +++ b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/024687ee-ad97-4e1f-98cd-0a55f51a1e57_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8278e5bf002722a49530a6755cdd4f8249d8c56aef882029bd8442fc2e893c9b +size 1407149 diff --git a/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/full.md b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a63a68a1aa9e33915c5674a8bcf61a62389cd2f --- /dev/null +++ b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/full.md @@ -0,0 +1,288 @@ +# OpenPSG: Open-set Panoptic Scene Graph Generation via Large Multimodal Models + +Zijian Zhou $^{1}$ Zheng Zhu $^{2}$ Holger Caesar $^{3}$ Miaojing Shi $^{4,5\boxtimes}$ + +$^{1}$ Department of Informatics, King's College London $^{2}$ GigaAI + +3Intelligent Vehicles Lab, Delft University of Technology + +4College of Electronic and Information Engineering, Tongji University +5Shanghai Institute of Intelligent Science and Technology, Tongji University + +Abstract. Panoptic Scene Graph Generation (PSG) aims to segment objects and recognize their relations, enabling the structured understanding of an image. Previous methods focus on predicting predefined object and relation categories, hence limiting their applications in the open world scenarios. With the rapid development of large multimodal models (LMMs), significant progress has been made in open-set object detection and segmentation, yet open-set relation prediction in PSG remains unexplored. In this paper, we focus on the task of open-set relation prediction integrated with a pretrained open-set panoptic segmentation model to achieve true open-set panoptic scene graph generation (OpenPSG). Our OpenPSG leverages LMMs to achieve open-set relation prediction in an autoregressive manner. We introduce a relation query transformer to efficiently extract visual features of object pairs and estimate the existence of relations between them. The latter can enhance the prediction efficiency by filtering irrelevant pairs. Finally, we design the generation and judgement instructions to perform open-set relation prediction in PSG autoregressively. To our knowledge, we are the first to propose the open-set PSG task. Extensive experiments demonstrate that our method achieves state-of-the-art performance in open-set relation prediction and panoptic scene graph generation. Code is available at https://github.com/franciszzj/OpenPSG. + +Keywords: Panoptic Scene Graph Generation $\cdot$ Open-set $\cdot$ Large Multimodal Models + +# 1 Introduction + +Panoptic scene graph generation (PSG) [36] aims to segment objects within an image and recognize the relations among them, thereby constructing a panoptic scene graph for a structured understanding of the image. Given its significant potential in applications such as visual question answering [13], image captioning [4, 10], and embodied navigation [29], PSG has attracted considerable attention from researchers ever since its emerging [19, 33, 35, 46-48]. + +![](images/39f4576820ef7e83836b9ae10496dfac3e2e398fbd2c3948dd76d7f41caf0287.jpg) +Image +Fig. 1: The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects. Our method can predict both known (close-set) relations, e.g., (0_person, playing, 1_skateboard), (2_person, looking at, 0_person), and unknown (open-set) relations, e.g., (0_person, pop shove-it, 1_skateboard), (2_person, recording, 0_person). + +![](images/af3d444f853e1b60198c8c5d2a37ab60ad3336fc0bb50681d4adabee86c03426.jpg) +Panoptic Segmentation + +![](images/1f18b16b500d07235addb206562c29206da4aebacedb346cbbceb2399ca831fe.jpg) +Relation Prediction + +Previous PSG methods [33, 47, 48] are only capable of predicting closed-set object and relation categories while failing to recognize objects/relations beyond predefined categories. Recently, with the advent of large multimodal models (LMMs) such as CLIP [27], BLIP-2 [18] etc., a significant number of open-set prediction methods for object detection [9, 21, 37, 42, 44] and segmentation [11, 20, 39, 44] are introduced, attributing to LMMs' rich understanding of language and strong connections between vision and language. Nevertheless, open-set prediction of relations has been largely unexplored so far. + +Compared to open-set object detection and segmentation, open-set relation prediction is more complex: the model is required to both understand different objects and recognize relations of object pairs based on their interactions; especially, the computation of the latter can be exponentially increased. To bridge the gap, in this paper, we focus on the open-set relation prediction. + +LLMs [18, 25, 49] have demonstrated exceptional semantic analysis and understanding abilities across various multimodal tasks. In particular with the text processing, LMMs are not only good at interpreting on nouns (i.e. representing objects) but also pay considerable attention on predicates (i.e. representing relations between objects), ensuring their generated contents to be sufficiently coherent [1]. Inspired by this, we propose the Open-set Panoptic Scene Graph Generation architecture, OpenPSG, leveraging the capabilities of LMMs for open-set relation prediction. + +To this end, we utilize a large multimodal model (e.g., BLIP-2 [18]) to achieve open-set relation prediction. Specifically, our model comprises three parts. First, the open-set panoptic segmenter, we adapt an existing model (e.g., OpenSeeD [44]) which is capable of extracting open-set object categories, masks, and visual features from the whole image, forming object pairs and pair masks. Second, the relation query transformer, which has two functions: extracting visual features of object pairs based on pair masks and with a special focus on pair interactions; judging the potential relations between object pairs. They are realized by two sets of queries, pair feature extraction query and relation existence estimation query. Only those object pairs that are judged to likely have + +relations are fed into the third part, the multimodal relation decoder. This decoder directly inherits from the LMM to predict the open-set relations given an object pair in an auto-regressive manner, on condition of specifically-designed text instructions and pre-extracted pair visual features. + +To the best of our knowledge, we are the first to propose the task of open-set panoptic scene graph generation, enabling the open-set prediction of both object masks and relations. Extensive experiments demonstrate that our OpenPSG achieves state-of-the-art results in the closed-set setting and exhibits outstanding performance in the open-set setting. + +# 2 Related Work + +# 2.1 Panoptic Scene Graph Generation + +Panoptic scene graph generation stems from scene graph generation (SGG) by replacing bounding boxes with panoptic segmentation masks to represent the objects, so as to achieve a more comprehensive understanding of scenes. Following the introduction of PSG [36], a series of related works [19, 33, 36, 46-48] emerge, significantly advancing its performance. For example, Yang et al. [36], based on DETR [3], introduce an end-to-end framework with learnable queries to generate panoptic scene graphs. Wang et al. [33] design a pair proposal network to filter irrespective subject-object pairs, achieving performance improvement of the PSG. Subsequently, Zhou et al. [47] build a HiLo architecture based on Mask2Former [6] and devise separate branches for high- and low-frequency relations respectively, hence achieving an unbiased relation prediction method. Li et al. [19] re-balance the relation prediction by adaptively transferring information from high-frequency to low-frequency relations. Additionally, Zhao et al. [46] implement a weakly-supervised PSG method given only image-text pairs as annotations, allowing for the learning of panoptic scene graphs from image-level ground truth. Recently, Zhou et al. [48] leverage the rich language information inherent in large language models [1] and design effective image-text interaction modules to assist unbiased PSG. All these works are learned in the closed-set setting, in this paper, we study the open-set PSG, which has been unexplored. + +# 2.2 Open-set Scene Graph Generation + +Open-set SGG has been studied in recent years. Earlier works [15, 40], often termed as zero-shot SGG, focus on transferring the knowledge from known relations to unknown relations given their prior connections; for example, Kan et al. [15] leverage the external commonsense knowledge while Yu et al. [40] use knowledge graphs for the knowledge transfer. Subsequently, with the advancement of large multimodal models, various open-set SGG works [5, 12, 38, 45] have emerged; for example, He et al. [12] and Zhang et al. [45] focus on predicting relations between unknown objects in the SGG using multimodal models. Yu et al. [38] and Chen et al. [5] on the other hand focus on the open-set relation prediction in PSG, sharing the same aim with us. Specifically, Yu et al. [38] utilize + +the CLIP model to match visual features with textual relation features for open-set relation prediction; Chen et al. [5] employ a student-teacher network to align visual concepts in multimodal models for predicting open-set relations. In this paper, different from previous works, we introduce an auto-regressive method based on the LMM to achieve open-set relation prediction. + +# 2.3 Large Multimodal Models + +Since the introduction of large models like the GPT series [1], recent years have witnessed rapid development in large multimodal models [18, 25, 27]. Benefiting from its nature of connecting vision and language, LMMs have significantly advanced various downstream tasks, ranging from computer vision [9, 39] to natural language processing [14, 28, 32]. Early multimodal models, such as CLIP [27], trained on image-text paired datasets through contrastive learning to align the visual and textual information. Subsequently, owing to enlightenment of the autoregressive prediction in large language models [1, 31], there has been an explosive growth of LMMs [18, 24, 25]; by introducing mechanisms that can transform visual information into large language models, they facilitate the communication between visual and textual information. Furthermore, this has endowed LMMs with the capability to generate free text, leading to substantial improvements in numerous multimodal tasks. In this paper, we leverage LMMs to design a multimodal relation decoder to predict relations in an open-set scenarios. + +# 3 Task Definition + +We define the task, open-set panoptic scene graph generation. Given an image $I \in \mathbb{R}^{H \times W \times 3}$ , the objective of this task is to extract an open-set panoptic scene graph $G = \{O, R\}$ from the image $I$ , where $H$ and $W$ are the height and width of the image. Here: + +- $O = \{o_i\}_{i=1}^N$ represents $N$ objects segmented from the image, each defined as $o_i = \{c, m\}$ , where $c$ is the object category that can belong to either predefined base object categories $C_{base}$ or undefined novel object categories $C_{novel}$ . $m$ represents the binary mask in $\{0, 1\}^{H \times W}$ of the object. +- $R = \{r_{i,j} \mid i, j \in \{1, 2, \dots, N\}, i \neq j\}$ represents the relations between objects, where $r_{i,j}$ denotes the relation between $o_i$ and $o_j$ , with $o_i$ as the subject and $o_j$ as the object. Each relation $r$ can belong to either predefined base relation categories $K_{base}$ or undefined novel relation categories $K_{novel}$ . + +# 4 Method + +As illustrated in Fig. 2, our OpenPSG comprises three components: object segmenter, relation query transformer (RelQ-Former), and multimodal relation decoder (RelDecoder). For the object segmenter (Sec. 4.1), we utilize a pretrained + +![](images/faeaf50c5ce4cd0274b1cf8fd14822d16869518baec9b9d0b087f96837661315.jpg) +Fig. 2: The overall framework of our OpenPSG, which comprises three components: object segmenter, relation query transformer and multimodal relation decoder. + +open-set panoptic segmentation model to transform the input image into object categories and masks, as well as visual feature representing the whole image. Subsequently, we input the object categories, masks, and visual feature into the RelQ-Former (Sec. 4.2). Through two sets of learnable queries and complemented by designed instructions, we obtain visual features of object pairs compatible with LMMs' input format as well as judgements on potential relation existence. Finally, only those object pairs being judged to likely have a relation are sent into the RelDecoder (Sec. 4.3) for open-set relation prediction, ultimately yielding an open-set panoptic scene graph. + +# 4.1 Object Segmenter + +Given an image $I$ , we utilize the pretrained open-set object segmenter (e.g., OpenSeeD [44]) to predict the objects $O$ within the image and the whole-image visual feature $F_{I} \in \mathbb{R}^{h \times w \times D}$ . Here, $h$ and $w$ represent the height and width of $F_{I}$ , and $D$ denotes the feature dimension. The segmenter has a similar architecture to Mask2Former [6] including a pixel decoder. The whole-image visual feature $F_{I}$ refers to the visual feature output by the pixel decoder. Below, we develop the patchify module and pairwise module to process the output of the segmenter, generating the input for RelQ-Former. + +Patchify Module. Patchify module aims to serialize visual feature $F_{I}$ and object masks $m$ , enabling them to be processed as inputs by the RelQ-Former (Sec. 4.2). Similar to the input patchify layer of vision transformer (ViT) [8], we utilize a single convolution layer to transform the extracted $F_{I}$ into a sequence of visual tokens $F_{Iseq} \in \mathbb{R}^{L \times D}$ , where $L$ is the number of patches and $D$ is the feature dimension. When the kernel size and stride of the convolution layer are both $p$ , $L$ is calculated as $L = \frac{h}{p} \times \frac{w}{p}$ . Simultaneously, we employ nearest neighbor interpolation to each extracted object's mask $m_{i}$ , where the size of $m_{i}$ + +is of height $\frac{h}{p}$ and width $\frac{w}{p}$ , and then reshape it to a one-dimensional vector with the length $L$ . After processing all masks in the same way, we obtain the mask sequence $m_{seq} \in \{0,1\}^{N \times L}$ for all objects. + +Pairwise Module. Pairwise module aims to construct subject-object pairs. Given $N$ objects in the image $I$ , we pairwise all objects $O$ into subject-object pairs $P = \{(o_i, o_j) | i, j \in \{1, 2, \dots, N\}, i \neq j\}$ . The number of subject-object pairs in $P$ is $N \times (N - 1)$ , which exhibits exponential growth as $N$ increases. Consequently, we also obtain the combined subject-object pair category set $c^{pair} \in \{(c_i, c_j) | i, j \in \{1, 2, \dots, N\}, i \neq j\}$ . We construct the mask sequences for the two objects corresponding to the indices $i$ and $j$ from $m_{seq}$ for each subject-object pair by using the logical $OR$ operation. This operation is performed for all subject-object pairs, resulting in the pair mask sequence $m_{seq}^{pair} \in \{0, 1\}^{N \times (N - 1) \times L}$ for the subject-object pairs, where $L$ is the number of patches. + +# 4.2 Relation Query Transformer + +Relation query transformer, leveraging the obtained $F_{Iseq}$ , $c^{pair}$ , and $m_{seq}^{pair}$ , employs two distinct types of queries, pair feature extraction query and relation existence estimation query, along with customized instructions. This approach facilitates the extraction of subject-object pair features and assesses which subject-object pairs likely have relations. + +Pair Feature Extraction Query. The objective of the pair feature extraction query is to extract corresponding subject-object pair features from the whole image visual feature based on the subject-object pair masks. A common extraction method involves mask pooling [7], which extracts features for the target subject-object pair, treating each area on the subject-object pair equally. However, for features used in relation prediction, they should focus more on areas where interactions between objects occur. By leveraging attention mechanisms, we facilitate interactions among visual tokens representing different areas within the visual feature sequence $F_{Iseq}$ of a subject-object pair. This way can enhance areas that are crucial for relation predictions. Furthermore, inspired by [24], we design an instruction to assist the this learnable query in understanding its purpose for extracting subject-object pair features. + +Specifically, for each subject-object pair $(o_i,o_j)$ , we first input the pair feature extraction query $Q^{feat}\in \mathbb{R}^{E\times D}$ into a self-attention layer $(SA(\cdot))$ , along with the pair instruction designed specifically for the pair feature extraction query. This pair instruction is processed through a tokenizer layer to obtain $F_{Inst}^{feat}\in \mathbb{R}^{X^{feat}\times D}$ , which specifies the function of the pair feature extraction query, namely "Extracting subject-object $(c_i,c_j)$ features from visual features according to the mask". Here $E$ is the token number of the pair feature extraction query, and $X^{feat}$ is the token number of the pair instruction. Note that we also incorporate the category names of the subject and object $(c_i,c_j)$ into this + +pair instruction. This operation is formulated as + +$$ +F _ {S A} ^ {\text {f e a t}} = \operatorname {T r u n c} \left(S A \left(\operatorname {C o n c a t} \left(Q ^ {\text {f e a t}}, F _ {\text {I n s t}} ^ {\text {f e a t}}\right)\right), E\right), \tag {1} +$$ + +where $\text{Concat}(\cdot)$ denotes the concatenation operation, $\text{Trunc}(\cdot)$ represents the truncation operation, and $E$ in this truncation operation indicates that we only extract the first $E$ features, namely the features corresponding to the pair feature extraction query. Next, we use a mask cross-attention layer ( $\text{MaskCA}(\cdot)$ ), with $F_{SA}^{\text{feat}}$ as the query, $F_{Iseq}$ as key and value, and $m_{seq}$ as the mask, to extract features corresponding to the subject-object pair, formulated as + +$$ +F _ {C A} ^ {f e a t} = \operatorname {M a s k C A} \left(F _ {S A} ^ {f e a t}, F _ {I s e q}, m _ {s e q}\right). \tag {2} +$$ + +The features $F_{CA}^{feat}$ are further refined through a feed-forward network $(FFN(\cdot))$ , formulated as $F_{FFN}^{feat} = FFN(F_{CA}^{feat})$ . + +By repeating this process twice, we obtain the visual features for the subject-object pair to be input into the multimodal relation decoder, $F_{I}^{pair(i,j)} \in \mathbb{R}^{E \times D}$ . We perform these operations in parallel for all subject-object pairs to obtain the corresponding features for all pairs. + +Relation Existence Estimation Query. In addition to the pair feature extraction query, we also design a relation existence estimation query to determine whether a relation likely exists between the subject $o_i$ and object $o_j$ , without predicting the specific relation category. The objective is to filter out irrelevant subject-object pairs to save the computation for subsequent LMM decoding. + +Specifically, for each subject-object pair $(o_i,o_j)$ , the relation existence estimation query $Q^{exist}\in \mathbb{R}^{1\times D}$ , similar to the pair feature extraction query, is input into the self-attention, mask cross-attention, and feed-forward network layers, interacting respectively with $F_{Iseq}$ , $m_{seq}$ and the specially designed relation instruction. The purpose of the relation instruction is to direct the relation existence estimation query towards determining whether a relation likely exists in the subject-object pair, e.g. "Is there a relation between $o_i$ and $o_j$ ?". The relation instruction, after being processed by the tokenizer, results in $F_{Inst}^{exist}\in \mathbb{R}^{X^{exist}\times D}$ , where $X^{exist}$ represents the number of tokens. Eventually, the extracted features are input into a relation existence prediction layer, which includes a 2-layer MLP, and the predicted scores are normalized to [0, 1] using the sigmoid function. It is worth noting that we train it using binary labels indicating whether a relation exists between objects, and a selector module specified below is utilized to perform filtering during inference. + +Selector. The selector module implemented by 2-layer MLP is set to filter irrelevant subject-object pairs. Only those with a score higher than the threshold $\theta$ can be input into the multimodal relation decoder. Compared to predicting for all subject-object pairs, this can enables a $20\times$ speedup in our experiment. + +# 4.3 Multimodal Relation Decoder + +Multimodal relation decoder aims to utilize the subject-object pair feature $F_{I}^{i,j}$ extracted by the aforementioned modules, combined with an instruction guiding it to achieve open-set relation prediction. Inspired by [39, 41], we first design a generation instruction to perform open-set relation prediction in an autoregressive manner. This works well, yet we find that it tends to favor common relations more or less. Therefore, we further design a judgement instruction, leveraging the LMMs' strong analytical and judgment capabilities. The judgement instruction also utilizes an autoregressive manner but to judge whether a specific relation exists between objects, thereby simplifying the complexity of open-set relation prediction. Next, we specify the two instructions, respectively. + +Generation Instruction. For the generation instruction, we follow the instruction design used in open-set object recognition [39], utilizing "What are the relations between $c_i$ and $c_j$ ?" Here, $c_i$ and $c_j$ respectively refers to the name of the subject and object. We convert this instruction into features $F_{inst}^{gen} \in \mathbb{R}^{X^{gen} \times D}$ using the tokenizer, where $X^{gen}$ is the token number of the generation instruction. We input the features of this generation instruction $F_{inst}^{gen}$ together with the subject-object pair features $F_I^{pair(i,j)}$ into the multimodal relation decoder $Dec(\cdot)$ , predicting all possible relations in an autoregressive way, formulated as + +$$ +r _ {i, j} = \operatorname {D e c} \left(\operatorname {C o n c a t} \left(F _ {I} ^ {\text {p a i r} (i, j)}, F _ {\text {i n s t}} ^ {\text {g e n}}\right)\right). \tag {3} +$$ + +If multiple relations are predicted, they are separated by the delimiter "[SEP]". + +Judgement Instruction. Unlike generation instruction, the judgement instruction guides relation decoder to judge, based on a given relation name, whether this relation exists between the subject and object. For example, "Please judge between $c_i$ and $c_j$ whether there is a relation $r_k$ ". In this case, we only need the multimodal relation decoder to answer "Yes" or "No" to determine the existence of this relation. Note that inputting the complete judgement instruction for each relation into the decoder can be costly. Therefore, we place the relation name at the end of the instruction. During inference we divide the judgement instruction into two parts: the section before the relation name, transformed into $F_{inst}^{judge}$ through the tokenizer, and the relation name itself, processed into $F_{inst}^{rel}$ . Benefiting from the autoregressive manner for open-set relation prediction, we initially input the subject-object pair feature $F_I^{pair(i,j)}$ and $F_{inst}^{judge}$ into the multimodal relation decoder, formulated as + +$$ +F _ {p r e f i x} ^ {(i, j)} = D e c (C o n c a t \left(F _ {I} ^ {p a i r (i, j)}, F _ {i n s t} ^ {j u d g e}\right)), \tag {4} +$$ + +which is then cached for subsequent calculations for each relation. For each relation $r_k$ , the multimodal relation decoder only needs to process $F_{prefix}$ and $F_{inst}^{rel(k)}$ to achieve relation prediction, formulated as: + +$$ +J _ {i, j, k} = \operatorname {D e c} \left(\operatorname {C o n c a t} \left(F _ {\text {p r e f i x}} ^ {(i, j)}, F _ {\text {i n s t}} ^ {\text {r e l} (k)}\right)\right), \tag {5} +$$ + +where $J_{i,j,k}$ represents the judgement for $(o_i, r_k, o_j)$ triplet. When $J_{i,j,k}$ is "Yes", it indicates that the relation $r_k$ exists between $o_i$ and $o_j$ ; otherwise, it does not exist. Through this approach, we can maintain the same prediction time as with the generation instruction. + +We perform the aforementioned process for all subject-object pairs that may have a relation, ultimately achieving open-set relation prediction. For method using generation instruction, we denote it as OpenPSG-G, and for those using judgement instruction, as OpenPSG-J. In next section, OpenPSG by default refers to the latter. + +# 4.4 Loss Function + +During the model training, there are two losses involved: the binary cross-entropy loss $\mathcal{L}_{\text{exist}}$ for estimating the existence of a relation using the relation existence estimation query in the relation query transformer, and the cross-entropy loss $\mathcal{L}_{LM}$ consistent with language model training used by the multimodal relation decoder. The total loss is: $\mathcal{L} = \lambda \mathcal{L}_{\text{exist}} + \mathcal{L}_{LM}$ , where $\lambda$ is a weight factor. + +# 5 Experiments + +# 5.1 Datasets + +Panoptic Scene Graph (PSG) dataset [36] is constructed based on the COCO dataset [2, 22], consisting of 48,749 annotated images: 46,563 for training and 2,186 for testing. It encompasses 80 "thing" object categories [22] and 53 "stuff" object categories [2], as well as 56 relation categories. + +Visual Genome (VG) dataset [17] is a widely used dataset in the SGG task. To further validate our method, we follow previous works [5, 38] and test our method on the VG-150 variant, which contains 150 object categories and 50 relation categories. + +# 5.2 Tasks and metrics + +Tasks. In both PSG and SGG, there are three distinct subtasks: Predicate Classification (PredCls), Scene Graph Classification (SGCls), and Scene Graph Detection (SGDet) [34]. In PredCls, the categories and locations of objects within the image are given, and only the relation categories between the objects need to be predicted. SGCls requires predicting both the categories of objects and the relations between them, given the locations of objects within the image. SGDet requires the simultaneous prediction of object categories, locations, and relations between objects, based on the given image. In this paper, we focus on the PredCls and SGDet subtasks. The PredCls excludes the influence of segmentation performance and only compares the relation prediction performance, while the SGDet considers the combined results for both object segmentation and relation prediction. + +Furthermore, to validate our method's capability in open-set relation prediction, we divide the dataset into base relations and novel relations at a ratio of 7:3. For the PSG dataset, please refer to the supplementary material for the specific division method. For the division of the VG dataset, we follow the practice in previous works [5, 38]. In the open-set scenarios, our model is trained with data only from base relations and tested on both base and novel relations. It is noteworthy that the test sets for open-set and closed-set are same. + +Metrics. Following previous works [36, 47], we use Recall@K (R@K) and mean Recall@K (mR@K) as our evaluation metrics. Additionally, in open-set scenarios, we also report the R@K and mR@K metrics for base and novel relations. + +# 5.3 Implementation details + +In our experiments, we utilize the pretrained OpenSeeD [44] as the open-set object segmenter. The patch size $p$ of the patchify module is set to 8. Within the relation query transformer, the length $E$ of the pair feature extraction query is 32, and the threshold $\theta$ used to filter subject-object pairs is set to 0.35. In the multimodal relation decoder, we employ the decoder of BLIP-2 [18]. During model training, the weight factor $\lambda$ for the loss is set to 10. We adopt the same data augmentation strategies as in previous methods [36,47]. To train our model, we use the AdamW [26] optimizer with a learning rate of $1e^{-4}$ and a weight decay of $5e^{-2}$ . Our model is trained for a total of 12 epochs, reducing the learning rate to $1e^{-5}$ at the $8_{th}$ epoch. The experimental platform uses four A100 GPUs. Note that during training we freeze the parameters of the object segmenter and multimodal relation decoder but only train the proposed RelQ-Former. + +# 5.4 Comparison to the state of the art + +PSG dataset. Tab. 1 consists of two parts, comparing the performance of our method against previous methods in closed-set and open-set scenarios under the subtasks of predicate classification and scene graph detection. For the first part, in the closed-set scenario, our method significantly surpasses previous methods. For the predicate classification subtask, only methods that predict segmentation and relation separately are applicable [36], and our method achieves a substantial improvement compared to previous best results, for instance, a $26.6\%$ increase in R@100 and a $25.0\%$ increase in mR@100. For scene graph detection, our method also achieves a significant increase of $9.0\%$ over the best previous method [47] in R@100 and a $17.0\%$ increase in mR@100. This indicates that in the closed-set scenario, our method demonstrates significant performance improvements. For the second part, in the open-set scenario, we train the model only on base relations and test it on all relations. Notably, for predicate classification, our method even outperforms previous methods trained on all relations, which demonstrates the superiority of our method. For example, compared with previous best results, we achieve a $39\%$ improvement in R@100 and a $7.2\%$ improvement in mR@100. For the scene graph detection subtask, our method with judgement instruction remains very competitive compared to [36] trained on all relations. + +Table 1: Comparison between our OpenPSG and other methods on the PSG dataset in both closed-set and open-set scenarios. Our method shows superior performance compared to all previous methods. + +
MethodPredicate ClassificationScene Graph Detection
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Train on all relations (closed-set)
IMP [34]31.9/9.636.8/10.938.9/11.616.5/6.518.2/7.118.6/7.2
Motifs [43]44.9/20.250.4/22.152.4/22.920.0/9.121.7/9.622.0/9.7
VCTree [30]45.3/20.550.8/22.652.7/23.320.6/9.722.1/10.222.5/10.2
GPSNet [23]31.5/13.239.9/16.444.7/18.317.8/7.019.6/7.520.1/7.7
PSGTR [36]- / -- / -- / -28.4/16.634.4/20.836.3/22.1
PSGFormer [36]- / -- / -- / -18.0/14.819.6/17.020.1/17.6
ADTrans [19]- / 29.0- / 36.2- / 38.826.0/26.429.6/29.730.0/30.0
PairNet [33]- / -- / -- / -29.6/24.735.6/28.539.6/30.6
HiLo [47]- / -- / -- / -34.1/23.740.7/30.343.0/33.1
OpenPSG55.1/39.270.6/53.879.3/63.838.1/32.346.8/40.952.0/50.1
Train on base relations (open-set)
OpenPSG45.1/29.155.5/38.761.5/46.025.9/20.931.6/24.036.7/25.4
+ +Table 2: Comparison between our OpenPSG and other methods on the VG dataset in both closed-set and open-set scenarios with predicate classification subtask. + +
MethodTrain on all relations (closed-set)Train on base relations (open-set)
R/mR@50R/mR@100R/mR@50R/mR@100
Motifs [43]65.2/15.967.1/17.2- / -- / -
VCTree [30]66.4/16.868.1/19.4- / -- / -
Cacao+Epic [38]- /39.0- /40.8- /16.5- /21.8
OvSGTR [5]36.4/ -42.4/ -22.9/ -26.7/ -
OpenPSG60.2/45.871.4/50.325.7/21.530.6/27.2
+ +VG dataset. To further validate our method on the VG dataset, we compare to two closed-set SGG methods [30, 43] and two recent open-set SGG works [5, 38]. Since our method relies on an object segmentation model while [5, 30, 38, 43] rely on an object detector, for a fair comparison, we only present the results for predicate classification subtask. Tab. 2 shows the results in both closed-set and open-set scenarios. In the closed-set scenario, compared with previous closed-set methods [30, 43], our method only performs a few points lower on R@50, yet yields significantly better results on the other metrics. In addition, compared with previous best open-set methods [5, 38], our method improves by $29.0\%$ in R@100 and by $9.5\%$ in mR@100. In open-set scenario, our method improves upon [5, 38] by $3.9\%$ in R@100 and by $5.4\%$ in mR@100. These results demonstrate the effectiveness of our method in both closed-set and open-set relation prediction. + +Table 3: Ablation study of comparison of different segmenters. + +
SegmenterPredicate ClassificationScene Graph Detection
PQR/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
OpenSeeD [44]55.155.1/39.270.6/53.879.3/63.838.1/32.346.8/40.951.9/50.1
Mask2Former [6]51.754.8/39.170.4/52.978.9/63.736.1/30.344.8/38.448.4/47.8
+ +# 5.5 Ablation study + +To validate the effectiveness of each module in our method, we conduct ablation studies, using the model trained on all relations with judgement instruction for relation prediction as the baseline unless otherwise specified. To eliminate interference, we validate the object segmenter and modules in RelQ-Former under closed-set settings, and test two types of instructions in the multimodal relation decoder under open-set settings. + +Different segmenters. Our method is based on a pretrained segmenter. To verify the impact of different segmenters on model performance, we experiment with two options: the closed-set Mask2Former [6] and the open-set OpenSeeD [44]. As shown in Tab. 3, OpenSeeD outperforms Mask2Former on the Panoptic Quality (PQ) [16] metric (55.1 vs. 51.7) in PSG test set in closed-set scenario. For the predicate classification subtask, the results show that using OpenSeeD is slightly better than using Mask2Former, with R@100 higher by $0.4\%$ and mR@100 by $0.1\%$ . For the scene graph detection subtask, OpenSeeD is a better segmenter than Mask2Former, with an increase of $3.5\%$ in R@100 and $2.3\%$ in mR@100, due to its superior object segmentation capability (see PQ in Tab. 3). + +Subject-object pair features extraction. To validate the effectiveness of the RelQ-Former in extracting subject-object pair features via the attention mechanism, we design an experiment by extracting subject-object pair features using mask pooling. Specifically, for mask pooling, we derive object features by applying mask pooling to their respective positions in the visual features and then concatenate these to form the subject-object pair features. As shown in Tab. 4, it demonstrates that our attention mechanism outperforms mask pooling based pair feature extraction method by $5.2\%$ in R@100 and by $4.7\%$ in mR@100. This indicates that our method using RelQ-Former can better focus the extracted features on the interactions between objects, thereby enhancing the performance for relation prediction. + +Selector in RelQ-Former. To validate the efficacy of the selector, we set $\theta$ to 0, meaning that during inference, we predict relations for all subject-object pairs. As shown in Tab. 4, we find that when we set $\theta$ to 0, the model performance is approximately the same. However, under these conditions, the model takes 20 times longer to run on the whole PSG test set. This indicates that our selector allows for a significant improvement in computational efficiency with a small sacrifice in performance. For more details, please refer to supplementary material. + +Relation Existence Loss. To evaluate the impact of the relation existence estimation loss (Sec. 4.4) on the model, we set the training $\lambda$ to 0, thereby + +Table 4: Ablation study of comparison of different designs in RelQ-Former. + +
MethodPredicate Classification
R@20mR@20R@50mR@50R@100mR@100
RelQ-Former55.139.270.653.879.363.8
Change to Mask Pooling51.536.365.949.374.159.1
Set θ = 0 in Selector56.840.770.954.079.663.9
Set λ = 0 in Loss54.938.769.853.678.563.2
+ +Table 5: Ablation study of different instruction types in multimodal relation decoder under various base-to-novel relation ratios. OpenPSG-G: using generation instruction. OpenPSG-J: using judgement instruction. + +
base:novelOpenPSG-GOpenPSG-J
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
7:340.1/24.549.4/32.457.1/36.845.1/29.155.5/38.761.5/46.0
6:434.9/21.346.3/29.252.9/32.040.6/25.153.9/35.357.0/43.5
5:525.4/14.038.6/23.240.1/24.236.4/19.248.9/29.350.3/40.8
4:619.8/11.232.9/14.333.2/18.829.8/17.340.9/21.644.3/34.7
3:713.0/10.717.6/13.018.5/14.422.8/15.728.3/19.735.1/23.7
+ +removing this loss for model training. We discover that this loss not only aids in training a relation existence classifier, but also has beneficial effect on the model. As shown in Tab. 4, setting $\lambda$ to 0 leads to a decrease in R@100 by $0.8\%$ and in mR@100 by $0.6\%$ . + +Analysis of instruction types in multimodal relation decoder. To further analyze the two types of instructions, generation instruction and judgement instruction of open-set relation prediction in our multimodal relation decoder (Sec. 4.3), we conduct a series of experiments by adjusting the proportion of novel relation categories, conducting tests with base:novel ratios of 7:3, 6:4, 5:5, 4:6, and 3:7, to validate the performance, and results shown in Tab. 5. First, under the same base:novel ratio, the method using judgement instruction (OpenPSG-J) consistently outperforms the one using generation instruction (OpenPSG-G). For example, at a base:novel ratio of 7:3, OpenPSG-J is $4.4\%$ higher than OpenPSG-G in R@100, and is $9.2\%$ higher in mR@100. Second, the results indicate that as the proportion of novel relations increases, the performance of both OpenPSG-G and OpenPSG-G gradually decreases. For OpenPSG-G, R@100 decreases from $57.1\%$ at a base:novel ratio of 7:3 to $18.5\%$ at a base:novel ratio of 3:7, a drop of $38.6\%$ . OpenPSG-J's R@100 decreases by $26.4\%$ , which is less than the decrease for OpenPSG-G, indicating that OpenPSG-J has a stronger capability for relation prediction in an open world. + +![](images/945e0dec144194d49b19bc00ffea3edede98fc965df40486901bd91fbade8f3a.jpg) + +![](images/43f82896c0cf6d296dee8e0ed5f3a47527d2c0f25367b9a2d389b575251be0fb.jpg) + +![](images/ea5ff6e6ef660d009ed346d9ad8071d523a51e9916ae6d1a33c7e9fdc6169e85.jpg) + +![](images/71276d8d5b425ea14af5a23887aee4543350e8b7b6b6e5e1b23e5ff279cabd5c.jpg) +Image +Fig. 3: Visualization results produced by our OpenPSG. The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects. + +![](images/67a9e9e758a42e6dd13f367a4c4a22f0368a4062381af332bd2bdf374f9913e0.jpg) +Panoptic Segmentation + +![](images/440cd1caf0ba872cd3fa1b9e07d3a7bc114939b763e4cccb263ae2cff9f0193b.jpg) +Relation Prediction + +# 5.6 Visualization + +As shown in Fig. 3, our OpenPSG method can predict relations not defined in the dataset, such as "flying", "observing", and "fighting", which qualitatively demonstrates the excellent open-set relation prediction capability of our method. + +# 6 Conclusion + +In this paper, we propose the open-set PSG task and introduce the OpenPSG to accomplish open-set relation prediction. With the help of large multimodal models, our method employs an autoregressive approach to predict open-set relations. Additionally, we have developed a relation query transformer which contains pair feature extraction and relation existence estimation queries, one for extracting features of subject-object pairs, the other for predicting the existence of relations between them to filter out irrelevant pairs. Furthermore, we design generation and judgement instructions to enable the multimodal relation decoder to predict open-set relations. Extensive experiments demonstrate that our method achieves excellent performance in open-set relation prediction. In the future, we plan to employ model distillation to reduce the model size, thus enhancing the practicality in various real-world scenarios while ensuring its ability to predict open-set relations. + +# Acknowledgments + +The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the Fundamental Research Funds for the Central Universities. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2, 3, 4 +2. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018) 9 +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020) 3 +4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: CVPR (2020) 1 +5. Chen, Z., Wu, J., Lei, Z., Zhang, Z., Chen, C.: Expanding scene graph boundaries: Fully open-vocabulary scene graph generation via visual-concept alignment and retention. arXiv preprint arXiv:2311.10988 (2023) 3, 4, 9, 10, 11 +6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 3, 5, 12 +7. Dai, J., He, K., Sun, J.: Convolutional feature masking for joint object and stuff segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3992-4000 (2015) 6 +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 5 +9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR. pp. 14084-14093 (2022) 2, 4 +0. Gao, L., Wang, B., Wang, W.: Image captioning with scene-graph based semantic concepts. In: ICMLC (2018) 1 +1. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV. pp. 540-557. Springer (2022) 2 +2. He, T., Gao, L., Song, J., Li, Y.F.: Towards open-vocabulary scene graph generation with prompt-based finetuning. In: ECCV (2022) 3 +3. Hildebrandt, M., Li, H., Koner, R., Tresp, V., Gunnemann, S.: Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072 (2020) 1 +4. Joshi, M., Levy, O., Weld, D.S., Zettlemoyer, L.: Bert for coreference resolution: Baselines and analysis. arXiv preprint arXiv:1908.09091 (2019) 4 +5. Kan, X., Cui, H., Yang, C.: Zero-shot scene graph relation prediction through commonsense knowledge integration. In: Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part II 21. pp. 466-482. Springer (2021) 3 + +16. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019) 12 +17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision pp. 32-73 (2017) 9 +18. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 2, 4, 10 +19. Li, L., Ji, W., Wu, Y., Li, M., Qin, Y., Wei, L., Zimmermann, R.: Panoptic scene graph generation with semantics-prototype learning. arXiv preprint arXiv:2307.15567 (2023) 1, 3, 11 +20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR. pp. 7061-7070 (2023) 2 +21. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843 (2022) 2 +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 9 +23. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: CVPR (2020) 11 +24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) 4, 6 +25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS 36 (2024) 2, 4 +26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 10 +27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) 2, 4 +28. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 4 +29. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: ICCV (2023) 1 +30. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: CVPR (2019) 11 +31. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4 +32. Wang, C., Liu, X., Song, D.: Language models are open knowledge graphs. arXiv preprint arXiv:2010.11967 (2020) 4 +33. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023) 1, 2, 3, 11 +34. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: CVPR (2017) 9, 11 + +35. Yang, J., Wang, C., Liu, Z., Wu, J., Wang, D., Yang, L., Cao, X.: Focusing on flexible masks: A novel framework for panoptic scene graph generation with relation constraints. In: ACM MM. pp. 4209-4218 (2023) 1 +36. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation. In: ECCV. pp. 178-196 (2022) 1, 3, 9, 10, 11 +37. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: CVPR. pp. 23497-23506 (2023) 2 +38. Yu, Q., Li, J., Wu, Y., Tang, S., Ji, W., Zhuang, Y.: Visually-prompted language model for fine-grained scene graph generation in an open world. arXiv preprint arXiv:2303.13233 (2023) 3, 9, 10, 11 +39. Yu, Q., Shen, X., Chen, L.C.: Towards open-ended visual recognition with large language model. arXiv preprint arXiv:2311.08400 (2023) 2, 4, 8 +40. Yu, X., Chen, R., Li, J., Sun, J., Yuan, S., Ji, H., Lu, X., Wu, C.: Zero-shot scene graph generation with knowledge graph completion. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022) 3 +41. Yue, K., Chen, B.C., Geiping, J., Li, H., Goldstein, T., Lim, S.N.: Object recognition as next token prediction. arXiv preprint arXiv:2312.02142 (2023) 8 +42. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 2 +43. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR (2018) 11 +44. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV. pp. 1020-1031 (2023) 2, 5, 10, 12 +45. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Learning to generate language-supervised and open-vocabulary scene graph using pre-trained visual-semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2915–2924 (2023) 3 +46. Zhao, C., Shen, Y., Chen, Z., Ding, M., Gan, C.: Textpsg: Panoptic scene graph generation from textual descriptions. In: ICCV. pp. 2839-2850 (2023) 1, 3 +47. Zhou, Z., Shi, M., Caesar, H.: Hilo: Exploiting high low frequency relations for unbiased panoptic scene graph generation. arXiv preprint arXiv:2303.15994 (2023) 1, 2, 3, 10, 11 +48. Zhou, Z., Shi, M., Caesar, H.: Vlprompt: Vision-language prompting for panoptic scene graph generation. arXiv preprint arXiv:2311.16492 (2023) 1, 2, 3 +49. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 2 \ No newline at end of file diff --git a/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/images.zip b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84f9d105a82300998771b8ad36fc8eaf197ac983 --- /dev/null +++ b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b585509c49bcbe781685661e99ec76dae7def4fc60028387838832c72238ebce +size 435661 diff --git a/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/layout.json b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..696b3f25c69dfa5f574b996fba65b7202d3e7b13 --- /dev/null +++ b/2024/OpenPSG_ Open-set Panoptic Scene Graph Generation via Large Multimodal Models/layout.json @@ -0,0 +1,10399 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "OpenPSG: Open-set Panoptic Scene Graph Generation via Large Multimodal Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "spans": [ + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "text", + "content": "Zijian Zhou" + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "text", + "content": " Zheng Zhu" + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "text", + "content": " Holger Caesar" + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "text", + "content": " Miaojing Shi" + }, + { + "bbox": [ + 162, + 167, + 451, + 181 + ], + "type": "inline_equation", + "content": "^{4,5\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 177, + 189, + 436, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 189, + 436, + 201 + ], + "spans": [ + { + "bbox": [ + 177, + 189, + 436, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 177, + 189, + 436, + 201 + ], + "type": "text", + "content": "Department of Informatics, King's College London " + }, + { + "bbox": [ + 177, + 189, + 436, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 177, + 189, + 436, + 201 + ], + "type": "text", + "content": "GigaAI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 190, + 201, + 423, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 201, + 423, + 212 + ], + "spans": [ + { + "bbox": [ + 190, + 201, + 423, + 212 + ], + "type": "text", + "content": "3Intelligent Vehicles Lab, Delft University of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 212, + 462, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 212, + 462, + 235 + ], + "spans": [ + { + "bbox": [ + 151, + 212, + 462, + 235 + ], + "type": "text", + "content": "4College of Electronic and Information Engineering, Tongji University \n5Shanghai Institute of Intelligent Science and Technology, Tongji University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 266, + 455, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 266, + 455, + 498 + ], + "spans": [ + { + "bbox": [ + 159, + 266, + 455, + 498 + ], + "type": "text", + "content": "Abstract. Panoptic Scene Graph Generation (PSG) aims to segment objects and recognize their relations, enabling the structured understanding of an image. Previous methods focus on predicting predefined object and relation categories, hence limiting their applications in the open world scenarios. With the rapid development of large multimodal models (LMMs), significant progress has been made in open-set object detection and segmentation, yet open-set relation prediction in PSG remains unexplored. In this paper, we focus on the task of open-set relation prediction integrated with a pretrained open-set panoptic segmentation model to achieve true open-set panoptic scene graph generation (OpenPSG). Our OpenPSG leverages LMMs to achieve open-set relation prediction in an autoregressive manner. We introduce a relation query transformer to efficiently extract visual features of object pairs and estimate the existence of relations between them. The latter can enhance the prediction efficiency by filtering irrelevant pairs. Finally, we design the generation and judgement instructions to perform open-set relation prediction in PSG autoregressively. To our knowledge, we are the first to propose the open-set PSG task. Extensive experiments demonstrate that our method achieves state-of-the-art performance in open-set relation prediction and panoptic scene graph generation. Code is available at https://github.com/franciszzj/OpenPSG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 502, + 453, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 502, + 453, + 523 + ], + "spans": [ + { + "bbox": [ + 159, + 502, + 453, + 523 + ], + "type": "text", + "content": "Keywords: Panoptic Scene Graph Generation " + }, + { + "bbox": [ + 159, + 502, + 453, + 523 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 502, + 453, + 523 + ], + "type": "text", + "content": " Open-set " + }, + { + "bbox": [ + 159, + 502, + 453, + 523 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 502, + 453, + 523 + ], + "type": "text", + "content": " Large Multimodal Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 546, + 230, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 230, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 230, + 559 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "text", + "content": "Panoptic scene graph generation (PSG) [36] aims to segment objects within an image and recognize the relations among them, thereby constructing a panoptic scene graph for a structured understanding of the image. Given its significant potential in applications such as visual question answering [13], image captioning [4, 10], and embodied navigation [29], PSG has attracted considerable attention from researchers ever since its emerging [19, 33, 35, 46-48]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 245, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 245, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 245, + 665 + ], + "type": "inline_equation", + "content": "\\boxdot" + }, + { + "bbox": [ + 141, + 653, + 245, + 665 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 247, + 193 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 247, + 193 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 247, + 193 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 247, + 193 + ], + "type": "image", + "image_path": "39f4576820ef7e83836b9ae10496dfac3e2e398fbd2c3948dd76d7f41caf0287.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 193, + 199, + 201 + ], + "lines": [ + { + "bbox": [ + 182, + 193, + 199, + 201 + ], + "spans": [ + { + "bbox": [ + 182, + 193, + 199, + 201 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 211, + 482, + 268 + ], + "lines": [ + { + "bbox": [ + 131, + 211, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 482, + 268 + ], + "type": "text", + "content": "Fig. 1: The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects. Our method can predict both known (close-set) relations, e.g., (0_person, playing, 1_skateboard), (2_person, looking at, 0_person), and unknown (open-set) relations, e.g., (0_person, pop shove-it, 1_skateboard), (2_person, recording, 0_person)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 116, + 364, + 193 + ], + "blocks": [ + { + "bbox": [ + 250, + 116, + 364, + 193 + ], + "lines": [ + { + "bbox": [ + 250, + 116, + 364, + 193 + ], + "spans": [ + { + "bbox": [ + 250, + 116, + 364, + 193 + ], + "type": "image", + "image_path": "af3d444f853e1b60198c8c5d2a37ab60ad3336fc0bb50681d4adabee86c03426.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 193, + 334, + 201 + ], + "lines": [ + { + "bbox": [ + 279, + 193, + 334, + 201 + ], + "spans": [ + { + "bbox": [ + 279, + 193, + 334, + 201 + ], + "type": "text", + "content": "Panoptic Segmentation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 116, + 481, + 193 + ], + "blocks": [ + { + "bbox": [ + 366, + 116, + 481, + 193 + ], + "lines": [ + { + "bbox": [ + 366, + 116, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 366, + 116, + 481, + 193 + ], + "type": "image", + "image_path": "1f18b16b500d07235addb206562c29206da4aebacedb346cbbceb2399ca831fe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 193, + 447, + 201 + ], + "lines": [ + { + "bbox": [ + 399, + 193, + 447, + 201 + ], + "spans": [ + { + "bbox": [ + 399, + 193, + 447, + 201 + ], + "type": "text", + "content": "Relation Prediction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 291, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 388 + ], + "type": "text", + "content": "Previous PSG methods [33, 47, 48] are only capable of predicting closed-set object and relation categories while failing to recognize objects/relations beyond predefined categories. Recently, with the advent of large multimodal models (LMMs) such as CLIP [27], BLIP-2 [18] etc., a significant number of open-set prediction methods for object detection [9, 21, 37, 42, 44] and segmentation [11, 20, 39, 44] are introduced, attributing to LMMs' rich understanding of language and strong connections between vision and language. Nevertheless, open-set prediction of relations has been largely unexplored so far." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": "Compared to open-set object detection and segmentation, open-set relation prediction is more complex: the model is required to both understand different objects and recognize relations of object pairs based on their interactions; especially, the computation of the latter can be exponentially increased. To bridge the gap, in this paper, we focus on the open-set relation prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "LLMs [18, 25, 49] have demonstrated exceptional semantic analysis and understanding abilities across various multimodal tasks. In particular with the text processing, LMMs are not only good at interpreting on nouns (i.e. representing objects) but also pay considerable attention on predicates (i.e. representing relations between objects), ensuring their generated contents to be sufficiently coherent [1]. Inspired by this, we propose the Open-set Panoptic Scene Graph Generation architecture, OpenPSG, leveraging the capabilities of LMMs for open-set relation prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "To this end, we utilize a large multimodal model (e.g., BLIP-2 [18]) to achieve open-set relation prediction. Specifically, our model comprises three parts. First, the open-set panoptic segmenter, we adapt an existing model (e.g., OpenSeeD [44]) which is capable of extracting open-set object categories, masks, and visual features from the whole image, forming object pairs and pair masks. Second, the relation query transformer, which has two functions: extracting visual features of object pairs based on pair masks and with a special focus on pair interactions; judging the potential relations between object pairs. They are realized by two sets of queries, pair feature extraction query and relation existence estimation query. Only those object pairs that are judged to likely have" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "relations are fed into the third part, the multimodal relation decoder. This decoder directly inherits from the LMM to predict the open-set relations given an object pair in an auto-regressive manner, on condition of specifically-designed text instructions and pre-extracted pair visual features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 224 + ], + "type": "text", + "content": "To the best of our knowledge, we are the first to propose the task of open-set panoptic scene graph generation, enabling the open-set prediction of both object masks and relations. Extensive experiments demonstrate that our OpenPSG achieves state-of-the-art results in the closed-set setting and exhibits outstanding performance in the open-set setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 241, + 237, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 241, + 237, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 237, + 253 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 265, + 332, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 332, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 332, + 277 + ], + "type": "text", + "content": "2.1 Panoptic Scene Graph Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 511 + ], + "type": "text", + "content": "Panoptic scene graph generation stems from scene graph generation (SGG) by replacing bounding boxes with panoptic segmentation masks to represent the objects, so as to achieve a more comprehensive understanding of scenes. Following the introduction of PSG [36], a series of related works [19, 33, 36, 46-48] emerge, significantly advancing its performance. For example, Yang et al. [36], based on DETR [3], introduce an end-to-end framework with learnable queries to generate panoptic scene graphs. Wang et al. [33] design a pair proposal network to filter irrespective subject-object pairs, achieving performance improvement of the PSG. Subsequently, Zhou et al. [47] build a HiLo architecture based on Mask2Former [6] and devise separate branches for high- and low-frequency relations respectively, hence achieving an unbiased relation prediction method. Li et al. [19] re-balance the relation prediction by adaptively transferring information from high-frequency to low-frequency relations. Additionally, Zhao et al. [46] implement a weakly-supervised PSG method given only image-text pairs as annotations, allowing for the learning of panoptic scene graphs from image-level ground truth. Recently, Zhou et al. [48] leverage the rich language information inherent in large language models [1] and design effective image-text interaction modules to assist unbiased PSG. All these works are learned in the closed-set setting, in this paper, we study the open-set PSG, which has been unexplored." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 527, + 334, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 334, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 334, + 539 + ], + "type": "text", + "content": "2.2 Open-set Scene Graph Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Open-set SGG has been studied in recent years. Earlier works [15, 40], often termed as zero-shot SGG, focus on transferring the knowledge from known relations to unknown relations given their prior connections; for example, Kan et al. [15] leverage the external commonsense knowledge while Yu et al. [40] use knowledge graphs for the knowledge transfer. Subsequently, with the advancement of large multimodal models, various open-set SGG works [5, 12, 38, 45] have emerged; for example, He et al. [12] and Zhang et al. [45] focus on predicting relations between unknown objects in the SGG using multimodal models. Yu et al. [38] and Chen et al. [5] on the other hand focus on the open-set relation prediction in PSG, sharing the same aim with us. Specifically, Yu et al. [38] utilize" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 175 + ], + "type": "text", + "content": "the CLIP model to match visual features with textual relation features for open-set relation prediction; Chen et al. [5] employ a student-teacher network to align visual concepts in multimodal models for predicting open-set relations. In this paper, different from previous works, we introduce an auto-regressive method based on the LMM to achieve open-set relation prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 291, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 291, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 291, + 205 + ], + "type": "text", + "content": "2.3 Large Multimodal Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 213, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 380 + ], + "type": "text", + "content": "Since the introduction of large models like the GPT series [1], recent years have witnessed rapid development in large multimodal models [18, 25, 27]. Benefiting from its nature of connecting vision and language, LMMs have significantly advanced various downstream tasks, ranging from computer vision [9, 39] to natural language processing [14, 28, 32]. Early multimodal models, such as CLIP [27], trained on image-text paired datasets through contrastive learning to align the visual and textual information. Subsequently, owing to enlightenment of the autoregressive prediction in large language models [1, 31], there has been an explosive growth of LMMs [18, 24, 25]; by introducing mechanisms that can transform visual information into large language models, they facilitate the communication between visual and textual information. Furthermore, this has endowed LMMs with the capability to generate free text, leading to substantial improvements in numerous multimodal tasks. In this paper, we leverage LMMs to design a multimodal relation decoder to predict relations in an open-set scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 400, + 244, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 244, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 244, + 412 + ], + "type": "text", + "content": "3 Task Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "text", + "content": "We define the task, open-set panoptic scene graph generation. Given an image " + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "text", + "content": ", the objective of this task is to extract an open-set panoptic scene graph " + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "inline_equation", + "content": "G = \\{O, R\\}" + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "text", + "content": " from the image " + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 132, + 426, + 480, + 473 + ], + "type": "text", + "content": " are the height and width of the image. Here:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 481, + 479, + 575 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "O = \\{o_i\\}_{i=1}^N" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": " objects segmented from the image, each defined as " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "o_i = \\{c, m\\}" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": " is the object category that can belong to either predefined base object categories " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "C_{base}" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": " or undefined novel object categories " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "C_{novel}" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": " represents the binary mask in " + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\{0, 1\\}^{H \\times W}" + }, + { + "bbox": [ + 136, + 481, + 479, + 528 + ], + "type": "text", + "content": " of the object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "spans": [ + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "R = \\{r_{i,j} \\mid i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": " represents the relations between objects, where " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "r_{i,j}" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": " denotes the relation between " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": " as the subject and " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": " as the object. Each relation " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": " can belong to either predefined base relation categories " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "K_{base}" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": " or undefined novel relation categories " + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "inline_equation", + "content": "K_{novel}" + }, + { + "bbox": [ + 137, + 529, + 479, + 575 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 604, + 201, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 201, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 201, + 616 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 479, + 665 + ], + "type": "text", + "content": "As illustrated in Fig. 2, our OpenPSG comprises three components: object segmenter, relation query transformer (RelQ-Former), and multimodal relation decoder (RelDecoder). For the object segmenter (Sec. 4.1), we utilize a pretrained" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 251 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 251 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 251 + ], + "type": "image", + "image_path": "faeaf50c5ce4cd0274b1cf8fd14822d16869518baec9b9d0b087f96837661315.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 259, + 479, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 281 + ], + "type": "text", + "content": "Fig. 2: The overall framework of our OpenPSG, which comprises three components: object segmenter, relation query transformer and multimodal relation decoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 306, + 480, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 414 + ], + "type": "text", + "content": "open-set panoptic segmentation model to transform the input image into object categories and masks, as well as visual feature representing the whole image. Subsequently, we input the object categories, masks, and visual feature into the RelQ-Former (Sec. 4.2). Through two sets of learnable queries and complemented by designed instructions, we obtain visual features of object pairs compatible with LMMs' input format as well as judgements on potential relation existence. Finally, only those object pairs being judged to likely have a relation are sent into the RelDecoder (Sec. 4.3) for open-set relation prediction, ultimately yielding an open-set panoptic scene graph." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 433, + 251, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 251, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 251, + 445 + ], + "type": "text", + "content": "4.1 Object Segmenter" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": ", we utilize the pretrained open-set object segmenter (e.g., OpenSeeD [44]) to predict the objects " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": " within the image and the whole-image visual feature " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "F_{I} \\in \\mathbb{R}^{h \\times w \\times D}" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": " represent the height and width of " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": " denotes the feature dimension. The segmenter has a similar architecture to Mask2Former [6] including a pixel decoder. The whole-image visual feature " + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 130, + 454, + 480, + 550 + ], + "type": "text", + "content": " refers to the visual feature output by the pixel decoder. Below, we develop the patchify module and pairwise module to process the output of the segmenter, generating the input for RelQ-Former." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": "Patchify Module. Patchify module aims to serialize visual feature " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": " and object masks " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": ", enabling them to be processed as inputs by the RelQ-Former (Sec. 4.2). Similar to the input patchify layer of vision transformer (ViT) [8], we utilize a single convolution layer to transform the extracted " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": " into a sequence of visual tokens " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "F_{Iseq} \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": " is the number of patches and " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": " is the feature dimension. When the kernel size and stride of the convolution layer are both " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": " is calculated as " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "L = \\frac{h}{p} \\times \\frac{w}{p}" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": ". Simultaneously, we employ nearest neighbor interpolation to each extracted object's mask " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": ", where the size of " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "m_{i}" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "text", + "content": "is of height " + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "inline_equation", + "content": "\\frac{h}{p}" + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "inline_equation", + "content": "\\frac{w}{p}" + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "text", + "content": ", and then reshape it to a one-dimensional vector with the length " + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "text", + "content": ". After processing all masks in the same way, we obtain the mask sequence " + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "inline_equation", + "content": "m_{seq} \\in \\{0,1\\}^{N \\times L}" + }, + { + "bbox": [ + 132, + 115, + 480, + 153 + ], + "type": "text", + "content": " for all objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": "Pairwise Module. Pairwise module aims to construct subject-object pairs. Given " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " objects in the image " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": ", we pairwise all objects " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " into subject-object pairs " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "P = \\{(o_i, o_j) | i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": ". The number of subject-object pairs in " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "N \\times (N - 1)" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": ", which exhibits exponential growth as " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " increases. Consequently, we also obtain the combined subject-object pair category set " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "c^{pair} \\in \\{(c_i, c_j) | i, j \\in \\{1, 2, \\dots, N\\}, i \\neq j\\}" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": ". We construct the mask sequences for the two objects corresponding to the indices " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "m_{seq}" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " for each subject-object pair by using the logical " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "OR" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " operation. This operation is performed for all subject-object pairs, resulting in the pair mask sequence " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "m_{seq}^{pair} \\in \\{0, 1\\}^{N \\times (N - 1) \\times L}" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " for the subject-object pairs, where " + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 171, + 481, + 293 + ], + "type": "text", + "content": " is the number of patches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 312, + 303, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 303, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 303, + 324 + ], + "type": "text", + "content": "4.2 Relation Query Transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "text", + "content": "Relation query transformer, leveraging the obtained " + }, + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "inline_equation", + "content": "F_{Iseq}" + }, + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "inline_equation", + "content": "c^{pair}" + }, + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "inline_equation", + "content": "m_{seq}^{pair}" + }, + { + "bbox": [ + 132, + 333, + 480, + 393 + ], + "type": "text", + "content": ", employs two distinct types of queries, pair feature extraction query and relation existence estimation query, along with customized instructions. This approach facilitates the extraction of subject-object pair features and assesses which subject-object pairs likely have relations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 412, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 481, + 555 + ], + "type": "text", + "content": "Pair Feature Extraction Query. The objective of the pair feature extraction query is to extract corresponding subject-object pair features from the whole image visual feature based on the subject-object pair masks. A common extraction method involves mask pooling [7], which extracts features for the target subject-object pair, treating each area on the subject-object pair equally. However, for features used in relation prediction, they should focus more on areas where interactions between objects occur. By leveraging attention mechanisms, we facilitate interactions among visual tokens representing different areas within the visual feature sequence " + }, + { + "bbox": [ + 132, + 412, + 481, + 555 + ], + "type": "inline_equation", + "content": "F_{Iseq}" + }, + { + "bbox": [ + 132, + 412, + 481, + 555 + ], + "type": "text", + "content": " of a subject-object pair. This way can enhance areas that are crucial for relation predictions. Furthermore, inspired by [24], we design an instruction to assist the this learnable query in understanding its purpose for extracting subject-object pair features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": "Specifically, for each subject-object pair " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "(o_i,o_j)" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": ", we first input the pair feature extraction query " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "Q^{feat}\\in \\mathbb{R}^{E\\times D}" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": " into a self-attention layer " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "(SA(\\cdot))" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": ", along with the pair instruction designed specifically for the pair feature extraction query. This pair instruction is processed through a tokenizer layer to obtain " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "F_{Inst}^{feat}\\in \\mathbb{R}^{X^{feat}\\times D}" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": ", which specifies the function of the pair feature extraction query, namely \"Extracting subject-object " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "(c_i,c_j)" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": " features from visual features according to the mask\". Here " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": " is the token number of the pair feature extraction query, and " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "X^{feat}" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": " is the token number of the pair instruction. Note that we also incorporate the category names of the subject and object " + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "inline_equation", + "content": "(c_i,c_j)" + }, + { + "bbox": [ + 132, + 556, + 481, + 666 + ], + "type": "text", + "content": " into this" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 346, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 346, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 346, + 128 + ], + "type": "text", + "content": "pair instruction. This operation is formulated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 140, + 481, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 140, + 481, + 156 + ], + "spans": [ + { + "bbox": [ + 206, + 140, + 481, + 156 + ], + "type": "interline_equation", + "content": "F _ {S A} ^ {\\text {f e a t}} = \\operatorname {T r u n c} \\left(S A \\left(\\operatorname {C o n c a t} \\left(Q ^ {\\text {f e a t}}, F _ {\\text {I n s t}} ^ {\\text {f e a t}}\\right)\\right), E\\right), \\tag {1}", + "image_path": "e88f2ea8d1160ad1e76ee4805970da66a84e064167522e70d94e61185c8f3146.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\text{Concat}(\\cdot)" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": " denotes the concatenation operation, " + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\text{Trunc}(\\cdot)" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": " represents the truncation operation, and " + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": " in this truncation operation indicates that we only extract the first " + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": " features, namely the features corresponding to the pair feature extraction query. Next, we use a mask cross-attention layer (" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\text{MaskCA}(\\cdot)" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": "), with " + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "F_{SA}^{\\text{feat}}" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": " as the query, " + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "F_{Iseq}" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": " as key and value, and " + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "inline_equation", + "content": "m_{seq}" + }, + { + "bbox": [ + 132, + 167, + 482, + 240 + ], + "type": "text", + "content": " as the mask, to extract features corresponding to the subject-object pair, formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 224, + 253, + 481, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 253, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 224, + 253, + 481, + 268 + ], + "type": "interline_equation", + "content": "F _ {C A} ^ {f e a t} = \\operatorname {M a s k C A} \\left(F _ {S A} ^ {f e a t}, F _ {I s e q}, m _ {s e q}\\right). \\tag {2}", + "image_path": "37a1f47e8706167fe618fc305a22c6e048d3ebf2ba95a6f7359ad2a6df24c062.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "text", + "content": "The features " + }, + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "inline_equation", + "content": "F_{CA}^{feat}" + }, + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "text", + "content": " are further refined through a feed-forward network " + }, + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "inline_equation", + "content": "(FFN(\\cdot))" + }, + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "text", + "content": ", formulated as " + }, + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "inline_equation", + "content": "F_{FFN}^{feat} = FFN(F_{CA}^{feat})" + }, + { + "bbox": [ + 132, + 280, + 480, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 357 + ], + "type": "text", + "content": "By repeating this process twice, we obtain the visual features for the subject-object pair to be input into the multimodal relation decoder, " + }, + { + "bbox": [ + 132, + 308, + 481, + 357 + ], + "type": "inline_equation", + "content": "F_{I}^{pair(i,j)} \\in \\mathbb{R}^{E \\times D}" + }, + { + "bbox": [ + 132, + 308, + 481, + 357 + ], + "type": "text", + "content": ". We perform these operations in parallel for all subject-object pairs to obtain the corresponding features for all pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 384, + 481, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 481, + 444 + ], + "type": "text", + "content": "Relation Existence Estimation Query. In addition to the pair feature extraction query, we also design a relation existence estimation query to determine whether a relation likely exists between the subject " + }, + { + "bbox": [ + 132, + 384, + 481, + 444 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 132, + 384, + 481, + 444 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 132, + 384, + 481, + 444 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 132, + 384, + 481, + 444 + ], + "type": "text", + "content": ", without predicting the specific relation category. The objective is to filter out irrelevant subject-object pairs to save the computation for subsequent LMM decoding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": "Specifically, for each subject-object pair " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "(o_i,o_j)" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": ", the relation existence estimation query " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "Q^{exist}\\in \\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": ", similar to the pair feature extraction query, is input into the self-attention, mask cross-attention, and feed-forward network layers, interacting respectively with " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "F_{Iseq}" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "m_{seq}" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": " and the specially designed relation instruction. The purpose of the relation instruction is to direct the relation existence estimation query towards determining whether a relation likely exists in the subject-object pair, e.g. \"Is there a relation between " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": "?\". The relation instruction, after being processed by the tokenizer, results in " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "F_{Inst}^{exist}\\in \\mathbb{R}^{X^{exist}\\times D}" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "inline_equation", + "content": "X^{exist}" + }, + { + "bbox": [ + 132, + 447, + 482, + 616 + ], + "type": "text", + "content": " represents the number of tokens. Eventually, the extracted features are input into a relation existence prediction layer, which includes a 2-layer MLP, and the predicted scores are normalized to [0, 1] using the sigmoid function. It is worth noting that we train it using binary labels indicating whether a relation exists between objects, and a selector module specified below is utilized to perform filtering during inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": "Selector. The selector module implemented by 2-layer MLP is set to filter irrelevant subject-object pairs. Only those with a score higher than the threshold " + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": " can be input into the multimodal relation decoder. Compared to predicting for all subject-object pairs, this can enables a " + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": " speedup in our experiment." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 312, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 312, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 312, + 127 + ], + "type": "text", + "content": "4.3 Multimodal Relation Decoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": "Multimodal relation decoder aims to utilize the subject-object pair feature " + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "inline_equation", + "content": "F_{I}^{i,j}" + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": " extracted by the aforementioned modules, combined with an instruction guiding it to achieve open-set relation prediction. Inspired by [39, 41], we first design a generation instruction to perform open-set relation prediction in an autoregressive manner. This works well, yet we find that it tends to favor common relations more or less. Therefore, we further design a judgement instruction, leveraging the LMMs' strong analytical and judgment capabilities. The judgement instruction also utilizes an autoregressive manner but to judge whether a specific relation exists between objects, thereby simplifying the complexity of open-set relation prediction. Next, we specify the two instructions, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": "Generation Instruction. For the generation instruction, we follow the instruction design used in open-set object recognition [39], utilizing \"What are the relations between " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": "?\" Here, " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": " respectively refers to the name of the subject and object. We convert this instruction into features " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "F_{inst}^{gen} \\in \\mathbb{R}^{X^{gen} \\times D}" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": " using the tokenizer, where " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "X^{gen}" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": " is the token number of the generation instruction. We input the features of this generation instruction " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "F_{inst}^{gen}" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": " together with the subject-object pair features " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "F_I^{pair(i,j)}" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": " into the multimodal relation decoder " + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "inline_equation", + "content": "Dec(\\cdot)" + }, + { + "bbox": [ + 131, + 269, + 482, + 368 + ], + "type": "text", + "content": ", predicting all possible relations in an autoregressive way, formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 373, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 373, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 225, + 373, + 481, + 389 + ], + "type": "interline_equation", + "content": "r _ {i, j} = \\operatorname {D e c} \\left(\\operatorname {C o n c a t} \\left(F _ {I} ^ {\\text {p a i r} (i, j)}, F _ {\\text {i n s t}} ^ {\\text {g e n}}\\right)\\right). \\tag {3}", + "image_path": "0893a9a132f6e26a2ab95a8d4c56c0eeaeaf9f8acb7095445cc21cd563cfc9e0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 394, + 478, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 478, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 478, + 407 + ], + "type": "text", + "content": "If multiple relations are predicted, they are separated by the delimiter \"[SEP]\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": "Judgement Instruction. Unlike generation instruction, the judgement instruction guides relation decoder to judge, based on a given relation name, whether this relation exists between the subject and object. For example, \"Please judge between " + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": " whether there is a relation " + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": "\". In this case, we only need the multimodal relation decoder to answer \"Yes\" or \"No\" to determine the existence of this relation. Note that inputting the complete judgement instruction for each relation into the decoder can be costly. Therefore, we place the relation name at the end of the instruction. During inference we divide the judgement instruction into two parts: the section before the relation name, transformed into " + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "inline_equation", + "content": "F_{inst}^{judge}" + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": " through the tokenizer, and the relation name itself, processed into " + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "inline_equation", + "content": "F_{inst}^{rel}" + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": ". Benefiting from the autoregressive manner for open-set relation prediction, we initially input the subject-object pair feature " + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "inline_equation", + "content": "F_I^{pair(i,j)}" + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "inline_equation", + "content": "F_{inst}^{judge}" + }, + { + "bbox": [ + 131, + 422, + 482, + 579 + ], + "type": "text", + "content": " into the multimodal relation decoder, formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 584, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 584, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 212, + 584, + 481, + 602 + ], + "type": "interline_equation", + "content": "F _ {p r e f i x} ^ {(i, j)} = D e c (C o n c a t \\left(F _ {I} ^ {p a i r (i, j)}, F _ {i n s t} ^ {j u d g e}\\right)), \\tag {4}", + "image_path": "ae9d31119c9b4f32cdd7d2d29f4492ae0efe060ab7d19b1179da971e8f147a98.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "text", + "content": "which is then cached for subsequent calculations for each relation. For each relation " + }, + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "text", + "content": ", the multimodal relation decoder only needs to process " + }, + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "inline_equation", + "content": "F_{prefix}" + }, + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "inline_equation", + "content": "F_{inst}^{rel(k)}" + }, + { + "bbox": [ + 131, + 606, + 482, + 647 + ], + "type": "text", + "content": " to achieve relation prediction, formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 220, + 651, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 651, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 220, + 651, + 481, + 668 + ], + "type": "interline_equation", + "content": "J _ {i, j, k} = \\operatorname {D e c} \\left(\\operatorname {C o n c a t} \\left(F _ {\\text {p r e f i x}} ^ {(i, j)}, F _ {\\text {i n s t}} ^ {\\text {r e l} (k)}\\right)\\right), \\tag {5}", + "image_path": "667e8091772fa7929c26ba80116dc9ce206feac83d206b2cf7ecce2cd12a8413.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 310, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 310, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 310, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "J_{i,j,k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " represents the judgement for " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "(o_i, r_k, o_j)" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " triplet. When " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "J_{i,j,k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " is \"Yes\", it indicates that the relation " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " exists between " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "; otherwise, it does not exist. Through this approach, we can maintain the same prediction time as with the generation instruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 479, + 224 + ], + "type": "text", + "content": "We perform the aforementioned process for all subject-object pairs that may have a relation, ultimately achieving open-set relation prediction. For method using generation instruction, we denote it as OpenPSG-G, and for those using judgement instruction, as OpenPSG-J. In next section, OpenPSG by default refers to the latter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 243, + 231, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 231, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 231, + 254 + ], + "type": "text", + "content": "4.4 Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "text", + "content": "During the model training, there are two losses involved: the binary cross-entropy loss " + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{exist}}" + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "text", + "content": " for estimating the existence of a relation using the relation existence estimation query in the relation query transformer, and the cross-entropy loss " + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LM}" + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "text", + "content": " consistent with language model training used by the multimodal relation decoder. The total loss is: " + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\lambda \\mathcal{L}_{\\text{exist}} + \\mathcal{L}_{LM}" + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 262, + 479, + 323 + ], + "type": "text", + "content": " is a weight factor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 342, + 230, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 230, + 355 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 230, + 355 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 367, + 205, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 367, + 205, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 367, + 205, + 378 + ], + "type": "text", + "content": "5.1 Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 387, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 479, + 434 + ], + "type": "text", + "content": "Panoptic Scene Graph (PSG) dataset [36] is constructed based on the COCO dataset [2, 22], consisting of 48,749 annotated images: 46,563 for training and 2,186 for testing. It encompasses 80 \"thing\" object categories [22] and 53 \"stuff\" object categories [2], as well as 56 relation categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 435, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 479, + 483 + ], + "type": "text", + "content": "Visual Genome (VG) dataset [17] is a widely used dataset in the SGG task. To further validate our method, we follow previous works [5, 38] and test our method on the VG-150 variant, which contains 150 object categories and 50 relation categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 502, + 252, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 252, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 252, + 513 + ], + "type": "text", + "content": "5.2 Tasks and metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 665 + ], + "type": "text", + "content": "Tasks. In both PSG and SGG, there are three distinct subtasks: Predicate Classification (PredCls), Scene Graph Classification (SGCls), and Scene Graph Detection (SGDet) [34]. In PredCls, the categories and locations of objects within the image are given, and only the relation categories between the objects need to be predicted. SGCls requires predicting both the categories of objects and the relations between them, given the locations of objects within the image. SGDet requires the simultaneous prediction of object categories, locations, and relations between objects, based on the given image. In this paper, we focus on the PredCls and SGDet subtasks. The PredCls excludes the influence of segmentation performance and only compares the relation prediction performance, while the SGDet considers the combined results for both object segmentation and relation prediction." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "Furthermore, to validate our method's capability in open-set relation prediction, we divide the dataset into base relations and novel relations at a ratio of 7:3. For the PSG dataset, please refer to the supplementary material for the specific division method. For the division of the VG dataset, we follow the practice in previous works [5, 38]. In the open-set scenarios, our model is trained with data only from base relations and tested on both base and novel relations. It is noteworthy that the test sets for open-set and closed-set are same." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 235 + ], + "type": "text", + "content": "Metrics. Following previous works [36, 47], we use Recall@K (R@K) and mean Recall@K (mR@K) as our evaluation metrics. Additionally, in open-set scenarios, we also report the R@K and mR@K metrics for base and novel relations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 252, + 277, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 252, + 277, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 277, + 264 + ], + "type": "text", + "content": "5.3 Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": "In our experiments, we utilize the pretrained OpenSeeD [44] as the open-set object segmenter. The patch size " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": " of the patchify module is set to 8. Within the relation query transformer, the length " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": " of the pair feature extraction query is 32, and the threshold " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": " used to filter subject-object pairs is set to 0.35. In the multimodal relation decoder, we employ the decoder of BLIP-2 [18]. During model training, the weight factor " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": " for the loss is set to 10. We adopt the same data augmentation strategies as in previous methods [36,47]. To train our model, we use the AdamW [26] optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "5e^{-2}" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": ". Our model is trained for a total of 12 epochs, reducing the learning rate to " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "inline_equation", + "content": "8_{th}" + }, + { + "bbox": [ + 130, + 271, + 479, + 415 + ], + "type": "text", + "content": " epoch. The experimental platform uses four A100 GPUs. Note that during training we freeze the parameters of the object segmenter and multimodal relation decoder but only train the proposed RelQ-Former." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 431, + 334, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 334, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 334, + 443 + ], + "type": "text", + "content": "5.4 Comparison to the state of the art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "content": "PSG dataset. Tab. 1 consists of two parts, comparing the performance of our method against previous methods in closed-set and open-set scenarios under the subtasks of predicate classification and scene graph detection. For the first part, in the closed-set scenario, our method significantly surpasses previous methods. For the predicate classification subtask, only methods that predict segmentation and relation separately are applicable [36], and our method achieves a substantial improvement compared to previous best results, for instance, a " + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "inline_equation", + "content": "26.6\\%" + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "content": " increase in R@100 and a " + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "inline_equation", + "content": "25.0\\%" + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "content": " increase in mR@100. For scene graph detection, our method also achieves a significant increase of " + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "inline_equation", + "content": "9.0\\%" + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "content": " over the best previous method [47] in R@100 and a " + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "inline_equation", + "content": "17.0\\%" + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "content": " increase in mR@100. This indicates that in the closed-set scenario, our method demonstrates significant performance improvements. For the second part, in the open-set scenario, we train the model only on base relations and test it on all relations. Notably, for predicate classification, our method even outperforms previous methods trained on all relations, which demonstrates the superiority of our method. For example, compared with previous best results, we achieve a " + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "content": " improvement in R@100 and a " + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 130, + 450, + 479, + 665 + ], + "type": "text", + "content": " improvement in mR@100. For the scene graph detection subtask, our method with judgement instruction remains very competitive compared to [36] trained on all relations." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 479, + 353 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison between our OpenPSG and other methods on the PSG dataset in both closed-set and open-set scenarios. Our method shows superior performance compared to all previous methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 479, + 353 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 479, + 353 + ], + "type": "table", + "html": "
MethodPredicate ClassificationScene Graph Detection
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Train on all relations (closed-set)
IMP [34]31.9/9.636.8/10.938.9/11.616.5/6.518.2/7.118.6/7.2
Motifs [43]44.9/20.250.4/22.152.4/22.920.0/9.121.7/9.622.0/9.7
VCTree [30]45.3/20.550.8/22.652.7/23.320.6/9.722.1/10.222.5/10.2
GPSNet [23]31.5/13.239.9/16.444.7/18.317.8/7.019.6/7.520.1/7.7
PSGTR [36]- / -- / -- / -28.4/16.634.4/20.836.3/22.1
PSGFormer [36]- / -- / -- / -18.0/14.819.6/17.020.1/17.6
ADTrans [19]- / 29.0- / 36.2- / 38.826.0/26.429.6/29.730.0/30.0
PairNet [33]- / -- / -- / -29.6/24.735.6/28.539.6/30.6
HiLo [47]- / -- / -- / -34.1/23.740.7/30.343.0/33.1
OpenPSG55.1/39.270.6/53.879.3/63.838.1/32.346.8/40.952.0/50.1
Train on base relations (open-set)
OpenPSG45.1/29.155.5/38.761.5/46.025.9/20.931.6/24.036.7/25.4
", + "image_path": "5f2e79004ffbe3f759de3159ab268baf6d7d40b0590c50f199ad4b73ec8d56dc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 151, + 399, + 460, + 485 + ], + "blocks": [ + { + "bbox": [ + 131, + 366, + 479, + 388 + ], + "lines": [ + { + "bbox": [ + 131, + 366, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 479, + 388 + ], + "type": "text", + "content": "Table 2: Comparison between our OpenPSG and other methods on the VG dataset in both closed-set and open-set scenarios with predicate classification subtask." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 399, + 460, + 485 + ], + "lines": [ + { + "bbox": [ + 151, + 399, + 460, + 485 + ], + "spans": [ + { + "bbox": [ + 151, + 399, + 460, + 485 + ], + "type": "table", + "html": "
MethodTrain on all relations (closed-set)Train on base relations (open-set)
R/mR@50R/mR@100R/mR@50R/mR@100
Motifs [43]65.2/15.967.1/17.2- / -- / -
VCTree [30]66.4/16.868.1/19.4- / -- / -
Cacao+Epic [38]- /39.0- /40.8- /16.5- /21.8
OvSGTR [5]36.4/ -42.4/ -22.9/ -26.7/ -
OpenPSG60.2/45.871.4/50.325.7/21.530.6/27.2
", + "image_path": "be488ea68c5071fc26d3dd32e8760236668cdfedc2e7c63f4e3ff2ec2b998686.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "VG dataset. To further validate our method on the VG dataset, we compare to two closed-set SGG methods [30, 43] and two recent open-set SGG works [5, 38]. Since our method relies on an object segmentation model while [5, 30, 38, 43] rely on an object detector, for a fair comparison, we only present the results for predicate classification subtask. Tab. 2 shows the results in both closed-set and open-set scenarios. In the closed-set scenario, compared with previous closed-set methods [30, 43], our method only performs a few points lower on R@50, yet yields significantly better results on the other metrics. In addition, compared with previous best open-set methods [5, 38], our method improves by " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "29.0\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " in R@100 and by " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " in mR@100. In open-set scenario, our method improves upon [5, 38] by " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " in R@100 and by " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " in mR@100. These results demonstrate the effectiveness of our method in both closed-set and open-set relation prediction." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 136, + 479, + 195 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "type": "text", + "content": "Table 3: Ablation study of comparison of different segmenters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 136, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 135, + 136, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 135, + 136, + 479, + 195 + ], + "type": "table", + "html": "
SegmenterPredicate ClassificationScene Graph Detection
PQR/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
OpenSeeD [44]55.155.1/39.270.6/53.879.3/63.838.1/32.346.8/40.951.9/50.1
Mask2Former [6]51.754.8/39.170.4/52.978.9/63.736.1/30.344.8/38.448.4/47.8
", + "image_path": "9b6443fb734edc22a9575f772882d6705261c7e9c5b9d8aac36e9d1855dbc18c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 236, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 236, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 236, + 228 + ], + "type": "text", + "content": "5.5 Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 479, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 479, + 305 + ], + "type": "text", + "content": "To validate the effectiveness of each module in our method, we conduct ablation studies, using the model trained on all relations with judgement instruction for relation prediction as the baseline unless otherwise specified. To eliminate interference, we validate the object segmenter and modules in RelQ-Former under closed-set settings, and test two types of instructions in the multimodal relation decoder under open-set settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "text", + "content": "Different segmenters. Our method is based on a pretrained segmenter. To verify the impact of different segmenters on model performance, we experiment with two options: the closed-set Mask2Former [6] and the open-set OpenSeeD [44]. As shown in Tab. 3, OpenSeeD outperforms Mask2Former on the Panoptic Quality (PQ) [16] metric (55.1 vs. 51.7) in PSG test set in closed-set scenario. For the predicate classification subtask, the results show that using OpenSeeD is slightly better than using Mask2Former, with R@100 higher by " + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "text", + "content": " and mR@100 by " + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "text", + "content": ". For the scene graph detection subtask, OpenSeeD is a better segmenter than Mask2Former, with an increase of " + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "text", + "content": " in R@100 and " + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 130, + 307, + 480, + 426 + ], + "type": "text", + "content": " in mR@100, due to its superior object segmentation capability (see PQ in Tab. 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "text", + "content": "Subject-object pair features extraction. To validate the effectiveness of the RelQ-Former in extracting subject-object pair features via the attention mechanism, we design an experiment by extracting subject-object pair features using mask pooling. Specifically, for mask pooling, we derive object features by applying mask pooling to their respective positions in the visual features and then concatenate these to form the subject-object pair features. As shown in Tab. 4, it demonstrates that our attention mechanism outperforms mask pooling based pair feature extraction method by " + }, + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "text", + "content": " in R@100 and by " + }, + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "text", + "content": " in mR@100. This indicates that our method using RelQ-Former can better focus the extracted features on the interactions between objects, thereby enhancing the performance for relation prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 641 + ], + "type": "text", + "content": "Selector in RelQ-Former. To validate the efficacy of the selector, we set " + }, + { + "bbox": [ + 130, + 558, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 558, + 481, + 641 + ], + "type": "text", + "content": " to 0, meaning that during inference, we predict relations for all subject-object pairs. As shown in Tab. 4, we find that when we set " + }, + { + "bbox": [ + 130, + 558, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 558, + 481, + 641 + ], + "type": "text", + "content": " to 0, the model performance is approximately the same. However, under these conditions, the model takes 20 times longer to run on the whole PSG test set. This indicates that our selector allows for a significant improvement in computational efficiency with a small sacrifice in performance. For more details, please refer to supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Relation Existence Loss. To evaluate the impact of the relation existence estimation loss (Sec. 4.4) on the model, we set the training " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " to 0, thereby" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 136, + 443, + 216 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 461, + 126 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 461, + 126 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 461, + 126 + ], + "type": "text", + "content": "Table 4: Ablation study of comparison of different designs in RelQ-Former." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 136, + 443, + 216 + ], + "lines": [ + { + "bbox": [ + 168, + 136, + 443, + 216 + ], + "spans": [ + { + "bbox": [ + 168, + 136, + 443, + 216 + ], + "type": "table", + "html": "
MethodPredicate Classification
R@20mR@20R@50mR@50R@100mR@100
RelQ-Former55.139.270.653.879.363.8
Change to Mask Pooling51.536.365.949.374.159.1
Set θ = 0 in Selector56.840.770.954.079.663.9
Set λ = 0 in Loss54.938.769.853.678.563.2
", + "image_path": "cf4e5fae65723027f38a5b0c5c72f15a22df0b31536858288efabe3c65248c2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 151, + 286, + 459, + 378 + ], + "blocks": [ + { + "bbox": [ + 132, + 243, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 243, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 480, + 277 + ], + "type": "text", + "content": "Table 5: Ablation study of different instruction types in multimodal relation decoder under various base-to-novel relation ratios. OpenPSG-G: using generation instruction. OpenPSG-J: using judgement instruction." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 286, + 459, + 378 + ], + "lines": [ + { + "bbox": [ + 151, + 286, + 459, + 378 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 459, + 378 + ], + "type": "table", + "html": "
base:novelOpenPSG-GOpenPSG-J
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
7:340.1/24.549.4/32.457.1/36.845.1/29.155.5/38.761.5/46.0
6:434.9/21.346.3/29.252.9/32.040.6/25.153.9/35.357.0/43.5
5:525.4/14.038.6/23.240.1/24.236.4/19.248.9/29.350.3/40.8
4:619.8/11.232.9/14.333.2/18.829.8/17.340.9/21.644.3/34.7
3:713.0/10.717.6/13.018.5/14.422.8/15.728.3/19.735.1/23.7
", + "image_path": "732c9dc4181813aa11d579d5683653b3441d3576a1c154a174a3d616f2c155cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "text", + "content": "removing this loss for model training. We discover that this loss not only aids in training a relation existence classifier, but also has beneficial effect on the model. As shown in Tab. 4, setting " + }, + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "text", + "content": " to 0 leads to a decrease in R@100 by " + }, + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "text", + "content": " and in mR@100 by " + }, + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 418, + 480, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Analysis of instruction types in multimodal relation decoder. To further analyze the two types of instructions, generation instruction and judgement instruction of open-set relation prediction in our multimodal relation decoder (Sec. 4.3), we conduct a series of experiments by adjusting the proportion of novel relation categories, conducting tests with base:novel ratios of 7:3, 6:4, 5:5, 4:6, and 3:7, to validate the performance, and results shown in Tab. 5. First, under the same base:novel ratio, the method using judgement instruction (OpenPSG-J) consistently outperforms the one using generation instruction (OpenPSG-G). For example, at a base:novel ratio of 7:3, OpenPSG-J is " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " higher than OpenPSG-G in R@100, and is " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " higher in mR@100. Second, the results indicate that as the proportion of novel relations increases, the performance of both OpenPSG-G and OpenPSG-G gradually decreases. For OpenPSG-G, R@100 decreases from " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " at a base:novel ratio of 7:3 to " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "18.5\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " at a base:novel ratio of 3:7, a drop of " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "38.6\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": ". OpenPSG-J's R@100 decreases by " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "26.4\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": ", which is less than the decrease for OpenPSG-G, indicating that OpenPSG-J has a stronger capability for relation prediction in an open world." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 248, + 193 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 248, + 193 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 248, + 193 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 248, + 193 + ], + "type": "image", + "image_path": "945e0dec144194d49b19bc00ffea3edede98fc965df40486901bd91fbade8f3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 117, + 363, + 194 + ], + "blocks": [ + { + "bbox": [ + 250, + 117, + 363, + 194 + ], + "lines": [ + { + "bbox": [ + 250, + 117, + 363, + 194 + ], + "spans": [ + { + "bbox": [ + 250, + 117, + 363, + 194 + ], + "type": "image", + "image_path": "43f82896c0cf6d296dee8e0ed5f3a47527d2c0f25367b9a2d389b575251be0fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 365, + 118, + 479, + 193 + ], + "blocks": [ + { + "bbox": [ + 365, + 118, + 479, + 193 + ], + "lines": [ + { + "bbox": [ + 365, + 118, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 365, + 118, + 479, + 193 + ], + "type": "image", + "image_path": "ea5ff6e6ef660d009ed346d9ad8071d523a51e9916ae6d1a33c7e9fdc6169e85.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 196, + 248, + 277 + ], + "blocks": [ + { + "bbox": [ + 135, + 196, + 248, + 277 + ], + "lines": [ + { + "bbox": [ + 135, + 196, + 248, + 277 + ], + "spans": [ + { + "bbox": [ + 135, + 196, + 248, + 277 + ], + "type": "image", + "image_path": "71276d8d5b425ea14af5a23887aee4543350e8b7b6b6e5e1b23e5ff279cabd5c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 277, + 200, + 285 + ], + "lines": [ + { + "bbox": [ + 182, + 277, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 182, + 277, + 200, + 285 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 297, + 480, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 480, + 331 + ], + "type": "text", + "content": "Fig. 3: Visualization results produced by our OpenPSG. The left image is the input to our OpenPSG, the middle one displays the panoptic segmentation result, and the right one shows the predicted relations between objects." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 250, + 196, + 363, + 277 + ], + "blocks": [ + { + "bbox": [ + 250, + 196, + 363, + 277 + ], + "lines": [ + { + "bbox": [ + 250, + 196, + 363, + 277 + ], + "spans": [ + { + "bbox": [ + 250, + 196, + 363, + 277 + ], + "type": "image", + "image_path": "67a9e9e758a42e6dd13f367a4c4a22f0368a4062381af332bd2bdf374f9913e0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 278, + 336, + 285 + ], + "lines": [ + { + "bbox": [ + 277, + 278, + 336, + 285 + ], + "spans": [ + { + "bbox": [ + 277, + 278, + 336, + 285 + ], + "type": "text", + "content": "Panoptic Segmentation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 196, + 479, + 277 + ], + "blocks": [ + { + "bbox": [ + 365, + 196, + 479, + 277 + ], + "lines": [ + { + "bbox": [ + 365, + 196, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 365, + 196, + 479, + 277 + ], + "type": "image", + "image_path": "440cd1caf0ba872cd3fa1b9e07d3a7bc114939b763e4cccb263ae2cff9f0193b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 277, + 447, + 285 + ], + "lines": [ + { + "bbox": [ + 397, + 277, + 447, + 285 + ], + "spans": [ + { + "bbox": [ + 397, + 277, + 447, + 285 + ], + "type": "text", + "content": "Relation Prediction" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 354, + 227, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 227, + 366 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 227, + 366 + ], + "type": "text", + "content": "5.6 Visualization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 374, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 480, + 411 + ], + "type": "text", + "content": "As shown in Fig. 3, our OpenPSG method can predict relations not defined in the dataset, such as \"flying\", \"observing\", and \"fighting\", which qualitatively demonstrates the excellent open-set relation prediction capability of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 429, + 220, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 429, + 220, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 429, + 220, + 441 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 454, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 610 + ], + "type": "text", + "content": "In this paper, we propose the open-set PSG task and introduce the OpenPSG to accomplish open-set relation prediction. With the help of large multimodal models, our method employs an autoregressive approach to predict open-set relations. Additionally, we have developed a relation query transformer which contains pair feature extraction and relation existence estimation queries, one for extracting features of subject-object pairs, the other for predicting the existence of relations between them to filter out irrelevant pairs. Furthermore, we design generation and judgement instructions to enable the multimodal relation decoder to predict open-set relations. Extensive experiments demonstrate that our method achieves excellent performance in open-set relation prediction. In the future, we plan to employ model distillation to reduce the model size, thus enhancing the practicality in various real-world scenarios while ensuring its ability to predict open-set relations." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "content": "The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "type": "text", + "content": "2. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 480, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 480, + 305 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "type": "text", + "content": "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: CVPR (2020) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 360 + ], + "type": "text", + "content": "5. Chen, Z., Wu, J., Lei, Z., Zhang, Z., Chen, C.: Expanding scene graph boundaries: Fully open-vocabulary scene graph generation via visual-concept alignment and retention. arXiv preprint arXiv:2311.10988 (2023) 3, 4, 9, 10, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "text", + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) 3, 5, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "text", + "content": "7. Dai, J., He, K., Sun, J.: Convolutional feature masking for joint object and stuff segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3992-4000 (2015) 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 469, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 502 + ], + "type": "text", + "content": "9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR. pp. 14084-14093 (2022) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "text", + "content": "0. Gao, L., Wang, B., Wang, W.: Image captioning with scene-graph based semantic concepts. In: ICMLC (2018) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "type": "text", + "content": "1. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV. pp. 540-557. Springer (2022) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "text", + "content": "2. He, T., Gao, L., Song, J., Li, Y.F.: Towards open-vocabulary scene graph generation with prompt-based finetuning. In: ECCV (2022) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "3. Hildebrandt, M., Li, H., Koner, R., Tresp, V., Gunnemann, S.: Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072 (2020) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "4. Joshi, M., Levy, O., Weld, D.S., Zettlemoyer, L.: Bert for coreference resolution: Baselines and analysis. arXiv preprint arXiv:1908.09091 (2019) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 665 + ], + "type": "text", + "content": "5. Kan, X., Cui, H., Yang, C.: Zero-shot scene graph relation prediction through commonsense knowledge integration. In: Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part II 21. pp. 466-482. Springer (2021) 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019) 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "17. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision pp. 32-73 (2017) 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "type": "text", + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 2, 4, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "19. Li, L., Ji, W., Wu, Y., Li, M., Qin, Y., Wei, L., Zimmermann, R.: Panoptic scene graph generation with semantics-prototype learning. arXiv preprint arXiv:2307.15567 (2023) 1, 3, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "20. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR. pp. 7061-7070 (2023) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "21. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843 (2022) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "23. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: CVPR (2020) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 412 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS 36 (2024) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 510 + ], + "type": "text", + "content": "28. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "29. Singh, K.P., Salvador, J., Weihs, L., Kembhavi, A.: Scene graph contrastive learning for embodied navigation. In: ICCV (2023) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "30. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: CVPR (2019) 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "text", + "content": "31. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "32. Wang, C., Liu, X., Song, D.: Language models are open knowledge graphs. arXiv preprint arXiv:2010.11967 (2020) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 642 + ], + "type": "text", + "content": "33. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023) 1, 2, 3, 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "34. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: CVPR (2017) 9, 11" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 101 + ], + "type": "text", + "content": "Z. Zhou, Z. Zhu, H. Caesar, M. Shi" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Yang, J., Wang, C., Liu, Z., Wu, J., Wang, D., Yang, L., Cao, X.: Focusing on flexible masks: A novel framework for panoptic scene graph generation with relation constraints. In: ACM MM. pp. 4209-4218 (2023) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "36. Yang, J., Ang, Y.Z., Guo, Z., Zhou, K., Zhang, W., Liu, Z.: Panoptic scene graph generation. In: ECCV. pp. 178-196 (2022) 1, 3, 9, 10, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "37. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: CVPR. pp. 23497-23506 (2023) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "38. Yu, Q., Li, J., Wu, Y., Tang, S., Ji, W., Zhuang, Y.: Visually-prompted language model for fine-grained scene graph generation in an open world. arXiv preprint arXiv:2303.13233 (2023) 3, 9, 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "39. Yu, Q., Shen, X., Chen, L.C.: Towards open-ended visual recognition with large language model. arXiv preprint arXiv:2311.08400 (2023) 2, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "40. Yu, X., Chen, R., Li, J., Sun, J., Yuan, S., Ji, H., Lu, X., Wu, C.: Zero-shot scene graph generation with knowledge graph completion. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "text", + "content": "41. Yue, K., Chen, B.C., Geiping, J., Li, H., Goldstein, T., Lim, S.N.: Object recognition as next token prediction. arXiv preprint arXiv:2312.02142 (2023) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "type": "text", + "content": "42. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "43. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR (2018) 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "44. Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Yang, J., Zhang, L.: A simple framework for open-vocabulary segmentation and detection. In: ICCV. pp. 1020-1031 (2023) 2, 5, 10, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 435 + ], + "type": "text", + "content": "45. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Learning to generate language-supervised and open-vocabulary scene graph using pre-trained visual-semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2915–2924 (2023) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "text", + "content": "46. Zhao, C., Shen, Y., Chen, Z., Ding, M., Gan, C.: Textpsg: Panoptic scene graph generation from textual descriptions. In: ICCV. pp. 2839-2850 (2023) 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "47. Zhou, Z., Shi, M., Caesar, H.: Hilo: Exploiting high low frequency relations for unbiased panoptic scene graph generation. arXiv preprint arXiv:2303.15994 (2023) 1, 2, 3, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "text", + "content": "48. Zhou, Z., Shi, M., Caesar, H.: Vlprompt: Vision-language prompting for panoptic scene graph generation. arXiv preprint arXiv:2311.16492 (2023) 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "49. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 102 + ], + "type": "text", + "content": "OpenPSG" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_content_list.json b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..167efb7a5df9968bf0935e8c1c2fd226b6f64556 --- /dev/null +++ b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_content_list.json @@ -0,0 +1,1946 @@ +[ + { + "type": "text", + "text": "OpenSight: A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hu Zhang $^{1}$ , Jianhua Xu $^{4}$ , Tao Tang $^{5}$ , Haiyang Sun $^{6}$", + "bbox": [ + 287, + 214, + 712, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin $\\mathrm{Yu}^{2*}$ , Zi Huang $^{2*}$ , and Kaicheng $\\mathrm{Yu}^{3}$", + "bbox": [ + 331, + 229, + 671, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 CSIRO DATA61, Australia", + "bbox": [ + 403, + 256, + 599, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Queensland, Australia", + "bbox": [ + 359, + 271, + 642, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Westlake University", + "bbox": [ + 426, + 285, + 575, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ DAMO Academy, Alibaba, China", + "bbox": [ + 382, + 299, + 620, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Shenzhen Campus of Sun Yat-sen University", + "bbox": [ + 344, + 313, + 656, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ LiAuto Inc.", + "bbox": [ + 454, + 325, + 547, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hu1.Zhang@csiro.au, xin.yu@uq.edu.au, kyu@westlake.edu.cn", + "bbox": [ + 285, + 340, + 715, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/huzhangcs/OpenSight", + "bbox": [ + 352, + 356, + 648, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Traditional LiDAR-based object detection research primarily focuses on closed-set scenarios, which falls short in complex real-world applications. Directly transferring existing 2D open-vocabulary models with some known LiDAR classes for open-vocabulary ability, however, tends to suffer from over-fitting problems: The obtained model will detect the known objects, even presented with a novel category. In this paper, we propose OpenSight, a more advanced 2D-3D modeling framework for LiDAR-based open-vocabulary detection. OpenSight utilizes 2D-3D geometric priors for the initial discernment and localization of generic objects, followed by a more specific semantic interpretation of the detected objects. The process begins by generating 2D boxes for generic objects from the accompanying camera images of LiDAR. These 2D boxes, together with LiDAR points, are then lifted back into the LiDAR space to estimate corresponding 3D boxes. For better generic object perception, our framework integrates both temporal and spatial-aware constraints. Temporal awareness correlates the predicted 3D boxes across consecutive timestamps, recalibrating the missed or inaccurate boxes. The spatial awareness randomly places some \"precisely\" estimated 3D boxes at varying distances, increasing the visibility of generic objects. To interpret the specific semantics of detected objects, we develop a cross-modal alignment and fusion module to first align 3D features with 2D image embeddings and then fuse the aligned 3D-2D features for semantic decoding. Our experiments indicate that our method establishes state-of-the-art open-vocabulary performance on widely used 3D detection benchmarks and effectively identifies objects for new categories of interest.", + "bbox": [ + 259, + 429, + 743, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: OpenSight $\\cdot$ Open-vocabulary $\\cdot$ 3D object detection", + "bbox": [ + 261, + 791, + 696, + 806 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 823, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d096d890463e46c7b2d841608429504d3e4e80ed044a6032ffcb977886f8b585.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 279, + 148, + 403, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ce7beea5336b8b0e18db1fe60cbbfdb0fcf1831bb0769225ed3c9a71509d8e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 148, + 537, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f9373995613eb9f30f897fbd57d7ac2b65d0cd9ff8401e853e810cb1ffdecf47.jpg", + "image_caption": [ + "(b)", + "Fig. 1: (a) A model trained on known class \"car\" correctly detects the target object (yellow) with query \"car\". However, with a new class query \"motorcycle\", the model incorrectly detects the object (red). (b) The average precision of known \"car\" and novel \"motorcycle\". The degraded performance in \"motorcycle\" (blue) indicates the limited generalization from the known classes." + ], + "image_footnote": [], + "bbox": [ + 555, + 152, + 710, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 359, + 375, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The growing prevalence of autonomous driving has significantly increased research interest in various 3D perception tasks [13,17,41,55], among which LiDAR-based object detection stands as a crucial component. Despite the advancements in this field, most existing research focuses on closed-set settings [19,25,36,48,64], where the obtained models are limited to predicting a set of predefined categories. This constraint makes these models insufficient for handling complex real-world scenarios, where diverse and previously unseen object categories may be encountered. Consequently, there is a growing demand to develop models capable of detecting novel categories of interest.", + "bbox": [ + 212, + 397, + 787, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the limited availability of annotated categories in existing LiDAR-based datasets [4,15,43], adapting established large-scale visual-language models (LVLs), e.g., CLIP [37] for open-vocabulary LiDAR detection appears promising. One paradigm is to co-embed known class object features with the image/text embeddings from LVLs, expecting alignment generalization to the novel classes. While effective in 2D image-based open-vocabulary detection, this paradigm exhibits significant overfitting in LiDAR applications, as models persist in detecting objects in known classes even when faced with a new category.", + "bbox": [ + 212, + 535, + 789, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider some known classes, e.g., \"car\", and follow the paradigm above to illustrate this phenomenon. When given a novel category \"motorcycle\", the obtained model erroneously detects the known object \"car\", as depicted in Figure 1 (a). An analysis of average precision (AP) further reveals the diminished efficacy in novel classes, as shown in Figure 1 (b). These results underscore the limitation of such a paradigm, emphasizing the necessity for a tailored approach for LiDAR-based open-vocabulary problems.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present \"OpenSight\", a more advanced 2D-3D modeling approach for LiDAR-based open-vocabulary detection. OpenSight primarily relies on 2D and 3D geometry priors to first acquire the concept and localization of generic objects. Then, it interprets the semantics of detected objects by coembedding LiDAR features with 2D LVLs. This framework addresses all poten", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tially novel objects explicitly and circumvents the reliance on predefined known classes, thereby avoiding the overfitting problem.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we consider both temporal-aware and spatial-aware constraints to acquire the concept of generic objects and their localization. Initially, leveraging the 2D detection model and the accompanying RGB images of LiDAR, we extract 2D geometric priors for generic objects, that is, 2D bounding boxes across successive frames. These boxes are then lifted back into 3D space with LiDAR points, resulting in a collection of 3D frustums. Point clustering is then performed in each frustum to estimate the 3D bounding boxes of objects. Temporally, we correlate predicted 3D bounding boxes across frames and project the predicted 3D boxes from previous or subsequent frames onto the current frame. The missed or inaccurate objects in the current frame are then identified through overlapping comparison. This constraint boosts the continuity and consistency of generic objects over time. Spatially, we first employ large language models (LLMs) to derive size priors for general objects and then maintain 3D boxes that closely align with these size priors. The maintained boxes are randomly selected and placed at varying distances within the current frame, with the LiDAR points inside these boxes being sampled by their distance. This spatial awareness notably increases the visibility of objects, particularly those located further from the sensor. By integrating the temporal and spatial constraints, the ability to recognize and localize generic objects is significantly augmented.", + "bbox": [ + 212, + 176, + 787, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To decode the specific semantics of detected boxes in LiDAR space, we correlate objects across both LiDAR and image domains. Initially, by back-projecting 3D LiDAR bounding boxes onto the image plane, we identify the most closely matching 2D boxes from the Grounding DINO [27]. We then aim to minimize the feature discrepancy between the 3D LiDAR objects and the 2D matching box features. Subsequently, we fuse the optimized 3D-2D features and align the unified features with textual category embeddings from Grounding DINO to decode semantics. This dual 2D-3D feature approach leverages salient patterns in both 2D images and 3D geometry, significantly enhancing the robustness of semantic decoding. Additionally, we integrate size priors from LLMs before, recalibrating predictions that diverge from the average box size, thus enhancing the accuracy of our predictive model.", + "bbox": [ + 212, + 464, + 787, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the experiments, we first establish a variety of baselines for comparison. Then, we test our method on the widely-used 3D detection benchmarks, demonstrating that our approach achieves state-of-the-art open-vocabulary performance and effectively identifies objects for new categories of interest.", + "bbox": [ + 212, + 646, + 785, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 729, + 397, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Closed-set 3D Object Detection. Extensive research has been proposed on LiDAR-based 3D object detection in autonomous driving. One group of works performs on LiDAR-only data [39,63], either by extracting features from raw point clouds to predict 3D bounding boxes [23,34-36,40,49] or converting the irregular LiDAR point clouds into a regular feature space [14,19,44,46,48,51,64].", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1c80c38be6ce7555c4a05f7995a94400e9932325d5737e08c25aee22ffc763d7.jpg", + "image_caption": [ + "Fig. 2: An overview of our OpenSight framework. In generic object perception, 2D bounding boxes are initially lifted back into 3D space (initial boxes estimation). This is followed by integrated temporal and spatial awareness in the estimated 3D boxes to retrieve the missed objects or increase their visibility. Subsequently, we utilize these enhanced estimates to compute the generic object loss and optimize the 3D detection model. In specific semantic interpretation, we project the generated 3D boxes of the optimized 3D model into the image plane to get projected 2D boxes. These projections are then matched with the nearest 2D boxes identified via Grounding DINO, facilitating the alignment of 3D LiDAR object features with corresponding 2D image features." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 640, + 291 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2743e696be55661dbd6a8a2368c33f7ab29b301dff040f0dc649f03a3440fd70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 143, + 785, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another group focuses on both LiDAR and camera modalities. Some works emphasize point-wise fusion [18,42,45,52], projecting raw LiDAR points to retrieve corresponding image features and concatenating them together. In contrast, others concentrate on feature-level fusion [22,24,53], extracting features or proposals based on LiDAR points and querying the associated image features. This line of methods has been proven to be more efficient. For example, TransFusion [3] utilizes proposals generated by LiDAR features to query the image feature for further refinement. BEVFusion [25, 28] projects the image features into BEV space by using lift-splat-shoot (LSS) and concatenates them with the LiDAR features. More recently, LoGoNet [21] constructs Local-to-Global fusion network to perform LiDAR-camera fusion at both local and global levels. However, all these approaches target the closed-set setting, restricting the range of detected object categories.", + "bbox": [ + 212, + 446, + 787, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-vocabulary 2D Object Detection. The advent of large-scale pretrained models, which exhibit good text-vision alignment and zero-shot transferability, has spurred researchers to investigate the open-vocabulary problem in object detection tasks. A variety of approaches have been proposed with 2D RGB images: OVR-CNN [57] initially pre-trains a Faster R-CNN detector [38] by using image-caption pairs and text extractor, e.g., BERT [12], then fine-tunes it on downstream detection datasets. ViLD [16] distills the knowledge from a CLIP-based [37] teacher to a student detector for generalized text and image embedding alignment. Detic [62] harnesses extensive image classification data to train the classifier within a detector while localizing objects. Similar to OVR-CNN, OWL [31] first pre-trains a transformer encoder by contrastively aligning image-text pairs, then transfers it to detection tasks by attaching lightweight object classification and localization heads. Following the DETR-based", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "structure, OV-DETR [56] employs CLIP-encoded text and image embeddings as queries for obtaining category-specific boxes. In contrast to the aforementioned methods, some alternatives have emerged to directly learn open-vocabulary capabilities on detection data. For example, GLIP [20] unifies the object detection and phrase grounding tasks, allowing the combined use of both detection and grounding datasets for aligning phrase semantics and regional images. Det-CLIP [50] extracts the general concepts from multiple data sources and then resorts to a description-enriched concept dictionary for enhancement. Recently, Grounding DINO [27] adopts a similar detection-grounding training strategy as GLIP but proposes a tighter text-vision fusion solution based on transformer-based detector DINO [58].", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot Learning in 3D Data. Exploration of open-vocabulary 3D object detection has been limited due to the challenges associated with acquiring large-scale point cloud-text pairs and the inherent complexity of 3D scenarios. In the confined indoor scenario, OV-3DET [29] addresses the open-vocabulary detection problem by first localizing 3D objects and then assigning labels for the objects with open-vocabulary 2D models. However, temporal and spatial constraints are overlooked. CoDA [5] proposes to use some base classes to assist in detecting novel classes. As illustrated in Figure 1, such a paradigm in LiDAR data easily overfits the base classes. Recently, UP-VL [32] also considers the open-vocabulary case in autonomous driving, however, the categories involved are rather limited (only 3 classes). Some other works [2,6,7,47] tackle a related but distinct task of differentiating unknown data in testing from known data in training, known as the out-of-distribution (OOD) problem. Similarly, MODEST [54], OYSTER [59], and LiSe [61] study the unsupervised object discovery problem. Meanwhile, some studies [26,30,33] attempt to transfer existing 2D large-scale pre-trained models for open-vocabulary scene segmentation, while others [8-11,60] leverage the open-set ability of pre-trained models for the 3D classification tasks.", + "bbox": [ + 212, + 313, + 787, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from the above works, in this paper, we investigate open-vocabulary ability in LiDAR-based 3D object detection in the outdoor scenario, representing the real-world scenario and a much more complex scene.", + "bbox": [ + 212, + 584, + 787, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 OpenSight", + "text_level": 1, + "bbox": [ + 215, + 648, + 354, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To elaborate on the LiDAR-based open-vocabulary object detection, we first briefly introduce the necessary notations for ease of understanding. Following this, we delve into the details of generic object perception in Section 3.2. Subsequently, we present the process of specific semantic interpretation in Section 3.3. The overall framework of OpenSight is presented in Figure 2.", + "bbox": [ + 212, + 678, + 787, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Notation and Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 772, + 486, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The LiDAR dataset is denoted as $\\mathcal{D} = \\{P_i, \\{y_k\\}_{k=1}^K\\}_{i=1}^{|\\mathcal{D}|}$ , where $P_i$ corresponds to the LiDAR point clouds in $i$ -th frame, basically composed of $(x_i, y_i, z_i)$ coordinates. $y_k$ represents the combination of each 3D bounding box $b_k$ and its", + "bbox": [ + 212, + 792, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "semantic label $c_k$ . In open-vocabulary scenarios, $P_i$ is visible while $y_k$ is not. Generally, in the context of autonomous driving, each LiDAR frame is accompanied by $m$ images captured from varying perspectives. We define the image dataset as $\\mathcal{I} = \\{(I_{ij})_{j=1}^m\\}_{i=1}^{|\\mathcal{D}|}$ here. Image $I_{ij}$ in $j$ -th view and LiDAR points $P_i$ could be connected by the intrinsic matrix $\\mathbf{K}_j$ and extrinsic matrix $\\mathbf{E}_j$ . For clarity and brevity in the rest of the paper, we utilize a single LiDAR frame as an example and abbreviate $I_{ij}$ as $I_j$ , respectively. $N$ denotes the number of 3D objects, with $f^{3D}$ and $f^{2D}$ denotes the extracted feature in 3D and 2D.", + "bbox": [ + 212, + 146, + 787, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Generic Objects Perception", + "text_level": 1, + "bbox": [ + 214, + 290, + 491, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discerning and localizing generic objects are fundamental in LiDAR-based open-vocabulary detection when ground-truth boxes and labels are absent. The model is expected to perceive the existence of a generic object before assigning specific labels. This section details how we leverage 2D and 3D geometry priors to enable this perception, including initial box estimation, and temporal-spatial awareness. Initial Box Estimation. In practical applications, each LiDAR frame is typically paired with $m$ images $\\{I_j\\}_{j=1}^m$ , captured from various perspectives. Each image is processed through the pre-trained detection model $\\mathcal{I}$ , yielding a set of 2D bounding boxes $\\{b_{jk}\\}_{k=1}^{V_j}$ , where $V_j$ represents the number of 2D objects identified in the $j$ -th viewpoint. By integrating the data from all viewpoints, we compile the collective set of detected boxes $\\{(b_{jk}^{2D})_{k=1}^{V_j}\\}_{j=1}^m$ .", + "bbox": [ + 212, + 314, + 787, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the obtained 2D boxes of generic objects, we try to lift them back in the LiDAR space. Specifically, we project the LiDAR point $P_{i}$ into each image view by:", + "bbox": [ + 212, + 484, + 787, + 527 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {i j} = \\mathbf {K} _ {j} \\cdot \\mathbf {E} _ {j} \\cdot P _ {i}, j = 1, \\dots , m \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 527, + 784, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{E}_j$ and $\\mathbf{K}_j$ are extrinsic and intrinsic matrix, respectively. The points whose projection falls within each of the predicted 2D bounding boxes $b_{jk}^{2D}$ essentially constitute the 3D frustum, a conical region perceivable from a given perspective. We then apply the region growing algorithm [1] to isolate the cluster with the densest LiDAR points. Such an algorithm first randomly selects one point in the frustum and then iteratively finds its adjacent points.", + "bbox": [ + 212, + 550, + 787, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To estimate 3D bounding boxes from LiDAR point clusters, we first determine the extremal points along each axis by calculating the minimum and maximum coordinates. Take the X-axis for example, this is achieved by performing $x_{\\mathrm{min}} = \\min (\\{x_i\\}_{i = 0}^{|N_c|})$ and $x_{\\mathrm{max}} = \\max (\\{x_i\\}_{i = 0}^{|N_c|})$ , where $|N_{c}|$ is the number of points in the obtained cluster. These extremal points, identified similarly for the Y and Z axis, define the vertices of the bounding box. From these vertices, we construct the maximal bounding box $b_{i}^{3D}$ , enclosing all points in the cluster. For the heading of the bounding box, we traverse different angles. In each angle, we can obtain the XYZ boundary. We then shrink the boundary by a scale of 0.95, some points will be out of the obtained bounding boxes. The angle with the most outside points is regarded as the angle of the object. It is based on the assumption that for the \"tightest\" bounding box, a slight change in the angle will incur a big change in the encompassed points. The 3D boxes for generic", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/110661097f1a82a690868c814dfe8dd6a228b8dc92ba39f79f60cd1ff952c13a.jpg", + "image_caption": [ + "Fig. 3: Temporal awareness: Project boxes from frame $t + 1$ to $t$ . By comparing with projections, we refine missed or inaccurately estimated boxes in frame $t$ ." + ], + "image_footnote": [], + "bbox": [ + 285, + 147, + 712, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "objects at timestamp $t$ are then finalized as $\\mathcal{B}^t = \\{b_i^{3D}\\}_{i=0}^{N_t}$ , by aggregating all boxes from $m$ perspectives.", + "bbox": [ + 212, + 305, + 782, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal Awareness. For the LiDAR point cloud, point density decreases sharply as distance increases, leading to less reliable estimations for distant objects. However, in autonomous driving scenarios, objects that are distant in the current frame may move closer in subsequent frames. To exploit this, we leverage the temporal continuity between frames by correlating 3D bounding box predictions across successive timestamps. We take frames at timestamp $t$ and $t + 1$ as an example. Specifically, given the predicted boxes $\\mathcal{B}^t = \\{b_i\\}_{i=0}^{N_t}$ and $\\mathcal{B}^{t+1} = \\{b_i\\}_{i=0}^{N_{t+1}}$ , where $N_{t+1}$ and $N_t$ represent the number of boxes in $t + 1$ and $t$ , respectively, we project each 3D box $b_i$ in $\\mathcal{B}^{t+1}$ back to frame $t$ using the transformation:", + "bbox": [ + 212, + 337, + 784, + 489 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {b} _ {i} = \\mathbf {E} _ {\\mathrm {l i d a r} \\leftarrow \\mathrm {e g o} _ {t}} \\cdot \\mathbf {E} _ {\\mathrm {e g o} _ {t} \\leftarrow \\mathrm {g l b}} \\cdot \\mathbf {E} _ {\\mathrm {g l b} \\leftarrow \\mathrm {e g o} _ {t + 1}} \\cdot \\mathbf {E} _ {\\mathrm {e g o} _ {t + 1} \\leftarrow \\mathrm {l i d a r}} \\cdot b _ {i},\n$$\n", + "text_format": "latex", + "bbox": [ + 289, + 502, + 710, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{E}_{\\mathrm{lidar}} \\leftarrow \\mathrm{ego}_t$ and $\\mathbf{E}_{\\mathrm{ego}_t} \\leftarrow \\mathrm{glb}$ are the extrinsic matrices for the ego to LiDAR and global to ego transformations at timestamp $t$ , and $\\mathbf{E}_{\\mathrm{glb}} \\leftarrow \\mathrm{ego}_{t+1}$ , $\\mathbf{E}_{\\mathrm{ego}_{t+1}} \\leftarrow \\mathrm{lidar}$ denotes ego to global and LiDAR to ego at timestamp $t+1$ , respectively.", + "bbox": [ + 212, + 532, + 782, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With all boxes in $\\mathcal{B}^{t + 1}$ considered, we obtain their projections as $\\hat{\\mathcal{B}}^{t + 1}$ and construct the overlap matrix $M\\in \\mathbb{R}^{N_{t + 1}\\times N_t}$ by calculating their Intersection over Union (IoU) with boxes in $\\mathcal{B}^t$ . We address two key scenarios: If a box $\\hat{b}_i$ has no overlap with boxes in $\\mathcal{B}^t$ , its counterpart $b_{i}$ in timestamp $t + 1$ is considered a missed detection in the current frame and is added to the prediction at timestamp $t$ . Conversely, if box $\\hat{b}_i$ overlaps with boxes in $\\mathcal{B}^t$ and its distance exceeds a certain threshold, we perform Non-Maximum Suppression (NMS) and retain the union of the two boxes to address potentially incomplete estimations at greater distances (See Figure 3).", + "bbox": [ + 212, + 579, + 784, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This process updates the predicted boxes of the current frame to $\\mathcal{B}_{\\mathrm{update}}^t = \\mathcal{B}^t \\bigcup \\mathcal{B}_{\\mathrm{temporal}}^t$ , which combines the original and temporally adjusted boxes. In practice, for the current frame $t$ , we consider both its subsequent frame at timespan $t + 1$ and its precedent frame at timestamp $t - 1$ . That means two adjacent frames are considered. Through such a constraint, we are likely to boost the continuity and consistency of generic objects over time.", + "bbox": [ + 212, + 717, + 784, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial Awareness. Considering the sparse object distribution in 3D LiDAR scans, where vast areas of the point cloud are often devoid of meaningful objects,", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 113, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5156ce522d22bed4fcf0cfc06f61d8699612134425bc3578851756d1313c215b.jpg", + "image_caption": [ + "Fig. 4: Spatial awareness: Size priors are first generated from GPT-4 to select \"precisely\" estimated boxes and build the object bank. Randomly select objects and points from the objects bank and place them in the current frame $t$ ." + ], + "image_footnote": [], + "bbox": [ + 287, + 143, + 709, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we also focus on the systematic integration of \"precisely\" estimated objects in each frame to enhance the detectability and representation of generic objects. To obtain these precise boxes from $\\mathcal{B}^t$ in the absence of any ground truth, we utilize large language models (LLMs), such as GPT-4, for extracting useful size priors of objects. GPT-4, when prompted appropriately, can provide valuable insights into the typical sizes of objects within a specific class, e.g., the width, length, and height of a class. For instance, prompts like \"What is the average size of a \\{\\#CLASS\\}\" in the autonomous driving scene?\" enable us to collect size statistics for common classes (Details in the Appendix). Moreover, leveraging the ability of GPT-4 to process images, we also input images of the relevant objects to validate these size priors (See Figure 4).", + "bbox": [ + 212, + 401, + 787, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Utilizing the extracted size priors, we selectively retain 3D boxes in $\\mathcal{B}^t$ that closely align with these priors. A box is reserved if its dimensions $w$ , $l$ , and $h$ fall within the specified range:", + "bbox": [ + 212, + 566, + 789, + 613 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw \\in [ (1 - \\tau) w _ {p r i o r}, (1 + \\tau) w _ {p r i o r} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 621, + 622, + 638 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\in [ (1 - \\tau) \\ell_ {p r i o r}, (1 + \\tau) \\ell_ {p r i o r} ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 638, + 784, + 656 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nh \\in [ (1 - \\tau) h _ {p r i o r}, (1 + \\tau) h _ {p r i o r} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 659, + 617, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tau$ represents the error margin. $w_{prior}$ , $l_{prior}$ and $h_{prior}$ are from LLMs-derived priors. This process results in an object bank $O = \\{(b_i^{3D}, P_i)\\}_{i=0}^{N_o}$ , where $b_i^{3D}$ is the reserved box, $P_i$ comprises the LiDAR points within each box and $N_o$ is the total number of reserved boxes.", + "bbox": [ + 212, + 683, + 787, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the current frame, we randomly select boxes from bank $O$ and integrate them at varying distances from the ego vehicle, ensuring diverse spatial distribution. The LiDAR points within each box are sampled according to their distance, with the sampling ratio defined as:", + "bbox": [ + 212, + 744, + 787, + 805 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr _ {s a m p l i n g} = \\min (1, \\frac {d - d _ {n e w}}{d - d _ {o r i}}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 811, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $d$ denotes the total range in the point cloud, $d_{ori}$ is the original distance of the object, $d_{new}$ is the distance post-placement. The selected boxes are denoted as $\\mathcal{B}_{\\mathrm{spatial}}^t$ and the collective set of boxes in the current frame is represented as $\\mathcal{B}_{\\mathrm{final}}^t = \\mathcal{B}^t \\cup \\mathcal{B}_{\\mathrm{temporal}}^t \\cup \\mathcal{B}_{\\mathrm{spatial}}^t$ . This strategy aims to substantially enhance the visibility of generic objects in LiDAR data and the perception ability of the obtained model.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss for Generic Object Perception. With the obtained 3D boxes $\\mathcal{B}_{\\mathrm{final}}^t = \\{b_k^{3D}\\}_{k=1}^{V_{\\mathrm{final}}}$ , we formulate the loss for generic object concept acquisition and localization. Utilizing a DETR-like structure, we extract features from LiDAR points and apply bipartite matching to determine the optimal permutation $\\hat{\\sigma}$ , the loss is thus expressed as:", + "bbox": [ + 214, + 238, + 782, + 316 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum^ {N} \\left[ \\mathcal {L} _ {\\text {f o c a l}} \\left(s _ {\\hat {\\sigma} _ {i} (k)}\\right) + \\mathcal {L} _ {\\mathrm {L} _ {1}} \\left(b _ {\\hat {\\sigma} _ {i} (k)}, b _ {k} ^ {3 D}\\right) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 325, + 784, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $s_{\\hat{\\sigma}_i(k)}$ denotes whether it is an object or not, $\\hat{\\sigma}_i(k)$ is the element in $\\hat{\\sigma}$ . $\\mathcal{L}_{\\mathrm{focal}}$ constrains the level of \"objectness\" of generated 3D boxes, $\\mathcal{L}_{\\mathrm{box}}$ considers the constraint on boxes. With this composite loss optimized, we obtain a model with the ability to perceive generic objects.", + "bbox": [ + 214, + 359, + 782, + 421 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Specific Semantics Interpretation", + "text_level": 1, + "bbox": [ + 214, + 439, + 535, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To interpret the specific semantics of detected generic objects in LiDAR, we correlate objects across both LiDAR and image domains. We apply the model after generic object perception to a given LiDAR frame, producing boxes $\\{b_i^{3D}\\}_{i=0}^N$ and their corresponding features $\\{f_i^{3D}\\}_{i=0}^N$ . Instead of using the image-level LVLs, e.g., CLIP, we opt for object-level LVLs, such as Grounding DINO [27], which provides 2D boxes $\\{b_{i=0}^{2D}\\}_{i=0}^{N_{2D}}$ , 2D object features $\\{f_i^{2D}\\}_{i=0}^{N_{2D}}$ and text embeddings $\\{t_i\\}_{i=0}^{N_{2D}}$ simultaneously.", + "bbox": [ + 214, + 460, + 782, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For each detected 3D box $b_{i}^{3D}$ , we project it to each view image with the corresponding intrinsic matrix $K_{j}$ and extrinsic matrix $E_{j}$ , formulated as:", + "bbox": [ + 212, + 569, + 782, + 599 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {b} _ {i j} ^ {2 D} = \\mathbf {K} _ {j} \\cdot \\mathbf {E} _ {j} \\cdot b _ {i} ^ {3 D}, j = 1, \\dots , m \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 608, + 784, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The $\\hat{b}_{ij}^{2D}$ is further filtered and only those lying in the image are reserved. For the reserved box $\\hat{b}_{ij}^{2D}$ , we compute the IoU with the 2D boxes $\\{b_{i}^{2D}\\}_{i=0}^{N_{2D}}$ from Grounding DINO and find the most close one $b_{j}^{2D}$ . The corresponding text embeddings $t_{j}^{2D}$ feature is retrieved. We then try to minimize the discrepancy between $f_{i}^{3D}$ and all 2D features that have the same text embeddings $t_{j}^{2D}$ . Mathematically, it is expressed as:", + "bbox": [ + 214, + 635, + 782, + 734 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {a l i g n}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\sum_ {j = 0} ^ {N _ {i}} e ^ {f _ {i} ^ {3 D} \\cdot f _ {j} ^ {2 D} / \\delta}}{\\sum_ {k = 0} ^ {N ^ {2 D}} e ^ {f _ {i} ^ {3 D} \\cdot f _ {k} ^ {2 D} / \\delta}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 742, + 782, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $N_{i}$ denotes all 2D features that have the same text embeddings $t_j^{2D}$ for $b_i^{3D}$ . Once the object-level alignment is finished, we can directly perform open-vocabulary LiDAR detection with the LiDAR input.", + "bbox": [ + 214, + 792, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "2D-3D Integration. Considering that paired RGB images with LiDAR are always available in autonomous driving, we thus can perform 2D-3D feature integration. For each detected object, we have optimized 3D features $f_{i}^{3D}$ and identified the best matched 2D boxes feature $f_{i}^{2D}$ , the similarity between 3D-2D features with the text embeddings in $\\{t_i\\}_{i = 0}^{N_{2D}}$ are computed as follows:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s _ {j} ^ {3 D} = \\cos \\left(f _ {i} ^ {3 D}, t _ {j}\\right), \\\\ 2 ^ {D} = \\left(3 ^ {D} + 1\\right) \\dots 0 \\quad \\text {v} ^ {3 D} \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 228, + 784, + 255 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\ns _ {j} ^ {2 D} = \\cos \\left(f _ {i} ^ {2 D}, t _ {j}\\right), j = 0, \\dots , N ^ {2 D}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 250, + 622, + 267 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We obtain the maximum value for 2D and 3D feature by $s^{3D} = \\max_j\\{s_j^{3D}\\}_{j = 0}^{N_{2D}}$ , $s^{2D} = \\max_j\\{s_j^{2D}\\}_{j = 0}^{N_{2D}}$ , along with their semantic labels $c^{3D}$ and $c^{2D}$ . The final label for object $i$ is determined by comparing $s_i^{2D}$ and $s_i^{3D}$ , selecting the label associated with the larger value.", + "bbox": [ + 212, + 273, + 787, + 340 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Large Language Model (LLMs) Post-processing. The size priors from LLMs can be used to assist in recalibrating predictions. By matching the 3D IoU of predicted boxes with these size priors, we can determine the most matching prior and the corresponding semantic label. We then incorporate the predictions from size prior with results from 2D-3D integration by selecting label with the highest similarity.", + "bbox": [ + 212, + 340, + 787, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.4 Inference", + "text_level": 1, + "bbox": [ + 214, + 450, + 339, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "we first input the LiDAR point cloud to the generic object detector and extract 3D object features. We then align these features with text embeddings from Grounding DINO [27] to determine semantic labels. Additionally, when RGB images are available, we further improve the alignment ability by integrating 3D and 2D features or use the LLMs prior (size priors from GPT-4) to calibrate the predicted labels, thereby improving the accuracy of obtained semantic labels.", + "bbox": [ + 212, + 470, + 787, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 580, + 375, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental settings", + "text_level": 1, + "bbox": [ + 214, + 607, + 442, + 623 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We conduct the open-vocabulary experiments on nuScenes [4], a large-scale, outdoor dataset featuring multimodal data modalities. Each frame in the dataset contains data from one 32-beam LiDAR and six cameras, offering a comprehensive view of the surroundings. The dataset comprises 1,000 scenes in total, with up to 1.4 million annotated 3D bounding boxes spanning 10 classes. We reserve all the classes as novel to evaluate the effectiveness of our method.", + "bbox": [ + 212, + 628, + 787, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. For quantitative evaluation, we initially employ Average Precision (AP) on each class and mean Average Precision (mAP) across all novel classes. Considering that LiDAR point density varies with the distance, we introduce distance-based metrics: mAP-near, mAP-midrange, and mAP-far. Similarly, acknowledging the size disparities among objects in different classes (e.g., 'bus' vs. 'traffic_cone'), we categorize classes into 'large', 'medium', and 'small' and propose mAP-large, mAP-medium, and mAP-small. More details are in the Appendix.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f3f3e75358e299ca73b2d78e5bfab19ab0c7536f2c772711393f1e24025e3404.jpg", + "table_caption": [ + "Table 1: Results on NuScenes dataset. All classes are reserved as novel classes. AP of each class and mAP across all classes are shown. \"Car\" in CoDA is the known class and is fully supervised ('Cone': 'Traffic_cone', 'C.V': 'Construction_ vehicle', 'Motor.'] 'Motorcycle', 'Ped.'] 'Pedestrian')." + ], + "table_footnote": [], + "table_body": "
MethodmAPNDSCarTruckC.V.BusTrailerBarrierBicycleConeMotor.Ped.
Supervised (box+label)57.361.786.038.818.363.334.270.546.770.259.585.2
Supervised (box only)34.239.053.220.17.420.50.750.137.852.639.160.8
Detic-3D [62]5.812.715.12.92.14.20.00.16.20.86.121.1
OWL-ViT-3D [31]5.412.414.22.51.84.00.00.15.21.15.519.4
OV-PointCLIP [60]0.15.00.00.00.00.00.00.00.70.20.00.1
OV-PointCLIPv2 [65]2.810.90.02.70.00.00.022.00.00.20.03.2
CLIP-3D [37]7.214.29.62.82.12.60.00.111.514.49.818.9
OV-3DET [29]5.712.016.12.90.31.30.00.31.811.02.021.1
SeCondPoint [26]8.215.015.12.32.02.00.00.010.914.915.219.3
3DGenZ [30]7.514.914.63.52.33.20.00.010.514.28.717.9
Openscene [33]8.815.417.23.92.23.00.10.212.814.010.823.4
UP-VL [32]11.317.219.47.94.33.10.40.514.115.616.830.8
CoDA [5]10.316.185.51.90.02.00.00.03.90.05.93.5
OpenSight (Ours)22.923.325.311.68.75.10.842.226.125.630.452.5
", + "bbox": [ + 217, + 204, + 782, + 354 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/cacdd256278e0e7a0b466974592fc2073c2fe642e7489b9ed5d5005ef7e58b7d.jpg", + "table_caption": [ + "Table 2: Performance w.r.t mAP-near, mAP-medium, mAP-far." + ], + "table_footnote": [], + "table_body": "
MethodmAP-nearmAP-midrangemAP-far
Detric-3D [62]18.59.35.8
OWL-ViT-3D [31]17.38.25.4
OV-PointCLIP [60]0.20.10.1
OV-PointCLIPv2 [65]4.53.62.8
CLIP-3D [37]18.811.27.2
OV-3DET [29]16.19.95.7
CoDA [5]18.216.610.3
OpenSight (Ours)28.026.522.9
", + "bbox": [ + 326, + 378, + 666, + 486 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Baselines", + "text_level": 1, + "bbox": [ + 215, + 505, + 339, + 520 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As there are no direct baselines addressing open-vocabulary LiDAR-based 3D detection, we primarily construct the baselines based on well-known works [5, 29, 37, 62, 65] and compare with OpenSight. Specifically, they are:", + "bbox": [ + 212, + 535, + 785, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fully-supervised training. Initially, we perform fully supervised training across all categories, establishing an upper bound for the open-vocabulary detection performance. We consider using both ground-truth boxes and labels ('Supervised (box+label)') or just using ground-truth boxes ('Supervised (box only)').", + "bbox": [ + 212, + 580, + 790, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "2D open-vocabulary methods. Detic [62] is a pioneer work in the realm of 2D open-vocabulary object detection tasks and can detect 21K classes. It is thus desirable for our study here. OWL-ViT [31] decodes the object labels and boxes directly from the pre-trained features and shows good robustness in real-world applications. With the 2D detection results of either Detic or OWL-ViT, we adopt the same procedure in Section 3.2 to obtain the corresponding 3D bounding boxes in LiDAR space. These two baselines are denoted as Detic-3D and OWL-ViT-3D.", + "bbox": [ + 212, + 643, + 787, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "3D open-vocabulary methods. We then consider the latest open-vocabulary point cloud classification methods and a variety of indoor open-vocabulary works to evaluate their performance in our tasks here. Specifically, we utilize a fully-supervised method to generate pseudo boxes and employ PointCLIP [60], Point-CLIPv2 [65], denoted as \"OV-PointCLIP\", \"OV-PointCLIPv2\". We also combine", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 113, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/aa92635ae5553547c73a86a0415e2d2bceccb3a6fac8f2a2244f5250db30095d.jpg", + "table_caption": [ + "Table 3: Performance w.r.t mAIP-large, mAIP-medium, mAIP-small." + ], + "table_footnote": [], + "table_body": "
MethodmAP-largemAP-mediummAP-small
Detric-3D [62]5.34.111.0
OWL-ViT-3D [31]4.93.710.3
OV-PointCLIP [60]0.00.30.2
OV-PointCLIPv2 [65]0.67.31.6
CLIP-3D [37]3.47.116.7
OV-3DET [29]4.11.416.0
CoDA [5]17.94.91.8
OpenSight (Ours)11.532.939.0
", + "bbox": [ + 323, + 161, + 671, + 268 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/60d497238061f65927b3263b7876938385c9abd934078ce3445571fdbd739161.jpg", + "table_caption": [ + "Table 4: Results with temporal and spatial awareness w.r.t mAP-{near, medium, far}." + ], + "table_footnote": [], + "table_body": "
#TemporalSpatialmAP-nearmAP-midrangemAP-far
(1)27.124.822.1
(2)✓27.425.922.5
(3)✓27.526.022.7
(4)✓✓28.026.522.9
", + "bbox": [ + 303, + 290, + 689, + 358 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "our predicted boxes and CLIP [37] for this task (\"CLIP-3D\"). Meanwhile, we adapt the latest works OV-3DET [29] and CoDA [5] in indoor settings and UP-VL [32] to our task here. We also consider the segmentation-based methods, e.g., SeCondPoint [26], 3DGenZ [30], and OpenScene [33], for comparison.", + "bbox": [ + 212, + 376, + 787, + 439 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Main results", + "text_level": 1, + "bbox": [ + 215, + 463, + 366, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we discuss the performance of our model on novel classes, as summarized in Table 1. Despite the lack of 3D ground truth supervision, our framework demonstrates strong overall open-vocabulary ability across classes. Notably, it achieves exceptional performance in some classes, for example, 52.5 AP in 'pedestrian' and 42.2 AP in 'Barrier'. Unfortunately, we also notice the limited performance in 'Trailer', with 0.8 AP. To analyze this, we compare with 'Supervised (box only)', where groundtruth boxes are used and only 0.7 AP is achieved in 'Trailer', we thus attribute the reason to label alignment with Grounding DINO rather than box inaccuracy. Grounding DINO offers limited help in fine-grained classes. Our model also shows superior performance when compared with the baselines. Specifically, we observe that PointCLIP is very limited in detecting novel objects and CLIP-3D achieves the best overall performance with our detected boxes. CoDA, while effective in known classes like 'car', falls short in novel class adaptation, averaging only $1.9\\%$ accuracy across novel categories.", + "bbox": [ + 212, + 491, + 787, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our more detailed evaluations, using distance-based metrics and size-based metrics, are presented in Tables 2 and 3, respectively. The evaluation in Table 2 reveals a trend where detection accuracy is higher for objects closer to the sensor. This is attributed to their denser point clouds and more defined geometric features. Conversely, detection accuracy decreases with increasing distance, likely due to sparser data and less distinct object outlines. Regarding size-based performance (Table 3), our model shows superior results for medium and small-sized objects. Large objects present lower detection rates, mainly due to their", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/db08133b340793c61862036eec9e8942c1739cedc4166b26370346b0c3f0a07a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 143, + 495, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b9faafb06d53ae2e42f602c6572d3f370a9374a8eef47e11f6dcdedeaee7a22a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 143, + 740, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/60f92f0039ec3803b4558a3cdf63f74369507c6db68f7f47729e348794a32ad0.jpg", + "image_caption": [ + "Fig. 5: Visualization of detected novel classes in different LiDAR frames. (a) Detected 'barrier', 'traffic_cone', and 'pedestrian'; (b) Detected 'car' and 'pedestrian'; (c) Detected 'car' and 'motorcycle'; (d) Detected 'barrier' and 'pedestrian'." + ], + "image_footnote": [], + "bbox": [ + 254, + 233, + 493, + 318 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6b2e1086d7a5632be171624e50694606b81d9882449c8aafc55ebefed85a223c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 234, + 740, + 316 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/039ece4aab9e9a44d9cc1d84af6f69d4298dd0c4f1d1531623210cb4f814b541.jpg", + "table_caption": [ + "Table 5: Results with direct alignment ('D.A.'), 3D-2D feature integration ('I.G.'), and LLMs priors." + ], + "table_footnote": [], + "table_body": "
#D.AI.G.PriorsmAP-largemAP-mediummAP-small
(1)✓9.431.838.1
(2)✓✓10.932.738.8
(3)✓✓✓11.532.939.0
", + "bbox": [ + 299, + 409, + 696, + 465 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "incomplete LiDAR scans. Additionally, the fine-grained categorization of large vehicles, such as 'Constructionvehicle', complicates their accurate detection.", + "bbox": [ + 214, + 470, + 782, + 502 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Qualitative results on novel classes", + "text_level": 1, + "bbox": [ + 214, + 521, + 549, + 536 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Figure 5, our framework demonstrates a strong ability to detect novel objects by inputting the interested categories. We visualize the results from both BEV and a specific perspective. For example, our method successfully identifies the 'pedestrian', 'barrier', and 'traffic_cone' in example (a), and the 'car' and 'motorcycle' in (c). These results also indicate that our method exhibits superior open-vocabulary detection ability across different scenes.", + "bbox": [ + 212, + 542, + 784, + 632 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation", + "text_level": 1, + "bbox": [ + 214, + 652, + 334, + 666 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this part, we present a comprehensive analysis of OpenSight and examine the effectiveness of each component.", + "bbox": [ + 212, + 672, + 784, + 703 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Components in Generic Object Perception. We decouple the components in generic object perception and record the results in Table 4. The overall improvement may not be so obvious, however, we improve more for mAP-medium and mAP-far with these techniques. For example, temporal improves 1.1 AP for mAP-medium while 0.3 AP for mAP-near. Both modules interact positively with each other and further improve the results when jointly introduced.", + "bbox": [ + 212, + 704, + 784, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Components in Semantic Interpretation. We analyze the modules of direct alignment, 3D-2D feature integration, and LLMs priors in semantic interpretation and record the results in Table 5. The performance is consistently improved", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/71a81f12604decfd6c85dcddf42055b3f118d74c1d5015a018485c89ea3c84cc.jpg", + "table_caption": [ + "Table 6: Results of image-level and object-level models." + ], + "table_footnote": [], + "table_body": "
mAP-near-midrange-far-large-medium-small
CLIP19.011.98.23.98.218.4
Grounding DINO28.026.522.911.532.939.0
", + "bbox": [ + 289, + 161, + 709, + 214 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b8fbe631153f030eea54a30f8001f92678cca0c243dd5d5ca768b4ede9683606.jpg", + "table_caption": [ + "Table 7: AP and AR of some classes under different threshold." + ], + "table_footnote": [], + "table_body": "
#CLASSAP-0.5AP-1AP-2AR-0.5AR-1AR-2
Car4.718.345.940.069.084.4
Truck0.23.013.313.230.051.1
Pedestrian38.851.657.158.469.374.2
", + "bbox": [ + 287, + 243, + 707, + 301 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "with integration and priors introduced. Compared with LLMs priors, 3D-2D integration brings more benefits, for example, 0.9 vs. 0.2 in mAP-medium.", + "bbox": [ + 212, + 310, + 782, + 339 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Image-level vs. Object-level. We also investigate the image-level CLIP [37] model in semantic interpretation. The results are recorded in Table 6. Compared with our adopted object-level Grounding DINO, we notice that object-level alignment shows much better results from either a distance perspective or a size perspective. For example, 28.0 AP of near objects is achieved in object-level alignment, 9 points higher than image-level alignment. We analyze that CLIP requires more accurate bounding box prediction, in comparison with object-level alignment which only requires a nearest box.", + "bbox": [ + 212, + 339, + 784, + 460 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Importance of Bounding Boxes Existence. Existing open-vocabulary methods focus more on the alignment between 3D features and 2D features. However, in LiDAR-based object detection, we emphasize that accurate bounding boxes are essential before alignment. We select some classes and investigate both average precision (AP) and average recall (AR) under different thresholds. The results are recorded in Table 7. Compared with AP, we observe a higher recall value in different threshold, indicating that although we can recall the correct objects, more objects with incorrect sizes are also recalled. Such results motivate that if we can obtain a more precise estimation of objects, the results will be improved further.", + "bbox": [ + 212, + 460, + 785, + 612 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 635, + 359, + 651 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present OpenSight, a simple and versatile open-vocabulary framework for LiDAR-Based object detection. In the absence of ground-truth boxes and semantic labels, OpenSight first learns to identify and localize a wide range of generic objects. To boost the ability of generic object perception, we incorporate temporal and spatial awareness techniques that recalibrate and improve the visibility of estimated 3D boxes. We also align LiDAR object features with the 2D features from the Grounding DINO for semantic interpretation. The extensive experiments demonstrate the effectiveness and generalization ability of the proposed framework, paving the way for further exploration in LiDAR-based open-vocabulary challenges.", + "bbox": [ + 212, + 667, + 787, + 819 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 393, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is funded in part by ARC-Discovery grant (DP220100800 to XY) and ARC-DECRA grant (DE230100477 to XY). We thank all anonymous reviewers and ACs for their constructive suggestions.", + "bbox": [ + 212, + 172, + 785, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 239, + 323, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994)", + "2. Alliegro, A., Cappio Borlino, F., Tommasi, T.: 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems 35, 21228-21240 (2022)", + "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022)", + "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)", + "5. Cao, Y., Zeng, Y., Xu, H., Xu, D.: Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. arXiv preprint arXiv:2310.02960 (2023)", + "6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Open-set 3d object detection. In: 2021 International Conference on 3D Vision (3DV). pp. 869-878. IEEE (2021)", + "7. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. pp. 318-334. Springer (2022)", + "8. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. arXiv preprint arXiv:1907.06371 (2019)", + "9. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 923-933 (2020)", + "0. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision 130(10), 2364-2384 (2022)", + "1. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019)", + "2. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "3. Douillard, B., Underwood, J., Kuntz, N., Vlaskine, V., Quadros, A., Morton, P., Frenkel, A.: On the segmentation of 3d lidar point clouds. In: 2011 IEEE International Conference on Robotics and Automation. pp. 2798-2805. IEEE (2011)", + "4. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021)" + ], + "bbox": [ + 225, + 266, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)", + "16. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=1L3lnMbR4WU", + "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)", + "18. Huang, T., Liu, Z., Chen, X., Bai, X.: Epnet: Enhancing point features with image semantics for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 35-52. Springer (2020)", + "19. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019)", + "20. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)", + "21. Li, X., Ma, T., Hou, Y., Shi, B., Yang, Y., Liu, Y., Wu, X., Chen, Q., Li, Y., Qiao, Y., et al.: Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. arXiv preprint arXiv:2303.03595 (2023)", + "22. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)", + "23. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021)", + "24. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)", + "25. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790 (2022)", + "26. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2021)", + "27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542 (2022)", + "29. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1190-1199 (2023)", + "30. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3d point clouds. In: 2021 International Conference on 3D Vision (3DV). pp. 992-1002. IEEE (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230 (2022)", + "32. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8602-8612 (2023)", + "33. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. arXiv preprint arXiv:2211.15654 (2022)", + "34. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 918-927 (2018)", + "35. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)", + "36. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)", + "39. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020)", + "40. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 770-779 (2019)", + "41. Simon, M., Amende, K., Kraus, A., Honer, J., Samann, T., Kaulbersch, H., Milz, S., Michael Gross, H.: Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "42. Sindagi, V.A., Zhou, Y., Tuzel, O.: Mvx-net: Multimodal voxelnet for 3d object detection. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7276-7282. IEEE (2019)", + "43. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)", + "44. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)", + "45. Wang, C., Ma, C., Zhu, M., Yang, X.: Pointaugmenting: Cross-modal augmentation for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11794-11803 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 18-34. Springer (2020)", + "47. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020)", + "48. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "49. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020)", + "50. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407 (2022)", + "51. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021)", + "52. Yin, T., Zhou, X., Krahenbuhl, P.: Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems 34, 16494-16507 (2021)", + "53. Yoo, J.H., Kim, Y., Kim, J., Choi, J.W.: 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 720-736. Springer (2020)", + "54. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1130-1140 (2022)", + "55. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access 8, 58443-58469 (2020)", + "56. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 106-122. Springer (2022)", + "57. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)", + "58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)", + "59. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9317-9328 (2023)", + "60. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)", + "61. Zhang, R., Zhang, H., Yu, H., Zheng, Z.: Approaching outside: Scaling unsupervised 3d object detection from 2d scene. In: Proceedings of the European conference on computer vision (ECCV) (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 350-368. Springer (2022)", + "63. Zhou, Y., Sun, P., Zhang, Y., Anguelov, D., Gao, J., Ouyang, T., Guo, J., Ngiam, J., Vasudevan, V.: End-to-end multi-view fusion for 3d object detection in lidar point clouds. In: Conference on Robot Learning. pp. 923-932. PMLR (2020)", + "64. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)", + "65. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 327 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection", + "bbox": [ + 230, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_model.json b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1de58305778b75e3279d99b3a5fa5aa91b2bff87 --- /dev/null +++ b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_model.json @@ -0,0 +1,2680 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "OpenSight: A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.215, + 0.714, + 0.231 + ], + "angle": 0, + "content": "Hu Zhang\\(^{1}\\), Jianhua Xu\\(^{4}\\), Tao Tang\\(^{5}\\), Haiyang Sun\\(^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.231, + 0.672, + 0.247 + ], + "angle": 0, + "content": "Xin \\(\\mathrm{Yu}^{2*}\\) , Zi Huang\\(^{2*}\\), and Kaicheng \\(\\mathrm{Yu}^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.404, + 0.257, + 0.601, + 0.271 + ], + "angle": 0, + "content": "1 CSIRO DATA61, Australia" + }, + { + "type": "text", + "bbox": [ + 0.36, + 0.272, + 0.643, + 0.286 + ], + "angle": 0, + "content": "2 The University of Queensland, Australia" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.286, + 0.576, + 0.299 + ], + "angle": 0, + "content": "3 Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.3, + 0.621, + 0.313 + ], + "angle": 0, + "content": "\\(^{4}\\) DAMO Academy, Alibaba, China" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.314, + 0.657, + 0.327 + ], + "angle": 0, + "content": "5 Shenzhen Campus of Sun Yat-sen University" + }, + { + "type": "text", + "bbox": [ + 0.455, + 0.327, + 0.548, + 0.34 + ], + "angle": 0, + "content": "\\(^{6}\\) LiAuto Inc." + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.342, + 0.717, + 0.356 + ], + "angle": 0, + "content": "Hu1.Zhang@csiro.au, xin.yu@uq.edu.au, kyu@westlake.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.357, + 0.649, + 0.37 + ], + "angle": 0, + "content": "https://github.com/huzhangcs/OpenSight" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.43, + 0.744, + 0.779 + ], + "angle": 0, + "content": "Abstract. Traditional LiDAR-based object detection research primarily focuses on closed-set scenarios, which falls short in complex real-world applications. Directly transferring existing 2D open-vocabulary models with some known LiDAR classes for open-vocabulary ability, however, tends to suffer from over-fitting problems: The obtained model will detect the known objects, even presented with a novel category. In this paper, we propose OpenSight, a more advanced 2D-3D modeling framework for LiDAR-based open-vocabulary detection. OpenSight utilizes 2D-3D geometric priors for the initial discernment and localization of generic objects, followed by a more specific semantic interpretation of the detected objects. The process begins by generating 2D boxes for generic objects from the accompanying camera images of LiDAR. These 2D boxes, together with LiDAR points, are then lifted back into the LiDAR space to estimate corresponding 3D boxes. For better generic object perception, our framework integrates both temporal and spatial-aware constraints. Temporal awareness correlates the predicted 3D boxes across consecutive timestamps, recalibrating the missed or inaccurate boxes. The spatial awareness randomly places some \"precisely\" estimated 3D boxes at varying distances, increasing the visibility of generic objects. To interpret the specific semantics of detected objects, we develop a cross-modal alignment and fusion module to first align 3D features with 2D image embeddings and then fuse the aligned 3D-2D features for semantic decoding. Our experiments indicate that our method establishes state-of-the-art open-vocabulary performance on widely used 3D detection benchmarks and effectively identifies objects for new categories of interest." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.792, + 0.697, + 0.807 + ], + "angle": 0, + "content": "Keywords: OpenSight \\(\\cdot\\) Open-vocabulary \\(\\cdot\\) 3D object detection" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.149, + 0.404, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.149, + 0.538, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.259, + 0.425, + 0.271 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.154, + 0.712, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.26, + 0.648, + 0.271 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.789, + 0.352 + ], + "angle": 0, + "content": "Fig. 1: (a) A model trained on known class \"car\" correctly detects the target object (yellow) with query \"car\". However, with a new class query \"motorcycle\", the model incorrectly detects the object (red). (b) The average precision of known \"car\" and novel \"motorcycle\". The degraded performance in \"motorcycle\" (blue) indicates the limited generalization from the known classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.36, + 0.376, + 0.377 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.789, + 0.534 + ], + "angle": 0, + "content": "The growing prevalence of autonomous driving has significantly increased research interest in various 3D perception tasks [13,17,41,55], among which LiDAR-based object detection stands as a crucial component. Despite the advancements in this field, most existing research focuses on closed-set settings [19,25,36,48,64], where the obtained models are limited to predicting a set of predefined categories. This constraint makes these models insufficient for handling complex real-world scenarios, where diverse and previously unseen object categories may be encountered. Consequently, there is a growing demand to develop models capable of detecting novel categories of interest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.79, + 0.656 + ], + "angle": 0, + "content": "Considering the limited availability of annotated categories in existing LiDAR-based datasets [4,15,43], adapting established large-scale visual-language models (LVLs), e.g., CLIP [37] for open-vocabulary LiDAR detection appears promising. One paradigm is to co-embed known class object features with the image/text embeddings from LVLs, expecting alignment generalization to the novel classes. While effective in 2D image-based open-vocabulary detection, this paradigm exhibits significant overfitting in LiDAR applications, as models persist in detecting objects in known classes even when faced with a new category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.789, + 0.764 + ], + "angle": 0, + "content": "We consider some known classes, e.g., \"car\", and follow the paradigm above to illustrate this phenomenon. When given a novel category \"motorcycle\", the obtained model erroneously detects the known object \"car\", as depicted in Figure 1 (a). An analysis of average precision (AP) further reveals the diminished efficacy in novel classes, as shown in Figure 1 (b). These results underscore the limitation of such a paradigm, emphasizing the necessity for a tailored approach for LiDAR-based open-vocabulary problems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this paper, we present \"OpenSight\", a more advanced 2D-3D modeling approach for LiDAR-based open-vocabulary detection. OpenSight primarily relies on 2D and 3D geometry priors to first acquire the concept and localization of generic objects. Then, it interprets the semantics of detected objects by coembedding LiDAR features with 2D LVLs. This framework addresses all poten" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "tially novel objects explicitly and circumvents the reliance on predefined known classes, thereby avoiding the overfitting problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Specifically, we consider both temporal-aware and spatial-aware constraints to acquire the concept of generic objects and their localization. Initially, leveraging the 2D detection model and the accompanying RGB images of LiDAR, we extract 2D geometric priors for generic objects, that is, 2D bounding boxes across successive frames. These boxes are then lifted back into 3D space with LiDAR points, resulting in a collection of 3D frustums. Point clustering is then performed in each frustum to estimate the 3D bounding boxes of objects. Temporally, we correlate predicted 3D bounding boxes across frames and project the predicted 3D boxes from previous or subsequent frames onto the current frame. The missed or inaccurate objects in the current frame are then identified through overlapping comparison. This constraint boosts the continuity and consistency of generic objects over time. Spatially, we first employ large language models (LLMs) to derive size priors for general objects and then maintain 3D boxes that closely align with these size priors. The maintained boxes are randomly selected and placed at varying distances within the current frame, with the LiDAR points inside these boxes being sampled by their distance. This spatial awareness notably increases the visibility of objects, particularly those located further from the sensor. By integrating the temporal and spatial constraints, the ability to recognize and localize generic objects is significantly augmented." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.645 + ], + "angle": 0, + "content": "To decode the specific semantics of detected boxes in LiDAR space, we correlate objects across both LiDAR and image domains. Initially, by back-projecting 3D LiDAR bounding boxes onto the image plane, we identify the most closely matching 2D boxes from the Grounding DINO [27]. We then aim to minimize the feature discrepancy between the 3D LiDAR objects and the 2D matching box features. Subsequently, we fuse the optimized 3D-2D features and align the unified features with textual category embeddings from Grounding DINO to decode semantics. This dual 2D-3D feature approach leverages salient patterns in both 2D images and 3D geometry, significantly enhancing the robustness of semantic decoding. Additionally, we integrate size priors from LLMs before, recalibrating predictions that diverge from the average box size, thus enhancing the accuracy of our predictive model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.787, + 0.707 + ], + "angle": 0, + "content": "In the experiments, we first establish a variety of baselines for comparison. Then, we test our method on the widely-used 3D detection benchmarks, demonstrating that our approach achieves state-of-the-art open-vocabulary performance and effectively identifies objects for new categories of interest." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.398, + 0.748 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Closed-set 3D Object Detection. Extensive research has been proposed on LiDAR-based 3D object detection in autonomous driving. One group of works performs on LiDAR-only data [39,63], either by extracting features from raw point clouds to predict 3D bounding boxes [23,34-36,40,49] or converting the irregular LiDAR point clouds into a regular feature space [14,19,44,46,48,51,64]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.642, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.145, + 0.787, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.789, + 0.43 + ], + "angle": 0, + "content": "Fig. 2: An overview of our OpenSight framework. In generic object perception, 2D bounding boxes are initially lifted back into 3D space (initial boxes estimation). This is followed by integrated temporal and spatial awareness in the estimated 3D boxes to retrieve the missed objects or increase their visibility. Subsequently, we utilize these enhanced estimates to compute the generic object loss and optimize the 3D detection model. In specific semantic interpretation, we project the generated 3D boxes of the optimized 3D model into the image plane to get projected 2D boxes. These projections are then matched with the nearest 2D boxes identified via Grounding DINO, facilitating the alignment of 3D LiDAR object features with corresponding 2D image features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.644 + ], + "angle": 0, + "content": "Another group focuses on both LiDAR and camera modalities. Some works emphasize point-wise fusion [18,42,45,52], projecting raw LiDAR points to retrieve corresponding image features and concatenating them together. In contrast, others concentrate on feature-level fusion [22,24,53], extracting features or proposals based on LiDAR points and querying the associated image features. This line of methods has been proven to be more efficient. For example, TransFusion [3] utilizes proposals generated by LiDAR features to query the image feature for further refinement. BEVFusion [25, 28] projects the image features into BEV space by using lift-splat-shoot (LSS) and concatenates them with the LiDAR features. More recently, LoGoNet [21] constructs Local-to-Global fusion network to perform LiDAR-camera fusion at both local and global levels. However, all these approaches target the closed-set setting, restricting the range of detected object categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Open-vocabulary 2D Object Detection. The advent of large-scale pretrained models, which exhibit good text-vision alignment and zero-shot transferability, has spurred researchers to investigate the open-vocabulary problem in object detection tasks. A variety of approaches have been proposed with 2D RGB images: OVR-CNN [57] initially pre-trains a Faster R-CNN detector [38] by using image-caption pairs and text extractor, e.g., BERT [12], then fine-tunes it on downstream detection datasets. ViLD [16] distills the knowledge from a CLIP-based [37] teacher to a student detector for generalized text and image embedding alignment. Detic [62] harnesses extensive image classification data to train the classifier within a detector while localizing objects. Similar to OVR-CNN, OWL [31] first pre-trains a transformer encoder by contrastively aligning image-text pairs, then transfers it to detection tasks by attaching lightweight object classification and localization heads. Following the DETR-based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "structure, OV-DETR [56] employs CLIP-encoded text and image embeddings as queries for obtaining category-specific boxes. In contrast to the aforementioned methods, some alternatives have emerged to directly learn open-vocabulary capabilities on detection data. For example, GLIP [20] unifies the object detection and phrase grounding tasks, allowing the combined use of both detection and grounding datasets for aligning phrase semantics and regional images. Det-CLIP [50] extracts the general concepts from multiple data sources and then resorts to a description-enriched concept dictionary for enhancement. Recently, Grounding DINO [27] adopts a similar detection-grounding training strategy as GLIP but proposes a tighter text-vision fusion solution based on transformer-based detector DINO [58]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.789, + 0.584 + ], + "angle": 0, + "content": "Zero-shot Learning in 3D Data. Exploration of open-vocabulary 3D object detection has been limited due to the challenges associated with acquiring large-scale point cloud-text pairs and the inherent complexity of 3D scenarios. In the confined indoor scenario, OV-3DET [29] addresses the open-vocabulary detection problem by first localizing 3D objects and then assigning labels for the objects with open-vocabulary 2D models. However, temporal and spatial constraints are overlooked. CoDA [5] proposes to use some base classes to assist in detecting novel classes. As illustrated in Figure 1, such a paradigm in LiDAR data easily overfits the base classes. Recently, UP-VL [32] also considers the open-vocabulary case in autonomous driving, however, the categories involved are rather limited (only 3 classes). Some other works [2,6,7,47] tackle a related but distinct task of differentiating unknown data in testing from known data in training, known as the out-of-distribution (OOD) problem. Similarly, MODEST [54], OYSTER [59], and LiSe [61] study the unsupervised object discovery problem. Meanwhile, some studies [26,30,33] attempt to transfer existing 2D large-scale pre-trained models for open-vocabulary scene segmentation, while others [8-11,60] leverage the open-set ability of pre-trained models for the 3D classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Different from the above works, in this paper, we investigate open-vocabulary ability in LiDAR-based 3D object detection in the outdoor scenario, representing the real-world scenario and a much more complex scene." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.355, + 0.668 + ], + "angle": 0, + "content": "3 OpenSight" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.755 + ], + "angle": 0, + "content": "To elaborate on the LiDAR-based open-vocabulary object detection, we first briefly introduce the necessary notations for ease of understanding. Following this, we delve into the details of generic object perception in Section 3.2. Subsequently, we present the process of specific semantic interpretation in Section 3.3. The overall framework of OpenSight is presented in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.773, + 0.488, + 0.788 + ], + "angle": 0, + "content": "3.1 Notation and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.793, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The LiDAR dataset is denoted as \\(\\mathcal{D} = \\{P_i, \\{y_k\\}_{k=1}^K\\}_{i=1}^{|\\mathcal{D}|}\\), where \\(P_i\\) corresponds to the LiDAR point clouds in \\(i\\)-th frame, basically composed of \\((x_i, y_i, z_i)\\) coordinates. \\(y_k\\) represents the combination of each 3D bounding box \\(b_k\\) and its" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.271 + ], + "angle": 0, + "content": "semantic label \\( c_k \\). In open-vocabulary scenarios, \\( P_i \\) is visible while \\( y_k \\) is not. Generally, in the context of autonomous driving, each LiDAR frame is accompanied by \\( m \\) images captured from varying perspectives. We define the image dataset as \\( \\mathcal{I} = \\{(I_{ij})_{j=1}^m\\}_{i=1}^{|\\mathcal{D}|} \\) here. Image \\( I_{ij} \\) in \\( j \\)-th view and LiDAR points \\( P_i \\) could be connected by the intrinsic matrix \\( \\mathbf{K}_j \\) and extrinsic matrix \\( \\mathbf{E}_j \\). For clarity and brevity in the rest of the paper, we utilize a single LiDAR frame as an example and abbreviate \\( I_{ij} \\) as \\( I_j \\), respectively. \\( N \\) denotes the number of 3D objects, with \\( f^{3D} \\) and \\( f^{2D} \\) denotes the extracted feature in 3D and 2D." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.291, + 0.492, + 0.307 + ], + "angle": 0, + "content": "3.2 Generic Objects Perception" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.789, + 0.485 + ], + "angle": 0, + "content": "Discerning and localizing generic objects are fundamental in LiDAR-based open-vocabulary detection when ground-truth boxes and labels are absent. The model is expected to perceive the existence of a generic object before assigning specific labels. This section details how we leverage 2D and 3D geometry priors to enable this perception, including initial box estimation, and temporal-spatial awareness. Initial Box Estimation. In practical applications, each LiDAR frame is typically paired with \\( m \\) images \\( \\{I_j\\}_{j=1}^m \\), captured from various perspectives. Each image is processed through the pre-trained detection model \\( \\mathcal{I} \\), yielding a set of 2D bounding boxes \\( \\{b_{jk}\\}_{k=1}^{V_j} \\), where \\( V_j \\) represents the number of 2D objects identified in the \\( j \\)-th viewpoint. By integrating the data from all viewpoints, we compile the collective set of detected boxes \\( \\{(b_{jk}^{2D})_{k=1}^{V_j}\\}_{j=1}^m \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.529 + ], + "angle": 0, + "content": "With the obtained 2D boxes of generic objects, we try to lift them back in the LiDAR space. Specifically, we project the LiDAR point \\( P_{i} \\) into each image view by:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.529, + 0.785, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {i j} = \\mathbf {K} _ {j} \\cdot \\mathbf {E} _ {j} \\cdot P _ {i}, j = 1, \\dots , m \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.643 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}_j\\) and \\(\\mathbf{K}_j\\) are extrinsic and intrinsic matrix, respectively. The points whose projection falls within each of the predicted 2D bounding boxes \\(b_{jk}^{2D}\\) essentially constitute the 3D frustum, a conical region perceivable from a given perspective. We then apply the region growing algorithm [1] to isolate the cluster with the densest LiDAR points. Such an algorithm first randomly selects one point in the frustum and then iteratively finds its adjacent points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To estimate 3D bounding boxes from LiDAR point clusters, we first determine the extremal points along each axis by calculating the minimum and maximum coordinates. Take the X-axis for example, this is achieved by performing \\( x_{\\mathrm{min}} = \\min (\\{x_i\\}_{i = 0}^{|N_c|}) \\) and \\( x_{\\mathrm{max}} = \\max (\\{x_i\\}_{i = 0}^{|N_c|}) \\), where \\( |N_{c}| \\) is the number of points in the obtained cluster. These extremal points, identified similarly for the Y and Z axis, define the vertices of the bounding box. From these vertices, we construct the maximal bounding box \\( b_{i}^{3D} \\), enclosing all points in the cluster. For the heading of the bounding box, we traverse different angles. In each angle, we can obtain the XYZ boundary. We then shrink the boundary by a scale of 0.95, some points will be out of the obtained bounding boxes. The angle with the most outside points is regarded as the angle of the object. It is based on the assumption that for the \"tightest\" bounding box, a slight change in the angle will incur a big change in the encompassed points. The 3D boxes for generic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.148, + 0.713, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.785, + 0.3 + ], + "angle": 0, + "content": "Fig. 3: Temporal awareness: Project boxes from frame \\( t + 1 \\) to \\( t \\). By comparing with projections, we refine missed or inaccurately estimated boxes in frame \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.784, + 0.337 + ], + "angle": 0, + "content": "objects at timestamp \\( t \\) are then finalized as \\( \\mathcal{B}^t = \\{b_i^{3D}\\}_{i=0}^{N_t} \\), by aggregating all boxes from \\( m \\) perspectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.785, + 0.491 + ], + "angle": 0, + "content": "Temporal Awareness. For the LiDAR point cloud, point density decreases sharply as distance increases, leading to less reliable estimations for distant objects. However, in autonomous driving scenarios, objects that are distant in the current frame may move closer in subsequent frames. To exploit this, we leverage the temporal continuity between frames by correlating 3D bounding box predictions across successive timestamps. We take frames at timestamp \\( t \\) and \\( t + 1 \\) as an example. Specifically, given the predicted boxes \\( \\mathcal{B}^t = \\{b_i\\}_{i=0}^{N_t} \\) and \\( \\mathcal{B}^{t+1} = \\{b_i\\}_{i=0}^{N_{t+1}} \\), where \\( N_{t+1} \\) and \\( N_t \\) represent the number of boxes in \\( t + 1 \\) and \\( t \\), respectively, we project each 3D box \\( b_i \\) in \\( \\mathcal{B}^{t+1} \\) back to frame \\( t \\) using the transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.29, + 0.503, + 0.712, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\hat {b} _ {i} = \\mathbf {E} _ {\\mathrm {l i d a r} \\leftarrow \\mathrm {e g o} _ {t}} \\cdot \\mathbf {E} _ {\\mathrm {e g o} _ {t} \\leftarrow \\mathrm {g l b}} \\cdot \\mathbf {E} _ {\\mathrm {g l b} \\leftarrow \\mathrm {e g o} _ {t + 1}} \\cdot \\mathbf {E} _ {\\mathrm {e g o} _ {t + 1} \\leftarrow \\mathrm {l i d a r}} \\cdot b _ {i},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.784, + 0.579 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}_{\\mathrm{lidar}} \\leftarrow \\mathrm{ego}_t\\) and \\(\\mathbf{E}_{\\mathrm{ego}_t} \\leftarrow \\mathrm{glb}\\) are the extrinsic matrices for the ego to LiDAR and global to ego transformations at timestamp \\(t\\), and \\(\\mathbf{E}_{\\mathrm{glb}} \\leftarrow \\mathrm{ego}_{t+1}\\), \\(\\mathbf{E}_{\\mathrm{ego}_{t+1}} \\leftarrow \\mathrm{lidar}\\) denotes ego to global and LiDAR to ego at timestamp \\(t+1\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.785, + 0.717 + ], + "angle": 0, + "content": "With all boxes in \\(\\mathcal{B}^{t + 1}\\) considered, we obtain their projections as \\(\\hat{\\mathcal{B}}^{t + 1}\\) and construct the overlap matrix \\(M\\in \\mathbb{R}^{N_{t + 1}\\times N_t}\\) by calculating their Intersection over Union (IoU) with boxes in \\(\\mathcal{B}^t\\). We address two key scenarios: If a box \\(\\hat{b}_i\\) has no overlap with boxes in \\(\\mathcal{B}^t\\), its counterpart \\(b_{i}\\) in timestamp \\(t + 1\\) is considered a missed detection in the current frame and is added to the prediction at timestamp \\(t\\). Conversely, if box \\(\\hat{b}_i\\) overlaps with boxes in \\(\\mathcal{B}^t\\) and its distance exceeds a certain threshold, we perform Non-Maximum Suppression (NMS) and retain the union of the two boxes to address potentially incomplete estimations at greater distances (See Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.785, + 0.809 + ], + "angle": 0, + "content": "This process updates the predicted boxes of the current frame to \\(\\mathcal{B}_{\\mathrm{update}}^t = \\mathcal{B}^t \\bigcup \\mathcal{B}_{\\mathrm{temporal}}^t\\), which combines the original and temporally adjusted boxes. In practice, for the current frame \\(t\\), we consider both its subsequent frame at timespan \\(t + 1\\) and its precedent frame at timestamp \\(t - 1\\). That means two adjacent frames are considered. Through such a constraint, we are likely to boost the continuity and consistency of generic objects over time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Spatial Awareness. Considering the sparse object distribution in 3D LiDAR scans, where vast areas of the point cloud are often devoid of meaningful objects," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.144, + 0.71, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Fig. 4: Spatial awareness: Size priors are first generated from GPT-4 to select \"precisely\" estimated boxes and build the object bank. Randomly select objects and points from the objects bank and place them in the current frame \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.568 + ], + "angle": 0, + "content": "we also focus on the systematic integration of \"precisely\" estimated objects in each frame to enhance the detectability and representation of generic objects. To obtain these precise boxes from \\(\\mathcal{B}^t\\) in the absence of any ground truth, we utilize large language models (LLMs), such as GPT-4, for extracting useful size priors of objects. GPT-4, when prompted appropriately, can provide valuable insights into the typical sizes of objects within a specific class, e.g., the width, length, and height of a class. For instance, prompts like \"What is the average size of a \\{\\#CLASS\\}\" in the autonomous driving scene?\" enable us to collect size statistics for common classes (Details in the Appendix). Moreover, leveraging the ability of GPT-4 to process images, we also input images of the relevant objects to validate these size priors (See Figure 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.79, + 0.614 + ], + "angle": 0, + "content": "Utilizing the extracted size priors, we selectively retain 3D boxes in \\(\\mathcal{B}^t\\) that closely align with these priors. A box is reserved if its dimensions \\(w\\), \\(l\\), and \\(h\\) fall within the specified range:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.622, + 0.623, + 0.639 + ], + "angle": 0, + "content": "\\[\nw \\in [ (1 - \\tau) w _ {p r i o r}, (1 + \\tau) w _ {p r i o r} ],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.64, + 0.785, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\ell \\in [ (1 - \\tau) \\ell_ {p r i o r}, (1 + \\tau) \\ell_ {p r i o r} ], \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.66, + 0.619, + 0.677 + ], + "angle": 0, + "content": "\\[\nh \\in [ (1 - \\tau) h _ {p r i o r}, (1 + \\tau) h _ {p r i o r} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.789, + 0.744 + ], + "angle": 0, + "content": "where \\(\\tau\\) represents the error margin. \\(w_{prior}\\), \\(l_{prior}\\) and \\(h_{prior}\\) are from LLMs-derived priors. This process results in an object bank \\(O = \\{(b_i^{3D}, P_i)\\}_{i=0}^{N_o}\\), where \\(b_i^{3D}\\) is the reserved box, \\(P_i\\) comprises the LiDAR points within each box and \\(N_o\\) is the total number of reserved boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.789, + 0.806 + ], + "angle": 0, + "content": "For the current frame, we randomly select boxes from bank \\( O \\) and integrate them at varying distances from the ego vehicle, ensuring diverse spatial distribution. The LiDAR points within each box are sampled according to their distance, with the sampling ratio defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.813, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nr _ {s a m p l i n g} = \\min (1, \\frac {d - d _ {n e w}}{d - d _ {o r i}}), \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "where \\(d\\) denotes the total range in the point cloud, \\(d_{ori}\\) is the original distance of the object, \\(d_{new}\\) is the distance post-placement. The selected boxes are denoted as \\(\\mathcal{B}_{\\mathrm{spatial}}^t\\) and the collective set of boxes in the current frame is represented as \\(\\mathcal{B}_{\\mathrm{final}}^t = \\mathcal{B}^t \\cup \\mathcal{B}_{\\mathrm{temporal}}^t \\cup \\mathcal{B}_{\\mathrm{spatial}}^t\\). This strategy aims to substantially enhance the visibility of generic objects in LiDAR data and the perception ability of the obtained model." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.239, + 0.784, + 0.318 + ], + "angle": 0, + "content": "Loss for Generic Object Perception. With the obtained 3D boxes \\(\\mathcal{B}_{\\mathrm{final}}^t = \\{b_k^{3D}\\}_{k=1}^{V_{\\mathrm{final}}}\\), we formulate the loss for generic object concept acquisition and localization. Utilizing a DETR-like structure, we extract features from LiDAR points and apply bipartite matching to determine the optimal permutation \\(\\hat{\\sigma}\\), the loss is thus expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.326, + 0.785, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum^ {N} \\left[ \\mathcal {L} _ {\\text {f o c a l}} \\left(s _ {\\hat {\\sigma} _ {i} (k)}\\right) + \\mathcal {L} _ {\\mathrm {L} _ {1}} \\left(b _ {\\hat {\\sigma} _ {i} (k)}, b _ {k} ^ {3 D}\\right) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.36, + 0.784, + 0.422 + ], + "angle": 0, + "content": "where \\( s_{\\hat{\\sigma}_i(k)} \\) denotes whether it is an object or not, \\( \\hat{\\sigma}_i(k) \\) is the element in \\( \\hat{\\sigma} \\). \\( \\mathcal{L}_{\\mathrm{focal}} \\) constrains the level of \"objectness\" of generated 3D boxes, \\( \\mathcal{L}_{\\mathrm{box}} \\) considers the constraint on boxes. With this composite loss optimized, we obtain a model with the ability to perceive generic objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.44, + 0.536, + 0.456 + ], + "angle": 0, + "content": "3.3 Specific Semantics Interpretation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.462, + 0.784, + 0.569 + ], + "angle": 0, + "content": "To interpret the specific semantics of detected generic objects in LiDAR, we correlate objects across both LiDAR and image domains. We apply the model after generic object perception to a given LiDAR frame, producing boxes \\(\\{b_i^{3D}\\}_{i=0}^N\\) and their corresponding features \\(\\{f_i^{3D}\\}_{i=0}^N\\). Instead of using the image-level LVLs, e.g., CLIP, we opt for object-level LVLs, such as Grounding DINO [27], which provides 2D boxes \\(\\{b_{i=0}^{2D}\\}_{i=0}^{N_{2D}}\\), 2D object features \\(\\{f_i^{2D}\\}_{i=0}^{N_{2D}}\\) and text embeddings \\(\\{t_i\\}_{i=0}^{N_{2D}}\\) simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.784, + 0.601 + ], + "angle": 0, + "content": "For each detected 3D box \\( b_{i}^{3D} \\), we project it to each view image with the corresponding intrinsic matrix \\( K_{j} \\) and extrinsic matrix \\( E_{j} \\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.609, + 0.785, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\hat {b} _ {i j} ^ {2 D} = \\mathbf {K} _ {j} \\cdot \\mathbf {E} _ {j} \\cdot b _ {i} ^ {3 D}, j = 1, \\dots , m \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.636, + 0.784, + 0.736 + ], + "angle": 0, + "content": "The \\(\\hat{b}_{ij}^{2D}\\) is further filtered and only those lying in the image are reserved. For the reserved box \\(\\hat{b}_{ij}^{2D}\\), we compute the IoU with the 2D boxes \\(\\{b_{i}^{2D}\\}_{i=0}^{N_{2D}}\\) from Grounding DINO and find the most close one \\(b_{j}^{2D}\\). The corresponding text embeddings \\(t_{j}^{2D}\\) feature is retrieved. We then try to minimize the discrepancy between \\(f_{i}^{3D}\\) and all 2D features that have the same text embeddings \\(t_{j}^{2D}\\). Mathematically, it is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.743, + 0.784, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {a l i g n}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\sum_ {j = 0} ^ {N _ {i}} e ^ {f _ {i} ^ {3 D} \\cdot f _ {j} ^ {2 D} / \\delta}}{\\sum_ {k = 0} ^ {N ^ {2 D}} e ^ {f _ {i} ^ {3 D} \\cdot f _ {k} ^ {2 D} / \\delta}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.793, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\(N_{i}\\) denotes all 2D features that have the same text embeddings \\(t_j^{2D}\\) for \\(b_i^{3D}\\). Once the object-level alignment is finished, we can directly perform open-vocabulary LiDAR detection with the LiDAR input." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "2D-3D Integration. Considering that paired RGB images with LiDAR are always available in autonomous driving, we thus can perform 2D-3D feature integration. For each detected object, we have optimized 3D features \\( f_{i}^{3D} \\) and identified the best matched 2D boxes feature \\( f_{i}^{2D} \\), the similarity between 3D-2D features with the text embeddings in \\( \\{t_i\\}_{i = 0}^{N_{2D}} \\) are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.229, + 0.785, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s _ {j} ^ {3 D} = \\cos \\left(f _ {i} ^ {3 D}, t _ {j}\\right), \\\\ 2 ^ {D} = \\left(3 ^ {D} + 1\\right) \\dots 0 \\quad \\text {v} ^ {3 D} \\end{array} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.25, + 0.623, + 0.268 + ], + "angle": 0, + "content": "\\[\ns _ {j} ^ {2 D} = \\cos \\left(f _ {i} ^ {2 D}, t _ {j}\\right), j = 0, \\dots , N ^ {2 D}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.341 + ], + "angle": 0, + "content": "We obtain the maximum value for 2D and 3D feature by \\( s^{3D} = \\max_j\\{s_j^{3D}\\}_{j = 0}^{N_{2D}} \\), \\( s^{2D} = \\max_j\\{s_j^{2D}\\}_{j = 0}^{N_{2D}} \\), along with their semantic labels \\( c^{3D} \\) and \\( c^{2D} \\). The final label for object \\( i \\) is determined by comparing \\( s_i^{2D} \\) and \\( s_i^{3D} \\), selecting the label associated with the larger value." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Large Language Model (LLMs) Post-processing. The size priors from LLMs can be used to assist in recalibrating predictions. By matching the 3D IoU of predicted boxes with these size priors, we can determine the most matching prior and the corresponding semantic label. We then incorporate the predictions from size prior with results from 2D-3D integration by selecting label with the highest similarity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.451, + 0.341, + 0.464 + ], + "angle": 0, + "content": "3.4 Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.563 + ], + "angle": 0, + "content": "we first input the LiDAR point cloud to the generic object detector and extract 3D object features. We then align these features with text embeddings from Grounding DINO [27] to determine semantic labels. Additionally, when RGB images are available, we further improve the alignment ability by integrating 3D and 2D features or use the LLMs prior (size priors from GPT-4) to calibrate the predicted labels, thereby improving the accuracy of obtained semantic labels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.377, + 0.598 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.608, + 0.444, + 0.624 + ], + "angle": 0, + "content": "4.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Dataset. We conduct the open-vocabulary experiments on nuScenes [4], a large-scale, outdoor dataset featuring multimodal data modalities. Each frame in the dataset contains data from one 32-beam LiDAR and six cameras, offering a comprehensive view of the surroundings. The dataset comprises 1,000 scenes in total, with up to 1.4 million annotated 3D bounding boxes spanning 10 classes. We reserve all the classes as novel to evaluate the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation Metrics. For quantitative evaluation, we initially employ Average Precision (AP) on each class and mean Average Precision (mAP) across all novel classes. Considering that LiDAR point density varies with the distance, we introduce distance-based metrics: mAP-near, mAP-midrange, and mAP-far. Similarly, acknowledging the size disparities among objects in different classes (e.g., 'bus' vs. 'traffic_cone'), we categorize classes into 'large', 'medium', and 'small' and propose mAP-large, mAP-medium, and mAP-small. More details are in the Appendix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Results on NuScenes dataset. All classes are reserved as novel classes. AP of each class and mAP across all classes are shown. \"Car\" in CoDA is the known class and is fully supervised ('Cone': 'Traffic_cone', 'C.V': 'Construction_ vehicle', 'Motor.'] 'Motorcycle', 'Ped.'] 'Pedestrian')." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.205, + 0.784, + 0.355 + ], + "angle": 0, + "content": "
MethodmAPNDSCarTruckC.V.BusTrailerBarrierBicycleConeMotor.Ped.
Supervised (box+label)57.361.786.038.818.363.334.270.546.770.259.585.2
Supervised (box only)34.239.053.220.17.420.50.750.137.852.639.160.8
Detic-3D [62]5.812.715.12.92.14.20.00.16.20.86.121.1
OWL-ViT-3D [31]5.412.414.22.51.84.00.00.15.21.15.519.4
OV-PointCLIP [60]0.15.00.00.00.00.00.00.00.70.20.00.1
OV-PointCLIPv2 [65]2.810.90.02.70.00.00.022.00.00.20.03.2
CLIP-3D [37]7.214.29.62.82.12.60.00.111.514.49.818.9
OV-3DET [29]5.712.016.12.90.31.30.00.31.811.02.021.1
SeCondPoint [26]8.215.015.12.32.02.00.00.010.914.915.219.3
3DGenZ [30]7.514.914.63.52.33.20.00.010.514.28.717.9
Openscene [33]8.815.417.23.92.23.00.10.212.814.010.823.4
UP-VL [32]11.317.219.47.94.33.10.40.514.115.616.830.8
CoDA [5]10.316.185.51.90.02.00.00.03.90.05.93.5
OpenSight (Ours)22.923.325.311.68.75.10.842.226.125.630.452.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.284, + 0.362, + 0.719, + 0.377 + ], + "angle": 0, + "content": "Table 2: Performance w.r.t mAP-near, mAP-medium, mAP-far." + }, + { + "type": "table", + "bbox": [ + 0.328, + 0.379, + 0.668, + 0.487 + ], + "angle": 0, + "content": "
MethodmAP-nearmAP-midrangemAP-far
Detric-3D [62]18.59.35.8
OWL-ViT-3D [31]17.38.25.4
OV-PointCLIP [60]0.20.10.1
OV-PointCLIPv2 [65]4.53.62.8
CLIP-3D [37]18.811.27.2
OV-3DET [29]16.19.95.7
CoDA [5]18.216.610.3
OpenSight (Ours)28.026.522.9
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.506, + 0.341, + 0.521 + ], + "angle": 0, + "content": "4.2 Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.581 + ], + "angle": 0, + "content": "As there are no direct baselines addressing open-vocabulary LiDAR-based 3D detection, we primarily construct the baselines based on well-known works [5, 29, 37, 62, 65] and compare with OpenSight. Specifically, they are:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.791, + 0.644 + ], + "angle": 0, + "content": "Fully-supervised training. Initially, we perform fully supervised training across all categories, establishing an upper bound for the open-vocabulary detection performance. We consider using both ground-truth boxes and labels ('Supervised (box+label)') or just using ground-truth boxes ('Supervised (box only)')." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.763 + ], + "angle": 0, + "content": "2D open-vocabulary methods. Detic [62] is a pioneer work in the realm of 2D open-vocabulary object detection tasks and can detect 21K classes. It is thus desirable for our study here. OWL-ViT [31] decodes the object labels and boxes directly from the pre-trained features and shows good robustness in real-world applications. With the 2D detection results of either Detic or OWL-ViT, we adopt the same procedure in Section 3.2 to obtain the corresponding 3D bounding boxes in LiDAR space. These two baselines are denoted as Detic-3D and OWL-ViT-3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3D open-vocabulary methods. We then consider the latest open-vocabulary point cloud classification methods and a variety of indoor open-vocabulary works to evaluate their performance in our tasks here. Specifically, we utilize a fully-supervised method to generate pseudo boxes and employ PointCLIP [60], Point-CLIPv2 [65], denoted as \"OV-PointCLIP\", \"OV-PointCLIPv2\". We also combine" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.274, + 0.145, + 0.728, + 0.159 + ], + "angle": 0, + "content": "Table 3: Performance w.r.t mAIP-large, mAIP-medium, mAIP-small." + }, + { + "type": "table", + "bbox": [ + 0.324, + 0.162, + 0.672, + 0.269 + ], + "angle": 0, + "content": "
MethodmAP-largemAP-mediummAP-small
Detric-3D [62]5.34.111.0
OWL-ViT-3D [31]4.93.710.3
OV-PointCLIP [60]0.00.30.2
OV-PointCLIPv2 [65]0.67.31.6
CLIP-3D [37]3.47.116.7
OV-3DET [29]4.11.416.0
CoDA [5]17.94.91.8
OpenSight (Ours)11.532.939.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.273, + 0.784, + 0.289 + ], + "angle": 0, + "content": "Table 4: Results with temporal and spatial awareness w.r.t mAP-{near, medium, far}." + }, + { + "type": "table", + "bbox": [ + 0.305, + 0.291, + 0.691, + 0.359 + ], + "angle": 0, + "content": "
#TemporalSpatialmAP-nearmAP-midrangemAP-far
(1)27.124.822.1
(2)✓27.425.922.5
(3)✓27.526.022.7
(4)✓✓28.026.522.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.44 + ], + "angle": 0, + "content": "our predicted boxes and CLIP [37] for this task (\"CLIP-3D\"). Meanwhile, we adapt the latest works OV-3DET [29] and CoDA [5] in indoor settings and UP-VL [32] to our task here. We also consider the segmentation-based methods, e.g., SeCondPoint [26], 3DGenZ [30], and OpenScene [33], for comparison." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.464, + 0.367, + 0.478 + ], + "angle": 0, + "content": "4.3 Main results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.719 + ], + "angle": 0, + "content": "In this section, we discuss the performance of our model on novel classes, as summarized in Table 1. Despite the lack of 3D ground truth supervision, our framework demonstrates strong overall open-vocabulary ability across classes. Notably, it achieves exceptional performance in some classes, for example, 52.5 AP in 'pedestrian' and 42.2 AP in 'Barrier'. Unfortunately, we also notice the limited performance in 'Trailer', with 0.8 AP. To analyze this, we compare with 'Supervised (box only)', where groundtruth boxes are used and only 0.7 AP is achieved in 'Trailer', we thus attribute the reason to label alignment with Grounding DINO rather than box inaccuracy. Grounding DINO offers limited help in fine-grained classes. Our model also shows superior performance when compared with the baselines. Specifically, we observe that PointCLIP is very limited in detecting novel objects and CLIP-3D achieves the best overall performance with our detected boxes. CoDA, while effective in known classes like 'car', falls short in novel class adaptation, averaging only \\(1.9\\%\\) accuracy across novel categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our more detailed evaluations, using distance-based metrics and size-based metrics, are presented in Tables 2 and 3, respectively. The evaluation in Table 2 reveals a trend where detection accuracy is higher for objects closer to the sensor. This is attributed to their denser point clouds and more defined geometric features. Conversely, detection accuracy decreases with increasing distance, likely due to sparser data and less distinct object outlines. Regarding size-based performance (Table 3), our model shows superior results for medium and small-sized objects. Large objects present lower detection rates, mainly due to their" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.145, + 0.496, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.145, + 0.741, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.234, + 0.495, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.235, + 0.741, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.785, + 0.368 + ], + "angle": 0, + "content": "Fig. 5: Visualization of detected novel classes in different LiDAR frames. (a) Detected 'barrier', 'traffic_cone', and 'pedestrian'; (b) Detected 'car' and 'pedestrian'; (c) Detected 'car' and 'motorcycle'; (d) Detected 'barrier' and 'pedestrian'." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.379, + 0.784, + 0.407 + ], + "angle": 0, + "content": "Table 5: Results with direct alignment ('D.A.'), 3D-2D feature integration ('I.G.'), and LLMs priors." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.41, + 0.697, + 0.467 + ], + "angle": 0, + "content": "
#D.AI.G.PriorsmAP-largemAP-mediummAP-small
(1)✓9.431.838.1
(2)✓✓10.932.738.8
(3)✓✓✓11.532.939.0
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.472, + 0.784, + 0.503 + ], + "angle": 0, + "content": "incomplete LiDAR scans. Additionally, the fine-grained categorization of large vehicles, such as 'Constructionvehicle', complicates their accurate detection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.522, + 0.55, + 0.537 + ], + "angle": 0, + "content": "4.4 Qualitative results on novel classes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.785, + 0.633 + ], + "angle": 0, + "content": "As shown in Figure 5, our framework demonstrates a strong ability to detect novel objects by inputting the interested categories. We visualize the results from both BEV and a specific perspective. For example, our method successfully identifies the 'pedestrian', 'barrier', and 'traffic_cone' in example (a), and the 'car' and 'motorcycle' in (c). These results also indicate that our method exhibits superior open-vocabulary detection ability across different scenes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.653, + 0.335, + 0.667 + ], + "angle": 0, + "content": "4.5 Ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.704 + ], + "angle": 0, + "content": "In this part, we present a comprehensive analysis of OpenSight and examine the effectiveness of each component." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.795 + ], + "angle": 0, + "content": "Components in Generic Object Perception. We decouple the components in generic object perception and record the results in Table 4. The overall improvement may not be so obvious, however, we improve more for mAP-medium and mAP-far with these techniques. For example, temporal improves 1.1 AP for mAP-medium while 0.3 AP for mAP-near. Both modules interact positively with each other and further improve the results when jointly introduced." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Components in Semantic Interpretation. We analyze the modules of direct alignment, 3D-2D feature integration, and LLMs priors in semantic interpretation and record the results in Table 5. The performance is consistently improved" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.145, + 0.689, + 0.159 + ], + "angle": 0, + "content": "Table 6: Results of image-level and object-level models." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.162, + 0.71, + 0.215 + ], + "angle": 0, + "content": "
mAP-near-midrange-far-large-medium-small
CLIP19.011.98.23.98.218.4
Grounding DINO28.026.522.911.532.939.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.228, + 0.712, + 0.241 + ], + "angle": 0, + "content": "Table 7: AP and AR of some classes under different threshold." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.244, + 0.709, + 0.302 + ], + "angle": 0, + "content": "
#CLASSAP-0.5AP-1AP-2AR-0.5AR-1AR-2
Car4.718.345.940.069.084.4
Truck0.23.013.313.230.051.1
Pedestrian38.851.657.158.469.374.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.34 + ], + "angle": 0, + "content": "with integration and priors introduced. Compared with LLMs priors, 3D-2D integration brings more benefits, for example, 0.9 vs. 0.2 in mAP-medium." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.785, + 0.461 + ], + "angle": 0, + "content": "Image-level vs. Object-level. We also investigate the image-level CLIP [37] model in semantic interpretation. The results are recorded in Table 6. Compared with our adopted object-level Grounding DINO, we notice that object-level alignment shows much better results from either a distance perspective or a size perspective. For example, 28.0 AP of near objects is achieved in object-level alignment, 9 points higher than image-level alignment. We analyze that CLIP requires more accurate bounding box prediction, in comparison with object-level alignment which only requires a nearest box." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.787, + 0.613 + ], + "angle": 0, + "content": "Importance of Bounding Boxes Existence. Existing open-vocabulary methods focus more on the alignment between 3D features and 2D features. However, in LiDAR-based object detection, we emphasize that accurate bounding boxes are essential before alignment. We select some classes and investigate both average precision (AP) and average recall (AR) under different thresholds. The results are recorded in Table 7. Compared with AP, we observe a higher recall value in different threshold, indicating that although we can recall the correct objects, more objects with incorrect sizes are also recalled. Such results motivate that if we can obtain a more precise estimation of objects, the results will be improved further." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.36, + 0.652 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.82 + ], + "angle": 0, + "content": "We present OpenSight, a simple and versatile open-vocabulary framework for LiDAR-Based object detection. In the absence of ground-truth boxes and semantic labels, OpenSight first learns to identify and localize a wide range of generic objects. To boost the ability of generic object perception, we incorporate temporal and spatial awareness techniques that recalibrate and improve the visibility of estimated 3D boxes. We also align LiDAR object features with the 2D features from the Grounding DINO for semantic interpretation. The extensive experiments demonstrate the effectiveness and generalization ability of the proposed framework, paving the way for further exploration in LiDAR-based open-vocabulary challenges." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.231, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.395, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.786, + 0.22 + ], + "angle": 0, + "content": "This research is funded in part by ARC-Discovery grant (DP220100800 to XY) and ARC-DECRA grant (DE230100477 to XY). We thank all anonymous reviewers and ACs for their constructive suggestions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.24, + 0.324, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.295 + ], + "angle": 0, + "content": "1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.336 + ], + "angle": 0, + "content": "2. Alliegro, A., Cappio Borlino, F., Tommasi, T.: 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems 35, 21228-21240 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.39 + ], + "angle": 0, + "content": "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "5. Cao, Y., Zeng, Y., Xu, H., Xu, D.: Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. arXiv preprint arXiv:2310.02960 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.5 + ], + "angle": 0, + "content": "6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Open-set 3d object detection. In: 2021 International Conference on 3D Vision (3DV). pp. 869-878. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.554 + ], + "angle": 0, + "content": "7. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. pp. 318-334. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "8. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. arXiv preprint arXiv:1907.06371 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.785, + 0.636 + ], + "angle": 0, + "content": "9. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 923-933 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.676 + ], + "angle": 0, + "content": "0. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision 130(10), 2364-2384 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "1. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Douillard, B., Underwood, J., Kuntz, N., Vlaskine, V., Quadros, A., Morton, P., Frenkel, A.: On the segmentation of 3d lidar point clouds. In: 2011 IEEE International Conference on Robotics and Automation. pp. 2798-2805. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "16. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=1L3lnMbR4WU" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "18. Huang, T., Liu, Z., Chen, X., Bai, X.: Epnet: Enhancing point features with image semantics for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 35-52. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "19. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.42 + ], + "angle": 0, + "content": "20. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "21. Li, X., Ma, T., Hou, Y., Shi, B., Yang, Y., Liu, Y., Wu, X., Chen, Q., Li, Y., Qiao, Y., et al.: Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. arXiv preprint arXiv:2303.03595 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.515 + ], + "angle": 0, + "content": "22. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.556 + ], + "angle": 0, + "content": "23. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.597 + ], + "angle": 0, + "content": "24. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "25. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "26. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.745 + ], + "angle": 0, + "content": "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "29. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1190-1199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3d point clouds. In: 2021 International Conference on 3D Vision (3DV). pp. 992-1002. IEEE (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.258 + ], + "angle": 0, + "content": "32. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8602-8612 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.787, + 0.298 + ], + "angle": 0, + "content": "33. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. arXiv preprint arXiv:2211.15654 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.339 + ], + "angle": 0, + "content": "34. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 918-927 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.379 + ], + "angle": 0, + "content": "35. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "36. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.473 + ], + "angle": 0, + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.556 + ], + "angle": 0, + "content": "39. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.596 + ], + "angle": 0, + "content": "40. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 770-779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.787, + 0.651 + ], + "angle": 0, + "content": "41. Simon, M., Amende, K., Kraus, A., Honer, J., Samann, T., Kaulbersch, H., Milz, S., Michael Gross, H.: Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "42. Sindagi, V.A., Zhou, Y., Tuzel, O.: Mvx-net: Multimodal voxelnet for 3d object detection. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7276-7282. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.746 + ], + "angle": 0, + "content": "43. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.787, + 0.8 + ], + "angle": 0, + "content": "44. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "45. Wang, C., Ma, C., Zhu, M., Yang, X.: Pointaugmenting: Cross-modal augmentation for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11794-11803 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "H. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "46. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 18-34. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.786, + 0.244 + ], + "angle": 0, + "content": "47. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.786, + 0.271 + ], + "angle": 0, + "content": "48. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.786, + 0.311 + ], + "angle": 0, + "content": "49. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.786, + 0.352 + ], + "angle": 0, + "content": "50. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.786, + 0.393 + ], + "angle": 0, + "content": "51. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.786, + 0.419 + ], + "angle": 0, + "content": "52. Yin, T., Zhou, X., Krahenbuhl, P.: Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems 34, 16494-16507 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.786, + 0.473 + ], + "angle": 0, + "content": "53. Yoo, J.H., Kim, Y., Kim, J., Choi, J.W.: 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 720-736. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.786, + 0.528 + ], + "angle": 0, + "content": "54. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1130-1140 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.786, + 0.569 + ], + "angle": 0, + "content": "55. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access 8, 58443-58469 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.786, + 0.623 + ], + "angle": 0, + "content": "56. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 106-122. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.786, + 0.664 + ], + "angle": 0, + "content": "57. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.786, + 0.704 + ], + "angle": 0, + "content": "58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.786, + 0.759 + ], + "angle": 0, + "content": "59. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9317-9328 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "60. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "61. Zhang, R., Zhang, H., Yu, H., Zheng, Z.: Approaching outside: Scaling unsupervised 3d object detection from 2d scene. In: Proceedings of the European conference on computer vision (ECCV) (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "62. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 350-368. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.788, + 0.245 + ], + "angle": 0, + "content": "63. Zhou, Y., Sun, P., Zhang, Y., Anguelov, D., Gao, J., Ouyang, T., Guo, J., Ngiam, J., Vasudevan, V.: End-to-end multi-view fusion for 3d object detection in lidar point clouds. In: Conference on Robot Learning. pp. 923-932. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "64. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "65. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_origin.pdf b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e9c7065c6f5f805800604cf95a7297cbea42546e --- /dev/null +++ b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/f9a42dd5-034f-4c08-a554-2875cca4969b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73688394c61519753e54a5ea81a5a3e6a9189284ab1b721c5d4e3146e90f804b +size 7287715 diff --git a/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/full.md b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..da861b08d68d1bb1efe8beafa9bc1a9a36c12a67 --- /dev/null +++ b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/full.md @@ -0,0 +1,350 @@ +# OpenSight: A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection + +Hu Zhang $^{1}$ , Jianhua Xu $^{4}$ , Tao Tang $^{5}$ , Haiyang Sun $^{6}$ + +Xin $\mathrm{Yu}^{2*}$ , Zi Huang $^{2*}$ , and Kaicheng $\mathrm{Yu}^{3}$ + +1 CSIRO DATA61, Australia + +2 The University of Queensland, Australia + +3 Westlake University + +$^{4}$ DAMO Academy, Alibaba, China + +5 Shenzhen Campus of Sun Yat-sen University + +$^{6}$ LiAuto Inc. + +Hu1.Zhang@csiro.au, xin.yu@uq.edu.au, kyu@westlake.edu.cn + +https://github.com/huzhangcs/OpenSight + +Abstract. Traditional LiDAR-based object detection research primarily focuses on closed-set scenarios, which falls short in complex real-world applications. Directly transferring existing 2D open-vocabulary models with some known LiDAR classes for open-vocabulary ability, however, tends to suffer from over-fitting problems: The obtained model will detect the known objects, even presented with a novel category. In this paper, we propose OpenSight, a more advanced 2D-3D modeling framework for LiDAR-based open-vocabulary detection. OpenSight utilizes 2D-3D geometric priors for the initial discernment and localization of generic objects, followed by a more specific semantic interpretation of the detected objects. The process begins by generating 2D boxes for generic objects from the accompanying camera images of LiDAR. These 2D boxes, together with LiDAR points, are then lifted back into the LiDAR space to estimate corresponding 3D boxes. For better generic object perception, our framework integrates both temporal and spatial-aware constraints. Temporal awareness correlates the predicted 3D boxes across consecutive timestamps, recalibrating the missed or inaccurate boxes. The spatial awareness randomly places some "precisely" estimated 3D boxes at varying distances, increasing the visibility of generic objects. To interpret the specific semantics of detected objects, we develop a cross-modal alignment and fusion module to first align 3D features with 2D image embeddings and then fuse the aligned 3D-2D features for semantic decoding. Our experiments indicate that our method establishes state-of-the-art open-vocabulary performance on widely used 3D detection benchmarks and effectively identifies objects for new categories of interest. + +Keywords: OpenSight $\cdot$ Open-vocabulary $\cdot$ 3D object detection + +![](images/d096d890463e46c7b2d841608429504d3e4e80ed044a6032ffcb977886f8b585.jpg) +(a) + +![](images/4ce7beea5336b8b0e18db1fe60cbbfdb0fcf1831bb0769225ed3c9a71509d8e5.jpg) + +![](images/f9373995613eb9f30f897fbd57d7ac2b65d0cd9ff8401e853e810cb1ffdecf47.jpg) +(b) +Fig. 1: (a) A model trained on known class "car" correctly detects the target object (yellow) with query "car". However, with a new class query "motorcycle", the model incorrectly detects the object (red). (b) The average precision of known "car" and novel "motorcycle". The degraded performance in "motorcycle" (blue) indicates the limited generalization from the known classes. + +# 1 Introduction + +The growing prevalence of autonomous driving has significantly increased research interest in various 3D perception tasks [13,17,41,55], among which LiDAR-based object detection stands as a crucial component. Despite the advancements in this field, most existing research focuses on closed-set settings [19,25,36,48,64], where the obtained models are limited to predicting a set of predefined categories. This constraint makes these models insufficient for handling complex real-world scenarios, where diverse and previously unseen object categories may be encountered. Consequently, there is a growing demand to develop models capable of detecting novel categories of interest. + +Considering the limited availability of annotated categories in existing LiDAR-based datasets [4,15,43], adapting established large-scale visual-language models (LVLs), e.g., CLIP [37] for open-vocabulary LiDAR detection appears promising. One paradigm is to co-embed known class object features with the image/text embeddings from LVLs, expecting alignment generalization to the novel classes. While effective in 2D image-based open-vocabulary detection, this paradigm exhibits significant overfitting in LiDAR applications, as models persist in detecting objects in known classes even when faced with a new category. + +We consider some known classes, e.g., "car", and follow the paradigm above to illustrate this phenomenon. When given a novel category "motorcycle", the obtained model erroneously detects the known object "car", as depicted in Figure 1 (a). An analysis of average precision (AP) further reveals the diminished efficacy in novel classes, as shown in Figure 1 (b). These results underscore the limitation of such a paradigm, emphasizing the necessity for a tailored approach for LiDAR-based open-vocabulary problems. + +In this paper, we present "OpenSight", a more advanced 2D-3D modeling approach for LiDAR-based open-vocabulary detection. OpenSight primarily relies on 2D and 3D geometry priors to first acquire the concept and localization of generic objects. Then, it interprets the semantics of detected objects by coembedding LiDAR features with 2D LVLs. This framework addresses all poten + +tially novel objects explicitly and circumvents the reliance on predefined known classes, thereby avoiding the overfitting problem. + +Specifically, we consider both temporal-aware and spatial-aware constraints to acquire the concept of generic objects and their localization. Initially, leveraging the 2D detection model and the accompanying RGB images of LiDAR, we extract 2D geometric priors for generic objects, that is, 2D bounding boxes across successive frames. These boxes are then lifted back into 3D space with LiDAR points, resulting in a collection of 3D frustums. Point clustering is then performed in each frustum to estimate the 3D bounding boxes of objects. Temporally, we correlate predicted 3D bounding boxes across frames and project the predicted 3D boxes from previous or subsequent frames onto the current frame. The missed or inaccurate objects in the current frame are then identified through overlapping comparison. This constraint boosts the continuity and consistency of generic objects over time. Spatially, we first employ large language models (LLMs) to derive size priors for general objects and then maintain 3D boxes that closely align with these size priors. The maintained boxes are randomly selected and placed at varying distances within the current frame, with the LiDAR points inside these boxes being sampled by their distance. This spatial awareness notably increases the visibility of objects, particularly those located further from the sensor. By integrating the temporal and spatial constraints, the ability to recognize and localize generic objects is significantly augmented. + +To decode the specific semantics of detected boxes in LiDAR space, we correlate objects across both LiDAR and image domains. Initially, by back-projecting 3D LiDAR bounding boxes onto the image plane, we identify the most closely matching 2D boxes from the Grounding DINO [27]. We then aim to minimize the feature discrepancy between the 3D LiDAR objects and the 2D matching box features. Subsequently, we fuse the optimized 3D-2D features and align the unified features with textual category embeddings from Grounding DINO to decode semantics. This dual 2D-3D feature approach leverages salient patterns in both 2D images and 3D geometry, significantly enhancing the robustness of semantic decoding. Additionally, we integrate size priors from LLMs before, recalibrating predictions that diverge from the average box size, thus enhancing the accuracy of our predictive model. + +In the experiments, we first establish a variety of baselines for comparison. Then, we test our method on the widely-used 3D detection benchmarks, demonstrating that our approach achieves state-of-the-art open-vocabulary performance and effectively identifies objects for new categories of interest. + +# 2 Related Works + +Closed-set 3D Object Detection. Extensive research has been proposed on LiDAR-based 3D object detection in autonomous driving. One group of works performs on LiDAR-only data [39,63], either by extracting features from raw point clouds to predict 3D bounding boxes [23,34-36,40,49] or converting the irregular LiDAR point clouds into a regular feature space [14,19,44,46,48,51,64]. + +![](images/1c80c38be6ce7555c4a05f7995a94400e9932325d5737e08c25aee22ffc763d7.jpg) +Fig. 2: An overview of our OpenSight framework. In generic object perception, 2D bounding boxes are initially lifted back into 3D space (initial boxes estimation). This is followed by integrated temporal and spatial awareness in the estimated 3D boxes to retrieve the missed objects or increase their visibility. Subsequently, we utilize these enhanced estimates to compute the generic object loss and optimize the 3D detection model. In specific semantic interpretation, we project the generated 3D boxes of the optimized 3D model into the image plane to get projected 2D boxes. These projections are then matched with the nearest 2D boxes identified via Grounding DINO, facilitating the alignment of 3D LiDAR object features with corresponding 2D image features. + +![](images/2743e696be55661dbd6a8a2368c33f7ab29b301dff040f0dc649f03a3440fd70.jpg) + +Another group focuses on both LiDAR and camera modalities. Some works emphasize point-wise fusion [18,42,45,52], projecting raw LiDAR points to retrieve corresponding image features and concatenating them together. In contrast, others concentrate on feature-level fusion [22,24,53], extracting features or proposals based on LiDAR points and querying the associated image features. This line of methods has been proven to be more efficient. For example, TransFusion [3] utilizes proposals generated by LiDAR features to query the image feature for further refinement. BEVFusion [25, 28] projects the image features into BEV space by using lift-splat-shoot (LSS) and concatenates them with the LiDAR features. More recently, LoGoNet [21] constructs Local-to-Global fusion network to perform LiDAR-camera fusion at both local and global levels. However, all these approaches target the closed-set setting, restricting the range of detected object categories. + +Open-vocabulary 2D Object Detection. The advent of large-scale pretrained models, which exhibit good text-vision alignment and zero-shot transferability, has spurred researchers to investigate the open-vocabulary problem in object detection tasks. A variety of approaches have been proposed with 2D RGB images: OVR-CNN [57] initially pre-trains a Faster R-CNN detector [38] by using image-caption pairs and text extractor, e.g., BERT [12], then fine-tunes it on downstream detection datasets. ViLD [16] distills the knowledge from a CLIP-based [37] teacher to a student detector for generalized text and image embedding alignment. Detic [62] harnesses extensive image classification data to train the classifier within a detector while localizing objects. Similar to OVR-CNN, OWL [31] first pre-trains a transformer encoder by contrastively aligning image-text pairs, then transfers it to detection tasks by attaching lightweight object classification and localization heads. Following the DETR-based + +structure, OV-DETR [56] employs CLIP-encoded text and image embeddings as queries for obtaining category-specific boxes. In contrast to the aforementioned methods, some alternatives have emerged to directly learn open-vocabulary capabilities on detection data. For example, GLIP [20] unifies the object detection and phrase grounding tasks, allowing the combined use of both detection and grounding datasets for aligning phrase semantics and regional images. Det-CLIP [50] extracts the general concepts from multiple data sources and then resorts to a description-enriched concept dictionary for enhancement. Recently, Grounding DINO [27] adopts a similar detection-grounding training strategy as GLIP but proposes a tighter text-vision fusion solution based on transformer-based detector DINO [58]. + +Zero-shot Learning in 3D Data. Exploration of open-vocabulary 3D object detection has been limited due to the challenges associated with acquiring large-scale point cloud-text pairs and the inherent complexity of 3D scenarios. In the confined indoor scenario, OV-3DET [29] addresses the open-vocabulary detection problem by first localizing 3D objects and then assigning labels for the objects with open-vocabulary 2D models. However, temporal and spatial constraints are overlooked. CoDA [5] proposes to use some base classes to assist in detecting novel classes. As illustrated in Figure 1, such a paradigm in LiDAR data easily overfits the base classes. Recently, UP-VL [32] also considers the open-vocabulary case in autonomous driving, however, the categories involved are rather limited (only 3 classes). Some other works [2,6,7,47] tackle a related but distinct task of differentiating unknown data in testing from known data in training, known as the out-of-distribution (OOD) problem. Similarly, MODEST [54], OYSTER [59], and LiSe [61] study the unsupervised object discovery problem. Meanwhile, some studies [26,30,33] attempt to transfer existing 2D large-scale pre-trained models for open-vocabulary scene segmentation, while others [8-11,60] leverage the open-set ability of pre-trained models for the 3D classification tasks. + +Different from the above works, in this paper, we investigate open-vocabulary ability in LiDAR-based 3D object detection in the outdoor scenario, representing the real-world scenario and a much more complex scene. + +# 3 OpenSight + +To elaborate on the LiDAR-based open-vocabulary object detection, we first briefly introduce the necessary notations for ease of understanding. Following this, we delve into the details of generic object perception in Section 3.2. Subsequently, we present the process of specific semantic interpretation in Section 3.3. The overall framework of OpenSight is presented in Figure 2. + +# 3.1 Notation and Preliminaries + +The LiDAR dataset is denoted as $\mathcal{D} = \{P_i, \{y_k\}_{k=1}^K\}_{i=1}^{|\mathcal{D}|}$ , where $P_i$ corresponds to the LiDAR point clouds in $i$ -th frame, basically composed of $(x_i, y_i, z_i)$ coordinates. $y_k$ represents the combination of each 3D bounding box $b_k$ and its + +semantic label $c_k$ . In open-vocabulary scenarios, $P_i$ is visible while $y_k$ is not. Generally, in the context of autonomous driving, each LiDAR frame is accompanied by $m$ images captured from varying perspectives. We define the image dataset as $\mathcal{I} = \{(I_{ij})_{j=1}^m\}_{i=1}^{|\mathcal{D}|}$ here. Image $I_{ij}$ in $j$ -th view and LiDAR points $P_i$ could be connected by the intrinsic matrix $\mathbf{K}_j$ and extrinsic matrix $\mathbf{E}_j$ . For clarity and brevity in the rest of the paper, we utilize a single LiDAR frame as an example and abbreviate $I_{ij}$ as $I_j$ , respectively. $N$ denotes the number of 3D objects, with $f^{3D}$ and $f^{2D}$ denotes the extracted feature in 3D and 2D. + +# 3.2 Generic Objects Perception + +Discerning and localizing generic objects are fundamental in LiDAR-based open-vocabulary detection when ground-truth boxes and labels are absent. The model is expected to perceive the existence of a generic object before assigning specific labels. This section details how we leverage 2D and 3D geometry priors to enable this perception, including initial box estimation, and temporal-spatial awareness. Initial Box Estimation. In practical applications, each LiDAR frame is typically paired with $m$ images $\{I_j\}_{j=1}^m$ , captured from various perspectives. Each image is processed through the pre-trained detection model $\mathcal{I}$ , yielding a set of 2D bounding boxes $\{b_{jk}\}_{k=1}^{V_j}$ , where $V_j$ represents the number of 2D objects identified in the $j$ -th viewpoint. By integrating the data from all viewpoints, we compile the collective set of detected boxes $\{(b_{jk}^{2D})_{k=1}^{V_j}\}_{j=1}^m$ . + +With the obtained 2D boxes of generic objects, we try to lift them back in the LiDAR space. Specifically, we project the LiDAR point $P_{i}$ into each image view by: + +$$ +\hat {P} _ {i j} = \mathbf {K} _ {j} \cdot \mathbf {E} _ {j} \cdot P _ {i}, j = 1, \dots , m \tag {1} +$$ + +where $\mathbf{E}_j$ and $\mathbf{K}_j$ are extrinsic and intrinsic matrix, respectively. The points whose projection falls within each of the predicted 2D bounding boxes $b_{jk}^{2D}$ essentially constitute the 3D frustum, a conical region perceivable from a given perspective. We then apply the region growing algorithm [1] to isolate the cluster with the densest LiDAR points. Such an algorithm first randomly selects one point in the frustum and then iteratively finds its adjacent points. + +To estimate 3D bounding boxes from LiDAR point clusters, we first determine the extremal points along each axis by calculating the minimum and maximum coordinates. Take the X-axis for example, this is achieved by performing $x_{\mathrm{min}} = \min (\{x_i\}_{i = 0}^{|N_c|})$ and $x_{\mathrm{max}} = \max (\{x_i\}_{i = 0}^{|N_c|})$ , where $|N_{c}|$ is the number of points in the obtained cluster. These extremal points, identified similarly for the Y and Z axis, define the vertices of the bounding box. From these vertices, we construct the maximal bounding box $b_{i}^{3D}$ , enclosing all points in the cluster. For the heading of the bounding box, we traverse different angles. In each angle, we can obtain the XYZ boundary. We then shrink the boundary by a scale of 0.95, some points will be out of the obtained bounding boxes. The angle with the most outside points is regarded as the angle of the object. It is based on the assumption that for the "tightest" bounding box, a slight change in the angle will incur a big change in the encompassed points. The 3D boxes for generic + +![](images/110661097f1a82a690868c814dfe8dd6a228b8dc92ba39f79f60cd1ff952c13a.jpg) +Fig. 3: Temporal awareness: Project boxes from frame $t + 1$ to $t$ . By comparing with projections, we refine missed or inaccurately estimated boxes in frame $t$ . + +objects at timestamp $t$ are then finalized as $\mathcal{B}^t = \{b_i^{3D}\}_{i=0}^{N_t}$ , by aggregating all boxes from $m$ perspectives. + +Temporal Awareness. For the LiDAR point cloud, point density decreases sharply as distance increases, leading to less reliable estimations for distant objects. However, in autonomous driving scenarios, objects that are distant in the current frame may move closer in subsequent frames. To exploit this, we leverage the temporal continuity between frames by correlating 3D bounding box predictions across successive timestamps. We take frames at timestamp $t$ and $t + 1$ as an example. Specifically, given the predicted boxes $\mathcal{B}^t = \{b_i\}_{i=0}^{N_t}$ and $\mathcal{B}^{t+1} = \{b_i\}_{i=0}^{N_{t+1}}$ , where $N_{t+1}$ and $N_t$ represent the number of boxes in $t + 1$ and $t$ , respectively, we project each 3D box $b_i$ in $\mathcal{B}^{t+1}$ back to frame $t$ using the transformation: + +$$ +\hat {b} _ {i} = \mathbf {E} _ {\mathrm {l i d a r} \leftarrow \mathrm {e g o} _ {t}} \cdot \mathbf {E} _ {\mathrm {e g o} _ {t} \leftarrow \mathrm {g l b}} \cdot \mathbf {E} _ {\mathrm {g l b} \leftarrow \mathrm {e g o} _ {t + 1}} \cdot \mathbf {E} _ {\mathrm {e g o} _ {t + 1} \leftarrow \mathrm {l i d a r}} \cdot b _ {i}, +$$ + +where $\mathbf{E}_{\mathrm{lidar}} \leftarrow \mathrm{ego}_t$ and $\mathbf{E}_{\mathrm{ego}_t} \leftarrow \mathrm{glb}$ are the extrinsic matrices for the ego to LiDAR and global to ego transformations at timestamp $t$ , and $\mathbf{E}_{\mathrm{glb}} \leftarrow \mathrm{ego}_{t+1}$ , $\mathbf{E}_{\mathrm{ego}_{t+1}} \leftarrow \mathrm{lidar}$ denotes ego to global and LiDAR to ego at timestamp $t+1$ , respectively. + +With all boxes in $\mathcal{B}^{t + 1}$ considered, we obtain their projections as $\hat{\mathcal{B}}^{t + 1}$ and construct the overlap matrix $M\in \mathbb{R}^{N_{t + 1}\times N_t}$ by calculating their Intersection over Union (IoU) with boxes in $\mathcal{B}^t$ . We address two key scenarios: If a box $\hat{b}_i$ has no overlap with boxes in $\mathcal{B}^t$ , its counterpart $b_{i}$ in timestamp $t + 1$ is considered a missed detection in the current frame and is added to the prediction at timestamp $t$ . Conversely, if box $\hat{b}_i$ overlaps with boxes in $\mathcal{B}^t$ and its distance exceeds a certain threshold, we perform Non-Maximum Suppression (NMS) and retain the union of the two boxes to address potentially incomplete estimations at greater distances (See Figure 3). + +This process updates the predicted boxes of the current frame to $\mathcal{B}_{\mathrm{update}}^t = \mathcal{B}^t \bigcup \mathcal{B}_{\mathrm{temporal}}^t$ , which combines the original and temporally adjusted boxes. In practice, for the current frame $t$ , we consider both its subsequent frame at timespan $t + 1$ and its precedent frame at timestamp $t - 1$ . That means two adjacent frames are considered. Through such a constraint, we are likely to boost the continuity and consistency of generic objects over time. + +Spatial Awareness. Considering the sparse object distribution in 3D LiDAR scans, where vast areas of the point cloud are often devoid of meaningful objects, + +![](images/5156ce522d22bed4fcf0cfc06f61d8699612134425bc3578851756d1313c215b.jpg) +Fig. 4: Spatial awareness: Size priors are first generated from GPT-4 to select "precisely" estimated boxes and build the object bank. Randomly select objects and points from the objects bank and place them in the current frame $t$ . + +we also focus on the systematic integration of "precisely" estimated objects in each frame to enhance the detectability and representation of generic objects. To obtain these precise boxes from $\mathcal{B}^t$ in the absence of any ground truth, we utilize large language models (LLMs), such as GPT-4, for extracting useful size priors of objects. GPT-4, when prompted appropriately, can provide valuable insights into the typical sizes of objects within a specific class, e.g., the width, length, and height of a class. For instance, prompts like "What is the average size of a \{\#CLASS\}" in the autonomous driving scene?" enable us to collect size statistics for common classes (Details in the Appendix). Moreover, leveraging the ability of GPT-4 to process images, we also input images of the relevant objects to validate these size priors (See Figure 4). + +Utilizing the extracted size priors, we selectively retain 3D boxes in $\mathcal{B}^t$ that closely align with these priors. A box is reserved if its dimensions $w$ , $l$ , and $h$ fall within the specified range: + +$$ +w \in [ (1 - \tau) w _ {p r i o r}, (1 + \tau) w _ {p r i o r} ], +$$ + +$$ +\ell \in [ (1 - \tau) \ell_ {p r i o r}, (1 + \tau) \ell_ {p r i o r} ], \tag {2} +$$ + +$$ +h \in [ (1 - \tau) h _ {p r i o r}, (1 + \tau) h _ {p r i o r} ], +$$ + +where $\tau$ represents the error margin. $w_{prior}$ , $l_{prior}$ and $h_{prior}$ are from LLMs-derived priors. This process results in an object bank $O = \{(b_i^{3D}, P_i)\}_{i=0}^{N_o}$ , where $b_i^{3D}$ is the reserved box, $P_i$ comprises the LiDAR points within each box and $N_o$ is the total number of reserved boxes. + +For the current frame, we randomly select boxes from bank $O$ and integrate them at varying distances from the ego vehicle, ensuring diverse spatial distribution. The LiDAR points within each box are sampled according to their distance, with the sampling ratio defined as: + +$$ +r _ {s a m p l i n g} = \min (1, \frac {d - d _ {n e w}}{d - d _ {o r i}}), \tag {3} +$$ + +where $d$ denotes the total range in the point cloud, $d_{ori}$ is the original distance of the object, $d_{new}$ is the distance post-placement. The selected boxes are denoted as $\mathcal{B}_{\mathrm{spatial}}^t$ and the collective set of boxes in the current frame is represented as $\mathcal{B}_{\mathrm{final}}^t = \mathcal{B}^t \cup \mathcal{B}_{\mathrm{temporal}}^t \cup \mathcal{B}_{\mathrm{spatial}}^t$ . This strategy aims to substantially enhance the visibility of generic objects in LiDAR data and the perception ability of the obtained model. + +Loss for Generic Object Perception. With the obtained 3D boxes $\mathcal{B}_{\mathrm{final}}^t = \{b_k^{3D}\}_{k=1}^{V_{\mathrm{final}}}$ , we formulate the loss for generic object concept acquisition and localization. Utilizing a DETR-like structure, we extract features from LiDAR points and apply bipartite matching to determine the optimal permutation $\hat{\sigma}$ , the loss is thus expressed as: + +$$ +\mathcal {L} = \sum^ {N} \left[ \mathcal {L} _ {\text {f o c a l}} \left(s _ {\hat {\sigma} _ {i} (k)}\right) + \mathcal {L} _ {\mathrm {L} _ {1}} \left(b _ {\hat {\sigma} _ {i} (k)}, b _ {k} ^ {3 D}\right) \right], \tag {4} +$$ + +where $s_{\hat{\sigma}_i(k)}$ denotes whether it is an object or not, $\hat{\sigma}_i(k)$ is the element in $\hat{\sigma}$ . $\mathcal{L}_{\mathrm{focal}}$ constrains the level of "objectness" of generated 3D boxes, $\mathcal{L}_{\mathrm{box}}$ considers the constraint on boxes. With this composite loss optimized, we obtain a model with the ability to perceive generic objects. + +# 3.3 Specific Semantics Interpretation + +To interpret the specific semantics of detected generic objects in LiDAR, we correlate objects across both LiDAR and image domains. We apply the model after generic object perception to a given LiDAR frame, producing boxes $\{b_i^{3D}\}_{i=0}^N$ and their corresponding features $\{f_i^{3D}\}_{i=0}^N$ . Instead of using the image-level LVLs, e.g., CLIP, we opt for object-level LVLs, such as Grounding DINO [27], which provides 2D boxes $\{b_{i=0}^{2D}\}_{i=0}^{N_{2D}}$ , 2D object features $\{f_i^{2D}\}_{i=0}^{N_{2D}}$ and text embeddings $\{t_i\}_{i=0}^{N_{2D}}$ simultaneously. + +For each detected 3D box $b_{i}^{3D}$ , we project it to each view image with the corresponding intrinsic matrix $K_{j}$ and extrinsic matrix $E_{j}$ , formulated as: + +$$ +\hat {b} _ {i j} ^ {2 D} = \mathbf {K} _ {j} \cdot \mathbf {E} _ {j} \cdot b _ {i} ^ {3 D}, j = 1, \dots , m \tag {5} +$$ + +The $\hat{b}_{ij}^{2D}$ is further filtered and only those lying in the image are reserved. For the reserved box $\hat{b}_{ij}^{2D}$ , we compute the IoU with the 2D boxes $\{b_{i}^{2D}\}_{i=0}^{N_{2D}}$ from Grounding DINO and find the most close one $b_{j}^{2D}$ . The corresponding text embeddings $t_{j}^{2D}$ feature is retrieved. We then try to minimize the discrepancy between $f_{i}^{3D}$ and all 2D features that have the same text embeddings $t_{j}^{2D}$ . Mathematically, it is expressed as: + +$$ +\ell_ {\text {a l i g n}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\sum_ {j = 0} ^ {N _ {i}} e ^ {f _ {i} ^ {3 D} \cdot f _ {j} ^ {2 D} / \delta}}{\sum_ {k = 0} ^ {N ^ {2 D}} e ^ {f _ {i} ^ {3 D} \cdot f _ {k} ^ {2 D} / \delta}}, \tag {6} +$$ + +where $N_{i}$ denotes all 2D features that have the same text embeddings $t_j^{2D}$ for $b_i^{3D}$ . Once the object-level alignment is finished, we can directly perform open-vocabulary LiDAR detection with the LiDAR input. + +2D-3D Integration. Considering that paired RGB images with LiDAR are always available in autonomous driving, we thus can perform 2D-3D feature integration. For each detected object, we have optimized 3D features $f_{i}^{3D}$ and identified the best matched 2D boxes feature $f_{i}^{2D}$ , the similarity between 3D-2D features with the text embeddings in $\{t_i\}_{i = 0}^{N_{2D}}$ are computed as follows: + +$$ +\begin{array}{l} s _ {j} ^ {3 D} = \cos \left(f _ {i} ^ {3 D}, t _ {j}\right), \\ 2 ^ {D} = \left(3 ^ {D} + 1\right) \dots 0 \quad \text {v} ^ {3 D} \end{array} \tag {7} +$$ + +$$ +s _ {j} ^ {2 D} = \cos \left(f _ {i} ^ {2 D}, t _ {j}\right), j = 0, \dots , N ^ {2 D} +$$ + +We obtain the maximum value for 2D and 3D feature by $s^{3D} = \max_j\{s_j^{3D}\}_{j = 0}^{N_{2D}}$ , $s^{2D} = \max_j\{s_j^{2D}\}_{j = 0}^{N_{2D}}$ , along with their semantic labels $c^{3D}$ and $c^{2D}$ . The final label for object $i$ is determined by comparing $s_i^{2D}$ and $s_i^{3D}$ , selecting the label associated with the larger value. + +Large Language Model (LLMs) Post-processing. The size priors from LLMs can be used to assist in recalibrating predictions. By matching the 3D IoU of predicted boxes with these size priors, we can determine the most matching prior and the corresponding semantic label. We then incorporate the predictions from size prior with results from 2D-3D integration by selecting label with the highest similarity. + +# 3.4 Inference + +we first input the LiDAR point cloud to the generic object detector and extract 3D object features. We then align these features with text embeddings from Grounding DINO [27] to determine semantic labels. Additionally, when RGB images are available, we further improve the alignment ability by integrating 3D and 2D features or use the LLMs prior (size priors from GPT-4) to calibrate the predicted labels, thereby improving the accuracy of obtained semantic labels. + +# 4 Experiments + +# 4.1 Experimental settings + +Dataset. We conduct the open-vocabulary experiments on nuScenes [4], a large-scale, outdoor dataset featuring multimodal data modalities. Each frame in the dataset contains data from one 32-beam LiDAR and six cameras, offering a comprehensive view of the surroundings. The dataset comprises 1,000 scenes in total, with up to 1.4 million annotated 3D bounding boxes spanning 10 classes. We reserve all the classes as novel to evaluate the effectiveness of our method. + +Evaluation Metrics. For quantitative evaluation, we initially employ Average Precision (AP) on each class and mean Average Precision (mAP) across all novel classes. Considering that LiDAR point density varies with the distance, we introduce distance-based metrics: mAP-near, mAP-midrange, and mAP-far. Similarly, acknowledging the size disparities among objects in different classes (e.g., 'bus' vs. 'traffic_cone'), we categorize classes into 'large', 'medium', and 'small' and propose mAP-large, mAP-medium, and mAP-small. More details are in the Appendix. + +Table 1: Results on NuScenes dataset. All classes are reserved as novel classes. AP of each class and mAP across all classes are shown. "Car" in CoDA is the known class and is fully supervised ('Cone': 'Traffic_cone', 'C.V': 'Construction_ vehicle', 'Motor.'] 'Motorcycle', 'Ped.'] 'Pedestrian'). + +
MethodmAPNDSCarTruckC.V.BusTrailerBarrierBicycleConeMotor.Ped.
Supervised (box+label)57.361.786.038.818.363.334.270.546.770.259.585.2
Supervised (box only)34.239.053.220.17.420.50.750.137.852.639.160.8
Detic-3D [62]5.812.715.12.92.14.20.00.16.20.86.121.1
OWL-ViT-3D [31]5.412.414.22.51.84.00.00.15.21.15.519.4
OV-PointCLIP [60]0.15.00.00.00.00.00.00.00.70.20.00.1
OV-PointCLIPv2 [65]2.810.90.02.70.00.00.022.00.00.20.03.2
CLIP-3D [37]7.214.29.62.82.12.60.00.111.514.49.818.9
OV-3DET [29]5.712.016.12.90.31.30.00.31.811.02.021.1
SeCondPoint [26]8.215.015.12.32.02.00.00.010.914.915.219.3
3DGenZ [30]7.514.914.63.52.33.20.00.010.514.28.717.9
Openscene [33]8.815.417.23.92.23.00.10.212.814.010.823.4
UP-VL [32]11.317.219.47.94.33.10.40.514.115.616.830.8
CoDA [5]10.316.185.51.90.02.00.00.03.90.05.93.5
OpenSight (Ours)22.923.325.311.68.75.10.842.226.125.630.452.5
+ +Table 2: Performance w.r.t mAP-near, mAP-medium, mAP-far. + +
MethodmAP-nearmAP-midrangemAP-far
Detric-3D [62]18.59.35.8
OWL-ViT-3D [31]17.38.25.4
OV-PointCLIP [60]0.20.10.1
OV-PointCLIPv2 [65]4.53.62.8
CLIP-3D [37]18.811.27.2
OV-3DET [29]16.19.95.7
CoDA [5]18.216.610.3
OpenSight (Ours)28.026.522.9
+ +# 4.2 Baselines + +As there are no direct baselines addressing open-vocabulary LiDAR-based 3D detection, we primarily construct the baselines based on well-known works [5, 29, 37, 62, 65] and compare with OpenSight. Specifically, they are: + +Fully-supervised training. Initially, we perform fully supervised training across all categories, establishing an upper bound for the open-vocabulary detection performance. We consider using both ground-truth boxes and labels ('Supervised (box+label)') or just using ground-truth boxes ('Supervised (box only)'). + +2D open-vocabulary methods. Detic [62] is a pioneer work in the realm of 2D open-vocabulary object detection tasks and can detect 21K classes. It is thus desirable for our study here. OWL-ViT [31] decodes the object labels and boxes directly from the pre-trained features and shows good robustness in real-world applications. With the 2D detection results of either Detic or OWL-ViT, we adopt the same procedure in Section 3.2 to obtain the corresponding 3D bounding boxes in LiDAR space. These two baselines are denoted as Detic-3D and OWL-ViT-3D. + +3D open-vocabulary methods. We then consider the latest open-vocabulary point cloud classification methods and a variety of indoor open-vocabulary works to evaluate their performance in our tasks here. Specifically, we utilize a fully-supervised method to generate pseudo boxes and employ PointCLIP [60], Point-CLIPv2 [65], denoted as "OV-PointCLIP", "OV-PointCLIPv2". We also combine + +Table 3: Performance w.r.t mAIP-large, mAIP-medium, mAIP-small. + +
MethodmAP-largemAP-mediummAP-small
Detric-3D [62]5.34.111.0
OWL-ViT-3D [31]4.93.710.3
OV-PointCLIP [60]0.00.30.2
OV-PointCLIPv2 [65]0.67.31.6
CLIP-3D [37]3.47.116.7
OV-3DET [29]4.11.416.0
CoDA [5]17.94.91.8
OpenSight (Ours)11.532.939.0
+ +Table 4: Results with temporal and spatial awareness w.r.t mAP-{near, medium, far}. + +
#TemporalSpatialmAP-nearmAP-midrangemAP-far
(1)27.124.822.1
(2)✓27.425.922.5
(3)✓27.526.022.7
(4)✓✓28.026.522.9
+ +our predicted boxes and CLIP [37] for this task ("CLIP-3D"). Meanwhile, we adapt the latest works OV-3DET [29] and CoDA [5] in indoor settings and UP-VL [32] to our task here. We also consider the segmentation-based methods, e.g., SeCondPoint [26], 3DGenZ [30], and OpenScene [33], for comparison. + +# 4.3 Main results + +In this section, we discuss the performance of our model on novel classes, as summarized in Table 1. Despite the lack of 3D ground truth supervision, our framework demonstrates strong overall open-vocabulary ability across classes. Notably, it achieves exceptional performance in some classes, for example, 52.5 AP in 'pedestrian' and 42.2 AP in 'Barrier'. Unfortunately, we also notice the limited performance in 'Trailer', with 0.8 AP. To analyze this, we compare with 'Supervised (box only)', where groundtruth boxes are used and only 0.7 AP is achieved in 'Trailer', we thus attribute the reason to label alignment with Grounding DINO rather than box inaccuracy. Grounding DINO offers limited help in fine-grained classes. Our model also shows superior performance when compared with the baselines. Specifically, we observe that PointCLIP is very limited in detecting novel objects and CLIP-3D achieves the best overall performance with our detected boxes. CoDA, while effective in known classes like 'car', falls short in novel class adaptation, averaging only $1.9\%$ accuracy across novel categories. + +Our more detailed evaluations, using distance-based metrics and size-based metrics, are presented in Tables 2 and 3, respectively. The evaluation in Table 2 reveals a trend where detection accuracy is higher for objects closer to the sensor. This is attributed to their denser point clouds and more defined geometric features. Conversely, detection accuracy decreases with increasing distance, likely due to sparser data and less distinct object outlines. Regarding size-based performance (Table 3), our model shows superior results for medium and small-sized objects. Large objects present lower detection rates, mainly due to their + +![](images/db08133b340793c61862036eec9e8942c1739cedc4166b26370346b0c3f0a07a.jpg) + +![](images/b9faafb06d53ae2e42f602c6572d3f370a9374a8eef47e11f6dcdedeaee7a22a.jpg) + +![](images/60f92f0039ec3803b4558a3cdf63f74369507c6db68f7f47729e348794a32ad0.jpg) +Fig. 5: Visualization of detected novel classes in different LiDAR frames. (a) Detected 'barrier', 'traffic_cone', and 'pedestrian'; (b) Detected 'car' and 'pedestrian'; (c) Detected 'car' and 'motorcycle'; (d) Detected 'barrier' and 'pedestrian'. + +![](images/6b2e1086d7a5632be171624e50694606b81d9882449c8aafc55ebefed85a223c.jpg) + +Table 5: Results with direct alignment ('D.A.'), 3D-2D feature integration ('I.G.'), and LLMs priors. + +
#D.AI.G.PriorsmAP-largemAP-mediummAP-small
(1)✓9.431.838.1
(2)✓✓10.932.738.8
(3)✓✓✓11.532.939.0
+ +incomplete LiDAR scans. Additionally, the fine-grained categorization of large vehicles, such as 'Constructionvehicle', complicates their accurate detection. + +# 4.4 Qualitative results on novel classes + +As shown in Figure 5, our framework demonstrates a strong ability to detect novel objects by inputting the interested categories. We visualize the results from both BEV and a specific perspective. For example, our method successfully identifies the 'pedestrian', 'barrier', and 'traffic_cone' in example (a), and the 'car' and 'motorcycle' in (c). These results also indicate that our method exhibits superior open-vocabulary detection ability across different scenes. + +# 4.5 Ablation + +In this part, we present a comprehensive analysis of OpenSight and examine the effectiveness of each component. + +Components in Generic Object Perception. We decouple the components in generic object perception and record the results in Table 4. The overall improvement may not be so obvious, however, we improve more for mAP-medium and mAP-far with these techniques. For example, temporal improves 1.1 AP for mAP-medium while 0.3 AP for mAP-near. Both modules interact positively with each other and further improve the results when jointly introduced. + +Components in Semantic Interpretation. We analyze the modules of direct alignment, 3D-2D feature integration, and LLMs priors in semantic interpretation and record the results in Table 5. The performance is consistently improved + +Table 6: Results of image-level and object-level models. + +
mAP-near-midrange-far-large-medium-small
CLIP19.011.98.23.98.218.4
Grounding DINO28.026.522.911.532.939.0
+ +Table 7: AP and AR of some classes under different threshold. + +
#CLASSAP-0.5AP-1AP-2AR-0.5AR-1AR-2
Car4.718.345.940.069.084.4
Truck0.23.013.313.230.051.1
Pedestrian38.851.657.158.469.374.2
+ +with integration and priors introduced. Compared with LLMs priors, 3D-2D integration brings more benefits, for example, 0.9 vs. 0.2 in mAP-medium. + +Image-level vs. Object-level. We also investigate the image-level CLIP [37] model in semantic interpretation. The results are recorded in Table 6. Compared with our adopted object-level Grounding DINO, we notice that object-level alignment shows much better results from either a distance perspective or a size perspective. For example, 28.0 AP of near objects is achieved in object-level alignment, 9 points higher than image-level alignment. We analyze that CLIP requires more accurate bounding box prediction, in comparison with object-level alignment which only requires a nearest box. + +Importance of Bounding Boxes Existence. Existing open-vocabulary methods focus more on the alignment between 3D features and 2D features. However, in LiDAR-based object detection, we emphasize that accurate bounding boxes are essential before alignment. We select some classes and investigate both average precision (AP) and average recall (AR) under different thresholds. The results are recorded in Table 7. Compared with AP, we observe a higher recall value in different threshold, indicating that although we can recall the correct objects, more objects with incorrect sizes are also recalled. Such results motivate that if we can obtain a more precise estimation of objects, the results will be improved further. + +# 5 Conclusion + +We present OpenSight, a simple and versatile open-vocabulary framework for LiDAR-Based object detection. In the absence of ground-truth boxes and semantic labels, OpenSight first learns to identify and localize a wide range of generic objects. To boost the ability of generic object perception, we incorporate temporal and spatial awareness techniques that recalibrate and improve the visibility of estimated 3D boxes. We also align LiDAR object features with the 2D features from the Grounding DINO for semantic interpretation. The extensive experiments demonstrate the effectiveness and generalization ability of the proposed framework, paving the way for further exploration in LiDAR-based open-vocabulary challenges. + +# Acknowledgement + +This research is funded in part by ARC-Discovery grant (DP220100800 to XY) and ARC-DECRA grant (DE230100477 to XY). We thank all anonymous reviewers and ACs for their constructive suggestions. + +# References + +1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994) +2. Alliegro, A., Cappio Borlino, F., Tommasi, T.: 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems 35, 21228-21240 (2022) +3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022) +4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) +5. Cao, Y., Zeng, Y., Xu, H., Xu, D.: Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. arXiv preprint arXiv:2310.02960 (2023) +6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Open-set 3d object detection. In: 2021 International Conference on 3D Vision (3DV). pp. 869-878. IEEE (2021) +7. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. pp. 318-334. Springer (2022) +8. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. arXiv preprint arXiv:1907.06371 (2019) +9. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 923-933 (2020) +0. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision 130(10), 2364-2384 (2022) +1. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019) +2. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +3. Douillard, B., Underwood, J., Kuntz, N., Vlaskine, V., Quadros, A., Morton, P., Frenkel, A.: On the segmentation of 3d lidar point clouds. In: 2011 IEEE International Conference on Robotics and Automation. pp. 2798-2805. IEEE (2011) +4. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021) + +15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012) +16. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=1L3lnMbR4WU +17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020) +18. Huang, T., Liu, Z., Chen, X., Bai, X.: Epnet: Enhancing point features with image semantics for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 35-52. Springer (2020) +19. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019) +20. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022) +21. Li, X., Ma, T., Hou, Y., Shi, B., Yang, Y., Liu, Y., Wu, X., Chen, Q., Li, Y., Qiao, Y., et al.: Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. arXiv preprint arXiv:2303.03595 (2023) +22. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022) +23. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021) +24. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018) +25. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790 (2022) +26. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2021) +27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542 (2022) +29. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1190-1199 (2023) +30. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3d point clouds. In: 2021 International Conference on 3D Vision (3DV). pp. 992-1002. IEEE (2021) + +31. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230 (2022) +32. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8602-8612 (2023) +33. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. arXiv preprint arXiv:2211.15654 (2022) +34. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 918-927 (2018) +35. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) +36. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) +39. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020) +40. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 770-779 (2019) +41. Simon, M., Amende, K., Kraus, A., Honer, J., Samann, T., Kaulbersch, H., Milz, S., Michael Gross, H.: Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +42. Sindagi, V.A., Zhou, Y., Tuzel, O.: Mvx-net: Multimodal voxelnet for 3d object detection. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7276-7282. IEEE (2019) +43. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) +44. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021) +45. Wang, C., Ma, C., Zhu, M., Yang, X.: Pointaugmenting: Cross-modal augmentation for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11794-11803 (2021) + +46. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 18-34. Springer (2020) +47. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020) +48. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +49. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020) +50. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407 (2022) +51. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021) +52. Yin, T., Zhou, X., Krahenbuhl, P.: Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems 34, 16494-16507 (2021) +53. Yoo, J.H., Kim, Y., Kim, J., Choi, J.W.: 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 720-736. Springer (2020) +54. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1130-1140 (2022) +55. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access 8, 58443-58469 (2020) +56. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 106-122. Springer (2022) +57. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021) +58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) +59. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9317-9328 (2023) +60. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022) +61. Zhang, R., Zhang, H., Yu, H., Zheng, Z.: Approaching outside: Scaling unsupervised 3d object detection from 2d scene. In: Proceedings of the European conference on computer vision (ECCV) (2024) + +62. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 350-368. Springer (2022) +63. Zhou, Y., Sun, P., Zhang, Y., Anguelov, D., Gao, J., Ouyang, T., Guo, J., Ngiam, J., Vasudevan, V.: End-to-end multi-view fusion for 3d object detection in lidar point clouds. In: Conference on Robot Learning. pp. 923-932. PMLR (2020) +64. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018) +65. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022) \ No newline at end of file diff --git a/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/images.zip b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..981efc478821ac47e01f55307290e6a8a378e385 --- /dev/null +++ b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:083b08789369db78807690817d65c446d923c4deee25481e97ce41b48f7204bf +size 474267 diff --git a/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/layout.json b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..14f5db44689abf2528db05b8b9dde8ef4e1bee13 --- /dev/null +++ b/2024/OpenSight_ A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection/layout.json @@ -0,0 +1,11054 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "text", + "content": "OpenSight: A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "spans": [ + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "text", + "content": "Hu Zhang" + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "text", + "content": ", Jianhua Xu" + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "text", + "content": ", Tao Tang" + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "text", + "content": ", Haiyang Sun" + }, + { + "bbox": [ + 176, + 170, + 436, + 182 + ], + "type": "inline_equation", + "content": "^{6}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "spans": [ + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "type": "text", + "content": "Xin " + }, + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{2*}" + }, + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "type": "text", + "content": " , Zi Huang" + }, + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "type": "text", + "content": ", and Kaicheng " + }, + { + "bbox": [ + 203, + 182, + 411, + 195 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{3}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 247, + 203, + 367, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 203, + 367, + 214 + ], + "spans": [ + { + "bbox": [ + 247, + 203, + 367, + 214 + ], + "type": "text", + "content": "1 CSIRO DATA61, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 215, + 393, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 215, + 393, + 226 + ], + "spans": [ + { + "bbox": [ + 220, + 215, + 393, + 226 + ], + "type": "text", + "content": "2 The University of Queensland, Australia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 226, + 352, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 226, + 352, + 236 + ], + "spans": [ + { + "bbox": [ + 261, + 226, + 352, + 236 + ], + "type": "text", + "content": "3 Westlake University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 234, + 237, + 380, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 237, + 380, + 247 + ], + "spans": [ + { + "bbox": [ + 234, + 237, + 380, + 247 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 234, + 237, + 380, + 247 + ], + "type": "text", + "content": " DAMO Academy, Alibaba, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 248, + 402, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 248, + 402, + 258 + ], + "spans": [ + { + "bbox": [ + 211, + 248, + 402, + 258 + ], + "type": "text", + "content": "5 Shenzhen Campus of Sun Yat-sen University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 278, + 258, + 335, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 258, + 335, + 269 + ], + "spans": [ + { + "bbox": [ + 278, + 258, + 335, + 269 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 278, + 258, + 335, + 269 + ], + "type": "text", + "content": " LiAuto Inc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 175, + 270, + 438, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 270, + 438, + 281 + ], + "spans": [ + { + "bbox": [ + 175, + 270, + 438, + 281 + ], + "type": "text", + "content": "Hu1.Zhang@csiro.au, xin.yu@uq.edu.au, kyu@westlake.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 282, + 397, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 282, + 397, + 293 + ], + "spans": [ + { + "bbox": [ + 216, + 282, + 397, + 293 + ], + "type": "text", + "content": "https://github.com/huzhangcs/OpenSight" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 159, + 340, + 455, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 340, + 455, + 616 + ], + "spans": [ + { + "bbox": [ + 159, + 340, + 455, + 616 + ], + "type": "text", + "content": "Abstract. Traditional LiDAR-based object detection research primarily focuses on closed-set scenarios, which falls short in complex real-world applications. Directly transferring existing 2D open-vocabulary models with some known LiDAR classes for open-vocabulary ability, however, tends to suffer from over-fitting problems: The obtained model will detect the known objects, even presented with a novel category. In this paper, we propose OpenSight, a more advanced 2D-3D modeling framework for LiDAR-based open-vocabulary detection. OpenSight utilizes 2D-3D geometric priors for the initial discernment and localization of generic objects, followed by a more specific semantic interpretation of the detected objects. The process begins by generating 2D boxes for generic objects from the accompanying camera images of LiDAR. These 2D boxes, together with LiDAR points, are then lifted back into the LiDAR space to estimate corresponding 3D boxes. For better generic object perception, our framework integrates both temporal and spatial-aware constraints. Temporal awareness correlates the predicted 3D boxes across consecutive timestamps, recalibrating the missed or inaccurate boxes. The spatial awareness randomly places some \"precisely\" estimated 3D boxes at varying distances, increasing the visibility of generic objects. To interpret the specific semantics of detected objects, we develop a cross-modal alignment and fusion module to first align 3D features with 2D image embeddings and then fuse the aligned 3D-2D features for semantic decoding. Our experiments indicate that our method establishes state-of-the-art open-vocabulary performance on widely used 3D detection benchmarks and effectively identifies objects for new categories of interest." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 160, + 627, + 426, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 627, + 426, + 639 + ], + "spans": [ + { + "bbox": [ + 160, + 627, + 426, + 639 + ], + "type": "text", + "content": "Keywords: OpenSight " + }, + { + "bbox": [ + 160, + 627, + 426, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 627, + 426, + 639 + ], + "type": "text", + "content": " Open-vocabulary " + }, + { + "bbox": [ + 160, + 627, + 426, + 639 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 627, + 426, + 639 + ], + "type": "text", + "content": " 3D object detection" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 652, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 118, + 247, + 204 + ], + "blocks": [ + { + "bbox": [ + 171, + 118, + 247, + 204 + ], + "lines": [ + { + "bbox": [ + 171, + 118, + 247, + 204 + ], + "spans": [ + { + "bbox": [ + 171, + 118, + 247, + 204 + ], + "type": "image", + "image_path": "d096d890463e46c7b2d841608429504d3e4e80ed044a6032ffcb977886f8b585.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 205, + 260, + 214 + ], + "lines": [ + { + "bbox": [ + 247, + 205, + 260, + 214 + ], + "spans": [ + { + "bbox": [ + 247, + 205, + 260, + 214 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 118, + 329, + 204 + ], + "blocks": [ + { + "bbox": [ + 251, + 118, + 329, + 204 + ], + "lines": [ + { + "bbox": [ + 251, + 118, + 329, + 204 + ], + "spans": [ + { + "bbox": [ + 251, + 118, + 329, + 204 + ], + "type": "image", + "image_path": "4ce7beea5336b8b0e18db1fe60cbbfdb0fcf1831bb0769225ed3c9a71509d8e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 340, + 121, + 435, + 207 + ], + "blocks": [ + { + "bbox": [ + 340, + 121, + 435, + 207 + ], + "lines": [ + { + "bbox": [ + 340, + 121, + 435, + 207 + ], + "spans": [ + { + "bbox": [ + 340, + 121, + 435, + 207 + ], + "type": "image", + "image_path": "f9373995613eb9f30f897fbd57d7ac2b65d0cd9ff8401e853e810cb1ffdecf47.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 205, + 396, + 214 + ], + "lines": [ + { + "bbox": [ + 383, + 205, + 396, + 214 + ], + "spans": [ + { + "bbox": [ + 383, + 205, + 396, + 214 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 223, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 278 + ], + "type": "text", + "content": "Fig. 1: (a) A model trained on known class \"car\" correctly detects the target object (yellow) with query \"car\". However, with a new class query \"motorcycle\", the model incorrectly detects the object (red). (b) The average precision of known \"car\" and novel \"motorcycle\". The degraded performance in \"motorcycle\" (blue) indicates the limited generalization from the known classes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 230, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 230, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 230, + 298 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 422 + ], + "type": "text", + "content": "The growing prevalence of autonomous driving has significantly increased research interest in various 3D perception tasks [13,17,41,55], among which LiDAR-based object detection stands as a crucial component. Despite the advancements in this field, most existing research focuses on closed-set settings [19,25,36,48,64], where the obtained models are limited to predicting a set of predefined categories. This constraint makes these models insufficient for handling complex real-world scenarios, where diverse and previously unseen object categories may be encountered. Consequently, there is a growing demand to develop models capable of detecting novel categories of interest." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 424, + 483, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 483, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 483, + 519 + ], + "type": "text", + "content": "Considering the limited availability of annotated categories in existing LiDAR-based datasets [4,15,43], adapting established large-scale visual-language models (LVLs), e.g., CLIP [37] for open-vocabulary LiDAR detection appears promising. One paradigm is to co-embed known class object features with the image/text embeddings from LVLs, expecting alignment generalization to the novel classes. While effective in 2D image-based open-vocabulary detection, this paradigm exhibits significant overfitting in LiDAR applications, as models persist in detecting objects in known classes even when faced with a new category." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "We consider some known classes, e.g., \"car\", and follow the paradigm above to illustrate this phenomenon. When given a novel category \"motorcycle\", the obtained model erroneously detects the known object \"car\", as depicted in Figure 1 (a). An analysis of average precision (AP) further reveals the diminished efficacy in novel classes, as shown in Figure 1 (b). These results underscore the limitation of such a paradigm, emphasizing the necessity for a tailored approach for LiDAR-based open-vocabulary problems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we present \"OpenSight\", a more advanced 2D-3D modeling approach for LiDAR-based open-vocabulary detection. OpenSight primarily relies on 2D and 3D geometry priors to first acquire the concept and localization of generic objects. Then, it interprets the semantics of detected objects by coembedding LiDAR features with 2D LVLs. This framework addresses all poten" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "tially novel objects explicitly and circumvents the reliance on predefined known classes, thereby avoiding the overfitting problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 368 + ], + "type": "text", + "content": "Specifically, we consider both temporal-aware and spatial-aware constraints to acquire the concept of generic objects and their localization. Initially, leveraging the 2D detection model and the accompanying RGB images of LiDAR, we extract 2D geometric priors for generic objects, that is, 2D bounding boxes across successive frames. These boxes are then lifted back into 3D space with LiDAR points, resulting in a collection of 3D frustums. Point clustering is then performed in each frustum to estimate the 3D bounding boxes of objects. Temporally, we correlate predicted 3D bounding boxes across frames and project the predicted 3D boxes from previous or subsequent frames onto the current frame. The missed or inaccurate objects in the current frame are then identified through overlapping comparison. This constraint boosts the continuity and consistency of generic objects over time. Spatially, we first employ large language models (LLMs) to derive size priors for general objects and then maintain 3D boxes that closely align with these size priors. The maintained boxes are randomly selected and placed at varying distances within the current frame, with the LiDAR points inside these boxes being sampled by their distance. This spatial awareness notably increases the visibility of objects, particularly those located further from the sensor. By integrating the temporal and spatial constraints, the ability to recognize and localize generic objects is significantly augmented." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 368, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 510 + ], + "type": "text", + "content": "To decode the specific semantics of detected boxes in LiDAR space, we correlate objects across both LiDAR and image domains. Initially, by back-projecting 3D LiDAR bounding boxes onto the image plane, we identify the most closely matching 2D boxes from the Grounding DINO [27]. We then aim to minimize the feature discrepancy between the 3D LiDAR objects and the 2D matching box features. Subsequently, we fuse the optimized 3D-2D features and align the unified features with textual category embeddings from Grounding DINO to decode semantics. This dual 2D-3D feature approach leverages salient patterns in both 2D images and 3D geometry, significantly enhancing the robustness of semantic decoding. Additionally, we integrate size priors from LLMs before, recalibrating predictions that diverge from the average box size, thus enhancing the accuracy of our predictive model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 512, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 481, + 559 + ], + "type": "text", + "content": "In the experiments, we first establish a variety of baselines for comparison. Then, we test our method on the widely-used 3D detection benchmarks, demonstrating that our approach achieves state-of-the-art open-vocabulary performance and effectively identifies objects for new categories of interest." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 578, + 243, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 243, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 243, + 592 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Closed-set 3D Object Detection. Extensive research has been proposed on LiDAR-based 3D object detection in autonomous driving. One group of works performs on LiDAR-only data [39,63], either by extracting features from raw point clouds to predict 3D bounding boxes [23,34-36,40,49] or converting the irregular LiDAR point clouds into a regular feature space [14,19,44,46,48,51,64]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 392, + 231 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 392, + 231 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 392, + 231 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 392, + 231 + ], + "type": "image", + "image_path": "1c80c38be6ce7555c4a05f7995a94400e9932325d5737e08c25aee22ffc763d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 340 + ], + "type": "text", + "content": "Fig. 2: An overview of our OpenSight framework. In generic object perception, 2D bounding boxes are initially lifted back into 3D space (initial boxes estimation). This is followed by integrated temporal and spatial awareness in the estimated 3D boxes to retrieve the missed objects or increase their visibility. Subsequently, we utilize these enhanced estimates to compute the generic object loss and optimize the 3D detection model. In specific semantic interpretation, we project the generated 3D boxes of the optimized 3D model into the image plane to get projected 2D boxes. These projections are then matched with the nearest 2D boxes identified via Grounding DINO, facilitating the alignment of 3D LiDAR object features with corresponding 2D image features." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 481, + 231 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 481, + 231 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 481, + 231 + ], + "type": "image", + "image_path": "2743e696be55661dbd6a8a2368c33f7ab29b301dff040f0dc649f03a3440fd70.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 510 + ], + "type": "text", + "content": "Another group focuses on both LiDAR and camera modalities. Some works emphasize point-wise fusion [18,42,45,52], projecting raw LiDAR points to retrieve corresponding image features and concatenating them together. In contrast, others concentrate on feature-level fusion [22,24,53], extracting features or proposals based on LiDAR points and querying the associated image features. This line of methods has been proven to be more efficient. For example, TransFusion [3] utilizes proposals generated by LiDAR features to query the image feature for further refinement. BEVFusion [25, 28] projects the image features into BEV space by using lift-splat-shoot (LSS) and concatenates them with the LiDAR features. More recently, LoGoNet [21] constructs Local-to-Global fusion network to perform LiDAR-camera fusion at both local and global levels. However, all these approaches target the closed-set setting, restricting the range of detected object categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Open-vocabulary 2D Object Detection. The advent of large-scale pretrained models, which exhibit good text-vision alignment and zero-shot transferability, has spurred researchers to investigate the open-vocabulary problem in object detection tasks. A variety of approaches have been proposed with 2D RGB images: OVR-CNN [57] initially pre-trains a Faster R-CNN detector [38] by using image-caption pairs and text extractor, e.g., BERT [12], then fine-tunes it on downstream detection datasets. ViLD [16] distills the knowledge from a CLIP-based [37] teacher to a student detector for generalized text and image embedding alignment. Detic [62] harnesses extensive image classification data to train the classifier within a detector while localizing objects. Similar to OVR-CNN, OWL [31] first pre-trains a transformer encoder by contrastively aligning image-text pairs, then transfers it to detection tasks by attaching lightweight object classification and localization heads. Following the DETR-based" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "structure, OV-DETR [56] employs CLIP-encoded text and image embeddings as queries for obtaining category-specific boxes. In contrast to the aforementioned methods, some alternatives have emerged to directly learn open-vocabulary capabilities on detection data. For example, GLIP [20] unifies the object detection and phrase grounding tasks, allowing the combined use of both detection and grounding datasets for aligning phrase semantics and regional images. Det-CLIP [50] extracts the general concepts from multiple data sources and then resorts to a description-enriched concept dictionary for enhancement. Recently, Grounding DINO [27] adopts a similar detection-grounding training strategy as GLIP but proposes a tighter text-vision fusion solution based on transformer-based detector DINO [58]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 462 + ], + "type": "text", + "content": "Zero-shot Learning in 3D Data. Exploration of open-vocabulary 3D object detection has been limited due to the challenges associated with acquiring large-scale point cloud-text pairs and the inherent complexity of 3D scenarios. In the confined indoor scenario, OV-3DET [29] addresses the open-vocabulary detection problem by first localizing 3D objects and then assigning labels for the objects with open-vocabulary 2D models. However, temporal and spatial constraints are overlooked. CoDA [5] proposes to use some base classes to assist in detecting novel classes. As illustrated in Figure 1, such a paradigm in LiDAR data easily overfits the base classes. Recently, UP-VL [32] also considers the open-vocabulary case in autonomous driving, however, the categories involved are rather limited (only 3 classes). Some other works [2,6,7,47] tackle a related but distinct task of differentiating unknown data in testing from known data in training, known as the out-of-distribution (OOD) problem. Similarly, MODEST [54], OYSTER [59], and LiSe [61] study the unsupervised object discovery problem. Meanwhile, some studies [26,30,33] attempt to transfer existing 2D large-scale pre-trained models for open-vocabulary scene segmentation, while others [8-11,60] leverage the open-set ability of pre-trained models for the 3D classification tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 463, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 499 + ], + "type": "text", + "content": "Different from the above works, in this paper, we investigate open-vocabulary ability in LiDAR-based 3D object detection in the outdoor scenario, representing the real-world scenario and a much more complex scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 217, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 217, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 217, + 529 + ], + "type": "text", + "content": "3 OpenSight" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 537, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 597 + ], + "type": "text", + "content": "To elaborate on the LiDAR-based open-vocabulary object detection, we first briefly introduce the necessary notations for ease of understanding. Following this, we delve into the details of generic object perception in Section 3.2. Subsequently, we present the process of specific semantic interpretation in Section 3.3. The overall framework of OpenSight is presented in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 612, + 298, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 298, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 298, + 624 + ], + "type": "text", + "content": "3.1 Notation and Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": "The LiDAR dataset is denoted as " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{P_i, \\{y_k\\}_{k=1}^K\\}_{i=1}^{|\\mathcal{D}|}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " corresponds to the LiDAR point clouds in " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": "-th frame, basically composed of " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "(x_i, y_i, z_i)" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " coordinates. " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_k" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " represents the combination of each 3D bounding box " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "b_k" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " and its" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": "semantic label " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": ". In open-vocabulary scenarios, " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " is visible while " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "y_k" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " is not. Generally, in the context of autonomous driving, each LiDAR frame is accompanied by " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " images captured from varying perspectives. We define the image dataset as " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{(I_{ij})_{j=1}^m\\}_{i=1}^{|\\mathcal{D}|}" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " here. Image " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "I_{ij}" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": "-th view and LiDAR points " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " could be connected by the intrinsic matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " and extrinsic matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": ". For clarity and brevity in the rest of the paper, we utilize a single LiDAR frame as an example and abbreviate " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "I_{ij}" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " denotes the number of 3D objects, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "f^{3D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "inline_equation", + "content": "f^{2D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 214 + ], + "type": "text", + "content": " denotes the extracted feature in 3D and 2D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 230, + 301, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 301, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 301, + 243 + ], + "type": "text", + "content": "3.2 Generic Objects Perception" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": "Discerning and localizing generic objects are fundamental in LiDAR-based open-vocabulary detection when ground-truth boxes and labels are absent. The model is expected to perceive the existence of a generic object before assigning specific labels. This section details how we leverage 2D and 3D geometry priors to enable this perception, including initial box estimation, and temporal-spatial awareness. Initial Box Estimation. In practical applications, each LiDAR frame is typically paired with " + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\{I_j\\}_{j=1}^m" + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": ", captured from various perspectives. Each image is processed through the pre-trained detection model " + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": ", yielding a set of 2D bounding boxes " + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\{b_{jk}\\}_{k=1}^{V_j}" + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "inline_equation", + "content": "V_j" + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": " represents the number of 2D objects identified in the " + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": "-th viewpoint. By integrating the data from all viewpoints, we compile the collective set of detected boxes " + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\{(b_{jk}^{2D})_{k=1}^{V_j}\\}_{j=1}^m" + }, + { + "bbox": [ + 130, + 249, + 482, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "type": "text", + "content": "With the obtained 2D boxes of generic objects, we try to lift them back in the LiDAR space. Specifically, we project the LiDAR point " + }, + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 384, + 482, + 418 + ], + "type": "text", + "content": " into each image view by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 418, + 480, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 418, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 241, + 418, + 480, + 432 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {i j} = \\mathbf {K} _ {j} \\cdot \\mathbf {E} _ {j} \\cdot P _ {i}, j = 1, \\dots , m \\tag {1}", + "image_path": "001f3173384f691d1bf0a60b22be53b56b1f3dc5f39eeeb821641611dfbf10fc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_j" + }, + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j" + }, + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "text", + "content": " are extrinsic and intrinsic matrix, respectively. The points whose projection falls within each of the predicted 2D bounding boxes " + }, + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "inline_equation", + "content": "b_{jk}^{2D}" + }, + { + "bbox": [ + 130, + 436, + 482, + 509 + ], + "type": "text", + "content": " essentially constitute the 3D frustum, a conical region perceivable from a given perspective. We then apply the region growing algorithm [1] to isolate the cluster with the densest LiDAR points. Such an algorithm first randomly selects one point in the frustum and then iteratively finds its adjacent points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "To estimate 3D bounding boxes from LiDAR point clusters, we first determine the extremal points along each axis by calculating the minimum and maximum coordinates. Take the X-axis for example, this is achieved by performing " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{min}} = \\min (\\{x_i\\}_{i = 0}^{|N_c|})" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{max}} = \\max (\\{x_i\\}_{i = 0}^{|N_c|})" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "|N_{c}|" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " is the number of points in the obtained cluster. These extremal points, identified similarly for the Y and Z axis, define the vertices of the bounding box. From these vertices, we construct the maximal bounding box " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "b_{i}^{3D}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": ", enclosing all points in the cluster. For the heading of the bounding box, we traverse different angles. In each angle, we can obtain the XYZ boundary. We then shrink the boundary by a scale of 0.95, some points will be out of the obtained bounding boxes. The angle with the most outside points is regarded as the angle of the object. It is based on the assumption that for the \"tightest\" bounding box, a slight change in the angle will incur a big change in the encompassed points. The 3D boxes for generic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 175, + 117, + 436, + 206 + ], + "blocks": [ + { + "bbox": [ + 175, + 117, + 436, + 206 + ], + "lines": [ + { + "bbox": [ + 175, + 117, + 436, + 206 + ], + "spans": [ + { + "bbox": [ + 175, + 117, + 436, + 206 + ], + "type": "image", + "image_path": "110661097f1a82a690868c814dfe8dd6a228b8dc92ba39f79f60cd1ff952c13a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "type": "text", + "content": "Fig. 3: Temporal awareness: Project boxes from frame " + }, + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "type": "text", + "content": ". By comparing with projections, we refine missed or inaccurately estimated boxes in frame " + }, + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 213, + 480, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "text", + "content": "objects at timestamp " + }, + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "text", + "content": " are then finalized as " + }, + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t = \\{b_i^{3D}\\}_{i=0}^{N_t}" + }, + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "text", + "content": ", by aggregating all boxes from " + }, + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 242, + 479, + 266 + ], + "type": "text", + "content": " perspectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": "Temporal Awareness. For the LiDAR point cloud, point density decreases sharply as distance increases, leading to less reliable estimations for distant objects. However, in autonomous driving scenarios, objects that are distant in the current frame may move closer in subsequent frames. To exploit this, we leverage the temporal continuity between frames by correlating 3D bounding box predictions across successive timestamps. We take frames at timestamp " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " as an example. Specifically, given the predicted boxes " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t = \\{b_i\\}_{i=0}^{N_t}" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{t+1} = \\{b_i\\}_{i=0}^{N_{t+1}}" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "N_{t+1}" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " represent the number of boxes in " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": ", respectively, we project each 3D box " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{t+1}" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " back to frame " + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 267, + 480, + 388 + ], + "type": "text", + "content": " using the transformation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 177, + 398, + 435, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 398, + 435, + 414 + ], + "spans": [ + { + "bbox": [ + 177, + 398, + 435, + 414 + ], + "type": "interline_equation", + "content": "\\hat {b} _ {i} = \\mathbf {E} _ {\\mathrm {l i d a r} \\leftarrow \\mathrm {e g o} _ {t}} \\cdot \\mathbf {E} _ {\\mathrm {e g o} _ {t} \\leftarrow \\mathrm {g l b}} \\cdot \\mathbf {E} _ {\\mathrm {g l b} \\leftarrow \\mathrm {e g o} _ {t + 1}} \\cdot \\mathbf {E} _ {\\mathrm {e g o} _ {t + 1} \\leftarrow \\mathrm {l i d a r}} \\cdot b _ {i},", + "image_path": "a7aa2c17894e78df1ee5b92d45df0a96e8a22b83e7ac25af497c9d251284663a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}} \\leftarrow \\mathrm{ego}_t" + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{ego}_t} \\leftarrow \\mathrm{glb}" + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "content": " are the extrinsic matrices for the ego to LiDAR and global to ego transformations at timestamp " + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{glb}} \\leftarrow \\mathrm{ego}_{t+1}" + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{ego}_{t+1}} \\leftarrow \\mathrm{lidar}" + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "content": " denotes ego to global and LiDAR to ego at timestamp " + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "inline_equation", + "content": "t+1" + }, + { + "bbox": [ + 130, + 422, + 479, + 458 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": "With all boxes in " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{t + 1}" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " considered, we obtain their projections as " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}^{t + 1}" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " and construct the overlap matrix " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "M\\in \\mathbb{R}^{N_{t + 1}\\times N_t}" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " by calculating their Intersection over Union (IoU) with boxes in " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": ". We address two key scenarios: If a box " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " has no overlap with boxes in " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": ", its counterpart " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " in timestamp " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " is considered a missed detection in the current frame and is added to the prediction at timestamp " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": ". Conversely, if box " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " overlaps with boxes in " + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t" + }, + { + "bbox": [ + 130, + 459, + 480, + 567 + ], + "type": "text", + "content": " and its distance exceeds a certain threshold, we perform Non-Maximum Suppression (NMS) and retain the union of the two boxes to address potentially incomplete estimations at greater distances (See Figure 3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "content": "This process updates the predicted boxes of the current frame to " + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{update}}^t = \\mathcal{B}^t \\bigcup \\mathcal{B}_{\\mathrm{temporal}}^t" + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "content": ", which combines the original and temporally adjusted boxes. In practice, for the current frame " + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "content": ", we consider both its subsequent frame at timespan " + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "content": " and its precedent frame at timestamp " + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "content": ". That means two adjacent frames are considered. Through such a constraint, we are likely to boost the continuity and consistency of generic objects over time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Spatial Awareness. Considering the sparse object distribution in 3D LiDAR scans, where vast areas of the point cloud are often devoid of meaningful objects," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 102 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 114, + 434, + 269 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 434, + 269 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 434, + 269 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 434, + 269 + ], + "type": "image", + "image_path": "5156ce522d22bed4fcf0cfc06f61d8699612134425bc3578851756d1313c215b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "text", + "content": "Fig. 4: Spatial awareness: Size priors are first generated from GPT-4 to select \"precisely\" estimated boxes and build the object bank. Randomly select objects and points from the objects bank and place them in the current frame " + }, + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "type": "text", + "content": "we also focus on the systematic integration of \"precisely\" estimated objects in each frame to enhance the detectability and representation of generic objects. To obtain these precise boxes from " + }, + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t" + }, + { + "bbox": [ + 130, + 318, + 482, + 449 + ], + "type": "text", + "content": " in the absence of any ground truth, we utilize large language models (LLMs), such as GPT-4, for extracting useful size priors of objects. GPT-4, when prompted appropriately, can provide valuable insights into the typical sizes of objects within a specific class, e.g., the width, length, and height of a class. For instance, prompts like \"What is the average size of a \\{\\#CLASS\\}\" in the autonomous driving scene?\" enable us to collect size statistics for common classes (Details in the Appendix). Moreover, leveraging the ability of GPT-4 to process images, we also input images of the relevant objects to validate these size priors (See Figure 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "content": "Utilizing the extracted size priors, we selectively retain 3D boxes in " + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t" + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "content": " that closely align with these priors. A box is reserved if its dimensions " + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 449, + 483, + 486 + ], + "type": "text", + "content": " fall within the specified range:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 492, + 381, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 492, + 381, + 506 + ], + "spans": [ + { + "bbox": [ + 231, + 492, + 381, + 506 + ], + "type": "interline_equation", + "content": "w \\in [ (1 - \\tau) w _ {p r i o r}, (1 + \\tau) w _ {p r i o r} ],", + "image_path": "a465917175f8895971467f24706cb27a4ea98a3d0629d072a7da6cf546ac1807.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 506, + 480, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 506, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 235, + 506, + 480, + 520 + ], + "type": "interline_equation", + "content": "\\ell \\in [ (1 - \\tau) \\ell_ {p r i o r}, (1 + \\tau) \\ell_ {p r i o r} ], \\tag {2}", + "image_path": "843dcc70abd7a3337b782a926a58cfaec825f072f7c6fadf538ee9512583b923.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 234, + 522, + 378, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 522, + 378, + 536 + ], + "spans": [ + { + "bbox": [ + 234, + 522, + 378, + 536 + ], + "type": "interline_equation", + "content": "h \\in [ (1 - \\tau) h _ {p r i o r}, (1 + \\tau) h _ {p r i o r} ],", + "image_path": "29b4cf448708c774d70fe2c05627fd9ef48d552ea9117bca1cffacafef1f5c51.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " represents the error margin. " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "w_{prior}" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "l_{prior}" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "h_{prior}" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " are from LLMs-derived priors. This process results in an object bank " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "O = \\{(b_i^{3D}, P_i)\\}_{i=0}^{N_o}" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "b_i^{3D}" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " is the reserved box, " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " comprises the LiDAR points within each box and " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "N_o" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " is the total number of reserved boxes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "type": "text", + "content": "For the current frame, we randomly select boxes from bank " + }, + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "type": "text", + "content": " and integrate them at varying distances from the ego vehicle, ensuring diverse spatial distribution. The LiDAR points within each box are sampled according to their distance, with the sampling ratio defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 239, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 239, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "r _ {s a m p l i n g} = \\min (1, \\frac {d - d _ {n e w}}{d - d _ {o r i}}), \\tag {3}", + "image_path": "0874e39569c1a152e05c537d002ecee698e9c980df408561652e75dbde07f3a6.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " denotes the total range in the point cloud, " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "d_{ori}" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " is the original distance of the object, " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "d_{new}" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " is the distance post-placement. The selected boxes are denoted as " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{spatial}}^t" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " and the collective set of boxes in the current frame is represented as " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{final}}^t = \\mathcal{B}^t \\cup \\mathcal{B}_{\\mathrm{temporal}}^t \\cup \\mathcal{B}_{\\mathrm{spatial}}^t" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": ". This strategy aims to substantially enhance the visibility of generic objects in LiDAR data and the perception ability of the obtained model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 189, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 189, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 189, + 479, + 251 + ], + "type": "text", + "content": "Loss for Generic Object Perception. With the obtained 3D boxes " + }, + { + "bbox": [ + 131, + 189, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{final}}^t = \\{b_k^{3D}\\}_{k=1}^{V_{\\mathrm{final}}}" + }, + { + "bbox": [ + 131, + 189, + 479, + 251 + ], + "type": "text", + "content": ", we formulate the loss for generic object concept acquisition and localization. Utilizing a DETR-like structure, we extract features from LiDAR points and apply bipartite matching to determine the optimal permutation " + }, + { + "bbox": [ + 131, + 189, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 131, + 189, + 479, + 251 + ], + "type": "text", + "content": ", the loss is thus expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 218, + 258, + 480, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 258, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 218, + 258, + 480, + 283 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum^ {N} \\left[ \\mathcal {L} _ {\\text {f o c a l}} \\left(s _ {\\hat {\\sigma} _ {i} (k)}\\right) + \\mathcal {L} _ {\\mathrm {L} _ {1}} \\left(b _ {\\hat {\\sigma} _ {i} (k)}, b _ {k} ^ {3 D}\\right) \\right], \\tag {4}", + "image_path": "3e2a11b40ebb8295c7f8a190c343d16907613f1e2cbe11faea388728d24612e8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "inline_equation", + "content": "s_{\\hat{\\sigma}_i(k)}" + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "text", + "content": " denotes whether it is an object or not, " + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}_i(k)" + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "text", + "content": " is the element in " + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{focal}}" + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "text", + "content": " constrains the level of \"objectness\" of generated 3D boxes, " + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 131, + 285, + 479, + 334 + ], + "type": "text", + "content": " considers the constraint on boxes. With this composite loss optimized, we obtain a model with the ability to perceive generic objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 348, + 328, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 328, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 328, + 361 + ], + "type": "text", + "content": "3.3 Specific Semantics Interpretation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "text", + "content": "To interpret the specific semantics of detected generic objects in LiDAR, we correlate objects across both LiDAR and image domains. We apply the model after generic object perception to a given LiDAR frame, producing boxes " + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "inline_equation", + "content": "\\{b_i^{3D}\\}_{i=0}^N" + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "text", + "content": " and their corresponding features " + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "inline_equation", + "content": "\\{f_i^{3D}\\}_{i=0}^N" + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "text", + "content": ". Instead of using the image-level LVLs, e.g., CLIP, we opt for object-level LVLs, such as Grounding DINO [27], which provides 2D boxes " + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "inline_equation", + "content": "\\{b_{i=0}^{2D}\\}_{i=0}^{N_{2D}}" + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "text", + "content": ", 2D object features " + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "inline_equation", + "content": "\\{f_i^{2D}\\}_{i=0}^{N_{2D}}" + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "text", + "content": " and text embeddings " + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "inline_equation", + "content": "\\{t_i\\}_{i=0}^{N_{2D}}" + }, + { + "bbox": [ + 131, + 365, + 479, + 450 + ], + "type": "text", + "content": " simultaneously." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "text", + "content": "For each detected 3D box " + }, + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "inline_equation", + "content": "b_{i}^{3D}" + }, + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "text", + "content": ", we project it to each view image with the corresponding intrinsic matrix " + }, + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "inline_equation", + "content": "K_{j}" + }, + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "text", + "content": " and extrinsic matrix " + }, + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "inline_equation", + "content": "E_{j}" + }, + { + "bbox": [ + 130, + 451, + 479, + 475 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 236, + 482, + 480, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 482, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 236, + 482, + 480, + 497 + ], + "type": "interline_equation", + "content": "\\hat {b} _ {i j} ^ {2 D} = \\mathbf {K} _ {j} \\cdot \\mathbf {E} _ {j} \\cdot b _ {i} ^ {3 D}, j = 1, \\dots , m \\tag {5}", + "image_path": "fc93fecd24f2b0bed7b5fee6ff6d1b38f99ded1a7ac88bcbe5d71bb15003087a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\hat{b}_{ij}^{2D}" + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": " is further filtered and only those lying in the image are reserved. For the reserved box " + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\hat{b}_{ij}^{2D}" + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": ", we compute the IoU with the 2D boxes " + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\{b_{i}^{2D}\\}_{i=0}^{N_{2D}}" + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": " from Grounding DINO and find the most close one " + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "inline_equation", + "content": "b_{j}^{2D}" + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": ". The corresponding text embeddings " + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "inline_equation", + "content": "t_{j}^{2D}" + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": " feature is retrieved. We then try to minimize the discrepancy between " + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "inline_equation", + "content": "f_{i}^{3D}" + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": " and all 2D features that have the same text embeddings " + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "inline_equation", + "content": "t_{j}^{2D}" + }, + { + "bbox": [ + 131, + 503, + 479, + 582 + ], + "type": "text", + "content": ". Mathematically, it is expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 223, + 588, + 479, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 588, + 479, + 622 + ], + "spans": [ + { + "bbox": [ + 223, + 588, + 479, + 622 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {a l i g n}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\sum_ {j = 0} ^ {N _ {i}} e ^ {f _ {i} ^ {3 D} \\cdot f _ {j} ^ {2 D} / \\delta}}{\\sum_ {k = 0} ^ {N ^ {2 D}} e ^ {f _ {i} ^ {3 D} \\cdot f _ {k} ^ {2 D} / \\delta}}, \\tag {6}", + "image_path": "ad33724a39c5393b5e2a5a7447452f6bb204eb78c05900574c5a4382e346cf59.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "text", + "content": " denotes all 2D features that have the same text embeddings " + }, + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "t_j^{2D}" + }, + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "inline_equation", + "content": "b_i^{3D}" + }, + { + "bbox": [ + 131, + 628, + 479, + 666 + ], + "type": "text", + "content": ". Once the object-level alignment is finished, we can directly perform open-vocabulary LiDAR detection with the LiDAR input." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 103 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "2D-3D Integration. Considering that paired RGB images with LiDAR are always available in autonomous driving, we thus can perform 2D-3D feature integration. For each detected object, we have optimized 3D features " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "f_{i}^{3D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and identified the best matched 2D boxes feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "f_{i}^{2D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", the similarity between 3D-2D features with the text embeddings in " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\{t_i\\}_{i = 0}^{N_{2D}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " are computed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 181, + 480, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 181, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 233, + 181, + 480, + 202 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s _ {j} ^ {3 D} = \\cos \\left(f _ {i} ^ {3 D}, t _ {j}\\right), \\\\ 2 ^ {D} = \\left(3 ^ {D} + 1\\right) \\dots 0 \\quad \\text {v} ^ {3 D} \\end{array} \\tag {7}", + "image_path": "0f47dccbcf6d222733a00f6e112c1c39b61579d72fbcda65d4baafc8513a2277.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 198, + 381, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 198, + 381, + 212 + ], + "spans": [ + { + "bbox": [ + 233, + 198, + 381, + 212 + ], + "type": "interline_equation", + "content": "s _ {j} ^ {2 D} = \\cos \\left(f _ {i} ^ {2 D}, t _ {j}\\right), j = 0, \\dots , N ^ {2 D}", + "image_path": "d04a0d16b4043ac08419dde7132e9f47591aacdfe2568790eede040773322932.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": "We obtain the maximum value for 2D and 3D feature by " + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "inline_equation", + "content": "s^{3D} = \\max_j\\{s_j^{3D}\\}_{j = 0}^{N_{2D}}" + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "inline_equation", + "content": "s^{2D} = \\max_j\\{s_j^{2D}\\}_{j = 0}^{N_{2D}}" + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": ", along with their semantic labels " + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "inline_equation", + "content": "c^{3D}" + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "inline_equation", + "content": "c^{2D}" + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": ". The final label for object " + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": " is determined by comparing " + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "inline_equation", + "content": "s_i^{2D}" + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "inline_equation", + "content": "s_i^{3D}" + }, + { + "bbox": [ + 130, + 217, + 482, + 270 + ], + "type": "text", + "content": ", selecting the label associated with the larger value." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 342 + ], + "type": "text", + "content": "Large Language Model (LLMs) Post-processing. The size priors from LLMs can be used to assist in recalibrating predictions. By matching the 3D IoU of predicted boxes with these size priors, we can determine the most matching prior and the corresponding semantic label. We then incorporate the predictions from size prior with results from 2D-3D integration by selecting label with the highest similarity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 357, + 208, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 208, + 367 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 208, + 367 + ], + "type": "text", + "content": "3.4 Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "content": "we first input the LiDAR point cloud to the generic object detector and extract 3D object features. We then align these features with text embeddings from Grounding DINO [27] to determine semantic labels. Additionally, when RGB images are available, we further improve the alignment ability by integrating 3D and 2D features or use the LLMs prior (size priors from GPT-4) to calibrate the predicted labels, thereby improving the accuracy of obtained semantic labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 460, + 230, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 230, + 473 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 230, + 473 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 481, + 271, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 271, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 271, + 494 + ], + "type": "text", + "content": "4.1 Experimental settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "content": "Dataset. We conduct the open-vocabulary experiments on nuScenes [4], a large-scale, outdoor dataset featuring multimodal data modalities. Each frame in the dataset contains data from one 32-beam LiDAR and six cameras, offering a comprehensive view of the surroundings. The dataset comprises 1,000 scenes in total, with up to 1.4 million annotated 3D bounding boxes spanning 10 classes. We reserve all the classes as novel to evaluate the effectiveness of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Metrics. For quantitative evaluation, we initially employ Average Precision (AP) on each class and mean Average Precision (mAP) across all novel classes. Considering that LiDAR point density varies with the distance, we introduce distance-based metrics: mAP-near, mAP-midrange, and mAP-far. Similarly, acknowledging the size disparities among objects in different classes (e.g., 'bus' vs. 'traffic_cone'), we categorize classes into 'large', 'medium', and 'small' and propose mAP-large, mAP-medium, and mAP-small. More details are in the Appendix." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 162, + 479, + 281 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Results on NuScenes dataset. All classes are reserved as novel classes. AP of each class and mAP across all classes are shown. \"Car\" in CoDA is the known class and is fully supervised ('Cone': 'Traffic_cone', 'C.V': 'Construction_ vehicle', 'Motor.'] 'Motorcycle', 'Ped.'] 'Pedestrian')." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 162, + 479, + 281 + ], + "lines": [ + { + "bbox": [ + 133, + 162, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 479, + 281 + ], + "type": "table", + "html": "
MethodmAPNDSCarTruckC.V.BusTrailerBarrierBicycleConeMotor.Ped.
Supervised (box+label)57.361.786.038.818.363.334.270.546.770.259.585.2
Supervised (box only)34.239.053.220.17.420.50.750.137.852.639.160.8
Detic-3D [62]5.812.715.12.92.14.20.00.16.20.86.121.1
OWL-ViT-3D [31]5.412.414.22.51.84.00.00.15.21.15.519.4
OV-PointCLIP [60]0.15.00.00.00.00.00.00.00.70.20.00.1
OV-PointCLIPv2 [65]2.810.90.02.70.00.00.022.00.00.20.03.2
CLIP-3D [37]7.214.29.62.82.12.60.00.111.514.49.818.9
OV-3DET [29]5.712.016.12.90.31.30.00.31.811.02.021.1
SeCondPoint [26]8.215.015.12.32.02.00.00.010.914.915.219.3
3DGenZ [30]7.514.914.63.52.33.20.00.010.514.28.717.9
Openscene [33]8.815.417.23.92.23.00.10.212.814.010.823.4
UP-VL [32]11.317.219.47.94.33.10.40.514.115.616.830.8
CoDA [5]10.316.185.51.90.02.00.00.03.90.05.93.5
OpenSight (Ours)22.923.325.311.68.75.10.842.226.125.630.452.5
", + "image_path": "f3f3e75358e299ca73b2d78e5bfab19ab0c7536f2c772711393f1e24025e3404.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 200, + 300, + 408, + 385 + ], + "blocks": [ + { + "bbox": [ + 173, + 286, + 440, + 298 + ], + "lines": [ + { + "bbox": [ + 173, + 286, + 440, + 298 + ], + "spans": [ + { + "bbox": [ + 173, + 286, + 440, + 298 + ], + "type": "text", + "content": "Table 2: Performance w.r.t mAP-near, mAP-medium, mAP-far." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 300, + 408, + 385 + ], + "lines": [ + { + "bbox": [ + 200, + 300, + 408, + 385 + ], + "spans": [ + { + "bbox": [ + 200, + 300, + 408, + 385 + ], + "type": "table", + "html": "
MethodmAP-nearmAP-midrangemAP-far
Detric-3D [62]18.59.35.8
OWL-ViT-3D [31]17.38.25.4
OV-PointCLIP [60]0.20.10.1
OV-PointCLIPv2 [65]4.53.62.8
CLIP-3D [37]18.811.27.2
OV-3DET [29]16.19.95.7
CoDA [5]18.216.610.3
OpenSight (Ours)28.026.522.9
", + "image_path": "cacdd256278e0e7a0b466974592fc2073c2fe642e7489b9ed5d5005ef7e58b7d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 400, + 208, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 208, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 208, + 412 + ], + "type": "text", + "content": "4.2 Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 424, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 460 + ], + "type": "text", + "content": "As there are no direct baselines addressing open-vocabulary LiDAR-based 3D detection, we primarily construct the baselines based on well-known works [5, 29, 37, 62, 65] and compare with OpenSight. Specifically, they are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 484, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 484, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 484, + 510 + ], + "type": "text", + "content": "Fully-supervised training. Initially, we perform fully supervised training across all categories, establishing an upper bound for the open-vocabulary detection performance. We consider using both ground-truth boxes and labels ('Supervised (box+label)') or just using ground-truth boxes ('Supervised (box only)')." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 604 + ], + "type": "text", + "content": "2D open-vocabulary methods. Detic [62] is a pioneer work in the realm of 2D open-vocabulary object detection tasks and can detect 21K classes. It is thus desirable for our study here. OWL-ViT [31] decodes the object labels and boxes directly from the pre-trained features and shows good robustness in real-world applications. With the 2D detection results of either Detic or OWL-ViT, we adopt the same procedure in Section 3.2 to obtain the corresponding 3D bounding boxes in LiDAR space. These two baselines are denoted as Detic-3D and OWL-ViT-3D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "3D open-vocabulary methods. We then consider the latest open-vocabulary point cloud classification methods and a variety of indoor open-vocabulary works to evaluate their performance in our tasks here. Specifically, we utilize a fully-supervised method to generate pseudo boxes and employ PointCLIP [60], Point-CLIPv2 [65], denoted as \"OV-PointCLIP\", \"OV-PointCLIPv2\". We also combine" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 141, + 90, + 447, + 103 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 198, + 128, + 411, + 213 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 445, + 125 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 445, + 125 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 445, + 125 + ], + "type": "text", + "content": "Table 3: Performance w.r.t mAIP-large, mAIP-medium, mAIP-small." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 198, + 128, + 411, + 213 + ], + "lines": [ + { + "bbox": [ + 198, + 128, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 198, + 128, + 411, + 213 + ], + "type": "table", + "html": "
MethodmAP-largemAP-mediummAP-small
Detric-3D [62]5.34.111.0
OWL-ViT-3D [31]4.93.710.3
OV-PointCLIP [60]0.00.30.2
OV-PointCLIPv2 [65]0.67.31.6
CLIP-3D [37]3.47.116.7
OV-3DET [29]4.11.416.0
CoDA [5]17.94.91.8
OpenSight (Ours)11.532.939.0
", + "image_path": "aa92635ae5553547c73a86a0415e2d2bceccb3a6fac8f2a2244f5250db30095d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 186, + 230, + 422, + 284 + ], + "blocks": [ + { + "bbox": [ + 132, + 216, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 216, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 479, + 228 + ], + "type": "text", + "content": "Table 4: Results with temporal and spatial awareness w.r.t mAP-{near, medium, far}." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 230, + 422, + 284 + ], + "lines": [ + { + "bbox": [ + 186, + 230, + 422, + 284 + ], + "spans": [ + { + "bbox": [ + 186, + 230, + 422, + 284 + ], + "type": "table", + "html": "
#TemporalSpatialmAP-nearmAP-midrangemAP-far
(1)27.124.822.1
(2)✓27.425.922.5
(3)✓27.526.022.7
(4)✓✓28.026.522.9
", + "image_path": "60d497238061f65927b3263b7876938385c9abd934078ce3445571fdbd739161.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 298, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 348 + ], + "type": "text", + "content": "our predicted boxes and CLIP [37] for this task (\"CLIP-3D\"). Meanwhile, we adapt the latest works OV-3DET [29] and CoDA [5] in indoor settings and UP-VL [32] to our task here. We also consider the segmentation-based methods, e.g., SeCondPoint [26], 3DGenZ [30], and OpenScene [33], for comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 367, + 224, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 224, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 224, + 378 + ], + "type": "text", + "content": "4.3 Main results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "type": "text", + "content": "In this section, we discuss the performance of our model on novel classes, as summarized in Table 1. Despite the lack of 3D ground truth supervision, our framework demonstrates strong overall open-vocabulary ability across classes. Notably, it achieves exceptional performance in some classes, for example, 52.5 AP in 'pedestrian' and 42.2 AP in 'Barrier'. Unfortunately, we also notice the limited performance in 'Trailer', with 0.8 AP. To analyze this, we compare with 'Supervised (box only)', where groundtruth boxes are used and only 0.7 AP is achieved in 'Trailer', we thus attribute the reason to label alignment with Grounding DINO rather than box inaccuracy. Grounding DINO offers limited help in fine-grained classes. Our model also shows superior performance when compared with the baselines. Specifically, we observe that PointCLIP is very limited in detecting novel objects and CLIP-3D achieves the best overall performance with our detected boxes. CoDA, while effective in known classes like 'car', falls short in novel class adaptation, averaging only " + }, + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 569 + ], + "type": "text", + "content": " accuracy across novel categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Our more detailed evaluations, using distance-based metrics and size-based metrics, are presented in Tables 2 and 3, respectively. The evaluation in Table 2 reveals a trend where detection accuracy is higher for objects closer to the sensor. This is attributed to their denser point clouds and more defined geometric features. Conversely, detection accuracy decreases with increasing distance, likely due to sparser data and less distinct object outlines. Regarding size-based performance (Table 3), our model shows superior results for medium and small-sized objects. Large objects present lower detection rates, mainly due to their" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 303, + 178 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 303, + 178 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 303, + 178 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 303, + 178 + ], + "type": "image", + "image_path": "db08133b340793c61862036eec9e8942c1739cedc4166b26370346b0c3f0a07a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 453, + 178 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 453, + 178 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 453, + 178 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 453, + 178 + ], + "type": "image", + "image_path": "b9faafb06d53ae2e42f602c6572d3f370a9374a8eef47e11f6dcdedeaee7a22a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 156, + 185, + 302, + 252 + ], + "blocks": [ + { + "bbox": [ + 156, + 185, + 302, + 252 + ], + "lines": [ + { + "bbox": [ + 156, + 185, + 302, + 252 + ], + "spans": [ + { + "bbox": [ + 156, + 185, + 302, + 252 + ], + "type": "image", + "image_path": "60f92f0039ec3803b4558a3cdf63f74369507c6db68f7f47729e348794a32ad0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 256, + 480, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 480, + 291 + ], + "type": "text", + "content": "Fig. 5: Visualization of detected novel classes in different LiDAR frames. (a) Detected 'barrier', 'traffic_cone', and 'pedestrian'; (b) Detected 'car' and 'pedestrian'; (c) Detected 'car' and 'motorcycle'; (d) Detected 'barrier' and 'pedestrian'." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 186, + 453, + 251 + ], + "blocks": [ + { + "bbox": [ + 308, + 186, + 453, + 251 + ], + "lines": [ + { + "bbox": [ + 308, + 186, + 453, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 453, + 251 + ], + "type": "image", + "image_path": "6b2e1086d7a5632be171624e50694606b81d9882449c8aafc55ebefed85a223c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 183, + 324, + 426, + 369 + ], + "blocks": [ + { + "bbox": [ + 131, + 300, + 479, + 322 + ], + "lines": [ + { + "bbox": [ + 131, + 300, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 479, + 322 + ], + "type": "text", + "content": "Table 5: Results with direct alignment ('D.A.'), 3D-2D feature integration ('I.G.'), and LLMs priors." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 324, + 426, + 369 + ], + "lines": [ + { + "bbox": [ + 183, + 324, + 426, + 369 + ], + "spans": [ + { + "bbox": [ + 183, + 324, + 426, + 369 + ], + "type": "table", + "html": "
#D.AI.G.PriorsmAP-largemAP-mediummAP-small
(1)✓9.431.838.1
(2)✓✓10.932.738.8
(3)✓✓✓11.532.939.0
", + "image_path": "039ece4aab9e9a44d9cc1d84af6f69d4298dd0c4f1d1531623210cb4f814b541.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 373, + 479, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 479, + 398 + ], + "type": "text", + "content": "incomplete LiDAR scans. Additionally, the fine-grained categorization of large vehicles, such as 'Constructionvehicle', complicates their accurate detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 413, + 336, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 336, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 336, + 425 + ], + "type": "text", + "content": "4.4 Qualitative results on novel classes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 430, + 480, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 501 + ], + "type": "text", + "content": "As shown in Figure 5, our framework demonstrates a strong ability to detect novel objects by inputting the interested categories. We visualize the results from both BEV and a specific perspective. For example, our method successfully identifies the 'pedestrian', 'barrier', and 'traffic_cone' in example (a), and the 'car' and 'motorcycle' in (c). These results also indicate that our method exhibits superior open-vocabulary detection ability across different scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 517, + 205, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 205, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 205, + 528 + ], + "type": "text", + "content": "4.5 Ablation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 557 + ], + "type": "text", + "content": "In this part, we present a comprehensive analysis of OpenSight and examine the effectiveness of each component." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 558, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 629 + ], + "type": "text", + "content": "Components in Generic Object Perception. We decouple the components in generic object perception and record the results in Table 4. The overall improvement may not be so obvious, however, we improve more for mAP-medium and mAP-far with these techniques. For example, temporal improves 1.1 AP for mAP-medium while 0.3 AP for mAP-near. Both modules interact positively with each other and further improve the results when jointly introduced." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Components in Semantic Interpretation. We analyze the modules of direct alignment, 3D-2D feature integration, and LLMs priors in semantic interpretation and record the results in Table 5. The performance is consistently improved" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 128, + 434, + 170 + ], + "blocks": [ + { + "bbox": [ + 190, + 114, + 421, + 125 + ], + "lines": [ + { + "bbox": [ + 190, + 114, + 421, + 125 + ], + "spans": [ + { + "bbox": [ + 190, + 114, + 421, + 125 + ], + "type": "text", + "content": "Table 6: Results of image-level and object-level models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 128, + 434, + 170 + ], + "lines": [ + { + "bbox": [ + 177, + 128, + 434, + 170 + ], + "spans": [ + { + "bbox": [ + 177, + 128, + 434, + 170 + ], + "type": "table", + "html": "
mAP-near-midrange-far-large-medium-small
CLIP19.011.98.23.98.218.4
Grounding DINO28.026.522.911.532.939.0
", + "image_path": "71a81f12604decfd6c85dcddf42055b3f118d74c1d5015a018485c89ea3c84cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 176, + 193, + 433, + 239 + ], + "blocks": [ + { + "bbox": [ + 177, + 180, + 435, + 190 + ], + "lines": [ + { + "bbox": [ + 177, + 180, + 435, + 190 + ], + "spans": [ + { + "bbox": [ + 177, + 180, + 435, + 190 + ], + "type": "text", + "content": "Table 7: AP and AR of some classes under different threshold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 193, + 433, + 239 + ], + "lines": [ + { + "bbox": [ + 176, + 193, + 433, + 239 + ], + "spans": [ + { + "bbox": [ + 176, + 193, + 433, + 239 + ], + "type": "table", + "html": "
#CLASSAP-0.5AP-1AP-2AR-0.5AR-1AR-2
Car4.718.345.940.069.084.4
Truck0.23.013.313.230.051.1
Pedestrian38.851.657.158.469.374.2
", + "image_path": "b8fbe631153f030eea54a30f8001f92678cca0c243dd5d5ca768b4ede9683606.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "content": "with integration and priors introduced. Compared with LLMs priors, 3D-2D integration brings more benefits, for example, 0.9 vs. 0.2 in mAP-medium." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 269, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 480, + 365 + ], + "type": "text", + "content": "Image-level vs. Object-level. We also investigate the image-level CLIP [37] model in semantic interpretation. The results are recorded in Table 6. Compared with our adopted object-level Grounding DINO, we notice that object-level alignment shows much better results from either a distance perspective or a size perspective. For example, 28.0 AP of near objects is achieved in object-level alignment, 9 points higher than image-level alignment. We analyze that CLIP requires more accurate bounding box prediction, in comparison with object-level alignment which only requires a nearest box." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 485 + ], + "type": "text", + "content": "Importance of Bounding Boxes Existence. Existing open-vocabulary methods focus more on the alignment between 3D features and 2D features. However, in LiDAR-based object detection, we emphasize that accurate bounding boxes are essential before alignment. We select some classes and investigate both average precision (AP) and average recall (AR) under different thresholds. The results are recorded in Table 7. Compared with AP, we observe a higher recall value in different threshold, indicating that although we can recall the correct objects, more objects with incorrect sizes are also recalled. Such results motivate that if we can obtain a more precise estimation of objects, the results will be improved further." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 503, + 220, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 220, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 220, + 516 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 529, + 482, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 649 + ], + "type": "text", + "content": "We present OpenSight, a simple and versatile open-vocabulary framework for LiDAR-Based object detection. In the absence of ground-truth boxes and semantic labels, OpenSight first learns to identify and localize a wide range of generic objects. To boost the ability of generic object perception, we incorporate temporal and spatial awareness techniques that recalibrate and improve the visibility of estimated 3D boxes. We also align LiDAR object features with the 2D features from the Grounding DINO for semantic interpretation. The extensive experiments demonstrate the effectiveness and generalization ability of the proposed framework, paving the way for further exploration in LiDAR-based open-vocabulary challenges." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "type": "text", + "content": "This research is funded in part by ARC-Discovery grant (DP220100800 to XY) and ARC-DECRA grant (DE230100477 to XY). We thank all anonymous reviewers and ACs for their constructive suggestions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 190, + 198, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 198, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 198, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "text", + "content": "1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "text", + "content": "2. Alliegro, A., Cappio Borlino, F., Tommasi, T.: 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems 35, 21228-21240 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 308 + ], + "type": "text", + "content": "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "5. Cao, Y., Zeng, Y., Xu, H., Xu, D.: Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. arXiv preprint arXiv:2310.02960 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "text", + "content": "6. Cen, J., Yun, P., Cai, J., Wang, M.Y., Liu, M.: Open-set 3d object detection. In: 2021 International Conference on 3D Vision (3DV). pp. 869-878. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "type": "text", + "content": "7. Cen, J., Yun, P., Zhang, S., Cai, J., Luan, D., Tang, M., Liu, M., Yu Wang, M.: Open-world semantic segmentation for lidar point clouds. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. pp. 318-334. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "text", + "content": "8. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Mitigating the hubness problem for zero-shot learning of 3d objects. arXiv preprint arXiv:1907.06371 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "text", + "content": "9. Cheraghian, A., Rahman, S., Campbell, D., Petersson, L.: Transductive zero-shot learning for 3d point cloud classification. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 923-933 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "type": "text", + "content": "0. Cheraghian, A., Rahman, S., Chowdhury, T.F., Campbell, D., Petersson, L.: Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision 130(10), 2364-2384 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "1. Cheraghian, A., Rahman, S., Petersson, L.: Zero-shot learning of 3d point cloud objects. In: 2019 16th International Conference on Machine Vision Applications (MVA). pp. 1-6. IEEE (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "2. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "3. Douillard, B., Underwood, J., Kuntz, N., Vlaskine, V., Quadros, A., Morton, P., Frenkel, A.: On the segmentation of 3d lidar point clouds. In: 2011 IEEE International Conference on Robotics and Automation. pp. 2798-2805. IEEE (2011)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2918-2927 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "16. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=1L3lnMbR4WU" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "type": "text", + "content": "18. Huang, T., Liu, Z., Chen, X., Bai, X.: Epnet: Enhancing point features with image semantics for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 35-52. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "text", + "content": "19. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 332 + ], + "type": "text", + "content": "20. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "21. Li, X., Ma, T., Hou, Y., Shi, B., Yang, Y., Liu, Y., Wu, X., Chen, Q., Li, Y., Qiao, Y., et al.: Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. arXiv preprint arXiv:2303.03595 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "text", + "content": "22. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "text", + "content": "23. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "type": "text", + "content": "24. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "25. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "26. Liu, B., Deng, S., Dong, Q., Hu, Z.: Language-level semantics conditioned 3d point cloud segmentation. arXiv preprint arXiv:2107.00430 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 632 + ], + "type": "text", + "content": "29. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1190-1199 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Michele, B., Boulch, A., Puy, G., Bucher, M., Marlet, R.: Generative zero-shot learning for semantic segmentation of 3d point clouds. In: 2021 International Conference on 3D Vision (3DV). pp. 992-1002. IEEE (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 204 + ], + "type": "text", + "content": "32. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8602-8612 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 204, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 481, + 236 + ], + "type": "text", + "content": "33. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. arXiv preprint arXiv:2211.15654 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 268 + ], + "type": "text", + "content": "34. Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J.: Frustum pointnets for 3d object detection from rgb-d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 918-927 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "text", + "content": "35. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "36. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 374 + ], + "type": "text", + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "text", + "content": "39. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10529–10538 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "text", + "content": "40. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 770-779 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 515 + ], + "type": "text", + "content": "41. Simon, M., Amende, K., Kraus, A., Honer, J., Samann, T., Kaulbersch, H., Milz, S., Michael Gross, H.: Complexer-yolo: Real-time 3d object detection and tracking on semantic point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "text", + "content": "42. Sindagi, V.A., Zhou, Y., Tuzel, O.: Mvx-net: Multimodal voxelnet for 3d object detection. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 7276-7282. IEEE (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 547, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 590 + ], + "type": "text", + "content": "43. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 590, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 633 + ], + "type": "text", + "content": "44. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5725-5734 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "45. Wang, C., Ma, C., Zhu, M., Yang, X.: Pointaugmenting: Cross-modal augmentation for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11794-11803 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "46. Wang, Y., Fathi, A., Kundu, A., Ross, D.A., Pantofaru, C., Funkhouser, T., Solomon, J.: Pillar-based object detection for autonomous driving. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 18-34. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 193 + ], + "type": "text", + "content": "47. Wong, K., Wang, S., Ren, M., Liang, M., Urtasun, R.: Identifying unknown instances for autonomous driving. In: Conference on Robot Learning. pp. 384-393. PMLR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 481, + 214 + ], + "type": "text", + "content": "48. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "text", + "content": "49. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "50. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "51. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784-11793 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 331 + ], + "type": "text", + "content": "52. Yin, T., Zhou, X., Krahenbuhl, P.: Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems 34, 16494-16507 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 374 + ], + "type": "text", + "content": "53. Yoo, J.H., Kim, Y., Kim, J., Choi, J.W.: 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 720-736. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "text", + "content": "54. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1130-1140 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "55. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access 8, 58443-58469 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 449, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 481, + 493 + ], + "type": "text", + "content": "56. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 106-122. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "57. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "58. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "text", + "content": "59. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9317-9328 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "60. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "61. Zhang, R., Zhang, H., Yu, H., Zheng, Z.: Approaching outside: Scaling unsupervised 3d object detection from 2d scene. In: Proceedings of the European conference on computer vision (ECCV) (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "H. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 259 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "62. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 350-368. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "text", + "content": "63. Zhou, Y., Sun, P., Zhang, Y., Anguelov, D., Gao, J., Ouyang, T., Guo, J., Ngiam, J., Vasudevan, V.: End-to-end multi-view fusion for 3d object detection in lidar point clouds. In: Conference on Robot Learning. pp. 923-932. PMLR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "64. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "text", + "content": "65. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 141, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Simple Open-Vocabulary Framework for LiDAR-Based Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_content_list.json b/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf0da24d57a3711a039925dd12a8ee4ebc1ca80d --- /dev/null +++ b/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_content_list.json @@ -0,0 +1,1751 @@ +[ + { + "type": "text", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "text_level": 1, + "bbox": [ + 279, + 140, + 725, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Steve Cruz1, Ryan Rabinowitz2, Manuel Günther3, and Terrance E. Boult2", + "bbox": [ + 336, + 212, + 666, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Notre Dame \n2 University of Colorado Colorado Springs", + "bbox": [ + 357, + 253, + 643, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Zurich", + "bbox": [ + 426, + 282, + 575, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stevecruz@nd.edu {rrabinow, tboult}@uccs.edu guenther@ifi.uzh.ch", + "bbox": [ + 254, + 297, + 746, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Open-Set Recognition (OSR) is a problem with mainly practical applications. However, recent evaluations have largely focused on small-scale data and tuning thresholds over the test set, which disregard the real-world operational needs of parameter selection. Thus, we revisit the original goals of OSR and propose a new evaluation metric, Operational Open-Set Accuracy (OOSA), which requires predicting an operationally relevant threshold from a validation set with known and a surrogate set with unknown samples, and then applying this threshold during testing. With this new measure in mind, we develop a large-scale evaluation protocol suited for operational scenarios. Additionally, we introduce the novel PostMax algorithm that performs post-processing refinement of the logit of the maximal class. This refinement involves normalizing logits by deep feature magnitudes and utilizing an extreme-value-based generalized Pareto distribution to map them into proper probabilities. We evaluate multiple pre-trained deep networks, including leading transformer and convolution-based architectures, on different selections of large-scale surrogate and test sets. Our experiments demonstrate that PostMax advances the state of the art in open-set recognition, showing statistically significant improvements in our novel OOSA metric as well as in previously used metrics such as AUROC, FPR95, and others.", + "bbox": [ + 259, + 349, + 743, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 655, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The original Open-Set Recognition (OSR) formulation [1,38,39] gained attention due to the flaws in existing recognition systems when handling inputs from outside the training classes. These systems lack control over inputs, and any unknown sample that is not rejected leads to a misclassification, thereby reducing accuracy. Therefore, early open-set work [1,3,35] introduced protocols tailored to emulate real-world systems, allowing for a proper evaluation of open-set performance.", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For a decade, researchers explored techniques that enhanced OSR performance [1,5,7,10,11,20,25,27,28,35,39,55,57]. While contributing novel ideas, recent works have diverged from the original goals of OSR. Instead of leveraging large-scale data as in early research on open-set deep networks [1], recent OSR evaluations have", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee72ea10b138a49268b1b1ed64f751f6311224ce748dfa177995a5f5af5983d8.jpg", + "image_caption": [ + "(a) Train" + ], + "image_footnote": [], + "bbox": [ + 251, + 157, + 401, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ffa717af6ff75c61ec29fa885705f87f51151c0316cbc16b292f1bd2501eedfb.jpg", + "image_caption": [ + "(b) Val", + "Fig. 1: POSTMAX PIPELINE. During train (a), post-processed maximum logits of correctly classified train samples are normalized by feature magnitude $\\ell$ ; the set is then used in a Maximum Likelihood Estimation (MLE) to obtain Generalized Pareto Distribution (GPD) [33] with parameters $\\mu, \\sigma$ , and $\\xi$ . During validation (b), we use a validation and a surrogate set where the cumulative distribution of the GPD $(\\Psi)$ gives the probability of the normalized max logit being from the train distribution and predict an operational threshold. During operation (c), we deploy each method's threshold and apply it to determine OSR and compute Operational Open Set Accuracy (OOSA). For evaluation we compare against Maximum Softmax Probability (MSP) [16,44], Maximum Logit (MaxLogit) [15,44], Nearest Neighbor Guidance (NNGuide) [32], and SCALE [50]. Operational Open-Set Accuracy (OOSA) performance is $(\\star)$ and $(\\diamond)$ is the optimal Open-Set Accuracy (OSA) for an oracle that knows the final test set (shown to highlight differences from OOSA). AUROC scores are shown for comparison." + ], + "image_footnote": [], + "bbox": [ + 408, + 157, + 560, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a6e86241cf88b5278f17d5bf7d74eb6e6f53a9f05afdaa8ee42ad43583b6bec9.jpg", + "image_caption": [ + "(c) Test (OOSA & AUROC)" + ], + "image_footnote": [], + "bbox": [ + 566, + 146, + 751, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "focused on small-scale datasets with few classes, with only a few notable exceptions [30,35,44]. Moreover, the prevalent reliance on metrics such as the Area Under the Receiver Operating Characteristics (AUROC) curve for performance evaluation has led to inopportune assessments since such metrics do not offer a realistic perspective on OSR. Fig. 1 illustrates an example where AUROC scores exhibit one ranking, yet in an operational setting, there is a slight difference. Inherently, common measures overlook the real-world operational needs of threshold selection. Some may argue for their use with a dedicated validation set; however, it is not clear how one would extract a threshold from these metrics and deploy it at test time. This is problematic as methods have largely ignored a fundamental practice of machine learning systems - having separate training, validation, and test sets to ensure that models generalize, avoid overfitting, and provide unbiased evaluation and reliable performance prediction on new data.", + "bbox": [ + 212, + 492, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Initial aspirations behind OSR aimed to enhance real systems [4]. Imagine an engineer tasked with implementing an OSR algorithm into their system. As they analyze results from various algorithms through tables and plots to inform decisions, a formidable challenge emerges. While AUROC and similar metrics assess a model's discriminative capability across all possible classification thresholds, they do not guide the engineer toward an optimal threshold for decision-making. There is a pressing need for a metric that enables confident and statistically rigorous evaluations with anticipated data, while also demonstrating robustness against varying numbers of unknowns. When deploying an open-set system, engineers must, at a minimum, consider two factors: (1) how an", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "approach scales with operational data, and (2) how to choose an operating point that balances their desired known classification accuracy and unknown rejection. These considerations underscore the importance of achieving overall accuracy with a predicted operational threshold. Therefore, we introduce a new metric called Operational Open-Set Accuracy (OOSA), illustrated in Fig. 1, which facilitates algorithm and operational threshold selection. To ensure real-world applicability, we implement a large-scale evaluation protocol with proper training/validation/test splits. We determine a threshold on the validation set and deploy it during testing on unseen unknowns. Additionally, for completeness, fairness, and robustness, we evaluate state-of-the-art algorithms on ImageNet-scale open-set recognition using common metrics (AUROC, FPR95, AUOSCR [6], and OpenAUC [46]), many of which are detailed in the supplemental.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before incorporating an operational threshold, it is natural to ask about the distribution of scores above this threshold. This naturally leads to a Peak-over-Threshold (POT) formulation of the distribution, best modeled via Extreme Value Theory (EVT). The Pickands-Balkema-De Haan theorem of EVT [33] yields a Generalized Pareto Distribution (GPD) over normalized scores, as depicted in Fig. 1. This GPD-based score transformation provides theoretically grounded probabilities rather than raw, network-dependent ad-hoc confidence scores (see supplemental for details). While GPD can be applied directly to Softmax confidences or logit values, our approach goes further. We introduce a novel algorithm, PostMax, which applies GPD to post-processed normalized maximum logits. Here, normalization by the deep feature magnitude enhances the separation of known and unknown classes. Prior observations and theoretical explanations [10,13,31,44] suggest that deep feature magnitudes extracted from inputs that the network was not trained on are generally smaller than those for known samples. However, our findings show the exact opposite, i.e., that modern networks trained on the large-scale ILSVRC2012 dataset generally exhibit larger deep feature magnitudes for unknown samples (details in supplemental). Therefore, applying GPD to logit values divided by the deep feature magnitude can enhance OSR beyond the strong baselines of using raw Softmax confidence [16], maximum logit values [15], overconfidence reduction with features and logits [32], or scaled logit values [50]. The contributions of this paper are as follows:", + "bbox": [ + 212, + 329, + 787, + 645 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel evaluation metric, Operational Open-Set Accuracy (OOSA), that emphasizes real-world usage by predicting an operational threshold on a validation set (knowns and unknowns surrogate), which is then deployed during testing on unseen data.", + "- We introduce a novel algorithm, PostMax, that post-processes maximum logits by normalizing them by deep feature magnitude and then applies GPD to provide probabilities. Our code is publicly available.4", + "- We develop new large-scale evaluation protocols suited for assessing algorithms in operational scenarios.", + "- We showcase that PostMax advances the state of the art in large-scale open-set recognition with statistical significance on OOSA and prior metrics." + ], + "bbox": [ + 225, + 647, + 784, + 811 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "4https://github.com/Vastlab/PostMax-00SA", + "bbox": [ + 236, + 824, + 550, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Improving OSR, distinct from OOD detection, anomaly detection, or novel category discovery, can be approached in two ways: providing better features through improved network training, or through post-processing where a pretrained network is trained for closed-set tasks and then adapted for OSR. Our work is rooted in the post-processing approaches.", + "bbox": [ + 212, + 175, + 787, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Related Problems", + "text_level": 1, + "bbox": [ + 215, + 272, + 410, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A common question regarding open-set is, What about a two-stage system with OOD followed by classification, does that not solve OSR? This hypothesis is rejected with an example from one of our operational systems where we have to recognize objects in novel contexts, e.g., in snowy or foggy conditions. Such conditions are OOD with respect to training samples, but they should not be rejected from an open-set point of view. Out-Of-Distribution (OOD) detection, though distinct from OSR, has been more consistent in its use of large-scale experimentation as were used in early OSR work [15,16,32,45,47,50,52,54]. Even when OOD is restricted to out-of-class, the two-stage system needs to be evaluated in that context, including a process for selecting an OOD threshold and then evaluating the resulting classification performance. The OpenOOD Benchmark [52,54] has implemented various OSR and OOD techniques such as OpenGan [20], MOS [19], ReAct [41], ViM [45], GEN [23], NNGuide [32], and SCALE [50]. While we caution the use of OOD algorithms in OSR evaluations/settings, based on the OpenOOD Benchmark leaderboard, we compare with the state-of-the-art, NNGuide & SCALE.", + "bbox": [ + 212, + 296, + 789, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Closed-Set Classifiers", + "text_level": 1, + "bbox": [ + 215, + 559, + 439, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, Vaze et al. [44] argued that closed-set classifiers are sufficient. This aligns with Hendrycks et al. [15, 16], who demonstrated thresholding on Softmax confidences or, especially, on logits provide unreasonably good baselines. Our work shows that such approaches/networks can be improved with normalization.", + "bbox": [ + 212, + 583, + 787, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate several leading pre-trained architectures trained on ILSVRC2012-1K with no additional data. Particularly, Meta's Vision Transformer Hiera-H [37], which strips non-essential components making it faster and more accurate during training/inference. Also, we use their Masked Autoencoders (MAE) model, ViT-H [14], which masks random patches of an image and reconstructs the missing pixels. To show generalization, we also utilize CNNs, including Meta's ConvNeXtV2-H [48] and ConvNeXt-L [24] which modernize a standard ResNet. Other networks are found in the supplemental material.", + "bbox": [ + 212, + 643, + 790, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Post-processing Approaches", + "text_level": 1, + "bbox": [ + 215, + 786, + 493, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides thresholding softmax scores or logits [15, 16, 44], OSR methods take features and use Weibull-calibrated Support Vector Machines (W-SVM) [38]", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "or Extreme Value Machines (EVM) [35] to estimate a probability of unknown. OpenMax [1] was the first method to add an artificial probability of unknown to closed-set networks by computing a logit score for the unknown class based on deep feature similarities to features of known classes. Other approaches include a small adaptor network [13, 42] that adds open-set capabilities [10] to features extracted from closed-set networks.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.4 Learning-based Approaches", + "text_level": 1, + "bbox": [ + 214, + 268, + 488, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most methods use negative aka. known unknown samples in training, which represent samples of no interest to the classifier and should be classified as unknown, with a few exemptions such as replacing the final Softmax classifier with a set of binary classifiers [40, 43]. While some use real negative samples [10, 13, 30], the majority of research tries to artificially create them. For example, combinations of knowns [57], added noise to knowns [47], or Generative Adversarial Networks (GANs) [11, 20]. In our approach, we avoid the use of any negative samples.", + "bbox": [ + 212, + 301, + 787, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.5 Evaluation Techniques", + "text_level": 1, + "bbox": [ + 214, + 439, + 447, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluating OSR correctly in a real-world setting is difficult and current measures do not satisfy operational requirements. For example, the widely used AUROC metric [5,7,27,28,44,51] only looks at the binary decision of known vs. unknown, but ignores the task of assigning the correct class for known samples. Such metrics are often combined with closed-set classification accuracy in order to evaluate OSR. These evaluations are reasonable, but provide no information under realistic scenarios, i.e., an operational threshold classifying samples as known or unknown.", + "bbox": [ + 212, + 473, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another metric often employed is the macro-averaged F1 score [1,27,35,53], which treats unknown as a separate class. This measure has many counter-intuitive properties that make it difficult to interpret the results. For example, a different threshold is required for each class so a sample could be classified as several known classes and unknown (at the same time). Thus, incorrect classification of unknown samples can easily be overlooked.", + "bbox": [ + 212, + 580, + 787, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recently, Dhamija et al. [10] introduced the Open-Set Classification Rate (OSCR) curve. This curve handles knowns and unknowns separately, evaluating the Correct Classification Rate (CCR) at various thresholds corresponding to specific False Positive Rates (FPR). While valuable for high-risk applications like open-set face recognition, its practicality in general OSR tasks is questionable. In scenarios where a threshold is chosen based on a specific FPR, the curve utilizes unknowns from the test set, raising doubts about its applicability to unseen classes. Moreover, the curve faces the challenge of not providing a single value for comparing different methods. Additionally, the proposed metrics, Area Under the OSCR (AUOSCR) [6] and Open Area Under the ROC Curve (OpenAUC) [46], do not provide an intuition on how to select an operating threshold.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/87d6f851dd352aa9e8462bda1bcfc8cfd879633a0dfc6b83807698c1099d5913.jpg", + "image_caption": [ + "Fig. 2: MOTIVATION FOR NORMALIZATION. We illustrate why PostMax divides by the feature magnitude norm before applying GPD. The comparison above shows squared deep feature values extracted from a pre-trained ILSVRC2012-1K Transformer (ViT-H [14]). We sorted the deep features of 20 randomly sampled images from class 990 based on increasing mean square magnitude. Next, we applied this sorted order to the features of 20 images misclassified as class 990 from each unknown dataset. Unknowns exhibit high feature values (bright colors) in regions where known sample activations are low. These large, uncorrelated responses accumulate to give it a high score for the class but also result in unknown dataset features having, on average, larger $L_{2}$ norms than those of the correct class, depicted in the magnitude histogram on the right. Therefore, dividing by magnitude enhances separation." + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 632, + 232 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3165a2e2b460cea80cc92ed43a899d77a49b290f0b66439493bc949ca32b1eed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 146, + 781, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 426, + 346, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several approaches advocate selecting the maximum class and then applying thresholding to certain scores, e.g., Softmax [16] or logits [15], to address OSR. While such thresholding proves effective in some recognition tasks [44], its relevance often extends only to establishing a ranked order of classes for individual samples. We focus on leveraging information from the features themselves and the formal distribution of these maxima. The result is PostMax, an effective approach with theoretically grounded probabilities to address OSR through (1) normalization and (2) score distribution.", + "bbox": [ + 212, + 459, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Normalization & GPD", + "text_level": 1, + "bbox": [ + 214, + 603, + 449, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From Fig. 2, unknown samples activate high dimensions in the feature space, which goes against our intuition that deep networks would only learn features necessary to classify known classes and disregard those belonging to other uninteresting ones. We attribute these findings to high logit values and Softmax scores, where positions with high feature magnitudes align with larger weights of certain classes, even when an input exhibits a large feature magnitude and the high-scoring class has a small or negative weight. The observed effect may depend on the network's loss function and the training dataset; in small-scale datasets such as MNIST or CIFAR, which only differentiate between 10 classes, the opposite has been observed [10, 18, 31, 44]. Further details regarding the normalizations in these works and how they differ from our findings are provided in the supplemental.", + "bbox": [ + 212, + 628, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we classify samples based on their \"maximum score\" threshold, a key question arises: Does this score belong to the training distribution of maximum scores? Statistical Extreme Value Theory (EVT) provides a grounded approach to", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 POSTMAX FITTING. This implements the probability distribution estimation of the proposed PostMax algorithm." + ], + "code_body": "Require: $D_{\\mathrm{train}}$ $\\operatorname {FV}(x)$ $\\operatorname {L}(x)$ $M\\gets \\{\\}$ \nfor each $(x,t)\\in D_{\\mathrm{train}}$ do if arg max $\\mathrm{L}(x) = t$ then $\\ell \\leftarrow \\frac{\\max\\mathrm{L}(x)}{\\|\\mathrm{FV}(x)\\|}$ $M\\gets M\\cup \\{\\ell \\}$ end if \nend for \n $\\mu ,\\sigma ,\\xi \\gets$ scipy.stats.genpareto.fit(M)", + "guess_lang": "txt", + "bbox": [ + 215, + 176, + 506, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "answering this question. Unlike ad-hoc Softmax confidences, transforming logits into real probabilities allows probabilities to be more interpretable compared to Softmax, probabilities do not need to sum up to one (images may contain multiple objects), and EVT models probabilities on are based on thousands/millions of samples rather than the current sample. Given that thresholding maximum logits or Softmax scores serves as a reasonable baseline [15,32,44], applying EVT to this problem is both theoretically justified and intuitive. Note, prior OSR research [1,26,35,38] utilizes the Fisher-Tippett-Gnedenko EVT [12], resulting in a Weibull distribution. However, practical effectiveness was limited due to challenges in parameter selection affecting EVT modeling. In contrast, we employ the Generalized Pareto Distribution (GPD) derived from the Peak-over-threshold (POT) approach [33], which models extreme values above a threshold.", + "bbox": [ + 212, + 340, + 787, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 PostMax", + "text_level": 1, + "bbox": [ + 215, + 542, + 336, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using these intuitions and observations, we define our Postnormalization of Maxima (PostMax) algorithm. We aim to find a general distribution $\\varPsi$ of logit values that is valid for all of our known classes $c\\in \\{1,\\ldots ,C\\}$ . As observed in Fig. 2, modeling raw logit values might not be fruitful since unknown samples have generally higher activation of deep features, which often leads to high logit values. For a sample, we utilize normalized logit values by dividing the original logit values by their deep feature magnitude via (1).", + "bbox": [ + 212, + 566, + 787, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on POT EVT [33], we understand that if we consider all maximum values above a threshold (such as the smallest correct logit observed during training), the resulting distribution follows a Generalized Pareto Distribution (GPD). Let $D_{\\mathrm{train}} = \\{(x_i, t_i)\\}$ be the collection of all samples $x_i$ with their respective ground-truth target label $t_i$ in the training set. Let $\\mathrm{FV}(x)$ be the function returning the feature vector of a sample $x$ . Let $\\mathrm{L}(x)$ be the function returning the logit vector of a sample. We collect the normalized logits from the target class for all correctly classified training samples, which are then used to model a Generalized Pareto Distribution $\\varPsi_{\\mu, \\sigma, \\xi}$ . Details can be found in Alg. 1.", + "bbox": [ + 212, + 672, + 787, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At inference for test sample $x$ we use $\\mathrm{FV}(x)$ and $\\mathrm{L}(x)$ to extract the deep features and logits of $x$ and then for class $c$ assign probability $p_c(x)$ using the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "normalized logit value:", + "bbox": [ + 215, + 146, + 382, + 161 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c} = \\frac {\\operatorname {L} (x) [ c ]}{\\| \\operatorname {F V} (x) \\|} \\quad p _ {c} (x) = \\Psi_ {\\mu , \\sigma , \\xi} (\\ell_ {c}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 172, + 785, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varPsi_{\\mu,\\sigma,\\xi}(\\ell)$ is the cumulative distribution of the GPD with location $\\mu$ , scale $\\sigma$ and shape $\\xi$ computed via Alg. 1. If we just want the top class (as all experiments), we compute the class with the max logit and apply the above to that logit.", + "bbox": [ + 212, + 215, + 785, + 261 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that we do not utilize any unknowns or negatives in this process, nor do we explicitly model a probability of unknowns. Instead, we threshold based on the cumulative GPD probability of known classes, computing the maximum value of normalized logits.", + "bbox": [ + 212, + 262, + 785, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Operational Open-Set Accuracy", + "text_level": 1, + "bbox": [ + 215, + 345, + 524, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By designing the OpenAUC metric, Wang et al. [46] proposed four different conditions that a good open-set evaluation metric should fulfill, which we rephrase:", + "bbox": [ + 212, + 372, + 785, + 402 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "P1 The metric needs to check that known samples are classified correctly with high probability.", + "P2 The metric needs to evaluate if unknown samples are assigned to known classes with low probability." + ], + "bbox": [ + 210, + 402, + 785, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "P3 The metric should be insensitive to a score threshold.", + "bbox": [ + 210, + 463, + 625, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "P4 The metric should be a single number.", + "bbox": [ + 210, + 478, + 519, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We totally agree with P1, P2 and P4. However, we reject P3 because it means that the metric cannot be used to select a classifier that works well in a specific operational setting, which requires obtaining an operational threshold. In our view, a useful open-set algorithm must include that step. Therefore, we reformulate condition P3' to be operationally relevant:", + "bbox": [ + 212, + 493, + 785, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "P3' The metric should indicate an operational threshold that is optimal under specified circumstances and can be applied to unseen data and unseen unknown classes.", + "bbox": [ + 212, + 569, + 785, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These four conditions (P1, P2, P3', P4) exclude all existing open-set evaluation metrics; we need a new metric to satisfy them. As discussed in Sec. 2.5, metrics such as AUROC, AUOSCR, and OpenAUC simply average over all possible operational thresholds, while F1 and normalized accuracy fail to use a single threshold [46]. On the other hand, the Open-Set Classification Rate (OSCR) curve [10] does not provide a single value for comparison, which makes it hard to compare different algorithms. Therefore, we introduce Operational Open-Set Accuracy (OOSA), which is inspired by OSCR but fulfills all four conditions.", + "bbox": [ + 212, + 614, + 785, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From an engineer's view, the real question is how well a specific method can perform under the presence of both known and unknown samples in their data. Most applications of open-set classification are not security-sensitive, so there is no need to have very low False Positive Rates (FPR) as typically evaluated in the OSCR [10,30].", + "bbox": [ + 212, + 736, + 785, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Usually, an engineer has a rough estimate of how many known and unknown samples the system will observe - so they are rather interested in the total accuracy", + "bbox": [ + 212, + 811, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the system will achieve and which kind of operational threshold will lead to this performance. To provide both, we exploit the FPR and CCR calculations used in OSCR [10] to define the Unknown Rejection Rate (URR) and the Open-Set Accuracy (OSA), both of which rely on the operational threshold $\\theta$ and only make use of the probabilities of known classes $P_{c}(x)$ :", + "bbox": [ + 214, + 146, + 787, + 218 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C C R} (\\theta) = \\frac {\\left| \\left\\{x \\mid x \\in D _ {c} \\wedge \\arg \\max _ {c} P _ {c} (x) = \\hat {c} \\wedge P _ {c} (x) \\geq \\theta \\right\\} \\right|}{\\left| D _ {c} \\right|}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 227, + 696, + 257 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {U R R} (\\theta) = \\frac {\\left| \\left\\{x \\mid x \\in D _ {a} \\wedge \\max _ {c} P _ {c} (x) < \\theta \\right\\} \\right|}{\\left| D _ {a} \\right|} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 260, + 785, + 291 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {O S A} (\\theta) = \\alpha \\cdot \\mathrm {C C R} (\\theta) + (1 - \\alpha) \\cdot \\mathrm {U R R} (\\theta)\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 295, + 637, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $D_{c}$ is the collection of all known test samples with class labels $\\hat{c}$ , while $D_{a}$ collects all unknown test samples [10]. Please note that $\\mathrm{URR}(\\theta) = 1 - \\mathrm{FPR}(\\theta)$ provides a positive view on the unknown samples, i.e., it calculates how many of them are correctly rejected under threshold $\\theta$ . OSA includes a weighted average over CCR and URR to provide a holistic view of the expected total accuracy, where the application-dependent $\\alpha$ can be used to provide a stronger focus on correctly classifying known or unknown samples. When both types of samples should be treated similarly, $\\alpha = \\frac{|D_{c}|}{|D_{c}| + |D_{a}|}$ can be selected, which we do in most of our experiments. In that case, OSA can be thought of as evaluating a $C$ class classifier using $C + 1$ class accuracy where the unknown class $C + 1$ is classified when the maximal probability does not exceed our operational threshold $\\theta$ .", + "bbox": [ + 214, + 319, + 787, + 487 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When plotting OSA over URR, by varying the operational threshold, we arrive at the plots seen in Fig. 1. This plot is not monotonic, but rather provides a peak where the OSA is maximized. The maximum value can be used to select a threshold $\\theta^{*}$ on a given set of scores $S$ effectively:", + "bbox": [ + 214, + 488, + 787, + 547 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\arg \\max \\mathrm {O S A} _ {S} (\\theta)\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 560, + 581, + 575 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\in S \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 473, + 575, + 785, + 590 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nS = \\left\\{P _ {\\hat {c}} (x) \\mid x \\in D _ {c} \\right\\} \\cup \\left\\{\\max _ {c} P _ {c} (x) \\mid x \\in D _ {a} \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 588, + 663, + 607 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $S$ is the collection of all maximum scores for known and unknown samples. When applying this calculation on the test set $S_{test}$ , we will find the optimal threshold $\\theta_{test}^{*}$ , indicated by $\\diamond$ in Fig. 1. However, since test set samples are unknown in operational settings, we make use of the validation set $D_{val}$ of known samples as $D_{c}$ and an additional surrogate set $D_{sur}$ of unknown samples to replace $D_{a}$ in (3) for obtaining the operational threshold $\\theta_{sur}^{*}$ , which we will then apply to the test set: $\\mathrm{OSA}_{\\text{test}}(\\theta_{sur}^{*})$ , which is indicated by $\\star$ in Fig. 1. How to optimally select this surrogate set is discussed below.", + "bbox": [ + 214, + 619, + 787, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 762, + 375, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although a few OSR methods have explored large-scale datasets [1,30,35,44], many recent techniques [5-7,10,11,20,25,27,28,55,57] still lack comprehensive large-scale evaluations. The performance on small-scale data such as MNIST [22],", + "bbox": [ + 214, + 794, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: OPERATIONAL OPEN-SET ACCURACY (OOSA). The mean OOSA $(\\uparrow)$ of all methods. To compute, we validate methods on ImageNetV2 [34] (10K images) as knowns and $20\\%$ of our surrogate 21K-P Hard [44] (9.8K images) as unknowns and predict an operational threshold. Then, we deploy each method's threshold and test on five different ILSVRC2012 val [36] splits (each 10K images) and specified unknowns. Standard deviation is omitted as it is $< 0.005$ for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012 - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold.", + "bbox": [ + 212, + 143, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1b5eade47e5a8af1c203f365afaef56c244289cc1bd659f0b083519a4aa06a05.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8250.8150.7800.7820.7270.8120.7960.7860.7130.772
NINCO [2] (~5.8K)0.7400.7390.7070.6250.6670.7320.7230.7180.5970.705
OpenImage-O [45] (~17.6K)0.8640.8140.7450.7960.6630.8490.8200.7920.7450.767
Places [56] (10K)0.7850.7540.6910.7450.6250.7790.7490.7100.6930.676
SUN [49] (10K)0.7950.7580.7090.7490.6510.7830.7510.7210.6930.690
Textures [8] (~5.1K)0.7640.7560.7360.6600.7150.7440.7360.7400.6010.740
21K-P Easy [44] (50K)0.8180.7400.6620.7440.5680.8130.7630.6910.7830.633
", + "bbox": [ + 217, + 282, + 787, + 385 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CIFAR [21], or SVHN [29] does not reflect operational scenarios and lacks generalizability due to small image crops and a limited number of classes [30].", + "bbox": [ + 212, + 410, + 784, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "An open-set protocol is essential and necessitates a large-scale setting for the proper evaluation of real-world scenarios [30, 44]. Additionally, it is paramount to employ proper machine-learning methodology, including a validation set (and a surrogate set) for operational threshold selection. Thus, we evaluate the robustness and generalization of PostMax on multiple large-scale datasets and splits with realistic and high-quality images. We also leverage the recently introduced large-scale ImageNet open-set splits [44]. Lastly, we conduct ablations to understand the performance impact of our approach.", + "bbox": [ + 212, + 441, + 787, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Details", + "text_level": 1, + "bbox": [ + 215, + 582, + 320, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Architectures. We use state-of-the-art pre-trained models on ILSVRC2012 [36] with standard $224 \\times 224$ image crops - Meta's Hiera-H [37] and ConvNeXtV2-H [48]. Meta fine-tuned these models with no additional/external data. Note, we do not perform additional training or fine-tuning.", + "bbox": [ + 212, + 606, + 787, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. For knowns, we utilize the large-scale ImageNet [9] subset ILSVRC2012 [36]. This dataset provides an advantage due to the extensive variety of available pre-trained architectures, enabling us to compare with more diverse deep networks. Our PostMax model is trained on the ILSVRC2012 train split, which contains $\\sim 1.28M$ images and $1K$ classes (PostMax training requires no additional negative/unknown data). For validation, we employ ImageNetV2 [34], which contains $10K$ images. For testing, we split ILSVRC2012 val into five equal splits of $10K$ images each.", + "bbox": [ + 212, + 688, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For unknowns, we use iNaturalist [17], NINCO [2], OpenImage-O [45], Places [56], SUN [49], Textures [8], and ImageNet-21K-P Open-Set splits [44]. The vision", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7682b3609bbda7d20e90252620d03098f0e15613de476215b67606ba63321ff8.jpg", + "table_caption": [ + "Table 2: COMMON METRICS. The mean AUROC $(\\uparrow)$ and FPR95 $(\\downarrow)$ of all methods. To compute, we test each method on five different ILSVRC2012 val [36] splits (each $10K$ images) as knowns and specified unknowns. Standard deviation is omitted as it is $< 0.01$ for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012-1K - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold." + ], + "table_footnote": [], + "table_body": "
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .21.92 / .37.88 / .36.90 / .50.81 / .51.95 / .26.91 / .42.90 / .37.86 / .58.88 / .44
NINCO [2] (~ 5.8K).87 / .58.83 / .62.79 / .61.69 / .78.73 / .67.86 / .56.83 / .64.82 / .60.71 / .81.80 / .62
OpenImage-O [45] (~ 17.6K).95 / .25.88 / .49.80 / .49.88 / .45.72 / .61.94 / .29.88 / .49.87 / .44.81 / .70.85 / .47
Places [56] (10K).89 / .49.83 / .60.75 / .61.85 / .55.67 / .71.89 / .50.84 / .62.79 / .60.82 / .64.76 / .64
SUN [49] (10K).90 / .46.84 / .57.78 / .57.85 / .56.71 / .66.89 / .47.84 / .59.81 / .57.83 / .60.77 / .61
Textures [8] (~ 5.1K).93 / .32.88 / .47.86 / .42.74 / .86.83 / .46.91 / .34.87 / .49.88 / .40.73 / .85.88 / .38
21K-P Easy [44] (50K).86 / .50.79 / .64.72 / .64.80 / .68.65 / .73.85 / .51.79 / .65.75 / .64.78 / .72.72 / .68
", + "bbox": [ + 217, + 239, + 787, + 339 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "community [15, 16, 32, 44, 45, 50, 52, 54] has utilized these datasets as they present unique challenges not seen in small-scale settings. For validation, to predict an operational threshold, we utilize $20\\%$ of ImageNet-21K-P Hard (9.8K images) as our surrogate unknowns. For testing, NINCO, OpenImage-O, Textures, and ImageNet-21K-P Easy are entirely used, while others sample $10K$ images from a manual selection of classes. Further descriptions of all datasets and splits are provided in the supplemental.", + "bbox": [ + 212, + 367, + 787, + 473 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. To align with existing literature, we utilized common metrics in Tab. 2, namely, Area Under the Receiver Operating Curve (AUROC) and False Positive Rate at a fixed True Positive Rate of $95\\%$ (FPR95). Additional metrics like Area Under Open-Set Classification Rate (AUOSCR) [6] and Open Area Under ROC Curve (OpenAUC) [46] are provided in the supplemental. Alongside these, we introduce Operational Open-Set Accuracy (OOSA) in Tab. 1, where we predict an operational threshold on a validation and a surrogate set and deploy it at test time. This measure, described in detail in Sec. 3.3, serves as a more realistic alternative to existing measures. When evaluating OSR methods in real-world scenarios, other measures are insufficient as they do not provide an operating point for deployment and are usually only evaluated with a test set.", + "bbox": [ + 212, + 494, + 789, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 215, + 684, + 323, + 698 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our main results are presented in Tab. 1 and 2, where we report performance on each unknown dataset mentioned in Sec. 4.1. We compare our approach with two commonly seen methods, Maximum Softmax Probability (MSP) [16,44] and Maximum Logit (MaxLogit) [15,44], along with the recently introduced Nearest Neighbor Guidance (NNGuide) [32] and SCALE [50]. We omit comparisons with older techniques such as OpenMax [1] and EVM [35] as they are much worse. According to the OpenOOD [52,54] Benchmark leaderboard for ImageNet-1K,", + "bbox": [ + 212, + 709, + 787, + 816 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "5https://zjysteven.github.io/Open00D", + "bbox": [ + 236, + 824, + 519, + 840 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/09b181ade35b4aabf1ab802d64d31db6e336ca4f42aa94eee45ddf47f8da556e.jpg", + "table_caption": [ + "Table 3: DIFFERENT VALIDATION - KNOWS & UNKNOWN'S SURROGATE. The Operational Open-Set Accuracy (OOSA) $(\\uparrow)$ of all methods on a different validation set. To compute, we validate methods on ILSVRC2012 val [36] (50K images) as knowns and 21K-P Easy [44] (50K images) as unknowns surrogate and predict an operational threshold. Then, we deploy each method's threshold and test on ImageNetV2 [34] (10K images) and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold." + ], + "table_footnote": [], + "table_body": "
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8630.8080.7800.8080.7160.8420.8060.7940.7590.763
NINCO [2] (~5.8K)0.7330.7300.7040.6550.6540.7590.7330.7260.6430.695
OpenImage-O [45] (~17.6K)0.8850.8100.7360.8090.6490.8650.8210.7860.7460.757
Places [56] (10K)0.8010.7480.6840.7660.6110.7950.7530.7080.7310.664
SUN [49] (10K)0.8120.7520.7020.7710.6350.8020.7570.7220.7380.679
Textures [8] (~5.1K)0.8130.7460.7390.6910.7060.7840.7500.7570.6570.732
21K-P Hard [44] (49K)0.6340.6730.5830.5350.5150.6470.6720.5930.5590.566
", + "bbox": [ + 217, + 253, + 787, + 357 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "NNGuide stands out as the leader in the OOD space, closely followed by SCALE. We caution against relying solely on OOD methods and metrics, as they may not accurately reflect OSR performance. Note, some OSA and ROC curves are found in the supplemental.", + "bbox": [ + 212, + 381, + 787, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "When utilizing both architectures (Hiera-H & ConvNeXtV2-H) as feature extractors, PostMax outperforms all other methods in every measure (OOSA, AUROC, FPR95) on each unknown dataset. PostMax demonstrates superior performance, emphasizing robustness and generalizability. Effectively, PostMax achieves an excellent trade-off between classification accuracy and unknown rejection. Overall, the results suggest that a system operator can seamlessly integrate PostMax, enhance the performance beyond good closed-set classifiers, and confidently select the operational threshold that best suits its operation.", + "bbox": [ + 212, + 441, + 789, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Statistical Testing. To underscore the significance of performance differences between algorithms in our experiments, we prioritized the assessment of statistical significance. As a result, we performed paired, two-tailed t-tests, with Bonferroni corrections, to compare PostMax with the four methods (MSP, MaxLogit, NNGuide, and SCALE) using both architectures (Hiera-H and ConvNeXtV2-H) across 5 fold runs for each metric (OOSA, AUROC, and FPR95). On Hiera-H, PostMax (1) OOSA scores improvements range from $2.5\\%$ to $12.4\\%$ with corresponding P-values all less than $1.0\\mathrm{E}^{-7}$ , (2) AUROC scores improvements range from $5.7\\%$ to $17.9\\%$ with corresponding P-values all less than $1.0\\mathrm{E}^{-8}$ , and (3) FPR95 scores improvements range from $12.7\\%$ to $22.5\\%$ with corresponding P-values all less than $1.0\\mathrm{E}^{-8}$ . Similarly, on ConvNeXtV2-H, PostMax (1) OOSA scores improvements range from $2.3\\%$ to $10.4\\%$ , (2) AUROC scores improvements range from $4.7\\%$ to $10.6\\%$ , and (3) FPR95 scores improvements range from $9.6\\%$ to $28.1\\%$ , where all corresponding P-values are less than $1.0\\mathrm{E}^{-8}$ for all tests on all three metrics. In summary, the t-tests conducted for PostMax, considering the combined scores of all metrics across all datasets and architectures, revealed very statistically significant differences compared to every other method.", + "bbox": [ + 212, + 583, + 789, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a4f1e8b2377a2415f45d31460ae8d5da627e05efeec2c632bd7c1e757187f831.jpg", + "table_caption": [ + "Table 4: DIFFERENT TESTING - PRIOR METRICS. The AUROC (↑) and FPR95 (↓) of all methods on different test sets. To compute, we test each method on ImageNetV2 [34] (10K images) as knowns and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold." + ], + "table_footnote": [], + "table_body": "
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .16.92 / .37.88 / .37.91 / .40.80 / .55.95 / .23.91 / .41.90 / .37.87 / .53.87 / .44
NINCO [2] (~5.8K).88 / .52.83 / .61.77 / .62.71 / .72.70 / .71.87 / .54.83 / .63.81 / .60.73 / .78.79 / .63
OpenImage-O [45] (~17.6K).96 / .20.87 / .48.79 / .49.90 / .38.70 / .64.94 / .27.88 / .49.86 / .43.82 / .65.84 / .48
Places [56] (10K).90 / .44.83 / .60.74 / .62.87 / .48.65 / .74.89 / .48.83 / .61.78 / .60.84 / .59.74 / .64
SUN [49] (10K).91 / .41.84 / .57.77 / .58.87 / .49.69 / .70.90 / .44.84 / .59.80 / .57.84 / .55.76 / .61
Textures [8] (~5.1K).94 / .27.88 / .47.85 / .42.77 / .79.81 / .49.92 / .32.87 / .49.88 / .39.75 / .82.87 / .39
21K-P Hard [44] (49K).77 / .68.73 / .73.68 / .73.68 / .76.62 / .82.76 / .72.74 / .75.70 / .74.67 / .82.67 / .79
", + "bbox": [ + 217, + 212, + 787, + 311 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 339, + 388, + 356 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 3 and 4, we investigate the effects of varying the validation datasets (knowns and unknowns surrogate). We interchanged ImageNetV2 with ILSVRC2012 val for knowns and 21K-P Hard with 21K-P Easy for the unknowns surrogate. Unlike in Tab. 1, this involves validating on a larger number of knowns and unknowns, with an easier unknowns surrogate. Performance remained consistent regardless of the datasets.", + "bbox": [ + 212, + 363, + 795, + 453 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Also, we investigated the impact of other architectures as extractors and interchanged pre-trained models (ILSVRC2012-1K only w/ 224 crops) with Meta's ViT-H [14] and ConvNeXt-L [24]. Results showcasing PostMax's robustness are in the supplemental.", + "bbox": [ + 212, + 454, + 787, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lastly, we explored the impact of varying the application-dependent $\\alpha$ from Eq. (2) and the normalization $(L_{2})$ within PostMax. Details and results are in the supplemental.", + "bbox": [ + 212, + 513, + 787, + 559 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 582, + 354, + 598 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This paper introduces the new evaluation metric Open-Set Accuracy (OSA) (2); we framed the evaluation of an open-set problem as predicting a threshold and then evaluating open-set accuracy at that threshold. This supports an engineer selecting an algorithm and a threshold to be applied in some Operational setting.", + "bbox": [ + 212, + 613, + 787, + 674 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In real-world scenarios, unknown samples can be drawn from an unlimited variety, unlike known samples, which are assumed to be related to the training data. Thus, the evaluation must consider the computation of the predictive threshold on validation data and a separate evaluation on test data. The quality of the predicted threshold will depend on many factors with the two most influential being (1) how representative are the surrogate samples of the real unknowns, and (2) the relative proportion of knowns and unknowns. In Sec. 4.3 and our supplemental, we explore some variations of those factors.", + "bbox": [ + 212, + 674, + 787, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Maybe not surprisingly, if we train with \"harder\" unknowns, we generally see better performance. As shown in Tab. 1 and 2, the novel PostMax algorithm is better on unknowns across two networks (other networks in the supplemental).", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "However, as shown in Sec. 4.3, when 21K-P Easy is used as our unknowns surrogate, PostMax does not exhibit an advantage on 21K-P Hard. In general, 21K-P Hard poses the most significant challenge for all algorithms. Notably, the sets close visual and semantic similarity to known samples suggests it represents a fine-grained categorization challenge rather than a typical open-set scenario.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "One can adjust the importance of knowns vs. unknowns either by explicitly weighting them or implicitly through their relative percentages in the data. Although performance degrades when there is a significant disparity in their importance during threshold prediction stages, PostMax consistently outperforms and demonstrates greater stability across all thresholds. Refer to the supplemental for further details.", + "bbox": [ + 212, + 224, + 787, + 314 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Finally, to account for inherent random variations in predictions and tests, we developed training/validation/test splits, which are available on the GitHub repo. This protocol enables rigorous experimentation to assess the statistical significance of our evaluation. While the paper demonstrates statistically significant improvements across all results, some ablations in the supplemental do not.", + "bbox": [ + 212, + 316, + 787, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 428, + 359, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper contributes to OSR research by shifting focus from common metrics on small-scale experiments to more realistic scenarios and datasets, thereby fostering more practically relevant research. We revisited and revised the criteria for effective open-set metrics as suggested by Wang et al. [46], making them more applicable to real-world scenarios. Our proposed Operational Open-Set Accuracy (OOSA), which incorporates known and unknown samples, offers a comprehensive metric and facilitates the selection of the most effective method and operating threshold for system operators.", + "bbox": [ + 212, + 472, + 787, + 593 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our investigation of deep feature magnitudes of unknown samples from various ILSVRC2012 pre-trained networks yielded surprising findings. Contrary to smaller-scale studies [10, 31], we observed that these magnitudes are often larger for unknown than for known samples. This observation was consistent across multiple networks and types of unknown samples. Thus, we introduced our novel PostMax algorithm, refining normalized logit values from closed-set architectures into real probabilities of class inclusion using a Generalized Pareto Distribution (GPD). This distribution method offers advantages over previous extreme-value-based techniques, including negating the need to select a tail size and minimizing the influence of outliers. Also, it eliminates the need to define negative training samples as representatives of unknown classes, a common approach in much of the OSR literature that has not yielded a sufficient solution.", + "bbox": [ + 212, + 595, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While a good closed-set classifier is an important place to start [44], our novel PostMax algorithm shows that one can do better than \"all you need\" and that for operational usage, one needs to predict a threshold to use. With this paper, we empower engineers to operationalize their usage of OSR.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2016)", + "2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing ImageNet out-of-distribution detection evaluation. In: ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models (2023)", + "3. Bodesheim, P., Freytag, A., Rodner, E., Denzler, J.: Local novelty detection in multiclass recognition problems. In: Winter Conference on Applications of Computer Vision (WACV) (2015)", + "4. Boult, T.E., Cruz, S., Dhamija, A.R., Günther, M., Henrydoss, J., Scheirer, W.J.: Learning and the unknown: Surveying steps toward open world recognition. In: AAAI Conference on Artificial Intelligence. vol. 33, pp. 9801-9807 (2019)", + "5. Cevikalp, H., Uzun, B., Salk, Y., Saribas, H., Köpukli, O.: From anomaly detection to open set recognition: Bridging the gap. Pattern Recognition 138, 109385 (2023)", + "6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 8065-8081 (2021)", + "7. Chen, G., Qiao, L., Shi, Y., Peng, P., Li, J., Huang, T., Pu, S., Tian, Y.: Learning open set network with discriminative reciprocal points. In: European Conference on Computer Vision (ECCV). Springer (2020)", + "8. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2014)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009)", + "0. Dhamija, A.R., Günther, M., Boult, T.: Reducing network agnostophobia. Advances in Neural Information Processing Systems (NeurIPS) 31 (2018)", + "1. Ge, Z., Demyanov, S., Garnavi, R.: Generative OpenMax for multi-class open set classification. In: British Machine Vision Conference (BMVC) (2017)", + "2. Gumbel, E.J.: Statistical Theory of Extreme Values and Some Practical Applications: A Series of Lectures, vol. 33. US Government Printing Office (1954)", + "3. Günther, M., Dhamija, A.R., Boult, T.E.: Watchlist adaptation: Protecting the innocent. In: International Conference of the Biometrics Special Interest Group (BIOSIG) (2020)", + "4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (2022)", + "5. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (ICML). PMLR (2022)", + "6. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR) (2017)", + "7. van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The INaturalist species classification and detection dataset In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "8. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems (NeurIPS) 34, 677-689 (2021)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8710-8719 (2021)", + "20. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: International Conference on Computer Vision (ICCV). pp. 813-822 (2021)", + "21. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009)", + "22. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)", + "23. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23946-23955 (2023)", + "24. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11976-11986 (2022)", + "25. Lu, J., Xu, Y., Li, H., Cheng, Z., Niu, Y.: Pmal: Open set recognition via robust prototype mining. In: AAAI Conference on Artificial Intelligence. vol. 36:2, pp. 1872-1880 (2022)", + "26. Lyu, Z., Gutierrez, N.B., Beksi, W.J.: Metamax: Improved open-set deep neural networks via weibull calibration. In: Winter Conference on Applications of Computer Vision Workshops (WACVW) (2023)", + "27. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. In: European Conference on Computer Vision (ECCV). pp. 365-381. Springer (2022)", + "28. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: European Conference on Computer Vision (ECCV) (2018)", + "29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. In: NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)", + "30. Palechor, A., Bhoumik, A., Günther, M.: Large-scale open-set classification protocols for imagenet. In: Winter Conference on Applications of Computer Vision (WACV). pp. 42-51. CVF/IEEE (2023)", + "31. Park, J., Chai, J.C.L., Yoon, J., Teoh, A.B.J.: Understanding the feature norm for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1557-1567 (2023)", + "32. Park, J., Jung, Y.G., Teoh, A.B.J.: Nearest neighbor guidance for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1686-1695 (2023)", + "33. Pickands III, J.: Statistical inference using extreme order statistics. The Annals of Statistics 3(1), 119 - 131 (1975)", + "34. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do Imagenet classifiers generalize to Imagenet? In: International Conference on Machine Learning (ICML). pp. 5389-5400. PMLR (2019)", + "35. Rudd, E.M., Jain, L.P., Scheirer, W.J., Boult, T.E.: The extreme value machine. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2017)", + "36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. International Conference on Machine Learning (ICML) (2023)", + "38. Scheirer, W.J., Jain, L.P., Boult, T.E.: Probability models for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 36(11), 2317-2324 (2014)", + "39. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Towards open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 35(7) (2013)", + "40. Shu, L., Xu, H., Liu, B.: DOC: Deep open classification of text documents. In: Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics (2017)", + "41. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems (NeurIPS) 34, 144-157 (2021)", + "42. Vareto, R.H., Linghu, Y., Boult, T.E., Schwartz, W.R., Günther, M.: Open-set face recognition with maximal entropy and Objectosphere loss. Image and Vision Computing (IMAVIS) 141 (2024)", + "43. Vareto, R.H., Schwartz, W.R., Günther, M.: Toward open-set face recognition with neural ensemble, maximal entropy loss and feature augmentation. In: Conference on Graphics, Patterns and Images (SIBGRAPI) (2023)", + "44. Vaze, S., Han, K., Vedaldi, A., Zissermann, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022)", + "45. Wang, H., Li, Z., Feng, L., Zhang, W.: Out-of-distribution with virtual-logit matching. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4921-4930 (2022)", + "46. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards a priori-oriented open-set recognition. Advances in Neural Information Processing Systems (NeurIPS) 35, 25033-25045 (2022)", + "47. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sünderhauf, N.: Safe: Sensitivity-aware features for out-of-distribution object detection. In: International Conference on Computer Vision (ICCV). pp. 23565-23576 (2023)", + "48. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16133-16142 (2023)", + "49. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3485-3492. IEEE (2010)", + "50. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. In: International Conference on Learning Representations (ICLR) (2024)", + "51. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)", + "52. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: OpenOOD: Benchmarking generalized out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS) 35, 32598-32611 (2022)" + ], + "bbox": [ + 212, + 146, + 789, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Operational Open-Set Recognition and PostMax Refinement", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "54. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: OpenOOD v1.5: Enhanced benchmark for out-of-distribution detection. In: NeurIPS Workshop on Distribution Shifts: New Frontiers with Foundation Models (2023)", + "55. Zhang, X., Cheng, X., Zhang, D., Bonnington, P., Ge, Z.: Learning network architecture for open-set recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36(3), pp. 3362-3370 (2022)", + "56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017)", + "57. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 354 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Cruz et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_model.json b/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e11e9d61e79dae822a7c09ac310a8666344bbaa --- /dev/null +++ b/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_model.json @@ -0,0 +1,2436 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.28, + 0.141, + 0.726, + 0.187 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.213, + 0.667, + 0.244 + ], + "angle": 0, + "content": "Steve Cruz1, Ryan Rabinowitz2, Manuel Günther3, and Terrance E. Boult2" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.255, + 0.644, + 0.283 + ], + "angle": 0, + "content": "1 University of Notre Dame \n2 University of Colorado Colorado Springs" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.284, + 0.576, + 0.297 + ], + "angle": 0, + "content": "3 University of Zurich" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.298, + 0.747, + 0.312 + ], + "angle": 0, + "content": "stevecruz@nd.edu {rrabinow, tboult}@uccs.edu guenther@ifi.uzh.ch" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.351, + 0.744, + 0.629 + ], + "angle": 0, + "content": "Abstract. Open-Set Recognition (OSR) is a problem with mainly practical applications. However, recent evaluations have largely focused on small-scale data and tuning thresholds over the test set, which disregard the real-world operational needs of parameter selection. Thus, we revisit the original goals of OSR and propose a new evaluation metric, Operational Open-Set Accuracy (OOSA), which requires predicting an operationally relevant threshold from a validation set with known and a surrogate set with unknown samples, and then applying this threshold during testing. With this new measure in mind, we develop a large-scale evaluation protocol suited for operational scenarios. Additionally, we introduce the novel PostMax algorithm that performs post-processing refinement of the logit of the maximal class. This refinement involves normalizing logits by deep feature magnitudes and utilizing an extreme-value-based generalized Pareto distribution to map them into proper probabilities. We evaluate multiple pre-trained deep networks, including leading transformer and convolution-based architectures, on different selections of large-scale surrogate and test sets. Our experiments demonstrate that PostMax advances the state of the art in open-set recognition, showing statistically significant improvements in our novel OOSA metric as well as in previously used metrics such as AUROC, FPR95, and others." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.656, + 0.377, + 0.673 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "The original Open-Set Recognition (OSR) formulation [1,38,39] gained attention due to the flaws in existing recognition systems when handling inputs from outside the training classes. These systems lack control over inputs, and any unknown sample that is not rejected leads to a misclassification, thereby reducing accuracy. Therefore, early open-set work [1,3,35] introduced protocols tailored to emulate real-world systems, allowing for a proper evaluation of open-set performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For a decade, researchers explored techniques that enhanced OSR performance [1,5,7,10,11,20,25,27,28,35,39,55,57]. While contributing novel ideas, recent works have diverged from the original goals of OSR. Instead of leveraging large-scale data as in early research on open-set deep networks [1], recent OSR evaluations have" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.159, + 0.403, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.258, + 0.357, + 0.27 + ], + "angle": 0, + "content": "(a) Train" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.159, + 0.561, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.258, + 0.511, + 0.27 + ], + "angle": 0, + "content": "(b) Val" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.147, + 0.753, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.258, + 0.743, + 0.27 + ], + "angle": 0, + "content": "(c) Test (OOSA & AUROC)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.281, + 0.789, + 0.462 + ], + "angle": 0, + "content": "Fig. 1: POSTMAX PIPELINE. During train (a), post-processed maximum logits of correctly classified train samples are normalized by feature magnitude \\(\\ell\\); the set is then used in a Maximum Likelihood Estimation (MLE) to obtain Generalized Pareto Distribution (GPD) [33] with parameters \\(\\mu, \\sigma\\), and \\(\\xi\\). During validation (b), we use a validation and a surrogate set where the cumulative distribution of the GPD \\((\\Psi)\\) gives the probability of the normalized max logit being from the train distribution and predict an operational threshold. During operation (c), we deploy each method's threshold and apply it to determine OSR and compute Operational Open Set Accuracy (OOSA). For evaluation we compare against Maximum Softmax Probability (MSP) [16,44], Maximum Logit (MaxLogit) [15,44], Nearest Neighbor Guidance (NNGuide) [32], and SCALE [50]. Operational Open-Set Accuracy (OOSA) performance is \\((\\star)\\) and \\((\\diamond)\\) is the optimal Open-Set Accuracy (OSA) for an oracle that knows the final test set (shown to highlight differences from OOSA). AUROC scores are shown for comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.689 + ], + "angle": 0, + "content": "focused on small-scale datasets with few classes, with only a few notable exceptions [30,35,44]. Moreover, the prevalent reliance on metrics such as the Area Under the Receiver Operating Characteristics (AUROC) curve for performance evaluation has led to inopportune assessments since such metrics do not offer a realistic perspective on OSR. Fig. 1 illustrates an example where AUROC scores exhibit one ranking, yet in an operational setting, there is a slight difference. Inherently, common measures overlook the real-world operational needs of threshold selection. Some may argue for their use with a dedicated validation set; however, it is not clear how one would extract a threshold from these metrics and deploy it at test time. This is problematic as methods have largely ignored a fundamental practice of machine learning systems - having separate training, validation, and test sets to ensure that models generalize, avoid overfitting, and provide unbiased evaluation and reliable performance prediction on new data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Initial aspirations behind OSR aimed to enhance real systems [4]. Imagine an engineer tasked with implementing an OSR algorithm into their system. As they analyze results from various algorithms through tables and plots to inform decisions, a formidable challenge emerges. While AUROC and similar metrics assess a model's discriminative capability across all possible classification thresholds, they do not guide the engineer toward an optimal threshold for decision-making. There is a pressing need for a metric that enables confident and statistically rigorous evaluations with anticipated data, while also demonstrating robustness against varying numbers of unknowns. When deploying an open-set system, engineers must, at a minimum, consider two factors: (1) how an" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "approach scales with operational data, and (2) how to choose an operating point that balances their desired known classification accuracy and unknown rejection. These considerations underscore the importance of achieving overall accuracy with a predicted operational threshold. Therefore, we introduce a new metric called Operational Open-Set Accuracy (OOSA), illustrated in Fig. 1, which facilitates algorithm and operational threshold selection. To ensure real-world applicability, we implement a large-scale evaluation protocol with proper training/validation/test splits. We determine a threshold on the validation set and deploy it during testing on unseen unknowns. Additionally, for completeness, fairness, and robustness, we evaluate state-of-the-art algorithms on ImageNet-scale open-set recognition using common metrics (AUROC, FPR95, AUOSCR [6], and OpenAUC [46]), many of which are detailed in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Before incorporating an operational threshold, it is natural to ask about the distribution of scores above this threshold. This naturally leads to a Peak-over-Threshold (POT) formulation of the distribution, best modeled via Extreme Value Theory (EVT). The Pickands-Balkema-De Haan theorem of EVT [33] yields a Generalized Pareto Distribution (GPD) over normalized scores, as depicted in Fig. 1. This GPD-based score transformation provides theoretically grounded probabilities rather than raw, network-dependent ad-hoc confidence scores (see supplemental for details). While GPD can be applied directly to Softmax confidences or logit values, our approach goes further. We introduce a novel algorithm, PostMax, which applies GPD to post-processed normalized maximum logits. Here, normalization by the deep feature magnitude enhances the separation of known and unknown classes. Prior observations and theoretical explanations [10,13,31,44] suggest that deep feature magnitudes extracted from inputs that the network was not trained on are generally smaller than those for known samples. However, our findings show the exact opposite, i.e., that modern networks trained on the large-scale ILSVRC2012 dataset generally exhibit larger deep feature magnitudes for unknown samples (details in supplemental). Therefore, applying GPD to logit values divided by the deep feature magnitude can enhance OSR beyond the strong baselines of using raw Softmax confidence [16], maximum logit values [15], overconfidence reduction with features and logits [32], or scaled logit values [50]. The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.707 + ], + "angle": 0, + "content": "- We design a novel evaluation metric, Operational Open-Set Accuracy (OOSA), that emphasizes real-world usage by predicting an operational threshold on a validation set (knowns and unknowns surrogate), which is then deployed during testing on unseen data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.708, + 0.785, + 0.751 + ], + "angle": 0, + "content": "- We introduce a novel algorithm, PostMax, that post-processes maximum logits by normalizing them by deep feature magnitude and then applies GPD to provide probabilities. Our code is publicly available.4" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.753, + 0.785, + 0.781 + ], + "angle": 0, + "content": "- We develop new large-scale evaluation protocols suited for assessing algorithms in operational scenarios." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.783, + 0.785, + 0.813 + ], + "angle": 0, + "content": "- We showcase that PostMax advances the state of the art in large-scale open-set recognition with statistical significance on OOSA and prior metrics." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "footer", + "bbox": [ + 0.238, + 0.825, + 0.551, + 0.84 + ], + "angle": 0, + "content": "4https://github.com/Vastlab/PostMax-00SA" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Improving OSR, distinct from OOD detection, anomaly detection, or novel category discovery, can be approached in two ways: providing better features through improved network training, or through post-processing where a pretrained network is trained for closed-set tasks and then adapted for OSR. Our work is rooted in the post-processing approaches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.273, + 0.411, + 0.287 + ], + "angle": 0, + "content": "2.1 Related Problems" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.79, + 0.538 + ], + "angle": 0, + "content": "A common question regarding open-set is, What about a two-stage system with OOD followed by classification, does that not solve OSR? This hypothesis is rejected with an example from one of our operational systems where we have to recognize objects in novel contexts, e.g., in snowy or foggy conditions. Such conditions are OOD with respect to training samples, but they should not be rejected from an open-set point of view. Out-Of-Distribution (OOD) detection, though distinct from OSR, has been more consistent in its use of large-scale experimentation as were used in early OSR work [15,16,32,45,47,50,52,54]. Even when OOD is restricted to out-of-class, the two-stage system needs to be evaluated in that context, including a process for selecting an OOD threshold and then evaluating the resulting classification performance. The OpenOOD Benchmark [52,54] has implemented various OSR and OOD techniques such as OpenGan [20], MOS [19], ReAct [41], ViM [45], GEN [23], NNGuide [32], and SCALE [50]. While we caution the use of OOD algorithms in OSR evaluations/settings, based on the OpenOOD Benchmark leaderboard, we compare with the state-of-the-art, NNGuide & SCALE." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.44, + 0.574 + ], + "angle": 0, + "content": "2.2 Closed-Set Classifiers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.644 + ], + "angle": 0, + "content": "Recently, Vaze et al. [44] argued that closed-set classifiers are sufficient. This aligns with Hendrycks et al. [15, 16], who demonstrated thresholding on Softmax confidences or, especially, on logits provide unreasonably good baselines. Our work shows that such approaches/networks can be improved with normalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.792, + 0.765 + ], + "angle": 0, + "content": "We evaluate several leading pre-trained architectures trained on ILSVRC2012-1K with no additional data. Particularly, Meta's Vision Transformer Hiera-H [37], which strips non-essential components making it faster and more accurate during training/inference. Also, we use their Masked Autoencoders (MAE) model, ViT-H [14], which masks random patches of an image and reconstructs the missing pixels. To show generalization, we also utilize CNNs, including Meta's ConvNeXtV2-H [48] and ConvNeXt-L [24] which modernize a standard ResNet. Other networks are found in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.494, + 0.803 + ], + "angle": 0, + "content": "2.3 Post-processing Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Besides thresholding softmax scores or logits [15, 16, 44], OSR methods take features and use Weibull-calibrated Support Vector Machines (W-SVM) [38]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "or Extreme Value Machines (EVM) [35] to estimate a probability of unknown. OpenMax [1] was the first method to add an artificial probability of unknown to closed-set networks by computing a logit score for the unknown class based on deep feature similarities to features of known classes. Other approaches include a small adaptor network [13, 42] that adds open-set capabilities [10] to features extracted from closed-set networks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.269, + 0.49, + 0.286 + ], + "angle": 0, + "content": "2.4 Learning-based Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.789, + 0.412 + ], + "angle": 0, + "content": "Most methods use negative aka. known unknown samples in training, which represent samples of no interest to the classifier and should be classified as unknown, with a few exemptions such as replacing the final Softmax classifier with a set of binary classifiers [40, 43]. While some use real negative samples [10, 13, 30], the majority of research tries to artificially create them. For example, combinations of knowns [57], added noise to knowns [47], or Generative Adversarial Networks (GANs) [11, 20]. In our approach, we avoid the use of any negative samples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.44, + 0.449, + 0.456 + ], + "angle": 0, + "content": "2.5 Evaluation Techniques" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.581 + ], + "angle": 0, + "content": "Evaluating OSR correctly in a real-world setting is difficult and current measures do not satisfy operational requirements. For example, the widely used AUROC metric [5,7,27,28,44,51] only looks at the binary decision of known vs. unknown, but ignores the task of assigning the correct class for known samples. Such metrics are often combined with closed-set classification accuracy in order to evaluate OSR. These evaluations are reasonable, but provide no information under realistic scenarios, i.e., an operational threshold classifying samples as known or unknown." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.673 + ], + "angle": 0, + "content": "Another metric often employed is the macro-averaged F1 score [1,27,35,53], which treats unknown as a separate class. This measure has many counter-intuitive properties that make it difficult to interpret the results. For example, a different threshold is required for each class so a sample could be classified as several known classes and unknown (at the same time). Thus, incorrect classification of unknown samples can easily be overlooked." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Recently, Dhamija et al. [10] introduced the Open-Set Classification Rate (OSCR) curve. This curve handles knowns and unknowns separately, evaluating the Correct Classification Rate (CCR) at various thresholds corresponding to specific False Positive Rates (FPR). While valuable for high-risk applications like open-set face recognition, its practicality in general OSR tasks is questionable. In scenarios where a threshold is chosen based on a specific FPR, the curve utilizes unknowns from the test set, raising doubts about its applicability to unseen classes. Moreover, the curve faces the challenge of not providing a single value for comparing different methods. Additionally, the proposed metrics, Area Under the OSCR (AUOSCR) [6] and Open Area Under the ROC Curve (OpenAUC) [46], do not provide an intuition on how to select an operating threshold." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.148, + 0.633, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.147, + 0.782, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.247, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Fig. 2: MOTIVATION FOR NORMALIZATION. We illustrate why PostMax divides by the feature magnitude norm before applying GPD. The comparison above shows squared deep feature values extracted from a pre-trained ILSVRC2012-1K Transformer (ViT-H [14]). We sorted the deep features of 20 randomly sampled images from class 990 based on increasing mean square magnitude. Next, we applied this sorted order to the features of 20 images misclassified as class 990 from each unknown dataset. Unknowns exhibit high feature values (bright colors) in regions where known sample activations are low. These large, uncorrelated responses accumulate to give it a high score for the class but also result in unknown dataset features having, on average, larger \\( L_{2} \\) norms than those of the correct class, depicted in the magnitude histogram on the right. Therefore, dividing by magnitude enhances separation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.428, + 0.347, + 0.446 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.789, + 0.581 + ], + "angle": 0, + "content": "Several approaches advocate selecting the maximum class and then applying thresholding to certain scores, e.g., Softmax [16] or logits [15], to address OSR. While such thresholding proves effective in some recognition tasks [44], its relevance often extends only to establishing a ranked order of classes for individual samples. We focus on leveraging information from the features themselves and the formal distribution of these maxima. The result is PostMax, an effective approach with theoretically grounded probabilities to address OSR through (1) normalization and (2) score distribution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.45, + 0.618 + ], + "angle": 0, + "content": "3.1 Normalization & GPD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.795 + ], + "angle": 0, + "content": "From Fig. 2, unknown samples activate high dimensions in the feature space, which goes against our intuition that deep networks would only learn features necessary to classify known classes and disregard those belonging to other uninteresting ones. We attribute these findings to high logit values and Softmax scores, where positions with high feature magnitudes align with larger weights of certain classes, even when an input exhibits a large feature magnitude and the high-scoring class has a small or negative weight. The observed effect may depend on the network's loss function and the training dataset; in small-scale datasets such as MNIST or CIFAR, which only differentiate between 10 classes, the opposite has been observed [10, 18, 31, 44]. Further details regarding the normalizations in these works and how they differ from our findings are provided in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As we classify samples based on their \"maximum score\" threshold, a key question arises: Does this score belong to the training distribution of maximum scores? Statistical Extreme Value Theory (EVT) provides a grounded approach to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "Algorithm 1 POSTMAX FITTING. This implements the probability distribution estimation of the proposed PostMax algorithm." + }, + { + "type": "code", + "bbox": [ + 0.217, + 0.178, + 0.507, + 0.309 + ], + "angle": 0, + "content": "Require: \\(D_{\\mathrm{train}}\\) \\(\\operatorname {FV}(x)\\) \\(\\operatorname {L}(x)\\) \\(M\\gets \\{\\}\\) \nfor each \\((x,t)\\in D_{\\mathrm{train}}\\) do if arg max \\(\\mathrm{L}(x) = t\\) then \\(\\ell \\leftarrow \\frac{\\max\\mathrm{L}(x)}{\\|\\mathrm{FV}(x)\\|}\\) \\(M\\gets M\\cup \\{\\ell \\}\\) end if \nend for \n\\(\\mu ,\\sigma ,\\xi \\gets\\) scipy.stats.genpareto.fit(M)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.523 + ], + "angle": 0, + "content": "answering this question. Unlike ad-hoc Softmax confidences, transforming logits into real probabilities allows probabilities to be more interpretable compared to Softmax, probabilities do not need to sum up to one (images may contain multiple objects), and EVT models probabilities on are based on thousands/millions of samples rather than the current sample. Given that thresholding maximum logits or Softmax scores serves as a reasonable baseline [15,32,44], applying EVT to this problem is both theoretically justified and intuitive. Note, prior OSR research [1,26,35,38] utilizes the Fisher-Tippett-Gnedenko EVT [12], resulting in a Weibull distribution. However, practical effectiveness was limited due to challenges in parameter selection affecting EVT modeling. In contrast, we employ the Generalized Pareto Distribution (GPD) derived from the Peak-over-threshold (POT) approach [33], which models extreme values above a threshold." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.338, + 0.558 + ], + "angle": 0, + "content": "3.2 PostMax" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Using these intuitions and observations, we define our Postnormalization of Maxima (PostMax) algorithm. We aim to find a general distribution \\(\\varPsi\\) of logit values that is valid for all of our known classes \\(c\\in \\{1,\\ldots ,C\\}\\). As observed in Fig. 2, modeling raw logit values might not be fruitful since unknown samples have generally higher activation of deep features, which often leads to high logit values. For a sample, we utilize normalized logit values by dividing the original logit values by their deep feature magnitude via (1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.811 + ], + "angle": 0, + "content": "Based on POT EVT [33], we understand that if we consider all maximum values above a threshold (such as the smallest correct logit observed during training), the resulting distribution follows a Generalized Pareto Distribution (GPD). Let \\( D_{\\mathrm{train}} = \\{(x_i, t_i)\\} \\) be the collection of all samples \\( x_i \\) with their respective ground-truth target label \\( t_i \\) in the training set. Let \\( \\mathrm{FV}(x) \\) be the function returning the feature vector of a sample \\( x \\). Let \\( \\mathrm{L}(x) \\) be the function returning the logit vector of a sample. We collect the normalized logits from the target class for all correctly classified training samples, which are then used to model a Generalized Pareto Distribution \\( \\varPsi_{\\mu, \\sigma, \\xi} \\). Details can be found in Alg. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "At inference for test sample \\( x \\) we use \\( \\mathrm{FV}(x) \\) and \\( \\mathrm{L}(x) \\) to extract the deep features and logits of \\( x \\) and then for class \\( c \\) assign probability \\( p_c(x) \\) using the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.128 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.383, + 0.162 + ], + "angle": 0, + "content": "normalized logit value:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.174, + 0.787, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c} = \\frac {\\operatorname {L} (x) [ c ]}{\\| \\operatorname {F V} (x) \\|} \\quad p _ {c} (x) = \\Psi_ {\\mu , \\sigma , \\xi} (\\ell_ {c}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.217, + 0.786, + 0.262 + ], + "angle": 0, + "content": "where \\(\\varPsi_{\\mu,\\sigma,\\xi}(\\ell)\\) is the cumulative distribution of the GPD with location \\(\\mu\\), scale \\(\\sigma\\) and shape \\(\\xi\\) computed via Alg. 1. If we just want the top class (as all experiments), we compute the class with the max logit and apply the above to that logit." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.263, + 0.787, + 0.324 + ], + "angle": 0, + "content": "Note that we do not utilize any unknowns or negatives in this process, nor do we explicitly model a probability of unknowns. Instead, we threshold based on the cumulative GPD probability of known classes, computing the maximum value of normalized logits." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.347, + 0.525, + 0.362 + ], + "angle": 0, + "content": "3.3 Operational Open-Set Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.786, + 0.403 + ], + "angle": 0, + "content": "By designing the OpenAUC metric, Wang et al. [46] proposed four different conditions that a good open-set evaluation metric should fulfill, which we rephrase:" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.404, + 0.786, + 0.433 + ], + "angle": 0, + "content": "P1 The metric needs to check that known samples are classified correctly with high probability." + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.434, + 0.786, + 0.463 + ], + "angle": 0, + "content": "P2 The metric needs to evaluate if unknown samples are assigned to known classes with low probability." + }, + { + "type": "list", + "bbox": [ + 0.212, + 0.404, + 0.786, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.464, + 0.627, + 0.478 + ], + "angle": 0, + "content": "P3 The metric should be insensitive to a score threshold." + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.479, + 0.521, + 0.493 + ], + "angle": 0, + "content": "P4 The metric should be a single number." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.569 + ], + "angle": 0, + "content": "We totally agree with P1, P2 and P4. However, we reject P3 because it means that the metric cannot be used to select a classifier that works well in a specific operational setting, which requires obtaining an operational threshold. In our view, a useful open-set algorithm must include that step. Therefore, we reformulate condition P3' to be operationally relevant:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.787, + 0.614 + ], + "angle": 0, + "content": "P3' The metric should indicate an operational threshold that is optimal under specified circumstances and can be applied to unseen data and unseen unknown classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.787, + 0.736 + ], + "angle": 0, + "content": "These four conditions (P1, P2, P3', P4) exclude all existing open-set evaluation metrics; we need a new metric to satisfy them. As discussed in Sec. 2.5, metrics such as AUROC, AUOSCR, and OpenAUC simply average over all possible operational thresholds, while F1 and normalized accuracy fail to use a single threshold [46]. On the other hand, the Open-Set Classification Rate (OSCR) curve [10] does not provide a single value for comparison, which makes it hard to compare different algorithms. Therefore, we introduce Operational Open-Set Accuracy (OOSA), which is inspired by OSCR but fulfills all four conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.737, + 0.787, + 0.812 + ], + "angle": 0, + "content": "From an engineer's view, the real question is how well a specific method can perform under the presence of both known and unknown samples in their data. Most applications of open-set classification are not security-sensitive, so there is no need to have very low False Positive Rates (FPR) as typically evaluated in the OSCR [10,30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Usually, an engineer has a rough estimate of how many known and unknown samples the system will observe - so they are rather interested in the total accuracy" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.219 + ], + "angle": 0, + "content": "the system will achieve and which kind of operational threshold will lead to this performance. To provide both, we exploit the FPR and CCR calculations used in OSCR [10] to define the Unknown Rejection Rate (URR) and the Open-Set Accuracy (OSA), both of which rely on the operational threshold \\(\\theta\\) and only make use of the probabilities of known classes \\(P_{c}(x)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.228, + 0.697, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C C R} (\\theta) = \\frac {\\left| \\left\\{x \\mid x \\in D _ {c} \\wedge \\arg \\max _ {c} P _ {c} (x) = \\hat {c} \\wedge P _ {c} (x) \\geq \\theta \\right\\} \\right|}{\\left| D _ {c} \\right|}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.261, + 0.786, + 0.292 + ], + "angle": 0, + "content": "\\[\n\\operatorname {U R R} (\\theta) = \\frac {\\left| \\left\\{x \\mid x \\in D _ {a} \\wedge \\max _ {c} P _ {c} (x) < \\theta \\right\\} \\right|}{\\left| D _ {a} \\right|} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.296, + 0.638, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\mathrm {O S A} (\\theta) = \\alpha \\cdot \\mathrm {C C R} (\\theta) + (1 - \\alpha) \\cdot \\mathrm {U R R} (\\theta)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.32, + 0.789, + 0.488 + ], + "angle": 0, + "content": "where \\( D_{c} \\) is the collection of all known test samples with class labels \\( \\hat{c} \\), while \\( D_{a} \\) collects all unknown test samples [10]. Please note that \\( \\mathrm{URR}(\\theta) = 1 - \\mathrm{FPR}(\\theta) \\) provides a positive view on the unknown samples, i.e., it calculates how many of them are correctly rejected under threshold \\( \\theta \\). OSA includes a weighted average over CCR and URR to provide a holistic view of the expected total accuracy, where the application-dependent \\( \\alpha \\) can be used to provide a stronger focus on correctly classifying known or unknown samples. When both types of samples should be treated similarly, \\( \\alpha = \\frac{|D_{c}|}{|D_{c}| + |D_{a}|} \\) can be selected, which we do in most of our experiments. In that case, OSA can be thought of as evaluating a \\( C \\) class classifier using \\( C + 1 \\) class accuracy where the unknown class \\( C + 1 \\) is classified when the maximal probability does not exceed our operational threshold \\( \\theta \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.489, + 0.788, + 0.549 + ], + "angle": 0, + "content": "When plotting OSA over URR, by varying the operational threshold, we arrive at the plots seen in Fig. 1. This plot is not monotonic, but rather provides a peak where the OSA is maximized. The maximum value can be used to select a threshold \\(\\theta^{*}\\) on a given set of scores \\(S\\) effectively:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.561, + 0.583, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\arg \\max \\mathrm {O S A} _ {S} (\\theta)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.475, + 0.576, + 0.786, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\theta \\in S \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.589, + 0.665, + 0.608 + ], + "angle": 0, + "content": "\\[\nS = \\left\\{P _ {\\hat {c}} (x) \\mid x \\in D _ {c} \\right\\} \\cup \\left\\{\\max _ {c} P _ {c} (x) \\mid x \\in D _ {a} \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.62, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Here, \\( S \\) is the collection of all maximum scores for known and unknown samples. When applying this calculation on the test set \\( S_{test} \\), we will find the optimal threshold \\( \\theta_{test}^{*} \\), indicated by \\( \\diamond \\) in Fig. 1. However, since test set samples are unknown in operational settings, we make use of the validation set \\( D_{val} \\) of known samples as \\( D_{c} \\) and an additional surrogate set \\( D_{sur} \\) of unknown samples to replace \\( D_{a} \\) in (3) for obtaining the operational threshold \\( \\theta_{sur}^{*} \\), which we will then apply to the test set: \\( \\mathrm{OSA}_{\\text{test}}(\\theta_{sur}^{*}) \\), which is indicated by \\( \\star \\) in Fig. 1. How to optimally select this surrogate set is discussed below." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.763, + 0.376, + 0.781 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Although a few OSR methods have explored large-scale datasets [1,30,35,44], many recent techniques [5-7,10,11,20,25,27,28,55,57] still lack comprehensive large-scale evaluations. The performance on small-scale data such as MNIST [22]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.27 + ], + "angle": 0, + "content": "Table 1: OPERATIONAL OPEN-SET ACCURACY (OOSA). The mean OOSA \\((\\uparrow)\\) of all methods. To compute, we validate methods on ImageNetV2 [34] (10K images) as knowns and \\(20\\%\\) of our surrogate 21K-P Hard [44] (9.8K images) as unknowns and predict an operational threshold. Then, we deploy each method's threshold and test on five different ILSVRC2012 val [36] splits (each 10K images) and specified unknowns. Standard deviation is omitted as it is \\(< 0.005\\) for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012 - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.283, + 0.789, + 0.386 + ], + "angle": 0, + "content": "
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8250.8150.7800.7820.7270.8120.7960.7860.7130.772
NINCO [2] (~5.8K)0.7400.7390.7070.6250.6670.7320.7230.7180.5970.705
OpenImage-O [45] (~17.6K)0.8640.8140.7450.7960.6630.8490.8200.7920.7450.767
Places [56] (10K)0.7850.7540.6910.7450.6250.7790.7490.7100.6930.676
SUN [49] (10K)0.7950.7580.7090.7490.6510.7830.7510.7210.6930.690
Textures [8] (~5.1K)0.7640.7560.7360.6600.7150.7440.7360.7400.6010.740
21K-P Easy [44] (50K)0.8180.7400.6620.7440.5680.8130.7630.6910.7830.633
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.44 + ], + "angle": 0, + "content": "CIFAR [21], or SVHN [29] does not reflect operational scenarios and lacks generalizability due to small image crops and a limited number of classes [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.562 + ], + "angle": 0, + "content": "An open-set protocol is essential and necessitates a large-scale setting for the proper evaluation of real-world scenarios [30, 44]. Additionally, it is paramount to employ proper machine-learning methodology, including a validation set (and a surrogate set) for operational threshold selection. Thus, we evaluate the robustness and generalization of PostMax on multiple large-scale datasets and splits with realistic and high-quality images. We also leverage the recently introduced large-scale ImageNet open-set splits [44]. Lastly, we conduct ablations to understand the performance impact of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.583, + 0.321, + 0.597 + ], + "angle": 0, + "content": "4.1 Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Architectures. We use state-of-the-art pre-trained models on ILSVRC2012 [36] with standard \\(224 \\times 224\\) image crops - Meta's Hiera-H [37] and ConvNeXtV2-H [48]. Meta fine-tuned these models with no additional/external data. Note, we do not perform additional training or fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Datasets. For knowns, we utilize the large-scale ImageNet [9] subset ILSVRC2012 [36]. This dataset provides an advantage due to the extensive variety of available pre-trained architectures, enabling us to compare with more diverse deep networks. Our PostMax model is trained on the ILSVRC2012 train split, which contains \\(\\sim 1.28M\\) images and \\(1K\\) classes (PostMax training requires no additional negative/unknown data). For validation, we employ ImageNetV2 [34], which contains \\(10K\\) images. For testing, we split ILSVRC2012 val into five equal splits of \\(10K\\) images each." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "For unknowns, we use iNaturalist [17], NINCO [2], OpenImage-O [45], Places [56], SUN [49], Textures [8], and ImageNet-21K-P Open-Set splits [44]. The vision" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: COMMON METRICS. The mean AUROC \\((\\uparrow)\\) and FPR95 \\((\\downarrow)\\) of all methods. To compute, we test each method on five different ILSVRC2012 val [36] splits (each \\(10K\\) images) as knowns and specified unknowns. Standard deviation is omitted as it is \\(< 0.01\\) for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012-1K - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.24, + 0.789, + 0.34 + ], + "angle": 0, + "content": "
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .21.92 / .37.88 / .36.90 / .50.81 / .51.95 / .26.91 / .42.90 / .37.86 / .58.88 / .44
NINCO [2] (~ 5.8K).87 / .58.83 / .62.79 / .61.69 / .78.73 / .67.86 / .56.83 / .64.82 / .60.71 / .81.80 / .62
OpenImage-O [45] (~ 17.6K).95 / .25.88 / .49.80 / .49.88 / .45.72 / .61.94 / .29.88 / .49.87 / .44.81 / .70.85 / .47
Places [56] (10K).89 / .49.83 / .60.75 / .61.85 / .55.67 / .71.89 / .50.84 / .62.79 / .60.82 / .64.76 / .64
SUN [49] (10K).90 / .46.84 / .57.78 / .57.85 / .56.71 / .66.89 / .47.84 / .59.81 / .57.83 / .60.77 / .61
Textures [8] (~ 5.1K).93 / .32.88 / .47.86 / .42.74 / .86.83 / .46.91 / .34.87 / .49.88 / .40.73 / .85.88 / .38
21K-P Easy [44] (50K).86 / .50.79 / .64.72 / .64.80 / .68.65 / .73.85 / .51.79 / .65.75 / .64.78 / .72.72 / .68
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.789, + 0.474 + ], + "angle": 0, + "content": "community [15, 16, 32, 44, 45, 50, 52, 54] has utilized these datasets as they present unique challenges not seen in small-scale settings. For validation, to predict an operational threshold, we utilize \\(20\\%\\) of ImageNet-21K-P Hard (9.8K images) as our surrogate unknowns. For testing, NINCO, OpenImage-O, Textures, and ImageNet-21K-P Easy are entirely used, while others sample \\(10K\\) images from a manual selection of classes. Further descriptions of all datasets and splits are provided in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.79, + 0.663 + ], + "angle": 0, + "content": "Evaluation Metrics. To align with existing literature, we utilized common metrics in Tab. 2, namely, Area Under the Receiver Operating Curve (AUROC) and False Positive Rate at a fixed True Positive Rate of \\(95\\%\\) (FPR95). Additional metrics like Area Under Open-Set Classification Rate (AUOSCR) [6] and Open Area Under ROC Curve (OpenAUC) [46] are provided in the supplemental. Alongside these, we introduce Operational Open-Set Accuracy (OOSA) in Tab. 1, where we predict an operational threshold on a validation and a surrogate set and deploy it at test time. This measure, described in detail in Sec. 3.3, serves as a more realistic alternative to existing measures. When evaluating OSR methods in real-world scenarios, other measures are insufficient as they do not provide an operating point for deployment and are usually only evaluated with a test set." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.325, + 0.699 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.789, + 0.817 + ], + "angle": 0, + "content": "Our main results are presented in Tab. 1 and 2, where we report performance on each unknown dataset mentioned in Sec. 4.1. We compare our approach with two commonly seen methods, Maximum Softmax Probability (MSP) [16,44] and Maximum Logit (MaxLogit) [15,44], along with the recently introduced Nearest Neighbor Guidance (NNGuide) [32] and SCALE [50]. We omit comparisons with older techniques such as OpenMax [1] and EVM [35] as they are much worse. According to the OpenOOD [52,54] Benchmark leaderboard for ImageNet-1K," + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.521, + 0.841 + ], + "angle": 0, + "content": "5https://zjysteven.github.io/Open00D" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 3: DIFFERENT VALIDATION - KNOWS & UNKNOWN'S SURROGATE. The Operational Open-Set Accuracy (OOSA) \\((\\uparrow)\\) of all methods on a different validation set. To compute, we validate methods on ILSVRC2012 val [36] (50K images) as knowns and 21K-P Easy [44] (50K images) as unknowns surrogate and predict an operational threshold. Then, we deploy each method's threshold and test on ImageNetV2 [34] (10K images) and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.255, + 0.789, + 0.358 + ], + "angle": 0, + "content": "
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8630.8080.7800.8080.7160.8420.8060.7940.7590.763
NINCO [2] (~5.8K)0.7330.7300.7040.6550.6540.7590.7330.7260.6430.695
OpenImage-O [45] (~17.6K)0.8850.8100.7360.8090.6490.8650.8210.7860.7460.757
Places [56] (10K)0.8010.7480.6840.7660.6110.7950.7530.7080.7310.664
SUN [49] (10K)0.8120.7520.7020.7710.6350.8020.7570.7220.7380.679
Textures [8] (~5.1K)0.8130.7460.7390.6910.7060.7840.7500.7570.6570.732
21K-P Hard [44] (49K)0.6340.6730.5830.5350.5150.6470.6720.5930.5590.566
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.789, + 0.442 + ], + "angle": 0, + "content": "NNGuide stands out as the leader in the OOD space, closely followed by SCALE. We caution against relying solely on OOD methods and metrics, as they may not accurately reflect OSR performance. Note, some OSA and ROC curves are found in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.79, + 0.563 + ], + "angle": 0, + "content": "When utilizing both architectures (Hiera-H & ConvNeXtV2-H) as feature extractors, PostMax outperforms all other methods in every measure (OOSA, AUROC, FPR95) on each unknown dataset. PostMax demonstrates superior performance, emphasizing robustness and generalizability. Effectively, PostMax achieves an excellent trade-off between classification accuracy and unknown rejection. Overall, the results suggest that a system operator can seamlessly integrate PostMax, enhance the performance beyond good closed-set classifiers, and confidently select the operational threshold that best suits its operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Statistical Testing. To underscore the significance of performance differences between algorithms in our experiments, we prioritized the assessment of statistical significance. As a result, we performed paired, two-tailed t-tests, with Bonferroni corrections, to compare PostMax with the four methods (MSP, MaxLogit, NNGuide, and SCALE) using both architectures (Hiera-H and ConvNeXtV2-H) across 5 fold runs for each metric (OOSA, AUROC, and FPR95). On Hiera-H, PostMax (1) OOSA scores improvements range from \\(2.5\\%\\) to \\(12.4\\%\\) with corresponding P-values all less than \\(1.0\\mathrm{E}^{-7}\\), (2) AUROC scores improvements range from \\(5.7\\%\\) to \\(17.9\\%\\) with corresponding P-values all less than \\(1.0\\mathrm{E}^{-8}\\), and (3) FPR95 scores improvements range from \\(12.7\\%\\) to \\(22.5\\%\\) with corresponding P-values all less than \\(1.0\\mathrm{E}^{-8}\\). Similarly, on ConvNeXtV2-H, PostMax (1) OOSA scores improvements range from \\(2.3\\%\\) to \\(10.4\\%\\), (2) AUROC scores improvements range from \\(4.7\\%\\) to \\(10.6\\%\\), and (3) FPR95 scores improvements range from \\(9.6\\%\\) to \\(28.1\\%\\), where all corresponding P-values are less than \\(1.0\\mathrm{E}^{-8}\\) for all tests on all three metrics. In summary, the t-tests conducted for PostMax, considering the combined scores of all metrics across all datasets and architectures, revealed very statistically significant differences compared to every other method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 4: DIFFERENT TESTING - PRIOR METRICS. The AUROC (↑) and FPR95 (↓) of all methods on different test sets. To compute, we test each method on ImageNetV2 [34] (10K images) as knowns and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.789, + 0.313 + ], + "angle": 0, + "content": "
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .16.92 / .37.88 / .37.91 / .40.80 / .55.95 / .23.91 / .41.90 / .37.87 / .53.87 / .44
NINCO [2] (~5.8K).88 / .52.83 / .61.77 / .62.71 / .72.70 / .71.87 / .54.83 / .63.81 / .60.73 / .78.79 / .63
OpenImage-O [45] (~17.6K).96 / .20.87 / .48.79 / .49.90 / .38.70 / .64.94 / .27.88 / .49.86 / .43.82 / .65.84 / .48
Places [56] (10K).90 / .44.83 / .60.74 / .62.87 / .48.65 / .74.89 / .48.83 / .61.78 / .60.84 / .59.74 / .64
SUN [49] (10K).91 / .41.84 / .57.77 / .58.87 / .49.69 / .70.90 / .44.84 / .59.80 / .57.84 / .55.76 / .61
Textures [8] (~5.1K).94 / .27.88 / .47.85 / .42.77 / .79.81 / .49.92 / .32.87 / .49.88 / .39.75 / .82.87 / .39
21K-P Hard [44] (49K).77 / .68.73 / .73.68 / .73.68 / .76.62 / .82.76 / .72.74 / .75.70 / .74.67 / .82.67 / .79
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.34, + 0.39, + 0.357 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.796, + 0.454 + ], + "angle": 0, + "content": "In Tab. 3 and 4, we investigate the effects of varying the validation datasets (knowns and unknowns surrogate). We interchanged ImageNetV2 with ILSVRC2012 val for knowns and 21K-P Hard with 21K-P Easy for the unknowns surrogate. Unlike in Tab. 1, this involves validating on a larger number of knowns and unknowns, with an easier unknowns surrogate. Performance remained consistent regardless of the datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Also, we investigated the impact of other architectures as extractors and interchanged pre-trained models (ILSVRC2012-1K only w/ 224 crops) with Meta's ViT-H [14] and ConvNeXt-L [24]. Results showcasing PostMax's robustness are in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Lastly, we explored the impact of varying the application-dependent \\(\\alpha\\) from Eq. (2) and the normalization \\((L_{2})\\) within PostMax. Details and results are in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.583, + 0.355, + 0.599 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.675 + ], + "angle": 0, + "content": "This paper introduces the new evaluation metric Open-Set Accuracy (OSA) (2); we framed the evaluation of an open-set problem as predicting a threshold and then evaluating open-set accuracy at that threshold. This supports an engineer selecting an algorithm and a threshold to be applied in some Operational setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In real-world scenarios, unknown samples can be drawn from an unlimited variety, unlike known samples, which are assumed to be related to the training data. Thus, the evaluation must consider the computation of the predictive threshold on validation data and a separate evaluation on test data. The quality of the predicted threshold will depend on many factors with the two most influential being (1) how representative are the surrogate samples of the real unknowns, and (2) the relative proportion of knowns and unknowns. In Sec. 4.3 and our supplemental, we explore some variations of those factors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Maybe not surprisingly, if we train with \"harder\" unknowns, we generally see better performance. As shown in Tab. 1 and 2, the novel PostMax algorithm is better on unknowns across two networks (other networks in the supplemental)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "However, as shown in Sec. 4.3, when 21K-P Easy is used as our unknowns surrogate, PostMax does not exhibit an advantage on 21K-P Hard. In general, 21K-P Hard poses the most significant challenge for all algorithms. Notably, the sets close visual and semantic similarity to known samples suggests it represents a fine-grained categorization challenge rather than a typical open-set scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.788, + 0.315 + ], + "angle": 0, + "content": "One can adjust the importance of knowns vs. unknowns either by explicitly weighting them or implicitly through their relative percentages in the data. Although performance degrades when there is a significant disparity in their importance during threshold prediction stages, PostMax consistently outperforms and demonstrates greater stability across all thresholds. Refer to the supplemental for further details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Finally, to account for inherent random variations in predictions and tests, we developed training/validation/test splits, which are available on the GitHub repo. This protocol enables rigorous experimentation to assess the statistical significance of our evaluation. While the paper demonstrates statistically significant improvements across all results, some ablations in the supplemental do not." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.429, + 0.36, + 0.445 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.594 + ], + "angle": 0, + "content": "This paper contributes to OSR research by shifting focus from common metrics on small-scale experiments to more realistic scenarios and datasets, thereby fostering more practically relevant research. We revisited and revised the criteria for effective open-set metrics as suggested by Wang et al. [46], making them more applicable to real-world scenarios. Our proposed Operational Open-Set Accuracy (OOSA), which incorporates known and unknown samples, offers a comprehensive metric and facilitates the selection of the most effective method and operating threshold for system operators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.778 + ], + "angle": 0, + "content": "Our investigation of deep feature magnitudes of unknown samples from various ILSVRC2012 pre-trained networks yielded surprising findings. Contrary to smaller-scale studies [10, 31], we observed that these magnitudes are often larger for unknown than for known samples. This observation was consistent across multiple networks and types of unknown samples. Thus, we introduced our novel PostMax algorithm, refining normalized logit values from closed-set architectures into real probabilities of class inclusion using a Generalized Pareto Distribution (GPD). This distribution method offers advantages over previous extreme-value-based techniques, including negating the need to select a tail size and minimizing the influence of outliers. Also, it eliminates the need to define negative training samples as representatives of unknown classes, a common approach in much of the OSR literature that has not yielded a sufficient solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While a good closed-set classifier is an important place to start [44], our novel PostMax algorithm shows that one can do better than \"all you need\" and that for operational usage, one needs to predict a threshold to use. With this paper, we empower engineers to operationalize their usage of OSR." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing ImageNet out-of-distribution detection evaluation. In: ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.787, + 0.283 + ], + "angle": 0, + "content": "3. Bodesheim, P., Freytag, A., Rodner, E., Denzler, J.: Local novelty detection in multiclass recognition problems. In: Winter Conference on Applications of Computer Vision (WACV) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.787, + 0.323 + ], + "angle": 0, + "content": "4. Boult, T.E., Cruz, S., Dhamija, A.R., Günther, M., Henrydoss, J., Scheirer, W.J.: Learning and the unknown: Surveying steps toward open world recognition. In: AAAI Conference on Artificial Intelligence. vol. 33, pp. 9801-9807 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.787, + 0.351 + ], + "angle": 0, + "content": "5. Cevikalp, H., Uzun, B., Salk, Y., Saribas, H., Köpukli, O.: From anomaly detection to open set recognition: Bridging the gap. Pattern Recognition 138, 109385 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.391 + ], + "angle": 0, + "content": "6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 8065-8081 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "7. Chen, G., Qiao, L., Shi, Y., Peng, P., Li, J., Huang, T., Pu, S., Tian, Y.: Learning open set network with discriminative reciprocal points. In: European Conference on Computer Vision (ECCV). Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "8. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "0. Dhamija, A.R., Günther, M., Boult, T.: Reducing network agnostophobia. Advances in Neural Information Processing Systems (NeurIPS) 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "1. Ge, Z., Demyanov, S., Garnavi, R.: Generative OpenMax for multi-class open set classification. In: British Machine Vision Conference (BMVC) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "2. Gumbel, E.J.: Statistical Theory of Extreme Values and Some Practical Applications: A Series of Lectures, vol. 33. US Government Printing Office (1954)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.636 + ], + "angle": 0, + "content": "3. Günther, M., Dhamija, A.R., Boult, T.E.: Watchlist adaptation: Protecting the innocent. In: International Conference of the Biometrics Special Interest Group (BIOSIG) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.677 + ], + "angle": 0, + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "5. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (ICML). PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "6. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "7. van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The INaturalist species classification and detection dataset In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems (NeurIPS) 34, 677-689 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8710-8719 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.218 + ], + "angle": 0, + "content": "20. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: International Conference on Computer Vision (ICCV). pp. 813-822 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.787, + 0.247 + ], + "angle": 0, + "content": "21. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.786, + 0.275 + ], + "angle": 0, + "content": "22. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.277, + 0.787, + 0.317 + ], + "angle": 0, + "content": "23. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23946-23955 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.36 + ], + "angle": 0, + "content": "24. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.402 + ], + "angle": 0, + "content": "25. Lu, J., Xu, Y., Li, H., Cheng, Z., Niu, Y.: Pmal: Open set recognition via robust prototype mining. In: AAAI Conference on Artificial Intelligence. vol. 36:2, pp. 1872-1880 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "26. Lyu, Z., Gutierrez, N.B., Beksi, W.J.: Metamax: Improved open-set deep neural networks via weibull calibration. In: Winter Conference on Applications of Computer Vision Workshops (WACVW) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.488 + ], + "angle": 0, + "content": "27. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. In: European Conference on Computer Vision (ECCV). pp. 365-381. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.516 + ], + "angle": 0, + "content": "28. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.558 + ], + "angle": 0, + "content": "29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. In: NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.6 + ], + "angle": 0, + "content": "30. Palechor, A., Bhoumik, A., Günther, M.: Large-scale open-set classification protocols for imagenet. In: Winter Conference on Applications of Computer Vision (WACV). pp. 42-51. CVF/IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "31. Park, J., Chai, J.C.L., Yoon, J., Teoh, A.B.J.: Understanding the feature norm for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1557-1567 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "32. Park, J., Jung, Y.G., Teoh, A.B.J.: Nearest neighbor guidance for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1686-1695 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.713 + ], + "angle": 0, + "content": "33. Pickands III, J.: Statistical inference using extreme order statistics. The Annals of Statistics 3(1), 119 - 131 (1975)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "34. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do Imagenet classifiers generalize to Imagenet? In: International Conference on Machine Learning (ICML). pp. 5389-5400. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.784 + ], + "angle": 0, + "content": "35. Rudd, E.M., Jain, L.P., Scheirer, W.J., Boult, T.E.: The extreme value machine. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Operational Open-Set Recognition and PostMax Refinement" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "37. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. International Conference on Machine Learning (ICML) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.79, + 0.245 + ], + "angle": 0, + "content": "38. Scheirer, W.J., Jain, L.P., Boult, T.E.: Probability models for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 36(11), 2317-2324 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "39. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Towards open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 35(7) (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.288, + 0.788, + 0.328 + ], + "angle": 0, + "content": "40. Shu, L., Xu, H., Liu, B.: DOC: Deep open classification of text documents. In: Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "41. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems (NeurIPS) 34, 144-157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.788, + 0.412 + ], + "angle": 0, + "content": "42. Vareto, R.H., Linghu, Y., Boult, T.E., Schwartz, W.R., Günther, M.: Open-set face recognition with maximal entropy and Objectosphere loss. Image and Vision Computing (IMAVIS) 141 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.788, + 0.453 + ], + "angle": 0, + "content": "43. Vareto, R.H., Schwartz, W.R., Günther, M.: Toward open-set face recognition with neural ensemble, maximal entropy loss and feature augmentation. In: Conference on Graphics, Patterns and Images (SIBGRAPI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.788, + 0.495 + ], + "angle": 0, + "content": "44. Vaze, S., Han, K., Vedaldi, A., Zissermann, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.788, + 0.536 + ], + "angle": 0, + "content": "45. Wang, H., Li, Z., Feng, L., Zhang, W.: Out-of-distribution with virtual-logit matching. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4921-4930 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.788, + 0.577 + ], + "angle": 0, + "content": "46. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards a priori-oriented open-set recognition. Advances in Neural Information Processing Systems (NeurIPS) 35, 25033-25045 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.577, + 0.788, + 0.619 + ], + "angle": 0, + "content": "47. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sünderhauf, N.: Safe: Sensitivity-aware features for out-of-distribution object detection. In: International Conference on Computer Vision (ICCV). pp. 23565-23576 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.62, + 0.788, + 0.661 + ], + "angle": 0, + "content": "48. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16133-16142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.661, + 0.788, + 0.702 + ], + "angle": 0, + "content": "49. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3485-3492. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.702, + 0.788, + 0.744 + ], + "angle": 0, + "content": "50. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.744, + 0.788, + 0.785 + ], + "angle": 0, + "content": "51. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.788, + 0.84 + ], + "angle": 0, + "content": "52. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: OpenOOD: Benchmarking generalized out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS) 35, 32598-32611 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.79, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.345, + 0.127 + ], + "angle": 0, + "content": "Cruz et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "53. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "54. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: OpenOOD v1.5: Enhanced benchmark for out-of-distribution detection. In: NeurIPS Workshop on Distribution Shifts: New Frontiers with Foundation Models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "55. Zhang, X., Cheng, X., Zhang, D., Bonnington, P., Ge, Z.: Learning network architecture for open-set recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36(3), pp. 3362-3370 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.788, + 0.356 + ], + "angle": 0, + "content": "57. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_origin.pdf b/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..487b42a6444f75fcc716ce279a350c368feffac4 --- /dev/null +++ b/2024/Operational Open-Set Recognition and PostMax Refinement/e1f6e331-c70e-4745-aa6d-8d2c67ea3455_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:398f0703bd82370327a73ca57204ef32d0a6148be3fe7d93f4afc0f4c1e2fbd9 +size 1957971 diff --git a/2024/Operational Open-Set Recognition and PostMax Refinement/full.md b/2024/Operational Open-Set Recognition and PostMax Refinement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c7c7309b17befef0f2a418f53f901c560bf817dc --- /dev/null +++ b/2024/Operational Open-Set Recognition and PostMax Refinement/full.md @@ -0,0 +1,308 @@ +# Operational Open-Set Recognition and PostMax Refinement + +Steve Cruz1, Ryan Rabinowitz2, Manuel Günther3, and Terrance E. Boult2 + +1 University of Notre Dame +2 University of Colorado Colorado Springs + +3 University of Zurich + +stevecruz@nd.edu {rrabinow, tboult}@uccs.edu guenther@ifi.uzh.ch + +Abstract. Open-Set Recognition (OSR) is a problem with mainly practical applications. However, recent evaluations have largely focused on small-scale data and tuning thresholds over the test set, which disregard the real-world operational needs of parameter selection. Thus, we revisit the original goals of OSR and propose a new evaluation metric, Operational Open-Set Accuracy (OOSA), which requires predicting an operationally relevant threshold from a validation set with known and a surrogate set with unknown samples, and then applying this threshold during testing. With this new measure in mind, we develop a large-scale evaluation protocol suited for operational scenarios. Additionally, we introduce the novel PostMax algorithm that performs post-processing refinement of the logit of the maximal class. This refinement involves normalizing logits by deep feature magnitudes and utilizing an extreme-value-based generalized Pareto distribution to map them into proper probabilities. We evaluate multiple pre-trained deep networks, including leading transformer and convolution-based architectures, on different selections of large-scale surrogate and test sets. Our experiments demonstrate that PostMax advances the state of the art in open-set recognition, showing statistically significant improvements in our novel OOSA metric as well as in previously used metrics such as AUROC, FPR95, and others. + +# 1 Introduction + +The original Open-Set Recognition (OSR) formulation [1,38,39] gained attention due to the flaws in existing recognition systems when handling inputs from outside the training classes. These systems lack control over inputs, and any unknown sample that is not rejected leads to a misclassification, thereby reducing accuracy. Therefore, early open-set work [1,3,35] introduced protocols tailored to emulate real-world systems, allowing for a proper evaluation of open-set performance. + +For a decade, researchers explored techniques that enhanced OSR performance [1,5,7,10,11,20,25,27,28,35,39,55,57]. While contributing novel ideas, recent works have diverged from the original goals of OSR. Instead of leveraging large-scale data as in early research on open-set deep networks [1], recent OSR evaluations have + +![](images/ee72ea10b138a49268b1b1ed64f751f6311224ce748dfa177995a5f5af5983d8.jpg) +(a) Train + +![](images/ffa717af6ff75c61ec29fa885705f87f51151c0316cbc16b292f1bd2501eedfb.jpg) +(b) Val +Fig. 1: POSTMAX PIPELINE. During train (a), post-processed maximum logits of correctly classified train samples are normalized by feature magnitude $\ell$ ; the set is then used in a Maximum Likelihood Estimation (MLE) to obtain Generalized Pareto Distribution (GPD) [33] with parameters $\mu, \sigma$ , and $\xi$ . During validation (b), we use a validation and a surrogate set where the cumulative distribution of the GPD $(\Psi)$ gives the probability of the normalized max logit being from the train distribution and predict an operational threshold. During operation (c), we deploy each method's threshold and apply it to determine OSR and compute Operational Open Set Accuracy (OOSA). For evaluation we compare against Maximum Softmax Probability (MSP) [16,44], Maximum Logit (MaxLogit) [15,44], Nearest Neighbor Guidance (NNGuide) [32], and SCALE [50]. Operational Open-Set Accuracy (OOSA) performance is $(\star)$ and $(\diamond)$ is the optimal Open-Set Accuracy (OSA) for an oracle that knows the final test set (shown to highlight differences from OOSA). AUROC scores are shown for comparison. + +![](images/a6e86241cf88b5278f17d5bf7d74eb6e6f53a9f05afdaa8ee42ad43583b6bec9.jpg) +(c) Test (OOSA & AUROC) + +focused on small-scale datasets with few classes, with only a few notable exceptions [30,35,44]. Moreover, the prevalent reliance on metrics such as the Area Under the Receiver Operating Characteristics (AUROC) curve for performance evaluation has led to inopportune assessments since such metrics do not offer a realistic perspective on OSR. Fig. 1 illustrates an example where AUROC scores exhibit one ranking, yet in an operational setting, there is a slight difference. Inherently, common measures overlook the real-world operational needs of threshold selection. Some may argue for their use with a dedicated validation set; however, it is not clear how one would extract a threshold from these metrics and deploy it at test time. This is problematic as methods have largely ignored a fundamental practice of machine learning systems - having separate training, validation, and test sets to ensure that models generalize, avoid overfitting, and provide unbiased evaluation and reliable performance prediction on new data. + +Initial aspirations behind OSR aimed to enhance real systems [4]. Imagine an engineer tasked with implementing an OSR algorithm into their system. As they analyze results from various algorithms through tables and plots to inform decisions, a formidable challenge emerges. While AUROC and similar metrics assess a model's discriminative capability across all possible classification thresholds, they do not guide the engineer toward an optimal threshold for decision-making. There is a pressing need for a metric that enables confident and statistically rigorous evaluations with anticipated data, while also demonstrating robustness against varying numbers of unknowns. When deploying an open-set system, engineers must, at a minimum, consider two factors: (1) how an + +approach scales with operational data, and (2) how to choose an operating point that balances their desired known classification accuracy and unknown rejection. These considerations underscore the importance of achieving overall accuracy with a predicted operational threshold. Therefore, we introduce a new metric called Operational Open-Set Accuracy (OOSA), illustrated in Fig. 1, which facilitates algorithm and operational threshold selection. To ensure real-world applicability, we implement a large-scale evaluation protocol with proper training/validation/test splits. We determine a threshold on the validation set and deploy it during testing on unseen unknowns. Additionally, for completeness, fairness, and robustness, we evaluate state-of-the-art algorithms on ImageNet-scale open-set recognition using common metrics (AUROC, FPR95, AUOSCR [6], and OpenAUC [46]), many of which are detailed in the supplemental. + +Before incorporating an operational threshold, it is natural to ask about the distribution of scores above this threshold. This naturally leads to a Peak-over-Threshold (POT) formulation of the distribution, best modeled via Extreme Value Theory (EVT). The Pickands-Balkema-De Haan theorem of EVT [33] yields a Generalized Pareto Distribution (GPD) over normalized scores, as depicted in Fig. 1. This GPD-based score transformation provides theoretically grounded probabilities rather than raw, network-dependent ad-hoc confidence scores (see supplemental for details). While GPD can be applied directly to Softmax confidences or logit values, our approach goes further. We introduce a novel algorithm, PostMax, which applies GPD to post-processed normalized maximum logits. Here, normalization by the deep feature magnitude enhances the separation of known and unknown classes. Prior observations and theoretical explanations [10,13,31,44] suggest that deep feature magnitudes extracted from inputs that the network was not trained on are generally smaller than those for known samples. However, our findings show the exact opposite, i.e., that modern networks trained on the large-scale ILSVRC2012 dataset generally exhibit larger deep feature magnitudes for unknown samples (details in supplemental). Therefore, applying GPD to logit values divided by the deep feature magnitude can enhance OSR beyond the strong baselines of using raw Softmax confidence [16], maximum logit values [15], overconfidence reduction with features and logits [32], or scaled logit values [50]. The contributions of this paper are as follows: + +- We design a novel evaluation metric, Operational Open-Set Accuracy (OOSA), that emphasizes real-world usage by predicting an operational threshold on a validation set (knowns and unknowns surrogate), which is then deployed during testing on unseen data. +- We introduce a novel algorithm, PostMax, that post-processes maximum logits by normalizing them by deep feature magnitude and then applies GPD to provide probabilities. Our code is publicly available.4 +- We develop new large-scale evaluation protocols suited for assessing algorithms in operational scenarios. +- We showcase that PostMax advances the state of the art in large-scale open-set recognition with statistical significance on OOSA and prior metrics. + +# 2 Related Work + +Improving OSR, distinct from OOD detection, anomaly detection, or novel category discovery, can be approached in two ways: providing better features through improved network training, or through post-processing where a pretrained network is trained for closed-set tasks and then adapted for OSR. Our work is rooted in the post-processing approaches. + +# 2.1 Related Problems + +A common question regarding open-set is, What about a two-stage system with OOD followed by classification, does that not solve OSR? This hypothesis is rejected with an example from one of our operational systems where we have to recognize objects in novel contexts, e.g., in snowy or foggy conditions. Such conditions are OOD with respect to training samples, but they should not be rejected from an open-set point of view. Out-Of-Distribution (OOD) detection, though distinct from OSR, has been more consistent in its use of large-scale experimentation as were used in early OSR work [15,16,32,45,47,50,52,54]. Even when OOD is restricted to out-of-class, the two-stage system needs to be evaluated in that context, including a process for selecting an OOD threshold and then evaluating the resulting classification performance. The OpenOOD Benchmark [52,54] has implemented various OSR and OOD techniques such as OpenGan [20], MOS [19], ReAct [41], ViM [45], GEN [23], NNGuide [32], and SCALE [50]. While we caution the use of OOD algorithms in OSR evaluations/settings, based on the OpenOOD Benchmark leaderboard, we compare with the state-of-the-art, NNGuide & SCALE. + +# 2.2 Closed-Set Classifiers + +Recently, Vaze et al. [44] argued that closed-set classifiers are sufficient. This aligns with Hendrycks et al. [15, 16], who demonstrated thresholding on Softmax confidences or, especially, on logits provide unreasonably good baselines. Our work shows that such approaches/networks can be improved with normalization. + +We evaluate several leading pre-trained architectures trained on ILSVRC2012-1K with no additional data. Particularly, Meta's Vision Transformer Hiera-H [37], which strips non-essential components making it faster and more accurate during training/inference. Also, we use their Masked Autoencoders (MAE) model, ViT-H [14], which masks random patches of an image and reconstructs the missing pixels. To show generalization, we also utilize CNNs, including Meta's ConvNeXtV2-H [48] and ConvNeXt-L [24] which modernize a standard ResNet. Other networks are found in the supplemental material. + +# 2.3 Post-processing Approaches + +Besides thresholding softmax scores or logits [15, 16, 44], OSR methods take features and use Weibull-calibrated Support Vector Machines (W-SVM) [38] + +or Extreme Value Machines (EVM) [35] to estimate a probability of unknown. OpenMax [1] was the first method to add an artificial probability of unknown to closed-set networks by computing a logit score for the unknown class based on deep feature similarities to features of known classes. Other approaches include a small adaptor network [13, 42] that adds open-set capabilities [10] to features extracted from closed-set networks. + +# 2.4 Learning-based Approaches + +Most methods use negative aka. known unknown samples in training, which represent samples of no interest to the classifier and should be classified as unknown, with a few exemptions such as replacing the final Softmax classifier with a set of binary classifiers [40, 43]. While some use real negative samples [10, 13, 30], the majority of research tries to artificially create them. For example, combinations of knowns [57], added noise to knowns [47], or Generative Adversarial Networks (GANs) [11, 20]. In our approach, we avoid the use of any negative samples. + +# 2.5 Evaluation Techniques + +Evaluating OSR correctly in a real-world setting is difficult and current measures do not satisfy operational requirements. For example, the widely used AUROC metric [5,7,27,28,44,51] only looks at the binary decision of known vs. unknown, but ignores the task of assigning the correct class for known samples. Such metrics are often combined with closed-set classification accuracy in order to evaluate OSR. These evaluations are reasonable, but provide no information under realistic scenarios, i.e., an operational threshold classifying samples as known or unknown. + +Another metric often employed is the macro-averaged F1 score [1,27,35,53], which treats unknown as a separate class. This measure has many counter-intuitive properties that make it difficult to interpret the results. For example, a different threshold is required for each class so a sample could be classified as several known classes and unknown (at the same time). Thus, incorrect classification of unknown samples can easily be overlooked. + +Recently, Dhamija et al. [10] introduced the Open-Set Classification Rate (OSCR) curve. This curve handles knowns and unknowns separately, evaluating the Correct Classification Rate (CCR) at various thresholds corresponding to specific False Positive Rates (FPR). While valuable for high-risk applications like open-set face recognition, its practicality in general OSR tasks is questionable. In scenarios where a threshold is chosen based on a specific FPR, the curve utilizes unknowns from the test set, raising doubts about its applicability to unseen classes. Moreover, the curve faces the challenge of not providing a single value for comparing different methods. Additionally, the proposed metrics, Area Under the OSCR (AUOSCR) [6] and Open Area Under the ROC Curve (OpenAUC) [46], do not provide an intuition on how to select an operating threshold. + +![](images/87d6f851dd352aa9e8462bda1bcfc8cfd879633a0dfc6b83807698c1099d5913.jpg) +Fig. 2: MOTIVATION FOR NORMALIZATION. We illustrate why PostMax divides by the feature magnitude norm before applying GPD. The comparison above shows squared deep feature values extracted from a pre-trained ILSVRC2012-1K Transformer (ViT-H [14]). We sorted the deep features of 20 randomly sampled images from class 990 based on increasing mean square magnitude. Next, we applied this sorted order to the features of 20 images misclassified as class 990 from each unknown dataset. Unknowns exhibit high feature values (bright colors) in regions where known sample activations are low. These large, uncorrelated responses accumulate to give it a high score for the class but also result in unknown dataset features having, on average, larger $L_{2}$ norms than those of the correct class, depicted in the magnitude histogram on the right. Therefore, dividing by magnitude enhances separation. + +![](images/3165a2e2b460cea80cc92ed43a899d77a49b290f0b66439493bc949ca32b1eed.jpg) + +# 3 Approach + +Several approaches advocate selecting the maximum class and then applying thresholding to certain scores, e.g., Softmax [16] or logits [15], to address OSR. While such thresholding proves effective in some recognition tasks [44], its relevance often extends only to establishing a ranked order of classes for individual samples. We focus on leveraging information from the features themselves and the formal distribution of these maxima. The result is PostMax, an effective approach with theoretically grounded probabilities to address OSR through (1) normalization and (2) score distribution. + +# 3.1 Normalization & GPD + +From Fig. 2, unknown samples activate high dimensions in the feature space, which goes against our intuition that deep networks would only learn features necessary to classify known classes and disregard those belonging to other uninteresting ones. We attribute these findings to high logit values and Softmax scores, where positions with high feature magnitudes align with larger weights of certain classes, even when an input exhibits a large feature magnitude and the high-scoring class has a small or negative weight. The observed effect may depend on the network's loss function and the training dataset; in small-scale datasets such as MNIST or CIFAR, which only differentiate between 10 classes, the opposite has been observed [10, 18, 31, 44]. Further details regarding the normalizations in these works and how they differ from our findings are provided in the supplemental. + +As we classify samples based on their "maximum score" threshold, a key question arises: Does this score belong to the training distribution of maximum scores? Statistical Extreme Value Theory (EVT) provides a grounded approach to + +Algorithm 1 POSTMAX FITTING. This implements the probability distribution estimation of the proposed PostMax algorithm. +```txt +Require: $D_{\mathrm{train}}$ $\operatorname {FV}(x)$ $\operatorname {L}(x)$ $M\gets \{\}$ +for each $(x,t)\in D_{\mathrm{train}}$ do if arg max $\mathrm{L}(x) = t$ then $\ell \leftarrow \frac{\max\mathrm{L}(x)}{\|\mathrm{FV}(x)\|}$ $M\gets M\cup \{\ell \}$ end if +end for + $\mu ,\sigma ,\xi \gets$ scipy.stats.genpareto.fit(M) +``` + +answering this question. Unlike ad-hoc Softmax confidences, transforming logits into real probabilities allows probabilities to be more interpretable compared to Softmax, probabilities do not need to sum up to one (images may contain multiple objects), and EVT models probabilities on are based on thousands/millions of samples rather than the current sample. Given that thresholding maximum logits or Softmax scores serves as a reasonable baseline [15,32,44], applying EVT to this problem is both theoretically justified and intuitive. Note, prior OSR research [1,26,35,38] utilizes the Fisher-Tippett-Gnedenko EVT [12], resulting in a Weibull distribution. However, practical effectiveness was limited due to challenges in parameter selection affecting EVT modeling. In contrast, we employ the Generalized Pareto Distribution (GPD) derived from the Peak-over-threshold (POT) approach [33], which models extreme values above a threshold. + +# 3.2 PostMax + +Using these intuitions and observations, we define our Postnormalization of Maxima (PostMax) algorithm. We aim to find a general distribution $\varPsi$ of logit values that is valid for all of our known classes $c\in \{1,\ldots ,C\}$ . As observed in Fig. 2, modeling raw logit values might not be fruitful since unknown samples have generally higher activation of deep features, which often leads to high logit values. For a sample, we utilize normalized logit values by dividing the original logit values by their deep feature magnitude via (1). + +Based on POT EVT [33], we understand that if we consider all maximum values above a threshold (such as the smallest correct logit observed during training), the resulting distribution follows a Generalized Pareto Distribution (GPD). Let $D_{\mathrm{train}} = \{(x_i, t_i)\}$ be the collection of all samples $x_i$ with their respective ground-truth target label $t_i$ in the training set. Let $\mathrm{FV}(x)$ be the function returning the feature vector of a sample $x$ . Let $\mathrm{L}(x)$ be the function returning the logit vector of a sample. We collect the normalized logits from the target class for all correctly classified training samples, which are then used to model a Generalized Pareto Distribution $\varPsi_{\mu, \sigma, \xi}$ . Details can be found in Alg. 1. + +At inference for test sample $x$ we use $\mathrm{FV}(x)$ and $\mathrm{L}(x)$ to extract the deep features and logits of $x$ and then for class $c$ assign probability $p_c(x)$ using the + +normalized logit value: + +$$ +\ell_ {c} = \frac {\operatorname {L} (x) [ c ]}{\| \operatorname {F V} (x) \|} \quad p _ {c} (x) = \Psi_ {\mu , \sigma , \xi} (\ell_ {c}) \tag {1} +$$ + +where $\varPsi_{\mu,\sigma,\xi}(\ell)$ is the cumulative distribution of the GPD with location $\mu$ , scale $\sigma$ and shape $\xi$ computed via Alg. 1. If we just want the top class (as all experiments), we compute the class with the max logit and apply the above to that logit. + +Note that we do not utilize any unknowns or negatives in this process, nor do we explicitly model a probability of unknowns. Instead, we threshold based on the cumulative GPD probability of known classes, computing the maximum value of normalized logits. + +# 3.3 Operational Open-Set Accuracy + +By designing the OpenAUC metric, Wang et al. [46] proposed four different conditions that a good open-set evaluation metric should fulfill, which we rephrase: + +P1 The metric needs to check that known samples are classified correctly with high probability. +P2 The metric needs to evaluate if unknown samples are assigned to known classes with low probability. + +P3 The metric should be insensitive to a score threshold. + +P4 The metric should be a single number. + +We totally agree with P1, P2 and P4. However, we reject P3 because it means that the metric cannot be used to select a classifier that works well in a specific operational setting, which requires obtaining an operational threshold. In our view, a useful open-set algorithm must include that step. Therefore, we reformulate condition P3' to be operationally relevant: + +P3' The metric should indicate an operational threshold that is optimal under specified circumstances and can be applied to unseen data and unseen unknown classes. + +These four conditions (P1, P2, P3', P4) exclude all existing open-set evaluation metrics; we need a new metric to satisfy them. As discussed in Sec. 2.5, metrics such as AUROC, AUOSCR, and OpenAUC simply average over all possible operational thresholds, while F1 and normalized accuracy fail to use a single threshold [46]. On the other hand, the Open-Set Classification Rate (OSCR) curve [10] does not provide a single value for comparison, which makes it hard to compare different algorithms. Therefore, we introduce Operational Open-Set Accuracy (OOSA), which is inspired by OSCR but fulfills all four conditions. + +From an engineer's view, the real question is how well a specific method can perform under the presence of both known and unknown samples in their data. Most applications of open-set classification are not security-sensitive, so there is no need to have very low False Positive Rates (FPR) as typically evaluated in the OSCR [10,30]. + +Usually, an engineer has a rough estimate of how many known and unknown samples the system will observe - so they are rather interested in the total accuracy + +the system will achieve and which kind of operational threshold will lead to this performance. To provide both, we exploit the FPR and CCR calculations used in OSCR [10] to define the Unknown Rejection Rate (URR) and the Open-Set Accuracy (OSA), both of which rely on the operational threshold $\theta$ and only make use of the probabilities of known classes $P_{c}(x)$ : + +$$ +\operatorname {C C R} (\theta) = \frac {\left| \left\{x \mid x \in D _ {c} \wedge \arg \max _ {c} P _ {c} (x) = \hat {c} \wedge P _ {c} (x) \geq \theta \right\} \right|}{\left| D _ {c} \right|} +$$ + +$$ +\operatorname {U R R} (\theta) = \frac {\left| \left\{x \mid x \in D _ {a} \wedge \max _ {c} P _ {c} (x) < \theta \right\} \right|}{\left| D _ {a} \right|} \tag {2} +$$ + +$$ +\mathrm {O S A} (\theta) = \alpha \cdot \mathrm {C C R} (\theta) + (1 - \alpha) \cdot \mathrm {U R R} (\theta) +$$ + +where $D_{c}$ is the collection of all known test samples with class labels $\hat{c}$ , while $D_{a}$ collects all unknown test samples [10]. Please note that $\mathrm{URR}(\theta) = 1 - \mathrm{FPR}(\theta)$ provides a positive view on the unknown samples, i.e., it calculates how many of them are correctly rejected under threshold $\theta$ . OSA includes a weighted average over CCR and URR to provide a holistic view of the expected total accuracy, where the application-dependent $\alpha$ can be used to provide a stronger focus on correctly classifying known or unknown samples. When both types of samples should be treated similarly, $\alpha = \frac{|D_{c}|}{|D_{c}| + |D_{a}|}$ can be selected, which we do in most of our experiments. In that case, OSA can be thought of as evaluating a $C$ class classifier using $C + 1$ class accuracy where the unknown class $C + 1$ is classified when the maximal probability does not exceed our operational threshold $\theta$ . + +When plotting OSA over URR, by varying the operational threshold, we arrive at the plots seen in Fig. 1. This plot is not monotonic, but rather provides a peak where the OSA is maximized. The maximum value can be used to select a threshold $\theta^{*}$ on a given set of scores $S$ effectively: + +$$ +\theta^ {*} = \arg \max \mathrm {O S A} _ {S} (\theta) +$$ + +$$ +\theta \in S \tag {3} +$$ + +$$ +S = \left\{P _ {\hat {c}} (x) \mid x \in D _ {c} \right\} \cup \left\{\max _ {c} P _ {c} (x) \mid x \in D _ {a} \right\} +$$ + +Here, $S$ is the collection of all maximum scores for known and unknown samples. When applying this calculation on the test set $S_{test}$ , we will find the optimal threshold $\theta_{test}^{*}$ , indicated by $\diamond$ in Fig. 1. However, since test set samples are unknown in operational settings, we make use of the validation set $D_{val}$ of known samples as $D_{c}$ and an additional surrogate set $D_{sur}$ of unknown samples to replace $D_{a}$ in (3) for obtaining the operational threshold $\theta_{sur}^{*}$ , which we will then apply to the test set: $\mathrm{OSA}_{\text{test}}(\theta_{sur}^{*})$ , which is indicated by $\star$ in Fig. 1. How to optimally select this surrogate set is discussed below. + +# 4 Experiments + +Although a few OSR methods have explored large-scale datasets [1,30,35,44], many recent techniques [5-7,10,11,20,25,27,28,55,57] still lack comprehensive large-scale evaluations. The performance on small-scale data such as MNIST [22], + +Table 1: OPERATIONAL OPEN-SET ACCURACY (OOSA). The mean OOSA $(\uparrow)$ of all methods. To compute, we validate methods on ImageNetV2 [34] (10K images) as knowns and $20\%$ of our surrogate 21K-P Hard [44] (9.8K images) as unknowns and predict an operational threshold. Then, we deploy each method's threshold and test on five different ILSVRC2012 val [36] splits (each 10K images) and specified unknowns. Standard deviation is omitted as it is $< 0.005$ for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012 - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold. + +
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8250.8150.7800.7820.7270.8120.7960.7860.7130.772
NINCO [2] (~5.8K)0.7400.7390.7070.6250.6670.7320.7230.7180.5970.705
OpenImage-O [45] (~17.6K)0.8640.8140.7450.7960.6630.8490.8200.7920.7450.767
Places [56] (10K)0.7850.7540.6910.7450.6250.7790.7490.7100.6930.676
SUN [49] (10K)0.7950.7580.7090.7490.6510.7830.7510.7210.6930.690
Textures [8] (~5.1K)0.7640.7560.7360.6600.7150.7440.7360.7400.6010.740
21K-P Easy [44] (50K)0.8180.7400.6620.7440.5680.8130.7630.6910.7830.633
+ +CIFAR [21], or SVHN [29] does not reflect operational scenarios and lacks generalizability due to small image crops and a limited number of classes [30]. + +An open-set protocol is essential and necessitates a large-scale setting for the proper evaluation of real-world scenarios [30, 44]. Additionally, it is paramount to employ proper machine-learning methodology, including a validation set (and a surrogate set) for operational threshold selection. Thus, we evaluate the robustness and generalization of PostMax on multiple large-scale datasets and splits with realistic and high-quality images. We also leverage the recently introduced large-scale ImageNet open-set splits [44]. Lastly, we conduct ablations to understand the performance impact of our approach. + +# 4.1 Details + +Architectures. We use state-of-the-art pre-trained models on ILSVRC2012 [36] with standard $224 \times 224$ image crops - Meta's Hiera-H [37] and ConvNeXtV2-H [48]. Meta fine-tuned these models with no additional/external data. Note, we do not perform additional training or fine-tuning. + +Datasets. For knowns, we utilize the large-scale ImageNet [9] subset ILSVRC2012 [36]. This dataset provides an advantage due to the extensive variety of available pre-trained architectures, enabling us to compare with more diverse deep networks. Our PostMax model is trained on the ILSVRC2012 train split, which contains $\sim 1.28M$ images and $1K$ classes (PostMax training requires no additional negative/unknown data). For validation, we employ ImageNetV2 [34], which contains $10K$ images. For testing, we split ILSVRC2012 val into five equal splits of $10K$ images each. + +For unknowns, we use iNaturalist [17], NINCO [2], OpenImage-O [45], Places [56], SUN [49], Textures [8], and ImageNet-21K-P Open-Set splits [44]. The vision + +Table 2: COMMON METRICS. The mean AUROC $(\uparrow)$ and FPR95 $(\downarrow)$ of all methods. To compute, we test each method on five different ILSVRC2012 val [36] splits (each $10K$ images) as knowns and specified unknowns. Standard deviation is omitted as it is $< 0.01$ for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012-1K - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold. + +
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .21.92 / .37.88 / .36.90 / .50.81 / .51.95 / .26.91 / .42.90 / .37.86 / .58.88 / .44
NINCO [2] (~ 5.8K).87 / .58.83 / .62.79 / .61.69 / .78.73 / .67.86 / .56.83 / .64.82 / .60.71 / .81.80 / .62
OpenImage-O [45] (~ 17.6K).95 / .25.88 / .49.80 / .49.88 / .45.72 / .61.94 / .29.88 / .49.87 / .44.81 / .70.85 / .47
Places [56] (10K).89 / .49.83 / .60.75 / .61.85 / .55.67 / .71.89 / .50.84 / .62.79 / .60.82 / .64.76 / .64
SUN [49] (10K).90 / .46.84 / .57.78 / .57.85 / .56.71 / .66.89 / .47.84 / .59.81 / .57.83 / .60.77 / .61
Textures [8] (~ 5.1K).93 / .32.88 / .47.86 / .42.74 / .86.83 / .46.91 / .34.87 / .49.88 / .40.73 / .85.88 / .38
21K-P Easy [44] (50K).86 / .50.79 / .64.72 / .64.80 / .68.65 / .73.85 / .51.79 / .65.75 / .64.78 / .72.72 / .68
+ +community [15, 16, 32, 44, 45, 50, 52, 54] has utilized these datasets as they present unique challenges not seen in small-scale settings. For validation, to predict an operational threshold, we utilize $20\%$ of ImageNet-21K-P Hard (9.8K images) as our surrogate unknowns. For testing, NINCO, OpenImage-O, Textures, and ImageNet-21K-P Easy are entirely used, while others sample $10K$ images from a manual selection of classes. Further descriptions of all datasets and splits are provided in the supplemental. + +Evaluation Metrics. To align with existing literature, we utilized common metrics in Tab. 2, namely, Area Under the Receiver Operating Curve (AUROC) and False Positive Rate at a fixed True Positive Rate of $95\%$ (FPR95). Additional metrics like Area Under Open-Set Classification Rate (AUOSCR) [6] and Open Area Under ROC Curve (OpenAUC) [46] are provided in the supplemental. Alongside these, we introduce Operational Open-Set Accuracy (OOSA) in Tab. 1, where we predict an operational threshold on a validation and a surrogate set and deploy it at test time. This measure, described in detail in Sec. 3.3, serves as a more realistic alternative to existing measures. When evaluating OSR methods in real-world scenarios, other measures are insufficient as they do not provide an operating point for deployment and are usually only evaluated with a test set. + +# 4.2 Results + +Our main results are presented in Tab. 1 and 2, where we report performance on each unknown dataset mentioned in Sec. 4.1. We compare our approach with two commonly seen methods, Maximum Softmax Probability (MSP) [16,44] and Maximum Logit (MaxLogit) [15,44], along with the recently introduced Nearest Neighbor Guidance (NNGuide) [32] and SCALE [50]. We omit comparisons with older techniques such as OpenMax [1] and EVM [35] as they are much worse. According to the OpenOOD [52,54] Benchmark leaderboard for ImageNet-1K, + +Table 3: DIFFERENT VALIDATION - KNOWS & UNKNOWN'S SURROGATE. The Operational Open-Set Accuracy (OOSA) $(\uparrow)$ of all methods on a different validation set. To compute, we validate methods on ILSVRC2012 val [36] (50K images) as knowns and 21K-P Easy [44] (50K images) as unknowns surrogate and predict an operational threshold. Then, we deploy each method's threshold and test on ImageNetV2 [34] (10K images) and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold. + +
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8630.8080.7800.8080.7160.8420.8060.7940.7590.763
NINCO [2] (~5.8K)0.7330.7300.7040.6550.6540.7590.7330.7260.6430.695
OpenImage-O [45] (~17.6K)0.8850.8100.7360.8090.6490.8650.8210.7860.7460.757
Places [56] (10K)0.8010.7480.6840.7660.6110.7950.7530.7080.7310.664
SUN [49] (10K)0.8120.7520.7020.7710.6350.8020.7570.7220.7380.679
Textures [8] (~5.1K)0.8130.7460.7390.6910.7060.7840.7500.7570.6570.732
21K-P Hard [44] (49K)0.6340.6730.5830.5350.5150.6470.6720.5930.5590.566
+ +NNGuide stands out as the leader in the OOD space, closely followed by SCALE. We caution against relying solely on OOD methods and metrics, as they may not accurately reflect OSR performance. Note, some OSA and ROC curves are found in the supplemental. + +When utilizing both architectures (Hiera-H & ConvNeXtV2-H) as feature extractors, PostMax outperforms all other methods in every measure (OOSA, AUROC, FPR95) on each unknown dataset. PostMax demonstrates superior performance, emphasizing robustness and generalizability. Effectively, PostMax achieves an excellent trade-off between classification accuracy and unknown rejection. Overall, the results suggest that a system operator can seamlessly integrate PostMax, enhance the performance beyond good closed-set classifiers, and confidently select the operational threshold that best suits its operation. + +Statistical Testing. To underscore the significance of performance differences between algorithms in our experiments, we prioritized the assessment of statistical significance. As a result, we performed paired, two-tailed t-tests, with Bonferroni corrections, to compare PostMax with the four methods (MSP, MaxLogit, NNGuide, and SCALE) using both architectures (Hiera-H and ConvNeXtV2-H) across 5 fold runs for each metric (OOSA, AUROC, and FPR95). On Hiera-H, PostMax (1) OOSA scores improvements range from $2.5\%$ to $12.4\%$ with corresponding P-values all less than $1.0\mathrm{E}^{-7}$ , (2) AUROC scores improvements range from $5.7\%$ to $17.9\%$ with corresponding P-values all less than $1.0\mathrm{E}^{-8}$ , and (3) FPR95 scores improvements range from $12.7\%$ to $22.5\%$ with corresponding P-values all less than $1.0\mathrm{E}^{-8}$ . Similarly, on ConvNeXtV2-H, PostMax (1) OOSA scores improvements range from $2.3\%$ to $10.4\%$ , (2) AUROC scores improvements range from $4.7\%$ to $10.6\%$ , and (3) FPR95 scores improvements range from $9.6\%$ to $28.1\%$ , where all corresponding P-values are less than $1.0\mathrm{E}^{-8}$ for all tests on all three metrics. In summary, the t-tests conducted for PostMax, considering the combined scores of all metrics across all datasets and architectures, revealed very statistically significant differences compared to every other method. + +Table 4: DIFFERENT TESTING - PRIOR METRICS. The AUROC (↑) and FPR95 (↓) of all methods on different test sets. To compute, we test each method on ImageNetV2 [34] (10K images) as knowns and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold. + +
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .16.92 / .37.88 / .37.91 / .40.80 / .55.95 / .23.91 / .41.90 / .37.87 / .53.87 / .44
NINCO [2] (~5.8K).88 / .52.83 / .61.77 / .62.71 / .72.70 / .71.87 / .54.83 / .63.81 / .60.73 / .78.79 / .63
OpenImage-O [45] (~17.6K).96 / .20.87 / .48.79 / .49.90 / .38.70 / .64.94 / .27.88 / .49.86 / .43.82 / .65.84 / .48
Places [56] (10K).90 / .44.83 / .60.74 / .62.87 / .48.65 / .74.89 / .48.83 / .61.78 / .60.84 / .59.74 / .64
SUN [49] (10K).91 / .41.84 / .57.77 / .58.87 / .49.69 / .70.90 / .44.84 / .59.80 / .57.84 / .55.76 / .61
Textures [8] (~5.1K).94 / .27.88 / .47.85 / .42.77 / .79.81 / .49.92 / .32.87 / .49.88 / .39.75 / .82.87 / .39
21K-P Hard [44] (49K).77 / .68.73 / .73.68 / .73.68 / .76.62 / .82.76 / .72.74 / .75.70 / .74.67 / .82.67 / .79
+ +# 4.3 Ablation Study + +In Tab. 3 and 4, we investigate the effects of varying the validation datasets (knowns and unknowns surrogate). We interchanged ImageNetV2 with ILSVRC2012 val for knowns and 21K-P Hard with 21K-P Easy for the unknowns surrogate. Unlike in Tab. 1, this involves validating on a larger number of knowns and unknowns, with an easier unknowns surrogate. Performance remained consistent regardless of the datasets. + +Also, we investigated the impact of other architectures as extractors and interchanged pre-trained models (ILSVRC2012-1K only w/ 224 crops) with Meta's ViT-H [14] and ConvNeXt-L [24]. Results showcasing PostMax's robustness are in the supplemental. + +Lastly, we explored the impact of varying the application-dependent $\alpha$ from Eq. (2) and the normalization $(L_{2})$ within PostMax. Details and results are in the supplemental. + +# 5 Discussion + +This paper introduces the new evaluation metric Open-Set Accuracy (OSA) (2); we framed the evaluation of an open-set problem as predicting a threshold and then evaluating open-set accuracy at that threshold. This supports an engineer selecting an algorithm and a threshold to be applied in some Operational setting. + +In real-world scenarios, unknown samples can be drawn from an unlimited variety, unlike known samples, which are assumed to be related to the training data. Thus, the evaluation must consider the computation of the predictive threshold on validation data and a separate evaluation on test data. The quality of the predicted threshold will depend on many factors with the two most influential being (1) how representative are the surrogate samples of the real unknowns, and (2) the relative proportion of knowns and unknowns. In Sec. 4.3 and our supplemental, we explore some variations of those factors. + +Maybe not surprisingly, if we train with "harder" unknowns, we generally see better performance. As shown in Tab. 1 and 2, the novel PostMax algorithm is better on unknowns across two networks (other networks in the supplemental). + +However, as shown in Sec. 4.3, when 21K-P Easy is used as our unknowns surrogate, PostMax does not exhibit an advantage on 21K-P Hard. In general, 21K-P Hard poses the most significant challenge for all algorithms. Notably, the sets close visual and semantic similarity to known samples suggests it represents a fine-grained categorization challenge rather than a typical open-set scenario. + +One can adjust the importance of knowns vs. unknowns either by explicitly weighting them or implicitly through their relative percentages in the data. Although performance degrades when there is a significant disparity in their importance during threshold prediction stages, PostMax consistently outperforms and demonstrates greater stability across all thresholds. Refer to the supplemental for further details. + +Finally, to account for inherent random variations in predictions and tests, we developed training/validation/test splits, which are available on the GitHub repo. This protocol enables rigorous experimentation to assess the statistical significance of our evaluation. While the paper demonstrates statistically significant improvements across all results, some ablations in the supplemental do not. + +# 6 Conclusion + +This paper contributes to OSR research by shifting focus from common metrics on small-scale experiments to more realistic scenarios and datasets, thereby fostering more practically relevant research. We revisited and revised the criteria for effective open-set metrics as suggested by Wang et al. [46], making them more applicable to real-world scenarios. Our proposed Operational Open-Set Accuracy (OOSA), which incorporates known and unknown samples, offers a comprehensive metric and facilitates the selection of the most effective method and operating threshold for system operators. + +Our investigation of deep feature magnitudes of unknown samples from various ILSVRC2012 pre-trained networks yielded surprising findings. Contrary to smaller-scale studies [10, 31], we observed that these magnitudes are often larger for unknown than for known samples. This observation was consistent across multiple networks and types of unknown samples. Thus, we introduced our novel PostMax algorithm, refining normalized logit values from closed-set architectures into real probabilities of class inclusion using a Generalized Pareto Distribution (GPD). This distribution method offers advantages over previous extreme-value-based techniques, including negating the need to select a tail size and minimizing the influence of outliers. Also, it eliminates the need to define negative training samples as representatives of unknown classes, a common approach in much of the OSR literature that has not yielded a sufficient solution. + +While a good closed-set classifier is an important place to start [44], our novel PostMax algorithm shows that one can do better than "all you need" and that for operational usage, one needs to predict a threshold to use. With this paper, we empower engineers to operationalize their usage of OSR. + +# References + +1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2016) +2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing ImageNet out-of-distribution detection evaluation. In: ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models (2023) +3. Bodesheim, P., Freytag, A., Rodner, E., Denzler, J.: Local novelty detection in multiclass recognition problems. In: Winter Conference on Applications of Computer Vision (WACV) (2015) +4. Boult, T.E., Cruz, S., Dhamija, A.R., Günther, M., Henrydoss, J., Scheirer, W.J.: Learning and the unknown: Surveying steps toward open world recognition. In: AAAI Conference on Artificial Intelligence. vol. 33, pp. 9801-9807 (2019) +5. Cevikalp, H., Uzun, B., Salk, Y., Saribas, H., Köpukli, O.: From anomaly detection to open set recognition: Bridging the gap. Pattern Recognition 138, 109385 (2023) +6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 8065-8081 (2021) +7. Chen, G., Qiao, L., Shi, Y., Peng, P., Li, J., Huang, T., Pu, S., Tian, Y.: Learning open set network with discriminative reciprocal points. In: European Conference on Computer Vision (ECCV). Springer (2020) +8. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2014) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009) +0. Dhamija, A.R., Günther, M., Boult, T.: Reducing network agnostophobia. Advances in Neural Information Processing Systems (NeurIPS) 31 (2018) +1. Ge, Z., Demyanov, S., Garnavi, R.: Generative OpenMax for multi-class open set classification. In: British Machine Vision Conference (BMVC) (2017) +2. Gumbel, E.J.: Statistical Theory of Extreme Values and Some Practical Applications: A Series of Lectures, vol. 33. US Government Printing Office (1954) +3. Günther, M., Dhamija, A.R., Boult, T.E.: Watchlist adaptation: Protecting the innocent. In: International Conference of the Biometrics Special Interest Group (BIOSIG) (2020) +4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (2022) +5. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (ICML). PMLR (2022) +6. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR) (2017) +7. van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The INaturalist species classification and detection dataset In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +8. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems (NeurIPS) 34, 677-689 (2021) + +19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8710-8719 (2021) +20. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: International Conference on Computer Vision (ICCV). pp. 813-822 (2021) +21. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009) +22. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) +23. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23946-23955 (2023) +24. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11976-11986 (2022) +25. Lu, J., Xu, Y., Li, H., Cheng, Z., Niu, Y.: Pmal: Open set recognition via robust prototype mining. In: AAAI Conference on Artificial Intelligence. vol. 36:2, pp. 1872-1880 (2022) +26. Lyu, Z., Gutierrez, N.B., Beksi, W.J.: Metamax: Improved open-set deep neural networks via weibull calibration. In: Winter Conference on Applications of Computer Vision Workshops (WACVW) (2023) +27. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. In: European Conference on Computer Vision (ECCV). pp. 365-381. Springer (2022) +28. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: European Conference on Computer Vision (ECCV) (2018) +29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. In: NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011) +30. Palechor, A., Bhoumik, A., Günther, M.: Large-scale open-set classification protocols for imagenet. In: Winter Conference on Applications of Computer Vision (WACV). pp. 42-51. CVF/IEEE (2023) +31. Park, J., Chai, J.C.L., Yoon, J., Teoh, A.B.J.: Understanding the feature norm for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1557-1567 (2023) +32. Park, J., Jung, Y.G., Teoh, A.B.J.: Nearest neighbor guidance for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1686-1695 (2023) +33. Pickands III, J.: Statistical inference using extreme order statistics. The Annals of Statistics 3(1), 119 - 131 (1975) +34. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do Imagenet classifiers generalize to Imagenet? In: International Conference on Machine Learning (ICML). pp. 5389-5400. PMLR (2019) +35. Rudd, E.M., Jain, L.P., Scheirer, W.J., Boult, T.E.: The extreme value machine. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2017) +36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015) + +37. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. International Conference on Machine Learning (ICML) (2023) +38. Scheirer, W.J., Jain, L.P., Boult, T.E.: Probability models for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 36(11), 2317-2324 (2014) +39. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Towards open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 35(7) (2013) +40. Shu, L., Xu, H., Liu, B.: DOC: Deep open classification of text documents. In: Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics (2017) +41. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems (NeurIPS) 34, 144-157 (2021) +42. Vareto, R.H., Linghu, Y., Boult, T.E., Schwartz, W.R., Günther, M.: Open-set face recognition with maximal entropy and Objectosphere loss. Image and Vision Computing (IMAVIS) 141 (2024) +43. Vareto, R.H., Schwartz, W.R., Günther, M.: Toward open-set face recognition with neural ensemble, maximal entropy loss and feature augmentation. In: Conference on Graphics, Patterns and Images (SIBGRAPI) (2023) +44. Vaze, S., Han, K., Vedaldi, A., Zissermann, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022) +45. Wang, H., Li, Z., Feng, L., Zhang, W.: Out-of-distribution with virtual-logit matching. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4921-4930 (2022) +46. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards a priori-oriented open-set recognition. Advances in Neural Information Processing Systems (NeurIPS) 35, 25033-25045 (2022) +47. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sünderhauf, N.: Safe: Sensitivity-aware features for out-of-distribution object detection. In: International Conference on Computer Vision (ICCV). pp. 23565-23576 (2023) +48. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16133-16142 (2023) +49. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3485-3492. IEEE (2010) +50. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. In: International Conference on Learning Representations (ICLR) (2024) +51. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020) +52. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: OpenOOD: Benchmarking generalized out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS) 35, 32598-32611 (2022) + +53. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +54. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: OpenOOD v1.5: Enhanced benchmark for out-of-distribution detection. In: NeurIPS Workshop on Distribution Shifts: New Frontiers with Foundation Models (2023) +55. Zhang, X., Cheng, X., Zhang, D., Bonnington, P., Ge, Z.: Learning network architecture for open-set recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36(3), pp. 3362-3370 (2022) +56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017) +57. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021) \ No newline at end of file diff --git a/2024/Operational Open-Set Recognition and PostMax Refinement/images.zip b/2024/Operational Open-Set Recognition and PostMax Refinement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01364facfcbbe6c4c0ea49085a663a8328f68f9b --- /dev/null +++ b/2024/Operational Open-Set Recognition and PostMax Refinement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e18a8d7a2114231b8a3d466e4d2d8c8bcbe37e7d3776a632e0cbdd38952fe925 +size 352865 diff --git a/2024/Operational Open-Set Recognition and PostMax Refinement/layout.json b/2024/Operational Open-Set Recognition and PostMax Refinement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fd3f4529390e087f0f48e3668e2cc0902e054acd --- /dev/null +++ b/2024/Operational Open-Set Recognition and PostMax Refinement/layout.json @@ -0,0 +1,9276 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 171, + 111, + 444, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 111, + 444, + 148 + ], + "spans": [ + { + "bbox": [ + 171, + 111, + 444, + 148 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 206, + 168, + 408, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 168, + 408, + 193 + ], + "spans": [ + { + "bbox": [ + 206, + 168, + 408, + 193 + ], + "type": "text", + "content": "Steve Cruz1, Ryan Rabinowitz2, Manuel Günther3, and Terrance E. Boult2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 219, + 201, + 394, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 201, + 394, + 224 + ], + "spans": [ + { + "bbox": [ + 219, + 201, + 394, + 224 + ], + "type": "text", + "content": "1 University of Notre Dame \n2 University of Colorado Colorado Springs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 224, + 352, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 224, + 352, + 235 + ], + "spans": [ + { + "bbox": [ + 261, + 224, + 352, + 235 + ], + "type": "text", + "content": "3 University of Zurich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 236, + 457, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 236, + 457, + 247 + ], + "spans": [ + { + "bbox": [ + 156, + 236, + 457, + 247 + ], + "type": "text", + "content": "stevecruz@nd.edu {rrabinow, tboult}@uccs.edu guenther@ifi.uzh.ch" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 277, + 455, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 277, + 455, + 498 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 455, + 498 + ], + "type": "text", + "content": "Abstract. Open-Set Recognition (OSR) is a problem with mainly practical applications. However, recent evaluations have largely focused on small-scale data and tuning thresholds over the test set, which disregard the real-world operational needs of parameter selection. Thus, we revisit the original goals of OSR and propose a new evaluation metric, Operational Open-Set Accuracy (OOSA), which requires predicting an operationally relevant threshold from a validation set with known and a surrogate set with unknown samples, and then applying this threshold during testing. With this new measure in mind, we develop a large-scale evaluation protocol suited for operational scenarios. Additionally, we introduce the novel PostMax algorithm that performs post-processing refinement of the logit of the maximal class. This refinement involves normalizing logits by deep feature magnitudes and utilizing an extreme-value-based generalized Pareto distribution to map them into proper probabilities. We evaluate multiple pre-trained deep networks, including leading transformer and convolution-based architectures, on different selections of large-scale surrogate and test sets. Our experiments demonstrate that PostMax advances the state of the art in open-set recognition, showing statistically significant improvements in our novel OOSA metric as well as in previously used metrics such as AUROC, FPR95, and others." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 519, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "The original Open-Set Recognition (OSR) formulation [1,38,39] gained attention due to the flaws in existing recognition systems when handling inputs from outside the training classes. These systems lack control over inputs, and any unknown sample that is not rejected leads to a misclassification, thereby reducing accuracy. Therefore, early open-set work [1,3,35] introduced protocols tailored to emulate real-world systems, allowing for a proper evaluation of open-set performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "For a decade, researchers explored techniques that enhanced OSR performance [1,5,7,10,11,20,25,27,28,35,39,55,57]. While contributing novel ideas, recent works have diverged from the original goals of OSR. Instead of leveraging large-scale data as in early research on open-set deep networks [1], recent OSR evaluations have" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 125, + 246, + 203 + ], + "blocks": [ + { + "bbox": [ + 154, + 125, + 246, + 203 + ], + "lines": [ + { + "bbox": [ + 154, + 125, + 246, + 203 + ], + "spans": [ + { + "bbox": [ + 154, + 125, + 246, + 203 + ], + "type": "image", + "image_path": "ee72ea10b138a49268b1b1ed64f751f6311224ce748dfa177995a5f5af5983d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 204, + 218, + 213 + ], + "lines": [ + { + "bbox": [ + 184, + 204, + 218, + 213 + ], + "spans": [ + { + "bbox": [ + 184, + 204, + 218, + 213 + ], + "type": "text", + "content": "(a) Train" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 125, + 343, + 203 + ], + "blocks": [ + { + "bbox": [ + 250, + 125, + 343, + 203 + ], + "lines": [ + { + "bbox": [ + 250, + 125, + 343, + 203 + ], + "spans": [ + { + "bbox": [ + 250, + 125, + 343, + 203 + ], + "type": "image", + "image_path": "ffa717af6ff75c61ec29fa885705f87f51151c0316cbc16b292f1bd2501eedfb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 204, + 312, + 213 + ], + "lines": [ + { + "bbox": [ + 284, + 204, + 312, + 213 + ], + "spans": [ + { + "bbox": [ + 284, + 204, + 312, + 213 + ], + "type": "text", + "content": "(b) Val" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "text", + "content": "Fig. 1: POSTMAX PIPELINE. During train (a), post-processed maximum logits of correctly classified train samples are normalized by feature magnitude " + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "text", + "content": "; the set is then used in a Maximum Likelihood Estimation (MLE) to obtain Generalized Pareto Distribution (GPD) [33] with parameters " + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mu, \\sigma" + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "text", + "content": ". During validation (b), we use a validation and a surrogate set where the cumulative distribution of the GPD " + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "inline_equation", + "content": "(\\Psi)" + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "text", + "content": " gives the probability of the normalized max logit being from the train distribution and predict an operational threshold. During operation (c), we deploy each method's threshold and apply it to determine OSR and compute Operational Open Set Accuracy (OOSA). For evaluation we compare against Maximum Softmax Probability (MSP) [16,44], Maximum Logit (MaxLogit) [15,44], Nearest Neighbor Guidance (NNGuide) [32], and SCALE [50]. Operational Open-Set Accuracy (OOSA) performance is " + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "inline_equation", + "content": "(\\star)" + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "inline_equation", + "content": "(\\diamond)" + }, + { + "bbox": [ + 130, + 222, + 482, + 365 + ], + "type": "text", + "content": " is the optimal Open-Set Accuracy (OSA) for an oracle that knows the final test set (shown to highlight differences from OOSA). AUROC scores are shown for comparison." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 116, + 460, + 201 + ], + "blocks": [ + { + "bbox": [ + 347, + 116, + 460, + 201 + ], + "lines": [ + { + "bbox": [ + 347, + 116, + 460, + 201 + ], + "spans": [ + { + "bbox": [ + 347, + 116, + 460, + 201 + ], + "type": "image", + "image_path": "a6e86241cf88b5278f17d5bf7d74eb6e6f53a9f05afdaa8ee42ad43583b6bec9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 204, + 454, + 213 + ], + "lines": [ + { + "bbox": [ + 352, + 204, + 454, + 213 + ], + "spans": [ + { + "bbox": [ + 352, + 204, + 454, + 213 + ], + "type": "text", + "content": "(c) Test (OOSA & AUROC)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "content": "focused on small-scale datasets with few classes, with only a few notable exceptions [30,35,44]. Moreover, the prevalent reliance on metrics such as the Area Under the Receiver Operating Characteristics (AUROC) curve for performance evaluation has led to inopportune assessments since such metrics do not offer a realistic perspective on OSR. Fig. 1 illustrates an example where AUROC scores exhibit one ranking, yet in an operational setting, there is a slight difference. Inherently, common measures overlook the real-world operational needs of threshold selection. Some may argue for their use with a dedicated validation set; however, it is not clear how one would extract a threshold from these metrics and deploy it at test time. This is problematic as methods have largely ignored a fundamental practice of machine learning systems - having separate training, validation, and test sets to ensure that models generalize, avoid overfitting, and provide unbiased evaluation and reliable performance prediction on new data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Initial aspirations behind OSR aimed to enhance real systems [4]. Imagine an engineer tasked with implementing an OSR algorithm into their system. As they analyze results from various algorithms through tables and plots to inform decisions, a formidable challenge emerges. While AUROC and similar metrics assess a model's discriminative capability across all possible classification thresholds, they do not guide the engineer toward an optimal threshold for decision-making. There is a pressing need for a metric that enables confident and statistically rigorous evaluations with anticipated data, while also demonstrating robustness against varying numbers of unknowns. When deploying an open-set system, engineers must, at a minimum, consider two factors: (1) how an" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "approach scales with operational data, and (2) how to choose an operating point that balances their desired known classification accuracy and unknown rejection. These considerations underscore the importance of achieving overall accuracy with a predicted operational threshold. Therefore, we introduce a new metric called Operational Open-Set Accuracy (OOSA), illustrated in Fig. 1, which facilitates algorithm and operational threshold selection. To ensure real-world applicability, we implement a large-scale evaluation protocol with proper training/validation/test splits. We determine a threshold on the validation set and deploy it during testing on unseen unknowns. Additionally, for completeness, fairness, and robustness, we evaluate state-of-the-art algorithms on ImageNet-scale open-set recognition using common metrics (AUROC, FPR95, AUOSCR [6], and OpenAUC [46]), many of which are detailed in the supplemental." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 511 + ], + "type": "text", + "content": "Before incorporating an operational threshold, it is natural to ask about the distribution of scores above this threshold. This naturally leads to a Peak-over-Threshold (POT) formulation of the distribution, best modeled via Extreme Value Theory (EVT). The Pickands-Balkema-De Haan theorem of EVT [33] yields a Generalized Pareto Distribution (GPD) over normalized scores, as depicted in Fig. 1. This GPD-based score transformation provides theoretically grounded probabilities rather than raw, network-dependent ad-hoc confidence scores (see supplemental for details). While GPD can be applied directly to Softmax confidences or logit values, our approach goes further. We introduce a novel algorithm, PostMax, which applies GPD to post-processed normalized maximum logits. Here, normalization by the deep feature magnitude enhances the separation of known and unknown classes. Prior observations and theoretical explanations [10,13,31,44] suggest that deep feature magnitudes extracted from inputs that the network was not trained on are generally smaller than those for known samples. However, our findings show the exact opposite, i.e., that modern networks trained on the large-scale ILSVRC2012 dataset generally exhibit larger deep feature magnitudes for unknown samples (details in supplemental). Therefore, applying GPD to logit values divided by the deep feature magnitude can enhance OSR beyond the strong baselines of using raw Softmax confidence [16], maximum logit values [15], overconfidence reduction with features and logits [32], or scaled logit values [50]. The contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 513, + 480, + 643 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 513, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 559 + ], + "type": "text", + "content": "- We design a novel evaluation metric, Operational Open-Set Accuracy (OOSA), that emphasizes real-world usage by predicting an operational threshold on a validation set (knowns and unknowns surrogate), which is then deployed during testing on unseen data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 560, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 560, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 138, + 560, + 480, + 594 + ], + "type": "text", + "content": "- We introduce a novel algorithm, PostMax, that post-processes maximum logits by normalizing them by deep feature magnitude and then applies GPD to provide probabilities. Our code is publicly available.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 596, + 480, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 596, + 480, + 618 + ], + "spans": [ + { + "bbox": [ + 138, + 596, + 480, + 618 + ], + "type": "text", + "content": "- We develop new large-scale evaluation protocols suited for assessing algorithms in operational scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 620, + 480, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 620, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 620, + 480, + 643 + ], + "type": "text", + "content": "- We showcase that PostMax advances the state of the art in large-scale open-set recognition with statistical significance on OOSA and prior metrics." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 337, + 665 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 337, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 337, + 665 + ], + "type": "text", + "content": "4https://github.com/Vastlab/PostMax-00SA" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "content": "Improving OSR, distinct from OOD detection, anomaly detection, or novel category discovery, can be approached in two ways: providing better features through improved network training, or through post-processing where a pretrained network is trained for closed-set tasks and then adapted for OSR. Our work is rooted in the post-processing approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 251, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 251, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 251, + 227 + ], + "type": "text", + "content": "2.1 Related Problems" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 483, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 483, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 483, + 426 + ], + "type": "text", + "content": "A common question regarding open-set is, What about a two-stage system with OOD followed by classification, does that not solve OSR? This hypothesis is rejected with an example from one of our operational systems where we have to recognize objects in novel contexts, e.g., in snowy or foggy conditions. Such conditions are OOD with respect to training samples, but they should not be rejected from an open-set point of view. Out-Of-Distribution (OOD) detection, though distinct from OSR, has been more consistent in its use of large-scale experimentation as were used in early OSR work [15,16,32,45,47,50,52,54]. Even when OOD is restricted to out-of-class, the two-stage system needs to be evaluated in that context, including a process for selecting an OOD threshold and then evaluating the resulting classification performance. The OpenOOD Benchmark [52,54] has implemented various OSR and OOD techniques such as OpenGan [20], MOS [19], ReAct [41], ViM [45], GEN [23], NNGuide [32], and SCALE [50]. While we caution the use of OOD algorithms in OSR evaluations/settings, based on the OpenOOD Benchmark leaderboard, we compare with the state-of-the-art, NNGuide & SCALE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 443, + 269, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 269, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 269, + 454 + ], + "type": "text", + "content": "2.2 Closed-Set Classifiers" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 510 + ], + "type": "text", + "content": "Recently, Vaze et al. [44] argued that closed-set classifiers are sufficient. This aligns with Hendrycks et al. [15, 16], who demonstrated thresholding on Softmax confidences or, especially, on logits provide unreasonably good baselines. Our work shows that such approaches/networks can be improved with normalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 484, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 484, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 484, + 605 + ], + "type": "text", + "content": "We evaluate several leading pre-trained architectures trained on ILSVRC2012-1K with no additional data. Particularly, Meta's Vision Transformer Hiera-H [37], which strips non-essential components making it faster and more accurate during training/inference. Also, we use their Masked Autoencoders (MAE) model, ViT-H [14], which masks random patches of an image and reconstructs the missing pixels. To show generalization, we also utilize CNNs, including Meta's ConvNeXtV2-H [48] and ConvNeXt-L [24] which modernize a standard ResNet. Other networks are found in the supplemental material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 623, + 302, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 302, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 302, + 635 + ], + "type": "text", + "content": "2.3 Post-processing Approaches" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Besides thresholding softmax scores or logits [15, 16, 44], OSR methods take features and use Weibull-calibrated Support Vector Machines (W-SVM) [38]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "or Extreme Value Machines (EVM) [35] to estimate a probability of unknown. OpenMax [1] was the first method to add an artificial probability of unknown to closed-set networks by computing a logit score for the unknown class based on deep feature similarities to features of known classes. Other approaches include a small adaptor network [13, 42] that adds open-set capabilities [10] to features extracted from closed-set networks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 213, + 299, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 299, + 226 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 299, + 226 + ], + "type": "text", + "content": "2.4 Learning-based Approaches" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 326 + ], + "type": "text", + "content": "Most methods use negative aka. known unknown samples in training, which represent samples of no interest to the classifier and should be classified as unknown, with a few exemptions such as replacing the final Softmax classifier with a set of binary classifiers [40, 43]. While some use real negative samples [10, 13, 30], the majority of research tries to artificially create them. For example, combinations of knowns [57], added noise to knowns [47], or Generative Adversarial Networks (GANs) [11, 20]. In our approach, we avoid the use of any negative samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 348, + 274, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 274, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 274, + 361 + ], + "type": "text", + "content": "2.5 Evaluation Techniques" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 460 + ], + "type": "text", + "content": "Evaluating OSR correctly in a real-world setting is difficult and current measures do not satisfy operational requirements. For example, the widely used AUROC metric [5,7,27,28,44,51] only looks at the binary decision of known vs. unknown, but ignores the task of assigning the correct class for known samples. Such metrics are often combined with closed-set classification accuracy in order to evaluate OSR. These evaluations are reasonable, but provide no information under realistic scenarios, i.e., an operational threshold classifying samples as known or unknown." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 533 + ], + "type": "text", + "content": "Another metric often employed is the macro-averaged F1 score [1,27,35,53], which treats unknown as a separate class. This measure has many counter-intuitive properties that make it difficult to interpret the results. For example, a different threshold is required for each class so a sample could be classified as several known classes and unknown (at the same time). Thus, incorrect classification of unknown samples can easily be overlooked." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Recently, Dhamija et al. [10] introduced the Open-Set Classification Rate (OSCR) curve. This curve handles knowns and unknowns separately, evaluating the Correct Classification Rate (CCR) at various thresholds corresponding to specific False Positive Rates (FPR). While valuable for high-risk applications like open-set face recognition, its practicality in general OSR tasks is questionable. In scenarios where a threshold is chosen based on a specific FPR, the curve utilizes unknowns from the test set, raising doubts about its applicability to unseen classes. Moreover, the curve faces the challenge of not providing a single value for comparing different methods. Additionally, the proposed metrics, Area Under the OSCR (AUOSCR) [6] and Open Area Under the ROC Curve (OpenAUC) [46], do not provide an intuition on how to select an operating threshold." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 387, + 184 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 387, + 184 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 387, + 184 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 387, + 184 + ], + "type": "image", + "image_path": "87d6f851dd352aa9e8462bda1bcfc8cfd879633a0dfc6b83807698c1099d5913.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 195, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 316 + ], + "type": "text", + "content": "Fig. 2: MOTIVATION FOR NORMALIZATION. We illustrate why PostMax divides by the feature magnitude norm before applying GPD. The comparison above shows squared deep feature values extracted from a pre-trained ILSVRC2012-1K Transformer (ViT-H [14]). We sorted the deep features of 20 randomly sampled images from class 990 based on increasing mean square magnitude. Next, we applied this sorted order to the features of 20 images misclassified as class 990 from each unknown dataset. Unknowns exhibit high feature values (bright colors) in regions where known sample activations are low. These large, uncorrelated responses accumulate to give it a high score for the class but also result in unknown dataset features having, on average, larger " + }, + { + "bbox": [ + 130, + 195, + 482, + 316 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 195, + 482, + 316 + ], + "type": "text", + "content": " norms than those of the correct class, depicted in the magnitude histogram on the right. Therefore, dividing by magnitude enhances separation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 387, + 116, + 478, + 184 + ], + "blocks": [ + { + "bbox": [ + 387, + 116, + 478, + 184 + ], + "lines": [ + { + "bbox": [ + 387, + 116, + 478, + 184 + ], + "spans": [ + { + "bbox": [ + 387, + 116, + 478, + 184 + ], + "type": "image", + "image_path": "3165a2e2b460cea80cc92ed43a899d77a49b290f0b66439493bc949ca32b1eed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 338, + 212, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 212, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 212, + 353 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 460 + ], + "type": "text", + "content": "Several approaches advocate selecting the maximum class and then applying thresholding to certain scores, e.g., Softmax [16] or logits [15], to address OSR. While such thresholding proves effective in some recognition tasks [44], its relevance often extends only to establishing a ranked order of classes for individual samples. We focus on leveraging information from the features themselves and the formal distribution of these maxima. The result is PostMax, an effective approach with theoretically grounded probabilities to address OSR through (1) normalization and (2) score distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 478, + 275, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 275, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 275, + 489 + ], + "type": "text", + "content": "3.1 Normalization & GPD" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "content": "From Fig. 2, unknown samples activate high dimensions in the feature space, which goes against our intuition that deep networks would only learn features necessary to classify known classes and disregard those belonging to other uninteresting ones. We attribute these findings to high logit values and Softmax scores, where positions with high feature magnitudes align with larger weights of certain classes, even when an input exhibits a large feature magnitude and the high-scoring class has a small or negative weight. The observed effect may depend on the network's loss function and the training dataset; in small-scale datasets such as MNIST or CIFAR, which only differentiate between 10 classes, the opposite has been observed [10, 18, 31, 44]. Further details regarding the normalizations in these works and how they differ from our findings are provided in the supplemental." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "As we classify samples based on their \"maximum score\" threshold, a key question arises: Does this score belong to the training distribution of maximum scores? Statistical Extreme Value Theory (EVT) provides a grounded approach to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 140, + 310, + 244 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 138 + ], + "type": "text", + "content": "Algorithm 1 POSTMAX FITTING. This implements the probability distribution estimation of the proposed PostMax algorithm." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "lines": [ + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{train}}" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "\\operatorname {FV}(x)" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "\\operatorname {L}(x)" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "M\\gets \\{\\}" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "text", + "content": " \nfor each " + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "(x,t)\\in D_{\\mathrm{train}}" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "text", + "content": " do if arg max " + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "\\mathrm{L}(x) = t" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "\\ell \\leftarrow \\frac{\\max\\mathrm{L}(x)}{\\|\\mathrm{FV}(x)\\|}" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "M\\gets M\\cup \\{\\ell \\}" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "text", + "content": " end if \nend for \n" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "inline_equation", + "content": "\\mu ,\\sigma ,\\xi \\gets" + }, + { + "bbox": [ + 132, + 140, + 310, + 244 + ], + "type": "text", + "content": " scipy.stats.genpareto.fit(M)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 414 + ], + "type": "text", + "content": "answering this question. Unlike ad-hoc Softmax confidences, transforming logits into real probabilities allows probabilities to be more interpretable compared to Softmax, probabilities do not need to sum up to one (images may contain multiple objects), and EVT models probabilities on are based on thousands/millions of samples rather than the current sample. Given that thresholding maximum logits or Softmax scores serves as a reasonable baseline [15,32,44], applying EVT to this problem is both theoretically justified and intuitive. Note, prior OSR research [1,26,35,38] utilizes the Fisher-Tippett-Gnedenko EVT [12], resulting in a Weibull distribution. However, practical effectiveness was limited due to challenges in parameter selection affecting EVT modeling. In contrast, we employ the Generalized Pareto Distribution (GPD) derived from the Peak-over-threshold (POT) approach [33], which models extreme values above a threshold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 430, + 206, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 206, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 206, + 441 + ], + "type": "text", + "content": "3.2 PostMax" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": "Using these intuitions and observations, we define our Postnormalization of Maxima (PostMax) algorithm. We aim to find a general distribution " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": " of logit values that is valid for all of our known classes " + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "inline_equation", + "content": "c\\in \\{1,\\ldots ,C\\}" + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": ". As observed in Fig. 2, modeling raw logit values might not be fruitful since unknown samples have generally higher activation of deep features, which often leads to high logit values. For a sample, we utilize normalized logit values by dividing the original logit values by their deep feature magnitude via (1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": "Based on POT EVT [33], we understand that if we consider all maximum values above a threshold (such as the smallest correct logit observed during training), the resulting distribution follows a Generalized Pareto Distribution (GPD). Let " + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{train}} = \\{(x_i, t_i)\\}" + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": " be the collection of all samples " + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": " with their respective ground-truth target label " + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": " in the training set. Let " + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathrm{FV}(x)" + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": " be the function returning the feature vector of a sample " + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathrm{L}(x)" + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": " be the function returning the logit vector of a sample. We collect the normalized logits from the target class for all correctly classified training samples, which are then used to model a Generalized Pareto Distribution " + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\varPsi_{\\mu, \\sigma, \\xi}" + }, + { + "bbox": [ + 130, + 533, + 482, + 642 + ], + "type": "text", + "content": ". Details can be found in Alg. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "At inference for test sample " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " we use " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{FV}(x)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{L}(x)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " to extract the deep features and logits of " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " and then for class " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " assign probability " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_c(x)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " using the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "type": "text", + "content": "normalized logit value:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 137, + 481, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 137, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 203, + 137, + 481, + 163 + ], + "type": "interline_equation", + "content": "\\ell_ {c} = \\frac {\\operatorname {L} (x) [ c ]}{\\| \\operatorname {F V} (x) \\|} \\quad p _ {c} (x) = \\Psi_ {\\mu , \\sigma , \\xi} (\\ell_ {c}) \\tag {1}", + "image_path": "42dbfea9649eb4c1cf0d626ff8f77c8eb033740089740c42de2635fbdde6d5a6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "inline_equation", + "content": "\\varPsi_{\\mu,\\sigma,\\xi}(\\ell)" + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "text", + "content": " is the cumulative distribution of the GPD with location " + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "text", + "content": " and shape " + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 130, + 171, + 481, + 207 + ], + "type": "text", + "content": " computed via Alg. 1. If we just want the top class (as all experiments), we compute the class with the max logit and apply the above to that logit." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 208, + 481, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 208, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 481, + 256 + ], + "type": "text", + "content": "Note that we do not utilize any unknowns or negatives in this process, nor do we explicitly model a probability of unknowns. Instead, we threshold based on the cumulative GPD probability of known classes, computing the maximum value of normalized logits." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 321, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 321, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 321, + 286 + ], + "type": "text", + "content": "3.3 Operational Open-Set Accuracy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 295, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 481, + 319 + ], + "type": "text", + "content": "By designing the OpenAUC metric, Wang et al. [46] proposed four different conditions that a good open-set evaluation metric should fulfill, which we rephrase:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 319, + 481, + 366 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 129, + 319, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 319, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 129, + 319, + 481, + 342 + ], + "type": "text", + "content": "P1 The metric needs to check that known samples are classified correctly with high probability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 343, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 343, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 129, + 343, + 481, + 366 + ], + "type": "text", + "content": "P2 The metric needs to evaluate if unknown samples are assigned to known classes with low probability." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 129, + 367, + 383, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 367, + 383, + 378 + ], + "spans": [ + { + "bbox": [ + 129, + 367, + 383, + 378 + ], + "type": "text", + "content": "P3 The metric should be insensitive to a score threshold." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 379, + 318, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 379, + 318, + 390 + ], + "spans": [ + { + "bbox": [ + 129, + 379, + 318, + 390 + ], + "type": "text", + "content": "P4 The metric should be a single number." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 391, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 450 + ], + "type": "text", + "content": "We totally agree with P1, P2 and P4. However, we reject P3 because it means that the metric cannot be used to select a classifier that works well in a specific operational setting, which requires obtaining an operational threshold. In our view, a useful open-set algorithm must include that step. Therefore, we reformulate condition P3' to be operationally relevant:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 451, + 481, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 486 + ], + "type": "text", + "content": "P3' The metric should indicate an operational threshold that is optimal under specified circumstances and can be applied to unseen data and unseen unknown classes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 487, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 582 + ], + "type": "text", + "content": "These four conditions (P1, P2, P3', P4) exclude all existing open-set evaluation metrics; we need a new metric to satisfy them. As discussed in Sec. 2.5, metrics such as AUROC, AUOSCR, and OpenAUC simply average over all possible operational thresholds, while F1 and normalized accuracy fail to use a single threshold [46]. On the other hand, the Open-Set Classification Rate (OSCR) curve [10] does not provide a single value for comparison, which makes it hard to compare different algorithms. Therefore, we introduce Operational Open-Set Accuracy (OOSA), which is inspired by OSCR but fulfills all four conditions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 583, + 481, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 583, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 583, + 481, + 643 + ], + "type": "text", + "content": "From an engineer's view, the real question is how well a specific method can perform under the presence of both known and unknown samples in their data. Most applications of open-set classification are not security-sensitive, so there is no need to have very low False Positive Rates (FPR) as typically evaluated in the OSCR [10,30]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "type": "text", + "content": "Usually, an engineer has a rough estimate of how many known and unknown samples the system will observe - so they are rather interested in the total accuracy" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 173 + ], + "type": "text", + "content": "the system will achieve and which kind of operational threshold will lead to this performance. To provide both, we exploit the FPR and CCR calculations used in OSCR [10] to define the Unknown Rejection Rate (URR) and the Open-Set Accuracy (OSA), both of which rely on the operational threshold " + }, + { + "bbox": [ + 131, + 116, + 482, + 173 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 116, + 482, + 173 + ], + "type": "text", + "content": " and only make use of the probabilities of known classes " + }, + { + "bbox": [ + 131, + 116, + 482, + 173 + ], + "type": "inline_equation", + "content": "P_{c}(x)" + }, + { + "bbox": [ + 131, + 116, + 482, + 173 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 188, + 180, + 426, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 180, + 426, + 204 + ], + "spans": [ + { + "bbox": [ + 188, + 180, + 426, + 204 + ], + "type": "interline_equation", + "content": "\\operatorname {C C R} (\\theta) = \\frac {\\left| \\left\\{x \\mid x \\in D _ {c} \\wedge \\arg \\max _ {c} P _ {c} (x) = \\hat {c} \\wedge P _ {c} (x) \\geq \\theta \\right\\} \\right|}{\\left| D _ {c} \\right|}", + "image_path": "be053941664aec9d6c2de173e97fec5ec6a8006a6246c524e5ff01832465905a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 206, + 481, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 206, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 219, + 206, + 481, + 231 + ], + "type": "interline_equation", + "content": "\\operatorname {U R R} (\\theta) = \\frac {\\left| \\left\\{x \\mid x \\in D _ {a} \\wedge \\max _ {c} P _ {c} (x) < \\theta \\right\\} \\right|}{\\left| D _ {a} \\right|} \\tag {2}", + "image_path": "26996e1516740ceffdd0a8e254a34b64a2509e07dcb5942d0d90a14338eb2b67.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 224, + 234, + 390, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 234, + 390, + 245 + ], + "spans": [ + { + "bbox": [ + 224, + 234, + 390, + 245 + ], + "type": "interline_equation", + "content": "\\mathrm {O S A} (\\theta) = \\alpha \\cdot \\mathrm {C C R} (\\theta) + (1 - \\alpha) \\cdot \\mathrm {U R R} (\\theta)", + "image_path": "93881b317f48caf6d776dd883158d19b28fd5320e6afae41aaf3223579e66c19.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " is the collection of all known test samples with class labels " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "D_{a}" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " collects all unknown test samples [10]. Please note that " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\mathrm{URR}(\\theta) = 1 - \\mathrm{FPR}(\\theta)" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " provides a positive view on the unknown samples, i.e., it calculates how many of them are correctly rejected under threshold " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": ". OSA includes a weighted average over CCR and URR to provide a holistic view of the expected total accuracy, where the application-dependent " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " can be used to provide a stronger focus on correctly classifying known or unknown samples. When both types of samples should be treated similarly, " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\alpha = \\frac{|D_{c}|}{|D_{c}| + |D_{a}|}" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " can be selected, which we do in most of our experiments. In that case, OSA can be thought of as evaluating a " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " class classifier using " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " class accuracy where the unknown class " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": " is classified when the maximal probability does not exceed our operational threshold " + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 253, + 482, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 387, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 482, + 434 + ], + "type": "text", + "content": "When plotting OSA over URR, by varying the operational threshold, we arrive at the plots seen in Fig. 1. This plot is not monotonic, but rather provides a peak where the OSA is maximized. The maximum value can be used to select a threshold " + }, + { + "bbox": [ + 131, + 387, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 131, + 387, + 482, + 434 + ], + "type": "text", + "content": " on a given set of scores " + }, + { + "bbox": [ + 131, + 387, + 482, + 434 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 387, + 482, + 434 + ], + "type": "text", + "content": " effectively:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 258, + 444, + 356, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 444, + 356, + 456 + ], + "spans": [ + { + "bbox": [ + 258, + 444, + 356, + 456 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\arg \\max \\mathrm {O S A} _ {S} (\\theta)", + "image_path": "2a6095c85e82cfc7e5042f4a2ca2cc3efb2e0f70c1910fd31cde647865dc6463.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 290, + 456, + 481, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 456, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 290, + 456, + 481, + 468 + ], + "type": "interline_equation", + "content": "\\theta \\in S \\tag {3}", + "image_path": "1a7c97ecbb1c2648a884dc7f2ee776d16b23320c3480fecac93c411acfb79b99.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 207, + 466, + 406, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 466, + 406, + 481 + ], + "spans": [ + { + "bbox": [ + 207, + 466, + 406, + 481 + ], + "type": "interline_equation", + "content": "S = \\left\\{P _ {\\hat {c}} (x) \\mid x \\in D _ {c} \\right\\} \\cup \\left\\{\\max _ {c} P _ {c} (x) \\mid x \\in D _ {a} \\right\\}", + "image_path": "684c7b90e91a90dc8568b68432291a18f0a4baecec21deab48ebf08fb3fa36fb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": " is the collection of all maximum scores for known and unknown samples. When applying this calculation on the test set " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "S_{test}" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": ", we will find the optimal threshold " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\theta_{test}^{*}" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": ", indicated by " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": " in Fig. 1. However, since test set samples are unknown in operational settings, we make use of the validation set " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "D_{val}" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": " of known samples as " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": " and an additional surrogate set " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "D_{sur}" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": " of unknown samples to replace " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "D_{a}" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": " in (3) for obtaining the operational threshold " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\theta_{sur}^{*}" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": ", which we will then apply to the test set: " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{OSA}_{\\text{test}}(\\theta_{sur}^{*})" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": ", which is indicated by " + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 131, + 491, + 482, + 586 + ], + "type": "text", + "content": " in Fig. 1. How to optimally select this surrogate set is discussed below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 604, + 230, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 230, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 230, + 618 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": "Although a few OSR methods have explored large-scale datasets [1,30,35,44], many recent techniques [5-7,10,11,20,25,27,28,55,57] still lack comprehensive large-scale evaluations. The performance on small-scale data such as MNIST [22]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "text", + "content": "Table 1: OPERATIONAL OPEN-SET ACCURACY (OOSA). The mean OOSA " + }, + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "text", + "content": " of all methods. To compute, we validate methods on ImageNetV2 [34] (10K images) as knowns and " + }, + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "text", + "content": " of our surrogate 21K-P Hard [44] (9.8K images) as unknowns and predict an operational threshold. Then, we deploy each method's threshold and test on five different ILSVRC2012 val [36] splits (each 10K images) and specified unknowns. Standard deviation is omitted as it is " + }, + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "inline_equation", + "content": "< 0.005" + }, + { + "bbox": [ + 130, + 114, + 482, + 213 + ], + "type": "text", + "content": " for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012 - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 224, + 482, + 305 + ], + "blocks": [ + { + "bbox": [ + 133, + 224, + 482, + 305 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 482, + 305 + ], + "type": "table", + "html": "
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8250.8150.7800.7820.7270.8120.7960.7860.7130.772
NINCO [2] (~5.8K)0.7400.7390.7070.6250.6670.7320.7230.7180.5970.705
OpenImage-O [45] (~17.6K)0.8640.8140.7450.7960.6630.8490.8200.7920.7450.767
Places [56] (10K)0.7850.7540.6910.7450.6250.7790.7490.7100.6930.676
SUN [49] (10K)0.7950.7580.7090.7490.6510.7830.7510.7210.6930.690
Textures [8] (~5.1K)0.7640.7560.7360.6600.7150.7440.7360.7400.6010.740
21K-P Easy [44] (50K)0.8180.7400.6620.7440.5680.8130.7630.6910.7830.633
", + "image_path": "1b5eade47e5a8af1c203f365afaef56c244289cc1bd659f0b083519a4aa06a05.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 325, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 348 + ], + "type": "text", + "content": "CIFAR [21], or SVHN [29] does not reflect operational scenarios and lacks generalizability due to small image crops and a limited number of classes [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "content": "An open-set protocol is essential and necessitates a large-scale setting for the proper evaluation of real-world scenarios [30, 44]. Additionally, it is paramount to employ proper machine-learning methodology, including a validation set (and a surrogate set) for operational threshold selection. Thus, we evaluate the robustness and generalization of PostMax on multiple large-scale datasets and splits with realistic and high-quality images. We also leverage the recently introduced large-scale ImageNet open-set splits [44]. Lastly, we conduct ablations to understand the performance impact of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 461, + 196, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 196, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 196, + 472 + ], + "type": "text", + "content": "4.1 Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "type": "text", + "content": "Architectures. We use state-of-the-art pre-trained models on ILSVRC2012 [36] with standard " + }, + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "type": "text", + "content": " image crops - Meta's Hiera-H [37] and ConvNeXtV2-H [48]. Meta fine-tuned these models with no additional/external data. Note, we do not perform additional training or fine-tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": "Datasets. For knowns, we utilize the large-scale ImageNet [9] subset ILSVRC2012 [36]. This dataset provides an advantage due to the extensive variety of available pre-trained architectures, enabling us to compare with more diverse deep networks. Our PostMax model is trained on the ILSVRC2012 train split, which contains " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\sim 1.28M" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": " images and " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "1K" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": " classes (PostMax training requires no additional negative/unknown data). For validation, we employ ImageNetV2 [34], which contains " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": " images. For testing, we split ILSVRC2012 val into five equal splits of " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": " images each." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "For unknowns, we use iNaturalist [17], NINCO [2], OpenImage-O [45], Places [56], SUN [49], Textures [8], and ImageNet-21K-P Open-Set splits [44]. The vision" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 190, + 482, + 269 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: COMMON METRICS. The mean AUROC " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " and FPR95 " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " of all methods. To compute, we test each method on five different ILSVRC2012 val [36] splits (each " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " images) as knowns and specified unknowns. Standard deviation is omitted as it is " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "< 0.01" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " for all methods/unknowns. OSR is performed on extractions from two state-of-the-art pre-trained architectures on ILSVRC2012-1K - (1) Transformer Hiera-H [37] (2) CNN ConvNeXtV2-H [48]. The best scores are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 190, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 133, + 190, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 482, + 269 + ], + "type": "table", + "html": "
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .21.92 / .37.88 / .36.90 / .50.81 / .51.95 / .26.91 / .42.90 / .37.86 / .58.88 / .44
NINCO [2] (~ 5.8K).87 / .58.83 / .62.79 / .61.69 / .78.73 / .67.86 / .56.83 / .64.82 / .60.71 / .81.80 / .62
OpenImage-O [45] (~ 17.6K).95 / .25.88 / .49.80 / .49.88 / .45.72 / .61.94 / .29.88 / .49.87 / .44.81 / .70.85 / .47
Places [56] (10K).89 / .49.83 / .60.75 / .61.85 / .55.67 / .71.89 / .50.84 / .62.79 / .60.82 / .64.76 / .64
SUN [49] (10K).90 / .46.84 / .57.78 / .57.85 / .56.71 / .66.89 / .47.84 / .59.81 / .57.83 / .60.77 / .61
Textures [8] (~ 5.1K).93 / .32.88 / .47.86 / .42.74 / .86.83 / .46.91 / .34.87 / .49.88 / .40.73 / .85.88 / .38
21K-P Easy [44] (50K).86 / .50.79 / .64.72 / .64.80 / .68.65 / .73.85 / .51.79 / .65.75 / .64.78 / .72.72 / .68
", + "image_path": "7682b3609bbda7d20e90252620d03098f0e15613de476215b67606ba63321ff8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 291, + 482, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 375 + ], + "type": "text", + "content": "community [15, 16, 32, 44, 45, 50, 52, 54] has utilized these datasets as they present unique challenges not seen in small-scale settings. For validation, to predict an operational threshold, we utilize " + }, + { + "bbox": [ + 130, + 291, + 482, + 375 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 291, + 482, + 375 + ], + "type": "text", + "content": " of ImageNet-21K-P Hard (9.8K images) as our surrogate unknowns. For testing, NINCO, OpenImage-O, Textures, and ImageNet-21K-P Easy are entirely used, while others sample " + }, + { + "bbox": [ + 130, + 291, + 482, + 375 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 130, + 291, + 482, + 375 + ], + "type": "text", + "content": " images from a manual selection of classes. Further descriptions of all datasets and splits are provided in the supplemental." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 483, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 483, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 483, + 525 + ], + "type": "text", + "content": "Evaluation Metrics. To align with existing literature, we utilized common metrics in Tab. 2, namely, Area Under the Receiver Operating Curve (AUROC) and False Positive Rate at a fixed True Positive Rate of " + }, + { + "bbox": [ + 130, + 392, + 483, + 525 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 392, + 483, + 525 + ], + "type": "text", + "content": " (FPR95). Additional metrics like Area Under Open-Set Classification Rate (AUOSCR) [6] and Open Area Under ROC Curve (OpenAUC) [46] are provided in the supplemental. Alongside these, we introduce Operational Open-Set Accuracy (OOSA) in Tab. 1, where we predict an operational threshold on a validation and a surrogate set and deploy it at test time. This measure, described in detail in Sec. 3.3, serves as a more realistic alternative to existing measures. When evaluating OSR methods in real-world scenarios, other measures are insufficient as they do not provide an operating point for deployment and are usually only evaluated with a test set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 542, + 198, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 198, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 198, + 553 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": "Our main results are presented in Tab. 1 and 2, where we report performance on each unknown dataset mentioned in Sec. 4.1. We compare our approach with two commonly seen methods, Maximum Softmax Probability (MSP) [16,44] and Maximum Logit (MaxLogit) [15,44], along with the recently introduced Nearest Neighbor Guidance (NNGuide) [32] and SCALE [50]. We omit comparisons with older techniques such as OpenMax [1] and EVM [35] as they are much worse. According to the OpenOOD [52,54] Benchmark leaderboard for ImageNet-1K," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 318, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 318, + 666 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 318, + 666 + ], + "type": "text", + "content": "5https://zjysteven.github.io/Open00D" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 201, + 482, + 283 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 3: DIFFERENT VALIDATION - KNOWS & UNKNOWN'S SURROGATE. The Operational Open-Set Accuracy (OOSA) " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": " of all methods on a different validation set. To compute, we validate methods on ILSVRC2012 val [36] (50K images) as knowns and 21K-P Easy [44] (50K images) as unknowns surrogate and predict an operational threshold. Then, we deploy each method's threshold and test on ImageNetV2 [34] (10K images) and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 201, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 133, + 201, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 482, + 283 + ], + "type": "table", + "html": "
Unknowns(# images)OOSA ↑
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K)0.8630.8080.7800.8080.7160.8420.8060.7940.7590.763
NINCO [2] (~5.8K)0.7330.7300.7040.6550.6540.7590.7330.7260.6430.695
OpenImage-O [45] (~17.6K)0.8850.8100.7360.8090.6490.8650.8210.7860.7460.757
Places [56] (10K)0.8010.7480.6840.7660.6110.7950.7530.7080.7310.664
SUN [49] (10K)0.8120.7520.7020.7710.6350.8020.7570.7220.7380.679
Textures [8] (~5.1K)0.8130.7460.7390.6910.7060.7840.7500.7570.6570.732
21K-P Hard [44] (49K)0.6340.6730.5830.5350.5150.6470.6720.5930.5590.566
", + "image_path": "09b181ade35b4aabf1ab802d64d31db6e336ca4f42aa94eee45ddf47f8da556e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "text", + "content": "NNGuide stands out as the leader in the OOD space, closely followed by SCALE. We caution against relying solely on OOD methods and metrics, as they may not accurately reflect OSR performance. Note, some OSA and ROC curves are found in the supplemental." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 483, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 483, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 483, + 445 + ], + "type": "text", + "content": "When utilizing both architectures (Hiera-H & ConvNeXtV2-H) as feature extractors, PostMax outperforms all other methods in every measure (OOSA, AUROC, FPR95) on each unknown dataset. PostMax demonstrates superior performance, emphasizing robustness and generalizability. Effectively, PostMax achieves an excellent trade-off between classification accuracy and unknown rejection. Overall, the results suggest that a system operator can seamlessly integrate PostMax, enhance the performance beyond good closed-set classifiers, and confidently select the operational threshold that best suits its operation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": "Statistical Testing. To underscore the significance of performance differences between algorithms in our experiments, we prioritized the assessment of statistical significance. As a result, we performed paired, two-tailed t-tests, with Bonferroni corrections, to compare PostMax with the four methods (MSP, MaxLogit, NNGuide, and SCALE) using both architectures (Hiera-H and ConvNeXtV2-H) across 5 fold runs for each metric (OOSA, AUROC, and FPR95). On Hiera-H, PostMax (1) OOSA scores improvements range from " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " with corresponding P-values all less than " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{E}^{-7}" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": ", (2) AUROC scores improvements range from " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "17.9\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " with corresponding P-values all less than " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{E}^{-8}" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": ", and (3) FPR95 scores improvements range from " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "22.5\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " with corresponding P-values all less than " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{E}^{-8}" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": ". Similarly, on ConvNeXtV2-H, PostMax (1) OOSA scores improvements range from " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": ", (2) AUROC scores improvements range from " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": ", and (3) FPR95 scores improvements range from " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "28.1\\%" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": ", where all corresponding P-values are less than " + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{E}^{-8}" + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": " for all tests on all three metrics. In summary, the t-tests conducted for PostMax, considering the combined scores of all metrics across all datasets and architectures, revealed very statistically significant differences compared to every other method." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 482, + 247 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 4: DIFFERENT TESTING - PRIOR METRICS. The AUROC (↑) and FPR95 (↓) of all methods on different test sets. To compute, we test each method on ImageNetV2 [34] (10K images) as knowns and specified unknowns. OSR is performed again on extractions from Hiera-H [37] and ConvNeXtV2-H [48]. The best scores are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 482, + 247 + ], + "type": "table", + "html": "
Unknowns(# images)AUROC ↑ / FPR95 ↓
Hiera-H [37]ConvNeXtV2-H [48]
PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]PostMax(Ours)MSP[16, 44]MaxLogit[15, 44]NNGuide[32]SCALE[50]
iNaturalist [17] (10K).96 / .16.92 / .37.88 / .37.91 / .40.80 / .55.95 / .23.91 / .41.90 / .37.87 / .53.87 / .44
NINCO [2] (~5.8K).88 / .52.83 / .61.77 / .62.71 / .72.70 / .71.87 / .54.83 / .63.81 / .60.73 / .78.79 / .63
OpenImage-O [45] (~17.6K).96 / .20.87 / .48.79 / .49.90 / .38.70 / .64.94 / .27.88 / .49.86 / .43.82 / .65.84 / .48
Places [56] (10K).90 / .44.83 / .60.74 / .62.87 / .48.65 / .74.89 / .48.83 / .61.78 / .60.84 / .59.74 / .64
SUN [49] (10K).91 / .41.84 / .57.77 / .58.87 / .49.69 / .70.90 / .44.84 / .59.80 / .57.84 / .55.76 / .61
Textures [8] (~5.1K).94 / .27.88 / .47.85 / .42.77 / .79.81 / .49.92 / .32.87 / .49.88 / .39.75 / .82.87 / .39
21K-P Hard [44] (49K).77 / .68.73 / .73.68 / .73.68 / .76.62 / .82.76 / .72.74 / .75.70 / .74.67 / .82.67 / .79
", + "image_path": "a4f1e8b2377a2415f45d31460ae8d5da627e05efeec2c632bd7c1e757187f831.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 269, + 238, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 238, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 238, + 282 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 487, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 487, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 487, + 359 + ], + "type": "text", + "content": "In Tab. 3 and 4, we investigate the effects of varying the validation datasets (knowns and unknowns surrogate). We interchanged ImageNetV2 with ILSVRC2012 val for knowns and 21K-P Hard with 21K-P Easy for the unknowns surrogate. Unlike in Tab. 1, this involves validating on a larger number of knowns and unknowns, with an easier unknowns surrogate. Performance remained consistent regardless of the datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 407 + ], + "type": "text", + "content": "Also, we investigated the impact of other architectures as extractors and interchanged pre-trained models (ILSVRC2012-1K only w/ 224 crops) with Meta's ViT-H [14] and ConvNeXt-L [24]. Results showcasing PostMax's robustness are in the supplemental." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 443 + ], + "type": "text", + "content": "Lastly, we explored the impact of varying the application-dependent " + }, + { + "bbox": [ + 130, + 407, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 407, + 482, + 443 + ], + "type": "text", + "content": " from Eq. (2) and the normalization " + }, + { + "bbox": [ + 130, + 407, + 482, + 443 + ], + "type": "inline_equation", + "content": "(L_{2})" + }, + { + "bbox": [ + 130, + 407, + 482, + 443 + ], + "type": "text", + "content": " within PostMax. Details and results are in the supplemental." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 461, + 217, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 217, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 217, + 474 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 534 + ], + "type": "text", + "content": "This paper introduces the new evaluation metric Open-Set Accuracy (OSA) (2); we framed the evaluation of an open-set problem as predicting a threshold and then evaluating open-set accuracy at that threshold. This supports an engineer selecting an algorithm and a threshold to be applied in some Operational setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "content": "In real-world scenarios, unknown samples can be drawn from an unlimited variety, unlike known samples, which are assumed to be related to the training data. Thus, the evaluation must consider the computation of the predictive threshold on validation data and a separate evaluation on test data. The quality of the predicted threshold will depend on many factors with the two most influential being (1) how representative are the surrogate samples of the real unknowns, and (2) the relative proportion of knowns and unknowns. In Sec. 4.3 and our supplemental, we explore some variations of those factors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Maybe not surprisingly, if we train with \"harder\" unknowns, we generally see better performance. As shown in Tab. 1 and 2, the novel PostMax algorithm is better on unknowns across two networks (other networks in the supplemental)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "However, as shown in Sec. 4.3, when 21K-P Easy is used as our unknowns surrogate, PostMax does not exhibit an advantage on 21K-P Hard. In general, 21K-P Hard poses the most significant challenge for all algorithms. Notably, the sets close visual and semantic similarity to known samples suggests it represents a fine-grained categorization challenge rather than a typical open-set scenario." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 178, + 482, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 249 + ], + "type": "text", + "content": "One can adjust the importance of knowns vs. unknowns either by explicitly weighting them or implicitly through their relative percentages in the data. Although performance degrades when there is a significant disparity in their importance during threshold prediction stages, PostMax consistently outperforms and demonstrates greater stability across all thresholds. Refer to the supplemental for further details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "content": "Finally, to account for inherent random variations in predictions and tests, we developed training/validation/test splits, which are available on the GitHub repo. This protocol enables rigorous experimentation to assess the statistical significance of our evaluation. While the paper demonstrates statistically significant improvements across all results, some ablations in the supplemental do not." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 339, + 220, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 220, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 220, + 352 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 374, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 470 + ], + "type": "text", + "content": "This paper contributes to OSR research by shifting focus from common metrics on small-scale experiments to more realistic scenarios and datasets, thereby fostering more practically relevant research. We revisited and revised the criteria for effective open-set metrics as suggested by Wang et al. [46], making them more applicable to real-world scenarios. Our proposed Operational Open-Set Accuracy (OOSA), which incorporates known and unknown samples, offers a comprehensive metric and facilitates the selection of the most effective method and operating threshold for system operators." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "type": "text", + "content": "Our investigation of deep feature magnitudes of unknown samples from various ILSVRC2012 pre-trained networks yielded surprising findings. Contrary to smaller-scale studies [10, 31], we observed that these magnitudes are often larger for unknown than for known samples. This observation was consistent across multiple networks and types of unknown samples. Thus, we introduced our novel PostMax algorithm, refining normalized logit values from closed-set architectures into real probabilities of class inclusion using a Generalized Pareto Distribution (GPD). This distribution method offers advantages over previous extreme-value-based techniques, including negating the need to select a tail size and minimizing the influence of outliers. Also, it eliminates the need to define negative training samples as representatives of unknown classes, a common approach in much of the OSR literature that has not yielded a sufficient solution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "While a good closed-set classifier is an important place to start [44], our novel PostMax algorithm shows that one can do better than \"all you need\" and that for operational usage, one needs to predict a threshold to use. With this paper, we empower engineers to operationalize their usage of OSR." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Conference on Computer Vision and Pattern Recognition (CVPR). IEEE (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "text", + "content": "2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing ImageNet out-of-distribution detection evaluation. In: ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 481, + 224 + ], + "type": "text", + "content": "3. Bodesheim, P., Freytag, A., Rodner, E., Denzler, J.: Local novelty detection in multiclass recognition problems. In: Winter Conference on Applications of Computer Vision (WACV) (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "text", + "content": "4. Boult, T.E., Cruz, S., Dhamija, A.R., Günther, M., Henrydoss, J., Scheirer, W.J.: Learning and the unknown: Surveying steps toward open world recognition. In: AAAI Conference on Artificial Intelligence. vol. 33, pp. 9801-9807 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 277 + ], + "type": "text", + "content": "5. Cevikalp, H., Uzun, B., Salk, Y., Saribas, H., Köpukli, O.: From anomaly detection to open set recognition: Bridging the gap. Pattern Recognition 138, 109385 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "type": "text", + "content": "6. Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(11), 8065-8081 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "7. Chen, G., Qiao, L., Shi, Y., Peng, P., Li, J., Huang, T., Pu, S., Tian, Y.: Learning open set network with discriminative reciprocal points. In: European Conference on Computer Vision (ECCV). Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "8. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 429 + ], + "type": "text", + "content": "0. Dhamija, A.R., Günther, M., Boult, T.: Reducing network agnostophobia. Advances in Neural Information Processing Systems (NeurIPS) 31 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "type": "text", + "content": "1. Ge, Z., Demyanov, S., Garnavi, R.: Generative OpenMax for multi-class open set classification. In: British Machine Vision Conference (BMVC) (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 471 + ], + "type": "text", + "content": "2. Gumbel, E.J.: Statistical Theory of Extreme Values and Some Practical Applications: A Series of Lectures, vol. 33. US Government Printing Office (1954)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 471, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 503 + ], + "type": "text", + "content": "3. Günther, M., Dhamija, A.R., Boult, T.E.: Watchlist adaptation: Protecting the innocent. In: International Conference of the Biometrics Special Interest Group (BIOSIG) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "type": "text", + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "5. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: International Conference on Machine Learning (ICML). PMLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "6. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR) (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "text", + "content": "7. van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The INaturalist species classification and detection dataset In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "8. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems (NeurIPS) 34, 677-689 (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8710-8719 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 172 + ], + "type": "text", + "content": "20. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: International Conference on Computer Vision (ICCV). pp. 813-822 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 195 + ], + "type": "text", + "content": "21. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 217 + ], + "type": "text", + "content": "22. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 219, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 481, + 251 + ], + "type": "text", + "content": "23. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23946-23955 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "type": "text", + "content": "24. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "25. Lu, J., Xu, Y., Li, H., Cheng, Z., Niu, Y.: Pmal: Open set recognition via robust prototype mining. In: AAAI Conference on Artificial Intelligence. vol. 36:2, pp. 1872-1880 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "26. Lyu, Z., Gutierrez, N.B., Beksi, W.J.: Metamax: Improved open-set deep neural networks via weibull calibration. In: Winter Conference on Applications of Computer Vision Workshops (WACVW) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "text", + "content": "27. Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. In: European Conference on Computer Vision (ECCV). pp. 365-381. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "text", + "content": "28. Neal, L., Olson, M., Fern, X., Wong, W.K., Li, F.: Open set learning with counterfactual images. In: European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "text", + "content": "29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. In: NIPS Workshop on Deep Learning and Unsupervised Feature Learning (2011)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "30. Palechor, A., Bhoumik, A., Günther, M.: Large-scale open-set classification protocols for imagenet. In: Winter Conference on Applications of Computer Vision (WACV). pp. 42-51. CVF/IEEE (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "31. Park, J., Chai, J.C.L., Yoon, J., Teoh, A.B.J.: Understanding the feature norm for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1557-1567 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "32. Park, J., Jung, Y.G., Teoh, A.B.J.: Nearest neighbor guidance for out-of-distribution detection. In: International Conference on Computer Vision (ICCV). pp. 1686-1695 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "type": "text", + "content": "33. Pickands III, J.: Statistical inference using extreme order statistics. The Annals of Statistics 3(1), 119 - 131 (1975)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "34. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do Imagenet classifiers generalize to Imagenet? In: International Conference on Machine Learning (ICML). pp. 5389-5400. PMLR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "type": "text", + "content": "35. Rudd, E.M., Jain, L.P., Scheirer, W.J., Boult, T.E.: The extreme value machine. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 483, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "37. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. International Conference on Machine Learning (ICML) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 483, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 483, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 483, + 194 + ], + "type": "text", + "content": "38. Scheirer, W.J., Jain, L.P., Boult, T.E.: Probability models for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 36(11), 2317-2324 (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "39. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Towards open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 35(7) (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 482, + 259 + ], + "type": "text", + "content": "40. Shu, L., Xu, H., Liu, B.: DOC: Deep open classification of text documents. In: Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "type": "text", + "content": "41. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems (NeurIPS) 34, 144-157 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 326 + ], + "type": "text", + "content": "42. Vareto, R.H., Linghu, Y., Boult, T.E., Schwartz, W.R., Günther, M.: Open-set face recognition with maximal entropy and Objectosphere loss. Image and Vision Computing (IMAVIS) 141 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 482, + 358 + ], + "type": "text", + "content": "43. Vareto, R.H., Schwartz, W.R., Günther, M.: Toward open-set face recognition with neural ensemble, maximal entropy loss and feature augmentation. In: Conference on Graphics, Patterns and Images (SIBGRAPI) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 482, + 392 + ], + "type": "text", + "content": "44. Vaze, S., Han, K., Vedaldi, A., Zissermann, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 482, + 424 + ], + "type": "text", + "content": "45. Wang, H., Li, Z., Feng, L., Zhang, W.: Out-of-distribution with virtual-logit matching. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4921-4930 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 482, + 456 + ], + "type": "text", + "content": "46. Wang, Z., Xu, Q., Yang, Z., He, Y., Cao, X., Huang, Q.: OpenAUC: Towards a priori-oriented open-set recognition. Advances in Neural Information Processing Systems (NeurIPS) 35, 25033-25045 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 482, + 490 + ], + "type": "text", + "content": "47. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sünderhauf, N.: Safe: Sensitivity-aware features for out-of-distribution object detection. In: International Conference on Computer Vision (ICCV). pp. 23565-23576 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 482, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 482, + 523 + ], + "type": "text", + "content": "48. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16133-16142 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 482, + 555 + ], + "type": "text", + "content": "49. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3485-3492. IEEE (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 482, + 589 + ], + "type": "text", + "content": "50. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 482, + 621 + ], + "type": "text", + "content": "51. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 482, + 665 + ], + "type": "text", + "content": "52. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: OpenOOD: Benchmarking generalized out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS) 35, 32598-32611 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Operational Open-Set Recognition and PostMax Refinement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 281 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "53. Yoshihashi, R., Shao, W., Kawakami, R., You, S., Iida, M., Naemura, T.: Classification-reconstruction learning for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "text", + "content": "54. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: OpenOOD v1.5: Enhanced benchmark for out-of-distribution detection. In: NeurIPS Workshop on Distribution Shifts: New Frontiers with Foundation Models (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "55. Zhang, X., Cheng, X., Zhang, D., Bonnington, P., Ge, Z.: Learning network architecture for open-set recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36(3), pp. 3362-3370 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "text", + "content": "56. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 40(6), 1452-1464 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 260, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 482, + 281 + ], + "type": "text", + "content": "57. Zhou, D.W., Ye, H.J., Zhan, D.C.: Learning placeholders for open-set recognition. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Cruz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_content_list.json b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..63d1bc951a3880e89489f0e09ea9a74a041cf37b --- /dev/null +++ b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_content_list.json @@ -0,0 +1,1611 @@ +[ + { + "type": "text", + "text": "OphNet: A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Hu $^{1,2,3*}$ , Peng Xia $^{1,3*}$ , Lin Wang $^{5*}$ , Siyuan Yan $^{1,2}$ , Feilong Tang $^{1,3}$ Zhongxing Xu $^{7}$ , Yimin Luo $^{6}$ , Kaimin Song $^{3}$ , Jurgen Leitner $^{2}$ Xuelian Cheng $^{1}$ , Jun Cheng $^{8}$ , Chi Liu $^{9}$ , Kaijing Zhou $^{4\\dagger}$ , and Zongyuan Ge $^{1,2,3\\dagger}$", + "bbox": [ + 212, + 210, + 785, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ AIM Lab, Faculty of IT, $^{2}$ Faculty of Engineering, Monash University \n $^{3}$ Airdoc-Monash Research, Airdoc $^{4}$ Eye Hospital, Wenzhou Medical University \n $^{5}$ Bosch Corporate Research $^{6}$ King's College London $^{7}$ Cornell University Institute for Infocomm Research, A*STAR \n $^{9}$ Faculty of Data Science, City University of Macau", + "bbox": [ + 230, + 268, + 771, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Surgical scene perception via videos is critical for advancing robotic surgery, telesurgery, and AI-assisted surgery, particularly in ophthalmology. However, the scarcity of diverse and richly annotated video datasets has hindered the development of intelligent systems for surgical workflow analysis. Existing datasets face challenges such as small scale, lack of diversity in surgery and phase categories, and absence of time-localized annotations. These limitations impede action understanding and model generalization validation in complex and diverse real-world surgical scenarios. To address this gap, we introduce OphNet, a large-scale, expert-annotated video benchmark for ophthalmic surgical workflow understanding. OphNet features: 1) A diverse collection of 2,278 surgical videos spanning 66 types of cataract, glaucoma, and corneal surgeries, with detailed annotations for 102 unique surgical phases and 150 fine-grained operations. 2) Sequential and hierarchical annotations for each surgery, phase, and operation, enabling comprehensive understanding and improved interpretability. 3) Time-localized annotations, facilitating temporal localization and prediction tasks within surgical workflows. With approximately 285 hours of surgical videos, OphNet is about 20 times larger than the largest existing surgical workflow analysis benchmark. Code and dataset are available at: https://minghuo830.github.io/OphNet-benchmark/.", + "bbox": [ + 261, + 369, + 743, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Surgical Workflow Understanding $\\cdot$ Ophthalmic Surgery $\\cdot$ Medical Image Analysis $\\cdot$ Video Benchmark", + "bbox": [ + 261, + 672, + 740, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 758, + 375, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As surgical robot platforms such as the da Vinci® surgical system become increasingly sophisticated, there is growing interest in integrating enhanced intelligi-", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding author", + "bbox": [ + 230, + 825, + 566, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gence into scenarios like minimally invasive surgery [20,32,68]. The advancements in machine vision perception empower these robotic systems to autonomously recognize and adapt to the intricacies of surgical environments, without relying on binary instrument usage signals, RFID tags, sensor data from tracking devices, or other signal information that necessitates laborious manual annotations or additional equipment installations [28]. This autonomy includes the capacity to identify anatomical structures, detect anomalies, and adjust surgical plans in real-time, which is crucial in dynamic and unpredictable surgical settings. In recent years, especially in endoscopy and ophthalmic surgery, the application of deep learning has demonstrated considerable promise in bolstering these autonomous capabilities. This encompasses the analysis of surgical workflows [6, 28, 74], segmentation of instruments and anatomy [4, 24, 43], and depth estimation [78], among others.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automatic video surgical workflow understanding is a fundamental yet challenging problem for developing computer-assisted and robotic-assisted surgery, which can be divided into internal (e.g., laparoscopic and endoscopic [28,47,57]) and external (e.g., operating room and nursing procedure [41]) analysis. In addition to promoting the development of intelligent surgery, it also greatly benefits surgical documentation, education, and training [11,12,65]. Baret et al. [6] showed networks, like Inflated 3D ConvNet (I3D) [10] that utilize spatiotemporal convolutions, require a relatively extensive dataset for effective training. In their study, the model achieves an accuracy exceeding $80\\%$ when trained on 100 videos, with a progressive improvement as the sample size surpasses 700. However, the highly efficient and rapidly evolving deep learning technologies for surgical workflow analysis are currently limited by the following shortcomings in current video benchmarks: 1) Small-scale: the majority of video datasets contain no more than 100 videos. For example, the CATARACTS [23] and CatRelDet [23] datasets contain only 50 and 21 surgical videos, respectively. These datasets are relatively small, insufficient for large-scale validation. 2) Limited categories of surgeries and phases: almost all ophthalmic surgical video datasets only include cataract surgery and do not further classify specific types of surgeries. Additionally, the number of phase categories is also limited, like CatRelDet [23] only contains 4 different phase labels, which is insufficient to meet the requirements for evaluation in real clinical environments. 3) Coarse-grained annotation: due to annotation costs, existing benchmarks often have coarse-grained action definitions. For example, adhesive injection may occur in two different phases: main incision and capsulorhexis, so it may be classified into different phase categories. Coarse-grained action definitions may lead to annotation bias. 4) Single time-boundary annotation: they only annotate designated phases in the videos, ignoring the continuity across different stages of ophthalmic surgery, as well as the hierarchical relationship between surgery, phase, and operation. Simpler datasets, such as LensID [22], are limited to binary classification tasks distinguishing lens implantation from other irrelevant phases. 5) Uniform domain: the videos are meticulously collected, and", + "bbox": [ + 212, + 371, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/57a423041021a8a715cf693772b19cdd4fe79d23df36d6ff6a7c98321c8ec8d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ProtocolDataset PropertiesTasks
DatasetsNo. of VideosNo. of Action SegmentsNo. of Surgery CategoriesNo. of Action CategoriesTotal DurationMulti-Surgery Presence RecognitionPhase RecognitionPhase LocalizationPhase Prediction
Endo&LapCholec120 [45]120-1776.2hX✓XX
SurgicalActions160 [54]1601601160.2hX✓XX
HeiCo [39, 50]30-3142.8hX✓✓✓
EndoVis 2021 [67]332501722.0hX✓XX
PitVis [3]2528711733.3hX✓XX
CholecT50 [46]50-11044.7hX✓✓✓
AutoLaparo [70]213001723.1hX✓✓✓
OphScopeLensID [22]1002,4401211.7hX✓XX
Cataract-101 [55]1011,26611014.0hX✓✓✓
CatRelDet [23]212,400142.0hX✓XX
CATARACTS [4]501,53611920.0hX✓✓✓
Cataract-1K [21]1,000931112118.7hX✓✓✓
OphNet(Ours)2,2789,79566150284.8h✓✓✓✓
", + "bbox": [ + 222, + 143, + 785, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: The statistics comparison among existing workflow analysis datasets and our OphNet. Compared to other datasets, OphNet focuses on more comprehensive coverage of various surgery, phase and operation categories, collects a large number of videos, totaling 284.8 hours, and also enables a variety of recognition, localization and prediction tasks. OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. For instance, Cholec120 [45], Cholec80 [65], m2cai-workflow and LapChole [62] form one series, whereas CholecT50 [46], CholecT45 [46], and CholecT40 [44] comprise another series. We have excluded the following scenarios from our comparison: (1) non-open-source datasets such as Bypass170 [66], ESD [30], Yu's [74], etc.; (2) a superset of multiple open-source or non-open-source datasets, like Cholec207 [6], etc.; (3) datasets employed for lesion, anatomy, and instrument classification and segmentation, such as SUN-SEG [27], CVC-ClinicDB [7], ROBUSTMIS [52], Mesejo's [40], Cata7 [43], etc., anomaly detection such as PolypDiag [63] (from Hyper-Kvasir [9] and LDPolypVideo [38]), Kvasir-Capsule [59], etc., and other datasets not dedicated to workflow analysis. It's worth mentioning that even in comparison with the above datasets, OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. Endo&Lap denotes the endoscopic and laparoscopic protocol, OphScope denotes the ophthalmic microscope protocol. We choose the latest version for comparison in cases where datasets have multiple supplementary updates.", + "bbox": [ + 215, + 406, + 785, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while this ensures video quality, the uniform style is not conducive to testing the model's domain generalization ability.", + "bbox": [ + 215, + 715, + 784, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While some works have explored semi-supervised and self-supervised learning strategies [8, 51, 73, 75] to alleviate the cost of annotations or use only a small fraction of available labels, these approaches still lack competitiveness in performance compared to fully supervised learning. This deficiency in performance is impeding the widespread clinical application of these strategies. To address the shortage of sufficient labeled datasets, we construct OphNet, a large-scale", + "bbox": [ + 215, + 750, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and expert-level video benchmark with high diversity, for ophthalmic surgical workflow understanding. The main advantages of OphNet are as follows:", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Largest scale and diversity: to the best of our knowledge, OphNet is currently the largest and most richly labeled dataset for surgical workflow analysis. It contains a number of videos 20 times greater than the current largest benchmark in ophthalmic surgery and far exceeds datasets in more established fields such as endoscopy. Additionally, OphNet includes the greatest variety of different types of surgeries, encompassing 66 different surgeries such as cataract, glaucoma, and corneal surgeries, along with 102 unique surgical phases and 150 distinct operations. This diversity significantly surpasses that of previous research.", + "- Fine-grained, sequential and hierarchical annotation: we have meticulously selected a subset of videos for annotation localization, with each video being annotated for an average of 22 operations. Additionally, we provide exquisite annotations at the levels of surgery, phase, and operation, catering to the requirements for training specific challenge models. This annotation design aims to offer a multifaceted understanding of surgical protocols, accommodate the nuances of each distinct surgery, and enhance the usability and interpretability of our dataset.", + "- Expert-level manual annotation: the annotation work for OphNet was completed by ten experienced ophthalmologists and five individuals with ophthalmic experience, encompassing, but not limited to, standardization of definitions for surgery, phase, and operation labels, video filtering, classification and localization annotations, and secondary verification, among others. Expert-level annotators ensure the quality and professionalism of OphNet." + ], + "bbox": [ + 215, + 185, + 785, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 553, + 385, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surgical Workflow Understanding. Beyond its therapeutic advantages, minimally invasive surgery also provides the capability for operative video recording. These videos can be stored and later utilized for various purposes such as cognitive training, skill assessment, and surgical workflow analysis [76, 77]. Techniques derived from the broader field of video content analysis and representation are increasingly being incorporated into the surgical realm [6, 36, 74]. A typical surgical workflow can be defined by a sequence of tasks or events, including patient positioning, incision, dissection, and suturing. These events are influenced not only by the specific type of surgery but also by the individual surgeon's proficiency and technique. Consequently, a comprehensive understanding of the surgical workflow necessitates a thorough examination of the temporal, spatial, and contextual facets of these tasks.", + "bbox": [ + 215, + 583, + 785, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weakly-Supervised Video Learning. Substantial pioneering work has been undertaken in the realms of video understanding [5,16,33,53,69,71,72,79]. Since even a small amount of videos easily comprises several million frames, methods that do not rely on a frame-level annotation are of special importance. Weakly-supervised video learning makes use of loosely labeled data to train models,", + "bbox": [ + 215, + 763, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "thereby obviating the necessity for exhaustively annotated training data. These weak labels might manifest in a variety of forms, including video-level labels or partial labels, which are less specific than frame-level or pixel-level annotations. The objective of weakly-supervised learning is to utilize these coarse labels to produce models capable of delivering fine-grained predictions, such as temporally precise action localization or detailed semantic segmentation [15,35]. The large-scale annotation of surgical videos demands a significant investment of invaluable medical time resources. In this context, weakly supervised learning emerges as a viable solution to this bottleneck.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Dataset Construction", + "text_level": 1, + "bbox": [ + 214, + 306, + 460, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we detail the construction of our dataset, which involved meticulous data collection and preprocessing. YouTube was leveraged as a primary source to circumvent privacy issues while ensuring a broad representation of ophthalmic surgeries. Our selection criteria aimed to capture a wide range of video qualities and styles, specifically targeting cataract, glaucoma, and corneal surgeries due to their prevalence in clinical settings. We refined the dataset by excluding videos of inadequate quality or those depicting non-human subjects. The annotation process was designed to reflect the complex nature of eye surgeries, incorporating hierarchical classification to account for multiple conditions often treated within a single procedure. This was complemented by detailed localization annotations, delineating the distinct phases and techniques characteristic of ophthalmic operations, all undertaken by a dedicated team of ophthalmologists to ensure accuracy and relevance.", + "bbox": [ + 212, + 338, + 789, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Data Collection & Preprocessing", + "text_level": 1, + "bbox": [ + 214, + 556, + 532, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collection. Medical data exhibit unique privacy considerations and tend to be of smaller data volumes, characteristics that are particularly salient in the case of video data. By capitalizing on the wealth of surgical videos available on YouTube, we are able to obviate potential ethical and privacy concerns while concurrently enabling the rapid procurement of a substantial corpus of videos for further screening [1,2,17,29]. To fulfill this objective, we deploy text-based search algorithms to probe each surgery on YouTube, obtaining videos with titles that incorporate the requisite surgical keywords. We select cataract, glaucoma, and corneal surgery—three of the most commonly performed ophthalmic surgeries in actual clinical environments—as the central subjects of our research. To expand our video collection, we bolster our search queries by integrating synonyms and abbreviations associated with each type of ophthalmic surgery. For instance, cataract surgery encompasses various types: Phacoemulsification (abbr. PHACO), Intraocular Lens implantation (abbr. IOL), and Extracapsular Cataract Extraction (abbr. ECCE), etc.", + "bbox": [ + 212, + 583, + 787, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preprocessing. Given the nature of text-based retrieval and the varying quality, style, and filming methods of surgical videos influenced by different YouTube", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6e3ce91aebec06a16e145c80ecb0120a0c0fee542ad3cc25d3237b18b57910a9.jpg", + "image_caption": [ + "Fig. 1: The figure shows two combined surgical videos, PHACO + IOL implantation and PHACO + Trabeculotomy + Transscleral Cyclophotocoagulation + IOL implantation. For each frame marked in color, we provide time-boundary annotations at surgical, phase and operation levels." + ], + "image_footnote": [], + "bbox": [ + 217, + 140, + 782, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sources, some videos may exhibit poor quality or deviations in surgical representation. We filter out low-resolution videos, black-and-white color schemes, and animated demonstrations. Furthermore, surgical videos featuring non-human eyes, such as pig eyes, rabbit eyes, or pseudo-human eyes, are also outside the scope of our annotation. The data preprocessing was jointly completed by five professionals with experience in ophthalmology for initial screening, followed by a review conducted by six attending ophthalmologists.", + "bbox": [ + 212, + 535, + 787, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Data Annotation", + "text_level": 1, + "bbox": [ + 215, + 662, + 401, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classification Annotation. Different from the single-label classification in most natural video datasets [17, 25, 58, 60, 80], in ophthalmic surgery, physicians typically consider and implement multiple types of surgeries and phases based on the specific conditions and needs of the patient. This multifaceted approach stems from the complexity of the eye as an organ, where numerous diseases often coexist. Consequently, multiple ocular problems may need to be addressed in a single surgery. As shown in Fig. 1, a patient with cataracts may also have glaucoma, necessitating the simultaneous treatment of both conditions within one surgery. Hence, we initially classified all videos based on the primary surgical categories into three principal groups: cataract, glaucoma, and corneal", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b52aaf4761e229370e29937aee0afdec046a214687e9c8f1afaa73289bf3a5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 143, + 460, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcd339a165035909f0c64709c414835b5183a983fcd5fd796b93494db91028e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 143, + 782, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5382f905733380884d8f3a80587ad66136d6c32d2e538c9aacfc4404a4bef176.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 223, + 311, + 362, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6bbd279c72da86950d440012c84e0729166b52cc2ee02691d1627955ac4675d.jpg", + "image_caption": [ + "Capsulorhexis" + ], + "image_footnote": [], + "bbox": [ + 367, + 310, + 504, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f53fade19af4480e7439ab961a92878738e07a1a7ce302ccdcefd63bcf56a13c.jpg", + "image_caption": [ + "(b)", + "Capsule Staining" + ], + "image_footnote": [], + "bbox": [ + 511, + 310, + 648, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/384b32f6ce0db429a9014cfe1b54426b62f743fd98ea4ea28d59c0b97219757e.jpg", + "image_caption": [ + "Hydrodissection and Float Nucleus" + ], + "image_footnote": [], + "bbox": [ + 655, + 310, + 792, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2e6c9e169eeb878b093e02d2e3546f7f344313126e52b0a16e736a10ddf5ac8.jpg", + "image_caption": [ + "Main Incision", + "Cortical Aspiration", + "Fig. 2: OphNet's composition, comparison with other datasets for the same task, and some phase examples: (a) an overview of the composition ratios at the levels of surgery, phase, and operation; (b) comparison among existing open-source $\\odot$ laparoscopic & endoscopic, and $\\odot$ ophthalmic microscope workflow analysis video datasets and $\\odot$ our OphNet. OphNet stands as the largest real-world video dataset for ophthalmic surgical workflow understanding, featuring the highest number of videos, longest duration, and diverse categories of surgeries and phases; (c) eight phase examples in OphNet." + ], + "image_footnote": [], + "bbox": [ + 223, + 390, + 362, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3f077bd2d823d8e6b0fe2ed77678c81ae2141c8569afde47d4f9692296618af.jpg", + "image_caption": [ + "IOL Implantation", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 367, + 391, + 504, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fa9450d546ce9bf0f96787f22bf2db0bc9f820be34ec429d30591de819b263a.jpg", + "image_caption": [ + "Cautery/Hemostasis" + ], + "image_footnote": [], + "bbox": [ + 511, + 391, + 648, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5633365ecc0620f91407220347e404338336258036c2a686fd8414bddf14b1f2.jpg", + "image_caption": [ + "Phacoemulsification" + ], + "image_footnote": [], + "bbox": [ + 655, + 391, + 792, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "surgeries. Subsequently, these were further allocated for the detailed annotation of primary and secondary surgery. It is important to note that there exists only a single type of primary surgery, whereas multiple secondary surgery may be present. The categorization and annotation were meticulously completed by a team comprised of 8 experienced ophthalmologists.", + "bbox": [ + 212, + 625, + 787, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hierarchical Localization Annotation. A single surgery is often multifaceted, involving a series of intricate phases, each requiring distinct techniques and instruments, and the transition between surgeries entails high precision and coordination. Routine cataract surgery involves several steps. It begins with the administration of anesthesia, followed by a small incision in the cornea. The surgeon then creates an opening in the lens capsule and uses ultrasonic vibrations to break up and remove the cataract. Afterward, an artificial intraocular lens (IOL) is inserted into the lens capsule, and the incision is sealed without stitches. We define the phases and operation for various surgeries based on the", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be64c5f57da297c9fbce7bc783237ab0b093c29b47faf1b30bccdf2c10f1511c.jpg", + "image_caption": [ + "(a) phase statistics" + ], + "image_footnote": [], + "bbox": [ + 225, + 142, + 500, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b436e3316441290ec1749f9081749dcef4e6f24d32f53b15df7b8bb989cd5f09.jpg", + "image_caption": [ + "(b) operation statistics", + "Fig. 3: We present the data statistics of trimmed videos at the levels of phase and operation, including the number of trimmed videos, average duration, and total duration. The IDs and corresponding names can be found in the appendix." + ], + "image_footnote": [], + "bbox": [ + 504, + 142, + 784, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "textbook Ophthalmic Surgery: Principles and Practice [61]. To ensure the quality of localization annotations, we assign each annotator videos of 2 different types of primary surgeries. For each action time-boundary annotation, we annotate at three different levels of granularity: surgery, phase, and operation. We also employed the complete linkage algorithm [13] to cluster and merge various temporal boundaries into stable boundaries that received multiple agreements. It's worth noting that an individual video may have multiple separate instances of the same or different phases, thereby leading to multiple boundary definitions. A team of 15 ophthalmologists completes the localization annotation.", + "bbox": [ + 212, + 424, + 787, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Dataset Statistics and Analysis", + "text_level": 1, + "bbox": [ + 214, + 582, + 519, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OphNet includes 2,278 surgical videos (284.8 hours), demonstrating 66 different types of ophthalmic surgeries: 13 types of cataract surgery, 14 types of glaucoma surgery, and 39 types of corneal surgery. There are 102 phases and 150 operations for recognition, detection and prediction tasks, summarized in Tab. 1. Over $77\\%$ of videos have high-definition resolutions of $1280 \\times 720$ pixels or higher. To facilitate algorithm development and evaluation, we selected 523 videos for localization annotations. Additionally, we trim videos according to annotated action boundaries, resulting in 7,320 phase instances and 9,795 operation instances, totaling 51.2 hours. The average duration of trimmed videos is 32 seconds, while untrimmed videos average 337 seconds.", + "bbox": [ + 212, + 604, + 787, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 779, + 375, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our study, we explore four potential tasks using the OphNet dataset: 1) primary surgery presence recognition, 2) phase and operation recognition, 3) phase", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "localization, and 4) phase anticipation. To establish robust baselines, we employ state-of-the-art models known for their effectiveness in human action recognition, detection and anticipation. For each task, we provide a detailed problem formulation and evaluate the baseline models' performance. Our findings offer valuable insights into video understanding within sequences and fine granularity, contributing to the field's knowledge of video tasks in medical contexts.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Surgery Presence, Phase and Operation Recognition", + "text_level": 1, + "bbox": [ + 214, + 258, + 694, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Task Description. Surgery presence recognition focuses on identifying various surgical types in untrimmed videos through weakly supervised methods. This requires the model to discern and capture distinct surgical action features within extensive video footage. In OphNet, all types of surgeries occurring in each video are annotated, but only a subset of the videos have time-boundary annotations, as detailed in Sec. 3.2. We identify the primary surgery based on the key objectives and durations of the surgeries. To streamline the process, our experiments are limited to recognizing the presence of these primary surgeries. Additionally, phase recognition segments the surgery into distinct phases using visual cues and movements, such as incision, lens removal, and implantation. Operation recognition involves identifying finer-grained surgical actions.", + "bbox": [ + 212, + 281, + 787, + 446 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselins. We compare the performance of I3D [10], SlowFast [19], X3D [18], and MViT V2 [31] models on this task. These models are evaluated in two versions: 1) random initialization training and 2) pretraining with weights from Kinetics 400 [29], which is a human action recognition dataset. In addition, we also explored the classification performance of X-CLIP [42] and ViFi-CLIP [49], two CLIP [48]-based models. We also compared the effects of different numbers of input frames on the models' performance, where the sub", + "bbox": [ + 212, + 448, + 444, + 719 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4c9c80c9bad3524905c6802bcbca78f7c080fde83c86f2dc37fd80cb46aa7fb5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselinesPrimary Surgery Classification
CataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
I3D [10]38.986.343.671.742.778.229.853.2
SlowFast [19]45.382.144.672.345.577.327.254.4
X3D [31]42.187.444.674.243.876.628.562.7
MViT V2 [18]43.284.245.581.845.575.929.160.1
I3D [10]36.884.748.281.848.676.527.250.6
SlowFast*49.083.747.380.949.275.627.250.6
X3D*47.486.346.481.548.378.235.461.4
MViT V2*44.285.349.181.847.777.328.563.3
X-CLIP16 [42]58.594.751.892.861.488.640.579.0
X-CLIP3260.692.653.583.756.884.158.981.0
ViFi-CLIP16 [49]59.688.352.480.261.481.858.979.8
ViFi-CLIP3259.688.351.584.850.075.056.377.2
", + "bbox": [ + 457, + 478, + 787, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Per-class Top-1 and Top-5 accuracy (%) for the primary surgery presence recognition on untrimmed videos. The best performance for each split has been highlighted in bold.", + "bbox": [ + 455, + 630, + 787, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "script _16 represents an input of 16 frames, and _32 represents an input of 32 frames.", + "bbox": [ + 212, + 719, + 787, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Setup. The dataset was randomly partitioned to ensure a balanced representation of examples for each surgery category. Specifically, we allocated $70\\%$ of the data for training (1,449 surgical videos), $10\\%$ for validation (205 surgical videos), and $20\\%$ for testing (424 surgical videos). For the phase and operation recognition experiments, we followed the same settings, with $70\\%$ of the data used for training (5,024 phase segments, 6,856 operations segments), $10\\%$ for", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "validation (730 phase segments, 975 operations segments), and $20\\%$ for testing (1,566 phase segments, 1,964 operations segments). The input for the surgery presence recognition experiment is untrimmed videos, while for the phase and operation recognition experiments, the input consists of trimmed segments. In all classification experiments, we set up the analysis and comparison from four perspectives: cataract surgery, glaucoma surgery, corneal surgery, and all surgical videos.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. We summarize the results in Tab. 2 and Tab. 3. In primary surgery classification, X-CLIP [42] achieved the highest overall Top-1 accuracy at $58.9\\%$ , leading in cataract and glaucoma surgeries with accuracies of $60.6\\%$ and $53.5\\%$ respectively. For corneal surgeries, X-CLIP also recorded the highest Top-1 and Top-5 accuracies of $61.4\\%$ and $88.6\\%$ . In phase classification, ViFi-CLIP [49] led with the highest Top-1 accuracy, particularly in cataract surgeries at $75.9\\%$ , and exhibited the best performance in both glaucoma and corneal surgeries. Furthermore, in operation classification, ViFi-CLIP outperformed all other models in all categories, especially noted in cataract and corneal surgeries with Top-1 accuracies of $75.1\\%$ and $83.7\\%$ and Top-5 accuracies of $93.8\\%$ and $85.2\\%$ respectively. Overall, ViFi-CLIP showed superior performance in both phase and operation classifications across various surgery types. Besides, in the phase and operation classification experiments, a higher number of input frames generally had a positive effect on the model. In Fig. 4, we present the heatmap visualization of four examples from the test set of phase recognition experiments using the ViFi-CLIP model. It can be observed that in a series of frame images, the model focuses on surgical instruments and the operated eye area, which is consistent with human experience. The instruments used in the same phase of ophthalmic surgery are often similar.", + "bbox": [ + 212, + 252, + 787, + 540 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Phase Localization", + "text_level": 1, + "bbox": [ + 215, + 566, + 415, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Task Description. Phase localization in ophthalmic surgical workflow analysis refers to the task of pinpointing the exact moments or time intervals within a surgical video where specific phases of the surgery begin and end. This involves the detailed temporal segmentation of the entire surgical procedure into its constituent phases based on visual cues, surgeon's actions, and the progression of the surgery. The objective of phase localization is to accurately identify the start and end times of different surgical stages, such as pre-operative preparation, incision, and removal of the lens facilitating a granular and precise understanding of the surgery timeline. This task is crucial for detailed surgical documentation, efficient surgical training, and the development of targeted interventions during specific stages of the surgery, enhancing overall surgical management and post-operative analysis.", + "bbox": [ + 212, + 597, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We conducted the experiments using phase-level labels and used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. Data split follows the setup of the primary surgery classification experiment.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bb89199ea3128527ce0316940d07464f1db6d1af112ee3db53f6ec72cd04d6d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselinesPhase ClassificationOperation Classification
CataractGlaucomaCorneaAllCataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1
I3D [10]27.255.724.157.518.952.125.758.226.854.923.556.018.051.225.057.1
SlowFast [19]26.556.523.156.524.249.126.760.125.855.222.945.923.548.526.059.0
X3D [31]27.058.321.055.521.828.526.662.326.447.220.544.621.327.826.161.5
MViT V2 [18]26.254.921.053.426.046.827.059.825.943.820.542.725.545.926.558.9
I3D*29.568.922.158.626.050.930.271.228.867.521.847.925.750.329.560.0
SlowFast*30.672.325.254.730.759.831.761.829.971.124.843.929.558.730.560.9
X3D*27.272.922.159.630.761.533.563.226.571.821.748.829.960.232.862.1
MViT V2*34.276.523.352.038.465.128.360.233.575.222.841.537.964.027.859.5
X-CLIP16 [42]68.392.247.389.853.077.463.485.367.591.046.578.982.276.162.584.0
X-CLIP3269.194.048.781.754.880.462.785.868.093.047.980.584.079.562.084.7
ViFi-CLIP16 [49]75.993.740.485.466.681.666.188.474.592.542.874.585.080.565.087.5
ViFi-CLIP3273.092.949.682.757.781.668.487.275.193.843.280.283.785.264.886.5
", + "bbox": [ + 218, + 143, + 787, + 303 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2b51ed963e3d1a99542736868cff811855359d1775a305368c08ac93a57693a0.jpg", + "table_caption": [ + "Table 3: Per-class Top-1 and Top-5 accuracy (\\%) for the primary surgery presence recognition on untrimmed videos and phase recognition on trimmed videos. * denotes the initialization from the model pre-trained on Kinetics 400 [29]. For the two CLIP models, we chose ViT-B/16 as the backbone and compared the performance of two different input frame numbers, 16 and 32. The best performance for each split has been highlighted in bold." + ], + "table_footnote": [ + "Table 4: The results for phase detection. Action-Former and TriDet are state-of-the-art models for human action detection tasks, and we use three different backbones for feature extraction and report mAP at the IoU thresholds of [0.1:0.2:0.9]. Average mAP is computed by averaging different IoU thresholds. The best performance for each split has been highlighted in bold." + ], + "table_body": "
BaselinesBackbonesmAP (%)
0.10.30.50.7Avg.
ActionFormerCSN [64]53.750.140.624.542.5
SwinViviT [34]59.354.743.326.346.4
SlowFast [19]60.055.945.126.047.5
TriDet [56]CSN56.153.043.129.446.2
SwinViviT61.057.147.133.150.4
SlowFast61.356.045.630.448.6
", + "bbox": [ + 215, + 430, + 555, + 537 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/91ccd9df8db4b79a6a78d0f3ba289c5a04b99f4d73038fd0c7d1f67691816b0c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5: The results for phase anticipation. We report top-1 accuracy at the observation ratios [0.1:0.2:0.9]. Average top-1 accuracy is computed by averaging different observation ratios. The best performance for each split has been highlighted in bold." + ], + "table_body": "
BaselinesTop-1 Acc. (%)
0.10.30.50.7Avg.
I3D [10]26.542.249.851.347.3
SlowFast [19]25.442.648.952.247.2
MViT V2 [18]25.643.749.352.347.5
I3D*27.343.550.151.447.6
SlowFast*27.543.249.952.347.8
MViT V2*27.843.850.551.748.2
", + "bbox": [ + 560, + 430, + 787, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Setup. We excluded Operation Gap and Invalid, and filtered out tags with fewer than 20 segments. We used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. To extract features from the videos, we first extracted RGB frames from each video at a rate of 25 frames per second. We also extracted optical flow using the TV-L1 [26,37] algorithm. We then fine-tuned an I3D [10] model that had been pre-trained on the ImageNet [14] dataset, and used it to generate features for each RGB and optical flow frame. Because each video has a variable duration, we performed uniform interpolation to generate 100 fixed-", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9c623f3ddfa6d3398ad7f45a6022f8c006b498813324c54bfce7ac270ce7be7b.jpg", + "image_caption": [ + "Fig. 4: Attention map visualizations of ViFi-CLIP [49] on four examples from OphNet's test set in the phase recognition task." + ], + "image_footnote": [], + "bbox": [ + 217, + 140, + 500, + 316 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/59a39a3f1b6cd39650adff69bb5d387e3de20e500fa779540ce4ad29512e2b1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 140, + 784, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "length features for each video. Finally, we concatenated the RGB and optical flow features into a 2048-dimensional embedding, which served as the input for our model.", + "bbox": [ + 212, + 385, + 784, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. The experimental results for phase localization are presented in the Tab. 4, showcasing the performance of different baseline models with various backbones in terms of mean Average Precision (mAP) at different Intersection over Union (IoU) thresholds [0.1:0.2:0.9]. The models evaluated include Action-Former with SwinViviT and SlowFast backbones, and TriDet with CSN, SwinViviT, and SlowFast backbones. The results indicate that the TriDet model with a SwinViviT backbone outperforms other combinations, achieving the highest mAP scores across most IoU thresholds, with notable scores of $61.0\\%$ (IoU=0.1), $57.1\\%$ (IoU=0.3), $47.1\\%$ (IoU=0.5), and $33.1\\%$ (IoU=0.7), resulting in an average mAP of $50.4\\%$ . This indicates that the TriDet model, especially when combined with the SwinViviT backbone, is particularly effective for phase localization in surgical videos. On the other hand, the TriDet model with a SlowFast backbone shows competitive performance, particularly achieving the highest mAP of $61.3\\%$ at the lowest IoU threshold (0.1). However, it falls slightly behind in performance at higher IoU thresholds compared to the SwinViviT backbone.", + "bbox": [ + 212, + 430, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Phase Anticipation", + "text_level": 1, + "bbox": [ + 215, + 679, + 419, + 694 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Task Description. This task requires the analysis of real-time or recorded video data to foresee the sequence of events based on current and past surgical activities. By understanding the typical progression of ophthalmic surgeries and recognizing patterns in the surgeon's actions and the use of instruments, the system aims to forecast the next phase of the surgery, allowing for proactive preparation and response. The objective of phase anticipation is to enhance the efficiency and safety of surgical procedures by providing the surgical team with advanced notice of upcoming steps, enabling better resource allocation, timing for critical tasks, and overall coordination within the operating room.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/57d2cb94b9d2a79c200a69e5e202df35fe28033b47254634482f624b18db31f8.jpg", + "image_caption": [ + "Fig. 5: Phase localization visualization of TriDet [56]. $GT$ represents the ground truth visualization for phases, while $Loc$ . visualizes the model's highest confidence phase category and the time-boundary results. Blank segments denote invalid segments or operation gaps." + ], + "image_footnote": [], + "bbox": [ + 217, + 140, + 495, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d476f929fee901f85ba529cf22f6620032dad1799437c48991626aebcd816a0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 140, + 784, + 220 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Setup. Following previous approaches of the primary surgery classification experiment in Sec. 4.1, We randomly mask phase sequences in the test video with different observation ratios.", + "bbox": [ + 212, + 316, + 782, + 359 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Baselines. We evaluate our datasets with the baseline models such as I3D [10], SlowFast [19], and MViT V2 [31]. For each model, we also adopted two training approaches: random initialization training and using pre-trained weights from Kinetics 400 [29].", + "bbox": [ + 212, + 362, + 784, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. The phase detection results are illustrated in Tab. 5. The results demonstrate that the baseline models pretrained on Kinetics 400 [29] generally outperform their original counterparts in terms of Top-1 accuracy across different observation ratios for phase anticipation. Specifically, the modified MViT V2* model exhibits the highest improvement, achieving the best average Top-1 accuracy of $48.2\\%$ . Moreover, while all models show increased accuracy with higher observation ratios, indicating that more observed data contributes to better performance, the consistent improvement across all ratios for the enhanced models suggests effective modifications.", + "bbox": [ + 212, + 422, + 787, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 582, + 362, + 597 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Dataset Bias. OphNet's videos are sourced from YouTube and exhibit diverse styles, clarity, and screen elements. This diversity can aid detection models in generalization but may affect their effectiveness and performance. Some videos in the dataset include subtitles or additional video windows, such as a little subtitle or watermark shown in Fig. 2. Similarly, additional video windows offer another perspective but can make the scene chaotic, making it harder to recognize primary surgical actions. The presence of these factors in OphNet reflects the complexity of real-world surgical environments, because an ophthalmic microscope may inherently display different windows or show parameters during recording. While they pose challenges, they also present opportunities for developing models that can better handle variability and unpredictability, which are crucial aspects of real-world surgical scenarios.", + "bbox": [ + 212, + 613, + 787, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Annotation Bias. OphNet is entirely annotated by ophthalmologists, there is a distinct possibility of annotation bias reflecting specific regional practices, terminologies, and interpretations. Despite the universal nature of many ophthalmic", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "procedures, subtle differences in surgical techniques, procedural preferences, and clinical terminologies could lead to inconsistencies in how surgeries are categorized and described across different regions. For instance, the terminology used to describe certain procedures might differ, with one region referring to a procedure as anterior vitrectomy while another uses pars plana vitrectomy. To reduce the possibility of biases in precise annotations, we have taken great care to establish a unified definition prior to describing the surgery, phase and operation. However, potential biases arising from regional variations and individual surgical practices are inevitable.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 305, + 357, + 321 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In response to the current challenges in ophthalmology, a surgical field apt for automation and remote control, we introduce OphNet, a large-scale, diverse, and expert-level video benchmark for understanding ophthalmic surgical workflows. OphNet is the most extensive dataset of its kind, containing a broad range of cataract, glaucoma, and corneal surgeries and detailed annotations for distinct surgical phases. OphNet comprises 2,278 surgical videos (284.8 hours), 7,320 phase segments and 9,795 operation segments (51.2 hours), showcasing 66 different types of ophthalmic surgeries: 13 cataract, 14 glaucoma, and 39 corneal. It is annotated with 102 phases and 150 operations. With OphNet, we explored primary surgery presence recognition, phase localization and phase anticipation on untrimmed videos, phase and operation recognition on trimmed videos. We employed state-of-the-art models to establish robust baselines and provided valuable insights into video understanding within sequences and fine granularity. Our work contributes to the broader understanding of surgical video tasks in medical contexts and promotes the integration of deep learning technologies into ophthalmic surgical procedures.", + "bbox": [ + 212, + 335, + 787, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 601, + 323, + 617 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Fair use on youtube. https://support.google.com/youtube/answer/9783148?hl=en#:~:text=If%20the%20use%20of%20copyright, copyright %20removal%20request%20to%20YouTube.", + "2. Youtube's copyright exception policy. https://www.youtube.com/howyoutubeworks/policies/copyright/#copyright-exceptions", + "3. Adrito, D., Danyal, Z.K., Dimitrios, P., Yitong, Z., John, G.H., Francisco, V., Sophia, B., Hani, J.M., Danail, S.: Pitvis: Workflow recognition in endoscopic pituitary surgery", + "4. Al Hajj, H., Lamard, M., Conze, P.H., Roychowdhury, S., Hu, X., Marsalkaite, G., Zisimopoulos, O., Dedmari, M.A., Zhao, F., Prellberg, J., Sahu, M., Galdran, A., Araujo, T., Vo, D.M., Panda, C., Dahiya, N., Kondo, S., Bian, Z., Vahdat, A., Bialopetravicius, J., Flouty, E., Qiu, C., Dill, S., Mukhopadhyay, A., Costa, P., Aresta, G., Ramamurthy, S., Lee, S.W., Campilho, A., Zachow, S., Xia, S., Conjeti, S., Stoyanov, D., Armaitis, J., Heng, P.A., Macready, W.G., Cochener, B., Quellec, G.: Cataracts: Challenge on automatic tool annotation for cataract" + ], + "bbox": [ + 222, + 631, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "surgery. Medical Image Analysis 52, 24-41 (2019). https://doi.org/https://doi.org/10.1016/j.media.2018.11.008, https://www.sciencedirect.com/science/article/pii/S136184151830865X", + "5. Alwassel, H., Giancola, S., Ghanem, B.: TSP: Temporally-sensitive pretraining of video encoders for localization tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 3173-3183 (2021)", + "6. Bar, O., Neimark, D., Zohar, M., Hager, G.D., Girshick, R., Fried, G.M., Wolf, T., Asselmann, D.: Impact of data on generalization of ai for surgical intelligence applications. Scientific Reports 10(1), 22208 (Dec 2020). https://doi.org/10.1038/s41598-020-79173-6, https://doi.org/10.1038/s41598-020-79173-6", + "7. Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilarino, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics 43, 99-111 (2015). https://doi.org/https://doi.org/10.1016/j.compmedimag.2015.02.007, https://www.sciencedirect.com/science/article/pii/S0895611115000567", + "8. Bodenstedt, S., Wagner, M., Katic, D., Mietkowski, P., Mayer, B., Kenngott, H., Müller-Stich, B., Dillmann, R., Speidel, S.: Unsupervised temporal context learning using convolutional neural networks for laparoscopic workflow analysis. arXiv preprint arXiv:1702.03684 (2017)", + "9. Borgli, H., Thambawita, V., Smedsrud, P.H., Hicks, S., Jha, D., Eskeland, S.L., Randel, K.R., Pogorelov, K., Lux, M., Nguyen, D.T.D., Johansen, D., Griwodz, C., Stensland, H.K., Garcia-Ceja, E., Schmidt, P.T., Hammer, H.L., Riegler, M.A., Halvorsen, P., de Lange, T.: HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific Data 7(1), 283 (2020). https://doi.org/10.1038/s41597-020-00622-y, https://doi.org/10.1038/s41597-020-00622-y", + "10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "11. Czempiel, T., Paschali, M., Keicher, M., Simson, W., Feussner, H., Kim, S.T., Navab, N.: Tecn: Surgical phase recognition with multi-stage temporal convolutional networks. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoeanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 343-352. Springer International Publishing, Cham (2020)", + "12. Czempiel, T., Paschali, M., Ostler, D., Kim, S.T., Busam, B., Navab, N.: Opera: Attention-regularized transformers for surgical phase recognition. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part IV 24. pp. 604–614. Springer (2021)", + "13. Defays, D.: An efficient algorithm for a complete link method. The Computer Journal 20(4), 364-366 (01 1977). https://doi.org/10.1093/comjnl/20.4.364, https://doi.org/10.1093/comjnl/20.4.364", + "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "15. Dong, S., Hu, H., Lian, D., Luo, W., Qian, Y., Gao, S.: Weakly supervised video representation learning with unaligned text for sequential videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2447 (2023)" + ], + "bbox": [ + 218, + 146, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Duong, H.T., Le, V.T., Hoang, V.T.: Deep learning-based anomaly detection in video surveillance: A survey. Sensors 23(11) (2023). https://doi.org/10.3390/s23115024, https://www.mdpi.com/1424-8220/23/11/5024", + "17. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015)", + "18. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)", + "19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)", + "20. Forslund Jacobsen, M., Konge, L., Alberti, M., la Cour, M., Park, Y.S., Thomsen, A.S.S.: ROBOT-ASSISTED VITREORETINAL SURGERY IMPROVES SURGICAL ACCURACY COMPARED WITH MANUAL SURGERY: A randomized trial in a simulated setting. Retina 40(11), 2091-2098 (Nov 2020)", + "21. Ghamsarian, N., El-Shabrawi, Y., Nasirihaghighi, S., Putzgruber-Adamitsch, D., Zinkernagel, M., Wolf, S., Schoeffmann, K., Sznitman, R.: Cataract-1k: Cataract surgery dataset for scene segmentation, phase recognition, and irregularity detection. arXiv preprint arXiv:2312.06295 (2023)", + "22. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., El-Shabrawi, Y., Schoeffmann, K.: Lensid: A cnn-rnn-based framework towards lens irregularity detection in cataract surgery videos. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part VIII. Lecture Notes in Computer Science, vol. 12908, pp. 76-86. Springer (2021). https://doi.org/10.1007/978-3-030-87237-3_8, https://doi.org/10.1007/978-3-030-87237-3_8", + "23. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., Schoeffmann, K.: Relevance detection in cataract surgery videos by spatio-temporal action localization. In: 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021. pp. 10720-10727. IEEE (2020). https://doi.org/10.1109/ICPR48806.2021.9412525, https://doi.org/10.1109/ICPR48806.2021.9412525", + "24. Grammatikopoulou, M., Flouty, E., Kadkhodamohammadi, A., Quellec, G., Chow, A., Nehme, J., Luengo, I., Stoyanov, D.: Cadis: Cataract dataset for image segmentation. arXiv preprint arXiv:1906.11586 (2019)", + "25. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: Ava: A video dataset of spatio-temporally localized atomic visual actions (2018)", + "26. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981)", + "27. Ji, G.P., Xiao, G., Chou, Y.C., Fan, D.P., Zhao, K., Chen, G., Van Gool, L.: Video polyp segmentation: A deep learning perspective. Machine Intelligence Research 19(6), 531-549 (Dec 2022). https://doi.org/10.1007/s11633-022-1371-y, https://doi.org/10.1007/s11633-022-1371-y", + "28. Jin, Y., Dou, Q., Chen, H., Yu, L., Qin, J., Fu, C.W., Heng, P.A.: SV-RCNet: Workflow recognition from surgical videos using recurrent convolutional network. IEEE Trans. Med. Imaging 37(5), 1114-1126 (May 2018)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)", + "30. Li, J., Jin, Y., Chen, Y., Yip, H.C., Scheppach, M., Chiu, P.W.Y., Yam, Y., Meng, H.M.L., Dou, Q.: Imitation learning from expert video data for dissection trajectory prediction in endoscopic surgical procedure. In: Greenspan, H., Madabhushi, A., Mousavi, P., Salcudean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2023. pp. 494-504. Springer Nature Switzerland, Cham (2023)", + "31. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022)", + "32. Lin, S., Miao, A.J., Lu, J., Yu, S., Chiu, Z.Y., Richter, F., Yip, M.C.: Semantic-super: A semantic-aware surgical perception framework for endoscopic tissue identification, reconstruction, and tracking. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4739-4746 (2023). https://doi.org/10.1109/ICRA48891.2023.10160746", + "33. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3889-3898 (2019)", + "34. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. arXiv preprint arXiv:2106.13230 (2021)", + "35. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Bi-calibration networks for weakly-supervised video representation learning. International Journal of Computer Vision 131(7), 1704-1721 (Jul 2023). https://doi.org/10.1007/s11263-023-01779-w, https://doi.org/10.1007/s11263-023-01779-w", + "36. Loukas, C.: Video content analysis of surgical procedures. Surgical Endoscopy 32(2), 553-568 (Feb 2018). https://doi.org/10.1007/s00464-017-5878-1, https://doi.org/10.1007/s00464-017-5878-1", + "37. Lucas, B.D., Kanade, T., et al.: An iterative image registration technique with an application to stereo vision, vol. 81. Vancouver (1981)", + "38. Ma, Y., Chen, X., Cheng, K., Li, Y., Sun, B.: Ldpolypvideo benchmark: A large-scale colonoscopy video dataset of diverse polyps. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021. pp. 387-396. Springer International Publishing, Cham (2021)", + "39. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific data 8(1), 101 (2021)", + "40. Mesejo, P., Pizarro, D., Abergel, A., Rouquette, O., Beorchia, S., Poincloux, L., Bartoli, A.: Computer-aided classification of gastrointestinal lesions in regular colonoscopy. IEEE Transactions on Medical Imaging 35(9), 2051-2063 (2016). https://doi.org/10.1109/TMI.2016.2547947", + "41. Ming, H., Lin, W., Siyuan, Y., Don, M., Qingli, R., Peng, X., Wei, F., Peibo, D., Lie, J., Zongyuan, G.: Nurvid: A large expert-level video database for nursing procedure activity understanding. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)", + "42. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Ni, Z.L., Bian, G.B., Zhou, X.H., Hou, Z.G., Xie, X.L., Wang, C., Zhou, Y.J., Li, R.Q., Li, Z.: Raunet: Residual attention u-net for semantic segmentation of cataract surgical instruments. In: International Conference on Neural Information Processing (ICONIP). pp. 139-149. Springer (2019)", + "44. Nwoye, C.I., Gonzalez, C., Yu, T., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23. pp. 364-374. Springer (2020)", + "45. Nwoye, C.I., Padoy, N.: Data splits and metrics for benchmarking methods on surgical action triplet datasets. arXiv preprint arXiv:2204.05235 (2022)", + "46. Nwoye, C.I., Yu, T., Gonzalez, C., Seeliger, B., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. Medical Image Analysis 78, 102433 (2022)", + "47. Pan, X., Gao, X., Wang, H., Zhang, W., Mu, Y., He, X.: Temporal-based swim transformer network for workflow recognition of surgical video. Int. J. Comput. Assist. Radiol. Surg. 18(1), 139-147 (Jan 2023)", + "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "49. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "50. Ross, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Comparative validation of multi-instance instrument segmentation in endoscopy: results of the robust-mis 2019 challenge. Medical image analysis 70, 101920 (2021)", + "51. Ross, T., Zimmerer, D., Vemuri, A., Isensee, F., Wiesenfarth, M., Bodenstedt, S., Both, F., Kessler, P., Wagner, M., Müller, B., et al.: Exploiting the potential of unlabeled endoscopic video data with self-supervised learning. International journal of computer assisted radiology and surgery 13, 925-933 (2018)", + "52. Rof, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Arbelaez, P., Bian, G.B., Bodenstedt, S., Bolmgren, J.L., Bravo-Sanchez, L., Chen, H.B., Gonzalez, C., Guo, D., Halvorsen, P., Heng, P.A., Hosgor, E., Hou, Z.G., Isensee, F., Jha, D., Jiang, T., Jin, Y., Kirtac, K., Kletz, S., Leger, S., Li, Z., Maier-Hein, K.H., Ni, Z.L., Riegler, M.A., Schoeffmann, K., Shi, R., Speidel, S., Stenzel, M., Twick, I., Wang, G., Wang, J., Wang, L., Wang, L., Zhang, Y., Zhou, Y.J., Zhu, L., Wiesenfarth, M., Kopp-Schneider, A., Müller-Stich, B.P., Maier-Hein, L.: Comparative validation of multi-instance instrument segmentation in endoscopy: Results of the robust-mis 2019 challenge. Medical Image Analysis 70, 101920 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101920, https://www.sciencedirect.com/science/article/pii/S136184152030284X", + "53. Sato, F., Hachiuma, R., Sekii, T.: Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6471-6480 (2023)", + "54. Schoeffmann, K., Husslein, H., Kletz, S., Petscharnig, S., Munzer, B., Beecks, C.: Video retrieval in laparoscopic video recordings with dynamic content descriptors." + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Multim. Tools Appl. 77(13), 16813-16832 (2018). https://doi.org/10.1007/s11042-017-5252-2, https://doi.org/10.1007/s11042-017-5252-2", + "55. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: César, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018). https://doi.org/10.1145/3204949.3208137, https://doi.org/10.1145/3204949.3208137", + "56. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18857-18866 (2023)", + "57. Shi, X., Jin, Y., Dou, Q., Heng, P.A.: LRTD: long-range temporal dependency based active learning for surgical workflow recognition. Int. J. Comput. Assist. Radiol. Surg. 15(9), 1573-1584 (Sep 2020)", + "58. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 510-526. Springer International Publishing, Cham (2016)", + "59. Smedsrud, P.H., Thambawita, V., Hicks, S.A., Gjestang, H., Nedrejord, O.O., Naess, E., Borgli, H., Jha, D., Berstad, T.J.D., Eskeland, S.L., Lux, M., Espeland, H., Petlund, A., Nguyen, D.T.D., Garcia-Ceja, E., Johansen, D., Schmidt, P.T., Toth, E., Hammer, H.L., de Lange, T., Riegler, M.A., Halvorsen, P.: Kvasir-Capsule, a video capsule endoscopy dataset. Scientific Data 8(1), 142 (2021). https://doi.org/10.1038/s41597-021-00920-z", + "60. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012)", + "61. Spaeth, G., Danesh-Meyer, H., Goldberg, I., Kampik, A.: Ophthalmic Surgery: Principles and Practice E-Book. Elsevier Health Sciences (2011), https://books.google.com.hk/books?id=wHWMUGH-5csC", + "62. Stauder, R., Ostler, D., Kranzfelder, M., Koller, S., Feußner, H., Navab, N.: The tum lapchole dataset for the m2cai 2016 workflow challenge. arXiv preprint arXiv:1610.09278 (2016)", + "63. Tian, Y., Pang, G., Liu, F., Liu, Y., Wang, C., Chen, Y., Verjans, J., Carneiro, G.: Contrastive transformer-based multiple instance learning for weakly supervised polyp frame detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 88-98. Springer (2022)", + "64. Tran, D., Wang, H., Torresani, L., Feiszli, M.: Video classification with channel-separated convolutional networks (2019)", + "65. Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., de Mathelin, M., Padoy, N.: Endonet: A deep architecture for recognition tasks on laparoscopic videos. IEEE Transactions on Medical Imaging (TMI) 36, 86-97 (2016), https://api-semanticscholar.org/CorpusID:5633749", + "66. Twinanda, A.P., Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Rsdnet: Learning to predict remaining surgery duration from laparoscopic videos without manual annotations. IEEE Transactions on Medical Imaging 38(4), 1069-1078 (2019). https://doi.org/10.1109/TMI.2018.2878055", + "67. Wagner, M., Müller-Stich, B.P., Kisilenko, A., Tran, D., Heger, P., Mündermann, L., Lubotsky, D.M., Müller, B., Davitashvili, T., Capek, M., et al.: Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis 86, 102770 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "68. Wang, T., Li, H., Pu, T., Yang, L.: Microsurgery robots: Applications, design, and development. Sensors 23(20) (2023). https://doi.org/10.3390/s23208503, https://www.mdpi.com/1424-8220/23/20/8503", + "69. Wang, X., Zhang, S., Qing, Z., Shao, Y., Gao, C., Sang, N.: Self-supervised learning for semi-supervised temporal action proposal. In: CVPR (2021)", + "70. Wang, Z., Lu, B., Long, Y., Zhong, F., Cheung, T.H., Dou, Q., Liu, Y.: Autolaparo: A new dataset of integrated multi-tasks for image-guided surgical automation in laparoscopic hysterectomy. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 486-496. Springer (2022)", + "71. Wu, W., Sun, Z., Ouyang, W.: Revisiting classifier: Transferring vision-language models for video recognition (2023)", + "72. Wu, W., Wang, X., Luo, H., Wang, J., Yang, Y., Ouyang, W.: Bidirectional cross-modal knowledge exploration for video recognition with pre-trained vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "73. Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Less is more: Surgical phase recognition with less annotations through self-supervised pre-training of cnn-lstm networks. arXiv preprint arXiv:1805.08569 (2018)", + "74. Yu, F., Silva Croso, G., Kim, T.S., Song, Z., Parker, F., Hager, G.D., Reiter, A., Vedula, S.S., Ali, H., Sikder, S.: Assessment of Automated Identification of Phases in Videos of Cataract Surgery Using Machine Learning and Deep Learning Techniques. JAMA Network Open 2(4), e191860-e191860 (04 2019). https://doi.org/10.1001/jamanetworkopen.2019.1860, https://doi.org/10.1001/jamanetworkopen.2019.1860", + "75. Yu, T., Mutter, D., Marescaux, J., Padoy, N.: Learning from a tiny dataset of manual annotations: a teacher/student approach for surgical phase recognition. arXiv preprint arXiv:1812.00033 (2018)", + "76. Yuan, K., Srivastav, V., Navab, N., Padoy, N.: Hecvl: Hierarchical videolanguage pretraining for zero-shot surgical phase recognition. arXiv preprint arXiv:2405.10075 (2024)", + "77. Yuan, K., Srivastav, V., Yu, T., Lavanchy, J., Mascagni, P., Navab, N., Padoy, N.: Learning multi-modal representations by watching hundreds of surgical video lectures (2023)", + "78. Zha, R., Cheng, X., Li, H., Harandi, M., Ge, Z.: Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 13-23. Springer (2023)", + "79. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: European Conference on Computer Vision. LNCS, vol. 13664, pp. 492-510 (2022)", + "80. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI Conference on Artificial Intelligence. pp. 7590-7598 (2018), https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/17344" + ], + "bbox": [ + 215, + 146, + 784, + 743 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "M. Hu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_model.json b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6d3bb42feee99ae4a67ba32fa981815a6ade79e3 --- /dev/null +++ b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_model.json @@ -0,0 +1,2616 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "OphNet: A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.786, + 0.26 + ], + "angle": 0, + "content": "Ming Hu\\(^{1,2,3*}\\), Peng Xia\\(^{1,3*}\\), Lin Wang\\(^{5*}\\), Siyuan Yan\\(^{1,2}\\), Feilong Tang\\(^{1,3}\\) Zhongxing Xu\\(^{7}\\), Yimin Luo\\(^{6}\\), Kaimin Song\\(^{3}\\), Jurgen Leitner\\(^{2}\\) Xuelian Cheng\\(^{1}\\), Jun Cheng\\(^{8}\\), Chi Liu\\(^{9}\\), Kaijing Zhou\\(^{4\\dagger}\\), and Zongyuan Ge\\(^{1,2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.269, + 0.772, + 0.341 + ], + "angle": 0, + "content": "\\(^{1}\\)AIM Lab, Faculty of IT, \\(^{2}\\)Faculty of Engineering, Monash University \n\\(^{3}\\)Airdoc-Monash Research, Airdoc \\(^{4}\\)Eye Hospital, Wenzhou Medical University \n\\(^{5}\\)Bosch Corporate Research \\(^{6}\\)King's College London \\(^{7}\\)Cornell University Institute for Infocomm Research, A*STAR \n\\(^{9}\\)Faculty of Data Science, City University of Macau" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.37, + 0.744, + 0.661 + ], + "angle": 0, + "content": "Abstract. Surgical scene perception via videos is critical for advancing robotic surgery, telesurgery, and AI-assisted surgery, particularly in ophthalmology. However, the scarcity of diverse and richly annotated video datasets has hindered the development of intelligent systems for surgical workflow analysis. Existing datasets face challenges such as small scale, lack of diversity in surgery and phase categories, and absence of time-localized annotations. These limitations impede action understanding and model generalization validation in complex and diverse real-world surgical scenarios. To address this gap, we introduce OphNet, a large-scale, expert-annotated video benchmark for ophthalmic surgical workflow understanding. OphNet features: 1) A diverse collection of 2,278 surgical videos spanning 66 types of cataract, glaucoma, and corneal surgeries, with detailed annotations for 102 unique surgical phases and 150 fine-grained operations. 2) Sequential and hierarchical annotations for each surgery, phase, and operation, enabling comprehensive understanding and improved interpretability. 3) Time-localized annotations, facilitating temporal localization and prediction tasks within surgical workflows. With approximately 285 hours of surgical videos, OphNet is about 20 times larger than the largest existing surgical workflow analysis benchmark. Code and dataset are available at: https://minghuo830.github.io/OphNet-benchmark/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.673, + 0.741, + 0.702 + ], + "angle": 0, + "content": "Keywords: Surgical Workflow Understanding \\(\\cdot\\) Ophthalmic Surgery \\(\\cdot\\) Medical Image Analysis \\(\\cdot\\) Video Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.76, + 0.377, + 0.775 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.819 + ], + "angle": 0, + "content": "As surgical robot platforms such as the da Vinci® surgical system become increasingly sophisticated, there is growing interest in integrating enhanced intelligi-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.567, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "gence into scenarios like minimally invasive surgery [20,32,68]. The advancements in machine vision perception empower these robotic systems to autonomously recognize and adapt to the intricacies of surgical environments, without relying on binary instrument usage signals, RFID tags, sensor data from tracking devices, or other signal information that necessitates laborious manual annotations or additional equipment installations [28]. This autonomy includes the capacity to identify anatomical structures, detect anomalies, and adjust surgical plans in real-time, which is crucial in dynamic and unpredictable surgical settings. In recent years, especially in endoscopy and ophthalmic surgery, the application of deep learning has demonstrated considerable promise in bolstering these autonomous capabilities. This encompasses the analysis of surgical workflows [6, 28, 74], segmentation of instruments and anatomy [4, 24, 43], and depth estimation [78], among others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Automatic video surgical workflow understanding is a fundamental yet challenging problem for developing computer-assisted and robotic-assisted surgery, which can be divided into internal (e.g., laparoscopic and endoscopic [28,47,57]) and external (e.g., operating room and nursing procedure [41]) analysis. In addition to promoting the development of intelligent surgery, it also greatly benefits surgical documentation, education, and training [11,12,65]. Baret et al. [6] showed networks, like Inflated 3D ConvNet (I3D) [10] that utilize spatiotemporal convolutions, require a relatively extensive dataset for effective training. In their study, the model achieves an accuracy exceeding \\(80\\%\\) when trained on 100 videos, with a progressive improvement as the sample size surpasses 700. However, the highly efficient and rapidly evolving deep learning technologies for surgical workflow analysis are currently limited by the following shortcomings in current video benchmarks: 1) Small-scale: the majority of video datasets contain no more than 100 videos. For example, the CATARACTS [23] and CatRelDet [23] datasets contain only 50 and 21 surgical videos, respectively. These datasets are relatively small, insufficient for large-scale validation. 2) Limited categories of surgeries and phases: almost all ophthalmic surgical video datasets only include cataract surgery and do not further classify specific types of surgeries. Additionally, the number of phase categories is also limited, like CatRelDet [23] only contains 4 different phase labels, which is insufficient to meet the requirements for evaluation in real clinical environments. 3) Coarse-grained annotation: due to annotation costs, existing benchmarks often have coarse-grained action definitions. For example, adhesive injection may occur in two different phases: main incision and capsulorhexis, so it may be classified into different phase categories. Coarse-grained action definitions may lead to annotation bias. 4) Single time-boundary annotation: they only annotate designated phases in the videos, ignoring the continuity across different stages of ophthalmic surgery, as well as the hierarchical relationship between surgery, phase, and operation. Simpler datasets, such as LensID [22], are limited to binary classification tasks distinguishing lens implantation from other irrelevant phases. 5) Uniform domain: the videos are meticulously collected, and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.145, + 0.787, + 0.396 + ], + "angle": 0, + "content": "
ProtocolDataset PropertiesTasks
DatasetsNo. of VideosNo. of Action SegmentsNo. of Surgery CategoriesNo. of Action CategoriesTotal DurationMulti-Surgery Presence RecognitionPhase RecognitionPhase LocalizationPhase Prediction
Endo&LapCholec120 [45]120-1776.2hX✓XX
SurgicalActions160 [54]1601601160.2hX✓XX
HeiCo [39, 50]30-3142.8hX✓✓✓
EndoVis 2021 [67]332501722.0hX✓XX
PitVis [3]2528711733.3hX✓XX
CholecT50 [46]50-11044.7hX✓✓✓
AutoLaparo [70]213001723.1hX✓✓✓
OphScopeLensID [22]1002,4401211.7hX✓XX
Cataract-101 [55]1011,26611014.0hX✓✓✓
CatRelDet [23]212,400142.0hX✓XX
CATARACTS [4]501,53611920.0hX✓✓✓
Cataract-1K [21]1,000931112118.7hX✓✓✓
OphNet(Ours)2,2789,79566150284.8h✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.671 + ], + "angle": 0, + "content": "Table 1: The statistics comparison among existing workflow analysis datasets and our OphNet. Compared to other datasets, OphNet focuses on more comprehensive coverage of various surgery, phase and operation categories, collects a large number of videos, totaling 284.8 hours, and also enables a variety of recognition, localization and prediction tasks. OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. For instance, Cholec120 [45], Cholec80 [65], m2cai-workflow and LapChole [62] form one series, whereas CholecT50 [46], CholecT45 [46], and CholecT40 [44] comprise another series. We have excluded the following scenarios from our comparison: (1) non-open-source datasets such as Bypass170 [66], ESD [30], Yu's [74], etc.; (2) a superset of multiple open-source or non-open-source datasets, like Cholec207 [6], etc.; (3) datasets employed for lesion, anatomy, and instrument classification and segmentation, such as SUN-SEG [27], CVC-ClinicDB [7], ROBUSTMIS [52], Mesejo's [40], Cata7 [43], etc., anomaly detection such as PolypDiag [63] (from Hyper-Kvasir [9] and LDPolypVideo [38]), Kvasir-Capsule [59], etc., and other datasets not dedicated to workflow analysis. It's worth mentioning that even in comparison with the above datasets, OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. Endo&Lap denotes the endoscopic and laparoscopic protocol, OphScope denotes the ophthalmic microscope protocol. We choose the latest version for comparison in cases where datasets have multiple supplementary updates." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.748 + ], + "angle": 0, + "content": "while this ensures video quality, the uniform style is not conducive to testing the model's domain generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "While some works have explored semi-supervised and self-supervised learning strategies [8, 51, 73, 75] to alleviate the cost of annotations or use only a small fraction of available labels, these approaches still lack competitiveness in performance compared to fully supervised learning. This deficiency in performance is impeding the widespread clinical application of these strategies. To address the shortage of sufficient labeled datasets, we construct OphNet, a large-scale" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "and expert-level video benchmark with high diversity, for ophthalmic surgical workflow understanding. The main advantages of OphNet are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.186, + 0.787, + 0.321 + ], + "angle": 0, + "content": "- Largest scale and diversity: to the best of our knowledge, OphNet is currently the largest and most richly labeled dataset for surgical workflow analysis. It contains a number of videos 20 times greater than the current largest benchmark in ophthalmic surgery and far exceeds datasets in more established fields such as endoscopy. Additionally, OphNet includes the greatest variety of different types of surgeries, encompassing 66 different surgeries such as cataract, glaucoma, and corneal surgeries, along with 102 unique surgical phases and 150 distinct operations. This diversity significantly surpasses that of previous research." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.322, + 0.787, + 0.441 + ], + "angle": 0, + "content": "- Fine-grained, sequential and hierarchical annotation: we have meticulously selected a subset of videos for annotation localization, with each video being annotated for an average of 22 operations. Additionally, we provide exquisite annotations at the levels of surgery, phase, and operation, catering to the requirements for training specific challenge models. This annotation design aims to offer a multifaceted understanding of surgical protocols, accommodate the nuances of each distinct surgery, and enhance the usability and interpretability of our dataset." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.443, + 0.787, + 0.533 + ], + "angle": 0, + "content": "- Expert-level manual annotation: the annotation work for OphNet was completed by ten experienced ophthalmologists and five individuals with ophthalmic experience, encompassing, but not limited to, standardization of definitions for surgery, phase, and operation labels, video filtering, classification and localization annotations, and secondary verification, among others. Expert-level annotators ensure the quality and professionalism of OphNet." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.186, + 0.787, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.554, + 0.387, + 0.57 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.584, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Surgical Workflow Understanding. Beyond its therapeutic advantages, minimally invasive surgery also provides the capability for operative video recording. These videos can be stored and later utilized for various purposes such as cognitive training, skill assessment, and surgical workflow analysis [76, 77]. Techniques derived from the broader field of video content analysis and representation are increasingly being incorporated into the surgical realm [6, 36, 74]. A typical surgical workflow can be defined by a sequence of tasks or events, including patient positioning, incision, dissection, and suturing. These events are influenced not only by the specific type of surgery but also by the individual surgeon's proficiency and technique. Consequently, a comprehensive understanding of the surgical workflow necessitates a thorough examination of the temporal, spatial, and contextual facets of these tasks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Weakly-Supervised Video Learning. Substantial pioneering work has been undertaken in the realms of video understanding [5,16,33,53,69,71,72,79]. Since even a small amount of videos easily comprises several million frames, methods that do not rely on a frame-level annotation are of special importance. Weakly-supervised video learning makes use of loosely labeled data to train models," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "thereby obviating the necessity for exhaustively annotated training data. These weak labels might manifest in a variety of forms, including video-level labels or partial labels, which are less specific than frame-level or pixel-level annotations. The objective of weakly-supervised learning is to utilize these coarse labels to produce models capable of delivering fine-grained predictions, such as temporally precise action localization or detailed semantic segmentation [15,35]. The large-scale annotation of surgical videos demands a significant investment of invaluable medical time resources. In this context, weakly supervised learning emerges as a viable solution to this bottleneck." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.461, + 0.323 + ], + "angle": 0, + "content": "3 Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.79, + 0.535 + ], + "angle": 0, + "content": "In this section, we detail the construction of our dataset, which involved meticulous data collection and preprocessing. YouTube was leveraged as a primary source to circumvent privacy issues while ensuring a broad representation of ophthalmic surgeries. Our selection criteria aimed to capture a wide range of video qualities and styles, specifically targeting cataract, glaucoma, and corneal surgeries due to their prevalence in clinical settings. We refined the dataset by excluding videos of inadequate quality or those depicting non-human subjects. The annotation process was designed to reflect the complex nature of eye surgeries, incorporating hierarchical classification to account for multiple conditions often treated within a single procedure. This was complemented by detailed localization annotations, delineating the distinct phases and techniques characteristic of ophthalmic operations, all undertaken by a dedicated team of ophthalmologists to ensure accuracy and relevance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.558, + 0.534, + 0.574 + ], + "angle": 0, + "content": "3.1 Data Collection & Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Collection. Medical data exhibit unique privacy considerations and tend to be of smaller data volumes, characteristics that are particularly salient in the case of video data. By capitalizing on the wealth of surgical videos available on YouTube, we are able to obviate potential ethical and privacy concerns while concurrently enabling the rapid procurement of a substantial corpus of videos for further screening [1,2,17,29]. To fulfill this objective, we deploy text-based search algorithms to probe each surgery on YouTube, obtaining videos with titles that incorporate the requisite surgical keywords. We select cataract, glaucoma, and corneal surgery—three of the most commonly performed ophthalmic surgeries in actual clinical environments—as the central subjects of our research. To expand our video collection, we bolster our search queries by integrating synonyms and abbreviations associated with each type of ophthalmic surgery. For instance, cataract surgery encompasses various types: Phacoemulsification (abbr. PHACO), Intraocular Lens implantation (abbr. IOL), and Extracapsular Cataract Extraction (abbr. ECCE), etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Preprocessing. Given the nature of text-based retrieval and the varying quality, style, and filming methods of surgical videos influenced by different YouTube" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.141, + 0.784, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 1: The figure shows two combined surgical videos, PHACO + IOL implantation and PHACO + Trabeculotomy + Transscleral Cyclophotocoagulation + IOL implantation. For each frame marked in color, we provide time-boundary annotations at surgical, phase and operation levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.643 + ], + "angle": 0, + "content": "sources, some videos may exhibit poor quality or deviations in surgical representation. We filter out low-resolution videos, black-and-white color schemes, and animated demonstrations. Furthermore, surgical videos featuring non-human eyes, such as pig eyes, rabbit eyes, or pseudo-human eyes, are also outside the scope of our annotation. The data preprocessing was jointly completed by five professionals with experience in ophthalmology for initial screening, followed by a review conducted by six attending ophthalmologists." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.403, + 0.679 + ], + "angle": 0, + "content": "3.2 Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Classification Annotation. Different from the single-label classification in most natural video datasets [17, 25, 58, 60, 80], in ophthalmic surgery, physicians typically consider and implement multiple types of surgeries and phases based on the specific conditions and needs of the patient. This multifaceted approach stems from the complexity of the eye as an organ, where numerous diseases often coexist. Consequently, multiple ocular problems may need to be addressed in a single surgery. As shown in Fig. 1, a patient with cataracts may also have glaucoma, necessitating the simultaneous treatment of both conditions within one surgery. Hence, we initially classified all videos based on the primary surgical categories into three principal groups: cataract, glaucoma, and corneal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.114, + 0.733, + 0.13 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.145, + 0.462, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.3, + 0.352, + 0.311 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.145, + 0.783, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.3, + 0.634, + 0.311 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.312, + 0.363, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.374, + 0.328, + 0.384 + ], + "angle": 0, + "content": "Main Incision" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.311, + 0.506, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.374, + 0.473, + 0.385 + ], + "angle": 0, + "content": "Capsulorhexis" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.311, + 0.65, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.374, + 0.622, + 0.386 + ], + "angle": 0, + "content": "Capsule Staining" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.311, + 0.794, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.372, + 0.775, + 0.392 + ], + "angle": 0, + "content": "Hydrodissection and Float Nucleus" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.391, + 0.363, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.455, + 0.34, + 0.465 + ], + "angle": 0, + "content": "Cortical Aspiration" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.392, + 0.506, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.455, + 0.48, + 0.465 + ], + "angle": 0, + "content": "IOL Implantation" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.392, + 0.65, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.455, + 0.63, + 0.465 + ], + "angle": 0, + "content": "Cautery/Hemostasis" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.392, + 0.794, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.455, + 0.774, + 0.465 + ], + "angle": 0, + "content": "Phacoemulsification" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.468, + 0.514, + 0.48 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Fig. 2: OphNet's composition, comparison with other datasets for the same task, and some phase examples: (a) an overview of the composition ratios at the levels of surgery, phase, and operation; (b) comparison among existing open-source \\(\\odot\\) laparoscopic & endoscopic, and \\(\\odot\\) ophthalmic microscope workflow analysis video datasets and \\(\\odot\\) our OphNet. OphNet stands as the largest real-world video dataset for ophthalmic surgical workflow understanding, featuring the highest number of videos, longest duration, and diverse categories of surgeries and phases; (c) eight phase examples in OphNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.702 + ], + "angle": 0, + "content": "surgeries. Subsequently, these were further allocated for the detailed annotation of primary and secondary surgery. It is important to note that there exists only a single type of primary surgery, whereas multiple secondary surgery may be present. The categorization and annotation were meticulously completed by a team comprised of 8 experienced ophthalmologists." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Hierarchical Localization Annotation. A single surgery is often multifaceted, involving a series of intricate phases, each requiring distinct techniques and instruments, and the transition between surgeries entails high precision and coordination. Routine cataract surgery involves several steps. It begins with the administration of anesthesia, followed by a small incision in the cornea. The surgeon then creates an opening in the lens capsule and uses ultrasonic vibrations to break up and remove the cataract. Afterward, an artificial intraocular lens (IOL) is inserted into the lens capsule, and the incision is sealed without stitches. We define the phases and operation for various surgeries based on the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.143, + 0.501, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.332, + 0.418, + 0.343 + ], + "angle": 0, + "content": "(a) phase statistics" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.785, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.332, + 0.711, + 0.343 + ], + "angle": 0, + "content": "(b) operation statistics" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.355, + 0.785, + 0.397 + ], + "angle": 0, + "content": "Fig. 3: We present the data statistics of trimmed videos at the levels of phase and operation, including the number of trimmed videos, average duration, and total duration. The IDs and corresponding names can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.562 + ], + "angle": 0, + "content": "textbook Ophthalmic Surgery: Principles and Practice [61]. To ensure the quality of localization annotations, we assign each annotator videos of 2 different types of primary surgeries. For each action time-boundary annotation, we annotate at three different levels of granularity: surgery, phase, and operation. We also employed the complete linkage algorithm [13] to cluster and merge various temporal boundaries into stable boundaries that received multiple agreements. It's worth noting that an individual video may have multiple separate instances of the same or different phases, thereby leading to multiple boundary definitions. A team of 15 ophthalmologists completes the localization annotation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.583, + 0.52, + 0.598 + ], + "angle": 0, + "content": "3.3 Dataset Statistics and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.758 + ], + "angle": 0, + "content": "OphNet includes 2,278 surgical videos (284.8 hours), demonstrating 66 different types of ophthalmic surgeries: 13 types of cataract surgery, 14 types of glaucoma surgery, and 39 types of corneal surgery. There are 102 phases and 150 operations for recognition, detection and prediction tasks, summarized in Tab. 1. Over \\(77\\%\\) of videos have high-definition resolutions of \\(1280 \\times 720\\) pixels or higher. To facilitate algorithm development and evaluation, we selected 523 videos for localization annotations. Additionally, we trim videos according to annotated action boundaries, resulting in 7,320 phase instances and 9,795 operation instances, totaling 51.2 hours. The average duration of trimmed videos is 32 seconds, while untrimmed videos average 337 seconds." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.78, + 0.377, + 0.797 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In our study, we explore four potential tasks using the OphNet dataset: 1) primary surgery presence recognition, 2) phase and operation recognition, 3) phase" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.114, + 0.732, + 0.13 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "localization, and 4) phase anticipation. To establish robust baselines, we employ state-of-the-art models known for their effectiveness in human action recognition, detection and anticipation. For each task, we provide a detailed problem formulation and evaluate the baseline models' performance. Our findings offer valuable insights into video understanding within sequences and fine granularity, contributing to the field's knowledge of video tasks in medical contexts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.259, + 0.696, + 0.275 + ], + "angle": 0, + "content": "4.1 Surgery Presence, Phase and Operation Recognition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.789, + 0.448 + ], + "angle": 0, + "content": "Task Description. Surgery presence recognition focuses on identifying various surgical types in untrimmed videos through weakly supervised methods. This requires the model to discern and capture distinct surgical action features within extensive video footage. In OphNet, all types of surgeries occurring in each video are annotated, but only a subset of the videos have time-boundary annotations, as detailed in Sec. 3.2. We identify the primary surgery based on the key objectives and durations of the surgeries. To streamline the process, our experiments are limited to recognizing the presence of these primary surgeries. Additionally, phase recognition segments the surgery into distinct phases using visual cues and movements, such as incision, lens removal, and implantation. Operation recognition involves identifying finer-grained surgical actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.446, + 0.72 + ], + "angle": 0, + "content": "Baselins. We compare the performance of I3D [10], SlowFast [19], X3D [18], and MViT V2 [31] models on this task. These models are evaluated in two versions: 1) random initialization training and 2) pretraining with weights from Kinetics 400 [29], which is a human action recognition dataset. In addition, we also explored the classification performance of X-CLIP [42] and ViFi-CLIP [49], two CLIP [48]-based models. We also compared the effects of different numbers of input frames on the models' performance, where the sub" + }, + { + "type": "table", + "bbox": [ + 0.458, + 0.479, + 0.789, + 0.621 + ], + "angle": 0, + "content": "
BaselinesPrimary Surgery Classification
CataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
I3D [10]38.986.343.671.742.778.229.853.2
SlowFast [19]45.382.144.672.345.577.327.254.4
X3D [31]42.187.444.674.243.876.628.562.7
MViT V2 [18]43.284.245.581.845.575.929.160.1
I3D [10]36.884.748.281.848.676.527.250.6
SlowFast*49.083.747.380.949.275.627.250.6
X3D*47.486.346.481.548.378.235.461.4
MViT V2*44.285.349.181.847.777.328.563.3
X-CLIP16 [42]58.594.751.892.861.488.640.579.0
X-CLIP3260.692.653.583.756.884.158.981.0
ViFi-CLIP16 [49]59.688.352.480.261.481.858.979.8
ViFi-CLIP3259.688.351.584.850.075.056.377.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.456, + 0.631, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Table 2: Per-class Top-1 and Top-5 accuracy (%) for the primary surgery presence recognition on untrimmed videos. The best performance for each split has been highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.748 + ], + "angle": 0, + "content": "script _16 represents an input of 16 frames, and _32 represents an input of 32 frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Setup. The dataset was randomly partitioned to ensure a balanced representation of examples for each surgery category. Specifically, we allocated \\(70\\%\\) of the data for training (1,449 surgical videos), \\(10\\%\\) for validation (205 surgical videos), and \\(20\\%\\) for testing (424 surgical videos). For the phase and operation recognition experiments, we followed the same settings, with \\(70\\%\\) of the data used for training (5,024 phase segments, 6,856 operations segments), \\(10\\%\\) for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "validation (730 phase segments, 975 operations segments), and \\(20\\%\\) for testing (1,566 phase segments, 1,964 operations segments). The input for the surgery presence recognition experiment is untrimmed videos, while for the phase and operation recognition experiments, the input consists of trimmed segments. In all classification experiments, we set up the analysis and comparison from four perspectives: cataract surgery, glaucoma surgery, corneal surgery, and all surgical videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Results. We summarize the results in Tab. 2 and Tab. 3. In primary surgery classification, X-CLIP [42] achieved the highest overall Top-1 accuracy at \\(58.9\\%\\), leading in cataract and glaucoma surgeries with accuracies of \\(60.6\\%\\) and \\(53.5\\%\\) respectively. For corneal surgeries, X-CLIP also recorded the highest Top-1 and Top-5 accuracies of \\(61.4\\%\\) and \\(88.6\\%\\). In phase classification, ViFi-CLIP [49] led with the highest Top-1 accuracy, particularly in cataract surgeries at \\(75.9\\%\\), and exhibited the best performance in both glaucoma and corneal surgeries. Furthermore, in operation classification, ViFi-CLIP outperformed all other models in all categories, especially noted in cataract and corneal surgeries with Top-1 accuracies of \\(75.1\\%\\) and \\(83.7\\%\\) and Top-5 accuracies of \\(93.8\\%\\) and \\(85.2\\%\\) respectively. Overall, ViFi-CLIP showed superior performance in both phase and operation classifications across various surgery types. Besides, in the phase and operation classification experiments, a higher number of input frames generally had a positive effect on the model. In Fig. 4, we present the heatmap visualization of four examples from the test set of phase recognition experiments using the ViFi-CLIP model. It can be observed that in a series of frame images, the model focuses on surgical instruments and the operated eye area, which is consistent with human experience. The instruments used in the same phase of ophthalmic surgery are often similar." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.416, + 0.582 + ], + "angle": 0, + "content": "4.2 Phase Localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Task Description. Phase localization in ophthalmic surgical workflow analysis refers to the task of pinpointing the exact moments or time intervals within a surgical video where specific phases of the surgery begin and end. This involves the detailed temporal segmentation of the entire surgical procedure into its constituent phases based on visual cues, surgeon's actions, and the progression of the surgery. The objective of phase localization is to accurately identify the start and end times of different surgical stages, such as pre-operative preparation, incision, and removal of the lens facilitating a granular and precise understanding of the surgery timeline. This task is crucial for detailed surgical documentation, efficient surgical training, and the development of targeted interventions during specific stages of the surgery, enhancing overall surgical management and post-operative analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baselines. We conducted the experiments using phase-level labels and used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. Data split follows the setup of the primary surgery classification experiment." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.145, + 0.788, + 0.304 + ], + "angle": 0, + "content": "
BaselinesPhase ClassificationOperation Classification
CataractGlaucomaCorneaAllCataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1
I3D [10]27.255.724.157.518.952.125.758.226.854.923.556.018.051.225.057.1
SlowFast [19]26.556.523.156.524.249.126.760.125.855.222.945.923.548.526.059.0
X3D [31]27.058.321.055.521.828.526.662.326.447.220.544.621.327.826.161.5
MViT V2 [18]26.254.921.053.426.046.827.059.825.943.820.542.725.545.926.558.9
I3D*29.568.922.158.626.050.930.271.228.867.521.847.925.750.329.560.0
SlowFast*30.672.325.254.730.759.831.761.829.971.124.843.929.558.730.560.9
X3D*27.272.922.159.630.761.533.563.226.571.821.748.829.960.232.862.1
MViT V2*34.276.523.352.038.465.128.360.233.575.222.841.537.964.027.859.5
X-CLIP16 [42]68.392.247.389.853.077.463.485.367.591.046.578.982.276.162.584.0
X-CLIP3269.194.048.781.754.880.462.785.868.093.047.980.584.079.562.084.7
ViFi-CLIP16 [49]75.993.740.485.466.681.666.188.474.592.542.874.585.080.565.087.5
ViFi-CLIP3273.092.949.682.757.781.668.487.275.193.843.280.283.785.264.886.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Table 3: Per-class Top-1 and Top-5 accuracy (\\%) for the primary surgery presence recognition on untrimmed videos and phase recognition on trimmed videos. * denotes the initialization from the model pre-trained on Kinetics 400 [29]. For the two CLIP models, we chose ViT-B/16 as the backbone and compared the performance of two different input frame numbers, 16 and 32. The best performance for each split has been highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.431, + 0.557, + 0.539 + ], + "angle": 0, + "content": "
BaselinesBackbonesmAP (%)
0.10.30.50.7Avg.
ActionFormerCSN [64]53.750.140.624.542.5
SwinViviT [34]59.354.743.326.346.4
SlowFast [19]60.055.945.126.047.5
TriDet [56]CSN56.153.043.129.446.2
SwinViviT61.057.147.133.150.4
SlowFast61.356.045.630.448.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.544, + 0.556, + 0.655 + ], + "angle": 0, + "content": "Table 4: The results for phase detection. Action-Former and TriDet are state-of-the-art models for human action detection tasks, and we use three different backbones for feature extraction and report mAP at the IoU thresholds of [0.1:0.2:0.9]. Average mAP is computed by averaging different IoU thresholds. The best performance for each split has been highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.431, + 0.789, + 0.539 + ], + "angle": 0, + "content": "
BaselinesTop-1 Acc. (%)
0.10.30.50.7Avg.
I3D [10]26.542.249.851.347.3
SlowFast [19]25.442.648.952.247.2
MViT V2 [18]25.643.749.352.347.5
I3D*27.343.550.151.447.6
SlowFast*27.543.249.952.347.8
MViT V2*27.843.850.551.748.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.559, + 0.544, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Table 5: The results for phase anticipation. We report top-1 accuracy at the observation ratios [0.1:0.2:0.9]. Average top-1 accuracy is computed by averaging different observation ratios. The best performance for each split has been highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Setup. We excluded Operation Gap and Invalid, and filtered out tags with fewer than 20 segments. We used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. To extract features from the videos, we first extracted RGB frames from each video at a rate of 25 frames per second. We also extracted optical flow using the TV-L1 [26,37] algorithm. We then fine-tuned an I3D [10] model that had been pre-trained on the ImageNet [14] dataset, and used it to generate features for each RGB and optical flow frame. Because each video has a variable duration, we performed uniform interpolation to generate 100 fixed-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.141, + 0.5, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.141, + 0.785, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Fig. 4: Attention map visualizations of ViFi-CLIP [49] on four examples from OphNet's test set in the phase recognition task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.785, + 0.429 + ], + "angle": 0, + "content": "length features for each video. Finally, we concatenated the RGB and optical flow features into a 2048-dimensional embedding, which served as the input for our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Results. The experimental results for phase localization are presented in the Tab. 4, showcasing the performance of different baseline models with various backbones in terms of mean Average Precision (mAP) at different Intersection over Union (IoU) thresholds [0.1:0.2:0.9]. The models evaluated include Action-Former with SwinViviT and SlowFast backbones, and TriDet with CSN, SwinViviT, and SlowFast backbones. The results indicate that the TriDet model with a SwinViviT backbone outperforms other combinations, achieving the highest mAP scores across most IoU thresholds, with notable scores of \\(61.0\\%\\) (IoU=0.1), \\(57.1\\%\\) (IoU=0.3), \\(47.1\\%\\) (IoU=0.5), and \\(33.1\\%\\) (IoU=0.7), resulting in an average mAP of \\(50.4\\%\\). This indicates that the TriDet model, especially when combined with the SwinViviT backbone, is particularly effective for phase localization in surgical videos. On the other hand, the TriDet model with a SlowFast backbone shows competitive performance, particularly achieving the highest mAP of \\(61.3\\%\\) at the lowest IoU threshold (0.1). However, it falls slightly behind in performance at higher IoU thresholds compared to the SwinViviT backbone." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.421, + 0.695 + ], + "angle": 0, + "content": "4.3 Phase Anticipation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Task Description. This task requires the analysis of real-time or recorded video data to foresee the sequence of events based on current and past surgical activities. By understanding the typical progression of ophthalmic surgeries and recognizing patterns in the surgeon's actions and the use of instruments, the system aims to forecast the next phase of the surgery, allowing for proactive preparation and response. The objective of phase anticipation is to enhance the efficiency and safety of surgical procedures by providing the surgical team with advanced notice of upcoming steps, enabling better resource allocation, timing for critical tasks, and overall coordination within the operating room." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.141, + 0.496, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.141, + 0.785, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.288 + ], + "angle": 0, + "content": "Fig. 5: Phase localization visualization of TriDet [56]. \\(GT\\) represents the ground truth visualization for phases, while \\(Loc\\). visualizes the model's highest confidence phase category and the time-boundary results. Blank segments denote invalid segments or operation gaps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.784, + 0.361 + ], + "angle": 0, + "content": "Setup. Following previous approaches of the primary surgery classification experiment in Sec. 4.1, We randomly mask phase sequences in the test video with different observation ratios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Baselines. We evaluate our datasets with the baseline models such as I3D [10], SlowFast [19], and MViT V2 [31]. For each model, we also adopted two training approaches: random initialization training and using pre-trained weights from Kinetics 400 [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Results. The phase detection results are illustrated in Tab. 5. The results demonstrate that the baseline models pretrained on Kinetics 400 [29] generally outperform their original counterparts in terms of Top-1 accuracy across different observation ratios for phase anticipation. Specifically, the modified MViT V2* model exhibits the highest improvement, achieving the best average Top-1 accuracy of \\(48.2\\%\\). Moreover, while all models show increased accuracy with higher observation ratios, indicating that more observed data contributes to better performance, the consistent improvement across all ratios for the enhanced models suggests effective modifications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.583, + 0.364, + 0.598 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Dataset Bias. OphNet's videos are sourced from YouTube and exhibit diverse styles, clarity, and screen elements. This diversity can aid detection models in generalization but may affect their effectiveness and performance. Some videos in the dataset include subtitles or additional video windows, such as a little subtitle or watermark shown in Fig. 2. Similarly, additional video windows offer another perspective but can make the scene chaotic, making it harder to recognize primary surgical actions. The presence of these factors in OphNet reflects the complexity of real-world surgical environments, because an ophthalmic microscope may inherently display different windows or show parameters during recording. While they pose challenges, they also present opportunities for developing models that can better handle variability and unpredictability, which are crucial aspects of real-world surgical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Annotation Bias. OphNet is entirely annotated by ophthalmologists, there is a distinct possibility of annotation bias reflecting specific regional practices, terminologies, and interpretations. Despite the universal nature of many ophthalmic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "procedures, subtle differences in surgical techniques, procedural preferences, and clinical terminologies could lead to inconsistencies in how surgeries are categorized and described across different regions. For instance, the terminology used to describe certain procedures might differ, with one region referring to a procedure as anterior vitrectomy while another uses pars plana vitrectomy. To reduce the possibility of biases in precise annotations, we have taken great care to establish a unified definition prior to describing the surgery, phase and operation. However, potential biases arising from regional variations and individual surgical practices are inevitable." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.306, + 0.359, + 0.323 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.789, + 0.58 + ], + "angle": 0, + "content": "In response to the current challenges in ophthalmology, a surgical field apt for automation and remote control, we introduce OphNet, a large-scale, diverse, and expert-level video benchmark for understanding ophthalmic surgical workflows. OphNet is the most extensive dataset of its kind, containing a broad range of cataract, glaucoma, and corneal surgeries and detailed annotations for distinct surgical phases. OphNet comprises 2,278 surgical videos (284.8 hours), 7,320 phase segments and 9,795 operation segments (51.2 hours), showcasing 66 different types of ophthalmic surgeries: 13 cataract, 14 glaucoma, and 39 corneal. It is annotated with 102 phases and 150 operations. With OphNet, we explored primary surgery presence recognition, phase localization and phase anticipation on untrimmed videos, phase and operation recognition on trimmed videos. We employed state-of-the-art models to establish robust baselines and provided valuable insights into video understanding within sequences and fine granularity. Our work contributes to the broader understanding of surgical video tasks in medical contexts and promotes the integration of deep learning technologies into ophthalmic surgical procedures." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.325, + 0.618 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.632, + 0.787, + 0.673 + ], + "angle": 0, + "content": "1. Fair use on youtube. https://support.google.com/youtube/answer/9783148?hl=en#:~:text=If%20the%20use%20of%20copyright, copyright %20removal%20request%20to%20YouTube." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.674, + 0.787, + 0.702 + ], + "angle": 0, + "content": "2. Youtube's copyright exception policy. https://www.youtube.com/howyoutubeworks/policies/copyright/#copyright-exceptions" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "3. Adrito, D., Danyal, Z.K., Dimitrios, P., Yitong, Z., John, G.H., Francisco, V., Sophia, B., Hani, J.M., Danail, S.: Pitvis: Workflow recognition in endoscopic pituitary surgery" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.743, + 0.787, + 0.84 + ], + "angle": 0, + "content": "4. Al Hajj, H., Lamard, M., Conze, P.H., Roychowdhury, S., Hu, X., Marsalkaite, G., Zisimopoulos, O., Dedmari, M.A., Zhao, F., Prellberg, J., Sahu, M., Galdran, A., Araujo, T., Vo, D.M., Panda, C., Dahiya, N., Kondo, S., Bian, Z., Vahdat, A., Bialopetravicius, J., Flouty, E., Qiu, C., Dill, S., Mukhopadhyay, A., Costa, P., Aresta, G., Ramamurthy, S., Lee, S.W., Campilho, A., Zachow, S., Xia, S., Conjeti, S., Stoyanov, D., Armaitis, J., Heng, P.A., Macready, W.G., Cochener, B., Quellec, G.: Cataracts: Challenge on automatic tool annotation for cataract" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.632, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "surgery. Medical Image Analysis 52, 24-41 (2019). https://doi.org/https://doi.org/10.1016/j.media.2018.11.008, https://www.sciencedirect.com/science/article/pii/S136184151830865X" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "5. Alwassel, H., Giancola, S., Ghanem, B.: TSP: Temporally-sensitive pretraining of video encoders for localization tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 3173-3183 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "6. Bar, O., Neimark, D., Zohar, M., Hager, G.D., Girshick, R., Fried, G.M., Wolf, T., Asselmann, D.: Impact of data on generalization of ai for surgical intelligence applications. Scientific Reports 10(1), 22208 (Dec 2020). https://doi.org/10.1038/s41598-020-79173-6, https://doi.org/10.1038/s41598-020-79173-6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.785, + 0.356 + ], + "angle": 0, + "content": "7. Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilarino, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics 43, 99-111 (2015). https://doi.org/https://doi.org/10.1016/j.compmedimag.2015.02.007, https://www.sciencedirect.com/science/article/pii/S0895611115000567" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.356, + 0.785, + 0.412 + ], + "angle": 0, + "content": "8. Bodenstedt, S., Wagner, M., Katic, D., Mietkowski, P., Mayer, B., Kenngott, H., Müller-Stich, B., Dillmann, R., Speidel, S.: Unsupervised temporal context learning using convolutional neural networks for laparoscopic workflow analysis. arXiv preprint arXiv:1702.03684 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.508 + ], + "angle": 0, + "content": "9. Borgli, H., Thambawita, V., Smedsrud, P.H., Hicks, S., Jha, D., Eskeland, S.L., Randel, K.R., Pogorelov, K., Lux, M., Nguyen, D.T.D., Johansen, D., Griwodz, C., Stensland, H.K., Garcia-Ceja, E., Schmidt, P.T., Hammer, H.L., Riegler, M.A., Halvorsen, P., de Lange, T.: HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific Data 7(1), 283 (2020). https://doi.org/10.1038/s41597-020-00622-y, https://doi.org/10.1038/s41597-020-00622-y" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.55, + 0.785, + 0.633 + ], + "angle": 0, + "content": "11. Czempiel, T., Paschali, M., Keicher, M., Simson, W., Feussner, H., Kim, S.T., Navab, N.: Tecn: Surgical phase recognition with multi-stage temporal convolutional networks. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoeanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 343-352. Springer International Publishing, Cham (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.633, + 0.785, + 0.702 + ], + "angle": 0, + "content": "12. Czempiel, T., Paschali, M., Ostler, D., Kim, S.T., Busam, B., Navab, N.: Opera: Attention-regularized transformers for surgical phase recognition. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part IV 24. pp. 604–614. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.744 + ], + "angle": 0, + "content": "13. Defays, D.: An efficient algorithm for a complete link method. The Computer Journal 20(4), 364-366 (01 1977). https://doi.org/10.1093/comjnl/20.4.364, https://doi.org/10.1093/comjnl/20.4.364" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.786 + ], + "angle": 0, + "content": "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Dong, S., Hu, H., Lian, D., Luo, W., Qian, Y., Gao, S.: Weakly supervised video representation learning with unaligned text for sequential videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2447 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Duong, H.T., Le, V.T., Hoang, V.T.: Deep learning-based anomaly detection in video surveillance: A survey. Sensors 23(11) (2023). https://doi.org/10.3390/s23115024, https://www.mdpi.com/1424-8220/23/11/5024" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "17. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "18. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.37 + ], + "angle": 0, + "content": "20. Forslund Jacobsen, M., Konge, L., Alberti, M., la Cour, M., Park, Y.S., Thomsen, A.S.S.: ROBOT-ASSISTED VITREORETINAL SURGERY IMPROVES SURGICAL ACCURACY COMPARED WITH MANUAL SURGERY: A randomized trial in a simulated setting. Retina 40(11), 2091-2098 (Nov 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.424 + ], + "angle": 0, + "content": "21. Ghamsarian, N., El-Shabrawi, Y., Nasirihaghighi, S., Putzgruber-Adamitsch, D., Zinkernagel, M., Wolf, S., Schoeffmann, K., Sznitman, R.: Cataract-1k: Cataract surgery dataset for scene segmentation, phase recognition, and irregularity detection. arXiv preprint arXiv:2312.06295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.548 + ], + "angle": 0, + "content": "22. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., El-Shabrawi, Y., Schoeffmann, K.: Lensid: A cnn-rnn-based framework towards lens irregularity detection in cataract surgery videos. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part VIII. Lecture Notes in Computer Science, vol. 12908, pp. 76-86. Springer (2021). https://doi.org/10.1007/978-3-030-87237-3_8, https://doi.org/10.1007/978-3-030-87237-3_8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.632 + ], + "angle": 0, + "content": "23. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., Schoeffmann, K.: Relevance detection in cataract surgery videos by spatio-temporal action localization. In: 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021. pp. 10720-10727. IEEE (2020). https://doi.org/10.1109/ICPR48806.2021.9412525, https://doi.org/10.1109/ICPR48806.2021.9412525" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "24. Grammatikopoulou, M., Flouty, E., Kadkhodamohammadi, A., Quellec, G., Chow, A., Nehme, J., Luengo, I., Stoyanov, D.: Cadis: Cataract dataset for image segmentation. arXiv preprint arXiv:1906.11586 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "25. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: Ava: A video dataset of spatio-temporally localized atomic visual actions (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "26. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "27. Ji, G.P., Xiao, G., Chou, Y.C., Fan, D.P., Zhao, K., Chen, G., Van Gool, L.: Video polyp segmentation: A deep learning perspective. Machine Intelligence Research 19(6), 531-549 (Dec 2022). https://doi.org/10.1007/s11633-022-1371-y, https://doi.org/10.1007/s11633-022-1371-y" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "28. Jin, Y., Dou, Q., Chen, H., Yu, L., Qin, J., Fu, C.W., Heng, P.A.: SV-RCNet: Workflow recognition from surgical videos using recurrent convolutional network. IEEE Trans. Med. Imaging 37(5), 1114-1126 (May 2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.271 + ], + "angle": 0, + "content": "30. Li, J., Jin, Y., Chen, Y., Yip, H.C., Scheppach, M., Chiu, P.W.Y., Yam, Y., Meng, H.M.L., Dou, Q.: Imitation learning from expert video data for dissection trajectory prediction in endoscopic surgical procedure. In: Greenspan, H., Madabhushi, A., Mousavi, P., Salcudean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2023. pp. 494-504. Springer Nature Switzerland, Cham (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "31. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.393 + ], + "angle": 0, + "content": "32. Lin, S., Miao, A.J., Lu, J., Yu, S., Chiu, Z.Y., Richter, F., Yip, M.C.: Semantic-super: A semantic-aware surgical perception framework for endoscopic tissue identification, reconstruction, and tracking. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4739-4746 (2023). https://doi.org/10.1109/ICRA48891.2023.10160746" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "33. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3889-3898 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.461 + ], + "angle": 0, + "content": "34. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. arXiv preprint arXiv:2106.13230 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.514 + ], + "angle": 0, + "content": "35. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Bi-calibration networks for weakly-supervised video representation learning. International Journal of Computer Vision 131(7), 1704-1721 (Jul 2023). https://doi.org/10.1007/s11263-023-01779-w, https://doi.org/10.1007/s11263-023-01779-w" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.556 + ], + "angle": 0, + "content": "36. Loukas, C.: Video content analysis of surgical procedures. Surgical Endoscopy 32(2), 553-568 (Feb 2018). https://doi.org/10.1007/s00464-017-5878-1, https://doi.org/10.1007/s00464-017-5878-1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "37. Lucas, B.D., Kanade, T., et al.: An iterative image registration technique with an application to stereo vision, vol. 81. Vancouver (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.65 + ], + "angle": 0, + "content": "38. Ma, Y., Chen, X., Cheng, K., Li, Y., Sun, B.: Ldpolypvideo benchmark: A large-scale colonoscopy video dataset of diverse polyps. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021. pp. 387-396. Springer International Publishing, Cham (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.704 + ], + "angle": 0, + "content": "39. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific data 8(1), 101 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.759 + ], + "angle": 0, + "content": "40. Mesejo, P., Pizarro, D., Abergel, A., Rouquette, O., Beorchia, S., Poincloux, L., Bartoli, A.: Computer-aided classification of gastrointestinal lesions in regular colonoscopy. IEEE Transactions on Medical Imaging 35(9), 2051-2063 (2016). https://doi.org/10.1109/TMI.2016.2547947" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "41. Ming, H., Lin, W., Siyuan, Y., Don, M., Qingli, R., Peng, X., Wei, F., Peibo, D., Lie, J., Zongyuan, G.: Nurvid: A large expert-level video database for nursing procedure activity understanding. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "43. Ni, Z.L., Bian, G.B., Zhou, X.H., Hou, Z.G., Xie, X.L., Wang, C., Zhou, Y.J., Li, R.Q., Li, Z.: Raunet: Residual attention u-net for semantic segmentation of cataract surgical instruments. In: International Conference on Neural Information Processing (ICONIP). pp. 139-149. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.272 + ], + "angle": 0, + "content": "44. Nwoye, C.I., Gonzalez, C., Yu, T., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23. pp. 364-374. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "45. Nwoye, C.I., Padoy, N.: Data splits and metrics for benchmarking methods on surgical action triplet datasets. arXiv preprint arXiv:2204.05235 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.355 + ], + "angle": 0, + "content": "46. Nwoye, C.I., Yu, T., Gonzalez, C., Seeliger, B., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. Medical Image Analysis 78, 102433 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.398 + ], + "angle": 0, + "content": "47. Pan, X., Gao, X., Wang, H., Zhang, W., Mu, Y., He, X.: Temporal-based swim transformer network for workflow recognition of surgical video. Int. J. Comput. Assist. Radiol. Surg. 18(1), 139-147 (Jan 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.452 + ], + "angle": 0, + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "49. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.55 + ], + "angle": 0, + "content": "50. Ross, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Comparative validation of multi-instance instrument segmentation in endoscopy: results of the robust-mis 2019 challenge. Medical image analysis 70, 101920 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "51. Ross, T., Zimmerer, D., Vemuri, A., Isensee, F., Wiesenfarth, M., Bodenstedt, S., Both, F., Kessler, P., Wagner, M., Müller, B., et al.: Exploiting the potential of unlabeled endoscopic video data with self-supervised learning. International journal of computer assisted radiology and surgery 13, 925-933 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.77 + ], + "angle": 0, + "content": "52. Rof, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Arbelaez, P., Bian, G.B., Bodenstedt, S., Bolmgren, J.L., Bravo-Sanchez, L., Chen, H.B., Gonzalez, C., Guo, D., Halvorsen, P., Heng, P.A., Hosgor, E., Hou, Z.G., Isensee, F., Jha, D., Jiang, T., Jin, Y., Kirtac, K., Kletz, S., Leger, S., Li, Z., Maier-Hein, K.H., Ni, Z.L., Riegler, M.A., Schoeffmann, K., Shi, R., Speidel, S., Stenzel, M., Twick, I., Wang, G., Wang, J., Wang, L., Wang, L., Zhang, Y., Zhou, Y.J., Zhu, L., Wiesenfarth, M., Kopp-Schneider, A., Müller-Stich, B.P., Maier-Hein, L.: Comparative validation of multi-instance instrument segmentation in endoscopy: Results of the robust-mis 2019 challenge. Medical Image Analysis 70, 101920 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101920, https://www.sciencedirect.com/science/article/pii/S136184152030284X" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "53. Sato, F., Hachiuma, R., Sekii, T.: Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6471-6480 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "54. Schoeffmann, K., Husslein, H., Kletz, S., Petscharnig, S., Munzer, B., Beecks, C.: Video retrieval in laparoscopic video recordings with dynamic content descriptors." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "Multim. Tools Appl. 77(13), 16813-16832 (2018). https://doi.org/10.1007/s11042-017-5252-2, https://doi.org/10.1007/s11042-017-5252-2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.786, + 0.258 + ], + "angle": 0, + "content": "55. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: César, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018). https://doi.org/10.1145/3204949.3208137, https://doi.org/10.1145/3204949.3208137" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "56. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18857-18866 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "57. Shi, X., Jin, Y., Dou, Q., Heng, P.A.: LRTD: long-range temporal dependency based active learning for surgical workflow recognition. Int. J. Comput. Assist. Radiol. Surg. 15(9), 1573-1584 (Sep 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.397 + ], + "angle": 0, + "content": "58. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 510-526. Springer International Publishing, Cham (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.48 + ], + "angle": 0, + "content": "59. Smedsrud, P.H., Thambawita, V., Hicks, S.A., Gjestang, H., Nedrejord, O.O., Naess, E., Borgli, H., Jha, D., Berstad, T.J.D., Eskeland, S.L., Lux, M., Espeland, H., Petlund, A., Nguyen, D.T.D., Garcia-Ceja, E., Johansen, D., Schmidt, P.T., Toth, E., Hammer, H.L., de Lange, T., Riegler, M.A., Halvorsen, P.: Kvasir-Capsule, a video capsule endoscopy dataset. Scientific Data 8(1), 142 (2021). https://doi.org/10.1038/s41597-021-00920-z" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.508 + ], + "angle": 0, + "content": "60. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.55 + ], + "angle": 0, + "content": "61. Spaeth, G., Danesh-Meyer, H., Goldberg, I., Kampik, A.: Ophthalmic Surgery: Principles and Practice E-Book. Elsevier Health Sciences (2011), https://books.google.com.hk/books?id=wHWMUGH-5csC" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.591 + ], + "angle": 0, + "content": "62. Stauder, R., Ostler, D., Kranzfelder, M., Koller, S., Feußner, H., Navab, N.: The tum lapchole dataset for the m2cai 2016 workflow challenge. arXiv preprint arXiv:1610.09278 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.647 + ], + "angle": 0, + "content": "63. Tian, Y., Pang, G., Liu, F., Liu, Y., Wang, C., Chen, Y., Verjans, J., Carneiro, G.: Contrastive transformer-based multiple instance learning for weakly supervised polyp frame detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 88-98. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "64. Tran, D., Wang, H., Torresani, L., Feiszli, M.: Video classification with channel-separated convolutional networks (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.73 + ], + "angle": 0, + "content": "65. Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., de Mathelin, M., Padoy, N.: Endonet: A deep architecture for recognition tasks on laparoscopic videos. IEEE Transactions on Medical Imaging (TMI) 36, 86-97 (2016), https://api-semanticscholar.org/CorpusID:5633749" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.785 + ], + "angle": 0, + "content": "66. Twinanda, A.P., Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Rsdnet: Learning to predict remaining surgery duration from laparoscopic videos without manual annotations. IEEE Transactions on Medical Imaging 38(4), 1069-1078 (2019). https://doi.org/10.1109/TMI.2018.2878055" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "67. Wagner, M., Müller-Stich, B.P., Kisilenko, A., Tran, D., Heger, P., Mündermann, L., Lubotsky, D.M., Müller, B., Davitashvili, T., Capek, M., et al.: Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis 86, 102770 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "M. Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "68. Wang, T., Li, H., Pu, T., Yang, L.: Microsurgery robots: Applications, design, and development. Sensors 23(20) (2023). https://doi.org/10.3390/s23208503, https://www.mdpi.com/1424-8220/23/20/8503" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "69. Wang, X., Zhang, S., Qing, Z., Shao, Y., Gao, C., Sang, N.: Self-supervised learning for semi-supervised temporal action proposal. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.273 + ], + "angle": 0, + "content": "70. Wang, Z., Lu, B., Long, Y., Zhong, F., Cheung, T.H., Dou, Q., Liu, Y.: Autolaparo: A new dataset of integrated multi-tasks for image-guided surgical automation in laparoscopic hysterectomy. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 486-496. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "71. Wu, W., Sun, Z., Ouyang, W.: Revisiting classifier: Transferring vision-language models for video recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.356 + ], + "angle": 0, + "content": "72. Wu, W., Wang, X., Luo, H., Wang, J., Yang, Y., Ouyang, W.: Bidirectional cross-modal knowledge exploration for video recognition with pre-trained vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.398 + ], + "angle": 0, + "content": "73. Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Less is more: Surgical phase recognition with less annotations through self-supervised pre-training of cnn-lstm networks. arXiv preprint arXiv:1805.08569 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.481 + ], + "angle": 0, + "content": "74. Yu, F., Silva Croso, G., Kim, T.S., Song, Z., Parker, F., Hager, G.D., Reiter, A., Vedula, S.S., Ali, H., Sikder, S.: Assessment of Automated Identification of Phases in Videos of Cataract Surgery Using Machine Learning and Deep Learning Techniques. JAMA Network Open 2(4), e191860-e191860 (04 2019). https://doi.org/10.1001/jamanetworkopen.2019.1860, https://doi.org/10.1001/jamanetworkopen.2019.1860" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "75. Yu, T., Mutter, D., Marescaux, J., Padoy, N.: Learning from a tiny dataset of manual annotations: a teacher/student approach for surgical phase recognition. arXiv preprint arXiv:1812.00033 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "76. Yuan, K., Srivastav, V., Navab, N., Padoy, N.: Hecvl: Hierarchical videolanguage pretraining for zero-shot surgical phase recognition. arXiv preprint arXiv:2405.10075 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "77. Yuan, K., Srivastav, V., Yu, T., Lavanchy, J., Mascagni, P., Navab, N., Padoy, N.: Learning multi-modal representations by watching hundreds of surgical video lectures (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.66 + ], + "angle": 0, + "content": "78. Zha, R., Cheng, X., Li, H., Harandi, M., Ge, Z.: Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 13-23. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "79. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: European Conference on Computer Vision. LNCS, vol. 13664, pp. 492-510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.744 + ], + "angle": 0, + "content": "80. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI Conference on Artificial Intelligence. pp. 7590-7598 (2018), https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/17344" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.744 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_origin.pdf b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c3f88b31ca015ffddbb7910249012de98ac17da1 --- /dev/null +++ b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/fe0409c2-8f5a-483f-b951-b1d166e6eb57_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:615b2ca2cfaaacfd7aadff9ebb42f7eeeabb11c0f7c9a8ae42b47b76821c08f7 +size 1445269 diff --git a/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/full.md b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..07e519248176b7650a9e429d83d3156c54adfbd4 --- /dev/null +++ b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/full.md @@ -0,0 +1,281 @@ +# OphNet: A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding + +Ming Hu $^{1,2,3*}$ , Peng Xia $^{1,3*}$ , Lin Wang $^{5*}$ , Siyuan Yan $^{1,2}$ , Feilong Tang $^{1,3}$ Zhongxing Xu $^{7}$ , Yimin Luo $^{6}$ , Kaimin Song $^{3}$ , Jurgen Leitner $^{2}$ Xuelian Cheng $^{1}$ , Jun Cheng $^{8}$ , Chi Liu $^{9}$ , Kaijing Zhou $^{4\dagger}$ , and Zongyuan Ge $^{1,2,3\dagger}$ + +$^{1}$ AIM Lab, Faculty of IT, $^{2}$ Faculty of Engineering, Monash University + $^{3}$ Airdoc-Monash Research, Airdoc $^{4}$ Eye Hospital, Wenzhou Medical University + $^{5}$ Bosch Corporate Research $^{6}$ King's College London $^{7}$ Cornell University Institute for Infocomm Research, A*STAR + $^{9}$ Faculty of Data Science, City University of Macau + +Abstract. Surgical scene perception via videos is critical for advancing robotic surgery, telesurgery, and AI-assisted surgery, particularly in ophthalmology. However, the scarcity of diverse and richly annotated video datasets has hindered the development of intelligent systems for surgical workflow analysis. Existing datasets face challenges such as small scale, lack of diversity in surgery and phase categories, and absence of time-localized annotations. These limitations impede action understanding and model generalization validation in complex and diverse real-world surgical scenarios. To address this gap, we introduce OphNet, a large-scale, expert-annotated video benchmark for ophthalmic surgical workflow understanding. OphNet features: 1) A diverse collection of 2,278 surgical videos spanning 66 types of cataract, glaucoma, and corneal surgeries, with detailed annotations for 102 unique surgical phases and 150 fine-grained operations. 2) Sequential and hierarchical annotations for each surgery, phase, and operation, enabling comprehensive understanding and improved interpretability. 3) Time-localized annotations, facilitating temporal localization and prediction tasks within surgical workflows. With approximately 285 hours of surgical videos, OphNet is about 20 times larger than the largest existing surgical workflow analysis benchmark. Code and dataset are available at: https://minghuo830.github.io/OphNet-benchmark/. + +Keywords: Surgical Workflow Understanding $\cdot$ Ophthalmic Surgery $\cdot$ Medical Image Analysis $\cdot$ Video Benchmark + +# 1 Introduction + +As surgical robot platforms such as the da Vinci® surgical system become increasingly sophisticated, there is growing interest in integrating enhanced intelligi- + +gence into scenarios like minimally invasive surgery [20,32,68]. The advancements in machine vision perception empower these robotic systems to autonomously recognize and adapt to the intricacies of surgical environments, without relying on binary instrument usage signals, RFID tags, sensor data from tracking devices, or other signal information that necessitates laborious manual annotations or additional equipment installations [28]. This autonomy includes the capacity to identify anatomical structures, detect anomalies, and adjust surgical plans in real-time, which is crucial in dynamic and unpredictable surgical settings. In recent years, especially in endoscopy and ophthalmic surgery, the application of deep learning has demonstrated considerable promise in bolstering these autonomous capabilities. This encompasses the analysis of surgical workflows [6, 28, 74], segmentation of instruments and anatomy [4, 24, 43], and depth estimation [78], among others. + +Automatic video surgical workflow understanding is a fundamental yet challenging problem for developing computer-assisted and robotic-assisted surgery, which can be divided into internal (e.g., laparoscopic and endoscopic [28,47,57]) and external (e.g., operating room and nursing procedure [41]) analysis. In addition to promoting the development of intelligent surgery, it also greatly benefits surgical documentation, education, and training [11,12,65]. Baret et al. [6] showed networks, like Inflated 3D ConvNet (I3D) [10] that utilize spatiotemporal convolutions, require a relatively extensive dataset for effective training. In their study, the model achieves an accuracy exceeding $80\%$ when trained on 100 videos, with a progressive improvement as the sample size surpasses 700. However, the highly efficient and rapidly evolving deep learning technologies for surgical workflow analysis are currently limited by the following shortcomings in current video benchmarks: 1) Small-scale: the majority of video datasets contain no more than 100 videos. For example, the CATARACTS [23] and CatRelDet [23] datasets contain only 50 and 21 surgical videos, respectively. These datasets are relatively small, insufficient for large-scale validation. 2) Limited categories of surgeries and phases: almost all ophthalmic surgical video datasets only include cataract surgery and do not further classify specific types of surgeries. Additionally, the number of phase categories is also limited, like CatRelDet [23] only contains 4 different phase labels, which is insufficient to meet the requirements for evaluation in real clinical environments. 3) Coarse-grained annotation: due to annotation costs, existing benchmarks often have coarse-grained action definitions. For example, adhesive injection may occur in two different phases: main incision and capsulorhexis, so it may be classified into different phase categories. Coarse-grained action definitions may lead to annotation bias. 4) Single time-boundary annotation: they only annotate designated phases in the videos, ignoring the continuity across different stages of ophthalmic surgery, as well as the hierarchical relationship between surgery, phase, and operation. Simpler datasets, such as LensID [22], are limited to binary classification tasks distinguishing lens implantation from other irrelevant phases. 5) Uniform domain: the videos are meticulously collected, and + +
ProtocolDataset PropertiesTasks
DatasetsNo. of VideosNo. of Action SegmentsNo. of Surgery CategoriesNo. of Action CategoriesTotal DurationMulti-Surgery Presence RecognitionPhase RecognitionPhase LocalizationPhase Prediction
Endo&LapCholec120 [45]120-1776.2hX✓XX
SurgicalActions160 [54]1601601160.2hX✓XX
HeiCo [39, 50]30-3142.8hX✓✓✓
EndoVis 2021 [67]332501722.0hX✓XX
PitVis [3]2528711733.3hX✓XX
CholecT50 [46]50-11044.7hX✓✓✓
AutoLaparo [70]213001723.1hX✓✓✓
OphScopeLensID [22]1002,4401211.7hX✓XX
Cataract-101 [55]1011,26611014.0hX✓✓✓
CatRelDet [23]212,400142.0hX✓XX
CATARACTS [4]501,53611920.0hX✓✓✓
Cataract-1K [21]1,000931112118.7hX✓✓✓
OphNet(Ours)2,2789,79566150284.8h✓✓✓✓
+ +Table 1: The statistics comparison among existing workflow analysis datasets and our OphNet. Compared to other datasets, OphNet focuses on more comprehensive coverage of various surgery, phase and operation categories, collects a large number of videos, totaling 284.8 hours, and also enables a variety of recognition, localization and prediction tasks. OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. For instance, Cholec120 [45], Cholec80 [65], m2cai-workflow and LapChole [62] form one series, whereas CholecT50 [46], CholecT45 [46], and CholecT40 [44] comprise another series. We have excluded the following scenarios from our comparison: (1) non-open-source datasets such as Bypass170 [66], ESD [30], Yu's [74], etc.; (2) a superset of multiple open-source or non-open-source datasets, like Cholec207 [6], etc.; (3) datasets employed for lesion, anatomy, and instrument classification and segmentation, such as SUN-SEG [27], CVC-ClinicDB [7], ROBUSTMIS [52], Mesejo's [40], Cata7 [43], etc., anomaly detection such as PolypDiag [63] (from Hyper-Kvasir [9] and LDPolypVideo [38]), Kvasir-Capsule [59], etc., and other datasets not dedicated to workflow analysis. It's worth mentioning that even in comparison with the above datasets, OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. Endo&Lap denotes the endoscopic and laparoscopic protocol, OphScope denotes the ophthalmic microscope protocol. We choose the latest version for comparison in cases where datasets have multiple supplementary updates. + +while this ensures video quality, the uniform style is not conducive to testing the model's domain generalization ability. + +While some works have explored semi-supervised and self-supervised learning strategies [8, 51, 73, 75] to alleviate the cost of annotations or use only a small fraction of available labels, these approaches still lack competitiveness in performance compared to fully supervised learning. This deficiency in performance is impeding the widespread clinical application of these strategies. To address the shortage of sufficient labeled datasets, we construct OphNet, a large-scale + +and expert-level video benchmark with high diversity, for ophthalmic surgical workflow understanding. The main advantages of OphNet are as follows: + +- Largest scale and diversity: to the best of our knowledge, OphNet is currently the largest and most richly labeled dataset for surgical workflow analysis. It contains a number of videos 20 times greater than the current largest benchmark in ophthalmic surgery and far exceeds datasets in more established fields such as endoscopy. Additionally, OphNet includes the greatest variety of different types of surgeries, encompassing 66 different surgeries such as cataract, glaucoma, and corneal surgeries, along with 102 unique surgical phases and 150 distinct operations. This diversity significantly surpasses that of previous research. +- Fine-grained, sequential and hierarchical annotation: we have meticulously selected a subset of videos for annotation localization, with each video being annotated for an average of 22 operations. Additionally, we provide exquisite annotations at the levels of surgery, phase, and operation, catering to the requirements for training specific challenge models. This annotation design aims to offer a multifaceted understanding of surgical protocols, accommodate the nuances of each distinct surgery, and enhance the usability and interpretability of our dataset. +- Expert-level manual annotation: the annotation work for OphNet was completed by ten experienced ophthalmologists and five individuals with ophthalmic experience, encompassing, but not limited to, standardization of definitions for surgery, phase, and operation labels, video filtering, classification and localization annotations, and secondary verification, among others. Expert-level annotators ensure the quality and professionalism of OphNet. + +# 2 Related Work + +Surgical Workflow Understanding. Beyond its therapeutic advantages, minimally invasive surgery also provides the capability for operative video recording. These videos can be stored and later utilized for various purposes such as cognitive training, skill assessment, and surgical workflow analysis [76, 77]. Techniques derived from the broader field of video content analysis and representation are increasingly being incorporated into the surgical realm [6, 36, 74]. A typical surgical workflow can be defined by a sequence of tasks or events, including patient positioning, incision, dissection, and suturing. These events are influenced not only by the specific type of surgery but also by the individual surgeon's proficiency and technique. Consequently, a comprehensive understanding of the surgical workflow necessitates a thorough examination of the temporal, spatial, and contextual facets of these tasks. + +Weakly-Supervised Video Learning. Substantial pioneering work has been undertaken in the realms of video understanding [5,16,33,53,69,71,72,79]. Since even a small amount of videos easily comprises several million frames, methods that do not rely on a frame-level annotation are of special importance. Weakly-supervised video learning makes use of loosely labeled data to train models, + +thereby obviating the necessity for exhaustively annotated training data. These weak labels might manifest in a variety of forms, including video-level labels or partial labels, which are less specific than frame-level or pixel-level annotations. The objective of weakly-supervised learning is to utilize these coarse labels to produce models capable of delivering fine-grained predictions, such as temporally precise action localization or detailed semantic segmentation [15,35]. The large-scale annotation of surgical videos demands a significant investment of invaluable medical time resources. In this context, weakly supervised learning emerges as a viable solution to this bottleneck. + +# 3 Dataset Construction + +In this section, we detail the construction of our dataset, which involved meticulous data collection and preprocessing. YouTube was leveraged as a primary source to circumvent privacy issues while ensuring a broad representation of ophthalmic surgeries. Our selection criteria aimed to capture a wide range of video qualities and styles, specifically targeting cataract, glaucoma, and corneal surgeries due to their prevalence in clinical settings. We refined the dataset by excluding videos of inadequate quality or those depicting non-human subjects. The annotation process was designed to reflect the complex nature of eye surgeries, incorporating hierarchical classification to account for multiple conditions often treated within a single procedure. This was complemented by detailed localization annotations, delineating the distinct phases and techniques characteristic of ophthalmic operations, all undertaken by a dedicated team of ophthalmologists to ensure accuracy and relevance. + +# 3.1 Data Collection & Preprocessing + +Collection. Medical data exhibit unique privacy considerations and tend to be of smaller data volumes, characteristics that are particularly salient in the case of video data. By capitalizing on the wealth of surgical videos available on YouTube, we are able to obviate potential ethical and privacy concerns while concurrently enabling the rapid procurement of a substantial corpus of videos for further screening [1,2,17,29]. To fulfill this objective, we deploy text-based search algorithms to probe each surgery on YouTube, obtaining videos with titles that incorporate the requisite surgical keywords. We select cataract, glaucoma, and corneal surgery—three of the most commonly performed ophthalmic surgeries in actual clinical environments—as the central subjects of our research. To expand our video collection, we bolster our search queries by integrating synonyms and abbreviations associated with each type of ophthalmic surgery. For instance, cataract surgery encompasses various types: Phacoemulsification (abbr. PHACO), Intraocular Lens implantation (abbr. IOL), and Extracapsular Cataract Extraction (abbr. ECCE), etc. + +Preprocessing. Given the nature of text-based retrieval and the varying quality, style, and filming methods of surgical videos influenced by different YouTube + +![](images/6e3ce91aebec06a16e145c80ecb0120a0c0fee542ad3cc25d3237b18b57910a9.jpg) +Fig. 1: The figure shows two combined surgical videos, PHACO + IOL implantation and PHACO + Trabeculotomy + Transscleral Cyclophotocoagulation + IOL implantation. For each frame marked in color, we provide time-boundary annotations at surgical, phase and operation levels. + +sources, some videos may exhibit poor quality or deviations in surgical representation. We filter out low-resolution videos, black-and-white color schemes, and animated demonstrations. Furthermore, surgical videos featuring non-human eyes, such as pig eyes, rabbit eyes, or pseudo-human eyes, are also outside the scope of our annotation. The data preprocessing was jointly completed by five professionals with experience in ophthalmology for initial screening, followed by a review conducted by six attending ophthalmologists. + +# 3.2 Data Annotation + +Classification Annotation. Different from the single-label classification in most natural video datasets [17, 25, 58, 60, 80], in ophthalmic surgery, physicians typically consider and implement multiple types of surgeries and phases based on the specific conditions and needs of the patient. This multifaceted approach stems from the complexity of the eye as an organ, where numerous diseases often coexist. Consequently, multiple ocular problems may need to be addressed in a single surgery. As shown in Fig. 1, a patient with cataracts may also have glaucoma, necessitating the simultaneous treatment of both conditions within one surgery. Hence, we initially classified all videos based on the primary surgical categories into three principal groups: cataract, glaucoma, and corneal + +![](images/7b52aaf4761e229370e29937aee0afdec046a214687e9c8f1afaa73289bf3a5b.jpg) + +![](images/fcd339a165035909f0c64709c414835b5183a983fcd5fd796b93494db91028e5.jpg) + +![](images/5382f905733380884d8f3a80587ad66136d6c32d2e538c9aacfc4404a4bef176.jpg) +(a) + +![](images/a6bbd279c72da86950d440012c84e0729166b52cc2ee02691d1627955ac4675d.jpg) +Capsulorhexis + +![](images/f53fade19af4480e7439ab961a92878738e07a1a7ce302ccdcefd63bcf56a13c.jpg) +(b) +Capsule Staining + +![](images/384b32f6ce0db429a9014cfe1b54426b62f743fd98ea4ea28d59c0b97219757e.jpg) +Hydrodissection and Float Nucleus + +![](images/b2e6c9e169eeb878b093e02d2e3546f7f344313126e52b0a16e736a10ddf5ac8.jpg) +Main Incision +Cortical Aspiration +Fig. 2: OphNet's composition, comparison with other datasets for the same task, and some phase examples: (a) an overview of the composition ratios at the levels of surgery, phase, and operation; (b) comparison among existing open-source $\odot$ laparoscopic & endoscopic, and $\odot$ ophthalmic microscope workflow analysis video datasets and $\odot$ our OphNet. OphNet stands as the largest real-world video dataset for ophthalmic surgical workflow understanding, featuring the highest number of videos, longest duration, and diverse categories of surgeries and phases; (c) eight phase examples in OphNet. + +![](images/c3f077bd2d823d8e6b0fe2ed77678c81ae2141c8569afde47d4f9692296618af.jpg) +IOL Implantation +(c) + +![](images/9fa9450d546ce9bf0f96787f22bf2db0bc9f820be34ec429d30591de819b263a.jpg) +Cautery/Hemostasis + +![](images/5633365ecc0620f91407220347e404338336258036c2a686fd8414bddf14b1f2.jpg) +Phacoemulsification + +surgeries. Subsequently, these were further allocated for the detailed annotation of primary and secondary surgery. It is important to note that there exists only a single type of primary surgery, whereas multiple secondary surgery may be present. The categorization and annotation were meticulously completed by a team comprised of 8 experienced ophthalmologists. + +Hierarchical Localization Annotation. A single surgery is often multifaceted, involving a series of intricate phases, each requiring distinct techniques and instruments, and the transition between surgeries entails high precision and coordination. Routine cataract surgery involves several steps. It begins with the administration of anesthesia, followed by a small incision in the cornea. The surgeon then creates an opening in the lens capsule and uses ultrasonic vibrations to break up and remove the cataract. Afterward, an artificial intraocular lens (IOL) is inserted into the lens capsule, and the incision is sealed without stitches. We define the phases and operation for various surgeries based on the + +![](images/be64c5f57da297c9fbce7bc783237ab0b093c29b47faf1b30bccdf2c10f1511c.jpg) +(a) phase statistics + +![](images/b436e3316441290ec1749f9081749dcef4e6f24d32f53b15df7b8bb989cd5f09.jpg) +(b) operation statistics +Fig. 3: We present the data statistics of trimmed videos at the levels of phase and operation, including the number of trimmed videos, average duration, and total duration. The IDs and corresponding names can be found in the appendix. + +textbook Ophthalmic Surgery: Principles and Practice [61]. To ensure the quality of localization annotations, we assign each annotator videos of 2 different types of primary surgeries. For each action time-boundary annotation, we annotate at three different levels of granularity: surgery, phase, and operation. We also employed the complete linkage algorithm [13] to cluster and merge various temporal boundaries into stable boundaries that received multiple agreements. It's worth noting that an individual video may have multiple separate instances of the same or different phases, thereby leading to multiple boundary definitions. A team of 15 ophthalmologists completes the localization annotation. + +# 3.3 Dataset Statistics and Analysis + +OphNet includes 2,278 surgical videos (284.8 hours), demonstrating 66 different types of ophthalmic surgeries: 13 types of cataract surgery, 14 types of glaucoma surgery, and 39 types of corneal surgery. There are 102 phases and 150 operations for recognition, detection and prediction tasks, summarized in Tab. 1. Over $77\%$ of videos have high-definition resolutions of $1280 \times 720$ pixels or higher. To facilitate algorithm development and evaluation, we selected 523 videos for localization annotations. Additionally, we trim videos according to annotated action boundaries, resulting in 7,320 phase instances and 9,795 operation instances, totaling 51.2 hours. The average duration of trimmed videos is 32 seconds, while untrimmed videos average 337 seconds. + +# 4 Experiments + +In our study, we explore four potential tasks using the OphNet dataset: 1) primary surgery presence recognition, 2) phase and operation recognition, 3) phase + +localization, and 4) phase anticipation. To establish robust baselines, we employ state-of-the-art models known for their effectiveness in human action recognition, detection and anticipation. For each task, we provide a detailed problem formulation and evaluate the baseline models' performance. Our findings offer valuable insights into video understanding within sequences and fine granularity, contributing to the field's knowledge of video tasks in medical contexts. + +# 4.1 Surgery Presence, Phase and Operation Recognition + +Task Description. Surgery presence recognition focuses on identifying various surgical types in untrimmed videos through weakly supervised methods. This requires the model to discern and capture distinct surgical action features within extensive video footage. In OphNet, all types of surgeries occurring in each video are annotated, but only a subset of the videos have time-boundary annotations, as detailed in Sec. 3.2. We identify the primary surgery based on the key objectives and durations of the surgeries. To streamline the process, our experiments are limited to recognizing the presence of these primary surgeries. Additionally, phase recognition segments the surgery into distinct phases using visual cues and movements, such as incision, lens removal, and implantation. Operation recognition involves identifying finer-grained surgical actions. + +Baselins. We compare the performance of I3D [10], SlowFast [19], X3D [18], and MViT V2 [31] models on this task. These models are evaluated in two versions: 1) random initialization training and 2) pretraining with weights from Kinetics 400 [29], which is a human action recognition dataset. In addition, we also explored the classification performance of X-CLIP [42] and ViFi-CLIP [49], two CLIP [48]-based models. We also compared the effects of different numbers of input frames on the models' performance, where the sub + +
BaselinesPrimary Surgery Classification
CataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
I3D [10]38.986.343.671.742.778.229.853.2
SlowFast [19]45.382.144.672.345.577.327.254.4
X3D [31]42.187.444.674.243.876.628.562.7
MViT V2 [18]43.284.245.581.845.575.929.160.1
I3D [10]36.884.748.281.848.676.527.250.6
SlowFast*49.083.747.380.949.275.627.250.6
X3D*47.486.346.481.548.378.235.461.4
MViT V2*44.285.349.181.847.777.328.563.3
X-CLIP16 [42]58.594.751.892.861.488.640.579.0
X-CLIP3260.692.653.583.756.884.158.981.0
ViFi-CLIP16 [49]59.688.352.480.261.481.858.979.8
ViFi-CLIP3259.688.351.584.850.075.056.377.2
+ +Table 2: Per-class Top-1 and Top-5 accuracy (%) for the primary surgery presence recognition on untrimmed videos. The best performance for each split has been highlighted in bold. + +script _16 represents an input of 16 frames, and _32 represents an input of 32 frames. + +Setup. The dataset was randomly partitioned to ensure a balanced representation of examples for each surgery category. Specifically, we allocated $70\%$ of the data for training (1,449 surgical videos), $10\%$ for validation (205 surgical videos), and $20\%$ for testing (424 surgical videos). For the phase and operation recognition experiments, we followed the same settings, with $70\%$ of the data used for training (5,024 phase segments, 6,856 operations segments), $10\%$ for + +validation (730 phase segments, 975 operations segments), and $20\%$ for testing (1,566 phase segments, 1,964 operations segments). The input for the surgery presence recognition experiment is untrimmed videos, while for the phase and operation recognition experiments, the input consists of trimmed segments. In all classification experiments, we set up the analysis and comparison from four perspectives: cataract surgery, glaucoma surgery, corneal surgery, and all surgical videos. + +Results. We summarize the results in Tab. 2 and Tab. 3. In primary surgery classification, X-CLIP [42] achieved the highest overall Top-1 accuracy at $58.9\%$ , leading in cataract and glaucoma surgeries with accuracies of $60.6\%$ and $53.5\%$ respectively. For corneal surgeries, X-CLIP also recorded the highest Top-1 and Top-5 accuracies of $61.4\%$ and $88.6\%$ . In phase classification, ViFi-CLIP [49] led with the highest Top-1 accuracy, particularly in cataract surgeries at $75.9\%$ , and exhibited the best performance in both glaucoma and corneal surgeries. Furthermore, in operation classification, ViFi-CLIP outperformed all other models in all categories, especially noted in cataract and corneal surgeries with Top-1 accuracies of $75.1\%$ and $83.7\%$ and Top-5 accuracies of $93.8\%$ and $85.2\%$ respectively. Overall, ViFi-CLIP showed superior performance in both phase and operation classifications across various surgery types. Besides, in the phase and operation classification experiments, a higher number of input frames generally had a positive effect on the model. In Fig. 4, we present the heatmap visualization of four examples from the test set of phase recognition experiments using the ViFi-CLIP model. It can be observed that in a series of frame images, the model focuses on surgical instruments and the operated eye area, which is consistent with human experience. The instruments used in the same phase of ophthalmic surgery are often similar. + +# 4.2 Phase Localization + +Task Description. Phase localization in ophthalmic surgical workflow analysis refers to the task of pinpointing the exact moments or time intervals within a surgical video where specific phases of the surgery begin and end. This involves the detailed temporal segmentation of the entire surgical procedure into its constituent phases based on visual cues, surgeon's actions, and the progression of the surgery. The objective of phase localization is to accurately identify the start and end times of different surgical stages, such as pre-operative preparation, incision, and removal of the lens facilitating a granular and precise understanding of the surgery timeline. This task is crucial for detailed surgical documentation, efficient surgical training, and the development of targeted interventions during specific stages of the surgery, enhancing overall surgical management and post-operative analysis. + +Baselines. We conducted the experiments using phase-level labels and used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. Data split follows the setup of the primary surgery classification experiment. + +
BaselinesPhase ClassificationOperation Classification
CataractGlaucomaCorneaAllCataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1
I3D [10]27.255.724.157.518.952.125.758.226.854.923.556.018.051.225.057.1
SlowFast [19]26.556.523.156.524.249.126.760.125.855.222.945.923.548.526.059.0
X3D [31]27.058.321.055.521.828.526.662.326.447.220.544.621.327.826.161.5
MViT V2 [18]26.254.921.053.426.046.827.059.825.943.820.542.725.545.926.558.9
I3D*29.568.922.158.626.050.930.271.228.867.521.847.925.750.329.560.0
SlowFast*30.672.325.254.730.759.831.761.829.971.124.843.929.558.730.560.9
X3D*27.272.922.159.630.761.533.563.226.571.821.748.829.960.232.862.1
MViT V2*34.276.523.352.038.465.128.360.233.575.222.841.537.964.027.859.5
X-CLIP16 [42]68.392.247.389.853.077.463.485.367.591.046.578.982.276.162.584.0
X-CLIP3269.194.048.781.754.880.462.785.868.093.047.980.584.079.562.084.7
ViFi-CLIP16 [49]75.993.740.485.466.681.666.188.474.592.542.874.585.080.565.087.5
ViFi-CLIP3273.092.949.682.757.781.668.487.275.193.843.280.283.785.264.886.5
+ +Table 3: Per-class Top-1 and Top-5 accuracy (\%) for the primary surgery presence recognition on untrimmed videos and phase recognition on trimmed videos. * denotes the initialization from the model pre-trained on Kinetics 400 [29]. For the two CLIP models, we chose ViT-B/16 as the backbone and compared the performance of two different input frame numbers, 16 and 32. The best performance for each split has been highlighted in bold. + +
BaselinesBackbonesmAP (%)
0.10.30.50.7Avg.
ActionFormerCSN [64]53.750.140.624.542.5
SwinViviT [34]59.354.743.326.346.4
SlowFast [19]60.055.945.126.047.5
TriDet [56]CSN56.153.043.129.446.2
SwinViviT61.057.147.133.150.4
SlowFast61.356.045.630.448.6
+ +Table 4: The results for phase detection. Action-Former and TriDet are state-of-the-art models for human action detection tasks, and we use three different backbones for feature extraction and report mAP at the IoU thresholds of [0.1:0.2:0.9]. Average mAP is computed by averaging different IoU thresholds. The best performance for each split has been highlighted in bold. + +
BaselinesTop-1 Acc. (%)
0.10.30.50.7Avg.
I3D [10]26.542.249.851.347.3
SlowFast [19]25.442.648.952.247.2
MViT V2 [18]25.643.749.352.347.5
I3D*27.343.550.151.447.6
SlowFast*27.543.249.952.347.8
MViT V2*27.843.850.551.748.2
+ +Table 5: The results for phase anticipation. We report top-1 accuracy at the observation ratios [0.1:0.2:0.9]. Average top-1 accuracy is computed by averaging different observation ratios. The best performance for each split has been highlighted in bold. + +Setup. We excluded Operation Gap and Invalid, and filtered out tags with fewer than 20 segments. We used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. To extract features from the videos, we first extracted RGB frames from each video at a rate of 25 frames per second. We also extracted optical flow using the TV-L1 [26,37] algorithm. We then fine-tuned an I3D [10] model that had been pre-trained on the ImageNet [14] dataset, and used it to generate features for each RGB and optical flow frame. Because each video has a variable duration, we performed uniform interpolation to generate 100 fixed- + +![](images/9c623f3ddfa6d3398ad7f45a6022f8c006b498813324c54bfce7ac270ce7be7b.jpg) +Fig. 4: Attention map visualizations of ViFi-CLIP [49] on four examples from OphNet's test set in the phase recognition task. + +![](images/59a39a3f1b6cd39650adff69bb5d387e3de20e500fa779540ce4ad29512e2b1f.jpg) + +length features for each video. Finally, we concatenated the RGB and optical flow features into a 2048-dimensional embedding, which served as the input for our model. + +Results. The experimental results for phase localization are presented in the Tab. 4, showcasing the performance of different baseline models with various backbones in terms of mean Average Precision (mAP) at different Intersection over Union (IoU) thresholds [0.1:0.2:0.9]. The models evaluated include Action-Former with SwinViviT and SlowFast backbones, and TriDet with CSN, SwinViviT, and SlowFast backbones. The results indicate that the TriDet model with a SwinViviT backbone outperforms other combinations, achieving the highest mAP scores across most IoU thresholds, with notable scores of $61.0\%$ (IoU=0.1), $57.1\%$ (IoU=0.3), $47.1\%$ (IoU=0.5), and $33.1\%$ (IoU=0.7), resulting in an average mAP of $50.4\%$ . This indicates that the TriDet model, especially when combined with the SwinViviT backbone, is particularly effective for phase localization in surgical videos. On the other hand, the TriDet model with a SlowFast backbone shows competitive performance, particularly achieving the highest mAP of $61.3\%$ at the lowest IoU threshold (0.1). However, it falls slightly behind in performance at higher IoU thresholds compared to the SwinViviT backbone. + +# 4.3 Phase Anticipation + +Task Description. This task requires the analysis of real-time or recorded video data to foresee the sequence of events based on current and past surgical activities. By understanding the typical progression of ophthalmic surgeries and recognizing patterns in the surgeon's actions and the use of instruments, the system aims to forecast the next phase of the surgery, allowing for proactive preparation and response. The objective of phase anticipation is to enhance the efficiency and safety of surgical procedures by providing the surgical team with advanced notice of upcoming steps, enabling better resource allocation, timing for critical tasks, and overall coordination within the operating room. + +![](images/57d2cb94b9d2a79c200a69e5e202df35fe28033b47254634482f624b18db31f8.jpg) +Fig. 5: Phase localization visualization of TriDet [56]. $GT$ represents the ground truth visualization for phases, while $Loc$ . visualizes the model's highest confidence phase category and the time-boundary results. Blank segments denote invalid segments or operation gaps. + +![](images/d476f929fee901f85ba529cf22f6620032dad1799437c48991626aebcd816a0e.jpg) + +Setup. Following previous approaches of the primary surgery classification experiment in Sec. 4.1, We randomly mask phase sequences in the test video with different observation ratios. + +Baselines. We evaluate our datasets with the baseline models such as I3D [10], SlowFast [19], and MViT V2 [31]. For each model, we also adopted two training approaches: random initialization training and using pre-trained weights from Kinetics 400 [29]. + +Results. The phase detection results are illustrated in Tab. 5. The results demonstrate that the baseline models pretrained on Kinetics 400 [29] generally outperform their original counterparts in terms of Top-1 accuracy across different observation ratios for phase anticipation. Specifically, the modified MViT V2* model exhibits the highest improvement, achieving the best average Top-1 accuracy of $48.2\%$ . Moreover, while all models show increased accuracy with higher observation ratios, indicating that more observed data contributes to better performance, the consistent improvement across all ratios for the enhanced models suggests effective modifications. + +# 5 Limitations + +Dataset Bias. OphNet's videos are sourced from YouTube and exhibit diverse styles, clarity, and screen elements. This diversity can aid detection models in generalization but may affect their effectiveness and performance. Some videos in the dataset include subtitles or additional video windows, such as a little subtitle or watermark shown in Fig. 2. Similarly, additional video windows offer another perspective but can make the scene chaotic, making it harder to recognize primary surgical actions. The presence of these factors in OphNet reflects the complexity of real-world surgical environments, because an ophthalmic microscope may inherently display different windows or show parameters during recording. While they pose challenges, they also present opportunities for developing models that can better handle variability and unpredictability, which are crucial aspects of real-world surgical scenarios. + +Annotation Bias. OphNet is entirely annotated by ophthalmologists, there is a distinct possibility of annotation bias reflecting specific regional practices, terminologies, and interpretations. Despite the universal nature of many ophthalmic + +procedures, subtle differences in surgical techniques, procedural preferences, and clinical terminologies could lead to inconsistencies in how surgeries are categorized and described across different regions. For instance, the terminology used to describe certain procedures might differ, with one region referring to a procedure as anterior vitrectomy while another uses pars plana vitrectomy. To reduce the possibility of biases in precise annotations, we have taken great care to establish a unified definition prior to describing the surgery, phase and operation. However, potential biases arising from regional variations and individual surgical practices are inevitable. + +# 6 Conclusion + +In response to the current challenges in ophthalmology, a surgical field apt for automation and remote control, we introduce OphNet, a large-scale, diverse, and expert-level video benchmark for understanding ophthalmic surgical workflows. OphNet is the most extensive dataset of its kind, containing a broad range of cataract, glaucoma, and corneal surgeries and detailed annotations for distinct surgical phases. OphNet comprises 2,278 surgical videos (284.8 hours), 7,320 phase segments and 9,795 operation segments (51.2 hours), showcasing 66 different types of ophthalmic surgeries: 13 cataract, 14 glaucoma, and 39 corneal. It is annotated with 102 phases and 150 operations. With OphNet, we explored primary surgery presence recognition, phase localization and phase anticipation on untrimmed videos, phase and operation recognition on trimmed videos. We employed state-of-the-art models to establish robust baselines and provided valuable insights into video understanding within sequences and fine granularity. Our work contributes to the broader understanding of surgical video tasks in medical contexts and promotes the integration of deep learning technologies into ophthalmic surgical procedures. + +# References + +1. Fair use on youtube. https://support.google.com/youtube/answer/9783148?hl=en#:~:text=If%20the%20use%20of%20copyright, copyright %20removal%20request%20to%20YouTube. +2. Youtube's copyright exception policy. https://www.youtube.com/howyoutubeworks/policies/copyright/#copyright-exceptions +3. Adrito, D., Danyal, Z.K., Dimitrios, P., Yitong, Z., John, G.H., Francisco, V., Sophia, B., Hani, J.M., Danail, S.: Pitvis: Workflow recognition in endoscopic pituitary surgery +4. Al Hajj, H., Lamard, M., Conze, P.H., Roychowdhury, S., Hu, X., Marsalkaite, G., Zisimopoulos, O., Dedmari, M.A., Zhao, F., Prellberg, J., Sahu, M., Galdran, A., Araujo, T., Vo, D.M., Panda, C., Dahiya, N., Kondo, S., Bian, Z., Vahdat, A., Bialopetravicius, J., Flouty, E., Qiu, C., Dill, S., Mukhopadhyay, A., Costa, P., Aresta, G., Ramamurthy, S., Lee, S.W., Campilho, A., Zachow, S., Xia, S., Conjeti, S., Stoyanov, D., Armaitis, J., Heng, P.A., Macready, W.G., Cochener, B., Quellec, G.: Cataracts: Challenge on automatic tool annotation for cataract + +surgery. Medical Image Analysis 52, 24-41 (2019). https://doi.org/https://doi.org/10.1016/j.media.2018.11.008, https://www.sciencedirect.com/science/article/pii/S136184151830865X +5. Alwassel, H., Giancola, S., Ghanem, B.: TSP: Temporally-sensitive pretraining of video encoders for localization tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 3173-3183 (2021) +6. Bar, O., Neimark, D., Zohar, M., Hager, G.D., Girshick, R., Fried, G.M., Wolf, T., Asselmann, D.: Impact of data on generalization of ai for surgical intelligence applications. Scientific Reports 10(1), 22208 (Dec 2020). https://doi.org/10.1038/s41598-020-79173-6, https://doi.org/10.1038/s41598-020-79173-6 +7. Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilarino, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics 43, 99-111 (2015). https://doi.org/https://doi.org/10.1016/j.compmedimag.2015.02.007, https://www.sciencedirect.com/science/article/pii/S0895611115000567 +8. Bodenstedt, S., Wagner, M., Katic, D., Mietkowski, P., Mayer, B., Kenngott, H., Müller-Stich, B., Dillmann, R., Speidel, S.: Unsupervised temporal context learning using convolutional neural networks for laparoscopic workflow analysis. arXiv preprint arXiv:1702.03684 (2017) +9. Borgli, H., Thambawita, V., Smedsrud, P.H., Hicks, S., Jha, D., Eskeland, S.L., Randel, K.R., Pogorelov, K., Lux, M., Nguyen, D.T.D., Johansen, D., Griwodz, C., Stensland, H.K., Garcia-Ceja, E., Schmidt, P.T., Hammer, H.L., Riegler, M.A., Halvorsen, P., de Lange, T.: HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific Data 7(1), 283 (2020). https://doi.org/10.1038/s41597-020-00622-y, https://doi.org/10.1038/s41597-020-00622-y +10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +11. Czempiel, T., Paschali, M., Keicher, M., Simson, W., Feussner, H., Kim, S.T., Navab, N.: Tecn: Surgical phase recognition with multi-stage temporal convolutional networks. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoeanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 343-352. Springer International Publishing, Cham (2020) +12. Czempiel, T., Paschali, M., Ostler, D., Kim, S.T., Busam, B., Navab, N.: Opera: Attention-regularized transformers for surgical phase recognition. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part IV 24. pp. 604–614. Springer (2021) +13. Defays, D.: An efficient algorithm for a complete link method. The Computer Journal 20(4), 364-366 (01 1977). https://doi.org/10.1093/comjnl/20.4.364, https://doi.org/10.1093/comjnl/20.4.364 +14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +15. Dong, S., Hu, H., Lian, D., Luo, W., Qian, Y., Gao, S.: Weakly supervised video representation learning with unaligned text for sequential videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2447 (2023) + +16. Duong, H.T., Le, V.T., Hoang, V.T.: Deep learning-based anomaly detection in video surveillance: A survey. Sensors 23(11) (2023). https://doi.org/10.3390/s23115024, https://www.mdpi.com/1424-8220/23/11/5024 +17. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015) +18. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020) +19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) +20. Forslund Jacobsen, M., Konge, L., Alberti, M., la Cour, M., Park, Y.S., Thomsen, A.S.S.: ROBOT-ASSISTED VITREORETINAL SURGERY IMPROVES SURGICAL ACCURACY COMPARED WITH MANUAL SURGERY: A randomized trial in a simulated setting. Retina 40(11), 2091-2098 (Nov 2020) +21. Ghamsarian, N., El-Shabrawi, Y., Nasirihaghighi, S., Putzgruber-Adamitsch, D., Zinkernagel, M., Wolf, S., Schoeffmann, K., Sznitman, R.: Cataract-1k: Cataract surgery dataset for scene segmentation, phase recognition, and irregularity detection. arXiv preprint arXiv:2312.06295 (2023) +22. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., El-Shabrawi, Y., Schoeffmann, K.: Lensid: A cnn-rnn-based framework towards lens irregularity detection in cataract surgery videos. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part VIII. Lecture Notes in Computer Science, vol. 12908, pp. 76-86. Springer (2021). https://doi.org/10.1007/978-3-030-87237-3_8, https://doi.org/10.1007/978-3-030-87237-3_8 +23. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., Schoeffmann, K.: Relevance detection in cataract surgery videos by spatio-temporal action localization. In: 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021. pp. 10720-10727. IEEE (2020). https://doi.org/10.1109/ICPR48806.2021.9412525, https://doi.org/10.1109/ICPR48806.2021.9412525 +24. Grammatikopoulou, M., Flouty, E., Kadkhodamohammadi, A., Quellec, G., Chow, A., Nehme, J., Luengo, I., Stoyanov, D.: Cadis: Cataract dataset for image segmentation. arXiv preprint arXiv:1906.11586 (2019) +25. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: Ava: A video dataset of spatio-temporally localized atomic visual actions (2018) +26. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981) +27. Ji, G.P., Xiao, G., Chou, Y.C., Fan, D.P., Zhao, K., Chen, G., Van Gool, L.: Video polyp segmentation: A deep learning perspective. Machine Intelligence Research 19(6), 531-549 (Dec 2022). https://doi.org/10.1007/s11633-022-1371-y, https://doi.org/10.1007/s11633-022-1371-y +28. Jin, Y., Dou, Q., Chen, H., Yu, L., Qin, J., Fu, C.W., Heng, P.A.: SV-RCNet: Workflow recognition from surgical videos using recurrent convolutional network. IEEE Trans. Med. Imaging 37(5), 1114-1126 (May 2018) + +29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) +30. Li, J., Jin, Y., Chen, Y., Yip, H.C., Scheppach, M., Chiu, P.W.Y., Yam, Y., Meng, H.M.L., Dou, Q.: Imitation learning from expert video data for dissection trajectory prediction in endoscopic surgical procedure. In: Greenspan, H., Madabhushi, A., Mousavi, P., Salcudean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2023. pp. 494-504. Springer Nature Switzerland, Cham (2023) +31. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022) +32. Lin, S., Miao, A.J., Lu, J., Yu, S., Chiu, Z.Y., Richter, F., Yip, M.C.: Semantic-super: A semantic-aware surgical perception framework for endoscopic tissue identification, reconstruction, and tracking. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4739-4746 (2023). https://doi.org/10.1109/ICRA48891.2023.10160746 +33. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3889-3898 (2019) +34. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. arXiv preprint arXiv:2106.13230 (2021) +35. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Bi-calibration networks for weakly-supervised video representation learning. International Journal of Computer Vision 131(7), 1704-1721 (Jul 2023). https://doi.org/10.1007/s11263-023-01779-w, https://doi.org/10.1007/s11263-023-01779-w +36. Loukas, C.: Video content analysis of surgical procedures. Surgical Endoscopy 32(2), 553-568 (Feb 2018). https://doi.org/10.1007/s00464-017-5878-1, https://doi.org/10.1007/s00464-017-5878-1 +37. Lucas, B.D., Kanade, T., et al.: An iterative image registration technique with an application to stereo vision, vol. 81. Vancouver (1981) +38. Ma, Y., Chen, X., Cheng, K., Li, Y., Sun, B.: Ldpolypvideo benchmark: A large-scale colonoscopy video dataset of diverse polyps. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021. pp. 387-396. Springer International Publishing, Cham (2021) +39. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific data 8(1), 101 (2021) +40. Mesejo, P., Pizarro, D., Abergel, A., Rouquette, O., Beorchia, S., Poincloux, L., Bartoli, A.: Computer-aided classification of gastrointestinal lesions in regular colonoscopy. IEEE Transactions on Medical Imaging 35(9), 2051-2063 (2016). https://doi.org/10.1109/TMI.2016.2547947 +41. Ming, H., Lin, W., Siyuan, Y., Don, M., Qingli, R., Peng, X., Wei, F., Peibo, D., Lie, J., Zongyuan, G.: Nurvid: A large expert-level video database for nursing procedure activity understanding. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023) +42. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022) + +43. Ni, Z.L., Bian, G.B., Zhou, X.H., Hou, Z.G., Xie, X.L., Wang, C., Zhou, Y.J., Li, R.Q., Li, Z.: Raunet: Residual attention u-net for semantic segmentation of cataract surgical instruments. In: International Conference on Neural Information Processing (ICONIP). pp. 139-149. Springer (2019) +44. Nwoye, C.I., Gonzalez, C., Yu, T., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23. pp. 364-374. Springer (2020) +45. Nwoye, C.I., Padoy, N.: Data splits and metrics for benchmarking methods on surgical action triplet datasets. arXiv preprint arXiv:2204.05235 (2022) +46. Nwoye, C.I., Yu, T., Gonzalez, C., Seeliger, B., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. Medical Image Analysis 78, 102433 (2022) +47. Pan, X., Gao, X., Wang, H., Zhang, W., Mu, Y., He, X.: Temporal-based swim transformer network for workflow recognition of surgical video. Int. J. Comput. Assist. Radiol. Surg. 18(1), 139-147 (Jan 2023) +48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +49. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +50. Ross, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Comparative validation of multi-instance instrument segmentation in endoscopy: results of the robust-mis 2019 challenge. Medical image analysis 70, 101920 (2021) +51. Ross, T., Zimmerer, D., Vemuri, A., Isensee, F., Wiesenfarth, M., Bodenstedt, S., Both, F., Kessler, P., Wagner, M., Müller, B., et al.: Exploiting the potential of unlabeled endoscopic video data with self-supervised learning. International journal of computer assisted radiology and surgery 13, 925-933 (2018) +52. Rof, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Arbelaez, P., Bian, G.B., Bodenstedt, S., Bolmgren, J.L., Bravo-Sanchez, L., Chen, H.B., Gonzalez, C., Guo, D., Halvorsen, P., Heng, P.A., Hosgor, E., Hou, Z.G., Isensee, F., Jha, D., Jiang, T., Jin, Y., Kirtac, K., Kletz, S., Leger, S., Li, Z., Maier-Hein, K.H., Ni, Z.L., Riegler, M.A., Schoeffmann, K., Shi, R., Speidel, S., Stenzel, M., Twick, I., Wang, G., Wang, J., Wang, L., Wang, L., Zhang, Y., Zhou, Y.J., Zhu, L., Wiesenfarth, M., Kopp-Schneider, A., Müller-Stich, B.P., Maier-Hein, L.: Comparative validation of multi-instance instrument segmentation in endoscopy: Results of the robust-mis 2019 challenge. Medical Image Analysis 70, 101920 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101920, https://www.sciencedirect.com/science/article/pii/S136184152030284X +53. Sato, F., Hachiuma, R., Sekii, T.: Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6471-6480 (2023) +54. Schoeffmann, K., Husslein, H., Kletz, S., Petscharnig, S., Munzer, B., Beecks, C.: Video retrieval in laparoscopic video recordings with dynamic content descriptors. + +Multim. Tools Appl. 77(13), 16813-16832 (2018). https://doi.org/10.1007/s11042-017-5252-2, https://doi.org/10.1007/s11042-017-5252-2 +55. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: César, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018). https://doi.org/10.1145/3204949.3208137, https://doi.org/10.1145/3204949.3208137 +56. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18857-18866 (2023) +57. Shi, X., Jin, Y., Dou, Q., Heng, P.A.: LRTD: long-range temporal dependency based active learning for surgical workflow recognition. Int. J. Comput. Assist. Radiol. Surg. 15(9), 1573-1584 (Sep 2020) +58. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 510-526. Springer International Publishing, Cham (2016) +59. Smedsrud, P.H., Thambawita, V., Hicks, S.A., Gjestang, H., Nedrejord, O.O., Naess, E., Borgli, H., Jha, D., Berstad, T.J.D., Eskeland, S.L., Lux, M., Espeland, H., Petlund, A., Nguyen, D.T.D., Garcia-Ceja, E., Johansen, D., Schmidt, P.T., Toth, E., Hammer, H.L., de Lange, T., Riegler, M.A., Halvorsen, P.: Kvasir-Capsule, a video capsule endoscopy dataset. Scientific Data 8(1), 142 (2021). https://doi.org/10.1038/s41597-021-00920-z +60. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012) +61. Spaeth, G., Danesh-Meyer, H., Goldberg, I., Kampik, A.: Ophthalmic Surgery: Principles and Practice E-Book. Elsevier Health Sciences (2011), https://books.google.com.hk/books?id=wHWMUGH-5csC +62. Stauder, R., Ostler, D., Kranzfelder, M., Koller, S., Feußner, H., Navab, N.: The tum lapchole dataset for the m2cai 2016 workflow challenge. arXiv preprint arXiv:1610.09278 (2016) +63. Tian, Y., Pang, G., Liu, F., Liu, Y., Wang, C., Chen, Y., Verjans, J., Carneiro, G.: Contrastive transformer-based multiple instance learning for weakly supervised polyp frame detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 88-98. Springer (2022) +64. Tran, D., Wang, H., Torresani, L., Feiszli, M.: Video classification with channel-separated convolutional networks (2019) +65. Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., de Mathelin, M., Padoy, N.: Endonet: A deep architecture for recognition tasks on laparoscopic videos. IEEE Transactions on Medical Imaging (TMI) 36, 86-97 (2016), https://api-semanticscholar.org/CorpusID:5633749 +66. Twinanda, A.P., Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Rsdnet: Learning to predict remaining surgery duration from laparoscopic videos without manual annotations. IEEE Transactions on Medical Imaging 38(4), 1069-1078 (2019). https://doi.org/10.1109/TMI.2018.2878055 +67. Wagner, M., Müller-Stich, B.P., Kisilenko, A., Tran, D., Heger, P., Mündermann, L., Lubotsky, D.M., Müller, B., Davitashvili, T., Capek, M., et al.: Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis 86, 102770 (2023) + +68. Wang, T., Li, H., Pu, T., Yang, L.: Microsurgery robots: Applications, design, and development. Sensors 23(20) (2023). https://doi.org/10.3390/s23208503, https://www.mdpi.com/1424-8220/23/20/8503 +69. Wang, X., Zhang, S., Qing, Z., Shao, Y., Gao, C., Sang, N.: Self-supervised learning for semi-supervised temporal action proposal. In: CVPR (2021) +70. Wang, Z., Lu, B., Long, Y., Zhong, F., Cheung, T.H., Dou, Q., Liu, Y.: Autolaparo: A new dataset of integrated multi-tasks for image-guided surgical automation in laparoscopic hysterectomy. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 486-496. Springer (2022) +71. Wu, W., Sun, Z., Ouyang, W.: Revisiting classifier: Transferring vision-language models for video recognition (2023) +72. Wu, W., Wang, X., Luo, H., Wang, J., Yang, Y., Ouyang, W.: Bidirectional cross-modal knowledge exploration for video recognition with pre-trained vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +73. Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Less is more: Surgical phase recognition with less annotations through self-supervised pre-training of cnn-lstm networks. arXiv preprint arXiv:1805.08569 (2018) +74. Yu, F., Silva Croso, G., Kim, T.S., Song, Z., Parker, F., Hager, G.D., Reiter, A., Vedula, S.S., Ali, H., Sikder, S.: Assessment of Automated Identification of Phases in Videos of Cataract Surgery Using Machine Learning and Deep Learning Techniques. JAMA Network Open 2(4), e191860-e191860 (04 2019). https://doi.org/10.1001/jamanetworkopen.2019.1860, https://doi.org/10.1001/jamanetworkopen.2019.1860 +75. Yu, T., Mutter, D., Marescaux, J., Padoy, N.: Learning from a tiny dataset of manual annotations: a teacher/student approach for surgical phase recognition. arXiv preprint arXiv:1812.00033 (2018) +76. Yuan, K., Srivastav, V., Navab, N., Padoy, N.: Hecvl: Hierarchical videolanguage pretraining for zero-shot surgical phase recognition. arXiv preprint arXiv:2405.10075 (2024) +77. Yuan, K., Srivastav, V., Yu, T., Lavanchy, J., Mascagni, P., Navab, N., Padoy, N.: Learning multi-modal representations by watching hundreds of surgical video lectures (2023) +78. Zha, R., Cheng, X., Li, H., Harandi, M., Ge, Z.: Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 13-23. Springer (2023) +79. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: European Conference on Computer Vision. LNCS, vol. 13664, pp. 492-510 (2022) +80. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI Conference on Artificial Intelligence. pp. 7590-7598 (2018), https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/17344 \ No newline at end of file diff --git a/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/images.zip b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..66777bfa8f61a663f71acf84a74c9a80634f24b1 --- /dev/null +++ b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2f87e4175230dbb16a50f3d21ca144547437ca49273035f57e5ad3ee94fa26a +size 626319 diff --git a/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/layout.json b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..59dd8ed2b653c6c2c6dbd62471627de3044d23f1 --- /dev/null +++ b/2024/OphNet_ A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding/layout.json @@ -0,0 +1,9165 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "text", + "content": "OphNet: A Large-Scale Video Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": "Ming Hu" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Peng Xia" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Lin Wang" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{5*}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Siyuan Yan" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Feilong Tang" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": " Zhongxing Xu" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Yimin Luo" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Kaimin Song" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Jurgen Leitner" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": " Xuelian Cheng" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Jun Cheng" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Chi Liu" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Kaijing Zhou" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "text", + "content": ", and Zongyuan Ge" + }, + { + "bbox": [ + 130, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "AIM Lab, Faculty of IT, " + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "Faculty of Engineering, Monash University \n" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "Airdoc-Monash Research, Airdoc " + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "Eye Hospital, Wenzhou Medical University \n" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "Bosch Corporate Research " + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "King's College London " + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "Cornell University Institute for Infocomm Research, A*STAR \n" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 141, + 213, + 472, + 270 + ], + "type": "text", + "content": "Faculty of Data Science, City University of Macau" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 293, + 455, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 293, + 455, + 523 + ], + "spans": [ + { + "bbox": [ + 160, + 293, + 455, + 523 + ], + "type": "text", + "content": "Abstract. Surgical scene perception via videos is critical for advancing robotic surgery, telesurgery, and AI-assisted surgery, particularly in ophthalmology. However, the scarcity of diverse and richly annotated video datasets has hindered the development of intelligent systems for surgical workflow analysis. Existing datasets face challenges such as small scale, lack of diversity in surgery and phase categories, and absence of time-localized annotations. These limitations impede action understanding and model generalization validation in complex and diverse real-world surgical scenarios. To address this gap, we introduce OphNet, a large-scale, expert-annotated video benchmark for ophthalmic surgical workflow understanding. OphNet features: 1) A diverse collection of 2,278 surgical videos spanning 66 types of cataract, glaucoma, and corneal surgeries, with detailed annotations for 102 unique surgical phases and 150 fine-grained operations. 2) Sequential and hierarchical annotations for each surgery, phase, and operation, enabling comprehensive understanding and improved interpretability. 3) Time-localized annotations, facilitating temporal localization and prediction tasks within surgical workflows. With approximately 285 hours of surgical videos, OphNet is about 20 times larger than the largest existing surgical workflow analysis benchmark. Code and dataset are available at: https://minghuo830.github.io/OphNet-benchmark/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "spans": [ + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "text", + "content": "Keywords: Surgical Workflow Understanding " + }, + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "text", + "content": " Ophthalmic Surgery " + }, + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "text", + "content": " Medical Image Analysis " + }, + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 453, + 555 + ], + "type": "text", + "content": " Video Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 601, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 230, + 613 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "As surgical robot platforms such as the da Vinci® surgical system become increasingly sophisticated, there is growing interest in integrating enhanced intelligi-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 347, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 347, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 347, + 666 + ], + "type": "text", + "content": "* Equal contribution † Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "gence into scenarios like minimally invasive surgery [20,32,68]. The advancements in machine vision perception empower these robotic systems to autonomously recognize and adapt to the intricacies of surgical environments, without relying on binary instrument usage signals, RFID tags, sensor data from tracking devices, or other signal information that necessitates laborious manual annotations or additional equipment installations [28]. This autonomy includes the capacity to identify anatomical structures, detect anomalies, and adjust surgical plans in real-time, which is crucial in dynamic and unpredictable surgical settings. In recent years, especially in endoscopy and ophthalmic surgery, the application of deep learning has demonstrated considerable promise in bolstering these autonomous capabilities. This encompasses the analysis of surgical workflows [6, 28, 74], segmentation of instruments and anatomy [4, 24, 43], and depth estimation [78], among others." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 294, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 668 + ], + "type": "text", + "content": "Automatic video surgical workflow understanding is a fundamental yet challenging problem for developing computer-assisted and robotic-assisted surgery, which can be divided into internal (e.g., laparoscopic and endoscopic [28,47,57]) and external (e.g., operating room and nursing procedure [41]) analysis. In addition to promoting the development of intelligent surgery, it also greatly benefits surgical documentation, education, and training [11,12,65]. Baret et al. [6] showed networks, like Inflated 3D ConvNet (I3D) [10] that utilize spatiotemporal convolutions, require a relatively extensive dataset for effective training. In their study, the model achieves an accuracy exceeding " + }, + { + "bbox": [ + 130, + 294, + 482, + 668 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 294, + 482, + 668 + ], + "type": "text", + "content": " when trained on 100 videos, with a progressive improvement as the sample size surpasses 700. However, the highly efficient and rapidly evolving deep learning technologies for surgical workflow analysis are currently limited by the following shortcomings in current video benchmarks: 1) Small-scale: the majority of video datasets contain no more than 100 videos. For example, the CATARACTS [23] and CatRelDet [23] datasets contain only 50 and 21 surgical videos, respectively. These datasets are relatively small, insufficient for large-scale validation. 2) Limited categories of surgeries and phases: almost all ophthalmic surgical video datasets only include cataract surgery and do not further classify specific types of surgeries. Additionally, the number of phase categories is also limited, like CatRelDet [23] only contains 4 different phase labels, which is insufficient to meet the requirements for evaluation in real clinical environments. 3) Coarse-grained annotation: due to annotation costs, existing benchmarks often have coarse-grained action definitions. For example, adhesive injection may occur in two different phases: main incision and capsulorhexis, so it may be classified into different phase categories. Coarse-grained action definitions may lead to annotation bias. 4) Single time-boundary annotation: they only annotate designated phases in the videos, ignoring the continuity across different stages of ophthalmic surgery, as well as the hierarchical relationship between surgery, phase, and operation. Simpler datasets, such as LensID [22], are limited to binary classification tasks distinguishing lens implantation from other irrelevant phases. 5) Uniform domain: the videos are meticulously collected, and" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 481, + 313 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 481, + 313 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 481, + 313 + ], + "type": "table", + "html": "
ProtocolDataset PropertiesTasks
DatasetsNo. of VideosNo. of Action SegmentsNo. of Surgery CategoriesNo. of Action CategoriesTotal DurationMulti-Surgery Presence RecognitionPhase RecognitionPhase LocalizationPhase Prediction
Endo&LapCholec120 [45]120-1776.2hX✓XX
SurgicalActions160 [54]1601601160.2hX✓XX
HeiCo [39, 50]30-3142.8hX✓✓✓
EndoVis 2021 [67]332501722.0hX✓XX
PitVis [3]2528711733.3hX✓XX
CholecT50 [46]50-11044.7hX✓✓✓
AutoLaparo [70]213001723.1hX✓✓✓
OphScopeLensID [22]1002,4401211.7hX✓XX
Cataract-101 [55]1011,26611014.0hX✓✓✓
CatRelDet [23]212,400142.0hX✓XX
CATARACTS [4]501,53611920.0hX✓✓✓
Cataract-1K [21]1,000931112118.7hX✓✓✓
OphNet(Ours)2,2789,79566150284.8h✓✓✓✓
", + "image_path": "57a423041021a8a715cf693772b19cdd4fe79d23df36d6ff6a7c98321c8ec8d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 322, + 481, + 531 + ], + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 531 + ], + "type": "text", + "content": "Table 1: The statistics comparison among existing workflow analysis datasets and our OphNet. Compared to other datasets, OphNet focuses on more comprehensive coverage of various surgery, phase and operation categories, collects a large number of videos, totaling 284.8 hours, and also enables a variety of recognition, localization and prediction tasks. OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. For instance, Cholec120 [45], Cholec80 [65], m2cai-workflow and LapChole [62] form one series, whereas CholecT50 [46], CholecT45 [46], and CholecT40 [44] comprise another series. We have excluded the following scenarios from our comparison: (1) non-open-source datasets such as Bypass170 [66], ESD [30], Yu's [74], etc.; (2) a superset of multiple open-source or non-open-source datasets, like Cholec207 [6], etc.; (3) datasets employed for lesion, anatomy, and instrument classification and segmentation, such as SUN-SEG [27], CVC-ClinicDB [7], ROBUSTMIS [52], Mesejo's [40], Cata7 [43], etc., anomaly detection such as PolypDiag [63] (from Hyper-Kvasir [9] and LDPolypVideo [38]), Kvasir-Capsule [59], etc., and other datasets not dedicated to workflow analysis. It's worth mentioning that even in comparison with the above datasets, OphNet demonstrates considerable competitiveness in both its scale and the richness of its labels. Endo&Lap denotes the endoscopic and laparoscopic protocol, OphScope denotes the ophthalmic microscope protocol. We choose the latest version for comparison in cases where datasets have multiple supplementary updates." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 567, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 592 + ], + "type": "text", + "content": "while this ensures video quality, the uniform style is not conducive to testing the model's domain generalization ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": "While some works have explored semi-supervised and self-supervised learning strategies [8, 51, 73, 75] to alleviate the cost of annotations or use only a small fraction of available labels, these approaches still lack competitiveness in performance compared to fully supervised learning. This deficiency in performance is impeding the widespread clinical application of these strategies. To address the shortage of sufficient labeled datasets, we construct OphNet, a large-scale" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "and expert-level video benchmark with high diversity, for ophthalmic surgical workflow understanding. The main advantages of OphNet are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 147, + 481, + 422 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 147, + 481, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 147, + 481, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 147, + 481, + 254 + ], + "type": "text", + "content": "- Largest scale and diversity: to the best of our knowledge, OphNet is currently the largest and most richly labeled dataset for surgical workflow analysis. It contains a number of videos 20 times greater than the current largest benchmark in ophthalmic surgery and far exceeds datasets in more established fields such as endoscopy. Additionally, OphNet includes the greatest variety of different types of surgeries, encompassing 66 different surgeries such as cataract, glaucoma, and corneal surgeries, along with 102 unique surgical phases and 150 distinct operations. This diversity significantly surpasses that of previous research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 255, + 481, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 481, + 349 + ], + "type": "text", + "content": "- Fine-grained, sequential and hierarchical annotation: we have meticulously selected a subset of videos for annotation localization, with each video being annotated for an average of 22 operations. Additionally, we provide exquisite annotations at the levels of surgery, phase, and operation, catering to the requirements for training specific challenge models. This annotation design aims to offer a multifaceted understanding of surgical protocols, accommodate the nuances of each distinct surgery, and enhance the usability and interpretability of our dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 350, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 481, + 422 + ], + "type": "text", + "content": "- Expert-level manual annotation: the annotation work for OphNet was completed by ten experienced ophthalmologists and five individuals with ophthalmic experience, encompassing, but not limited to, standardization of definitions for surgery, phase, and operation labels, video filtering, classification and localization annotations, and secondary verification, among others. Expert-level annotators ensure the quality and professionalism of OphNet." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 438, + 236, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 236, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 236, + 451 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 462, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 481, + 604 + ], + "type": "text", + "content": "Surgical Workflow Understanding. Beyond its therapeutic advantages, minimally invasive surgery also provides the capability for operative video recording. These videos can be stored and later utilized for various purposes such as cognitive training, skill assessment, and surgical workflow analysis [76, 77]. Techniques derived from the broader field of video content analysis and representation are increasingly being incorporated into the surgical realm [6, 36, 74]. A typical surgical workflow can be defined by a sequence of tasks or events, including patient positioning, incision, dissection, and suturing. These events are influenced not only by the specific type of surgery but also by the individual surgeon's proficiency and technique. Consequently, a comprehensive understanding of the surgical workflow necessitates a thorough examination of the temporal, spatial, and contextual facets of these tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "content": "Weakly-Supervised Video Learning. Substantial pioneering work has been undertaken in the realms of video understanding [5,16,33,53,69,71,72,79]. Since even a small amount of videos easily comprises several million frames, methods that do not rely on a frame-level annotation are of special importance. Weakly-supervised video learning makes use of loosely labeled data to train models," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "thereby obviating the necessity for exhaustively annotated training data. These weak labels might manifest in a variety of forms, including video-level labels or partial labels, which are less specific than frame-level or pixel-level annotations. The objective of weakly-supervised learning is to utilize these coarse labels to produce models capable of delivering fine-grained predictions, such as temporally precise action localization or detailed semantic segmentation [15,35]. The large-scale annotation of surgical videos demands a significant investment of invaluable medical time resources. In this context, weakly supervised learning emerges as a viable solution to this bottleneck." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 243, + 282, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 282, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 282, + 255 + ], + "type": "text", + "content": "3 Dataset Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 483, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 483, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 483, + 423 + ], + "type": "text", + "content": "In this section, we detail the construction of our dataset, which involved meticulous data collection and preprocessing. YouTube was leveraged as a primary source to circumvent privacy issues while ensuring a broad representation of ophthalmic surgeries. Our selection criteria aimed to capture a wide range of video qualities and styles, specifically targeting cataract, glaucoma, and corneal surgeries due to their prevalence in clinical settings. We refined the dataset by excluding videos of inadequate quality or those depicting non-human subjects. The annotation process was designed to reflect the complex nature of eye surgeries, incorporating hierarchical classification to account for multiple conditions often treated within a single procedure. This was complemented by detailed localization annotations, delineating the distinct phases and techniques characteristic of ophthalmic operations, all undertaken by a dedicated team of ophthalmologists to ensure accuracy and relevance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 441, + 326, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 326, + 454 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 326, + 454 + ], + "type": "text", + "content": "3.1 Data Collection & Preprocessing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": "Collection. Medical data exhibit unique privacy considerations and tend to be of smaller data volumes, characteristics that are particularly salient in the case of video data. By capitalizing on the wealth of surgical videos available on YouTube, we are able to obviate potential ethical and privacy concerns while concurrently enabling the rapid procurement of a substantial corpus of videos for further screening [1,2,17,29]. To fulfill this objective, we deploy text-based search algorithms to probe each surgery on YouTube, obtaining videos with titles that incorporate the requisite surgical keywords. We select cataract, glaucoma, and corneal surgery—three of the most commonly performed ophthalmic surgeries in actual clinical environments—as the central subjects of our research. To expand our video collection, we bolster our search queries by integrating synonyms and abbreviations associated with each type of ophthalmic surgery. For instance, cataract surgery encompasses various types: Phacoemulsification (abbr. PHACO), Intraocular Lens implantation (abbr. IOL), and Extracapsular Cataract Extraction (abbr. ECCE), etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Preprocessing. Given the nature of text-based retrieval and the varying quality, style, and filming methods of surgical videos influenced by different YouTube" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 103 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 111, + 479, + 348 + ], + "blocks": [ + { + "bbox": [ + 133, + 111, + 479, + 348 + ], + "lines": [ + { + "bbox": [ + 133, + 111, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 479, + 348 + ], + "type": "image", + "image_path": "6e3ce91aebec06a16e145c80ecb0120a0c0fee542ad3cc25d3237b18b57910a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 356, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 401 + ], + "type": "text", + "content": "Fig. 1: The figure shows two combined surgical videos, PHACO + IOL implantation and PHACO + Trabeculotomy + Transscleral Cyclophotocoagulation + IOL implantation. For each frame marked in color, we provide time-boundary annotations at surgical, phase and operation levels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 509 + ], + "type": "text", + "content": "sources, some videos may exhibit poor quality or deviations in surgical representation. We filter out low-resolution videos, black-and-white color schemes, and animated demonstrations. Furthermore, surgical videos featuring non-human eyes, such as pig eyes, rabbit eyes, or pseudo-human eyes, are also outside the scope of our annotation. The data preprocessing was jointly completed by five professionals with experience in ophthalmology for initial screening, followed by a review conducted by six attending ophthalmologists." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 525, + 246, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 246, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 246, + 537 + ], + "type": "text", + "content": "3.2 Data Annotation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Classification Annotation. Different from the single-label classification in most natural video datasets [17, 25, 58, 60, 80], in ophthalmic surgery, physicians typically consider and implement multiple types of surgeries and phases based on the specific conditions and needs of the patient. This multifaceted approach stems from the complexity of the eye as an organ, where numerous diseases often coexist. Consequently, multiple ocular problems may need to be addressed in a single surgery. As shown in Fig. 1, a patient with cataracts may also have glaucoma, necessitating the simultaneous treatment of both conditions within one surgery. Hence, we initially classified all videos based on the primary surgical categories into three principal groups: cataract, glaucoma, and corneal" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 282, + 232 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 282, + 232 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 282, + 232 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 282, + 232 + ], + "type": "image", + "image_path": "7b52aaf4761e229370e29937aee0afdec046a214687e9c8f1afaa73289bf3a5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 289, + 114, + 479, + 236 + ], + "blocks": [ + { + "bbox": [ + 289, + 114, + 479, + 236 + ], + "lines": [ + { + "bbox": [ + 289, + 114, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 289, + 114, + 479, + 236 + ], + "type": "image", + "image_path": "fcd339a165035909f0c64709c414835b5183a983fcd5fd796b93494db91028e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 247, + 222, + 294 + ], + "blocks": [ + { + "bbox": [ + 203, + 237, + 215, + 246 + ], + "lines": [ + { + "bbox": [ + 203, + 237, + 215, + 246 + ], + "spans": [ + { + "bbox": [ + 203, + 237, + 215, + 246 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 247, + 222, + 294 + ], + "lines": [ + { + "bbox": [ + 137, + 247, + 222, + 294 + ], + "spans": [ + { + "bbox": [ + 137, + 247, + 222, + 294 + ], + "type": "image", + "image_path": "5382f905733380884d8f3a80587ad66136d6c32d2e538c9aacfc4404a4bef176.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 246, + 309, + 294 + ], + "blocks": [ + { + "bbox": [ + 225, + 246, + 309, + 294 + ], + "lines": [ + { + "bbox": [ + 225, + 246, + 309, + 294 + ], + "spans": [ + { + "bbox": [ + 225, + 246, + 309, + 294 + ], + "type": "image", + "image_path": "a6bbd279c72da86950d440012c84e0729166b52cc2ee02691d1627955ac4675d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 296, + 289, + 304 + ], + "lines": [ + { + "bbox": [ + 246, + 296, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 246, + 296, + 289, + 304 + ], + "type": "text", + "content": "Capsulorhexis" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 246, + 397, + 294 + ], + "blocks": [ + { + "bbox": [ + 375, + 237, + 388, + 246 + ], + "lines": [ + { + "bbox": [ + 375, + 237, + 388, + 246 + ], + "spans": [ + { + "bbox": [ + 375, + 237, + 388, + 246 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 246, + 397, + 294 + ], + "lines": [ + { + "bbox": [ + 313, + 246, + 397, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 397, + 294 + ], + "type": "image", + "image_path": "f53fade19af4480e7439ab961a92878738e07a1a7ce302ccdcefd63bcf56a13c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 296, + 380, + 305 + ], + "lines": [ + { + "bbox": [ + 331, + 296, + 380, + 305 + ], + "spans": [ + { + "bbox": [ + 331, + 296, + 380, + 305 + ], + "type": "text", + "content": "Capsule Staining" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 401, + 246, + 485, + 294 + ], + "blocks": [ + { + "bbox": [ + 401, + 246, + 485, + 294 + ], + "lines": [ + { + "bbox": [ + 401, + 246, + 485, + 294 + ], + "spans": [ + { + "bbox": [ + 401, + 246, + 485, + 294 + ], + "type": "image", + "image_path": "384b32f6ce0db429a9014cfe1b54426b62f743fd98ea4ea28d59c0b97219757e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 294, + 474, + 310 + ], + "lines": [ + { + "bbox": [ + 413, + 294, + 474, + 310 + ], + "spans": [ + { + "bbox": [ + 413, + 294, + 474, + 310 + ], + "type": "text", + "content": "Hydrodissection and Float Nucleus" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 137, + 309, + 222, + 357 + ], + "blocks": [ + { + "bbox": [ + 159, + 296, + 200, + 304 + ], + "lines": [ + { + "bbox": [ + 159, + 296, + 200, + 304 + ], + "spans": [ + { + "bbox": [ + 159, + 296, + 200, + 304 + ], + "type": "text", + "content": "Main Incision" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 309, + 222, + 357 + ], + "lines": [ + { + "bbox": [ + 137, + 309, + 222, + 357 + ], + "spans": [ + { + "bbox": [ + 137, + 309, + 222, + 357 + ], + "type": "image", + "image_path": "b2e6c9e169eeb878b093e02d2e3546f7f344313126e52b0a16e736a10ddf5ac8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 360, + 208, + 368 + ], + "lines": [ + { + "bbox": [ + 151, + 360, + 208, + 368 + ], + "spans": [ + { + "bbox": [ + 151, + 360, + 208, + 368 + ], + "type": "text", + "content": "Cortical Aspiration" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "type": "text", + "content": "Fig. 2: OphNet's composition, comparison with other datasets for the same task, and some phase examples: (a) an overview of the composition ratios at the levels of surgery, phase, and operation; (b) comparison among existing open-source " + }, + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "type": "text", + "content": " laparoscopic & endoscopic, and " + }, + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "type": "text", + "content": " ophthalmic microscope workflow analysis video datasets and " + }, + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 388, + 482, + 467 + ], + "type": "text", + "content": " our OphNet. OphNet stands as the largest real-world video dataset for ophthalmic surgical workflow understanding, featuring the highest number of videos, longest duration, and diverse categories of surgeries and phases; (c) eight phase examples in OphNet." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 225, + 310, + 309, + 357 + ], + "blocks": [ + { + "bbox": [ + 225, + 310, + 309, + 357 + ], + "lines": [ + { + "bbox": [ + 225, + 310, + 309, + 357 + ], + "spans": [ + { + "bbox": [ + 225, + 310, + 309, + 357 + ], + "type": "image", + "image_path": "c3f077bd2d823d8e6b0fe2ed77678c81ae2141c8569afde47d4f9692296618af.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 360, + 293, + 368 + ], + "lines": [ + { + "bbox": [ + 241, + 360, + 293, + 368 + ], + "spans": [ + { + "bbox": [ + 241, + 360, + 293, + 368 + ], + "type": "text", + "content": "IOL Implantation" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 370, + 314, + 380 + ], + "lines": [ + { + "bbox": [ + 302, + 370, + 314, + 380 + ], + "spans": [ + { + "bbox": [ + 302, + 370, + 314, + 380 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 313, + 310, + 397, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 310, + 397, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 397, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 397, + 357 + ], + "type": "image", + "image_path": "9fa9450d546ce9bf0f96787f22bf2db0bc9f820be34ec429d30591de819b263a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 360, + 385, + 368 + ], + "lines": [ + { + "bbox": [ + 326, + 360, + 385, + 368 + ], + "spans": [ + { + "bbox": [ + 326, + 360, + 385, + 368 + ], + "type": "text", + "content": "Cautery/Hemostasis" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 401, + 310, + 485, + 357 + ], + "blocks": [ + { + "bbox": [ + 401, + 310, + 485, + 357 + ], + "lines": [ + { + "bbox": [ + 401, + 310, + 485, + 357 + ], + "spans": [ + { + "bbox": [ + 401, + 310, + 485, + 357 + ], + "type": "image", + "image_path": "5633365ecc0620f91407220347e404338336258036c2a686fd8414bddf14b1f2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 360, + 473, + 368 + ], + "lines": [ + { + "bbox": [ + 413, + 360, + 473, + 368 + ], + "spans": [ + { + "bbox": [ + 413, + 360, + 473, + 368 + ], + "type": "text", + "content": "Phacoemulsification" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "content": "surgeries. Subsequently, these were further allocated for the detailed annotation of primary and secondary surgery. It is important to note that there exists only a single type of primary surgery, whereas multiple secondary surgery may be present. The categorization and annotation were meticulously completed by a team comprised of 8 experienced ophthalmologists." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Hierarchical Localization Annotation. A single surgery is often multifaceted, involving a series of intricate phases, each requiring distinct techniques and instruments, and the transition between surgeries entails high precision and coordination. Routine cataract surgery involves several steps. It begins with the administration of anesthesia, followed by a small incision in the cornea. The surgeon then creates an opening in the lens capsule and uses ultrasonic vibrations to break up and remove the cataract. Afterward, an artificial intraocular lens (IOL) is inserted into the lens capsule, and the incision is sealed without stitches. We define the phases and operation for various surgeries based on the" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 113, + 306, + 259 + ], + "blocks": [ + { + "bbox": [ + 138, + 113, + 306, + 259 + ], + "lines": [ + { + "bbox": [ + 138, + 113, + 306, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 113, + 306, + 259 + ], + "type": "image", + "image_path": "be64c5f57da297c9fbce7bc783237ab0b093c29b47faf1b30bccdf2c10f1511c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 262, + 255, + 271 + ], + "lines": [ + { + "bbox": [ + 187, + 262, + 255, + 271 + ], + "spans": [ + { + "bbox": [ + 187, + 262, + 255, + 271 + ], + "type": "text", + "content": "(a) phase statistics" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 480, + 259 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 480, + 259 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 480, + 259 + ], + "type": "image", + "image_path": "b436e3316441290ec1749f9081749dcef4e6f24d32f53b15df7b8bb989cd5f09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 262, + 435, + 271 + ], + "lines": [ + { + "bbox": [ + 352, + 262, + 435, + 271 + ], + "spans": [ + { + "bbox": [ + 352, + 262, + 435, + 271 + ], + "type": "text", + "content": "(b) operation statistics" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 281, + 480, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 314 + ], + "type": "text", + "content": "Fig. 3: We present the data statistics of trimmed videos at the levels of phase and operation, including the number of trimmed videos, average duration, and total duration. The IDs and corresponding names can be found in the appendix." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 445 + ], + "type": "text", + "content": "textbook Ophthalmic Surgery: Principles and Practice [61]. To ensure the quality of localization annotations, we assign each annotator videos of 2 different types of primary surgeries. For each action time-boundary annotation, we annotate at three different levels of granularity: surgery, phase, and operation. We also employed the complete linkage algorithm [13] to cluster and merge various temporal boundaries into stable boundaries that received multiple agreements. It's worth noting that an individual video may have multiple separate instances of the same or different phases, thereby leading to multiple boundary definitions. A team of 15 ophthalmologists completes the localization annotation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 461, + 318, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 318, + 473 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 318, + 473 + ], + "type": "text", + "content": "3.3 Dataset Statistics and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 479, + 482, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 600 + ], + "type": "text", + "content": "OphNet includes 2,278 surgical videos (284.8 hours), demonstrating 66 different types of ophthalmic surgeries: 13 types of cataract surgery, 14 types of glaucoma surgery, and 39 types of corneal surgery. There are 102 phases and 150 operations for recognition, detection and prediction tasks, summarized in Tab. 1. Over " + }, + { + "bbox": [ + 130, + 479, + 482, + 600 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 130, + 479, + 482, + 600 + ], + "type": "text", + "content": " of videos have high-definition resolutions of " + }, + { + "bbox": [ + 130, + 479, + 482, + 600 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 130, + 479, + 482, + 600 + ], + "type": "text", + "content": " pixels or higher. To facilitate algorithm development and evaluation, we selected 523 videos for localization annotations. Additionally, we trim videos according to annotated action boundaries, resulting in 7,320 phase instances and 9,795 operation instances, totaling 51.2 hours. The average duration of trimmed videos is 32 seconds, while untrimmed videos average 337 seconds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 230, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 230, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 230, + 631 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In our study, we explore four potential tasks using the OphNet dataset: 1) primary surgery presence recognition, 2) phase and operation recognition, 3) phase" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "localization, and 4) phase anticipation. To establish robust baselines, we employ state-of-the-art models known for their effectiveness in human action recognition, detection and anticipation. For each task, we provide a detailed problem formulation and evaluate the baseline models' performance. Our findings offer valuable insights into video understanding within sequences and fine granularity, contributing to the field's knowledge of video tasks in medical contexts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 425, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 425, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 425, + 217 + ], + "type": "text", + "content": "4.1 Surgery Presence, Phase and Operation Recognition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 354 + ], + "type": "text", + "content": "Task Description. Surgery presence recognition focuses on identifying various surgical types in untrimmed videos through weakly supervised methods. This requires the model to discern and capture distinct surgical action features within extensive video footage. In OphNet, all types of surgeries occurring in each video are annotated, but only a subset of the videos have time-boundary annotations, as detailed in Sec. 3.2. We identify the primary surgery based on the key objectives and durations of the surgeries. To streamline the process, our experiments are limited to recognizing the presence of these primary surgeries. Additionally, phase recognition segments the surgery into distinct phases using visual cues and movements, such as incision, lens removal, and implantation. Operation recognition involves identifying finer-grained surgical actions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 355, + 272, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 272, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 272, + 570 + ], + "type": "text", + "content": "Baselins. We compare the performance of I3D [10], SlowFast [19], X3D [18], and MViT V2 [31] models on this task. These models are evaluated in two versions: 1) random initialization training and 2) pretraining with weights from Kinetics 400 [29], which is a human action recognition dataset. In addition, we also explored the classification performance of X-CLIP [42] and ViFi-CLIP [49], two CLIP [48]-based models. We also compared the effects of different numbers of input frames on the models' performance, where the sub" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 280, + 379, + 482, + 491 + ], + "blocks": [ + { + "bbox": [ + 280, + 379, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 280, + 379, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 280, + 379, + 482, + 491 + ], + "type": "table", + "html": "
BaselinesPrimary Surgery Classification
CataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
I3D [10]38.986.343.671.742.778.229.853.2
SlowFast [19]45.382.144.672.345.577.327.254.4
X3D [31]42.187.444.674.243.876.628.562.7
MViT V2 [18]43.284.245.581.845.575.929.160.1
I3D [10]36.884.748.281.848.676.527.250.6
SlowFast*49.083.747.380.949.275.627.250.6
X3D*47.486.346.481.548.378.235.461.4
MViT V2*44.285.349.181.847.777.328.563.3
X-CLIP16 [42]58.594.751.892.861.488.640.579.0
X-CLIP3260.692.653.583.756.884.158.981.0
ViFi-CLIP16 [49]59.688.352.480.261.481.858.979.8
ViFi-CLIP3259.688.351.584.850.075.056.377.2
", + "image_path": "4c9c80c9bad3524905c6802bcbca78f7c080fde83c86f2dc37fd80cb46aa7fb5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 279, + 499, + 482, + 544 + ], + "lines": [ + { + "bbox": [ + 279, + 499, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 279, + 499, + 482, + 544 + ], + "type": "text", + "content": "Table 2: Per-class Top-1 and Top-5 accuracy (%) for the primary surgery presence recognition on untrimmed videos. The best performance for each split has been highlighted in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 570, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 592 + ], + "type": "text", + "content": "script _16 represents an input of 16 frames, and _32 represents an input of 32 frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Setup. The dataset was randomly partitioned to ensure a balanced representation of examples for each surgery category. Specifically, we allocated " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of the data for training (1,449 surgical videos), " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for validation (205 surgical videos), and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for testing (424 surgical videos). For the phase and operation recognition experiments, we followed the same settings, with " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of the data used for training (5,024 phase segments, 6,856 operations segments), " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "validation (730 phase segments, 975 operations segments), and " + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": " for testing (1,566 phase segments, 1,964 operations segments). The input for the surgery presence recognition experiment is untrimmed videos, while for the phase and operation recognition experiments, the input consists of trimmed segments. In all classification experiments, we set up the analysis and comparison from four perspectives: cataract surgery, glaucoma surgery, corneal surgery, and all surgical videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": "Results. We summarize the results in Tab. 2 and Tab. 3. In primary surgery classification, X-CLIP [42] achieved the highest overall Top-1 accuracy at " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "58.9\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": ", leading in cataract and glaucoma surgeries with accuracies of " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "60.6\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "53.5\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": " respectively. For corneal surgeries, X-CLIP also recorded the highest Top-1 and Top-5 accuracies of " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "61.4\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "88.6\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": ". In phase classification, ViFi-CLIP [49] led with the highest Top-1 accuracy, particularly in cataract surgeries at " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "75.9\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": ", and exhibited the best performance in both glaucoma and corneal surgeries. Furthermore, in operation classification, ViFi-CLIP outperformed all other models in all categories, especially noted in cataract and corneal surgeries with Top-1 accuracies of " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "83.7\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": " and Top-5 accuracies of " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "93.8\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": " respectively. Overall, ViFi-CLIP showed superior performance in both phase and operation classifications across various surgery types. Besides, in the phase and operation classification experiments, a higher number of input frames generally had a positive effect on the model. In Fig. 4, we present the heatmap visualization of four examples from the test set of phase recognition experiments using the ViFi-CLIP model. It can be observed that in a series of frame images, the model focuses on surgical instruments and the operated eye area, which is consistent with human experience. The instruments used in the same phase of ophthalmic surgery are often similar." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 449, + 254, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 254, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 254, + 460 + ], + "type": "text", + "content": "4.2 Phase Localization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "type": "text", + "content": "Task Description. Phase localization in ophthalmic surgical workflow analysis refers to the task of pinpointing the exact moments or time intervals within a surgical video where specific phases of the surgery begin and end. This involves the detailed temporal segmentation of the entire surgical procedure into its constituent phases based on visual cues, surgeon's actions, and the progression of the surgery. The objective of phase localization is to accurately identify the start and end times of different surgical stages, such as pre-operative preparation, incision, and removal of the lens facilitating a granular and precise understanding of the surgery timeline. This task is crucial for detailed surgical documentation, efficient surgical training, and the development of targeted interventions during specific stages of the surgery, enhancing overall surgical management and post-operative analysis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Baselines. We conducted the experiments using phase-level labels and used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. Data split follows the setup of the primary surgery classification experiment." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 482, + 240 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 240 + ], + "type": "table", + "html": "
BaselinesPhase ClassificationOperation Classification
CataractGlaucomaCorneaAllCataractGlaucomaCorneaAll
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1
I3D [10]27.255.724.157.518.952.125.758.226.854.923.556.018.051.225.057.1
SlowFast [19]26.556.523.156.524.249.126.760.125.855.222.945.923.548.526.059.0
X3D [31]27.058.321.055.521.828.526.662.326.447.220.544.621.327.826.161.5
MViT V2 [18]26.254.921.053.426.046.827.059.825.943.820.542.725.545.926.558.9
I3D*29.568.922.158.626.050.930.271.228.867.521.847.925.750.329.560.0
SlowFast*30.672.325.254.730.759.831.761.829.971.124.843.929.558.730.560.9
X3D*27.272.922.159.630.761.533.563.226.571.821.748.829.960.232.862.1
MViT V2*34.276.523.352.038.465.128.360.233.575.222.841.537.964.027.859.5
X-CLIP16 [42]68.392.247.389.853.077.463.485.367.591.046.578.982.276.162.584.0
X-CLIP3269.194.048.781.754.880.462.785.868.093.047.980.584.079.562.084.7
ViFi-CLIP16 [49]75.993.740.485.466.681.666.188.474.592.542.874.585.080.565.087.5
ViFi-CLIP3273.092.949.682.757.781.668.487.275.193.843.280.283.785.264.886.5
", + "image_path": "bb89199ea3128527ce0316940d07464f1db6d1af112ee3db53f6ec72cd04d6d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 341, + 340, + 426 + ], + "blocks": [ + { + "bbox": [ + 130, + 248, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 315 + ], + "type": "text", + "content": "Table 3: Per-class Top-1 and Top-5 accuracy (\\%) for the primary surgery presence recognition on untrimmed videos and phase recognition on trimmed videos. * denotes the initialization from the model pre-trained on Kinetics 400 [29]. For the two CLIP models, we chose ViT-B/16 as the backbone and compared the performance of two different input frame numbers, 16 and 32. The best performance for each split has been highlighted in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 341, + 340, + 426 + ], + "lines": [ + { + "bbox": [ + 132, + 341, + 340, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 340, + 426 + ], + "type": "table", + "html": "
BaselinesBackbonesmAP (%)
0.10.30.50.7Avg.
ActionFormerCSN [64]53.750.140.624.542.5
SwinViviT [34]59.354.743.326.346.4
SlowFast [19]60.055.945.126.047.5
TriDet [56]CSN56.153.043.129.446.2
SwinViviT61.057.147.133.150.4
SlowFast61.356.045.630.448.6
", + "image_path": "2b51ed963e3d1a99542736868cff811855359d1775a305368c08ac93a57693a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 430, + 340, + 518 + ], + "lines": [ + { + "bbox": [ + 130, + 430, + 340, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 340, + 518 + ], + "type": "text", + "content": "Table 4: The results for phase detection. Action-Former and TriDet are state-of-the-art models for human action detection tasks, and we use three different backbones for feature extraction and report mAP at the IoU thresholds of [0.1:0.2:0.9]. Average mAP is computed by averaging different IoU thresholds. The best performance for each split has been highlighted in bold." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 343, + 341, + 482, + 426 + ], + "blocks": [ + { + "bbox": [ + 343, + 341, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 343, + 341, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 343, + 341, + 482, + 426 + ], + "type": "table", + "html": "
BaselinesTop-1 Acc. (%)
0.10.30.50.7Avg.
I3D [10]26.542.249.851.347.3
SlowFast [19]25.442.648.952.247.2
MViT V2 [18]25.643.749.352.347.5
I3D*27.343.550.151.447.6
SlowFast*27.543.249.952.347.8
MViT V2*27.843.850.551.748.2
", + "image_path": "91ccd9df8db4b79a6a78d0f3ba289c5a04b99f4d73038fd0c7d1f67691816b0c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 342, + 430, + 482, + 518 + ], + "lines": [ + { + "bbox": [ + 342, + 430, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 342, + 430, + 482, + 518 + ], + "type": "text", + "content": "Table 5: The results for phase anticipation. We report top-1 accuracy at the observation ratios [0.1:0.2:0.9]. Average top-1 accuracy is computed by averaging different observation ratios. The best performance for each split has been highlighted in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Setup. We excluded Operation Gap and Invalid, and filtered out tags with fewer than 20 segments. We used two baseline models, ActionFormer [79] and TriDet [56], with backbone networks configured as CSN [64], SwinViviT [34], and SlowFast [19]. To extract features from the videos, we first extracted RGB frames from each video at a rate of 25 frames per second. We also extracted optical flow using the TV-L1 [26,37] algorithm. We then fine-tuned an I3D [10] model that had been pre-trained on the ImageNet [14] dataset, and used it to generate features for each RGB and optical flow frame. Because each video has a variable duration, we performed uniform interpolation to generate 100 fixed-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 111, + 306, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 111, + 306, + 251 + ], + "lines": [ + { + "bbox": [ + 133, + 111, + 306, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 306, + 251 + ], + "type": "image", + "image_path": "9c623f3ddfa6d3398ad7f45a6022f8c006b498813324c54bfce7ac270ce7be7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 281 + ], + "type": "text", + "content": "Fig. 4: Attention map visualizations of ViFi-CLIP [49] on four examples from OphNet's test set in the phase recognition task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 111, + 480, + 251 + ], + "blocks": [ + { + "bbox": [ + 307, + 111, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 307, + 111, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 111, + 480, + 251 + ], + "type": "image", + "image_path": "59a39a3f1b6cd39650adff69bb5d387e3de20e500fa779540ce4ad29512e2b1f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 305, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 480, + 339 + ], + "type": "text", + "content": "length features for each video. Finally, we concatenated the RGB and optical flow features into a 2048-dimensional embedding, which served as the input for our model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "content": "Results. The experimental results for phase localization are presented in the Tab. 4, showcasing the performance of different baseline models with various backbones in terms of mean Average Precision (mAP) at different Intersection over Union (IoU) thresholds [0.1:0.2:0.9]. The models evaluated include Action-Former with SwinViviT and SlowFast backbones, and TriDet with CSN, SwinViviT, and SlowFast backbones. The results indicate that the TriDet model with a SwinViviT backbone outperforms other combinations, achieving the highest mAP scores across most IoU thresholds, with notable scores of " + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "inline_equation", + "content": "61.0\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "content": " (IoU=0.1), " + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "content": " (IoU=0.3), " + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "inline_equation", + "content": "47.1\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "content": " (IoU=0.5), and " + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "inline_equation", + "content": "33.1\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "content": " (IoU=0.7), resulting in an average mAP of " + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "inline_equation", + "content": "50.4\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "content": ". This indicates that the TriDet model, especially when combined with the SwinViviT backbone, is particularly effective for phase localization in surgical videos. On the other hand, the TriDet model with a SlowFast backbone shows competitive performance, particularly achieving the highest mAP of " + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "inline_equation", + "content": "61.3\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 521 + ], + "type": "text", + "content": " at the lowest IoU threshold (0.1). However, it falls slightly behind in performance at higher IoU thresholds compared to the SwinViviT backbone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 538, + 257, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 257, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 257, + 550 + ], + "type": "text", + "content": "4.3 Phase Anticipation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Task Description. This task requires the analysis of real-time or recorded video data to foresee the sequence of events based on current and past surgical activities. By understanding the typical progression of ophthalmic surgeries and recognizing patterns in the surgeon's actions and the use of instruments, the system aims to forecast the next phase of the surgery, allowing for proactive preparation and response. The objective of phase anticipation is to enhance the efficiency and safety of surgical procedures by providing the surgical team with advanced notice of upcoming steps, enabling better resource allocation, timing for critical tasks, and overall coordination within the operating room." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 111, + 303, + 175 + ], + "blocks": [ + { + "bbox": [ + 133, + 111, + 303, + 175 + ], + "lines": [ + { + "bbox": [ + 133, + 111, + 303, + 175 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 303, + 175 + ], + "type": "image", + "image_path": "57d2cb94b9d2a79c200a69e5e202df35fe28033b47254634482f624b18db31f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 183, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 228 + ], + "type": "text", + "content": "Fig. 5: Phase localization visualization of TriDet [56]. " + }, + { + "bbox": [ + 130, + 183, + 482, + 228 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 130, + 183, + 482, + 228 + ], + "type": "text", + "content": " represents the ground truth visualization for phases, while " + }, + { + "bbox": [ + 130, + 183, + 482, + 228 + ], + "type": "inline_equation", + "content": "Loc" + }, + { + "bbox": [ + 130, + 183, + 482, + 228 + ], + "type": "text", + "content": ". visualizes the model's highest confidence phase category and the time-boundary results. Blank segments denote invalid segments or operation gaps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 111, + 480, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 111, + 480, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 111, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 111, + 480, + 175 + ], + "type": "image", + "image_path": "d476f929fee901f85ba529cf22f6620032dad1799437c48991626aebcd816a0e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 285 + ], + "type": "text", + "content": "Setup. Following previous approaches of the primary surgery classification experiment in Sec. 4.1, We randomly mask phase sequences in the test video with different observation ratios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 287, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 334 + ], + "type": "text", + "content": "Baselines. We evaluate our datasets with the baseline models such as I3D [10], SlowFast [19], and MViT V2 [31]. For each model, we also adopted two training approaches: random initialization training and using pre-trained weights from Kinetics 400 [29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "type": "text", + "content": "Results. The phase detection results are illustrated in Tab. 5. The results demonstrate that the baseline models pretrained on Kinetics 400 [29] generally outperform their original counterparts in terms of Top-1 accuracy across different observation ratios for phase anticipation. Specifically, the modified MViT V2* model exhibits the highest improvement, achieving the best average Top-1 accuracy of " + }, + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "type": "inline_equation", + "content": "48.2\\%" + }, + { + "bbox": [ + 130, + 335, + 482, + 441 + ], + "type": "text", + "content": ". Moreover, while all models show increased accuracy with higher observation ratios, indicating that more observed data contributes to better performance, the consistent improvement across all ratios for the enhanced models suggests effective modifications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 461, + 222, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 222, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 222, + 473 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": "Dataset Bias. OphNet's videos are sourced from YouTube and exhibit diverse styles, clarity, and screen elements. This diversity can aid detection models in generalization but may affect their effectiveness and performance. Some videos in the dataset include subtitles or additional video windows, such as a little subtitle or watermark shown in Fig. 2. Similarly, additional video windows offer another perspective but can make the scene chaotic, making it harder to recognize primary surgical actions. The presence of these factors in OphNet reflects the complexity of real-world surgical environments, because an ophthalmic microscope may inherently display different windows or show parameters during recording. While they pose challenges, they also present opportunities for developing models that can better handle variability and unpredictability, which are crucial aspects of real-world surgical scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Annotation Bias. OphNet is entirely annotated by ophthalmologists, there is a distinct possibility of annotation bias reflecting specific regional practices, terminologies, and interpretations. Despite the universal nature of many ophthalmic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "procedures, subtle differences in surgical techniques, procedural preferences, and clinical terminologies could lead to inconsistencies in how surgeries are categorized and described across different regions. For instance, the terminology used to describe certain procedures might differ, with one region referring to a procedure as anterior vitrectomy while another uses pars plana vitrectomy. To reduce the possibility of biases in precise annotations, we have taken great care to establish a unified definition prior to describing the surgery, phase and operation. However, potential biases arising from regional variations and individual surgical practices are inevitable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 242, + 219, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 219, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 219, + 255 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 459 + ], + "type": "text", + "content": "In response to the current challenges in ophthalmology, a surgical field apt for automation and remote control, we introduce OphNet, a large-scale, diverse, and expert-level video benchmark for understanding ophthalmic surgical workflows. OphNet is the most extensive dataset of its kind, containing a broad range of cataract, glaucoma, and corneal surgeries and detailed annotations for distinct surgical phases. OphNet comprises 2,278 surgical videos (284.8 hours), 7,320 phase segments and 9,795 operation segments (51.2 hours), showcasing 66 different types of ophthalmic surgeries: 13 cataract, 14 glaucoma, and 39 corneal. It is annotated with 102 phases and 150 operations. With OphNet, we explored primary surgery presence recognition, phase localization and phase anticipation on untrimmed videos, phase and operation recognition on trimmed videos. We employed state-of-the-art models to establish robust baselines and provided valuable insights into video understanding within sequences and fine granularity. Our work contributes to the broader understanding of surgical video tasks in medical contexts and promotes the integration of deep learning technologies into ophthalmic surgical procedures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 476, + 198, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 198, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 198, + 489 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 500, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 500, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 500, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 136, + 500, + 481, + 533 + ], + "type": "text", + "content": "1. Fair use on youtube. https://support.google.com/youtube/answer/9783148?hl=en#:~:text=If%20the%20use%20of%20copyright, copyright %20removal%20request%20to%20YouTube." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 136, + 533, + 481, + 555 + ], + "type": "text", + "content": "2. Youtube's copyright exception policy. https://www.youtube.com/howyoutubeworks/policies/copyright/#copyright-exceptions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 136, + 555, + 481, + 588 + ], + "type": "text", + "content": "3. Adrito, D., Danyal, Z.K., Dimitrios, P., Yitong, Z., John, G.H., Francisco, V., Sophia, B., Hani, J.M., Danail, S.: Pitvis: Workflow recognition in endoscopic pituitary surgery" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 588, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 588, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 588, + 481, + 665 + ], + "type": "text", + "content": "4. Al Hajj, H., Lamard, M., Conze, P.H., Roychowdhury, S., Hu, X., Marsalkaite, G., Zisimopoulos, O., Dedmari, M.A., Zhao, F., Prellberg, J., Sahu, M., Galdran, A., Araujo, T., Vo, D.M., Panda, C., Dahiya, N., Kondo, S., Bian, Z., Vahdat, A., Bialopetravicius, J., Flouty, E., Qiu, C., Dill, S., Mukhopadhyay, A., Costa, P., Aresta, G., Ramamurthy, S., Lee, S.W., Campilho, A., Zachow, S., Xia, S., Conjeti, S., Stoyanov, D., Armaitis, J., Heng, P.A., Macready, W.G., Cochener, B., Quellec, G.: Cataracts: Challenge on automatic tool annotation for cataract" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "type": "text", + "content": "surgery. Medical Image Analysis 52, 24-41 (2019). https://doi.org/https://doi.org/10.1016/j.media.2018.11.008, https://www.sciencedirect.com/science/article/pii/S136184151830865X" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 481, + 183 + ], + "type": "text", + "content": "5. Alwassel, H., Giancola, S., Ghanem, B.: TSP: Temporally-sensitive pretraining of video encoders for localization tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 3173-3183 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 226 + ], + "type": "text", + "content": "6. Bar, O., Neimark, D., Zohar, M., Hager, G.D., Girshick, R., Fried, G.M., Wolf, T., Asselmann, D.: Impact of data on generalization of ai for surgical intelligence applications. Scientific Reports 10(1), 22208 (Dec 2020). https://doi.org/10.1038/s41598-020-79173-6, https://doi.org/10.1038/s41598-020-79173-6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 227, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 480, + 281 + ], + "type": "text", + "content": "7. Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilarino, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics 43, 99-111 (2015). https://doi.org/https://doi.org/10.1016/j.compmedimag.2015.02.007, https://www.sciencedirect.com/science/article/pii/S0895611115000567" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 281, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 480, + 326 + ], + "type": "text", + "content": "8. Bodenstedt, S., Wagner, M., Katic, D., Mietkowski, P., Mayer, B., Kenngott, H., Müller-Stich, B., Dillmann, R., Speidel, S.: Unsupervised temporal context learning using convolutional neural networks for laparoscopic workflow analysis. arXiv preprint arXiv:1702.03684 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 326, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 402 + ], + "type": "text", + "content": "9. Borgli, H., Thambawita, V., Smedsrud, P.H., Hicks, S., Jha, D., Eskeland, S.L., Randel, K.R., Pogorelov, K., Lux, M., Nguyen, D.T.D., Johansen, D., Griwodz, C., Stensland, H.K., Garcia-Ceja, E., Schmidt, P.T., Hammer, H.L., Riegler, M.A., Halvorsen, P., de Lange, T.: HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific Data 7(1), 283 (2020). https://doi.org/10.1038/s41597-020-00622-y, https://doi.org/10.1038/s41597-020-00622-y" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 435 + ], + "type": "text", + "content": "10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 435, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 435, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 134, + 435, + 480, + 501 + ], + "type": "text", + "content": "11. Czempiel, T., Paschali, M., Keicher, M., Simson, W., Feussner, H., Kim, S.T., Navab, N.: Tecn: Surgical phase recognition with multi-stage temporal convolutional networks. In: Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou, S.K., Racoeanu, D., Joskowicz, L. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2020. pp. 343-352. Springer International Publishing, Cham (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 501, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 501, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 501, + 480, + 555 + ], + "type": "text", + "content": "12. Czempiel, T., Paschali, M., Ostler, D., Kim, S.T., Busam, B., Navab, N.: Opera: Attention-regularized transformers for surgical phase recognition. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part IV 24. pp. 604–614. Springer (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 555, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 589 + ], + "type": "text", + "content": "13. Defays, D.: An efficient algorithm for a complete link method. The Computer Journal 20(4), 364-366 (01 1977). https://doi.org/10.1093/comjnl/20.4.364, https://doi.org/10.1093/comjnl/20.4.364" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 589, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 622 + ], + "type": "text", + "content": "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "15. Dong, S., Hu, H., Lian, D., Luo, W., Qian, Y., Gao, S.: Weakly supervised video representation learning with unaligned text for sequential videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2447 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Duong, H.T., Le, V.T., Hoang, V.T.: Deep learning-based anomaly detection in video surveillance: A survey. Sensors 23(11) (2023). https://doi.org/10.3390/s23115024, https://www.mdpi.com/1424-8220/23/11/5024" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "17. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "18. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "text", + "content": "19. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "text", + "content": "20. Forslund Jacobsen, M., Konge, L., Alberti, M., la Cour, M., Park, Y.S., Thomsen, A.S.S.: ROBOT-ASSISTED VITREORETINAL SURGERY IMPROVES SURGICAL ACCURACY COMPARED WITH MANUAL SURGERY: A randomized trial in a simulated setting. Retina 40(11), 2091-2098 (Nov 2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "text", + "content": "21. Ghamsarian, N., El-Shabrawi, Y., Nasirihaghighi, S., Putzgruber-Adamitsch, D., Zinkernagel, M., Wolf, S., Schoeffmann, K., Sznitman, R.: Cataract-1k: Cataract surgery dataset for scene segmentation, phase recognition, and irregularity detection. arXiv preprint arXiv:2312.06295 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 434 + ], + "type": "text", + "content": "22. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., El-Shabrawi, Y., Schoeffmann, K.: Lensid: A cnn-rnn-based framework towards lens irregularity detection in cataract surgery videos. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part VIII. Lecture Notes in Computer Science, vol. 12908, pp. 76-86. Springer (2021). https://doi.org/10.1007/978-3-030-87237-3_8, https://doi.org/10.1007/978-3-030-87237-3_8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 435, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 500 + ], + "type": "text", + "content": "23. Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., Schoeffmann, K.: Relevance detection in cataract surgery videos by spatio-temporal action localization. In: 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021. pp. 10720-10727. IEEE (2020). https://doi.org/10.1109/ICPR48806.2021.9412525, https://doi.org/10.1109/ICPR48806.2021.9412525" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "24. Grammatikopoulou, M., Flouty, E., Kadkhodamohammadi, A., Quellec, G., Chow, A., Nehme, J., Luengo, I., Stoyanov, D.: Cadis: Cataract dataset for image segmentation. arXiv preprint arXiv:1906.11586 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "25. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: Ava: A video dataset of spatio-temporally localized atomic visual actions (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "26. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "27. Ji, G.P., Xiao, G., Chou, Y.C., Fan, D.P., Zhao, K., Chen, G., Van Gool, L.: Video polyp segmentation: A deep learning perspective. Machine Intelligence Research 19(6), 531-549 (Dec 2022). https://doi.org/10.1007/s11633-022-1371-y, https://doi.org/10.1007/s11633-022-1371-y" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "28. Jin, Y., Dou, Q., Chen, H., Yu, L., Qin, J., Fu, C.W., Heng, P.A.: SV-RCNet: Workflow recognition from surgical videos using recurrent convolutional network. IEEE Trans. Med. Imaging 37(5), 1114-1126 (May 2018)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 214 + ], + "type": "text", + "content": "30. Li, J., Jin, Y., Chen, Y., Yip, H.C., Scheppach, M., Chiu, P.W.Y., Yam, Y., Meng, H.M.L., Dou, Q.: Imitation learning from expert video data for dissection trajectory prediction in endoscopic surgical procedure. In: Greenspan, H., Madabhushi, A., Mousavi, P., Salcudean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2023. pp. 494-504. Springer Nature Switzerland, Cham (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "type": "text", + "content": "31. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 257, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 311 + ], + "type": "text", + "content": "32. Lin, S., Miao, A.J., Lu, J., Yu, S., Chiu, Z.Y., Richter, F., Yip, M.C.: Semantic-super: A semantic-aware surgical perception framework for endoscopic tissue identification, reconstruction, and tracking. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4739-4746 (2023). https://doi.org/10.1109/ICRA48891.2023.10160746" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "text", + "content": "33. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3889-3898 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "text", + "content": "34. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. arXiv preprint arXiv:2106.13230 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "text", + "content": "35. Long, F., Yao, T., Qiu, Z., Tian, X., Luo, J., Mei, T.: Bi-calibration networks for weakly-supervised video representation learning. International Journal of Computer Vision 131(7), 1704-1721 (Jul 2023). https://doi.org/10.1007/s11263-023-01779-w, https://doi.org/10.1007/s11263-023-01779-w" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "text", + "content": "36. Loukas, C.: Video content analysis of surgical procedures. Surgical Endoscopy 32(2), 553-568 (Feb 2018). https://doi.org/10.1007/s00464-017-5878-1, https://doi.org/10.1007/s00464-017-5878-1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "37. Lucas, B.D., Kanade, T., et al.: An iterative image registration technique with an application to stereo vision, vol. 81. Vancouver (1981)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 460, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 514 + ], + "type": "text", + "content": "38. Ma, Y., Chen, X., Cheng, K., Li, Y., Sun, B.: Ldpolypvideo benchmark: A large-scale colonoscopy video dataset of diverse polyps. In: de Bruijne, M., Cattin, P.C., Cotin, S., Padoy, N., Speidel, S., Zheng, Y., Essert, C. (eds.) Medical Image Computing and Computer Assisted Intervention - MICCAI 2021. pp. 387-396. Springer International Publishing, Cham (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "type": "text", + "content": "39. Maier-Hein, L., Wagner, M., Ross, T., Reinke, A., Bodenstedt, S., Full, P.M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Heidelberg colorectal data set for surgical data science in the sensor operating room. Scientific data 8(1), 101 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "text", + "content": "40. Mesejo, P., Pizarro, D., Abergel, A., Rouquette, O., Beorchia, S., Poincloux, L., Bartoli, A.: Computer-aided classification of gastrointestinal lesions in regular colonoscopy. IEEE Transactions on Medical Imaging 35(9), 2051-2063 (2016). https://doi.org/10.1109/TMI.2016.2547947" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 643 + ], + "type": "text", + "content": "41. Ming, H., Lin, W., Siyuan, Y., Don, M., Qingli, R., Peng, X., Wei, F., Peibo, D., Lie, J., Zongyuan, G.: Nurvid: A large expert-level video database for nursing procedure activity understanding. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "42. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "43. Ni, Z.L., Bian, G.B., Zhou, X.H., Hou, Z.G., Xie, X.L., Wang, C., Zhou, Y.J., Li, R.Q., Li, Z.: Raunet: Residual attention u-net for semantic segmentation of cataract surgical instruments. In: International Conference on Neural Information Processing (ICONIP). pp. 139-149. Springer (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 215 + ], + "type": "text", + "content": "44. Nwoye, C.I., Gonzalez, C., Yu, T., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23. pp. 364-374. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "text", + "content": "45. Nwoye, C.I., Padoy, N.: Data splits and metrics for benchmarking methods on surgical action triplet datasets. arXiv preprint arXiv:2204.05235 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 237, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 281 + ], + "type": "text", + "content": "46. Nwoye, C.I., Yu, T., Gonzalez, C., Seeliger, B., Mascagni, P., Mutter, D., Marescaux, J., Padoy, N.: Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. Medical Image Analysis 78, 102433 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "text", + "content": "47. Pan, X., Gao, X., Wang, H., Zhang, W., Mu, Y., He, X.: Temporal-based swim transformer network for workflow recognition of surgical video. Int. J. Comput. Assist. Radiol. Surg. 18(1), 139-147 (Jan 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "49. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "text", + "content": "50. Ross, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., et al.: Comparative validation of multi-instance instrument segmentation in endoscopy: results of the robust-mis 2019 challenge. Medical image analysis 70, 101920 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "51. Ross, T., Zimmerer, D., Vemuri, A., Isensee, F., Wiesenfarth, M., Bodenstedt, S., Both, F., Kessler, P., Wagner, M., Müller, B., et al.: Exploiting the potential of unlabeled endoscopic video data with self-supervised learning. International journal of computer assisted radiology and surgery 13, 925-933 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 609 + ], + "type": "text", + "content": "52. Rof, T., Reinke, A., Full, P.M., Wagner, M., Kenngott, H., Apitz, M., Hempe, H., Mindroc-Filimon, D., Scholz, P., Tran, T.N., Bruno, P., Arbelaez, P., Bian, G.B., Bodenstedt, S., Bolmgren, J.L., Bravo-Sanchez, L., Chen, H.B., Gonzalez, C., Guo, D., Halvorsen, P., Heng, P.A., Hosgor, E., Hou, Z.G., Isensee, F., Jha, D., Jiang, T., Jin, Y., Kirtac, K., Kletz, S., Leger, S., Li, Z., Maier-Hein, K.H., Ni, Z.L., Riegler, M.A., Schoeffmann, K., Shi, R., Speidel, S., Stenzel, M., Twick, I., Wang, G., Wang, J., Wang, L., Wang, L., Zhang, Y., Zhou, Y.J., Zhu, L., Wiesenfarth, M., Kopp-Schneider, A., Müller-Stich, B.P., Maier-Hein, L.: Comparative validation of multi-instance instrument segmentation in endoscopy: Results of the robust-mis 2019 challenge. Medical Image Analysis 70, 101920 (2021). https://doi.org/https://doi.org/10.1016/j.media.2020.101920, https://www.sciencedirect.com/science/article/pii/S136184152030284X" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "53. Sato, F., Hachiuma, R., Sekii, T.: Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6471-6480 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "54. Schoeffmann, K., Husslein, H., Kletz, S., Petscharnig, S., Munzer, B., Beecks, C.: Video retrieval in laparoscopic video recordings with dynamic content descriptors." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "text", + "content": "Multim. Tools Appl. 77(13), 16813-16832 (2018). https://doi.org/10.1007/s11042-017-5252-2, https://doi.org/10.1007/s11042-017-5252-2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 204 + ], + "type": "text", + "content": "55. Schoeffmann, K., Taschwer, M., Sarny, S., Munzer, B., Primus, M.J., Putzgruber, D.: Cataract-101: video dataset of 101 cataract surgeries. In: César, P., Zink, M., Murray, N. (eds.) Proceedings of the 9th ACM Multimedia Systems Conference, MMSys 2018, Amsterdam, The Netherlands, June 12-15, 2018. pp. 421-425. ACM (2018). https://doi.org/10.1145/3204949.3208137, https://doi.org/10.1145/3204949.3208137" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "56. Shi, D., Zhong, Y., Cao, Q., Ma, L., Li, J., Tao, D.: Tridet: Temporal action detection with relative boundary modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18857-18866 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "57. Shi, X., Jin, Y., Dou, Q., Heng, P.A.: LRTD: long-range temporal dependency based active learning for surgical workflow recognition. Int. J. Comput. Assist. Radiol. Surg. 15(9), 1573-1584 (Sep 2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "text", + "content": "58. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 510-526. Springer International Publishing, Cham (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 380 + ], + "type": "text", + "content": "59. Smedsrud, P.H., Thambawita, V., Hicks, S.A., Gjestang, H., Nedrejord, O.O., Naess, E., Borgli, H., Jha, D., Berstad, T.J.D., Eskeland, S.L., Lux, M., Espeland, H., Petlund, A., Nguyen, D.T.D., Garcia-Ceja, E., Johansen, D., Schmidt, P.T., Toth, E., Hammer, H.L., de Lange, T., Riegler, M.A., Halvorsen, P.: Kvasir-Capsule, a video capsule endoscopy dataset. Scientific Data 8(1), 142 (2021). https://doi.org/10.1038/s41597-021-00920-z" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "60. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "61. Spaeth, G., Danesh-Meyer, H., Goldberg, I., Kampik, A.: Ophthalmic Surgery: Principles and Practice E-Book. Elsevier Health Sciences (2011), https://books.google.com.hk/books?id=wHWMUGH-5csC" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "62. Stauder, R., Ostler, D., Kranzfelder, M., Koller, S., Feußner, H., Navab, N.: The tum lapchole dataset for the m2cai 2016 workflow challenge. arXiv preprint arXiv:1610.09278 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "text", + "content": "63. Tian, Y., Pang, G., Liu, F., Liu, Y., Wang, C., Chen, Y., Verjans, J., Carneiro, G.: Contrastive transformer-based multiple instance learning for weakly supervised polyp frame detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 88-98. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "64. Tran, D., Wang, H., Torresani, L., Feiszli, M.: Video classification with channel-separated convolutional networks (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "text", + "content": "65. Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., de Mathelin, M., Padoy, N.: Endonet: A deep architecture for recognition tasks on laparoscopic videos. IEEE Transactions on Medical Imaging (TMI) 36, 86-97 (2016), https://api-semanticscholar.org/CorpusID:5633749" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "text", + "content": "66. Twinanda, A.P., Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Rsdnet: Learning to predict remaining surgery duration from laparoscopic videos without manual annotations. IEEE Transactions on Medical Imaging 38(4), 1069-1078 (2019). https://doi.org/10.1109/TMI.2018.2878055" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "67. Wagner, M., Müller-Stich, B.P., Kisilenko, A., Tran, D., Heger, P., Mündermann, L., Lubotsky, D.M., Müller, B., Davitashvili, T., Capek, M., et al.: Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis 86, 102770 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "OphNet: A Benchmark for Ophthalmic Surgical Workflow Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 589 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "68. Wang, T., Li, H., Pu, T., Yang, L.: Microsurgery robots: Applications, design, and development. Sensors 23(20) (2023). https://doi.org/10.3390/s23208503, https://www.mdpi.com/1424-8220/23/20/8503" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "69. Wang, X., Zhang, S., Qing, Z., Shao, Y., Gao, C., Sang, N.: Self-supervised learning for semi-supervised temporal action proposal. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 216 + ], + "type": "text", + "content": "70. Wang, Z., Lu, B., Long, Y., Zhong, F., Cheung, T.H., Dou, Q., Liu, Y.: Autolaparo: A new dataset of integrated multi-tasks for image-guided surgical automation in laparoscopic hysterectomy. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 486-496. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "71. Wu, W., Sun, Z., Ouyang, W.: Revisiting classifier: Transferring vision-language models for video recognition (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 281 + ], + "type": "text", + "content": "72. Wu, W., Wang, X., Luo, H., Wang, J., Yang, Y., Ouyang, W.: Bidirectional cross-modal knowledge exploration for video recognition with pre-trained vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "text", + "content": "73. Yengera, G., Mutter, D., Marescaux, J., Padoy, N.: Less is more: Surgical phase recognition with less annotations through self-supervised pre-training of cnn-lstm networks. arXiv preprint arXiv:1805.08569 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 380 + ], + "type": "text", + "content": "74. Yu, F., Silva Croso, G., Kim, T.S., Song, Z., Parker, F., Hager, G.D., Reiter, A., Vedula, S.S., Ali, H., Sikder, S.: Assessment of Automated Identification of Phases in Videos of Cataract Surgery Using Machine Learning and Deep Learning Techniques. JAMA Network Open 2(4), e191860-e191860 (04 2019). https://doi.org/10.1001/jamanetworkopen.2019.1860, https://doi.org/10.1001/jamanetworkopen.2019.1860" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "75. Yu, T., Mutter, D., Marescaux, J., Padoy, N.: Learning from a tiny dataset of manual annotations: a teacher/student approach for surgical phase recognition. arXiv preprint arXiv:1812.00033 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "76. Yuan, K., Srivastav, V., Navab, N., Padoy, N.: Hecvl: Hierarchical videolanguage pretraining for zero-shot surgical phase recognition. arXiv preprint arXiv:2405.10075 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "77. Yuan, K., Srivastav, V., Yu, T., Lavanchy, J., Mascagni, P., Navab, N., Padoy, N.: Learning multi-modal representations by watching hundreds of surgical video lectures (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "78. Zha, R., Cheng, X., Li, H., Harandi, M., Ge, Z.: Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 13-23. Springer (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "79. Zhang, C.L., Wu, J., Li, Y.: Actionformer: Localizing moments of actions with transformers. In: European Conference on Computer Vision. LNCS, vol. 13664, pp. 492-510 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 589 + ], + "type": "text", + "content": "80. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI Conference on Artificial Intelligence. pp. 7590-7598 (2018), https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/17344" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "M. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_content_list.json b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1051102fa139110e070e2272a467bd90109584d9 --- /dev/null +++ b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_content_list.json @@ -0,0 +1,1843 @@ +[ + { + "type": "text", + "text": "Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaofan Que1 and Qi Yu1", + "bbox": [ + 393, + 210, + 609, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rochester Institute of Technology", + "bbox": [ + 388, + 239, + 614, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xq5054,qi.yu}@rit.edu", + "bbox": [ + 416, + 253, + 588, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. While data cleansing offers a viable solution to address noisy labels in the general learning settings, it exacerbates information loss in FSL due to limited training data, resulting in inadequate model training. To best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed unsupervised auxiliary tasks to compensate for information loss. Using unsupervised tasks can effectively avoid additional annotation costs and minimize the risk of introducing additional label noises. However, a randomly constructed unsupervised task may misguide the model to learn sample-specific features that are likely to compromise the primary few-shot learning task due to the noisy weak learning signals. We propose to conduct novel auxiliary task selection to ensure the intra-diversity among the unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold. We conduct a theoretical analysis to derive novel generalization bounds for learning with auxiliary tasks. Extensive experiments are conducted to demonstrate that our method outperforms existing noisy few-shot learning methods under various in-domain and cross-domain few-shot classification benchmarks.", + "bbox": [ + 261, + 292, + 743, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Noisy few-shot learning $\\cdot$ Auxiliary task $\\cdot$ Optimal transport", + "bbox": [ + 261, + 595, + 740, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 631, + 375, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot learning (FSL) aims to utilize knowledge acquired from a large amount of diverse tasks during training and apply it to new, unseen tasks from the same task distribution with only a small amount of labeled data [10,43,45]. Therefore, learning an accurate representation of the task manifold that effectively captures the underlying structure and relationships among tasks is essential to ensure good few-shot generalization. On the other hand, noisy labels commonly exist in the task distribution due to the ever-increasing demand for large-scale annotated datasets, where labeling errors can be introduced in various means that are impossible to control [4,30,34,47,58]. Given the few-shot nature of each training task, memorization of noisy labels introduced along with limited training data can exacerbate overfitting, which will disrupt the representation space and severely hinder few-shot generalization.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2614a906af7f80077a56aeb94198149d1d9514148a3960dc8af6f6619e283466.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 143, + 444, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/39eda97b665f88caefd54193ed4dad36fb05f025e587095472ea9c90e76db570.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 147, + 535, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c4d380a4768522a57180e6d5788fc8d265b96db48cd0eb49eb30324fba0e2ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 148, + 651, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c26f4556282ddb153d152d93bde461d3b6abdcbf33c4926b7acdf643180d2c29.jpg", + "image_caption": [ + "Data with different classes" + ], + "image_footnote": [], + "bbox": [ + 326, + 220, + 442, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b44827c96832d1e1ad99fb930421e6ee5afbe6aee990458d1e9225d5f62a322.jpg", + "image_caption": [ + "Fig. 1: (a) Data manifold of the primary task distribution. (b) Data manifold after using data cleansing to remove some noisy labels, which is incomplete due to information loss. (c) Using augmentation can make up for the information loss but it is limited by the diversity of existing data. (d) Data manifold of the auxiliary task distribution. (e) Joint data manifold of primary and auxiliary data distribution, in which the domain-specific knowledge of the auxiliary manifold is also included. (f) Through optimal transport of diverse unsupervised tasks, the learned shared representation only includes rich domain-invariant information from the auxiliary manifold." + ], + "image_footnote": [], + "bbox": [ + 457, + 222, + 547, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/066b9ed187915da508c3e67f51403bea56694dbe21373a297f8623774f866889.jpg", + "image_caption": [ + "Missing data Augmented data", + "Joint data manifold" + ], + "image_footnote": [], + "bbox": [ + 560, + 220, + 676, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While training robust models from noisy data has been intensively investigated in standard settings, learning from noisy few-shot data is only sparsely explored in existing literature [28,31]. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. Within each training task, FSL operates with few training samples per class, making each data point highly valuable. Hence, the model tends to be more sensitive to variations in the training data and noisy labels can more easily cause the model to learn incorrect patterns, compromising its ability to accurately generalize to new tasks. Furthermore, identifying and rectifying label noises also becomes more challenging. The scarcity of data makes it harder to distinguish between true class characteristics and erroneous annotations.", + "bbox": [ + 212, + 473, + 787, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Techniques, including robust optimization [28] and data augmentation [31] have been leveraged to handle noises in limited data. However, these techniques, while beneficial, do not entirely resolve the problem. Conventional methods for learning with noisy labels, such as data cleansing, data augmentations, and robust loss functions, are inadequate to tackle the challenges as outlined above. As shown in Fig. [1(a)-(b)], data cleansing introduces information loss, further diminishing the already scarce training data in each task. This poses a greater challenge for the model to learn a meaningful and generalizable representation of the task manifold. Data augmentation could alleviate the information loss introduced through data cleansing as shown in Fig. [1(c)]. Nevertheless, augmentation is performed on the limited training samples, so it may not cover a broad task manifold. Additionally, robust optimization algorithms disregard the concern of information loss, making them less suitable for learning from noisy few-shot data.", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6959258b7b5d6adb712e53cd6dca8e3ea1efe8d4cf7b902563392ed28ddd249a.jpg", + "image_caption": [ + "Fig. 2: In DUNT, the feature extractor extract information from both primary and auxiliary tasks, the classifier predicts the labels of the primary and auxiliary examples, and the domain critic outputs the source of the examples. The primary tasks (top) are sampled and cleansed by episode cleansing, the auxiliary tasks (bottom) are constructed and filtered by their diverseness. (Best viewed in color.)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 756, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the unique challenges arising from NFSL, we propose novel extensions to existing techniques designed for standard noisy learning scenarios to more effectively handle the information loss in the few-shot setting. In particular, to best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed auxiliary tasks to compensate information loss. The auxiliary tasks are entirely unsupervised as using supervised tasks can be counterproductive in the NFSL setting for three main reasons: (1) They may incur additional annotation overhead. (2) Noisy labels may continue to be introduced that further compromise the manifold structure. (3) Even clean labels can be guaranteed, these labels may not totally align with the primary tasks that can potentially overwhelm that weak learning signals carried by the limited training samples.", + "bbox": [ + 212, + 380, + 787, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While unsupervised auxiliary tasks have been leveraged in diverse learning scenarios [19,24,32,36,52], including the few-shot setting, the noisy labels and potential information loss introduced through necessary data cleansing makes NFSL much more vulnerable to the potential noises introduced by the auxiliary data. Since data samples are unlabeled, they are usually randomly sampled into an auxiliary task. When a contrastive loss is used to differentiate unlabeled data samples from the same underlying class, the model may be forced to learn sample-specific (spurious) features that cause overfitting. To this end, we propose to conduct novel auxiliary task selection to ensure the intra-diversity among the chosen unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold as shown in Fig. 1(d)-(g).", + "bbox": [ + 212, + 565, + 787, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the proposed framework, referred to as Diverse UNsupervised Tasks for NFSL (DUNT4NFSL or DUNT for short). The feature extractor of DUNT undergoes updates through three distinct loss components: 1) A cross-entropy loss, integral for extracting discriminative and effective features for classification from the primary tasks; 2) A contrastive loss tailored for the unsupervised auxiliary tasks, designed to encourage an embedding space where similar sample", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pairs remain close while dissimilar pairs are distanced; 3) An adversarial loss, implemented to align the unsupervised auxiliary tasks with the primary tasks, employing a domain critic head to reduce the covariate shift across multiple domains. To ensure that the feature representation remains unaffected by label noise, we apply an episode cleaning technique to the primary task, removing noisy examples in each task. For the auxiliary tasks, we introduce a novel criterion for diverseness to guarantee that the unsupervised auxiliary tasks contribute informative and valuable insights. Our contributions are threefolds:", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present DUNT, a framework designed to adeptly utilize unsupervised data from a distinct domain in order to counteract the information loss resulting from data cleansing in the challenging scenario of NFSL.", + "- We introduce a novel strategy to perform diverse auxiliary task selection to avoid learning sample-specific spurious features from the unlabeled auxiliary data samples.", + "- We conduct an in-depth theoretical analysis of the auxiliary tasks introduced in the DUNT framework, establishing novel generalization bounds, offering valuable insights into the contributions of these auxiliary tasks." + ], + "bbox": [ + 225, + 282, + 785, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have conducted comprehensive experiments on five benchmark datasets with synthetic label noises to demonstrate the effectiveness and robustness of DUNT and its advantage over SOTA methods by a large margin.", + "bbox": [ + 212, + 433, + 787, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 508, + 387, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Noisy few-shot learning. Existing methods for training DNNs with noisy labels usually fall into the following categories: robust loss [12,49,50,57,60] designs robust cost functions in the presence of noisy labels; noise transition estimation [5,20,53,54] aims to estimate the probabilities that clean labels flip into noisy labels; loss correction [17,22,23,39,44,55,59] either explicitly or implicitly relabels or reweights the noisy samples, a.k.a. data cleansing. Noisy label in FSL is an important yet relatively under-explored topic. The scarcity of data in FSL not only heightens its susceptibility to label noise but also adds complexity to the challenge of identifying and correcting such noise. A few previous attempts such as [28,29,31] address mislabeled issue in FSL settings using re-weighting strategy or generating robust prototypes by feature hybrid and augmentation techniques. Other works such as RW-MAML [25], AQ [14], DFSL [27] tackle the out-of-distribution tasks or adversarial attacks in few-shot learning. In this work, we particularly focus on solving the limited data problem by utilizing auxiliary unlabeled data as unsupervised tasks to enhance the robustness of FSL.", + "bbox": [ + 212, + 546, + 787, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimal transport and Wasserstein distance. The optimal transport (OT) cost is used to measure the difference between distributions supported on high-dimensional space using SGD [1,46]. For a constant $\\xi \\geq 1$ , the $\\xi$ -Wasserstein", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "metric between distributions $\\mathbb{P}_X$ and $\\mathbb{P}_{X'}$ is defined as:", + "bbox": [ + 215, + 146, + 612, + 161 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {\\xi} \\left(\\mathbb {P} _ {X}, \\mathbb {P} _ {X ^ {\\prime}}\\right) = \\left(\\inf _ {\\gamma \\in \\Pi \\left(\\mathbb {P} _ {X}, \\mathbb {P} _ {X ^ {\\prime}}\\right)} \\mathbb {E} _ {\\gamma \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right)} \\left[ c ^ {\\xi} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right) \\right]\\right) ^ {\\frac {1}{\\xi}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 172, + 785, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x},\\mathbf{x}^{\\prime}\\in \\mathbb{R}^{d}$ denote the random variables in which the distributions $\\mathbb{P}_X$ and $\\mathbb{P}_{X^{\\prime}}$ are defined. $\\prod (\\mathbb{P}_X,\\mathbb{P}_{X'})$ is the set of all joint distributions (i.e. couplings) whose marginals are $\\mathbb{P}_X$ and $\\mathbb{P}_{X'}$ , that is, $\\int \\gamma (\\mathbf{x},\\mathbf{x}')d\\mathbf{x} = \\mathbb{P}_{X'}$ and $\\int \\gamma (\\mathbf{x},\\mathbf{x}')d\\mathbf{x}' = \\mathbb{P}_X$ . $c(\\mathbf{x},\\mathbf{x}'): \\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}^+$ is the cost function for moving from $\\mathbf{x}$ to $\\mathbf{x}'$ . Intuitively, $\\gamma (\\mathbf{x},\\mathbf{x}')$ indicates how much \"mass\" must be transported from $\\mathbf{x}$ to $\\mathbf{x}'$ in order to transform the distributions $\\mathbb{P}_X$ into $\\mathbb{P}_{X'}$ . The Wasserstein metric then is the \"cost\" of the optimal transport plan [1]. Eq. [1] is the primal form of the Wasserstein metric, particularly for the case of $\\xi = 1$ with $c(\\mathbf{x},\\mathbf{x}') = \\| \\mathbf{x} - \\mathbf{x}'\\|$ [1]. Some existing efforts propose domain adaptation algorithms using ideas from OT theory, which aims to reduce the divergence between two domains by minimizing the Wasserstein distance between their distributions [7,8,11]. Works that use domain adaptation techniques in FSL [6,18,33,40,56] share the same goal that aims to learn transferable knowledge. In this work, we achieve this goal by leveraging extra data first and then minimize the Wasserstein distance between the primary and auxiliary distributions to ensure the learned knowledge are domain-invariant and discriminative to future unseen tasks.", + "bbox": [ + 215, + 220, + 787, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 487, + 380, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem formulation. In FSL, the goal is to learn a mapping $f: \\mathcal{X} \\to \\mathcal{Y}$ that maps an observation $\\mathbf{x}$ to its label $\\mathbf{y}$ , where $\\mathcal{X}$ is the feature space and $\\mathcal{Y}$ is the label space. To train the model with the quick adaptation ability to the new tasks with only a small number of examples, a set of tasks $\\mathcal{T}_i = \\{\\mathcal{S}_i^{\\mathrm{sup}},\\mathcal{S}_i^{\\mathrm{que}}\\}$ is sampled from the data distribution $\\mathcal{D}$ over $\\mathcal{X} \\times \\mathcal{Y}$ , where $S_{i}^{\\mathrm{sup}} = \\{(x_{i}^{k},y_{i}^{k})\\}_{k = 1}^{NK}$ denotes the support set and $S_{i}^{\\mathrm{que}} = \\{(x_{i}^{k},y_{i}^{k})\\}_{k = 1}^{NQ}$ denotes the query set. $N$ is the number of classes, $K$ is the number of examples in each class of the support set, and $Q$ is that of the query set. In DUNT, the meta-training tasks are sampled from the two domains: the primary tasks are sampled from the noisy primary distribution $\\tilde{\\mathcal{D}}^{\\mathrm{pri}}$ over $\\mathcal{X} \\times \\tilde{\\mathcal{Y}}$ , where the label space $\\tilde{\\mathcal{V}}$ is considered noisy, and the unsupervised auxiliary tasks are sampled from the auxiliary distribution $\\mathcal{D}^{\\mathrm{aux}}$ over $\\mathcal{X}'$ , where the entire label space is missing. We assume that the primary and the auxiliary distributions are from different domains: $\\tilde{\\mathcal{D}}^{\\mathrm{pri}} \\neq \\mathcal{D}^{\\mathrm{aux}}$ . During meta-testing, the tasks are sampled from a new unseen target domain $\\mathcal{D}^{\\mathrm{tgt}}$ over $\\mathcal{X}^{\\mathrm{tgt}} \\times \\mathcal{Y}^{\\mathrm{tgt}}$ , the label is only used for evaluation purpose. The goal is to learn a meta-model $\\theta$ trained on $\\tilde{\\mathcal{D}}^{\\mathrm{pri}}$ and $\\mathcal{D}^{\\mathrm{aux}}$ that can quickly adapt to tasks in $\\mathcal{D}^{\\mathrm{tgt}}$ using only a few examples.", + "bbox": [ + 215, + 520, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Framework overview. The DUNT framework (see Fig. 2) is composed of a shared feature extractor $\\theta$ , the task classifier $\\theta_i^c$ for task $i$ and the domain critic $\\theta^d$ . The feature extractor takes input from both the primary tasks and the auxiliary", + "bbox": [ + 215, + 794, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tasks; the classifiers predict the labels for the input and the critic predict the source of the input. The gradient reversal layer (GRL) ensures that the feature distributions over the two domains are made similar.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Learning from the Primary Tasks", + "text_level": 1, + "bbox": [ + 215, + 217, + 539, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One commonly used FSL strategy that takes advantage of task-based formulation is to learn a meta-model through a bi-level optimization. By defining the tasks-specific model parameter for task $\\mathcal{T}_i$ as $\\theta_i'$ , the meta model $\\theta$ are trained by optimizing for the performance of $\\theta_i'$ with respect to $\\theta$ across tasks sampled from $\\mathcal{D}^{\\mathrm{pri}}$ . The loss is", + "bbox": [ + 212, + 244, + 787, + 320 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {p r i}} = \\sum_ {\\mathcal {T} _ {i} \\left(\\mathcal {S} _ {i} ^ {\\sup }, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}\\right) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}} \\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(\\theta_ {i} ^ {\\prime}, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 332, + 785, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta_i' \\gets \\theta - \\alpha \\nabla_\\theta \\mathcal{L}_{\\mathcal{T}_i}(\\theta, \\mathcal{S}_i^{\\mathrm{sup}})$ is the task-specific model adapted from $\\theta$ within one gradient step. This is a dual-level optimization where in the outer-level, the meta-model $\\theta$ is updated with the summation of a series of tasks evaluated on the query sets $\\{\\mathcal{S}_i^{\\mathrm{que}}\\}$ , with $\\mathcal{L}_{\\mathcal{T}_i}(\\theta_i', \\mathcal{S}_i^{\\mathrm{que}}) = \\sum_{(\\mathbf{x}_i^k, \\mathbf{y}_i^k) \\in \\mathcal{S}_i^{\\mathrm{que}}} \\ell(\\theta_i^c(\\theta_i'(\\mathbf{x}_i^k)), \\mathbf{y}_i^k)$ . The inner-level objective is computed using the task-specific model $\\theta_i'$ updated with support set $\\mathcal{S}_i^{\\mathrm{sup}}$ using loss $\\mathcal{L}_{\\mathcal{T}_i}(\\theta, \\mathcal{S}_i^{\\mathrm{sup}}) = \\sum_{(\\mathbf{x}_i^k, \\mathbf{y}_i^k) \\in \\mathcal{S}_i^{\\mathrm{sup}}} \\ell(\\theta_i^c(\\theta(\\mathbf{x}_i^k)), \\mathbf{y}_i^k)$ , where $\\ell$ is the cross-entropy loss, $\\theta_i^c$ is the $i$ -th task classifier.", + "bbox": [ + 214, + 378, + 787, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Episode cleansing. In the dual-level optimization, the meta-model $\\theta$ is updated by first adapting to different task-specific models $\\theta_{i}^{\\prime}$ with the corresponding support set and then updated using the summation of the loss of the task-specific models $\\theta_{i}^{\\prime}$ on their corresponding query sets. Since the dataset is noisy, the error may be accumulated through the noisy query set and cause the overfitting of the meta-model. Therefore, we propose episode cleansing to remove the noisy data in the query sets based on the memorization effect which indicates that noisy examples have higher losses than the clean ones. The classification loss $\\mathcal{L}_{\\mathcal{T}_i}(\\theta_i', S_i^{\\mathrm{que}})$ becomes", + "bbox": [ + 214, + 508, + 789, + 647 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(f _ {\\theta_ {i} ^ {\\prime}}, \\hat {\\mathcal {S}} _ {i} ^ {\\mathrm {q u e}}\\right) = \\sum_ {\\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {i} ^ {k} \\in \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}} w _ {i} ^ {k} \\ell_ {i} ^ {k}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 657, + 785, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{S}_i^{\\mathrm{que}}$ is the cleansed (weighted) query set for task $\\mathcal{T}_i$ ; for simplification, we use $\\ell_i^k$ to denote $\\ell (\\theta_i^c (\\theta_i'(x_i^k)),y_i^k)$ . $w_{i}^{k}$ is the self-paced sample weight for each example pair, and", + "bbox": [ + 214, + 707, + 787, + 753 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {k} = \\mathbb {1} (\\ell_ {i} ^ {k} < \\gamma_ {t h}),\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 753, + 566, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is an indicator function whose value is 1 when $\\ell_i^k < \\gamma_{th}$ and 0 otherwise; and $\\gamma_{th}$ is a predefined hyperparameter used to filter out high-loss examples. We use the percentage of the ordered loss to avoid extensive hyperparameter search, whose ablation is provided in Appendix C.", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The primary loss the becomes:", + "bbox": [ + 238, + 146, + 464, + 161 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {p r i}} = \\sum_ {\\mathcal {T} _ {i} (\\mathcal {S} _ {i} ^ {\\sup }, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}} \\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(\\theta_ {i} ^ {\\prime}, \\hat {\\mathcal {S}} _ {i} ^ {\\mathrm {q u e}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 172, + 785, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similarly, for the support set label noise, we use meta-model to compute the loss values and remove the noisy ones in the support set.", + "bbox": [ + 214, + 218, + 784, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Learning from Diverse Unsupervised Tasks", + "text_level": 1, + "bbox": [ + 215, + 270, + 617, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the general unsupervised auxiliary task $\\mathcal{T}_j$ is constructed by i) first randomly sampling $N$ images from the dataset for the support set $\\mathcal{S}_j^{\\mathrm{sup}} = \\{\\mathbf{x}_j^k\\}_{k=1}^N$ ; ii) the query set $\\mathcal{S}_j^{\\mathrm{que}} = \\{A_1(\\mathbf{x}_j^k), \\dots, A_Q(\\mathbf{x}_j^k)\\}_{k=1}^N$ is obtained by augmenting the $N$ images in $Q$ different ways, where $A_1(\\cdot), \\dots, A_Q(\\cdot)$ denote $Q$ different augmentation techniques, such as random cropping, translation, flipping. The auxiliary loss is the summation of the contrastive loss of a series of auxiliary tasks:", + "bbox": [ + 214, + 294, + 787, + 400 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {a u x}} = \\sum_ {\\mathcal {T} _ {j} \\left(\\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right) \\sim \\mathcal {D} ^ {\\text {a u x}}} \\mathcal {L} _ {\\mathcal {T} _ {j}} \\left(\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 412, + 785, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The contrastive loss is utilized for the unsupervised tasks:", + "bbox": [ + 215, + 457, + 633, + 470 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {T} _ {j}} \\left(\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right) = - \\frac {1}{N Q} \\sum_ {k = 1} ^ {N} \\sum_ {q = 1} ^ {Q} \\log \\frac {\\exp \\left(- d \\left[ \\theta \\left(A _ {q} \\left(\\mathbf {x} _ {j} ^ {k}\\right)\\right) , \\theta \\left(\\mathbf {x} _ {j} ^ {k}\\right) \\right]\\right)}{\\sum_ {l = 1} ^ {N} \\exp \\left(- d \\left[ \\theta \\left(A _ {q} \\left(\\mathbf {x} _ {j} ^ {k}\\right)\\right) , \\theta \\left(\\mathbf {x} _ {j} ^ {l}\\right) \\right]\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 484, + 785, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $d[\\cdot, \\cdot]$ is a distance metric, such as Euclidean distance.", + "bbox": [ + 215, + 536, + 650, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Auxiliary task selection. In contrastive learning, the positive pairs are formulated by the original example and its augmentation, and the negative pairs are formulated by images from two different classes. The contrastive loss is designed to pull the positive pairs close to each other and push the negative ones away. However, since each image in the unsupervised task are randomly sampled and treated as a class, there are chances that the images are sampled from the same class. In some extreme case, all the images are sampled from the same class, which means that the positive pairs being treated as negative pairs and being push away from each other. This would cause the model being forced to overfit to the unnecessary details and creates spurious correlations and harm the generalization ability. We denote those tasks as confusing tasks, as a comparison, the tasks with images sampled from different classes are denoted as valuable tasks. The intricate nature of confusing tasks appears to pose a significant threat to the resilience and robustness of FSL, as shown in our later ablation studies. In this work, we propose a novel task selection metric to remove these confusing unsupervised tasks.", + "bbox": [ + 214, + 566, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given two images $\\mathbf{x}_j^k$ , $\\mathbf{x}_j^l$ , the model parameterized by $\\theta$ lead to class probabilities vectors $\\mathbf{p}_j^k = \\mathrm{softmax}(\\theta (\\mathbf{x}_j^k))$ and $\\mathbf{p}_j^l = \\mathrm{softmax}(\\theta (\\mathbf{x}_j^l))$ . If the two images", + "bbox": [ + 214, + 806, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are from the same class, $\\mathbf{p}_j^k$ and $\\mathbf{p}_j^l$ are naturally encouraged to be close to the assignment vector $\\mathbf{e}_y$ in $\\{0,1\\}^N$ with a single non-zero entry at position $y$ , where $y$ is the class label associated to $\\mathbf{x}$ and $N$ is the number of classes. Therefore, we consider the following diversity criterion:", + "bbox": [ + 215, + 146, + 787, + 209 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd i v _ {j} = \\frac {1}{C (N , 2)} \\sum_ {i = 1} ^ {N - 1} \\sum_ {j = i + 1} ^ {N} \\frac {1}{\\cos (\\mathbf {p} _ {j} ^ {k} , \\mathbf {p} _ {j} ^ {l})}.\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 220, + 637, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The diverseness of the task $\\mathcal{T}_j$ is defined as the average of the reciprocal of the cosine similarity between the class probabilities of all example pairs in the task, where $C(N,2) = \\frac{N!}{(N - 2)!2!}$ is the 2-combination of set $N$ . According to the definition, the higher the value of $div_j$ , the more diverse the task.", + "bbox": [ + 215, + 273, + 787, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Auxiliary loss function. Auxiliary loss $\\mathcal{L}^{\\mathrm{aux}}$ is updated as", + "bbox": [ + 215, + 349, + 630, + 366 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {a u x}} = \\sum_ {\\mathcal {T} _ {j} (\\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\mathrm {q u e}}) \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} v _ {j} \\mathcal {L} _ {\\mathcal {T} _ {j}} (\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\mathrm {q u e}})\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 378, + 658, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $v_{j}$ is the diversity weight defined as:", + "bbox": [ + 215, + 424, + 527, + 440 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nv _ {j} = \\mathbb {1} (d i v _ {j} > \\gamma_ {d i v}),\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 453, + 573, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\gamma_{div}$ is a predefined hyperparameter, which filter out those less diverse, a.k.a., confusing tasks. In practice, we simplify the hyperparameter searching by first upsampling six auxiliary tasks, computes each task diversity metric, and only keep the most diverse task in $\\{\\mathcal{T}_j\\}^{\\mathrm{val}}$ for training. The corresponding ablations about the numbers of extra tasks sampled are presented in Sec. 4.3", + "bbox": [ + 215, + 479, + 787, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Adversarial Training and the Critic Head", + "text_level": 1, + "bbox": [ + 215, + 579, + 602, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the auxiliary tasks are drawn from a different data distribution, it is inevitable that there exist some domain shift from the primary domain to the auxiliary one. While learning from the domain-specific knowledge from the auxiliary domain main cause the learned model biased and deteriorate the model's generalization performance, learning from the domain-invariant knowledge helps improve the generalization ability. To enforce the learning of the domain-invariant representation of the meta-model, we propose a critic model $\\theta^c$ to discriminate the source of the input data, either primary or auxiliary. By training the model in an adversarial way, the critic model $\\theta^c$ try to distinguish which domain that the input features come from, while the meta-model $\\theta$ aims to confuse the critic. In this way, when the training process converge, $\\theta$ tends to extract domain-invariant knowledge from both domains. Formally, the overall adversarial objective is a min-max problem as follows:", + "bbox": [ + 215, + 604, + 787, + 801 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , \\theta^ {c}} \\max _ {\\theta^ {d}} \\alpha_ {1} \\mathcal {L} ^ {\\text {p r i}} + (1 - \\alpha_ {1}) \\mathcal {L} ^ {\\text {a u x}} + \\alpha_ {2} \\mathcal {E} (\\theta , \\theta^ {d}, \\tilde {\\mathcal {D}} ^ {\\text {p r i}}, \\mathcal {D} ^ {\\text {a u x}}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 811, + 784, + 835 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the critic model $\\theta^d$ assigns 1 to the example from the primary domain and 0 otherwise based on the feature representation extracted by $\\theta$ . $\\mathcal{E}(\\cdot)$ is the adversarial loss", + "bbox": [ + 214, + 146, + 787, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} (\\theta , \\theta^ {d}, \\hat {\\mathcal {D}} ^ {\\mathrm {p r i}}, \\mathcal {D} ^ {\\mathrm {a u x}}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\hat {\\mathcal {D}} ^ {\\mathrm {p r i}}} [ g (\\mathbf {x}) ] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} [ g (\\mathbf {x}) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 202, + 689, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $g(\\mathbf{x}) = \\theta^d (\\theta (\\mathbf{x}))$ , with its empirical counterpart as", + "bbox": [ + 214, + 232, + 625, + 250 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {E}} (g _ {\\theta , \\theta^ {d}}, \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}, \\mathcal {D} ^ {\\mathrm {a u x}}) = \\frac {1}{| \\mathcal {D} ^ {\\mathrm {p r i}} |} \\sum_ {x \\sim \\mathcal {D} ^ {\\mathrm {p r i}}} [ g (\\mathbf {x}) ] - \\frac {1}{| \\mathcal {D} ^ {\\mathrm {a u x}} |} \\sum_ {\\mathbf {x} \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} [ g (\\mathbf {x}) ]\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 258, + 723, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The optimization of Eq. (7) is conducted through an alternative training between the optimization of the meta-model feature extractor $\\theta$ , classifier parameter $\\theta^c$ and the critic model parameter $\\theta^d$ while keeping the other fixed. Parameters of $\\theta$ , $\\theta^c$ and $\\theta^d$ are updated alternatively. In practice, $\\theta$ and $\\theta^c$ are connected by the GRL, which reverses the gradients that flow through $\\theta^c$ . The pseudo code of the algorithm can be found in Appendix A.", + "bbox": [ + 214, + 306, + 787, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Theoretical Analysis", + "text_level": 1, + "bbox": [ + 215, + 420, + 431, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct a theoretical analysis of DUNT to show: 1) the equivalence between the adversarial regularizer and the Wasserstein distance, which captures the diversity property of the auxiliary distribution while enforcing the learning of a domain invariant representation; 2) the relationship between the primary and auxiliary distributions; 3) a tight generalization bound of the target based on the expected risk of the primary and auxiliary distributions.", + "bbox": [ + 214, + 446, + 787, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Equivalence to Wasserstein distance. The Kantorovich-Rubinstein of Wasserstein distance takes the form [1]:", + "bbox": [ + 214, + 560, + 803, + 593 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} (\\mathbb {P} _ {1}, \\mathbb {P} _ {2}) = \\sup _ {\\| h \\| _ {L} \\leq 1} \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {1}} [ h (\\mathbf {x}) ] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {2}} [ h (\\mathbf {x}) ] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 604, + 784, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the supremum is over all the 1-Lipschitz functions $h: \\mathcal{X} \\to \\mathbb{R}$ . Therefore, if we have a parameterized family of functions $\\{h_w\\}_{w \\in \\mathcal{W}}$ that are all 1-Lipschitz, we could consider solving the problem", + "bbox": [ + 214, + 640, + 787, + 686 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\max \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {1}} \\left[ h _ {w} (\\mathbf {x}) \\right] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {2}} \\left[ h _ {w} (\\mathbf {x}) \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 699, + 784, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "which has the same form of the regularizer in Sec. 3.3. This indicates that our adversarial regularizer has the same theoretical foundation of Wasserstein distance optimal transport. To satisfy the 1-Lipschitz constraint, the critic $\\theta^d$ is implemented with gradient penalty [15]. The overall objective can be reformulated as", + "bbox": [ + 214, + 726, + 787, + 789 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\alpha_ {1} \\mathcal {L} ^ {\\text {p r i}} + (1 - \\alpha_ {1}) \\mathcal {L} ^ {\\text {a u x}} + \\alpha_ {2} \\mathcal {W} _ {1} (\\mathcal {D} ^ {\\text {p r i}}, \\mathcal {D} ^ {\\text {a u x}}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 813, + 784, + 837 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the first and second term are the primary and auxiliary loss, resp., and the third term is the Wasserstein distance based regularizer, which aims to extract domain-invariant feature from the primary and auxiliary domains. Wasserstein distance reflects an optimal transport cost for moving one distribution to another. A smaller transport cost means a better coverage of the distribution, which captures the property of diversity [42].", + "bbox": [ + 212, + 146, + 787, + 239 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Relationship of the primary and auxiliary distributions. By viewing this problem as a distribution matching that matches the auxiliary distribution with the primary one, we have the generalization property that quantify the relationship between the primary and the auxiliary distribution. Given the primary distribution $\\mathcal{D}^{\\mathrm{pri}}$ and the auxiliary distribution $\\mathcal{D}^{\\mathrm{aux}}$ , the expected risk w.r.t. is defined as $\\mathcal{D}^{\\mathrm{pri}}R_{\\mathcal{D}^{\\mathrm{pri}}} = \\mathbb{E}_{x\\sim \\mathcal{D}^{\\mathrm{pri}}}\\ell (h(x),h^{*}(x))$ , and the expected risk w.r.t. $\\mathcal{D}^{\\mathrm{aux}}$ is defined as $R_{\\mathcal{D}^{\\mathrm{aux}}} = \\mathbb{E}_{x\\sim \\mathcal{D}^{\\mathrm{aux}}}\\ell (h(x),h^{*}(x))$ . We have the following Lemma 1:", + "bbox": [ + 212, + 258, + 787, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lemma 1. If the loss $\\ell$ is symmetric and $L$ -Lipschitz, $\\forall h \\in \\mathcal{H}$ is at most $H$ -Lipschitz function and the underlying mapping function $h^*$ is $\\phi(\\lambda) - (\\mathcal{D}^{pri}, \\mathcal{D}^{aux})$ joint probabilistic Lipschitz; then the expected risk w.r.t. $\\mathcal{D}^{pri}$ is upper bounded", + "bbox": [ + 212, + 372, + 787, + 417 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathcal {D} ^ {p r i}} \\leq R _ {\\mathcal {D} ^ {a u x}} + L (H + \\lambda) \\mathcal {W} _ {1} (\\mathcal {D} ^ {p r i}, \\mathcal {D} ^ {a u x}) + L \\phi (\\lambda), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 428, + 785, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\phi : \\mathbb{R} \\to [0,1]$ and a mapping $h^*$ is $\\phi(\\lambda) - (\\mathcal{D}^{pri},\\mathcal{D}^{aux})$ joint probabilistic Lipschitz if $\\text{supp}(\\mathcal{D}^{aux}) \\subseteq \\text{supp}(\\mathcal{D}^{pri})$ ; $\\forall \\lambda > 0$ , all coupling $\\gamma \\in \\prod(\\mathcal{D}^{pri},\\mathcal{D}^{aux}) : \\mathbb{P}_{(x_{\\mathcal{D}^{pri}},x_{\\mathcal{D}^{aux}})\\sim \\gamma}[|h^{*}(x_{\\mathcal{D}^{pri}}) - h^{*}(x_{\\mathcal{D}^{aux}})| > \\lambda \\| x_{\\mathcal{D}^{pri}} - x_{\\mathcal{D}^{aux}}\\|_2] \\leq \\phi(\\lambda)$ .", + "bbox": [ + 214, + 455, + 787, + 503 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lemma 1 shows that the expected risk of the primary distributions is upper bounded by the expected risk w.r.t. the auxiliary distribution, the Wasserstein distance, and the labeling function property. The proof of Lemma 1 can be derived with Theorem 1 in [42]. Eq. (10) follows the upper bound from Lemma 1, which means a desirable auxiliary distribution with a better matching with the primary distribution which holds a small expected risk leads to tighter generalization bound for the primary distribution. This further verifies the rationality of our design of using the WD based regularizer.", + "bbox": [ + 212, + 508, + 787, + 630 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Generalization bound of DUNT. Furthermore, in the following Lemma 2 we derive the upper bound for the target risk $R_{\\mathcal{D}^{\\mathrm{tgt}}} = \\mathbb{E}_{(\\mathbf{x},y)\\sim \\mathcal{D}^{\\mathrm{tgt}}}\\ell (\\mathbf{x},y)$ .", + "bbox": [ + 212, + 651, + 785, + 683 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lemma 2. Let $\\mathcal{D}^{tgt}$ be the target distribution, $h$ is $K$ -Lipschitz w.r.t. the feature $\\mathbf{x}$ , and the setting of loss and hypothesis be the same as Lemma 1, the expected risk w.r.t. the target distribution is upper-bounded by:", + "bbox": [ + 212, + 691, + 787, + 738 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R _ {\\mathcal {D} ^ {t g t}} \\leq \\alpha_ {1} \\left(R _ {\\mathcal {D} ^ {p r i}} + L K \\mathcal {W} _ {1} \\left(\\mathcal {D} ^ {p r i}, \\mathcal {D} ^ {t g t}\\right)\\right) \\tag {12} \\\\ + (1 - \\alpha_ {1}) \\left(R _ {\\mathcal {D} ^ {a u x}} + L K \\mathcal {W} _ {1} \\left(\\mathcal {D} ^ {a u x}, \\mathcal {D} ^ {t g t}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 747, + 784, + 782 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lemma 2 shows that the expected risk w.r.t. the target distribution is upper bounded by the expected risk w.r.t. the primary distribution and the auxiliary distribution, and the corresponding Wasserstein distances. The proof of Lemma 2", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "can be derived with Theorem 2 in [41]. Lemma 2 provide solid theoretical foundation for the design of DUNT with a tight generalization bound to the unseen target domain with primary and auxiliary distributions, which can be extended to multiple sources.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 231, + 375, + 248 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Datasets, Baselines, and Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 263, + 650, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments using the following benchmark datasets for FSL: CUB 51, miniImageNet (mini) 48, tieredImageNet (tiered) 38, CIFAR-FS (CF) 3, and Cross-Domain dataset 16. Each datasets contains synthetic labels noise with different ratios. More details are in the Appendix.", + "bbox": [ + 214, + 289, + 785, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We use ProtoNet [43] as the base few-shot learning method and consider comparing with improving methods with general data cleansing methods including SPL [26] and FSR [61], denoted as $PN + SPL$ and $PN + FSR$ , respectively. We also consider to improve ProtoNet in transductive settings including CoTeaching [17], CoTeaching+ [59], denoted as $PN + Co$ , and $PN + Co+$ , respectively, STARTUP [35], and DDNet [21]. Last but not least, we consider to compare with noisy FSL methods RNNP [31], TraNFS [28] and DCML [37]. The detailed descriptions and settings are provided in the Appendix.", + "bbox": [ + 212, + 366, + 787, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental settings. The feature extractor follows the same architecture as the ProtoNet used by [43], which has 4 modules with a $3 \\times 3$ convolutions and 64 filters, followed by batch normalization, a ReLU nonlinearity, and $2 \\times 2$ max-pooling, a.k.a., Conv4. The architecture of classifier have the design with a fully connected layer followed by a ReLU nonlinearity, a dropout layer, and another fully connected layer. The critic architecture has the same architecture with an extra sigmoid function. We've considered transductive and non-transductive in the settings. For the former, unlabeled test data of the same dataset is used as auxiliary data during training. This is applied to our method and other transductive methods including $\\mathrm{PN} + \\mathrm{Co}$ , $\\mathrm{PN} + \\mathrm{Co}+$ , STARTUP, and DDNet. For the latter, we utilize unlabeled datasets from other different datasets as auxiliary. This setting is applied to all the mentioned methods. During training, we use a learning rate $=0.001$ , and use Adam as the optimizer. The batch size is 4. The meta-model is trained with 40000 randomly selected primary tasks and 40000 randomly selected auxiliary tasks. During evaluation, the average accuracy is computed based on 600 randomly sampled tasks.", + "bbox": [ + 212, + 503, + 787, + 746 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Experimental Results", + "text_level": 1, + "bbox": [ + 214, + 768, + 490, + 782 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following the experiment design of [48], we report the average accuracy of 5 runs of each 5-way 1-shot and 5-way 5-shots experiments in Tab. [1] to Tab. [2]. We use $\\alpha_{1} = 0.5$ , $\\alpha_{2} = 0.02$ , $\\gamma_{th} = 50\\%$ , $\\gamma_{div}$ as one out of six tasks for our methods", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4fc70739048743b7d1b73b1a22b10e5d9e924ef46c005f70882e1a4b449f511a.jpg", + "table_caption": [ + "Table 1: Non-transductive auxiliary datasets" + ], + "table_footnote": [], + "table_body": "
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+SPL48.55 ± 1.5143.44 ± 1.4937.06 ± 0.6342.11 ± 0.7155.20 ± 0.69
PN+FSR37.14 ± 0.6044.81 ± 0.5939.82 ± 0.6830.65 ± 0.4937.39 ± 0.63
RNNP55.39 ± 0.4950.18 ± 0.6241.75 ± 0.7243.76 ± 0.4860.15 ± 0.72
TraNFS53.02 ± 0.5246.66 ± 0.6338.25 ± 0.7940.32 ± 0.8860.86 ± 0.63
DCML54.39 ± 0.8749.33 ± 1.2241.77 ± 0.8048.33 ± 0.8060.28 ± 0.82
DUNT56.94 ± 1.0352.22 ± 1.1043.01 ± 0.5149.12 ± 0.6862.80 ± 0.49
Task5-way 5-shot
PN+SPL68.47 ± 0.7458.67 ± 0.8654.94 ± 1.0359.62 ± 1.3170.93 ± 0.83
PN+FSR51.48 ± 0.9161.89 ± 0.8057.38 ± 1.1050.40 ± 1.2144.68 ± 1.09
RNNP70.37 ± 1.3363.57 ± 1.2858.79 ± 2.3655.07 ± 1.2174.31 ± 1.36
TraNFS68.77 ± 1.1761.06 ± 1.1351.24 ± 1.1648.74 ± 1.2074.15 ± 0.83
DCML71.28 ± 1.2362.77 ± 1.0057.77 ± 1.3361.19 ± 1.0172.33 ± 0.90
DUNT72.17 ± 1.1165.10 ± 1.2358.50 ± 1.1162.01 ± 1.0974.44 ± 1.42
", + "bbox": [ + 320, + 171, + 687, + 340 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "in all datasets. In Tab. 1 and Tab. 2 we have provided the test accuracy all indomain datasets and one cross-domain dataset: mini-CUB for both transductive and non-transductive settings. The performance of cross-domain dataset 16 is reported in Tab. 3", + "bbox": [ + 212, + 368, + 787, + 433 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Non-transductive results. For Tab. DUNT uses miniimagenet as auxiliary data for CUB datasets, uses tieredimagnet as auxiliary data for the cross-domain mini-CUB, and uses CUB as auxiliary data for the rest datasets. The results show that, with extra auxiliary datasets, DUNT has outperformed the general data-cleansing techniques and other noisy FSL methods in general. In Tab. the performance of SPL and FSR are generally worse than that of TraNFS and RNNP, which confirms our assumption that by removing potential noisy data, the data cleansing technique might harm the performance of FSL due to the lack of information, which can be compensated with our auxiliary data design. And compared to TraNFS and RNNP, DUNT has reduced the negative impact of the noisy labels more effectively.", + "bbox": [ + 212, + 445, + 787, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Transductive results. In Tab. 2, we consider the transductive setting where the auxiliary dataset for all the methods are unlabeled test data of the same dataset. The results show that our method ranks the top among all the transductive methods. Compared to Tab. 1, some of the transductive methods underperform non-transductive ones due to the following potential reasons: i) the test dataset is smaller than other auxiliary datasets. Therefore, the non-transductive dataset has more information to provide. Take miniImageNet as an example, the test data of the miniImageNet dataset only contains 1000 images and 10 classes, while the auxiliary dataset CUB contains 200 categories and 12000 images, the abundance and diversity of the auxiliary helps greatly improve the model's performance over the primary tasks more than the close test data does; and ii) the auxiliary dataset provides more valuable fine-grained tasks that benefits representation learning. For instance, when training miniImageNet with CUB as an auxiliary, both datasets contain natural images which are closely related, and", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cd308ae470bfe9533e4b31bb663e01724ad4a5b58a7f20bf4a5f16a5040319df.jpg", + "table_caption": [ + "Table 2: Transductive auxiliary datasets" + ], + "table_footnote": [], + "table_body": "
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+Co49.10 ± 1.2144.44 ± 0.9636.48 ± 0.9840.36 ± 0.8152.51 ± 0.92
PN+Co+49.45 ± 1.1945.08 ± 0.9139.54 ± 1.0940.19 ± 1.1051.27 ± 1.11
STARTUP52.02 ± 1.2048.79 ± 1.1840.08 ± 1.2244.48 ± 0.9260.69 ± 1.31
DDNet52.55 ± 1.3950.82 ± 0.8741.11 ± 1.8945.11 ± 1.3360.25 ± 1.46
DUNT55.01 ± 1.1252.38 ± 0.8542.07 ± 1.1049.29 ± 1.2661.90 ± 1.14
Task5-way 5-shot
PN+Co65.42 ± 1.2259.57 ± 1.2056.90 ± 1.2955.10 ± 1.2166.89 ± 1.26
PN+Co+67.92 ± 1.1160.43 ± 1.1354.29 ± 1.1054.54 ± 0.8467.22 ± 0.71
STARTUP68.74 ± 1.2162.00 ± 1.2451.07 ± 1.0155.92 ± 0.9974.04 ± 0.84
DDNet69.21 ± 1.3463.70 ± 2.0153.87 ± 0.4457.22 ± 0.7573.92 ± 1.20
DUNT71.20 ± 1.1266.46 ± 0.9859.05 ± 1.0161.09 ± 0.9076.11 ± 0.98
", + "bbox": [ + 318, + 171, + 687, + 316 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c50f63ef4a795b4dae068d57e2695e5857024b7b19ace1c888fada52c4ea5f2e.jpg", + "table_caption": [ + "Table 3: Cross-domain Datasets" + ], + "table_footnote": [], + "table_body": "
MethodsEuroSATCropDiseaseChestXISICEuroSATCropDiseaseChestXISIC
5-way 1-shot5-way 5-shot
ProtoNet52.52±0.5256.32±0.8928.90±0.4330.11±1.0061.60±0.7078.89±0.8333.04±0.7136.95±0.90
PN+SPL52.89±1.2158.95±2.0427.68±0.5231.79±1.0967.60±1.2380.25±1.1931.13±1.1837.05±1.15
PN+FSR47.53±1.2156.97±1.2328.40±0.6133.58±0.9261.96±0.9075.93±0.8631.43±1.1936.42±1.22
RNNP51.86±1.2056.56±1.2127.39±0.6131.26±0.8067.97±1.3478.43±1.2931.16±1.2537.84±1.19
TraNFS50.61±1.1058.39±0.9128.26±0.9832.84±1.1167.93±1.1181.51±1.2332.16±1.1636.95±1.18
DCML52.33±1.0258.36±0.4627.54±0.6931.96±0.8068.01±0.9079.11±0.9831.64±1.0035.44±1.10
PN+Co56.56±1.0859.13±1.1426.67±0.7329.16±1.1372.08±1.1578.59±2.1130.45±1.1734.74±1.09
PN+Co+52.18±1.1158.18±1.1028.23±0.8730.63±1.0665.57±1.1078.47±0.9030.77±0.8233.68±0.91
STARTUP52.74±1.2259.43±1.2327.75±0.9532.84±1.1369.78±1.1780.12±1.2231.69±1.3039.58±0.99
DDNet55.21±1.8360.09±1.0928.66±0.9834.10±1.0170.09±1.9081.90±2.0932.01±1.8939.44±1.13
DUNT57.10±1.2961.49±1.0129.78±0.9035.80±0.7772.81±1.1083.20±1.0833.94±1.2040.31±0.90
", + "bbox": [ + 220, + 354, + 787, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "CUB dataset is a collection of bird species, which provide fine-grained tasks for learning a better feature representation, while the miniImageNet test data only provide coarse-grained task, which makes it less useful in some extend. As for DUNT, the performance stays stable regardless the two settings, showing that our framework has the potential to fully utilize any extra information provided.", + "bbox": [ + 212, + 526, + 787, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Cross-domain results. In Tab. 3, we present the performance on the cross-domain datasets. The training procedure follows the setting in [16], which uses miniImageNet as base dataset for all the methods. For CoTeaching, CoTeaching+, STARTUP, DDNet and DUNT, we use the unlabeled test data EuroSAT, CropDisease, ChestX, and ISIC as auxiliary dataset, respectively. In cross-domain datasets, the transductive methods generally outperform the non-transductive ones due to the utilization of the unlabeled test data that reduces the domain gap between the source domain and the target domain, which is critical when the they are extremely different. DUNT has used the Wasserstein regularization to further reduce the differences. Hence, it ranks the highest in the cross-domain datasets.", + "bbox": [ + 212, + 614, + 795, + 765 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 786, + 388, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we study the effectiveness of different components and impact of hyperparameters. For more ablations, please refer to Appendix C.", + "bbox": [ + 212, + 809, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of different components. Tab. 4 shows the ablation study with different elements of DUNT on CUB and miniImageNet datasets under nontransductive setting. We first show the performance of only using the primary data, which is vanilla ProtoNet (Primary). It can be observed that each of these components have a significant impact on the overall performance.", + "bbox": [ + 215, + 146, + 498, + 296 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/46f7933062526486d7174b0fe53a1ae0c994bb130ca686c842397be0c466c999.jpg", + "table_caption": [ + "Table 4: Impact of different components" + ], + "table_footnote": [], + "table_body": "
PrimaryAuxiliaryRegECTSCUBmini
✓XXXX42.57 ± 0.3439.95 ± 1.06
✓XX✓X48.55 ± 1.5143.44 ± 1.49
✓✓XXX50.19 ± 0.7148.90 ± 0.77
✓✓X✓X53.55 ± 0.9750.02 ± 0.81
✓✓XX✓52.70 ± 1.0149.50 ± 1.66
✓✓X✓✓54.31 ± 0.9950.18 ± 0.87
✓✓✓✓X51.90 ± 0.8949.18 ± 0.76
✓✓✓✓X55.47 ± 0.7950.73 ± 0.60
✓✓✓X✓54.79 ± 1.1049.91 ± 1.00
✓✓✓✓✓56.94 ± 1.0352.22 ± 1.10
", + "bbox": [ + 511, + 162, + 777, + 286 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Auxiliary data (Auxiliary) and episode cleansing (EC) serve pivotal functions, with the former introducing additional data and the latter eliminating noisy data. The regularization term (Reg) aids in extracting domain-invariant features, contributing to further performance improvement. Task selection (TS) mitigates the risk of overfitting, thereby enhancing overall performance.", + "bbox": [ + 212, + 297, + 784, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of task selection. We study the impact of task selection through the threshold $\\gamma_{div}$ . As mentioned earlier, to streamline the hyperparameter search process, we opt for an approach that involves upsampling $M$ unsupervised tasks. Subsequently,", + "bbox": [ + 212, + 383, + 431, + 520 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/119c14515f988d49b3272c985d6487b8da064817c7b0187310459b8a882185c9.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 447, + 383, + 604, + 452 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c2eed60c02f5ad14d0614b1639173f300d880f92f1f4932967b26cdfbec4e02d.jpg", + "image_caption": [ + "(b)", + "Fig. 3: DUNT on CUB with different $\\gamma_{div}$ (a); DUNT with different task selection (b)" + ], + "image_footnote": [], + "bbox": [ + 617, + 383, + 769, + 452 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "we select only the most diverse task to be included in the valuable task group, instead of specifying a real value for the diverseness threshold. In this ablation, the number of upsampled tasks varies from 2-10. As depicted in Fig. 3 (a), the performance improves alongside an increase in the parameter $M$ , but this comes at the cost of heightened computational demands. Consequently, considering efficiency, we opt for $\\gamma_{\\mathrm{div}} = 6$ for all experiments. In Fig. 3 (b), we test the effect of using different unsupervised task: select confusing tasks and select valuable tasks. It is evident from the results that the inclusion of confusing tasks significantly deteriorates performance, whereas the utilization of valuable tasks leads to a substantial improvement.", + "bbox": [ + 212, + 520, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 690, + 359, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose a novel FSL framework with auxiliary tasks that utilize carefully selected unlabeled data under noisy settings. To better align the auxiliary distribution with the primary one, we propose a regularization term based on the Wasserstein distance for learning a domain-invariant representation. To enhance robustness against label noise, we introduce episode cleansing for examples in the primary task and adopt a diverse task selection strategy for the unsupervised auxiliary tasks. Our framework is theoretically and experimentally proved to be effective and beneficial.", + "bbox": [ + 212, + 718, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017)", + "2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009)", + "3. Bertinetto, L., Henriques, J.F., Torr, P.H., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: Proceedings of the International Conference on Learning Representations (2019)", + "4. Beyer, L., Henaff, O.J., Kolesnikov, A., Zhai, X., Oord, A.v.d.: Are we done with imagenet? arXiv preprint arXiv:2006.07159 (2020)", + "5. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022)", + "6. Chobola, T., Vašata, D., Kordík, P.: Transfer learning based few-shot classification using optimal transport mapping from preprocessed latent space of backbone neural network. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. pp. 29-37. PMLR (2021)", + "7. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Advances in Neural Information Processing Systems 30 (2017)", + "8. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 274-289. Springer (2014)", + "9. Deleu, T., Würfl, T., Samiei, M., Cohen, J.P., Bengio, Y.: Torchmeta: A Meta-Learning library for PyTorch (2019), https://arxiv.org/abs/1909.06576, available at: https://github.com/tristandeleu/pytorch-meta", + "10. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. pp. 1126-1135 (2017)", + "1. Flamary, R., Courty, N., Tuia, D., Rakotomamonjy, A.: Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell 1 (2016)", + "2. Ghosh, A., Kumar, H., Sastry, P.S.: Robust loss functions under label noise for deep neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)", + "3. Gidaris, S., Bursuc, A., Komodakis, N., Pérez, P., Cord, M.: Boosting few-shot visual learning with self-supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8059-8068 (2019)", + "4. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020)", + "5. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)", + "6. Guo, Y., Codella, N.C., Karlinsky, L., Codella, J.V., Smith, J.R., Saenko, K., Rosing, T., Feris, R.: A broader study of cross-domain few-shot learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 124-141. Springer (2020)" + ], + "bbox": [ + 225, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Proceedings of the Conference on Neural Information Processing Systems (2018)", + "18. Han, C., Fan, Z., Zhang, D., Qiu, M., Gao, M., Zhou, A.: Meta-learning adversarial domain adaptation network for few-shot text classification. In: Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021. Findings of ACL, vol. ACL/IJCNLP 2021, pp. 1664-1673. Association for Computational Linguistics (2021)", + "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020)", + "20. Hendrycks, D., Mazeika, M., Wilson, D., Gimpel, K.: Using trusted data to train deep networks on labels corrupted by severe noise. Advances in neural information processing systems 31 (2018)", + "21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Feris, R., Radke, R.J.: Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems 34, 3584-3595 (2021)", + "22. Jiang, L., Meng, D., Mitamura, T., Hauptmann, A.G.: Easy samples first: Self-paced reranking for zero-example multimedia search. In: Proceedings of the 22nd ACM international conference on Multimedia. pp. 547-556 (2014)", + "23. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018)", + "24. Khodadadeh, S., Boloni, L., Shah, M.: Unsupervised meta-learning for few-shot image classification. Advances in neural information processing systems 32 (2019)", + "25. Killamsetty, K., Li, C., Zhao, C., Iyer, R., Chen, F.: A reweighted meta learning framework for robust few shot learning. arXiv preprint arXiv:2011.06782 (2020)", + "26. Kumar, M.P., Packer, B., Koller, D.: Self-paced learning for latent variable models. In: Proceedings of the Conference on Neural Information Processing Systems. vol. 1, p. 2 (2010)", + "27. Li, W., Wang, L., Zhang, X., Qi, L., Huo, J., Gao, Y., Luo, J.: Defensive few-shot learning. IEEE transactions on pattern analysis and machine intelligence (2022)", + "28. Liang, K.J., Rangrej, S.B., Petrovic, V., Hassner, T.: Few-shot learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9089-9098 (2022)", + "29. Lu, J., Jin, S., Liang, J., Zhang, C.: Robust few-shot learning for user-provided data. IEEE transactions on neural networks and learning systems 32(4), 1433-1447 (2020)", + "30. Masi, I., Trn, A.T., Hassner, T., Sahin, G., Medioni, G.: Face-specific data augmentation for unconstrained face recognition. International Journal of Computer Vision 127, 642-667 (2019)", + "31. Mazumder, P., Singh, P., Namboodiri, V.P.: Rnnp: A robust few-shot learning approach. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2664-2673 (2021)", + "32. Medina, C., Devos, A., Grossglauer, M.: Self-Supervised Prototypical Transfer Learning for Few-Shot Classification. arXiv preprint arXiv:2006.11325 (2020)", + "33. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in neural information processing systems 30 (2017)", + "34. Northcutt, C.G., Athalye, A., Mueller, J.: Pervasive label errors in test sets destabilize machine learning benchmarks. arXiv preprint arXiv:2103.14749 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Phoo, C.P., Hariharan, B.: Self-training for few-shot transfer across extreme task differences. In: Proceedings of the International Conference on Learning Representations (2021)", + "36. Qin, T., Li, W., Shi, Y., Gao, Y.: Diversity helps: Unsupervised few-shot learning via distribution shift-based data augmentation. arXiv preprint arXiv:2004.05805 (2020)", + "37. Que, X., Yu, Q.: Dual-level curriculum meta-learning for noisy few-shot learning tasks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 14740-14748 (2024)", + "38. Ren, M., Triantafillou, E., Ravi, S., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: Proceedings of the International Conference on Learning Representations (2018)", + "39. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: International Conference on Machine Learning. pp. 4334-4343. PMLR (2018)", + "40. Sahoo, D., Le, H., Liu, C., Hoi, S.C.: Meta-learning with domain adaptation for few-shot learning under domain shift (2018)", + "41. Shui, C., Li, Z., Li, J., Gagné, C., Ling, C.X., Wang, B.: Aggregating from multiple target-shifted sources. In: International Conference on Machine Learning. pp. 9638-9648. PMLR (2021)", + "42. Shui, C., Zhou, F., Gagné, C., Wang, B.: Deep active learning: Unified and principled method for query and training. In: International Conference on Artificial Intelligence and Statistics. pp. 1308-1318. PMLR (2020)", + "43. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 4077-4087 (2017)", + "44. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: Advances in neural information processing systems (2020)", + "45. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019)", + "46. Tolstikhin, I., Bousquet, O., Gelly, S., Schoelkopf, B.: Wasserstein auto-encoders. In: International Conference on Machine Learning (2017)", + "47. Tsipras, D., Santurkar, S., Engstrom, L., Ilyas, A., Madry, A.: From imagenet to image classification: Contextualizing progress on benchmarks. In: International Conference on Machine Learning. pp. 9625-9635. PMLR (2020)", + "48. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. In: Advances in neural information processing systems. pp. 3630-3638 (2016)", + "49. Wang, X., Hua, Y., Kodirov, E., Robertson, N.M.: Imae for noise-robust learning: Mean absolute error does not treat examples equally and gradient magnitude's variance matters. arXiv preprint arXiv:1903.12141 (2019)", + "50. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019)", + "51. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-ucsd birds 200 (2010)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Robust Learning from Noisy Few-Shot Data", + "bbox": [ + 436, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018)", + "53. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems 33, 7597-7610 (2020)", + "54. Xia, X., Liu, T., Wang, N., Han, B., Gong, C., Niu, G., Sugiyama, M.: Are anchor points really indispensable in label-noise learning? Advances in neural information processing systems 32 (2019)", + "55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020)", + "56. Xu, R., Yang, X., Liu, B., Zhang, K., Liu, W.: Selecting task with optimal transport self-supervised learning for few-shot classification. arXiv preprint arXiv:2204.00289 (2022)", + "57. Xu, Y., Cao, P., Kong, Y., Wang, Y.: L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems 32 (2019)", + "58. Yang, Y., Liang, K.J., Carin, L.: Object detection as a positive-unlabeled problem. arXiv preprint arXiv:2002.04672 (2020)", + "59. Yu, X., Han, B., Yao, J., Niu, G., Tsang, I., Sugiyama, M.: How does disagreement help generalization against label corruption? In: International Conference on Machine Learning. pp. 7164-7173. PMLR (2019)", + "60. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems 31 (2018)", + "61. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 549 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Que and Yu", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_model.json b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..efa75b93d582cb86fdcf6efe3245ef1011a5f821 --- /dev/null +++ b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_model.json @@ -0,0 +1,2557 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.212, + 0.61, + 0.23 + ], + "angle": 0, + "content": "Xiaofan Que1 and Qi Yu1" + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.24, + 0.615, + 0.255 + ], + "angle": 0, + "content": "Rochester Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.255, + 0.589, + 0.269 + ], + "angle": 0, + "content": "{xq5054,qi.yu}@rit.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.293, + 0.744, + 0.584 + ], + "angle": 0, + "content": "Abstract. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. While data cleansing offers a viable solution to address noisy labels in the general learning settings, it exacerbates information loss in FSL due to limited training data, resulting in inadequate model training. To best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed unsupervised auxiliary tasks to compensate for information loss. Using unsupervised tasks can effectively avoid additional annotation costs and minimize the risk of introducing additional label noises. However, a randomly constructed unsupervised task may misguide the model to learn sample-specific features that are likely to compromise the primary few-shot learning task due to the noisy weak learning signals. We propose to conduct novel auxiliary task selection to ensure the intra-diversity among the unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold. We conduct a theoretical analysis to derive novel generalization bounds for learning with auxiliary tasks. Extensive experiments are conducted to demonstrate that our method outperforms existing noisy few-shot learning methods under various in-domain and cross-domain few-shot classification benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.597, + 0.741, + 0.612 + ], + "angle": 0, + "content": "Keywords: Noisy few-shot learning \\(\\cdot\\) Auxiliary task \\(\\cdot\\) Optimal transport" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.376, + 0.646 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Few-shot learning (FSL) aims to utilize knowledge acquired from a large amount of diverse tasks during training and apply it to new, unseen tasks from the same task distribution with only a small amount of labeled data [10,43,45]. Therefore, learning an accurate representation of the task manifold that effectively captures the underlying structure and relationships among tasks is essential to ensure good few-shot generalization. On the other hand, noisy labels commonly exist in the task distribution due to the ever-increasing demand for large-scale annotated datasets, where labeling errors can be introduced in various means that are impossible to control [4,30,34,47,58]. Given the few-shot nature of each training task, memorization of noisy labels introduced along with limited training data can exacerbate overfitting, which will disrupt the representation space and severely hinder few-shot generalization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.144, + 0.445, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.148, + 0.537, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.149, + 0.653, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.222, + 0.444, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.223, + 0.549, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.222, + 0.678, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.312, + 0.435, + 0.322 + ], + "angle": 0, + "content": "Data with different classes" + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.313, + 0.583, + 0.322 + ], + "angle": 0, + "content": "Missing data Augmented data" + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.313, + 0.675, + 0.322 + ], + "angle": 0, + "content": "Joint data manifold" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Fig. 1: (a) Data manifold of the primary task distribution. (b) Data manifold after using data cleansing to remove some noisy labels, which is incomplete due to information loss. (c) Using augmentation can make up for the information loss but it is limited by the diversity of existing data. (d) Data manifold of the auxiliary task distribution. (e) Joint data manifold of primary and auxiliary data distribution, in which the domain-specific knowledge of the auxiliary manifold is also included. (f) Through optimal transport of diverse unsupervised tasks, the learned shared representation only includes rich domain-invariant information from the auxiliary manifold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.64 + ], + "angle": 0, + "content": "While training robust models from noisy data has been intensively investigated in standard settings, learning from noisy few-shot data is only sparsely explored in existing literature [28,31]. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. Within each training task, FSL operates with few training samples per class, making each data point highly valuable. Hence, the model tends to be more sensitive to variations in the training data and noisy labels can more easily cause the model to learn incorrect patterns, compromising its ability to accurately generalize to new tasks. Furthermore, identifying and rectifying label noises also becomes more challenging. The scarcity of data makes it harder to distinguish between true class characteristics and erroneous annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Techniques, including robust optimization [28] and data augmentation [31] have been leveraged to handle noises in limited data. However, these techniques, while beneficial, do not entirely resolve the problem. Conventional methods for learning with noisy labels, such as data cleansing, data augmentations, and robust loss functions, are inadequate to tackle the challenges as outlined above. As shown in Fig. [1(a)-(b)], data cleansing introduces information loss, further diminishing the already scarce training data in each task. This poses a greater challenge for the model to learn a meaningful and generalizable representation of the task manifold. Data augmentation could alleviate the information loss introduced through data cleansing as shown in Fig. [1(c)]. Nevertheless, augmentation is performed on the limited training samples, so it may not cover a broad task manifold. Additionally, robust optimization algorithms disregard the concern of information loss, making them less suitable for learning from noisy few-shot data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.758, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.789, + 0.361 + ], + "angle": 0, + "content": "Fig. 2: In DUNT, the feature extractor extract information from both primary and auxiliary tasks, the classifier predicts the labels of the primary and auxiliary examples, and the domain critic outputs the source of the examples. The primary tasks (top) are sampled and cleansed by episode cleansing, the auxiliary tasks (bottom) are constructed and filtered by their diverseness. (Best viewed in color.)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.562 + ], + "angle": 0, + "content": "To tackle the unique challenges arising from NFSL, we propose novel extensions to existing techniques designed for standard noisy learning scenarios to more effectively handle the information loss in the few-shot setting. In particular, to best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed auxiliary tasks to compensate information loss. The auxiliary tasks are entirely unsupervised as using supervised tasks can be counterproductive in the NFSL setting for three main reasons: (1) They may incur additional annotation overhead. (2) Noisy labels may continue to be introduced that further compromise the manifold structure. (3) Even clean labels can be guaranteed, these labels may not totally align with the primary tasks that can potentially overwhelm that weak learning signals carried by the limited training samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.748 + ], + "angle": 0, + "content": "While unsupervised auxiliary tasks have been leveraged in diverse learning scenarios [19,24,32,36,52], including the few-shot setting, the noisy labels and potential information loss introduced through necessary data cleansing makes NFSL much more vulnerable to the potential noises introduced by the auxiliary data. Since data samples are unlabeled, they are usually randomly sampled into an auxiliary task. When a contrastive loss is used to differentiate unlabeled data samples from the same underlying class, the model may be forced to learn sample-specific (spurious) features that cause overfitting. To this end, we propose to conduct novel auxiliary task selection to ensure the intra-diversity among the chosen unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold as shown in Fig. 1(d)-(g)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Fig. 2 shows the proposed framework, referred to as Diverse UNsupervised Tasks for NFSL (DUNT4NFSL or DUNT for short). The feature extractor of DUNT undergoes updates through three distinct loss components: 1) A cross-entropy loss, integral for extracting discriminative and effective features for classification from the primary tasks; 2) A contrastive loss tailored for the unsupervised auxiliary tasks, designed to encourage an embedding space where similar sample" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "pairs remain close while dissimilar pairs are distanced; 3) An adversarial loss, implemented to align the unsupervised auxiliary tasks with the primary tasks, employing a domain critic head to reduce the covariate shift across multiple domains. To ensure that the feature representation remains unaffected by label noise, we apply an episode cleaning technique to the primary task, removing noisy examples in each task. For the auxiliary tasks, we introduce a novel criterion for diverseness to guarantee that the unsupervised auxiliary tasks contribute informative and valuable insights. Our contributions are threefolds:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.327 + ], + "angle": 0, + "content": "- We present DUNT, a framework designed to adeptly utilize unsupervised data from a distinct domain in order to counteract the information loss resulting from data cleansing in the challenging scenario of NFSL." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.374 + ], + "angle": 0, + "content": "- We introduce a novel strategy to perform diverse auxiliary task selection to avoid learning sample-specific spurious features from the unlabeled auxiliary data samples." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.376, + 0.786, + 0.42 + ], + "angle": 0, + "content": "- We conduct an in-depth theoretical analysis of the auxiliary tasks introduced in the DUNT framework, establishing novel generalization bounds, offering valuable insights into the contributions of these auxiliary tasks." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.482 + ], + "angle": 0, + "content": "We have conducted comprehensive experiments on five benchmark datasets with synthetic label noises to demonstrate the effectiveness and robustness of DUNT and its advantage over SOTA methods by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.51, + 0.388, + 0.525 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Noisy few-shot learning. Existing methods for training DNNs with noisy labels usually fall into the following categories: robust loss [12,49,50,57,60] designs robust cost functions in the presence of noisy labels; noise transition estimation [5,20,53,54] aims to estimate the probabilities that clean labels flip into noisy labels; loss correction [17,22,23,39,44,55,59] either explicitly or implicitly relabels or reweights the noisy samples, a.k.a. data cleansing. Noisy label in FSL is an important yet relatively under-explored topic. The scarcity of data in FSL not only heightens its susceptibility to label noise but also adds complexity to the challenge of identifying and correcting such noise. A few previous attempts such as [28,29,31] address mislabeled issue in FSL settings using re-weighting strategy or generating robust prototypes by feature hybrid and augmentation techniques. Other works such as RW-MAML [25], AQ [14], DFSL [27] tackle the out-of-distribution tasks or adversarial attacks in few-shot learning. In this work, we particularly focus on solving the limited data problem by utilizing auxiliary unlabeled data as unsupervised tasks to enhance the robustness of FSL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Optimal transport and Wasserstein distance. The optimal transport (OT) cost is used to measure the difference between distributions supported on high-dimensional space using SGD [1,46]. For a constant \\(\\xi \\geq 1\\), the \\(\\xi\\)-Wasserstein" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.614, + 0.162 + ], + "angle": 0, + "content": "metric between distributions \\(\\mathbb{P}_X\\) and \\(\\mathbb{P}_{X'}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.174, + 0.786, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {\\xi} \\left(\\mathbb {P} _ {X}, \\mathbb {P} _ {X ^ {\\prime}}\\right) = \\left(\\inf _ {\\gamma \\in \\Pi \\left(\\mathbb {P} _ {X}, \\mathbb {P} _ {X ^ {\\prime}}\\right)} \\mathbb {E} _ {\\gamma \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right)} \\left[ c ^ {\\xi} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right) \\right]\\right) ^ {\\frac {1}{\\xi}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.222, + 0.788, + 0.465 + ], + "angle": 0, + "content": "where \\(\\mathbf{x},\\mathbf{x}^{\\prime}\\in \\mathbb{R}^{d}\\) denote the random variables in which the distributions \\(\\mathbb{P}_X\\) and \\(\\mathbb{P}_{X^{\\prime}}\\) are defined. \\(\\prod (\\mathbb{P}_X,\\mathbb{P}_{X'})\\) is the set of all joint distributions (i.e. couplings) whose marginals are \\(\\mathbb{P}_X\\) and \\(\\mathbb{P}_{X'}\\), that is, \\(\\int \\gamma (\\mathbf{x},\\mathbf{x}')d\\mathbf{x} = \\mathbb{P}_{X'}\\) and \\(\\int \\gamma (\\mathbf{x},\\mathbf{x}')d\\mathbf{x}' = \\mathbb{P}_X\\). \\(c(\\mathbf{x},\\mathbf{x}'): \\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}^+\\) is the cost function for moving from \\(\\mathbf{x}\\) to \\(\\mathbf{x}'\\). Intuitively, \\(\\gamma (\\mathbf{x},\\mathbf{x}')\\) indicates how much \"mass\" must be transported from \\(\\mathbf{x}\\) to \\(\\mathbf{x}'\\) in order to transform the distributions \\(\\mathbb{P}_X\\) into \\(\\mathbb{P}_{X'}\\). The Wasserstein metric then is the \"cost\" of the optimal transport plan [1]. Eq. [1] is the primal form of the Wasserstein metric, particularly for the case of \\(\\xi = 1\\) with \\(c(\\mathbf{x},\\mathbf{x}') = \\| \\mathbf{x} - \\mathbf{x}'\\|\\) [1]. Some existing efforts propose domain adaptation algorithms using ideas from OT theory, which aims to reduce the divergence between two domains by minimizing the Wasserstein distance between their distributions [7,8,11]. Works that use domain adaptation techniques in FSL [6,18,33,40,56] share the same goal that aims to learn transferable knowledge. In this work, we achieve this goal by leveraging extra data first and then minimize the Wasserstein distance between the primary and auxiliary distributions to ensure the learned knowledge are domain-invariant and discriminative to future unseen tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.381, + 0.507 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.521, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Problem formulation. In FSL, the goal is to learn a mapping \\( f: \\mathcal{X} \\to \\mathcal{Y} \\) that maps an observation \\( \\mathbf{x} \\) to its label \\( \\mathbf{y} \\), where \\( \\mathcal{X} \\) is the feature space and \\( \\mathcal{Y} \\) is the label space. To train the model with the quick adaptation ability to the new tasks with only a small number of examples, a set of tasks \\( \\mathcal{T}_i = \\{\\mathcal{S}_i^{\\mathrm{sup}},\\mathcal{S}_i^{\\mathrm{que}}\\} \\) is sampled from the data distribution \\( \\mathcal{D} \\) over \\( \\mathcal{X} \\times \\mathcal{Y} \\), where \\( S_{i}^{\\mathrm{sup}} = \\{(x_{i}^{k},y_{i}^{k})\\}_{k = 1}^{NK} \\) denotes the support set and \\( S_{i}^{\\mathrm{que}} = \\{(x_{i}^{k},y_{i}^{k})\\}_{k = 1}^{NQ} \\) denotes the query set. \\( N \\) is the number of classes, \\( K \\) is the number of examples in each class of the support set, and \\( Q \\) is that of the query set. In DUNT, the meta-training tasks are sampled from the two domains: the primary tasks are sampled from the noisy primary distribution \\( \\tilde{\\mathcal{D}}^{\\mathrm{pri}} \\) over \\( \\mathcal{X} \\times \\tilde{\\mathcal{Y}} \\), where the label space \\( \\tilde{\\mathcal{V}} \\) is considered noisy, and the unsupervised auxiliary tasks are sampled from the auxiliary distribution \\( \\mathcal{D}^{\\mathrm{aux}} \\) over \\( \\mathcal{X}' \\), where the entire label space is missing. We assume that the primary and the auxiliary distributions are from different domains: \\( \\tilde{\\mathcal{D}}^{\\mathrm{pri}} \\neq \\mathcal{D}^{\\mathrm{aux}} \\). During meta-testing, the tasks are sampled from a new unseen target domain \\( \\mathcal{D}^{\\mathrm{tgt}} \\) over \\( \\mathcal{X}^{\\mathrm{tgt}} \\times \\mathcal{Y}^{\\mathrm{tgt}} \\), the label is only used for evaluation purpose. The goal is to learn a meta-model \\( \\theta \\) trained on \\( \\tilde{\\mathcal{D}}^{\\mathrm{pri}} \\) and \\( \\mathcal{D}^{\\mathrm{aux}} \\) that can quickly adapt to tasks in \\( \\mathcal{D}^{\\mathrm{tgt}} \\) using only a few examples." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Framework overview. The DUNT framework (see Fig. 2) is composed of a shared feature extractor \\(\\theta\\), the task classifier \\(\\theta_i^c\\) for task \\(i\\) and the domain critic \\(\\theta^d\\). The feature extractor takes input from both the primary tasks and the auxiliary" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "tasks; the classifiers predict the labels for the input and the critic predict the source of the input. The gradient reversal layer (GRL) ensures that the feature distributions over the two domains are made similar." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.218, + 0.54, + 0.234 + ], + "angle": 0, + "content": "3.1 Learning from the Primary Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.321 + ], + "angle": 0, + "content": "One commonly used FSL strategy that takes advantage of task-based formulation is to learn a meta-model through a bi-level optimization. By defining the tasks-specific model parameter for task \\(\\mathcal{T}_i\\) as \\(\\theta_i'\\), the meta model \\(\\theta\\) are trained by optimizing for the performance of \\(\\theta_i'\\) with respect to \\(\\theta\\) across tasks sampled from \\(\\mathcal{D}^{\\mathrm{pri}}\\). The loss is" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.333, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {p r i}} = \\sum_ {\\mathcal {T} _ {i} \\left(\\mathcal {S} _ {i} ^ {\\sup }, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}\\right) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}} \\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(\\theta_ {i} ^ {\\prime}, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.38, + 0.789, + 0.492 + ], + "angle": 0, + "content": "where \\(\\theta_i' \\gets \\theta - \\alpha \\nabla_\\theta \\mathcal{L}_{\\mathcal{T}_i}(\\theta, \\mathcal{S}_i^{\\mathrm{sup}})\\) is the task-specific model adapted from \\(\\theta\\) within one gradient step. This is a dual-level optimization where in the outer-level, the meta-model \\(\\theta\\) is updated with the summation of a series of tasks evaluated on the query sets \\(\\{\\mathcal{S}_i^{\\mathrm{que}}\\}\\), with \\(\\mathcal{L}_{\\mathcal{T}_i}(\\theta_i', \\mathcal{S}_i^{\\mathrm{que}}) = \\sum_{(\\mathbf{x}_i^k, \\mathbf{y}_i^k) \\in \\mathcal{S}_i^{\\mathrm{que}}} \\ell(\\theta_i^c(\\theta_i'(\\mathbf{x}_i^k)), \\mathbf{y}_i^k)\\). The inner-level objective is computed using the task-specific model \\(\\theta_i'\\) updated with support set \\(\\mathcal{S}_i^{\\mathrm{sup}}\\) using loss \\(\\mathcal{L}_{\\mathcal{T}_i}(\\theta, \\mathcal{S}_i^{\\mathrm{sup}}) = \\sum_{(\\mathbf{x}_i^k, \\mathbf{y}_i^k) \\in \\mathcal{S}_i^{\\mathrm{sup}}} \\ell(\\theta_i^c(\\theta(\\mathbf{x}_i^k)), \\mathbf{y}_i^k)\\), where \\(\\ell\\) is the cross-entropy loss, \\(\\theta_i^c\\) is the \\(i\\)-th task classifier." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.509, + 0.79, + 0.648 + ], + "angle": 0, + "content": "Episode cleansing. In the dual-level optimization, the meta-model \\(\\theta\\) is updated by first adapting to different task-specific models \\(\\theta_{i}^{\\prime}\\) with the corresponding support set and then updated using the summation of the loss of the task-specific models \\(\\theta_{i}^{\\prime}\\) on their corresponding query sets. Since the dataset is noisy, the error may be accumulated through the noisy query set and cause the overfitting of the meta-model. Therefore, we propose episode cleansing to remove the noisy data in the query sets based on the memorization effect which indicates that noisy examples have higher losses than the clean ones. The classification loss \\(\\mathcal{L}_{\\mathcal{T}_i}(\\theta_i', S_i^{\\mathrm{que}})\\) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.659, + 0.786, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(f _ {\\theta_ {i} ^ {\\prime}}, \\hat {\\mathcal {S}} _ {i} ^ {\\mathrm {q u e}}\\right) = \\sum_ {\\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {i} ^ {k} \\in \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}} w _ {i} ^ {k} \\ell_ {i} ^ {k}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.708, + 0.789, + 0.754 + ], + "angle": 0, + "content": "where \\(\\hat{S}_i^{\\mathrm{que}}\\) is the cleansed (weighted) query set for task \\(\\mathcal{T}_i\\); for simplification, we use \\(\\ell_i^k\\) to denote \\(\\ell (\\theta_i^c (\\theta_i'(x_i^k)),y_i^k)\\). \\(w_{i}^{k}\\) is the self-paced sample weight for each example pair, and" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.755, + 0.568, + 0.773 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {k} = \\mathbb {1} (\\ell_ {i} ^ {k} < \\gamma_ {t h}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is an indicator function whose value is 1 when \\(\\ell_i^k < \\gamma_{th}\\) and 0 otherwise; and \\(\\gamma_{th}\\) is a predefined hyperparameter used to filter out high-loss examples. We use the percentage of the ordered loss to avoid extensive hyperparameter search, whose ablation is provided in Appendix C." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.147, + 0.465, + 0.162 + ], + "angle": 0, + "content": "The primary loss the becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.174, + 0.786, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {p r i}} = \\sum_ {\\mathcal {T} _ {i} (\\mathcal {S} _ {i} ^ {\\sup }, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}} \\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(\\theta_ {i} ^ {\\prime}, \\hat {\\mathcal {S}} _ {i} ^ {\\mathrm {q u e}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.219, + 0.785, + 0.25 + ], + "angle": 0, + "content": "Similarly, for the support set label noise, we use meta-model to compute the loss values and remove the noisy ones in the support set." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.271, + 0.618, + 0.286 + ], + "angle": 0, + "content": "3.2 Learning from Diverse Unsupervised Tasks" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.295, + 0.789, + 0.401 + ], + "angle": 0, + "content": "As shown in Fig. 2, the general unsupervised auxiliary task \\(\\mathcal{T}_j\\) is constructed by i) first randomly sampling \\(N\\) images from the dataset for the support set \\(\\mathcal{S}_j^{\\mathrm{sup}} = \\{\\mathbf{x}_j^k\\}_{k=1}^N\\); ii) the query set \\(\\mathcal{S}_j^{\\mathrm{que}} = \\{A_1(\\mathbf{x}_j^k), \\dots, A_Q(\\mathbf{x}_j^k)\\}_{k=1}^N\\) is obtained by augmenting the \\(N\\) images in \\(Q\\) different ways, where \\(A_1(\\cdot), \\dots, A_Q(\\cdot)\\) denote \\(Q\\) different augmentation techniques, such as random cropping, translation, flipping. The auxiliary loss is the summation of the contrastive loss of a series of auxiliary tasks:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.413, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {a u x}} = \\sum_ {\\mathcal {T} _ {j} \\left(\\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right) \\sim \\mathcal {D} ^ {\\text {a u x}}} \\mathcal {L} _ {\\mathcal {T} _ {j}} \\left(\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.458, + 0.634, + 0.472 + ], + "angle": 0, + "content": "The contrastive loss is utilized for the unsupervised tasks:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.485, + 0.786, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {T} _ {j}} \\left(\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right) = - \\frac {1}{N Q} \\sum_ {k = 1} ^ {N} \\sum_ {q = 1} ^ {Q} \\log \\frac {\\exp \\left(- d \\left[ \\theta \\left(A _ {q} \\left(\\mathbf {x} _ {j} ^ {k}\\right)\\right) , \\theta \\left(\\mathbf {x} _ {j} ^ {k}\\right) \\right]\\right)}{\\sum_ {l = 1} ^ {N} \\exp \\left(- d \\left[ \\theta \\left(A _ {q} \\left(\\mathbf {x} _ {j} ^ {k}\\right)\\right) , \\theta \\left(\\mathbf {x} _ {j} ^ {l}\\right) \\right]\\right)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.537, + 0.651, + 0.553 + ], + "angle": 0, + "content": "where \\(d[\\cdot, \\cdot]\\) is a distance metric, such as Euclidean distance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.567, + 0.789, + 0.807 + ], + "angle": 0, + "content": "Auxiliary task selection. In contrastive learning, the positive pairs are formulated by the original example and its augmentation, and the negative pairs are formulated by images from two different classes. The contrastive loss is designed to pull the positive pairs close to each other and push the negative ones away. However, since each image in the unsupervised task are randomly sampled and treated as a class, there are chances that the images are sampled from the same class. In some extreme case, all the images are sampled from the same class, which means that the positive pairs being treated as negative pairs and being push away from each other. This would cause the model being forced to overfit to the unnecessary details and creates spurious correlations and harm the generalization ability. We denote those tasks as confusing tasks, as a comparison, the tasks with images sampled from different classes are denoted as valuable tasks. The intricate nature of confusing tasks appears to pose a significant threat to the resilience and robustness of FSL, as shown in our later ablation studies. In this work, we propose a novel task selection metric to remove these confusing unsupervised tasks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.808, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Given two images \\(\\mathbf{x}_j^k\\), \\(\\mathbf{x}_j^l\\), the model parameterized by \\(\\theta\\) lead to class probabilities vectors \\(\\mathbf{p}_j^k = \\mathrm{softmax}(\\theta (\\mathbf{x}_j^k))\\) and \\(\\mathbf{p}_j^l = \\mathrm{softmax}(\\theta (\\mathbf{x}_j^l))\\). If the two images" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.21 + ], + "angle": 0, + "content": "are from the same class, \\(\\mathbf{p}_j^k\\) and \\(\\mathbf{p}_j^l\\) are naturally encouraged to be close to the assignment vector \\(\\mathbf{e}_y\\) in \\(\\{0,1\\}^N\\) with a single non-zero entry at position \\(y\\), where \\(y\\) is the class label associated to \\(\\mathbf{x}\\) and \\(N\\) is the number of classes. Therefore, we consider the following diversity criterion:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.221, + 0.638, + 0.263 + ], + "angle": 0, + "content": "\\[\nd i v _ {j} = \\frac {1}{C (N , 2)} \\sum_ {i = 1} ^ {N - 1} \\sum_ {j = i + 1} ^ {N} \\frac {1}{\\cos (\\mathbf {p} _ {j} ^ {k} , \\mathbf {p} _ {j} ^ {l})}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.274, + 0.789, + 0.337 + ], + "angle": 0, + "content": "The diverseness of the task \\(\\mathcal{T}_j\\) is defined as the average of the reciprocal of the cosine similarity between the class probabilities of all example pairs in the task, where \\(C(N,2) = \\frac{N!}{(N - 2)!2!}\\) is the 2-combination of set \\(N\\). According to the definition, the higher the value of \\(div_j\\), the more diverse the task." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.351, + 0.632, + 0.367 + ], + "angle": 0, + "content": "Auxiliary loss function. Auxiliary loss \\(\\mathcal{L}^{\\mathrm{aux}}\\) is updated as" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.379, + 0.66, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {a u x}} = \\sum_ {\\mathcal {T} _ {j} (\\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\mathrm {q u e}}) \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} v _ {j} \\mathcal {L} _ {\\mathcal {T} _ {j}} (\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\mathrm {q u e}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.425, + 0.528, + 0.441 + ], + "angle": 0, + "content": "where \\( v_{j} \\) is the diversity weight defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.454, + 0.574, + 0.47 + ], + "angle": 0, + "content": "\\[\nv _ {j} = \\mathbb {1} (d i v _ {j} > \\gamma_ {d i v}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.481, + 0.788, + 0.558 + ], + "angle": 0, + "content": "where \\(\\gamma_{div}\\) is a predefined hyperparameter, which filter out those less diverse, a.k.a., confusing tasks. In practice, we simplify the hyperparameter searching by first upsampling six auxiliary tasks, computes each task diversity metric, and only keep the most diverse task in \\(\\{\\mathcal{T}_j\\}^{\\mathrm{val}}\\) for training. The corresponding ablations about the numbers of extra tasks sampled are presented in Sec. 4.3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.58, + 0.604, + 0.594 + ], + "angle": 0, + "content": "3.3 Adversarial Training and the Critic Head" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.605, + 0.788, + 0.802 + ], + "angle": 0, + "content": "Since the auxiliary tasks are drawn from a different data distribution, it is inevitable that there exist some domain shift from the primary domain to the auxiliary one. While learning from the domain-specific knowledge from the auxiliary domain main cause the learned model biased and deteriorate the model's generalization performance, learning from the domain-invariant knowledge helps improve the generalization ability. To enforce the learning of the domain-invariant representation of the meta-model, we propose a critic model \\(\\theta^c\\) to discriminate the source of the input data, either primary or auxiliary. By training the model in an adversarial way, the critic model \\(\\theta^c\\) try to distinguish which domain that the input features come from, while the meta-model \\(\\theta\\) aims to confuse the critic. In this way, when the training process converge, \\(\\theta\\) tends to extract domain-invariant knowledge from both domains. Formally, the overall adversarial objective is a min-max problem as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.813, + 0.785, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , \\theta^ {c}} \\max _ {\\theta^ {d}} \\alpha_ {1} \\mathcal {L} ^ {\\text {p r i}} + (1 - \\alpha_ {1}) \\mathcal {L} ^ {\\text {a u x}} + \\alpha_ {2} \\mathcal {E} (\\theta , \\theta^ {d}, \\tilde {\\mathcal {D}} ^ {\\text {p r i}}, \\mathcal {D} ^ {\\text {a u x}}), \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "where the critic model \\(\\theta^d\\) assigns 1 to the example from the primary domain and 0 otherwise based on the feature representation extracted by \\(\\theta\\). \\(\\mathcal{E}(\\cdot)\\) is the adversarial loss" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.203, + 0.69, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} (\\theta , \\theta^ {d}, \\hat {\\mathcal {D}} ^ {\\mathrm {p r i}}, \\mathcal {D} ^ {\\mathrm {a u x}}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\hat {\\mathcal {D}} ^ {\\mathrm {p r i}}} [ g (\\mathbf {x}) ] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} [ g (\\mathbf {x}) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.233, + 0.627, + 0.25 + ], + "angle": 0, + "content": "where \\(g(\\mathbf{x}) = \\theta^d (\\theta (\\mathbf{x}))\\), with its empirical counterpart as" + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.259, + 0.724, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {E}} (g _ {\\theta , \\theta^ {d}}, \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}, \\mathcal {D} ^ {\\mathrm {a u x}}) = \\frac {1}{| \\mathcal {D} ^ {\\mathrm {p r i}} |} \\sum_ {x \\sim \\mathcal {D} ^ {\\mathrm {p r i}}} [ g (\\mathbf {x}) ] - \\frac {1}{| \\mathcal {D} ^ {\\mathrm {a u x}} |} \\sum_ {\\mathbf {x} \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} [ g (\\mathbf {x}) ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.307, + 0.789, + 0.4 + ], + "angle": 0, + "content": "The optimization of Eq. (7) is conducted through an alternative training between the optimization of the meta-model feature extractor \\(\\theta\\), classifier parameter \\(\\theta^c\\) and the critic model parameter \\(\\theta^d\\) while keeping the other fixed. Parameters of \\(\\theta\\), \\(\\theta^c\\) and \\(\\theta^d\\) are updated alternatively. In practice, \\(\\theta\\) and \\(\\theta^c\\) are connected by the GRL, which reverses the gradients that flow through \\(\\theta^c\\). The pseudo code of the algorithm can be found in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.432, + 0.436 + ], + "angle": 0, + "content": "3.4 Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.447, + 0.789, + 0.538 + ], + "angle": 0, + "content": "We conduct a theoretical analysis of DUNT to show: 1) the equivalence between the adversarial regularizer and the Wasserstein distance, which captures the diversity property of the auxiliary distribution while enforcing the learning of a domain invariant representation; 2) the relationship between the primary and auxiliary distributions; 3) a tight generalization bound of the target based on the expected risk of the primary and auxiliary distributions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.561, + 0.804, + 0.594 + ], + "angle": 0, + "content": "Equivalence to Wasserstein distance. The Kantorovich-Rubinstein of Wasserstein distance takes the form [1]:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.605, + 0.785, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} (\\mathbb {P} _ {1}, \\mathbb {P} _ {2}) = \\sup _ {\\| h \\| _ {L} \\leq 1} \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {1}} [ h (\\mathbf {x}) ] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {2}} [ h (\\mathbf {x}) ] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.641, + 0.788, + 0.687 + ], + "angle": 0, + "content": "where the supremum is over all the 1-Lipschitz functions \\( h: \\mathcal{X} \\to \\mathbb{R} \\). Therefore, if we have a parameterized family of functions \\( \\{h_w\\}_{w \\in \\mathcal{W}} \\) that are all 1-Lipschitz, we could consider solving the problem" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.7, + 0.785, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\max \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {1}} \\left[ h _ {w} (\\mathbf {x}) \\right] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {2}} \\left[ h _ {w} (\\mathbf {x}) \\right] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.727, + 0.788, + 0.79 + ], + "angle": 0, + "content": "which has the same form of the regularizer in Sec. 3.3. This indicates that our adversarial regularizer has the same theoretical foundation of Wasserstein distance optimal transport. To satisfy the 1-Lipschitz constraint, the critic \\(\\theta^d\\) is implemented with gradient penalty [15]. The overall objective can be reformulated as" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.814, + 0.785, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\alpha_ {1} \\mathcal {L} ^ {\\text {p r i}} + (1 - \\alpha_ {1}) \\mathcal {L} ^ {\\text {a u x}} + \\alpha_ {2} \\mathcal {W} _ {1} (\\mathcal {D} ^ {\\text {p r i}}, \\mathcal {D} ^ {\\text {a u x}}), \\tag {10}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.24 + ], + "angle": 0, + "content": "where the first and second term are the primary and auxiliary loss, resp., and the third term is the Wasserstein distance based regularizer, which aims to extract domain-invariant feature from the primary and auxiliary domains. Wasserstein distance reflects an optimal transport cost for moving one distribution to another. A smaller transport cost means a better coverage of the distribution, which captures the property of diversity [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Relationship of the primary and auxiliary distributions. By viewing this problem as a distribution matching that matches the auxiliary distribution with the primary one, we have the generalization property that quantify the relationship between the primary and the auxiliary distribution. Given the primary distribution \\(\\mathcal{D}^{\\mathrm{pri}}\\) and the auxiliary distribution \\(\\mathcal{D}^{\\mathrm{aux}}\\), the expected risk w.r.t. is defined as \\(\\mathcal{D}^{\\mathrm{pri}}R_{\\mathcal{D}^{\\mathrm{pri}}} = \\mathbb{E}_{x\\sim \\mathcal{D}^{\\mathrm{pri}}}\\ell (h(x),h^{*}(x))\\), and the expected risk w.r.t. \\(\\mathcal{D}^{\\mathrm{aux}}\\) is defined as \\(R_{\\mathcal{D}^{\\mathrm{aux}}} = \\mathbb{E}_{x\\sim \\mathcal{D}^{\\mathrm{aux}}}\\ell (h(x),h^{*}(x))\\). We have the following Lemma 1:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Lemma 1. If the loss \\(\\ell\\) is symmetric and \\(L\\)-Lipschitz, \\(\\forall h \\in \\mathcal{H}\\) is at most \\(H\\)-Lipschitz function and the underlying mapping function \\(h^*\\) is \\(\\phi(\\lambda) - (\\mathcal{D}^{pri}, \\mathcal{D}^{aux})\\) joint probabilistic Lipschitz; then the expected risk w.r.t. \\(\\mathcal{D}^{pri}\\) is upper bounded" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.429, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathcal {D} ^ {p r i}} \\leq R _ {\\mathcal {D} ^ {a u x}} + L (H + \\lambda) \\mathcal {W} _ {1} (\\mathcal {D} ^ {p r i}, \\mathcal {D} ^ {a u x}) + L \\phi (\\lambda), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.456, + 0.788, + 0.504 + ], + "angle": 0, + "content": "where \\(\\phi : \\mathbb{R} \\to [0,1]\\) and a mapping \\(h^*\\) is \\(\\phi(\\lambda) - (\\mathcal{D}^{pri},\\mathcal{D}^{aux})\\) joint probabilistic Lipschitz if \\(\\text{supp}(\\mathcal{D}^{aux}) \\subseteq \\text{supp}(\\mathcal{D}^{pri})\\); \\(\\forall \\lambda > 0\\), all coupling \\(\\gamma \\in \\prod(\\mathcal{D}^{pri},\\mathcal{D}^{aux}) : \\mathbb{P}_{(x_{\\mathcal{D}^{pri}},x_{\\mathcal{D}^{aux}})\\sim \\gamma}[|h^{*}(x_{\\mathcal{D}^{pri}}) - h^{*}(x_{\\mathcal{D}^{aux}})| > \\lambda \\| x_{\\mathcal{D}^{pri}} - x_{\\mathcal{D}^{aux}}\\|_2] \\leq \\phi(\\lambda)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Lemma 1 shows that the expected risk of the primary distributions is upper bounded by the expected risk w.r.t. the auxiliary distribution, the Wasserstein distance, and the labeling function property. The proof of Lemma 1 can be derived with Theorem 1 in [42]. Eq. (10) follows the upper bound from Lemma 1, which means a desirable auxiliary distribution with a better matching with the primary distribution which holds a small expected risk leads to tighter generalization bound for the primary distribution. This further verifies the rationality of our design of using the WD based regularizer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.787, + 0.684 + ], + "angle": 0, + "content": "Generalization bound of DUNT. Furthermore, in the following Lemma 2 we derive the upper bound for the target risk \\( R_{\\mathcal{D}^{\\mathrm{tgt}}} = \\mathbb{E}_{(\\mathbf{x},y)\\sim \\mathcal{D}^{\\mathrm{tgt}}}\\ell (\\mathbf{x},y) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.739 + ], + "angle": 0, + "content": "Lemma 2. Let \\(\\mathcal{D}^{tgt}\\) be the target distribution, \\(h\\) is \\(K\\)-Lipschitz w.r.t. the feature \\(\\mathbf{x}\\), and the setting of loss and hypothesis be the same as Lemma 1, the expected risk w.r.t. the target distribution is upper-bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.748, + 0.785, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R _ {\\mathcal {D} ^ {t g t}} \\leq \\alpha_ {1} \\left(R _ {\\mathcal {D} ^ {p r i}} + L K \\mathcal {W} _ {1} \\left(\\mathcal {D} ^ {p r i}, \\mathcal {D} ^ {t g t}\\right)\\right) \\tag {12} \\\\ + (1 - \\alpha_ {1}) \\left(R _ {\\mathcal {D} ^ {a u x}} + L K \\mathcal {W} _ {1} \\left(\\mathcal {D} ^ {a u x}, \\mathcal {D} ^ {t g t}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Lemma 2 shows that the expected risk w.r.t. the target distribution is upper bounded by the expected risk w.r.t. the primary distribution and the auxiliary distribution, and the corresponding Wasserstein distances. The proof of Lemma 2" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "can be derived with Theorem 2 in [41]. Lemma 2 provide solid theoretical foundation for the design of DUNT with a tight generalization bound to the unseen target domain with primary and auxiliary distributions, which can be extended to multiple sources." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.232, + 0.376, + 0.249 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.265, + 0.651, + 0.28 + ], + "angle": 0, + "content": "4.1 Datasets, Baselines, and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.29, + 0.787, + 0.352 + ], + "angle": 0, + "content": "Datasets. We conduct experiments using the following benchmark datasets for FSL: CUB 51, miniImageNet (mini) 48, tieredImageNet (tiered) 38, CIFAR-FS (CF) 3, and Cross-Domain dataset 16. Each datasets contains synthetic labels noise with different ratios. More details are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Baselines. We use ProtoNet [43] as the base few-shot learning method and consider comparing with improving methods with general data cleansing methods including SPL [26] and FSR [61], denoted as \\(PN + SPL\\) and \\(PN + FSR\\), respectively. We also consider to improve ProtoNet in transductive settings including CoTeaching [17], CoTeaching+ [59], denoted as \\(PN + Co\\), and \\(PN + Co+\\), respectively, STARTUP [35], and DDNet [21]. Last but not least, we consider to compare with noisy FSL methods RNNP [31], TraNFS [28] and DCML [37]. The detailed descriptions and settings are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Experimental settings. The feature extractor follows the same architecture as the ProtoNet used by [43], which has 4 modules with a \\(3 \\times 3\\) convolutions and 64 filters, followed by batch normalization, a ReLU nonlinearity, and \\(2 \\times 2\\) max-pooling, a.k.a., Conv4. The architecture of classifier have the design with a fully connected layer followed by a ReLU nonlinearity, a dropout layer, and another fully connected layer. The critic architecture has the same architecture with an extra sigmoid function. We've considered transductive and non-transductive in the settings. For the former, unlabeled test data of the same dataset is used as auxiliary data during training. This is applied to our method and other transductive methods including \\(\\mathrm{PN} + \\mathrm{Co}\\), \\(\\mathrm{PN} + \\mathrm{Co}+\\), STARTUP, and DDNet. For the latter, we utilize unlabeled datasets from other different datasets as auxiliary. This setting is applied to all the mentioned methods. During training, we use a learning rate \\(=0.001\\), and use Adam as the optimizer. The batch size is 4. The meta-model is trained with 40000 randomly selected primary tasks and 40000 randomly selected auxiliary tasks. During evaluation, the average accuracy is computed based on 600 randomly sampled tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.491, + 0.784 + ], + "angle": 0, + "content": "4.2 Main Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Following the experiment design of [48], we report the average accuracy of 5 runs of each 5-way 1-shot and 5-way 5-shots experiments in Tab. [1] to Tab. [2]. We use \\(\\alpha_{1} = 0.5\\), \\(\\alpha_{2} = 0.02\\), \\(\\gamma_{th} = 50\\%\\), \\(\\gamma_{div}\\) as one out of six tasks for our methods" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "table_caption", + "bbox": [ + 0.348, + 0.145, + 0.656, + 0.16 + ], + "angle": 0, + "content": "Table 1: Non-transductive auxiliary datasets" + }, + { + "type": "table", + "bbox": [ + 0.321, + 0.172, + 0.688, + 0.341 + ], + "angle": 0, + "content": "
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+SPL48.55 ± 1.5143.44 ± 1.4937.06 ± 0.6342.11 ± 0.7155.20 ± 0.69
PN+FSR37.14 ± 0.6044.81 ± 0.5939.82 ± 0.6830.65 ± 0.4937.39 ± 0.63
RNNP55.39 ± 0.4950.18 ± 0.6241.75 ± 0.7243.76 ± 0.4860.15 ± 0.72
TraNFS53.02 ± 0.5246.66 ± 0.6338.25 ± 0.7940.32 ± 0.8860.86 ± 0.63
DCML54.39 ± 0.8749.33 ± 1.2241.77 ± 0.8048.33 ± 0.8060.28 ± 0.82
DUNT56.94 ± 1.0352.22 ± 1.1043.01 ± 0.5149.12 ± 0.6862.80 ± 0.49
Task5-way 5-shot
PN+SPL68.47 ± 0.7458.67 ± 0.8654.94 ± 1.0359.62 ± 1.3170.93 ± 0.83
PN+FSR51.48 ± 0.9161.89 ± 0.8057.38 ± 1.1050.40 ± 1.2144.68 ± 1.09
RNNP70.37 ± 1.3363.57 ± 1.2858.79 ± 2.3655.07 ± 1.2174.31 ± 1.36
TraNFS68.77 ± 1.1761.06 ± 1.1351.24 ± 1.1648.74 ± 1.2074.15 ± 0.83
DCML71.28 ± 1.2362.77 ± 1.0057.77 ± 1.3361.19 ± 1.0172.33 ± 0.90
DUNT72.17 ± 1.1165.10 ± 1.2358.50 ± 1.1162.01 ± 1.0974.44 ± 1.42
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.434 + ], + "angle": 0, + "content": "in all datasets. In Tab. 1 and Tab. 2 we have provided the test accuracy all indomain datasets and one cross-domain dataset: mini-CUB for both transductive and non-transductive settings. The performance of cross-domain dataset 16 is reported in Tab. 3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Non-transductive results. For Tab. DUNT uses miniimagenet as auxiliary data for CUB datasets, uses tieredimagnet as auxiliary data for the cross-domain mini-CUB, and uses CUB as auxiliary data for the rest datasets. The results show that, with extra auxiliary datasets, DUNT has outperformed the general data-cleansing techniques and other noisy FSL methods in general. In Tab. the performance of SPL and FSR are generally worse than that of TraNFS and RNNP, which confirms our assumption that by removing potential noisy data, the data cleansing technique might harm the performance of FSL due to the lack of information, which can be compensated with our auxiliary data design. And compared to TraNFS and RNNP, DUNT has reduced the negative impact of the noisy labels more effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Transductive results. In Tab. 2, we consider the transductive setting where the auxiliary dataset for all the methods are unlabeled test data of the same dataset. The results show that our method ranks the top among all the transductive methods. Compared to Tab. 1, some of the transductive methods underperform non-transductive ones due to the following potential reasons: i) the test dataset is smaller than other auxiliary datasets. Therefore, the non-transductive dataset has more information to provide. Take miniImageNet as an example, the test data of the miniImageNet dataset only contains 1000 images and 10 classes, while the auxiliary dataset CUB contains 200 categories and 12000 images, the abundance and diversity of the auxiliary helps greatly improve the model's performance over the primary tasks more than the close test data does; and ii) the auxiliary dataset provides more valuable fine-grained tasks that benefits representation learning. For instance, when training miniImageNet with CUB as an auxiliary, both datasets contain natural images which are closely related, and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.362, + 0.145, + 0.642, + 0.159 + ], + "angle": 0, + "content": "Table 2: Transductive auxiliary datasets" + }, + { + "type": "table", + "bbox": [ + 0.32, + 0.172, + 0.688, + 0.317 + ], + "angle": 0, + "content": "
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+Co49.10 ± 1.2144.44 ± 0.9636.48 ± 0.9840.36 ± 0.8152.51 ± 0.92
PN+Co+49.45 ± 1.1945.08 ± 0.9139.54 ± 1.0940.19 ± 1.1051.27 ± 1.11
STARTUP52.02 ± 1.2048.79 ± 1.1840.08 ± 1.2244.48 ± 0.9260.69 ± 1.31
DDNet52.55 ± 1.3950.82 ± 0.8741.11 ± 1.8945.11 ± 1.3360.25 ± 1.46
DUNT55.01 ± 1.1252.38 ± 0.8542.07 ± 1.1049.29 ± 1.2661.90 ± 1.14
Task5-way 5-shot
PN+Co65.42 ± 1.2259.57 ± 1.2056.90 ± 1.2955.10 ± 1.2166.89 ± 1.26
PN+Co+67.92 ± 1.1160.43 ± 1.1354.29 ± 1.1054.54 ± 0.8467.22 ± 0.71
STARTUP68.74 ± 1.2162.00 ± 1.2451.07 ± 1.0155.92 ± 0.9974.04 ± 0.84
DDNet69.21 ± 1.3463.70 ± 2.0153.87 ± 0.4457.22 ± 0.7573.92 ± 1.20
DUNT71.20 ± 1.1266.46 ± 0.9859.05 ± 1.0161.09 ± 0.9076.11 ± 0.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.39, + 0.33, + 0.613, + 0.342 + ], + "angle": 0, + "content": "Table 3: Cross-domain Datasets" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.356, + 0.789, + 0.513 + ], + "angle": 0, + "content": "
MethodsEuroSATCropDiseaseChestXISICEuroSATCropDiseaseChestXISIC
5-way 1-shot5-way 5-shot
ProtoNet52.52±0.5256.32±0.8928.90±0.4330.11±1.0061.60±0.7078.89±0.8333.04±0.7136.95±0.90
PN+SPL52.89±1.2158.95±2.0427.68±0.5231.79±1.0967.60±1.2380.25±1.1931.13±1.1837.05±1.15
PN+FSR47.53±1.2156.97±1.2328.40±0.6133.58±0.9261.96±0.9075.93±0.8631.43±1.1936.42±1.22
RNNP51.86±1.2056.56±1.2127.39±0.6131.26±0.8067.97±1.3478.43±1.2931.16±1.2537.84±1.19
TraNFS50.61±1.1058.39±0.9128.26±0.9832.84±1.1167.93±1.1181.51±1.2332.16±1.1636.95±1.18
DCML52.33±1.0258.36±0.4627.54±0.6931.96±0.8068.01±0.9079.11±0.9831.64±1.0035.44±1.10
PN+Co56.56±1.0859.13±1.1426.67±0.7329.16±1.1372.08±1.1578.59±2.1130.45±1.1734.74±1.09
PN+Co+52.18±1.1158.18±1.1028.23±0.8730.63±1.0665.57±1.1078.47±0.9030.77±0.8233.68±0.91
STARTUP52.74±1.2259.43±1.2327.75±0.9532.84±1.1369.78±1.1780.12±1.2231.69±1.3039.58±0.99
DDNet55.21±1.8360.09±1.0928.66±0.9834.10±1.0170.09±1.9081.90±2.0932.01±1.8939.44±1.13
DUNT57.10±1.2961.49±1.0129.78±0.9035.80±0.7772.81±1.1083.20±1.0833.94±1.2040.31±0.90
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.789, + 0.603 + ], + "angle": 0, + "content": "CUB dataset is a collection of bird species, which provide fine-grained tasks for learning a better feature representation, while the miniImageNet test data only provide coarse-grained task, which makes it less useful in some extend. As for DUNT, the performance stays stable regardless the two settings, showing that our framework has the potential to fully utilize any extra information provided." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.796, + 0.766 + ], + "angle": 0, + "content": "Cross-domain results. In Tab. 3, we present the performance on the cross-domain datasets. The training procedure follows the setting in [16], which uses miniImageNet as base dataset for all the methods. For CoTeaching, CoTeaching+, STARTUP, DDNet and DUNT, we use the unlabeled test data EuroSAT, CropDisease, ChestX, and ISIC as auxiliary dataset, respectively. In cross-domain datasets, the transductive methods generally outperform the non-transductive ones due to the utilization of the unlabeled test data that reduces the domain gap between the source domain and the target domain, which is critical when the they are extremely different. DUNT has used the Wasserstein regularization to further reduce the differences. Hence, it ranks the highest in the cross-domain datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.39, + 0.803 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.843 + ], + "angle": 0, + "content": "In this section, we study the effectiveness of different components and impact of hyperparameters. For more ablations, please refer to Appendix C." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.499, + 0.297 + ], + "angle": 0, + "content": "Effectiveness of different components. Tab. 4 shows the ablation study with different elements of DUNT on CUB and miniImageNet datasets under nontransductive setting. We first show the performance of only using the primary data, which is vanilla ProtoNet (Primary). It can be observed that each of these components have a significant impact on the overall performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.148, + 0.787, + 0.162 + ], + "angle": 0, + "content": "Table 4: Impact of different components" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.163, + 0.778, + 0.287 + ], + "angle": 0, + "content": "
PrimaryAuxiliaryRegECTSCUBmini
✓XXXX42.57 ± 0.3439.95 ± 1.06
✓XX✓X48.55 ± 1.5143.44 ± 1.49
✓✓XXX50.19 ± 0.7148.90 ± 0.77
✓✓X✓X53.55 ± 0.9750.02 ± 0.81
✓✓XX✓52.70 ± 1.0149.50 ± 1.66
✓✓X✓✓54.31 ± 0.9950.18 ± 0.87
✓✓✓✓X51.90 ± 0.8949.18 ± 0.76
✓✓✓✓X55.47 ± 0.7950.73 ± 0.60
✓✓✓X✓54.79 ± 1.1049.91 ± 1.00
✓✓✓✓✓56.94 ± 1.0352.22 ± 1.10
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Auxiliary data (Auxiliary) and episode cleansing (EC) serve pivotal functions, with the former introducing additional data and the latter eliminating noisy data. The regularization term (Reg) aids in extracting domain-invariant features, contributing to further performance improvement. Task selection (TS) mitigates the risk of overfitting, thereby enhancing overall performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.432, + 0.521 + ], + "angle": 0, + "content": "Effectiveness of task selection. We study the impact of task selection through the threshold \\(\\gamma_{div}\\). As mentioned earlier, to streamline the hyperparameter search process, we opt for an approach that involves upsampling \\(M\\) unsupervised tasks. Subsequently," + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.385, + 0.605, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.459, + 0.537, + 0.471 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.384, + 0.771, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.459, + 0.705, + 0.471 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.482, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Fig. 3: DUNT on CUB with different \\(\\gamma_{div}\\) (a); DUNT with different task selection (b)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.673 + ], + "angle": 0, + "content": "we select only the most diverse task to be included in the valuable task group, instead of specifying a real value for the diverseness threshold. In this ablation, the number of upsampled tasks varies from 2-10. As depicted in Fig. 3 (a), the performance improves alongside an increase in the parameter \\( M \\), but this comes at the cost of heightened computational demands. Consequently, considering efficiency, we opt for \\( \\gamma_{\\mathrm{div}} = 6 \\) for all experiments. In Fig. 3 (b), we test the effect of using different unsupervised task: select confusing tasks and select valuable tasks. It is evident from the results that the inclusion of confusing tasks significantly deteriorates performance, whereas the utilization of valuable tasks leads to a substantial improvement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.36, + 0.708 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We propose a novel FSL framework with auxiliary tasks that utilize carefully selected unlabeled data under noisy settings. To better align the auxiliary distribution with the primary one, we propose a regularization term based on the Wasserstein distance for learning a domain-invariant representation. To enhance robustness against label noise, we introduce episode cleansing for examples in the primary task and adopt a diverse task selection strategy for the unsupervised auxiliary tasks. Our framework is theoretically and experimentally proved to be effective and beneficial." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Bertinetto, L., Henriques, J.F., Torr, P.H., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: Proceedings of the International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "4. Beyer, L., Henaff, O.J., Kolesnikov, A., Zhai, X., Oord, A.v.d.: Are we done with imagenet? arXiv preprint arXiv:2006.07159 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.424 + ], + "angle": 0, + "content": "6. Chobola, T., Vašata, D., Kordík, P.: Transfer learning based few-shot classification using optimal transport mapping from preprocessed latent space of backbone neural network. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. pp. 29-37. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.467 + ], + "angle": 0, + "content": "7. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Advances in Neural Information Processing Systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.509 + ], + "angle": 0, + "content": "8. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 274-289. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Deleu, T., Würfl, T., Samiei, M., Cohen, J.P., Bengio, Y.: Torchmeta: A Meta-Learning library for PyTorch (2019), https://arxiv.org/abs/1909.06576, available at: https://github.com/tristandeleu/pytorch-meta" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.592 + ], + "angle": 0, + "content": "10. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. pp. 1126-1135 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.619 + ], + "angle": 0, + "content": "1. Flamary, R., Courty, N., Tuia, D., Rakotomamonjy, A.: Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell 1 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "2. Ghosh, A., Kumar, H., Sastry, P.S.: Robust loss functions under label noise for deep neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "3. Gidaris, S., Bursuc, A., Komodakis, N., Pérez, P., Cord, M.: Boosting few-shot visual learning with self-supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8059-8068 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "4. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "5. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Guo, Y., Codella, N.C., Karlinsky, L., Codella, J.V., Smith, J.R., Saenko, K., Rosing, T., Feris, R.: A broader study of cross-domain few-shot learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 124-141. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "17. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Proceedings of the Conference on Neural Information Processing Systems (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.259 + ], + "angle": 0, + "content": "18. Han, C., Fan, Z., Zhang, D., Qiu, M., Gao, M., Zhou, A.: Meta-learning adversarial domain adaptation network for few-shot text classification. In: Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021. Findings of ACL, vol. ACL/IJCNLP 2021, pp. 1664-1673. Association for Computational Linguistics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "20. Hendrycks, D., Mazeika, M., Wilson, D., Gimpel, K.: Using trusted data to train deep networks on labels corrupted by severe noise. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Feris, R., Radke, R.J.: Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems 34, 3584-3595 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "22. Jiang, L., Meng, D., Mitamura, T., Hauptmann, A.G.: Easy samples first: Self-paced reranking for zero-example multimedia search. In: Proceedings of the 22nd ACM international conference on Multimedia. pp. 547-556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.787, + 0.467 + ], + "angle": 0, + "content": "23. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.495 + ], + "angle": 0, + "content": "24. Khodadadeh, S., Boloni, L., Shah, M.: Unsupervised meta-learning for few-shot image classification. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.496, + 0.787, + 0.522 + ], + "angle": 0, + "content": "25. Killamsetty, K., Li, C., Zhao, C., Iyer, R., Chen, F.: A reweighted meta learning framework for robust few shot learning. arXiv preprint arXiv:2011.06782 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "26. Kumar, M.P., Packer, B., Koller, D.: Self-paced learning for latent variable models. In: Proceedings of the Conference on Neural Information Processing Systems. vol. 1, p. 2 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "27. Li, W., Wang, L., Zhang, X., Qi, L., Huo, J., Gao, Y., Luo, J.: Defensive few-shot learning. IEEE transactions on pattern analysis and machine intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "28. Liang, K.J., Rangrej, S.B., Petrovic, V., Hassner, T.: Few-shot learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9089-9098 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.674 + ], + "angle": 0, + "content": "29. Lu, J., Jin, S., Liang, J., Zhang, C.: Robust few-shot learning for user-provided data. IEEE transactions on neural networks and learning systems 32(4), 1433-1447 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "30. Masi, I., Trn, A.T., Hassner, T., Sahin, G., Medioni, G.: Face-specific data augmentation for unconstrained face recognition. International Journal of Computer Vision 127, 642-667 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "31. Mazumder, P., Singh, P., Namboodiri, V.P.: Rnnp: A robust few-shot learning approach. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2664-2673 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.785 + ], + "angle": 0, + "content": "32. Medina, C., Devos, A., Grossglauer, M.: Self-Supervised Prototypical Transfer Learning for Few-Shot Classification. arXiv preprint arXiv:2006.11325 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "33. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Northcutt, C.G., Athalye, A., Mueller, J.: Pervasive label errors in test sets destabilize machine learning benchmarks. arXiv preprint arXiv:2103.14749 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.437, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Robust Learning from Noisy Few-Shot Data" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Phoo, C.P., Hariharan, B.: Self-training for few-shot transfer across extreme task differences. In: Proceedings of the International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "36. Qin, T., Li, W., Shi, Y., Gao, Y.: Diversity helps: Unsupervised few-shot learning via distribution shift-based data augmentation. arXiv preprint arXiv:2004.05805 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.274 + ], + "angle": 0, + "content": "37. Que, X., Yu, Q.: Dual-level curriculum meta-learning for noisy few-shot learning tasks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 14740-14748 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.331 + ], + "angle": 0, + "content": "38. Ren, M., Triantafillou, E., Ravi, S., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: Proceedings of the International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.373 + ], + "angle": 0, + "content": "39. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: International Conference on Machine Learning. pp. 4334-4343. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.402 + ], + "angle": 0, + "content": "40. Sahoo, D., Le, H., Liu, C., Hoi, S.C.: Meta-learning with domain adaptation for few-shot learning under domain shift (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.444 + ], + "angle": 0, + "content": "41. Shui, C., Li, Z., Li, J., Gagné, C., Ling, C.X., Wang, B.: Aggregating from multiple target-shifted sources. In: International Conference on Machine Learning. pp. 9638-9648. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "42. Shui, C., Zhou, F., Gagné, C., Wang, B.: Deep active learning: Unified and principled method for query and training. In: International Conference on Artificial Intelligence and Statistics. pp. 1308-1318. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.515 + ], + "angle": 0, + "content": "43. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 4077-4087 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.571 + ], + "angle": 0, + "content": "44. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: Advances in neural information processing systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.613 + ], + "angle": 0, + "content": "45. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.642 + ], + "angle": 0, + "content": "46. Tolstikhin, I., Bousquet, O., Gelly, S., Schoelkopf, B.: Wasserstein auto-encoders. In: International Conference on Machine Learning (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.685 + ], + "angle": 0, + "content": "47. Tsipras, D., Santurkar, S., Engstrom, L., Ilyas, A., Madry, A.: From imagenet to image classification: Contextualizing progress on benchmarks. In: International Conference on Machine Learning. pp. 9625-9635. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.787, + 0.727 + ], + "angle": 0, + "content": "48. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. In: Advances in neural information processing systems. pp. 3630-3638 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.769 + ], + "angle": 0, + "content": "49. Wang, X., Hua, Y., Kodirov, E., Robertson, N.M.: Imae for noise-robust learning: Mean absolute error does not treat examples equally and gradient magnitude's variance matters. arXiv preprint arXiv:1903.12141 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "50. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "51. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-ucsd birds 200 (2010)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Que and Yu" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "52. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "53. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems 33, 7597-7610 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "54. Xia, X., Liu, T., Wang, N., Han, B., Gong, C., Niu, G., Sugiyama, M.: Are anchor points really indispensable in label-noise learning? Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.788, + 0.355 + ], + "angle": 0, + "content": "56. Xu, R., Yang, X., Liu, B., Zhang, K., Liu, W.: Selecting task with optimal transport self-supervised learning for few-shot classification. arXiv preprint arXiv:2204.00289 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "57. Xu, Y., Cao, P., Kong, Y., Wang, Y.: L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.788, + 0.424 + ], + "angle": 0, + "content": "58. Yang, Y., Liang, K.J., Carin, L.: Object detection as a positive-unlabeled problem. arXiv preprint arXiv:2002.04672 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.788, + 0.466 + ], + "angle": 0, + "content": "59. Yu, X., Han, B., Yao, J., Niu, G., Tsang, I., Sugiyama, M.: How does disagreement help generalization against label corruption? In: International Conference on Machine Learning. pp. 7164-7173. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.788, + 0.508 + ], + "angle": 0, + "content": "60. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "61. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.55 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_origin.pdf b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..53d2ae8a412ceafaee031063057532b551162bf1 --- /dev/null +++ b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/d819cfb1-7fec-4b5f-9ee3-f5114805f9f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c5f2fb4118713abb321e91bd6c7a94762f32b836fbf25a58992057ff6d88e38 +size 2901674 diff --git a/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/full.md b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..92c47f26350c35cb72474a2c8c83e9565517bbd8 --- /dev/null +++ b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/full.md @@ -0,0 +1,348 @@ +# Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data + +Xiaofan Que1 and Qi Yu1 + +Rochester Institute of Technology + +{xq5054,qi.yu}@rit.edu + +Abstract. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. While data cleansing offers a viable solution to address noisy labels in the general learning settings, it exacerbates information loss in FSL due to limited training data, resulting in inadequate model training. To best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed unsupervised auxiliary tasks to compensate for information loss. Using unsupervised tasks can effectively avoid additional annotation costs and minimize the risk of introducing additional label noises. However, a randomly constructed unsupervised task may misguide the model to learn sample-specific features that are likely to compromise the primary few-shot learning task due to the noisy weak learning signals. We propose to conduct novel auxiliary task selection to ensure the intra-diversity among the unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold. We conduct a theoretical analysis to derive novel generalization bounds for learning with auxiliary tasks. Extensive experiments are conducted to demonstrate that our method outperforms existing noisy few-shot learning methods under various in-domain and cross-domain few-shot classification benchmarks. + +Keywords: Noisy few-shot learning $\cdot$ Auxiliary task $\cdot$ Optimal transport + +# 1 Introduction + +Few-shot learning (FSL) aims to utilize knowledge acquired from a large amount of diverse tasks during training and apply it to new, unseen tasks from the same task distribution with only a small amount of labeled data [10,43,45]. Therefore, learning an accurate representation of the task manifold that effectively captures the underlying structure and relationships among tasks is essential to ensure good few-shot generalization. On the other hand, noisy labels commonly exist in the task distribution due to the ever-increasing demand for large-scale annotated datasets, where labeling errors can be introduced in various means that are impossible to control [4,30,34,47,58]. Given the few-shot nature of each training task, memorization of noisy labels introduced along with limited training data can exacerbate overfitting, which will disrupt the representation space and severely hinder few-shot generalization. + +![](images/2614a906af7f80077a56aeb94198149d1d9514148a3960dc8af6f6619e283466.jpg) + +![](images/39eda97b665f88caefd54193ed4dad36fb05f025e587095472ea9c90e76db570.jpg) + +![](images/0c4d380a4768522a57180e6d5788fc8d265b96db48cd0eb49eb30324fba0e2ec.jpg) + +![](images/c26f4556282ddb153d152d93bde461d3b6abdcbf33c4926b7acdf643180d2c29.jpg) +Data with different classes + +![](images/5b44827c96832d1e1ad99fb930421e6ee5afbe6aee990458d1e9225d5f62a322.jpg) +Fig. 1: (a) Data manifold of the primary task distribution. (b) Data manifold after using data cleansing to remove some noisy labels, which is incomplete due to information loss. (c) Using augmentation can make up for the information loss but it is limited by the diversity of existing data. (d) Data manifold of the auxiliary task distribution. (e) Joint data manifold of primary and auxiliary data distribution, in which the domain-specific knowledge of the auxiliary manifold is also included. (f) Through optimal transport of diverse unsupervised tasks, the learned shared representation only includes rich domain-invariant information from the auxiliary manifold. + +![](images/066b9ed187915da508c3e67f51403bea56694dbe21373a297f8623774f866889.jpg) +Missing data Augmented data +Joint data manifold + +While training robust models from noisy data has been intensively investigated in standard settings, learning from noisy few-shot data is only sparsely explored in existing literature [28,31]. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. Within each training task, FSL operates with few training samples per class, making each data point highly valuable. Hence, the model tends to be more sensitive to variations in the training data and noisy labels can more easily cause the model to learn incorrect patterns, compromising its ability to accurately generalize to new tasks. Furthermore, identifying and rectifying label noises also becomes more challenging. The scarcity of data makes it harder to distinguish between true class characteristics and erroneous annotations. + +Techniques, including robust optimization [28] and data augmentation [31] have been leveraged to handle noises in limited data. However, these techniques, while beneficial, do not entirely resolve the problem. Conventional methods for learning with noisy labels, such as data cleansing, data augmentations, and robust loss functions, are inadequate to tackle the challenges as outlined above. As shown in Fig. [1(a)-(b)], data cleansing introduces information loss, further diminishing the already scarce training data in each task. This poses a greater challenge for the model to learn a meaningful and generalizable representation of the task manifold. Data augmentation could alleviate the information loss introduced through data cleansing as shown in Fig. [1(c)]. Nevertheless, augmentation is performed on the limited training samples, so it may not cover a broad task manifold. Additionally, robust optimization algorithms disregard the concern of information loss, making them less suitable for learning from noisy few-shot data. + +![](images/6959258b7b5d6adb712e53cd6dca8e3ea1efe8d4cf7b902563392ed28ddd249a.jpg) +Fig. 2: In DUNT, the feature extractor extract information from both primary and auxiliary tasks, the classifier predicts the labels of the primary and auxiliary examples, and the domain critic outputs the source of the examples. The primary tasks (top) are sampled and cleansed by episode cleansing, the auxiliary tasks (bottom) are constructed and filtered by their diverseness. (Best viewed in color.) + +To tackle the unique challenges arising from NFSL, we propose novel extensions to existing techniques designed for standard noisy learning scenarios to more effectively handle the information loss in the few-shot setting. In particular, to best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed auxiliary tasks to compensate information loss. The auxiliary tasks are entirely unsupervised as using supervised tasks can be counterproductive in the NFSL setting for three main reasons: (1) They may incur additional annotation overhead. (2) Noisy labels may continue to be introduced that further compromise the manifold structure. (3) Even clean labels can be guaranteed, these labels may not totally align with the primary tasks that can potentially overwhelm that weak learning signals carried by the limited training samples. + +While unsupervised auxiliary tasks have been leveraged in diverse learning scenarios [19,24,32,36,52], including the few-shot setting, the noisy labels and potential information loss introduced through necessary data cleansing makes NFSL much more vulnerable to the potential noises introduced by the auxiliary data. Since data samples are unlabeled, they are usually randomly sampled into an auxiliary task. When a contrastive loss is used to differentiate unlabeled data samples from the same underlying class, the model may be forced to learn sample-specific (spurious) features that cause overfitting. To this end, we propose to conduct novel auxiliary task selection to ensure the intra-diversity among the chosen unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold as shown in Fig. 1(d)-(g). + +Fig. 2 shows the proposed framework, referred to as Diverse UNsupervised Tasks for NFSL (DUNT4NFSL or DUNT for short). The feature extractor of DUNT undergoes updates through three distinct loss components: 1) A cross-entropy loss, integral for extracting discriminative and effective features for classification from the primary tasks; 2) A contrastive loss tailored for the unsupervised auxiliary tasks, designed to encourage an embedding space where similar sample + +pairs remain close while dissimilar pairs are distanced; 3) An adversarial loss, implemented to align the unsupervised auxiliary tasks with the primary tasks, employing a domain critic head to reduce the covariate shift across multiple domains. To ensure that the feature representation remains unaffected by label noise, we apply an episode cleaning technique to the primary task, removing noisy examples in each task. For the auxiliary tasks, we introduce a novel criterion for diverseness to guarantee that the unsupervised auxiliary tasks contribute informative and valuable insights. Our contributions are threefolds: + +- We present DUNT, a framework designed to adeptly utilize unsupervised data from a distinct domain in order to counteract the information loss resulting from data cleansing in the challenging scenario of NFSL. +- We introduce a novel strategy to perform diverse auxiliary task selection to avoid learning sample-specific spurious features from the unlabeled auxiliary data samples. +- We conduct an in-depth theoretical analysis of the auxiliary tasks introduced in the DUNT framework, establishing novel generalization bounds, offering valuable insights into the contributions of these auxiliary tasks. + +We have conducted comprehensive experiments on five benchmark datasets with synthetic label noises to demonstrate the effectiveness and robustness of DUNT and its advantage over SOTA methods by a large margin. + +# 2 Related Work + +Noisy few-shot learning. Existing methods for training DNNs with noisy labels usually fall into the following categories: robust loss [12,49,50,57,60] designs robust cost functions in the presence of noisy labels; noise transition estimation [5,20,53,54] aims to estimate the probabilities that clean labels flip into noisy labels; loss correction [17,22,23,39,44,55,59] either explicitly or implicitly relabels or reweights the noisy samples, a.k.a. data cleansing. Noisy label in FSL is an important yet relatively under-explored topic. The scarcity of data in FSL not only heightens its susceptibility to label noise but also adds complexity to the challenge of identifying and correcting such noise. A few previous attempts such as [28,29,31] address mislabeled issue in FSL settings using re-weighting strategy or generating robust prototypes by feature hybrid and augmentation techniques. Other works such as RW-MAML [25], AQ [14], DFSL [27] tackle the out-of-distribution tasks or adversarial attacks in few-shot learning. In this work, we particularly focus on solving the limited data problem by utilizing auxiliary unlabeled data as unsupervised tasks to enhance the robustness of FSL. + +Optimal transport and Wasserstein distance. The optimal transport (OT) cost is used to measure the difference between distributions supported on high-dimensional space using SGD [1,46]. For a constant $\xi \geq 1$ , the $\xi$ -Wasserstein + +metric between distributions $\mathbb{P}_X$ and $\mathbb{P}_{X'}$ is defined as: + +$$ +\mathcal {W} _ {\xi} \left(\mathbb {P} _ {X}, \mathbb {P} _ {X ^ {\prime}}\right) = \left(\inf _ {\gamma \in \Pi \left(\mathbb {P} _ {X}, \mathbb {P} _ {X ^ {\prime}}\right)} \mathbb {E} _ {\gamma \left(\mathbf {x}, \mathbf {x} ^ {\prime}\right)} \left[ c ^ {\xi} \left(\mathbf {x}, \mathbf {x} ^ {\prime}\right) \right]\right) ^ {\frac {1}{\xi}} \tag {1} +$$ + +where $\mathbf{x},\mathbf{x}^{\prime}\in \mathbb{R}^{d}$ denote the random variables in which the distributions $\mathbb{P}_X$ and $\mathbb{P}_{X^{\prime}}$ are defined. $\prod (\mathbb{P}_X,\mathbb{P}_{X'})$ is the set of all joint distributions (i.e. couplings) whose marginals are $\mathbb{P}_X$ and $\mathbb{P}_{X'}$ , that is, $\int \gamma (\mathbf{x},\mathbf{x}')d\mathbf{x} = \mathbb{P}_{X'}$ and $\int \gamma (\mathbf{x},\mathbf{x}')d\mathbf{x}' = \mathbb{P}_X$ . $c(\mathbf{x},\mathbf{x}'): \mathbb{R}^d\times \mathbb{R}^d\to \mathbb{R}^+$ is the cost function for moving from $\mathbf{x}$ to $\mathbf{x}'$ . Intuitively, $\gamma (\mathbf{x},\mathbf{x}')$ indicates how much "mass" must be transported from $\mathbf{x}$ to $\mathbf{x}'$ in order to transform the distributions $\mathbb{P}_X$ into $\mathbb{P}_{X'}$ . The Wasserstein metric then is the "cost" of the optimal transport plan [1]. Eq. [1] is the primal form of the Wasserstein metric, particularly for the case of $\xi = 1$ with $c(\mathbf{x},\mathbf{x}') = \| \mathbf{x} - \mathbf{x}'\|$ [1]. Some existing efforts propose domain adaptation algorithms using ideas from OT theory, which aims to reduce the divergence between two domains by minimizing the Wasserstein distance between their distributions [7,8,11]. Works that use domain adaptation techniques in FSL [6,18,33,40,56] share the same goal that aims to learn transferable knowledge. In this work, we achieve this goal by leveraging extra data first and then minimize the Wasserstein distance between the primary and auxiliary distributions to ensure the learned knowledge are domain-invariant and discriminative to future unseen tasks. + +# 3 Methodology + +Problem formulation. In FSL, the goal is to learn a mapping $f: \mathcal{X} \to \mathcal{Y}$ that maps an observation $\mathbf{x}$ to its label $\mathbf{y}$ , where $\mathcal{X}$ is the feature space and $\mathcal{Y}$ is the label space. To train the model with the quick adaptation ability to the new tasks with only a small number of examples, a set of tasks $\mathcal{T}_i = \{\mathcal{S}_i^{\mathrm{sup}},\mathcal{S}_i^{\mathrm{que}}\}$ is sampled from the data distribution $\mathcal{D}$ over $\mathcal{X} \times \mathcal{Y}$ , where $S_{i}^{\mathrm{sup}} = \{(x_{i}^{k},y_{i}^{k})\}_{k = 1}^{NK}$ denotes the support set and $S_{i}^{\mathrm{que}} = \{(x_{i}^{k},y_{i}^{k})\}_{k = 1}^{NQ}$ denotes the query set. $N$ is the number of classes, $K$ is the number of examples in each class of the support set, and $Q$ is that of the query set. In DUNT, the meta-training tasks are sampled from the two domains: the primary tasks are sampled from the noisy primary distribution $\tilde{\mathcal{D}}^{\mathrm{pri}}$ over $\mathcal{X} \times \tilde{\mathcal{Y}}$ , where the label space $\tilde{\mathcal{V}}$ is considered noisy, and the unsupervised auxiliary tasks are sampled from the auxiliary distribution $\mathcal{D}^{\mathrm{aux}}$ over $\mathcal{X}'$ , where the entire label space is missing. We assume that the primary and the auxiliary distributions are from different domains: $\tilde{\mathcal{D}}^{\mathrm{pri}} \neq \mathcal{D}^{\mathrm{aux}}$ . During meta-testing, the tasks are sampled from a new unseen target domain $\mathcal{D}^{\mathrm{tgt}}$ over $\mathcal{X}^{\mathrm{tgt}} \times \mathcal{Y}^{\mathrm{tgt}}$ , the label is only used for evaluation purpose. The goal is to learn a meta-model $\theta$ trained on $\tilde{\mathcal{D}}^{\mathrm{pri}}$ and $\mathcal{D}^{\mathrm{aux}}$ that can quickly adapt to tasks in $\mathcal{D}^{\mathrm{tgt}}$ using only a few examples. + +Framework overview. The DUNT framework (see Fig. 2) is composed of a shared feature extractor $\theta$ , the task classifier $\theta_i^c$ for task $i$ and the domain critic $\theta^d$ . The feature extractor takes input from both the primary tasks and the auxiliary + +tasks; the classifiers predict the labels for the input and the critic predict the source of the input. The gradient reversal layer (GRL) ensures that the feature distributions over the two domains are made similar. + +# 3.1 Learning from the Primary Tasks + +One commonly used FSL strategy that takes advantage of task-based formulation is to learn a meta-model through a bi-level optimization. By defining the tasks-specific model parameter for task $\mathcal{T}_i$ as $\theta_i'$ , the meta model $\theta$ are trained by optimizing for the performance of $\theta_i'$ with respect to $\theta$ across tasks sampled from $\mathcal{D}^{\mathrm{pri}}$ . The loss is + +$$ +\mathcal {L} ^ {\mathrm {p r i}} = \sum_ {\mathcal {T} _ {i} \left(\mathcal {S} _ {i} ^ {\sup }, \mathcal {S} _ {i} ^ {\mathrm {q u e}}\right) \sim \tilde {\mathcal {D}} ^ {\mathrm {p r i}}} \mathcal {L} _ {\mathcal {T} _ {i}} \left(\theta_ {i} ^ {\prime}, \mathcal {S} _ {i} ^ {\mathrm {q u e}}\right) \tag {2} +$$ + +where $\theta_i' \gets \theta - \alpha \nabla_\theta \mathcal{L}_{\mathcal{T}_i}(\theta, \mathcal{S}_i^{\mathrm{sup}})$ is the task-specific model adapted from $\theta$ within one gradient step. This is a dual-level optimization where in the outer-level, the meta-model $\theta$ is updated with the summation of a series of tasks evaluated on the query sets $\{\mathcal{S}_i^{\mathrm{que}}\}$ , with $\mathcal{L}_{\mathcal{T}_i}(\theta_i', \mathcal{S}_i^{\mathrm{que}}) = \sum_{(\mathbf{x}_i^k, \mathbf{y}_i^k) \in \mathcal{S}_i^{\mathrm{que}}} \ell(\theta_i^c(\theta_i'(\mathbf{x}_i^k)), \mathbf{y}_i^k)$ . The inner-level objective is computed using the task-specific model $\theta_i'$ updated with support set $\mathcal{S}_i^{\mathrm{sup}}$ using loss $\mathcal{L}_{\mathcal{T}_i}(\theta, \mathcal{S}_i^{\mathrm{sup}}) = \sum_{(\mathbf{x}_i^k, \mathbf{y}_i^k) \in \mathcal{S}_i^{\mathrm{sup}}} \ell(\theta_i^c(\theta(\mathbf{x}_i^k)), \mathbf{y}_i^k)$ , where $\ell$ is the cross-entropy loss, $\theta_i^c$ is the $i$ -th task classifier. + +Episode cleansing. In the dual-level optimization, the meta-model $\theta$ is updated by first adapting to different task-specific models $\theta_{i}^{\prime}$ with the corresponding support set and then updated using the summation of the loss of the task-specific models $\theta_{i}^{\prime}$ on their corresponding query sets. Since the dataset is noisy, the error may be accumulated through the noisy query set and cause the overfitting of the meta-model. Therefore, we propose episode cleansing to remove the noisy data in the query sets based on the memorization effect which indicates that noisy examples have higher losses than the clean ones. The classification loss $\mathcal{L}_{\mathcal{T}_i}(\theta_i', S_i^{\mathrm{que}})$ becomes + +$$ +\mathcal {L} _ {\mathcal {T} _ {i}} \left(f _ {\theta_ {i} ^ {\prime}}, \hat {\mathcal {S}} _ {i} ^ {\mathrm {q u e}}\right) = \sum_ {\mathbf {x} _ {i} ^ {k}, \mathbf {y} _ {i} ^ {k} \in \mathcal {S} _ {i} ^ {\mathrm {q u e}}} w _ {i} ^ {k} \ell_ {i} ^ {k}, \tag {3} +$$ + +where $\hat{S}_i^{\mathrm{que}}$ is the cleansed (weighted) query set for task $\mathcal{T}_i$ ; for simplification, we use $\ell_i^k$ to denote $\ell (\theta_i^c (\theta_i'(x_i^k)),y_i^k)$ . $w_{i}^{k}$ is the self-paced sample weight for each example pair, and + +$$ +w _ {i} ^ {k} = \mathbb {1} (\ell_ {i} ^ {k} < \gamma_ {t h}), +$$ + +where $\mathbb{1}(\cdot)$ is an indicator function whose value is 1 when $\ell_i^k < \gamma_{th}$ and 0 otherwise; and $\gamma_{th}$ is a predefined hyperparameter used to filter out high-loss examples. We use the percentage of the ordered loss to avoid extensive hyperparameter search, whose ablation is provided in Appendix C. + +The primary loss the becomes: + +$$ +\mathcal {L} ^ {\mathrm {p r i}} = \sum_ {\mathcal {T} _ {i} (\mathcal {S} _ {i} ^ {\sup }, \mathcal {S} _ {i} ^ {\mathrm {q u e}}) \sim \tilde {\mathcal {D}} ^ {\mathrm {p r i}}} \mathcal {L} _ {\mathcal {T} _ {i}} \left(\theta_ {i} ^ {\prime}, \hat {\mathcal {S}} _ {i} ^ {\mathrm {q u e}}\right). \tag {4} +$$ + +Similarly, for the support set label noise, we use meta-model to compute the loss values and remove the noisy ones in the support set. + +# 3.2 Learning from Diverse Unsupervised Tasks + +As shown in Fig. 2, the general unsupervised auxiliary task $\mathcal{T}_j$ is constructed by i) first randomly sampling $N$ images from the dataset for the support set $\mathcal{S}_j^{\mathrm{sup}} = \{\mathbf{x}_j^k\}_{k=1}^N$ ; ii) the query set $\mathcal{S}_j^{\mathrm{que}} = \{A_1(\mathbf{x}_j^k), \dots, A_Q(\mathbf{x}_j^k)\}_{k=1}^N$ is obtained by augmenting the $N$ images in $Q$ different ways, where $A_1(\cdot), \dots, A_Q(\cdot)$ denote $Q$ different augmentation techniques, such as random cropping, translation, flipping. The auxiliary loss is the summation of the contrastive loss of a series of auxiliary tasks: + +$$ +\mathcal {L} ^ {\text {a u x}} = \sum_ {\mathcal {T} _ {j} \left(\mathcal {S} _ {j} ^ {\sup }, \mathcal {S} _ {j} ^ {\text {q u e}}\right) \sim \mathcal {D} ^ {\text {a u x}}} \mathcal {L} _ {\mathcal {T} _ {j}} \left(\theta , \mathcal {S} _ {j} ^ {\sup }, \mathcal {S} _ {j} ^ {\text {q u e}}\right). \tag {5} +$$ + +The contrastive loss is utilized for the unsupervised tasks: + +$$ +\mathcal {L} _ {\mathcal {T} _ {j}} \left(\theta , \mathcal {S} _ {j} ^ {\sup }, \mathcal {S} _ {j} ^ {\text {q u e}}\right) = - \frac {1}{N Q} \sum_ {k = 1} ^ {N} \sum_ {q = 1} ^ {Q} \log \frac {\exp \left(- d \left[ \theta \left(A _ {q} \left(\mathbf {x} _ {j} ^ {k}\right)\right) , \theta \left(\mathbf {x} _ {j} ^ {k}\right) \right]\right)}{\sum_ {l = 1} ^ {N} \exp \left(- d \left[ \theta \left(A _ {q} \left(\mathbf {x} _ {j} ^ {k}\right)\right) , \theta \left(\mathbf {x} _ {j} ^ {l}\right) \right]\right)} \tag {6} +$$ + +where $d[\cdot, \cdot]$ is a distance metric, such as Euclidean distance. + +Auxiliary task selection. In contrastive learning, the positive pairs are formulated by the original example and its augmentation, and the negative pairs are formulated by images from two different classes. The contrastive loss is designed to pull the positive pairs close to each other and push the negative ones away. However, since each image in the unsupervised task are randomly sampled and treated as a class, there are chances that the images are sampled from the same class. In some extreme case, all the images are sampled from the same class, which means that the positive pairs being treated as negative pairs and being push away from each other. This would cause the model being forced to overfit to the unnecessary details and creates spurious correlations and harm the generalization ability. We denote those tasks as confusing tasks, as a comparison, the tasks with images sampled from different classes are denoted as valuable tasks. The intricate nature of confusing tasks appears to pose a significant threat to the resilience and robustness of FSL, as shown in our later ablation studies. In this work, we propose a novel task selection metric to remove these confusing unsupervised tasks. + +Given two images $\mathbf{x}_j^k$ , $\mathbf{x}_j^l$ , the model parameterized by $\theta$ lead to class probabilities vectors $\mathbf{p}_j^k = \mathrm{softmax}(\theta (\mathbf{x}_j^k))$ and $\mathbf{p}_j^l = \mathrm{softmax}(\theta (\mathbf{x}_j^l))$ . If the two images + +are from the same class, $\mathbf{p}_j^k$ and $\mathbf{p}_j^l$ are naturally encouraged to be close to the assignment vector $\mathbf{e}_y$ in $\{0,1\}^N$ with a single non-zero entry at position $y$ , where $y$ is the class label associated to $\mathbf{x}$ and $N$ is the number of classes. Therefore, we consider the following diversity criterion: + +$$ +d i v _ {j} = \frac {1}{C (N , 2)} \sum_ {i = 1} ^ {N - 1} \sum_ {j = i + 1} ^ {N} \frac {1}{\cos (\mathbf {p} _ {j} ^ {k} , \mathbf {p} _ {j} ^ {l})}. +$$ + +The diverseness of the task $\mathcal{T}_j$ is defined as the average of the reciprocal of the cosine similarity between the class probabilities of all example pairs in the task, where $C(N,2) = \frac{N!}{(N - 2)!2!}$ is the 2-combination of set $N$ . According to the definition, the higher the value of $div_j$ , the more diverse the task. + +Auxiliary loss function. Auxiliary loss $\mathcal{L}^{\mathrm{aux}}$ is updated as + +$$ +\mathcal {L} ^ {\mathrm {a u x}} = \sum_ {\mathcal {T} _ {j} (\mathcal {S} _ {j} ^ {\sup }, \mathcal {S} _ {j} ^ {\mathrm {q u e}}) \sim \mathcal {D} ^ {\mathrm {a u x}}} v _ {j} \mathcal {L} _ {\mathcal {T} _ {j}} (\theta , \mathcal {S} _ {j} ^ {\sup }, \mathcal {S} _ {j} ^ {\mathrm {q u e}}) +$$ + +where $v_{j}$ is the diversity weight defined as: + +$$ +v _ {j} = \mathbb {1} (d i v _ {j} > \gamma_ {d i v}), +$$ + +where $\gamma_{div}$ is a predefined hyperparameter, which filter out those less diverse, a.k.a., confusing tasks. In practice, we simplify the hyperparameter searching by first upsampling six auxiliary tasks, computes each task diversity metric, and only keep the most diverse task in $\{\mathcal{T}_j\}^{\mathrm{val}}$ for training. The corresponding ablations about the numbers of extra tasks sampled are presented in Sec. 4.3 + +# 3.3 Adversarial Training and the Critic Head + +Since the auxiliary tasks are drawn from a different data distribution, it is inevitable that there exist some domain shift from the primary domain to the auxiliary one. While learning from the domain-specific knowledge from the auxiliary domain main cause the learned model biased and deteriorate the model's generalization performance, learning from the domain-invariant knowledge helps improve the generalization ability. To enforce the learning of the domain-invariant representation of the meta-model, we propose a critic model $\theta^c$ to discriminate the source of the input data, either primary or auxiliary. By training the model in an adversarial way, the critic model $\theta^c$ try to distinguish which domain that the input features come from, while the meta-model $\theta$ aims to confuse the critic. In this way, when the training process converge, $\theta$ tends to extract domain-invariant knowledge from both domains. Formally, the overall adversarial objective is a min-max problem as follows: + +$$ +\min _ {\theta , \theta^ {c}} \max _ {\theta^ {d}} \alpha_ {1} \mathcal {L} ^ {\text {p r i}} + (1 - \alpha_ {1}) \mathcal {L} ^ {\text {a u x}} + \alpha_ {2} \mathcal {E} (\theta , \theta^ {d}, \tilde {\mathcal {D}} ^ {\text {p r i}}, \mathcal {D} ^ {\text {a u x}}), \tag {7} +$$ + +where the critic model $\theta^d$ assigns 1 to the example from the primary domain and 0 otherwise based on the feature representation extracted by $\theta$ . $\mathcal{E}(\cdot)$ is the adversarial loss + +$$ +\mathcal {E} (\theta , \theta^ {d}, \hat {\mathcal {D}} ^ {\mathrm {p r i}}, \mathcal {D} ^ {\mathrm {a u x}}) = \mathbb {E} _ {\mathbf {x} \sim \hat {\mathcal {D}} ^ {\mathrm {p r i}}} [ g (\mathbf {x}) ] - \mathbb {E} _ {\mathbf {x} \sim \mathcal {D} ^ {\mathrm {a u x}}} [ g (\mathbf {x}) ], +$$ + +where $g(\mathbf{x}) = \theta^d (\theta (\mathbf{x}))$ , with its empirical counterpart as + +$$ +\hat {\mathcal {E}} (g _ {\theta , \theta^ {d}}, \tilde {\mathcal {D}} ^ {\mathrm {p r i}}, \mathcal {D} ^ {\mathrm {a u x}}) = \frac {1}{| \mathcal {D} ^ {\mathrm {p r i}} |} \sum_ {x \sim \mathcal {D} ^ {\mathrm {p r i}}} [ g (\mathbf {x}) ] - \frac {1}{| \mathcal {D} ^ {\mathrm {a u x}} |} \sum_ {\mathbf {x} \sim \mathcal {D} ^ {\mathrm {a u x}}} [ g (\mathbf {x}) ] +$$ + +The optimization of Eq. (7) is conducted through an alternative training between the optimization of the meta-model feature extractor $\theta$ , classifier parameter $\theta^c$ and the critic model parameter $\theta^d$ while keeping the other fixed. Parameters of $\theta$ , $\theta^c$ and $\theta^d$ are updated alternatively. In practice, $\theta$ and $\theta^c$ are connected by the GRL, which reverses the gradients that flow through $\theta^c$ . The pseudo code of the algorithm can be found in Appendix A. + +# 3.4 Theoretical Analysis + +We conduct a theoretical analysis of DUNT to show: 1) the equivalence between the adversarial regularizer and the Wasserstein distance, which captures the diversity property of the auxiliary distribution while enforcing the learning of a domain invariant representation; 2) the relationship between the primary and auxiliary distributions; 3) a tight generalization bound of the target based on the expected risk of the primary and auxiliary distributions. + +Equivalence to Wasserstein distance. The Kantorovich-Rubinstein of Wasserstein distance takes the form [1]: + +$$ +\mathcal {W} (\mathbb {P} _ {1}, \mathbb {P} _ {2}) = \sup _ {\| h \| _ {L} \leq 1} \mathbb {E} _ {\mathbf {x} \sim \mathbb {P} _ {1}} [ h (\mathbf {x}) ] - \mathbb {E} _ {\mathbf {x} \sim \mathbb {P} _ {2}} [ h (\mathbf {x}) ] \tag {8} +$$ + +where the supremum is over all the 1-Lipschitz functions $h: \mathcal{X} \to \mathbb{R}$ . Therefore, if we have a parameterized family of functions $\{h_w\}_{w \in \mathcal{W}}$ that are all 1-Lipschitz, we could consider solving the problem + +$$ +\max \mathbb {E} _ {\mathbf {x} \sim \mathbb {P} _ {1}} \left[ h _ {w} (\mathbf {x}) \right] - \mathbb {E} _ {\mathbf {x} \sim \mathbb {P} _ {2}} \left[ h _ {w} (\mathbf {x}) \right] \tag {9} +$$ + +which has the same form of the regularizer in Sec. 3.3. This indicates that our adversarial regularizer has the same theoretical foundation of Wasserstein distance optimal transport. To satisfy the 1-Lipschitz constraint, the critic $\theta^d$ is implemented with gradient penalty [15]. The overall objective can be reformulated as + +$$ +\min _ {\theta} \alpha_ {1} \mathcal {L} ^ {\text {p r i}} + (1 - \alpha_ {1}) \mathcal {L} ^ {\text {a u x}} + \alpha_ {2} \mathcal {W} _ {1} (\mathcal {D} ^ {\text {p r i}}, \mathcal {D} ^ {\text {a u x}}), \tag {10} +$$ + +where the first and second term are the primary and auxiliary loss, resp., and the third term is the Wasserstein distance based regularizer, which aims to extract domain-invariant feature from the primary and auxiliary domains. Wasserstein distance reflects an optimal transport cost for moving one distribution to another. A smaller transport cost means a better coverage of the distribution, which captures the property of diversity [42]. + +Relationship of the primary and auxiliary distributions. By viewing this problem as a distribution matching that matches the auxiliary distribution with the primary one, we have the generalization property that quantify the relationship between the primary and the auxiliary distribution. Given the primary distribution $\mathcal{D}^{\mathrm{pri}}$ and the auxiliary distribution $\mathcal{D}^{\mathrm{aux}}$ , the expected risk w.r.t. is defined as $\mathcal{D}^{\mathrm{pri}}R_{\mathcal{D}^{\mathrm{pri}}} = \mathbb{E}_{x\sim \mathcal{D}^{\mathrm{pri}}}\ell (h(x),h^{*}(x))$ , and the expected risk w.r.t. $\mathcal{D}^{\mathrm{aux}}$ is defined as $R_{\mathcal{D}^{\mathrm{aux}}} = \mathbb{E}_{x\sim \mathcal{D}^{\mathrm{aux}}}\ell (h(x),h^{*}(x))$ . We have the following Lemma 1: + +Lemma 1. If the loss $\ell$ is symmetric and $L$ -Lipschitz, $\forall h \in \mathcal{H}$ is at most $H$ -Lipschitz function and the underlying mapping function $h^*$ is $\phi(\lambda) - (\mathcal{D}^{pri}, \mathcal{D}^{aux})$ joint probabilistic Lipschitz; then the expected risk w.r.t. $\mathcal{D}^{pri}$ is upper bounded + +$$ +R _ {\mathcal {D} ^ {p r i}} \leq R _ {\mathcal {D} ^ {a u x}} + L (H + \lambda) \mathcal {W} _ {1} (\mathcal {D} ^ {p r i}, \mathcal {D} ^ {a u x}) + L \phi (\lambda), \tag {11} +$$ + +where $\phi : \mathbb{R} \to [0,1]$ and a mapping $h^*$ is $\phi(\lambda) - (\mathcal{D}^{pri},\mathcal{D}^{aux})$ joint probabilistic Lipschitz if $\text{supp}(\mathcal{D}^{aux}) \subseteq \text{supp}(\mathcal{D}^{pri})$ ; $\forall \lambda > 0$ , all coupling $\gamma \in \prod(\mathcal{D}^{pri},\mathcal{D}^{aux}) : \mathbb{P}_{(x_{\mathcal{D}^{pri}},x_{\mathcal{D}^{aux}})\sim \gamma}[|h^{*}(x_{\mathcal{D}^{pri}}) - h^{*}(x_{\mathcal{D}^{aux}})| > \lambda \| x_{\mathcal{D}^{pri}} - x_{\mathcal{D}^{aux}}\|_2] \leq \phi(\lambda)$ . + +Lemma 1 shows that the expected risk of the primary distributions is upper bounded by the expected risk w.r.t. the auxiliary distribution, the Wasserstein distance, and the labeling function property. The proof of Lemma 1 can be derived with Theorem 1 in [42]. Eq. (10) follows the upper bound from Lemma 1, which means a desirable auxiliary distribution with a better matching with the primary distribution which holds a small expected risk leads to tighter generalization bound for the primary distribution. This further verifies the rationality of our design of using the WD based regularizer. + +Generalization bound of DUNT. Furthermore, in the following Lemma 2 we derive the upper bound for the target risk $R_{\mathcal{D}^{\mathrm{tgt}}} = \mathbb{E}_{(\mathbf{x},y)\sim \mathcal{D}^{\mathrm{tgt}}}\ell (\mathbf{x},y)$ . + +Lemma 2. Let $\mathcal{D}^{tgt}$ be the target distribution, $h$ is $K$ -Lipschitz w.r.t. the feature $\mathbf{x}$ , and the setting of loss and hypothesis be the same as Lemma 1, the expected risk w.r.t. the target distribution is upper-bounded by: + +$$ +\begin{array}{l} R _ {\mathcal {D} ^ {t g t}} \leq \alpha_ {1} \left(R _ {\mathcal {D} ^ {p r i}} + L K \mathcal {W} _ {1} \left(\mathcal {D} ^ {p r i}, \mathcal {D} ^ {t g t}\right)\right) \tag {12} \\ + (1 - \alpha_ {1}) \left(R _ {\mathcal {D} ^ {a u x}} + L K \mathcal {W} _ {1} \left(\mathcal {D} ^ {a u x}, \mathcal {D} ^ {t g t}\right)\right), \\ \end{array} +$$ + +Lemma 2 shows that the expected risk w.r.t. the target distribution is upper bounded by the expected risk w.r.t. the primary distribution and the auxiliary distribution, and the corresponding Wasserstein distances. The proof of Lemma 2 + +can be derived with Theorem 2 in [41]. Lemma 2 provide solid theoretical foundation for the design of DUNT with a tight generalization bound to the unseen target domain with primary and auxiliary distributions, which can be extended to multiple sources. + +# 4 Experiments + +# 4.1 Datasets, Baselines, and Experimental Settings + +Datasets. We conduct experiments using the following benchmark datasets for FSL: CUB 51, miniImageNet (mini) 48, tieredImageNet (tiered) 38, CIFAR-FS (CF) 3, and Cross-Domain dataset 16. Each datasets contains synthetic labels noise with different ratios. More details are in the Appendix. + +Baselines. We use ProtoNet [43] as the base few-shot learning method and consider comparing with improving methods with general data cleansing methods including SPL [26] and FSR [61], denoted as $PN + SPL$ and $PN + FSR$ , respectively. We also consider to improve ProtoNet in transductive settings including CoTeaching [17], CoTeaching+ [59], denoted as $PN + Co$ , and $PN + Co+$ , respectively, STARTUP [35], and DDNet [21]. Last but not least, we consider to compare with noisy FSL methods RNNP [31], TraNFS [28] and DCML [37]. The detailed descriptions and settings are provided in the Appendix. + +Experimental settings. The feature extractor follows the same architecture as the ProtoNet used by [43], which has 4 modules with a $3 \times 3$ convolutions and 64 filters, followed by batch normalization, a ReLU nonlinearity, and $2 \times 2$ max-pooling, a.k.a., Conv4. The architecture of classifier have the design with a fully connected layer followed by a ReLU nonlinearity, a dropout layer, and another fully connected layer. The critic architecture has the same architecture with an extra sigmoid function. We've considered transductive and non-transductive in the settings. For the former, unlabeled test data of the same dataset is used as auxiliary data during training. This is applied to our method and other transductive methods including $\mathrm{PN} + \mathrm{Co}$ , $\mathrm{PN} + \mathrm{Co}+$ , STARTUP, and DDNet. For the latter, we utilize unlabeled datasets from other different datasets as auxiliary. This setting is applied to all the mentioned methods. During training, we use a learning rate $=0.001$ , and use Adam as the optimizer. The batch size is 4. The meta-model is trained with 40000 randomly selected primary tasks and 40000 randomly selected auxiliary tasks. During evaluation, the average accuracy is computed based on 600 randomly sampled tasks. + +# 4.2 Main Experimental Results + +Following the experiment design of [48], we report the average accuracy of 5 runs of each 5-way 1-shot and 5-way 5-shots experiments in Tab. [1] to Tab. [2]. We use $\alpha_{1} = 0.5$ , $\alpha_{2} = 0.02$ , $\gamma_{th} = 50\%$ , $\gamma_{div}$ as one out of six tasks for our methods + +Table 1: Non-transductive auxiliary datasets + +
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+SPL48.55 ± 1.5143.44 ± 1.4937.06 ± 0.6342.11 ± 0.7155.20 ± 0.69
PN+FSR37.14 ± 0.6044.81 ± 0.5939.82 ± 0.6830.65 ± 0.4937.39 ± 0.63
RNNP55.39 ± 0.4950.18 ± 0.6241.75 ± 0.7243.76 ± 0.4860.15 ± 0.72
TraNFS53.02 ± 0.5246.66 ± 0.6338.25 ± 0.7940.32 ± 0.8860.86 ± 0.63
DCML54.39 ± 0.8749.33 ± 1.2241.77 ± 0.8048.33 ± 0.8060.28 ± 0.82
DUNT56.94 ± 1.0352.22 ± 1.1043.01 ± 0.5149.12 ± 0.6862.80 ± 0.49
Task5-way 5-shot
PN+SPL68.47 ± 0.7458.67 ± 0.8654.94 ± 1.0359.62 ± 1.3170.93 ± 0.83
PN+FSR51.48 ± 0.9161.89 ± 0.8057.38 ± 1.1050.40 ± 1.2144.68 ± 1.09
RNNP70.37 ± 1.3363.57 ± 1.2858.79 ± 2.3655.07 ± 1.2174.31 ± 1.36
TraNFS68.77 ± 1.1761.06 ± 1.1351.24 ± 1.1648.74 ± 1.2074.15 ± 0.83
DCML71.28 ± 1.2362.77 ± 1.0057.77 ± 1.3361.19 ± 1.0172.33 ± 0.90
DUNT72.17 ± 1.1165.10 ± 1.2358.50 ± 1.1162.01 ± 1.0974.44 ± 1.42
+ +in all datasets. In Tab. 1 and Tab. 2 we have provided the test accuracy all indomain datasets and one cross-domain dataset: mini-CUB for both transductive and non-transductive settings. The performance of cross-domain dataset 16 is reported in Tab. 3 + +Non-transductive results. For Tab. DUNT uses miniimagenet as auxiliary data for CUB datasets, uses tieredimagnet as auxiliary data for the cross-domain mini-CUB, and uses CUB as auxiliary data for the rest datasets. The results show that, with extra auxiliary datasets, DUNT has outperformed the general data-cleansing techniques and other noisy FSL methods in general. In Tab. the performance of SPL and FSR are generally worse than that of TraNFS and RNNP, which confirms our assumption that by removing potential noisy data, the data cleansing technique might harm the performance of FSL due to the lack of information, which can be compensated with our auxiliary data design. And compared to TraNFS and RNNP, DUNT has reduced the negative impact of the noisy labels more effectively. + +Transductive results. In Tab. 2, we consider the transductive setting where the auxiliary dataset for all the methods are unlabeled test data of the same dataset. The results show that our method ranks the top among all the transductive methods. Compared to Tab. 1, some of the transductive methods underperform non-transductive ones due to the following potential reasons: i) the test dataset is smaller than other auxiliary datasets. Therefore, the non-transductive dataset has more information to provide. Take miniImageNet as an example, the test data of the miniImageNet dataset only contains 1000 images and 10 classes, while the auxiliary dataset CUB contains 200 categories and 12000 images, the abundance and diversity of the auxiliary helps greatly improve the model's performance over the primary tasks more than the close test data does; and ii) the auxiliary dataset provides more valuable fine-grained tasks that benefits representation learning. For instance, when training miniImageNet with CUB as an auxiliary, both datasets contain natural images which are closely related, and + +Table 2: Transductive auxiliary datasets + +
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+Co49.10 ± 1.2144.44 ± 0.9636.48 ± 0.9840.36 ± 0.8152.51 ± 0.92
PN+Co+49.45 ± 1.1945.08 ± 0.9139.54 ± 1.0940.19 ± 1.1051.27 ± 1.11
STARTUP52.02 ± 1.2048.79 ± 1.1840.08 ± 1.2244.48 ± 0.9260.69 ± 1.31
DDNet52.55 ± 1.3950.82 ± 0.8741.11 ± 1.8945.11 ± 1.3360.25 ± 1.46
DUNT55.01 ± 1.1252.38 ± 0.8542.07 ± 1.1049.29 ± 1.2661.90 ± 1.14
Task5-way 5-shot
PN+Co65.42 ± 1.2259.57 ± 1.2056.90 ± 1.2955.10 ± 1.2166.89 ± 1.26
PN+Co+67.92 ± 1.1160.43 ± 1.1354.29 ± 1.1054.54 ± 0.8467.22 ± 0.71
STARTUP68.74 ± 1.2162.00 ± 1.2451.07 ± 1.0155.92 ± 0.9974.04 ± 0.84
DDNet69.21 ± 1.3463.70 ± 2.0153.87 ± 0.4457.22 ± 0.7573.92 ± 1.20
DUNT71.20 ± 1.1266.46 ± 0.9859.05 ± 1.0161.09 ± 0.9076.11 ± 0.98
+ +Table 3: Cross-domain Datasets + +
MethodsEuroSATCropDiseaseChestXISICEuroSATCropDiseaseChestXISIC
5-way 1-shot5-way 5-shot
ProtoNet52.52±0.5256.32±0.8928.90±0.4330.11±1.0061.60±0.7078.89±0.8333.04±0.7136.95±0.90
PN+SPL52.89±1.2158.95±2.0427.68±0.5231.79±1.0967.60±1.2380.25±1.1931.13±1.1837.05±1.15
PN+FSR47.53±1.2156.97±1.2328.40±0.6133.58±0.9261.96±0.9075.93±0.8631.43±1.1936.42±1.22
RNNP51.86±1.2056.56±1.2127.39±0.6131.26±0.8067.97±1.3478.43±1.2931.16±1.2537.84±1.19
TraNFS50.61±1.1058.39±0.9128.26±0.9832.84±1.1167.93±1.1181.51±1.2332.16±1.1636.95±1.18
DCML52.33±1.0258.36±0.4627.54±0.6931.96±0.8068.01±0.9079.11±0.9831.64±1.0035.44±1.10
PN+Co56.56±1.0859.13±1.1426.67±0.7329.16±1.1372.08±1.1578.59±2.1130.45±1.1734.74±1.09
PN+Co+52.18±1.1158.18±1.1028.23±0.8730.63±1.0665.57±1.1078.47±0.9030.77±0.8233.68±0.91
STARTUP52.74±1.2259.43±1.2327.75±0.9532.84±1.1369.78±1.1780.12±1.2231.69±1.3039.58±0.99
DDNet55.21±1.8360.09±1.0928.66±0.9834.10±1.0170.09±1.9081.90±2.0932.01±1.8939.44±1.13
DUNT57.10±1.2961.49±1.0129.78±0.9035.80±0.7772.81±1.1083.20±1.0833.94±1.2040.31±0.90
+ +CUB dataset is a collection of bird species, which provide fine-grained tasks for learning a better feature representation, while the miniImageNet test data only provide coarse-grained task, which makes it less useful in some extend. As for DUNT, the performance stays stable regardless the two settings, showing that our framework has the potential to fully utilize any extra information provided. + +Cross-domain results. In Tab. 3, we present the performance on the cross-domain datasets. The training procedure follows the setting in [16], which uses miniImageNet as base dataset for all the methods. For CoTeaching, CoTeaching+, STARTUP, DDNet and DUNT, we use the unlabeled test data EuroSAT, CropDisease, ChestX, and ISIC as auxiliary dataset, respectively. In cross-domain datasets, the transductive methods generally outperform the non-transductive ones due to the utilization of the unlabeled test data that reduces the domain gap between the source domain and the target domain, which is critical when the they are extremely different. DUNT has used the Wasserstein regularization to further reduce the differences. Hence, it ranks the highest in the cross-domain datasets. + +# 4.3 Ablation Study + +In this section, we study the effectiveness of different components and impact of hyperparameters. For more ablations, please refer to Appendix C. + +Effectiveness of different components. Tab. 4 shows the ablation study with different elements of DUNT on CUB and miniImageNet datasets under nontransductive setting. We first show the performance of only using the primary data, which is vanilla ProtoNet (Primary). It can be observed that each of these components have a significant impact on the overall performance. + +Table 4: Impact of different components + +
PrimaryAuxiliaryRegECTSCUBmini
✓XXXX42.57 ± 0.3439.95 ± 1.06
✓XX✓X48.55 ± 1.5143.44 ± 1.49
✓✓XXX50.19 ± 0.7148.90 ± 0.77
✓✓X✓X53.55 ± 0.9750.02 ± 0.81
✓✓XX✓52.70 ± 1.0149.50 ± 1.66
✓✓X✓✓54.31 ± 0.9950.18 ± 0.87
✓✓✓✓X51.90 ± 0.8949.18 ± 0.76
✓✓✓✓X55.47 ± 0.7950.73 ± 0.60
✓✓✓X✓54.79 ± 1.1049.91 ± 1.00
✓✓✓✓✓56.94 ± 1.0352.22 ± 1.10
+ +Auxiliary data (Auxiliary) and episode cleansing (EC) serve pivotal functions, with the former introducing additional data and the latter eliminating noisy data. The regularization term (Reg) aids in extracting domain-invariant features, contributing to further performance improvement. Task selection (TS) mitigates the risk of overfitting, thereby enhancing overall performance. + +Effectiveness of task selection. We study the impact of task selection through the threshold $\gamma_{div}$ . As mentioned earlier, to streamline the hyperparameter search process, we opt for an approach that involves upsampling $M$ unsupervised tasks. Subsequently, + +![](images/119c14515f988d49b3272c985d6487b8da064817c7b0187310459b8a882185c9.jpg) +(a) + +![](images/c2eed60c02f5ad14d0614b1639173f300d880f92f1f4932967b26cdfbec4e02d.jpg) +(b) +Fig. 3: DUNT on CUB with different $\gamma_{div}$ (a); DUNT with different task selection (b) + +we select only the most diverse task to be included in the valuable task group, instead of specifying a real value for the diverseness threshold. In this ablation, the number of upsampled tasks varies from 2-10. As depicted in Fig. 3 (a), the performance improves alongside an increase in the parameter $M$ , but this comes at the cost of heightened computational demands. Consequently, considering efficiency, we opt for $\gamma_{\mathrm{div}} = 6$ for all experiments. In Fig. 3 (b), we test the effect of using different unsupervised task: select confusing tasks and select valuable tasks. It is evident from the results that the inclusion of confusing tasks significantly deteriorates performance, whereas the utilization of valuable tasks leads to a substantial improvement. + +# 5 Conclusion + +We propose a novel FSL framework with auxiliary tasks that utilize carefully selected unlabeled data under noisy settings. To better align the auxiliary distribution with the primary one, we propose a regularization term based on the Wasserstein distance for learning a domain-invariant representation. To enhance robustness against label noise, we introduce episode cleansing for examples in the primary task and adopt a diverse task selection strategy for the unsupervised auxiliary tasks. Our framework is theoretically and experimentally proved to be effective and beneficial. + +# References + +1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017) +2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009) +3. Bertinetto, L., Henriques, J.F., Torr, P.H., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: Proceedings of the International Conference on Learning Representations (2019) +4. Beyer, L., Henaff, O.J., Kolesnikov, A., Zhai, X., Oord, A.v.d.: Are we done with imagenet? arXiv preprint arXiv:2006.07159 (2020) +5. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022) +6. Chobola, T., Vašata, D., Kordík, P.: Transfer learning based few-shot classification using optimal transport mapping from preprocessed latent space of backbone neural network. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. pp. 29-37. PMLR (2021) +7. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Advances in Neural Information Processing Systems 30 (2017) +8. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 274-289. Springer (2014) +9. Deleu, T., Würfl, T., Samiei, M., Cohen, J.P., Bengio, Y.: Torchmeta: A Meta-Learning library for PyTorch (2019), https://arxiv.org/abs/1909.06576, available at: https://github.com/tristandeleu/pytorch-meta +10. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. pp. 1126-1135 (2017) +1. Flamary, R., Courty, N., Tuia, D., Rakotomamonjy, A.: Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell 1 (2016) +2. Ghosh, A., Kumar, H., Sastry, P.S.: Robust loss functions under label noise for deep neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017) +3. Gidaris, S., Bursuc, A., Komodakis, N., Pérez, P., Cord, M.: Boosting few-shot visual learning with self-supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8059-8068 (2019) +4. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020) +5. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017) +6. Guo, Y., Codella, N.C., Karlinsky, L., Codella, J.V., Smith, J.R., Saenko, K., Rosing, T., Feris, R.: A broader study of cross-domain few-shot learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 124-141. Springer (2020) + +17. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Proceedings of the Conference on Neural Information Processing Systems (2018) +18. Han, C., Fan, Z., Zhang, D., Qiu, M., Gao, M., Zhou, A.: Meta-learning adversarial domain adaptation network for few-shot text classification. In: Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021. Findings of ACL, vol. ACL/IJCNLP 2021, pp. 1664-1673. Association for Computational Linguistics (2021) +19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020) +20. Hendrycks, D., Mazeika, M., Wilson, D., Gimpel, K.: Using trusted data to train deep networks on labels corrupted by severe noise. Advances in neural information processing systems 31 (2018) +21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Feris, R., Radke, R.J.: Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems 34, 3584-3595 (2021) +22. Jiang, L., Meng, D., Mitamura, T., Hauptmann, A.G.: Easy samples first: Self-paced reranking for zero-example multimedia search. In: Proceedings of the 22nd ACM international conference on Multimedia. pp. 547-556 (2014) +23. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018) +24. Khodadadeh, S., Boloni, L., Shah, M.: Unsupervised meta-learning for few-shot image classification. Advances in neural information processing systems 32 (2019) +25. Killamsetty, K., Li, C., Zhao, C., Iyer, R., Chen, F.: A reweighted meta learning framework for robust few shot learning. arXiv preprint arXiv:2011.06782 (2020) +26. Kumar, M.P., Packer, B., Koller, D.: Self-paced learning for latent variable models. In: Proceedings of the Conference on Neural Information Processing Systems. vol. 1, p. 2 (2010) +27. Li, W., Wang, L., Zhang, X., Qi, L., Huo, J., Gao, Y., Luo, J.: Defensive few-shot learning. IEEE transactions on pattern analysis and machine intelligence (2022) +28. Liang, K.J., Rangrej, S.B., Petrovic, V., Hassner, T.: Few-shot learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9089-9098 (2022) +29. Lu, J., Jin, S., Liang, J., Zhang, C.: Robust few-shot learning for user-provided data. IEEE transactions on neural networks and learning systems 32(4), 1433-1447 (2020) +30. Masi, I., Trn, A.T., Hassner, T., Sahin, G., Medioni, G.: Face-specific data augmentation for unconstrained face recognition. International Journal of Computer Vision 127, 642-667 (2019) +31. Mazumder, P., Singh, P., Namboodiri, V.P.: Rnnp: A robust few-shot learning approach. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2664-2673 (2021) +32. Medina, C., Devos, A., Grossglauer, M.: Self-Supervised Prototypical Transfer Learning for Few-Shot Classification. arXiv preprint arXiv:2006.11325 (2020) +33. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in neural information processing systems 30 (2017) +34. Northcutt, C.G., Athalye, A., Mueller, J.: Pervasive label errors in test sets destabilize machine learning benchmarks. arXiv preprint arXiv:2103.14749 (2021) + +35. Phoo, C.P., Hariharan, B.: Self-training for few-shot transfer across extreme task differences. In: Proceedings of the International Conference on Learning Representations (2021) +36. Qin, T., Li, W., Shi, Y., Gao, Y.: Diversity helps: Unsupervised few-shot learning via distribution shift-based data augmentation. arXiv preprint arXiv:2004.05805 (2020) +37. Que, X., Yu, Q.: Dual-level curriculum meta-learning for noisy few-shot learning tasks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 14740-14748 (2024) +38. Ren, M., Triantafillou, E., Ravi, S., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: Proceedings of the International Conference on Learning Representations (2018) +39. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: International Conference on Machine Learning. pp. 4334-4343. PMLR (2018) +40. Sahoo, D., Le, H., Liu, C., Hoi, S.C.: Meta-learning with domain adaptation for few-shot learning under domain shift (2018) +41. Shui, C., Li, Z., Li, J., Gagné, C., Ling, C.X., Wang, B.: Aggregating from multiple target-shifted sources. In: International Conference on Machine Learning. pp. 9638-9648. PMLR (2021) +42. Shui, C., Zhou, F., Gagné, C., Wang, B.: Deep active learning: Unified and principled method for query and training. In: International Conference on Artificial Intelligence and Statistics. pp. 1308-1318. PMLR (2020) +43. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 4077-4087 (2017) +44. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: Advances in neural information processing systems (2020) +45. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019) +46. Tolstikhin, I., Bousquet, O., Gelly, S., Schoelkopf, B.: Wasserstein auto-encoders. In: International Conference on Machine Learning (2017) +47. Tsipras, D., Santurkar, S., Engstrom, L., Ilyas, A., Madry, A.: From imagenet to image classification: Contextualizing progress on benchmarks. In: International Conference on Machine Learning. pp. 9625-9635. PMLR (2020) +48. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. In: Advances in neural information processing systems. pp. 3630-3638 (2016) +49. Wang, X., Hua, Y., Kodirov, E., Robertson, N.M.: Imae for noise-robust learning: Mean absolute error does not treat examples equally and gradient magnitude's variance matters. arXiv preprint arXiv:1903.12141 (2019) +50. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019) +51. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-ucsd birds 200 (2010) + +52. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018) +53. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems 33, 7597-7610 (2020) +54. Xia, X., Liu, T., Wang, N., Han, B., Gong, C., Niu, G., Sugiyama, M.: Are anchor points really indispensable in label-noise learning? Advances in neural information processing systems 32 (2019) +55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020) +56. Xu, R., Yang, X., Liu, B., Zhang, K., Liu, W.: Selecting task with optimal transport self-supervised learning for few-shot classification. arXiv preprint arXiv:2204.00289 (2022) +57. Xu, Y., Cao, P., Kong, Y., Wang, Y.: L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems 32 (2019) +58. Yang, Y., Liang, K.J., Carin, L.: Object detection as a positive-unlabeled problem. arXiv preprint arXiv:2002.04672 (2020) +59. Yu, X., Han, B., Yao, J., Niu, G., Tsang, I., Sugiyama, M.: How does disagreement help generalization against label corruption? In: International Conference on Machine Learning. pp. 7164-7173. PMLR (2019) +60. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems 31 (2018) +61. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021) \ No newline at end of file diff --git a/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/images.zip b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e32756ba017c6c9804f547e6f3400f5f6efc2bc8 --- /dev/null +++ b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bad3c3c294c58a0f788998ef9fa7845d2b4a34d9dc3c2989a1b1a9343470b4bd +size 440519 diff --git a/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/layout.json b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1573b2880083128003fe358398ce643cc58843d3 --- /dev/null +++ b/2024/Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data/layout.json @@ -0,0 +1,11303 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Optimal Transport of Diverse Unsupervised Tasks for Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 241, + 167, + 373, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 167, + 373, + 182 + ], + "spans": [ + { + "bbox": [ + 241, + 167, + 373, + 182 + ], + "type": "text", + "content": "Xiaofan Que1 and Qi Yu1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 238, + 190, + 376, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 190, + 376, + 201 + ], + "spans": [ + { + "bbox": [ + 238, + 190, + 376, + 201 + ], + "type": "text", + "content": "Rochester Institute of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 201, + 360, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 201, + 360, + 213 + ], + "spans": [ + { + "bbox": [ + 255, + 201, + 360, + 213 + ], + "type": "text", + "content": "{xq5054,qi.yu}@rit.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 232, + 455, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 232, + 455, + 462 + ], + "spans": [ + { + "bbox": [ + 160, + 232, + 455, + 462 + ], + "type": "text", + "content": "Abstract. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. While data cleansing offers a viable solution to address noisy labels in the general learning settings, it exacerbates information loss in FSL due to limited training data, resulting in inadequate model training. To best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed unsupervised auxiliary tasks to compensate for information loss. Using unsupervised tasks can effectively avoid additional annotation costs and minimize the risk of introducing additional label noises. However, a randomly constructed unsupervised task may misguide the model to learn sample-specific features that are likely to compromise the primary few-shot learning task due to the noisy weak learning signals. We propose to conduct novel auxiliary task selection to ensure the intra-diversity among the unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold. We conduct a theoretical analysis to derive novel generalization bounds for learning with auxiliary tasks. Extensive experiments are conducted to demonstrate that our method outperforms existing noisy few-shot learning methods under various in-domain and cross-domain few-shot classification benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 472, + 453, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 472, + 453, + 484 + ], + "spans": [ + { + "bbox": [ + 160, + 472, + 453, + 484 + ], + "type": "text", + "content": "Keywords: Noisy few-shot learning " + }, + { + "bbox": [ + 160, + 472, + 453, + 484 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 453, + 484 + ], + "type": "text", + "content": " Auxiliary task " + }, + { + "bbox": [ + 160, + 472, + 453, + 484 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 453, + 484 + ], + "type": "text", + "content": " Optimal transport" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 500, + 230, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 230, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 230, + 511 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Few-shot learning (FSL) aims to utilize knowledge acquired from a large amount of diverse tasks during training and apply it to new, unseen tasks from the same task distribution with only a small amount of labeled data [10,43,45]. Therefore, learning an accurate representation of the task manifold that effectively captures the underlying structure and relationships among tasks is essential to ensure good few-shot generalization. On the other hand, noisy labels commonly exist in the task distribution due to the ever-increasing demand for large-scale annotated datasets, where labeling errors can be introduced in various means that are impossible to control [4,30,34,47,58]. Given the few-shot nature of each training task, memorization of noisy labels introduced along with limited training data can exacerbate overfitting, which will disrupt the representation space and severely hinder few-shot generalization." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 200, + 114, + 272, + 171 + ], + "blocks": [ + { + "bbox": [ + 200, + 114, + 272, + 171 + ], + "lines": [ + { + "bbox": [ + 200, + 114, + 272, + 171 + ], + "spans": [ + { + "bbox": [ + 200, + 114, + 272, + 171 + ], + "type": "image", + "image_path": "2614a906af7f80077a56aeb94198149d1d9514148a3960dc8af6f6619e283466.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 284, + 117, + 328, + 169 + ], + "blocks": [ + { + "bbox": [ + 284, + 117, + 328, + 169 + ], + "lines": [ + { + "bbox": [ + 284, + 117, + 328, + 169 + ], + "spans": [ + { + "bbox": [ + 284, + 117, + 328, + 169 + ], + "type": "image", + "image_path": "39eda97b665f88caefd54193ed4dad36fb05f025e587095472ea9c90e76db570.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 358, + 118, + 399, + 169 + ], + "blocks": [ + { + "bbox": [ + 358, + 118, + 399, + 169 + ], + "lines": [ + { + "bbox": [ + 358, + 118, + 399, + 169 + ], + "spans": [ + { + "bbox": [ + 358, + 118, + 399, + 169 + ], + "type": "image", + "image_path": "0c4d380a4768522a57180e6d5788fc8d265b96db48cd0eb49eb30324fba0e2ec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 200, + 175, + 271, + 242 + ], + "blocks": [ + { + "bbox": [ + 200, + 175, + 271, + 242 + ], + "lines": [ + { + "bbox": [ + 200, + 175, + 271, + 242 + ], + "spans": [ + { + "bbox": [ + 200, + 175, + 271, + 242 + ], + "type": "image", + "image_path": "c26f4556282ddb153d152d93bde461d3b6abdcbf33c4926b7acdf643180d2c29.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 247, + 266, + 255 + ], + "lines": [ + { + "bbox": [ + 198, + 247, + 266, + 255 + ], + "spans": [ + { + "bbox": [ + 198, + 247, + 266, + 255 + ], + "type": "text", + "content": "Data with different classes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 280, + 176, + 335, + 239 + ], + "blocks": [ + { + "bbox": [ + 280, + 176, + 335, + 239 + ], + "lines": [ + { + "bbox": [ + 280, + 176, + 335, + 239 + ], + "spans": [ + { + "bbox": [ + 280, + 176, + 335, + 239 + ], + "type": "image", + "image_path": "5b44827c96832d1e1ad99fb930421e6ee5afbe6aee990458d1e9225d5f62a322.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 355 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 355 + ], + "type": "text", + "content": "Fig. 1: (a) Data manifold of the primary task distribution. (b) Data manifold after using data cleansing to remove some noisy labels, which is incomplete due to information loss. (c) Using augmentation can make up for the information loss but it is limited by the diversity of existing data. (d) Data manifold of the auxiliary task distribution. (e) Joint data manifold of primary and auxiliary data distribution, in which the domain-specific knowledge of the auxiliary manifold is also included. (f) Through optimal transport of diverse unsupervised tasks, the learned shared representation only includes rich domain-invariant information from the auxiliary manifold." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 343, + 175, + 414, + 239 + ], + "blocks": [ + { + "bbox": [ + 343, + 175, + 414, + 239 + ], + "lines": [ + { + "bbox": [ + 343, + 175, + 414, + 239 + ], + "spans": [ + { + "bbox": [ + 343, + 175, + 414, + 239 + ], + "type": "image", + "image_path": "066b9ed187915da508c3e67f51403bea56694dbe21373a297f8623774f866889.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 247, + 356, + 255 + ], + "lines": [ + { + "bbox": [ + 268, + 247, + 356, + 255 + ], + "spans": [ + { + "bbox": [ + 268, + 247, + 356, + 255 + ], + "type": "text", + "content": "Missing data Augmented data" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 247, + 413, + 255 + ], + "lines": [ + { + "bbox": [ + 359, + 247, + 413, + 255 + ], + "spans": [ + { + "bbox": [ + 359, + 247, + 413, + 255 + ], + "type": "text", + "content": "Joint data manifold" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 506 + ], + "type": "text", + "content": "While training robust models from noisy data has been intensively investigated in standard settings, learning from noisy few-shot data is only sparsely explored in existing literature [28,31]. Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. Within each training task, FSL operates with few training samples per class, making each data point highly valuable. Hence, the model tends to be more sensitive to variations in the training data and noisy labels can more easily cause the model to learn incorrect patterns, compromising its ability to accurately generalize to new tasks. Furthermore, identifying and rectifying label noises also becomes more challenging. The scarcity of data makes it harder to distinguish between true class characteristics and erroneous annotations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "Techniques, including robust optimization [28] and data augmentation [31] have been leveraged to handle noises in limited data. However, these techniques, while beneficial, do not entirely resolve the problem. Conventional methods for learning with noisy labels, such as data cleansing, data augmentations, and robust loss functions, are inadequate to tackle the challenges as outlined above. As shown in Fig. [1(a)-(b)], data cleansing introduces information loss, further diminishing the already scarce training data in each task. This poses a greater challenge for the model to learn a meaningful and generalizable representation of the task manifold. Data augmentation could alleviate the information loss introduced through data cleansing as shown in Fig. [1(c)]. Nevertheless, augmentation is performed on the limited training samples, so it may not cover a broad task manifold. Additionally, robust optimization algorithms disregard the concern of information loss, making them less suitable for learning from noisy few-shot data." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 463, + 219 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 463, + 219 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 463, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 463, + 219 + ], + "type": "image", + "image_path": "6959258b7b5d6adb712e53cd6dca8e3ea1efe8d4cf7b902563392ed28ddd249a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 285 + ], + "type": "text", + "content": "Fig. 2: In DUNT, the feature extractor extract information from both primary and auxiliary tasks, the classifier predicts the labels of the primary and auxiliary examples, and the domain critic outputs the source of the examples. The primary tasks (top) are sampled and cleansed by episode cleansing, the auxiliary tasks (bottom) are constructed and filtered by their diverseness. (Best viewed in color.)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 445 + ], + "type": "text", + "content": "To tackle the unique challenges arising from NFSL, we propose novel extensions to existing techniques designed for standard noisy learning scenarios to more effectively handle the information loss in the few-shot setting. In particular, to best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed auxiliary tasks to compensate information loss. The auxiliary tasks are entirely unsupervised as using supervised tasks can be counterproductive in the NFSL setting for three main reasons: (1) They may incur additional annotation overhead. (2) Noisy labels may continue to be introduced that further compromise the manifold structure. (3) Even clean labels can be guaranteed, these labels may not totally align with the primary tasks that can potentially overwhelm that weak learning signals carried by the limited training samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 592 + ], + "type": "text", + "content": "While unsupervised auxiliary tasks have been leveraged in diverse learning scenarios [19,24,32,36,52], including the few-shot setting, the noisy labels and potential information loss introduced through necessary data cleansing makes NFSL much more vulnerable to the potential noises introduced by the auxiliary data. Since data samples are unlabeled, they are usually randomly sampled into an auxiliary task. When a contrastive loss is used to differentiate unlabeled data samples from the same underlying class, the model may be forced to learn sample-specific (spurious) features that cause overfitting. To this end, we propose to conduct novel auxiliary task selection to ensure the intra-diversity among the chosen unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold as shown in Fig. 1(d)-(g)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Fig. 2 shows the proposed framework, referred to as Diverse UNsupervised Tasks for NFSL (DUNT4NFSL or DUNT for short). The feature extractor of DUNT undergoes updates through three distinct loss components: 1) A cross-entropy loss, integral for extracting discriminative and effective features for classification from the primary tasks; 2) A contrastive loss tailored for the unsupervised auxiliary tasks, designed to encourage an embedding space where similar sample" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "pairs remain close while dissimilar pairs are distanced; 3) An adversarial loss, implemented to align the unsupervised auxiliary tasks with the primary tasks, employing a domain critic head to reduce the covariate shift across multiple domains. To ensure that the feature representation remains unaffected by label noise, we apply an episode cleaning technique to the primary task, removing noisy examples in each task. For the auxiliary tasks, we introduce a novel criterion for diverseness to guarantee that the unsupervised auxiliary tasks contribute informative and valuable insights. Our contributions are threefolds:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 224, + 481, + 332 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 224, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 258 + ], + "type": "text", + "content": "- We present DUNT, a framework designed to adeptly utilize unsupervised data from a distinct domain in order to counteract the information loss resulting from data cleansing in the challenging scenario of NFSL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 261, + 481, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 296 + ], + "type": "text", + "content": "- We introduce a novel strategy to perform diverse auxiliary task selection to avoid learning sample-specific spurious features from the unlabeled auxiliary data samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 297, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 481, + 332 + ], + "type": "text", + "content": "- We conduct an in-depth theoretical analysis of the auxiliary tasks introduced in the DUNT framework, establishing novel generalization bounds, offering valuable insights into the contributions of these auxiliary tasks." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 343, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 381 + ], + "type": "text", + "content": "We have conducted comprehensive experiments on five benchmark datasets with synthetic label noises to demonstrate the effectiveness and robustness of DUNT and its advantage over SOTA methods by a large margin." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 403, + 237, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 237, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 237, + 415 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 433, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 613 + ], + "type": "text", + "content": "Noisy few-shot learning. Existing methods for training DNNs with noisy labels usually fall into the following categories: robust loss [12,49,50,57,60] designs robust cost functions in the presence of noisy labels; noise transition estimation [5,20,53,54] aims to estimate the probabilities that clean labels flip into noisy labels; loss correction [17,22,23,39,44,55,59] either explicitly or implicitly relabels or reweights the noisy samples, a.k.a. data cleansing. Noisy label in FSL is an important yet relatively under-explored topic. The scarcity of data in FSL not only heightens its susceptibility to label noise but also adds complexity to the challenge of identifying and correcting such noise. A few previous attempts such as [28,29,31] address mislabeled issue in FSL settings using re-weighting strategy or generating robust prototypes by feature hybrid and augmentation techniques. Other works such as RW-MAML [25], AQ [14], DFSL [27] tackle the out-of-distribution tasks or adversarial attacks in few-shot learning. In this work, we particularly focus on solving the limited data problem by utilizing auxiliary unlabeled data as unsupervised tasks to enhance the robustness of FSL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Optimal transport and Wasserstein distance. The optimal transport (OT) cost is used to measure the difference between distributions supported on high-dimensional space using SGD [1,46]. For a constant " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\xi \\geq 1" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "-Wasserstein" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 375, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 375, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 375, + 128 + ], + "type": "text", + "content": "metric between distributions " + }, + { + "bbox": [ + 132, + 116, + 375, + 128 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_X" + }, + { + "bbox": [ + 132, + 116, + 375, + 128 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 375, + 128 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{X'}" + }, + { + "bbox": [ + 132, + 116, + 375, + 128 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 137, + 481, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 137, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 198, + 137, + 481, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {\\xi} \\left(\\mathbb {P} _ {X}, \\mathbb {P} _ {X ^ {\\prime}}\\right) = \\left(\\inf _ {\\gamma \\in \\Pi \\left(\\mathbb {P} _ {X}, \\mathbb {P} _ {X ^ {\\prime}}\\right)} \\mathbb {E} _ {\\gamma \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right)} \\left[ c ^ {\\xi} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right) \\right]\\right) ^ {\\frac {1}{\\xi}} \\tag {1}", + "image_path": "09d0209176cdde766f46bde050c09572ccc42a30400150881a97b218050ed217.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{x}^{\\prime}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " denote the random variables in which the distributions " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_X" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{X^{\\prime}}" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " are defined. " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\prod (\\mathbb{P}_X,\\mathbb{P}_{X'})" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " is the set of all joint distributions (i.e. couplings) whose marginals are " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_X" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{X'}" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": ", that is, " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\int \\gamma (\\mathbf{x},\\mathbf{x}')d\\mathbf{x} = \\mathbb{P}_{X'}" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\int \\gamma (\\mathbf{x},\\mathbf{x}')d\\mathbf{x}' = \\mathbb{P}_X" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "c(\\mathbf{x},\\mathbf{x}'): \\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}^+" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " is the cost function for moving from " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x},\\mathbf{x}')" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " indicates how much \"mass\" must be transported from " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " in order to transform the distributions " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_X" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{X'}" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": ". The Wasserstein metric then is the \"cost\" of the optimal transport plan [1]. Eq. [1] is the primal form of the Wasserstein metric, particularly for the case of " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\xi = 1" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "inline_equation", + "content": "c(\\mathbf{x},\\mathbf{x}') = \\| \\mathbf{x} - \\mathbf{x}'\\|" + }, + { + "bbox": [ + 132, + 175, + 482, + 368 + ], + "type": "text", + "content": " [1]. Some existing efforts propose domain adaptation algorithms using ideas from OT theory, which aims to reduce the divergence between two domains by minimizing the Wasserstein distance between their distributions [7,8,11]. Works that use domain adaptation techniques in FSL [6,18,33,40,56] share the same goal that aims to learn transferable knowledge. In this work, we achieve this goal by leveraging extra data first and then minimize the Wasserstein distance between the primary and auxiliary distributions to ensure the learned knowledge are domain-invariant and discriminative to future unseen tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 386, + 233, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 233, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 233, + 401 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": "Problem formulation. In FSL, the goal is to learn a mapping " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " that maps an observation " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " to its label " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " is the feature space and " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " is the label space. To train the model with the quick adaptation ability to the new tasks with only a small number of examples, a set of tasks " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i = \\{\\mathcal{S}_i^{\\mathrm{sup}},\\mathcal{S}_i^{\\mathrm{que}}\\}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " is sampled from the data distribution " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\times \\mathcal{Y}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "S_{i}^{\\mathrm{sup}} = \\{(x_{i}^{k},y_{i}^{k})\\}_{k = 1}^{NK}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " denotes the support set and " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "S_{i}^{\\mathrm{que}} = \\{(x_{i}^{k},y_{i}^{k})\\}_{k = 1}^{NQ}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " denotes the query set. " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " is the number of examples in each class of the support set, and " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " is that of the query set. In DUNT, the meta-training tasks are sampled from the two domains: the primary tasks are sampled from the noisy primary distribution " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}^{\\mathrm{pri}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\times \\tilde{\\mathcal{Y}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": ", where the label space " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " is considered noisy, and the unsupervised auxiliary tasks are sampled from the auxiliary distribution " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{X}'" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": ", where the entire label space is missing. We assume that the primary and the auxiliary distributions are from different domains: " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}^{\\mathrm{pri}} \\neq \\mathcal{D}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": ". During meta-testing, the tasks are sampled from a new unseen target domain " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{\\mathrm{tgt}} \\times \\mathcal{Y}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": ", the label is only used for evaluation purpose. The goal is to learn a meta-model " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " trained on " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}^{\\mathrm{pri}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " that can quickly adapt to tasks in " + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 132, + 412, + 482, + 617 + ], + "type": "text", + "content": " using only a few examples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "content": "Framework overview. The DUNT framework (see Fig. 2) is composed of a shared feature extractor " + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "content": ", the task classifier " + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta_i^c" + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "content": " for task " + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "content": " and the domain critic " + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta^d" + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "content": ". The feature extractor takes input from both the primary tasks and the auxiliary" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "tasks; the classifiers predict the labels for the input and the critic predict the source of the input. The gradient reversal layer (GRL) ensures that the feature distributions over the two domains are made similar." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 330, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 330, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 330, + 185 + ], + "type": "text", + "content": "3.1 Learning from the Primary Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "content": "One commonly used FSL strategy that takes advantage of task-based formulation is to learn a meta-model through a bi-level optimization. By defining the tasks-specific model parameter for task " + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "inline_equation", + "content": "\\theta_i'" + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "content": ", the meta model " + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "content": " are trained by optimizing for the performance of " + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "inline_equation", + "content": "\\theta_i'" + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "content": " across tasks sampled from " + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{pri}}" + }, + { + "bbox": [ + 130, + 194, + 482, + 254 + ], + "type": "text", + "content": ". The loss is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 229, + 263, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 263, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 229, + 263, + 481, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {p r i}} = \\sum_ {\\mathcal {T} _ {i} \\left(\\mathcal {S} _ {i} ^ {\\sup }, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}\\right) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}} \\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(\\theta_ {i} ^ {\\prime}, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}\\right) \\tag {2}", + "image_path": "4da6efd5f3182264fd41b760b0a8f7a507949283125ad9d8eaf10cb656a45aed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\theta_i' \\gets \\theta - \\alpha \\nabla_\\theta \\mathcal{L}_{\\mathcal{T}_i}(\\theta, \\mathcal{S}_i^{\\mathrm{sup}})" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": " is the task-specific model adapted from " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": " within one gradient step. This is a dual-level optimization where in the outer-level, the meta-model " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": " is updated with the summation of a series of tasks evaluated on the query sets " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{S}_i^{\\mathrm{que}}\\}" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{T}_i}(\\theta_i', \\mathcal{S}_i^{\\mathrm{que}}) = \\sum_{(\\mathbf{x}_i^k, \\mathbf{y}_i^k) \\in \\mathcal{S}_i^{\\mathrm{que}}} \\ell(\\theta_i^c(\\theta_i'(\\mathbf{x}_i^k)), \\mathbf{y}_i^k)" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": ". The inner-level objective is computed using the task-specific model " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\theta_i'" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": " updated with support set " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_i^{\\mathrm{sup}}" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": " using loss " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{T}_i}(\\theta, \\mathcal{S}_i^{\\mathrm{sup}}) = \\sum_{(\\mathbf{x}_i^k, \\mathbf{y}_i^k) \\in \\mathcal{S}_i^{\\mathrm{sup}}} \\ell(\\theta_i^c(\\theta(\\mathbf{x}_i^k)), \\mathbf{y}_i^k)" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": " is the cross-entropy loss, " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\theta_i^c" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 300, + 482, + 389 + ], + "type": "text", + "content": "-th task classifier." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "text", + "content": "Episode cleansing. In the dual-level optimization, the meta-model " + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "text", + "content": " is updated by first adapting to different task-specific models " + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{\\prime}" + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "text", + "content": " with the corresponding support set and then updated using the summation of the loss of the task-specific models " + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{\\prime}" + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "text", + "content": " on their corresponding query sets. Since the dataset is noisy, the error may be accumulated through the noisy query set and cause the overfitting of the meta-model. Therefore, we propose episode cleansing to remove the noisy data in the query sets based on the memorization effect which indicates that noisy examples have higher losses than the clean ones. The classification loss " + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{T}_i}(\\theta_i', S_i^{\\mathrm{que}})" + }, + { + "bbox": [ + 131, + 403, + 483, + 513 + ], + "type": "text", + "content": " becomes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 521, + 481, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 521, + 481, + 549 + ], + "spans": [ + { + "bbox": [ + 236, + 521, + 481, + 549 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(f _ {\\theta_ {i} ^ {\\prime}}, \\hat {\\mathcal {S}} _ {i} ^ {\\mathrm {q u e}}\\right) = \\sum_ {\\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {i} ^ {k} \\in \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}} w _ {i} ^ {k} \\ell_ {i} ^ {k}, \\tag {3}", + "image_path": "af0a258fcf452eb3f19c501536abd0d4a5fc9e6992c01dc78d56363f09643c3c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\hat{S}_i^{\\mathrm{que}}" + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "text", + "content": " is the cleansed (weighted) query set for task " + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "text", + "content": "; for simplification, we use " + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\ell_i^k" + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "text", + "content": " to denote " + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\ell (\\theta_i^c (\\theta_i'(x_i^k)),y_i^k)" + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "inline_equation", + "content": "w_{i}^{k}" + }, + { + "bbox": [ + 131, + 560, + 482, + 597 + ], + "type": "text", + "content": " is the self-paced sample weight for each example pair, and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 268, + 597, + 347, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 597, + 347, + 612 + ], + "spans": [ + { + "bbox": [ + 268, + 597, + 347, + 612 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {k} = \\mathbb {1} (\\ell_ {i} ^ {k} < \\gamma_ {t h}),", + "image_path": "2fbc439826c34f38dafc952e36e4c2b2850d41f14e40dec1c1b3662b875adecb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " is an indicator function whose value is 1 when " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\ell_i^k < \\gamma_{th}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " and 0 otherwise; and " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\gamma_{th}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " is a predefined hyperparameter used to filter out high-loss examples. We use the percentage of the ordered loss to avoid extensive hyperparameter search, whose ablation is provided in Appendix C." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 146, + 116, + 284, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 116, + 284, + 128 + ], + "spans": [ + { + "bbox": [ + 146, + 116, + 284, + 128 + ], + "type": "text", + "content": "The primary loss the becomes:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 137, + 481, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 137, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 227, + 137, + 481, + 166 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {p r i}} = \\sum_ {\\mathcal {T} _ {i} (\\mathcal {S} _ {i} ^ {\\sup }, \\mathcal {S} _ {i} ^ {\\mathrm {q u e}}) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}} \\mathcal {L} _ {\\mathcal {T} _ {i}} \\left(\\theta_ {i} ^ {\\prime}, \\hat {\\mathcal {S}} _ {i} ^ {\\mathrm {q u e}}\\right). \\tag {4}", + "image_path": "2e33e44dd0c90906ebb1b78184c2a7ab9e4d8bf70fd86b8987acec143c474834.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 173, + 480, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 173, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 131, + 173, + 480, + 198 + ], + "type": "text", + "content": "Similarly, for the support set label noise, we use meta-model to compute the loss values and remove the noisy ones in the support set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 378, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 378, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 378, + 226 + ], + "type": "text", + "content": "3.2 Learning from Diverse Unsupervised Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": "As shown in Fig. 2, the general unsupervised auxiliary task " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_j" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": " is constructed by i) first randomly sampling " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": " images from the dataset for the support set " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_j^{\\mathrm{sup}} = \\{\\mathbf{x}_j^k\\}_{k=1}^N" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": "; ii) the query set " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_j^{\\mathrm{que}} = \\{A_1(\\mathbf{x}_j^k), \\dots, A_Q(\\mathbf{x}_j^k)\\}_{k=1}^N" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": " is obtained by augmenting the " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": " images in " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": " different ways, where " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "A_1(\\cdot), \\dots, A_Q(\\cdot)" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 233, + 482, + 317 + ], + "type": "text", + "content": " different augmentation techniques, such as random cropping, translation, flipping. The auxiliary loss is the summation of the contrastive loss of a series of auxiliary tasks:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 327, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 327, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 214, + 327, + 481, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {a u x}} = \\sum_ {\\mathcal {T} _ {j} \\left(\\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right) \\sim \\mathcal {D} ^ {\\text {a u x}}} \\mathcal {L} _ {\\mathcal {T} _ {j}} \\left(\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right). \\tag {5}", + "image_path": "71f630141c4c542e7ad49dbc220d636f73b8a07bb8f95f461ef8307763c2092b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 362, + 388, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 388, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 388, + 373 + ], + "type": "text", + "content": "The contrastive loss is utilized for the unsupervised tasks:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 148, + 384, + 481, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 384, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 148, + 384, + 481, + 417 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {T} _ {j}} \\left(\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\text {q u e}}\\right) = - \\frac {1}{N Q} \\sum_ {k = 1} ^ {N} \\sum_ {q = 1} ^ {Q} \\log \\frac {\\exp \\left(- d \\left[ \\theta \\left(A _ {q} \\left(\\mathbf {x} _ {j} ^ {k}\\right)\\right) , \\theta \\left(\\mathbf {x} _ {j} ^ {k}\\right) \\right]\\right)}{\\sum_ {l = 1} ^ {N} \\exp \\left(- d \\left[ \\theta \\left(A _ {q} \\left(\\mathbf {x} _ {j} ^ {k}\\right)\\right) , \\theta \\left(\\mathbf {x} _ {j} ^ {l}\\right) \\right]\\right)} \\tag {6}", + "image_path": "aa442a06c8882de68976c257e8f93e2591b3b2964289fbfb1ab257cdeabcc4d7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 425, + 398, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 398, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 398, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 425, + 398, + 437 + ], + "type": "inline_equation", + "content": "d[\\cdot, \\cdot]" + }, + { + "bbox": [ + 132, + 425, + 398, + 437 + ], + "type": "text", + "content": " is a distance metric, such as Euclidean distance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 449, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 482, + 639 + ], + "type": "text", + "content": "Auxiliary task selection. In contrastive learning, the positive pairs are formulated by the original example and its augmentation, and the negative pairs are formulated by images from two different classes. The contrastive loss is designed to pull the positive pairs close to each other and push the negative ones away. However, since each image in the unsupervised task are randomly sampled and treated as a class, there are chances that the images are sampled from the same class. In some extreme case, all the images are sampled from the same class, which means that the positive pairs being treated as negative pairs and being push away from each other. This would cause the model being forced to overfit to the unnecessary details and creates spurious correlations and harm the generalization ability. We denote those tasks as confusing tasks, as a comparison, the tasks with images sampled from different classes are denoted as valuable tasks. The intricate nature of confusing tasks appears to pose a significant threat to the resilience and robustness of FSL, as shown in our later ablation studies. In this work, we propose a novel task selection metric to remove these confusing unsupervised tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": "Given two images " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^k" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^l" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": ", the model parameterized by " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": " lead to class probabilities vectors " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^k = \\mathrm{softmax}(\\theta (\\mathbf{x}_j^k))" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^l = \\mathrm{softmax}(\\theta (\\mathbf{x}_j^l))" + }, + { + "bbox": [ + 131, + 639, + 482, + 667 + ], + "type": "text", + "content": ". If the two images" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": "are from the same class, " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^k" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^l" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": " are naturally encouraged to be close to the assignment vector " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_y" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\{0,1\\}^N" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": " with a single non-zero entry at position " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": " is the class label associated to " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 116, + 482, + 166 + ], + "type": "text", + "content": " is the number of classes. Therefore, we consider the following diversity criterion:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 175, + 390, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 175, + 390, + 208 + ], + "spans": [ + { + "bbox": [ + 223, + 175, + 390, + 208 + ], + "type": "interline_equation", + "content": "d i v _ {j} = \\frac {1}{C (N , 2)} \\sum_ {i = 1} ^ {N - 1} \\sum_ {j = i + 1} ^ {N} \\frac {1}{\\cos (\\mathbf {p} _ {j} ^ {k} , \\mathbf {p} _ {j} ^ {l})}.", + "image_path": "f93e29dcce0ccf6e6c6eeae153b822d2515015b8cf94260e0931a297f9af4bbb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "text", + "content": "The diverseness of the task " + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_j" + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "text", + "content": " is defined as the average of the reciprocal of the cosine similarity between the class probabilities of all example pairs in the task, where " + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "C(N,2) = \\frac{N!}{(N - 2)!2!}" + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "text", + "content": " is the 2-combination of set " + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "text", + "content": ". According to the definition, the higher the value of " + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "div_j" + }, + { + "bbox": [ + 132, + 217, + 482, + 266 + ], + "type": "text", + "content": ", the more diverse the task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 277, + 386, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 386, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 386, + 290 + ], + "type": "text", + "content": "Auxiliary loss function. Auxiliary loss " + }, + { + "bbox": [ + 132, + 277, + 386, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 132, + 277, + 386, + 290 + ], + "type": "text", + "content": " is updated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 300, + 403, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 300, + 403, + 328 + ], + "spans": [ + { + "bbox": [ + 211, + 300, + 403, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {a u x}} = \\sum_ {\\mathcal {T} _ {j} (\\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\mathrm {q u e}}) \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} v _ {j} \\mathcal {L} _ {\\mathcal {T} _ {j}} (\\theta , \\mathcal {S} _ {j} ^ {\\sup }, \\mathcal {S} _ {j} ^ {\\mathrm {q u e}})", + "image_path": "27d25db6db14b4925f6d918e83046cd752c24d17325aeaa4bed01ff0fc2d7032.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 336, + 323, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 323, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 323, + 349 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 336, + 323, + 349 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 132, + 336, + 323, + 349 + ], + "type": "text", + "content": " is the diversity weight defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 263, + 359, + 351, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 359, + 351, + 372 + ], + "spans": [ + { + "bbox": [ + 263, + 359, + 351, + 372 + ], + "type": "interline_equation", + "content": "v _ {j} = \\mathbb {1} (d i v _ {j} > \\gamma_ {d i v}),", + "image_path": "6ef6d1041f293003d5669a98e994e9454d8f52056b641a45f12f8407d1cfe9d4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 482, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 380, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\gamma_{div}" + }, + { + "bbox": [ + 132, + 380, + 482, + 441 + ], + "type": "text", + "content": " is a predefined hyperparameter, which filter out those less diverse, a.k.a., confusing tasks. In practice, we simplify the hyperparameter searching by first upsampling six auxiliary tasks, computes each task diversity metric, and only keep the most diverse task in " + }, + { + "bbox": [ + 132, + 380, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_j\\}^{\\mathrm{val}}" + }, + { + "bbox": [ + 132, + 380, + 482, + 441 + ], + "type": "text", + "content": " for training. The corresponding ablations about the numbers of extra tasks sampled are presented in Sec. 4.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 459, + 369, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 369, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 369, + 470 + ], + "type": "text", + "content": "3.3 Adversarial Training and the Critic Head" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "text", + "content": "Since the auxiliary tasks are drawn from a different data distribution, it is inevitable that there exist some domain shift from the primary domain to the auxiliary one. While learning from the domain-specific knowledge from the auxiliary domain main cause the learned model biased and deteriorate the model's generalization performance, learning from the domain-invariant knowledge helps improve the generalization ability. To enforce the learning of the domain-invariant representation of the meta-model, we propose a critic model " + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\theta^c" + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "text", + "content": " to discriminate the source of the input data, either primary or auxiliary. By training the model in an adversarial way, the critic model " + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\theta^c" + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "text", + "content": " try to distinguish which domain that the input features come from, while the meta-model " + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "text", + "content": " aims to confuse the critic. In this way, when the training process converge, " + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 479, + 482, + 635 + ], + "type": "text", + "content": " tends to extract domain-invariant knowledge from both domains. Formally, the overall adversarial objective is a min-max problem as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 187, + 643, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 643, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 187, + 643, + 480, + 662 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , \\theta^ {c}} \\max _ {\\theta^ {d}} \\alpha_ {1} \\mathcal {L} ^ {\\text {p r i}} + (1 - \\alpha_ {1}) \\mathcal {L} ^ {\\text {a u x}} + \\alpha_ {2} \\mathcal {E} (\\theta , \\theta^ {d}, \\tilde {\\mathcal {D}} ^ {\\text {p r i}}, \\mathcal {D} ^ {\\text {a u x}}), \\tag {7}", + "image_path": "1c104891ca1a85def869a2c698865883e9b9e9c5e096cdb29546cb088ba58c6d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": "where the critic model " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\theta^d" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": " assigns 1 to the example from the primary domain and 0 otherwise based on the feature representation extracted by " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": " is the adversarial loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 160, + 422, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 160, + 422, + 175 + ], + "spans": [ + { + "bbox": [ + 192, + 160, + 422, + 175 + ], + "type": "interline_equation", + "content": "\\mathcal {E} (\\theta , \\theta^ {d}, \\hat {\\mathcal {D}} ^ {\\mathrm {p r i}}, \\mathcal {D} ^ {\\mathrm {a u x}}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\hat {\\mathcal {D}} ^ {\\mathrm {p r i}}} [ g (\\mathbf {x}) ] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} [ g (\\mathbf {x}) ],", + "image_path": "48635e1ee5b63803705d4a5fcd593ab4c3c548d9e797d190d2e40a241d507001.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 184, + 383, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 184, + 383, + 198 + ], + "spans": [ + { + "bbox": [ + 131, + 184, + 383, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 184, + 383, + 198 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x}) = \\theta^d (\\theta (\\mathbf{x}))" + }, + { + "bbox": [ + 131, + 184, + 383, + 198 + ], + "type": "text", + "content": ", with its empirical counterpart as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 205, + 443, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 205, + 443, + 235 + ], + "spans": [ + { + "bbox": [ + 171, + 205, + 443, + 235 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {E}} (g _ {\\theta , \\theta^ {d}}, \\tilde {\\mathcal {D}} ^ {\\mathrm {p r i}}, \\mathcal {D} ^ {\\mathrm {a u x}}) = \\frac {1}{| \\mathcal {D} ^ {\\mathrm {p r i}} |} \\sum_ {x \\sim \\mathcal {D} ^ {\\mathrm {p r i}}} [ g (\\mathbf {x}) ] - \\frac {1}{| \\mathcal {D} ^ {\\mathrm {a u x}} |} \\sum_ {\\mathbf {x} \\sim \\mathcal {D} ^ {\\mathrm {a u x}}} [ g (\\mathbf {x}) ]", + "image_path": "170a215cc9f49a6e4c9dab2b65536b4f0d9a815c28fbd52e1ea155836730b6ff.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": "The optimization of Eq. (7) is conducted through an alternative training between the optimization of the meta-model feature extractor " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": ", classifier parameter " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta^c" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": " and the critic model parameter " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta^d" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": " while keeping the other fixed. Parameters of " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta^c" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta^d" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": " are updated alternatively. In practice, " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta^c" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": " are connected by the GRL, which reverses the gradients that flow through " + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\theta^c" + }, + { + "bbox": [ + 131, + 243, + 482, + 316 + ], + "type": "text", + "content": ". The pseudo code of the algorithm can be found in Appendix A." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 333, + 264, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 264, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 264, + 345 + ], + "type": "text", + "content": "3.4 Theoretical Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "text", + "content": "We conduct a theoretical analysis of DUNT to show: 1) the equivalence between the adversarial regularizer and the Wasserstein distance, which captures the diversity property of the auxiliary distribution while enforcing the learning of a domain invariant representation; 2) the relationship between the primary and auxiliary distributions; 3) a tight generalization bound of the target based on the expected risk of the primary and auxiliary distributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 444, + 492, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 492, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 492, + 470 + ], + "type": "text", + "content": "Equivalence to Wasserstein distance. The Kantorovich-Rubinstein of Wasserstein distance takes the form [1]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 206, + 479, + 480, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 479, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 206, + 479, + 480, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {W} (\\mathbb {P} _ {1}, \\mathbb {P} _ {2}) = \\sup _ {\\| h \\| _ {L} \\leq 1} \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {1}} [ h (\\mathbf {x}) ] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {2}} [ h (\\mathbf {x}) ] \\tag {8}", + "image_path": "349f38b77b39188b387f8a5b536c4c10987bc15ebdd760cb029d742fcfa7eeb0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 507, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 507, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 507, + 482, + 544 + ], + "type": "text", + "content": "where the supremum is over all the 1-Lipschitz functions " + }, + { + "bbox": [ + 131, + 507, + 482, + 544 + ], + "type": "inline_equation", + "content": "h: \\mathcal{X} \\to \\mathbb{R}" + }, + { + "bbox": [ + 131, + 507, + 482, + 544 + ], + "type": "text", + "content": ". Therefore, if we have a parameterized family of functions " + }, + { + "bbox": [ + 131, + 507, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\{h_w\\}_{w \\in \\mathcal{W}}" + }, + { + "bbox": [ + 131, + 507, + 482, + 544 + ], + "type": "text", + "content": " that are all 1-Lipschitz, we could consider solving the problem" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 233, + 554, + 480, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 554, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 233, + 554, + 480, + 567 + ], + "type": "interline_equation", + "content": "\\max \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {1}} \\left[ h _ {w} (\\mathbf {x}) \\right] - \\mathbb {E} _ {\\mathbf {x} \\sim \\mathbb {P} _ {2}} \\left[ h _ {w} (\\mathbf {x}) \\right] \\tag {9}", + "image_path": "ac422fd777b7bf040a0f8bf9faa36f8897d11ccb04f6a68ad281c597d16e2a93.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 575, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 482, + 625 + ], + "type": "text", + "content": "which has the same form of the regularizer in Sec. 3.3. This indicates that our adversarial regularizer has the same theoretical foundation of Wasserstein distance optimal transport. To satisfy the 1-Lipschitz constraint, the critic " + }, + { + "bbox": [ + 131, + 575, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\theta^d" + }, + { + "bbox": [ + 131, + 575, + 482, + 625 + ], + "type": "text", + "content": " is implemented with gradient penalty [15]. The overall objective can be reformulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 204, + 644, + 480, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 644, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 204, + 644, + 480, + 663 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\alpha_ {1} \\mathcal {L} ^ {\\text {p r i}} + (1 - \\alpha_ {1}) \\mathcal {L} ^ {\\text {a u x}} + \\alpha_ {2} \\mathcal {W} _ {1} (\\mathcal {D} ^ {\\text {p r i}}, \\mathcal {D} ^ {\\text {a u x}}), \\tag {10}", + "image_path": "9b9c480fcc7df3d2e83f2b9ffcd2053e327d60b317bc4192fc1eac00177d076f.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 448, + 102 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 190 + ], + "type": "text", + "content": "where the first and second term are the primary and auxiliary loss, resp., and the third term is the Wasserstein distance based regularizer, which aims to extract domain-invariant feature from the primary and auxiliary domains. Wasserstein distance reflects an optimal transport cost for moving one distribution to another. A smaller transport cost means a better coverage of the distribution, which captures the property of diversity [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "text", + "content": "Relationship of the primary and auxiliary distributions. By viewing this problem as a distribution matching that matches the auxiliary distribution with the primary one, we have the generalization property that quantify the relationship between the primary and the auxiliary distribution. Given the primary distribution " + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{pri}}" + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "text", + "content": " and the auxiliary distribution " + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "text", + "content": ", the expected risk w.r.t. is defined as " + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{pri}}R_{\\mathcal{D}^{\\mathrm{pri}}} = \\mathbb{E}_{x\\sim \\mathcal{D}^{\\mathrm{pri}}}\\ell (h(x),h^{*}(x))" + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "text", + "content": ", and the expected risk w.r.t. " + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "inline_equation", + "content": "R_{\\mathcal{D}^{\\mathrm{aux}}} = \\mathbb{E}_{x\\sim \\mathcal{D}^{\\mathrm{aux}}}\\ell (h(x),h^{*}(x))" + }, + { + "bbox": [ + 130, + 205, + 482, + 291 + ], + "type": "text", + "content": ". We have the following Lemma 1:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": "Lemma 1. If the loss " + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": " is symmetric and " + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": "-Lipschitz, " + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\forall h \\in \\mathcal{H}" + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": " is at most " + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": "-Lipschitz function and the underlying mapping function " + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "inline_equation", + "content": "h^*" + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\phi(\\lambda) - (\\mathcal{D}^{pri}, \\mathcal{D}^{aux})" + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": " joint probabilistic Lipschitz; then the expected risk w.r.t. " + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{pri}" + }, + { + "bbox": [ + 130, + 295, + 482, + 331 + ], + "type": "text", + "content": " is upper bounded" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 339, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 339, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 194, + 339, + 481, + 354 + ], + "type": "interline_equation", + "content": "R _ {\\mathcal {D} ^ {p r i}} \\leq R _ {\\mathcal {D} ^ {a u x}} + L (H + \\lambda) \\mathcal {W} _ {1} (\\mathcal {D} ^ {p r i}, \\mathcal {D} ^ {a u x}) + L \\phi (\\lambda), \\tag {11}", + "image_path": "9cdeb4471d2f464112e4ac0fe1249e3a04f269afa4e81884d0e7d0e257fe8350.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\phi : \\mathbb{R} \\to [0,1]" + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "content": " and a mapping " + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "inline_equation", + "content": "h^*" + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\phi(\\lambda) - (\\mathcal{D}^{pri},\\mathcal{D}^{aux})" + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "content": " joint probabilistic Lipschitz if " + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\text{supp}(\\mathcal{D}^{aux}) \\subseteq \\text{supp}(\\mathcal{D}^{pri})" + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\forall \\lambda > 0" + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "content": ", all coupling " + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\prod(\\mathcal{D}^{pri},\\mathcal{D}^{aux}) : \\mathbb{P}_{(x_{\\mathcal{D}^{pri}},x_{\\mathcal{D}^{aux}})\\sim \\gamma}[|h^{*}(x_{\\mathcal{D}^{pri}}) - h^{*}(x_{\\mathcal{D}^{aux}})| > \\lambda \\| x_{\\mathcal{D}^{pri}} - x_{\\mathcal{D}^{aux}}\\|_2] \\leq \\phi(\\lambda)" + }, + { + "bbox": [ + 131, + 361, + 482, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "Lemma 1 shows that the expected risk of the primary distributions is upper bounded by the expected risk w.r.t. the auxiliary distribution, the Wasserstein distance, and the labeling function property. The proof of Lemma 1 can be derived with Theorem 1 in [42]. Eq. (10) follows the upper bound from Lemma 1, which means a desirable auxiliary distribution with a better matching with the primary distribution which holds a small expected risk leads to tighter generalization bound for the primary distribution. This further verifies the rationality of our design of using the WD based regularizer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 516, + 481, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 481, + 541 + ], + "type": "text", + "content": "Generalization bound of DUNT. Furthermore, in the following Lemma 2 we derive the upper bound for the target risk " + }, + { + "bbox": [ + 130, + 516, + 481, + 541 + ], + "type": "inline_equation", + "content": "R_{\\mathcal{D}^{\\mathrm{tgt}}} = \\mathbb{E}_{(\\mathbf{x},y)\\sim \\mathcal{D}^{\\mathrm{tgt}}}\\ell (\\mathbf{x},y)" + }, + { + "bbox": [ + 130, + 516, + 481, + 541 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "content": "Lemma 2. Let " + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{tgt}" + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "content": " be the target distribution, " + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "content": "-Lipschitz w.r.t. the feature " + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 548, + 482, + 585 + ], + "type": "text", + "content": ", and the setting of loss and hypothesis be the same as Lemma 1, the expected risk w.r.t. the target distribution is upper-bounded by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 592, + 480, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 592, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 205, + 592, + 480, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R _ {\\mathcal {D} ^ {t g t}} \\leq \\alpha_ {1} \\left(R _ {\\mathcal {D} ^ {p r i}} + L K \\mathcal {W} _ {1} \\left(\\mathcal {D} ^ {p r i}, \\mathcal {D} ^ {t g t}\\right)\\right) \\tag {12} \\\\ + (1 - \\alpha_ {1}) \\left(R _ {\\mathcal {D} ^ {a u x}} + L K \\mathcal {W} _ {1} \\left(\\mathcal {D} ^ {a u x}, \\mathcal {D} ^ {t g t}\\right)\\right), \\\\ \\end{array}", + "image_path": "0e3c129fad9c8488b44ecef8cddda0ed31809c74878ee9362aa32773501253ab.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Lemma 2 shows that the expected risk w.r.t. the target distribution is upper bounded by the expected risk w.r.t. the primary distribution and the auxiliary distribution, and the corresponding Wasserstein distances. The proof of Lemma 2" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "can be derived with Theorem 2 in [41]. Lemma 2 provide solid theoretical foundation for the design of DUNT with a tight generalization bound to the unseen target domain with primary and auxiliary distributions, which can be extended to multiple sources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 183, + 230, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 183, + 230, + 197 + ], + "spans": [ + { + "bbox": [ + 131, + 183, + 230, + 197 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 209, + 398, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 209, + 398, + 221 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 398, + 221 + ], + "type": "text", + "content": "4.1 Datasets, Baselines, and Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 229, + 481, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 481, + 278 + ], + "type": "text", + "content": "Datasets. We conduct experiments using the following benchmark datasets for FSL: CUB 51, miniImageNet (mini) 48, tieredImageNet (tiered) 38, CIFAR-FS (CF) 3, and Cross-Domain dataset 16. Each datasets contains synthetic labels noise with different ratios. More details are in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": "Baselines. We use ProtoNet [43] as the base few-shot learning method and consider comparing with improving methods with general data cleansing methods including SPL [26] and FSR [61], denoted as " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "PN + SPL" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "PN + FSR" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": ", respectively. We also consider to improve ProtoNet in transductive settings including CoTeaching [17], CoTeaching+ [59], denoted as " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "PN + Co" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "PN + Co+" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": ", respectively, STARTUP [35], and DDNet [21]. Last but not least, we consider to compare with noisy FSL methods RNNP [31], TraNFS [28] and DCML [37]. The detailed descriptions and settings are provided in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "text", + "content": "Experimental settings. The feature extractor follows the same architecture as the ProtoNet used by [43], which has 4 modules with a " + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "text", + "content": " convolutions and 64 filters, followed by batch normalization, a ReLU nonlinearity, and " + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "text", + "content": " max-pooling, a.k.a., Conv4. The architecture of classifier have the design with a fully connected layer followed by a ReLU nonlinearity, a dropout layer, and another fully connected layer. The critic architecture has the same architecture with an extra sigmoid function. We've considered transductive and non-transductive in the settings. For the former, unlabeled test data of the same dataset is used as auxiliary data during training. This is applied to our method and other transductive methods including " + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathrm{PN} + \\mathrm{Co}" + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathrm{PN} + \\mathrm{Co}+" + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "text", + "content": ", STARTUP, and DDNet. For the latter, we utilize unlabeled datasets from other different datasets as auxiliary. This setting is applied to all the mentioned methods. During training, we use a learning rate " + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "inline_equation", + "content": "=0.001" + }, + { + "bbox": [ + 130, + 399, + 482, + 591 + ], + "type": "text", + "content": ", and use Adam as the optimizer. The batch size is 4. The meta-model is trained with 40000 randomly selected primary tasks and 40000 randomly selected auxiliary tasks. During evaluation, the average accuracy is computed based on 600 randomly sampled tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 609, + 300, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 300, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 300, + 620 + ], + "type": "text", + "content": "4.2 Main Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Following the experiment design of [48], we report the average accuracy of 5 runs of each 5-way 1-shot and 5-way 5-shots experiments in Tab. [1] to Tab. [2]. We use " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 0.5" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 0.02" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\gamma_{th} = 50\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\gamma_{div}" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": " as one out of six tasks for our methods" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 196, + 136, + 421, + 270 + ], + "blocks": [ + { + "bbox": [ + 212, + 114, + 401, + 126 + ], + "lines": [ + { + "bbox": [ + 212, + 114, + 401, + 126 + ], + "spans": [ + { + "bbox": [ + 212, + 114, + 401, + 126 + ], + "type": "text", + "content": "Table 1: Non-transductive auxiliary datasets" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 196, + 136, + 421, + 270 + ], + "lines": [ + { + "bbox": [ + 196, + 136, + 421, + 270 + ], + "spans": [ + { + "bbox": [ + 196, + 136, + 421, + 270 + ], + "type": "table", + "html": "
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+SPL48.55 ± 1.5143.44 ± 1.4937.06 ± 0.6342.11 ± 0.7155.20 ± 0.69
PN+FSR37.14 ± 0.6044.81 ± 0.5939.82 ± 0.6830.65 ± 0.4937.39 ± 0.63
RNNP55.39 ± 0.4950.18 ± 0.6241.75 ± 0.7243.76 ± 0.4860.15 ± 0.72
TraNFS53.02 ± 0.5246.66 ± 0.6338.25 ± 0.7940.32 ± 0.8860.86 ± 0.63
DCML54.39 ± 0.8749.33 ± 1.2241.77 ± 0.8048.33 ± 0.8060.28 ± 0.82
DUNT56.94 ± 1.0352.22 ± 1.1043.01 ± 0.5149.12 ± 0.6862.80 ± 0.49
Task5-way 5-shot
PN+SPL68.47 ± 0.7458.67 ± 0.8654.94 ± 1.0359.62 ± 1.3170.93 ± 0.83
PN+FSR51.48 ± 0.9161.89 ± 0.8057.38 ± 1.1050.40 ± 1.2144.68 ± 1.09
RNNP70.37 ± 1.3363.57 ± 1.2858.79 ± 2.3655.07 ± 1.2174.31 ± 1.36
TraNFS68.77 ± 1.1761.06 ± 1.1351.24 ± 1.1648.74 ± 1.2074.15 ± 0.83
DCML71.28 ± 1.2362.77 ± 1.0057.77 ± 1.3361.19 ± 1.0172.33 ± 0.90
DUNT72.17 ± 1.1165.10 ± 1.2358.50 ± 1.1162.01 ± 1.0974.44 ± 1.42
", + "image_path": "4fc70739048743b7d1b73b1a22b10e5d9e924ef46c005f70882e1a4b449f511a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 292, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 343 + ], + "type": "text", + "content": "in all datasets. In Tab. 1 and Tab. 2 we have provided the test accuracy all indomain datasets and one cross-domain dataset: mini-CUB for both transductive and non-transductive settings. The performance of cross-domain dataset 16 is reported in Tab. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 485 + ], + "type": "text", + "content": "Non-transductive results. For Tab. DUNT uses miniimagenet as auxiliary data for CUB datasets, uses tieredimagnet as auxiliary data for the cross-domain mini-CUB, and uses CUB as auxiliary data for the rest datasets. The results show that, with extra auxiliary datasets, DUNT has outperformed the general data-cleansing techniques and other noisy FSL methods in general. In Tab. the performance of SPL and FSR are generally worse than that of TraNFS and RNNP, which confirms our assumption that by removing potential noisy data, the data cleansing technique might harm the performance of FSL due to the lack of information, which can be compensated with our auxiliary data design. And compared to TraNFS and RNNP, DUNT has reduced the negative impact of the noisy labels more effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Transductive results. In Tab. 2, we consider the transductive setting where the auxiliary dataset for all the methods are unlabeled test data of the same dataset. The results show that our method ranks the top among all the transductive methods. Compared to Tab. 1, some of the transductive methods underperform non-transductive ones due to the following potential reasons: i) the test dataset is smaller than other auxiliary datasets. Therefore, the non-transductive dataset has more information to provide. Take miniImageNet as an example, the test data of the miniImageNet dataset only contains 1000 images and 10 classes, while the auxiliary dataset CUB contains 200 categories and 12000 images, the abundance and diversity of the auxiliary helps greatly improve the model's performance over the primary tasks more than the close test data does; and ii) the auxiliary dataset provides more valuable fine-grained tasks that benefits representation learning. For instance, when training miniImageNet with CUB as an auxiliary, both datasets contain natural images which are closely related, and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 195, + 136, + 421, + 251 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 392, + 125 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 392, + 125 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 392, + 125 + ], + "type": "text", + "content": "Table 2: Transductive auxiliary datasets" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 136, + 421, + 251 + ], + "lines": [ + { + "bbox": [ + 195, + 136, + 421, + 251 + ], + "spans": [ + { + "bbox": [ + 195, + 136, + 421, + 251 + ], + "type": "table", + "html": "
MethodsCUBminimini-CUBtieredCIFAR-FS
Task5-way 1-shot
PN+Co49.10 ± 1.2144.44 ± 0.9636.48 ± 0.9840.36 ± 0.8152.51 ± 0.92
PN+Co+49.45 ± 1.1945.08 ± 0.9139.54 ± 1.0940.19 ± 1.1051.27 ± 1.11
STARTUP52.02 ± 1.2048.79 ± 1.1840.08 ± 1.2244.48 ± 0.9260.69 ± 1.31
DDNet52.55 ± 1.3950.82 ± 0.8741.11 ± 1.8945.11 ± 1.3360.25 ± 1.46
DUNT55.01 ± 1.1252.38 ± 0.8542.07 ± 1.1049.29 ± 1.2661.90 ± 1.14
Task5-way 5-shot
PN+Co65.42 ± 1.2259.57 ± 1.2056.90 ± 1.2955.10 ± 1.2166.89 ± 1.26
PN+Co+67.92 ± 1.1160.43 ± 1.1354.29 ± 1.1054.54 ± 0.8467.22 ± 0.71
STARTUP68.74 ± 1.2162.00 ± 1.2451.07 ± 1.0155.92 ± 0.9974.04 ± 0.84
DDNet69.21 ± 1.3463.70 ± 2.0153.87 ± 0.4457.22 ± 0.7573.92 ± 1.20
DUNT71.20 ± 1.1266.46 ± 0.9859.05 ± 1.0161.09 ± 0.9076.11 ± 0.98
", + "image_path": "cd308ae470bfe9533e4b31bb663e01724ad4a5b58a7f20bf4a5f16a5040319df.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 281, + 482, + 406 + ], + "blocks": [ + { + "bbox": [ + 238, + 261, + 375, + 270 + ], + "lines": [ + { + "bbox": [ + 238, + 261, + 375, + 270 + ], + "spans": [ + { + "bbox": [ + 238, + 261, + 375, + 270 + ], + "type": "text", + "content": "Table 3: Cross-domain Datasets" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 281, + 482, + 406 + ], + "lines": [ + { + "bbox": [ + 135, + 281, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 135, + 281, + 482, + 406 + ], + "type": "table", + "html": "
MethodsEuroSATCropDiseaseChestXISICEuroSATCropDiseaseChestXISIC
5-way 1-shot5-way 5-shot
ProtoNet52.52±0.5256.32±0.8928.90±0.4330.11±1.0061.60±0.7078.89±0.8333.04±0.7136.95±0.90
PN+SPL52.89±1.2158.95±2.0427.68±0.5231.79±1.0967.60±1.2380.25±1.1931.13±1.1837.05±1.15
PN+FSR47.53±1.2156.97±1.2328.40±0.6133.58±0.9261.96±0.9075.93±0.8631.43±1.1936.42±1.22
RNNP51.86±1.2056.56±1.2127.39±0.6131.26±0.8067.97±1.3478.43±1.2931.16±1.2537.84±1.19
TraNFS50.61±1.1058.39±0.9128.26±0.9832.84±1.1167.93±1.1181.51±1.2332.16±1.1636.95±1.18
DCML52.33±1.0258.36±0.4627.54±0.6931.96±0.8068.01±0.9079.11±0.9831.64±1.0035.44±1.10
PN+Co56.56±1.0859.13±1.1426.67±0.7329.16±1.1372.08±1.1578.59±2.1130.45±1.1734.74±1.09
PN+Co+52.18±1.1158.18±1.1028.23±0.8730.63±1.0665.57±1.1078.47±0.9030.77±0.8233.68±0.91
STARTUP52.74±1.2259.43±1.2327.75±0.9532.84±1.1369.78±1.1780.12±1.2231.69±1.3039.58±0.99
DDNet55.21±1.8360.09±1.0928.66±0.9834.10±1.0170.09±1.9081.90±2.0932.01±1.8939.44±1.13
DUNT57.10±1.2961.49±1.0129.78±0.9035.80±0.7772.81±1.1083.20±1.0833.94±1.2040.31±0.90
", + "image_path": "c50f63ef4a795b4dae068d57e2695e5857024b7b19ace1c888fada52c4ea5f2e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 417, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 477 + ], + "type": "text", + "content": "CUB dataset is a collection of bird species, which provide fine-grained tasks for learning a better feature representation, while the miniImageNet test data only provide coarse-grained task, which makes it less useful in some extend. As for DUNT, the performance stays stable regardless the two settings, showing that our framework has the potential to fully utilize any extra information provided." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 487, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 487, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 487, + 606 + ], + "type": "text", + "content": "Cross-domain results. In Tab. 3, we present the performance on the cross-domain datasets. The training procedure follows the setting in [16], which uses miniImageNet as base dataset for all the methods. For CoTeaching, CoTeaching+, STARTUP, DDNet and DUNT, we use the unlabeled test data EuroSAT, CropDisease, ChestX, and ISIC as auxiliary dataset, respectively. In cross-domain datasets, the transductive methods generally outperform the non-transductive ones due to the utilization of the unlabeled test data that reduces the domain gap between the source domain and the target domain, which is critical when the they are extremely different. DUNT has used the Wasserstein regularization to further reduce the differences. Hence, it ranks the highest in the cross-domain datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 623, + 238, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 238, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 238, + 635 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 667 + ], + "type": "text", + "content": "In this section, we study the effectiveness of different components and impact of hyperparameters. For more ablations, please refer to Appendix C." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 447, + 102 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 305, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 305, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 305, + 235 + ], + "type": "text", + "content": "Effectiveness of different components. Tab. 4 shows the ablation study with different elements of DUNT on CUB and miniImageNet datasets under nontransductive setting. We first show the performance of only using the primary data, which is vanilla ProtoNet (Primary). It can be observed that each of these components have a significant impact on the overall performance." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 313, + 129, + 476, + 227 + ], + "blocks": [ + { + "bbox": [ + 312, + 117, + 481, + 128 + ], + "lines": [ + { + "bbox": [ + 312, + 117, + 481, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 117, + 481, + 128 + ], + "type": "text", + "content": "Table 4: Impact of different components" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 129, + 476, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 129, + 476, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 476, + 227 + ], + "type": "table", + "html": "
PrimaryAuxiliaryRegECTSCUBmini
✓XXXX42.57 ± 0.3439.95 ± 1.06
✓XX✓X48.55 ± 1.5143.44 ± 1.49
✓✓XXX50.19 ± 0.7148.90 ± 0.77
✓✓X✓X53.55 ± 0.9750.02 ± 0.81
✓✓XX✓52.70 ± 1.0149.50 ± 1.66
✓✓X✓✓54.31 ± 0.9950.18 ± 0.87
✓✓✓✓X51.90 ± 0.8949.18 ± 0.76
✓✓✓✓X55.47 ± 0.7950.73 ± 0.60
✓✓✓X✓54.79 ± 1.1049.91 ± 1.00
✓✓✓✓✓56.94 ± 1.0352.22 ± 1.10
", + "image_path": "46f7933062526486d7174b0fe53a1ae0c994bb130ca686c842397be0c466c999.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 297 + ], + "type": "text", + "content": "Auxiliary data (Auxiliary) and episode cleansing (EC) serve pivotal functions, with the former introducing additional data and the latter eliminating noisy data. The regularization term (Reg) aids in extracting domain-invariant features, contributing to further performance improvement. Task selection (TS) mitigates the risk of overfitting, thereby enhancing overall performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 264, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 264, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 264, + 412 + ], + "type": "text", + "content": "Effectiveness of task selection. We study the impact of task selection through the threshold " + }, + { + "bbox": [ + 130, + 304, + 264, + 412 + ], + "type": "inline_equation", + "content": "\\gamma_{div}" + }, + { + "bbox": [ + 130, + 304, + 264, + 412 + ], + "type": "text", + "content": ". As mentioned earlier, to streamline the hyperparameter search process, we opt for an approach that involves upsampling " + }, + { + "bbox": [ + 130, + 304, + 264, + 412 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 304, + 264, + 412 + ], + "type": "text", + "content": " unsupervised tasks. Subsequently," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 274, + 304, + 370, + 358 + ], + "blocks": [ + { + "bbox": [ + 274, + 304, + 370, + 358 + ], + "lines": [ + { + "bbox": [ + 274, + 304, + 370, + 358 + ], + "spans": [ + { + "bbox": [ + 274, + 304, + 370, + 358 + ], + "type": "image", + "image_path": "119c14515f988d49b3272c985d6487b8da064817c7b0187310459b8a882185c9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 363, + 328, + 373 + ], + "lines": [ + { + "bbox": [ + 316, + 363, + 328, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 328, + 373 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 378, + 304, + 471, + 358 + ], + "blocks": [ + { + "bbox": [ + 378, + 304, + 471, + 358 + ], + "lines": [ + { + "bbox": [ + 378, + 304, + 471, + 358 + ], + "spans": [ + { + "bbox": [ + 378, + 304, + 471, + 358 + ], + "type": "image", + "image_path": "c2eed60c02f5ad14d0614b1639173f300d880f92f1f4932967b26cdfbec4e02d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 363, + 431, + 373 + ], + "lines": [ + { + "bbox": [ + 419, + 363, + 431, + 373 + ], + "spans": [ + { + "bbox": [ + 419, + 363, + 431, + 373 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 381, + 482, + 405 + ], + "lines": [ + { + "bbox": [ + 270, + 381, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 270, + 381, + 482, + 405 + ], + "type": "text", + "content": "Fig. 3: DUNT on CUB with different " + }, + { + "bbox": [ + 270, + 381, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\gamma_{div}" + }, + { + "bbox": [ + 270, + 381, + 482, + 405 + ], + "type": "text", + "content": " (a); DUNT with different task selection (b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": "we select only the most diverse task to be included in the valuable task group, instead of specifying a real value for the diverseness threshold. In this ablation, the number of upsampled tasks varies from 2-10. As depicted in Fig. 3 (a), the performance improves alongside an increase in the parameter " + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": ", but this comes at the cost of heightened computational demands. Consequently, considering efficiency, we opt for " + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{div}} = 6" + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": " for all experiments. In Fig. 3 (b), we test the effect of using different unsupervised task: select confusing tasks and select valuable tasks. It is evident from the results that the inclusion of confusing tasks significantly deteriorates performance, whereas the utilization of valuable tasks leads to a substantial improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "We propose a novel FSL framework with auxiliary tasks that utilize carefully selected unlabeled data under noisy settings. To better align the auxiliary distribution with the primary one, we propose a regularization term based on the Wasserstein distance for learning a domain-invariant representation. To enhance robustness against label noise, we introduce episode cleansing for examples in the primary task and adopt a diverse task selection strategy for the unsupervised auxiliary tasks. Our framework is theoretically and experimentally proved to be effective and beneficial." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "text", + "content": "1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 194 + ], + "type": "text", + "content": "2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Bertinetto, L., Henriques, J.F., Torr, P.H., Vedaldi, A.: Meta-learning with differentiable closed-form solvers. In: Proceedings of the International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "text", + "content": "4. Beyer, L., Henaff, O.J., Kolesnikov, A., Zhai, X., Oord, A.v.d.: Are we done with imagenet? arXiv preprint arXiv:2006.07159 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 250, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 293 + ], + "type": "text", + "content": "5. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 335 + ], + "type": "text", + "content": "6. Chobola, T., Vašata, D., Kordík, P.: Transfer learning based few-shot classification using optimal transport mapping from preprocessed latent space of backbone neural network. In: AAAI Workshop on Meta-Learning and MetaDL Challenge. pp. 29-37. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "text", + "content": "7. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Advances in Neural Information Processing Systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "type": "text", + "content": "8. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 274-289. Springer (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "9. Deleu, T., Würfl, T., Samiei, M., Cohen, J.P., Bengio, Y.: Torchmeta: A Meta-Learning library for PyTorch (2019), https://arxiv.org/abs/1909.06576, available at: https://github.com/tristandeleu/pytorch-meta" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "text", + "content": "10. Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: Proceedings of the 34th International Conference on Machine Learning. pp. 1126-1135 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "text", + "content": "1. Flamary, R., Courty, N., Tuia, D., Rakotomamonjy, A.: Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell 1 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "2. Ghosh, A., Kumar, H., Sastry, P.S.: Robust loss functions under label noise for deep neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "3. Gidaris, S., Bursuc, A., Komodakis, N., Pérez, P., Cord, M.: Boosting few-shot visual learning with self-supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8059-8068 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 588 + ], + "type": "text", + "content": "4. Goldblum, M., Fowl, L., Goldstein, T.: Adversarily robust few-shot learning: A meta-learning approach. Advances in Neural Information Processing Systems 33, 17886-17895 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "text", + "content": "5. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "6. Guo, Y., Codella, N.C., Karlinsky, L., Codella, J.V., Smith, J.R., Saenko, K., Rosing, T., Feris, R.: A broader study of cross-domain few-shot learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16. pp. 124-141. Springer (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 447, + 101 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "17. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Proceedings of the Conference on Neural Information Processing Systems (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "text", + "content": "18. Han, C., Fan, Z., Zhang, D., Qiu, M., Gao, M., Zhou, A.: Meta-learning adversarial domain adaptation network for few-shot text classification. In: Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021. Findings of ACL, vol. ACL/IJCNLP 2021, pp. 1664-1673. Association for Computational Linguistics (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "20. Hendrycks, D., Mazeika, M., Wilson, D., Gimpel, K.: Using trusted data to train deep networks on labels corrupted by severe noise. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "21. Islam, A., Chen, C.F.R., Panda, R., Karlinsky, L., Feris, R., Radke, R.J.: Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems 34, 3584-3595 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "22. Jiang, L., Meng, D., Mitamura, T., Hauptmann, A.G.: Easy samples first: Self-paced reranking for zero-example multimedia search. In: Proceedings of the 22nd ACM international conference on Multimedia. pp. 547-556 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "type": "text", + "content": "23. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "type": "text", + "content": "24. Khodadadeh, S., Boloni, L., Shah, M.: Unsupervised meta-learning for few-shot image classification. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "25. Killamsetty, K., Li, C., Zhao, C., Iyer, R., Chen, F.: A reweighted meta learning framework for robust few shot learning. arXiv preprint arXiv:2011.06782 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "26. Kumar, M.P., Packer, B., Koller, D.: Self-paced learning for latent variable models. In: Proceedings of the Conference on Neural Information Processing Systems. vol. 1, p. 2 (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "27. Li, W., Wang, L., Zhang, X., Qi, L., Huo, J., Gao, Y., Luo, J.: Defensive few-shot learning. IEEE transactions on pattern analysis and machine intelligence (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "28. Liang, K.J., Rangrej, S.B., Petrovic, V., Hassner, T.: Few-shot learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9089-9098 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "text", + "content": "29. Lu, J., Jin, S., Liang, J., Zhang, C.: Robust few-shot learning for user-provided data. IEEE transactions on neural networks and learning systems 32(4), 1433-1447 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "30. Masi, I., Trn, A.T., Hassner, T., Sahin, G., Medioni, G.: Face-specific data augmentation for unconstrained face recognition. International Journal of Computer Vision 127, 642-667 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "31. Mazumder, P., Singh, P., Namboodiri, V.P.: Rnnp: A robust few-shot learning approach. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2664-2673 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "32. Medina, C., Devos, A., Grossglauer, M.: Self-Supervised Prototypical Transfer Learning for Few-Shot Classification. arXiv preprint arXiv:2006.11325 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "33. Motiian, S., Jones, Q., Iranmanesh, S., Doretto, G.: Few-shot adversarial domain adaptation. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "34. Northcutt, C.G., Athalye, A., Mueller, J.: Pervasive label errors in test sets destabilize machine learning benchmarks. arXiv preprint arXiv:2103.14749 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Phoo, C.P., Hariharan, B.: Self-training for few-shot transfer across extreme task differences. In: Proceedings of the International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "36. Qin, T., Li, W., Shi, Y., Gao, Y.: Diversity helps: Unsupervised few-shot learning via distribution shift-based data augmentation. arXiv preprint arXiv:2004.05805 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "37. Que, X., Yu, Q.: Dual-level curriculum meta-learning for noisy few-shot learning tasks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 14740-14748 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "text", + "content": "38. Ren, M., Triantafillou, E., Ravi, S., Snell, J., Swersky, K., Tenenbaum, J.B., Larochelle, H., Zemel, R.S.: Meta-learning for semi-supervised few-shot classification. In: Proceedings of the International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 295 + ], + "type": "text", + "content": "39. Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: International Conference on Machine Learning. pp. 4334-4343. PMLR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "type": "text", + "content": "40. Sahoo, D., Le, H., Liu, C., Hoi, S.C.: Meta-learning with domain adaptation for few-shot learning under domain shift (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "type": "text", + "content": "41. Shui, C., Li, Z., Li, J., Gagné, C., Ling, C.X., Wang, B.: Aggregating from multiple target-shifted sources. In: International Conference on Machine Learning. pp. 9638-9648. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "text", + "content": "42. Shui, C., Zhou, F., Gagné, C., Wang, B.: Deep active learning: Unified and principled method for query and training. In: International Conference on Artificial Intelligence and Statistics. pp. 1308-1318. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "43. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems. pp. 4077-4087 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 409, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 452 + ], + "type": "text", + "content": "44. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin, A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: Advances in neural information processing systems (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "text", + "content": "45. Sun, Q., Liu, Y., Chua, T.S., Schiele, B.: Meta-transfer learning for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 403-412 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 508 + ], + "type": "text", + "content": "46. Tolstikhin, I., Bousquet, O., Gelly, S., Schoelkopf, B.: Wasserstein auto-encoders. In: International Conference on Machine Learning (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "text", + "content": "47. Tsipras, D., Santurkar, S., Engstrom, L., Ilyas, A., Madry, A.: From imagenet to image classification: Contextualizing progress on benchmarks. In: International Conference on Machine Learning. pp. 9625-9635. PMLR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 575 + ], + "type": "text", + "content": "48. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for one shot learning. In: Advances in neural information processing systems. pp. 3630-3638 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "text", + "content": "49. Wang, X., Hua, Y., Kodirov, E., Robertson, N.M.: Imae for noise-robust learning: Mean absolute error does not treat examples equally and gradient magnitude's variance matters. arXiv preprint arXiv:1903.12141 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "50. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322-330 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "51. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-ucsd birds 200 (2010)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 267, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 267, + 91, + 447, + 101 + ], + "type": "text", + "content": "Robust Learning from Noisy Few-Shot Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 435 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "52. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "53. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems 33, 7597-7610 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 482, + 215 + ], + "type": "text", + "content": "54. Xia, X., Liu, T., Wang, N., Han, B., Gong, C., Niu, G., Sugiyama, M.: Are anchor points really indispensable in label-noise learning? Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 482, + 248 + ], + "type": "text", + "content": "55. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10687-10698 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 482, + 281 + ], + "type": "text", + "content": "56. Xu, R., Yang, X., Liu, B., Zhang, K., Liu, W.: Selecting task with optimal transport self-supervised learning for few-shot classification. arXiv preprint arXiv:2204.00289 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "type": "text", + "content": "57. Xu, Y., Cao, P., Kong, Y., Wang, Y.: L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 482, + 335 + ], + "type": "text", + "content": "58. Yang, Y., Liang, K.J., Carin, L.: Object detection as a positive-unlabeled problem. arXiv preprint arXiv:2002.04672 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "type": "text", + "content": "59. Yu, X., Han, B., Yao, J., Niu, G., Tsang, I., Sugiyama, M.: How does disagreement help generalization against label corruption? In: International Conference on Machine Learning. pp. 7164-7173. PMLR (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 482, + 402 + ], + "type": "text", + "content": "60. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 482, + 435 + ], + "type": "text", + "content": "61. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Que and Yu" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_content_list.json b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f053888917248038f69aaaa78a4b611d6781034b --- /dev/null +++ b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_content_list.json @@ -0,0 +1,1394 @@ +[ + { + "type": "text", + "text": "Optimization-based Uncertainty Attribution Via Learning Informative Perturbations", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanjing Wang $^{1}$ , Bashirul Azam Biswas $^{1}$ , and Qiang Ji $^{1}$", + "bbox": [ + 287, + 210, + 712, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rensselaer Polytechnic Institute, Troy NY 12180, USA {wangh36, biswab, jiq}@rpi.edu", + "bbox": [ + 316, + 239, + 683, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Uncertainty attribution (UA) aims to identify key contributors to predictive uncertainty in deep learning models. To improve the faithfulness of existing UA methods, we formulate UA as an optimization problem to learn a binary mask on the input. The learned mask identifies regions that significantly contribute to output uncertainty and allows uncertainty reduction through learning informative perturbations on the masked input. Our method enhances UA interpretability and maintains high efficiency by integrating three key improvements: Segment Anything Model (SAM)-guided mask parameterization for efficient and interpretable mask learning; learnable perturbations that adaptively target and refine problematic regions specific to each input without manually tuning the perturbation parameters; and a novel application of Gumbel-sigmoid reparameterization for efficiently learning Bernoulli-distributed binary masks under continuous optimization. Our experiments on problematic region detection and faithfulness tests demonstrate our method's superiority over state-of-the-art UA methods.", + "bbox": [ + 261, + 297, + 743, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Uncertainty Attribution $\\cdot$ Model Explainability", + "bbox": [ + 261, + 532, + 666, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 569, + 375, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While conventional deep learning has made remarkable strides in various domains, they often face challenges with overconfidence in their predictions, which can be risky in safety-critical applications. Hence, uncertainty quantification (UQ) is paramount, as it directly impacts the safety, reliability, and trustworthiness of AI systems, especially for various computer vision applications [8,15,48]. Predictive uncertainties in deep learning are categorized into aleatoric uncertainty, reflecting inherent data randomness, and epistemic uncertainty, resulting from limited model knowledge. Bayesian Neural Networks (BNNs) serve as a foundational approach for quantifying uncertainty, encompassing techniques such as Variational Inference (VI) [7, 19, 22] and Markov Chain Monte Carlo (MCMC) [36, 42, 47] methods to estimate the complex posterior distribution of model parameters. Following this, Ensemble methods [16, 37, 38, 43] offer an alternative strategy, aggregating predictions from multiple models to capture uncertainty. While numerous studies have focused on enhancing the efficiency and accuracy of UQ, comparatively little effort has been devoted to understanding the sources of uncertainty. Identifying the sources of uncertainty is essential,", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as it reveals where and why certain inputs are associated with high uncertainty for enhanced explainability, thereby informing more targeted improvements in model design and training.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uncertainty attribution (UA) [1, 26, 39, 40] systematically identifies and analyzes the specific inputs or features that contribute significantly to the uncertainty in the model's predictions. Traditional explainable AI methods highlight areas of inputs leading to specific predictions, which are unreliable when applied directly to UA for identifying problematic regions in input data. The primary reason is that they focus on explaining the model's predictions, assuming confident predictions, while UA aims to pinpoint the sources of input imperfections that contribute to high predictive uncertainties. Additionally, these methods are often class-discriminative when applied to explain the classification score. Consequently, they may struggle to explain inputs with incorrect predictions and large uncertainties [26]. Uncertainty attribution methods are broadly categorized into two types: gradient-based and perturbation-based methods. Gradient-based methods leverage the gradients of the model output with respect to the input to identify the areas contributing to uncertainty. On the other hand, perturbation-based methods involve systematically altering parts of the input and observing the effect on uncertainty, allowing for a more direct assessment of how variations in the input influence the uncertainty. While gradient-based UA methods are often efficient, they only consider the uncertainty sensitivity to very small input changes, leading to noisy and sometimes misleading indications of uncertainty. Perturbation-based methods can be computationally intensive and may be less accurate if an insufficient number of perturbations are performed.", + "bbox": [ + 212, + 196, + 787, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by existing literature [3,5,45], we treat UA as an optimization task aimed at learning a binary mask to identify problematic regions of the input. This mask targets specific areas for maximum uncertainty reduction, achieved by applying learnable perturbations to these selected areas. This optimization-based UA framework can potentially improve the faithfulness of UA, which quantifies how accurately a method's explanations reflect the real impact of input features on the uncertainty. A method is considered faithful if the identified areas truly correspond to regions in the input that, when perturbed, have a significant impact on the predictive uncertainty. Our approach ensures faithfulness through a direct linkage between the uncertainty and the identification of uncertain regions in the optimization objective. Moreover, compared to perturbation-based methods, our approach utilizes gradients to iteratively update the mask through continuous optimization, avoiding the need for computationally intensive, pixelwise input perturbations. This strategy may reduce the cost associated with exploring an exponential number of pixel combinations. Unlike gradient-based methods, which directly employ gradients to generate uncertainty maps, our method applies actual perturbations to inputs, not just minor input changes, potentially yielding more accurate explanations.", + "bbox": [ + 212, + 518, + 787, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance UA performance, we introduce three key adjustments. First, we use a linear combination of segments from the pre-trained Segment Anything Model (SAM) [12] to parameterize the mask. Using SAM segments as a basis", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enables targeted analysis of semantic regional effects and simplifies the learning process to linear combination coefficients instead of pixel-level masks. Second, traditional perturbations used to explain model predictions often fall short in UA. For example, widely used mask-out perturbations invariably result in increased uncertainty in UA, even when applied to problematic regions. This underscores the need for advanced, learnable perturbations in UA, which are essential for accurately reducing uncertainty by modifying problematic regions. Finally, given the challenge of learning a binary mask directly, we adopt a Bernoulli distribution for the masks and employ the Gumbel-sigmoid reparameterization trick [10], enabling continuous optimization with high efficiency.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 321, + 387, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gradient-based Attribution Methods. Gradient-based methods [11,29,31-35,44,46] leverage gradient information as input attribution. However, raw gradients can be noisy, leading researchers to develop various approaches for refining gradient information. SmoothGrad [32] addresses the issue of noisy gradients by aggregating attributions from multiple noisy images. FullGrad [34] combines the gradients of the output with respect to both the input and the bias variable in each convolutional or fully-connected layer. Integrated Gradients (IG) [35] and its variants [11, 44] create a path integral between a reference image and the input. They then aggregate the gradients of the model output for the artificial images along this path, providing a more comprehensive understanding of the input-output relationship. Grad-CAM [29] focuses on the gradients of the latent feature maps, offering more interpretable insights. Layer-wise relevance propagation (LRP) methods [2, 23, 30] construct attributions by backpropagating the relevance score layer-by-layer from the output to the input. While most gradient-based methods can be applied for both model prediction and uncertainty attribution, LRP methods often need specific neural network architectures [39]. The direct application of these attribution methods might not be sufficient for UA, leading Wang et al. [39] to develop a backpropagation method specifically for UA. In summary, gradient-based methods are often computationally efficient to analyze neural network input-output relationships. Yet, they mainly model output changes from very small input perturbations, which may not yield precise UA results. This limitation is not effectively overcome by refining gradient information through smoothing, aggregation, and path integration techniques.", + "bbox": [ + 212, + 356, + 787, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perturbation-based Attribution Methods. Perturbation-based methods [3, 5, 6, 27, 45, 46] estimate attributions by modifying the input and observing the subsequent changes in the model's output. For example, LIME [28] approximates the model's behavior locally by fitting an interpretable model to the perturbed data. RISE [27] generates saliency maps by randomly masking the input and aggregating the model's responses. SHAP and its variants [20, 21] leverage Shapley values from cooperative game theory to interpret complex model predictions, assigning importance to each feature by observing output changes upon feature", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "perturbations. The SHAP-based methods are extended to UA with theoretical studies [41]. While perturbation-based methods provide a more intuitive understanding of the model's behavior and allow for better interpretability, they can also be computationally expensive. This is primarily due to the need for an exponentially increasing number of perturbations and subsequent model evaluations to explore the input-output relationship effectively. To improve efficiency, Dabkowski et al. [3], Fong et al. [5], and Yang et al. [45] focus on identifying the smallest regions for perturbation that can significantly change the model's prediction, leveraging continuous optimization techniques to speed up the attribution process. In recent years, several methods have been specifically designed for UA. For instance, CLUE [1] and its variants [17,18] focus on creating an improved image with minimal uncertainty by modifying the uncertain input using a generative model. The attribution map is then generated by measuring the difference between the original and modified input. To further enhance pixel-wise attributions, Perez et al. [26] combined the CLUE method with the path integral technique. However, CLUE-based methods require a pre-trained generative model, which may be challenging to acquire for practical applications.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 435, + 426, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 477, + 361, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bayesian Neural Networks. Distinct from point-estimated neural networks, BNNs treat neural network parameters as random variables and calculate the posterior distribution of the parameters $\\pmb{\\theta}$ , denoted as $p(\\pmb{\\theta} \\mid \\mathcal{D})$ where $\\mathcal{D}$ represents the training data. Utilizing Bayes' rule, the posterior distribution can be computed using $p(\\pmb{\\theta} \\mid \\mathcal{D}) = \\frac{p(\\mathcal{D}|\\pmb{\\theta})p(\\pmb{\\theta})}{p(\\mathcal{D})}$ with $p(\\mathcal{D}|\\pmb{\\theta})$ as the likelihood and $p(\\pmb{\\theta})$ as the prior. Since the exact posterior distribution is often intractable, various techniques have been developed, including MCMC sampling, VI and Ensemble methods. These approaches either directly sample from the posterior or approximate it using a simpler distribution for efficient sampling.", + "bbox": [ + 212, + 511, + 787, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty Quantification. In this study, we concentrate on image classification tasks, where $\\mathbf{x}$ denotes the input and $\\mathbf{y}$ denotes the output. The neural network predicts $p(\\mathbf{y}|\\mathbf{x},\\boldsymbol{\\theta})$ , which is a categorical distribution parameterized by learnable softmax probability. Entropy is a common method for quantifying the uncertainty or randomness in a probability distribution. In entropy-based UQ, the total uncertainty of predictions is decomposed as follows [4]:", + "bbox": [ + 212, + 651, + 787, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x} , \\mathcal {D}) ]} _ {\\text {T o t a l U n c e r t a i n t y} U _ {t} (\\mathbf {x})} = \\underbrace {\\mathcal {I} [ \\mathbf {y} , \\boldsymbol {\\theta} | \\mathbf {x} , \\mathcal {D} ]} _ {\\text {E p i s t e m i c U n c e r t a i n t y}} + \\underbrace {\\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} \\left[ \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x} , \\boldsymbol {\\theta}) ] \\right]} _ {\\text {A l e a t o r i c U n c e r t a i n t y} U _ {a} (\\mathbf {x})} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 757, + 785, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{H}$ , $\\mathcal{I}$ , and $\\mathbb{E}$ stand for entropy, mutual information, and expectation, respectively. By employing a Monte Carlo approximation of the posterior, we can", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "approximate the uncertainties as:", + "bbox": [ + 215, + 147, + 460, + 161 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\mathcal {D}) ] = \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta}) ] \\right] \\approx \\mathcal {H} \\left[ \\frac {1}{S} \\sum_ {s = 1} ^ {S} p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta} ^ {s}) \\right] \\tag {2a}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 172, + 785, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} \\left[ \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta}) ] \\right] \\approx \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta} ^ {s}) ] \\tag {2b}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 217, + 785, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{\\theta}^s\\sim p(\\pmb {\\theta}|\\mathcal{D})$ represents the sampling of parameters $\\pmb{\\theta}$ from their posterior distribution, and $S$ is the sample size.", + "bbox": [ + 212, + 267, + 784, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Optimization-based Uncertainty Attribution", + "text_level": 1, + "bbox": [ + 214, + 319, + 627, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach defines UA as an optimization problem, aiming to learn a binary mask, $\\mathbf{M}$ , with the same dimensions as $\\mathbf{x}$ , that highlights areas in the inputs significantly contributing to uncertainty. For each element $u$ in $\\mathbf{x}$ , we assume $\\mathbf{M}(u) \\in \\{0,1\\}$ , with $\\mathbf{M}(u) = 1$ indicating strong contributions to predictive uncertainty and $\\mathbf{M}(u) = 0$ otherwise. To estimate each element's contribution, we perturb the regions indicated by $\\mathbf{M}$ . Specifically, with $\\hat{\\mathbf{x}}$ representing the perturbed input with reduced uncertainty, we replace the regions where $\\mathbf{M}(u) = 1$ in $\\mathbf{x}$ with those from $\\hat{\\mathbf{x}}$ , resulting in $\\mathbf{x}' = (1 - \\mathbf{M}) \\odot \\mathbf{x} + \\mathbf{M} \\odot \\hat{\\mathbf{x}}$ to denote the masked input. The goal is to identify a small subset of $\\mathbf{x}$ 's regions, $\\mathbf{M}$ , that significantly reduce predictive uncertainty through the masked input. The uncertainty, denoted as $U$ , could be aleatoric $(U_a)$ , epistemic $(U_e)$ , or total $(U_t)$ , as indicated in Eq. (1). The optimal $\\mathbf{M}^*$ is determined by:", + "bbox": [ + 212, + 343, + 787, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {*} = \\arg \\min _ {\\mathbf {M}} U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {M} | | \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 541, + 785, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\odot$ denotes element-wise multiplication, $||\\mathbf{M}||$ is the norm of $\\mathbf{M}$ , and $\\lambda$ is a hyperparameter balancing uncertainty reduction and mask sparsity, usually prioritizing uncertainty loss over sparsity regularization. As $U$ utilizes all ensemble models sampled from $p(\\boldsymbol{\\theta}|\\mathcal{D})$ to compute uncertainty, the UA maps are learned by jointly backpropagating through these models. Perturbations in identified problematic regions decrease predictive uncertainty, while those in high-certain areas will increase uncertainty, guiding $\\mathbf{M}$ to select critical regions for perturbation to reduce uncertainty. Background or non-informative regions will not impact uncertainty after perturbation, necessitating the inclusion of a sparsity term to ensure these areas are assigned zero values in $\\mathbf{M}$ .", + "bbox": [ + 212, + 566, + 787, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 SAM-guided Mask Parameterization", + "text_level": 1, + "bbox": [ + 214, + 739, + 563, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the explainability of our proposed method and reduce the optimization complexity, we aim to parameterize the uncertainty mask, $\\mathbf{M}$ , through a linear combination of segments derived from the pre-trained SAM model, denoted as $\\{\\mathbf{M}_i\\}_{i=1}^N$ . These $N$ segments are binary masks of the same dimension as $\\mathbf{M}$ , each highlighting distinct regions of the input space. Generated by the", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 302, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SAM model, these segments are inherently binary and delineate areas corresponding to semantically meaningful and human-understandable concepts. The parameterization of $\\mathbf{M}$ is $\\mathbf{M} = \\sum_{i=1}^{N} w_i \\mathbf{M}_i$ where $w_i \\in \\{0,1\\}$ represents the binary weights. By employing these segments as the basis for constructing $\\mathbf{M}$ , our method not only identifies regions contributing to uncertainty but also ensures that these regions are aligned with recognizable semantic features. This alignment significantly boosts the explainability of our method, as the identified areas of uncertainty are directly tied to concepts that are easily interpretable by humans. For instance, in an image classification task, rather than indicating abstract or non-intuitive areas of uncertainty, our method could highlight specific objects or parts of an image that contribute to the uncertainty, such as \"eyes\" in a facial recognition task or \"vehicles\" in an autonomous driving context.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the effectiveness of our methods depends on the accuracy of SAM, we believe our approach can still be effective with imperfect SAM segmentation. For instance, if problematic regions are either included in a large SAM segment or spread across multiple segments, our method can sometimes still identify these areas as potentially problematic. Furthermore, our method is compatible with state-of-the-art segmentation models, ensuring adaptability and improved performance across diverse applications with domain-specific segmentation models.", + "bbox": [ + 212, + 328, + 787, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Learnable Perturbation", + "text_level": 1, + "bbox": [ + 214, + 454, + 455, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Traditional perturbation-based methods often use fixed or random perturbations, which may not effectively reduce uncertainty and could unintentionally increase it, even in problematic regions. Therefore, we dynamically adjust learnable perturbations during the optimization process, specifically targeting and modifying areas in the input $\\mathbf{x}$ to effectively reduce uncertainty.", + "bbox": [ + 212, + 478, + 785, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach updates the perturbed input $\\hat{\\mathbf{x}}$ , as described in Eq. (3), through iterative blurring, represented as $\\hat{\\mathbf{x}} = g_{\\phi}(\\mathbf{x})$ , where $g_{\\phi}(\\cdot)$ is a blurring function parameterized by $\\phi$ . This strategy hypothesizes that blurring problematic regions decreases uncertainty, while blurring good areas could increase it by obscuring important information for prediction. Recognizing that a universal blurring strategy does not cater to the varied demands of different inputs, we aim to learn the blurring function's parameters $\\phi$ within the optimization process. This approach allows for the precise, dynamic adjustment of perturbations, blurring each region as needed to optimally reduce uncertainty.", + "bbox": [ + 212, + 554, + 787, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5 Gumbel-sigmoid Reparameterization For Binary Weights", + "text_level": 1, + "bbox": [ + 214, + 710, + 732, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the challenge of modeling binary weights for our uncertainty mask, $\\mathbf{M}$ , we confront the complexities of discrete optimization, which demands extensive search efforts and complicates direct learning of these weights. While continuous optimization through iterative gradient descent offers a potential solution, simply relaxing binary weights to a continuous range between [0, 1] undermines the mask's binary nature, which is crucial for pinpointing problematic areas and facilitating accurate input region replacement through perturbations.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On one hand, adhering strictly to binary weights ensures clarity in identifying areas contributing to uncertainty but, on the other hand, it restricts our ability to convey the nuanced relative importance of these regions.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this paper, we propose maintaining the binary conception of weights while adopting a probabilistic approach for learning them. Specifically, we model each weight as an independent Bernoulli distribution, denoted as $w_{i} \\sim \\mathrm{Ber}(p_{i})$ with probability parameter $p_i \\in [0,1]$ . This strategy enables the use of continuous optimization techniques to learn the probabilities, thereby circumventing the challenges associated with direct discrete optimization. Consequently, our loss function is designed to account for the expected loss over the distribution of weights. At each step of gradient descent, we sample from this distribution, using the Gumbel-sigmoid function [10, 25] as a smooth approximation to generate these samples. Let $w_{i}^{s}$ denote the $s$ -th sample from $\\mathrm{Ber}(p_i)$ . The Gumbel-sigmoid reparameterization trick approximates the binary sample $w_{i}^{s}$ by", + "bbox": [ + 212, + 191, + 787, + 359 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {s} = \\frac {1}{1 + \\exp \\left[ - \\frac {\\log \\left(\\frac {1 - p _ {i}}{p _ {i}}\\right) + g _ {1} ^ {s} - g _ {0} ^ {s}}{\\tau} \\right]} = \\sigma \\left(\\frac {l + t ^ {s}}{\\tau}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 369, + 785, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$g_{1}^{s}$ and $g_{0}^{s}$ are independent samples from standard Gumbel distribution, i.e., Gumbel(0,1), and $\\tau$ is the temperature controlling how close $w_{i}^{s}$ is to either 0 or 1. As shown by [25], $\\tau$ is often chosen from 0.01 to 0.2 to achieve good binary approximation and optimization stability. We define the logit $l = \\log \\left(\\frac{1 - p_i}{p_i}\\right)$ , $t^s = g_1^s - g_0^s$ , and $\\sigma(\\cdot)$ as the sigmoid function. It can be shown that $t^s$ is equally sampled from the standard logistic distribution (Logistic(0,1)) [25]. With a small temperature, $w_{i}^{s}$ is close to binary. The Gumbel-sigmoid reparameterization trick facilitates the use of gradient-based optimization while preserving the binary nature of the weights. Moreover, it enhances the expressiveness of the mask by utilizing the probabilities to indicate the relative importance of each region.", + "bbox": [ + 214, + 430, + 787, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.6 Overall Process", + "text_level": 1, + "bbox": [ + 215, + 604, + 392, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this paper, we incorporate a SAM-guided mask, learnable perturbation, and Gumbel-sigmoid reparameterization into Eq. (3) as follows:", + "bbox": [ + 212, + 630, + 782, + 661 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {p} ^ {*}, \\phi^ {*} = \\arg \\min _ {\\mathbf {p}, \\phi} \\mathbb {E} _ {\\mathbf {w} \\sim \\operatorname {B e r} (\\mathbf {p})} \\left[ U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {w} | | \\right] \\\\ \\approx \\arg \\min _ {\\mathbf {p}, \\phi} \\mathbb {E} _ {\\mathbf {t} \\sim \\text {L o g i s t i c} (0, 1)} \\left[ U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {w} | | \\right] \\tag {5} \\\\ \\text {w h e r e} \\quad \\mathbf {M} = \\sum_ {i = 1} ^ {N} \\sigma \\left(\\frac {\\log \\frac {1 - p _ {i}}{p _ {i}} + t}{\\tau}\\right) \\mathbf {M} _ {i} \\quad \\text {a n d} \\quad \\hat {\\mathbf {x}} = g _ {\\phi} (\\mathbf {x}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 674, + 784, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbb{E}$ represents the expectation, and $\\mathbf{p} = [p_1,p_2,\\dots ,p_n]^T$ are the probabilities for the Bernoulli distributions. $\\mathbf{M}_i$ is the binary segment mask pre-generated by SAM, and $g_{\\phi}(\\cdot)$ is a blur function parameterized by $\\phi$ . In this paper, we adopt a Gaussian blur function, and $\\phi$ will represent the standard deviation that governs", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the spread of the blur. To ensure $p_i \\in [0,1]$ , we apply a sigmoid function to $p_i$ denoted as $p_i = \\sigma(v_i)$ , and learn the real number $\\{v_i\\}_{i=1}^N$ instead. To solve the optimization problem, we initialize $v_i = 0, \\phi = 0$ for all indices $i$ and utilize gradient descent to update them. To estimate the expectation term, we leverage the stochastic optimization method. At each gradient step, we get one sample from Logistic(0,1) to update the parameters until convergence. Compared to directly solving Eq. (3), applying a SAM-guided mask could potentially speed up the optimization process since only a small set of weight parameters need to be learned. Adding the learnable perturbation parameter $\\phi$ will not theoretically increase complexity, as $\\phi$ can be updated in parallel with $\\{w_i\\}_{i=1}^N$ . If the SAM model is not available, our method can also be applied at the pixel level, treating $\\mathbf{M}_i$ as a segment solely for the $i$ -th pixel. Although pixel-wise UA is less interpretable by humans, it often contains more detailed information.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Eq. (5), the uncertainty function directly serves as the optimization objective, and the mask is learned to get the optimal uncertainty reduction, potentially yielding more accurate and faithful explanations. Compared to standard perturbation-based methods, the proposed framework avoids performing an exponential number of perturbations on the input and manually tuning the perturbation parameters. Perturbation-based methods are accurate only if the perturbations are representative. However, it is hard to randomly obtain representative perturbations for a large input space, such as high-resolution images.", + "bbox": [ + 212, + 343, + 787, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 488, + 375, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset. We evaluate the proposed method on the benchmark image classification datasets including CIFAR-10 (C10) [14], CIFAR-100 (C100) [13], and SVHN [24]. We also evaluate our method on high-quality PartImageNet [9].", + "bbox": [ + 212, + 522, + 785, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "UQ Model. We adopted the deep ensemble technique [16] for quantifying uncertainty. This method involves the training of several deep neural networks, each beginning from different initialization. Notably, this technique has proven highly effective in UQ, surpassing various approximate BNNs [16]. Specifically, the UQ model for each dataset consists of 5 ensemble components, each trained independently with random initialization. We use ResNet18 and ResNet152 for C10 and C100, a two-layer CNN for SVHN, and DinoV2 with fully connected layers for PartImageNet. The implementation details of model architecture and training hyperparameters for all datasets are provided in Appendix A.1.", + "bbox": [ + 212, + 568, + 787, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We compare our proposed method with both gradient-based and perturbation-based methods. Following the experiment settings of [39], the gradient-based baselines include Grad [31], SmoothGrad [32], IG [35], and UA-Backprop [39]. The perturbation-based baselines include KernelSHAP [21], LIME [28], and Occlusion [46]. Those methods, except UA-Backprop, are general approaches for analyzing the input-output relationship and can be directly extended to UA. They have been extended to UA in [26, 39], serving as standard baselines. It is worth noting that effectively applying SAM to gradient-based attribution methods is challenging due to the dependency of uncertainty on the entire input. Yet,", + "bbox": [ + 212, + 704, + 805, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SAM may fit well with some perturbation-based methods, allowing for the perturbation of segment combinations generated by SAM and avoiding the need to perturb thousands of pixel combinations. The results of the integration of KernelSHAP with SAM are also reported for the detection of problematic regions. Implementation details for all methods are shown in Appendix A.2.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Tasks. In Sec. 4.1, we assess all methods for detecting problematic regions on semi-synthetic data, where the ground-truth problematic regions are known. In Sec. 4.2, we perform faithfulness tests on the methods to observe the reduction in uncertainty achieved by altering only the identified problematic regions. The detailed experiment settings are provided in Appendix A.3.", + "bbox": [ + 212, + 222, + 784, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Detection of Problematic Regions", + "text_level": 1, + "bbox": [ + 214, + 318, + 540, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we aim to evaluate methods for detecting image anomalies using semi-synthetic data with known problematic areas. We introduce controlled mismatches that mimic anomalies by randomly replacing a patch in each test image with a corresponding patch from the training dataset at the same location, thereby generating ground-truth problematic regions. Knowing these regions allows us to quantitatively assess our method's precision in identifying artificially induced problematic regions. We conduct experiments on four datasets: C10, C100, SVHN, and PartImageNet. For the C10, C100, and SVHN datasets, we create ground-truth problematic regions by substituting a 10 by 10 patch in test images with a matching sample from the training data at the same location. For the high-resolution PartImageNet, we replace a 100 by 100 patch in each image. We then select 500 images from each dataset that exhibit the most significant increase in uncertainty after modification as our test subjects, indicating the presence of problematic regions. To identify these, we generate uncertainty maps for total uncertainty $U_{t}$ that illustrate each pixel's contribution to it. We use these maps to locate the problematic regions by fitting a bounding box of the same size as the ground truth (10 by 10 for C10, C100, and SVHN; 100 by 100 for PartImageNet) around the area with the highest average attribution score, employing a straightforward brute-force search method to pinpoint the most likely location of the altered patch. Besides qualitative evaluation, our quantitative metrics include Intersection over Union (IoU) and anomaly detection accuracy (ADA). IoU calculates the overlap percentage between detected and actual anomalies divided by the union of these regions, while ADA measures the percentage of detections with an IoU above 0.5.", + "bbox": [ + 212, + 342, + 787, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative Evaluation. The examples of uncertainty maps for all methods are shown in Figure 1. We observe that our method can clearly detect problematic regions that match the ground truth. It is worth noting that gradient-based methods, such as Grad, SmoothGrad, IG, and UA-Backprop, usually produce noisy \"scattered\" visualizations. Although other perturbation-based baselines may yield clear regions, they often lead to misleading results by pointing to inaccurate areas. Appendix B will provide additional examples across datasets, uncertainty maps for real images with high uncertainty, and some failure cases.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 303, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f94464d8baa5541c8feece0a2f2a807d4511a40a9c9acf32002b1d6b208cefa9.jpg", + "image_caption": [ + "Fig. 1: Examples of uncertainty maps for identifying problematic regions on PartImageNet. Darker regions indicate a higher contribution to the uncertainty." + ], + "image_footnote": [], + "bbox": [ + 250, + 141, + 753, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Results. The experiment results are shown in Table 1. Our method significantly outperforms all baselines in detecting problematic regions across all datasets. For instance, on the SVHN dataset, our method achieves an IoU of 0.829 and an ADA of 0.830, marking a considerable improvement of approximately $50\\%$ in both IoU and ADA over LIME and Occlusion. It is worth noting that perturbation-based methods generally perform better than gradient-based methods in detecting problematic regions. When KernelSHAP is enhanced with SAM (KernelSHAP + SAM), it shows notable improvements, potentially because SAM provides valuable prior knowledge about potential problematic areas, offering a solid foundation for perturbation-based methods to identify candidates for perturbation. However, our method still surpasses KernelSHAP + SAM. If SAM is not available, our method can be applied at the pixel level, as discussed in Sec. 3.6. To avoid any confusion, results comparing our method with other baselines at the pixel level will be provided in Appendix C.1. Our method still consistently outperforms other baselines in the pixel-level setting.", + "bbox": [ + 212, + 613, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e1882e2cc3500b5cfe0277e77c2f33027f7489be9253bfa03d43f2f7a06f6dd2.jpg", + "table_caption": [ + "Table 1: IoU $\\uparrow$ and ADA $\\uparrow$ for detection of problematic regions. Bold values indicate the best performance." + ], + "table_footnote": [], + "table_body": "
MethodC10C100SVHNPartImageNet
IoUADAIoUADAIoUADAIoUADA
Grad0.1090.0540.2820.2500.2880.1800.1470.065
SmoothGrad0.2970.2260.3220.1320.3120.1840.1610.075
IG0.2080.1520.2240.2500.3100.2560.4870.505
UA-Backprop0.2710.1780.3600.2500.3360.2580.6190.695
KernelSHAP0.3240.3060.2860.2500.3820.340--
KernelSHAP + SAM0.6760.6700.6670.6620.7980.7940.3220.280
LIME0.2810.2580.5530.5000.3750.3300.2420.190
Occlusion0.2180.1600.2450.2500.3500.3020.3010.240
Ours0.7130.7000.7500.7500.8290.8300.6330.635
", + "bbox": [ + 245, + 183, + 756, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Faithfulness Test", + "text_level": 1, + "bbox": [ + 215, + 383, + 401, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Faithfulness in uncertainty attribution quantifies the accuracy with which a method's explanations reflect the actual influence of input features on the model's uncertainty. A method exhibits high faithfulness if the regions it identifies as significant truly correspond to areas in the input that, upon alteration, substantially affect the model's uncertainty. In this section, we will refine the most problematic pixels of the input, for example, by updating $2\\%$ of the pixels with the highest UA scores, and then observe the reduction in uncertainty after the alteration. To select the pixels with the highest UA scores, we compare all baselines at the pixel level. This is necessary because if one segment shares one score, the relative importance of each pixel remains unknown, preventing us from selecting the most problematic pixels.", + "bbox": [ + 212, + 414, + 787, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The experiments are conducted on images with high uncertainty from the C10, C100, and SVHN testing datasets. Specifically, we select 500 samples with the highest total uncertainty from the testing data for each dataset. We exclude PartImageNet from the comparison because KernelSHAP and LIME are difficult to apply to high-resolution images at the pixel level. The linear regression used in these methods requires that the number of perturbations be at least larger than the number of pixels, which takes a significant amount of time and substantial memory. We consider two methods to refine the original inputs in their problematic regions. The first method utilizes gradient descent with uncertainty as the sole loss function. Specifically, we iteratively update the input $\\mathbf{x}$ through $\\mathbf{x} = \\mathbf{x} + \\eta \\cdot \\mathbf{M}(\\mathbf{x}) \\cdot \\frac{\\partial U(\\mathbf{x})}{\\partial \\mathbf{x}}$ where $\\eta$ is the learning rate, $\\mathbf{M}(\\mathbf{x})$ is a binary map of the same size as $x$ which highlights the pixels (i.e., $1\\%$ of the total pixels) with largest UA score, and $U(\\cdot)$ represents the function for estimating uncertainty. Through iteratively gradient descent, the uncertainty should be significantly reduced if $\\mathbf{M}(\\mathbf{x})$ shows the true problematic areas. The second method directly blurs the most problematic regions indicated by $\\mathbf{M}(\\mathbf{x})$ , (i.e., $2\\%$ of the total pixels) following the experiment settings of [39] for the blurring test. After alter", + "bbox": [ + 212, + 580, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 303, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fd37f3dadad16c9aac693bddf7c848014d3bade56d067691fa152d3eb002b14d.jpg", + "table_caption": [ + "Table 2: We evaluate attribution performance in terms of URR $\\uparrow$ across four datasets, refining images by targeting the top $0.5\\%$ or $1\\%$ of pixels with high total uncertainty using iterative gradient descent, and the top $2\\%$ or $5\\%$ of pixels using blurring techniques. Bold values denote the best performance." + ], + "table_footnote": [], + "table_body": "
MethodRefining using Iterative Gradient Descent
C10C100SVHNAvg. Performance
0.5%1%0.5%1%0.5%1%0.5% + 1%
Grad0.6620.8850.6150.7780.6810.8080.738
SmoothGrad0.6230.8630.5780.7600.7250.8600.735
IG0.6210.8650.5810.7610.7000.8380.728
UA-Backprop0.7060.9030.2610.4360.7140.8530.646
KernelSHAP0.7310.9040.6350.7790.7330.8380.770
LIME0.7870.9270.4580.5730.2560.4110.569
Occlusion0.8220.9400.2970.4530.4200.5920.587
Ours0.8250.9240.7660.8080.9880.9920.884
MethodRefining using Gaussian Blurring
C10C100SVHNAvg. Performance
2%5%2%5%2%5%2% + 5%
Grad0.3430.4400.0990.1360.0530.0310.184
SmoothGrad0.3310.4180.0790.1260.1090.1050.195
IG0.3920.4810.0960.1410.0320.0660.201
UA-Backprop0.3740.4530.0880.1180.0760.1510.210
KernelSHAP0.8230.9020.3190.4480.1220.0980.452
LIME0.3780.3550.4150.5470.1310.2220.341
Occlusion0.7300.7860.2970.3900.4230.5390.528
Ours0.8600.9180.3860.4870.8690.9390.743
", + "bbox": [ + 259, + 210, + 743, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ing the original input $\\mathbf{x}$ to get the updated input $\\tilde{\\mathbf{x}}$ , we compare the uncertainty of them by calculating the Uncertainty Reduction Rate (URR): $1 - \\frac{U(\\tilde{\\mathbf{x}})}{U(\\mathbf{x})}$ . For gradient descent refinement, we aim to update $0.5\\%$ and $1\\%$ of the total pixels, whereas for the blurring test, we refine $2\\%$ and $5\\%$ of the total pixels. More pixels are refined in the blurring test to achieve noticeable uncertainty reduction, as blurring is not as effective as directly optimizing uncertainty to attain the optimal value. The detailed settings will be shown in Appendix A.3.", + "bbox": [ + 212, + 579, + 787, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The experiment results are shown in Table 2. our method stands out remarkably, achieving the highest average URR in both settings. Specifically, with an impressive average URR of 0.884 using iterative gradient descent, our method outperforms its closest competitor, KernelSHAP, by an impressive $11.4\\%$ . Similarly, in the blurring refinement, our method's average URR significantly surpasses that of Occlusion, the next best performer, by $21.5\\%$ . It is worth noting that perturbation-based LIME and Occlusion perform a little bit worse than gradient-based methods if the refinement is conducted using gradient descent. This may be because gradient-based methods provide the directions in terms of gradients for refining the images. If the refinement is conducted using blur-", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5760238ee87a06acdb9b13332c028ab90ec7152978b1e95fe394b79ff1188dfb.jpg", + "table_caption": [ + "Table 3: Problematic region detection results for the contribution of each component." + ], + "table_footnote": [], + "table_body": "
MethodC10SVHN
IoUADAIoUADA
Eq. (3)0.3330.2940.3710.325
Eq. (3) + SAM0.6480.6380.7090.712
Eq. (3) + SAM + LP0.6750.6640.7650.764
Eq. (3) + SAM + LP + Gumbel0.7130.7000.8290.832
", + "bbox": [ + 300, + 169, + 702, + 265 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ring, perturbation-based methods usually works better since they consider the blurring perturbation in the attribution process to estimate the contribution of each pixel. Overall, our method consistently outputs other baselines, leveraging both advantages of perturbation-based and gradient-based methods for enhanced faithfulness.", + "bbox": [ + 212, + 292, + 787, + 369 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 391, + 401, + 405 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Contribution of Each Component. In this section, we aim to highlight the contribution of each component to the performance in detecting problematic regions. The results are presented in Table 3. The method, described as \"Eq. (3)\", solves Eq. (3) by relaxing $\\mathbf{M}$ into a probability mask with a sigmoid function, $\\mathbf{M} = \\sigma (\\mathbf{t})\\in [0,1]^{\\left|\\mathbf{M}\\right|}$ , where $\\mathbf{t}\\in \\mathcal{R}^{\\left|\\mathbf{M}\\right|}$ are the learnable parameters. The results indicate that all three components positively contribute to the improvement in detecting problematic regions. SAM contributes the most significantly, and Gumbel-sigmoid reparameterization is the second most significant contributor. More analysis will be shown in Appendix C.2.", + "bbox": [ + 212, + 416, + 787, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hyperparameter Sensitivity. Two hyperparameters, $\\lambda$ and $\\tau$ , are essential for tuning our method. $\\lambda$ balances uncertainty loss with sparsity loss in Eq. (3), typically chosen to prioritize uncertainty loss over sparsity regularization. $\\tau$ in Eq. (4) adjusts the approximation level for a binary mask, with a small $\\tau$ (0.05 to 0.2) effectively achieving good binary approximations. Empirical results in Table 4 demonstrate that the performance is relatively stable across a range of these hyperparameters. Results for more datasets will be shown in Appendix C.3.", + "bbox": [ + 212, + 575, + 539, + 772 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Complexity Analysis. The empirical runtime for attributing one input is shown in Table 5. While our proposed methods can", + "bbox": [ + 212, + 794, + 539, + 839 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/618d39a19a70c0335600494349d6f558a230a95101a6ab575ce76c62a526df8a.jpg", + "table_caption": [ + "Table 4: Hyperparameter sensitivity for detecting problematic regions on C10 dataset." + ], + "table_footnote": [], + "table_body": "
MethodC10\nIoU ADA
λ = 3, σ = 0.10.713 0.700
λ = 3, σ = 0.070.679 0.686
λ = 3, σ = 0.150.688 0.682
λ = 3, σ = 0.10.713 0.700
λ = 2.5, σ = 0.10.677 0.680
λ = 2, σ = 0.10.682 0.680
", + "bbox": [ + 573, + 660, + 764, + 785 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 302, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "achieve better UA performance, their runtime complexity depends on the number of gradient descent iterations needed for convergence when solving (5).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Typically, about 100-300 forward and backward passes are needed to generate one uncertainty map. Further analysis of the convergence of our method will be provided in Appendix C.4. Among the baselines, Grad is the most efficient, requiring only one forward and backward propagation. The time complexity of SmoothGrad and IG depends on the number of samples used for aggregating the gradients, which usually involves hundreds of samples. Although UA-Backprop theoretically only requires one forward and backward pass, in practice, it usually needs one pass per class. Compared", + "bbox": [ + 212, + 176, + 468, + 448 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/cd3731701d6c7a5e24aba3fb1b427b7764cc4217a0f1f120f62d65a06d6b767f.jpg", + "table_caption": [ + "Table 5: Average runtime of attributing one input in seconds (s)." + ], + "table_footnote": [], + "table_body": "
MethodC10RuntimePartImageNetRuntime
Grad0.02s5.94s
SmoothGrad2.10s33.3s
IG2.14s35.6s
UA-Backprop0.26s9.75s
KernelSHAP1.63s>200s
KernelSHAP + SAM1.08s113s
LIME1.16s>200s
Occlusion0.15s>200s
Ours1.54s103s
", + "bbox": [ + 480, + 247, + 787, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "to perturbation-based methods such as KernelSHAP and LIME, our method generally requires less time when applied to high-resolution images. KernelSHAP and LIME encounter challenges due to their linear regression component, which demands a sufficient number of perturbations and can be slow when dealing with a high variable dimension. It is worth noting that our method scales well with large images and can generate pixel-level UA maps for these images with complexity that increases linearly.", + "bbox": [ + 212, + 449, + 787, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 579, + 359, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This research aims to develop a more faithful and explainable uncertainty attribution method for identifying problematic regions of the input. We formulate UA as an optimization problem to learn a binary mask that indicates problematic regions, thus enhancing our method's faithfulness by directly optimizing uncertainty as the objective. Three key enhancements are introduced. First, we propose SAM-based mask parameterization to generate more human-understandable uncertainty maps, potentially making the optimization process more efficient. Second, we introduce learnable perturbations to adaptively learn the perturbation parameters without the need for manual tuning. Third, we apply a probabilistic approach to learn the binary mask through Gumbel-sigmoid reparameterization under continuous optimization. For evaluation, we test our method's ability to detect synthetic problematic regions and estimate its faithfulness with real, problematic images, demonstrating our method's superior performance. We also discuss the limitations and future work of our method in Appendix D.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the National Science Foundation award IIS 2236026 and in part by the Rensselaer-IBM AI Research Collaboration (http://airc.rpi.edu), part of the IBM AI Horizons Network.", + "bbox": [ + 215, + 146, + 787, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 217, + 321, + 233 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Antorán, J., Bhatt, U., Adel, T., Weller, A., Hernández-Lobato, J.M.: Getting a clue: A method for explaining uncertainty estimates. arXiv preprint arXiv:2006.06848 (2020)", + "2. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015)", + "3. Dabkowski, P., Gal, Y.: Real time image saliency for black box classifiers. Advances in neural information processing systems 30 (2017)", + "4. Depeweg, S., Hernandez-Lobato, J.M., Doshi-Velez, F., Udluft, S.: Decomposition of uncertainty in bayesian deep learning for efficient and risk-sensitive learning. In: International Conference on Machine Learning. pp. 1184-1193. PMLR (2018)", + "5. Fong, R., Patrick, M., Vedaldi, A.: Understanding deep networks via extremal perturbations and smooth masks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2950-2958 (2019)", + "6. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017)", + "7. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020)", + "8. Guo, H., Wang, H., Ji, Q.: Uncertainty-guided probabilistic transformer for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20052-20061 (2022)", + "9. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933 (2021)", + "0. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)", + "1. Kapishnikov, A., Venugopalan, S., Avci, B., Wedin, B., Terry, M., Bolukbasi, T.: Guided integrated gradients: An adaptive path method for removing noise. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5050-5058 (2021)", + "2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "3. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "4. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html 55(5) (2014)", + "5. Kwon, Y., Won, J.H., Kim, B.J., Paik, M.C.: Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis 142, 106816 (2020)" + ], + "bbox": [ + 225, + 250, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 303, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.01474", + "17. Ley, D., Bhatt, U., Weller, A.: $\\{\\backslash$ delta\\} -clue: Diverse sets of explanations for uncertainty estimates. arXiv preprint arXiv:2104.06323 (2021)", + "18. Ley, D., Bhatt, U., Weller, A.: Diverse, global and amortised counterfactual explanations for uncertainty estimates. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7390-7398 (2022)", + "19. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017)", + "20. Lundberg, S.M., Erion, G.G., Lee, S.I.: Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:1802.03888 (2018)", + "21. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)", + "22. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf", + "23. Montavon, G., Lapuschkin, S., Binder, A., Samek, W., Müller, K.R.: Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition 65, 211-222 (2017)", + "24. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)", + "25. Ng, I., Zhu, S., Fang, Z., Li, H., Chen, Z., Wang, J.: Masked gradient-based causal structure learning. In: Proceedings of the 2022 SIAM International Conference on Data Mining (SDM). pp. 424-432. SIAM (2022)", + "26. Perez, I., Skalski, P., Barns-Graham, A., Wong, J., Sutton, D.: Attribution of predictive uncertainties in classification models. In: The 38th Conference on Uncertainty in Artificial Intelligence (2022)", + "27. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)", + "28. Ribeiro, M.T., Singh, S., Guestrin, C.: \" why should i trust you?\" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016)", + "29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "30. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017)", + "31. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013)", + "32. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825 (2017)", + "33. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019)", + "35. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017)", + "36. Tierney, L.: Markov chains for exploring posterior distributions. the Annals of Statistics pp. 1701-1728 (1994)", + "37. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019)", + "38. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023)", + "39. Wang, H., Joshi, D., Wang, S., Ji, Q.: Gradient-based uncertainty attribution for explainable bayesian deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12044-12053 (2023)", + "40. Wang, H., Wang, S., Ji, Q.: Semantic attribution for explainable uncertainty quantification. In: International Workshop on Epistemic Uncertainty in Artificial Intelligence. pp. 101-112. Springer (2023)", + "41. Watson, D., O'Hara, J., Tax, N., Mudd, R., Guy, I.: Explaining predictive uncertainty with information theoretic shapley values. Advances in Neural Information Processing Systems 36 (2024)", + "42. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics. In: Proceedings of the 28th international conference on machine learning (ICML-11). pp. 681-688 (2011)", + "43. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020)", + "44. Xu, S., Venugopalan, S., Sundararajan, M.: Attribution in scale and space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9680-9689 (2020)", + "45. Yang, Q., Zhu, X., Fwu, J.K., Ye, Y., You, G., Zhu, Y.: Mfpp: Morphological fragmental perturbation pyramid for black-box model explanations. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 1376-1383. IEEE (2021)", + "46. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European conference on computer vision. pp. 818-833. Springer (2014)", + "47. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019)", + "48. Zhang, Y., Wang, H., Kephart, J.O., Ji, Q.: Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9032 (2023)" + ], + "bbox": [ + 212, + 146, + 785, + 674 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Optimization-based UA Via Learning Informative Perturbations", + "bbox": [ + 303, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_model.json b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a8419e388d95f010b239addda2d44f1c66215807 --- /dev/null +++ b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_model.json @@ -0,0 +1,1917 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Optimization-based Uncertainty Attribution Via Learning Informative Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.212, + 0.714, + 0.228 + ], + "angle": 0, + "content": "Hanjing Wang\\(^{1}\\), Bashirul Azam Biswas\\(^{1}\\), and Qiang Ji\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.24, + 0.684, + 0.269 + ], + "angle": 0, + "content": "Rensselaer Polytechnic Institute, Troy NY 12180, USA {wangh36, biswab, jiq}@rpi.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.298, + 0.744, + 0.522 + ], + "angle": 0, + "content": "Abstract. Uncertainty attribution (UA) aims to identify key contributors to predictive uncertainty in deep learning models. To improve the faithfulness of existing UA methods, we formulate UA as an optimization problem to learn a binary mask on the input. The learned mask identifies regions that significantly contribute to output uncertainty and allows uncertainty reduction through learning informative perturbations on the masked input. Our method enhances UA interpretability and maintains high efficiency by integrating three key improvements: Segment Anything Model (SAM)-guided mask parameterization for efficient and interpretable mask learning; learnable perturbations that adaptively target and refine problematic regions specific to each input without manually tuning the perturbation parameters; and a novel application of Gumbel-sigmoid reparameterization for efficiently learning Bernoulli-distributed binary masks under continuous optimization. Our experiments on problematic region detection and faithfulness tests demonstrate our method's superiority over state-of-the-art UA methods." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.533, + 0.668, + 0.548 + ], + "angle": 0, + "content": "Keywords: Uncertainty Attribution \\(\\cdot\\) Model Explainability" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.377, + 0.587 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "While conventional deep learning has made remarkable strides in various domains, they often face challenges with overconfidence in their predictions, which can be risky in safety-critical applications. Hence, uncertainty quantification (UQ) is paramount, as it directly impacts the safety, reliability, and trustworthiness of AI systems, especially for various computer vision applications [8,15,48]. Predictive uncertainties in deep learning are categorized into aleatoric uncertainty, reflecting inherent data randomness, and epistemic uncertainty, resulting from limited model knowledge. Bayesian Neural Networks (BNNs) serve as a foundational approach for quantifying uncertainty, encompassing techniques such as Variational Inference (VI) [7, 19, 22] and Markov Chain Monte Carlo (MCMC) [36, 42, 47] methods to estimate the complex posterior distribution of model parameters. Following this, Ensemble methods [16, 37, 38, 43] offer an alternative strategy, aggregating predictions from multiple models to capture uncertainty. While numerous studies have focused on enhancing the efficiency and accuracy of UQ, comparatively little effort has been devoted to understanding the sources of uncertainty. Identifying the sources of uncertainty is essential," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "as it reveals where and why certain inputs are associated with high uncertainty for enhanced explainability, thereby informing more targeted improvements in model design and training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Uncertainty attribution (UA) [1, 26, 39, 40] systematically identifies and analyzes the specific inputs or features that contribute significantly to the uncertainty in the model's predictions. Traditional explainable AI methods highlight areas of inputs leading to specific predictions, which are unreliable when applied directly to UA for identifying problematic regions in input data. The primary reason is that they focus on explaining the model's predictions, assuming confident predictions, while UA aims to pinpoint the sources of input imperfections that contribute to high predictive uncertainties. Additionally, these methods are often class-discriminative when applied to explain the classification score. Consequently, they may struggle to explain inputs with incorrect predictions and large uncertainties [26]. Uncertainty attribution methods are broadly categorized into two types: gradient-based and perturbation-based methods. Gradient-based methods leverage the gradients of the model output with respect to the input to identify the areas contributing to uncertainty. On the other hand, perturbation-based methods involve systematically altering parts of the input and observing the effect on uncertainty, allowing for a more direct assessment of how variations in the input influence the uncertainty. While gradient-based UA methods are often efficient, they only consider the uncertainty sensitivity to very small input changes, leading to noisy and sometimes misleading indications of uncertainty. Perturbation-based methods can be computationally intensive and may be less accurate if an insufficient number of perturbations are performed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.791 + ], + "angle": 0, + "content": "Inspired by existing literature [3,5,45], we treat UA as an optimization task aimed at learning a binary mask to identify problematic regions of the input. This mask targets specific areas for maximum uncertainty reduction, achieved by applying learnable perturbations to these selected areas. This optimization-based UA framework can potentially improve the faithfulness of UA, which quantifies how accurately a method's explanations reflect the real impact of input features on the uncertainty. A method is considered faithful if the identified areas truly correspond to regions in the input that, when perturbed, have a significant impact on the predictive uncertainty. Our approach ensures faithfulness through a direct linkage between the uncertainty and the identification of uncertain regions in the optimization objective. Moreover, compared to perturbation-based methods, our approach utilizes gradients to iteratively update the mask through continuous optimization, avoiding the need for computationally intensive, pixelwise input perturbations. This strategy may reduce the cost associated with exploring an exponential number of pixel combinations. Unlike gradient-based methods, which directly employ gradients to generate uncertainty maps, our method applies actual perturbations to inputs, not just minor input changes, potentially yielding more accurate explanations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To enhance UA performance, we introduce three key adjustments. First, we use a linear combination of segments from the pre-trained Segment Anything Model (SAM) [12] to parameterize the mask. Using SAM segments as a basis" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.304, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "enables targeted analysis of semantic regional effects and simplifies the learning process to linear combination coefficients instead of pixel-level masks. Second, traditional perturbations used to explain model predictions often fall short in UA. For example, widely used mask-out perturbations invariably result in increased uncertainty in UA, even when applied to problematic regions. This underscores the need for advanced, learnable perturbations in UA, which are essential for accurately reducing uncertainty by modifying problematic regions. Finally, given the challenge of learning a binary mask directly, we adopt a Bernoulli distribution for the masks and employ the Gumbel-sigmoid reparameterization trick [10], enabling continuous optimization with high efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.388, + 0.34 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.789, + 0.706 + ], + "angle": 0, + "content": "Gradient-based Attribution Methods. Gradient-based methods [11,29,31-35,44,46] leverage gradient information as input attribution. However, raw gradients can be noisy, leading researchers to develop various approaches for refining gradient information. SmoothGrad [32] addresses the issue of noisy gradients by aggregating attributions from multiple noisy images. FullGrad [34] combines the gradients of the output with respect to both the input and the bias variable in each convolutional or fully-connected layer. Integrated Gradients (IG) [35] and its variants [11, 44] create a path integral between a reference image and the input. They then aggregate the gradients of the model output for the artificial images along this path, providing a more comprehensive understanding of the input-output relationship. Grad-CAM [29] focuses on the gradients of the latent feature maps, offering more interpretable insights. Layer-wise relevance propagation (LRP) methods [2, 23, 30] construct attributions by backpropagating the relevance score layer-by-layer from the output to the input. While most gradient-based methods can be applied for both model prediction and uncertainty attribution, LRP methods often need specific neural network architectures [39]. The direct application of these attribution methods might not be sufficient for UA, leading Wang et al. [39] to develop a backpropagation method specifically for UA. In summary, gradient-based methods are often computationally efficient to analyze neural network input-output relationships. Yet, they mainly model output changes from very small input perturbations, which may not yield precise UA results. This limitation is not effectively overcome by refining gradient information through smoothing, aggregation, and path integration techniques." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Perturbation-based Attribution Methods. Perturbation-based methods [3, 5, 6, 27, 45, 46] estimate attributions by modifying the input and observing the subsequent changes in the model's output. For example, LIME [28] approximates the model's behavior locally by fitting an interpretable model to the perturbed data. RISE [27] generates saliency maps by randomly masking the input and aggregating the model's responses. SHAP and its variants [20, 21] leverage Shapley values from cooperative game theory to interpret complex model predictions, assigning importance to each feature by observing output changes upon feature" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.404 + ], + "angle": 0, + "content": "perturbations. The SHAP-based methods are extended to UA with theoretical studies [41]. While perturbation-based methods provide a more intuitive understanding of the model's behavior and allow for better interpretability, they can also be computationally expensive. This is primarily due to the need for an exponentially increasing number of perturbations and subsequent model evaluations to explore the input-output relationship effectively. To improve efficiency, Dabkowski et al. [3], Fong et al. [5], and Yang et al. [45] focus on identifying the smallest regions for perturbation that can significantly change the model's prediction, leveraging continuous optimization techniques to speed up the attribution process. In recent years, several methods have been specifically designed for UA. For instance, CLUE [1] and its variants [17,18] focus on creating an improved image with minimal uncertainty by modifying the uncertain input using a generative model. The attribution map is then generated by measuring the difference between the original and modified input. To further enhance pixel-wise attributions, Perez et al. [26] combined the CLUE method with the path integral technique. However, CLUE-based methods require a pre-trained generative model, which may be challenging to acquire for practical applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.427, + 0.454 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.478, + 0.362, + 0.494 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Bayesian Neural Networks. Distinct from point-estimated neural networks, BNNs treat neural network parameters as random variables and calculate the posterior distribution of the parameters \\(\\pmb{\\theta}\\), denoted as \\(p(\\pmb{\\theta} \\mid \\mathcal{D})\\) where \\(\\mathcal{D}\\) represents the training data. Utilizing Bayes' rule, the posterior distribution can be computed using \\(p(\\pmb{\\theta} \\mid \\mathcal{D}) = \\frac{p(\\mathcal{D}|\\pmb{\\theta})p(\\pmb{\\theta})}{p(\\mathcal{D})}\\) with \\(p(\\mathcal{D}|\\pmb{\\theta})\\) as the likelihood and \\(p(\\pmb{\\theta})\\) as the prior. Since the exact posterior distribution is often intractable, various techniques have been developed, including MCMC sampling, VI and Ensemble methods. These approaches either directly sample from the posterior or approximate it using a simpler distribution for efficient sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Uncertainty Quantification. In this study, we concentrate on image classification tasks, where \\(\\mathbf{x}\\) denotes the input and \\(\\mathbf{y}\\) denotes the output. The neural network predicts \\(p(\\mathbf{y}|\\mathbf{x},\\boldsymbol{\\theta})\\), which is a categorical distribution parameterized by learnable softmax probability. Entropy is a common method for quantifying the uncertainty or randomness in a probability distribution. In entropy-based UQ, the total uncertainty of predictions is decomposed as follows [4]:" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.758, + 0.787, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x} , \\mathcal {D}) ]} _ {\\text {T o t a l U n c e r t a i n t y} U _ {t} (\\mathbf {x})} = \\underbrace {\\mathcal {I} [ \\mathbf {y} , \\boldsymbol {\\theta} | \\mathbf {x} , \\mathcal {D} ]} _ {\\text {E p i s t e m i c U n c e r t a i n t y}} + \\underbrace {\\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} \\left[ \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x} , \\boldsymbol {\\theta}) ] \\right]} _ {\\text {A l e a t o r i c U n c e r t a i n t y} U _ {a} (\\mathbf {x})} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{H}\\), \\(\\mathcal{I}\\), and \\(\\mathbb{E}\\) stand for entropy, mutual information, and expectation, respectively. By employing a Monte Carlo approximation of the posterior, we can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.148, + 0.461, + 0.162 + ], + "angle": 0, + "content": "approximate the uncertainties as:" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.174, + 0.786, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\mathcal {D}) ] = \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta}) ] \\right] \\approx \\mathcal {H} \\left[ \\frac {1}{S} \\sum_ {s = 1} ^ {S} p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta} ^ {s}) \\right] \\tag {2a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.218, + 0.786, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} \\left[ \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta}) ] \\right] \\approx \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta} ^ {s}) ] \\tag {2b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.299 + ], + "angle": 0, + "content": "where \\(\\pmb{\\theta}^s\\sim p(\\pmb {\\theta}|\\mathcal{D})\\) represents the sampling of parameters \\(\\pmb{\\theta}\\) from their posterior distribution, and \\(S\\) is the sample size." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.32, + 0.628, + 0.336 + ], + "angle": 0, + "content": "3.2 Optimization-based Uncertainty Attribution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Our approach defines UA as an optimization problem, aiming to learn a binary mask, \\(\\mathbf{M}\\), with the same dimensions as \\(\\mathbf{x}\\), that highlights areas in the inputs significantly contributing to uncertainty. For each element \\(u\\) in \\(\\mathbf{x}\\), we assume \\(\\mathbf{M}(u) \\in \\{0,1\\}\\), with \\(\\mathbf{M}(u) = 1\\) indicating strong contributions to predictive uncertainty and \\(\\mathbf{M}(u) = 0\\) otherwise. To estimate each element's contribution, we perturb the regions indicated by \\(\\mathbf{M}\\). Specifically, with \\(\\hat{\\mathbf{x}}\\) representing the perturbed input with reduced uncertainty, we replace the regions where \\(\\mathbf{M}(u) = 1\\) in \\(\\mathbf{x}\\) with those from \\(\\hat{\\mathbf{x}}\\), resulting in \\(\\mathbf{x}' = (1 - \\mathbf{M}) \\odot \\mathbf{x} + \\mathbf{M} \\odot \\hat{\\mathbf{x}}\\) to denote the masked input. The goal is to identify a small subset of \\(\\mathbf{x}\\)'s regions, \\(\\mathbf{M}\\), that significantly reduce predictive uncertainty through the masked input. The uncertainty, denoted as \\(U\\), could be aleatoric \\((U_a)\\), epistemic \\((U_e)\\), or total \\((U_t)\\), as indicated in Eq. (1). The optimal \\(\\mathbf{M}^*\\) is determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.542, + 0.786, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {*} = \\arg \\min _ {\\mathbf {M}} U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {M} | | \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Here, \\(\\odot\\) denotes element-wise multiplication, \\(||\\mathbf{M}||\\) is the norm of \\(\\mathbf{M}\\), and \\(\\lambda\\) is a hyperparameter balancing uncertainty reduction and mask sparsity, usually prioritizing uncertainty loss over sparsity regularization. As \\(U\\) utilizes all ensemble models sampled from \\(p(\\boldsymbol{\\theta}|\\mathcal{D})\\) to compute uncertainty, the UA maps are learned by jointly backpropagating through these models. Perturbations in identified problematic regions decrease predictive uncertainty, while those in high-certain areas will increase uncertainty, guiding \\(\\mathbf{M}\\) to select critical regions for perturbation to reduce uncertainty. Background or non-informative regions will not impact uncertainty after perturbation, necessitating the inclusion of a sparsity term to ensure these areas are assigned zero values in \\(\\mathbf{M}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.564, + 0.756 + ], + "angle": 0, + "content": "3.3 SAM-guided Mask Parameterization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To enhance the explainability of our proposed method and reduce the optimization complexity, we aim to parameterize the uncertainty mask, \\(\\mathbf{M}\\), through a linear combination of segments derived from the pre-trained SAM model, denoted as \\(\\{\\mathbf{M}_i\\}_{i=1}^N\\). These \\(N\\) segments are binary masks of the same dimension as \\(\\mathbf{M}\\), each highlighting distinct regions of the input space. Generated by the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "SAM model, these segments are inherently binary and delineate areas corresponding to semantically meaningful and human-understandable concepts. The parameterization of \\(\\mathbf{M}\\) is \\(\\mathbf{M} = \\sum_{i=1}^{N} w_i \\mathbf{M}_i\\) where \\(w_i \\in \\{0,1\\}\\) represents the binary weights. By employing these segments as the basis for constructing \\(\\mathbf{M}\\), our method not only identifies regions contributing to uncertainty but also ensures that these regions are aligned with recognizable semantic features. This alignment significantly boosts the explainability of our method, as the identified areas of uncertainty are directly tied to concepts that are easily interpretable by humans. For instance, in an image classification task, rather than indicating abstract or non-intuitive areas of uncertainty, our method could highlight specific objects or parts of an image that contribute to the uncertainty, such as \"eyes\" in a facial recognition task or \"vehicles\" in an autonomous driving context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.435 + ], + "angle": 0, + "content": "While the effectiveness of our methods depends on the accuracy of SAM, we believe our approach can still be effective with imperfect SAM segmentation. For instance, if problematic regions are either included in a large SAM segment or spread across multiple segments, our method can sometimes still identify these areas as potentially problematic. Furthermore, our method is compatible with state-of-the-art segmentation models, ensuring adaptability and improved performance across diverse applications with domain-specific segmentation models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.455, + 0.457, + 0.47 + ], + "angle": 0, + "content": "3.4 Learnable Perturbation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.787, + 0.554 + ], + "angle": 0, + "content": "Traditional perturbation-based methods often use fixed or random perturbations, which may not effectively reduce uncertainty and could unintentionally increase it, even in problematic regions. Therefore, we dynamically adjust learnable perturbations during the optimization process, specifically targeting and modifying areas in the input \\(\\mathbf{x}\\) to effectively reduce uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Our approach updates the perturbed input \\(\\hat{\\mathbf{x}}\\), as described in Eq. (3), through iterative blurring, represented as \\(\\hat{\\mathbf{x}} = g_{\\phi}(\\mathbf{x})\\), where \\(g_{\\phi}(\\cdot)\\) is a blurring function parameterized by \\(\\phi\\). This strategy hypothesizes that blurring problematic regions decreases uncertainty, while blurring good areas could increase it by obscuring important information for prediction. Recognizing that a universal blurring strategy does not cater to the varied demands of different inputs, we aim to learn the blurring function's parameters \\(\\phi\\) within the optimization process. This approach allows for the precise, dynamic adjustment of perturbations, blurring each region as needed to optimally reduce uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.733, + 0.727 + ], + "angle": 0, + "content": "3.5 Gumbel-sigmoid Reparameterization For Binary Weights" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To address the challenge of modeling binary weights for our uncertainty mask, \\(\\mathbf{M}\\), we confront the complexities of discrete optimization, which demands extensive search efforts and complicates direct learning of these weights. While continuous optimization through iterative gradient descent offers a potential solution, simply relaxing binary weights to a continuous range between [0, 1] undermines the mask's binary nature, which is crucial for pinpointing problematic areas and facilitating accurate input region replacement through perturbations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.304, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "On one hand, adhering strictly to binary weights ensures clarity in identifying areas contributing to uncertainty but, on the other hand, it restricts our ability to convey the nuanced relative importance of these regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.36 + ], + "angle": 0, + "content": "In this paper, we propose maintaining the binary conception of weights while adopting a probabilistic approach for learning them. Specifically, we model each weight as an independent Bernoulli distribution, denoted as \\( w_{i} \\sim \\mathrm{Ber}(p_{i}) \\) with probability parameter \\( p_i \\in [0,1] \\). This strategy enables the use of continuous optimization techniques to learn the probabilities, thereby circumventing the challenges associated with direct discrete optimization. Consequently, our loss function is designed to account for the expected loss over the distribution of weights. At each step of gradient descent, we sample from this distribution, using the Gumbel-sigmoid function [10, 25] as a smooth approximation to generate these samples. Let \\( w_{i}^{s} \\) denote the \\( s \\)-th sample from \\( \\mathrm{Ber}(p_i) \\). The Gumbel-sigmoid reparameterization trick approximates the binary sample \\( w_{i}^{s} \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.371, + 0.786, + 0.422 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {s} = \\frac {1}{1 + \\exp \\left[ - \\frac {\\log \\left(\\frac {1 - p _ {i}}{p _ {i}}\\right) + g _ {1} ^ {s} - g _ {0} ^ {s}}{\\tau} \\right]} = \\sigma \\left(\\frac {l + t ^ {s}}{\\tau}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.431, + 0.788, + 0.585 + ], + "angle": 0, + "content": "\\(g_{1}^{s}\\) and \\(g_{0}^{s}\\) are independent samples from standard Gumbel distribution, i.e., Gumbel(0,1), and \\(\\tau\\) is the temperature controlling how close \\(w_{i}^{s}\\) is to either 0 or 1. As shown by [25], \\(\\tau\\) is often chosen from 0.01 to 0.2 to achieve good binary approximation and optimization stability. We define the logit \\(l = \\log \\left(\\frac{1 - p_i}{p_i}\\right)\\), \\(t^s = g_1^s - g_0^s\\), and \\(\\sigma(\\cdot)\\) as the sigmoid function. It can be shown that \\(t^s\\) is equally sampled from the standard logistic distribution (Logistic(0,1)) [25]. With a small temperature, \\(w_{i}^{s}\\) is close to binary. The Gumbel-sigmoid reparameterization trick facilitates the use of gradient-based optimization while preserving the binary nature of the weights. Moreover, it enhances the expressiveness of the mask by utilizing the probabilities to indicate the relative importance of each region." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.393, + 0.62 + ], + "angle": 0, + "content": "3.6 Overall Process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.784, + 0.662 + ], + "angle": 0, + "content": "In this paper, we incorporate a SAM-guided mask, learnable perturbation, and Gumbel-sigmoid reparameterization into Eq. (3) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.675, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {p} ^ {*}, \\phi^ {*} = \\arg \\min _ {\\mathbf {p}, \\phi} \\mathbb {E} _ {\\mathbf {w} \\sim \\operatorname {B e r} (\\mathbf {p})} \\left[ U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {w} | | \\right] \\\\ \\approx \\arg \\min _ {\\mathbf {p}, \\phi} \\mathbb {E} _ {\\mathbf {t} \\sim \\text {L o g i s t i c} (0, 1)} \\left[ U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {w} | | \\right] \\tag {5} \\\\ \\text {w h e r e} \\quad \\mathbf {M} = \\sum_ {i = 1} ^ {N} \\sigma \\left(\\frac {\\log \\frac {1 - p _ {i}}{p _ {i}} + t}{\\tau}\\right) \\mathbf {M} _ {i} \\quad \\text {a n d} \\quad \\hat {\\mathbf {x}} = g _ {\\phi} (\\mathbf {x}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "\\(\\mathbb{E}\\) represents the expectation, and \\(\\mathbf{p} = [p_1,p_2,\\dots ,p_n]^T\\) are the probabilities for the Bernoulli distributions. \\(\\mathbf{M}_i\\) is the binary segment mask pre-generated by SAM, and \\(g_{\\phi}(\\cdot)\\) is a blur function parameterized by \\(\\phi\\). In this paper, we adopt a Gaussian blur function, and \\(\\phi\\) will represent the standard deviation that governs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "the spread of the blur. To ensure \\( p_i \\in [0,1] \\), we apply a sigmoid function to \\( p_i \\) denoted as \\( p_i = \\sigma(v_i) \\), and learn the real number \\( \\{v_i\\}_{i=1}^N \\) instead. To solve the optimization problem, we initialize \\( v_i = 0, \\phi = 0 \\) for all indices \\( i \\) and utilize gradient descent to update them. To estimate the expectation term, we leverage the stochastic optimization method. At each gradient step, we get one sample from Logistic(0,1) to update the parameters until convergence. Compared to directly solving Eq. (3), applying a SAM-guided mask could potentially speed up the optimization process since only a small set of weight parameters need to be learned. Adding the learnable perturbation parameter \\( \\phi \\) will not theoretically increase complexity, as \\( \\phi \\) can be updated in parallel with \\( \\{w_i\\}_{i=1}^N \\). If the SAM model is not available, our method can also be applied at the pixel level, treating \\( \\mathbf{M}_i \\) as a segment solely for the \\( i \\)-th pixel. Although pixel-wise UA is less interpretable by humans, it often contains more detailed information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.466 + ], + "angle": 0, + "content": "As shown in Eq. (5), the uncertainty function directly serves as the optimization objective, and the mask is learned to get the optimal uncertainty reduction, potentially yielding more accurate and faithful explanations. Compared to standard perturbation-based methods, the proposed framework avoids performing an exponential number of perturbations on the input and manually tuning the perturbation parameters. Perturbation-based methods are accurate only if the perturbations are representative. However, it is hard to randomly obtain representative perturbations for a large input space, such as high-resolution images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.377, + 0.508 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Dataset. We evaluate the proposed method on the benchmark image classification datasets including CIFAR-10 (C10) [14], CIFAR-100 (C100) [13], and SVHN [24]. We also evaluate our method on high-quality PartImageNet [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.704 + ], + "angle": 0, + "content": "UQ Model. We adopted the deep ensemble technique [16] for quantifying uncertainty. This method involves the training of several deep neural networks, each beginning from different initialization. Notably, this technique has proven highly effective in UQ, surpassing various approximate BNNs [16]. Specifically, the UQ model for each dataset consists of 5 ensemble components, each trained independently with random initialization. We use ResNet18 and ResNet152 for C10 and C100, a two-layer CNN for SVHN, and DinoV2 with fully connected layers for PartImageNet. The implementation details of model architecture and training hyperparameters for all datasets are provided in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.806, + 0.842 + ], + "angle": 0, + "content": "Baselines. We compare our proposed method with both gradient-based and perturbation-based methods. Following the experiment settings of [39], the gradient-based baselines include Grad [31], SmoothGrad [32], IG [35], and UA-Backprop [39]. The perturbation-based baselines include KernelSHAP [21], LIME [28], and Occlusion [46]. Those methods, except UA-Backprop, are general approaches for analyzing the input-output relationship and can be directly extended to UA. They have been extended to UA in [26, 39], serving as standard baselines. It is worth noting that effectively applying SAM to gradient-based attribution methods is challenging due to the dependency of uncertainty on the entire input. Yet," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.304, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.223 + ], + "angle": 0, + "content": "SAM may fit well with some perturbation-based methods, allowing for the perturbation of segment combinations generated by SAM and avoiding the need to perturb thousands of pixel combinations. The results of the integration of KernelSHAP with SAM are also reported for the detection of problematic regions. Implementation details for all methods are shown in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Evaluation Tasks. In Sec. 4.1, we assess all methods for detecting problematic regions on semi-synthetic data, where the ground-truth problematic regions are known. In Sec. 4.2, we perform faithfulness tests on the methods to observe the reduction in uncertainty achieved by altering only the identified problematic regions. The detailed experiment settings are provided in Appendix A.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.319, + 0.542, + 0.334 + ], + "angle": 0, + "content": "4.1 Detection of Problematic Regions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.706 + ], + "angle": 0, + "content": "In this section, we aim to evaluate methods for detecting image anomalies using semi-synthetic data with known problematic areas. We introduce controlled mismatches that mimic anomalies by randomly replacing a patch in each test image with a corresponding patch from the training dataset at the same location, thereby generating ground-truth problematic regions. Knowing these regions allows us to quantitatively assess our method's precision in identifying artificially induced problematic regions. We conduct experiments on four datasets: C10, C100, SVHN, and PartImageNet. For the C10, C100, and SVHN datasets, we create ground-truth problematic regions by substituting a 10 by 10 patch in test images with a matching sample from the training data at the same location. For the high-resolution PartImageNet, we replace a 100 by 100 patch in each image. We then select 500 images from each dataset that exhibit the most significant increase in uncertainty after modification as our test subjects, indicating the presence of problematic regions. To identify these, we generate uncertainty maps for total uncertainty \\( U_{t} \\) that illustrate each pixel's contribution to it. We use these maps to locate the problematic regions by fitting a bounding box of the same size as the ground truth (10 by 10 for C10, C100, and SVHN; 100 by 100 for PartImageNet) around the area with the highest average attribution score, employing a straightforward brute-force search method to pinpoint the most likely location of the altered patch. Besides qualitative evaluation, our quantitative metrics include Intersection over Union (IoU) and anomaly detection accuracy (ADA). IoU calculates the overlap percentage between detected and actual anomalies divided by the union of these regions, while ADA measures the percentage of detections with an IoU above 0.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Evaluation. The examples of uncertainty maps for all methods are shown in Figure 1. We observe that our method can clearly detect problematic regions that match the ground truth. It is worth noting that gradient-based methods, such as Grad, SmoothGrad, IG, and UA-Backprop, usually produce noisy \"scattered\" visualizations. Although other perturbation-based baselines may yield clear regions, they often lead to misleading results by pointing to inaccurate areas. Appendix B will provide additional examples across datasets, uncertainty maps for real images with high uncertainty, and some failure cases." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.142, + 0.754, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.553, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Fig. 1: Examples of uncertainty maps for identifying problematic regions on PartImageNet. Darker regions indicate a higher contribution to the uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Quantitative Results. The experiment results are shown in Table 1. Our method significantly outperforms all baselines in detecting problematic regions across all datasets. For instance, on the SVHN dataset, our method achieves an IoU of 0.829 and an ADA of 0.830, marking a considerable improvement of approximately \\(50\\%\\) in both IoU and ADA over LIME and Occlusion. It is worth noting that perturbation-based methods generally perform better than gradient-based methods in detecting problematic regions. When KernelSHAP is enhanced with SAM (KernelSHAP + SAM), it shows notable improvements, potentially because SAM provides valuable prior knowledge about potential problematic areas, offering a solid foundation for perturbation-based methods to identify candidates for perturbation. However, our method still surpasses KernelSHAP + SAM. If SAM is not available, our method can be applied at the pixel level, as discussed in Sec. 3.6. To avoid any confusion, results comparing our method with other baselines at the pixel level will be provided in Appendix C.1. Our method still consistently outperforms other baselines in the pixel-level setting." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.304, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 1: IoU \\(\\uparrow\\) and ADA \\(\\uparrow\\) for detection of problematic regions. Bold values indicate the best performance." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.184, + 0.758, + 0.356 + ], + "angle": 0, + "content": "
MethodC10C100SVHNPartImageNet
IoUADAIoUADAIoUADAIoUADA
Grad0.1090.0540.2820.2500.2880.1800.1470.065
SmoothGrad0.2970.2260.3220.1320.3120.1840.1610.075
IG0.2080.1520.2240.2500.3100.2560.4870.505
UA-Backprop0.2710.1780.3600.2500.3360.2580.6190.695
KernelSHAP0.3240.3060.2860.2500.3820.340--
KernelSHAP + SAM0.6760.6700.6670.6620.7980.7940.3220.280
LIME0.2810.2580.5530.5000.3750.3300.2420.190
Occlusion0.2180.1600.2450.2500.3500.3020.3010.240
Ours0.7130.7000.7500.7500.8290.8300.6330.635
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.385, + 0.403, + 0.4 + ], + "angle": 0, + "content": "4.2 Faithfulness Test" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Faithfulness in uncertainty attribution quantifies the accuracy with which a method's explanations reflect the actual influence of input features on the model's uncertainty. A method exhibits high faithfulness if the regions it identifies as significant truly correspond to areas in the input that, upon alteration, substantially affect the model's uncertainty. In this section, we will refine the most problematic pixels of the input, for example, by updating \\(2\\%\\) of the pixels with the highest UA scores, and then observe the reduction in uncertainty after the alteration. To select the pixels with the highest UA scores, we compare all baselines at the pixel level. This is necessary because if one segment shares one score, the relative importance of each pixel remains unknown, preventing us from selecting the most problematic pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The experiments are conducted on images with high uncertainty from the C10, C100, and SVHN testing datasets. Specifically, we select 500 samples with the highest total uncertainty from the testing data for each dataset. We exclude PartImageNet from the comparison because KernelSHAP and LIME are difficult to apply to high-resolution images at the pixel level. The linear regression used in these methods requires that the number of perturbations be at least larger than the number of pixels, which takes a significant amount of time and substantial memory. We consider two methods to refine the original inputs in their problematic regions. The first method utilizes gradient descent with uncertainty as the sole loss function. Specifically, we iteratively update the input \\(\\mathbf{x}\\) through \\(\\mathbf{x} = \\mathbf{x} + \\eta \\cdot \\mathbf{M}(\\mathbf{x}) \\cdot \\frac{\\partial U(\\mathbf{x})}{\\partial \\mathbf{x}}\\) where \\(\\eta\\) is the learning rate, \\(\\mathbf{M}(\\mathbf{x})\\) is a binary map of the same size as \\(x\\) which highlights the pixels (i.e., \\(1\\%\\) of the total pixels) with largest UA score, and \\(U(\\cdot)\\) represents the function for estimating uncertainty. Through iteratively gradient descent, the uncertainty should be significantly reduced if \\(\\mathbf{M}(\\mathbf{x})\\) shows the true problematic areas. The second method directly blurs the most problematic regions indicated by \\(\\mathbf{M}(\\mathbf{x})\\), (i.e., \\(2\\%\\) of the total pixels) following the experiment settings of [39] for the blurring test. After alter" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: We evaluate attribution performance in terms of URR \\(\\uparrow\\) across four datasets, refining images by targeting the top \\(0.5\\%\\) or \\(1\\%\\) of pixels with high total uncertainty using iterative gradient descent, and the top \\(2\\%\\) or \\(5\\%\\) of pixels using blurring techniques. Bold values denote the best performance." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.212, + 0.744, + 0.552 + ], + "angle": 0, + "content": "
MethodRefining using Iterative Gradient Descent
C10C100SVHNAvg. Performance
0.5%1%0.5%1%0.5%1%0.5% + 1%
Grad0.6620.8850.6150.7780.6810.8080.738
SmoothGrad0.6230.8630.5780.7600.7250.8600.735
IG0.6210.8650.5810.7610.7000.8380.728
UA-Backprop0.7060.9030.2610.4360.7140.8530.646
KernelSHAP0.7310.9040.6350.7790.7330.8380.770
LIME0.7870.9270.4580.5730.2560.4110.569
Occlusion0.8220.9400.2970.4530.4200.5920.587
Ours0.8250.9240.7660.8080.9880.9920.884
MethodRefining using Gaussian Blurring
C10C100SVHNAvg. Performance
2%5%2%5%2%5%2% + 5%
Grad0.3430.4400.0990.1360.0530.0310.184
SmoothGrad0.3310.4180.0790.1260.1090.1050.195
IG0.3920.4810.0960.1410.0320.0660.201
UA-Backprop0.3740.4530.0880.1180.0760.1510.210
KernelSHAP0.8230.9020.3190.4480.1220.0980.452
LIME0.3780.3550.4150.5470.1310.2220.341
Occlusion0.7300.7860.2970.3900.4230.5390.528
Ours0.8600.9180.3860.4870.8690.9390.743
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.689 + ], + "angle": 0, + "content": "ing the original input \\(\\mathbf{x}\\) to get the updated input \\(\\tilde{\\mathbf{x}}\\), we compare the uncertainty of them by calculating the Uncertainty Reduction Rate (URR): \\(1 - \\frac{U(\\tilde{\\mathbf{x}})}{U(\\mathbf{x})}\\). For gradient descent refinement, we aim to update \\(0.5\\%\\) and \\(1\\%\\) of the total pixels, whereas for the blurring test, we refine \\(2\\%\\) and \\(5\\%\\) of the total pixels. More pixels are refined in the blurring test to achieve noticeable uncertainty reduction, as blurring is not as effective as directly optimizing uncertainty to attain the optimal value. The detailed settings will be shown in Appendix A.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The experiment results are shown in Table 2. our method stands out remarkably, achieving the highest average URR in both settings. Specifically, with an impressive average URR of 0.884 using iterative gradient descent, our method outperforms its closest competitor, KernelSHAP, by an impressive \\(11.4\\%\\). Similarly, in the blurring refinement, our method's average URR significantly surpasses that of Occlusion, the next best performer, by \\(21.5\\%\\). It is worth noting that perturbation-based LIME and Occlusion perform a little bit worse than gradient-based methods if the refinement is conducted using gradient descent. This may be because gradient-based methods provide the directions in terms of gradients for refining the images. If the refinement is conducted using blur-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.784, + 0.159 + ], + "angle": 0, + "content": "Table 3: Problematic region detection results for the contribution of each component." + }, + { + "type": "table", + "bbox": [ + 0.301, + 0.17, + 0.703, + 0.266 + ], + "angle": 0, + "content": "
MethodC10SVHN
IoUADAIoUADA
Eq. (3)0.3330.2940.3710.325
Eq. (3) + SAM0.6480.6380.7090.712
Eq. (3) + SAM + LP0.6750.6640.7650.764
Eq. (3) + SAM + LP + Gumbel0.7130.7000.8290.832
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.788, + 0.37 + ], + "angle": 0, + "content": "ring, perturbation-based methods usually works better since they consider the blurring perturbation in the attribution process to estimate the contribution of each pixel. Overall, our method consistently outputs other baselines, leveraging both advantages of perturbation-based and gradient-based methods for enhanced faithfulness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.392, + 0.402, + 0.406 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Contribution of Each Component. In this section, we aim to highlight the contribution of each component to the performance in detecting problematic regions. The results are presented in Table 3. The method, described as \"Eq. (3)\", solves Eq. (3) by relaxing \\(\\mathbf{M}\\) into a probability mask with a sigmoid function, \\(\\mathbf{M} = \\sigma (\\mathbf{t})\\in [0,1]^{\\left|\\mathbf{M}\\right|}\\), where \\(\\mathbf{t}\\in \\mathcal{R}^{\\left|\\mathbf{M}\\right|}\\) are the learnable parameters. The results indicate that all three components positively contribute to the improvement in detecting problematic regions. SAM contributes the most significantly, and Gumbel-sigmoid reparameterization is the second most significant contributor. More analysis will be shown in Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.54, + 0.773 + ], + "angle": 0, + "content": "Hyperparameter Sensitivity. Two hyperparameters, \\(\\lambda\\) and \\(\\tau\\), are essential for tuning our method. \\(\\lambda\\) balances uncertainty loss with sparsity loss in Eq. (3), typically chosen to prioritize uncertainty loss over sparsity regularization. \\(\\tau\\) in Eq. (4) adjusts the approximation level for a binary mask, with a small \\(\\tau\\) (0.05 to 0.2) effectively achieving good binary approximations. Empirical results in Table 4 demonstrate that the performance is relatively stable across a range of these hyperparameters. Results for more datasets will be shown in Appendix C.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.54, + 0.84 + ], + "angle": 0, + "content": "Complexity Analysis. The empirical runtime for attributing one input is shown in Table 5. While our proposed methods can" + }, + { + "type": "table_caption", + "bbox": [ + 0.549, + 0.607, + 0.788, + 0.649 + ], + "angle": 0, + "content": "Table 4: Hyperparameter sensitivity for detecting problematic regions on C10 dataset." + }, + { + "type": "table", + "bbox": [ + 0.574, + 0.661, + 0.766, + 0.786 + ], + "angle": 0, + "content": "
MethodC10\nIoU ADA
λ = 3, σ = 0.10.713 0.700
λ = 3, σ = 0.070.679 0.686
λ = 3, σ = 0.150.688 0.682
λ = 3, σ = 0.10.713 0.700
λ = 2.5, σ = 0.10.677 0.680
λ = 2, σ = 0.10.682 0.680
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "achieve better UA performance, their runtime complexity depends on the number of gradient descent iterations needed for convergence when solving (5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.47, + 0.449 + ], + "angle": 0, + "content": "Typically, about 100-300 forward and backward passes are needed to generate one uncertainty map. Further analysis of the convergence of our method will be provided in Appendix C.4. Among the baselines, Grad is the most efficient, requiring only one forward and backward propagation. The time complexity of SmoothGrad and IG depends on the number of samples used for aggregating the gradients, which usually involves hundreds of samples. Although UA-Backprop theoretically only requires one forward and backward pass, in practice, it usually needs one pass per class. Compared" + }, + { + "type": "table_caption", + "bbox": [ + 0.48, + 0.209, + 0.788, + 0.238 + ], + "angle": 0, + "content": "Table 5: Average runtime of attributing one input in seconds (s)." + }, + { + "type": "table", + "bbox": [ + 0.481, + 0.248, + 0.789, + 0.419 + ], + "angle": 0, + "content": "
MethodC10RuntimePartImageNetRuntime
Grad0.02s5.94s
SmoothGrad2.10s33.3s
IG2.14s35.6s
UA-Backprop0.26s9.75s
KernelSHAP1.63s>200s
KernelSHAP + SAM1.08s113s
LIME1.16s>200s
Occlusion0.15s>200s
Ours1.54s103s
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.556 + ], + "angle": 0, + "content": "to perturbation-based methods such as KernelSHAP and LIME, our method generally requires less time when applied to high-resolution images. KernelSHAP and LIME encounter challenges due to their linear regression component, which demands a sufficient number of perturbations and can be slow when dealing with a high variable dimension. It is worth noting that our method scales well with large images and can generate pixel-level UA maps for these images with complexity that increases linearly." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.58, + 0.36, + 0.597 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "This research aims to develop a more faithful and explainable uncertainty attribution method for identifying problematic regions of the input. We formulate UA as an optimization problem to learn a binary mask that indicates problematic regions, thus enhancing our method's faithfulness by directly optimizing uncertainty as the objective. Three key enhancements are introduced. First, we propose SAM-based mask parameterization to generate more human-understandable uncertainty maps, potentially making the optimization process more efficient. Second, we introduce learnable perturbations to adaptively learn the perturbation parameters without the need for manual tuning. Third, we apply a probabilistic approach to learn the binary mask through Gumbel-sigmoid reparameterization under continuous optimization. For evaluation, we test our method's ability to detect synthetic problematic regions and estimate its faithfulness with real, problematic images, demonstrating our method's superior performance. We also discuss the limitations and future work of our method in Appendix D." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.304, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the National Science Foundation award IIS 2236026 and in part by the Rensselaer-IBM AI Research Collaboration (http://airc.rpi.edu), part of the IBM AI Horizons Network." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.218, + 0.323, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.787, + 0.294 + ], + "angle": 0, + "content": "1. Antorán, J., Bhatt, U., Adel, T., Weller, A., Hernández-Lobato, J.M.: Getting a clue: A method for explaining uncertainty estimates. arXiv preprint arXiv:2006.06848 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.787, + 0.336 + ], + "angle": 0, + "content": "2. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.787, + 0.364 + ], + "angle": 0, + "content": "3. Dabkowski, P., Gal, Y.: Real time image saliency for black box classifiers. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.787, + 0.406 + ], + "angle": 0, + "content": "4. Depeweg, S., Hernandez-Lobato, J.M., Doshi-Velez, F., Udluft, S.: Decomposition of uncertainty in bayesian deep learning for efficient and risk-sensitive learning. In: International Conference on Machine Learning. pp. 1184-1193. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.787, + 0.449 + ], + "angle": 0, + "content": "5. Fong, R., Patrick, M., Vedaldi, A.: Understanding deep networks via extremal perturbations and smooth masks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2950-2958 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.787, + 0.49 + ], + "angle": 0, + "content": "6. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.787, + 0.532 + ], + "angle": 0, + "content": "7. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.787, + 0.574 + ], + "angle": 0, + "content": "8. Guo, H., Wang, H., Ji, Q.: Uncertainty-guided probabilistic transformer for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20052-20061 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.787, + 0.616 + ], + "angle": 0, + "content": "9. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.787, + 0.644 + ], + "angle": 0, + "content": "0. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.787, + 0.699 + ], + "angle": 0, + "content": "1. Kapishnikov, A., Venugopalan, S., Avci, B., Wedin, B., Terry, M., Bolukbasi, T.: Guided integrated gradients: An adaptive path method for removing noise. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5050-5058 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.787, + 0.771 + ], + "angle": 0, + "content": "3. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "4. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html 55(5) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "5. Kwon, Y., Won, J.H., Kim, B.J., Paik, M.C.: Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis 142, 106816 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.251, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.188 + ], + "angle": 0, + "content": "16. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.01474" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "17. Ley, D., Bhatt, U., Weller, A.: \\(\\{\\backslash\\) delta\\} -clue: Diverse sets of explanations for uncertainty estimates. arXiv preprint arXiv:2104.06323 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "18. Ley, D., Bhatt, U., Weller, A.: Diverse, global and amortised counterfactual explanations for uncertainty estimates. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7390-7398 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "19. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.325 + ], + "angle": 0, + "content": "20. Lundberg, S.M., Erion, G.G., Lee, S.I.: Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:1802.03888 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.352 + ], + "angle": 0, + "content": "21. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.434 + ], + "angle": 0, + "content": "22. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "23. Montavon, G., Lapuschkin, S., Binder, A., Samek, W., Müller, K.R.: Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition 65, 211-222 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.502 + ], + "angle": 0, + "content": "24. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.542 + ], + "angle": 0, + "content": "25. Ng, I., Zhu, S., Fang, Z., Li, H., Chen, Z., Wang, J.: Masked gradient-based causal structure learning. In: Proceedings of the 2022 SIAM International Conference on Data Mining (SDM). pp. 424-432. SIAM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "26. Perez, I., Skalski, P., Barns-Graham, A., Wong, J., Sutton, D.: Attribution of predictive uncertainties in classification models. In: The 38th Conference on Uncertainty in Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.61 + ], + "angle": 0, + "content": "27. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.651 + ], + "angle": 0, + "content": "28. Ribeiro, M.T., Singh, S., Guestrin, C.: \" why should i trust you?\" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.704 + ], + "angle": 0, + "content": "29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.746 + ], + "angle": 0, + "content": "30. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.787, + 0.786 + ], + "angle": 0, + "content": "31. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.814 + ], + "angle": 0, + "content": "32. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.841 + ], + "angle": 0, + "content": "33. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.304, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Optimization-based UA Via Learning Informative Perturbations" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "34. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "35. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.231 + ], + "angle": 0, + "content": "36. Tierney, L.: Markov chains for exploring posterior distributions. the Annals of Statistics pp. 1701-1728 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.231, + 0.785, + 0.259 + ], + "angle": 0, + "content": "37. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "38. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.286, + 0.785, + 0.327 + ], + "angle": 0, + "content": "39. Wang, H., Joshi, D., Wang, S., Ji, Q.: Gradient-based uncertainty attribution for explainable bayesian deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12044-12053 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "40. Wang, H., Wang, S., Ji, Q.: Semantic attribution for explainable uncertainty quantification. In: International Workshop on Epistemic Uncertainty in Artificial Intelligence. pp. 101-112. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "41. Watson, D., O'Hara, J., Tax, N., Mudd, R., Guy, I.: Explaining predictive uncertainty with information theoretic shapley values. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.411, + 0.785, + 0.453 + ], + "angle": 0, + "content": "42. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics. In: Proceedings of the 28th international conference on machine learning (ICML-11). pp. 681-688 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "43. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.48, + 0.785, + 0.522 + ], + "angle": 0, + "content": "44. Xu, S., Venugopalan, S., Sundararajan, M.: Attribution in scale and space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9680-9689 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.522, + 0.785, + 0.577 + ], + "angle": 0, + "content": "45. Yang, Q., Zhu, X., Fwu, J.K., Ye, Y., You, G., Zhu, Y.: Mfpp: Morphological fragmental perturbation pyramid for black-box model explanations. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 1376-1383. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.577, + 0.785, + 0.605 + ], + "angle": 0, + "content": "46. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European conference on computer vision. pp. 818-833. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.605, + 0.785, + 0.633 + ], + "angle": 0, + "content": "47. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "48. Zhang, Y., Wang, H., Kephart, J.O., Ji, Q.: Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9032 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_origin.pdf b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6152a87e498a52b2e4cfff0ea22d857c1fc0c8ba --- /dev/null +++ b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/fd2d699f-9d6a-496b-a01f-797dcd4bf8a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75fcc79fb2b6148d8bdbc2c68cb64739561e0b14a0d9144250a1a46489299d61 +size 1897788 diff --git a/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/full.md b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c577431d528da0fb8940a39ccfac7e108e9fa7d0 --- /dev/null +++ b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/full.md @@ -0,0 +1,238 @@ +# Optimization-based Uncertainty Attribution Via Learning Informative Perturbations + +Hanjing Wang $^{1}$ , Bashirul Azam Biswas $^{1}$ , and Qiang Ji $^{1}$ + +Rensselaer Polytechnic Institute, Troy NY 12180, USA {wangh36, biswab, jiq}@rpi.edu + +Abstract. Uncertainty attribution (UA) aims to identify key contributors to predictive uncertainty in deep learning models. To improve the faithfulness of existing UA methods, we formulate UA as an optimization problem to learn a binary mask on the input. The learned mask identifies regions that significantly contribute to output uncertainty and allows uncertainty reduction through learning informative perturbations on the masked input. Our method enhances UA interpretability and maintains high efficiency by integrating three key improvements: Segment Anything Model (SAM)-guided mask parameterization for efficient and interpretable mask learning; learnable perturbations that adaptively target and refine problematic regions specific to each input without manually tuning the perturbation parameters; and a novel application of Gumbel-sigmoid reparameterization for efficiently learning Bernoulli-distributed binary masks under continuous optimization. Our experiments on problematic region detection and faithfulness tests demonstrate our method's superiority over state-of-the-art UA methods. + +Keywords: Uncertainty Attribution $\cdot$ Model Explainability + +# 1 Introduction + +While conventional deep learning has made remarkable strides in various domains, they often face challenges with overconfidence in their predictions, which can be risky in safety-critical applications. Hence, uncertainty quantification (UQ) is paramount, as it directly impacts the safety, reliability, and trustworthiness of AI systems, especially for various computer vision applications [8,15,48]. Predictive uncertainties in deep learning are categorized into aleatoric uncertainty, reflecting inherent data randomness, and epistemic uncertainty, resulting from limited model knowledge. Bayesian Neural Networks (BNNs) serve as a foundational approach for quantifying uncertainty, encompassing techniques such as Variational Inference (VI) [7, 19, 22] and Markov Chain Monte Carlo (MCMC) [36, 42, 47] methods to estimate the complex posterior distribution of model parameters. Following this, Ensemble methods [16, 37, 38, 43] offer an alternative strategy, aggregating predictions from multiple models to capture uncertainty. While numerous studies have focused on enhancing the efficiency and accuracy of UQ, comparatively little effort has been devoted to understanding the sources of uncertainty. Identifying the sources of uncertainty is essential, + +as it reveals where and why certain inputs are associated with high uncertainty for enhanced explainability, thereby informing more targeted improvements in model design and training. + +Uncertainty attribution (UA) [1, 26, 39, 40] systematically identifies and analyzes the specific inputs or features that contribute significantly to the uncertainty in the model's predictions. Traditional explainable AI methods highlight areas of inputs leading to specific predictions, which are unreliable when applied directly to UA for identifying problematic regions in input data. The primary reason is that they focus on explaining the model's predictions, assuming confident predictions, while UA aims to pinpoint the sources of input imperfections that contribute to high predictive uncertainties. Additionally, these methods are often class-discriminative when applied to explain the classification score. Consequently, they may struggle to explain inputs with incorrect predictions and large uncertainties [26]. Uncertainty attribution methods are broadly categorized into two types: gradient-based and perturbation-based methods. Gradient-based methods leverage the gradients of the model output with respect to the input to identify the areas contributing to uncertainty. On the other hand, perturbation-based methods involve systematically altering parts of the input and observing the effect on uncertainty, allowing for a more direct assessment of how variations in the input influence the uncertainty. While gradient-based UA methods are often efficient, they only consider the uncertainty sensitivity to very small input changes, leading to noisy and sometimes misleading indications of uncertainty. Perturbation-based methods can be computationally intensive and may be less accurate if an insufficient number of perturbations are performed. + +Inspired by existing literature [3,5,45], we treat UA as an optimization task aimed at learning a binary mask to identify problematic regions of the input. This mask targets specific areas for maximum uncertainty reduction, achieved by applying learnable perturbations to these selected areas. This optimization-based UA framework can potentially improve the faithfulness of UA, which quantifies how accurately a method's explanations reflect the real impact of input features on the uncertainty. A method is considered faithful if the identified areas truly correspond to regions in the input that, when perturbed, have a significant impact on the predictive uncertainty. Our approach ensures faithfulness through a direct linkage between the uncertainty and the identification of uncertain regions in the optimization objective. Moreover, compared to perturbation-based methods, our approach utilizes gradients to iteratively update the mask through continuous optimization, avoiding the need for computationally intensive, pixelwise input perturbations. This strategy may reduce the cost associated with exploring an exponential number of pixel combinations. Unlike gradient-based methods, which directly employ gradients to generate uncertainty maps, our method applies actual perturbations to inputs, not just minor input changes, potentially yielding more accurate explanations. + +To enhance UA performance, we introduce three key adjustments. First, we use a linear combination of segments from the pre-trained Segment Anything Model (SAM) [12] to parameterize the mask. Using SAM segments as a basis + +enables targeted analysis of semantic regional effects and simplifies the learning process to linear combination coefficients instead of pixel-level masks. Second, traditional perturbations used to explain model predictions often fall short in UA. For example, widely used mask-out perturbations invariably result in increased uncertainty in UA, even when applied to problematic regions. This underscores the need for advanced, learnable perturbations in UA, which are essential for accurately reducing uncertainty by modifying problematic regions. Finally, given the challenge of learning a binary mask directly, we adopt a Bernoulli distribution for the masks and employ the Gumbel-sigmoid reparameterization trick [10], enabling continuous optimization with high efficiency. + +# 2 Related Work + +Gradient-based Attribution Methods. Gradient-based methods [11,29,31-35,44,46] leverage gradient information as input attribution. However, raw gradients can be noisy, leading researchers to develop various approaches for refining gradient information. SmoothGrad [32] addresses the issue of noisy gradients by aggregating attributions from multiple noisy images. FullGrad [34] combines the gradients of the output with respect to both the input and the bias variable in each convolutional or fully-connected layer. Integrated Gradients (IG) [35] and its variants [11, 44] create a path integral between a reference image and the input. They then aggregate the gradients of the model output for the artificial images along this path, providing a more comprehensive understanding of the input-output relationship. Grad-CAM [29] focuses on the gradients of the latent feature maps, offering more interpretable insights. Layer-wise relevance propagation (LRP) methods [2, 23, 30] construct attributions by backpropagating the relevance score layer-by-layer from the output to the input. While most gradient-based methods can be applied for both model prediction and uncertainty attribution, LRP methods often need specific neural network architectures [39]. The direct application of these attribution methods might not be sufficient for UA, leading Wang et al. [39] to develop a backpropagation method specifically for UA. In summary, gradient-based methods are often computationally efficient to analyze neural network input-output relationships. Yet, they mainly model output changes from very small input perturbations, which may not yield precise UA results. This limitation is not effectively overcome by refining gradient information through smoothing, aggregation, and path integration techniques. + +Perturbation-based Attribution Methods. Perturbation-based methods [3, 5, 6, 27, 45, 46] estimate attributions by modifying the input and observing the subsequent changes in the model's output. For example, LIME [28] approximates the model's behavior locally by fitting an interpretable model to the perturbed data. RISE [27] generates saliency maps by randomly masking the input and aggregating the model's responses. SHAP and its variants [20, 21] leverage Shapley values from cooperative game theory to interpret complex model predictions, assigning importance to each feature by observing output changes upon feature + +perturbations. The SHAP-based methods are extended to UA with theoretical studies [41]. While perturbation-based methods provide a more intuitive understanding of the model's behavior and allow for better interpretability, they can also be computationally expensive. This is primarily due to the need for an exponentially increasing number of perturbations and subsequent model evaluations to explore the input-output relationship effectively. To improve efficiency, Dabkowski et al. [3], Fong et al. [5], and Yang et al. [45] focus on identifying the smallest regions for perturbation that can significantly change the model's prediction, leveraging continuous optimization techniques to speed up the attribution process. In recent years, several methods have been specifically designed for UA. For instance, CLUE [1] and its variants [17,18] focus on creating an improved image with minimal uncertainty by modifying the uncertain input using a generative model. The attribution map is then generated by measuring the difference between the original and modified input. To further enhance pixel-wise attributions, Perez et al. [26] combined the CLUE method with the path integral technique. However, CLUE-based methods require a pre-trained generative model, which may be challenging to acquire for practical applications. + +# 3 Proposed Method + +# 3.1 Preliminary + +Bayesian Neural Networks. Distinct from point-estimated neural networks, BNNs treat neural network parameters as random variables and calculate the posterior distribution of the parameters $\pmb{\theta}$ , denoted as $p(\pmb{\theta} \mid \mathcal{D})$ where $\mathcal{D}$ represents the training data. Utilizing Bayes' rule, the posterior distribution can be computed using $p(\pmb{\theta} \mid \mathcal{D}) = \frac{p(\mathcal{D}|\pmb{\theta})p(\pmb{\theta})}{p(\mathcal{D})}$ with $p(\mathcal{D}|\pmb{\theta})$ as the likelihood and $p(\pmb{\theta})$ as the prior. Since the exact posterior distribution is often intractable, various techniques have been developed, including MCMC sampling, VI and Ensemble methods. These approaches either directly sample from the posterior or approximate it using a simpler distribution for efficient sampling. + +Uncertainty Quantification. In this study, we concentrate on image classification tasks, where $\mathbf{x}$ denotes the input and $\mathbf{y}$ denotes the output. The neural network predicts $p(\mathbf{y}|\mathbf{x},\boldsymbol{\theta})$ , which is a categorical distribution parameterized by learnable softmax probability. Entropy is a common method for quantifying the uncertainty or randomness in a probability distribution. In entropy-based UQ, the total uncertainty of predictions is decomposed as follows [4]: + +$$ +\underbrace {\mathcal {H} [ p (\mathbf {y} | \mathbf {x} , \mathcal {D}) ]} _ {\text {T o t a l U n c e r t a i n t y} U _ {t} (\mathbf {x})} = \underbrace {\mathcal {I} [ \mathbf {y} , \boldsymbol {\theta} | \mathbf {x} , \mathcal {D} ]} _ {\text {E p i s t e m i c U n c e r t a i n t y}} + \underbrace {\mathbb {E} _ {p (\boldsymbol {\theta} | \mathcal {D})} \left[ \mathcal {H} [ p (\mathbf {y} | \mathbf {x} , \boldsymbol {\theta}) ] \right]} _ {\text {A l e a t o r i c U n c e r t a i n t y} U _ {a} (\mathbf {x})} \tag {1} +$$ + +Here, $\mathcal{H}$ , $\mathcal{I}$ , and $\mathbb{E}$ stand for entropy, mutual information, and expectation, respectively. By employing a Monte Carlo approximation of the posterior, we can + +approximate the uncertainties as: + +$$ +\mathcal {H} [ p (\mathbf {y} | \mathbf {x}, \mathcal {D}) ] = \mathcal {H} \left[ \mathbb {E} _ {p (\boldsymbol {\theta} | \mathcal {D})} [ p (\mathbf {y} | \mathbf {x}, \boldsymbol {\theta}) ] \right] \approx \mathcal {H} \left[ \frac {1}{S} \sum_ {s = 1} ^ {S} p (\mathbf {y} | \mathbf {x}, \boldsymbol {\theta} ^ {s}) \right] \tag {2a} +$$ + +$$ +\mathbb {E} _ {p (\boldsymbol {\theta} | \mathcal {D})} \left[ \mathcal {H} [ p (\mathbf {y} | \mathbf {x}, \boldsymbol {\theta}) ] \right] \approx \frac {1}{S} \sum_ {s = 1} ^ {S} \mathcal {H} [ p (\mathbf {y} | \mathbf {x}, \boldsymbol {\theta} ^ {s}) ] \tag {2b} +$$ + +where $\pmb{\theta}^s\sim p(\pmb {\theta}|\mathcal{D})$ represents the sampling of parameters $\pmb{\theta}$ from their posterior distribution, and $S$ is the sample size. + +# 3.2 Optimization-based Uncertainty Attribution + +Our approach defines UA as an optimization problem, aiming to learn a binary mask, $\mathbf{M}$ , with the same dimensions as $\mathbf{x}$ , that highlights areas in the inputs significantly contributing to uncertainty. For each element $u$ in $\mathbf{x}$ , we assume $\mathbf{M}(u) \in \{0,1\}$ , with $\mathbf{M}(u) = 1$ indicating strong contributions to predictive uncertainty and $\mathbf{M}(u) = 0$ otherwise. To estimate each element's contribution, we perturb the regions indicated by $\mathbf{M}$ . Specifically, with $\hat{\mathbf{x}}$ representing the perturbed input with reduced uncertainty, we replace the regions where $\mathbf{M}(u) = 1$ in $\mathbf{x}$ with those from $\hat{\mathbf{x}}$ , resulting in $\mathbf{x}' = (1 - \mathbf{M}) \odot \mathbf{x} + \mathbf{M} \odot \hat{\mathbf{x}}$ to denote the masked input. The goal is to identify a small subset of $\mathbf{x}$ 's regions, $\mathbf{M}$ , that significantly reduce predictive uncertainty through the masked input. The uncertainty, denoted as $U$ , could be aleatoric $(U_a)$ , epistemic $(U_e)$ , or total $(U_t)$ , as indicated in Eq. (1). The optimal $\mathbf{M}^*$ is determined by: + +$$ +\mathbf {M} ^ {*} = \arg \min _ {\mathbf {M}} U ((1 - \mathbf {M}) \odot \mathbf {x} + \mathbf {M} \odot \hat {\mathbf {x}}) + \lambda | | \mathbf {M} | | \tag {3} +$$ + +Here, $\odot$ denotes element-wise multiplication, $||\mathbf{M}||$ is the norm of $\mathbf{M}$ , and $\lambda$ is a hyperparameter balancing uncertainty reduction and mask sparsity, usually prioritizing uncertainty loss over sparsity regularization. As $U$ utilizes all ensemble models sampled from $p(\boldsymbol{\theta}|\mathcal{D})$ to compute uncertainty, the UA maps are learned by jointly backpropagating through these models. Perturbations in identified problematic regions decrease predictive uncertainty, while those in high-certain areas will increase uncertainty, guiding $\mathbf{M}$ to select critical regions for perturbation to reduce uncertainty. Background or non-informative regions will not impact uncertainty after perturbation, necessitating the inclusion of a sparsity term to ensure these areas are assigned zero values in $\mathbf{M}$ . + +# 3.3 SAM-guided Mask Parameterization + +To enhance the explainability of our proposed method and reduce the optimization complexity, we aim to parameterize the uncertainty mask, $\mathbf{M}$ , through a linear combination of segments derived from the pre-trained SAM model, denoted as $\{\mathbf{M}_i\}_{i=1}^N$ . These $N$ segments are binary masks of the same dimension as $\mathbf{M}$ , each highlighting distinct regions of the input space. Generated by the + +SAM model, these segments are inherently binary and delineate areas corresponding to semantically meaningful and human-understandable concepts. The parameterization of $\mathbf{M}$ is $\mathbf{M} = \sum_{i=1}^{N} w_i \mathbf{M}_i$ where $w_i \in \{0,1\}$ represents the binary weights. By employing these segments as the basis for constructing $\mathbf{M}$ , our method not only identifies regions contributing to uncertainty but also ensures that these regions are aligned with recognizable semantic features. This alignment significantly boosts the explainability of our method, as the identified areas of uncertainty are directly tied to concepts that are easily interpretable by humans. For instance, in an image classification task, rather than indicating abstract or non-intuitive areas of uncertainty, our method could highlight specific objects or parts of an image that contribute to the uncertainty, such as "eyes" in a facial recognition task or "vehicles" in an autonomous driving context. + +While the effectiveness of our methods depends on the accuracy of SAM, we believe our approach can still be effective with imperfect SAM segmentation. For instance, if problematic regions are either included in a large SAM segment or spread across multiple segments, our method can sometimes still identify these areas as potentially problematic. Furthermore, our method is compatible with state-of-the-art segmentation models, ensuring adaptability and improved performance across diverse applications with domain-specific segmentation models. + +# 3.4 Learnable Perturbation + +Traditional perturbation-based methods often use fixed or random perturbations, which may not effectively reduce uncertainty and could unintentionally increase it, even in problematic regions. Therefore, we dynamically adjust learnable perturbations during the optimization process, specifically targeting and modifying areas in the input $\mathbf{x}$ to effectively reduce uncertainty. + +Our approach updates the perturbed input $\hat{\mathbf{x}}$ , as described in Eq. (3), through iterative blurring, represented as $\hat{\mathbf{x}} = g_{\phi}(\mathbf{x})$ , where $g_{\phi}(\cdot)$ is a blurring function parameterized by $\phi$ . This strategy hypothesizes that blurring problematic regions decreases uncertainty, while blurring good areas could increase it by obscuring important information for prediction. Recognizing that a universal blurring strategy does not cater to the varied demands of different inputs, we aim to learn the blurring function's parameters $\phi$ within the optimization process. This approach allows for the precise, dynamic adjustment of perturbations, blurring each region as needed to optimally reduce uncertainty. + +# 3.5 Gumbel-sigmoid Reparameterization For Binary Weights + +To address the challenge of modeling binary weights for our uncertainty mask, $\mathbf{M}$ , we confront the complexities of discrete optimization, which demands extensive search efforts and complicates direct learning of these weights. While continuous optimization through iterative gradient descent offers a potential solution, simply relaxing binary weights to a continuous range between [0, 1] undermines the mask's binary nature, which is crucial for pinpointing problematic areas and facilitating accurate input region replacement through perturbations. + +On one hand, adhering strictly to binary weights ensures clarity in identifying areas contributing to uncertainty but, on the other hand, it restricts our ability to convey the nuanced relative importance of these regions. + +In this paper, we propose maintaining the binary conception of weights while adopting a probabilistic approach for learning them. Specifically, we model each weight as an independent Bernoulli distribution, denoted as $w_{i} \sim \mathrm{Ber}(p_{i})$ with probability parameter $p_i \in [0,1]$ . This strategy enables the use of continuous optimization techniques to learn the probabilities, thereby circumventing the challenges associated with direct discrete optimization. Consequently, our loss function is designed to account for the expected loss over the distribution of weights. At each step of gradient descent, we sample from this distribution, using the Gumbel-sigmoid function [10, 25] as a smooth approximation to generate these samples. Let $w_{i}^{s}$ denote the $s$ -th sample from $\mathrm{Ber}(p_i)$ . The Gumbel-sigmoid reparameterization trick approximates the binary sample $w_{i}^{s}$ by + +$$ +w _ {i} ^ {s} = \frac {1}{1 + \exp \left[ - \frac {\log \left(\frac {1 - p _ {i}}{p _ {i}}\right) + g _ {1} ^ {s} - g _ {0} ^ {s}}{\tau} \right]} = \sigma \left(\frac {l + t ^ {s}}{\tau}\right). \tag {4} +$$ + +$g_{1}^{s}$ and $g_{0}^{s}$ are independent samples from standard Gumbel distribution, i.e., Gumbel(0,1), and $\tau$ is the temperature controlling how close $w_{i}^{s}$ is to either 0 or 1. As shown by [25], $\tau$ is often chosen from 0.01 to 0.2 to achieve good binary approximation and optimization stability. We define the logit $l = \log \left(\frac{1 - p_i}{p_i}\right)$ , $t^s = g_1^s - g_0^s$ , and $\sigma(\cdot)$ as the sigmoid function. It can be shown that $t^s$ is equally sampled from the standard logistic distribution (Logistic(0,1)) [25]. With a small temperature, $w_{i}^{s}$ is close to binary. The Gumbel-sigmoid reparameterization trick facilitates the use of gradient-based optimization while preserving the binary nature of the weights. Moreover, it enhances the expressiveness of the mask by utilizing the probabilities to indicate the relative importance of each region. + +# 3.6 Overall Process + +In this paper, we incorporate a SAM-guided mask, learnable perturbation, and Gumbel-sigmoid reparameterization into Eq. (3) as follows: + +$$ +\begin{array}{l} \mathbf {p} ^ {*}, \phi^ {*} = \arg \min _ {\mathbf {p}, \phi} \mathbb {E} _ {\mathbf {w} \sim \operatorname {B e r} (\mathbf {p})} \left[ U ((1 - \mathbf {M}) \odot \mathbf {x} + \mathbf {M} \odot \hat {\mathbf {x}}) + \lambda | | \mathbf {w} | | \right] \\ \approx \arg \min _ {\mathbf {p}, \phi} \mathbb {E} _ {\mathbf {t} \sim \text {L o g i s t i c} (0, 1)} \left[ U ((1 - \mathbf {M}) \odot \mathbf {x} + \mathbf {M} \odot \hat {\mathbf {x}}) + \lambda | | \mathbf {w} | | \right] \tag {5} \\ \text {w h e r e} \quad \mathbf {M} = \sum_ {i = 1} ^ {N} \sigma \left(\frac {\log \frac {1 - p _ {i}}{p _ {i}} + t}{\tau}\right) \mathbf {M} _ {i} \quad \text {a n d} \quad \hat {\mathbf {x}} = g _ {\phi} (\mathbf {x}) \\ \end{array} +$$ + +$\mathbb{E}$ represents the expectation, and $\mathbf{p} = [p_1,p_2,\dots ,p_n]^T$ are the probabilities for the Bernoulli distributions. $\mathbf{M}_i$ is the binary segment mask pre-generated by SAM, and $g_{\phi}(\cdot)$ is a blur function parameterized by $\phi$ . In this paper, we adopt a Gaussian blur function, and $\phi$ will represent the standard deviation that governs + +the spread of the blur. To ensure $p_i \in [0,1]$ , we apply a sigmoid function to $p_i$ denoted as $p_i = \sigma(v_i)$ , and learn the real number $\{v_i\}_{i=1}^N$ instead. To solve the optimization problem, we initialize $v_i = 0, \phi = 0$ for all indices $i$ and utilize gradient descent to update them. To estimate the expectation term, we leverage the stochastic optimization method. At each gradient step, we get one sample from Logistic(0,1) to update the parameters until convergence. Compared to directly solving Eq. (3), applying a SAM-guided mask could potentially speed up the optimization process since only a small set of weight parameters need to be learned. Adding the learnable perturbation parameter $\phi$ will not theoretically increase complexity, as $\phi$ can be updated in parallel with $\{w_i\}_{i=1}^N$ . If the SAM model is not available, our method can also be applied at the pixel level, treating $\mathbf{M}_i$ as a segment solely for the $i$ -th pixel. Although pixel-wise UA is less interpretable by humans, it often contains more detailed information. + +As shown in Eq. (5), the uncertainty function directly serves as the optimization objective, and the mask is learned to get the optimal uncertainty reduction, potentially yielding more accurate and faithful explanations. Compared to standard perturbation-based methods, the proposed framework avoids performing an exponential number of perturbations on the input and manually tuning the perturbation parameters. Perturbation-based methods are accurate only if the perturbations are representative. However, it is hard to randomly obtain representative perturbations for a large input space, such as high-resolution images. + +# 4 Experiments + +Dataset. We evaluate the proposed method on the benchmark image classification datasets including CIFAR-10 (C10) [14], CIFAR-100 (C100) [13], and SVHN [24]. We also evaluate our method on high-quality PartImageNet [9]. + +UQ Model. We adopted the deep ensemble technique [16] for quantifying uncertainty. This method involves the training of several deep neural networks, each beginning from different initialization. Notably, this technique has proven highly effective in UQ, surpassing various approximate BNNs [16]. Specifically, the UQ model for each dataset consists of 5 ensemble components, each trained independently with random initialization. We use ResNet18 and ResNet152 for C10 and C100, a two-layer CNN for SVHN, and DinoV2 with fully connected layers for PartImageNet. The implementation details of model architecture and training hyperparameters for all datasets are provided in Appendix A.1. + +Baselines. We compare our proposed method with both gradient-based and perturbation-based methods. Following the experiment settings of [39], the gradient-based baselines include Grad [31], SmoothGrad [32], IG [35], and UA-Backprop [39]. The perturbation-based baselines include KernelSHAP [21], LIME [28], and Occlusion [46]. Those methods, except UA-Backprop, are general approaches for analyzing the input-output relationship and can be directly extended to UA. They have been extended to UA in [26, 39], serving as standard baselines. It is worth noting that effectively applying SAM to gradient-based attribution methods is challenging due to the dependency of uncertainty on the entire input. Yet, + +SAM may fit well with some perturbation-based methods, allowing for the perturbation of segment combinations generated by SAM and avoiding the need to perturb thousands of pixel combinations. The results of the integration of KernelSHAP with SAM are also reported for the detection of problematic regions. Implementation details for all methods are shown in Appendix A.2. + +Evaluation Tasks. In Sec. 4.1, we assess all methods for detecting problematic regions on semi-synthetic data, where the ground-truth problematic regions are known. In Sec. 4.2, we perform faithfulness tests on the methods to observe the reduction in uncertainty achieved by altering only the identified problematic regions. The detailed experiment settings are provided in Appendix A.3. + +# 4.1 Detection of Problematic Regions + +In this section, we aim to evaluate methods for detecting image anomalies using semi-synthetic data with known problematic areas. We introduce controlled mismatches that mimic anomalies by randomly replacing a patch in each test image with a corresponding patch from the training dataset at the same location, thereby generating ground-truth problematic regions. Knowing these regions allows us to quantitatively assess our method's precision in identifying artificially induced problematic regions. We conduct experiments on four datasets: C10, C100, SVHN, and PartImageNet. For the C10, C100, and SVHN datasets, we create ground-truth problematic regions by substituting a 10 by 10 patch in test images with a matching sample from the training data at the same location. For the high-resolution PartImageNet, we replace a 100 by 100 patch in each image. We then select 500 images from each dataset that exhibit the most significant increase in uncertainty after modification as our test subjects, indicating the presence of problematic regions. To identify these, we generate uncertainty maps for total uncertainty $U_{t}$ that illustrate each pixel's contribution to it. We use these maps to locate the problematic regions by fitting a bounding box of the same size as the ground truth (10 by 10 for C10, C100, and SVHN; 100 by 100 for PartImageNet) around the area with the highest average attribution score, employing a straightforward brute-force search method to pinpoint the most likely location of the altered patch. Besides qualitative evaluation, our quantitative metrics include Intersection over Union (IoU) and anomaly detection accuracy (ADA). IoU calculates the overlap percentage between detected and actual anomalies divided by the union of these regions, while ADA measures the percentage of detections with an IoU above 0.5. + +Qualitative Evaluation. The examples of uncertainty maps for all methods are shown in Figure 1. We observe that our method can clearly detect problematic regions that match the ground truth. It is worth noting that gradient-based methods, such as Grad, SmoothGrad, IG, and UA-Backprop, usually produce noisy "scattered" visualizations. Although other perturbation-based baselines may yield clear regions, they often lead to misleading results by pointing to inaccurate areas. Appendix B will provide additional examples across datasets, uncertainty maps for real images with high uncertainty, and some failure cases. + +![](images/f94464d8baa5541c8feece0a2f2a807d4511a40a9c9acf32002b1d6b208cefa9.jpg) +Fig. 1: Examples of uncertainty maps for identifying problematic regions on PartImageNet. Darker regions indicate a higher contribution to the uncertainty. + +Quantitative Results. The experiment results are shown in Table 1. Our method significantly outperforms all baselines in detecting problematic regions across all datasets. For instance, on the SVHN dataset, our method achieves an IoU of 0.829 and an ADA of 0.830, marking a considerable improvement of approximately $50\%$ in both IoU and ADA over LIME and Occlusion. It is worth noting that perturbation-based methods generally perform better than gradient-based methods in detecting problematic regions. When KernelSHAP is enhanced with SAM (KernelSHAP + SAM), it shows notable improvements, potentially because SAM provides valuable prior knowledge about potential problematic areas, offering a solid foundation for perturbation-based methods to identify candidates for perturbation. However, our method still surpasses KernelSHAP + SAM. If SAM is not available, our method can be applied at the pixel level, as discussed in Sec. 3.6. To avoid any confusion, results comparing our method with other baselines at the pixel level will be provided in Appendix C.1. Our method still consistently outperforms other baselines in the pixel-level setting. + +Table 1: IoU $\uparrow$ and ADA $\uparrow$ for detection of problematic regions. Bold values indicate the best performance. + +
MethodC10C100SVHNPartImageNet
IoUADAIoUADAIoUADAIoUADA
Grad0.1090.0540.2820.2500.2880.1800.1470.065
SmoothGrad0.2970.2260.3220.1320.3120.1840.1610.075
IG0.2080.1520.2240.2500.3100.2560.4870.505
UA-Backprop0.2710.1780.3600.2500.3360.2580.6190.695
KernelSHAP0.3240.3060.2860.2500.3820.340--
KernelSHAP + SAM0.6760.6700.6670.6620.7980.7940.3220.280
LIME0.2810.2580.5530.5000.3750.3300.2420.190
Occlusion0.2180.1600.2450.2500.3500.3020.3010.240
Ours0.7130.7000.7500.7500.8290.8300.6330.635
+ +# 4.2 Faithfulness Test + +Faithfulness in uncertainty attribution quantifies the accuracy with which a method's explanations reflect the actual influence of input features on the model's uncertainty. A method exhibits high faithfulness if the regions it identifies as significant truly correspond to areas in the input that, upon alteration, substantially affect the model's uncertainty. In this section, we will refine the most problematic pixels of the input, for example, by updating $2\%$ of the pixels with the highest UA scores, and then observe the reduction in uncertainty after the alteration. To select the pixels with the highest UA scores, we compare all baselines at the pixel level. This is necessary because if one segment shares one score, the relative importance of each pixel remains unknown, preventing us from selecting the most problematic pixels. + +The experiments are conducted on images with high uncertainty from the C10, C100, and SVHN testing datasets. Specifically, we select 500 samples with the highest total uncertainty from the testing data for each dataset. We exclude PartImageNet from the comparison because KernelSHAP and LIME are difficult to apply to high-resolution images at the pixel level. The linear regression used in these methods requires that the number of perturbations be at least larger than the number of pixels, which takes a significant amount of time and substantial memory. We consider two methods to refine the original inputs in their problematic regions. The first method utilizes gradient descent with uncertainty as the sole loss function. Specifically, we iteratively update the input $\mathbf{x}$ through $\mathbf{x} = \mathbf{x} + \eta \cdot \mathbf{M}(\mathbf{x}) \cdot \frac{\partial U(\mathbf{x})}{\partial \mathbf{x}}$ where $\eta$ is the learning rate, $\mathbf{M}(\mathbf{x})$ is a binary map of the same size as $x$ which highlights the pixels (i.e., $1\%$ of the total pixels) with largest UA score, and $U(\cdot)$ represents the function for estimating uncertainty. Through iteratively gradient descent, the uncertainty should be significantly reduced if $\mathbf{M}(\mathbf{x})$ shows the true problematic areas. The second method directly blurs the most problematic regions indicated by $\mathbf{M}(\mathbf{x})$ , (i.e., $2\%$ of the total pixels) following the experiment settings of [39] for the blurring test. After alter + +Table 2: We evaluate attribution performance in terms of URR $\uparrow$ across four datasets, refining images by targeting the top $0.5\%$ or $1\%$ of pixels with high total uncertainty using iterative gradient descent, and the top $2\%$ or $5\%$ of pixels using blurring techniques. Bold values denote the best performance. + +
MethodRefining using Iterative Gradient Descent
C10C100SVHNAvg. Performance
0.5%1%0.5%1%0.5%1%0.5% + 1%
Grad0.6620.8850.6150.7780.6810.8080.738
SmoothGrad0.6230.8630.5780.7600.7250.8600.735
IG0.6210.8650.5810.7610.7000.8380.728
UA-Backprop0.7060.9030.2610.4360.7140.8530.646
KernelSHAP0.7310.9040.6350.7790.7330.8380.770
LIME0.7870.9270.4580.5730.2560.4110.569
Occlusion0.8220.9400.2970.4530.4200.5920.587
Ours0.8250.9240.7660.8080.9880.9920.884
MethodRefining using Gaussian Blurring
C10C100SVHNAvg. Performance
2%5%2%5%2%5%2% + 5%
Grad0.3430.4400.0990.1360.0530.0310.184
SmoothGrad0.3310.4180.0790.1260.1090.1050.195
IG0.3920.4810.0960.1410.0320.0660.201
UA-Backprop0.3740.4530.0880.1180.0760.1510.210
KernelSHAP0.8230.9020.3190.4480.1220.0980.452
LIME0.3780.3550.4150.5470.1310.2220.341
Occlusion0.7300.7860.2970.3900.4230.5390.528
Ours0.8600.9180.3860.4870.8690.9390.743
+ +ing the original input $\mathbf{x}$ to get the updated input $\tilde{\mathbf{x}}$ , we compare the uncertainty of them by calculating the Uncertainty Reduction Rate (URR): $1 - \frac{U(\tilde{\mathbf{x}})}{U(\mathbf{x})}$ . For gradient descent refinement, we aim to update $0.5\%$ and $1\%$ of the total pixels, whereas for the blurring test, we refine $2\%$ and $5\%$ of the total pixels. More pixels are refined in the blurring test to achieve noticeable uncertainty reduction, as blurring is not as effective as directly optimizing uncertainty to attain the optimal value. The detailed settings will be shown in Appendix A.3. + +The experiment results are shown in Table 2. our method stands out remarkably, achieving the highest average URR in both settings. Specifically, with an impressive average URR of 0.884 using iterative gradient descent, our method outperforms its closest competitor, KernelSHAP, by an impressive $11.4\%$ . Similarly, in the blurring refinement, our method's average URR significantly surpasses that of Occlusion, the next best performer, by $21.5\%$ . It is worth noting that perturbation-based LIME and Occlusion perform a little bit worse than gradient-based methods if the refinement is conducted using gradient descent. This may be because gradient-based methods provide the directions in terms of gradients for refining the images. If the refinement is conducted using blur- + +Table 3: Problematic region detection results for the contribution of each component. + +
MethodC10SVHN
IoUADAIoUADA
Eq. (3)0.3330.2940.3710.325
Eq. (3) + SAM0.6480.6380.7090.712
Eq. (3) + SAM + LP0.6750.6640.7650.764
Eq. (3) + SAM + LP + Gumbel0.7130.7000.8290.832
+ +ring, perturbation-based methods usually works better since they consider the blurring perturbation in the attribution process to estimate the contribution of each pixel. Overall, our method consistently outputs other baselines, leveraging both advantages of perturbation-based and gradient-based methods for enhanced faithfulness. + +# 4.3 Ablation Studies + +Contribution of Each Component. In this section, we aim to highlight the contribution of each component to the performance in detecting problematic regions. The results are presented in Table 3. The method, described as "Eq. (3)", solves Eq. (3) by relaxing $\mathbf{M}$ into a probability mask with a sigmoid function, $\mathbf{M} = \sigma (\mathbf{t})\in [0,1]^{\left|\mathbf{M}\right|}$ , where $\mathbf{t}\in \mathcal{R}^{\left|\mathbf{M}\right|}$ are the learnable parameters. The results indicate that all three components positively contribute to the improvement in detecting problematic regions. SAM contributes the most significantly, and Gumbel-sigmoid reparameterization is the second most significant contributor. More analysis will be shown in Appendix C.2. + +Hyperparameter Sensitivity. Two hyperparameters, $\lambda$ and $\tau$ , are essential for tuning our method. $\lambda$ balances uncertainty loss with sparsity loss in Eq. (3), typically chosen to prioritize uncertainty loss over sparsity regularization. $\tau$ in Eq. (4) adjusts the approximation level for a binary mask, with a small $\tau$ (0.05 to 0.2) effectively achieving good binary approximations. Empirical results in Table 4 demonstrate that the performance is relatively stable across a range of these hyperparameters. Results for more datasets will be shown in Appendix C.3. + +Complexity Analysis. The empirical runtime for attributing one input is shown in Table 5. While our proposed methods can + +Table 4: Hyperparameter sensitivity for detecting problematic regions on C10 dataset. + +
MethodC10 +IoU ADA
λ = 3, σ = 0.10.713 0.700
λ = 3, σ = 0.070.679 0.686
λ = 3, σ = 0.150.688 0.682
λ = 3, σ = 0.10.713 0.700
λ = 2.5, σ = 0.10.677 0.680
λ = 2, σ = 0.10.682 0.680
+ +achieve better UA performance, their runtime complexity depends on the number of gradient descent iterations needed for convergence when solving (5). + +Typically, about 100-300 forward and backward passes are needed to generate one uncertainty map. Further analysis of the convergence of our method will be provided in Appendix C.4. Among the baselines, Grad is the most efficient, requiring only one forward and backward propagation. The time complexity of SmoothGrad and IG depends on the number of samples used for aggregating the gradients, which usually involves hundreds of samples. Although UA-Backprop theoretically only requires one forward and backward pass, in practice, it usually needs one pass per class. Compared + +Table 5: Average runtime of attributing one input in seconds (s). + +
MethodC10RuntimePartImageNetRuntime
Grad0.02s5.94s
SmoothGrad2.10s33.3s
IG2.14s35.6s
UA-Backprop0.26s9.75s
KernelSHAP1.63s>200s
KernelSHAP + SAM1.08s113s
LIME1.16s>200s
Occlusion0.15s>200s
Ours1.54s103s
+ +to perturbation-based methods such as KernelSHAP and LIME, our method generally requires less time when applied to high-resolution images. KernelSHAP and LIME encounter challenges due to their linear regression component, which demands a sufficient number of perturbations and can be slow when dealing with a high variable dimension. It is worth noting that our method scales well with large images and can generate pixel-level UA maps for these images with complexity that increases linearly. + +# 5 Conclusion + +This research aims to develop a more faithful and explainable uncertainty attribution method for identifying problematic regions of the input. We formulate UA as an optimization problem to learn a binary mask that indicates problematic regions, thus enhancing our method's faithfulness by directly optimizing uncertainty as the objective. Three key enhancements are introduced. First, we propose SAM-based mask parameterization to generate more human-understandable uncertainty maps, potentially making the optimization process more efficient. Second, we introduce learnable perturbations to adaptively learn the perturbation parameters without the need for manual tuning. Third, we apply a probabilistic approach to learn the binary mask through Gumbel-sigmoid reparameterization under continuous optimization. For evaluation, we test our method's ability to detect synthetic problematic regions and estimate its faithfulness with real, problematic images, demonstrating our method's superior performance. We also discuss the limitations and future work of our method in Appendix D. + +Acknowledgement. This work was supported in part by the National Science Foundation award IIS 2236026 and in part by the Rensselaer-IBM AI Research Collaboration (http://airc.rpi.edu), part of the IBM AI Horizons Network. + +# References + +1. Antorán, J., Bhatt, U., Adel, T., Weller, A., Hernández-Lobato, J.M.: Getting a clue: A method for explaining uncertainty estimates. arXiv preprint arXiv:2006.06848 (2020) +2. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015) +3. Dabkowski, P., Gal, Y.: Real time image saliency for black box classifiers. Advances in neural information processing systems 30 (2017) +4. Depeweg, S., Hernandez-Lobato, J.M., Doshi-Velez, F., Udluft, S.: Decomposition of uncertainty in bayesian deep learning for efficient and risk-sensitive learning. In: International Conference on Machine Learning. pp. 1184-1193. PMLR (2018) +5. Fong, R., Patrick, M., Vedaldi, A.: Understanding deep networks via extremal perturbations and smooth masks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2950-2958 (2019) +6. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017) +7. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020) +8. Guo, H., Wang, H., Ji, Q.: Uncertainty-guided probabilistic transformer for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20052-20061 (2022) +9. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933 (2021) +0. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016) +1. Kapishnikov, A., Venugopalan, S., Avci, B., Wedin, B., Terry, M., Bolukbasi, T.: Guided integrated gradients: An adaptive path method for removing noise. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5050-5058 (2021) +2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +3. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +4. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html 55(5) (2014) +5. Kwon, Y., Won, J.H., Kim, B.J., Paik, M.C.: Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis 142, 106816 (2020) + +16. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.01474 +17. Ley, D., Bhatt, U., Weller, A.: $\{\backslash$ delta\} -clue: Diverse sets of explanations for uncertainty estimates. arXiv preprint arXiv:2104.06323 (2021) +18. Ley, D., Bhatt, U., Weller, A.: Diverse, global and amortised counterfactual explanations for uncertainty estimates. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7390-7398 (2022) +19. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017) +20. Lundberg, S.M., Erion, G.G., Lee, S.I.: Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:1802.03888 (2018) +21. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017) +22. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf +23. Montavon, G., Lapuschkin, S., Binder, A., Samek, W., Müller, K.R.: Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition 65, 211-222 (2017) +24. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) +25. Ng, I., Zhu, S., Fang, Z., Li, H., Chen, Z., Wang, J.: Masked gradient-based causal structure learning. In: Proceedings of the 2022 SIAM International Conference on Data Mining (SDM). pp. 424-432. SIAM (2022) +26. Perez, I., Skalski, P., Barns-Graham, A., Wong, J., Sutton, D.: Attribution of predictive uncertainties in classification models. In: The 38th Conference on Uncertainty in Artificial Intelligence (2022) +27. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018) +28. Ribeiro, M.T., Singh, S., Guestrin, C.: " why should i trust you?" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016) +29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +30. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017) +31. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013) +32. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825 (2017) +33. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014) + +34. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019) +35. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017) +36. Tierney, L.: Markov chains for exploring posterior distributions. the Annals of Statistics pp. 1701-1728 (1994) +37. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019) +38. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023) +39. Wang, H., Joshi, D., Wang, S., Ji, Q.: Gradient-based uncertainty attribution for explainable bayesian deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12044-12053 (2023) +40. Wang, H., Wang, S., Ji, Q.: Semantic attribution for explainable uncertainty quantification. In: International Workshop on Epistemic Uncertainty in Artificial Intelligence. pp. 101-112. Springer (2023) +41. Watson, D., O'Hara, J., Tax, N., Mudd, R., Guy, I.: Explaining predictive uncertainty with information theoretic shapley values. Advances in Neural Information Processing Systems 36 (2024) +42. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics. In: Proceedings of the 28th international conference on machine learning (ICML-11). pp. 681-688 (2011) +43. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020) +44. Xu, S., Venugopalan, S., Sundararajan, M.: Attribution in scale and space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9680-9689 (2020) +45. Yang, Q., Zhu, X., Fwu, J.K., Ye, Y., You, G., Zhu, Y.: Mfpp: Morphological fragmental perturbation pyramid for black-box model explanations. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 1376-1383. IEEE (2021) +46. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European conference on computer vision. pp. 818-833. Springer (2014) +47. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019) +48. Zhang, Y., Wang, H., Kephart, J.O., Ji, Q.: Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9032 (2023) \ No newline at end of file diff --git a/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/images.zip b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab1f95f9b0efeb37994a209ce3ea3a6aed7c30af --- /dev/null +++ b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8bca27eb350b018a576354931dc355e088341cabb2977ecd6a55d661251281a +size 467713 diff --git a/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/layout.json b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a571bc31fb0fde0a6e09ee968084a09dab80a9e1 --- /dev/null +++ b/2024/Optimization-based Uncertainty Attribution Via Learning Informative Perturbations/layout.json @@ -0,0 +1,8491 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "Optimization-based Uncertainty Attribution Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "spans": [ + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "type": "text", + "content": "Hanjing Wang" + }, + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "type": "text", + "content": ", Bashirul Azam Biswas" + }, + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "type": "text", + "content": ", and Qiang Ji" + }, + { + "bbox": [ + 176, + 167, + 436, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 190, + 418, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 190, + 418, + 213 + ], + "spans": [ + { + "bbox": [ + 194, + 190, + 418, + 213 + ], + "type": "text", + "content": "Rensselaer Polytechnic Institute, Troy NY 12180, USA {wangh36, biswab, jiq}@rpi.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 236, + 455, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 455, + 413 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 455, + 413 + ], + "type": "text", + "content": "Abstract. Uncertainty attribution (UA) aims to identify key contributors to predictive uncertainty in deep learning models. To improve the faithfulness of existing UA methods, we formulate UA as an optimization problem to learn a binary mask on the input. The learned mask identifies regions that significantly contribute to output uncertainty and allows uncertainty reduction through learning informative perturbations on the masked input. Our method enhances UA interpretability and maintains high efficiency by integrating three key improvements: Segment Anything Model (SAM)-guided mask parameterization for efficient and interpretable mask learning; learnable perturbations that adaptively target and refine problematic regions specific to each input without manually tuning the perturbation parameters; and a novel application of Gumbel-sigmoid reparameterization for efficiently learning Bernoulli-distributed binary masks under continuous optimization. Our experiments on problematic region detection and faithfulness tests demonstrate our method's superiority over state-of-the-art UA methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 422, + 408, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 422, + 408, + 434 + ], + "spans": [ + { + "bbox": [ + 160, + 422, + 408, + 434 + ], + "type": "text", + "content": "Keywords: Uncertainty Attribution " + }, + { + "bbox": [ + 160, + 422, + 408, + 434 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 422, + 408, + 434 + ], + "type": "text", + "content": " Model Explainability" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 451, + 230, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 230, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 230, + 464 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "While conventional deep learning has made remarkable strides in various domains, they often face challenges with overconfidence in their predictions, which can be risky in safety-critical applications. Hence, uncertainty quantification (UQ) is paramount, as it directly impacts the safety, reliability, and trustworthiness of AI systems, especially for various computer vision applications [8,15,48]. Predictive uncertainties in deep learning are categorized into aleatoric uncertainty, reflecting inherent data randomness, and epistemic uncertainty, resulting from limited model knowledge. Bayesian Neural Networks (BNNs) serve as a foundational approach for quantifying uncertainty, encompassing techniques such as Variational Inference (VI) [7, 19, 22] and Markov Chain Monte Carlo (MCMC) [36, 42, 47] methods to estimate the complex posterior distribution of model parameters. Following this, Ensemble methods [16, 37, 38, 43] offer an alternative strategy, aggregating predictions from multiple models to capture uncertainty. While numerous studies have focused on enhancing the efficiency and accuracy of UQ, comparatively little effort has been devoted to understanding the sources of uncertainty. Identifying the sources of uncertainty is essential," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "as it reveals where and why certain inputs are associated with high uncertainty for enhanced explainability, thereby informing more targeted improvements in model design and training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 407 + ], + "type": "text", + "content": "Uncertainty attribution (UA) [1, 26, 39, 40] systematically identifies and analyzes the specific inputs or features that contribute significantly to the uncertainty in the model's predictions. Traditional explainable AI methods highlight areas of inputs leading to specific predictions, which are unreliable when applied directly to UA for identifying problematic regions in input data. The primary reason is that they focus on explaining the model's predictions, assuming confident predictions, while UA aims to pinpoint the sources of input imperfections that contribute to high predictive uncertainties. Additionally, these methods are often class-discriminative when applied to explain the classification score. Consequently, they may struggle to explain inputs with incorrect predictions and large uncertainties [26]. Uncertainty attribution methods are broadly categorized into two types: gradient-based and perturbation-based methods. Gradient-based methods leverage the gradients of the model output with respect to the input to identify the areas contributing to uncertainty. On the other hand, perturbation-based methods involve systematically altering parts of the input and observing the effect on uncertainty, allowing for a more direct assessment of how variations in the input influence the uncertainty. While gradient-based UA methods are often efficient, they only consider the uncertainty sensitivity to very small input changes, leading to noisy and sometimes misleading indications of uncertainty. Perturbation-based methods can be computationally intensive and may be less accurate if an insufficient number of perturbations are performed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 411, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 626 + ], + "type": "text", + "content": "Inspired by existing literature [3,5,45], we treat UA as an optimization task aimed at learning a binary mask to identify problematic regions of the input. This mask targets specific areas for maximum uncertainty reduction, achieved by applying learnable perturbations to these selected areas. This optimization-based UA framework can potentially improve the faithfulness of UA, which quantifies how accurately a method's explanations reflect the real impact of input features on the uncertainty. A method is considered faithful if the identified areas truly correspond to regions in the input that, when perturbed, have a significant impact on the predictive uncertainty. Our approach ensures faithfulness through a direct linkage between the uncertainty and the identification of uncertain regions in the optimization objective. Moreover, compared to perturbation-based methods, our approach utilizes gradients to iteratively update the mask through continuous optimization, avoiding the need for computationally intensive, pixelwise input perturbations. This strategy may reduce the cost associated with exploring an exponential number of pixel combinations. Unlike gradient-based methods, which directly employ gradients to generate uncertainty maps, our method applies actual perturbations to inputs, not just minor input changes, potentially yielding more accurate explanations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "To enhance UA performance, we introduce three key adjustments. First, we use a linear combination of segments from the pre-trained Segment Anything Model (SAM) [12] to parameterize the mask. Using SAM segments as a basis" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "enables targeted analysis of semantic regional effects and simplifies the learning process to linear combination coefficients instead of pixel-level masks. Second, traditional perturbations used to explain model predictions often fall short in UA. For example, widely used mask-out perturbations invariably result in increased uncertainty in UA, even when applied to problematic regions. This underscores the need for advanced, learnable perturbations in UA, which are essential for accurately reducing uncertainty by modifying problematic regions. Finally, given the challenge of learning a binary mask directly, we adopt a Bernoulli distribution for the masks and employ the Gumbel-sigmoid reparameterization trick [10], enabling continuous optimization with high efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 255, + 237, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 237, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 237, + 269 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 282, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 559 + ], + "type": "text", + "content": "Gradient-based Attribution Methods. Gradient-based methods [11,29,31-35,44,46] leverage gradient information as input attribution. However, raw gradients can be noisy, leading researchers to develop various approaches for refining gradient information. SmoothGrad [32] addresses the issue of noisy gradients by aggregating attributions from multiple noisy images. FullGrad [34] combines the gradients of the output with respect to both the input and the bias variable in each convolutional or fully-connected layer. Integrated Gradients (IG) [35] and its variants [11, 44] create a path integral between a reference image and the input. They then aggregate the gradients of the model output for the artificial images along this path, providing a more comprehensive understanding of the input-output relationship. Grad-CAM [29] focuses on the gradients of the latent feature maps, offering more interpretable insights. Layer-wise relevance propagation (LRP) methods [2, 23, 30] construct attributions by backpropagating the relevance score layer-by-layer from the output to the input. While most gradient-based methods can be applied for both model prediction and uncertainty attribution, LRP methods often need specific neural network architectures [39]. The direct application of these attribution methods might not be sufficient for UA, leading Wang et al. [39] to develop a backpropagation method specifically for UA. In summary, gradient-based methods are often computationally efficient to analyze neural network input-output relationships. Yet, they mainly model output changes from very small input perturbations, which may not yield precise UA results. This limitation is not effectively overcome by refining gradient information through smoothing, aggregation, and path integration techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Perturbation-based Attribution Methods. Perturbation-based methods [3, 5, 6, 27, 45, 46] estimate attributions by modifying the input and observing the subsequent changes in the model's output. For example, LIME [28] approximates the model's behavior locally by fitting an interpretable model to the perturbed data. RISE [27] generates saliency maps by randomly masking the input and aggregating the model's responses. SHAP and its variants [20, 21] leverage Shapley values from cooperative game theory to interpret complex model predictions, assigning importance to each feature by observing output changes upon feature" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "perturbations. The SHAP-based methods are extended to UA with theoretical studies [41]. While perturbation-based methods provide a more intuitive understanding of the model's behavior and allow for better interpretability, they can also be computationally expensive. This is primarily due to the need for an exponentially increasing number of perturbations and subsequent model evaluations to explore the input-output relationship effectively. To improve efficiency, Dabkowski et al. [3], Fong et al. [5], and Yang et al. [45] focus on identifying the smallest regions for perturbation that can significantly change the model's prediction, leveraging continuous optimization techniques to speed up the attribution process. In recent years, several methods have been specifically designed for UA. For instance, CLUE [1] and its variants [17,18] focus on creating an improved image with minimal uncertainty by modifying the uncertain input using a generative model. The attribution map is then generated by measuring the difference between the original and modified input. To further enhance pixel-wise attributions, Perez et al. [26] combined the CLUE method with the path integral technique. However, CLUE-based methods require a pre-trained generative model, which may be challenging to acquire for practical applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 345, + 261, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 261, + 359 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 261, + 359 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 378, + 221, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 221, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 221, + 391 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "content": "Bayesian Neural Networks. Distinct from point-estimated neural networks, BNNs treat neural network parameters as random variables and calculate the posterior distribution of the parameters " + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "inline_equation", + "content": "p(\\pmb{\\theta} \\mid \\mathcal{D})" + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "content": " represents the training data. Utilizing Bayes' rule, the posterior distribution can be computed using " + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "inline_equation", + "content": "p(\\pmb{\\theta} \\mid \\mathcal{D}) = \\frac{p(\\mathcal{D}|\\pmb{\\theta})p(\\pmb{\\theta})}{p(\\mathcal{D})}" + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "inline_equation", + "content": "p(\\mathcal{D}|\\pmb{\\theta})" + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "content": " as the likelihood and " + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "inline_equation", + "content": "p(\\pmb{\\theta})" + }, + { + "bbox": [ + 130, + 405, + 482, + 515 + ], + "type": "text", + "content": " as the prior. Since the exact posterior distribution is often intractable, various techniques have been developed, including MCMC sampling, VI and Ensemble methods. These approaches either directly sample from the posterior or approximate it using a simpler distribution for efficient sampling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": "Uncertainty Quantification. In this study, we concentrate on image classification tasks, where " + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": " denotes the input and " + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": " denotes the output. The neural network predicts " + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\mathbf{x},\\boldsymbol{\\theta})" + }, + { + "bbox": [ + 130, + 516, + 482, + 589 + ], + "type": "text", + "content": ", which is a categorical distribution parameterized by learnable softmax probability. Entropy is a common method for quantifying the uncertainty or randomness in a probability distribution. In entropy-based UQ, the total uncertainty of predictions is decomposed as follows [4]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 600, + 481, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 600, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 142, + 600, + 481, + 630 + ], + "type": "interline_equation", + "content": "\\underbrace {\\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x} , \\mathcal {D}) ]} _ {\\text {T o t a l U n c e r t a i n t y} U _ {t} (\\mathbf {x})} = \\underbrace {\\mathcal {I} [ \\mathbf {y} , \\boldsymbol {\\theta} | \\mathbf {x} , \\mathcal {D} ]} _ {\\text {E p i s t e m i c U n c e r t a i n t y}} + \\underbrace {\\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} \\left[ \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x} , \\boldsymbol {\\theta}) ] \\right]} _ {\\text {A l e a t o r i c U n c e r t a i n t y} U _ {a} (\\mathbf {x})} \\tag {1}", + "image_path": "819a8a1f8aa7d23fca541ec6e1715102ea74cc2ae3d92bf026a377785ee457c7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " stand for entropy, mutual information, and expectation, respectively. By employing a Monte Carlo approximation of the posterior, we can" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 282, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 282, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 282, + 128 + ], + "type": "text", + "content": "approximate the uncertainties as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 174, + 137, + 481, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 174, + 137, + 481, + 169 + ], + "type": "interline_equation", + "content": "\\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\mathcal {D}) ] = \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta}) ] \\right] \\approx \\mathcal {H} \\left[ \\frac {1}{S} \\sum_ {s = 1} ^ {S} p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta} ^ {s}) \\right] \\tag {2a}", + "image_path": "125d1d45dd7a6c3fafc2aabf823163d6aaa094b54e17f46210f6dc6ac61741fb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 172, + 481, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 172, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 173, + 172, + 481, + 203 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {p (\\boldsymbol {\\theta} | \\mathcal {D})} \\left[ \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta}) ] \\right] \\approx \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\mathcal {H} [ p (\\mathbf {y} | \\mathbf {x}, \\boldsymbol {\\theta} ^ {s}) ] \\tag {2b}", + "image_path": "d64fd4d17ed5063a8bdefab5fa75aca7b5b69b3ae39311e6aaef3f224ccfd2ed.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^s\\sim p(\\pmb {\\theta}|\\mathcal{D})" + }, + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "text", + "content": " represents the sampling of parameters " + }, + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "text", + "content": " from their posterior distribution, and " + }, + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 212, + 480, + 236 + ], + "type": "text", + "content": " is the sample size." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 253, + 384, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 253, + 384, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 384, + 266 + ], + "type": "text", + "content": "3.2 Optimization-based Uncertainty Attribution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": "Our approach defines UA as an optimization problem, aiming to learn a binary mask, " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", with the same dimensions as " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", that highlights areas in the inputs significantly contributing to uncertainty. For each element " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", we assume " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(u) \\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(u) = 1" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " indicating strong contributions to predictive uncertainty and " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(u) = 0" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " otherwise. To estimate each element's contribution, we perturb the regions indicated by " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ". Specifically, with " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " representing the perturbed input with reduced uncertainty, we replace the regions where " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(u) = 1" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " with those from " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' = (1 - \\mathbf{M}) \\odot \\mathbf{x} + \\mathbf{M} \\odot \\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " to denote the masked input. The goal is to identify a small subset of " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": "'s regions, " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", that significantly reduce predictive uncertainty through the masked input. The uncertainty, denoted as " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", could be aleatoric " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "(U_a)" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", epistemic " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "(U_e)" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", or total " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "(U_t)" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": ", as indicated in Eq. (1). The optimal " + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^*" + }, + { + "bbox": [ + 130, + 272, + 482, + 416 + ], + "type": "text", + "content": " is determined by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 429, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 429, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 200, + 429, + 481, + 444 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {*} = \\arg \\min _ {\\mathbf {M}} U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {M} | | \\tag {3}", + "image_path": "91c15a2b4892e647b1086765225ff2867c3fbf05fe5da426d90e3544ba32cb93.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": " denotes element-wise multiplication, " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "||\\mathbf{M}||" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": " is the norm of " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": " is a hyperparameter balancing uncertainty reduction and mask sparsity, usually prioritizing uncertainty loss over sparsity regularization. As " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": " utilizes all ensemble models sampled from " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{\\theta}|\\mathcal{D})" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": " to compute uncertainty, the UA maps are learned by jointly backpropagating through these models. Perturbations in identified problematic regions decrease predictive uncertainty, while those in high-certain areas will increase uncertainty, guiding " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": " to select critical regions for perturbation to reduce uncertainty. Background or non-informative regions will not impact uncertainty after perturbation, necessitating the inclusion of a sparsity term to ensure these areas are assigned zero values in " + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 586, + 345, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 345, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 345, + 598 + ], + "type": "text", + "content": "3.3 SAM-guided Mask Parameterization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "To enhance the explainability of our proposed method and reduce the optimization complexity, we aim to parameterize the uncertainty mask, " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", through a linear combination of segments derived from the pre-trained SAM model, denoted as " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". These " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " segments are binary masks of the same dimension as " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", each highlighting distinct regions of the input space. Generated by the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 448, + 103 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "SAM model, these segments are inherently binary and delineate areas corresponding to semantically meaningful and human-understandable concepts. The parameterization of " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\sum_{i=1}^{N} w_i \\mathbf{M}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "w_i \\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " represents the binary weights. By employing these segments as the basis for constructing " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ", our method not only identifies regions contributing to uncertainty but also ensures that these regions are aligned with recognizable semantic features. This alignment significantly boosts the explainability of our method, as the identified areas of uncertainty are directly tied to concepts that are easily interpretable by humans. For instance, in an image classification task, rather than indicating abstract or non-intuitive areas of uncertainty, our method could highlight specific objects or parts of an image that contribute to the uncertainty, such as \"eyes\" in a facial recognition task or \"vehicles\" in an autonomous driving context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 344 + ], + "type": "text", + "content": "While the effectiveness of our methods depends on the accuracy of SAM, we believe our approach can still be effective with imperfect SAM segmentation. For instance, if problematic regions are either included in a large SAM segment or spread across multiple segments, our method can sometimes still identify these areas as potentially problematic. Furthermore, our method is compatible with state-of-the-art segmentation models, ensuring adaptability and improved performance across diverse applications with domain-specific segmentation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 360, + 279, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 360, + 279, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 360, + 279, + 372 + ], + "type": "text", + "content": "3.4 Learnable Perturbation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 481, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 438 + ], + "type": "text", + "content": "Traditional perturbation-based methods often use fixed or random perturbations, which may not effectively reduce uncertainty and could unintentionally increase it, even in problematic regions. Therefore, we dynamically adjust learnable perturbations during the optimization process, specifically targeting and modifying areas in the input " + }, + { + "bbox": [ + 130, + 379, + 481, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 379, + 481, + 438 + ], + "type": "text", + "content": " to effectively reduce uncertainty." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "content": "Our approach updates the perturbed input " + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "content": ", as described in Eq. (3), through iterative blurring, represented as " + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = g_{\\phi}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "inline_equation", + "content": "g_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "content": " is a blurring function parameterized by " + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "content": ". This strategy hypothesizes that blurring problematic regions decreases uncertainty, while blurring good areas could increase it by obscuring important information for prediction. Recognizing that a universal blurring strategy does not cater to the varied demands of different inputs, we aim to learn the blurring function's parameters " + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "content": " within the optimization process. This approach allows for the precise, dynamic adjustment of perturbations, blurring each region as needed to optimally reduce uncertainty." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 563, + 448, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 448, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 448, + 575 + ], + "type": "text", + "content": "3.5 Gumbel-sigmoid Reparameterization For Binary Weights" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To address the challenge of modeling binary weights for our uncertainty mask, " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", we confront the complexities of discrete optimization, which demands extensive search efforts and complicates direct learning of these weights. While continuous optimization through iterative gradient descent offers a potential solution, simply relaxing binary weights to a continuous range between [0, 1] undermines the mask's binary nature, which is crucial for pinpointing problematic areas and facilitating accurate input region replacement through perturbations." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "On one hand, adhering strictly to binary weights ensures clarity in identifying areas contributing to uncertainty but, on the other hand, it restricts our ability to convey the nuanced relative importance of these regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "content": "In this paper, we propose maintaining the binary conception of weights while adopting a probabilistic approach for learning them. Specifically, we model each weight as an independent Bernoulli distribution, denoted as " + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "inline_equation", + "content": "w_{i} \\sim \\mathrm{Ber}(p_{i})" + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "content": " with probability parameter " + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "inline_equation", + "content": "p_i \\in [0,1]" + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "content": ". This strategy enables the use of continuous optimization techniques to learn the probabilities, thereby circumventing the challenges associated with direct discrete optimization. Consequently, our loss function is designed to account for the expected loss over the distribution of weights. At each step of gradient descent, we sample from this distribution, using the Gumbel-sigmoid function [10, 25] as a smooth approximation to generate these samples. Let " + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "inline_equation", + "content": "w_{i}^{s}" + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "content": "-th sample from " + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathrm{Ber}(p_i)" + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "content": ". The Gumbel-sigmoid reparameterization trick approximates the binary sample " + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "inline_equation", + "content": "w_{i}^{s}" + }, + { + "bbox": [ + 130, + 152, + 482, + 285 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 293, + 481, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 293, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 200, + 293, + 481, + 334 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {s} = \\frac {1}{1 + \\exp \\left[ - \\frac {\\log \\left(\\frac {1 - p _ {i}}{p _ {i}}\\right) + g _ {1} ^ {s} - g _ {0} ^ {s}}{\\tau} \\right]} = \\sigma \\left(\\frac {l + t ^ {s}}{\\tau}\\right). \\tag {4}", + "image_path": "6728c0d7d9ba358449fe03a36ada13e24f6e53be8385beb69558c718e9c3ac15.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "g_{1}^{s}" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "g_{0}^{s}" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " are independent samples from standard Gumbel distribution, i.e., Gumbel(0,1), and " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " is the temperature controlling how close " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "w_{i}^{s}" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " is to either 0 or 1. As shown by [25], " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " is often chosen from 0.01 to 0.2 to achieve good binary approximation and optimization stability. We define the logit " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "l = \\log \\left(\\frac{1 - p_i}{p_i}\\right)" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "t^s = g_1^s - g_0^s" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " as the sigmoid function. It can be shown that " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "t^s" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " is equally sampled from the standard logistic distribution (Logistic(0,1)) [25]. With a small temperature, " + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "inline_equation", + "content": "w_{i}^{s}" + }, + { + "bbox": [ + 131, + 341, + 482, + 463 + ], + "type": "text", + "content": " is close to binary. The Gumbel-sigmoid reparameterization trick facilitates the use of gradient-based optimization while preserving the binary nature of the weights. Moreover, it enhances the expressiveness of the mask by utilizing the probabilities to indicate the relative importance of each region." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 479, + 240, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 240, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 240, + 491 + ], + "type": "text", + "content": "3.6 Overall Process" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "type": "text", + "content": "In this paper, we incorporate a SAM-guided mask, learnable perturbation, and Gumbel-sigmoid reparameterization into Eq. (3) as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 162, + 534, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 534, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 162, + 534, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {p} ^ {*}, \\phi^ {*} = \\arg \\min _ {\\mathbf {p}, \\phi} \\mathbb {E} _ {\\mathbf {w} \\sim \\operatorname {B e r} (\\mathbf {p})} \\left[ U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {w} | | \\right] \\\\ \\approx \\arg \\min _ {\\mathbf {p}, \\phi} \\mathbb {E} _ {\\mathbf {t} \\sim \\text {L o g i s t i c} (0, 1)} \\left[ U ((1 - \\mathbf {M}) \\odot \\mathbf {x} + \\mathbf {M} \\odot \\hat {\\mathbf {x}}) + \\lambda | | \\mathbf {w} | | \\right] \\tag {5} \\\\ \\text {w h e r e} \\quad \\mathbf {M} = \\sum_ {i = 1} ^ {N} \\sigma \\left(\\frac {\\log \\frac {1 - p _ {i}}{p _ {i}} + t}{\\tau}\\right) \\mathbf {M} _ {i} \\quad \\text {a n d} \\quad \\hat {\\mathbf {x}} = g _ {\\phi} (\\mathbf {x}) \\\\ \\end{array}", + "image_path": "0c0b3ac07d1b004458e1c174f93bfa0939ab80d8b20502c1b2873b96c5730f4e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " represents the expectation, and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [p_1,p_2,\\dots ,p_n]^T" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " are the probabilities for the Bernoulli distributions. " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the binary segment mask pre-generated by SAM, and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "g_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is a blur function parameterized by " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". In this paper, we adopt a Gaussian blur function, and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " will represent the standard deviation that governs" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "the spread of the blur. To ensure " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "p_i \\in [0,1]" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", we apply a sigmoid function to " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "p_i = \\sigma(v_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", and learn the real number " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\{v_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " instead. To solve the optimization problem, we initialize " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "v_i = 0, \\phi = 0" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " for all indices " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " and utilize gradient descent to update them. To estimate the expectation term, we leverage the stochastic optimization method. At each gradient step, we get one sample from Logistic(0,1) to update the parameters until convergence. Compared to directly solving Eq. (3), applying a SAM-guided mask could potentially speed up the optimization process since only a small set of weight parameters need to be learned. Adding the learnable perturbation parameter " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " will not theoretically increase complexity, as " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " can be updated in parallel with " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\{w_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ". If the SAM model is not available, our method can also be applied at the pixel level, treating " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " as a segment solely for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "-th pixel. Although pixel-wise UA is less interpretable by humans, it often contains more detailed information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 369 + ], + "type": "text", + "content": "As shown in Eq. (5), the uncertainty function directly serves as the optimization objective, and the mask is learned to get the optimal uncertainty reduction, potentially yielding more accurate and faithful explanations. Compared to standard perturbation-based methods, the proposed framework avoids performing an exponential number of perturbations on the input and manually tuning the perturbation parameters. Perturbation-based methods are accurate only if the perturbations are representative. However, it is hard to randomly obtain representative perturbations for a large input space, such as high-resolution images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 387, + 230, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 230, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 230, + 402 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 450 + ], + "type": "text", + "content": "Dataset. We evaluate the proposed method on the benchmark image classification datasets including CIFAR-10 (C10) [14], CIFAR-100 (C100) [13], and SVHN [24]. We also evaluate our method on high-quality PartImageNet [9]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 557 + ], + "type": "text", + "content": "UQ Model. We adopted the deep ensemble technique [16] for quantifying uncertainty. This method involves the training of several deep neural networks, each beginning from different initialization. Notably, this technique has proven highly effective in UQ, surpassing various approximate BNNs [16]. Specifically, the UQ model for each dataset consists of 5 ensemble components, each trained independently with random initialization. We use ResNet18 and ResNet152 for C10 and C100, a two-layer CNN for SVHN, and DinoV2 with fully connected layers for PartImageNet. The implementation details of model architecture and training hyperparameters for all datasets are provided in Appendix A.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 493, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 493, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 493, + 666 + ], + "type": "text", + "content": "Baselines. We compare our proposed method with both gradient-based and perturbation-based methods. Following the experiment settings of [39], the gradient-based baselines include Grad [31], SmoothGrad [32], IG [35], and UA-Backprop [39]. The perturbation-based baselines include KernelSHAP [21], LIME [28], and Occlusion [46]. Those methods, except UA-Backprop, are general approaches for analyzing the input-output relationship and can be directly extended to UA. They have been extended to UA in [26, 39], serving as standard baselines. It is worth noting that effectively applying SAM to gradient-based attribution methods is challenging due to the dependency of uncertainty on the entire input. Yet," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "SAM may fit well with some perturbation-based methods, allowing for the perturbation of segment combinations generated by SAM and avoiding the need to perturb thousands of pixel combinations. The results of the integration of KernelSHAP with SAM are also reported for the detection of problematic regions. Implementation details for all methods are shown in Appendix A.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 236 + ], + "type": "text", + "content": "Evaluation Tasks. In Sec. 4.1, we assess all methods for detecting problematic regions on semi-synthetic data, where the ground-truth problematic regions are known. In Sec. 4.2, we perform faithfulness tests on the methods to observe the reduction in uncertainty achieved by altering only the identified problematic regions. The detailed experiment settings are provided in Appendix A.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 252, + 331, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 252, + 331, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 331, + 264 + ], + "type": "text", + "content": "4.1 Detection of Problematic Regions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 559 + ], + "type": "text", + "content": "In this section, we aim to evaluate methods for detecting image anomalies using semi-synthetic data with known problematic areas. We introduce controlled mismatches that mimic anomalies by randomly replacing a patch in each test image with a corresponding patch from the training dataset at the same location, thereby generating ground-truth problematic regions. Knowing these regions allows us to quantitatively assess our method's precision in identifying artificially induced problematic regions. We conduct experiments on four datasets: C10, C100, SVHN, and PartImageNet. For the C10, C100, and SVHN datasets, we create ground-truth problematic regions by substituting a 10 by 10 patch in test images with a matching sample from the training data at the same location. For the high-resolution PartImageNet, we replace a 100 by 100 patch in each image. We then select 500 images from each dataset that exhibit the most significant increase in uncertainty after modification as our test subjects, indicating the presence of problematic regions. To identify these, we generate uncertainty maps for total uncertainty " + }, + { + "bbox": [ + 130, + 271, + 482, + 559 + ], + "type": "inline_equation", + "content": "U_{t}" + }, + { + "bbox": [ + 130, + 271, + 482, + 559 + ], + "type": "text", + "content": " that illustrate each pixel's contribution to it. We use these maps to locate the problematic regions by fitting a bounding box of the same size as the ground truth (10 by 10 for C10, C100, and SVHN; 100 by 100 for PartImageNet) around the area with the highest average attribution score, employing a straightforward brute-force search method to pinpoint the most likely location of the altered patch. Besides qualitative evaluation, our quantitative metrics include Intersection over Union (IoU) and anomaly detection accuracy (ADA). IoU calculates the overlap percentage between detected and actual anomalies divided by the union of these regions, while ADA measures the percentage of detections with an IoU above 0.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Evaluation. The examples of uncertainty maps for all methods are shown in Figure 1. We observe that our method can clearly detect problematic regions that match the ground truth. It is worth noting that gradient-based methods, such as Grad, SmoothGrad, IG, and UA-Backprop, usually produce noisy \"scattered\" visualizations. Although other perturbation-based baselines may yield clear regions, they often lead to misleading results by pointing to inaccurate areas. Appendix B will provide additional examples across datasets, uncertainty maps for real images with high uncertainty, and some failure cases." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 103 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 112, + 461, + 430 + ], + "blocks": [ + { + "bbox": [ + 153, + 112, + 461, + 430 + ], + "lines": [ + { + "bbox": [ + 153, + 112, + 461, + 430 + ], + "spans": [ + { + "bbox": [ + 153, + 112, + 461, + 430 + ], + "type": "image", + "image_path": "f94464d8baa5541c8feece0a2f2a807d4511a40a9c9acf32002b1d6b208cefa9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 437, + 482, + 461 + ], + "lines": [ + { + "bbox": [ + 132, + 437, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 482, + 461 + ], + "type": "text", + "content": "Fig. 1: Examples of uncertainty maps for identifying problematic regions on PartImageNet. Darker regions indicate a higher contribution to the uncertainty." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "Quantitative Results. The experiment results are shown in Table 1. Our method significantly outperforms all baselines in detecting problematic regions across all datasets. For instance, on the SVHN dataset, our method achieves an IoU of 0.829 and an ADA of 0.830, marking a considerable improvement of approximately " + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": " in both IoU and ADA over LIME and Occlusion. It is worth noting that perturbation-based methods generally perform better than gradient-based methods in detecting problematic regions. When KernelSHAP is enhanced with SAM (KernelSHAP + SAM), it shows notable improvements, potentially because SAM provides valuable prior knowledge about potential problematic areas, offering a solid foundation for perturbation-based methods to identify candidates for perturbation. However, our method still surpasses KernelSHAP + SAM. If SAM is not available, our method can be applied at the pixel level, as discussed in Sec. 3.6. To avoid any confusion, results comparing our method with other baselines at the pixel level will be provided in Appendix C.1. Our method still consistently outperforms other baselines in the pixel-level setting." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 145, + 463, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: IoU " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " and ADA " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " for detection of problematic regions. Bold values indicate the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 145, + 463, + 281 + ], + "lines": [ + { + "bbox": [ + 150, + 145, + 463, + 281 + ], + "spans": [ + { + "bbox": [ + 150, + 145, + 463, + 281 + ], + "type": "table", + "html": "
MethodC10C100SVHNPartImageNet
IoUADAIoUADAIoUADAIoUADA
Grad0.1090.0540.2820.2500.2880.1800.1470.065
SmoothGrad0.2970.2260.3220.1320.3120.1840.1610.075
IG0.2080.1520.2240.2500.3100.2560.4870.505
UA-Backprop0.2710.1780.3600.2500.3360.2580.6190.695
KernelSHAP0.3240.3060.2860.2500.3820.340--
KernelSHAP + SAM0.6760.6700.6670.6620.7980.7940.3220.280
LIME0.2810.2580.5530.5000.3750.3300.2420.190
Occlusion0.2180.1600.2450.2500.3500.3020.3010.240
Ours0.7130.7000.7500.7500.8290.8300.6330.635
", + "image_path": "e1882e2cc3500b5cfe0277e77c2f33027f7489be9253bfa03d43f2f7a06f6dd2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 304, + 246, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 246, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 246, + 316 + ], + "type": "text", + "content": "4.2 Faithfulness Test" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 328, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 460 + ], + "type": "text", + "content": "Faithfulness in uncertainty attribution quantifies the accuracy with which a method's explanations reflect the actual influence of input features on the model's uncertainty. A method exhibits high faithfulness if the regions it identifies as significant truly correspond to areas in the input that, upon alteration, substantially affect the model's uncertainty. In this section, we will refine the most problematic pixels of the input, for example, by updating " + }, + { + "bbox": [ + 130, + 328, + 482, + 460 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 460 + ], + "type": "text", + "content": " of the pixels with the highest UA scores, and then observe the reduction in uncertainty after the alteration. To select the pixels with the highest UA scores, we compare all baselines at the pixel level. This is necessary because if one segment shares one score, the relative importance of each pixel remains unknown, preventing us from selecting the most problematic pixels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": "The experiments are conducted on images with high uncertainty from the C10, C100, and SVHN testing datasets. Specifically, we select 500 samples with the highest total uncertainty from the testing data for each dataset. We exclude PartImageNet from the comparison because KernelSHAP and LIME are difficult to apply to high-resolution images at the pixel level. The linear regression used in these methods requires that the number of perturbations be at least larger than the number of pixels, which takes a significant amount of time and substantial memory. We consider two methods to refine the original inputs in their problematic regions. The first method utilizes gradient descent with uncertainty as the sole loss function. Specifically, we iteratively update the input " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\mathbf{x} + \\eta \\cdot \\mathbf{M}(\\mathbf{x}) \\cdot \\frac{\\partial U(\\mathbf{x})}{\\partial \\mathbf{x}}" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " is the learning rate, " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " is a binary map of the same size as " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " which highlights the pixels (i.e., " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " of the total pixels) with largest UA score, and " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "U(\\cdot)" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " represents the function for estimating uncertainty. Through iteratively gradient descent, the uncertainty should be significantly reduced if " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " shows the true problematic areas. The second method directly blurs the most problematic regions indicated by " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": ", (i.e., " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " of the total pixels) following the experiment settings of [39] for the blurring test. After alter" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 448, + 102 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 167, + 455, + 437 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: We evaluate attribution performance in terms of URR " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " across four datasets, refining images by targeting the top " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " of pixels with high total uncertainty using iterative gradient descent, and the top " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " of pixels using blurring techniques. Bold values denote the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 167, + 455, + 437 + ], + "lines": [ + { + "bbox": [ + 159, + 167, + 455, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 167, + 455, + 437 + ], + "type": "table", + "html": "
MethodRefining using Iterative Gradient Descent
C10C100SVHNAvg. Performance
0.5%1%0.5%1%0.5%1%0.5% + 1%
Grad0.6620.8850.6150.7780.6810.8080.738
SmoothGrad0.6230.8630.5780.7600.7250.8600.735
IG0.6210.8650.5810.7610.7000.8380.728
UA-Backprop0.7060.9030.2610.4360.7140.8530.646
KernelSHAP0.7310.9040.6350.7790.7330.8380.770
LIME0.7870.9270.4580.5730.2560.4110.569
Occlusion0.8220.9400.2970.4530.4200.5920.587
Ours0.8250.9240.7660.8080.9880.9920.884
MethodRefining using Gaussian Blurring
C10C100SVHNAvg. Performance
2%5%2%5%2%5%2% + 5%
Grad0.3430.4400.0990.1360.0530.0310.184
SmoothGrad0.3310.4180.0790.1260.1090.1050.195
IG0.3920.4810.0960.1410.0320.0660.201
UA-Backprop0.3740.4530.0880.1180.0760.1510.210
KernelSHAP0.8230.9020.3190.4480.1220.0980.452
LIME0.3780.3550.4150.5470.1310.2220.341
Occlusion0.7300.7860.2970.3900.4230.5390.528
Ours0.8600.9180.3860.4870.8690.9390.743
", + "image_path": "fd37f3dadad16c9aac693bddf7c848014d3bade56d067691fa152d3eb002b14d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": "ing the original input " + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": " to get the updated input " + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": ", we compare the uncertainty of them by calculating the Uncertainty Reduction Rate (URR): " + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "inline_equation", + "content": "1 - \\frac{U(\\tilde{\\mathbf{x}})}{U(\\mathbf{x})}" + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": ". For gradient descent refinement, we aim to update " + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": " of the total pixels, whereas for the blurring test, we refine " + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 459, + 482, + 545 + ], + "type": "text", + "content": " of the total pixels. More pixels are refined in the blurring test to achieve noticeable uncertainty reduction, as blurring is not as effective as directly optimizing uncertainty to attain the optimal value. The detailed settings will be shown in Appendix A.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "The experiment results are shown in Table 2. our method stands out remarkably, achieving the highest average URR in both settings. Specifically, with an impressive average URR of 0.884 using iterative gradient descent, our method outperforms its closest competitor, KernelSHAP, by an impressive " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". Similarly, in the blurring refinement, our method's average URR significantly surpasses that of Occlusion, the next best performer, by " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "21.5\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". It is worth noting that perturbation-based LIME and Occlusion perform a little bit worse than gradient-based methods if the refinement is conducted using gradient descent. This may be because gradient-based methods provide the directions in terms of gradients for refining the images. If the refinement is conducted using blur-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 134, + 430, + 210 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "type": "text", + "content": "Table 3: Problematic region detection results for the contribution of each component." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 134, + 430, + 210 + ], + "lines": [ + { + "bbox": [ + 184, + 134, + 430, + 210 + ], + "spans": [ + { + "bbox": [ + 184, + 134, + 430, + 210 + ], + "type": "table", + "html": "
MethodC10SVHN
IoUADAIoUADA
Eq. (3)0.3330.2940.3710.325
Eq. (3) + SAM0.6480.6380.7090.712
Eq. (3) + SAM + LP0.6750.6640.7650.764
Eq. (3) + SAM + LP + Gumbel0.7130.7000.8290.832
", + "image_path": "5760238ee87a06acdb9b13332c028ab90ec7152978b1e95fe394b79ff1188dfb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 293 + ], + "type": "text", + "content": "ring, perturbation-based methods usually works better since they consider the blurring perturbation in the attribution process to estimate the contribution of each pixel. Overall, our method consistently outputs other baselines, leveraging both advantages of perturbation-based and gradient-based methods for enhanced faithfulness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 310, + 246, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 246, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 246, + 321 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "text", + "content": "Contribution of Each Component. In this section, we aim to highlight the contribution of each component to the performance in detecting problematic regions. The results are presented in Table 3. The method, described as \"Eq. (3)\", solves Eq. (3) by relaxing " + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "text", + "content": " into a probability mask with a sigmoid function, " + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\sigma (\\mathbf{t})\\in [0,1]^{\\left|\\mathbf{M}\\right|}" + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{t}\\in \\mathcal{R}^{\\left|\\mathbf{M}\\right|}" + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "text", + "content": " are the learnable parameters. The results indicate that all three components positively contribute to the improvement in detecting problematic regions. SAM contributes the most significantly, and Gumbel-sigmoid reparameterization is the second most significant contributor. More analysis will be shown in Appendix C.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "text", + "content": "Hyperparameter Sensitivity. Two hyperparameters, " + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "text", + "content": ", are essential for tuning our method. " + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "text", + "content": " balances uncertainty loss with sparsity loss in Eq. (3), typically chosen to prioritize uncertainty loss over sparsity regularization. " + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "text", + "content": " in Eq. (4) adjusts the approximation level for a binary mask, with a small " + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 456, + 330, + 612 + ], + "type": "text", + "content": " (0.05 to 0.2) effectively achieving good binary approximations. Empirical results in Table 4 demonstrate that the performance is relatively stable across a range of these hyperparameters. Results for more datasets will be shown in Appendix C.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 330, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 330, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 330, + 665 + ], + "type": "text", + "content": "Complexity Analysis. The empirical runtime for attributing one input is shown in Table 5. While our proposed methods can" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 351, + 523, + 468, + 622 + ], + "blocks": [ + { + "bbox": [ + 335, + 480, + 482, + 514 + ], + "lines": [ + { + "bbox": [ + 335, + 480, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 335, + 480, + 482, + 514 + ], + "type": "text", + "content": "Table 4: Hyperparameter sensitivity for detecting problematic regions on C10 dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 523, + 468, + 622 + ], + "lines": [ + { + "bbox": [ + 351, + 523, + 468, + 622 + ], + "spans": [ + { + "bbox": [ + 351, + 523, + 468, + 622 + ], + "type": "table", + "html": "
MethodC10\nIoU ADA
λ = 3, σ = 0.10.713 0.700
λ = 3, σ = 0.070.679 0.686
λ = 3, σ = 0.150.688 0.682
λ = 3, σ = 0.10.713 0.700
λ = 2.5, σ = 0.10.677 0.680
λ = 2, σ = 0.10.682 0.680
", + "image_path": "618d39a19a70c0335600494349d6f558a230a95101a6ab575ce76c62a526df8a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 90, + 447, + 102 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "achieve better UA performance, their runtime complexity depends on the number of gradient descent iterations needed for convergence when solving (5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 287, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 287, + 355 + ], + "type": "text", + "content": "Typically, about 100-300 forward and backward passes are needed to generate one uncertainty map. Further analysis of the convergence of our method will be provided in Appendix C.4. Among the baselines, Grad is the most efficient, requiring only one forward and backward propagation. The time complexity of SmoothGrad and IG depends on the number of samples used for aggregating the gradients, which usually involves hundreds of samples. Although UA-Backprop theoretically only requires one forward and backward pass, in practice, it usually needs one pass per class. Compared" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 294, + 196, + 482, + 331 + ], + "blocks": [ + { + "bbox": [ + 293, + 165, + 482, + 188 + ], + "lines": [ + { + "bbox": [ + 293, + 165, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 293, + 165, + 482, + 188 + ], + "type": "text", + "content": "Table 5: Average runtime of attributing one input in seconds (s)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 294, + 196, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 294, + 196, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 294, + 196, + 482, + 331 + ], + "type": "table", + "html": "
MethodC10RuntimePartImageNetRuntime
Grad0.02s5.94s
SmoothGrad2.10s33.3s
IG2.14s35.6s
UA-Backprop0.26s9.75s
KernelSHAP1.63s>200s
KernelSHAP + SAM1.08s113s
LIME1.16s>200s
Occlusion0.15s>200s
Ours1.54s103s
", + "image_path": "cd3731701d6c7a5e24aba3fb1b427b7764cc4217a0f1f120f62d65a06d6b767f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "content": "to perturbation-based methods such as KernelSHAP and LIME, our method generally requires less time when applied to high-resolution images. KernelSHAP and LIME encounter challenges due to their linear regression component, which demands a sufficient number of perturbations and can be slow when dealing with a high variable dimension. It is worth noting that our method scales well with large images and can generate pixel-level UA maps for these images with complexity that increases linearly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 459, + 220, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 220, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 220, + 472 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "This research aims to develop a more faithful and explainable uncertainty attribution method for identifying problematic regions of the input. We formulate UA as an optimization problem to learn a binary mask that indicates problematic regions, thus enhancing our method's faithfulness by directly optimizing uncertainty as the objective. Three key enhancements are introduced. First, we propose SAM-based mask parameterization to generate more human-understandable uncertainty maps, potentially making the optimization process more efficient. Second, we introduce learnable perturbations to adaptively learn the perturbation parameters without the need for manual tuning. Third, we apply a probabilistic approach to learn the binary mask through Gumbel-sigmoid reparameterization under continuous optimization. For evaluation, we test our method's ability to detect synthetic problematic regions and estimate its faithfulness with real, problematic images, demonstrating our method's superior performance. We also discuss the limitations and future work of our method in Appendix D." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the National Science Foundation award IIS 2236026 and in part by the Rensselaer-IBM AI Research Collaboration (http://airc.rpi.edu), part of the IBM AI Horizons Network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 198, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 198, + 481, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 481, + 232 + ], + "type": "text", + "content": "1. Antorán, J., Bhatt, U., Adel, T., Weller, A., Hernández-Lobato, J.M.: Getting a clue: A method for explaining uncertainty estimates. arXiv preprint arXiv:2006.06848 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "type": "text", + "content": "2. Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.R., Samek, W.: On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one 10(7), e0130140 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 266, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 288 + ], + "type": "text", + "content": "3. Dabkowski, P., Gal, Y.: Real time image saliency for black box classifiers. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 321 + ], + "type": "text", + "content": "4. Depeweg, S., Hernandez-Lobato, J.M., Doshi-Velez, F., Udluft, S.: Decomposition of uncertainty in bayesian deep learning for efficient and risk-sensitive learning. In: International Conference on Machine Learning. pp. 1184-1193. PMLR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 481, + 355 + ], + "type": "text", + "content": "5. Fong, R., Patrick, M., Vedaldi, A.: Understanding deep networks via extremal perturbations and smooth masks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2950-2958 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 355, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 481, + 388 + ], + "type": "text", + "content": "6. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: Proceedings of the IEEE international conference on computer vision. pp. 3429-3437 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 481, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 481, + 421 + ], + "type": "text", + "content": "7. Franchi, G., Bursuc, A., Aldea, E., Dubuisson, S., Bloch, I.: Encoding the latent posterior of bayesian neural networks for uncertainty quantification. arXiv preprint arXiv:2012.02818 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 422, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 481, + 454 + ], + "type": "text", + "content": "8. Guo, H., Wang, H., Ji, Q.: Uncertainty-guided probabilistic transformer for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20052-20061 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 455, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 481, + 487 + ], + "type": "text", + "content": "9. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 488, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 510 + ], + "type": "text", + "content": "0. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 510, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 553 + ], + "type": "text", + "content": "1. Kapishnikov, A., Venugopalan, S., Avci, B., Wedin, B., Terry, M., Bolukbasi, T.: Guided integrated gradients: An adaptive path method for removing noise. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5050-5058 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "text", + "content": "2. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 588, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 610 + ], + "type": "text", + "content": "3. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "type": "text", + "content": "4. Krizhevsky, A., Nair, V., Hinton, G.: The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html 55(5) (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "5. Kwon, Y., Won, J.H., Kim, B.J., Paik, M.C.: Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis 142, 106816 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 148 + ], + "type": "text", + "content": "16. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016), http://arxiv.org/abs/1612.01474" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "text", + "content": "17. Ley, D., Bhatt, U., Weller, A.: " + }, + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "inline_equation", + "content": "\\{\\backslash" + }, + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "text", + "content": " delta\\} -clue: Diverse sets of explanations for uncertainty estimates. arXiv preprint arXiv:2104.06323 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "text", + "content": "18. Ley, D., Bhatt, U., Weller, A.: Diverse, global and amortised counterfactual explanations for uncertainty estimates. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7390-7398 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "19. Louizos, C., Welling, M.: Multiplicative normalizing flows for variational bayesian neural networks. In: International Conference on Machine Learning. pp. 2218-2227. PMLR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "text", + "content": "20. Lundberg, S.M., Erion, G.G., Lee, S.I.: Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:1802.03888 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": "21. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 343 + ], + "type": "text", + "content": "22. Maddox, W.J., Izmailov, P., Garipov, T., Vetrov, D.P., Wilson, A.G.: A simple baseline for bayesian uncertainty in deep learning. In: Wallach, H., Larochelle, H., Beygelzimer, A., e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32, pp. 13132-13143. Curran Associates, Inc. (2019), http://papers.nips.cc/paper/9472-a-simple-baseline-for-bayesian-uncertainty-in-deep-learning.pdf" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "text", + "content": "23. Montavon, G., Lapuschkin, S., Binder, A., Samek, W., Müller, K.R.: Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition 65, 211-222 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "text", + "content": "24. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 429 + ], + "type": "text", + "content": "25. Ng, I., Zhu, S., Fang, Z., Li, H., Chen, Z., Wang, J.: Masked gradient-based causal structure learning. In: Proceedings of the 2022 SIAM International Conference on Data Mining (SDM). pp. 424-432. SIAM (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "text", + "content": "26. Perez, I., Skalski, P., Barns-Graham, A., Wong, J., Sutton, D.: Attribution of predictive uncertainties in classification models. In: The 38th Conference on Uncertainty in Artificial Intelligence (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 483 + ], + "type": "text", + "content": "27. Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "text", + "content": "28. Ribeiro, M.T., Singh, S., Guestrin, C.: \" why should i trust you?\" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. pp. 1135-1144 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 557 + ], + "type": "text", + "content": "29. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "text", + "content": "30. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learning. pp. 3145-3153. PMLR (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "31. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 644 + ], + "type": "text", + "content": "32. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "text", + "content": "33. Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 534 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "34. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "text", + "content": "35. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International conference on machine learning. pp. 3319-3328. PMLR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "36. Tierney, L.: Markov chains for exploring posterior distributions. the Annals of Statistics pp. 1701-1728 (1994)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "37. Valdenegro-Toro, M.: Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "38. Wang, H., Ji, Q.: Diversity-enhanced probabilistic ensemble for uncertainty estimation. In: Uncertainty in Artificial Intelligence. pp. 2214-2225. PMLR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 226, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 480, + 258 + ], + "type": "text", + "content": "39. Wang, H., Joshi, D., Wang, S., Ji, Q.: Gradient-based uncertainty attribution for explainable bayesian deep learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12044-12053 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "40. Wang, H., Wang, S., Ji, Q.: Semantic attribution for explainable uncertainty quantification. In: International Workshop on Epistemic Uncertainty in Artificial Intelligence. pp. 101-112. Springer (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "41. Watson, D., O'Hara, J., Tax, N., Mudd, R., Guy, I.: Explaining predictive uncertainty with information theoretic shapley values. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "text", + "content": "42. Welling, M., Teh, Y.W.: Bayesian learning via stochastic gradient Langevin dynamics. In: Proceedings of the 28th international conference on machine learning (ICML-11). pp. 681-688 (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "43. Wen, Y., Tran, D., Ba, J.: Batchsemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "44. Xu, S., Venugopalan, S., Sundararajan, M.: Attribution in scale and space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9680-9689 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "type": "text", + "content": "45. Yang, Q., Zhu, X., Fwu, J.K., Ye, Y., You, G., Zhu, Y.: Mfpp: Morphological fragmental perturbation pyramid for black-box model explanations. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 1376-1383. IEEE (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 479 + ], + "type": "text", + "content": "46. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European conference on computer vision. pp. 818-833. Springer (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "47. Zhang, R., Li, C., Zhang, J., Chen, C., Wilson, A.G.: Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "48. Zhang, Y., Wang, H., Kephart, J.O., Ji, Q.: Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020-9032 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 186, + 91, + 447, + 102 + ], + "type": "text", + "content": "Optimization-based UA Via Learning Informative Perturbations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_content_list.json b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..492304cb6955c8b3bf4b14d94c2585647a0f8fd3 --- /dev/null +++ b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_content_list.json @@ -0,0 +1,1777 @@ +[ + { + "type": "text", + "text": "Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation", + "text_level": 1, + "bbox": [ + 220, + 140, + 781, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixiao Wang $^{1}$ , Chen Tang $^{1,2}$ , Lingfeng Sun $^{1}$ , Simone Rossi $^{3}$ , Yichen Xie $^{1}$ , Chensheng Peng $^{1}$ , Thomas Hannagan $^{3}$ , Stefano Sabatini $^{3}$ , Nicola Poerio $^{4}$ , Masayoshi Tomizuka $^{1}$ , and Wei Zhan $^{1}$", + "bbox": [ + 233, + 210, + 769, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of California, Berkeley, USA", + "bbox": [ + 359, + 268, + 640, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Texas at Austin, USA", + "bbox": [ + 359, + 282, + 640, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Stellantis, France", + "bbox": [ + 436, + 297, + 566, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Stellantis, Italy", + "bbox": [ + 442, + 311, + 560, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yixiao_wang, wzhan}@berkeley.edu", + "bbox": [ + 372, + 327, + 630, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Diffusion models are promising for joint trajectory prediction and controllable generation in autonomous driving, but they face challenges of inefficient inference steps and high computational demands. To tackle these challenges, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance. OGD optimizes the prior distribution for a small diffusion time $T$ and starts the reverse diffusion process from it. ECM directly injects guidance gradients to the estimated clean manifold, eliminating extensive gradient backpropagation throughout the network. Our methodology streamlines the generative process, enabling practical applications with reduced computational overhead. Experimental validation on the large-scale Argoverse 2 dataset demonstrates our approach's superior performance, offering a viable solution for computationally efficient, high-quality joint trajectory prediction and controllable generation for autonomous driving. Our project webpage is at https://yixiaowang7.github.io/OptTrajDiff_Page/", + "bbox": [ + 259, + 371, + 743, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion model $\\cdot$ Autonomous driving $\\cdot$ Trajectory Prediction $\\cdot$ Controllable Trajectory Generation", + "bbox": [ + 259, + 592, + 740, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 643, + 375, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The diffusion model is a class of generative models capable of representing high-dimensional data distributions. In particular, it has demonstrated strong performance in trajectory prediction and generation for autonomous driving [10, 11, 16, 28, 49]. In contrast to traditional trajectory prediction [18, 50] and generative models [41, 45], the unique advantage of diffusion models lies in their ability to deform the generated trajectory distribution to comply with additional requirements at inference stage via gradient-based guided sampling. Notably, it is achieved without extra model training costs. This ability to perform controllable trajectory generation enables various useful applications, such as enforcing additional realism constraints on predicted trajectories, generating directed and user-specified simulation scenarios.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, computational efficiency is a crucial bottleneck hindering the practical application of diffusion models in autonomous driving. Real-time inference is essential for trajectory prediction, as it provides timely forecasts of surrounding agents' behavior, enabling safe and efficient planning in dynamic traffic scenarios. The high demand for inference speed, coupled with limited onboard computational resources, makes it infeasible to deploy diffusion models for onboard trajectory prediction. While simulations do not occur onboard, lightweight models are still preferred to streamline the closed-loop training and evaluation pipelines. The heavy computational cost is mainly attributed to the following two aspects:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computationally Expensive Reverse Diffusion. At the inference stage, the diffusion model samples from standard Gaussian distribution and gradually denoises the sample through dynamics described by a Stochastic Differential Equation (SDE) [39], aiming to eventually obtain a sample as if drawn from a target data distribution. The target data distribution can be significantly different from a standard Gaussian distribution, necessitating a large number of denoising steps to yield good performance. Prior works have attempted to reduce the reverse diffusion steps through adaptive noise schedule [17, 32], fast samplers [21,35,43,46], or distillation [31,37]. However, the fixed standard Gaussian prior poses a challenge in accelerating the reverse diffusion process without violating the SDE, which can inevitably compromise the generation quality.", + "bbox": [ + 212, + 287, + 789, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computationally Expensive Guided Sampling. Controllable generation is typically achieved by guiding the denoising process with the gradient of a guidance cost function. The guidance cost function encodes the desired characteristics of the generated data, which is typically defined on the clean data manifold in trajectory prediction and controllable generation problems. However, guided sampling intends to inject the gradient of the guidance cost function into the series of noisy data at intermediate diffusion steps. It requires a forward pass to estimate the clean data first and then back-propagating throughout the entire network to estimate the gradient with respect to the intermediate noisy data [16, 20, 28], which is very computationally intensive.", + "bbox": [ + 212, + 457, + 789, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Targeted at these challenges, we take a step further to improve the computational efficiency of diffusion models while maintaining their performance for joint trajectory prediction and controllable generation tasks. Specifically, we propose two novel solutions for efficient reverse diffusion and guided sampling respectively. First, we present Optimal Gaussian Diffusion (OGD) to accelerate the reverse diffusion process. At the inference stage, instead of a standard Gaussian distribution far away from the desired data distribution, OGD starts from an optimal Gaussian distribution, which minimizes the distance to intermediate data distribution at a specific noise level, cutting down the diffusion steps before that. We show that we can analytically estimate such an optimal Gaussian distribution and also, an optimal perturbation kernel distribution, at any noise level from the statistics of the data distribution. It allows flexible tuning of the diffusion steps at the inference stage, without the need to train any additional models [48]. We further derive a practical implementation of OGD for joint trajectory prediction and generation, where the optimal Gaussian prior is computed", + "bbox": [ + 212, + 613, + 789, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/084ebe88e07559bafe43edf81a5887443d8208b771aa65c76fb05067f5ddf346.jpg", + "image_caption": [ + "(a) Optimal Gaussian Diffusion (OGD)" + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 501, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/478d8a8314156abe4e50cacc8cd7679f2ec49c114e6b21600fa39eee88cb2aec.jpg", + "image_caption": [ + "Fig. 1: Overview of Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold Guidance (ECM). (a) OGD uses the mean and variance of the data distribution to calculate the optimal prior distribution at a small $T$ . It can largely reduce the diffusion time compared with vanilla diffusion. (b) ECM directly injects the gradient of guidance into the clean data manifold to mitigate computational complexity." + ], + "image_footnote": [], + "bbox": [ + 501, + 148, + 681, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3e5c50c1457d8655f9228e597d7e61812dc7b9f955f88e6ac8ae4d56cd786193.jpg", + "image_caption": [ + "(b) Estimated Clean Manifold Guidance (ECM)" + ], + "image_footnote": [], + "bbox": [ + 681, + 152, + 784, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using the mean and variance of marginal trajectory distributions estimated with a pre-trained marginal trajectory prediction [18, 24, 50] model.", + "bbox": [ + 212, + 398, + 784, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Second, we propose Estimated Clean Manifold Guidance (ECM) to accelerate guided sampling for controllable generation. To save the computational cost due to estimating the guided gradient on the noisy data, we aim to directly inject the gradient into the clean data manifold without lengthy backpropagation. ECM is motivated by the insight that guided sampling can be regarded as a multi-objective optimization problem on the clean data manifold: The first objective is to maximize the likelihood of the samples on the estimated real data distribution; the second objective is to achieve low guidance cost. ECM hierarchically solves this multi-objective problem without backpropagation throughout the entire diffusion model. We show that it leads to faster inference time and much better performance than existing approaches. Also, to tackle the challenges imposed by the multi-modal nature of vehicle interactions, we propose to warm-start the multi-objective optimization problem with reference joint trajectory points estimated using a marginal trajectory predictor. We refer to the complete algorithm as Estimated Clean Manifold with Reference Joint Trajectory (ECMR).", + "bbox": [ + 212, + 430, + 787, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the proposed OGD and ECM methods on real-world tasks, we implement the OGD model leveraging a pre-trained marginal prediction model, QCNet [50], and conduct extensive experiments on the Argoverse 2 dataset. We show that OGD can achieve better joint trajectory prediction performance than vanilla diffusion with significantly fewer diffusion steps—it only takes $1/12$ of the diffusion steps used by the vanilla diffusion model. OGD also achieves outstanding prediction accuracy compared to non-diffusion joint prediction models. In addition, ECMR, coupled with OGD, can generate samples with significantly lower guidance costs with the same level of realism scores compared to conducting controllable generation on the vanilla diffusion model using existing guided sampling approaches used in autonomous driving [16, 28], but using around $1/5$ of their inference step.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 113, + 733, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models have demonstrated their ability to produce high-quality, diverse samples in a variety of applications, such as image, video, and 3D generation [13, 22, 29]. Recently, diffusion models have been applied to trajectory prediction in autonomous driving. It shows great performance on representing future trajectory distribution [10, 16]. However, diffusion models need to run lengthy reverse diffusion processes to generate high-quality samples [5, 10]. This makes it hard to apply diffusion models to trajectory prediction in autonomous driving since it requires in-time prediction for the downstream planning module to promptly respond to dynamically changing traffic scenarios. Previous works [17, 21, 31, 32, 35, 37, 43, 46] mostly focused on mitigating this issue by investigating how to solve reverse SDE in a faster manner. In addition, similar to ours, some works sought alternative initialization of the reverse diffusion process to achieve faster inference. For example, [3] initialized the reverse diffusion process with samples generated by another generative backbone network. However, the backbone network is not deliberately trained for the reverse diffusion process. [48] proposed to learn the initial diffusion with a generative adversarial network (GAN) [8]. However, it requires training a large additional model with a complex training procedure. Also, it requires specifying the reverse diffusion time as a hyperparameter prior to GAN training, which is hard to tune.", + "bbox": [ + 212, + 174, + 787, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Guided sampling. Diffusion models have been successfully used to tackle controllable tasks through guided sampling, such as image inpainting [36] and motion planning [15]. A notable feature of diffusion models with guided sampling is their ability to condition the generation process on the user's preference, which was not available during the training phase. In the driving domain, recent works used guided sampling to generate controllable traffic [16, 49], and controllable pedestrian animation [28]. Our work belongs to the prior category, where user-specified guidance cost is used to guide generation in the trajectory space. In this case, the guidance cost function encodes certain desired properties of the generated trajectories. Thus, it is normally defined on the realistic trajectory samples, which are on the so-called clean manifold, instead of the noisy manifold containing the intermediate noisy data. Some works attempted to learn the guidance cost function on the noisy manifold [15]. Otherwise, it will lead to numerical instability when evaluating guidance cost at intermediate noisy data [28, 49]. To avoid the additional computational costs introduced by a learned guidance cost function, [16, 28] proposed to project the intermediate noisy data into the clean manifold through the diffusion model, and evaluate the guidance cost on the projected point. This approach requires back-propagating throughout the entire diffusion model, which is also computationally intensive.", + "bbox": [ + 212, + 462, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory prediction. In autonomous driving [26, 27], it is vital to precisely forecast how other participants in traffic will move in the future so that ego vehicle can plan a safe and efficient trajectory to execute in the future. Marginal trajectory prediction is used to predict the trajectory distribution for single vehicle and recent works involve kinematic constraints of the vehicles, restrictions of complex topology of roads, and interaction from the surrounding", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vehicles [18,24,50]. Recently, joint trajectory prediction has attracted attention from several researchers. It consists in predicting the joint future trajectories for all agents so that these trajectories are consistent with one another [25], an aspect which marginal trajectory prediction does not consider. Scene-Transformer [25] uses a fixed set of learnable scene embeddings to generate corresponding joint trajectories for all the vehicles in the given scene. Models like M2I [40], and FJMP [30] adopt a conditional approach, predicting the motions of other agents based on the movements of controlled agents. Diffusion model [16] has also been used to predict the joint trajectory. However, joint trajectory prediction is still a challenging problem since the complexity increases exponentially with the number of vehicles in the scene. The efficiency problem becomes more serious when using diffusion model to predict the joint trajectory distribution [16], and largely limits the application of diffusion models in autonomous driving.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 361, + 380, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 388, + 401, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The diffusion process continuously perturbs the unknown data distribution $p_{data}$ with a perturbation kernel and generates intermediate data with a given diffusion time $T$ . Denote the distribution of the time-dependent intermediate nosiy data $\\boldsymbol{x}_t$ as $p_t(\\boldsymbol{x}_t)$ , $t \\in [0, T]$ , where $p_0(\\boldsymbol{x}_0) = p_{data}$ is the clean data distribution and $\\boldsymbol{x}_0$ is the clean data. The series of intermediate data $\\boldsymbol{x}_t$ are generated through the Stochastic Differential Equation (SDE) [39]:", + "bbox": [ + 212, + 407, + 787, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\boldsymbol {x} _ {t} = f (\\boldsymbol {x} _ {t}, t) \\mathrm {d} t + g (t) \\mathrm {d} \\boldsymbol {w}, \\boldsymbol {x} _ {0} \\sim p _ {\\text {d a t a}} = p _ {0} (\\boldsymbol {x} _ {0}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 506, + 784, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f(\\cdot, t)$ is the drift coefficient, $g(t)$ is the diffusion coefficient, and $\\boldsymbol{w}$ is the Wiener process. We can recover $p_{data}$ from reverse-time SDE", + "bbox": [ + 212, + 527, + 785, + 558 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\boldsymbol {x} _ {t} = \\left[ f (\\boldsymbol {x} _ {t}, t) - g (t) ^ {2} \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t}) \\right] \\mathrm {d} t + g (t) \\mathrm {d} \\bar {\\boldsymbol {w}}, \\boldsymbol {x} _ {T} \\sim p _ {T} (\\boldsymbol {x} _ {T}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 564, + 784, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\boldsymbol{w}}$ is another Wiener process, dt is negative timestep.", + "bbox": [ + 214, + 585, + 648, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to solve Eq. (2), we first need to approximate $p_T(\\pmb{x}_T)$ . In previous works, $p_T(\\pmb{x}_T)$ is typically approximated by some prior distributions $p_{prior}$ which contain no information of $p_{data}$ . In this paper, we adopt the setting of Variance Preserving (VP) SDE [12, 39]. In VP-SDE, $p_T(\\pmb{x}_T) \\approx p_{prior}$ when $T \\to \\infty$ . The perturbation kernels are in the form of $p_t(\\pmb{x}_t|\\pmb{x}_0) = \\mathcal{N}(\\pmb{x}_t; \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0, (1 - \\bar{\\alpha}_t)\\pmb{\\Sigma}_p)$ where scalar $\\bar{\\alpha}_t$ is diffusion schedule parameter, $|\\pmb{\\Sigma}_p| = 1$ . A common choice for $\\pmb{\\Sigma}_p$ is the identity matrix $\\mathbf{I}$ [12, 39]. Second, we need to approximate $\\nabla_{\\pmb{x}_t}\\log p_t(\\pmb{x}_t)$ for all $t \\in (0,T]$ . Some works solve score-matching problem [14, 38] and learn a score function $s_\\theta(s_t,t)$ to approximate $\\nabla_{\\pmb{x}_t}\\log p_t(\\pmb{x}_t)$ . In this paper, we follow the practice in DDPM [12] to learn the noise $\\epsilon$ using a network $\\epsilon_\\theta(\\pmb{x}_t,t)$ :", + "bbox": [ + 212, + 601, + 787, + 752 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {\\boldsymbol {x} _ {0} \\sim p _ {d a t a}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {\\Sigma} _ {p})} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 766, + 784, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}_t = \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon},\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t) = -\\sqrt{1 - \\bar{\\alpha}_t}\\pmb{s}_{\\theta}(\\pmb{x}_t,t)$ . With these two approximations, we can learn $q_{\\theta}(\\pmb{x}_0)$ to estimate unknown data distribution $p_0(\\pmb{x}_0)$ solving Eq. (2) from $t = T$ to $t = 0$ .", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Guided Sampling", + "text_level": 1, + "bbox": [ + 215, + 146, + 405, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In prior diffusion-based controllable generation frameworks [15, 16, 20, 28, 49], controllable generation is achieved via biasing of the score function for sampling:", + "bbox": [ + 212, + 174, + 787, + 205 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {x} _ {t}} \\log \\left[ p _ {t} (\\boldsymbol {x} _ {t}) \\exp (- \\mathcal {C} (\\boldsymbol {x} _ {t})) \\right] = \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\mathcal {C} (\\boldsymbol {x} _ {t}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 220, + 785, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{C}(\\cdot)$ is the guidance function. It requires estimating the guidance gradient with respect to the noisy data $\\pmb{x}_t$ . Some approaches introduce an additional neural network to approximate $\\mathcal{C}(\\cdot)$ at different noisy levels [4,15,49]. The additional neural network imposes additional training costs and heavier computational burden at the inference stage. To this end, some works define an analytical guidance function $\\mathcal{J}(\\cdot)$ on the clean data $\\pmb{x}_0$ . They first estimate $\\hat{\\pmb{x}}_0 = f_\\theta(\\pmb{x}_t)$ based on $\\pmb{x}_t$ with the diffusion model, then calculate $\\mathcal{C}(\\pmb{x}_t)$ as $\\mathcal{J}(\\hat{\\pmb{x}}_0)$ . However, when taking the gradient of $\\mathcal{J}(\\hat{\\pmb{x}}_0)$ with respect to $\\pmb{x}_t$ , we get", + "bbox": [ + 212, + 244, + 787, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {x} _ {t}} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0}) = \\nabla_ {\\hat {\\boldsymbol {x}} _ {0}} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0}) \\cdot \\nabla_ {\\boldsymbol {x} _ {t}} f _ {\\theta} (\\boldsymbol {x} _ {t}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 380, + 785, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Estimating $\\nabla_{\\pmb{x}_t}f_\\theta (\\pmb{x}_t)$ requires backpropagating throughout the entire diffusion model, i.e., $f_{\\theta}(\\pmb {x}_t)$ . It requires heavy computing resources and GPU memory.", + "bbox": [ + 214, + 407, + 785, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Trajectory Prediction and Controllable Generation", + "text_level": 1, + "bbox": [ + 214, + 463, + 681, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Joint trajectory prediction aims to predict the future joint trajectories $\\pmb{x}_0$ for all the vehicles in the scene, conditioned on context information $\\pmb{c}$ . It can be regarded as a conditional generation task where the goal is to train a generative model to approximate the distribution $p_0(\\pmb{x}_0|\\pmb{c})$ . For simplicity, we omit $\\pmb{c}$ and represent $p_0(\\pmb{x}_0|\\pmb{c})$ as $p_0(\\pmb{x}_0)$ . We denote $n$ as the number of vehicles in the same scene and $\\pmb{x}_{0,i}$ as the future trajectory for vehicle $i$ , $i \\in \\{1,2,\\dots,n\\}$ , so $\\pmb{x}_0 = [\\pmb{x}_{0,1},\\pmb{x}_{0,2},\\dots,\\pmb{x}_{0,n}]$ . Joint trajectory prediction can be very challenging. The complex interactions among vehicles, especially in highly interactive and dense traffic, result in a complicated high-dimensional $p_0(\\pmb{x}_0)$ , which is difficult to accurately model with lightweight and computationally efficient models. A simplified solution is to approximately decompose the joint trajectory distribution into marginal ones, i.e., $p_0(\\pmb{x}_0) \\approx \\prod_i^n p_0(\\pmb{x}_{0,i})$ . It leads to the marginal trajectory prediction task, which has been extensively studied with mature solutions [18, 24, 50]. One drawback is that it omits the interactions among vehicles in the predicted horizon, which leads to large errors in highly interactive scenes.", + "bbox": [ + 212, + 492, + 787, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Controllable trajectory generation is closely related to trajectory prediction. In addition to generating realistic trajectory samples resembling the ground-truth $p_0(\\pmb{x}_0)$ , controllable generation imposes an additional objective—the generated trajectories should comply with specified guidance cost function $\\mathcal{J}(\\cdot)$ . We term the former objective as realism and the latter as guidance effectiveness. The guidance cost function $\\mathcal{J}(\\cdot)$ can be some goal points of the vehicles, kinematic constraints, etc, which are mostly defined on the clean manifold rather than on the noisy data when it comes to diffusion-based controllable generation.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 143, + 379, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Optimal Gaussian Diffusion", + "text_level": 1, + "bbox": [ + 215, + 175, + 488, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.1, diffusion models typically select a non-informative prior distribution $p_{prior}$ , such as a standard Gaussian, as the initial data distribution for the reverse diffusion process. Such a non-informative $p_{prior}$ is reasonable since $p_T(\\pmb{x}_T)$ converges to it when $T \\to \\infty$ . However, it also means that a large $T$ is required at inference time to yield good performance, which undermines computational efficiency and limits its wide real-time applications in autonomous driving. In this section, we aim to investigate a practical solution to tackle this challenge for joint trajectory prediction. Given the inherent limitation imposed by a non-informative prior, the key question we look into is: can we instead adopt an informative prior so that we can obtain the same level of performance with much smaller reverse diffusion steps?", + "bbox": [ + 212, + 198, + 787, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we still consider a Gaussian prior, but with learnable parameters, i.e., we parameterize $p_{prior}$ as $q_{\\phi}(T) = \\mathcal{N}(\\pmb{\\mu}, \\pmb{\\Sigma})$ , where $\\pmb{\\mu}$ and $\\pmb{\\Sigma}$ are learnable. We aim to optimize $\\pmb{\\mu}$ and $\\pmb{\\Sigma}$ to enhance the generation performance for small $T$ . We got inspiration from [5], where an upper bound of the Kullback-Leibler divergence of the clean data distribution $p_0(\\pmb{x}_0)$ and learned distribution $q_{\\theta}(\\pmb{x}_0)$ was derived as a function of the diffusion time $T$ :", + "bbox": [ + 214, + 364, + 787, + 455 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} \\left[ p _ {0} \\left(\\boldsymbol {x} _ {0}\\right) \\mid \\mid q _ {\\theta} \\left(\\boldsymbol {x} _ {0}\\right) \\right] \\leq \\mathcal {G} \\left(\\boldsymbol {x} _ {\\theta}, T\\right) + \\operatorname {K L} \\left[ p _ {T} \\left(\\boldsymbol {x} _ {T}\\right) \\mid \\mid p _ {\\text {p r i o r}} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 467, + 784, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{G}(\\pmb{x}_{\\theta},T)$ is the positive accumulated error between $\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)$ and $\\pmb{s}_{\\theta}(\\pmb{x}_t,t)$ from 0 to $T$ [5]. Note that $\\mathcal{G}(\\pmb{x}_{\\theta},T)$ is an accumulated error so $\\mathcal{G}(\\pmb{x}_{\\theta},T_1)\\leq \\mathcal{G}(\\pmb{x}_{\\theta},T_2), T_1\\leq T_2$ . If we can achieve lower KL $[p_{T_1}(\\pmb{x}_{T_1})||q_\\phi (T_1)]$ , then a tighter upper bound can be obtained. This opens up the possibility to achieve better performance with less diffusion time. Thus, we propose to optimize the prior distribution by minimizing KL $[p_T(\\pmb{x}_T)||q_\\phi (T)]$ . As shown in Proposition 1 (See proof in Appendix A), it turns out that the optimal $\\mu$ and $\\Sigma$ can be expressed analytically as functions of the ground-truth data statistics. In addition, we find that we can further minimize the target KL divergence if we set a learnable $\\pmb{\\Sigma}_{p}$ in the perturbation kernel $p_t(\\pmb {x}_t|\\pmb {x}_0) = \\mathcal{N}(\\pmb {x}_t;\\sqrt{\\bar{\\alpha}_t}\\pmb {x}_0,(1 - \\bar{\\alpha}_t)\\pmb {\\Sigma}_p)$ , whose optimal value can also be expressed as a function of the data statistics.", + "bbox": [ + 214, + 492, + 789, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proposition 1. Denote $\\pmb{\\mu}_d$ and $\\pmb{\\Sigma}_d$ as the mean and variance of $p_{data}$ . Denote $\\pmb{\\Sigma}^{*}(i,j)$ and $\\pmb{\\Sigma}_{p}^{*}(i,j)$ as the element at ith row and jth column of matrix $\\pmb{\\Sigma}^*$ and $\\pmb{\\Sigma}_{p}^{*}$ . The optimal solution to $\\min KL[p_T(\\pmb{x}_T)||q_\\phi (\\pmb{x}_T,T)]$ is", + "bbox": [ + 215, + 667, + 787, + 715 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\boldsymbol {\\mu} ^ {*} \\approx \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} \\\\ \\boldsymbol {\\Sigma} _ {p} ^ {*} (i, j) \\approx \\frac {1}{| \\boldsymbol {\\Sigma} _ {d} |} \\boldsymbol {\\Sigma} _ {d} (i, j) \\\\ \\boldsymbol {\\Sigma} ^ {*} (i, j) \\approx \\bar {\\alpha} _ {T} \\boldsymbol {\\Sigma} _ {d} + (1 - \\bar {\\alpha} _ {T}) ^ {2} \\boldsymbol {\\Sigma} _ {p} ^ {*} = (\\bar {\\alpha} _ {T} ^ {2} + \\frac {(1 - \\bar {\\alpha} _ {T}) ^ {2}}{| \\boldsymbol {\\Sigma} _ {d} |}) \\boldsymbol {\\Sigma} _ {d} (i, j) \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 729, + 784, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thus, if we are able to estimate the mean and variance of the ground-truth data distribution, we can then analytically determine the optimal prior distribution $\\mathcal{N}(\\pmb{\\mu}^*,\\pmb{\\Sigma}^*)$ and the optimal perturbation kernel variance $\\pmb{\\Sigma}_p^*$ at any noise", + "bbox": [ + 214, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "level $T \\in [0, \\infty)$ . It leads to a crucial advantage against prior efforts on prior learning [48], where a target noise level has to be determined at training time in order to train an additional neural network to represent the prior distribution at the pre-selected noise level. In contrast, our method enables flexible tuning of the number of diffusion steps at inference time, without additional training costs. Since the learnable $q_{\\phi}(T)$ and perturbation kernel are both Gaussian, we refer to our proposed diffusion framework as Optimal Gaussian Diffusion (OGD).", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For joint trajectory prediction and generation, we need to estimate the mean and variance of the joint trajectory distribution $p_0(\\pmb{x}_0)$ . It is not straightforward as only a limited number of trajectory samples exist in the dataset under the same context. Considering there exists mature and accurate marginal trajectory prediction models [18, 24, 50], we can conveniently extract statistics of the marginal trajectory distributions, i.e., $p_0(\\pmb{x}_{0,i}), i \\in \\{1,2,\\dots,n\\}$ from a pretrained marginal trajectory predictor. Since Proposition 1 provides element-wise optimal value, we can easily get:", + "bbox": [ + 212, + 252, + 789, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Corollary 1. Denote $\\pmb{\\mu}_d(\\pmb{x}_{0,i})$ and $\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})$ as the mean and variance of marginal distribution for vehicle $i$ and set both $\\pmb{\\Sigma}$ and $\\pmb{\\Sigma}_p$ are block-diagonal matrices where each block represents the marginal characteristics of one single vehicle. The optimal solution to $\\min KL[p_T(\\pmb{x}_T)||q_\\phi (\\pmb{x}_T,T)]$ is", + "bbox": [ + 215, + 378, + 802, + 441 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\boldsymbol {\\mu} ^ {*} = \\left[ \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} (\\boldsymbol {x} _ {0, n}) \\right] ^ {T} \\\\ \\boldsymbol {\\Sigma} _ {p} ^ {*} = \\frac {1}{\\sum_ {i = 1} ^ {n} | \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0 , i}) |} d i a g [ \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, n}) ] \\\\ \\boldsymbol {\\Sigma} ^ {*} = \\bar {\\alpha} _ {T} d i a g [ \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, n}) ] + (1 - \\bar {\\alpha} _ {T}) ^ {2} \\boldsymbol {\\Sigma} _ {p} ^ {*} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 453, + 785, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Corollary 1 implies that, if we further confine $\\pmb{\\Sigma}$ and $\\pmb{\\Sigma}_p$ to be block-diagonal without covariance between the states of different vehicles, then we can determine their optimal values purely from the estimated marginal statistics $\\pmb{\\mu}_d(\\pmb{x}_{0,i})$ and $\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})$ , $i \\in \\{1,2,\\dots,n\\}$ , which enables a practical implementation of the proposed OGD model for joint trajectory prediction and generation tasks. Specifically, for vehicle $i$ , we leverage a pre-trained marginal trajectory predictor [18,24,50], predict diverse marginal trajectory sample set $\\mathcal{R}_i = \\{\\pmb{r}_i^l\\}_{l=1}^L$ and corresponding likelihood set $\\{p(\\pmb{r}_i^l)\\}_{l=1}^L$ , and estimate $\\pmb{\\mu}_d(\\pmb{x}_{0,i})$ and $\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})$ . For example, $\\pmb{\\mu}_d(\\pmb{x}_{0,i})$ can be estimated as $\\frac{1}{L}\\sum_{l=L}p(\\pmb{r}_i^l)\\pmb{r}_i^l$ .", + "bbox": [ + 212, + 513, + 789, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Estimated Clean Manifold Guidance with Reference", + "text_level": 1, + "bbox": [ + 215, + 667, + 692, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.2, the intensive computation required for guided sampling comes from the calculation of $\\nabla_{\\boldsymbol{x}_t}f_\\theta (\\boldsymbol{x}_t)$ . Previous guided sampling approaches bias the score function defined on the intermediate noisy data $\\boldsymbol{x}_t$ . In this section, we aim to investigate whether we can inject the gradient directly into $\\boldsymbol{x}_0$ to avoid the gradient propagation process. We first reformulate the controllable generation as a multi-objective optimization problem directly over $\\boldsymbol{x}_0$ and propose an iterative algorithm to solve the formulated problem. In addition, we use reference trajectory points to create the region of interest, which helps solve the local optimal problem caused by multi-modal joint trajectory distribution and accelerate the guided sampling process.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Estimated Clean Manifold Guidance. The objective of controllable generation of sample $\\pmb{x}_0$ includes two different objectives. The most important is the negative likelihood, ensuring the sample lies in the clean manifold. The second important is the guidance cost representing the user preference on the generated sample $\\pmb{x}_0$ . This multi-objective optimization problem can be represented as", + "bbox": [ + 215, + 146, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {x} _ {0}} \\left[ - \\log q _ {\\theta} (\\boldsymbol {x} _ {0}), \\mathcal {J} (\\boldsymbol {x} _ {0}) \\right] ^ {T} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 234, + 782, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inspired by lexicographic optimization [33], we solve this multi-objective optimization problem hierarchically. The main idea is to optimize each objective in the order of importance regardless of the degradation of the other less significant objectives. We first optimize the most important objective, $-\\log q_{\\theta}(\\pmb{x}_0)$ , to generate realistic and high-likelihood samples. The diffusion model achieves this goal effectively by reversing the diffusion process from noisy samples at specific noise level. However, exact noise level of current sample $\\pmb{x}_0$ is unknown. To address this, we inject noise at level $t$ into $\\pmb{x}_0$ , and then denoise it from $t$ with learned diffusion model. This approach, similar to noise injection and denoising [1,23,37], improves the desired sample quality. We iteratively repeat this process $K$ times, injecting guidance at each iteration to strengthen user preference.", + "bbox": [ + 212, + 268, + 784, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, denote $\\pmb{x}_0(k)$ as the sample at iteration $k$ . We first regenerate high-likelihood sample $\\hat{\\pmb{x}}_0(k)$ by diffusion model:", + "bbox": [ + 212, + 434, + 782, + 465 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {0} (k) \\leftarrow \\mathbb {E} \\left[ q _ {\\theta} \\left(\\boldsymbol {x} _ {0} (k) \\mid \\boldsymbol {x} _ {t _ {k}} (k)\\right) \\right] = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t _ {k}}}} \\left(\\boldsymbol {x} _ {t _ {k}} - \\sqrt {1 - \\bar {\\alpha} _ {t _ {k}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t _ {k}} (k), t _ {k}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 477, + 782, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\pmb{x}_{t_k}(k) = \\sqrt{\\bar{\\alpha}_{t_k}}\\pmb{x}_0(k) + \\sqrt{1 - \\bar{\\alpha}_{t_k}}\\pmb{\\epsilon}, \\pmb{\\epsilon} \\sim \\mathcal{N}(\\pmb{0},\\pmb{\\Sigma}_p)$ . $t_k \\in [0,T]$ is a tunable parameter. Then we minimize the guidance cost function with a small degradation of the most important objective $-\\log q_{\\theta}(\\pmb{x}_0)$ ,", + "bbox": [ + 214, + 518, + 782, + 565 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {0} (k - 1) \\leftarrow \\hat {\\boldsymbol {x}} _ {0} (k) - \\zeta \\nabla_ {\\hat {\\boldsymbol {x}} _ {0} (k)} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0} (k)) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 579, + 782, + 593 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\zeta$ is the step size. Small degradation of $-\\log q_{\\theta}(\\pmb{x}_0)$ is realized by one-step gradient update and proper step size $\\zeta$ . See the derivation of the optimization process and the parameter tuning in Appendix B.", + "bbox": [ + 212, + 604, + 782, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During the iterations, $\\pmb{x}_0(k), \\forall k \\in \\{0,1,\\dots,K - 1\\}$ are not exactly on the clean manifold. We minimize its negative log-likelihood through diffusion model, resulting in it lying on an estimated clean manifold. Thus, we call our method Estimated Clean Manifold (ECM) Guidance.", + "bbox": [ + 212, + 651, + 782, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Reference Joint Trajectories. To generate trajectories with low guidance cost, we are essentially searching low-cost trajectories within the high-likelihood region. At the same time, joint trajectory distribution is a multi-modal distribution resulting from road topologies and different decision variables, meaning the likelihood has multiple peaks. This leads to the optimal solution of Problem 9 having multiple local optimals, and each optimal is far away from the other. Guided sampling methods [15,16,28,49], including our method ECM, suffer from", + "bbox": [ + 212, + 734, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/35e3bb102b3b53c1b1e68f1e89625642576cf47a0668139d544a34789a2fcb50.jpg", + "image_caption": [ + "Fig. 2: Two challenges with multi-peak function optimization: 1) Gradients may lead to suboptimal local optima (left); 2) There exist regions with low likelihood but high guidance cost uncertainty, leading to instability (right). Our approach can bypass the lengthy paths between peaks, search for better optima, and avoid uncertain areas." + ], + "image_footnote": [], + "bbox": [ + 243, + 146, + 756, + 265 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "two challenges (See Fig. 2): 1) it can be trapped at the local optimal around the initial position; 2) it takes massive efforts to drag the sample from one peak to another and transferring from one modal to another will need to pass through the region of low-likelihood (off clean-manifold), leading to numerical instability.", + "bbox": [ + 212, + 361, + 784, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To overcome this, we generate high-likelihood reference joint trajectories, choose the best one as the initialization. Note that combinations of samples with high marginal likelihood tend to exhibit high joint likelihood. Therefore, we can utilize the marginal sample set $\\mathcal{R} = \\{\\mathcal{R}_i\\}_{i=1}^n$ obtained from pre-trained marginal models to generate the references. Specifically, for iteration $k$ , we", + "bbox": [ + 212, + 422, + 429, + 632 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "construct candidate joint trajectory set $\\mathcal{R} \\otimes \\hat{\\pmb{x}}_0(k) = \\{[\\pmb{w}_1, \\pmb{w}_2, \\dots, \\pmb{w}_n] | \\pmb{w}_i \\in \\mathcal{R}_i \\cup \\{\\hat{\\pmb{x}}_{0,i}\\}, i = 1, 2, \\dots, n\\}$ . We calculate the guidance cost of all possible combinations, $\\mathcal{J}(w)$ , $w \\in \\mathcal{R} \\otimes \\hat{\\pmb{x}}_0(k)$ , and choose the minimal-cost one as the reference. The guided sampling algorithm, ECM with reference joint trajectories (ECMR), is introduced in Algorithm 1.", + "bbox": [ + 212, + 632, + 784, + 709 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 729, + 375, + 744 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 757, + 424, + 772 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We use Argoverse 2 [44], a widely used and large-scale trajectory prediction dataset, to test the effectiveness of our approaches for joint trajectory prediction and controllable generation. It has a large observation window of 5s and a long prediction horizon of 6s.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/48a5a8eee9a6c4266adf53c7fb8417137de55d41071e0edf6f540852a6a665ce.jpg", + "table_caption": [ + "Algorithm 1: ECMR" + ], + "table_footnote": [], + "table_body": "
Input: J(·), {αt,βt}T-1t=0, {tk}K-1k=0
1: x0(K) ~ N(0, Σp)
2: for k = K-1,..., 1 do
3: ε ~ N(0, Σp)
4: xtk(k) = √αtkx0(k) + √1-αtkε
5: x0(k) = 1/√αtk(xtk(k) - √1-αtk)εθ(xtk(k),t))
6: x0(k) = arg min J(w), w ∈ R⊗ x0(k)
7: x0(k-1) ← x0(k) - ζ∇x0(k)J(x0(k))
8: end for
Output: x0(0)
", + "bbox": [ + 439, + 474, + 785, + 597 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. We use the fixed scene context encoder of pretrained QCNet [50] to extract compact and representative context features from context information $\\pmb{c}$ . Then, we utilize a cross-attention layer to update the intermediate noisy data $\\pmb{x}_t$ with multiple contexts, including the history encodings of the target agent, the map encodings, the neighboring agents' encodings. Inspired by [7], we also add a cross-attention layer to update $\\pmb{x}_t$ with the diverse marginal trajectory samples $\\pmb{r}_i^l$ and its corresponding likelihood. In addition, we use self-attention to allow the interaction between $\\pmb{x}_{t,i}$ and $\\pmb{x}_{t,j}$ . Then the model predicts the noise $\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t)$ . According to [16], compact trajectory representation helps the diffusion model to generate high-quality trajectories efficiently. Inspired by this, we also learn a linear mapping between the 10-dimensional latent and 120-dimensional trajectories. Similar to [16], we design a rapid sample clustering algorithm so that we can generate a representative joint trajectory set. To increase the efficiency of sampling, we use DDIM [35] to accelerate the inference, and the DDIM step stride is 10. See Appendix C for details and analysis.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Joint trajectory prediction", + "text_level": 1, + "bbox": [ + 215, + 400, + 482, + 416 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We now evaluate OGD for joint trajectory prediction. Given $K$ joint trajectories, the evaluation metrics are 1) avgMinFDE $_K$ /avgMinADE $_K$ : the average of lowest final/average displacement error (FDE/ADE) of joint trajectory samples; 2) actorMR $_K$ : the rate of trajectory predictions that are considered to be \"missed\" ( $>2m$ FDE) in the lowest minFDE joint trajectory samples; 3) actorCR $_K$ : the rate of collisions across \"best\" (lowest avgMinFDE) joint trajectory samples; 4) avgBrierMinFDE $_K$ : calculated similarly to avgMinFDE $_K$ but scaled by the probability score of joint trajectory samples. We denote metrics with superscript \"\\*\" as those after sample clustering (see Appendix C.3).", + "bbox": [ + 212, + 430, + 787, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As a baseline, we train a vanilla diffusion (VD) baseline that shares the same neural network architecture as OGD. Specifically, we train an Optimal Gaussian Diffusion model with diffusion time $T_{train} = 100$ (OGD), vanilla diffusion with two different diffusion times $T_{train} = 100,500$ ( $\\mathrm{VD}_{100}, \\mathrm{VD}_{500}$ ). Note that $T$ is denoted as the diffusion time from which the reverse diffusion process starts. First, during the inference, we change $T$ and evaluate OGD and $\\mathrm{VD}_{100}$ who have the same $T_{train} = 100$ . Figure 3 shows that, with the decrease of the reverse steps, avgMinFDE128 of OGD keeps lower than $\\mathrm{VD}_{100}$ and it is more stable with the change of $T$ . It is also interesting to note that the performance of OGD even becomes better in the early stage. We hypothesize that it is because that $\\mathrm{KL}[p_T(\\boldsymbol{x}_T)||p_{prior}]$ of OGD is small and the accumulated error of $\\mathcal{G}(\\boldsymbol{x}_\\theta, T)$ is sufficiently reduced by a small $T$ . It also shows an advantage of OGD, which is that it is easy for OGD to tune a suitable $T$ for reverse diffusion without training an additional model for every $T$ [48]. Table 1 shows that OGD outperforms $\\mathrm{VD}_{500}$ with only 40 diffusion steps.", + "bbox": [ + 212, + 566, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compare our OGD with the other state-of-the-art methods on the Argoverse multi-world leaderboard; our approach OGD ranks $4^{\\mathrm{th}}$ on the leaderboard ranked by avgBrierMinFDE $_K$ , which demonstrates the effectiveness of our", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 113, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b007635a1852e6cad01ff697f59805204a1361e9b002412795d7953320f36e96.jpg", + "image_caption": [ + "Fig.3: Evaluation of Optimal Gaussian Diffusion and vanilla diffusion over reverse steps $T$ ." + ], + "image_footnote": [], + "bbox": [ + 225, + 157, + 415, + 272 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8308f5bc1680249b79ca240f51227f0425b79156f8376521c53f99ee4ff5a903.jpg", + "table_caption": [ + "Table 1: Evaluation on joint trajectory prediction task. For each metric, the best result is in **bold** and the second best result is **underlined**. $T = 70$ is the best $T$ from Fig. 3. $T = 40$ is the minimal diffusion time when OGD outperforms $\\mathrm{VD}_{500}$ on all metrics." + ], + "table_footnote": [], + "table_body": "
ModelTavgMinFDE*avgMinADE*avgMinFDE128avgMinADE128
VD1001000.621.380.490.97
VD5005000.611.360.480.91
OGD1000.601.320.430.81
OGD700.591.310.420.75
OGD400.611.340.470.77
", + "bbox": [ + 444, + 237, + 782, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "OGD framework. Note that we only list the entries with publications or technical reports in Tab. 2. Please refer to the official website for the full leaderboard.", + "bbox": [ + 212, + 352, + 782, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Controllable Generation", + "text_level": 1, + "bbox": [ + 215, + 402, + 460, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tasks. Future behaviors of vehicles can be effectively can be effectively represented by a set of goal points [9, 47] such as acceleration, braking, and right or left turn. Generating such diverse modes of joint trajectories is a typical motivation for using controllable generation. Thus, we study the controllable generation task where the goal is to reach diverse goal points at a specific time to fully test our guided sampling method. Considering that goal points should lie in realistic routes, and the trajectories to reach such goal points should also lie in such routes, we generate such target goal points in some realistic routes $\\mathrm{Route}(\\tau_g)$ . The guidance cost function can be expressed as", + "bbox": [ + 212, + 428, + 787, + 564 + ], + "page_idx": 11 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {J} \\left(\\boldsymbol {x} _ {0}\\right) = \\frac {1}{n} \\left\\| \\operatorname {P o s i t i o n} \\left(\\boldsymbol {x} _ {0}, \\tau_ {d}\\right) - \\operatorname {R o u t e} \\left(\\tau_ {g}\\right) \\right\\| _ {2} ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 571, + 784, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $\\mathrm{Position}(\\pmb{x}_0, \\tau_d)$ is the positions of $\\pmb{x}_0$ at time $\\tau_d$ . We choose ground-truth trajectories and a random combination set of diverse marginal samples, i.e., $\\mathcal{U} = \\{[\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_n] | \\pmb{u}_i \\in \\mathcal{R}_i, i = 1, 2, \\dots, n\\}$ as the realistic routes. We denote", + "bbox": [ + 212, + 609, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/df3971bf348b37f04daa90a3a07e3a1cf03b57f73d5fd57dfc17d46b2d49541a.jpg", + "table_caption": [ + "Table 2: Quantitative results on the Argoverse 2 Multi-world Forecasting leaderboard. For each metric, the best result is in bold and the second best result is underlined." + ], + "table_footnote": [], + "table_body": "
ModelavgMinFDE*6avgMinFDE1*actorMR6avgMinADE*6avgMinADE1*avgBrierMinFDE*6actorCR6
QCXet [51]1.022.290.130.500.941.650.01
Gnet [6]1.463.050.190.691.232.120.01
Forecast-MAE [2]1.553.330.190.691.302.240.01
FJMP [30]1.894.000.230.811.522.590.01
OGD (Ours)1.312.710.170.601.081.950.01
", + "bbox": [ + 220, + 744, + 787, + 825 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cd2c4fc275eb274906f664ebf2d8d0a023fbc099e1021823ade9e6ce6ac94c8f.jpg", + "image_caption": [ + "Fig. 4: Evaluation on controllable generation: route set $\\mathbf{U}$ and Deceleration. Magenta diamonds represent goal points. In the first (second) row, goal points are set at the fork lane (right lane). NNM [49] and SF [16, 28] struggle to drag samples from one modal to another. Our methods can achieve better guidance effectiveness and realism." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 352, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0d617d43fdb73080ada75572219dd3e8fb80437fc7d31d06249af1e3ab6d25d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 146, + 496, + 244 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f2826df3fa6be27717cf866bb6e0d50a84b0dd9747e514d6b0866fb580ee98ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 146, + 643, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f5820e3377fcdc4969a8b6dc291ffff9c9a6ee1effda1a9146dd7bf9bbe84d4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 145, + 785, + 244 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the former as GT set and the latter as $\\mathbf{U}$ set. The underlying assumption is that diverse samples from a good marginal trajectory predictor are realistic. We design different velocity settings to cover the diverse controllable generation tasks in autonomous driving: first is Normal Speed (N), $\\tau_{d} = \\tau_{g} = 6s$ ; second is Acceleration (A), $\\tau_{d} = 5s < \\tau_{g} = 6s$ ; third is Deceleration (D), $\\tau_{d} = 6s < \\tau_{g} = 5s$ .", + "bbox": [ + 212, + 340, + 787, + 430 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Metrics. We use the following metrics to evaluate controllable generation performance: Joint Route Deviation Error (JRDE), which measures the displacements to the routes to evaluate the realism, and Joint Final Displacement Error (JFDE), which evaluates the guidance effectiveness. We also evaluate from the \"min\" and \"mean\" perspectives: The \"min\" metric considers the best sample's performance, while the \"mean\" metric assesses the ratio of valid samples.", + "bbox": [ + 212, + 431, + 787, + 521 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Baselines. Guided sampling in controllable generation is mainly divided into two approaches: the first is to directly calculate $\\nabla_{\\boldsymbol{x}_t}\\mathcal{J}(\\boldsymbol{x}_t)$ [49]; the second is to calculate $\\nabla_{\\boldsymbol{x}_t}\\mathcal{J}(\\hat{\\boldsymbol{x}}_0)$ [16, 28]. We denote the former as Next Noisy Mean Guidance (NNM) and the latter as Score Function Guidance (SF). For a fair comparison, we use one guidance step followed by one DDIM step. We also tune the gradient step size for different guided sampling with Optimal Gaussian Diffusion and vanilla diffusion and report the results with the optimal step size. See Appendix D for the details of baseline derivation and step size tuning. We evaluate the following experiments with 128 joint trajectory samples.", + "bbox": [ + 212, + 522, + 787, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Evaluation. First, we evaluate the performance and efficiency with our diffusion model (OGD) and guided sampling method (ECM and ECMR) in Tab. 3, which demonstrates our methods can generate more realistic and effective samples with 5 times less DDIM steps. In addition, with reference joint trajectories, ECM significantly improves 'mean' metrics, indicating it addresses the issues discussed in Sec. 4.2 to a certain extent. Second, we compare solely on different guided sampling methods using the same OGD model shown in Fig. 5. Our ECM achieves better performance both in guidance effectiveness and realism. In Fig. 4, our ECM can generate trajectories that reach goal points more closely than NNM and SF. And with reference joint trajectory, ECMR can easily move samples from one modal to another. Third, we evaluate the average inference time and average GPU memory usage in Tab. 4. Our methods can generate", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 113, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/26364905998bd8627052369ae1edfe20eee12ec9554b483e2f66014c433264c7.jpg", + "table_caption": [ + "Table 3: Evaluation on controllable generation: route set U and Deceleration. For each metric, the best result is in bold and the second best result is underlined." + ], + "table_footnote": [], + "table_body": "
ModelSamplingDDIM StepsGuidance minJFDEEffectiveness meanJFDERealism minJRDEmeanJRDE
VD500No Guid501.9615.2290.1650.492
VD500NNM [49]500.7782.9130.1300.309
VD500SF [16, 28]500.5382.3390.1580.500
OGDNo Guid101.7725.1720.1380.469
OGDECM (Ours)100.0720.2370.1280.236
OGDECMR (Ours)100.0530.1460.1100.154
", + "bbox": [ + 228, + 196, + 776, + 330 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/37ff5e2a75e2db6356b4fabdda5bd2949af8aafc47e61308f7aa0cd4c341a16f.jpg", + "image_caption": [ + "Fig. 5: Evaluation of different guided sampling methods on various tasks. See Appendix E for other metrics." + ], + "image_footnote": [], + "bbox": [ + 217, + 359, + 475, + 459 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/4193b86ee1a4e92ad77051a649d0130efea584c828aa3aace4f6eec13116e957.jpg", + "table_caption": [ + "Table 4: We evaluate the average inference time per step and GPU incremental memory on $\\mathbf{U} +$ Deceleration. We test on single RTX A6000, batch size is 16 and number of samples is 128." + ], + "table_footnote": [], + "table_body": "
SamplingminJFDEminJRDEtime(ms)memory (GB)
NNM [49]0.7240.1191133.21
SF [16,28]0.1550.1262477.96
ECM(Ours)0.0750.1161113.21
ECMR(Ours)0.0530.1101163.22
", + "bbox": [ + 503, + 444, + 782, + 510 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "realistic trajectories satisfying guidance quite well with low inference time and GPU memory usage. More results on controllable generation can be found in Appendix E.", + "bbox": [ + 212, + 545, + 787, + 592 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 619, + 359, + 636 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance to significantly improve the computational efficiency and performance of diffusion models in autonomous driving. These methodologies enable a substantial reduction in inference steps and computational demands while ensuring enhanced joint trajectory prediction and controllable generation capabilities. Our approaches and experimental results underscore the potential of diffusion models for real-time applications in dynamic environments, marking a pivotal advancement in the deployment of diffusion models for autonomous driving. One limitation of the current implementation is that the performance is affected by the accuracy of marginal trajectory predictors. Enhancements could be achieved with superior marginal models or by directly learning joint predictions' mean and variance, areas for future work.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by Berkeley DeepDrive.", + "bbox": [ + 240, + 176, + 609, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 215, + 321, + 231 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18208-18218 (2022)", + "2. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023)", + "3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)", + "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "5. Franzese, G., Rossi, S., Yang, L., Finamore, A., Rossi, D., Filippone, M., Michiardi, P.: How much is enough? a study on diffusion times in score-based generative models. Entropy 25(4), 633 (2023)", + "6. Gao, X., Jia, X., Li, Y., Xiong, H.: Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolutional recurrent networks. IEEE Robotics and Automation Letters 8(5), 2946-2953 (2023)", + "7. Gilles, T., Sabatini, S., Tsishkou, D., Stanciulescu, B., Moutarde, F.: Thomas: Trajectory heatmap output with learned multi-agent sampling. arXiv preprint arXiv:2110.06607 (2021)", + "8. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural information processing systems. pp. 2672-2680 (2014)", + "9. Gu, J., Sun, C., Zhao, H.: Densetnt: End-to-end trajectory prediction from dense goal sets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15303-15312 (2021)", + "0. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022)", + "1. Guo, Z., Gao, X., Zhou, J., Cai, X., Shi, B.: Scenedm: Scene-level multiagent trajectory generation with consistent diffusion models. arXiv preprint arXiv:2311.15736 (2023)", + "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "3. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022)", + "4. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005)", + "5. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022)" + ], + "bbox": [ + 225, + 247, + 785, + 815 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 114, + 732, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_footnote", + "text": "$^{5}$ https://deepdrive.berkeley.edu", + "bbox": [ + 218, + 824, + 437, + 839 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motion-diffuser: Controllable multi-agent motion prediction using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9644-9653 (2023)", + "17. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. Advances in neural information processing systems 34, 21696-21707 (2021)", + "18. Lan, Z., Jiang, Y., Mu, Y., Chen, C., Li, S.E., Zhao, H., Li, K.: Sept: Towards efficient scene representation learning for motion prediction. arXiv preprint arXiv:2309.15289 (2023)", + "19. Laumanns, M., Thiele, L., Zitzler, E.: An efficient, adaptive parameter variation scheme for metaheuristics based on the epsilon-constraint method. European Journal of Operational Research 169(3), 932-942 (2006)", + "20. Lin, H., Wang, Y., Huo, M., Peng, C., Liu, Z., Tomizuka, M.: Joint pedestrian trajectory prediction through posterior sampling. arXiv preprint arXiv:2404.00237 (2024)", + "21. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022)", + "22. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021)", + "23. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022), https://arxiv.org/abs/2108.01073", + "24. Nayakanti, N., Al-Rfou, R., Zhou, A., Goel, K., Refaat, K.S., Sapp, B.: Wayformer: Motion forecasting via simple & efficient attention networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2980-2987. IEEE (2023)", + "25. Ngiam, J., Caine, B., Vasudevan, V., Zhang, Z., Chiang, H.T.L., Ling, J., Roelofs, R., Bewley, A., Liu, C., Venugopal, A., et al.: Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417 (2021)", + "26. Peng, C., Wang, G., Lo, X.W., Wu, X., Xu, C., Tomizuka, M., Zhan, W., Wang, H.: Delflow: Dense efficient learning of scene flow for large-scale point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16901-16910 (2023)", + "27. Peng, C., Xu, C., Wang, Y., Ding, M., Yang, H., Tomizuka, M., Keutzer, K., Pavone, M., Zhan, W.: Q-slam: Quadric representations for monocular slam. arXiv preprint arXiv:2403.08125 (2024)", + "28. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023)", + "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "30. Rowe, L., Ethier, M., Dykhne, E.H., Czarnecki, K.: Fjmp: Factorized joint multiagent motion prediction over learned directed acyclic interaction graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13745-13755 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512 (2022)", + "32. San-Roman, R., Nachmani, E., Wolf, L.: Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600 (2021)", + "33. Sherali, H.D., Soyster, A.L.: Preemptive and nonpreemptive multi-objective programming: Relationship and counterexamples. Journal of Optimization Theory and Applications 39, 173-186 (1983)", + "34. Shi, S., Jiang, L., Dai, D., Schiele, B.: Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. arXiv preprint arXiv:2306.17770 (2023)", + "35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "36. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: International Conference on Learning Representations (2022)", + "37. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. arXiv preprint arXiv:2303.01469 (2023)", + "38. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)", + "39. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "40. Sun, Q., Huang, X., Gu, J., Williams, B.C., Zhao, H.: M2i: From factored marginal trajectory prediction to interactive prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6543-6552 (2022)", + "41. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)", + "42. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022)", + "43. Watson, D., Chan, W., Ho, J., Norouzi, M.: Learning fast samplers for diffusion models by differentiating through sample quality. arXiv preprint arXiv:2202.05830 (2022)", + "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)", + "45. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023)", + "46. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902 (2022)", + "47. Zhao, H., Gao, J., Lan, T., Sun, C., Sapp, B., Varadarajan, B., Shen, Y., Shen, Y., Chai, Y., Schmid, C., et al.: Tnt: Target-driven trajectory prediction. In: Conference on Robot Learning. pp. 895-904. PMLR (2021)", + "48. Zheng, H., He, P., Chen, W., Zhou, M.: Truncated diffusion probabilistic models. arXiv preprint arXiv:2202.09671 (2022)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OptTrajDiff", + "bbox": [ + 648, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)", + "50. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023)", + "51. Zhou, Z., Wen, Z., Wang, J., Li, Y.H., Huang, Y.K.: Qcnext: A next-generation framework for joint multi-agent trajectory prediction (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 273 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_model.json b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d6d45d2bd7024619aaa981452efd155af22b5d4 --- /dev/null +++ b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_model.json @@ -0,0 +1,2359 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.782, + 0.186 + ], + "angle": 0, + "content": "Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.212, + 0.77, + 0.259 + ], + "angle": 0, + "content": "Yixiao Wang\\(^{1}\\), Chen Tang\\(^{1,2}\\), Lingfeng Sun\\(^{1}\\), Simone Rossi\\(^{3}\\), Yichen Xie\\(^{1}\\), Chensheng Peng\\(^{1}\\), Thomas Hannagan\\(^{3}\\), Stefano Sabatini\\(^{3}\\), Nicola Poerio\\(^{4}\\), Masayoshi Tomizuka\\(^{1}\\), and Wei Zhan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.27, + 0.642, + 0.285 + ], + "angle": 0, + "content": "1 University of California, Berkeley, USA" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.284, + 0.642, + 0.298 + ], + "angle": 0, + "content": "2 The University of Texas at Austin, USA" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.298, + 0.567, + 0.312 + ], + "angle": 0, + "content": "3 Stellantis, France" + }, + { + "type": "text", + "bbox": [ + 0.444, + 0.312, + 0.562, + 0.326 + ], + "angle": 0, + "content": "4 Stellantis, Italy" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.328, + 0.631, + 0.34 + ], + "angle": 0, + "content": "{yixiao_wang, wzhan}@berkeley.edu" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.372, + 0.744, + 0.583 + ], + "angle": 0, + "content": "Abstract. Diffusion models are promising for joint trajectory prediction and controllable generation in autonomous driving, but they face challenges of inefficient inference steps and high computational demands. To tackle these challenges, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance. OGD optimizes the prior distribution for a small diffusion time \\( T \\) and starts the reverse diffusion process from it. ECM directly injects guidance gradients to the estimated clean manifold, eliminating extensive gradient backpropagation throughout the network. Our methodology streamlines the generative process, enabling practical applications with reduced computational overhead. Experimental validation on the large-scale Argoverse 2 dataset demonstrates our approach's superior performance, offering a viable solution for computationally efficient, high-quality joint trajectory prediction and controllable generation for autonomous driving. Our project webpage is at https://yixiaowang7.github.io/OptTrajDiff_Page/" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.593, + 0.741, + 0.622 + ], + "angle": 0, + "content": "Keywords: Diffusion model \\(\\cdot\\) Autonomous driving \\(\\cdot\\) Trajectory Prediction \\(\\cdot\\) Controllable Trajectory Generation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.377, + 0.66 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The diffusion model is a class of generative models capable of representing high-dimensional data distributions. In particular, it has demonstrated strong performance in trajectory prediction and generation for autonomous driving [10, 11, 16, 28, 49]. In contrast to traditional trajectory prediction [18, 50] and generative models [41, 45], the unique advantage of diffusion models lies in their ability to deform the generated trajectory distribution to comply with additional requirements at inference stage via gradient-based guided sampling. Notably, it is achieved without extra model training costs. This ability to perform controllable trajectory generation enables various useful applications, such as enforcing additional realism constraints on predicted trajectories, generating directed and user-specified simulation scenarios." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "However, computational efficiency is a crucial bottleneck hindering the practical application of diffusion models in autonomous driving. Real-time inference is essential for trajectory prediction, as it provides timely forecasts of surrounding agents' behavior, enabling safe and efficient planning in dynamic traffic scenarios. The high demand for inference speed, coupled with limited onboard computational resources, makes it infeasible to deploy diffusion models for onboard trajectory prediction. While simulations do not occur onboard, lightweight models are still preferred to streamline the closed-loop training and evaluation pipelines. The heavy computational cost is mainly attributed to the following two aspects:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.79, + 0.455 + ], + "angle": 0, + "content": "Computationally Expensive Reverse Diffusion. At the inference stage, the diffusion model samples from standard Gaussian distribution and gradually denoises the sample through dynamics described by a Stochastic Differential Equation (SDE) [39], aiming to eventually obtain a sample as if drawn from a target data distribution. The target data distribution can be significantly different from a standard Gaussian distribution, necessitating a large number of denoising steps to yield good performance. Prior works have attempted to reduce the reverse diffusion steps through adaptive noise schedule [17, 32], fast samplers [21,35,43,46], or distillation [31,37]. However, the fixed standard Gaussian prior poses a challenge in accelerating the reverse diffusion process without violating the SDE, which can inevitably compromise the generation quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.79, + 0.61 + ], + "angle": 0, + "content": "Computationally Expensive Guided Sampling. Controllable generation is typically achieved by guiding the denoising process with the gradient of a guidance cost function. The guidance cost function encodes the desired characteristics of the generated data, which is typically defined on the clean data manifold in trajectory prediction and controllable generation problems. However, guided sampling intends to inject the gradient of the guidance cost function into the series of noisy data at intermediate diffusion steps. It requires a forward pass to estimate the clean data first and then back-propagating throughout the entire network to estimate the gradient with respect to the intermediate noisy data [16, 20, 28], which is very computationally intensive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.79, + 0.844 + ], + "angle": 0, + "content": "Targeted at these challenges, we take a step further to improve the computational efficiency of diffusion models while maintaining their performance for joint trajectory prediction and controllable generation tasks. Specifically, we propose two novel solutions for efficient reverse diffusion and guided sampling respectively. First, we present Optimal Gaussian Diffusion (OGD) to accelerate the reverse diffusion process. At the inference stage, instead of a standard Gaussian distribution far away from the desired data distribution, OGD starts from an optimal Gaussian distribution, which minimizes the distance to intermediate data distribution at a specific noise level, cutting down the diffusion steps before that. We show that we can analytically estimate such an optimal Gaussian distribution and also, an optimal perturbation kernel distribution, at any noise level from the statistics of the data distribution. It allows flexible tuning of the diffusion steps at the inference stage, without the need to train any additional models [48]. We further derive a practical implementation of OGD for joint trajectory prediction and generation, where the optimal Gaussian prior is computed" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.114, + 0.734, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.502, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.272, + 0.474, + 0.285 + ], + "angle": 0, + "content": "(a) Optimal Gaussian Diffusion (OGD)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.149, + 0.683, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.272, + 0.782, + 0.285 + ], + "angle": 0, + "content": "(b) Estimated Clean Manifold Guidance (ECM)" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.153, + 0.785, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 1: Overview of Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold Guidance (ECM). (a) OGD uses the mean and variance of the data distribution to calculate the optimal prior distribution at a small \\( T \\). It can largely reduce the diffusion time compared with vanilla diffusion. (b) ECM directly injects the gradient of guidance into the clean data manifold to mitigate computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.429 + ], + "angle": 0, + "content": "using the mean and variance of marginal trajectory distributions estimated with a pre-trained marginal trajectory prediction [18, 24, 50] model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Second, we propose Estimated Clean Manifold Guidance (ECM) to accelerate guided sampling for controllable generation. To save the computational cost due to estimating the guided gradient on the noisy data, we aim to directly inject the gradient into the clean data manifold without lengthy backpropagation. ECM is motivated by the insight that guided sampling can be regarded as a multi-objective optimization problem on the clean data manifold: The first objective is to maximize the likelihood of the samples on the estimated real data distribution; the second objective is to achieve low guidance cost. ECM hierarchically solves this multi-objective problem without backpropagation throughout the entire diffusion model. We show that it leads to faster inference time and much better performance than existing approaches. Also, to tackle the challenges imposed by the multi-modal nature of vehicle interactions, we propose to warm-start the multi-objective optimization problem with reference joint trajectory points estimated using a marginal trajectory predictor. We refer to the complete algorithm as Estimated Clean Manifold with Reference Joint Trajectory (ECMR)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To evaluate the proposed OGD and ECM methods on real-world tasks, we implement the OGD model leveraging a pre-trained marginal prediction model, QCNet [50], and conduct extensive experiments on the Argoverse 2 dataset. We show that OGD can achieve better joint trajectory prediction performance than vanilla diffusion with significantly fewer diffusion steps—it only takes \\(1/12\\) of the diffusion steps used by the vanilla diffusion model. OGD also achieves outstanding prediction accuracy compared to non-diffusion joint prediction models. In addition, ECMR, coupled with OGD, can generate samples with significantly lower guidance costs with the same level of realism scores compared to conducting controllable generation on the vanilla diffusion model using existing guided sampling approaches used in autonomous driving [16, 28], but using around \\(1/5\\) of their inference step." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models have demonstrated their ability to produce high-quality, diverse samples in a variety of applications, such as image, video, and 3D generation [13, 22, 29]. Recently, diffusion models have been applied to trajectory prediction in autonomous driving. It shows great performance on representing future trajectory distribution [10, 16]. However, diffusion models need to run lengthy reverse diffusion processes to generate high-quality samples [5, 10]. This makes it hard to apply diffusion models to trajectory prediction in autonomous driving since it requires in-time prediction for the downstream planning module to promptly respond to dynamically changing traffic scenarios. Previous works [17, 21, 31, 32, 35, 37, 43, 46] mostly focused on mitigating this issue by investigating how to solve reverse SDE in a faster manner. In addition, similar to ours, some works sought alternative initialization of the reverse diffusion process to achieve faster inference. For example, [3] initialized the reverse diffusion process with samples generated by another generative backbone network. However, the backbone network is not deliberately trained for the reverse diffusion process. [48] proposed to learn the initial diffusion with a generative adversarial network (GAN) [8]. However, it requires training a large additional model with a complex training procedure. Also, it requires specifying the reverse diffusion time as a hyperparameter prior to GAN training, which is hard to tune." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Guided sampling. Diffusion models have been successfully used to tackle controllable tasks through guided sampling, such as image inpainting [36] and motion planning [15]. A notable feature of diffusion models with guided sampling is their ability to condition the generation process on the user's preference, which was not available during the training phase. In the driving domain, recent works used guided sampling to generate controllable traffic [16, 49], and controllable pedestrian animation [28]. Our work belongs to the prior category, where user-specified guidance cost is used to guide generation in the trajectory space. In this case, the guidance cost function encodes certain desired properties of the generated trajectories. Thus, it is normally defined on the realistic trajectory samples, which are on the so-called clean manifold, instead of the noisy manifold containing the intermediate noisy data. Some works attempted to learn the guidance cost function on the noisy manifold [15]. Otherwise, it will lead to numerical instability when evaluating guidance cost at intermediate noisy data [28, 49]. To avoid the additional computational costs introduced by a learned guidance cost function, [16, 28] proposed to project the intermediate noisy data into the clean manifold through the diffusion model, and evaluate the guidance cost on the projected point. This approach requires back-propagating throughout the entire diffusion model, which is also computationally intensive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Trajectory prediction. In autonomous driving [26, 27], it is vital to precisely forecast how other participants in traffic will move in the future so that ego vehicle can plan a safe and efficient trajectory to execute in the future. Marginal trajectory prediction is used to predict the trajectory distribution for single vehicle and recent works involve kinematic constraints of the vehicles, restrictions of complex topology of roads, and interaction from the surrounding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "vehicles [18,24,50]. Recently, joint trajectory prediction has attracted attention from several researchers. It consists in predicting the joint future trajectories for all agents so that these trajectories are consistent with one another [25], an aspect which marginal trajectory prediction does not consider. Scene-Transformer [25] uses a fixed set of learnable scene embeddings to generate corresponding joint trajectories for all the vehicles in the given scene. Models like M2I [40], and FJMP [30] adopt a conditional approach, predicting the motions of other agents based on the movements of controlled agents. Diffusion model [16] has also been used to predict the joint trajectory. However, joint trajectory prediction is still a challenging problem since the complexity increases exponentially with the number of vehicles in the scene. The efficiency problem becomes more serious when using diffusion model to predict the joint trajectory distribution [16], and largely limits the application of diffusion models in autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.362, + 0.382, + 0.377 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.389, + 0.403, + 0.403 + ], + "angle": 0, + "content": "3.1 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.501 + ], + "angle": 0, + "content": "The diffusion process continuously perturbs the unknown data distribution \\( p_{data} \\) with a perturbation kernel and generates intermediate data with a given diffusion time \\( T \\). Denote the distribution of the time-dependent intermediate nosiy data \\( \\boldsymbol{x}_t \\) as \\( p_t(\\boldsymbol{x}_t) \\), \\( t \\in [0, T] \\), where \\( p_0(\\boldsymbol{x}_0) = p_{data} \\) is the clean data distribution and \\( \\boldsymbol{x}_0 \\) is the clean data. The series of intermediate data \\( \\boldsymbol{x}_t \\) are generated through the Stochastic Differential Equation (SDE) [39]:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.507, + 0.785, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\boldsymbol {x} _ {t} = f (\\boldsymbol {x} _ {t}, t) \\mathrm {d} t + g (t) \\mathrm {d} \\boldsymbol {w}, \\boldsymbol {x} _ {0} \\sim p _ {\\text {d a t a}} = p _ {0} (\\boldsymbol {x} _ {0}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.787, + 0.559 + ], + "angle": 0, + "content": "where \\( f(\\cdot, t) \\) is the drift coefficient, \\( g(t) \\) is the diffusion coefficient, and \\( \\boldsymbol{w} \\) is the Wiener process. We can recover \\( p_{data} \\) from reverse-time SDE" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.565, + 0.785, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\boldsymbol {x} _ {t} = \\left[ f (\\boldsymbol {x} _ {t}, t) - g (t) ^ {2} \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t}) \\right] \\mathrm {d} t + g (t) \\mathrm {d} \\bar {\\boldsymbol {w}}, \\boldsymbol {x} _ {T} \\sim p _ {T} (\\boldsymbol {x} _ {T}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.587, + 0.649, + 0.601 + ], + "angle": 0, + "content": "where \\(\\bar{\\boldsymbol{w}}\\) is another Wiener process, dt is negative timestep." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.753 + ], + "angle": 0, + "content": "In order to solve Eq. (2), we first need to approximate \\( p_T(\\pmb{x}_T) \\). In previous works, \\( p_T(\\pmb{x}_T) \\) is typically approximated by some prior distributions \\( p_{prior} \\) which contain no information of \\( p_{data} \\). In this paper, we adopt the setting of Variance Preserving (VP) SDE [12, 39]. In VP-SDE, \\( p_T(\\pmb{x}_T) \\approx p_{prior} \\) when \\( T \\to \\infty \\). The perturbation kernels are in the form of \\( p_t(\\pmb{x}_t|\\pmb{x}_0) = \\mathcal{N}(\\pmb{x}_t; \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0, (1 - \\bar{\\alpha}_t)\\pmb{\\Sigma}_p) \\) where scalar \\( \\bar{\\alpha}_t \\) is diffusion schedule parameter, \\( |\\pmb{\\Sigma}_p| = 1 \\). A common choice for \\( \\pmb{\\Sigma}_p \\) is the identity matrix \\( \\mathbf{I} \\) [12, 39]. Second, we need to approximate \\( \\nabla_{\\pmb{x}_t}\\log p_t(\\pmb{x}_t) \\) for all \\( t \\in (0,T] \\). Some works solve score-matching problem [14, 38] and learn a score function \\( s_\\theta(s_t,t) \\) to approximate \\( \\nabla_{\\pmb{x}_t}\\log p_t(\\pmb{x}_t) \\). In this paper, we follow the practice in DDPM [12] to learn the noise \\( \\epsilon \\) using a network \\( \\epsilon_\\theta(\\pmb{x}_t,t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.767, + 0.785, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {\\boldsymbol {x} _ {0} \\sim p _ {d a t a}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {\\Sigma} _ {p})} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_t = \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon},\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t) = -\\sqrt{1 - \\bar{\\alpha}_t}\\pmb{s}_{\\theta}(\\pmb{x}_t,t)\\). With these two approximations, we can learn \\(q_{\\theta}(\\pmb{x}_0)\\) to estimate unknown data distribution \\(p_0(\\pmb{x}_0)\\) solving Eq. (2) from \\(t = T\\) to \\(t = 0\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.406, + 0.163 + ], + "angle": 0, + "content": "3.2 Guided Sampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.207 + ], + "angle": 0, + "content": "In prior diffusion-based controllable generation frameworks [15, 16, 20, 28, 49], controllable generation is achieved via biasing of the score function for sampling:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.222, + 0.786, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {x} _ {t}} \\log \\left[ p _ {t} (\\boldsymbol {x} _ {t}) \\exp (- \\mathcal {C} (\\boldsymbol {x} _ {t})) \\right] = \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\mathcal {C} (\\boldsymbol {x} _ {t}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.368 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}(\\cdot)\\) is the guidance function. It requires estimating the guidance gradient with respect to the noisy data \\(\\pmb{x}_t\\). Some approaches introduce an additional neural network to approximate \\(\\mathcal{C}(\\cdot)\\) at different noisy levels [4,15,49]. The additional neural network imposes additional training costs and heavier computational burden at the inference stage. To this end, some works define an analytical guidance function \\(\\mathcal{J}(\\cdot)\\) on the clean data \\(\\pmb{x}_0\\). They first estimate \\(\\hat{\\pmb{x}}_0 = f_\\theta(\\pmb{x}_t)\\) based on \\(\\pmb{x}_t\\) with the diffusion model, then calculate \\(\\mathcal{C}(\\pmb{x}_t)\\) as \\(\\mathcal{J}(\\hat{\\pmb{x}}_0)\\). However, when taking the gradient of \\(\\mathcal{J}(\\hat{\\pmb{x}}_0)\\) with respect to \\(\\pmb{x}_t\\), we get" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.381, + 0.786, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {x} _ {t}} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0}) = \\nabla_ {\\hat {\\boldsymbol {x}} _ {0}} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0}) \\cdot \\nabla_ {\\boldsymbol {x} _ {t}} f _ {\\theta} (\\boldsymbol {x} _ {t}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.409, + 0.787, + 0.44 + ], + "angle": 0, + "content": "Estimating \\(\\nabla_{\\pmb{x}_t}f_\\theta (\\pmb{x}_t)\\) requires backpropagating throughout the entire diffusion model, i.e., \\(f_{\\theta}(\\pmb {x}_t)\\). It requires heavy computing resources and GPU memory." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.464, + 0.682, + 0.48 + ], + "angle": 0, + "content": "3.3 Trajectory Prediction and Controllable Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Joint trajectory prediction aims to predict the future joint trajectories \\( \\pmb{x}_0 \\) for all the vehicles in the scene, conditioned on context information \\( \\pmb{c} \\). It can be regarded as a conditional generation task where the goal is to train a generative model to approximate the distribution \\( p_0(\\pmb{x}_0|\\pmb{c}) \\). For simplicity, we omit \\( \\pmb{c} \\) and represent \\( p_0(\\pmb{x}_0|\\pmb{c}) \\) as \\( p_0(\\pmb{x}_0) \\). We denote \\( n \\) as the number of vehicles in the same scene and \\( \\pmb{x}_{0,i} \\) as the future trajectory for vehicle \\( i \\), \\( i \\in \\{1,2,\\dots,n\\} \\), so \\( \\pmb{x}_0 = [\\pmb{x}_{0,1},\\pmb{x}_{0,2},\\dots,\\pmb{x}_{0,n}] \\). Joint trajectory prediction can be very challenging. The complex interactions among vehicles, especially in highly interactive and dense traffic, result in a complicated high-dimensional \\( p_0(\\pmb{x}_0) \\), which is difficult to accurately model with lightweight and computationally efficient models. A simplified solution is to approximately decompose the joint trajectory distribution into marginal ones, i.e., \\( p_0(\\pmb{x}_0) \\approx \\prod_i^n p_0(\\pmb{x}_{0,i}) \\). It leads to the marginal trajectory prediction task, which has been extensively studied with mature solutions [18, 24, 50]. One drawback is that it omits the interactions among vehicles in the predicted horizon, which leads to large errors in highly interactive scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Controllable trajectory generation is closely related to trajectory prediction. In addition to generating realistic trajectory samples resembling the ground-truth \\( p_0(\\pmb{x}_0) \\), controllable generation imposes an additional objective—the generated trajectories should comply with specified guidance cost function \\( \\mathcal{J}(\\cdot) \\). We term the former objective as realism and the latter as guidance effectiveness. The guidance cost function \\( \\mathcal{J}(\\cdot) \\) can be some goal points of the vehicles, kinematic constraints, etc, which are mostly defined on the clean manifold rather than on the noisy data when it comes to diffusion-based controllable generation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.38, + 0.163 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.176, + 0.49, + 0.192 + ], + "angle": 0, + "content": "4.1 Optimal Gaussian Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.788, + 0.365 + ], + "angle": 0, + "content": "As discussed in Sec. 3.1, diffusion models typically select a non-informative prior distribution \\( p_{prior} \\), such as a standard Gaussian, as the initial data distribution for the reverse diffusion process. Such a non-informative \\( p_{prior} \\) is reasonable since \\( p_T(\\pmb{x}_T) \\) converges to it when \\( T \\to \\infty \\). However, it also means that a large \\( T \\) is required at inference time to yield good performance, which undermines computational efficiency and limits its wide real-time applications in autonomous driving. In this section, we aim to investigate a practical solution to tackle this challenge for joint trajectory prediction. Given the inherent limitation imposed by a non-informative prior, the key question we look into is: can we instead adopt an informative prior so that we can obtain the same level of performance with much smaller reverse diffusion steps?" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.788, + 0.456 + ], + "angle": 0, + "content": "First, we still consider a Gaussian prior, but with learnable parameters, i.e., we parameterize \\( p_{prior} \\) as \\( q_{\\phi}(T) = \\mathcal{N}(\\pmb{\\mu}, \\pmb{\\Sigma}) \\), where \\( \\pmb{\\mu} \\) and \\( \\pmb{\\Sigma} \\) are learnable. We aim to optimize \\( \\pmb{\\mu} \\) and \\( \\pmb{\\Sigma} \\) to enhance the generation performance for small \\( T \\). We got inspiration from [5], where an upper bound of the Kullback-Leibler divergence of the clean data distribution \\( p_0(\\pmb{x}_0) \\) and learned distribution \\( q_{\\theta}(\\pmb{x}_0) \\) was derived as a function of the diffusion time \\( T \\):" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.468, + 0.785, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} \\left[ p _ {0} \\left(\\boldsymbol {x} _ {0}\\right) \\mid \\mid q _ {\\theta} \\left(\\boldsymbol {x} _ {0}\\right) \\right] \\leq \\mathcal {G} \\left(\\boldsymbol {x} _ {\\theta}, T\\right) + \\operatorname {K L} \\left[ p _ {T} \\left(\\boldsymbol {x} _ {T}\\right) \\mid \\mid p _ {\\text {p r i o r}} \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.493, + 0.79, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}(\\pmb{x}_{\\theta},T)\\) is the positive accumulated error between \\(\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)\\) and \\(\\pmb{s}_{\\theta}(\\pmb{x}_t,t)\\) from 0 to \\(T\\) [5]. Note that \\(\\mathcal{G}(\\pmb{x}_{\\theta},T)\\) is an accumulated error so \\(\\mathcal{G}(\\pmb{x}_{\\theta},T_1)\\leq \\mathcal{G}(\\pmb{x}_{\\theta},T_2), T_1\\leq T_2\\). If we can achieve lower KL \\([p_{T_1}(\\pmb{x}_{T_1})||q_\\phi (T_1)]\\), then a tighter upper bound can be obtained. This opens up the possibility to achieve better performance with less diffusion time. Thus, we propose to optimize the prior distribution by minimizing KL \\([p_T(\\pmb{x}_T)||q_\\phi (T)]\\). As shown in Proposition 1 (See proof in Appendix A), it turns out that the optimal \\(\\mu\\) and \\(\\Sigma\\) can be expressed analytically as functions of the ground-truth data statistics. In addition, we find that we can further minimize the target KL divergence if we set a learnable \\(\\pmb{\\Sigma}_{p}\\) in the perturbation kernel \\(p_t(\\pmb {x}_t|\\pmb {x}_0) = \\mathcal{N}(\\pmb {x}_t;\\sqrt{\\bar{\\alpha}_t}\\pmb {x}_0,(1 - \\bar{\\alpha}_t)\\pmb {\\Sigma}_p)\\), whose optimal value can also be expressed as a function of the data statistics." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.669, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Proposition 1. Denote \\(\\pmb{\\mu}_d\\) and \\(\\pmb{\\Sigma}_d\\) as the mean and variance of \\(p_{data}\\). Denote \\(\\pmb{\\Sigma}^{*}(i,j)\\) and \\(\\pmb{\\Sigma}_{p}^{*}(i,j)\\) as the element at ith row and jth column of matrix \\(\\pmb{\\Sigma}^*\\) and \\(\\pmb{\\Sigma}_{p}^{*}\\). The optimal solution to \\(\\min KL[p_T(\\pmb{x}_T)||q_\\phi (\\pmb{x}_T,T)]\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.279, + 0.73, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\boldsymbol {\\mu} ^ {*} \\approx \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} \\\\ \\boldsymbol {\\Sigma} _ {p} ^ {*} (i, j) \\approx \\frac {1}{| \\boldsymbol {\\Sigma} _ {d} |} \\boldsymbol {\\Sigma} _ {d} (i, j) \\\\ \\boldsymbol {\\Sigma} ^ {*} (i, j) \\approx \\bar {\\alpha} _ {T} \\boldsymbol {\\Sigma} _ {d} + (1 - \\bar {\\alpha} _ {T}) ^ {2} \\boldsymbol {\\Sigma} _ {p} ^ {*} = (\\bar {\\alpha} _ {T} ^ {2} + \\frac {(1 - \\bar {\\alpha} _ {T}) ^ {2}}{| \\boldsymbol {\\Sigma} _ {d} |}) \\boldsymbol {\\Sigma} _ {d} (i, j) \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Thus, if we are able to estimate the mean and variance of the ground-truth data distribution, we can then analytically determine the optimal prior distribution \\(\\mathcal{N}(\\pmb{\\mu}^*,\\pmb{\\Sigma}^*)\\) and the optimal perturbation kernel variance \\(\\pmb{\\Sigma}_p^*\\) at any noise" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "level \\( T \\in [0, \\infty) \\). It leads to a crucial advantage against prior efforts on prior learning [48], where a target noise level has to be determined at training time in order to train an additional neural network to represent the prior distribution at the pre-selected noise level. In contrast, our method enables flexible tuning of the number of diffusion steps at inference time, without additional training costs. Since the learnable \\( q_{\\phi}(T) \\) and perturbation kernel are both Gaussian, we refer to our proposed diffusion framework as Optimal Gaussian Diffusion (OGD)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.79, + 0.374 + ], + "angle": 0, + "content": "For joint trajectory prediction and generation, we need to estimate the mean and variance of the joint trajectory distribution \\( p_0(\\pmb{x}_0) \\). It is not straightforward as only a limited number of trajectory samples exist in the dataset under the same context. Considering there exists mature and accurate marginal trajectory prediction models [18, 24, 50], we can conveniently extract statistics of the marginal trajectory distributions, i.e., \\( p_0(\\pmb{x}_{0,i}), i \\in \\{1,2,\\dots,n\\} \\) from a pretrained marginal trajectory predictor. Since Proposition 1 provides element-wise optimal value, we can easily get:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.379, + 0.803, + 0.442 + ], + "angle": 0, + "content": "Corollary 1. Denote \\(\\pmb{\\mu}_d(\\pmb{x}_{0,i})\\) and \\(\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})\\) as the mean and variance of marginal distribution for vehicle \\(i\\) and set both \\(\\pmb{\\Sigma}\\) and \\(\\pmb{\\Sigma}_p\\) are block-diagonal matrices where each block represents the marginal characteristics of one single vehicle. The optimal solution to \\(\\min KL[p_T(\\pmb{x}_T)||q_\\phi (\\pmb{x}_T,T)]\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.454, + 0.786, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\boldsymbol {\\mu} ^ {*} = \\left[ \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} (\\boldsymbol {x} _ {0, n}) \\right] ^ {T} \\\\ \\boldsymbol {\\Sigma} _ {p} ^ {*} = \\frac {1}{\\sum_ {i = 1} ^ {n} | \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0 , i}) |} d i a g [ \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, n}) ] \\\\ \\boldsymbol {\\Sigma} ^ {*} = \\bar {\\alpha} _ {T} d i a g [ \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, n}) ] + (1 - \\bar {\\alpha} _ {T}) ^ {2} \\boldsymbol {\\Sigma} _ {p} ^ {*} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.79, + 0.653 + ], + "angle": 0, + "content": "Corollary 1 implies that, if we further confine \\(\\pmb{\\Sigma}\\) and \\(\\pmb{\\Sigma}_p\\) to be block-diagonal without covariance between the states of different vehicles, then we can determine their optimal values purely from the estimated marginal statistics \\(\\pmb{\\mu}_d(\\pmb{x}_{0,i})\\) and \\(\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})\\), \\(i \\in \\{1,2,\\dots,n\\}\\), which enables a practical implementation of the proposed OGD model for joint trajectory prediction and generation tasks. Specifically, for vehicle \\(i\\), we leverage a pre-trained marginal trajectory predictor [18,24,50], predict diverse marginal trajectory sample set \\(\\mathcal{R}_i = \\{\\pmb{r}_i^l\\}_{l=1}^L\\) and corresponding likelihood set \\(\\{p(\\pmb{r}_i^l)\\}_{l=1}^L\\), and estimate \\(\\pmb{\\mu}_d(\\pmb{x}_{0,i})\\) and \\(\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})\\). For example, \\(\\pmb{\\mu}_d(\\pmb{x}_{0,i})\\) can be estimated as \\(\\frac{1}{L}\\sum_{l=L}p(\\pmb{r}_i^l)\\pmb{r}_i^l\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.668, + 0.694, + 0.683 + ], + "angle": 0, + "content": "4.2 Estimated Clean Manifold Guidance with Reference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "As discussed in Sec. 3.2, the intensive computation required for guided sampling comes from the calculation of \\(\\nabla_{\\boldsymbol{x}_t}f_\\theta (\\boldsymbol{x}_t)\\). Previous guided sampling approaches bias the score function defined on the intermediate noisy data \\(\\boldsymbol{x}_t\\). In this section, we aim to investigate whether we can inject the gradient directly into \\(\\boldsymbol{x}_0\\) to avoid the gradient propagation process. We first reformulate the controllable generation as a multi-objective optimization problem directly over \\(\\boldsymbol{x}_0\\) and propose an iterative algorithm to solve the formulated problem. In addition, we use reference trajectory points to create the region of interest, which helps solve the local optimal problem caused by multi-modal joint trajectory distribution and accelerate the guided sampling process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Estimated Clean Manifold Guidance. The objective of controllable generation of sample \\( \\pmb{x}_0 \\) includes two different objectives. The most important is the negative likelihood, ensuring the sample lies in the clean manifold. The second important is the guidance cost representing the user preference on the generated sample \\( \\pmb{x}_0 \\). This multi-objective optimization problem can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.235, + 0.784, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {x} _ {0}} \\left[ - \\log q _ {\\theta} (\\boldsymbol {x} _ {0}), \\mathcal {J} (\\boldsymbol {x} _ {0}) \\right] ^ {T} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.785, + 0.435 + ], + "angle": 0, + "content": "Inspired by lexicographic optimization [33], we solve this multi-objective optimization problem hierarchically. The main idea is to optimize each objective in the order of importance regardless of the degradation of the other less significant objectives. We first optimize the most important objective, \\(-\\log q_{\\theta}(\\pmb{x}_0)\\), to generate realistic and high-likelihood samples. The diffusion model achieves this goal effectively by reversing the diffusion process from noisy samples at specific noise level. However, exact noise level of current sample \\(\\pmb{x}_0\\) is unknown. To address this, we inject noise at level \\(t\\) into \\(\\pmb{x}_0\\), and then denoise it from \\(t\\) with learned diffusion model. This approach, similar to noise injection and denoising [1,23,37], improves the desired sample quality. We iteratively repeat this process \\(K\\) times, injecting guidance at each iteration to strengthen user preference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.784, + 0.466 + ], + "angle": 0, + "content": "Specifically, denote \\( \\pmb{x}_0(k) \\) as the sample at iteration \\( k \\). We first regenerate high-likelihood sample \\( \\hat{\\pmb{x}}_0(k) \\) by diffusion model:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.478, + 0.784, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {0} (k) \\leftarrow \\mathbb {E} \\left[ q _ {\\theta} \\left(\\boldsymbol {x} _ {0} (k) \\mid \\boldsymbol {x} _ {t _ {k}} (k)\\right) \\right] = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t _ {k}}}} \\left(\\boldsymbol {x} _ {t _ {k}} - \\sqrt {1 - \\bar {\\alpha} _ {t _ {k}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t _ {k}} (k), t _ {k}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.784, + 0.566 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_{t_k}(k) = \\sqrt{\\bar{\\alpha}_{t_k}}\\pmb{x}_0(k) + \\sqrt{1 - \\bar{\\alpha}_{t_k}}\\pmb{\\epsilon}, \\pmb{\\epsilon} \\sim \\mathcal{N}(\\pmb{0},\\pmb{\\Sigma}_p)\\). \\(t_k \\in [0,T]\\) is a tunable parameter. Then we minimize the guidance cost function with a small degradation of the most important objective \\(-\\log q_{\\theta}(\\pmb{x}_0)\\)," + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.58, + 0.784, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {0} (k - 1) \\leftarrow \\hat {\\boldsymbol {x}} _ {0} (k) - \\zeta \\nabla_ {\\hat {\\boldsymbol {x}} _ {0} (k)} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0} (k)) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.784, + 0.65 + ], + "angle": 0, + "content": "where \\(\\zeta\\) is the step size. Small degradation of \\(-\\log q_{\\theta}(\\pmb{x}_0)\\) is realized by one-step gradient update and proper step size \\(\\zeta\\). See the derivation of the optimization process and the parameter tuning in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.784, + 0.711 + ], + "angle": 0, + "content": "During the iterations, \\( \\pmb{x}_0(k), \\forall k \\in \\{0,1,\\dots,K - 1\\} \\) are not exactly on the clean manifold. We minimize its negative log-likelihood through diffusion model, resulting in it lying on an estimated clean manifold. Thus, we call our method Estimated Clean Manifold (ECM) Guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Reference Joint Trajectories. To generate trajectories with low guidance cost, we are essentially searching low-cost trajectories within the high-likelihood region. At the same time, joint trajectory distribution is a multi-modal distribution resulting from road topologies and different decision variables, meaning the likelihood has multiple peaks. This leads to the optimal solution of Problem 9 having multiple local optimals, and each optimal is far away from the other. Guided sampling methods [15,16,28,49], including our method ECM, suffer from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.147, + 0.758, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 2: Two challenges with multi-peak function optimization: 1) Gradients may lead to suboptimal local optima (left); 2) There exist regions with low likelihood but high guidance cost uncertainty, leading to instability (right). Our approach can bypass the lengthy paths between peaks, search for better optima, and avoid uncertain areas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.423 + ], + "angle": 0, + "content": "two challenges (See Fig. 2): 1) it can be trapped at the local optimal around the initial position; 2) it takes massive efforts to drag the sample from one peak to another and transferring from one modal to another will need to pass through the region of low-likelihood (off clean-manifold), leading to numerical instability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.431, + 0.633 + ], + "angle": 0, + "content": "To overcome this, we generate high-likelihood reference joint trajectories, choose the best one as the initialization. Note that combinations of samples with high marginal likelihood tend to exhibit high joint likelihood. Therefore, we can utilize the marginal sample set \\(\\mathcal{R} = \\{\\mathcal{R}_i\\}_{i=1}^n\\) obtained from pre-trained marginal models to generate the references. Specifically, for iteration \\(k\\), we" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.785, + 0.71 + ], + "angle": 0, + "content": "construct candidate joint trajectory set \\(\\mathcal{R} \\otimes \\hat{\\pmb{x}}_0(k) = \\{[\\pmb{w}_1, \\pmb{w}_2, \\dots, \\pmb{w}_n] | \\pmb{w}_i \\in \\mathcal{R}_i \\cup \\{\\hat{\\pmb{x}}_{0,i}\\}, i = 1, 2, \\dots, n\\}\\). We calculate the guidance cost of all possible combinations, \\(\\mathcal{J}(w)\\), \\(w \\in \\mathcal{R} \\otimes \\hat{\\pmb{x}}_0(k)\\), and choose the minimal-cost one as the reference. The guided sampling algorithm, ECM with reference joint trajectories (ECMR), is introduced in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.73, + 0.376, + 0.746 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.425, + 0.773 + ], + "angle": 0, + "content": "5.1 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Dataset. We use Argoverse 2 [44], a widely used and large-scale trajectory prediction dataset, to test the effectiveness of our approaches for joint trajectory prediction and controllable generation. It has a large observation window of 5s and a long prediction horizon of 6s." + }, + { + "type": "table_caption", + "bbox": [ + 0.455, + 0.459, + 0.62, + 0.473 + ], + "angle": 0, + "content": "Algorithm 1: ECMR" + }, + { + "type": "table", + "bbox": [ + 0.441, + 0.476, + 0.787, + 0.598 + ], + "angle": 0, + "content": "
Input: J(·), {αt,βt}T-1t=0, {tk}K-1k=0
1: x0(K) ~ N(0, Σp)
2: for k = K-1,..., 1 do
3: ε ~ N(0, Σp)
4: xtk(k) = √αtkx0(k) + √1-αtkε
5: x0(k) = 1/√αtk(xtk(k) - √1-αtk)εθ(xtk(k),t))
6: x0(k) = arg min J(w), w ∈ R⊗ x0(k)
7: x0(k-1) ← x0(k) - ζ∇x0(k)J(x0(k))
8: end for
Output: x0(0)
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Implementation Details. We use the fixed scene context encoder of pretrained QCNet [50] to extract compact and representative context features from context information \\( \\pmb{c} \\). Then, we utilize a cross-attention layer to update the intermediate noisy data \\( \\pmb{x}_t \\) with multiple contexts, including the history encodings of the target agent, the map encodings, the neighboring agents' encodings. Inspired by [7], we also add a cross-attention layer to update \\( \\pmb{x}_t \\) with the diverse marginal trajectory samples \\( \\pmb{r}_i^l \\) and its corresponding likelihood. In addition, we use self-attention to allow the interaction between \\( \\pmb{x}_{t,i} \\) and \\( \\pmb{x}_{t,j} \\). Then the model predicts the noise \\( \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t) \\). According to [16], compact trajectory representation helps the diffusion model to generate high-quality trajectories efficiently. Inspired by this, we also learn a linear mapping between the 10-dimensional latent and 120-dimensional trajectories. Similar to [16], we design a rapid sample clustering algorithm so that we can generate a representative joint trajectory set. To increase the efficiency of sampling, we use DDIM [35] to accelerate the inference, and the DDIM step stride is 10. See Appendix C for details and analysis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.401, + 0.483, + 0.417 + ], + "angle": 0, + "content": "5.2 Joint trajectory prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.567 + ], + "angle": 0, + "content": "We now evaluate OGD for joint trajectory prediction. Given \\(K\\) joint trajectories, the evaluation metrics are 1) avgMinFDE\\(_K\\)/avgMinADE\\(_K\\): the average of lowest final/average displacement error (FDE/ADE) of joint trajectory samples; 2) actorMR\\(_K\\): the rate of trajectory predictions that are considered to be \"missed\" (\\(>2m\\) FDE) in the lowest minFDE joint trajectory samples; 3) actorCR\\(_K\\): the rate of collisions across \"best\" (lowest avgMinFDE) joint trajectory samples; 4) avgBrierMinFDE\\(_K\\): calculated similarly to avgMinFDE\\(_K\\) but scaled by the probability score of joint trajectory samples. We denote metrics with superscript \"\\*\" as those after sample clustering (see Appendix C.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.794 + ], + "angle": 0, + "content": "As a baseline, we train a vanilla diffusion (VD) baseline that shares the same neural network architecture as OGD. Specifically, we train an Optimal Gaussian Diffusion model with diffusion time \\( T_{train} = 100 \\) (OGD), vanilla diffusion with two different diffusion times \\( T_{train} = 100,500 \\) (\\( \\mathrm{VD}_{100}, \\mathrm{VD}_{500} \\)). Note that \\( T \\) is denoted as the diffusion time from which the reverse diffusion process starts. First, during the inference, we change \\( T \\) and evaluate OGD and \\( \\mathrm{VD}_{100} \\) who have the same \\( T_{train} = 100 \\). Figure 3 shows that, with the decrease of the reverse steps, avgMinFDE128 of OGD keeps lower than \\( \\mathrm{VD}_{100} \\) and it is more stable with the change of \\( T \\). It is also interesting to note that the performance of OGD even becomes better in the early stage. We hypothesize that it is because that \\( \\mathrm{KL}[p_T(\\boldsymbol{x}_T)||p_{prior}] \\) of OGD is small and the accumulated error of \\( \\mathcal{G}(\\boldsymbol{x}_\\theta, T) \\) is sufficiently reduced by a small \\( T \\). It also shows an advantage of OGD, which is that it is easy for OGD to tune a suitable \\( T \\) for reverse diffusion without training an additional model for every \\( T \\) [48]. Table 1 shows that OGD outperforms \\( \\mathrm{VD}_{500} \\) with only 40 diffusion steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "We also compare our OGD with the other state-of-the-art methods on the Argoverse multi-world leaderboard; our approach OGD ranks \\(4^{\\mathrm{th}}\\) on the leaderboard ranked by avgBrierMinFDE\\(_K\\), which demonstrates the effectiveness of our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.158, + 0.416, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.282, + 0.436, + 0.324 + ], + "angle": 0, + "content": "Fig.3: Evaluation of Optimal Gaussian Diffusion and vanilla diffusion over reverse steps \\(T\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.442, + 0.157, + 0.788, + 0.227 + ], + "angle": 0, + "content": "Table 1: Evaluation on joint trajectory prediction task. For each metric, the best result is in **bold** and the second best result is **underlined**. \\( T = 70 \\) is the best \\( T \\) from Fig. 3. \\( T = 40 \\) is the minimal diffusion time when OGD outperforms \\( \\mathrm{VD}_{500} \\) on all metrics." + }, + { + "type": "table", + "bbox": [ + 0.446, + 0.238, + 0.784, + 0.314 + ], + "angle": 0, + "content": "
ModelTavgMinFDE*avgMinADE*avgMinFDE128avgMinADE128
VD1001000.621.380.490.97
VD5005000.611.360.480.91
OGD1000.601.320.430.81
OGD700.591.310.420.75
OGD400.611.340.470.77
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.385 + ], + "angle": 0, + "content": "OGD framework. Note that we only list the entries with publications or technical reports in Tab. 2. Please refer to the official website for the full leaderboard." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.462, + 0.419 + ], + "angle": 0, + "content": "5.3 Controllable Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.565 + ], + "angle": 0, + "content": "Tasks. Future behaviors of vehicles can be effectively can be effectively represented by a set of goal points [9, 47] such as acceleration, braking, and right or left turn. Generating such diverse modes of joint trajectories is a typical motivation for using controllable generation. Thus, we study the controllable generation task where the goal is to reach diverse goal points at a specific time to fully test our guided sampling method. Considering that goal points should lie in realistic routes, and the trajectories to reach such goal points should also lie in such routes, we generate such target goal points in some realistic routes \\(\\mathrm{Route}(\\tau_g)\\). The guidance cost function can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.573, + 0.785, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathcal {J} \\left(\\boldsymbol {x} _ {0}\\right) = \\frac {1}{n} \\left\\| \\operatorname {P o s i t i o n} \\left(\\boldsymbol {x} _ {0}, \\tau_ {d}\\right) - \\operatorname {R o u t e} \\left(\\tau_ {g}\\right) \\right\\| _ {2} ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathrm{Position}(\\pmb{x}_0, \\tau_d)\\) is the positions of \\(\\pmb{x}_0\\) at time \\(\\tau_d\\). We choose ground-truth trajectories and a random combination set of diverse marginal samples, i.e., \\(\\mathcal{U} = \\{[\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_n] | \\pmb{u}_i \\in \\mathcal{R}_i, i = 1, 2, \\dots, n\\}\\) as the realistic routes. We denote" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.682, + 0.784, + 0.71 + ], + "angle": 0, + "content": "Table 2: Quantitative results on the Argoverse 2 Multi-world Forecasting leaderboard. For each metric, the best result is in bold and the second best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.745, + 0.788, + 0.826 + ], + "angle": 0, + "content": "
ModelavgMinFDE*6avgMinFDE1*actorMR6avgMinADE*6avgMinADE1*avgBrierMinFDE*6actorCR6
QCXet [51]1.022.290.130.500.941.650.01
Gnet [6]1.463.050.190.691.232.120.01
Forecast-MAE [2]1.553.330.190.691.302.240.01
FJMP [30]1.894.000.230.811.522.590.01
OGD (Ours)1.312.710.170.601.081.950.01
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.353, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.147, + 0.497, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.147, + 0.644, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.146, + 0.787, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Fig. 4: Evaluation on controllable generation: route set \\(\\mathbf{U}\\) and Deceleration. Magenta diamonds represent goal points. In the first (second) row, goal points are set at the fork lane (right lane). NNM [49] and SF [16, 28] struggle to drag samples from one modal to another. Our methods can achieve better guidance effectiveness and realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.431 + ], + "angle": 0, + "content": "the former as GT set and the latter as \\(\\mathbf{U}\\) set. The underlying assumption is that diverse samples from a good marginal trajectory predictor are realistic. We design different velocity settings to cover the diverse controllable generation tasks in autonomous driving: first is Normal Speed (N), \\(\\tau_{d} = \\tau_{g} = 6s\\); second is Acceleration (A), \\(\\tau_{d} = 5s < \\tau_{g} = 6s\\); third is Deceleration (D), \\(\\tau_{d} = 6s < \\tau_{g} = 5s\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Metrics. We use the following metrics to evaluate controllable generation performance: Joint Route Deviation Error (JRDE), which measures the displacements to the routes to evaluate the realism, and Joint Final Displacement Error (JFDE), which evaluates the guidance effectiveness. We also evaluate from the \"min\" and \"mean\" perspectives: The \"min\" metric considers the best sample's performance, while the \"mean\" metric assesses the ratio of valid samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Baselines. Guided sampling in controllable generation is mainly divided into two approaches: the first is to directly calculate \\(\\nabla_{\\boldsymbol{x}_t}\\mathcal{J}(\\boldsymbol{x}_t)\\) [49]; the second is to calculate \\(\\nabla_{\\boldsymbol{x}_t}\\mathcal{J}(\\hat{\\boldsymbol{x}}_0)\\) [16, 28]. We denote the former as Next Noisy Mean Guidance (NNM) and the latter as Score Function Guidance (SF). For a fair comparison, we use one guidance step followed by one DDIM step. We also tune the gradient step size for different guided sampling with Optimal Gaussian Diffusion and vanilla diffusion and report the results with the optimal step size. See Appendix D for the details of baseline derivation and step size tuning. We evaluate the following experiments with 128 joint trajectory samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation. First, we evaluate the performance and efficiency with our diffusion model (OGD) and guided sampling method (ECM and ECMR) in Tab. 3, which demonstrates our methods can generate more realistic and effective samples with 5 times less DDIM steps. In addition, with reference joint trajectories, ECM significantly improves 'mean' metrics, indicating it addresses the issues discussed in Sec. 4.2 to a certain extent. Second, we compare solely on different guided sampling methods using the same OGD model shown in Fig. 5. Our ECM achieves better performance both in guidance effectiveness and realism. In Fig. 4, our ECM can generate trajectories that reach goal points more closely than NNM and SF. And with reference joint trajectory, ECMR can easily move samples from one modal to another. Third, we evaluate the average inference time and average GPU memory usage in Tab. 4. Our methods can generate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Evaluation on controllable generation: route set U and Deceleration. For each metric, the best result is in bold and the second best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.198, + 0.777, + 0.332 + ], + "angle": 0, + "content": "
ModelSamplingDDIM StepsGuidance minJFDEEffectiveness meanJFDERealism minJRDEmeanJRDE
VD500No Guid501.9615.2290.1650.492
VD500NNM [49]500.7782.9130.1300.309
VD500SF [16, 28]500.5382.3390.1580.500
OGDNo Guid101.7725.1720.1380.469
OGDECM (Ours)100.0720.2370.1280.236
OGDECMR (Ours)100.0530.1460.1100.154
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.361, + 0.476, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.471, + 0.476, + 0.514 + ], + "angle": 0, + "content": "Fig. 5: Evaluation of different guided sampling methods on various tasks. See Appendix E for other metrics." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.364, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Table 4: We evaluate the average inference time per step and GPU incremental memory on \\(\\mathbf{U} +\\) Deceleration. We test on single RTX A6000, batch size is 16 and number of samples is 128." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.445, + 0.784, + 0.511 + ], + "angle": 0, + "content": "
SamplingminJFDEminJRDEtime(ms)memory (GB)
NNM [49]0.7240.1191133.21
SF [16,28]0.1550.1262477.96
ECM(Ours)0.0750.1161113.21
ECMR(Ours)0.0530.1101163.22
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.593 + ], + "angle": 0, + "content": "realistic trajectories satisfying guidance quite well with low inference time and GPU memory usage. More results on controllable generation can be found in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.36, + 0.637 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this work, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance to significantly improve the computational efficiency and performance of diffusion models in autonomous driving. These methodologies enable a substantial reduction in inference steps and computational demands while ensuring enhanced joint trajectory prediction and controllable generation capabilities. Our approaches and experimental results underscore the potential of diffusion models for real-time applications in dynamic environments, marking a pivotal advancement in the deployment of diffusion models for autonomous driving. One limitation of the current implementation is that the performance is affected by the accuracy of marginal trajectory predictors. Enhancements could be achieved with superior marginal models or by directly learning joint predictions' mean and variance, areas for future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.177, + 0.61, + 0.193 + ], + "angle": 0, + "content": "This work was supported by Berkeley DeepDrive." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.216, + 0.323, + 0.232 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.248, + 0.787, + 0.289 + ], + "angle": 0, + "content": "1. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18208-18218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.291, + 0.786, + 0.331 + ], + "angle": 0, + "content": "2. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.786, + 0.387 + ], + "angle": 0, + "content": "3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.786, + 0.414 + ], + "angle": 0, + "content": "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.416, + 0.786, + 0.455 + ], + "angle": 0, + "content": "5. Franzese, G., Rossi, S., Yang, L., Finamore, A., Rossi, D., Filippone, M., Michiardi, P.: How much is enough? a study on diffusion times in score-based generative models. Entropy 25(4), 633 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.786, + 0.497 + ], + "angle": 0, + "content": "6. Gao, X., Jia, X., Li, Y., Xiong, H.: Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolutional recurrent networks. IEEE Robotics and Automation Letters 8(5), 2946-2953 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.499, + 0.786, + 0.539 + ], + "angle": 0, + "content": "7. Gilles, T., Sabatini, S., Tsishkou, D., Stanciulescu, B., Moutarde, F.: Thomas: Trajectory heatmap output with learned multi-agent sampling. arXiv preprint arXiv:2110.06607 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.786, + 0.58 + ], + "angle": 0, + "content": "8. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural information processing systems. pp. 2672-2680 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.622 + ], + "angle": 0, + "content": "9. Gu, J., Sun, C., Zhao, H.: Densetnt: End-to-end trajectory prediction from dense goal sets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15303-15312 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.786, + 0.663 + ], + "angle": 0, + "content": "0. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.665, + 0.786, + 0.704 + ], + "angle": 0, + "content": "1. Guo, Z., Gao, X., Zhou, J., Cai, X., Shi, B.: Scenedm: Scene-level multiagent trajectory generation with consistent diffusion models. arXiv preprint arXiv:2311.15736 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.707, + 0.786, + 0.733 + ], + "angle": 0, + "content": "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.735, + 0.786, + 0.76 + ], + "angle": 0, + "content": "3. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.763, + 0.786, + 0.788 + ], + "angle": 0, + "content": "4. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.79, + 0.786, + 0.816 + ], + "angle": 0, + "content": "5. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.248, + 0.787, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.438, + 0.84 + ], + "angle": 0, + "content": "\\(^{5}\\) https://deepdrive.berkeley.edu" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "16. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motion-diffuser: Controllable multi-agent motion prediction using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9644-9653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.231 + ], + "angle": 0, + "content": "17. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. Advances in neural information processing systems 34, 21696-21707 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "18. Lan, Z., Jiang, Y., Mu, Y., Chen, C., Li, S.E., Zhao, H., Li, K.: Sept: Towards efficient scene representation learning for motion prediction. arXiv preprint arXiv:2309.15289 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "19. Laumanns, M., Thiele, L., Zitzler, E.: An efficient, adaptive parameter variation scheme for metaheuristics based on the epsilon-constraint method. European Journal of Operational Research 169(3), 932-942 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "20. Lin, H., Wang, Y., Huo, M., Peng, C., Liu, Z., Tomizuka, M.: Joint pedestrian trajectory prediction through posterior sampling. arXiv preprint arXiv:2404.00237 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "21. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "22. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.48 + ], + "angle": 0, + "content": "23. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022), https://arxiv.org/abs/2108.01073" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.535 + ], + "angle": 0, + "content": "24. Nayakanti, N., Al-Rfou, R., Zhou, A., Goel, K., Refaat, K.S., Sapp, B.: Wayformer: Motion forecasting via simple & efficient attention networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2980-2987. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.59 + ], + "angle": 0, + "content": "25. Ngiam, J., Caine, B., Vasudevan, V., Zhang, Z., Chiang, H.T.L., Ling, J., Roelofs, R., Bewley, A., Liu, C., Venugopal, A., et al.: Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.646 + ], + "angle": 0, + "content": "26. Peng, C., Wang, G., Lo, X.W., Wu, X., Xu, C., Tomizuka, M., Zhan, W., Wang, H.: Delflow: Dense efficient learning of scene flow for large-scale point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16901-16910 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "27. Peng, C., Xu, C., Wang, Y., Ding, M., Yang, H., Tomizuka, M., Keutzer, K., Pavone, M., Zhan, W.: Q-slam: Quadric representations for monocular slam. arXiv preprint arXiv:2403.08125 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "28. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Rowe, L., Ethier, M., Dykhne, E.H., Czarnecki, K.: Fjmp: Factorized joint multiagent motion prediction over learned directed acyclic interaction graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13745-13755 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.65, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OptTrajDiff" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "31. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "32. San-Roman, R., Nachmani, E., Wolf, L.: Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "33. Sherali, H.D., Soyster, A.L.: Preemptive and nonpreemptive multi-objective programming: Relationship and counterexamples. Journal of Optimization Theory and Applications 39, 173-186 (1983)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.249, + 0.785, + 0.288 + ], + "angle": 0, + "content": "34. Shi, S., Jiang, L., Dai, D., Schiele, B.: Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. arXiv preprint arXiv:2306.17770 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.291, + 0.785, + 0.317 + ], + "angle": 0, + "content": "35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.319, + 0.785, + 0.359 + ], + "angle": 0, + "content": "36. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.361, + 0.785, + 0.387 + ], + "angle": 0, + "content": "37. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. arXiv preprint arXiv:2303.01469 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.39, + 0.785, + 0.416 + ], + "angle": 0, + "content": "38. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "39. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.785, + 0.501 + ], + "angle": 0, + "content": "40. Sun, Q., Huang, X., Gu, J., Williams, B.C., Zhao, H.: M2i: From factored marginal trajectory prediction to interactive prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6543-6552 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.503, + 0.785, + 0.543 + ], + "angle": 0, + "content": "41. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.546, + 0.785, + 0.613 + ], + "angle": 0, + "content": "42. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.615, + 0.785, + 0.655 + ], + "angle": 0, + "content": "43. Watson, D., Chan, W., Ho, J., Norouzi, M.: Learning fast samplers for diffusion models by differentiating through sample quality. arXiv preprint arXiv:2202.05830 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.657, + 0.785, + 0.698 + ], + "angle": 0, + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.7, + 0.785, + 0.74 + ], + "angle": 0, + "content": "45. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.742, + 0.785, + 0.769 + ], + "angle": 0, + "content": "46. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.771, + 0.785, + 0.811 + ], + "angle": 0, + "content": "47. Zhao, H., Gao, J., Lan, T., Sun, C., Sapp, B., Varadarajan, B., Shen, Y., Shen, Y., Chai, Y., Schmid, C., et al.: Tnt: Target-driven trajectory prediction. In: Conference on Robot Learning. pp. 895-904. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "48. Zheng, H., He, P., Chen, W., Zhou, M.: Truncated diffusion probabilistic models. arXiv preprint arXiv:2202.09671 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "49. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "50. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.274 + ], + "angle": 0, + "content": "51. Zhou, Z., Wen, Z., Wang, J., Li, Y.H., Huang, Y.K.: Qcnext: A next-generation framework for joint multi-agent trajectory prediction (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.274 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_origin.pdf b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76208d054769b40d68b82f284fd77e52e4810313 --- /dev/null +++ b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/595522ed-58d9-4a45-8703-3a11caa62b1f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:275ad4a5707a737216a24e8f41ca65983e43a275f4e8b0afb021cb8c03376e28 +size 5468914 diff --git a/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/full.md b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0ad95bd18157c0d965ffb576318c162a22627c86 --- /dev/null +++ b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/full.md @@ -0,0 +1,313 @@ +# Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation + +Yixiao Wang $^{1}$ , Chen Tang $^{1,2}$ , Lingfeng Sun $^{1}$ , Simone Rossi $^{3}$ , Yichen Xie $^{1}$ , Chensheng Peng $^{1}$ , Thomas Hannagan $^{3}$ , Stefano Sabatini $^{3}$ , Nicola Poerio $^{4}$ , Masayoshi Tomizuka $^{1}$ , and Wei Zhan $^{1}$ + +1 University of California, Berkeley, USA + +2 The University of Texas at Austin, USA + +3 Stellantis, France + +4 Stellantis, Italy + +{yixiao_wang, wzhan}@berkeley.edu + +Abstract. Diffusion models are promising for joint trajectory prediction and controllable generation in autonomous driving, but they face challenges of inefficient inference steps and high computational demands. To tackle these challenges, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance. OGD optimizes the prior distribution for a small diffusion time $T$ and starts the reverse diffusion process from it. ECM directly injects guidance gradients to the estimated clean manifold, eliminating extensive gradient backpropagation throughout the network. Our methodology streamlines the generative process, enabling practical applications with reduced computational overhead. Experimental validation on the large-scale Argoverse 2 dataset demonstrates our approach's superior performance, offering a viable solution for computationally efficient, high-quality joint trajectory prediction and controllable generation for autonomous driving. Our project webpage is at https://yixiaowang7.github.io/OptTrajDiff_Page/ + +Keywords: Diffusion model $\cdot$ Autonomous driving $\cdot$ Trajectory Prediction $\cdot$ Controllable Trajectory Generation + +# 1 Introduction + +The diffusion model is a class of generative models capable of representing high-dimensional data distributions. In particular, it has demonstrated strong performance in trajectory prediction and generation for autonomous driving [10, 11, 16, 28, 49]. In contrast to traditional trajectory prediction [18, 50] and generative models [41, 45], the unique advantage of diffusion models lies in their ability to deform the generated trajectory distribution to comply with additional requirements at inference stage via gradient-based guided sampling. Notably, it is achieved without extra model training costs. This ability to perform controllable trajectory generation enables various useful applications, such as enforcing additional realism constraints on predicted trajectories, generating directed and user-specified simulation scenarios. + +However, computational efficiency is a crucial bottleneck hindering the practical application of diffusion models in autonomous driving. Real-time inference is essential for trajectory prediction, as it provides timely forecasts of surrounding agents' behavior, enabling safe and efficient planning in dynamic traffic scenarios. The high demand for inference speed, coupled with limited onboard computational resources, makes it infeasible to deploy diffusion models for onboard trajectory prediction. While simulations do not occur onboard, lightweight models are still preferred to streamline the closed-loop training and evaluation pipelines. The heavy computational cost is mainly attributed to the following two aspects: + +Computationally Expensive Reverse Diffusion. At the inference stage, the diffusion model samples from standard Gaussian distribution and gradually denoises the sample through dynamics described by a Stochastic Differential Equation (SDE) [39], aiming to eventually obtain a sample as if drawn from a target data distribution. The target data distribution can be significantly different from a standard Gaussian distribution, necessitating a large number of denoising steps to yield good performance. Prior works have attempted to reduce the reverse diffusion steps through adaptive noise schedule [17, 32], fast samplers [21,35,43,46], or distillation [31,37]. However, the fixed standard Gaussian prior poses a challenge in accelerating the reverse diffusion process without violating the SDE, which can inevitably compromise the generation quality. + +Computationally Expensive Guided Sampling. Controllable generation is typically achieved by guiding the denoising process with the gradient of a guidance cost function. The guidance cost function encodes the desired characteristics of the generated data, which is typically defined on the clean data manifold in trajectory prediction and controllable generation problems. However, guided sampling intends to inject the gradient of the guidance cost function into the series of noisy data at intermediate diffusion steps. It requires a forward pass to estimate the clean data first and then back-propagating throughout the entire network to estimate the gradient with respect to the intermediate noisy data [16, 20, 28], which is very computationally intensive. + +Targeted at these challenges, we take a step further to improve the computational efficiency of diffusion models while maintaining their performance for joint trajectory prediction and controllable generation tasks. Specifically, we propose two novel solutions for efficient reverse diffusion and guided sampling respectively. First, we present Optimal Gaussian Diffusion (OGD) to accelerate the reverse diffusion process. At the inference stage, instead of a standard Gaussian distribution far away from the desired data distribution, OGD starts from an optimal Gaussian distribution, which minimizes the distance to intermediate data distribution at a specific noise level, cutting down the diffusion steps before that. We show that we can analytically estimate such an optimal Gaussian distribution and also, an optimal perturbation kernel distribution, at any noise level from the statistics of the data distribution. It allows flexible tuning of the diffusion steps at the inference stage, without the need to train any additional models [48]. We further derive a practical implementation of OGD for joint trajectory prediction and generation, where the optimal Gaussian prior is computed + +![](images/084ebe88e07559bafe43edf81a5887443d8208b771aa65c76fb05067f5ddf346.jpg) +(a) Optimal Gaussian Diffusion (OGD) + +![](images/478d8a8314156abe4e50cacc8cd7679f2ec49c114e6b21600fa39eee88cb2aec.jpg) +Fig. 1: Overview of Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold Guidance (ECM). (a) OGD uses the mean and variance of the data distribution to calculate the optimal prior distribution at a small $T$ . It can largely reduce the diffusion time compared with vanilla diffusion. (b) ECM directly injects the gradient of guidance into the clean data manifold to mitigate computational complexity. + +![](images/3e5c50c1457d8655f9228e597d7e61812dc7b9f955f88e6ac8ae4d56cd786193.jpg) +(b) Estimated Clean Manifold Guidance (ECM) + +using the mean and variance of marginal trajectory distributions estimated with a pre-trained marginal trajectory prediction [18, 24, 50] model. + +Second, we propose Estimated Clean Manifold Guidance (ECM) to accelerate guided sampling for controllable generation. To save the computational cost due to estimating the guided gradient on the noisy data, we aim to directly inject the gradient into the clean data manifold without lengthy backpropagation. ECM is motivated by the insight that guided sampling can be regarded as a multi-objective optimization problem on the clean data manifold: The first objective is to maximize the likelihood of the samples on the estimated real data distribution; the second objective is to achieve low guidance cost. ECM hierarchically solves this multi-objective problem without backpropagation throughout the entire diffusion model. We show that it leads to faster inference time and much better performance than existing approaches. Also, to tackle the challenges imposed by the multi-modal nature of vehicle interactions, we propose to warm-start the multi-objective optimization problem with reference joint trajectory points estimated using a marginal trajectory predictor. We refer to the complete algorithm as Estimated Clean Manifold with Reference Joint Trajectory (ECMR). + +To evaluate the proposed OGD and ECM methods on real-world tasks, we implement the OGD model leveraging a pre-trained marginal prediction model, QCNet [50], and conduct extensive experiments on the Argoverse 2 dataset. We show that OGD can achieve better joint trajectory prediction performance than vanilla diffusion with significantly fewer diffusion steps—it only takes $1/12$ of the diffusion steps used by the vanilla diffusion model. OGD also achieves outstanding prediction accuracy compared to non-diffusion joint prediction models. In addition, ECMR, coupled with OGD, can generate samples with significantly lower guidance costs with the same level of realism scores compared to conducting controllable generation on the vanilla diffusion model using existing guided sampling approaches used in autonomous driving [16, 28], but using around $1/5$ of their inference step. + +# 2 Related Work + +Diffusion models. Diffusion models have demonstrated their ability to produce high-quality, diverse samples in a variety of applications, such as image, video, and 3D generation [13, 22, 29]. Recently, diffusion models have been applied to trajectory prediction in autonomous driving. It shows great performance on representing future trajectory distribution [10, 16]. However, diffusion models need to run lengthy reverse diffusion processes to generate high-quality samples [5, 10]. This makes it hard to apply diffusion models to trajectory prediction in autonomous driving since it requires in-time prediction for the downstream planning module to promptly respond to dynamically changing traffic scenarios. Previous works [17, 21, 31, 32, 35, 37, 43, 46] mostly focused on mitigating this issue by investigating how to solve reverse SDE in a faster manner. In addition, similar to ours, some works sought alternative initialization of the reverse diffusion process to achieve faster inference. For example, [3] initialized the reverse diffusion process with samples generated by another generative backbone network. However, the backbone network is not deliberately trained for the reverse diffusion process. [48] proposed to learn the initial diffusion with a generative adversarial network (GAN) [8]. However, it requires training a large additional model with a complex training procedure. Also, it requires specifying the reverse diffusion time as a hyperparameter prior to GAN training, which is hard to tune. + +Guided sampling. Diffusion models have been successfully used to tackle controllable tasks through guided sampling, such as image inpainting [36] and motion planning [15]. A notable feature of diffusion models with guided sampling is their ability to condition the generation process on the user's preference, which was not available during the training phase. In the driving domain, recent works used guided sampling to generate controllable traffic [16, 49], and controllable pedestrian animation [28]. Our work belongs to the prior category, where user-specified guidance cost is used to guide generation in the trajectory space. In this case, the guidance cost function encodes certain desired properties of the generated trajectories. Thus, it is normally defined on the realistic trajectory samples, which are on the so-called clean manifold, instead of the noisy manifold containing the intermediate noisy data. Some works attempted to learn the guidance cost function on the noisy manifold [15]. Otherwise, it will lead to numerical instability when evaluating guidance cost at intermediate noisy data [28, 49]. To avoid the additional computational costs introduced by a learned guidance cost function, [16, 28] proposed to project the intermediate noisy data into the clean manifold through the diffusion model, and evaluate the guidance cost on the projected point. This approach requires back-propagating throughout the entire diffusion model, which is also computationally intensive. + +Trajectory prediction. In autonomous driving [26, 27], it is vital to precisely forecast how other participants in traffic will move in the future so that ego vehicle can plan a safe and efficient trajectory to execute in the future. Marginal trajectory prediction is used to predict the trajectory distribution for single vehicle and recent works involve kinematic constraints of the vehicles, restrictions of complex topology of roads, and interaction from the surrounding + +vehicles [18,24,50]. Recently, joint trajectory prediction has attracted attention from several researchers. It consists in predicting the joint future trajectories for all agents so that these trajectories are consistent with one another [25], an aspect which marginal trajectory prediction does not consider. Scene-Transformer [25] uses a fixed set of learnable scene embeddings to generate corresponding joint trajectories for all the vehicles in the given scene. Models like M2I [40], and FJMP [30] adopt a conditional approach, predicting the motions of other agents based on the movements of controlled agents. Diffusion model [16] has also been used to predict the joint trajectory. However, joint trajectory prediction is still a challenging problem since the complexity increases exponentially with the number of vehicles in the scene. The efficiency problem becomes more serious when using diffusion model to predict the joint trajectory distribution [16], and largely limits the application of diffusion models in autonomous driving. + +# 3 Preliminaries + +# 3.1 Diffusion Models + +The diffusion process continuously perturbs the unknown data distribution $p_{data}$ with a perturbation kernel and generates intermediate data with a given diffusion time $T$ . Denote the distribution of the time-dependent intermediate nosiy data $\boldsymbol{x}_t$ as $p_t(\boldsymbol{x}_t)$ , $t \in [0, T]$ , where $p_0(\boldsymbol{x}_0) = p_{data}$ is the clean data distribution and $\boldsymbol{x}_0$ is the clean data. The series of intermediate data $\boldsymbol{x}_t$ are generated through the Stochastic Differential Equation (SDE) [39]: + +$$ +\mathrm {d} \boldsymbol {x} _ {t} = f (\boldsymbol {x} _ {t}, t) \mathrm {d} t + g (t) \mathrm {d} \boldsymbol {w}, \boldsymbol {x} _ {0} \sim p _ {\text {d a t a}} = p _ {0} (\boldsymbol {x} _ {0}), \tag {1} +$$ + +where $f(\cdot, t)$ is the drift coefficient, $g(t)$ is the diffusion coefficient, and $\boldsymbol{w}$ is the Wiener process. We can recover $p_{data}$ from reverse-time SDE + +$$ +\mathrm {d} \boldsymbol {x} _ {t} = \left[ f (\boldsymbol {x} _ {t}, t) - g (t) ^ {2} \nabla_ {\boldsymbol {x} _ {t}} \log p _ {t} (\boldsymbol {x} _ {t}) \right] \mathrm {d} t + g (t) \mathrm {d} \bar {\boldsymbol {w}}, \boldsymbol {x} _ {T} \sim p _ {T} (\boldsymbol {x} _ {T}), \tag {2} +$$ + +where $\bar{\boldsymbol{w}}$ is another Wiener process, dt is negative timestep. + +In order to solve Eq. (2), we first need to approximate $p_T(\pmb{x}_T)$ . In previous works, $p_T(\pmb{x}_T)$ is typically approximated by some prior distributions $p_{prior}$ which contain no information of $p_{data}$ . In this paper, we adopt the setting of Variance Preserving (VP) SDE [12, 39]. In VP-SDE, $p_T(\pmb{x}_T) \approx p_{prior}$ when $T \to \infty$ . The perturbation kernels are in the form of $p_t(\pmb{x}_t|\pmb{x}_0) = \mathcal{N}(\pmb{x}_t; \sqrt{\bar{\alpha}_t}\pmb{x}_0, (1 - \bar{\alpha}_t)\pmb{\Sigma}_p)$ where scalar $\bar{\alpha}_t$ is diffusion schedule parameter, $|\pmb{\Sigma}_p| = 1$ . A common choice for $\pmb{\Sigma}_p$ is the identity matrix $\mathbf{I}$ [12, 39]. Second, we need to approximate $\nabla_{\pmb{x}_t}\log p_t(\pmb{x}_t)$ for all $t \in (0,T]$ . Some works solve score-matching problem [14, 38] and learn a score function $s_\theta(s_t,t)$ to approximate $\nabla_{\pmb{x}_t}\log p_t(\pmb{x}_t)$ . In this paper, we follow the practice in DDPM [12] to learn the noise $\epsilon$ using a network $\epsilon_\theta(\pmb{x}_t,t)$ : + +$$ +\underset {\theta} {\arg \min } \mathbb {E} _ {\boldsymbol {x} _ {0} \sim p _ {d a t a}, \boldsymbol {\epsilon} \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {\Sigma} _ {p})} \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {t}, t) \| _ {2} ^ {2} \tag {3} +$$ + +where $\pmb{x}_t = \sqrt{\bar{\alpha}_t}\pmb{x}_0 + \sqrt{1 - \bar{\alpha}_t}\pmb{\epsilon},\pmb{\epsilon}_{\theta}(\pmb{x}_t,t) = -\sqrt{1 - \bar{\alpha}_t}\pmb{s}_{\theta}(\pmb{x}_t,t)$ . With these two approximations, we can learn $q_{\theta}(\pmb{x}_0)$ to estimate unknown data distribution $p_0(\pmb{x}_0)$ solving Eq. (2) from $t = T$ to $t = 0$ . + +# 3.2 Guided Sampling + +In prior diffusion-based controllable generation frameworks [15, 16, 20, 28, 49], controllable generation is achieved via biasing of the score function for sampling: + +$$ +\nabla_ {\boldsymbol {x} _ {t}} \log \left[ p _ {t} (\boldsymbol {x} _ {t}) \exp (- \mathcal {C} (\boldsymbol {x} _ {t})) \right] = \nabla_ {\boldsymbol {x} _ {t}} \log p _ {t} (\boldsymbol {x} _ {t}) - \nabla_ {\boldsymbol {x} _ {t}} \mathcal {C} (\boldsymbol {x} _ {t}) \tag {4} +$$ + +where $\mathcal{C}(\cdot)$ is the guidance function. It requires estimating the guidance gradient with respect to the noisy data $\pmb{x}_t$ . Some approaches introduce an additional neural network to approximate $\mathcal{C}(\cdot)$ at different noisy levels [4,15,49]. The additional neural network imposes additional training costs and heavier computational burden at the inference stage. To this end, some works define an analytical guidance function $\mathcal{J}(\cdot)$ on the clean data $\pmb{x}_0$ . They first estimate $\hat{\pmb{x}}_0 = f_\theta(\pmb{x}_t)$ based on $\pmb{x}_t$ with the diffusion model, then calculate $\mathcal{C}(\pmb{x}_t)$ as $\mathcal{J}(\hat{\pmb{x}}_0)$ . However, when taking the gradient of $\mathcal{J}(\hat{\pmb{x}}_0)$ with respect to $\pmb{x}_t$ , we get + +$$ +\nabla_ {\boldsymbol {x} _ {t}} \mathcal {J} (\hat {\boldsymbol {x}} _ {0}) = \nabla_ {\hat {\boldsymbol {x}} _ {0}} \mathcal {J} (\hat {\boldsymbol {x}} _ {0}) \cdot \nabla_ {\boldsymbol {x} _ {t}} f _ {\theta} (\boldsymbol {x} _ {t}). \tag {5} +$$ + +Estimating $\nabla_{\pmb{x}_t}f_\theta (\pmb{x}_t)$ requires backpropagating throughout the entire diffusion model, i.e., $f_{\theta}(\pmb {x}_t)$ . It requires heavy computing resources and GPU memory. + +# 3.3 Trajectory Prediction and Controllable Generation + +Joint trajectory prediction aims to predict the future joint trajectories $\pmb{x}_0$ for all the vehicles in the scene, conditioned on context information $\pmb{c}$ . It can be regarded as a conditional generation task where the goal is to train a generative model to approximate the distribution $p_0(\pmb{x}_0|\pmb{c})$ . For simplicity, we omit $\pmb{c}$ and represent $p_0(\pmb{x}_0|\pmb{c})$ as $p_0(\pmb{x}_0)$ . We denote $n$ as the number of vehicles in the same scene and $\pmb{x}_{0,i}$ as the future trajectory for vehicle $i$ , $i \in \{1,2,\dots,n\}$ , so $\pmb{x}_0 = [\pmb{x}_{0,1},\pmb{x}_{0,2},\dots,\pmb{x}_{0,n}]$ . Joint trajectory prediction can be very challenging. The complex interactions among vehicles, especially in highly interactive and dense traffic, result in a complicated high-dimensional $p_0(\pmb{x}_0)$ , which is difficult to accurately model with lightweight and computationally efficient models. A simplified solution is to approximately decompose the joint trajectory distribution into marginal ones, i.e., $p_0(\pmb{x}_0) \approx \prod_i^n p_0(\pmb{x}_{0,i})$ . It leads to the marginal trajectory prediction task, which has been extensively studied with mature solutions [18, 24, 50]. One drawback is that it omits the interactions among vehicles in the predicted horizon, which leads to large errors in highly interactive scenes. + +Controllable trajectory generation is closely related to trajectory prediction. In addition to generating realistic trajectory samples resembling the ground-truth $p_0(\pmb{x}_0)$ , controllable generation imposes an additional objective—the generated trajectories should comply with specified guidance cost function $\mathcal{J}(\cdot)$ . We term the former objective as realism and the latter as guidance effectiveness. The guidance cost function $\mathcal{J}(\cdot)$ can be some goal points of the vehicles, kinematic constraints, etc, which are mostly defined on the clean manifold rather than on the noisy data when it comes to diffusion-based controllable generation. + +# 4 Methodology + +# 4.1 Optimal Gaussian Diffusion + +As discussed in Sec. 3.1, diffusion models typically select a non-informative prior distribution $p_{prior}$ , such as a standard Gaussian, as the initial data distribution for the reverse diffusion process. Such a non-informative $p_{prior}$ is reasonable since $p_T(\pmb{x}_T)$ converges to it when $T \to \infty$ . However, it also means that a large $T$ is required at inference time to yield good performance, which undermines computational efficiency and limits its wide real-time applications in autonomous driving. In this section, we aim to investigate a practical solution to tackle this challenge for joint trajectory prediction. Given the inherent limitation imposed by a non-informative prior, the key question we look into is: can we instead adopt an informative prior so that we can obtain the same level of performance with much smaller reverse diffusion steps? + +First, we still consider a Gaussian prior, but with learnable parameters, i.e., we parameterize $p_{prior}$ as $q_{\phi}(T) = \mathcal{N}(\pmb{\mu}, \pmb{\Sigma})$ , where $\pmb{\mu}$ and $\pmb{\Sigma}$ are learnable. We aim to optimize $\pmb{\mu}$ and $\pmb{\Sigma}$ to enhance the generation performance for small $T$ . We got inspiration from [5], where an upper bound of the Kullback-Leibler divergence of the clean data distribution $p_0(\pmb{x}_0)$ and learned distribution $q_{\theta}(\pmb{x}_0)$ was derived as a function of the diffusion time $T$ : + +$$ +\operatorname {K L} \left[ p _ {0} \left(\boldsymbol {x} _ {0}\right) \mid \mid q _ {\theta} \left(\boldsymbol {x} _ {0}\right) \right] \leq \mathcal {G} \left(\boldsymbol {x} _ {\theta}, T\right) + \operatorname {K L} \left[ p _ {T} \left(\boldsymbol {x} _ {T}\right) \mid \mid p _ {\text {p r i o r}} \right] \tag {6} +$$ + +where $\mathcal{G}(\pmb{x}_{\theta},T)$ is the positive accumulated error between $\nabla_{\pmb{x}}\log p_t(\pmb{x}_t)$ and $\pmb{s}_{\theta}(\pmb{x}_t,t)$ from 0 to $T$ [5]. Note that $\mathcal{G}(\pmb{x}_{\theta},T)$ is an accumulated error so $\mathcal{G}(\pmb{x}_{\theta},T_1)\leq \mathcal{G}(\pmb{x}_{\theta},T_2), T_1\leq T_2$ . If we can achieve lower KL $[p_{T_1}(\pmb{x}_{T_1})||q_\phi (T_1)]$ , then a tighter upper bound can be obtained. This opens up the possibility to achieve better performance with less diffusion time. Thus, we propose to optimize the prior distribution by minimizing KL $[p_T(\pmb{x}_T)||q_\phi (T)]$ . As shown in Proposition 1 (See proof in Appendix A), it turns out that the optimal $\mu$ and $\Sigma$ can be expressed analytically as functions of the ground-truth data statistics. In addition, we find that we can further minimize the target KL divergence if we set a learnable $\pmb{\Sigma}_{p}$ in the perturbation kernel $p_t(\pmb {x}_t|\pmb {x}_0) = \mathcal{N}(\pmb {x}_t;\sqrt{\bar{\alpha}_t}\pmb {x}_0,(1 - \bar{\alpha}_t)\pmb {\Sigma}_p)$ , whose optimal value can also be expressed as a function of the data statistics. + +Proposition 1. Denote $\pmb{\mu}_d$ and $\pmb{\Sigma}_d$ as the mean and variance of $p_{data}$ . Denote $\pmb{\Sigma}^{*}(i,j)$ and $\pmb{\Sigma}_{p}^{*}(i,j)$ as the element at ith row and jth column of matrix $\pmb{\Sigma}^*$ and $\pmb{\Sigma}_{p}^{*}$ . The optimal solution to $\min KL[p_T(\pmb{x}_T)||q_\phi (\pmb{x}_T,T)]$ is + +$$ +\left\{ \begin{array}{l} \boldsymbol {\mu} ^ {*} \approx \sqrt {\bar {\alpha} _ {T}} \boldsymbol {\mu} _ {d} \\ \boldsymbol {\Sigma} _ {p} ^ {*} (i, j) \approx \frac {1}{| \boldsymbol {\Sigma} _ {d} |} \boldsymbol {\Sigma} _ {d} (i, j) \\ \boldsymbol {\Sigma} ^ {*} (i, j) \approx \bar {\alpha} _ {T} \boldsymbol {\Sigma} _ {d} + (1 - \bar {\alpha} _ {T}) ^ {2} \boldsymbol {\Sigma} _ {p} ^ {*} = (\bar {\alpha} _ {T} ^ {2} + \frac {(1 - \bar {\alpha} _ {T}) ^ {2}}{| \boldsymbol {\Sigma} _ {d} |}) \boldsymbol {\Sigma} _ {d} (i, j) \end{array} \right. \tag {7} +$$ + +Thus, if we are able to estimate the mean and variance of the ground-truth data distribution, we can then analytically determine the optimal prior distribution $\mathcal{N}(\pmb{\mu}^*,\pmb{\Sigma}^*)$ and the optimal perturbation kernel variance $\pmb{\Sigma}_p^*$ at any noise + +level $T \in [0, \infty)$ . It leads to a crucial advantage against prior efforts on prior learning [48], where a target noise level has to be determined at training time in order to train an additional neural network to represent the prior distribution at the pre-selected noise level. In contrast, our method enables flexible tuning of the number of diffusion steps at inference time, without additional training costs. Since the learnable $q_{\phi}(T)$ and perturbation kernel are both Gaussian, we refer to our proposed diffusion framework as Optimal Gaussian Diffusion (OGD). + +For joint trajectory prediction and generation, we need to estimate the mean and variance of the joint trajectory distribution $p_0(\pmb{x}_0)$ . It is not straightforward as only a limited number of trajectory samples exist in the dataset under the same context. Considering there exists mature and accurate marginal trajectory prediction models [18, 24, 50], we can conveniently extract statistics of the marginal trajectory distributions, i.e., $p_0(\pmb{x}_{0,i}), i \in \{1,2,\dots,n\}$ from a pretrained marginal trajectory predictor. Since Proposition 1 provides element-wise optimal value, we can easily get: + +Corollary 1. Denote $\pmb{\mu}_d(\pmb{x}_{0,i})$ and $\pmb{\Sigma}_d(\pmb{x}_{0,i})$ as the mean and variance of marginal distribution for vehicle $i$ and set both $\pmb{\Sigma}$ and $\pmb{\Sigma}_p$ are block-diagonal matrices where each block represents the marginal characteristics of one single vehicle. The optimal solution to $\min KL[p_T(\pmb{x}_T)||q_\phi (\pmb{x}_T,T)]$ is + +$$ +\left\{ \begin{array}{l} \boldsymbol {\mu} ^ {*} = \left[ \sqrt {\bar {\alpha} _ {T}} \boldsymbol {\mu} _ {d} (\boldsymbol {x} _ {0, 1}), \dots , \sqrt {\bar {\alpha} _ {T}} \boldsymbol {\mu} _ {d} (\boldsymbol {x} _ {0, n}) \right] ^ {T} \\ \boldsymbol {\Sigma} _ {p} ^ {*} = \frac {1}{\sum_ {i = 1} ^ {n} | \boldsymbol {\Sigma} _ {d} (\boldsymbol {x} _ {0 , i}) |} d i a g [ \boldsymbol {\Sigma} _ {d} (\boldsymbol {x} _ {0, 1}), \dots , \boldsymbol {\Sigma} _ {d} (\boldsymbol {x} _ {0, n}) ] \\ \boldsymbol {\Sigma} ^ {*} = \bar {\alpha} _ {T} d i a g [ \boldsymbol {\Sigma} _ {d} (\boldsymbol {x} _ {0, 1}), \dots , \boldsymbol {\Sigma} _ {d} (\boldsymbol {x} _ {0, n}) ] + (1 - \bar {\alpha} _ {T}) ^ {2} \boldsymbol {\Sigma} _ {p} ^ {*} \end{array} \right. \tag {8} +$$ + +Corollary 1 implies that, if we further confine $\pmb{\Sigma}$ and $\pmb{\Sigma}_p$ to be block-diagonal without covariance between the states of different vehicles, then we can determine their optimal values purely from the estimated marginal statistics $\pmb{\mu}_d(\pmb{x}_{0,i})$ and $\pmb{\Sigma}_d(\pmb{x}_{0,i})$ , $i \in \{1,2,\dots,n\}$ , which enables a practical implementation of the proposed OGD model for joint trajectory prediction and generation tasks. Specifically, for vehicle $i$ , we leverage a pre-trained marginal trajectory predictor [18,24,50], predict diverse marginal trajectory sample set $\mathcal{R}_i = \{\pmb{r}_i^l\}_{l=1}^L$ and corresponding likelihood set $\{p(\pmb{r}_i^l)\}_{l=1}^L$ , and estimate $\pmb{\mu}_d(\pmb{x}_{0,i})$ and $\pmb{\Sigma}_d(\pmb{x}_{0,i})$ . For example, $\pmb{\mu}_d(\pmb{x}_{0,i})$ can be estimated as $\frac{1}{L}\sum_{l=L}p(\pmb{r}_i^l)\pmb{r}_i^l$ . + +# 4.2 Estimated Clean Manifold Guidance with Reference + +As discussed in Sec. 3.2, the intensive computation required for guided sampling comes from the calculation of $\nabla_{\boldsymbol{x}_t}f_\theta (\boldsymbol{x}_t)$ . Previous guided sampling approaches bias the score function defined on the intermediate noisy data $\boldsymbol{x}_t$ . In this section, we aim to investigate whether we can inject the gradient directly into $\boldsymbol{x}_0$ to avoid the gradient propagation process. We first reformulate the controllable generation as a multi-objective optimization problem directly over $\boldsymbol{x}_0$ and propose an iterative algorithm to solve the formulated problem. In addition, we use reference trajectory points to create the region of interest, which helps solve the local optimal problem caused by multi-modal joint trajectory distribution and accelerate the guided sampling process. + +Estimated Clean Manifold Guidance. The objective of controllable generation of sample $\pmb{x}_0$ includes two different objectives. The most important is the negative likelihood, ensuring the sample lies in the clean manifold. The second important is the guidance cost representing the user preference on the generated sample $\pmb{x}_0$ . This multi-objective optimization problem can be represented as + +$$ +\min _ {\boldsymbol {x} _ {0}} \left[ - \log q _ {\theta} (\boldsymbol {x} _ {0}), \mathcal {J} (\boldsymbol {x} _ {0}) \right] ^ {T} \tag {9} +$$ + +Inspired by lexicographic optimization [33], we solve this multi-objective optimization problem hierarchically. The main idea is to optimize each objective in the order of importance regardless of the degradation of the other less significant objectives. We first optimize the most important objective, $-\log q_{\theta}(\pmb{x}_0)$ , to generate realistic and high-likelihood samples. The diffusion model achieves this goal effectively by reversing the diffusion process from noisy samples at specific noise level. However, exact noise level of current sample $\pmb{x}_0$ is unknown. To address this, we inject noise at level $t$ into $\pmb{x}_0$ , and then denoise it from $t$ with learned diffusion model. This approach, similar to noise injection and denoising [1,23,37], improves the desired sample quality. We iteratively repeat this process $K$ times, injecting guidance at each iteration to strengthen user preference. + +Specifically, denote $\pmb{x}_0(k)$ as the sample at iteration $k$ . We first regenerate high-likelihood sample $\hat{\pmb{x}}_0(k)$ by diffusion model: + +$$ +\hat {\boldsymbol {x}} _ {0} (k) \leftarrow \mathbb {E} \left[ q _ {\theta} \left(\boldsymbol {x} _ {0} (k) \mid \boldsymbol {x} _ {t _ {k}} (k)\right) \right] = \frac {1}{\sqrt {\bar {\alpha} _ {t _ {k}}}} \left(\boldsymbol {x} _ {t _ {k}} - \sqrt {1 - \bar {\alpha} _ {t _ {k}}} \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {x} _ {t _ {k}} (k), t _ {k}\right)\right), \tag {10} +$$ + +where $\pmb{x}_{t_k}(k) = \sqrt{\bar{\alpha}_{t_k}}\pmb{x}_0(k) + \sqrt{1 - \bar{\alpha}_{t_k}}\pmb{\epsilon}, \pmb{\epsilon} \sim \mathcal{N}(\pmb{0},\pmb{\Sigma}_p)$ . $t_k \in [0,T]$ is a tunable parameter. Then we minimize the guidance cost function with a small degradation of the most important objective $-\log q_{\theta}(\pmb{x}_0)$ , + +$$ +\boldsymbol {x} _ {0} (k - 1) \leftarrow \hat {\boldsymbol {x}} _ {0} (k) - \zeta \nabla_ {\hat {\boldsymbol {x}} _ {0} (k)} \mathcal {J} (\hat {\boldsymbol {x}} _ {0} (k)) \tag {11} +$$ + +where $\zeta$ is the step size. Small degradation of $-\log q_{\theta}(\pmb{x}_0)$ is realized by one-step gradient update and proper step size $\zeta$ . See the derivation of the optimization process and the parameter tuning in Appendix B. + +During the iterations, $\pmb{x}_0(k), \forall k \in \{0,1,\dots,K - 1\}$ are not exactly on the clean manifold. We minimize its negative log-likelihood through diffusion model, resulting in it lying on an estimated clean manifold. Thus, we call our method Estimated Clean Manifold (ECM) Guidance. + +Reference Joint Trajectories. To generate trajectories with low guidance cost, we are essentially searching low-cost trajectories within the high-likelihood region. At the same time, joint trajectory distribution is a multi-modal distribution resulting from road topologies and different decision variables, meaning the likelihood has multiple peaks. This leads to the optimal solution of Problem 9 having multiple local optimals, and each optimal is far away from the other. Guided sampling methods [15,16,28,49], including our method ECM, suffer from + +![](images/35e3bb102b3b53c1b1e68f1e89625642576cf47a0668139d544a34789a2fcb50.jpg) +Fig. 2: Two challenges with multi-peak function optimization: 1) Gradients may lead to suboptimal local optima (left); 2) There exist regions with low likelihood but high guidance cost uncertainty, leading to instability (right). Our approach can bypass the lengthy paths between peaks, search for better optima, and avoid uncertain areas. + +two challenges (See Fig. 2): 1) it can be trapped at the local optimal around the initial position; 2) it takes massive efforts to drag the sample from one peak to another and transferring from one modal to another will need to pass through the region of low-likelihood (off clean-manifold), leading to numerical instability. + +To overcome this, we generate high-likelihood reference joint trajectories, choose the best one as the initialization. Note that combinations of samples with high marginal likelihood tend to exhibit high joint likelihood. Therefore, we can utilize the marginal sample set $\mathcal{R} = \{\mathcal{R}_i\}_{i=1}^n$ obtained from pre-trained marginal models to generate the references. Specifically, for iteration $k$ , we + +construct candidate joint trajectory set $\mathcal{R} \otimes \hat{\pmb{x}}_0(k) = \{[\pmb{w}_1, \pmb{w}_2, \dots, \pmb{w}_n] | \pmb{w}_i \in \mathcal{R}_i \cup \{\hat{\pmb{x}}_{0,i}\}, i = 1, 2, \dots, n\}$ . We calculate the guidance cost of all possible combinations, $\mathcal{J}(w)$ , $w \in \mathcal{R} \otimes \hat{\pmb{x}}_0(k)$ , and choose the minimal-cost one as the reference. The guided sampling algorithm, ECM with reference joint trajectories (ECMR), is introduced in Algorithm 1. + +# 5 Experiments + +# 5.1 Experimental setup + +Dataset. We use Argoverse 2 [44], a widely used and large-scale trajectory prediction dataset, to test the effectiveness of our approaches for joint trajectory prediction and controllable generation. It has a large observation window of 5s and a long prediction horizon of 6s. + +Algorithm 1: ECMR + +
Input: J(·), {αt,βt}T-1t=0, {tk}K-1k=0
1: x0(K) ~ N(0, Σp)
2: for k = K-1,..., 1 do
3: ε ~ N(0, Σp)
4: xtk(k) = √αtkx0(k) + √1-αtkε
5: x0(k) = 1/√αtk(xtk(k) - √1-αtk)εθ(xtk(k),t))
6: x0(k) = arg min J(w), w ∈ R⊗ x0(k)
7: x0(k-1) ← x0(k) - ζ∇x0(k)J(x0(k))
8: end for
Output: x0(0)
+ +Implementation Details. We use the fixed scene context encoder of pretrained QCNet [50] to extract compact and representative context features from context information $\pmb{c}$ . Then, we utilize a cross-attention layer to update the intermediate noisy data $\pmb{x}_t$ with multiple contexts, including the history encodings of the target agent, the map encodings, the neighboring agents' encodings. Inspired by [7], we also add a cross-attention layer to update $\pmb{x}_t$ with the diverse marginal trajectory samples $\pmb{r}_i^l$ and its corresponding likelihood. In addition, we use self-attention to allow the interaction between $\pmb{x}_{t,i}$ and $\pmb{x}_{t,j}$ . Then the model predicts the noise $\pmb{\epsilon}_{\theta}(\pmb{x}_t,t)$ . According to [16], compact trajectory representation helps the diffusion model to generate high-quality trajectories efficiently. Inspired by this, we also learn a linear mapping between the 10-dimensional latent and 120-dimensional trajectories. Similar to [16], we design a rapid sample clustering algorithm so that we can generate a representative joint trajectory set. To increase the efficiency of sampling, we use DDIM [35] to accelerate the inference, and the DDIM step stride is 10. See Appendix C for details and analysis. + +# 5.2 Joint trajectory prediction + +We now evaluate OGD for joint trajectory prediction. Given $K$ joint trajectories, the evaluation metrics are 1) avgMinFDE $_K$ /avgMinADE $_K$ : the average of lowest final/average displacement error (FDE/ADE) of joint trajectory samples; 2) actorMR $_K$ : the rate of trajectory predictions that are considered to be "missed" ( $>2m$ FDE) in the lowest minFDE joint trajectory samples; 3) actorCR $_K$ : the rate of collisions across "best" (lowest avgMinFDE) joint trajectory samples; 4) avgBrierMinFDE $_K$ : calculated similarly to avgMinFDE $_K$ but scaled by the probability score of joint trajectory samples. We denote metrics with superscript "\*" as those after sample clustering (see Appendix C.3). + +As a baseline, we train a vanilla diffusion (VD) baseline that shares the same neural network architecture as OGD. Specifically, we train an Optimal Gaussian Diffusion model with diffusion time $T_{train} = 100$ (OGD), vanilla diffusion with two different diffusion times $T_{train} = 100,500$ ( $\mathrm{VD}_{100}, \mathrm{VD}_{500}$ ). Note that $T$ is denoted as the diffusion time from which the reverse diffusion process starts. First, during the inference, we change $T$ and evaluate OGD and $\mathrm{VD}_{100}$ who have the same $T_{train} = 100$ . Figure 3 shows that, with the decrease of the reverse steps, avgMinFDE128 of OGD keeps lower than $\mathrm{VD}_{100}$ and it is more stable with the change of $T$ . It is also interesting to note that the performance of OGD even becomes better in the early stage. We hypothesize that it is because that $\mathrm{KL}[p_T(\boldsymbol{x}_T)||p_{prior}]$ of OGD is small and the accumulated error of $\mathcal{G}(\boldsymbol{x}_\theta, T)$ is sufficiently reduced by a small $T$ . It also shows an advantage of OGD, which is that it is easy for OGD to tune a suitable $T$ for reverse diffusion without training an additional model for every $T$ [48]. Table 1 shows that OGD outperforms $\mathrm{VD}_{500}$ with only 40 diffusion steps. + +We also compare our OGD with the other state-of-the-art methods on the Argoverse multi-world leaderboard; our approach OGD ranks $4^{\mathrm{th}}$ on the leaderboard ranked by avgBrierMinFDE $_K$ , which demonstrates the effectiveness of our + +![](images/b007635a1852e6cad01ff697f59805204a1361e9b002412795d7953320f36e96.jpg) +Fig.3: Evaluation of Optimal Gaussian Diffusion and vanilla diffusion over reverse steps $T$ . + +Table 1: Evaluation on joint trajectory prediction task. For each metric, the best result is in **bold** and the second best result is **underlined**. $T = 70$ is the best $T$ from Fig. 3. $T = 40$ is the minimal diffusion time when OGD outperforms $\mathrm{VD}_{500}$ on all metrics. + +
ModelTavgMinFDE*avgMinADE*avgMinFDE128avgMinADE128
VD1001000.621.380.490.97
VD5005000.611.360.480.91
OGD1000.601.320.430.81
OGD700.591.310.420.75
OGD400.611.340.470.77
+ +OGD framework. Note that we only list the entries with publications or technical reports in Tab. 2. Please refer to the official website for the full leaderboard. + +# 5.3 Controllable Generation + +Tasks. Future behaviors of vehicles can be effectively can be effectively represented by a set of goal points [9, 47] such as acceleration, braking, and right or left turn. Generating such diverse modes of joint trajectories is a typical motivation for using controllable generation. Thus, we study the controllable generation task where the goal is to reach diverse goal points at a specific time to fully test our guided sampling method. Considering that goal points should lie in realistic routes, and the trajectories to reach such goal points should also lie in such routes, we generate such target goal points in some realistic routes $\mathrm{Route}(\tau_g)$ . The guidance cost function can be expressed as + +$$ +\mathcal {J} \left(\boldsymbol {x} _ {0}\right) = \frac {1}{n} \left\| \operatorname {P o s i t i o n} \left(\boldsymbol {x} _ {0}, \tau_ {d}\right) - \operatorname {R o u t e} \left(\tau_ {g}\right) \right\| _ {2} ^ {2}, \tag {12} +$$ + +where $\mathrm{Position}(\pmb{x}_0, \tau_d)$ is the positions of $\pmb{x}_0$ at time $\tau_d$ . We choose ground-truth trajectories and a random combination set of diverse marginal samples, i.e., $\mathcal{U} = \{[\pmb{u}_1, \pmb{u}_2, \dots, \pmb{u}_n] | \pmb{u}_i \in \mathcal{R}_i, i = 1, 2, \dots, n\}$ as the realistic routes. We denote + +Table 2: Quantitative results on the Argoverse 2 Multi-world Forecasting leaderboard. For each metric, the best result is in bold and the second best result is underlined. + +
ModelavgMinFDE*6avgMinFDE1*actorMR6avgMinADE*6avgMinADE1*avgBrierMinFDE*6actorCR6
QCXet [51]1.022.290.130.500.941.650.01
Gnet [6]1.463.050.190.691.232.120.01
Forecast-MAE [2]1.553.330.190.691.302.240.01
FJMP [30]1.894.000.230.811.522.590.01
OGD (Ours)1.312.710.170.601.081.950.01
+ +![](images/cd2c4fc275eb274906f664ebf2d8d0a023fbc099e1021823ade9e6ce6ac94c8f.jpg) +Fig. 4: Evaluation on controllable generation: route set $\mathbf{U}$ and Deceleration. Magenta diamonds represent goal points. In the first (second) row, goal points are set at the fork lane (right lane). NNM [49] and SF [16, 28] struggle to drag samples from one modal to another. Our methods can achieve better guidance effectiveness and realism. + +![](images/0d617d43fdb73080ada75572219dd3e8fb80437fc7d31d06249af1e3ab6d25d0.jpg) + +![](images/f2826df3fa6be27717cf866bb6e0d50a84b0dd9747e514d6b0866fb580ee98ac.jpg) + +![](images/f5820e3377fcdc4969a8b6dc291ffff9c9a6ee1effda1a9146dd7bf9bbe84d4e.jpg) + +the former as GT set and the latter as $\mathbf{U}$ set. The underlying assumption is that diverse samples from a good marginal trajectory predictor are realistic. We design different velocity settings to cover the diverse controllable generation tasks in autonomous driving: first is Normal Speed (N), $\tau_{d} = \tau_{g} = 6s$ ; second is Acceleration (A), $\tau_{d} = 5s < \tau_{g} = 6s$ ; third is Deceleration (D), $\tau_{d} = 6s < \tau_{g} = 5s$ . + +Metrics. We use the following metrics to evaluate controllable generation performance: Joint Route Deviation Error (JRDE), which measures the displacements to the routes to evaluate the realism, and Joint Final Displacement Error (JFDE), which evaluates the guidance effectiveness. We also evaluate from the "min" and "mean" perspectives: The "min" metric considers the best sample's performance, while the "mean" metric assesses the ratio of valid samples. + +Baselines. Guided sampling in controllable generation is mainly divided into two approaches: the first is to directly calculate $\nabla_{\boldsymbol{x}_t}\mathcal{J}(\boldsymbol{x}_t)$ [49]; the second is to calculate $\nabla_{\boldsymbol{x}_t}\mathcal{J}(\hat{\boldsymbol{x}}_0)$ [16, 28]. We denote the former as Next Noisy Mean Guidance (NNM) and the latter as Score Function Guidance (SF). For a fair comparison, we use one guidance step followed by one DDIM step. We also tune the gradient step size for different guided sampling with Optimal Gaussian Diffusion and vanilla diffusion and report the results with the optimal step size. See Appendix D for the details of baseline derivation and step size tuning. We evaluate the following experiments with 128 joint trajectory samples. + +Evaluation. First, we evaluate the performance and efficiency with our diffusion model (OGD) and guided sampling method (ECM and ECMR) in Tab. 3, which demonstrates our methods can generate more realistic and effective samples with 5 times less DDIM steps. In addition, with reference joint trajectories, ECM significantly improves 'mean' metrics, indicating it addresses the issues discussed in Sec. 4.2 to a certain extent. Second, we compare solely on different guided sampling methods using the same OGD model shown in Fig. 5. Our ECM achieves better performance both in guidance effectiveness and realism. In Fig. 4, our ECM can generate trajectories that reach goal points more closely than NNM and SF. And with reference joint trajectory, ECMR can easily move samples from one modal to another. Third, we evaluate the average inference time and average GPU memory usage in Tab. 4. Our methods can generate + +Table 3: Evaluation on controllable generation: route set U and Deceleration. For each metric, the best result is in bold and the second best result is underlined. + +
ModelSamplingDDIM StepsGuidance minJFDEEffectiveness meanJFDERealism minJRDEmeanJRDE
VD500No Guid501.9615.2290.1650.492
VD500NNM [49]500.7782.9130.1300.309
VD500SF [16, 28]500.5382.3390.1580.500
OGDNo Guid101.7725.1720.1380.469
OGDECM (Ours)100.0720.2370.1280.236
OGDECMR (Ours)100.0530.1460.1100.154
+ +![](images/37ff5e2a75e2db6356b4fabdda5bd2949af8aafc47e61308f7aa0cd4c341a16f.jpg) +Fig. 5: Evaluation of different guided sampling methods on various tasks. See Appendix E for other metrics. + +Table 4: We evaluate the average inference time per step and GPU incremental memory on $\mathbf{U} +$ Deceleration. We test on single RTX A6000, batch size is 16 and number of samples is 128. + +
SamplingminJFDEminJRDEtime(ms)memory (GB)
NNM [49]0.7240.1191133.21
SF [16,28]0.1550.1262477.96
ECM(Ours)0.0750.1161113.21
ECMR(Ours)0.0530.1101163.22
+ +realistic trajectories satisfying guidance quite well with low inference time and GPU memory usage. More results on controllable generation can be found in Appendix E. + +# 6 Conclusion + +In this work, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance to significantly improve the computational efficiency and performance of diffusion models in autonomous driving. These methodologies enable a substantial reduction in inference steps and computational demands while ensuring enhanced joint trajectory prediction and controllable generation capabilities. Our approaches and experimental results underscore the potential of diffusion models for real-time applications in dynamic environments, marking a pivotal advancement in the deployment of diffusion models for autonomous driving. One limitation of the current implementation is that the performance is affected by the accuracy of marginal trajectory predictors. Enhancements could be achieved with superior marginal models or by directly learning joint predictions' mean and variance, areas for future work. + +# Acknowledgement + +This work was supported by Berkeley DeepDrive. + +# References + +1. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18208-18218 (2022) +2. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023) +3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022) +4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +5. Franzese, G., Rossi, S., Yang, L., Finamore, A., Rossi, D., Filippone, M., Michiardi, P.: How much is enough? a study on diffusion times in score-based generative models. Entropy 25(4), 633 (2023) +6. Gao, X., Jia, X., Li, Y., Xiong, H.: Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolutional recurrent networks. IEEE Robotics and Automation Letters 8(5), 2946-2953 (2023) +7. Gilles, T., Sabatini, S., Tsishkou, D., Stanciulescu, B., Moutarde, F.: Thomas: Trajectory heatmap output with learned multi-agent sampling. arXiv preprint arXiv:2110.06607 (2021) +8. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural information processing systems. pp. 2672-2680 (2014) +9. Gu, J., Sun, C., Zhao, H.: Densetnt: End-to-end trajectory prediction from dense goal sets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15303-15312 (2021) +0. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022) +1. Guo, Z., Gao, X., Zhou, J., Cai, X., Shi, B.: Scenedm: Scene-level multiagent trajectory generation with consistent diffusion models. arXiv preprint arXiv:2311.15736 (2023) +2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +3. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022) +4. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005) +5. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022) + +16. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motion-diffuser: Controllable multi-agent motion prediction using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9644-9653 (2023) +17. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. Advances in neural information processing systems 34, 21696-21707 (2021) +18. Lan, Z., Jiang, Y., Mu, Y., Chen, C., Li, S.E., Zhao, H., Li, K.: Sept: Towards efficient scene representation learning for motion prediction. arXiv preprint arXiv:2309.15289 (2023) +19. Laumanns, M., Thiele, L., Zitzler, E.: An efficient, adaptive parameter variation scheme for metaheuristics based on the epsilon-constraint method. European Journal of Operational Research 169(3), 932-942 (2006) +20. Lin, H., Wang, Y., Huo, M., Peng, C., Liu, Z., Tomizuka, M.: Joint pedestrian trajectory prediction through posterior sampling. arXiv preprint arXiv:2404.00237 (2024) +21. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022) +22. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021) +23. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022), https://arxiv.org/abs/2108.01073 +24. Nayakanti, N., Al-Rfou, R., Zhou, A., Goel, K., Refaat, K.S., Sapp, B.: Wayformer: Motion forecasting via simple & efficient attention networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2980-2987. IEEE (2023) +25. Ngiam, J., Caine, B., Vasudevan, V., Zhang, Z., Chiang, H.T.L., Ling, J., Roelofs, R., Bewley, A., Liu, C., Venugopal, A., et al.: Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417 (2021) +26. Peng, C., Wang, G., Lo, X.W., Wu, X., Xu, C., Tomizuka, M., Zhan, W., Wang, H.: Delflow: Dense efficient learning of scene flow for large-scale point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16901-16910 (2023) +27. Peng, C., Xu, C., Wang, Y., Ding, M., Yang, H., Tomizuka, M., Keutzer, K., Pavone, M., Zhan, W.: Q-slam: Quadric representations for monocular slam. arXiv preprint arXiv:2403.08125 (2024) +28. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023) +29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +30. Rowe, L., Ethier, M., Dykhne, E.H., Czarnecki, K.: Fjmp: Factorized joint multiagent motion prediction over learned directed acyclic interaction graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13745-13755 (2023) + +31. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512 (2022) +32. San-Roman, R., Nachmani, E., Wolf, L.: Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600 (2021) +33. Sherali, H.D., Soyster, A.L.: Preemptive and nonpreemptive multi-objective programming: Relationship and counterexamples. Journal of Optimization Theory and Applications 39, 173-186 (1983) +34. Shi, S., Jiang, L., Dai, D., Schiele, B.: Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. arXiv preprint arXiv:2306.17770 (2023) +35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +36. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: International Conference on Learning Representations (2022) +37. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. arXiv preprint arXiv:2303.01469 (2023) +38. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019) +39. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +40. Sun, Q., Huang, X., Gu, J., Williams, B.C., Zhao, H.: M2i: From factored marginal trajectory prediction to interactive prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6543-6552 (2022) +41. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021) +42. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022) +43. Watson, D., Chan, W., Ho, J., Norouzi, M.: Learning fast samplers for diffusion models by differentiating through sample quality. arXiv preprint arXiv:2202.05830 (2022) +44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023) +45. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023) +46. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902 (2022) +47. Zhao, H., Gao, J., Lan, T., Sun, C., Sapp, B., Varadarajan, B., Shen, Y., Shen, Y., Chai, Y., Schmid, C., et al.: Tnt: Target-driven trajectory prediction. In: Conference on Robot Learning. pp. 895-904. PMLR (2021) +48. Zheng, H., He, P., Chen, W., Zhou, M.: Truncated diffusion probabilistic models. arXiv preprint arXiv:2202.09671 (2022) + +49. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023) +50. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023) +51. Zhou, Z., Wen, Z., Wang, J., Li, Y.H., Huang, Y.K.: Qcnext: A next-generation framework for joint multi-agent trajectory prediction (2023) \ No newline at end of file diff --git a/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/images.zip b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25ec1f97ab3a5f6d762ff5e8a9dc3faf41482fae --- /dev/null +++ b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e440b63d1a38396ebaccaae806dd39db03687d337e84782ca4859cbb55ddae00 +size 359494 diff --git a/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/layout.json b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5a44408d1055055de134f07b8c8c1881a67e7017 --- /dev/null +++ b/2024/Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation/layout.json @@ -0,0 +1,11647 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 478, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 478, + 147 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 478, + 147 + ], + "type": "text", + "content": "Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": "Yixiao Wang" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Chen Tang" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Lingfeng Sun" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Simone Rossi" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Yichen Xie" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Chensheng Peng" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Thomas Hannagan" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Stefano Sabatini" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Nicola Poerio" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", Masayoshi Tomizuka" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "text", + "content": ", and Wei Zhan" + }, + { + "bbox": [ + 143, + 167, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 220, + 213, + 392, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 213, + 392, + 225 + ], + "spans": [ + { + "bbox": [ + 220, + 213, + 392, + 225 + ], + "type": "text", + "content": "1 University of California, Berkeley, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 224, + 392, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 224, + 392, + 236 + ], + "spans": [ + { + "bbox": [ + 220, + 224, + 392, + 236 + ], + "type": "text", + "content": "2 The University of Texas at Austin, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 267, + 236, + 347, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 236, + 347, + 247 + ], + "spans": [ + { + "bbox": [ + 267, + 236, + 347, + 247 + ], + "type": "text", + "content": "3 Stellantis, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 271, + 247, + 343, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 247, + 343, + 258 + ], + "spans": [ + { + "bbox": [ + 271, + 247, + 343, + 258 + ], + "type": "text", + "content": "4 Stellantis, Italy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 259, + 386, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 259, + 386, + 269 + ], + "spans": [ + { + "bbox": [ + 228, + 259, + 386, + 269 + ], + "type": "text", + "content": "{yixiao_wang, wzhan}@berkeley.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 294, + 455, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 294, + 455, + 461 + ], + "spans": [ + { + "bbox": [ + 159, + 294, + 455, + 461 + ], + "type": "text", + "content": "Abstract. Diffusion models are promising for joint trajectory prediction and controllable generation in autonomous driving, but they face challenges of inefficient inference steps and high computational demands. To tackle these challenges, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance. OGD optimizes the prior distribution for a small diffusion time " + }, + { + "bbox": [ + 159, + 294, + 455, + 461 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 159, + 294, + 455, + 461 + ], + "type": "text", + "content": " and starts the reverse diffusion process from it. ECM directly injects guidance gradients to the estimated clean manifold, eliminating extensive gradient backpropagation throughout the network. Our methodology streamlines the generative process, enabling practical applications with reduced computational overhead. Experimental validation on the large-scale Argoverse 2 dataset demonstrates our approach's superior performance, offering a viable solution for computationally efficient, high-quality joint trajectory prediction and controllable generation for autonomous driving. Our project webpage is at https://yixiaowang7.github.io/OptTrajDiff_Page/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "spans": [ + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "text", + "content": "Keywords: Diffusion model " + }, + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "text", + "content": " Autonomous driving " + }, + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "text", + "content": " Trajectory Prediction " + }, + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 469, + 453, + 492 + ], + "type": "text", + "content": " Controllable Trajectory Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 510, + 230, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 230, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 230, + 522 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The diffusion model is a class of generative models capable of representing high-dimensional data distributions. In particular, it has demonstrated strong performance in trajectory prediction and generation for autonomous driving [10, 11, 16, 28, 49]. In contrast to traditional trajectory prediction [18, 50] and generative models [41, 45], the unique advantage of diffusion models lies in their ability to deform the generated trajectory distribution to comply with additional requirements at inference stage via gradient-based guided sampling. Notably, it is achieved without extra model training costs. This ability to perform controllable trajectory generation enables various useful applications, such as enforcing additional realism constraints on predicted trajectories, generating directed and user-specified simulation scenarios." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "However, computational efficiency is a crucial bottleneck hindering the practical application of diffusion models in autonomous driving. Real-time inference is essential for trajectory prediction, as it provides timely forecasts of surrounding agents' behavior, enabling safe and efficient planning in dynamic traffic scenarios. The high demand for inference speed, coupled with limited onboard computational resources, makes it infeasible to deploy diffusion models for onboard trajectory prediction. While simulations do not occur onboard, lightweight models are still preferred to streamline the closed-loop training and evaluation pipelines. The heavy computational cost is mainly attributed to the following two aspects:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "content": "Computationally Expensive Reverse Diffusion. At the inference stage, the diffusion model samples from standard Gaussian distribution and gradually denoises the sample through dynamics described by a Stochastic Differential Equation (SDE) [39], aiming to eventually obtain a sample as if drawn from a target data distribution. The target data distribution can be significantly different from a standard Gaussian distribution, necessitating a large number of denoising steps to yield good performance. Prior works have attempted to reduce the reverse diffusion steps through adaptive noise schedule [17, 32], fast samplers [21,35,43,46], or distillation [31,37]. However, the fixed standard Gaussian prior poses a challenge in accelerating the reverse diffusion process without violating the SDE, which can inevitably compromise the generation quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 362, + 483, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 483, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 483, + 483 + ], + "type": "text", + "content": "Computationally Expensive Guided Sampling. Controllable generation is typically achieved by guiding the denoising process with the gradient of a guidance cost function. The guidance cost function encodes the desired characteristics of the generated data, which is typically defined on the clean data manifold in trajectory prediction and controllable generation problems. However, guided sampling intends to inject the gradient of the guidance cost function into the series of noisy data at intermediate diffusion steps. It requires a forward pass to estimate the clean data first and then back-propagating throughout the entire network to estimate the gradient with respect to the intermediate noisy data [16, 20, 28], which is very computationally intensive." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 483, + 668 + ], + "type": "text", + "content": "Targeted at these challenges, we take a step further to improve the computational efficiency of diffusion models while maintaining their performance for joint trajectory prediction and controllable generation tasks. Specifically, we propose two novel solutions for efficient reverse diffusion and guided sampling respectively. First, we present Optimal Gaussian Diffusion (OGD) to accelerate the reverse diffusion process. At the inference stage, instead of a standard Gaussian distribution far away from the desired data distribution, OGD starts from an optimal Gaussian distribution, which minimizes the distance to intermediate data distribution at a specific noise level, cutting down the diffusion steps before that. We show that we can analytically estimate such an optimal Gaussian distribution and also, an optimal perturbation kernel distribution, at any noise level from the statistics of the data distribution. It allows flexible tuning of the diffusion steps at the inference stage, without the need to train any additional models [48]. We further derive a practical implementation of OGD for joint trajectory prediction and generation, where the optimal Gaussian prior is computed" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 307, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 307, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 307, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 307, + 213 + ], + "type": "image", + "image_path": "084ebe88e07559bafe43edf81a5887443d8208b771aa65c76fb05067f5ddf346.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 215, + 290, + 225 + ], + "lines": [ + { + "bbox": [ + 150, + 215, + 290, + 225 + ], + "spans": [ + { + "bbox": [ + 150, + 215, + 290, + 225 + ], + "type": "text", + "content": "(a) Optimal Gaussian Diffusion (OGD)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 118, + 417, + 213 + ], + "blocks": [ + { + "bbox": [ + 307, + 118, + 417, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 118, + 417, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 417, + 213 + ], + "type": "image", + "image_path": "478d8a8314156abe4e50cacc8cd7679f2ec49c114e6b21600fa39eee88cb2aec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "type": "text", + "content": "Fig. 1: Overview of Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold Guidance (ECM). (a) OGD uses the mean and variance of the data distribution to calculate the optimal prior distribution at a small " + }, + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 233, + 482, + 289 + ], + "type": "text", + "content": ". It can largely reduce the diffusion time compared with vanilla diffusion. (b) ECM directly injects the gradient of guidance into the clean data manifold to mitigate computational complexity." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 417, + 121, + 480, + 213 + ], + "blocks": [ + { + "bbox": [ + 308, + 215, + 478, + 225 + ], + "lines": [ + { + "bbox": [ + 308, + 215, + 478, + 225 + ], + "spans": [ + { + "bbox": [ + 308, + 215, + 478, + 225 + ], + "type": "text", + "content": "(b) Estimated Clean Manifold Guidance (ECM)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 121, + 480, + 213 + ], + "lines": [ + { + "bbox": [ + 417, + 121, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 417, + 121, + 480, + 213 + ], + "type": "image", + "image_path": "3e5c50c1457d8655f9228e597d7e61812dc7b9f955f88e6ac8ae4d56cd786193.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 316, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 339 + ], + "type": "text", + "content": "using the mean and variance of marginal trajectory distributions estimated with a pre-trained marginal trajectory prediction [18, 24, 50] model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 341, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 520 + ], + "type": "text", + "content": "Second, we propose Estimated Clean Manifold Guidance (ECM) to accelerate guided sampling for controllable generation. To save the computational cost due to estimating the guided gradient on the noisy data, we aim to directly inject the gradient into the clean data manifold without lengthy backpropagation. ECM is motivated by the insight that guided sampling can be regarded as a multi-objective optimization problem on the clean data manifold: The first objective is to maximize the likelihood of the samples on the estimated real data distribution; the second objective is to achieve low guidance cost. ECM hierarchically solves this multi-objective problem without backpropagation throughout the entire diffusion model. We show that it leads to faster inference time and much better performance than existing approaches. Also, to tackle the challenges imposed by the multi-modal nature of vehicle interactions, we propose to warm-start the multi-objective optimization problem with reference joint trajectory points estimated using a marginal trajectory predictor. We refer to the complete algorithm as Estimated Clean Manifold with Reference Joint Trajectory (ECMR)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "To evaluate the proposed OGD and ECM methods on real-world tasks, we implement the OGD model leveraging a pre-trained marginal prediction model, QCNet [50], and conduct extensive experiments on the Argoverse 2 dataset. We show that OGD can achieve better joint trajectory prediction performance than vanilla diffusion with significantly fewer diffusion steps—it only takes " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/12" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of the diffusion steps used by the vanilla diffusion model. OGD also achieves outstanding prediction accuracy compared to non-diffusion joint prediction models. In addition, ECMR, coupled with OGD, can generate samples with significantly lower guidance costs with the same level of realism scores compared to conducting controllable generation on the vanilla diffusion model using existing guided sampling approaches used in autonomous driving [16, 28], but using around " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/5" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of their inference step." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 90, + 449, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 90, + 449, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 90, + 449, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 365 + ], + "type": "text", + "content": "Diffusion models. Diffusion models have demonstrated their ability to produce high-quality, diverse samples in a variety of applications, such as image, video, and 3D generation [13, 22, 29]. Recently, diffusion models have been applied to trajectory prediction in autonomous driving. It shows great performance on representing future trajectory distribution [10, 16]. However, diffusion models need to run lengthy reverse diffusion processes to generate high-quality samples [5, 10]. This makes it hard to apply diffusion models to trajectory prediction in autonomous driving since it requires in-time prediction for the downstream planning module to promptly respond to dynamically changing traffic scenarios. Previous works [17, 21, 31, 32, 35, 37, 43, 46] mostly focused on mitigating this issue by investigating how to solve reverse SDE in a faster manner. In addition, similar to ours, some works sought alternative initialization of the reverse diffusion process to achieve faster inference. For example, [3] initialized the reverse diffusion process with samples generated by another generative backbone network. However, the backbone network is not deliberately trained for the reverse diffusion process. [48] proposed to learn the initial diffusion with a generative adversarial network (GAN) [8]. However, it requires training a large additional model with a complex training procedure. Also, it requires specifying the reverse diffusion time as a hyperparameter prior to GAN training, which is hard to tune." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 366, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 594 + ], + "type": "text", + "content": "Guided sampling. Diffusion models have been successfully used to tackle controllable tasks through guided sampling, such as image inpainting [36] and motion planning [15]. A notable feature of diffusion models with guided sampling is their ability to condition the generation process on the user's preference, which was not available during the training phase. In the driving domain, recent works used guided sampling to generate controllable traffic [16, 49], and controllable pedestrian animation [28]. Our work belongs to the prior category, where user-specified guidance cost is used to guide generation in the trajectory space. In this case, the guidance cost function encodes certain desired properties of the generated trajectories. Thus, it is normally defined on the realistic trajectory samples, which are on the so-called clean manifold, instead of the noisy manifold containing the intermediate noisy data. Some works attempted to learn the guidance cost function on the noisy manifold [15]. Otherwise, it will lead to numerical instability when evaluating guidance cost at intermediate noisy data [28, 49]. To avoid the additional computational costs introduced by a learned guidance cost function, [16, 28] proposed to project the intermediate noisy data into the clean manifold through the diffusion model, and evaluate the guidance cost on the projected point. This approach requires back-propagating throughout the entire diffusion model, which is also computationally intensive." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Trajectory prediction. In autonomous driving [26, 27], it is vital to precisely forecast how other participants in traffic will move in the future so that ego vehicle can plan a safe and efficient trajectory to execute in the future. Marginal trajectory prediction is used to predict the trajectory distribution for single vehicle and recent works involve kinematic constraints of the vehicles, restrictions of complex topology of roads, and interaction from the surrounding" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "vehicles [18,24,50]. Recently, joint trajectory prediction has attracted attention from several researchers. It consists in predicting the joint future trajectories for all agents so that these trajectories are consistent with one another [25], an aspect which marginal trajectory prediction does not consider. Scene-Transformer [25] uses a fixed set of learnable scene embeddings to generate corresponding joint trajectories for all the vehicles in the given scene. Models like M2I [40], and FJMP [30] adopt a conditional approach, predicting the motions of other agents based on the movements of controlled agents. Diffusion model [16] has also been used to predict the joint trajectory. However, joint trajectory prediction is still a challenging problem since the complexity increases exponentially with the number of vehicles in the scene. The efficiency problem becomes more serious when using diffusion model to predict the joint trajectory distribution [16], and largely limits the application of diffusion models in autonomous driving." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 286, + 233, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 233, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 233, + 298 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 308, + 246, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 246, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 246, + 319 + ], + "type": "text", + "content": "3.1 Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": "The diffusion process continuously perturbs the unknown data distribution " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "p_{data}" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": " with a perturbation kernel and generates intermediate data with a given diffusion time " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": ". Denote the distribution of the time-dependent intermediate nosiy data " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "p_t(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "p_0(\\boldsymbol{x}_0) = p_{data}" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": " is the clean data distribution and " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": " is the clean data. The series of intermediate data " + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 130, + 323, + 482, + 396 + ], + "type": "text", + "content": " are generated through the Stochastic Differential Equation (SDE) [39]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 203, + 401, + 480, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 401, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 203, + 401, + 480, + 414 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\boldsymbol {x} _ {t} = f (\\boldsymbol {x} _ {t}, t) \\mathrm {d} t + g (t) \\mathrm {d} \\boldsymbol {w}, \\boldsymbol {x} _ {0} \\sim p _ {\\text {d a t a}} = p _ {0} (\\boldsymbol {x} _ {0}), \\tag {1}", + "image_path": "0d7372b573e73ce69a53587ae01b00ef4aace9a2fe610a208471f4242d205867.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "inline_equation", + "content": "f(\\cdot, t)" + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "text", + "content": " is the drift coefficient, " + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "text", + "content": " is the diffusion coefficient, and " + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "text", + "content": " is the Wiener process. We can recover " + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "inline_equation", + "content": "p_{data}" + }, + { + "bbox": [ + 130, + 418, + 481, + 442 + ], + "type": "text", + "content": " from reverse-time SDE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 447, + 480, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 447, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 169, + 447, + 480, + 460 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\boldsymbol {x} _ {t} = \\left[ f (\\boldsymbol {x} _ {t}, t) - g (t) ^ {2} \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t}) \\right] \\mathrm {d} t + g (t) \\mathrm {d} \\bar {\\boldsymbol {w}}, \\boldsymbol {x} _ {T} \\sim p _ {T} (\\boldsymbol {x} _ {T}), \\tag {2}", + "image_path": "a5c2585e60b149a487d6897d847aaaf550ecada103c0f6600f122dfaeb3b6d84.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 464, + 397, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 397, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 397, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 464, + 397, + 475 + ], + "type": "inline_equation", + "content": "\\bar{\\boldsymbol{w}}" + }, + { + "bbox": [ + 131, + 464, + 397, + 475 + ], + "type": "text", + "content": " is another Wiener process, dt is negative timestep." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": "In order to solve Eq. (2), we first need to approximate " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "p_T(\\pmb{x}_T)" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": ". In previous works, " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "p_T(\\pmb{x}_T)" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " is typically approximated by some prior distributions " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "p_{prior}" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " which contain no information of " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "p_{data}" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": ". In this paper, we adopt the setting of Variance Preserving (VP) SDE [12, 39]. In VP-SDE, " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "p_T(\\pmb{x}_T) \\approx p_{prior}" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "T \\to \\infty" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": ". The perturbation kernels are in the form of " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "p_t(\\pmb{x}_t|\\pmb{x}_0) = \\mathcal{N}(\\pmb{x}_t; \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0, (1 - \\bar{\\alpha}_t)\\pmb{\\Sigma}_p)" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " where scalar " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " is diffusion schedule parameter, " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "|\\pmb{\\Sigma}_p| = 1" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": ". A common choice for " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_p" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " is the identity matrix " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " [12, 39]. Second, we need to approximate " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}_t}\\log p_t(\\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "t \\in (0,T]" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": ". Some works solve score-matching problem [14, 38] and learn a score function " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "s_\\theta(s_t,t)" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}_t}\\log p_t(\\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": ". In this paper, we follow the practice in DDPM [12] to learn the noise " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": " using a network " + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 130, + 476, + 482, + 596 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 213, + 607, + 480, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 607, + 480, + 626 + ], + "spans": [ + { + "bbox": [ + 213, + 607, + 480, + 626 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {\\boldsymbol {x} _ {0} \\sim p _ {d a t a}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {\\Sigma} _ {p})} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\tag {3}", + "image_path": "892a49b3ecc440a9ad8517fe60f317660c4dca85995bc6ea90117b8648dbb0ba.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon},\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t) = -\\sqrt{1 - \\bar{\\alpha}_t}\\pmb{s}_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": ". With these two approximations, we can learn " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": " to estimate unknown data distribution " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": " solving Eq. (2) from " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 248, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 248, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 248, + 129 + ], + "type": "text", + "content": "3.2 Guided Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 163 + ], + "type": "text", + "content": "In prior diffusion-based controllable generation frameworks [15, 16, 20, 28, 49], controllable generation is achieved via biasing of the score function for sampling:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 175, + 481, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 175, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 182, + 175, + 481, + 189 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {x} _ {t}} \\log \\left[ p _ {t} (\\boldsymbol {x} _ {t}) \\exp (- \\mathcal {C} (\\boldsymbol {x} _ {t})) \\right] = \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\mathcal {C} (\\boldsymbol {x} _ {t}) \\tag {4}", + "image_path": "bf50a22c044590f724030f6a772380f1571675afe990b88c612550eeb779b01f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": " is the guidance function. It requires estimating the guidance gradient with respect to the noisy data " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": ". Some approaches introduce an additional neural network to approximate " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": " at different noisy levels [4,15,49]. The additional neural network imposes additional training costs and heavier computational burden at the inference stage. To this end, some works define an analytical guidance function " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(\\cdot)" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": " on the clean data " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": ". They first estimate " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0 = f_\\theta(\\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": " with the diffusion model, then calculate " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(\\hat{\\pmb{x}}_0)" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": ". However, when taking the gradient of " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(\\hat{\\pmb{x}}_0)" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 194, + 482, + 291 + ], + "type": "text", + "content": ", we get" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 301, + 481, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 301, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 228, + 301, + 481, + 314 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {x} _ {t}} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0}) = \\nabla_ {\\hat {\\boldsymbol {x}} _ {0}} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0}) \\cdot \\nabla_ {\\boldsymbol {x} _ {t}} f _ {\\theta} (\\boldsymbol {x} _ {t}). \\tag {5}", + "image_path": "0cbe33b9b89b7657115b27b792dc7fd195fb22715fdc45e7b83664b74ebde570.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 323, + 481, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 481, + 348 + ], + "type": "text", + "content": "Estimating " + }, + { + "bbox": [ + 131, + 323, + 481, + 348 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}_t}f_\\theta (\\pmb{x}_t)" + }, + { + "bbox": [ + 131, + 323, + 481, + 348 + ], + "type": "text", + "content": " requires backpropagating throughout the entire diffusion model, i.e., " + }, + { + "bbox": [ + 131, + 323, + 481, + 348 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\pmb {x}_t)" + }, + { + "bbox": [ + 131, + 323, + 481, + 348 + ], + "type": "text", + "content": ". It requires heavy computing resources and GPU memory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 367, + 417, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 367, + 417, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 367, + 417, + 380 + ], + "type": "text", + "content": "3.3 Trajectory Prediction and Controllable Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": "Joint trajectory prediction aims to predict the future joint trajectories " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": " for all the vehicles in the scene, conditioned on context information " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ". It can be regarded as a conditional generation task where the goal is to train a generative model to approximate the distribution " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0|\\pmb{c})" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ". For simplicity, we omit " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": " and represent " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0|\\pmb{c})" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": " as the number of vehicles in the same scene and " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{0,i}" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": " as the future trajectory for vehicle " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,n\\}" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ", so " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 = [\\pmb{x}_{0,1},\\pmb{x}_{0,2},\\dots,\\pmb{x}_{0,n}]" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ". Joint trajectory prediction can be very challenging. The complex interactions among vehicles, especially in highly interactive and dense traffic, result in a complicated high-dimensional " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ", which is difficult to accurately model with lightweight and computationally efficient models. A simplified solution is to approximately decompose the joint trajectory distribution into marginal ones, i.e., " + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0) \\approx \\prod_i^n p_0(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": ". It leads to the marginal trajectory prediction task, which has been extensively studied with mature solutions [18, 24, 50]. One drawback is that it omits the interactions among vehicles in the predicted horizon, which leads to large errors in highly interactive scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Controllable trajectory generation is closely related to trajectory prediction. In addition to generating realistic trajectory samples resembling the ground-truth " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", controllable generation imposes an additional objective—the generated trajectories should comply with specified guidance cost function " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". We term the former objective as realism and the latter as guidance effectiveness. The guidance cost function " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " can be some goal points of the vehicles, kinematic constraints, etc, which are mostly defined on the clean manifold rather than on the noisy data when it comes to diffusion-based controllable generation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 232, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 232, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 232, + 129 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 299, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 299, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 299, + 152 + ], + "type": "text", + "content": "4.1 Optimal Gaussian Diffusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "text", + "content": "As discussed in Sec. 3.1, diffusion models typically select a non-informative prior distribution " + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "inline_equation", + "content": "p_{prior}" + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "text", + "content": ", such as a standard Gaussian, as the initial data distribution for the reverse diffusion process. Such a non-informative " + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "inline_equation", + "content": "p_{prior}" + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "text", + "content": " is reasonable since " + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "inline_equation", + "content": "p_T(\\pmb{x}_T)" + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "text", + "content": " converges to it when " + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "inline_equation", + "content": "T \\to \\infty" + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "text", + "content": ". However, it also means that a large " + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 157, + 482, + 289 + ], + "type": "text", + "content": " is required at inference time to yield good performance, which undermines computational efficiency and limits its wide real-time applications in autonomous driving. In this section, we aim to investigate a practical solution to tackle this challenge for joint trajectory prediction. Given the inherent limitation imposed by a non-informative prior, the key question we look into is: can we instead adopt an informative prior so that we can obtain the same level of performance with much smaller reverse diffusion steps?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": "First, we still consider a Gaussian prior, but with learnable parameters, i.e., we parameterize " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "p_{prior}" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "q_{\\phi}(T) = \\mathcal{N}(\\pmb{\\mu}, \\pmb{\\Sigma})" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": " are learnable. We aim to optimize " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": " to enhance the generation performance for small " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": ". We got inspiration from [5], where an upper bound of the Kullback-Leibler divergence of the clean data distribution " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0)" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": " and learned distribution " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": " was derived as a function of the diffusion time " + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 289, + 482, + 361 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 370, + 480, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 370, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 194, + 370, + 480, + 383 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} \\left[ p _ {0} \\left(\\boldsymbol {x} _ {0}\\right) \\mid \\mid q _ {\\theta} \\left(\\boldsymbol {x} _ {0}\\right) \\right] \\leq \\mathcal {G} \\left(\\boldsymbol {x} _ {\\theta}, T\\right) + \\operatorname {K L} \\left[ p _ {T} \\left(\\boldsymbol {x} _ {T}\\right) \\mid \\mid p _ {\\text {p r i o r}} \\right] \\tag {6}", + "image_path": "473671d35bef9b59152e3b476fe8d4fa0cb166667e03f5d872d14e8eef95f280.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\pmb{x}_{\\theta},T)" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " is the positive accumulated error between " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{s}_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " from 0 to " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " [5]. Note that " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\pmb{x}_{\\theta},T)" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " is an accumulated error so " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\pmb{x}_{\\theta},T_1)\\leq \\mathcal{G}(\\pmb{x}_{\\theta},T_2), T_1\\leq T_2" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": ". If we can achieve lower KL " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "[p_{T_1}(\\pmb{x}_{T_1})||q_\\phi (T_1)]" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": ", then a tighter upper bound can be obtained. This opens up the possibility to achieve better performance with less diffusion time. Thus, we propose to optimize the prior distribution by minimizing KL " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "[p_T(\\pmb{x}_T)||q_\\phi (T)]" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": ". As shown in Proposition 1 (See proof in Appendix A), it turns out that the optimal " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " can be expressed analytically as functions of the ground-truth data statistics. In addition, we find that we can further minimize the target KL divergence if we set a learnable " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{p}" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": " in the perturbation kernel " + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "inline_equation", + "content": "p_t(\\pmb {x}_t|\\pmb {x}_0) = \\mathcal{N}(\\pmb {x}_t;\\sqrt{\\bar{\\alpha}_t}\\pmb {x}_0,(1 - \\bar{\\alpha}_t)\\pmb {\\Sigma}_p)" + }, + { + "bbox": [ + 131, + 390, + 483, + 521 + ], + "type": "text", + "content": ", whose optimal value can also be expressed as a function of the data statistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": "Proposition 1. Denote " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_d" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_d" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": " as the mean and variance of " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "p_{data}" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": ". Denote " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{*}(i,j)" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{p}^{*}(i,j)" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": " as the element at ith row and jth column of matrix " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^*" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{p}^{*}" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": ". The optimal solution to " + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\min KL[p_T(\\pmb{x}_T)||q_\\phi (\\pmb{x}_T,T)]" + }, + { + "bbox": [ + 132, + 529, + 482, + 567 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 170, + 578, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 578, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 170, + 578, + 480, + 622 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\boldsymbol {\\mu} ^ {*} \\approx \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} \\\\ \\boldsymbol {\\Sigma} _ {p} ^ {*} (i, j) \\approx \\frac {1}{| \\boldsymbol {\\Sigma} _ {d} |} \\boldsymbol {\\Sigma} _ {d} (i, j) \\\\ \\boldsymbol {\\Sigma} ^ {*} (i, j) \\approx \\bar {\\alpha} _ {T} \\boldsymbol {\\Sigma} _ {d} + (1 - \\bar {\\alpha} _ {T}) ^ {2} \\boldsymbol {\\Sigma} _ {p} ^ {*} = (\\bar {\\alpha} _ {T} ^ {2} + \\frac {(1 - \\bar {\\alpha} _ {T}) ^ {2}}{| \\boldsymbol {\\Sigma} _ {d} |}) \\boldsymbol {\\Sigma} _ {d} (i, j) \\end{array} \\right. \\tag {7}", + "image_path": "4ea15c3dc2aaccc7d9771797dd6f610caafe798b907fb5cb8816594963a6f52c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": "Thus, if we are able to estimate the mean and variance of the ground-truth data distribution, we can then analytically determine the optimal prior distribution " + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\pmb{\\mu}^*,\\pmb{\\Sigma}^*)" + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": " and the optimal perturbation kernel variance " + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_p^*" + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": " at any noise" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "level " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "T \\in [0, \\infty)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". It leads to a crucial advantage against prior efforts on prior learning [48], where a target noise level has to be determined at training time in order to train an additional neural network to represent the prior distribution at the pre-selected noise level. In contrast, our method enables flexible tuning of the number of diffusion steps at inference time, without additional training costs. Since the learnable " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "q_{\\phi}(T)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and perturbation kernel are both Gaussian, we refer to our proposed diffusion framework as Optimal Gaussian Diffusion (OGD)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 483, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 483, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 483, + 296 + ], + "type": "text", + "content": "For joint trajectory prediction and generation, we need to estimate the mean and variance of the joint trajectory distribution " + }, + { + "bbox": [ + 130, + 200, + 483, + 296 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 200, + 483, + 296 + ], + "type": "text", + "content": ". It is not straightforward as only a limited number of trajectory samples exist in the dataset under the same context. Considering there exists mature and accurate marginal trajectory prediction models [18, 24, 50], we can conveniently extract statistics of the marginal trajectory distributions, i.e., " + }, + { + "bbox": [ + 130, + 200, + 483, + 296 + ], + "type": "inline_equation", + "content": "p_0(\\pmb{x}_{0,i}), i \\in \\{1,2,\\dots,n\\}" + }, + { + "bbox": [ + 130, + 200, + 483, + 296 + ], + "type": "text", + "content": " from a pretrained marginal trajectory predictor. Since Proposition 1 provides element-wise optimal value, we can easily get:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "content": "Corollary 1. Denote " + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_d(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "content": " as the mean and variance of marginal distribution for vehicle " + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "content": " and set both " + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_p" + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "content": " are block-diagonal matrices where each block represents the marginal characteristics of one single vehicle. The optimal solution to " + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "inline_equation", + "content": "\\min KL[p_T(\\pmb{x}_T)||q_\\phi (\\pmb{x}_T,T)]" + }, + { + "bbox": [ + 132, + 300, + 491, + 350 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 359, + 481, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 359, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 184, + 359, + 481, + 403 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\boldsymbol {\\mu} ^ {*} = \\left[ \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\sqrt {\\bar {\\alpha} _ {T}} \\boldsymbol {\\mu} _ {d} (\\boldsymbol {x} _ {0, n}) \\right] ^ {T} \\\\ \\boldsymbol {\\Sigma} _ {p} ^ {*} = \\frac {1}{\\sum_ {i = 1} ^ {n} | \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0 , i}) |} d i a g [ \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, n}) ] \\\\ \\boldsymbol {\\Sigma} ^ {*} = \\bar {\\alpha} _ {T} d i a g [ \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, 1}), \\dots , \\boldsymbol {\\Sigma} _ {d} (\\boldsymbol {x} _ {0, n}) ] + (1 - \\bar {\\alpha} _ {T}) ^ {2} \\boldsymbol {\\Sigma} _ {p} ^ {*} \\end{array} \\right. \\tag {8}", + "image_path": "5103db7de823c7a9af6d2a2fa7be931b790327ffd11db1b570239eca90341fab.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": "Corollary 1 implies that, if we further confine " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_p" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": " to be block-diagonal without covariance between the states of different vehicles, then we can determine their optimal values purely from the estimated marginal statistics " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_d(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,n\\}" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": ", which enables a practical implementation of the proposed OGD model for joint trajectory prediction and generation tasks. Specifically, for vehicle " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": ", we leverage a pre-trained marginal trajectory predictor [18,24,50], predict diverse marginal trajectory sample set " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i = \\{\\pmb{r}_i^l\\}_{l=1}^L" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": " and corresponding likelihood set " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\{p(\\pmb{r}_i^l)\\}_{l=1}^L" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": ", and estimate " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_d(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_d(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": ". For example, " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_d(\\pmb{x}_{0,i})" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": " can be estimated as " + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "inline_equation", + "content": "\\frac{1}{L}\\sum_{l=L}p(\\pmb{r}_i^l)\\pmb{r}_i^l" + }, + { + "bbox": [ + 130, + 407, + 483, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 529, + 424, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 424, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 424, + 540 + ], + "type": "text", + "content": "4.2 Estimated Clean Manifold Guidance with Reference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "As discussed in Sec. 3.2, the intensive computation required for guided sampling comes from the calculation of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{x}_t}f_\\theta (\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". Previous guided sampling approaches bias the score function defined on the intermediate noisy data " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". In this section, we aim to investigate whether we can inject the gradient directly into " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " to avoid the gradient propagation process. We first reformulate the controllable generation as a multi-objective optimization problem directly over " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and propose an iterative algorithm to solve the formulated problem. In addition, we use reference trajectory points to create the region of interest, which helps solve the local optimal problem caused by multi-modal joint trajectory distribution and accelerate the guided sampling process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 175 + ], + "type": "text", + "content": "Estimated Clean Manifold Guidance. The objective of controllable generation of sample " + }, + { + "bbox": [ + 132, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 132, + 116, + 479, + 175 + ], + "type": "text", + "content": " includes two different objectives. The most important is the negative likelihood, ensuring the sample lies in the clean manifold. The second important is the guidance cost representing the user preference on the generated sample " + }, + { + "bbox": [ + 132, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 132, + 116, + 479, + 175 + ], + "type": "text", + "content": ". This multi-objective optimization problem can be represented as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 186, + 479, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 186, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 246, + 186, + 479, + 202 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {x} _ {0}} \\left[ - \\log q _ {\\theta} (\\boldsymbol {x} _ {0}), \\mathcal {J} (\\boldsymbol {x} _ {0}) \\right] ^ {T} \\tag {9}", + "image_path": "97d843e8251bb57f6005ed64d30dd17b6fca191d9c8af81d86bb5e54822e0661.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": "Inspired by lexicographic optimization [33], we solve this multi-objective optimization problem hierarchically. The main idea is to optimize each objective in the order of importance regardless of the degradation of the other less significant objectives. We first optimize the most important objective, " + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "inline_equation", + "content": "-\\log q_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": ", to generate realistic and high-likelihood samples. The diffusion model achieves this goal effectively by reversing the diffusion process from noisy samples at specific noise level. However, exact noise level of current sample " + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": " is unknown. To address this, we inject noise at level " + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": ", and then denoise it from " + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": " with learned diffusion model. This approach, similar to noise injection and denoising [1,23,37], improves the desired sample quality. We iteratively repeat this process " + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": " times, injecting guidance at each iteration to strengthen user preference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "text", + "content": "Specifically, denote " + }, + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0(k)" + }, + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "text", + "content": " as the sample at iteration " + }, + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "text", + "content": ". We first regenerate high-likelihood sample " + }, + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0(k)" + }, + { + "bbox": [ + 130, + 344, + 479, + 369 + ], + "type": "text", + "content": " by diffusion model:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 378, + 479, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 378, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 149, + 378, + 479, + 402 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {0} (k) \\leftarrow \\mathbb {E} \\left[ q _ {\\theta} \\left(\\boldsymbol {x} _ {0} (k) \\mid \\boldsymbol {x} _ {t _ {k}} (k)\\right) \\right] = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t _ {k}}}} \\left(\\boldsymbol {x} _ {t _ {k}} - \\sqrt {1 - \\bar {\\alpha} _ {t _ {k}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t _ {k}} (k), t _ {k}\\right)\\right), \\tag {10}", + "image_path": "8b210b2d0631414bea8c54187f8f0a04bd5b073f782a302325a107bc0b80353f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t_k}(k) = \\sqrt{\\bar{\\alpha}_{t_k}}\\pmb{x}_0(k) + \\sqrt{1 - \\bar{\\alpha}_{t_k}}\\pmb{\\epsilon}, \\pmb{\\epsilon} \\sim \\mathcal{N}(\\pmb{0},\\pmb{\\Sigma}_p)" + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "inline_equation", + "content": "t_k \\in [0,T]" + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "text", + "content": " is a tunable parameter. Then we minimize the guidance cost function with a small degradation of the most important objective " + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "inline_equation", + "content": "-\\log q_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 223, + 459, + 479, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 459, + 479, + 470 + ], + "spans": [ + { + "bbox": [ + 223, + 459, + 479, + 470 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {0} (k - 1) \\leftarrow \\hat {\\boldsymbol {x}} _ {0} (k) - \\zeta \\nabla_ {\\hat {\\boldsymbol {x}} _ {0} (k)} \\mathcal {J} (\\hat {\\boldsymbol {x}} _ {0} (k)) \\tag {11}", + "image_path": "f98909488b8d0b3acc9068d6edfd56f0cb492954767debe273340d917d439536.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "text", + "content": " is the step size. Small degradation of " + }, + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "inline_equation", + "content": "-\\log q_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "text", + "content": " is realized by one-step gradient update and proper step size " + }, + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 479, + 479, + 514 + ], + "type": "text", + "content": ". See the derivation of the optimization process and the parameter tuning in Appendix B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 516, + 479, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 479, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 479, + 563 + ], + "type": "text", + "content": "During the iterations, " + }, + { + "bbox": [ + 130, + 516, + 479, + 563 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0(k), \\forall k \\in \\{0,1,\\dots,K - 1\\}" + }, + { + "bbox": [ + 130, + 516, + 479, + 563 + ], + "type": "text", + "content": " are not exactly on the clean manifold. We minimize its negative log-likelihood through diffusion model, resulting in it lying on an estimated clean manifold. Thus, we call our method Estimated Clean Manifold (ECM) Guidance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": "Reference Joint Trajectories. To generate trajectories with low guidance cost, we are essentially searching low-cost trajectories within the high-likelihood region. At the same time, joint trajectory distribution is a multi-modal distribution resulting from road topologies and different decision variables, meaning the likelihood has multiple peaks. This leads to the optimal solution of Problem 9 having multiple local optimals, and each optimal is far away from the other. Guided sampling methods [15,16,28,49], including our method ECM, suffer from" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 116, + 463, + 210 + ], + "blocks": [ + { + "bbox": [ + 149, + 116, + 463, + 210 + ], + "lines": [ + { + "bbox": [ + 149, + 116, + 463, + 210 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 463, + 210 + ], + "type": "image", + "image_path": "35e3bb102b3b53c1b1e68f1e89625642576cf47a0668139d544a34789a2fcb50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 266 + ], + "type": "text", + "content": "Fig. 2: Two challenges with multi-peak function optimization: 1) Gradients may lead to suboptimal local optima (left); 2) There exist regions with low likelihood but high guidance cost uncertainty, leading to instability (right). Our approach can bypass the lengthy paths between peaks, search for better optima, and avoid uncertain areas." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 286, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 335 + ], + "type": "text", + "content": "two challenges (See Fig. 2): 1) it can be trapped at the local optimal around the initial position; 2) it takes massive efforts to drag the sample from one peak to another and transferring from one modal to another will need to pass through the region of low-likelihood (off clean-manifold), leading to numerical instability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 263, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 263, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 263, + 501 + ], + "type": "text", + "content": "To overcome this, we generate high-likelihood reference joint trajectories, choose the best one as the initialization. Note that combinations of samples with high marginal likelihood tend to exhibit high joint likelihood. Therefore, we can utilize the marginal sample set " + }, + { + "bbox": [ + 130, + 335, + 263, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{\\mathcal{R}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 335, + 263, + 501 + ], + "type": "text", + "content": " obtained from pre-trained marginal models to generate the references. Specifically, for iteration " + }, + { + "bbox": [ + 130, + 335, + 263, + 501 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 335, + 263, + 501 + ], + "type": "text", + "content": ", we" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "text", + "content": "construct candidate joint trajectory set " + }, + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\otimes \\hat{\\pmb{x}}_0(k) = \\{[\\pmb{w}_1, \\pmb{w}_2, \\dots, \\pmb{w}_n] | \\pmb{w}_i \\in \\mathcal{R}_i \\cup \\{\\hat{\\pmb{x}}_{0,i}\\}, i = 1, 2, \\dots, n\\}" + }, + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "text", + "content": ". We calculate the guidance cost of all possible combinations, " + }, + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(w)" + }, + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "inline_equation", + "content": "w \\in \\mathcal{R} \\otimes \\hat{\\pmb{x}}_0(k)" + }, + { + "bbox": [ + 130, + 501, + 480, + 562 + ], + "type": "text", + "content": ", and choose the minimal-cost one as the reference. The guided sampling algorithm, ECM with reference joint trajectories (ECMR), is introduced in Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 600, + 260, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 260, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 260, + 612 + ], + "type": "text", + "content": "5.1 Experimental setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Dataset. We use Argoverse 2 [44], a widely used and large-scale trajectory prediction dataset, to test the effectiveness of our approaches for joint trajectory prediction and controllable generation. It has a large observation window of 5s and a long prediction horizon of 6s." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 269, + 376, + 481, + 473 + ], + "blocks": [ + { + "bbox": [ + 278, + 363, + 379, + 374 + ], + "lines": [ + { + "bbox": [ + 278, + 363, + 379, + 374 + ], + "spans": [ + { + "bbox": [ + 278, + 363, + 379, + 374 + ], + "type": "text", + "content": "Algorithm 1: ECMR" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 269, + 376, + 481, + 473 + ], + "lines": [ + { + "bbox": [ + 269, + 376, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 269, + 376, + 481, + 473 + ], + "type": "table", + "html": "
Input: J(·), {αt,βt}T-1t=0, {tk}K-1k=0
1: x0(K) ~ N(0, Σp)
2: for k = K-1,..., 1 do
3: ε ~ N(0, Σp)
4: xtk(k) = √αtkx0(k) + √1-αtkε
5: x0(k) = 1/√αtk(xtk(k) - √1-αtk)εθ(xtk(k),t))
6: x0(k) = arg min J(w), w ∈ R⊗ x0(k)
7: x0(k-1) ← x0(k) - ζ∇x0(k)J(x0(k))
8: end for
Output: x0(0)
", + "image_path": "48a5a8eee9a6c4266adf53c7fb8417137de55d41071e0edf6f540852a6a665ce.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "Implementation Details. We use the fixed scene context encoder of pretrained QCNet [50] to extract compact and representative context features from context information " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". Then, we utilize a cross-attention layer to update the intermediate noisy data " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " with multiple contexts, including the history encodings of the target agent, the map encodings, the neighboring agents' encodings. Inspired by [7], we also add a cross-attention layer to update " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " with the diverse marginal trajectory samples " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{r}_i^l" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and its corresponding likelihood. In addition, we use self-attention to allow the interaction between " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t,i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t,j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". Then the model predicts the noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". According to [16], compact trajectory representation helps the diffusion model to generate high-quality trajectories efficiently. Inspired by this, we also learn a linear mapping between the 10-dimensional latent and 120-dimensional trajectories. Similar to [16], we design a rapid sample clustering algorithm so that we can generate a representative joint trajectory set. To increase the efficiency of sampling, we use DDIM [35] to accelerate the inference, and the DDIM step stride is 10. See Appendix C for details and analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 317, + 295, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 317, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 317, + 295, + 330 + ], + "type": "text", + "content": "5.2 Joint trajectory prediction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": "We now evaluate OGD for joint trajectory prediction. Given " + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": " joint trajectories, the evaluation metrics are 1) avgMinFDE" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": "_K" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": "/avgMinADE" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": "_K" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": ": the average of lowest final/average displacement error (FDE/ADE) of joint trajectory samples; 2) actorMR" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": "_K" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": ": the rate of trajectory predictions that are considered to be \"missed\" (" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": ">2m" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": " FDE) in the lowest minFDE joint trajectory samples; 3) actorCR" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": "_K" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": ": the rate of collisions across \"best\" (lowest avgMinFDE) joint trajectory samples; 4) avgBrierMinFDE" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": "_K" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": ": calculated similarly to avgMinFDE" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "inline_equation", + "content": "_K" + }, + { + "bbox": [ + 130, + 341, + 482, + 449 + ], + "type": "text", + "content": " but scaled by the probability score of joint trajectory samples. We denote metrics with superscript \"\\*\" as those after sample clustering (see Appendix C.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": "As a baseline, we train a vanilla diffusion (VD) baseline that shares the same neural network architecture as OGD. Specifically, we train an Optimal Gaussian Diffusion model with diffusion time " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T_{train} = 100" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " (OGD), vanilla diffusion with two different diffusion times " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T_{train} = 100,500" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{VD}_{100}, \\mathrm{VD}_{500}" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": "). Note that " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " is denoted as the diffusion time from which the reverse diffusion process starts. First, during the inference, we change " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " and evaluate OGD and " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{VD}_{100}" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " who have the same " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T_{train} = 100" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": ". Figure 3 shows that, with the decrease of the reverse steps, avgMinFDE128 of OGD keeps lower than " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{VD}_{100}" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " and it is more stable with the change of " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": ". It is also interesting to note that the performance of OGD even becomes better in the early stage. We hypothesize that it is because that " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}[p_T(\\boldsymbol{x}_T)||p_{prior}]" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " of OGD is small and the accumulated error of " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\boldsymbol{x}_\\theta, T)" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " is sufficiently reduced by a small " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": ". It also shows an advantage of OGD, which is that it is easy for OGD to tune a suitable " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " for reverse diffusion without training an additional model for every " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " [48]. Table 1 shows that OGD outperforms " + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{VD}_{500}" + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": " with only 40 diffusion steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "We also compare our OGD with the other state-of-the-art methods on the Argoverse multi-world leaderboard; our approach OGD ranks " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " on the leaderboard ranked by avgBrierMinFDE" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "_K" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", which demonstrates the effectiveness of our" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 125, + 254, + 216 + ], + "blocks": [ + { + "bbox": [ + 138, + 125, + 254, + 216 + ], + "lines": [ + { + "bbox": [ + 138, + 125, + 254, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 125, + 254, + 216 + ], + "type": "image", + "image_path": "b007635a1852e6cad01ff697f59805204a1361e9b002412795d7953320f36e96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 223, + 266, + 256 + ], + "lines": [ + { + "bbox": [ + 132, + 223, + 266, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 223, + 266, + 256 + ], + "type": "text", + "content": "Fig.3: Evaluation of Optimal Gaussian Diffusion and vanilla diffusion over reverse steps " + }, + { + "bbox": [ + 132, + 223, + 266, + 256 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 223, + 266, + 256 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 272, + 188, + 479, + 248 + ], + "blocks": [ + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "lines": [ + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "spans": [ + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "text", + "content": "Table 1: Evaluation on joint trajectory prediction task. For each metric, the best result is in **bold** and the second best result is **underlined**. " + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "inline_equation", + "content": "T = 70" + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "text", + "content": " is the best " + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "text", + "content": " from Fig. 3. " + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "inline_equation", + "content": "T = 40" + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "text", + "content": " is the minimal diffusion time when OGD outperforms " + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{VD}_{500}" + }, + { + "bbox": [ + 270, + 124, + 482, + 179 + ], + "type": "text", + "content": " on all metrics." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 272, + 188, + 479, + 248 + ], + "lines": [ + { + "bbox": [ + 272, + 188, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 272, + 188, + 479, + 248 + ], + "type": "table", + "html": "
ModelTavgMinFDE*avgMinADE*avgMinFDE128avgMinADE128
VD1001000.621.380.490.97
VD5005000.611.360.480.91
OGD1000.601.320.430.81
OGD700.591.310.420.75
OGD400.611.340.470.77
", + "image_path": "8308f5bc1680249b79ca240f51227f0425b79156f8376521c53f99ee4ff5a903.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "content": "OGD framework. Note that we only list the entries with publications or technical reports in Tab. 2. Please refer to the official website for the full leaderboard." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 319, + 282, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 282, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 282, + 331 + ], + "type": "text", + "content": "5.3 Controllable Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 339, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 447 + ], + "type": "text", + "content": "Tasks. Future behaviors of vehicles can be effectively can be effectively represented by a set of goal points [9, 47] such as acceleration, braking, and right or left turn. Generating such diverse modes of joint trajectories is a typical motivation for using controllable generation. Thus, we study the controllable generation task where the goal is to reach diverse goal points at a specific time to fully test our guided sampling method. Considering that goal points should lie in realistic routes, and the trajectories to reach such goal points should also lie in such routes, we generate such target goal points in some realistic routes " + }, + { + "bbox": [ + 130, + 339, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathrm{Route}(\\tau_g)" + }, + { + "bbox": [ + 130, + 339, + 482, + 447 + ], + "type": "text", + "content": ". The guidance cost function can be expressed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 209, + 453, + 480, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 453, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 209, + 453, + 480, + 477 + ], + "type": "interline_equation", + "content": "\\mathcal {J} \\left(\\boldsymbol {x} _ {0}\\right) = \\frac {1}{n} \\left\\| \\operatorname {P o s i t i o n} \\left(\\boldsymbol {x} _ {0}, \\tau_ {d}\\right) - \\operatorname {R o u t e} \\left(\\tau_ {g}\\right) \\right\\| _ {2} ^ {2}, \\tag {12}", + "image_path": "1915345d93a127d59e2aaa933d09a23855ce18bd53b14770c8ce4ab5c32a484f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{Position}(\\pmb{x}_0, \\tau_d)" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": " is the positions of " + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\tau_d" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": ". We choose ground-truth trajectories and a random combination set of diverse marginal samples, i.e., " + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\{[\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_n] | \\pmb{u}_i \\in \\mathcal{R}_i, i = 1, 2, \\dots, n\\}" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": " as the realistic routes. We denote" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 135, + 590, + 482, + 654 + ], + "blocks": [ + { + "bbox": [ + 130, + 540, + 479, + 562 + ], + "lines": [ + { + "bbox": [ + 130, + 540, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 479, + 562 + ], + "type": "text", + "content": "Table 2: Quantitative results on the Argoverse 2 Multi-world Forecasting leaderboard. For each metric, the best result is in bold and the second best result is underlined." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 590, + 482, + 654 + ], + "lines": [ + { + "bbox": [ + 135, + 590, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 135, + 590, + 482, + 654 + ], + "type": "table", + "html": "
ModelavgMinFDE*6avgMinFDE1*actorMR6avgMinADE*6avgMinADE1*avgBrierMinFDE*6actorCR6
QCXet [51]1.022.290.130.500.941.650.01
Gnet [6]1.463.050.190.691.232.120.01
Forecast-MAE [2]1.553.330.190.691.302.240.01
FJMP [30]1.894.000.230.811.522.590.01
OGD (Ours)1.312.710.170.601.081.950.01
", + "image_path": "df3971bf348b37f04daa90a3a07e3a1cf03b57f73d5fd57dfc17d46b2d49541a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 216, + 193 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 216, + 193 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 216, + 193 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 216, + 193 + ], + "type": "image", + "image_path": "cd2c4fc275eb274906f664ebf2d8d0a023fbc099e1021823ade9e6ce6ac94c8f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "text", + "content": "Fig. 4: Evaluation on controllable generation: route set " + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 130, + 201, + 482, + 246 + ], + "type": "text", + "content": " and Deceleration. Magenta diamonds represent goal points. In the first (second) row, goal points are set at the fork lane (right lane). NNM [49] and SF [16, 28] struggle to drag samples from one modal to another. Our methods can achieve better guidance effectiveness and realism." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 224, + 116, + 304, + 194 + ], + "blocks": [ + { + "bbox": [ + 224, + 116, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 224, + 116, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 224, + 116, + 304, + 194 + ], + "type": "image", + "image_path": "0d617d43fdb73080ada75572219dd3e8fb80437fc7d31d06249af1e3ab6d25d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 116, + 394, + 193 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 394, + 193 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 394, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 394, + 193 + ], + "type": "image", + "image_path": "f2826df3fa6be27717cf866bb6e0d50a84b0dd9747e514d6b0866fb580ee98ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 403, + 115, + 481, + 194 + ], + "blocks": [ + { + "bbox": [ + 403, + 115, + 481, + 194 + ], + "lines": [ + { + "bbox": [ + 403, + 115, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 403, + 115, + 481, + 194 + ], + "type": "image", + "image_path": "f5820e3377fcdc4969a8b6dc291ffff9c9a6ee1effda1a9146dd7bf9bbe84d4e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "text", + "content": "the former as GT set and the latter as " + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "text", + "content": " set. The underlying assumption is that diverse samples from a good marginal trajectory predictor are realistic. We design different velocity settings to cover the diverse controllable generation tasks in autonomous driving: first is Normal Speed (N), " + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\tau_{d} = \\tau_{g} = 6s" + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "text", + "content": "; second is Acceleration (A), " + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\tau_{d} = 5s < \\tau_{g} = 6s" + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "text", + "content": "; third is Deceleration (D), " + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\tau_{d} = 6s < \\tau_{g} = 5s" + }, + { + "bbox": [ + 130, + 270, + 482, + 341 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "type": "text", + "content": "Metrics. We use the following metrics to evaluate controllable generation performance: Joint Route Deviation Error (JRDE), which measures the displacements to the routes to evaluate the realism, and Joint Final Displacement Error (JFDE), which evaluates the guidance effectiveness. We also evaluate from the \"min\" and \"mean\" perspectives: The \"min\" metric considers the best sample's performance, while the \"mean\" metric assesses the ratio of valid samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "content": "Baselines. Guided sampling in controllable generation is mainly divided into two approaches: the first is to directly calculate " + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{x}_t}\\mathcal{J}(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "content": " [49]; the second is to calculate " + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{x}_t}\\mathcal{J}(\\hat{\\boldsymbol{x}}_0)" + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "content": " [16, 28]. We denote the former as Next Noisy Mean Guidance (NNM) and the latter as Score Function Guidance (SF). For a fair comparison, we use one guidance step followed by one DDIM step. We also tune the gradient step size for different guided sampling with Optimal Gaussian Diffusion and vanilla diffusion and report the results with the optimal step size. See Appendix D for the details of baseline derivation and step size tuning. We evaluate the following experiments with 128 joint trajectory samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Evaluation. First, we evaluate the performance and efficiency with our diffusion model (OGD) and guided sampling method (ECM and ECMR) in Tab. 3, which demonstrates our methods can generate more realistic and effective samples with 5 times less DDIM steps. In addition, with reference joint trajectories, ECM significantly improves 'mean' metrics, indicating it addresses the issues discussed in Sec. 4.2 to a certain extent. Second, we compare solely on different guided sampling methods using the same OGD model shown in Fig. 5. Our ECM achieves better performance both in guidance effectiveness and realism. In Fig. 4, our ECM can generate trajectories that reach goal points more closely than NNM and SF. And with reference joint trajectory, ECMR can easily move samples from one modal to another. Third, we evaluate the average inference time and average GPU memory usage in Tab. 4. Our methods can generate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 90, + 448, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 156, + 475, + 262 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Evaluation on controllable generation: route set U and Deceleration. For each metric, the best result is in bold and the second best result is underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 156, + 475, + 262 + ], + "lines": [ + { + "bbox": [ + 140, + 156, + 475, + 262 + ], + "spans": [ + { + "bbox": [ + 140, + 156, + 475, + 262 + ], + "type": "table", + "html": "
ModelSamplingDDIM StepsGuidance minJFDEEffectiveness meanJFDERealism minJRDEmeanJRDE
VD500No Guid501.9615.2290.1650.492
VD500NNM [49]500.7782.9130.1300.309
VD500SF [16, 28]500.5382.3390.1580.500
OGDNo Guid101.7725.1720.1380.469
OGDECM (Ours)100.0720.2370.1280.236
OGDECMR (Ours)100.0530.1460.1100.154
", + "image_path": "26364905998bd8627052369ae1edfe20eee12ec9554b483e2f66014c433264c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 285, + 291, + 364 + ], + "blocks": [ + { + "bbox": [ + 133, + 285, + 291, + 364 + ], + "lines": [ + { + "bbox": [ + 133, + 285, + 291, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 291, + 364 + ], + "type": "image", + "image_path": "37ff5e2a75e2db6356b4fabdda5bd2949af8aafc47e61308f7aa0cd4c341a16f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 373, + 291, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 373, + 291, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 291, + 407 + ], + "type": "text", + "content": "Fig. 5: Evaluation of different guided sampling methods on various tasks. See Appendix E for other metrics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 308, + 352, + 479, + 404 + ], + "blocks": [ + { + "bbox": [ + 304, + 288, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 288, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 482, + 344 + ], + "type": "text", + "content": "Table 4: We evaluate the average inference time per step and GPU incremental memory on " + }, + { + "bbox": [ + 304, + 288, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{U} +" + }, + { + "bbox": [ + 304, + 288, + 482, + 344 + ], + "type": "text", + "content": " Deceleration. We test on single RTX A6000, batch size is 16 and number of samples is 128." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 352, + 479, + 404 + ], + "lines": [ + { + "bbox": [ + 308, + 352, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 479, + 404 + ], + "type": "table", + "html": "
SamplingminJFDEminJRDEtime(ms)memory (GB)
NNM [49]0.7240.1191133.21
SF [16,28]0.1550.1262477.96
ECM(Ours)0.0750.1161113.21
ECMR(Ours)0.0530.1101163.22
", + "image_path": "4193b86ee1a4e92ad77051a649d0130efea584c828aa3aace4f6eec13116e957.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 469 + ], + "type": "text", + "content": "realistic trajectories satisfying guidance quite well with low inference time and GPU memory usage. More results on controllable generation can be found in Appendix E." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 491, + 220, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 220, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 220, + 504 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce Optimal Gaussian Diffusion (OGD) and Estimated Clean Manifold (ECM) Guidance to significantly improve the computational efficiency and performance of diffusion models in autonomous driving. These methodologies enable a substantial reduction in inference steps and computational demands while ensuring enhanced joint trajectory prediction and controllable generation capabilities. Our approaches and experimental results underscore the potential of diffusion models for real-time applications in dynamic environments, marking a pivotal advancement in the deployment of diffusion models for autonomous driving. One limitation of the current implementation is that the performance is affected by the accuracy of marginal trajectory predictors. Enhancements could be achieved with superior marginal models or by directly learning joint predictions' mean and variance, areas for future work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 140, + 373, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 140, + 373, + 152 + ], + "spans": [ + { + "bbox": [ + 147, + 140, + 373, + 152 + ], + "type": "text", + "content": "This work was supported by Berkeley DeepDrive." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 197, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 197, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 197, + 183 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 196, + 481, + 646 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "text", + "content": "1. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18208-18218 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 230, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 230, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 230, + 481, + 262 + ], + "type": "text", + "content": "2. Cheng, J., Mei, X., Liu, M.: Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8679-8689 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 263, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 481, + 306 + ], + "type": "text", + "content": "3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 307, + 481, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 481, + 327 + ], + "type": "text", + "content": "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 329, + 481, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 481, + 360 + ], + "type": "text", + "content": "5. Franzese, G., Rossi, S., Yang, L., Finamore, A., Rossi, D., Filippone, M., Michiardi, P.: How much is enough? a study on diffusion times in score-based generative models. Entropy 25(4), 633 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "type": "text", + "content": "6. Gao, X., Jia, X., Li, Y., Xiong, H.: Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolutional recurrent networks. IEEE Robotics and Automation Letters 8(5), 2946-2953 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 395, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 395, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 395, + 481, + 426 + ], + "type": "text", + "content": "7. Gilles, T., Sabatini, S., Tsishkou, D., Stanciulescu, B., Moutarde, F.: Thomas: Trajectory heatmap output with learned multi-agent sampling. arXiv preprint arXiv:2110.06607 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 481, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 459 + ], + "type": "text", + "content": "8. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural information processing systems. pp. 2672-2680 (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 492 + ], + "type": "text", + "content": "9. Gu, J., Sun, C., Zhao, H.: Densetnt: End-to-end trajectory prediction from dense goal sets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15303-15312 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 494, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 481, + 525 + ], + "type": "text", + "content": "0. Gu, T., Chen, G., Li, J., Lin, C., Rao, Y., Zhou, J., Lu, J.: Stochastic trajectory prediction via motion indeterminacy diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17113-17122 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 526, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 526, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 526, + 481, + 557 + ], + "type": "text", + "content": "1. Guo, Z., Gao, X., Zhou, J., Cai, X., Shi, B.: Scenedm: Scene-level multiagent trajectory generation with consistent diffusion models. arXiv preprint arXiv:2311.15736 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 559, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 559, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 138, + 559, + 481, + 580 + ], + "type": "text", + "content": "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 582, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 582, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 582, + 481, + 601 + ], + "type": "text", + "content": "3. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 604, + 481, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 604, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 138, + 604, + 481, + 624 + ], + "type": "text", + "content": "4. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 625, + 481, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 625, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 138, + 625, + 481, + 646 + ], + "type": "text", + "content": "5. Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 448, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 653, + 268, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 268, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 268, + 665 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 134, + 653, + 268, + 665 + ], + "type": "text", + "content": " https://deepdrive.berkeley.edu" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "16. Jiang, C., Cornman, A., Park, C., Sapp, B., Zhou, Y., Anguelov, D., et al.: Motion-diffuser: Controllable multi-agent motion prediction using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9644-9653 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 182 + ], + "type": "text", + "content": "17. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. Advances in neural information processing systems 34, 21696-21707 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 134, + 183, + 481, + 215 + ], + "type": "text", + "content": "18. Lan, Z., Jiang, Y., Mu, Y., Chen, C., Li, S.E., Zhao, H., Li, K.: Sept: Towards efficient scene representation learning for motion prediction. arXiv preprint arXiv:2309.15289 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 134, + 216, + 481, + 248 + ], + "type": "text", + "content": "19. Laumanns, M., Thiele, L., Zitzler, E.: An efficient, adaptive parameter variation scheme for metaheuristics based on the epsilon-constraint method. European Journal of Operational Research 169(3), 932-942 (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "text", + "content": "20. Lin, H., Wang, Y., Huo, M., Peng, C., Liu, Z., Tomizuka, M.: Joint pedestrian trajectory prediction through posterior sampling. arXiv preprint arXiv:2404.00237 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "21. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35, 5775-5787 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "22. Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837-2845 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "23. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022), https://arxiv.org/abs/2108.01073" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "text", + "content": "24. Nayakanti, N., Al-Rfou, R., Zhou, A., Goel, K., Refaat, K.S., Sapp, B.: Wayformer: Motion forecasting via simple & efficient attention networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2980-2987. IEEE (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "text", + "content": "25. Ngiam, J., Caine, B., Vasudevan, V., Zhang, Z., Chiang, H.T.L., Ling, J., Roelofs, R., Bewley, A., Liu, C., Venugopal, A., et al.: Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "type": "text", + "content": "26. Peng, C., Wang, G., Lo, X.W., Wu, X., Xu, C., Tomizuka, M., Zhan, W., Wang, H.: Delflow: Dense efficient learning of scene flow for large-scale point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16901-16910 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "27. Peng, C., Xu, C., Wang, Y., Ding, M., Yang, H., Tomizuka, M., Keutzer, K., Pavone, M., Zhan, W.: Q-slam: Quadric representations for monocular slam. arXiv preprint arXiv:2403.08125 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "28. Rempe, D., Luo, Z., Bin Peng, X., Yuan, Y., Kitani, K., Kreis, K., Fidler, S., Litany, O.: Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13756-13766 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "30. Rowe, L., Ethier, M., Dykhne, E.H., Czarnecki, K.: Fjmp: Factorized joint multiagent motion prediction over learned directed acyclic interaction graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13745-13755 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "31. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "32. San-Roman, R., Nachmani, E., Wolf, L.: Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "text", + "content": "33. Sherali, H.D., Soyster, A.L.: Preemptive and nonpreemptive multi-objective programming: Relationship and counterexamples. Journal of Optimization Theory and Applications 39, 173-186 (1983)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 197, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 480, + 228 + ], + "type": "text", + "content": "34. Shi, S., Jiang, L., Dai, D., Schiele, B.: Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. arXiv preprint arXiv:2306.17770 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 230, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 230, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 230, + 480, + 251 + ], + "type": "text", + "content": "35. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 252, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 252, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 133, + 252, + 480, + 284 + ], + "type": "text", + "content": "36. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 480, + 306 + ], + "type": "text", + "content": "37. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. arXiv preprint arXiv:2303.01469 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 480, + 329 + ], + "type": "text", + "content": "38. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 480, + 363 + ], + "type": "text", + "content": "39. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "type": "text", + "content": "40. Sun, Q., Huang, X., Gu, J., Williams, B.C., Zhao, H.: M2i: From factored marginal trajectory prediction to interactive prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6543-6552 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 398, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 398, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 398, + 480, + 430 + ], + "type": "text", + "content": "41. Suo, S., Regalado, S., Casas, S., Urtasun, R.: Trafficsim: Learning to simulate realistic multi-agent behaviors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10400-10409 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 432, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 432, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 133, + 432, + 480, + 485 + ], + "type": "text", + "content": "42. Varadarajan, B., Hefny, A., Srivastava, A., Refaat, K.S., Nayakanti, N., Cornman, A., Chen, K., Douillard, B., Lam, C.P., Anguelov, D., et al.: Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 7814-7821. IEEE (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 487, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 487, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 133, + 487, + 480, + 518 + ], + "type": "text", + "content": "43. Watson, D., Chan, W., Ho, J., Norouzi, M.: Learning fast samplers for diffusion models by differentiating through sample quality. arXiv preprint arXiv:2202.05830 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 520, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 520, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 133, + 520, + 480, + 552 + ], + "type": "text", + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 554, + 480, + 586 + ], + "type": "text", + "content": "45. Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929-2936. IEEE (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 587, + 480, + 609 + ], + "type": "text", + "content": "46. Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 610, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 610, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 133, + 610, + 480, + 642 + ], + "type": "text", + "content": "47. Zhao, H., Gao, J., Lan, T., Sun, C., Sapp, B., Varadarajan, B., Shen, Y., Shen, Y., Chai, Y., Schmid, C., et al.: Tnt: Target-driven trajectory prediction. In: Conference on Robot Learning. pp. 895-904. PMLR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "48. Zheng, H., He, P., Chen, W., Zhou, M.: Truncated diffusion probabilistic models. arXiv preprint arXiv:2202.09671 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 448, + 102 + ], + "type": "text", + "content": "OptTrajDiff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 217 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "49. Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560-3566. IEEE (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "50. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 217 + ], + "type": "text", + "content": "51. Zhou, Z., Wen, Z., Wang, J., Li, Y.H., Huang, Y.K.: Qcnext: A next-generation framework for joint multi-agent trajectory prediction (2023)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_content_list.json b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89b04725b2a13bcf55c458bdd82bdc63ce2bfd7c --- /dev/null +++ b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_content_list.json @@ -0,0 +1,1534 @@ +[ + { + "type": "text", + "text": "Optimizing Factorized Encoder Models: Time and Memory Reduction for Scalable and Efficient Action Recognition", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shreyank N Gowda1, Anurag Arnab2, and Jonathan Huang2", + "bbox": [ + 281, + 234, + 720, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Oxford", + "bbox": [ + 423, + 261, + 578, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Google Research", + "bbox": [ + 437, + 277, + 563, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we address the challenges posed by the substantial training time and memory consumption associated with video transformers, focusing on the ViViT (Video Vision Transformer) model, in particular the Factorised Encoder version, as our baseline for action recognition tasks. The factorised encoder variant follows the late-fusion approach that is adopted by many state of the art approaches. Despite standing out for its favorable speed/accuracy tradeoffs among the different variants of ViViT, its considerable training time and memory requirements still pose a significant barrier to entry. Our method is designed to lower this barrier and is based on the idea of freezing the spatial transformer during training. This leads to a low accuracy model if naively done. But we show that by (1) appropriately initializing the temporal transformer (a module responsible for processing temporal information) (2) introducing a compact adapter model connecting frozen spatial representations (a module that selectively focuses on regions of the input image) to the temporal transformer, we can enjoy the benefits of freezing the spatial transformer without sacrificing accuracy. Through extensive experimentation over 6 benchmarks, we demonstrate that our proposed training strategy significantly reduces training costs (by) and memory consumption while maintaining or slightly improving performance by up to $1.79\\%$ compared to the baseline model. Our approach additionally unlocks the capability to utilize larger image transformer models as our spatial transformer and access more frames with the same memory consumption. We also show the generalization of this approach to other factorized encoder models. The advancements made in this work have the potential to advance research in the video understanding domain and provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in the action recognition field.", + "bbox": [ + 261, + 323, + 743, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 750, + 375, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Action recognition focuses on understanding and identifying actions in video sequences with applications in surveillance, human-computer interaction, and video content analysis. The field has advanced significantly due to large-scale annotated datasets [5] and a shift from hand-crafted features [24, 41] to deep", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/605c40cbe22011785ca026fef4526aa157f4015630ea40ee6639e44b6c7af1af.jpg", + "image_caption": [ + "Fig. 1: Comparison of our initialization method vs conventional training of ViViT. Training time is scaled relative to setting ViViT-B training time to '1 unit' (7.93 hours). We see clear time saving using our initialization scheme and for larger models, the training time saved is much larger." + ], + "image_footnote": [], + "bbox": [ + 259, + 160, + 736, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learning models like convolutional networks (CNNs) [5, 13, 22, 36]. Recently, transformers have revolutionized computer vision by offering an alternative to traditional CNNs, leading to the development of many new state-of-the-art architectures [3, 10, 28]. Moreover, the flexibility of transformers have inspired researchers to adapt these models to more complex problems, including emotion recognition [12, 14], video understanding and action recognition [1, 29].", + "bbox": [ + 212, + 483, + 787, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers, however, are notoriously expensive [15], and Video Transformer-based architectures [1,2,29], which integrate information across space and time are even more so. And memory consumption and training times become even more significant when working with large-scale video datasets with long sequences [5]. These high computational costs present a particular challenge for researchers with limited resources, especially those from universities and smaller companies. The goal of our work is therefore to cut the cost of training: we want to train transformer-based video models with fewer resources or use larger model variants and handle more frames with the same resources.", + "bbox": [ + 212, + 579, + 789, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our primary focus is on a diverse range of factorized encoder models, among which ViViT [1] stands out as a prominent example. Specifically, we focus on the \"factorized encoder\" variant of ViViT which has separate spatial and temporal transformer stages, where the spatial transformer is responsible for extracting features from individual frames, while the temporal transformer processes the temporal dynamics across frames. We choose this factorized encoder design because it is more efficient compared to, e.g., the variant of ViViT using all-to-all spatiotemporal attention, while still achieving high accuracies and has thus been", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adopted as the building block for recent state-of-the-art architectures on various tasks [6, 19, 42, 46-48].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the challenge of reducing training time and memory usage without compromising the sophistication and accuracy of the original model, our approach is based on the simple idea of freezing the spatial backbone. Freezing the spatial backbone has many advantages: by not backpropagating through this transformer, training is faster and requires less memory (allowing for the model to handle more frames). We also inherit the benefits of pretraining the spatial transformer on a large dataset (such as JFT [20]). Naively implemented, however, we show that this approach falls very short in accuracy. Instead, with a few simple (but important) tweaks to the above idea, we propose a method that has the same advantages of freezing the spatial transformer, but does not compromise on accuracy.", + "bbox": [ + 212, + 176, + 785, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method proceeds in two stages. In the first stage we pretrain a cheap version of the model using fewer frames, e.g., 8 frames as opposed to, e.g., 32 frames. In the second stage we fine tune this model with more frames, which is more expensive but in this stage we freeze the spatial encoder and introduce a compact \"adapter\" model connecting frozen spatial representations to the temporal transformer, negating the need for end-to-end training of the spatial transformer. Critically this includes pre-training the temporal transformer (by initializing from stage 1) which is often overlooked in current video models which typically initialize this component from scratch. However our experiments show that this step is critical if we wish to not sacrifice performance.", + "bbox": [ + 212, + 343, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With our training recipe, we match or slightly outperform conventional training of ViViT at roughly half the cost as seen in Figure 1. A notable benefit of our training recipe, is its ability to process up to 80 frames (ViViT-B) on typical university-grade GPUs, a significant leap from the previous capacity of 16 frames. This expansion in processing power broadens the range of video data manageable under resource-constrained settings. As we elaborate in Section 4.7, our research underscores the potential to democratize access to advanced video transformer models. Another notable benefit is the model's ability to now use even larger models as the spatial transformer, we introduce ViViT-g as seen in Section 4.6. This accessibility paves the way for future video action recognition research, irrespective of resource constraints. We also show in Section 4.9 that our method generalizes to non-ViViT models such as Dualformer [26] and VideoSwin [29]. Hereafter, we refer to our version of ViViT as SFA-ViViT, where SFA denotes 'Spatial Frozen and Adapter Initialized'.", + "bbox": [ + 212, + 494, + 787, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 729, + 387, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformers for Videos Action recognition is a key research area in computer vision, addressed by many traditional [24, 41] and CNN based approaches [5, 16, 17, 21, 27, 36, 39, 44] aided by the release of large-scale datasets [5, 23, 37]. Since we focus on transformer based architectures, a thorough review of earlier methods are out of this scope. More recently, the transformer architecture, initially devel", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "oped for NLP tasks [40], has been adapted for video understanding and action recognition tasks, leading to state-of-the-art models such as TimeSformer [2], ViViT [1], VideoSwin [29], and Uniformer [25] These transformer-based models leverage self-attention mechanisms to capture complex spatiotemporal patterns in action recognition tasks. TimeSformer [2] is one of the first transformer-based models for video understanding, adapting the transformer architecture to video by treating it as a sequence of flattened image patches. ViViT [1] integrates spatial and temporal transformers to efficiently capture spatiotemporal information in video sequences. VideoSwin [29] is a hierarchical transformer that applies local windowing for efficiency, enabling the model to handle longer video sequences. Similar to VideoSwin, Dualformer [26] also adopts a hierarchical architecture but uses a local-global stratified transformer design to capture both fine-grained and global video features efficiently. More recently, Uniformer [25] integrates 3D convolution and spatiotemporal self-attention, MTV [47] proposes a multi-view transformer model using distinct encoders for each video \"view\", improving accuracy as the number of views increases. The Multiscale Vision Transformers (MViT) [11] model streamlines computation and memory usage by operating at different resolutions, focusing on high-level features at lower resolutions and low-level details at higher ones, effectively leveraging both spatial and temporal information in visual tasks. TubeViT [33] introduces a method of sparsely sampling different-sized 3D segments from videos, facilitating efficient joint image and video learning, and allowing the adaptation of larger models to videos with less computational resources. Typically these models have FLOPs in the range of TFLOPs and training times that last more than days on the largest of GPUs/TPUs available, making them infeasible to train or use in lower resourced settings such as academia. It is critical that we find a way to train these models with limited resources while maintaining their performance. To this end, we focus on the factorised encoder version of ViViT as the late-fusion approach followed is used as a foundation for state-of-the-art approaches of various tasks [6, 19, 42, 46–48] and hence believe that the initialization scheme proposed can be used for future methods working on similar architectures. Similar to the approach for ViViT factorized encoder, the hierarchical architectures of VideoSwin and DualFormer can also benefit from staged training. By freezing the early layers that capture local spatial-temporal features, and adding an adapter before fine-tuning the later layers focused on global context, the overall training can be made more efficient.", + "bbox": [ + 212, + 145, + 787, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Transformers in Videos Efficiency is a nuanced topic [9], as there are multiple cost indicators of efficiency (for example, GFLOPs, inference time, training time, memory usage), and models which improve efficiency in one dimension, are not necessarily better in other dimensions [9]. TokenLearner [35] proposes a method that adaptively learns tokens for efficient image and video understanding tasks, enabling effective modeling of pairwise attention over longer temporal horizons or spatial content. TokenLearner reduces the GFLOPs required by ViViT by about half, but does not significantly change the training time or the inference time of ViViT. Spatial Temporal Token Selection (STTS) [43] proposes", + "bbox": [ + 212, + 703, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a dynamic token selection framework for spatial and temporal dimensions that ranks token importance using a lightweight scorer network, selecting top-scoring tokens for downstream evaluation in an end-to-end training process. STTS again reduces the GFLOPs, but the training time and inference time do not change significantly. TokShift [49], a zero-parameter, zero-FLOPs operator that models temporal relations in transformer encoders by temporally shifting partial token features across adjacent frames but again requires the same training time as the original model. By densely integrating TokShift into a plain 2D vision transformer, a computationally efficient, convolution-free video transformer is created for video understanding. Most similar to our work is the ST-Adapter [32], that utilizes built-in spatio-temporal reasoning in a compact design, allowing pretrained image models to reason about dynamic video content with a small pertask parameter cost, surpassing existing methods in both parameter-efficiency and performance. However, it does not change FLOPs or inference time at all. Unlike ST-Adapter, we use a spatial only adapter which we show is enough to reproduce the performance of the baseline model at close to half the training time. In particular, our proposed method improves the training time and training memory usage, addressing the key problem of researchers and practitioners being able to train video models. It does not, however, change the inference time compared to a standard ViViT model. We consider overall train time for the same hyperparameters and use the same hardware for a direct comparison. We consider efficiency in this paper as the time saved in the overall training of the model.", + "bbox": [ + 217, + 146, + 787, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 515, + 380, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Revisiting ViViT", + "text_level": 1, + "bbox": [ + 215, + 544, + 401, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Video Vision Transformer (ViViT) extends the Vision Transformer architecture to handle video data by incorporating spatio-temporal reasoning. The idea behind ViViT is to process video input as a sequence of image patches, combining spatial and temporal information through a series of transformer layers, which include multi-head self-attention, layer normalization, and feed-forward networks. The output is used for video classification.", + "bbox": [ + 215, + 566, + 785, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the \"vanilla\" variant of ViViT, one extracts spatio-temporal tokens from a video then forwards all tokens through a transformer encoder which explicitly models all pairwise interactions between all spatio-temporal tokens. We build off of the more efficient \"Factorized Encoder\" variant of ViViT whose architecture consists of two separate transformer encoders, a spatial transformer modeling interactions between tokens from the same temporal index and a temporal transformer modeling interactions between tokens from different temporal indices. Despite having more parameters, it requires fewer floating point operations (FLOPs) than vanilla ViViT. Because the Factorised Encoder variant strikes a good balance point between accuracy and processing speed, it has also been adopted as the foundation for other architectures [6,19,42,46-48], reinforcing its utility and robustness.", + "bbox": [ + 215, + 659, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Our training strategy", + "text_level": 1, + "bbox": [ + 215, + 146, + 439, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We concentrate on the factorised encoder variant of ViViT as it is already the most efficient version of the baseline. Henceforth, when we talk about ViViT we refer to this variant of ViViT. Consider the ViViT model that contains a spatial transformer with parameters $\\theta_{\\text{spatial}}$ and a temporal transformer with parameters $\\theta_{\\text{temporal}}$ :", + "bbox": [ + 212, + 170, + 787, + 247 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\text {s p a t i a l}} = T _ {\\text {s p a t i a l}} \\left(X _ {\\text {i n}}; \\theta_ {\\text {s p a t i a l}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 258, + 785, + 273 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {o u t} = T _ {t e m p o r a l} (X _ {s p a t i a l}; \\theta_ {t e m p o r a l}).\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 277, + 643, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In conventional ViViT training, $\\theta_{\\text{spatial}}$ is initialized from an image pre-trained checkpoint such as ImageNet-21k [34] or JFT [20] and the $\\theta_{\\text{temporal}}$ is initialized from scratch. During backpropagation, the gradient flows through the entire model. This entails training two sizable transformer models end-to-end, which is a highly resource-intensive process, as the transformer architecture is inherently computationally demanding, especially with more frames and larger ViViT variants (e.g., ViViT-H).", + "bbox": [ + 212, + 303, + 782, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One approach to reducing training time is to freeze the parameters of the spatial transformer $\\theta_{\\text{spatial}}$ . By not backpropagating through $\\theta_{\\text{spatial}}$ , gradient updates are faster and require less memory, allowing us to access more frames without encountering out-of-memory issues. But as we show in experiments, the accuracy of the resulting model with frozen $\\theta_{\\text{spatial}}$ is not competitive (in accuracy) with the baseline training approach.", + "bbox": [ + 212, + 409, + 784, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present a two stage approach (see Fig. 2) to training ViViT models that inherits the same benefits of freezing the spatial transformer, while not compromising on model quality.", + "bbox": [ + 212, + 500, + 784, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 1. In Stage 1, we pretrain our ViViT model on a reduced number of frames initializing the spatial transformer using a pre-trained image checkpoint. We do not freeze the spatial transformer during this stage, but critically, Stage 1 serves to also initialize the temporal transformer.", + "bbox": [ + 212, + 559, + 782, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To set the number of frames at this stage, we must balance the goal of efficiency (using fewer frames) against our finding in experiments that pre-training on too few frames can lead to suboptimal results. In our ablations, we identify a sweet spot at 8 frames.", + "bbox": [ + 212, + 619, + 782, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 2. In Stage 2, we fine tune our ViViT model on the full frame count (e.g. 128 frames) initializing both spatial and temporal transformer parameters learned in Stage 1. Because this stage is significantly more expensive, in stage 2, we freeze the spatial transformer parameters $\\theta_{\\text{spatial}}$ and add a lightweight adapter module with parameters $\\theta_{\\text{adapter}}$ following the spatial transformer:", + "bbox": [ + 212, + 695, + 787, + 772 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {s p a t i a l} = T _ {s p a t i a l} (X _ {i n}; \\theta_ {s p a t i a l})\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 782, + 588, + 799 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\text {a d a p t e r}} = A _ {\\text {a d a p t e r}} \\left(X _ {\\text {s p a t i a l}}; \\theta_ {\\text {a d a p t e r}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 801, + 784, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {o u t} = T _ {t e m p o r a l} (X _ {a d a p t e r}; \\theta_ {t e m p o r a l})\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 820, + 645, + 837 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this setting, by backpropagating only through the temporal transformer and the lightweight adapter module (in our experiments, a two layer MLP), we effectively cut total training time by half.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9684e11d88c86e01c18153a848d98f9e34482da6431f64f0305b8a0bf38b378f.jpg", + "image_caption": [ + "Fig. 2: STAGE 1: We first use the full ViViT-FE model on 8 frames by initializing the spatial transformer from an image checkpoint and the temporal transformer from scratch. STAGE 2: We then use this as our checkpoint to initialize the spatial and temporal transformer for models using more frames (such as 32, 64 or 128). We then freeze the spatial transformer and add an adapter model to finetune spatial transformer features. The temporal transformer is finetuned from the same checkpoint." + ], + "image_footnote": [], + "bbox": [ + 295, + 215, + 707, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The crucial finding here is that the spatial transformer requires only short-term context for initialization (after which it remains frozen), whereas the temporal transformer necessitates long-term context to achieve its optimal performance. Further details and empirical analysis can be found in the next section.", + "bbox": [ + 212, + 526, + 787, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Analysis", + "text_level": 1, + "bbox": [ + 214, + 609, + 472, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Through a series of comprehensive experiments which we now present, we investigate the significance of the spatial transformer, examining the impact of pre-training datasets and how larger models affect action recognition performance. We also explore the importance of initializing the temporal transformer by employing various initialization schemes and datasets, assessing whether the number of frames is critical for initializing larger models, initializing full ViViT models, and initializing models on one dataset while fine-tuning on another.", + "bbox": [ + 212, + 641, + 787, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 768, + 334, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate on all the datasets considered in [1] (specifically, Kinetics-400 [5], Kinetics-600 [4], EPIC-Kitchens [8], Something-something v2 [18] and Moments-in-time [31]) as well as the Something-Else [30] dataset. As these datasets are", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "common in the community, we include further details in the supplementary along with the implementation details.", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 198, + 390, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first address two critical aspects: the significance of fine-tuning the spatial transformer and the importance of initializing the temporal transformer. To do so, we conduct a series of experiments in various scenarios, which are detailed below. Our analysis focuses on the Something-something dataset, utilizing the large version of the ViViT model, referred to as ViViT-L.", + "bbox": [ + 212, + 220, + 787, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examine four main elements that modify the structure of conventional ViViT training and these are mentioned with indices in Table 1 namely: I. The freezing of the spatial transformer ( $\\theta_{\\text{spatial}}$ is initialized and then frozen), II. The freezing of the temporal transformer ( $\\theta_{\\text{temporal}}$ is frozen), III. The addition of an adapter (lightweight module with parameters $\\theta_{\\text{adapter}}$ ), IV. Next, we initialize the temporal transformer using VideoMAE [38], while keeping the spatial transformer frozen and the adapter incorporated and V. The initialization of the temporal transformer ( $\\theta_{\\text{spatial}}$ and $\\theta_{\\text{temporal}}$ are initialized using the ViViT $_{8f}$ ).", + "bbox": [ + 212, + 297, + 787, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is important to note that the VideoMAE training is an extremely expensive process as can be seen in the table. But combined with the line below it, these two models, which significantly outperform lines I, II and III, show that properly initializing the temporal transformer is critical.", + "bbox": [ + 212, + 417, + 787, + 478 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bc52d5cf3ce247c3369e396ad575c897678c71af26708647128262c32e3cfe3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IndexSpatialFrozenAdapterTemporalFrozenTemporal InitTop-1AccTop-5AccTrain Time
-××××××64.4587.4814.17 h
I✓×××××27.7556.730.5x()
II××✓×××25.8053.070.5x()
III✓✓××××38.7768.930.53x()
IV✓✓×VMAE58.5485.832.51x()
V✓✓×ViViT-8f63.8587.620.62x()
VI×✓×ViViT-8f64.1387.351.37x()
", + "bbox": [ + 223, + 503, + 781, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1: Ablation study results illustrating the impact of various modifications to the ViViT-B model, including spatial and temporal transformer freezing, adapter addition, and initialization methods, on top-1 and top-5 accuracy. Dataset is Something-something v2.", + "bbox": [ + 212, + 612, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, initializing the spatial transformer yields further improvement. The adapter plays a vital role in augmenting performance when the spatial transformer is frozen, and due to its lightweight nature, it will be an essential component of our training methodology.", + "bbox": [ + 212, + 688, + 787, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 How many frames should we use for Stage 1?", + "text_level": 1, + "bbox": [ + 214, + 770, + 635, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next we experiment with various frame counts for stage 1 training, We test seven variants: JFT [20] checkpoint (image-based), 2, 4, 8, 16, 32, and 48-frame ViViT checkpoints. We then fine-tune these with a frozen spatial transformer and add", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d5c317ddb1f7b3a0aabbf997b8b620449519d929e656203aa8e93b4f0aafa1d.jpg", + "image_caption": [ + "Fig. 3: The effect of initializing with different numbers of frames (JFT, 2, 4, 8, 16, 32, and 48), freezing the spatial transformer and adding an adapter model and fine-tuning using 64, 96, and 128 frames. Results on Kinetics400 dataset, 'f' refers to frames." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 472, + 303 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fe08a86088348e9a74136a447c7c9de0bf4d4007e5f6187b2ffcf33f2efd0040.jpg", + "image_caption": [ + "Fig.4: Comparison of our initialization method vs conventional training of ViViT on Top-1 accuracy and loss on the Kinetics-400 dataset using 64 and 128 frames. We see that our initialization gives a significant headstart to the models." + ], + "image_footnote": [], + "bbox": [ + 514, + 150, + 746, + 294 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "an adapter model using 64, 96, and 128 frames (see Figure 3). Results show that using too few frames for Stage 1 training can underperform (with image-only initialization from a JFT [20] checkpoint performing the worst). Thus we deduce that short term temporal context is essential for initializing the spatial transformer. Performance also plateaus after 8 frames, and given that using more frames increases training time, we settle on using 8 frames as our \"sweet spot\" for Stage 1 training. We consider this an ablation, as this is done only on Kinetics and kept fixed for all models and datasets.", + "bbox": [ + 212, + 441, + 787, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Does the proposed training increase convergence speed?", + "text_level": 1, + "bbox": [ + 212, + 585, + 723, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Another potential question that may arise concerns the impact of this initialization method in terms of convergence speed, if any. This specific aspect holds considerable significance owing to its potential ability to drastically curtail the duration of time required for training and the number of epochs necessary to effectively train the model. Moreover, an important element to take into account is the effect of freezing the spatial transformer. This approach decreases the memory needed to store the model but also considerably enhances the training speed. To provide a clearer picture, we have plotted the validation curves with and without initialization, which can be seen in Figure 4. Note that with the proposed initialization, we get a significant head start in overall accuracy.", + "bbox": [ + 212, + 611, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 What about initializing on one dataset and finetuning on other?", + "text_level": 1, + "bbox": [ + 212, + 784, + 785, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our study, we find training an 8-frame version of the standard ViViT model affordable. We consider training this on a temporally dependent dataset like", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Something-something, then fine-tuning on other datasets like Kinetics-400. We examine two scenarios: first, the standard ViViT model trained on Kinetics-400 using 8-frames, and second, the same model trained on Something-something. Post-training, we freeze the spatial transformer, add an adapter, and fine-tune the models on the alternate dataset with more frames. We contrast this with models fine-tuned on their original datasets (see Table 2). Results favor initializing larger frame models from the 8-frame version on the same dataset. Thus, for the final comparison in Sec 4.8, we initialize models with 8-frame versions of the baseline.", + "bbox": [ + 212, + 146, + 787, + 280 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/60a2537233231d7a0faec47811afb5b6b9cad60246f39545051dbe77dac39e58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CheckpointSSv2K400
K400-init44.71/74.5382.81/93.98
SSv2-init63.85/87.6276.79/92.35
", + "bbox": [ + 241, + 282, + 527, + 349 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ac4e06bac995354864e24d6aa527455d3db9bf762d500cb7f49e40018c43495c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneTop-1Top-5Steps
ViViT-L79.6491.7348k steps
ViViT-H81.0293.0939k steps
ViViT-g81.8194.5529k steps
", + "bbox": [ + 535, + 282, + 779, + 354 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: A summary of cross-dataset initialization of the proposed model and performance comparison. We use Kinetics400 and Something-something v2 as our datasets.", + "bbox": [ + 238, + 358, + 500, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 3: A comparison of top-1 and top-5 accuracies for the ViViT-g model with the proposed training strategy, which incorporates a larger spatial transformer backbone. All models use 48 frames for fair comparison. Results are on Kinetics400 dataset.", + "bbox": [ + 521, + 366, + 782, + 462 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.6 Extending the image backbone to 1.5B parameters", + "text_level": 1, + "bbox": [ + 215, + 483, + 679, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "An intriguing consequence of our approach is the ability to incorporate larger backbones into the spatial transformer, made possible by the additional memory available to us as a result of freezing the spatial transformer during training. Consequently, we introduce ViViT-g, which integrates the ViT-g model (with 1.5B parameters) as its backbone. To ensure a fair comparison, we focus solely on training and inference using 48 frames, and abstain from employing multiview or multicrop testing. Our objective is to investigate the potential impact of a more substantial spatial transformer backbone on the overall performance and show the potential of larger spatial backbones that are possible due to our training process. It is essential to note that the full ViViT-g model could not process more than 8 frames due to memory limitations. However, our proposed strategy allows processing up to 48 frames. A comparison of accuracies is presented in Table 3 along with the number of steps needed to reach the best performance. Dataset used is Kinetics-400 and all the ViT checkpoints are JFT-pretrained [20].", + "bbox": [ + 212, + 507, + 787, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.7 Comparison of Memory Usage with Standard ViViT Training and Proposed Method", + "text_level": 1, + "bbox": [ + 214, + 739, + 772, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this study, we compare the number of frames that can be accessed using the standard ViViT training scheme against our proposed scheme, employing a set of 64 v3 TPUs that have 16 GB each. We further evaluate the performance of ViViT variants, including H, and g, in comparison with the SFA-ViViT using the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "same variant configurations. Maintaining identical hyperparameters, we ensure a local batch size of 1. Our findings indicate that the conventional ViViT training approach restricts frame accessibility to 96 frames for the ViViT-H model, and a mere 8 frames for the ViViT-g model, before reaching memory limitations. Conversely, our proposed method enables access to 128 frames for ViViT-H, and up to 48 frames when utilizing ViViT-g with the same hardware. Furthermore, we investigate the impact of utilizing university grade GPUs by conducting ViViT experiments on an NVIDIA RTX 2080 Ti GPU farm equipped with 8 GPUs having 12 GB each. Under these circumstances, ViViT can only process 16 frames using a local batch size of 1. However, our proposed training strategy enables a notable improvement, expanding the frame capacity to 80 frames (ViViT-B) helping us reproduce ViViT results on lower end GPUs. This enhancement provides a valuable opportunity for researchers with limited resources to attain performance levels comparable to those with extensive resources. We show a comparison of number of frames accessible with and without our training recipe in Table 4.", + "bbox": [ + 212, + 146, + 787, + 387 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4adfc635ed495e70e89c018e4ea64ab1d8e0e2fc5b525e58e6bcdae25bd21961.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model481632486496128
ViViT-H✓✓✓✓✓✓✓×
SFA-ViViT-H✓✓✓✓✓✓✓✓
ViViT-g✓✓××××××
SFA-ViViT-g✓✓✓✓✓×××
", + "bbox": [ + 372, + 411, + 630, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: Memory usage in different ViViT training schemes is compared using the Kinetics400 dataset on 64 TPUs v3 with 16GB memory each. A $\\checkmark$ indicates accessible frames given hardware constraints, while a $\\times$ signals an out-of-memory (OOM) error.", + "bbox": [ + 212, + 494, + 787, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.8 Comparison on all benchmarks to the baseline model", + "text_level": 1, + "bbox": [ + 214, + 604, + 699, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we present a comprehensive comparative analysis, focusing on the proposed approach and the baseline model. We report the Top-1 accuracy, Top-5 accuracy and the overall training time. The evaluation is conducted on the large and huge variants of ViViT across three datasets, namely Kinetics400, Kinetics600, and Moments in Time (MiT), with the summarized results tabulated in Table 5. The findings indicate a slight enhancement in accuracy for both Kinetics400 and Kinetics600 datasets, whereas a notable $1.79\\%$ increase in top-1 accuracy is observed for the MiT dataset using the proposed method. Furthermore, the proposed approach showcases a significant reduction in training time, accounting for approximately $56\\%$ of the original duration. This reduction emphasizes the advantageous nature of the proposed approach. To calculate the total training time for the SFA version, the train time of the 8 frame (Stage 1) ViViT model is combined with the train time of the (Stage 2) SFA-ViViT model. Conversely, the total training time for the standard ViViT encompasses", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the total train time for the same number of frames that SFA-ViViT is trained on for fair comparison.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c6b4ebaa83ad040738e284e4785b496332f88ec54c2baa6098bdcdb511f18b6c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelKinetics-400Kinetics-600Moments in Time
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f82.59/93.091x(21.57 h)83.29/95.821x(26.14 h)--
ViViT-L128f + SFA8f82.78/94.030.56x()83.47/95.290.56x()--
ViViT-H96f84.21/94.661x(56.71 h)84.18/95.681x(60.45 h)38.17/62.841x(110.79 h)
ViViT-H96f + SFA8f84.42/94.720.57x()84.39/96.200.57x()39.96/64.390.59x()
", + "bbox": [ + 222, + 204, + 781, + 282 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8d4be734e59d8fbc1a5dd60a936b51cb5e86714749c884823e9fe4f78f8a6912.jpg", + "table_caption": [ + "Table 5: Performance Comparison of various versions of ViViT with the proposed training strategy for Kinetics-400, Kinetics-600 and Moments in Time. Accuracies listed as Top-1/Top-5." + ], + "table_footnote": [], + "table_body": "
ModelSomething-somethingSomething-ElseEpic-Kitchens
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f64.45/87.481x(14.17 h)53.14/73.981x(3.84 h)43.53/56.55/65.401x(5.61 h)
ViViT-L128f + SFA8f63.85/87.620.62x()53.60/74.470.62x()43.54/56.78/65.160.63x()
", + "bbox": [ + 222, + 337, + 781, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Performance Comparison of ViViT-L with the proposed training strategy for Something-something v2, Something-Else and Epic-Kitchens. Accuracies listed as Top-1/Top-5, for Epic Kitchens Top-1 noun-verb/ Top-1 noun/ Top-1 Verb.", + "bbox": [ + 212, + 393, + 784, + 435 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further examine the performance of ViViT-L incorporating our proposed training strategy in comparison to the original version on three additional datasets: Something-something, Something-Else, and Epic-Kitchens. A consistent trend is observed, with the modified approach outperforming the baseline model, at only a $62\\%$ cost of the baseline training time. In summary, our proposed training strategy demonstrates promising potential by yielding comparable or slightly improved performance across all datasets. This is obtained while maintaining a training cost ranging from $56\\%$ to $62\\%$ of the original model, thus highlighting its effectiveness. Results can be seen in Table 6.", + "bbox": [ + 212, + 479, + 787, + 614 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.9 Beyond the ViViT model", + "text_level": 1, + "bbox": [ + 215, + 643, + 472, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We adapted our approach to tiny versions of Dualformer [26], VideoSwin [29] and Video-FocalNets [45] on the Kinetics 400 dataset, demonstrating up to $2\\mathrm{x}$ faster training with minimal accuracy loss. Thus, our method exhibits consistent improvements in training efficiency across multiple video architectures. We believe these comparisons better showcase the generalized benefits of our freeze-and-finetune approach for efficient video model training. These results can be seen in Table 7. While these models do not directly fall under the factorized encoder models, we can follow a similar pipeline to train them under the SFA training regime. For the Dualformer-T and VideoSwin-T, we first train 8-frame versions of these models and then freeze the first 3 blocks of the models (Stage 1 to Stage 3) and only finetune the last block.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/88fe4d8e721a36d6b2d93f582c1579d97f3a804b5bd46df2cbe12e2a822885f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTotal Train Time Accuracy
Dualformer-T32f1x(30.72 h)79.53
Dualformer-T32f-SFA8f0.57x()79.48
Dualformer-B32f1x(68.57 h)81.07
Dualformer-B32f-SFA8f0.59x()81.12
VideoSwin-T32f1x(41.45 h)78.62
VideoSwin-T32f-SFA8f0.52x()78.68
VideoSwin-B32f1x(115.23 h)80.45
VideoSwin-T32f-SFA8f0.53x()80.52
Video-FocalNet-T32f1x(11.28 h)79.76
Video-FocalNet-T32f-SFA8f0.52x()79.95
Video-FocalNet-B32f1x(28.84 h)83.55
Video-FocalNet-B32f-SFA8f0.51x()84.25
", + "bbox": [ + 300, + 143, + 702, + 339 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.10 Comparing with other efficient training strategies", + "text_level": 1, + "bbox": [ + 215, + 428, + 679, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "While we have already mentioned that strategies such as ST-Adapter, STTS etc are efficient in parameters and GFLOPs and not in training time. We demonstrate the results of training these in comparison to the proposed training method in terms of training time and inference time. The results demonstrate that our proposed SFA training pipeline confers substantial training time savings compared to these methods. For example, SFA-ViViT-B attains similar accuracy as ST-Adapter at approximately 0.58 times the total training time on Kinetics-400. We would like to point out that only STTS does better than the proposed strategy in terms of throughput, but as shown earlier requires twice the amount of time to train the model. All methods use 32 frames. These results are reported in Table 8.", + "bbox": [ + 212, + 460, + 787, + 625 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/fe067e5a51c6dcb0c88f3df7fd45cb4f6fbbafc3eae7ccef52cb19028f4f3198.jpg", + "table_caption": [ + "Table 7: Training efficiency improvements from SFA approach on tiny Dualformer and VideoSwin models. Our method achieves $>2\\mathrm{x}$ faster training with minimal accuracy drop in Kinetics-400." + ], + "table_footnote": [], + "table_body": "
ModelTotal Train Time ThroughputAccuracy
ViT-B + ST-Adapter [32]1x(34.18 hours)9082.0
ViViT-B [1]1.03x()8882.6
STTS [43]1.07x()12980.7
SBP [7]0.92x()8982.2
SFA8f-ViViT-B32f (Ours)0.58x()8882.8
", + "bbox": [ + 267, + 652, + 736, + 751 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 8: Training efficiency improvements from SFA approach in comparison to other efficient training strategies. We require approximately half of the training time in comparison to other efficient training strategies.", + "bbox": [ + 212, + 751, + 787, + 792 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Evaluations Discussion", + "text_level": 1, + "bbox": [ + 217, + 143, + 470, + 160 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our evaluations on ViViT was not to obtain SOTA performance using ViViT, but to show the reduced time required to train the model. In the ViViT paper [1], the reported results are different from what we report. This is due to some videos being removed from the Kinetics dataset. We re-run ViViT using the official codebase and report results on the datasets using this.", + "bbox": [ + 217, + 181, + 785, + 257 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The point of using 128 frames was not to show that using more frames improves performance, but instead to show that for larger models, using our training strategy we can actually access more frames due to memory being saved. The logic behind the training strategy holds for lower number of frames such as 32 as well. Hence, using multiview inference would still save time. We would still need only $0.6x$ of the overall training time. Also, as pointed out in Section 4.7, larger models such as ViViT-g can only access 8 frames using standard ViViT-FE training. However, using our approach, we can access 48 frames. A notable benefit of accessing more frames is in long-form video understanding tasks and while such experiments have not been conducted, this training approach promises improved results in the task.", + "bbox": [ + 217, + 258, + 785, + 424 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 454, + 356, + 469 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have investigated the challenges posed by the substantial training time and memory consumption of video transformers, particularly focusing on the factorised encoder variant of the ViViT model as our baseline. To address these challenges, we proposed two effective strategies: utilizing a compact adapter model for fine-tuning image representations instead of end-to-end training of the spatial transformer, and initializing the temporal transformer using the baseline model trained with 8 frames. Our proposed training strategy has demonstrated the potential to significantly reduce training costs and memory consumption while maintaining, or even slightly improving, performance compared to the baseline model. Furthermore, we observed that with proper initialization, our baseline model can achieve near-peak performance within the first $10\\%$ of training epochs. While we focus on the factorized encoder variant of ViViT, we also show the generalization power of the training regime by comparing results of DualFormer, VideoSwin and Video-FocalNets. We also show that the proposed method requires approximately half the time required in comparison to other efficient training strategies. The advancements made in this work could help research in the video understanding domain by enabling access to more frames and the utilization of larger image models as the spatial transformer, all while maintaining the same memory consumption. Our findings provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in action recognition. Future work may focus on further optimizing and refining these strategies, and exploring their application to other video transformer architectures.", + "bbox": [ + 217, + 492, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)", + "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020)", + "4. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. arXiv preprint arXiv:1808.01340 (2018)", + "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "6. Chen, J., Ho, C.M.: Mm-vit: Multi-modal video transformer for compressed video action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1910-1921 (2022)", + "7. Cheng, F., Xu, M., Xiong, Y., Chen, H., Li, X., Li, W., Xia, W.: Stochastic backpropagation: A memory efficient strategy for training video models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8301-8310 (2022)", + "8. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 720-736 (2018)", + "9. Dehghani, M., Arnab, A., Beyer, L., Vaswani, A., Tay, Y.: The efficiency misnomer. arXiv preprint arXiv:2110.12894 (2021)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021)", + "2. Foteinopoulou, N.M., Patras, I.: Emoclip: A vision-language method for zero-shot video facial expression recognition. arXiv preprint arXiv:2310.16640 (2023)", + "3. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017)", + "4. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024)", + "5. Gowda, S.N., Hao, X., Li, G., Sevilla-Lara, L., Gowda, S.N.: Watt for what: Rethinking deep learning's energy-performance relationship. arXiv preprint arXiv:2310.06522 (2023)", + "6. Gowda, S.N., Rohrbach, M., Keller, F., Sevilla-Lara, L.: Learn2augment: Learning to composite videos for data augmentation in action recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 242-259. Springer (2022)" + ], + "bbox": [ + 225, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021)", + "18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)", + "19. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lučić, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023)", + "20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "21. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012)", + "22. Kim, K., Gowda, S.N., Mac Aodha, O., Sevilla-Lara, L.: Capturing temporal information in a single frame: Channel sampling strategies for action recognition. arXiv preprint arXiv:2201.10394 (2022)", + "23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)", + "24. Laptev, I.: On space-time interest points. International journal of computer vision 64, 107-123 (2005)", + "25. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022)", + "26. Liang, Y., Zhou, P., Zimmermann, R., Yan, S.: Dualformer: Local-global stratified transformer for efficient video recognition. In: European Conference on Computer Vision. pp. 577-595. Springer (2022)", + "27. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)", + "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "29. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)", + "30. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020)", + "31. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019)", + "32. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Shreyank N Gowda et al.", + "bbox": [ + 271, + 114, + 439, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. arXiv preprint arXiv:2212.03229 (2022)", + "34. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021)", + "35. Ryoo, M., Piergiovanni, A., Arnab, A., Dehghani, M., Angelova, A.: Tokenlearner: Adaptive space-time tokenization for videos. Advances in Neural Information Processing Systems 34, 12786-12797 (2021)", + "36. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)", + "37. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)", + "38. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602 (2022)", + "39. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)", + "40. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "41. Wang, H., Kläser, A., Schmid, C., Liu, C.L.: Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision 103, 60-79 (2013)", + "42. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. Transactions of Machine Learning Research", + "43. Wang, J., Yang, X., Li, H., Liu, L., Wu, Z., Jiang, Y.G.: Efficient video transformers with spatial-temporal token selection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 69-86. Springer (2022)", + "44. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)", + "45. Wasim, S.T., Khattak, M.U., Naseer, M., Khan, S., Shah, M., Khan, F.S.: Videofocalnets: Spatio-temporal focal modulation for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13778-13789 (2023)", + "46. Xiong, X., Arnab, A., Nagrani, A., Schmid, C.: M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852 (2022)", + "47. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022)", + "48. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "49. Zhang, H., Hao, Y., Ngo, C.W.: Token shift transformer for video classification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 917-925 (2021)" + ], + "bbox": [ + 212, + 146, + 785, + 825 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_model.json b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..862e4366b0ff36588655faee3bd95adedf781d42 --- /dev/null +++ b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_model.json @@ -0,0 +1,2049 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.21 + ], + "angle": 0, + "content": "Optimizing Factorized Encoder Models: Time and Memory Reduction for Scalable and Efficient Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.235, + 0.721, + 0.252 + ], + "angle": 0, + "content": "Shreyank N Gowda1, Anurag Arnab2, and Jonathan Huang2" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.262, + 0.58, + 0.277 + ], + "angle": 0, + "content": "1 University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.278, + 0.565, + 0.291 + ], + "angle": 0, + "content": "2 Google Research" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.324, + 0.744, + 0.726 + ], + "angle": 0, + "content": "Abstract. In this paper, we address the challenges posed by the substantial training time and memory consumption associated with video transformers, focusing on the ViViT (Video Vision Transformer) model, in particular the Factorised Encoder version, as our baseline for action recognition tasks. The factorised encoder variant follows the late-fusion approach that is adopted by many state of the art approaches. Despite standing out for its favorable speed/accuracy tradeoffs among the different variants of ViViT, its considerable training time and memory requirements still pose a significant barrier to entry. Our method is designed to lower this barrier and is based on the idea of freezing the spatial transformer during training. This leads to a low accuracy model if naively done. But we show that by (1) appropriately initializing the temporal transformer (a module responsible for processing temporal information) (2) introducing a compact adapter model connecting frozen spatial representations (a module that selectively focuses on regions of the input image) to the temporal transformer, we can enjoy the benefits of freezing the spatial transformer without sacrificing accuracy. Through extensive experimentation over 6 benchmarks, we demonstrate that our proposed training strategy significantly reduces training costs (by) and memory consumption while maintaining or slightly improving performance by up to \\(1.79\\%\\) compared to the baseline model. Our approach additionally unlocks the capability to utilize larger image transformer models as our spatial transformer and access more frames with the same memory consumption. We also show the generalization of this approach to other factorized encoder models. The advancements made in this work have the potential to advance research in the video understanding domain and provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in the action recognition field." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.75, + 0.377, + 0.767 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Action recognition focuses on understanding and identifying actions in video sequences with applications in surveillance, human-computer interaction, and video content analysis. The field has advanced significantly due to large-scale annotated datasets [5] and a shift from hand-crafted features [24, 41] to deep" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.161, + 0.737, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Fig. 1: Comparison of our initialization method vs conventional training of ViViT. Training time is scaled relative to setting ViViT-B training time to '1 unit' (7.93 hours). We see clear time saving using our initialization scheme and for larger models, the training time saved is much larger." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.788, + 0.577 + ], + "angle": 0, + "content": "learning models like convolutional networks (CNNs) [5, 13, 22, 36]. Recently, transformers have revolutionized computer vision by offering an alternative to traditional CNNs, leading to the development of many new state-of-the-art architectures [3, 10, 28]. Moreover, the flexibility of transformers have inspired researchers to adapt these models to more complex problems, including emotion recognition [12, 14], video understanding and action recognition [1, 29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.79, + 0.716 + ], + "angle": 0, + "content": "Transformers, however, are notoriously expensive [15], and Video Transformer-based architectures [1,2,29], which integrate information across space and time are even more so. And memory consumption and training times become even more significant when working with large-scale video datasets with long sequences [5]. These high computational costs present a particular challenge for researchers with limited resources, especially those from universities and smaller companies. The goal of our work is therefore to cut the cost of training: we want to train transformer-based video models with fewer resources or use larger model variants and handle more frames with the same resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our primary focus is on a diverse range of factorized encoder models, among which ViViT [1] stands out as a prominent example. Specifically, we focus on the \"factorized encoder\" variant of ViViT which has separate spatial and temporal transformer stages, where the spatial transformer is responsible for extracting features from individual frames, while the temporal transformer processes the temporal dynamics across frames. We choose this factorized encoder design because it is more efficient compared to, e.g., the variant of ViViT using all-to-all spatiotemporal attention, while still achieving high accuracies and has thus been" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "adopted as the building block for recent state-of-the-art architectures on various tasks [6, 19, 42, 46-48]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.344 + ], + "angle": 0, + "content": "To address the challenge of reducing training time and memory usage without compromising the sophistication and accuracy of the original model, our approach is based on the simple idea of freezing the spatial backbone. Freezing the spatial backbone has many advantages: by not backpropagating through this transformer, training is faster and requires less memory (allowing for the model to handle more frames). We also inherit the benefits of pretraining the spatial transformer on a large dataset (such as JFT [20]). Naively implemented, however, we show that this approach falls very short in accuracy. Instead, with a few simple (but important) tweaks to the above idea, we propose a method that has the same advantages of freezing the spatial transformer, but does not compromise on accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Our method proceeds in two stages. In the first stage we pretrain a cheap version of the model using fewer frames, e.g., 8 frames as opposed to, e.g., 32 frames. In the second stage we fine tune this model with more frames, which is more expensive but in this stage we freeze the spatial encoder and introduce a compact \"adapter\" model connecting frozen spatial representations to the temporal transformer, negating the need for end-to-end training of the spatial transformer. Critically this includes pre-training the temporal transformer (by initializing from stage 1) which is often overlooked in current video models which typically initialize this component from scratch. However our experiments show that this step is critical if we wish to not sacrifice performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.707 + ], + "angle": 0, + "content": "With our training recipe, we match or slightly outperform conventional training of ViViT at roughly half the cost as seen in Figure 1. A notable benefit of our training recipe, is its ability to process up to 80 frames (ViViT-B) on typical university-grade GPUs, a significant leap from the previous capacity of 16 frames. This expansion in processing power broadens the range of video data manageable under resource-constrained settings. As we elaborate in Section 4.7, our research underscores the potential to democratize access to advanced video transformer models. Another notable benefit is the model's ability to now use even larger models as the spatial transformer, we introduce ViViT-g as seen in Section 4.6. This accessibility paves the way for future video action recognition research, irrespective of resource constraints. We also show in Section 4.9 that our method generalizes to non-ViViT models such as Dualformer [26] and VideoSwin [29]. Hereafter, we refer to our version of ViViT as SFA-ViViT, where SFA denotes 'Spatial Frozen and Adapter Initialized'." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.388, + 0.748 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Transformers for Videos Action recognition is a key research area in computer vision, addressed by many traditional [24, 41] and CNN based approaches [5, 16, 17, 21, 27, 36, 39, 44] aided by the release of large-scale datasets [5, 23, 37]. Since we focus on transformer based architectures, a thorough review of earlier methods are out of this scope. More recently, the transformer architecture, initially devel" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.692 + ], + "angle": 0, + "content": "oped for NLP tasks [40], has been adapted for video understanding and action recognition tasks, leading to state-of-the-art models such as TimeSformer [2], ViViT [1], VideoSwin [29], and Uniformer [25] These transformer-based models leverage self-attention mechanisms to capture complex spatiotemporal patterns in action recognition tasks. TimeSformer [2] is one of the first transformer-based models for video understanding, adapting the transformer architecture to video by treating it as a sequence of flattened image patches. ViViT [1] integrates spatial and temporal transformers to efficiently capture spatiotemporal information in video sequences. VideoSwin [29] is a hierarchical transformer that applies local windowing for efficiency, enabling the model to handle longer video sequences. Similar to VideoSwin, Dualformer [26] also adopts a hierarchical architecture but uses a local-global stratified transformer design to capture both fine-grained and global video features efficiently. More recently, Uniformer [25] integrates 3D convolution and spatiotemporal self-attention, MTV [47] proposes a multi-view transformer model using distinct encoders for each video \"view\", improving accuracy as the number of views increases. The Multiscale Vision Transformers (MViT) [11] model streamlines computation and memory usage by operating at different resolutions, focusing on high-level features at lower resolutions and low-level details at higher ones, effectively leveraging both spatial and temporal information in visual tasks. TubeViT [33] introduces a method of sparsely sampling different-sized 3D segments from videos, facilitating efficient joint image and video learning, and allowing the adaptation of larger models to videos with less computational resources. Typically these models have FLOPs in the range of TFLOPs and training times that last more than days on the largest of GPUs/TPUs available, making them infeasible to train or use in lower resourced settings such as academia. It is critical that we find a way to train these models with limited resources while maintaining their performance. To this end, we focus on the factorised encoder version of ViViT as the late-fusion approach followed is used as a foundation for state-of-the-art approaches of various tasks [6, 19, 42, 46–48] and hence believe that the initialization scheme proposed can be used for future methods working on similar architectures. Similar to the approach for ViViT factorized encoder, the hierarchical architectures of VideoSwin and DualFormer can also benefit from staged training. By freezing the early layers that capture local spatial-temporal features, and adding an adapter before fine-tuning the later layers focused on global context, the overall training can be made more efficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Efficient Transformers in Videos Efficiency is a nuanced topic [9], as there are multiple cost indicators of efficiency (for example, GFLOPs, inference time, training time, memory usage), and models which improve efficiency in one dimension, are not necessarily better in other dimensions [9]. TokenLearner [35] proposes a method that adaptively learns tokens for efficient image and video understanding tasks, enabling effective modeling of pairwise attention over longer temporal horizons or spatial content. TokenLearner reduces the GFLOPs required by ViViT by about half, but does not significantly change the training time or the inference time of ViViT. Spatial Temporal Token Selection (STTS) [43] proposes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.493 + ], + "angle": 0, + "content": "a dynamic token selection framework for spatial and temporal dimensions that ranks token importance using a lightweight scorer network, selecting top-scoring tokens for downstream evaluation in an end-to-end training process. STTS again reduces the GFLOPs, but the training time and inference time do not change significantly. TokShift [49], a zero-parameter, zero-FLOPs operator that models temporal relations in transformer encoders by temporally shifting partial token features across adjacent frames but again requires the same training time as the original model. By densely integrating TokShift into a plain 2D vision transformer, a computationally efficient, convolution-free video transformer is created for video understanding. Most similar to our work is the ST-Adapter [32], that utilizes built-in spatio-temporal reasoning in a compact design, allowing pretrained image models to reason about dynamic video content with a small pertask parameter cost, surpassing existing methods in both parameter-efficiency and performance. However, it does not change FLOPs or inference time at all. Unlike ST-Adapter, we use a spatial only adapter which we show is enough to reproduce the performance of the baseline model at close to half the training time. In particular, our proposed method improves the training time and training memory usage, addressing the key problem of researchers and practitioners being able to train video models. It does not, however, change the inference time compared to a standard ViViT model. We consider overall train time for the same hyperparameters and use the same hardware for a direct comparison. We consider efficiency in this paper as the time saved in the overall training of the model." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.516, + 0.381, + 0.534 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.545, + 0.403, + 0.56 + ], + "angle": 0, + "content": "3.1 Revisiting ViViT" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.568, + 0.787, + 0.658 + ], + "angle": 0, + "content": "The Video Vision Transformer (ViViT) extends the Vision Transformer architecture to handle video data by incorporating spatio-temporal reasoning. The idea behind ViViT is to process video input as a sequence of image patches, combining spatial and temporal information through a series of transformer layers, which include multi-head self-attention, layer normalization, and feed-forward networks. The output is used for video classification." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.66, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the \"vanilla\" variant of ViViT, one extracts spatio-temporal tokens from a video then forwards all tokens through a transformer encoder which explicitly models all pairwise interactions between all spatio-temporal tokens. We build off of the more efficient \"Factorized Encoder\" variant of ViViT whose architecture consists of two separate transformer encoders, a spatial transformer modeling interactions between tokens from the same temporal index and a temporal transformer modeling interactions between tokens from different temporal indices. Despite having more parameters, it requires fewer floating point operations (FLOPs) than vanilla ViViT. Because the Factorised Encoder variant strikes a good balance point between accuracy and processing speed, it has also been adopted as the foundation for other architectures [6,19,42,46-48], reinforcing its utility and robustness." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.44, + 0.163 + ], + "angle": 0, + "content": "3.2 Our training strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.248 + ], + "angle": 0, + "content": "We concentrate on the factorised encoder variant of ViViT as it is already the most efficient version of the baseline. Henceforth, when we talk about ViViT we refer to this variant of ViViT. Consider the ViViT model that contains a spatial transformer with parameters \\(\\theta_{\\text{spatial}}\\) and a temporal transformer with parameters \\(\\theta_{\\text{temporal}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.259, + 0.786, + 0.274 + ], + "angle": 0, + "content": "\\[\nX _ {\\text {s p a t i a l}} = T _ {\\text {s p a t i a l}} \\left(X _ {\\text {i n}}; \\theta_ {\\text {s p a t i a l}}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.278, + 0.644, + 0.294 + ], + "angle": 0, + "content": "\\[\nX _ {o u t} = T _ {t e m p o r a l} (X _ {s p a t i a l}; \\theta_ {t e m p o r a l}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.784, + 0.409 + ], + "angle": 0, + "content": "In conventional ViViT training, \\(\\theta_{\\text{spatial}}\\) is initialized from an image pre-trained checkpoint such as ImageNet-21k [34] or JFT [20] and the \\(\\theta_{\\text{temporal}}\\) is initialized from scratch. During backpropagation, the gradient flows through the entire model. This entails training two sizable transformer models end-to-end, which is a highly resource-intensive process, as the transformer architecture is inherently computationally demanding, especially with more frames and larger ViViT variants (e.g., ViViT-H)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.501 + ], + "angle": 0, + "content": "One approach to reducing training time is to freeze the parameters of the spatial transformer \\(\\theta_{\\text{spatial}}\\). By not backpropagating through \\(\\theta_{\\text{spatial}}\\), gradient updates are faster and require less memory, allowing us to access more frames without encountering out-of-memory issues. But as we show in experiments, the accuracy of the resulting model with frozen \\(\\theta_{\\text{spatial}}\\) is not competitive (in accuracy) with the baseline training approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.785, + 0.547 + ], + "angle": 0, + "content": "We present a two stage approach (see Fig. 2) to training ViViT models that inherits the same benefits of freezing the spatial transformer, while not compromising on model quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.784, + 0.62 + ], + "angle": 0, + "content": "Stage 1. In Stage 1, we pretrain our ViViT model on a reduced number of frames initializing the spatial transformer using a pre-trained image checkpoint. We do not freeze the spatial transformer during this stage, but critically, Stage 1 serves to also initialize the temporal transformer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.784, + 0.681 + ], + "angle": 0, + "content": "To set the number of frames at this stage, we must balance the goal of efficiency (using fewer frames) against our finding in experiments that pre-training on too few frames can lead to suboptimal results. In our ablations, we identify a sweet spot at 8 frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Stage 2. In Stage 2, we fine tune our ViViT model on the full frame count (e.g. 128 frames) initializing both spatial and temporal transformer parameters learned in Stage 1. Because this stage is significantly more expensive, in stage 2, we freeze the spatial transformer parameters \\(\\theta_{\\text{spatial}}\\) and add a lightweight adapter module with parameters \\(\\theta_{\\text{adapter}}\\) following the spatial transformer:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.784, + 0.589, + 0.8 + ], + "angle": 0, + "content": "\\[\nX _ {s p a t i a l} = T _ {s p a t i a l} (X _ {i n}; \\theta_ {s p a t i a l})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.802, + 0.785, + 0.818 + ], + "angle": 0, + "content": "\\[\nX _ {\\text {a d a p t e r}} = A _ {\\text {a d a p t e r}} \\left(X _ {\\text {s p a t i a l}}; \\theta_ {\\text {a d a p t e r}}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.821, + 0.647, + 0.838 + ], + "angle": 0, + "content": "\\[\nX _ {o u t} = T _ {t e m p o r a l} (X _ {a d a p t e r}; \\theta_ {t e m p o r a l})\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "In this setting, by backpropagating only through the temporal transformer and the lightweight adapter module (in our experiments, a two layer MLP), we effectively cut total training time by half." + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.217, + 0.709, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.498 + ], + "angle": 0, + "content": "Fig. 2: STAGE 1: We first use the full ViViT-FE model on 8 frames by initializing the spatial transformer from an image checkpoint and the temporal transformer from scratch. STAGE 2: We then use this as our checkpoint to initialize the spatial and temporal transformer for models using more frames (such as 32, 64 or 128). We then freeze the spatial transformer and add an adapter model to finetune spatial transformer features. The temporal transformer is finetuned from the same checkpoint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.589 + ], + "angle": 0, + "content": "The crucial finding here is that the spatial transformer requires only short-term context for initialization (after which it remains frozen), whereas the temporal transformer necessitates long-term context to achieve its optimal performance. Further details and empirical analysis can be found in the next section." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.61, + 0.473, + 0.628 + ], + "angle": 0, + "content": "4 Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Through a series of comprehensive experiments which we now present, we investigate the significance of the spatial transformer, examining the impact of pre-training datasets and how larger models affect action recognition performance. We also explore the importance of initializing the temporal transformer by employing various initialization schemes and datasets, assessing whether the number of frames is critical for initializing larger models, initializing full ViViT models, and initializing models on one dataset while fine-tuning on another." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.335, + 0.785 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "We evaluate on all the datasets considered in [1] (specifically, Kinetics-400 [5], Kinetics-600 [4], EPIC-Kitchens [8], Something-something v2 [18] and Moments-in-time [31]) as well as the Something-Else [30] dataset. As these datasets are" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "common in the community, we include further details in the supplementary along with the implementation details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.199, + 0.391, + 0.214 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.298 + ], + "angle": 0, + "content": "We first address two critical aspects: the significance of fine-tuning the spatial transformer and the importance of initializing the temporal transformer. To do so, we conduct a series of experiments in various scenarios, which are detailed below. Our analysis focuses on the Something-something dataset, utilizing the large version of the ViViT model, referred to as ViViT-L." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.419 + ], + "angle": 0, + "content": "We examine four main elements that modify the structure of conventional ViViT training and these are mentioned with indices in Table 1 namely: I. The freezing of the spatial transformer (\\(\\theta_{\\text{spatial}}\\) is initialized and then frozen), II. The freezing of the temporal transformer (\\(\\theta_{\\text{temporal}}\\) is frozen), III. The addition of an adapter (lightweight module with parameters \\(\\theta_{\\text{adapter}}\\)), IV. Next, we initialize the temporal transformer using VideoMAE [38], while keeping the spatial transformer frozen and the adapter incorporated and V. The initialization of the temporal transformer (\\(\\theta_{\\text{spatial}}\\) and \\(\\theta_{\\text{temporal}}\\) are initialized using the ViViT\\(_{8f}\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.479 + ], + "angle": 0, + "content": "It is important to note that the VideoMAE training is an extremely expensive process as can be seen in the table. But combined with the line below it, these two models, which significantly outperform lines I, II and III, show that properly initializing the temporal transformer is critical." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.504, + 0.782, + 0.612 + ], + "angle": 0, + "content": "
IndexSpatialFrozenAdapterTemporalFrozenTemporal InitTop-1AccTop-5AccTrain Time
-××××××64.4587.4814.17 h
I✓×××××27.7556.730.5x()
II××✓×××25.8053.070.5x()
III✓✓××××38.7768.930.53x()
IV✓✓×VMAE58.5485.832.51x()
V✓✓×ViViT-8f63.8587.620.62x()
VI×✓×ViViT-8f64.1387.351.37x()
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Table 1: Ablation study results illustrating the impact of various modifications to the ViViT-B model, including spatial and temporal transformer freezing, adapter addition, and initialization methods, on top-1 and top-5 accuracy. Dataset is Something-something v2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Additionally, initializing the spatial transformer yields further improvement. The adapter plays a vital role in augmenting performance when the spatial transformer is frozen, and due to its lightweight nature, it will be an essential component of our training methodology." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.637, + 0.787 + ], + "angle": 0, + "content": "4.3 How many frames should we use for Stage 1?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Next we experiment with various frame counts for stage 1 training, We test seven variants: JFT [20] checkpoint (image-based), 2, 4, 8, 16, 32, and 48-frame ViViT checkpoints. We then fine-tune these with a frozen spatial transformer and add" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.473, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.315, + 0.479, + 0.414 + ], + "angle": 0, + "content": "Fig. 3: The effect of initializing with different numbers of frames (JFT, 2, 4, 8, 16, 32, and 48), freezing the spatial transformer and adding an adapter model and fine-tuning using 64, 96, and 128 frames. Results on Kinetics400 dataset, 'f' refers to frames." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.151, + 0.747, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.311, + 0.761, + 0.407 + ], + "angle": 0, + "content": "Fig.4: Comparison of our initialization method vs conventional training of ViViT on Top-1 accuracy and loss on the Kinetics-400 dataset using 64 and 128 frames. We see that our initialization gives a significant headstart to the models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.565 + ], + "angle": 0, + "content": "an adapter model using 64, 96, and 128 frames (see Figure 3). Results show that using too few frames for Stage 1 training can underperform (with image-only initialization from a JFT [20] checkpoint performing the worst). Thus we deduce that short term temporal context is essential for initializing the spatial transformer. Performance also plateaus after 8 frames, and given that using more frames increases training time, we settle on using 8 frames as our \"sweet spot\" for Stage 1 training. We consider this an ablation, as this is done only on Kinetics and kept fixed for all models and datasets." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.587, + 0.725, + 0.604 + ], + "angle": 0, + "content": "4.4 Does the proposed training increase convergence speed?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Another potential question that may arise concerns the impact of this initialization method in terms of convergence speed, if any. This specific aspect holds considerable significance owing to its potential ability to drastically curtail the duration of time required for training and the number of epochs necessary to effectively train the model. Moreover, an important element to take into account is the effect of freezing the spatial transformer. This approach decreases the memory needed to store the model but also considerably enhances the training speed. To provide a clearer picture, we have plotted the validation curves with and without initialization, which can be seen in Figure 4. Note that with the proposed initialization, we get a significant head start in overall accuracy." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.785, + 0.787, + 0.801 + ], + "angle": 0, + "content": "4.5 What about initializing on one dataset and finetuning on other?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In our study, we find training an 8-frame version of the standard ViViT model affordable. We consider training this on a temporally dependent dataset like" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.281 + ], + "angle": 0, + "content": "Something-something, then fine-tuning on other datasets like Kinetics-400. We examine two scenarios: first, the standard ViViT model trained on Kinetics-400 using 8-frames, and second, the same model trained on Something-something. Post-training, we freeze the spatial transformer, add an adapter, and fine-tune the models on the alternate dataset with more frames. We contrast this with models fine-tuned on their original datasets (see Table 2). Results favor initializing larger frame models from the 8-frame version on the same dataset. Thus, for the final comparison in Sec 4.8, we initialize models with 8-frame versions of the baseline." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.283, + 0.528, + 0.35 + ], + "angle": 0, + "content": "
CheckpointSSv2K400
K400-init44.71/74.5382.81/93.98
SSv2-init63.85/87.6276.79/92.35
" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.283, + 0.78, + 0.356 + ], + "angle": 0, + "content": "
BackboneTop-1Top-5Steps
ViViT-L79.6491.7348k steps
ViViT-H81.0293.0939k steps
ViViT-g81.8194.5529k steps
" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.359, + 0.501, + 0.428 + ], + "angle": 0, + "content": "Table 2: A summary of cross-dataset initialization of the proposed model and performance comparison. We use Kinetics400 and Something-something v2 as our datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.367, + 0.784, + 0.463 + ], + "angle": 0, + "content": "Table 3: A comparison of top-1 and top-5 accuracies for the ViViT-g model with the proposed training strategy, which incorporates a larger spatial transformer backbone. All models use 48 frames for fair comparison. Results are on Kinetics400 dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.681, + 0.5 + ], + "angle": 0, + "content": "4.6 Extending the image backbone to 1.5B parameters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.721 + ], + "angle": 0, + "content": "An intriguing consequence of our approach is the ability to incorporate larger backbones into the spatial transformer, made possible by the additional memory available to us as a result of freezing the spatial transformer during training. Consequently, we introduce ViViT-g, which integrates the ViT-g model (with 1.5B parameters) as its backbone. To ensure a fair comparison, we focus solely on training and inference using 48 frames, and abstain from employing multiview or multicrop testing. Our objective is to investigate the potential impact of a more substantial spatial transformer backbone on the overall performance and show the potential of larger spatial backbones that are possible due to our training process. It is essential to note that the full ViViT-g model could not process more than 8 frames due to memory limitations. However, our proposed strategy allows processing up to 48 frames. A comparison of accuracies is presented in Table 3 along with the number of steps needed to reach the best performance. Dataset used is Kinetics-400 and all the ViT checkpoints are JFT-pretrained [20]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.773, + 0.772 + ], + "angle": 0, + "content": "4.7 Comparison of Memory Usage with Standard ViViT Training and Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this study, we compare the number of frames that can be accessed using the standard ViViT training scheme against our proposed scheme, employing a set of 64 v3 TPUs that have 16 GB each. We further evaluate the performance of ViViT variants, including H, and g, in comparison with the SFA-ViViT using the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.388 + ], + "angle": 0, + "content": "same variant configurations. Maintaining identical hyperparameters, we ensure a local batch size of 1. Our findings indicate that the conventional ViViT training approach restricts frame accessibility to 96 frames for the ViViT-H model, and a mere 8 frames for the ViViT-g model, before reaching memory limitations. Conversely, our proposed method enables access to 128 frames for ViViT-H, and up to 48 frames when utilizing ViViT-g with the same hardware. Furthermore, we investigate the impact of utilizing university grade GPUs by conducting ViViT experiments on an NVIDIA RTX 2080 Ti GPU farm equipped with 8 GPUs having 12 GB each. Under these circumstances, ViViT can only process 16 frames using a local batch size of 1. However, our proposed training strategy enables a notable improvement, expanding the frame capacity to 80 frames (ViViT-B) helping us reproduce ViViT results on lower end GPUs. This enhancement provides a valuable opportunity for researchers with limited resources to attain performance levels comparable to those with extensive resources. We show a comparison of number of frames accessible with and without our training recipe in Table 4." + }, + { + "type": "table", + "bbox": [ + 0.373, + 0.412, + 0.632, + 0.496 + ], + "angle": 0, + "content": "
Model481632486496128
ViViT-H✓✓✓✓✓✓✓×
SFA-ViViT-H✓✓✓✓✓✓✓✓
ViViT-g✓✓××××××
SFA-ViViT-g✓✓✓✓✓×××
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.539 + ], + "angle": 0, + "content": "Table 4: Memory usage in different ViViT training schemes is compared using the Kinetics400 dataset on 64 TPUs v3 with 16GB memory each. A \\(\\checkmark\\) indicates accessible frames given hardware constraints, while a \\(\\times\\) signals an out-of-memory (OOM) error." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.7, + 0.62 + ], + "angle": 0, + "content": "4.8 Comparison on all benchmarks to the baseline model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we present a comprehensive comparative analysis, focusing on the proposed approach and the baseline model. We report the Top-1 accuracy, Top-5 accuracy and the overall training time. The evaluation is conducted on the large and huge variants of ViViT across three datasets, namely Kinetics400, Kinetics600, and Moments in Time (MiT), with the summarized results tabulated in Table 5. The findings indicate a slight enhancement in accuracy for both Kinetics400 and Kinetics600 datasets, whereas a notable \\(1.79\\%\\) increase in top-1 accuracy is observed for the MiT dataset using the proposed method. Furthermore, the proposed approach showcases a significant reduction in training time, accounting for approximately \\(56\\%\\) of the original duration. This reduction emphasizes the advantageous nature of the proposed approach. To calculate the total training time for the SFA version, the train time of the 8 frame (Stage 1) ViViT model is combined with the train time of the (Stage 2) SFA-ViViT model. Conversely, the total training time for the standard ViViT encompasses" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "the total train time for the same number of frames that SFA-ViViT is trained on for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.205, + 0.782, + 0.284 + ], + "angle": 0, + "content": "
ModelKinetics-400Kinetics-600Moments in Time
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f82.59/93.091x(21.57 h)83.29/95.821x(26.14 h)--
ViViT-L128f + SFA8f82.78/94.030.56x()83.47/95.290.56x()--
ViViT-H96f84.21/94.661x(56.71 h)84.18/95.681x(60.45 h)38.17/62.841x(110.79 h)
ViViT-H96f + SFA8f84.42/94.720.57x()84.39/96.200.57x()39.96/64.390.59x()
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.284, + 0.785, + 0.326 + ], + "angle": 0, + "content": "Table 5: Performance Comparison of various versions of ViViT with the proposed training strategy for Kinetics-400, Kinetics-600 and Moments in Time. Accuracies listed as Top-1/Top-5." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.338, + 0.782, + 0.394 + ], + "angle": 0, + "content": "
ModelSomething-somethingSomething-ElseEpic-Kitchens
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f64.45/87.481x(14.17 h)53.14/73.981x(3.84 h)43.53/56.55/65.401x(5.61 h)
ViViT-L128f + SFA8f63.85/87.620.62x()53.60/74.470.62x()43.54/56.78/65.160.63x()
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.436 + ], + "angle": 0, + "content": "Table 6: Performance Comparison of ViViT-L with the proposed training strategy for Something-something v2, Something-Else and Epic-Kitchens. Accuracies listed as Top-1/Top-5, for Epic Kitchens Top-1 noun-verb/ Top-1 noun/ Top-1 Verb." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.616 + ], + "angle": 0, + "content": "We further examine the performance of ViViT-L incorporating our proposed training strategy in comparison to the original version on three additional datasets: Something-something, Something-Else, and Epic-Kitchens. A consistent trend is observed, with the modified approach outperforming the baseline model, at only a \\(62\\%\\) cost of the baseline training time. In summary, our proposed training strategy demonstrates promising potential by yielding comparable or slightly improved performance across all datasets. This is obtained while maintaining a training cost ranging from \\(56\\%\\) to \\(62\\%\\) of the original model, thus highlighting its effectiveness. Results can be seen in Table 6." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.473, + 0.659 + ], + "angle": 0, + "content": "4.9 Beyond the ViViT model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We adapted our approach to tiny versions of Dualformer [26], VideoSwin [29] and Video-FocalNets [45] on the Kinetics 400 dataset, demonstrating up to \\(2\\mathrm{x}\\) faster training with minimal accuracy loss. Thus, our method exhibits consistent improvements in training efficiency across multiple video architectures. We believe these comparisons better showcase the generalized benefits of our freeze-and-finetune approach for efficient video model training. These results can be seen in Table 7. While these models do not directly fall under the factorized encoder models, we can follow a similar pipeline to train them under the SFA training regime. For the Dualformer-T and VideoSwin-T, we first train 8-frame versions of these models and then freeze the first 3 blocks of the models (Stage 1 to Stage 3) and only finetune the last block." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.302, + 0.144, + 0.703, + 0.34 + ], + "angle": 0, + "content": "
ModelTotal Train Time Accuracy
Dualformer-T32f1x(30.72 h)79.53
Dualformer-T32f-SFA8f0.57x()79.48
Dualformer-B32f1x(68.57 h)81.07
Dualformer-B32f-SFA8f0.59x()81.12
VideoSwin-T32f1x(41.45 h)78.62
VideoSwin-T32f-SFA8f0.52x()78.68
VideoSwin-B32f1x(115.23 h)80.45
VideoSwin-T32f-SFA8f0.53x()80.52
Video-FocalNet-T32f1x(11.28 h)79.76
Video-FocalNet-T32f-SFA8f0.52x()79.95
Video-FocalNet-B32f1x(28.84 h)83.55
Video-FocalNet-B32f-SFA8f0.51x()84.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Table 7: Training efficiency improvements from SFA approach on tiny Dualformer and VideoSwin models. Our method achieves \\(>2\\mathrm{x}\\) faster training with minimal accuracy drop in Kinetics-400." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.429, + 0.681, + 0.446 + ], + "angle": 0, + "content": "4.10 Comparing with other efficient training strategies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.626 + ], + "angle": 0, + "content": "While we have already mentioned that strategies such as ST-Adapter, STTS etc are efficient in parameters and GFLOPs and not in training time. We demonstrate the results of training these in comparison to the proposed training method in terms of training time and inference time. The results demonstrate that our proposed SFA training pipeline confers substantial training time savings compared to these methods. For example, SFA-ViViT-B attains similar accuracy as ST-Adapter at approximately 0.58 times the total training time on Kinetics-400. We would like to point out that only STTS does better than the proposed strategy in terms of throughput, but as shown earlier requires twice the amount of time to train the model. All methods use 32 frames. These results are reported in Table 8." + }, + { + "type": "table", + "bbox": [ + 0.268, + 0.654, + 0.737, + 0.752 + ], + "angle": 0, + "content": "
ModelTotal Train Time ThroughputAccuracy
ViT-B + ST-Adapter [32]1x(34.18 hours)9082.0
ViViT-B [1]1.03x()8882.6
STTS [43]1.07x()12980.7
SBP [7]0.92x()8982.2
SFA8f-ViViT-B32f (Ours)0.58x()8882.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.752, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Table 8: Training efficiency improvements from SFA approach in comparison to other efficient training strategies. We require approximately half of the training time in comparison to other efficient training strategies." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.472, + 0.161 + ], + "angle": 0, + "content": "5 Evaluations Discussion" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.183, + 0.786, + 0.258 + ], + "angle": 0, + "content": "Our evaluations on ViViT was not to obtain SOTA performance using ViViT, but to show the reduced time required to train the model. In the ViViT paper [1], the reported results are different from what we report. This is due to some videos being removed from the Kinetics dataset. We re-run ViViT using the official codebase and report results on the datasets using this." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.26, + 0.787, + 0.425 + ], + "angle": 0, + "content": "The point of using 128 frames was not to show that using more frames improves performance, but instead to show that for larger models, using our training strategy we can actually access more frames due to memory being saved. The logic behind the training strategy holds for lower number of frames such as 32 as well. Hence, using multiview inference would still save time. We would still need only \\(0.6x\\) of the overall training time. Also, as pointed out in Section 4.7, larger models such as ViViT-g can only access 8 frames using standard ViViT-FE training. However, using our approach, we can access 48 frames. A notable benefit of accessing more frames is in long-form video understanding tasks and while such experiments have not been conducted, this training approach promises improved results in the task." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.455, + 0.357, + 0.47 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.493, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We have investigated the challenges posed by the substantial training time and memory consumption of video transformers, particularly focusing on the factorised encoder variant of the ViViT model as our baseline. To address these challenges, we proposed two effective strategies: utilizing a compact adapter model for fine-tuning image representations instead of end-to-end training of the spatial transformer, and initializing the temporal transformer using the baseline model trained with 8 frames. Our proposed training strategy has demonstrated the potential to significantly reduce training costs and memory consumption while maintaining, or even slightly improving, performance compared to the baseline model. Furthermore, we observed that with proper initialization, our baseline model can achieve near-peak performance within the first \\(10\\%\\) of training epochs. While we focus on the factorized encoder variant of ViViT, we also show the generalization power of the training regime by comparing results of DualFormer, VideoSwin and Video-FocalNets. We also show that the proposed method requires approximately half the time required in comparison to other efficient training strategies. The advancements made in this work could help research in the video understanding domain by enabling access to more frames and the utilization of larger image models as the spatial transformer, all while maintaining the same memory consumption. Our findings provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in action recognition. Future work may focus on further optimizing and refining these strategies, and exploring their application to other video transformer architectures." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.301 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.329 + ], + "angle": 0, + "content": "4. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. arXiv preprint arXiv:1808.01340 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "6. Chen, J., Ho, C.M.: Mm-vit: Multi-modal video transformer for compressed video action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1910-1921 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.467 + ], + "angle": 0, + "content": "7. Cheng, F., Xu, M., Xiong, Y., Chen, H., Li, X., Li, W., Xia, W.: Stochastic backpropagation: A memory efficient strategy for training video models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8301-8310 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.523 + ], + "angle": 0, + "content": "8. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 720-736 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Dehghani, M., Arnab, A., Beyer, L., Vaswani, A., Tay, Y.: The efficiency misnomer. arXiv preprint arXiv:2110.12894 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.605 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "2. Foteinopoulou, N.M., Patras, I.: Emoclip: A vision-language method for zero-shot video facial expression recognition. arXiv preprint arXiv:2310.16640 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "3. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "4. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "5. Gowda, S.N., Hao, X., Li, G., Sevilla-Lara, L., Gowda, S.N.: Watt for what: Rethinking deep learning's energy-performance relationship. arXiv preprint arXiv:2310.06522 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Gowda, S.N., Rohrbach, M., Keller, F., Sevilla-Lara, L.: Learn2augment: Learning to composite videos for data augmentation in action recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 242-259. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.44, + 0.128 + ], + "angle": 0, + "content": "Shreyank N Gowda et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.257 + ], + "angle": 0, + "content": "18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "19. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lučić, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "21. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.406 + ], + "angle": 0, + "content": "22. Kim, K., Gowda, S.N., Mac Aodha, O., Sevilla-Lara, L.: Capturing temporal information in a single frame: Channel sampling strategies for action recognition. arXiv preprint arXiv:2201.10394 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.447 + ], + "angle": 0, + "content": "23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "24. Laptev, I.: On space-time interest points. International journal of computer vision 64, 107-123 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "25. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.554 + ], + "angle": 0, + "content": "26. Liang, Y., Zhou, P., Zimmermann, R., Yan, S.: Dualformer: Local-global stratified transformer for efficient video recognition. In: European Conference on Computer Vision. pp. 577-595. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.785, + 0.595 + ], + "angle": 0, + "content": "27. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.65 + ], + "angle": 0, + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "29. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.785, + 0.744 + ], + "angle": 0, + "content": "30. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "31. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "33. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. arXiv preprint arXiv:2212.03229 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.203 + ], + "angle": 0, + "content": "34. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.203, + 0.785, + 0.245 + ], + "angle": 0, + "content": "35. Ryoo, M., Piergiovanni, A., Arnab, A., Dehghani, M., Angelova, A.: Tokenlearner: Adaptive space-time tokenization for videos. Advances in Neural Information Processing Systems 34, 12786-12797 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.245, + 0.785, + 0.273 + ], + "angle": 0, + "content": "36. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "37. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "38. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "39. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.383, + 0.785, + 0.424 + ], + "angle": 0, + "content": "40. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.424, + 0.785, + 0.467 + ], + "angle": 0, + "content": "41. Wang, H., Kläser, A., Schmid, C., Liu, C.L.: Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision 103, 60-79 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "42. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. Transactions of Machine Learning Research" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.508, + 0.785, + 0.564 + ], + "angle": 0, + "content": "43. Wang, J., Yang, X., Li, H., Liu, L., Wu, Z., Jiang, Y.G.: Efficient video transformers with spatial-temporal token selection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 69-86. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "44. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.605, + 0.785, + 0.66 + ], + "angle": 0, + "content": "45. Wasim, S.T., Khattak, M.U., Naseer, M., Khan, S., Shah, M., Khan, F.S.: Videofocalnets: Spatio-temporal focal modulation for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13778-13789 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "46. Xiong, X., Arnab, A., Nagrani, A., Schmid, C.: M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.688, + 0.785, + 0.731 + ], + "angle": 0, + "content": "47. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.73, + 0.785, + 0.785 + ], + "angle": 0, + "content": "48. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.785, + 0.827 + ], + "angle": 0, + "content": "49. Zhang, H., Hao, Y., Ngo, C.W.: Token shift transformer for video classification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 917-925 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_origin.pdf b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23b3978a34dcbd5e092e46f18868d94d4a6a547c --- /dev/null +++ b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/c4bb7a55-de6e-4483-9b92-2d2938c2d99c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b160e5014543a157b5b8f21f8571d91d638de09ade6b6261e88dbb7727df2e9 +size 621844 diff --git a/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/full.md b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..644c2a4cc48acaeea5eeef9b45b4fa90c1e1644e --- /dev/null +++ b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/full.md @@ -0,0 +1,255 @@ +# Optimizing Factorized Encoder Models: Time and Memory Reduction for Scalable and Efficient Action Recognition + +Shreyank N Gowda1, Anurag Arnab2, and Jonathan Huang2 + +1 University of Oxford + +2 Google Research + +Abstract. In this paper, we address the challenges posed by the substantial training time and memory consumption associated with video transformers, focusing on the ViViT (Video Vision Transformer) model, in particular the Factorised Encoder version, as our baseline for action recognition tasks. The factorised encoder variant follows the late-fusion approach that is adopted by many state of the art approaches. Despite standing out for its favorable speed/accuracy tradeoffs among the different variants of ViViT, its considerable training time and memory requirements still pose a significant barrier to entry. Our method is designed to lower this barrier and is based on the idea of freezing the spatial transformer during training. This leads to a low accuracy model if naively done. But we show that by (1) appropriately initializing the temporal transformer (a module responsible for processing temporal information) (2) introducing a compact adapter model connecting frozen spatial representations (a module that selectively focuses on regions of the input image) to the temporal transformer, we can enjoy the benefits of freezing the spatial transformer without sacrificing accuracy. Through extensive experimentation over 6 benchmarks, we demonstrate that our proposed training strategy significantly reduces training costs (by) and memory consumption while maintaining or slightly improving performance by up to $1.79\%$ compared to the baseline model. Our approach additionally unlocks the capability to utilize larger image transformer models as our spatial transformer and access more frames with the same memory consumption. We also show the generalization of this approach to other factorized encoder models. The advancements made in this work have the potential to advance research in the video understanding domain and provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in the action recognition field. + +# 1 Introduction + +Action recognition focuses on understanding and identifying actions in video sequences with applications in surveillance, human-computer interaction, and video content analysis. The field has advanced significantly due to large-scale annotated datasets [5] and a shift from hand-crafted features [24, 41] to deep + +![](images/605c40cbe22011785ca026fef4526aa157f4015630ea40ee6639e44b6c7af1af.jpg) +Fig. 1: Comparison of our initialization method vs conventional training of ViViT. Training time is scaled relative to setting ViViT-B training time to '1 unit' (7.93 hours). We see clear time saving using our initialization scheme and for larger models, the training time saved is much larger. + +learning models like convolutional networks (CNNs) [5, 13, 22, 36]. Recently, transformers have revolutionized computer vision by offering an alternative to traditional CNNs, leading to the development of many new state-of-the-art architectures [3, 10, 28]. Moreover, the flexibility of transformers have inspired researchers to adapt these models to more complex problems, including emotion recognition [12, 14], video understanding and action recognition [1, 29]. + +Transformers, however, are notoriously expensive [15], and Video Transformer-based architectures [1,2,29], which integrate information across space and time are even more so. And memory consumption and training times become even more significant when working with large-scale video datasets with long sequences [5]. These high computational costs present a particular challenge for researchers with limited resources, especially those from universities and smaller companies. The goal of our work is therefore to cut the cost of training: we want to train transformer-based video models with fewer resources or use larger model variants and handle more frames with the same resources. + +Our primary focus is on a diverse range of factorized encoder models, among which ViViT [1] stands out as a prominent example. Specifically, we focus on the "factorized encoder" variant of ViViT which has separate spatial and temporal transformer stages, where the spatial transformer is responsible for extracting features from individual frames, while the temporal transformer processes the temporal dynamics across frames. We choose this factorized encoder design because it is more efficient compared to, e.g., the variant of ViViT using all-to-all spatiotemporal attention, while still achieving high accuracies and has thus been + +adopted as the building block for recent state-of-the-art architectures on various tasks [6, 19, 42, 46-48]. + +To address the challenge of reducing training time and memory usage without compromising the sophistication and accuracy of the original model, our approach is based on the simple idea of freezing the spatial backbone. Freezing the spatial backbone has many advantages: by not backpropagating through this transformer, training is faster and requires less memory (allowing for the model to handle more frames). We also inherit the benefits of pretraining the spatial transformer on a large dataset (such as JFT [20]). Naively implemented, however, we show that this approach falls very short in accuracy. Instead, with a few simple (but important) tweaks to the above idea, we propose a method that has the same advantages of freezing the spatial transformer, but does not compromise on accuracy. + +Our method proceeds in two stages. In the first stage we pretrain a cheap version of the model using fewer frames, e.g., 8 frames as opposed to, e.g., 32 frames. In the second stage we fine tune this model with more frames, which is more expensive but in this stage we freeze the spatial encoder and introduce a compact "adapter" model connecting frozen spatial representations to the temporal transformer, negating the need for end-to-end training of the spatial transformer. Critically this includes pre-training the temporal transformer (by initializing from stage 1) which is often overlooked in current video models which typically initialize this component from scratch. However our experiments show that this step is critical if we wish to not sacrifice performance. + +With our training recipe, we match or slightly outperform conventional training of ViViT at roughly half the cost as seen in Figure 1. A notable benefit of our training recipe, is its ability to process up to 80 frames (ViViT-B) on typical university-grade GPUs, a significant leap from the previous capacity of 16 frames. This expansion in processing power broadens the range of video data manageable under resource-constrained settings. As we elaborate in Section 4.7, our research underscores the potential to democratize access to advanced video transformer models. Another notable benefit is the model's ability to now use even larger models as the spatial transformer, we introduce ViViT-g as seen in Section 4.6. This accessibility paves the way for future video action recognition research, irrespective of resource constraints. We also show in Section 4.9 that our method generalizes to non-ViViT models such as Dualformer [26] and VideoSwin [29]. Hereafter, we refer to our version of ViViT as SFA-ViViT, where SFA denotes 'Spatial Frozen and Adapter Initialized'. + +# 2 Related Work + +Transformers for Videos Action recognition is a key research area in computer vision, addressed by many traditional [24, 41] and CNN based approaches [5, 16, 17, 21, 27, 36, 39, 44] aided by the release of large-scale datasets [5, 23, 37]. Since we focus on transformer based architectures, a thorough review of earlier methods are out of this scope. More recently, the transformer architecture, initially devel + +oped for NLP tasks [40], has been adapted for video understanding and action recognition tasks, leading to state-of-the-art models such as TimeSformer [2], ViViT [1], VideoSwin [29], and Uniformer [25] These transformer-based models leverage self-attention mechanisms to capture complex spatiotemporal patterns in action recognition tasks. TimeSformer [2] is one of the first transformer-based models for video understanding, adapting the transformer architecture to video by treating it as a sequence of flattened image patches. ViViT [1] integrates spatial and temporal transformers to efficiently capture spatiotemporal information in video sequences. VideoSwin [29] is a hierarchical transformer that applies local windowing for efficiency, enabling the model to handle longer video sequences. Similar to VideoSwin, Dualformer [26] also adopts a hierarchical architecture but uses a local-global stratified transformer design to capture both fine-grained and global video features efficiently. More recently, Uniformer [25] integrates 3D convolution and spatiotemporal self-attention, MTV [47] proposes a multi-view transformer model using distinct encoders for each video "view", improving accuracy as the number of views increases. The Multiscale Vision Transformers (MViT) [11] model streamlines computation and memory usage by operating at different resolutions, focusing on high-level features at lower resolutions and low-level details at higher ones, effectively leveraging both spatial and temporal information in visual tasks. TubeViT [33] introduces a method of sparsely sampling different-sized 3D segments from videos, facilitating efficient joint image and video learning, and allowing the adaptation of larger models to videos with less computational resources. Typically these models have FLOPs in the range of TFLOPs and training times that last more than days on the largest of GPUs/TPUs available, making them infeasible to train or use in lower resourced settings such as academia. It is critical that we find a way to train these models with limited resources while maintaining their performance. To this end, we focus on the factorised encoder version of ViViT as the late-fusion approach followed is used as a foundation for state-of-the-art approaches of various tasks [6, 19, 42, 46–48] and hence believe that the initialization scheme proposed can be used for future methods working on similar architectures. Similar to the approach for ViViT factorized encoder, the hierarchical architectures of VideoSwin and DualFormer can also benefit from staged training. By freezing the early layers that capture local spatial-temporal features, and adding an adapter before fine-tuning the later layers focused on global context, the overall training can be made more efficient. + +Efficient Transformers in Videos Efficiency is a nuanced topic [9], as there are multiple cost indicators of efficiency (for example, GFLOPs, inference time, training time, memory usage), and models which improve efficiency in one dimension, are not necessarily better in other dimensions [9]. TokenLearner [35] proposes a method that adaptively learns tokens for efficient image and video understanding tasks, enabling effective modeling of pairwise attention over longer temporal horizons or spatial content. TokenLearner reduces the GFLOPs required by ViViT by about half, but does not significantly change the training time or the inference time of ViViT. Spatial Temporal Token Selection (STTS) [43] proposes + +a dynamic token selection framework for spatial and temporal dimensions that ranks token importance using a lightweight scorer network, selecting top-scoring tokens for downstream evaluation in an end-to-end training process. STTS again reduces the GFLOPs, but the training time and inference time do not change significantly. TokShift [49], a zero-parameter, zero-FLOPs operator that models temporal relations in transformer encoders by temporally shifting partial token features across adjacent frames but again requires the same training time as the original model. By densely integrating TokShift into a plain 2D vision transformer, a computationally efficient, convolution-free video transformer is created for video understanding. Most similar to our work is the ST-Adapter [32], that utilizes built-in spatio-temporal reasoning in a compact design, allowing pretrained image models to reason about dynamic video content with a small pertask parameter cost, surpassing existing methods in both parameter-efficiency and performance. However, it does not change FLOPs or inference time at all. Unlike ST-Adapter, we use a spatial only adapter which we show is enough to reproduce the performance of the baseline model at close to half the training time. In particular, our proposed method improves the training time and training memory usage, addressing the key problem of researchers and practitioners being able to train video models. It does not, however, change the inference time compared to a standard ViViT model. We consider overall train time for the same hyperparameters and use the same hardware for a direct comparison. We consider efficiency in this paper as the time saved in the overall training of the model. + +# 3 Methodology + +# 3.1 Revisiting ViViT + +The Video Vision Transformer (ViViT) extends the Vision Transformer architecture to handle video data by incorporating spatio-temporal reasoning. The idea behind ViViT is to process video input as a sequence of image patches, combining spatial and temporal information through a series of transformer layers, which include multi-head self-attention, layer normalization, and feed-forward networks. The output is used for video classification. + +In the "vanilla" variant of ViViT, one extracts spatio-temporal tokens from a video then forwards all tokens through a transformer encoder which explicitly models all pairwise interactions between all spatio-temporal tokens. We build off of the more efficient "Factorized Encoder" variant of ViViT whose architecture consists of two separate transformer encoders, a spatial transformer modeling interactions between tokens from the same temporal index and a temporal transformer modeling interactions between tokens from different temporal indices. Despite having more parameters, it requires fewer floating point operations (FLOPs) than vanilla ViViT. Because the Factorised Encoder variant strikes a good balance point between accuracy and processing speed, it has also been adopted as the foundation for other architectures [6,19,42,46-48], reinforcing its utility and robustness. + +# 3.2 Our training strategy + +We concentrate on the factorised encoder variant of ViViT as it is already the most efficient version of the baseline. Henceforth, when we talk about ViViT we refer to this variant of ViViT. Consider the ViViT model that contains a spatial transformer with parameters $\theta_{\text{spatial}}$ and a temporal transformer with parameters $\theta_{\text{temporal}}$ : + +$$ +X _ {\text {s p a t i a l}} = T _ {\text {s p a t i a l}} \left(X _ {\text {i n}}; \theta_ {\text {s p a t i a l}}\right) \tag {1} +$$ + +$$ +X _ {o u t} = T _ {t e m p o r a l} (X _ {s p a t i a l}; \theta_ {t e m p o r a l}). +$$ + +In conventional ViViT training, $\theta_{\text{spatial}}$ is initialized from an image pre-trained checkpoint such as ImageNet-21k [34] or JFT [20] and the $\theta_{\text{temporal}}$ is initialized from scratch. During backpropagation, the gradient flows through the entire model. This entails training two sizable transformer models end-to-end, which is a highly resource-intensive process, as the transformer architecture is inherently computationally demanding, especially with more frames and larger ViViT variants (e.g., ViViT-H). + +One approach to reducing training time is to freeze the parameters of the spatial transformer $\theta_{\text{spatial}}$ . By not backpropagating through $\theta_{\text{spatial}}$ , gradient updates are faster and require less memory, allowing us to access more frames without encountering out-of-memory issues. But as we show in experiments, the accuracy of the resulting model with frozen $\theta_{\text{spatial}}$ is not competitive (in accuracy) with the baseline training approach. + +We present a two stage approach (see Fig. 2) to training ViViT models that inherits the same benefits of freezing the spatial transformer, while not compromising on model quality. + +Stage 1. In Stage 1, we pretrain our ViViT model on a reduced number of frames initializing the spatial transformer using a pre-trained image checkpoint. We do not freeze the spatial transformer during this stage, but critically, Stage 1 serves to also initialize the temporal transformer. + +To set the number of frames at this stage, we must balance the goal of efficiency (using fewer frames) against our finding in experiments that pre-training on too few frames can lead to suboptimal results. In our ablations, we identify a sweet spot at 8 frames. + +Stage 2. In Stage 2, we fine tune our ViViT model on the full frame count (e.g. 128 frames) initializing both spatial and temporal transformer parameters learned in Stage 1. Because this stage is significantly more expensive, in stage 2, we freeze the spatial transformer parameters $\theta_{\text{spatial}}$ and add a lightweight adapter module with parameters $\theta_{\text{adapter}}$ following the spatial transformer: + +$$ +X _ {s p a t i a l} = T _ {s p a t i a l} (X _ {i n}; \theta_ {s p a t i a l}) +$$ + +$$ +X _ {\text {a d a p t e r}} = A _ {\text {a d a p t e r}} \left(X _ {\text {s p a t i a l}}; \theta_ {\text {a d a p t e r}}\right) \tag {2} +$$ + +$$ +X _ {o u t} = T _ {t e m p o r a l} (X _ {a d a p t e r}; \theta_ {t e m p o r a l}) +$$ + +In this setting, by backpropagating only through the temporal transformer and the lightweight adapter module (in our experiments, a two layer MLP), we effectively cut total training time by half. + +![](images/9684e11d88c86e01c18153a848d98f9e34482da6431f64f0305b8a0bf38b378f.jpg) +Fig. 2: STAGE 1: We first use the full ViViT-FE model on 8 frames by initializing the spatial transformer from an image checkpoint and the temporal transformer from scratch. STAGE 2: We then use this as our checkpoint to initialize the spatial and temporal transformer for models using more frames (such as 32, 64 or 128). We then freeze the spatial transformer and add an adapter model to finetune spatial transformer features. The temporal transformer is finetuned from the same checkpoint. + +The crucial finding here is that the spatial transformer requires only short-term context for initialization (after which it remains frozen), whereas the temporal transformer necessitates long-term context to achieve its optimal performance. Further details and empirical analysis can be found in the next section. + +# 4 Experimental Analysis + +Through a series of comprehensive experiments which we now present, we investigate the significance of the spatial transformer, examining the impact of pre-training datasets and how larger models affect action recognition performance. We also explore the importance of initializing the temporal transformer by employing various initialization schemes and datasets, assessing whether the number of frames is critical for initializing larger models, initializing full ViViT models, and initializing models on one dataset while fine-tuning on another. + +# 4.1 Datasets + +We evaluate on all the datasets considered in [1] (specifically, Kinetics-400 [5], Kinetics-600 [4], EPIC-Kitchens [8], Something-something v2 [18] and Moments-in-time [31]) as well as the Something-Else [30] dataset. As these datasets are + +common in the community, we include further details in the supplementary along with the implementation details. + +# 4.2 Ablation Study + +We first address two critical aspects: the significance of fine-tuning the spatial transformer and the importance of initializing the temporal transformer. To do so, we conduct a series of experiments in various scenarios, which are detailed below. Our analysis focuses on the Something-something dataset, utilizing the large version of the ViViT model, referred to as ViViT-L. + +We examine four main elements that modify the structure of conventional ViViT training and these are mentioned with indices in Table 1 namely: I. The freezing of the spatial transformer ( $\theta_{\text{spatial}}$ is initialized and then frozen), II. The freezing of the temporal transformer ( $\theta_{\text{temporal}}$ is frozen), III. The addition of an adapter (lightweight module with parameters $\theta_{\text{adapter}}$ ), IV. Next, we initialize the temporal transformer using VideoMAE [38], while keeping the spatial transformer frozen and the adapter incorporated and V. The initialization of the temporal transformer ( $\theta_{\text{spatial}}$ and $\theta_{\text{temporal}}$ are initialized using the ViViT $_{8f}$ ). + +It is important to note that the VideoMAE training is an extremely expensive process as can be seen in the table. But combined with the line below it, these two models, which significantly outperform lines I, II and III, show that properly initializing the temporal transformer is critical. + +
IndexSpatialFrozenAdapterTemporalFrozenTemporal InitTop-1AccTop-5AccTrain Time
-××××××64.4587.4814.17 h
I✓×××××27.7556.730.5x()
II××✓×××25.8053.070.5x()
III✓✓××××38.7768.930.53x()
IV✓✓×VMAE58.5485.832.51x()
V✓✓×ViViT-8f63.8587.620.62x()
VI×✓×ViViT-8f64.1387.351.37x()
+ +Table 1: Ablation study results illustrating the impact of various modifications to the ViViT-B model, including spatial and temporal transformer freezing, adapter addition, and initialization methods, on top-1 and top-5 accuracy. Dataset is Something-something v2. + +Additionally, initializing the spatial transformer yields further improvement. The adapter plays a vital role in augmenting performance when the spatial transformer is frozen, and due to its lightweight nature, it will be an essential component of our training methodology. + +# 4.3 How many frames should we use for Stage 1? + +Next we experiment with various frame counts for stage 1 training, We test seven variants: JFT [20] checkpoint (image-based), 2, 4, 8, 16, 32, and 48-frame ViViT checkpoints. We then fine-tune these with a frozen spatial transformer and add + +![](images/7d5c317ddb1f7b3a0aabbf997b8b620449519d929e656203aa8e93b4f0aafa1d.jpg) +Fig. 3: The effect of initializing with different numbers of frames (JFT, 2, 4, 8, 16, 32, and 48), freezing the spatial transformer and adding an adapter model and fine-tuning using 64, 96, and 128 frames. Results on Kinetics400 dataset, 'f' refers to frames. + +![](images/fe08a86088348e9a74136a447c7c9de0bf4d4007e5f6187b2ffcf33f2efd0040.jpg) +Fig.4: Comparison of our initialization method vs conventional training of ViViT on Top-1 accuracy and loss on the Kinetics-400 dataset using 64 and 128 frames. We see that our initialization gives a significant headstart to the models. + +an adapter model using 64, 96, and 128 frames (see Figure 3). Results show that using too few frames for Stage 1 training can underperform (with image-only initialization from a JFT [20] checkpoint performing the worst). Thus we deduce that short term temporal context is essential for initializing the spatial transformer. Performance also plateaus after 8 frames, and given that using more frames increases training time, we settle on using 8 frames as our "sweet spot" for Stage 1 training. We consider this an ablation, as this is done only on Kinetics and kept fixed for all models and datasets. + +# 4.4 Does the proposed training increase convergence speed? + +Another potential question that may arise concerns the impact of this initialization method in terms of convergence speed, if any. This specific aspect holds considerable significance owing to its potential ability to drastically curtail the duration of time required for training and the number of epochs necessary to effectively train the model. Moreover, an important element to take into account is the effect of freezing the spatial transformer. This approach decreases the memory needed to store the model but also considerably enhances the training speed. To provide a clearer picture, we have plotted the validation curves with and without initialization, which can be seen in Figure 4. Note that with the proposed initialization, we get a significant head start in overall accuracy. + +# 4.5 What about initializing on one dataset and finetuning on other? + +In our study, we find training an 8-frame version of the standard ViViT model affordable. We consider training this on a temporally dependent dataset like + +Something-something, then fine-tuning on other datasets like Kinetics-400. We examine two scenarios: first, the standard ViViT model trained on Kinetics-400 using 8-frames, and second, the same model trained on Something-something. Post-training, we freeze the spatial transformer, add an adapter, and fine-tune the models on the alternate dataset with more frames. We contrast this with models fine-tuned on their original datasets (see Table 2). Results favor initializing larger frame models from the 8-frame version on the same dataset. Thus, for the final comparison in Sec 4.8, we initialize models with 8-frame versions of the baseline. + +
CheckpointSSv2K400
K400-init44.71/74.5382.81/93.98
SSv2-init63.85/87.6276.79/92.35
+ +
BackboneTop-1Top-5Steps
ViViT-L79.6491.7348k steps
ViViT-H81.0293.0939k steps
ViViT-g81.8194.5529k steps
+ +Table 2: A summary of cross-dataset initialization of the proposed model and performance comparison. We use Kinetics400 and Something-something v2 as our datasets. + +Table 3: A comparison of top-1 and top-5 accuracies for the ViViT-g model with the proposed training strategy, which incorporates a larger spatial transformer backbone. All models use 48 frames for fair comparison. Results are on Kinetics400 dataset. + +# 4.6 Extending the image backbone to 1.5B parameters + +An intriguing consequence of our approach is the ability to incorporate larger backbones into the spatial transformer, made possible by the additional memory available to us as a result of freezing the spatial transformer during training. Consequently, we introduce ViViT-g, which integrates the ViT-g model (with 1.5B parameters) as its backbone. To ensure a fair comparison, we focus solely on training and inference using 48 frames, and abstain from employing multiview or multicrop testing. Our objective is to investigate the potential impact of a more substantial spatial transformer backbone on the overall performance and show the potential of larger spatial backbones that are possible due to our training process. It is essential to note that the full ViViT-g model could not process more than 8 frames due to memory limitations. However, our proposed strategy allows processing up to 48 frames. A comparison of accuracies is presented in Table 3 along with the number of steps needed to reach the best performance. Dataset used is Kinetics-400 and all the ViT checkpoints are JFT-pretrained [20]. + +# 4.7 Comparison of Memory Usage with Standard ViViT Training and Proposed Method + +In this study, we compare the number of frames that can be accessed using the standard ViViT training scheme against our proposed scheme, employing a set of 64 v3 TPUs that have 16 GB each. We further evaluate the performance of ViViT variants, including H, and g, in comparison with the SFA-ViViT using the + +same variant configurations. Maintaining identical hyperparameters, we ensure a local batch size of 1. Our findings indicate that the conventional ViViT training approach restricts frame accessibility to 96 frames for the ViViT-H model, and a mere 8 frames for the ViViT-g model, before reaching memory limitations. Conversely, our proposed method enables access to 128 frames for ViViT-H, and up to 48 frames when utilizing ViViT-g with the same hardware. Furthermore, we investigate the impact of utilizing university grade GPUs by conducting ViViT experiments on an NVIDIA RTX 2080 Ti GPU farm equipped with 8 GPUs having 12 GB each. Under these circumstances, ViViT can only process 16 frames using a local batch size of 1. However, our proposed training strategy enables a notable improvement, expanding the frame capacity to 80 frames (ViViT-B) helping us reproduce ViViT results on lower end GPUs. This enhancement provides a valuable opportunity for researchers with limited resources to attain performance levels comparable to those with extensive resources. We show a comparison of number of frames accessible with and without our training recipe in Table 4. + +
Model481632486496128
ViViT-H✓✓✓✓✓✓✓×
SFA-ViViT-H✓✓✓✓✓✓✓✓
ViViT-g✓✓××××××
SFA-ViViT-g✓✓✓✓✓×××
+ +Table 4: Memory usage in different ViViT training schemes is compared using the Kinetics400 dataset on 64 TPUs v3 with 16GB memory each. A $\checkmark$ indicates accessible frames given hardware constraints, while a $\times$ signals an out-of-memory (OOM) error. + +# 4.8 Comparison on all benchmarks to the baseline model + +In this section, we present a comprehensive comparative analysis, focusing on the proposed approach and the baseline model. We report the Top-1 accuracy, Top-5 accuracy and the overall training time. The evaluation is conducted on the large and huge variants of ViViT across three datasets, namely Kinetics400, Kinetics600, and Moments in Time (MiT), with the summarized results tabulated in Table 5. The findings indicate a slight enhancement in accuracy for both Kinetics400 and Kinetics600 datasets, whereas a notable $1.79\%$ increase in top-1 accuracy is observed for the MiT dataset using the proposed method. Furthermore, the proposed approach showcases a significant reduction in training time, accounting for approximately $56\%$ of the original duration. This reduction emphasizes the advantageous nature of the proposed approach. To calculate the total training time for the SFA version, the train time of the 8 frame (Stage 1) ViViT model is combined with the train time of the (Stage 2) SFA-ViViT model. Conversely, the total training time for the standard ViViT encompasses + +the total train time for the same number of frames that SFA-ViViT is trained on for fair comparison. + +
ModelKinetics-400Kinetics-600Moments in Time
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f82.59/93.091x(21.57 h)83.29/95.821x(26.14 h)--
ViViT-L128f + SFA8f82.78/94.030.56x()83.47/95.290.56x()--
ViViT-H96f84.21/94.661x(56.71 h)84.18/95.681x(60.45 h)38.17/62.841x(110.79 h)
ViViT-H96f + SFA8f84.42/94.720.57x()84.39/96.200.57x()39.96/64.390.59x()
+ +Table 5: Performance Comparison of various versions of ViViT with the proposed training strategy for Kinetics-400, Kinetics-600 and Moments in Time. Accuracies listed as Top-1/Top-5. + +
ModelSomething-somethingSomething-ElseEpic-Kitchens
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f64.45/87.481x(14.17 h)53.14/73.981x(3.84 h)43.53/56.55/65.401x(5.61 h)
ViViT-L128f + SFA8f63.85/87.620.62x()53.60/74.470.62x()43.54/56.78/65.160.63x()
+ +Table 6: Performance Comparison of ViViT-L with the proposed training strategy for Something-something v2, Something-Else and Epic-Kitchens. Accuracies listed as Top-1/Top-5, for Epic Kitchens Top-1 noun-verb/ Top-1 noun/ Top-1 Verb. + +We further examine the performance of ViViT-L incorporating our proposed training strategy in comparison to the original version on three additional datasets: Something-something, Something-Else, and Epic-Kitchens. A consistent trend is observed, with the modified approach outperforming the baseline model, at only a $62\%$ cost of the baseline training time. In summary, our proposed training strategy demonstrates promising potential by yielding comparable or slightly improved performance across all datasets. This is obtained while maintaining a training cost ranging from $56\%$ to $62\%$ of the original model, thus highlighting its effectiveness. Results can be seen in Table 6. + +# 4.9 Beyond the ViViT model + +We adapted our approach to tiny versions of Dualformer [26], VideoSwin [29] and Video-FocalNets [45] on the Kinetics 400 dataset, demonstrating up to $2\mathrm{x}$ faster training with minimal accuracy loss. Thus, our method exhibits consistent improvements in training efficiency across multiple video architectures. We believe these comparisons better showcase the generalized benefits of our freeze-and-finetune approach for efficient video model training. These results can be seen in Table 7. While these models do not directly fall under the factorized encoder models, we can follow a similar pipeline to train them under the SFA training regime. For the Dualformer-T and VideoSwin-T, we first train 8-frame versions of these models and then freeze the first 3 blocks of the models (Stage 1 to Stage 3) and only finetune the last block. + +
ModelTotal Train Time Accuracy
Dualformer-T32f1x(30.72 h)79.53
Dualformer-T32f-SFA8f0.57x()79.48
Dualformer-B32f1x(68.57 h)81.07
Dualformer-B32f-SFA8f0.59x()81.12
VideoSwin-T32f1x(41.45 h)78.62
VideoSwin-T32f-SFA8f0.52x()78.68
VideoSwin-B32f1x(115.23 h)80.45
VideoSwin-T32f-SFA8f0.53x()80.52
Video-FocalNet-T32f1x(11.28 h)79.76
Video-FocalNet-T32f-SFA8f0.52x()79.95
Video-FocalNet-B32f1x(28.84 h)83.55
Video-FocalNet-B32f-SFA8f0.51x()84.25
+ +# 4.10 Comparing with other efficient training strategies + +While we have already mentioned that strategies such as ST-Adapter, STTS etc are efficient in parameters and GFLOPs and not in training time. We demonstrate the results of training these in comparison to the proposed training method in terms of training time and inference time. The results demonstrate that our proposed SFA training pipeline confers substantial training time savings compared to these methods. For example, SFA-ViViT-B attains similar accuracy as ST-Adapter at approximately 0.58 times the total training time on Kinetics-400. We would like to point out that only STTS does better than the proposed strategy in terms of throughput, but as shown earlier requires twice the amount of time to train the model. All methods use 32 frames. These results are reported in Table 8. + +Table 7: Training efficiency improvements from SFA approach on tiny Dualformer and VideoSwin models. Our method achieves $>2\mathrm{x}$ faster training with minimal accuracy drop in Kinetics-400. + +
ModelTotal Train Time ThroughputAccuracy
ViT-B + ST-Adapter [32]1x(34.18 hours)9082.0
ViViT-B [1]1.03x()8882.6
STTS [43]1.07x()12980.7
SBP [7]0.92x()8982.2
SFA8f-ViViT-B32f (Ours)0.58x()8882.8
+ +Table 8: Training efficiency improvements from SFA approach in comparison to other efficient training strategies. We require approximately half of the training time in comparison to other efficient training strategies. + +# 5 Evaluations Discussion + +Our evaluations on ViViT was not to obtain SOTA performance using ViViT, but to show the reduced time required to train the model. In the ViViT paper [1], the reported results are different from what we report. This is due to some videos being removed from the Kinetics dataset. We re-run ViViT using the official codebase and report results on the datasets using this. + +The point of using 128 frames was not to show that using more frames improves performance, but instead to show that for larger models, using our training strategy we can actually access more frames due to memory being saved. The logic behind the training strategy holds for lower number of frames such as 32 as well. Hence, using multiview inference would still save time. We would still need only $0.6x$ of the overall training time. Also, as pointed out in Section 4.7, larger models such as ViViT-g can only access 8 frames using standard ViViT-FE training. However, using our approach, we can access 48 frames. A notable benefit of accessing more frames is in long-form video understanding tasks and while such experiments have not been conducted, this training approach promises improved results in the task. + +# 6 Conclusion + +We have investigated the challenges posed by the substantial training time and memory consumption of video transformers, particularly focusing on the factorised encoder variant of the ViViT model as our baseline. To address these challenges, we proposed two effective strategies: utilizing a compact adapter model for fine-tuning image representations instead of end-to-end training of the spatial transformer, and initializing the temporal transformer using the baseline model trained with 8 frames. Our proposed training strategy has demonstrated the potential to significantly reduce training costs and memory consumption while maintaining, or even slightly improving, performance compared to the baseline model. Furthermore, we observed that with proper initialization, our baseline model can achieve near-peak performance within the first $10\%$ of training epochs. While we focus on the factorized encoder variant of ViViT, we also show the generalization power of the training regime by comparing results of DualFormer, VideoSwin and Video-FocalNets. We also show that the proposed method requires approximately half the time required in comparison to other efficient training strategies. The advancements made in this work could help research in the video understanding domain by enabling access to more frames and the utilization of larger image models as the spatial transformer, all while maintaining the same memory consumption. Our findings provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in action recognition. Future work may focus on further optimizing and refining these strategies, and exploring their application to other video transformer architectures. + +# References + +1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021) +2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) +4. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. arXiv preprint arXiv:1808.01340 (2018) +5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +6. Chen, J., Ho, C.M.: Mm-vit: Multi-modal video transformer for compressed video action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1910-1921 (2022) +7. Cheng, F., Xu, M., Xiong, Y., Chen, H., Li, X., Li, W., Xia, W.: Stochastic backpropagation: A memory efficient strategy for training video models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8301-8310 (2022) +8. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 720-736 (2018) +9. Dehghani, M., Arnab, A., Beyer, L., Vaswani, A., Tay, Y.: The efficiency misnomer. arXiv preprint arXiv:2110.12894 (2021) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021) +2. Foteinopoulou, N.M., Patras, I.: Emoclip: A vision-language method for zero-shot video facial expression recognition. arXiv preprint arXiv:2310.16640 (2023) +3. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017) +4. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024) +5. Gowda, S.N., Hao, X., Li, G., Sevilla-Lara, L., Gowda, S.N.: Watt for what: Rethinking deep learning's energy-performance relationship. arXiv preprint arXiv:2310.06522 (2023) +6. Gowda, S.N., Rohrbach, M., Keller, F., Sevilla-Lara, L.: Learn2augment: Learning to composite videos for data augmentation in action recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 242-259. Springer (2022) + +17. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021) +18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The" something something" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017) +19. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lučić, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023) +20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +21. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012) +22. Kim, K., Gowda, S.N., Mac Aodha, O., Sevilla-Lara, L.: Capturing temporal information in a single frame: Channel sampling strategies for action recognition. arXiv preprint arXiv:2201.10394 (2022) +23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011) +24. Laptev, I.: On space-time interest points. International journal of computer vision 64, 107-123 (2005) +25. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022) +26. Liang, Y., Zhou, P., Zimmermann, R., Yan, S.: Dualformer: Local-global stratified transformer for efficient video recognition. In: European Conference on Computer Vision. pp. 577-595. Springer (2022) +27. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019) +28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +29. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022) +30. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020) +31. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019) +32. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022) + +33. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. arXiv preprint arXiv:2212.03229 (2022) +34. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021) +35. Ryoo, M., Piergiovanni, A., Arnab, A., Dehghani, M., Angelova, A.: Tokenlearner: Adaptive space-time tokenization for videos. Advances in Neural Information Processing Systems 34, 12786-12797 (2021) +36. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014) +37. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) +38. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602 (2022) +39. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015) +40. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +41. Wang, H., Kläser, A., Schmid, C., Liu, C.L.: Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision 103, 60-79 (2013) +42. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. Transactions of Machine Learning Research +43. Wang, J., Yang, X., Li, H., Liu, L., Wu, Z., Jiang, Y.G.: Efficient video transformers with spatial-temporal token selection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 69-86. Springer (2022) +44. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016) +45. Wasim, S.T., Khattak, M.U., Naseer, M., Khan, S., Shah, M., Khan, F.S.: Videofocalnets: Spatio-temporal focal modulation for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13778-13789 (2023) +46. Xiong, X., Arnab, A., Nagrani, A., Schmid, C.: M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852 (2022) +47. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022) +48. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +49. Zhang, H., Hao, Y., Ngo, C.W.: Token shift transformer for video classification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 917-925 (2021) \ No newline at end of file diff --git a/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/images.zip b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e499dc3ec51daac850b764d87aa083b64bcefb69 --- /dev/null +++ b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cbe1e35d25524669eabc51794dc16259c6a82ccfeb70be069afc588dd172bee +size 328643 diff --git a/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/layout.json b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a8c55e67427fed16755b180fd8c35a9456987114 --- /dev/null +++ b/2024/Optimizing Factorized Encoder Models_ Time and Memory Reduction for Scalable and Efficient Action Recognition/layout.json @@ -0,0 +1,6871 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 166 + ], + "type": "text", + "content": "Optimizing Factorized Encoder Models: Time and Memory Reduction for Scalable and Efficient Action Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 172, + 186, + 441, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 186, + 441, + 199 + ], + "spans": [ + { + "bbox": [ + 172, + 186, + 441, + 199 + ], + "type": "text", + "content": "Shreyank N Gowda1, Anurag Arnab2, and Jonathan Huang2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 259, + 207, + 354, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 207, + 354, + 219 + ], + "spans": [ + { + "bbox": [ + 259, + 207, + 354, + 219 + ], + "type": "text", + "content": "1 University of Oxford" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 268, + 220, + 345, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 220, + 345, + 230 + ], + "spans": [ + { + "bbox": [ + 268, + 220, + 345, + 230 + ], + "type": "text", + "content": "2 Google Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 256, + 455, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 455, + 574 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 455, + 574 + ], + "type": "text", + "content": "Abstract. In this paper, we address the challenges posed by the substantial training time and memory consumption associated with video transformers, focusing on the ViViT (Video Vision Transformer) model, in particular the Factorised Encoder version, as our baseline for action recognition tasks. The factorised encoder variant follows the late-fusion approach that is adopted by many state of the art approaches. Despite standing out for its favorable speed/accuracy tradeoffs among the different variants of ViViT, its considerable training time and memory requirements still pose a significant barrier to entry. Our method is designed to lower this barrier and is based on the idea of freezing the spatial transformer during training. This leads to a low accuracy model if naively done. But we show that by (1) appropriately initializing the temporal transformer (a module responsible for processing temporal information) (2) introducing a compact adapter model connecting frozen spatial representations (a module that selectively focuses on regions of the input image) to the temporal transformer, we can enjoy the benefits of freezing the spatial transformer without sacrificing accuracy. Through extensive experimentation over 6 benchmarks, we demonstrate that our proposed training strategy significantly reduces training costs (by) and memory consumption while maintaining or slightly improving performance by up to " + }, + { + "bbox": [ + 160, + 256, + 455, + 574 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 574 + ], + "type": "text", + "content": " compared to the baseline model. Our approach additionally unlocks the capability to utilize larger image transformer models as our spatial transformer and access more frames with the same memory consumption. We also show the generalization of this approach to other factorized encoder models. The advancements made in this work have the potential to advance research in the video understanding domain and provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in the action recognition field." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 594, + 230, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 230, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 230, + 607 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Action recognition focuses on understanding and identifying actions in video sequences with applications in surveillance, human-computer interaction, and video content analysis. The field has advanced significantly due to large-scale annotated datasets [5] and a shift from hand-crafted features [24, 41] to deep" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 127, + 451, + 297 + ], + "blocks": [ + { + "bbox": [ + 159, + 127, + 451, + 297 + ], + "lines": [ + { + "bbox": [ + 159, + 127, + 451, + 297 + ], + "spans": [ + { + "bbox": [ + 159, + 127, + 451, + 297 + ], + "type": "image", + "image_path": "605c40cbe22011785ca026fef4526aa157f4015630ea40ee6639e44b6c7af1af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "content": "Fig. 1: Comparison of our initialization method vs conventional training of ViViT. Training time is scaled relative to setting ViViT-B training time to '1 unit' (7.93 hours). We see clear time saving using our initialization scheme and for larger models, the training time saved is much larger." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "text", + "content": "learning models like convolutional networks (CNNs) [5, 13, 22, 36]. Recently, transformers have revolutionized computer vision by offering an alternative to traditional CNNs, leading to the development of many new state-of-the-art architectures [3, 10, 28]. Moreover, the flexibility of transformers have inspired researchers to adapt these models to more complex problems, including emotion recognition [12, 14], video understanding and action recognition [1, 29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 459, + 483, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 483, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 483, + 567 + ], + "type": "text", + "content": "Transformers, however, are notoriously expensive [15], and Video Transformer-based architectures [1,2,29], which integrate information across space and time are even more so. And memory consumption and training times become even more significant when working with large-scale video datasets with long sequences [5]. These high computational costs present a particular challenge for researchers with limited resources, especially those from universities and smaller companies. The goal of our work is therefore to cut the cost of training: we want to train transformer-based video models with fewer resources or use larger model variants and handle more frames with the same resources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Our primary focus is on a diverse range of factorized encoder models, among which ViViT [1] stands out as a prominent example. Specifically, we focus on the \"factorized encoder\" variant of ViViT which has separate spatial and temporal transformer stages, where the spatial transformer is responsible for extracting features from individual frames, while the temporal transformer processes the temporal dynamics across frames. We choose this factorized encoder design because it is more efficient compared to, e.g., the variant of ViViT using all-to-all spatiotemporal attention, while still achieving high accuracies and has thus been" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "adopted as the building block for recent state-of-the-art architectures on various tasks [6, 19, 42, 46-48]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "content": "To address the challenge of reducing training time and memory usage without compromising the sophistication and accuracy of the original model, our approach is based on the simple idea of freezing the spatial backbone. Freezing the spatial backbone has many advantages: by not backpropagating through this transformer, training is faster and requires less memory (allowing for the model to handle more frames). We also inherit the benefits of pretraining the spatial transformer on a large dataset (such as JFT [20]). Naively implemented, however, we show that this approach falls very short in accuracy. Instead, with a few simple (but important) tweaks to the above idea, we propose a method that has the same advantages of freezing the spatial transformer, but does not compromise on accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "content": "Our method proceeds in two stages. In the first stage we pretrain a cheap version of the model using fewer frames, e.g., 8 frames as opposed to, e.g., 32 frames. In the second stage we fine tune this model with more frames, which is more expensive but in this stage we freeze the spatial encoder and introduce a compact \"adapter\" model connecting frozen spatial representations to the temporal transformer, negating the need for end-to-end training of the spatial transformer. Critically this includes pre-training the temporal transformer (by initializing from stage 1) which is often overlooked in current video models which typically initialize this component from scratch. However our experiments show that this step is critical if we wish to not sacrifice performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "type": "text", + "content": "With our training recipe, we match or slightly outperform conventional training of ViViT at roughly half the cost as seen in Figure 1. A notable benefit of our training recipe, is its ability to process up to 80 frames (ViViT-B) on typical university-grade GPUs, a significant leap from the previous capacity of 16 frames. This expansion in processing power broadens the range of video data manageable under resource-constrained settings. As we elaborate in Section 4.7, our research underscores the potential to democratize access to advanced video transformer models. Another notable benefit is the model's ability to now use even larger models as the spatial transformer, we introduce ViViT-g as seen in Section 4.6. This accessibility paves the way for future video action recognition research, irrespective of resource constraints. We also show in Section 4.9 that our method generalizes to non-ViViT models such as Dualformer [26] and VideoSwin [29]. Hereafter, we refer to our version of ViViT as SFA-ViViT, where SFA denotes 'Spatial Frozen and Adapter Initialized'." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 578, + 237, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 237, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 237, + 592 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Transformers for Videos Action recognition is a key research area in computer vision, addressed by many traditional [24, 41] and CNN based approaches [5, 16, 17, 21, 27, 36, 39, 44] aided by the release of large-scale datasets [5, 23, 37]. Since we focus on transformer based architectures, a thorough review of earlier methods are out of this scope. More recently, the transformer architecture, initially devel" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 548 + ], + "type": "text", + "content": "oped for NLP tasks [40], has been adapted for video understanding and action recognition tasks, leading to state-of-the-art models such as TimeSformer [2], ViViT [1], VideoSwin [29], and Uniformer [25] These transformer-based models leverage self-attention mechanisms to capture complex spatiotemporal patterns in action recognition tasks. TimeSformer [2] is one of the first transformer-based models for video understanding, adapting the transformer architecture to video by treating it as a sequence of flattened image patches. ViViT [1] integrates spatial and temporal transformers to efficiently capture spatiotemporal information in video sequences. VideoSwin [29] is a hierarchical transformer that applies local windowing for efficiency, enabling the model to handle longer video sequences. Similar to VideoSwin, Dualformer [26] also adopts a hierarchical architecture but uses a local-global stratified transformer design to capture both fine-grained and global video features efficiently. More recently, Uniformer [25] integrates 3D convolution and spatiotemporal self-attention, MTV [47] proposes a multi-view transformer model using distinct encoders for each video \"view\", improving accuracy as the number of views increases. The Multiscale Vision Transformers (MViT) [11] model streamlines computation and memory usage by operating at different resolutions, focusing on high-level features at lower resolutions and low-level details at higher ones, effectively leveraging both spatial and temporal information in visual tasks. TubeViT [33] introduces a method of sparsely sampling different-sized 3D segments from videos, facilitating efficient joint image and video learning, and allowing the adaptation of larger models to videos with less computational resources. Typically these models have FLOPs in the range of TFLOPs and training times that last more than days on the largest of GPUs/TPUs available, making them infeasible to train or use in lower resourced settings such as academia. It is critical that we find a way to train these models with limited resources while maintaining their performance. To this end, we focus on the factorised encoder version of ViViT as the late-fusion approach followed is used as a foundation for state-of-the-art approaches of various tasks [6, 19, 42, 46–48] and hence believe that the initialization scheme proposed can be used for future methods working on similar architectures. Similar to the approach for ViViT factorized encoder, the hierarchical architectures of VideoSwin and DualFormer can also benefit from staged training. By freezing the early layers that capture local spatial-temporal features, and adding an adapter before fine-tuning the later layers focused on global context, the overall training can be made more efficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 483, + 666 + ], + "type": "text", + "content": "Efficient Transformers in Videos Efficiency is a nuanced topic [9], as there are multiple cost indicators of efficiency (for example, GFLOPs, inference time, training time, memory usage), and models which improve efficiency in one dimension, are not necessarily better in other dimensions [9]. TokenLearner [35] proposes a method that adaptively learns tokens for efficient image and video understanding tasks, enabling effective modeling of pairwise attention over longer temporal horizons or spatial content. TokenLearner reduces the GFLOPs required by ViViT by about half, but does not significantly change the training time or the inference time of ViViT. Spatial Temporal Token Selection (STTS) [43] proposes" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 390 + ], + "type": "text", + "content": "a dynamic token selection framework for spatial and temporal dimensions that ranks token importance using a lightweight scorer network, selecting top-scoring tokens for downstream evaluation in an end-to-end training process. STTS again reduces the GFLOPs, but the training time and inference time do not change significantly. TokShift [49], a zero-parameter, zero-FLOPs operator that models temporal relations in transformer encoders by temporally shifting partial token features across adjacent frames but again requires the same training time as the original model. By densely integrating TokShift into a plain 2D vision transformer, a computationally efficient, convolution-free video transformer is created for video understanding. Most similar to our work is the ST-Adapter [32], that utilizes built-in spatio-temporal reasoning in a compact design, allowing pretrained image models to reason about dynamic video content with a small pertask parameter cost, surpassing existing methods in both parameter-efficiency and performance. However, it does not change FLOPs or inference time at all. Unlike ST-Adapter, we use a spatial only adapter which we show is enough to reproduce the performance of the baseline model at close to half the training time. In particular, our proposed method improves the training time and training memory usage, addressing the key problem of researchers and practitioners being able to train video models. It does not, however, change the inference time compared to a standard ViViT model. We consider overall train time for the same hyperparameters and use the same hardware for a direct comparison. We consider efficiency in this paper as the time saved in the overall training of the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 408, + 233, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 233, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 233, + 422 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 431, + 246, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 246, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 246, + 443 + ], + "type": "text", + "content": "3.1 Revisiting ViViT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 449, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 481, + 521 + ], + "type": "text", + "content": "The Video Vision Transformer (ViViT) extends the Vision Transformer architecture to handle video data by incorporating spatio-temporal reasoning. The idea behind ViViT is to process video input as a sequence of image patches, combining spatial and temporal information through a series of transformer layers, which include multi-head self-attention, layer normalization, and feed-forward networks. The output is used for video classification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 522, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 482, + 665 + ], + "type": "text", + "content": "In the \"vanilla\" variant of ViViT, one extracts spatio-temporal tokens from a video then forwards all tokens through a transformer encoder which explicitly models all pairwise interactions between all spatio-temporal tokens. We build off of the more efficient \"Factorized Encoder\" variant of ViViT whose architecture consists of two separate transformer encoders, a spatial transformer modeling interactions between tokens from the same temporal index and a temporal transformer modeling interactions between tokens from different temporal indices. Despite having more parameters, it requires fewer floating point operations (FLOPs) than vanilla ViViT. Because the Factorised Encoder variant strikes a good balance point between accuracy and processing speed, it has also been adopted as the foundation for other architectures [6,19,42,46-48], reinforcing its utility and robustness." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 269, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 269, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 269, + 129 + ], + "type": "text", + "content": "3.2 Our training strategy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 196 + ], + "type": "text", + "content": "We concentrate on the factorised encoder variant of ViViT as it is already the most efficient version of the baseline. Henceforth, when we talk about ViViT we refer to this variant of ViViT. Consider the ViViT model that contains a spatial transformer with parameters " + }, + { + "bbox": [ + 130, + 135, + 482, + 196 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 135, + 482, + 196 + ], + "type": "text", + "content": " and a temporal transformer with parameters " + }, + { + "bbox": [ + 130, + 135, + 482, + 196 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{temporal}}" + }, + { + "bbox": [ + 130, + 135, + 482, + 196 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 205, + 481, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 205, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 220, + 205, + 481, + 217 + ], + "type": "interline_equation", + "content": "X _ {\\text {s p a t i a l}} = T _ {\\text {s p a t i a l}} \\left(X _ {\\text {i n}}; \\theta_ {\\text {s p a t i a l}}\\right) \\tag {1}", + "image_path": "3f5bc7400f4dc3c3cb4cd15ea6a35b5025746acd1adba905fa8a5d13ba5bacb6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 220, + 394, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 220, + 394, + 232 + ], + "spans": [ + { + "bbox": [ + 233, + 220, + 394, + 232 + ], + "type": "interline_equation", + "content": "X _ {o u t} = T _ {t e m p o r a l} (X _ {s p a t i a l}; \\theta_ {t e m p o r a l}).", + "image_path": "136ba163a9fe23607db21e8607ee910a531c553f2a67ce3889a29aa65ebedb31.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 240, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 479, + 323 + ], + "type": "text", + "content": "In conventional ViViT training, " + }, + { + "bbox": [ + 130, + 240, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 240, + 479, + 323 + ], + "type": "text", + "content": " is initialized from an image pre-trained checkpoint such as ImageNet-21k [34] or JFT [20] and the " + }, + { + "bbox": [ + 130, + 240, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{temporal}}" + }, + { + "bbox": [ + 130, + 240, + 479, + 323 + ], + "type": "text", + "content": " is initialized from scratch. During backpropagation, the gradient flows through the entire model. This entails training two sizable transformer models end-to-end, which is a highly resource-intensive process, as the transformer architecture is inherently computationally demanding, especially with more frames and larger ViViT variants (e.g., ViViT-H)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "text", + "content": "One approach to reducing training time is to freeze the parameters of the spatial transformer " + }, + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "text", + "content": ". By not backpropagating through " + }, + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "text", + "content": ", gradient updates are faster and require less memory, allowing us to access more frames without encountering out-of-memory issues. But as we show in experiments, the accuracy of the resulting model with frozen " + }, + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 324, + 480, + 396 + ], + "type": "text", + "content": " is not competitive (in accuracy) with the baseline training approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 433 + ], + "type": "text", + "content": "We present a two stage approach (see Fig. 2) to training ViViT models that inherits the same benefits of freezing the spatial transformer, while not compromising on model quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "text", + "content": "Stage 1. In Stage 1, we pretrain our ViViT model on a reduced number of frames initializing the spatial transformer using a pre-trained image checkpoint. We do not freeze the spatial transformer during this stage, but critically, Stage 1 serves to also initialize the temporal transformer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 491, + 479, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 479, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 479, + 539 + ], + "type": "text", + "content": "To set the number of frames at this stage, we must balance the goal of efficiency (using fewer frames) against our finding in experiments that pre-training on too few frames can lead to suboptimal results. In our ablations, we identify a sweet spot at 8 frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 551, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 612 + ], + "type": "text", + "content": "Stage 2. In Stage 2, we fine tune our ViViT model on the full frame count (e.g. 128 frames) initializing both spatial and temporal transformer parameters learned in Stage 1. Because this stage is significantly more expensive, in stage 2, we freeze the spatial transformer parameters " + }, + { + "bbox": [ + 130, + 551, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 551, + 482, + 612 + ], + "type": "text", + "content": " and add a lightweight adapter module with parameters " + }, + { + "bbox": [ + 130, + 551, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{adapter}}" + }, + { + "bbox": [ + 130, + 551, + 482, + 612 + ], + "type": "text", + "content": " following the spatial transformer:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 222, + 620, + 360, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 620, + 360, + 633 + ], + "spans": [ + { + "bbox": [ + 222, + 620, + 360, + 633 + ], + "type": "interline_equation", + "content": "X _ {s p a t i a l} = T _ {s p a t i a l} (X _ {i n}; \\theta_ {s p a t i a l})", + "image_path": "e88ece2b67a88acad4321bdfb732e2c31b1fc7e22658a66d1273f8b3cc21c04c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 219, + 635, + 480, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 635, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 219, + 635, + 480, + 647 + ], + "type": "interline_equation", + "content": "X _ {\\text {a d a p t e r}} = A _ {\\text {a d a p t e r}} \\left(X _ {\\text {s p a t i a l}}; \\theta_ {\\text {a d a p t e r}}\\right) \\tag {2}", + "image_path": "99e65519e6482cd1cc70fabbf4c521d8422450d6510349595142cf4ff2260f42.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 235, + 650, + 395, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 650, + 395, + 663 + ], + "spans": [ + { + "bbox": [ + 235, + 650, + 395, + 663 + ], + "type": "interline_equation", + "content": "X _ {o u t} = T _ {t e m p o r a l} (X _ {a d a p t e r}; \\theta_ {t e m p o r a l})", + "image_path": "0b4dab62bf761b14cabb8f6e1054aefdcb8d01e11e35e7afb99827fb31a44a9c.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "In this setting, by backpropagating only through the temporal transformer and the lightweight adapter module (in our experiments, a two layer MLP), we effectively cut total training time by half." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 171, + 433, + 318 + ], + "blocks": [ + { + "bbox": [ + 181, + 171, + 433, + 318 + ], + "lines": [ + { + "bbox": [ + 181, + 171, + 433, + 318 + ], + "spans": [ + { + "bbox": [ + 181, + 171, + 433, + 318 + ], + "type": "image", + "image_path": "9684e11d88c86e01c18153a848d98f9e34482da6431f64f0305b8a0bf38b378f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 327, + 482, + 394 + ], + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 394 + ], + "type": "text", + "content": "Fig. 2: STAGE 1: We first use the full ViViT-FE model on 8 frames by initializing the spatial transformer from an image checkpoint and the temporal transformer from scratch. STAGE 2: We then use this as our checkpoint to initialize the spatial and temporal transformer for models using more frames (such as 32, 64 or 128). We then freeze the spatial transformer and add an adapter model to finetune spatial transformer features. The temporal transformer is finetuned from the same checkpoint." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "content": "The crucial finding here is that the spatial transformer requires only short-term context for initialization (after which it remains frozen), whereas the temporal transformer necessitates long-term context to achieve its optimal performance. Further details and empirical analysis can be found in the next section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 483, + 289, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 289, + 497 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 289, + 497 + ], + "type": "text", + "content": "4 Experimental Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 593 + ], + "type": "text", + "content": "Through a series of comprehensive experiments which we now present, we investigate the significance of the spatial transformer, examining the impact of pre-training datasets and how larger models affect action recognition performance. We also explore the importance of initializing the temporal transformer by employing various initialization schemes and datasets, assessing whether the number of frames is critical for initializing larger models, initializing full ViViT models, and initializing models on one dataset while fine-tuning on another." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 609, + 205, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 205, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 205, + 621 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "We evaluate on all the datasets considered in [1] (specifically, Kinetics-400 [5], Kinetics-600 [4], EPIC-Kitchens [8], Something-something v2 [18] and Moments-in-time [31]) as well as the Something-Else [30] dataset. As these datasets are" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "common in the community, we include further details in the supplementary along with the implementation details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 157, + 239, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 157, + 239, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 239, + 169 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 236 + ], + "type": "text", + "content": "We first address two critical aspects: the significance of fine-tuning the spatial transformer and the importance of initializing the temporal transformer. To do so, we conduct a series of experiments in various scenarios, which are detailed below. Our analysis focuses on the Something-something dataset, utilizing the large version of the ViViT model, referred to as ViViT-L." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": "We examine four main elements that modify the structure of conventional ViViT training and these are mentioned with indices in Table 1 namely: I. The freezing of the spatial transformer (" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": " is initialized and then frozen), II. The freezing of the temporal transformer (" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{temporal}}" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": " is frozen), III. The addition of an adapter (lightweight module with parameters " + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{adapter}}" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": "), IV. Next, we initialize the temporal transformer using VideoMAE [38], while keeping the spatial transformer frozen and the adapter incorporated and V. The initialization of the temporal transformer (" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{spatial}}" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{temporal}}" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": " are initialized using the ViViT" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "inline_equation", + "content": "_{8f}" + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 379 + ], + "type": "text", + "content": "It is important to note that the VideoMAE training is an extremely expensive process as can be seen in the table. But combined with the line below it, these two models, which significantly outperform lines I, II and III, show that properly initializing the temporal transformer is critical." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 137, + 399, + 478, + 484 + ], + "blocks": [ + { + "bbox": [ + 137, + 399, + 478, + 484 + ], + "lines": [ + { + "bbox": [ + 137, + 399, + 478, + 484 + ], + "spans": [ + { + "bbox": [ + 137, + 399, + 478, + 484 + ], + "type": "table", + "html": "
IndexSpatialFrozenAdapterTemporalFrozenTemporal InitTop-1AccTop-5AccTrain Time
-××××××64.4587.4814.17 h
I✓×××××27.7556.730.5x()
II××✓×××25.8053.070.5x()
III✓✓××××38.7768.930.53x()
IV✓✓×VMAE58.5485.832.51x()
V✓✓×ViViT-8f63.8587.620.62x()
VI×✓×ViViT-8f64.1387.351.37x()
", + "image_path": "bc52d5cf3ce247c3369e396ad575c897678c71af26708647128262c32e3cfe3e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 529 + ], + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 529 + ], + "type": "text", + "content": "Table 1: Ablation study results illustrating the impact of various modifications to the ViViT-B model, including spatial and temporal transformer freezing, adapter addition, and initialization methods, on top-1 and top-5 accuracy. Dataset is Something-something v2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 594 + ], + "type": "text", + "content": "Additionally, initializing the spatial transformer yields further improvement. The adapter plays a vital role in augmenting performance when the spatial transformer is frozen, and due to its lightweight nature, it will be an essential component of our training methodology." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 610, + 389, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 389, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 389, + 623 + ], + "type": "text", + "content": "4.3 How many frames should we use for Stage 1?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Next we experiment with various frame counts for stage 1 training, We test seven variants: JFT [20] checkpoint (image-based), 2, 4, 8, 16, 32, and 48-frame ViViT checkpoints. We then fine-tune these with a frozen spatial transformer and add" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 289, + 240 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 289, + 240 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 289, + 240 + ], + "type": "image", + "image_path": "7d5c317ddb1f7b3a0aabbf997b8b620449519d929e656203aa8e93b4f0aafa1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 249, + 293, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 293, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 293, + 327 + ], + "type": "text", + "content": "Fig. 3: The effect of initializing with different numbers of frames (JFT, 2, 4, 8, 16, 32, and 48), freezing the spatial transformer and adding an adapter model and fine-tuning using 64, 96, and 128 frames. Results on Kinetics400 dataset, 'f' refers to frames." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 119, + 457, + 233 + ], + "blocks": [ + { + "bbox": [ + 315, + 119, + 457, + 233 + ], + "lines": [ + { + "bbox": [ + 315, + 119, + 457, + 233 + ], + "spans": [ + { + "bbox": [ + 315, + 119, + 457, + 233 + ], + "type": "image", + "image_path": "fe08a86088348e9a74136a447c7c9de0bf4d4007e5f6187b2ffcf33f2efd0040.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 246, + 465, + 322 + ], + "lines": [ + { + "bbox": [ + 304, + 246, + 465, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 465, + 322 + ], + "type": "text", + "content": "Fig.4: Comparison of our initialization method vs conventional training of ViViT on Top-1 accuracy and loss on the Kinetics-400 dataset using 64 and 128 frames. We see that our initialization gives a significant headstart to the models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 447 + ], + "type": "text", + "content": "an adapter model using 64, 96, and 128 frames (see Figure 3). Results show that using too few frames for Stage 1 training can underperform (with image-only initialization from a JFT [20] checkpoint performing the worst). Thus we deduce that short term temporal context is essential for initializing the spatial transformer. Performance also plateaus after 8 frames, and given that using more frames increases training time, we settle on using 8 frames as our \"sweet spot\" for Stage 1 training. We consider this an ablation, as this is done only on Kinetics and kept fixed for all models and datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 443, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 443, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 443, + 478 + ], + "type": "text", + "content": "4.4 Does the proposed training increase convergence speed?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 484, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 605 + ], + "type": "text", + "content": "Another potential question that may arise concerns the impact of this initialization method in terms of convergence speed, if any. This specific aspect holds considerable significance owing to its potential ability to drastically curtail the duration of time required for training and the number of epochs necessary to effectively train the model. Moreover, an important element to take into account is the effect of freezing the spatial transformer. This approach decreases the memory needed to store the model but also considerably enhances the training speed. To provide a clearer picture, we have plotted the validation curves with and without initialization, which can be seen in Figure 4. Note that with the proposed initialization, we get a significant head start in overall accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 621, + 481, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 481, + 634 + ], + "type": "text", + "content": "4.5 What about initializing on one dataset and finetuning on other?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In our study, we find training an 8-frame version of the standard ViViT model affordable. We consider training this on a temporally dependent dataset like" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "type": "text", + "content": "Something-something, then fine-tuning on other datasets like Kinetics-400. We examine two scenarios: first, the standard ViViT model trained on Kinetics-400 using 8-frames, and second, the same model trained on Something-something. Post-training, we freeze the spatial transformer, add an adapter, and fine-tune the models on the alternate dataset with more frames. We contrast this with models fine-tuned on their original datasets (see Table 2). Results favor initializing larger frame models from the 8-frame version on the same dataset. Thus, for the final comparison in Sec 4.8, we initialize models with 8-frame versions of the baseline." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 148, + 224, + 323, + 277 + ], + "blocks": [ + { + "bbox": [ + 148, + 224, + 323, + 277 + ], + "lines": [ + { + "bbox": [ + 148, + 224, + 323, + 277 + ], + "spans": [ + { + "bbox": [ + 148, + 224, + 323, + 277 + ], + "type": "table", + "html": "
CheckpointSSv2K400
K400-init44.71/74.5382.81/93.98
SSv2-init63.85/87.6276.79/92.35
", + "image_path": "60a2537233231d7a0faec47811afb5b6b9cad60246f39545051dbe77dac39e58.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 224, + 477, + 281 + ], + "blocks": [ + { + "bbox": [ + 328, + 224, + 477, + 281 + ], + "lines": [ + { + "bbox": [ + 328, + 224, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 328, + 224, + 477, + 281 + ], + "type": "table", + "html": "
BackboneTop-1Top-5Steps
ViViT-L79.6491.7348k steps
ViViT-H81.0293.0939k steps
ViViT-g81.8194.5529k steps
", + "image_path": "ac4e06bac995354864e24d6aa527455d3db9bf762d500cb7f49e40018c43495c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 284, + 306, + 338 + ], + "lines": [ + { + "bbox": [ + 146, + 284, + 306, + 338 + ], + "spans": [ + { + "bbox": [ + 146, + 284, + 306, + 338 + ], + "type": "text", + "content": "Table 2: A summary of cross-dataset initialization of the proposed model and performance comparison. We use Kinetics400 and Something-something v2 as our datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 319, + 290, + 479, + 366 + ], + "lines": [ + { + "bbox": [ + 319, + 290, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 319, + 290, + 479, + 366 + ], + "type": "text", + "content": "Table 3: A comparison of top-1 and top-5 accuracies for the ViViT-g model with the proposed training strategy, which incorporates a larger spatial transformer backbone. All models use 48 frames for fair comparison. Results are on Kinetics400 dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 383, + 416, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 416, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 416, + 396 + ], + "type": "text", + "content": "4.6 Extending the image backbone to 1.5B parameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 571 + ], + "type": "text", + "content": "An intriguing consequence of our approach is the ability to incorporate larger backbones into the spatial transformer, made possible by the additional memory available to us as a result of freezing the spatial transformer during training. Consequently, we introduce ViViT-g, which integrates the ViT-g model (with 1.5B parameters) as its backbone. To ensure a fair comparison, we focus solely on training and inference using 48 frames, and abstain from employing multiview or multicrop testing. Our objective is to investigate the potential impact of a more substantial spatial transformer backbone on the overall performance and show the potential of larger spatial backbones that are possible due to our training process. It is essential to note that the full ViViT-g model could not process more than 8 frames due to memory limitations. However, our proposed strategy allows processing up to 48 frames. A comparison of accuracies is presented in Table 3 along with the number of steps needed to reach the best performance. Dataset used is Kinetics-400 and all the ViT checkpoints are JFT-pretrained [20]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 473, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 473, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 473, + 611 + ], + "type": "text", + "content": "4.7 Comparison of Memory Usage with Standard ViViT Training and Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this study, we compare the number of frames that can be accessed using the standard ViViT training scheme against our proposed scheme, employing a set of 64 v3 TPUs that have 16 GB each. We further evaluate the performance of ViViT variants, including H, and g, in comparison with the SFA-ViViT using the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "content": "same variant configurations. Maintaining identical hyperparameters, we ensure a local batch size of 1. Our findings indicate that the conventional ViViT training approach restricts frame accessibility to 96 frames for the ViViT-H model, and a mere 8 frames for the ViViT-g model, before reaching memory limitations. Conversely, our proposed method enables access to 128 frames for ViViT-H, and up to 48 frames when utilizing ViViT-g with the same hardware. Furthermore, we investigate the impact of utilizing university grade GPUs by conducting ViViT experiments on an NVIDIA RTX 2080 Ti GPU farm equipped with 8 GPUs having 12 GB each. Under these circumstances, ViViT can only process 16 frames using a local batch size of 1. However, our proposed training strategy enables a notable improvement, expanding the frame capacity to 80 frames (ViViT-B) helping us reproduce ViViT results on lower end GPUs. This enhancement provides a valuable opportunity for researchers with limited resources to attain performance levels comparable to those with extensive resources. We show a comparison of number of frames accessible with and without our training recipe in Table 4." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 228, + 326, + 386, + 392 + ], + "blocks": [ + { + "bbox": [ + 228, + 326, + 386, + 392 + ], + "lines": [ + { + "bbox": [ + 228, + 326, + 386, + 392 + ], + "spans": [ + { + "bbox": [ + 228, + 326, + 386, + 392 + ], + "type": "table", + "html": "
Model481632486496128
ViViT-H✓✓✓✓✓✓✓×
SFA-ViViT-H✓✓✓✓✓✓✓✓
ViViT-g✓✓××××××
SFA-ViViT-g✓✓✓✓✓×××
", + "image_path": "4adfc635ed495e70e89c018e4ea64ab1d8e0e2fc5b525e58e6bcdae25bd21961.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 426 + ], + "type": "text", + "content": "Table 4: Memory usage in different ViViT training schemes is compared using the Kinetics400 dataset on 64 TPUs v3 with 16GB memory each. A " + }, + { + "bbox": [ + 130, + 392, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 130, + 392, + 482, + 426 + ], + "type": "text", + "content": " indicates accessible frames given hardware constraints, while a " + }, + { + "bbox": [ + 130, + 392, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 392, + 482, + 426 + ], + "type": "text", + "content": " signals an out-of-memory (OOM) error." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 479, + 428, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 428, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 428, + 491 + ], + "type": "text", + "content": "4.8 Comparison on all benchmarks to the baseline model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In this section, we present a comprehensive comparative analysis, focusing on the proposed approach and the baseline model. We report the Top-1 accuracy, Top-5 accuracy and the overall training time. The evaluation is conducted on the large and huge variants of ViViT across three datasets, namely Kinetics400, Kinetics600, and Moments in Time (MiT), with the summarized results tabulated in Table 5. The findings indicate a slight enhancement in accuracy for both Kinetics400 and Kinetics600 datasets, whereas a notable " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " increase in top-1 accuracy is observed for the MiT dataset using the proposed method. Furthermore, the proposed approach showcases a significant reduction in training time, accounting for approximately " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " of the original duration. This reduction emphasizes the advantageous nature of the proposed approach. To calculate the total training time for the SFA version, the train time of the 8 frame (Stage 1) ViViT model is combined with the train time of the (Stage 2) SFA-ViViT model. Conversely, the total training time for the standard ViViT encompasses" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "the total train time for the same number of frames that SFA-ViViT is trained on for fair comparison." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 136, + 162, + 478, + 224 + ], + "blocks": [ + { + "bbox": [ + 136, + 162, + 478, + 224 + ], + "lines": [ + { + "bbox": [ + 136, + 162, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 136, + 162, + 478, + 224 + ], + "type": "table", + "html": "
ModelKinetics-400Kinetics-600Moments in Time
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f82.59/93.091x(21.57 h)83.29/95.821x(26.14 h)--
ViViT-L128f + SFA8f82.78/94.030.56x()83.47/95.290.56x()--
ViViT-H96f84.21/94.661x(56.71 h)84.18/95.681x(60.45 h)38.17/62.841x(110.79 h)
ViViT-H96f + SFA8f84.42/94.720.57x()84.39/96.200.57x()39.96/64.390.59x()
", + "image_path": "c6b4ebaa83ad040738e284e4785b496332f88ec54c2baa6098bdcdb511f18b6c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 267, + 478, + 312 + ], + "blocks": [ + { + "bbox": [ + 130, + 224, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 258 + ], + "type": "text", + "content": "Table 5: Performance Comparison of various versions of ViViT with the proposed training strategy for Kinetics-400, Kinetics-600 and Moments in Time. Accuracies listed as Top-1/Top-5." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 267, + 478, + 312 + ], + "lines": [ + { + "bbox": [ + 136, + 267, + 478, + 312 + ], + "spans": [ + { + "bbox": [ + 136, + 267, + 478, + 312 + ], + "type": "table", + "html": "
ModelSomething-somethingSomething-ElseEpic-Kitchens
AccuracyTrain TimeAccuracyTrain TimeAccuracyTrain Time
ViViT-L128f64.45/87.481x(14.17 h)53.14/73.981x(3.84 h)43.53/56.55/65.401x(5.61 h)
ViViT-L128f + SFA8f63.85/87.620.62x()53.60/74.470.62x()43.54/56.78/65.160.63x()
", + "image_path": "8d4be734e59d8fbc1a5dd60a936b51cb5e86714749c884823e9fe4f78f8a6912.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 480, + 345 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 345 + ], + "type": "text", + "content": "Table 6: Performance Comparison of ViViT-L with the proposed training strategy for Something-something v2, Something-Else and Epic-Kitchens. Accuracies listed as Top-1/Top-5, for Epic Kitchens Top-1 noun-verb/ Top-1 noun/ Top-1 Verb." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "text", + "content": "We further examine the performance of ViViT-L incorporating our proposed training strategy in comparison to the original version on three additional datasets: Something-something, Something-Else, and Epic-Kitchens. A consistent trend is observed, with the modified approach outperforming the baseline model, at only a " + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "text", + "content": " cost of the baseline training time. In summary, our proposed training strategy demonstrates promising potential by yielding comparable or slightly improved performance across all datasets. This is obtained while maintaining a training cost ranging from " + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 130, + 380, + 482, + 487 + ], + "type": "text", + "content": " of the original model, thus highlighting its effectiveness. Results can be seen in Table 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 289, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 289, + 521 + ], + "type": "text", + "content": "4.9 Beyond the ViViT model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We adapted our approach to tiny versions of Dualformer [26], VideoSwin [29] and Video-FocalNets [45] on the Kinetics 400 dataset, demonstrating up to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " faster training with minimal accuracy loss. Thus, our method exhibits consistent improvements in training efficiency across multiple video architectures. We believe these comparisons better showcase the generalized benefits of our freeze-and-finetune approach for efficient video model training. These results can be seen in Table 7. While these models do not directly fall under the factorized encoder models, we can follow a similar pipeline to train them under the SFA training regime. For the Dualformer-T and VideoSwin-T, we first train 8-frame versions of these models and then freeze the first 3 blocks of the models (Stage 1 to Stage 3) and only finetune the last block." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 114, + 430, + 269 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 430, + 269 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 430, + 269 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 430, + 269 + ], + "type": "table", + "html": "
ModelTotal Train Time Accuracy
Dualformer-T32f1x(30.72 h)79.53
Dualformer-T32f-SFA8f0.57x()79.48
Dualformer-B32f1x(68.57 h)81.07
Dualformer-B32f-SFA8f0.59x()81.12
VideoSwin-T32f1x(41.45 h)78.62
VideoSwin-T32f-SFA8f0.52x()78.68
VideoSwin-B32f1x(115.23 h)80.45
VideoSwin-T32f-SFA8f0.53x()80.52
Video-FocalNet-T32f1x(11.28 h)79.76
Video-FocalNet-T32f-SFA8f0.52x()79.95
Video-FocalNet-B32f1x(28.84 h)83.55
Video-FocalNet-B32f-SFA8f0.51x()84.25
", + "image_path": "88fe4d8e721a36d6b2d93f582c1579d97f3a804b5bd46df2cbe12e2a822885f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 339, + 416, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 416, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 416, + 353 + ], + "type": "text", + "content": "4.10 Comparing with other efficient training strategies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 495 + ], + "type": "text", + "content": "While we have already mentioned that strategies such as ST-Adapter, STTS etc are efficient in parameters and GFLOPs and not in training time. We demonstrate the results of training these in comparison to the proposed training method in terms of training time and inference time. The results demonstrate that our proposed SFA training pipeline confers substantial training time savings compared to these methods. For example, SFA-ViViT-B attains similar accuracy as ST-Adapter at approximately 0.58 times the total training time on Kinetics-400. We would like to point out that only STTS does better than the proposed strategy in terms of throughput, but as shown earlier requires twice the amount of time to train the model. All methods use 32 frames. These results are reported in Table 8." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 164, + 517, + 451, + 595 + ], + "blocks": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "Table 7: Training efficiency improvements from SFA approach on tiny Dualformer and VideoSwin models. Our method achieves " + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "inline_equation", + "content": ">2\\mathrm{x}" + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": " faster training with minimal accuracy drop in Kinetics-400." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 517, + 451, + 595 + ], + "lines": [ + { + "bbox": [ + 164, + 517, + 451, + 595 + ], + "spans": [ + { + "bbox": [ + 164, + 517, + 451, + 595 + ], + "type": "table", + "html": "
ModelTotal Train Time ThroughputAccuracy
ViT-B + ST-Adapter [32]1x(34.18 hours)9082.0
ViViT-B [1]1.03x()8882.6
STTS [43]1.07x()12980.7
SBP [7]0.92x()8982.2
SFA8f-ViViT-B32f (Ours)0.58x()8882.8
", + "image_path": "fe067e5a51c6dcb0c88f3df7fd45cb4f6fbbafc3eae7ccef52cb19028f4f3198.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 595, + 482, + 628 + ], + "lines": [ + { + "bbox": [ + 130, + 595, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 482, + 628 + ], + "type": "text", + "content": "Table 8: Training efficiency improvements from SFA approach in comparison to other efficient training strategies. We require approximately half of the training time in comparison to other efficient training strategies." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 288, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 288, + 127 + ], + "type": "text", + "content": "5 Evaluations Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 144, + 481, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 144, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 144, + 481, + 204 + ], + "type": "text", + "content": "Our evaluations on ViViT was not to obtain SOTA performance using ViViT, but to show the reduced time required to train the model. In the ViViT paper [1], the reported results are different from what we report. This is due to some videos being removed from the Kinetics dataset. We re-run ViViT using the official codebase and report results on the datasets using this." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 481, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 481, + 336 + ], + "type": "text", + "content": "The point of using 128 frames was not to show that using more frames improves performance, but instead to show that for larger models, using our training strategy we can actually access more frames due to memory being saved. The logic behind the training strategy holds for lower number of frames such as 32 as well. Hence, using multiview inference would still save time. We would still need only " + }, + { + "bbox": [ + 133, + 205, + 481, + 336 + ], + "type": "inline_equation", + "content": "0.6x" + }, + { + "bbox": [ + 133, + 205, + 481, + 336 + ], + "type": "text", + "content": " of the overall training time. Also, as pointed out in Section 4.7, larger models such as ViViT-g can only access 8 frames using standard ViViT-FE training. However, using our approach, we can access 48 frames. A notable benefit of accessing more frames is in long-form video understanding tasks and while such experiments have not been conducted, this training approach promises improved results in the task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 360, + 218, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 360, + 218, + 372 + ], + "spans": [ + { + "bbox": [ + 133, + 360, + 218, + 372 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 390, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 390, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 390, + 481, + 665 + ], + "type": "text", + "content": "We have investigated the challenges posed by the substantial training time and memory consumption of video transformers, particularly focusing on the factorised encoder variant of the ViViT model as our baseline. To address these challenges, we proposed two effective strategies: utilizing a compact adapter model for fine-tuning image representations instead of end-to-end training of the spatial transformer, and initializing the temporal transformer using the baseline model trained with 8 frames. Our proposed training strategy has demonstrated the potential to significantly reduce training costs and memory consumption while maintaining, or even slightly improving, performance compared to the baseline model. Furthermore, we observed that with proper initialization, our baseline model can achieve near-peak performance within the first " + }, + { + "bbox": [ + 133, + 390, + 481, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 133, + 390, + 481, + 665 + ], + "type": "text", + "content": " of training epochs. While we focus on the factorized encoder variant of ViViT, we also show the generalization power of the training regime by comparing results of DualFormer, VideoSwin and Video-FocalNets. We also show that the proposed method requires approximately half the time required in comparison to other efficient training strategies. The advancements made in this work could help research in the video understanding domain by enabling access to more frames and the utilization of larger image models as the spatial transformer, all while maintaining the same memory consumption. Our findings provide valuable insights for researchers and practitioners with limited resources, paving the way for more efficient and scalable alternatives in action recognition. Future work may focus on further optimizing and refining these strategies, and exploring their application to other video transformer architectures." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "text", + "content": "1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 238 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "type": "text", + "content": "4. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. arXiv preprint arXiv:1808.01340 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "text", + "content": "6. Chen, J., Ho, C.M.: Mm-vit: Multi-modal video transformer for compressed video action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1910-1921 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 369 + ], + "type": "text", + "content": "7. Cheng, F., Xu, M., Xiong, Y., Chen, H., Li, X., Li, W., Xia, W.: Stochastic backpropagation: A memory efficient strategy for training video models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8301-8310 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 414 + ], + "type": "text", + "content": "8. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 720-736 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "text", + "content": "9. Dehghani, M., Arnab, A., Beyer, L., Vaswani, A., Tay, Y.: The efficiency misnomer. arXiv preprint arXiv:2110.12894 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 435, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 479 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "text", + "content": "1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "2. Foteinopoulou, N.M., Patras, I.: Emoclip: A vision-language method for zero-shot video facial expression recognition. arXiv preprint arXiv:2310.16640 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "text", + "content": "3. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "4. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "text", + "content": "5. Gowda, S.N., Hao, X., Li, G., Sevilla-Lara, L., Gowda, S.N.: Watt for what: Rethinking deep learning's energy-performance relationship. arXiv preprint arXiv:2310.06522 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "6. Gowda, S.N., Rohrbach, M., Keller, F., Sevilla-Lara, L.: Learn2augment: Learning to composite videos for data augmentation in action recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI. pp. 242-259. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "17. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 203 + ], + "type": "text", + "content": "18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "text", + "content": "19. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lučić, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "text", + "content": "20. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "21. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence 35(1), 221-231 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "text", + "content": "22. Kim, K., Gowda, S.N., Mac Aodha, O., Sevilla-Lara, L.: Capturing temporal information in a single frame: Channel sampling strategies for action recognition. arXiv preprint arXiv:2201.10394 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "type": "text", + "content": "23. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "24. Laptev, I.: On space-time interest points. International journal of computer vision 64, 107-123 (2005)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "25. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "text", + "content": "26. Liang, Y., Zhou, P., Zimmermann, R., Yan, S.: Dualformer: Local-global stratified transformer for efficient video recognition. In: European Conference on Computer Vision. pp. 577-595. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "type": "text", + "content": "27. Lin, J., Gan, C., Han, S.: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7083-7093 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 514 + ], + "type": "text", + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "29. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 546, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 480, + 589 + ], + "type": "text", + "content": "30. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "31. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "32. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 269, + 101 + ], + "type": "text", + "content": "Shreyank N Gowda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 654 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "33. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. arXiv preprint arXiv:2212.03229 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "text", + "content": "34. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "text", + "content": "35. Ryoo, M., Piergiovanni, A., Arnab, A., Dehghani, M., Angelova, A.: Tokenlearner: Adaptive space-time tokenization for videos. Advances in Neural Information Processing Systems 34, 12786-12797 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "36. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "37. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "38. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "39. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 303, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 335 + ], + "type": "text", + "content": "40. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "text", + "content": "41. Wang, H., Kläser, A., Schmid, C., Liu, C.L.: Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision 103, 60-79 (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "42. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. Transactions of Machine Learning Research" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 446 + ], + "type": "text", + "content": "43. Wang, J., Yang, X., Li, H., Liu, L., Wu, Z., Jiang, Y.G.: Efficient video transformers with spatial-temporal token selection. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 69-86. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "44. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: European conference on computer vision. pp. 20-36. Springer (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "45. Wasim, S.T., Khattak, M.U., Naseer, M., Khan, S., Shah, M., Khan, F.S.: Videofocalnets: Spatio-temporal focal modulation for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13778-13789 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "46. Xiong, X., Arnab, A., Nagrani, A., Schmid, C.: M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "47. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "text", + "content": "48. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR 2023-IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 654 + ], + "type": "text", + "content": "49. Zhang, H., Hao, Y., Ngo, C.W.: Token shift transformer for video classification. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 917-925 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_content_list.json b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98110802b101f0ded2b4de1746885f5f16245ada --- /dev/null +++ b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_content_list.json @@ -0,0 +1,1395 @@ +[ + { + "type": "text", + "text": "Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging", + "text_level": 1, + "bbox": [ + 292, + 140, + 710, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mahmoud Affifi, Zhenhua Hu, and Liang Liang", + "bbox": [ + 321, + 212, + 679, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google {mafifi, zhenhuahu, annyliang}@google.com", + "bbox": [ + 341, + 239, + 661, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. High dynamic range (HDR) imaging involves capturing a series of frames of the same scene, each with different exposure settings, to broaden the dynamic range of light. This can be achieved through burst capturing or using staggered HDR sensors that capture long and short exposures simultaneously in the camera image signal processor (ISP). Within camera ISP pipeline, illuminant estimation is a crucial step aiming to estimate the color of the global illuminant in the scene. This estimation is used in camera ISP white-balance module to remove undesirable color cast in the final image. Despite the multiple frames captured in the HDR pipeline, conventional illuminant estimation methods often rely only on a single frame of the scene. In this paper, we explore leveraging information from frames captured with different exposure times. Specifically, we introduce a simple feature extracted from dual-exposure images to guide illuminant estimators, referred to as the dual-exposure feature (DEF). To validate the efficiency of DEF, we employed two illuminant estimators using the proposed DEF: 1) a multilayer perceptron network (MLP), referred to as exposure-based MLP (EMLP), and 2) a modified version of the convolutional color constancy (CCC) to integrate our DEF, that we call ECCC. Both EMLP and ECCC achieve promising results, in some cases surpassing prior methods that require hundreds of thousands or millions of parameters, with only a few hundred parameters for EMLP and a few thousand parameters for ECCC.", + "bbox": [ + 261, + 301, + 738, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Computational color constancy $\\cdot$ Illuminant estimation $\\cdot$ HDR imaging", + "bbox": [ + 261, + 619, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction and Related Work", + "text_level": 1, + "bbox": [ + 215, + 672, + 558, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera image signal processor (ISP) comprises several modules, each dedicated to enhancing specific aspects of the quality of captured raw images by the camera sensor [16]. One key component among these modules is the white-balance module, which aims to eliminate undesirable color casts introduced by the combination of scene lighting and camera sensitivity. To achieve this, the auto white-balance module runs an illuminant estimator onboard the camera ISP to estimate the RGB color of the illuminant, under the assumption that a single global illuminant illuminates the scene for simplicity [30, 44]. Image white balancing can thus be described as follows:", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/624c42bbf13ed50c54d147571f7754ba36c46b57c60088198a8fc60090b33093.jpg", + "image_caption": [ + "Full-resolution raw image", + "Raw frames with different exposure times", + "Fig. 1: Conventional illuminant estimators often rely on a single frame for illuminant color estimation. Although the HDR camera pipeline includes at least two frames of the same scene, conventional methods usually consider only a single frame (either with auto exposure, shown in the dashed green line, merged frame, or one frame of the burst capture) [12,16,44,62]. This paper proposes leveraging information from dual-exposure capturing in multi-exposure HDR imaging to enhance illuminant estimation in camera pipelines. Our method uses frames captured at long and short exposures, available in multi-exposure bursts [28] or staggered HDR sensors [52, 58]. Achieving comparable or superior results, our method employs lightweight models ( $\\sim$ 300-6000 parameters) compared to those using hundreds of thousands or millions of parameters. In this figure and the following figures, all raw images have the gamma operator applied to aid visualization, and all sRGB images are rendered using the HDR+ camera pipeline [44]." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 401, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/adfd9146708b0ec3c480ce71a33c61177aee412bcc5386c04114f28be22b6875.jpg", + "image_caption": [ + "Dual-exposure frames with long exposure times" + ], + "image_footnote": [], + "bbox": [ + 401, + 143, + 470, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6fb6fabfbc74ff841dde9dddc176556fbfcf223f5f5d2cc28aaeb2491330dca5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 143, + 586, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c614014ad931b16fbcffe465d0a06c1479ebb18cc9a2ccf855600573a4b5d44.jpg", + "image_caption": [ + "Final rendered image with our white-balance correction" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 738, + 224 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {W B}} = \\operatorname {d i a g} \\left(\\left[ \\frac {\\hat {\\ell} _ {G}}{\\hat {\\ell} _ {R}}, 1, \\frac {\\hat {\\ell} _ {G}}{\\hat {\\ell} _ {B}} \\right] ^ {T}\\right) I, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 465, + 785, + 507 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\hat {\\ell} _ {R}, \\hat {\\ell} _ {G}, \\hat {\\ell} _ {B} \\right] ^ {T} = f (I), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 522, + 784, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $I$ and $I_{\\mathrm{WB}}$ are $3 \\times k$ RGB colors of raw image and white-balanced raw image, respectively, with $k$ refers to the total number of pixels in the image, $\\operatorname{diag}(.)$ creates a $3 \\times 3$ diagonal matrix from a given vector, $\\hat{\\ell} \\in \\mathbb{R}^3$ refers to the estimated illuminant color vector, $T$ refers to vector transpose, and $f$ is an illuminant estimator function.", + "bbox": [ + 212, + 551, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modern camera ISPs capture several frames of each scene with different exposure times to enhance final image dynamic range [28,38]. This can be achieved through burst capturing, involving rapid capture with varying exposure times in quick succession, or by using staggered high dynamic range (sHDR) sensors that simultaneously capture long and short exposures [52, 58]. The captured images are then combined, incorporating multiple exposures of the same scene at different levels, resulting in a greater dynamic range than what would be possible with a single image [43, 49, 51, 55]. This HDR imaging camera pipeline, therefore, includes more than a single frame of the captured scene, and internal camera ISP modules, such as the auto white-balance module, can access these additional frames. While such additional frames may have beneficial information to help illuminant estimators, conventional illuminant estimation methods often rely on a single frame (e.g., [9, 36, 41, 47, 54, 63, 64]) in both traditional single-frame camera pipelines [39] and multi-frame HDR ISPs [44].", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such single-frame illuminant estimators can be categorized into: 1) statistical methods that rely on statistics computed from input raw image information (e.g., colors, edges) [13, 14, 25, 31, 50, 56, 57] and 2) learning-based methods that learn, from a set of training images labeled with ground-truth illuminant colors, to map from input raw color information to the corresponding illuminant color [6, 8, 10, 11, 24, 36, 42, 45, 47, 48, 54, 63, 64]. While the latter category typically achieves better results than the statistical methods, most learning-based methods are camera-dependent, meaning that they require domain adaptation or fine-tuning when deployed on new cameras to achieve similar accuracy on cameras used for training due to the influence of the camera response function on both raw image colors and ground-truth colors [3-5, 35, 53, 59].", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A limited number of attempts have proposed learning-based methods that go beyond the single-frame input scheme. For example, the cross-camera convolutional color constancy (C5) [4] suggests utilizing additional unlabeled images captured by the testing camera, in addition to the primary single-frame image of the scene being captured to improve the generalization for cameras that were not included in the training phase. Xing et al. [60] proposed the use of a depth map, captured by a time-of-flight (ToF) sensor, along with the primary raw image to predict illuminant color in the scene by leveraging the geometry information obtained from depth map. Abdelhamed et al., [2] proposed leveraging the presence of two cameras in most modern mobile phone devices. Assuming dual streaming from both cameras, they derived a feature within the framework of chromagenic color constancy theory [21-23] to enhance the accuracy of illuminant estimation, yielding promising results.", + "bbox": [ + 212, + 313, + 787, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method, in contrast to the majority of prior work, adopts the strategy of benefiting from multiple frames available in the camera ISP (similar to [2]). However, unlike [2, 60], which requires streaming from dual cameras and may lead to impractical high power consumption, our method relies on two frames captured of the same scene under different exposure settings, already present in the HDR imaging pipeline, to estimate the illuminant color of the captured scene (see Fig. 1). Specifically, we utilize: 1) a frame captured with short exposure time, that we refer to as short-exposure image $(I_s)$ , and 2) another frame of the same scene captured with long exposure time, that we refer to as long-exposure image $(I_l)$ . Having dual-exposure images in the HDR imaging pipeline is feasible, making our method practical for most HDR imaging pipelines. Accordingly, Eq. 2 can be modified to include our dual-exposure input images as follows:", + "bbox": [ + 212, + 508, + 787, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n[ \\hat {\\ell} _ {R}, \\hat {\\ell} _ {G}, \\hat {\\ell} _ {B} ] ^ {T} = f (I _ {l}, I _ {s}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 702, + 784, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective in this work is to design the function $f$ in a manner that enables the effective utilization of the additional information provided by the dual-exposure images, $I_{l}$ and $I_{s}$ .", + "bbox": [ + 212, + 727, + 787, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contribution In this paper, we present a feature inspired by the chromagenic color constancy theory [21-23], termed the dual-exposure feature (DEF), that is derived from images captured with both short exposure time $(I_s)$ and long", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "exposure time $(I_l)$ . DEF leverages the variations in chromatic information between these dual-exposure images, providing valuable guidance for illuminant estimator methods. To assess its effectiveness, we trained a lightweight multi-layer perceptron network (MLP) for illuminant estimation that utilizes solely our DEF as input, departing from the conventional approach of using actual RGB values of the captured scene colors. Additionally, we explored the integration of DEF into an established color constancy framework, specifically the convolutional color constancy (CCC) [4,8,9,37], which we refer to as exposure-based CCC (ECCC). The experimental results on a multi-exposure dataset, collected to evaluate our work empirically, show that these models - namely, EMLP and ECC - achieve promising results with a reasonable number of parameters - 354 learnable parameters for EMLP and 6,156 learnable parameters for ECC. This outperformance across diverse evaluation metrics is observed when comparing to prior methods that require significantly higher number of parameters, ranging from tens to hundreds of thousands, or even millions.", + "bbox": [ + 212, + 146, + 787, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Illuminant-Linked Dual Exposure Feature", + "text_level": 1, + "bbox": [ + 215, + 395, + 656, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To develop an efficient illuminant estimator that benefits from both $I_{l}$ and $I_{s}$ , we introduce a compact feature that aims to capture the correlation between these dual-exposure images. Our feature is inspired by the chromagenic color constancy theory [21-23]. Specifically, we explore the analogy between long and short exposure images, $I_{l}$ and $I_{s}$ , and aligned images captured by two cameras [2] under the chromagenic color constancy theory, leading us to develop our dual-exposure feature (DEF). To begin, we review the chromagenic color constancy theory under the Lambertian reflectance model with a single illuminant assumption. The captured raw image $I$ can mathematically be described by:", + "bbox": [ + 212, + 426, + 787, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\rho} ^ {(y)} = \\int_ {\\gamma} S _ {\\rho} (x) D (x) R (y, x) d x + z, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 575, + 785, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S(\\cdot), D(\\cdot)$ , and $R(\\cdot)$ represent the camera response function (typically represented by camera sensitivity, infrared cut-off filter, and spectral lens transmission), the spectral power distribution of light, and scene reflectance, respectively. Here, $x$ refers to a wavelength within the visible range $\\gamma$ , $y$ refers to the pixel location in image $I$ , $\\rho \\in \\{R,G,B\\}$ refers to the color channel, and $z$ denotes the undesired noise, typically represented by signal-dependent and additive components [1]. According to the chromagenic color constancy theory [22], if we capture two images of the same scene with a specially chosen chromagenic filter, $Q$ , applied between image captures, the linear color matrix that maps between those captured images is unique to the illuminant color present in this scene. That is, given an image, $I$ , that is captured by the main camera, and a filtered image, $I_f$ , that can be described as:", + "bbox": [ + 212, + 617, + 787, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {f _ {\\rho}} ^ {(y)} = \\int_ {\\gamma} S _ {\\rho} (x) Q _ {\\rho} (x) D (x) R (y, x) d x + z, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 811, + 785, + 844 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the $3 \\times 3$ color matrix $C_c$ that maps between the colors of $I$ and $I_f$ is indicative of the scene illuminant. This color matrix can be computed by minimizing the following equation:", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {C _ {c}} {\\arg \\min } \\| C _ {c} I _ {f} - I \\| _ {F}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 210, + 785, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _F$ is the Frobenius norm. The closed-form solution for Eq. 6 can be obtained using the pseudoinverse (i.e., $C_c = II_f^\\dagger$ ). While theoretically validated, the chromagenic filter conditions required to obtain a unique mapping matrix per illuminant are challenging to meet in practice. Empirically, Finlayson et al., [23] found that most filters exhibit a reasonable level of correlation between the computed matrix and the illuminant color, excluding neutral filters, which consistently result in a scaling relationship between the unfiltered image, $I$ , and the filtered image, $I_{f}$ . Relaxing the conditions to include normal colored filter, Abdelhamed et al., [2] proposed using two cameras, with the second camera serving as the main camera after applying a colored filter—i.e., $S(\\cdot)Q(\\cdot)$ . Surprisingly, even when chromagenic filter conditions are not met, using a normal colored filter [23] or another camera with a different response function [2] still shows a correlation between the computed matrix $C_c$ and the illuminant color to some extent, achieving promising results.", + "bbox": [ + 212, + 246, + 787, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We argue that this correlation arises due to the variations, or what can be considered a form of \"distortion\", in the colors captured by the second (filtered) camera when compared to the original colors captured by the main camera. This color distortion varies based on the interplay between the scene irradiance, $D(\\cdot)R(\\cdot)$ , and the camera response functions - namely, the main camera's $S(\\cdot)$ and the filtered/second camera's $S(\\cdot)Q(\\cdot)$ in Eqs. 4 and 5. In the context of machine learning (ML) models, the camera response functions remain fixed across the entire dataset. Consequently, ML models, such as the one proposed in [2], learn the correlation between the ground-truth illuminant color and the scene irradiance through the matrix $C_c$ that represents the level of \"distortion\" between the colors of captured scene images.", + "bbox": [ + 212, + 460, + 787, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a dual-exposure setup, we have a long-exposure image, $I_{l}$ , and a short-exposure image, $I_{s}$ , both capturing the same scene. The extent of color distortion in each image varies based on the scene irradiance; for instance, $I_{s}$ may exhibit a higher level of color distortion and noise than $I_{l}$ under suboptimal lighting conditions (e.g., indoor light), while $I_{l}$ may have a higher level of color distortion and over-saturated colors than $I_{s}$ in well-lit scenes (e.g., outdoor light). This difference in color distortion arises because the non-linear camera response function receives a different number of photons to form each image colors based on the exposure time and scene irradiance [32]. As a result, $I_{l}$ and $I_{s}$ can exhibit color variations that differ across individual color channels [15], that are somewhat akin to those caused by a color filter (though to a lesser extent). Even within the same color channel, different levels of color differences between $I_{l}$ and $I_{s}$ can be observed spatially due to the interaction between the non-linear camera response function of that channel and the spatially varying scene irradiance (see", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a50bb5a388c976a7d741d29a232d1e3442348dcc5b03e515c91b834bfc5a225.jpg", + "image_caption": [ + "Fig. 2: Cameras perceive different amounts of photons when capturing the same scene with different exposure times. Images taken with both long and short exposure times exhibit variations in each channel due to the camera response function and scene irradiance. Additionally, spatial variations, in each color channel, can be observed based on object reflectance, as the interplay of lighting, object reflectance, and the camera response function leads to different outcomes. (A) and (B) show raw images of scenes captured under indoor and outdoor lighting, respectively. In (C) and (D), we present the average $rg$ -chromaticity histogram and aggregated red, green, and blue pixel values from 25 images sharing similar lighting conditions in (A) and (B), respectively." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 743, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 2). In Fig. 2, we show two examples of dual-exposure images captured under different lighting conditions. As can be seen, the differences between the colors in $I_{l}$ and $I_{s}$ under each lighting condition exhibit variations, which are observable in the histogram similarity (here we use the Bhattacharyya distance similarity metric) and variations of the ratios in each color channel between $I_{l}$ and $I_{s}$ . That is, the correlation between $I_{l}$ and $I_{s}$ is not always a proportional scaling, as in the case of a neutral filter. We observed similar patterns as shown in Fig. 2 when studying examples from the two-camera dataset in [2] (see supplemental materials).", + "bbox": [ + 212, + 470, + 787, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this discussion, we propose to compute the color matrix $C_c$ to map between the $rgb$ -chromaticity values (i.e., $[R / \\kappa, G / \\kappa, B / \\kappa]^T$ , with $\\kappa = R + G + B$ ) of $I_s$ and $I_l$ . The reason for not using the RGB triples, similar to chromagenic color constancy, is that we aim to reduce the influence of intensity differences between $I_l$ and $I_s$ when computing $C_c$ . In addition, we compute the covariance matrix, $C_v$ , of the ratio image $X \\in \\mathbb{R}^{3 \\times k}$ , where $X_{\\rho}^{(y)} = I_{s_{\\rho}}^{(y)} / \\left(I_{l_{\\rho}}^{(y)} + \\epsilon\\right)$ , and $\\epsilon$ is a small number added for numerical stability. Computing $C_v$ is performed as described in Eq. 7 to measure the variance in each color channel between $I_s$ and $I_l$ and the joint variability across channels.", + "bbox": [ + 212, + 607, + 787, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC _ {v} (X) = \\mathbb {E} \\left[ (X - \\mathbb {E} [ X ]) (X - \\mathbb {E} [ X ]) ^ {T} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 765, + 785, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{E}(\\cdot)$ is the expected value (mean) of the matrix. Both $C_c$ and $C_v$ form our dual-exposure feature (DEF) that represents the differences in color distortion in $I_l$ and $I_s$ . Our DEF is represented as a vector $\\in \\mathbb{R}^{15}$ , where we exclude", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b3d06d705088c112421a9f1450023c45d66abf1a10489659fced4a1da18ee52.jpg", + "image_caption": [ + "Fig. 3: We present an illuminant-related dual-exposure feature (DEF), derived from a pair of images captured with short and long exposures. Using DEF, we deploy a simple multilayer perceptron network (MLP) with only 354 parameters, referred to as exposure-based MLP or EMLP, for illuminant estimation, as shown in (A). We further explore the integration of DEF into the CCC framework, as shown in (B), by dynamically generating bias map based on DEF. We denote this modified CCC framework as exposure-based CCC or ECCC." + ], + "image_footnote": [], + "bbox": [ + 263, + 142, + 743, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the redundant values in $C_v$ over the symmetric positions. To evaluate DEF's effectiveness as a clue for scene illuminant colors, we employed a lightweight MLP for scene illuminant estimation (Fig. 3-A), that we call exposure-based MLP (EMLP). EMLP relies solely on our DEF as input. It comprises an input fully connected (fc) layer with 9 output neurons, followed by leaky ReLU (LReLU) [61], two hidden fc layers, each with 9 output neurons with LReLU in between, and an output fc layer with three neurons. EMLP achieves results comparable to complex models with thousands or millions of parameters (refer to Tables 1 and 2), while maintaining a lightweight design with just a few hundred parameters. Consequently, DEF is empirically shown as a valuable feature providing strong insights into the illuminant scene color.", + "bbox": [ + 212, + 434, + 787, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3 Integration with Convolutional Color Constancy", + "text_level": 1, + "bbox": [ + 215, + 625, + 718, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we incorporate the proposed DEF into one of the most established frameworks for illuminant estimation. Specifically, we introduce modifications to the convolutional color constancy (CCC) framework [4, 8, 9, 37] to leverage the benefits of the DEF, which we refer to as exposure-based CCC or ECCC for short. It is important to note that ECCC does not introduce a new CCC method; instead, it serves as an illustration of how the DEF can be seamlessly integrated into existing, well-established illuminant estimation frameworks.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let's begin with a brief description of the CCC [8, 9]. Given a single raw image, the CCC operates by learning one or more convolutional 2D filters, denoted as $\\{F_j\\}$ , which convolve over the 2D histogram(s), $\\{H_j\\}$ , of the image colors in the $uv$ color space (i.e., the log of $G / R$ and $G / B$ chroma values of pixel colors) [8, 20]. This convolutional operation can be accelerated by FFTs as", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proposed in [9]. Afterwards, a 2D bias map, $B$ , is added, followed by a softmax operation, $\\sigma$ , to compute the \"probability\" map, $P$ , of the illuminant bin in the uv 2D histogram space. This simplified version of the CCC can be described as follows:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nP = \\sigma \\left(\\sum_ {j} \\left(F _ {j} * H _ {j}\\right) + B\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 220, + 785, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In CCC [8,9], the filters and bias are learned across the entire training dataset. This means it consists of a single bias and one or more filters (the number depends on the histograms used, either a single histogram for image colors or two histograms, including the edge color histogram). Later, C5 [4] proposes a hypernetwork that dynamically generates filters and bias based on the input raw image and additional images taken from the same camera, aiming to improve the generalization of CCC across cameras. While cross-camera color constancy is out of the scope of this paper, we propose to use our DEF to dynamically \"generate\" a bias map based on the input image. That is, we learn a bank of biases, $\\{B_i\\}$ , where $i \\in \\{1, \\dots, n\\}$ , such that we linearly interpolate between them based on blending weights emitted from an MLP network that processes our DEF. In this way, the DEF controls the bias of the CCC model, acting as \"candidate\" illuminant priors within the uv space for the input image (see Fig. 3-B). To implement this, we need to have multiple bias maps, which definitely will lead to an impractical increase in model size. Thus, we propose to use a downsampled size $(1/4$ of histogram size) for the learnable biases, $\\{B_i\\}$ . Furthermore, we propose to feed two histograms (i.e., $j \\in \\{l, s\\}$ ) of both $I_l$ and $I_s$ , denoted as $H_l$ and $H_s$ , respectively, into the model. This leads to the learning of two downsampled filters, $\\{F_j\\}$ , corresponding to the histograms of the long-exposure image $(I_l)$ and the short-exposure image $(I_s)$ . Learning small-sized filters and biases has the following benefits. First, we can learn many biases within an affordable model size (e.g., $n = 20$ requires $\\sim 6\\mathrm{K}$ parameters, while FFCC [9] with a single bias requires $\\sim 12\\mathrm{K}$ parameters). Second, it implicitly produces smooth filters and biases, which are desirable to avoid overfitting [4,9]. With this modification, Eq. 8 can be rewritten as follows:", + "bbox": [ + 212, + 277, + 787, + 654 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nP = \\sigma \\left(\\sum_ {j} \\left(\\uparrow \\left(F _ {j}\\right) * H _ {j}\\right) + B _ {\\mathrm {u p}} ^ {\\prime}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 670, + 785, + 718 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nB _ {\\mathrm {u p}} ^ {\\prime} = \\uparrow \\left(\\sum_ {i = 1} ^ {n} w _ {i} B _ {i}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 729, + 785, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\uparrow (\\cdot)$ refers to upscaling through bilinear interpolation, $[w_{1},\\dots,w_{n}]^{T}$ is a weighting vector produced by a lightweight MLP (similar to EMLP, but with $n$ output neurons) that processes our DEF. After computing $P$ , the estimated illuminant can be obtained by:", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/618580631a98426f579bf158e013215a7ee2148b60ded767f5544d3053a0a5fb.jpg", + "image_caption": [ + "Fig. 4: Examples from the dataset used in this work. For each scene, we captured the scene with a gray calibration object placed in the scene to obtain the ground-truth illuminant (A) and captured the scene using different exposure settings without the gray object (B-H). The terms 'short /e' (C-E) and 'long $\\times e$ (F-H) refer to multiplying and dividing auto exposure time by a factor $e$ , respectively. The first image in (A) is displayed in sRGB, while the rest are shown in raw RGB space." + ], + "image_footnote": [], + "bbox": [ + 261, + 146, + 743, + 296 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\ell} _ {u} = \\sum_ {u, v} u P (u, v), \\hat {\\ell} _ {v} = \\sum_ {u, v} v P (u, v), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 431, + 785, + 465 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\ell} = \\left[ \\exp \\left(- \\hat {\\ell_ {u}}\\right) / q, 1 / q, \\exp \\left(- \\hat {\\ell_ {v}}\\right) / q \\right] ^ {T}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 473, + 785, + 503 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nq = \\sqrt {\\exp \\left(- \\hat {\\ell_ {u}}\\right) ^ {2} + \\exp \\left(- \\hat {\\ell_ {v}}\\right) ^ {2} + 1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 516, + 784, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 570, + 375, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Data", + "text_level": 1, + "bbox": [ + 215, + 602, + 303, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To the best of our knowledge, the existing HDR multi-exposure datasets (e.g., [17, 33]) do not provide ground-truth illuminant colors for training and evaluation of our method. Thus, we compiled a dataset of scenes captured with both auto exposure and various multiple-exposure settings for each scene using Pixel 7 Pro camera (see Fig. 4). Images captured with auto exposure are used for training (in the case of learning-based methods) and evaluating other methods, while those captured with multiple exposure settings are employed for validating our method. Our set comprises 558 scenes, each captured using auto exposure and six additional exposure settings: $\\times 2$ , $\\times 4$ , $\\times 8$ , $1/2$ , $1/4$ , and $1/8$ , indicating adjustments to the original exposure time. Specifically, $\\times 2$ signifies doubling the original exposure time, $\\times 4$ and $\\times 8$ denote quadrupling and octupling, respectively, while $1/2$ , $1/4$ , and $1/8$ represent dividing the original exposure time by 2, 4, and 8, respectively. These adjustments are dynamically computed as ratios from the initial auto exposure, following the common practice in multi-exposure", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "HDR imaging on smartphone cameras [28, 38]. The camera was fixed on a tripod to minimize potential misalignment between images captured by different exposure settings. In total, we captured 3,906 raw images ( $558 \\times 7$ captures of each scene). The dataset is organized as follows: 83 and 86 scenes were randomly selected for validation and testing, respectively, with the remaining 387 scenes were used for training. Additionally, each scene was captured with a gray calibration object to obtain the ground-truth illuminant. The images with the gray calibration object are not included in any of the training, validation, nor testing sets.", + "bbox": [ + 212, + 146, + 787, + 280 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During training, we optionally augment training data using chromatic adaptation. Specifically, we employ clustering on the training data using our DEF as query feature. We use K-means [34] with L2 to create 80 clusters. Subsequently, we augment the training data by generating three additional images for each sample. For each augmented image, we randomly select an illuminant color from the cluster to which the original image belongs. We then apply Von Kries transform [18] to remove the original illuminant from the image, using the actual ground truth, and apply the newly selected illuminant color to the image using the diagonal scaling operator. This new illuminant serves as the ground truth for the augmented image. As our method is designed to handle two images, namely $I_{s}$ and $I_{l}$ , we use, without loss of generality, a single exposure factor, $e$ , for both images, such that $I_{s}$ was captured with $1 / e$ of the auto-exposure time, while $I_{l}$ was captured with $\\times e$ of the auto-exposure time, for simplicity. In Sec. 4.3, we present an ablation study on the impact of different values of $e$ on our final results.", + "bbox": [ + 212, + 282, + 787, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Training", + "text_level": 1, + "bbox": [ + 215, + 530, + 333, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We trained EMLP and ECCC using the Adam optimizer [40] with a mini-batch size set to 32 for 1000 and 200 epochs, respectively. For EMLP, the learning rate was $10^{-3}$ , while ECCC was trained with an incremental mini-batch size (similar to [4]) with a learning rate of $5 \\times 10^{-3}$ with a cosine annealing schedule [46], and the weight decay (i.e., L2 regularization penalty) was set to $10^{-5}$ . For ECCC, the filter weights were initialized to zeros, and the biases were initialized to $n$ 2D histograms of training ground-truth illuminant colors after clustering into $n$ clusters using our DEF as a query feature with K-means. The $n$ histograms of training ground-truth illuminants were first processed by morphological dilation using a $3 \\times 3$ diamond-shaped structuring element before being used as initial values for our learnable biases. This initialization improves the results, as it assists in establishing reasonable \"candidate\" illuminant priors linked to the corresponding DEFs (see Sec. 4.3 for an ablation study). Both models were trained using the angular error, $L(\\cdot)$ , between the predicted illuminant color $\\hat{\\ell}$ and the ground-truth illuminant $\\ell$ , as described in the following equation [19]:", + "bbox": [ + 212, + 554, + 787, + 785 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL (\\ell , \\hat {\\ell}) = \\cos^ {- 1} \\left(\\frac {\\ell \\cdot \\hat {\\ell}}{\\| \\ell \\| \\| \\hat {\\ell} \\|}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 795, + 785, + 844 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\|\\cdot\\|$ denotes the Euclidean norm, and $(\\cdot)$ represents the vector dot product. For ECCC, we further added two smoothness loss terms to encourage smoothness in the learned filters and biases, similar to [4]. See supplemental materials for more details.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Results", + "text_level": 1, + "bbox": [ + 215, + 231, + 325, + 244 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conducted a comprehensive comparison of our proposed methods against various existing techniques, which include: training-free statistical methods [13, 14, 25, 50], camera-independent learning-based methods [4, 5], and camera-dependent learning-based methods [6, 9, 29, 41, 42, 54, 63, 64]. For the camera-dependent learning-based methods, we trained each model on our data using the provided code and recommended parameters by the respective authors. Meanwhile, for sensor-independent learning methods, we utilized the provided pre-trained models on other datasets (e.g., [7, 14, 27]).", + "bbox": [ + 212, + 258, + 787, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For the sake of conducting realistic experiments, we assessed all methods, including ours, on $384 \\times 256$ images—a suitable size for evaluating illuminant estimation methods designed for embedded hardware devices; such devices may utilize even smaller image sizes in white-balance modules [9]. For CCC methods [4,9], including the ECCC, $64 \\times 64$ histograms were used. The ECCC incorporates the proposed DEF into the CCC by generating an interpolated bias map based on the DEF. This can be seen as a spatial case analogous to C5 [4], where C5 utilizes a neural network to emit bias and convolution filters based on additional images taken by the same camera to enhance generalization across cameras. To validate our modification against C5, we trained different versions of C5 not aimed at improving cross-camera generalization (as our method also focuses on a single camera). We refer to these modified versions as follows:", + "bbox": [ + 212, + 378, + 787, + 559 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- C5 (model A): We use the histogram of the averaged short and long exposure images without additional images. That is, the C5 network uses a single histogram to generate CCC model parameters.", + "- C5 (model B): We use histograms of both images taken by long and short exposures as input to the C5 network to generate CCC model parameters. The CCC mode is then applied to the histogram of the averaged long-short images.", + "- C5 (models C and D): Both Model C and Model D use histograms of both images taken by long and short exposures to feed the C5 network and generate CCC model that is then applied to histogram of short exposure image (model C) and long exposure image (model D)." + ], + "bbox": [ + 225, + 571, + 784, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The results are given in Tables 1 and 2 on the validation and testing sets, respectively. We report the mean, median, tri-mean, best $25\\%$ , worst $25\\%$ , worst $5\\%$ , and max angular errors in each set. We also report the total number of parameters for other methods, including ours. Additionally, we present the results of a set of ablation studies conducted to examine the impact of the exposure factor $e$ in both EMLP and ECCC. We also studied the impact of different color", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Angular errors on the validation set. $\\diamond$ and $\\star$ denote camera-independent models and training-free statistical methods, respectively. Methods are listed chronologically by publication year, with the top and second-best results highlighted in yellow and red. We present the results of camera-dependent trained models for various versions of C5 [4], identified as models A, B, C, and D (see main text for more details). Our results are reported using EMLP, ECCC, and the ensemble model (EMLP + ECCC). Ablation studies are included using different mapping matrices (CM: $3 \\times 3$ color mapping matrix, TM: $3 \\times 4$ affine transformation matrix, HM: $3 \\times 3$ homography matrix), different color representations (rgb: raw RGB triplet as used in [2,21-23,26], rg/rgb-chroma: rg-chromaticity/rgb-chromaticity of the raw RGB triplet), different exposure ratios ( $e \\in \\{2,4,8\\}$ ), different input image sizes and histograms ( $s$ indicates $48 \\times 32$ input images, $s^2$ indicates $48 \\times 32$ input images and $32 \\times 32$ histograms), using different input histograms for ECCC of both long and short exposure images, average image, and long/short image, different number of learned biases ( $n$ ), and results with (w/) and without (w/o) augmentation, w/ and w/o the covariance matrix's parameters (for EMLP), w/ and w/o DEF and bias initialization (BI) (for ECCC).", + "bbox": [ + 212, + 143, + 787, + 367 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6922385beefb8ee5a2b01ffdab8104b106a0875c0282cda64583fb137add6b50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodValidation setParams
MeanMed.Trl.Best 25%Worst 25%Worst 5%Max
Grayworld* [13]5.543.134.110.9712.8319.4528.82-
Shades of Gray* [25]7.166.216.430.9915.0618.3919.21-
PCA* [14]6.134.535.020.9113.6618.2820.33-
Gamut (pixels) [29]7.537.387.032.0413.9618.3421.01636
Gamut (edges) [29]7.036.066.161.4914.5419.5121.98324
FCC[3]3.771.732.320.549.5616.0828.9912,288
Gray Index * [50]6.675.786.890.9815.0624.4528.82-
APAP-LUT [6]4.843.305.871.1210.8416.9227.53289
SIE[5]4.704.044.231.359.2113.2716.241,008,044
BoCF [41]4.843.305.871.1210.8416.9227.5359,354
C4 [63]3.913.263.241.068.5012.7617.355,115,657
CWCC [42]4.493.353.481.619.2113.3116.11100,830
C5* [4]4.012.813.141.088.5112.4721.81411,711
C5 (model A) [4]3.932.292.901.019.4013.3318.75171,511
C5 (model B) [4]4.492.843.240.8810.5215.2320.55213,831
C5 (model C) [4]4.252.673.190.9310.1214.5719.82213,831
C5 (model D) [4]4.512.943.380.8710.7416.0120.74213,831
EMLP [4]4.162.743.270.8810.7414.0521.2992,910,186
PCCC [64]4.614.194.151.758.5012.0617.86450
EMLP [c = 8, rg-chroma, HM, w/o cov]5.864.264.961.3612.316.3620.44-
EMLP [c = 8, rg-chroma, TM, w/o cov]4.783.944.271.149.5713.2016.54363
EMLP [c = 8, rgb, CM, w/o cov]4.543.533.871.269.5113.7419.22300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.813.984.221.299.3212.7514.99255
EMLP [c = 2, rg-chroma, CM, w/o cov]4.353.333.571.119.4413.6017.14300
EMLP [c = 4, rg-chroma, CM, w/o cov]4.163.563.771.008.3812.1616.79300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.073.513.641.237.7710.3813.29300
EMLP [c = 8, rg-chroma, CM, w/o cov]3.772.892.920.868.5911.4812.71354
EMLP [c = 8, n = 8, w/ DEF, both]3.642.792.860.868.6811.4213.54354
EMLP [c = 8, rg-chroma, CM, w/ aug, w/ cov]3.922.432.850.868.6811.3213.11354
ECCC [c = 8, n = 5, w/ DEF, both]3.692.783.110.797.8310.6212.372,166
ECCC [c = 8, n = 10, w/ DEF, both]3.662.692.920.958.0311.0612.323,496
ECCC [c = 8, n = 15, w/ DEF, both]3.582.712.910.917.9611.8513.514,826
ECCC [c = 2, n = 20, w/ DEF, both]3.912.993.231.078.3311.8415.236,156
ECCC [c = 4, n = 20, w/ DEF, both]4.283.193.360.999.6313.7214.726,156
ECCC [c = 8, n = 20, w/o DEF, both]4.023.233.691.229.1712.1414.474,608
ECCC [c = 8, n = 20, w/ BI, both]3.993.133.510.978.5311.4113.666,156
ECCC [c = 8, n = 20, w/ DEF, avg]4.032.973.341.038.6812.0914.715,900
ECCC [c = 8, n = 20, w/ DEF, short]3.712.742.931.008.0712.2015.435,900
ECCC [c = 8, n = 20, w/ DEF, long]3.702.742.931.008.0711.1115.435,900
ECCC [c = 8, n = 20, rg-chroma, w/ DEF, both]3.792.733.230.828.4111.6315.446,156
ECCC [c = 8, n = 20, w/ DEF, both, w/ aug]3.612.973.140.797.7710.5411.416,156
ECCC [c = 4, n = 20, w/ DEF, both]3.692.623.070.98.1210.9515.16,156
ECCC *c = 8, n = 20, w/ DEF, both)4.043.343.500.668.7511.2812.961,932
ECCC [c = 8, n = 20, w/ DEF, both]3.562.502.790.877.9311.3213.476,156
EMPLP + ECCC3.242.372.530.847.1110.6412.066,510
", + "bbox": [ + 351, + 378, + 651, + 664 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "spaces of input images before computing our DEF. Specifically, we used the $rg$ -chromaticity and raw RGB values (similar to [2, 21-23, 26]), in addition to our main design that uses $rgb$ -chromaticity. We also studied different mapping matrices, in addition to the $3 \\times 3$ color matrix we used in Sec. 2. Specifically, we examined using the geometric $3 \\times 4$ affine transformation matrix and the $3 \\times 3$ homography matrix.", + "bbox": [ + 212, + 700, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We report results of our EMLP with and without the covariance matrix parameters in addition to the results of ECCC with different numbers of learnable biases, $n$ , and with and without the proposed bias initialization. Furthermore, we", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d49fa48b8cfd814e9da81e236472b768919c0a1770f4ee2986b7a3077f760524.jpg", + "table_caption": [ + "Table 2: Angular errors on the testing set. See caption of Table 1 for abbreviations." + ], + "table_footnote": [], + "table_body": "
MethodTesting set
MeanMed.Tri.Best 25%Worst 25%Worst 50%Max
Grayworld* [13]5.323.704.510.9511.3414.4117.75
Shades of Gray* [25]6.775.776.081.0714.2416.6417.50
PCA* [14]5.914.995.221.0712.6217.1019.08
Gamut [pixels] [29]7.497.257.372.3012.9616.6219.00
Gamut [edges] [29]5.904.804.931.1312.2017.6120.09
FFPC [9]3.181.832.280.397.9713.1320.58
Gray Index* [50]5.183.474.180.6012.3216.9019.67
APAP-LUT [6]3.942.823.320.818.5411.9016.36
SHE* [5]3.512.352.740.807.5611.0612.39
BoCf 441 [9]3.913.313.401.187.7211.2412.81
C4 [63]3.792.763.061.057.9311.3313.39
CWCC [42]3.712.853.061.117.7911.2112.83
CS* [4]3.452.933.030.957.0010.3814.01
CS (model A) [4]3.822.332.810.729.0815.6920.68
CS (model B) [4]3.302.052.420.697.5811.2314.02
CS (model C) [4]3.312.012.400.637.6712.0014.23
CS (model D) [4]3.332.202.420.647.6710.8313.35
TLCX [54]3.732.893.210.957.5411.5514.78
PCC [64]4.373.663.640.898.1415.1723.34
EMLP (ε = 2, w/ aug.)5.642.662.970.887.8711.518.82
EMLP (ε = 4, w/ aug.)3.502.632.780.927.5711.6213.42
EMLP+A (ε = 8, w/ aug.)3.372.342.740.677.2410.1614.10
EMLP (ε = 8, w/ aug.)3.362.732.840.767.039.5311.85
ECCC (ε = 8)3.332.922.920.966.569.3810.26
ECCC-ε2(ε = 8)3.412.873.030.816.879.1211.03
ECCC (ε = 2)2.942.102.380.696.318.139.25
ECCC (ε = 4)3.182.392.600.816.9910.0110.56
ECCC (ε = 8)3.002.312.450.746.669.8512.23
ECCC (ε = 8, w/ aug.)2.952.092.370.766.068.149.32
ELMP + ECCC2.912.332.440.726.118.389.83
", + "bbox": [ + 408, + 171, + 594, + 368 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "show the results of ECCC without utilizing the DEF feature. Lastly, we demonstrate the influence of different sizes of input sizes on the inference accuracy of both EMLP and ECCC. The results show that the best results are obtained with $e = 8$ . This choice is intuitively sensible, as a higher exposure factor increases the differences between the dual-exposure images, resulting in greater distinction based on scene lighting conditions. The augmentation is found to be useful for EMLP; however, we did not observe consistent improvement in the case of ECCC. This is likely because EMLP has a limited number of input features, making it more susceptible to overfitting. Thus, augmentation helps in generalization. Conversely, in the case of ECCC, the inclusion of histogram features alongside the small DEF feature suggests that the model may not consistently derive benefits from augmentation. We also noted that in ECCC, using raw RGB colors of dual-exposure images (without chromaticity conversion) tends to yield better results, while in EMLP, the chromaticity conversion tends to enhance performance (see Table 1). Results presented in Table 2 default to $rgb$ -chromaticity for EMLP and raw RGB values for ECCC.", + "bbox": [ + 212, + 398, + 787, + 641 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "From the results presented in Tables 1 and 2, it is evident that our proposed DEF serves as a promising feature for guiding illuminant estimators. This is demonstrated by the performance of our lightweight EMLP model (354 parameters) compared to complex models that rely solely on raw RGB image colors (e.g., TLCC [54] with 32 million parameters). Moreover, incorporating DEF into ECCC reduces the maximum error by over $50\\%$ and achieves comparable results across various evaluation metrics when compared to FFCC [9].", + "bbox": [ + 212, + 642, + 787, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since both models - namely, EMLP and ECCC - have a reasonable number of parameters, we can combine their predictions to create an ensemble model by averaging the predicted illuminant colors from both models. The results of this ensemble model are reported in Tables 1 and 2 under 'ECCC + EMLP'. The combined efforts of EMLP and ECCC, guided by our DEF, demonstrate promising results, outperforming several state-of-the-art approaches across vari-", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/84a7ef8cc5e191d03786ac653100b37e086159e16609c229ed2a4a065f06f630.jpg", + "image_caption": [ + "Fig. 5: Randomly selected examples from our worst $25\\%$ results of ECCC (top two rows are from validation set and remaining rows are from testing set). (A) Input pair of raw images captured with long and short exposures (note that other methods use a single image captured with auto exposure). (B-G) Images corrected with the estimated illuminant by: (B) FFCC [9], (C) C5 [4] (chosen the best results among all variations discussed in Sec. 4.3), (D) TLCC [54], (E) EMLP, (F) ECCC, and (G) Ground-truth illuminant. The estimated illuminant of each method is shown on the right side of each image, along with the angular error written in the top-left corner of the image." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 743, + 308 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ous evaluation metrics while utilizing considerably fewer parameters. Qualitative examples of our results, randomly selected from the worst $25\\%$ of ECCC results, are presented in Fig. 5 alongside results from other methods.", + "bbox": [ + 212, + 460, + 787, + 506 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 215, + 532, + 532, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced DEF, a feature derived from dual exposure images to enhance illuminant estimation. Our DEF achieves comparable results with state-of-the-art methods that employ thousands or millions of parameters (e.g., [54, 63]), using only 354 parameters in a straightforward MLP network. We further discuss incorporating the proposed DEF into the established CCC framework, referred to as ECCC. ECCC achieves comparable or better results on different evaluation metrics than classic CCC approaches while remaining lightweight, requiring only 6,156 parameters (50% reduction compared to FFCC [9]), approximately 30 KB of memory, and running in $\\sim 0.25$ milliseconds per image on CPU. Our solution focuses on the single-camera case, intending testing on the same camera used for training. Future work includes studying the stability of this feature across different cameras. We discussed integrating the proposed DEF into the CCC framework. Further exploration may involve examining DEF's benefits for other illuminant estimation techniques that rely on convNets and raw image pixels as input. Another research direction could involve developing a spatially varying version of DEF, rather than our global feature, to utilize for spatially varying illuminant estimation and image white balancing.", + "bbox": [ + 212, + 566, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018)", + "2. Abdelhamed, A., Punnappurath, A., Brown, M.S.: Leveraging the availability of two cameras for illuminant estimation. In: CVPR (2021)", + "3. Affi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021)", + "4. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: ICCV (2021)", + "5. Affifi, M., Brown, M.S.: Sensor-independent illumination estimation for DNN models. In: BMVC (2019)", + "6. Affi, M., Punnappurath, A., Finlayson, G., Brown, M.S.: As-projective-as-possible bias correction for illumination estimation algorithms. Journal of the Optical Society of America A 36(1), 71-78 (2019)", + "7. Banić, N., Koščević, K., Lončarić, S.: Unsupervised learning for color constancy. arXiv preprint arXiv:1712.00436 (2017)", + "8. Barron, J.T.: Convolutional color constancy. In: ICCV (2015)", + "9. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017)", + "0. Bianco, S., Schettini, R.: Color constancy using faces. In: CVPR (2012)", + "1. Brainard, D.H., Freeman, W.T.: Bayesian color constancy. Journal of the Optical Society of America A 14(7), 1393-1411 (1997)", + "2. Brown, M.: Color processing for digital cameras, chap. 5, pp. 81-98. John Wiley & Sons, Ltd (2023)", + "3. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin Institute 310(1), 1-26 (1980)", + "4. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. Journal of the Optical Society of America A 31(5), 1049-1058 (2014)", + "5. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: Annual Conference on Computer Graphics and Interactive Techniques (1997)", + "6. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021)", + "7. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Transactions on Graphics 36(6) (2017)", + "8. Fairchild, M.D.: Color appearance models. John Wiley & Sons (2013)", + "9. Finlayson, G.D., Funt, B.V., Barnard, K.: Color constancy under varying illumination. In: ICCV (1995)", + "20. Finlayson, G.D., Hordley, S.D.: Color constancy at a pixel. Journal of the Optical Society of America A 18(2), 253-264 (2001)", + "21. Finlayson, G.D., Hordley, S.D.: The chromagenic colour camera and illuminant estimation. In: Color Imaging Conference (2005)", + "22. Finlayson, G.D., Hordley, S.D., Morovic, P.: Chromagenic colour constancy. In: Congress of the International Colour Association (2005)", + "23. Finlayson, G.D., Hordley, S.D., Morovic, P.: Colour constancy using the chromogenic constraint. In: CVPR (2005)", + "24. Finlayson, G.D., Hordley, S.D., Tastl, I.: Gamut constrained illuminant estimation. International Journal of Computer Vision 67(1), 93-109 (2006)", + "25. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference (2004)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Fredembach, C., Finlayson, G.: The bright-chromagenic algorithm for illuminant estimation. Journal of Imaging Science and Technology 52, 40906:1-40906:11 (2008)", + "27. Gehler, P.V., Rother, C., Blake, A., Minka, T., Sharp, T.: Bayesian color constancy revisited. In: CVPR (2008)", + "28. Gelfand, N., Adams, A., Park, S.H., Pulli, K.: Multi-exposure imaging on mobile devices. In: ACM International Conference on Multimedia (2010)", + "29. Gijsenij, A., Gevers, T., Van De Weijer, J.: Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision 86(2), 127-139 (2010)", + "30. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing 20(9), 2475-2489 (2011)", + "31. Gijsenij, A., Gevers, T., Van De Weijer, J.: Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(5), 918-929 (2011)", + "32. Granados, M., Ajdin, B., Wand, M., Theobalt, C., Seidel, H.P., Lensch, H.P.: Optimal HDR reconstruction with linear digital cameras. In: CVPR (2010)", + "33. Hanji, P., Mantiuk, R.K., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: SIGGRAPH (2022)", + "34. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A K-means clustering algorithm. Journal of the royal statistical society. Series C (applied statistics) 28(1), 100-108 (1979)", + "35. Hernandez-Juarez, D., Parisot, S., Busam, B., Leonardis, A., Slabaugh, G., McDonagh, S.: A multi-hypothesis approach to color constancy. In: CVPR (2020)", + "36. Hu, Y., Wang, B., Lin, S.: FC4: Fully convolutional color constancy with confidence-weighted pooling. In: CVPR (2017)", + "37. Hubel, P.M., Finlayson, G.D., Hordley, S.D.: White point estimation using color by convolution (2007), US Patent 7,200,264", + "38. Jung, C., Yang, Y., Jiao, L.: High dynamic range imaging on mobile devices using fusion of multiexposure images. Optical Engineering 52(10), 102004-102004 (2013)", + "39. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: ECCV (2016)", + "40. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "41. Laakom, F., Passalis, N., Raitoharju, J., Nikkanen, J., Tefas, A., Iosifidis, A., Gabbouj, M.: Bag of color features for color constancy. IEEE Transactions on Image Processing 29, 7722-7734 (2020)", + "42. Laakom, F., Raitoharju, J., Nikkanen, J., Iosifidis, A., Gabbouj, M.: Robust channel-wise illumination estimation. In: BMVC (2021)", + "43. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: WACV (2023)", + "44. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., et al.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 164-1 (2019)", + "45. Lo, Y.C., Chang, C.C., Chiu, H.C., Huang, Y.H., Chen, C.P., Chang, Y.L., Jou, K.: CLCC: Contrastive learning for color constancy. In: CVPR (2021)", + "46. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "47. Lou, Z., Gevers, T., Hu, N., Lucassen, M.P., et al.: Color constancy by deep learning. In: BMVC (2015)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Afifi et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Oh, S.W., Kim, S.J.: Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition 61, 405-416 (2017)", + "49. Qi, G., Chang, L., Luo, Y., Chen, Y., Zhu, Z., Wang, S.: A precise multi-exposure image fusion method based on low-level features. Sensors 20(6), 1597 (2020)", + "50. Qian, Y., Kamarainen, J.K., Nikkanen, J., Matas, J.: On finding gray pixels. In: CVPR (2019)", + "51. Shen, R., Cheng, I., Shi, J., Basu, A.: Generalized random walks for fusion of multi-exposure images. IEEE Transactions on Image Processing 20(12), 3634-3646 (2011)", + "52. Solhusvik, J., Hu, S., Johansson, R., Lin, Z., Ma, S., Mabuchi, K., Manabe, S., Mao, D., Phan, B., Rhodes, H., et al.: A 1392x976 2.8 $\\mu$ m 120db CIS with per-pixel controlled conversion gain. In: International Image Sensor Workshop (2017)", + "53. Solomatov, G., Akkaynak, D.: Spectral sensitivity estimation without a camera. In: ICCP (2023)", + "54. Tang, Y., Kang, X., Li, C., Lin, Z., Ming, A.: Transfer learning for color constancy via statistic perspective. In: AAAI (2022)", + "55. Ulucan, O., Karakaya, D., Turkan, M.: Multi-exposure image fusion based on linear embeddings and watershed masking. Signal Processing 178, 107791 (2021)", + "56. Ulucan, O., Ulucan, D., Ebner, M.: Block-based color constancy: The deviation of salient pixels. In: International Conference on Acoustics, Speech and Signal Processing (2023)", + "57. Van De Weijer, J., Gevers, T., Gijsenij, A.: Edge-based color constancy. IEEE Transactions on Image Processing 16(9), 2207-2214 (2007)", + "58. Willassen, T., Solhusvik, J., Johansson, R., Yaghmai, S., Rhodes, H., Manabe, S., Mao, D., Lin, Z., Yang, D., Cellek, O., et al.: A $1280 \\times 1080$ 4.2 $\\mu$ m split-diode pixel HDR sensor in $110\\mathrm{nm}$ BSI CMOS process. In: International Image Sensor Workshop (2015)", + "59. Xiao, J., Gu, S., Zhang, L.: Multi-domain learning for accurate and few-shot color constancy. In: CVPR (2020)", + "60. Xing, X., Qian, Y., Feng, S., Dong, Y., Matas, J.: Point cloud color constancy. In: CVPR (2022)", + "61. Xu, B., Wang, N., Chen, T., Li, M.: Empirical evaluation of rectified activations in convolutional network. arXiv preprint arXiv:1505.00853 (2015)", + "62. Yahiaoui, L., Horgan, J., Deegan, B., Yogamani, S., Hughes, C., Denny, P.: Overview and empirical analysis of ISP parameter tuning for visual perception in autonomous driving. Journal of Imaging 5(10), 78 (2019)", + "63. Yu, H., Chen, K., Wang, K., Qian, Y., Zhang, Z., Jia, K.: Cascading convolutional color constancy. In: AAAI (2020)", + "64. Yue, S., Wei, M.: Color constancy from a pure color view. Journal of the Optical Society of America A 40(3), 602-610 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 702 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "HDR illuminant estimation", + "bbox": [ + 547, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_model.json b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_model.json new file mode 100644 index 0000000000000000000000000000000000000000..832769b8f906b5fab4e96b604670b99a4b914b95 --- /dev/null +++ b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_model.json @@ -0,0 +1,2137 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.294, + 0.141, + 0.711, + 0.187 + ], + "angle": 0, + "content": "Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.213, + 0.681, + 0.229 + ], + "angle": 0, + "content": "Mahmoud Affifi, Zhenhua Hu, and Liang Liang" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.24, + 0.662, + 0.269 + ], + "angle": 0, + "content": "Google {mafifi, zhenhuahu, annyliang}@google.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.303, + 0.74, + 0.608 + ], + "angle": 0, + "content": "Abstract. High dynamic range (HDR) imaging involves capturing a series of frames of the same scene, each with different exposure settings, to broaden the dynamic range of light. This can be achieved through burst capturing or using staggered HDR sensors that capture long and short exposures simultaneously in the camera image signal processor (ISP). Within camera ISP pipeline, illuminant estimation is a crucial step aiming to estimate the color of the global illuminant in the scene. This estimation is used in camera ISP white-balance module to remove undesirable color cast in the final image. Despite the multiple frames captured in the HDR pipeline, conventional illuminant estimation methods often rely only on a single frame of the scene. In this paper, we explore leveraging information from frames captured with different exposure times. Specifically, we introduce a simple feature extracted from dual-exposure images to guide illuminant estimators, referred to as the dual-exposure feature (DEF). To validate the efficiency of DEF, we employed two illuminant estimators using the proposed DEF: 1) a multilayer perceptron network (MLP), referred to as exposure-based MLP (EMLP), and 2) a modified version of the convolutional color constancy (CCC) to integrate our DEF, that we call ECCC. Both EMLP and ECCC achieve promising results, in some cases surpassing prior methods that require hundreds of thousands or millions of parameters, with only a few hundred parameters for EMLP and a few thousand parameters for ECCC." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.621, + 0.741, + 0.649 + ], + "angle": 0, + "content": "Keywords: Computational color constancy \\(\\cdot\\) Illuminant estimation \\(\\cdot\\) HDR imaging" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.674, + 0.56, + 0.691 + ], + "angle": 0, + "content": "1 Introduction and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Camera image signal processor (ISP) comprises several modules, each dedicated to enhancing specific aspects of the quality of captured raw images by the camera sensor [16]. One key component among these modules is the white-balance module, which aims to eliminate undesirable color casts introduced by the combination of scene lighting and camera sensitivity. To achieve this, the auto white-balance module runs an illuminant estimator onboard the camera ISP to estimate the RGB color of the illuminant, under the assumption that a single global illuminant illuminates the scene for simplicity [30, 44]. Image white balancing can thus be described as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.144, + 0.402, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.226, + 0.379, + 0.234 + ], + "angle": 0, + "content": "Full-resolution raw image" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.144, + 0.471, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.226, + 0.471, + 0.234 + ], + "angle": 0, + "content": "Raw frames with different exposure times" + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.144, + 0.588, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.226, + 0.57, + 0.234 + ], + "angle": 0, + "content": "Dual-exposure frames with long exposure times" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.144, + 0.74, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.226, + 0.734, + 0.242 + ], + "angle": 0, + "content": "Final rendered image with our white-balance correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 1: Conventional illuminant estimators often rely on a single frame for illuminant color estimation. Although the HDR camera pipeline includes at least two frames of the same scene, conventional methods usually consider only a single frame (either with auto exposure, shown in the dashed green line, merged frame, or one frame of the burst capture) [12,16,44,62]. This paper proposes leveraging information from dual-exposure capturing in multi-exposure HDR imaging to enhance illuminant estimation in camera pipelines. Our method uses frames captured at long and short exposures, available in multi-exposure bursts [28] or staggered HDR sensors [52, 58]. Achieving comparable or superior results, our method employs lightweight models (\\(\\sim\\)300-6000 parameters) compared to those using hundreds of thousands or millions of parameters. In this figure and the following figures, all raw images have the gamma operator applied to aid visualization, and all sRGB images are rendered using the HDR+ camera pipeline [44]." + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {W B}} = \\operatorname {d i a g} \\left(\\left[ \\frac {\\hat {\\ell} _ {G}}{\\hat {\\ell} _ {R}}, 1, \\frac {\\hat {\\ell} _ {G}}{\\hat {\\ell} _ {B}} \\right] ^ {T}\\right) I, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.523, + 0.785, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\left[ \\hat {\\ell} _ {R}, \\hat {\\ell} _ {G}, \\hat {\\ell} _ {B} \\right] ^ {T} = f (I), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.627 + ], + "angle": 0, + "content": "where \\(I\\) and \\(I_{\\mathrm{WB}}\\) are \\(3 \\times k\\) RGB colors of raw image and white-balanced raw image, respectively, with \\(k\\) refers to the total number of pixels in the image, \\(\\operatorname{diag}(.)\\) creates a \\(3 \\times 3\\) diagonal matrix from a given vector, \\(\\hat{\\ell} \\in \\mathbb{R}^3\\) refers to the estimated illuminant color vector, \\(T\\) refers to vector transpose, and \\(f\\) is an illuminant estimator function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Modern camera ISPs capture several frames of each scene with different exposure times to enhance final image dynamic range [28,38]. This can be achieved through burst capturing, involving rapid capture with varying exposure times in quick succession, or by using staggered high dynamic range (sHDR) sensors that simultaneously capture long and short exposures [52, 58]. The captured images are then combined, incorporating multiple exposures of the same scene at different levels, resulting in a greater dynamic range than what would be possible with a single image [43, 49, 51, 55]. This HDR imaging camera pipeline, therefore, includes more than a single frame of the captured scene, and internal camera ISP modules, such as the auto white-balance module, can access these additional frames. While such additional frames may have beneficial information to help illuminant estimators, conventional illuminant estimation methods often rely on a single frame (e.g., [9, 36, 41, 47, 54, 63, 64]) in both traditional single-frame camera pipelines [39] and multi-frame HDR ISPs [44]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Such single-frame illuminant estimators can be categorized into: 1) statistical methods that rely on statistics computed from input raw image information (e.g., colors, edges) [13, 14, 25, 31, 50, 56, 57] and 2) learning-based methods that learn, from a set of training images labeled with ground-truth illuminant colors, to map from input raw color information to the corresponding illuminant color [6, 8, 10, 11, 24, 36, 42, 45, 47, 48, 54, 63, 64]. While the latter category typically achieves better results than the statistical methods, most learning-based methods are camera-dependent, meaning that they require domain adaptation or fine-tuning when deployed on new cameras to achieve similar accuracy on cameras used for training due to the influence of the camera response function on both raw image colors and ground-truth colors [3-5, 35, 53, 59]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.789, + 0.509 + ], + "angle": 0, + "content": "A limited number of attempts have proposed learning-based methods that go beyond the single-frame input scheme. For example, the cross-camera convolutional color constancy (C5) [4] suggests utilizing additional unlabeled images captured by the testing camera, in addition to the primary single-frame image of the scene being captured to improve the generalization for cameras that were not included in the training phase. Xing et al. [60] proposed the use of a depth map, captured by a time-of-flight (ToF) sensor, along with the primary raw image to predict illuminant color in the scene by leveraging the geometry information obtained from depth map. Abdelhamed et al., [2] proposed leveraging the presence of two cameras in most modern mobile phone devices. Assuming dual streaming from both cameras, they derived a feature within the framework of chromagenic color constancy theory [21-23] to enhance the accuracy of illuminant estimation, yielding promising results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.789, + 0.691 + ], + "angle": 0, + "content": "Our method, in contrast to the majority of prior work, adopts the strategy of benefiting from multiple frames available in the camera ISP (similar to [2]). However, unlike [2, 60], which requires streaming from dual cameras and may lead to impractical high power consumption, our method relies on two frames captured of the same scene under different exposure settings, already present in the HDR imaging pipeline, to estimate the illuminant color of the captured scene (see Fig. 1). Specifically, we utilize: 1) a frame captured with short exposure time, that we refer to as short-exposure image \\((I_s)\\), and 2) another frame of the same scene captured with long exposure time, that we refer to as long-exposure image \\((I_l)\\). Having dual-exposure images in the HDR imaging pipeline is feasible, making our method practical for most HDR imaging pipelines. Accordingly, Eq. 2 can be modified to include our dual-exposure input images as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.703, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\n[ \\hat {\\ell} _ {R}, \\hat {\\ell} _ {G}, \\hat {\\ell} _ {B} ] ^ {T} = f (I _ {l}, I _ {s}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Our objective in this work is to design the function \\( f \\) in a manner that enables the effective utilization of the additional information provided by the dual-exposure images, \\( I_{l} \\) and \\( I_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Contribution In this paper, we present a feature inspired by the chromagenic color constancy theory [21-23], termed the dual-exposure feature (DEF), that is derived from images captured with both short exposure time \\((I_s)\\) and long" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.374 + ], + "angle": 0, + "content": "exposure time \\((I_l)\\). DEF leverages the variations in chromatic information between these dual-exposure images, providing valuable guidance for illuminant estimator methods. To assess its effectiveness, we trained a lightweight multi-layer perceptron network (MLP) for illuminant estimation that utilizes solely our DEF as input, departing from the conventional approach of using actual RGB values of the captured scene colors. Additionally, we explored the integration of DEF into an established color constancy framework, specifically the convolutional color constancy (CCC) [4,8,9,37], which we refer to as exposure-based CCC (ECCC). The experimental results on a multi-exposure dataset, collected to evaluate our work empirically, show that these models - namely, EMLP and ECC - achieve promising results with a reasonable number of parameters - 354 learnable parameters for EMLP and 6,156 learnable parameters for ECC. This outperformance across diverse evaluation metrics is observed when comparing to prior methods that require significantly higher number of parameters, ranging from tens to hundreds of thousands, or even millions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.396, + 0.657, + 0.414 + ], + "angle": 0, + "content": "2 Illuminant-Linked Dual Exposure Feature" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.565 + ], + "angle": 0, + "content": "To develop an efficient illuminant estimator that benefits from both \\( I_{l} \\) and \\( I_{s} \\), we introduce a compact feature that aims to capture the correlation between these dual-exposure images. Our feature is inspired by the chromagenic color constancy theory [21-23]. Specifically, we explore the analogy between long and short exposure images, \\( I_{l} \\) and \\( I_{s} \\), and aligned images captured by two cameras [2] under the chromagenic color constancy theory, leading us to develop our dual-exposure feature (DEF). To begin, we review the chromagenic color constancy theory under the Lambertian reflectance model with a single illuminant assumption. The captured raw image \\( I \\) can mathematically be described by:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.576, + 0.787, + 0.609 + ], + "angle": 0, + "content": "\\[\nI _ {\\rho} ^ {(y)} = \\int_ {\\gamma} S _ {\\rho} (x) D (x) R (y, x) d x + z, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.8 + ], + "angle": 0, + "content": "where \\( S(\\cdot), D(\\cdot) \\), and \\( R(\\cdot) \\) represent the camera response function (typically represented by camera sensitivity, infrared cut-off filter, and spectral lens transmission), the spectral power distribution of light, and scene reflectance, respectively. Here, \\( x \\) refers to a wavelength within the visible range \\( \\gamma \\), \\( y \\) refers to the pixel location in image \\( I \\), \\( \\rho \\in \\{R,G,B\\} \\) refers to the color channel, and \\( z \\) denotes the undesired noise, typically represented by signal-dependent and additive components [1]. According to the chromagenic color constancy theory [22], if we capture two images of the same scene with a specially chosen chromagenic filter, \\( Q \\), applied between image captures, the linear color matrix that maps between those captured images is unique to the illuminant color present in this scene. That is, given an image, \\( I \\), that is captured by the main camera, and a filtered image, \\( I_f \\), that can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.812, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nI _ {f _ {\\rho}} ^ {(y)} = \\int_ {\\gamma} S _ {\\rho} (x) Q _ {\\rho} (x) D (x) R (y, x) d x + z, \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "the \\(3 \\times 3\\) color matrix \\(C_c\\) that maps between the colors of \\(I\\) and \\(I_f\\) is indicative of the scene illuminant. This color matrix can be computed by minimizing the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.212, + 0.786, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\underset {C _ {c}} {\\arg \\min } \\| C _ {c} I _ {f} - I \\| _ {F}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.247, + 0.789, + 0.461 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _F\\) is the Frobenius norm. The closed-form solution for Eq. 6 can be obtained using the pseudoinverse (i.e., \\(C_c = II_f^\\dagger\\)). While theoretically validated, the chromagenic filter conditions required to obtain a unique mapping matrix per illuminant are challenging to meet in practice. Empirically, Finlayson et al., [23] found that most filters exhibit a reasonable level of correlation between the computed matrix and the illuminant color, excluding neutral filters, which consistently result in a scaling relationship between the unfiltered image, \\(I\\), and the filtered image, \\(I_{f}\\). Relaxing the conditions to include normal colored filter, Abdelhamed et al., [2] proposed using two cameras, with the second camera serving as the main camera after applying a colored filter—i.e., \\(S(\\cdot)Q(\\cdot)\\). Surprisingly, even when chromagenic filter conditions are not met, using a normal colored filter [23] or another camera with a different response function [2] still shows a correlation between the computed matrix \\(C_c\\) and the illuminant color to some extent, achieving promising results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.789, + 0.629 + ], + "angle": 0, + "content": "We argue that this correlation arises due to the variations, or what can be considered a form of \"distortion\", in the colors captured by the second (filtered) camera when compared to the original colors captured by the main camera. This color distortion varies based on the interplay between the scene irradiance, \\( D(\\cdot)R(\\cdot) \\), and the camera response functions - namely, the main camera's \\( S(\\cdot) \\) and the filtered/second camera's \\( S(\\cdot)Q(\\cdot) \\) in Eqs. 4 and 5. In the context of machine learning (ML) models, the camera response functions remain fixed across the entire dataset. Consequently, ML models, such as the one proposed in [2], learn the correlation between the ground-truth illuminant color and the scene irradiance through the matrix \\( C_c \\) that represents the level of \"distortion\" between the colors of captured scene images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In a dual-exposure setup, we have a long-exposure image, \\( I_{l} \\), and a short-exposure image, \\( I_{s} \\), both capturing the same scene. The extent of color distortion in each image varies based on the scene irradiance; for instance, \\( I_{s} \\) may exhibit a higher level of color distortion and noise than \\( I_{l} \\) under suboptimal lighting conditions (e.g., indoor light), while \\( I_{l} \\) may have a higher level of color distortion and over-saturated colors than \\( I_{s} \\) in well-lit scenes (e.g., outdoor light). This difference in color distortion arises because the non-linear camera response function receives a different number of photons to form each image colors based on the exposure time and scene irradiance [32]. As a result, \\( I_{l} \\) and \\( I_{s} \\) can exhibit color variations that differ across individual color channels [15], that are somewhat akin to those caused by a color filter (though to a lesser extent). Even within the same color channel, different levels of color differences between \\( I_{l} \\) and \\( I_{s} \\) can be observed spatially due to the interaction between the non-linear camera response function of that channel and the spatially varying scene irradiance (see" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.145, + 0.744, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.44 + ], + "angle": 0, + "content": "Fig. 2: Cameras perceive different amounts of photons when capturing the same scene with different exposure times. Images taken with both long and short exposure times exhibit variations in each channel due to the camera response function and scene irradiance. Additionally, spatial variations, in each color channel, can be observed based on object reflectance, as the interplay of lighting, object reflectance, and the camera response function leads to different outcomes. (A) and (B) show raw images of scenes captured under indoor and outdoor lighting, respectively. In (C) and (D), we present the average \\( rg \\)-chromaticity histogram and aggregated red, green, and blue pixel values from 25 images sharing similar lighting conditions in (A) and (B), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.606 + ], + "angle": 0, + "content": "Fig. 2). In Fig. 2, we show two examples of dual-exposure images captured under different lighting conditions. As can be seen, the differences between the colors in \\( I_{l} \\) and \\( I_{s} \\) under each lighting condition exhibit variations, which are observable in the histogram similarity (here we use the Bhattacharyya distance similarity metric) and variations of the ratios in each color channel between \\( I_{l} \\) and \\( I_{s} \\). That is, the correlation between \\( I_{l} \\) and \\( I_{s} \\) is not always a proportional scaling, as in the case of a neutral filter. We observed similar patterns as shown in Fig. 2 when studying examples from the two-camera dataset in [2] (see supplemental materials)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Based on this discussion, we propose to compute the color matrix \\( C_c \\) to map between the \\( rgb \\)-chromaticity values (i.e., \\( [R / \\kappa, G / \\kappa, B / \\kappa]^T \\), with \\( \\kappa = R + G + B \\)) of \\( I_s \\) and \\( I_l \\). The reason for not using the RGB triples, similar to chromagenic color constancy, is that we aim to reduce the influence of intensity differences between \\( I_l \\) and \\( I_s \\) when computing \\( C_c \\). In addition, we compute the covariance matrix, \\( C_v \\), of the ratio image \\( X \\in \\mathbb{R}^{3 \\times k} \\), where \\( X_{\\rho}^{(y)} = I_{s_{\\rho}}^{(y)} / \\left(I_{l_{\\rho}}^{(y)} + \\epsilon\\right) \\), and \\( \\epsilon \\) is a small number added for numerical stability. Computing \\( C_v \\) is performed as described in Eq. 7 to measure the variance in each color channel between \\( I_s \\) and \\( I_l \\) and the joint variability across channels." + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.766, + 0.787, + 0.786 + ], + "angle": 0, + "content": "\\[\nC _ {v} (X) = \\mathbb {E} \\left[ (X - \\mathbb {E} [ X ]) (X - \\mathbb {E} [ X ]) ^ {T} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathbb{E}(\\cdot)\\) is the expected value (mean) of the matrix. Both \\(C_c\\) and \\(C_v\\) form our dual-exposure feature (DEF) that represents the differences in color distortion in \\(I_l\\) and \\(I_s\\). Our DEF is represented as a vector \\(\\in \\mathbb{R}^{15}\\), where we exclude" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.549, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.143, + 0.744, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 3: We present an illuminant-related dual-exposure feature (DEF), derived from a pair of images captured with short and long exposures. Using DEF, we deploy a simple multilayer perceptron network (MLP) with only 354 parameters, referred to as exposure-based MLP or EMLP, for illuminant estimation, as shown in (A). We further explore the integration of DEF into the CCC framework, as shown in (B), by dynamically generating bias map based on DEF. We denote this modified CCC framework as exposure-based CCC or ECCC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.603 + ], + "angle": 0, + "content": "the redundant values in \\( C_v \\) over the symmetric positions. To evaluate DEF's effectiveness as a clue for scene illuminant colors, we employed a lightweight MLP for scene illuminant estimation (Fig. 3-A), that we call exposure-based MLP (EMLP). EMLP relies solely on our DEF as input. It comprises an input fully connected (fc) layer with 9 output neurons, followed by leaky ReLU (LReLU) [61], two hidden fc layers, each with 9 output neurons with LReLU in between, and an output fc layer with three neurons. EMLP achieves results comparable to complex models with thousands or millions of parameters (refer to Tables 1 and 2), while maintaining a lightweight design with just a few hundred parameters. Consequently, DEF is empirically shown as a valuable feature providing strong insights into the illuminant scene color." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.626, + 0.72, + 0.644 + ], + "angle": 0, + "content": "3 Integration with Convolutional Color Constancy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "In this section, we incorporate the proposed DEF into one of the most established frameworks for illuminant estimation. Specifically, we introduce modifications to the convolutional color constancy (CCC) framework [4, 8, 9, 37] to leverage the benefits of the DEF, which we refer to as exposure-based CCC or ECCC for short. It is important to note that ECCC does not introduce a new CCC method; instead, it serves as an illustration of how the DEF can be seamlessly integrated into existing, well-established illuminant estimation frameworks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Let's begin with a brief description of the CCC [8, 9]. Given a single raw image, the CCC operates by learning one or more convolutional 2D filters, denoted as \\(\\{F_j\\}\\), which convolve over the 2D histogram(s), \\(\\{H_j\\}\\), of the image colors in the \\(uv\\) color space (i.e., the log of \\(G / R\\) and \\(G / B\\) chroma values of pixel colors) [8, 20]. This convolutional operation can be accelerated by FFTs as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "proposed in [9]. Afterwards, a 2D bias map, \\( B \\), is added, followed by a softmax operation, \\( \\sigma \\), to compute the \"probability\" map, \\( P \\), of the illuminant bin in the uv 2D histogram space. This simplified version of the CCC can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.222, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\nP = \\sigma \\left(\\sum_ {j} \\left(F _ {j} * H _ {j}\\right) + B\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.789, + 0.655 + ], + "angle": 0, + "content": "In CCC [8,9], the filters and bias are learned across the entire training dataset. This means it consists of a single bias and one or more filters (the number depends on the histograms used, either a single histogram for image colors or two histograms, including the edge color histogram). Later, C5 [4] proposes a hypernetwork that dynamically generates filters and bias based on the input raw image and additional images taken from the same camera, aiming to improve the generalization of CCC across cameras. While cross-camera color constancy is out of the scope of this paper, we propose to use our DEF to dynamically \"generate\" a bias map based on the input image. That is, we learn a bank of biases, \\(\\{B_i\\}\\), where \\(i \\in \\{1, \\dots, n\\}\\), such that we linearly interpolate between them based on blending weights emitted from an MLP network that processes our DEF. In this way, the DEF controls the bias of the CCC model, acting as \"candidate\" illuminant priors within the uv space for the input image (see Fig. 3-B). To implement this, we need to have multiple bias maps, which definitely will lead to an impractical increase in model size. Thus, we propose to use a downsampled size \\((1/4\\) of histogram size) for the learnable biases, \\(\\{B_i\\}\\). Furthermore, we propose to feed two histograms (i.e., \\(j \\in \\{l, s\\}\\)) of both \\(I_l\\) and \\(I_s\\), denoted as \\(H_l\\) and \\(H_s\\), respectively, into the model. This leads to the learning of two downsampled filters, \\(\\{F_j\\}\\), corresponding to the histograms of the long-exposure image \\((I_l)\\) and the short-exposure image \\((I_s)\\). Learning small-sized filters and biases has the following benefits. First, we can learn many biases within an affordable model size (e.g., \\(n = 20\\) requires \\(\\sim 6\\mathrm{K}\\) parameters, while FFCC [9] with a single bias requires \\(\\sim 12\\mathrm{K}\\) parameters). Second, it implicitly produces smooth filters and biases, which are desirable to avoid overfitting [4,9]. With this modification, Eq. 8 can be rewritten as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.671, + 0.786, + 0.719 + ], + "angle": 0, + "content": "\\[\nP = \\sigma \\left(\\sum_ {j} \\left(\\uparrow \\left(F _ {j}\\right) * H _ {j}\\right) + B _ {\\mathrm {u p}} ^ {\\prime}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.731, + 0.786, + 0.769 + ], + "angle": 0, + "content": "\\[\nB _ {\\mathrm {u p}} ^ {\\prime} = \\uparrow \\left(\\sum_ {i = 1} ^ {n} w _ {i} B _ {i}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\uparrow (\\cdot)\\) refers to upscaling through bilinear interpolation, \\([w_{1},\\dots,w_{n}]^{T}\\) is a weighting vector produced by a lightweight MLP (similar to EMLP, but with \\(n\\) output neurons) that processes our DEF. After computing \\(P\\), the estimated illuminant can be obtained by:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.549, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.147, + 0.744, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 4: Examples from the dataset used in this work. For each scene, we captured the scene with a gray calibration object placed in the scene to obtain the ground-truth illuminant (A) and captured the scene using different exposure settings without the gray object (B-H). The terms 'short /e' (C-E) and 'long \\(\\times e\\) (F-H) refer to multiplying and dividing auto exposure time by a factor \\(e\\), respectively. The first image in (A) is displayed in sRGB, while the rest are shown in raw RGB space." + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.433, + 0.786, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\hat {\\ell} _ {u} = \\sum_ {u, v} u P (u, v), \\hat {\\ell} _ {v} = \\sum_ {u, v} v P (u, v), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.474, + 0.786, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\hat {\\ell} = \\left[ \\exp \\left(- \\hat {\\ell_ {u}}\\right) / q, 1 / q, \\exp \\left(- \\hat {\\ell_ {v}}\\right) / q \\right] ^ {T}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.517, + 0.785, + 0.552 + ], + "angle": 0, + "content": "\\[\nq = \\sqrt {\\exp \\left(- \\hat {\\ell_ {u}}\\right) ^ {2} + \\exp \\left(- \\hat {\\ell_ {v}}\\right) ^ {2} + 1}. \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.571, + 0.376, + 0.589 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.603, + 0.304, + 0.617 + ], + "angle": 0, + "content": "4.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To the best of our knowledge, the existing HDR multi-exposure datasets (e.g., [17, 33]) do not provide ground-truth illuminant colors for training and evaluation of our method. Thus, we compiled a dataset of scenes captured with both auto exposure and various multiple-exposure settings for each scene using Pixel 7 Pro camera (see Fig. 4). Images captured with auto exposure are used for training (in the case of learning-based methods) and evaluating other methods, while those captured with multiple exposure settings are employed for validating our method. Our set comprises 558 scenes, each captured using auto exposure and six additional exposure settings: \\(\\times 2\\), \\(\\times 4\\), \\(\\times 8\\), \\(1/2\\), \\(1/4\\), and \\(1/8\\), indicating adjustments to the original exposure time. Specifically, \\(\\times 2\\) signifies doubling the original exposure time, \\(\\times 4\\) and \\(\\times 8\\) denote quadrupling and octupling, respectively, while \\(1/2\\), \\(1/4\\), and \\(1/8\\) represent dividing the original exposure time by 2, 4, and 8, respectively. These adjustments are dynamically computed as ratios from the initial auto exposure, following the common practice in multi-exposure" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.281 + ], + "angle": 0, + "content": "HDR imaging on smartphone cameras [28, 38]. The camera was fixed on a tripod to minimize potential misalignment between images captured by different exposure settings. In total, we captured 3,906 raw images (\\(558 \\times 7\\) captures of each scene). The dataset is organized as follows: 83 and 86 scenes were randomly selected for validation and testing, respectively, with the remaining 387 scenes were used for training. Additionally, each scene was captured with a gray calibration object to obtain the ground-truth illuminant. The images with the gray calibration object are not included in any of the training, validation, nor testing sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.789, + 0.508 + ], + "angle": 0, + "content": "During training, we optionally augment training data using chromatic adaptation. Specifically, we employ clustering on the training data using our DEF as query feature. We use K-means [34] with L2 to create 80 clusters. Subsequently, we augment the training data by generating three additional images for each sample. For each augmented image, we randomly select an illuminant color from the cluster to which the original image belongs. We then apply Von Kries transform [18] to remove the original illuminant from the image, using the actual ground truth, and apply the newly selected illuminant color to the image using the diagonal scaling operator. This new illuminant serves as the ground truth for the augmented image. As our method is designed to handle two images, namely \\( I_{s} \\) and \\( I_{l} \\), we use, without loss of generality, a single exposure factor, \\( e \\), for both images, such that \\( I_{s} \\) was captured with \\( 1 / e \\) of the auto-exposure time, while \\( I_{l} \\) was captured with \\( \\times e \\) of the auto-exposure time, for simplicity. In Sec. 4.3, we present an ablation study on the impact of different values of \\( e \\) on our final results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.334, + 0.547 + ], + "angle": 0, + "content": "4.2 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.789, + 0.786 + ], + "angle": 0, + "content": "We trained EMLP and ECCC using the Adam optimizer [40] with a mini-batch size set to 32 for 1000 and 200 epochs, respectively. For EMLP, the learning rate was \\(10^{-3}\\), while ECCC was trained with an incremental mini-batch size (similar to [4]) with a learning rate of \\(5 \\times 10^{-3}\\) with a cosine annealing schedule [46], and the weight decay (i.e., L2 regularization penalty) was set to \\(10^{-5}\\). For ECCC, the filter weights were initialized to zeros, and the biases were initialized to \\(n\\) 2D histograms of training ground-truth illuminant colors after clustering into \\(n\\) clusters using our DEF as a query feature with K-means. The \\(n\\) histograms of training ground-truth illuminants were first processed by morphological dilation using a \\(3 \\times 3\\) diamond-shaped structuring element before being used as initial values for our learnable biases. This initialization improves the results, as it assists in establishing reasonable \"candidate\" illuminant priors linked to the corresponding DEFs (see Sec. 4.3 for an ablation study). Both models were trained using the angular error, \\(L(\\cdot)\\), between the predicted illuminant color \\(\\hat{\\ell}\\) and the ground-truth illuminant \\(\\ell\\), as described in the following equation [19]:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.796, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nL (\\ell , \\hat {\\ell}) = \\cos^ {- 1} \\left(\\frac {\\ell \\cdot \\hat {\\ell}}{\\| \\ell \\| \\| \\hat {\\ell} \\|}\\right), \\tag {14}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.548, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "where \\( \\|\\cdot\\| \\) denotes the Euclidean norm, and \\( (\\cdot) \\) represents the vector dot product. For ECCC, we further added two smoothness loss terms to encourage smoothness in the learned filters and biases, similar to [4]. See supplemental materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.232, + 0.326, + 0.246 + ], + "angle": 0, + "content": "4.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.379 + ], + "angle": 0, + "content": "We conducted a comprehensive comparison of our proposed methods against various existing techniques, which include: training-free statistical methods [13, 14, 25, 50], camera-independent learning-based methods [4, 5], and camera-dependent learning-based methods [6, 9, 29, 41, 42, 54, 63, 64]. For the camera-dependent learning-based methods, we trained each model on our data using the provided code and recommended parameters by the respective authors. Meanwhile, for sensor-independent learning methods, we utilized the provided pre-trained models on other datasets (e.g., [7, 14, 27])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.56 + ], + "angle": 0, + "content": "For the sake of conducting realistic experiments, we assessed all methods, including ours, on \\(384 \\times 256\\) images—a suitable size for evaluating illuminant estimation methods designed for embedded hardware devices; such devices may utilize even smaller image sizes in white-balance modules [9]. For CCC methods [4,9], including the ECCC, \\(64 \\times 64\\) histograms were used. The ECCC incorporates the proposed DEF into the CCC by generating an interpolated bias map based on the DEF. This can be seen as a spatial case analogous to C5 [4], where C5 utilizes a neural network to emit bias and convolution filters based on additional images taken by the same camera to enhance generalization across cameras. To validate our modification against C5, we trained different versions of C5 not aimed at improving cross-camera generalization (as our method also focuses on a single camera). We refer to these modified versions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.572, + 0.785, + 0.615 + ], + "angle": 0, + "content": "- C5 (model A): We use the histogram of the averaged short and long exposure images without additional images. That is, the C5 network uses a single histogram to generate CCC model parameters." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.675 + ], + "angle": 0, + "content": "- C5 (model B): We use histograms of both images taken by long and short exposures as input to the C5 network to generate CCC model parameters. The CCC mode is then applied to the histogram of the averaged long-short images." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.678, + 0.785, + 0.736 + ], + "angle": 0, + "content": "- C5 (models C and D): Both Model C and Model D use histograms of both images taken by long and short exposures to feed the C5 network and generate CCC model that is then applied to histogram of short exposure image (model C) and long exposure image (model D)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.572, + 0.785, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The results are given in Tables 1 and 2 on the validation and testing sets, respectively. We report the mean, median, tri-mean, best \\(25\\%\\), worst \\(25\\%\\), worst \\(5\\%\\), and max angular errors in each set. We also report the total number of parameters for other methods, including ours. Additionally, we present the results of a set of ablation studies conducted to examine the impact of the exposure factor \\(e\\) in both EMLP and ECCC. We also studied the impact of different color" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Table 1: Angular errors on the validation set. \\(\\diamond\\) and \\(\\star\\) denote camera-independent models and training-free statistical methods, respectively. Methods are listed chronologically by publication year, with the top and second-best results highlighted in yellow and red. We present the results of camera-dependent trained models for various versions of C5 [4], identified as models A, B, C, and D (see main text for more details). Our results are reported using EMLP, ECCC, and the ensemble model (EMLP + ECCC). Ablation studies are included using different mapping matrices (CM: \\(3 \\times 3\\) color mapping matrix, TM: \\(3 \\times 4\\) affine transformation matrix, HM: \\(3 \\times 3\\) homography matrix), different color representations (rgb: raw RGB triplet as used in [2,21-23,26], rg/rgb-chroma: rg-chromaticity/rgb-chromaticity of the raw RGB triplet), different exposure ratios (\\(e \\in \\{2,4,8\\}\\)), different input image sizes and histograms (\\(s\\) indicates \\(48 \\times 32\\) input images, \\(s^2\\) indicates \\(48 \\times 32\\) input images and \\(32 \\times 32\\) histograms), using different input histograms for ECCC of both long and short exposure images, average image, and long/short image, different number of learned biases (\\(n\\)), and results with (w/) and without (w/o) augmentation, w/ and w/o the covariance matrix's parameters (for EMLP), w/ and w/o DEF and bias initialization (BI) (for ECCC)." + }, + { + "type": "table", + "bbox": [ + 0.352, + 0.379, + 0.653, + 0.665 + ], + "angle": 0, + "content": "
MethodValidation setParams
MeanMed.Trl.Best 25%Worst 25%Worst 5%Max
Grayworld* [13]5.543.134.110.9712.8319.4528.82-
Shades of Gray* [25]7.166.216.430.9915.0618.3919.21-
PCA* [14]6.134.535.020.9113.6618.2820.33-
Gamut (pixels) [29]7.537.387.032.0413.9618.3421.01636
Gamut (edges) [29]7.036.066.161.4914.5419.5121.98324
FCC[3]3.771.732.320.549.5616.0828.9912,288
Gray Index * [50]6.675.786.890.9815.0624.4528.82-
APAP-LUT [6]4.843.305.871.1210.8416.9227.53289
SIE[5]4.704.044.231.359.2113.2716.241,008,044
BoCF [41]4.843.305.871.1210.8416.9227.5359,354
C4 [63]3.913.263.241.068.5012.7617.355,115,657
CWCC [42]4.493.353.481.619.2113.3116.11100,830
C5* [4]4.012.813.141.088.5112.4721.81411,711
C5 (model A) [4]3.932.292.901.019.4013.3318.75171,511
C5 (model B) [4]4.492.843.240.8810.5215.2320.55213,831
C5 (model C) [4]4.252.673.190.9310.1214.5719.82213,831
C5 (model D) [4]4.512.943.380.8710.7416.0120.74213,831
EMLP [4]4.162.743.270.8810.7414.0521.2992,910,186
PCCC [64]4.614.194.151.758.5012.0617.86450
EMLP [c = 8, rg-chroma, HM, w/o cov]5.864.264.961.3612.316.3620.44-
EMLP [c = 8, rg-chroma, TM, w/o cov]4.783.944.271.149.5713.2016.54363
EMLP [c = 8, rgb, CM, w/o cov]4.543.533.871.269.5113.7419.22300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.813.984.221.299.3212.7514.99255
EMLP [c = 2, rg-chroma, CM, w/o cov]4.353.333.571.119.4413.6017.14300
EMLP [c = 4, rg-chroma, CM, w/o cov]4.163.563.771.008.3812.1616.79300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.073.513.641.237.7710.3813.29300
EMLP [c = 8, rg-chroma, CM, w/o cov]3.772.892.920.868.5911.4812.71354
EMLP [c = 8, n = 8, w/ DEF, both]3.642.792.860.868.6811.4213.54354
EMLP [c = 8, rg-chroma, CM, w/ aug, w/ cov]3.922.432.850.868.6811.3213.11354
ECCC [c = 8, n = 5, w/ DEF, both]3.692.783.110.797.8310.6212.372,166
ECCC [c = 8, n = 10, w/ DEF, both]3.662.692.920.958.0311.0612.323,496
ECCC [c = 8, n = 15, w/ DEF, both]3.582.712.910.917.9611.8513.514,826
ECCC [c = 2, n = 20, w/ DEF, both]3.912.993.231.078.3311.8415.236,156
ECCC [c = 4, n = 20, w/ DEF, both]4.283.193.360.999.6313.7214.726,156
ECCC [c = 8, n = 20, w/o DEF, both]4.023.233.691.229.1712.1414.474,608
ECCC [c = 8, n = 20, w/ BI, both]3.993.133.510.978.5311.4113.666,156
ECCC [c = 8, n = 20, w/ DEF, avg]4.032.973.341.038.6812.0914.715,900
ECCC [c = 8, n = 20, w/ DEF, short]3.712.742.931.008.0712.2015.435,900
ECCC [c = 8, n = 20, w/ DEF, long]3.702.742.931.008.0711.1115.435,900
ECCC [c = 8, n = 20, rg-chroma, w/ DEF, both]3.792.733.230.828.4111.6315.446,156
ECCC [c = 8, n = 20, w/ DEF, both, w/ aug]3.612.973.140.797.7710.5411.416,156
ECCC [c = 4, n = 20, w/ DEF, both]3.692.623.070.98.1210.9515.16,156
ECCC *c = 8, n = 20, w/ DEF, both)4.043.343.500.668.7511.2812.961,932
ECCC [c = 8, n = 20, w/ DEF, both]3.562.502.790.877.9311.3213.476,156
EMPLP + ECCC3.242.372.530.847.1110.6412.066,510
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.793 + ], + "angle": 0, + "content": "spaces of input images before computing our DEF. Specifically, we used the \\( rg \\)-chromaticity and raw RGB values (similar to [2, 21-23, 26]), in addition to our main design that uses \\( rgb \\)-chromaticity. We also studied different mapping matrices, in addition to the \\( 3 \\times 3 \\) color matrix we used in Sec. 2. Specifically, we examined using the geometric \\( 3 \\times 4 \\) affine transformation matrix and the \\( 3 \\times 3 \\) homography matrix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We report results of our EMLP with and without the covariance matrix parameters in addition to the results of ECCC with different numbers of learnable biases, \\( n \\), and with and without the proposed bias initialization. Furthermore, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.549, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.159 + ], + "angle": 0, + "content": "Table 2: Angular errors on the testing set. See caption of Table 1 for abbreviations." + }, + { + "type": "table", + "bbox": [ + 0.409, + 0.172, + 0.596, + 0.369 + ], + "angle": 0, + "content": "
MethodTesting set
MeanMed.Tri.Best 25%Worst 25%Worst 50%Max
Grayworld* [13]5.323.704.510.9511.3414.4117.75
Shades of Gray* [25]6.775.776.081.0714.2416.6417.50
PCA* [14]5.914.995.221.0712.6217.1019.08
Gamut [pixels] [29]7.497.257.372.3012.9616.6219.00
Gamut [edges] [29]5.904.804.931.1312.2017.6120.09
FFPC [9]3.181.832.280.397.9713.1320.58
Gray Index* [50]5.183.474.180.6012.3216.9019.67
APAP-LUT [6]3.942.823.320.818.5411.9016.36
SHE* [5]3.512.352.740.807.5611.0612.39
BoCf 441 [9]3.913.313.401.187.7211.2412.81
C4 [63]3.792.763.061.057.9311.3313.39
CWCC [42]3.712.853.061.117.7911.2112.83
CS* [4]3.452.933.030.957.0010.3814.01
CS (model A) [4]3.822.332.810.729.0815.6920.68
CS (model B) [4]3.302.052.420.697.5811.2314.02
CS (model C) [4]3.312.012.400.637.6712.0014.23
CS (model D) [4]3.332.202.420.647.6710.8313.35
TLCX [54]3.732.893.210.957.5411.5514.78
PCC [64]4.373.663.640.898.1415.1723.34
EMLP (ε = 2, w/ aug.)5.642.662.970.887.8711.518.82
EMLP (ε = 4, w/ aug.)3.502.632.780.927.5711.6213.42
EMLP+A (ε = 8, w/ aug.)3.372.342.740.677.2410.1614.10
EMLP (ε = 8, w/ aug.)3.362.732.840.767.039.5311.85
ECCC (ε = 8)3.332.922.920.966.569.3810.26
ECCC-ε2(ε = 8)3.412.873.030.816.879.1211.03
ECCC (ε = 2)2.942.102.380.696.318.139.25
ECCC (ε = 4)3.182.392.600.816.9910.0110.56
ECCC (ε = 8)3.002.312.450.746.669.8512.23
ECCC (ε = 8, w/ aug.)2.952.092.370.766.068.149.32
ELMP + ECCC2.912.332.440.726.118.389.83
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.642 + ], + "angle": 0, + "content": "show the results of ECCC without utilizing the DEF feature. Lastly, we demonstrate the influence of different sizes of input sizes on the inference accuracy of both EMLP and ECCC. The results show that the best results are obtained with \\( e = 8 \\). This choice is intuitively sensible, as a higher exposure factor increases the differences between the dual-exposure images, resulting in greater distinction based on scene lighting conditions. The augmentation is found to be useful for EMLP; however, we did not observe consistent improvement in the case of ECCC. This is likely because EMLP has a limited number of input features, making it more susceptible to overfitting. Thus, augmentation helps in generalization. Conversely, in the case of ECCC, the inclusion of histogram features alongside the small DEF feature suggests that the model may not consistently derive benefits from augmentation. We also noted that in ECCC, using raw RGB colors of dual-exposure images (without chromaticity conversion) tends to yield better results, while in EMLP, the chromaticity conversion tends to enhance performance (see Table 1). Results presented in Table 2 default to \\( rgb \\)-chromaticity for EMLP and raw RGB values for ECCC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.749 + ], + "angle": 0, + "content": "From the results presented in Tables 1 and 2, it is evident that our proposed DEF serves as a promising feature for guiding illuminant estimators. This is demonstrated by the performance of our lightweight EMLP model (354 parameters) compared to complex models that rely solely on raw RGB image colors (e.g., TLCC [54] with 32 million parameters). Moreover, incorporating DEF into ECCC reduces the maximum error by over \\(50\\%\\) and achieves comparable results across various evaluation metrics when compared to FFCC [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Since both models - namely, EMLP and ECCC - have a reasonable number of parameters, we can combine their predictions to create an ensemble model by averaging the predicted illuminant colors from both models. The results of this ensemble model are reported in Tables 1 and 2 under 'ECCC + EMLP'. The combined efforts of EMLP and ECCC, guided by our DEF, demonstrate promising results, outperforming several state-of-the-art approaches across vari-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.145, + 0.744, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 5: Randomly selected examples from our worst \\(25\\%\\) results of ECCC (top two rows are from validation set and remaining rows are from testing set). (A) Input pair of raw images captured with long and short exposures (note that other methods use a single image captured with auto exposure). (B-G) Images corrected with the estimated illuminant by: (B) FFCC [9], (C) C5 [4] (chosen the best results among all variations discussed in Sec. 4.3), (D) TLCC [54], (E) EMLP, (F) ECCC, and (G) Ground-truth illuminant. The estimated illuminant of each method is shown on the right side of each image, along with the angular error written in the top-left corner of the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.507 + ], + "angle": 0, + "content": "ous evaluation metrics while utilizing considerably fewer parameters. Qualitative examples of our results, randomly selected from the worst \\(25\\%\\) of ECCC results, are presented in Fig. 5 alongside results from other methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.533, + 0.55 + ], + "angle": 0, + "content": "5 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduced DEF, a feature derived from dual exposure images to enhance illuminant estimation. Our DEF achieves comparable results with state-of-the-art methods that employ thousands or millions of parameters (e.g., [54, 63]), using only 354 parameters in a straightforward MLP network. We further discuss incorporating the proposed DEF into the established CCC framework, referred to as ECCC. ECCC achieves comparable or better results on different evaluation metrics than classic CCC approaches while remaining lightweight, requiring only 6,156 parameters (50% reduction compared to FFCC [9]), approximately 30 KB of memory, and running in \\(\\sim 0.25\\) milliseconds per image on CPU. Our solution focuses on the single-camera case, intending testing on the same camera used for training. Future work includes studying the stability of this feature across different cameras. We discussed integrating the proposed DEF into the CCC framework. Further exploration may involve examining DEF's benefits for other illuminant estimation techniques that rely on convNets and raw image pixels as input. Another research direction could involve developing a spatially varying version of DEF, rather than our global feature, to utilize for spatially varying illuminant estimation and image white balancing." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.549, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Abdelhamed, A., Punnappurath, A., Brown, M.S.: Leveraging the availability of two cameras for illuminant estimation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.779, + 0.246 + ], + "angle": 0, + "content": "3. Affi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.273 + ], + "angle": 0, + "content": "4. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.3 + ], + "angle": 0, + "content": "5. Affifi, M., Brown, M.S.: Sensor-independent illumination estimation for DNN models. In: BMVC (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "6. Affi, M., Punnappurath, A., Finlayson, G., Brown, M.S.: As-projective-as-possible bias correction for illumination estimation algorithms. Journal of the Optical Society of America A 36(1), 71-78 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "7. Banić, N., Koščević, K., Lončarić, S.: Unsupervised learning for color constancy. arXiv preprint arXiv:1712.00436 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.653, + 0.384 + ], + "angle": 0, + "content": "8. Barron, J.T.: Convolutional color constancy. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.723, + 0.397 + ], + "angle": 0, + "content": "9. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.72, + 0.412 + ], + "angle": 0, + "content": "0. Bianco, S., Schettini, R.: Color constancy using faces. In: CVPR (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "1. Brainard, D.H., Freeman, W.T.: Bayesian color constancy. Journal of the Optical Society of America A 14(7), 1393-1411 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.467 + ], + "angle": 0, + "content": "2. Brown, M.: Color processing for digital cameras, chap. 5, pp. 81-98. John Wiley & Sons, Ltd (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "3. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin Institute 310(1), 1-26 (1980)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.536 + ], + "angle": 0, + "content": "4. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. Journal of the Optical Society of America A 31(5), 1049-1058 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "5. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: Annual Conference on Computer Graphics and Interactive Techniques (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "6. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "7. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Transactions on Graphics 36(6) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.706, + 0.647 + ], + "angle": 0, + "content": "8. Fairchild, M.D.: Color appearance models. John Wiley & Sons (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.674 + ], + "angle": 0, + "content": "9. Finlayson, G.D., Funt, B.V., Barnard, K.: Color constancy under varying illumination. In: ICCV (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "20. Finlayson, G.D., Hordley, S.D.: Color constancy at a pixel. Journal of the Optical Society of America A 18(2), 253-264 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "21. Finlayson, G.D., Hordley, S.D.: The chromagenic colour camera and illuminant estimation. In: Color Imaging Conference (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "22. Finlayson, G.D., Hordley, S.D., Morovic, P.: Chromagenic colour constancy. In: Congress of the International Colour Association (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "23. Finlayson, G.D., Hordley, S.D., Morovic, P.: Colour constancy using the chromogenic constraint. In: CVPR (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "24. Finlayson, G.D., Hordley, S.D., Tastl, I.: Gamut constrained illuminant estimation. International Journal of Computer Vision 67(1), 93-109 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "25. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference (2004)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "M. Afifi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "26. Fredembach, C., Finlayson, G.: The bright-chromagenic algorithm for illuminant estimation. Journal of Imaging Science and Technology 52, 40906:1-40906:11 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "27. Gehler, P.V., Rother, C., Blake, A., Minka, T., Sharp, T.: Bayesian color constancy revisited. In: CVPR (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "28. Gelfand, N., Adams, A., Park, S.H., Pulli, K.: Multi-exposure imaging on mobile devices. In: ACM International Conference on Multimedia (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "29. Gijsenij, A., Gevers, T., Van De Weijer, J.: Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision 86(2), 127-139 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "30. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing 20(9), 2475-2489 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "31. Gijsenij, A., Gevers, T., Van De Weijer, J.: Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(5), 918-929 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.379 + ], + "angle": 0, + "content": "32. Granados, M., Ajdin, B., Wand, M., Theobalt, C., Seidel, H.P., Lensch, H.P.: Optimal HDR reconstruction with linear digital cameras. In: CVPR (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "33. Hanji, P., Mantiuk, R.K., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "34. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A K-means clustering algorithm. Journal of the royal statistical society. Series C (applied statistics) 28(1), 100-108 (1979)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.488 + ], + "angle": 0, + "content": "35. Hernandez-Juarez, D., Parisot, S., Busam, B., Leonardis, A., Slabaugh, G., McDonagh, S.: A multi-hypothesis approach to color constancy. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.515 + ], + "angle": 0, + "content": "36. Hu, Y., Wang, B., Lin, S.: FC4: Fully convolutional color constancy with confidence-weighted pooling. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.542 + ], + "angle": 0, + "content": "37. Hubel, P.M., Finlayson, G.D., Hordley, S.D.: White point estimation using color by convolution (2007), US Patent 7,200,264" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.569 + ], + "angle": 0, + "content": "38. Jung, C., Yang, Y., Jiao, L.: High dynamic range imaging on mobile devices using fusion of multiexposure images. Optical Engineering 52(10), 102004-102004 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.596 + ], + "angle": 0, + "content": "39. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.623 + ], + "angle": 0, + "content": "40. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "41. Laakom, F., Passalis, N., Raitoharju, J., Nikkanen, J., Tefas, A., Iosifidis, A., Gabbouj, M.: Bag of color features for color constancy. IEEE Transactions on Image Processing 29, 7722-7734 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.691 + ], + "angle": 0, + "content": "42. Laakom, F., Raitoharju, J., Nikkanen, J., Iosifidis, A., Gabbouj, M.: Robust channel-wise illumination estimation. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "43. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: WACV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "44. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., et al.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 164-1 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "45. Lo, Y.C., Chang, C.C., Chiu, H.C., Huang, Y.H., Chen, C.P., Chang, Y.L., Jou, K.: CLCC: Contrastive learning for color constancy. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "46. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Lou, Z., Gevers, T., Hu, N., Lucassen, M.P., et al.: Color constancy by deep learning. In: BMVC (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.549, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "HDR illuminant estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "48. Oh, S.W., Kim, S.J.: Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition 61, 405-416 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "49. Qi, G., Chang, L., Luo, Y., Chen, Y., Zhu, Z., Wang, S.: A precise multi-exposure image fusion method based on low-level features. Sensors 20(6), 1597 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.231 + ], + "angle": 0, + "content": "50. Qian, Y., Kamarainen, J.K., Nikkanen, J., Matas, J.: On finding gray pixels. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "51. Shen, R., Cheng, I., Shi, J., Basu, A.: Generalized random walks for fusion of multi-exposure images. IEEE Transactions on Image Processing 20(12), 3634-3646 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "52. Solhusvik, J., Hu, S., Johansson, R., Lin, Z., Ma, S., Mabuchi, K., Manabe, S., Mao, D., Phan, B., Rhodes, H., et al.: A 1392x976 2.8 \\(\\mu\\)m 120db CIS with per-pixel controlled conversion gain. In: International Image Sensor Workshop (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.342 + ], + "angle": 0, + "content": "53. Solomatov, G., Akkaynak, D.: Spectral sensitivity estimation without a camera. In: ICCP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "54. Tang, Y., Kang, X., Li, C., Lin, Z., Ming, A.: Transfer learning for color constancy via statistic perspective. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.397 + ], + "angle": 0, + "content": "55. Ulucan, O., Karakaya, D., Turkan, M.: Multi-exposure image fusion based on linear embeddings and watershed masking. Signal Processing 178, 107791 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.438 + ], + "angle": 0, + "content": "56. Ulucan, O., Ulucan, D., Ebner, M.: Block-based color constancy: The deviation of salient pixels. In: International Conference on Acoustics, Speech and Signal Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.466 + ], + "angle": 0, + "content": "57. Van De Weijer, J., Gevers, T., Gijsenij, A.: Edge-based color constancy. IEEE Transactions on Image Processing 16(9), 2207-2214 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.522 + ], + "angle": 0, + "content": "58. Willassen, T., Solhusvik, J., Johansson, R., Yaghmai, S., Rhodes, H., Manabe, S., Mao, D., Lin, Z., Yang, D., Cellek, O., et al.: A \\(1280 \\times 1080\\) 4.2 \\(\\mu\\)m split-diode pixel HDR sensor in \\(110\\mathrm{nm}\\) BSI CMOS process. In: International Image Sensor Workshop (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.549 + ], + "angle": 0, + "content": "59. Xiao, J., Gu, S., Zhang, L.: Multi-domain learning for accurate and few-shot color constancy. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.577 + ], + "angle": 0, + "content": "60. Xing, X., Qian, Y., Feng, S., Dong, Y., Matas, J.: Point cloud color constancy. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.605 + ], + "angle": 0, + "content": "61. Xu, B., Wang, N., Chen, T., Li, M.: Empirical evaluation of rectified activations in convolutional network. arXiv preprint arXiv:1505.00853 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.646 + ], + "angle": 0, + "content": "62. Yahiaoui, L., Horgan, J., Deegan, B., Yogamani, S., Hughes, C., Denny, P.: Overview and empirical analysis of ISP parameter tuning for visual perception in autonomous driving. Journal of Imaging 5(10), 78 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.674 + ], + "angle": 0, + "content": "63. Yu, H., Chen, K., Wang, K., Qian, Y., Zhang, Z., Jia, K.: Cascading convolutional color constancy. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.703 + ], + "angle": 0, + "content": "64. Yue, S., Wei, M.: Color constancy from a pure color view. Journal of the Optical Society of America A 40(3), 602-610 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.703 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_origin.pdf b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..17481a0f1587197f1ba488ccde27e1e9e9de7016 --- /dev/null +++ b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/34e042fa-0cc0-4e4a-bee5-6c9ff75dce26_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50251f9f1b738bd47fb0bfef6dca486fa3383e631b60712904b4ad44d5391545 +size 2870012 diff --git a/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/full.md b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d71d497cff63694413b7aaf1ff06eeebb279c031 --- /dev/null +++ b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/full.md @@ -0,0 +1,265 @@ +# Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging + +Mahmoud Affifi, Zhenhua Hu, and Liang Liang + +Google {mafifi, zhenhuahu, annyliang}@google.com + +Abstract. High dynamic range (HDR) imaging involves capturing a series of frames of the same scene, each with different exposure settings, to broaden the dynamic range of light. This can be achieved through burst capturing or using staggered HDR sensors that capture long and short exposures simultaneously in the camera image signal processor (ISP). Within camera ISP pipeline, illuminant estimation is a crucial step aiming to estimate the color of the global illuminant in the scene. This estimation is used in camera ISP white-balance module to remove undesirable color cast in the final image. Despite the multiple frames captured in the HDR pipeline, conventional illuminant estimation methods often rely only on a single frame of the scene. In this paper, we explore leveraging information from frames captured with different exposure times. Specifically, we introduce a simple feature extracted from dual-exposure images to guide illuminant estimators, referred to as the dual-exposure feature (DEF). To validate the efficiency of DEF, we employed two illuminant estimators using the proposed DEF: 1) a multilayer perceptron network (MLP), referred to as exposure-based MLP (EMLP), and 2) a modified version of the convolutional color constancy (CCC) to integrate our DEF, that we call ECCC. Both EMLP and ECCC achieve promising results, in some cases surpassing prior methods that require hundreds of thousands or millions of parameters, with only a few hundred parameters for EMLP and a few thousand parameters for ECCC. + +Keywords: Computational color constancy $\cdot$ Illuminant estimation $\cdot$ HDR imaging + +# 1 Introduction and Related Work + +Camera image signal processor (ISP) comprises several modules, each dedicated to enhancing specific aspects of the quality of captured raw images by the camera sensor [16]. One key component among these modules is the white-balance module, which aims to eliminate undesirable color casts introduced by the combination of scene lighting and camera sensitivity. To achieve this, the auto white-balance module runs an illuminant estimator onboard the camera ISP to estimate the RGB color of the illuminant, under the assumption that a single global illuminant illuminates the scene for simplicity [30, 44]. Image white balancing can thus be described as follows: + +![](images/624c42bbf13ed50c54d147571f7754ba36c46b57c60088198a8fc60090b33093.jpg) +Full-resolution raw image +Raw frames with different exposure times +Fig. 1: Conventional illuminant estimators often rely on a single frame for illuminant color estimation. Although the HDR camera pipeline includes at least two frames of the same scene, conventional methods usually consider only a single frame (either with auto exposure, shown in the dashed green line, merged frame, or one frame of the burst capture) [12,16,44,62]. This paper proposes leveraging information from dual-exposure capturing in multi-exposure HDR imaging to enhance illuminant estimation in camera pipelines. Our method uses frames captured at long and short exposures, available in multi-exposure bursts [28] or staggered HDR sensors [52, 58]. Achieving comparable or superior results, our method employs lightweight models ( $\sim$ 300-6000 parameters) compared to those using hundreds of thousands or millions of parameters. In this figure and the following figures, all raw images have the gamma operator applied to aid visualization, and all sRGB images are rendered using the HDR+ camera pipeline [44]. + +![](images/adfd9146708b0ec3c480ce71a33c61177aee412bcc5386c04114f28be22b6875.jpg) +Dual-exposure frames with long exposure times + +![](images/6fb6fabfbc74ff841dde9dddc176556fbfcf223f5f5d2cc28aaeb2491330dca5.jpg) + +![](images/1c614014ad931b16fbcffe465d0a06c1479ebb18cc9a2ccf855600573a4b5d44.jpg) +Final rendered image with our white-balance correction + +$$ +I _ {\mathrm {W B}} = \operatorname {d i a g} \left(\left[ \frac {\hat {\ell} _ {G}}{\hat {\ell} _ {R}}, 1, \frac {\hat {\ell} _ {G}}{\hat {\ell} _ {B}} \right] ^ {T}\right) I, \tag {1} +$$ + +$$ +\left[ \hat {\ell} _ {R}, \hat {\ell} _ {G}, \hat {\ell} _ {B} \right] ^ {T} = f (I), \tag {2} +$$ + +where $I$ and $I_{\mathrm{WB}}$ are $3 \times k$ RGB colors of raw image and white-balanced raw image, respectively, with $k$ refers to the total number of pixels in the image, $\operatorname{diag}(.)$ creates a $3 \times 3$ diagonal matrix from a given vector, $\hat{\ell} \in \mathbb{R}^3$ refers to the estimated illuminant color vector, $T$ refers to vector transpose, and $f$ is an illuminant estimator function. + +Modern camera ISPs capture several frames of each scene with different exposure times to enhance final image dynamic range [28,38]. This can be achieved through burst capturing, involving rapid capture with varying exposure times in quick succession, or by using staggered high dynamic range (sHDR) sensors that simultaneously capture long and short exposures [52, 58]. The captured images are then combined, incorporating multiple exposures of the same scene at different levels, resulting in a greater dynamic range than what would be possible with a single image [43, 49, 51, 55]. This HDR imaging camera pipeline, therefore, includes more than a single frame of the captured scene, and internal camera ISP modules, such as the auto white-balance module, can access these additional frames. While such additional frames may have beneficial information to help illuminant estimators, conventional illuminant estimation methods often rely on a single frame (e.g., [9, 36, 41, 47, 54, 63, 64]) in both traditional single-frame camera pipelines [39] and multi-frame HDR ISPs [44]. + +Such single-frame illuminant estimators can be categorized into: 1) statistical methods that rely on statistics computed from input raw image information (e.g., colors, edges) [13, 14, 25, 31, 50, 56, 57] and 2) learning-based methods that learn, from a set of training images labeled with ground-truth illuminant colors, to map from input raw color information to the corresponding illuminant color [6, 8, 10, 11, 24, 36, 42, 45, 47, 48, 54, 63, 64]. While the latter category typically achieves better results than the statistical methods, most learning-based methods are camera-dependent, meaning that they require domain adaptation or fine-tuning when deployed on new cameras to achieve similar accuracy on cameras used for training due to the influence of the camera response function on both raw image colors and ground-truth colors [3-5, 35, 53, 59]. + +A limited number of attempts have proposed learning-based methods that go beyond the single-frame input scheme. For example, the cross-camera convolutional color constancy (C5) [4] suggests utilizing additional unlabeled images captured by the testing camera, in addition to the primary single-frame image of the scene being captured to improve the generalization for cameras that were not included in the training phase. Xing et al. [60] proposed the use of a depth map, captured by a time-of-flight (ToF) sensor, along with the primary raw image to predict illuminant color in the scene by leveraging the geometry information obtained from depth map. Abdelhamed et al., [2] proposed leveraging the presence of two cameras in most modern mobile phone devices. Assuming dual streaming from both cameras, they derived a feature within the framework of chromagenic color constancy theory [21-23] to enhance the accuracy of illuminant estimation, yielding promising results. + +Our method, in contrast to the majority of prior work, adopts the strategy of benefiting from multiple frames available in the camera ISP (similar to [2]). However, unlike [2, 60], which requires streaming from dual cameras and may lead to impractical high power consumption, our method relies on two frames captured of the same scene under different exposure settings, already present in the HDR imaging pipeline, to estimate the illuminant color of the captured scene (see Fig. 1). Specifically, we utilize: 1) a frame captured with short exposure time, that we refer to as short-exposure image $(I_s)$ , and 2) another frame of the same scene captured with long exposure time, that we refer to as long-exposure image $(I_l)$ . Having dual-exposure images in the HDR imaging pipeline is feasible, making our method practical for most HDR imaging pipelines. Accordingly, Eq. 2 can be modified to include our dual-exposure input images as follows: + +$$ +[ \hat {\ell} _ {R}, \hat {\ell} _ {G}, \hat {\ell} _ {B} ] ^ {T} = f (I _ {l}, I _ {s}). \tag {3} +$$ + +Our objective in this work is to design the function $f$ in a manner that enables the effective utilization of the additional information provided by the dual-exposure images, $I_{l}$ and $I_{s}$ . + +Contribution In this paper, we present a feature inspired by the chromagenic color constancy theory [21-23], termed the dual-exposure feature (DEF), that is derived from images captured with both short exposure time $(I_s)$ and long + +exposure time $(I_l)$ . DEF leverages the variations in chromatic information between these dual-exposure images, providing valuable guidance for illuminant estimator methods. To assess its effectiveness, we trained a lightweight multi-layer perceptron network (MLP) for illuminant estimation that utilizes solely our DEF as input, departing from the conventional approach of using actual RGB values of the captured scene colors. Additionally, we explored the integration of DEF into an established color constancy framework, specifically the convolutional color constancy (CCC) [4,8,9,37], which we refer to as exposure-based CCC (ECCC). The experimental results on a multi-exposure dataset, collected to evaluate our work empirically, show that these models - namely, EMLP and ECC - achieve promising results with a reasonable number of parameters - 354 learnable parameters for EMLP and 6,156 learnable parameters for ECC. This outperformance across diverse evaluation metrics is observed when comparing to prior methods that require significantly higher number of parameters, ranging from tens to hundreds of thousands, or even millions. + +# 2 Illuminant-Linked Dual Exposure Feature + +To develop an efficient illuminant estimator that benefits from both $I_{l}$ and $I_{s}$ , we introduce a compact feature that aims to capture the correlation between these dual-exposure images. Our feature is inspired by the chromagenic color constancy theory [21-23]. Specifically, we explore the analogy between long and short exposure images, $I_{l}$ and $I_{s}$ , and aligned images captured by two cameras [2] under the chromagenic color constancy theory, leading us to develop our dual-exposure feature (DEF). To begin, we review the chromagenic color constancy theory under the Lambertian reflectance model with a single illuminant assumption. The captured raw image $I$ can mathematically be described by: + +$$ +I _ {\rho} ^ {(y)} = \int_ {\gamma} S _ {\rho} (x) D (x) R (y, x) d x + z, \tag {4} +$$ + +where $S(\cdot), D(\cdot)$ , and $R(\cdot)$ represent the camera response function (typically represented by camera sensitivity, infrared cut-off filter, and spectral lens transmission), the spectral power distribution of light, and scene reflectance, respectively. Here, $x$ refers to a wavelength within the visible range $\gamma$ , $y$ refers to the pixel location in image $I$ , $\rho \in \{R,G,B\}$ refers to the color channel, and $z$ denotes the undesired noise, typically represented by signal-dependent and additive components [1]. According to the chromagenic color constancy theory [22], if we capture two images of the same scene with a specially chosen chromagenic filter, $Q$ , applied between image captures, the linear color matrix that maps between those captured images is unique to the illuminant color present in this scene. That is, given an image, $I$ , that is captured by the main camera, and a filtered image, $I_f$ , that can be described as: + +$$ +I _ {f _ {\rho}} ^ {(y)} = \int_ {\gamma} S _ {\rho} (x) Q _ {\rho} (x) D (x) R (y, x) d x + z, \tag {5} +$$ + +the $3 \times 3$ color matrix $C_c$ that maps between the colors of $I$ and $I_f$ is indicative of the scene illuminant. This color matrix can be computed by minimizing the following equation: + +$$ +\underset {C _ {c}} {\arg \min } \| C _ {c} I _ {f} - I \| _ {F}, \tag {6} +$$ + +where $\| \cdot \| _F$ is the Frobenius norm. The closed-form solution for Eq. 6 can be obtained using the pseudoinverse (i.e., $C_c = II_f^\dagger$ ). While theoretically validated, the chromagenic filter conditions required to obtain a unique mapping matrix per illuminant are challenging to meet in practice. Empirically, Finlayson et al., [23] found that most filters exhibit a reasonable level of correlation between the computed matrix and the illuminant color, excluding neutral filters, which consistently result in a scaling relationship between the unfiltered image, $I$ , and the filtered image, $I_{f}$ . Relaxing the conditions to include normal colored filter, Abdelhamed et al., [2] proposed using two cameras, with the second camera serving as the main camera after applying a colored filter—i.e., $S(\cdot)Q(\cdot)$ . Surprisingly, even when chromagenic filter conditions are not met, using a normal colored filter [23] or another camera with a different response function [2] still shows a correlation between the computed matrix $C_c$ and the illuminant color to some extent, achieving promising results. + +We argue that this correlation arises due to the variations, or what can be considered a form of "distortion", in the colors captured by the second (filtered) camera when compared to the original colors captured by the main camera. This color distortion varies based on the interplay between the scene irradiance, $D(\cdot)R(\cdot)$ , and the camera response functions - namely, the main camera's $S(\cdot)$ and the filtered/second camera's $S(\cdot)Q(\cdot)$ in Eqs. 4 and 5. In the context of machine learning (ML) models, the camera response functions remain fixed across the entire dataset. Consequently, ML models, such as the one proposed in [2], learn the correlation between the ground-truth illuminant color and the scene irradiance through the matrix $C_c$ that represents the level of "distortion" between the colors of captured scene images. + +In a dual-exposure setup, we have a long-exposure image, $I_{l}$ , and a short-exposure image, $I_{s}$ , both capturing the same scene. The extent of color distortion in each image varies based on the scene irradiance; for instance, $I_{s}$ may exhibit a higher level of color distortion and noise than $I_{l}$ under suboptimal lighting conditions (e.g., indoor light), while $I_{l}$ may have a higher level of color distortion and over-saturated colors than $I_{s}$ in well-lit scenes (e.g., outdoor light). This difference in color distortion arises because the non-linear camera response function receives a different number of photons to form each image colors based on the exposure time and scene irradiance [32]. As a result, $I_{l}$ and $I_{s}$ can exhibit color variations that differ across individual color channels [15], that are somewhat akin to those caused by a color filter (though to a lesser extent). Even within the same color channel, different levels of color differences between $I_{l}$ and $I_{s}$ can be observed spatially due to the interaction between the non-linear camera response function of that channel and the spatially varying scene irradiance (see + +![](images/8a50bb5a388c976a7d741d29a232d1e3442348dcc5b03e515c91b834bfc5a225.jpg) +Fig. 2: Cameras perceive different amounts of photons when capturing the same scene with different exposure times. Images taken with both long and short exposure times exhibit variations in each channel due to the camera response function and scene irradiance. Additionally, spatial variations, in each color channel, can be observed based on object reflectance, as the interplay of lighting, object reflectance, and the camera response function leads to different outcomes. (A) and (B) show raw images of scenes captured under indoor and outdoor lighting, respectively. In (C) and (D), we present the average $rg$ -chromaticity histogram and aggregated red, green, and blue pixel values from 25 images sharing similar lighting conditions in (A) and (B), respectively. + +Fig. 2). In Fig. 2, we show two examples of dual-exposure images captured under different lighting conditions. As can be seen, the differences between the colors in $I_{l}$ and $I_{s}$ under each lighting condition exhibit variations, which are observable in the histogram similarity (here we use the Bhattacharyya distance similarity metric) and variations of the ratios in each color channel between $I_{l}$ and $I_{s}$ . That is, the correlation between $I_{l}$ and $I_{s}$ is not always a proportional scaling, as in the case of a neutral filter. We observed similar patterns as shown in Fig. 2 when studying examples from the two-camera dataset in [2] (see supplemental materials). + +Based on this discussion, we propose to compute the color matrix $C_c$ to map between the $rgb$ -chromaticity values (i.e., $[R / \kappa, G / \kappa, B / \kappa]^T$ , with $\kappa = R + G + B$ ) of $I_s$ and $I_l$ . The reason for not using the RGB triples, similar to chromagenic color constancy, is that we aim to reduce the influence of intensity differences between $I_l$ and $I_s$ when computing $C_c$ . In addition, we compute the covariance matrix, $C_v$ , of the ratio image $X \in \mathbb{R}^{3 \times k}$ , where $X_{\rho}^{(y)} = I_{s_{\rho}}^{(y)} / \left(I_{l_{\rho}}^{(y)} + \epsilon\right)$ , and $\epsilon$ is a small number added for numerical stability. Computing $C_v$ is performed as described in Eq. 7 to measure the variance in each color channel between $I_s$ and $I_l$ and the joint variability across channels. + +$$ +C _ {v} (X) = \mathbb {E} \left[ (X - \mathbb {E} [ X ]) (X - \mathbb {E} [ X ]) ^ {T} \right], \tag {7} +$$ + +where $\mathbb{E}(\cdot)$ is the expected value (mean) of the matrix. Both $C_c$ and $C_v$ form our dual-exposure feature (DEF) that represents the differences in color distortion in $I_l$ and $I_s$ . Our DEF is represented as a vector $\in \mathbb{R}^{15}$ , where we exclude + +![](images/4b3d06d705088c112421a9f1450023c45d66abf1a10489659fced4a1da18ee52.jpg) +Fig. 3: We present an illuminant-related dual-exposure feature (DEF), derived from a pair of images captured with short and long exposures. Using DEF, we deploy a simple multilayer perceptron network (MLP) with only 354 parameters, referred to as exposure-based MLP or EMLP, for illuminant estimation, as shown in (A). We further explore the integration of DEF into the CCC framework, as shown in (B), by dynamically generating bias map based on DEF. We denote this modified CCC framework as exposure-based CCC or ECCC. + +the redundant values in $C_v$ over the symmetric positions. To evaluate DEF's effectiveness as a clue for scene illuminant colors, we employed a lightweight MLP for scene illuminant estimation (Fig. 3-A), that we call exposure-based MLP (EMLP). EMLP relies solely on our DEF as input. It comprises an input fully connected (fc) layer with 9 output neurons, followed by leaky ReLU (LReLU) [61], two hidden fc layers, each with 9 output neurons with LReLU in between, and an output fc layer with three neurons. EMLP achieves results comparable to complex models with thousands or millions of parameters (refer to Tables 1 and 2), while maintaining a lightweight design with just a few hundred parameters. Consequently, DEF is empirically shown as a valuable feature providing strong insights into the illuminant scene color. + +# 3 Integration with Convolutional Color Constancy + +In this section, we incorporate the proposed DEF into one of the most established frameworks for illuminant estimation. Specifically, we introduce modifications to the convolutional color constancy (CCC) framework [4, 8, 9, 37] to leverage the benefits of the DEF, which we refer to as exposure-based CCC or ECCC for short. It is important to note that ECCC does not introduce a new CCC method; instead, it serves as an illustration of how the DEF can be seamlessly integrated into existing, well-established illuminant estimation frameworks. + +Let's begin with a brief description of the CCC [8, 9]. Given a single raw image, the CCC operates by learning one or more convolutional 2D filters, denoted as $\{F_j\}$ , which convolve over the 2D histogram(s), $\{H_j\}$ , of the image colors in the $uv$ color space (i.e., the log of $G / R$ and $G / B$ chroma values of pixel colors) [8, 20]. This convolutional operation can be accelerated by FFTs as + +proposed in [9]. Afterwards, a 2D bias map, $B$ , is added, followed by a softmax operation, $\sigma$ , to compute the "probability" map, $P$ , of the illuminant bin in the uv 2D histogram space. This simplified version of the CCC can be described as follows: + +$$ +P = \sigma \left(\sum_ {j} \left(F _ {j} * H _ {j}\right) + B\right). \tag {8} +$$ + +In CCC [8,9], the filters and bias are learned across the entire training dataset. This means it consists of a single bias and one or more filters (the number depends on the histograms used, either a single histogram for image colors or two histograms, including the edge color histogram). Later, C5 [4] proposes a hypernetwork that dynamically generates filters and bias based on the input raw image and additional images taken from the same camera, aiming to improve the generalization of CCC across cameras. While cross-camera color constancy is out of the scope of this paper, we propose to use our DEF to dynamically "generate" a bias map based on the input image. That is, we learn a bank of biases, $\{B_i\}$ , where $i \in \{1, \dots, n\}$ , such that we linearly interpolate between them based on blending weights emitted from an MLP network that processes our DEF. In this way, the DEF controls the bias of the CCC model, acting as "candidate" illuminant priors within the uv space for the input image (see Fig. 3-B). To implement this, we need to have multiple bias maps, which definitely will lead to an impractical increase in model size. Thus, we propose to use a downsampled size $(1/4$ of histogram size) for the learnable biases, $\{B_i\}$ . Furthermore, we propose to feed two histograms (i.e., $j \in \{l, s\}$ ) of both $I_l$ and $I_s$ , denoted as $H_l$ and $H_s$ , respectively, into the model. This leads to the learning of two downsampled filters, $\{F_j\}$ , corresponding to the histograms of the long-exposure image $(I_l)$ and the short-exposure image $(I_s)$ . Learning small-sized filters and biases has the following benefits. First, we can learn many biases within an affordable model size (e.g., $n = 20$ requires $\sim 6\mathrm{K}$ parameters, while FFCC [9] with a single bias requires $\sim 12\mathrm{K}$ parameters). Second, it implicitly produces smooth filters and biases, which are desirable to avoid overfitting [4,9]. With this modification, Eq. 8 can be rewritten as follows: + +$$ +P = \sigma \left(\sum_ {j} \left(\uparrow \left(F _ {j}\right) * H _ {j}\right) + B _ {\mathrm {u p}} ^ {\prime}\right), \tag {9} +$$ + +$$ +B _ {\mathrm {u p}} ^ {\prime} = \uparrow \left(\sum_ {i = 1} ^ {n} w _ {i} B _ {i}\right). \tag {10} +$$ + +where $\uparrow (\cdot)$ refers to upscaling through bilinear interpolation, $[w_{1},\dots,w_{n}]^{T}$ is a weighting vector produced by a lightweight MLP (similar to EMLP, but with $n$ output neurons) that processes our DEF. After computing $P$ , the estimated illuminant can be obtained by: + +![](images/618580631a98426f579bf158e013215a7ee2148b60ded767f5544d3053a0a5fb.jpg) +Fig. 4: Examples from the dataset used in this work. For each scene, we captured the scene with a gray calibration object placed in the scene to obtain the ground-truth illuminant (A) and captured the scene using different exposure settings without the gray object (B-H). The terms 'short /e' (C-E) and 'long $\times e$ (F-H) refer to multiplying and dividing auto exposure time by a factor $e$ , respectively. The first image in (A) is displayed in sRGB, while the rest are shown in raw RGB space. + +$$ +\hat {\ell} _ {u} = \sum_ {u, v} u P (u, v), \hat {\ell} _ {v} = \sum_ {u, v} v P (u, v), \tag {11} +$$ + +$$ +\hat {\ell} = \left[ \exp \left(- \hat {\ell_ {u}}\right) / q, 1 / q, \exp \left(- \hat {\ell_ {v}}\right) / q \right] ^ {T}, \tag {12} +$$ + +$$ +q = \sqrt {\exp \left(- \hat {\ell_ {u}}\right) ^ {2} + \exp \left(- \hat {\ell_ {v}}\right) ^ {2} + 1}. \tag {13} +$$ + +# 4 Experiments + +# 4.1 Data + +To the best of our knowledge, the existing HDR multi-exposure datasets (e.g., [17, 33]) do not provide ground-truth illuminant colors for training and evaluation of our method. Thus, we compiled a dataset of scenes captured with both auto exposure and various multiple-exposure settings for each scene using Pixel 7 Pro camera (see Fig. 4). Images captured with auto exposure are used for training (in the case of learning-based methods) and evaluating other methods, while those captured with multiple exposure settings are employed for validating our method. Our set comprises 558 scenes, each captured using auto exposure and six additional exposure settings: $\times 2$ , $\times 4$ , $\times 8$ , $1/2$ , $1/4$ , and $1/8$ , indicating adjustments to the original exposure time. Specifically, $\times 2$ signifies doubling the original exposure time, $\times 4$ and $\times 8$ denote quadrupling and octupling, respectively, while $1/2$ , $1/4$ , and $1/8$ represent dividing the original exposure time by 2, 4, and 8, respectively. These adjustments are dynamically computed as ratios from the initial auto exposure, following the common practice in multi-exposure + +HDR imaging on smartphone cameras [28, 38]. The camera was fixed on a tripod to minimize potential misalignment between images captured by different exposure settings. In total, we captured 3,906 raw images ( $558 \times 7$ captures of each scene). The dataset is organized as follows: 83 and 86 scenes were randomly selected for validation and testing, respectively, with the remaining 387 scenes were used for training. Additionally, each scene was captured with a gray calibration object to obtain the ground-truth illuminant. The images with the gray calibration object are not included in any of the training, validation, nor testing sets. + +During training, we optionally augment training data using chromatic adaptation. Specifically, we employ clustering on the training data using our DEF as query feature. We use K-means [34] with L2 to create 80 clusters. Subsequently, we augment the training data by generating three additional images for each sample. For each augmented image, we randomly select an illuminant color from the cluster to which the original image belongs. We then apply Von Kries transform [18] to remove the original illuminant from the image, using the actual ground truth, and apply the newly selected illuminant color to the image using the diagonal scaling operator. This new illuminant serves as the ground truth for the augmented image. As our method is designed to handle two images, namely $I_{s}$ and $I_{l}$ , we use, without loss of generality, a single exposure factor, $e$ , for both images, such that $I_{s}$ was captured with $1 / e$ of the auto-exposure time, while $I_{l}$ was captured with $\times e$ of the auto-exposure time, for simplicity. In Sec. 4.3, we present an ablation study on the impact of different values of $e$ on our final results. + +# 4.2 Training + +We trained EMLP and ECCC using the Adam optimizer [40] with a mini-batch size set to 32 for 1000 and 200 epochs, respectively. For EMLP, the learning rate was $10^{-3}$ , while ECCC was trained with an incremental mini-batch size (similar to [4]) with a learning rate of $5 \times 10^{-3}$ with a cosine annealing schedule [46], and the weight decay (i.e., L2 regularization penalty) was set to $10^{-5}$ . For ECCC, the filter weights were initialized to zeros, and the biases were initialized to $n$ 2D histograms of training ground-truth illuminant colors after clustering into $n$ clusters using our DEF as a query feature with K-means. The $n$ histograms of training ground-truth illuminants were first processed by morphological dilation using a $3 \times 3$ diamond-shaped structuring element before being used as initial values for our learnable biases. This initialization improves the results, as it assists in establishing reasonable "candidate" illuminant priors linked to the corresponding DEFs (see Sec. 4.3 for an ablation study). Both models were trained using the angular error, $L(\cdot)$ , between the predicted illuminant color $\hat{\ell}$ and the ground-truth illuminant $\ell$ , as described in the following equation [19]: + +$$ +L (\ell , \hat {\ell}) = \cos^ {- 1} \left(\frac {\ell \cdot \hat {\ell}}{\| \ell \| \| \hat {\ell} \|}\right), \tag {14} +$$ + +where $\|\cdot\|$ denotes the Euclidean norm, and $(\cdot)$ represents the vector dot product. For ECCC, we further added two smoothness loss terms to encourage smoothness in the learned filters and biases, similar to [4]. See supplemental materials for more details. + +# 4.3 Results + +We conducted a comprehensive comparison of our proposed methods against various existing techniques, which include: training-free statistical methods [13, 14, 25, 50], camera-independent learning-based methods [4, 5], and camera-dependent learning-based methods [6, 9, 29, 41, 42, 54, 63, 64]. For the camera-dependent learning-based methods, we trained each model on our data using the provided code and recommended parameters by the respective authors. Meanwhile, for sensor-independent learning methods, we utilized the provided pre-trained models on other datasets (e.g., [7, 14, 27]). + +For the sake of conducting realistic experiments, we assessed all methods, including ours, on $384 \times 256$ images—a suitable size for evaluating illuminant estimation methods designed for embedded hardware devices; such devices may utilize even smaller image sizes in white-balance modules [9]. For CCC methods [4,9], including the ECCC, $64 \times 64$ histograms were used. The ECCC incorporates the proposed DEF into the CCC by generating an interpolated bias map based on the DEF. This can be seen as a spatial case analogous to C5 [4], where C5 utilizes a neural network to emit bias and convolution filters based on additional images taken by the same camera to enhance generalization across cameras. To validate our modification against C5, we trained different versions of C5 not aimed at improving cross-camera generalization (as our method also focuses on a single camera). We refer to these modified versions as follows: + +- C5 (model A): We use the histogram of the averaged short and long exposure images without additional images. That is, the C5 network uses a single histogram to generate CCC model parameters. +- C5 (model B): We use histograms of both images taken by long and short exposures as input to the C5 network to generate CCC model parameters. The CCC mode is then applied to the histogram of the averaged long-short images. +- C5 (models C and D): Both Model C and Model D use histograms of both images taken by long and short exposures to feed the C5 network and generate CCC model that is then applied to histogram of short exposure image (model C) and long exposure image (model D). + +The results are given in Tables 1 and 2 on the validation and testing sets, respectively. We report the mean, median, tri-mean, best $25\%$ , worst $25\%$ , worst $5\%$ , and max angular errors in each set. We also report the total number of parameters for other methods, including ours. Additionally, we present the results of a set of ablation studies conducted to examine the impact of the exposure factor $e$ in both EMLP and ECCC. We also studied the impact of different color + +Table 1: Angular errors on the validation set. $\diamond$ and $\star$ denote camera-independent models and training-free statistical methods, respectively. Methods are listed chronologically by publication year, with the top and second-best results highlighted in yellow and red. We present the results of camera-dependent trained models for various versions of C5 [4], identified as models A, B, C, and D (see main text for more details). Our results are reported using EMLP, ECCC, and the ensemble model (EMLP + ECCC). Ablation studies are included using different mapping matrices (CM: $3 \times 3$ color mapping matrix, TM: $3 \times 4$ affine transformation matrix, HM: $3 \times 3$ homography matrix), different color representations (rgb: raw RGB triplet as used in [2,21-23,26], rg/rgb-chroma: rg-chromaticity/rgb-chromaticity of the raw RGB triplet), different exposure ratios ( $e \in \{2,4,8\}$ ), different input image sizes and histograms ( $s$ indicates $48 \times 32$ input images, $s^2$ indicates $48 \times 32$ input images and $32 \times 32$ histograms), using different input histograms for ECCC of both long and short exposure images, average image, and long/short image, different number of learned biases ( $n$ ), and results with (w/) and without (w/o) augmentation, w/ and w/o the covariance matrix's parameters (for EMLP), w/ and w/o DEF and bias initialization (BI) (for ECCC). + +
MethodValidation setParams
MeanMed.Trl.Best 25%Worst 25%Worst 5%Max
Grayworld* [13]5.543.134.110.9712.8319.4528.82-
Shades of Gray* [25]7.166.216.430.9915.0618.3919.21-
PCA* [14]6.134.535.020.9113.6618.2820.33-
Gamut (pixels) [29]7.537.387.032.0413.9618.3421.01636
Gamut (edges) [29]7.036.066.161.4914.5419.5121.98324
FCC[3]3.771.732.320.549.5616.0828.9912,288
Gray Index * [50]6.675.786.890.9815.0624.4528.82-
APAP-LUT [6]4.843.305.871.1210.8416.9227.53289
SIE[5]4.704.044.231.359.2113.2716.241,008,044
BoCF [41]4.843.305.871.1210.8416.9227.5359,354
C4 [63]3.913.263.241.068.5012.7617.355,115,657
CWCC [42]4.493.353.481.619.2113.3116.11100,830
C5* [4]4.012.813.141.088.5112.4721.81411,711
C5 (model A) [4]3.932.292.901.019.4013.3318.75171,511
C5 (model B) [4]4.492.843.240.8810.5215.2320.55213,831
C5 (model C) [4]4.252.673.190.9310.1214.5719.82213,831
C5 (model D) [4]4.512.943.380.8710.7416.0120.74213,831
EMLP [4]4.162.743.270.8810.7414.0521.2992,910,186
PCCC [64]4.614.194.151.758.5012.0617.86450
EMLP [c = 8, rg-chroma, HM, w/o cov]5.864.264.961.3612.316.3620.44-
EMLP [c = 8, rg-chroma, TM, w/o cov]4.783.944.271.149.5713.2016.54363
EMLP [c = 8, rgb, CM, w/o cov]4.543.533.871.269.5113.7419.22300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.813.984.221.299.3212.7514.99255
EMLP [c = 2, rg-chroma, CM, w/o cov]4.353.333.571.119.4413.6017.14300
EMLP [c = 4, rg-chroma, CM, w/o cov]4.163.563.771.008.3812.1616.79300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.073.513.641.237.7710.3813.29300
EMLP [c = 8, rg-chroma, CM, w/o cov]3.772.892.920.868.5911.4812.71354
EMLP [c = 8, n = 8, w/ DEF, both]3.642.792.860.868.6811.4213.54354
EMLP [c = 8, rg-chroma, CM, w/ aug, w/ cov]3.922.432.850.868.6811.3213.11354
ECCC [c = 8, n = 5, w/ DEF, both]3.692.783.110.797.8310.6212.372,166
ECCC [c = 8, n = 10, w/ DEF, both]3.662.692.920.958.0311.0612.323,496
ECCC [c = 8, n = 15, w/ DEF, both]3.582.712.910.917.9611.8513.514,826
ECCC [c = 2, n = 20, w/ DEF, both]3.912.993.231.078.3311.8415.236,156
ECCC [c = 4, n = 20, w/ DEF, both]4.283.193.360.999.6313.7214.726,156
ECCC [c = 8, n = 20, w/o DEF, both]4.023.233.691.229.1712.1414.474,608
ECCC [c = 8, n = 20, w/ BI, both]3.993.133.510.978.5311.4113.666,156
ECCC [c = 8, n = 20, w/ DEF, avg]4.032.973.341.038.6812.0914.715,900
ECCC [c = 8, n = 20, w/ DEF, short]3.712.742.931.008.0712.2015.435,900
ECCC [c = 8, n = 20, w/ DEF, long]3.702.742.931.008.0711.1115.435,900
ECCC [c = 8, n = 20, rg-chroma, w/ DEF, both]3.792.733.230.828.4111.6315.446,156
ECCC [c = 8, n = 20, w/ DEF, both, w/ aug]3.612.973.140.797.7710.5411.416,156
ECCC [c = 4, n = 20, w/ DEF, both]3.692.623.070.98.1210.9515.16,156
ECCC *c = 8, n = 20, w/ DEF, both)4.043.343.500.668.7511.2812.961,932
ECCC [c = 8, n = 20, w/ DEF, both]3.562.502.790.877.9311.3213.476,156
EMPLP + ECCC3.242.372.530.847.1110.6412.066,510
+ +spaces of input images before computing our DEF. Specifically, we used the $rg$ -chromaticity and raw RGB values (similar to [2, 21-23, 26]), in addition to our main design that uses $rgb$ -chromaticity. We also studied different mapping matrices, in addition to the $3 \times 3$ color matrix we used in Sec. 2. Specifically, we examined using the geometric $3 \times 4$ affine transformation matrix and the $3 \times 3$ homography matrix. + +We report results of our EMLP with and without the covariance matrix parameters in addition to the results of ECCC with different numbers of learnable biases, $n$ , and with and without the proposed bias initialization. Furthermore, we + +Table 2: Angular errors on the testing set. See caption of Table 1 for abbreviations. + +
MethodTesting set
MeanMed.Tri.Best 25%Worst 25%Worst 50%Max
Grayworld* [13]5.323.704.510.9511.3414.4117.75
Shades of Gray* [25]6.775.776.081.0714.2416.6417.50
PCA* [14]5.914.995.221.0712.6217.1019.08
Gamut [pixels] [29]7.497.257.372.3012.9616.6219.00
Gamut [edges] [29]5.904.804.931.1312.2017.6120.09
FFPC [9]3.181.832.280.397.9713.1320.58
Gray Index* [50]5.183.474.180.6012.3216.9019.67
APAP-LUT [6]3.942.823.320.818.5411.9016.36
SHE* [5]3.512.352.740.807.5611.0612.39
BoCf 441 [9]3.913.313.401.187.7211.2412.81
C4 [63]3.792.763.061.057.9311.3313.39
CWCC [42]3.712.853.061.117.7911.2112.83
CS* [4]3.452.933.030.957.0010.3814.01
CS (model A) [4]3.822.332.810.729.0815.6920.68
CS (model B) [4]3.302.052.420.697.5811.2314.02
CS (model C) [4]3.312.012.400.637.6712.0014.23
CS (model D) [4]3.332.202.420.647.6710.8313.35
TLCX [54]3.732.893.210.957.5411.5514.78
PCC [64]4.373.663.640.898.1415.1723.34
EMLP (ε = 2, w/ aug.)5.642.662.970.887.8711.518.82
EMLP (ε = 4, w/ aug.)3.502.632.780.927.5711.6213.42
EMLP+A (ε = 8, w/ aug.)3.372.342.740.677.2410.1614.10
EMLP (ε = 8, w/ aug.)3.362.732.840.767.039.5311.85
ECCC (ε = 8)3.332.922.920.966.569.3810.26
ECCC-ε2(ε = 8)3.412.873.030.816.879.1211.03
ECCC (ε = 2)2.942.102.380.696.318.139.25
ECCC (ε = 4)3.182.392.600.816.9910.0110.56
ECCC (ε = 8)3.002.312.450.746.669.8512.23
ECCC (ε = 8, w/ aug.)2.952.092.370.766.068.149.32
ELMP + ECCC2.912.332.440.726.118.389.83
+ +show the results of ECCC without utilizing the DEF feature. Lastly, we demonstrate the influence of different sizes of input sizes on the inference accuracy of both EMLP and ECCC. The results show that the best results are obtained with $e = 8$ . This choice is intuitively sensible, as a higher exposure factor increases the differences between the dual-exposure images, resulting in greater distinction based on scene lighting conditions. The augmentation is found to be useful for EMLP; however, we did not observe consistent improvement in the case of ECCC. This is likely because EMLP has a limited number of input features, making it more susceptible to overfitting. Thus, augmentation helps in generalization. Conversely, in the case of ECCC, the inclusion of histogram features alongside the small DEF feature suggests that the model may not consistently derive benefits from augmentation. We also noted that in ECCC, using raw RGB colors of dual-exposure images (without chromaticity conversion) tends to yield better results, while in EMLP, the chromaticity conversion tends to enhance performance (see Table 1). Results presented in Table 2 default to $rgb$ -chromaticity for EMLP and raw RGB values for ECCC. + +From the results presented in Tables 1 and 2, it is evident that our proposed DEF serves as a promising feature for guiding illuminant estimators. This is demonstrated by the performance of our lightweight EMLP model (354 parameters) compared to complex models that rely solely on raw RGB image colors (e.g., TLCC [54] with 32 million parameters). Moreover, incorporating DEF into ECCC reduces the maximum error by over $50\%$ and achieves comparable results across various evaluation metrics when compared to FFCC [9]. + +Since both models - namely, EMLP and ECCC - have a reasonable number of parameters, we can combine their predictions to create an ensemble model by averaging the predicted illuminant colors from both models. The results of this ensemble model are reported in Tables 1 and 2 under 'ECCC + EMLP'. The combined efforts of EMLP and ECCC, guided by our DEF, demonstrate promising results, outperforming several state-of-the-art approaches across vari- + +![](images/84a7ef8cc5e191d03786ac653100b37e086159e16609c229ed2a4a065f06f630.jpg) +Fig. 5: Randomly selected examples from our worst $25\%$ results of ECCC (top two rows are from validation set and remaining rows are from testing set). (A) Input pair of raw images captured with long and short exposures (note that other methods use a single image captured with auto exposure). (B-G) Images corrected with the estimated illuminant by: (B) FFCC [9], (C) C5 [4] (chosen the best results among all variations discussed in Sec. 4.3), (D) TLCC [54], (E) EMLP, (F) ECCC, and (G) Ground-truth illuminant. The estimated illuminant of each method is shown on the right side of each image, along with the angular error written in the top-left corner of the image. + +ous evaluation metrics while utilizing considerably fewer parameters. Qualitative examples of our results, randomly selected from the worst $25\%$ of ECCC results, are presented in Fig. 5 alongside results from other methods. + +# 5 Conclusion and Future Work + +In this paper, we introduced DEF, a feature derived from dual exposure images to enhance illuminant estimation. Our DEF achieves comparable results with state-of-the-art methods that employ thousands or millions of parameters (e.g., [54, 63]), using only 354 parameters in a straightforward MLP network. We further discuss incorporating the proposed DEF into the established CCC framework, referred to as ECCC. ECCC achieves comparable or better results on different evaluation metrics than classic CCC approaches while remaining lightweight, requiring only 6,156 parameters (50% reduction compared to FFCC [9]), approximately 30 KB of memory, and running in $\sim 0.25$ milliseconds per image on CPU. Our solution focuses on the single-camera case, intending testing on the same camera used for training. Future work includes studying the stability of this feature across different cameras. We discussed integrating the proposed DEF into the CCC framework. Further exploration may involve examining DEF's benefits for other illuminant estimation techniques that rely on convNets and raw image pixels as input. Another research direction could involve developing a spatially varying version of DEF, rather than our global feature, to utilize for spatially varying illuminant estimation and image white balancing. + +# References + +1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018) +2. Abdelhamed, A., Punnappurath, A., Brown, M.S.: Leveraging the availability of two cameras for illuminant estimation. In: CVPR (2021) +3. Affi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021) +4. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: ICCV (2021) +5. Affifi, M., Brown, M.S.: Sensor-independent illumination estimation for DNN models. In: BMVC (2019) +6. Affi, M., Punnappurath, A., Finlayson, G., Brown, M.S.: As-projective-as-possible bias correction for illumination estimation algorithms. Journal of the Optical Society of America A 36(1), 71-78 (2019) +7. Banić, N., Koščević, K., Lončarić, S.: Unsupervised learning for color constancy. arXiv preprint arXiv:1712.00436 (2017) +8. Barron, J.T.: Convolutional color constancy. In: ICCV (2015) +9. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017) +0. Bianco, S., Schettini, R.: Color constancy using faces. In: CVPR (2012) +1. Brainard, D.H., Freeman, W.T.: Bayesian color constancy. Journal of the Optical Society of America A 14(7), 1393-1411 (1997) +2. Brown, M.: Color processing for digital cameras, chap. 5, pp. 81-98. John Wiley & Sons, Ltd (2023) +3. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin Institute 310(1), 1-26 (1980) +4. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. Journal of the Optical Society of America A 31(5), 1049-1058 (2014) +5. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: Annual Conference on Computer Graphics and Interactive Techniques (1997) +6. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021) +7. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Transactions on Graphics 36(6) (2017) +8. Fairchild, M.D.: Color appearance models. John Wiley & Sons (2013) +9. Finlayson, G.D., Funt, B.V., Barnard, K.: Color constancy under varying illumination. In: ICCV (1995) +20. Finlayson, G.D., Hordley, S.D.: Color constancy at a pixel. Journal of the Optical Society of America A 18(2), 253-264 (2001) +21. Finlayson, G.D., Hordley, S.D.: The chromagenic colour camera and illuminant estimation. In: Color Imaging Conference (2005) +22. Finlayson, G.D., Hordley, S.D., Morovic, P.: Chromagenic colour constancy. In: Congress of the International Colour Association (2005) +23. Finlayson, G.D., Hordley, S.D., Morovic, P.: Colour constancy using the chromogenic constraint. In: CVPR (2005) +24. Finlayson, G.D., Hordley, S.D., Tastl, I.: Gamut constrained illuminant estimation. International Journal of Computer Vision 67(1), 93-109 (2006) +25. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference (2004) + +26. Fredembach, C., Finlayson, G.: The bright-chromagenic algorithm for illuminant estimation. Journal of Imaging Science and Technology 52, 40906:1-40906:11 (2008) +27. Gehler, P.V., Rother, C., Blake, A., Minka, T., Sharp, T.: Bayesian color constancy revisited. In: CVPR (2008) +28. Gelfand, N., Adams, A., Park, S.H., Pulli, K.: Multi-exposure imaging on mobile devices. In: ACM International Conference on Multimedia (2010) +29. Gijsenij, A., Gevers, T., Van De Weijer, J.: Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision 86(2), 127-139 (2010) +30. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing 20(9), 2475-2489 (2011) +31. Gijsenij, A., Gevers, T., Van De Weijer, J.: Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(5), 918-929 (2011) +32. Granados, M., Ajdin, B., Wand, M., Theobalt, C., Seidel, H.P., Lensch, H.P.: Optimal HDR reconstruction with linear digital cameras. In: CVPR (2010) +33. Hanji, P., Mantiuk, R.K., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: SIGGRAPH (2022) +34. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A K-means clustering algorithm. Journal of the royal statistical society. Series C (applied statistics) 28(1), 100-108 (1979) +35. Hernandez-Juarez, D., Parisot, S., Busam, B., Leonardis, A., Slabaugh, G., McDonagh, S.: A multi-hypothesis approach to color constancy. In: CVPR (2020) +36. Hu, Y., Wang, B., Lin, S.: FC4: Fully convolutional color constancy with confidence-weighted pooling. In: CVPR (2017) +37. Hubel, P.M., Finlayson, G.D., Hordley, S.D.: White point estimation using color by convolution (2007), US Patent 7,200,264 +38. Jung, C., Yang, Y., Jiao, L.: High dynamic range imaging on mobile devices using fusion of multiexposure images. Optical Engineering 52(10), 102004-102004 (2013) +39. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: ECCV (2016) +40. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +41. Laakom, F., Passalis, N., Raitoharju, J., Nikkanen, J., Tefas, A., Iosifidis, A., Gabbouj, M.: Bag of color features for color constancy. IEEE Transactions on Image Processing 29, 7722-7734 (2020) +42. Laakom, F., Raitoharju, J., Nikkanen, J., Iosifidis, A., Gabbouj, M.: Robust channel-wise illumination estimation. In: BMVC (2021) +43. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: WACV (2023) +44. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., et al.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 164-1 (2019) +45. Lo, Y.C., Chang, C.C., Chiu, H.C., Huang, Y.H., Chen, C.P., Chang, Y.L., Jou, K.: CLCC: Contrastive learning for color constancy. In: CVPR (2021) +46. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +47. Lou, Z., Gevers, T., Hu, N., Lucassen, M.P., et al.: Color constancy by deep learning. In: BMVC (2015) + +48. Oh, S.W., Kim, S.J.: Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition 61, 405-416 (2017) +49. Qi, G., Chang, L., Luo, Y., Chen, Y., Zhu, Z., Wang, S.: A precise multi-exposure image fusion method based on low-level features. Sensors 20(6), 1597 (2020) +50. Qian, Y., Kamarainen, J.K., Nikkanen, J., Matas, J.: On finding gray pixels. In: CVPR (2019) +51. Shen, R., Cheng, I., Shi, J., Basu, A.: Generalized random walks for fusion of multi-exposure images. IEEE Transactions on Image Processing 20(12), 3634-3646 (2011) +52. Solhusvik, J., Hu, S., Johansson, R., Lin, Z., Ma, S., Mabuchi, K., Manabe, S., Mao, D., Phan, B., Rhodes, H., et al.: A 1392x976 2.8 $\mu$ m 120db CIS with per-pixel controlled conversion gain. In: International Image Sensor Workshop (2017) +53. Solomatov, G., Akkaynak, D.: Spectral sensitivity estimation without a camera. In: ICCP (2023) +54. Tang, Y., Kang, X., Li, C., Lin, Z., Ming, A.: Transfer learning for color constancy via statistic perspective. In: AAAI (2022) +55. Ulucan, O., Karakaya, D., Turkan, M.: Multi-exposure image fusion based on linear embeddings and watershed masking. Signal Processing 178, 107791 (2021) +56. Ulucan, O., Ulucan, D., Ebner, M.: Block-based color constancy: The deviation of salient pixels. In: International Conference on Acoustics, Speech and Signal Processing (2023) +57. Van De Weijer, J., Gevers, T., Gijsenij, A.: Edge-based color constancy. IEEE Transactions on Image Processing 16(9), 2207-2214 (2007) +58. Willassen, T., Solhusvik, J., Johansson, R., Yaghmai, S., Rhodes, H., Manabe, S., Mao, D., Lin, Z., Yang, D., Cellek, O., et al.: A $1280 \times 1080$ 4.2 $\mu$ m split-diode pixel HDR sensor in $110\mathrm{nm}$ BSI CMOS process. In: International Image Sensor Workshop (2015) +59. Xiao, J., Gu, S., Zhang, L.: Multi-domain learning for accurate and few-shot color constancy. In: CVPR (2020) +60. Xing, X., Qian, Y., Feng, S., Dong, Y., Matas, J.: Point cloud color constancy. In: CVPR (2022) +61. Xu, B., Wang, N., Chen, T., Li, M.: Empirical evaluation of rectified activations in convolutional network. arXiv preprint arXiv:1505.00853 (2015) +62. Yahiaoui, L., Horgan, J., Deegan, B., Yogamani, S., Hughes, C., Denny, P.: Overview and empirical analysis of ISP parameter tuning for visual perception in autonomous driving. Journal of Imaging 5(10), 78 (2019) +63. Yu, H., Chen, K., Wang, K., Qian, Y., Zhang, Z., Jia, K.: Cascading convolutional color constancy. In: AAAI (2020) +64. Yue, S., Wei, M.: Color constancy from a pure color view. Journal of the Optical Society of America A 40(3), 602-610 (2023) \ No newline at end of file diff --git a/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/images.zip b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b5f08db7e4d9c5e29587e1c38a88efc38e8bba9 --- /dev/null +++ b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:334c7863d87ed424a4caa1e7812535d85000242ea9f1e501411c68ca07972165 +size 505920 diff --git a/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/layout.json b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1ea8ea7ab873a2e6eee1e374c23ddcf744b93ba --- /dev/null +++ b/2024/Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging/layout.json @@ -0,0 +1,10255 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 179, + 111, + 435, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 111, + 435, + 148 + ], + "spans": [ + { + "bbox": [ + 179, + 111, + 435, + 148 + ], + "type": "text", + "content": "Optimizing Illuminant Estimation in Dual-Exposure HDR Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 197, + 168, + 416, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 168, + 416, + 181 + ], + "spans": [ + { + "bbox": [ + 197, + 168, + 416, + 181 + ], + "type": "text", + "content": "Mahmoud Affifi, Zhenhua Hu, and Liang Liang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 190, + 405, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 190, + 405, + 213 + ], + "spans": [ + { + "bbox": [ + 209, + 190, + 405, + 213 + ], + "type": "text", + "content": "Google {mafifi, zhenhuahu, annyliang}@google.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 239, + 452, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 452, + 481 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 452, + 481 + ], + "type": "text", + "content": "Abstract. High dynamic range (HDR) imaging involves capturing a series of frames of the same scene, each with different exposure settings, to broaden the dynamic range of light. This can be achieved through burst capturing or using staggered HDR sensors that capture long and short exposures simultaneously in the camera image signal processor (ISP). Within camera ISP pipeline, illuminant estimation is a crucial step aiming to estimate the color of the global illuminant in the scene. This estimation is used in camera ISP white-balance module to remove undesirable color cast in the final image. Despite the multiple frames captured in the HDR pipeline, conventional illuminant estimation methods often rely only on a single frame of the scene. In this paper, we explore leveraging information from frames captured with different exposure times. Specifically, we introduce a simple feature extracted from dual-exposure images to guide illuminant estimators, referred to as the dual-exposure feature (DEF). To validate the efficiency of DEF, we employed two illuminant estimators using the proposed DEF: 1) a multilayer perceptron network (MLP), referred to as exposure-based MLP (EMLP), and 2) a modified version of the convolutional color constancy (CCC) to integrate our DEF, that we call ECCC. Both EMLP and ECCC achieve promising results, in some cases surpassing prior methods that require hundreds of thousands or millions of parameters, with only a few hundred parameters for EMLP and a few thousand parameters for ECCC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "content": "Keywords: Computational color constancy " + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "content": " Illuminant estimation " + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 491, + 453, + 514 + ], + "type": "text", + "content": " HDR imaging" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 533, + 342, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 342, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 342, + 547 + ], + "type": "text", + "content": "1 Introduction and Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Camera image signal processor (ISP) comprises several modules, each dedicated to enhancing specific aspects of the quality of captured raw images by the camera sensor [16]. One key component among these modules is the white-balance module, which aims to eliminate undesirable color casts introduced by the combination of scene lighting and camera sensitivity. To achieve this, the auto white-balance module runs an illuminant estimator onboard the camera ISP to estimate the RGB color of the illuminant, under the assumption that a single global illuminant illuminates the scene for simplicity [30, 44]. Image white balancing can thus be described as follows:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 246, + 178 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 246, + 178 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 246, + 178 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 246, + 178 + ], + "type": "image", + "image_path": "624c42bbf13ed50c54d147571f7754ba36c46b57c60088198a8fc60090b33093.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 178, + 231, + 185 + ], + "lines": [ + { + "bbox": [ + 176, + 178, + 231, + 185 + ], + "spans": [ + { + "bbox": [ + 176, + 178, + 231, + 185 + ], + "type": "text", + "content": "Full-resolution raw image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 178, + 288, + 185 + ], + "lines": [ + { + "bbox": [ + 246, + 178, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 246, + 178, + 288, + 185 + ], + "type": "text", + "content": "Raw frames with different exposure times" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 201, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 334 + ], + "type": "text", + "content": "Fig. 1: Conventional illuminant estimators often rely on a single frame for illuminant color estimation. Although the HDR camera pipeline includes at least two frames of the same scene, conventional methods usually consider only a single frame (either with auto exposure, shown in the dashed green line, merged frame, or one frame of the burst capture) [12,16,44,62]. This paper proposes leveraging information from dual-exposure capturing in multi-exposure HDR imaging to enhance illuminant estimation in camera pipelines. Our method uses frames captured at long and short exposures, available in multi-exposure bursts [28] or staggered HDR sensors [52, 58]. Achieving comparable or superior results, our method employs lightweight models (" + }, + { + "bbox": [ + 130, + 201, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 201, + 482, + 334 + ], + "type": "text", + "content": "300-6000 parameters) compared to those using hundreds of thousands or millions of parameters. In this figure and the following figures, all raw images have the gamma operator applied to aid visualization, and all sRGB images are rendered using the HDR+ camera pipeline [44]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 114, + 288, + 178 + ], + "blocks": [ + { + "bbox": [ + 246, + 114, + 288, + 178 + ], + "lines": [ + { + "bbox": [ + 246, + 114, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 246, + 114, + 288, + 178 + ], + "type": "image", + "image_path": "adfd9146708b0ec3c480ce71a33c61177aee412bcc5386c04114f28be22b6875.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 178, + 348, + 185 + ], + "lines": [ + { + "bbox": [ + 288, + 178, + 348, + 185 + ], + "spans": [ + { + "bbox": [ + 288, + 178, + 348, + 185 + ], + "type": "text", + "content": "Dual-exposure frames with long exposure times" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 288, + 114, + 359, + 178 + ], + "blocks": [ + { + "bbox": [ + 288, + 114, + 359, + 178 + ], + "lines": [ + { + "bbox": [ + 288, + 114, + 359, + 178 + ], + "spans": [ + { + "bbox": [ + 288, + 114, + 359, + 178 + ], + "type": "image", + "image_path": "6fb6fabfbc74ff841dde9dddc176556fbfcf223f5f5d2cc28aaeb2491330dca5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 452, + 178 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 452, + 178 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 452, + 178 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 452, + 178 + ], + "type": "image", + "image_path": "1c614014ad931b16fbcffe465d0a06c1479ebb18cc9a2ccf855600573a4b5d44.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 178, + 449, + 191 + ], + "lines": [ + { + "bbox": [ + 369, + 178, + 449, + 191 + ], + "spans": [ + { + "bbox": [ + 369, + 178, + 449, + 191 + ], + "type": "text", + "content": "Final rendered image with our white-balance correction" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 243, + 369, + 481, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 243, + 369, + 481, + 402 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {W B}} = \\operatorname {d i a g} \\left(\\left[ \\frac {\\hat {\\ell} _ {G}}{\\hat {\\ell} _ {R}}, 1, \\frac {\\hat {\\ell} _ {G}}{\\hat {\\ell} _ {B}} \\right] ^ {T}\\right) I, \\tag {1}", + "image_path": "b141044f401637da4a6b19e96555e634c1d9ed42b40448d10158ed16deea196f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 261, + 414, + 480, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 414, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 261, + 414, + 480, + 430 + ], + "type": "interline_equation", + "content": "\\left[ \\hat {\\ell} _ {R}, \\hat {\\ell} _ {G}, \\hat {\\ell} _ {B} \\right] ^ {T} = f (I), \\tag {2}", + "image_path": "093b68ebef7cd94803a07a9eff48004266ecef1384e65419cd50ae58efe8a5f5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{WB}}" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "3 \\times k" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " RGB colors of raw image and white-balanced raw image, respectively, with " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " refers to the total number of pixels in the image, " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\operatorname{diag}(.)" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " creates a " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " diagonal matrix from a given vector, " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "\\hat{\\ell} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " refers to the estimated illuminant color vector, " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " refers to vector transpose, and " + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": " is an illuminant estimator function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Modern camera ISPs capture several frames of each scene with different exposure times to enhance final image dynamic range [28,38]. This can be achieved through burst capturing, involving rapid capture with varying exposure times in quick succession, or by using staggered high dynamic range (sHDR) sensors that simultaneously capture long and short exposures [52, 58]. The captured images are then combined, incorporating multiple exposures of the same scene at different levels, resulting in a greater dynamic range than what would be possible with a single image [43, 49, 51, 55]. This HDR imaging camera pipeline, therefore, includes more than a single frame of the captured scene, and internal camera ISP modules, such as the auto white-balance module, can access these additional frames. While such additional frames may have beneficial information to help illuminant estimators, conventional illuminant estimation methods often rely on a single frame (e.g., [9, 36, 41, 47, 54, 63, 64]) in both traditional single-frame camera pipelines [39] and multi-frame HDR ISPs [44]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "Such single-frame illuminant estimators can be categorized into: 1) statistical methods that rely on statistics computed from input raw image information (e.g., colors, edges) [13, 14, 25, 31, 50, 56, 57] and 2) learning-based methods that learn, from a set of training images labeled with ground-truth illuminant colors, to map from input raw color information to the corresponding illuminant color [6, 8, 10, 11, 24, 36, 42, 45, 47, 48, 54, 63, 64]. While the latter category typically achieves better results than the statistical methods, most learning-based methods are camera-dependent, meaning that they require domain adaptation or fine-tuning when deployed on new cameras to achieve similar accuracy on cameras used for training due to the influence of the camera response function on both raw image colors and ground-truth colors [3-5, 35, 53, 59]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "type": "text", + "content": "A limited number of attempts have proposed learning-based methods that go beyond the single-frame input scheme. For example, the cross-camera convolutional color constancy (C5) [4] suggests utilizing additional unlabeled images captured by the testing camera, in addition to the primary single-frame image of the scene being captured to improve the generalization for cameras that were not included in the training phase. Xing et al. [60] proposed the use of a depth map, captured by a time-of-flight (ToF) sensor, along with the primary raw image to predict illuminant color in the scene by leveraging the geometry information obtained from depth map. Abdelhamed et al., [2] proposed leveraging the presence of two cameras in most modern mobile phone devices. Assuming dual streaming from both cameras, they derived a feature within the framework of chromagenic color constancy theory [21-23] to enhance the accuracy of illuminant estimation, yielding promising results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "text", + "content": "Our method, in contrast to the majority of prior work, adopts the strategy of benefiting from multiple frames available in the camera ISP (similar to [2]). However, unlike [2, 60], which requires streaming from dual cameras and may lead to impractical high power consumption, our method relies on two frames captured of the same scene under different exposure settings, already present in the HDR imaging pipeline, to estimate the illuminant color of the captured scene (see Fig. 1). Specifically, we utilize: 1) a frame captured with short exposure time, that we refer to as short-exposure image " + }, + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "inline_equation", + "content": "(I_s)" + }, + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "text", + "content": ", and 2) another frame of the same scene captured with long exposure time, that we refer to as long-exposure image " + }, + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "inline_equation", + "content": "(I_l)" + }, + { + "bbox": [ + 130, + 403, + 482, + 547 + ], + "type": "text", + "content": ". Having dual-exposure images in the HDR imaging pipeline is feasible, making our method practical for most HDR imaging pipelines. Accordingly, Eq. 2 can be modified to include our dual-exposure input images as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 556, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 556, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 255, + 556, + 480, + 572 + ], + "type": "interline_equation", + "content": "[ \\hat {\\ell} _ {R}, \\hat {\\ell} _ {G}, \\hat {\\ell} _ {B} ] ^ {T} = f (I _ {l}, I _ {s}). \\tag {3}", + "image_path": "84aae96246d9df4b6eda9a70eea76beba8657d213326ade1d474840a330bee1b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "text", + "content": "Our objective in this work is to design the function " + }, + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "text", + "content": " in a manner that enables the effective utilization of the additional information provided by the dual-exposure images, " + }, + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 576, + 482, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Contribution In this paper, we present a feature inspired by the chromagenic color constancy theory [21-23], termed the dual-exposure feature (DEF), that is derived from images captured with both short exposure time " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "(I_s)" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and long" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": "exposure time " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "(I_l)" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": ". DEF leverages the variations in chromatic information between these dual-exposure images, providing valuable guidance for illuminant estimator methods. To assess its effectiveness, we trained a lightweight multi-layer perceptron network (MLP) for illuminant estimation that utilizes solely our DEF as input, departing from the conventional approach of using actual RGB values of the captured scene colors. Additionally, we explored the integration of DEF into an established color constancy framework, specifically the convolutional color constancy (CCC) [4,8,9,37], which we refer to as exposure-based CCC (ECCC). The experimental results on a multi-exposure dataset, collected to evaluate our work empirically, show that these models - namely, EMLP and ECC - achieve promising results with a reasonable number of parameters - 354 learnable parameters for EMLP and 6,156 learnable parameters for ECC. This outperformance across diverse evaluation metrics is observed when comparing to prior methods that require significantly higher number of parameters, ranging from tens to hundreds of thousands, or even millions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 313, + 402, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 402, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 402, + 327 + ], + "type": "text", + "content": "2 Illuminant-Linked Dual Exposure Feature" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "content": "To develop an efficient illuminant estimator that benefits from both " + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "content": ", we introduce a compact feature that aims to capture the correlation between these dual-exposure images. Our feature is inspired by the chromagenic color constancy theory [21-23]. Specifically, we explore the analogy between long and short exposure images, " + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "content": ", and aligned images captured by two cameras [2] under the chromagenic color constancy theory, leading us to develop our dual-exposure feature (DEF). To begin, we review the chromagenic color constancy theory under the Lambertian reflectance model with a single illuminant assumption. The captured raw image " + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "content": " can mathematically be described by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 456, + 481, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 456, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 231, + 456, + 481, + 482 + ], + "type": "interline_equation", + "content": "I _ {\\rho} ^ {(y)} = \\int_ {\\gamma} S _ {\\rho} (x) D (x) R (y, x) d x + z, \\tag {4}", + "image_path": "4ab45c0cae23bbf2b54276c613420ff68dd4f85cd62d5619f409ade053693ea0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "S(\\cdot), D(\\cdot)" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": " represent the camera response function (typically represented by camera sensitivity, infrared cut-off filter, and spectral lens transmission), the spectral power distribution of light, and scene reflectance, respectively. Here, " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": " refers to a wavelength within the visible range " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": " refers to the pixel location in image " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "\\rho \\in \\{R,G,B\\}" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": " refers to the color channel, and " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": " denotes the undesired noise, typically represented by signal-dependent and additive components [1]. According to the chromagenic color constancy theory [22], if we capture two images of the same scene with a specially chosen chromagenic filter, " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": ", applied between image captures, the linear color matrix that maps between those captured images is unique to the illuminant color present in this scene. That is, given an image, " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": ", that is captured by the main camera, and a filtered image, " + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 130, + 489, + 482, + 633 + ], + "type": "text", + "content": ", that can be described as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 218, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 218, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "I _ {f _ {\\rho}} ^ {(y)} = \\int_ {\\gamma} S _ {\\rho} (x) Q _ {\\rho} (x) D (x) R (y, x) d x + z, \\tag {5}", + "image_path": "c2565e515bc3d7d0a7937428874ffe057a814e2c95efa967c9990eef5c38b7a5.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " color matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " that maps between the colors of " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is indicative of the scene illuminant. This color matrix can be computed by minimizing the following equation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 260, + 167, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 167, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 260, + 167, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\underset {C _ {c}} {\\arg \\min } \\| C _ {c} I _ {f} - I \\| _ {F}, \\tag {6}", + "image_path": "ffd8fd00ef82b4f78f9959c5e0b5cd4ae65daa5d069bc6907a0d006dc9b0178a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F" + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "content": " is the Frobenius norm. The closed-form solution for Eq. 6 can be obtained using the pseudoinverse (i.e., " + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "inline_equation", + "content": "C_c = II_f^\\dagger" + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "content": "). While theoretically validated, the chromagenic filter conditions required to obtain a unique mapping matrix per illuminant are challenging to meet in practice. Empirically, Finlayson et al., [23] found that most filters exhibit a reasonable level of correlation between the computed matrix and the illuminant color, excluding neutral filters, which consistently result in a scaling relationship between the unfiltered image, " + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "content": ", and the filtered image, " + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "content": ". Relaxing the conditions to include normal colored filter, Abdelhamed et al., [2] proposed using two cameras, with the second camera serving as the main camera after applying a colored filter—i.e., " + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "inline_equation", + "content": "S(\\cdot)Q(\\cdot)" + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "content": ". Surprisingly, even when chromagenic filter conditions are not met, using a normal colored filter [23] or another camera with a different response function [2] still shows a correlation between the computed matrix " + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 130, + 195, + 482, + 365 + ], + "type": "text", + "content": " and the illuminant color to some extent, achieving promising results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "text", + "content": "We argue that this correlation arises due to the variations, or what can be considered a form of \"distortion\", in the colors captured by the second (filtered) camera when compared to the original colors captured by the main camera. This color distortion varies based on the interplay between the scene irradiance, " + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "inline_equation", + "content": "D(\\cdot)R(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "text", + "content": ", and the camera response functions - namely, the main camera's " + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "text", + "content": " and the filtered/second camera's " + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "inline_equation", + "content": "S(\\cdot)Q(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "text", + "content": " in Eqs. 4 and 5. In the context of machine learning (ML) models, the camera response functions remain fixed across the entire dataset. Consequently, ML models, such as the one proposed in [2], learn the correlation between the ground-truth illuminant color and the scene irradiance through the matrix " + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 130, + 365, + 482, + 498 + ], + "type": "text", + "content": " that represents the level of \"distortion\" between the colors of captured scene images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In a dual-exposure setup, we have a long-exposure image, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", and a short-exposure image, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", both capturing the same scene. The extent of color distortion in each image varies based on the scene irradiance; for instance, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " may exhibit a higher level of color distortion and noise than " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " under suboptimal lighting conditions (e.g., indoor light), while " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " may have a higher level of color distortion and over-saturated colors than " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in well-lit scenes (e.g., outdoor light). This difference in color distortion arises because the non-linear camera response function receives a different number of photons to form each image colors based on the exposure time and scene irradiance [32]. As a result, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " can exhibit color variations that differ across individual color channels [15], that are somewhat akin to those caused by a color filter (though to a lesser extent). Even within the same color channel, different levels of color differences between " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " can be observed spatially due to the interaction between the non-linear camera response function of that channel and the spatially varying scene irradiance (see" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 455, + 241 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 455, + 241 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 455, + 241 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 455, + 241 + ], + "type": "image", + "image_path": "8a50bb5a388c976a7d741d29a232d1e3442348dcc5b03e515c91b834bfc5a225.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 248, + 482, + 348 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 348 + ], + "type": "text", + "content": "Fig. 2: Cameras perceive different amounts of photons when capturing the same scene with different exposure times. Images taken with both long and short exposure times exhibit variations in each channel due to the camera response function and scene irradiance. Additionally, spatial variations, in each color channel, can be observed based on object reflectance, as the interplay of lighting, object reflectance, and the camera response function leads to different outcomes. (A) and (B) show raw images of scenes captured under indoor and outdoor lighting, respectively. In (C) and (D), we present the average " + }, + { + "bbox": [ + 130, + 248, + 482, + 348 + ], + "type": "inline_equation", + "content": "rg" + }, + { + "bbox": [ + 130, + 248, + 482, + 348 + ], + "type": "text", + "content": "-chromaticity histogram and aggregated red, green, and blue pixel values from 25 images sharing similar lighting conditions in (A) and (B), respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "content": "Fig. 2). In Fig. 2, we show two examples of dual-exposure images captured under different lighting conditions. As can be seen, the differences between the colors in " + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "content": " under each lighting condition exhibit variations, which are observable in the histogram similarity (here we use the Bhattacharyya distance similarity metric) and variations of the ratios in each color channel between " + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "content": ". That is, the correlation between " + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 373, + 482, + 479 + ], + "type": "text", + "content": " is not always a proportional scaling, as in the case of a neutral filter. We observed similar patterns as shown in Fig. 2 when studying examples from the two-camera dataset in [2] (see supplemental materials)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": "Based on this discussion, we propose to compute the color matrix " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " to map between the " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "rgb" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": "-chromaticity values (i.e., " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "[R / \\kappa, G / \\kappa, B / \\kappa]^T" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\kappa = R + G + B" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": ") of " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": ". The reason for not using the RGB triples, similar to chromagenic color constancy, is that we aim to reduce the influence of intensity differences between " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " when computing " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": ". In addition, we compute the covariance matrix, " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "C_v" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": ", of the ratio image " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{3 \\times k}" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "X_{\\rho}^{(y)} = I_{s_{\\rho}}^{(y)} / \\left(I_{l_{\\rho}}^{(y)} + \\epsilon\\right)" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " is a small number added for numerical stability. Computing " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "C_v" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " is performed as described in Eq. 7 to measure the variance in each color channel between " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 130, + 481, + 482, + 594 + ], + "type": "text", + "content": " and the joint variability across channels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 606, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 606, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 219, + 606, + 481, + 622 + ], + "type": "interline_equation", + "content": "C _ {v} (X) = \\mathbb {E} \\left[ (X - \\mathbb {E} [ X ]) (X - \\mathbb {E} [ X ]) ^ {T} \\right], \\tag {7}", + "image_path": "4a032db345f28cc42157dd9a4c9e6a04dc7d9c488392284c94ed9372a97e48d0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{E}(\\cdot)" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is the expected value (mean) of the matrix. Both " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_v" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " form our dual-exposure feature (DEF) that represents the differences in color distortion in " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ". Our DEF is represented as a vector " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{15}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", where we exclude" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 113, + 455, + 235 + ], + "blocks": [ + { + "bbox": [ + 161, + 113, + 455, + 235 + ], + "lines": [ + { + "bbox": [ + 161, + 113, + 455, + 235 + ], + "spans": [ + { + "bbox": [ + 161, + 113, + 455, + 235 + ], + "type": "image", + "image_path": "4b3d06d705088c112421a9f1450023c45d66abf1a10489659fced4a1da18ee52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 319 + ], + "type": "text", + "content": "Fig. 3: We present an illuminant-related dual-exposure feature (DEF), derived from a pair of images captured with short and long exposures. Using DEF, we deploy a simple multilayer perceptron network (MLP) with only 354 parameters, referred to as exposure-based MLP or EMLP, for illuminant estimation, as shown in (A). We further explore the integration of DEF into the CCC framework, as shown in (B), by dynamically generating bias map based on DEF. We denote this modified CCC framework as exposure-based CCC or ECCC." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 344, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 477 + ], + "type": "text", + "content": "the redundant values in " + }, + { + "bbox": [ + 130, + 344, + 482, + 477 + ], + "type": "inline_equation", + "content": "C_v" + }, + { + "bbox": [ + 130, + 344, + 482, + 477 + ], + "type": "text", + "content": " over the symmetric positions. To evaluate DEF's effectiveness as a clue for scene illuminant colors, we employed a lightweight MLP for scene illuminant estimation (Fig. 3-A), that we call exposure-based MLP (EMLP). EMLP relies solely on our DEF as input. It comprises an input fully connected (fc) layer with 9 output neurons, followed by leaky ReLU (LReLU) [61], two hidden fc layers, each with 9 output neurons with LReLU in between, and an output fc layer with three neurons. EMLP achieves results comparable to complex models with thousands or millions of parameters (refer to Tables 1 and 2), while maintaining a lightweight design with just a few hundred parameters. Consequently, DEF is empirically shown as a valuable feature providing strong insights into the illuminant scene color." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 495, + 440, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 440, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 440, + 510 + ], + "type": "text", + "content": "3 Integration with Convolutional Color Constancy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "In this section, we incorporate the proposed DEF into one of the most established frameworks for illuminant estimation. Specifically, we introduce modifications to the convolutional color constancy (CCC) framework [4, 8, 9, 37] to leverage the benefits of the DEF, which we refer to as exposure-based CCC or ECCC for short. It is important to note that ECCC does not introduce a new CCC method; instead, it serves as an illustration of how the DEF can be seamlessly integrated into existing, well-established illuminant estimation frameworks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Let's begin with a brief description of the CCC [8, 9]. Given a single raw image, the CCC operates by learning one or more convolutional 2D filters, denoted as " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\{F_j\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", which convolve over the 2D histogram(s), " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\{H_j\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", of the image colors in the " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " color space (i.e., the log of " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "G / R" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "G / B" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " chroma values of pixel colors) [8, 20]. This convolutional operation can be accelerated by FFTs as" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 100 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "proposed in [9]. Afterwards, a 2D bias map, " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", is added, followed by a softmax operation, " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", to compute the \"probability\" map, " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", of the illuminant bin in the uv 2D histogram space. This simplified version of the CCC can be described as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 175, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 175, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 244, + 175, + 481, + 213 + ], + "type": "interline_equation", + "content": "P = \\sigma \\left(\\sum_ {j} \\left(F _ {j} * H _ {j}\\right) + B\\right). \\tag {8}", + "image_path": "695309f0dbfc01790cdc0f18030abe22ae5601f9168585aa675cd52ad254c52d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": "In CCC [8,9], the filters and bias are learned across the entire training dataset. This means it consists of a single bias and one or more filters (the number depends on the histograms used, either a single histogram for image colors or two histograms, including the edge color histogram). Later, C5 [4] proposes a hypernetwork that dynamically generates filters and bias based on the input raw image and additional images taken from the same camera, aiming to improve the generalization of CCC across cameras. While cross-camera color constancy is out of the scope of this paper, we propose to use our DEF to dynamically \"generate\" a bias map based on the input image. That is, we learn a bank of biases, " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\{B_i\\}" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, n\\}" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ", such that we linearly interpolate between them based on blending weights emitted from an MLP network that processes our DEF. In this way, the DEF controls the bias of the CCC model, acting as \"candidate\" illuminant priors within the uv space for the input image (see Fig. 3-B). To implement this, we need to have multiple bias maps, which definitely will lead to an impractical increase in model size. Thus, we propose to use a downsampled size " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "(1/4" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": " of histogram size) for the learnable biases, " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\{B_i\\}" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ". Furthermore, we propose to feed two histograms (i.e., " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "j \\in \\{l, s\\}" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ") of both " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "H_l" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "H_s" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ", respectively, into the model. This leads to the learning of two downsampled filters, " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\{F_j\\}" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ", corresponding to the histograms of the long-exposure image " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "(I_l)" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": " and the short-exposure image " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "(I_s)" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": ". Learning small-sized filters and biases has the following benefits. First, we can learn many biases within an affordable model size (e.g., " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "n = 20" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": " requires " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\sim 6\\mathrm{K}" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": " parameters, while FFCC [9] with a single bias requires " + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\sim 12\\mathrm{K}" + }, + { + "bbox": [ + 130, + 220, + 482, + 518 + ], + "type": "text", + "content": " parameters). Second, it implicitly produces smooth filters and biases, which are desirable to avoid overfitting [4,9]. With this modification, Eq. 8 can be rewritten as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 531, + 481, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 531, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 231, + 531, + 481, + 569 + ], + "type": "interline_equation", + "content": "P = \\sigma \\left(\\sum_ {j} \\left(\\uparrow \\left(F _ {j}\\right) * H _ {j}\\right) + B _ {\\mathrm {u p}} ^ {\\prime}\\right), \\tag {9}", + "image_path": "5b7336183077f8f5cc180feba3463c6a1b867fcfc0fb16a2ced5134552618b36.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 266, + 578, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 266, + 578, + 481, + 609 + ], + "type": "interline_equation", + "content": "B _ {\\mathrm {u p}} ^ {\\prime} = \\uparrow \\left(\\sum_ {i = 1} ^ {n} w _ {i} B _ {i}\\right). \\tag {10}", + "image_path": "67f11c68c1a53a948f1771c49d33af0dce8ae77e739cca579c1a8ff55a04e54d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\uparrow (\\cdot)" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " refers to upscaling through bilinear interpolation, " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "[w_{1},\\dots,w_{n}]^{T}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " is a weighting vector produced by a lightweight MLP (similar to EMLP, but with " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " output neurons) that processes our DEF. After computing " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", the estimated illuminant can be obtained by:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 116, + 455, + 235 + ], + "blocks": [ + { + "bbox": [ + 160, + 116, + 455, + 235 + ], + "lines": [ + { + "bbox": [ + 160, + 116, + 455, + 235 + ], + "spans": [ + { + "bbox": [ + 160, + 116, + 455, + 235 + ], + "type": "image", + "image_path": "618580631a98426f579bf158e013215a7ee2148b60ded767f5544d3053a0a5fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": "Fig. 4: Examples from the dataset used in this work. For each scene, we captured the scene with a gray calibration object placed in the scene to obtain the ground-truth illuminant (A) and captured the scene using different exposure settings without the gray object (B-H). The terms 'short /e' (C-E) and 'long " + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\times e" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": " (F-H) refer to multiplying and dividing auto exposure time by a factor " + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": ", respectively. The first image in (A) is displayed in sRGB, while the rest are shown in raw RGB space." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 342, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 342, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 225, + 342, + 481, + 369 + ], + "type": "interline_equation", + "content": "\\hat {\\ell} _ {u} = \\sum_ {u, v} u P (u, v), \\hat {\\ell} _ {v} = \\sum_ {u, v} v P (u, v), \\tag {11}", + "image_path": "2312acfe39e322aae8cdfce5ec3b704cc6d3336434fbb830ff91425daff86bef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 375, + 481, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 375, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 219, + 375, + 481, + 399 + ], + "type": "interline_equation", + "content": "\\hat {\\ell} = \\left[ \\exp \\left(- \\hat {\\ell_ {u}}\\right) / q, 1 / q, \\exp \\left(- \\hat {\\ell_ {v}}\\right) / q \\right] ^ {T}, \\tag {12}", + "image_path": "2cad851a7e3975a73dd1f33595eb92a87a4e42ba977d59daae2df284631607bb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 224, + 409, + 480, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 409, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 224, + 409, + 480, + 437 + ], + "type": "interline_equation", + "content": "q = \\sqrt {\\exp \\left(- \\hat {\\ell_ {u}}\\right) ^ {2} + \\exp \\left(- \\hat {\\ell_ {v}}\\right) ^ {2} + 1}. \\tag {13}", + "image_path": "eb992f04d71d78bc7ea24c0aad99912ccec9ad50f4c7f17655404541ae7b9324.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 452, + 230, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 230, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 230, + 466 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 477, + 186, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 477, + 186, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 186, + 488 + ], + "type": "text", + "content": "4.1 Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "To the best of our knowledge, the existing HDR multi-exposure datasets (e.g., [17, 33]) do not provide ground-truth illuminant colors for training and evaluation of our method. Thus, we compiled a dataset of scenes captured with both auto exposure and various multiple-exposure settings for each scene using Pixel 7 Pro camera (see Fig. 4). Images captured with auto exposure are used for training (in the case of learning-based methods) and evaluating other methods, while those captured with multiple exposure settings are employed for validating our method. Our set comprises 558 scenes, each captured using auto exposure and six additional exposure settings: " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", indicating adjustments to the original exposure time. Specifically, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " signifies doubling the original exposure time, " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " denote quadrupling and octupling, respectively, while " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " represent dividing the original exposure time by 2, 4, and 8, respectively. These adjustments are dynamically computed as ratios from the initial auto exposure, following the common practice in multi-exposure" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 100 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "type": "text", + "content": "HDR imaging on smartphone cameras [28, 38]. The camera was fixed on a tripod to minimize potential misalignment between images captured by different exposure settings. In total, we captured 3,906 raw images (" + }, + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "type": "inline_equation", + "content": "558 \\times 7" + }, + { + "bbox": [ + 130, + 116, + 482, + 222 + ], + "type": "text", + "content": " captures of each scene). The dataset is organized as follows: 83 and 86 scenes were randomly selected for validation and testing, respectively, with the remaining 387 scenes were used for training. Additionally, each scene was captured with a gray calibration object to obtain the ground-truth illuminant. The images with the gray calibration object are not included in any of the training, validation, nor testing sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": "During training, we optionally augment training data using chromatic adaptation. Specifically, we employ clustering on the training data using our DEF as query feature. We use K-means [34] with L2 to create 80 clusters. Subsequently, we augment the training data by generating three additional images for each sample. For each augmented image, we randomly select an illuminant color from the cluster to which the original image belongs. We then apply Von Kries transform [18] to remove the original illuminant from the image, using the actual ground truth, and apply the newly selected illuminant color to the image using the diagonal scaling operator. This new illuminant serves as the ground truth for the augmented image. As our method is designed to handle two images, namely " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": ", we use, without loss of generality, a single exposure factor, " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": ", for both images, such that " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": " was captured with " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "1 / e" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": " of the auto-exposure time, while " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": " was captured with " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\times e" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": " of the auto-exposure time, for simplicity. In Sec. 4.3, we present an ablation study on the impact of different values of " + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 224, + 482, + 402 + ], + "type": "text", + "content": " on our final results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 420, + 204, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 204, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 204, + 433 + ], + "type": "text", + "content": "4.2 Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": "We trained EMLP and ECCC using the Adam optimizer [40] with a mini-batch size set to 32 for 1000 and 200 epochs, respectively. For EMLP, the learning rate was " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": ", while ECCC was trained with an incremental mini-batch size (similar to [4]) with a learning rate of " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": " with a cosine annealing schedule [46], and the weight decay (i.e., L2 regularization penalty) was set to " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": ". For ECCC, the filter weights were initialized to zeros, and the biases were initialized to " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": " 2D histograms of training ground-truth illuminant colors after clustering into " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": " clusters using our DEF as a query feature with K-means. The " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": " histograms of training ground-truth illuminants were first processed by morphological dilation using a " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": " diamond-shaped structuring element before being used as initial values for our learnable biases. This initialization improves the results, as it assists in establishing reasonable \"candidate\" illuminant priors linked to the corresponding DEFs (see Sec. 4.3 for an ablation study). Both models were trained using the angular error, " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "L(\\cdot)" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": ", between the predicted illuminant color " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\hat{\\ell}" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": " and the ground-truth illuminant " + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 439, + 482, + 622 + ], + "type": "text", + "content": ", as described in the following equation [19]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 246, + 630, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 630, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 246, + 630, + 481, + 669 + ], + "type": "interline_equation", + "content": "L (\\ell , \\hat {\\ell}) = \\cos^ {- 1} \\left(\\frac {\\ell \\cdot \\hat {\\ell}}{\\| \\ell \\| \\| \\hat {\\ell} \\|}\\right), \\tag {14}", + "image_path": "a0546d49264704ffa82231583e73105812ed8e72e035565581c95cbf94dc5cbd.jpg" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " denotes the Euclidean norm, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " represents the vector dot product. For ECCC, we further added two smoothness loss terms to encourage smoothness in the learned filters and biases, similar to [4]. See supplemental materials for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 199, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 199, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 199, + 194 + ], + "type": "text", + "content": "4.3 Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 300 + ], + "type": "text", + "content": "We conducted a comprehensive comparison of our proposed methods against various existing techniques, which include: training-free statistical methods [13, 14, 25, 50], camera-independent learning-based methods [4, 5], and camera-dependent learning-based methods [6, 9, 29, 41, 42, 54, 63, 64]. For the camera-dependent learning-based methods, we trained each model on our data using the provided code and recommended parameters by the respective authors. Meanwhile, for sensor-independent learning methods, we utilized the provided pre-trained models on other datasets (e.g., [7, 14, 27])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 443 + ], + "type": "text", + "content": "For the sake of conducting realistic experiments, we assessed all methods, including ours, on " + }, + { + "bbox": [ + 130, + 300, + 482, + 443 + ], + "type": "inline_equation", + "content": "384 \\times 256" + }, + { + "bbox": [ + 130, + 300, + 482, + 443 + ], + "type": "text", + "content": " images—a suitable size for evaluating illuminant estimation methods designed for embedded hardware devices; such devices may utilize even smaller image sizes in white-balance modules [9]. For CCC methods [4,9], including the ECCC, " + }, + { + "bbox": [ + 130, + 300, + 482, + 443 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 300, + 482, + 443 + ], + "type": "text", + "content": " histograms were used. The ECCC incorporates the proposed DEF into the CCC by generating an interpolated bias map based on the DEF. This can be seen as a spatial case analogous to C5 [4], where C5 utilizes a neural network to emit bias and convolution filters based on additional images taken by the same camera to enhance generalization across cameras. To validate our modification against C5, we trained different versions of C5 not aimed at improving cross-camera generalization (as our method also focuses on a single camera). We refer to these modified versions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 453, + 480, + 582 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 453, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 480, + 487 + ], + "type": "text", + "content": "- C5 (model A): We use the histogram of the averaged short and long exposure images without additional images. That is, the C5 network uses a single histogram to generate CCC model parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 489, + 480, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 534 + ], + "type": "text", + "content": "- C5 (model B): We use histograms of both images taken by long and short exposures as input to the C5 network to generate CCC model parameters. The CCC mode is then applied to the histogram of the averaged long-short images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 536, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 582 + ], + "type": "text", + "content": "- C5 (models C and D): Both Model C and Model D use histograms of both images taken by long and short exposures to feed the C5 network and generate CCC model that is then applied to histogram of short exposure image (model C) and long exposure image (model D)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "The results are given in Tables 1 and 2 on the validation and testing sets, respectively. We report the mean, median, tri-mean, best " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", worst " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", worst " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", and max angular errors in each set. We also report the total number of parameters for other methods, including ours. Additionally, we present the results of a set of ablation studies conducted to examine the impact of the exposure factor " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " in both EMLP and ECCC. We also studied the impact of different color" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": "Table 1: Angular errors on the validation set. " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " denote camera-independent models and training-free statistical methods, respectively. Methods are listed chronologically by publication year, with the top and second-best results highlighted in yellow and red. We present the results of camera-dependent trained models for various versions of C5 [4], identified as models A, B, C, and D (see main text for more details). Our results are reported using EMLP, ECCC, and the ensemble model (EMLP + ECCC). Ablation studies are included using different mapping matrices (CM: " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " color mapping matrix, TM: " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "3 \\times 4" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " affine transformation matrix, HM: " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " homography matrix), different color representations (rgb: raw RGB triplet as used in [2,21-23,26], rg/rgb-chroma: rg-chromaticity/rgb-chromaticity of the raw RGB triplet), different exposure ratios (" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "e \\in \\{2,4,8\\}" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": "), different input image sizes and histograms (" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "48 \\times 32" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " input images, " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "48 \\times 32" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " input images and " + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": " histograms), using different input histograms for ECCC of both long and short exposure images, average image, and long/short image, different number of learned biases (" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 114, + 482, + 291 + ], + "type": "text", + "content": "), and results with (w/) and without (w/o) augmentation, w/ and w/o the covariance matrix's parameters (for EMLP), w/ and w/o DEF and bias initialization (BI) (for ECCC)." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 215, + 300, + 399, + 526 + ], + "blocks": [ + { + "bbox": [ + 215, + 300, + 399, + 526 + ], + "lines": [ + { + "bbox": [ + 215, + 300, + 399, + 526 + ], + "spans": [ + { + "bbox": [ + 215, + 300, + 399, + 526 + ], + "type": "table", + "html": "
MethodValidation setParams
MeanMed.Trl.Best 25%Worst 25%Worst 5%Max
Grayworld* [13]5.543.134.110.9712.8319.4528.82-
Shades of Gray* [25]7.166.216.430.9915.0618.3919.21-
PCA* [14]6.134.535.020.9113.6618.2820.33-
Gamut (pixels) [29]7.537.387.032.0413.9618.3421.01636
Gamut (edges) [29]7.036.066.161.4914.5419.5121.98324
FCC[3]3.771.732.320.549.5616.0828.9912,288
Gray Index * [50]6.675.786.890.9815.0624.4528.82-
APAP-LUT [6]4.843.305.871.1210.8416.9227.53289
SIE[5]4.704.044.231.359.2113.2716.241,008,044
BoCF [41]4.843.305.871.1210.8416.9227.5359,354
C4 [63]3.913.263.241.068.5012.7617.355,115,657
CWCC [42]4.493.353.481.619.2113.3116.11100,830
C5* [4]4.012.813.141.088.5112.4721.81411,711
C5 (model A) [4]3.932.292.901.019.4013.3318.75171,511
C5 (model B) [4]4.492.843.240.8810.5215.2320.55213,831
C5 (model C) [4]4.252.673.190.9310.1214.5719.82213,831
C5 (model D) [4]4.512.943.380.8710.7416.0120.74213,831
EMLP [4]4.162.743.270.8810.7414.0521.2992,910,186
PCCC [64]4.614.194.151.758.5012.0617.86450
EMLP [c = 8, rg-chroma, HM, w/o cov]5.864.264.961.3612.316.3620.44-
EMLP [c = 8, rg-chroma, TM, w/o cov]4.783.944.271.149.5713.2016.54363
EMLP [c = 8, rgb, CM, w/o cov]4.543.533.871.269.5113.7419.22300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.813.984.221.299.3212.7514.99255
EMLP [c = 2, rg-chroma, CM, w/o cov]4.353.333.571.119.4413.6017.14300
EMLP [c = 4, rg-chroma, CM, w/o cov]4.163.563.771.008.3812.1616.79300
EMLP [c = 8, rg-chroma, CM, w/o cov]4.073.513.641.237.7710.3813.29300
EMLP [c = 8, rg-chroma, CM, w/o cov]3.772.892.920.868.5911.4812.71354
EMLP [c = 8, n = 8, w/ DEF, both]3.642.792.860.868.6811.4213.54354
EMLP [c = 8, rg-chroma, CM, w/ aug, w/ cov]3.922.432.850.868.6811.3213.11354
ECCC [c = 8, n = 5, w/ DEF, both]3.692.783.110.797.8310.6212.372,166
ECCC [c = 8, n = 10, w/ DEF, both]3.662.692.920.958.0311.0612.323,496
ECCC [c = 8, n = 15, w/ DEF, both]3.582.712.910.917.9611.8513.514,826
ECCC [c = 2, n = 20, w/ DEF, both]3.912.993.231.078.3311.8415.236,156
ECCC [c = 4, n = 20, w/ DEF, both]4.283.193.360.999.6313.7214.726,156
ECCC [c = 8, n = 20, w/o DEF, both]4.023.233.691.229.1712.1414.474,608
ECCC [c = 8, n = 20, w/ BI, both]3.993.133.510.978.5311.4113.666,156
ECCC [c = 8, n = 20, w/ DEF, avg]4.032.973.341.038.6812.0914.715,900
ECCC [c = 8, n = 20, w/ DEF, short]3.712.742.931.008.0712.2015.435,900
ECCC [c = 8, n = 20, w/ DEF, long]3.702.742.931.008.0711.1115.435,900
ECCC [c = 8, n = 20, rg-chroma, w/ DEF, both]3.792.733.230.828.4111.6315.446,156
ECCC [c = 8, n = 20, w/ DEF, both, w/ aug]3.612.973.140.797.7710.5411.416,156
ECCC [c = 4, n = 20, w/ DEF, both]3.692.623.070.98.1210.9515.16,156
ECCC *c = 8, n = 20, w/ DEF, both)4.043.343.500.668.7511.2812.961,932
ECCC [c = 8, n = 20, w/ DEF, both]3.562.502.790.877.9311.3213.476,156
EMPLP + ECCC3.242.372.530.847.1110.6412.066,510
", + "image_path": "6922385beefb8ee5a2b01ffdab8104b106a0875c0282cda64583fb137add6b50.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "spaces of input images before computing our DEF. Specifically, we used the " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "rg" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "-chromaticity and raw RGB values (similar to [2, 21-23, 26]), in addition to our main design that uses " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "rgb" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": "-chromaticity. We also studied different mapping matrices, in addition to the " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " color matrix we used in Sec. 2. Specifically, we examined using the geometric " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "3 \\times 4" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " affine transformation matrix and the " + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 555, + 482, + 628 + ], + "type": "text", + "content": " homography matrix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "We report results of our EMLP with and without the covariance matrix parameters in addition to the results of ECCC with different numbers of learnable biases, " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", and with and without the proposed bias initialization. Furthermore, we" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 250, + 136, + 364, + 292 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "type": "text", + "content": "Table 2: Angular errors on the testing set. See caption of Table 1 for abbreviations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 250, + 136, + 364, + 292 + ], + "lines": [ + { + "bbox": [ + 250, + 136, + 364, + 292 + ], + "spans": [ + { + "bbox": [ + 250, + 136, + 364, + 292 + ], + "type": "table", + "html": "
MethodTesting set
MeanMed.Tri.Best 25%Worst 25%Worst 50%Max
Grayworld* [13]5.323.704.510.9511.3414.4117.75
Shades of Gray* [25]6.775.776.081.0714.2416.6417.50
PCA* [14]5.914.995.221.0712.6217.1019.08
Gamut [pixels] [29]7.497.257.372.3012.9616.6219.00
Gamut [edges] [29]5.904.804.931.1312.2017.6120.09
FFPC [9]3.181.832.280.397.9713.1320.58
Gray Index* [50]5.183.474.180.6012.3216.9019.67
APAP-LUT [6]3.942.823.320.818.5411.9016.36
SHE* [5]3.512.352.740.807.5611.0612.39
BoCf 441 [9]3.913.313.401.187.7211.2412.81
C4 [63]3.792.763.061.057.9311.3313.39
CWCC [42]3.712.853.061.117.7911.2112.83
CS* [4]3.452.933.030.957.0010.3814.01
CS (model A) [4]3.822.332.810.729.0815.6920.68
CS (model B) [4]3.302.052.420.697.5811.2314.02
CS (model C) [4]3.312.012.400.637.6712.0014.23
CS (model D) [4]3.332.202.420.647.6710.8313.35
TLCX [54]3.732.893.210.957.5411.5514.78
PCC [64]4.373.663.640.898.1415.1723.34
EMLP (ε = 2, w/ aug.)5.642.662.970.887.8711.518.82
EMLP (ε = 4, w/ aug.)3.502.632.780.927.5711.6213.42
EMLP+A (ε = 8, w/ aug.)3.372.342.740.677.2410.1614.10
EMLP (ε = 8, w/ aug.)3.362.732.840.767.039.5311.85
ECCC (ε = 8)3.332.922.920.966.569.3810.26
ECCC-ε2(ε = 8)3.412.873.030.816.879.1211.03
ECCC (ε = 2)2.942.102.380.696.318.139.25
ECCC (ε = 4)3.182.392.600.816.9910.0110.56
ECCC (ε = 8)3.002.312.450.746.669.8512.23
ECCC (ε = 8, w/ aug.)2.952.092.370.766.068.149.32
ELMP + ECCC2.912.332.440.726.118.389.83
", + "image_path": "d49fa48b8cfd814e9da81e236472b768919c0a1770f4ee2986b7a3077f760524.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "text", + "content": "show the results of ECCC without utilizing the DEF feature. Lastly, we demonstrate the influence of different sizes of input sizes on the inference accuracy of both EMLP and ECCC. The results show that the best results are obtained with " + }, + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "inline_equation", + "content": "e = 8" + }, + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "text", + "content": ". This choice is intuitively sensible, as a higher exposure factor increases the differences between the dual-exposure images, resulting in greater distinction based on scene lighting conditions. The augmentation is found to be useful for EMLP; however, we did not observe consistent improvement in the case of ECCC. This is likely because EMLP has a limited number of input features, making it more susceptible to overfitting. Thus, augmentation helps in generalization. Conversely, in the case of ECCC, the inclusion of histogram features alongside the small DEF feature suggests that the model may not consistently derive benefits from augmentation. We also noted that in ECCC, using raw RGB colors of dual-exposure images (without chromaticity conversion) tends to yield better results, while in EMLP, the chromaticity conversion tends to enhance performance (see Table 1). Results presented in Table 2 default to " + }, + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "inline_equation", + "content": "rgb" + }, + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "text", + "content": "-chromaticity for EMLP and raw RGB values for ECCC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "type": "text", + "content": "From the results presented in Tables 1 and 2, it is evident that our proposed DEF serves as a promising feature for guiding illuminant estimators. This is demonstrated by the performance of our lightweight EMLP model (354 parameters) compared to complex models that rely solely on raw RGB image colors (e.g., TLCC [54] with 32 million parameters). Moreover, incorporating DEF into ECCC reduces the maximum error by over " + }, + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 593 + ], + "type": "text", + "content": " and achieves comparable results across various evaluation metrics when compared to FFCC [9]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Since both models - namely, EMLP and ECCC - have a reasonable number of parameters, we can combine their predictions to create an ensemble model by averaging the predicted illuminant colors from both models. The results of this ensemble model are reported in Tables 1 and 2 under 'ECCC + EMLP'. The combined efforts of EMLP and ECCC, guided by our DEF, demonstrate promising results, outperforming several state-of-the-art approaches across vari-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 455, + 244 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 455, + 244 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 455, + 244 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 455, + 244 + ], + "type": "image", + "image_path": "84a7ef8cc5e191d03786ac653100b37e086159e16609c229ed2a4a065f06f630.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 340 + ], + "type": "text", + "content": "Fig. 5: Randomly selected examples from our worst " + }, + { + "bbox": [ + 130, + 251, + 482, + 340 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 251, + 482, + 340 + ], + "type": "text", + "content": " results of ECCC (top two rows are from validation set and remaining rows are from testing set). (A) Input pair of raw images captured with long and short exposures (note that other methods use a single image captured with auto exposure). (B-G) Images corrected with the estimated illuminant by: (B) FFCC [9], (C) C5 [4] (chosen the best results among all variations discussed in Sec. 4.3), (D) TLCC [54], (E) EMLP, (F) ECCC, and (G) Ground-truth illuminant. The estimated illuminant of each method is shown on the right side of each image, along with the angular error written in the top-left corner of the image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 401 + ], + "type": "text", + "content": "ous evaluation metrics while utilizing considerably fewer parameters. Qualitative examples of our results, randomly selected from the worst " + }, + { + "bbox": [ + 130, + 365, + 482, + 401 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 401 + ], + "type": "text", + "content": " of ECCC results, are presented in Fig. 5 alongside results from other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 422, + 326, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 326, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 326, + 435 + ], + "type": "text", + "content": "5 Conclusion and Future Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduced DEF, a feature derived from dual exposure images to enhance illuminant estimation. Our DEF achieves comparable results with state-of-the-art methods that employ thousands or millions of parameters (e.g., [54, 63]), using only 354 parameters in a straightforward MLP network. We further discuss incorporating the proposed DEF into the established CCC framework, referred to as ECCC. ECCC achieves comparable or better results on different evaluation metrics than classic CCC approaches while remaining lightweight, requiring only 6,156 parameters (50% reduction compared to FFCC [9]), approximately 30 KB of memory, and running in " + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 0.25" + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": " milliseconds per image on CPU. Our solution focuses on the single-camera case, intending testing on the same camera used for training. Future work includes studying the stability of this feature across different cameras. We discussed integrating the proposed DEF into the CCC framework. Further exploration may involve examining DEF's benefits for other illuminant estimation techniques that rely on convNets and raw image pixels as input. Another research direction could involve developing a spatially varying version of DEF, rather than our global feature, to utilize for spatially varying illuminant estimation and image white balancing." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: CVPR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Abdelhamed, A., Punnappurath, A., Brown, M.S.: Leveraging the availability of two cameras for illuminant estimation. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 476, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 476, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 476, + 194 + ], + "type": "text", + "content": "3. Affi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. In: BMVC (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "text", + "content": "4. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: ICCV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "text", + "content": "5. Affifi, M., Brown, M.S.: Sensor-independent illumination estimation for DNN models. In: BMVC (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "6. Affi, M., Punnappurath, A., Finlayson, G., Brown, M.S.: As-projective-as-possible bias correction for illumination estimation algorithms. Journal of the Optical Society of America A 36(1), 71-78 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "7. Banić, N., Koščević, K., Lončarić, S.: Unsupervised learning for color constancy. arXiv preprint arXiv:1712.00436 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 293, + 399, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 399, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 399, + 304 + ], + "type": "text", + "content": "8. Barron, J.T.: Convolutional color constancy. In: ICCV (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 304, + 442, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 442, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 442, + 314 + ], + "type": "text", + "content": "9. Barron, J.T., Tsai, Y.T.: Fast Fourier color constancy. In: CVPR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 315, + 440, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 440, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 440, + 326 + ], + "type": "text", + "content": "0. Bianco, S., Schettini, R.: Color constancy using faces. In: CVPR (2012)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "1. Brainard, D.H., Freeman, W.T.: Bayesian color constancy. Journal of the Optical Society of America A 14(7), 1393-1411 (1997)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "text", + "content": "2. Brown, M.: Color processing for digital cameras, chap. 5, pp. 81-98. John Wiley & Sons, Ltd (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "text", + "content": "3. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin Institute 310(1), 1-26 (1980)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "4. Cheng, D., Prasad, D.K., Brown, M.S.: Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. Journal of the Optical Society of America A 31(5), 1049-1058 (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "text", + "content": "5. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: Annual Conference on Computer Graphics and Interactive Techniques (1997)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "text", + "content": "6. Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Mobile computational photography: A tour. Annual Review of Vision Science 7, 571-604 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "text", + "content": "7. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Transactions on Graphics 36(6) (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 502, + 432, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 432, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 432, + 512 + ], + "type": "text", + "content": "8. Fairchild, M.D.: Color appearance models. John Wiley & Sons (2013)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 533 + ], + "type": "text", + "content": "9. Finlayson, G.D., Funt, B.V., Barnard, K.: Color constancy under varying illumination. In: ICCV (1995)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "20. Finlayson, G.D., Hordley, S.D.: Color constancy at a pixel. Journal of the Optical Society of America A 18(2), 253-264 (2001)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "text", + "content": "21. Finlayson, G.D., Hordley, S.D.: The chromagenic colour camera and illuminant estimation. In: Color Imaging Conference (2005)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "22. Finlayson, G.D., Hordley, S.D., Morovic, P.: Chromagenic colour constancy. In: Congress of the International Colour Association (2005)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "23. Finlayson, G.D., Hordley, S.D., Morovic, P.: Colour constancy using the chromogenic constraint. In: CVPR (2005)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "24. Finlayson, G.D., Hordley, S.D., Tastl, I.: Gamut constrained illuminant estimation. International Journal of Computer Vision 67(1), 93-109 (2006)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "25. Finlayson, G.D., Trezzi, E.: Shades of gray and colour constancy. In: Color and Imaging Conference (2004)" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "26. Fredembach, C., Finlayson, G.: The bright-chromagenic algorithm for illuminant estimation. Journal of Imaging Science and Technology 52, 40906:1-40906:11 (2008)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "27. Gehler, P.V., Rother, C., Blake, A., Minka, T., Sharp, T.: Bayesian color constancy revisited. In: CVPR (2008)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "28. Gelfand, N., Adams, A., Park, S.H., Pulli, K.: Multi-exposure imaging on mobile devices. In: ACM International Conference on Multimedia (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "29. Gijsenij, A., Gevers, T., Van De Weijer, J.: Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision 86(2), 127-139 (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "30. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing 20(9), 2475-2489 (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "text", + "content": "31. Gijsenij, A., Gevers, T., Van De Weijer, J.: Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(5), 918-929 (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "32. Granados, M., Ajdin, B., Wand, M., Theobalt, C., Seidel, H.P., Lensch, H.P.: Optimal HDR reconstruction with linear digital cameras. In: CVPR (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "33. Hanji, P., Mantiuk, R.K., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: SIGGRAPH (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "34. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A K-means clustering algorithm. Journal of the royal statistical society. Series C (applied statistics) 28(1), 100-108 (1979)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "text", + "content": "35. Hernandez-Juarez, D., Parisot, S., Busam, B., Leonardis, A., Slabaugh, G., McDonagh, S.: A multi-hypothesis approach to color constancy. In: CVPR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "36. Hu, Y., Wang, B., Lin, S.: FC4: Fully convolutional color constancy with confidence-weighted pooling. In: CVPR (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "37. Hubel, P.M., Finlayson, G.D., Hordley, S.D.: White point estimation using color by convolution (2007), US Patent 7,200,264" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "text", + "content": "38. Jung, C., Yang, Y., Jiao, L.: High dynamic range imaging on mobile devices using fusion of multiexposure images. Optical Engineering 52(10), 102004-102004 (2013)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "text", + "content": "39. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: ECCV (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "40. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "41. Laakom, F., Passalis, N., Raitoharju, J., Nikkanen, J., Tefas, A., Iosifidis, A., Gabbouj, M.: Bag of color features for color constancy. IEEE Transactions on Image Processing 29, 7722-7734 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "text", + "content": "42. Laakom, F., Raitoharju, J., Nikkanen, J., Iosifidis, A., Gabbouj, M.: Robust channel-wise illumination estimation. In: BMVC (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "43. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: WACV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "44. Liba, O., Murthy, K., Tsai, Y.T., Brooks, T., Xue, T., Karnad, N., He, Q., Barron, J.T., Sharlet, D., Geiss, R., et al.: Handheld mobile photography in very low light. ACM Transactions on Graphics 38(6), 164-1 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "45. Lo, Y.C., Chang, C.C., Chiu, H.C., Huang, Y.H., Chen, C.P., Chang, Y.L., Jou, K.: CLCC: Contrastive learning for color constancy. In: CVPR (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "46. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "47. Lou, Z., Gevers, T., Hu, N., Lucassen, M.P., et al.: Color constancy by deep learning. In: BMVC (2015)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "M. Afifi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 556 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "48. Oh, S.W., Kim, S.J.: Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition 61, 405-416 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "49. Qi, G., Chang, L., Luo, Y., Chen, Y., Zhu, Z., Wang, S.: A precise multi-exposure image fusion method based on low-level features. Sensors 20(6), 1597 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "text", + "content": "50. Qian, Y., Kamarainen, J.K., Nikkanen, J., Matas, J.: On finding gray pixels. In: CVPR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "51. Shen, R., Cheng, I., Shi, J., Basu, A.: Generalized random walks for fusion of multi-exposure images. IEEE Transactions on Image Processing 20(12), 3634-3646 (2011)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "52. Solhusvik, J., Hu, S., Johansson, R., Lin, Z., Ma, S., Mabuchi, K., Manabe, S., Mao, D., Phan, B., Rhodes, H., et al.: A 1392x976 2.8 " + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "m 120db CIS with per-pixel controlled conversion gain. In: International Image Sensor Workshop (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "text", + "content": "53. Solomatov, G., Akkaynak, D.: Spectral sensitivity estimation without a camera. In: ICCP (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "54. Tang, Y., Kang, X., Li, C., Lin, Z., Ming, A.: Transfer learning for color constancy via statistic perspective. In: AAAI (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "text", + "content": "55. Ulucan, O., Karakaya, D., Turkan, M.: Multi-exposure image fusion based on linear embeddings and watershed masking. Signal Processing 178, 107791 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "text", + "content": "56. Ulucan, O., Ulucan, D., Ebner, M.: Block-based color constancy: The deviation of salient pixels. In: International Conference on Acoustics, Speech and Signal Processing (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 369 + ], + "type": "text", + "content": "57. Van De Weijer, J., Gevers, T., Gijsenij, A.: Edge-based color constancy. IEEE Transactions on Image Processing 16(9), 2207-2214 (2007)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "text", + "content": "58. Willassen, T., Solhusvik, J., Johansson, R., Yaghmai, S., Rhodes, H., Manabe, S., Mao, D., Lin, Z., Yang, D., Cellek, O., et al.: A " + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "inline_equation", + "content": "1280 \\times 1080" + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "text", + "content": " 4.2 " + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "text", + "content": "m split-diode pixel HDR sensor in " + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "inline_equation", + "content": "110\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "text", + "content": " BSI CMOS process. In: International Image Sensor Workshop (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 413, + 482, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 434 + ], + "type": "text", + "content": "59. Xiao, J., Gu, S., Zhang, L.: Multi-domain learning for accurate and few-shot color constancy. In: CVPR (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "type": "text", + "content": "60. Xing, X., Qian, Y., Feng, S., Dong, Y., Matas, J.: Point cloud color constancy. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "type": "text", + "content": "61. Xu, B., Wang, N., Chen, T., Li, M.: Empirical evaluation of rectified activations in convolutional network. arXiv preprint arXiv:1505.00853 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "text", + "content": "62. Yahiaoui, L., Horgan, J., Deegan, B., Yogamani, S., Hughes, C., Denny, P.: Overview and empirical analysis of ISP parameter tuning for visual perception in autonomous driving. Journal of Imaging 5(10), 78 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "text", + "content": "63. Yu, H., Chen, K., Wang, K., Qian, Y., Zhang, Z., Jia, K.: Cascading convolutional color constancy. In: AAAI (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 533, + 482, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 556 + ], + "type": "text", + "content": "64. Yue, S., Wei, M.: Color constancy from a pure color view. Journal of the Optical Society of America A 40(3), 602-610 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 447, + 101 + ], + "type": "text", + "content": "HDR illuminant estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_content_list.json b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..70683b9f67fa6d25dea02fda3fdcd30b75b6b245 --- /dev/null +++ b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "text_level": 1, + "bbox": [ + 228, + 140, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Opher Bar Nathan $^{1}$ , Deborah Levy $^{1}$ , Tali Treibitz $^{1}$ , and Dan Rosenbaum $^{2}$", + "bbox": [ + 279, + 210, + 722, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hatter Department of Marine Technologies, Charney School of Marine Sciences", + "bbox": [ + 228, + 253, + 772, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, University of Haifa, Haifa, Israel osmosis-diffusion.github.io", + "bbox": [ + 380, + 268, + 620, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Underwater image restoration is a challenging task because of water effects that increase dramatically with distance. This is worsened by lack of ground truth data of clean scenes without water. Diffusion priors have emerged as strong image restoration priors. However, they are often trained with a dataset of the desired restored output, which is not available in our case. We also observe that using only color data is insufficient, and therefore augment the prior with a depth channel. We train an unconditional diffusion model prior on the joint space of color and depth, using standard RGBD datasets of natural outdoor scenes in air. Using this prior together with a novel guidance method based on the underwater image formation model, we generate posterior samples of clean images, removing the water effects. Even though our prior did not see any underwater images during training, our method outperforms state-of-the-art baselines for image restoration on very challenging scenes. Our code, models and data are available on the project's website.", + "bbox": [ + 259, + 345, + 743, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion Models $\\cdot$ Physics-Based Computer Vision $\\cdot$ Underwater Image Restoration", + "bbox": [ + 259, + 566, + 740, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 619, + 375, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Underwater images are used in many applications, such as, underwater construction and maintenance, marine sciences, and fisheries. However, their automatic analysis is hindered because of the optical effects of the water that strongly attenuates and scatters light in a wavelength dependent manner. This causes color distortion and loss of contrast that exponentially increase with depth3. With growing human activity in the oceans, clear underwater vision becomes increasingly important.", + "bbox": [ + 212, + 651, + 787, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Restoring underwater scenes is still a very challenging ill-posed problem. Classic approaches are based on designing priors for clean images or inverting the water formation model, and are limited by the ability to form strong priors. On", + "bbox": [ + 212, + 758, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "3 For consistency with computer vision literature the term depth is used throughout to refer to the distance from the camera rather than water depth.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b4cb06e8a2c39b20c10752b8793299cde98dfbcf28485bfb197e4c2d0735bd3.jpg", + "image_caption": [ + "Fig. 1: Our method receives as an input a single underwater image and outputs the restored clean image and an estimated depth map. The output is estimated using a diffusion prior trained on RGBD images and the physical image formation model." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the other hand, learning based approaches are limited by the lack of supervised training data of clean underwater images. This is a critical issue, as the ocean cannot be emptied for the sake of data collection.", + "bbox": [ + 212, + 376, + 784, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We suggest an unsupervised restoration method (Figs. 1,2) based on an inverse problem approach using a diffusion prior for both color and depth, coupled with the underwater image formation model. Restoring an image is formulated as posterior inference, computed using a natural image prior, and a likelihood term that is based on the underwater image formation model. The challenge with applying this approach for underwater image restoration directly is that (i) the degradation for each pixel depends on its depth, and other unknown parameters; and (ii) there is no ground truth clean underwater data to train the prior.", + "bbox": [ + 212, + 422, + 784, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve this, we replace the image prior with a prior on the joint space of color and depth of natural images. Adding depth to the prior allows us to formulate the forward model that forms the likelihood term over the observed corrupted underwater image, and apply posterior sampling. Moreover, this leverages the high capacity of diffusion models, in capturing the strong correlations between color and depth in natural scenes.", + "bbox": [ + 212, + 542, + 784, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to train a prior model of RGBD images using available datasets of natural outdoor scenes that were collected in air. Using in-air scenes for underwater images may be counter-intuitive, but actually has strong benefits: 1) it overcomes the lack of clean underwater image data; 2) it leads to a strong prior that captures the joint statistics governing color and depth in natural scenes, where, as opposed to underwater images, color does not fade with distance; 3) it prevents overfitting to specific types of underwater images.", + "bbox": [ + 212, + 633, + 784, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We then use this prior together with the underwater physical image formation model to simultaneously estimate the clean image, its depth, and the model parameters, all from a single underwater image (summarized in Fig. 2). We show that our method outperforms models that were trained on underwater data.", + "bbox": [ + 212, + 739, + 785, + 799 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66431529cd214904f8ef4ab142f8c8ab8cff33156409f1e8e3279aedda391347.jpg", + "image_caption": [ + "Fig. 2: The iterative sampling process starts in $t = T$ with random noise in 4 channels. The denoising step outputs denoised samples $\\hat{x}_0 = (\\hat{J}_0, \\hat{D}_0)$ . We use the underwater physical image formation model together with $\\hat{x}_0$ to optimize the water parameters $\\hat{\\phi}$ , and to guide the sampling towards the observed image. This process repeats itself, gradually updating both the estimated image and depth, until $t = 0$ , in which $x_0$ holds the method's estimate for both the reconstructed scene $J_0$ and its depth $D_0$ ." + ], + "image_footnote": [], + "bbox": [ + 225, + 141, + 777, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this paper are:", + "text_level": 1, + "bbox": [ + 215, + 503, + 560, + 518 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We train an RGBD prior, and demonstrate that modeling color and depth and jointly sampling them, provides a stronger diffusion prior for underwater image restoration.", + "2. We propose a new method that combines the RGBD prior of in-air data with the underwater image formation model, leading to a diffusion guidance method that generates the restored underwater images as posterior samples.", + "3. We demonstrate that our method outperforms state-of-the-art underwater image restoration methods both qualitatively and quantitatively on real and simulated data." + ], + "bbox": [ + 212, + 518, + 784, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We publish all our code, the new trained RGBD prior, images and results.", + "bbox": [ + 215, + 654, + 750, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 693, + 395, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Underwater Image Restoration and Depth Estimation", + "text_level": 1, + "bbox": [ + 215, + 724, + 705, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we review recent works that are most relevant to our method. See [61] for a recent more comprehensive review.", + "bbox": [ + 212, + 750, + 784, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classic underwater image restoration. To cope with the ill-posed nature of the problem, earlier works introduced tailored image priors. Some priors aim to estimate a depth or transmission map of the scene to reduce the number of the unknowns and then use the estimated depth to restore the scenes [7, 41, 42].", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised learning. Clean underwater image data is very scarce. Several datasets aim to mitigate this issue. In UIEB [35], the images denoted as ground-truth are generated from enhancement results of 9 different baseline methods, and having human observers vote for the best one. The LSUI dataset [40] is larger, produced using the same methodology, based on choosing results from 18 different methods. These datasets enable supervised learning methods, but are still limited by their scale, the quality of the baseline methods, and the human bias to choose visually pleasing rather than physically consistent images.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CWR [24] introduced the HICRD dataset, where the images are restored using optical parameters impressively measured using ocean optics instruments. Unfortunately, the images are acquired in a downward-looking position, and thus their depth range is very limited. In FUnIE-GAN [27] a training dataset is generated by having humans select good images from a large set of unprocessed underwater images. The humans are instructed to choose images where the foreground objects are identifiable. These are then distorted by a GAN to produce the paired poor images. A synthetic dataset was generated in [34]. The dataset is synthesized from the NYU-v2 RGBD dataset [46], using the image formation model equation and several sets of values for the water parameters.", + "bbox": [ + 212, + 267, + 787, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unsupervised learning. USUIR [19] aims to restore images without supervision, by separately estimating the image components (clean image, transmission, backscatter) and using them to construct an underwater image that is used for supervision against the original one. Subsequent frames were used in [4] for self-supervising monocular depth estimation, and in [54] for self-supervising both depth and restoration. In UW-NET [23] a cycle-GAN is used for learning mapping from RGBD in-air datasets to underwater images. As opposed to all these methods, we present the first prior based on diffusion models that does not rely on ground truth underwater supervision, and uses the physical model for inference.", + "bbox": [ + 212, + 417, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Diffusion Model Prior", + "text_level": 1, + "bbox": [ + 214, + 589, + 444, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models have emerged as a powerful type of generative models. In the last few years several formulations and variations have been developed [25,47,50,52], most of which use a U-Net architecture [43] as a noise predictor. Because the training relies on very large datasets and is extremely time-consuming, significant work has been devoted to the setup where models are first pre-trained on large datasets and only later fine-tuned to more specific data, closer to the tasks at hand [44, 48]. In DepthGen [45] this approach is taken forward, by using a pre-trained model to kickstart a new model trained on different modalities using a different architecture. This is done by replacing the input and output layers of the pre-trained U-Net. We take a similar approach for training our RGBD prior, by using a pretrained diffusion model that was trained on RGB only.", + "bbox": [ + 212, + 613, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional diffusion models have been used for various image restoration tasks, by training models using different levels of supervision. Examples of similar tasks as ours include dehazing and deraining [8,39,56], and shadow removal [22].", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional diffusion models for underwater image enhancement have also been proposed [37,53]. Our approach differs from these by focusing on image restoration that inverts the physical model, rather than relying on supervised data that is optimized for image appearance.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion model as a prior for clean images. In addition to conditional generation of images, there is a growing body of research where diffusion models are being used as clean image priors, and image restoration is formulated as posterior sampling [5, 9, 11-14, 18, 21, 28, 29, 49, 51]. These include tasks like denoising, inpainting, deblurring, and more general tasks. The limited access to clean data in many cases, has led to research on training a prior of clean images, using noisy training data only. In [1, 15, 30] a diffusion model prior is trained using noisy data, assuming a known degradation model. Since this setup is not directly applicable in our case, we chose instead to train our prior on clean data that was not taken underwater.", + "bbox": [ + 212, + 207, + 784, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion model prior for blind image restoration. A more challenging task, is to use diffusion models as a clean image prior when the degradation model is unknown or depends on unknown parameters. Several papers have proposed to tackle this problem in different setups. In [17,38,59] the unknown parameters are optimized during the sampling process with a reconstruction objective function. In [10] a separate prior is trained for the unknown parameters, and sampled in parallel with image sampling. Our method differs from the above in that we learn a prior of the main unknown aspect of the degradation model, namely the depth, together with the prior of the variables we want to infer - the image color. This is done by training a single diffusion model on the joint space of color and depth.", + "bbox": [ + 212, + 358, + 785, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 547, + 380, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Underwater Image Formation", + "text_level": 1, + "bbox": [ + 214, + 579, + 506, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In water, we observe two wavelength- and distance-dependent effects. First, the direct signal reflected from the object is attenuated. Second, light is scattered onto the object's line-of-sight (LOS), creating an additive signal termed backscatter that increases with distance. The occluding backscatter layer is independent of the scene content. Thus, the visibility and contrast of further objects is significantly reduced and their colors are distorted.", + "bbox": [ + 212, + 603, + 784, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the revised underwater image formation model [2,32], under ambient illumination image intensity (per pixel, per color channel) is given as:", + "bbox": [ + 212, + 694, + 782, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI = J \\cdot e ^ {- \\phi_ {a} \\cdot D} + \\phi^ {\\infty} \\cdot \\left(1 - e ^ {- \\phi_ {b} \\cdot D}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 736, + 784, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I$ is the linear image captured by the camera of a scene with range $D$ , $J$ is the clear scene that would have been captured had there been no water along the LOS, and $\\phi^{\\infty}$ is the water color at infinity, i.e., the backscatter at areas that contain no objects. The two parameters $\\phi_{a}$ and $\\phi_{b}$ are the attenuation and backscatter coefficients, respectively.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 146, + 401, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffusion models have proven to be very effective in capturing the distribution of natural images, given in training data. We describe here the formulation that we use in our work, adopted from [16]. A diffusion model is defined by a Markov chain designed to transform the distribution of real images $x_0$ to a Gaussian distribution $x_{T}$ , by gradually scaling down the image values and adding Gaussian noise, using a schedule determined by the scalar parameters $\\alpha_{t}$ ,", + "bbox": [ + 212, + 170, + 787, + 261 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 272, + 785, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this forward process, an inverse process is trained to gradually denoise images starting from a Gaussian distribution, back into the distribution of the original image dataset. This can be formulated as a factorization of the joint distribution over the images in reverse order, $p(x_{0:T}) = \\prod_{t=1}^{T} p_t(x_{t-1} \\mid x_t)$ , and $p_t(x_{t-1} \\mid x_t)$ is approximated by a Gaussian", + "bbox": [ + 214, + 297, + 787, + 376 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} \\left(x _ {t - 1} \\mid x _ {t}\\right) \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) ， \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 385, + 785, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with parameters $\\mu_{\\theta}$ and $\\Sigma_{\\theta}$ predicted by a trained neural network conditioned on $x_{t}$ and $t$ . Simulating this process results in samples form the approximated data distribution $p(x_0)$ . A popular implementation first predicts the noise at each time step, and then computes the mean by", + "bbox": [ + 214, + 411, + 787, + 472 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\theta} (x _ {t}, t) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 482, + 785, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(x_t,t)$ is the neural network trained to approximate $\\epsilon$ , and $\\bar{\\alpha}_t = \\prod_{s=0}^t \\alpha_s$ . The same neural network can also be used to predict a diagonal covariance $\\Sigma_{\\theta}(x_t,t)$ . At any iteration, an estimate of the clean image $x_0$ can be derived by computing the mean of $p(x_0 \\mid x_t)$ :", + "bbox": [ + 214, + 525, + 787, + 587 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {0} (x _ {t}, t) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t}, t)\\right). \\qquad \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 597, + 785, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above formulation can also be developed from a score-based modeling approach, where it can be shown that $\\epsilon_{\\theta}(x_t,t)$ is an approximation to the score function $\\nabla_{x_t}\\log p_t(x_t)$ .", + "bbox": [ + 214, + 636, + 787, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Posterior Sampling", + "text_level": 1, + "bbox": [ + 215, + 703, + 419, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Under the score-based view, given an observation $y$ and a likelihood function $p(y|x)$ , we can use the same mechanism to sample from the posterior $p(x|y)$ using the posterior score function", + "bbox": [ + 214, + 727, + 785, + 772 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {x _ {t}} \\log p _ {t} (x _ {t} | y) = \\nabla_ {x _ {t}} \\log p _ {t} (x _ {t}) + \\nabla_ {x _ {t}} \\log p _ {t} (y | x _ {t}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 784, + 785, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This idea of adding a conditional signal to the score is also called guidance. The challenge with the second term in Eq. 6, is that usually we are given a likelihood", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "O.Bar Nathan et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model based on the clean image $x_0$ , and not an intermediary noisy image $x_t$ . In our case the underwater image formation model transforms a clean image to an underwater observation. Connecting the model to the noisy sample $x_t$ leads to the intractable integral $p_t(y|x_t) = \\int p(y|x_0)p(x_0|x_t)dx_0$ .", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The various works on posterior sampling propose different approximations of this integral. Some propose to collapse the uncertainty in $y$ [9, 12-14, 21, 51]. In [18] a variational inference approximation is proposed, and in [5, 11, 28, 29, 49] the likelihood model is computed on either $x_{t}$ directly, or on the mean $\\hat{x}_0 = \\mathbb{E}[x_0|x_t]$ , as given in Eq. 5. After some experimentation, we decided to use the latter approximation as formulated in DPS [11]:", + "bbox": [ + 212, + 207, + 784, + 297 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {x _ {t}} \\log p _ {t} (x _ {t} \\mid y) \\approx \\nabla_ {x _ {t}} \\log p _ {t} (x _ {t}) + s \\nabla_ {x _ {t}} \\log p (y \\mid \\hat {x} _ {0} (x _ {t}, t)), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 309, + 784, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $s$ is the guidance scale used to control the weight of the approximated likelihood term.", + "bbox": [ + 212, + 335, + 782, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 388, + 330, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the underwater image formation model (Eq. 1) as a forward model that maps the space of natural images $x$ to the space of underwater observed images $y$ . We aim to use this forward model to construct a likelihood term $p(y \\mid x)$ , and use it to sample from the posterior distribution of images (Eq. 7). In our case this cannot be implemented directly since the image formation model contains unknown parameters, the depth $D$ at each pixel, and the water parameters, $\\phi_{a}, \\phi_{b}, \\phi^{\\infty}$ , which can differ between scenes. One way to deal with the unknown depth, is to use a monocular depth estimator in a separate first stage, and then use it as part of the forward model. In the results we discuss this approach and show that it is suboptimal (method variant termed DA-Osmosis).", + "bbox": [ + 212, + 420, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another approach is to optimize the unknown parameters, including the depth, during sampling. In some recent work, different methods to do this were proposed [17, 38, 59]. However, these do not work properly when the unknown parameter is high dimensional, highly complex, and strongly correlated with the target image. In our early experimentation we found that in order to separate the image into a clean image and the effects governed by the depth, we need more than a good image prior.", + "bbox": [ + 212, + 571, + 784, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, one of our main contributions, is to train a joint prior on both the color and depth in clean images, and show how it can be used for underwater image restoration. Following this we define $x_0 = (J, D)$ , where $J$ represents a clear image and consists of the 3 color channels, and $D$ is the depth image. This allows us to use the following forward model (based on Eq. 1) conditioned on the estimated clean image and depth $\\hat{x}_0 = (\\hat{J}_0, \\hat{D}_0)$ :", + "bbox": [ + 212, + 676, + 784, + 768 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\phi} \\left(\\hat {x} _ {0}\\right) = \\hat {J} _ {0} e ^ {- \\phi_ {a} \\hat {D} _ {0}} + \\phi^ {\\infty} \\left(1 - e ^ {- \\phi_ {b} \\hat {D} _ {0}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 780, + 784, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\phi = (\\phi_{a},\\phi_{b},\\phi^{\\infty})$ are the remaining 9 unknown water parameters (a parameter per color channel for each of $\\phi = (\\phi_a,\\phi_b,\\phi^\\infty)$ ). We optimize $\\phi$ during", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e86b82f8a475dc3e49e2c4717a28416d5ea7469eb3c4c8d8317064a08b3868c.jpg", + "image_caption": [ + "Fig. 3: [Left] Example images from outdoor RGBD datasets used for training our prior. From left to right: DIODE [55], ReDWeb-S [36], HR-WSI [57], KITTI [20]. [Right] Samples from the trained RGBD prior. The samples demonstrate the inherent correlation between RGB image and depth in our trained RGBD prior." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sampling, in a similar way to GDP [17], using gradient descent on the image reconstruction loss. The likelihood term used for guidance is defined as a Gaussian around $f_{\\phi}(\\hat{x}_0)$ with a fixed variance.", + "bbox": [ + 212, + 345, + 782, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training a joint prior on color and depth, not only allows us to use the above forward model, but also exploits the diffusion model capacity to capture the complexity of the depth image and its correlations with the clean color image.", + "bbox": [ + 212, + 393, + 784, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Training the Prior", + "text_level": 1, + "bbox": [ + 215, + 469, + 413, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To learn a natural image prior, we train the joint prior model on both color and depth using public RGBD datasets of outdoor scenes that were taken in air. While differing from underwater images, the ability to use large amounts of high quality training data leads to a prior that captures the correlation between color and depth in natural scenes, which as we show, is an important aspect for underwater image restoration.", + "bbox": [ + 212, + 503, + 784, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the sake of data efficiency and training time, we start with a pretrained diffusion model trained on ImageNet (we take the unconditional model from [16]), and finetune it on RGBD data. We implement this, inspired by DepthGen [45], by replacing the first and last layers of the U-Net to 4 channels rather than 3, and initializing those layers randomly. The datasets we use for finetuning are (see Fig. 3[left]): DIODE [55], RedWeb-S [36], HR-WSI [57], KITTI [20], with 16884, 2179, 20378, 23946 pictures, respectively.", + "bbox": [ + 212, + 595, + 784, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A major challenge in working with RGBD data is to turn the available depth information into a proper image. This includes filling holes and scaling the values to a standard range. Since each of the above datasets were collected in a different manner, we treat each of them differently. This is detailed in the supplementary material. Fig. 3[right] shows two samples from the prior that show corresponding color and depth images. While prior samples demonstrate an evident domain gap between in-air and underwater data, our results show that posterior samples are not affected by the gap, and can leverage the strong correlation between depth and color in natural scenes.", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Sampling from the Posterior", + "text_level": 1, + "bbox": [ + 215, + 146, + 496, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given a trained prior, we perform underwater image restoration by sampling from the posterior with guidance of the image formation model. The sampling process is described in Figs. 2, 4. In each iteration we generate samples of both the image and the depth, while optimizing the water parameters. This results in a gradual update to the estimates of the clean image, and the depth, demonstrated in Fig. 4[right]. We adopt the sampling method in DPS [11], using the reconstruction loss, $\\left\\| y - f_{\\phi}(\\hat{x}_0)\\right\\|_2^2$ , which is the negative log-likelihood formed from the model in Eq. 8. Similar to GDP [17], the remaining unknown parameters $\\phi$ are optimized in parallel to the inverse sampling process, using gradient descent on the same loss used for guidance.", + "bbox": [ + 212, + 172, + 787, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that one of our main novelties is that the RGBD prior is used as an inherent part of the iterative method and not as a stand-alone depth estimator. The depth is estimated together with the image in every iteration, guided by the underwater model, see, e.g., Fig. 4[right]. As demonstrated in the results, this improves the restoration quality over using a fixed estimated depth of the scene.", + "bbox": [ + 212, + 325, + 787, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Running posterior sampling, guided by a reconstruction loss only, we observe two problems. First, for pixels with large depth values, the reconstructed image is dominated by the backscatter, making the estimation of the clean pixel color values unstable. Second, the color values can shift outside the valid range, causing color saturation. To overcome the first issue we multiply the reconstruction loss in every pixel by the estimated depth value (without passing the gradients of this operation, denoted by $sg$ ):", + "bbox": [ + 212, + 402, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\left\\| s g (\\hat {D} _ {0}) \\cdot \\left(y - f _ {\\phi} (\\hat {x} _ {0})\\right) \\right\\| _ {2} ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 521, + 784, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In order to overcome the second issue, we introduce two auxiliary losses. The first penalizes RGB values outside the valid range $[-1, 1]$ , and the second encourages the average values of each channel to approach the middle of the color range (in line with the gray world assumption). We implement the two auxiliary losses as:", + "bbox": [ + 212, + 551, + 787, + 612 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v a l}} = \\lambda_ {v} \\sum_ {i, c} \\max \\left(\\left| \\hat {J} _ {0} (i, c) \\right| - T _ {v}, 0\\right) ^ {2}, \\quad \\mathcal {L} _ {\\mathrm {a v r g}} = \\lambda_ {a} \\sum_ {c} \\left| \\sum_ {i} \\hat {J} _ {0} (i, c) - T _ {a} \\right|\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 625, + 759, + 665 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{J}_0(i,c)$ is the value of color channel $c$ of pixel $i$ , assuming the valid color range is $[-1,1]$ , $T_v$ and $T_a$ are thresholds set to values close to 1 and 0 respectively, and $\\lambda_v,\\lambda_a$ are the scalar weights of both losses. The total loss is then:", + "bbox": [ + 212, + 679, + 787, + 724 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {v a l}} + \\mathcal {L} _ {\\text {a v r g}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 742, + 784, + 757 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "At each iteration we compute the gradient of the loss with respect to $x_{t}$ , which forms the log-likelihood gradient in the posterior score (Eq. 7), and the gradient with respect to $\\phi$ which is used to update the parameters in the underwater model. In order to stabilize the optimization, we apply gradient clipping by value, to the gradients of $x_{t}$ , and we run the optimization of $\\phi$ only in some of", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Osmosis - Sampling" + ], + "code_body": "Input: y - Underwater Degraded Image (RGB image) $f(\\hat{x}_0)$ - Underwater Physical Degradation Model Initialization of $\\phi = \\{\\phi_a,\\phi_b,\\phi^\\infty \\}$ // size = [9] Output: $\\hat{x}_0 = [\\hat{J}_0,\\hat{D}_0]$ .. $\\hat{J}_0$ - Restored RGB Image, $\\hat{D}_0$ - Depth Estimation \n1 $x_{T}\\sim \\mathcal{N}(0,I)$ // $x_{T}$ size = [h,w,4] \n2 for $t = T:1$ do \n3 $\\epsilon \\gets \\epsilon_{\\theta}(x_t,t)$ \n4 $\\mu = \\frac{1}{\\sqrt{\\alpha_t}} (x_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\cdot \\epsilon)$ \n5 $\\hat{x}_0 = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}} (x_t - \\sqrt{1 - \\bar{\\alpha}_t}\\cdot \\epsilon)$ \n6 $f_{\\phi}(\\hat{x}_0) = \\hat{J}_0e^{-\\phi_a\\hat{D}_0} + \\phi^{\\infty}\\left(1 - e^{-\\phi_b\\hat{D}_0}\\right)$ \n7 $\\mathcal{L} =$ loss from Eq. 10 \n8 if Optimstart $\\geq t\\geq$ Optimend then \n9 for $i = 1:N$ do \n10 | $\\phi \\gets \\phi -\\eta \\nabla_{\\phi}\\mathcal{L}$ \n11 end \n12 end \n13 $x_{t - 1}\\sim \\mathcal{N}(\\mu -s\\cdot \\nabla_{x_t}\\mathcal{L},\\Sigma)$ \n14 end", + "bbox": [ + 225, + 164, + 609, + 478 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/60362b871aea45a00a5c8a56e41f258b858e99b217edf629e8f74acc0c4ae836.jpg", + "image_caption": [ + "Fig. 4: Our algorithm. [Left] Detailed steps of our algorithm. [Right] Example of how $\\hat{J}_0, \\hat{D}_0$ change during the iterations." + ], + "image_footnote": [], + "bbox": [ + 625, + 143, + 782, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the sampling steps (defined by the values $\\mathrm{Optim}_{\\mathrm{start}}$ and $\\mathrm{Optim}_{\\mathrm{end}}$ ), running $N$ iterations of gradient descent in each step. When the gradient of $x_{t}$ is applied in the sampling step, it is multiplied by a guidance scale $s$ . We find that using a smaller scale for the depth channel leads to better results. This makes sense as while all channels are treated the same in the prior, the depth has a different role in the forward model (specifically it is used inside an exponent), and this can lead to a different gradient scale. More details on the implementation are given in the supplementary.", + "bbox": [ + 212, + 550, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 215, + 694, + 323, + 709 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use a prior on $256 \\times 256$ images, with sampling time of about 3 minutes per image on an Nvidia A100 GPU. We present here selected results and analysis. Please refer to the supplementary material for a complete set.", + "bbox": [ + 212, + 726, + 785, + 772 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Real-World Scenes. We present an extensive comparison on real-world linear images from the datasets SQUID [7], SeaThru [3], SeaThruNerf [32], and additional images acquired in different locations in the world, the Indian and", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 410, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pacific Oceans, and Mediterranean, and Red Seas. Images are white-balanced as pre-processing. All runs use the same set of parameters. Fig. 1 shows several results of both image restoration and depth estimation.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare image restoration with the following methods: CWR [24], DM [53], FUnIE-GAN [27], GDCP [41], IBLA [42], MMLE [60], semi-UIR [26], Ucolor [33], USUIR [19], UW-Net [23], waternet [35], and USe-ReDI-Net [54]. For compactness, Fig. 5 summarizes the results for a chosen subset of methods. The complete comparison is shown in the supplementary material. Among previous methods, we found GDCP to be the most consistent. Our result recovers the full range of the scenes, specifically improving contrast of objects that are further away (note the zoomed-in objects in the far areas, e.g., the diver in row 1). Our method recovers vibrant and consistent colors also in further areas. For example, in rows 2,3,4 note that the background rocks and sand appear bluish in all methods except of ours.", + "bbox": [ + 212, + 191, + 785, + 357 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compare to a method that restores the RGB using a fixed depth that is pre-estimated using an external SOTA method, depth anything (DA) [58]. We use the pre-estimated depth in our method's pipeline instead of our gradually estimated depth. We term this method $DA$ -Osmosis, and the results show that our method restores the far areas with better color, emphasizing the advantage of using our joint RGBD prior.", + "bbox": [ + 212, + 359, + 785, + 449 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare depth estimation with GDCP [41], IBLA [42], unveiling [6], UW-Net [23], and monoUWnet [4], and depth anything (DA) [58]. In GDCP [41] and unveiling [6] we compute the depth from the output transmission maps. Our depth estimations have more details and better explain the scenes (Fig. 5).", + "bbox": [ + 212, + 450, + 785, + 511 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition, we conducted a non-reference quantitative comparison using the MUSIQ [31] measure on 50 real-world images presented in the paper and supplementary material. Our method achieved the highest score. The quantitative results are summarized in Table 1.", + "bbox": [ + 212, + 511, + 785, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The UIEB dataset [35] is sometimes used for quantitative evaluation of image restoration. However, it is collected from various sources and is not linear. Thus it is not suitable for physics-based methods like ours. In addition, the 'ground truth' of the UIEB dataset is chosen from results of previous methods and it is not the real ground truth. Fig. 6 shows four examples (out of many) where our result removes more water effects than the ground truth, even despite the non-linearity of the input.", + "bbox": [ + 212, + 571, + 785, + 679 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Simulation. We conducted a simulation following [26, 34] using all the 1449 images from the NYU-v2 RGBD dataset [46] with randomly varying water parameters and generated the underwater images with the image formation model in Eq. 1. For fairness in comparison with other methods we applied $\\phi_{a} = \\phi_{b}$ .", + "bbox": [ + 212, + 702, + 785, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The quantitative results are summarized in Table 2. Our method substantially outperforms other methods in the image restoration metrics of PSNR, SSIM, and LPIPS. We emphasize that our prior was not trained on this data, or indoor data at all. Moreover, two of the methods [26,34] we outperform were trained on this data.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f4b142b6cdff793a8bd591999f1024301558cf19c1d87c7a599d39925154352d.jpg", + "image_caption": [ + "Fig. 5: Real-world restoration results. From left to right: white-balanced input, contrast stretch, GDCP [41], Ucolor [33], CWR [24], semi-UIR [26], DM [53], Depth Anything [58] - Osmosis, Osmosis (ours). Zoom-in colored rectangles emphasize far objects that have higher contrast in our results. Real-world depth results. From left to right: GDCP [41], IBLA [42], unveiling [6], UW-Net [23], monoUWnet [4], Depth Anything [58], Osmosis (ours). Our depth results are smoother and less affected by object gradients. The reader is encouraged to zoom-in." + ], + "image_footnote": [], + "bbox": [ + 218, + 170, + 784, + 700 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9b38183a0241158bcbbb723669f1d50af5b3358eb4b5a8805773054b93d2cece.jpg", + "image_caption": [ + "Fig. 6: Example results on UIEB [35], a dataset comprised from non-linear images with ground truth produced by different enhancement algorithms. Our method yields better results than the dataset's ground truth. See for example the color of the sand, the orange color of the fish in the top-left, and the divers' authentic skin color in the bottom-right." + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 782, + 303 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a61fe0eaf08a26298a39d1d1ebb64075ba9f8dca4163305b8b53b0112cba9e9.jpg", + "image_caption": [ + "Fig. 7: Ablation. Ablations $\\#1, \\#5, \\#6$ show the importance of the losses we added to the reconstruction loss. Without $\\mathcal{L}_{\\mathrm{val}} (\\#6)$ , the colors tend to \"explode\" and oversaturate. Without $\\mathcal{L}_{\\mathrm{avg}} (\\#5)$ the colors sometimes skew towards pink/purple. In $\\#2$ , the further areas are not restored well, because the loss is not weighted by the depth $D$ . In $\\#3$ the guidance scale is the same for the RGB and depth channels. This harms the depth reconstruction. In $\\#4$ we set $\\phi_{a} = \\phi_{b}$ (separately per color channel). Since this is an inaccurate model, the restoration in further areas is harmed." + ], + "image_footnote": [], + "bbox": [ + 220, + 398, + 782, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation. To demonstrate the effect of different components in our methods, we conduct an ablation study of the following variants: 1. $\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}}$ (instead of Eq. 10); 2. Removing weighting by $\\hat{D}_0$ in Eq. 9 3. Same guidance scale $s$ for all channels; 4. $\\phi_{a} = \\phi_{b}$ ; 5. $\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}} + \\mathcal{L}_{\\mathrm{val}}$ (remove $\\mathcal{L}_{\\mathrm{avg}}$ from Eq. 10); 6. $\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}} + \\mathcal{L}_{\\mathrm{avg}}$ (remove $\\mathcal{L}_{\\mathrm{val}}$ from Eq. 10). Numerical results on the simulation for variants $\\# 1 - \\# 3$ are presented in Table 2. Ablations $\\# 4 - \\# 6$ are shown only on real world images since in the simulation we do not use $\\mathcal{L}_{\\mathrm{avg}}$ , and $\\phi_{a} = \\phi_{b}$ . Fig. 7 presents the results of all variants on one of the scenes. We see that the additional losses are important to prevent color saturation and shift. Increasing the loss weight with depth improves restoration in further areas. Separating guidance scales between the RGB and depth channels improves depth reconstruction. Setting $\\phi_{a} \\neq \\phi_{b}$ extends the range of the restoration.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/80675c08c4d0b032b28ef1e331d6b0fb20c43d32389eccc896c2c4ffaadc94a6.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Non-reference quantitative comparison on 50 real-world images using MUSIQ [31]. Our method achieves the highest score." + ], + "table_body": "
MethodMUSIQ↑
input51.59
contrast stretch53.75
CWR [24]39.63
DM [53]54.64
FUnIE-GAN [27]42.70
GDCP [41]56.25
IBLA [42]54.42
MMLE [60]56.26
semi-UIR [26]54.99
Ucolor [33]47.73
USUIR [19]52.44
UW-Net [23]46.02
waternet [35]53.64
unveiling [6]50.83
DA-osmosis [58]51.23
osmosis (ours)56.62
", + "bbox": [ + 254, + 150, + 439, + 388 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/fd23a0af7ef71ba85ff983c25bb0c33b6e362d4b27e8c4fe46f4cca0be88ddab.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Quantitative comparison on the simulation. Our method achieves best scores in image restoration." + ], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
contrast stretch17.130.830.11
CWR [24]16.930.790.20
DM [53]17.410.820.12
FUnIE-GAN [27]17.640.770.21
GDCP [41]12.410.710.16
IBLA [42]15.070.700.19
MMLE [60]17.000.740.17
semi-UIR [26]17.820.830.12
Ucolor [33]17.920.830.10
USUIR [19]16.760.800.18
UW-Net [23]18.040.750.26
waternet [35]17.270.820.11
unveiling [6]16.340.790.18
ablation #121.090.860.09
ablation #222.170.880.07
ablation #322.000.880.06
osmosis (ours)22.740.890.06
", + "bbox": [ + 522, + 143, + 802, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 486, + 354, + 502 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work we demonstrated how to harness the strength of a new RGBD diffusion prior to achieve state-of-the-art results on underwater image restoration. To do this we solved several challenges: i) because of lack of clean underwater image data we use datasets of scenes in air; ii) we notice that the color prior does not suffice to guide restoration, therefore, we add the scene depth to the prior; iii) We use the physical image formation model to guide restoration, and also estimate the water parameters in the process. This results in the most comprehensive single underwater image restoration method to-date. It does not train on any underwater images and therefore does not overfit to any. It gives higher weight to further objects and therefore is superior in reconstructing all the details of the scene. Jointly solving for the depth results in excellent depth estimation from monocular images.", + "bbox": [ + 212, + 521, + 785, + 702 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "There is a domain gap between the prior (trained on in-air data) and underwater data. This is intentional. The beauty of our method is that by incorporating the underwater model in the sampling it succeeds even without training on underwater data (that is very hard to obtain). Thus, it is not specialized for a certain type of water or objects, and it learns the color distribution of natural images not degraded by water.", + "bbox": [ + 212, + 703, + 785, + 794 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Like all diffusion models with U-Nets, our method is limited by a fixed resolution and long running time. This could potentially be improved using other network architectures and sampling methods.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 410, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. The research was funded by Israel Science Foundation grant #1951/23, Israeli Ministry of Science and Technology grants #1001577600 & #1001593851, EU Horizon 2020 research and innovation programme GA 101094924 (ANERIS), the Leona M. and Harry B. Helmsley Charitable Trust, and the Maurice Hatter Foundation. We thank Dr. Derya Akkaynak and Dr. Matan Yuval for substantial data contribution, Amir Dayan for the paper's name and Meirav Keidar for graphics design.", + "bbox": [ + 215, + 147, + 810, + 229 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 251, + 323, + 268 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aali, A., Arvinte, M., Kumar, S., Tamir, J.I.: Solving inverse problems with score-based generative priors learned from noisy data. arXiv preprint arXiv:2305.01166 (2023)", + "2. Akkaynak, D., Treibitz, T.: A revised underwater image formation model. In: CVPR. pp. 6723-6732 (2018)", + "3. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: CVPR. pp. 1682-1691 (2019)", + "4. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: IEEE Int. Conf. Robotics and Automation (ICRA). pp. 1098-1104 (2023)", + "5. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: CVPR. pp. 843-852 (2023)", + "6. Bekerman, Y., Avidan, S., Treibitz, T.: Unveiling optical properties in underwater images. In: ICCP (2020)", + "7. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE TPAMI 43(8), 2822-2837 (2020)", + "8. Chan, M.A., Young, S.I., Metzler, C.A.: Sud $^2$ : Supervision by denoising diffusion models for image reconstruction. arXiv preprint arXiv:2303.09642 (2023)", + "9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938 (2021)", + "0. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: CVPR (2023)", + "1. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGT0k", + "2. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. NeurIPS 35, 25683-25696 (2022)", + "3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022)", + "4. Chung, H., Ye, J.C.: Score-based diffusion models for accelerated mri. Medical image analysis 80, 102479 (2022)", + "5. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A.G., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. arXiv preprint arXiv:2305.19256 (2023)", + "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS 34, 8780-8794 (2021)", + "7. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: CVPR. pp. 9935-9946 (2023)" + ], + "bbox": [ + 225, + 280, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Feng, B.T., Smith, J., Rubinstein, M., Chang, H., Bouman, K.L., Freeman, W.T.: Score-based diffusion models as principled priors for inverse imaging. arXiv preprint arXiv:2304.11751 (2023)", + "19. Fu, Z., Lin, H., Yang, Y., Chai, S., Sun, L., Huang, Y., Ding, X.: Unsupervised underwater image restoration: From a homology perspective. In: Proc. AAAI Conf. on Artificial Intelligence. vol. 36, pp. 643-651 (2022)", + "20. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. Int. J. Robotics Research (IJRR) (2013)", + "21. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: NeurIPS (2022)", + "22. Guo, L., Wang, C., Yang, W., Huang, S., Wang, Y., Pfister, H., Wen, B.: Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In: CVPR. pp. 14049-14058 (2023)", + "23. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: ICIP. pp. 624-628 (2019)", + "24. Han, J., Shoeiby, M., Malthus, T., Botha, E., Anstee, J., Anwar, S., Wei, R., Armin, M.A., Li, H., Petersson, L.: Underwater image restoration via contrastive learning and a real-world dataset. Remote Sensing 14(17), 4297 (2022)", + "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)", + "26. Huang, S., Wang, K., Liu, H., Chen, J., Li, Y.: Contrastive semi-supervised learning for underwater image restoration via reliable bank. In: CVPR. pp. 18145-18155 (2023)", + "27. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters 5(2), 3227-3234 (2020)", + "28. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. NeurIPS 34, 14938-14954 (2021)", + "29. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. NeurIPS 35, 23593-23606 (2022)", + "30. Kawar, B., Elata, N., Michaeli, T., Elad, M.: Gsure-based diffusion model training with corrupted data. arXiv preprint arXiv:2305.13128 (2023)", + "31. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5148-5157 (October 2021)", + "32. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: CVPR. pp. 56-65 (2023)", + "33. Li, C., Anwar, S., Hou, J., Cong, R., Guo, C., Ren, W.: Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE TIP 30, 4985-5000 (2021)", + "34. Li, C., Anwar, S., Porikli, F.: Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition 98, 107038 (2020)", + "35. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE TIP 29, 4376-4389 (2019)", + "36. Liu, N., Zhang, N., Shao, L., Han, J.: Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI (2021)", + "37. Lu, S., Guan, F., Zhang, H., Lai, H.: Underwater image enhancement method based on denoising diffusion probabilistic model. J. of Visual Communication and Image Representation 96, 103926 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 410, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: GibbsDDRM: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. In: Int. Conf. on Machine Learning (2023)", + "39. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023)", + "40. Peng, L., Zhu, C., Bian, L.: U-shape transformer for underwater image enhancement. IEEE TIP (2023)", + "41. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE TIP 27(6), 2856-2868 (2018)", + "42. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE TIP 26(4), 1579-1594 (2017)", + "43. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015)", + "44. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR. pp. 22500-22510 (2023)", + "45. Saxena, S., Kar, A., Norouzi, M., Fleet, D.J.: Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816 (2023)", + "46. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760 (2012)", + "47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Int. conf. machine learning. pp. 2256-2265 (2015)", + "48. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983 (2023)", + "49. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: ICLR (2023), https://openreview.net/forum?id=9GsMA8MRKQ", + "50. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS. vol. 32 (2019)", + "51. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005 (2021)", + "52. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "53. Tang, Y., Kawasaki, H., Iwaguchi, T.: Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In: Proc. ACM Int. Conf. Multimedia. pp. 5419-5427 (2023)", + "54. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: ICCV. pp. 12248-12258 (2023)", + "55. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463", + "56. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "57. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR (June 2020)", + "58. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)", + "59. Yang, P., Zhou, S., Tao, Q., Loy, C.C.: PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In: NeurIPS (2023)", + "60. Zhang, W., Zhuang, P., Sun, H.H., Li, G., Kwong, S., Li, C.: Underwater image enhancement via minimal color loss and locally adaptive contrast enhancement. IEEE TIP 31, 3997-4010 (2022)", + "61. Zhou, J., Yang, T., Zhang, W.: Underwater vision enhancement technologies: a comprehensive review, challenges, and recent trends. Applied Intelligence 53(3), 3594-3621 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 327 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "O. Bar Nathan et al.", + "bbox": [ + 271, + 114, + 410, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_model.json b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa630995883fbb3ddff98062f201549c870aa8b5 --- /dev/null +++ b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_model.json @@ -0,0 +1,2436 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.777, + 0.187 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.212, + 0.723, + 0.243 + ], + "angle": 0, + "content": "Opher Bar Nathan\\(^{1}\\), Deborah Levy\\(^{1}\\), Tali Treibitz\\(^{1}\\), and Dan Rosenbaum\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.255, + 0.773, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Hatter Department of Marine Technologies, Charney School of Marine Sciences" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.27, + 0.621, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computer Science, University of Haifa, Haifa, Israel osmosis-diffusion.github.io" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.347, + 0.744, + 0.555 + ], + "angle": 0, + "content": "Abstract. Underwater image restoration is a challenging task because of water effects that increase dramatically with distance. This is worsened by lack of ground truth data of clean scenes without water. Diffusion priors have emerged as strong image restoration priors. However, they are often trained with a dataset of the desired restored output, which is not available in our case. We also observe that using only color data is insufficient, and therefore augment the prior with a depth channel. We train an unconditional diffusion model prior on the joint space of color and depth, using standard RGBD datasets of natural outdoor scenes in air. Using this prior together with a novel guidance method based on the underwater image formation model, we generate posterior samples of clean images, removing the water effects. Even though our prior did not see any underwater images during training, our method outperforms state-of-the-art baselines for image restoration on very challenging scenes. Our code, models and data are available on the project's website." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.568, + 0.741, + 0.596 + ], + "angle": 0, + "content": "Keywords: Diffusion Models \\(\\cdot\\) Physics-Based Computer Vision \\(\\cdot\\) Underwater Image Restoration" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.377, + 0.637 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.758 + ], + "angle": 0, + "content": "Underwater images are used in many applications, such as, underwater construction and maintenance, marine sciences, and fisheries. However, their automatic analysis is hindered because of the optical effects of the water that strongly attenuates and scatters light in a wavelength dependent manner. This causes color distortion and loss of contrast that exponentially increase with depth3. With growing human activity in the oceans, clear underwater vision becomes increasingly important." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Restoring underwater scenes is still a very challenging ill-posed problem. Classic approaches are based on designing priors for clean images or inverting the water formation model, and are limited by the ability to form strong priors. On" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3 For consistency with computer vision literature the term depth is used throughout to refer to the distance from the camera rather than water depth." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.785, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 1: Our method receives as an input a single underwater image and outputs the restored clean image and an estimated depth map. The output is estimated using a diffusion prior trained on RGBD images and the physical image formation model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.785, + 0.422 + ], + "angle": 0, + "content": "the other hand, learning based approaches are limited by the lack of supervised training data of clean underwater images. This is a critical issue, as the ocean cannot be emptied for the sake of data collection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.543 + ], + "angle": 0, + "content": "We suggest an unsupervised restoration method (Figs. 1,2) based on an inverse problem approach using a diffusion prior for both color and depth, coupled with the underwater image formation model. Restoring an image is formulated as posterior inference, computed using a natural image prior, and a likelihood term that is based on the underwater image formation model. The challenge with applying this approach for underwater image restoration directly is that (i) the degradation for each pixel depends on its depth, and other unknown parameters; and (ii) there is no ground truth clean underwater data to train the prior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.785, + 0.633 + ], + "angle": 0, + "content": "To solve this, we replace the image prior with a prior on the joint space of color and depth of natural images. Adding depth to the prior allows us to formulate the forward model that forms the likelihood term over the observed corrupted underwater image, and apply posterior sampling. Moreover, this leverages the high capacity of diffusion models, in capturing the strong correlations between color and depth in natural scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.785, + 0.739 + ], + "angle": 0, + "content": "We propose to train a prior model of RGBD images using available datasets of natural outdoor scenes that were collected in air. Using in-air scenes for underwater images may be counter-intuitive, but actually has strong benefits: 1) it overcomes the lack of clean underwater image data; 2) it leads to a strong prior that captures the joint statistics governing color and depth in natural scenes, where, as opposed to underwater images, color does not fade with distance; 3) it prevents overfitting to specific types of underwater images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.787, + 0.8 + ], + "angle": 0, + "content": "We then use this prior together with the underwater physical image formation model to simultaneously estimate the clean image, its depth, and the model parameters, all from a single underwater image (summarized in Fig. 2). We show that our method outperforms models that were trained on underwater data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.142, + 0.778, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.474 + ], + "angle": 0, + "content": "Fig. 2: The iterative sampling process starts in \\( t = T \\) with random noise in 4 channels. The denoising step outputs denoised samples \\( \\hat{x}_0 = (\\hat{J}_0, \\hat{D}_0) \\). We use the underwater physical image formation model together with \\( \\hat{x}_0 \\) to optimize the water parameters \\( \\hat{\\phi} \\), and to guide the sampling towards the observed image. This process repeats itself, gradually updating both the estimated image and depth, until \\( t = 0 \\), in which \\( x_0 \\) holds the method's estimate for both the reconstructed scene \\( J_0 \\) and its depth \\( D_0 \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.562, + 0.519 + ], + "angle": 0, + "content": "The main contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.785, + 0.564 + ], + "angle": 0, + "content": "1. We train an RGBD prior, and demonstrate that modeling color and depth and jointly sampling them, provides a stronger diffusion prior for underwater image restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.785, + 0.61 + ], + "angle": 0, + "content": "2. We propose a new method that combines the RGBD prior of in-air data with the underwater image formation model, leading to a diffusion guidance method that generates the restored underwater images as posterior samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.653 + ], + "angle": 0, + "content": "3. We demonstrate that our method outperforms state-of-the-art underwater image restoration methods both qualitatively and quantitatively on real and simulated data." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.52, + 0.785, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.655, + 0.75, + 0.671 + ], + "angle": 0, + "content": "We publish all our code, the new trained RGBD prior, images and results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.396, + 0.709 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.707, + 0.74 + ], + "angle": 0, + "content": "2.1 Underwater Image Restoration and Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Here we review recent works that are most relevant to our method. See [61] for a recent more comprehensive review." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Classic underwater image restoration. To cope with the ill-posed nature of the problem, earlier works introduced tailored image priors. Some priors aim to estimate a depth or transmission map of the scene to reduce the number of the unknowns and then use the estimated depth to restore the scenes [7, 41, 42]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Supervised learning. Clean underwater image data is very scarce. Several datasets aim to mitigate this issue. In UIEB [35], the images denoted as ground-truth are generated from enhancement results of 9 different baseline methods, and having human observers vote for the best one. The LSUI dataset [40] is larger, produced using the same methodology, based on choosing results from 18 different methods. These datasets enable supervised learning methods, but are still limited by their scale, the quality of the baseline methods, and the human bias to choose visually pleasing rather than physically consistent images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.419 + ], + "angle": 0, + "content": "CWR [24] introduced the HICRD dataset, where the images are restored using optical parameters impressively measured using ocean optics instruments. Unfortunately, the images are acquired in a downward-looking position, and thus their depth range is very limited. In FUnIE-GAN [27] a training dataset is generated by having humans select good images from a large set of unprocessed underwater images. The humans are instructed to choose images where the foreground objects are identifiable. These are then distorted by a GAN to produce the paired poor images. A synthetic dataset was generated in [34]. The dataset is synthesized from the NYU-v2 RGBD dataset [46], using the image formation model equation and several sets of values for the water parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Unsupervised learning. USUIR [19] aims to restore images without supervision, by separately estimating the image components (clean image, transmission, backscatter) and using them to construct an underwater image that is used for supervision against the original one. Subsequent frames were used in [4] for self-supervising monocular depth estimation, and in [54] for self-supervising both depth and restoration. In UW-NET [23] a cycle-GAN is used for learning mapping from RGBD in-air datasets to underwater images. As opposed to all these methods, we present the first prior based on diffusion models that does not rely on ground truth underwater supervision, and uses the physical model for inference." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.445, + 0.605 + ], + "angle": 0, + "content": "2.2 Diffusion Model Prior" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models have emerged as a powerful type of generative models. In the last few years several formulations and variations have been developed [25,47,50,52], most of which use a U-Net architecture [43] as a noise predictor. Because the training relies on very large datasets and is extremely time-consuming, significant work has been devoted to the setup where models are first pre-trained on large datasets and only later fine-tuned to more specific data, closer to the tasks at hand [44, 48]. In DepthGen [45] this approach is taken forward, by using a pre-trained model to kickstart a new model trained on different modalities using a different architecture. This is done by replacing the input and output layers of the pre-trained U-Net. We take a similar approach for training our RGBD prior, by using a pretrained diffusion model that was trained on RGB only." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Conditional diffusion models have been used for various image restoration tasks, by training models using different levels of supervision. Examples of similar tasks as ours include dehazing and deraining [8,39,56], and shadow removal [22]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Conditional diffusion models for underwater image enhancement have also been proposed [37,53]. Our approach differs from these by focusing on image restoration that inverts the physical model, rather than relying on supervised data that is optimized for image appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Diffusion model as a prior for clean images. In addition to conditional generation of images, there is a growing body of research where diffusion models are being used as clean image priors, and image restoration is formulated as posterior sampling [5, 9, 11-14, 18, 21, 28, 29, 49, 51]. These include tasks like denoising, inpainting, deblurring, and more general tasks. The limited access to clean data in many cases, has led to research on training a prior of clean images, using noisy training data only. In [1, 15, 30] a diffusion model prior is trained using noisy data, assuming a known degradation model. Since this setup is not directly applicable in our case, we chose instead to train our prior on clean data that was not taken underwater." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Diffusion model prior for blind image restoration. A more challenging task, is to use diffusion models as a clean image prior when the degradation model is unknown or depends on unknown parameters. Several papers have proposed to tackle this problem in different setups. In [17,38,59] the unknown parameters are optimized during the sampling process with a reconstruction objective function. In [10] a separate prior is trained for the unknown parameters, and sampled in parallel with image sampling. Our method differs from the above in that we learn a prior of the main unknown aspect of the degradation model, namely the depth, together with the prior of the variables we want to infer - the image color. This is done by training a single diffusion model on the joint space of color and depth." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.382, + 0.563 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.58, + 0.507, + 0.594 + ], + "angle": 0, + "content": "3.1 Underwater Image Formation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.785, + 0.693 + ], + "angle": 0, + "content": "In water, we observe two wavelength- and distance-dependent effects. First, the direct signal reflected from the object is attenuated. Second, light is scattered onto the object's line-of-sight (LOS), creating an additive signal termed backscatter that increases with distance. The occluding backscatter layer is independent of the scene content. Thus, the visibility and contrast of further objects is significantly reduced and their colors are distorted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.784, + 0.726 + ], + "angle": 0, + "content": "Following the revised underwater image formation model [2,32], under ambient illumination image intensity (per pixel, per color channel) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.737, + 0.785, + 0.755 + ], + "angle": 0, + "content": "\\[\nI = J \\cdot e ^ {- \\phi_ {a} \\cdot D} + \\phi^ {\\infty} \\cdot \\left(1 - e ^ {- \\phi_ {b} \\cdot D}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(I\\) is the linear image captured by the camera of a scene with range \\(D\\), \\(J\\) is the clear scene that would have been captured had there been no water along the LOS, and \\(\\phi^{\\infty}\\) is the water color at infinity, i.e., the backscatter at areas that contain no objects. The two parameters \\(\\phi_{a}\\) and \\(\\phi_{b}\\) are the attenuation and backscatter coefficients, respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "O.Bar Nathan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.403, + 0.16 + ], + "angle": 0, + "content": "3.2 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.262 + ], + "angle": 0, + "content": "Diffusion models have proven to be very effective in capturing the distribution of natural images, given in training data. We describe here the formulation that we use in our work, adopted from [16]. A diffusion model is defined by a Markov chain designed to transform the distribution of real images \\( x_0 \\) to a Gaussian distribution \\( x_{T} \\), by gradually scaling down the image values and adding Gaussian noise, using a schedule determined by the scalar parameters \\( \\alpha_{t} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.273, + 0.786, + 0.29 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.299, + 0.789, + 0.377 + ], + "angle": 0, + "content": "Based on this forward process, an inverse process is trained to gradually denoise images starting from a Gaussian distribution, back into the distribution of the original image dataset. This can be formulated as a factorization of the joint distribution over the images in reverse order, \\( p(x_{0:T}) = \\prod_{t=1}^{T} p_t(x_{t-1} \\mid x_t) \\), and \\( p_t(x_{t-1} \\mid x_t) \\) is approximated by a Gaussian" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.386, + 0.786, + 0.403 + ], + "angle": 0, + "content": "\\[\np _ {t} \\left(x _ {t - 1} \\mid x _ {t}\\right) \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) ， \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.412, + 0.789, + 0.473 + ], + "angle": 0, + "content": "with parameters \\(\\mu_{\\theta}\\) and \\(\\Sigma_{\\theta}\\) predicted by a trained neural network conditioned on \\(x_{t}\\) and \\(t\\). Simulating this process results in samples form the approximated data distribution \\(p(x_0)\\). A popular implementation first predicts the noise at each time step, and then computes the mean by" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.483, + 0.787, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\theta} (x _ {t}, t) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.526, + 0.788, + 0.588 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(x_t,t)\\) is the neural network trained to approximate \\(\\epsilon\\), and \\(\\bar{\\alpha}_t = \\prod_{s=0}^t \\alpha_s\\). The same neural network can also be used to predict a diagonal covariance \\(\\Sigma_{\\theta}(x_t,t)\\). At any iteration, an estimate of the clean image \\(x_0\\) can be derived by computing the mean of \\(p(x_0 \\mid x_t)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.598, + 0.787, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {0} (x _ {t}, t) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t}, t)\\right). \\qquad \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.637, + 0.788, + 0.684 + ], + "angle": 0, + "content": "The above formulation can also be developed from a score-based modeling approach, where it can be shown that \\(\\epsilon_{\\theta}(x_t,t)\\) is an approximation to the score function \\(\\nabla_{x_t}\\log p_t(x_t)\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.421, + 0.719 + ], + "angle": 0, + "content": "3.3 Posterior Sampling" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.728, + 0.787, + 0.773 + ], + "angle": 0, + "content": "Under the score-based view, given an observation \\(y\\) and a likelihood function \\(p(y|x)\\), we can use the same mechanism to sample from the posterior \\(p(x|y)\\) using the posterior score function" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.785, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {x _ {t}} \\log p _ {t} (x _ {t} | y) = \\nabla_ {x _ {t}} \\log p _ {t} (x _ {t}) + \\nabla_ {x _ {t}} \\log p _ {t} (y | x _ {t}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This idea of adding a conditional signal to the score is also called guidance. The challenge with the second term in Eq. 6, is that usually we are given a likelihood" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "model based on the clean image \\( x_0 \\), and not an intermediary noisy image \\( x_t \\). In our case the underwater image formation model transforms a clean image to an underwater observation. Connecting the model to the noisy sample \\( x_t \\) leads to the intractable integral \\( p_t(y|x_t) = \\int p(y|x_0)p(x_0|x_t)dx_0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.299 + ], + "angle": 0, + "content": "The various works on posterior sampling propose different approximations of this integral. Some propose to collapse the uncertainty in \\( y \\) [9, 12-14, 21, 51]. In [18] a variational inference approximation is proposed, and in [5, 11, 28, 29, 49] the likelihood model is computed on either \\( x_{t} \\) directly, or on the mean \\( \\hat{x}_0 = \\mathbb{E}[x_0|x_t] \\), as given in Eq. 5. After some experimentation, we decided to use the latter approximation as formulated in DPS [11]:" + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.31, + 0.785, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {x _ {t}} \\log p _ {t} (x _ {t} \\mid y) \\approx \\nabla_ {x _ {t}} \\log p _ {t} (x _ {t}) + s \\nabla_ {x _ {t}} \\log p (y \\mid \\hat {x} _ {0} (x _ {t}, t)), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.366 + ], + "angle": 0, + "content": "where \\( s \\) is the guidance scale used to control the weight of the approximated likelihood term." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.39, + 0.331, + 0.406 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.572 + ], + "angle": 0, + "content": "We use the underwater image formation model (Eq. 1) as a forward model that maps the space of natural images \\( x \\) to the space of underwater observed images \\( y \\). We aim to use this forward model to construct a likelihood term \\( p(y \\mid x) \\), and use it to sample from the posterior distribution of images (Eq. 7). In our case this cannot be implemented directly since the image formation model contains unknown parameters, the depth \\( D \\) at each pixel, and the water parameters, \\( \\phi_{a}, \\phi_{b}, \\phi^{\\infty} \\), which can differ between scenes. One way to deal with the unknown depth, is to use a monocular depth estimator in a separate first stage, and then use it as part of the forward model. In the results we discuss this approach and show that it is suboptimal (method variant termed DA-Osmosis)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.785, + 0.678 + ], + "angle": 0, + "content": "Another approach is to optimize the unknown parameters, including the depth, during sampling. In some recent work, different methods to do this were proposed [17, 38, 59]. However, these do not work properly when the unknown parameter is high dimensional, highly complex, and strongly correlated with the target image. In our early experimentation we found that in order to separate the image into a clean image and the effects governed by the depth, we need more than a good image prior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.785, + 0.77 + ], + "angle": 0, + "content": "Therefore, one of our main contributions, is to train a joint prior on both the color and depth in clean images, and show how it can be used for underwater image restoration. Following this we define \\( x_0 = (J, D) \\), where \\( J \\) represents a clear image and consists of the 3 color channels, and \\( D \\) is the depth image. This allows us to use the following forward model (based on Eq. 1) conditioned on the estimated clean image and depth \\( \\hat{x}_0 = (\\hat{J}_0, \\hat{D}_0) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.781, + 0.785, + 0.8 + ], + "angle": 0, + "content": "\\[\nf _ {\\phi} \\left(\\hat {x} _ {0}\\right) = \\hat {J} _ {0} e ^ {- \\phi_ {a} \\hat {D} _ {0}} + \\phi^ {\\infty} \\left(1 - e ^ {- \\phi_ {b} \\hat {D} _ {0}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\phi = (\\phi_{a},\\phi_{b},\\phi^{\\infty})\\) are the remaining 9 unknown water parameters (a parameter per color channel for each of \\(\\phi = (\\phi_a,\\phi_b,\\phi^\\infty)\\)). We optimize \\(\\phi\\) during" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 3: [Left] Example images from outdoor RGBD datasets used for training our prior. From left to right: DIODE [55], ReDWeb-S [36], HR-WSI [57], KITTI [20]. [Right] Samples from the trained RGBD prior. The samples demonstrate the inherent correlation between RGB image and depth in our trained RGBD prior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.784, + 0.392 + ], + "angle": 0, + "content": "sampling, in a similar way to GDP [17], using gradient descent on the image reconstruction loss. The likelihood term used for guidance is defined as a Gaussian around \\( f_{\\phi}(\\hat{x}_0) \\) with a fixed variance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.44 + ], + "angle": 0, + "content": "Training a joint prior on color and depth, not only allows us to use the above forward model, but also exploits the diffusion model capacity to capture the complexity of the depth image and its correlations with the clean color image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.414, + 0.485 + ], + "angle": 0, + "content": "4.1 Training the Prior" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.595 + ], + "angle": 0, + "content": "To learn a natural image prior, we train the joint prior model on both color and depth using public RGBD datasets of outdoor scenes that were taken in air. While differing from underwater images, the ability to use large amounts of high quality training data leads to a prior that captures the correlation between color and depth in natural scenes, which as we show, is an important aspect for underwater image restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.785, + 0.703 + ], + "angle": 0, + "content": "For the sake of data efficiency and training time, we start with a pretrained diffusion model trained on ImageNet (we take the unconditional model from [16]), and finetune it on RGBD data. We implement this, inspired by DepthGen [45], by replacing the first and last layers of the U-Net to 4 channels rather than 3, and initializing those layers randomly. The datasets we use for finetuning are (see Fig. 3[left]): DIODE [55], RedWeb-S [36], HR-WSI [57], KITTI [20], with 16884, 2179, 20378, 23946 pictures, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "A major challenge in working with RGBD data is to turn the available depth information into a proper image. This includes filling holes and scaling the values to a standard range. Since each of the above datasets were collected in a different manner, we treat each of them differently. This is detailed in the supplementary material. Fig. 3[right] shows two samples from the prior that show corresponding color and depth images. While prior samples demonstrate an evident domain gap between in-air and underwater data, our results show that posterior samples are not affected by the gap, and can leverage the strong correlation between depth and color in natural scenes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.498, + 0.163 + ], + "angle": 0, + "content": "4.2 Sampling from the Posterior" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.326 + ], + "angle": 0, + "content": "Given a trained prior, we perform underwater image restoration by sampling from the posterior with guidance of the image formation model. The sampling process is described in Figs. 2, 4. In each iteration we generate samples of both the image and the depth, while optimizing the water parameters. This results in a gradual update to the estimates of the clean image, and the depth, demonstrated in Fig. 4[right]. We adopt the sampling method in DPS [11], using the reconstruction loss, \\(\\left\\| y - f_{\\phi}(\\hat{x}_0)\\right\\|_2^2\\), which is the negative log-likelihood formed from the model in Eq. 8. Similar to GDP [17], the remaining unknown parameters \\(\\phi\\) are optimized in parallel to the inverse sampling process, using gradient descent on the same loss used for guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Note that one of our main novelties is that the RGBD prior is used as an inherent part of the iterative method and not as a stand-alone depth estimator. The depth is estimated together with the image in every iteration, guided by the underwater model, see, e.g., Fig. 4[right]. As demonstrated in the results, this improves the restoration quality over using a fixed estimated depth of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Running posterior sampling, guided by a reconstruction loss only, we observe two problems. First, for pixels with large depth values, the reconstructed image is dominated by the backscatter, making the estimation of the clean pixel color values unstable. Second, the color values can shift outside the valid range, causing color saturation. To overcome the first issue we multiply the reconstruction loss in every pixel by the estimated depth value (without passing the gradients of this operation, denoted by \\( sg \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.522, + 0.785, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\left\\| s g (\\hat {D} _ {0}) \\cdot \\left(y - f _ {\\phi} (\\hat {x} _ {0})\\right) \\right\\| _ {2} ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.613 + ], + "angle": 0, + "content": "In order to overcome the second issue, we introduce two auxiliary losses. The first penalizes RGB values outside the valid range \\([-1, 1]\\), and the second encourages the average values of each channel to approach the middle of the color range (in line with the gray world assumption). We implement the two auxiliary losses as:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.625, + 0.76, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v a l}} = \\lambda_ {v} \\sum_ {i, c} \\max \\left(\\left| \\hat {J} _ {0} (i, c) \\right| - T _ {v}, 0\\right) ^ {2}, \\quad \\mathcal {L} _ {\\mathrm {a v r g}} = \\lambda_ {a} \\sum_ {c} \\left| \\sum_ {i} \\hat {J} _ {0} (i, c) - T _ {a} \\right|\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.726 + ], + "angle": 0, + "content": "where \\(\\hat{J}_0(i,c)\\) is the value of color channel \\(c\\) of pixel \\(i\\), assuming the valid color range is \\([-1,1]\\), \\(T_v\\) and \\(T_a\\) are thresholds set to values close to 1 and 0 respectively, and \\(\\lambda_v,\\lambda_a\\) are the scalar weights of both losses. The total loss is then:" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.743, + 0.785, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {v a l}} + \\mathcal {L} _ {\\text {a v r g}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "At each iteration we compute the gradient of the loss with respect to \\( x_{t} \\), which forms the log-likelihood gradient in the posterior score (Eq. 7), and the gradient with respect to \\( \\phi \\) which is used to update the parameters in the underwater model. In order to stabilize the optimization, we apply gradient clipping by value, to the gradients of \\( x_{t} \\), and we run the optimization of \\( \\phi \\) only in some of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.127 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.235, + 0.147, + 0.494, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1: Osmosis - Sampling" + }, + { + "type": "algorithm", + "bbox": [ + 0.227, + 0.165, + 0.61, + 0.479 + ], + "angle": 0, + "content": "Input: y - Underwater Degraded Image (RGB image) \\(f(\\hat{x}_0)\\) - Underwater Physical Degradation Model Initialization of \\(\\phi = \\{\\phi_a,\\phi_b,\\phi^\\infty \\}\\) // size = [9] Output: \\(\\hat{x}_0 = [\\hat{J}_0,\\hat{D}_0]\\) .. \\(\\hat{J}_0\\) - Restored RGB Image, \\(\\hat{D}_0\\) - Depth Estimation \n1 \\(x_{T}\\sim \\mathcal{N}(0,I)\\) // \\(x_{T}\\) size = [h,w,4] \n2 for \\(t = T:1\\) do \n3 \\(\\epsilon \\gets \\epsilon_{\\theta}(x_t,t)\\) \n4 \\(\\mu = \\frac{1}{\\sqrt{\\alpha_t}} (x_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\cdot \\epsilon)\\) \n5 \\(\\hat{x}_0 = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}} (x_t - \\sqrt{1 - \\bar{\\alpha}_t}\\cdot \\epsilon)\\) \n6 \\(f_{\\phi}(\\hat{x}_0) = \\hat{J}_0e^{-\\phi_a\\hat{D}_0} + \\phi^{\\infty}\\left(1 - e^{-\\phi_b\\hat{D}_0}\\right)\\) \n7 \\(\\mathcal{L} =\\) loss from Eq. 10 \n8 if Optimstart \\(\\geq t\\geq\\) Optimend then \n9 for \\(i = 1:N\\) do \n10 | \\(\\phi \\gets \\phi -\\eta \\nabla_{\\phi}\\mathcal{L}\\) \n11 end \n12 end \n13 \\(x_{t - 1}\\sim \\mathcal{N}(\\mu -s\\cdot \\nabla_{x_t}\\mathcal{L},\\Sigma)\\) \n14 end" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.145, + 0.784, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 4: Our algorithm. [Left] Detailed steps of our algorithm. [Right] Example of how \\(\\hat{J}_0, \\hat{D}_0\\) change during the iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.673 + ], + "angle": 0, + "content": "the sampling steps (defined by the values \\(\\mathrm{Optim}_{\\mathrm{start}}\\) and \\(\\mathrm{Optim}_{\\mathrm{end}}\\)), running \\(N\\) iterations of gradient descent in each step. When the gradient of \\(x_{t}\\) is applied in the sampling step, it is multiplied by a guidance scale \\(s\\). We find that using a smaller scale for the depth channel leads to better results. This makes sense as while all channels are treated the same in the prior, the depth has a different role in the forward model (specifically it is used inside an exponent), and this can lead to a different gradient scale. More details on the implementation are given in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.325, + 0.71 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.787, + 0.773 + ], + "angle": 0, + "content": "We use a prior on \\(256 \\times 256\\) images, with sampling time of about 3 minutes per image on an Nvidia A100 GPU. We present here selected results and analysis. Please refer to the supplementary material for a complete set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Real-World Scenes. We present an extensive comparison on real-world linear images from the datasets SQUID [7], SeaThru [3], SeaThruNerf [32], and additional images acquired in different locations in the world, the Indian and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "Pacific Oceans, and Mediterranean, and Red Seas. Images are white-balanced as pre-processing. All runs use the same set of parameters. Fig. 1 shows several results of both image restoration and depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.358 + ], + "angle": 0, + "content": "We compare image restoration with the following methods: CWR [24], DM [53], FUnIE-GAN [27], GDCP [41], IBLA [42], MMLE [60], semi-UIR [26], Ucolor [33], USUIR [19], UW-Net [23], waternet [35], and USe-ReDI-Net [54]. For compactness, Fig. 5 summarizes the results for a chosen subset of methods. The complete comparison is shown in the supplementary material. Among previous methods, we found GDCP to be the most consistent. Our result recovers the full range of the scenes, specifically improving contrast of objects that are further away (note the zoomed-in objects in the far areas, e.g., the diver in row 1). Our method recovers vibrant and consistent colors also in further areas. For example, in rows 2,3,4 note that the background rocks and sand appear bluish in all methods except of ours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.787, + 0.45 + ], + "angle": 0, + "content": "We also compare to a method that restores the RGB using a fixed depth that is pre-estimated using an external SOTA method, depth anything (DA) [58]. We use the pre-estimated depth in our method's pipeline instead of our gradually estimated depth. We term this method \\( DA \\)-Osmosis, and the results show that our method restores the far areas with better color, emphasizing the advantage of using our joint RGBD prior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.787, + 0.512 + ], + "angle": 0, + "content": "We compare depth estimation with GDCP [41], IBLA [42], unveiling [6], UW-Net [23], and monoUWnet [4], and depth anything (DA) [58]. In GDCP [41] and unveiling [6] we compute the depth from the output transmission maps. Our depth estimations have more details and better explain the scenes (Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.787, + 0.572 + ], + "angle": 0, + "content": "In addition, we conducted a non-reference quantitative comparison using the MUSIQ [31] measure on 50 real-world images presented in the paper and supplementary material. Our method achieved the highest score. The quantitative results are summarized in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.787, + 0.68 + ], + "angle": 0, + "content": "The UIEB dataset [35] is sometimes used for quantitative evaluation of image restoration. However, it is collected from various sources and is not linear. Thus it is not suitable for physics-based methods like ours. In addition, the 'ground truth' of the UIEB dataset is chosen from results of previous methods and it is not the real ground truth. Fig. 6 shows four examples (out of many) where our result removes more water effects than the ground truth, even despite the non-linearity of the input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.787, + 0.765 + ], + "angle": 0, + "content": "Simulation. We conducted a simulation following [26, 34] using all the 1449 images from the NYU-v2 RGBD dataset [46] with randomly varying water parameters and generated the underwater images with the image formation model in Eq. 1. For fairness in comparison with other methods we applied \\(\\phi_{a} = \\phi_{b}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The quantitative results are summarized in Table 2. Our method substantially outperforms other methods in the image restoration metrics of PSNR, SSIM, and LPIPS. We emphasize that our prior was not trained on this data, or indoor data at all. Moreover, two of the methods [26,34] we outperform were trained on this data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.171, + 0.785, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Fig. 5: Real-world restoration results. From left to right: white-balanced input, contrast stretch, GDCP [41], Ucolor [33], CWR [24], semi-UIR [26], DM [53], Depth Anything [58] - Osmosis, Osmosis (ours). Zoom-in colored rectangles emphasize far objects that have higher contrast in our results. Real-world depth results. From left to right: GDCP [41], IBLA [42], unveiling [6], UW-Net [23], monoUWnet [4], Depth Anything [58], Osmosis (ours). Our depth results are smoother and less affected by object gradients. The reader is encouraged to zoom-in." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.143, + 0.784, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.785, + 0.384 + ], + "angle": 0, + "content": "Fig. 6: Example results on UIEB [35], a dataset comprised from non-linear images with ground truth produced by different enhancement algorithms. Our method yields better results than the dataset's ground truth. See for example the color of the sand, the orange color of the fish in the top-left, and the divers' authentic skin color in the bottom-right." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.399, + 0.784, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Fig. 7: Ablation. Ablations \\(\\#1, \\#5, \\#6\\) show the importance of the losses we added to the reconstruction loss. Without \\(\\mathcal{L}_{\\mathrm{val}} (\\#6)\\), the colors tend to \"explode\" and oversaturate. Without \\(\\mathcal{L}_{\\mathrm{avg}} (\\#5)\\) the colors sometimes skew towards pink/purple. In \\(\\#2\\), the further areas are not restored well, because the loss is not weighted by the depth \\(D\\). In \\(\\#3\\) the guidance scale is the same for the RGB and depth channels. This harms the depth reconstruction. In \\(\\#4\\) we set \\(\\phi_{a} = \\phi_{b}\\) (separately per color channel). Since this is an inaccurate model, the restoration in further areas is harmed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablation. To demonstrate the effect of different components in our methods, we conduct an ablation study of the following variants: 1. \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}}\\) (instead of Eq. 10); 2. Removing weighting by \\(\\hat{D}_0\\) in Eq. 9 3. Same guidance scale \\(s\\) for all channels; 4. \\(\\phi_{a} = \\phi_{b}\\); 5. \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}} + \\mathcal{L}_{\\mathrm{val}}\\) (remove \\(\\mathcal{L}_{\\mathrm{avg}}\\) from Eq. 10); 6. \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}} + \\mathcal{L}_{\\mathrm{avg}}\\) (remove \\(\\mathcal{L}_{\\mathrm{val}}\\) from Eq. 10). Numerical results on the simulation for variants \\(\\# 1 - \\# 3\\) are presented in Table 2. Ablations \\(\\# 4 - \\# 6\\) are shown only on real world images since in the simulation we do not use \\(\\mathcal{L}_{\\mathrm{avg}}\\), and \\(\\phi_{a} = \\phi_{b}\\). Fig. 7 presents the results of all variants on one of the scenes. We see that the additional losses are important to prevent color saturation and shift. Increasing the loss weight with depth improves restoration in further areas. Separating guidance scales between the RGB and depth channels improves depth reconstruction. Setting \\(\\phi_{a} \\neq \\phi_{b}\\) extends the range of the restoration." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.127 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.151, + 0.44, + 0.389 + ], + "angle": 0, + "content": "
MethodMUSIQ↑
input51.59
contrast stretch53.75
CWR [24]39.63
DM [53]54.64
FUnIE-GAN [27]42.70
GDCP [41]56.25
IBLA [42]54.42
MMLE [60]56.26
semi-UIR [26]54.99
Ucolor [33]47.73
USUIR [19]52.44
UW-Net [23]46.02
waternet [35]53.64
unveiling [6]50.83
DA-osmosis [58]51.23
osmosis (ours)56.62
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.39, + 0.477, + 0.445 + ], + "angle": 0, + "content": "Table 1: Non-reference quantitative comparison on 50 real-world images using MUSIQ [31]. Our method achieves the highest score." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.144, + 0.803, + 0.397 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
contrast stretch17.130.830.11
CWR [24]16.930.790.20
DM [53]17.410.820.12
FUnIE-GAN [27]17.640.770.21
GDCP [41]12.410.710.16
IBLA [42]15.070.700.19
MMLE [60]17.000.740.17
semi-UIR [26]17.820.830.12
Ucolor [33]17.920.830.10
USUIR [19]16.760.800.18
UW-Net [23]18.040.750.26
waternet [35]17.270.820.11
unveiling [6]16.340.790.18
ablation #121.090.860.09
ablation #222.170.880.07
ablation #322.000.880.06
osmosis (ours)22.740.890.06
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.521, + 0.398, + 0.782, + 0.438 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on the simulation. Our method achieves best scores in image restoration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.487, + 0.355, + 0.503 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.703 + ], + "angle": 0, + "content": "In this work we demonstrated how to harness the strength of a new RGBD diffusion prior to achieve state-of-the-art results on underwater image restoration. To do this we solved several challenges: i) because of lack of clean underwater image data we use datasets of scenes in air; ii) we notice that the color prior does not suffice to guide restoration, therefore, we add the scene depth to the prior; iii) We use the physical image formation model to guide restoration, and also estimate the water parameters in the process. This results in the most comprehensive single underwater image restoration method to-date. It does not train on any underwater images and therefore does not overfit to any. It gives higher weight to further objects and therefore is superior in reconstructing all the details of the scene. Jointly solving for the depth results in excellent depth estimation from monocular images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.795 + ], + "angle": 0, + "content": "There is a domain gap between the prior (trained on in-air data) and underwater data. This is intentional. The beauty of our method is that by incorporating the underwater model in the sampling it succeeds even without training on underwater data (that is very hard to obtain). Thus, it is not specialized for a certain type of water or objects, and it learns the color distribution of natural images not degraded by water." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Like all diffusion models with U-Nets, our method is limited by a fixed resolution and long running time. This could potentially be improved using other network architectures and sampling methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.148, + 0.812, + 0.231 + ], + "angle": 0, + "content": "Acknowledgements. The research was funded by Israel Science Foundation grant #1951/23, Israeli Ministry of Science and Technology grants #1001577600 & #1001593851, EU Horizon 2020 research and innovation programme GA 101094924 (ANERIS), the Leona M. and Harry B. Helmsley Charitable Trust, and the Maurice Hatter Foundation. We thank Dr. Derya Akkaynak and Dr. Matan Yuval for substantial data contribution, Amir Dayan for the paper's name and Meirav Keidar for graphics design." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.252, + 0.324, + 0.269 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.786, + 0.322 + ], + "angle": 0, + "content": "1. Aali, A., Arvinte, M., Kumar, S., Tamir, J.I.: Solving inverse problems with score-based generative priors learned from noisy data. arXiv preprint arXiv:2305.01166 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.786, + 0.35 + ], + "angle": 0, + "content": "2. Akkaynak, D., Treibitz, T.: A revised underwater image formation model. In: CVPR. pp. 6723-6732 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.786, + 0.377 + ], + "angle": 0, + "content": "3. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: CVPR. pp. 1682-1691 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.786, + 0.404 + ], + "angle": 0, + "content": "4. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: IEEE Int. Conf. Robotics and Automation (ICRA). pp. 1098-1104 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.445 + ], + "angle": 0, + "content": "5. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: CVPR. pp. 843-852 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.786, + 0.472 + ], + "angle": 0, + "content": "6. Bekerman, Y., Avidan, S., Treibitz, T.: Unveiling optical properties in underwater images. In: ICCP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.786, + 0.513 + ], + "angle": 0, + "content": "7. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE TPAMI 43(8), 2822-2837 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.786, + 0.541 + ], + "angle": 0, + "content": "8. Chan, M.A., Young, S.I., Metzler, C.A.: Sud\\(^2\\): Supervision by denoising diffusion models for image reconstruction. arXiv preprint arXiv:2303.09642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.786, + 0.568 + ], + "angle": 0, + "content": "9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.786, + 0.595 + ], + "angle": 0, + "content": "0. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.786, + 0.635 + ], + "angle": 0, + "content": "1. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGT0k" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.786, + 0.663 + ], + "angle": 0, + "content": "2. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. NeurIPS 35, 25683-25696 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.704 + ], + "angle": 0, + "content": "3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.786, + 0.731 + ], + "angle": 0, + "content": "4. Chung, H., Ye, J.C.: Score-based diffusion models for accelerated mri. Medical image analysis 80, 102479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.786, + 0.772 + ], + "angle": 0, + "content": "5. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A.G., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. arXiv preprint arXiv:2305.19256 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.799 + ], + "angle": 0, + "content": "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "7. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: CVPR. pp. 9935-9946 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.281, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.127 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Feng, B.T., Smith, J., Rubinstein, M., Chang, H., Bouman, K.L., Freeman, W.T.: Score-based diffusion models as principled priors for inverse imaging. arXiv preprint arXiv:2304.11751 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "19. Fu, Z., Lin, H., Yang, Y., Chai, S., Sun, L., Huang, Y., Ding, X.: Unsupervised underwater image restoration: From a homology perspective. In: Proc. AAAI Conf. on Artificial Intelligence. vol. 36, pp. 643-651 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "20. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. Int. J. Robotics Research (IJRR) (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "21. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.325 + ], + "angle": 0, + "content": "22. Guo, L., Wang, C., Yang, W., Huang, S., Wang, Y., Pfister, H., Wen, B.: Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In: CVPR. pp. 14049-14058 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "23. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: ICIP. pp. 624-628 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "24. Han, J., Shoeiby, M., Malthus, T., Botha, E., Anstee, J., Anwar, S., Wei, R., Armin, M.A., Li, H., Petersson, L.: Underwater image restoration via contrastive learning and a real-world dataset. Remote Sensing 14(17), 4297 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "26. Huang, S., Wang, K., Liu, H., Chen, J., Li, Y.: Contrastive semi-supervised learning for underwater image restoration via reliable bank. In: CVPR. pp. 18145-18155 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "27. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters 5(2), 3227-3234 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "28. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. NeurIPS 34, 14938-14954 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.555 + ], + "angle": 0, + "content": "29. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. NeurIPS 35, 23593-23606 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.582 + ], + "angle": 0, + "content": "30. Kawar, B., Elata, N., Michaeli, T., Elad, M.: Gsure-based diffusion model training with corrupted data. arXiv preprint arXiv:2305.13128 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "31. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5148-5157 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "32. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: CVPR. pp. 56-65 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.704 + ], + "angle": 0, + "content": "33. Li, C., Anwar, S., Hou, J., Cong, R., Guo, C., Ren, W.: Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE TIP 30, 4985-5000 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.732 + ], + "angle": 0, + "content": "34. Li, C., Anwar, S., Porikli, F.: Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition 98, 107038 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "35. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE TIP 29, 4376-4389 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "36. Liu, N., Zhang, N., Shao, L., Han, J.: Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Lu, S., Guan, F., Zhang, H., Lai, H.: Underwater image enhancement method based on denoising diffusion probabilistic model. J. of Visual Communication and Image Representation 96, 103926 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "38. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: GibbsDDRM: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. In: Int. Conf. on Machine Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "39. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "40. Peng, L., Zhu, C., Bian, L.: U-shape transformer for underwater image enhancement. IEEE TIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "41. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE TIP 27(6), 2856-2868 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.297 + ], + "angle": 0, + "content": "42. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE TIP 26(4), 1579-1594 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.338 + ], + "angle": 0, + "content": "43. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "44. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "45. Saxena, S., Kar, A., Norouzi, M., Fleet, D.J.: Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.432 + ], + "angle": 0, + "content": "46. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.472 + ], + "angle": 0, + "content": "47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Int. conf. machine learning. pp. 2256-2265 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.472, + 0.785, + 0.514 + ], + "angle": 0, + "content": "48. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.554 + ], + "angle": 0, + "content": "49. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: ICLR (2023), https://openreview.net/forum?id=9GsMA8MRKQ" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.785, + 0.595 + ], + "angle": 0, + "content": "50. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS. vol. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.622 + ], + "angle": 0, + "content": "51. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.785, + 0.662 + ], + "angle": 0, + "content": "52. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.662, + 0.785, + 0.703 + ], + "angle": 0, + "content": "53. Tang, Y., Kawasaki, H., Iwaguchi, T.: Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In: Proc. ACM Int. Conf. Multimedia. pp. 5419-5427 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.744 + ], + "angle": 0, + "content": "54. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: ICCV. pp. 12248-12258 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "55. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "56. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.127 + ], + "angle": 0, + "content": "O. Bar Nathan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "57. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "58. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "59. Yang, P., Zhou, S., Tao, Q., Loy, C.C.: PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "60. Zhang, W., Zhuang, P., Sun, H.H., Li, G., Kwong, S., Li, C.: Underwater image enhancement via minimal color loss and locally adaptive contrast enhancement. IEEE TIP 31, 3997-4010 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "61. Zhou, J., Yang, T., Zhang, W.: Underwater vision enhancement technologies: a comprehensive review, challenges, and recent trends. Applied Intelligence 53(3), 3594-3621 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_origin.pdf b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..444e0a54b70dac90baa890dc9b92932aa2136841 --- /dev/null +++ b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/25e6c8d0-63b3-4857-86b1-de08c18761d6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34a4a5de4c17f9145d1bf89e8d0d52c49198df849c5b943969c2e6542940d43d +size 5958966 diff --git a/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/full.md b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f8fba911a6c4c3a4fc9a0438c23bf6c2bab93723 --- /dev/null +++ b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/full.md @@ -0,0 +1,324 @@ +# Osmosis: RGBD Diffusion Prior for Underwater Image Restoration + +Opher Bar Nathan $^{1}$ , Deborah Levy $^{1}$ , Tali Treibitz $^{1}$ , and Dan Rosenbaum $^{2}$ + +$^{1}$ Hatter Department of Marine Technologies, Charney School of Marine Sciences + +$^{2}$ Department of Computer Science, University of Haifa, Haifa, Israel osmosis-diffusion.github.io + +Abstract. Underwater image restoration is a challenging task because of water effects that increase dramatically with distance. This is worsened by lack of ground truth data of clean scenes without water. Diffusion priors have emerged as strong image restoration priors. However, they are often trained with a dataset of the desired restored output, which is not available in our case. We also observe that using only color data is insufficient, and therefore augment the prior with a depth channel. We train an unconditional diffusion model prior on the joint space of color and depth, using standard RGBD datasets of natural outdoor scenes in air. Using this prior together with a novel guidance method based on the underwater image formation model, we generate posterior samples of clean images, removing the water effects. Even though our prior did not see any underwater images during training, our method outperforms state-of-the-art baselines for image restoration on very challenging scenes. Our code, models and data are available on the project's website. + +Keywords: Diffusion Models $\cdot$ Physics-Based Computer Vision $\cdot$ Underwater Image Restoration + +# 1 Introduction + +Underwater images are used in many applications, such as, underwater construction and maintenance, marine sciences, and fisheries. However, their automatic analysis is hindered because of the optical effects of the water that strongly attenuates and scatters light in a wavelength dependent manner. This causes color distortion and loss of contrast that exponentially increase with depth3. With growing human activity in the oceans, clear underwater vision becomes increasingly important. + +Restoring underwater scenes is still a very challenging ill-posed problem. Classic approaches are based on designing priors for clean images or inverting the water formation model, and are limited by the ability to form strong priors. On + +![](images/4b4cb06e8a2c39b20c10752b8793299cde98dfbcf28485bfb197e4c2d0735bd3.jpg) +Fig. 1: Our method receives as an input a single underwater image and outputs the restored clean image and an estimated depth map. The output is estimated using a diffusion prior trained on RGBD images and the physical image formation model. + +the other hand, learning based approaches are limited by the lack of supervised training data of clean underwater images. This is a critical issue, as the ocean cannot be emptied for the sake of data collection. + +We suggest an unsupervised restoration method (Figs. 1,2) based on an inverse problem approach using a diffusion prior for both color and depth, coupled with the underwater image formation model. Restoring an image is formulated as posterior inference, computed using a natural image prior, and a likelihood term that is based on the underwater image formation model. The challenge with applying this approach for underwater image restoration directly is that (i) the degradation for each pixel depends on its depth, and other unknown parameters; and (ii) there is no ground truth clean underwater data to train the prior. + +To solve this, we replace the image prior with a prior on the joint space of color and depth of natural images. Adding depth to the prior allows us to formulate the forward model that forms the likelihood term over the observed corrupted underwater image, and apply posterior sampling. Moreover, this leverages the high capacity of diffusion models, in capturing the strong correlations between color and depth in natural scenes. + +We propose to train a prior model of RGBD images using available datasets of natural outdoor scenes that were collected in air. Using in-air scenes for underwater images may be counter-intuitive, but actually has strong benefits: 1) it overcomes the lack of clean underwater image data; 2) it leads to a strong prior that captures the joint statistics governing color and depth in natural scenes, where, as opposed to underwater images, color does not fade with distance; 3) it prevents overfitting to specific types of underwater images. + +We then use this prior together with the underwater physical image formation model to simultaneously estimate the clean image, its depth, and the model parameters, all from a single underwater image (summarized in Fig. 2). We show that our method outperforms models that were trained on underwater data. + +![](images/66431529cd214904f8ef4ab142f8c8ab8cff33156409f1e8e3279aedda391347.jpg) +Fig. 2: The iterative sampling process starts in $t = T$ with random noise in 4 channels. The denoising step outputs denoised samples $\hat{x}_0 = (\hat{J}_0, \hat{D}_0)$ . We use the underwater physical image formation model together with $\hat{x}_0$ to optimize the water parameters $\hat{\phi}$ , and to guide the sampling towards the observed image. This process repeats itself, gradually updating both the estimated image and depth, until $t = 0$ , in which $x_0$ holds the method's estimate for both the reconstructed scene $J_0$ and its depth $D_0$ . + +# The main contributions of this paper are: + +1. We train an RGBD prior, and demonstrate that modeling color and depth and jointly sampling them, provides a stronger diffusion prior for underwater image restoration. +2. We propose a new method that combines the RGBD prior of in-air data with the underwater image formation model, leading to a diffusion guidance method that generates the restored underwater images as posterior samples. +3. We demonstrate that our method outperforms state-of-the-art underwater image restoration methods both qualitatively and quantitatively on real and simulated data. + +We publish all our code, the new trained RGBD prior, images and results. + +# 2 Related Works + +# 2.1 Underwater Image Restoration and Depth Estimation + +Here we review recent works that are most relevant to our method. See [61] for a recent more comprehensive review. + +Classic underwater image restoration. To cope with the ill-posed nature of the problem, earlier works introduced tailored image priors. Some priors aim to estimate a depth or transmission map of the scene to reduce the number of the unknowns and then use the estimated depth to restore the scenes [7, 41, 42]. + +Supervised learning. Clean underwater image data is very scarce. Several datasets aim to mitigate this issue. In UIEB [35], the images denoted as ground-truth are generated from enhancement results of 9 different baseline methods, and having human observers vote for the best one. The LSUI dataset [40] is larger, produced using the same methodology, based on choosing results from 18 different methods. These datasets enable supervised learning methods, but are still limited by their scale, the quality of the baseline methods, and the human bias to choose visually pleasing rather than physically consistent images. + +CWR [24] introduced the HICRD dataset, where the images are restored using optical parameters impressively measured using ocean optics instruments. Unfortunately, the images are acquired in a downward-looking position, and thus their depth range is very limited. In FUnIE-GAN [27] a training dataset is generated by having humans select good images from a large set of unprocessed underwater images. The humans are instructed to choose images where the foreground objects are identifiable. These are then distorted by a GAN to produce the paired poor images. A synthetic dataset was generated in [34]. The dataset is synthesized from the NYU-v2 RGBD dataset [46], using the image formation model equation and several sets of values for the water parameters. + +Unsupervised learning. USUIR [19] aims to restore images without supervision, by separately estimating the image components (clean image, transmission, backscatter) and using them to construct an underwater image that is used for supervision against the original one. Subsequent frames were used in [4] for self-supervising monocular depth estimation, and in [54] for self-supervising both depth and restoration. In UW-NET [23] a cycle-GAN is used for learning mapping from RGBD in-air datasets to underwater images. As opposed to all these methods, we present the first prior based on diffusion models that does not rely on ground truth underwater supervision, and uses the physical model for inference. + +# 2.2 Diffusion Model Prior + +Diffusion models. Diffusion models have emerged as a powerful type of generative models. In the last few years several formulations and variations have been developed [25,47,50,52], most of which use a U-Net architecture [43] as a noise predictor. Because the training relies on very large datasets and is extremely time-consuming, significant work has been devoted to the setup where models are first pre-trained on large datasets and only later fine-tuned to more specific data, closer to the tasks at hand [44, 48]. In DepthGen [45] this approach is taken forward, by using a pre-trained model to kickstart a new model trained on different modalities using a different architecture. This is done by replacing the input and output layers of the pre-trained U-Net. We take a similar approach for training our RGBD prior, by using a pretrained diffusion model that was trained on RGB only. + +Conditional diffusion models have been used for various image restoration tasks, by training models using different levels of supervision. Examples of similar tasks as ours include dehazing and deraining [8,39,56], and shadow removal [22]. + +Conditional diffusion models for underwater image enhancement have also been proposed [37,53]. Our approach differs from these by focusing on image restoration that inverts the physical model, rather than relying on supervised data that is optimized for image appearance. + +Diffusion model as a prior for clean images. In addition to conditional generation of images, there is a growing body of research where diffusion models are being used as clean image priors, and image restoration is formulated as posterior sampling [5, 9, 11-14, 18, 21, 28, 29, 49, 51]. These include tasks like denoising, inpainting, deblurring, and more general tasks. The limited access to clean data in many cases, has led to research on training a prior of clean images, using noisy training data only. In [1, 15, 30] a diffusion model prior is trained using noisy data, assuming a known degradation model. Since this setup is not directly applicable in our case, we chose instead to train our prior on clean data that was not taken underwater. + +Diffusion model prior for blind image restoration. A more challenging task, is to use diffusion models as a clean image prior when the degradation model is unknown or depends on unknown parameters. Several papers have proposed to tackle this problem in different setups. In [17,38,59] the unknown parameters are optimized during the sampling process with a reconstruction objective function. In [10] a separate prior is trained for the unknown parameters, and sampled in parallel with image sampling. Our method differs from the above in that we learn a prior of the main unknown aspect of the degradation model, namely the depth, together with the prior of the variables we want to infer - the image color. This is done by training a single diffusion model on the joint space of color and depth. + +# 3 Preliminaries + +# 3.1 Underwater Image Formation + +In water, we observe two wavelength- and distance-dependent effects. First, the direct signal reflected from the object is attenuated. Second, light is scattered onto the object's line-of-sight (LOS), creating an additive signal termed backscatter that increases with distance. The occluding backscatter layer is independent of the scene content. Thus, the visibility and contrast of further objects is significantly reduced and their colors are distorted. + +Following the revised underwater image formation model [2,32], under ambient illumination image intensity (per pixel, per color channel) is given as: + +$$ +I = J \cdot e ^ {- \phi_ {a} \cdot D} + \phi^ {\infty} \cdot \left(1 - e ^ {- \phi_ {b} \cdot D}\right), \tag {1} +$$ + +where $I$ is the linear image captured by the camera of a scene with range $D$ , $J$ is the clear scene that would have been captured had there been no water along the LOS, and $\phi^{\infty}$ is the water color at infinity, i.e., the backscatter at areas that contain no objects. The two parameters $\phi_{a}$ and $\phi_{b}$ are the attenuation and backscatter coefficients, respectively. + +# 3.2 Diffusion Models + +Diffusion models have proven to be very effective in capturing the distribution of natural images, given in training data. We describe here the formulation that we use in our work, adopted from [16]. A diffusion model is defined by a Markov chain designed to transform the distribution of real images $x_0$ to a Gaussian distribution $x_{T}$ , by gradually scaling down the image values and adding Gaussian noise, using a schedule determined by the scalar parameters $\alpha_{t}$ , + +$$ +x _ {t} = \sqrt {\alpha_ {t}} x _ {t - 1} + \sqrt {1 - \alpha_ {t}} \epsilon , \quad \epsilon \sim \mathcal {N} (0, I). \tag {2} +$$ + +Based on this forward process, an inverse process is trained to gradually denoise images starting from a Gaussian distribution, back into the distribution of the original image dataset. This can be formulated as a factorization of the joint distribution over the images in reverse order, $p(x_{0:T}) = \prod_{t=1}^{T} p_t(x_{t-1} \mid x_t)$ , and $p_t(x_{t-1} \mid x_t)$ is approximated by a Gaussian + +$$ +p _ {t} \left(x _ {t - 1} \mid x _ {t}\right) \sim \mathcal {N} \left(\mu_ {\theta} \left(x _ {t}, t\right), \Sigma_ {\theta} \left(x _ {t}, t\right)\right) ， \tag {3} +$$ + +with parameters $\mu_{\theta}$ and $\Sigma_{\theta}$ predicted by a trained neural network conditioned on $x_{t}$ and $t$ . Simulating this process results in samples form the approximated data distribution $p(x_0)$ . A popular implementation first predicts the noise at each time step, and then computes the mean by + +$$ +\mu_ {\theta} (x _ {t}, t) = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (x _ {t}, t)\right), \tag {4} +$$ + +where $\epsilon_{\theta}(x_t,t)$ is the neural network trained to approximate $\epsilon$ , and $\bar{\alpha}_t = \prod_{s=0}^t \alpha_s$ . The same neural network can also be used to predict a diagonal covariance $\Sigma_{\theta}(x_t,t)$ . At any iteration, an estimate of the clean image $x_0$ can be derived by computing the mean of $p(x_0 \mid x_t)$ : + +$$ +\hat {x} _ {0} (x _ {t}, t) = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} (x _ {t}, t)\right). \qquad \qquad (5) +$$ + +The above formulation can also be developed from a score-based modeling approach, where it can be shown that $\epsilon_{\theta}(x_t,t)$ is an approximation to the score function $\nabla_{x_t}\log p_t(x_t)$ . + +# 3.3 Posterior Sampling + +Under the score-based view, given an observation $y$ and a likelihood function $p(y|x)$ , we can use the same mechanism to sample from the posterior $p(x|y)$ using the posterior score function + +$$ +\nabla_ {x _ {t}} \log p _ {t} (x _ {t} | y) = \nabla_ {x _ {t}} \log p _ {t} (x _ {t}) + \nabla_ {x _ {t}} \log p _ {t} (y | x _ {t}). \tag {6} +$$ + +This idea of adding a conditional signal to the score is also called guidance. The challenge with the second term in Eq. 6, is that usually we are given a likelihood + +model based on the clean image $x_0$ , and not an intermediary noisy image $x_t$ . In our case the underwater image formation model transforms a clean image to an underwater observation. Connecting the model to the noisy sample $x_t$ leads to the intractable integral $p_t(y|x_t) = \int p(y|x_0)p(x_0|x_t)dx_0$ . + +The various works on posterior sampling propose different approximations of this integral. Some propose to collapse the uncertainty in $y$ [9, 12-14, 21, 51]. In [18] a variational inference approximation is proposed, and in [5, 11, 28, 29, 49] the likelihood model is computed on either $x_{t}$ directly, or on the mean $\hat{x}_0 = \mathbb{E}[x_0|x_t]$ , as given in Eq. 5. After some experimentation, we decided to use the latter approximation as formulated in DPS [11]: + +$$ +\nabla_ {x _ {t}} \log p _ {t} (x _ {t} \mid y) \approx \nabla_ {x _ {t}} \log p _ {t} (x _ {t}) + s \nabla_ {x _ {t}} \log p (y \mid \hat {x} _ {0} (x _ {t}, t)), \tag {7} +$$ + +where $s$ is the guidance scale used to control the weight of the approximated likelihood term. + +# 4 Method + +We use the underwater image formation model (Eq. 1) as a forward model that maps the space of natural images $x$ to the space of underwater observed images $y$ . We aim to use this forward model to construct a likelihood term $p(y \mid x)$ , and use it to sample from the posterior distribution of images (Eq. 7). In our case this cannot be implemented directly since the image formation model contains unknown parameters, the depth $D$ at each pixel, and the water parameters, $\phi_{a}, \phi_{b}, \phi^{\infty}$ , which can differ between scenes. One way to deal with the unknown depth, is to use a monocular depth estimator in a separate first stage, and then use it as part of the forward model. In the results we discuss this approach and show that it is suboptimal (method variant termed DA-Osmosis). + +Another approach is to optimize the unknown parameters, including the depth, during sampling. In some recent work, different methods to do this were proposed [17, 38, 59]. However, these do not work properly when the unknown parameter is high dimensional, highly complex, and strongly correlated with the target image. In our early experimentation we found that in order to separate the image into a clean image and the effects governed by the depth, we need more than a good image prior. + +Therefore, one of our main contributions, is to train a joint prior on both the color and depth in clean images, and show how it can be used for underwater image restoration. Following this we define $x_0 = (J, D)$ , where $J$ represents a clear image and consists of the 3 color channels, and $D$ is the depth image. This allows us to use the following forward model (based on Eq. 1) conditioned on the estimated clean image and depth $\hat{x}_0 = (\hat{J}_0, \hat{D}_0)$ : + +$$ +f _ {\phi} \left(\hat {x} _ {0}\right) = \hat {J} _ {0} e ^ {- \phi_ {a} \hat {D} _ {0}} + \phi^ {\infty} \left(1 - e ^ {- \phi_ {b} \hat {D} _ {0}}\right), \tag {8} +$$ + +where $\phi = (\phi_{a},\phi_{b},\phi^{\infty})$ are the remaining 9 unknown water parameters (a parameter per color channel for each of $\phi = (\phi_a,\phi_b,\phi^\infty)$ ). We optimize $\phi$ during + +![](images/3e86b82f8a475dc3e49e2c4717a28416d5ea7469eb3c4c8d8317064a08b3868c.jpg) +Fig. 3: [Left] Example images from outdoor RGBD datasets used for training our prior. From left to right: DIODE [55], ReDWeb-S [36], HR-WSI [57], KITTI [20]. [Right] Samples from the trained RGBD prior. The samples demonstrate the inherent correlation between RGB image and depth in our trained RGBD prior. + +sampling, in a similar way to GDP [17], using gradient descent on the image reconstruction loss. The likelihood term used for guidance is defined as a Gaussian around $f_{\phi}(\hat{x}_0)$ with a fixed variance. + +Training a joint prior on color and depth, not only allows us to use the above forward model, but also exploits the diffusion model capacity to capture the complexity of the depth image and its correlations with the clean color image. + +# 4.1 Training the Prior + +To learn a natural image prior, we train the joint prior model on both color and depth using public RGBD datasets of outdoor scenes that were taken in air. While differing from underwater images, the ability to use large amounts of high quality training data leads to a prior that captures the correlation between color and depth in natural scenes, which as we show, is an important aspect for underwater image restoration. + +For the sake of data efficiency and training time, we start with a pretrained diffusion model trained on ImageNet (we take the unconditional model from [16]), and finetune it on RGBD data. We implement this, inspired by DepthGen [45], by replacing the first and last layers of the U-Net to 4 channels rather than 3, and initializing those layers randomly. The datasets we use for finetuning are (see Fig. 3[left]): DIODE [55], RedWeb-S [36], HR-WSI [57], KITTI [20], with 16884, 2179, 20378, 23946 pictures, respectively. + +A major challenge in working with RGBD data is to turn the available depth information into a proper image. This includes filling holes and scaling the values to a standard range. Since each of the above datasets were collected in a different manner, we treat each of them differently. This is detailed in the supplementary material. Fig. 3[right] shows two samples from the prior that show corresponding color and depth images. While prior samples demonstrate an evident domain gap between in-air and underwater data, our results show that posterior samples are not affected by the gap, and can leverage the strong correlation between depth and color in natural scenes. + +# 4.2 Sampling from the Posterior + +Given a trained prior, we perform underwater image restoration by sampling from the posterior with guidance of the image formation model. The sampling process is described in Figs. 2, 4. In each iteration we generate samples of both the image and the depth, while optimizing the water parameters. This results in a gradual update to the estimates of the clean image, and the depth, demonstrated in Fig. 4[right]. We adopt the sampling method in DPS [11], using the reconstruction loss, $\left\| y - f_{\phi}(\hat{x}_0)\right\|_2^2$ , which is the negative log-likelihood formed from the model in Eq. 8. Similar to GDP [17], the remaining unknown parameters $\phi$ are optimized in parallel to the inverse sampling process, using gradient descent on the same loss used for guidance. + +Note that one of our main novelties is that the RGBD prior is used as an inherent part of the iterative method and not as a stand-alone depth estimator. The depth is estimated together with the image in every iteration, guided by the underwater model, see, e.g., Fig. 4[right]. As demonstrated in the results, this improves the restoration quality over using a fixed estimated depth of the scene. + +Running posterior sampling, guided by a reconstruction loss only, we observe two problems. First, for pixels with large depth values, the reconstructed image is dominated by the backscatter, making the estimation of the clean pixel color values unstable. Second, the color values can shift outside the valid range, causing color saturation. To overcome the first issue we multiply the reconstruction loss in every pixel by the estimated depth value (without passing the gradients of this operation, denoted by $sg$ ): + +$$ +\mathcal {L} _ {\text {r e c}} = \left\| s g (\hat {D} _ {0}) \cdot \left(y - f _ {\phi} (\hat {x} _ {0})\right) \right\| _ {2} ^ {2}. \tag {9} +$$ + +In order to overcome the second issue, we introduce two auxiliary losses. The first penalizes RGB values outside the valid range $[-1, 1]$ , and the second encourages the average values of each channel to approach the middle of the color range (in line with the gray world assumption). We implement the two auxiliary losses as: + +$$ +\mathcal {L} _ {\mathrm {v a l}} = \lambda_ {v} \sum_ {i, c} \max \left(\left| \hat {J} _ {0} (i, c) \right| - T _ {v}, 0\right) ^ {2}, \quad \mathcal {L} _ {\mathrm {a v r g}} = \lambda_ {a} \sum_ {c} \left| \sum_ {i} \hat {J} _ {0} (i, c) - T _ {a} \right| +$$ + +where $\hat{J}_0(i,c)$ is the value of color channel $c$ of pixel $i$ , assuming the valid color range is $[-1,1]$ , $T_v$ and $T_a$ are thresholds set to values close to 1 and 0 respectively, and $\lambda_v,\lambda_a$ are the scalar weights of both losses. The total loss is then: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {\text {v a l}} + \mathcal {L} _ {\text {a v r g}}. \tag {10} +$$ + +At each iteration we compute the gradient of the loss with respect to $x_{t}$ , which forms the log-likelihood gradient in the posterior score (Eq. 7), and the gradient with respect to $\phi$ which is used to update the parameters in the underwater model. In order to stabilize the optimization, we apply gradient clipping by value, to the gradients of $x_{t}$ , and we run the optimization of $\phi$ only in some of + +Algorithm 1: Osmosis - Sampling +Input: y - Underwater Degraded Image (RGB image) $f(\hat{x}_0)$ - Underwater Physical Degradation Model Initialization of $\phi = \{\phi_a,\phi_b,\phi^\infty \}$ // size = [9] Output: $\hat{x}_0 = [\hat{J}_0,\hat{D}_0]$ .. $\hat{J}_0$ - Restored RGB Image, $\hat{D}_0$ - Depth Estimation +1 $x_{T}\sim \mathcal{N}(0,I)$ // $x_{T}$ size = [h,w,4] +2 for $t = T:1$ do +3 $\epsilon \gets \epsilon_{\theta}(x_t,t)$ +4 $\mu = \frac{1}{\sqrt{\alpha_t}} (x_t - \frac{\beta_t}{\sqrt{1 - \bar{\alpha}_t}}\cdot \epsilon)$ +5 $\hat{x}_0 = \frac{1}{\sqrt{\bar{\alpha}_t}} (x_t - \sqrt{1 - \bar{\alpha}_t}\cdot \epsilon)$ +6 $f_{\phi}(\hat{x}_0) = \hat{J}_0e^{-\phi_a\hat{D}_0} + \phi^{\infty}\left(1 - e^{-\phi_b\hat{D}_0}\right)$ +7 $\mathcal{L} =$ loss from Eq. 10 +8 if Optimstart $\geq t\geq$ Optimend then +9 for $i = 1:N$ do +10 | $\phi \gets \phi -\eta \nabla_{\phi}\mathcal{L}$ +11 end +12 end +13 $x_{t - 1}\sim \mathcal{N}(\mu -s\cdot \nabla_{x_t}\mathcal{L},\Sigma)$ +14 end + +![](images/60362b871aea45a00a5c8a56e41f258b858e99b217edf629e8f74acc0c4ae836.jpg) +Fig. 4: Our algorithm. [Left] Detailed steps of our algorithm. [Right] Example of how $\hat{J}_0, \hat{D}_0$ change during the iterations. + +the sampling steps (defined by the values $\mathrm{Optim}_{\mathrm{start}}$ and $\mathrm{Optim}_{\mathrm{end}}$ ), running $N$ iterations of gradient descent in each step. When the gradient of $x_{t}$ is applied in the sampling step, it is multiplied by a guidance scale $s$ . We find that using a smaller scale for the depth channel leads to better results. This makes sense as while all channels are treated the same in the prior, the depth has a different role in the forward model (specifically it is used inside an exponent), and this can lead to a different gradient scale. More details on the implementation are given in the supplementary. + +# 5 Results + +We use a prior on $256 \times 256$ images, with sampling time of about 3 minutes per image on an Nvidia A100 GPU. We present here selected results and analysis. Please refer to the supplementary material for a complete set. + +Real-World Scenes. We present an extensive comparison on real-world linear images from the datasets SQUID [7], SeaThru [3], SeaThruNerf [32], and additional images acquired in different locations in the world, the Indian and + +Pacific Oceans, and Mediterranean, and Red Seas. Images are white-balanced as pre-processing. All runs use the same set of parameters. Fig. 1 shows several results of both image restoration and depth estimation. + +We compare image restoration with the following methods: CWR [24], DM [53], FUnIE-GAN [27], GDCP [41], IBLA [42], MMLE [60], semi-UIR [26], Ucolor [33], USUIR [19], UW-Net [23], waternet [35], and USe-ReDI-Net [54]. For compactness, Fig. 5 summarizes the results for a chosen subset of methods. The complete comparison is shown in the supplementary material. Among previous methods, we found GDCP to be the most consistent. Our result recovers the full range of the scenes, specifically improving contrast of objects that are further away (note the zoomed-in objects in the far areas, e.g., the diver in row 1). Our method recovers vibrant and consistent colors also in further areas. For example, in rows 2,3,4 note that the background rocks and sand appear bluish in all methods except of ours. + +We also compare to a method that restores the RGB using a fixed depth that is pre-estimated using an external SOTA method, depth anything (DA) [58]. We use the pre-estimated depth in our method's pipeline instead of our gradually estimated depth. We term this method $DA$ -Osmosis, and the results show that our method restores the far areas with better color, emphasizing the advantage of using our joint RGBD prior. + +We compare depth estimation with GDCP [41], IBLA [42], unveiling [6], UW-Net [23], and monoUWnet [4], and depth anything (DA) [58]. In GDCP [41] and unveiling [6] we compute the depth from the output transmission maps. Our depth estimations have more details and better explain the scenes (Fig. 5). + +In addition, we conducted a non-reference quantitative comparison using the MUSIQ [31] measure on 50 real-world images presented in the paper and supplementary material. Our method achieved the highest score. The quantitative results are summarized in Table 1. + +The UIEB dataset [35] is sometimes used for quantitative evaluation of image restoration. However, it is collected from various sources and is not linear. Thus it is not suitable for physics-based methods like ours. In addition, the 'ground truth' of the UIEB dataset is chosen from results of previous methods and it is not the real ground truth. Fig. 6 shows four examples (out of many) where our result removes more water effects than the ground truth, even despite the non-linearity of the input. + +Simulation. We conducted a simulation following [26, 34] using all the 1449 images from the NYU-v2 RGBD dataset [46] with randomly varying water parameters and generated the underwater images with the image formation model in Eq. 1. For fairness in comparison with other methods we applied $\phi_{a} = \phi_{b}$ . + +The quantitative results are summarized in Table 2. Our method substantially outperforms other methods in the image restoration metrics of PSNR, SSIM, and LPIPS. We emphasize that our prior was not trained on this data, or indoor data at all. Moreover, two of the methods [26,34] we outperform were trained on this data. + +![](images/f4b142b6cdff793a8bd591999f1024301558cf19c1d87c7a599d39925154352d.jpg) +Fig. 5: Real-world restoration results. From left to right: white-balanced input, contrast stretch, GDCP [41], Ucolor [33], CWR [24], semi-UIR [26], DM [53], Depth Anything [58] - Osmosis, Osmosis (ours). Zoom-in colored rectangles emphasize far objects that have higher contrast in our results. Real-world depth results. From left to right: GDCP [41], IBLA [42], unveiling [6], UW-Net [23], monoUWnet [4], Depth Anything [58], Osmosis (ours). Our depth results are smoother and less affected by object gradients. The reader is encouraged to zoom-in. + +![](images/9b38183a0241158bcbbb723669f1d50af5b3358eb4b5a8805773054b93d2cece.jpg) +Fig. 6: Example results on UIEB [35], a dataset comprised from non-linear images with ground truth produced by different enhancement algorithms. Our method yields better results than the dataset's ground truth. See for example the color of the sand, the orange color of the fish in the top-left, and the divers' authentic skin color in the bottom-right. + +![](images/9a61fe0eaf08a26298a39d1d1ebb64075ba9f8dca4163305b8b53b0112cba9e9.jpg) +Fig. 7: Ablation. Ablations $\#1, \#5, \#6$ show the importance of the losses we added to the reconstruction loss. Without $\mathcal{L}_{\mathrm{val}} (\#6)$ , the colors tend to "explode" and oversaturate. Without $\mathcal{L}_{\mathrm{avg}} (\#5)$ the colors sometimes skew towards pink/purple. In $\#2$ , the further areas are not restored well, because the loss is not weighted by the depth $D$ . In $\#3$ the guidance scale is the same for the RGB and depth channels. This harms the depth reconstruction. In $\#4$ we set $\phi_{a} = \phi_{b}$ (separately per color channel). Since this is an inaccurate model, the restoration in further areas is harmed. + +Ablation. To demonstrate the effect of different components in our methods, we conduct an ablation study of the following variants: 1. $\mathcal{L} = \mathcal{L}_{\mathrm{rec}}$ (instead of Eq. 10); 2. Removing weighting by $\hat{D}_0$ in Eq. 9 3. Same guidance scale $s$ for all channels; 4. $\phi_{a} = \phi_{b}$ ; 5. $\mathcal{L} = \mathcal{L}_{\mathrm{rec}} + \mathcal{L}_{\mathrm{val}}$ (remove $\mathcal{L}_{\mathrm{avg}}$ from Eq. 10); 6. $\mathcal{L} = \mathcal{L}_{\mathrm{rec}} + \mathcal{L}_{\mathrm{avg}}$ (remove $\mathcal{L}_{\mathrm{val}}$ from Eq. 10). Numerical results on the simulation for variants $\# 1 - \# 3$ are presented in Table 2. Ablations $\# 4 - \# 6$ are shown only on real world images since in the simulation we do not use $\mathcal{L}_{\mathrm{avg}}$ , and $\phi_{a} = \phi_{b}$ . Fig. 7 presents the results of all variants on one of the scenes. We see that the additional losses are important to prevent color saturation and shift. Increasing the loss weight with depth improves restoration in further areas. Separating guidance scales between the RGB and depth channels improves depth reconstruction. Setting $\phi_{a} \neq \phi_{b}$ extends the range of the restoration. + +
MethodMUSIQ↑
input51.59
contrast stretch53.75
CWR [24]39.63
DM [53]54.64
FUnIE-GAN [27]42.70
GDCP [41]56.25
IBLA [42]54.42
MMLE [60]56.26
semi-UIR [26]54.99
Ucolor [33]47.73
USUIR [19]52.44
UW-Net [23]46.02
waternet [35]53.64
unveiling [6]50.83
DA-osmosis [58]51.23
osmosis (ours)56.62
+ +Table 1: Non-reference quantitative comparison on 50 real-world images using MUSIQ [31]. Our method achieves the highest score. + +
MethodPSNR↑SSIM↑LPIPS↓
contrast stretch17.130.830.11
CWR [24]16.930.790.20
DM [53]17.410.820.12
FUnIE-GAN [27]17.640.770.21
GDCP [41]12.410.710.16
IBLA [42]15.070.700.19
MMLE [60]17.000.740.17
semi-UIR [26]17.820.830.12
Ucolor [33]17.920.830.10
USUIR [19]16.760.800.18
UW-Net [23]18.040.750.26
waternet [35]17.270.820.11
unveiling [6]16.340.790.18
ablation #121.090.860.09
ablation #222.170.880.07
ablation #322.000.880.06
osmosis (ours)22.740.890.06
+ +Table 2: Quantitative comparison on the simulation. Our method achieves best scores in image restoration. + +# 6 Discussion + +In this work we demonstrated how to harness the strength of a new RGBD diffusion prior to achieve state-of-the-art results on underwater image restoration. To do this we solved several challenges: i) because of lack of clean underwater image data we use datasets of scenes in air; ii) we notice that the color prior does not suffice to guide restoration, therefore, we add the scene depth to the prior; iii) We use the physical image formation model to guide restoration, and also estimate the water parameters in the process. This results in the most comprehensive single underwater image restoration method to-date. It does not train on any underwater images and therefore does not overfit to any. It gives higher weight to further objects and therefore is superior in reconstructing all the details of the scene. Jointly solving for the depth results in excellent depth estimation from monocular images. + +There is a domain gap between the prior (trained on in-air data) and underwater data. This is intentional. The beauty of our method is that by incorporating the underwater model in the sampling it succeeds even without training on underwater data (that is very hard to obtain). Thus, it is not specialized for a certain type of water or objects, and it learns the color distribution of natural images not degraded by water. + +Like all diffusion models with U-Nets, our method is limited by a fixed resolution and long running time. This could potentially be improved using other network architectures and sampling methods. + +Acknowledgements. The research was funded by Israel Science Foundation grant #1951/23, Israeli Ministry of Science and Technology grants #1001577600 & #1001593851, EU Horizon 2020 research and innovation programme GA 101094924 (ANERIS), the Leona M. and Harry B. Helmsley Charitable Trust, and the Maurice Hatter Foundation. We thank Dr. Derya Akkaynak and Dr. Matan Yuval for substantial data contribution, Amir Dayan for the paper's name and Meirav Keidar for graphics design. + +# References + +1. Aali, A., Arvinte, M., Kumar, S., Tamir, J.I.: Solving inverse problems with score-based generative priors learned from noisy data. arXiv preprint arXiv:2305.01166 (2023) +2. Akkaynak, D., Treibitz, T.: A revised underwater image formation model. In: CVPR. pp. 6723-6732 (2018) +3. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: CVPR. pp. 1682-1691 (2019) +4. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: IEEE Int. Conf. Robotics and Automation (ICRA). pp. 1098-1104 (2023) +5. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: CVPR. pp. 843-852 (2023) +6. Bekerman, Y., Avidan, S., Treibitz, T.: Unveiling optical properties in underwater images. In: ICCP (2020) +7. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE TPAMI 43(8), 2822-2837 (2020) +8. Chan, M.A., Young, S.I., Metzler, C.A.: Sud $^2$ : Supervision by denoising diffusion models for image reconstruction. arXiv preprint arXiv:2303.09642 (2023) +9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938 (2021) +0. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: CVPR (2023) +1. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGT0k +2. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. NeurIPS 35, 25683-25696 (2022) +3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022) +4. Chung, H., Ye, J.C.: Score-based diffusion models for accelerated mri. Medical image analysis 80, 102479 (2022) +5. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A.G., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. arXiv preprint arXiv:2305.19256 (2023) +6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS 34, 8780-8794 (2021) +7. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: CVPR. pp. 9935-9946 (2023) + +18. Feng, B.T., Smith, J., Rubinstein, M., Chang, H., Bouman, K.L., Freeman, W.T.: Score-based diffusion models as principled priors for inverse imaging. arXiv preprint arXiv:2304.11751 (2023) +19. Fu, Z., Lin, H., Yang, Y., Chai, S., Sun, L., Huang, Y., Ding, X.: Unsupervised underwater image restoration: From a homology perspective. In: Proc. AAAI Conf. on Artificial Intelligence. vol. 36, pp. 643-651 (2022) +20. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. Int. J. Robotics Research (IJRR) (2013) +21. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: NeurIPS (2022) +22. Guo, L., Wang, C., Yang, W., Huang, S., Wang, Y., Pfister, H., Wen, B.: Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In: CVPR. pp. 14049-14058 (2023) +23. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: ICIP. pp. 624-628 (2019) +24. Han, J., Shoeiby, M., Malthus, T., Botha, E., Anstee, J., Anwar, S., Wei, R., Armin, M.A., Li, H., Petersson, L.: Underwater image restoration via contrastive learning and a real-world dataset. Remote Sensing 14(17), 4297 (2022) +25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020) +26. Huang, S., Wang, K., Liu, H., Chen, J., Li, Y.: Contrastive semi-supervised learning for underwater image restoration via reliable bank. In: CVPR. pp. 18145-18155 (2023) +27. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters 5(2), 3227-3234 (2020) +28. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. NeurIPS 34, 14938-14954 (2021) +29. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. NeurIPS 35, 23593-23606 (2022) +30. Kawar, B., Elata, N., Michaeli, T., Elad, M.: Gsure-based diffusion model training with corrupted data. arXiv preprint arXiv:2305.13128 (2023) +31. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5148-5157 (October 2021) +32. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: CVPR. pp. 56-65 (2023) +33. Li, C., Anwar, S., Hou, J., Cong, R., Guo, C., Ren, W.: Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE TIP 30, 4985-5000 (2021) +34. Li, C., Anwar, S., Porikli, F.: Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition 98, 107038 (2020) +35. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE TIP 29, 4376-4389 (2019) +36. Liu, N., Zhang, N., Shao, L., Han, J.: Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI (2021) +37. Lu, S., Guan, F., Zhang, H., Lai, H.: Underwater image enhancement method based on denoising diffusion probabilistic model. J. of Visual Communication and Image Representation 96, 103926 (2023) + +38. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: GibbsDDRM: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. In: Int. Conf. on Machine Learning (2023) +39. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023) +40. Peng, L., Zhu, C., Bian, L.: U-shape transformer for underwater image enhancement. IEEE TIP (2023) +41. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE TIP 27(6), 2856-2868 (2018) +42. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE TIP 26(4), 1579-1594 (2017) +43. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015) +44. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR. pp. 22500-22510 (2023) +45. Saxena, S., Kar, A., Norouzi, M., Fleet, D.J.: Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816 (2023) +46. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760 (2012) +47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Int. conf. machine learning. pp. 2256-2265 (2015) +48. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983 (2023) +49. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: ICLR (2023), https://openreview.net/forum?id=9GsMA8MRKQ +50. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS. vol. 32 (2019) +51. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005 (2021) +52. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +53. Tang, Y., Kawasaki, H., Iwaguchi, T.: Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In: Proc. ACM Int. Conf. Multimedia. pp. 5419-5427 (2023) +54. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: ICCV. pp. 12248-12258 (2023) +55. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463 +56. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023) + +57. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR (June 2020) +58. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) +59. Yang, P., Zhou, S., Tao, Q., Loy, C.C.: PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In: NeurIPS (2023) +60. Zhang, W., Zhuang, P., Sun, H.H., Li, G., Kwong, S., Li, C.: Underwater image enhancement via minimal color loss and locally adaptive contrast enhancement. IEEE TIP 31, 3997-4010 (2022) +61. Zhou, J., Yang, T., Zhang, W.: Underwater vision enhancement technologies: a comprehensive review, challenges, and recent trends. Applied Intelligence 53(3), 3594-3621 (2023) \ No newline at end of file diff --git a/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/images.zip b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..741793ed5a03e1f6d0f7098468a80c22c5aa4f34 --- /dev/null +++ b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:695c0b101dd59e32c1163bb70fe2841df096259b246c42fe6050bc74366c357a +size 677780 diff --git a/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/layout.json b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..405e4f06df47e9cf06027a544e16b0754b104f80 --- /dev/null +++ b/2024/Osmosis_ RGBD Diffusion Prior for Underwater Image Restoration/layout.json @@ -0,0 +1,9947 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "spans": [ + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "text", + "content": "Opher Bar Nathan" + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "text", + "content": ", Deborah Levy" + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "text", + "content": ", Tali Treibitz" + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "text", + "content": ", and Dan Rosenbaum" + }, + { + "bbox": [ + 171, + 167, + 442, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 201, + 473, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 201, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 473, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 201, + 473, + 213 + ], + "type": "text", + "content": " Hatter Department of Marine Technologies, Charney School of Marine Sciences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 213, + 380, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 213, + 380, + 246 + ], + "spans": [ + { + "bbox": [ + 233, + 213, + 380, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 233, + 213, + 380, + 246 + ], + "type": "text", + "content": " Department of Computer Science, University of Haifa, Haifa, Israel osmosis-diffusion.github.io" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 274, + 455, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 274, + 455, + 439 + ], + "spans": [ + { + "bbox": [ + 159, + 274, + 455, + 439 + ], + "type": "text", + "content": "Abstract. Underwater image restoration is a challenging task because of water effects that increase dramatically with distance. This is worsened by lack of ground truth data of clean scenes without water. Diffusion priors have emerged as strong image restoration priors. However, they are often trained with a dataset of the desired restored output, which is not available in our case. We also observe that using only color data is insufficient, and therefore augment the prior with a depth channel. We train an unconditional diffusion model prior on the joint space of color and depth, using standard RGBD datasets of natural outdoor scenes in air. Using this prior together with a novel guidance method based on the underwater image formation model, we generate posterior samples of clean images, removing the water effects. Even though our prior did not see any underwater images during training, our method outperforms state-of-the-art baselines for image restoration on very challenging scenes. Our code, models and data are available on the project's website." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 449, + 453, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 449, + 453, + 472 + ], + "spans": [ + { + "bbox": [ + 159, + 449, + 453, + 472 + ], + "type": "text", + "content": "Keywords: Diffusion Models " + }, + { + "bbox": [ + 159, + 449, + 453, + 472 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 449, + 453, + 472 + ], + "type": "text", + "content": " Physics-Based Computer Vision " + }, + { + "bbox": [ + 159, + 449, + 453, + 472 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 449, + 453, + 472 + ], + "type": "text", + "content": " Underwater Image Restoration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 491, + 230, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 230, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 230, + 504 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 516, + 482, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 600 + ], + "type": "text", + "content": "Underwater images are used in many applications, such as, underwater construction and maintenance, marine sciences, and fisheries. However, their automatic analysis is hindered because of the optical effects of the water that strongly attenuates and scatters light in a wavelength dependent manner. This causes color distortion and loss of contrast that exponentially increase with depth3. With growing human activity in the oceans, clear underwater vision becomes increasingly important." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 601, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 636 + ], + "type": "text", + "content": "Restoring underwater scenes is still a very challenging ill-posed problem. Classic approaches are based on designing priors for clean images or inverting the water formation model, and are limited by the ability to form strong priors. On" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "3 For consistency with computer vision literature the term depth is used throughout to refer to the distance from the camera rather than water depth." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 235 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 235 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 235 + ], + "type": "image", + "image_path": "4b4cb06e8a2c39b20c10752b8793299cde98dfbcf28485bfb197e4c2d0735bd3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 241, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 275 + ], + "type": "text", + "content": "Fig. 1: Our method receives as an input a single underwater image and outputs the restored clean image and an estimated depth map. The output is estimated using a diffusion prior trained on RGBD images and the physical image formation model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 298, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 480, + 334 + ], + "type": "text", + "content": "the other hand, learning based approaches are limited by the lack of supervised training data of clean underwater images. This is a critical issue, as the ocean cannot be emptied for the sake of data collection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "text", + "content": "We suggest an unsupervised restoration method (Figs. 1,2) based on an inverse problem approach using a diffusion prior for both color and depth, coupled with the underwater image formation model. Restoring an image is formulated as posterior inference, computed using a natural image prior, and a likelihood term that is based on the underwater image formation model. The challenge with applying this approach for underwater image restoration directly is that (i) the degradation for each pixel depends on its depth, and other unknown parameters; and (ii) there is no ground truth clean underwater data to train the prior." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 480, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 501 + ], + "type": "text", + "content": "To solve this, we replace the image prior with a prior on the joint space of color and depth of natural images. Adding depth to the prior allows us to formulate the forward model that forms the likelihood term over the observed corrupted underwater image, and apply posterior sampling. Moreover, this leverages the high capacity of diffusion models, in capturing the strong correlations between color and depth in natural scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 480, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 480, + 585 + ], + "type": "text", + "content": "We propose to train a prior model of RGBD images using available datasets of natural outdoor scenes that were collected in air. Using in-air scenes for underwater images may be counter-intuitive, but actually has strong benefits: 1) it overcomes the lack of clean underwater image data; 2) it leads to a strong prior that captures the joint statistics governing color and depth in natural scenes, where, as opposed to underwater images, color does not fade with distance; 3) it prevents overfitting to specific types of underwater images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 586, + 481, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 481, + 633 + ], + "type": "text", + "content": "We then use this prior together with the underwater physical image formation model to simultaneously estimate the clean image, its depth, and the model parameters, all from a single underwater image (summarized in Fig. 2). We show that our method outperforms models that were trained on underwater data." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 112, + 476, + 291 + ], + "blocks": [ + { + "bbox": [ + 138, + 112, + 476, + 291 + ], + "lines": [ + { + "bbox": [ + 138, + 112, + 476, + 291 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 476, + 291 + ], + "type": "image", + "image_path": "66431529cd214904f8ef4ab142f8c8ab8cff33156409f1e8e3279aedda391347.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": "Fig. 2: The iterative sampling process starts in " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": " with random noise in 4 channels. The denoising step outputs denoised samples " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = (\\hat{J}_0, \\hat{D}_0)" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": ". We use the underwater physical image formation model together with " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": " to optimize the water parameters " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": ", and to guide the sampling towards the observed image. This process repeats itself, gradually updating both the estimated image and depth, until " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": " holds the method's estimate for both the reconstructed scene " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "J_0" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": " and its depth " + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 130, + 299, + 482, + 375 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 399, + 343, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 343, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 343, + 411 + ], + "type": "text", + "content": "The main contributions of this paper are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 480, + 517 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 411, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 446 + ], + "type": "text", + "content": "1. We train an RGBD prior, and demonstrate that modeling color and depth and jointly sampling them, provides a stronger diffusion prior for underwater image restoration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 480, + 483 + ], + "type": "text", + "content": "2. We propose a new method that combines the RGBD prior of in-air data with the underwater image formation model, leading to a diffusion guidance method that generates the restored underwater images as posterior samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 483, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 517 + ], + "type": "text", + "content": "3. We demonstrate that our method outperforms state-of-the-art underwater image restoration methods both qualitatively and quantitatively on real and simulated data." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 518, + 459, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 459, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 459, + 531 + ], + "type": "text", + "content": "We publish all our code, the new trained RGBD prior, images and results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 549, + 242, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 242, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 242, + 561 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 574, + 432, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 432, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 432, + 586 + ], + "type": "text", + "content": "2.1 Underwater Image Restoration and Depth Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": "Here we review recent works that are most relevant to our method. See [61] for a recent more comprehensive review." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Classic underwater image restoration. To cope with the ill-posed nature of the problem, earlier works introduced tailored image priors. Some priors aim to estimate a depth or transmission map of the scene to reduce the number of the unknowns and then use the estimated depth to restore the scenes [7, 41, 42]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "Supervised learning. Clean underwater image data is very scarce. Several datasets aim to mitigate this issue. In UIEB [35], the images denoted as ground-truth are generated from enhancement results of 9 different baseline methods, and having human observers vote for the best one. The LSUI dataset [40] is larger, produced using the same methodology, based on choosing results from 18 different methods. These datasets enable supervised learning methods, but are still limited by their scale, the quality of the baseline methods, and the human bias to choose visually pleasing rather than physically consistent images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 331 + ], + "type": "text", + "content": "CWR [24] introduced the HICRD dataset, where the images are restored using optical parameters impressively measured using ocean optics instruments. Unfortunately, the images are acquired in a downward-looking position, and thus their depth range is very limited. In FUnIE-GAN [27] a training dataset is generated by having humans select good images from a large set of unprocessed underwater images. The humans are instructed to choose images where the foreground objects are identifiable. These are then distorted by a GAN to produce the paired poor images. A synthetic dataset was generated in [34]. The dataset is synthesized from the NYU-v2 RGBD dataset [46], using the image formation model equation and several sets of values for the water parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": "Unsupervised learning. USUIR [19] aims to restore images without supervision, by separately estimating the image components (clean image, transmission, backscatter) and using them to construct an underwater image that is used for supervision against the original one. Subsequent frames were used in [4] for self-supervising monocular depth estimation, and in [54] for self-supervising both depth and restoration. In UW-NET [23] a cycle-GAN is used for learning mapping from RGBD in-air datasets to underwater images. As opposed to all these methods, we present the first prior based on diffusion models that does not rely on ground truth underwater supervision, and uses the physical model for inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 467, + 272, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 272, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 272, + 479 + ], + "type": "text", + "content": "2.2 Diffusion Model Prior" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "content": "Diffusion models. Diffusion models have emerged as a powerful type of generative models. In the last few years several formulations and variations have been developed [25,47,50,52], most of which use a U-Net architecture [43] as a noise predictor. Because the training relies on very large datasets and is extremely time-consuming, significant work has been devoted to the setup where models are first pre-trained on large datasets and only later fine-tuned to more specific data, closer to the tasks at hand [44, 48]. In DepthGen [45] this approach is taken forward, by using a pre-trained model to kickstart a new model trained on different modalities using a different architecture. This is done by replacing the input and output layers of the pre-trained U-Net. We take a similar approach for training our RGBD prior, by using a pretrained diffusion model that was trained on RGB only." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Conditional diffusion models have been used for various image restoration tasks, by training models using different levels of supervision. Examples of similar tasks as ours include dehazing and deraining [8,39,56], and shadow removal [22]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Conditional diffusion models for underwater image enhancement have also been proposed [37,53]. Our approach differs from these by focusing on image restoration that inverts the physical model, rather than relying on supervised data that is optimized for image appearance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 281 + ], + "type": "text", + "content": "Diffusion model as a prior for clean images. In addition to conditional generation of images, there is a growing body of research where diffusion models are being used as clean image priors, and image restoration is formulated as posterior sampling [5, 9, 11-14, 18, 21, 28, 29, 49, 51]. These include tasks like denoising, inpainting, deblurring, and more general tasks. The limited access to clean data in many cases, has led to research on training a prior of clean images, using noisy training data only. In [1, 15, 30] a diffusion model prior is trained using noisy data, assuming a known degradation model. Since this setup is not directly applicable in our case, we chose instead to train our prior on clean data that was not taken underwater." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "text", + "content": "Diffusion model prior for blind image restoration. A more challenging task, is to use diffusion models as a clean image prior when the degradation model is unknown or depends on unknown parameters. Several papers have proposed to tackle this problem in different setups. In [17,38,59] the unknown parameters are optimized during the sampling process with a reconstruction objective function. In [10] a separate prior is trained for the unknown parameters, and sampled in parallel with image sampling. Our method differs from the above in that we learn a prior of the main unknown aspect of the degradation model, namely the depth, together with the prior of the variables we want to infer - the image color. This is done by training a single diffusion model on the joint space of color and depth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 434, + 233, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 233, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 233, + 445 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 459, + 310, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 310, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 310, + 470 + ], + "type": "text", + "content": "3.1 Underwater Image Formation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 478, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 480, + 548 + ], + "type": "text", + "content": "In water, we observe two wavelength- and distance-dependent effects. First, the direct signal reflected from the object is attenuated. Second, light is scattered onto the object's line-of-sight (LOS), creating an additive signal termed backscatter that increases with distance. The occluding backscatter layer is independent of the scene content. Thus, the visibility and contrast of further objects is significantly reduced and their colors are distorted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 550, + 479, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 479, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 479, + 574 + ], + "type": "text", + "content": "Following the revised underwater image formation model [2,32], under ambient illumination image intensity (per pixel, per color channel) is given as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 583, + 480, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 583, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 227, + 583, + 480, + 597 + ], + "type": "interline_equation", + "content": "I = J \\cdot e ^ {- \\phi_ {a} \\cdot D} + \\phi^ {\\infty} \\cdot \\left(1 - e ^ {- \\phi_ {b} \\cdot D}\\right), \\tag {1}", + "image_path": "ca6a124e36301a290acc4137e8a854bafad11313dc44b517f6c9e346c20b3bc6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " is the linear image captured by the camera of a scene with range " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " is the clear scene that would have been captured had there been no water along the LOS, and " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\phi^{\\infty}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " is the water color at infinity, i.e., the backscatter at areas that contain no objects. The two parameters " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\phi_{a}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\phi_{b}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " are the attenuation and backscatter coefficients, respectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "type": "text", + "content": "3.2 Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "text", + "content": "Diffusion models have proven to be very effective in capturing the distribution of natural images, given in training data. We describe here the formulation that we use in our work, adopted from [16]. A diffusion model is defined by a Markov chain designed to transform the distribution of real images " + }, + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "text", + "content": " to a Gaussian distribution " + }, + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "text", + "content": ", by gradually scaling down the image values and adding Gaussian noise, using a schedule determined by the scalar parameters " + }, + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 135, + 482, + 207 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 216, + 481, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 216, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 214, + 216, + 481, + 229 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I). \\tag {2}", + "image_path": "ff6834ca17bbdaded5223066f1250ebe84164101eee1f702b0c1ca1f06d9cd8d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 236, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 482, + 298 + ], + "type": "text", + "content": "Based on this forward process, an inverse process is trained to gradually denoise images starting from a Gaussian distribution, back into the distribution of the original image dataset. This can be formulated as a factorization of the joint distribution over the images in reverse order, " + }, + { + "bbox": [ + 131, + 236, + 482, + 298 + ], + "type": "inline_equation", + "content": "p(x_{0:T}) = \\prod_{t=1}^{T} p_t(x_{t-1} \\mid x_t)" + }, + { + "bbox": [ + 131, + 236, + 482, + 298 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 236, + 482, + 298 + ], + "type": "inline_equation", + "content": "p_t(x_{t-1} \\mid x_t)" + }, + { + "bbox": [ + 131, + 236, + 482, + 298 + ], + "type": "text", + "content": " is approximated by a Gaussian" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 305, + 481, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 305, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 220, + 305, + 481, + 319 + ], + "type": "interline_equation", + "content": "p _ {t} \\left(x _ {t - 1} \\mid x _ {t}\\right) \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) ， \\tag {3}", + "image_path": "161fd465a5333215a97728e9ec6ad82237614f322c1c5bcf6be21a88998b16eb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "text", + "content": "with parameters " + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}" + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "text", + "content": " predicted by a trained neural network conditioned on " + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "text", + "content": ". Simulating this process results in samples form the approximated data distribution " + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "inline_equation", + "content": "p(x_0)" + }, + { + "bbox": [ + 131, + 326, + 482, + 374 + ], + "type": "text", + "content": ". A popular implementation first predicts the noise at each time step, and then computes the mean by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 382, + 481, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 382, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 212, + 382, + 481, + 409 + ], + "type": "interline_equation", + "content": "\\mu_ {\\theta} (x _ {t}, t) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right), \\tag {4}", + "image_path": "1e2a94177be86fa265f90995e7d7f2ec184473803a81754968f4ef35d6f91d7e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "content": " is the neural network trained to approximate " + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=0}^t \\alpha_s" + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "content": ". The same neural network can also be used to predict a diagonal covariance " + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "content": ". At any iteration, an estimate of the clean image " + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "content": " can be derived by computing the mean of " + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "inline_equation", + "content": "p(x_0 \\mid x_t)" + }, + { + "bbox": [ + 131, + 416, + 482, + 465 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 473, + 481, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 473, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 216, + 473, + 481, + 498 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {0} (x _ {t}, t) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t}, t)\\right). \\qquad \\qquad (5)", + "image_path": "a02923ce2f6169cf0e929be4dd71bf9ee241f20365525841b42948baf2b8ccd7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 504, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 504, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 504, + 482, + 541 + ], + "type": "text", + "content": "The above formulation can also be developed from a score-based modeling approach, where it can be shown that " + }, + { + "bbox": [ + 131, + 504, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 131, + 504, + 482, + 541 + ], + "type": "text", + "content": " is an approximation to the score function " + }, + { + "bbox": [ + 131, + 504, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p_t(x_t)" + }, + { + "bbox": [ + 131, + 504, + 482, + 541 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 557, + 257, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 257, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 257, + 569 + ], + "type": "text", + "content": "3.3 Posterior Sampling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "text", + "content": "Under the score-based view, given an observation " + }, + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "text", + "content": " and a likelihood function " + }, + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "inline_equation", + "content": "p(y|x)" + }, + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "text", + "content": ", we can use the same mechanism to sample from the posterior " + }, + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "inline_equation", + "content": "p(x|y)" + }, + { + "bbox": [ + 131, + 576, + 481, + 612 + ], + "type": "text", + "content": " using the posterior score function" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 194, + 621, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 621, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 194, + 621, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\nabla_ {x _ {t}} \\log p _ {t} (x _ {t} | y) = \\nabla_ {x _ {t}} \\log p _ {t} (x _ {t}) + \\nabla_ {x _ {t}} \\log p _ {t} (y | x _ {t}). \\tag {6}", + "image_path": "ff3a38c2813d9d5ee3fa5f2ee74e6fe3517b4186efc37ac2a17f82d9b8d0e7f6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "This idea of adding a conditional signal to the score is also called guidance. The challenge with the second term in Eq. 6, is that usually we are given a likelihood" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "O.Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "model based on the clean image " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": ", and not an intermediary noisy image " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": ". In our case the underwater image formation model transforms a clean image to an underwater observation. Connecting the model to the noisy sample " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": " leads to the intractable integral " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "p_t(y|x_t) = \\int p(y|x_0)p(x_0|x_t)dx_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": "The various works on posterior sampling propose different approximations of this integral. Some propose to collapse the uncertainty in " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": " [9, 12-14, 21, 51]. In [18] a variational inference approximation is proposed, and in [5, 11, 28, 29, 49] the likelihood model is computed on either " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": " directly, or on the mean " + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = \\mathbb{E}[x_0|x_t]" + }, + { + "bbox": [ + 130, + 164, + 480, + 236 + ], + "type": "text", + "content": ", as given in Eq. 5. After some experimentation, we decided to use the latter approximation as formulated in DPS [11]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 245, + 480, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 245, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 175, + 245, + 480, + 258 + ], + "type": "interline_equation", + "content": "\\nabla_ {x _ {t}} \\log p _ {t} (x _ {t} \\mid y) \\approx \\nabla_ {x _ {t}} \\log p _ {t} (x _ {t}) + s \\nabla_ {x _ {t}} \\log p (y \\mid \\hat {x} _ {0} (x _ {t}, t)), \\tag {7}", + "image_path": "285dce9c00fec577b1d18db89dc015928ef8ab3b881643904ce39242d5077ad0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "content": " is the guidance scale used to control the weight of the approximated likelihood term." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 308, + 202, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 202, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 202, + 321 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "content": "We use the underwater image formation model (Eq. 1) as a forward model that maps the space of natural images " + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "content": " to the space of underwater observed images " + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "content": ". We aim to use this forward model to construct a likelihood term " + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "inline_equation", + "content": "p(y \\mid x)" + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "content": ", and use it to sample from the posterior distribution of images (Eq. 7). In our case this cannot be implemented directly since the image formation model contains unknown parameters, the depth " + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "content": " at each pixel, and the water parameters, " + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\phi_{a}, \\phi_{b}, \\phi^{\\infty}" + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "content": ", which can differ between scenes. One way to deal with the unknown depth, is to use a monocular depth estimator in a separate first stage, and then use it as part of the forward model. In the results we discuss this approach and show that it is suboptimal (method variant termed DA-Osmosis)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 480, + 536 + ], + "type": "text", + "content": "Another approach is to optimize the unknown parameters, including the depth, during sampling. In some recent work, different methods to do this were proposed [17, 38, 59]. However, these do not work properly when the unknown parameter is high dimensional, highly complex, and strongly correlated with the target image. In our early experimentation we found that in order to separate the image into a clean image and the effects governed by the depth, we need more than a good image prior." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "content": "Therefore, one of our main contributions, is to train a joint prior on both the color and depth in clean images, and show how it can be used for underwater image restoration. Following this we define " + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "inline_equation", + "content": "x_0 = (J, D)" + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "content": " represents a clear image and consists of the 3 color channels, and " + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "content": " is the depth image. This allows us to use the following forward model (based on Eq. 1) conditioned on the estimated clean image and depth " + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = (\\hat{J}_0, \\hat{D}_0)" + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 220, + 618, + 480, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 618, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 220, + 618, + 480, + 633 + ], + "type": "interline_equation", + "content": "f _ {\\phi} \\left(\\hat {x} _ {0}\\right) = \\hat {J} _ {0} e ^ {- \\phi_ {a} \\hat {D} _ {0}} + \\phi^ {\\infty} \\left(1 - e ^ {- \\phi_ {b} \\hat {D} _ {0}}\\right), \\tag {8}", + "image_path": "fae185c907b021aff064e527157c63964cc7c9650d16b1c90530a12f84b61e27.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\phi = (\\phi_{a},\\phi_{b},\\phi^{\\infty})" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " are the remaining 9 unknown water parameters (a parameter per color channel for each of " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\phi = (\\phi_a,\\phi_b,\\phi^\\infty)" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "). We optimize " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " during" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 195 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 195 + ], + "type": "image", + "image_path": "3e86b82f8a475dc3e49e2c4717a28416d5ea7469eb3c4c8d8317064a08b3868c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 203, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 248 + ], + "type": "text", + "content": "Fig. 3: [Left] Example images from outdoor RGBD datasets used for training our prior. From left to right: DIODE [55], ReDWeb-S [36], HR-WSI [57], KITTI [20]. [Right] Samples from the trained RGBD prior. The samples demonstrate the inherent correlation between RGB image and depth in our trained RGBD prior." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "text", + "content": "sampling, in a similar way to GDP [17], using gradient descent on the image reconstruction loss. The likelihood term used for guidance is defined as a Gaussian around " + }, + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "inline_equation", + "content": "f_{\\phi}(\\hat{x}_0)" + }, + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "text", + "content": " with a fixed variance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 348 + ], + "type": "text", + "content": "Training a joint prior on color and depth, not only allows us to use the above forward model, but also exploits the diffusion model capacity to capture the complexity of the depth image and its correlations with the clean color image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 372, + 253, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 253, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 253, + 384 + ], + "type": "text", + "content": "4.1 Training the Prior" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 471 + ], + "type": "text", + "content": "To learn a natural image prior, we train the joint prior model on both color and depth using public RGBD datasets of outdoor scenes that were taken in air. While differing from underwater images, the ability to use large amounts of high quality training data leads to a prior that captures the correlation between color and depth in natural scenes, which as we show, is an important aspect for underwater image restoration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 556 + ], + "type": "text", + "content": "For the sake of data efficiency and training time, we start with a pretrained diffusion model trained on ImageNet (we take the unconditional model from [16]), and finetune it on RGBD data. We implement this, inspired by DepthGen [45], by replacing the first and last layers of the U-Net to 4 channels rather than 3, and initializing those layers randomly. The datasets we use for finetuning are (see Fig. 3[left]): DIODE [55], RedWeb-S [36], HR-WSI [57], KITTI [20], with 16884, 2179, 20378, 23946 pictures, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "A major challenge in working with RGBD data is to turn the available depth information into a proper image. This includes filling holes and scaling the values to a standard range. Since each of the above datasets were collected in a different manner, we treat each of them differently. This is detailed in the supplementary material. Fig. 3[right] shows two samples from the prior that show corresponding color and depth images. While prior samples demonstrate an evident domain gap between in-air and underwater data, our results show that posterior samples are not affected by the gap, and can leverage the strong correlation between depth and color in natural scenes." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 304, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 304, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 304, + 129 + ], + "type": "text", + "content": "4.2 Sampling from the Posterior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "text", + "content": "Given a trained prior, we perform underwater image restoration by sampling from the posterior with guidance of the image formation model. The sampling process is described in Figs. 2, 4. In each iteration we generate samples of both the image and the depth, while optimizing the water parameters. This results in a gradual update to the estimates of the clean image, and the depth, demonstrated in Fig. 4[right]. We adopt the sampling method in DPS [11], using the reconstruction loss, " + }, + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\left\\| y - f_{\\phi}(\\hat{x}_0)\\right\\|_2^2" + }, + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "text", + "content": ", which is the negative log-likelihood formed from the model in Eq. 8. Similar to GDP [17], the remaining unknown parameters " + }, + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 137, + 482, + 258 + ], + "type": "text", + "content": " are optimized in parallel to the inverse sampling process, using gradient descent on the same loss used for guidance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "text", + "content": "Note that one of our main novelties is that the RGBD prior is used as an inherent part of the iterative method and not as a stand-alone depth estimator. The depth is estimated together with the image in every iteration, guided by the underwater model, see, e.g., Fig. 4[right]. As demonstrated in the results, this improves the restoration quality over using a fixed estimated depth of the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 403 + ], + "type": "text", + "content": "Running posterior sampling, guided by a reconstruction loss only, we observe two problems. First, for pixels with large depth values, the reconstructed image is dominated by the backscatter, making the estimation of the clean pixel color values unstable. Second, the color values can shift outside the valid range, causing color saturation. To overcome the first issue we multiply the reconstruction loss in every pixel by the estimated depth value (without passing the gradients of this operation, denoted by " + }, + { + "bbox": [ + 130, + 319, + 482, + 403 + ], + "type": "inline_equation", + "content": "sg" + }, + { + "bbox": [ + 130, + 319, + 482, + 403 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 226, + 413, + 480, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 413, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 226, + 413, + 480, + 429 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\left\\| s g (\\hat {D} _ {0}) \\cdot \\left(y - f _ {\\phi} (\\hat {x} _ {0})\\right) \\right\\| _ {2} ^ {2}. \\tag {9}", + "image_path": "02aa510e6771099a1911e5f83e66bd4f115a04057cdbc9b3f6db93d49e2413a5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "content": "In order to overcome the second issue, we introduce two auxiliary losses. The first penalizes RGB values outside the valid range " + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "content": ", and the second encourages the average values of each channel to approach the middle of the color range (in line with the gray world assumption). We implement the two auxiliary losses as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 495, + 465, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 495, + 465, + 527 + ], + "spans": [ + { + "bbox": [ + 144, + 495, + 465, + 527 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v a l}} = \\lambda_ {v} \\sum_ {i, c} \\max \\left(\\left| \\hat {J} _ {0} (i, c) \\right| - T _ {v}, 0\\right) ^ {2}, \\quad \\mathcal {L} _ {\\mathrm {a v r g}} = \\lambda_ {a} \\sum_ {c} \\left| \\sum_ {i} \\hat {J} _ {0} (i, c) - T _ {a} \\right|", + "image_path": "6b9fc03ae2d3149e016fb32e482b50dd35f898208e2f3734f64bef6eeaa3ec71.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\hat{J}_0(i,c)" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " is the value of color channel " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " of pixel " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": ", assuming the valid color range is " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "T_v" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "T_a" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " are thresholds set to values close to 1 and 0 respectively, and " + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\lambda_v,\\lambda_a" + }, + { + "bbox": [ + 130, + 538, + 482, + 574 + ], + "type": "text", + "content": " are the scalar weights of both losses. The total loss is then:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 252, + 588, + 480, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 588, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 252, + 588, + 480, + 600 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {v a l}} + \\mathcal {L} _ {\\text {a v r g}}. \\tag {10}", + "image_path": "0ddfc64d86e94b45066c339ab1b57968580db0a02b5f6cd5c13240b245c761d1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "At each iteration we compute the gradient of the loss with respect to " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", which forms the log-likelihood gradient in the posterior score (Eq. 7), and the gradient with respect to " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " which is used to update the parameters in the underwater model. In order to stabilize the optimization, we apply gradient clipping by value, to the gradients of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", and we run the optimization of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " only in some of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 138, + 130, + 373, + 379 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 302, + 129 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 302, + 129 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 302, + 129 + ], + "type": "text", + "content": "Algorithm 1: Osmosis - Sampling" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "lines": [ + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": "Input: y - Underwater Degraded Image (RGB image) " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "f(\\hat{x}_0)" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " - Underwater Physical Degradation Model Initialization of " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\phi_a,\\phi_b,\\phi^\\infty \\}" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " // size = [9] Output: " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = [\\hat{J}_0,\\hat{D}_0]" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\hat{J}_0" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " - Restored RGB Image, " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\hat{D}_0" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " - Depth Estimation \n1 " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "x_{T}\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " // " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " size = [h,w,4] \n2 for " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "t = T:1" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " do \n3 " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\epsilon \\gets \\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\mu = \\frac{1}{\\sqrt{\\alpha_t}} (x_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\cdot \\epsilon)" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " \n5 " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}} (x_t - \\sqrt{1 - \\bar{\\alpha}_t}\\cdot \\epsilon)" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " \n6 " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "f_{\\phi}(\\hat{x}_0) = \\hat{J}_0e^{-\\phi_a\\hat{D}_0} + \\phi^{\\infty}\\left(1 - e^{-\\phi_b\\hat{D}_0}\\right)" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " \n7 " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{L} =" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " loss from Eq. 10 \n8 if Optimstart " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\geq t\\geq" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " Optimend then \n9 for " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "i = 1:N" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " do \n10 | " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "\\phi \\gets \\phi -\\eta \\nabla_{\\phi}\\mathcal{L}" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " \n11 end \n12 end \n13 " + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "inline_equation", + "content": "x_{t - 1}\\sim \\mathcal{N}(\\mu -s\\cdot \\nabla_{x_t}\\mathcal{L},\\Sigma)" + }, + { + "bbox": [ + 138, + 130, + 373, + 379 + ], + "type": "text", + "content": " \n14 end" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "type": "image", + "bbox": [ + 383, + 114, + 479, + 380 + ], + "blocks": [ + { + "bbox": [ + 383, + 114, + 479, + 380 + ], + "lines": [ + { + "bbox": [ + 383, + 114, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 383, + 114, + 479, + 380 + ], + "type": "image", + "image_path": "60362b871aea45a00a5c8a56e41f258b858e99b217edf629e8f74acc0c4ae836.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 389, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 413 + ], + "type": "text", + "content": "Fig. 4: Our algorithm. [Left] Detailed steps of our algorithm. [Right] Example of how " + }, + { + "bbox": [ + 130, + 389, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\hat{J}_0, \\hat{D}_0" + }, + { + "bbox": [ + 130, + 389, + 482, + 413 + ], + "type": "text", + "content": " change during the iterations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": "the sampling steps (defined by the values " + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{Optim}_{\\mathrm{start}}" + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{Optim}_{\\mathrm{end}}" + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": "), running " + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": " iterations of gradient descent in each step. When the gradient of " + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": " is applied in the sampling step, it is multiplied by a guidance scale " + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 436, + 482, + 533 + ], + "type": "text", + "content": ". We find that using a smaller scale for the depth channel leads to better results. This makes sense as while all channels are treated the same in the prior, the depth has a different role in the forward model (specifically it is used inside an exponent), and this can lead to a different gradient scale. More details on the implementation are given in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 550, + 198, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 198, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 198, + 562 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 575, + 481, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 481, + 612 + ], + "type": "text", + "content": "We use a prior on " + }, + { + "bbox": [ + 130, + 575, + 481, + 612 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 575, + 481, + 612 + ], + "type": "text", + "content": " images, with sampling time of about 3 minutes per image on an Nvidia A100 GPU. We present here selected results and analysis. Please refer to the supplementary material for a complete set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Real-World Scenes. We present an extensive comparison on real-world linear images from the datasets SQUID [7], SeaThru [3], SeaThruNerf [32], and additional images acquired in different locations in the world, the Indian and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "Pacific Oceans, and Mediterranean, and Red Seas. Images are white-balanced as pre-processing. All runs use the same set of parameters. Fig. 1 shows several results of both image restoration and depth estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 283 + ], + "type": "text", + "content": "We compare image restoration with the following methods: CWR [24], DM [53], FUnIE-GAN [27], GDCP [41], IBLA [42], MMLE [60], semi-UIR [26], Ucolor [33], USUIR [19], UW-Net [23], waternet [35], and USe-ReDI-Net [54]. For compactness, Fig. 5 summarizes the results for a chosen subset of methods. The complete comparison is shown in the supplementary material. Among previous methods, we found GDCP to be the most consistent. Our result recovers the full range of the scenes, specifically improving contrast of objects that are further away (note the zoomed-in objects in the far areas, e.g., the diver in row 1). Our method recovers vibrant and consistent colors also in further areas. For example, in rows 2,3,4 note that the background rocks and sand appear bluish in all methods except of ours." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 481, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 481, + 356 + ], + "type": "text", + "content": "We also compare to a method that restores the RGB using a fixed depth that is pre-estimated using an external SOTA method, depth anything (DA) [58]. We use the pre-estimated depth in our method's pipeline instead of our gradually estimated depth. We term this method " + }, + { + "bbox": [ + 130, + 285, + 481, + 356 + ], + "type": "inline_equation", + "content": "DA" + }, + { + "bbox": [ + 130, + 285, + 481, + 356 + ], + "type": "text", + "content": "-Osmosis, and the results show that our method restores the far areas with better color, emphasizing the advantage of using our joint RGBD prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 481, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 405 + ], + "type": "text", + "content": "We compare depth estimation with GDCP [41], IBLA [42], unveiling [6], UW-Net [23], and monoUWnet [4], and depth anything (DA) [58]. In GDCP [41] and unveiling [6] we compute the depth from the output transmission maps. Our depth estimations have more details and better explain the scenes (Fig. 5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 405, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 481, + 453 + ], + "type": "text", + "content": "In addition, we conducted a non-reference quantitative comparison using the MUSIQ [31] measure on 50 real-world images presented in the paper and supplementary material. Our method achieved the highest score. The quantitative results are summarized in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 481, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 538 + ], + "type": "text", + "content": "The UIEB dataset [35] is sometimes used for quantitative evaluation of image restoration. However, it is collected from various sources and is not linear. Thus it is not suitable for physics-based methods like ours. In addition, the 'ground truth' of the UIEB dataset is chosen from results of previous methods and it is not the real ground truth. Fig. 6 shows four examples (out of many) where our result removes more water effects than the ground truth, even despite the non-linearity of the input." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "type": "text", + "content": "Simulation. We conducted a simulation following [26, 34] using all the 1449 images from the NYU-v2 RGBD dataset [46] with randomly varying water parameters and generated the underwater images with the image formation model in Eq. 1. For fairness in comparison with other methods we applied " + }, + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{a} = \\phi_{b}" + }, + { + "bbox": [ + 130, + 556, + 481, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "The quantitative results are summarized in Table 2. Our method substantially outperforms other methods in the image restoration metrics of PSNR, SSIM, and LPIPS. We emphasize that our prior was not trained on this data, or indoor data at all. Moreover, two of the methods [26,34] we outperform were trained on this data." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 135, + 480, + 555 + ], + "blocks": [ + { + "bbox": [ + 134, + 135, + 480, + 555 + ], + "lines": [ + { + "bbox": [ + 134, + 135, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 135, + 480, + 555 + ], + "type": "image", + "image_path": "f4b142b6cdff793a8bd591999f1024301558cf19c1d87c7a599d39925154352d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 563, + 482, + 641 + ], + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 641 + ], + "type": "text", + "content": "Fig. 5: Real-world restoration results. From left to right: white-balanced input, contrast stretch, GDCP [41], Ucolor [33], CWR [24], semi-UIR [26], DM [53], Depth Anything [58] - Osmosis, Osmosis (ours). Zoom-in colored rectangles emphasize far objects that have higher contrast in our results. Real-world depth results. From left to right: GDCP [41], IBLA [42], unveiling [6], UW-Net [23], monoUWnet [4], Depth Anything [58], Osmosis (ours). Our depth results are smoother and less affected by object gradients. The reader is encouraged to zoom-in." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 479, + 240 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 479, + 240 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 479, + 240 + ], + "type": "image", + "image_path": "9b38183a0241158bcbbb723669f1d50af5b3358eb4b5a8805773054b93d2cece.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 304 + ], + "type": "text", + "content": "Fig. 6: Example results on UIEB [35], a dataset comprised from non-linear images with ground truth produced by different enhancement algorithms. Our method yields better results than the dataset's ground truth. See for example the color of the sand, the orange color of the fish in the top-left, and the divers' authentic skin color in the bottom-right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 316, + 479, + 411 + ], + "blocks": [ + { + "bbox": [ + 135, + 316, + 479, + 411 + ], + "lines": [ + { + "bbox": [ + 135, + 316, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 135, + 316, + 479, + 411 + ], + "type": "image", + "image_path": "9a61fe0eaf08a26298a39d1d1ebb64075ba9f8dca4163305b8b53b0112cba9e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": "Fig. 7: Ablation. Ablations " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\#1, \\#5, \\#6" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": " show the importance of the losses we added to the reconstruction loss. Without " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{val}} (\\#6)" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": ", the colors tend to \"explode\" and oversaturate. Without " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{avg}} (\\#5)" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": " the colors sometimes skew towards pink/purple. In " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\#2" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": ", the further areas are not restored well, because the loss is not weighted by the depth " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\#3" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": " the guidance scale is the same for the RGB and depth channels. This harms the depth reconstruction. In " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\#4" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": " we set " + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\phi_{a} = \\phi_{b}" + }, + { + "bbox": [ + 130, + 419, + 482, + 495 + ], + "type": "text", + "content": " (separately per color channel). Since this is an inaccurate model, the restoration in further areas is harmed." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Ablation. To demonstrate the effect of different components in our methods, we conduct an ablation study of the following variants: 1. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " (instead of Eq. 10); 2. Removing weighting by " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{D}_0" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " in Eq. 9 3. Same guidance scale " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " for all channels; 4. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi_{a} = \\phi_{b}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "; 5. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}} + \\mathcal{L}_{\\mathrm{val}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " (remove " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{avg}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " from Eq. 10); 6. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{rec}} + \\mathcal{L}_{\\mathrm{avg}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " (remove " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{val}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " from Eq. 10). Numerical results on the simulation for variants " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\# 1 - \\# 3" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " are presented in Table 2. Ablations " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\# 4 - \\# 6" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " are shown only on real world images since in the simulation we do not use " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{avg}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi_{a} = \\phi_{b}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". Fig. 7 presents the results of all variants on one of the scenes. We see that the additional losses are important to prevent color saturation and shift. Increasing the loss weight with depth improves restoration in further areas. Separating guidance scales between the RGB and depth channels improves depth reconstruction. Setting " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi_{a} \\neq \\phi_{b}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " extends the range of the restoration." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 119, + 269, + 308 + ], + "blocks": [ + { + "bbox": [ + 156, + 119, + 269, + 308 + ], + "lines": [ + { + "bbox": [ + 156, + 119, + 269, + 308 + ], + "spans": [ + { + "bbox": [ + 156, + 119, + 269, + 308 + ], + "type": "table", + "html": "
MethodMUSIQ↑
input51.59
contrast stretch53.75
CWR [24]39.63
DM [53]54.64
FUnIE-GAN [27]42.70
GDCP [41]56.25
IBLA [42]54.42
MMLE [60]56.26
semi-UIR [26]54.99
Ucolor [33]47.73
USUIR [19]52.44
UW-Net [23]46.02
waternet [35]53.64
unveiling [6]50.83
DA-osmosis [58]51.23
osmosis (ours)56.62
", + "image_path": "80675c08c4d0b032b28ef1e331d6b0fb20c43d32389eccc896c2c4ffaadc94a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 308, + 291, + 352 + ], + "lines": [ + { + "bbox": [ + 132, + 308, + 291, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 291, + 352 + ], + "type": "text", + "content": "Table 1: Non-reference quantitative comparison on 50 real-world images using MUSIQ [31]. Our method achieves the highest score." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 320, + 114, + 491, + 314 + ], + "blocks": [ + { + "bbox": [ + 320, + 114, + 491, + 314 + ], + "lines": [ + { + "bbox": [ + 320, + 114, + 491, + 314 + ], + "spans": [ + { + "bbox": [ + 320, + 114, + 491, + 314 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
contrast stretch17.130.830.11
CWR [24]16.930.790.20
DM [53]17.410.820.12
FUnIE-GAN [27]17.640.770.21
GDCP [41]12.410.710.16
IBLA [42]15.070.700.19
MMLE [60]17.000.740.17
semi-UIR [26]17.820.830.12
Ucolor [33]17.920.830.10
USUIR [19]16.760.800.18
UW-Net [23]18.040.750.26
waternet [35]17.270.820.11
unveiling [6]16.340.790.18
ablation #121.090.860.09
ablation #222.170.880.07
ablation #322.000.880.06
osmosis (ours)22.740.890.06
", + "image_path": "fd23a0af7ef71ba85ff983c25bb0c33b6e362d4b27e8c4fe46f4cca0be88ddab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 318, + 315, + 478, + 346 + ], + "lines": [ + { + "bbox": [ + 318, + 315, + 478, + 346 + ], + "spans": [ + { + "bbox": [ + 318, + 315, + 478, + 346 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on the simulation. Our method achieves best scores in image restoration." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 385, + 217, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 217, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 217, + 398 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 556 + ], + "type": "text", + "content": "In this work we demonstrated how to harness the strength of a new RGBD diffusion prior to achieve state-of-the-art results on underwater image restoration. To do this we solved several challenges: i) because of lack of clean underwater image data we use datasets of scenes in air; ii) we notice that the color prior does not suffice to guide restoration, therefore, we add the scene depth to the prior; iii) We use the physical image formation model to guide restoration, and also estimate the water parameters in the process. This results in the most comprehensive single underwater image restoration method to-date. It does not train on any underwater images and therefore does not overfit to any. It gives higher weight to further objects and therefore is superior in reconstructing all the details of the scene. Jointly solving for the depth results in excellent depth estimation from monocular images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "content": "There is a domain gap between the prior (trained on in-air data) and underwater data. This is intentional. The beauty of our method is that by incorporating the underwater model in the sampling it succeeds even without training on underwater data (that is very hard to obtain). Thus, it is not specialized for a certain type of water or objects, and it learns the color distribution of natural images not degraded by water." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Like all diffusion models with U-Nets, our method is limited by a fixed resolution and long running time. This could potentially be improved using other network architectures and sampling methods." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 496, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 496, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 496, + 182 + ], + "type": "text", + "content": "Acknowledgements. The research was funded by Israel Science Foundation grant #1951/23, Israeli Ministry of Science and Technology grants #1001577600 & #1001593851, EU Horizon 2020 research and innovation programme GA 101094924 (ANERIS), the Leona M. and Harry B. Helmsley Charitable Trust, and the Maurice Hatter Foundation. We thank Dr. Derya Akkaynak and Dr. Matan Yuval for substantial data contribution, Amir Dayan for the paper's name and Meirav Keidar for graphics design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 199, + 198, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 199, + 198, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 198, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 222, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 222, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 481, + 255 + ], + "type": "text", + "content": "1. Aali, A., Arvinte, M., Kumar, S., Tamir, J.I.: Solving inverse problems with score-based generative priors learned from noisy data. arXiv preprint arXiv:2305.01166 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 255, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 277 + ], + "type": "text", + "content": "2. Akkaynak, D., Treibitz, T.: A revised underwater image formation model. In: CVPR. pp. 6723-6732 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 277, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 298 + ], + "type": "text", + "content": "3. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: CVPR. pp. 1682-1691 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "type": "text", + "content": "4. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: IEEE Int. Conf. Robotics and Automation (ICRA). pp. 1098-1104 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "text", + "content": "5. Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: CVPR. pp. 843-852 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 352, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 481, + 373 + ], + "type": "text", + "content": "6. Bekerman, Y., Avidan, S., Treibitz, T.: Unveiling optical properties in underwater images. In: ICCP (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 481, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 406 + ], + "type": "text", + "content": "7. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE TPAMI 43(8), 2822-2837 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 406, + 481, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 481, + 428 + ], + "type": "text", + "content": "8. Chan, M.A., Young, S.I., Metzler, C.A.: Sud" + }, + { + "bbox": [ + 138, + 406, + 481, + 428 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 138, + 406, + 481, + 428 + ], + "type": "text", + "content": ": Supervision by denoising diffusion models for image reconstruction. arXiv preprint arXiv:2303.09642 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 449 + ], + "type": "text", + "content": "9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 449, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 471 + ], + "type": "text", + "content": "0. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 471, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 502 + ], + "type": "text", + "content": "1. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGT0k" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 502, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 525 + ], + "type": "text", + "content": "2. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. NeurIPS 35, 25683-25696 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "text", + "content": "3. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "type": "text", + "content": "4. Chung, H., Ye, J.C.: Score-based diffusion models for accelerated mri. Medical image analysis 80, 102479 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "text", + "content": "5. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A.G., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. arXiv preprint arXiv:2305.19256 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "7. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: CVPR. pp. 9935-9946 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Feng, B.T., Smith, J., Rubinstein, M., Chang, H., Bouman, K.L., Freeman, W.T.: Score-based diffusion models as principled priors for inverse imaging. arXiv preprint arXiv:2304.11751 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "19. Fu, Z., Lin, H., Yang, Y., Chai, S., Sun, L., Huang, Y., Ding, X.: Unsupervised underwater image restoration: From a homology perspective. In: Proc. AAAI Conf. on Artificial Intelligence. vol. 36, pp. 643-651 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "20. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. Int. J. Robotics Research (IJRR) (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "21. Graikos, A., Malkin, N., Jojic, N., Samaras, D.: Diffusion models as plug-and-play priors. In: NeurIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "text", + "content": "22. Guo, L., Wang, C., Yang, W., Huang, S., Wang, Y., Pfister, H., Wen, B.: Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In: CVPR. pp. 14049-14058 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "23. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: ICIP. pp. 624-628 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "text", + "content": "24. Han, J., Shoeiby, M., Malthus, T., Botha, E., Anstee, J., Anwar, S., Wei, R., Armin, M.A., Li, H., Petersson, L.: Underwater image restoration via contrastive learning and a real-world dataset. Remote Sensing 14(17), 4297 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "26. Huang, S., Wang, K., Liu, H., Chen, J., Li, Y.: Contrastive semi-supervised learning for underwater image restoration via reliable bank. In: CVPR. pp. 18145-18155 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "27. Islam, M.J., Xia, Y., Sattar, J.: Fast underwater image enhancement for improved visual perception. IEEE Robotics and Automation Letters 5(2), 3227-3234 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "text", + "content": "28. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. NeurIPS 34, 14938-14954 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "29. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. NeurIPS 35, 23593-23606 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 460 + ], + "type": "text", + "content": "30. Kawar, B., Elata, N., Michaeli, T., Elad, M.: Gsure-based diffusion model training with corrupted data. arXiv preprint arXiv:2305.13128 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "text", + "content": "31. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5148-5157 (October 2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "32. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: CVPR. pp. 56-65 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "33. Li, C., Anwar, S., Hou, J., Cong, R., Guo, C., Ren, W.: Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE TIP 30, 4985-5000 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 579 + ], + "type": "text", + "content": "34. Li, C., Anwar, S., Porikli, F.: Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition 98, 107038 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "35. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE TIP 29, 4376-4389 (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "36. Liu, N., Zhang, N., Shao, L., Han, J.: Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "37. Lu, S., Guan, F., Zhang, H., Lai, H.: Underwater image enhancement method based on denoising diffusion probabilistic model. J. of Visual Communication and Image Representation 96, 103926 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "38. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: GibbsDDRM: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. In: Int. Conf. on Machine Learning (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "39. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "40. Peng, L., Zhu, C., Bian, L.: U-shape transformer for underwater image enhancement. IEEE TIP (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "text", + "content": "41. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE TIP 27(6), 2856-2868 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "type": "text", + "content": "42. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE TIP 26(4), 1579-1594 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 267 + ], + "type": "text", + "content": "43. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "44. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "text", + "content": "45. Saxena, S., Kar, A., Norouzi, M., Fleet, D.J.: Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "type": "text", + "content": "46. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760 (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "text", + "content": "47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Int. conf. machine learning. pp. 2256-2265 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 373, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 480, + 407 + ], + "type": "text", + "content": "48. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "text", + "content": "49. Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion models for inverse problems. In: ICLR (2023), https://openreview.net/forum?id=9GsMA8MRKQ" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "type": "text", + "content": "50. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS. vol. 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 492 + ], + "type": "text", + "content": "51. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 492, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 480, + 524 + ], + "type": "text", + "content": "52. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 524, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 480, + 556 + ], + "type": "text", + "content": "53. Tang, Y., Kawasaki, H., Iwaguchi, T.: Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In: Proc. ACM Int. Conf. Multimedia. pp. 5419-5427 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 556, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 589 + ], + "type": "text", + "content": "54. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: ICCV. pp. 12248-12258 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "55. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "56. Wei, M., Shen, Y., Wang, Y., Xie, H., Wang, F.L.: Raindiffusion: When unsupervised learning meets diffusion models for real-world image deraining. arXiv preprint arXiv:2301.09430 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Osmosis: RGBD Diffusion Prior for Underwater Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 259 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "57. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR (June 2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "58. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "text", + "content": "59. Yang, P., Zhou, S., Tao, Q., Loy, C.C.: PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In: NeurIPS (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "60. Zhang, W., Zhuang, P., Sun, H.H., Li, G., Kwong, S., Li, C.: Underwater image enhancement via minimal color loss and locally adaptive contrast enhancement. IEEE TIP 31, 3997-4010 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "61. Zhou, J., Yang, T., Zhang, W.: Underwater vision enhancement technologies: a comprehensive review, challenges, and recent trends. Applied Intelligence 53(3), 3594-3621 (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "text", + "content": "O. Bar Nathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_content_list.json b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..802d2e4d879bca36e8027457bfce7151cb40ec21 --- /dev/null +++ b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_content_list.json @@ -0,0 +1,1338 @@ +[ + { + "type": "text", + "text": "Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)", + "text_level": 1, + "bbox": [ + 240, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marko Savic1 and Guoying Zhao★1", + "bbox": [ + 362, + 212, + 638, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland {marko.savic, guoying.zhao}@oulu.fi", + "bbox": [ + 227, + 239, + 774, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Remote photoplethysmography (rPPG) is an emerging technology that can detect the pulse rate remotely from face videos. However, it is easily influenced by the recording environment, as robustness to noise is still an open problem. This vulnerability can therefore be exploited to inject fake signals or impair predictions physically. In this study we propose the first dataset containing a wide set of physical domain attack scenarios divided in three categories (illumination, movement, concealment) that directly target the main weaknesses of rPPG. We propose the Oulu rPPG Physical Domain Attacks Database (ORPDAD) as a benchmark for evaluation of robustness to physical attacks. We perform extensive experiments on conventional hand-crafted and deep learning (end-to-end, non-end-to-end, CNN, transformer, self-supervised) methods and study their susceptibility to the attacks. We conclude by discussing the most critical vulnerabilities discovered and stress the importance of designing more secure solutions. Our code and instructions for requesting the dataset (with trained models) are available at: https://github.com/marukosan93/ORPDAD/.", + "bbox": [ + 261, + 301, + 743, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Remote Photoplethysmography $\\cdot$ Facial Video $\\cdot$ Physical-Domain Attack $\\cdot$ Dataset", + "bbox": [ + 259, + 550, + 740, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 604, + 375, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Physiological signals measured via photoplethysmography (PPG) and electrocardiogram (ECG) are used to determine vital healthcare parameters such as heart rate (HR), heart rate variability (HRV), respiratory rate (RR) and oxygen saturation (SpO2). However, they are limited by the necessity for physical contact that can make them uncomfortable in long term monitoring scenarios, especially with acute skin condition patients or fragile skin neonates, and very inconvenient in natural interactions. On the other hand, Remote photoplethysmography (rPPG) is a technology that allows extracting a signal akin to the blood volume pulse (BVP) without any contact, by utilising ubiquitous RGB cameras to record facial videos. Relying on the same principle as contact PPG [2], rPPG exploits subtle variations in reflected light from the skin that are caused by the cardiac cycle [43]. Several early methods had been proposed that relied", + "bbox": [ + 212, + 633, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates corresponding author", + "bbox": [ + 217, + 825, + 441, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on hand-crafted features or blind source separation [8, 16, 28, 29, 43, 44]. However, they lacked environmental robustness and were surpassed by deep learning approaches. Most deep learning methods employed supervised learning via Convolutional Neural Network (CNN) based models [5, 7, 21, 24, 25, 37, 46, 47] or more recently, self-attention based transformer architectures [13, 20, 31, 49]. Self-supervised methods such as [11, 35, 40] have also been a recently proposed to learn without annotations.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Factors such as illumination variation, subject movement, camera sensor noise and occlusion give rise to a poor signal-to-noise ratio (SNR), making robustness to environmental noise challenging. Most rPPG methods developed in recent years try to solve this environmental noise issue, and they are evaluated on benchmark datasets designed to contain a certain degree of noise, e.g. different head motions [38], facial expressions [50], different illumination conditions [1, 15, 42], and also challenging motion and lighting in indoors [23] or in car driving scenarios [27]. A few studies also consider skin occlusions either via opaque masks [33] or translucent make-up [45]. Despite numerous methods and increasing number of datasets, robustness to environmental conditions is still an open research problem. To be widely used and applicable in real world settings, rPPG needs to be both robust and secure. From a security standpoint, lack of environmental robustness can be considered a critical vulnerability that should be addressed. Thus far, security related research in rPPG has been limited to the digital domain, as most works focus on editing the underlying physiological signals [4,41] (considered a digital attack) or de-identifying the data while retaining physiological features [30]. Digital injection of an attack is studied in [34], where a limited physical attack involving lighting is proposed.", + "bbox": [ + 212, + 252, + 785, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The lack of robustness to environmental conditions is a key issue, and we wish to draw attention to the fact that it is a significant vulnerability that attackers can potentially exploit. As rPPG is in its early stages, exposing and patching security vulnerabilities is crucial for the development of robust and secure solutions. We propose a new dataset aiming specifically at potential physical domain attacks that can either spoof the rPPG with a target signal or deteriorate it. Our novel Oulu rPPG Physical Domain Attacks Database (ORPDAD) consists of three different types of attacks (i.e. illumination, movement and concealment) that result in 26 different recording scenarios across 30 subjects. Through these attacks, we physically inject periodical signals that can spoof the rPPG (via a pulsating light and rhythmical movements) or conceal skin to deteriorate the rPPG. The purpose of this dataset is to aid in developing more secure solutions to physical domain attacks and to provide a new benchmark for rPPG robustness to physical attacks. Our contributions are as follows:", + "bbox": [ + 212, + 525, + 787, + 736 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ORPDAD is the first dataset to consider a comprehensive range of physical domain rPPG attacks, that can spoof rPPG signals or deteriorate them.", + "- Ours is the first work to introduce rhythmical movement as a potential attack, and expands greatly on illumination and concealment based attacks.", + "- We carry out an extensive comparative study for each attack's effectiveness, with experiments encompassing thirteen hand-crafted, deep learning based" + ], + "bbox": [ + 223, + 747, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervised and self-supervised rPPG methods. We analyse and share our findings on attack susceptibility and uncovered weaknesses.", + "bbox": [ + 240, + 146, + 784, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We establish a novel benchmark for physical domain rPPG attacks and share our dataset and materials (data, evaluation protocols, code and baselines).", + "bbox": [ + 225, + 176, + 784, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 232, + 390, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand-crafted rPPG: The first rPPG method [43] extracted a signal from the green channel via averaged facial video pixels in a region of interest (e.g. forehead and cheeks). Model based methods have been proposed to better discriminate rPPG signals from motion noise such as CHROM [8] and POS [44], that leverage optical/physiological considerations expressed through mathematical models. Common blind source separation approaches such as PCA [16] and ICA [29] have been applied as well. In LGI [28] a new feature space is proposed that is more robust to motion and illumination changes via differentiable local transformations. However, the assumptions and hand-crafted features of the aforementioned methods did not hold in less constrained scenarios. For example, CHROM [8] assumes a standardized skin color profile, ICA [29] assumes independence of source signals and PCA [16] assumes their uncorrelation. Deep learning has shown to be more effective, especially in less controlled scenarios.", + "bbox": [ + 212, + 265, + 787, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep learning based rPPG: The first 2D-CNN models extracted the HR from two adjacent frames, such as HR-CNN [37] and DeepPhys [5], only taking spatial information into account. End-to-end spatial-temporal 3D-CNN models such as PhysNet [46], rPPGNet [47] and TS-CAN [19] have been used to exploit the temporal information. To suppress the HR unrelated information spatial-temporal maps are introduced as input, like in non-end-to-end methods such as Niu et al. [26], RhythmNet [24], Dual-GAN [21], BVPNet [7], and CVD [25]. Recently, Efficientphys [20], PhysFormer [49], RADIANT [13] and PhySU-Net [31] have shown promising improvements by leveraging self-attention based transformers. To overcome the need for (expensive) annotated data, self-supervised methods that leverage the data alone have been proposed. They learn contrastively with artificial augmentations [11], via instance-wise sampling Contrast-Phys [40], exploit inter-data differences in a non-contrastive manner SiNC [35] or use a large set of diverse priors RS-rPPG [32]. In addition, rPPG has been widely utilised as an important face anti-spoofing cue [14, 18], as the presence of an rPPG signal can be an effective face liveness indicator.", + "bbox": [ + 212, + 462, + 787, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rPPG Datasets: Given the crucial role of data, both for evaluation and for training supervised methods, rPPG datasets have been essential in the development of new methods. The challenge of the rPPG task on a specific dataset is highly dependent on the recording environment, as less constrained (variable illumination, large motion, unstable fps, compression, etc.) scenarios lead to higher error. PURE [39] contains facial videos with PPG groundtruth, with a total of 60 one-minute-long $640 \\times 480$ 30fps videos from 10 subjects. It was recorded under ambient lighting with small motion tasks (steady, talking, slow translation, fast translation, slow rotation, medium rotation). MMSE-HR [50] has 102 videos", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(with PPG, blood pressure, RR) of 20-70s recorded with stable lighting from 40 subjects in emotion elicitation experiments, it contains spontaneous facial expressions and head motions. COHFACE [15] consists of 40 participants with 160 total 1-minute $640 \\times 480$ 20fps recordings (with PPG), where controlled studio and natural ambient lighting are used. OBF [17] contains 200 5-min $1920 \\times 1080$ 60fps RGB videos (with PPG, ECG, RR) recorded from 100 diverse subjects, under stable lighting and with minimal movement, with both resting and elevated HR conditions. UBFC-rPPG [1] is a dataset of 50 participants with 2-min $640 \\times 480$ 30fps recordings each (with PPG). It has variable indoor lighting and contains both 8 at-rest recordings and 42 recordings where participants play a time-sensitive mathematical game to increase their heart rate. VIPL-HR [23] contains 2,378 RGB videos of 20-30s length recorded from 108 subjects, it was recorded in a challenging environment with: different devices, varied and unstable frame rates, large movements and variable lighting. MR-NIRP [27] is the first rPPG dataset containing driving scenarios, it was recorded from 19 subject with 190 total $640 \\times 640$ 60fps videos (with PPG), both during driving and while parked. On top of the variable lighting, subjects were asked to speak or perform random head movements. MMPD [42] is the first dataset recorded entirely with smartphone cameras (with PPG). It includes 33 subjects with 330 one-minute videos total, recorded under different lighting conditions (LED high, LED low, incandescent, natural), and different motions such as: resting, head rotation, conversation, and walking. Currently, there are numerous rPPG datasets containing different kind of environmental noise. Nevertheless, none of these large scale datasets contain intentional periodical noise as a physical attack medium, via movement or illumination. Furthermore, skin concealment via opaque cloth or transparent make-up is also not present in the aforementioned datasets.", + "bbox": [ + 212, + 145, + 787, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Security in rPPG: As rPPG deals with sensitive biometric facial videos and physiological signals, security and privacy based research has also been a focus in recent years. Protecting privacy by removing identity related features while preserving physiological features has been attempted first via a Gaussian decomposition that maintains the low frequency spatial features [3]. A visual utility preserving method has also been proposed, that removes identity related features but preserves the physiological and visual utility via a semi-adversarial framework [30]. In contrast to rPPG preservation, PulseEdit [4] and PrivacyPhys [41] were devised to alter the rPPG signal in videos by using small perturbations. The original purpose of the aforementioned works is to conceal identity or the physiological signals, but they can be repurposed to alter the rPPG signal digitally and thus employed as a digital attacks that alter the underlying physiological data to result in a target signal. However, all these works are limited to the digital domain. In [34] digital-domain rPPG attacks are explicitly studied by injecting a target periodical signal. Furthermore, the authors propose a simple pulsating single LED attack by collecting a small dataset. However, most attack frequencies were outside the normal resting heart rate (filtered out during post-processing), the data contained only five two-minute videos from four subjects, attack intensity and range were limited, and the attack was a simple sinusoidal", + "bbox": [ + 212, + 539, + 787, + 825 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "wave. A similar attack was proposed in MSPM [36] involving pulsating light from a monitor. Apart from illumination variation, motion is also a large noise source that datasets try to incorporate in varying degrees. However, no studies have yet tried to inject a specific target signal by utilising periodical motion as a medium. Finally, as rPPG is measured from light reflected from the facial skin, concealing it can be considered as physical domain attack with the aim to deteriorate the signal. In [33], the authors propose the MPM dataset containing 159 3-minute recordings of 54 subjects wearing facial masks, and show the importance of having methods robust to face occlusions, but only consider opaque face masks. In [45] foundation is considered as a translucent mask with 22 subjects, as make-up can affect rPPG readings with a notable decrease in SNR, but the authors only create one scenario involving foundation alone. Current rPPG solutions are not robust to noise and struggle in challenging scenarios, their robustness to physical domain attacks lacks sufficient investigation, as current physical domain attack research is limited in scope and only considers specific instances.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Oulu rPPG Physical Domain Attacks Database", + "text_level": 1, + "bbox": [ + 214, + 411, + 710, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As rPPG is an emerging technology that deals with sensitive data and could be employed in applications where security is critical (clinical, remote driver monitoring, anti-spoofing, authentication, etc.), it is crucial to critically assess its vulnerabilities early and work towards mitigating potential security exploits. In this work, we focus on the scarcely studied physical-domain attacks. A physical-domain attack assumes that the attacker cannot influence the digital data after acquisition or the rPPG method used for estimation, namely the attacker can only inject noise during the acquisition process using physical means. We are inspired by previous works [33, 34, 36, 45] that deal with a subset of potential attacks in a limited scope. To fully understand and mitigate potential physical attacks to rPPG, we propose the ORPDAD dataset that aims to study a large set of diverse physical attacks within the same recording environment, thus isolating the variable of interest associated with the specific attack. The attacks we propose are simple to execute, but can have a large influence on the rPPG estimation.", + "bbox": [ + 212, + 444, + 787, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Motivation and reasons for setup", + "text_level": 1, + "bbox": [ + 214, + 693, + 535, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We identify vulnerabilities based on the properties of rPPG measuring, in Fig. 1 we display three major vulnerabilities. Firstly, as rPPG is measured from reflected light, a possible attack vector is the light source. Adding a weak perturbation signal to the incident light source means that it will be present in the estimated rPPG signal as well. Secondly, subject movement induces noise to the extracted signal, consequently having the subject move at a set rhythm will make this added motion perturbation strongly periodical. Therefore, we can leverage both motion and illumination to inject strongly periodical (rPPG like) signals", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2568bf67fec847deda912322786522e2599299977088f778485d85c2d9b5e350.jpg", + "image_caption": [ + "Fig. 1: (Left) Major rPPG measurement environment vulnerabilities and related potential attack vectors. (Right) Data recording setup with equipment as listed in Tab. 1" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 557, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/793d37da45e10660132ca2ed54f4d0ab5cd646506741d2456b546be52fcdb4ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 142, + 785, + 313 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa32596eb79328c39dc8c0f5607cf2aedbc81a04974d4242d9d242e5eead67a2.jpg", + "table_caption": [ + "Table 1: Equipment used for data collection, also shown in Fig. 1 (Right)" + ], + "table_footnote": [], + "table_body": "
Device/ToolDescription/Specifications
1) RGB CameraBlackmagic Ursa Mini 4K EF1920x1080 30fps 500Mbps f8.7 360° ISO400 WB4000K fl24mm
2) 2×Studio lightsAputure LightStorm LS 1c 3500lux 4000K
3) 2×Custom LED panels1600lm 11.5W 4000K (96×2835SMD LED) 20×13.3cm²
4) LED Control CircuitRPi 4B + IRL540N MOSFET + Resistors + 12V3A PSU
5) Finger-oximeter PPGBeurer PO80 120Hz
6) ECGBittium Faros 180 1Khz 1ch
7) MetronomeGoogle Metronome Web App
8) Concealment (C1-C3)Blue Medical Mask and Black Beanie Hat
9) Concealment (C4-C6)Maybelline Fit Me Matte Poreless \nPrimer + Foundation 220 + Powder 220
", + "bbox": [ + 240, + 391, + 763, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "into the data. Thirdly, the light measured is reflected from the skin, so either blocking the light (with an opaque cloth) or attenuating it (with make-up) can affect the rPPG negatively. Consequently, we define these three physical-domain attack categories: Illumination (I), Movement (M) and Concealment (C).", + "bbox": [ + 212, + 565, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Study design", + "text_level": 1, + "bbox": [ + 214, + 648, + 370, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We record data from 30 healthy (no history of heart disease) volunteers under different scenarios based on the three aforementioned categories of attacks. All the participants gave written informed consent. For the active attacks we choose $50~\\mathrm{bpm}$ and $100~\\mathrm{bpm}$ as attack frequencies as they can be considered borderline resting heart rates, thus plausible but different from the real HR in most recordings. We avoid results being biased by the real HR trend by averaging results over both frequencies. We record 70 second videos for each scenario, for a total of 26 scenarios per subject, resulting in 780 videos equivalent to $\\approx 15.2\\mathrm{h}$ of footage. For the synchronised ground truth physiological signals, we record both PPG waveforms via a finger-oximeter and ECG waveforms via an ECG device. Based on the defined attack categories, we define the following scenarios:", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81e662409a576a7160298d48a912ddf8f99f101b4c74ebf23d4bba2e4568f40d.jpg", + "image_caption": [ + "Fig. 2: Raw signal (GREEN [43]) and its spectrogram for diverse attacks. For active illumination (I1, I6) and motion based attacks (M1, M9), the target attack frequency is clearly visible in the spectrogram and can take over the prediction. On the other hand, passive concealment attacks (C1, C6) significantly affect the estimation's SNR." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 480, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04fb62e317ab8ff284ecdd5d2f5174836b124212928b2d92068f754c2741cb61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 146, + 784, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Still (S1-S3): Subject is recorded at a still and resting state before each attack category. Motion is kept at a comfortable minimum.", + "bbox": [ + 212, + 414, + 784, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Illumination (I1-I6): We combine the two attack frequencies (50bpm, 100bpm) and three different intensity settings (S=7%, M=14% and L=21% of the attack device max brightness), resulting in I1=50S, I2=50M, I3=50L, I4=100S, I5=100M, I6=100L. The intensities are much weaker than the base lighting, and cause perturbations are visually imperceptible in the recordings.", + "bbox": [ + 212, + 444, + 807, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Movement (M1-M11): We explore simple movements such as small vertical (SV), large vertical (LV), small horizontal (SH), large horizontal (LH), mouth open/close (M) at the two target frequencies (50bpm, 100bpm). Resulting in M1=50SV, M2=50LV, M3=50SH, M4=50LH, M5=50M, M6=100SV, M7=100LV, M8=100SH, M9=100LH, M10=100M. We also include natural talking motion M11 that is not at a set attack frequency.", + "bbox": [ + 212, + 520, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concealment (C1-C6): The first three opaque (C1-C3) scenarios involve wearing a forehead concealing beanie C1, a facial mask C2 and then both C3. The transparent concealment scenarios (C4-C6) contain progressive application of make-up with primer C4, foundation C5 and setting powder C6.", + "bbox": [ + 212, + 611, + 787, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 2 we visualise a subset of the aforementioned attacks and their effect on the raw signal from the GREEN [43] method.", + "bbox": [ + 212, + 671, + 787, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Equipment and setup", + "text_level": 1, + "bbox": [ + 214, + 724, + 439, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To study each attack vector in an isolated manner, we set up a controlled recording environment as to minimise the influence of external factors that are not related to the attack. Thus, we keep the same base environment for all subjects and scenarios, enabling us to study the specific contribution induced by the attacks. For the base video recordings, we use a professional RGB camera and two fixed LED lights. We record the physiological signals with a finger PPG oximeter", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and a portable ECG belt device. All the recording devices are synchronised by keeping internal clock times consistent. The subjects are seated at $\\approx 1\\mathrm{m}$ from the camera and the lights are positioned above the camera and illuminate the subject at a $45^{\\circ}$ angle at $\\approx 1.5\\mathrm{m}$ distance.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the illumination attacks, we design two modulated light panels that each operate at $1600\\mathrm{lm}$ maximum, by using LED strips and a control circuit. These panels allow the injected light signal to be more uniformly distributed over the face surface, instead of attacking only part of it (single led case). We pilot the panels with Pulse Width Modulation (PWM) signals, thus being able to create a light source that can output any waveform and at any fraction of the max intensity. For this experiment, we simulate a typical rPPG waveform, with the main frequency component placed at the target frequency and a second harmonic at $1/4$ of the peak intensity. For the L, M, H settings we use $7\\%$ , $14\\%$ , $21\\%$ of the max power and set the lights at $\\approx 1\\mathrm{m}$ distance from the subject. For the movement scenarios, we simply utilise an audio metronome and instruct the subjects to move according to the set rhythm and movement type. For opaque concealment, we use medical face masks and black beanie hats. For translucent concealment we choose a primer, foundation and setting powder, and keep the brand and shade (220 natural beige) consistent among all subjects.", + "bbox": [ + 212, + 207, + 785, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide a detailed list of all the equipment and settings utilised for recording in Tab. 1, and show all the equipment in Fig. 1 (right).", + "bbox": [ + 212, + 433, + 785, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 487, + 375, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assess the effects of each physical attack on a large set of rPPG methods that represent the past and current state of the art, as they include both common hand-crafted and CNN based methods, as well more recent transformer based and self-supervised works. We choose the following hand-crafted non-learning based methods: GREEN [43], CHROM [8], POS [44], LGI [28], ICA [29] and PCA [16]. As for learning based CNN methods, we choose the end-to-end methods DeepPhys [5], TS-CAN [19] and PhysNet [46], and the non-end-to-end method CVD [25]. We also include two transformer based methods that are Physformer [49] (end-to-end) and PhySU-Net [31] (non-end-to-end). Finally, we include an end-to-end self-supervised method Contrast-Phys [40]. We faithfully re-implement the aforementioned methods and fine-tune them to obtain the best performance on our dataset. All our code can be found at https://github.com/marukosan93/ORPDAD/, with instructions for requesting the dataset (containing both raw data, preprocessed data and trained models).", + "bbox": [ + 212, + 518, + 787, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 753, + 426, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Preprocessing: Firstly, we utilise PyFeat [6] with RetinaFace [9] model for face detection and PFLD [12] model for landmark localisation, the landmarks are then stabilised with a 5-point moving average filter. For the end-to-end methods, we crop the images to $128 \\times 128$ so that the face is centered. For hand-crafted", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods we use the landmarks to define a ROI (including forehead, cheeks and chin) and average the pixels to the raw signal input. Finally, as both non-end-to-end methods CVD [25] and PhySU-Net [31] utilise multi-scale spatial-temporal maps (MSTmaps) as input, we follow the suggested processing [25] to generate the input MSTmaps. For the ground-truth signal, as ECG is considered the gold standard, we choose to primarily use ECG for experimentation. We first filter the ECG with the method in [10] via NeuroKit2 [22], and then apply the 10-point average method from [48] to get a resulting signal containing the same physiological features but with a shape better correlated to the rPPG signal. All ground-truth signals, original videos and pre-processed videos (landmarks, MSTmaps, cropped) will be included in the shared dataset version.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics: We follow previous works by using mean absolute error (MAE), root-mean-square error (RMSE), and Pearson's correlation coefficient (R). Furthermore, we calculate the SNR, defined as the ratio between the power in the dominant frequency and second harmonics versus the remaining power [8]. Note that good SNR does not coincide with accurate prediction, as it measures the spectrum sparsity around the predicted rPPG (without considering the true label). For active attacks, we also calculate success rate, defined as the percentage of times when the predicted HR is closer to the two attack frequencies than the ground truth, meaning that the attack successfully hijacked the prediction.", + "bbox": [ + 212, + 311, + 787, + 446 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Protocol: To ensure fair validation we adopt a 5-fold subject exclusive protocol for all experiments. Additionally, we consider two different training conditions. Attack Unaware (AU), where the learning based methods are trained only on the still (S) scenarios and are not exposed to attacks during training, but only during testing. Attack Aware (AA), in this case, the training data contains the attack scenarios so that we can assess whether awareness improves robustness. For AA, we add an attack frequency exclusive cross-validation to the protocol, we train on all passive scenarios and active scenarios with one attack frequency, and then test on the non-seen attack frequency. As the attack frequencies are fixed, supervised methods could easily cheat by learning to not predict the seen specific attack frequency, without learning any useful features from the attacks. Our AA protocol is aware of the attacks, but not of the specific attack frequency. In testing 30 second non-overlapping segments are evaluated, and we calculate the HR by finding the highest PSD peak.", + "bbox": [ + 212, + 449, + 787, + 660 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 680, + 441, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We analyse each attack category separately, regarding the signal deterioration in terms of accuracy and SNR. Subsequently, we study the success rate of the active attacks.", + "bbox": [ + 212, + 704, + 785, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Firstly, we evaluate all the studied methods on the illumination scenarios, as shown in Tab. 2. On the base scenarios (S1-S3) almost all methods (except GREEN [43]) have very good performance ( $\\leq 1$ RMSE) as the base task is not very challenging by design, due to the controlled environment. It stands out noticeably that hand-crafted methods are very susceptible to the illumination attack as most methods have a large drop in accuracy, with only the skin", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/25ee8e94706d714e1ebdb9e916c4a2c61ceff14d3595b21b0ebb0d890e8e4d3d.jpg", + "table_caption": [ + "Table 2: Results on illumination attacks I1-I6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + ], + "table_footnote": [], + "table_body": "
RMSE↓ |SNR↑S1-S3 (still)I1+I4 (small)I2+I5 (med.)I3+I6 (large)
GREEN [43]□8.57 -7.425.7 -6.727.9 -4.127.7 -2.5
CHROM [8]□0.91 -3.71.87 -4.713.3 -4.419.5 -4.3
POS [44]□0.70 -3.45.77 -4.411.4 -4.813.8 -4.8
LGI [28]□1.05 -3.58.16 -4.420.7 -4.225.9 -3.3
ICA [29]□1.14 -2.922.3 -3.727.6 -1.427.7 0.2
PCA [16]□1.35 -5.110.5 -6.215.6 -6.513.1 -7.2
DeepPhys [5]◆AU1.02 -2.75.53 -3.922.6 -3.426.8 -2.4
AA0.79 -2.80.63 -3.78.23 -4.013.9 -4.4
TS-CAN [19]◆AU1.16 -3.66.31 -4.723.5 -4.127.4 -3.1
AA0.88 -3.50.76 -3.91.09 -4.25.90 -4.5
PhysNet [46]◆AU0.88 -3.14.58 -4.213.2 -4.616.1 -4.7
AA0.55 -2.80.51 -3.10.46 -3.51.61 -4.1
CVD [25]◇AU1.03 -3.11.51 -3.95.47 -4.67.17 -5.2
AA0.79 -2.41.10 -2.70.95 -3.04.08 -3.4
Physformer [49]◆AU0.95 -3.01.01 -3.810.0 -4.514.8 -4.6
AA1.34 -2.41.59 -2.70.57 -3.17.51 -3.8
PhySU-Net [31]◇AU0.74 -1.90.62 -2.95.71 -3.76.24 -4.0
AA0.50 -1.80.93 -2.11.41 -2.63.15 -2.7
Contrast-Phys [40]▼AU0.95 -2.81.51 -3.813.9 -4.317.2 -4.4
AA20.1 -3.620.9 -3.721.6 -4.121.9 -4.5
", + "bbox": [ + 305, + 198, + 697, + 486 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "model based CHROM [8] and POS [44] having comparable performance only with the weakest attack (I1+I4). Moreover, decrease in accuracy does not necessarily mean decrease in SNR, e.g. ICA [29] improves in SNR by having a higher intensity attack, as the output spectrum contains more power in the dominant attack frequency. For supervised deep learning methods there is an overall better degree of innate robustness as the learned features are more complex and physiology specific. However, in the AU case, except for the weakest setting (I1+I4), the supervised methods are significantly affected by the attack. The illumination based attacks affect non-end-to-end methods CVD [25] and PhySU-Net [31] less, as the spatial-temporal map inputs are more resistant to the illumination variation. When adding the attacks into the training data (AA), the performance significantly improves for the supervised methods, as the networks learn to ignore the features specific to the attack, showing the important role of new data in protection against physical attacks. On the other hand, self-supervised Contrast-Phys [40] is very receptive to the attacks as it focuses on salient frequency features without supervision. In the AA case, the learned features lose their physiological meaning as the method learns from inter-data differences that are most pronounced between the attacks, leading to inaccurate predictions even for the non-challenging still scenarios.", + "bbox": [ + 212, + 518, + 787, + 806 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Secondly, we study the movement based attacks as shown in Tab. 3. The hand-crafted methods all have a significant performance drop, that is closely cor-", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/43ce97e7aeafb570fc04f5fdf47dc57e291018a02abdae9e575f9e73137ef45a.jpg", + "table_caption": [ + "Table 3: Results on movement attacks M1-M11. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + ], + "table_footnote": [], + "table_body": "
RMSE↓ |SNR↑S1-S3(still)M1+M6(s. ver.)M2+M7(l. ver.)M3+M8(s. hor.)M4+M9(l. hor.)M5+M10(mouth)M11(talk)
GREEN [43]□8.57 -7.426.6 -4.426.8 -2.823.5 -6.731.0 -5.626.9 -5.223.1 -8.7
CHROM [8]□0.91 -3.710.4 -4.911.5 -5.86.34 -5.019.3 -5.910.6 -5.51.35 -6.8
POS [44]□0.70 -3.48.60 -4.813.8 -5.78.22 -4.814.5 -6.110.3 -5.41.20 -6.5
LGI [28]□1.05 -3.518.0 -4.926.1 -3.811.4 -4.924.2 -5.112.7 -5.35.45 -6.7
ICA [29]□1.14 -2.925.5 -3.324.9 -2.721.8 -4.726.3 -4.123.4 -4.03.99 -6.4
PCA [16]□1.35 -5.113.4 -5.416.6 -7.314.5 -5.923.9 -6.413.5 -5.76.13 -7.1
DeepPhys [5]◆AU1.02 -2.719.4 -4.626.3 -3.523.7 -4.625.1 -4.717.2 -4.95.65 -6.6
AA0.79 -2.86.16 -3.915.9 -5.115.0 -4.620.6 -5.39.38 -4.17.47 -5.4
TS-CAN [19]◆AU1.16 -3.624.2 -4.526.8 -3.224.4 -5.326.0 -4.819.8 -5.01.79 -7.2
AA0.88 -3.55.99 -4.318.5 -5.315.2 -5.021.0 -5.811.7 -4.71.45 -5.7
PhysNet [46]◆AU0.88 -3.117.1 -5.022.6 -4.99.45 -5.320.6 -6.38.86 -4.56.21 -5.0
AA0.55 -2.83.61 -3.311.5 -4.43.62 -3.79.11 -5.50.71 -3.63.37 -4.2
CVD [25]◇AU1.03 -3.110.6 -5.516.8 -7.46.62 -4.211.2 -6.15.23 -5.18.74 -4.9
AA0.79 -2.44.64 -3.412.9 -5.80.83 -2.75.96 -4.52.97 -3.66.30 -3.7
Physformer [49]◆AU0.95 -3.011.0 -4.318.3 -5.012.0 -4.820.5 -5.78.55 -4.68.68 -5.2
AA1.34 -2.40.69 -2.610.8 -4.07.30 -3.111.7 -4.85.94 -3.25.67 -3.9
PhySU-Net [31]◇AU0.74 -1.95.93 -4.215.2 -6.53.58 -3.99.48 -6.20.69 -3.67.71 -4.2
AA0.50 -1.81.31 -2.39.17 -3.80.69 -2.27.04 -4.20.62 -2.27.56 -3.1
Contrast-Phys [40]▼AU0.95 -2.815.0 -4.821.6 -4.66.84 -4.918.1 -5.99.54 -4.63.41 -5.0
AA20.1 -3.625.5 -4.627.1 -5.827.8 -5.228.8 -5.422.3 -4.424.1 -5.0
", + "bbox": [ + 215, + 198, + 785, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "related to the movement intensity, as larger movements yield larger errors. There is a contrast between the strongly periodical mouth active attack (M5+M10) and the normal talking motion (M11), as the performance is better when the noise is not periodical (e.g. CHROM [8] and POS [44] have comparable performance to the still scenarios), even though the SNR is lower signifying a noisier spectrum. All the supervised deep learning methods are severely affected by the attack, with non-end-to-end PhySU-Net [31] and CVD [25] having lower accuracy drops overall. Similarly to the illumination attacks, in the AA case the performance significantly improves for supervised methods, as the networks learn to ignore attack specific features to a certain degree. On the other hand, for the AU susceptible Contrast-Phys [40] the addition of attack awareness is even aggravating, similarly as in the illumination case.", + "bbox": [ + 217, + 520, + 787, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Thirdly, we study the concealment based attacks as shown in Tab. 4 and visualise the SNR changes in Fig. 3. In this case, hand-crafted methods obtain variable performance on partial concealment (C1 and C2), with C2 having strong drops in accuracy for LGI [28], CHROM [8] and PCA [16]. This is due to the medical mask covering a larger surface of the ROI than the forehead covering hat. In case of total coverage C3, due to the small portion of uncovered skin the task is extremely challenging and leads to low accuracy. There is an overall decrease in SNR when applying the concealment, as the physiologically irrelevant surface is larger. In the make-up related attacks (C4-C6), the performance only", + "bbox": [ + 217, + 704, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b218a268a7696c855f1acf0281e83bef019ab2045bea4ab2615f62a32a4cfa09.jpg", + "table_caption": [ + "Table 4: Results on concealment attacks C1-C6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + ], + "table_footnote": [], + "table_body": "
RMSE↓|SNR↑S1-S3(still)C1(hat)C2(mask)C3(both)C4(primer)C5(found.)C6(powder)
GREEN [43]□8.57 -7.414.0 -8.718.9 -9.121.9 -9.29.46 -7.712.1 -8.813.7 -8.7
CHROM [8]□0.91 -3.70.72 -5.311.1 -7.817.6 -9.50.87 -4.11.16 -5.20.95 -5.3
POS [44]□0.70 -3.40.69 -4.70.96 -6.023.7 -10.0.57 -3.50.94 -4.60.80 -4.6
LGI [28]□1.05 -3.50.70 -4.914.1 -8.722.0 -9.60.84 -3.61.13 -4.90.92 -4.9
ICA [29]□1.14 -2.90.81 -3.81.03 -5.217.7 -9.20.75 -3.01.05 -4.40.87 -4.2
PCA [16]□1.35 -5.11.37 -5.424.1 -9.422.0 -9.01.08 -4.87.78 -6.08.32 -6.1
DeepPhys [5]AU1.02 -2.70.77 -3.415.3 -9.418.9 -9.30.76 -2.80.84 -3.60.75 -3.4
AA0.79 -2.80.82 -3.31.18 -4.00.85 -5.80.73 -2.80.80 -3.80.81 -3.7
TS-CAN [19]◆AU1.16 -3.61.05 -4.918.4 -9.017.7 -8.90.77 -3.70.91 -4.91.21 -4.5
AA0.88 -3.50.93 -4.01.48 -4.43.35 -6.10.76 -3.41.01 -4.81.38 -4.7
PhysNet [46]◆AU0.88 -3.10.50 -3.30.67 -4.514.8 -7.60.71 -3.10.76 -4.00.60 -3.7
AA0.55 -2.80.50 -3.10.51 -3.20.79 -3.90.54 -2.80.60 -3.60.52 -3.3
CVD [25]◇AU1.03 -3.11.24 -3.57.07 -7.510.4 -8.57.76 -3.11.05 -3.70.75 -3.8
AA0.79 -2.40.82 -2.91.16 -3.88.43 -7.10.87 -2.31.03 -2.90.73 -3.0
Physformer [49]◆AU0.95 -3.00.94 -3.410.8 -6.616.1 -7.90.71 -3.00.68 -3.90.66 -3.6
AA1.34 -2.41.87 -2.81.09 -3.12.33 -3.90.54 -2.40.69 -3.20.59 -2.8
PhySU-Net [31]◇AU0.74 -1.90.58 -2.86.47 -7.218.7 -8.70.67 -1.90.57 -3.00.55 -2.7
AA0.50 -1.80.62 -2.30.84 -2.810.3 -6.70.62 -1.80.63 -2.50.51 -2.1
Contrast-Phys [40]▼AU0.95 -2.80.73 -3.413.8 -8.713.6 -9.00.81 -2.80.72 -4.30.86 -4.4
AA20.1 -3.618.9 -3.437.0 -6.335.8 -6.020.1 -3.418.3 -3.616.7 -3.5
", + "bbox": [ + 217, + 198, + 785, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "significantly drops for GREEN [43] and PCA [16] that are more susceptible to the drop in SNR. For the fully transparent primer C4 (optically similar to a facial creme), the drop in SNR is negligible not significantly affecting the rPPG signal. For the foundation C5 and powder C6 scenarios, the drop in SNR is very noticeable ( $\\approx 1dB$ on average) due to the changed optical properties of the skin. Between C5 and C6 (same shade, but C6 has an additional layer) the difference in negligible, meaning that with regards to the rPPG the shade is most relevant. Similarly, for AU supervised deep learning methods the facial mask is more damaging to the accuracy of non-end-to-end methods, as they rely on the ROI only, and C3 is too challenging of a task. For make-up the behaviour is analogous to the hand-crafted case, with C4 being negligible, and C5 and C6 having a similar overall SNR drop ( $\\approx 1dB$ ). In the AA case, all methods can adapt to the partial attack (C1, C2), and surprisingly end-to-end methods are even able to obtain good performance on the challenging C3 scenario, as they are able to find alternative physiologically relevant patches of skin. Regarding make-up C5 and C6, in the AA case SNR increases noticeably ( $\\approx 0.5dB$ ) as the networks partially mitigate the change in skin shade. As for the self-supervised Contrast-Phys [40], it is vulnerable to C2 and C3 in the AU case. Analogously to the previous two attack categories, in the AA case the features decline in physiological meaning.", + "bbox": [ + 215, + 537, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/65f411554251c62ee182cc2713a978effde1b57c30c8448ded6477c04a751200.jpg", + "image_caption": [ + "Fig. 3: SNR for passive concealment attacks (C1-C6) and success rates for active attacks (I1-I6, M1-M10). Higher SNR and lower success rates (lighter color intensity) correspond to better robustness." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 750, + 383 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b46a47b0c8a2817e0f4c3890db7f47d3098084324a4bc87fcf00714fd50a584f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 146, + 782, + 315 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lastly, we also analyse the success rate of the active attacks, as shown in Fig. 3, using the success rate metric (from Sec. 4.1) we can effectively determine with what rate the attacks overtake the real predictions. In line with the accuracy, we can see that CHROM [8] and POS [44] are the overall the more robust hand-crafted methods. The non-end-to-end (PhySU-Net [31] and CVD [25]) AU supervised methods overall are more robust with the illumination and motion attacks, with large vertical movement as the most challenging attack. The AA case makes the success rate significantly lower. In line with the accuracy, for the non-robust Contrast-Phys [40] the AA case seems to worsen, as the addition of attacks in the training data has the negative effect of increasing noise related inter-data differences that guide the method to learn non-physiological features.", + "bbox": [ + 212, + 469, + 787, + 636 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Discussion", + "text_level": 1, + "bbox": [ + 215, + 660, + 349, + 674 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our experiments benchmark the current state of rPPG in relation to physical domain attacks, in hand-crafted, supervised and self-supervised settings (AU, AA). Active attacks (I1-I6, M1-M10) have a strong impact on hand-crafted methods as they are less robust to environmental noise, with CHROM [8] and POS [44] being less susceptible as they are based on physiology specific mathematical models. Supervised methods offer more resistance to the attacks, as they are guided by strong supervision to ignore non-physiological features and learn features strongly correlated to the ground truth, thus are able to offer better robustness. Non-end-to-end methods (CVD [25], PhySU-Net [31]) using signal map inputs show the best robustness out of supervised methods, due to the noise", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "suppression capabilities of the signal maps. However, non-end-to-end methods depend on face ROI locations, making them less effective with Movement attacks compared to Illumination. For the end-to-end methods, DeepPhys [5] and TS-CAN [19] are less robust than Physnet [46] and Physformer [49], due to their short temporal context architectures. Having the attacks present during training (AA) greatly improves robustness of supervised methods. Most methods follow a similar trend to the AU case, as the innate robustness is supplemented by the data awareness, meaning that the same design factors lead to effectiveness in the AA case as in the AU case. Self-supervised methods (ContrastPhys [40]) on the other hand are very susceptible to attacks, both in the AU case and especially in the AA case, as the attacks negatively influence the training. Opaque Concealment attacks (C1-C3) are more effective on ROI dependant methods such as hand-crafted and non-end-to-end, with the facial mask being more aggravating as it covers a larger physiologically relevant surface, and C3 (mask and hat) being too challenging of a task, except for end-to-end methods in AA case. Regarding make-up (C4-C6), the transparent primer has negligible effect, and that addition of foundation leads to a prominent drop in SNR, but further addition of powder is negligible, meaning that shade is more impactful. Most supervised methods partially adapt to the make-up and have improved SNR in the AA case. The addition of attacks into the training data significantly boost the performance as the networks learn to ignore the specific features related to the attacks (that are not correlated with the ground truth), thus showing the strong necessity of having such physical attacks present in the training set. Nonetheless, depending only on data to defend from such attacks is not feasible and future rPPG methods need to be more secure and innately robust to physical attacks.", + "bbox": [ + 212, + 147, + 787, + 525 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 547, + 359, + 564 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose ORPDAD, the first rPPG dataset to include comprehensive physical domain attacks. We extensively evaluate thirteen diverse rPPG methods on the attack scenarios. We point out that hand-crafted and self-supervised methods are most susceptible, supervised methods (particularly non-end-to-end) have a certain degree of innate robustness that can be improved with data containing attacks. However, only having more data cannot mitigate this issue, as secure rPPG methods need to be designed to overcome these vulnerabilities. Thus, we hope that our dataset will encourage novel secure solutions. Future work includes the exploration of new robust methods against general physical attacks, and the extension of the study and dataset to include face spoofing attacks, as the studied attacks have the potential to deceive liveness detection based anti-spoofing.", + "bbox": [ + 212, + 579, + 787, + 746 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 768, + 375, + 784 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by the Research Council of Finland Academy Professor project EmotionAI (grants 336116, 345122) and ICT 2023 project TrustFace (grant 345948), and the University of Oulu & Research Council of Finland Profi", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 (grant 352788), and by Infotech Oulu. As well, the authors acknowledge CSC - IT Center for Science, Finland, for computational resources. We also thank Mengting Wei and all the study volunteers for their contribution.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 214, + 321, + 229 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019)", + "2. Challoner, A., Ramsay, C.: A photoelectric plethysmograph for the measurement of cutaneous blood flow. Physics in Medicine & Biology 19(3), 317 (1974)", + "3. Chen, M., Jayaweera, A., Wong, C.W., Wu, M.: Identity-Privacy Protection for Facial-rPPG Based Smart Health Research (Jul 2022). https://doi.org/10.36227/techrxiv.20352843.v1, https://www.techrxiv.org/articles/preprint/Identity-Privacy_Protection_for_Facial-rPPG-Based_Smart_Health_Research/20352843/1", + "4. Chen, M., Liao, X., Wu, M.: Pulseedit: Editing physiological signals in facial videos for privacy protection. IEEE Transactions on Information Forensics and Security 17, 457-471 (2022)", + "5. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 349-365 (2018)", + "6. Cheong, J.H., Xie, T., Byrne, S., Chang, L.J.: Py-feat: Python facial expression analysis toolbox. CoRR abs/2104.03509 (2021), https://arxiv.org/abs/2104.03509", + "7. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021). pp. 01-08. IEEE (2021)", + "8. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering 60(10), 2878-2886 (2013)", + "9. Deng, J., Guo, J., Ververas, E., Kotsia, I., Zafeiriou, S.: Retinaface: Single-shot multi-level face localisation in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5203-5212 (2020)", + "0. Elgendi, M., Jonkman, M., De Boer, F.: Frequency bands effects on qrs detection. Biosignals 2003, 2002 (2010)", + "1. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021)", + "2. Guo, X., Li, S., Yu, J., Zhang, J., Ma, J., Ma, L., Liu, W., Ling, H.: Pfld: A practical facial landmark detector. arXiv preprint arXiv:1902.10859 (2019)", + "3. Gupta, A.K., Kumar, R., Birla, L., Gupta, P.: Radiant: Better rppg estimation using signal embeddings and transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4976-4986 (2023)", + "4. Hernandez-Ortega, J., Fierrez, J., Morales, A., Tome, P.: Time analysis of pulse-based face anti-spoofing in visible and nir. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-552 (2018)", + "5. Heusch, G., Anjos, A., Marcel, S.: A reproducible study on remote heart rate measurement. arXiv preprint arXiv:1709.00962 (2017)" + ], + "bbox": [ + 225, + 244, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Lewandowska, M., Rumiński, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: 2011 federated conference on computer science and information systems (FedCSIS). pp. 405-410. IEEE (2011)", + "17. Li, X., Alikhani, I., Shi, J., Seppanen, T., Junttila, J., Majamaa-Voltti, K., Tulppo, M., Zhao, G.: The obf database: A large face video database for remote physiological signal measurement and atrial fibrillation detection. In: 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018). pp. 242-249. IEEE (2018)", + "18. Liu, S., Yuen, P.C., Zhang, S., Zhao, G.: 3d mask face anti-spoofing with remote photoplethysmography. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 85-100. Springer (2016)", + "19. Liu, X., Fromm, J., Patel, S., McDuff, D.: Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems 33, 19400-19411 (2020)", + "20. Liu, X., Hill, B., Jiang, Z., Patel, S., McDuff, D.: Efficientphys: Enabling simple, fast and accurate camera-based cardiac measurement. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5008-5017 (2023)", + "21. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12404-12413 (2021)", + "22. Makowski, D., Pham, T., Lau, Z.J., Brammer, J.C., Lespinasse, F., Pham, H., Schölzel, C., S H Chen, A.: Neurokit2: A python toolbox for neurophysiological signal processing (2020). https://doi.org/10.5281/ZENODO.3597887, https://github.com/neuropsychology/NeuroKit", + "23. Niu, X., Han, H., Shan, S., Chen, X.: Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14. pp. 562-576. Springer (2019)", + "24. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing 29, 2409-2423 (2019)", + "25. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. pp. 295–310. Springer (2020)", + "26. Niu, X., Zhao, X., Han, H., Das, A., Dantcheva, A., Shan, S., Chen, X.: Robust remote heart rate estimation from face utilizing spatial-temporal attention. In: 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019). pp. 1-8 (2019). https://doi.org/10.1109/FG.2019.8756554", + "27. Nowara, E.M., Marks, T.K., Mansour, H., Veeraraghavan, A.: Near-infrared imaging photoplethysmography during driving. IEEE transactions on intelligent transportation systems 23(4), 3589-3600 (2020)", + "28. Pilz, C.S., Zaunseder, S., Krajewski, J., Blazek, V.: Local group invariance for heart rate estimation from face videos in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1254-1262 (2018)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Poh, M.Z., McDuff, D.J., Picard, R.W.: Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE transactions on biomedical engineering 58(1), 7-11 (2010)", + "30. Savic, M., Zhao, G.: De-identification of facial videos while preserving remote physiological utility. In: 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023 (2023)", + "31. Savic, M., Zhao, G.: Physu-net: Long temporal context transformer for rppg with self-supervised pre-training (2024)", + "32. Savic, M., Zhao, G.: Rs-rppg: Robust self-supervised learning for rppg. In: 2024 18th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2024). IEEE (2024)", + "33. Speth, J., Vance, N., Flynn, P., Bowyer, K., Czajka, A.: Remote pulse estimation in the presence of face masks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2086-2095 (2022)", + "34. Speth, J., Vance, N., Flynn, P., Bowyer, K.W., Czajka, A.: Digital and physical-world attacks on remote pulse detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2407-2416 (2022)", + "35. Speth, J., Vance, N., Flynn, P., Czajka, A.: Non-contrastive unsupervised learning of physiological signals from video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14464-14474 (2023)", + "36. Speth, J., Vance, N., Sporrer, B., Niu, L., Flynn, P., Czajka, A.: Mspm: A multi-site physiological monitoring dataset for remote pulse, respiration, and blood pressure estimation. arXiv preprint arXiv:2402.02224 (2024)", + "37. Špetlík, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proceedings of the British machine vision conference, Newcastle, UK. pp. 3-6 (2018)", + "38. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062. IEEE (2014)", + "39. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062 (2014)", + "40. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XII. pp. 492–510. Springer (2022)", + "41. Sun, Z., Li, X.: Privacy-phys: Facial video-based physiological modification for privacy protection. IEEE Signal Processing Letters 29, 1507-1511 (2022). https://doi.org/10.1109/LSP.2022.3185964", + "42. Tang, J., Chen, K., Wang, Y., Shi, Y., Patel, S., McDuff, D., Liu, X.: Mmpd: Multi-domain mobile video physiology dataset. arXiv preprint arXiv:2302.03840 (2023)", + "43. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Optics express 16(26), 21434-21445 (2008)", + "44. Wang, W., den Brinker, A.C., Stuijk, S., De Haan, G.: Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering 64(7), 1479-1491 (2016)", + "45. Wang, W., Shan, C.: Impact of makeup on remote-ppg monitoring. Biomedical Physics & Engineering Express 6(3), 035004 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ORPDAD", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Yu, Z., Li, X., Zhao, G.: Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In: British Machine Vision Conference (2019)", + "47. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 151-160 (2019)", + "48. Yu, Z., Shen, Y., Shi, J., Zhao, H., Cui, Y., Zhang, J., Torr, P., Zhao, G.: Physformer++: Facial video-based physiological measurement with slowfast temporal difference transformer. International Journal of Computer Vision 131(6), 1307-1330 (2023)", + "49. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P.H., Zhao, G.: Physformer: facial video-based physiological measurement with temporal difference transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4186-4196 (2022)", + "50. Zhang, Z., Girard, J.M., Wu, Y., Zhang, X., Liu, P., Ciftci, U., Canavan, S., Reale, M., Horowitz, A., Yang, H., et al.: Multimodal spontaneous emotion corpus for human behavior analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3438-3446 (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 411 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M. Savic and G. Zhao", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_model.json b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87e2be500b6b3af25103341048059fc0bf1c54f9 --- /dev/null +++ b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_model.json @@ -0,0 +1,1919 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.241, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.213, + 0.64, + 0.228 + ], + "angle": 0, + "content": "Marko Savic1 and Guoying Zhao★1" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.24, + 0.776, + 0.269 + ], + "angle": 0, + "content": "Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland {marko.savic, guoying.zhao}@oulu.fi" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.303, + 0.744, + 0.54 + ], + "angle": 0, + "content": "Abstract. Remote photoplethysmography (rPPG) is an emerging technology that can detect the pulse rate remotely from face videos. However, it is easily influenced by the recording environment, as robustness to noise is still an open problem. This vulnerability can therefore be exploited to inject fake signals or impair predictions physically. In this study we propose the first dataset containing a wide set of physical domain attack scenarios divided in three categories (illumination, movement, concealment) that directly target the main weaknesses of rPPG. We propose the Oulu rPPG Physical Domain Attacks Database (ORPDAD) as a benchmark for evaluation of robustness to physical attacks. We perform extensive experiments on conventional hand-crafted and deep learning (end-to-end, non-end-to-end, CNN, transformer, self-supervised) methods and study their susceptibility to the attacks. We conclude by discussing the most critical vulnerabilities discovered and stress the importance of designing more secure solutions. Our code and instructions for requesting the dataset (with trained models) are available at: https://github.com/marukosan93/ORPDAD/." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.551, + 0.741, + 0.58 + ], + "angle": 0, + "content": "Keywords: Remote Photoplethysmography \\(\\cdot\\) Facial Video \\(\\cdot\\) Physical-Domain Attack \\(\\cdot\\) Dataset" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.377, + 0.621 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Physiological signals measured via photoplethysmography (PPG) and electrocardiogram (ECG) are used to determine vital healthcare parameters such as heart rate (HR), heart rate variability (HRV), respiratory rate (RR) and oxygen saturation (SpO2). However, they are limited by the necessity for physical contact that can make them uncomfortable in long term monitoring scenarios, especially with acute skin condition patients or fragile skin neonates, and very inconvenient in natural interactions. On the other hand, Remote photoplethysmography (rPPG) is a technology that allows extracting a signal akin to the blood volume pulse (BVP) without any contact, by utilising ubiquitous RGB cameras to record facial videos. Relying on the same principle as contact PPG [2], rPPG exploits subtle variations in reflected light from the skin that are caused by the cardiac cycle [43]. Several early methods had been proposed that relied" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.442, + 0.84 + ], + "angle": 0, + "content": "* indicates corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.128 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "on hand-crafted features or blind source separation [8, 16, 28, 29, 43, 44]. However, they lacked environmental robustness and were surpassed by deep learning approaches. Most deep learning methods employed supervised learning via Convolutional Neural Network (CNN) based models [5, 7, 21, 24, 25, 37, 46, 47] or more recently, self-attention based transformer architectures [13, 20, 31, 49]. Self-supervised methods such as [11, 35, 40] have also been a recently proposed to learn without annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Factors such as illumination variation, subject movement, camera sensor noise and occlusion give rise to a poor signal-to-noise ratio (SNR), making robustness to environmental noise challenging. Most rPPG methods developed in recent years try to solve this environmental noise issue, and they are evaluated on benchmark datasets designed to contain a certain degree of noise, e.g. different head motions [38], facial expressions [50], different illumination conditions [1, 15, 42], and also challenging motion and lighting in indoors [23] or in car driving scenarios [27]. A few studies also consider skin occlusions either via opaque masks [33] or translucent make-up [45]. Despite numerous methods and increasing number of datasets, robustness to environmental conditions is still an open research problem. To be widely used and applicable in real world settings, rPPG needs to be both robust and secure. From a security standpoint, lack of environmental robustness can be considered a critical vulnerability that should be addressed. Thus far, security related research in rPPG has been limited to the digital domain, as most works focus on editing the underlying physiological signals [4,41] (considered a digital attack) or de-identifying the data while retaining physiological features [30]. Digital injection of an attack is studied in [34], where a limited physical attack involving lighting is proposed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.737 + ], + "angle": 0, + "content": "The lack of robustness to environmental conditions is a key issue, and we wish to draw attention to the fact that it is a significant vulnerability that attackers can potentially exploit. As rPPG is in its early stages, exposing and patching security vulnerabilities is crucial for the development of robust and secure solutions. We propose a new dataset aiming specifically at potential physical domain attacks that can either spoof the rPPG with a target signal or deteriorate it. Our novel Oulu rPPG Physical Domain Attacks Database (ORPDAD) consists of three different types of attacks (i.e. illumination, movement and concealment) that result in 26 different recording scenarios across 30 subjects. Through these attacks, we physically inject periodical signals that can spoof the rPPG (via a pulsating light and rhythmical movements) or conceal skin to deteriorate the rPPG. The purpose of this dataset is to aid in developing more secure solutions to physical domain attacks and to provide a new benchmark for rPPG robustness to physical attacks. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.748, + 0.785, + 0.778 + ], + "angle": 0, + "content": "- ORPDAD is the first dataset to consider a comprehensive range of physical domain rPPG attacks, that can spoof rPPG signals or deteriorate them." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.785, + 0.809 + ], + "angle": 0, + "content": "- Ours is the first work to introduce rhythmical movement as a potential attack, and expands greatly on illumination and concealment based attacks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- We carry out an extensive comparative study for each attack's effectiveness, with experiments encompassing thirteen hand-crafted, deep learning based" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.748, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "supervised and self-supervised rPPG methods. We analyse and share our findings on attack susceptibility and uncovered weaknesses." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.785, + 0.209 + ], + "angle": 0, + "content": "- We establish a novel benchmark for physical domain rPPG attacks and share our dataset and materials (data, evaluation protocols, code and baselines)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.233, + 0.392, + 0.249 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Hand-crafted rPPG: The first rPPG method [43] extracted a signal from the green channel via averaged facial video pixels in a region of interest (e.g. forehead and cheeks). Model based methods have been proposed to better discriminate rPPG signals from motion noise such as CHROM [8] and POS [44], that leverage optical/physiological considerations expressed through mathematical models. Common blind source separation approaches such as PCA [16] and ICA [29] have been applied as well. In LGI [28] a new feature space is proposed that is more robust to motion and illumination changes via differentiable local transformations. However, the assumptions and hand-crafted features of the aforementioned methods did not hold in less constrained scenarios. For example, CHROM [8] assumes a standardized skin color profile, ICA [29] assumes independence of source signals and PCA [16] assumes their uncorrelation. Deep learning has shown to be more effective, especially in less controlled scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Deep learning based rPPG: The first 2D-CNN models extracted the HR from two adjacent frames, such as HR-CNN [37] and DeepPhys [5], only taking spatial information into account. End-to-end spatial-temporal 3D-CNN models such as PhysNet [46], rPPGNet [47] and TS-CAN [19] have been used to exploit the temporal information. To suppress the HR unrelated information spatial-temporal maps are introduced as input, like in non-end-to-end methods such as Niu et al. [26], RhythmNet [24], Dual-GAN [21], BVPNet [7], and CVD [25]. Recently, Efficientphys [20], PhysFormer [49], RADIANT [13] and PhySU-Net [31] have shown promising improvements by leveraging self-attention based transformers. To overcome the need for (expensive) annotated data, self-supervised methods that leverage the data alone have been proposed. They learn contrastively with artificial augmentations [11], via instance-wise sampling Contrast-Phys [40], exploit inter-data differences in a non-contrastive manner SiNC [35] or use a large set of diverse priors RS-rPPG [32]. In addition, rPPG has been widely utilised as an important face anti-spoofing cue [14, 18], as the presence of an rPPG signal can be an effective face liveness indicator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "rPPG Datasets: Given the crucial role of data, both for evaluation and for training supervised methods, rPPG datasets have been essential in the development of new methods. The challenge of the rPPG task on a specific dataset is highly dependent on the recording environment, as less constrained (variable illumination, large motion, unstable fps, compression, etc.) scenarios lead to higher error. PURE [39] contains facial videos with PPG groundtruth, with a total of 60 one-minute-long \\(640 \\times 480\\) 30fps videos from 10 subjects. It was recorded under ambient lighting with small motion tasks (steady, talking, slow translation, fast translation, slow rotation, medium rotation). MMSE-HR [50] has 102 videos" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.127 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.538 + ], + "angle": 0, + "content": "(with PPG, blood pressure, RR) of 20-70s recorded with stable lighting from 40 subjects in emotion elicitation experiments, it contains spontaneous facial expressions and head motions. COHFACE [15] consists of 40 participants with 160 total 1-minute \\(640 \\times 480\\) 20fps recordings (with PPG), where controlled studio and natural ambient lighting are used. OBF [17] contains 200 5-min \\(1920 \\times 1080\\) 60fps RGB videos (with PPG, ECG, RR) recorded from 100 diverse subjects, under stable lighting and with minimal movement, with both resting and elevated HR conditions. UBFC-rPPG [1] is a dataset of 50 participants with 2-min \\(640 \\times 480\\) 30fps recordings each (with PPG). It has variable indoor lighting and contains both 8 at-rest recordings and 42 recordings where participants play a time-sensitive mathematical game to increase their heart rate. VIPL-HR [23] contains 2,378 RGB videos of 20-30s length recorded from 108 subjects, it was recorded in a challenging environment with: different devices, varied and unstable frame rates, large movements and variable lighting. MR-NIRP [27] is the first rPPG dataset containing driving scenarios, it was recorded from 19 subject with 190 total \\(640 \\times 640\\) 60fps videos (with PPG), both during driving and while parked. On top of the variable lighting, subjects were asked to speak or perform random head movements. MMPD [42] is the first dataset recorded entirely with smartphone cameras (with PPG). It includes 33 subjects with 330 one-minute videos total, recorded under different lighting conditions (LED high, LED low, incandescent, natural), and different motions such as: resting, head rotation, conversation, and walking. Currently, there are numerous rPPG datasets containing different kind of environmental noise. Nevertheless, none of these large scale datasets contain intentional periodical noise as a physical attack medium, via movement or illumination. Furthermore, skin concealment via opaque cloth or transparent make-up is also not present in the aforementioned datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.827 + ], + "angle": 0, + "content": "Security in rPPG: As rPPG deals with sensitive biometric facial videos and physiological signals, security and privacy based research has also been a focus in recent years. Protecting privacy by removing identity related features while preserving physiological features has been attempted first via a Gaussian decomposition that maintains the low frequency spatial features [3]. A visual utility preserving method has also been proposed, that removes identity related features but preserves the physiological and visual utility via a semi-adversarial framework [30]. In contrast to rPPG preservation, PulseEdit [4] and PrivacyPhys [41] were devised to alter the rPPG signal in videos by using small perturbations. The original purpose of the aforementioned works is to conceal identity or the physiological signals, but they can be repurposed to alter the rPPG signal digitally and thus employed as a digital attacks that alter the underlying physiological data to result in a target signal. However, all these works are limited to the digital domain. In [34] digital-domain rPPG attacks are explicitly studied by injecting a target periodical signal. Furthermore, the authors propose a simple pulsating single LED attack by collecting a small dataset. However, most attack frequencies were outside the normal resting heart rate (filtered out during post-processing), the data contained only five two-minute videos from four subjects, attack intensity and range were limited, and the attack was a simple sinusoidal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "wave. A similar attack was proposed in MSPM [36] involving pulsating light from a monitor. Apart from illumination variation, motion is also a large noise source that datasets try to incorporate in varying degrees. However, no studies have yet tried to inject a specific target signal by utilising periodical motion as a medium. Finally, as rPPG is measured from light reflected from the facial skin, concealing it can be considered as physical domain attack with the aim to deteriorate the signal. In [33], the authors propose the MPM dataset containing 159 3-minute recordings of 54 subjects wearing facial masks, and show the importance of having methods robust to face occlusions, but only consider opaque face masks. In [45] foundation is considered as a translucent mask with 22 subjects, as make-up can affect rPPG readings with a notable decrease in SNR, but the authors only create one scenario involving foundation alone. Current rPPG solutions are not robust to noise and struggle in challenging scenarios, their robustness to physical domain attacks lacks sufficient investigation, as current physical domain attack research is limited in scope and only considers specific instances." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.412, + 0.712, + 0.43 + ], + "angle": 0, + "content": "3 Oulu rPPG Physical Domain Attacks Database" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.671 + ], + "angle": 0, + "content": "As rPPG is an emerging technology that deals with sensitive data and could be employed in applications where security is critical (clinical, remote driver monitoring, anti-spoofing, authentication, etc.), it is crucial to critically assess its vulnerabilities early and work towards mitigating potential security exploits. In this work, we focus on the scarcely studied physical-domain attacks. A physical-domain attack assumes that the attacker cannot influence the digital data after acquisition or the rPPG method used for estimation, namely the attacker can only inject noise during the acquisition process using physical means. We are inspired by previous works [33, 34, 36, 45] that deal with a subset of potential attacks in a limited scope. To fully understand and mitigate potential physical attacks to rPPG, we propose the ORPDAD dataset that aims to study a large set of diverse physical attacks within the same recording environment, thus isolating the variable of interest associated with the specific attack. The attacks we propose are simple to execute, but can have a large influence on the rPPG estimation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.536, + 0.71 + ], + "angle": 0, + "content": "3.1 Motivation and reasons for setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We identify vulnerabilities based on the properties of rPPG measuring, in Fig. 1 we display three major vulnerabilities. Firstly, as rPPG is measured from reflected light, a possible attack vector is the light source. Adding a weak perturbation signal to the incident light source means that it will be present in the estimated rPPG signal as well. Secondly, subject movement induces noise to the extracted signal, consequently having the subject move at a set rhythm will make this added motion perturbation strongly periodical. Therefore, we can leverage both motion and illumination to inject strongly periodical (rPPG like) signals" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.421, + 0.127 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.558, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.143, + 0.787, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.323, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 1: (Left) Major rPPG measurement environment vulnerabilities and related potential attack vectors. (Right) Data recording setup with equipment as listed in Tab. 1" + }, + { + "type": "table_caption", + "bbox": [ + 0.252, + 0.366, + 0.749, + 0.382 + ], + "angle": 0, + "content": "Table 1: Equipment used for data collection, also shown in Fig. 1 (Right)" + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.392, + 0.764, + 0.539 + ], + "angle": 0, + "content": "
Device/ToolDescription/Specifications
1) RGB CameraBlackmagic Ursa Mini 4K EF1920x1080 30fps 500Mbps f8.7 360° ISO400 WB4000K fl24mm
2) 2×Studio lightsAputure LightStorm LS 1c 3500lux 4000K
3) 2×Custom LED panels1600lm 11.5W 4000K (96×2835SMD LED) 20×13.3cm²
4) LED Control CircuitRPi 4B + IRL540N MOSFET + Resistors + 12V3A PSU
5) Finger-oximeter PPGBeurer PO80 120Hz
6) ECGBittium Faros 180 1Khz 1ch
7) MetronomeGoogle Metronome Web App
8) Concealment (C1-C3)Blue Medical Mask and Black Beanie Hat
9) Concealment (C4-C6)Maybelline Fit Me Matte Poreless \nPrimer + Foundation 220 + Powder 220
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.629 + ], + "angle": 0, + "content": "into the data. Thirdly, the light measured is reflected from the skin, so either blocking the light (with an opaque cloth) or attenuating it (with make-up) can affect the rPPG negatively. Consequently, we define these three physical-domain attack categories: Illumination (I), Movement (M) and Concealment (C)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.649, + 0.371, + 0.665 + ], + "angle": 0, + "content": "3.2 Study design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We record data from 30 healthy (no history of heart disease) volunteers under different scenarios based on the three aforementioned categories of attacks. All the participants gave written informed consent. For the active attacks we choose \\(50~\\mathrm{bpm}\\) and \\(100~\\mathrm{bpm}\\) as attack frequencies as they can be considered borderline resting heart rates, thus plausible but different from the real HR in most recordings. We avoid results being biased by the real HR trend by averaging results over both frequencies. We record 70 second videos for each scenario, for a total of 26 scenarios per subject, resulting in 780 videos equivalent to \\(\\approx 15.2\\mathrm{h}\\) of footage. For the synchronised ground truth physiological signals, we record both PPG waveforms via a finger-oximeter and ECG waveforms via an ECG device. Based on the defined attack categories, we define the following scenarios:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.482, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.147, + 0.785, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Fig. 2: Raw signal (GREEN [43]) and its spectrogram for diverse attacks. For active illumination (I1, I6) and motion based attacks (M1, M9), the target attack frequency is clearly visible in the spectrogram and can take over the prediction. On the other hand, passive concealment attacks (C1, C6) significantly affect the estimation's SNR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.445 + ], + "angle": 0, + "content": "Still (S1-S3): Subject is recorded at a still and resting state before each attack category. Motion is kept at a comfortable minimum." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.808, + 0.52 + ], + "angle": 0, + "content": "Illumination (I1-I6): We combine the two attack frequencies (50bpm, 100bpm) and three different intensity settings (S=7%, M=14% and L=21% of the attack device max brightness), resulting in I1=50S, I2=50M, I3=50L, I4=100S, I5=100M, I6=100L. The intensities are much weaker than the base lighting, and cause perturbations are visually imperceptible in the recordings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Movement (M1-M11): We explore simple movements such as small vertical (SV), large vertical (LV), small horizontal (SH), large horizontal (LH), mouth open/close (M) at the two target frequencies (50bpm, 100bpm). Resulting in M1=50SV, M2=50LV, M3=50SH, M4=50LH, M5=50M, M6=100SV, M7=100LV, M8=100SH, M9=100LH, M10=100M. We also include natural talking motion M11 that is not at a set attack frequency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Concealment (C1-C6): The first three opaque (C1-C3) scenarios involve wearing a forehead concealing beanie C1, a facial mask C2 and then both C3. The transparent concealment scenarios (C4-C6) contain progressive application of make-up with primer C4, foundation C5 and setting powder C6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.703 + ], + "angle": 0, + "content": "In Fig. 2 we visualise a subset of the aforementioned attacks and their effect on the raw signal from the GREEN [43] method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.441, + 0.74 + ], + "angle": 0, + "content": "3.3 Equipment and setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To study each attack vector in an isolated manner, we set up a controlled recording environment as to minimise the influence of external factors that are not related to the attack. Thus, we keep the same base environment for all subjects and scenarios, enabling us to study the specific contribution induced by the attacks. For the base video recordings, we use a professional RGB camera and two fixed LED lights. We record the physiological signals with a finger PPG oximeter" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.421, + 0.128 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "and a portable ECG belt device. All the recording devices are synchronised by keeping internal clock times consistent. The subjects are seated at \\(\\approx 1\\mathrm{m}\\) from the camera and the lights are positioned above the camera and illuminate the subject at a \\(45^{\\circ}\\) angle at \\(\\approx 1.5\\mathrm{m}\\) distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.433 + ], + "angle": 0, + "content": "For the illumination attacks, we design two modulated light panels that each operate at \\(1600\\mathrm{lm}\\) maximum, by using LED strips and a control circuit. These panels allow the injected light signal to be more uniformly distributed over the face surface, instead of attacking only part of it (single led case). We pilot the panels with Pulse Width Modulation (PWM) signals, thus being able to create a light source that can output any waveform and at any fraction of the max intensity. For this experiment, we simulate a typical rPPG waveform, with the main frequency component placed at the target frequency and a second harmonic at \\(1/4\\) of the peak intensity. For the L, M, H settings we use \\(7\\%\\), \\(14\\%\\), \\(21\\%\\) of the max power and set the lights at \\(\\approx 1\\mathrm{m}\\) distance from the subject. For the movement scenarios, we simply utilise an audio metronome and instruct the subjects to move according to the set rhythm and movement type. For opaque concealment, we use medical face masks and black beanie hats. For translucent concealment we choose a primer, foundation and setting powder, and keep the brand and shade (220 natural beige) consistent among all subjects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.786, + 0.465 + ], + "angle": 0, + "content": "We provide a detailed list of all the equipment and settings utilised for recording in Tab. 1, and show all the equipment in Fig. 1 (right)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.377, + 0.506 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.732 + ], + "angle": 0, + "content": "We assess the effects of each physical attack on a large set of rPPG methods that represent the past and current state of the art, as they include both common hand-crafted and CNN based methods, as well more recent transformer based and self-supervised works. We choose the following hand-crafted non-learning based methods: GREEN [43], CHROM [8], POS [44], LGI [28], ICA [29] and PCA [16]. As for learning based CNN methods, we choose the end-to-end methods DeepPhys [5], TS-CAN [19] and PhysNet [46], and the non-end-to-end method CVD [25]. We also include two transformer based methods that are Physformer [49] (end-to-end) and PhySU-Net [31] (non-end-to-end). Finally, we include an end-to-end self-supervised method Contrast-Phys [40]. We faithfully re-implement the aforementioned methods and fine-tune them to obtain the best performance on our dataset. All our code can be found at https://github.com/marukosan93/ORPDAD/, with instructions for requesting the dataset (containing both raw data, preprocessed data and trained models)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.427, + 0.77 + ], + "angle": 0, + "content": "4.1 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Preprocessing: Firstly, we utilise PyFeat [6] with RetinaFace [9] model for face detection and PFLD [12] model for landmark localisation, the landmarks are then stabilised with a 5-point moving average filter. For the end-to-end methods, we crop the images to \\(128 \\times 128\\) so that the face is centered. For hand-crafted" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "methods we use the landmarks to define a ROI (including forehead, cheeks and chin) and average the pixels to the raw signal input. Finally, as both non-end-to-end methods CVD [25] and PhySU-Net [31] utilise multi-scale spatial-temporal maps (MSTmaps) as input, we follow the suggested processing [25] to generate the input MSTmaps. For the ground-truth signal, as ECG is considered the gold standard, we choose to primarily use ECG for experimentation. We first filter the ECG with the method in [10] via NeuroKit2 [22], and then apply the 10-point average method from [48] to get a resulting signal containing the same physiological features but with a shape better correlated to the rPPG signal. All ground-truth signals, original videos and pre-processed videos (landmarks, MSTmaps, cropped) will be included in the shared dataset version." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Metrics: We follow previous works by using mean absolute error (MAE), root-mean-square error (RMSE), and Pearson's correlation coefficient (R). Furthermore, we calculate the SNR, defined as the ratio between the power in the dominant frequency and second harmonics versus the remaining power [8]. Note that good SNR does not coincide with accurate prediction, as it measures the spectrum sparsity around the predicted rPPG (without considering the true label). For active attacks, we also calculate success rate, defined as the percentage of times when the predicted HR is closer to the two attack frequencies than the ground truth, meaning that the attack successfully hijacked the prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Protocol: To ensure fair validation we adopt a 5-fold subject exclusive protocol for all experiments. Additionally, we consider two different training conditions. Attack Unaware (AU), where the learning based methods are trained only on the still (S) scenarios and are not exposed to attacks during training, but only during testing. Attack Aware (AA), in this case, the training data contains the attack scenarios so that we can assess whether awareness improves robustness. For AA, we add an attack frequency exclusive cross-validation to the protocol, we train on all passive scenarios and active scenarios with one attack frequency, and then test on the non-seen attack frequency. As the attack frequencies are fixed, supervised methods could easily cheat by learning to not predict the seen specific attack frequency, without learning any useful features from the attacks. Our AA protocol is aware of the attacks, but not of the specific attack frequency. In testing 30 second non-overlapping segments are evaluated, and we calculate the HR by finding the highest PSD peak." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.442, + 0.696 + ], + "angle": 0, + "content": "4.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.748 + ], + "angle": 0, + "content": "We analyse each attack category separately, regarding the signal deterioration in terms of accuracy and SNR. Subsequently, we study the success rate of the active attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Firstly, we evaluate all the studied methods on the illumination scenarios, as shown in Tab. 2. On the base scenarios (S1-S3) almost all methods (except GREEN [43]) have very good performance (\\(\\leq 1\\) RMSE) as the base task is not very challenging by design, due to the controlled environment. It stands out noticeably that hand-crafted methods are very susceptible to the illumination attack as most methods have a large drop in accuracy, with only the skin" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.127 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Results on illumination attacks I1-I6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.199, + 0.699, + 0.487 + ], + "angle": 0, + "content": "
RMSE↓ |SNR↑S1-S3 (still)I1+I4 (small)I2+I5 (med.)I3+I6 (large)
GREEN [43]□8.57 -7.425.7 -6.727.9 -4.127.7 -2.5
CHROM [8]□0.91 -3.71.87 -4.713.3 -4.419.5 -4.3
POS [44]□0.70 -3.45.77 -4.411.4 -4.813.8 -4.8
LGI [28]□1.05 -3.58.16 -4.420.7 -4.225.9 -3.3
ICA [29]□1.14 -2.922.3 -3.727.6 -1.427.7 0.2
PCA [16]□1.35 -5.110.5 -6.215.6 -6.513.1 -7.2
DeepPhys [5]◆AU1.02 -2.75.53 -3.922.6 -3.426.8 -2.4
AA0.79 -2.80.63 -3.78.23 -4.013.9 -4.4
TS-CAN [19]◆AU1.16 -3.66.31 -4.723.5 -4.127.4 -3.1
AA0.88 -3.50.76 -3.91.09 -4.25.90 -4.5
PhysNet [46]◆AU0.88 -3.14.58 -4.213.2 -4.616.1 -4.7
AA0.55 -2.80.51 -3.10.46 -3.51.61 -4.1
CVD [25]◇AU1.03 -3.11.51 -3.95.47 -4.67.17 -5.2
AA0.79 -2.41.10 -2.70.95 -3.04.08 -3.4
Physformer [49]◆AU0.95 -3.01.01 -3.810.0 -4.514.8 -4.6
AA1.34 -2.41.59 -2.70.57 -3.17.51 -3.8
PhySU-Net [31]◇AU0.74 -1.90.62 -2.95.71 -3.76.24 -4.0
AA0.50 -1.80.93 -2.11.41 -2.63.15 -2.7
Contrast-Phys [40]▼AU0.95 -2.81.51 -3.813.9 -4.317.2 -4.4
AA20.1 -3.620.9 -3.721.6 -4.121.9 -4.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.807 + ], + "angle": 0, + "content": "model based CHROM [8] and POS [44] having comparable performance only with the weakest attack (I1+I4). Moreover, decrease in accuracy does not necessarily mean decrease in SNR, e.g. ICA [29] improves in SNR by having a higher intensity attack, as the output spectrum contains more power in the dominant attack frequency. For supervised deep learning methods there is an overall better degree of innate robustness as the learned features are more complex and physiology specific. However, in the AU case, except for the weakest setting (I1+I4), the supervised methods are significantly affected by the attack. The illumination based attacks affect non-end-to-end methods CVD [25] and PhySU-Net [31] less, as the spatial-temporal map inputs are more resistant to the illumination variation. When adding the attacks into the training data (AA), the performance significantly improves for the supervised methods, as the networks learn to ignore the features specific to the attack, showing the important role of new data in protection against physical attacks. On the other hand, self-supervised Contrast-Phys [40] is very receptive to the attacks as it focuses on salient frequency features without supervision. In the AA case, the learned features lose their physiological meaning as the method learns from inter-data differences that are most pronounced between the attacks, leading to inaccurate predictions even for the non-challenging still scenarios." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Secondly, we study the movement based attacks as shown in Tab. 3. The hand-crafted methods all have a significant performance drop, that is closely cor-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Results on movement attacks M1-M11. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.787, + 0.486 + ], + "angle": 0, + "content": "
RMSE↓ |SNR↑S1-S3(still)M1+M6(s. ver.)M2+M7(l. ver.)M3+M8(s. hor.)M4+M9(l. hor.)M5+M10(mouth)M11(talk)
GREEN [43]□8.57 -7.426.6 -4.426.8 -2.823.5 -6.731.0 -5.626.9 -5.223.1 -8.7
CHROM [8]□0.91 -3.710.4 -4.911.5 -5.86.34 -5.019.3 -5.910.6 -5.51.35 -6.8
POS [44]□0.70 -3.48.60 -4.813.8 -5.78.22 -4.814.5 -6.110.3 -5.41.20 -6.5
LGI [28]□1.05 -3.518.0 -4.926.1 -3.811.4 -4.924.2 -5.112.7 -5.35.45 -6.7
ICA [29]□1.14 -2.925.5 -3.324.9 -2.721.8 -4.726.3 -4.123.4 -4.03.99 -6.4
PCA [16]□1.35 -5.113.4 -5.416.6 -7.314.5 -5.923.9 -6.413.5 -5.76.13 -7.1
DeepPhys [5]◆AU1.02 -2.719.4 -4.626.3 -3.523.7 -4.625.1 -4.717.2 -4.95.65 -6.6
AA0.79 -2.86.16 -3.915.9 -5.115.0 -4.620.6 -5.39.38 -4.17.47 -5.4
TS-CAN [19]◆AU1.16 -3.624.2 -4.526.8 -3.224.4 -5.326.0 -4.819.8 -5.01.79 -7.2
AA0.88 -3.55.99 -4.318.5 -5.315.2 -5.021.0 -5.811.7 -4.71.45 -5.7
PhysNet [46]◆AU0.88 -3.117.1 -5.022.6 -4.99.45 -5.320.6 -6.38.86 -4.56.21 -5.0
AA0.55 -2.83.61 -3.311.5 -4.43.62 -3.79.11 -5.50.71 -3.63.37 -4.2
CVD [25]◇AU1.03 -3.110.6 -5.516.8 -7.46.62 -4.211.2 -6.15.23 -5.18.74 -4.9
AA0.79 -2.44.64 -3.412.9 -5.80.83 -2.75.96 -4.52.97 -3.66.30 -3.7
Physformer [49]◆AU0.95 -3.011.0 -4.318.3 -5.012.0 -4.820.5 -5.78.55 -4.68.68 -5.2
AA1.34 -2.40.69 -2.610.8 -4.07.30 -3.111.7 -4.85.94 -3.25.67 -3.9
PhySU-Net [31]◇AU0.74 -1.95.93 -4.215.2 -6.53.58 -3.99.48 -6.20.69 -3.67.71 -4.2
AA0.50 -1.81.31 -2.39.17 -3.80.69 -2.27.04 -4.20.62 -2.27.56 -3.1
Contrast-Phys [40]▼AU0.95 -2.815.0 -4.821.6 -4.66.84 -4.918.1 -5.99.54 -4.63.41 -5.0
AA20.1 -3.625.5 -4.627.1 -5.827.8 -5.228.8 -5.422.3 -4.424.1 -5.0
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.521, + 0.788, + 0.701 + ], + "angle": 0, + "content": "related to the movement intensity, as larger movements yield larger errors. There is a contrast between the strongly periodical mouth active attack (M5+M10) and the normal talking motion (M11), as the performance is better when the noise is not periodical (e.g. CHROM [8] and POS [44] have comparable performance to the still scenarios), even though the SNR is lower signifying a noisier spectrum. All the supervised deep learning methods are severely affected by the attack, with non-end-to-end PhySU-Net [31] and CVD [25] having lower accuracy drops overall. Similarly to the illumination attacks, in the AA case the performance significantly improves for supervised methods, as the networks learn to ignore attack specific features to a certain degree. On the other hand, for the AU susceptible Contrast-Phys [40] the addition of attack awareness is even aggravating, similarly as in the illumination case." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Thirdly, we study the concealment based attacks as shown in Tab. 4 and visualise the SNR changes in Fig. 3. In this case, hand-crafted methods obtain variable performance on partial concealment (C1 and C2), with C2 having strong drops in accuracy for LGI [28], CHROM [8] and PCA [16]. This is due to the medical mask covering a larger surface of the ROI than the forehead covering hat. In case of total coverage C3, due to the small portion of uncovered skin the task is extremely challenging and leads to low accuracy. There is an overall decrease in SNR when applying the concealment, as the physiologically irrelevant surface is larger. In the make-up related attacks (C4-C6), the performance only" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.127 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Table 4: Results on concealment attacks C1-C6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.787, + 0.486 + ], + "angle": 0, + "content": "
RMSE↓|SNR↑S1-S3(still)C1(hat)C2(mask)C3(both)C4(primer)C5(found.)C6(powder)
GREEN [43]□8.57 -7.414.0 -8.718.9 -9.121.9 -9.29.46 -7.712.1 -8.813.7 -8.7
CHROM [8]□0.91 -3.70.72 -5.311.1 -7.817.6 -9.50.87 -4.11.16 -5.20.95 -5.3
POS [44]□0.70 -3.40.69 -4.70.96 -6.023.7 -10.0.57 -3.50.94 -4.60.80 -4.6
LGI [28]□1.05 -3.50.70 -4.914.1 -8.722.0 -9.60.84 -3.61.13 -4.90.92 -4.9
ICA [29]□1.14 -2.90.81 -3.81.03 -5.217.7 -9.20.75 -3.01.05 -4.40.87 -4.2
PCA [16]□1.35 -5.11.37 -5.424.1 -9.422.0 -9.01.08 -4.87.78 -6.08.32 -6.1
DeepPhys [5]AU1.02 -2.70.77 -3.415.3 -9.418.9 -9.30.76 -2.80.84 -3.60.75 -3.4
AA0.79 -2.80.82 -3.31.18 -4.00.85 -5.80.73 -2.80.80 -3.80.81 -3.7
TS-CAN [19]◆AU1.16 -3.61.05 -4.918.4 -9.017.7 -8.90.77 -3.70.91 -4.91.21 -4.5
AA0.88 -3.50.93 -4.01.48 -4.43.35 -6.10.76 -3.41.01 -4.81.38 -4.7
PhysNet [46]◆AU0.88 -3.10.50 -3.30.67 -4.514.8 -7.60.71 -3.10.76 -4.00.60 -3.7
AA0.55 -2.80.50 -3.10.51 -3.20.79 -3.90.54 -2.80.60 -3.60.52 -3.3
CVD [25]◇AU1.03 -3.11.24 -3.57.07 -7.510.4 -8.57.76 -3.11.05 -3.70.75 -3.8
AA0.79 -2.40.82 -2.91.16 -3.88.43 -7.10.87 -2.31.03 -2.90.73 -3.0
Physformer [49]◆AU0.95 -3.00.94 -3.410.8 -6.616.1 -7.90.71 -3.00.68 -3.90.66 -3.6
AA1.34 -2.41.87 -2.81.09 -3.12.33 -3.90.54 -2.40.69 -3.20.59 -2.8
PhySU-Net [31]◇AU0.74 -1.90.58 -2.86.47 -7.218.7 -8.70.67 -1.90.57 -3.00.55 -2.7
AA0.50 -1.80.62 -2.30.84 -2.810.3 -6.70.62 -1.80.63 -2.50.51 -2.1
Contrast-Phys [40]▼AU0.95 -2.80.73 -3.413.8 -8.713.6 -9.00.81 -2.80.72 -4.30.86 -4.4
AA20.1 -3.618.9 -3.437.0 -6.335.8 -6.020.1 -3.418.3 -3.616.7 -3.5
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.538, + 0.787, + 0.841 + ], + "angle": 0, + "content": "significantly drops for GREEN [43] and PCA [16] that are more susceptible to the drop in SNR. For the fully transparent primer C4 (optically similar to a facial creme), the drop in SNR is negligible not significantly affecting the rPPG signal. For the foundation C5 and powder C6 scenarios, the drop in SNR is very noticeable (\\(\\approx 1dB\\) on average) due to the changed optical properties of the skin. Between C5 and C6 (same shade, but C6 has an additional layer) the difference in negligible, meaning that with regards to the rPPG the shade is most relevant. Similarly, for AU supervised deep learning methods the facial mask is more damaging to the accuracy of non-end-to-end methods, as they rely on the ROI only, and C3 is too challenging of a task. For make-up the behaviour is analogous to the hand-crafted case, with C4 being negligible, and C5 and C6 having a similar overall SNR drop (\\(\\approx 1dB\\)). In the AA case, all methods can adapt to the partial attack (C1, C2), and surprisingly end-to-end methods are even able to obtain good performance on the challenging C3 scenario, as they are able to find alternative physiologically relevant patches of skin. Regarding make-up C5 and C6, in the AA case SNR increases noticeably (\\(\\approx 0.5dB\\)) as the networks partially mitigate the change in skin shade. As for the self-supervised Contrast-Phys [40], it is vulnerable to C2 and C3 in the AU case. Analogously to the previous two attack categories, in the AA case the features decline in physiological meaning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.75, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.147, + 0.783, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.439 + ], + "angle": 0, + "content": "Fig. 3: SNR for passive concealment attacks (C1-C6) and success rates for active attacks (I1-I6, M1-M10). Higher SNR and lower success rates (lighter color intensity) correspond to better robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.637 + ], + "angle": 0, + "content": "Lastly, we also analyse the success rate of the active attacks, as shown in Fig. 3, using the success rate metric (from Sec. 4.1) we can effectively determine with what rate the attacks overtake the real predictions. In line with the accuracy, we can see that CHROM [8] and POS [44] are the overall the more robust hand-crafted methods. The non-end-to-end (PhySU-Net [31] and CVD [25]) AU supervised methods overall are more robust with the illumination and motion attacks, with large vertical movement as the most challenging attack. The AA case makes the success rate significantly lower. In line with the accuracy, for the non-robust Contrast-Phys [40] the AA case seems to worsen, as the addition of attacks in the training data has the negative effect of increasing noise related inter-data differences that guide the method to learn non-physiological features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.35, + 0.675 + ], + "angle": 0, + "content": "4.3 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our experiments benchmark the current state of rPPG in relation to physical domain attacks, in hand-crafted, supervised and self-supervised settings (AU, AA). Active attacks (I1-I6, M1-M10) have a strong impact on hand-crafted methods as they are less robust to environmental noise, with CHROM [8] and POS [44] being less susceptible as they are based on physiology specific mathematical models. Supervised methods offer more resistance to the attacks, as they are guided by strong supervision to ignore non-physiological features and learn features strongly correlated to the ground truth, thus are able to offer better robustness. Non-end-to-end methods (CVD [25], PhySU-Net [31]) using signal map inputs show the best robustness out of supervised methods, due to the noise" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.127 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.148, + 0.788, + 0.526 + ], + "angle": 0, + "content": "suppression capabilities of the signal maps. However, non-end-to-end methods depend on face ROI locations, making them less effective with Movement attacks compared to Illumination. For the end-to-end methods, DeepPhys [5] and TS-CAN [19] are less robust than Physnet [46] and Physformer [49], due to their short temporal context architectures. Having the attacks present during training (AA) greatly improves robustness of supervised methods. Most methods follow a similar trend to the AU case, as the innate robustness is supplemented by the data awareness, meaning that the same design factors lead to effectiveness in the AA case as in the AU case. Self-supervised methods (ContrastPhys [40]) on the other hand are very susceptible to attacks, both in the AU case and especially in the AA case, as the attacks negatively influence the training. Opaque Concealment attacks (C1-C3) are more effective on ROI dependant methods such as hand-crafted and non-end-to-end, with the facial mask being more aggravating as it covers a larger physiologically relevant surface, and C3 (mask and hat) being too challenging of a task, except for end-to-end methods in AA case. Regarding make-up (C4-C6), the transparent primer has negligible effect, and that addition of foundation leads to a prominent drop in SNR, but further addition of powder is negligible, meaning that shade is more impactful. Most supervised methods partially adapt to the make-up and have improved SNR in the AA case. The addition of attacks into the training data significantly boost the performance as the networks learn to ignore the specific features related to the attacks (that are not correlated with the ground truth), thus showing the strong necessity of having such physical attacks present in the training set. Nonetheless, depending only on data to defend from such attacks is not feasible and future rPPG methods need to be more secure and innately robust to physical attacks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.36, + 0.565 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.747 + ], + "angle": 0, + "content": "We propose ORPDAD, the first rPPG dataset to include comprehensive physical domain attacks. We extensively evaluate thirteen diverse rPPG methods on the attack scenarios. We point out that hand-crafted and self-supervised methods are most susceptible, supervised methods (particularly non-end-to-end) have a certain degree of innate robustness that can be improved with data containing attacks. However, only having more data cannot mitigate this issue, as secure rPPG methods need to be designed to overcome these vulnerabilities. Thus, we hope that our dataset will encourage novel secure solutions. Future work includes the exploration of new robust methods against general physical attacks, and the extension of the study and dataset to include face spoofing attacks, as the studied attacks have the potential to deceive liveness detection based anti-spoofing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.377, + 0.785 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This work was supported by the Research Council of Finland Academy Professor project EmotionAI (grants 336116, 345122) and ICT 2023 project TrustFace (grant 345948), and the University of Oulu & Research Council of Finland Profi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "7 (grant 352788), and by Infotech Oulu. As well, the authors acknowledge CSC - IT Center for Science, Finland, for computational resources. We also thank Mengting Wei and all the study volunteers for their contribution." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.215, + 0.323, + 0.231 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.287 + ], + "angle": 0, + "content": "1. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.316 + ], + "angle": 0, + "content": "2. Challoner, A., Ramsay, C.: A photoelectric plethysmograph for the measurement of cutaneous blood flow. Physics in Medicine & Biology 19(3), 317 (1974)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.785, + 0.384 + ], + "angle": 0, + "content": "3. Chen, M., Jayaweera, A., Wong, C.W., Wu, M.: Identity-Privacy Protection for Facial-rPPG Based Smart Health Research (Jul 2022). https://doi.org/10.36227/techrxiv.20352843.v1, https://www.techrxiv.org/articles/preprint/Identity-Privacy_Protection_for_Facial-rPPG-Based_Smart_Health_Research/20352843/1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.426 + ], + "angle": 0, + "content": "4. Chen, M., Liao, X., Wu, M.: Pulseedit: Editing physiological signals in facial videos for privacy protection. IEEE Transactions on Information Forensics and Security 17, 457-471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.785, + 0.468 + ], + "angle": 0, + "content": "5. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 349-365 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.507 + ], + "angle": 0, + "content": "6. Cheong, J.H., Xie, T., Byrne, S., Chang, L.J.: Py-feat: Python facial expression analysis toolbox. CoRR abs/2104.03509 (2021), https://arxiv.org/abs/2104.03509" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.564 + ], + "angle": 0, + "content": "7. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021). pp. 01-08. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "8. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering 60(10), 2878-2886 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.633 + ], + "angle": 0, + "content": "9. Deng, J., Guo, J., Ververas, E., Kotsia, I., Zafeiriou, S.: Retinaface: Single-shot multi-level face localisation in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5203-5212 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.661 + ], + "angle": 0, + "content": "0. Elgendi, M., Jonkman, M., De Boer, F.: Frequency bands effects on qrs detection. Biosignals 2003, 2002 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "1. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.731 + ], + "angle": 0, + "content": "2. Guo, X., Li, S., Yu, J., Zhang, J., Ma, J., Ma, L., Liu, W., Ling, H.: Pfld: A practical facial landmark detector. arXiv preprint arXiv:1902.10859 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "3. Gupta, A.K., Kumar, R., Birla, L., Gupta, P.: Radiant: Better rppg estimation using signal embeddings and transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4976-4986 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "4. Hernandez-Ortega, J., Fierrez, J., Morales, A., Tome, P.: Time analysis of pulse-based face anti-spoofing in visible and nir. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-552 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Heusch, G., Anjos, A., Marcel, S.: A reproducible study on remote heart rate measurement. arXiv preprint arXiv:1709.00962 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.127 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.204 + ], + "angle": 0, + "content": "16. Lewandowska, M., Rumiński, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: 2011 federated conference on computer science and information systems (FedCSIS). pp. 405-410. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.274 + ], + "angle": 0, + "content": "17. Li, X., Alikhani, I., Shi, J., Seppanen, T., Junttila, J., Majamaa-Voltti, K., Tulppo, M., Zhao, G.: The obf database: A large face video database for remote physiological signal measurement and atrial fibrillation detection. In: 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018). pp. 242-249. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.331 + ], + "angle": 0, + "content": "18. Liu, S., Yuen, P.C., Zhang, S., Zhao, G.: 3d mask face anti-spoofing with remote photoplethysmography. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 85-100. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "19. Liu, X., Fromm, J., Patel, S., McDuff, D.: Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems 33, 19400-19411 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.43 + ], + "angle": 0, + "content": "20. Liu, X., Hill, B., Jiang, Z., Patel, S., McDuff, D.: Efficientphys: Enabling simple, fast and accurate camera-based cardiac measurement. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5008-5017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "21. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12404-12413 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.529 + ], + "angle": 0, + "content": "22. Makowski, D., Pham, T., Lau, Z.J., Brammer, J.C., Lespinasse, F., Pham, H., Schölzel, C., S H Chen, A.: Neurokit2: A python toolbox for neurophysiological signal processing (2020). https://doi.org/10.5281/ZENODO.3597887, https://github.com/neuropsychology/NeuroKit" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.586 + ], + "angle": 0, + "content": "23. Niu, X., Han, H., Shan, S., Chen, X.: Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14. pp. 562-576. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "24. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing 29, 2409-2423 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.787, + 0.685 + ], + "angle": 0, + "content": "25. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. pp. 295–310. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.741 + ], + "angle": 0, + "content": "26. Niu, X., Zhao, X., Han, H., Das, A., Dantcheva, A., Shan, S., Chen, X.: Robust remote heart rate estimation from face utilizing spatial-temporal attention. In: 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019). pp. 1-8 (2019). https://doi.org/10.1109/FG.2019.8756554" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.784 + ], + "angle": 0, + "content": "27. Nowara, E.M., Marks, T.K., Mansour, H., Veeraraghavan, A.: Near-infrared imaging photoplethysmography during driving. IEEE transactions on intelligent transportation systems 23(4), 3589-3600 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Pilz, C.S., Zaunseder, S., Krajewski, J., Blazek, V.: Local group invariance for heart rate estimation from face videos in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1254-1262 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.662, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ORPDAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Poh, M.Z., McDuff, D.J., Picard, R.W.: Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE transactions on biomedical engineering 58(1), 7-11 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "30. Savic, M., Zhao, G.: De-identification of facial videos while preserving remote physiological utility. In: 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.786, + 0.261 + ], + "angle": 0, + "content": "31. Savic, M., Zhao, G.: Physu-net: Long temporal context transformer for rppg with self-supervised pre-training (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.303 + ], + "angle": 0, + "content": "32. Savic, M., Zhao, G.: Rs-rppg: Robust self-supervised learning for rppg. In: 2024 18th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2024). IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "33. Speth, J., Vance, N., Flynn, P., Bowyer, K., Czajka, A.: Remote pulse estimation in the presence of face masks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2086-2095 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "34. Speth, J., Vance, N., Flynn, P., Bowyer, K.W., Czajka, A.: Digital and physical-world attacks on remote pulse detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2407-2416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.431 + ], + "angle": 0, + "content": "35. Speth, J., Vance, N., Flynn, P., Czajka, A.: Non-contrastive unsupervised learning of physiological signals from video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14464-14474 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.787, + 0.473 + ], + "angle": 0, + "content": "36. Speth, J., Vance, N., Sporrer, B., Niu, L., Flynn, P., Czajka, A.: Mspm: A multi-site physiological monitoring dataset for remote pulse, respiration, and blood pressure estimation. arXiv preprint arXiv:2402.02224 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "37. Špetlík, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proceedings of the British machine vision conference, Newcastle, UK. pp. 3-6 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.558 + ], + "angle": 0, + "content": "38. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062. IEEE (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.6 + ], + "angle": 0, + "content": "39. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.657 + ], + "angle": 0, + "content": "40. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XII. pp. 492–510. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "41. Sun, Z., Li, X.: Privacy-phys: Facial video-based physiological modification for privacy protection. IEEE Signal Processing Letters 29, 1507-1511 (2022). https://doi.org/10.1109/LSP.2022.3185964" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "42. Tang, J., Chen, K., Wang, Y., Shi, Y., Patel, S., McDuff, D., Liu, X.: Mmpd: Multi-domain mobile video physiology dataset. arXiv preprint arXiv:2302.03840 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.77 + ], + "angle": 0, + "content": "43. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Optics express 16(26), 21434-21445 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "44. Wang, W., den Brinker, A.C., Stuijk, S., De Haan, G.: Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering 64(7), 1479-1491 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "45. Wang, W., Shan, C.: Impact of makeup on remote-ppg monitoring. Biomedical Physics & Engineering Express 6(3), 035004 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.127 + ], + "angle": 0, + "content": "M. Savic and G. Zhao" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "46. Yu, Z., Li, X., Zhao, G.: Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In: British Machine Vision Conference (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "47. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 151-160 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "48. Yu, Z., Shen, Y., Shi, J., Zhao, H., Cui, Y., Zhang, J., Torr, P., Zhao, G.: Physformer++: Facial video-based physiological measurement with slowfast temporal difference transformer. International Journal of Computer Vision 131(6), 1307-1330 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.357 + ], + "angle": 0, + "content": "49. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P.H., Zhao, G.: Physformer: facial video-based physiological measurement with temporal difference transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4186-4196 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.412 + ], + "angle": 0, + "content": "50. Zhang, Z., Girard, J.M., Wu, Y., Zhang, X., Liu, P., Ciftci, U., Canavan, S., Reale, M., Horowitz, A., Yang, H., et al.: Multimodal spontaneous emotion corpus for human behavior analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3438-3446 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.412 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_origin.pdf b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f3dabe6163b9f5bd05b77907c6466a74bba8e1b5 --- /dev/null +++ b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/7e9c5f3a-b0e6-4921-9bc2-ea9d60f57b77_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcf2d750c49a8b3f4830a5e13a5bbb7b420d81d2bb2452aa760524a35b7dc96d +size 6129396 diff --git a/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/full.md b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/full.md new file mode 100644 index 0000000000000000000000000000000000000000..faba008f3cdcfc240a42bfdeb6be0d67bb709d01 --- /dev/null +++ b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/full.md @@ -0,0 +1,211 @@ +# Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD) + +Marko Savic1 and Guoying Zhao★1 + +Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland {marko.savic, guoying.zhao}@oulu.fi + +Abstract. Remote photoplethysmography (rPPG) is an emerging technology that can detect the pulse rate remotely from face videos. However, it is easily influenced by the recording environment, as robustness to noise is still an open problem. This vulnerability can therefore be exploited to inject fake signals or impair predictions physically. In this study we propose the first dataset containing a wide set of physical domain attack scenarios divided in three categories (illumination, movement, concealment) that directly target the main weaknesses of rPPG. We propose the Oulu rPPG Physical Domain Attacks Database (ORPDAD) as a benchmark for evaluation of robustness to physical attacks. We perform extensive experiments on conventional hand-crafted and deep learning (end-to-end, non-end-to-end, CNN, transformer, self-supervised) methods and study their susceptibility to the attacks. We conclude by discussing the most critical vulnerabilities discovered and stress the importance of designing more secure solutions. Our code and instructions for requesting the dataset (with trained models) are available at: https://github.com/marukosan93/ORPDAD/. + +Keywords: Remote Photoplethysmography $\cdot$ Facial Video $\cdot$ Physical-Domain Attack $\cdot$ Dataset + +# 1 Introduction + +Physiological signals measured via photoplethysmography (PPG) and electrocardiogram (ECG) are used to determine vital healthcare parameters such as heart rate (HR), heart rate variability (HRV), respiratory rate (RR) and oxygen saturation (SpO2). However, they are limited by the necessity for physical contact that can make them uncomfortable in long term monitoring scenarios, especially with acute skin condition patients or fragile skin neonates, and very inconvenient in natural interactions. On the other hand, Remote photoplethysmography (rPPG) is a technology that allows extracting a signal akin to the blood volume pulse (BVP) without any contact, by utilising ubiquitous RGB cameras to record facial videos. Relying on the same principle as contact PPG [2], rPPG exploits subtle variations in reflected light from the skin that are caused by the cardiac cycle [43]. Several early methods had been proposed that relied + +on hand-crafted features or blind source separation [8, 16, 28, 29, 43, 44]. However, they lacked environmental robustness and were surpassed by deep learning approaches. Most deep learning methods employed supervised learning via Convolutional Neural Network (CNN) based models [5, 7, 21, 24, 25, 37, 46, 47] or more recently, self-attention based transformer architectures [13, 20, 31, 49]. Self-supervised methods such as [11, 35, 40] have also been a recently proposed to learn without annotations. + +Factors such as illumination variation, subject movement, camera sensor noise and occlusion give rise to a poor signal-to-noise ratio (SNR), making robustness to environmental noise challenging. Most rPPG methods developed in recent years try to solve this environmental noise issue, and they are evaluated on benchmark datasets designed to contain a certain degree of noise, e.g. different head motions [38], facial expressions [50], different illumination conditions [1, 15, 42], and also challenging motion and lighting in indoors [23] or in car driving scenarios [27]. A few studies also consider skin occlusions either via opaque masks [33] or translucent make-up [45]. Despite numerous methods and increasing number of datasets, robustness to environmental conditions is still an open research problem. To be widely used and applicable in real world settings, rPPG needs to be both robust and secure. From a security standpoint, lack of environmental robustness can be considered a critical vulnerability that should be addressed. Thus far, security related research in rPPG has been limited to the digital domain, as most works focus on editing the underlying physiological signals [4,41] (considered a digital attack) or de-identifying the data while retaining physiological features [30]. Digital injection of an attack is studied in [34], where a limited physical attack involving lighting is proposed. + +The lack of robustness to environmental conditions is a key issue, and we wish to draw attention to the fact that it is a significant vulnerability that attackers can potentially exploit. As rPPG is in its early stages, exposing and patching security vulnerabilities is crucial for the development of robust and secure solutions. We propose a new dataset aiming specifically at potential physical domain attacks that can either spoof the rPPG with a target signal or deteriorate it. Our novel Oulu rPPG Physical Domain Attacks Database (ORPDAD) consists of three different types of attacks (i.e. illumination, movement and concealment) that result in 26 different recording scenarios across 30 subjects. Through these attacks, we physically inject periodical signals that can spoof the rPPG (via a pulsating light and rhythmical movements) or conceal skin to deteriorate the rPPG. The purpose of this dataset is to aid in developing more secure solutions to physical domain attacks and to provide a new benchmark for rPPG robustness to physical attacks. Our contributions are as follows: + +- ORPDAD is the first dataset to consider a comprehensive range of physical domain rPPG attacks, that can spoof rPPG signals or deteriorate them. +- Ours is the first work to introduce rhythmical movement as a potential attack, and expands greatly on illumination and concealment based attacks. +- We carry out an extensive comparative study for each attack's effectiveness, with experiments encompassing thirteen hand-crafted, deep learning based + +supervised and self-supervised rPPG methods. We analyse and share our findings on attack susceptibility and uncovered weaknesses. + +- We establish a novel benchmark for physical domain rPPG attacks and share our dataset and materials (data, evaluation protocols, code and baselines). + +# 2 Related works + +Hand-crafted rPPG: The first rPPG method [43] extracted a signal from the green channel via averaged facial video pixels in a region of interest (e.g. forehead and cheeks). Model based methods have been proposed to better discriminate rPPG signals from motion noise such as CHROM [8] and POS [44], that leverage optical/physiological considerations expressed through mathematical models. Common blind source separation approaches such as PCA [16] and ICA [29] have been applied as well. In LGI [28] a new feature space is proposed that is more robust to motion and illumination changes via differentiable local transformations. However, the assumptions and hand-crafted features of the aforementioned methods did not hold in less constrained scenarios. For example, CHROM [8] assumes a standardized skin color profile, ICA [29] assumes independence of source signals and PCA [16] assumes their uncorrelation. Deep learning has shown to be more effective, especially in less controlled scenarios. + +Deep learning based rPPG: The first 2D-CNN models extracted the HR from two adjacent frames, such as HR-CNN [37] and DeepPhys [5], only taking spatial information into account. End-to-end spatial-temporal 3D-CNN models such as PhysNet [46], rPPGNet [47] and TS-CAN [19] have been used to exploit the temporal information. To suppress the HR unrelated information spatial-temporal maps are introduced as input, like in non-end-to-end methods such as Niu et al. [26], RhythmNet [24], Dual-GAN [21], BVPNet [7], and CVD [25]. Recently, Efficientphys [20], PhysFormer [49], RADIANT [13] and PhySU-Net [31] have shown promising improvements by leveraging self-attention based transformers. To overcome the need for (expensive) annotated data, self-supervised methods that leverage the data alone have been proposed. They learn contrastively with artificial augmentations [11], via instance-wise sampling Contrast-Phys [40], exploit inter-data differences in a non-contrastive manner SiNC [35] or use a large set of diverse priors RS-rPPG [32]. In addition, rPPG has been widely utilised as an important face anti-spoofing cue [14, 18], as the presence of an rPPG signal can be an effective face liveness indicator. + +rPPG Datasets: Given the crucial role of data, both for evaluation and for training supervised methods, rPPG datasets have been essential in the development of new methods. The challenge of the rPPG task on a specific dataset is highly dependent on the recording environment, as less constrained (variable illumination, large motion, unstable fps, compression, etc.) scenarios lead to higher error. PURE [39] contains facial videos with PPG groundtruth, with a total of 60 one-minute-long $640 \times 480$ 30fps videos from 10 subjects. It was recorded under ambient lighting with small motion tasks (steady, talking, slow translation, fast translation, slow rotation, medium rotation). MMSE-HR [50] has 102 videos + +(with PPG, blood pressure, RR) of 20-70s recorded with stable lighting from 40 subjects in emotion elicitation experiments, it contains spontaneous facial expressions and head motions. COHFACE [15] consists of 40 participants with 160 total 1-minute $640 \times 480$ 20fps recordings (with PPG), where controlled studio and natural ambient lighting are used. OBF [17] contains 200 5-min $1920 \times 1080$ 60fps RGB videos (with PPG, ECG, RR) recorded from 100 diverse subjects, under stable lighting and with minimal movement, with both resting and elevated HR conditions. UBFC-rPPG [1] is a dataset of 50 participants with 2-min $640 \times 480$ 30fps recordings each (with PPG). It has variable indoor lighting and contains both 8 at-rest recordings and 42 recordings where participants play a time-sensitive mathematical game to increase their heart rate. VIPL-HR [23] contains 2,378 RGB videos of 20-30s length recorded from 108 subjects, it was recorded in a challenging environment with: different devices, varied and unstable frame rates, large movements and variable lighting. MR-NIRP [27] is the first rPPG dataset containing driving scenarios, it was recorded from 19 subject with 190 total $640 \times 640$ 60fps videos (with PPG), both during driving and while parked. On top of the variable lighting, subjects were asked to speak or perform random head movements. MMPD [42] is the first dataset recorded entirely with smartphone cameras (with PPG). It includes 33 subjects with 330 one-minute videos total, recorded under different lighting conditions (LED high, LED low, incandescent, natural), and different motions such as: resting, head rotation, conversation, and walking. Currently, there are numerous rPPG datasets containing different kind of environmental noise. Nevertheless, none of these large scale datasets contain intentional periodical noise as a physical attack medium, via movement or illumination. Furthermore, skin concealment via opaque cloth or transparent make-up is also not present in the aforementioned datasets. + +Security in rPPG: As rPPG deals with sensitive biometric facial videos and physiological signals, security and privacy based research has also been a focus in recent years. Protecting privacy by removing identity related features while preserving physiological features has been attempted first via a Gaussian decomposition that maintains the low frequency spatial features [3]. A visual utility preserving method has also been proposed, that removes identity related features but preserves the physiological and visual utility via a semi-adversarial framework [30]. In contrast to rPPG preservation, PulseEdit [4] and PrivacyPhys [41] were devised to alter the rPPG signal in videos by using small perturbations. The original purpose of the aforementioned works is to conceal identity or the physiological signals, but they can be repurposed to alter the rPPG signal digitally and thus employed as a digital attacks that alter the underlying physiological data to result in a target signal. However, all these works are limited to the digital domain. In [34] digital-domain rPPG attacks are explicitly studied by injecting a target periodical signal. Furthermore, the authors propose a simple pulsating single LED attack by collecting a small dataset. However, most attack frequencies were outside the normal resting heart rate (filtered out during post-processing), the data contained only five two-minute videos from four subjects, attack intensity and range were limited, and the attack was a simple sinusoidal + +wave. A similar attack was proposed in MSPM [36] involving pulsating light from a monitor. Apart from illumination variation, motion is also a large noise source that datasets try to incorporate in varying degrees. However, no studies have yet tried to inject a specific target signal by utilising periodical motion as a medium. Finally, as rPPG is measured from light reflected from the facial skin, concealing it can be considered as physical domain attack with the aim to deteriorate the signal. In [33], the authors propose the MPM dataset containing 159 3-minute recordings of 54 subjects wearing facial masks, and show the importance of having methods robust to face occlusions, but only consider opaque face masks. In [45] foundation is considered as a translucent mask with 22 subjects, as make-up can affect rPPG readings with a notable decrease in SNR, but the authors only create one scenario involving foundation alone. Current rPPG solutions are not robust to noise and struggle in challenging scenarios, their robustness to physical domain attacks lacks sufficient investigation, as current physical domain attack research is limited in scope and only considers specific instances. + +# 3 Oulu rPPG Physical Domain Attacks Database + +As rPPG is an emerging technology that deals with sensitive data and could be employed in applications where security is critical (clinical, remote driver monitoring, anti-spoofing, authentication, etc.), it is crucial to critically assess its vulnerabilities early and work towards mitigating potential security exploits. In this work, we focus on the scarcely studied physical-domain attacks. A physical-domain attack assumes that the attacker cannot influence the digital data after acquisition or the rPPG method used for estimation, namely the attacker can only inject noise during the acquisition process using physical means. We are inspired by previous works [33, 34, 36, 45] that deal with a subset of potential attacks in a limited scope. To fully understand and mitigate potential physical attacks to rPPG, we propose the ORPDAD dataset that aims to study a large set of diverse physical attacks within the same recording environment, thus isolating the variable of interest associated with the specific attack. The attacks we propose are simple to execute, but can have a large influence on the rPPG estimation. + +# 3.1 Motivation and reasons for setup + +We identify vulnerabilities based on the properties of rPPG measuring, in Fig. 1 we display three major vulnerabilities. Firstly, as rPPG is measured from reflected light, a possible attack vector is the light source. Adding a weak perturbation signal to the incident light source means that it will be present in the estimated rPPG signal as well. Secondly, subject movement induces noise to the extracted signal, consequently having the subject move at a set rhythm will make this added motion perturbation strongly periodical. Therefore, we can leverage both motion and illumination to inject strongly periodical (rPPG like) signals + +![](images/2568bf67fec847deda912322786522e2599299977088f778485d85c2d9b5e350.jpg) +Fig. 1: (Left) Major rPPG measurement environment vulnerabilities and related potential attack vectors. (Right) Data recording setup with equipment as listed in Tab. 1 + +![](images/793d37da45e10660132ca2ed54f4d0ab5cd646506741d2456b546be52fcdb4ab.jpg) + +Table 1: Equipment used for data collection, also shown in Fig. 1 (Right) + +
Device/ToolDescription/Specifications
1) RGB CameraBlackmagic Ursa Mini 4K EF1920x1080 30fps 500Mbps f8.7 360° ISO400 WB4000K fl24mm
2) 2×Studio lightsAputure LightStorm LS 1c 3500lux 4000K
3) 2×Custom LED panels1600lm 11.5W 4000K (96×2835SMD LED) 20×13.3cm²
4) LED Control CircuitRPi 4B + IRL540N MOSFET + Resistors + 12V3A PSU
5) Finger-oximeter PPGBeurer PO80 120Hz
6) ECGBittium Faros 180 1Khz 1ch
7) MetronomeGoogle Metronome Web App
8) Concealment (C1-C3)Blue Medical Mask and Black Beanie Hat
9) Concealment (C4-C6)Maybelline Fit Me Matte Poreless +Primer + Foundation 220 + Powder 220
+ +into the data. Thirdly, the light measured is reflected from the skin, so either blocking the light (with an opaque cloth) or attenuating it (with make-up) can affect the rPPG negatively. Consequently, we define these three physical-domain attack categories: Illumination (I), Movement (M) and Concealment (C). + +# 3.2 Study design + +We record data from 30 healthy (no history of heart disease) volunteers under different scenarios based on the three aforementioned categories of attacks. All the participants gave written informed consent. For the active attacks we choose $50~\mathrm{bpm}$ and $100~\mathrm{bpm}$ as attack frequencies as they can be considered borderline resting heart rates, thus plausible but different from the real HR in most recordings. We avoid results being biased by the real HR trend by averaging results over both frequencies. We record 70 second videos for each scenario, for a total of 26 scenarios per subject, resulting in 780 videos equivalent to $\approx 15.2\mathrm{h}$ of footage. For the synchronised ground truth physiological signals, we record both PPG waveforms via a finger-oximeter and ECG waveforms via an ECG device. Based on the defined attack categories, we define the following scenarios: + +![](images/81e662409a576a7160298d48a912ddf8f99f101b4c74ebf23d4bba2e4568f40d.jpg) +Fig. 2: Raw signal (GREEN [43]) and its spectrogram for diverse attacks. For active illumination (I1, I6) and motion based attacks (M1, M9), the target attack frequency is clearly visible in the spectrogram and can take over the prediction. On the other hand, passive concealment attacks (C1, C6) significantly affect the estimation's SNR. + +![](images/04fb62e317ab8ff284ecdd5d2f5174836b124212928b2d92068f754c2741cb61.jpg) + +Still (S1-S3): Subject is recorded at a still and resting state before each attack category. Motion is kept at a comfortable minimum. + +Illumination (I1-I6): We combine the two attack frequencies (50bpm, 100bpm) and three different intensity settings (S=7%, M=14% and L=21% of the attack device max brightness), resulting in I1=50S, I2=50M, I3=50L, I4=100S, I5=100M, I6=100L. The intensities are much weaker than the base lighting, and cause perturbations are visually imperceptible in the recordings. + +Movement (M1-M11): We explore simple movements such as small vertical (SV), large vertical (LV), small horizontal (SH), large horizontal (LH), mouth open/close (M) at the two target frequencies (50bpm, 100bpm). Resulting in M1=50SV, M2=50LV, M3=50SH, M4=50LH, M5=50M, M6=100SV, M7=100LV, M8=100SH, M9=100LH, M10=100M. We also include natural talking motion M11 that is not at a set attack frequency. + +Concealment (C1-C6): The first three opaque (C1-C3) scenarios involve wearing a forehead concealing beanie C1, a facial mask C2 and then both C3. The transparent concealment scenarios (C4-C6) contain progressive application of make-up with primer C4, foundation C5 and setting powder C6. + +In Fig. 2 we visualise a subset of the aforementioned attacks and their effect on the raw signal from the GREEN [43] method. + +# 3.3 Equipment and setup + +To study each attack vector in an isolated manner, we set up a controlled recording environment as to minimise the influence of external factors that are not related to the attack. Thus, we keep the same base environment for all subjects and scenarios, enabling us to study the specific contribution induced by the attacks. For the base video recordings, we use a professional RGB camera and two fixed LED lights. We record the physiological signals with a finger PPG oximeter + +and a portable ECG belt device. All the recording devices are synchronised by keeping internal clock times consistent. The subjects are seated at $\approx 1\mathrm{m}$ from the camera and the lights are positioned above the camera and illuminate the subject at a $45^{\circ}$ angle at $\approx 1.5\mathrm{m}$ distance. + +For the illumination attacks, we design two modulated light panels that each operate at $1600\mathrm{lm}$ maximum, by using LED strips and a control circuit. These panels allow the injected light signal to be more uniformly distributed over the face surface, instead of attacking only part of it (single led case). We pilot the panels with Pulse Width Modulation (PWM) signals, thus being able to create a light source that can output any waveform and at any fraction of the max intensity. For this experiment, we simulate a typical rPPG waveform, with the main frequency component placed at the target frequency and a second harmonic at $1/4$ of the peak intensity. For the L, M, H settings we use $7\%$ , $14\%$ , $21\%$ of the max power and set the lights at $\approx 1\mathrm{m}$ distance from the subject. For the movement scenarios, we simply utilise an audio metronome and instruct the subjects to move according to the set rhythm and movement type. For opaque concealment, we use medical face masks and black beanie hats. For translucent concealment we choose a primer, foundation and setting powder, and keep the brand and shade (220 natural beige) consistent among all subjects. + +We provide a detailed list of all the equipment and settings utilised for recording in Tab. 1, and show all the equipment in Fig. 1 (right). + +# 4 Experiments + +We assess the effects of each physical attack on a large set of rPPG methods that represent the past and current state of the art, as they include both common hand-crafted and CNN based methods, as well more recent transformer based and self-supervised works. We choose the following hand-crafted non-learning based methods: GREEN [43], CHROM [8], POS [44], LGI [28], ICA [29] and PCA [16]. As for learning based CNN methods, we choose the end-to-end methods DeepPhys [5], TS-CAN [19] and PhysNet [46], and the non-end-to-end method CVD [25]. We also include two transformer based methods that are Physformer [49] (end-to-end) and PhySU-Net [31] (non-end-to-end). Finally, we include an end-to-end self-supervised method Contrast-Phys [40]. We faithfully re-implement the aforementioned methods and fine-tune them to obtain the best performance on our dataset. All our code can be found at https://github.com/marukosan93/ORPDAD/, with instructions for requesting the dataset (containing both raw data, preprocessed data and trained models). + +# 4.1 Experimental setup + +Preprocessing: Firstly, we utilise PyFeat [6] with RetinaFace [9] model for face detection and PFLD [12] model for landmark localisation, the landmarks are then stabilised with a 5-point moving average filter. For the end-to-end methods, we crop the images to $128 \times 128$ so that the face is centered. For hand-crafted + +methods we use the landmarks to define a ROI (including forehead, cheeks and chin) and average the pixels to the raw signal input. Finally, as both non-end-to-end methods CVD [25] and PhySU-Net [31] utilise multi-scale spatial-temporal maps (MSTmaps) as input, we follow the suggested processing [25] to generate the input MSTmaps. For the ground-truth signal, as ECG is considered the gold standard, we choose to primarily use ECG for experimentation. We first filter the ECG with the method in [10] via NeuroKit2 [22], and then apply the 10-point average method from [48] to get a resulting signal containing the same physiological features but with a shape better correlated to the rPPG signal. All ground-truth signals, original videos and pre-processed videos (landmarks, MSTmaps, cropped) will be included in the shared dataset version. + +Metrics: We follow previous works by using mean absolute error (MAE), root-mean-square error (RMSE), and Pearson's correlation coefficient (R). Furthermore, we calculate the SNR, defined as the ratio between the power in the dominant frequency and second harmonics versus the remaining power [8]. Note that good SNR does not coincide with accurate prediction, as it measures the spectrum sparsity around the predicted rPPG (without considering the true label). For active attacks, we also calculate success rate, defined as the percentage of times when the predicted HR is closer to the two attack frequencies than the ground truth, meaning that the attack successfully hijacked the prediction. + +Protocol: To ensure fair validation we adopt a 5-fold subject exclusive protocol for all experiments. Additionally, we consider two different training conditions. Attack Unaware (AU), where the learning based methods are trained only on the still (S) scenarios and are not exposed to attacks during training, but only during testing. Attack Aware (AA), in this case, the training data contains the attack scenarios so that we can assess whether awareness improves robustness. For AA, we add an attack frequency exclusive cross-validation to the protocol, we train on all passive scenarios and active scenarios with one attack frequency, and then test on the non-seen attack frequency. As the attack frequencies are fixed, supervised methods could easily cheat by learning to not predict the seen specific attack frequency, without learning any useful features from the attacks. Our AA protocol is aware of the attacks, but not of the specific attack frequency. In testing 30 second non-overlapping segments are evaluated, and we calculate the HR by finding the highest PSD peak. + +# 4.2 Experimental Results + +We analyse each attack category separately, regarding the signal deterioration in terms of accuracy and SNR. Subsequently, we study the success rate of the active attacks. + +Firstly, we evaluate all the studied methods on the illumination scenarios, as shown in Tab. 2. On the base scenarios (S1-S3) almost all methods (except GREEN [43]) have very good performance ( $\leq 1$ RMSE) as the base task is not very challenging by design, due to the controlled environment. It stands out noticeably that hand-crafted methods are very susceptible to the illumination attack as most methods have a large drop in accuracy, with only the skin + +Table 2: Results on illumination attacks I1-I6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement. + +
RMSE↓ |SNR↑S1-S3 (still)I1+I4 (small)I2+I5 (med.)I3+I6 (large)
GREEN [43]□8.57 -7.425.7 -6.727.9 -4.127.7 -2.5
CHROM [8]□0.91 -3.71.87 -4.713.3 -4.419.5 -4.3
POS [44]□0.70 -3.45.77 -4.411.4 -4.813.8 -4.8
LGI [28]□1.05 -3.58.16 -4.420.7 -4.225.9 -3.3
ICA [29]□1.14 -2.922.3 -3.727.6 -1.427.7 0.2
PCA [16]□1.35 -5.110.5 -6.215.6 -6.513.1 -7.2
DeepPhys [5]◆AU1.02 -2.75.53 -3.922.6 -3.426.8 -2.4
AA0.79 -2.80.63 -3.78.23 -4.013.9 -4.4
TS-CAN [19]◆AU1.16 -3.66.31 -4.723.5 -4.127.4 -3.1
AA0.88 -3.50.76 -3.91.09 -4.25.90 -4.5
PhysNet [46]◆AU0.88 -3.14.58 -4.213.2 -4.616.1 -4.7
AA0.55 -2.80.51 -3.10.46 -3.51.61 -4.1
CVD [25]◇AU1.03 -3.11.51 -3.95.47 -4.67.17 -5.2
AA0.79 -2.41.10 -2.70.95 -3.04.08 -3.4
Physformer [49]◆AU0.95 -3.01.01 -3.810.0 -4.514.8 -4.6
AA1.34 -2.41.59 -2.70.57 -3.17.51 -3.8
PhySU-Net [31]◇AU0.74 -1.90.62 -2.95.71 -3.76.24 -4.0
AA0.50 -1.80.93 -2.11.41 -2.63.15 -2.7
Contrast-Phys [40]▼AU0.95 -2.81.51 -3.813.9 -4.317.2 -4.4
AA20.1 -3.620.9 -3.721.6 -4.121.9 -4.5
+ +model based CHROM [8] and POS [44] having comparable performance only with the weakest attack (I1+I4). Moreover, decrease in accuracy does not necessarily mean decrease in SNR, e.g. ICA [29] improves in SNR by having a higher intensity attack, as the output spectrum contains more power in the dominant attack frequency. For supervised deep learning methods there is an overall better degree of innate robustness as the learned features are more complex and physiology specific. However, in the AU case, except for the weakest setting (I1+I4), the supervised methods are significantly affected by the attack. The illumination based attacks affect non-end-to-end methods CVD [25] and PhySU-Net [31] less, as the spatial-temporal map inputs are more resistant to the illumination variation. When adding the attacks into the training data (AA), the performance significantly improves for the supervised methods, as the networks learn to ignore the features specific to the attack, showing the important role of new data in protection against physical attacks. On the other hand, self-supervised Contrast-Phys [40] is very receptive to the attacks as it focuses on salient frequency features without supervision. In the AA case, the learned features lose their physiological meaning as the method learns from inter-data differences that are most pronounced between the attacks, leading to inaccurate predictions even for the non-challenging still scenarios. + +Secondly, we study the movement based attacks as shown in Tab. 3. The hand-crafted methods all have a significant performance drop, that is closely cor- + +Table 3: Results on movement attacks M1-M11. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement. + +
RMSE↓ |SNR↑S1-S3(still)M1+M6(s. ver.)M2+M7(l. ver.)M3+M8(s. hor.)M4+M9(l. hor.)M5+M10(mouth)M11(talk)
GREEN [43]□8.57 -7.426.6 -4.426.8 -2.823.5 -6.731.0 -5.626.9 -5.223.1 -8.7
CHROM [8]□0.91 -3.710.4 -4.911.5 -5.86.34 -5.019.3 -5.910.6 -5.51.35 -6.8
POS [44]□0.70 -3.48.60 -4.813.8 -5.78.22 -4.814.5 -6.110.3 -5.41.20 -6.5
LGI [28]□1.05 -3.518.0 -4.926.1 -3.811.4 -4.924.2 -5.112.7 -5.35.45 -6.7
ICA [29]□1.14 -2.925.5 -3.324.9 -2.721.8 -4.726.3 -4.123.4 -4.03.99 -6.4
PCA [16]□1.35 -5.113.4 -5.416.6 -7.314.5 -5.923.9 -6.413.5 -5.76.13 -7.1
DeepPhys [5]◆AU1.02 -2.719.4 -4.626.3 -3.523.7 -4.625.1 -4.717.2 -4.95.65 -6.6
AA0.79 -2.86.16 -3.915.9 -5.115.0 -4.620.6 -5.39.38 -4.17.47 -5.4
TS-CAN [19]◆AU1.16 -3.624.2 -4.526.8 -3.224.4 -5.326.0 -4.819.8 -5.01.79 -7.2
AA0.88 -3.55.99 -4.318.5 -5.315.2 -5.021.0 -5.811.7 -4.71.45 -5.7
PhysNet [46]◆AU0.88 -3.117.1 -5.022.6 -4.99.45 -5.320.6 -6.38.86 -4.56.21 -5.0
AA0.55 -2.83.61 -3.311.5 -4.43.62 -3.79.11 -5.50.71 -3.63.37 -4.2
CVD [25]◇AU1.03 -3.110.6 -5.516.8 -7.46.62 -4.211.2 -6.15.23 -5.18.74 -4.9
AA0.79 -2.44.64 -3.412.9 -5.80.83 -2.75.96 -4.52.97 -3.66.30 -3.7
Physformer [49]◆AU0.95 -3.011.0 -4.318.3 -5.012.0 -4.820.5 -5.78.55 -4.68.68 -5.2
AA1.34 -2.40.69 -2.610.8 -4.07.30 -3.111.7 -4.85.94 -3.25.67 -3.9
PhySU-Net [31]◇AU0.74 -1.95.93 -4.215.2 -6.53.58 -3.99.48 -6.20.69 -3.67.71 -4.2
AA0.50 -1.81.31 -2.39.17 -3.80.69 -2.27.04 -4.20.62 -2.27.56 -3.1
Contrast-Phys [40]▼AU0.95 -2.815.0 -4.821.6 -4.66.84 -4.918.1 -5.99.54 -4.63.41 -5.0
AA20.1 -3.625.5 -4.627.1 -5.827.8 -5.228.8 -5.422.3 -4.424.1 -5.0
+ +related to the movement intensity, as larger movements yield larger errors. There is a contrast between the strongly periodical mouth active attack (M5+M10) and the normal talking motion (M11), as the performance is better when the noise is not periodical (e.g. CHROM [8] and POS [44] have comparable performance to the still scenarios), even though the SNR is lower signifying a noisier spectrum. All the supervised deep learning methods are severely affected by the attack, with non-end-to-end PhySU-Net [31] and CVD [25] having lower accuracy drops overall. Similarly to the illumination attacks, in the AA case the performance significantly improves for supervised methods, as the networks learn to ignore attack specific features to a certain degree. On the other hand, for the AU susceptible Contrast-Phys [40] the addition of attack awareness is even aggravating, similarly as in the illumination case. + +Thirdly, we study the concealment based attacks as shown in Tab. 4 and visualise the SNR changes in Fig. 3. In this case, hand-crafted methods obtain variable performance on partial concealment (C1 and C2), with C2 having strong drops in accuracy for LGI [28], CHROM [8] and PCA [16]. This is due to the medical mask covering a larger surface of the ROI than the forehead covering hat. In case of total coverage C3, due to the small portion of uncovered skin the task is extremely challenging and leads to low accuracy. There is an overall decrease in SNR when applying the concealment, as the physiologically irrelevant surface is larger. In the make-up related attacks (C4-C6), the performance only + +Table 4: Results on concealment attacks C1-C6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement. + +
RMSE↓|SNR↑S1-S3(still)C1(hat)C2(mask)C3(both)C4(primer)C5(found.)C6(powder)
GREEN [43]□8.57 -7.414.0 -8.718.9 -9.121.9 -9.29.46 -7.712.1 -8.813.7 -8.7
CHROM [8]□0.91 -3.70.72 -5.311.1 -7.817.6 -9.50.87 -4.11.16 -5.20.95 -5.3
POS [44]□0.70 -3.40.69 -4.70.96 -6.023.7 -10.0.57 -3.50.94 -4.60.80 -4.6
LGI [28]□1.05 -3.50.70 -4.914.1 -8.722.0 -9.60.84 -3.61.13 -4.90.92 -4.9
ICA [29]□1.14 -2.90.81 -3.81.03 -5.217.7 -9.20.75 -3.01.05 -4.40.87 -4.2
PCA [16]□1.35 -5.11.37 -5.424.1 -9.422.0 -9.01.08 -4.87.78 -6.08.32 -6.1
DeepPhys [5]AU1.02 -2.70.77 -3.415.3 -9.418.9 -9.30.76 -2.80.84 -3.60.75 -3.4
AA0.79 -2.80.82 -3.31.18 -4.00.85 -5.80.73 -2.80.80 -3.80.81 -3.7
TS-CAN [19]◆AU1.16 -3.61.05 -4.918.4 -9.017.7 -8.90.77 -3.70.91 -4.91.21 -4.5
AA0.88 -3.50.93 -4.01.48 -4.43.35 -6.10.76 -3.41.01 -4.81.38 -4.7
PhysNet [46]◆AU0.88 -3.10.50 -3.30.67 -4.514.8 -7.60.71 -3.10.76 -4.00.60 -3.7
AA0.55 -2.80.50 -3.10.51 -3.20.79 -3.90.54 -2.80.60 -3.60.52 -3.3
CVD [25]◇AU1.03 -3.11.24 -3.57.07 -7.510.4 -8.57.76 -3.11.05 -3.70.75 -3.8
AA0.79 -2.40.82 -2.91.16 -3.88.43 -7.10.87 -2.31.03 -2.90.73 -3.0
Physformer [49]◆AU0.95 -3.00.94 -3.410.8 -6.616.1 -7.90.71 -3.00.68 -3.90.66 -3.6
AA1.34 -2.41.87 -2.81.09 -3.12.33 -3.90.54 -2.40.69 -3.20.59 -2.8
PhySU-Net [31]◇AU0.74 -1.90.58 -2.86.47 -7.218.7 -8.70.67 -1.90.57 -3.00.55 -2.7
AA0.50 -1.80.62 -2.30.84 -2.810.3 -6.70.62 -1.80.63 -2.50.51 -2.1
Contrast-Phys [40]▼AU0.95 -2.80.73 -3.413.8 -8.713.6 -9.00.81 -2.80.72 -4.30.86 -4.4
AA20.1 -3.618.9 -3.437.0 -6.335.8 -6.020.1 -3.418.3 -3.616.7 -3.5
+ +significantly drops for GREEN [43] and PCA [16] that are more susceptible to the drop in SNR. For the fully transparent primer C4 (optically similar to a facial creme), the drop in SNR is negligible not significantly affecting the rPPG signal. For the foundation C5 and powder C6 scenarios, the drop in SNR is very noticeable ( $\approx 1dB$ on average) due to the changed optical properties of the skin. Between C5 and C6 (same shade, but C6 has an additional layer) the difference in negligible, meaning that with regards to the rPPG the shade is most relevant. Similarly, for AU supervised deep learning methods the facial mask is more damaging to the accuracy of non-end-to-end methods, as they rely on the ROI only, and C3 is too challenging of a task. For make-up the behaviour is analogous to the hand-crafted case, with C4 being negligible, and C5 and C6 having a similar overall SNR drop ( $\approx 1dB$ ). In the AA case, all methods can adapt to the partial attack (C1, C2), and surprisingly end-to-end methods are even able to obtain good performance on the challenging C3 scenario, as they are able to find alternative physiologically relevant patches of skin. Regarding make-up C5 and C6, in the AA case SNR increases noticeably ( $\approx 0.5dB$ ) as the networks partially mitigate the change in skin shade. As for the self-supervised Contrast-Phys [40], it is vulnerable to C2 and C3 in the AU case. Analogously to the previous two attack categories, in the AA case the features decline in physiological meaning. + +![](images/65f411554251c62ee182cc2713a978effde1b57c30c8448ded6477c04a751200.jpg) +Fig. 3: SNR for passive concealment attacks (C1-C6) and success rates for active attacks (I1-I6, M1-M10). Higher SNR and lower success rates (lighter color intensity) correspond to better robustness. + +![](images/b46a47b0c8a2817e0f4c3890db7f47d3098084324a4bc87fcf00714fd50a584f.jpg) + +Lastly, we also analyse the success rate of the active attacks, as shown in Fig. 3, using the success rate metric (from Sec. 4.1) we can effectively determine with what rate the attacks overtake the real predictions. In line with the accuracy, we can see that CHROM [8] and POS [44] are the overall the more robust hand-crafted methods. The non-end-to-end (PhySU-Net [31] and CVD [25]) AU supervised methods overall are more robust with the illumination and motion attacks, with large vertical movement as the most challenging attack. The AA case makes the success rate significantly lower. In line with the accuracy, for the non-robust Contrast-Phys [40] the AA case seems to worsen, as the addition of attacks in the training data has the negative effect of increasing noise related inter-data differences that guide the method to learn non-physiological features. + +# 4.3 Discussion + +Our experiments benchmark the current state of rPPG in relation to physical domain attacks, in hand-crafted, supervised and self-supervised settings (AU, AA). Active attacks (I1-I6, M1-M10) have a strong impact on hand-crafted methods as they are less robust to environmental noise, with CHROM [8] and POS [44] being less susceptible as they are based on physiology specific mathematical models. Supervised methods offer more resistance to the attacks, as they are guided by strong supervision to ignore non-physiological features and learn features strongly correlated to the ground truth, thus are able to offer better robustness. Non-end-to-end methods (CVD [25], PhySU-Net [31]) using signal map inputs show the best robustness out of supervised methods, due to the noise + +suppression capabilities of the signal maps. However, non-end-to-end methods depend on face ROI locations, making them less effective with Movement attacks compared to Illumination. For the end-to-end methods, DeepPhys [5] and TS-CAN [19] are less robust than Physnet [46] and Physformer [49], due to their short temporal context architectures. Having the attacks present during training (AA) greatly improves robustness of supervised methods. Most methods follow a similar trend to the AU case, as the innate robustness is supplemented by the data awareness, meaning that the same design factors lead to effectiveness in the AA case as in the AU case. Self-supervised methods (ContrastPhys [40]) on the other hand are very susceptible to attacks, both in the AU case and especially in the AA case, as the attacks negatively influence the training. Opaque Concealment attacks (C1-C3) are more effective on ROI dependant methods such as hand-crafted and non-end-to-end, with the facial mask being more aggravating as it covers a larger physiologically relevant surface, and C3 (mask and hat) being too challenging of a task, except for end-to-end methods in AA case. Regarding make-up (C4-C6), the transparent primer has negligible effect, and that addition of foundation leads to a prominent drop in SNR, but further addition of powder is negligible, meaning that shade is more impactful. Most supervised methods partially adapt to the make-up and have improved SNR in the AA case. The addition of attacks into the training data significantly boost the performance as the networks learn to ignore the specific features related to the attacks (that are not correlated with the ground truth), thus showing the strong necessity of having such physical attacks present in the training set. Nonetheless, depending only on data to defend from such attacks is not feasible and future rPPG methods need to be more secure and innately robust to physical attacks. + +# 5 Conclusion + +We propose ORPDAD, the first rPPG dataset to include comprehensive physical domain attacks. We extensively evaluate thirteen diverse rPPG methods on the attack scenarios. We point out that hand-crafted and self-supervised methods are most susceptible, supervised methods (particularly non-end-to-end) have a certain degree of innate robustness that can be improved with data containing attacks. However, only having more data cannot mitigate this issue, as secure rPPG methods need to be designed to overcome these vulnerabilities. Thus, we hope that our dataset will encourage novel secure solutions. Future work includes the exploration of new robust methods against general physical attacks, and the extension of the study and dataset to include face spoofing attacks, as the studied attacks have the potential to deceive liveness detection based anti-spoofing. + +# Acknowledgements + +This work was supported by the Research Council of Finland Academy Professor project EmotionAI (grants 336116, 345122) and ICT 2023 project TrustFace (grant 345948), and the University of Oulu & Research Council of Finland Profi + +7 (grant 352788), and by Infotech Oulu. As well, the authors acknowledge CSC - IT Center for Science, Finland, for computational resources. We also thank Mengting Wei and all the study volunteers for their contribution. + +# References + +1. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019) +2. Challoner, A., Ramsay, C.: A photoelectric plethysmograph for the measurement of cutaneous blood flow. Physics in Medicine & Biology 19(3), 317 (1974) +3. Chen, M., Jayaweera, A., Wong, C.W., Wu, M.: Identity-Privacy Protection for Facial-rPPG Based Smart Health Research (Jul 2022). https://doi.org/10.36227/techrxiv.20352843.v1, https://www.techrxiv.org/articles/preprint/Identity-Privacy_Protection_for_Facial-rPPG-Based_Smart_Health_Research/20352843/1 +4. Chen, M., Liao, X., Wu, M.: Pulseedit: Editing physiological signals in facial videos for privacy protection. IEEE Transactions on Information Forensics and Security 17, 457-471 (2022) +5. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 349-365 (2018) +6. Cheong, J.H., Xie, T., Byrne, S., Chang, L.J.: Py-feat: Python facial expression analysis toolbox. CoRR abs/2104.03509 (2021), https://arxiv.org/abs/2104.03509 +7. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021). pp. 01-08. IEEE (2021) +8. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering 60(10), 2878-2886 (2013) +9. Deng, J., Guo, J., Ververas, E., Kotsia, I., Zafeiriou, S.: Retinaface: Single-shot multi-level face localisation in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5203-5212 (2020) +0. Elgendi, M., Jonkman, M., De Boer, F.: Frequency bands effects on qrs detection. Biosignals 2003, 2002 (2010) +1. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021) +2. Guo, X., Li, S., Yu, J., Zhang, J., Ma, J., Ma, L., Liu, W., Ling, H.: Pfld: A practical facial landmark detector. arXiv preprint arXiv:1902.10859 (2019) +3. Gupta, A.K., Kumar, R., Birla, L., Gupta, P.: Radiant: Better rppg estimation using signal embeddings and transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4976-4986 (2023) +4. Hernandez-Ortega, J., Fierrez, J., Morales, A., Tome, P.: Time analysis of pulse-based face anti-spoofing in visible and nir. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-552 (2018) +5. Heusch, G., Anjos, A., Marcel, S.: A reproducible study on remote heart rate measurement. arXiv preprint arXiv:1709.00962 (2017) + +16. Lewandowska, M., Rumiński, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: 2011 federated conference on computer science and information systems (FedCSIS). pp. 405-410. IEEE (2011) +17. Li, X., Alikhani, I., Shi, J., Seppanen, T., Junttila, J., Majamaa-Voltti, K., Tulppo, M., Zhao, G.: The obf database: A large face video database for remote physiological signal measurement and atrial fibrillation detection. In: 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018). pp. 242-249. IEEE (2018) +18. Liu, S., Yuen, P.C., Zhang, S., Zhao, G.: 3d mask face anti-spoofing with remote photoplethysmography. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 85-100. Springer (2016) +19. Liu, X., Fromm, J., Patel, S., McDuff, D.: Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems 33, 19400-19411 (2020) +20. Liu, X., Hill, B., Jiang, Z., Patel, S., McDuff, D.: Efficientphys: Enabling simple, fast and accurate camera-based cardiac measurement. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5008-5017 (2023) +21. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12404-12413 (2021) +22. Makowski, D., Pham, T., Lau, Z.J., Brammer, J.C., Lespinasse, F., Pham, H., Schölzel, C., S H Chen, A.: Neurokit2: A python toolbox for neurophysiological signal processing (2020). https://doi.org/10.5281/ZENODO.3597887, https://github.com/neuropsychology/NeuroKit +23. Niu, X., Han, H., Shan, S., Chen, X.: Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14. pp. 562-576. Springer (2019) +24. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing 29, 2409-2423 (2019) +25. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. pp. 295–310. Springer (2020) +26. Niu, X., Zhao, X., Han, H., Das, A., Dantcheva, A., Shan, S., Chen, X.: Robust remote heart rate estimation from face utilizing spatial-temporal attention. In: 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019). pp. 1-8 (2019). https://doi.org/10.1109/FG.2019.8756554 +27. Nowara, E.M., Marks, T.K., Mansour, H., Veeraraghavan, A.: Near-infrared imaging photoplethysmography during driving. IEEE transactions on intelligent transportation systems 23(4), 3589-3600 (2020) +28. Pilz, C.S., Zaunseder, S., Krajewski, J., Blazek, V.: Local group invariance for heart rate estimation from face videos in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1254-1262 (2018) + +29. Poh, M.Z., McDuff, D.J., Picard, R.W.: Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE transactions on biomedical engineering 58(1), 7-11 (2010) +30. Savic, M., Zhao, G.: De-identification of facial videos while preserving remote physiological utility. In: 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023 (2023) +31. Savic, M., Zhao, G.: Physu-net: Long temporal context transformer for rppg with self-supervised pre-training (2024) +32. Savic, M., Zhao, G.: Rs-rppg: Robust self-supervised learning for rppg. In: 2024 18th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2024). IEEE (2024) +33. Speth, J., Vance, N., Flynn, P., Bowyer, K., Czajka, A.: Remote pulse estimation in the presence of face masks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2086-2095 (2022) +34. Speth, J., Vance, N., Flynn, P., Bowyer, K.W., Czajka, A.: Digital and physical-world attacks on remote pulse detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2407-2416 (2022) +35. Speth, J., Vance, N., Flynn, P., Czajka, A.: Non-contrastive unsupervised learning of physiological signals from video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14464-14474 (2023) +36. Speth, J., Vance, N., Sporrer, B., Niu, L., Flynn, P., Czajka, A.: Mspm: A multi-site physiological monitoring dataset for remote pulse, respiration, and blood pressure estimation. arXiv preprint arXiv:2402.02224 (2024) +37. Špetlík, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proceedings of the British machine vision conference, Newcastle, UK. pp. 3-6 (2018) +38. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062. IEEE (2014) +39. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062 (2014) +40. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XII. pp. 492–510. Springer (2022) +41. Sun, Z., Li, X.: Privacy-phys: Facial video-based physiological modification for privacy protection. IEEE Signal Processing Letters 29, 1507-1511 (2022). https://doi.org/10.1109/LSP.2022.3185964 +42. Tang, J., Chen, K., Wang, Y., Shi, Y., Patel, S., McDuff, D., Liu, X.: Mmpd: Multi-domain mobile video physiology dataset. arXiv preprint arXiv:2302.03840 (2023) +43. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Optics express 16(26), 21434-21445 (2008) +44. Wang, W., den Brinker, A.C., Stuijk, S., De Haan, G.: Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering 64(7), 1479-1491 (2016) +45. Wang, W., Shan, C.: Impact of makeup on remote-ppg monitoring. Biomedical Physics & Engineering Express 6(3), 035004 (2020) + +46. Yu, Z., Li, X., Zhao, G.: Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In: British Machine Vision Conference (2019) +47. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 151-160 (2019) +48. Yu, Z., Shen, Y., Shi, J., Zhao, H., Cui, Y., Zhang, J., Torr, P., Zhao, G.: Physformer++: Facial video-based physiological measurement with slowfast temporal difference transformer. International Journal of Computer Vision 131(6), 1307-1330 (2023) +49. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P.H., Zhao, G.: Physformer: facial video-based physiological measurement with temporal difference transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4186-4196 (2022) +50. Zhang, Z., Girard, J.M., Wu, Y., Zhang, X., Liu, P., Ciftci, U., Canavan, S., Reale, M., Horowitz, A., Yang, H., et al.: Multimodal spontaneous emotion corpus for human behavior analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3438-3446 (2016) \ No newline at end of file diff --git a/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/images.zip b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..449f36cdcd99c197c8bf54ae009223ec07182174 --- /dev/null +++ b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ef0582939f4e81cbd9bf7659eaeb9a61f463740de1b6503284309875057f6ee +size 745227 diff --git a/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/layout.json b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a52153e0cea79d7fb7b35ac36321b5f2d0df331 --- /dev/null +++ b/2024/Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)/layout.json @@ -0,0 +1,6376 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 147, + 111, + 467, + 148 + ], + "type": "text", + "content": "Oulu Remote-photoplethysmography Physical Domain Attacks Database (ORPDAD)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 222, + 168, + 391, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 168, + 391, + 180 + ], + "spans": [ + { + "bbox": [ + 222, + 168, + 391, + 180 + ], + "type": "text", + "content": "Marko Savic1 and Guoying Zhao★1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 190, + 474, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 190, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 139, + 190, + 474, + 213 + ], + "type": "text", + "content": "Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland {marko.savic, guoying.zhao}@oulu.fi" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 239, + 455, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 455, + 427 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 455, + 427 + ], + "type": "text", + "content": "Abstract. Remote photoplethysmography (rPPG) is an emerging technology that can detect the pulse rate remotely from face videos. However, it is easily influenced by the recording environment, as robustness to noise is still an open problem. This vulnerability can therefore be exploited to inject fake signals or impair predictions physically. In this study we propose the first dataset containing a wide set of physical domain attack scenarios divided in three categories (illumination, movement, concealment) that directly target the main weaknesses of rPPG. We propose the Oulu rPPG Physical Domain Attacks Database (ORPDAD) as a benchmark for evaluation of robustness to physical attacks. We perform extensive experiments on conventional hand-crafted and deep learning (end-to-end, non-end-to-end, CNN, transformer, self-supervised) methods and study their susceptibility to the attacks. We conclude by discussing the most critical vulnerabilities discovered and stress the importance of designing more secure solutions. Our code and instructions for requesting the dataset (with trained models) are available at: https://github.com/marukosan93/ORPDAD/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "spans": [ + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "text", + "content": "Keywords: Remote Photoplethysmography " + }, + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "text", + "content": " Facial Video " + }, + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "text", + "content": " Physical-Domain Attack " + }, + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 436, + 453, + 459 + ], + "type": "text", + "content": " Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 502, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 647 + ], + "type": "text", + "content": "Physiological signals measured via photoplethysmography (PPG) and electrocardiogram (ECG) are used to determine vital healthcare parameters such as heart rate (HR), heart rate variability (HRV), respiratory rate (RR) and oxygen saturation (SpO2). However, they are limited by the necessity for physical contact that can make them uncomfortable in long term monitoring scenarios, especially with acute skin condition patients or fragile skin neonates, and very inconvenient in natural interactions. On the other hand, Remote photoplethysmography (rPPG) is a technology that allows extracting a signal akin to the blood volume pulse (BVP) without any contact, by utilising ubiquitous RGB cameras to record facial videos. Relying on the same principle as contact PPG [2], rPPG exploits subtle variations in reflected light from the skin that are caused by the cardiac cycle [43]. Several early methods had been proposed that relied" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 270, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 270, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 270, + 665 + ], + "type": "text", + "content": "* indicates corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "on hand-crafted features or blind source separation [8, 16, 28, 29, 43, 44]. However, they lacked environmental robustness and were surpassed by deep learning approaches. Most deep learning methods employed supervised learning via Convolutional Neural Network (CNN) based models [5, 7, 21, 24, 25, 37, 46, 47] or more recently, self-attention based transformer architectures [13, 20, 31, 49]. Self-supervised methods such as [11, 35, 40] have also been a recently proposed to learn without annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 415 + ], + "type": "text", + "content": "Factors such as illumination variation, subject movement, camera sensor noise and occlusion give rise to a poor signal-to-noise ratio (SNR), making robustness to environmental noise challenging. Most rPPG methods developed in recent years try to solve this environmental noise issue, and they are evaluated on benchmark datasets designed to contain a certain degree of noise, e.g. different head motions [38], facial expressions [50], different illumination conditions [1, 15, 42], and also challenging motion and lighting in indoors [23] or in car driving scenarios [27]. A few studies also consider skin occlusions either via opaque masks [33] or translucent make-up [45]. Despite numerous methods and increasing number of datasets, robustness to environmental conditions is still an open research problem. To be widely used and applicable in real world settings, rPPG needs to be both robust and secure. From a security standpoint, lack of environmental robustness can be considered a critical vulnerability that should be addressed. Thus far, security related research in rPPG has been limited to the digital domain, as most works focus on editing the underlying physiological signals [4,41] (considered a digital attack) or de-identifying the data while retaining physiological features [30]. Digital injection of an attack is studied in [34], where a limited physical attack involving lighting is proposed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 416, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 583 + ], + "type": "text", + "content": "The lack of robustness to environmental conditions is a key issue, and we wish to draw attention to the fact that it is a significant vulnerability that attackers can potentially exploit. As rPPG is in its early stages, exposing and patching security vulnerabilities is crucial for the development of robust and secure solutions. We propose a new dataset aiming specifically at potential physical domain attacks that can either spoof the rPPG with a target signal or deteriorate it. Our novel Oulu rPPG Physical Domain Attacks Database (ORPDAD) consists of three different types of attacks (i.e. illumination, movement and concealment) that result in 26 different recording scenarios across 30 subjects. Through these attacks, we physically inject periodical signals that can spoof the rPPG (via a pulsating light and rhythmical movements) or conceal skin to deteriorate the rPPG. The purpose of this dataset is to aid in developing more secure solutions to physical domain attacks and to provide a new benchmark for rPPG robustness to physical attacks. Our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 592, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 137, + 592, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 592, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 137, + 592, + 480, + 616 + ], + "type": "text", + "content": "- ORPDAD is the first dataset to consider a comprehensive range of physical domain rPPG attacks, that can spoof rPPG signals or deteriorate them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 617, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 480, + 640 + ], + "type": "text", + "content": "- Ours is the first work to introduce rhythmical movement as a potential attack, and expands greatly on illumination and concealment based attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 642, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 642, + 480, + 665 + ], + "type": "text", + "content": "- We carry out an extensive comparative study for each attack's effectiveness, with experiments encompassing thirteen hand-crafted, deep learning based" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "type": "text", + "content": "supervised and self-supervised rPPG methods. We analyse and share our findings on attack susceptibility and uncovered weaknesses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 165 + ], + "type": "text", + "content": "- We establish a novel benchmark for physical domain rPPG attacks and share our dataset and materials (data, evaluation protocols, code and baselines)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 239, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 239, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 239, + 197 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 210, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 365 + ], + "type": "text", + "content": "Hand-crafted rPPG: The first rPPG method [43] extracted a signal from the green channel via averaged facial video pixels in a region of interest (e.g. forehead and cheeks). Model based methods have been proposed to better discriminate rPPG signals from motion noise such as CHROM [8] and POS [44], that leverage optical/physiological considerations expressed through mathematical models. Common blind source separation approaches such as PCA [16] and ICA [29] have been applied as well. In LGI [28] a new feature space is proposed that is more robust to motion and illumination changes via differentiable local transformations. However, the assumptions and hand-crafted features of the aforementioned methods did not hold in less constrained scenarios. For example, CHROM [8] assumes a standardized skin color profile, ICA [29] assumes independence of source signals and PCA [16] assumes their uncorrelation. Deep learning has shown to be more effective, especially in less controlled scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 557 + ], + "type": "text", + "content": "Deep learning based rPPG: The first 2D-CNN models extracted the HR from two adjacent frames, such as HR-CNN [37] and DeepPhys [5], only taking spatial information into account. End-to-end spatial-temporal 3D-CNN models such as PhysNet [46], rPPGNet [47] and TS-CAN [19] have been used to exploit the temporal information. To suppress the HR unrelated information spatial-temporal maps are introduced as input, like in non-end-to-end methods such as Niu et al. [26], RhythmNet [24], Dual-GAN [21], BVPNet [7], and CVD [25]. Recently, Efficientphys [20], PhysFormer [49], RADIANT [13] and PhySU-Net [31] have shown promising improvements by leveraging self-attention based transformers. To overcome the need for (expensive) annotated data, self-supervised methods that leverage the data alone have been proposed. They learn contrastively with artificial augmentations [11], via instance-wise sampling Contrast-Phys [40], exploit inter-data differences in a non-contrastive manner SiNC [35] or use a large set of diverse priors RS-rPPG [32]. In addition, rPPG has been widely utilised as an important face anti-spoofing cue [14, 18], as the presence of an rPPG signal can be an effective face liveness indicator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "rPPG Datasets: Given the crucial role of data, both for evaluation and for training supervised methods, rPPG datasets have been essential in the development of new methods. The challenge of the rPPG task on a specific dataset is highly dependent on the recording environment, as less constrained (variable illumination, large motion, unstable fps, compression, etc.) scenarios lead to higher error. PURE [39] contains facial videos with PPG groundtruth, with a total of 60 one-minute-long " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " 30fps videos from 10 subjects. It was recorded under ambient lighting with small motion tasks (steady, talking, slow translation, fast translation, slow rotation, medium rotation). MMSE-HR [50] has 102 videos" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "text", + "content": "(with PPG, blood pressure, RR) of 20-70s recorded with stable lighting from 40 subjects in emotion elicitation experiments, it contains spontaneous facial expressions and head motions. COHFACE [15] consists of 40 participants with 160 total 1-minute " + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "text", + "content": " 20fps recordings (with PPG), where controlled studio and natural ambient lighting are used. OBF [17] contains 200 5-min " + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "text", + "content": " 60fps RGB videos (with PPG, ECG, RR) recorded from 100 diverse subjects, under stable lighting and with minimal movement, with both resting and elevated HR conditions. UBFC-rPPG [1] is a dataset of 50 participants with 2-min " + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "text", + "content": " 30fps recordings each (with PPG). It has variable indoor lighting and contains both 8 at-rest recordings and 42 recordings where participants play a time-sensitive mathematical game to increase their heart rate. VIPL-HR [23] contains 2,378 RGB videos of 20-30s length recorded from 108 subjects, it was recorded in a challenging environment with: different devices, varied and unstable frame rates, large movements and variable lighting. MR-NIRP [27] is the first rPPG dataset containing driving scenarios, it was recorded from 19 subject with 190 total " + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 115, + 482, + 426 + ], + "type": "text", + "content": " 60fps videos (with PPG), both during driving and while parked. On top of the variable lighting, subjects were asked to speak or perform random head movements. MMPD [42] is the first dataset recorded entirely with smartphone cameras (with PPG). It includes 33 subjects with 330 one-minute videos total, recorded under different lighting conditions (LED high, LED low, incandescent, natural), and different motions such as: resting, head rotation, conversation, and walking. Currently, there are numerous rPPG datasets containing different kind of environmental noise. Nevertheless, none of these large scale datasets contain intentional periodical noise as a physical attack medium, via movement or illumination. Furthermore, skin concealment via opaque cloth or transparent make-up is also not present in the aforementioned datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 427, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 654 + ], + "type": "text", + "content": "Security in rPPG: As rPPG deals with sensitive biometric facial videos and physiological signals, security and privacy based research has also been a focus in recent years. Protecting privacy by removing identity related features while preserving physiological features has been attempted first via a Gaussian decomposition that maintains the low frequency spatial features [3]. A visual utility preserving method has also been proposed, that removes identity related features but preserves the physiological and visual utility via a semi-adversarial framework [30]. In contrast to rPPG preservation, PulseEdit [4] and PrivacyPhys [41] were devised to alter the rPPG signal in videos by using small perturbations. The original purpose of the aforementioned works is to conceal identity or the physiological signals, but they can be repurposed to alter the rPPG signal digitally and thus employed as a digital attacks that alter the underlying physiological data to result in a target signal. However, all these works are limited to the digital domain. In [34] digital-domain rPPG attacks are explicitly studied by injecting a target periodical signal. Furthermore, the authors propose a simple pulsating single LED attack by collecting a small dataset. However, most attack frequencies were outside the normal resting heart rate (filtered out during post-processing), the data contained only five two-minute videos from four subjects, attack intensity and range were limited, and the attack was a simple sinusoidal" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "wave. A similar attack was proposed in MSPM [36] involving pulsating light from a monitor. Apart from illumination variation, motion is also a large noise source that datasets try to incorporate in varying degrees. However, no studies have yet tried to inject a specific target signal by utilising periodical motion as a medium. Finally, as rPPG is measured from light reflected from the facial skin, concealing it can be considered as physical domain attack with the aim to deteriorate the signal. In [33], the authors propose the MPM dataset containing 159 3-minute recordings of 54 subjects wearing facial masks, and show the importance of having methods robust to face occlusions, but only consider opaque face masks. In [45] foundation is considered as a translucent mask with 22 subjects, as make-up can affect rPPG readings with a notable decrease in SNR, but the authors only create one scenario involving foundation alone. Current rPPG solutions are not robust to noise and struggle in challenging scenarios, their robustness to physical domain attacks lacks sufficient investigation, as current physical domain attack research is limited in scope and only considers specific instances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 326, + 435, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 435, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 435, + 340 + ], + "type": "text", + "content": "3 Oulu rPPG Physical Domain Attacks Database" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 352, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 531 + ], + "type": "text", + "content": "As rPPG is an emerging technology that deals with sensitive data and could be employed in applications where security is critical (clinical, remote driver monitoring, anti-spoofing, authentication, etc.), it is crucial to critically assess its vulnerabilities early and work towards mitigating potential security exploits. In this work, we focus on the scarcely studied physical-domain attacks. A physical-domain attack assumes that the attacker cannot influence the digital data after acquisition or the rPPG method used for estimation, namely the attacker can only inject noise during the acquisition process using physical means. We are inspired by previous works [33, 34, 36, 45] that deal with a subset of potential attacks in a limited scope. To fully understand and mitigate potential physical attacks to rPPG, we propose the ORPDAD dataset that aims to study a large set of diverse physical attacks within the same recording environment, thus isolating the variable of interest associated with the specific attack. The attacks we propose are simple to execute, but can have a large influence on the rPPG estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 549, + 328, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 328, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 328, + 562 + ], + "type": "text", + "content": "3.1 Motivation and reasons for setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We identify vulnerabilities based on the properties of rPPG measuring, in Fig. 1 we display three major vulnerabilities. Firstly, as rPPG is measured from reflected light, a possible attack vector is the light source. Adding a weak perturbation signal to the incident light source means that it will be present in the estimated rPPG signal as well. Secondly, subject movement induces noise to the extracted signal, consequently having the subject move at a set rhythm will make this added motion perturbation strongly periodical. Therefore, we can leverage both motion and illumination to inject strongly periodical (rPPG like) signals" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 341, + 240 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 341, + 240 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 341, + 240 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 341, + 240 + ], + "type": "image", + "image_path": "2568bf67fec847deda912322786522e2599299977088f778485d85c2d9b5e350.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "type": "text", + "content": "Fig. 1: (Left) Major rPPG measurement environment vulnerabilities and related potential attack vectors. (Right) Data recording setup with equipment as listed in Tab. 1" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 352, + 113, + 481, + 248 + ], + "blocks": [ + { + "bbox": [ + 352, + 113, + 481, + 248 + ], + "lines": [ + { + "bbox": [ + 352, + 113, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 352, + 113, + 481, + 248 + ], + "type": "image", + "image_path": "793d37da45e10660132ca2ed54f4d0ab5cd646506741d2456b546be52fcdb4ab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 147, + 310, + 467, + 426 + ], + "blocks": [ + { + "bbox": [ + 154, + 289, + 458, + 302 + ], + "lines": [ + { + "bbox": [ + 154, + 289, + 458, + 302 + ], + "spans": [ + { + "bbox": [ + 154, + 289, + 458, + 302 + ], + "type": "text", + "content": "Table 1: Equipment used for data collection, also shown in Fig. 1 (Right)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 310, + 467, + 426 + ], + "lines": [ + { + "bbox": [ + 147, + 310, + 467, + 426 + ], + "spans": [ + { + "bbox": [ + 147, + 310, + 467, + 426 + ], + "type": "table", + "html": "
Device/ToolDescription/Specifications
1) RGB CameraBlackmagic Ursa Mini 4K EF1920x1080 30fps 500Mbps f8.7 360° ISO400 WB4000K fl24mm
2) 2×Studio lightsAputure LightStorm LS 1c 3500lux 4000K
3) 2×Custom LED panels1600lm 11.5W 4000K (96×2835SMD LED) 20×13.3cm²
4) LED Control CircuitRPi 4B + IRL540N MOSFET + Resistors + 12V3A PSU
5) Finger-oximeter PPGBeurer PO80 120Hz
6) ECGBittium Faros 180 1Khz 1ch
7) MetronomeGoogle Metronome Web App
8) Concealment (C1-C3)Blue Medical Mask and Black Beanie Hat
9) Concealment (C4-C6)Maybelline Fit Me Matte Poreless \nPrimer + Foundation 220 + Powder 220
", + "image_path": "fa32596eb79328c39dc8c0f5607cf2aedbc81a04974d4242d9d242e5eead67a2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 448, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 498 + ], + "type": "text", + "content": "into the data. Thirdly, the light measured is reflected from the skin, so either blocking the light (with an opaque cloth) or attenuating it (with make-up) can affect the rPPG negatively. Consequently, we define these three physical-domain attack categories: Illumination (I), Movement (M) and Concealment (C)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 514, + 227, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 227, + 526 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 227, + 526 + ], + "type": "text", + "content": "3.2 Study design" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We record data from 30 healthy (no history of heart disease) volunteers under different scenarios based on the three aforementioned categories of attacks. All the participants gave written informed consent. For the active attacks we choose " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "50~\\mathrm{bpm}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "100~\\mathrm{bpm}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " as attack frequencies as they can be considered borderline resting heart rates, thus plausible but different from the real HR in most recordings. We avoid results being biased by the real HR trend by averaging results over both frequencies. We record 70 second videos for each scenario, for a total of 26 scenarios per subject, resulting in 780 videos equivalent to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\approx 15.2\\mathrm{h}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " of footage. For the synchronised ground truth physiological signals, we record both PPG waveforms via a finger-oximeter and ECG waveforms via an ECG device. Based on the defined attack categories, we define the following scenarios:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 294, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 294, + 251 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 294, + 251 + ], + "type": "image", + "image_path": "81e662409a576a7160298d48a912ddf8f99f101b4c74ebf23d4bba2e4568f40d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 260, + 482, + 305 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 305 + ], + "type": "text", + "content": "Fig. 2: Raw signal (GREEN [43]) and its spectrogram for diverse attacks. For active illumination (I1, I6) and motion based attacks (M1, M9), the target attack frequency is clearly visible in the spectrogram and can take over the prediction. On the other hand, passive concealment attacks (C1, C6) significantly affect the estimation's SNR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 116, + 480, + 251 + ], + "blocks": [ + { + "bbox": [ + 317, + 116, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 317, + 116, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 480, + 251 + ], + "type": "image", + "image_path": "04fb62e317ab8ff284ecdd5d2f5174836b124212928b2d92068f754c2741cb61.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 328, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 352 + ], + "type": "text", + "content": "Still (S1-S3): Subject is recorded at a still and resting state before each attack category. Motion is kept at a comfortable minimum." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 494, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 494, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 494, + 411 + ], + "type": "text", + "content": "Illumination (I1-I6): We combine the two attack frequencies (50bpm, 100bpm) and three different intensity settings (S=7%, M=14% and L=21% of the attack device max brightness), resulting in I1=50S, I2=50M, I3=50L, I4=100S, I5=100M, I6=100L. The intensities are much weaker than the base lighting, and cause perturbations are visually imperceptible in the recordings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 412, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 483 + ], + "type": "text", + "content": "Movement (M1-M11): We explore simple movements such as small vertical (SV), large vertical (LV), small horizontal (SH), large horizontal (LH), mouth open/close (M) at the two target frequencies (50bpm, 100bpm). Resulting in M1=50SV, M2=50LV, M3=50SH, M4=50LH, M5=50M, M6=100SV, M7=100LV, M8=100SH, M9=100LH, M10=100M. We also include natural talking motion M11 that is not at a set attack frequency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 484, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 532 + ], + "type": "text", + "content": "Concealment (C1-C6): The first three opaque (C1-C3) scenarios involve wearing a forehead concealing beanie C1, a facial mask C2 and then both C3. The transparent concealment scenarios (C4-C6) contain progressive application of make-up with primer C4, foundation C5 and setting powder C6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 532, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 556 + ], + "type": "text", + "content": "In Fig. 2 we visualise a subset of the aforementioned attacks and their effect on the raw signal from the GREEN [43] method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 574, + 269, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 269, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 269, + 586 + ], + "type": "text", + "content": "3.3 Equipment and setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To study each attack vector in an isolated manner, we set up a controlled recording environment as to minimise the influence of external factors that are not related to the attack. Thus, we keep the same base environment for all subjects and scenarios, enabling us to study the specific contribution induced by the attacks. For the base video recordings, we use a professional RGB camera and two fixed LED lights. We record the physiological signals with a finger PPG oximeter" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "and a portable ECG belt device. All the recording devices are synchronised by keeping internal clock times consistent. The subjects are seated at " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\approx 1\\mathrm{m}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " from the camera and the lights are positioned above the camera and illuminate the subject at a " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " angle at " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\approx 1.5\\mathrm{m}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " distance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "content": "For the illumination attacks, we design two modulated light panels that each operate at " + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "inline_equation", + "content": "1600\\mathrm{lm}" + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "content": " maximum, by using LED strips and a control circuit. These panels allow the injected light signal to be more uniformly distributed over the face surface, instead of attacking only part of it (single led case). We pilot the panels with Pulse Width Modulation (PWM) signals, thus being able to create a light source that can output any waveform and at any fraction of the max intensity. For this experiment, we simulate a typical rPPG waveform, with the main frequency component placed at the target frequency and a second harmonic at " + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "content": " of the peak intensity. For the L, M, H settings we use " + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "content": " of the max power and set the lights at " + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "inline_equation", + "content": "\\approx 1\\mathrm{m}" + }, + { + "bbox": [ + 130, + 164, + 481, + 342 + ], + "type": "text", + "content": " distance from the subject. For the movement scenarios, we simply utilise an audio metronome and instruct the subjects to move according to the set rhythm and movement type. For opaque concealment, we use medical face masks and black beanie hats. For translucent concealment we choose a primer, foundation and setting powder, and keep the brand and shade (220 natural beige) consistent among all subjects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 481, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 368 + ], + "type": "text", + "content": "We provide a detailed list of all the equipment and settings utilised for recording in Tab. 1, and show all the equipment in Fig. 1 (right)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 386, + 230, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 230, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 230, + 400 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "content": "We assess the effects of each physical attack on a large set of rPPG methods that represent the past and current state of the art, as they include both common hand-crafted and CNN based methods, as well more recent transformer based and self-supervised works. We choose the following hand-crafted non-learning based methods: GREEN [43], CHROM [8], POS [44], LGI [28], ICA [29] and PCA [16]. As for learning based CNN methods, we choose the end-to-end methods DeepPhys [5], TS-CAN [19] and PhysNet [46], and the non-end-to-end method CVD [25]. We also include two transformer based methods that are Physformer [49] (end-to-end) and PhySU-Net [31] (non-end-to-end). Finally, we include an end-to-end self-supervised method Contrast-Phys [40]. We faithfully re-implement the aforementioned methods and fine-tune them to obtain the best performance on our dataset. All our code can be found at https://github.com/marukosan93/ORPDAD/, with instructions for requesting the dataset (containing both raw data, preprocessed data and trained models)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 597, + 261, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 261, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 261, + 609 + ], + "type": "text", + "content": "4.1 Experimental setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Preprocessing: Firstly, we utilise PyFeat [6] with RetinaFace [9] model for face detection and PFLD [12] model for landmark localisation, the landmarks are then stabilised with a 5-point moving average filter. For the end-to-end methods, we crop the images to " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " so that the face is centered. For hand-crafted" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "methods we use the landmarks to define a ROI (including forehead, cheeks and chin) and average the pixels to the raw signal input. Finally, as both non-end-to-end methods CVD [25] and PhySU-Net [31] utilise multi-scale spatial-temporal maps (MSTmaps) as input, we follow the suggested processing [25] to generate the input MSTmaps. For the ground-truth signal, as ECG is considered the gold standard, we choose to primarily use ECG for experimentation. We first filter the ECG with the method in [10] via NeuroKit2 [22], and then apply the 10-point average method from [48] to get a resulting signal containing the same physiological features but with a shape better correlated to the rPPG signal. All ground-truth signals, original videos and pre-processed videos (landmarks, MSTmaps, cropped) will be included in the shared dataset version." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 354 + ], + "type": "text", + "content": "Metrics: We follow previous works by using mean absolute error (MAE), root-mean-square error (RMSE), and Pearson's correlation coefficient (R). Furthermore, we calculate the SNR, defined as the ratio between the power in the dominant frequency and second harmonics versus the remaining power [8]. Note that good SNR does not coincide with accurate prediction, as it measures the spectrum sparsity around the predicted rPPG (without considering the true label). For active attacks, we also calculate success rate, defined as the percentage of times when the predicted HR is closer to the two attack frequencies than the ground truth, meaning that the attack successfully hijacked the prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 523 + ], + "type": "text", + "content": "Protocol: To ensure fair validation we adopt a 5-fold subject exclusive protocol for all experiments. Additionally, we consider two different training conditions. Attack Unaware (AU), where the learning based methods are trained only on the still (S) scenarios and are not exposed to attacks during training, but only during testing. Attack Aware (AA), in this case, the training data contains the attack scenarios so that we can assess whether awareness improves robustness. For AA, we add an attack frequency exclusive cross-validation to the protocol, we train on all passive scenarios and active scenarios with one attack frequency, and then test on the non-seen attack frequency. As the attack frequencies are fixed, supervised methods could easily cheat by learning to not predict the seen specific attack frequency, without learning any useful features from the attacks. Our AA protocol is aware of the attacks, but not of the specific attack frequency. In testing 30 second non-overlapping segments are evaluated, and we calculate the HR by finding the highest PSD peak." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 539, + 270, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 270, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 270, + 551 + ], + "type": "text", + "content": "4.2 Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 592 + ], + "type": "text", + "content": "We analyse each attack category separately, regarding the signal deterioration in terms of accuracy and SNR. Subsequently, we study the success rate of the active attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Firstly, we evaluate all the studied methods on the illumination scenarios, as shown in Tab. 2. On the base scenarios (S1-S3) almost all methods (except GREEN [43]) have very good performance (" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\leq 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " RMSE) as the base task is not very challenging by design, due to the controlled environment. It stands out noticeably that hand-crafted methods are very susceptible to the illumination attack as most methods have a large drop in accuracy, with only the skin" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 157, + 427, + 385 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Results on illumination attacks I1-I6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 157, + 427, + 385 + ], + "lines": [ + { + "bbox": [ + 187, + 157, + 427, + 385 + ], + "spans": [ + { + "bbox": [ + 187, + 157, + 427, + 385 + ], + "type": "table", + "html": "
RMSE↓ |SNR↑S1-S3 (still)I1+I4 (small)I2+I5 (med.)I3+I6 (large)
GREEN [43]□8.57 -7.425.7 -6.727.9 -4.127.7 -2.5
CHROM [8]□0.91 -3.71.87 -4.713.3 -4.419.5 -4.3
POS [44]□0.70 -3.45.77 -4.411.4 -4.813.8 -4.8
LGI [28]□1.05 -3.58.16 -4.420.7 -4.225.9 -3.3
ICA [29]□1.14 -2.922.3 -3.727.6 -1.427.7 0.2
PCA [16]□1.35 -5.110.5 -6.215.6 -6.513.1 -7.2
DeepPhys [5]◆AU1.02 -2.75.53 -3.922.6 -3.426.8 -2.4
AA0.79 -2.80.63 -3.78.23 -4.013.9 -4.4
TS-CAN [19]◆AU1.16 -3.66.31 -4.723.5 -4.127.4 -3.1
AA0.88 -3.50.76 -3.91.09 -4.25.90 -4.5
PhysNet [46]◆AU0.88 -3.14.58 -4.213.2 -4.616.1 -4.7
AA0.55 -2.80.51 -3.10.46 -3.51.61 -4.1
CVD [25]◇AU1.03 -3.11.51 -3.95.47 -4.67.17 -5.2
AA0.79 -2.41.10 -2.70.95 -3.04.08 -3.4
Physformer [49]◆AU0.95 -3.01.01 -3.810.0 -4.514.8 -4.6
AA1.34 -2.41.59 -2.70.57 -3.17.51 -3.8
PhySU-Net [31]◇AU0.74 -1.90.62 -2.95.71 -3.76.24 -4.0
AA0.50 -1.80.93 -2.11.41 -2.63.15 -2.7
Contrast-Phys [40]▼AU0.95 -2.81.51 -3.813.9 -4.317.2 -4.4
AA20.1 -3.620.9 -3.721.6 -4.121.9 -4.5
", + "image_path": "25ee8e94706d714e1ebdb9e916c4a2c61ceff14d3595b21b0ebb0d890e8e4d3d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 411, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 639 + ], + "type": "text", + "content": "model based CHROM [8] and POS [44] having comparable performance only with the weakest attack (I1+I4). Moreover, decrease in accuracy does not necessarily mean decrease in SNR, e.g. ICA [29] improves in SNR by having a higher intensity attack, as the output spectrum contains more power in the dominant attack frequency. For supervised deep learning methods there is an overall better degree of innate robustness as the learned features are more complex and physiology specific. However, in the AU case, except for the weakest setting (I1+I4), the supervised methods are significantly affected by the attack. The illumination based attacks affect non-end-to-end methods CVD [25] and PhySU-Net [31] less, as the spatial-temporal map inputs are more resistant to the illumination variation. When adding the attacks into the training data (AA), the performance significantly improves for the supervised methods, as the networks learn to ignore the features specific to the attack, showing the important role of new data in protection against physical attacks. On the other hand, self-supervised Contrast-Phys [40] is very receptive to the attacks as it focuses on salient frequency features without supervision. In the AA case, the learned features lose their physiological meaning as the method learns from inter-data differences that are most pronounced between the attacks, leading to inaccurate predictions even for the non-challenging still scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Secondly, we study the movement based attacks as shown in Tab. 3. The hand-crafted methods all have a significant performance drop, that is closely cor-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 481, + 384 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Results on movement attacks M1-M11. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 481, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 481, + 384 + ], + "type": "table", + "html": "
RMSE↓ |SNR↑S1-S3(still)M1+M6(s. ver.)M2+M7(l. ver.)M3+M8(s. hor.)M4+M9(l. hor.)M5+M10(mouth)M11(talk)
GREEN [43]□8.57 -7.426.6 -4.426.8 -2.823.5 -6.731.0 -5.626.9 -5.223.1 -8.7
CHROM [8]□0.91 -3.710.4 -4.911.5 -5.86.34 -5.019.3 -5.910.6 -5.51.35 -6.8
POS [44]□0.70 -3.48.60 -4.813.8 -5.78.22 -4.814.5 -6.110.3 -5.41.20 -6.5
LGI [28]□1.05 -3.518.0 -4.926.1 -3.811.4 -4.924.2 -5.112.7 -5.35.45 -6.7
ICA [29]□1.14 -2.925.5 -3.324.9 -2.721.8 -4.726.3 -4.123.4 -4.03.99 -6.4
PCA [16]□1.35 -5.113.4 -5.416.6 -7.314.5 -5.923.9 -6.413.5 -5.76.13 -7.1
DeepPhys [5]◆AU1.02 -2.719.4 -4.626.3 -3.523.7 -4.625.1 -4.717.2 -4.95.65 -6.6
AA0.79 -2.86.16 -3.915.9 -5.115.0 -4.620.6 -5.39.38 -4.17.47 -5.4
TS-CAN [19]◆AU1.16 -3.624.2 -4.526.8 -3.224.4 -5.326.0 -4.819.8 -5.01.79 -7.2
AA0.88 -3.55.99 -4.318.5 -5.315.2 -5.021.0 -5.811.7 -4.71.45 -5.7
PhysNet [46]◆AU0.88 -3.117.1 -5.022.6 -4.99.45 -5.320.6 -6.38.86 -4.56.21 -5.0
AA0.55 -2.83.61 -3.311.5 -4.43.62 -3.79.11 -5.50.71 -3.63.37 -4.2
CVD [25]◇AU1.03 -3.110.6 -5.516.8 -7.46.62 -4.211.2 -6.15.23 -5.18.74 -4.9
AA0.79 -2.44.64 -3.412.9 -5.80.83 -2.75.96 -4.52.97 -3.66.30 -3.7
Physformer [49]◆AU0.95 -3.011.0 -4.318.3 -5.012.0 -4.820.5 -5.78.55 -4.68.68 -5.2
AA1.34 -2.40.69 -2.610.8 -4.07.30 -3.111.7 -4.85.94 -3.25.67 -3.9
PhySU-Net [31]◇AU0.74 -1.95.93 -4.215.2 -6.53.58 -3.99.48 -6.20.69 -3.67.71 -4.2
AA0.50 -1.81.31 -2.39.17 -3.80.69 -2.27.04 -4.20.62 -2.27.56 -3.1
Contrast-Phys [40]▼AU0.95 -2.815.0 -4.821.6 -4.66.84 -4.918.1 -5.99.54 -4.63.41 -5.0
AA20.1 -3.625.5 -4.627.1 -5.827.8 -5.228.8 -5.422.3 -4.424.1 -5.0
", + "image_path": "43ce97e7aeafb570fc04f5fdf47dc57e291018a02abdae9e575f9e73137ef45a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 412, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 412, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 133, + 412, + 482, + 555 + ], + "type": "text", + "content": "related to the movement intensity, as larger movements yield larger errors. There is a contrast between the strongly periodical mouth active attack (M5+M10) and the normal talking motion (M11), as the performance is better when the noise is not periodical (e.g. CHROM [8] and POS [44] have comparable performance to the still scenarios), even though the SNR is lower signifying a noisier spectrum. All the supervised deep learning methods are severely affected by the attack, with non-end-to-end PhySU-Net [31] and CVD [25] having lower accuracy drops overall. Similarly to the illumination attacks, in the AA case the performance significantly improves for supervised methods, as the networks learn to ignore attack specific features to a certain degree. On the other hand, for the AU susceptible Contrast-Phys [40] the addition of attack awareness is even aggravating, similarly as in the illumination case." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 558, + 482, + 665 + ], + "type": "text", + "content": "Thirdly, we study the concealment based attacks as shown in Tab. 4 and visualise the SNR changes in Fig. 3. In this case, hand-crafted methods obtain variable performance on partial concealment (C1 and C2), with C2 having strong drops in accuracy for LGI [28], CHROM [8] and PCA [16]. This is due to the medical mask covering a larger surface of the ROI than the forehead covering hat. In case of total coverage C3, due to the small portion of uncovered skin the task is extremely challenging and leads to low accuracy. There is an overall decrease in SNR when applying the concealment, as the physiologically irrelevant surface is larger. In the make-up related attacks (C4-C6), the performance only" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 481, + 384 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 4: Results on concealment attacks C1-C6. Symbols denote: □ Hand-crafted, Supervised ◆ End-to-end and ◇ Non-end-to-end, ▼ Self-supervised. RMSE and SNR are shown, complete metrics including MAE and R are in supplement." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 481, + 384 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 481, + 384 + ], + "type": "table", + "html": "
RMSE↓|SNR↑S1-S3(still)C1(hat)C2(mask)C3(both)C4(primer)C5(found.)C6(powder)
GREEN [43]□8.57 -7.414.0 -8.718.9 -9.121.9 -9.29.46 -7.712.1 -8.813.7 -8.7
CHROM [8]□0.91 -3.70.72 -5.311.1 -7.817.6 -9.50.87 -4.11.16 -5.20.95 -5.3
POS [44]□0.70 -3.40.69 -4.70.96 -6.023.7 -10.0.57 -3.50.94 -4.60.80 -4.6
LGI [28]□1.05 -3.50.70 -4.914.1 -8.722.0 -9.60.84 -3.61.13 -4.90.92 -4.9
ICA [29]□1.14 -2.90.81 -3.81.03 -5.217.7 -9.20.75 -3.01.05 -4.40.87 -4.2
PCA [16]□1.35 -5.11.37 -5.424.1 -9.422.0 -9.01.08 -4.87.78 -6.08.32 -6.1
DeepPhys [5]AU1.02 -2.70.77 -3.415.3 -9.418.9 -9.30.76 -2.80.84 -3.60.75 -3.4
AA0.79 -2.80.82 -3.31.18 -4.00.85 -5.80.73 -2.80.80 -3.80.81 -3.7
TS-CAN [19]◆AU1.16 -3.61.05 -4.918.4 -9.017.7 -8.90.77 -3.70.91 -4.91.21 -4.5
AA0.88 -3.50.93 -4.01.48 -4.43.35 -6.10.76 -3.41.01 -4.81.38 -4.7
PhysNet [46]◆AU0.88 -3.10.50 -3.30.67 -4.514.8 -7.60.71 -3.10.76 -4.00.60 -3.7
AA0.55 -2.80.50 -3.10.51 -3.20.79 -3.90.54 -2.80.60 -3.60.52 -3.3
CVD [25]◇AU1.03 -3.11.24 -3.57.07 -7.510.4 -8.57.76 -3.11.05 -3.70.75 -3.8
AA0.79 -2.40.82 -2.91.16 -3.88.43 -7.10.87 -2.31.03 -2.90.73 -3.0
Physformer [49]◆AU0.95 -3.00.94 -3.410.8 -6.616.1 -7.90.71 -3.00.68 -3.90.66 -3.6
AA1.34 -2.41.87 -2.81.09 -3.12.33 -3.90.54 -2.40.69 -3.20.59 -2.8
PhySU-Net [31]◇AU0.74 -1.90.58 -2.86.47 -7.218.7 -8.70.67 -1.90.57 -3.00.55 -2.7
AA0.50 -1.80.62 -2.30.84 -2.810.3 -6.70.62 -1.80.63 -2.50.51 -2.1
Contrast-Phys [40]▼AU0.95 -2.80.73 -3.413.8 -8.713.6 -9.00.81 -2.80.72 -4.30.86 -4.4
AA20.1 -3.618.9 -3.437.0 -6.335.8 -6.020.1 -3.418.3 -3.616.7 -3.5
", + "image_path": "b218a268a7696c855f1acf0281e83bef019ab2045bea4ab2615f62a32a4cfa09.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "text", + "content": "significantly drops for GREEN [43] and PCA [16] that are more susceptible to the drop in SNR. For the fully transparent primer C4 (optically similar to a facial creme), the drop in SNR is negligible not significantly affecting the rPPG signal. For the foundation C5 and powder C6 scenarios, the drop in SNR is very noticeable (" + }, + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\approx 1dB" + }, + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "text", + "content": " on average) due to the changed optical properties of the skin. Between C5 and C6 (same shade, but C6 has an additional layer) the difference in negligible, meaning that with regards to the rPPG the shade is most relevant. Similarly, for AU supervised deep learning methods the facial mask is more damaging to the accuracy of non-end-to-end methods, as they rely on the ROI only, and C3 is too challenging of a task. For make-up the behaviour is analogous to the hand-crafted case, with C4 being negligible, and C5 and C6 having a similar overall SNR drop (" + }, + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\approx 1dB" + }, + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "text", + "content": "). In the AA case, all methods can adapt to the partial attack (C1, C2), and surprisingly end-to-end methods are even able to obtain good performance on the challenging C3 scenario, as they are able to find alternative physiologically relevant patches of skin. Regarding make-up C5 and C6, in the AA case SNR increases noticeably (" + }, + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\approx 0.5dB" + }, + { + "bbox": [ + 132, + 426, + 481, + 666 + ], + "type": "text", + "content": ") as the networks partially mitigate the change in skin shade. As for the self-supervised Contrast-Phys [40], it is vulnerable to C2 and C3 in the AU case. Analogously to the previous two attack categories, in the AA case the features decline in physiological meaning." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 459, + 304 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 459, + 304 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 459, + 304 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 459, + 304 + ], + "type": "image", + "image_path": "65f411554251c62ee182cc2713a978effde1b57c30c8448ded6477c04a751200.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 313, + 482, + 347 + ], + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 347 + ], + "type": "text", + "content": "Fig. 3: SNR for passive concealment attacks (C1-C6) and success rates for active attacks (I1-I6, M1-M10). Higher SNR and lower success rates (lighter color intensity) correspond to better robustness." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 453, + 116, + 479, + 250 + ], + "blocks": [ + { + "bbox": [ + 453, + 116, + 479, + 250 + ], + "lines": [ + { + "bbox": [ + 453, + 116, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 453, + 116, + 479, + 250 + ], + "type": "image", + "image_path": "b46a47b0c8a2817e0f4c3890db7f47d3098084324a4bc87fcf00714fd50a584f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 372, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 504 + ], + "type": "text", + "content": "Lastly, we also analyse the success rate of the active attacks, as shown in Fig. 3, using the success rate metric (from Sec. 4.1) we can effectively determine with what rate the attacks overtake the real predictions. In line with the accuracy, we can see that CHROM [8] and POS [44] are the overall the more robust hand-crafted methods. The non-end-to-end (PhySU-Net [31] and CVD [25]) AU supervised methods overall are more robust with the illumination and motion attacks, with large vertical movement as the most challenging attack. The AA case makes the success rate significantly lower. In line with the accuracy, for the non-robust Contrast-Phys [40] the AA case seems to worsen, as the addition of attacks in the training data has the negative effect of increasing noise related inter-data differences that guide the method to learn non-physiological features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 523, + 214, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 214, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 214, + 534 + ], + "type": "text", + "content": "4.3 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Our experiments benchmark the current state of rPPG in relation to physical domain attacks, in hand-crafted, supervised and self-supervised settings (AU, AA). Active attacks (I1-I6, M1-M10) have a strong impact on hand-crafted methods as they are less robust to environmental noise, with CHROM [8] and POS [44] being less susceptible as they are based on physiology specific mathematical models. Supervised methods offer more resistance to the attacks, as they are guided by strong supervision to ignore non-physiological features and learn features strongly correlated to the ground truth, thus are able to offer better robustness. Non-end-to-end methods (CVD [25], PhySU-Net [31]) using signal map inputs show the best robustness out of supervised methods, due to the noise" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 117, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 117, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 117, + 482, + 416 + ], + "type": "text", + "content": "suppression capabilities of the signal maps. However, non-end-to-end methods depend on face ROI locations, making them less effective with Movement attacks compared to Illumination. For the end-to-end methods, DeepPhys [5] and TS-CAN [19] are less robust than Physnet [46] and Physformer [49], due to their short temporal context architectures. Having the attacks present during training (AA) greatly improves robustness of supervised methods. Most methods follow a similar trend to the AU case, as the innate robustness is supplemented by the data awareness, meaning that the same design factors lead to effectiveness in the AA case as in the AU case. Self-supervised methods (ContrastPhys [40]) on the other hand are very susceptible to attacks, both in the AU case and especially in the AA case, as the attacks negatively influence the training. Opaque Concealment attacks (C1-C3) are more effective on ROI dependant methods such as hand-crafted and non-end-to-end, with the facial mask being more aggravating as it covers a larger physiologically relevant surface, and C3 (mask and hat) being too challenging of a task, except for end-to-end methods in AA case. Regarding make-up (C4-C6), the transparent primer has negligible effect, and that addition of foundation leads to a prominent drop in SNR, but further addition of powder is negligible, meaning that shade is more impactful. Most supervised methods partially adapt to the make-up and have improved SNR in the AA case. The addition of attacks into the training data significantly boost the performance as the networks learn to ignore the specific features related to the attacks (that are not correlated with the ground truth), thus showing the strong necessity of having such physical attacks present in the training set. Nonetheless, depending only on data to defend from such attacks is not feasible and future rPPG methods need to be more secure and innately robust to physical attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 434, + 220, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 220, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 220, + 447 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 459, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 591 + ], + "type": "text", + "content": "We propose ORPDAD, the first rPPG dataset to include comprehensive physical domain attacks. We extensively evaluate thirteen diverse rPPG methods on the attack scenarios. We point out that hand-crafted and self-supervised methods are most susceptible, supervised methods (particularly non-end-to-end) have a certain degree of innate robustness that can be improved with data containing attacks. However, only having more data cannot mitigate this issue, as secure rPPG methods need to be designed to overcome these vulnerabilities. Thus, we hope that our dataset will encourage novel secure solutions. Future work includes the exploration of new robust methods against general physical attacks, and the extension of the study and dataset to include face spoofing attacks, as the studied attacks have the potential to deceive liveness detection based anti-spoofing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 609, + 230, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 230, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 230, + 621 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "This work was supported by the Research Council of Finland Academy Professor project EmotionAI (grants 336116, 345122) and ICT 2023 project TrustFace (grant 345948), and the University of Oulu & Research Council of Finland Profi" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "7 (grant 352788), and by Infotech Oulu. As well, the authors acknowledge CSC - IT Center for Science, Finland, for computational resources. We also thank Mengting Wei and all the study volunteers for their contribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 194, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 227 + ], + "type": "text", + "content": "1. Bobbia, S., Macwan, R., Benezeth, Y., Mansouri, A., Dubois, J.: Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters 124, 82-90 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 250 + ], + "type": "text", + "content": "2. Challoner, A., Ramsay, C.: A photoelectric plethysmograph for the measurement of cutaneous blood flow. Physics in Medicine & Biology 19(3), 317 (1974)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 304 + ], + "type": "text", + "content": "3. Chen, M., Jayaweera, A., Wong, C.W., Wu, M.: Identity-Privacy Protection for Facial-rPPG Based Smart Health Research (Jul 2022). https://doi.org/10.36227/techrxiv.20352843.v1, https://www.techrxiv.org/articles/preprint/Identity-Privacy_Protection_for_Facial-rPPG-Based_Smart_Health_Research/20352843/1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "type": "text", + "content": "4. Chen, M., Liao, X., Wu, M.: Pulseedit: Editing physiological signals in facial videos for privacy protection. IEEE Transactions on Information Forensics and Security 17, 457-471 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "type": "text", + "content": "5. Chen, W., McDuff, D.: Deepphys: Video-based physiological measurement using convolutional attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 349-365 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 371, + 480, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 401 + ], + "type": "text", + "content": "6. Cheong, J.H., Xie, T., Byrne, S., Chang, L.J.: Py-feat: Python facial expression analysis toolbox. CoRR abs/2104.03509 (2021), https://arxiv.org/abs/2104.03509" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 403, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 446 + ], + "type": "text", + "content": "7. Das, A., Lu, H., Han, H., Dantcheva, A., Shan, S., Chen, X.: Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In: 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021). pp. 01-08. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "8. De Haan, G., Jeanne, V.: Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering 60(10), 2878-2886 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 469, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 501 + ], + "type": "text", + "content": "9. Deng, J., Guo, J., Ververas, E., Kotsia, I., Zafeiriou, S.: Retinaface: Single-shot multi-level face localisation in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5203-5212 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "text", + "content": "0. Elgendi, M., Jonkman, M., De Boer, F.: Frequency bands effects on qrs detection. Biosignals 2003, 2002 (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "1. Gideon, J., Stent, S.: The way to my heart is through contrastive learning: Remote photoplethysmography from unlabeled video. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3995-4004 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "text", + "content": "2. Guo, X., Li, S., Yu, J., Zhang, J., Ma, J., Ma, L., Liu, W., Ling, H.: Pfld: A practical facial landmark detector. arXiv preprint arXiv:1902.10859 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "text", + "content": "3. Gupta, A.K., Kumar, R., Birla, L., Gupta, P.: Radiant: Better rppg estimation using signal embeddings and transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4976-4986 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "4. Hernandez-Ortega, J., Fierrez, J., Morales, A., Tome, P.: Time analysis of pulse-based face anti-spoofing in visible and nir. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-552 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "5. Heusch, G., Anjos, A., Marcel, S.: A reproducible study on remote heart rate measurement. arXiv preprint arXiv:1709.00962 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "text", + "content": "16. Lewandowska, M., Rumiński, J., Kocejko, T., Nowak, J.: Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: 2011 federated conference on computer science and information systems (FedCSIS). pp. 405-410. IEEE (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 217 + ], + "type": "text", + "content": "17. Li, X., Alikhani, I., Shi, J., Seppanen, T., Junttila, J., Majamaa-Voltti, K., Tulppo, M., Zhao, G.: The obf database: A large face video database for remote physiological signal measurement and atrial fibrillation detection. In: 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018). pp. 242-249. IEEE (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "text", + "content": "18. Liu, S., Yuen, P.C., Zhang, S., Zhao, G.: 3d mask face anti-spoofing with remote photoplethysmography. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14. pp. 85-100. Springer (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "19. Liu, X., Fromm, J., Patel, S., McDuff, D.: Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems 33, 19400-19411 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 297, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 340 + ], + "type": "text", + "content": "20. Liu, X., Hill, B., Jiang, Z., Patel, S., McDuff, D.: Efficientphys: Enabling simple, fast and accurate camera-based cardiac measurement. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5008-5017 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "21. Lu, H., Han, H., Zhou, S.K.: Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12404-12413 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 375, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 418 + ], + "type": "text", + "content": "22. Makowski, D., Pham, T., Lau, Z.J., Brammer, J.C., Lespinasse, F., Pham, H., Schölzel, C., S H Chen, A.: Neurokit2: A python toolbox for neurophysiological signal processing (2020). https://doi.org/10.5281/ZENODO.3597887, https://github.com/neuropsychology/NeuroKit" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 419, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 464 + ], + "type": "text", + "content": "23. Niu, X., Han, H., Shan, S., Chen, X.: Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14. pp. 562-576. Springer (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "24. Niu, X., Shan, S., Han, H., Chen, X.: Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing 29, 2409-2423 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "text", + "content": "25. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Video-based remote physiological measurement via cross-verified feature disentangling. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. pp. 295–310. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "text", + "content": "26. Niu, X., Zhao, X., Han, H., Das, A., Dantcheva, A., Shan, S., Chen, X.: Robust remote heart rate estimation from face utilizing spatial-temporal attention. In: 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019). pp. 1-8 (2019). https://doi.org/10.1109/FG.2019.8756554" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "type": "text", + "content": "27. Nowara, E.M., Marks, T.K., Mansour, H., Veeraraghavan, A.: Near-infrared imaging photoplethysmography during driving. IEEE transactions on intelligent transportation systems 23(4), 3589-3600 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "28. Pilz, C.S., Zaunseder, S., Krajewski, J., Blazek, V.: Local group invariance for heart rate estimation from face videos in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1254-1262 (2018)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Poh, M.Z., McDuff, D.J., Picard, R.W.: Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE transactions on biomedical engineering 58(1), 7-11 (2010)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "30. Savic, M., Zhao, G.: De-identification of facial videos while preserving remote physiological utility. In: 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "31. Savic, M., Zhao, G.: Physu-net: Long temporal context transformer for rppg with self-supervised pre-training (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "text", + "content": "32. Savic, M., Zhao, G.: Rs-rppg: Robust self-supervised learning for rppg. In: 2024 18th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2024). IEEE (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "33. Speth, J., Vance, N., Flynn, P., Bowyer, K., Czajka, A.: Remote pulse estimation in the presence of face masks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2086-2095 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "34. Speth, J., Vance, N., Flynn, P., Bowyer, K.W., Czajka, A.: Digital and physical-world attacks on remote pulse detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2407-2416 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "35. Speth, J., Vance, N., Flynn, P., Czajka, A.: Non-contrastive unsupervised learning of physiological signals from video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14464-14474 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "36. Speth, J., Vance, N., Sporrer, B., Niu, L., Flynn, P., Czajka, A.: Mspm: A multi-site physiological monitoring dataset for remote pulse, respiration, and blood pressure estimation. arXiv preprint arXiv:2402.02224 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "37. Špetlík, R., Franc, V., Matas, J.: Visual heart rate estimation with convolutional neural network. In: Proceedings of the British machine vision conference, Newcastle, UK. pp. 3-6 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "38. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062. IEEE (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "39. Stricker, R., Müller, S., Gross, H.M.: Non-contact video-based pulse rate measurement on a mobile service robot. In: The 23rd IEEE International Symposium on Robot and Human Interactive Communication. pp. 1056-1062 (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "text", + "content": "40. Sun, Z., Li, X.: Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XII. pp. 492–510. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "41. Sun, Z., Li, X.: Privacy-phys: Facial video-based physiological modification for privacy protection. IEEE Signal Processing Letters 29, 1507-1511 (2022). https://doi.org/10.1109/LSP.2022.3185964" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "text", + "content": "42. Tang, J., Chen, K., Wang, Y., Shi, Y., Patel, S., McDuff, D., Liu, X.: Mmpd: Multi-domain mobile video physiology dataset. arXiv preprint arXiv:2302.03840 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "43. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Optics express 16(26), 21434-21445 (2008)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "44. Wang, W., den Brinker, A.C., Stuijk, S., De Haan, G.: Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering 64(7), 1479-1491 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "45. Wang, W., Shan, C.: Impact of makeup on remote-ppg monitoring. Biomedical Physics & Engineering Express 6(3), 035004 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "ORPDAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 326 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "46. Yu, Z., Li, X., Zhao, G.: Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In: British Machine Vision Conference (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "47. Yu, Z., Peng, W., Li, X., Hong, X., Zhao, G.: Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 151-160 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "48. Yu, Z., Shen, Y., Shi, J., Zhao, H., Cui, Y., Zhang, J., Torr, P., Zhao, G.: Physformer++: Facial video-based physiological measurement with slowfast temporal difference transformer. International Journal of Computer Vision 131(6), 1307-1330 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "text", + "content": "49. Yu, Z., Shen, Y., Shi, J., Zhao, H., Torr, P.H., Zhao, G.: Physformer: facial video-based physiological measurement with temporal difference transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4186-4196 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "text", + "content": "50. Zhang, Z., Girard, J.M., Wu, Y., Zhang, X., Liu, P., Ciftci, U., Canavan, S., Reale, M., Horowitz, A., Yang, H., et al.: Multimodal spontaneous emotion corpus for human behavior analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3438-3446 (2016)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "M. Savic and G. Zhao" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_content_list.json b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd177053b2501c1b6073c9f4175c2c9315d1efc6 --- /dev/null +++ b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_content_list.json @@ -0,0 +1,2278 @@ +[ + { + "type": "text", + "text": "Out-of-Bounding-Box Triggers: A Stealthy Approach to Cheat Object Detectors", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Lin $^{1,2}$ , Lijia Yu $^{1}$ , Gaojie Jin $^{1,2}$ , Renjue Li $^{1,2}$ , Peng Wu $^{1,2,\\text{图}}$ , and Lijun Zhang $^{1,2,3,\\text{图}}$", + "bbox": [ + 230, + 218, + 771, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China", + "bbox": [ + 215, + 268, + 785, + 311 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{2}$ University of Chinese Academy of Sciences, Beijing, China", + "3 Automotive Software Innovation Center, Chongqing, China {lintao,yulijia,gaojie,lirj19,wp,zhanglj}@ios.ac.cn" + ], + "bbox": [ + 295, + 311, + 704, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In recent years, the study of adversarial robustness in object detection systems, particularly those based on deep neural networks (DNNs), has become a pivotal area of research. Traditional physical attacks targeting object detectors, such as adversarial patches and texture manipulations, directly manipulate the surface of the object. While these methods are effective, their overt manipulation of objects may draw attention in real-world applications. To address this, this paper introduces a more subtle approach: an inconspicuous adversarial trigger that operates outside the bounding boxes, rendering the object undetectable to the model. We further enhance this approach by proposing the Feature Guidance (FG) technique and the Universal AutoPGD (UAPGD) optimization strategy for crafting high-quality triggers. The effectiveness of our method is validated through extensive empirical testing, demonstrating its high performance in both digital and physical environments. The code and video will be available at: https://github.com/linToTao/Out-of-bbox-attack.", + "bbox": [ + 259, + 386, + 743, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Universal physical adversarial attack $\\cdot$ Out-of-bounding-box trigger $\\cdot$ Object detection", + "bbox": [ + 259, + 621, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection aims to identify instances of semantic objects typically within images or video clips. It finds widespread applications across a range of domains, including face recognition [10, 11, 23, 49, 58], object tracking [34, 35, 60], and autonomous driving [6, 18, 68]. Especially in autonomous driving systems, object detectors play a crucial role in tasks such as recognizing traffic signs, pedestrians, vehicles, and traffic lights. Nonetheless, in recent years, security concerns have emerged regarding object detectors, primarily due to the susceptibility of deep neural networks (DNNs) to adversarial examples (AEs). AEs are intentionally crafted inputs designed to deceive DNNs into making incorrect predictions.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d27a9aa680fd619b1b3f75df85313255adc96627a4aca5e08809f66f56ca564a.jpg", + "image_caption": [ + "Fig. 1: An illustration of vanilla object attack methods and our novel approach. Typically, vanilla attack methods target the most informative regions of an object, such as the central, highly visible areas of the target object, e.g., the stop sign in (a) [16], (b) [69], (c) [15], and the car in (d) [19], (e) [59], (f) [22]. In contrast, our method adopts a more covert strategy, focusing on attacking the peripheral edge areas surrounding the stop sign." + ], + "image_footnote": [], + "bbox": [ + 232, + 142, + 772, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early research mainly focused on studying AEs like adversarial patch [16, 20, 33, 40, 73], optical [43, 70] and texture [15, 21], which are targeted at the objects. These AEs involve the computation of perturbations applied directly to the objects to mislead a DNN-based detector. However, a direct attack on the object as shown in Figure 1, along with its subsequent attempt to deceive the detector, is overly conspicuous and cannot be deployed without raising suspicion in real-world practical scenarios.", + "bbox": [ + 212, + 434, + 787, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we provide an alternative perspective: is it possible to introduce an adversarial trigger in an inconspicuous area outside the object while still successfully deceiving the detector?", + "bbox": [ + 212, + 544, + 789, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Actually, crafting an adversarial trigger within the areas beyond the object boundaries is even more demanding. On one hand, for dense prediction networks like object detectors, the output results are more reliant on the red regions of the heatmap in Figure 1. This implies that tailoring the adversarial trigger by utilizing irrelevant information, like the blue regions of the heatmap, is sharply challenging. More difficultly, our trigger is designed to conform to practical size and shape constraints. Nevertheless, for exactly the same reasons, our attack method gets more delusive and harder to defense.", + "bbox": [ + 212, + 594, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper is focused on the development of robust and inconspicuous triggers to attack typical object detectors deployed in practical scenarios. To tackle the challenge mentioned above, we present a novel technique called Feature Guidance (FG). In this approach, we guide the feature layers with an adversarial trigger to resemble those with the object hidden. And we propose a optimization strategy, called Universal Auto-PGD (UAPGD), applicable to universal attacks. This optimization strategy allows for the selection of an appropriate timing to halve the step size, based on the oscillation of the loss function.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To assess the effectiveness of our approach in both digital and physical domains, we trained the adversarial trigger and evaluated the corresponding metrics on COCO [39] dataset and Carla [14], respectively. We also evaluated the robustness of the trigger by recording videos using an onboard camera in the real world.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, this paper makes the following contributions:", + "bbox": [ + 238, + 222, + 627, + 237 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- This work studies the realm of adversarial robustness in object detection, providing a novel perspective by exploring the concept of an adversarial trigger applied outside an object.", + "- We propose a novel attack method designed to generate triggers that achieve excellent continuous attack effects and demonstrate high robustness in real world. The key enabler of our method is the utilization of similarity between feature layers and adaptive step size decay to endow the trigger with the ability to mislead the detector.", + "- An extensive set of empirical results are provided to demonstrate that our attack exhibits strong performance and robustness in both the COCO dataset and Carla simulator, as well as in real-world scenarios." + ], + "bbox": [ + 225, + 244, + 784, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 431, + 382, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial examples (AEs), first discovered by Szegedy et al. [53], involve carefully crafted inputs inducing model misclassifications. Transferability [26, 61, 64, 66] of AEs, allows them to deceive networks with different hyperparameters or architectures.", + "bbox": [ + 212, + 460, + 787, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Digital adversarial examples are generated by directly manipulating any pixels in digital images prior to feeding them into the model. Adversarial attacks can be classified into black-box attacks and white-box attacks based on whether the internal structure and parameters of the target model are known. White-box attacks [5, 12, 17, 27, 28, 31, 44, 46] primarily rely on the computation of model gradients. Black-box methods, on the other hand, depend on querying the unknown model by observing the prediction [12, 13, 38, 47, 62–64, 71] or leveraging an agent model to imitate the threaten model [3, 7, 8, 24, 30, 67]. Recently, Croce et al. [9] proposed AutoAttack, a parameter-free method that combines novel white-box and black-box attack strategies. The works of applying digital adversarial examples to object detectors have been explored in [25] and [50] that introduce perturbations to input images to mislead detectors. [52] generate more subtle perturbations via reinforcement learning, successfully attacking the detector.", + "bbox": [ + 212, + 522, + 787, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physical adversarial examples are crafted to be printed out and captured by sensory equipment like camera [31,36] and LiDAR [4,41,55]. These examples pose a real-world threat due to their ability to adapt to camera and scenario variations. Early studies focused on attacking recognition models [16,51]. Recent research has explored physical attacks in the more challenging context of object detection. Perturbations in this domain include adversarial patches [1,20,54,65, 69], adversarial camouflage textures [21,22,56,59], and optical attacks [37,43,57, 72]. To improve robustness against various transformations in physical settings,", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Athalye et al. [2] introduced Expectation Over Transformation (EOT), widely used in physical attacks, to enhance the effectiveness of adversarial attacks.", + "bbox": [ + 215, + 145, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior physical attack methods need to be applied directly to the objects so that these methods are expected to yield favorable results but they are prone to raising suspicion in practical scenarios. Unlike these works, our method constructs a universal attack trigger that can deceive the detector disregarding the specific object without altering the appearance of itself, and can be discreetly deployed in the real world.", + "bbox": [ + 212, + 176, + 785, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The works that exhibit the closest alignment with our motivation are Dpatch [42] and [32] (after this referred to as Dpatch2). However, we must emphasize that our work is independent of them, highlighting the difference between our work and theirs in Appendix F. Furthermore, in the experiments section, we compare the attack effectiveness of our method with Dpatch2 [32] on the COCO dataset.", + "bbox": [ + 212, + 267, + 787, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 381, + 380, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we aim to generate physical adversarial triggers that can be naturally deployed in an inconsequential position. These adversarial triggers are designed not to destroy the object information but still deceive the model effectively. In this section, we commence by presenting the problem formulation in Section 3.1. Subsequently, Fig. 2a provides a comprehensive illustration of our framework. Our primary focus is on the white-box attack paradigm, which necessitates an intricate optimization process. Given an object detection model, we design a loss function to determine the gradient direction for adversarial attacks in Section 3.2. This loss function is then utilized iteratively to update the adversarial trigger. In the pursuit of strengthening attack performance and enhancing the continuity of attacks in video-based scenarios, we introduce a novel method named \"Feature Guidance\" (FG) and an additional loss term $L_{FG}$ in Section 3.2. To bolster the overall robustness and deployability of our approach, we also employ two indispensable techniques, EOT and total variation loss in Section 3.2. Furthermore, our methodology draws inspiration from the work of [9] in their Auto PGD framework during the training process, we thus modify the conditions of step size decay based on the universal training mode (Section 3.3).", + "bbox": [ + 212, + 412, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 693, + 439, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial attacks on object detection can have diverse objectives, encompassing but not limited to inducing misalignment in bounding boxes, reducing classification confidence, or provoking misclassification. Our goal is to generate adversarial examples for a specific class, with the aim of causing the object detection algorithm to fail in identifying objects belonging to that class.", + "bbox": [ + 212, + 719, + 785, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input image $x \\in \\mathbb{R}^d$ with the ground truth $y = \\{(y_i^{\\mathrm{coor}}, y_i^{\\mathrm{cls}})\\}_{i \\in [n]} \\subset \\mathbb{R}_+^4 \\times \\mathbb{Z}_+$ , $n$ is the number of objects in the image, $y_i^{\\mathrm{coor}}$ contains top-left and bottom-right coordinates of the bounding box in the image, and $y_i^{\\mathrm{cls}}$ is the class of", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6a78b26b4e945d1d75311757f6bbaed5ebcfc963d7903bd4d7028b09970f320.jpg", + "image_caption": [ + "Fig.2: (a) shows the overview of our adversarial trigger generation framework. We deploy the adversarial trigger onto each image in the dataset with robustness transformation. With a fixed detector, our method optimizes the adversarial trigger through the obtained adversarial gradient. (b) illustrates the motivation and calculation method of Feature Guidance. Taking a stop sign as an example, for a given image $x$ , minimizing $\\| f_z(T(x,\\mathcal{A}_t)) - f_z(x')\\|_2$ can mislead a detector to classify the stop sign image $T(x,\\mathcal{A}_t)$ as none stop." + ], + "image_footnote": [], + "bbox": [ + 218, + 141, + 500, + 276 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/040c7978640aa6ffc12a9b83a1f8b8b143351261c8a3c6cb9427c04db7f9f4b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 172, + 767, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "object. Then, the DNN-based object detector can be represented as the following form:", + "bbox": [ + 212, + 412, + 784, + 440 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf: \\mathbb {R} ^ {d} \\to 2 ^ {\\mathbb {R} _ {+} ^ {4} \\times \\mathbb {Z} _ {+} \\times [ 0, 1 ] ^ {2}}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 440, + 583, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a given image $x$ , the output of $f$ can be represented as $\\tilde{y} = \\{(\\tilde{y}_i^{\\mathrm{coor}},\\tilde{y}_i^{\\mathrm{cls}},\\tilde{y}_i^{\\mathrm{conf}})\\}_{i\\in [n]} \\subset \\mathbb{R}_+^4\\times \\mathbb{Z}_+ \\times [0,1]^2$ , where $\\tilde{y}_i^{\\mathrm{conf}}$ denotes the confidence associated with the predicted box coordinates $\\tilde{y}_i^{\\mathrm{coor}}$ and the predicted class $\\tilde{y}_i^{\\mathrm{cls}}$ for $i$ -th object. We aim to generate a universal adversarial trigger $\\mathcal{A}_t\\in [0,1]^{3\\times U\\times V}$ , where $U,V\\in \\mathbb{Z}_{+}$ are the given dimensions, for the given specific class $t$ , such that these adversarial triggers can cause $f$ to fail in recognizing objects of class $t$ . Then, the problem can be formulated as an optimization task, represented as follows:", + "bbox": [ + 212, + 465, + 787, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a r g m i n} _ {\\mathcal {A} _ {t}} \\mathbb {E} _ {x \\sim D} [ | f (T (x, \\mathcal {A} _ {t})) \\cap (\\mathbb {R} _ {+} ^ {4} \\times \\{t \\} \\times [ 0, 1 ] ^ {2}) | ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 588, + 785, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|A|$ denotes the cardinality of the set $A$ , $D$ represents the input distribution, and $T$ (which will be elaborated in the following) is a transform operation to incorporate the adversarial trigger into the original image. In contrast to prior studies based on adversarial patches, our approach avoids placing the adversarial trigger on the informative surface, e.g., traffic signs, thereby preventing any unnecessary attention. We would like to point out that this kind of attack method is challenging because the adversarial trigger can only be stuck on the unimportant places of the image.", + "bbox": [ + 212, + 612, + 787, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transform operation. We utilize a differentiable operation $T$ to render the adversarial trigger $\\mathcal{A}_t$ onto an image $x$ . Let the object with class $t$ be located in the box $B$ of the image, we use the following two-step to add the adversarial trigger to the image:", + "bbox": [ + 212, + 734, + 787, + 794 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We can select an appropriate area $C(B)$ in the image where $C$ is a given function, such position does not overlap with the surface of the object;", + "(2) We cover the box $C(B)$ part of the image $x$ with adversarial trigger $\\mathcal{A}_t$ , but" + ], + "bbox": [ + 214, + 794, + 782, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering that the adversarial trigger and the box are not the same size, we need to apply the following affine matrix to $\\mathcal{A}_t$ :", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\left( \\begin{array}{c c c} S _ {h} \\cos \\alpha & - S _ {h} \\sin \\alpha & T _ {h} \\\\ S _ {v} \\sin \\alpha & S _ {v} \\cos \\alpha & T _ {v} \\\\ 0 & 0 & 1 \\end{array} \\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 188, + 785, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$S_{h}$ and $S_{v}$ denotes the scaling factors utilized to resize the adversarial trigger, $\\alpha$ represents the angle of rotation, $T_{h}$ and $T_{v}$ are the distance adversarial trigger need to shift. These parameters are all computed by $C(B)$ . The more specific remark regarding $C(B)$ will be detailed in Appendix B.", + "bbox": [ + 212, + 244, + 782, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that our attack method has two characteristics: Our attack is out of the object; our attack area changes depending on the position of the object. We were the first to attack target detection with such a method.", + "bbox": [ + 212, + 306, + 784, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Adversarial gradient with Feature Guidance", + "text_level": 1, + "bbox": [ + 214, + 373, + 624, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we focus on utilizing white-box attack methods to search for adversarial gradients and optimize the adversarial trigger. Additionally, we provide black-box empirical results in Appendix D that the trigger is generated by a surrogate model. As illustrated above, we render the adversarial trigger onto images and feed them to the detectors, take the derivative of the loss function, and use it to update the adversarial trigger. The loss function consists of the following parts:", + "bbox": [ + 212, + 398, + 782, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adversarial loss of detector output. As mentioned above, the detector outputs the predicted class and bounding boxes and their confidence. Considering most of the detectors will not output the predicted class and boxes with a confidence lower than a threshold, we want images with the adversarial trigger to make the confidence as low as possible.", + "bbox": [ + 212, + 505, + 782, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $f_{i,coor}^{\\mathrm{conf}}(x)$ and $f_{i,cls}^{\\mathrm{conf}}(x)$ is the confidence of the $i$ -th object on image $x$ , we can define the first part of our loss function as follows:", + "bbox": [ + 212, + 579, + 782, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d e t} \\left(\\mathcal {A} _ {t}\\right) = \\mathbb {E} _ {x \\sim D} \\left[ \\max _ {i: \\tilde {y} _ {i} ^ {c l s} = t} f _ {i, c o o r} ^ {\\text {c o n f}} (T (x, \\mathcal {A} _ {t})) f _ {i, c l s} ^ {\\text {c o n f}} (T (x, \\mathcal {A} _ {t})) \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 621, + 784, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature guidance. The biggest difficulty with an attack outside the object is that the features of the original object on the image will not be destroyed, so the network will still notice the presence of the object.", + "bbox": [ + 212, + 657, + 782, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We gain the following insights from the network structure to address this issue. A DNN-based detector is usually designed with three modules: Backbone, Neck, and Head. The Backbone module extracts global features from an input image. These features are then fused through the Neck module to produce multiple feature maps with varying receptive fields. The Head module subsequently makes predictions based on the feature maps. This modular structure imitates how people often perceive a comprehensive horizon (like the Backbone module) and focus on varying attractive areas (like the Neck module). Subsequently, the attractive areas will contribute to making perceptions (like the Head module).", + "bbox": [ + 212, + 704, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, in order to provoke model errors, it is imperative to deliberately induce information misleading prior to inputting the head module.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To achieve that, we use the following method, called Feature guidance, as shown in Fig. 2b. For a given image $x$ with a target object $t$ , we can get the image $x'$ , which differs from $x$ only with the object $t$ missing. Then, if we can elaborate our adversarial trigger $\\mathcal{A}_t$ to guide the feature map $f_z(T(x, \\mathcal{A}_t))$ as close to $f_z(x')$ generated by the Neck module of the detector, it will have a desirable impact on the Head module such that the stop sign cannot be detected.", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Figure 2a, for the sake of convenience, we derive $x_{mask}$ that exhibits an analogous effect to $x'$ simply by covering the bounding box position in the ground truth of $x$ with a gray rectangle.", + "bbox": [ + 212, + 268, + 784, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to achieve Feature guidance, we need to make $\\mathcal{A}_t$ to minimum such loss function:", + "bbox": [ + 212, + 313, + 782, + 340 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {F G} (\\mathcal {A} _ {t}) = E _ {x \\sim D} [ \\| f _ {z} (T (x, \\mathcal {A} _ {t})) - f _ {z} (x _ {\\text {m a s k}}) \\| _ {2} ]\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 354, + 671, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Other skills To enhance the robustness of our adversarial trigger $\\mathcal{A}_t$ , we incorporate diverse variations, including random noise, contrast, brightness, and rotation, into the generated trigger prior to its application on images, thus simulating real-world conditions. To address printing challenges, we employ TV loss [45] on the adversarial trigger to promote smoothness. TV loss is defined as", + "bbox": [ + 212, + 383, + 784, + 460 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t v} \\left(\\mathcal {A} _ {t}\\right) = \\sum_ {i, j} \\sqrt {\\left(\\mathcal {A} _ {t} ^ {i , j} - \\mathcal {A} _ {t} ^ {i + 1 , j}\\right) ^ {2} + \\left(\\mathcal {A} _ {t} ^ {i , j} - \\mathcal {A} _ {t} ^ {i , j + 1}\\right) ^ {2}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 315, + 472, + 784, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the subindices $i$ and $j$ represent the pixel coordinate of the trigger $\\mathcal{A}_t$", + "bbox": [ + 212, + 513, + 764, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, we will get the adversarial trigger $\\mathcal{A}_t$ by optimizing following problem:", + "bbox": [ + 212, + 529, + 782, + 554 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {A} _ {t} \\in [ 0, 1 ] ^ {3 \\times U \\times V}} L _ {d e t} (\\mathcal {A} _ {t}) + \\lambda_ {F G} L _ {F G} (\\mathcal {A} _ {t}) + \\lambda_ {t v} L _ {t v} (\\mathcal {A} _ {t}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 556, + 784, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{FG}$ and $\\lambda_{tv}$ are hyperparameter.", + "bbox": [ + 218, + 583, + 511, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Universal Auto-PGD", + "text_level": 1, + "bbox": [ + 214, + 621, + 434, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let $L_{all} = L_{det} + \\lambda_{FG}L_{FG} + \\lambda_{tv}L_{tv}$ , we will use gradient descent to find the adversarial trigger via minimizing the loss function $L_{all}$ .", + "bbox": [ + 212, + 647, + 782, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering the instability that appeared in the experiment when optimizing multiple terms of $L_{all}$ with conventional methods, we propose the following optimization strategy called Universal Auto-PGD (UAPGD):", + "bbox": [ + 212, + 678, + 784, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At the beginning of optimizing, we fix the following values: comparison length $l_{c}$ , observation length $l_{o}$ , the slack variable $\\epsilon_{1}, \\epsilon_{2}$ , let the initial step size be $\\eta$ and we record the following values:", + "bbox": [ + 212, + 723, + 782, + 767 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1): After epoch $j$ , record the value of the loss function, name it $L_{j}$ ;", + "(2): After epoch $i * l_{o}$ where $i \\in Z_{+}$ , record:" + ], + "bbox": [ + 214, + 768, + 705, + 797 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\min , i} = \\min _ {j \\in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \\left\\{L _ {j} \\right\\} \\\\ V _ {i} = \\operatorname {V a r i a n c e} \\left[ \\left\\{L _ {j} \\right\\} _ {j \\in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 810, + 643, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: UAPGD" + ], + "code_body": "Input: Threaten network $f$ , initial step size $\\eta$ , the number of epoch $N_{\\mathrm{epoch}}$ dataset $D = \\cup_{j = 1}^{k}D_{j}$ which was divided into k batches, the hyperparameter in UAPGD $l_c,l_o$ , initial adversarial trigger $\\mathcal{A}_t^0$ Output: Adversarial trigger $\\mathcal{A}_{best}$ \n1 $S_{c}\\gets \\phi$ $S_{o}\\gets \\phi$ $L_{best}\\gets \\infty$ $\\mathcal{A}_{best}\\leftarrow \\mathcal{A}_t^0$ \n2 for $i\\in [N_{\\mathrm{epoch}}]$ do \n3 $L_{i}\\gets 0$ \n4 $\\mathcal{A}_{t}^{i,0}\\gets \\mathcal{A}_{t}^{i - 1}$ \n5 for $j\\in [k]$ do \n6 Get $L_{all}(\\mathcal{A}_t^{i,j - 1},D_j)$ via $f$ \n7 $L_{i}\\gets L_{i} + L_{all}(\\mathcal{A}_{t}^{i,j - 1},D_{j}) / k$ \n8 $\\mathcal{A}_{t}^{i,j}\\gets \\mathcal{A}_{t}^{i,j - 1} - \\eta \\cdot \\mathrm{sign}(\\frac{\\nabla L_{all}(\\mathcal{A}_{t}^{i,j - 1},D_{j})}{\\nabla\\mathcal{A}_{t}^{i,j - 1}})$ \n9 $\\mathcal{A}_{t}^{i,j}\\gets Clamp(\\mathcal{A}_{t}^{i,j},0,1)$ \n10 $\\mathcal{A}_{t}^{i}\\gets \\mathcal{A}_{t}^{i,k}$ \n11 if $L_{i} < L_{best}$ then \n12 $\\begin{array}{r}\\big{\\lfloor}\\mathcal{A}_{best}\\leftarrow \\mathcal{A}_{t}^{i}\\text{and} L_{best}\\leftarrow L_{i} \\end{array}$ \n13 $S_{o} = S_{o}\\cup L_{i}$ \n14 if $|S_{o}| = l_{o}$ then \n15 $\\begin{array}{rl} & L_{min}\\gets \\min_{L\\in S_o\\backslash \\{S_o\\}}\\\\ & V\\gets \\operatorname{Var}_{L\\in S_o\\backslash \\{S_o\\}}\\\\ & S_c = S_c\\cup (L_{min},V)\\\\ & S_o\\gets \\phi \\end{array}$ \n16 \n17 \n18 \n19 if $|S_{c}| = l_{c}$ then // Condition 1 and Condition 2 are introduced in section 3.3 if Condition 1(Sc) and Condition 2(Sc) then \n21 $\\begin{array}{r}\\lfloor \\eta \\leftarrow \\eta /2\\text{and}\\mathcal{A}_t^i\\leftarrow \\mathcal{A}_{best}\\\\ S_c\\leftarrow \\phi \\end{array}$ \n22", + "bbox": [ + 217, + 166, + 785, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to prevent violent oscillations around the local optimum during the optimization process, we will halve the step size when the following conditions are met:", + "bbox": [ + 212, + 654, + 787, + 696 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(Con 1): $L_{min,j} \\geq L_{min,q} - \\epsilon_1$ for all $(j - l_c) < q < j$ ;", + "(Con 2): $V_{j}\\geq V_{q} - \\epsilon_{2}$ for all $j - l_{c} < q < j$" + ], + "bbox": [ + 215, + 698, + 602, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the stochasticity from the robustness transformation mentioned in Other skills, the value of the loss function may fluctuate so that we incorporate the slack variable $\\epsilon_{1},\\epsilon_{2}$ into numerical comparison.", + "bbox": [ + 212, + 729, + 787, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, these two conditions respectively describe the exploration and exploitation aspects in the optimization process. Con 1 means that the value of the loss function has not decreased much, indicating the necessity of a transition from exploring the entire feasible space to focusing on local optimization. Con", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2 characterizes the oscillation of the loss function, of which the amplitude per $l_{c}$ steps has not decreased much. This implies the imminent conclusion of the algorithm's exploitation in a local space. The simultaneous occurrence of c1 and c2 means that the adversarial trigger jumps near one of the optimal points, but the step size is too large to fall steadily near the best advantages. Therefore, at this point, we will halve the step size.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our approach is somewhat similar to AutoAttack [9], but there are several differences: (1): Our method is for universal attack, AutoAttack is for attacking a single image; (2): We consider the amplitude of the oscillation during the optimization process and use this as an indicator of decline; (3): We include the slack variables in order to mitigate potential misleading phenomena caused by experimental randomness. For example, the oscillation of the loss function often easily cause $L_{j}$ to show a downward trend.", + "bbox": [ + 212, + 237, + 787, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall process of our algorithm is shown in algorithm 1. And we use $L_{all}(\\mathcal{A}_t, D_j)$ to represent calculating $L_{all}$ on an input batch $D_j$ .", + "bbox": [ + 214, + 342, + 785, + 375 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 396, + 364, + 414 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we choose the stop sign as the target object to evaluate the effectiveness of our adversarial attack method. The stop sign is a commonly used object in the evaluation of physical adversarial attacks, as it is critical to traffic safety. We begin by detailing the implementation settings of our evaluation. This is followed by an extensive set of experiments, utilizing the proposed adversarial trigger in the digital space, specifically on the COCO dataset [39]. Furthermore, we extend the application of our method to real-world scenarios, i.e., we record videos to demonstrate that our method can attack the object detector in both simulators and real-world scenes. Meanwhile, we evaluate our attack method across varying distances and angles. See Appendix A and B for the basic experimental setting, the results of additional attacks on other objects, and the detail of selecting C(B) to place the trigger on the stop sign.", + "bbox": [ + 212, + 429, + 787, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Digital experiment", + "text_level": 1, + "bbox": [ + 214, + 633, + 418, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this subsection, our experiments are carried out using the COCO dataset [39]. From this dataset, we extract a total of 1100 images featuring stop signs. Of these, 600 images constitute the training set, which is utilized to generate the adversarial trigger. The remaining 500 images are designated as the test set, serving to evaluate the effectiveness of the generated adversarial trigger.", + "bbox": [ + 212, + 657, + 787, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To demonstrate the superiority of our algorithm, we adopt the following methods to generate adversarial triggers: optimizing the adversarial trigger using PGD [44] or UAPGD with a minimization of the loss function $L_{det} + \\lambda_{tv}L_{tv}$ or $L_{all}$ . This included comparisons between our method and the normal PGD under the condition of placing the perturbation outside the object, as well as ablation experiments with varying hyperparameters. Specifically, for the YOLOv3(YOLOv5) model, we choose $\\lambda_{FG}$ values of 0.1, 0.5, and 1.0(0.05,", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "0.1, and 0.2). The optimization process utilized the initial step size of 8/255 and 16/255. We calculate the number of undetected images among all the images as a metric called attack success rate(ASR).", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/387eee5861047cbcf18a2e8f69887e106531443e69063696bef5bfe074c2f98e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Threaten ModelYOLOv3YOLOv5
DatasetTrain(%)Test(%)Train(%)Test(%)
Dpatch226.6724.007.176.67
Dpatch2 (Ldet)35.5335.8032.8332.00
Adam (16/255)65.1662.2043.8340.40
Adam (8/255)64.3360.8052.1650.80
PGD(16/255)56.6754.0043.6544.60
PGD(8/255)65.1762.2053.8351.00
UAPGD(16/255)66.0063.0057.8352.60
UAPGD(8/255)67.8364.8054.1750.40
FG(λF, FG, 16/255)57.6755.4051.8348.20
FG(λF, FG, 16/255)56.0054.6047.8344.60
FG(λF, FG, 16/255)54.0051.0043.8340.00
FG(λF, FG, 8/255)60.0055.2059.1755.60
FG(λF, FG, 8/255)64.8362.6052.1750.60
FG(λF, FG, 8/255)61.6759.8056.6756.40
Ensemble(λF, FG, 16/255)67.8366.2058.3356.40
Ensemble(λF, FG, 16/255)67.8364.8060.1757.60
Ensemble(λF, FG, 16/255)64.8361.6058.0056.60
Ensemble(λF, FG, 8/255)68.3365.0062.6758.60
Ensemble(λF, FG, 8/255)69.3366.6056.6753.20
Ensemble(λF, FG, 8/255)66.1662.2060.1757.80
", + "bbox": [ + 238, + 224, + 524, + 412 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e500f78b3fb3b0e15738097e9d4b2a293b6a94619633331edbc4f41b2060a1fb.jpg", + "image_caption": [ + "PGD" + ], + "image_footnote": [], + "bbox": [ + 542, + 224, + 635, + 316 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/eff3da90302496f2f5287b95ee4853e25e7c8e65d25fed96b2d7ca11422f39c0.jpg", + "image_caption": [ + "UAPGD" + ], + "image_footnote": [], + "bbox": [ + 656, + 226, + 748, + 316 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1183652b95f0fa74a19b6e9a11fd4eff6384308eb44440906d6254e24f780695.jpg", + "image_caption": [ + "FG", + "Fig. 3: The table on the left displays the ASR values of triggers, which are generated by various settings, on the training and testing sets of the COCO dataset. The values in parentheses represent the attack step size, and we employ different $\\lambda_{FG}^{i}$ values for YOLOv3 and YOLOv5. For YOLOv3, we set $\\lambda_{FG}^{1} = 0.1$ , $\\lambda_{FG}^{2} = 0.5$ , $\\lambda_{FG}^{3} = 1.0$ ; for YOLOv5, we set $\\lambda_{FG}^{1} = 0.05$ , $\\lambda_{FG}^{2} = 0.1$ , $\\lambda_{FG}^{3} = 0.2$ . Ensemble denotes the combined usage of FG and UAPGD. Dpatch2 is the original approach from [32] and Dpatch2( $L_{det}$ ) denotes the original approach adopts the $l_{det}$ as the optimization objective. The bolded values indicate the optimal results for each respective column in the table. There are some results of triggers on the right generated under different settings." + ], + "image_footnote": [], + "bbox": [ + 540, + 328, + 635, + 419 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/72a324c1c3af2b29d9de062c92e1853f662fd94512d7fdfcf55c36d481b79427.jpg", + "image_caption": [ + "Ensemble" + ], + "image_footnote": [], + "bbox": [ + 656, + 328, + 748, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Figure 3, We compare our method with the more reasonable and advanced Dpatch2 [32]. Considering the divergent optimization objectives, we substitute the maximization of the model loss function mentioned in [32] with the minimization of the proposed $L_{det}$ referred to as Dpatch2( $L_{det}$ ). Undoubtedly, our method achieves the highest ASR in different experimental settings.", + "bbox": [ + 212, + 606, + 787, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A comparative analysis of the effectiveness between PGD and UAPGD reveals notable differences. Specifically, we observe that UAPGD, when employed with an identical attack step size as PGD, frequently results in a higher ASR. Moreover, it can be observed that UAPGD yields better performance compared to adaptive step size optimizers like Adam.", + "bbox": [ + 212, + 685, + 787, + 760 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 4a and Figure 4b present a detailed depiction of the variations in the detection loss $L_{det}$ throughout the training process when generating adversarial triggers using both PGD and UAPGD techniques, specifically over YOLOv3 and YOLOv5 models. Notably, when operating under an equivalent number of epochs, UAPGD demonstrates a more pronounced capability to reduce $L_{det}$", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/216198102cf21938cf28389a83992fa67fac2aa58b3dce6a9b8db735f7b00ce6.jpg", + "image_caption": [ + "(a) YOLOv3" + ], + "image_footnote": [], + "bbox": [ + 246, + 150, + 359, + 215 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c08f344bdf0c70c5ec8686666c02d292668611c99437150d75fd5b2f1305ec09.jpg", + "image_caption": [ + "(b) YOLOv5", + "Fig. 4: (a) and (b) illustrate the variations of $L_{det}$ , during the training process of generating adversarial triggers using PGD and UAPGD for attacking the YOLOv3 and YOLOv5 models. (c) and (d) present the confidence of some samples from the COCO dataset, outputted by the detector, during the training process of individual attacks. The blue, green, and red curved lines represent the utilization of $L_{FG}$ , $L_{det}$ , and both as the loss function respectively. For individual attacks, the optimization process is relatively easier so that step size decay is unnecessary." + ], + "image_footnote": [], + "bbox": [ + 375, + 150, + 488, + 215 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b3b0d2844550071cd4739d33a5a239e6198c81c6f3fcc7113b7a1a95a1bb235b.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 506, + 150, + 617, + 215 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fc29a5f10440eede6d74adbed270de542c47a7e05ec61406d34ee959305ee6e1.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 635, + 151, + 750, + 215 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "compared to PGD. The effectiveness of the loss monitoring in UAPGD will be analyzed in detail in Appendix E.", + "bbox": [ + 212, + 369, + 784, + 400 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Furthermore, as shown in Figure 3, UAPGD exhibits a superior capability in generating smoother adversarial triggers. This increased smoothness can be attributed to the implementation of step size decay, which allows for a more refined optimization process in the generation of triggers. Consequently, adversarial triggers crafted via UAPGD display enhanced suitability for practical applications, such as printing and deployment in the physical world.", + "bbox": [ + 212, + 401, + 784, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Based on results in Figure 3, we can observe that integrating $L_{FG}$ with suitable step size and weight into the loss function, coupled with using either PGD or UAPGD, can enhance the ASR. This improvement is particularly noticeable when applied to the YOLOv5 model. This observation highlights that employing appropriate feature guidance can improve the performance of universal attacks. However, incorporating FG did not always achieve satisfactory results under certain weights and step sizes. This phenomenon is because in universal attacks, the inclusion of $L_{FG}$ introduces additional complexities to the optimization process. It requires appropriate adjustment of hyperparameters to fully leverage its effectiveness. Hence, the optimal outcome can be achieved by ensemble UAPGD with it, which allows for finer control over the step size to facilitate the optimization process. The empirical results indicate that the combination of UAPGD and FG in generating triggers yields the most effective attack performance.", + "bbox": [ + 212, + 492, + 785, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To further investigate the potential role of Feature Guidance, we conducted individual adversarial attacks on each image with an adversarial trigger (fixed step size of 8/255, 500iteration) in the training dataset. We identified two primary roles of FG: 1, as shown in Figure 4c, our observations reveal that FG can significantly influence the optimization direction during the training phase. This influence is evidenced by the expedited convergence in adversarial trigger training, indicating the advantageous role of FG in enhancing training efficiency. 2, As shown in Figure 4d, when some images are challenging to attack with out-of-box adversarial triggers, introducing FG would enhance the attack effectiveness at this time.", + "bbox": [ + 212, + 688, + 785, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experiment in simulator", + "text_level": 1, + "bbox": [ + 215, + 146, + 464, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we demonstrate the effects of our attack in autonomous driving scenarios, we use the Carla(0.9.14) simulator [14] to simulate the real driving process. In the simulator, we utilize the YOLOv3 model as the detection network.", + "bbox": [ + 212, + 172, + 542, + 263 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Due to the significant differences between images from the COCO dataset and those generated by the Carla simulator, the adversarial trigger developed in Section 4.1 is not directly applicable to Carla. To tailor an adversarial trigger specifically for the Carla simulator, we compile a training dataset comprising 242 images featuring stop signs from the simulator. The process of attacking the Carla simulator", + "bbox": [ + 212, + 263, + 545, + 400 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "is conducted in four distinct manners: employing either PGD or UAPGD, combined with a loss function of either $L_{det} + \\lambda_{tv}L_{tv}$ or $L_{all}$ , where $\\lambda_{FG} = 0.1$ and the initial step size is 16/255.", + "bbox": [ + 212, + 400, + 785, + 445 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the attack, we implement a procedure where the adversarial trigger is placed under the stop sign in the simulator. Subsequently, a vehicle is navigated towards this stop sign, and a 9-second video is recorded using an onboard camera. The key metric is the percentage of time during which the detection system fails to recognize the stop sign.", + "bbox": [ + 212, + 446, + 544, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Table 1, without any adversarial trigger, the car consistently recognizes the stop sign. When the attack is initiated using the PGD method alone, the attack duration is approximately 3 to 4 seconds. However, the integration of either UAPGD or FG methods into the attack strategy results in a slight increase in the attack duration by roughly an additional second. Notably, the implementation of an ensemble approach, which combines", + "bbox": [ + 212, + 582, + 545, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "these methods, further enhances the attack performance, prolonging the duration to over 5 seconds.", + "bbox": [ + 212, + 733, + 784, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Moreover, to provide a more direct illustration of the attack impact, we have plotted the trend of the stop sign detection confidence as the car approaches it, shown as in Figure 5.", + "bbox": [ + 212, + 763, + 785, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The empirical results reveal that the efficacy of our attack method consistently surpasses that of the vanilla PGD approach. Particularly in the final 5", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5f2afb937c2977ca3cbc7dc45959a5ebf3204655db07389b357e21f189570558.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
attack methodProportion
No attack0%
PGD36.67%
UAPGD45.56%
PGD,FG46.67%
UAPGD,FG58.89%
", + "bbox": [ + 555, + 205, + 750, + 303 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1: The proportion of time in the video duration during which the detector failed to recognize the stop sign.", + "bbox": [ + 553, + 314, + 787, + 371 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1573d79a60744cdd7bd24480190fc083636faa2152ec8561a176d560f8381904.jpg", + "image_caption": [ + "Fig.5: We uniformly extracted approximately 90 frames from the 9s video, which is recorded by the simulated vehicle's onboard camera, to showcase the confidence variations of the detector's predictions" + ], + "image_footnote": [], + "bbox": [ + 576, + 489, + 754, + 595 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b8e45a5dd8cba39c41d276472d2bc1939fb261b8a1afc662ea6fdb21dbf8a847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 305, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/020d5516646b7171a2f025ba9ad8498d13860622eee940bafb812b165d50dc88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 143, + 401, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fb46028851858f4f27fb21816c751721a2a879e2c011fd88dc4a393528d58713.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 143, + 496, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/80a91c18a7098f6271510abfd1fa494a7ff388a763d416c0b350cc715b965be0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 143, + 589, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7ce192eb37439234d236016e572a1cd17a104cec4ad0de0ac13780167dd3ab21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 143, + 684, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d505e6ab38c8190e4c4f9fcfded3417de960d5ed23cfa3b2c09c09fc92656142.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 143, + 779, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/57c7373b99afc96b087ddd9f70348bdec3b104ad0b8ca2f8b3dda05b6a103985.jpg", + "image_caption": [ + "$conf = 0.98$" + ], + "image_footnote": [], + "bbox": [ + 222, + 220, + 305, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/97daaf357441bc729c5b8bdd4ff84d0737309f6b31b0be13e138b497536c93aa.jpg", + "image_caption": [ + "$conf = 0.98$" + ], + "image_footnote": [], + "bbox": [ + 316, + 220, + 401, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dce6c46738c2e26110444f0f82fc486cb2f788ea3d111490695d1b6ddf6e8e62.jpg", + "image_caption": [ + "$conf = 0.99$" + ], + "image_footnote": [], + "bbox": [ + 411, + 220, + 496, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/71b36cb6d44365d25874b895739252d1a378d99a409eca9e863f79db513acdee.jpg", + "image_caption": [ + "$conf = 1.0$" + ], + "image_footnote": [], + "bbox": [ + 506, + 220, + 589, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c26c9dd08335a303d179cbbe2630761be8f4ab3f48ad6b73f2b08276409c32fb.jpg", + "image_caption": [ + "$conf = 1.0$" + ], + "image_footnote": [], + "bbox": [ + 602, + 220, + 684, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dbfe6e49f9e35fdffe1203b773593c2b106ede580426fa9700678f2f185f6cad.jpg", + "image_caption": [ + "conf $= 1.0$" + ], + "image_footnote": [], + "bbox": [ + 697, + 220, + 779, + 286 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b29838f6ef7e6212ab2899bc44bab014f8952a5a327e3a5136d28f056e1fd0ce.jpg", + "image_caption": [ + "conf = 0.94", + "conf = 0.82", + "Fig. 6: Six frames are extracted uniformly from the video. Row 1 represents a clean stop sign with the detector's output. In Row 2 and Row 3, we deploy the trigger generated by the PGD and ensemble method respectively." + ], + "image_footnote": [], + "bbox": [ + 222, + 299, + 305, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3211e5fbf980c00e5dec0844fc08ee8a80e3bce9781c29159e934f983d16ffdd.jpg", + "image_caption": [ + "conf = 0.72", + "conf $< 0.5$" + ], + "image_footnote": [], + "bbox": [ + 316, + 299, + 401, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f2f1a8163f9747eb6f4e979e25ffde3526c132a9cda55c3dc569cb7deb73a46a.jpg", + "image_caption": [ + "conf = 0.70", + "conf $< 0.5$" + ], + "image_footnote": [], + "bbox": [ + 411, + 299, + 496, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a6cad8c464753a06f75fe998044cb74428ce366d2df841b1b724f1e07627e41c.jpg", + "image_caption": [ + "conf $< 0.5$", + "conf $< 0.5$" + ], + "image_footnote": [], + "bbox": [ + 506, + 299, + 589, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2fdc8d2b620c383bfc67247a2a0dc97fcde66944e285c05397d3375f7f78f700.jpg", + "image_caption": [ + "conf $< 0.5$", + "conf $< 0.5$" + ], + "image_footnote": [], + "bbox": [ + 602, + 299, + 684, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/30b02c088302f1994a6b46405fe11ca2f32ee3811b98ca22c95710c304108388.jpg", + "image_caption": [ + "$conf = 0.6$", + "conf $< 0.5$" + ], + "image_footnote": [], + "bbox": [ + 697, + 299, + 779, + 364 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "seconds of the video, during which the adversarial trigger persistently causes the detection system to lose sight of the stop sign. This effect illustrates the potential of our method to compromise the functionality of autonomous driving systems.", + "bbox": [ + 212, + 463, + 784, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in the Figure 6, we extract six frames uniformly from the video as examples to demonstrate the effects of this attack in driving scenarios.", + "bbox": [ + 212, + 508, + 784, + 540 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The observations imply that normally the car will always recognize the stop sign (Row 1 of Figure 6, each stop is marked by a red box). However, after the adversarial trigger created by our attack method (Row 3) is added to the stop sign, the car fails to recognize the stop sign(no red box). Compared with the attack effect of the adversarial trigger generated by PGD (Row 2), it can be seen that our method is clearly superior to the PGD. It is worth mentioning that the perturbation generated by Dpatch2 is suboptimal in the simulator, as demonstrated in Appendix F.", + "bbox": [ + 212, + 541, + 787, + 661 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Robustness of our attack in real world", + "text_level": 1, + "bbox": [ + 214, + 688, + 578, + 704 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To observe the attack effectiveness and robustness of the universal adversarial trigger obtained through algorithm 1 in the real world, we print out the adversarial trigger and post it under a stop sign. Moreover, in the case where clean images serve as a reference, we compare our method with the normal PGD.", + "bbox": [ + 212, + 718, + 784, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We examined the variation trends in the confidence of the detector's output for recognizing the stop sign using an onboard camera under different conditions: as the distance between the onboard camera and the stop sign decreased from far to near, and as the onboard camera's perspective towards the stop sign shifted", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dc06d519f6941854aa3b1382d908a39aac8d08daaa9d114fe402187acf191d5e.jpg", + "image_caption": [ + "Fig. 7: We print and deploy the adversarial trigger in the physical world. (a) illustrates the confidence variation of the stop sign as the distance increases. (b) depicts the confidence variation of the stop sign when facing different angles changing." + ], + "image_footnote": [], + "bbox": [ + 236, + 151, + 302, + 276 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5f29f0e1b609f41b75d9cd1e99a879bf61b186d69656e2d35881f4567dea5563.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 326, + 157, + 529, + 275 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4a4ffe0d3a4463fc4445b23c01801f7b5c8af1d023e8b6742f67de06f022499b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 558, + 157, + 759, + 275 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "from -70-degree side view to a 70-degree side view. The real-world experiment results are shown in Figure 7.", + "bbox": [ + 212, + 375, + 782, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "According to Figure 7a, due to the relatively long distance between the vehicle and the stop position in the first half of the curve, the trigger occupies a small proportion of the field of view, which diminishes the stability of successful attacks. In the second half of the curve, where the distance between the stop and the vehicle is closer, the attacks become more stable. Our method consistently maintains a higher success rate in this scenario. Examining Figure 7b, it can be observed that when facing the stop sign directly, our method misleads the detector to output a remarkably low confidence. As the angle approaches approximately $\\pm 50$ degrees, our method becomes ineffective. In contrast, the triggers generated by PGD are not smooth, which prevents them from effectively imprinting adversarial features into the real world.", + "bbox": [ + 212, + 405, + 787, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Overall, in the real world, our attack consistently exhibits robust and effective performance when the onboard camera is approaching a stop sign. This means that our attack can successfully mislead an autonomous driving system to disregard the stop sign unconsciously.", + "bbox": [ + 212, + 571, + 787, + 632 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 656, + 359, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present a novel adversarial attack methodology that extends beyond the bounding boxes of objects. This method poses greater challenges compared to previous studies due to its tendency to focus on less informative areas. Thus, we propose the FG techniques and UAPGD optimization strategy, specifically designed to augment the efficacy of this attack in both digital and physical realms. Furthermore, we validate the effectiveness of our attack method through its application in autonomous driving scenarios and real-world conditions. Discussion of limitations and potential negative impact of our works will be placed in Appendix G and immediately following that we also briefly address future work.", + "bbox": [ + 212, + 689, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank the anonymous referees for their valuable and insightful comments. This work is supported by the National Key R&D Program of China under Grant No. 2022YFA1005100, 2022YFA1005101, and 2022YFA1005104, the CAS Project for Young Scientists in Basic Research, under Grant No.YSBR-040, ISCAS New Cultivation Project ISCAS-PYFX-202201, ISCAS Basic Research ISCAS-JCZD-202302, and the Major Project of ISCAS under Grant No. ISCAS-ZD-202302.", + "bbox": [ + 212, + 174, + 787, + 280 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 316 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aich, A., Li, S., Song, C., Asif, M.S., Krishnamurthy, S.V., Roy-Chowdhury, A.K.: Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1308-1318 (2023) 3", + "2. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018) 4", + "3. Brendel, W., Rauber, J., Bethge, M.: Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248 (2017) 3", + "4. Cao, Y., Xiao, C., Cyr, B., Zhou, Y., Park, W., Rampazzi, S., Chen, Q.A., Fu, K., Mao, Z.M.: Adversarial sensor attack on lidar-based perception in autonomous driving. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 2267-2281 (2019) 3", + "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017) 3", + "6. Chen, C., Seff, A., Kornhauser, A., Xiao, J.: Deepdriving: Learning affordance for direct perception in autonomous driving. In: Proceedings of the IEEE international conference on computer vision. pp. 2722-2730 (2015) 1", + "7. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 IEEE symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020) 3", + "8. Chen, P.Y., Zhang, H., Sharma, Y., Yi, J., Hsieh, C.J.: Zoo: Zeroth order optimization based black-box attacks to deep neural networks without training substitute models. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 15-26 (2017) 3", + "9. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020) 3, 4, 9", + "0. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019) 1", + "1. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019) 1", + "2. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018) 3" + ], + "bbox": [ + 225, + 330, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019) 3", + "14. Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: Carla: An open urban driving simulator. In: Conference on robot learning. pp. 1-16. PMLR (2017) 3, 12", + "15. Duan, R., Ma, X., Wang, Y., Bailey, J., Qin, A.K., Yang, Y.: Adversarial camouflage: Hiding physical-world attacks with natural styles. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1000-1008 (2020) 2", + "16. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018) 2, 3", + "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) 3", + "18. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020) 1", + "19. Hoory, S., Shapira, T., Shabtai, A., Elovici, Y.: Dynamic adversarial patch for evading object detection models. arXiv preprint arXiv:2010.13070 (2020) 2", + "20. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7848-7857 (2021) 2, 3", + "21. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13307-13316 (2022) 2, 3", + "22. Huang, L., Gao, C., Zhou, Y., Xie, C., Yuille, A.L., Zou, C., Liu, N.: Universal physical camouflage attacks on object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 720-729 (2020) 2, 3", + "23. Huang, W., Zhao, X., Jin, G., Huang, X.: Safari: Versatile and efficient evaluations for robustness of interpretability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1988-1998 (2023) 1", + "24. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018) 3", + "25. Im Choi, J., Tian, Q.: Adversarial attack and defense of yolo detectors in autonomous driving scenarios. In: 2022 IEEE Intelligent Vehicles Symposium (IV). pp. 1011-1017. IEEE (2022) 3", + "26. Jia, S., Yin, B., Yao, T., Ding, S., Shen, C., Yang, X., Ma, C.: Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. Advances in Neural Information Processing Systems 35, 34136-34147 (2022) 3", + "27. Jin, G., Yi, X., Huang, W., Schewe, S., Huang, X.: Enhancing adversarial training with second-order statistics of weights. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15273-15283 (2022) 3", + "28. Jin, G., Yi, X., Wu, D., Mu, R., Huang, X.: Randomized adversarial training via taylor expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16447-16457 (2023) 3" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Michael, K., Fang, J., Yifu, Z., Wong, C., Montes, D., et al.: ultralytics/yolov5: v7. 0-yolov5 sota realtime instance segmentation. Zenodo (2022) 1", + "30. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022) 3", + "31. Kurakin, A., Goodfellow, I., Bengio, S., et al.: Adversarial examples in the physical world (2016) 3", + "32. Lee, M., Kolter, Z.: On physical adversarial patches for object detection. arXiv preprint arXiv:1906.11897 (2019) 4, 10, 11, 16", + "33. Lei, X., Cai, X., Lu, C., Jiang, Z., Gong, Z., Lu, L.: Using frequency attention to make adversarial patch powerful against person detector. IEEE Access 11, 27217-27225 (2022) 2", + "34. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4282-4291 (2019) 1", + "35. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8971-8980 (2018) 1", + "36. Li, J., Schmidt, F., Kolter, Z.: Adversarial camera stickers: A physical camera-based attack on deep learning systems. In: International Conference on Machine Learning. pp. 3896-3904. PMLR (2019) 3", + "37. Li, Y., Li, Y., Dai, X., Guo, S., Xiao, B.: Physical-world optical adversarial attacks on 3d face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24699-24708 (2023) 3", + "38. Liang, S., Wu, B., Fan, Y., Wei, X., Cao, X.: Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970 (2022) 3", + "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 3, 9", + "40. Liu, A., Liu, X., Fan, J., Ma, Y., Zhang, A., Xie, H., Tao, D.: Perceptual-sensitive gan for generating adversarial patches. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1028-1035 (2019) 2", + "41. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023) 3", + "42. Liu, X., Yang, H., Liu, Z., Song, L., Li, H., Chen, Y.: Dpatch: An adversarial patch attack on object detectors. arXiv preprint arXiv:1806.02299 (2018) 4, 11", + "43. Lovisotto, G., Turner, H., Sluganovic, I., Strohmeier, M., Martinovic, I.: {SLAP}: Improving physical adversarial examples with {Short-Lived} adversarial perturbations. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1865-1882 (2021) 2, 3", + "44. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 9", + "45. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 7" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016) 3", + "47. Pomponi, J., Scardapane, S., Uncini, A.: Pixle: a fast and effective black-box attack based on rearranging pixels. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2022) 3", + "48. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 1", + "49. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) 1", + "50. Shapira, A., Zolfi, A., Demetrio, L., Biggio, B., Shabtai, A.: Phantom sponges: Exploiting non-maximum suppression to attack deep object detectors. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4571-4580 (2023) 3", + "51. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In: Proceedings of the 2016 acm sig sac conference on computer and communications security. pp. 1528-1540 (2016) 3", + "52. Shi, Z., Yang, W., Xu, Z., Yu, Z., Huang, L.: Reinforcement learning-based adversarial attacks on object detectors using reward shaping. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8424-8432 (2023) 3", + "53. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013) 3", + "54. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) 3", + "55. Tu, J., Ren, M., Manivasagam, S., Liang, M., Yang, B., Du, R., Cheng, F., Urtasun, R.: Physically realizable adversarial examples for lidar object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13716-13725 (2020) 3", + "56. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2414-2422 (2022) 3", + "57. Wang, D., Yao, W., Jiang, T., Li, C., Chen, X.: Rfla: A stealthy reflected light adversarial attack in the physical world. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4455-4465 (2023) 3", + "58. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018) 1", + "59. Wang, J., Liu, A., Yin, Z., Liu, S., Tang, S., Liu, X.: Dual attention suppression attack: Generate adversarial camouflage in physical world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8565-8574 (2021) 2, 3", + "60. Wang, Q., Zhang, L., Bertinetto, L., Hu, W., Torr, P.H.: Fast online object tracking and segmentation: A unifying approach. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 1328-1338 (2019) 1" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021) 3", + "62. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020) 3", + "63. Wu, W., Su, Y., Chen, X., Zhao, S., King, I., Lyu, M.R., Tai, Y.W.: Boosting the transferability of adversarial samples via attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1161-1170 (2020) 3", + "64. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019) 3", + "65. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 665-681. Springer (2020) 3", + "66. Zhang, J., Wu, W., Huang, J.t., Huang, Y., Wang, W., Su, Y., Lyu, M.R.: Improving adversarial transferability via neuron attribution-based attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14993-15002 (2022) 3", + "67. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15115-15125 (2022) 3", + "68. Zhang, L., Xu, N., Yang, P., Jin, G., Huang, C.C., Zhang, L.: Trajpac: Towards robustness verification of pedestrian trajectory prediction models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8327-8339 (2023) 1", + "69. Zhao, Y., Zhu, H., Liang, R., Shen, Q., Zhang, S., Chen, K.: Seeing isn't believing: Towards more robust adversarial attack against real world object detectors. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 1989-2004 (2019) 2, 3", + "70. Zhong, Y., Liu, X., Zhai, D., Jiang, J., Ji, X.: Shadows can be dangerous: Stealthy and effective physical-world adversarial attack by natural phenomenon. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15345-15354 (2022) 2", + "71. Zhu, L., Wang, T., Li, J., Zhang, Z., Shen, J., Wang, X.: Efficient query-based black-box attack against cross-modal hashing retrieval. ACM Transactions on Information Systems 41(3), 1-25 (2023) 3", + "72. Zhu, X., Li, X., Li, J., Wang, Z., Hu, X.: Fooling thermal infrared pedestrian detectors in real world using small bulbs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3616-3624 (2021) 3", + "73. Zolfi, A., Kravchik, M., Elovici, Y., Shabtai, A.: The translucent patch: A physical and universal attack on object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15232-15241 (2021) 2" + ], + "bbox": [ + 212, + 146, + 787, + 797 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Out-of-Bounding-Box Triggers", + "bbox": [ + 527, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_model.json b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ea24df8fe73f2f05994ebdb9f9b77fb97685af4 --- /dev/null +++ b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_model.json @@ -0,0 +1,3373 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.187 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers: A Stealthy Approach to Cheat Object Detectors" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.219, + 0.772, + 0.259 + ], + "angle": 0, + "content": "Tao Lin\\(^{1,2}\\), Lijia Yu\\(^{1}\\), Gaojie Jin\\(^{1,2}\\), Renjue Li\\(^{1,2}\\), Peng Wu\\(^{1,2,\\text{图}}\\), and Lijun Zhang\\(^{1,2,3,\\text{图}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.269, + 0.786, + 0.312 + ], + "angle": 0, + "content": "1 Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.312, + 0.703, + 0.326 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Chinese Academy of Sciences, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.326, + 0.705, + 0.354 + ], + "angle": 0, + "content": "3 Automotive Software Innovation Center, Chongqing, China {lintao,yulijia,gaojie,lirj19,wp,zhanglj}@ios.ac.cn" + }, + { + "type": "list", + "bbox": [ + 0.297, + 0.312, + 0.705, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.387, + 0.744, + 0.61 + ], + "angle": 0, + "content": "Abstract. In recent years, the study of adversarial robustness in object detection systems, particularly those based on deep neural networks (DNNs), has become a pivotal area of research. Traditional physical attacks targeting object detectors, such as adversarial patches and texture manipulations, directly manipulate the surface of the object. While these methods are effective, their overt manipulation of objects may draw attention in real-world applications. To address this, this paper introduces a more subtle approach: an inconspicuous adversarial trigger that operates outside the bounding boxes, rendering the object undetectable to the model. We further enhance this approach by proposing the Feature Guidance (FG) technique and the Universal AutoPGD (UAPGD) optimization strategy for crafting high-quality triggers. The effectiveness of our method is validated through extensive empirical testing, demonstrating its high performance in both digital and physical environments. The code and video will be available at: https://github.com/linToTao/Out-of-bbox-attack." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.622, + 0.741, + 0.65 + ], + "angle": 0, + "content": "Keywords: Universal physical adversarial attack \\(\\cdot\\) Out-of-bounding-box trigger \\(\\cdot\\) Object detection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Object detection aims to identify instances of semantic objects typically within images or video clips. It finds widespread applications across a range of domains, including face recognition [10, 11, 23, 49, 58], object tracking [34, 35, 60], and autonomous driving [6, 18, 68]. Especially in autonomous driving systems, object detectors play a crucial role in tasks such as recognizing traffic signs, pedestrians, vehicles, and traffic lights. Nonetheless, in recent years, security concerns have emerged regarding object detectors, primarily due to the susceptibility of deep neural networks (DNNs) to adversarial examples (AEs). AEs are intentionally crafted inputs designed to deceive DNNs into making incorrect predictions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.143, + 0.774, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 1: An illustration of vanilla object attack methods and our novel approach. Typically, vanilla attack methods target the most informative regions of an object, such as the central, highly visible areas of the target object, e.g., the stop sign in (a) [16], (b) [69], (c) [15], and the car in (d) [19], (e) [59], (f) [22]. In contrast, our method adopts a more covert strategy, focusing on attacking the peripheral edge areas surrounding the stop sign." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Early research mainly focused on studying AEs like adversarial patch [16, 20, 33, 40, 73], optical [43, 70] and texture [15, 21], which are targeted at the objects. These AEs involve the computation of perturbations applied directly to the objects to mislead a DNN-based detector. However, a direct attack on the object as shown in Figure 1, along with its subsequent attempt to deceive the detector, is overly conspicuous and cannot be deployed without raising suspicion in real-world practical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.79, + 0.59 + ], + "angle": 0, + "content": "In this work, we provide an alternative perspective: is it possible to introduce an adversarial trigger in an inconspicuous area outside the object while still successfully deceiving the detector?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Actually, crafting an adversarial trigger within the areas beyond the object boundaries is even more demanding. On one hand, for dense prediction networks like object detectors, the output results are more reliant on the red regions of the heatmap in Figure 1. This implies that tailoring the adversarial trigger by utilizing irrelevant information, like the blue regions of the heatmap, is sharply challenging. More difficultly, our trigger is designed to conform to practical size and shape constraints. Nevertheless, for exactly the same reasons, our attack method gets more delusive and harder to defense." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paper is focused on the development of robust and inconspicuous triggers to attack typical object detectors deployed in practical scenarios. To tackle the challenge mentioned above, we present a novel technique called Feature Guidance (FG). In this approach, we guide the feature layers with an adversarial trigger to resemble those with the object hidden. And we propose a optimization strategy, called Universal Auto-PGD (UAPGD), applicable to universal attacks. This optimization strategy allows for the selection of an appropriate timing to halve the step size, based on the oscillation of the loss function." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.221 + ], + "angle": 0, + "content": "To assess the effectiveness of our approach in both digital and physical domains, we trained the adversarial trigger and evaluated the corresponding metrics on COCO [39] dataset and Carla [14], respectively. We also evaluated the robustness of the trigger by recording videos using an onboard camera in the real world." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.223, + 0.629, + 0.238 + ], + "angle": 0, + "content": "Overall, this paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.245, + 0.784, + 0.289 + ], + "angle": 0, + "content": "- This work studies the realm of adversarial robustness in object detection, providing a novel perspective by exploring the concept of an adversarial trigger applied outside an object." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.291, + 0.785, + 0.363 + ], + "angle": 0, + "content": "- We propose a novel attack method designed to generate triggers that achieve excellent continuous attack effects and demonstrate high robustness in real world. The key enabler of our method is the utilization of similarity between feature layers and adaptive step size decay to endow the trigger with the ability to mislead the detector." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.367, + 0.785, + 0.409 + ], + "angle": 0, + "content": "- An extensive set of empirical results are provided to demonstrate that our attack exhibits strong performance and robustness in both the COCO dataset and Carla simulator, as well as in real-world scenarios." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.245, + 0.785, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.383, + 0.449 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Adversarial examples (AEs), first discovered by Szegedy et al. [53], involve carefully crafted inputs inducing model misclassifications. Transferability [26, 61, 64, 66] of AEs, allows them to deceive networks with different hyperparameters or architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Digital adversarial examples are generated by directly manipulating any pixels in digital images prior to feeding them into the model. Adversarial attacks can be classified into black-box attacks and white-box attacks based on whether the internal structure and parameters of the target model are known. White-box attacks [5, 12, 17, 27, 28, 31, 44, 46] primarily rely on the computation of model gradients. Black-box methods, on the other hand, depend on querying the unknown model by observing the prediction [12, 13, 38, 47, 62–64, 71] or leveraging an agent model to imitate the threaten model [3, 7, 8, 24, 30, 67]. Recently, Croce et al. [9] proposed AutoAttack, a parameter-free method that combines novel white-box and black-box attack strategies. The works of applying digital adversarial examples to object detectors have been explored in [25] and [50] that introduce perturbations to input images to mislead detectors. [52] generate more subtle perturbations via reinforcement learning, successfully attacking the detector." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Physical adversarial examples are crafted to be printed out and captured by sensory equipment like camera [31,36] and LiDAR [4,41,55]. These examples pose a real-world threat due to their ability to adapt to camera and scenario variations. Early studies focused on attacking recognition models [16,51]. Recent research has explored physical attacks in the more challenging context of object detection. Perturbations in this domain include adversarial patches [1,20,54,65, 69], adversarial camouflage textures [21,22,56,59], and optical attacks [37,43,57, 72]. To improve robustness against various transformations in physical settings," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.177 + ], + "angle": 0, + "content": "Athalye et al. [2] introduced Expectation Over Transformation (EOT), widely used in physical attacks, to enhance the effectiveness of adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.268 + ], + "angle": 0, + "content": "Prior physical attack methods need to be applied directly to the objects so that these methods are expected to yield favorable results but they are prone to raising suspicion in practical scenarios. Unlike these works, our method constructs a universal attack trigger that can deceive the detector disregarding the specific object without altering the appearance of itself, and can be discreetly deployed in the real world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.357 + ], + "angle": 0, + "content": "The works that exhibit the closest alignment with our motivation are Dpatch [42] and [32] (after this referred to as Dpatch2). However, we must emphasize that our work is independent of them, highlighting the difference between our work and theirs in Appendix F. Furthermore, in the experiments section, we compare the attack effectiveness of our method with Dpatch2 [32] on the COCO dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.382, + 0.381, + 0.4 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.673 + ], + "angle": 0, + "content": "In this work, we aim to generate physical adversarial triggers that can be naturally deployed in an inconsequential position. These adversarial triggers are designed not to destroy the object information but still deceive the model effectively. In this section, we commence by presenting the problem formulation in Section 3.1. Subsequently, Fig. 2a provides a comprehensive illustration of our framework. Our primary focus is on the white-box attack paradigm, which necessitates an intricate optimization process. Given an object detection model, we design a loss function to determine the gradient direction for adversarial attacks in Section 3.2. This loss function is then utilized iteratively to update the adversarial trigger. In the pursuit of strengthening attack performance and enhancing the continuity of attacks in video-based scenarios, we introduce a novel method named \"Feature Guidance\" (FG) and an additional loss term \\( L_{FG} \\) in Section 3.2. To bolster the overall robustness and deployability of our approach, we also employ two indispensable techniques, EOT and total variation loss in Section 3.2. Furthermore, our methodology draws inspiration from the work of [9] in their Auto PGD framework during the training process, we thus modify the conditions of step size decay based on the universal training mode (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.44, + 0.709 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Adversarial attacks on object detection can have diverse objectives, encompassing but not limited to inducing misalignment in bounding boxes, reducing classification confidence, or provoking misclassification. Our goal is to generate adversarial examples for a specific class, with the aim of causing the object detection algorithm to fail in identifying objects belonging to that class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Given an input image \\( x \\in \\mathbb{R}^d \\) with the ground truth \\( y = \\{(y_i^{\\mathrm{coor}}, y_i^{\\mathrm{cls}})\\}_{i \\in [n]} \\subset \\mathbb{R}_+^4 \\times \\mathbb{Z}_+ \\), \\( n \\) is the number of objects in the image, \\( y_i^{\\mathrm{coor}} \\) contains top-left and bottom-right coordinates of the bounding box in the image, and \\( y_i^{\\mathrm{cls}} \\) is the class of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.142, + 0.501, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.174, + 0.768, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Fig.2: (a) shows the overview of our adversarial trigger generation framework. We deploy the adversarial trigger onto each image in the dataset with robustness transformation. With a fixed detector, our method optimizes the adversarial trigger through the obtained adversarial gradient. (b) illustrates the motivation and calculation method of Feature Guidance. Taking a stop sign as an example, for a given image \\( x \\), minimizing \\( \\| f_z(T(x,\\mathcal{A}_t)) - f_z(x')\\|_2 \\) can mislead a detector to classify the stop sign image \\( T(x,\\mathcal{A}_t) \\) as none stop." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.441 + ], + "angle": 0, + "content": "object. Then, the DNN-based object detector can be represented as the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.441, + 0.584, + 0.459 + ], + "angle": 0, + "content": "\\[\nf: \\mathbb {R} ^ {d} \\to 2 ^ {\\mathbb {R} _ {+} ^ {4} \\times \\mathbb {Z} _ {+} \\times [ 0, 1 ] ^ {2}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.588 + ], + "angle": 0, + "content": "For a given image \\( x \\), the output of \\( f \\) can be represented as \\( \\tilde{y} = \\{(\\tilde{y}_i^{\\mathrm{coor}},\\tilde{y}_i^{\\mathrm{cls}},\\tilde{y}_i^{\\mathrm{conf}})\\}_{i\\in [n]} \\subset \\mathbb{R}_+^4\\times \\mathbb{Z}_+ \\times [0,1]^2 \\), where \\( \\tilde{y}_i^{\\mathrm{conf}} \\) denotes the confidence associated with the predicted box coordinates \\( \\tilde{y}_i^{\\mathrm{coor}} \\) and the predicted class \\( \\tilde{y}_i^{\\mathrm{cls}} \\) for \\( i \\)-th object. We aim to generate a universal adversarial trigger \\( \\mathcal{A}_t\\in [0,1]^{3\\times U\\times V} \\), where \\( U,V\\in \\mathbb{Z}_{+} \\) are the given dimensions, for the given specific class \\( t \\), such that these adversarial triggers can cause \\( f \\) to fail in recognizing objects of class \\( t \\). Then, the problem can be formulated as an optimization task, represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.589, + 0.787, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a r g m i n} _ {\\mathcal {A} _ {t}} \\mathbb {E} _ {x \\sim D} [ | f (T (x, \\mathcal {A} _ {t})) \\cap (\\mathbb {R} _ {+} ^ {4} \\times \\{t \\} \\times [ 0, 1 ] ^ {2}) | ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.734 + ], + "angle": 0, + "content": "where \\( |A| \\) denotes the cardinality of the set \\( A \\), \\( D \\) represents the input distribution, and \\( T \\) (which will be elaborated in the following) is a transform operation to incorporate the adversarial trigger into the original image. In contrast to prior studies based on adversarial patches, our approach avoids placing the adversarial trigger on the informative surface, e.g., traffic signs, thereby preventing any unnecessary attention. We would like to point out that this kind of attack method is challenging because the adversarial trigger can only be stuck on the unimportant places of the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Transform operation. We utilize a differentiable operation \\( T \\) to render the adversarial trigger \\( \\mathcal{A}_t \\) onto an image \\( x \\). Let the object with class \\( t \\) be located in the box \\( B \\) of the image, we use the following two-step to add the adversarial trigger to the image:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.825 + ], + "angle": 0, + "content": "(1) We can select an appropriate area \\( C(B) \\) in the image where \\( C \\) is a given function, such position does not overlap with the surface of the object;" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.784, + 0.842 + ], + "angle": 0, + "content": "(2) We cover the box \\( C(B) \\) part of the image \\( x \\) with adversarial trigger \\( \\mathcal{A}_t \\), but" + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Considering that the adversarial trigger and the box are not the same size, we need to apply the following affine matrix to \\(\\mathcal{A}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.189, + 0.786, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\theta = \\left( \\begin{array}{c c c} S _ {h} \\cos \\alpha & - S _ {h} \\sin \\alpha & T _ {h} \\\\ S _ {v} \\sin \\alpha & S _ {v} \\cos \\alpha & T _ {v} \\\\ 0 & 0 & 1 \\end{array} \\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.784, + 0.306 + ], + "angle": 0, + "content": "\\(S_{h}\\) and \\(S_{v}\\) denotes the scaling factors utilized to resize the adversarial trigger, \\(\\alpha\\) represents the angle of rotation, \\(T_{h}\\) and \\(T_{v}\\) are the distance adversarial trigger need to shift. These parameters are all computed by \\(C(B)\\). The more specific remark regarding \\(C(B)\\) will be detailed in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Note that our attack method has two characteristics: Our attack is out of the object; our attack area changes depending on the position of the object. We were the first to attack target detection with such a method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.374, + 0.625, + 0.389 + ], + "angle": 0, + "content": "3.2 Adversarial gradient with Feature Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.502 + ], + "angle": 0, + "content": "In this paper, we focus on utilizing white-box attack methods to search for adversarial gradients and optimize the adversarial trigger. Additionally, we provide black-box empirical results in Appendix D that the trigger is generated by a surrogate model. As illustrated above, we render the adversarial trigger onto images and feed them to the detectors, take the derivative of the loss function, and use it to update the adversarial trigger. The loss function consists of the following parts:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.784, + 0.579 + ], + "angle": 0, + "content": "Adversarial loss of detector output. As mentioned above, the detector outputs the predicted class and bounding boxes and their confidence. Considering most of the detectors will not output the predicted class and boxes with a confidence lower than a threshold, we want images with the adversarial trigger to make the confidence as low as possible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.784, + 0.61 + ], + "angle": 0, + "content": "Let \\( f_{i,coor}^{\\mathrm{conf}}(x) \\) and \\( f_{i,cls}^{\\mathrm{conf}}(x) \\) is the confidence of the \\( i \\)-th object on image \\( x \\), we can define the first part of our loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.622, + 0.785, + 0.647 + ], + "angle": 0, + "content": "\\[\nL _ {d e t} \\left(\\mathcal {A} _ {t}\\right) = \\mathbb {E} _ {x \\sim D} \\left[ \\max _ {i: \\tilde {y} _ {i} ^ {c l s} = t} f _ {i, c o o r} ^ {\\text {c o n f}} (T (x, \\mathcal {A} _ {t})) f _ {i, c l s} ^ {\\text {c o n f}} (T (x, \\mathcal {A} _ {t})) \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.704 + ], + "angle": 0, + "content": "Feature guidance. The biggest difficulty with an attack outside the object is that the features of the original object on the image will not be destroyed, so the network will still notice the presence of the object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We gain the following insights from the network structure to address this issue. A DNN-based detector is usually designed with three modules: Backbone, Neck, and Head. The Backbone module extracts global features from an input image. These features are then fused through the Neck module to produce multiple feature maps with varying receptive fields. The Head module subsequently makes predictions based on the feature maps. This modular structure imitates how people often perceive a comprehensive horizon (like the Backbone module) and focus on varying attractive areas (like the Neck module). Subsequently, the attractive areas will contribute to making perceptions (like the Head module)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Hence, in order to provoke model errors, it is imperative to deliberately induce information misleading prior to inputting the head module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "To achieve that, we use the following method, called Feature guidance, as shown in Fig. 2b. For a given image \\( x \\) with a target object \\( t \\), we can get the image \\( x' \\), which differs from \\( x \\) only with the object \\( t \\) missing. Then, if we can elaborate our adversarial trigger \\( \\mathcal{A}_t \\) to guide the feature map \\( f_z(T(x, \\mathcal{A}_t)) \\) as close to \\( f_z(x') \\) generated by the Neck module of the detector, it will have a desirable impact on the Head module such that the stop sign cannot be detected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.785, + 0.313 + ], + "angle": 0, + "content": "As illustrated in Figure 2a, for the sake of convenience, we derive \\( x_{mask} \\) that exhibits an analogous effect to \\( x' \\) simply by covering the bounding box position in the ground truth of \\( x \\) with a gray rectangle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.342 + ], + "angle": 0, + "content": "In order to achieve Feature guidance, we need to make \\(\\mathcal{A}_t\\) to minimum such loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.356, + 0.673, + 0.373 + ], + "angle": 0, + "content": "\\[\nL _ {F G} (\\mathcal {A} _ {t}) = E _ {x \\sim D} [ \\| f _ {z} (T (x, \\mathcal {A} _ {t})) - f _ {z} (x _ {\\text {m a s k}}) \\| _ {2} ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.785, + 0.461 + ], + "angle": 0, + "content": "Other skills To enhance the robustness of our adversarial trigger \\(\\mathcal{A}_t\\), we incorporate diverse variations, including random noise, contrast, brightness, and rotation, into the generated trigger prior to its application on images, thus simulating real-world conditions. To address printing challenges, we employ TV loss [45] on the adversarial trigger to promote smoothness. TV loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.316, + 0.473, + 0.785, + 0.509 + ], + "angle": 0, + "content": "\\[\nL _ {t v} \\left(\\mathcal {A} _ {t}\\right) = \\sum_ {i, j} \\sqrt {\\left(\\mathcal {A} _ {t} ^ {i , j} - \\mathcal {A} _ {t} ^ {i + 1 , j}\\right) ^ {2} + \\left(\\mathcal {A} _ {t} ^ {i , j} - \\mathcal {A} _ {t} ^ {i , j + 1}\\right) ^ {2}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.766, + 0.53 + ], + "angle": 0, + "content": "where the subindices \\(i\\) and \\(j\\) represent the pixel coordinate of the trigger \\(\\mathcal{A}_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.784, + 0.555 + ], + "angle": 0, + "content": "Overall, we will get the adversarial trigger \\(\\mathcal{A}_t\\) by optimizing following problem:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.557, + 0.785, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {A} _ {t} \\in [ 0, 1 ] ^ {3 \\times U \\times V}} L _ {d e t} (\\mathcal {A} _ {t}) + \\lambda_ {F G} L _ {F G} (\\mathcal {A} _ {t}) + \\lambda_ {t v} L _ {t v} (\\mathcal {A} _ {t}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.584, + 0.512, + 0.599 + ], + "angle": 0, + "content": "where \\(\\lambda_{FG}\\) and \\(\\lambda_{tv}\\) are hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.622, + 0.436, + 0.636 + ], + "angle": 0, + "content": "3.3 Universal Auto-PGD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.784, + 0.678 + ], + "angle": 0, + "content": "Let \\( L_{all} = L_{det} + \\lambda_{FG}L_{FG} + \\lambda_{tv}L_{tv} \\), we will use gradient descent to find the adversarial trigger via minimizing the loss function \\( L_{all} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.785, + 0.723 + ], + "angle": 0, + "content": "Considering the instability that appeared in the experiment when optimizing multiple terms of \\( L_{all} \\) with conventional methods, we propose the following optimization strategy called Universal Auto-PGD (UAPGD):" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.784, + 0.768 + ], + "angle": 0, + "content": "At the beginning of optimizing, we fix the following values: comparison length \\( l_{c} \\), observation length \\( l_{o} \\), the slack variable \\( \\epsilon_{1}, \\epsilon_{2} \\), let the initial step size be \\( \\eta \\) and we record the following values:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.769, + 0.707, + 0.785 + ], + "angle": 0, + "content": "(1): After epoch \\(j\\), record the value of the loss function, name it \\(L_{j}\\);" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.785, + 0.533, + 0.799 + ], + "angle": 0, + "content": "(2): After epoch \\( i * l_{o} \\) where \\( i \\in Z_{+} \\), record:" + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.769, + 0.707, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.811, + 0.645, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\min , i} = \\min _ {j \\in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \\left\\{L _ {j} \\right\\} \\\\ V _ {i} = \\operatorname {V a r i a n c e} \\left[ \\left\\{L _ {j} \\right\\} _ {j \\in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \\right] \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.23, + 0.15, + 0.404, + 0.164 + ], + "angle": 0, + "content": "Algorithm 1: UAPGD" + }, + { + "type": "algorithm", + "bbox": [ + 0.218, + 0.167, + 0.787, + 0.61 + ], + "angle": 0, + "content": "Input: Threaten network \\(f\\) , initial step size \\(\\eta\\) , the number of epoch \\(N_{\\mathrm{epoch}}\\) dataset \\(D = \\cup_{j = 1}^{k}D_{j}\\) which was divided into k batches, the hyperparameter in UAPGD \\(l_c,l_o\\) , initial adversarial trigger \\(\\mathcal{A}_t^0\\) Output: Adversarial trigger \\(\\mathcal{A}_{best}\\) \n1 \\(S_{c}\\gets \\phi\\) \\(S_{o}\\gets \\phi\\) \\(L_{best}\\gets \\infty\\) \\(\\mathcal{A}_{best}\\leftarrow \\mathcal{A}_t^0\\) \n2 for \\(i\\in [N_{\\mathrm{epoch}}]\\) do \n3 \\(L_{i}\\gets 0\\) \n4 \\(\\mathcal{A}_{t}^{i,0}\\gets \\mathcal{A}_{t}^{i - 1}\\) \n5 for \\(j\\in [k]\\) do \n6 Get \\(L_{all}(\\mathcal{A}_t^{i,j - 1},D_j)\\) via \\(f\\) \n7 \\(L_{i}\\gets L_{i} + L_{all}(\\mathcal{A}_{t}^{i,j - 1},D_{j}) / k\\) \n8 \\(\\mathcal{A}_{t}^{i,j}\\gets \\mathcal{A}_{t}^{i,j - 1} - \\eta \\cdot \\mathrm{sign}(\\frac{\\nabla L_{all}(\\mathcal{A}_{t}^{i,j - 1},D_{j})}{\\nabla\\mathcal{A}_{t}^{i,j - 1}})\\) \n9 \\(\\mathcal{A}_{t}^{i,j}\\gets Clamp(\\mathcal{A}_{t}^{i,j},0,1)\\) \n10 \\(\\mathcal{A}_{t}^{i}\\gets \\mathcal{A}_{t}^{i,k}\\) \n11 if \\(L_{i} < L_{best}\\) then \n12 \\(\\begin{array}{r}\\big{\\lfloor}\\mathcal{A}_{best}\\leftarrow \\mathcal{A}_{t}^{i}\\text{and} L_{best}\\leftarrow L_{i} \\end{array}\\) \n13 \\(S_{o} = S_{o}\\cup L_{i}\\) \n14 if \\(|S_{o}| = l_{o}\\) then \n15 \\(\\begin{array}{rl} & L_{min}\\gets \\min_{L\\in S_o\\backslash \\{S_o\\}}\\\\ & V\\gets \\operatorname{Var}_{L\\in S_o\\backslash \\{S_o\\}}\\\\ & S_c = S_c\\cup (L_{min},V)\\\\ & S_o\\gets \\phi \\end{array}\\) \n16 \n17 \n18 \n19 if \\(|S_{c}| = l_{c}\\) then // Condition 1 and Condition 2 are introduced in section 3.3 if Condition 1(Sc) and Condition 2(Sc) then \n21 \\(\\begin{array}{r}\\lfloor \\eta \\leftarrow \\eta /2\\text{and}\\mathcal{A}_t^i\\leftarrow \\mathcal{A}_{best}\\\\ S_c\\leftarrow \\phi \\end{array}\\) \n22" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.697 + ], + "angle": 0, + "content": "In order to prevent violent oscillations around the local optimum during the optimization process, we will halve the step size when the following conditions are met:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.699, + 0.604, + 0.716 + ], + "angle": 0, + "content": "(Con 1): \\( L_{min,j} \\geq L_{min,q} - \\epsilon_1 \\) for all \\( (j - l_c) < q < j \\);" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.715, + 0.53, + 0.731 + ], + "angle": 0, + "content": "(Con 2): \\(V_{j}\\geq V_{q} - \\epsilon_{2}\\) for all \\(j - l_{c} < q < j\\)" + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.699, + 0.604, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.776 + ], + "angle": 0, + "content": "Due to the stochasticity from the robustness transformation mentioned in Other skills, the value of the loss function may fluctuate so that we incorporate the slack variable \\(\\epsilon_{1},\\epsilon_{2}\\) into numerical comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Intuitively, these two conditions respectively describe the exploration and exploitation aspects in the optimization process. Con 1 means that the value of the loss function has not decreased much, indicating the necessity of a transition from exploring the entire feasible space to focusing on local optimization. Con" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "2 characterizes the oscillation of the loss function, of which the amplitude per \\( l_{c} \\) steps has not decreased much. This implies the imminent conclusion of the algorithm's exploitation in a local space. The simultaneous occurrence of c1 and c2 means that the adversarial trigger jumps near one of the optimal points, but the step size is too large to fall steadily near the best advantages. Therefore, at this point, we will halve the step size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Our approach is somewhat similar to AutoAttack [9], but there are several differences: (1): Our method is for universal attack, AutoAttack is for attacking a single image; (2): We consider the amplitude of the oscillation during the optimization process and use this as an indicator of decline; (3): We include the slack variables in order to mitigate potential misleading phenomena caused by experimental randomness. For example, the oscillation of the loss function often easily cause \\( L_{j} \\) to show a downward trend." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.343, + 0.787, + 0.375 + ], + "angle": 0, + "content": "The overall process of our algorithm is shown in algorithm 1. And we use \\( L_{all}(\\mathcal{A}_t, D_j) \\) to represent calculating \\( L_{all} \\) on an input batch \\( D_j \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.397, + 0.366, + 0.415 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.611 + ], + "angle": 0, + "content": "In this section, we choose the stop sign as the target object to evaluate the effectiveness of our adversarial attack method. The stop sign is a commonly used object in the evaluation of physical adversarial attacks, as it is critical to traffic safety. We begin by detailing the implementation settings of our evaluation. This is followed by an extensive set of experiments, utilizing the proposed adversarial trigger in the digital space, specifically on the COCO dataset [39]. Furthermore, we extend the application of our method to real-world scenarios, i.e., we record videos to demonstrate that our method can attack the object detector in both simulators and real-world scenes. Meanwhile, we evaluate our attack method across varying distances and angles. See Appendix A and B for the basic experimental setting, the results of additional attacks on other objects, and the detail of selecting C(B) to place the trigger on the stop sign." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.419, + 0.649 + ], + "angle": 0, + "content": "4.1 Digital experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.735 + ], + "angle": 0, + "content": "In this subsection, our experiments are carried out using the COCO dataset [39]. From this dataset, we extract a total of 1100 images featuring stop signs. Of these, 600 images constitute the training set, which is utilized to generate the adversarial trigger. The remaining 500 images are designated as the test set, serving to evaluate the effectiveness of the generated adversarial trigger." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To demonstrate the superiority of our algorithm, we adopt the following methods to generate adversarial triggers: optimizing the adversarial trigger using PGD [44] or UAPGD with a minimization of the loss function \\( L_{det} + \\lambda_{tv}L_{tv} \\) or \\( L_{all} \\). This included comparisons between our method and the normal PGD under the condition of placing the perturbation outside the object, as well as ablation experiments with varying hyperparameters. Specifically, for the YOLOv3(YOLOv5) model, we choose \\( \\lambda_{FG} \\) values of 0.1, 0.5, and 1.0(0.05," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "0.1, and 0.2). The optimization process utilized the initial step size of 8/255 and 16/255. We calculate the number of undetected images among all the images as a metric called attack success rate(ASR)." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.226, + 0.526, + 0.413 + ], + "angle": 0, + "content": "
Threaten ModelYOLOv3YOLOv5
DatasetTrain(%)Test(%)Train(%)Test(%)
Dpatch226.6724.007.176.67
Dpatch2 (Ldet)35.5335.8032.8332.00
Adam (16/255)65.1662.2043.8340.40
Adam (8/255)64.3360.8052.1650.80
PGD(16/255)56.6754.0043.6544.60
PGD(8/255)65.1762.2053.8351.00
UAPGD(16/255)66.0063.0057.8352.60
UAPGD(8/255)67.8364.8054.1750.40
FG(λF, FG, 16/255)57.6755.4051.8348.20
FG(λF, FG, 16/255)56.0054.6047.8344.60
FG(λF, FG, 16/255)54.0051.0043.8340.00
FG(λF, FG, 8/255)60.0055.2059.1755.60
FG(λF, FG, 8/255)64.8362.6052.1750.60
FG(λF, FG, 8/255)61.6759.8056.6756.40
Ensemble(λF, FG, 16/255)67.8366.2058.3356.40
Ensemble(λF, FG, 16/255)67.8364.8060.1757.60
Ensemble(λF, FG, 16/255)64.8361.6058.0056.60
Ensemble(λF, FG, 8/255)68.3365.0062.6758.60
Ensemble(λF, FG, 8/255)69.3366.6056.6753.20
Ensemble(λF, FG, 8/255)66.1662.2060.1757.80
" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.226, + 0.637, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.319, + 0.604, + 0.328 + ], + "angle": 0, + "content": "PGD" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.227, + 0.749, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.319, + 0.728, + 0.328 + ], + "angle": 0, + "content": "UAPGD" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.329, + 0.636, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.422, + 0.599, + 0.432 + ], + "angle": 0, + "content": "FG" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.329, + 0.749, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.422, + 0.731, + 0.432 + ], + "angle": 0, + "content": "Ensemble" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Fig. 3: The table on the left displays the ASR values of triggers, which are generated by various settings, on the training and testing sets of the COCO dataset. The values in parentheses represent the attack step size, and we employ different \\(\\lambda_{FG}^{i}\\) values for YOLOv3 and YOLOv5. For YOLOv3, we set \\(\\lambda_{FG}^{1} = 0.1\\), \\(\\lambda_{FG}^{2} = 0.5\\), \\(\\lambda_{FG}^{3} = 1.0\\); for YOLOv5, we set \\(\\lambda_{FG}^{1} = 0.05\\), \\(\\lambda_{FG}^{2} = 0.1\\), \\(\\lambda_{FG}^{3} = 0.2\\). Ensemble denotes the combined usage of FG and UAPGD. Dpatch2 is the original approach from [32] and Dpatch2(\\(L_{det}\\)) denotes the original approach adopts the \\(l_{det}\\) as the optimization objective. The bolded values indicate the optimal results for each respective column in the table. There are some results of triggers on the right generated under different settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.682 + ], + "angle": 0, + "content": "As shown in Figure 3, We compare our method with the more reasonable and advanced Dpatch2 [32]. Considering the divergent optimization objectives, we substitute the maximization of the model loss function mentioned in [32] with the minimization of the proposed \\( L_{det} \\) referred to as Dpatch2(\\( L_{det} \\)). Undoubtedly, our method achieves the highest ASR in different experimental settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.761 + ], + "angle": 0, + "content": "A comparative analysis of the effectiveness between PGD and UAPGD reveals notable differences. Specifically, we observe that UAPGD, when employed with an identical attack step size as PGD, frequently results in a higher ASR. Moreover, it can be observed that UAPGD yields better performance compared to adaptive step size optimizers like Adam." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Figure 4a and Figure 4b present a detailed depiction of the variations in the detection loss \\( L_{det} \\) throughout the training process when generating adversarial triggers using both PGD and UAPGD techniques, specifically over YOLOv3 and YOLOv5 models. Notably, when operating under an equivalent number of epochs, UAPGD demonstrates a more pronounced capability to reduce \\( L_{det} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.151, + 0.361, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.22, + 0.347, + 0.232 + ], + "angle": 0, + "content": "(a) YOLOv3" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.151, + 0.49, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.22, + 0.476, + 0.232 + ], + "angle": 0, + "content": "(b) YOLOv5" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.151, + 0.619, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.22, + 0.576, + 0.232 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.152, + 0.75, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.22, + 0.706, + 0.232 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.243, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 4: (a) and (b) illustrate the variations of \\( L_{det} \\), during the training process of generating adversarial triggers using PGD and UAPGD for attacking the YOLOv3 and YOLOv5 models. (c) and (d) present the confidence of some samples from the COCO dataset, outputted by the detector, during the training process of individual attacks. The blue, green, and red curved lines represent the utilization of \\( L_{FG} \\), \\( L_{det} \\), and both as the loss function respectively. For individual attacks, the optimization process is relatively easier so that step size decay is unnecessary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.401 + ], + "angle": 0, + "content": "compared to PGD. The effectiveness of the loss monitoring in UAPGD will be analyzed in detail in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.492 + ], + "angle": 0, + "content": "Furthermore, as shown in Figure 3, UAPGD exhibits a superior capability in generating smoother adversarial triggers. This increased smoothness can be attributed to the implementation of step size decay, which allows for a more refined optimization process in the generation of triggers. Consequently, adversarial triggers crafted via UAPGD display enhanced suitability for practical applications, such as printing and deployment in the physical world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.786, + 0.689 + ], + "angle": 0, + "content": "Based on results in Figure 3, we can observe that integrating \\( L_{FG} \\) with suitable step size and weight into the loss function, coupled with using either PGD or UAPGD, can enhance the ASR. This improvement is particularly noticeable when applied to the YOLOv5 model. This observation highlights that employing appropriate feature guidance can improve the performance of universal attacks. However, incorporating FG did not always achieve satisfactory results under certain weights and step sizes. This phenomenon is because in universal attacks, the inclusion of \\( L_{FG} \\) introduces additional complexities to the optimization process. It requires appropriate adjustment of hyperparameters to fully leverage its effectiveness. Hence, the optimal outcome can be achieved by ensemble UAPGD with it, which allows for finer control over the step size to facilitate the optimization process. The empirical results indicate that the combination of UAPGD and FG in generating triggers yields the most effective attack performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.839 + ], + "angle": 0, + "content": "To further investigate the potential role of Feature Guidance, we conducted individual adversarial attacks on each image with an adversarial trigger (fixed step size of 8/255, 500iteration) in the training dataset. We identified two primary roles of FG: 1, as shown in Figure 4c, our observations reveal that FG can significantly influence the optimization direction during the training phase. This influence is evidenced by the expedited convergence in adversarial trigger training, indicating the advantageous role of FG in enhancing training efficiency. 2, As shown in Figure 4d, when some images are challenging to attack with out-of-box adversarial triggers, introducing FG would enhance the attack effectiveness at this time." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.465, + 0.162 + ], + "angle": 0, + "content": "4.2 Experiment in simulator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.544, + 0.264 + ], + "angle": 0, + "content": "In this section, we demonstrate the effects of our attack in autonomous driving scenarios, we use the Carla(0.9.14) simulator [14] to simulate the real driving process. In the simulator, we utilize the YOLOv3 model as the detection network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.546, + 0.401 + ], + "angle": 0, + "content": "Due to the significant differences between images from the COCO dataset and those generated by the Carla simulator, the adversarial trigger developed in Section 4.1 is not directly applicable to Carla. To tailor an adversarial trigger specifically for the Carla simulator, we compile a training dataset comprising 242 images featuring stop signs from the simulator. The process of attacking the Carla simulator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.787, + 0.446 + ], + "angle": 0, + "content": "is conducted in four distinct manners: employing either PGD or UAPGD, combined with a loss function of either \\( L_{det} + \\lambda_{tv}L_{tv} \\) or \\( L_{all} \\), where \\( \\lambda_{FG} = 0.1 \\) and the initial step size is 16/255." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.545, + 0.581 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the attack, we implement a procedure where the adversarial trigger is placed under the stop sign in the simulator. Subsequently, a vehicle is navigated towards this stop sign, and a 9-second video is recorded using an onboard camera. The key metric is the percentage of time during which the detection system fails to recognize the stop sign." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.546, + 0.734 + ], + "angle": 0, + "content": "As shown in Table 1, without any adversarial trigger, the car consistently recognizes the stop sign. When the attack is initiated using the PGD method alone, the attack duration is approximately 3 to 4 seconds. However, the integration of either UAPGD or FG methods into the attack strategy results in a slight increase in the attack duration by roughly an additional second. Notably, the implementation of an ensemble approach, which combines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.763 + ], + "angle": 0, + "content": "these methods, further enhances the attack performance, prolonging the duration to over 5 seconds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Moreover, to provide a more direct illustration of the attack impact, we have plotted the trend of the stop sign detection confidence as the car approaches it, shown as in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The empirical results reveal that the efficacy of our attack method consistently surpasses that of the vanilla PGD approach. Particularly in the final 5" + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.206, + 0.751, + 0.304 + ], + "angle": 0, + "content": "
attack methodProportion
No attack0%
PGD36.67%
UAPGD45.56%
PGD,FG46.67%
UAPGD,FG58.89%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.315, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Table 1: The proportion of time in the video duration during which the detector failed to recognize the stop sign." + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.49, + 0.756, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.608, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Fig.5: We uniformly extracted approximately 90 frames from the 9s video, which is recorded by the simulated vehicle's onboard camera, to showcase the confidence variations of the detector's predictions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.307, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.212, + 0.301, + 0.221 + ], + "angle": 0, + "content": "\\(conf = 0.98\\)" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.144, + 0.402, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.213, + 0.396, + 0.221 + ], + "angle": 0, + "content": "\\(conf = 0.98\\)" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.144, + 0.497, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.212, + 0.491, + 0.221 + ], + "angle": 0, + "content": "\\(conf = 0.99\\)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.144, + 0.591, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.212, + 0.582, + 0.221 + ], + "angle": 0, + "content": "\\(conf = 1.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.144, + 0.686, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.212, + 0.676, + 0.221 + ], + "angle": 0, + "content": "\\(conf = 1.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.144, + 0.781, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.212, + 0.771, + 0.221 + ], + "angle": 0, + "content": "conf \\(= 1.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.222, + 0.307, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.29, + 0.301, + 0.299 + ], + "angle": 0, + "content": "conf = 0.94" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.222, + 0.402, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.29, + 0.396, + 0.299 + ], + "angle": 0, + "content": "conf = 0.72" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.222, + 0.497, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.29, + 0.49, + 0.299 + ], + "angle": 0, + "content": "conf = 0.70" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.222, + 0.591, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.29, + 0.582, + 0.299 + ], + "angle": 0, + "content": "conf \\(< 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.222, + 0.686, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.29, + 0.678, + 0.299 + ], + "angle": 0, + "content": "conf \\(< 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.222, + 0.781, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.29, + 0.771, + 0.299 + ], + "angle": 0, + "content": "\\(conf = 0.6\\)" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.3, + 0.307, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.368, + 0.301, + 0.378 + ], + "angle": 0, + "content": "conf = 0.82" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.3, + 0.402, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.368, + 0.392, + 0.377 + ], + "angle": 0, + "content": "conf \\(< 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.3, + 0.497, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.368, + 0.487, + 0.377 + ], + "angle": 0, + "content": "conf \\(< 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.3, + 0.591, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.368, + 0.581, + 0.377 + ], + "angle": 0, + "content": "conf \\(< 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.3, + 0.686, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.368, + 0.676, + 0.377 + ], + "angle": 0, + "content": "conf \\(< 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.3, + 0.781, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.368, + 0.771, + 0.377 + ], + "angle": 0, + "content": "conf \\(< 0.5\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Fig. 6: Six frames are extracted uniformly from the video. Row 1 represents a clean stop sign with the detector's output. In Row 2 and Row 3, we deploy the trigger generated by the PGD and ensemble method respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.785, + 0.51 + ], + "angle": 0, + "content": "seconds of the video, during which the adversarial trigger persistently causes the detection system to lose sight of the stop sign. This effect illustrates the potential of our method to compromise the functionality of autonomous driving systems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.785, + 0.541 + ], + "angle": 0, + "content": "As shown in the Figure 6, we extract six frames uniformly from the video as examples to demonstrate the effects of this attack in driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.662 + ], + "angle": 0, + "content": "The observations imply that normally the car will always recognize the stop sign (Row 1 of Figure 6, each stop is marked by a red box). However, after the adversarial trigger created by our attack method (Row 3) is added to the stop sign, the car fails to recognize the stop sign(no red box). Compared with the attack effect of the adversarial trigger generated by PGD (Row 2), it can be seen that our method is clearly superior to the PGD. It is worth mentioning that the perturbation generated by Dpatch2 is suboptimal in the simulator, as demonstrated in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.58, + 0.705 + ], + "angle": 0, + "content": "4.3 Robustness of our attack in real world" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.785, + 0.779 + ], + "angle": 0, + "content": "To observe the attack effectiveness and robustness of the universal adversarial trigger obtained through algorithm 1 in the real world, we print out the adversarial trigger and post it under a stop sign. Moreover, in the case where clean images serve as a reference, we compare our method with the normal PGD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We examined the variation trends in the confidence of the detector's output for recognizing the stop sign using an onboard camera under different conditions: as the distance between the onboard camera and the stop sign decreased from far to near, and as the onboard camera's perspective towards the stop sign shifted" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.152, + 0.303, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.158, + 0.53, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.28, + 0.445, + 0.292 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.159, + 0.761, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.279, + 0.675, + 0.291 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.789, + 0.346 + ], + "angle": 0, + "content": "Fig. 7: We print and deploy the adversarial trigger in the physical world. (a) illustrates the confidence variation of the stop sign as the distance increases. (b) depicts the confidence variation of the stop sign when facing different angles changing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.784, + 0.404 + ], + "angle": 0, + "content": "from -70-degree side view to a 70-degree side view. The real-world experiment results are shown in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.571 + ], + "angle": 0, + "content": "According to Figure 7a, due to the relatively long distance between the vehicle and the stop position in the first half of the curve, the trigger occupies a small proportion of the field of view, which diminishes the stability of successful attacks. In the second half of the curve, where the distance between the stop and the vehicle is closer, the attacks become more stable. Our method consistently maintains a higher success rate in this scenario. Examining Figure 7b, it can be observed that when facing the stop sign directly, our method misleads the detector to output a remarkably low confidence. As the angle approaches approximately \\(\\pm 50\\) degrees, our method becomes ineffective. In contrast, the triggers generated by PGD are not smooth, which prevents them from effectively imprinting adversarial features into the real world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.789, + 0.633 + ], + "angle": 0, + "content": "Overall, in the real world, our attack consistently exhibits robust and effective performance when the onboard camera is approaching a stop sign. This means that our attack can successfully mislead an autonomous driving system to disregard the stop sign unconsciously." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.657, + 0.36, + 0.673 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this work, we present a novel adversarial attack methodology that extends beyond the bounding boxes of objects. This method poses greater challenges compared to previous studies due to its tendency to focus on less informative areas. Thus, we propose the FG techniques and UAPGD optimization strategy, specifically designed to augment the efficacy of this attack in both digital and physical realms. Furthermore, we validate the effectiveness of our attack method through its application in autonomous driving scenarios and real-world conditions. Discussion of limitations and potential negative impact of our works will be placed in Appendix G and immediately following that we also briefly address future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.281 + ], + "angle": 0, + "content": "We thank the anonymous referees for their valuable and insightful comments. This work is supported by the National Key R&D Program of China under Grant No. 2022YFA1005100, 2022YFA1005101, and 2022YFA1005104, the CAS Project for Young Scientists in Basic Research, under Grant No.YSBR-040, ISCAS New Cultivation Project ISCAS-PYFX-202201, ISCAS Basic Research ISCAS-JCZD-202302, and the Major Project of ISCAS under Grant No. ISCAS-ZD-202302." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.303, + 0.323, + 0.318 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.388 + ], + "angle": 0, + "content": "1. Aich, A., Li, S., Song, C., Asif, M.S., Krishnamurthy, S.V., Roy-Chowdhury, A.K.: Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1308-1318 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.787, + 0.429 + ], + "angle": 0, + "content": "2. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.787, + 0.47 + ], + "angle": 0, + "content": "3. Brendel, W., Rauber, J., Bethge, M.: Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.787, + 0.525 + ], + "angle": 0, + "content": "4. Cao, Y., Xiao, C., Cyr, B., Zhou, Y., Park, W., Rampazzi, S., Chen, Q.A., Fu, K., Mao, Z.M.: Adversarial sensor attack on lidar-based perception in autonomous driving. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 2267-2281 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.787, + 0.553 + ], + "angle": 0, + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.787, + 0.594 + ], + "angle": 0, + "content": "6. Chen, C., Seff, A., Kornhauser, A., Xiao, J.: Deepdriving: Learning affordance for direct perception in autonomous driving. In: Proceedings of the IEEE international conference on computer vision. pp. 2722-2730 (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.787, + 0.635 + ], + "angle": 0, + "content": "7. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 IEEE symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.787, + 0.69 + ], + "angle": 0, + "content": "8. Chen, P.Y., Zhang, H., Sharma, Y., Yi, J., Hsieh, C.J.: Zoo: Zeroth order optimization based black-box attacks to deep neural networks without training substitute models. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 15-26 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "9. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020) 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "0. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "1. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "2. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.23 + ], + "angle": 0, + "content": "14. Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: Carla: An open urban driving simulator. In: Conference on robot learning. pp. 1-16. PMLR (2017) 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.286 + ], + "angle": 0, + "content": "15. Duan, R., Ma, X., Wang, Y., Bailey, J., Qin, A.K., Yang, Y.: Adversarial camouflage: Hiding physical-world attacks with natural styles. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1000-1008 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "16. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "18. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "19. Hoory, S., Shapira, T., Shabtai, A., Elovici, Y.: Dynamic adversarial patch for evading object detection models. arXiv preprint arXiv:2010.13070 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.493 + ], + "angle": 0, + "content": "20. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7848-7857 (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.548 + ], + "angle": 0, + "content": "21. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13307-13316 (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.591 + ], + "angle": 0, + "content": "22. Huang, L., Gao, C., Zhou, Y., Xie, C., Yuille, A.L., Zou, C., Liu, N.: Universal physical camouflage attacks on object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 720-729 (2020) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "23. Huang, W., Zhao, X., Jin, G., Huang, X.: Safari: Versatile and efficient evaluations for robustness of interpretability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1988-1998 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.674 + ], + "angle": 0, + "content": "24. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "25. Im Choi, J., Tian, Q.: Adversarial attack and defense of yolo detectors in autonomous driving scenarios. In: 2022 IEEE Intelligent Vehicles Symposium (IV). pp. 1011-1017. IEEE (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "26. Jia, S., Yin, B., Yao, T., Ding, S., Shen, C., Yang, X., Ma, C.: Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. Advances in Neural Information Processing Systems 35, 34136-34147 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "27. Jin, G., Yi, X., Huang, W., Schewe, S., Huang, X.: Enhancing adversarial training with second-order statistics of weights. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15273-15283 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Jin, G., Yi, X., Wu, D., Mu, R., Huang, X.: Randomized adversarial training via taylor expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16447-16457 (2023) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Michael, K., Fang, J., Yifu, Z., Wong, C., Montes, D., et al.: ultralytics/yolov5: v7. 0-yolov5 sota realtime instance segmentation. Zenodo (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "30. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.256 + ], + "angle": 0, + "content": "31. Kurakin, A., Goodfellow, I., Bengio, S., et al.: Adversarial examples in the physical world (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.284 + ], + "angle": 0, + "content": "32. Lee, M., Kolter, Z.: On physical adversarial patches for object detection. arXiv preprint arXiv:1906.11897 (2019) 4, 10, 11, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.324 + ], + "angle": 0, + "content": "33. Lei, X., Cai, X., Lu, C., Jiang, Z., Gong, Z., Lu, L.: Using frequency attention to make adversarial patch powerful against person detector. IEEE Access 11, 27217-27225 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.365 + ], + "angle": 0, + "content": "34. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4282-4291 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.405 + ], + "angle": 0, + "content": "35. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8971-8980 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.446 + ], + "angle": 0, + "content": "36. Li, J., Schmidt, F., Kolter, Z.: Adversarial camera stickers: A physical camera-based attack on deep learning systems. In: International Conference on Machine Learning. pp. 3896-3904. PMLR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.487 + ], + "angle": 0, + "content": "37. Li, Y., Li, Y., Dai, X., Guo, S., Xiao, B.: Physical-world optical adversarial attacks on 3d face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24699-24708 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.528 + ], + "angle": 0, + "content": "38. Liang, S., Wu, B., Fan, Y., Wei, X., Cao, X.: Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.582 + ], + "angle": 0, + "content": "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.623 + ], + "angle": 0, + "content": "40. Liu, A., Liu, X., Fan, J., Ma, Y., Zhang, A., Xie, H., Tao, D.: Perceptual-sensitive gan for generating adversarial patches. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1028-1035 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.677 + ], + "angle": 0, + "content": "41. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.704 + ], + "angle": 0, + "content": "42. Liu, X., Yang, H., Liu, Z., Song, L., Li, H., Chen, Y.: Dpatch: An adversarial patch attack on object detectors. arXiv preprint arXiv:1806.02299 (2018) 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.759 + ], + "angle": 0, + "content": "43. Lovisotto, G., Turner, H., Sluganovic, I., Strohmeier, M., Martinovic, I.: {SLAP}: Improving physical adversarial examples with {Short-Lived} adversarial perturbations. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1865-1882 (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "44. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "45. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 7" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "T. Lin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "47. Pomponi, J., Scardapane, S., Uncini, A.: Pixle: a fast and effective black-box attack based on rearranging pixels. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "48. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "49. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.357 + ], + "angle": 0, + "content": "50. Shapira, A., Zolfi, A., Demetrio, L., Biggio, B., Shabtai, A.: Phantom sponges: Exploiting non-maximum suppression to attack deep object detectors. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4571-4580 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.411 + ], + "angle": 0, + "content": "51. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In: Proceedings of the 2016 acm sig sac conference on computer and communications security. pp. 1528-1540 (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "52. Shi, Z., Yang, W., Xu, Z., Yu, Z., Huang, L.: Reinforcement learning-based adversarial attacks on object detectors using reward shaping. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8424-8432 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "53. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.548 + ], + "angle": 0, + "content": "54. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.605 + ], + "angle": 0, + "content": "55. Tu, J., Ren, M., Manivasagam, S., Liang, M., Yang, B., Du, R., Cheng, F., Urtasun, R.: Physically realizable adversarial examples for lidar object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13716-13725 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.661 + ], + "angle": 0, + "content": "56. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2414-2422 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "57. Wang, D., Yao, W., Jiang, T., Li, C., Chen, X.: Rfla: A stealthy reflected light adversarial attack in the physical world. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4455-4465 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.744 + ], + "angle": 0, + "content": "58. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "59. Wang, J., Liu, A., Yin, Z., Liu, S., Tang, S., Liu, X.: Dual attention suppression attack: Generate adversarial camouflage in physical world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8565-8574 (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "60. Wang, Q., Zhang, L., Bertinetto, L., Hu, W., Torr, P.H.: Fast online object tracking and segmentation: A unifying approach. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 1328-1338 (2019) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.528, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Out-of-Bounding-Box Triggers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "61. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "62. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.286 + ], + "angle": 0, + "content": "63. Wu, W., Su, Y., Chen, X., Zhao, S., King, I., Lyu, M.R., Tai, Y.W.: Boosting the transferability of adversarial samples via attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1161-1170 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "64. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.398 + ], + "angle": 0, + "content": "65. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 665-681. Springer (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.453 + ], + "angle": 0, + "content": "66. Zhang, J., Wu, W., Huang, J.t., Huang, Y., Wang, W., Su, Y., Lyu, M.R.: Improving adversarial transferability via neuron attribution-based attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14993-15002 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.493 + ], + "angle": 0, + "content": "67. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15115-15125 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.55 + ], + "angle": 0, + "content": "68. Zhang, L., Xu, N., Yang, P., Jin, G., Huang, C.C., Zhang, L.: Trajpac: Towards robustness verification of pedestrian trajectory prediction models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8327-8339 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.605 + ], + "angle": 0, + "content": "69. Zhao, Y., Zhu, H., Liang, R., Shen, Q., Zhang, S., Chen, K.: Seeing isn't believing: Towards more robust adversarial attack against real world object detectors. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 1989-2004 (2019) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.66 + ], + "angle": 0, + "content": "70. Zhong, Y., Liu, X., Zhai, D., Jiang, J., Ji, X.: Shadows can be dangerous: Stealthy and effective physical-world adversarial attack by natural phenomenon. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15345-15354 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "71. Zhu, L., Wang, T., Li, J., Zhang, Z., Shen, J., Wang, X.: Efficient query-based black-box attack against cross-modal hashing retrieval. ACM Transactions on Information Systems 41(3), 1-25 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "72. Zhu, X., Li, X., Li, J., Wang, Z., Hu, X.: Fooling thermal infrared pedestrian detectors in real world using small bulbs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3616-3624 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.798 + ], + "angle": 0, + "content": "73. Zolfi, A., Kravchik, M., Elovici, Y., Shabtai, A.: The translucent patch: A physical and universal attack on object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15232-15241 (2021) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.798 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_origin.pdf b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..838415d65cddbb2862787ff8730bf5c6510cc355 --- /dev/null +++ b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/e1f02941-0449-4cca-83a5-ef88a2d74a7d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff38f211f6b177c2125b0d5ddbcfab96d27b54675a4a2958fac6deb7fe551bcc +size 8281770 diff --git a/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/full.md b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd07b28ae415021deadfe18b54a299a1ca7c28a6 --- /dev/null +++ b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/full.md @@ -0,0 +1,440 @@ +# Out-of-Bounding-Box Triggers: A Stealthy Approach to Cheat Object Detectors + +Tao Lin $^{1,2}$ , Lijia Yu $^{1}$ , Gaojie Jin $^{1,2}$ , Renjue Li $^{1,2}$ , Peng Wu $^{1,2,\text{图}}$ , and Lijun Zhang $^{1,2,3,\text{图}}$ + +1 Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China + +$^{2}$ University of Chinese Academy of Sciences, Beijing, China +3 Automotive Software Innovation Center, Chongqing, China {lintao,yulijia,gaojie,lirj19,wp,zhanglj}@ios.ac.cn + +Abstract. In recent years, the study of adversarial robustness in object detection systems, particularly those based on deep neural networks (DNNs), has become a pivotal area of research. Traditional physical attacks targeting object detectors, such as adversarial patches and texture manipulations, directly manipulate the surface of the object. While these methods are effective, their overt manipulation of objects may draw attention in real-world applications. To address this, this paper introduces a more subtle approach: an inconspicuous adversarial trigger that operates outside the bounding boxes, rendering the object undetectable to the model. We further enhance this approach by proposing the Feature Guidance (FG) technique and the Universal AutoPGD (UAPGD) optimization strategy for crafting high-quality triggers. The effectiveness of our method is validated through extensive empirical testing, demonstrating its high performance in both digital and physical environments. The code and video will be available at: https://github.com/linToTao/Out-of-bbox-attack. + +Keywords: Universal physical adversarial attack $\cdot$ Out-of-bounding-box trigger $\cdot$ Object detection + +# 1 Introduction + +Object detection aims to identify instances of semantic objects typically within images or video clips. It finds widespread applications across a range of domains, including face recognition [10, 11, 23, 49, 58], object tracking [34, 35, 60], and autonomous driving [6, 18, 68]. Especially in autonomous driving systems, object detectors play a crucial role in tasks such as recognizing traffic signs, pedestrians, vehicles, and traffic lights. Nonetheless, in recent years, security concerns have emerged regarding object detectors, primarily due to the susceptibility of deep neural networks (DNNs) to adversarial examples (AEs). AEs are intentionally crafted inputs designed to deceive DNNs into making incorrect predictions. + +![](images/d27a9aa680fd619b1b3f75df85313255adc96627a4aca5e08809f66f56ca564a.jpg) +Fig. 1: An illustration of vanilla object attack methods and our novel approach. Typically, vanilla attack methods target the most informative regions of an object, such as the central, highly visible areas of the target object, e.g., the stop sign in (a) [16], (b) [69], (c) [15], and the car in (d) [19], (e) [59], (f) [22]. In contrast, our method adopts a more covert strategy, focusing on attacking the peripheral edge areas surrounding the stop sign. + +Early research mainly focused on studying AEs like adversarial patch [16, 20, 33, 40, 73], optical [43, 70] and texture [15, 21], which are targeted at the objects. These AEs involve the computation of perturbations applied directly to the objects to mislead a DNN-based detector. However, a direct attack on the object as shown in Figure 1, along with its subsequent attempt to deceive the detector, is overly conspicuous and cannot be deployed without raising suspicion in real-world practical scenarios. + +In this work, we provide an alternative perspective: is it possible to introduce an adversarial trigger in an inconspicuous area outside the object while still successfully deceiving the detector? + +Actually, crafting an adversarial trigger within the areas beyond the object boundaries is even more demanding. On one hand, for dense prediction networks like object detectors, the output results are more reliant on the red regions of the heatmap in Figure 1. This implies that tailoring the adversarial trigger by utilizing irrelevant information, like the blue regions of the heatmap, is sharply challenging. More difficultly, our trigger is designed to conform to practical size and shape constraints. Nevertheless, for exactly the same reasons, our attack method gets more delusive and harder to defense. + +This paper is focused on the development of robust and inconspicuous triggers to attack typical object detectors deployed in practical scenarios. To tackle the challenge mentioned above, we present a novel technique called Feature Guidance (FG). In this approach, we guide the feature layers with an adversarial trigger to resemble those with the object hidden. And we propose a optimization strategy, called Universal Auto-PGD (UAPGD), applicable to universal attacks. This optimization strategy allows for the selection of an appropriate timing to halve the step size, based on the oscillation of the loss function. + +To assess the effectiveness of our approach in both digital and physical domains, we trained the adversarial trigger and evaluated the corresponding metrics on COCO [39] dataset and Carla [14], respectively. We also evaluated the robustness of the trigger by recording videos using an onboard camera in the real world. + +Overall, this paper makes the following contributions: + +- This work studies the realm of adversarial robustness in object detection, providing a novel perspective by exploring the concept of an adversarial trigger applied outside an object. +- We propose a novel attack method designed to generate triggers that achieve excellent continuous attack effects and demonstrate high robustness in real world. The key enabler of our method is the utilization of similarity between feature layers and adaptive step size decay to endow the trigger with the ability to mislead the detector. +- An extensive set of empirical results are provided to demonstrate that our attack exhibits strong performance and robustness in both the COCO dataset and Carla simulator, as well as in real-world scenarios. + +# 2 Related work + +Adversarial examples (AEs), first discovered by Szegedy et al. [53], involve carefully crafted inputs inducing model misclassifications. Transferability [26, 61, 64, 66] of AEs, allows them to deceive networks with different hyperparameters or architectures. + +Digital adversarial examples are generated by directly manipulating any pixels in digital images prior to feeding them into the model. Adversarial attacks can be classified into black-box attacks and white-box attacks based on whether the internal structure and parameters of the target model are known. White-box attacks [5, 12, 17, 27, 28, 31, 44, 46] primarily rely on the computation of model gradients. Black-box methods, on the other hand, depend on querying the unknown model by observing the prediction [12, 13, 38, 47, 62–64, 71] or leveraging an agent model to imitate the threaten model [3, 7, 8, 24, 30, 67]. Recently, Croce et al. [9] proposed AutoAttack, a parameter-free method that combines novel white-box and black-box attack strategies. The works of applying digital adversarial examples to object detectors have been explored in [25] and [50] that introduce perturbations to input images to mislead detectors. [52] generate more subtle perturbations via reinforcement learning, successfully attacking the detector. + +Physical adversarial examples are crafted to be printed out and captured by sensory equipment like camera [31,36] and LiDAR [4,41,55]. These examples pose a real-world threat due to their ability to adapt to camera and scenario variations. Early studies focused on attacking recognition models [16,51]. Recent research has explored physical attacks in the more challenging context of object detection. Perturbations in this domain include adversarial patches [1,20,54,65, 69], adversarial camouflage textures [21,22,56,59], and optical attacks [37,43,57, 72]. To improve robustness against various transformations in physical settings, + +Athalye et al. [2] introduced Expectation Over Transformation (EOT), widely used in physical attacks, to enhance the effectiveness of adversarial attacks. + +Prior physical attack methods need to be applied directly to the objects so that these methods are expected to yield favorable results but they are prone to raising suspicion in practical scenarios. Unlike these works, our method constructs a universal attack trigger that can deceive the detector disregarding the specific object without altering the appearance of itself, and can be discreetly deployed in the real world. + +The works that exhibit the closest alignment with our motivation are Dpatch [42] and [32] (after this referred to as Dpatch2). However, we must emphasize that our work is independent of them, highlighting the difference between our work and theirs in Appendix F. Furthermore, in the experiments section, we compare the attack effectiveness of our method with Dpatch2 [32] on the COCO dataset. + +# 3 Methodology + +In this work, we aim to generate physical adversarial triggers that can be naturally deployed in an inconsequential position. These adversarial triggers are designed not to destroy the object information but still deceive the model effectively. In this section, we commence by presenting the problem formulation in Section 3.1. Subsequently, Fig. 2a provides a comprehensive illustration of our framework. Our primary focus is on the white-box attack paradigm, which necessitates an intricate optimization process. Given an object detection model, we design a loss function to determine the gradient direction for adversarial attacks in Section 3.2. This loss function is then utilized iteratively to update the adversarial trigger. In the pursuit of strengthening attack performance and enhancing the continuity of attacks in video-based scenarios, we introduce a novel method named "Feature Guidance" (FG) and an additional loss term $L_{FG}$ in Section 3.2. To bolster the overall robustness and deployability of our approach, we also employ two indispensable techniques, EOT and total variation loss in Section 3.2. Furthermore, our methodology draws inspiration from the work of [9] in their Auto PGD framework during the training process, we thus modify the conditions of step size decay based on the universal training mode (Section 3.3). + +# 3.1 Problem Formulation + +Adversarial attacks on object detection can have diverse objectives, encompassing but not limited to inducing misalignment in bounding boxes, reducing classification confidence, or provoking misclassification. Our goal is to generate adversarial examples for a specific class, with the aim of causing the object detection algorithm to fail in identifying objects belonging to that class. + +Given an input image $x \in \mathbb{R}^d$ with the ground truth $y = \{(y_i^{\mathrm{coor}}, y_i^{\mathrm{cls}})\}_{i \in [n]} \subset \mathbb{R}_+^4 \times \mathbb{Z}_+$ , $n$ is the number of objects in the image, $y_i^{\mathrm{coor}}$ contains top-left and bottom-right coordinates of the bounding box in the image, and $y_i^{\mathrm{cls}}$ is the class of + +![](images/b6a78b26b4e945d1d75311757f6bbaed5ebcfc963d7903bd4d7028b09970f320.jpg) +Fig.2: (a) shows the overview of our adversarial trigger generation framework. We deploy the adversarial trigger onto each image in the dataset with robustness transformation. With a fixed detector, our method optimizes the adversarial trigger through the obtained adversarial gradient. (b) illustrates the motivation and calculation method of Feature Guidance. Taking a stop sign as an example, for a given image $x$ , minimizing $\| f_z(T(x,\mathcal{A}_t)) - f_z(x')\|_2$ can mislead a detector to classify the stop sign image $T(x,\mathcal{A}_t)$ as none stop. + +![](images/040c7978640aa6ffc12a9b83a1f8b8b143351261c8a3c6cb9427c04db7f9f4b4.jpg) + +object. Then, the DNN-based object detector can be represented as the following form: + +$$ +f: \mathbb {R} ^ {d} \to 2 ^ {\mathbb {R} _ {+} ^ {4} \times \mathbb {Z} _ {+} \times [ 0, 1 ] ^ {2}} +$$ + +For a given image $x$ , the output of $f$ can be represented as $\tilde{y} = \{(\tilde{y}_i^{\mathrm{coor}},\tilde{y}_i^{\mathrm{cls}},\tilde{y}_i^{\mathrm{conf}})\}_{i\in [n]} \subset \mathbb{R}_+^4\times \mathbb{Z}_+ \times [0,1]^2$ , where $\tilde{y}_i^{\mathrm{conf}}$ denotes the confidence associated with the predicted box coordinates $\tilde{y}_i^{\mathrm{coor}}$ and the predicted class $\tilde{y}_i^{\mathrm{cls}}$ for $i$ -th object. We aim to generate a universal adversarial trigger $\mathcal{A}_t\in [0,1]^{3\times U\times V}$ , where $U,V\in \mathbb{Z}_{+}$ are the given dimensions, for the given specific class $t$ , such that these adversarial triggers can cause $f$ to fail in recognizing objects of class $t$ . Then, the problem can be formulated as an optimization task, represented as follows: + +$$ +\operatorname {a r g m i n} _ {\mathcal {A} _ {t}} \mathbb {E} _ {x \sim D} [ | f (T (x, \mathcal {A} _ {t})) \cap (\mathbb {R} _ {+} ^ {4} \times \{t \} \times [ 0, 1 ] ^ {2}) | ] \tag {1} +$$ + +where $|A|$ denotes the cardinality of the set $A$ , $D$ represents the input distribution, and $T$ (which will be elaborated in the following) is a transform operation to incorporate the adversarial trigger into the original image. In contrast to prior studies based on adversarial patches, our approach avoids placing the adversarial trigger on the informative surface, e.g., traffic signs, thereby preventing any unnecessary attention. We would like to point out that this kind of attack method is challenging because the adversarial trigger can only be stuck on the unimportant places of the image. + +Transform operation. We utilize a differentiable operation $T$ to render the adversarial trigger $\mathcal{A}_t$ onto an image $x$ . Let the object with class $t$ be located in the box $B$ of the image, we use the following two-step to add the adversarial trigger to the image: + +(1) We can select an appropriate area $C(B)$ in the image where $C$ is a given function, such position does not overlap with the surface of the object; +(2) We cover the box $C(B)$ part of the image $x$ with adversarial trigger $\mathcal{A}_t$ , but + +Considering that the adversarial trigger and the box are not the same size, we need to apply the following affine matrix to $\mathcal{A}_t$ : + +$$ +\theta = \left( \begin{array}{c c c} S _ {h} \cos \alpha & - S _ {h} \sin \alpha & T _ {h} \\ S _ {v} \sin \alpha & S _ {v} \cos \alpha & T _ {v} \\ 0 & 0 & 1 \end{array} \right) \tag {2} +$$ + +$S_{h}$ and $S_{v}$ denotes the scaling factors utilized to resize the adversarial trigger, $\alpha$ represents the angle of rotation, $T_{h}$ and $T_{v}$ are the distance adversarial trigger need to shift. These parameters are all computed by $C(B)$ . The more specific remark regarding $C(B)$ will be detailed in Appendix B. + +Note that our attack method has two characteristics: Our attack is out of the object; our attack area changes depending on the position of the object. We were the first to attack target detection with such a method. + +# 3.2 Adversarial gradient with Feature Guidance + +In this paper, we focus on utilizing white-box attack methods to search for adversarial gradients and optimize the adversarial trigger. Additionally, we provide black-box empirical results in Appendix D that the trigger is generated by a surrogate model. As illustrated above, we render the adversarial trigger onto images and feed them to the detectors, take the derivative of the loss function, and use it to update the adversarial trigger. The loss function consists of the following parts: + +Adversarial loss of detector output. As mentioned above, the detector outputs the predicted class and bounding boxes and their confidence. Considering most of the detectors will not output the predicted class and boxes with a confidence lower than a threshold, we want images with the adversarial trigger to make the confidence as low as possible. + +Let $f_{i,coor}^{\mathrm{conf}}(x)$ and $f_{i,cls}^{\mathrm{conf}}(x)$ is the confidence of the $i$ -th object on image $x$ , we can define the first part of our loss function as follows: + +$$ +L _ {d e t} \left(\mathcal {A} _ {t}\right) = \mathbb {E} _ {x \sim D} \left[ \max _ {i: \tilde {y} _ {i} ^ {c l s} = t} f _ {i, c o o r} ^ {\text {c o n f}} (T (x, \mathcal {A} _ {t})) f _ {i, c l s} ^ {\text {c o n f}} (T (x, \mathcal {A} _ {t})) \right] \tag {3} +$$ + +Feature guidance. The biggest difficulty with an attack outside the object is that the features of the original object on the image will not be destroyed, so the network will still notice the presence of the object. + +We gain the following insights from the network structure to address this issue. A DNN-based detector is usually designed with three modules: Backbone, Neck, and Head. The Backbone module extracts global features from an input image. These features are then fused through the Neck module to produce multiple feature maps with varying receptive fields. The Head module subsequently makes predictions based on the feature maps. This modular structure imitates how people often perceive a comprehensive horizon (like the Backbone module) and focus on varying attractive areas (like the Neck module). Subsequently, the attractive areas will contribute to making perceptions (like the Head module). + +Hence, in order to provoke model errors, it is imperative to deliberately induce information misleading prior to inputting the head module. + +To achieve that, we use the following method, called Feature guidance, as shown in Fig. 2b. For a given image $x$ with a target object $t$ , we can get the image $x'$ , which differs from $x$ only with the object $t$ missing. Then, if we can elaborate our adversarial trigger $\mathcal{A}_t$ to guide the feature map $f_z(T(x, \mathcal{A}_t))$ as close to $f_z(x')$ generated by the Neck module of the detector, it will have a desirable impact on the Head module such that the stop sign cannot be detected. + +As illustrated in Figure 2a, for the sake of convenience, we derive $x_{mask}$ that exhibits an analogous effect to $x'$ simply by covering the bounding box position in the ground truth of $x$ with a gray rectangle. + +In order to achieve Feature guidance, we need to make $\mathcal{A}_t$ to minimum such loss function: + +$$ +L _ {F G} (\mathcal {A} _ {t}) = E _ {x \sim D} [ \| f _ {z} (T (x, \mathcal {A} _ {t})) - f _ {z} (x _ {\text {m a s k}}) \| _ {2} ] +$$ + +Other skills To enhance the robustness of our adversarial trigger $\mathcal{A}_t$ , we incorporate diverse variations, including random noise, contrast, brightness, and rotation, into the generated trigger prior to its application on images, thus simulating real-world conditions. To address printing challenges, we employ TV loss [45] on the adversarial trigger to promote smoothness. TV loss is defined as + +$$ +L _ {t v} \left(\mathcal {A} _ {t}\right) = \sum_ {i, j} \sqrt {\left(\mathcal {A} _ {t} ^ {i , j} - \mathcal {A} _ {t} ^ {i + 1 , j}\right) ^ {2} + \left(\mathcal {A} _ {t} ^ {i , j} - \mathcal {A} _ {t} ^ {i , j + 1}\right) ^ {2}} \tag {4} +$$ + +where the subindices $i$ and $j$ represent the pixel coordinate of the trigger $\mathcal{A}_t$ + +Overall, we will get the adversarial trigger $\mathcal{A}_t$ by optimizing following problem: + +$$ +\min _ {\mathcal {A} _ {t} \in [ 0, 1 ] ^ {3 \times U \times V}} L _ {d e t} (\mathcal {A} _ {t}) + \lambda_ {F G} L _ {F G} (\mathcal {A} _ {t}) + \lambda_ {t v} L _ {t v} (\mathcal {A} _ {t}) \tag {5} +$$ + +where $\lambda_{FG}$ and $\lambda_{tv}$ are hyperparameter. + +# 3.3 Universal Auto-PGD + +Let $L_{all} = L_{det} + \lambda_{FG}L_{FG} + \lambda_{tv}L_{tv}$ , we will use gradient descent to find the adversarial trigger via minimizing the loss function $L_{all}$ . + +Considering the instability that appeared in the experiment when optimizing multiple terms of $L_{all}$ with conventional methods, we propose the following optimization strategy called Universal Auto-PGD (UAPGD): + +At the beginning of optimizing, we fix the following values: comparison length $l_{c}$ , observation length $l_{o}$ , the slack variable $\epsilon_{1}, \epsilon_{2}$ , let the initial step size be $\eta$ and we record the following values: + +(1): After epoch $j$ , record the value of the loss function, name it $L_{j}$ ; +(2): After epoch $i * l_{o}$ where $i \in Z_{+}$ , record: + +$$ +\begin{array}{l} L _ {\min , i} = \min _ {j \in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \left\{L _ {j} \right\} \\ V _ {i} = \operatorname {V a r i a n c e} \left[ \left\{L _ {j} \right\} _ {j \in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \right] \\ \end{array} +$$ + +Algorithm 1: UAPGD +Input: Threaten network $f$ , initial step size $\eta$ , the number of epoch $N_{\mathrm{epoch}}$ dataset $D = \cup_{j = 1}^{k}D_{j}$ which was divided into k batches, the hyperparameter in UAPGD $l_c,l_o$ , initial adversarial trigger $\mathcal{A}_t^0$ Output: Adversarial trigger $\mathcal{A}_{best}$ +1 $S_{c}\gets \phi$ $S_{o}\gets \phi$ $L_{best}\gets \infty$ $\mathcal{A}_{best}\leftarrow \mathcal{A}_t^0$ +2 for $i\in [N_{\mathrm{epoch}}]$ do +3 $L_{i}\gets 0$ +4 $\mathcal{A}_{t}^{i,0}\gets \mathcal{A}_{t}^{i - 1}$ +5 for $j\in [k]$ do +6 Get $L_{all}(\mathcal{A}_t^{i,j - 1},D_j)$ via $f$ +7 $L_{i}\gets L_{i} + L_{all}(\mathcal{A}_{t}^{i,j - 1},D_{j}) / k$ +8 $\mathcal{A}_{t}^{i,j}\gets \mathcal{A}_{t}^{i,j - 1} - \eta \cdot \mathrm{sign}(\frac{\nabla L_{all}(\mathcal{A}_{t}^{i,j - 1},D_{j})}{\nabla\mathcal{A}_{t}^{i,j - 1}})$ +9 $\mathcal{A}_{t}^{i,j}\gets Clamp(\mathcal{A}_{t}^{i,j},0,1)$ +10 $\mathcal{A}_{t}^{i}\gets \mathcal{A}_{t}^{i,k}$ +11 if $L_{i} < L_{best}$ then +12 $\begin{array}{r}\big{\lfloor}\mathcal{A}_{best}\leftarrow \mathcal{A}_{t}^{i}\text{and} L_{best}\leftarrow L_{i} \end{array}$ +13 $S_{o} = S_{o}\cup L_{i}$ +14 if $|S_{o}| = l_{o}$ then +15 $\begin{array}{rl} & L_{min}\gets \min_{L\in S_o\backslash \{S_o\}}\\ & V\gets \operatorname{Var}_{L\in S_o\backslash \{S_o\}}\\ & S_c = S_c\cup (L_{min},V)\\ & S_o\gets \phi \end{array}$ +16 +17 +18 +19 if $|S_{c}| = l_{c}$ then // Condition 1 and Condition 2 are introduced in section 3.3 if Condition 1(Sc) and Condition 2(Sc) then +21 $\begin{array}{r}\lfloor \eta \leftarrow \eta /2\text{and}\mathcal{A}_t^i\leftarrow \mathcal{A}_{best}\\ S_c\leftarrow \phi \end{array}$ +22 + +In order to prevent violent oscillations around the local optimum during the optimization process, we will halve the step size when the following conditions are met: + +(Con 1): $L_{min,j} \geq L_{min,q} - \epsilon_1$ for all $(j - l_c) < q < j$ ; +(Con 2): $V_{j}\geq V_{q} - \epsilon_{2}$ for all $j - l_{c} < q < j$ + +Due to the stochasticity from the robustness transformation mentioned in Other skills, the value of the loss function may fluctuate so that we incorporate the slack variable $\epsilon_{1},\epsilon_{2}$ into numerical comparison. + +Intuitively, these two conditions respectively describe the exploration and exploitation aspects in the optimization process. Con 1 means that the value of the loss function has not decreased much, indicating the necessity of a transition from exploring the entire feasible space to focusing on local optimization. Con + +2 characterizes the oscillation of the loss function, of which the amplitude per $l_{c}$ steps has not decreased much. This implies the imminent conclusion of the algorithm's exploitation in a local space. The simultaneous occurrence of c1 and c2 means that the adversarial trigger jumps near one of the optimal points, but the step size is too large to fall steadily near the best advantages. Therefore, at this point, we will halve the step size. + +Our approach is somewhat similar to AutoAttack [9], but there are several differences: (1): Our method is for universal attack, AutoAttack is for attacking a single image; (2): We consider the amplitude of the oscillation during the optimization process and use this as an indicator of decline; (3): We include the slack variables in order to mitigate potential misleading phenomena caused by experimental randomness. For example, the oscillation of the loss function often easily cause $L_{j}$ to show a downward trend. + +The overall process of our algorithm is shown in algorithm 1. And we use $L_{all}(\mathcal{A}_t, D_j)$ to represent calculating $L_{all}$ on an input batch $D_j$ . + +# 4 Experiment + +In this section, we choose the stop sign as the target object to evaluate the effectiveness of our adversarial attack method. The stop sign is a commonly used object in the evaluation of physical adversarial attacks, as it is critical to traffic safety. We begin by detailing the implementation settings of our evaluation. This is followed by an extensive set of experiments, utilizing the proposed adversarial trigger in the digital space, specifically on the COCO dataset [39]. Furthermore, we extend the application of our method to real-world scenarios, i.e., we record videos to demonstrate that our method can attack the object detector in both simulators and real-world scenes. Meanwhile, we evaluate our attack method across varying distances and angles. See Appendix A and B for the basic experimental setting, the results of additional attacks on other objects, and the detail of selecting C(B) to place the trigger on the stop sign. + +# 4.1 Digital experiment + +In this subsection, our experiments are carried out using the COCO dataset [39]. From this dataset, we extract a total of 1100 images featuring stop signs. Of these, 600 images constitute the training set, which is utilized to generate the adversarial trigger. The remaining 500 images are designated as the test set, serving to evaluate the effectiveness of the generated adversarial trigger. + +To demonstrate the superiority of our algorithm, we adopt the following methods to generate adversarial triggers: optimizing the adversarial trigger using PGD [44] or UAPGD with a minimization of the loss function $L_{det} + \lambda_{tv}L_{tv}$ or $L_{all}$ . This included comparisons between our method and the normal PGD under the condition of placing the perturbation outside the object, as well as ablation experiments with varying hyperparameters. Specifically, for the YOLOv3(YOLOv5) model, we choose $\lambda_{FG}$ values of 0.1, 0.5, and 1.0(0.05, + +0.1, and 0.2). The optimization process utilized the initial step size of 8/255 and 16/255. We calculate the number of undetected images among all the images as a metric called attack success rate(ASR). + +
Threaten ModelYOLOv3YOLOv5
DatasetTrain(%)Test(%)Train(%)Test(%)
Dpatch226.6724.007.176.67
Dpatch2 (Ldet)35.5335.8032.8332.00
Adam (16/255)65.1662.2043.8340.40
Adam (8/255)64.3360.8052.1650.80
PGD(16/255)56.6754.0043.6544.60
PGD(8/255)65.1762.2053.8351.00
UAPGD(16/255)66.0063.0057.8352.60
UAPGD(8/255)67.8364.8054.1750.40
FG(λF, FG, 16/255)57.6755.4051.8348.20
FG(λF, FG, 16/255)56.0054.6047.8344.60
FG(λF, FG, 16/255)54.0051.0043.8340.00
FG(λF, FG, 8/255)60.0055.2059.1755.60
FG(λF, FG, 8/255)64.8362.6052.1750.60
FG(λF, FG, 8/255)61.6759.8056.6756.40
Ensemble(λF, FG, 16/255)67.8366.2058.3356.40
Ensemble(λF, FG, 16/255)67.8364.8060.1757.60
Ensemble(λF, FG, 16/255)64.8361.6058.0056.60
Ensemble(λF, FG, 8/255)68.3365.0062.6758.60
Ensemble(λF, FG, 8/255)69.3366.6056.6753.20
Ensemble(λF, FG, 8/255)66.1662.2060.1757.80
+ +![](images/e500f78b3fb3b0e15738097e9d4b2a293b6a94619633331edbc4f41b2060a1fb.jpg) +PGD + +![](images/eff3da90302496f2f5287b95ee4853e25e7c8e65d25fed96b2d7ca11422f39c0.jpg) +UAPGD + +![](images/1183652b95f0fa74a19b6e9a11fd4eff6384308eb44440906d6254e24f780695.jpg) +FG +Fig. 3: The table on the left displays the ASR values of triggers, which are generated by various settings, on the training and testing sets of the COCO dataset. The values in parentheses represent the attack step size, and we employ different $\lambda_{FG}^{i}$ values for YOLOv3 and YOLOv5. For YOLOv3, we set $\lambda_{FG}^{1} = 0.1$ , $\lambda_{FG}^{2} = 0.5$ , $\lambda_{FG}^{3} = 1.0$ ; for YOLOv5, we set $\lambda_{FG}^{1} = 0.05$ , $\lambda_{FG}^{2} = 0.1$ , $\lambda_{FG}^{3} = 0.2$ . Ensemble denotes the combined usage of FG and UAPGD. Dpatch2 is the original approach from [32] and Dpatch2( $L_{det}$ ) denotes the original approach adopts the $l_{det}$ as the optimization objective. The bolded values indicate the optimal results for each respective column in the table. There are some results of triggers on the right generated under different settings. + +![](images/72a324c1c3af2b29d9de062c92e1853f662fd94512d7fdfcf55c36d481b79427.jpg) +Ensemble + +As shown in Figure 3, We compare our method with the more reasonable and advanced Dpatch2 [32]. Considering the divergent optimization objectives, we substitute the maximization of the model loss function mentioned in [32] with the minimization of the proposed $L_{det}$ referred to as Dpatch2( $L_{det}$ ). Undoubtedly, our method achieves the highest ASR in different experimental settings. + +A comparative analysis of the effectiveness between PGD and UAPGD reveals notable differences. Specifically, we observe that UAPGD, when employed with an identical attack step size as PGD, frequently results in a higher ASR. Moreover, it can be observed that UAPGD yields better performance compared to adaptive step size optimizers like Adam. + +Figure 4a and Figure 4b present a detailed depiction of the variations in the detection loss $L_{det}$ throughout the training process when generating adversarial triggers using both PGD and UAPGD techniques, specifically over YOLOv3 and YOLOv5 models. Notably, when operating under an equivalent number of epochs, UAPGD demonstrates a more pronounced capability to reduce $L_{det}$ + +![](images/216198102cf21938cf28389a83992fa67fac2aa58b3dce6a9b8db735f7b00ce6.jpg) +(a) YOLOv3 + +![](images/c08f344bdf0c70c5ec8686666c02d292668611c99437150d75fd5b2f1305ec09.jpg) +(b) YOLOv5 +Fig. 4: (a) and (b) illustrate the variations of $L_{det}$ , during the training process of generating adversarial triggers using PGD and UAPGD for attacking the YOLOv3 and YOLOv5 models. (c) and (d) present the confidence of some samples from the COCO dataset, outputted by the detector, during the training process of individual attacks. The blue, green, and red curved lines represent the utilization of $L_{FG}$ , $L_{det}$ , and both as the loss function respectively. For individual attacks, the optimization process is relatively easier so that step size decay is unnecessary. + +![](images/b3b0d2844550071cd4739d33a5a239e6198c81c6f3fcc7113b7a1a95a1bb235b.jpg) +(c) + +![](images/fc29a5f10440eede6d74adbed270de542c47a7e05ec61406d34ee959305ee6e1.jpg) +(d) + +compared to PGD. The effectiveness of the loss monitoring in UAPGD will be analyzed in detail in Appendix E. + +Furthermore, as shown in Figure 3, UAPGD exhibits a superior capability in generating smoother adversarial triggers. This increased smoothness can be attributed to the implementation of step size decay, which allows for a more refined optimization process in the generation of triggers. Consequently, adversarial triggers crafted via UAPGD display enhanced suitability for practical applications, such as printing and deployment in the physical world. + +Based on results in Figure 3, we can observe that integrating $L_{FG}$ with suitable step size and weight into the loss function, coupled with using either PGD or UAPGD, can enhance the ASR. This improvement is particularly noticeable when applied to the YOLOv5 model. This observation highlights that employing appropriate feature guidance can improve the performance of universal attacks. However, incorporating FG did not always achieve satisfactory results under certain weights and step sizes. This phenomenon is because in universal attacks, the inclusion of $L_{FG}$ introduces additional complexities to the optimization process. It requires appropriate adjustment of hyperparameters to fully leverage its effectiveness. Hence, the optimal outcome can be achieved by ensemble UAPGD with it, which allows for finer control over the step size to facilitate the optimization process. The empirical results indicate that the combination of UAPGD and FG in generating triggers yields the most effective attack performance. + +To further investigate the potential role of Feature Guidance, we conducted individual adversarial attacks on each image with an adversarial trigger (fixed step size of 8/255, 500iteration) in the training dataset. We identified two primary roles of FG: 1, as shown in Figure 4c, our observations reveal that FG can significantly influence the optimization direction during the training phase. This influence is evidenced by the expedited convergence in adversarial trigger training, indicating the advantageous role of FG in enhancing training efficiency. 2, As shown in Figure 4d, when some images are challenging to attack with out-of-box adversarial triggers, introducing FG would enhance the attack effectiveness at this time. + +# 4.2 Experiment in simulator + +In this section, we demonstrate the effects of our attack in autonomous driving scenarios, we use the Carla(0.9.14) simulator [14] to simulate the real driving process. In the simulator, we utilize the YOLOv3 model as the detection network. + +Due to the significant differences between images from the COCO dataset and those generated by the Carla simulator, the adversarial trigger developed in Section 4.1 is not directly applicable to Carla. To tailor an adversarial trigger specifically for the Carla simulator, we compile a training dataset comprising 242 images featuring stop signs from the simulator. The process of attacking the Carla simulator + +is conducted in four distinct manners: employing either PGD or UAPGD, combined with a loss function of either $L_{det} + \lambda_{tv}L_{tv}$ or $L_{all}$ , where $\lambda_{FG} = 0.1$ and the initial step size is 16/255. + +To evaluate the effectiveness of the attack, we implement a procedure where the adversarial trigger is placed under the stop sign in the simulator. Subsequently, a vehicle is navigated towards this stop sign, and a 9-second video is recorded using an onboard camera. The key metric is the percentage of time during which the detection system fails to recognize the stop sign. + +As shown in Table 1, without any adversarial trigger, the car consistently recognizes the stop sign. When the attack is initiated using the PGD method alone, the attack duration is approximately 3 to 4 seconds. However, the integration of either UAPGD or FG methods into the attack strategy results in a slight increase in the attack duration by roughly an additional second. Notably, the implementation of an ensemble approach, which combines + +these methods, further enhances the attack performance, prolonging the duration to over 5 seconds. + +Moreover, to provide a more direct illustration of the attack impact, we have plotted the trend of the stop sign detection confidence as the car approaches it, shown as in Figure 5. + +The empirical results reveal that the efficacy of our attack method consistently surpasses that of the vanilla PGD approach. Particularly in the final 5 + +
attack methodProportion
No attack0%
PGD36.67%
UAPGD45.56%
PGD,FG46.67%
UAPGD,FG58.89%
+ +Table 1: The proportion of time in the video duration during which the detector failed to recognize the stop sign. + +![](images/1573d79a60744cdd7bd24480190fc083636faa2152ec8561a176d560f8381904.jpg) +Fig.5: We uniformly extracted approximately 90 frames from the 9s video, which is recorded by the simulated vehicle's onboard camera, to showcase the confidence variations of the detector's predictions + +![](images/b8e45a5dd8cba39c41d276472d2bc1939fb261b8a1afc662ea6fdb21dbf8a847.jpg) + +![](images/020d5516646b7171a2f025ba9ad8498d13860622eee940bafb812b165d50dc88.jpg) + +![](images/fb46028851858f4f27fb21816c751721a2a879e2c011fd88dc4a393528d58713.jpg) + +![](images/80a91c18a7098f6271510abfd1fa494a7ff388a763d416c0b350cc715b965be0.jpg) + +![](images/7ce192eb37439234d236016e572a1cd17a104cec4ad0de0ac13780167dd3ab21.jpg) + +![](images/d505e6ab38c8190e4c4f9fcfded3417de960d5ed23cfa3b2c09c09fc92656142.jpg) + +![](images/57c7373b99afc96b087ddd9f70348bdec3b104ad0b8ca2f8b3dda05b6a103985.jpg) +$conf = 0.98$ + +![](images/97daaf357441bc729c5b8bdd4ff84d0737309f6b31b0be13e138b497536c93aa.jpg) +$conf = 0.98$ + +![](images/dce6c46738c2e26110444f0f82fc486cb2f788ea3d111490695d1b6ddf6e8e62.jpg) +$conf = 0.99$ + +![](images/71b36cb6d44365d25874b895739252d1a378d99a409eca9e863f79db513acdee.jpg) +$conf = 1.0$ + +![](images/c26c9dd08335a303d179cbbe2630761be8f4ab3f48ad6b73f2b08276409c32fb.jpg) +$conf = 1.0$ + +![](images/dbfe6e49f9e35fdffe1203b773593c2b106ede580426fa9700678f2f185f6cad.jpg) +conf $= 1.0$ + +![](images/b29838f6ef7e6212ab2899bc44bab014f8952a5a327e3a5136d28f056e1fd0ce.jpg) +conf = 0.94 +conf = 0.82 +Fig. 6: Six frames are extracted uniformly from the video. Row 1 represents a clean stop sign with the detector's output. In Row 2 and Row 3, we deploy the trigger generated by the PGD and ensemble method respectively. + +![](images/3211e5fbf980c00e5dec0844fc08ee8a80e3bce9781c29159e934f983d16ffdd.jpg) +conf = 0.72 +conf $< 0.5$ + +![](images/f2f1a8163f9747eb6f4e979e25ffde3526c132a9cda55c3dc569cb7deb73a46a.jpg) +conf = 0.70 +conf $< 0.5$ + +![](images/a6cad8c464753a06f75fe998044cb74428ce366d2df841b1b724f1e07627e41c.jpg) +conf $< 0.5$ +conf $< 0.5$ + +![](images/2fdc8d2b620c383bfc67247a2a0dc97fcde66944e285c05397d3375f7f78f700.jpg) +conf $< 0.5$ +conf $< 0.5$ + +![](images/30b02c088302f1994a6b46405fe11ca2f32ee3811b98ca22c95710c304108388.jpg) +$conf = 0.6$ +conf $< 0.5$ + +seconds of the video, during which the adversarial trigger persistently causes the detection system to lose sight of the stop sign. This effect illustrates the potential of our method to compromise the functionality of autonomous driving systems. + +As shown in the Figure 6, we extract six frames uniformly from the video as examples to demonstrate the effects of this attack in driving scenarios. + +The observations imply that normally the car will always recognize the stop sign (Row 1 of Figure 6, each stop is marked by a red box). However, after the adversarial trigger created by our attack method (Row 3) is added to the stop sign, the car fails to recognize the stop sign(no red box). Compared with the attack effect of the adversarial trigger generated by PGD (Row 2), it can be seen that our method is clearly superior to the PGD. It is worth mentioning that the perturbation generated by Dpatch2 is suboptimal in the simulator, as demonstrated in Appendix F. + +# 4.3 Robustness of our attack in real world + +To observe the attack effectiveness and robustness of the universal adversarial trigger obtained through algorithm 1 in the real world, we print out the adversarial trigger and post it under a stop sign. Moreover, in the case where clean images serve as a reference, we compare our method with the normal PGD. + +We examined the variation trends in the confidence of the detector's output for recognizing the stop sign using an onboard camera under different conditions: as the distance between the onboard camera and the stop sign decreased from far to near, and as the onboard camera's perspective towards the stop sign shifted + +![](images/dc06d519f6941854aa3b1382d908a39aac8d08daaa9d114fe402187acf191d5e.jpg) +Fig. 7: We print and deploy the adversarial trigger in the physical world. (a) illustrates the confidence variation of the stop sign as the distance increases. (b) depicts the confidence variation of the stop sign when facing different angles changing. + +![](images/5f29f0e1b609f41b75d9cd1e99a879bf61b186d69656e2d35881f4567dea5563.jpg) +(a) + +![](images/4a4ffe0d3a4463fc4445b23c01801f7b5c8af1d023e8b6742f67de06f022499b.jpg) +(b) + +from -70-degree side view to a 70-degree side view. The real-world experiment results are shown in Figure 7. + +According to Figure 7a, due to the relatively long distance between the vehicle and the stop position in the first half of the curve, the trigger occupies a small proportion of the field of view, which diminishes the stability of successful attacks. In the second half of the curve, where the distance between the stop and the vehicle is closer, the attacks become more stable. Our method consistently maintains a higher success rate in this scenario. Examining Figure 7b, it can be observed that when facing the stop sign directly, our method misleads the detector to output a remarkably low confidence. As the angle approaches approximately $\pm 50$ degrees, our method becomes ineffective. In contrast, the triggers generated by PGD are not smooth, which prevents them from effectively imprinting adversarial features into the real world. + +Overall, in the real world, our attack consistently exhibits robust and effective performance when the onboard camera is approaching a stop sign. This means that our attack can successfully mislead an autonomous driving system to disregard the stop sign unconsciously. + +# 5 Conclusion + +In this work, we present a novel adversarial attack methodology that extends beyond the bounding boxes of objects. This method poses greater challenges compared to previous studies due to its tendency to focus on less informative areas. Thus, we propose the FG techniques and UAPGD optimization strategy, specifically designed to augment the efficacy of this attack in both digital and physical realms. Furthermore, we validate the effectiveness of our attack method through its application in autonomous driving scenarios and real-world conditions. Discussion of limitations and potential negative impact of our works will be placed in Appendix G and immediately following that we also briefly address future work. + +# Acknowledgement + +We thank the anonymous referees for their valuable and insightful comments. This work is supported by the National Key R&D Program of China under Grant No. 2022YFA1005100, 2022YFA1005101, and 2022YFA1005104, the CAS Project for Young Scientists in Basic Research, under Grant No.YSBR-040, ISCAS New Cultivation Project ISCAS-PYFX-202201, ISCAS Basic Research ISCAS-JCZD-202302, and the Major Project of ISCAS under Grant No. ISCAS-ZD-202302. + +# References + +1. Aich, A., Li, S., Song, C., Asif, M.S., Krishnamurthy, S.V., Roy-Chowdhury, A.K.: Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1308-1318 (2023) 3 +2. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018) 4 +3. Brendel, W., Rauber, J., Bethge, M.: Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248 (2017) 3 +4. Cao, Y., Xiao, C., Cyr, B., Zhou, Y., Park, W., Rampazzi, S., Chen, Q.A., Fu, K., Mao, Z.M.: Adversarial sensor attack on lidar-based perception in autonomous driving. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 2267-2281 (2019) 3 +5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017) 3 +6. Chen, C., Seff, A., Kornhauser, A., Xiao, J.: Deepdriving: Learning affordance for direct perception in autonomous driving. In: Proceedings of the IEEE international conference on computer vision. pp. 2722-2730 (2015) 1 +7. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 IEEE symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020) 3 +8. Chen, P.Y., Zhang, H., Sharma, Y., Yi, J., Hsieh, C.J.: Zoo: Zeroth order optimization based black-box attacks to deep neural networks without training substitute models. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 15-26 (2017) 3 +9. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020) 3, 4, 9 +0. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019) 1 +1. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019) 1 +2. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018) 3 + +13. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019) 3 +14. Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: Carla: An open urban driving simulator. In: Conference on robot learning. pp. 1-16. PMLR (2017) 3, 12 +15. Duan, R., Ma, X., Wang, Y., Bailey, J., Qin, A.K., Yang, Y.: Adversarial camouflage: Hiding physical-world attacks with natural styles. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1000-1008 (2020) 2 +16. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018) 2, 3 +17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) 3 +18. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020) 1 +19. Hoory, S., Shapira, T., Shabtai, A., Elovici, Y.: Dynamic adversarial patch for evading object detection models. arXiv preprint arXiv:2010.13070 (2020) 2 +20. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7848-7857 (2021) 2, 3 +21. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13307-13316 (2022) 2, 3 +22. Huang, L., Gao, C., Zhou, Y., Xie, C., Yuille, A.L., Zou, C., Liu, N.: Universal physical camouflage attacks on object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 720-729 (2020) 2, 3 +23. Huang, W., Zhao, X., Jin, G., Huang, X.: Safari: Versatile and efficient evaluations for robustness of interpretability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1988-1998 (2023) 1 +24. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018) 3 +25. Im Choi, J., Tian, Q.: Adversarial attack and defense of yolo detectors in autonomous driving scenarios. In: 2022 IEEE Intelligent Vehicles Symposium (IV). pp. 1011-1017. IEEE (2022) 3 +26. Jia, S., Yin, B., Yao, T., Ding, S., Shen, C., Yang, X., Ma, C.: Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. Advances in Neural Information Processing Systems 35, 34136-34147 (2022) 3 +27. Jin, G., Yi, X., Huang, W., Schewe, S., Huang, X.: Enhancing adversarial training with second-order statistics of weights. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15273-15283 (2022) 3 +28. Jin, G., Yi, X., Wu, D., Mu, R., Huang, X.: Randomized adversarial training via taylor expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16447-16457 (2023) 3 + +29. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Michael, K., Fang, J., Yifu, Z., Wong, C., Montes, D., et al.: ultralytics/yolov5: v7. 0-yolov5 sota realtime instance segmentation. Zenodo (2022) 1 +30. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022) 3 +31. Kurakin, A., Goodfellow, I., Bengio, S., et al.: Adversarial examples in the physical world (2016) 3 +32. Lee, M., Kolter, Z.: On physical adversarial patches for object detection. arXiv preprint arXiv:1906.11897 (2019) 4, 10, 11, 16 +33. Lei, X., Cai, X., Lu, C., Jiang, Z., Gong, Z., Lu, L.: Using frequency attention to make adversarial patch powerful against person detector. IEEE Access 11, 27217-27225 (2022) 2 +34. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4282-4291 (2019) 1 +35. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8971-8980 (2018) 1 +36. Li, J., Schmidt, F., Kolter, Z.: Adversarial camera stickers: A physical camera-based attack on deep learning systems. In: International Conference on Machine Learning. pp. 3896-3904. PMLR (2019) 3 +37. Li, Y., Li, Y., Dai, X., Guo, S., Xiao, B.: Physical-world optical adversarial attacks on 3d face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24699-24708 (2023) 3 +38. Liang, S., Wu, B., Fan, Y., Wei, X., Cao, X.: Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970 (2022) 3 +39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 3, 9 +40. Liu, A., Liu, X., Fan, J., Ma, Y., Zhang, A., Xie, H., Tao, D.: Perceptual-sensitive gan for generating adversarial patches. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1028-1035 (2019) 2 +41. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023) 3 +42. Liu, X., Yang, H., Liu, Z., Song, L., Li, H., Chen, Y.: Dpatch: An adversarial patch attack on object detectors. arXiv preprint arXiv:1806.02299 (2018) 4, 11 +43. Lovisotto, G., Turner, H., Sluganovic, I., Strohmeier, M., Martinovic, I.: {SLAP}: Improving physical adversarial examples with {Short-Lived} adversarial perturbations. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1865-1882 (2021) 2, 3 +44. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 9 +45. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 7 + +46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016) 3 +47. Pomponi, J., Scardapane, S., Uncini, A.: Pixle: a fast and effective black-box attack based on rearranging pixels. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2022) 3 +48. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 1 +49. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) 1 +50. Shapira, A., Zolfi, A., Demetrio, L., Biggio, B., Shabtai, A.: Phantom sponges: Exploiting non-maximum suppression to attack deep object detectors. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4571-4580 (2023) 3 +51. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In: Proceedings of the 2016 acm sig sac conference on computer and communications security. pp. 1528-1540 (2016) 3 +52. Shi, Z., Yang, W., Xu, Z., Yu, Z., Huang, L.: Reinforcement learning-based adversarial attacks on object detectors using reward shaping. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8424-8432 (2023) 3 +53. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013) 3 +54. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) 3 +55. Tu, J., Ren, M., Manivasagam, S., Liang, M., Yang, B., Du, R., Cheng, F., Urtasun, R.: Physically realizable adversarial examples for lidar object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13716-13725 (2020) 3 +56. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2414-2422 (2022) 3 +57. Wang, D., Yao, W., Jiang, T., Li, C., Chen, X.: Rfla: A stealthy reflected light adversarial attack in the physical world. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4455-4465 (2023) 3 +58. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018) 1 +59. Wang, J., Liu, A., Yin, Z., Liu, S., Tang, S., Liu, X.: Dual attention suppression attack: Generate adversarial camouflage in physical world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8565-8574 (2021) 2, 3 +60. Wang, Q., Zhang, L., Bertinetto, L., Hu, W., Torr, P.H.: Fast online object tracking and segmentation: A unifying approach. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 1328-1338 (2019) 1 + +61. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021) 3 +62. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020) 3 +63. Wu, W., Su, Y., Chen, X., Zhao, S., King, I., Lyu, M.R., Tai, Y.W.: Boosting the transferability of adversarial samples via attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1161-1170 (2020) 3 +64. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019) 3 +65. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 665-681. Springer (2020) 3 +66. Zhang, J., Wu, W., Huang, J.t., Huang, Y., Wang, W., Su, Y., Lyu, M.R.: Improving adversarial transferability via neuron attribution-based attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14993-15002 (2022) 3 +67. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15115-15125 (2022) 3 +68. Zhang, L., Xu, N., Yang, P., Jin, G., Huang, C.C., Zhang, L.: Trajpac: Towards robustness verification of pedestrian trajectory prediction models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8327-8339 (2023) 1 +69. Zhao, Y., Zhu, H., Liang, R., Shen, Q., Zhang, S., Chen, K.: Seeing isn't believing: Towards more robust adversarial attack against real world object detectors. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 1989-2004 (2019) 2, 3 +70. Zhong, Y., Liu, X., Zhai, D., Jiang, J., Ji, X.: Shadows can be dangerous: Stealthy and effective physical-world adversarial attack by natural phenomenon. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15345-15354 (2022) 2 +71. Zhu, L., Wang, T., Li, J., Zhang, Z., Shen, J., Wang, X.: Efficient query-based black-box attack against cross-modal hashing retrieval. ACM Transactions on Information Systems 41(3), 1-25 (2023) 3 +72. Zhu, X., Li, X., Li, J., Wang, Z., Hu, X.: Fooling thermal infrared pedestrian detectors in real world using small bulbs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3616-3624 (2021) 3 +73. Zolfi, A., Kravchik, M., Elovici, Y., Shabtai, A.: The translucent patch: A physical and universal attack on object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15232-15241 (2021) 2 \ No newline at end of file diff --git a/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/images.zip b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15a1401f31dbfbe1b6c98c4374a9fe3058b684a9 --- /dev/null +++ b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f20e0cb70d4665ccd12f45d930b7bd2934b3fbc9de7214519d1f03655476d43f +size 407923 diff --git a/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/layout.json b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7276880d61af1fe9d20638536b5fd35b1a97e16b --- /dev/null +++ b/2024/Out-of-Bounding-Box Triggers_ A Stealthy Approach to Cheat Object Detectors/layout.json @@ -0,0 +1,13436 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers: A Stealthy Approach to Cheat Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "spans": [ + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "text", + "content": "Tao Lin" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "text", + "content": ", Lijia Yu" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "text", + "content": ", Gaojie Jin" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "text", + "content": ", Renjue Li" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "text", + "content": ", Peng Wu" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "inline_equation", + "content": "^{1,2,\\text{图}}" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "text", + "content": ", and Lijun Zhang" + }, + { + "bbox": [ + 141, + 173, + 472, + 205 + ], + "type": "inline_equation", + "content": "^{1,2,3,\\text{图}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 247 + ], + "type": "text", + "content": "1 Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 247, + 431, + 280 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 181, + 247, + 430, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 247, + 430, + 258 + ], + "spans": [ + { + "bbox": [ + 181, + 247, + 430, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 247, + 430, + 258 + ], + "type": "text", + "content": " University of Chinese Academy of Sciences, Beijing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 258, + 431, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 258, + 431, + 280 + ], + "spans": [ + { + "bbox": [ + 181, + 258, + 431, + 280 + ], + "type": "text", + "content": "3 Automotive Software Innovation Center, Chongqing, China {lintao,yulijia,gaojie,lirj19,wp,zhanglj}@ios.ac.cn" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 306, + 455, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 306, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 159, + 306, + 455, + 483 + ], + "type": "text", + "content": "Abstract. In recent years, the study of adversarial robustness in object detection systems, particularly those based on deep neural networks (DNNs), has become a pivotal area of research. Traditional physical attacks targeting object detectors, such as adversarial patches and texture manipulations, directly manipulate the surface of the object. While these methods are effective, their overt manipulation of objects may draw attention in real-world applications. To address this, this paper introduces a more subtle approach: an inconspicuous adversarial trigger that operates outside the bounding boxes, rendering the object undetectable to the model. We further enhance this approach by proposing the Feature Guidance (FG) technique and the Universal AutoPGD (UAPGD) optimization strategy for crafting high-quality triggers. The effectiveness of our method is validated through extensive empirical testing, demonstrating its high performance in both digital and physical environments. The code and video will be available at: https://github.com/linToTao/Out-of-bbox-attack." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "content": "Keywords: Universal physical adversarial attack " + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "content": " Out-of-bounding-box trigger " + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "content": " Object detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Object detection aims to identify instances of semantic objects typically within images or video clips. It finds widespread applications across a range of domains, including face recognition [10, 11, 23, 49, 58], object tracking [34, 35, 60], and autonomous driving [6, 18, 68]. Especially in autonomous driving systems, object detectors play a crucial role in tasks such as recognizing traffic signs, pedestrians, vehicles, and traffic lights. Nonetheless, in recent years, security concerns have emerged regarding object detectors, primarily due to the susceptibility of deep neural networks (DNNs) to adversarial examples (AEs). AEs are intentionally crafted inputs designed to deceive DNNs into making incorrect predictions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 113, + 473, + 238 + ], + "blocks": [ + { + "bbox": [ + 142, + 113, + 473, + 238 + ], + "lines": [ + { + "bbox": [ + 142, + 113, + 473, + 238 + ], + "spans": [ + { + "bbox": [ + 142, + 113, + 473, + 238 + ], + "type": "image", + "image_path": "d27a9aa680fd619b1b3f75df85313255adc96627a4aca5e08809f66f56ca564a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "text", + "content": "Fig. 1: An illustration of vanilla object attack methods and our novel approach. Typically, vanilla attack methods target the most informative regions of an object, such as the central, highly visible areas of the target object, e.g., the stop sign in (a) [16], (b) [69], (c) [15], and the car in (d) [19], (e) [59], (f) [22]. In contrast, our method adopts a more covert strategy, focusing on attacking the peripheral edge areas surrounding the stop sign." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": "Early research mainly focused on studying AEs like adversarial patch [16, 20, 33, 40, 73], optical [43, 70] and texture [15, 21], which are targeted at the objects. These AEs involve the computation of perturbations applied directly to the objects to mislead a DNN-based detector. However, a direct attack on the object as shown in Figure 1, along with its subsequent attempt to deceive the detector, is overly conspicuous and cannot be deployed without raising suspicion in real-world practical scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 483, + 467 + ], + "type": "text", + "content": "In this work, we provide an alternative perspective: is it possible to introduce an adversarial trigger in an inconspicuous area outside the object while still successfully deceiving the detector?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": "Actually, crafting an adversarial trigger within the areas beyond the object boundaries is even more demanding. On one hand, for dense prediction networks like object detectors, the output results are more reliant on the red regions of the heatmap in Figure 1. This implies that tailoring the adversarial trigger by utilizing irrelevant information, like the blue regions of the heatmap, is sharply challenging. More difficultly, our trigger is designed to conform to practical size and shape constraints. Nevertheless, for exactly the same reasons, our attack method gets more delusive and harder to defense." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "This paper is focused on the development of robust and inconspicuous triggers to attack typical object detectors deployed in practical scenarios. To tackle the challenge mentioned above, we present a novel technique called Feature Guidance (FG). In this approach, we guide the feature layers with an adversarial trigger to resemble those with the object hidden. And we propose a optimization strategy, called Universal Auto-PGD (UAPGD), applicable to universal attacks. This optimization strategy allows for the selection of an appropriate timing to halve the step size, based on the oscillation of the loss function." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "To assess the effectiveness of our approach in both digital and physical domains, we trained the adversarial trigger and evaluated the corresponding metrics on COCO [39] dataset and Carla [14], respectively. We also evaluated the robustness of the trigger by recording videos using an onboard camera in the real world." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 176, + 384, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 176, + 384, + 188 + ], + "spans": [ + { + "bbox": [ + 146, + 176, + 384, + 188 + ], + "type": "text", + "content": "Overall, this paper makes the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 480, + 323 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 194, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 479, + 228 + ], + "type": "text", + "content": "- This work studies the realm of adversarial robustness in object detection, providing a novel perspective by exploring the concept of an adversarial trigger applied outside an object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 230, + 480, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 230, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 230, + 480, + 287 + ], + "type": "text", + "content": "- We propose a novel attack method designed to generate triggers that achieve excellent continuous attack effects and demonstrate high robustness in real world. The key enabler of our method is the utilization of similarity between feature layers and adaptive step size decay to endow the trigger with the ability to mislead the detector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 290, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 480, + 323 + ], + "type": "text", + "content": "- An extensive set of empirical results are provided to demonstrate that our attack exhibits strong performance and robustness in both the COCO dataset and Carla simulator, as well as in real-world scenarios." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 342, + 234, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 234, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 234, + 355 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 413 + ], + "type": "text", + "content": "Adversarial examples (AEs), first discovered by Szegedy et al. [53], involve carefully crafted inputs inducing model misclassifications. Transferability [26, 61, 64, 66] of AEs, allows them to deceive networks with different hyperparameters or architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "text", + "content": "Digital adversarial examples are generated by directly manipulating any pixels in digital images prior to feeding them into the model. Adversarial attacks can be classified into black-box attacks and white-box attacks based on whether the internal structure and parameters of the target model are known. White-box attacks [5, 12, 17, 27, 28, 31, 44, 46] primarily rely on the computation of model gradients. Black-box methods, on the other hand, depend on querying the unknown model by observing the prediction [12, 13, 38, 47, 62–64, 71] or leveraging an agent model to imitate the threaten model [3, 7, 8, 24, 30, 67]. Recently, Croce et al. [9] proposed AutoAttack, a parameter-free method that combines novel white-box and black-box attack strategies. The works of applying digital adversarial examples to object detectors have been explored in [25] and [50] that introduce perturbations to input images to mislead detectors. [52] generate more subtle perturbations via reinforcement learning, successfully attacking the detector." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Physical adversarial examples are crafted to be printed out and captured by sensory equipment like camera [31,36] and LiDAR [4,41,55]. These examples pose a real-world threat due to their ability to adapt to camera and scenario variations. Early studies focused on attacking recognition models [16,51]. Recent research has explored physical attacks in the more challenging context of object detection. Perturbations in this domain include adversarial patches [1,20,54,65, 69], adversarial camouflage textures [21,22,56,59], and optical attacks [37,43,57, 72]. To improve robustness against various transformations in physical settings," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": "Athalye et al. [2] introduced Expectation Over Transformation (EOT), widely used in physical attacks, to enhance the effectiveness of adversarial attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "text", + "content": "Prior physical attack methods need to be applied directly to the objects so that these methods are expected to yield favorable results but they are prone to raising suspicion in practical scenarios. Unlike these works, our method constructs a universal attack trigger that can deceive the detector disregarding the specific object without altering the appearance of itself, and can be discreetly deployed in the real world." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 282 + ], + "type": "text", + "content": "The works that exhibit the closest alignment with our motivation are Dpatch [42] and [32] (after this referred to as Dpatch2). However, we must emphasize that our work is independent of them, highlighting the difference between our work and theirs in Appendix F. Furthermore, in the experiments section, we compare the attack effectiveness of our method with Dpatch2 [32] on the COCO dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 302, + 233, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 302, + 233, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 233, + 316 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 327, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 533 + ], + "type": "text", + "content": "In this work, we aim to generate physical adversarial triggers that can be naturally deployed in an inconsequential position. These adversarial triggers are designed not to destroy the object information but still deceive the model effectively. In this section, we commence by presenting the problem formulation in Section 3.1. Subsequently, Fig. 2a provides a comprehensive illustration of our framework. Our primary focus is on the white-box attack paradigm, which necessitates an intricate optimization process. Given an object detection model, we design a loss function to determine the gradient direction for adversarial attacks in Section 3.2. This loss function is then utilized iteratively to update the adversarial trigger. In the pursuit of strengthening attack performance and enhancing the continuity of attacks in video-based scenarios, we introduce a novel method named \"Feature Guidance\" (FG) and an additional loss term " + }, + { + "bbox": [ + 130, + 327, + 482, + 533 + ], + "type": "inline_equation", + "content": "L_{FG}" + }, + { + "bbox": [ + 130, + 327, + 482, + 533 + ], + "type": "text", + "content": " in Section 3.2. To bolster the overall robustness and deployability of our approach, we also employ two indispensable techniques, EOT and total variation loss in Section 3.2. Furthermore, our methodology draws inspiration from the work of [9] in their Auto PGD framework during the training process, we thus modify the conditions of step size decay based on the universal training mode (Section 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 549, + 269, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 269, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 269, + 561 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 629 + ], + "type": "text", + "content": "Adversarial attacks on object detection can have diverse objectives, encompassing but not limited to inducing misalignment in bounding boxes, reducing classification confidence, or provoking misclassification. Our goal is to generate adversarial examples for a specific class, with the aim of causing the object detection algorithm to fail in identifying objects belonging to that class." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " with the ground truth " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "y = \\{(y_i^{\\mathrm{coor}}, y_i^{\\mathrm{cls}})\\}_{i \\in [n]} \\subset \\mathbb{R}_+^4 \\times \\mathbb{Z}_+" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " is the number of objects in the image, " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{coor}}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " contains top-left and bottom-right coordinates of the bounding box in the image, and " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{cls}}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " is the class of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 112, + 306, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 112, + 306, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 112, + 306, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 306, + 219 + ], + "type": "image", + "image_path": "b6a78b26b4e945d1d75311757f6bbaed5ebcfc963d7903bd4d7028b09970f320.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "type": "text", + "content": "Fig.2: (a) shows the overview of our adversarial trigger generation framework. We deploy the adversarial trigger onto each image in the dataset with robustness transformation. With a fixed detector, our method optimizes the adversarial trigger through the obtained adversarial gradient. (b) illustrates the motivation and calculation method of Feature Guidance. Taking a stop sign as an example, for a given image " + }, + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "type": "text", + "content": ", minimizing " + }, + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\| f_z(T(x,\\mathcal{A}_t)) - f_z(x')\\|_2" + }, + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "type": "text", + "content": " can mislead a detector to classify the stop sign image " + }, + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "type": "inline_equation", + "content": "T(x,\\mathcal{A}_t)" + }, + { + "bbox": [ + 130, + 225, + 482, + 304 + ], + "type": "text", + "content": " as none stop." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 137, + 470, + 217 + ], + "blocks": [ + { + "bbox": [ + 315, + 137, + 470, + 217 + ], + "lines": [ + { + "bbox": [ + 315, + 137, + 470, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 470, + 217 + ], + "type": "image", + "image_path": "040c7978640aa6ffc12a9b83a1f8b8b143351261c8a3c6cb9427c04db7f9f4b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 480, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 349 + ], + "type": "text", + "content": "object. Then, the DNN-based object detector can be represented as the following form:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 349, + 357, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 349, + 357, + 363 + ], + "spans": [ + { + "bbox": [ + 255, + 349, + 357, + 363 + ], + "type": "interline_equation", + "content": "f: \\mathbb {R} ^ {d} \\to 2 ^ {\\mathbb {R} _ {+} ^ {4} \\times \\mathbb {Z} _ {+} \\times [ 0, 1 ] ^ {2}}", + "image_path": "d082ee5670c2fa46275917a69767b3e40d070dc0421d24c79fef53e950647571.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": "For a given image " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": ", the output of " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": " can be represented as " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\tilde{y} = \\{(\\tilde{y}_i^{\\mathrm{coor}},\\tilde{y}_i^{\\mathrm{cls}},\\tilde{y}_i^{\\mathrm{conf}})\\}_{i\\in [n]} \\subset \\mathbb{R}_+^4\\times \\mathbb{Z}_+ \\times [0,1]^2" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i^{\\mathrm{conf}}" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": " denotes the confidence associated with the predicted box coordinates " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i^{\\mathrm{coor}}" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": " and the predicted class " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i^{\\mathrm{cls}}" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": "-th object. We aim to generate a universal adversarial trigger " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t\\in [0,1]^{3\\times U\\times V}" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "U,V\\in \\mathbb{Z}_{+}" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": " are the given dimensions, for the given specific class " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": ", such that these adversarial triggers can cause " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": " to fail in recognizing objects of class " + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 369, + 482, + 465 + ], + "type": "text", + "content": ". Then, the problem can be formulated as an optimization task, represented as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 192, + 466, + 481, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 466, + 481, + 481 + ], + "spans": [ + { + "bbox": [ + 192, + 466, + 481, + 481 + ], + "type": "interline_equation", + "content": "\\operatorname {a r g m i n} _ {\\mathcal {A} _ {t}} \\mathbb {E} _ {x \\sim D} [ | f (T (x, \\mathcal {A} _ {t})) \\cap (\\mathbb {R} _ {+} ^ {4} \\times \\{t \\} \\times [ 0, 1 ] ^ {2}) | ] \\tag {1}", + "image_path": "1eb6ff0d07bbe0ded1e5bcc5b5ea25ac69c3a8a27573ca6c6715f7d6bbdd6d18.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "inline_equation", + "content": "|A|" + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": " denotes the cardinality of the set " + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": " represents the input distribution, and " + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": " (which will be elaborated in the following) is a transform operation to incorporate the adversarial trigger into the original image. In contrast to prior studies based on adversarial patches, our approach avoids placing the adversarial trigger on the informative surface, e.g., traffic signs, thereby preventing any unnecessary attention. We would like to point out that this kind of attack method is challenging because the adversarial trigger can only be stuck on the unimportant places of the image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": "Transform operation. We utilize a differentiable operation " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " to render the adversarial trigger " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " onto an image " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": ". Let the object with class " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " be located in the box " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " of the image, we use the following two-step to add the adversarial trigger to the image:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 131, + 629, + 479, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 653 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 653 + ], + "type": "text", + "content": "(1) We can select an appropriate area " + }, + { + "bbox": [ + 131, + 629, + 479, + 653 + ], + "type": "inline_equation", + "content": "C(B)" + }, + { + "bbox": [ + 131, + 629, + 479, + 653 + ], + "type": "text", + "content": " in the image where " + }, + { + "bbox": [ + 131, + 629, + 479, + 653 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 629, + 479, + 653 + ], + "type": "text", + "content": " is a given function, such position does not overlap with the surface of the object;" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": "(2) We cover the box " + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "inline_equation", + "content": "C(B)" + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": " part of the image " + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": " with adversarial trigger " + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 131, + 653, + 479, + 666 + ], + "type": "text", + "content": ", but" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Considering that the adversarial trigger and the box are not the same size, we need to apply the following affine matrix to " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 149, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 149, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 244, + 149, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\theta = \\left( \\begin{array}{c c c} S _ {h} \\cos \\alpha & - S _ {h} \\sin \\alpha & T _ {h} \\\\ S _ {v} \\sin \\alpha & S _ {v} \\cos \\alpha & T _ {v} \\\\ 0 & 0 & 1 \\end{array} \\right) \\tag {2}", + "image_path": "ae225647a60cdff225bc9b4dd984b10731cf231bc12a538dd4242420d2fd6213.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "inline_equation", + "content": "S_{h}" + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "inline_equation", + "content": "S_{v}" + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "content": " denotes the scaling factors utilized to resize the adversarial trigger, " + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "content": " represents the angle of rotation, " + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "inline_equation", + "content": "T_{h}" + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "inline_equation", + "content": "T_{v}" + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "content": " are the distance adversarial trigger need to shift. These parameters are all computed by " + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "inline_equation", + "content": "C(B)" + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "content": ". The more specific remark regarding " + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "inline_equation", + "content": "C(B)" + }, + { + "bbox": [ + 130, + 194, + 479, + 242 + ], + "type": "text", + "content": " will be detailed in Appendix B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 480, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 278 + ], + "type": "text", + "content": "Note that our attack method has two characteristics: Our attack is out of the object; our attack area changes depending on the position of the object. We were the first to attack target detection with such a method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 296, + 382, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 382, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 382, + 308 + ], + "type": "text", + "content": "3.2 Adversarial gradient with Feature Guidance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 316, + 479, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 397 + ], + "type": "text", + "content": "In this paper, we focus on utilizing white-box attack methods to search for adversarial gradients and optimize the adversarial trigger. Additionally, we provide black-box empirical results in Appendix D that the trigger is generated by a surrogate model. As illustrated above, we render the adversarial trigger onto images and feed them to the detectors, take the derivative of the loss function, and use it to update the adversarial trigger. The loss function consists of the following parts:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 400, + 479, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 479, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 479, + 458 + ], + "type": "text", + "content": "Adversarial loss of detector output. As mentioned above, the detector outputs the predicted class and bounding boxes and their confidence. Considering most of the detectors will not output the predicted class and boxes with a confidence lower than a threshold, we want images with the adversarial trigger to make the confidence as low as possible." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "inline_equation", + "content": "f_{i,coor}^{\\mathrm{conf}}(x)" + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "inline_equation", + "content": "f_{i,cls}^{\\mathrm{conf}}(x)" + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "text", + "content": " is the confidence of the " + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "text", + "content": "-th object on image " + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 459, + 479, + 483 + ], + "type": "text", + "content": ", we can define the first part of our loss function as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 183, + 492, + 480, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 492, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 183, + 492, + 480, + 512 + ], + "type": "interline_equation", + "content": "L _ {d e t} \\left(\\mathcal {A} _ {t}\\right) = \\mathbb {E} _ {x \\sim D} \\left[ \\max _ {i: \\tilde {y} _ {i} ^ {c l s} = t} f _ {i, c o o r} ^ {\\text {c o n f}} (T (x, \\mathcal {A} _ {t})) f _ {i, c l s} ^ {\\text {c o n f}} (T (x, \\mathcal {A} _ {t})) \\right] \\tag {3}", + "image_path": "b1abd8b645100a388bf9ca0136260f9002ead4b91fd881cc10838fb52196e245.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 479, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 557 + ], + "type": "text", + "content": "Feature guidance. The biggest difficulty with an attack outside the object is that the features of the original object on the image will not be destroyed, so the network will still notice the presence of the object." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": "We gain the following insights from the network structure to address this issue. A DNN-based detector is usually designed with three modules: Backbone, Neck, and Head. The Backbone module extracts global features from an input image. These features are then fused through the Neck module to produce multiple feature maps with varying receptive fields. The Head module subsequently makes predictions based on the feature maps. This modular structure imitates how people often perceive a comprehensive horizon (like the Backbone module) and focus on varying attractive areas (like the Neck module). Subsequently, the attractive areas will contribute to making perceptions (like the Head module)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Hence, in order to provoke model errors, it is imperative to deliberately induce information misleading prior to inputting the head module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "To achieve that, we use the following method, called Feature guidance, as shown in Fig. 2b. For a given image " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " with a target object " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", we can get the image " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", which differs from " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " only with the object " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " missing. Then, if we can elaborate our adversarial trigger " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " to guide the feature map " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "f_z(T(x, \\mathcal{A}_t))" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " as close to " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "f_z(x')" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " generated by the Neck module of the detector, it will have a desirable impact on the Head module such that the stop sign cannot be detected." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "text", + "content": "As illustrated in Figure 2a, for the sake of convenience, we derive " + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "inline_equation", + "content": "x_{mask}" + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "text", + "content": " that exhibits an analogous effect to " + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "text", + "content": " simply by covering the bounding box position in the ground truth of " + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "text", + "content": " with a gray rectangle." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 270 + ], + "type": "text", + "content": "In order to achieve Feature guidance, we need to make " + }, + { + "bbox": [ + 130, + 248, + 479, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 130, + 248, + 479, + 270 + ], + "type": "text", + "content": " to minimum such loss function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 281, + 411, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 281, + 411, + 295 + ], + "spans": [ + { + "bbox": [ + 201, + 281, + 411, + 295 + ], + "type": "interline_equation", + "content": "L _ {F G} (\\mathcal {A} _ {t}) = E _ {x \\sim D} [ \\| f _ {z} (T (x, \\mathcal {A} _ {t})) - f _ {z} (x _ {\\text {m a s k}}) \\| _ {2} ]", + "image_path": "42faf9215ed9ca2bc8a7faef00f995d617e917c7dbdec5b5eebdf3d258c89ccd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "type": "text", + "content": "Other skills To enhance the robustness of our adversarial trigger " + }, + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 130, + 304, + 480, + 365 + ], + "type": "text", + "content": ", we incorporate diverse variations, including random noise, contrast, brightness, and rotation, into the generated trigger prior to its application on images, thus simulating real-world conditions. To address printing challenges, we employ TV loss [45] on the adversarial trigger to promote smoothness. TV loss is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 193, + 374, + 480, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 374, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 193, + 374, + 480, + 403 + ], + "type": "interline_equation", + "content": "L _ {t v} \\left(\\mathcal {A} _ {t}\\right) = \\sum_ {i, j} \\sqrt {\\left(\\mathcal {A} _ {t} ^ {i , j} - \\mathcal {A} _ {t} ^ {i + 1 , j}\\right) ^ {2} + \\left(\\mathcal {A} _ {t} ^ {i , j} - \\mathcal {A} _ {t} ^ {i , j + 1}\\right) ^ {2}} \\tag {4}", + "image_path": "aea3936f0d646766c41be38baf9b52aa320c3f38b63b870b991adc7131a2a847.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "type": "text", + "content": "where the subindices " + }, + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "type": "text", + "content": " represent the pixel coordinate of the trigger " + }, + { + "bbox": [ + 130, + 407, + 468, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 419, + 479, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 479, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 479, + 439 + ], + "type": "text", + "content": "Overall, we will get the adversarial trigger " + }, + { + "bbox": [ + 130, + 419, + 479, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 130, + 419, + 479, + 439 + ], + "type": "text", + "content": " by optimizing following problem:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 200, + 441, + 480, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 441, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 200, + 441, + 480, + 459 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {A} _ {t} \\in [ 0, 1 ] ^ {3 \\times U \\times V}} L _ {d e t} (\\mathcal {A} _ {t}) + \\lambda_ {F G} L _ {F G} (\\mathcal {A} _ {t}) + \\lambda_ {t v} L _ {t v} (\\mathcal {A} _ {t}) \\tag {5}", + "image_path": "453cb92dc536d9795ebbb1c5369b33377dd63a594fb9e1740a75405727b16e67.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 462, + 313, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 462, + 313, + 474 + ], + "spans": [ + { + "bbox": [ + 134, + 462, + 313, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 134, + 462, + 313, + 474 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}" + }, + { + "bbox": [ + 134, + 462, + 313, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 462, + 313, + 474 + ], + "type": "inline_equation", + "content": "\\lambda_{tv}" + }, + { + "bbox": [ + 134, + 462, + 313, + 474 + ], + "type": "text", + "content": " are hyperparameter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 492, + 266, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 266, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 266, + 503 + ], + "type": "text", + "content": "3.3 Universal Auto-PGD" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 513, + 479, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 479, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 479, + 536 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 513, + 479, + 536 + ], + "type": "inline_equation", + "content": "L_{all} = L_{det} + \\lambda_{FG}L_{FG} + \\lambda_{tv}L_{tv}" + }, + { + "bbox": [ + 130, + 513, + 479, + 536 + ], + "type": "text", + "content": ", we will use gradient descent to find the adversarial trigger via minimizing the loss function " + }, + { + "bbox": [ + 130, + 513, + 479, + 536 + ], + "type": "inline_equation", + "content": "L_{all}" + }, + { + "bbox": [ + 130, + 513, + 479, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "type": "text", + "content": "Considering the instability that appeared in the experiment when optimizing multiple terms of " + }, + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "type": "inline_equation", + "content": "L_{all}" + }, + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "type": "text", + "content": " with conventional methods, we propose the following optimization strategy called Universal Auto-PGD (UAPGD):" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "text", + "content": "At the beginning of optimizing, we fix the following values: comparison length " + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "inline_equation", + "content": "l_{c}" + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "text", + "content": ", observation length " + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "inline_equation", + "content": "l_{o}" + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "text", + "content": ", the slack variable " + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon_{1}, \\epsilon_{2}" + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "text", + "content": ", let the initial step size be " + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 573, + 479, + 608 + ], + "type": "text", + "content": " and we record the following values:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 609, + 432, + 632 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 131, + 609, + 432, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 432, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 432, + 621 + ], + "type": "text", + "content": "(1): After epoch " + }, + { + "bbox": [ + 131, + 609, + 432, + 621 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 609, + 432, + 621 + ], + "type": "text", + "content": ", record the value of the loss function, name it " + }, + { + "bbox": [ + 131, + 609, + 432, + 621 + ], + "type": "inline_equation", + "content": "L_{j}" + }, + { + "bbox": [ + 131, + 609, + 432, + 621 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 621, + 326, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 326, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 326, + 632 + ], + "type": "text", + "content": "(2): After epoch " + }, + { + "bbox": [ + 131, + 621, + 326, + 632 + ], + "type": "inline_equation", + "content": "i * l_{o}" + }, + { + "bbox": [ + 131, + 621, + 326, + 632 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 621, + 326, + 632 + ], + "type": "inline_equation", + "content": "i \\in Z_{+}" + }, + { + "bbox": [ + 131, + 621, + 326, + 632 + ], + "type": "text", + "content": ", record:" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 217, + 642, + 394, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 642, + 394, + 668 + ], + "spans": [ + { + "bbox": [ + 217, + 642, + 394, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\min , i} = \\min _ {j \\in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \\left\\{L _ {j} \\right\\} \\\\ V _ {i} = \\operatorname {V a r i a n c e} \\left[ \\left\\{L _ {j} \\right\\} _ {j \\in [ (i - 1) * l _ {o} + 1, i * l _ {o} ]} \\right] \\\\ \\end{array}", + "image_path": "134553a7ffee1542cd0e30f8fc1364939340e6c0025ab1c9b82c2ca7f68cf11c.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 133, + 132, + 481, + 483 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 247, + 129 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 247, + 129 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 247, + 129 + ], + "type": "text", + "content": "Algorithm 1: UAPGD" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "lines": [ + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": "Input: Threaten network " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " , initial step size " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " , the number of epoch " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{epoch}}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " dataset " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "D = \\cup_{j = 1}^{k}D_{j}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " which was divided into k batches, the hyperparameter in UAPGD " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "l_c,l_o" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " , initial adversarial trigger " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t^0" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " Output: Adversarial trigger " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{best}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "S_{c}\\gets \\phi" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "S_{o}\\gets \\phi" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "L_{best}\\gets \\infty" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{best}\\leftarrow \\mathcal{A}_t^0" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n2 for " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "i\\in [N_{\\mathrm{epoch}}]" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " do \n3 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "L_{i}\\gets 0" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{t}^{i,0}\\gets \\mathcal{A}_{t}^{i - 1}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n5 for " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "j\\in [k]" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " do \n6 Get " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "L_{all}(\\mathcal{A}_t^{i,j - 1},D_j)" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n7 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "L_{i}\\gets L_{i} + L_{all}(\\mathcal{A}_{t}^{i,j - 1},D_{j}) / k" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n8 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{t}^{i,j}\\gets \\mathcal{A}_{t}^{i,j - 1} - \\eta \\cdot \\mathrm{sign}(\\frac{\\nabla L_{all}(\\mathcal{A}_{t}^{i,j - 1},D_{j})}{\\nabla\\mathcal{A}_{t}^{i,j - 1}})" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n9 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{t}^{i,j}\\gets Clamp(\\mathcal{A}_{t}^{i,j},0,1)" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n10 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{t}^{i}\\gets \\mathcal{A}_{t}^{i,k}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n11 if " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "L_{i} < L_{best}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " then \n12 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\big{\\lfloor}\\mathcal{A}_{best}\\leftarrow \\mathcal{A}_{t}^{i}\\text{and} L_{best}\\leftarrow L_{i} \\end{array}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n13 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "S_{o} = S_{o}\\cup L_{i}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n14 if " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "|S_{o}| = l_{o}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " then \n15 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl} & L_{min}\\gets \\min_{L\\in S_o\\backslash \\{S_o\\}}\\\\ & V\\gets \\operatorname{Var}_{L\\in S_o\\backslash \\{S_o\\}}\\\\ & S_c = S_c\\cup (L_{min},V)\\\\ & S_o\\gets \\phi \\end{array}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n16 \n17 \n18 \n19 if " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "|S_{c}| = l_{c}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " then // Condition 1 and Condition 2 are introduced in section 3.3 if Condition 1(Sc) and Condition 2(Sc) then \n21 " + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\lfloor \\eta \\leftarrow \\eta /2\\text{and}\\mathcal{A}_t^i\\leftarrow \\mathcal{A}_{best}\\\\ S_c\\leftarrow \\phi \\end{array}" + }, + { + "bbox": [ + 133, + 132, + 481, + 483 + ], + "type": "text", + "content": " \n22" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 518, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 552 + ], + "type": "text", + "content": "In order to prevent violent oscillations around the local optimum during the optimization process, we will halve the step size when the following conditions are met:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 553, + 369, + 578 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 553, + 369, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 369, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 369, + 567 + ], + "type": "text", + "content": "(Con 1): " + }, + { + "bbox": [ + 132, + 553, + 369, + 567 + ], + "type": "inline_equation", + "content": "L_{min,j} \\geq L_{min,q} - \\epsilon_1" + }, + { + "bbox": [ + 132, + 553, + 369, + 567 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 132, + 553, + 369, + 567 + ], + "type": "inline_equation", + "content": "(j - l_c) < q < j" + }, + { + "bbox": [ + 132, + 553, + 369, + 567 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 566, + 324, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 324, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 324, + 578 + ], + "type": "text", + "content": "(Con 2): " + }, + { + "bbox": [ + 132, + 566, + 324, + 578 + ], + "type": "inline_equation", + "content": "V_{j}\\geq V_{q} - \\epsilon_{2}" + }, + { + "bbox": [ + 132, + 566, + 324, + 578 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 132, + 566, + 324, + 578 + ], + "type": "inline_equation", + "content": "j - l_{c} < q < j" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 578, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 614 + ], + "type": "text", + "content": "Due to the stochasticity from the robustness transformation mentioned in Other skills, the value of the loss function may fluctuate so that we incorporate the slack variable " + }, + { + "bbox": [ + 130, + 578, + 482, + 614 + ], + "type": "inline_equation", + "content": "\\epsilon_{1},\\epsilon_{2}" + }, + { + "bbox": [ + 130, + 578, + 482, + 614 + ], + "type": "text", + "content": " into numerical comparison." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Intuitively, these two conditions respectively describe the exploration and exploitation aspects in the optimization process. Con 1 means that the value of the loss function has not decreased much, indicating the necessity of a transition from exploring the entire feasible space to focusing on local optimization. Con" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "2 characterizes the oscillation of the loss function, of which the amplitude per " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "l_{c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " steps has not decreased much. This implies the imminent conclusion of the algorithm's exploitation in a local space. The simultaneous occurrence of c1 and c2 means that the adversarial trigger jumps near one of the optimal points, but the step size is too large to fall steadily near the best advantages. Therefore, at this point, we will halve the step size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "type": "text", + "content": "Our approach is somewhat similar to AutoAttack [9], but there are several differences: (1): Our method is for universal attack, AutoAttack is for attacking a single image; (2): We consider the amplitude of the oscillation during the optimization process and use this as an indicator of decline; (3): We include the slack variables in order to mitigate potential misleading phenomena caused by experimental randomness. For example, the oscillation of the loss function often easily cause " + }, + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "type": "inline_equation", + "content": "L_{j}" + }, + { + "bbox": [ + 130, + 188, + 482, + 270 + ], + "type": "text", + "content": " to show a downward trend." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "text", + "content": "The overall process of our algorithm is shown in algorithm 1. And we use " + }, + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "inline_equation", + "content": "L_{all}(\\mathcal{A}_t, D_j)" + }, + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "text", + "content": " to represent calculating " + }, + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "inline_equation", + "content": "L_{all}" + }, + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "text", + "content": " on an input batch " + }, + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "inline_equation", + "content": "D_j" + }, + { + "bbox": [ + 131, + 271, + 481, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 314, + 223, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 314, + 223, + 328 + ], + "spans": [ + { + "bbox": [ + 131, + 314, + 223, + 328 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": "In this section, we choose the stop sign as the target object to evaluate the effectiveness of our adversarial attack method. The stop sign is a commonly used object in the evaluation of physical adversarial attacks, as it is critical to traffic safety. We begin by detailing the implementation settings of our evaluation. This is followed by an extensive set of experiments, utilizing the proposed adversarial trigger in the digital space, specifically on the COCO dataset [39]. Furthermore, we extend the application of our method to real-world scenarios, i.e., we record videos to demonstrate that our method can attack the object detector in both simulators and real-world scenes. Meanwhile, we evaluate our attack method across varying distances and angles. See Appendix A and B for the basic experimental setting, the results of additional attacks on other objects, and the detail of selecting C(B) to place the trigger on the stop sign." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 502, + 256, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 256, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 256, + 514 + ], + "type": "text", + "content": "4.1 Digital experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "In this subsection, our experiments are carried out using the COCO dataset [39]. From this dataset, we extract a total of 1100 images featuring stop signs. Of these, 600 images constitute the training set, which is utilized to generate the adversarial trigger. The remaining 500 images are designated as the test set, serving to evaluate the effectiveness of the generated adversarial trigger." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "To demonstrate the superiority of our algorithm, we adopt the following methods to generate adversarial triggers: optimizing the adversarial trigger using PGD [44] or UAPGD with a minimization of the loss function " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{det} + \\lambda_{tv}L_{tv}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{all}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ". This included comparisons between our method and the normal PGD under the condition of placing the perturbation outside the object, as well as ablation experiments with varying hyperparameters. Specifically, for the YOLOv3(YOLOv5) model, we choose " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " values of 0.1, 0.5, and 1.0(0.05," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "0.1, and 0.2). The optimization process utilized the initial step size of 8/255 and 16/255. We calculate the number of undetected images among all the images as a metric called attack success rate(ASR)." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 146, + 178, + 321, + 327 + ], + "blocks": [ + { + "bbox": [ + 146, + 178, + 321, + 327 + ], + "lines": [ + { + "bbox": [ + 146, + 178, + 321, + 327 + ], + "spans": [ + { + "bbox": [ + 146, + 178, + 321, + 327 + ], + "type": "table", + "html": "
Threaten ModelYOLOv3YOLOv5
DatasetTrain(%)Test(%)Train(%)Test(%)
Dpatch226.6724.007.176.67
Dpatch2 (Ldet)35.5335.8032.8332.00
Adam (16/255)65.1662.2043.8340.40
Adam (8/255)64.3360.8052.1650.80
PGD(16/255)56.6754.0043.6544.60
PGD(8/255)65.1762.2053.8351.00
UAPGD(16/255)66.0063.0057.8352.60
UAPGD(8/255)67.8364.8054.1750.40
FG(λF, FG, 16/255)57.6755.4051.8348.20
FG(λF, FG, 16/255)56.0054.6047.8344.60
FG(λF, FG, 16/255)54.0051.0043.8340.00
FG(λF, FG, 8/255)60.0055.2059.1755.60
FG(λF, FG, 8/255)64.8362.6052.1750.60
FG(λF, FG, 8/255)61.6759.8056.6756.40
Ensemble(λF, FG, 16/255)67.8366.2058.3356.40
Ensemble(λF, FG, 16/255)67.8364.8060.1757.60
Ensemble(λF, FG, 16/255)64.8361.6058.0056.60
Ensemble(λF, FG, 8/255)68.3365.0062.6758.60
Ensemble(λF, FG, 8/255)69.3366.6056.6753.20
Ensemble(λF, FG, 8/255)66.1662.2060.1757.80
", + "image_path": "387eee5861047cbcf18a2e8f69887e106531443e69063696bef5bfe074c2f98e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 332, + 178, + 389, + 251 + ], + "blocks": [ + { + "bbox": [ + 332, + 178, + 389, + 251 + ], + "lines": [ + { + "bbox": [ + 332, + 178, + 389, + 251 + ], + "spans": [ + { + "bbox": [ + 332, + 178, + 389, + 251 + ], + "type": "image", + "image_path": "e500f78b3fb3b0e15738097e9d4b2a293b6a94619633331edbc4f41b2060a1fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 252, + 369, + 259 + ], + "lines": [ + { + "bbox": [ + 351, + 252, + 369, + 259 + ], + "spans": [ + { + "bbox": [ + 351, + 252, + 369, + 259 + ], + "type": "text", + "content": "PGD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 402, + 179, + 458, + 251 + ], + "blocks": [ + { + "bbox": [ + 402, + 179, + 458, + 251 + ], + "lines": [ + { + "bbox": [ + 402, + 179, + 458, + 251 + ], + "spans": [ + { + "bbox": [ + 402, + 179, + 458, + 251 + ], + "type": "image", + "image_path": "eff3da90302496f2f5287b95ee4853e25e7c8e65d25fed96b2d7ca11422f39c0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 252, + 445, + 259 + ], + "lines": [ + { + "bbox": [ + 414, + 252, + 445, + 259 + ], + "spans": [ + { + "bbox": [ + 414, + 252, + 445, + 259 + ], + "type": "text", + "content": "UAPGD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 260, + 389, + 332 + ], + "blocks": [ + { + "bbox": [ + 331, + 260, + 389, + 332 + ], + "lines": [ + { + "bbox": [ + 331, + 260, + 389, + 332 + ], + "spans": [ + { + "bbox": [ + 331, + 260, + 389, + 332 + ], + "type": "image", + "image_path": "1183652b95f0fa74a19b6e9a11fd4eff6384308eb44440906d6254e24f780695.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 334, + 366, + 342 + ], + "lines": [ + { + "bbox": [ + 354, + 334, + 366, + 342 + ], + "spans": [ + { + "bbox": [ + 354, + 334, + 366, + 342 + ], + "type": "text", + "content": "FG" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": "Fig. 3: The table on the left displays the ASR values of triggers, which are generated by various settings, on the training and testing sets of the COCO dataset. The values in parentheses represent the attack step size, and we employ different " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}^{i}" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": " values for YOLOv3 and YOLOv5. For YOLOv3, we set " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}^{1} = 0.1" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}^{2} = 0.5" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}^{3} = 1.0" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": "; for YOLOv5, we set " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}^{1} = 0.05" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}^{2} = 0.1" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{FG}^{3} = 0.2" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": ". Ensemble denotes the combined usage of FG and UAPGD. Dpatch2 is the original approach from [32] and Dpatch2(" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": ") denotes the original approach adopts the " + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "inline_equation", + "content": "l_{det}" + }, + { + "bbox": [ + 130, + 352, + 482, + 451 + ], + "type": "text", + "content": " as the optimization objective. The bolded values indicate the optimal results for each respective column in the table. There are some results of triggers on the right generated under different settings." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 402, + 260, + 458, + 332 + ], + "blocks": [ + { + "bbox": [ + 402, + 260, + 458, + 332 + ], + "lines": [ + { + "bbox": [ + 402, + 260, + 458, + 332 + ], + "spans": [ + { + "bbox": [ + 402, + 260, + 458, + 332 + ], + "type": "image", + "image_path": "72a324c1c3af2b29d9de062c92e1853f662fd94512d7fdfcf55c36d481b79427.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 334, + 447, + 342 + ], + "lines": [ + { + "bbox": [ + 413, + 334, + 447, + 342 + ], + "spans": [ + { + "bbox": [ + 413, + 334, + 447, + 342 + ], + "type": "text", + "content": "Ensemble" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "content": "As shown in Figure 3, We compare our method with the more reasonable and advanced Dpatch2 [32]. Considering the divergent optimization objectives, we substitute the maximization of the model loss function mentioned in [32] with the minimization of the proposed " + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "content": " referred to as Dpatch2(" + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "content": "). Undoubtedly, our method achieves the highest ASR in different experimental settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 602 + ], + "type": "text", + "content": "A comparative analysis of the effectiveness between PGD and UAPGD reveals notable differences. Specifically, we observe that UAPGD, when employed with an identical attack step size as PGD, frequently results in a higher ASR. Moreover, it can be observed that UAPGD yields better performance compared to adaptive step size optimizers like Adam." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Figure 4a and Figure 4b present a detailed depiction of the variations in the detection loss " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " throughout the training process when generating adversarial triggers using both PGD and UAPGD techniques, specifically over YOLOv3 and YOLOv5 models. Notably, when operating under an equivalent number of epochs, UAPGD demonstrates a more pronounced capability to reduce " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{det}" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 119, + 220, + 171 + ], + "blocks": [ + { + "bbox": [ + 151, + 119, + 220, + 171 + ], + "lines": [ + { + "bbox": [ + 151, + 119, + 220, + 171 + ], + "spans": [ + { + "bbox": [ + 151, + 119, + 220, + 171 + ], + "type": "image", + "image_path": "216198102cf21938cf28389a83992fa67fac2aa58b3dce6a9b8db735f7b00ce6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 174, + 212, + 183 + ], + "lines": [ + { + "bbox": [ + 165, + 174, + 212, + 183 + ], + "spans": [ + { + "bbox": [ + 165, + 174, + 212, + 183 + ], + "type": "text", + "content": "(a) YOLOv3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 119, + 299, + 171 + ], + "blocks": [ + { + "bbox": [ + 230, + 119, + 299, + 171 + ], + "lines": [ + { + "bbox": [ + 230, + 119, + 299, + 171 + ], + "spans": [ + { + "bbox": [ + 230, + 119, + 299, + 171 + ], + "type": "image", + "image_path": "c08f344bdf0c70c5ec8686666c02d292668611c99437150d75fd5b2f1305ec09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 174, + 291, + 183 + ], + "lines": [ + { + "bbox": [ + 244, + 174, + 291, + 183 + ], + "spans": [ + { + "bbox": [ + 244, + 174, + 291, + 183 + ], + "type": "text", + "content": "(b) YOLOv5" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "type": "text", + "content": "Fig. 4: (a) and (b) illustrate the variations of " + }, + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "type": "text", + "content": ", during the training process of generating adversarial triggers using PGD and UAPGD for attacking the YOLOv3 and YOLOv5 models. (c) and (d) present the confidence of some samples from the COCO dataset, outputted by the detector, during the training process of individual attacks. The blue, green, and red curved lines represent the utilization of " + }, + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "type": "inline_equation", + "content": "L_{FG}" + }, + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 131, + 192, + 482, + 269 + ], + "type": "text", + "content": ", and both as the loss function respectively. For individual attacks, the optimization process is relatively easier so that step size decay is unnecessary." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 119, + 378, + 171 + ], + "blocks": [ + { + "bbox": [ + 310, + 119, + 378, + 171 + ], + "lines": [ + { + "bbox": [ + 310, + 119, + 378, + 171 + ], + "spans": [ + { + "bbox": [ + 310, + 119, + 378, + 171 + ], + "type": "image", + "image_path": "b3b0d2844550071cd4739d33a5a239e6198c81c6f3fcc7113b7a1a95a1bb235b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 174, + 352, + 183 + ], + "lines": [ + { + "bbox": [ + 340, + 174, + 352, + 183 + ], + "spans": [ + { + "bbox": [ + 340, + 174, + 352, + 183 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 120, + 459, + 171 + ], + "blocks": [ + { + "bbox": [ + 389, + 120, + 459, + 171 + ], + "lines": [ + { + "bbox": [ + 389, + 120, + 459, + 171 + ], + "spans": [ + { + "bbox": [ + 389, + 120, + 459, + 171 + ], + "type": "image", + "image_path": "fc29a5f10440eede6d74adbed270de542c47a7e05ec61406d34ee959305ee6e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 174, + 432, + 183 + ], + "lines": [ + { + "bbox": [ + 419, + 174, + 432, + 183 + ], + "spans": [ + { + "bbox": [ + 419, + 174, + 432, + 183 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 480, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 317 + ], + "type": "text", + "content": "compared to PGD. The effectiveness of the loss monitoring in UAPGD will be analyzed in detail in Appendix E." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 318, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 389 + ], + "type": "text", + "content": "Furthermore, as shown in Figure 3, UAPGD exhibits a superior capability in generating smoother adversarial triggers. This increased smoothness can be attributed to the implementation of step size decay, which allows for a more refined optimization process in the generation of triggers. Consequently, adversarial triggers crafted via UAPGD display enhanced suitability for practical applications, such as printing and deployment in the physical world." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "text", + "content": "Based on results in Figure 3, we can observe that integrating " + }, + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "inline_equation", + "content": "L_{FG}" + }, + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "text", + "content": " with suitable step size and weight into the loss function, coupled with using either PGD or UAPGD, can enhance the ASR. This improvement is particularly noticeable when applied to the YOLOv5 model. This observation highlights that employing appropriate feature guidance can improve the performance of universal attacks. However, incorporating FG did not always achieve satisfactory results under certain weights and step sizes. This phenomenon is because in universal attacks, the inclusion of " + }, + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "inline_equation", + "content": "L_{FG}" + }, + { + "bbox": [ + 130, + 390, + 481, + 545 + ], + "type": "text", + "content": " introduces additional complexities to the optimization process. It requires appropriate adjustment of hyperparameters to fully leverage its effectiveness. Hence, the optimal outcome can be achieved by ensemble UAPGD with it, which allows for finer control over the step size to facilitate the optimization process. The empirical results indicate that the combination of UAPGD and FG in generating triggers yields the most effective attack performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 664 + ], + "type": "text", + "content": "To further investigate the potential role of Feature Guidance, we conducted individual adversarial attacks on each image with an adversarial trigger (fixed step size of 8/255, 500iteration) in the training dataset. We identified two primary roles of FG: 1, as shown in Figure 4c, our observations reveal that FG can significantly influence the optimization direction during the training phase. This influence is evidenced by the expedited convergence in adversarial trigger training, indicating the advantageous role of FG in enhancing training efficiency. 2, As shown in Figure 4d, when some images are challenging to attack with out-of-box adversarial triggers, introducing FG would enhance the attack effectiveness at this time." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 284, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 284, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 284, + 128 + ], + "type": "text", + "content": "4.2 Experiment in simulator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 332, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 332, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 332, + 209 + ], + "type": "text", + "content": "In this section, we demonstrate the effects of our attack in autonomous driving scenarios, we use the Carla(0.9.14) simulator [14] to simulate the real driving process. In the simulator, we utilize the YOLOv3 model as the detection network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 334, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 334, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 334, + 317 + ], + "type": "text", + "content": "Due to the significant differences between images from the COCO dataset and those generated by the Carla simulator, the adversarial trigger developed in Section 4.1 is not directly applicable to Carla. To tailor an adversarial trigger specifically for the Carla simulator, we compile a training dataset comprising 242 images featuring stop signs from the simulator. The process of attacking the Carla simulator" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "text", + "content": "is conducted in four distinct manners: employing either PGD or UAPGD, combined with a loss function of either " + }, + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "inline_equation", + "content": "L_{det} + \\lambda_{tv}L_{tv}" + }, + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "inline_equation", + "content": "L_{all}" + }, + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "inline_equation", + "content": "\\lambda_{FG} = 0.1" + }, + { + "bbox": [ + 130, + 317, + 481, + 353 + ], + "type": "text", + "content": " and the initial step size is 16/255." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 333, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 333, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 333, + 460 + ], + "type": "text", + "content": "To evaluate the effectiveness of the attack, we implement a procedure where the adversarial trigger is placed under the stop sign in the simulator. Subsequently, a vehicle is navigated towards this stop sign, and a 9-second video is recorded using an onboard camera. The key metric is the percentage of time during which the detection system fails to recognize the stop sign." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 334, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 334, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 334, + 581 + ], + "type": "text", + "content": "As shown in Table 1, without any adversarial trigger, the car consistently recognizes the stop sign. When the attack is initiated using the PGD method alone, the attack duration is approximately 3 to 4 seconds. However, the integration of either UAPGD or FG methods into the attack strategy results in a slight increase in the attack duration by roughly an additional second. Notably, the implementation of an ensemble approach, which combines" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 604 + ], + "type": "text", + "content": "these methods, further enhances the attack performance, prolonging the duration to over 5 seconds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 640 + ], + "type": "text", + "content": "Moreover, to provide a more direct illustration of the attack impact, we have plotted the trend of the stop sign detection confidence as the car approaches it, shown as in Figure 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "The empirical results reveal that the efficacy of our attack method consistently surpasses that of the vanilla PGD approach. Particularly in the final 5" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 340, + 163, + 459, + 240 + ], + "blocks": [ + { + "bbox": [ + 340, + 163, + 459, + 240 + ], + "lines": [ + { + "bbox": [ + 340, + 163, + 459, + 240 + ], + "spans": [ + { + "bbox": [ + 340, + 163, + 459, + 240 + ], + "type": "table", + "html": "
attack methodProportion
No attack0%
PGD36.67%
UAPGD45.56%
PGD,FG46.67%
UAPGD,FG58.89%
", + "image_path": "5f2afb937c2977ca3cbc7dc45959a5ebf3204655db07389b357e21f189570558.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 249, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 339, + 249, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 339, + 249, + 482, + 294 + ], + "type": "text", + "content": "Table 1: The proportion of time in the video duration during which the detector failed to recognize the stop sign." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 353, + 388, + 462, + 472 + ], + "blocks": [ + { + "bbox": [ + 353, + 388, + 462, + 472 + ], + "lines": [ + { + "bbox": [ + 353, + 388, + 462, + 472 + ], + "spans": [ + { + "bbox": [ + 353, + 388, + 462, + 472 + ], + "type": "image", + "image_path": "1573d79a60744cdd7bd24480190fc083636faa2152ec8561a176d560f8381904.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 481, + 482, + 558 + ], + "lines": [ + { + "bbox": [ + 339, + 481, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 339, + 481, + 482, + 558 + ], + "type": "text", + "content": "Fig.5: We uniformly extracted approximately 90 frames from the 9s video, which is recorded by the simulated vehicle's onboard camera, to showcase the confidence variations of the detector's predictions" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 187, + 166 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 187, + 166 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 187, + 166 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 187, + 166 + ], + "type": "image", + "image_path": "b8e45a5dd8cba39c41d276472d2bc1939fb261b8a1afc662ea6fdb21dbf8a847.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 114, + 246, + 166 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 246, + 166 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 246, + 166 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 246, + 166 + ], + "type": "image", + "image_path": "020d5516646b7171a2f025ba9ad8498d13860622eee940bafb812b165d50dc88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 252, + 114, + 304, + 166 + ], + "blocks": [ + { + "bbox": [ + 252, + 114, + 304, + 166 + ], + "lines": [ + { + "bbox": [ + 252, + 114, + 304, + 166 + ], + "spans": [ + { + "bbox": [ + 252, + 114, + 304, + 166 + ], + "type": "image", + "image_path": "fb46028851858f4f27fb21816c751721a2a879e2c011fd88dc4a393528d58713.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 361, + 166 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 361, + 166 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 361, + 166 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 361, + 166 + ], + "type": "image", + "image_path": "80a91c18a7098f6271510abfd1fa494a7ff388a763d416c0b350cc715b965be0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 419, + 166 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 419, + 166 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 419, + 166 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 419, + 166 + ], + "type": "image", + "image_path": "7ce192eb37439234d236016e572a1cd17a104cec4ad0de0ac13780167dd3ab21.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 114, + 477, + 166 + ], + "blocks": [ + { + "bbox": [ + 427, + 114, + 477, + 166 + ], + "lines": [ + { + "bbox": [ + 427, + 114, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 427, + 114, + 477, + 166 + ], + "type": "image", + "image_path": "d505e6ab38c8190e4c4f9fcfded3417de960d5ed23cfa3b2c09c09fc92656142.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 136, + 175, + 187, + 227 + ], + "blocks": [ + { + "bbox": [ + 139, + 167, + 184, + 175 + ], + "lines": [ + { + "bbox": [ + 139, + 167, + 184, + 175 + ], + "spans": [ + { + "bbox": [ + 139, + 167, + 184, + 175 + ], + "type": "inline_equation", + "content": "conf = 0.98" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 175, + 187, + 227 + ], + "lines": [ + { + "bbox": [ + 136, + 175, + 187, + 227 + ], + "spans": [ + { + "bbox": [ + 136, + 175, + 187, + 227 + ], + "type": "image", + "image_path": "57c7373b99afc96b087ddd9f70348bdec3b104ad0b8ca2f8b3dda05b6a103985.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 194, + 175, + 246, + 227 + ], + "blocks": [ + { + "bbox": [ + 198, + 168, + 242, + 175 + ], + "lines": [ + { + "bbox": [ + 198, + 168, + 242, + 175 + ], + "spans": [ + { + "bbox": [ + 198, + 168, + 242, + 175 + ], + "type": "inline_equation", + "content": "conf = 0.98" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 175, + 246, + 227 + ], + "lines": [ + { + "bbox": [ + 194, + 175, + 246, + 227 + ], + "spans": [ + { + "bbox": [ + 194, + 175, + 246, + 227 + ], + "type": "image", + "image_path": "97daaf357441bc729c5b8bdd4ff84d0737309f6b31b0be13e138b497536c93aa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 252, + 175, + 304, + 227 + ], + "blocks": [ + { + "bbox": [ + 255, + 167, + 300, + 175 + ], + "lines": [ + { + "bbox": [ + 255, + 167, + 300, + 175 + ], + "spans": [ + { + "bbox": [ + 255, + 167, + 300, + 175 + ], + "type": "inline_equation", + "content": "conf = 0.99" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 175, + 304, + 227 + ], + "lines": [ + { + "bbox": [ + 252, + 175, + 304, + 227 + ], + "spans": [ + { + "bbox": [ + 252, + 175, + 304, + 227 + ], + "type": "image", + "image_path": "dce6c46738c2e26110444f0f82fc486cb2f788ea3d111490695d1b6ddf6e8e62.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 175, + 361, + 227 + ], + "blocks": [ + { + "bbox": [ + 317, + 167, + 356, + 175 + ], + "lines": [ + { + "bbox": [ + 317, + 167, + 356, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 356, + 175 + ], + "type": "inline_equation", + "content": "conf = 1.0" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 175, + 361, + 227 + ], + "lines": [ + { + "bbox": [ + 310, + 175, + 361, + 227 + ], + "spans": [ + { + "bbox": [ + 310, + 175, + 361, + 227 + ], + "type": "image", + "image_path": "71b36cb6d44365d25874b895739252d1a378d99a409eca9e863f79db513acdee.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 369, + 175, + 419, + 227 + ], + "blocks": [ + { + "bbox": [ + 375, + 167, + 413, + 175 + ], + "lines": [ + { + "bbox": [ + 375, + 167, + 413, + 175 + ], + "spans": [ + { + "bbox": [ + 375, + 167, + 413, + 175 + ], + "type": "inline_equation", + "content": "conf = 1.0" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 175, + 419, + 227 + ], + "lines": [ + { + "bbox": [ + 369, + 175, + 419, + 227 + ], + "spans": [ + { + "bbox": [ + 369, + 175, + 419, + 227 + ], + "type": "image", + "image_path": "c26c9dd08335a303d179cbbe2630761be8f4ab3f48ad6b73f2b08276409c32fb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 427, + 175, + 477, + 227 + ], + "blocks": [ + { + "bbox": [ + 432, + 167, + 471, + 175 + ], + "lines": [ + { + "bbox": [ + 432, + 167, + 471, + 175 + ], + "spans": [ + { + "bbox": [ + 432, + 167, + 471, + 175 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 432, + 167, + 471, + 175 + ], + "type": "inline_equation", + "content": "= 1.0" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 175, + 477, + 227 + ], + "lines": [ + { + "bbox": [ + 427, + 175, + 477, + 227 + ], + "spans": [ + { + "bbox": [ + 427, + 175, + 477, + 227 + ], + "type": "image", + "image_path": "dbfe6e49f9e35fdffe1203b773593c2b106ede580426fa9700678f2f185f6cad.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 136, + 237, + 187, + 289 + ], + "blocks": [ + { + "bbox": [ + 139, + 229, + 184, + 236 + ], + "lines": [ + { + "bbox": [ + 139, + 229, + 184, + 236 + ], + "spans": [ + { + "bbox": [ + 139, + 229, + 184, + 236 + ], + "type": "text", + "content": "conf = 0.94" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 237, + 187, + 289 + ], + "lines": [ + { + "bbox": [ + 136, + 237, + 187, + 289 + ], + "spans": [ + { + "bbox": [ + 136, + 237, + 187, + 289 + ], + "type": "image", + "image_path": "b29838f6ef7e6212ab2899bc44bab014f8952a5a327e3a5136d28f056e1fd0ce.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 291, + 184, + 299 + ], + "lines": [ + { + "bbox": [ + 139, + 291, + 184, + 299 + ], + "spans": [ + { + "bbox": [ + 139, + 291, + 184, + 299 + ], + "type": "text", + "content": "conf = 0.82" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 308, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 343 + ], + "type": "text", + "content": "Fig. 6: Six frames are extracted uniformly from the video. Row 1 represents a clean stop sign with the detector's output. In Row 2 and Row 3, we deploy the trigger generated by the PGD and ensemble method respectively." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 194, + 237, + 246, + 289 + ], + "blocks": [ + { + "bbox": [ + 197, + 229, + 242, + 236 + ], + "lines": [ + { + "bbox": [ + 197, + 229, + 242, + 236 + ], + "spans": [ + { + "bbox": [ + 197, + 229, + 242, + 236 + ], + "type": "text", + "content": "conf = 0.72" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 237, + 246, + 289 + ], + "lines": [ + { + "bbox": [ + 194, + 237, + 246, + 289 + ], + "spans": [ + { + "bbox": [ + 194, + 237, + 246, + 289 + ], + "type": "image", + "image_path": "3211e5fbf980c00e5dec0844fc08ee8a80e3bce9781c29159e934f983d16ffdd.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 291, + 239, + 298 + ], + "lines": [ + { + "bbox": [ + 199, + 291, + 239, + 298 + ], + "spans": [ + { + "bbox": [ + 199, + 291, + 239, + 298 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 199, + 291, + 239, + 298 + ], + "type": "inline_equation", + "content": "< 0.5" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 252, + 237, + 304, + 289 + ], + "blocks": [ + { + "bbox": [ + 255, + 229, + 299, + 236 + ], + "lines": [ + { + "bbox": [ + 255, + 229, + 299, + 236 + ], + "spans": [ + { + "bbox": [ + 255, + 229, + 299, + 236 + ], + "type": "text", + "content": "conf = 0.70" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 237, + 304, + 289 + ], + "lines": [ + { + "bbox": [ + 252, + 237, + 304, + 289 + ], + "spans": [ + { + "bbox": [ + 252, + 237, + 304, + 289 + ], + "type": "image", + "image_path": "f2f1a8163f9747eb6f4e979e25ffde3526c132a9cda55c3dc569cb7deb73a46a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 291, + 298, + 298 + ], + "lines": [ + { + "bbox": [ + 257, + 291, + 298, + 298 + ], + "spans": [ + { + "bbox": [ + 257, + 291, + 298, + 298 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 257, + 291, + 298, + 298 + ], + "type": "inline_equation", + "content": "< 0.5" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 310, + 237, + 361, + 289 + ], + "blocks": [ + { + "bbox": [ + 316, + 229, + 356, + 236 + ], + "lines": [ + { + "bbox": [ + 316, + 229, + 356, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 356, + 236 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 316, + 229, + 356, + 236 + ], + "type": "inline_equation", + "content": "< 0.5" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 237, + 361, + 289 + ], + "lines": [ + { + "bbox": [ + 310, + 237, + 361, + 289 + ], + "spans": [ + { + "bbox": [ + 310, + 237, + 361, + 289 + ], + "type": "image", + "image_path": "a6cad8c464753a06f75fe998044cb74428ce366d2df841b1b724f1e07627e41c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 291, + 355, + 298 + ], + "lines": [ + { + "bbox": [ + 316, + 291, + 355, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 355, + 298 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 316, + 291, + 355, + 298 + ], + "type": "inline_equation", + "content": "< 0.5" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 369, + 237, + 419, + 289 + ], + "blocks": [ + { + "bbox": [ + 374, + 229, + 414, + 236 + ], + "lines": [ + { + "bbox": [ + 374, + 229, + 414, + 236 + ], + "spans": [ + { + "bbox": [ + 374, + 229, + 414, + 236 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 374, + 229, + 414, + 236 + ], + "type": "inline_equation", + "content": "< 0.5" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 237, + 419, + 289 + ], + "lines": [ + { + "bbox": [ + 369, + 237, + 419, + 289 + ], + "spans": [ + { + "bbox": [ + 369, + 237, + 419, + 289 + ], + "type": "image", + "image_path": "2fdc8d2b620c383bfc67247a2a0dc97fcde66944e285c05397d3375f7f78f700.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 291, + 413, + 298 + ], + "lines": [ + { + "bbox": [ + 374, + 291, + 413, + 298 + ], + "spans": [ + { + "bbox": [ + 374, + 291, + 413, + 298 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 374, + 291, + 413, + 298 + ], + "type": "inline_equation", + "content": "< 0.5" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 427, + 237, + 477, + 289 + ], + "blocks": [ + { + "bbox": [ + 432, + 229, + 471, + 236 + ], + "lines": [ + { + "bbox": [ + 432, + 229, + 471, + 236 + ], + "spans": [ + { + "bbox": [ + 432, + 229, + 471, + 236 + ], + "type": "inline_equation", + "content": "conf = 0.6" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 237, + 477, + 289 + ], + "lines": [ + { + "bbox": [ + 427, + 237, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 427, + 237, + 477, + 289 + ], + "type": "image", + "image_path": "30b02c088302f1994a6b46405fe11ca2f32ee3811b98ca22c95710c304108388.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 291, + 471, + 298 + ], + "lines": [ + { + "bbox": [ + 432, + 291, + 471, + 298 + ], + "spans": [ + { + "bbox": [ + 432, + 291, + 471, + 298 + ], + "type": "text", + "content": "conf " + }, + { + "bbox": [ + 432, + 291, + 471, + 298 + ], + "type": "inline_equation", + "content": "< 0.5" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 367, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 480, + 403 + ], + "type": "text", + "content": "seconds of the video, during which the adversarial trigger persistently causes the detection system to lose sight of the stop sign. This effect illustrates the potential of our method to compromise the functionality of autonomous driving systems." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 130, + 403, + 480, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 428 + ], + "type": "text", + "content": "As shown in the Figure 6, we extract six frames uniformly from the video as examples to demonstrate the effects of this attack in driving scenarios." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 524 + ], + "type": "text", + "content": "The observations imply that normally the car will always recognize the stop sign (Row 1 of Figure 6, each stop is marked by a red box). However, after the adversarial trigger created by our attack method (Row 3) is added to the stop sign, the car fails to recognize the stop sign(no red box). Compared with the attack effect of the adversarial trigger generated by PGD (Row 2), it can be seen that our method is clearly superior to the PGD. It is worth mentioning that the perturbation generated by Dpatch2 is suboptimal in the simulator, as demonstrated in Appendix F." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 131, + 545, + 354, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 354, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 354, + 558 + ], + "type": "text", + "content": "4.3 Robustness of our attack in real world" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 130, + 569, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 480, + 616 + ], + "type": "text", + "content": "To observe the attack effectiveness and robustness of the universal adversarial trigger obtained through algorithm 1 in the real world, we print out the adversarial trigger and post it under a stop sign. Moreover, in the case where clean images serve as a reference, we compare our method with the normal PGD." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We examined the variation trends in the confidence of the detector's output for recognizing the stop sign using an onboard camera under different conditions: as the distance between the onboard camera and the stop sign decreased from far to near, and as the onboard camera's perspective towards the stop sign shifted" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 448, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 120, + 185, + 219 + ], + "blocks": [ + { + "bbox": [ + 145, + 120, + 185, + 219 + ], + "lines": [ + { + "bbox": [ + 145, + 120, + 185, + 219 + ], + "spans": [ + { + "bbox": [ + 145, + 120, + 185, + 219 + ], + "type": "image", + "image_path": "dc06d519f6941854aa3b1382d908a39aac8d08daaa9d114fe402187acf191d5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "type": "text", + "content": "Fig. 7: We print and deploy the adversarial trigger in the physical world. (a) illustrates the confidence variation of the stop sign as the distance increases. (b) depicts the confidence variation of the stop sign when facing different angles changing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 200, + 125, + 324, + 218 + ], + "blocks": [ + { + "bbox": [ + 200, + 125, + 324, + 218 + ], + "lines": [ + { + "bbox": [ + 200, + 125, + 324, + 218 + ], + "spans": [ + { + "bbox": [ + 200, + 125, + 324, + 218 + ], + "type": "image", + "image_path": "5f29f0e1b609f41b75d9cd1e99a879bf61b186d69656e2d35881f4567dea5563.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 221, + 272, + 231 + ], + "lines": [ + { + "bbox": [ + 260, + 221, + 272, + 231 + ], + "spans": [ + { + "bbox": [ + 260, + 221, + 272, + 231 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 342, + 125, + 465, + 218 + ], + "blocks": [ + { + "bbox": [ + 342, + 125, + 465, + 218 + ], + "lines": [ + { + "bbox": [ + 342, + 125, + 465, + 218 + ], + "spans": [ + { + "bbox": [ + 342, + 125, + 465, + 218 + ], + "type": "image", + "image_path": "4a4ffe0d3a4463fc4445b23c01801f7b5c8af1d023e8b6742f67de06f022499b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 220, + 413, + 230 + ], + "lines": [ + { + "bbox": [ + 400, + 220, + 413, + 230 + ], + "spans": [ + { + "bbox": [ + 400, + 220, + 413, + 230 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "text", + "content": "from -70-degree side view to a 70-degree side view. The real-world experiment results are shown in Figure 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 321, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 452 + ], + "type": "text", + "content": "According to Figure 7a, due to the relatively long distance between the vehicle and the stop position in the first half of the curve, the trigger occupies a small proportion of the field of view, which diminishes the stability of successful attacks. In the second half of the curve, where the distance between the stop and the vehicle is closer, the attacks become more stable. Our method consistently maintains a higher success rate in this scenario. Examining Figure 7b, it can be observed that when facing the stop sign directly, our method misleads the detector to output a remarkably low confidence. As the angle approaches approximately " + }, + { + "bbox": [ + 130, + 321, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\pm 50" + }, + { + "bbox": [ + 130, + 321, + 482, + 452 + ], + "type": "text", + "content": " degrees, our method becomes ineffective. In contrast, the triggers generated by PGD are not smooth, which prevents them from effectively imprinting adversarial features into the real world." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 453, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 501 + ], + "type": "text", + "content": "Overall, in the real world, our attack consistently exhibits robust and effective performance when the onboard camera is approaching a stop sign. This means that our attack can successfully mislead an autonomous driving system to disregard the stop sign unconsciously." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 520, + 220, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 520, + 220, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 520, + 220, + 533 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "In this work, we present a novel adversarial attack methodology that extends beyond the bounding boxes of objects. This method poses greater challenges compared to previous studies due to its tendency to focus on less informative areas. Thus, we propose the FG techniques and UAPGD optimization strategy, specifically designed to augment the efficacy of this attack in both digital and physical realms. Furthermore, we validate the effectiveness of our attack method through its application in autonomous driving scenarios and real-world conditions. Discussion of limitations and potential negative impact of our works will be placed in Appendix G and immediately following that we also briefly address future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 222 + ], + "type": "text", + "content": "We thank the anonymous referees for their valuable and insightful comments. This work is supported by the National Key R&D Program of China under Grant No. 2022YFA1005100, 2022YFA1005101, and 2022YFA1005104, the CAS Project for Young Scientists in Basic Research, under Grant No.YSBR-040, ISCAS New Cultivation Project ISCAS-PYFX-202201, ISCAS Basic Research ISCAS-JCZD-202302, and the Major Project of ISCAS under Grant No. ISCAS-ZD-202302." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 262, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 307 + ], + "type": "text", + "content": "1. Aich, A., Li, S., Song, C., Asif, M.S., Krishnamurthy, S.V., Roy-Chowdhury, A.K.: Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1308-1318 (2023) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 307, + 481, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 481, + 339 + ], + "type": "text", + "content": "2. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 339, + 481, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 481, + 372 + ], + "type": "text", + "content": "3. Brendel, W., Rauber, J., Bethge, M.: Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248 (2017) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 372, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 481, + 415 + ], + "type": "text", + "content": "4. Cao, Y., Xiao, C., Cyr, B., Zhou, Y., Park, W., Rampazzi, S., Chen, Q.A., Fu, K., Mao, Z.M.: Adversarial sensor attack on lidar-based perception in autonomous driving. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 2267-2281 (2019) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 415, + 481, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 481, + 437 + ], + "type": "text", + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 437, + 481, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 470 + ], + "type": "text", + "content": "6. Chen, C., Seff, A., Kornhauser, A., Xiao, J.: Deepdriving: Learning affordance for direct perception in autonomous driving. In: Proceedings of the IEEE international conference on computer vision. pp. 2722-2730 (2015) 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "type": "text", + "content": "7. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 IEEE symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 502, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 546 + ], + "type": "text", + "content": "8. Chen, P.Y., Zhang, H., Sharma, Y., Yi, J., Hsieh, C.J.: Zoo: Zeroth order optimization based black-box attacks to deep neural networks without training substitute models. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 15-26 (2017) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "type": "text", + "content": "9. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216. PMLR (2020) 3, 4, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "text", + "content": "0. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "1. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "2. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018) 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "13. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "14. Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: Carla: An open urban driving simulator. In: Conference on robot learning. pp. 1-16. PMLR (2017) 3, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "type": "text", + "content": "15. Duan, R., Ma, X., Wang, Y., Bailey, J., Qin, A.K., Yang, Y.: Adversarial camouflage: Hiding physical-world attacks with natural styles. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1000-1008 (2020) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "16. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018) 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "18. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "19. Hoory, S., Shapira, T., Shabtai, A., Elovici, Y.: Dynamic adversarial patch for evading object detection models. arXiv preprint arXiv:2010.13070 (2020) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "text", + "content": "20. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7848-7857 (2021) 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "text", + "content": "21. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13307-13316 (2022) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "text", + "content": "22. Huang, L., Gao, C., Zhou, Y., Xie, C., Yuille, A.L., Zou, C., Liu, N.: Universal physical camouflage attacks on object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 720-729 (2020) 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "23. Huang, W., Zhao, X., Jin, G., Huang, X.: Safari: Versatile and efficient evaluations for robustness of interpretability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1988-1998 (2023) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "text", + "content": "24. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "25. Im Choi, J., Tian, Q.: Adversarial attack and defense of yolo detectors in autonomous driving scenarios. In: 2022 IEEE Intelligent Vehicles Symposium (IV). pp. 1011-1017. IEEE (2022) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "26. Jia, S., Yin, B., Yao, T., Ding, S., Shen, C., Yang, X., Ma, C.: Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. Advances in Neural Information Processing Systems 35, 34136-34147 (2022) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "27. Jin, G., Yi, X., Huang, W., Schewe, S., Huang, X.: Enhancing adversarial training with second-order statistics of weights. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15273-15283 (2022) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "28. Jin, G., Yi, X., Wu, D., Mu, R., Huang, X.: Randomized adversarial training via taylor expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16447-16457 (2023) 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Michael, K., Fang, J., Yifu, Z., Wong, C., Montes, D., et al.: ultralytics/yolov5: v7. 0-yolov5 sota realtime instance segmentation. Zenodo (2022) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "30. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15045-15053 (2022) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 202 + ], + "type": "text", + "content": "31. Kurakin, A., Goodfellow, I., Bengio, S., et al.: Adversarial examples in the physical world (2016) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 202, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 224 + ], + "type": "text", + "content": "32. Lee, M., Kolter, Z.: On physical adversarial patches for object detection. arXiv preprint arXiv:1906.11897 (2019) 4, 10, 11, 16" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "type": "text", + "content": "33. Lei, X., Cai, X., Lu, C., Jiang, Z., Gong, Z., Lu, L.: Using frequency attention to make adversarial patch powerful against person detector. IEEE Access 11, 27217-27225 (2022) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "text", + "content": "34. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4282-4291 (2019) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "type": "text", + "content": "35. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8971-8980 (2018) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 320, + 482, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 353 + ], + "type": "text", + "content": "36. Li, J., Schmidt, F., Kolter, Z.: Adversarial camera stickers: A physical camera-based attack on deep learning systems. In: International Conference on Machine Learning. pp. 3896-3904. PMLR (2019) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 353, + 482, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 385 + ], + "type": "text", + "content": "37. Li, Y., Li, Y., Dai, X., Guo, S., Xiao, B.: Physical-world optical adversarial attacks on 3d face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24699-24708 (2023) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 385, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 418 + ], + "type": "text", + "content": "38. Liang, S., Wu, B., Fan, Y., Wei, X., Cao, X.: Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970 (2022) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 418, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 460 + ], + "type": "text", + "content": "39. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 3, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "type": "text", + "content": "40. Liu, A., Liu, X., Fan, J., Ma, Y., Zhang, A., Xie, H., Tao, D.: Perceptual-sensitive gan for generating adversarial patches. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1028-1035 (2019) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 493, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 536 + ], + "type": "text", + "content": "41. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 536, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 557 + ], + "type": "text", + "content": "42. Liu, X., Yang, H., Liu, Z., Song, L., Li, H., Chen, Y.: Dpatch: An adversarial patch attack on object detectors. arXiv preprint arXiv:1806.02299 (2018) 4, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "text", + "content": "43. Lovisotto, G., Turner, H., Sluganovic, I., Strohmeier, M., Martinovic, I.: {SLAP}: Improving physical adversarial examples with {Short-Lived} adversarial perturbations. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1865-1882 (2021) 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "text", + "content": "44. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) 3, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "45. Mahendran, A., Vedaldi, A.: Understanding deep image representations by inverting them. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5188-5196 (2015) 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "47. Pomponi, J., Scardapane, S., Uncini, A.: Pixle: a fast and effective black-box attack based on rearranging pixels. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 1-7. IEEE (2022) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "48. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "49. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "text", + "content": "50. Shapira, A., Zolfi, A., Demetrio, L., Biggio, B., Shabtai, A.: Phantom sponges: Exploiting non-maximum suppression to attack deep object detectors. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4571-4580 (2023) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "text", + "content": "51. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In: Proceedings of the 2016 acm sig sac conference on computer and communications security. pp. 1528-1540 (2016) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "52. Shi, Z., Yang, W., Xu, Z., Yu, Z., Huang, L.: Reinforcement learning-based adversarial attacks on object detectors using reward shaping. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8424-8432 (2023) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "53. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "text", + "content": "54. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "type": "text", + "content": "55. Tu, J., Ren, M., Manivasagam, S., Liang, M., Yang, B., Du, R., Cheng, F., Urtasun, R.: Physically realizable adversarial examples for lidar object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13716-13725 (2020) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 523 + ], + "type": "text", + "content": "56. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2414-2422 (2022) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "57. Wang, D., Yao, W., Jiang, T., Li, C., Chen, X.: Rfla: A stealthy reflected light adversarial attack in the physical world. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4455-4465 (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 589 + ], + "type": "text", + "content": "58. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265-5274 (2018) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "59. Wang, J., Liu, A., Yin, Z., Liu, S., Tang, S., Liu, X.: Dual attention suppression attack: Generate adversarial camouflage in physical world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8565-8574 (2021) 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "60. Wang, Q., Zhang, L., Bertinetto, L., Hu, W., Torr, P.H.: Fast online object tracking and segmentation: A unifying approach. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 1328-1338 (2019) 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "T. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 632 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "61. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "62. Wu, D., Wang, Y., Xia, S.T., Bailey, J., Ma, X.: Skip connections matter: On the transferability of adversarial examples generated with resnets. arXiv preprint arXiv:2002.05990 (2020) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "text", + "content": "63. Wu, W., Su, Y., Chen, X., Zhao, S., King, I., Lyu, M.R., Tai, Y.W.: Boosting the transferability of adversarial samples via attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1161-1170 (2020) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "64. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "type": "text", + "content": "65. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 665-681. Springer (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 358 + ], + "type": "text", + "content": "66. Zhang, J., Wu, W., Huang, J.t., Huang, Y., Wang, W., Su, Y., Lyu, M.R.: Improving adversarial transferability via neuron attribution-based attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14993-15002 (2022) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "text", + "content": "67. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15115-15125 (2022) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 390, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 435 + ], + "type": "text", + "content": "68. Zhang, L., Xu, N., Yang, P., Jin, G., Huang, C.C., Zhang, L.: Trajpac: Towards robustness verification of pedestrian trajectory prediction models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8327-8339 (2023) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 435, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 479 + ], + "type": "text", + "content": "69. Zhao, Y., Zhu, H., Liang, R., Shen, Q., Zhang, S., Chen, K.: Seeing isn't believing: Towards more robust adversarial attack against real world object detectors. In: Proceedings of the 2019 ACM SIGSAC conference on computer and communications security. pp. 1989-2004 (2019) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "text", + "content": "70. Zhong, Y., Liu, X., Zhai, D., Jiang, J., Ji, X.: Shadows can be dangerous: Stealthy and effective physical-world adversarial attack by natural phenomenon. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15345-15354 (2022) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "71. Zhu, L., Wang, T., Li, J., Zhang, Z., Shen, J., Wang, X.: Efficient query-based black-box attack against cross-modal hashing retrieval. ACM Transactions on Information Systems 41(3), 1-25 (2023) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "72. Zhu, X., Li, X., Li, J., Wang, Z., Hu, X.: Fooling thermal infrared pedestrian detectors in real world using small bulbs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3616-3624 (2021) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "text", + "content": "73. Zolfi, A., Kravchik, M., Elovici, Y., Shabtai, A.: The translucent patch: A physical and universal attack on object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15232-15241 (2021) 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 323, + 91, + 447, + 102 + ], + "type": "text", + "content": "Out-of-Bounding-Box Triggers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_content_list.json b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ec7dfa1fc80bee8dcc79d8cd90df74d9681e9f9c --- /dev/null +++ b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_content_list.json @@ -0,0 +1,1834 @@ +[ + { + "type": "text", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyang Xiang $^{1}$ , Zuohui Chen $^{2}$ , Siqi Li $^{1}$ , Qing Wu $^{3}$ , and Yong Liu $^{1,4}$", + "bbox": [ + 223, + 210, + 779, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ APRIL Lab, Zhejiang University", + "bbox": [ + 382, + 239, + 617, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Zhejiang University of Technology", + "bbox": [ + 380, + 253, + 622, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 College of Computer Science, Hangzhou Dianzi University", + "bbox": [ + 300, + 267, + 700, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Huzhou Institute, Zhejiang University", + "bbox": [ + 367, + 281, + 633, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Binary Neural Networks (BNNs) have been proven to be highly effective for deploying deep neural networks on mobile and embedded platforms. Most existing works focus on minimizing quantization errors, improving representation ability, or designing gradient approximations to alleviate gradient mismatch in BNNs, while leaving the weight sign flipping, a critical factor for achieving powerful BNNs, untouched. In this paper, we investigate the efficiency of weight sign updates in BNNs. We observe that, for vanilla BNNs, over $50\\%$ of the weights remain their signs unchanged during training, and these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. We refer to these weights as \"silent weights\", which slow down convergence and lead to a significant accuracy degradation. Theoretically, we reveal this is due to the independence of the BNNs gradient from the latent weight distribution. To address the issue, we propose Overcome Silent Weights (OvSW). OvSW first employs Adaptive Gradient Scaling (AGS) to establish a relationship between the gradient and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Additionally, we design Silence Awareness Decaying (SAD) to automatically identify \"silent weights\" by tracking weight flipping state, and apply an additional penalty to \"silent weights\" to facilitate their flipping. By efficiently updating weight signs, our method achieves faster convergence and state-of-the-art performance on CIFAR10 and ImageNet1K dataset with various architectures. For example, OvSW obtains $61.6\\%$ and $65.5\\%$ top-1 accuracy on the ImageNet1K using binarized ResNet18 and ResNet34 architecture respectively. Codes are available at https://github.com/JingyangXiang/OvSW.", + "bbox": [ + 261, + 323, + 743, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Binary Neural Networks $\\cdot$ Silent Weights $\\cdot$ Adaptive Gradient Scaling $\\cdot$ Silence Awareness Decaying", + "bbox": [ + 259, + 709, + 740, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 758, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have shown tremendous success in various computer vision tasks, including image classification [15, 23], object detection [10, 14,", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: yongliu@iipc.zju.edu.cn.", + "bbox": [ + 217, + 825, + 550, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa0599de141f25a3e675a8e7cc2915ecd394dde387fde65361b145a01d95251f.jpg", + "image_caption": [ + "(a) Vanilla" + ], + "image_footnote": [], + "bbox": [ + 236, + 151, + 395, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3619b96a085883ae65264adcd23e8d5335158e452f910b9157e61590abc48866.jpg", + "image_caption": [ + "(b) OvSW", + "Fig. 1: (a) and (b): Histogram of the initialized weight distribution (blue) and the weights that never update signs throughout training (orange). $54.07\\%$ and $2.03\\%$ represent the ratio of the corresponding orange area to the blue. (c) Top-1 Accuracy (solid lines) and Flip Ratio (dashed lines) are for Vanilla (green) and OvSW (red). (a), (b) and Flip Ratio in (c) is for layer4(conv2.weight." + ], + "image_footnote": [], + "bbox": [ + 421, + 152, + 578, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/835fe892beb883b41ac6b6b6f4cf79680a2fe929679a91a43a7219685a569510.jpg", + "image_caption": [ + "(c) Top-1 Acc & Flip Ratio" + ], + "image_footnote": [], + "bbox": [ + 598, + 152, + 781, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "43], and semantic segmentation [12, 35]. However, the remarkable performance is always attributed to deeper and wider architectures [15, 46], which comes with expensive memory and computational overhead and makes it challenging to deploy DNNs on resource-constrained edge platforms.", + "bbox": [ + 212, + 378, + 784, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The community has been delving into model compression, which aims to reduce inference overhead for DNNs while preserving their performance. Typical techniques include, but are not limited to, efficient architecture design [7, 58], neural architecture search [8, 47], network pruning [28, 36], knowledge distillation [18, 27], and network quantization [4, 11, 61]. Among them, network quantization is widely suggested as a promising solution. It effectively enhances memory efficiency and execution speed on embedding platforms by reducing the weight and activation bits and replacing expensive floating-point arithmetic with relatively inexpensive fixed-point arithmetic. Extensive studies [11, 20, 54, 57] have demonstrated its effectiveness and yielded light and efficient DNNs.", + "bbox": [ + 212, + 441, + 787, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on studying binary neural networks (BNNs), which are an aggressive form of quantized neural networks (QNNs). BNNs binarize both weights and activations to discrete values $\\{+1, -1\\}$ , which can be represented by 1-bit and reduce memory usage by $32 \\times$ . On the other hand, by employing efficient bitwise operations instead of floating-point ones, BNNs also significantly reduce computation complexity, providing $58 \\times$ speedup as reported by XNOR-Net [42]. These characteristics make them well-suited for deployment on low-power embedding platforms, including FPGA, ASICs, and IoT devices [19].", + "bbox": [ + 212, + 594, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the application of BNNs is still limited due to constraining both weights and activations to 1-bit significantly reducing accuracy compared to full-precision models. To address this problem, many methods have been developed to reduce quantization error and enhance the representation capability, thereby closing the performance gap between BNNs and their real-valued counterparts.", + "bbox": [ + 212, + 717, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these methods have effectively narrowed the performance gap, they still overlook the essence of BNN optimization is to update the signs of latent weights [17], instead of updating their values. In other words, if the", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "signs of the latent weights don't change, the BNNs are hardly ever updated. To demonstrate the weight sign flipping of vanilla BNNs, we train binarized ResNet18 on CIFAR100 with 120 epochs and plot the flipping information for layer4 conv2 weight. As shown in Fig. 1a, more than $50\\%$ of the weights never change their signs throughout training process, which slows down convergence and leads to a significant accuracy degradation for BNNs. We denote the part of weights that hardly change their signs during training as \"silent weights\".", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It's worth noting that, as a study mostly related to ours, ReCU [53] proposes a rectified clamp unit to revive the \"dead weights\" for updating, which refers to a group of weights that are distributed at the tails of the weight distribution and barely update their signs during training. They claim that the magnitudes of latent weights do not contribute to the forward propagation. If two weights have the same sign, they have the same effect on the forward propagation. Intuitively, the signs of the weights around the zero are easily changed, while weights distributed at the tails of the weight distribution tend to have difficulties in changing their signs. Although ReCU facilitates the updating for \"dead weights\", it ignores the fact, the gradient distribution across layers and weights have great distinction. Weights with larger magnitudes are intuitively more difficult to flip their signs, but small weights may also suffer from silence once their gradients are relatively small compared to their magnitudes. As shown in Fig. 1a, for a vanilla BNN, these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. As a result, it is still sub-optimal and remains to be improved.", + "bbox": [ + 212, + 252, + 787, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we reveal that the root cause of the large number of \"silent weights\" is attributed to the independence of the BNN gradients from the latent weight distribution via a systematical and theoretical analysis. To this end, two simple yet effective techniques, including Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are introduced to Overcome Silent Weights (OvSW). Specifically, AGS adaptively scales the gradients by establishing a relationship between the gradients and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Meanwhile, SAD automatically identifies \"silent weights\" by tracking weight sign flipping state and applies an additional penalty to them, further promoting the updating efficiency of their signs. Benefiting from AGS and SAD, OvSW enables efficient flipping of weight signs, significantly accelerates the convergence and promotes the performance for BNNs as shown in Fig. 1b and Fig. 1c. In comparison to the expensive computational cost of forward and backward propagation, OvSW introduces negligible extra overhead to the training process, maintaining its efficiency. Furthermore, since OvSW aims to facilitate the weight sign flipping for BNNs, which is orthogonal to the previous work, it has good compatibility and can be used as plug-and-play modules to enhance other BNNs' performance.", + "bbox": [ + 212, + 494, + 787, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of our work are highlighted as follows:", + "bbox": [ + 238, + 767, + 650, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We are the first to find that a large number of weights in BNNs fail to update their signs throughout the training. These \"silent weights\" are not only distributed at the tails of the weight distribution but also universally", + "bbox": [ + 223, + 794, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 113, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "present in the vicinity of zero. Theoretically, we reveal this is attributed to the independence of the gradients from the distribution of the latent weights.", + "bbox": [ + 240, + 146, + 785, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We propose to overcome \"silent weights\" with two novel techniques: (1) an adaptive gradient scaling method to scale the gradients according to the distribution of the latent weights, which improves overall efficiency in updating signs for weights; (2) a silence awareness decaying strategy to identify \"silent weights\" by tracking weight sign flipping state and introduce an additional penalty to them, further facilitating the flipping of their signs.", + "bbox": [ + 225, + 176, + 785, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Extensive experiments for various BNNs including ResNet18/20 [15] and VGGsmall [46] on CIFAR10 [22] and ResNet18/34 [15] on ImageNet1K [44] demonstrate the effectiveness of our method. For example, OvSW achieves $61.6\\%$ and $65.5\\%$ top-1 accuracy on the ImageNet for binarized ResNet18 and ResNet34, improving ReCU [53] by $0.6\\%$ and $0.4\\%$ respectively. Besides, for ResNet18, OvSW achieves $2.00\\%$ and $2.83\\%$ improvement on CIFAR100 [22] when combined with AdaBin (enhancing representation ability) and RBNN (training-aware gradient approximation) respectively, demonstrating its good compatibility with the other methods.", + "bbox": [ + 225, + 265, + 785, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 419, + 387, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Great efforts have been put into reducing the performance gap between BNNs and their real-valued counterparts. The pioneering BNN work dates back to Courbariaux et al. [5], which binarizes weights and activations to $+1$ or $-1$ by sign function. However, this aggressive approach limits the representation ability of BNNs to the binary space, resulting in a significant accuracy degradation. To reduce the quantization error and improve their accuracy, XNOR-Net [42] introduces a scaling factor obtained through the $\\ell_1$ -norm of weights or activations. Furthermore, XNOR-Net++ [3] merges two scale factors from the weights and activations into a trainable parameter and optimizes them via backpropagation. ABC-Net [31] employs a linear combination of multiple binary weight sets to approximate the real-valued weights and alleviate the information loss. Bi-Real Net [34] connects the real-valued activations to the consecutive block via an identity shortcut, which significantly enhances network representation ability while incurring negligible computational overhead. AdaBin [48] introduces adaptive binary sets to fit different distributions, enhancing binarized representation ability. UaBNN [59] introduces an uncertainty-aware BNN to reduce these weights' uncertainties. INSTA-BNN [24] controls the quantization threshold in an input instance-aware manner, taking higher-order statistics into account.", + "bbox": [ + 212, + 446, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from enhancing the representation ability of BNNs, gradient estimation is also one of the critical research directions, since gradients in the sign function are almost zero everywhere. Straight through estimator (STE) [1] is the most widely used function to enable the gradient to backpropagate. However, the gradient error is huge for STE and will accumulate during backpropagation, leading to instability in training and severe accuracy degradation. To alleviate this, Bi-Real Net [34] utilizes a piecewise polynomial function as the approximation function. IR-Net [40] proposes an error decay estimator and RBNN [30]", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "employs a training-aware approximation function to replace the sign function. EWGS [25] takes discretization error between input and output into account and introduces element-wise gradient scaling to adaptively scale up or down each gradient element. ReSTE [49] revises the original STE and balance the estimating error and the gradient stability well. All of these methods effectively reduce the gradient error and achieve consistent improvement in both training stability and accuracy compared to the vanilla STE.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Background", + "text_level": 1, + "bbox": [ + 215, + 273, + 367, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We briefly review the optimization process of BNNs in this section. Given a DNN, we denote $\\mathcal{W}_j \\in \\mathbb{R}^{C_{\\mathrm{out}}^j \\times C_{\\mathrm{in}}^j \\times K_{\\mathrm{h}}^j \\times K_{\\mathrm{w}}^j}$ as the real-valued weight in the $j$ -th layer, $C_{\\mathrm{out}}^j$ , $C_{\\mathrm{in}}^j$ , $K_{\\mathrm{h}}^j$ and $K_{\\mathrm{w}}^j$ are the number of output channels, input channels, kernel height, and kernel width, respectively. Let the real-valued activation be $\\mathcal{A}_j$ , then the convolution process can be formulated as", + "bbox": [ + 212, + 301, + 782, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {j + 1} = \\mathcal {W} _ {j} * \\mathcal {A} _ {j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 392, + 784, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $*$ represents the standard convolution operation.", + "bbox": [ + 212, + 417, + 612, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BNNs aim to binarize weights $\\mathcal{W}_j$ and activations $\\mathcal{A}_j$ to discrete values $(\\{+1, -1\\})$ through sign function:", + "bbox": [ + 212, + 431, + 813, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\operatorname {s i g n} (x) = \\left\\{ \\begin{array}{l} + 1, \\text {i f} x \\geq 0, \\\\ - 1, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 470, + 784, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reduce the quantization error in BNNs, XNOR-Net [42] introduces two scale factors to approximate the binarized weights $\\hat{\\mathcal{W}}_j^b$ and activation $\\hat{\\mathcal{A}}_j^b$ . Furthermore, XNOR-Net++ [3] proposes fusing the activation and weight scaling factors into one and optimizing it via backpropagation. This approach significantly outperforms XNOR-Net within the same computational budget and has been widely adopted by recent works [33, 34, 40, 53]. Following them, we denote the scaling factor as $\\alpha_{j}$ . Then the binary convolution operation can be formulated as", + "bbox": [ + 212, + 510, + 782, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {j + 1} = \\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j}\\right) \\odot \\alpha_ {j}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 628, + 784, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\circ$ is the efficient XNOR and Bitcount operation and $\\odot$ is theadamard product. In the implementation of BNNs, $\\mathcal{A}_{j + 1}$ will be processed through several layers, e.g., Batch Normalization (BN) layer, non-linear activation layer, and max-pooling layer. In this section, we omit these operations for simplicity.", + "bbox": [ + 212, + 655, + 782, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train a BNN, the forward propagation includes Eq. (2) and Eq. (3), where their real-value counterparts $\\mathcal{W}_j$ and $\\mathcal{A}_j$ are used for calculating gradients and updating during the backpropagation. However, the sign is not differentiable, thereby gradient estimation is important in BNNs. Following the previous studies, we use straight-through estimator (STE) [1] to approximate the gradient of the loss w.r.t. the weight $w\\in \\mathcal{W}$ :", + "bbox": [ + 212, + 715, + 782, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial w} = \\frac {\\partial \\mathcal {L}}{\\partial \\hat {w}} \\cdot \\frac {\\partial \\hat {w}}{\\partial w} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\hat {w}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 814, + 784, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59c277c91b6d9d4e2a2ff49e77ebfbbcb06174c22dbbaaa2575d510eb86a770a.jpg", + "image_caption": [ + "Fig. 2: Forward and backward computation graph for binary convolutional operation with quantization aware training." + ], + "image_footnote": [], + "bbox": [ + 271, + 146, + 733, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ denotes the loss function. For the gradient w.r.t. the activation $a\\in \\mathcal{A}$ , we adopt the piece-wise polynomial gradient estimation function [34] as follows:", + "bbox": [ + 214, + 316, + 782, + 348 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial a} = \\frac {\\partial \\mathcal {L}}{\\partial \\hat {a}} \\cdot \\frac {\\partial \\hat {a}}{\\partial a} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\hat {a}} \\cdot \\frac {\\partial F (a)}{\\partial a}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 357, + 785, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 392, + 261, + 405 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial F (a)}{\\partial a} = \\left\\{ \\begin{array}{l l} 2 + 2 a, & \\text {i f} - 1 \\leq a < 0, \\\\ 2 - 2 a, & \\text {i f} 0 \\leq a < 1, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 402, + 785, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 468, + 380, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we show that the distribution of gradients and weights for BNNs is independent by systematical and theoretical analysis. Then, we propose Adaptive Gradient Scaling (AGS) to scale the gradient and introduce Silence Awareness Decaying (SAD) to detect \"silent weights\", moving them towards zero. Both of them can enhance the efficiency of weight sign flipping.", + "bbox": [ + 214, + 497, + 787, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 The Independence of the Gradient and Weight Distribution", + "text_level": 1, + "bbox": [ + 214, + 594, + 754, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 2 demonstrates forward and backward computation graph for binary convolutional (Bin-Conv) operation. As seen, Bin-Conv is often followed by a BN layer. We can conclude that once two BNNs $\\mathcal{N}$ and $\\mathcal{N}'$ satisfy:", + "bbox": [ + 214, + 618, + 782, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {W}} _ {j} = \\hat {\\mathcal {W}} _ {j} ^ {\\prime}, \\alpha_ {j} = \\alpha_ {j} ^ {\\prime}, \\mathrm {B N} _ {j} = \\mathrm {B N} _ {j} ^ {\\prime}, \\forall j, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 670, + 785, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the same input $\\hat{\\mathcal{A}}_j$ , the output $\\mathcal{A}_{j + 1}$ and $\\mathcal{A}_{j + 1}^{\\prime}$ can be formulated as:", + "bbox": [ + 214, + 700, + 740, + 720 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {j + 1} = \\mathrm {B N} \\left(\\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j}\\right) \\odot \\alpha_ {j}\\right) = \\mathrm {B N} ^ {\\prime} \\left(\\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j} ^ {\\prime}\\right) \\odot \\alpha_ {j} ^ {\\prime}\\right) = \\mathcal {A} _ {j + 1} ^ {\\prime}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 727, + 785, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through mathematical induction, we can know that the $\\mathcal{N}$ and $\\mathcal{N}'$ will have the same output and loss, i.e. $\\mathcal{L} = \\mathcal{L}'$ . Therefore, the backpropagation for them is:", + "bbox": [ + 214, + 763, + 787, + 795 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W} _ {j}} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A} _ {j + 1}} \\frac {\\partial \\mathcal {A} _ {j + 1}}{\\partial \\hat {\\mathcal {W}} _ {j}} \\frac {\\partial \\hat {\\mathcal {W}} _ {j}}{\\partial \\mathcal {W} _ {j}} = \\frac {\\partial \\mathcal {L} ^ {\\prime}}{\\partial \\mathcal {A} _ {j + 1} ^ {\\prime}} \\frac {\\partial \\mathcal {A} _ {j + 1} ^ {\\prime}}{\\partial \\hat {\\mathcal {W}} _ {j} ^ {\\prime}} \\frac {\\hat {\\mathcal {W}} _ {j} ^ {\\prime}}{\\partial \\mathcal {W} _ {j} ^ {\\prime}} = \\frac {\\partial \\mathcal {L} ^ {\\prime}}{\\partial \\mathcal {W} _ {j} ^ {\\prime}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 803, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Without loss of generality, we can obtain that once two networks satisfy the conditions in Eq. (7), the gradients obtained from the backpropagation are the same (a more generalized scenario are extrapolated in ??). The gradient has nothing to do with the magnitude of weight and weight with larger magnitude does not necessarily corresponding to larger gradient. Therefore, although intuitively, weight with large magnitude tends to remain its sign unchanged, it is also potential for weight with small magnitude to suffer from silence if its gradient is relatively small compared to its magnitude. Furthermore, if we assume the weights further satisfy $\\mathcal{W}_j' = \\gamma \\mathcal{W}_j$ , where $\\gamma > 1$ and networks are optimized via vanilla SGD, the optimization of $t$ -step can be formulated as:", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\mathcal {W} _ {j} ^ {\\prime} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} ^ {\\prime} (t)}{\\partial \\mathcal {W} _ {j} ^ {\\prime} (t)} = \\gamma \\mathcal {W} _ {j} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} \\tag {10} \\\\ = (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} = (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t + 1), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 310, + 784, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\beta(t)$ is the learning rate at $t$ -step. Equation (10) can be viewed as a variant of exponential moving averages and the effect of $\\mathcal{W}_j(t)$ on $\\mathcal{W}_j'(t + 1)$ will gradually increase as $\\gamma$ increases. Once $k$ is large enough:", + "bbox": [ + 212, + 392, + 787, + 439 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {\\gamma \\rightarrow \\infty} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\lim _ {\\gamma \\rightarrow \\infty} [ (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t + 1) ] = (\\gamma - 1) \\mathcal {W} _ {j} (t). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 449, + 784, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The signs of $\\mathcal{W}_j^{\\prime}(t + 1)$ will be the same as $\\mathcal{W}_j(t)$ .", + "bbox": [ + 214, + 487, + 580, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, BNNs binarize $\\mathcal{W}_j'$ to $\\{+1, -1\\}$ at each training step. Once the signs of $\\mathcal{W}_j'$ do not change, the BNNs hardly ever update. It not only slows down convergence, but also leads to significant accuracy degradation. Please see ?? for a detailed ablation analysis for $\\gamma$ .", + "bbox": [ + 212, + 506, + 787, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Adaptive Gradient Scaling", + "text_level": 1, + "bbox": [ + 214, + 593, + 480, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To improve the efficiency of the update, an intuitive approach is to scale $\\beta(t)$ to $\\gamma\\beta(t)$ . In this case, $\\mathcal{W}_j'(t + 1)$ can be formulated as:", + "bbox": [ + 212, + 619, + 785, + 654 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\mathcal {W} _ {j} ^ {\\prime} (t) - \\gamma \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} ^ {\\prime} (t)} \\tag {12} \\\\ = \\gamma \\mathcal {W} _ {j} (t) - \\gamma \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} = \\gamma \\mathcal {W} _ {j} (t + 1). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 666, + 784, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, selecting the appropriate scaling factor is tricky. Meanwhile, an inappropriate scaling factor may make weights with small magnitudes correspond to large gradients, leading them to oscillate frequently in $\\{+1, -1\\}$ and introducing instability to BNNs training.", + "bbox": [ + 212, + 744, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To overcome this issue, we introduce \"Adaptive Gradient Scaling\" (AGS). Let $\\mathcal{G}_j\\in \\mathbb{R}^{C_j^j\\times C_j^j\\times K_h^j\\times K_w^j}$ denote the gradient with respect to $\\mathcal{W}_j$ , i.e., $\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{W}_j}$", + "bbox": [ + 212, + 806, + 787, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and $\\| \\cdot \\| _F$ denote the Forbenius norm, i.e., the $k$ -th filter norm in $\\mathcal{W}_j$ can be formulated as:", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\mathcal {W} _ {j} ^ {k} \\right\\| _ {F} = \\sqrt {\\sum_ {l = 1} ^ {C _ {\\mathrm {i n}} ^ {j}} \\sum_ {m = 1} ^ {K _ {\\mathrm {h}} ^ {j}} \\sum_ {n = 1} ^ {K _ {\\mathrm {w}} ^ {j}} \\mathcal {W} _ {j} ^ {k , l , m , n}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 186, + 785, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The AGS algorithm is motivated by the observation that the ratio of the norm of $\\mathcal{G}_j^k$ to $\\mathcal{W}_j^k\\left(\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\right)$ provides a simple measure of how much a single gradient descent step will change the original weight $\\mathcal{W}_j$ . For instance, if we train BNNs via vanilla SGD without momentum, then $\\frac{\\|\\triangle\\mathcal{W}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F} = \\beta \\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}$ , where the parameter update for the $\\mathcal{W}_j^k$ is given by $\\triangle \\mathcal{W}_j^k = -\\beta \\mathcal{G}_j^k$ . We can conclude that the small value of $\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}$ during training is the root cause why a large number of weights fail to flip signs and adaptively scaling $\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}$ plays a crucial role in promoting the update efficiency for BNNs. Specifically, in AGS algorithm, $\\mathcal{G}_j^{k,l,m,n}$ is scaled as:", + "bbox": [ + 212, + 229, + 787, + 409 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} = \\left\\{ \\begin{array}{l l} \\lambda \\frac {\\| \\mathcal {W} _ {j} ^ {k} \\| _ {F}}{\\| \\mathcal {G} _ {j} ^ {k} \\| _ {F}} \\mathcal {G} _ {j} ^ {k, l, m, n} & \\text {i f} \\frac {\\| \\mathcal {G} _ {j} ^ {k} \\| _ {F}}{\\| \\mathcal {W} _ {j} ^ {k} \\| _ {F}} < \\lambda , \\\\ \\mathcal {G} _ {j} ^ {k, l, m, n} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 421, + 784, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ is a scalar scaling threshold to limits the lower bound of $\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}$ . Ablation analysis for $\\lambda$ can be found in Sec. 5.3.", + "bbox": [ + 212, + 481, + 782, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AGS and \"Adaptive Gradient Clipping\" (AGC) [2] are closely related but fundamentally different, as the former restricts the lower bound of $\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}$ to enhance the flipping efficiency of BNN weight signs, while the latter restricts upper bound and is designed to improve training stability. AGS also can be viewed as a adaptive variant to Layer-wise Adaptive Rate Scaling (LARS) [56], which sets the norm of update parameter to a fixed ratio of the parameter norm, and completely ignores the gradient magnitude to real-valued networks. Although LARS is also able to improve the accuracy for BNNs, we find that LARS and AGS is much different in the update of gradient momentum and doing so degrades performance compared to AGS. More details for comparing AGS with LARS and ablation analysis for LARS can be found in ??", + "bbox": [ + 212, + 520, + 787, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Silence Awareness Decaying", + "text_level": 1, + "bbox": [ + 215, + 719, + 493, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose another approach (Silence Awareness Decaying, SAD) orthogonal to AGS to detect and prevent \"silent weights\". Specifically, we track the flipping state of $\\mathcal{W}_j$ over time using an exponential moving average (EMA) strategy, which is formulated as:", + "bbox": [ + 212, + 744, + 782, + 805 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {j} (t) = m \\cdot \\mathcal {S} _ {j} (t - 1) + (1 - m) \\cdot \\frac {\\left| \\operatorname {s i g n} \\left(\\mathcal {W} _ {j} (t)\\right) - \\operatorname {s i g n} \\left(\\mathcal {W} _ {j} (t - 1)\\right) \\right| _ {a b s}}{2}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 814, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Overview of the OvSW method." + ], + "code_body": "Input: A minibatch of inputs and their labels, real-valued weights $\\mathcal{W}(t)$ , scaling factor $\\alpha(t)$ , $\\lambda$ for AGS, $\\tau$ for SAD, $(S(t), m, \\sigma)$ for flipping state detection. \nOutput: Updated $\\mathcal{W}(t+1)$ , $\\alpha(t+1)$ and $S(t+1)$ . \n1: while Forward propagation do \n2: $\\hat{\\mathcal{A}}_j \\gets \\mathrm{sign}(\\mathcal{A}_j)$ . \n3: $\\hat{\\mathcal{W}}_j \\gets \\mathrm{sign}(\\mathcal{W}_j)$ . \n4: Computing features via Eq. (2) and Eq. (3). \n5: Computing loss $\\mathcal{L}$ . \n6: end while \n7: while Backward propagation do \n8: Computing $\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{W}_j}$ , i.e. $\\mathcal{G}_j$ , and $\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{A}_j}$ via Eq. (4) and Eq. (5)&Eq. (6) \n9: Scaling $\\mathcal{G}_j$ to $\\overline{\\mathcal{G}}_j$ via Eq. (14). \n10: Adding penalties to $\\overline{\\mathcal{G}}_j$ via Eq. (16). \n11: Update $\\mathcal{W}_j(t+1)$ and $\\alpha_j(t+1)$ via SGD optimizer. \n12: Update $\\mathcal{S}_j(t+1)$ via Eq. (15). \n13: end while", + "bbox": [ + 215, + 162, + 785, + 392 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $m$ is the momentum and $S_{j}$ is the auxiliary variable. Nagel et al. [38] employ this technique to identify and dampen oscillations prematurely while we employ it to identify \"silent weights\" and dynamically introduce additional weight penalties to move them towards zero. In our algorithm, we think if $S_{j}^{k,l,m,n}$ is less than a pre-defined threshold $\\sigma$ , its corresponding weight $\\mathcal{W}_{j}^{k,l,m,n}$ is considered as a \"silent weight\" and will be applied with an additional penalty. The silence awareness decaying process is formulated as:", + "bbox": [ + 212, + 421, + 787, + 532 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) = \\left\\{ \\begin{array}{l l} \\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) + \\gamma \\mathcal {W} _ {j} ^ {k, l, m, n} (t), & \\text {i f} \\mathcal {S} _ {j} ^ {k, l, m, n} (t) < \\sigma , \\\\ \\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 539, + 785, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\gamma$ is the proportion of penalty term.", + "bbox": [ + 214, + 590, + 522, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "It is worth noting that while both AGS and SAD improve the efficiency of weight sign flipping, they solve the problem in fundamentally different ways. AGS facilitates sign flipping for the whole weights by adaptively scaling the gradient, while SAD detects \"silent weights\" by tracking their flipping state and applies additional penalties. In Sec. 5.3, we show that both methods achieve significant performance improvement and they are complementary to each other. As an algorithm guideline, the pseudo-code of OvSW is provided in Algorithm 1.", + "bbox": [ + 212, + 604, + 787, + 712 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 733, + 366, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we conduct extensive image classification experiments for OvSW and compare it to state-of-the-art (SOTA) methods on CIFAR10 and ImageNet1K with various architectures. Then, we discuss the hyperparameter settings for OvSW, including $\\lambda$ for AGS and $\\sigma$ for SAD and convergence speed on CIFAR100. We also conduct ablation study to demonstrate the compatibility", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0651b0b017d3239b04e09a8c123a518d5b4f2b8cea03116f71e238f8241a065c.jpg", + "table_caption": [ + "Table 1: Performance comparison with SOTAs on CIFAR10. We report the Top-1 Accuracy performance on ResNet18, ResNet20, and VGGsmall. W/A denotes the bit-width of weights/activations." + ], + "table_footnote": [], + "table_body": "
NetworksMethodsBit-width (W/A)Top-1 Acc.(%)
ResNet18Full-precision32/3294.8
IR-Net [40]1/191.5
RBNN [30]1/192.2
CMIM [45]1/192.2
SiMaN [29]1/192.5
ReCU [53]1/192.8
OvSW (Ours)1/193.2
ResNet20Full-precision32/3292.1
SLB [55]1/185.5
FDA-BNN [52]1/186.2
IR-Net [40]1/186.5
CMIM [45]1/187.3
SiMaN [29]1/187.4
ReCU [53]1/187.4
OvSW (Ours)1/187.7
VGGsmallFull-precision32/3294.1
DoReFa [60]1/190.2
RAD [6]1/190.5
RBNN [30]1/191.3
DSQ [11]1/191.7
Proxy-BNN [16]1/191.8
SLB [55]1/192.0
ReCU [53]1/192.2
FDA-BNN [52]1/192.5
SiMaN [29]1/192.5
OvSW (Ours)1/192.8
", + "bbox": [ + 305, + 210, + 694, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and visualize the loss landscape for OvSW. Finally, we deploy OvSW to a real-world mobile device to exhibit its efficiency. To train OvSW, we use one NVIDIA RTX 3090 on the CIFAR10 and CIFAR100 and four on the ImageNet1K. All experiments are implemented on PyTorch [39].", + "bbox": [ + 212, + 551, + 787, + 616 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Results on CIFAR10", + "text_level": 1, + "bbox": [ + 215, + 633, + 433, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We trained OvSW for CIFAR10 with 600 epochs, where the batch size is set to 256 and the initial learning rate to 0.1, decaying with CosineAnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 5e-4 and employ the same data augmentation in ReCU [53]. $\\lambda$ and $\\sigma$ are set to 0.04 and 9e-4 respectively. We compare OvSW with IR-Net [40], RBNN [30], CMIM [45], SiMaN [29], ReCU [53], SLB [55], FDA-BNN [52], DoReFa [60], RAD [6], DSQ [11], and Proxy-BNN [16]. As shown in Tab. 1, OvSW achieves the best performance among all methods. For ResNet18, OvSW obtains $93.2\\%$ top-1 accuracy, which outperforms SiMaN and ReCU by $0.7\\%$ and $0.4\\%$ respectively, reducing the accuracy gap between BNNs and full-precision model to $1.6\\%$ . In addition, it yields $87.7\\%$ and $92.8\\%$ top-1 accuracy on ResNet20 and VGGsmall, which succeeds ReCU and SiMaN by $0.3\\%$ and $0.3\\%$ respectively.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/291ebf58c2a51a69c414de09bf3712661f2e5b9102666ef82caa61912dba91a7.jpg", + "table_caption": [ + "Table 2: Performance comparison with SOTAs on ImageNet1K. We report the Top-1 and Top-5 Accuracy performance on ResNet18 and ResNet34. W/A denotes the bit-width of weights/activations. * means using the two-step training setting as ReActNet." + ], + "table_footnote": [], + "table_body": "
ModelMethodBit-width (W/A)Top-1 Acc.(%)Top-5 Acc.(%)
ResNet18Full-precision32/3269.689.2
XNOR [42]1/151.273.2
BiReal [34]1/156.479.5
IR-Net [40]1/158.180.0
RBNN [30]1/159.981.9
SiMaN [29]1/160.182.3
FDA-BNN [52]1/160.282.3
ReCU [53]1/161.082.6
OvSW (Ours)1/161.683.1
ReActNet [33]1/165.986.1
ReCU [53]1/166.486.5
OvSW* (Ours)1/166.686.7
ResNet34Full-precision32/3273.391.3
XNOR++ [3]1/157.179.9
BiReal [34]1/162.283.9
IR-Net [40]1/162.984.1
RBNN [30]1/163.184.4
SiMaN [29]1/163.984.8
CMIM [45]1/165.085.7
ReCU [53]1/165.185.8
OvSW (Ours)1/165.586.1
", + "bbox": [ + 256, + 210, + 743, + 474 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Results on ImageNet1K", + "text_level": 1, + "bbox": [ + 215, + 532, + 460, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "On ImageNet1K, OvSW is trained from scratch. We train OvSW with 200 epochs, where the batch size is set to 512 and the initial learning rate is set to 0.1, decaying with CosineAnnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 1e-4 and the data augmentation is the same as ReCU [53]. $\\lambda$ and $\\sigma$ are set to 0.02 and 2e-5 respectively. We demonstrate the ImageNet1K performance of ResNet18/34 and compare OvSW with SOTA methods, including one-stage training methods XNOR [42], BiReal [34], IR-Net [40], RBNN [30], SiMaN [29], FDA-BNN [52], ReCU [53], CMIM [45], and two-stage training method [37] adopted by ReActNet [33]. As shown in Tab. 2, OvSW also achieves the best performance. For ResNet18, OvSW achieves $61.6\\%$ top-1 and $83.1\\%$ top-5 accuracy compared to ReCU's $61.0\\%$ top-1 and $82.6\\%$ top-5 accuracy, which demonstrates the efficiency of overcoming the \"silent weights\". For ResNet34, OvSW achieves $65.5\\%$ top-1 accuracy, which is also better than ReCU. We further compare OvSW with the two-stage training method ReActNet. OvSW obtains $66.6\\%$ top-1 accuracy, succeeding ReActNet and ReCU by $0.7\\%$ and $0.2\\%$ respectively.", + "bbox": [ + 212, + 597, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6895c2fb511bc40f4d87993ac9633d85fe403fdc0196c3f3675e68239cff5347.jpg", + "image_caption": [ + "(a) AGS vs Vanilla" + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 390, + 241 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d6e1112df788cf6c6d3eba2d6b1d6d289c81b0a64489042b7a0ca4984cab13fb.jpg", + "image_caption": [ + "(b) AGS+SAD (OvSW) vs AGS" + ], + "image_footnote": [], + "bbox": [ + 408, + 146, + 575, + 241 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/809ac9f4f5de1f838857ad4a76ee314db15488ad2979afb8648143bb561e8706.jpg", + "image_caption": [ + "(c) Convergence", + "Fig. 3: Top-1 accuracy (mean±std) of binarized ResNet18 w.r.t. different values of $\\lambda$ (a), $\\sigma$ (b) and epoch (c) on CIFAR100." + ], + "image_footnote": [], + "bbox": [ + 596, + 145, + 759, + 243 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Analysis", + "text_level": 1, + "bbox": [ + 215, + 332, + 408, + 347 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We investigate the effectiveness of hyper-parameters, including $\\lambda$ and $\\sigma$ , convergence speed, different components, and compatibility through ablation analysis. All the following results are based on binarized ResNet18 for CIFAR100.", + "bbox": [ + 212, + 366, + 784, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$\\lambda$ for AGS and $\\sigma$ for SAD. We first compare AGS with vanilla BNNs and analyze the $\\lambda$ for AGS. In Fig. 3a, we present the results for nine different settings of $\\lambda$ , which vary from 0.01 to 0.09. As seen, the appropriate $\\lambda$ can significantly improve the performance of BNNs, and $\\lambda = 0.04$ achieves the best performance. If $\\lambda$ is too small, the sign of the weights can flip inefficiently; while $\\lambda$ is too large, the sign of the weights can flip dramatically, introducing instability to the training. Based on this result, we further introduce SAD to AGS ( $\\lambda = 0.04$ ) to analyze $\\sigma$ . Figure 3b demonstrates that the performance of the model can be further improved by identifying \"silent weights\" and applying additional penalties to them via SAD ( $\\sigma = 0.0009$ ).", + "bbox": [ + 212, + 412, + 785, + 564 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Convergence for OvSW. To verify that OvSW facilitates the flipping of weight signs and thus improves the efficiency of convergence, we fix the $\\lambda$ and $\\sigma$ to 0.04 and 0.0009 respectively, and record the top-1 accuracy over the training epoch from 60 to 500. As shown in Fig. 3c, OvSW effectively accelerates training convergence and achieves better performance. For example, OvSW achieves $66.37 \\pm 0.35\\%$ top-1 accuracy with only 60 training epochs, while vanilla BNNs only reaches $55.28 \\pm 0.07\\%$ and $65.23 \\pm 0.21\\%$ top-1 accuracy with 60 and 120 training epochs respectively. It indicates that OvSW can achieve significant performance gains in scenarios with limited training resources, such as training the network on edge devices. Meanwhile, although increasing epochs improve the final performance for both OvSW and vanilla BNNs, OvSW consistently outperforms the vanilla BNNs.", + "bbox": [ + 212, + 566, + 787, + 746 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Components. AGS and SAD promote sign flipping for overall weights and \"silent weights\" respectively. To prove that they are orthogonal to each other, we conduct components ablation study for different modules and show the results in Tab. 3 (left). As seen, AGS and SAD achieve $69.61 \\pm 0.22\\%$ and $69.45 \\pm 0.18\\%$ top-1 accuracy respectively and both of them succeed vanilla BNNs. It shows that they are effective in improving the efficiency of weight signs flipping. By", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cb3ca3d4596a072411d3873ed7b622133007f541d5e5d07c956eeb4113310070.jpg", + "table_caption": [ + "Table 3: Left: Ablation study of different components in OvSW. Right: Applying OvSW as a plug-and-play module to other methods." + ], + "table_footnote": [], + "table_body": "
AGSSADmean±std (%)
XX65.23±0.21
✓X69.61±0.22
X✓69.45±0.18
✓✓69.77±0.15
", + "bbox": [ + 259, + 184, + 485, + 258 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/518ce1e59d72404acf9e2a0cbc12b619a7790a9a20500fb608948c2294f3a6f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodmean±std (%)
AdaBin70.56±0.19
AdaBin+OvSW72.56±0.11
RBNN67.15±0.23
RBNN+OvSW69.98±0.13
", + "bbox": [ + 504, + 184, + 738, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "combining them, OvSW (AGS+SAD) further achieves the best top-1 accuracy with $69.77 \\pm 0.15\\%$ .", + "bbox": [ + 212, + 284, + 782, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Compatibility. We show the good compatibility of OvSW by inserting it as a plug-and-play module into the current state-of-the-art methods, including AdaBin [48] and RBNN [30]. The former introduces an adaptive binary set to enhance the feature representation of BNNs, while the latter proposes a training-aware approximation function to reduce the gradient estimation error during training. Table 3 (right) demonstrates OvSW achieves $72.56 \\pm 0.11\\%$ top-1 accuracy on AdaBin and $69.98 \\pm 0.13\\%$ on RBNN, succeeding their original performance. This result indicates that OvSW has good compatibility and can effectively enhance the performance of existing methods.", + "bbox": [ + 212, + 315, + 784, + 450 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/279443e141fae0a2bf64c41a77474b9af62136279d270fa9b5d0bd2c7ec75f5f.jpg", + "image_caption": [ + "(a) Full-precision" + ], + "image_footnote": [], + "bbox": [ + 232, + 486, + 405, + 593 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6eebd3e9d4cd4094fd46d4f4305b5710935c510020cc0a81ba8fed42e1781c7c.jpg", + "image_caption": [ + "(b) XNOR-Net++", + "Fig. 4: 3D visualization of the loss surfaces of ResNet18 on CIFAR100, which is used to enable comparisons of sharpness/flatness of different methods." + ], + "image_footnote": [], + "bbox": [ + 418, + 486, + 589, + 592 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2f3830e5b67c1b66adf1cfa49777dcaf2d99080c4b17e14e376d973f70a8db95.jpg", + "image_caption": [ + "(c) OvSW" + ], + "image_footnote": [], + "bbox": [ + 602, + 486, + 777, + 592 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Loss Landscape Visualization", + "text_level": 1, + "bbox": [ + 214, + 696, + 503, + 710 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "BNNs restrict the weights and activations to discrete values, which naturally limits the representational capacity of the model and further result in disparate optimization landscapes compared to real-valued ones [32]. As illustrated in Fig. 4, we follow the method in [26] to plot the actual optimization landscape about our OvSW and compare it with the same architecture to real-valued and XNOR-Net++. As seen, our OvSW has a significantly smoother loss-landscape and minor loss elevation compared to XNOR-Net++, which confirms the effectiveness of OvSW in the BNN optimization.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 113, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a8ca2ca6beef932bae7616679e16bc71461baaa4ae018f85b08dffad9159a27c.jpg", + "table_caption": [ + "Table 4: Comparing OvSW (1-bit) with 32-bit and 16-bit backbones on M1 Pro." + ], + "table_footnote": [], + "table_body": "
NetworkW/ASize (MB)Memory SavingLatency (ms)Acceleration
ResNet1832/3246.76-27.67-
16/1623.382×16.651.66×
1/12.8116.64×7.973.47×
ResNet3432/3287.19-48.82-
16/1643.602×29.011.68×
1/14.1221.16×14.283.42×
", + "bbox": [ + 258, + 170, + 743, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.5 Deployment Efficiency", + "text_level": 1, + "bbox": [ + 215, + 305, + 446, + 321 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We implement 1-bit models on the M1 Pro, which features 8 high-performance Firestorm cores and 2 efficient Icestorm cores in a hybrid design. The Firestorm cores can be clocked up to $3.2\\mathrm{GHz}$ , while the Icestorm cores can reach 2.1 GHz. Our OvSW shows significant efficiency gains when deployed on real-world mobile devices, as evidenced by practical speed evaluations. To make our inference framework BOLT [9] compatible with OvSW, we leverage the ARM NEON SIMD instruction SSHL. We compare OvSW with 32-bit and 16-bit backbones. As shown in Tab. 4, OvSW inference speed is substantially faster with the highly efficient BOLT framework in a single thread. For instance, OvSW achieves an acceleration rate of about $3.47\\times$ on ResNet18 compared to its 32-bit counterpart. For the ResNet34 backbone, OvSW achieves a $3.42\\times$ acceleration rate with the BOLT framework on hardware, which is significant for computer vision applications on real-world edge devices. At the same time, OvSW can save memory by a factor of $16.64\\times$ and $21.16\\times$ , which demonstrates its potential for applications with limited memory resources.", + "bbox": [ + 212, + 335, + 787, + 563 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 590, + 359, + 607 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "BNN is a crucial method to compress deep learning models and reduce inference overhead. In this paper, systematically and theoretically, we prove the distribution of gradients is independent of latent weights, which is the root cause of inefficient updating and performance degradation of BNNs. To this end, Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are proposed to Overcome Silent Weights (OvSW) and achieve SOTA performance on BNNs. Specifically, AGS adaptively scales the gradient based on the distribution of weights, improving the efficiency of sign flipping for the overall weights. SAD can effectively measure the states of weight flipping, detect the \"silent weights\" and introduce additional penalty to them to facilitate their flipping. Extensive experiments demonstrate that OvSW can achieve notable performance gains over the SOTA methods on various datasets and networks. In addition, OvSW has better convergence efficiency and excellent compatibility, which can be combined with existing methods to further enhance the performance of BNNs.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013) 4, 5", + "2. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: Int. Conf. Mach. Learn. pp. 1059-1071. PMLR (2021) 8", + "3. Bulat, A., Tzimiropoulos, G.: Xnor-net++: Improved binary neural networks. arXiv preprint arXiv:1909.13863 (2019) 4, 5, 11", + "4. Chen, J., Liu, L., Liu, Y., Zeng, X.: A learning framework for n-bit quantized neural networks toward fpgas. IEEE Transactions on Neural Networks and Learning Systems 32(3), 1067-1081 (2020) 2", + "5. Courbariaux, M., Hubara, I., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks: Training deep neural networks with weights and activations constrained to + 1 or -1. arXiv preprint arXiv:1602.02830 (2016) 4", + "6. Ding, R., Chin, T.W., Liu, Z., Marculescu, D.: Regularizing activation distribution for training binarized deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11408-11417 (2019) 10", + "7. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13733-13742 (2021) 2", + "8. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench101: Improving transferability and generalizability of cross-task neural architecture search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5251-5260 (2021) 2", + "9. Feng, J.: Bolt. https://github.com/huawei-noah/bolt (2021) 14", + "0. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 580-587 (2014) 1", + "1. Gong, R., Liu, X., Jiang, S., Li, T., Hu, P., Lin, J., Yu, F., Yan, J.: Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In: Int. Conf. Comput. Vis. pp. 4852–4861 (2019) 2, 10", + "2. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Int. Conf. Comput. Vis. pp. 2961-2969 (2017) 2", + "3. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Int. Conf. Comput. Vis. pp. 1026-1034 (2015)", + "4. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans. Pattern Anal. Mach. Intell. 37(9), 1904-1916 (2015) 1", + "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 1, 2, 4", + "6. He, X., Mo, Z., Cheng, K., Xu, W., Hu, Q., Wang, P., Liu, Q., Cheng, J.: Proxybnn: Learning binarized neural networks via proxy matrices. In: Eur. Conf. Comput. Vis. pp. 223-241. Springer (2020) 10", + "7. Helwegen, K., Widdicombe, J., Geiger, L., Liu, Z., Cheng, K.T., Nusselder, R.: Latent weights do not exist: Rethinking binarized neural network optimization. Adv. Neural Inform. Process. Syst. 32 (2019) 2", + "8. Hinton, G., Vinyals, O., Dean, J., et al.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 2(7) (2015) 2" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC). pp. 10-14. IEEE (2014) 2", + "20. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2704-2713 (2018) 2", + "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 4", + "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst. 25 (2012) 1", + "24. Lee, C., Kim, H., Park, E., Kim, J.J.: Insta-bnn: Binary neural network with instance-aware threshold. In: Int. Conf. Comput. Vis. pp. 17325-17334 (2023) 4", + "25. Lee, J., Kim, D., Ham, B.: Network quantization with element-wise gradient scaling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6448-6457 (2021) 5", + "26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. In: Adv. Neural Inform. Process. Syst. (2018) 13", + "27. Li, S., Lin, M., Wang, Y., Fei, C., Shao, L., Ji, R.: Learning efficient gans for image translation via differentiable masks and co-attention distillation. IEEE Trans. Multimedia (2022) 2", + "28. Lin, M., Ji, R., Wang, Y., Zhang, Y., Zhang, B., Tian, Y., Shao, L.: Hrank: Filter pruning using high-rank feature map. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1529-1538 (2020) 2", + "29. Lin, M., Ji, R., Xu, Z., Zhang, B., Chao, F., Lin, C.W., Shao, L.: Siman: Sign-to-magnitude network binarization. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 10, 11", + "30. Lin, M., Ji, R., Xu, Z., Zhang, B., Wang, Y., Wu, Y., Huang, F., Lin, C.W.: Rotated binary neural network. Adv. Neural Inform. Process. Syst. 33, 7474-7485 (2020) 4, 10, 11, 13", + "31. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Adv. Neural Inform. Process. Syst. 30 (2017) 4", + "32. Liu, Z., Shen, Z., Li, S., Helwegen, K., Huang, D., Cheng, K.T.: How do adam and training strategies help bnns optimization. In: Int. Conf. Mach. Learn. pp. 6936-6946. PMLR (2021) 13", + "33. Liu, Z., Shen, Z., Savvides, M., Cheng, K.T.: Reactnet: Towards precise binary neural network with generalized activation functions. In: Eur. Conf. Comput. Vis. pp. 143-159. Springer (2020) 5, 11", + "34. Liu, Z., Wu, B., Luo, W., Yang, X., Liu, W., Cheng, K.T.: Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In: Eur. Conf. Comput. Vis. pp. 722-737 (2018) 4, 5, 6, 11", + "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3431-3440 (2015) 2", + "36. Luo, J.H., Wu, J.: Autopruner: An end-to-end trainable filter pruning method for efficient deep model inference. Pattern Recognition 107, 107461 (2020) 2", + "37. Martinez, B., Yang, J., Bulat, A., Tzimiropoulos, G.: Training binary neural networks with real-to-binary convolutions. arXiv preprint arXiv:2003.11535 (2020) 11" + ], + "bbox": [ + 215, + 147, + 784, + 837 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: Int. Conf. Mach. Learn. pp. 16318-16330. PMLR (2022) 9", + "39. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché Buc, F., Fox, E., Garnett, R. (eds.) Adv. Neural Inform. Process. Syst. pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf 10", + "40. Qin, H., Gong, R., Liu, X., Shen, M., Wei, Z., Yu, F., Song, J.: Forward and backward information retention for accurate binary neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2250-2259 (2020) 4, 5, 10, 11", + "41. Rajbhandari, S., Rasley, J., Ruwase, O., He, Y.: Zero: Memory optimizations toward training trillion parameter models. In: SC20: International Conference for High Performance Computing, Networking, Storage and Analysis. pp. 1-16. IEEE (2020)", + "42. Rastegari, M., Ordonez, V., Redmon, J., Farhadi, A.: Xnor-net: Imagenet classification using binary convolutional neural networks. In: Eur. Conf. Comput. Vis. pp. 525-542. Springer (2016) 2, 4, 5, 11", + "43. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2016) 1", + "44. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015) 4", + "45. Shang, Y., Xu, D., Zong, Z., Yan, Y.: Network binarization via contrastive learning. arXiv preprint arXiv:2207.02970 (2022) 10, 11", + "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 2, 4", + "47. Su, X., Huang, T., Li, Y., You, S., Wang, F., Qian, C., Zhang, C., Xu, C.: Prioritized architecture sampling with monto-carlo tree search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10968-10977 (2021) 2", + "48. Tu, Z., Chen, X., Ren, P., Wang, Y.: Adabin: Improving binary neural networks with adaptive binary sets. In: Eur. Conf. Comput. Vis. pp. 379-395. Springer (2022) 4, 13", + "49. Wu, X.M., Zheng, D., Liu, Z., Zheng, W.S.: Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17055-17064 (2023) 5", + "50. Xu, S., Li, Y., Ma, T., Lin, M., Dong, H., Zhang, B., Gao, P., Lu, J.: Resilient binary neural network. In: AAAI. vol. 37, pp. 10620-10628 (2023)", + "51. Xu, S., Li, Y., Wang, T., Ma, T., Zhang, B., Gao, P., Qiao, Y., Lü, J., Guo, G.: Recurrent bilinear optimization for binary neural networks. In: Eur. Conf. Comput. Vis. pp. 19-35. Springer (2022)", + "52. Xu, Y., Han, K., Xu, C., Tang, Y., Xu, C., Wang, Y.: Learning frequency domain approximation for binary neural networks. Adv. Neural Inform. Process. Syst. 34, 25553-25565 (2021) 10, 11" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Xu, Z., Lin, M., Liu, J., Chen, J., Shao, L., Gao, Y., Tian, Y., Ji, R.: Recu: Reviving the dead weights in binary neural networks. In: Int. Conf. Comput. Vis. pp. 5198-5208 (2021) 3, 4, 5, 10, 11", + "54. Yang, J., Shen, X., Xing, J., Tian, X., Li, H., Deng, B., Huang, J., Hua, X.s.: Quantization networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7308-7316 (2019) 2", + "55. Yang, Z., Wang, Y., Han, K., Xu, C., Xu, C., Tao, D., Xu, C.: Searching for low-bit weights in quantized neural networks. Adv. Neural Inform. Process. Syst. 33, 4091-4102 (2020) 10", + "56. You, Y., Gitman, I., Ginsburg, B.: Large batch training of convolutional networks. arXiv preprint arXiv:1708.03888 (2017) 8", + "57. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Eur. Conf. Comput. Vis. pp. 365-382 (2018) 2", + "58. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6848-6856 (2018) 2", + "59. Zhao, J., Yang, L., Zhang, B., Guo, G., Doermann, D.S.: Uncertainty-aware binary neural networks. In: IJCAI. pp. 3441-3447 (2021) 4", + "60. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016) 10", + "61. Zhu, F., Gong, R., Yu, F., Liu, X., Wang, Y., Li, Z., Yang, X., Yan, J.: Towards unified int8 training for convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1969-1979 (2020) 2" + ], + "bbox": [ + 215, + 146, + 784, + 494 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Xiang et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_model.json b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_model.json new file mode 100644 index 0000000000000000000000000000000000000000..95a1a21d3aebf3b6bbbf3da4611583d9e9767575 --- /dev/null +++ b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_model.json @@ -0,0 +1,2568 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.212, + 0.78, + 0.229 + ], + "angle": 0, + "content": "Jingyang Xiang\\(^{1}\\), Zuohui Chen\\(^{2}\\), Siqi Li\\(^{1}\\), Qing Wu\\(^{3}\\), and Yong Liu\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.24, + 0.619, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) APRIL Lab, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.255, + 0.623, + 0.268 + ], + "angle": 0, + "content": "2 Zhejiang University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.268, + 0.702, + 0.281 + ], + "angle": 0, + "content": "3 College of Computer Science, Hangzhou Dianzi University" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.282, + 0.635, + 0.296 + ], + "angle": 0, + "content": "4 Huzhou Institute, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.324, + 0.744, + 0.699 + ], + "angle": 0, + "content": "Abstract. Binary Neural Networks (BNNs) have been proven to be highly effective for deploying deep neural networks on mobile and embedded platforms. Most existing works focus on minimizing quantization errors, improving representation ability, or designing gradient approximations to alleviate gradient mismatch in BNNs, while leaving the weight sign flipping, a critical factor for achieving powerful BNNs, untouched. In this paper, we investigate the efficiency of weight sign updates in BNNs. We observe that, for vanilla BNNs, over \\(50\\%\\) of the weights remain their signs unchanged during training, and these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. We refer to these weights as \"silent weights\", which slow down convergence and lead to a significant accuracy degradation. Theoretically, we reveal this is due to the independence of the BNNs gradient from the latent weight distribution. To address the issue, we propose Overcome Silent Weights (OvSW). OvSW first employs Adaptive Gradient Scaling (AGS) to establish a relationship between the gradient and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Additionally, we design Silence Awareness Decaying (SAD) to automatically identify \"silent weights\" by tracking weight flipping state, and apply an additional penalty to \"silent weights\" to facilitate their flipping. By efficiently updating weight signs, our method achieves faster convergence and state-of-the-art performance on CIFAR10 and ImageNet1K dataset with various architectures. For example, OvSW obtains \\(61.6\\%\\) and \\(65.5\\%\\) top-1 accuracy on the ImageNet1K using binarized ResNet18 and ResNet34 architecture respectively. Codes are available at https://github.com/JingyangXiang/OvSW." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.71, + 0.741, + 0.739 + ], + "angle": 0, + "content": "Keywords: Binary Neural Networks \\(\\cdot\\) Silent Weights \\(\\cdot\\) Adaptive Gradient Scaling \\(\\cdot\\) Silence Awareness Decaying" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.76, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have shown tremendous success in various computer vision tasks, including image classification [15, 23], object detection [10, 14," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.551, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author: yongliu@iipc.zju.edu.cn." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.152, + 0.396, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.252, + 0.35, + 0.265 + ], + "angle": 0, + "content": "(a) Vanilla" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.153, + 0.58, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.253, + 0.534, + 0.265 + ], + "angle": 0, + "content": "(b) OvSW" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.153, + 0.782, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.253, + 0.768, + 0.266 + ], + "angle": 0, + "content": "(c) Top-1 Acc & Flip Ratio" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 1: (a) and (b): Histogram of the initialized weight distribution (blue) and the weights that never update signs throughout training (orange). \\(54.07\\%\\) and \\(2.03\\%\\) represent the ratio of the corresponding orange area to the blue. (c) Top-1 Accuracy (solid lines) and Flip Ratio (dashed lines) are for Vanilla (green) and OvSW (red). (a), (b) and Flip Ratio in (c) is for layer4(conv2.weight." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.44 + ], + "angle": 0, + "content": "43], and semantic segmentation [12, 35]. However, the remarkable performance is always attributed to deeper and wider architectures [15, 46], which comes with expensive memory and computational overhead and makes it challenging to deploy DNNs on resource-constrained edge platforms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.594 + ], + "angle": 0, + "content": "The community has been delving into model compression, which aims to reduce inference overhead for DNNs while preserving their performance. Typical techniques include, but are not limited to, efficient architecture design [7, 58], neural architecture search [8, 47], network pruning [28, 36], knowledge distillation [18, 27], and network quantization [4, 11, 61]. Among them, network quantization is widely suggested as a promising solution. It effectively enhances memory efficiency and execution speed on embedding platforms by reducing the weight and activation bits and replacing expensive floating-point arithmetic with relatively inexpensive fixed-point arithmetic. Extensive studies [11, 20, 54, 57] have demonstrated its effectiveness and yielded light and efficient DNNs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.716 + ], + "angle": 0, + "content": "In this paper, we focus on studying binary neural networks (BNNs), which are an aggressive form of quantized neural networks (QNNs). BNNs binarize both weights and activations to discrete values \\(\\{+1, -1\\}\\), which can be represented by 1-bit and reduce memory usage by \\(32 \\times\\). On the other hand, by employing efficient bitwise operations instead of floating-point ones, BNNs also significantly reduce computation complexity, providing \\(58 \\times\\) speedup as reported by XNOR-Net [42]. These characteristics make them well-suited for deployment on low-power embedding platforms, including FPGA, ASICs, and IoT devices [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.794 + ], + "angle": 0, + "content": "However, the application of BNNs is still limited due to constraining both weights and activations to 1-bit significantly reducing accuracy compared to full-precision models. To address this problem, many methods have been developed to reduce quantization error and enhance the representation capability, thereby closing the performance gap between BNNs and their real-valued counterparts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Although these methods have effectively narrowed the performance gap, they still overlook the essence of BNN optimization is to update the signs of latent weights [17], instead of updating their values. In other words, if the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "signs of the latent weights don't change, the BNNs are hardly ever updated. To demonstrate the weight sign flipping of vanilla BNNs, we train binarized ResNet18 on CIFAR100 with 120 epochs and plot the flipping information for layer4 conv2 weight. As shown in Fig. 1a, more than \\(50\\%\\) of the weights never change their signs throughout training process, which slows down convergence and leads to a significant accuracy degradation for BNNs. We denote the part of weights that hardly change their signs during training as \"silent weights\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.789, + 0.494 + ], + "angle": 0, + "content": "It's worth noting that, as a study mostly related to ours, ReCU [53] proposes a rectified clamp unit to revive the \"dead weights\" for updating, which refers to a group of weights that are distributed at the tails of the weight distribution and barely update their signs during training. They claim that the magnitudes of latent weights do not contribute to the forward propagation. If two weights have the same sign, they have the same effect on the forward propagation. Intuitively, the signs of the weights around the zero are easily changed, while weights distributed at the tails of the weight distribution tend to have difficulties in changing their signs. Although ReCU facilitates the updating for \"dead weights\", it ignores the fact, the gradient distribution across layers and weights have great distinction. Weights with larger magnitudes are intuitively more difficult to flip their signs, but small weights may also suffer from silence once their gradients are relatively small compared to their magnitudes. As shown in Fig. 1a, for a vanilla BNN, these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. As a result, it is still sub-optimal and remains to be improved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.789, + 0.767 + ], + "angle": 0, + "content": "In this paper, we reveal that the root cause of the large number of \"silent weights\" is attributed to the independence of the BNN gradients from the latent weight distribution via a systematical and theoretical analysis. To this end, two simple yet effective techniques, including Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are introduced to Overcome Silent Weights (OvSW). Specifically, AGS adaptively scales the gradients by establishing a relationship between the gradients and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Meanwhile, SAD automatically identifies \"silent weights\" by tracking weight sign flipping state and applies an additional penalty to them, further promoting the updating efficiency of their signs. Benefiting from AGS and SAD, OvSW enables efficient flipping of weight signs, significantly accelerates the convergence and promotes the performance for BNNs as shown in Fig. 1b and Fig. 1c. In comparison to the expensive computational cost of forward and backward propagation, OvSW introduces negligible extra overhead to the training process, maintaining its efficiency. Furthermore, since OvSW aims to facilitate the weight sign flipping for BNNs, which is orthogonal to the previous work, it has good compatibility and can be used as plug-and-play modules to enhance other BNNs' performance." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.768, + 0.651, + 0.783 + ], + "angle": 0, + "content": "The contributions of our work are highlighted as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "- We are the first to find that a large number of weights in BNNs fail to update their signs throughout the training. These \"silent weights\" are not only distributed at the tails of the weight distribution but also universally" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.177 + ], + "angle": 0, + "content": "present in the vicinity of zero. Theoretically, we reveal this is attributed to the independence of the gradients from the distribution of the latent weights." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.265 + ], + "angle": 0, + "content": "- We propose to overcome \"silent weights\" with two novel techniques: (1) an adaptive gradient scaling method to scale the gradients according to the distribution of the latent weights, which improves overall efficiency in updating signs for weights; (2) a silence awareness decaying strategy to identify \"silent weights\" by tracking weight sign flipping state and introduce an additional penalty to them, further facilitating the flipping of their signs." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.266, + 0.786, + 0.401 + ], + "angle": 0, + "content": "- Extensive experiments for various BNNs including ResNet18/20 [15] and VGGsmall [46] on CIFAR10 [22] and ResNet18/34 [15] on ImageNet1K [44] demonstrate the effectiveness of our method. For example, OvSW achieves \\(61.6\\%\\) and \\(65.5\\%\\) top-1 accuracy on the ImageNet for binarized ResNet18 and ResNet34, improving ReCU [53] by \\(0.6\\%\\) and \\(0.4\\%\\) respectively. Besides, for ResNet18, OvSW achieves \\(2.00\\%\\) and \\(2.83\\%\\) improvement on CIFAR100 [22] when combined with AdaBin (enhancing representation ability) and RBNN (training-aware gradient approximation) respectively, demonstrating its good compatibility with the other methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.42, + 0.388, + 0.436 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Great efforts have been put into reducing the performance gap between BNNs and their real-valued counterparts. The pioneering BNN work dates back to Courbariaux et al. [5], which binarizes weights and activations to \\(+1\\) or \\(-1\\) by sign function. However, this aggressive approach limits the representation ability of BNNs to the binary space, resulting in a significant accuracy degradation. To reduce the quantization error and improve their accuracy, XNOR-Net [42] introduces a scaling factor obtained through the \\(\\ell_1\\)-norm of weights or activations. Furthermore, XNOR-Net++ [3] merges two scale factors from the weights and activations into a trainable parameter and optimizes them via backpropagation. ABC-Net [31] employs a linear combination of multiple binary weight sets to approximate the real-valued weights and alleviate the information loss. Bi-Real Net [34] connects the real-valued activations to the consecutive block via an identity shortcut, which significantly enhances network representation ability while incurring negligible computational overhead. AdaBin [48] introduces adaptive binary sets to fit different distributions, enhancing binarized representation ability. UaBNN [59] introduces an uncertainty-aware BNN to reduce these weights' uncertainties. INSTA-BNN [24] controls the quantization threshold in an input instance-aware manner, taking higher-order statistics into account." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Apart from enhancing the representation ability of BNNs, gradient estimation is also one of the critical research directions, since gradients in the sign function are almost zero everywhere. Straight through estimator (STE) [1] is the most widely used function to enable the gradient to backpropagate. However, the gradient error is huge for STE and will accumulate during backpropagation, leading to instability in training and severe accuracy degradation. To alleviate this, Bi-Real Net [34] utilizes a piecewise polynomial function as the approximation function. IR-Net [40] proposes an error decay estimator and RBNN [30]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "employs a training-aware approximation function to replace the sign function. EWGS [25] takes discretization error between input and output into account and introduces element-wise gradient scaling to adaptively scale up or down each gradient element. ReSTE [49] revises the original STE and balance the estimating error and the gradient stability well. All of these methods effectively reduce the gradient error and achieve consistent improvement in both training stability and accuracy compared to the vanilla STE." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.368, + 0.291 + ], + "angle": 0, + "content": "3 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.784, + 0.382 + ], + "angle": 0, + "content": "We briefly review the optimization process of BNNs in this section. Given a DNN, we denote \\(\\mathcal{W}_j \\in \\mathbb{R}^{C_{\\mathrm{out}}^j \\times C_{\\mathrm{in}}^j \\times K_{\\mathrm{h}}^j \\times K_{\\mathrm{w}}^j}\\) as the real-valued weight in the \\(j\\)-th layer, \\(C_{\\mathrm{out}}^j\\), \\(C_{\\mathrm{in}}^j\\), \\(K_{\\mathrm{h}}^j\\) and \\(K_{\\mathrm{w}}^j\\) are the number of output channels, input channels, kernel height, and kernel width, respectively. Let the real-valued activation be \\(\\mathcal{A}_j\\), then the convolution process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.393, + 0.785, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {j + 1} = \\mathcal {W} _ {j} * \\mathcal {A} _ {j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.613, + 0.432 + ], + "angle": 0, + "content": "where \\(*\\) represents the standard convolution operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.814, + 0.462 + ], + "angle": 0, + "content": "BNNs aim to binarize weights \\(\\mathcal{W}_j\\) and activations \\(\\mathcal{A}_j\\) to discrete values \\((\\{+1, -1\\})\\) through sign function:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.471, + 0.785, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\operatorname {s i g n} (x) = \\left\\{ \\begin{array}{l} + 1, \\text {i f} x \\geq 0, \\\\ - 1, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.784, + 0.629 + ], + "angle": 0, + "content": "To reduce the quantization error in BNNs, XNOR-Net [42] introduces two scale factors to approximate the binarized weights \\(\\hat{\\mathcal{W}}_j^b\\) and activation \\(\\hat{\\mathcal{A}}_j^b\\). Furthermore, XNOR-Net++ [3] proposes fusing the activation and weight scaling factors into one and optimizing it via backpropagation. This approach significantly outperforms XNOR-Net within the same computational budget and has been widely adopted by recent works [33, 34, 40, 53]. Following them, we denote the scaling factor as \\(\\alpha_{j}\\). Then the binary convolution operation can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.629, + 0.785, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {j + 1} = \\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j}\\right) \\odot \\alpha_ {j}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.784, + 0.716 + ], + "angle": 0, + "content": "where \\(\\circ\\) is the efficient XNOR and Bitcount operation and \\(\\odot\\) is theadamard product. In the implementation of BNNs, \\(\\mathcal{A}_{j + 1}\\) will be processed through several layers, e.g., Batch Normalization (BN) layer, non-linear activation layer, and max-pooling layer. In this section, we omit these operations for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.784, + 0.807 + ], + "angle": 0, + "content": "To train a BNN, the forward propagation includes Eq. (2) and Eq. (3), where their real-value counterparts \\(\\mathcal{W}_j\\) and \\(\\mathcal{A}_j\\) are used for calculating gradients and updating during the backpropagation. However, the sign is not differentiable, thereby gradient estimation is important in BNNs. Following the previous studies, we use straight-through estimator (STE) [1] to approximate the gradient of the loss w.r.t. the weight \\(w\\in \\mathcal{W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.815, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial w} = \\frac {\\partial \\mathcal {L}}{\\partial \\hat {w}} \\cdot \\frac {\\partial \\hat {w}}{\\partial w} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\hat {w}}, \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.147, + 0.734, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.261, + 0.788, + 0.292 + ], + "angle": 0, + "content": "Fig. 2: Forward and backward computation graph for binary convolutional operation with quantization aware training." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.318, + 0.784, + 0.349 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) denotes the loss function. For the gradient w.r.t. the activation \\(a\\in \\mathcal{A}\\), we adopt the piece-wise polynomial gradient estimation function [34] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.358, + 0.786, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial a} = \\frac {\\partial \\mathcal {L}}{\\partial \\hat {a}} \\cdot \\frac {\\partial \\hat {a}}{\\partial a} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\hat {a}} \\cdot \\frac {\\partial F (a)}{\\partial a}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.393, + 0.263, + 0.406 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.403, + 0.786, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial F (a)}{\\partial a} = \\left\\{ \\begin{array}{l l} 2 + 2 a, & \\text {i f} - 1 \\leq a < 0, \\\\ 2 - 2 a, & \\text {i f} 0 \\leq a < 1, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.381, + 0.487 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.788, + 0.577 + ], + "angle": 0, + "content": "In this section, we show that the distribution of gradients and weights for BNNs is independent by systematical and theoretical analysis. Then, we propose Adaptive Gradient Scaling (AGS) to scale the gradient and introduce Silence Awareness Decaying (SAD) to detect \"silent weights\", moving them towards zero. Both of them can enhance the efficiency of weight sign flipping." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.595, + 0.755, + 0.611 + ], + "angle": 0, + "content": "4.1 The Independence of the Gradient and Weight Distribution" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.619, + 0.784, + 0.664 + ], + "angle": 0, + "content": "Figure 2 demonstrates forward and backward computation graph for binary convolutional (Bin-Conv) operation. As seen, Bin-Conv is often followed by a BN layer. We can conclude that once two BNNs \\(\\mathcal{N}\\) and \\(\\mathcal{N}'\\) satisfy:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.671, + 0.787, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {W}} _ {j} = \\hat {\\mathcal {W}} _ {j} ^ {\\prime}, \\alpha_ {j} = \\alpha_ {j} ^ {\\prime}, \\mathrm {B N} _ {j} = \\mathrm {B N} _ {j} ^ {\\prime}, \\forall j, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.702, + 0.741, + 0.721 + ], + "angle": 0, + "content": "for the same input \\(\\hat{\\mathcal{A}}_j\\), the output \\(\\mathcal{A}_{j + 1}\\) and \\(\\mathcal{A}_{j + 1}^{\\prime}\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.728, + 0.787, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {j + 1} = \\mathrm {B N} \\left(\\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j}\\right) \\odot \\alpha_ {j}\\right) = \\mathrm {B N} ^ {\\prime} \\left(\\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j} ^ {\\prime}\\right) \\odot \\alpha_ {j} ^ {\\prime}\\right) = \\mathcal {A} _ {j + 1} ^ {\\prime}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.796 + ], + "angle": 0, + "content": "Through mathematical induction, we can know that the \\(\\mathcal{N}\\) and \\(\\mathcal{N}'\\) will have the same output and loss, i.e. \\(\\mathcal{L} = \\mathcal{L}'\\). Therefore, the backpropagation for them is:" + }, + { + "type": "equation", + "bbox": [ + 0.283, + 0.804, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W} _ {j}} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A} _ {j + 1}} \\frac {\\partial \\mathcal {A} _ {j + 1}}{\\partial \\hat {\\mathcal {W}} _ {j}} \\frac {\\partial \\hat {\\mathcal {W}} _ {j}}{\\partial \\mathcal {W} _ {j}} = \\frac {\\partial \\mathcal {L} ^ {\\prime}}{\\partial \\mathcal {A} _ {j + 1} ^ {\\prime}} \\frac {\\partial \\mathcal {A} _ {j + 1} ^ {\\prime}}{\\partial \\hat {\\mathcal {W}} _ {j} ^ {\\prime}} \\frac {\\hat {\\mathcal {W}} _ {j} ^ {\\prime}}{\\partial \\mathcal {W} _ {j} ^ {\\prime}} = \\frac {\\partial \\mathcal {L} ^ {\\prime}}{\\partial \\mathcal {W} _ {j} ^ {\\prime}}. \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Without loss of generality, we can obtain that once two networks satisfy the conditions in Eq. (7), the gradients obtained from the backpropagation are the same (a more generalized scenario are extrapolated in ??). The gradient has nothing to do with the magnitude of weight and weight with larger magnitude does not necessarily corresponding to larger gradient. Therefore, although intuitively, weight with large magnitude tends to remain its sign unchanged, it is also potential for weight with small magnitude to suffer from silence if its gradient is relatively small compared to its magnitude. Furthermore, if we assume the weights further satisfy \\(\\mathcal{W}_j' = \\gamma \\mathcal{W}_j\\), where \\(\\gamma > 1\\) and networks are optimized via vanilla SGD, the optimization of \\(t\\)-step can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.311, + 0.785, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\mathcal {W} _ {j} ^ {\\prime} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} ^ {\\prime} (t)}{\\partial \\mathcal {W} _ {j} ^ {\\prime} (t)} = \\gamma \\mathcal {W} _ {j} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} \\tag {10} \\\\ = (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} = (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t + 1), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.44 + ], + "angle": 0, + "content": "where \\(\\beta(t)\\) is the learning rate at \\(t\\)-step. Equation (10) can be viewed as a variant of exponential moving averages and the effect of \\(\\mathcal{W}_j(t)\\) on \\(\\mathcal{W}_j'(t + 1)\\) will gradually increase as \\(\\gamma\\) increases. Once \\(k\\) is large enough:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.45, + 0.785, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\lim _ {\\gamma \\rightarrow \\infty} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\lim _ {\\gamma \\rightarrow \\infty} [ (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t + 1) ] = (\\gamma - 1) \\mathcal {W} _ {j} (t). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.488, + 0.581, + 0.505 + ], + "angle": 0, + "content": "The signs of \\(\\mathcal{W}_j^{\\prime}(t + 1)\\) will be the same as \\(\\mathcal{W}_j(t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.572 + ], + "angle": 0, + "content": "However, BNNs binarize \\(\\mathcal{W}_j'\\) to \\(\\{+1, -1\\}\\) at each training step. Once the signs of \\(\\mathcal{W}_j'\\) do not change, the BNNs hardly ever update. It not only slows down convergence, but also leads to significant accuracy degradation. Please see ?? for a detailed ablation analysis for \\(\\gamma\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.594, + 0.482, + 0.61 + ], + "angle": 0, + "content": "4.2 Adaptive Gradient Scaling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.787, + 0.655 + ], + "angle": 0, + "content": "To improve the efficiency of the update, an intuitive approach is to scale \\(\\beta(t)\\) to \\(\\gamma\\beta(t)\\). In this case, \\(\\mathcal{W}_j'(t + 1)\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.667, + 0.785, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\mathcal {W} _ {j} ^ {\\prime} (t) - \\gamma \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} ^ {\\prime} (t)} \\tag {12} \\\\ = \\gamma \\mathcal {W} _ {j} (t) - \\gamma \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} = \\gamma \\mathcal {W} _ {j} (t + 1). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.788, + 0.807 + ], + "angle": 0, + "content": "However, selecting the appropriate scaling factor is tricky. Meanwhile, an inappropriate scaling factor may make weights with small magnitudes correspond to large gradients, leading them to oscillate frequently in \\(\\{+1, -1\\}\\) and introducing instability to BNNs training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.807, + 0.788, + 0.845 + ], + "angle": 0, + "content": "To overcome this issue, we introduce \"Adaptive Gradient Scaling\" (AGS). Let \\(\\mathcal{G}_j\\in \\mathbb{R}^{C_j^j\\times C_j^j\\times K_h^j\\times K_w^j}\\) denote the gradient with respect to \\(\\mathcal{W}_j\\), i.e., \\(\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{W}_j}\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "and \\(\\| \\cdot \\| _F\\) denote the Forbenius norm, i.e., the \\(k\\)-th filter norm in \\(\\mathcal{W}_j\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.187, + 0.786, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\mathcal {W} _ {j} ^ {k} \\right\\| _ {F} = \\sqrt {\\sum_ {l = 1} ^ {C _ {\\mathrm {i n}} ^ {j}} \\sum_ {m = 1} ^ {K _ {\\mathrm {h}} ^ {j}} \\sum_ {n = 1} ^ {K _ {\\mathrm {w}} ^ {j}} \\mathcal {W} _ {j} ^ {k , l , m , n}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.41 + ], + "angle": 0, + "content": "The AGS algorithm is motivated by the observation that the ratio of the norm of \\(\\mathcal{G}_j^k\\) to \\(\\mathcal{W}_j^k\\left(\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\right)\\) provides a simple measure of how much a single gradient descent step will change the original weight \\(\\mathcal{W}_j\\). For instance, if we train BNNs via vanilla SGD without momentum, then \\(\\frac{\\|\\triangle\\mathcal{W}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F} = \\beta \\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\), where the parameter update for the \\(\\mathcal{W}_j^k\\) is given by \\(\\triangle \\mathcal{W}_j^k = -\\beta \\mathcal{G}_j^k\\). We can conclude that the small value of \\(\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\) during training is the root cause why a large number of weights fail to flip signs and adaptively scaling \\(\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\) plays a crucial role in promoting the update efficiency for BNNs. Specifically, in AGS algorithm, \\(\\mathcal{G}_j^{k,l,m,n}\\) is scaled as:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.422, + 0.785, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} = \\left\\{ \\begin{array}{l l} \\lambda \\frac {\\| \\mathcal {W} _ {j} ^ {k} \\| _ {F}}{\\| \\mathcal {G} _ {j} ^ {k} \\| _ {F}} \\mathcal {G} _ {j} ^ {k, l, m, n} & \\text {i f} \\frac {\\| \\mathcal {G} _ {j} ^ {k} \\| _ {F}}{\\| \\mathcal {W} _ {j} ^ {k} \\| _ {F}} < \\lambda , \\\\ \\mathcal {G} _ {j} ^ {k, l, m, n} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.784, + 0.519 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a scalar scaling threshold to limits the lower bound of \\(\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\). Ablation analysis for \\(\\lambda\\) can be found in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.697 + ], + "angle": 0, + "content": "AGS and \"Adaptive Gradient Clipping\" (AGC) [2] are closely related but fundamentally different, as the former restricts the lower bound of \\(\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\) to enhance the flipping efficiency of BNN weight signs, while the latter restricts upper bound and is designed to improve training stability. AGS also can be viewed as a adaptive variant to Layer-wise Adaptive Rate Scaling (LARS) [56], which sets the norm of update parameter to a fixed ratio of the parameter norm, and completely ignores the gradient magnitude to real-valued networks. Although LARS is also able to improve the accuracy for BNNs, we find that LARS and AGS is much different in the update of gradient momentum and doing so degrades performance compared to AGS. More details for comparing AGS with LARS and ablation analysis for LARS can be found in ??" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.494, + 0.736 + ], + "angle": 0, + "content": "4.3 Silence Awareness Decaying" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.745, + 0.784, + 0.806 + ], + "angle": 0, + "content": "We propose another approach (Silence Awareness Decaying, SAD) orthogonal to AGS to detect and prevent \"silent weights\". Specifically, we track the flipping state of \\(\\mathcal{W}_j\\) over time using an exponential moving average (EMA) strategy, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.815, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {j} (t) = m \\cdot \\mathcal {S} _ {j} (t - 1) + (1 - m) \\cdot \\frac {\\left| \\operatorname {s i g n} \\left(\\mathcal {W} _ {j} (t)\\right) - \\operatorname {s i g n} \\left(\\mathcal {W} _ {j} (t - 1)\\right) \\right| _ {a b s}}{2}, \\tag {15}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.556, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Overview of the OvSW method." + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.164, + 0.787, + 0.393 + ], + "angle": 0, + "content": "Input: A minibatch of inputs and their labels, real-valued weights \\(\\mathcal{W}(t)\\), scaling factor \\(\\alpha(t)\\), \\(\\lambda\\) for AGS, \\(\\tau\\) for SAD, \\((S(t), m, \\sigma)\\) for flipping state detection. \nOutput: Updated \\(\\mathcal{W}(t+1)\\), \\(\\alpha(t+1)\\) and \\(S(t+1)\\). \n1: while Forward propagation do \n2: \\(\\hat{\\mathcal{A}}_j \\gets \\mathrm{sign}(\\mathcal{A}_j)\\). \n3: \\(\\hat{\\mathcal{W}}_j \\gets \\mathrm{sign}(\\mathcal{W}_j)\\). \n4: Computing features via Eq. (2) and Eq. (3). \n5: Computing loss \\(\\mathcal{L}\\). \n6: end while \n7: while Backward propagation do \n8: Computing \\(\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{W}_j}\\), i.e. \\(\\mathcal{G}_j\\), and \\(\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{A}_j}\\) via Eq. (4) and Eq. (5)&Eq. (6) \n9: Scaling \\(\\mathcal{G}_j\\) to \\(\\overline{\\mathcal{G}}_j\\) via Eq. (14). \n10: Adding penalties to \\(\\overline{\\mathcal{G}}_j\\) via Eq. (16). \n11: Update \\(\\mathcal{W}_j(t+1)\\) and \\(\\alpha_j(t+1)\\) via SGD optimizer. \n12: Update \\(\\mathcal{S}_j(t+1)\\) via Eq. (15). \n13: end while" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.533 + ], + "angle": 0, + "content": "where \\( m \\) is the momentum and \\( S_{j} \\) is the auxiliary variable. Nagel et al. [38] employ this technique to identify and dampen oscillations prematurely while we employ it to identify \"silent weights\" and dynamically introduce additional weight penalties to move them towards zero. In our algorithm, we think if \\( S_{j}^{k,l,m,n} \\) is less than a pre-defined threshold \\( \\sigma \\), its corresponding weight \\( \\mathcal{W}_{j}^{k,l,m,n} \\) is considered as a \"silent weight\" and will be applied with an additional penalty. The silence awareness decaying process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.279, + 0.54, + 0.786, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) = \\left\\{ \\begin{array}{l l} \\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) + \\gamma \\mathcal {W} _ {j} ^ {k, l, m, n} (t), & \\text {i f} \\mathcal {S} _ {j} ^ {k, l, m, n} (t) < \\sigma , \\\\ \\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.592, + 0.524, + 0.607 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the proportion of penalty term." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.789, + 0.713 + ], + "angle": 0, + "content": "It is worth noting that while both AGS and SAD improve the efficiency of weight sign flipping, they solve the problem in fundamentally different ways. AGS facilitates sign flipping for the whole weights by adaptively scaling the gradient, while SAD detects \"silent weights\" by tracking their flipping state and applies additional penalties. In Sec. 5.3, we show that both methods achieve significant performance improvement and they are complementary to each other. As an algorithm guideline, the pseudo-code of OvSW is provided in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.367, + 0.751 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this section, we conduct extensive image classification experiments for OvSW and compare it to state-of-the-art (SOTA) methods on CIFAR10 and ImageNet1K with various architectures. Then, we discuss the hyperparameter settings for OvSW, including \\(\\lambda\\) for AGS and \\(\\sigma\\) for SAD and convergence speed on CIFAR100. We also conduct ablation study to demonstrate the compatibility" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Performance comparison with SOTAs on CIFAR10. We report the Top-1 Accuracy performance on ResNet18, ResNet20, and VGGsmall. W/A denotes the bit-width of weights/activations." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.212, + 0.695, + 0.514 + ], + "angle": 0, + "content": "
NetworksMethodsBit-width (W/A)Top-1 Acc.(%)
ResNet18Full-precision32/3294.8
IR-Net [40]1/191.5
RBNN [30]1/192.2
CMIM [45]1/192.2
SiMaN [29]1/192.5
ReCU [53]1/192.8
OvSW (Ours)1/193.2
ResNet20Full-precision32/3292.1
SLB [55]1/185.5
FDA-BNN [52]1/186.2
IR-Net [40]1/186.5
CMIM [45]1/187.3
SiMaN [29]1/187.4
ReCU [53]1/187.4
OvSW (Ours)1/187.7
VGGsmallFull-precision32/3294.1
DoReFa [60]1/190.2
RAD [6]1/190.5
RBNN [30]1/191.3
DSQ [11]1/191.7
Proxy-BNN [16]1/191.8
SLB [55]1/192.0
ReCU [53]1/192.2
FDA-BNN [52]1/192.5
SiMaN [29]1/192.5
OvSW (Ours)1/192.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.617 + ], + "angle": 0, + "content": "and visualize the loss landscape for OvSW. Finally, we deploy OvSW to a real-world mobile device to exhibit its efficiency. To train OvSW, we use one NVIDIA RTX 3090 on the CIFAR10 and CIFAR100 and four on the ImageNet1K. All experiments are implemented on PyTorch [39]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.434, + 0.649 + ], + "angle": 0, + "content": "5.1 Results on CIFAR10" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We trained OvSW for CIFAR10 with 600 epochs, where the batch size is set to 256 and the initial learning rate to 0.1, decaying with CosineAnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 5e-4 and employ the same data augmentation in ReCU [53]. \\(\\lambda\\) and \\(\\sigma\\) are set to 0.04 and 9e-4 respectively. We compare OvSW with IR-Net [40], RBNN [30], CMIM [45], SiMaN [29], ReCU [53], SLB [55], FDA-BNN [52], DoReFa [60], RAD [6], DSQ [11], and Proxy-BNN [16]. As shown in Tab. 1, OvSW achieves the best performance among all methods. For ResNet18, OvSW obtains \\(93.2\\%\\) top-1 accuracy, which outperforms SiMaN and ReCU by \\(0.7\\%\\) and \\(0.4\\%\\) respectively, reducing the accuracy gap between BNNs and full-precision model to \\(1.6\\%\\). In addition, it yields \\(87.7\\%\\) and \\(92.8\\%\\) top-1 accuracy on ResNet20 and VGGsmall, which succeeds ReCU and SiMaN by \\(0.3\\%\\) and \\(0.3\\%\\) respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Performance comparison with SOTAs on ImageNet1K. We report the Top-1 and Top-5 Accuracy performance on ResNet18 and ResNet34. W/A denotes the bit-width of weights/activations. * means using the two-step training setting as ReActNet." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.212, + 0.744, + 0.476 + ], + "angle": 0, + "content": "
ModelMethodBit-width (W/A)Top-1 Acc.(%)Top-5 Acc.(%)
ResNet18Full-precision32/3269.689.2
XNOR [42]1/151.273.2
BiReal [34]1/156.479.5
IR-Net [40]1/158.180.0
RBNN [30]1/159.981.9
SiMaN [29]1/160.182.3
FDA-BNN [52]1/160.282.3
ReCU [53]1/161.082.6
OvSW (Ours)1/161.683.1
ReActNet [33]1/165.986.1
ReCU [53]1/166.486.5
OvSW* (Ours)1/166.686.7
ResNet34Full-precision32/3273.391.3
XNOR++ [3]1/157.179.9
BiReal [34]1/162.283.9
IR-Net [40]1/162.984.1
RBNN [30]1/163.184.4
SiMaN [29]1/163.984.8
CMIM [45]1/165.085.7
ReCU [53]1/165.185.8
OvSW (Ours)1/165.586.1
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.461, + 0.548 + ], + "angle": 0, + "content": "5.2 Results on ImageNet1K" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.841 + ], + "angle": 0, + "content": "On ImageNet1K, OvSW is trained from scratch. We train OvSW with 200 epochs, where the batch size is set to 512 and the initial learning rate is set to 0.1, decaying with CosineAnnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 1e-4 and the data augmentation is the same as ReCU [53]. \\(\\lambda\\) and \\(\\sigma\\) are set to 0.02 and 2e-5 respectively. We demonstrate the ImageNet1K performance of ResNet18/34 and compare OvSW with SOTA methods, including one-stage training methods XNOR [42], BiReal [34], IR-Net [40], RBNN [30], SiMaN [29], FDA-BNN [52], ReCU [53], CMIM [45], and two-stage training method [37] adopted by ReActNet [33]. As shown in Tab. 2, OvSW also achieves the best performance. For ResNet18, OvSW achieves \\(61.6\\%\\) top-1 and \\(83.1\\%\\) top-5 accuracy compared to ReCU's \\(61.0\\%\\) top-1 and \\(82.6\\%\\) top-5 accuracy, which demonstrates the efficiency of overcoming the \"silent weights\". For ResNet34, OvSW achieves \\(65.5\\%\\) top-1 accuracy, which is also better than ReCU. We further compare OvSW with the two-stage training method ReActNet. OvSW obtains \\(66.6\\%\\) top-1 accuracy, succeeding ReActNet and ReCU by \\(0.7\\%\\) and \\(0.2\\%\\) respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.391, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.246, + 0.373, + 0.259 + ], + "angle": 0, + "content": "(a) AGS vs Vanilla" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.147, + 0.576, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.244, + 0.593, + 0.258 + ], + "angle": 0, + "content": "(b) AGS+SAD (OvSW) vs AGS" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.146, + 0.761, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.247, + 0.736, + 0.26 + ], + "angle": 0, + "content": "(c) Convergence" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Fig. 3: Top-1 accuracy (mean±std) of binarized ResNet18 w.r.t. different values of \\(\\lambda\\) (a), \\(\\sigma\\) (b) and epoch (c) on CIFAR100." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.333, + 0.41, + 0.348 + ], + "angle": 0, + "content": "5.3 Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.412 + ], + "angle": 0, + "content": "We investigate the effectiveness of hyper-parameters, including \\(\\lambda\\) and \\(\\sigma\\), convergence speed, different components, and compatibility through ablation analysis. All the following results are based on binarized ResNet18 for CIFAR100." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.787, + 0.565 + ], + "angle": 0, + "content": "\\(\\lambda\\) for AGS and \\(\\sigma\\) for SAD. We first compare AGS with vanilla BNNs and analyze the \\(\\lambda\\) for AGS. In Fig. 3a, we present the results for nine different settings of \\(\\lambda\\), which vary from 0.01 to 0.09. As seen, the appropriate \\(\\lambda\\) can significantly improve the performance of BNNs, and \\(\\lambda = 0.04\\) achieves the best performance. If \\(\\lambda\\) is too small, the sign of the weights can flip inefficiently; while \\(\\lambda\\) is too large, the sign of the weights can flip dramatically, introducing instability to the training. Based on this result, we further introduce SAD to AGS (\\(\\lambda = 0.04\\)) to analyze \\(\\sigma\\). Figure 3b demonstrates that the performance of the model can be further improved by identifying \"silent weights\" and applying additional penalties to them via SAD (\\(\\sigma = 0.0009\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Convergence for OvSW. To verify that OvSW facilitates the flipping of weight signs and thus improves the efficiency of convergence, we fix the \\(\\lambda\\) and \\(\\sigma\\) to 0.04 and 0.0009 respectively, and record the top-1 accuracy over the training epoch from 60 to 500. As shown in Fig. 3c, OvSW effectively accelerates training convergence and achieves better performance. For example, OvSW achieves \\(66.37 \\pm 0.35\\%\\) top-1 accuracy with only 60 training epochs, while vanilla BNNs only reaches \\(55.28 \\pm 0.07\\%\\) and \\(65.23 \\pm 0.21\\%\\) top-1 accuracy with 60 and 120 training epochs respectively. It indicates that OvSW can achieve significant performance gains in scenarios with limited training resources, such as training the network on edge devices. Meanwhile, although increasing epochs improve the final performance for both OvSW and vanilla BNNs, OvSW consistently outperforms the vanilla BNNs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Components. AGS and SAD promote sign flipping for overall weights and \"silent weights\" respectively. To prove that they are orthogonal to each other, we conduct components ablation study for different modules and show the results in Tab. 3 (left). As seen, AGS and SAD achieve \\(69.61 \\pm 0.22\\%\\) and \\(69.45 \\pm 0.18\\%\\) top-1 accuracy respectively and both of them succeed vanilla BNNs. It shows that they are effective in improving the efficiency of weight signs flipping. By" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Left: Ablation study of different components in OvSW. Right: Applying OvSW as a plug-and-play module to other methods." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.185, + 0.486, + 0.259 + ], + "angle": 0, + "content": "
AGSSADmean±std (%)
XX65.23±0.21
✓X69.61±0.22
X✓69.45±0.18
✓✓69.77±0.15
" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.185, + 0.74, + 0.259 + ], + "angle": 0, + "content": "
Methodmean±std (%)
AdaBin70.56±0.19
AdaBin+OvSW72.56±0.11
RBNN67.15±0.23
RBNN+OvSW69.98±0.13
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.784, + 0.314 + ], + "angle": 0, + "content": "combining them, OvSW (AGS+SAD) further achieves the best top-1 accuracy with \\(69.77 \\pm 0.15\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Compatibility. We show the good compatibility of OvSW by inserting it as a plug-and-play module into the current state-of-the-art methods, including AdaBin [48] and RBNN [30]. The former introduces an adaptive binary set to enhance the feature representation of BNNs, while the latter proposes a training-aware approximation function to reduce the gradient estimation error during training. Table 3 (right) demonstrates OvSW achieves \\(72.56 \\pm 0.11\\%\\) top-1 accuracy on AdaBin and \\(69.98 \\pm 0.13\\%\\) on RBNN, succeeding their original performance. This result indicates that OvSW has good compatibility and can effectively enhance the performance of existing methods." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.487, + 0.406, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.596, + 0.367, + 0.608 + ], + "angle": 0, + "content": "(a) Full-precision" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.487, + 0.591, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.596, + 0.557, + 0.609 + ], + "angle": 0, + "content": "(b) XNOR-Net++" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.487, + 0.779, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.596, + 0.72, + 0.608 + ], + "angle": 0, + "content": "(c) OvSW" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.62, + 0.784, + 0.647 + ], + "angle": 0, + "content": "Fig. 4: 3D visualization of the loss surfaces of ResNet18 on CIFAR100, which is used to enable comparisons of sharpness/flatness of different methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.504, + 0.712 + ], + "angle": 0, + "content": "5.4 Loss Landscape Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "BNNs restrict the weights and activations to discrete values, which naturally limits the representational capacity of the model and further result in disparate optimization landscapes compared to real-valued ones [32]. As illustrated in Fig. 4, we follow the method in [26] to plot the actual optimization landscape about our OvSW and compare it with the same architecture to real-valued and XNOR-Net++. As seen, our OvSW has a significantly smoother loss-landscape and minor loss elevation compared to XNOR-Net++, which confirms the effectiveness of OvSW in the BNN optimization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.144, + 0.772, + 0.16 + ], + "angle": 0, + "content": "Table 4: Comparing OvSW (1-bit) with 32-bit and 16-bit backbones on M1 Pro." + }, + { + "type": "table", + "bbox": [ + 0.259, + 0.171, + 0.744, + 0.266 + ], + "angle": 0, + "content": "
NetworkW/ASize (MB)Memory SavingLatency (ms)Acceleration
ResNet1832/3246.76-27.67-
16/1623.382×16.651.66×
1/12.8116.64×7.973.47×
ResNet3432/3287.19-48.82-
16/1643.602×29.011.68×
1/14.1221.16×14.283.42×
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.306, + 0.447, + 0.323 + ], + "angle": 0, + "content": "5.5 Deployment Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.564 + ], + "angle": 0, + "content": "We implement 1-bit models on the M1 Pro, which features 8 high-performance Firestorm cores and 2 efficient Icestorm cores in a hybrid design. The Firestorm cores can be clocked up to \\(3.2\\mathrm{GHz}\\), while the Icestorm cores can reach 2.1 GHz. Our OvSW shows significant efficiency gains when deployed on real-world mobile devices, as evidenced by practical speed evaluations. To make our inference framework BOLT [9] compatible with OvSW, we leverage the ARM NEON SIMD instruction SSHL. We compare OvSW with 32-bit and 16-bit backbones. As shown in Tab. 4, OvSW inference speed is substantially faster with the highly efficient BOLT framework in a single thread. For instance, OvSW achieves an acceleration rate of about \\(3.47\\times\\) on ResNet18 compared to its 32-bit counterpart. For the ResNet34 backbone, OvSW achieves a \\(3.42\\times\\) acceleration rate with the BOLT framework on hardware, which is significant for computer vision applications on real-world edge devices. At the same time, OvSW can save memory by a factor of \\(16.64\\times\\) and \\(21.16\\times\\), which demonstrates its potential for applications with limited memory resources." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.592, + 0.36, + 0.608 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "BNN is a crucial method to compress deep learning models and reduce inference overhead. In this paper, systematically and theoretically, we prove the distribution of gradients is independent of latent weights, which is the root cause of inefficient updating and performance degradation of BNNs. To this end, Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are proposed to Overcome Silent Weights (OvSW) and achieve SOTA performance on BNNs. Specifically, AGS adaptively scales the gradient based on the distribution of weights, improving the efficiency of sign flipping for the overall weights. SAD can effectively measure the states of weight flipping, detect the \"silent weights\" and introduce additional penalty to them to facilitate their flipping. Extensive experiments demonstrate that OvSW can achieve notable performance gains over the SOTA methods on various datasets and networks. In addition, OvSW has better convergence efficiency and excellent compatibility, which can be combined with existing methods to further enhance the performance of BNNs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: Int. Conf. Mach. Learn. pp. 1059-1071. PMLR (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Bulat, A., Tzimiropoulos, G.: Xnor-net++: Improved binary neural networks. arXiv preprint arXiv:1909.13863 (2019) 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Chen, J., Liu, L., Liu, Y., Zeng, X.: A learning framework for n-bit quantized neural networks toward fpgas. IEEE Transactions on Neural Networks and Learning Systems 32(3), 1067-1081 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Courbariaux, M., Hubara, I., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks: Training deep neural networks with weights and activations constrained to + 1 or -1. arXiv preprint arXiv:1602.02830 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "6. Ding, R., Chin, T.W., Liu, Z., Marculescu, D.: Regularizing activation distribution for training binarized deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11408-11417 (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13733-13742 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "8. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench101: Improving transferability and generalizability of cross-task neural architecture search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5251-5260 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.688, + 0.508 + ], + "angle": 0, + "content": "9. Feng, J.: Bolt. https://github.com/huawei-noah/bolt (2021) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "0. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 580-587 (2014) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "1. Gong, R., Liu, X., Jiang, S., Li, T., Hu, P., Lin, J., Yu, F., Yan, J.: Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In: Int. Conf. Comput. Vis. pp. 4852–4861 (2019) 2, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "2. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Int. Conf. Comput. Vis. pp. 2961-2969 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Int. Conf. Comput. Vis. pp. 1026-1034 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "4. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans. Pattern Anal. Mach. Intell. 37(9), 1904-1916 (2015) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "6. He, X., Mo, Z., Cheng, K., Xu, W., Hu, Q., Wang, P., Liu, Q., Cheng, J.: Proxybnn: Learning binarized neural networks via proxy matrices. In: Eur. Conf. Comput. Vis. pp. 223-241. Springer (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "7. Helwegen, K., Widdicombe, J., Geiger, L., Liu, Z., Cheng, K.T., Nusselder, R.: Latent weights do not exist: Rethinking binarized neural network optimization. Adv. Neural Inform. Process. Syst. 32 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Hinton, G., Vinyals, O., Dean, J., et al.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 2(7) (2015) 2" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC). pp. 10-14. IEEE (2014) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.243 + ], + "angle": 0, + "content": "20. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2704-2713 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.271 + ], + "angle": 0, + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.324 + ], + "angle": 0, + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst. 25 (2012) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "24. Lee, C., Kim, H., Park, E., Kim, J.J.: Insta-bnn: Binary neural network with instance-aware threshold. In: Int. Conf. Comput. Vis. pp. 17325-17334 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.378 + ], + "angle": 0, + "content": "25. Lee, J., Kim, D., Ham, B.: Network quantization with element-wise gradient scaling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6448-6457 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.378, + 0.785, + 0.405 + ], + "angle": 0, + "content": "26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. In: Adv. Neural Inform. Process. Syst. (2018) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "27. Li, S., Lin, M., Wang, Y., Fei, C., Shao, L., Ji, R.: Learning efficient gans for image translation via differentiable masks and co-attention distillation. IEEE Trans. Multimedia (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.487 + ], + "angle": 0, + "content": "28. Lin, M., Ji, R., Wang, Y., Zhang, Y., Zhang, B., Tian, Y., Shao, L.: Hrank: Filter pruning using high-rank feature map. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1529-1538 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.527 + ], + "angle": 0, + "content": "29. Lin, M., Ji, R., Xu, Z., Zhang, B., Chao, F., Lin, C.W., Shao, L.: Siman: Sign-to-magnitude network binarization. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.527, + 0.785, + 0.568 + ], + "angle": 0, + "content": "30. Lin, M., Ji, R., Xu, Z., Zhang, B., Wang, Y., Wu, Y., Huang, F., Lin, C.W.: Rotated binary neural network. Adv. Neural Inform. Process. Syst. 33, 7474-7485 (2020) 4, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "31. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Adv. Neural Inform. Process. Syst. 30 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.636 + ], + "angle": 0, + "content": "32. Liu, Z., Shen, Z., Li, S., Helwegen, K., Huang, D., Cheng, K.T.: How do adam and training strategies help bnns optimization. In: Int. Conf. Mach. Learn. pp. 6936-6946. PMLR (2021) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.785, + 0.677 + ], + "angle": 0, + "content": "33. Liu, Z., Shen, Z., Savvides, M., Cheng, K.T.: Reactnet: Towards precise binary neural network with generalized activation functions. In: Eur. Conf. Comput. Vis. pp. 143-159. Springer (2020) 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.731 + ], + "angle": 0, + "content": "34. Liu, Z., Wu, B., Luo, W., Yang, X., Liu, W., Cheng, K.T.: Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In: Eur. Conf. Comput. Vis. pp. 722-737 (2018) 4, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3431-3440 (2015) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "36. Luo, J.H., Wu, J.: Autopruner: An end-to-end trainable filter pruning method for efficient deep model inference. Pattern Recognition 107, 107461 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.838 + ], + "angle": 0, + "content": "37. Martinez, B., Yang, J., Bulat, A., Tzimiropoulos, G.: Training binary neural networks with real-to-binary convolutions. arXiv preprint arXiv:2003.11535 (2020) 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "38. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: Int. Conf. Mach. Learn. pp. 16318-16330. PMLR (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.302 + ], + "angle": 0, + "content": "39. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché Buc, F., Fox, E., Garnett, R. (eds.) Adv. Neural Inform. Process. Syst. pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.302, + 0.787, + 0.344 + ], + "angle": 0, + "content": "40. Qin, H., Gong, R., Liu, X., Shen, M., Wei, Z., Yu, F., Song, J.: Forward and backward information retention for accurate binary neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2250-2259 (2020) 4, 5, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.344, + 0.787, + 0.4 + ], + "angle": 0, + "content": "41. Rajbhandari, S., Rasley, J., Ruwase, O., He, Y.: Zero: Memory optimizations toward training trillion parameter models. In: SC20: International Conference for High Performance Computing, Networking, Storage and Analysis. pp. 1-16. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.4, + 0.787, + 0.443 + ], + "angle": 0, + "content": "42. Rastegari, M., Ordonez, V., Redmon, J., Farhadi, A.: Xnor-net: Imagenet classification using binary convolutional neural networks. In: Eur. Conf. Comput. Vis. pp. 525-542. Springer (2016) 2, 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.444, + 0.787, + 0.486 + ], + "angle": 0, + "content": "43. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.486, + 0.787, + 0.528 + ], + "angle": 0, + "content": "44. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.556 + ], + "angle": 0, + "content": "45. Shang, Y., Xu, D., Zong, Z., Yan, Y.: Network binarization via contrastive learning. arXiv preprint arXiv:2207.02970 (2022) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.787, + 0.585 + ], + "angle": 0, + "content": "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.586, + 0.787, + 0.628 + ], + "angle": 0, + "content": "47. Su, X., Huang, T., Li, Y., You, S., Wang, F., Qian, C., Zhang, C., Xu, C.: Prioritized architecture sampling with monto-carlo tree search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10968-10977 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.669 + ], + "angle": 0, + "content": "48. Tu, Z., Chen, X., Ren, P., Wang, Y.: Adabin: Improving binary neural networks with adaptive binary sets. In: Eur. Conf. Comput. Vis. pp. 379-395. Springer (2022) 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.727 + ], + "angle": 0, + "content": "49. Wu, X.M., Zheng, D., Liu, Z., Zheng, W.S.: Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17055-17064 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "50. Xu, S., Li, Y., Ma, T., Lin, M., Dong, H., Zhang, B., Gao, P., Lu, J.: Resilient binary neural network. In: AAAI. vol. 37, pp. 10620-10628 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.798 + ], + "angle": 0, + "content": "51. Xu, S., Li, Y., Wang, T., Ma, T., Zhang, B., Gao, P., Qiao, Y., Lü, J., Guo, G.: Recurrent bilinear optimization for binary neural networks. In: Eur. Conf. Comput. Vis. pp. 19-35. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "52. Xu, Y., Han, K., Xu, C., Tang, Y., Xu, C., Wang, Y.: Learning frequency domain approximation for binary neural networks. Adv. Neural Inform. Process. Syst. 34, 25553-25565 (2021) 10, 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Xiang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "53. Xu, Z., Lin, M., Liu, J., Chen, J., Shao, L., Gao, Y., Tian, Y., Ji, R.: Recu: Reviving the dead weights in binary neural networks. In: Int. Conf. Comput. Vis. pp. 5198-5208 (2021) 3, 4, 5, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "54. Yang, J., Shen, X., Xing, J., Tian, X., Li, H., Deng, B., Huang, J., Hua, X.s.: Quantization networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7308-7316 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "55. Yang, Z., Wang, Y., Han, K., Xu, C., Xu, C., Tao, D., Xu, C.: Searching for low-bit weights in quantized neural networks. Adv. Neural Inform. Process. Syst. 33, 4091-4102 (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.785, + 0.3 + ], + "angle": 0, + "content": "56. You, Y., Gitman, I., Ginsburg, B.: Large batch training of convolutional networks. arXiv preprint arXiv:1708.03888 (2017) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "57. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Eur. Conf. Comput. Vis. pp. 365-382 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "58. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6848-6856 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "59. Zhao, J., Yang, L., Zhang, B., Guo, G., Doermann, D.S.: Uncertainty-aware binary neural networks. In: IJCAI. pp. 3441-3447 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "60. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.785, + 0.495 + ], + "angle": 0, + "content": "61. Zhu, F., Gong, R., Yu, F., Liu, X., Wang, Y., Li, Z., Yang, X., Yan, J.: Towards unified int8 training for convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1969-1979 (2020) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.495 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_origin.pdf b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2a64cb3e00bc84077e399118e0bbfbc6383b6d8d --- /dev/null +++ b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/43eb14f0-154a-457a-9941-4861ded25366_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1c972fcac67e8889dc29c3bc3f61468f3d5bee36ee7bf948e386308324d9f13 +size 857331 diff --git a/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/full.md b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4d4993a2dacfbe3e151a1fa207d92dae05a671c2 --- /dev/null +++ b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/full.md @@ -0,0 +1,356 @@ +# OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks + +Jingyang Xiang $^{1}$ , Zuohui Chen $^{2}$ , Siqi Li $^{1}$ , Qing Wu $^{3}$ , and Yong Liu $^{1,4}$ + +$^{1}$ APRIL Lab, Zhejiang University + +2 Zhejiang University of Technology + +3 College of Computer Science, Hangzhou Dianzi University + +4 Huzhou Institute, Zhejiang University + +Abstract. Binary Neural Networks (BNNs) have been proven to be highly effective for deploying deep neural networks on mobile and embedded platforms. Most existing works focus on minimizing quantization errors, improving representation ability, or designing gradient approximations to alleviate gradient mismatch in BNNs, while leaving the weight sign flipping, a critical factor for achieving powerful BNNs, untouched. In this paper, we investigate the efficiency of weight sign updates in BNNs. We observe that, for vanilla BNNs, over $50\%$ of the weights remain their signs unchanged during training, and these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. We refer to these weights as "silent weights", which slow down convergence and lead to a significant accuracy degradation. Theoretically, we reveal this is due to the independence of the BNNs gradient from the latent weight distribution. To address the issue, we propose Overcome Silent Weights (OvSW). OvSW first employs Adaptive Gradient Scaling (AGS) to establish a relationship between the gradient and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Additionally, we design Silence Awareness Decaying (SAD) to automatically identify "silent weights" by tracking weight flipping state, and apply an additional penalty to "silent weights" to facilitate their flipping. By efficiently updating weight signs, our method achieves faster convergence and state-of-the-art performance on CIFAR10 and ImageNet1K dataset with various architectures. For example, OvSW obtains $61.6\%$ and $65.5\%$ top-1 accuracy on the ImageNet1K using binarized ResNet18 and ResNet34 architecture respectively. Codes are available at https://github.com/JingyangXiang/OvSW. + +Keywords: Binary Neural Networks $\cdot$ Silent Weights $\cdot$ Adaptive Gradient Scaling $\cdot$ Silence Awareness Decaying + +# 1 Introduction + +Deep neural networks (DNNs) have shown tremendous success in various computer vision tasks, including image classification [15, 23], object detection [10, 14, + +![](images/aa0599de141f25a3e675a8e7cc2915ecd394dde387fde65361b145a01d95251f.jpg) +(a) Vanilla + +![](images/3619b96a085883ae65264adcd23e8d5335158e452f910b9157e61590abc48866.jpg) +(b) OvSW +Fig. 1: (a) and (b): Histogram of the initialized weight distribution (blue) and the weights that never update signs throughout training (orange). $54.07\%$ and $2.03\%$ represent the ratio of the corresponding orange area to the blue. (c) Top-1 Accuracy (solid lines) and Flip Ratio (dashed lines) are for Vanilla (green) and OvSW (red). (a), (b) and Flip Ratio in (c) is for layer4(conv2.weight. + +![](images/835fe892beb883b41ac6b6b6f4cf79680a2fe929679a91a43a7219685a569510.jpg) +(c) Top-1 Acc & Flip Ratio + +43], and semantic segmentation [12, 35]. However, the remarkable performance is always attributed to deeper and wider architectures [15, 46], which comes with expensive memory and computational overhead and makes it challenging to deploy DNNs on resource-constrained edge platforms. + +The community has been delving into model compression, which aims to reduce inference overhead for DNNs while preserving their performance. Typical techniques include, but are not limited to, efficient architecture design [7, 58], neural architecture search [8, 47], network pruning [28, 36], knowledge distillation [18, 27], and network quantization [4, 11, 61]. Among them, network quantization is widely suggested as a promising solution. It effectively enhances memory efficiency and execution speed on embedding platforms by reducing the weight and activation bits and replacing expensive floating-point arithmetic with relatively inexpensive fixed-point arithmetic. Extensive studies [11, 20, 54, 57] have demonstrated its effectiveness and yielded light and efficient DNNs. + +In this paper, we focus on studying binary neural networks (BNNs), which are an aggressive form of quantized neural networks (QNNs). BNNs binarize both weights and activations to discrete values $\{+1, -1\}$ , which can be represented by 1-bit and reduce memory usage by $32 \times$ . On the other hand, by employing efficient bitwise operations instead of floating-point ones, BNNs also significantly reduce computation complexity, providing $58 \times$ speedup as reported by XNOR-Net [42]. These characteristics make them well-suited for deployment on low-power embedding platforms, including FPGA, ASICs, and IoT devices [19]. + +However, the application of BNNs is still limited due to constraining both weights and activations to 1-bit significantly reducing accuracy compared to full-precision models. To address this problem, many methods have been developed to reduce quantization error and enhance the representation capability, thereby closing the performance gap between BNNs and their real-valued counterparts. + +Although these methods have effectively narrowed the performance gap, they still overlook the essence of BNN optimization is to update the signs of latent weights [17], instead of updating their values. In other words, if the + +signs of the latent weights don't change, the BNNs are hardly ever updated. To demonstrate the weight sign flipping of vanilla BNNs, we train binarized ResNet18 on CIFAR100 with 120 epochs and plot the flipping information for layer4 conv2 weight. As shown in Fig. 1a, more than $50\%$ of the weights never change their signs throughout training process, which slows down convergence and leads to a significant accuracy degradation for BNNs. We denote the part of weights that hardly change their signs during training as "silent weights". + +It's worth noting that, as a study mostly related to ours, ReCU [53] proposes a rectified clamp unit to revive the "dead weights" for updating, which refers to a group of weights that are distributed at the tails of the weight distribution and barely update their signs during training. They claim that the magnitudes of latent weights do not contribute to the forward propagation. If two weights have the same sign, they have the same effect on the forward propagation. Intuitively, the signs of the weights around the zero are easily changed, while weights distributed at the tails of the weight distribution tend to have difficulties in changing their signs. Although ReCU facilitates the updating for "dead weights", it ignores the fact, the gradient distribution across layers and weights have great distinction. Weights with larger magnitudes are intuitively more difficult to flip their signs, but small weights may also suffer from silence once their gradients are relatively small compared to their magnitudes. As shown in Fig. 1a, for a vanilla BNN, these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. As a result, it is still sub-optimal and remains to be improved. + +In this paper, we reveal that the root cause of the large number of "silent weights" is attributed to the independence of the BNN gradients from the latent weight distribution via a systematical and theoretical analysis. To this end, two simple yet effective techniques, including Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are introduced to Overcome Silent Weights (OvSW). Specifically, AGS adaptively scales the gradients by establishing a relationship between the gradients and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Meanwhile, SAD automatically identifies "silent weights" by tracking weight sign flipping state and applies an additional penalty to them, further promoting the updating efficiency of their signs. Benefiting from AGS and SAD, OvSW enables efficient flipping of weight signs, significantly accelerates the convergence and promotes the performance for BNNs as shown in Fig. 1b and Fig. 1c. In comparison to the expensive computational cost of forward and backward propagation, OvSW introduces negligible extra overhead to the training process, maintaining its efficiency. Furthermore, since OvSW aims to facilitate the weight sign flipping for BNNs, which is orthogonal to the previous work, it has good compatibility and can be used as plug-and-play modules to enhance other BNNs' performance. + +The contributions of our work are highlighted as follows: + +- We are the first to find that a large number of weights in BNNs fail to update their signs throughout the training. These "silent weights" are not only distributed at the tails of the weight distribution but also universally + +present in the vicinity of zero. Theoretically, we reveal this is attributed to the independence of the gradients from the distribution of the latent weights. + +- We propose to overcome "silent weights" with two novel techniques: (1) an adaptive gradient scaling method to scale the gradients according to the distribution of the latent weights, which improves overall efficiency in updating signs for weights; (2) a silence awareness decaying strategy to identify "silent weights" by tracking weight sign flipping state and introduce an additional penalty to them, further facilitating the flipping of their signs. + +- Extensive experiments for various BNNs including ResNet18/20 [15] and VGGsmall [46] on CIFAR10 [22] and ResNet18/34 [15] on ImageNet1K [44] demonstrate the effectiveness of our method. For example, OvSW achieves $61.6\%$ and $65.5\%$ top-1 accuracy on the ImageNet for binarized ResNet18 and ResNet34, improving ReCU [53] by $0.6\%$ and $0.4\%$ respectively. Besides, for ResNet18, OvSW achieves $2.00\%$ and $2.83\%$ improvement on CIFAR100 [22] when combined with AdaBin (enhancing representation ability) and RBNN (training-aware gradient approximation) respectively, demonstrating its good compatibility with the other methods. + +# 2 Related Work + +Great efforts have been put into reducing the performance gap between BNNs and their real-valued counterparts. The pioneering BNN work dates back to Courbariaux et al. [5], which binarizes weights and activations to $+1$ or $-1$ by sign function. However, this aggressive approach limits the representation ability of BNNs to the binary space, resulting in a significant accuracy degradation. To reduce the quantization error and improve their accuracy, XNOR-Net [42] introduces a scaling factor obtained through the $\ell_1$ -norm of weights or activations. Furthermore, XNOR-Net++ [3] merges two scale factors from the weights and activations into a trainable parameter and optimizes them via backpropagation. ABC-Net [31] employs a linear combination of multiple binary weight sets to approximate the real-valued weights and alleviate the information loss. Bi-Real Net [34] connects the real-valued activations to the consecutive block via an identity shortcut, which significantly enhances network representation ability while incurring negligible computational overhead. AdaBin [48] introduces adaptive binary sets to fit different distributions, enhancing binarized representation ability. UaBNN [59] introduces an uncertainty-aware BNN to reduce these weights' uncertainties. INSTA-BNN [24] controls the quantization threshold in an input instance-aware manner, taking higher-order statistics into account. + +Apart from enhancing the representation ability of BNNs, gradient estimation is also one of the critical research directions, since gradients in the sign function are almost zero everywhere. Straight through estimator (STE) [1] is the most widely used function to enable the gradient to backpropagate. However, the gradient error is huge for STE and will accumulate during backpropagation, leading to instability in training and severe accuracy degradation. To alleviate this, Bi-Real Net [34] utilizes a piecewise polynomial function as the approximation function. IR-Net [40] proposes an error decay estimator and RBNN [30] + +employs a training-aware approximation function to replace the sign function. EWGS [25] takes discretization error between input and output into account and introduces element-wise gradient scaling to adaptively scale up or down each gradient element. ReSTE [49] revises the original STE and balance the estimating error and the gradient stability well. All of these methods effectively reduce the gradient error and achieve consistent improvement in both training stability and accuracy compared to the vanilla STE. + +# 3 Background + +We briefly review the optimization process of BNNs in this section. Given a DNN, we denote $\mathcal{W}_j \in \mathbb{R}^{C_{\mathrm{out}}^j \times C_{\mathrm{in}}^j \times K_{\mathrm{h}}^j \times K_{\mathrm{w}}^j}$ as the real-valued weight in the $j$ -th layer, $C_{\mathrm{out}}^j$ , $C_{\mathrm{in}}^j$ , $K_{\mathrm{h}}^j$ and $K_{\mathrm{w}}^j$ are the number of output channels, input channels, kernel height, and kernel width, respectively. Let the real-valued activation be $\mathcal{A}_j$ , then the convolution process can be formulated as + +$$ +\mathcal {A} _ {j + 1} = \mathcal {W} _ {j} * \mathcal {A} _ {j}, \tag {1} +$$ + +where $*$ represents the standard convolution operation. + +BNNs aim to binarize weights $\mathcal{W}_j$ and activations $\mathcal{A}_j$ to discrete values $(\{+1, -1\})$ through sign function: + +$$ +\hat {x} = \operatorname {s i g n} (x) = \left\{ \begin{array}{l} + 1, \text {i f} x \geq 0, \\ - 1, \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +To reduce the quantization error in BNNs, XNOR-Net [42] introduces two scale factors to approximate the binarized weights $\hat{\mathcal{W}}_j^b$ and activation $\hat{\mathcal{A}}_j^b$ . Furthermore, XNOR-Net++ [3] proposes fusing the activation and weight scaling factors into one and optimizing it via backpropagation. This approach significantly outperforms XNOR-Net within the same computational budget and has been widely adopted by recent works [33, 34, 40, 53]. Following them, we denote the scaling factor as $\alpha_{j}$ . Then the binary convolution operation can be formulated as + +$$ +\mathcal {A} _ {j + 1} = \left(\hat {\mathcal {A}} _ {j} * \hat {\mathcal {W}} _ {j}\right) \odot \alpha_ {j}, \tag {3} +$$ + +where $\circ$ is the efficient XNOR and Bitcount operation and $\odot$ is theadamard product. In the implementation of BNNs, $\mathcal{A}_{j + 1}$ will be processed through several layers, e.g., Batch Normalization (BN) layer, non-linear activation layer, and max-pooling layer. In this section, we omit these operations for simplicity. + +To train a BNN, the forward propagation includes Eq. (2) and Eq. (3), where their real-value counterparts $\mathcal{W}_j$ and $\mathcal{A}_j$ are used for calculating gradients and updating during the backpropagation. However, the sign is not differentiable, thereby gradient estimation is important in BNNs. Following the previous studies, we use straight-through estimator (STE) [1] to approximate the gradient of the loss w.r.t. the weight $w\in \mathcal{W}$ : + +$$ +\frac {\partial \mathcal {L}}{\partial w} = \frac {\partial \mathcal {L}}{\partial \hat {w}} \cdot \frac {\partial \hat {w}}{\partial w} \approx \frac {\partial \mathcal {L}}{\partial \hat {w}}, \tag {4} +$$ + +![](images/59c277c91b6d9d4e2a2ff49e77ebfbbcb06174c22dbbaaa2575d510eb86a770a.jpg) +Fig. 2: Forward and backward computation graph for binary convolutional operation with quantization aware training. + +where $\mathcal{L}$ denotes the loss function. For the gradient w.r.t. the activation $a\in \mathcal{A}$ , we adopt the piece-wise polynomial gradient estimation function [34] as follows: + +$$ +\frac {\partial \mathcal {L}}{\partial a} = \frac {\partial \mathcal {L}}{\partial \hat {a}} \cdot \frac {\partial \hat {a}}{\partial a} \approx \frac {\partial \mathcal {L}}{\partial \hat {a}} \cdot \frac {\partial F (a)}{\partial a}, \tag {5} +$$ + +where + +$$ +\frac {\partial F (a)}{\partial a} = \left\{ \begin{array}{l l} 2 + 2 a, & \text {i f} - 1 \leq a < 0, \\ 2 - 2 a, & \text {i f} 0 \leq a < 1, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +# 4 Methodology + +In this section, we show that the distribution of gradients and weights for BNNs is independent by systematical and theoretical analysis. Then, we propose Adaptive Gradient Scaling (AGS) to scale the gradient and introduce Silence Awareness Decaying (SAD) to detect "silent weights", moving them towards zero. Both of them can enhance the efficiency of weight sign flipping. + +# 4.1 The Independence of the Gradient and Weight Distribution + +Figure 2 demonstrates forward and backward computation graph for binary convolutional (Bin-Conv) operation. As seen, Bin-Conv is often followed by a BN layer. We can conclude that once two BNNs $\mathcal{N}$ and $\mathcal{N}'$ satisfy: + +$$ +\hat {\mathcal {W}} _ {j} = \hat {\mathcal {W}} _ {j} ^ {\prime}, \alpha_ {j} = \alpha_ {j} ^ {\prime}, \mathrm {B N} _ {j} = \mathrm {B N} _ {j} ^ {\prime}, \forall j, \tag {7} +$$ + +for the same input $\hat{\mathcal{A}}_j$ , the output $\mathcal{A}_{j + 1}$ and $\mathcal{A}_{j + 1}^{\prime}$ can be formulated as: + +$$ +\mathcal {A} _ {j + 1} = \mathrm {B N} \left(\left(\hat {\mathcal {A}} _ {j} * \hat {\mathcal {W}} _ {j}\right) \odot \alpha_ {j}\right) = \mathrm {B N} ^ {\prime} \left(\left(\hat {\mathcal {A}} _ {j} * \hat {\mathcal {W}} _ {j} ^ {\prime}\right) \odot \alpha_ {j} ^ {\prime}\right) = \mathcal {A} _ {j + 1} ^ {\prime}. \tag {8} +$$ + +Through mathematical induction, we can know that the $\mathcal{N}$ and $\mathcal{N}'$ will have the same output and loss, i.e. $\mathcal{L} = \mathcal{L}'$ . Therefore, the backpropagation for them is: + +$$ +\frac {\partial \mathcal {L}}{\partial \mathcal {W} _ {j}} = \frac {\partial \mathcal {L}}{\partial \mathcal {A} _ {j + 1}} \frac {\partial \mathcal {A} _ {j + 1}}{\partial \hat {\mathcal {W}} _ {j}} \frac {\partial \hat {\mathcal {W}} _ {j}}{\partial \mathcal {W} _ {j}} = \frac {\partial \mathcal {L} ^ {\prime}}{\partial \mathcal {A} _ {j + 1} ^ {\prime}} \frac {\partial \mathcal {A} _ {j + 1} ^ {\prime}}{\partial \hat {\mathcal {W}} _ {j} ^ {\prime}} \frac {\hat {\mathcal {W}} _ {j} ^ {\prime}}{\partial \mathcal {W} _ {j} ^ {\prime}} = \frac {\partial \mathcal {L} ^ {\prime}}{\partial \mathcal {W} _ {j} ^ {\prime}}. \tag {9} +$$ + +Without loss of generality, we can obtain that once two networks satisfy the conditions in Eq. (7), the gradients obtained from the backpropagation are the same (a more generalized scenario are extrapolated in ??). The gradient has nothing to do with the magnitude of weight and weight with larger magnitude does not necessarily corresponding to larger gradient. Therefore, although intuitively, weight with large magnitude tends to remain its sign unchanged, it is also potential for weight with small magnitude to suffer from silence if its gradient is relatively small compared to its magnitude. Furthermore, if we assume the weights further satisfy $\mathcal{W}_j' = \gamma \mathcal{W}_j$ , where $\gamma > 1$ and networks are optimized via vanilla SGD, the optimization of $t$ -step can be formulated as: + +$$ +\begin{array}{l} \mathcal {W} _ {j} ^ {\prime} (t + 1) = \mathcal {W} _ {j} ^ {\prime} (t) - \beta (t) \frac {\partial \mathcal {L} ^ {\prime} (t)}{\partial \mathcal {W} _ {j} ^ {\prime} (t)} = \gamma \mathcal {W} _ {j} (t) - \beta (t) \frac {\partial \mathcal {L} (t)}{\partial \mathcal {W} _ {j} (t)} \tag {10} \\ = (\gamma - 1) \mathcal {W} _ {j} (t) + \mathcal {W} _ {j} (t) - \beta (t) \frac {\partial \mathcal {L} (t)}{\partial \mathcal {W} _ {j} (t)} = (\gamma - 1) \mathcal {W} _ {j} (t) + \mathcal {W} _ {j} (t + 1), \\ \end{array} +$$ + +where $\beta(t)$ is the learning rate at $t$ -step. Equation (10) can be viewed as a variant of exponential moving averages and the effect of $\mathcal{W}_j(t)$ on $\mathcal{W}_j'(t + 1)$ will gradually increase as $\gamma$ increases. Once $k$ is large enough: + +$$ +\lim _ {\gamma \rightarrow \infty} \mathcal {W} _ {j} ^ {\prime} (t + 1) = \lim _ {\gamma \rightarrow \infty} [ (\gamma - 1) \mathcal {W} _ {j} (t) + \mathcal {W} _ {j} (t + 1) ] = (\gamma - 1) \mathcal {W} _ {j} (t). \tag {11} +$$ + +The signs of $\mathcal{W}_j^{\prime}(t + 1)$ will be the same as $\mathcal{W}_j(t)$ . + +However, BNNs binarize $\mathcal{W}_j'$ to $\{+1, -1\}$ at each training step. Once the signs of $\mathcal{W}_j'$ do not change, the BNNs hardly ever update. It not only slows down convergence, but also leads to significant accuracy degradation. Please see ?? for a detailed ablation analysis for $\gamma$ . + +# 4.2 Adaptive Gradient Scaling + +To improve the efficiency of the update, an intuitive approach is to scale $\beta(t)$ to $\gamma\beta(t)$ . In this case, $\mathcal{W}_j'(t + 1)$ can be formulated as: + +$$ +\begin{array}{l} \mathcal {W} _ {j} ^ {\prime} (t + 1) = \mathcal {W} _ {j} ^ {\prime} (t) - \gamma \beta (t) \frac {\partial \mathcal {L} (t)}{\partial \mathcal {W} _ {j} ^ {\prime} (t)} \tag {12} \\ = \gamma \mathcal {W} _ {j} (t) - \gamma \beta (t) \frac {\partial \mathcal {L} (t)}{\partial \mathcal {W} _ {j} (t)} = \gamma \mathcal {W} _ {j} (t + 1). \\ \end{array} +$$ + +However, selecting the appropriate scaling factor is tricky. Meanwhile, an inappropriate scaling factor may make weights with small magnitudes correspond to large gradients, leading them to oscillate frequently in $\{+1, -1\}$ and introducing instability to BNNs training. + +To overcome this issue, we introduce "Adaptive Gradient Scaling" (AGS). Let $\mathcal{G}_j\in \mathbb{R}^{C_j^j\times C_j^j\times K_h^j\times K_w^j}$ denote the gradient with respect to $\mathcal{W}_j$ , i.e., $\frac{\partial\mathcal{L}}{\partial\mathcal{W}_j}$ + +and $\| \cdot \| _F$ denote the Forbenius norm, i.e., the $k$ -th filter norm in $\mathcal{W}_j$ can be formulated as: + +$$ +\left\| \mathcal {W} _ {j} ^ {k} \right\| _ {F} = \sqrt {\sum_ {l = 1} ^ {C _ {\mathrm {i n}} ^ {j}} \sum_ {m = 1} ^ {K _ {\mathrm {h}} ^ {j}} \sum_ {n = 1} ^ {K _ {\mathrm {w}} ^ {j}} \mathcal {W} _ {j} ^ {k , l , m , n}}. \tag {13} +$$ + +The AGS algorithm is motivated by the observation that the ratio of the norm of $\mathcal{G}_j^k$ to $\mathcal{W}_j^k\left(\frac{\|\mathcal{G}_j^k\|_F}{\|\mathcal{W}_j^k\|_F}\right)$ provides a simple measure of how much a single gradient descent step will change the original weight $\mathcal{W}_j$ . For instance, if we train BNNs via vanilla SGD without momentum, then $\frac{\|\triangle\mathcal{W}_j^k\|_F}{\|\mathcal{W}_j^k\|_F} = \beta \frac{\|\mathcal{G}_j^k\|_F}{\|\mathcal{W}_j^k\|_F}$ , where the parameter update for the $\mathcal{W}_j^k$ is given by $\triangle \mathcal{W}_j^k = -\beta \mathcal{G}_j^k$ . We can conclude that the small value of $\frac{\|\mathcal{G}_j^k\|_F}{\|\mathcal{W}_j^k\|_F}$ during training is the root cause why a large number of weights fail to flip signs and adaptively scaling $\frac{\|\mathcal{G}_j^k\|_F}{\|\mathcal{W}_j^k\|_F}$ plays a crucial role in promoting the update efficiency for BNNs. Specifically, in AGS algorithm, $\mathcal{G}_j^{k,l,m,n}$ is scaled as: + +$$ +\overline {{\mathcal {G}}} _ {j} ^ {k, l, m, n} = \left\{ \begin{array}{l l} \lambda \frac {\| \mathcal {W} _ {j} ^ {k} \| _ {F}}{\| \mathcal {G} _ {j} ^ {k} \| _ {F}} \mathcal {G} _ {j} ^ {k, l, m, n} & \text {i f} \frac {\| \mathcal {G} _ {j} ^ {k} \| _ {F}}{\| \mathcal {W} _ {j} ^ {k} \| _ {F}} < \lambda , \\ \mathcal {G} _ {j} ^ {k, l, m, n} & \text {o t h e r w i s e .} \end{array} \right. \tag {14} +$$ + +where $\lambda$ is a scalar scaling threshold to limits the lower bound of $\frac{\|\mathcal{G}_j^k\|_F}{\|\mathcal{W}_j^k\|_F}$ . Ablation analysis for $\lambda$ can be found in Sec. 5.3. + +AGS and "Adaptive Gradient Clipping" (AGC) [2] are closely related but fundamentally different, as the former restricts the lower bound of $\frac{\|\mathcal{G}_j^k\|_F}{\|\mathcal{W}_j^k\|_F}$ to enhance the flipping efficiency of BNN weight signs, while the latter restricts upper bound and is designed to improve training stability. AGS also can be viewed as a adaptive variant to Layer-wise Adaptive Rate Scaling (LARS) [56], which sets the norm of update parameter to a fixed ratio of the parameter norm, and completely ignores the gradient magnitude to real-valued networks. Although LARS is also able to improve the accuracy for BNNs, we find that LARS and AGS is much different in the update of gradient momentum and doing so degrades performance compared to AGS. More details for comparing AGS with LARS and ablation analysis for LARS can be found in ?? + +# 4.3 Silence Awareness Decaying + +We propose another approach (Silence Awareness Decaying, SAD) orthogonal to AGS to detect and prevent "silent weights". Specifically, we track the flipping state of $\mathcal{W}_j$ over time using an exponential moving average (EMA) strategy, which is formulated as: + +$$ +\mathcal {S} _ {j} (t) = m \cdot \mathcal {S} _ {j} (t - 1) + (1 - m) \cdot \frac {\left| \operatorname {s i g n} \left(\mathcal {W} _ {j} (t)\right) - \operatorname {s i g n} \left(\mathcal {W} _ {j} (t - 1)\right) \right| _ {a b s}}{2}, \tag {15} +$$ + +Algorithm 1 Overview of the OvSW method. +Input: A minibatch of inputs and their labels, real-valued weights $\mathcal{W}(t)$ , scaling factor $\alpha(t)$ , $\lambda$ for AGS, $\tau$ for SAD, $(S(t), m, \sigma)$ for flipping state detection. +Output: Updated $\mathcal{W}(t+1)$ , $\alpha(t+1)$ and $S(t+1)$ . +1: while Forward propagation do +2: $\hat{\mathcal{A}}_j \gets \mathrm{sign}(\mathcal{A}_j)$ . +3: $\hat{\mathcal{W}}_j \gets \mathrm{sign}(\mathcal{W}_j)$ . +4: Computing features via Eq. (2) and Eq. (3). +5: Computing loss $\mathcal{L}$ . +6: end while +7: while Backward propagation do +8: Computing $\frac{\partial\mathcal{L}}{\partial\mathcal{W}_j}$ , i.e. $\mathcal{G}_j$ , and $\frac{\partial\mathcal{L}}{\partial\mathcal{A}_j}$ via Eq. (4) and Eq. (5)&Eq. (6) +9: Scaling $\mathcal{G}_j$ to $\overline{\mathcal{G}}_j$ via Eq. (14). +10: Adding penalties to $\overline{\mathcal{G}}_j$ via Eq. (16). +11: Update $\mathcal{W}_j(t+1)$ and $\alpha_j(t+1)$ via SGD optimizer. +12: Update $\mathcal{S}_j(t+1)$ via Eq. (15). +13: end while + +where $m$ is the momentum and $S_{j}$ is the auxiliary variable. Nagel et al. [38] employ this technique to identify and dampen oscillations prematurely while we employ it to identify "silent weights" and dynamically introduce additional weight penalties to move them towards zero. In our algorithm, we think if $S_{j}^{k,l,m,n}$ is less than a pre-defined threshold $\sigma$ , its corresponding weight $\mathcal{W}_{j}^{k,l,m,n}$ is considered as a "silent weight" and will be applied with an additional penalty. The silence awareness decaying process is formulated as: + +$$ +\overline {{\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) = \left\{ \begin{array}{l l} \overline {{\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) + \gamma \mathcal {W} _ {j} ^ {k, l, m, n} (t), & \text {i f} \mathcal {S} _ {j} ^ {k, l, m, n} (t) < \sigma , \\ \overline {{\mathcal {G}}} _ {j} ^ {k, l, m, n} (t), & \text {o t h e r w i s e}, \end{array} \right. \tag {16} +$$ + +where $\gamma$ is the proportion of penalty term. + +It is worth noting that while both AGS and SAD improve the efficiency of weight sign flipping, they solve the problem in fundamentally different ways. AGS facilitates sign flipping for the whole weights by adaptively scaling the gradient, while SAD detects "silent weights" by tracking their flipping state and applies additional penalties. In Sec. 5.3, we show that both methods achieve significant performance improvement and they are complementary to each other. As an algorithm guideline, the pseudo-code of OvSW is provided in Algorithm 1. + +# 5 Experiment + +In this section, we conduct extensive image classification experiments for OvSW and compare it to state-of-the-art (SOTA) methods on CIFAR10 and ImageNet1K with various architectures. Then, we discuss the hyperparameter settings for OvSW, including $\lambda$ for AGS and $\sigma$ for SAD and convergence speed on CIFAR100. We also conduct ablation study to demonstrate the compatibility + +Table 1: Performance comparison with SOTAs on CIFAR10. We report the Top-1 Accuracy performance on ResNet18, ResNet20, and VGGsmall. W/A denotes the bit-width of weights/activations. + +
NetworksMethodsBit-width (W/A)Top-1 Acc.(%)
ResNet18Full-precision32/3294.8
IR-Net [40]1/191.5
RBNN [30]1/192.2
CMIM [45]1/192.2
SiMaN [29]1/192.5
ReCU [53]1/192.8
OvSW (Ours)1/193.2
ResNet20Full-precision32/3292.1
SLB [55]1/185.5
FDA-BNN [52]1/186.2
IR-Net [40]1/186.5
CMIM [45]1/187.3
SiMaN [29]1/187.4
ReCU [53]1/187.4
OvSW (Ours)1/187.7
VGGsmallFull-precision32/3294.1
DoReFa [60]1/190.2
RAD [6]1/190.5
RBNN [30]1/191.3
DSQ [11]1/191.7
Proxy-BNN [16]1/191.8
SLB [55]1/192.0
ReCU [53]1/192.2
FDA-BNN [52]1/192.5
SiMaN [29]1/192.5
OvSW (Ours)1/192.8
+ +and visualize the loss landscape for OvSW. Finally, we deploy OvSW to a real-world mobile device to exhibit its efficiency. To train OvSW, we use one NVIDIA RTX 3090 on the CIFAR10 and CIFAR100 and four on the ImageNet1K. All experiments are implemented on PyTorch [39]. + +# 5.1 Results on CIFAR10 + +We trained OvSW for CIFAR10 with 600 epochs, where the batch size is set to 256 and the initial learning rate to 0.1, decaying with CosineAnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 5e-4 and employ the same data augmentation in ReCU [53]. $\lambda$ and $\sigma$ are set to 0.04 and 9e-4 respectively. We compare OvSW with IR-Net [40], RBNN [30], CMIM [45], SiMaN [29], ReCU [53], SLB [55], FDA-BNN [52], DoReFa [60], RAD [6], DSQ [11], and Proxy-BNN [16]. As shown in Tab. 1, OvSW achieves the best performance among all methods. For ResNet18, OvSW obtains $93.2\%$ top-1 accuracy, which outperforms SiMaN and ReCU by $0.7\%$ and $0.4\%$ respectively, reducing the accuracy gap between BNNs and full-precision model to $1.6\%$ . In addition, it yields $87.7\%$ and $92.8\%$ top-1 accuracy on ResNet20 and VGGsmall, which succeeds ReCU and SiMaN by $0.3\%$ and $0.3\%$ respectively. + +Table 2: Performance comparison with SOTAs on ImageNet1K. We report the Top-1 and Top-5 Accuracy performance on ResNet18 and ResNet34. W/A denotes the bit-width of weights/activations. * means using the two-step training setting as ReActNet. + +
ModelMethodBit-width (W/A)Top-1 Acc.(%)Top-5 Acc.(%)
ResNet18Full-precision32/3269.689.2
XNOR [42]1/151.273.2
BiReal [34]1/156.479.5
IR-Net [40]1/158.180.0
RBNN [30]1/159.981.9
SiMaN [29]1/160.182.3
FDA-BNN [52]1/160.282.3
ReCU [53]1/161.082.6
OvSW (Ours)1/161.683.1
ReActNet [33]1/165.986.1
ReCU [53]1/166.486.5
OvSW* (Ours)1/166.686.7
ResNet34Full-precision32/3273.391.3
XNOR++ [3]1/157.179.9
BiReal [34]1/162.283.9
IR-Net [40]1/162.984.1
RBNN [30]1/163.184.4
SiMaN [29]1/163.984.8
CMIM [45]1/165.085.7
ReCU [53]1/165.185.8
OvSW (Ours)1/165.586.1
+ +# 5.2 Results on ImageNet1K + +On ImageNet1K, OvSW is trained from scratch. We train OvSW with 200 epochs, where the batch size is set to 512 and the initial learning rate is set to 0.1, decaying with CosineAnnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 1e-4 and the data augmentation is the same as ReCU [53]. $\lambda$ and $\sigma$ are set to 0.02 and 2e-5 respectively. We demonstrate the ImageNet1K performance of ResNet18/34 and compare OvSW with SOTA methods, including one-stage training methods XNOR [42], BiReal [34], IR-Net [40], RBNN [30], SiMaN [29], FDA-BNN [52], ReCU [53], CMIM [45], and two-stage training method [37] adopted by ReActNet [33]. As shown in Tab. 2, OvSW also achieves the best performance. For ResNet18, OvSW achieves $61.6\%$ top-1 and $83.1\%$ top-5 accuracy compared to ReCU's $61.0\%$ top-1 and $82.6\%$ top-5 accuracy, which demonstrates the efficiency of overcoming the "silent weights". For ResNet34, OvSW achieves $65.5\%$ top-1 accuracy, which is also better than ReCU. We further compare OvSW with the two-stage training method ReActNet. OvSW obtains $66.6\%$ top-1 accuracy, succeeding ReActNet and ReCU by $0.7\%$ and $0.2\%$ respectively. + +![](images/6895c2fb511bc40f4d87993ac9633d85fe403fdc0196c3f3675e68239cff5347.jpg) +(a) AGS vs Vanilla + +![](images/d6e1112df788cf6c6d3eba2d6b1d6d289c81b0a64489042b7a0ca4984cab13fb.jpg) +(b) AGS+SAD (OvSW) vs AGS + +![](images/809ac9f4f5de1f838857ad4a76ee314db15488ad2979afb8648143bb561e8706.jpg) +(c) Convergence +Fig. 3: Top-1 accuracy (mean±std) of binarized ResNet18 w.r.t. different values of $\lambda$ (a), $\sigma$ (b) and epoch (c) on CIFAR100. + +# 5.3 Ablation Analysis + +We investigate the effectiveness of hyper-parameters, including $\lambda$ and $\sigma$ , convergence speed, different components, and compatibility through ablation analysis. All the following results are based on binarized ResNet18 for CIFAR100. + +$\lambda$ for AGS and $\sigma$ for SAD. We first compare AGS with vanilla BNNs and analyze the $\lambda$ for AGS. In Fig. 3a, we present the results for nine different settings of $\lambda$ , which vary from 0.01 to 0.09. As seen, the appropriate $\lambda$ can significantly improve the performance of BNNs, and $\lambda = 0.04$ achieves the best performance. If $\lambda$ is too small, the sign of the weights can flip inefficiently; while $\lambda$ is too large, the sign of the weights can flip dramatically, introducing instability to the training. Based on this result, we further introduce SAD to AGS ( $\lambda = 0.04$ ) to analyze $\sigma$ . Figure 3b demonstrates that the performance of the model can be further improved by identifying "silent weights" and applying additional penalties to them via SAD ( $\sigma = 0.0009$ ). + +Convergence for OvSW. To verify that OvSW facilitates the flipping of weight signs and thus improves the efficiency of convergence, we fix the $\lambda$ and $\sigma$ to 0.04 and 0.0009 respectively, and record the top-1 accuracy over the training epoch from 60 to 500. As shown in Fig. 3c, OvSW effectively accelerates training convergence and achieves better performance. For example, OvSW achieves $66.37 \pm 0.35\%$ top-1 accuracy with only 60 training epochs, while vanilla BNNs only reaches $55.28 \pm 0.07\%$ and $65.23 \pm 0.21\%$ top-1 accuracy with 60 and 120 training epochs respectively. It indicates that OvSW can achieve significant performance gains in scenarios with limited training resources, such as training the network on edge devices. Meanwhile, although increasing epochs improve the final performance for both OvSW and vanilla BNNs, OvSW consistently outperforms the vanilla BNNs. + +Components. AGS and SAD promote sign flipping for overall weights and "silent weights" respectively. To prove that they are orthogonal to each other, we conduct components ablation study for different modules and show the results in Tab. 3 (left). As seen, AGS and SAD achieve $69.61 \pm 0.22\%$ and $69.45 \pm 0.18\%$ top-1 accuracy respectively and both of them succeed vanilla BNNs. It shows that they are effective in improving the efficiency of weight signs flipping. By + +Table 3: Left: Ablation study of different components in OvSW. Right: Applying OvSW as a plug-and-play module to other methods. + +
AGSSADmean±std (%)
XX65.23±0.21
✓X69.61±0.22
X✓69.45±0.18
✓✓69.77±0.15
+ +
Methodmean±std (%)
AdaBin70.56±0.19
AdaBin+OvSW72.56±0.11
RBNN67.15±0.23
RBNN+OvSW69.98±0.13
+ +combining them, OvSW (AGS+SAD) further achieves the best top-1 accuracy with $69.77 \pm 0.15\%$ . + +Compatibility. We show the good compatibility of OvSW by inserting it as a plug-and-play module into the current state-of-the-art methods, including AdaBin [48] and RBNN [30]. The former introduces an adaptive binary set to enhance the feature representation of BNNs, while the latter proposes a training-aware approximation function to reduce the gradient estimation error during training. Table 3 (right) demonstrates OvSW achieves $72.56 \pm 0.11\%$ top-1 accuracy on AdaBin and $69.98 \pm 0.13\%$ on RBNN, succeeding their original performance. This result indicates that OvSW has good compatibility and can effectively enhance the performance of existing methods. + +![](images/279443e141fae0a2bf64c41a77474b9af62136279d270fa9b5d0bd2c7ec75f5f.jpg) +(a) Full-precision + +![](images/6eebd3e9d4cd4094fd46d4f4305b5710935c510020cc0a81ba8fed42e1781c7c.jpg) +(b) XNOR-Net++ +Fig. 4: 3D visualization of the loss surfaces of ResNet18 on CIFAR100, which is used to enable comparisons of sharpness/flatness of different methods. + +![](images/2f3830e5b67c1b66adf1cfa49777dcaf2d99080c4b17e14e376d973f70a8db95.jpg) +(c) OvSW + +# 5.4 Loss Landscape Visualization + +BNNs restrict the weights and activations to discrete values, which naturally limits the representational capacity of the model and further result in disparate optimization landscapes compared to real-valued ones [32]. As illustrated in Fig. 4, we follow the method in [26] to plot the actual optimization landscape about our OvSW and compare it with the same architecture to real-valued and XNOR-Net++. As seen, our OvSW has a significantly smoother loss-landscape and minor loss elevation compared to XNOR-Net++, which confirms the effectiveness of OvSW in the BNN optimization. + +Table 4: Comparing OvSW (1-bit) with 32-bit and 16-bit backbones on M1 Pro. + +
NetworkW/ASize (MB)Memory SavingLatency (ms)Acceleration
ResNet1832/3246.76-27.67-
16/1623.382×16.651.66×
1/12.8116.64×7.973.47×
ResNet3432/3287.19-48.82-
16/1643.602×29.011.68×
1/14.1221.16×14.283.42×
+ +# 5.5 Deployment Efficiency + +We implement 1-bit models on the M1 Pro, which features 8 high-performance Firestorm cores and 2 efficient Icestorm cores in a hybrid design. The Firestorm cores can be clocked up to $3.2\mathrm{GHz}$ , while the Icestorm cores can reach 2.1 GHz. Our OvSW shows significant efficiency gains when deployed on real-world mobile devices, as evidenced by practical speed evaluations. To make our inference framework BOLT [9] compatible with OvSW, we leverage the ARM NEON SIMD instruction SSHL. We compare OvSW with 32-bit and 16-bit backbones. As shown in Tab. 4, OvSW inference speed is substantially faster with the highly efficient BOLT framework in a single thread. For instance, OvSW achieves an acceleration rate of about $3.47\times$ on ResNet18 compared to its 32-bit counterpart. For the ResNet34 backbone, OvSW achieves a $3.42\times$ acceleration rate with the BOLT framework on hardware, which is significant for computer vision applications on real-world edge devices. At the same time, OvSW can save memory by a factor of $16.64\times$ and $21.16\times$ , which demonstrates its potential for applications with limited memory resources. + +# 6 Conclusion + +BNN is a crucial method to compress deep learning models and reduce inference overhead. In this paper, systematically and theoretically, we prove the distribution of gradients is independent of latent weights, which is the root cause of inefficient updating and performance degradation of BNNs. To this end, Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are proposed to Overcome Silent Weights (OvSW) and achieve SOTA performance on BNNs. Specifically, AGS adaptively scales the gradient based on the distribution of weights, improving the efficiency of sign flipping for the overall weights. SAD can effectively measure the states of weight flipping, detect the "silent weights" and introduce additional penalty to them to facilitate their flipping. Extensive experiments demonstrate that OvSW can achieve notable performance gains over the SOTA methods on various datasets and networks. In addition, OvSW has better convergence efficiency and excellent compatibility, which can be combined with existing methods to further enhance the performance of BNNs. + +# References + +1. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013) 4, 5 +2. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: Int. Conf. Mach. Learn. pp. 1059-1071. PMLR (2021) 8 +3. Bulat, A., Tzimiropoulos, G.: Xnor-net++: Improved binary neural networks. arXiv preprint arXiv:1909.13863 (2019) 4, 5, 11 +4. Chen, J., Liu, L., Liu, Y., Zeng, X.: A learning framework for n-bit quantized neural networks toward fpgas. IEEE Transactions on Neural Networks and Learning Systems 32(3), 1067-1081 (2020) 2 +5. Courbariaux, M., Hubara, I., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks: Training deep neural networks with weights and activations constrained to + 1 or -1. arXiv preprint arXiv:1602.02830 (2016) 4 +6. Ding, R., Chin, T.W., Liu, Z., Marculescu, D.: Regularizing activation distribution for training binarized deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11408-11417 (2019) 10 +7. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13733-13742 (2021) 2 +8. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench101: Improving transferability and generalizability of cross-task neural architecture search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5251-5260 (2021) 2 +9. Feng, J.: Bolt. https://github.com/huawei-noah/bolt (2021) 14 +0. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 580-587 (2014) 1 +1. Gong, R., Liu, X., Jiang, S., Li, T., Hu, P., Lin, J., Yu, F., Yan, J.: Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In: Int. Conf. Comput. Vis. pp. 4852–4861 (2019) 2, 10 +2. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Int. Conf. Comput. Vis. pp. 2961-2969 (2017) 2 +3. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Int. Conf. Comput. Vis. pp. 1026-1034 (2015) +4. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans. Pattern Anal. Mach. Intell. 37(9), 1904-1916 (2015) 1 +5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 1, 2, 4 +6. He, X., Mo, Z., Cheng, K., Xu, W., Hu, Q., Wang, P., Liu, Q., Cheng, J.: Proxybnn: Learning binarized neural networks via proxy matrices. In: Eur. Conf. Comput. Vis. pp. 223-241. Springer (2020) 10 +7. Helwegen, K., Widdicombe, J., Geiger, L., Liu, Z., Cheng, K.T., Nusselder, R.: Latent weights do not exist: Rethinking binarized neural network optimization. Adv. Neural Inform. Process. Syst. 32 (2019) 2 +8. Hinton, G., Vinyals, O., Dean, J., et al.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 2(7) (2015) 2 + +19. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC). pp. 10-14. IEEE (2014) 2 +20. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2704-2713 (2018) 2 +21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 4 +23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst. 25 (2012) 1 +24. Lee, C., Kim, H., Park, E., Kim, J.J.: Insta-bnn: Binary neural network with instance-aware threshold. In: Int. Conf. Comput. Vis. pp. 17325-17334 (2023) 4 +25. Lee, J., Kim, D., Ham, B.: Network quantization with element-wise gradient scaling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6448-6457 (2021) 5 +26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. In: Adv. Neural Inform. Process. Syst. (2018) 13 +27. Li, S., Lin, M., Wang, Y., Fei, C., Shao, L., Ji, R.: Learning efficient gans for image translation via differentiable masks and co-attention distillation. IEEE Trans. Multimedia (2022) 2 +28. Lin, M., Ji, R., Wang, Y., Zhang, Y., Zhang, B., Tian, Y., Shao, L.: Hrank: Filter pruning using high-rank feature map. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1529-1538 (2020) 2 +29. Lin, M., Ji, R., Xu, Z., Zhang, B., Chao, F., Lin, C.W., Shao, L.: Siman: Sign-to-magnitude network binarization. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 10, 11 +30. Lin, M., Ji, R., Xu, Z., Zhang, B., Wang, Y., Wu, Y., Huang, F., Lin, C.W.: Rotated binary neural network. Adv. Neural Inform. Process. Syst. 33, 7474-7485 (2020) 4, 10, 11, 13 +31. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Adv. Neural Inform. Process. Syst. 30 (2017) 4 +32. Liu, Z., Shen, Z., Li, S., Helwegen, K., Huang, D., Cheng, K.T.: How do adam and training strategies help bnns optimization. In: Int. Conf. Mach. Learn. pp. 6936-6946. PMLR (2021) 13 +33. Liu, Z., Shen, Z., Savvides, M., Cheng, K.T.: Reactnet: Towards precise binary neural network with generalized activation functions. In: Eur. Conf. Comput. Vis. pp. 143-159. Springer (2020) 5, 11 +34. Liu, Z., Wu, B., Luo, W., Yang, X., Liu, W., Cheng, K.T.: Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In: Eur. Conf. Comput. Vis. pp. 722-737 (2018) 4, 5, 6, 11 +35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3431-3440 (2015) 2 +36. Luo, J.H., Wu, J.: Autopruner: An end-to-end trainable filter pruning method for efficient deep model inference. Pattern Recognition 107, 107461 (2020) 2 +37. Martinez, B., Yang, J., Bulat, A., Tzimiropoulos, G.: Training binary neural networks with real-to-binary convolutions. arXiv preprint arXiv:2003.11535 (2020) 11 + +38. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: Int. Conf. Mach. Learn. pp. 16318-16330. PMLR (2022) 9 +39. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché Buc, F., Fox, E., Garnett, R. (eds.) Adv. Neural Inform. Process. Syst. pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf 10 +40. Qin, H., Gong, R., Liu, X., Shen, M., Wei, Z., Yu, F., Song, J.: Forward and backward information retention for accurate binary neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2250-2259 (2020) 4, 5, 10, 11 +41. Rajbhandari, S., Rasley, J., Ruwase, O., He, Y.: Zero: Memory optimizations toward training trillion parameter models. In: SC20: International Conference for High Performance Computing, Networking, Storage and Analysis. pp. 1-16. IEEE (2020) +42. Rastegari, M., Ordonez, V., Redmon, J., Farhadi, A.: Xnor-net: Imagenet classification using binary convolutional neural networks. In: Eur. Conf. Comput. Vis. pp. 525-542. Springer (2016) 2, 4, 5, 11 +43. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2016) 1 +44. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015) 4 +45. Shang, Y., Xu, D., Zong, Z., Yan, Y.: Network binarization via contrastive learning. arXiv preprint arXiv:2207.02970 (2022) 10, 11 +46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 2, 4 +47. Su, X., Huang, T., Li, Y., You, S., Wang, F., Qian, C., Zhang, C., Xu, C.: Prioritized architecture sampling with monto-carlo tree search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10968-10977 (2021) 2 +48. Tu, Z., Chen, X., Ren, P., Wang, Y.: Adabin: Improving binary neural networks with adaptive binary sets. In: Eur. Conf. Comput. Vis. pp. 379-395. Springer (2022) 4, 13 +49. Wu, X.M., Zheng, D., Liu, Z., Zheng, W.S.: Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17055-17064 (2023) 5 +50. Xu, S., Li, Y., Ma, T., Lin, M., Dong, H., Zhang, B., Gao, P., Lu, J.: Resilient binary neural network. In: AAAI. vol. 37, pp. 10620-10628 (2023) +51. Xu, S., Li, Y., Wang, T., Ma, T., Zhang, B., Gao, P., Qiao, Y., Lü, J., Guo, G.: Recurrent bilinear optimization for binary neural networks. In: Eur. Conf. Comput. Vis. pp. 19-35. Springer (2022) +52. Xu, Y., Han, K., Xu, C., Tang, Y., Xu, C., Wang, Y.: Learning frequency domain approximation for binary neural networks. Adv. Neural Inform. Process. Syst. 34, 25553-25565 (2021) 10, 11 + +53. Xu, Z., Lin, M., Liu, J., Chen, J., Shao, L., Gao, Y., Tian, Y., Ji, R.: Recu: Reviving the dead weights in binary neural networks. In: Int. Conf. Comput. Vis. pp. 5198-5208 (2021) 3, 4, 5, 10, 11 +54. Yang, J., Shen, X., Xing, J., Tian, X., Li, H., Deng, B., Huang, J., Hua, X.s.: Quantization networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7308-7316 (2019) 2 +55. Yang, Z., Wang, Y., Han, K., Xu, C., Xu, C., Tao, D., Xu, C.: Searching for low-bit weights in quantized neural networks. Adv. Neural Inform. Process. Syst. 33, 4091-4102 (2020) 10 +56. You, Y., Gitman, I., Ginsburg, B.: Large batch training of convolutional networks. arXiv preprint arXiv:1708.03888 (2017) 8 +57. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Eur. Conf. Comput. Vis. pp. 365-382 (2018) 2 +58. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6848-6856 (2018) 2 +59. Zhao, J., Yang, L., Zhang, B., Guo, G., Doermann, D.S.: Uncertainty-aware binary neural networks. In: IJCAI. pp. 3441-3447 (2021) 4 +60. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016) 10 +61. Zhu, F., Gong, R., Yu, F., Liu, X., Wang, Y., Li, Z., Yang, X., Yan, J.: Towards unified int8 training for convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1969-1979 (2020) 2 \ No newline at end of file diff --git a/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/images.zip b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a74f6a7cbaf8bbe020c2df5d1330d5b252b4148 --- /dev/null +++ b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d81e3fecef7afdcb8fa58a23941d50f34a4810b153508e46d8b66168abf72cd4 +size 410024 diff --git a/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/layout.json b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..514493a7fefd61b0b7509b84632e77a51b28d96b --- /dev/null +++ b/2024/OvSW_ Overcoming Silent Weights for Accurate Binary Neural Networks/layout.json @@ -0,0 +1,11372 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "text", + "content": "Jingyang Xiang" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "text", + "content": ", Zuohui Chen" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "text", + "content": ", Siqi Li" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "text", + "content": ", Qing Wu" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "text", + "content": ", and Yong Liu" + }, + { + "bbox": [ + 137, + 167, + 477, + 181 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 234, + 190, + 378, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 190, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 234, + 190, + 378, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 234, + 190, + 378, + 201 + ], + "type": "text", + "content": " APRIL Lab, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 201, + 381, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 201, + 381, + 212 + ], + "spans": [ + { + "bbox": [ + 233, + 201, + 381, + 212 + ], + "type": "text", + "content": "2 Zhejiang University of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 212, + 429, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 212, + 429, + 222 + ], + "spans": [ + { + "bbox": [ + 184, + 212, + 429, + 222 + ], + "type": "text", + "content": "3 College of Computer Science, Hangzhou Dianzi University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 223, + 388, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 223, + 388, + 234 + ], + "spans": [ + { + "bbox": [ + 225, + 223, + 388, + 234 + ], + "type": "text", + "content": "4 Huzhou Institute, Zhejiang University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "content": "Abstract. Binary Neural Networks (BNNs) have been proven to be highly effective for deploying deep neural networks on mobile and embedded platforms. Most existing works focus on minimizing quantization errors, improving representation ability, or designing gradient approximations to alleviate gradient mismatch in BNNs, while leaving the weight sign flipping, a critical factor for achieving powerful BNNs, untouched. In this paper, we investigate the efficiency of weight sign updates in BNNs. We observe that, for vanilla BNNs, over " + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "content": " of the weights remain their signs unchanged during training, and these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. We refer to these weights as \"silent weights\", which slow down convergence and lead to a significant accuracy degradation. Theoretically, we reveal this is due to the independence of the BNNs gradient from the latent weight distribution. To address the issue, we propose Overcome Silent Weights (OvSW). OvSW first employs Adaptive Gradient Scaling (AGS) to establish a relationship between the gradient and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Additionally, we design Silence Awareness Decaying (SAD) to automatically identify \"silent weights\" by tracking weight flipping state, and apply an additional penalty to \"silent weights\" to facilitate their flipping. By efficiently updating weight signs, our method achieves faster convergence and state-of-the-art performance on CIFAR10 and ImageNet1K dataset with various architectures. For example, OvSW obtains " + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "inline_equation", + "content": "61.6\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "inline_equation", + "content": "65.5\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "content": " top-1 accuracy on the ImageNet1K using binarized ResNet18 and ResNet34 architecture respectively. Codes are available at https://github.com/JingyangXiang/OvSW." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "spans": [ + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "text", + "content": "Keywords: Binary Neural Networks " + }, + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "text", + "content": " Silent Weights " + }, + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "text", + "content": " Adaptive Gradient Scaling " + }, + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 562, + 453, + 585 + ], + "type": "text", + "content": " Silence Awareness Decaying" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have shown tremendous success in various computer vision tasks, including image classification [15, 23], object detection [10, 14," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 337, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 337, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 337, + 666 + ], + "type": "text", + "content": "* Corresponding author: yongliu@iipc.zju.edu.cn." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 120, + 242, + 198 + ], + "blocks": [ + { + "bbox": [ + 145, + 120, + 242, + 198 + ], + "lines": [ + { + "bbox": [ + 145, + 120, + 242, + 198 + ], + "spans": [ + { + "bbox": [ + 145, + 120, + 242, + 198 + ], + "type": "image", + "image_path": "aa0599de141f25a3e675a8e7cc2915ecd394dde387fde65361b145a01d95251f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 199, + 214, + 209 + ], + "lines": [ + { + "bbox": [ + 173, + 199, + 214, + 209 + ], + "spans": [ + { + "bbox": [ + 173, + 199, + 214, + 209 + ], + "type": "text", + "content": "(a) Vanilla" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 121, + 354, + 199 + ], + "blocks": [ + { + "bbox": [ + 258, + 121, + 354, + 199 + ], + "lines": [ + { + "bbox": [ + 258, + 121, + 354, + 199 + ], + "spans": [ + { + "bbox": [ + 258, + 121, + 354, + 199 + ], + "type": "image", + "image_path": "3619b96a085883ae65264adcd23e8d5335158e452f910b9157e61590abc48866.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 200, + 326, + 209 + ], + "lines": [ + { + "bbox": [ + 287, + 200, + 326, + 209 + ], + "spans": [ + { + "bbox": [ + 287, + 200, + 326, + 209 + ], + "type": "text", + "content": "(b) OvSW" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "type": "text", + "content": "Fig. 1: (a) and (b): Histogram of the initialized weight distribution (blue) and the weights that never update signs throughout training (orange). " + }, + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "type": "inline_equation", + "content": "54.07\\%" + }, + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "type": "inline_equation", + "content": "2.03\\%" + }, + { + "bbox": [ + 130, + 218, + 482, + 275 + ], + "type": "text", + "content": " represent the ratio of the corresponding orange area to the blue. (c) Top-1 Accuracy (solid lines) and Flip Ratio (dashed lines) are for Vanilla (green) and OvSW (red). (a), (b) and Flip Ratio in (c) is for layer4(conv2.weight." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 121, + 478, + 199 + ], + "blocks": [ + { + "bbox": [ + 366, + 121, + 478, + 199 + ], + "lines": [ + { + "bbox": [ + 366, + 121, + 478, + 199 + ], + "spans": [ + { + "bbox": [ + 366, + 121, + 478, + 199 + ], + "type": "image", + "image_path": "835fe892beb883b41ac6b6b6f4cf79680a2fe929679a91a43a7219685a569510.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 200, + 470, + 210 + ], + "lines": [ + { + "bbox": [ + 372, + 200, + 470, + 210 + ], + "spans": [ + { + "bbox": [ + 372, + 200, + 470, + 210 + ], + "type": "text", + "content": "(c) Top-1 Acc & Flip Ratio" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 348 + ], + "type": "text", + "content": "43], and semantic segmentation [12, 35]. However, the remarkable performance is always attributed to deeper and wider architectures [15, 46], which comes with expensive memory and computational overhead and makes it challenging to deploy DNNs on resource-constrained edge platforms." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 350, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 470 + ], + "type": "text", + "content": "The community has been delving into model compression, which aims to reduce inference overhead for DNNs while preserving their performance. Typical techniques include, but are not limited to, efficient architecture design [7, 58], neural architecture search [8, 47], network pruning [28, 36], knowledge distillation [18, 27], and network quantization [4, 11, 61]. Among them, network quantization is widely suggested as a promising solution. It effectively enhances memory efficiency and execution speed on embedding platforms by reducing the weight and activation bits and replacing expensive floating-point arithmetic with relatively inexpensive fixed-point arithmetic. Extensive studies [11, 20, 54, 57] have demonstrated its effectiveness and yielded light and efficient DNNs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": "In this paper, we focus on studying binary neural networks (BNNs), which are an aggressive form of quantized neural networks (QNNs). BNNs binarize both weights and activations to discrete values " + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\{+1, -1\\}" + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": ", which can be represented by 1-bit and reduce memory usage by " + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": ". On the other hand, by employing efficient bitwise operations instead of floating-point ones, BNNs also significantly reduce computation complexity, providing " + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "inline_equation", + "content": "58 \\times" + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": " speedup as reported by XNOR-Net [42]. These characteristics make them well-suited for deployment on low-power embedding platforms, including FPGA, ASICs, and IoT devices [19]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "content": "However, the application of BNNs is still limited due to constraining both weights and activations to 1-bit significantly reducing accuracy compared to full-precision models. To address this problem, many methods have been developed to reduce quantization error and enhance the representation capability, thereby closing the performance gap between BNNs and their real-valued counterparts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Although these methods have effectively narrowed the performance gap, they still overlook the essence of BNN optimization is to update the signs of latent weights [17], instead of updating their values. In other words, if the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "signs of the latent weights don't change, the BNNs are hardly ever updated. To demonstrate the weight sign flipping of vanilla BNNs, we train binarized ResNet18 on CIFAR100 with 120 epochs and plot the flipping information for layer4 conv2 weight. As shown in Fig. 1a, more than " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " of the weights never change their signs throughout training process, which slows down convergence and leads to a significant accuracy degradation for BNNs. We denote the part of weights that hardly change their signs during training as \"silent weights\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 391 + ], + "type": "text", + "content": "It's worth noting that, as a study mostly related to ours, ReCU [53] proposes a rectified clamp unit to revive the \"dead weights\" for updating, which refers to a group of weights that are distributed at the tails of the weight distribution and barely update their signs during training. They claim that the magnitudes of latent weights do not contribute to the forward propagation. If two weights have the same sign, they have the same effect on the forward propagation. Intuitively, the signs of the weights around the zero are easily changed, while weights distributed at the tails of the weight distribution tend to have difficulties in changing their signs. Although ReCU facilitates the updating for \"dead weights\", it ignores the fact, the gradient distribution across layers and weights have great distinction. Weights with larger magnitudes are intuitively more difficult to flip their signs, but small weights may also suffer from silence once their gradients are relatively small compared to their magnitudes. As shown in Fig. 1a, for a vanilla BNN, these weights are not only distributed at the tails of the weight distribution but also universally present in the vicinity of zero. As a result, it is still sub-optimal and remains to be improved." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 607 + ], + "type": "text", + "content": "In this paper, we reveal that the root cause of the large number of \"silent weights\" is attributed to the independence of the BNN gradients from the latent weight distribution via a systematical and theoretical analysis. To this end, two simple yet effective techniques, including Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are introduced to Overcome Silent Weights (OvSW). Specifically, AGS adaptively scales the gradients by establishing a relationship between the gradients and the latent weight distribution, thereby improving the overall efficiency of weight sign updates. Meanwhile, SAD automatically identifies \"silent weights\" by tracking weight sign flipping state and applies an additional penalty to them, further promoting the updating efficiency of their signs. Benefiting from AGS and SAD, OvSW enables efficient flipping of weight signs, significantly accelerates the convergence and promotes the performance for BNNs as shown in Fig. 1b and Fig. 1c. In comparison to the expensive computational cost of forward and backward propagation, OvSW introduces negligible extra overhead to the training process, maintaining its efficiency. Furthermore, since OvSW aims to facilitate the weight sign flipping for BNNs, which is orthogonal to the previous work, it has good compatibility and can be used as plug-and-play modules to enhance other BNNs' performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 608, + 398, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 608, + 398, + 620 + ], + "spans": [ + { + "bbox": [ + 146, + 608, + 398, + 620 + ], + "type": "text", + "content": "The contributions of our work are highlighted as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "content": "- We are the first to find that a large number of weights in BNNs fail to update their signs throughout the training. These \"silent weights\" are not only distributed at the tails of the weight distribution but also universally" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 140 + ], + "type": "text", + "content": "present in the vicinity of zero. Theoretically, we reveal this is attributed to the independence of the gradients from the distribution of the latent weights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 209 + ], + "type": "text", + "content": "- We propose to overcome \"silent weights\" with two novel techniques: (1) an adaptive gradient scaling method to scale the gradients according to the distribution of the latent weights, which improves overall efficiency in updating signs for weights; (2) a silence awareness decaying strategy to identify \"silent weights\" by tracking weight sign flipping state and introduce an additional penalty to them, further facilitating the flipping of their signs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "content": "- Extensive experiments for various BNNs including ResNet18/20 [15] and VGGsmall [46] on CIFAR10 [22] and ResNet18/34 [15] on ImageNet1K [44] demonstrate the effectiveness of our method. For example, OvSW achieves " + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "inline_equation", + "content": "61.6\\%" + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "inline_equation", + "content": "65.5\\%" + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "content": " top-1 accuracy on the ImageNet for binarized ResNet18 and ResNet34, improving ReCU [53] by " + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "content": " respectively. Besides, for ResNet18, OvSW achieves " + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "inline_equation", + "content": "2.00\\%" + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "inline_equation", + "content": "2.83\\%" + }, + { + "bbox": [ + 138, + 210, + 481, + 317 + ], + "type": "text", + "content": " improvement on CIFAR100 [22] when combined with AdaBin (enhancing representation ability) and RBNN (training-aware gradient approximation) respectively, demonstrating its good compatibility with the other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 332, + 237, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 237, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 237, + 345 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": "Great efforts have been put into reducing the performance gap between BNNs and their real-valued counterparts. The pioneering BNN work dates back to Courbariaux et al. [5], which binarizes weights and activations to " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "+1" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": " by sign function. However, this aggressive approach limits the representation ability of BNNs to the binary space, resulting in a significant accuracy degradation. To reduce the quantization error and improve their accuracy, XNOR-Net [42] introduces a scaling factor obtained through the " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": "-norm of weights or activations. Furthermore, XNOR-Net++ [3] merges two scale factors from the weights and activations into a trainable parameter and optimizes them via backpropagation. ABC-Net [31] employs a linear combination of multiple binary weight sets to approximate the real-valued weights and alleviate the information loss. Bi-Real Net [34] connects the real-valued activations to the consecutive block via an identity shortcut, which significantly enhances network representation ability while incurring negligible computational overhead. AdaBin [48] introduces adaptive binary sets to fit different distributions, enhancing binarized representation ability. UaBNN [59] introduces an uncertainty-aware BNN to reduce these weights' uncertainties. INSTA-BNN [24] controls the quantization threshold in an input instance-aware manner, taking higher-order statistics into account." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Apart from enhancing the representation ability of BNNs, gradient estimation is also one of the critical research directions, since gradients in the sign function are almost zero everywhere. Straight through estimator (STE) [1] is the most widely used function to enable the gradient to backpropagate. However, the gradient error is huge for STE and will accumulate during backpropagation, leading to instability in training and severe accuracy degradation. To alleviate this, Bi-Real Net [34] utilizes a piecewise polynomial function as the approximation function. IR-Net [40] proposes an error decay estimator and RBNN [30]" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "employs a training-aware approximation function to replace the sign function. EWGS [25] takes discretization error between input and output into account and introduces element-wise gradient scaling to adaptively scale up or down each gradient element. ReSTE [49] revises the original STE and balance the estimating error and the gradient stability well. All of these methods effectively reduce the gradient error and achieve consistent improvement in both training stability and accuracy compared to the vanilla STE." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 225, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 225, + 230 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 225, + 230 + ], + "type": "text", + "content": "3 Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": "We briefly review the optimization process of BNNs in this section. Given a DNN, we denote " + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j \\in \\mathbb{R}^{C_{\\mathrm{out}}^j \\times C_{\\mathrm{in}}^j \\times K_{\\mathrm{h}}^j \\times K_{\\mathrm{w}}^j}" + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": " as the real-valued weight in the " + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{out}}^j" + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{in}}^j" + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{h}}^j" + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{w}}^j" + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": " are the number of output channels, input channels, kernel height, and kernel width, respectively. Let the real-valued activation be " + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_j" + }, + { + "bbox": [ + 130, + 239, + 479, + 302 + ], + "type": "text", + "content": ", then the convolution process can be formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 269, + 311, + 480, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 311, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 269, + 311, + 480, + 323 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {j + 1} = \\mathcal {W} _ {j} * \\mathcal {A} _ {j}, \\tag {1}", + "image_path": "8a26f9106caf38fe1114d125e474285a98f45dc26fa84892f984a30d999ccb30.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 375, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 375, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 375, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 331, + 375, + 342 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 130, + 331, + 375, + 342 + ], + "type": "text", + "content": " represents the standard convolution operation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "text", + "content": "BNNs aim to binarize weights " + }, + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "text", + "content": " and activations " + }, + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_j" + }, + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "text", + "content": " to discrete values " + }, + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "inline_equation", + "content": "(\\{+1, -1\\})" + }, + { + "bbox": [ + 130, + 342, + 498, + 365 + ], + "type": "text", + "content": " through sign function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 373, + 480, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 373, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 238, + 373, + 480, + 398 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\operatorname {s i g n} (x) = \\left\\{ \\begin{array}{l} + 1, \\text {i f} x \\geq 0, \\\\ - 1, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "1804b30b01fef78c4d916d6de110a01b7bb83e68149af6d8ab1fb89940fc52b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "text", + "content": "To reduce the quantization error in BNNs, XNOR-Net [42] introduces two scale factors to approximate the binarized weights " + }, + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{W}}_j^b" + }, + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "text", + "content": " and activation " + }, + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}_j^b" + }, + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "text", + "content": ". Furthermore, XNOR-Net++ [3] proposes fusing the activation and weight scaling factors into one and optimizing it via backpropagation. This approach significantly outperforms XNOR-Net within the same computational budget and has been widely adopted by recent works [33, 34, 40, 53]. Following them, we denote the scaling factor as " + }, + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 130, + 404, + 479, + 498 + ], + "type": "text", + "content": ". Then the binary convolution operation can be formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 251, + 498, + 480, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 498, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 251, + 498, + 480, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {j + 1} = \\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j}\\right) \\odot \\alpha_ {j}, \\tag {3}", + "image_path": "c028a1a9b351b7f1ab2c40a24b7d244d22b32351ed54db66e0208047cc85ac93.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "text", + "content": " is the efficient XNOR and Bitcount operation and " + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "text", + "content": " is theadamard product. In the implementation of BNNs, " + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{j + 1}" + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "text", + "content": " will be processed through several layers, e.g., Batch Normalization (BN) layer, non-linear activation layer, and max-pooling layer. In this section, we omit these operations for simplicity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "text", + "content": "To train a BNN, the forward propagation includes Eq. (2) and Eq. (3), where their real-value counterparts " + }, + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_j" + }, + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "text", + "content": " are used for calculating gradients and updating during the backpropagation. However, the sign is not differentiable, thereby gradient estimation is important in BNNs. Following the previous studies, we use straight-through estimator (STE) [1] to approximate the gradient of the loss w.r.t. the weight " + }, + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "inline_equation", + "content": "w\\in \\mathcal{W}" + }, + { + "bbox": [ + 130, + 567, + 479, + 639 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 257, + 645, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 645, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 257, + 645, + 480, + 668 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial w} = \\frac {\\partial \\mathcal {L}}{\\partial \\hat {w}} \\cdot \\frac {\\partial \\hat {w}}{\\partial w} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\hat {w}}, \\tag {4}", + "image_path": "c721eeec5b328cb486d296c911662bad0e9d82499155786da34475cb9c27a8c4.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 116, + 449, + 199 + ], + "blocks": [ + { + "bbox": [ + 166, + 116, + 449, + 199 + ], + "lines": [ + { + "bbox": [ + 166, + 116, + 449, + 199 + ], + "spans": [ + { + "bbox": [ + 166, + 116, + 449, + 199 + ], + "type": "image", + "image_path": "59c277c91b6d9d4e2a2ff49e77ebfbbcb06174c22dbbaaa2575d510eb86a770a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 206, + 482, + 231 + ], + "lines": [ + { + "bbox": [ + 131, + 206, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 482, + 231 + ], + "type": "text", + "content": "Fig. 2: Forward and backward computation graph for binary convolutional operation with quantization aware training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 251, + 479, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 479, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 251, + 479, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 251, + 479, + 276 + ], + "type": "text", + "content": " denotes the loss function. For the gradient w.r.t. the activation " + }, + { + "bbox": [ + 131, + 251, + 479, + 276 + ], + "type": "inline_equation", + "content": "a\\in \\mathcal{A}" + }, + { + "bbox": [ + 131, + 251, + 479, + 276 + ], + "type": "text", + "content": ", we adopt the piece-wise polynomial gradient estimation function [34] as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 283, + 481, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 283, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 238, + 283, + 481, + 308 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial a} = \\frac {\\partial \\mathcal {L}}{\\partial \\hat {a}} \\cdot \\frac {\\partial \\hat {a}}{\\partial a} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\hat {a}} \\cdot \\frac {\\partial F (a)}{\\partial a}, \\tag {5}", + "image_path": "002c4555180e2ebb3e8235341a5985f38b8c0f80fb29a23b13d7b0203f3e12a4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 311, + 160, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 160, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 160, + 321 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 319, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 319, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 228, + 319, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\frac {\\partial F (a)}{\\partial a} = \\left\\{ \\begin{array}{l l} 2 + 2 a, & \\text {i f} - 1 \\leq a < 0, \\\\ 2 - 2 a, & \\text {i f} 0 \\leq a < 1, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}", + "image_path": "c2ee7754462a7776693c6925f8003025fdefa635f14c9cb7e025b6aef98f4595.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 371, + 233, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 233, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 233, + 385 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 394, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 482, + 456 + ], + "type": "text", + "content": "In this section, we show that the distribution of gradients and weights for BNNs is independent by systematical and theoretical analysis. Then, we propose Adaptive Gradient Scaling (AGS) to scale the gradient and introduce Silence Awareness Decaying (SAD) to detect \"silent weights\", moving them towards zero. Both of them can enhance the efficiency of weight sign flipping." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 471, + 462, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 471, + 462, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 471, + 462, + 483 + ], + "type": "text", + "content": "4.1 The Independence of the Gradient and Weight Distribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 490, + 479, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 490, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 490, + 479, + 525 + ], + "type": "text", + "content": "Figure 2 demonstrates forward and backward computation graph for binary convolutional (Bin-Conv) operation. As seen, Bin-Conv is often followed by a BN layer. We can conclude that once two BNNs " + }, + { + "bbox": [ + 131, + 490, + 479, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 131, + 490, + 479, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 490, + 479, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{N}'" + }, + { + "bbox": [ + 131, + 490, + 479, + 525 + ], + "type": "text", + "content": " satisfy:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 228, + 531, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 531, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 228, + 531, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {W}} _ {j} = \\hat {\\mathcal {W}} _ {j} ^ {\\prime}, \\alpha_ {j} = \\alpha_ {j} ^ {\\prime}, \\mathrm {B N} _ {j} = \\mathrm {B N} _ {j} ^ {\\prime}, \\forall j, \\tag {7}", + "image_path": "8abceccedee2178c8f82ebdc8c648a542c1935314ce91307ab34b0854a68453b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "text", + "content": "for the same input " + }, + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}_j" + }, + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{j + 1}" + }, + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{j + 1}^{\\prime}" + }, + { + "bbox": [ + 131, + 555, + 453, + 571 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 146, + 576, + 481, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 576, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 146, + 576, + 481, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {j + 1} = \\mathrm {B N} \\left(\\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j}\\right) \\odot \\alpha_ {j}\\right) = \\mathrm {B N} ^ {\\prime} \\left(\\left(\\hat {\\mathcal {A}} _ {j} * \\hat {\\mathcal {W}} _ {j} ^ {\\prime}\\right) \\odot \\alpha_ {j} ^ {\\prime}\\right) = \\mathcal {A} _ {j + 1} ^ {\\prime}. \\tag {8}", + "image_path": "c1d3354e63c89b2e6031a05835a914a4f641e0a2a25a9437d97ac44b78ef6fda.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "text", + "content": "Through mathematical induction, we can know that the " + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{N}'" + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "text", + "content": " will have the same output and loss, i.e. " + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}'" + }, + { + "bbox": [ + 131, + 605, + 482, + 630 + ], + "type": "text", + "content": ". Therefore, the backpropagation for them is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 173, + 636, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 636, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 173, + 636, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {W} _ {j}} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {A} _ {j + 1}} \\frac {\\partial \\mathcal {A} _ {j + 1}}{\\partial \\hat {\\mathcal {W}} _ {j}} \\frac {\\partial \\hat {\\mathcal {W}} _ {j}}{\\partial \\mathcal {W} _ {j}} = \\frac {\\partial \\mathcal {L} ^ {\\prime}}{\\partial \\mathcal {A} _ {j + 1} ^ {\\prime}} \\frac {\\partial \\mathcal {A} _ {j + 1} ^ {\\prime}}{\\partial \\hat {\\mathcal {W}} _ {j} ^ {\\prime}} \\frac {\\hat {\\mathcal {W}} _ {j} ^ {\\prime}}{\\partial \\mathcal {W} _ {j} ^ {\\prime}} = \\frac {\\partial \\mathcal {L} ^ {\\prime}}{\\partial \\mathcal {W} _ {j} ^ {\\prime}}. \\tag {9}", + "image_path": "03142fec8856c49f2fa376b2bfba0ee2210693333a97d56ac0e9ae9abf98a176.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Without loss of generality, we can obtain that once two networks satisfy the conditions in Eq. (7), the gradients obtained from the backpropagation are the same (a more generalized scenario are extrapolated in ??). The gradient has nothing to do with the magnitude of weight and weight with larger magnitude does not necessarily corresponding to larger gradient. Therefore, although intuitively, weight with large magnitude tends to remain its sign unchanged, it is also potential for weight with small magnitude to suffer from silence if its gradient is relatively small compared to its magnitude. Furthermore, if we assume the weights further satisfy " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j' = \\gamma \\mathcal{W}_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " and networks are optimized via vanilla SGD, the optimization of " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "-step can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 246, + 480, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 246, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 144, + 246, + 480, + 302 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\mathcal {W} _ {j} ^ {\\prime} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} ^ {\\prime} (t)}{\\partial \\mathcal {W} _ {j} ^ {\\prime} (t)} = \\gamma \\mathcal {W} _ {j} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} \\tag {10} \\\\ = (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t) - \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} = (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t + 1), \\\\ \\end{array}", + "image_path": "a2def31cc3c2f2541fbfa0491eadc1558bfcf18a0da7a710d1b4b54563a66b52.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\beta(t)" + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "content": " is the learning rate at " + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "content": "-step. Equation (10) can be viewed as a variant of exponential moving averages and the effect of " + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j(t)" + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j'(t + 1)" + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "content": " will gradually increase as " + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "content": " increases. Once " + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 311, + 482, + 348 + ], + "type": "text", + "content": " is large enough:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 356, + 480, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 356, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 148, + 356, + 480, + 376 + ], + "type": "interline_equation", + "content": "\\lim _ {\\gamma \\rightarrow \\infty} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\lim _ {\\gamma \\rightarrow \\infty} [ (\\gamma - 1) \\mathcal {W} _ {j} (t) + \\mathcal {W} _ {j} (t + 1) ] = (\\gamma - 1) \\mathcal {W} _ {j} (t). \\tag {11}", + "image_path": "e4f4f1f7536c63c13e33115da7695aa8c2600d5ac7e8241e878baeadcd44dddf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 386, + 355, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 355, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 355, + 399 + ], + "type": "text", + "content": "The signs of " + }, + { + "bbox": [ + 131, + 386, + 355, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j^{\\prime}(t + 1)" + }, + { + "bbox": [ + 131, + 386, + 355, + 399 + ], + "type": "text", + "content": " will be the same as " + }, + { + "bbox": [ + 131, + 386, + 355, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j(t)" + }, + { + "bbox": [ + 131, + 386, + 355, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "text", + "content": "However, BNNs binarize " + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j'" + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\{+1, -1\\}" + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "text", + "content": " at each training step. Once the signs of " + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j'" + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "text", + "content": " do not change, the BNNs hardly ever update. It not only slows down convergence, but also leads to significant accuracy degradation. Please see ?? for a detailed ablation analysis for " + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 401, + 482, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 470, + 294, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 294, + 483 + ], + "type": "text", + "content": "4.2 Adaptive Gradient Scaling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "text", + "content": "To improve the efficiency of the update, an intuitive approach is to scale " + }, + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "inline_equation", + "content": "\\beta(t)" + }, + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "inline_equation", + "content": "\\gamma\\beta(t)" + }, + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "text", + "content": ". In this case, " + }, + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j'(t + 1)" + }, + { + "bbox": [ + 130, + 491, + 481, + 518 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 194, + 528, + 480, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 528, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 194, + 528, + 480, + 582 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {W} _ {j} ^ {\\prime} (t + 1) = \\mathcal {W} _ {j} ^ {\\prime} (t) - \\gamma \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} ^ {\\prime} (t)} \\tag {12} \\\\ = \\gamma \\mathcal {W} _ {j} (t) - \\gamma \\beta (t) \\frac {\\partial \\mathcal {L} (t)}{\\partial \\mathcal {W} _ {j} (t)} = \\gamma \\mathcal {W} _ {j} (t + 1). \\\\ \\end{array}", + "image_path": "cc218585a2fac2c0f5e7de4a1bfe383df37fb7819b72f590ab7737e07d189d30.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 590, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 639 + ], + "type": "text", + "content": "However, selecting the appropriate scaling factor is tricky. Meanwhile, an inappropriate scaling factor may make weights with small magnitudes correspond to large gradients, leading them to oscillate frequently in " + }, + { + "bbox": [ + 130, + 590, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\{+1, -1\\}" + }, + { + "bbox": [ + 130, + 590, + 482, + 639 + ], + "type": "text", + "content": " and introducing instability to BNNs training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "type": "text", + "content": "To overcome this issue, we introduce \"Adaptive Gradient Scaling\" (AGS). Let " + }, + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j\\in \\mathbb{R}^{C_j^j\\times C_j^j\\times K_h^j\\times K_w^j}" + }, + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "type": "text", + "content": " denote the gradient with respect to " + }, + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 639, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{W}_j}" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " denote the Forbenius norm, i.e., the " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "-th filter norm in " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 148, + 481, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 148, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 219, + 148, + 481, + 174 + ], + "type": "interline_equation", + "content": "\\left\\| \\mathcal {W} _ {j} ^ {k} \\right\\| _ {F} = \\sqrt {\\sum_ {l = 1} ^ {C _ {\\mathrm {i n}} ^ {j}} \\sum_ {m = 1} ^ {K _ {\\mathrm {h}} ^ {j}} \\sum_ {n = 1} ^ {K _ {\\mathrm {w}} ^ {j}} \\mathcal {W} _ {j} ^ {k , l , m , n}}. \\tag {13}", + "image_path": "ac60409cc847b7a54dc53f3179a6fc8a016afba655a6c6787373cccaa30632ab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": "The AGS algorithm is motivated by the observation that the ratio of the norm of " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j^k" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j^k\\left(\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}\\right)" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": " provides a simple measure of how much a single gradient descent step will change the original weight " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": ". For instance, if we train BNNs via vanilla SGD without momentum, then " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\frac{\\|\\triangle\\mathcal{W}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F} = \\beta \\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": ", where the parameter update for the " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j^k" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\triangle \\mathcal{W}_j^k = -\\beta \\mathcal{G}_j^k" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": ". We can conclude that the small value of " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": " during training is the root cause why a large number of weights fail to flip signs and adaptively scaling " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": " plays a crucial role in promoting the update efficiency for BNNs. Specifically, in AGS algorithm, " + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j^{k,l,m,n}" + }, + { + "bbox": [ + 130, + 182, + 482, + 324 + ], + "type": "text", + "content": " is scaled as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 334, + 480, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 334, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 209, + 334, + 480, + 372 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} = \\left\\{ \\begin{array}{l l} \\lambda \\frac {\\| \\mathcal {W} _ {j} ^ {k} \\| _ {F}}{\\| \\mathcal {G} _ {j} ^ {k} \\| _ {F}} \\mathcal {G} _ {j} ^ {k, l, m, n} & \\text {i f} \\frac {\\| \\mathcal {G} _ {j} ^ {k} \\| _ {F}}{\\| \\mathcal {W} _ {j} ^ {k} \\| _ {F}} < \\lambda , \\\\ \\mathcal {G} _ {j} ^ {k, l, m, n} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {14}", + "image_path": "a616bdc1158464ad44e75366598374758d2c2fa08ee0059280a43bfd2c8a9941.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "text", + "content": " is a scalar scaling threshold to limits the lower bound of " + }, + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}" + }, + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "text", + "content": ". Ablation analysis for " + }, + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 381, + 479, + 411 + ], + "type": "text", + "content": " can be found in Sec. 5.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 412, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 552 + ], + "type": "text", + "content": "AGS and \"Adaptive Gradient Clipping\" (AGC) [2] are closely related but fundamentally different, as the former restricts the lower bound of " + }, + { + "bbox": [ + 130, + 412, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\frac{\\|\\mathcal{G}_j^k\\|_F}{\\|\\mathcal{W}_j^k\\|_F}" + }, + { + "bbox": [ + 130, + 412, + 482, + 552 + ], + "type": "text", + "content": " to enhance the flipping efficiency of BNN weight signs, while the latter restricts upper bound and is designed to improve training stability. AGS also can be viewed as a adaptive variant to Layer-wise Adaptive Rate Scaling (LARS) [56], which sets the norm of update parameter to a fixed ratio of the parameter norm, and completely ignores the gradient magnitude to real-valued networks. Although LARS is also able to improve the accuracy for BNNs, we find that LARS and AGS is much different in the update of gradient momentum and doing so degrades performance compared to AGS. More details for comparing AGS with LARS and ablation analysis for LARS can be found in ??" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 570, + 302, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 302, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 302, + 582 + ], + "type": "text", + "content": "4.3 Silence Awareness Decaying" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 590, + 479, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 479, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 479, + 638 + ], + "type": "text", + "content": "We propose another approach (Silence Awareness Decaying, SAD) orthogonal to AGS to detect and prevent \"silent weights\". Specifically, we track the flipping state of " + }, + { + "bbox": [ + 130, + 590, + 479, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 130, + 590, + 479, + 638 + ], + "type": "text", + "content": " over time using an exponential moving average (EMA) strategy, which is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 645, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 645, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 145, + 645, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {j} (t) = m \\cdot \\mathcal {S} _ {j} (t - 1) + (1 - m) \\cdot \\frac {\\left| \\operatorname {s i g n} \\left(\\mathcal {W} _ {j} (t)\\right) - \\operatorname {s i g n} \\left(\\mathcal {W} _ {j} (t - 1)\\right) \\right| _ {a b s}}{2}, \\tag {15}", + "image_path": "02d7adf6126c2a93796549d82545f1e08e1fafcc5530cd243b85a2dcf1367dc1.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 129, + 481, + 311 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 340, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 340, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 340, + 128 + ], + "type": "text", + "content": "Algorithm 1 Overview of the OvSW method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": "Input: A minibatch of inputs and their labels, real-valued weights " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(t)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ", scaling factor " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\alpha(t)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " for AGS, " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " for SAD, " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "(S(t), m, \\sigma)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " for flipping state detection. \nOutput: Updated " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(t+1)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\alpha(t+1)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "S(t+1)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ". \n1: while Forward propagation do \n2: " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}_j \\gets \\mathrm{sign}(\\mathcal{A}_j)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ". \n3: " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{W}}_j \\gets \\mathrm{sign}(\\mathcal{W}_j)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ". \n4: Computing features via Eq. (2) and Eq. (3). \n5: Computing loss " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ". \n6: end while \n7: while Backward propagation do \n8: Computing " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{W}_j}" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\mathcal{A}_j}" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " via Eq. (4) and Eq. (5)&Eq. (6) \n9: Scaling " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{G}}_j" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " via Eq. (14). \n10: Adding penalties to " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{G}}_j" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " via Eq. (16). \n11: Update " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j(t+1)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\alpha_j(t+1)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " via SGD optimizer. \n12: Update " + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_j(t+1)" + }, + { + "bbox": [ + 132, + 129, + 481, + 311 + ], + "type": "text", + "content": " via Eq. (15). \n13: end while" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "text", + "content": " is the momentum and " + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "text", + "content": " is the auxiliary variable. Nagel et al. [38] employ this technique to identify and dampen oscillations prematurely while we employ it to identify \"silent weights\" and dynamically introduce additional weight penalties to move them towards zero. In our algorithm, we think if " + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "inline_equation", + "content": "S_{j}^{k,l,m,n}" + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "text", + "content": " is less than a pre-defined threshold " + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "text", + "content": ", its corresponding weight " + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{j}^{k,l,m,n}" + }, + { + "bbox": [ + 130, + 334, + 482, + 422 + ], + "type": "text", + "content": " is considered as a \"silent weight\" and will be applied with an additional penalty. The silence awareness decaying process is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 427, + 481, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 427, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 170, + 427, + 481, + 462 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) = \\left\\{ \\begin{array}{l l} \\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t) + \\gamma \\mathcal {W} _ {j} ^ {k, l, m, n} (t), & \\text {i f} \\mathcal {S} _ {j} ^ {k, l, m, n} (t) < \\sigma , \\\\ \\overline {{\\mathcal {G}}} _ {j} ^ {k, l, m, n} (t), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {16}", + "image_path": "78fb7396aa42af5f843f9ffb2d4d66a8960d4917d55c594503799ec6532f970e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 468, + 320, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 320, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 320, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 468, + 320, + 480 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 468, + 320, + 480 + ], + "type": "text", + "content": " is the proportion of penalty term." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 564 + ], + "type": "text", + "content": "It is worth noting that while both AGS and SAD improve the efficiency of weight sign flipping, they solve the problem in fundamentally different ways. AGS facilitates sign flipping for the whole weights by adaptively scaling the gradient, while SAD detects \"silent weights\" by tracking their flipping state and applies additional penalties. In Sec. 5.3, we show that both methods achieve significant performance improvement and they are complementary to each other. As an algorithm guideline, the pseudo-code of OvSW is provided in Algorithm 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 224, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 224, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 224, + 594 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this section, we conduct extensive image classification experiments for OvSW and compare it to state-of-the-art (SOTA) methods on CIFAR10 and ImageNet1K with various architectures. Then, we discuss the hyperparameter settings for OvSW, including " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " for AGS and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " for SAD and convergence speed on CIFAR100. We also conduct ablation study to demonstrate the compatibility" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 167, + 425, + 407 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Performance comparison with SOTAs on CIFAR10. We report the Top-1 Accuracy performance on ResNet18, ResNet20, and VGGsmall. W/A denotes the bit-width of weights/activations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 167, + 425, + 407 + ], + "lines": [ + { + "bbox": [ + 187, + 167, + 425, + 407 + ], + "spans": [ + { + "bbox": [ + 187, + 167, + 425, + 407 + ], + "type": "table", + "html": "
NetworksMethodsBit-width (W/A)Top-1 Acc.(%)
ResNet18Full-precision32/3294.8
IR-Net [40]1/191.5
RBNN [30]1/192.2
CMIM [45]1/192.2
SiMaN [29]1/192.5
ReCU [53]1/192.8
OvSW (Ours)1/193.2
ResNet20Full-precision32/3292.1
SLB [55]1/185.5
FDA-BNN [52]1/186.2
IR-Net [40]1/186.5
CMIM [45]1/187.3
SiMaN [29]1/187.4
ReCU [53]1/187.4
OvSW (Ours)1/187.7
VGGsmallFull-precision32/3294.1
DoReFa [60]1/190.2
RAD [6]1/190.5
RBNN [30]1/191.3
DSQ [11]1/191.7
Proxy-BNN [16]1/191.8
SLB [55]1/192.0
ReCU [53]1/192.2
FDA-BNN [52]1/192.5
SiMaN [29]1/192.5
OvSW (Ours)1/192.8
", + "image_path": "0651b0b017d3239b04e09a8c123a518d5b4f2b8cea03116f71e238f8241a065c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 437, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 488 + ], + "type": "text", + "content": "and visualize the loss landscape for OvSW. Finally, we deploy OvSW to a real-world mobile device to exhibit its efficiency. To train OvSW, we use one NVIDIA RTX 3090 on the CIFAR10 and CIFAR100 and four on the ImageNet1K. All experiments are implemented on PyTorch [39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 502, + 265, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 265, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 265, + 514 + ], + "type": "text", + "content": "5.1 Results on CIFAR10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We trained OvSW for CIFAR10 with 600 epochs, where the batch size is set to 256 and the initial learning rate to 0.1, decaying with CosineAnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 5e-4 and employ the same data augmentation in ReCU [53]. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " are set to 0.04 and 9e-4 respectively. We compare OvSW with IR-Net [40], RBNN [30], CMIM [45], SiMaN [29], ReCU [53], SLB [55], FDA-BNN [52], DoReFa [60], RAD [6], DSQ [11], and Proxy-BNN [16]. As shown in Tab. 1, OvSW achieves the best performance among all methods. For ResNet18, OvSW obtains " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "93.2\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " top-1 accuracy, which outperforms SiMaN and ReCU by " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " respectively, reducing the accuracy gap between BNNs and full-precision model to " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". In addition, it yields " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "87.7\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "92.8\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " top-1 accuracy on ResNet20 and VGGsmall, which succeeds ReCU and SiMaN by " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 167, + 455, + 376 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Performance comparison with SOTAs on ImageNet1K. We report the Top-1 and Top-5 Accuracy performance on ResNet18 and ResNet34. W/A denotes the bit-width of weights/activations. * means using the two-step training setting as ReActNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 167, + 455, + 376 + ], + "lines": [ + { + "bbox": [ + 157, + 167, + 455, + 376 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 455, + 376 + ], + "type": "table", + "html": "
ModelMethodBit-width (W/A)Top-1 Acc.(%)Top-5 Acc.(%)
ResNet18Full-precision32/3269.689.2
XNOR [42]1/151.273.2
BiReal [34]1/156.479.5
IR-Net [40]1/158.180.0
RBNN [30]1/159.981.9
SiMaN [29]1/160.182.3
FDA-BNN [52]1/160.282.3
ReCU [53]1/161.082.6
OvSW (Ours)1/161.683.1
ReActNet [33]1/165.986.1
ReCU [53]1/166.486.5
OvSW* (Ours)1/166.686.7
ResNet34Full-precision32/3273.391.3
XNOR++ [3]1/157.179.9
BiReal [34]1/162.283.9
IR-Net [40]1/162.984.1
RBNN [30]1/163.184.4
SiMaN [29]1/163.984.8
CMIM [45]1/165.085.7
ReCU [53]1/165.185.8
OvSW (Ours)1/165.586.1
", + "image_path": "291ebf58c2a51a69c414de09bf3712661f2e5b9102666ef82caa61912dba91a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 422, + 282, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 282, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 282, + 434 + ], + "type": "text", + "content": "5.2 Results on ImageNet1K" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": "On ImageNet1K, OvSW is trained from scratch. We train OvSW with 200 epochs, where the batch size is set to 512 and the initial learning rate is set to 0.1, decaying with CosineAnnealing. We adopt SGD optimizer with a momentum of 0.9 and weight decay of 1e-4 and the data augmentation is the same as ReCU [53]. " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " are set to 0.02 and 2e-5 respectively. We demonstrate the ImageNet1K performance of ResNet18/34 and compare OvSW with SOTA methods, including one-stage training methods XNOR [42], BiReal [34], IR-Net [40], RBNN [30], SiMaN [29], FDA-BNN [52], ReCU [53], CMIM [45], and two-stage training method [37] adopted by ReActNet [33]. As shown in Tab. 2, OvSW also achieves the best performance. For ResNet18, OvSW achieves " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "61.6\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " top-1 and " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "83.1\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " top-5 accuracy compared to ReCU's " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "61.0\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " top-1 and " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "82.6\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " top-5 accuracy, which demonstrates the efficiency of overcoming the \"silent weights\". For ResNet34, OvSW achieves " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "65.5\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " top-1 accuracy, which is also better than ReCU. We further compare OvSW with the two-stage training method ReActNet. OvSW obtains " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "66.6\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " top-1 accuracy, succeeding ReActNet and ReCU by " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 239, + 191 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 239, + 191 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 239, + 191 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 239, + 191 + ], + "type": "image", + "image_path": "6895c2fb511bc40f4d87993ac9633d85fe403fdc0196c3f3675e68239cff5347.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 194, + 228, + 205 + ], + "lines": [ + { + "bbox": [ + 158, + 194, + 228, + 205 + ], + "spans": [ + { + "bbox": [ + 158, + 194, + 228, + 205 + ], + "type": "text", + "content": "(a) AGS vs Vanilla" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 116, + 352, + 191 + ], + "blocks": [ + { + "bbox": [ + 250, + 116, + 352, + 191 + ], + "lines": [ + { + "bbox": [ + 250, + 116, + 352, + 191 + ], + "spans": [ + { + "bbox": [ + 250, + 116, + 352, + 191 + ], + "type": "image", + "image_path": "d6e1112df788cf6c6d3eba2d6b1d6d289c81b0a64489042b7a0ca4984cab13fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 193, + 362, + 204 + ], + "lines": [ + { + "bbox": [ + 251, + 193, + 362, + 204 + ], + "spans": [ + { + "bbox": [ + 251, + 193, + 362, + 204 + ], + "type": "text", + "content": "(b) AGS+SAD (OvSW) vs AGS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 115, + 465, + 193 + ], + "blocks": [ + { + "bbox": [ + 365, + 115, + 465, + 193 + ], + "lines": [ + { + "bbox": [ + 365, + 115, + 465, + 193 + ], + "spans": [ + { + "bbox": [ + 365, + 115, + 465, + 193 + ], + "type": "image", + "image_path": "809ac9f4f5de1f838857ad4a76ee314db15488ad2979afb8648143bb561e8706.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 195, + 450, + 205 + ], + "lines": [ + { + "bbox": [ + 392, + 195, + 450, + 205 + ], + "spans": [ + { + "bbox": [ + 392, + 195, + 450, + 205 + ], + "type": "text", + "content": "(c) Convergence" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "type": "text", + "content": "Fig. 3: Top-1 accuracy (mean±std) of binarized ResNet18 w.r.t. different values of " + }, + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "type": "text", + "content": " (a), " + }, + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "type": "text", + "content": " (b) and epoch (c) on CIFAR100." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 250, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 250, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 250, + 275 + ], + "type": "text", + "content": "5.3 Ablation Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "text", + "content": "We investigate the effectiveness of hyper-parameters, including " + }, + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 290, + 480, + 326 + ], + "type": "text", + "content": ", convergence speed, different components, and compatibility through ablation analysis. All the following results are based on binarized ResNet18 for CIFAR100." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": " for AGS and " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": " for SAD. We first compare AGS with vanilla BNNs and analyze the " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": " for AGS. In Fig. 3a, we present the results for nine different settings of " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": ", which vary from 0.01 to 0.09. As seen, the appropriate " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": " can significantly improve the performance of BNNs, and " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda = 0.04" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": " achieves the best performance. If " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": " is too small, the sign of the weights can flip inefficiently; while " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": " is too large, the sign of the weights can flip dramatically, introducing instability to the training. Based on this result, we further introduce SAD to AGS (" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\lambda = 0.04" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": ") to analyze " + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": ". Figure 3b demonstrates that the performance of the model can be further improved by identifying \"silent weights\" and applying additional penalties to them via SAD (" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "inline_equation", + "content": "\\sigma = 0.0009" + }, + { + "bbox": [ + 130, + 327, + 481, + 447 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "text", + "content": "Convergence for OvSW. To verify that OvSW facilitates the flipping of weight signs and thus improves the efficiency of convergence, we fix the " + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "text", + "content": " to 0.04 and 0.0009 respectively, and record the top-1 accuracy over the training epoch from 60 to 500. As shown in Fig. 3c, OvSW effectively accelerates training convergence and achieves better performance. For example, OvSW achieves " + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "inline_equation", + "content": "66.37 \\pm 0.35\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "text", + "content": " top-1 accuracy with only 60 training epochs, while vanilla BNNs only reaches " + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "inline_equation", + "content": "55.28 \\pm 0.07\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "inline_equation", + "content": "65.23 \\pm 0.21\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 591 + ], + "type": "text", + "content": " top-1 accuracy with 60 and 120 training epochs respectively. It indicates that OvSW can achieve significant performance gains in scenarios with limited training resources, such as training the network on edge devices. Meanwhile, although increasing epochs improve the final performance for both OvSW and vanilla BNNs, OvSW consistently outperforms the vanilla BNNs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Components. AGS and SAD promote sign flipping for overall weights and \"silent weights\" respectively. To prove that they are orthogonal to each other, we conduct components ablation study for different modules and show the results in Tab. 3 (left). As seen, AGS and SAD achieve " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "69.61 \\pm 0.22\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "69.45 \\pm 0.18\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " top-1 accuracy respectively and both of them succeed vanilla BNNs. It shows that they are effective in improving the efficiency of weight signs flipping. By" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 146, + 297, + 205 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Left: Ablation study of different components in OvSW. Right: Applying OvSW as a plug-and-play module to other methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 146, + 297, + 205 + ], + "lines": [ + { + "bbox": [ + 159, + 146, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 159, + 146, + 297, + 205 + ], + "type": "table", + "html": "
AGSSADmean±std (%)
XX65.23±0.21
✓X69.61±0.22
X✓69.45±0.18
✓✓69.77±0.15
", + "image_path": "cb3ca3d4596a072411d3873ed7b622133007f541d5e5d07c956eeb4113310070.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 309, + 146, + 452, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 146, + 452, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 146, + 452, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 146, + 452, + 205 + ], + "type": "table", + "html": "
Methodmean±std (%)
AdaBin70.56±0.19
AdaBin+OvSW72.56±0.11
RBNN67.15±0.23
RBNN+OvSW69.98±0.13
", + "image_path": "518ce1e59d72404acf9e2a0cbc12b619a7790a9a20500fb608948c2294f3a6f3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 479, + 248 + ], + "type": "text", + "content": "combining them, OvSW (AGS+SAD) further achieves the best top-1 accuracy with " + }, + { + "bbox": [ + 130, + 225, + 479, + 248 + ], + "type": "inline_equation", + "content": "69.77 \\pm 0.15\\%" + }, + { + "bbox": [ + 130, + 225, + 479, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": "Compatibility. We show the good compatibility of OvSW by inserting it as a plug-and-play module into the current state-of-the-art methods, including AdaBin [48] and RBNN [30]. The former introduces an adaptive binary set to enhance the feature representation of BNNs, while the latter proposes a training-aware approximation function to reduce the gradient estimation error during training. Table 3 (right) demonstrates OvSW achieves " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "72.56 \\pm 0.11\\%" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": " top-1 accuracy on AdaBin and " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "69.98 \\pm 0.13\\%" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": " on RBNN, succeeding their original performance. This result indicates that OvSW has good compatibility and can effectively enhance the performance of existing methods." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 142, + 385, + 248, + 470 + ], + "blocks": [ + { + "bbox": [ + 142, + 385, + 248, + 470 + ], + "lines": [ + { + "bbox": [ + 142, + 385, + 248, + 470 + ], + "spans": [ + { + "bbox": [ + 142, + 385, + 248, + 470 + ], + "type": "image", + "image_path": "279443e141fae0a2bf64c41a77474b9af62136279d270fa9b5d0bd2c7ec75f5f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 472, + 224, + 481 + ], + "lines": [ + { + "bbox": [ + 162, + 472, + 224, + 481 + ], + "spans": [ + { + "bbox": [ + 162, + 472, + 224, + 481 + ], + "type": "text", + "content": "(a) Full-precision" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 256, + 385, + 361, + 469 + ], + "blocks": [ + { + "bbox": [ + 256, + 385, + 361, + 469 + ], + "lines": [ + { + "bbox": [ + 256, + 385, + 361, + 469 + ], + "spans": [ + { + "bbox": [ + 256, + 385, + 361, + 469 + ], + "type": "image", + "image_path": "6eebd3e9d4cd4094fd46d4f4305b5710935c510020cc0a81ba8fed42e1781c7c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 472, + 340, + 482 + ], + "lines": [ + { + "bbox": [ + 273, + 472, + 340, + 482 + ], + "spans": [ + { + "bbox": [ + 273, + 472, + 340, + 482 + ], + "type": "text", + "content": "(b) XNOR-Net++" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 491, + 479, + 512 + ], + "lines": [ + { + "bbox": [ + 131, + 491, + 479, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 479, + 512 + ], + "type": "text", + "content": "Fig. 4: 3D visualization of the loss surfaces of ResNet18 on CIFAR100, which is used to enable comparisons of sharpness/flatness of different methods." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 385, + 476, + 469 + ], + "blocks": [ + { + "bbox": [ + 369, + 385, + 476, + 469 + ], + "lines": [ + { + "bbox": [ + 369, + 385, + 476, + 469 + ], + "spans": [ + { + "bbox": [ + 369, + 385, + 476, + 469 + ], + "type": "image", + "image_path": "2f3830e5b67c1b66adf1cfa49777dcaf2d99080c4b17e14e376d973f70a8db95.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 472, + 440, + 481 + ], + "lines": [ + { + "bbox": [ + 402, + 472, + 440, + 481 + ], + "spans": [ + { + "bbox": [ + 402, + 472, + 440, + 481 + ], + "type": "text", + "content": "(c) OvSW" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 552, + 308, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 308, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 308, + 563 + ], + "type": "text", + "content": "5.4 Loss Landscape Visualization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "BNNs restrict the weights and activations to discrete values, which naturally limits the representational capacity of the model and further result in disparate optimization landscapes compared to real-valued ones [32]. As illustrated in Fig. 4, we follow the method in [26] to plot the actual optimization landscape about our OvSW and compare it with the same architecture to real-valued and XNOR-Net++. As seen, our OvSW has a significantly smoother loss-landscape and minor loss elevation compared to XNOR-Net++, which confirms the effectiveness of OvSW in the BNN optimization." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 158, + 135, + 455, + 210 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "type": "text", + "content": "Table 4: Comparing OvSW (1-bit) with 32-bit and 16-bit backbones on M1 Pro." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 158, + 135, + 455, + 210 + ], + "lines": [ + { + "bbox": [ + 158, + 135, + 455, + 210 + ], + "spans": [ + { + "bbox": [ + 158, + 135, + 455, + 210 + ], + "type": "table", + "html": "
NetworkW/ASize (MB)Memory SavingLatency (ms)Acceleration
ResNet1832/3246.76-27.67-
16/1623.382×16.651.66×
1/12.8116.64×7.973.47×
ResNet3432/3287.19-48.82-
16/1643.602×29.011.68×
1/14.1221.16×14.283.42×
", + "image_path": "a8ca2ca6beef932bae7616679e16bc71461baaa4ae018f85b08dffad9159a27c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 242, + 273, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 273, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 273, + 255 + ], + "type": "text", + "content": "5.5 Deployment Efficiency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "text", + "content": "We implement 1-bit models on the M1 Pro, which features 8 high-performance Firestorm cores and 2 efficient Icestorm cores in a hybrid design. The Firestorm cores can be clocked up to " + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "inline_equation", + "content": "3.2\\mathrm{GHz}" + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "text", + "content": ", while the Icestorm cores can reach 2.1 GHz. Our OvSW shows significant efficiency gains when deployed on real-world mobile devices, as evidenced by practical speed evaluations. To make our inference framework BOLT [9] compatible with OvSW, we leverage the ARM NEON SIMD instruction SSHL. We compare OvSW with 32-bit and 16-bit backbones. As shown in Tab. 4, OvSW inference speed is substantially faster with the highly efficient BOLT framework in a single thread. For instance, OvSW achieves an acceleration rate of about " + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "inline_equation", + "content": "3.47\\times" + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "text", + "content": " on ResNet18 compared to its 32-bit counterpart. For the ResNet34 backbone, OvSW achieves a " + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "inline_equation", + "content": "3.42\\times" + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "text", + "content": " acceleration rate with the BOLT framework on hardware, which is significant for computer vision applications on real-world edge devices. At the same time, OvSW can save memory by a factor of " + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "inline_equation", + "content": "16.64\\times" + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "inline_equation", + "content": "21.16\\times" + }, + { + "bbox": [ + 130, + 266, + 482, + 446 + ], + "type": "text", + "content": ", which demonstrates its potential for applications with limited memory resources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 468, + 220, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 220, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 220, + 481 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "BNN is a crucial method to compress deep learning models and reduce inference overhead. In this paper, systematically and theoretically, we prove the distribution of gradients is independent of latent weights, which is the root cause of inefficient updating and performance degradation of BNNs. To this end, Adaptive Gradient Scaling (AGS) and Silence Awareness Decaying (SAD), are proposed to Overcome Silent Weights (OvSW) and achieve SOTA performance on BNNs. Specifically, AGS adaptively scales the gradient based on the distribution of weights, improving the efficiency of sign flipping for the overall weights. SAD can effectively measure the states of weight flipping, detect the \"silent weights\" and introduce additional penalty to them to facilitate their flipping. Extensive experiments demonstrate that OvSW can achieve notable performance gains over the SOTA methods on various datasets and networks. In addition, OvSW has better convergence efficiency and excellent compatibility, which can be combined with existing methods to further enhance the performance of BNNs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013) 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Brock, A., De, S., Smith, S.L., Simonyan, K.: High-performance large-scale image recognition without normalization. In: Int. Conf. Mach. Learn. pp. 1059-1071. PMLR (2021) 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "text", + "content": "3. Bulat, A., Tzimiropoulos, G.: Xnor-net++: Improved binary neural networks. arXiv preprint arXiv:1909.13863 (2019) 4, 5, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "4. Chen, J., Liu, L., Liu, Y., Zeng, X.: A learning framework for n-bit quantized neural networks toward fpgas. IEEE Transactions on Neural Networks and Learning Systems 32(3), 1067-1081 (2020) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "5. Courbariaux, M., Hubara, I., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks: Training deep neural networks with weights and activations constrained to + 1 or -1. arXiv preprint arXiv:1602.02830 (2016) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "text", + "content": "6. Ding, R., Chin, T.W., Liu, Z., Marculescu, D.: Regularizing activation distribution for training binarized deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11408-11417 (2019) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": "7. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13733-13742 (2021) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 392 + ], + "type": "text", + "content": "8. Duan, Y., Chen, X., Xu, H., Chen, Z., Liang, X., Zhang, T., Li, Z.: Transnas-bench101: Improving transferability and generalizability of cross-task neural architecture search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5251-5260 (2021) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 421, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 421, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 421, + 402 + ], + "type": "text", + "content": "9. Feng, J.: Bolt. https://github.com/huawei-noah/bolt (2021) 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "0. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 580-587 (2014) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "1. Gong, R., Liu, X., Jiang, S., Li, T., Hu, P., Lin, J., Yu, F., Yan, J.: Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In: Int. Conf. Comput. Vis. pp. 4852–4861 (2019) 2, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "text", + "content": "2. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Int. Conf. Comput. Vis. pp. 2961-2969 (2017) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Int. Conf. Comput. Vis. pp. 1026-1034 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "4. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans. Pattern Anal. Mach. Intell. 37(9), 1904-1916 (2015) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "text", + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 1, 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "text", + "content": "6. He, X., Mo, Z., Cheng, K., Xu, W., Hu, Q., Wang, P., Liu, Q., Cheng, J.: Proxybnn: Learning binarized neural networks via proxy matrices. In: Eur. Conf. Comput. Vis. pp. 223-241. Springer (2020) 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "7. Helwegen, K., Widdicombe, J., Geiger, L., Liu, Z., Cheng, K.T., Nusselder, R.: Latent weights do not exist: Rethinking binarized neural network optimization. Adv. Neural Inform. Process. Syst. 32 (2019) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "8. Hinton, G., Vinyals, O., Dean, J., et al.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 2(7) (2015) 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "19. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC). pp. 10-14. IEEE (2014) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 192 + ], + "type": "text", + "content": "20. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2704-2713 (2018) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "text", + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "text", + "content": "22. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "text", + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst. 25 (2012) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "24. Lee, C., Kim, H., Park, E., Kim, J.J.: Insta-bnn: Binary neural network with instance-aware threshold. In: Int. Conf. Comput. Vis. pp. 17325-17334 (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "text", + "content": "25. Lee, J., Kim, D., Ham, B.: Network quantization with element-wise gradient scaling. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6448-6457 (2021) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 299, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 480, + 320 + ], + "type": "text", + "content": "26. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. In: Adv. Neural Inform. Process. Syst. (2018) 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 352 + ], + "type": "text", + "content": "27. Li, S., Lin, M., Wang, Y., Fei, C., Shao, L., Ji, R.: Learning efficient gans for image translation via differentiable masks and co-attention distillation. IEEE Trans. Multimedia (2022) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "text", + "content": "28. Lin, M., Ji, R., Wang, Y., Zhang, Y., Zhang, B., Tian, Y., Shao, L.: Hrank: Filter pruning using high-rank feature map. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1529-1538 (2020) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 385, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 417 + ], + "type": "text", + "content": "29. Lin, M., Ji, R., Xu, Z., Zhang, B., Chao, F., Lin, C.W., Shao, L.: Siman: Sign-to-magnitude network binarization. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 10, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 417, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 480, + 449 + ], + "type": "text", + "content": "30. Lin, M., Ji, R., Xu, Z., Zhang, B., Wang, Y., Wu, Y., Huang, F., Lin, C.W.: Rotated binary neural network. Adv. Neural Inform. Process. Syst. 33, 7474-7485 (2020) 4, 10, 11, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 471 + ], + "type": "text", + "content": "31. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Adv. Neural Inform. Process. Syst. 30 (2017) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 503 + ], + "type": "text", + "content": "32. Liu, Z., Shen, Z., Li, S., Helwegen, K., Huang, D., Cheng, K.T.: How do adam and training strategies help bnns optimization. In: Int. Conf. Mach. Learn. pp. 6936-6946. PMLR (2021) 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 503, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 480, + 536 + ], + "type": "text", + "content": "33. Liu, Z., Shen, Z., Savvides, M., Cheng, K.T.: Reactnet: Towards precise binary neural network with generalized activation functions. In: Eur. Conf. Comput. Vis. pp. 143-159. Springer (2020) 5, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 578 + ], + "type": "text", + "content": "34. Liu, Z., Wu, B., Luo, W., Yang, X., Liu, W., Cheng, K.T.: Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In: Eur. Conf. Comput. Vis. pp. 722-737 (2018) 4, 5, 6, 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3431-3440 (2015) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "36. Luo, J.H., Wu, J.: Autopruner: An end-to-end trainable filter pruning method for efficient deep model inference. Pattern Recognition 107, 107461 (2020) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 663 + ], + "type": "text", + "content": "37. Martinez, B., Yang, J., Bulat, A., Tzimiropoulos, G.: Training binary neural networks with real-to-binary convolutions. arXiv preprint arXiv:2003.11535 (2020) 11" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "38. Nagel, M., Fournarakis, M., Bondarenko, Y., Blankevoort, T.: Overcoming oscillations in quantization-aware training. In: Int. Conf. Mach. Learn. pp. 16318-16330. PMLR (2022) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 239 + ], + "type": "text", + "content": "39. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché Buc, F., Fox, E., Garnett, R. (eds.) Adv. Neural Inform. Process. Syst. pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 239, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 272 + ], + "type": "text", + "content": "40. Qin, H., Gong, R., Liu, X., Shen, M., Wei, Z., Yu, F., Song, J.: Forward and backward information retention for accurate binary neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2250-2259 (2020) 4, 5, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 272, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 272, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 481, + 316 + ], + "type": "text", + "content": "41. Rajbhandari, S., Rasley, J., Ruwase, O., He, Y.: Zero: Memory optimizations toward training trillion parameter models. In: SC20: International Conference for High Performance Computing, Networking, Storage and Analysis. pp. 1-16. IEEE (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 316, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 481, + 350 + ], + "type": "text", + "content": "42. Rastegari, M., Ordonez, V., Redmon, J., Farhadi, A.: Xnor-net: Imagenet classification using binary convolutional neural networks. In: Eur. Conf. Comput. Vis. pp. 525-542. Springer (2016) 2, 4, 5, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 351, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 481, + 384 + ], + "type": "text", + "content": "43. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2016) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 384, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 481, + 418 + ], + "type": "text", + "content": "44. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (2015) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "type": "text", + "content": "45. Shang, Y., Xu, D., Zong, Z., Yan, Y.: Network binarization via contrastive learning. arXiv preprint arXiv:2207.02970 (2022) 10, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 441, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 463 + ], + "type": "text", + "content": "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "47. Su, X., Huang, T., Li, Y., You, S., Wang, F., Qian, C., Zhang, C., Xu, C.: Prioritized architecture sampling with monto-carlo tree search. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10968-10977 (2021) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 498, + 481, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 529 + ], + "type": "text", + "content": "48. Tu, Z., Chen, X., Ren, P., Wang, Y.: Adabin: Improving binary neural networks with adaptive binary sets. In: Eur. Conf. Comput. Vis. pp. 379-395. Springer (2022) 4, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 531, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 575 + ], + "type": "text", + "content": "49. Wu, X.M., Zheng, D., Liu, Z., Zheng, W.S.: Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17055-17064 (2023) 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "50. Xu, S., Li, Y., Ma, T., Lin, M., Dong, H., Zhang, B., Gao, P., Lu, J.: Resilient binary neural network. In: AAAI. vol. 37, pp. 10620-10628 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "text", + "content": "51. Xu, S., Li, Y., Wang, T., Ma, T., Zhang, B., Gao, P., Qiao, Y., Lü, J., Guo, G.: Recurrent bilinear optimization for binary neural networks. In: Eur. Conf. Comput. Vis. pp. 19-35. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "52. Xu, Y., Han, K., Xu, C., Tang, Y., Xu, C., Wang, Y.: Learning frequency domain approximation for binary neural networks. Adv. Neural Inform. Process. Syst. 34, 25553-25565 (2021) 10, 11" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 392 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "53. Xu, Z., Lin, M., Liu, J., Chen, J., Shao, L., Gao, Y., Tian, Y., Ji, R.: Recu: Reviving the dead weights in binary neural networks. In: Int. Conf. Comput. Vis. pp. 5198-5208 (2021) 3, 4, 5, 10, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "54. Yang, J., Shen, X., Xing, J., Tian, X., Li, H., Deng, B., Huang, J., Hua, X.s.: Quantization networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7308-7316 (2019) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "55. Yang, Z., Wang, Y., Han, K., Xu, C., Xu, C., Tao, D., Xu, C.: Searching for low-bit weights in quantized neural networks. Adv. Neural Inform. Process. Syst. 33, 4091-4102 (2020) 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 237 + ], + "type": "text", + "content": "56. You, Y., Gitman, I., Ginsburg, B.: Large batch training of convolutional networks. arXiv preprint arXiv:1708.03888 (2017) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "57. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Eur. Conf. Comput. Vis. pp. 365-382 (2018) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "58. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6848-6856 (2018) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "59. Zhao, J., Yang, L., Zhang, B., Guo, G., Doermann, D.S.: Uncertainty-aware binary neural networks. In: IJCAI. pp. 3441-3447 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "60. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 357, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 480, + 392 + ], + "type": "text", + "content": "61. Zhu, F., Gong, R., Yu, F., Liu, X., Wang, Y., Li, Z., Yang, X., Yan, J.: Towards unified int8 training for convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1969-1979 (2020) 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Xiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_content_list.json b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b2f4aa4efe7d0b388f3c94441f6e3099a6cb2f49 --- /dev/null +++ b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_content_list.json @@ -0,0 +1,1706 @@ +[ + { + "type": "text", + "text": "Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunfeng Fan $^{1}$ , Wenchao Xu $^{1(\\boxdot)}$ , Haozhao Wang $^{2}$ , Fushuo Huo $^{1}$ , Jinyu Chen $^{1}$ , and Song Guo $^{3}$", + "bbox": [ + 225, + 212, + 779, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Department of Computing, The Hong Kong Polytechnic University, Hong Kong", + "$^{2}$ School of Computer Science and Technology, Huazhong University of Science and Technology, Hubei, China", + "3 Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong" + ], + "bbox": [ + 225, + 253, + 777, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yunfeng.fan, fushuo.huo, jinyu.chen}@connect.polyu.hk \nwenchao.xu@polyu.edu.hk, hz_wang@hust.edu.cn, songguo@cse.ust.hk \nhttps://github.com/fanyunfeng-bit/Balanced-Modality-Selection-in-MFL", + "bbox": [ + 233, + 325, + 766, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Selecting proper clients to participate in each federated learning (FL) round is critical to effectively harness a broad range of distributed data. Existing client selection methods simply consider the mining of distributed uni-modal data, yet, their effectiveness may diminish in multi-modal FL (MFL) as the modality imbalance problem not only impedes the collaborative local training but also leads to a severe global modality-level bias. We empirically reveal that local training with a certain single modality may contribute more to the global model than training with all local modalities. To effectively exploit the distributed multiple modalities, we propose a novel Balanced Modality Selection framework for MFL (BMSFed) to overcome the modal bias. On the one hand, we introduce a modal enhancement loss during local training to alleviate local imbalance based on the aggregated global prototypes. On the other hand, we propose the modality selection aiming to select subsets of local modalities with great diversity and achieving global modal balance simultaneously. Our extensive experiments on audio-visual, colored-gray, and front-back datasets showcase the superiority of BMSFed over baselines and its effectiveness in multi-modal data exploitation.", + "bbox": [ + 259, + 409, + 743, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multi-modal federated learning $\\cdot$ modality imbalance $\\cdot$ modality selection", + "bbox": [ + 259, + 672, + 751, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 729, + 375, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) [23] aims to collaboratively learn data that has been collected by, and resides on, a number of remote devices or servers. FL stands to develop top-performing models by aggregating knowledge from numerous edge clients [8,36], which relies on the iterative interaction among participating clients and the server. However, comprehensively employing the information from all", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/93439485a3d596deee8af2d8086c2d99f2de57511d8c9dd41795e03d8837f1c7.jpg", + "table_caption": [ + "Table 1: Performance of various client selection methods in MFL under IID setting. A and V denote uni-audio and uni-visual while A-V means the multi-modal result. 'Local' represents that a client is trained based its local data without aggregation. A strong modal bias of global model exists on the two datasets." + ], + "table_footnote": [], + "table_body": "
DatasetCREMA-D [4]AVE [33]
MethodAVA-VAVA-V
Local41.920.439.633.416.735.2
FedAvg51.220.650.761.126.862.2
pow-d [6]51.520.450.561.926.962.5
DivFL [3]52.321.151.762.725.363.3
FedAvg-0.250.628.652.460.629.663.4
FedAvg-0.550.534.655.758.730.060.7
FedAvg-0.848.150.961.256.431.858.5
BMSFed51.041.964.559.740.264.7
", + "bbox": [ + 251, + 212, + 746, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "clients can be exceptionally difficult due to the client heterogeneity and resource limitations [10, 29].", + "bbox": [ + 212, + 378, + 782, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Random sampling [22, 34] from available clients has been widely used in FL to satisfy some practical restrictions, e.g., limited communication bandwidth [26, 35] and computing capacities [15]. To improve the information exploitation of all clients, extensive research has been conducted on effective client selection strategies [9, 43]. Despite the success of traditional client selection methods in uni-modal FL, their effectiveness diminishes when dealing with clients with multi-modal data as the inter-modal interactions during the MFL training are neglected. According to [14, 37], there may exist inconsistent learning paces for different modalities in multi-modal joint training, i.e., modality imbalance, which not only impedes the collaborative local training but also leads to a severe modal bias for global model in MFL. As illustrated in Tab. 1, audio modality significantly outperforms visual modality in CREMA-D and AVE datasets during local training and the aggregated model still suffers from it. However, two well-designed client selection methods (pow-d [6] and DivFL [3]) only obtain severely limited improvement over random sampling (FedAvg) on the multi-modal global model. We can see that client selection methods achieve the best uni-audio performance while uni-visual performance even drops sometimes, which means existing client selection scheme heavily relies on the better modality, while ignoring the importance of improving weak modalities that also has potential for global model aggregation. Based on the above analysis, a pivotal question arises: Can we design a new selection scheme in MFL that can overcome the modal bias and exploit each modality comprehensively?", + "bbox": [ + 212, + 412, + 787, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer this question, we investigate the interactions between different modalities via randomly discarding the data from one modality (audio or visual) on part of clients, which is inspired from modality dropout [2, 40] that drops a specific modal data during training for regularization. The results are shown in rows 7-9 in Tab. 1, where $-x$ denotes randomly discarding a modality on a client with probability $x$ . We can see that dropping with a certain prob", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/120a71489c7f4011cf5f68f0ab0ba1f38a566c955a660b2fcc2ccf8631830f73.jpg", + "image_caption": [ + "Fig. 1: Left: Traditional client selection in FL aims to sample a client subset in each round while our modality selection considers each local modality as the sampling unit. Right: The paradigm of BMSFed with four clients. The global prototypes are used to enhance the weak modality during local update. Only networks corresponding to the selected modalities will be uploaded to the server for aggregation." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 781, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ability can improve the global multi-modal performance on both datasets (e.g., FedAvg-0.2), and the main reason comes from the dramatic improvement of visual modality. This phenomenon suggests that performing uni-modal training can unleash its potential without being inhibited by another modality, and uni-modal local training may contribute more to the global model than multi-modal training on some clients. As the dropping ratio increases, although the visual modality still improves, multi-modal performance may decline as the audio modality declines as shown in columns 5 and 7, suggesting that we should carefully control which modalities on each client should be involved in training and aggregation to make the contribution most to the global model.", + "bbox": [ + 212, + 388, + 787, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to the above investigations, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to mitigating the modal bias and comprehensively exploit the diverse information from all modalities. Specifically, instead of selecting a subset of clients, we treat each modality on the local side as a selection unit, as demonstrated in Fig. 1. Our BMSFed mainly contains two parts: Firstly, we intend to alleviate the local modality imbalance by introducing a modal enhancement loss based on aggregated global prototypes to promote the performance of weak modality. Secondly, we complete the modality selection by building two separated submodular functions for selecting multi-modal clients (training with multi-modal data) and uni-modal clients (training with selected uni-modal data) respectively. Inspired from [3], the criterion is to select modalities that are most representative on the gradients while also alleviate the global modal bias. A simple yet effective conflict resolution strategy is devised to ensure the validity of modality selection and keep modal balance on global model simultaneously.", + "bbox": [ + 212, + 544, + 787, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of the paper are summarized as follows:", + "bbox": [ + 238, + 773, + 702, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. We empirically analyze the modality imbalance problem in MFL and reveal that uni-modal training on some clients may contribute more to the global model than multi-modal training.", + "bbox": [ + 222, + 794, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 478, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Based on the analysis, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to comprehensively exploit all modalities via a modal enhancement loss and representative modality selection to overcome the global modal bias.", + "bbox": [ + 220, + 146, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. We conduct comprehensive experiments on audio-visual, colored-gray, and front-back datasets, and considering the statistical heterogeneity and modality incongruity problems in MFL, to validate the superiority of our BMSFed.", + "bbox": [ + 220, + 205, + 784, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Background and Related Works", + "text_level": 1, + "bbox": [ + 215, + 273, + 560, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Multi-modal federated learning", + "text_level": 1, + "bbox": [ + 215, + 296, + 521, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In MFL, each client has one or multiple modalities of data. Without loss of generality, we consider two input modalities, which are denoted by $A$ and $I$ , respectively in MFL. There are a set $V$ of $N$ clients, $V = [N]$ , respectively owning datasets $\\mathcal{D}_i = \\left\\{\\mathbf{X}_i^A, \\mathbf{X}_i^I, \\mathbf{y}\\right\\}$ , $i \\in [N]$ . A typical federated learning objective is the average of each client's local loss function:", + "bbox": [ + 212, + 318, + 782, + 401 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min f (\\theta) = \\sum_ {k = 1} ^ {N} p _ {k} F _ {k} (\\theta) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 410, + 784, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta = \\{\\theta^A, \\theta^I, \\omega\\}$ denotes the model parameters. $\\theta^A$ and $\\theta^I$ represent the encoder parameters of modality $A$ and $I$ . $\\omega$ is the parameter of fusion classifier. $p_k$ is a pre-defined weight. $F_k$ is each client's local loss (cross entropy (CE) loss for classification task in this paper).", + "bbox": [ + 212, + 446, + 782, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Statistical heterogeneity [16,20] is a widely concerned challenge in uni-modal FL. To tackle this issue, FedProx [21] uses a proximal term to stabilize model aggregation. FedProto [32] shares class prototypes to regularize the learning of local models. In MFL, modality incongruity [44,45] (clients consist of different modalities combinations), as well as statistical heterogeneity [41], are all considered. Yu et al. [44] propose CreamFL to align the representations between different clients and different modalities via communicating knowledge on a public dataset. Chen et al. [5] introduce FedMSplit to split local models into several components and aggregate them by their correlations. However, they still focus on the heterogeneity, but ignore the interaction between private data of different modalities, which limits their information exploitation.", + "bbox": [ + 212, + 507, + 784, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Client selection and submodular function", + "text_level": 1, + "bbox": [ + 214, + 686, + 602, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Client selection [9,43] is a critical issue for FL especially when the communication cost with all devices is prohibitively high, which has been extensively studied in uni-modal FL. Cho et al. [6] propose Power-of-Choice to select clients with targets local loss. Balakrishnan et al. [3] propose to select a subset of clients with great diversity.", + "bbox": [ + 212, + 704, + 782, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diverse client selection via submodularity. Maximizing a submodular function is reported to improve the diversity and reduce the redundancy of a subset. This property makes it appropriate for client selection in FL. If a function $F$ is submodular, it should satisfy: given a finite ground set $V$ of size", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$N$ , $F(A \\cup \\{v\\}) - F(A) \\geqslant F(B \\cup \\{v\\}) - F(B)$ , for any $A \\subseteq B \\subseteq V$ and $v \\in V \\setminus B$ . The marginal utility of an element $v$ w.r.t. a subset $A$ is denoted as $F(v|A) = F(A \\cup \\{v\\}) - F(A)$ , which can represent the importance of $v$ to $A$ . The client selection via submodular maximization can be expressed following [3]: find a subset $S$ of clients whose aggregated gradients can approximate the full aggregation from all clients:", + "bbox": [ + 214, + 145, + 782, + 237 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {k \\in [ N ]} \\nabla F _ {k} (v ^ {k}) = \\sum_ {k \\in [ N ]} \\left[ \\nabla F _ {k} (v ^ {k}) - \\nabla F _ {\\sigma (k)} (v ^ {\\sigma (k)}) \\right] + \\sum_ {k \\in S} \\gamma_ {k} \\nabla F _ {k} (v ^ {k}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 250, + 785, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ maps $V\\to S$ and the gradient $\\nabla F_k(v^k)$ from client $k$ is approximated by the gradient from a selected client $\\sigma (k)\\in S$ . For $i\\in S$ , let $C_i\\triangleq \\{k\\in V|\\sigma (k) = i\\}$ , and therefore $\\gamma_{i}\\triangleq |C_{i}|$ . Take the norms and apply triangular inequality after subtracting the second term from both sides, we can obtain a relaxed objective $G(S)$ for minimizing the approximation error:", + "bbox": [ + 214, + 296, + 782, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\sum_ {k \\in [ N ]} \\nabla F _ {k} \\left(v ^ {k}\\right) - \\sum_ {k \\in S} \\gamma_ {k} \\nabla F _ {k} \\left(v ^ {k}\\right) \\right\\| \\leqslant \\sum_ {k \\in [ N ]} \\min _ {i \\in S} \\left\\| \\nabla F _ {k} \\left(v ^ {k}\\right) - \\nabla F _ {i} \\left(v ^ {i}\\right) \\right\\| \\triangleq G (S) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 387, + 785, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minimizing $G(S)$ can be seen as maximizing the well-known submodular function, i.e., the facility location function [7]. The submodular maximizing problem is NP-hard but can be approximated via the greedy [25] or stochastic greedy algorithm [24]:", + "bbox": [ + 214, + 462, + 782, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS \\leftarrow S \\cup k ^ {*}, k ^ {*} \\in \\underset {k \\in \\operatorname {r a n d} (V \\backslash S, \\operatorname {s i z e} = s)} {\\arg \\max } \\left[ \\bar {G} (S) - \\bar {G} (\\{k \\} \\cup S) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 535, + 784, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\bar{G}$ represents a constant minus the negation of $G$", + "bbox": [ + 214, + 575, + 571, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although these methods make great improvement in uni-modal FL, the selection strategy is under-explored in MFL and we reveal that traditional client selection approaches cannot address the severe modal bias in MFL.", + "bbox": [ + 212, + 590, + 782, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Imbalanced multi-modal learning", + "text_level": 1, + "bbox": [ + 214, + 652, + 535, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modality imbalance indicates the inconsistent learning progress of different modalities in multi-modal learning [14,37]. Peng et al. [28] propose OGM-GE to alleviate the inhibitory effect on weak modality by slowing down the dominant modality. Fan et al. [11] further build a non-parametric classifier by class centroids to adjust the update direction of weak modality. In this paper, we aim to power each modality of all clients by a meticulously designed modality selection strategy in each round of training.", + "bbox": [ + 212, + 674, + 787, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 795, + 330, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce BMSFed that contains the local imbalance alleviation and balanced modality selection.", + "bbox": [ + 212, + 820, + 782, + 849 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 477, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Local Imbalance Alleviation", + "text_level": 1, + "bbox": [ + 215, + 146, + 493, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed in Sec. 2.3 and Tab. 1, the multi-modal training on each client may suffer from severe modality imbalance, leading to inadequate information exploitation and consequently incurring the modal bias on the global model. Therefore, we first try to alleviate the imbalance during local training.", + "bbox": [ + 212, + 169, + 782, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "According to [11], the prototypes (i.e., class centroids) are suitable to calibrate the gradient directions to avoid the interference from other modalities. Therefore, we choose to use prototypes to facilitate learning of weak modality. For the $i$ -th client with the data $\\mathcal{D}_i = \\left\\{\\boldsymbol{X}_i^A, \\boldsymbol{X}_i^I, \\boldsymbol{y}\\right\\}$ , the local prototype for class $j$ of each modality is defined as the mean value of representations:", + "bbox": [ + 212, + 229, + 784, + 313 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i, j} ^ {I} = \\frac {1}{\\left| \\mathcal {D} _ {i , j} \\right|} \\sum_ {x ^ {I} \\in \\mathcal {D} _ {i, j}} h _ {i} \\left(\\theta_ {i} ^ {I}; x ^ {I}\\right), c _ {i, j} ^ {A} = \\frac {1}{\\left| \\mathcal {D} _ {i , j} \\right|} \\sum_ {x ^ {A} \\in \\mathcal {D} _ {i, j}} h _ {i} \\left(\\theta_ {i} ^ {A}; x ^ {A}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 321, + 785, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{i,j}$ denotes the samples belonging to $j$ -th class in client $i$ . $h_i$ is the function of encoder. Considering the heterogeneity across clients, we aggregate the local prototypes to a global prototype as:", + "bbox": [ + 212, + 366, + 782, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {j} ^ {G I} = \\frac {1}{| \\mathcal {N} _ {j} |} \\sum_ {i \\in \\mathcal {N} _ {j}} \\frac {| \\mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {I}, c _ {j} ^ {G A} = \\frac {1}{| \\mathcal {N} _ {j} |} \\sum_ {i \\in \\mathcal {N} _ {j}} \\frac {| \\mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {A} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 421, + 785, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_j$ denotes the set of clients that have class $j$ and $N_{j}$ is the number of instances belonging to class $j$ over all clients. Then, we introduce a modal enhancement loss (ME) based on global prototype to adjust local training:", + "bbox": [ + 212, + 465, + 782, + 513 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M E} ^ {k} \\left(v _ {I} ^ {k}\\right) = - \\mathbb {E} _ {\\left(x _ {i} ^ {I}, y\\right) \\in \\mathcal {D} _ {k}} \\log \\left[ \\frac {\\exp \\left(- d \\left(z _ {i} ^ {I} , c _ {y} ^ {G I}\\right)\\right)}{\\sum_ {j = 1} ^ {Y} \\exp \\left(- d \\left(z _ {i} ^ {I} , c _ {j} ^ {G I}\\right)\\right)} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 523, + 785, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M E} ^ {k} \\left(v _ {A} ^ {k}\\right) = - \\mathbb {E} _ {\\left(x _ {i} ^ {A}, y\\right) \\in \\mathcal {D} _ {k}} \\log \\left[ \\frac {\\exp \\left(- d \\left(z _ {i} ^ {A} , c _ {y} ^ {G A}\\right)\\right)}{\\sum_ {j = 1} ^ {Y} \\exp \\left(- d \\left(z _ {i} ^ {A} , c _ {j} ^ {G A}\\right)\\right)} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 566, + 705, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ is the distance function (Euclidean distance), $z_i^I$ is the representation of $x_i^I$ , i.e., $z_i^I = h_i(\\theta_i^I; x_i^I)$ . $Y$ is the class number. $v_A^k$ and $v_I^k$ indicate the corresponding modal data in client $k$ . Hence, the local loss should be (data superscripts are omitted for simplicity):", + "bbox": [ + 212, + 614, + 782, + 676 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {k} \\left(v _ {A}, v _ {I}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\gamma^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {A}\\right), \\rho_ {I} ^ {k} \\leqslant 1 \\\\ \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right), \\rho_ {I} ^ {k} > 1 \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 686, + 785, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\rho_I^k$ indicates the local imbalance ratio for client $k$ calculated based on the modality-wise ground truth prediction, and $\\rho_I^k > 1$ means modality $A$ outperforms modality $I$ and vice versa. $\\gamma^k, \\beta^k \\in (0,1)$ are the modulation coefficients. The calculation details for $\\gamma^k, \\beta^k$ and $\\rho_I^k$ are in Appendix.", + "bbox": [ + 212, + 728, + 782, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different from [11], we apply the global prototype instead of local prototypes to aggregate the knowledge from different clients, and the ME loss is not only applied to multi-modal clients, but also to uni-modal clients (See details in Sec. 3.2). Now, the weak modality will be stimulated through local update.", + "bbox": [ + 212, + 789, + 782, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Balanced Modality Selection", + "text_level": 1, + "bbox": [ + 215, + 146, + 498, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering that different types of modal combinations may be selected to participate in local training in our scheme, we define the participated clients as multi-modal clients and uni-modal clients (e.g., client 2 in Fig. 1 Right is a uni-modal client with modal $I$ for training while client 1 and 4 are multi-modal clients with both modalities for training). In Sec. 3.1, we propose the ME loss to facilitate the local learning of weak modality, which is originally designed for multi-modal clients. Here, we also apply it on uni-weak-modal clients to realize gradient consistency on different types of clients.", + "bbox": [ + 212, + 164, + 787, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Assume modality $I$ is weak here. Hence, the local loss for multi-modal and uni-modal clients should be:", + "bbox": [ + 214, + 284, + 785, + 314 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {m u l t i - m o d a l}: F _ {k} \\left(v _ {A}, v _ {I}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right) \\\\ \\text {u n i - m o d a l}: F _ {k} \\left(v _ {A}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}\\right), F _ {k} \\left(v _ {I}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right) \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 320, + 785, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, following Eq. (2), we define the paradigm of modality selection for MFL, aiming to approximate the full gradient aggregation from all clients by the gradient from the customized multi-modal and uni-modal clients:", + "bbox": [ + 214, + 368, + 785, + 414 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {k \\in [ N ]} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) = \\sum_ {k \\in [ N ]} \\left[ \\begin{array}{c} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) - \\nabla F _ {\\sigma_ {M} (k)} \\left(v _ {A}, v _ {I}\\right) \\\\ - \\nabla F _ {\\sigma_ {A} (k)} \\left(v _ {A}\\right) - \\nabla F _ {\\sigma_ {I} (k)} \\left(v _ {I}\\right) \\end{array} \\right] \\\\ + \\sum_ {k \\in S _ {M}} \\gamma_ {k} ^ {M} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) + \\sum_ {k \\in S _ {A}} \\gamma_ {k} ^ {A} \\nabla F _ {k} \\left(v _ {A}\\right) + \\sum_ {k \\in S _ {I}} \\gamma_ {k} ^ {I} \\nabla F _ {k} \\left(v _ {I}\\right) \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 422, + 784, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sigma_{M},\\sigma_{A}$ and $\\sigma_I$ map $V\\to S_M,S_A,S_I$ , the client sets who use multi-modal, uni- $A$ , and uni- $I$ data for training respectively, and $S_{M}\\cap S_{A} = S_{A}\\cap S_{I} = S_{M}\\cap S_{I} = \\emptyset$ . $\\gamma_k^M$ , $\\gamma_k^A$ and $\\gamma_k^I$ have the similar meaning as $\\gamma_{k}$ in Eq. (2).", + "bbox": [ + 214, + 507, + 785, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since modality $I$ is weak here, we omit the uni- $A$ clients as the multi-modal gradient is dominated by modality $A$ [28], which means we do not need to select uni- $A$ clients for its enhancement. Then, bring Eq. (9) to Eq. (10) and follow the operations from Eq. (2) to Eq. (3), we can obtain:", + "bbox": [ + 212, + 551, + 785, + 613 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) - \\gamma_ {i} ^ {M} \\nabla F _ {i} \\left(v _ {A}, v _ {I}\\right) - \\gamma_ {j} ^ {I} \\nabla F _ {j} \\left(v _ {I}\\right) \\right\\| \\\\ = \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\begin{array}{c} \\nabla \\mathcal {L} _ {C E} ^ {k} (v _ {A}, v _ {I}) + \\nabla \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} (v _ {I}) - \\nabla \\mathcal {L} _ {C E} ^ {i} (v _ {A}, v _ {I}) \\\\ - \\nabla \\beta^ {i} \\mathcal {L} _ {M E} ^ {i} (v _ {I}) - \\nabla \\mathcal {L} _ {C E} ^ {j} (v _ {I}) - \\nabla \\beta^ {j} \\mathcal {L} _ {M E} ^ {j} (v _ {I}) \\end{array} \\right\\| \\\\ \\leqslant \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}} \\left\\| \\nabla \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) - \\nabla \\mathcal {L} _ {C E} ^ {i} \\left(v _ {A}, v _ {I}\\right) \\right\\| \\tag {11} \\\\ + \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\begin{array}{c} \\nabla \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} (v _ {I}) - \\nabla \\beta^ {i} \\mathcal {L} _ {M E} ^ {i} (v _ {I}) \\\\ - \\nabla \\mathcal {L} _ {C E} ^ {j} (v _ {I}) - \\nabla \\beta^ {j} \\mathcal {L} _ {M E} ^ {j} (v _ {I}) \\end{array} \\right\\| \\\\ \\triangleq G \\left(S _ {M}\\right) + G \\left(S _ {M} \\cup S _ {I}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 623, + 784, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The right-hand side of the first equation is the modality selection formula that aims to select a group of multi-modal clients $S_{M}$ and uni-modal clients", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 477, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: BMSFed." + ], + "code_body": "Input: Input data $\\mathcal{D}_i = \\{\\pmb {X}_i^A,\\pmb {X}_i^I,\\pmb {y}\\} ,i\\in [N],$ initial model $\\theta$ hyper-parameters $\\chi$ global communication epochs $E$ $e = 1$ while $e < E$ do if $e = 1$ then Send $\\theta$ to all clients; Perform one-step local update for gradients, prototypes and $\\rho_I^k$ Aggregate global prototypes and $\\rho_{I}$ else $\\begin{array}{rl} & {\\mathrm{Aggrecate~global~model~}\\theta ,\\mathrm{prototypes~}c^{G}\\mathrm{~and~}\\rho_{I};}\\\\ & {\\mathrm{S e l e c t~multi - mod a l i t y~for~}S_{M}\\mathrm{~a n d~u n i - m o d a l i t y~for~}S_{I}\\mathrm{~(o r~}S_{A})\\mathrm{~u s i n g}}\\\\ & {\\mathrm{E q s.~(12)~a n d~(13);}}\\\\ & {\\mathrm{S e n d~}\\theta ,c^{G}\\mathrm{~a n d~}\\rho_{I}\\mathrm{~t o~s e l e c t e d~c l i e n s;}}\\\\ & {\\mathrm{f o r e a c h~c l i e n t~i n s e l e c t e d~c l i e n s i n~p a r a l l e l d o}}\\\\ & {\\mathrm{P r o p e r m~m u l t i - m o d a l ~l a n e a i n g~i n~}S_{M}\\mathrm{~a n d~u n i - m o d a l ~l e a n i n g~i n~}S_{I}\\mathrm{~(o r}}\\\\ & {\\mathrm{S}_{A})\\mathrm{b y w i t h~E q s.~(8)~a n d~(9);}}\\\\ & {\\mathrm{S e n d~g r a d i e n t s,~l o c a l~p r o t e t y p e s a n d~}\\rho_{I}^{k}\\mathrm{~t o~s e r v e r};} \\end{array}$", + "bbox": [ + 228, + 166, + 759, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$S_I$ to approximate the aggregated gradients from all clients. However, the joint selection for $S_M$ and $S_I$ is a complex joint optimization problem. Therefore, we decouple this objective into two submodular functions $G(S_M)$ and $G(S_M \\cup S_I)$ according to triangle inequality, while the full gradient approximation is divided into two parts: the first part uses selected multi-modal CE gradient to fit fully multi-modal CE gradient aggregation, and the second part approximates the fully multi-modal ME gradient aggregation via selected uni-modal CE gradient and both selected multi- and uni-modal ME gradient. The modality-level gradient decoupling converts the complex joint selection to two simply separated selection problems.", + "bbox": [ + 212, + 435, + 787, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although we can solve the two submodular functions with the stochastic greedy algorithm [24], there are still two issues: (1) the selected client according to $G(S_M \\cup S_I)$ should be specified whether it is uni-modal client or multi-modal client; (2) the separated selection strategy pays less attention to the global modal bias. To address the two problems, we perform the stochastic greedy algorithm for two submodular functions in parallel and propose a simple yet effective conflict resolution strategy to ensure $S_M \\cap S_I = \\emptyset$ as well as, more importantly, balance the learning of different modalities on global model:", + "bbox": [ + 212, + 587, + 787, + 708 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS _ {M} \\leftarrow S _ {M} \\cup k _ {1} ^ {*}, k _ {1} ^ {*} \\in \\underset {k \\in \\operatorname {r a n d} (V \\backslash S _ {M} \\backslash S _ {I}, \\mathrm {s})} {\\arg \\max } \\left[ \\bar {G} \\left(S _ {M}\\right) - \\bar {G} \\left(\\{k \\} \\cup S _ {M}\\right) \\right] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 254, + 713, + 785, + 741 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} i f k _ {1} ^ {*} = k _ {2} ^ {*}, S _ {M} \\cup k _ {2} ^ {*}; \\\\ i f k _ {1} ^ {*} \\neq k _ {2} ^ {*}, \\left\\{ \\begin{array}{l} S _ {I} \\cup k _ {2} ^ {*}, i f \\rho_ {I} ^ {k} > \\chi \\\\ S _ {M} \\cup k _ {2} ^ {*}, i f \\rho_ {I} ^ {k} \\leqslant \\chi \\end{array} \\right. \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 756, + 785, + 813 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nk _ {2} ^ {*} \\in \\operatorname * {a r g m a x} _ {k \\in \\operatorname {r a n d} (V \\setminus S _ {M} \\setminus S _ {I}, \\mathrm {s})} \\left[ \\bar {G} \\left(S _ {M} \\cup S _ {I}\\right) - \\bar {G} \\left(\\{k \\} \\cup S _ {M} \\cup S _ {I}\\right) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 815, + 707, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For every selection, we randomly sample a subset of clients $s$ . A multi-modal client $k_{1}^{*}$ is selected from $s$ according to $G(S_{M})$ while $k_{2}^{*}$ is also selected from $s$ according to $G(S_{M} \\cup S_{I})$ . $k_{1}^{*} = k_{2}^{*}$ means using multi-modal data from this client can contribute most to the global model. When $k_{1}^{*} \\neq k_{2}^{*}$ , we allocate it to uni-modal or multi-modal client according to its local imbalance ratio: if it is severely imbalanced, we use its uni-weak-modal data for training and aggregation to alleviate the global modal bias, otherwise we believe that training with its multi-modal data contributes more than uni-modal data. $\\chi$ is a hyper-parameter.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Discussion. (1) Overcoming the modal bias in our method are twofold: the ME loss alleviates imbalance at local side and the selected uni-modal clients further promote balanced learning of global model. Meanwhile, the diversity coming from two submodular functions ensures the representative information for the global model. (2) We assume $I$ is the weak modality above while in practice, we can determine the weak modality before modality selection via the aggregated global imbalance ratio $\\rho_{I} = \\frac{1}{\\sum_{k=1}^{N} n_{k}} \\sum_{k=1}^{N} \\rho_{I}^{k} \\cdot n_{k}$ . Overall, the pseudo-code of BMSFed is provided in Algorithm 1. (3) Only the gradients, prototypes and $\\rho_{I}^{k}$ participate in communication, so there is no privacy issue and similar communication overheads as in traditional FL.", + "bbox": [ + 212, + 268, + 787, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 458, + 356, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and baselines", + "text_level": 1, + "bbox": [ + 215, + 500, + 449, + 515 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on four datasets: (1) CREMA-D [4] is an audio-visual dataset for emotion recognition task with total six categories for emotional states. (2) AVE [33] is an audio-visual dataset for event localization with 28 event classes, and here we use it to construct a labeled multimodal classification dataset following [11]. (3) Colored-and-gray MNIST (CG-MNIST) [17] is a synthetic dataset based on MNIST [18] with gray-scale and monochromatic images as two modalities, following [38]. (4) ModelNet40 is one of the Princeton ModelNet datasets [39] with 3D objects of 40 categories. The front and back [31] views are considered as two modalities, following [5]. Details of these four datasets are in the Appendix.", + "bbox": [ + 212, + 535, + 787, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We choose eight baselines for comparison from four categories: (1) three uni-modal FL methods designed for statistical heterogeneity are extended to multi-modal scenarios: FedAvg [23], FedProx [21] and FedProto [32]. (2) Integrating OGM-GE [28] and PMR [11], the solutions for modality imbalance, with FedAvg forms two MFL methods: FedOGM and FedPMR. (3) Two client selection method: Power-of-choice (pow-d) [6] and DivFL [3], evolved from its uni-modal version directly. (4) One MFL method, FedMSplit [5], especially designed for modality incongruity. Compared with these baselines, we demonstrate that an elaborate modality selection strategy is essential to realize comprehensive information exploitation in MFL.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 478, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/70d9710578f202fcb8609f7f2b60aaf80944a7d16374a053242d212d1a40e42a.jpg", + "table_caption": [ + "Table 2: Comparison results on four datasets. The metric is the top-1 accuracy (\\%). The best is in **bold**, and the second best is **underlined**. Our method achieves significant improvement on both IID and non-IID scenarios." + ], + "table_footnote": [], + "table_body": "
DatasetCREMA-DAVECG-MNISTModelNet40
MethodIIDnon-IIDIIDnon-IIDIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.742.341.787.286.5
FedProx51.049.062.659.942.943.686.987.1
FedProto58.754.061.758.851.551.487.587.2
FedOGM56.956.462.859.357.253.087.687.0
FedPMR55.555.163.161.666.163.387.687.7
pow-d50.550.762.560.041.240.386.886.2
DivFL51.750.863.359.643.042.186.586.4
FedMSplit52.451.662.460.843.550.987.587.4
BMSFed64.561.664.762.170.266.788.787.5
", + "bbox": [ + 217, + 198, + 792, + 359 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental settings", + "text_level": 1, + "bbox": [ + 215, + 388, + 441, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For CREMA-D, AVE and ModelNet40, we use ResNet18 [12] as the backbone for audio, visual and flow modalities. Audio data is converted to a spectrogram of size 257x299 for CREMA-D and $257 \\times 1,004$ for AVE. We randomly choose 3 frames and 4 frames to build image training sets for CREMA-D and AVE respectively. For CG-MNIST, we build a neural network with 4 convolution layers and 1 average pool layer as the encoder, following the setting as in [11]. We choose the simple yet effective fusion method, concatenation [27], to build fusion classifier for all the datasets. We set 20 clients for CREMA-D, AVE and ModelNet40 while the number for CG-MNIST is 30. 5 clients are selected in each communication round for CREMA-D, AVE, ModelNet40 and 6 for CG-MNIST. For IID setting, training data is uniformly distributed to all clients. For non-IID scenarios, we use Dirichlet distribution [13] $Dir(\\alpha)$ to split data ( $\\alpha = 3$ for CREMA-D, AVE, ModelNet40, $\\alpha = 2$ for CG-MNIST). The optimizer is SGD [30] for all datasets. Learning rate is initialized at 1e-3 or 1e-2 for CEAMA-D, AVE and ModelNet40 or CG-MNIST and becomes 1e-4 or 1e-3 in the later training stage. The hyper-parameter $\\chi$ is set to 1.2-2.5 according to datasets and settings. To complete stochastic greedy algorithm for Eqs. (12) and (13), we use the gradients from the selected clients at current round to update part of the similarity matrix, which is named \"no-overheads\" in [3]. Except for pow-d, DivFL and BMSFed, other baselines select clients randomly. Each client has two modal data by default. We do all experiments on a workstation with an RTX 3090 GPU, a 3.9-GHZ Intel Core i9-12900K CPU and 64GB of RAM.", + "bbox": [ + 215, + 419, + 787, + 751 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison with baselines", + "text_level": 1, + "bbox": [ + 215, + 779, + 480, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "BMSFed effectively improves the performance. As demonstrated in Tab. 2, our BMSFed achieves the best results on the four datasets under both IID and", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7274319b65aa763598ff2d5925193705ec461b62a41061d3a100987b44781a05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 143, + 477, + 258 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9b3d7afbe8bc08c2edd2791e64d64337f539822ddbbe88cc01e4b0616ba75aea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 143, + 733, + 258 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3f9ee423b51e7bc5c78e5f48750c24fdf3f3cc47347e01cf107b65d20afd33c1.jpg", + "image_caption": [ + "(a) CREMA-D under IID", + "(c) CREMA-D under non-HID", + "Fig. 2: Test accuracy of BMSFed compared with other baselines on CREMA-D and AVE. BMSFed converges to more accurate solutions than all baselines." + ], + "image_footnote": [], + "bbox": [ + 243, + 273, + 473, + 388 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8247a651b91c8a250e5f9e8347c93ced39ef28b450d87b76185fff7957efb5fd.jpg", + "image_caption": [ + "(b) AVE under IID", + "(d) AVE under non-IID" + ], + "image_footnote": [], + "bbox": [ + 503, + 273, + 733, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "non-IID settings (by up to $5.8\\%$ on CREMA-D). Client sampling here (powd and DivFL) cannot fully exploit information for all modalities, making its improvement limited or even worse than FedAvg in CG-MNIST. Although FedOGM and FedPMR accomplish modest improvement because of their ability to alleviate modality imbalance, they are not as good as BMSFed since they do not consider the overall performance of the global model. Traditional uni-modal FL methods for statistical heterogeneity (e.g. FedProx) and MFL method for modality incongruity (FedMSplit) only obtain slight improvement. We also illustrate the trend of test accuracy versus the number of communication rounds on CREMA-D and AVE in Fig. 2. BMSFed can realize comparable or even faster convergence speeds in CREMA-D and AVE.", + "bbox": [ + 212, + 484, + 787, + 652 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "BMSFed exploits all modalities comprehensively. To show the effect of our method on addressing modal bias, we report the performance of each modality on CREMA-D and AVE as shown in Tab. 3. The uni-modal performance evaluation follows [11]: a sample is classified into the class corresponding to its nearest prototype. It is clear that BMSFed could considerably improve the performance of weak modality (visual) and mitigate the modality-level bias. Besides, compared with randomly modality abandoning, which significantly reduces audio performance as illustrated in Tab. 1, BMSFed achieves comparable audio performance with other baselines. Although FedProto, FedOGM and FedPMR also alleviate the imbalance, they mainly focus on local optimization, resulting in the performance gap between them and our BMSFed on the aggregated model, which further indicates that in MFL, it is important to take both local opti", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 478, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a57ca85f65fb0714fb6281faa536afdfb2da60e50bbd33c90ccd56c9bde76091.jpg", + "table_caption": [ + "Table 3: The uni-modal performance comparison on CREMA-D and AVE. The metric is the top-1 accuracy (\\%). Our method achieves the best uni-visual (the weak modality) performance on all settings and maintains the comparable uni-audio performance." + ], + "table_footnote": [], + "table_body": "
DatasetCREMA-DAVE
SettingIIDnon-IIDIIDnon-IID
MethodAVAVAVAV
FedAvg51.220.650.720.261.126.861.426.4
FedProx51.320.250.122.060.427.161.226.9
FedProto50.235.348.639.155.736.859.732.8
FedOGM50.535.748.830.258.728.859.429.4
FedPMR51.538.750.135.961.739.661.735.3
pow-d51.520.451.618.861.926.960.127.1
DivFL52.321.152.122.762.725.361.626.3
FedMSplit52.021.850.821.661.326.962.328.7
BMSFed51.041.949.341.459.740.260.238.6
", + "bbox": [ + 243, + 198, + 754, + 373 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fc3afd223b8a37577ab080260dc2797e123b8e99271b24392990abfbe94de898.jpg", + "table_caption": [ + "Table 4: Ablation study. 'BMSFed-local' uses local prototypes rather than global prototypes for ME loss." + ], + "table_footnote": [], + "table_body": "
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.7
DivFL51.750.863.359.6
FedAvg-0.252.450.163.461.1
FedAvg-0.555.755.160.759.4
FedAvg-0.861.258.158.558.7
FedAvg+ME55.854.562.860.7
DivFL+ME57.155.663.061.1
BMSFed-local63.760.363.460.1
BMSFed64.561.664.762.1
", + "bbox": [ + 287, + 425, + 712, + 584 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "mization for each modality and the overall performance for global model into consideration simultaneously.", + "bbox": [ + 215, + 612, + 785, + 643 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 664, + 388, + 680 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of each component. Tab. 4 studies the effect of each BMSFed component. Applying ME loss Eq. (7) on random sampling $(\\mathrm{FedAvg} + \\mathrm{ME})$ and the well-designed client selection $(\\mathrm{DivFL} + \\mathrm{ME})$ surpasses their vanilla strategies $(\\mathrm{FedAvg}, \\mathrm{DivFL})$ by a large margin, demonstrating its effectiveness on local enhancement. Comparing BMSFed (64.5% on IID CREMA-D) with 'DivFL+ME' (57.1% on the same setting) also denotes the necessity of balancing different modalities considering the global model via modality selection. To show the importance of aligning feature spaces of weak modality, we replace the global prototypes with local prototypes (BMSFed-local). Global alignment achieves notable improvement (by up to 2% on non-IID AVE). The performance improvement", + "bbox": [ + 215, + 688, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/415f2b607fb3f72dca6976bbcca677abe4bd373b6aee558d7cc1dc0a31073c83.jpg", + "image_caption": [ + "(a) CREMA-D under IID" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 390, + 239 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dad87f0c4c177161d2c4c78b1c41cc82d49aa19581bf69f9e6edc4c98b807eab.jpg", + "image_caption": [ + "(b) CREMA-D under IID" + ], + "image_footnote": [], + "bbox": [ + 406, + 145, + 576, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f823299f830c257182854f75cfb233c70db6a5d6161f8a3d1929ab37dcbdd28b.jpg", + "image_caption": [ + "(c) CG-MNIST under IID" + ], + "image_footnote": [], + "bbox": [ + 594, + 143, + 767, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3ebb8aae80912edba076b03f45e9c62fa05be448518a3a27122e06eaeefc3413.jpg", + "image_caption": [ + "Fig. 3: Robustness validation on data size, local epoch and client number. Our BMSFed consistently outperforms baseline (FedAvg) under various scenarios.", + "Fig. 4: Proportional change of audio and visual respectively and the curve of global imbalance ratio during training on CREMA-D under IID setting." + ], + "image_footnote": [], + "bbox": [ + 359, + 316, + 643, + 452 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "compared with 'FedAvg-0.2,0.5,0.8' exhibits that randomly sampling modalities does not always lead to improvement and further demonstrates the need of meticulously selecting modalities for information exploitation.", + "bbox": [ + 212, + 513, + 784, + 559 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Robustness test. To verify the robustness of our method, we vary three key hyperparameters to build various scenarios: (1) change the data size $|\\mathcal{D}_i|$ to allow each client to hold a small amount of data, (2) set different local training epochs, and (3) vary the total client number $N$ . As illustrated in Figs. 3a and 3b, our BMSFed consistently outperforms baseline (FedAvg) under various scenarios. More data as well as more local training epochs can bring further improvements to our method, indicating that exploiting the weak modality need more training efforts. Based on the results in Fig. 3c, our approach can also be generalised to clients with larger scales.", + "bbox": [ + 212, + 564, + 787, + 699 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The relationship between modality selection and imbalance degree. In Eqs. (12) and (13), we use a conflict resolution strategy based on local imbalance ratio to realize balanced modality selection. We visualize the proportions of audio and visual modalities selected in each round and the global imbalance ratio. It is clear from Fig. 4 that audio is the dominant modality (imbalance ratio is always greater than 1) and modality imbalance is gradually alleviated as training progresses (global imbalance ratio shows a downward trend). In addition, the proportion of selected visual modality follows the same trend (larger in the early stage of training and becomes smaller later), implying the rationality of our", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 478, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c6ca3d171ef9610b6c518f6857c57e5296518ac806c97c4d8b52f5928be35af8.jpg", + "table_caption": [ + "Table 5: Performance on CREMA-D and AVE with modality incongruity. $50\\%$ of clients have all modal data and $50\\%$ of clients only retain data with a single modality (audio or visual)." + ], + "table_footnote": [], + "table_body": "
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg55.755.160.758.7
FedProx56.856.061.258.5
FedProto58.757.061.359.7
FedOGM58.657.460.158.5
FedPMR56.455.561.960.3
DivFL57.455.661.158.6
FedMSplit58.956.961.760.0
BMSFed62.459.863.560.9
", + "bbox": [ + 297, + 198, + 699, + 357 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "selection strategy based on imbalance ratio and its effectiveness on mitigating bias.", + "bbox": [ + 215, + 388, + 784, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness on modality incongruity scenario. All above experiments assume that each client initially has complete modal data. Here, we build the stimulation of modality incongruity scenario, in which half of the clients have data with two modalities, and the other half only have data in one modality: random audio or visual. The results are shown in Tab. 5 (The results of pow-d is not available because it is not applicable to this scenario). Our BMSFed still makes impressive improvement compared with all other baselines (by up to $3.5\\%$ on IID CREMA-D), illustrating the good generalization ability of our method in different scenarios. It is worth mentioning that FedMSplit performs better than before because it is specifically designed for modality incongruity.", + "bbox": [ + 215, + 420, + 785, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 603, + 357, + 618 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we analyze traditional client selections and find their ineffectiveness in MFL. We further reveal that there exists strong modality-level bias due to the modality imbalance during the training iterations and uni-modal training on some clients may contribute more to the global model than multi-modal training. To address this issue, we propose the balanced modality selection scheme for MFL (BMSFed) with modality-level gradient decoupling to release the potential of all modalities and maximize the gradient diversities to improve global aggregation. We also introduce a modal enhancement loss to optimize the local update process. Our method does not introduce additional local training costs and communication overheads compared with previous methods. Extensive experiments on four datasets demonstrate the superiority of our method in performance and applicability under different modal combinations, data distributions and modality incongruity scenarios.", + "bbox": [ + 215, + 643, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Our work was supported by two grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. PolyU15222621, PolyU15225023) and the funding from National Natural Science Foundation of China under grants 62302184.", + "bbox": [ + 212, + 172, + 787, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 270 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abavisani, M., Wu, L., Hu, S., Tetreault, J., Jaimes, A.: Multimodal categorization of crisis events in social media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14679-14689 (2020)", + "2. Alfasly, S., Lu, J., Xu, C., Zou, Y.: Learnable irrelevant modality dropout for multimodal action recognition on modality-specific annotated videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20208-20217 (2022)", + "3. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (2022)", + "4. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing 5(4), 377-390 (2014)", + "5. Chen, J., Zhang, A.: Fedmsplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 87-96 (2022)", + "6. Cho, Y.J., Wang, J., Joshi, G.: Client selection in federated learning: Convergence analysis and power-of-choice selection strategies. arXiv preprint arXiv:2010.01243 (2020)", + "7. Cornuejols, G., Fisher, M., Nemhauser, G.L.: On the uncapacitated location problem. In: Annals of Discrete Mathematics, vol. 1, pp. 163-177. Elsevier (1977)", + "8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023)", + "9. Deng, Y., Lyu, F., Ren, J., Wu, H., Zhou, Y., Zhang, Y., Shen, X.: Auction: Automated and quality-aware client selection framework for efficient federated learning. IEEE Transactions on Parallel and Distributed Systems 33(8), 1996-2009 (2021)", + "0. Duan, J.h., Li, W., Zou, D., Li, R., Lu, S.: Federated learning with data-agnostic distribution fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8074–8083 (2023)", + "11. Fan, Y., Xu, W., Wang, H., Wang, J., Guo, S.: Pmr: Prototypical modal rebalance for multimodal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20029-20038 (2023)", + "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "13. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019)", + "14. Huang, Y., Lin, J., Zhou, C., Yang, H., Huang, L.: Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In: International Conference on Machine Learning. pp. 9226-9259. PMLR (2022)" + ], + "bbox": [ + 218, + 282, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 478, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Imteaj, A., Thakker, U., Wang, S., Li, J., Amini, M.H.: A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal 9(1), 1-24 (2021)", + "16. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020)", + "17. Kim, B., Kim, H., Kim, K., Kim, S., Kim, J.: Learning not to learn: Training deep neural networks with biased data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9012-9020 (2019)", + "18. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)", + "19. Li, H., Li, X., Hu, P., Lei, Y., Li, C., Zhou, Y.: Boosting multi-modal model performance with adaptive gradient modulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22214-22224 (2023)", + "20. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020)", + "21. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)", + "22. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189 (2019)", + "23. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)", + "24. Mirzasoleiman, B., Badanidiyuru, A., Karbasi, A., Vondrák, J., Krause, A.: Lazier than lazy greedy. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 29 (2015)", + "25. Nemhauser, G.L., Wolsey, L.A., Fisher, M.L.: An analysis of approximations for maximizing submodular set functions--i. Mathematical programming 14, 265-294 (1978)", + "26. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020)", + "27. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European conference on computer vision (ECCV). pp. 631-648 (2018)", + "28. Peng, X., Wei, Y., Deng, A., Wang, D., Hu, D.: Balanced multimodal learning via on-the-fly gradient modulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8238-8247 (2022)", + "29. Reisizadeh, A., Tziotis, I., Hassani, H., Mokhtari, A., Pedarsani, R.: Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity. IEEE Journal on Selected Areas in Information Theory 3(2), 197-205 (2022)", + "30. Robbins, H., Monro, S.: A stochastic approximation method. The annals of mathematical statistics pp. 400-407 (1951)", + "31. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)", + "32. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Fan et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 247-263 (2018)", + "34. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023)", + "35. Wang, H., Qu, Z., Guo, S., Wang, N., Li, R., Zhuang, W.: Losp: Overlap synchronization parallel with local compensation for fast distributed training. IEEE Journal on Selected Areas in Communications 39(8), 2541-2557 (2021)", + "36. Wang, H., Xu, W., Fan, Y., Li, R., Zhou, P.: Aocc-fl: Federated learning with aligned overlapping via calibrated compensation. In: IEEE INFOCOM 2023-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2023)", + "37. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12695-12705 (2020)", + "38. Wu, N., Jastrzebski, S., Cho, K., Geras, K.J.: Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. In: International Conference on Machine Learning. pp. 24043-24055. PMLR (2022)", + "39. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)", + "40. Xiao, F., Lee, Y.J., Grauman, K., Malik, J., Feichtenhofer, C.: Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740 (2020)", + "41. Xiong, B., Yang, X., Qi, F., Xu, C.: A unified framework for multi-modal federated learning. Neurocomputing 480, 110-118 (2022)", + "42. Xu, J., Tong, X., Huang, S.L.: Personalized federated learning with feature alignment and classifier collaboration. arXiv preprint arXiv:2306.11867 (2023)", + "43. Xu, J., Wang, H.: Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective. IEEE Transactions on Wireless Communications 20(2), 1188-1200 (2020)", + "44. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023)", + "45. Zhao, Y., Barnaghi, P., Haddadi, H.: Multimodal federated learning on IoT data. In: 2022 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation (IoTDI). pp. 43-54. IEEE (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 632 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Balanced Modality Selection for MFL", + "bbox": [ + 478, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_model.json b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..014f32d0138a5a18748dcdb2d9e448ae5da09f41 --- /dev/null +++ b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_model.json @@ -0,0 +1,2302 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.213, + 0.78, + 0.245 + ], + "angle": 0, + "content": "Yunfeng Fan\\(^{1}\\), Wenchao Xu\\(^{1(\\boxdot)}\\), Haozhao Wang\\(^{2}\\), Fushuo Huo\\(^{1}\\), Jinyu Chen\\(^{1}\\), and Song Guo\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.255, + 0.774, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Computing, The Hong Kong Polytechnic University, Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.27, + 0.777, + 0.297 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Computer Science and Technology, Huazhong University of Science and Technology, Hubei, China" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.297, + 0.779, + 0.325 + ], + "angle": 0, + "content": "3 Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.255, + 0.779, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.327, + 0.767, + 0.367 + ], + "angle": 0, + "content": "{yunfeng.fan, fushuo.huo, jinyu.chen}@connect.polyu.hk \nwenchao.xu@polyu.edu.hk, hz_wang@hust.edu.cn, songguo@cse.ust.hk \nhttps://github.com/fanyunfeng-bit/Balanced-Modality-Selection-in-MFL" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.41, + 0.744, + 0.659 + ], + "angle": 0, + "content": "Abstract. Selecting proper clients to participate in each federated learning (FL) round is critical to effectively harness a broad range of distributed data. Existing client selection methods simply consider the mining of distributed uni-modal data, yet, their effectiveness may diminish in multi-modal FL (MFL) as the modality imbalance problem not only impedes the collaborative local training but also leads to a severe global modality-level bias. We empirically reveal that local training with a certain single modality may contribute more to the global model than training with all local modalities. To effectively exploit the distributed multiple modalities, we propose a novel Balanced Modality Selection framework for MFL (BMSFed) to overcome the modal bias. On the one hand, we introduce a modal enhancement loss during local training to alleviate local imbalance based on the aggregated global prototypes. On the other hand, we propose the modality selection aiming to select subsets of local modalities with great diversity and achieving global modal balance simultaneously. Our extensive experiments on audio-visual, colored-gray, and front-back datasets showcase the superiority of BMSFed over baselines and its effectiveness in multi-modal data exploitation." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.673, + 0.752, + 0.701 + ], + "angle": 0, + "content": "Keywords: Multi-modal federated learning \\(\\cdot\\) modality imbalance \\(\\cdot\\) modality selection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.73, + 0.377, + 0.746 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Federated learning (FL) [23] aims to collaboratively learn data that has been collected by, and resides on, a number of remote devices or servers. FL stands to develop top-performing models by aggregating knowledge from numerous edge clients [8,36], which relies on the iterative interaction among participating clients and the server. However, comprehensively employing the information from all" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.2 + ], + "angle": 0, + "content": "Table 1: Performance of various client selection methods in MFL under IID setting. A and V denote uni-audio and uni-visual while A-V means the multi-modal result. 'Local' represents that a client is trained based its local data without aggregation. A strong modal bias of global model exists on the two datasets." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.213, + 0.747, + 0.359 + ], + "angle": 0, + "content": "
DatasetCREMA-D [4]AVE [33]
MethodAVA-VAVA-V
Local41.920.439.633.416.735.2
FedAvg51.220.650.761.126.862.2
pow-d [6]51.520.450.561.926.962.5
DivFL [3]52.321.151.762.725.363.3
FedAvg-0.250.628.652.460.629.663.4
FedAvg-0.550.534.655.758.730.060.7
FedAvg-0.848.150.961.256.431.858.5
BMSFed51.041.964.559.740.264.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.784, + 0.41 + ], + "angle": 0, + "content": "clients can be exceptionally difficult due to the client heterogeneity and resource limitations [10, 29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Random sampling [22, 34] from available clients has been widely used in FL to satisfy some practical restrictions, e.g., limited communication bandwidth [26, 35] and computing capacities [15]. To improve the information exploitation of all clients, extensive research has been conducted on effective client selection strategies [9, 43]. Despite the success of traditional client selection methods in uni-modal FL, their effectiveness diminishes when dealing with clients with multi-modal data as the inter-modal interactions during the MFL training are neglected. According to [14, 37], there may exist inconsistent learning paces for different modalities in multi-modal joint training, i.e., modality imbalance, which not only impedes the collaborative local training but also leads to a severe modal bias for global model in MFL. As illustrated in Tab. 1, audio modality significantly outperforms visual modality in CREMA-D and AVE datasets during local training and the aggregated model still suffers from it. However, two well-designed client selection methods (pow-d [6] and DivFL [3]) only obtain severely limited improvement over random sampling (FedAvg) on the multi-modal global model. We can see that client selection methods achieve the best uni-audio performance while uni-visual performance even drops sometimes, which means existing client selection scheme heavily relies on the better modality, while ignoring the importance of improving weak modalities that also has potential for global model aggregation. Based on the above analysis, a pivotal question arises: Can we design a new selection scheme in MFL that can overcome the modal bias and exploit each modality comprehensively?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To answer this question, we investigate the interactions between different modalities via randomly discarding the data from one modality (audio or visual) on part of clients, which is inspired from modality dropout [2, 40] that drops a specific modal data during training for regularization. The results are shown in rows 7-9 in Tab. 1, where \\(-x\\) denotes randomly discarding a modality on a client with probability \\(x\\). We can see that dropping with a certain prob" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.479, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.782, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 1: Left: Traditional client selection in FL aims to sample a client subset in each round while our modality selection considers each local modality as the sampling unit. Right: The paradigm of BMSFed with four clients. The global prototypes are used to enhance the weak modality during local update. Only networks corresponding to the selected modalities will be uploaded to the server for aggregation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.789, + 0.542 + ], + "angle": 0, + "content": "ability can improve the global multi-modal performance on both datasets (e.g., FedAvg-0.2), and the main reason comes from the dramatic improvement of visual modality. This phenomenon suggests that performing uni-modal training can unleash its potential without being inhibited by another modality, and uni-modal local training may contribute more to the global model than multi-modal training on some clients. As the dropping ratio increases, although the visual modality still improves, multi-modal performance may decline as the audio modality declines as shown in columns 5 and 7, suggesting that we should carefully control which modalities on each client should be involved in training and aggregation to make the contribution most to the global model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.789, + 0.771 + ], + "angle": 0, + "content": "According to the above investigations, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to mitigating the modal bias and comprehensively exploit the diverse information from all modalities. Specifically, instead of selecting a subset of clients, we treat each modality on the local side as a selection unit, as demonstrated in Fig. 1. Our BMSFed mainly contains two parts: Firstly, we intend to alleviate the local modality imbalance by introducing a modal enhancement loss based on aggregated global prototypes to promote the performance of weak modality. Secondly, we complete the modality selection by building two separated submodular functions for selecting multi-modal clients (training with multi-modal data) and uni-modal clients (training with selected uni-modal data) respectively. Inspired from [3], the criterion is to select modalities that are most representative on the gradients while also alleviate the global modal bias. A simple yet effective conflict resolution strategy is devised to ensure the validity of modality selection and keep modal balance on global model simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.774, + 0.703, + 0.789 + ], + "angle": 0, + "content": "The main contributions of the paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "1. We empirically analyze the modality imbalance problem in MFL and reveal that uni-modal training on some clients may contribute more to the global model than multi-modal training." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "2. Based on the analysis, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to comprehensively exploit all modalities via a modal enhancement loss and representative modality selection to overcome the global modal bias." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.207, + 0.785, + 0.252 + ], + "angle": 0, + "content": "3. We conduct comprehensive experiments on audio-visual, colored-gray, and front-back datasets, and considering the statistical heterogeneity and modality incongruity problems in MFL, to validate the superiority of our BMSFed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.561, + 0.291 + ], + "angle": 0, + "content": "2 Background and Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.297, + 0.522, + 0.313 + ], + "angle": 0, + "content": "2.1 Multi-modal federated learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.784, + 0.402 + ], + "angle": 0, + "content": "In MFL, each client has one or multiple modalities of data. Without loss of generality, we consider two input modalities, which are denoted by \\( A \\) and \\( I \\), respectively in MFL. There are a set \\( V \\) of \\( N \\) clients, \\( V = [N] \\), respectively owning datasets \\( \\mathcal{D}_i = \\left\\{\\mathbf{X}_i^A, \\mathbf{X}_i^I, \\mathbf{y}\\right\\} \\), \\( i \\in [N] \\). A typical federated learning objective is the average of each client's local loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\min f (\\theta) = \\sum_ {k = 1} ^ {N} p _ {k} F _ {k} (\\theta) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.784, + 0.508 + ], + "angle": 0, + "content": "where \\(\\theta = \\{\\theta^A, \\theta^I, \\omega\\}\\) denotes the model parameters. \\(\\theta^A\\) and \\(\\theta^I\\) represent the encoder parameters of modality \\(A\\) and \\(I\\). \\(\\omega\\) is the parameter of fusion classifier. \\(p_k\\) is a pre-defined weight. \\(F_k\\) is each client's local loss (cross entropy (CE) loss for classification task in this paper)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.675 + ], + "angle": 0, + "content": "Statistical heterogeneity [16,20] is a widely concerned challenge in uni-modal FL. To tackle this issue, FedProx [21] uses a proximal term to stabilize model aggregation. FedProto [32] shares class prototypes to regularize the learning of local models. In MFL, modality incongruity [44,45] (clients consist of different modalities combinations), as well as statistical heterogeneity [41], are all considered. Yu et al. [44] propose CreamFL to align the representations between different clients and different modalities via communicating knowledge on a public dataset. Chen et al. [5] introduce FedMSplit to split local models into several components and aggregate them by their correlations. However, they still focus on the heterogeneity, but ignore the interaction between private data of different modalities, which limits their information exploitation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.688, + 0.603, + 0.702 + ], + "angle": 0, + "content": "2.2 Client selection and submodular function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.784, + 0.779 + ], + "angle": 0, + "content": "Client selection [9,43] is a critical issue for FL especially when the communication cost with all devices is prohibitively high, which has been extensively studied in uni-modal FL. Cho et al. [6] propose Power-of-Choice to select clients with targets local loss. Balakrishnan et al. [3] propose to select a subset of clients with great diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Diverse client selection via submodularity. Maximizing a submodular function is reported to improve the diversity and reduce the redundancy of a subset. This property makes it appropriate for client selection in FL. If a function \\( F \\) is submodular, it should satisfy: given a finite ground set \\( V \\) of size" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.478, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.784, + 0.238 + ], + "angle": 0, + "content": "\\(N\\), \\(F(A \\cup \\{v\\}) - F(A) \\geqslant F(B \\cup \\{v\\}) - F(B)\\), for any \\(A \\subseteq B \\subseteq V\\) and \\(v \\in V \\setminus B\\). The marginal utility of an element \\(v\\) w.r.t. a subset \\(A\\) is denoted as \\(F(v|A) = F(A \\cup \\{v\\}) - F(A)\\), which can represent the importance of \\(v\\) to \\(A\\). The client selection via submodular maximization can be expressed following [3]: find a subset \\(S\\) of clients whose aggregated gradients can approximate the full aggregation from all clients:" + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.25, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\sum_ {k \\in [ N ]} \\nabla F _ {k} (v ^ {k}) = \\sum_ {k \\in [ N ]} \\left[ \\nabla F _ {k} (v ^ {k}) - \\nabla F _ {\\sigma (k)} (v ^ {\\sigma (k)}) \\right] + \\sum_ {k \\in S} \\gamma_ {k} \\nabla F _ {k} (v ^ {k}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.297, + 0.784, + 0.378 + ], + "angle": 0, + "content": "where \\(\\sigma\\) maps \\(V\\to S\\) and the gradient \\(\\nabla F_k(v^k)\\) from client \\(k\\) is approximated by the gradient from a selected client \\(\\sigma (k)\\in S\\). For \\(i\\in S\\), let \\(C_i\\triangleq \\{k\\in V|\\sigma (k) = i\\}\\), and therefore \\(\\gamma_{i}\\triangleq |C_{i}|\\). Take the norms and apply triangular inequality after subtracting the second term from both sides, we can obtain a relaxed objective \\(G(S)\\) for minimizing the approximation error:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.388, + 0.786, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\sum_ {k \\in [ N ]} \\nabla F _ {k} \\left(v ^ {k}\\right) - \\sum_ {k \\in S} \\gamma_ {k} \\nabla F _ {k} \\left(v ^ {k}\\right) \\right\\| \\leqslant \\sum_ {k \\in [ N ]} \\min _ {i \\in S} \\left\\| \\nabla F _ {k} \\left(v ^ {k}\\right) - \\nabla F _ {i} \\left(v ^ {i}\\right) \\right\\| \\triangleq G (S) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.463, + 0.784, + 0.524 + ], + "angle": 0, + "content": "Minimizing \\( G(S) \\) can be seen as maximizing the well-known submodular function, i.e., the facility location function [7]. The submodular maximizing problem is NP-hard but can be approximated via the greedy [25] or stochastic greedy algorithm [24]:" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.536, + 0.785, + 0.564 + ], + "angle": 0, + "content": "\\[\nS \\leftarrow S \\cup k ^ {*}, k ^ {*} \\in \\underset {k \\in \\operatorname {r a n d} (V \\backslash S, \\operatorname {s i z e} = s)} {\\arg \\max } \\left[ \\bar {G} (S) - \\bar {G} (\\{k \\} \\cup S) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.576, + 0.573, + 0.59 + ], + "angle": 0, + "content": "\\(\\bar{G}\\) represents a constant minus the negation of \\(G\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.784, + 0.637 + ], + "angle": 0, + "content": "Although these methods make great improvement in uni-modal FL, the selection strategy is under-explored in MFL and we reveal that traditional client selection approaches cannot address the severe modal bias in MFL." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.654, + 0.536, + 0.67 + ], + "angle": 0, + "content": "2.3 Imbalanced multi-modal learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.781 + ], + "angle": 0, + "content": "Modality imbalance indicates the inconsistent learning progress of different modalities in multi-modal learning [14,37]. Peng et al. [28] propose OGM-GE to alleviate the inhibitory effect on weak modality by slowing down the dominant modality. Fan et al. [11] further build a non-parametric classifier by class centroids to adjust the update direction of weak modality. In this paper, we aim to power each modality of all clients by a meticulously designed modality selection strategy in each round of training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.796, + 0.331, + 0.811 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.821, + 0.784, + 0.851 + ], + "angle": 0, + "content": "In this section, we introduce BMSFed that contains the local imbalance alleviation and balanced modality selection." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.494, + 0.162 + ], + "angle": 0, + "content": "3.1 Local Imbalance Alleviation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.784, + 0.23 + ], + "angle": 0, + "content": "As discussed in Sec. 2.3 and Tab. 1, the multi-modal training on each client may suffer from severe modality imbalance, leading to inadequate information exploitation and consequently incurring the modal bias on the global model. Therefore, we first try to alleviate the imbalance during local training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.785, + 0.314 + ], + "angle": 0, + "content": "According to [11], the prototypes (i.e., class centroids) are suitable to calibrate the gradient directions to avoid the interference from other modalities. Therefore, we choose to use prototypes to facilitate learning of weak modality. For the \\( i \\)-th client with the data \\( \\mathcal{D}_i = \\left\\{\\boldsymbol{X}_i^A, \\boldsymbol{X}_i^I, \\boldsymbol{y}\\right\\} \\), the local prototype for class \\( j \\) of each modality is defined as the mean value of representations:" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.322, + 0.786, + 0.36 + ], + "angle": 0, + "content": "\\[\nc _ {i, j} ^ {I} = \\frac {1}{\\left| \\mathcal {D} _ {i , j} \\right|} \\sum_ {x ^ {I} \\in \\mathcal {D} _ {i, j}} h _ {i} \\left(\\theta_ {i} ^ {I}; x ^ {I}\\right), c _ {i, j} ^ {A} = \\frac {1}{\\left| \\mathcal {D} _ {i , j} \\right|} \\sum_ {x ^ {A} \\in \\mathcal {D} _ {i, j}} h _ {i} \\left(\\theta_ {i} ^ {A}; x ^ {A}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.784, + 0.413 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{i,j}\\) denotes the samples belonging to \\(j\\)-th class in client \\(i\\). \\(h_i\\) is the function of encoder. Considering the heterogeneity across clients, we aggregate the local prototypes to a global prototype as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.422, + 0.786, + 0.461 + ], + "angle": 0, + "content": "\\[\nc _ {j} ^ {G I} = \\frac {1}{| \\mathcal {N} _ {j} |} \\sum_ {i \\in \\mathcal {N} _ {j}} \\frac {| \\mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {I}, c _ {j} ^ {G A} = \\frac {1}{| \\mathcal {N} _ {j} |} \\sum_ {i \\in \\mathcal {N} _ {j}} \\frac {| \\mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {A} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.784, + 0.515 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_j\\) denotes the set of clients that have class \\(j\\) and \\(N_{j}\\) is the number of instances belonging to class \\(j\\) over all clients. Then, we introduce a modal enhancement loss (ME) based on global prototype to adjust local training:" + }, + { + "type": "equation", + "bbox": [ + 0.298, + 0.524, + 0.786, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M E} ^ {k} \\left(v _ {I} ^ {k}\\right) = - \\mathbb {E} _ {\\left(x _ {i} ^ {I}, y\\right) \\in \\mathcal {D} _ {k}} \\log \\left[ \\frac {\\exp \\left(- d \\left(z _ {i} ^ {I} , c _ {y} ^ {G I}\\right)\\right)}{\\sum_ {j = 1} ^ {Y} \\exp \\left(- d \\left(z _ {i} ^ {I} , c _ {j} ^ {G I}\\right)\\right)} \\right] \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.567, + 0.707, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M E} ^ {k} \\left(v _ {A} ^ {k}\\right) = - \\mathbb {E} _ {\\left(x _ {i} ^ {A}, y\\right) \\in \\mathcal {D} _ {k}} \\log \\left[ \\frac {\\exp \\left(- d \\left(z _ {i} ^ {A} , c _ {y} ^ {G A}\\right)\\right)}{\\sum_ {j = 1} ^ {Y} \\exp \\left(- d \\left(z _ {i} ^ {A} , c _ {j} ^ {G A}\\right)\\right)} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.784, + 0.678 + ], + "angle": 0, + "content": "where \\( d(\\cdot, \\cdot) \\) is the distance function (Euclidean distance), \\( z_i^I \\) is the representation of \\( x_i^I \\), i.e., \\( z_i^I = h_i(\\theta_i^I; x_i^I) \\). \\( Y \\) is the class number. \\( v_A^k \\) and \\( v_I^k \\) indicate the corresponding modal data in client \\( k \\). Hence, the local loss should be (data superscripts are omitted for simplicity):" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.687, + 0.786, + 0.72 + ], + "angle": 0, + "content": "\\[\nF _ {k} \\left(v _ {A}, v _ {I}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\gamma^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {A}\\right), \\rho_ {I} ^ {k} \\leqslant 1 \\\\ \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right), \\rho_ {I} ^ {k} > 1 \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.784, + 0.789 + ], + "angle": 0, + "content": "where \\(\\rho_I^k\\) indicates the local imbalance ratio for client \\(k\\) calculated based on the modality-wise ground truth prediction, and \\(\\rho_I^k > 1\\) means modality \\(A\\) outperforms modality \\(I\\) and vice versa. \\(\\gamma^k, \\beta^k \\in (0,1)\\) are the modulation coefficients. The calculation details for \\(\\gamma^k, \\beta^k\\) and \\(\\rho_I^k\\) are in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.79, + 0.784, + 0.851 + ], + "angle": 0, + "content": "Different from [11], we apply the global prototype instead of local prototypes to aggregate the knowledge from different clients, and the ME loss is not only applied to multi-modal clients, but also to uni-modal clients (See details in Sec. 3.2). Now, the weak modality will be stimulated through local update." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.478, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.499, + 0.162 + ], + "angle": 0, + "content": "3.2 Balanced Modality Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.165, + 0.788, + 0.285 + ], + "angle": 0, + "content": "Considering that different types of modal combinations may be selected to participate in local training in our scheme, we define the participated clients as multi-modal clients and uni-modal clients (e.g., client 2 in Fig. 1 Right is a uni-modal client with modal \\( I \\) for training while client 1 and 4 are multi-modal clients with both modalities for training). In Sec. 3.1, we propose the ME loss to facilitate the local learning of weak modality, which is originally designed for multi-modal clients. Here, we also apply it on uni-weak-modal clients to realize gradient consistency on different types of clients." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.285, + 0.787, + 0.315 + ], + "angle": 0, + "content": "Assume modality \\( I \\) is weak here. Hence, the local loss for multi-modal and uni-modal clients should be:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.321, + 0.786, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {m u l t i - m o d a l}: F _ {k} \\left(v _ {A}, v _ {I}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right) \\\\ \\text {u n i - m o d a l}: F _ {k} \\left(v _ {A}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}\\right), F _ {k} \\left(v _ {I}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right) \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.369, + 0.787, + 0.415 + ], + "angle": 0, + "content": "Next, following Eq. (2), we define the paradigm of modality selection for MFL, aiming to approximate the full gradient aggregation from all clients by the gradient from the customized multi-modal and uni-modal clients:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.424, + 0.785, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {k \\in [ N ]} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) = \\sum_ {k \\in [ N ]} \\left[ \\begin{array}{c} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) - \\nabla F _ {\\sigma_ {M} (k)} \\left(v _ {A}, v _ {I}\\right) \\\\ - \\nabla F _ {\\sigma_ {A} (k)} \\left(v _ {A}\\right) - \\nabla F _ {\\sigma_ {I} (k)} \\left(v _ {I}\\right) \\end{array} \\right] \\\\ + \\sum_ {k \\in S _ {M}} \\gamma_ {k} ^ {M} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) + \\sum_ {k \\in S _ {A}} \\gamma_ {k} ^ {A} \\nabla F _ {k} \\left(v _ {A}\\right) + \\sum_ {k \\in S _ {I}} \\gamma_ {k} ^ {I} \\nabla F _ {k} \\left(v _ {I}\\right) \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.508, + 0.787, + 0.554 + ], + "angle": 0, + "content": "where \\(\\sigma_{M},\\sigma_{A}\\) and \\(\\sigma_I\\) map \\(V\\to S_M,S_A,S_I\\), the client sets who use multi-modal, uni-\\(A\\), and uni-\\(I\\) data for training respectively, and \\(S_{M}\\cap S_{A} = S_{A}\\cap S_{I} = S_{M}\\cap S_{I} = \\emptyset\\). \\(\\gamma_k^M\\), \\(\\gamma_k^A\\) and \\(\\gamma_k^I\\) have the similar meaning as \\(\\gamma_{k}\\) in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.787, + 0.614 + ], + "angle": 0, + "content": "Since modality \\( I \\) is weak here, we omit the uni-\\( A \\) clients as the multi-modal gradient is dominated by modality \\( A \\) [28], which means we do not need to select uni-\\( A \\) clients for its enhancement. Then, bring Eq. (9) to Eq. (10) and follow the operations from Eq. (2) to Eq. (3), we can obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.624, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) - \\gamma_ {i} ^ {M} \\nabla F _ {i} \\left(v _ {A}, v _ {I}\\right) - \\gamma_ {j} ^ {I} \\nabla F _ {j} \\left(v _ {I}\\right) \\right\\| \\\\ = \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\begin{array}{c} \\nabla \\mathcal {L} _ {C E} ^ {k} (v _ {A}, v _ {I}) + \\nabla \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} (v _ {I}) - \\nabla \\mathcal {L} _ {C E} ^ {i} (v _ {A}, v _ {I}) \\\\ - \\nabla \\beta^ {i} \\mathcal {L} _ {M E} ^ {i} (v _ {I}) - \\nabla \\mathcal {L} _ {C E} ^ {j} (v _ {I}) - \\nabla \\beta^ {j} \\mathcal {L} _ {M E} ^ {j} (v _ {I}) \\end{array} \\right\\| \\\\ \\leqslant \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}} \\left\\| \\nabla \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) - \\nabla \\mathcal {L} _ {C E} ^ {i} \\left(v _ {A}, v _ {I}\\right) \\right\\| \\tag {11} \\\\ + \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\begin{array}{c} \\nabla \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} (v _ {I}) - \\nabla \\beta^ {i} \\mathcal {L} _ {M E} ^ {i} (v _ {I}) \\\\ - \\nabla \\mathcal {L} _ {C E} ^ {j} (v _ {I}) - \\nabla \\beta^ {j} \\mathcal {L} _ {M E} ^ {j} (v _ {I}) \\end{array} \\right\\| \\\\ \\triangleq G \\left(S _ {M}\\right) + G \\left(S _ {M} \\cup S _ {I}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The right-hand side of the first equation is the modality selection formula that aims to select a group of multi-modal clients \\( S_{M} \\) and uni-modal clients" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.23, + 0.15, + 0.409, + 0.165 + ], + "angle": 0, + "content": "Algorithm 1: BMSFed." + }, + { + "type": "algorithm", + "bbox": [ + 0.23, + 0.167, + 0.761, + 0.401 + ], + "angle": 0, + "content": "Input: Input data \\(\\mathcal{D}_i = \\{\\pmb {X}_i^A,\\pmb {X}_i^I,\\pmb {y}\\} ,i\\in [N],\\) initial model \\(\\theta\\) hyper-parameters \\(\\chi\\) global communication epochs \\(E\\) \\(e = 1\\) while \\(e < E\\) do if \\(e = 1\\) then Send \\(\\theta\\) to all clients; Perform one-step local update for gradients, prototypes and \\(\\rho_I^k\\) Aggregate global prototypes and \\(\\rho_{I}\\) else \\(\\begin{array}{rl} & {\\mathrm{Aggrecate~global~model~}\\theta ,\\mathrm{prototypes~}c^{G}\\mathrm{~and~}\\rho_{I};}\\\\ & {\\mathrm{S e l e c t~multi - mod a l i t y~for~}S_{M}\\mathrm{~a n d~u n i - m o d a l i t y~for~}S_{I}\\mathrm{~(o r~}S_{A})\\mathrm{~u s i n g}}\\\\ & {\\mathrm{E q s.~(12)~a n d~(13);}}\\\\ & {\\mathrm{S e n d~}\\theta ,c^{G}\\mathrm{~a n d~}\\rho_{I}\\mathrm{~t o~s e l e c t e d~c l i e n s;}}\\\\ & {\\mathrm{f o r e a c h~c l i e n t~i n s e l e c t e d~c l i e n s i n~p a r a l l e l d o}}\\\\ & {\\mathrm{P r o p e r m~m u l t i - m o d a l ~l a n e a i n g~i n~}S_{M}\\mathrm{~a n d~u n i - m o d a l ~l e a n i n g~i n~}S_{I}\\mathrm{~(o r}}\\\\ & {\\mathrm{S}_{A})\\mathrm{b y w i t h~E q s.~(8)~a n d~(9);}}\\\\ & {\\mathrm{S e n d~g r a d i e n t s,~l o c a l~p r o t e t y p e s a n d~}\\rho_{I}^{k}\\mathrm{~t o~s e r v e r};} \\end{array}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.587 + ], + "angle": 0, + "content": "\\(S_I\\) to approximate the aggregated gradients from all clients. However, the joint selection for \\(S_M\\) and \\(S_I\\) is a complex joint optimization problem. Therefore, we decouple this objective into two submodular functions \\(G(S_M)\\) and \\(G(S_M \\cup S_I)\\) according to triangle inequality, while the full gradient approximation is divided into two parts: the first part uses selected multi-modal CE gradient to fit fully multi-modal CE gradient aggregation, and the second part approximates the fully multi-modal ME gradient aggregation via selected uni-modal CE gradient and both selected multi- and uni-modal ME gradient. The modality-level gradient decoupling converts the complex joint selection to two simply separated selection problems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Although we can solve the two submodular functions with the stochastic greedy algorithm [24], there are still two issues: (1) the selected client according to \\( G(S_M \\cup S_I) \\) should be specified whether it is uni-modal client or multi-modal client; (2) the separated selection strategy pays less attention to the global modal bias. To address the two problems, we perform the stochastic greedy algorithm for two submodular functions in parallel and propose a simple yet effective conflict resolution strategy to ensure \\( S_M \\cap S_I = \\emptyset \\) as well as, more importantly, balance the learning of different modalities on global model:" + }, + { + "type": "equation", + "bbox": [ + 0.255, + 0.714, + 0.786, + 0.742 + ], + "angle": 0, + "content": "\\[\nS _ {M} \\leftarrow S _ {M} \\cup k _ {1} ^ {*}, k _ {1} ^ {*} \\in \\underset {k \\in \\operatorname {r a n d} (V \\backslash S _ {M} \\backslash S _ {I}, \\mathrm {s})} {\\arg \\max } \\left[ \\bar {G} \\left(S _ {M}\\right) - \\bar {G} \\left(\\{k \\} \\cup S _ {M}\\right) \\right] \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.757, + 0.786, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} i f k _ {1} ^ {*} = k _ {2} ^ {*}, S _ {M} \\cup k _ {2} ^ {*}; \\\\ i f k _ {1} ^ {*} \\neq k _ {2} ^ {*}, \\left\\{ \\begin{array}{l} S _ {I} \\cup k _ {2} ^ {*}, i f \\rho_ {I} ^ {k} > \\chi \\\\ S _ {M} \\cup k _ {2} ^ {*}, i f \\rho_ {I} ^ {k} \\leqslant \\chi \\end{array} \\right. \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.816, + 0.709, + 0.843 + ], + "angle": 0, + "content": "\\[\nk _ {2} ^ {*} \\in \\operatorname * {a r g m a x} _ {k \\in \\operatorname {r a n d} (V \\setminus S _ {M} \\setminus S _ {I}, \\mathrm {s})} \\left[ \\bar {G} \\left(S _ {M} \\cup S _ {I}\\right) - \\bar {G} \\left(\\{k \\} \\cup S _ {M} \\cup S _ {I}\\right) \\right]\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.479, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "For every selection, we randomly sample a subset of clients \\(s\\). A multi-modal client \\(k_{1}^{*}\\) is selected from \\(s\\) according to \\(G(S_{M})\\) while \\(k_{2}^{*}\\) is also selected from \\(s\\) according to \\(G(S_{M} \\cup S_{I})\\). \\(k_{1}^{*} = k_{2}^{*}\\) means using multi-modal data from this client can contribute most to the global model. When \\(k_{1}^{*} \\neq k_{2}^{*}\\), we allocate it to uni-modal or multi-modal client according to its local imbalance ratio: if it is severely imbalanced, we use its uni-weak-modal data for training and aggregation to alleviate the global modal bias, otherwise we believe that training with its multi-modal data contributes more than uni-modal data. \\(\\chi\\) is a hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.789, + 0.427 + ], + "angle": 0, + "content": "Discussion. (1) Overcoming the modal bias in our method are twofold: the ME loss alleviates imbalance at local side and the selected uni-modal clients further promote balanced learning of global model. Meanwhile, the diversity coming from two submodular functions ensures the representative information for the global model. (2) We assume \\( I \\) is the weak modality above while in practice, we can determine the weak modality before modality selection via the aggregated global imbalance ratio \\( \\rho_{I} = \\frac{1}{\\sum_{k=1}^{N} n_{k}} \\sum_{k=1}^{N} \\rho_{I}^{k} \\cdot n_{k} \\). Overall, the pseudo-code of BMSFed is provided in Algorithm 1. (3) Only the gradients, prototypes and \\( \\rho_{I}^{k} \\) participate in communication, so there is no privacy issue and similar communication overheads as in traditional FL." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.459, + 0.357, + 0.475 + ], + "angle": 0, + "content": "4 Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.45, + 0.516 + ], + "angle": 0, + "content": "4.1 Datasets and baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on four datasets: (1) CREMA-D [4] is an audio-visual dataset for emotion recognition task with total six categories for emotional states. (2) AVE [33] is an audio-visual dataset for event localization with 28 event classes, and here we use it to construct a labeled multimodal classification dataset following [11]. (3) Colored-and-gray MNIST (CG-MNIST) [17] is a synthetic dataset based on MNIST [18] with gray-scale and monochromatic images as two modalities, following [38]. (4) ModelNet40 is one of the Princeton ModelNet datasets [39] with 3D objects of 40 categories. The front and back [31] views are considered as two modalities, following [5]. Details of these four datasets are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baselines. We choose eight baselines for comparison from four categories: (1) three uni-modal FL methods designed for statistical heterogeneity are extended to multi-modal scenarios: FedAvg [23], FedProx [21] and FedProto [32]. (2) Integrating OGM-GE [28] and PMR [11], the solutions for modality imbalance, with FedAvg forms two MFL methods: FedOGM and FedPMR. (3) Two client selection method: Power-of-choice (pow-d) [6] and DivFL [3], evolved from its uni-modal version directly. (4) One MFL method, FedMSplit [5], especially designed for modality incongruity. Compared with these baselines, we demonstrate that an elaborate modality selection strategy is essential to realize comprehensive information exploitation in MFL." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison results on four datasets. The metric is the top-1 accuracy (\\%). The best is in **bold**, and the second best is **underlined**. Our method achieves significant improvement on both IID and non-IID scenarios." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.794, + 0.36 + ], + "angle": 0, + "content": "
DatasetCREMA-DAVECG-MNISTModelNet40
MethodIIDnon-IIDIIDnon-IIDIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.742.341.787.286.5
FedProx51.049.062.659.942.943.686.987.1
FedProto58.754.061.758.851.551.487.587.2
FedOGM56.956.462.859.357.253.087.687.0
FedPMR55.555.163.161.666.163.387.687.7
pow-d50.550.762.560.041.240.386.886.2
DivFL51.750.863.359.643.042.186.586.4
FedMSplit52.451.662.460.843.550.987.587.4
BMSFed64.561.664.762.170.266.788.787.5
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.39, + 0.442, + 0.406 + ], + "angle": 0, + "content": "4.2 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.42, + 0.788, + 0.752 + ], + "angle": 0, + "content": "For CREMA-D, AVE and ModelNet40, we use ResNet18 [12] as the backbone for audio, visual and flow modalities. Audio data is converted to a spectrogram of size 257x299 for CREMA-D and \\(257 \\times 1,004\\) for AVE. We randomly choose 3 frames and 4 frames to build image training sets for CREMA-D and AVE respectively. For CG-MNIST, we build a neural network with 4 convolution layers and 1 average pool layer as the encoder, following the setting as in [11]. We choose the simple yet effective fusion method, concatenation [27], to build fusion classifier for all the datasets. We set 20 clients for CREMA-D, AVE and ModelNet40 while the number for CG-MNIST is 30. 5 clients are selected in each communication round for CREMA-D, AVE, ModelNet40 and 6 for CG-MNIST. For IID setting, training data is uniformly distributed to all clients. For non-IID scenarios, we use Dirichlet distribution [13] \\(Dir(\\alpha)\\) to split data (\\(\\alpha = 3\\) for CREMA-D, AVE, ModelNet40, \\(\\alpha = 2\\) for CG-MNIST). The optimizer is SGD [30] for all datasets. Learning rate is initialized at 1e-3 or 1e-2 for CEAMA-D, AVE and ModelNet40 or CG-MNIST and becomes 1e-4 or 1e-3 in the later training stage. The hyper-parameter \\(\\chi\\) is set to 1.2-2.5 according to datasets and settings. To complete stochastic greedy algorithm for Eqs. (12) and (13), we use the gradients from the selected clients at current round to update part of the similarity matrix, which is named \"no-overheads\" in [3]. Except for pow-d, DivFL and BMSFed, other baselines select clients randomly. Each client has two modal data by default. We do all experiments on a workstation with an RTX 3090 GPU, a 3.9-GHZ Intel Core i9-12900K CPU and 64GB of RAM." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.78, + 0.482, + 0.795 + ], + "angle": 0, + "content": "4.3 Comparison with baselines" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "BMSFed effectively improves the performance. As demonstrated in Tab. 2, our BMSFed achieves the best results on the four datasets under both IID and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.48, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.478, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.262, + 0.446, + 0.274 + ], + "angle": 0, + "content": "(a) CREMA-D under IID" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.145, + 0.734, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.262, + 0.689, + 0.274 + ], + "angle": 0, + "content": "(b) AVE under IID" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.274, + 0.475, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.391, + 0.459, + 0.402 + ], + "angle": 0, + "content": "(c) CREMA-D under non-HID" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.275, + 0.734, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.391, + 0.701, + 0.402 + ], + "angle": 0, + "content": "(d) AVE under non-IID" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.414, + 0.787, + 0.443 + ], + "angle": 0, + "content": "Fig. 2: Test accuracy of BMSFed compared with other baselines on CREMA-D and AVE. BMSFed converges to more accurate solutions than all baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.653 + ], + "angle": 0, + "content": "non-IID settings (by up to \\(5.8\\%\\) on CREMA-D). Client sampling here (powd and DivFL) cannot fully exploit information for all modalities, making its improvement limited or even worse than FedAvg in CG-MNIST. Although FedOGM and FedPMR accomplish modest improvement because of their ability to alleviate modality imbalance, they are not as good as BMSFed since they do not consider the overall performance of the global model. Traditional uni-modal FL methods for statistical heterogeneity (e.g. FedProx) and MFL method for modality incongruity (FedMSplit) only obtain slight improvement. We also illustrate the trend of test accuracy versus the number of communication rounds on CREMA-D and AVE in Fig. 2. BMSFed can realize comparable or even faster convergence speeds in CREMA-D and AVE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "BMSFed exploits all modalities comprehensively. To show the effect of our method on addressing modal bias, we report the performance of each modality on CREMA-D and AVE as shown in Tab. 3. The uni-modal performance evaluation follows [11]: a sample is classified into the class corresponding to its nearest prototype. It is clear that BMSFed could considerably improve the performance of weak modality (visual) and mitigate the modality-level bias. Besides, compared with randomly modality abandoning, which significantly reduces audio performance as illustrated in Tab. 1, BMSFed achieves comparable audio performance with other baselines. Although FedProto, FedOGM and FedPMR also alleviate the imbalance, they mainly focus on local optimization, resulting in the performance gap between them and our BMSFed on the aggregated model, which further indicates that in MFL, it is important to take both local opti" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 3: The uni-modal performance comparison on CREMA-D and AVE. The metric is the top-1 accuracy (\\%). Our method achieves the best uni-visual (the weak modality) performance on all settings and maintains the comparable uni-audio performance." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.199, + 0.755, + 0.374 + ], + "angle": 0, + "content": "
DatasetCREMA-DAVE
SettingIIDnon-IIDIIDnon-IID
MethodAVAVAVAV
FedAvg51.220.650.720.261.126.861.426.4
FedProx51.320.250.122.060.427.161.226.9
FedProto50.235.348.639.155.736.859.732.8
FedOGM50.535.748.830.258.728.859.429.4
FedPMR51.538.750.135.961.739.661.735.3
pow-d51.520.451.618.861.926.960.127.1
DivFL52.321.152.122.762.725.361.626.3
FedMSplit52.021.850.821.661.326.962.328.7
BMSFed51.041.949.341.459.740.260.238.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.386, + 0.787, + 0.414 + ], + "angle": 0, + "content": "Table 4: Ablation study. 'BMSFed-local' uses local prototypes rather than global prototypes for ME loss." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.426, + 0.713, + 0.585 + ], + "angle": 0, + "content": "
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.7
DivFL51.750.863.359.6
FedAvg-0.252.450.163.461.1
FedAvg-0.555.755.160.759.4
FedAvg-0.861.258.158.558.7
FedAvg+ME55.854.562.860.7
DivFL+ME57.155.663.061.1
BMSFed-local63.760.363.460.1
BMSFed64.561.664.762.1
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.613, + 0.787, + 0.644 + ], + "angle": 0, + "content": "mization for each modality and the overall performance for global model into consideration simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.389, + 0.681 + ], + "angle": 0, + "content": "4.4 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of each component. Tab. 4 studies the effect of each BMSFed component. Applying ME loss Eq. (7) on random sampling \\((\\mathrm{FedAvg} + \\mathrm{ME})\\) and the well-designed client selection \\((\\mathrm{DivFL} + \\mathrm{ME})\\) surpasses their vanilla strategies \\((\\mathrm{FedAvg}, \\mathrm{DivFL})\\) by a large margin, demonstrating its effectiveness on local enhancement. Comparing BMSFed (64.5% on IID CREMA-D) with 'DivFL+ME' (57.1% on the same setting) also denotes the necessity of balancing different modalities considering the global model via modality selection. To show the importance of aligning feature spaces of weak modality, we replace the global prototypes with local prototypes (BMSFed-local). Global alignment achieves notable improvement (by up to 2% on non-IID AVE). The performance improvement" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.479, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.391, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.243, + 0.388, + 0.255 + ], + "angle": 0, + "content": "(a) CREMA-D under IID" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.146, + 0.578, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.243, + 0.577, + 0.255 + ], + "angle": 0, + "content": "(b) CREMA-D under IID" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.144, + 0.768, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.243, + 0.766, + 0.255 + ], + "angle": 0, + "content": "(c) CG-MNIST under IID" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.267, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig. 3: Robustness validation on data size, local epoch and client number. Our BMSFed consistently outperforms baseline (FedAvg) under various scenarios." + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.317, + 0.644, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.466, + 0.785, + 0.494 + ], + "angle": 0, + "content": "Fig. 4: Proportional change of audio and visual respectively and the curve of global imbalance ratio during training on CREMA-D under IID setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.56 + ], + "angle": 0, + "content": "compared with 'FedAvg-0.2,0.5,0.8' exhibits that randomly sampling modalities does not always lead to improvement and further demonstrates the need of meticulously selecting modalities for information exploitation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Robustness test. To verify the robustness of our method, we vary three key hyperparameters to build various scenarios: (1) change the data size \\( |\\mathcal{D}_i| \\) to allow each client to hold a small amount of data, (2) set different local training epochs, and (3) vary the total client number \\( N \\). As illustrated in Figs. 3a and 3b, our BMSFed consistently outperforms baseline (FedAvg) under various scenarios. More data as well as more local training epochs can bring further improvements to our method, indicating that exploiting the weak modality need more training efforts. Based on the results in Fig. 3c, our approach can also be generalised to clients with larger scales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The relationship between modality selection and imbalance degree. In Eqs. (12) and (13), we use a conflict resolution strategy based on local imbalance ratio to realize balanced modality selection. We visualize the proportions of audio and visual modalities selected in each round and the global imbalance ratio. It is clear from Fig. 4 that audio is the dominant modality (imbalance ratio is always greater than 1) and modality imbalance is gradually alleviated as training progresses (global imbalance ratio shows a downward trend). In addition, the proportion of selected visual modality follows the same trend (larger in the early stage of training and becomes smaller later), implying the rationality of our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Performance on CREMA-D and AVE with modality incongruity. \\(50\\%\\) of clients have all modal data and \\(50\\%\\) of clients only retain data with a single modality (audio or visual)." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.199, + 0.7, + 0.358 + ], + "angle": 0, + "content": "
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg55.755.160.758.7
FedProx56.856.061.258.5
FedProto58.757.061.359.7
FedOGM58.657.460.158.5
FedPMR56.455.561.960.3
DivFL57.455.661.158.6
FedMSplit58.956.961.760.0
BMSFed62.459.863.560.9
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.39, + 0.785, + 0.418 + ], + "angle": 0, + "content": "selection strategy based on imbalance ratio and its effectiveness on mitigating bias." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.573 + ], + "angle": 0, + "content": "Effectiveness on modality incongruity scenario. All above experiments assume that each client initially has complete modal data. Here, we build the stimulation of modality incongruity scenario, in which half of the clients have data with two modalities, and the other half only have data in one modality: random audio or visual. The results are shown in Tab. 5 (The results of pow-d is not available because it is not applicable to this scenario). Our BMSFed still makes impressive improvement compared with all other baselines (by up to \\(3.5\\%\\) on IID CREMA-D), illustrating the good generalization ability of our method in different scenarios. It is worth mentioning that FedMSplit performs better than before because it is specifically designed for modality incongruity." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.604, + 0.358, + 0.619 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we analyze traditional client selections and find their ineffectiveness in MFL. We further reveal that there exists strong modality-level bias due to the modality imbalance during the training iterations and uni-modal training on some clients may contribute more to the global model than multi-modal training. To address this issue, we propose the balanced modality selection scheme for MFL (BMSFed) with modality-level gradient decoupling to release the potential of all modalities and maximize the gradient diversities to improve global aggregation. We also introduce a modal enhancement loss to optimize the local update process. Our method does not introduce additional local training costs and communication overheads compared with previous methods. Extensive experiments on four datasets demonstrate the superiority of our method in performance and applicability under different modal combinations, data distributions and modality incongruity scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.479, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.234 + ], + "angle": 0, + "content": "Our work was supported by two grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. PolyU15222621, PolyU15225023) and the funding from National Natural Science Foundation of China under grants 62302184." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.255, + 0.323, + 0.271 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.323 + ], + "angle": 0, + "content": "1. Abavisani, M., Wu, L., Hu, S., Tetreault, J., Jaimes, A.: Multimodal categorization of crisis events in social media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14679-14689 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.377 + ], + "angle": 0, + "content": "2. Alfasly, S., Lu, J., Xu, C., Zou, Y.: Learnable irrelevant modality dropout for multimodal action recognition on modality-specific annotated videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20208-20217 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.418 + ], + "angle": 0, + "content": "3. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.459 + ], + "angle": 0, + "content": "4. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing 5(4), 377-390 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "5. Chen, J., Zhang, A.: Fedmsplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 87-96 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.54 + ], + "angle": 0, + "content": "6. Cho, Y.J., Wang, J., Joshi, G.: Client selection in federated learning: Convergence analysis and power-of-choice selection strategies. arXiv preprint arXiv:2010.01243 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.567 + ], + "angle": 0, + "content": "7. Cornuejols, G., Fisher, M., Nemhauser, G.L.: On the uncapacitated location problem. In: Annals of Discrete Mathematics, vol. 1, pp. 163-177. Elsevier (1977)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.608 + ], + "angle": 0, + "content": "8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.649 + ], + "angle": 0, + "content": "9. Deng, Y., Lyu, F., Ren, J., Wu, H., Zhou, Y., Zhang, Y., Shen, X.: Auction: Automated and quality-aware client selection framework for efficient federated learning. IEEE Transactions on Parallel and Distributed Systems 33(8), 1996-2009 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.651, + 0.785, + 0.69 + ], + "angle": 0, + "content": "0. Duan, J.h., Li, W., Zou, D., Li, R., Lu, S.: Federated learning with data-agnostic distribution fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8074–8083 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "11. Fan, Y., Xu, W., Wang, H., Wang, J., Guo, S.: Pmr: Prototypical modal rebalance for multimodal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20029-20038 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.732, + 0.785, + 0.771 + ], + "angle": 0, + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.798 + ], + "angle": 0, + "content": "13. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.839 + ], + "angle": 0, + "content": "14. Huang, Y., Lin, J., Zhou, C., Yang, H., Huang, L.: Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In: International Conference on Machine Learning. pp. 9226-9259. PMLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.283, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Fan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Imteaj, A., Thakker, U., Wang, S., Li, J., Amini, M.H.: A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal 9(1), 1-24 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "16. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "17. Kim, B., Kim, H., Kim, K., Kim, S., Kim, J.: Learning not to learn: Training deep neural networks with biased data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9012-9020 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "18. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "19. Li, H., Li, X., Hu, P., Lei, Y., Li, C., Zhou, Y.: Boosting multi-modal model performance with adaptive gradient modulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22214-22224 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.366 + ], + "angle": 0, + "content": "20. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.405 + ], + "angle": 0, + "content": "21. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.787, + 0.433 + ], + "angle": 0, + "content": "22. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "23. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "24. Mirzasoleiman, B., Badanidiyuru, A., Karbasi, A., Vondrák, J., Krause, A.: Lazier than lazy greedy. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 29 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "25. Nemhauser, G.L., Wolsey, L.A., Fisher, M.L.: An analysis of approximations for maximizing submodular set functions--i. Mathematical programming 14, 265-294 (1978)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.595 + ], + "angle": 0, + "content": "26. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.637 + ], + "angle": 0, + "content": "27. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European conference on computer vision (ECCV). pp. 631-648 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.787, + 0.677 + ], + "angle": 0, + "content": "28. Peng, X., Wei, Y., Deng, A., Wang, D., Hu, D.: Balanced multimodal learning via on-the-fly gradient modulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8238-8247 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.732 + ], + "angle": 0, + "content": "29. Reisizadeh, A., Tziotis, I., Hassani, H., Mokhtari, A., Pedarsani, R.: Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity. IEEE Journal on Selected Areas in Information Theory 3(2), 197-205 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "30. Robbins, H., Monro, S.: A stochastic approximation method. The annals of mathematical statistics pp. 400-407 (1951)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "31. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.479, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Balanced Modality Selection for MFL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 247-263 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "35. Wang, H., Qu, Z., Guo, S., Wang, N., Li, R., Zhuang, W.: Losp: Overlap synchronization parallel with local compensation for fast distributed training. IEEE Journal on Selected Areas in Communications 39(8), 2541-2557 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "36. Wang, H., Xu, W., Fan, Y., Li, R., Zhou, P.: Aocc-fl: Federated learning with aligned overlapping via calibrated compensation. In: IEEE INFOCOM 2023-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.355 + ], + "angle": 0, + "content": "37. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12695-12705 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "38. Wu, N., Jastrzebski, S., Cho, K., Geras, K.J.: Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. In: International Conference on Machine Learning. pp. 24043-24055. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "39. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.467 + ], + "angle": 0, + "content": "40. Xiao, F., Lee, Y.J., Grauman, K., Malik, J., Feichtenhofer, C.: Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.494 + ], + "angle": 0, + "content": "41. Xiong, B., Yang, X., Qi, F., Xu, C.: A unified framework for multi-modal federated learning. Neurocomputing 480, 110-118 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.522 + ], + "angle": 0, + "content": "42. Xu, J., Tong, X., Huang, S.L.: Personalized federated learning with feature alignment and classifier collaboration. arXiv preprint arXiv:2306.11867 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "43. Xu, J., Wang, H.: Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective. IEEE Transactions on Wireless Communications 20(2), 1188-1200 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.592 + ], + "angle": 0, + "content": "44. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.633 + ], + "angle": 0, + "content": "45. Zhao, Y., Barnaghi, P., Haddadi, H.: Multimodal federated learning on IoT data. In: 2022 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation (IoTDI). pp. 43-54. IEEE (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.633 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_origin.pdf b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b229102ce8a903af282466d16dd4e20f5a43d9af --- /dev/null +++ b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/0a3f63d7-3f37-4257-852d-e86bee89a98f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:107cb31161cea28a260cc4a8df41c9e35d971233eaf147f8f3eed31e2f853acd +size 908736 diff --git a/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/full.md b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c29a3945f2b04364b0e3a0f29ad8b0d45292a4bc --- /dev/null +++ b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/full.md @@ -0,0 +1,315 @@ +# Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection + +Yunfeng Fan $^{1}$ , Wenchao Xu $^{1(\boxdot)}$ , Haozhao Wang $^{2}$ , Fushuo Huo $^{1}$ , Jinyu Chen $^{1}$ , and Song Guo $^{3}$ + +$^{1}$ Department of Computing, The Hong Kong Polytechnic University, Hong Kong +$^{2}$ School of Computer Science and Technology, Huazhong University of Science and Technology, Hubei, China +3 Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong + +{yunfeng.fan, fushuo.huo, jinyu.chen}@connect.polyu.hk +wenchao.xu@polyu.edu.hk, hz_wang@hust.edu.cn, songguo@cse.ust.hk +https://github.com/fanyunfeng-bit/Balanced-Modality-Selection-in-MFL + +Abstract. Selecting proper clients to participate in each federated learning (FL) round is critical to effectively harness a broad range of distributed data. Existing client selection methods simply consider the mining of distributed uni-modal data, yet, their effectiveness may diminish in multi-modal FL (MFL) as the modality imbalance problem not only impedes the collaborative local training but also leads to a severe global modality-level bias. We empirically reveal that local training with a certain single modality may contribute more to the global model than training with all local modalities. To effectively exploit the distributed multiple modalities, we propose a novel Balanced Modality Selection framework for MFL (BMSFed) to overcome the modal bias. On the one hand, we introduce a modal enhancement loss during local training to alleviate local imbalance based on the aggregated global prototypes. On the other hand, we propose the modality selection aiming to select subsets of local modalities with great diversity and achieving global modal balance simultaneously. Our extensive experiments on audio-visual, colored-gray, and front-back datasets showcase the superiority of BMSFed over baselines and its effectiveness in multi-modal data exploitation. + +Keywords: Multi-modal federated learning $\cdot$ modality imbalance $\cdot$ modality selection + +# 1 Introduction + +Federated learning (FL) [23] aims to collaboratively learn data that has been collected by, and resides on, a number of remote devices or servers. FL stands to develop top-performing models by aggregating knowledge from numerous edge clients [8,36], which relies on the iterative interaction among participating clients and the server. However, comprehensively employing the information from all + +Table 1: Performance of various client selection methods in MFL under IID setting. A and V denote uni-audio and uni-visual while A-V means the multi-modal result. 'Local' represents that a client is trained based its local data without aggregation. A strong modal bias of global model exists on the two datasets. + +
DatasetCREMA-D [4]AVE [33]
MethodAVA-VAVA-V
Local41.920.439.633.416.735.2
FedAvg51.220.650.761.126.862.2
pow-d [6]51.520.450.561.926.962.5
DivFL [3]52.321.151.762.725.363.3
FedAvg-0.250.628.652.460.629.663.4
FedAvg-0.550.534.655.758.730.060.7
FedAvg-0.848.150.961.256.431.858.5
BMSFed51.041.964.559.740.264.7
+ +clients can be exceptionally difficult due to the client heterogeneity and resource limitations [10, 29]. + +Random sampling [22, 34] from available clients has been widely used in FL to satisfy some practical restrictions, e.g., limited communication bandwidth [26, 35] and computing capacities [15]. To improve the information exploitation of all clients, extensive research has been conducted on effective client selection strategies [9, 43]. Despite the success of traditional client selection methods in uni-modal FL, their effectiveness diminishes when dealing with clients with multi-modal data as the inter-modal interactions during the MFL training are neglected. According to [14, 37], there may exist inconsistent learning paces for different modalities in multi-modal joint training, i.e., modality imbalance, which not only impedes the collaborative local training but also leads to a severe modal bias for global model in MFL. As illustrated in Tab. 1, audio modality significantly outperforms visual modality in CREMA-D and AVE datasets during local training and the aggregated model still suffers from it. However, two well-designed client selection methods (pow-d [6] and DivFL [3]) only obtain severely limited improvement over random sampling (FedAvg) on the multi-modal global model. We can see that client selection methods achieve the best uni-audio performance while uni-visual performance even drops sometimes, which means existing client selection scheme heavily relies on the better modality, while ignoring the importance of improving weak modalities that also has potential for global model aggregation. Based on the above analysis, a pivotal question arises: Can we design a new selection scheme in MFL that can overcome the modal bias and exploit each modality comprehensively? + +To answer this question, we investigate the interactions between different modalities via randomly discarding the data from one modality (audio or visual) on part of clients, which is inspired from modality dropout [2, 40] that drops a specific modal data during training for regularization. The results are shown in rows 7-9 in Tab. 1, where $-x$ denotes randomly discarding a modality on a client with probability $x$ . We can see that dropping with a certain prob + +![](images/120a71489c7f4011cf5f68f0ab0ba1f38a566c955a660b2fcc2ccf8631830f73.jpg) +Fig. 1: Left: Traditional client selection in FL aims to sample a client subset in each round while our modality selection considers each local modality as the sampling unit. Right: The paradigm of BMSFed with four clients. The global prototypes are used to enhance the weak modality during local update. Only networks corresponding to the selected modalities will be uploaded to the server for aggregation. + +ability can improve the global multi-modal performance on both datasets (e.g., FedAvg-0.2), and the main reason comes from the dramatic improvement of visual modality. This phenomenon suggests that performing uni-modal training can unleash its potential without being inhibited by another modality, and uni-modal local training may contribute more to the global model than multi-modal training on some clients. As the dropping ratio increases, although the visual modality still improves, multi-modal performance may decline as the audio modality declines as shown in columns 5 and 7, suggesting that we should carefully control which modalities on each client should be involved in training and aggregation to make the contribution most to the global model. + +According to the above investigations, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to mitigating the modal bias and comprehensively exploit the diverse information from all modalities. Specifically, instead of selecting a subset of clients, we treat each modality on the local side as a selection unit, as demonstrated in Fig. 1. Our BMSFed mainly contains two parts: Firstly, we intend to alleviate the local modality imbalance by introducing a modal enhancement loss based on aggregated global prototypes to promote the performance of weak modality. Secondly, we complete the modality selection by building two separated submodular functions for selecting multi-modal clients (training with multi-modal data) and uni-modal clients (training with selected uni-modal data) respectively. Inspired from [3], the criterion is to select modalities that are most representative on the gradients while also alleviate the global modal bias. A simple yet effective conflict resolution strategy is devised to ensure the validity of modality selection and keep modal balance on global model simultaneously. + +The main contributions of the paper are summarized as follows: + +1. We empirically analyze the modality imbalance problem in MFL and reveal that uni-modal training on some clients may contribute more to the global model than multi-modal training. + +2. Based on the analysis, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to comprehensively exploit all modalities via a modal enhancement loss and representative modality selection to overcome the global modal bias. + +3. We conduct comprehensive experiments on audio-visual, colored-gray, and front-back datasets, and considering the statistical heterogeneity and modality incongruity problems in MFL, to validate the superiority of our BMSFed. + +# 2 Background and Related Works + +# 2.1 Multi-modal federated learning + +In MFL, each client has one or multiple modalities of data. Without loss of generality, we consider two input modalities, which are denoted by $A$ and $I$ , respectively in MFL. There are a set $V$ of $N$ clients, $V = [N]$ , respectively owning datasets $\mathcal{D}_i = \left\{\mathbf{X}_i^A, \mathbf{X}_i^I, \mathbf{y}\right\}$ , $i \in [N]$ . A typical federated learning objective is the average of each client's local loss function: + +$$ +\min f (\theta) = \sum_ {k = 1} ^ {N} p _ {k} F _ {k} (\theta) \tag {1} +$$ + +where $\theta = \{\theta^A, \theta^I, \omega\}$ denotes the model parameters. $\theta^A$ and $\theta^I$ represent the encoder parameters of modality $A$ and $I$ . $\omega$ is the parameter of fusion classifier. $p_k$ is a pre-defined weight. $F_k$ is each client's local loss (cross entropy (CE) loss for classification task in this paper). + +Statistical heterogeneity [16,20] is a widely concerned challenge in uni-modal FL. To tackle this issue, FedProx [21] uses a proximal term to stabilize model aggregation. FedProto [32] shares class prototypes to regularize the learning of local models. In MFL, modality incongruity [44,45] (clients consist of different modalities combinations), as well as statistical heterogeneity [41], are all considered. Yu et al. [44] propose CreamFL to align the representations between different clients and different modalities via communicating knowledge on a public dataset. Chen et al. [5] introduce FedMSplit to split local models into several components and aggregate them by their correlations. However, they still focus on the heterogeneity, but ignore the interaction between private data of different modalities, which limits their information exploitation. + +# 2.2 Client selection and submodular function + +Client selection [9,43] is a critical issue for FL especially when the communication cost with all devices is prohibitively high, which has been extensively studied in uni-modal FL. Cho et al. [6] propose Power-of-Choice to select clients with targets local loss. Balakrishnan et al. [3] propose to select a subset of clients with great diversity. + +Diverse client selection via submodularity. Maximizing a submodular function is reported to improve the diversity and reduce the redundancy of a subset. This property makes it appropriate for client selection in FL. If a function $F$ is submodular, it should satisfy: given a finite ground set $V$ of size + +$N$ , $F(A \cup \{v\}) - F(A) \geqslant F(B \cup \{v\}) - F(B)$ , for any $A \subseteq B \subseteq V$ and $v \in V \setminus B$ . The marginal utility of an element $v$ w.r.t. a subset $A$ is denoted as $F(v|A) = F(A \cup \{v\}) - F(A)$ , which can represent the importance of $v$ to $A$ . The client selection via submodular maximization can be expressed following [3]: find a subset $S$ of clients whose aggregated gradients can approximate the full aggregation from all clients: + +$$ +\sum_ {k \in [ N ]} \nabla F _ {k} (v ^ {k}) = \sum_ {k \in [ N ]} \left[ \nabla F _ {k} (v ^ {k}) - \nabla F _ {\sigma (k)} (v ^ {\sigma (k)}) \right] + \sum_ {k \in S} \gamma_ {k} \nabla F _ {k} (v ^ {k}) \tag {2} +$$ + +where $\sigma$ maps $V\to S$ and the gradient $\nabla F_k(v^k)$ from client $k$ is approximated by the gradient from a selected client $\sigma (k)\in S$ . For $i\in S$ , let $C_i\triangleq \{k\in V|\sigma (k) = i\}$ , and therefore $\gamma_{i}\triangleq |C_{i}|$ . Take the norms and apply triangular inequality after subtracting the second term from both sides, we can obtain a relaxed objective $G(S)$ for minimizing the approximation error: + +$$ +\left\| \sum_ {k \in [ N ]} \nabla F _ {k} \left(v ^ {k}\right) - \sum_ {k \in S} \gamma_ {k} \nabla F _ {k} \left(v ^ {k}\right) \right\| \leqslant \sum_ {k \in [ N ]} \min _ {i \in S} \left\| \nabla F _ {k} \left(v ^ {k}\right) - \nabla F _ {i} \left(v ^ {i}\right) \right\| \triangleq G (S) \tag {3} +$$ + +Minimizing $G(S)$ can be seen as maximizing the well-known submodular function, i.e., the facility location function [7]. The submodular maximizing problem is NP-hard but can be approximated via the greedy [25] or stochastic greedy algorithm [24]: + +$$ +S \leftarrow S \cup k ^ {*}, k ^ {*} \in \underset {k \in \operatorname {r a n d} (V \backslash S, \operatorname {s i z e} = s)} {\arg \max } \left[ \bar {G} (S) - \bar {G} (\{k \} \cup S) \right] \tag {4} +$$ + +$\bar{G}$ represents a constant minus the negation of $G$ + +Although these methods make great improvement in uni-modal FL, the selection strategy is under-explored in MFL and we reveal that traditional client selection approaches cannot address the severe modal bias in MFL. + +# 2.3 Imbalanced multi-modal learning + +Modality imbalance indicates the inconsistent learning progress of different modalities in multi-modal learning [14,37]. Peng et al. [28] propose OGM-GE to alleviate the inhibitory effect on weak modality by slowing down the dominant modality. Fan et al. [11] further build a non-parametric classifier by class centroids to adjust the update direction of weak modality. In this paper, we aim to power each modality of all clients by a meticulously designed modality selection strategy in each round of training. + +# 3 Method + +In this section, we introduce BMSFed that contains the local imbalance alleviation and balanced modality selection. + +# 3.1 Local Imbalance Alleviation + +As discussed in Sec. 2.3 and Tab. 1, the multi-modal training on each client may suffer from severe modality imbalance, leading to inadequate information exploitation and consequently incurring the modal bias on the global model. Therefore, we first try to alleviate the imbalance during local training. + +According to [11], the prototypes (i.e., class centroids) are suitable to calibrate the gradient directions to avoid the interference from other modalities. Therefore, we choose to use prototypes to facilitate learning of weak modality. For the $i$ -th client with the data $\mathcal{D}_i = \left\{\boldsymbol{X}_i^A, \boldsymbol{X}_i^I, \boldsymbol{y}\right\}$ , the local prototype for class $j$ of each modality is defined as the mean value of representations: + +$$ +c _ {i, j} ^ {I} = \frac {1}{\left| \mathcal {D} _ {i , j} \right|} \sum_ {x ^ {I} \in \mathcal {D} _ {i, j}} h _ {i} \left(\theta_ {i} ^ {I}; x ^ {I}\right), c _ {i, j} ^ {A} = \frac {1}{\left| \mathcal {D} _ {i , j} \right|} \sum_ {x ^ {A} \in \mathcal {D} _ {i, j}} h _ {i} \left(\theta_ {i} ^ {A}; x ^ {A}\right) \tag {5} +$$ + +where $\mathcal{D}_{i,j}$ denotes the samples belonging to $j$ -th class in client $i$ . $h_i$ is the function of encoder. Considering the heterogeneity across clients, we aggregate the local prototypes to a global prototype as: + +$$ +c _ {j} ^ {G I} = \frac {1}{| \mathcal {N} _ {j} |} \sum_ {i \in \mathcal {N} _ {j}} \frac {| \mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {I}, c _ {j} ^ {G A} = \frac {1}{| \mathcal {N} _ {j} |} \sum_ {i \in \mathcal {N} _ {j}} \frac {| \mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {A} \tag {6} +$$ + +where $\mathcal{N}_j$ denotes the set of clients that have class $j$ and $N_{j}$ is the number of instances belonging to class $j$ over all clients. Then, we introduce a modal enhancement loss (ME) based on global prototype to adjust local training: + +$$ +\mathcal {L} _ {M E} ^ {k} \left(v _ {I} ^ {k}\right) = - \mathbb {E} _ {\left(x _ {i} ^ {I}, y\right) \in \mathcal {D} _ {k}} \log \left[ \frac {\exp \left(- d \left(z _ {i} ^ {I} , c _ {y} ^ {G I}\right)\right)}{\sum_ {j = 1} ^ {Y} \exp \left(- d \left(z _ {i} ^ {I} , c _ {j} ^ {G I}\right)\right)} \right] \tag {7} +$$ + +$$ +\mathcal {L} _ {M E} ^ {k} \left(v _ {A} ^ {k}\right) = - \mathbb {E} _ {\left(x _ {i} ^ {A}, y\right) \in \mathcal {D} _ {k}} \log \left[ \frac {\exp \left(- d \left(z _ {i} ^ {A} , c _ {y} ^ {G A}\right)\right)}{\sum_ {j = 1} ^ {Y} \exp \left(- d \left(z _ {i} ^ {A} , c _ {j} ^ {G A}\right)\right)} \right] +$$ + +where $d(\cdot, \cdot)$ is the distance function (Euclidean distance), $z_i^I$ is the representation of $x_i^I$ , i.e., $z_i^I = h_i(\theta_i^I; x_i^I)$ . $Y$ is the class number. $v_A^k$ and $v_I^k$ indicate the corresponding modal data in client $k$ . Hence, the local loss should be (data superscripts are omitted for simplicity): + +$$ +F _ {k} \left(v _ {A}, v _ {I}\right) = \left\{ \begin{array}{l} \mathcal {L} _ {C E} ^ {k} \left(v _ {A}, v _ {I}\right) + \gamma^ {k} \mathcal {L} _ {M E} ^ {k} \left(v _ {A}\right), \rho_ {I} ^ {k} \leqslant 1 \\ \mathcal {L} _ {C E} ^ {k} \left(v _ {A}, v _ {I}\right) + \beta^ {k} \mathcal {L} _ {M E} ^ {k} \left(v _ {I}\right), \rho_ {I} ^ {k} > 1 \end{array} \right. \tag {8} +$$ + +where $\rho_I^k$ indicates the local imbalance ratio for client $k$ calculated based on the modality-wise ground truth prediction, and $\rho_I^k > 1$ means modality $A$ outperforms modality $I$ and vice versa. $\gamma^k, \beta^k \in (0,1)$ are the modulation coefficients. The calculation details for $\gamma^k, \beta^k$ and $\rho_I^k$ are in Appendix. + +Different from [11], we apply the global prototype instead of local prototypes to aggregate the knowledge from different clients, and the ME loss is not only applied to multi-modal clients, but also to uni-modal clients (See details in Sec. 3.2). Now, the weak modality will be stimulated through local update. + +# 3.2 Balanced Modality Selection + +Considering that different types of modal combinations may be selected to participate in local training in our scheme, we define the participated clients as multi-modal clients and uni-modal clients (e.g., client 2 in Fig. 1 Right is a uni-modal client with modal $I$ for training while client 1 and 4 are multi-modal clients with both modalities for training). In Sec. 3.1, we propose the ME loss to facilitate the local learning of weak modality, which is originally designed for multi-modal clients. Here, we also apply it on uni-weak-modal clients to realize gradient consistency on different types of clients. + +Assume modality $I$ is weak here. Hence, the local loss for multi-modal and uni-modal clients should be: + +$$ +\begin{array}{l} \text {m u l t i - m o d a l}: F _ {k} \left(v _ {A}, v _ {I}\right) = \mathcal {L} _ {C E} ^ {k} \left(v _ {A}, v _ {I}\right) + \beta^ {k} \mathcal {L} _ {M E} ^ {k} \left(v _ {I}\right) \\ \text {u n i - m o d a l}: F _ {k} \left(v _ {A}\right) = \mathcal {L} _ {C E} ^ {k} \left(v _ {A}\right), F _ {k} \left(v _ {I}\right) = \mathcal {L} _ {C E} ^ {k} \left(v _ {I}\right) + \beta^ {k} \mathcal {L} _ {M E} ^ {k} \left(v _ {I}\right) \tag {9} \\ \end{array} +$$ + +Next, following Eq. (2), we define the paradigm of modality selection for MFL, aiming to approximate the full gradient aggregation from all clients by the gradient from the customized multi-modal and uni-modal clients: + +$$ +\begin{array}{l} \sum_ {k \in [ N ]} \nabla F _ {k} \left(v _ {A}, v _ {I}\right) = \sum_ {k \in [ N ]} \left[ \begin{array}{c} \nabla F _ {k} \left(v _ {A}, v _ {I}\right) - \nabla F _ {\sigma_ {M} (k)} \left(v _ {A}, v _ {I}\right) \\ - \nabla F _ {\sigma_ {A} (k)} \left(v _ {A}\right) - \nabla F _ {\sigma_ {I} (k)} \left(v _ {I}\right) \end{array} \right] \\ + \sum_ {k \in S _ {M}} \gamma_ {k} ^ {M} \nabla F _ {k} \left(v _ {A}, v _ {I}\right) + \sum_ {k \in S _ {A}} \gamma_ {k} ^ {A} \nabla F _ {k} \left(v _ {A}\right) + \sum_ {k \in S _ {I}} \gamma_ {k} ^ {I} \nabla F _ {k} \left(v _ {I}\right) \tag {10} \\ \end{array} +$$ + +where $\sigma_{M},\sigma_{A}$ and $\sigma_I$ map $V\to S_M,S_A,S_I$ , the client sets who use multi-modal, uni- $A$ , and uni- $I$ data for training respectively, and $S_{M}\cap S_{A} = S_{A}\cap S_{I} = S_{M}\cap S_{I} = \emptyset$ . $\gamma_k^M$ , $\gamma_k^A$ and $\gamma_k^I$ have the similar meaning as $\gamma_{k}$ in Eq. (2). + +Since modality $I$ is weak here, we omit the uni- $A$ clients as the multi-modal gradient is dominated by modality $A$ [28], which means we do not need to select uni- $A$ clients for its enhancement. Then, bring Eq. (9) to Eq. (10) and follow the operations from Eq. (2) to Eq. (3), we can obtain: + +$$ +\begin{array}{l} \sum_ {k \in [ N ]} \min _ {i \in S _ {M}, j \in S _ {I}} \left\| \nabla F _ {k} \left(v _ {A}, v _ {I}\right) - \gamma_ {i} ^ {M} \nabla F _ {i} \left(v _ {A}, v _ {I}\right) - \gamma_ {j} ^ {I} \nabla F _ {j} \left(v _ {I}\right) \right\| \\ = \sum_ {k \in [ N ]} \min _ {i \in S _ {M}, j \in S _ {I}} \left\| \begin{array}{c} \nabla \mathcal {L} _ {C E} ^ {k} (v _ {A}, v _ {I}) + \nabla \beta^ {k} \mathcal {L} _ {M E} ^ {k} (v _ {I}) - \nabla \mathcal {L} _ {C E} ^ {i} (v _ {A}, v _ {I}) \\ - \nabla \beta^ {i} \mathcal {L} _ {M E} ^ {i} (v _ {I}) - \nabla \mathcal {L} _ {C E} ^ {j} (v _ {I}) - \nabla \beta^ {j} \mathcal {L} _ {M E} ^ {j} (v _ {I}) \end{array} \right\| \\ \leqslant \sum_ {k \in [ N ]} \min _ {i \in S _ {M}} \left\| \nabla \mathcal {L} _ {C E} ^ {k} \left(v _ {A}, v _ {I}\right) - \nabla \mathcal {L} _ {C E} ^ {i} \left(v _ {A}, v _ {I}\right) \right\| \tag {11} \\ + \sum_ {k \in [ N ]} \min _ {i \in S _ {M}, j \in S _ {I}} \left\| \begin{array}{c} \nabla \beta^ {k} \mathcal {L} _ {M E} ^ {k} (v _ {I}) - \nabla \beta^ {i} \mathcal {L} _ {M E} ^ {i} (v _ {I}) \\ - \nabla \mathcal {L} _ {C E} ^ {j} (v _ {I}) - \nabla \beta^ {j} \mathcal {L} _ {M E} ^ {j} (v _ {I}) \end{array} \right\| \\ \triangleq G \left(S _ {M}\right) + G \left(S _ {M} \cup S _ {I}\right) \\ \end{array} +$$ + +The right-hand side of the first equation is the modality selection formula that aims to select a group of multi-modal clients $S_{M}$ and uni-modal clients + +Algorithm 1: BMSFed. +Input: Input data $\mathcal{D}_i = \{\pmb {X}_i^A,\pmb {X}_i^I,\pmb {y}\} ,i\in [N],$ initial model $\theta$ hyper-parameters $\chi$ global communication epochs $E$ $e = 1$ while $e < E$ do if $e = 1$ then Send $\theta$ to all clients; Perform one-step local update for gradients, prototypes and $\rho_I^k$ Aggregate global prototypes and $\rho_{I}$ else $\begin{array}{rl} & {\mathrm{Aggrecate~global~model~}\theta ,\mathrm{prototypes~}c^{G}\mathrm{~and~}\rho_{I};}\\ & {\mathrm{S e l e c t~multi - mod a l i t y~for~}S_{M}\mathrm{~a n d~u n i - m o d a l i t y~for~}S_{I}\mathrm{~(o r~}S_{A})\mathrm{~u s i n g}}\\ & {\mathrm{E q s.~(12)~a n d~(13);}}\\ & {\mathrm{S e n d~}\theta ,c^{G}\mathrm{~a n d~}\rho_{I}\mathrm{~t o~s e l e c t e d~c l i e n s;}}\\ & {\mathrm{f o r e a c h~c l i e n t~i n s e l e c t e d~c l i e n s i n~p a r a l l e l d o}}\\ & {\mathrm{P r o p e r m~m u l t i - m o d a l ~l a n e a i n g~i n~}S_{M}\mathrm{~a n d~u n i - m o d a l ~l e a n i n g~i n~}S_{I}\mathrm{~(o r}}\\ & {\mathrm{S}_{A})\mathrm{b y w i t h~E q s.~(8)~a n d~(9);}}\\ & {\mathrm{S e n d~g r a d i e n t s,~l o c a l~p r o t e t y p e s a n d~}\rho_{I}^{k}\mathrm{~t o~s e r v e r};} \end{array}$ + +$S_I$ to approximate the aggregated gradients from all clients. However, the joint selection for $S_M$ and $S_I$ is a complex joint optimization problem. Therefore, we decouple this objective into two submodular functions $G(S_M)$ and $G(S_M \cup S_I)$ according to triangle inequality, while the full gradient approximation is divided into two parts: the first part uses selected multi-modal CE gradient to fit fully multi-modal CE gradient aggregation, and the second part approximates the fully multi-modal ME gradient aggregation via selected uni-modal CE gradient and both selected multi- and uni-modal ME gradient. The modality-level gradient decoupling converts the complex joint selection to two simply separated selection problems. + +Although we can solve the two submodular functions with the stochastic greedy algorithm [24], there are still two issues: (1) the selected client according to $G(S_M \cup S_I)$ should be specified whether it is uni-modal client or multi-modal client; (2) the separated selection strategy pays less attention to the global modal bias. To address the two problems, we perform the stochastic greedy algorithm for two submodular functions in parallel and propose a simple yet effective conflict resolution strategy to ensure $S_M \cap S_I = \emptyset$ as well as, more importantly, balance the learning of different modalities on global model: + +$$ +S _ {M} \leftarrow S _ {M} \cup k _ {1} ^ {*}, k _ {1} ^ {*} \in \underset {k \in \operatorname {r a n d} (V \backslash S _ {M} \backslash S _ {I}, \mathrm {s})} {\arg \max } \left[ \bar {G} \left(S _ {M}\right) - \bar {G} \left(\{k \} \cup S _ {M}\right) \right] \tag {12} +$$ + +$$ +\left\{ \begin{array}{l} i f k _ {1} ^ {*} = k _ {2} ^ {*}, S _ {M} \cup k _ {2} ^ {*}; \\ i f k _ {1} ^ {*} \neq k _ {2} ^ {*}, \left\{ \begin{array}{l} S _ {I} \cup k _ {2} ^ {*}, i f \rho_ {I} ^ {k} > \chi \\ S _ {M} \cup k _ {2} ^ {*}, i f \rho_ {I} ^ {k} \leqslant \chi \end{array} \right. \end{array} \right. \tag {13} +$$ + +$$ +k _ {2} ^ {*} \in \operatorname * {a r g m a x} _ {k \in \operatorname {r a n d} (V \setminus S _ {M} \setminus S _ {I}, \mathrm {s})} \left[ \bar {G} \left(S _ {M} \cup S _ {I}\right) - \bar {G} \left(\{k \} \cup S _ {M} \cup S _ {I}\right) \right] +$$ + +For every selection, we randomly sample a subset of clients $s$ . A multi-modal client $k_{1}^{*}$ is selected from $s$ according to $G(S_{M})$ while $k_{2}^{*}$ is also selected from $s$ according to $G(S_{M} \cup S_{I})$ . $k_{1}^{*} = k_{2}^{*}$ means using multi-modal data from this client can contribute most to the global model. When $k_{1}^{*} \neq k_{2}^{*}$ , we allocate it to uni-modal or multi-modal client according to its local imbalance ratio: if it is severely imbalanced, we use its uni-weak-modal data for training and aggregation to alleviate the global modal bias, otherwise we believe that training with its multi-modal data contributes more than uni-modal data. $\chi$ is a hyper-parameter. + +Discussion. (1) Overcoming the modal bias in our method are twofold: the ME loss alleviates imbalance at local side and the selected uni-modal clients further promote balanced learning of global model. Meanwhile, the diversity coming from two submodular functions ensures the representative information for the global model. (2) We assume $I$ is the weak modality above while in practice, we can determine the weak modality before modality selection via the aggregated global imbalance ratio $\rho_{I} = \frac{1}{\sum_{k=1}^{N} n_{k}} \sum_{k=1}^{N} \rho_{I}^{k} \cdot n_{k}$ . Overall, the pseudo-code of BMSFed is provided in Algorithm 1. (3) Only the gradients, prototypes and $\rho_{I}^{k}$ participate in communication, so there is no privacy issue and similar communication overheads as in traditional FL. + +# 4 Evaluation + +# 4.1 Datasets and baselines + +Datasets. We conduct experiments on four datasets: (1) CREMA-D [4] is an audio-visual dataset for emotion recognition task with total six categories for emotional states. (2) AVE [33] is an audio-visual dataset for event localization with 28 event classes, and here we use it to construct a labeled multimodal classification dataset following [11]. (3) Colored-and-gray MNIST (CG-MNIST) [17] is a synthetic dataset based on MNIST [18] with gray-scale and monochromatic images as two modalities, following [38]. (4) ModelNet40 is one of the Princeton ModelNet datasets [39] with 3D objects of 40 categories. The front and back [31] views are considered as two modalities, following [5]. Details of these four datasets are in the Appendix. + +Baselines. We choose eight baselines for comparison from four categories: (1) three uni-modal FL methods designed for statistical heterogeneity are extended to multi-modal scenarios: FedAvg [23], FedProx [21] and FedProto [32]. (2) Integrating OGM-GE [28] and PMR [11], the solutions for modality imbalance, with FedAvg forms two MFL methods: FedOGM and FedPMR. (3) Two client selection method: Power-of-choice (pow-d) [6] and DivFL [3], evolved from its uni-modal version directly. (4) One MFL method, FedMSplit [5], especially designed for modality incongruity. Compared with these baselines, we demonstrate that an elaborate modality selection strategy is essential to realize comprehensive information exploitation in MFL. + +Table 2: Comparison results on four datasets. The metric is the top-1 accuracy (\%). The best is in **bold**, and the second best is **underlined**. Our method achieves significant improvement on both IID and non-IID scenarios. + +
DatasetCREMA-DAVECG-MNISTModelNet40
MethodIIDnon-IIDIIDnon-IIDIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.742.341.787.286.5
FedProx51.049.062.659.942.943.686.987.1
FedProto58.754.061.758.851.551.487.587.2
FedOGM56.956.462.859.357.253.087.687.0
FedPMR55.555.163.161.666.163.387.687.7
pow-d50.550.762.560.041.240.386.886.2
DivFL51.750.863.359.643.042.186.586.4
FedMSplit52.451.662.460.843.550.987.587.4
BMSFed64.561.664.762.170.266.788.787.5
+ +# 4.2 Experimental settings + +For CREMA-D, AVE and ModelNet40, we use ResNet18 [12] as the backbone for audio, visual and flow modalities. Audio data is converted to a spectrogram of size 257x299 for CREMA-D and $257 \times 1,004$ for AVE. We randomly choose 3 frames and 4 frames to build image training sets for CREMA-D and AVE respectively. For CG-MNIST, we build a neural network with 4 convolution layers and 1 average pool layer as the encoder, following the setting as in [11]. We choose the simple yet effective fusion method, concatenation [27], to build fusion classifier for all the datasets. We set 20 clients for CREMA-D, AVE and ModelNet40 while the number for CG-MNIST is 30. 5 clients are selected in each communication round for CREMA-D, AVE, ModelNet40 and 6 for CG-MNIST. For IID setting, training data is uniformly distributed to all clients. For non-IID scenarios, we use Dirichlet distribution [13] $Dir(\alpha)$ to split data ( $\alpha = 3$ for CREMA-D, AVE, ModelNet40, $\alpha = 2$ for CG-MNIST). The optimizer is SGD [30] for all datasets. Learning rate is initialized at 1e-3 or 1e-2 for CEAMA-D, AVE and ModelNet40 or CG-MNIST and becomes 1e-4 or 1e-3 in the later training stage. The hyper-parameter $\chi$ is set to 1.2-2.5 according to datasets and settings. To complete stochastic greedy algorithm for Eqs. (12) and (13), we use the gradients from the selected clients at current round to update part of the similarity matrix, which is named "no-overheads" in [3]. Except for pow-d, DivFL and BMSFed, other baselines select clients randomly. Each client has two modal data by default. We do all experiments on a workstation with an RTX 3090 GPU, a 3.9-GHZ Intel Core i9-12900K CPU and 64GB of RAM. + +# 4.3 Comparison with baselines + +BMSFed effectively improves the performance. As demonstrated in Tab. 2, our BMSFed achieves the best results on the four datasets under both IID and + +![](images/7274319b65aa763598ff2d5925193705ec461b62a41061d3a100987b44781a05.jpg) + +![](images/9b3d7afbe8bc08c2edd2791e64d64337f539822ddbbe88cc01e4b0616ba75aea.jpg) + +![](images/3f9ee423b51e7bc5c78e5f48750c24fdf3f3cc47347e01cf107b65d20afd33c1.jpg) +(a) CREMA-D under IID +(c) CREMA-D under non-HID +Fig. 2: Test accuracy of BMSFed compared with other baselines on CREMA-D and AVE. BMSFed converges to more accurate solutions than all baselines. + +![](images/8247a651b91c8a250e5f9e8347c93ced39ef28b450d87b76185fff7957efb5fd.jpg) +(b) AVE under IID +(d) AVE under non-IID + +non-IID settings (by up to $5.8\%$ on CREMA-D). Client sampling here (powd and DivFL) cannot fully exploit information for all modalities, making its improvement limited or even worse than FedAvg in CG-MNIST. Although FedOGM and FedPMR accomplish modest improvement because of their ability to alleviate modality imbalance, they are not as good as BMSFed since they do not consider the overall performance of the global model. Traditional uni-modal FL methods for statistical heterogeneity (e.g. FedProx) and MFL method for modality incongruity (FedMSplit) only obtain slight improvement. We also illustrate the trend of test accuracy versus the number of communication rounds on CREMA-D and AVE in Fig. 2. BMSFed can realize comparable or even faster convergence speeds in CREMA-D and AVE. + +BMSFed exploits all modalities comprehensively. To show the effect of our method on addressing modal bias, we report the performance of each modality on CREMA-D and AVE as shown in Tab. 3. The uni-modal performance evaluation follows [11]: a sample is classified into the class corresponding to its nearest prototype. It is clear that BMSFed could considerably improve the performance of weak modality (visual) and mitigate the modality-level bias. Besides, compared with randomly modality abandoning, which significantly reduces audio performance as illustrated in Tab. 1, BMSFed achieves comparable audio performance with other baselines. Although FedProto, FedOGM and FedPMR also alleviate the imbalance, they mainly focus on local optimization, resulting in the performance gap between them and our BMSFed on the aggregated model, which further indicates that in MFL, it is important to take both local opti + +Table 3: The uni-modal performance comparison on CREMA-D and AVE. The metric is the top-1 accuracy (\%). Our method achieves the best uni-visual (the weak modality) performance on all settings and maintains the comparable uni-audio performance. + +
DatasetCREMA-DAVE
SettingIIDnon-IIDIIDnon-IID
MethodAVAVAVAV
FedAvg51.220.650.720.261.126.861.426.4
FedProx51.320.250.122.060.427.161.226.9
FedProto50.235.348.639.155.736.859.732.8
FedOGM50.535.748.830.258.728.859.429.4
FedPMR51.538.750.135.961.739.661.735.3
pow-d51.520.451.618.861.926.960.127.1
DivFL52.321.152.122.762.725.361.626.3
FedMSplit52.021.850.821.661.326.962.328.7
BMSFed51.041.949.341.459.740.260.238.6
+ +Table 4: Ablation study. 'BMSFed-local' uses local prototypes rather than global prototypes for ME loss. + +
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.7
DivFL51.750.863.359.6
FedAvg-0.252.450.163.461.1
FedAvg-0.555.755.160.759.4
FedAvg-0.861.258.158.558.7
FedAvg+ME55.854.562.860.7
DivFL+ME57.155.663.061.1
BMSFed-local63.760.363.460.1
BMSFed64.561.664.762.1
+ +mization for each modality and the overall performance for global model into consideration simultaneously. + +# 4.4 Ablation study + +Effectiveness of each component. Tab. 4 studies the effect of each BMSFed component. Applying ME loss Eq. (7) on random sampling $(\mathrm{FedAvg} + \mathrm{ME})$ and the well-designed client selection $(\mathrm{DivFL} + \mathrm{ME})$ surpasses their vanilla strategies $(\mathrm{FedAvg}, \mathrm{DivFL})$ by a large margin, demonstrating its effectiveness on local enhancement. Comparing BMSFed (64.5% on IID CREMA-D) with 'DivFL+ME' (57.1% on the same setting) also denotes the necessity of balancing different modalities considering the global model via modality selection. To show the importance of aligning feature spaces of weak modality, we replace the global prototypes with local prototypes (BMSFed-local). Global alignment achieves notable improvement (by up to 2% on non-IID AVE). The performance improvement + +![](images/415f2b607fb3f72dca6976bbcca677abe4bd373b6aee558d7cc1dc0a31073c83.jpg) +(a) CREMA-D under IID + +![](images/dad87f0c4c177161d2c4c78b1c41cc82d49aa19581bf69f9e6edc4c98b807eab.jpg) +(b) CREMA-D under IID + +![](images/f823299f830c257182854f75cfb233c70db6a5d6161f8a3d1929ab37dcbdd28b.jpg) +(c) CG-MNIST under IID + +![](images/3ebb8aae80912edba076b03f45e9c62fa05be448518a3a27122e06eaeefc3413.jpg) +Fig. 3: Robustness validation on data size, local epoch and client number. Our BMSFed consistently outperforms baseline (FedAvg) under various scenarios. +Fig. 4: Proportional change of audio and visual respectively and the curve of global imbalance ratio during training on CREMA-D under IID setting. + +compared with 'FedAvg-0.2,0.5,0.8' exhibits that randomly sampling modalities does not always lead to improvement and further demonstrates the need of meticulously selecting modalities for information exploitation. + +Robustness test. To verify the robustness of our method, we vary three key hyperparameters to build various scenarios: (1) change the data size $|\mathcal{D}_i|$ to allow each client to hold a small amount of data, (2) set different local training epochs, and (3) vary the total client number $N$ . As illustrated in Figs. 3a and 3b, our BMSFed consistently outperforms baseline (FedAvg) under various scenarios. More data as well as more local training epochs can bring further improvements to our method, indicating that exploiting the weak modality need more training efforts. Based on the results in Fig. 3c, our approach can also be generalised to clients with larger scales. + +The relationship between modality selection and imbalance degree. In Eqs. (12) and (13), we use a conflict resolution strategy based on local imbalance ratio to realize balanced modality selection. We visualize the proportions of audio and visual modalities selected in each round and the global imbalance ratio. It is clear from Fig. 4 that audio is the dominant modality (imbalance ratio is always greater than 1) and modality imbalance is gradually alleviated as training progresses (global imbalance ratio shows a downward trend). In addition, the proportion of selected visual modality follows the same trend (larger in the early stage of training and becomes smaller later), implying the rationality of our + +Table 5: Performance on CREMA-D and AVE with modality incongruity. $50\%$ of clients have all modal data and $50\%$ of clients only retain data with a single modality (audio or visual). + +
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg55.755.160.758.7
FedProx56.856.061.258.5
FedProto58.757.061.359.7
FedOGM58.657.460.158.5
FedPMR56.455.561.960.3
DivFL57.455.661.158.6
FedMSplit58.956.961.760.0
BMSFed62.459.863.560.9
+ +selection strategy based on imbalance ratio and its effectiveness on mitigating bias. + +Effectiveness on modality incongruity scenario. All above experiments assume that each client initially has complete modal data. Here, we build the stimulation of modality incongruity scenario, in which half of the clients have data with two modalities, and the other half only have data in one modality: random audio or visual. The results are shown in Tab. 5 (The results of pow-d is not available because it is not applicable to this scenario). Our BMSFed still makes impressive improvement compared with all other baselines (by up to $3.5\%$ on IID CREMA-D), illustrating the good generalization ability of our method in different scenarios. It is worth mentioning that FedMSplit performs better than before because it is specifically designed for modality incongruity. + +# 5 Conclusion + +In this paper, we analyze traditional client selections and find their ineffectiveness in MFL. We further reveal that there exists strong modality-level bias due to the modality imbalance during the training iterations and uni-modal training on some clients may contribute more to the global model than multi-modal training. To address this issue, we propose the balanced modality selection scheme for MFL (BMSFed) with modality-level gradient decoupling to release the potential of all modalities and maximize the gradient diversities to improve global aggregation. We also introduce a modal enhancement loss to optimize the local update process. Our method does not introduce additional local training costs and communication overheads compared with previous methods. Extensive experiments on four datasets demonstrate the superiority of our method in performance and applicability under different modal combinations, data distributions and modality incongruity scenarios. + +# Acknowledgements + +Our work was supported by two grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. PolyU15222621, PolyU15225023) and the funding from National Natural Science Foundation of China under grants 62302184. + +# References + +1. Abavisani, M., Wu, L., Hu, S., Tetreault, J., Jaimes, A.: Multimodal categorization of crisis events in social media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14679-14689 (2020) +2. Alfasly, S., Lu, J., Xu, C., Zou, Y.: Learnable irrelevant modality dropout for multimodal action recognition on modality-specific annotated videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20208-20217 (2022) +3. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (2022) +4. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing 5(4), 377-390 (2014) +5. Chen, J., Zhang, A.: Fedmsplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 87-96 (2022) +6. Cho, Y.J., Wang, J., Joshi, G.: Client selection in federated learning: Convergence analysis and power-of-choice selection strategies. arXiv preprint arXiv:2010.01243 (2020) +7. Cornuejols, G., Fisher, M., Nemhauser, G.L.: On the uncapacitated location problem. In: Annals of Discrete Mathematics, vol. 1, pp. 163-177. Elsevier (1977) +8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023) +9. Deng, Y., Lyu, F., Ren, J., Wu, H., Zhou, Y., Zhang, Y., Shen, X.: Auction: Automated and quality-aware client selection framework for efficient federated learning. IEEE Transactions on Parallel and Distributed Systems 33(8), 1996-2009 (2021) +0. Duan, J.h., Li, W., Zou, D., Li, R., Lu, S.: Federated learning with data-agnostic distribution fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8074–8083 (2023) +11. Fan, Y., Xu, W., Wang, H., Wang, J., Guo, S.: Pmr: Prototypical modal rebalance for multimodal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20029-20038 (2023) +12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +13. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019) +14. Huang, Y., Lin, J., Zhou, C., Yang, H., Huang, L.: Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In: International Conference on Machine Learning. pp. 9226-9259. PMLR (2022) + +15. Imteaj, A., Thakker, U., Wang, S., Li, J., Amini, M.H.: A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal 9(1), 1-24 (2021) +16. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020) +17. Kim, B., Kim, H., Kim, K., Kim, S., Kim, J.: Learning not to learn: Training deep neural networks with biased data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9012-9020 (2019) +18. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) +19. Li, H., Li, X., Hu, P., Lei, Y., Li, C., Zhou, Y.: Boosting multi-modal model performance with adaptive gradient modulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22214-22224 (2023) +20. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020) +21. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020) +22. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189 (2019) +23. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) +24. Mirzasoleiman, B., Badanidiyuru, A., Karbasi, A., Vondrák, J., Krause, A.: Lazier than lazy greedy. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 29 (2015) +25. Nemhauser, G.L., Wolsey, L.A., Fisher, M.L.: An analysis of approximations for maximizing submodular set functions--i. Mathematical programming 14, 265-294 (1978) +26. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020) +27. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European conference on computer vision (ECCV). pp. 631-648 (2018) +28. Peng, X., Wei, Y., Deng, A., Wang, D., Hu, D.: Balanced multimodal learning via on-the-fly gradient modulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8238-8247 (2022) +29. Reisizadeh, A., Tziotis, I., Hassani, H., Mokhtari, A., Pedarsani, R.: Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity. IEEE Journal on Selected Areas in Information Theory 3(2), 197-205 (2022) +30. Robbins, H., Monro, S.: A stochastic approximation method. The annals of mathematical statistics pp. 400-407 (1951) +31. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015) +32. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022) + +33. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 247-263 (2018) +34. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023) +35. Wang, H., Qu, Z., Guo, S., Wang, N., Li, R., Zhuang, W.: Losp: Overlap synchronization parallel with local compensation for fast distributed training. IEEE Journal on Selected Areas in Communications 39(8), 2541-2557 (2021) +36. Wang, H., Xu, W., Fan, Y., Li, R., Zhou, P.: Aocc-fl: Federated learning with aligned overlapping via calibrated compensation. In: IEEE INFOCOM 2023-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2023) +37. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12695-12705 (2020) +38. Wu, N., Jastrzebski, S., Cho, K., Geras, K.J.: Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. In: International Conference on Machine Learning. pp. 24043-24055. PMLR (2022) +39. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015) +40. Xiao, F., Lee, Y.J., Grauman, K., Malik, J., Feichtenhofer, C.: Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740 (2020) +41. Xiong, B., Yang, X., Qi, F., Xu, C.: A unified framework for multi-modal federated learning. Neurocomputing 480, 110-118 (2022) +42. Xu, J., Tong, X., Huang, S.L.: Personalized federated learning with feature alignment and classifier collaboration. arXiv preprint arXiv:2306.11867 (2023) +43. Xu, J., Wang, H.: Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective. IEEE Transactions on Wireless Communications 20(2), 1188-1200 (2020) +44. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023) +45. Zhao, Y., Barnaghi, P., Haddadi, H.: Multimodal federated learning on IoT data. In: 2022 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation (IoTDI). pp. 43-54. IEEE (2022) \ No newline at end of file diff --git a/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/images.zip b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..affad69fa42974c05b8e53d0cda94a6b29b9a544 --- /dev/null +++ b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbb4c4964f51add62cc9215053f8c081e255f36c2d787d01d8d6b059a38b911c +size 610994 diff --git a/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/layout.json b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a0addb029cc915da27e8fc3e743b9b5a09c6f77 --- /dev/null +++ b/2024/Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection/layout.json @@ -0,0 +1,9825 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "Overcome Modal Bias in Multi-modal Federated Learning via Balanced Modality Selection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "text", + "content": "Yunfeng Fan" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "text", + "content": ", Wenchao Xu" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1(\\boxdot)}" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "text", + "content": ", Haozhao Wang" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "text", + "content": ", Fushuo Huo" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "text", + "content": ", Jinyu Chen" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "text", + "content": ", and Song Guo" + }, + { + "bbox": [ + 138, + 168, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 201, + 476, + 257 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 138, + 201, + 473, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 473, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 201, + 473, + 213 + ], + "type": "text", + "content": " Department of Computing, The Hong Kong Polytechnic University, Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 213, + 475, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 475, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 475, + 235 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 213, + 475, + 235 + ], + "type": "text", + "content": " School of Computer Science and Technology, Huazhong University of Science and Technology, Hubei, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 235, + 476, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 476, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 476, + 257 + ], + "type": "text", + "content": "3 Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 143, + 258, + 469, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 469, + 290 + ], + "type": "text", + "content": "{yunfeng.fan, fushuo.huo, jinyu.chen}@connect.polyu.hk \nwenchao.xu@polyu.edu.hk, hz_wang@hust.edu.cn, songguo@cse.ust.hk \nhttps://github.com/fanyunfeng-bit/Balanced-Modality-Selection-in-MFL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 324, + 455, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 324, + 455, + 521 + ], + "spans": [ + { + "bbox": [ + 159, + 324, + 455, + 521 + ], + "type": "text", + "content": "Abstract. Selecting proper clients to participate in each federated learning (FL) round is critical to effectively harness a broad range of distributed data. Existing client selection methods simply consider the mining of distributed uni-modal data, yet, their effectiveness may diminish in multi-modal FL (MFL) as the modality imbalance problem not only impedes the collaborative local training but also leads to a severe global modality-level bias. We empirically reveal that local training with a certain single modality may contribute more to the global model than training with all local modalities. To effectively exploit the distributed multiple modalities, we propose a novel Balanced Modality Selection framework for MFL (BMSFed) to overcome the modal bias. On the one hand, we introduce a modal enhancement loss during local training to alleviate local imbalance based on the aggregated global prototypes. On the other hand, we propose the modality selection aiming to select subsets of local modalities with great diversity and achieving global modal balance simultaneously. Our extensive experiments on audio-visual, colored-gray, and front-back datasets showcase the superiority of BMSFed over baselines and its effectiveness in multi-modal data exploitation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 533, + 460, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 533, + 460, + 555 + ], + "spans": [ + { + "bbox": [ + 159, + 533, + 460, + 555 + ], + "type": "text", + "content": "Keywords: Multi-modal federated learning " + }, + { + "bbox": [ + 159, + 533, + 460, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 533, + 460, + 555 + ], + "type": "text", + "content": " modality imbalance " + }, + { + "bbox": [ + 159, + 533, + 460, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 533, + 460, + 555 + ], + "type": "text", + "content": " modality selection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Federated learning (FL) [23] aims to collaboratively learn data that has been collected by, and resides on, a number of remote devices or servers. FL stands to develop top-performing models by aggregating knowledge from numerous edge clients [8,36], which relies on the iterative interaction among participating clients and the server. However, comprehensively employing the information from all" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 168, + 457, + 284 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 158 + ], + "type": "text", + "content": "Table 1: Performance of various client selection methods in MFL under IID setting. A and V denote uni-audio and uni-visual while A-V means the multi-modal result. 'Local' represents that a client is trained based its local data without aggregation. A strong modal bias of global model exists on the two datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 168, + 457, + 284 + ], + "lines": [ + { + "bbox": [ + 154, + 168, + 457, + 284 + ], + "spans": [ + { + "bbox": [ + 154, + 168, + 457, + 284 + ], + "type": "table", + "html": "
DatasetCREMA-D [4]AVE [33]
MethodAVA-VAVA-V
Local41.920.439.633.416.735.2
FedAvg51.220.650.761.126.862.2
pow-d [6]51.520.450.561.926.962.5
DivFL [3]52.321.151.762.725.363.3
FedAvg-0.250.628.652.460.629.663.4
FedAvg-0.550.534.655.758.730.060.7
FedAvg-0.848.150.961.256.431.858.5
BMSFed51.041.964.559.740.264.7
", + "image_path": "93439485a3d596deee8af2d8086c2d99f2de57511d8c9dd41795e03d8837f1c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 324 + ], + "type": "text", + "content": "clients can be exceptionally difficult due to the client heterogeneity and resource limitations [10, 29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 591 + ], + "type": "text", + "content": "Random sampling [22, 34] from available clients has been widely used in FL to satisfy some practical restrictions, e.g., limited communication bandwidth [26, 35] and computing capacities [15]. To improve the information exploitation of all clients, extensive research has been conducted on effective client selection strategies [9, 43]. Despite the success of traditional client selection methods in uni-modal FL, their effectiveness diminishes when dealing with clients with multi-modal data as the inter-modal interactions during the MFL training are neglected. According to [14, 37], there may exist inconsistent learning paces for different modalities in multi-modal joint training, i.e., modality imbalance, which not only impedes the collaborative local training but also leads to a severe modal bias for global model in MFL. As illustrated in Tab. 1, audio modality significantly outperforms visual modality in CREMA-D and AVE datasets during local training and the aggregated model still suffers from it. However, two well-designed client selection methods (pow-d [6] and DivFL [3]) only obtain severely limited improvement over random sampling (FedAvg) on the multi-modal global model. We can see that client selection methods achieve the best uni-audio performance while uni-visual performance even drops sometimes, which means existing client selection scheme heavily relies on the better modality, while ignoring the importance of improving weak modalities that also has potential for global model aggregation. Based on the above analysis, a pivotal question arises: Can we design a new selection scheme in MFL that can overcome the modal bias and exploit each modality comprehensively?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To answer this question, we investigate the interactions between different modalities via randomly discarding the data from one modality (audio or visual) on part of clients, which is inspired from modality dropout [2, 40] that drops a specific modal data during training for regularization. The results are shown in rows 7-9 in Tab. 1, where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "-x" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " denotes randomly discarding a modality on a client with probability " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". We can see that dropping with a certain prob" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 478, + 227 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 478, + 227 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 478, + 227 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 478, + 227 + ], + "type": "image", + "image_path": "120a71489c7f4011cf5f68f0ab0ba1f38a566c955a660b2fcc2ccf8631830f73.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 291 + ], + "type": "text", + "content": "Fig. 1: Left: Traditional client selection in FL aims to sample a client subset in each round while our modality selection considers each local modality as the sampling unit. Right: The paradigm of BMSFed with four clients. The global prototypes are used to enhance the weak modality during local update. Only networks corresponding to the selected modalities will be uploaded to the server for aggregation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": "ability can improve the global multi-modal performance on both datasets (e.g., FedAvg-0.2), and the main reason comes from the dramatic improvement of visual modality. This phenomenon suggests that performing uni-modal training can unleash its potential without being inhibited by another modality, and uni-modal local training may contribute more to the global model than multi-modal training on some clients. As the dropping ratio increases, although the visual modality still improves, multi-modal performance may decline as the audio modality declines as shown in columns 5 and 7, suggesting that we should carefully control which modalities on each client should be involved in training and aggregation to make the contribution most to the global model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 431, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 610 + ], + "type": "text", + "content": "According to the above investigations, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to mitigating the modal bias and comprehensively exploit the diverse information from all modalities. Specifically, instead of selecting a subset of clients, we treat each modality on the local side as a selection unit, as demonstrated in Fig. 1. Our BMSFed mainly contains two parts: Firstly, we intend to alleviate the local modality imbalance by introducing a modal enhancement loss based on aggregated global prototypes to promote the performance of weak modality. Secondly, we complete the modality selection by building two separated submodular functions for selecting multi-modal clients (training with multi-modal data) and uni-modal clients (training with selected uni-modal data) respectively. Inspired from [3], the criterion is to select modalities that are most representative on the gradients while also alleviate the global modal bias. A simple yet effective conflict resolution strategy is devised to ensure the validity of modality selection and keep modal balance on global model simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 613, + 430, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 613, + 430, + 624 + ], + "spans": [ + { + "bbox": [ + 146, + 613, + 430, + 624 + ], + "type": "text", + "content": "The main contributions of the paper are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 629, + 482, + 665 + ], + "type": "text", + "content": "1. We empirically analyze the modality imbalance problem in MFL and reveal that uni-modal training on some clients may contribute more to the global model than multi-modal training." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 479, + 163 + ], + "type": "text", + "content": "2. Based on the analysis, we propose a novel Balanced Modality Selection scheme for MFL (BMSFed) to comprehensively exploit all modalities via a modal enhancement loss and representative modality selection to overcome the global modal bias." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 163, + 480, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 163, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 135, + 163, + 480, + 199 + ], + "type": "text", + "content": "3. We conduct comprehensive experiments on audio-visual, colored-gray, and front-back datasets, and considering the statistical heterogeneity and modality incongruity problems in MFL, to validate the superiority of our BMSFed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 343, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 343, + 230 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 343, + 230 + ], + "type": "text", + "content": "2 Background and Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 235, + 319, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 319, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 319, + 247 + ], + "type": "text", + "content": "2.1 Multi-modal federated learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": "In MFL, each client has one or multiple modalities of data. Without loss of generality, we consider two input modalities, which are denoted by " + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": ", respectively in MFL. There are a set " + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": " clients, " + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "inline_equation", + "content": "V = [N]" + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": ", respectively owning datasets " + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i = \\left\\{\\mathbf{X}_i^A, \\mathbf{X}_i^I, \\mathbf{y}\\right\\}" + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "inline_equation", + "content": "i \\in [N]" + }, + { + "bbox": [ + 130, + 252, + 479, + 318 + ], + "type": "text", + "content": ". A typical federated learning objective is the average of each client's local loss function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 325, + 480, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 255, + 325, + 480, + 357 + ], + "type": "interline_equation", + "content": "\\min f (\\theta) = \\sum_ {k = 1} ^ {N} p _ {k} F _ {k} (\\theta) \\tag {1}", + "image_path": "3f16accac1d50ecd95f2c774fab0d06eb380204df84116428969ef47554b2fbe.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\theta^A, \\theta^I, \\omega\\}" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": " denotes the model parameters. " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "\\theta^A" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "\\theta^I" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": " represent the encoder parameters of modality " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": " is the parameter of fusion classifier. " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": " is a pre-defined weight. " + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "inline_equation", + "content": "F_k" + }, + { + "bbox": [ + 130, + 354, + 479, + 402 + ], + "type": "text", + "content": " is each client's local loss (cross entropy (CE) loss for classification task in this paper)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 480, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 534 + ], + "type": "text", + "content": "Statistical heterogeneity [16,20] is a widely concerned challenge in uni-modal FL. To tackle this issue, FedProx [21] uses a proximal term to stabilize model aggregation. FedProto [32] shares class prototypes to regularize the learning of local models. In MFL, modality incongruity [44,45] (clients consist of different modalities combinations), as well as statistical heterogeneity [41], are all considered. Yu et al. [44] propose CreamFL to align the representations between different clients and different modalities via communicating knowledge on a public dataset. Chen et al. [5] introduce FedMSplit to split local models into several components and aggregate them by their correlations. However, they still focus on the heterogeneity, but ignore the interaction between private data of different modalities, which limits their information exploitation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 544, + 369, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 369, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 369, + 555 + ], + "type": "text", + "content": "2.2 Client selection and submodular function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 479, + 616 + ], + "type": "text", + "content": "Client selection [9,43] is a critical issue for FL especially when the communication cost with all devices is prohibitively high, which has been extensively studied in uni-modal FL. Cho et al. [6] propose Power-of-Choice to select clients with targets local loss. Balakrishnan et al. [3] propose to select a subset of clients with great diversity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Diverse client selection via submodularity. Maximizing a submodular function is reported to improve the diversity and reduce the redundancy of a subset. This property makes it appropriate for client selection in FL. If a function " + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": " is submodular, it should satisfy: given a finite ground set " + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": " of size" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "F(A \\cup \\{v\\}) - F(A) \\geqslant F(B \\cup \\{v\\}) - F(B)" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": ", for any " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "A \\subseteq B \\subseteq V" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "v \\in V \\setminus B" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": ". The marginal utility of an element " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": " w.r.t. a subset " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "F(v|A) = F(A \\cup \\{v\\}) - F(A)" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": ", which can represent the importance of " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": ". The client selection via submodular maximization can be expressed following [3]: find a subset " + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 115, + 479, + 188 + ], + "type": "text", + "content": " of clients whose aggregated gradients can approximate the full aggregation from all clients:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 198, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 198, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 147, + 198, + 481, + 225 + ], + "type": "interline_equation", + "content": "\\sum_ {k \\in [ N ]} \\nabla F _ {k} (v ^ {k}) = \\sum_ {k \\in [ N ]} \\left[ \\nabla F _ {k} (v ^ {k}) - \\nabla F _ {\\sigma (k)} (v ^ {\\sigma (k)}) \\right] + \\sum_ {k \\in S} \\gamma_ {k} \\nabla F _ {k} (v ^ {k}) \\tag {2}", + "image_path": "613b1404034aed729c503bfcfe7a8bf25ae8b1c128cfe2be515dd3df0e978374.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": " maps " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "V\\to S" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": " and the gradient " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\nabla F_k(v^k)" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": " from client " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": " is approximated by the gradient from a selected client " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\sigma (k)\\in S" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "i\\in S" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "C_i\\triangleq \\{k\\in V|\\sigma (k) = i\\}" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": ", and therefore " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\gamma_{i}\\triangleq |C_{i}|" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": ". Take the norms and apply triangular inequality after subtracting the second term from both sides, we can obtain a relaxed objective " + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "inline_equation", + "content": "G(S)" + }, + { + "bbox": [ + 131, + 235, + 479, + 299 + ], + "type": "text", + "content": " for minimizing the approximation error:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 307, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 307, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 134, + 307, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\left\\| \\sum_ {k \\in [ N ]} \\nabla F _ {k} \\left(v ^ {k}\\right) - \\sum_ {k \\in S} \\gamma_ {k} \\nabla F _ {k} \\left(v ^ {k}\\right) \\right\\| \\leqslant \\sum_ {k \\in [ N ]} \\min _ {i \\in S} \\left\\| \\nabla F _ {k} \\left(v ^ {k}\\right) - \\nabla F _ {i} \\left(v ^ {i}\\right) \\right\\| \\triangleq G (S) \\tag {3}", + "image_path": "c0663fee30603eee641a5e671279d5782a1097c74a3f5fc90a3f3bc5cf94fd79.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 366, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 366, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 479, + 415 + ], + "type": "text", + "content": "Minimizing " + }, + { + "bbox": [ + 131, + 366, + 479, + 415 + ], + "type": "inline_equation", + "content": "G(S)" + }, + { + "bbox": [ + 131, + 366, + 479, + 415 + ], + "type": "text", + "content": " can be seen as maximizing the well-known submodular function, i.e., the facility location function [7]. The submodular maximizing problem is NP-hard but can be approximated via the greedy [25] or stochastic greedy algorithm [24]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 184, + 424, + 480, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 424, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 184, + 424, + 480, + 446 + ], + "type": "interline_equation", + "content": "S \\leftarrow S \\cup k ^ {*}, k ^ {*} \\in \\underset {k \\in \\operatorname {r a n d} (V \\backslash S, \\operatorname {s i z e} = s)} {\\arg \\max } \\left[ \\bar {G} (S) - \\bar {G} (\\{k \\} \\cup S) \\right] \\tag {4}", + "image_path": "6fc53f2ee498697c037ad12ce2e88f6712a94296fd98c4d40c8532cb1859a810.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 456, + 350, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 350, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 350, + 467 + ], + "type": "inline_equation", + "content": "\\bar{G}" + }, + { + "bbox": [ + 131, + 456, + 350, + 467 + ], + "type": "text", + "content": " represents a constant minus the negation of " + }, + { + "bbox": [ + 131, + 456, + 350, + 467 + ], + "type": "inline_equation", + "content": "G" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 468, + 479, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 479, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 479, + 504 + ], + "type": "text", + "content": "Although these methods make great improvement in uni-modal FL, the selection strategy is under-explored in MFL and we reveal that traditional client selection approaches cannot address the severe modal bias in MFL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 517, + 328, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 328, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 328, + 530 + ], + "type": "text", + "content": "2.3 Imbalanced multi-modal learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 618 + ], + "type": "text", + "content": "Modality imbalance indicates the inconsistent learning progress of different modalities in multi-modal learning [14,37]. Peng et al. [28] propose OGM-GE to alleviate the inhibitory effect on weak modality by slowing down the dominant modality. Fan et al. [11] further build a non-parametric classifier by class centroids to adjust the update direction of weak modality. In this paper, we aim to power each modality of all clients by a meticulously designed modality selection strategy in each round of training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 630, + 202, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 202, + 642 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 202, + 642 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 650, + 479, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 650, + 479, + 673 + ], + "spans": [ + { + "bbox": [ + 130, + 650, + 479, + 673 + ], + "type": "text", + "content": "In this section, we introduce BMSFed that contains the local imbalance alleviation and balanced modality selection." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 292, + 91, + 447, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 302, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 302, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 302, + 128 + ], + "type": "text", + "content": "3.1 Local Imbalance Alleviation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 479, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 479, + 182 + ], + "type": "text", + "content": "As discussed in Sec. 2.3 and Tab. 1, the multi-modal training on each client may suffer from severe modality imbalance, leading to inadequate information exploitation and consequently incurring the modal bias on the global model. Therefore, we first try to alleviate the imbalance during local training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "text", + "content": "According to [11], the prototypes (i.e., class centroids) are suitable to calibrate the gradient directions to avoid the interference from other modalities. Therefore, we choose to use prototypes to facilitate learning of weak modality. For the " + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "text", + "content": "-th client with the data " + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i = \\left\\{\\boldsymbol{X}_i^A, \\boldsymbol{X}_i^I, \\boldsymbol{y}\\right\\}" + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "text", + "content": ", the local prototype for class " + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 182, + 480, + 248 + ], + "type": "text", + "content": " of each modality is defined as the mean value of representations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 255, + 481, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 255, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 174, + 255, + 481, + 285 + ], + "type": "interline_equation", + "content": "c _ {i, j} ^ {I} = \\frac {1}{\\left| \\mathcal {D} _ {i , j} \\right|} \\sum_ {x ^ {I} \\in \\mathcal {D} _ {i, j}} h _ {i} \\left(\\theta_ {i} ^ {I}; x ^ {I}\\right), c _ {i, j} ^ {A} = \\frac {1}{\\left| \\mathcal {D} _ {i , j} \\right|} \\sum_ {x ^ {A} \\in \\mathcal {D} _ {i, j}} h _ {i} \\left(\\theta_ {i} ^ {A}; x ^ {A}\\right) \\tag {5}", + "image_path": "cdd51be603a484871ba37e7b97cf6fe9ae2988a26a85ba62a11b14c3674e7439.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{i,j}" + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "text", + "content": " denotes the samples belonging to " + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "text", + "content": "-th class in client " + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 130, + 290, + 479, + 327 + ], + "type": "text", + "content": " is the function of encoder. Considering the heterogeneity across clients, we aggregate the local prototypes to a global prototype as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 334, + 481, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 334, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 194, + 334, + 481, + 365 + ], + "type": "interline_equation", + "content": "c _ {j} ^ {G I} = \\frac {1}{| \\mathcal {N} _ {j} |} \\sum_ {i \\in \\mathcal {N} _ {j}} \\frac {| \\mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {I}, c _ {j} ^ {G A} = \\frac {1}{| \\mathcal {N} _ {j} |} \\sum_ {i \\in \\mathcal {N} _ {j}} \\frac {| \\mathcal {D} _ {i , j} |}{N _ {j}} c _ {i, j} ^ {A} \\tag {6}", + "image_path": "806107fd8e7441cdc6c8d27d559ec9ddc5c6a218ec93fac86daeacb60e16ec56.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_j" + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "text", + "content": " denotes the set of clients that have class " + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "text", + "content": " is the number of instances belonging to class " + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 369, + 479, + 407 + ], + "type": "text", + "content": " over all clients. Then, we introduce a modal enhancement loss (ME) based on global prototype to adjust local training:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 182, + 415, + 481, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 415, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 182, + 415, + 481, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M E} ^ {k} \\left(v _ {I} ^ {k}\\right) = - \\mathbb {E} _ {\\left(x _ {i} ^ {I}, y\\right) \\in \\mathcal {D} _ {k}} \\log \\left[ \\frac {\\exp \\left(- d \\left(z _ {i} ^ {I} , c _ {y} ^ {G I}\\right)\\right)}{\\sum_ {j = 1} ^ {Y} \\exp \\left(- d \\left(z _ {i} ^ {I} , c _ {j} ^ {G I}\\right)\\right)} \\right] \\tag {7}", + "image_path": "eae2345f47fa19e727e95a9feed6b9bb6a558e1e39f1962ba3da5fbce9f2b30b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 181, + 449, + 432, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 449, + 432, + 482 + ], + "spans": [ + { + "bbox": [ + 181, + 449, + 432, + 482 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M E} ^ {k} \\left(v _ {A} ^ {k}\\right) = - \\mathbb {E} _ {\\left(x _ {i} ^ {A}, y\\right) \\in \\mathcal {D} _ {k}} \\log \\left[ \\frac {\\exp \\left(- d \\left(z _ {i} ^ {A} , c _ {y} ^ {G A}\\right)\\right)}{\\sum_ {j = 1} ^ {Y} \\exp \\left(- d \\left(z _ {i} ^ {A} , c _ {j} ^ {G A}\\right)\\right)} \\right]", + "image_path": "842e739b5a63781bf0229af0a0f57b5c8c828b8eeb0311205f2eb1482634f1f1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": " is the distance function (Euclidean distance), " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "z_i^I" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": " is the representation of " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "x_i^I" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "z_i^I = h_i(\\theta_i^I; x_i^I)" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": " is the class number. " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "v_A^k" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "v_I^k" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": " indicate the corresponding modal data in client " + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 487, + 479, + 536 + ], + "type": "text", + "content": ". Hence, the local loss should be (data superscripts are omitted for simplicity):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 194, + 544, + 481, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 544, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 194, + 544, + 481, + 570 + ], + "type": "interline_equation", + "content": "F _ {k} \\left(v _ {A}, v _ {I}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\gamma^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {A}\\right), \\rho_ {I} ^ {k} \\leqslant 1 \\\\ \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right), \\rho_ {I} ^ {k} > 1 \\end{array} \\right. \\tag {8}", + "image_path": "25f082174c192850e414ef464873334147e738bf57d34e0fc58eb33319a0eb7d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\rho_I^k" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " indicates the local imbalance ratio for client " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " calculated based on the modality-wise ground truth prediction, and " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\rho_I^k > 1" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " means modality " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " outperforms modality " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " and vice versa. " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\gamma^k, \\beta^k \\in (0,1)" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " are the modulation coefficients. The calculation details for " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\gamma^k, \\beta^k" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\rho_I^k" + }, + { + "bbox": [ + 130, + 577, + 479, + 624 + ], + "type": "text", + "content": " are in Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 625, + 479, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 625, + 479, + 673 + ], + "spans": [ + { + "bbox": [ + 130, + 625, + 479, + 673 + ], + "type": "text", + "content": "Different from [11], we apply the global prototype instead of local prototypes to aggregate the knowledge from different clients, and the ME loss is not only applied to multi-modal clients, but also to uni-modal clients (See details in Sec. 3.2). Now, the weak modality will be stimulated through local update." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 305, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 305, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 305, + 128 + ], + "type": "text", + "content": "3.2 Balanced Modality Selection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 130, + 482, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 130, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 130, + 482, + 225 + ], + "type": "text", + "content": "Considering that different types of modal combinations may be selected to participate in local training in our scheme, we define the participated clients as multi-modal clients and uni-modal clients (e.g., client 2 in Fig. 1 Right is a uni-modal client with modal " + }, + { + "bbox": [ + 130, + 130, + 482, + 225 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 130, + 482, + 225 + ], + "type": "text", + "content": " for training while client 1 and 4 are multi-modal clients with both modalities for training). In Sec. 3.1, we propose the ME loss to facilitate the local learning of weak modality, which is originally designed for multi-modal clients. Here, we also apply it on uni-weak-modal clients to realize gradient consistency on different types of clients." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 225, + 481, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 481, + 249 + ], + "type": "text", + "content": "Assume modality " + }, + { + "bbox": [ + 131, + 225, + 481, + 249 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 225, + 481, + 249 + ], + "type": "text", + "content": " is weak here. Hence, the local loss for multi-modal and uni-modal clients should be:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 254, + 481, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 254, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 148, + 254, + 481, + 286 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {m u l t i - m o d a l}: F _ {k} \\left(v _ {A}, v _ {I}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right) \\\\ \\text {u n i - m o d a l}: F _ {k} \\left(v _ {A}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}\\right), F _ {k} \\left(v _ {I}\\right) = \\mathcal {L} _ {C E} ^ {k} \\left(v _ {I}\\right) + \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} \\left(v _ {I}\\right) \\tag {9} \\\\ \\end{array}", + "image_path": "3ff00bae886cf5e7424220fa99b938d059efea6fe294b8a0b7c326c7879a655f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 292, + 481, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 292, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 481, + 328 + ], + "type": "text", + "content": "Next, following Eq. (2), we define the paradigm of modality selection for MFL, aiming to approximate the full gradient aggregation from all clients by the gradient from the customized multi-modal and uni-modal clients:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 172, + 335, + 480, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 335, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 172, + 335, + 480, + 395 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {k \\in [ N ]} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) = \\sum_ {k \\in [ N ]} \\left[ \\begin{array}{c} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) - \\nabla F _ {\\sigma_ {M} (k)} \\left(v _ {A}, v _ {I}\\right) \\\\ - \\nabla F _ {\\sigma_ {A} (k)} \\left(v _ {A}\\right) - \\nabla F _ {\\sigma_ {I} (k)} \\left(v _ {I}\\right) \\end{array} \\right] \\\\ + \\sum_ {k \\in S _ {M}} \\gamma_ {k} ^ {M} \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) + \\sum_ {k \\in S _ {A}} \\gamma_ {k} ^ {A} \\nabla F _ {k} \\left(v _ {A}\\right) + \\sum_ {k \\in S _ {I}} \\gamma_ {k} ^ {I} \\nabla F _ {k} \\left(v _ {I}\\right) \\tag {10} \\\\ \\end{array}", + "image_path": "f92d22f4cd29af93e8db36f9ae13b979d5b8324a2fde573d801ebe4667cc0670.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "\\sigma_{M},\\sigma_{A}" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "\\sigma_I" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": " map " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "V\\to S_M,S_A,S_I" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": ", the client sets who use multi-modal, uni-" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": ", and uni-" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": " data for training respectively, and " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "S_{M}\\cap S_{A} = S_{A}\\cap S_{I} = S_{M}\\cap S_{I} = \\emptyset" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "\\gamma_k^M" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "\\gamma_k^A" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "\\gamma_k^I" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": " have the similar meaning as " + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 131, + 402, + 481, + 438 + ], + "type": "text", + "content": " in Eq. (2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "text", + "content": "Since modality " + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "text", + "content": " is weak here, we omit the uni-" + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "text", + "content": " clients as the multi-modal gradient is dominated by modality " + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "text", + "content": " [28], which means we do not need to select uni-" + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 437, + 481, + 486 + ], + "type": "text", + "content": " clients for its enhancement. Then, bring Eq. (9) to Eq. (10) and follow the operations from Eq. (2) to Eq. (3), we can obtain:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 148, + 494, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 494, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 148, + 494, + 480, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\nabla F _ {k} \\left(v _ {A}, v _ {I}\\right) - \\gamma_ {i} ^ {M} \\nabla F _ {i} \\left(v _ {A}, v _ {I}\\right) - \\gamma_ {j} ^ {I} \\nabla F _ {j} \\left(v _ {I}\\right) \\right\\| \\\\ = \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\begin{array}{c} \\nabla \\mathcal {L} _ {C E} ^ {k} (v _ {A}, v _ {I}) + \\nabla \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} (v _ {I}) - \\nabla \\mathcal {L} _ {C E} ^ {i} (v _ {A}, v _ {I}) \\\\ - \\nabla \\beta^ {i} \\mathcal {L} _ {M E} ^ {i} (v _ {I}) - \\nabla \\mathcal {L} _ {C E} ^ {j} (v _ {I}) - \\nabla \\beta^ {j} \\mathcal {L} _ {M E} ^ {j} (v _ {I}) \\end{array} \\right\\| \\\\ \\leqslant \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}} \\left\\| \\nabla \\mathcal {L} _ {C E} ^ {k} \\left(v _ {A}, v _ {I}\\right) - \\nabla \\mathcal {L} _ {C E} ^ {i} \\left(v _ {A}, v _ {I}\\right) \\right\\| \\tag {11} \\\\ + \\sum_ {k \\in [ N ]} \\min _ {i \\in S _ {M}, j \\in S _ {I}} \\left\\| \\begin{array}{c} \\nabla \\beta^ {k} \\mathcal {L} _ {M E} ^ {k} (v _ {I}) - \\nabla \\beta^ {i} \\mathcal {L} _ {M E} ^ {i} (v _ {I}) \\\\ - \\nabla \\mathcal {L} _ {C E} ^ {j} (v _ {I}) - \\nabla \\beta^ {j} \\mathcal {L} _ {M E} ^ {j} (v _ {I}) \\end{array} \\right\\| \\\\ \\triangleq G \\left(S _ {M}\\right) + G \\left(S _ {M} \\cup S _ {I}\\right) \\\\ \\end{array}", + "image_path": "2f23a814745772cd775c58de9feb6233db4d3d62d01724881d6e71ef6a3b8d01.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "The right-hand side of the first equation is the modality selection formula that aims to select a group of multi-modal clients " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "S_{M}" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " and uni-modal clients" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 292, + 91, + 448, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 140, + 132, + 465, + 317 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 250, + 130 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 250, + 130 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 250, + 130 + ], + "type": "text", + "content": "Algorithm 1: BMSFed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "lines": [ + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "spans": [ + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": "Input: Input data " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i = \\{\\pmb {X}_i^A,\\pmb {X}_i^I,\\pmb {y}\\} ,i\\in [N]," + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " initial model " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " hyper-parameters " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " global communication epochs " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "e = 1" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "e < E" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " do if " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "e = 1" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " then Send " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " to all clients; Perform one-step local update for gradients, prototypes and " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "\\rho_I^k" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " Aggregate global prototypes and " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "\\rho_{I}" + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "text", + "content": " else " + }, + { + "bbox": [ + 140, + 132, + 465, + 317 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl} & {\\mathrm{Aggrecate~global~model~}\\theta ,\\mathrm{prototypes~}c^{G}\\mathrm{~and~}\\rho_{I};}\\\\ & {\\mathrm{S e l e c t~multi - mod a l i t y~for~}S_{M}\\mathrm{~a n d~u n i - m o d a l i t y~for~}S_{I}\\mathrm{~(o r~}S_{A})\\mathrm{~u s i n g}}\\\\ & {\\mathrm{E q s.~(12)~a n d~(13);}}\\\\ & {\\mathrm{S e n d~}\\theta ,c^{G}\\mathrm{~a n d~}\\rho_{I}\\mathrm{~t o~s e l e c t e d~c l i e n s;}}\\\\ & {\\mathrm{f o r e a c h~c l i e n t~i n s e l e c t e d~c l i e n s i n~p a r a l l e l d o}}\\\\ & {\\mathrm{P r o p e r m~m u l t i - m o d a l ~l a n e a i n g~i n~}S_{M}\\mathrm{~a n d~u n i - m o d a l ~l e a n i n g~i n~}S_{I}\\mathrm{~(o r}}\\\\ & {\\mathrm{S}_{A})\\mathrm{b y w i t h~E q s.~(8)~a n d~(9);}}\\\\ & {\\mathrm{S e n d~g r a d i e n t s,~l o c a l~p r o t e t y p e s a n d~}\\rho_{I}^{k}\\mathrm{~t o~s e r v e r};} \\end{array}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "inline_equation", + "content": "S_I" + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "text", + "content": " to approximate the aggregated gradients from all clients. However, the joint selection for " + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "inline_equation", + "content": "S_M" + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "inline_equation", + "content": "S_I" + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "text", + "content": " is a complex joint optimization problem. Therefore, we decouple this objective into two submodular functions " + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "inline_equation", + "content": "G(S_M)" + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "inline_equation", + "content": "G(S_M \\cup S_I)" + }, + { + "bbox": [ + 130, + 345, + 482, + 464 + ], + "type": "text", + "content": " according to triangle inequality, while the full gradient approximation is divided into two parts: the first part uses selected multi-modal CE gradient to fit fully multi-modal CE gradient aggregation, and the second part approximates the fully multi-modal ME gradient aggregation via selected uni-modal CE gradient and both selected multi- and uni-modal ME gradient. The modality-level gradient decoupling converts the complex joint selection to two simply separated selection problems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 465, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 561 + ], + "type": "text", + "content": "Although we can solve the two submodular functions with the stochastic greedy algorithm [24], there are still two issues: (1) the selected client according to " + }, + { + "bbox": [ + 130, + 465, + 482, + 561 + ], + "type": "inline_equation", + "content": "G(S_M \\cup S_I)" + }, + { + "bbox": [ + 130, + 465, + 482, + 561 + ], + "type": "text", + "content": " should be specified whether it is uni-modal client or multi-modal client; (2) the separated selection strategy pays less attention to the global modal bias. To address the two problems, we perform the stochastic greedy algorithm for two submodular functions in parallel and propose a simple yet effective conflict resolution strategy to ensure " + }, + { + "bbox": [ + 130, + 465, + 482, + 561 + ], + "type": "inline_equation", + "content": "S_M \\cap S_I = \\emptyset" + }, + { + "bbox": [ + 130, + 465, + 482, + 561 + ], + "type": "text", + "content": " as well as, more importantly, balance the learning of different modalities on global model:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 565, + 481, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 565, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 156, + 565, + 481, + 587 + ], + "type": "interline_equation", + "content": "S _ {M} \\leftarrow S _ {M} \\cup k _ {1} ^ {*}, k _ {1} ^ {*} \\in \\underset {k \\in \\operatorname {r a n d} (V \\backslash S _ {M} \\backslash S _ {I}, \\mathrm {s})} {\\arg \\max } \\left[ \\bar {G} \\left(S _ {M}\\right) - \\bar {G} \\left(\\{k \\} \\cup S _ {M}\\right) \\right] \\tag {12}", + "image_path": "40f93ed42905edad9a67d0292b00cd1b474b5fb58d94b2bba3ca1af2132ae51c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 181, + 599, + 481, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 599, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 181, + 599, + 481, + 644 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} i f k _ {1} ^ {*} = k _ {2} ^ {*}, S _ {M} \\cup k _ {2} ^ {*}; \\\\ i f k _ {1} ^ {*} \\neq k _ {2} ^ {*}, \\left\\{ \\begin{array}{l} S _ {I} \\cup k _ {2} ^ {*}, i f \\rho_ {I} ^ {k} > \\chi \\\\ S _ {M} \\cup k _ {2} ^ {*}, i f \\rho_ {I} ^ {k} \\leqslant \\chi \\end{array} \\right. \\end{array} \\right. \\tag {13}", + "image_path": "a46e2f85586e52eaf98f0f79aa3582ade2bdd8bf21ead216b050c03775e37f2c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 179, + 646, + 433, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 646, + 433, + 667 + ], + "spans": [ + { + "bbox": [ + 179, + 646, + 433, + 667 + ], + "type": "interline_equation", + "content": "k _ {2} ^ {*} \\in \\operatorname * {a r g m a x} _ {k \\in \\operatorname {r a n d} (V \\setminus S _ {M} \\setminus S _ {I}, \\mathrm {s})} \\left[ \\bar {G} \\left(S _ {M} \\cup S _ {I}\\right) - \\bar {G} \\left(\\{k \\} \\cup S _ {M} \\cup S _ {I}\\right) \\right]", + "image_path": "677d4201c58690eeb03623ea71fd5b6a382bb2db2764fc6472e9f5dcb3f85a34.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "For every selection, we randomly sample a subset of clients " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ". A multi-modal client " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "k_{1}^{*}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " is selected from " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "G(S_{M})" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "k_{2}^{*}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " is also selected from " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "G(S_{M} \\cup S_{I})" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "k_{1}^{*} = k_{2}^{*}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " means using multi-modal data from this client can contribute most to the global model. When " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "k_{1}^{*} \\neq k_{2}^{*}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", we allocate it to uni-modal or multi-modal client according to its local imbalance ratio: if it is severely imbalanced, we use its uni-weak-modal data for training and aggregation to alleviate the global modal bias, otherwise we believe that training with its multi-modal data contributes more than uni-modal data. " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " is a hyper-parameter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "text", + "content": "Discussion. (1) Overcoming the modal bias in our method are twofold: the ME loss alleviates imbalance at local side and the selected uni-modal clients further promote balanced learning of global model. Meanwhile, the diversity coming from two submodular functions ensures the representative information for the global model. (2) We assume " + }, + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "text", + "content": " is the weak modality above while in practice, we can determine the weak modality before modality selection via the aggregated global imbalance ratio " + }, + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\rho_{I} = \\frac{1}{\\sum_{k=1}^{N} n_{k}} \\sum_{k=1}^{N} \\rho_{I}^{k} \\cdot n_{k}" + }, + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "text", + "content": ". Overall, the pseudo-code of BMSFed is provided in Algorithm 1. (3) Only the gradients, prototypes and " + }, + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\rho_{I}^{k}" + }, + { + "bbox": [ + 130, + 213, + 482, + 338 + ], + "type": "text", + "content": " participate in communication, so there is no privacy issue and similar communication overheads as in traditional FL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 363, + 218, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 218, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 218, + 376 + ], + "type": "text", + "content": "4 Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 396, + 275, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 275, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 275, + 408 + ], + "type": "text", + "content": "4.1 Datasets and baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 424, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 544 + ], + "type": "text", + "content": "Datasets. We conduct experiments on four datasets: (1) CREMA-D [4] is an audio-visual dataset for emotion recognition task with total six categories for emotional states. (2) AVE [33] is an audio-visual dataset for event localization with 28 event classes, and here we use it to construct a labeled multimodal classification dataset following [11]. (3) Colored-and-gray MNIST (CG-MNIST) [17] is a synthetic dataset based on MNIST [18] with gray-scale and monochromatic images as two modalities, following [38]. (4) ModelNet40 is one of the Princeton ModelNet datasets [39] with 3D objects of 40 categories. The front and back [31] views are considered as two modalities, following [5]. Details of these four datasets are in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Baselines. We choose eight baselines for comparison from four categories: (1) three uni-modal FL methods designed for statistical heterogeneity are extended to multi-modal scenarios: FedAvg [23], FedProx [21] and FedProto [32]. (2) Integrating OGM-GE [28] and PMR [11], the solutions for modality imbalance, with FedAvg forms two MFL methods: FedOGM and FedPMR. (3) Two client selection method: Power-of-choice (pow-d) [6] and DivFL [3], evolved from its uni-modal version directly. (4) One MFL method, FedMSplit [5], especially designed for modality incongruity. Compared with these baselines, we demonstrate that an elaborate modality selection strategy is essential to realize comprehensive information exploitation in MFL." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 485, + 285 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 2: Comparison results on four datasets. The metric is the top-1 accuracy (\\%). The best is in **bold**, and the second best is **underlined**. Our method achieves significant improvement on both IID and non-IID scenarios." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 485, + 285 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 485, + 285 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 485, + 285 + ], + "type": "table", + "html": "
DatasetCREMA-DAVECG-MNISTModelNet40
MethodIIDnon-IIDIIDnon-IIDIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.742.341.787.286.5
FedProx51.049.062.659.942.943.686.987.1
FedProto58.754.061.758.851.551.487.587.2
FedOGM56.956.462.859.357.253.087.687.0
FedPMR55.555.163.161.666.163.387.687.7
pow-d50.550.762.560.041.240.386.886.2
DivFL51.750.863.359.643.042.186.586.4
FedMSplit52.451.662.460.843.550.987.587.4
BMSFed64.561.664.762.170.266.788.787.5
", + "image_path": "70d9710578f202fcb8609f7f2b60aaf80944a7d16374a053242d212d1a40e42a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 308, + 270, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 270, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 270, + 321 + ], + "type": "text", + "content": "4.2 Experimental settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "text", + "content": "For CREMA-D, AVE and ModelNet40, we use ResNet18 [12] as the backbone for audio, visual and flow modalities. Audio data is converted to a spectrogram of size 257x299 for CREMA-D and " + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "inline_equation", + "content": "257 \\times 1,004" + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "text", + "content": " for AVE. We randomly choose 3 frames and 4 frames to build image training sets for CREMA-D and AVE respectively. For CG-MNIST, we build a neural network with 4 convolution layers and 1 average pool layer as the encoder, following the setting as in [11]. We choose the simple yet effective fusion method, concatenation [27], to build fusion classifier for all the datasets. We set 20 clients for CREMA-D, AVE and ModelNet40 while the number for CG-MNIST is 30. 5 clients are selected in each communication round for CREMA-D, AVE, ModelNet40 and 6 for CG-MNIST. For IID setting, training data is uniformly distributed to all clients. For non-IID scenarios, we use Dirichlet distribution [13] " + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "inline_equation", + "content": "Dir(\\alpha)" + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "text", + "content": " to split data (" + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\alpha = 3" + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "text", + "content": " for CREMA-D, AVE, ModelNet40, " + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\alpha = 2" + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "text", + "content": " for CG-MNIST). The optimizer is SGD [30] for all datasets. Learning rate is initialized at 1e-3 or 1e-2 for CEAMA-D, AVE and ModelNet40 or CG-MNIST and becomes 1e-4 or 1e-3 in the later training stage. The hyper-parameter " + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 132, + 332, + 482, + 595 + ], + "type": "text", + "content": " is set to 1.2-2.5 according to datasets and settings. To complete stochastic greedy algorithm for Eqs. (12) and (13), we use the gradients from the selected clients at current round to update part of the similarity matrix, which is named \"no-overheads\" in [3]. Except for pow-d, DivFL and BMSFed, other baselines select clients randomly. Each client has two modal data by default. We do all experiments on a workstation with an RTX 3090 GPU, a 3.9-GHZ Intel Core i9-12900K CPU and 64GB of RAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 617, + 294, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 294, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 294, + 629 + ], + "type": "text", + "content": "4.3 Comparison with baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "BMSFed effectively improves the performance. As demonstrated in Tab. 2, our BMSFed achieves the best results on the four datasets under both IID and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 292, + 205 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 292, + 205 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 292, + 205 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 292, + 205 + ], + "type": "image", + "image_path": "7274319b65aa763598ff2d5925193705ec461b62a41061d3a100987b44781a05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 449, + 205 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 449, + 205 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 449, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 449, + 205 + ], + "type": "image", + "image_path": "9b3d7afbe8bc08c2edd2791e64d64337f539822ddbbe88cc01e4b0616ba75aea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 217, + 290, + 308 + ], + "blocks": [ + { + "bbox": [ + 181, + 207, + 272, + 217 + ], + "lines": [ + { + "bbox": [ + 181, + 207, + 272, + 217 + ], + "spans": [ + { + "bbox": [ + 181, + 207, + 272, + 217 + ], + "type": "text", + "content": "(a) CREMA-D under IID" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 217, + 290, + 308 + ], + "lines": [ + { + "bbox": [ + 149, + 217, + 290, + 308 + ], + "spans": [ + { + "bbox": [ + 149, + 217, + 290, + 308 + ], + "type": "image", + "image_path": "3f9ee423b51e7bc5c78e5f48750c24fdf3f3cc47347e01cf107b65d20afd33c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 309, + 280, + 318 + ], + "lines": [ + { + "bbox": [ + 174, + 309, + 280, + 318 + ], + "spans": [ + { + "bbox": [ + 174, + 309, + 280, + 318 + ], + "type": "text", + "content": "(c) CREMA-D under non-HID" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 327, + 481, + 350 + ], + "lines": [ + { + "bbox": [ + 132, + 327, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 481, + 350 + ], + "type": "text", + "content": "Fig. 2: Test accuracy of BMSFed compared with other baselines on CREMA-D and AVE. BMSFed converges to more accurate solutions than all baselines." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 217, + 449, + 308 + ], + "blocks": [ + { + "bbox": [ + 351, + 207, + 421, + 217 + ], + "lines": [ + { + "bbox": [ + 351, + 207, + 421, + 217 + ], + "spans": [ + { + "bbox": [ + 351, + 207, + 421, + 217 + ], + "type": "text", + "content": "(b) AVE under IID" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 217, + 449, + 308 + ], + "lines": [ + { + "bbox": [ + 308, + 217, + 449, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 449, + 308 + ], + "type": "image", + "image_path": "8247a651b91c8a250e5f9e8347c93ced39ef28b450d87b76185fff7957efb5fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 309, + 429, + 318 + ], + "lines": [ + { + "bbox": [ + 343, + 309, + 429, + 318 + ], + "spans": [ + { + "bbox": [ + 343, + 309, + 429, + 318 + ], + "type": "text", + "content": "(d) AVE under non-IID" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "type": "text", + "content": "non-IID settings (by up to " + }, + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 517 + ], + "type": "text", + "content": " on CREMA-D). Client sampling here (powd and DivFL) cannot fully exploit information for all modalities, making its improvement limited or even worse than FedAvg in CG-MNIST. Although FedOGM and FedPMR accomplish modest improvement because of their ability to alleviate modality imbalance, they are not as good as BMSFed since they do not consider the overall performance of the global model. Traditional uni-modal FL methods for statistical heterogeneity (e.g. FedProx) and MFL method for modality incongruity (FedMSplit) only obtain slight improvement. We also illustrate the trend of test accuracy versus the number of communication rounds on CREMA-D and AVE in Fig. 2. BMSFed can realize comparable or even faster convergence speeds in CREMA-D and AVE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "BMSFed exploits all modalities comprehensively. To show the effect of our method on addressing modal bias, we report the performance of each modality on CREMA-D and AVE as shown in Tab. 3. The uni-modal performance evaluation follows [11]: a sample is classified into the class corresponding to its nearest prototype. It is clear that BMSFed could considerably improve the performance of weak modality (visual) and mitigate the modality-level bias. Besides, compared with randomly modality abandoning, which significantly reduces audio performance as illustrated in Tab. 1, BMSFed achieves comparable audio performance with other baselines. Although FedProto, FedOGM and FedPMR also alleviate the imbalance, they mainly focus on local optimization, resulting in the performance gap between them and our BMSFed on the aggregated model, which further indicates that in MFL, it is important to take both local opti" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 157, + 462, + 296 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: The uni-modal performance comparison on CREMA-D and AVE. The metric is the top-1 accuracy (\\%). Our method achieves the best uni-visual (the weak modality) performance on all settings and maintains the comparable uni-audio performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 157, + 462, + 296 + ], + "lines": [ + { + "bbox": [ + 149, + 157, + 462, + 296 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 462, + 296 + ], + "type": "table", + "html": "
DatasetCREMA-DAVE
SettingIIDnon-IIDIIDnon-IID
MethodAVAVAVAV
FedAvg51.220.650.720.261.126.861.426.4
FedProx51.320.250.122.060.427.161.226.9
FedProto50.235.348.639.155.736.859.732.8
FedOGM50.535.748.830.258.728.859.429.4
FedPMR51.538.750.135.961.739.661.735.3
pow-d51.520.451.618.861.926.960.127.1
DivFL52.321.152.122.762.725.361.626.3
FedMSplit52.021.850.821.661.326.962.328.7
BMSFed51.041.949.341.459.740.260.238.6
", + "image_path": "a57ca85f65fb0714fb6281faa536afdfb2da60e50bbd33c90ccd56c9bde76091.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 176, + 337, + 436, + 463 + ], + "blocks": [ + { + "bbox": [ + 132, + 305, + 481, + 327 + ], + "lines": [ + { + "bbox": [ + 132, + 305, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 481, + 327 + ], + "type": "text", + "content": "Table 4: Ablation study. 'BMSFed-local' uses local prototypes rather than global prototypes for ME loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 337, + 436, + 463 + ], + "lines": [ + { + "bbox": [ + 176, + 337, + 436, + 463 + ], + "spans": [ + { + "bbox": [ + 176, + 337, + 436, + 463 + ], + "type": "table", + "html": "
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg50.749.862.259.7
DivFL51.750.863.359.6
FedAvg-0.252.450.163.461.1
FedAvg-0.555.755.160.759.4
FedAvg-0.861.258.158.558.7
FedAvg+ME55.854.562.860.7
DivFL+ME57.155.663.061.1
BMSFed-local63.760.363.460.1
BMSFed64.561.664.762.1
", + "image_path": "fc3afd223b8a37577ab080260dc2797e123b8e99271b24392990abfbe94de898.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 485, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 481, + 510 + ], + "type": "text", + "content": "mization for each modality and the overall performance for global model into consideration simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 526, + 238, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 238, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 238, + 539 + ], + "type": "text", + "content": "4.4 Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of each component. Tab. 4 studies the effect of each BMSFed component. Applying ME loss Eq. (7) on random sampling " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{FedAvg} + \\mathrm{ME})" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " and the well-designed client selection " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{DivFL} + \\mathrm{ME})" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " surpasses their vanilla strategies " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{FedAvg}, \\mathrm{DivFL})" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " by a large margin, demonstrating its effectiveness on local enhancement. Comparing BMSFed (64.5% on IID CREMA-D) with 'DivFL+ME' (57.1% on the same setting) also denotes the necessity of balancing different modalities considering the global model via modality selection. To show the importance of aligning feature spaces of weak modality, we replace the global prototypes with local prototypes (BMSFed-local). Global alignment achieves notable improvement (by up to 2% on non-IID AVE). The performance improvement" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 239, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 239, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 239, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 239, + 190 + ], + "type": "image", + "image_path": "415f2b607fb3f72dca6976bbcca677abe4bd373b6aee558d7cc1dc0a31073c83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 192, + 237, + 201 + ], + "lines": [ + { + "bbox": [ + 145, + 192, + 237, + 201 + ], + "spans": [ + { + "bbox": [ + 145, + 192, + 237, + 201 + ], + "type": "text", + "content": "(a) CREMA-D under IID" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 115, + 353, + 191 + ], + "blocks": [ + { + "bbox": [ + 249, + 115, + 353, + 191 + ], + "lines": [ + { + "bbox": [ + 249, + 115, + 353, + 191 + ], + "spans": [ + { + "bbox": [ + 249, + 115, + 353, + 191 + ], + "type": "image", + "image_path": "dad87f0c4c177161d2c4c78b1c41cc82d49aa19581bf69f9e6edc4c98b807eab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 192, + 353, + 201 + ], + "lines": [ + { + "bbox": [ + 261, + 192, + 353, + 201 + ], + "spans": [ + { + "bbox": [ + 261, + 192, + 353, + 201 + ], + "type": "text", + "content": "(b) CREMA-D under IID" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 364, + 114, + 470, + 191 + ], + "blocks": [ + { + "bbox": [ + 364, + 114, + 470, + 191 + ], + "lines": [ + { + "bbox": [ + 364, + 114, + 470, + 191 + ], + "spans": [ + { + "bbox": [ + 364, + 114, + 470, + 191 + ], + "type": "image", + "image_path": "f823299f830c257182854f75cfb233c70db6a5d6161f8a3d1929ab37dcbdd28b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 192, + 468, + 201 + ], + "lines": [ + { + "bbox": [ + 376, + 192, + 468, + 201 + ], + "spans": [ + { + "bbox": [ + 376, + 192, + 468, + 201 + ], + "type": "text", + "content": "(c) CG-MNIST under IID" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 251, + 394, + 358 + ], + "blocks": [ + { + "bbox": [ + 131, + 211, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 211, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 480, + 232 + ], + "type": "text", + "content": "Fig. 3: Robustness validation on data size, local epoch and client number. Our BMSFed consistently outperforms baseline (FedAvg) under various scenarios." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 251, + 394, + 358 + ], + "lines": [ + { + "bbox": [ + 220, + 251, + 394, + 358 + ], + "spans": [ + { + "bbox": [ + 220, + 251, + 394, + 358 + ], + "type": "image", + "image_path": "3ebb8aae80912edba076b03f45e9c62fa05be448518a3a27122e06eaeefc3413.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 369, + 480, + 391 + ], + "lines": [ + { + "bbox": [ + 131, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 480, + 391 + ], + "type": "text", + "content": "Fig. 4: Proportional change of audio and visual respectively and the curve of global imbalance ratio during training on CREMA-D under IID setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 443 + ], + "type": "text", + "content": "compared with 'FedAvg-0.2,0.5,0.8' exhibits that randomly sampling modalities does not always lead to improvement and further demonstrates the need of meticulously selecting modalities for information exploitation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 447, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 554 + ], + "type": "text", + "content": "Robustness test. To verify the robustness of our method, we vary three key hyperparameters to build various scenarios: (1) change the data size " + }, + { + "bbox": [ + 130, + 447, + 482, + 554 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_i|" + }, + { + "bbox": [ + 130, + 447, + 482, + 554 + ], + "type": "text", + "content": " to allow each client to hold a small amount of data, (2) set different local training epochs, and (3) vary the total client number " + }, + { + "bbox": [ + 130, + 447, + 482, + 554 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 447, + 482, + 554 + ], + "type": "text", + "content": ". As illustrated in Figs. 3a and 3b, our BMSFed consistently outperforms baseline (FedAvg) under various scenarios. More data as well as more local training epochs can bring further improvements to our method, indicating that exploiting the weak modality need more training efforts. Based on the results in Fig. 3c, our approach can also be generalised to clients with larger scales." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The relationship between modality selection and imbalance degree. In Eqs. (12) and (13), we use a conflict resolution strategy based on local imbalance ratio to realize balanced modality selection. We visualize the proportions of audio and visual modalities selected in each round and the global imbalance ratio. It is clear from Fig. 4 that audio is the dominant modality (imbalance ratio is always greater than 1) and modality imbalance is gradually alleviated as training progresses (global imbalance ratio shows a downward trend). In addition, the proportion of selected visual modality follows the same trend (larger in the early stage of training and becomes smaller later), implying the rationality of our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 293, + 91, + 448, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 157, + 428, + 283 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Performance on CREMA-D and AVE with modality incongruity. " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " of clients have all modal data and " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " of clients only retain data with a single modality (audio or visual)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 157, + 428, + 283 + ], + "lines": [ + { + "bbox": [ + 182, + 157, + 428, + 283 + ], + "spans": [ + { + "bbox": [ + 182, + 157, + 428, + 283 + ], + "type": "table", + "html": "
DatasetCREMA-DAVE
settingIIDnon-IIDIIDnon-IID
FedAvg55.755.160.758.7
FedProx56.856.061.258.5
FedProto58.757.061.359.7
FedOGM58.657.460.158.5
FedPMR56.455.561.960.3
DivFL57.455.661.158.6
FedMSplit58.956.961.760.0
BMSFed62.459.863.560.9
", + "image_path": "c6ca3d171ef9610b6c518f6857c57e5296518ac806c97c4d8b52f5928be35af8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 308, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 331 + ], + "type": "text", + "content": "selection strategy based on imbalance ratio and its effectiveness on mitigating bias." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 333, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 453 + ], + "type": "text", + "content": "Effectiveness on modality incongruity scenario. All above experiments assume that each client initially has complete modal data. Here, we build the stimulation of modality incongruity scenario, in which half of the clients have data with two modalities, and the other half only have data in one modality: random audio or visual. The results are shown in Tab. 5 (The results of pow-d is not available because it is not applicable to this scenario). Our BMSFed still makes impressive improvement compared with all other baselines (by up to " + }, + { + "bbox": [ + 132, + 333, + 481, + 453 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 132, + 333, + 481, + 453 + ], + "type": "text", + "content": " on IID CREMA-D), illustrating the good generalization ability of our method in different scenarios. It is worth mentioning that FedMSplit performs better than before because it is specifically designed for modality incongruity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 478, + 219, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 478, + 219, + 490 + ], + "spans": [ + { + "bbox": [ + 133, + 478, + 219, + 490 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we analyze traditional client selections and find their ineffectiveness in MFL. We further reveal that there exists strong modality-level bias due to the modality imbalance during the training iterations and uni-modal training on some clients may contribute more to the global model than multi-modal training. To address this issue, we propose the balanced modality selection scheme for MFL (BMSFed) with modality-level gradient decoupling to release the potential of all modalities and maximize the gradient diversities to improve global aggregation. We also introduce a modal enhancement loss to optimize the local update process. Our method does not introduce additional local training costs and communication overheads compared with previous methods. Extensive experiments on four datasets demonstrate the superiority of our method in performance and applicability under different modal combinations, data distributions and modality incongruity scenarios." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "type": "text", + "content": "Our work was supported by two grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. PolyU15222621, PolyU15225023) and the funding from National Natural Science Foundation of China under grants 62302184." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 224, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "1. Abavisani, M., Wu, L., Hu, S., Tetreault, J., Jaimes, A.: Multimodal categorization of crisis events in social media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14679-14689 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "type": "text", + "content": "2. Alfasly, S., Lu, J., Xu, C., Zou, Y.: Learnable irrelevant modality dropout for multimodal action recognition on modality-specific annotated videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20208-20217 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "text", + "content": "3. Balakrishnan, R., Li, T., Zhou, T., Himayat, N., Smith, V., Bilmes, J.: Diverse client selection for federated learning via submodular maximization. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "text", + "content": "4. Cao, H., Cooper, D.G., Keutmann, M.K., Gur, R.C., Nenkova, A., Verma, R.: Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing 5(4), 377-390 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "text", + "content": "5. Chen, J., Zhang, A.: Fedmsplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 87-96 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 427 + ], + "type": "text", + "content": "6. Cho, Y.J., Wang, J., Joshi, G.: Client selection in federated learning: Convergence analysis and power-of-choice selection strategies. arXiv preprint arXiv:2010.01243 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "type": "text", + "content": "7. Cornuejols, G., Fisher, M., Nemhauser, G.L.: On the uncapacitated location problem. In: Annals of Discrete Mathematics, vol. 1, pp. 163-177. Elsevier (1977)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 450, + 480, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 481 + ], + "type": "text", + "content": "8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "text", + "content": "9. Deng, Y., Lyu, F., Ren, J., Wu, H., Zhou, Y., Zhang, Y., Shen, X.: Auction: Automated and quality-aware client selection framework for efficient federated learning. IEEE Transactions on Parallel and Distributed Systems 33(8), 1996-2009 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 515, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 480, + 546 + ], + "type": "text", + "content": "0. Duan, J.h., Li, W., Zou, D., Li, R., Lu, S.: Federated learning with data-agnostic distribution fusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8074–8083 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 547, + 480, + 578 + ], + "type": "text", + "content": "11. Fan, Y., Xu, W., Wang, H., Wang, J., Guo, S.: Pmr: Prototypical modal rebalance for multimodal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20029-20038 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 579, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 480, + 610 + ], + "type": "text", + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "13. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 664 + ], + "type": "text", + "content": "14. Huang, Y., Lin, J., Zhou, C., Yang, H., Huang, L.: Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In: International Conference on Machine Learning. pp. 9226-9259. PMLR (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Imteaj, A., Thakker, U., Wang, S., Li, J., Amini, M.H.: A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal 9(1), 1-24 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "16. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "17. Kim, B., Kim, H., Kim, K., Kim, S., Kim, J.: Learning not to learn: Training deep neural networks with biased data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9012-9020 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 236 + ], + "type": "text", + "content": "18. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "text", + "content": "19. Li, H., Li, X., Hu, P., Lei, Y., Li, C., Zhou, Y.: Boosting multi-modal model performance with adaptive gradient modulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22214-22224 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "type": "text", + "content": "20. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "type": "text", + "content": "21. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 342 + ], + "type": "text", + "content": "22. Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "23. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "24. Mirzasoleiman, B., Badanidiyuru, A., Karbasi, A., Vondrák, J., Krause, A.: Lazier than lazy greedy. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 29 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "25. Nemhauser, G.L., Wolsey, L.A., Fisher, M.L.: An analysis of approximations for maximizing submodular set functions--i. Mathematical programming 14, 265-294 (1978)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "text", + "content": "26. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 504 + ], + "type": "text", + "content": "27. Owens, A., Efros, A.A.: Audio-visual scene analysis with self-supervised multisensory features. In: Proceedings of the European conference on computer vision (ECCV). pp. 631-648 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "type": "text", + "content": "28. Peng, X., Wei, Y., Deng, A., Wang, D., Hu, D.: Balanced multimodal learning via on-the-fly gradient modulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8238-8247 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 579 + ], + "type": "text", + "content": "29. Reisizadeh, A., Tziotis, I., Hassani, H., Mokhtari, A., Pedarsani, R.: Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity. IEEE Journal on Selected Areas in Information Theory 3(2), 197-205 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 601 + ], + "type": "text", + "content": "30. Robbins, H., Monro, S.: A stochastic approximation method. The annals of mathematical statistics pp. 400-407 (1951)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "31. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "32. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Fan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 501 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 247-263 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "34. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "35. Wang, H., Qu, Z., Guo, S., Wang, N., Li, R., Zhuang, W.: Losp: Overlap synchronization parallel with local compensation for fast distributed training. IEEE Journal on Selected Areas in Communications 39(8), 2541-2557 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "36. Wang, H., Xu, W., Fan, Y., Li, R., Zhou, P.: Aocc-fl: Federated learning with aligned overlapping via calibrated compensation. In: IEEE INFOCOM 2023-IEEE Conference on Computer Communications. pp. 1-10. IEEE (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "text", + "content": "37. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12695-12705 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "38. Wu, N., Jastrzebski, S., Cho, K., Geras, K.J.: Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. In: International Conference on Machine Learning. pp. 24043-24055. PMLR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "39. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1912-1920 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "type": "text", + "content": "40. Xiao, F., Lee, Y.J., Grauman, K., Malik, J., Feichtenhofer, C.: Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 370, + 482, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 391 + ], + "type": "text", + "content": "41. Xiong, B., Yang, X., Qi, F., Xu, C.: A unified framework for multi-modal federated learning. Neurocomputing 480, 110-118 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "text", + "content": "42. Xu, J., Tong, X., Huang, S.L.: Personalized federated learning with feature alignment and classifier collaboration. arXiv preprint arXiv:2306.11867 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "text", + "content": "43. Xu, J., Wang, H.: Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective. IEEE Transactions on Wireless Communications 20(2), 1188-1200 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "text", + "content": "44. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "45. Zhao, Y., Barnaghi, P., Haddadi, H.: Multimodal federated learning on IoT data. In: 2022 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation (IoTDI). pp. 43-54. IEEE (2022)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 293, + 91, + 447, + 102 + ], + "type": "text", + "content": "Balanced Modality Selection for MFL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_content_list.json b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bbeba0bed7ccf336fabbc23d90eb3c08cb06f692 --- /dev/null +++ b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_content_list.json @@ -0,0 +1,1999 @@ +[ + { + "type": "text", + "text": "Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks", + "text_level": 1, + "bbox": [ + 254, + 141, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheeun Hong1 and Kyoung Mu Lee1,2", + "bbox": [ + 359, + 212, + 640, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dept. of ECE & ASRI, {cheeun914, kyoungmu}@snu.ac.kr", + "bbox": [ + 285, + 239, + 715, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ IPAI, Seoul National University", + "bbox": [ + 398, + 253, + 602, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Although quantization has emerged as a promising approach to reducing computational complexity across various high-level vision tasks, it inevitably leads to accuracy loss in image super-resolution (SR) networks. This is due to the significantly divergent feature distributions across different channels and input images of the SR networks, which complicates the selection of a fixed quantization range. Existing works address this distribution mismatch problem by dynamically adapting quantization ranges to the varying distributions during test time. However, such a dynamic adaptation incurs additional computational costs during inference. In contrast, we propose a new quantization-aware training scheme that effectively Overcomes the Distribution Mismatch problem in SR networks without the need for dynamic adaptation. Intuitively, this mismatch can be mitigated by regularizing the distance between the feature and a fixed quantization range. However, we observe that such regularization can conflict with the reconstruction loss during training, negatively impacting SR accuracy. Therefore, we opt to regularize the mismatch only when the gradients of the regularization are aligned with those of the reconstruction loss. Additionally, we introduce a layer-wise weight clipping correction scheme to determine a more suitable quantization range for layer-wise weights. Experimental results demonstrate that our framework effectively reduces the distribution mismatch and achieves state-of-the-art performance with minimal computational overhead. Codes are available at https://github.com/Cheeun/ODM.", + "bbox": [ + 261, + 309, + 743, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image super-resolution $\\cdot$ Network quantization $\\cdot$ Quantization-aware training", + "bbox": [ + 261, + 614, + 743, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 670, + 357, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image super-resolution (SR) is a core low-level vision task aimed at reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. Recent advances in deep learning [7, 11, 31, 36, 37, 43, 45, 54, 55] have led to astonishing achievements in producing high-fidelity images. However, the remarkable performance is based on heavy network architectures that incur significant computational costs, limiting practical viability, such as mobile deployment.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate the computational complexity of neural networks, quantization has emerged as a promising avenue. Network quantization has proven effective in reducing computational costs without significant accuracy loss, particularly in high-level vision tasks,", + "bbox": [ + 212, + 795, + 807, + 839 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/4f6bfbf9bc8581da0579778a9a898eb0a74f74974949eb8bbe82a50b5b528b16.jpg", + "table_caption": [ + "Table 1: Quantization methods on SR. For performance, existing methods rely on channel-wise quantization or input-adaptive module, which incur computational overhead. Our method achieves high accuracy without utilizing channel-wise quantization or input-adaptive modules." + ], + "table_footnote": [], + "table_body": "
MethodChannel-wise QInput-adaptive ModulesPSNRSSIM
PAMS [33]XX29.510.835
DDTB [57]X✓30.970.876
DAQ [18]✓✓31.010.871
OursXX31.500.882
", + "bbox": [ + 316, + 198, + 683, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such as image classification [8, 20, 58]. However, when it comes to quantizing SR networks to lower bit-widths, substantial performance degradation [24] often occurs, posing a persistent and challenging problem to be addressed.", + "bbox": [ + 215, + 306, + 785, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This degradation can be attributed to the significant variance in the feature (activation) distributions of SR networks. The feature distribution of a layer exhibits substantial discrepancies across different channels and images, which makes it difficult to determine a single quantization range for a layer. Early approaches to SR quantization [33] adopt a training scheme to learn the quantization range of each layer. However, despite careful selection, the quantization ranges do not align with the varied values within the channel and image dimensions, which we refer to as distribution mismatch in features.", + "bbox": [ + 215, + 352, + 785, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent approaches aim to address this issue by incorporating dynamic adaptation modules that accommodate the varying distributions. For example, the quantization range is dynamically adjusted by directly leveraging the distribution mean and variance at test time [18] or by employing input-adaptive dynamic modules [57]. Although adapting the quantization function to each image during inference might handle variable distributions, the dynamic modules introduce considerable computational overhead, potentially compromising the computational benefits of quantization.", + "bbox": [ + 215, + 459, + 785, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we propose a novel quantization-aware training framework that addresses the distribution mismatch problem with a loss term that regulates the mismatch in distributions. Although directly minimizing the feature mismatch presents the potential for quantization-friendliness, whether it preserves reconstruction accuracy is questionable. We observe that concurrent optimization with the mismatch regularization and the original reconstruction loss can disrupt the image reconstruction process. Therefore, we introduce a cooperative mismatch regularization strategy, where the mismatch is regulated only when it collaborates harmoniously with the reconstruction loss. To determine the cooperative behavior, we assess the cosine similarity of the gradients from each loss, then we weigh the gradients of mismatch regularization based on this similarity. Consequently, we effectively update the SR network to hold both quantization-friendliness and reconstruction accuracy.", + "bbox": [ + 215, + 566, + 785, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we identify the distribution mismatch among the weights of different layers. We discover that employing a fixed policy to determine the layer-wise weight quantization range [17,18,33,47] can be suboptimal and that a further precise range can be obtained by considering both the current distribution of weights and the distinct tendencies of each layer. Therefore, we additionally incorporate layer-specific variations using a correction parameter for each layer. This strategy allows us to accurately find", + "bbox": [ + 215, + 750, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the quantization range for weights while incurring only a minimal overhead (0.01% additional storage size and no additional bitOPs)—a significantly smaller impact compared to methods that use dynamic modules. Overall, the contributions of our work include the following:", + "bbox": [ + 218, + 146, + 785, + 205 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the first quantization framework to address the distribution mismatch problem in SR networks without dynamic modules, as compared in Table 1. Our framework updates the SR network to be quantization-friendly and accurate simultaneously.", + "- Based on the observations of the distribution mismatch in SR networks, we effectively reduce the mismatch by introducing a cooperative mismatch regularization term and a weight clipping correction term.", + "- Compared to existing approaches to SR quantization, ours achieves state-of-the-art performance with similar or fewer computations." + ], + "bbox": [ + 225, + 215, + 782, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 217, + 375, + 374, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Image Super-Resolution", + "text_level": 1, + "bbox": [ + 217, + 405, + 424, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Convolutional Neural Network (CNN)-based approaches have demonstrated remarkable advancements in the image super-resolution (SR) task [32, 37], but at the cost of substantial computational resources. The intensive computations required by SR networks have spurred interest in developing lightweight SR architectures [10, 22, 23, 28, 54]. Furthermore, various strategies for lightweight SR networks have been explored, including neural architecture search [9, 30, 34, 35, 46], knowledge distillation [22, 23, 53], and pruning [26, 42, 56]. While these methods predominantly focus on reducing network depth or the number of channels, our work specifically aims to lower the precision of floating-point operations through network quantization.", + "bbox": [ + 217, + 430, + 785, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Network Quantization", + "text_level": 1, + "bbox": [ + 217, + 587, + 413, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By mapping 32-bit floating point values of features and weights in convolutional layers to lower-bit representations, network quantization provides a dramatic reduction in computational resources [5,8,14,29,58,59]. Recent works have successfully quantized various networks to low bit-widths with minimal compromise in accuracy [6,12,16, 27,39,49,52]. However, these efforts primarily target high-level vision tasks, whereas networks for low-level vision tasks remain vulnerable to low-bit quantization.", + "bbox": [ + 217, + 612, + 785, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Quantized Super-Resolution Networks", + "text_level": 1, + "bbox": [ + 217, + 724, + 524, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to high-level vision tasks, SR poses different challenges due to its inherent high sensitivity to quantization [24, 40, 48, 51]. Some works have attempted to recover accuracy by modifying the network architecture [2, 25, 51] or by assigning different bits for each image [17, 19, 47] or network stage [38]. However, the primary challenge in quantizing SR networks lies in the vastly distinct feature distributions. To address this, Li et al. [33] adopted a learnable quantization range for different layers. Subsequently,", + "bbox": [ + 217, + 750, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 125 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "recognizing that the distributions vary not only by layer, but also by channel and input, Hong et al. [18] introduced a channel-wise dynamic quantization function. Additionally, Zhong et al. [57] utilized an input-adaptive dynamic module to tailor quantization ranges for each specific input image. However, such dynamic adaptations of quantization functions during test-time incur non-negligible computational overheads. Instead of relying on test-time adaptive modules, our approach focuses on mitigating the feature mismatch before quantization. Our framework reduces the inherent distribution mismatch in SR networks with minimal overhead, accurately quantizing SR networks without dynamic modules. More recently, Qin et al. [44] introduced additional transformation functions in both the forward and backward processes. However, performance degradation is still evident in ultra-low bit (e.g., 2-bit) scenarios.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 334, + 398, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, after a brief introduction to network quantization (Section 3.1), we first analyze the mismatch in the features of SR networks (Section 3.2). Subsequently, we propose a solution to reduce this feature mismatch during training (Section 3.3). Additionally, we examine the mismatch in weights across SR networks (Section 3.4) and introduce a quantization range selection scheme that addresses the weight mismatch (Section 3.5). The overall training process is summarized in Algorithm 1.", + "bbox": [ + 212, + 364, + 787, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 477, + 352, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reduce the heavy computations of convolutional and linear layers in neural networks, the input features (activations) and weights of each convolution/linear layer are quantized to low-bit values [5,8,15,29]. Given the input feature of the $i$ -th convolution/linear layer $\\mathbf{X}_i \\in \\mathbb{R}^{B \\times C \\times H \\times W}$ , where $B, C, H$ , and $W$ represent the dimensions of the input batch, channel, height, and width, respectively, a quantization operator $q(\\cdot)$ quantizes the feature $\\mathbf{X}_i$ with bit-width $b$ :", + "bbox": [ + 212, + 501, + 787, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {X} _ {i}; l, u\\right) = \\operatorname {I n t} \\left(\\frac {\\operatorname {c l i p} \\left(\\boldsymbol {X} _ {i} , l , u\\right) - l}{s}\\right) \\cdot s + l, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 599, + 785, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{clip}(\\cdot ,l,u)$ truncates the input into the quantization range of $[l,u]$ and $s = \\frac{u - l}{2^b - 1}$ . After truncation, the truncated feature is scaled to $[0,2^b -1]$ , then rounded to integer values with $\\mathrm{Int}(\\cdot)$ , and rescaled to range $[l,u]$ .", + "bbox": [ + 212, + 640, + 785, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For activation quantization, the quantization range is defined by $[l_a, u_a]$ . To obtain better quantization ranges in SR networks, the clipping parameters $l_a, u_a$ for each layer are typically learned through quantization-aware training [33, 57]. Since the rounding function is not differentiable, a straight-through estimator (STE) [3] is used to train the clipping parameters in an end-to-end manner. Following [57], we initialize $l_a$ and $u_a$ as the $(100-j)$ -th and $j$ -th percentile values of the feature, averaged among the training data, where $j$ is set to 99 in our experiments to avoid outliers that corrupt the quantization range. For weights, as their distributions tend to be symmetric and the mean approximates zero, we utilize a symmetric quantization function. Thus, for the weight $W$ of each convolution/linear layer, the quantization range is defined by $[-u_w, u_w]$ .", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d92441ffc2ec62642ec83b5e93c005e8a1a3b5ea3320f7b650a5d30bf62c2413.jpg", + "image_caption": [ + "(a) Channel-wise features" + ], + "image_footnote": [], + "bbox": [ + 228, + 150, + 395, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4fb8a124108b508df44192265acd63ee451d89a09dca2232802e0d91f15135cd.jpg", + "image_caption": [ + "(b) Image-wise features" + ], + "image_footnote": [], + "bbox": [ + 398, + 150, + 573, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0caab5668e1ca1407062e3e85501a6508542407dee4bbca90d549c4f25eea568.jpg", + "image_caption": [ + "Figure 1: Distribution mismatch in SR networks. Compared to a classification network (e.g., ResNet-18), an SR network (e.g., EDSR) exhibits significant mismatches within the feature distributions across channel and image dimensions. The large distribution mismatch complicates the selection of an appropriate quantization range. Channels and images from the 2nd layer are randomly selected for visualization. Additional results are available in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 576, + 156, + 774, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Distribution Mismatch in SR Networks", + "text_level": 1, + "bbox": [ + 215, + 393, + 529, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The unfriendliness to quantization in SR networks arises from diverse feature (activation) distributions, as reported in previous studies [18, 33, 57], primarily due to the absence of batch normalization layers in SR networks. As illustrated in Figure 1, where there are notable discrepancies between the channel and image distributions, quantization grids are unnecessarily allocated to regions with minimal feature density. Early SR quantization methods tackled this issue by employing learnable quantization range parameters [33] for each feature. However, even though the quantization-aware training process strives to find the optimal range for each feature, it fails to account for the channel-wise and input-wise variance in distributions. This mismatch results in a large quantization error that can impair SR performance. To deal with these distribution mismatches, existing methods adopt different quantization ranges for each channel [18] or input image [18, 57]. Nevertheless, the test-time adaptation modules used to determine these ranges introduce unwanted computational overhead during inference. Thus, our straightforward solution is to pre-adjust the distributions to be quantization-friendly, thereby eliminating the need for additional adaptations at inference. The following sections will introduce a new quantization-aware training scheme designed to resolve the distribution mismatch problem.", + "bbox": [ + 212, + 419, + 787, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Cooperative Mismatch Regularization", + "text_level": 1, + "bbox": [ + 215, + 698, + 522, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of trying to identify a quantization range capable of accommodating diverse feature distributions, our approach aims to regularize the distribution mismatch beforehand. Obtaining an appropriate quantization range for a feature with high image- and channel-wise variance is difficult, as a certain number of channels or images will invariably be distant from the selected quantization range. In this work, we refer to the total distance of each feature from the selected quantization grid as the mismatch,", + "bbox": [ + 212, + 722, + 787, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM \\left(\\boldsymbol {X} _ {i}\\right) = \\left| \\left| \\boldsymbol {X} _ {i} - q \\left(\\boldsymbol {X} _ {i}; l _ {a}, u _ {a}\\right) \\right| \\right| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 824, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4202b0a395949f8651e631f3770f7db57f4318e0e989692d6c27c8c40890ff9e.jpg", + "image_caption": [ + "(a) Gradient conflict [13]" + ], + "image_footnote": [], + "bbox": [ + 245, + 145, + 411, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0c48f56800665b41efc99d9f91fc9d9a3ea69907a9ecd99714b4a3c7ab8dcdef.jpg", + "image_caption": [ + "Figure 2: Conflict between mismatch regularization and reconstruction loss. Mismatch regularization updates a number of parameters in the contradictory direction to the reconstruction loss, which we refer to as gradient conflict. (b) When the two losses are jointly used, gradient conflict consistently occurs during training, outputting a negative cosine similarity value. (c) We plot the ratio of conflicted gradients during training. Nearly half of the parameters undergo gradient conflict, which indicates that merely combining mismatch regularization with the reconstruction loss can impair SR accuracy. Visualizations are done on EDSR." + ], + "image_footnote": [], + "bbox": [ + 416, + 145, + 584, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/577d111d740facb7db319f01807c78e6686e30d2d7571d279f6ed47a1d1f8f5d.jpg", + "image_caption": [ + "(b) Cosine similarity of two gradients", + "(c) Gradient conflict ratio" + ], + "image_footnote": [], + "bbox": [ + 591, + 146, + 759, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $||\\cdot||_2$ calculates the Frobenius norm. Further analyses of the definition of mismatch are provided in the supplementary material. We can reduce the overall mismatch by directly regularizing the mismatch of each feature to be quantized. The mismatch regularization loss is obtained by summing the mismatch over all quantized features:", + "bbox": [ + 214, + 393, + 787, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M} = \\sum_ {i} ^ {\\# \\text {l a y e r s}} M \\left(\\boldsymbol {X} _ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 463, + 785, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The mismatch regularization loss can be used in line with the original reconstruction loss typically used in the general quantization-aware training pipeline for SR networks:", + "bbox": [ + 214, + 512, + 787, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\mathcal {L} _ {1} \\left(\\mathcal {Q} \\left(\\boldsymbol {I} _ {L R}\\right), \\boldsymbol {I} _ {H R}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 558, + 785, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_1$ loss indicates the $l_{1}$ distance between the reconstructed image using the quantized network $\\mathcal{Q}$ and the ground-truth HR image $I_{HR}$ . Then, the optimization of the parameter $\\theta^t$ is formulated as:", + "bbox": [ + 214, + 579, + 787, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {t + 1} = \\theta^ {t} - \\beta^ {t} \\cdot \\left(\\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right) + \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 632, + 785, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\nabla_{\\theta}\\mathcal{L}_R(\\theta^t)$ denotes the gradient from the original reconstruction loss and $\\nabla_{\\theta}\\mathcal{L}_M(\\theta^t)$ is the gradient from mismatch regularization loss, and $\\beta^{t}$ refers to the learning rate. Updating the network to minimize the mismatch regularization loss will reduce the overall error from feature quantization.", + "bbox": [ + 212, + 657, + 803, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, then a question arises: does reducing the quantization error of each feature lead to improved reconstruction accuracy? The answer is, according to our observation in Figure 2, not necessarily. During the training process, the mismatch regularization loss can collide with the original reconstruction loss. That is, for some parameters, the direction of the gradient from reconstruction loss and that of the mismatch regularization are opposing, referred to as gradient conflict [13]. As in Figure 2b, the cosine similarity of two gradients oscillates between positive and negative values during training, indicating that the directions of two gradients do not converge and the gradient", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 410, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64f1105e59ced2ec4a37d7b67bd46c76c604069032719521e323cb524ff0eb57.jpg", + "image_caption": [ + "(a) Layer-wise QE using max policy" + ], + "image_footnote": [], + "bbox": [ + 248, + 145, + 410, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/761c64155d31b35d0453355c3a91ea0cb50014020ce0fa63629d7ab7f6973068.jpg", + "image_caption": [ + "(b) 4-bit quantization grid", + "Figure 3: Layer-wise variation in error from weight quantization. (a) Quantization error (QE) varies across different layers when a fixed global policy (i.e., max) is used to determine the quantization range, particularly for low bits. For some layers, using max does not effectively serve as a proper policy for quantization range selection. (b) Outliers often dominate the quantization range, leading to quantization grids being wasted on low-density areas. (c) At low bits, quantization grids fail to cover high-density regions adequately. Therefore, the quantization range should be adjusted for certain layers." + ], + "image_footnote": [], + "bbox": [ + 421, + 146, + 581, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/505d8e18b256eb5946f6611527585c4ba7a19fcdb11468ceab4808cef71b9080.jpg", + "image_caption": [ + "(c) 2-bit quantization grid" + ], + "image_footnote": [], + "bbox": [ + 594, + 146, + 754, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "occasionally conflicts. Furthermore, as shown in Figure 2c, the proportion of parameters undergoing gradient conflict is not minor, implying that the regularization loss can severely hinder the reconstruction loss.", + "bbox": [ + 212, + 390, + 787, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We aim to avoid the conflict between these two losses, in other words, to minimize the mismatch as long as it does not hinder the reconstruction loss. Thus, we dismiss the mismatch regularization term when it is not cooperative with reconstruction loss and make more use of it when it is cooperative. Specifically, we determine whether the two losses are cooperative by examining the cosine similarity of the gradients of each loss and then simply weigh the gradient of mismatch regularization by the gradient similarity. Our cooperative mismatch regularization can be formulated as follows:", + "bbox": [ + 212, + 436, + 787, + 541 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {t + 1} = \\theta^ {t} - \\beta^ {t} \\cdot \\left(\\lambda_ {R} \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right) + \\lambda_ {M} \\cdot s i m \\left(\\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right), \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right) \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 550, + 787, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the underlined term, gradient similarity, is defined as $\\text{sim}(\\pmb{v}_a, \\pmb{v}_b) = \\frac{\\cos(\\pmb{v}_a, \\pmb{v}_b) + 1}{2}$ and $\\cos(\\cdot, \\cdot) \\in [-1, 1]$ calculates the cosine similarity between two vectors. $\\lambda_R, \\lambda_M$ are hyper-parameters to balance the two gradients. If the directions of two gradients are similar (i.e., smaller than $90^\\circ$ and closer to $0^\\circ$ ), the gradient similarity is a large value, then the parameter is updated in the direction where mismatch regularization is also substantially considered. On the contrary, if the two gradients point in the other direction (i.e., larger than $90^\\circ$ and closer to $180^\\circ$ ), the two losses restrain each other, and the gradient similarity is close to 0. In this case, we follow the gradient of reconstruction loss. This allows the network to reduce the quantization error cooperatively with the reconstruction error. Details on gradient similarity are in the supplementary material.", + "bbox": [ + 212, + 580, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Distribution Mismatch in Weights", + "text_level": 1, + "bbox": [ + 214, + 755, + 495, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The weight distributions of SR networks have remained relatively unexplored in previous literature. This is because the weights in SR networks typically exhibit bell-shaped distributions, which are considered easier to quantize compared to the long-tailed, input-wise and channel-wise distinct activation distributions. Consequently, many", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Quantization-aware training process of ODM" + ], + "code_body": "Input: Pre-trained 32-bit network $\\mathcal{P}$ \nOutput: Quantized network $\\mathcal{Q}$ \nfor $t = 1,\\dots ,$ # iters do for $i = 1,\\dots ,$ # layers do if $t = 1$ then Initialize activation quantization range $[l_a,u_a]$ Initialize weight quantization range $[-u_w,u_w]$ Given quantization range, obtain $q(\\pmb {X}_i;l_a,u_a)$ using Eq. (1) Given $\\pmb {X}_i$ and $q(\\pmb {X}_i;l_a,u_a)$ , obtain mismatch using Eq. (2) Adjust weight quantization range $[-u_w,u_w]$ using Eq. (9) Given quantization range, obtain $q(\\pmb {W}_i; - u_w,u_w)$ using Eq. (1) Replace $\\pmb {X}_i$ $\\pmb {W}_i$ in $\\mathcal{P}$ with $q(\\pmb {X}_i),q(\\pmb {W}_i)$ Calculate mismatch regularization loss (Eq. (3)) and reconstruction loss (Eq. (4)) Update parameters of $\\mathcal{Q}$ with two losses cooperatively using Eq. (6)", + "bbox": [ + 215, + 162, + 785, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "studies [17, 33, 47] simply adopt max quantization for weights, setting the quantization range with the maximum value of the current weight distribution. However, we notice that this is suboptimal and may contribute to low performance when SR networks are quantized to ultra-low bits (e.g., 2-bit). The issue arises because, in some layers, the outliers are far from the distribution mean, as visualized in Figures 3b and 3c. Thus, when the maximum value is used to determine the quantization range for such distributions, the quantization range is dominated by outliers, with quantization grids not allocated to high-density regions (e.g., near 0). This leads to substantial quantization errors that can accumulate and degrade restoration performance. In the case of 4-bit quantization (Figure 3b), although grids still cover high-density regions to an extent, a number of grids are wasted on low-density areas. The problem intensifies with low-bit (2-bit) quantization (Figure 3c), where quantization errors become significantly larger. This observation underscores the need for careful selection of the quantization range for layer-wise weights, particularly in low-bit quantization scenarios.", + "bbox": [ + 212, + 415, + 789, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Weight Clipping Correction", + "text_level": 1, + "bbox": [ + 215, + 646, + 454, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also, we notice that the error from weight quantization varies across different layers, as shown in Figure 3a. For instance, employing the maximum value as the quantization range can be an effective policy for certain layers, yet this policy proves inadequate for others (e.g., the layer shown in Figure 3c). Given the unique tendency of each layer, applying a uniform global policy for selecting the quantization range across all layers is suboptimal. Existing methods [17, 33, 47] utilize a fixed global policy throughout training to set the quantization range clipping parameter $u_{w}$ as follows:", + "bbox": [ + 212, + 669, + 787, + 776 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nu _ {w} ^ {t} = f \\left(\\boldsymbol {W} ^ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 782, + 785, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the global policy $f(\\cdot)$ is the same function for all layers (e.g., $\\max(\\cdot)$ ). A simple solution to accommodate layer-specific variations is to make the clipping parameter $u_w$", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/601fb6c7c1f802781b422fa3656a700b51a560506fcbca99c9b16709a9c21be2.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons on EDSR of scale 4" + ], + "table_footnote": [ + "† We note that the reported results of DAQ [18] are obtained using EDSR of 32 residual blocks. For a fair comparison, we reproduce EDSR-DAQ using EDSR of 16 residual blocks (i.e., EDSR-baseline)." + ], + "table_body": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
EDSR3232.100.89428.580.78127.560.73626.040.785
EDSR-PAMS431.590.88528.200.77327.320.72825.320.762
EDSR-DAQ431.850.88728.380.77627.420.73225.730.772
EDSR-DDTB431.850.88928.390.77727.440.73225.690.774
EDSR-ODM (Ours)432.000.89128.470.77927.510.73525.800.778
EDSR-PAMS327.250.78025.240.67325.380.64422.760.641
EDSR-DAQ331.660.88428.190.77127.280.72825.400.762
EDSR-DDTB331.520.88328.180.77127.300.72725.330.761
EDSR-ODM (Ours)331.850.88828.380.77627.430.73225.590.771
EDSR-PAMS229.510.83526.790.73426.450.69623.720.688
EDSR-DAQ†231.010.87127.890.76227.090.71924.880.740
EDSR-DDTB230.970.87627.870.76427.090.71924.820.742
EDSR-ODM (Ours)231.500.88228.140.77027.270.72625.170.755
", + "bbox": [ + 256, + 170, + 746, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "for each layer a learnable parameter [14]:", + "bbox": [ + 215, + 436, + 488, + 450 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nu _ {w} ^ {t + 1} = u _ {w} ^ {t} - \\beta^ {t} \\cdot \\nabla_ {u _ {w}} \\mathcal {L} _ {R} \\left(u _ {w} ^ {t}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 460, + 785, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\beta^t$ denotes the learning rate. This process determines the clipping parameter $u_w^t$ to quantize $\\mathbf{W}^t$ based on the weight of the previous iteration, $\\mathbf{W}^{t-1}$ . However, since the weight is also updated at iteration step $t$ ( $\\mathbf{W}^{t-1} \\to \\mathbf{W}^t$ ), a mismatch occurs between the current weight and the weight quantization range derived from the previous weight. To address this, we first obtain the quantization range by applying the global policy to the current weight, then adjust the range with a learnable parameter that accounts for layer-specific tendencies. Our clipping parameter is formulated as follows:", + "bbox": [ + 212, + 489, + 787, + 595 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nu _ {w} ^ {t + 1} = f \\left(\\boldsymbol {W} ^ {t + 1}\\right) \\cdot \\left(\\gamma_ {w} ^ {t} - \\beta^ {t} \\cdot \\nabla_ {\\gamma_ {w}} \\mathcal {L} _ {R} \\left(\\gamma_ {w} ^ {t}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 604, + 785, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\gamma_w$ is the learnable parameter for each layer representing the layer-wise adjustment. Each $\\gamma_w$ is initially set to 1. To prevent outliers from dominating the initial quantization range, we set the global policy $f(\\cdot)$ as the $j$ -th percentile function. Our clipping correction scheme introduces only one additional parameter per layer; thus, the overall computational overhead is minimal. For further details, please refer to Section 4.3.", + "bbox": [ + 212, + 633, + 787, + 712 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 732, + 356, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The efficacy and adaptability of the proposed quantization framework, ODM, are assessed through its application across several SR networks. The experimental settings are described (Section 4.1), and quantitative (Section 4.2), qualitative (Section 4.4), and complexity (Section 4.3) evaluations are conducted on various SR networks. Ablation studies are conducted to examine each component of the framework (Section 4.5).", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3313582ce5ea6b7f65e274bddcd4b112ab128e24d7f42db284e859a9de38f75c.jpg", + "table_caption": [ + "Table 3: Quantitative comparisons on RDN of scale 4" + ], + "table_footnote": [], + "table_body": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
RDN3232.240.89628.670.78427.630.73826.290.792
RDN-PAMS430.440.86227.540.75326.870.71024.520.726
RDN-DAQ431.910.88928.380.77527.380.73325.810.779
RDN-DDTB431.970.89128.490.78027.490.73525.900.783
RDN-ODM (Ours)432.060.89228.490.78027.520.73625.880.783
RDN-PAMS329.540.83826.820.73426.470.69623.830.692
RDN-DAQ331.570.88328.180.77127.270.72825.470.765
RDN-DDTB331.490.88328.170.77227.300.72825.350.764
RDN-ODM (Ours)331.790.88728.330.77627.420.73225.510.770
RDN-PAMS229.730.84326.960.73926.570.70023.870.696
RDN-DAQ230.710.86627.610.75526.930.71524.710.731
RDN-DDTB230.570.86727.560.75726.910.71424.500.728
RDN-ODM (Ours)231.370.88028.080.77027.240.72725.090.755
", + "bbox": [ + 259, + 171, + 741, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 412, + 419, + 428 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Models and Training. The proposed framework is applied directly to existing representative SR networks that produce satisfactory SR results, but involve heavy computations: EDSR (baseline) [37] and RDN [55]. Furthermore, we apply our method to the Transformer-based SR model, SwinIR-S [36]. Following prior works on SR quantization [17, 18, 33, 40, 51, 57], weights and activations of the high-level feature extraction module are quantized, which is the most computationally demanding. Training and validation are conducted using the DIV2K [1] dataset. ODM trains the network for 60K iterations with a batch size of 8. The weights are updated with an initial learning rate of $\\beta^0 = 10^{-4}$ . For cooperative update, we update clipping parameters with $10\\cdot \\beta^0$ initial learning rate. The learning rates are halved every 15K iteration. The hyperparameter for the percentile is set to $j = 99$ , and to balance the gradient of the loss terms, we set $\\lambda_{R} = 1$ and $\\lambda_{M} = 10^{-5}$ . Specially, we set $\\lambda_{M} = 10^{-6}$ for RDN whose overall mismatch is large. Ablation studies on the hyperparameters are in the supplementary material. All our experiments are implemented using PyTorch and run on an RTX 2080Ti GPU.", + "bbox": [ + 215, + 436, + 787, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation. We evaluate our framework on the standard benchmark (Set5 [4], Set14 [32], BSD100 [41], and Urban100 [21]) by measuring the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM [50]). To assess the computational complexity of our framework, we measure bitOPs and storage size. BitOPs refers to the number of operations weighted by the bit-widths of the two operands. Storage size is calculated as the number of stored parameters weighted by the precision of each parameter value.", + "bbox": [ + 215, + 671, + 795, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 784, + 398, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed scheme, we compare the results with existing SR quantization works using their official codes: PAMS [33], DAQ [18], and", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 218, + 116, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3a692a0cd6455135f82ba27e37f82d65a2c22d714dcbaf6d97f76202130d4b89.jpg", + "table_caption": [ + "Table 4: Quantitative comparisons on SwinIR of scale 4" + ], + "table_footnote": [], + "table_body": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-PAMS431.990.89028.500.77927.490.73525.860.780
SwinIR-DAQ431.820.88728.340.77527.370.73025.680.772
SwinIR-DDTB432.090.89128.550.78027.540.73526.010.783
SwinIR-ODM (Ours)432.170.89228.590.78127.560.73626.060.785
SwinIR-PAMS331.620.88428.230.77127.310.72825.380.762
SwinIR-DAQ331.500.88227.990.77027.120.72725.290.761
SwinIR-DDTB331.800.88728.340.77527.400.73125.630.771
SwinIR-ODM (Ours)331.940.88928.390.77727.450.73325.720.775
SwinIR-PAMS229.480.83426.790.73326.460.69823.720.688
SwinIR-DAQ229.100.82426.550.72526.300.69123.510.678
SwinIR-DDTB231.010.87327.800.76227.040.71924.790.739
SwinIR-ODM (Ours)231.440.88028.060.76927.230.72525.140.754
", + "bbox": [ + 254, + 171, + 746, + 383 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "DDTB [57]. For a fair comparison, we reproduce other methods using the same training iterations, 60K iterations. The supplementary materials provide additional experiments that further demonstrate the applicability of ODM, including results on 300K iterations, scale 2, and fully quantized settings.", + "bbox": [ + 217, + 434, + 785, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "EDSR. As shown in Table 2, ODM outperforms other methods in the 4, 3, and 2-bit settings, and notably, the improvement is significant for 2-bit, achieving a gain of more than $0.49\\mathrm{dB}$ for Set5. We notice that 4-bit EDSR-ODM achieves closer accuracy to 32-bit EDSR, with a marginal difference of $0.1\\mathrm{dB}$ for Set5. This indicates that ODM can effectively bridge the gap between the quantized network and the 32-bit network.", + "bbox": [ + 217, + 505, + 785, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "RDN. Similarly, Table 3 compares the results on RDN, whose computational complexity is more burdensome than EDSR. The results show that ODM consistently achieves superior performance on 4, 3, and 2-bit quantization. The gain over existing methods is especially large for the 2-bit setting, where it exceeds 0.66 dB for Set5.", + "bbox": [ + 217, + 590, + 785, + 652 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SwinIR. Furthermore, we evaluate our framework on the Transformer-based architecture, SwinIR. The linear and convolutional layers of SwinIR are quantized. According to Table 4, ODM is also proven effective in quantizing SwinIR across all bit settings, where the improvement is most notable in the 2-bit setting (0.43 dB).", + "bbox": [ + 217, + 662, + 785, + 723 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with QuantSR. We also compare our method with the concurrent work, QuantSR [44]. As the training code of QuantSR has not been released, we base our comparison on the reported performance. For a fair comparison with QuantSR's reported performance, we also train our model for 300K iterations on SRResNet [32] and SwinIR [36]. In Table 5, the results demonstrate that our method achieves better results than QuantSR; compared to QuantSR, ours shows a gain of 0.51 dB on 2-bit SRResNet and a gain of 0.14 dB on 2-bit SwinIR for Set5.", + "bbox": [ + 217, + 733, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c4e523073a7c9532acb32136b83aed3d88406966bafc75a98546239cc1e5eaa8.jpg", + "table_caption": [ + "Table 5: Quantitative comparisons with QuantSR on SRResNet and SwinIR of scale 4. For a fair comparison, our model (ODM*) is trained for 300K iterations following QuantSR." + ], + "table_footnote": [], + "table_body": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SRResNet3232.070.89328.500.78027.520.73525.860.779
SRResNet-QuantSR231.300.88228.080.76927.230.72525.130.754
SRResNet-ODM*(Ours)231.810.88828.320.77427.380.73025.540.767
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-QuantSR231.530.88528.160.77227.280.72725.260.761
SwinIR-ODM*(Ours)231.670.88528.230.77227.330.72825.360.762
", + "bbox": [ + 243, + 184, + 756, + 320 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d52149543286d58e35f3522e09e57386533aab09bf120df26eac15a8f9725d44.jpg", + "table_caption": [ + "Table 6: Computational complexity comparison with SR quantization methods" + ], + "table_footnote": [ + "(a) EDSR" + ], + "table_body": "
ModelBitStorage sizeBitOPsPSNRSSIM
EDSR321517.6K527.1T32.100.894
EDSR-PAMS2411.7K215.1T29.510.835
EDSR-DAQ2411.7K215.6T31.010.871
EDSR-DDTB2413.6K215.1T30.970.876
EDSR-ODM (Ours)2411.8K215.1T31.500.882
", + "bbox": [ + 225, + 361, + 506, + 434 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8c25a971270eeb087a805d8743f5b7c54f6963edcbe149932dcadd91753dd3d1.jpg", + "table_caption": [], + "table_footnote": [ + "(b) RDN" + ], + "table_body": "
ModelBitStorage sizeBitOPsPSNRSSIM
RDN3222271.1K6032.9T32.240.896
RDN-PAMS21715.9K236.6T29.730.843
RDN-DAQ21715.9K287.7T30.710.866
RDN-DDTB21761.6K236.6T30.570.867
RDN-ODM (Ours)21716.1K236.6T31.370.880
", + "bbox": [ + 516, + 361, + 797, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Complexity Analysis", + "text_level": 1, + "bbox": [ + 215, + 489, + 401, + 506 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Along with the accuracy of SR, we also evaluate the computational complexity of our framework in Table 6. We measure the storage size for the model weights and the bitOPs required for generating a $1920 \\times 1080$ image with $\\times 4$ SR network. Overall, our framework, ODM, achieves higher restoration accuracy with similar or fewer computational resources. Specifically, our weight-clipping correction involves an additional storage size of $0.06\\mathrm{K} / 0.15\\mathrm{K}$ for EDSR/RDN. Since the correction process can be predetermined before test time, no extra bitOPs are required. Compared to existing methods, our method achieves $\\times 31.7$ smaller storage size overhead than DDTB and $0.5\\mathrm{T}$ fewer bitOPs than DAQ on EDSR. For RDN, the gap is larger; our method's overhead is $\\times 304.7$ smaller in storage size than DDTB and $51.1\\mathrm{T}$ fewer in bitOPs than DAQ. Moreover, we note that DAQ adopts a channel-wise dynamic quantization function and DDTB an asymmetric weight quantization function, which are not favorably supported by hardware. Despite incurring a minor additional storage size of $0.06\\mathrm{K} / 0.15\\mathrm{K}$ over PAMS, the accuracy improvement over PAMS is significant $(+1.99\\mathrm{dB} / +1.64\\mathrm{dB})$ . The results prove that our method achieves significant accuracy gain with minimal or no computational overhead.", + "bbox": [ + 212, + 517, + 789, + 760 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 782, + 390, + 797 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figure 4 provides qualitative results and comparisons with the output images of quantized EDSR, RDN, and SwinIR. Our method, ODM, produces visually cleaner output", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/48f31ccff2149a662116a591268ea1d85671aa8092085a36e6c3288afb605d3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 143, + 331, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bd9fc4437f955de8ad27668a738c32de34abbc2716ae7ac98327e46e63b3e1ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 143, + 442, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6bd5970ab69c00ebe20b77452ec46237d89a94deeb9d0fec53fc720c48e58c1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 143, + 553, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e6ee8476d336b356eb837369dd161a8185082c8dd1eed02580f940ada0041802.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 143, + 665, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d3d8666199992305fd506fe9f9c341f5ed64987eafbac29a6e2056e9082d5c02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 776, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a50e368c941ce706fb8ce96038b518d158e5424ab6a3b527191de290e3da13a.jpg", + "image_caption": [ + "GT (img052)" + ], + "image_footnote": [], + "bbox": [ + 225, + 241, + 331, + 323 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/94ac52e687adaec83ce44802358a1af6b281543c2eb634705d46b9e899b6822d.jpg", + "image_caption": [ + "EDSR-PAMS" + ], + "image_footnote": [], + "bbox": [ + 338, + 241, + 442, + 321 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f5bb72cfabcf01540dafd10e4205225b8ccd6f774e898e3769955e95ea515ce5.jpg", + "image_caption": [ + "EDSR-DAQ" + ], + "image_footnote": [], + "bbox": [ + 449, + 241, + 553, + 321 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/87009efb5c6a912e285bad0cf882d8fc15bbf626bfb9c69c4e74699fd96d39f4.jpg", + "image_caption": [ + "EDSR-DDTB" + ], + "image_footnote": [], + "bbox": [ + 558, + 241, + 663, + 321 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/42374a7692ca11d70464905bf91d12d9cb58183504d900c0fed7489e1c5c9607.jpg", + "image_caption": [ + "EDSR-ODM (Ours)" + ], + "image_footnote": [], + "bbox": [ + 669, + 241, + 776, + 321 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8148fe4b70330995e0f2858bc25e553e1867d20f02ab11318201ba2ed4260740.jpg", + "image_caption": [ + "GT (img060)", + "GT (img024)", + "Figure 4: Qualitative results on Urban100 with EDSR, SwinIR, and RDN-based models" + ], + "image_footnote": [], + "bbox": [ + 225, + 340, + 331, + 421 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/957c19027a7689b2252fbfebd423a03b41af2f32a44ce9408fe80b793fa4e098.jpg", + "image_caption": [ + "SwinIR-PAMS", + "RDN-PAMS" + ], + "image_footnote": [], + "bbox": [ + 338, + 340, + 442, + 421 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/102846df4210d3a34ea014f5e39104b5e8821d2b5229dc78f9eb68be84bde719.jpg", + "image_caption": [ + "SwinIR-DAQ", + "RDN-DAQ" + ], + "image_footnote": [], + "bbox": [ + 449, + 340, + 553, + 421 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/31c11b324d8fd9c3c087792ad483b28c3f2a8928111d655ff31b99b40854ba3e.jpg", + "image_caption": [ + "SwinIR-DDTB", + "RDN-DDTB" + ], + "image_footnote": [], + "bbox": [ + 558, + 340, + 665, + 421 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a9377a151552375a06ebd3bc5b814aac447ba04db828c248af4eb8b8f3f0b28a.jpg", + "image_caption": [ + "SwinIR-ODM (Ours)", + "RDN-ODM (Ours)" + ], + "image_footnote": [], + "bbox": [ + 669, + 340, + 776, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "images compared to existing methods. In contrast, existing methods, especially PAMS, suffer from blurred lines or artifacts. These qualitative results stress the importance of alleviating the distribution mismatch problem in SR networks.", + "bbox": [ + 214, + 507, + 787, + 554 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 574, + 362, + 589 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Table 7, we verify the importance of each attribute of our framework: cooperative mismatch regularization and weight clipping correction. According to the results, each attribute individually improves baseline accuracy. Weight clipping correction improves the baseline by $+1.18$ dB for Set5, indicating that considering both the layer-wise trend and the current weight distribution is important for weight quantization. Also, while directly integrating mismatch regularization with the reconstruction loss (Model (d)) rather degrades performance by $-0.13$ dB, our cooperative scheme (Model (e)) improves the SR accuracy by $+0.51$ dB. This shows that reducing the mismatch in both activations and weights is important for accurately quantizing SR networks.", + "bbox": [ + 212, + 598, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Furthermore, we visualize the feature distributions in Figure 5 to validate the importance of our cooperative regularization term. After training only with the reconstruction loss, the outliers remain far from the quantization grid. When mismatch regularization loss is simply added to the original reconstruction loss, the activation distribution falls into a narrower range and resembles a multi-modal distribution near the quantization grids. Although this distribution is quantization-friendly, it significantly deviates from the original activation of the 32-bit network and removes originally dense values (near", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b69669853f569d6450d28da44ef5f78ddd9a49f301e4c968a0a66a4f6b63812c.jpg", + "table_caption": [ + "Table 7: Ablation study on each attribute of our framework. WCC refers to weight clipping correction and MR refers to mismatch regularization. Cooperative denotes whether the mismatch regularization is cooperatively used with the reconstruction loss. Non-cooperative MR denotes that the two losses are simply used together." + ], + "table_footnote": [], + "table_body": "
ModelWCCCooperativeMRSet5 (PSNR / SSIM)Urban100 (PSNR / SSIM)
(a)---29.94 / 0.84823.99 / 0.703
(b)-✓✓30.34 / 0.85924.27 / 0.715
(c)✓--31.12 / 0.87624.91 / 0.746
(d)✓-✓30.99 / 0.87124.79 / 0.735
(e)✓✓✓31.50 / 0.88225.17 / 0.755
", + "bbox": [ + 261, + 212, + 736, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cb1d982ff5ed9d4b32ea978e8f62584334404d66cc635ced9f85cddf219e85dd.jpg", + "image_caption": [ + "32-bit network" + ], + "image_footnote": [], + "bbox": [ + 243, + 316, + 370, + 387 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9421d72fd81ed07469ad22e1c1dd5c21da9645ca017142de6d76d72e53001947.jpg", + "image_caption": [ + "$\\mathcal{L}_R$", + "Figure 5: Distribution of activations before quantization. Using our cooperative mismatch regularization results in distributions more robust to low-bit quantization. 8th conv layer of EDSR-ODM (2-bit) on 'baby' (Set5) are visualized." + ], + "image_footnote": [], + "bbox": [ + 374, + 316, + 500, + 386 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3fb76831ab786402541c5738c45d160c5883dc0554c15cc9181a0eff6f1bedaa.jpg", + "image_caption": [ + "$\\mathcal{L}_R + \\mathcal{L}_M$" + ], + "image_footnote": [], + "bbox": [ + 504, + 316, + 627, + 386 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/57b31d47aa8a7e73a362d9f0e8c69193b3cda3e294d574c690803cd669c0c670.jpg", + "image_caption": [ + "Cooperative MR" + ], + "image_footnote": [], + "bbox": [ + 633, + 316, + 756, + 386 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "0). This can lead to an accuracy drop, as demonstrated in Table 7 (d). In contrast, our cooperative mismatch regularization results in quantization-friendly distribution while largely preserving the activations in the dense region of the original 32-bit network.", + "bbox": [ + 212, + 503, + 787, + 550 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 571, + 344, + 588 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SR networks suffer accuracy loss from quantization due to the inherent mismatch in feature distributions. Instead of employing resource-demanding dynamic modules to handle distinct distributions during test time, we introduce a new quantization-aware training technique that alleviates this mismatch through distribution optimization. We utilize cooperative mismatch regularization to update the SR network to be quantization-friendly and accurate. Additionally, to address the mismatch in layer-wise weights, we propose a weight-clipping correction strategy. These straightforward solutions effectively reduce the distribution mismatch with minimal computational overhead.", + "bbox": [ + 212, + 604, + 787, + 726 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 215, + 748, + 366, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A9E4080782] funded by the Korean government (MSIT).", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image super-resolution: Dataset and study. In: CVPR Workshops (2017)", + "2. Ayazoglu, M.: Extremely lightweight quantization robust real-time single-image super resolution for mobile devices. In: CVPR Workshops (2021)", + "3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)", + "4. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012)", + "5. Cai, Z., He, X., Sun, J., Vasconcelos, N.: Deep learning with low precision by half-wave gaussian quantization. In: CVPR (2017)", + "6. Cai, Z., Vasconcelos, N.: Rethinking differentiable search for mixed-precision neural networks. In: CVPR (2020)", + "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023)", + "8. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)", + "9. Chu, X., Zhang, B., Ma, H., Xu, R., Li, Q.: Fast, accurate and lightweight super-resolution with neural architecture search. In: ICPR (2021)", + "10. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014)", + "1. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE TPAMI 38(2), 295-307 (2015)", + "2. Dong, Z., Yao, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Hawq: Hessian aware quantization of neural networks with mixed-precision. In: ICCV (2019)", + "3. Du, Y., Czarnecki, W.M., Jayakumar, S.M., Farajtabar, M., Pascanu, R., Lakshminarayanan, B.: Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224 (2018)", + "4. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: ICLR (2020)", + "5. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022)", + "6. Habi, H.V., Jennings, R.H., Netzer, A.: Hmq: Hardware friendly mixed precision quantization block for cnns. In: ECCV (2020)", + "7. Hong, C., Baik, S., Kim, H., Nah, S., Lee, K.M.: Cadyq: Content-aware dynamic quantization for image super-resolution. In: ECCV (2022)", + "8. Hong, C., Kim, H., Baik, S., Oh, J., Lee, K.M.: Daq: Channel-wise distribution-aware quantization for deep image super-resolution networks. In: WACV (2022)", + "9. Hong, C., Lee, K.M.: Adabm: On-the-fly adaptive bit mapping for image super-resolution. In: CVPR (2024)", + "20. Hou, L., Kwok, J.T.: Loss-aware weight quantization of deep networks. In: ICLR (2018)", + "21. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015)", + "22. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACMMM (2019)", + "23. Hui, Z., Wang, X., Gao, X.: Fast and accurate single image super-resolution via information distillation network. In: CVPR (2018)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "24. Ignatov, A., Timofte, R., Denna, M., Younes, A.: Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In: CVPR Workshops (2021)", + "25. Jiang, X., Wang, N., Xin, J., Li, K., Yang, X., Gao, X.: Training binary neural network without batch normalization for image super-resolution. In: AAAI (2021)", + "26. Jiang, X., Wang, N., Xin, J., Xia, X., Yang, X., Gao, X.: Learning lightweight superresolution networks with weight pruning. Neural Networks 144, 21-32 (2021)", + "27. Jin, Q., Yang, L., Liao, Z.: Adabits: Neural network quantization with adaptive bit-widths. In: CVPR (2020)", + "28. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: CVPR (2021)", + "29. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: CVPR (2019)", + "30. Kim, H., Hong, S., Han, B., Myeong, H., Lee, K.M.: Fine-grained neural architecture search for image super-resolution. Journal of Visual Communication and Image Representation 89, 103654 (2022)", + "31. Kim, J., Lee, J., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016)", + "32. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super-resolution using a generative adversarial network. In: CVPR (2017)", + "33. Li, H., Yan, C., Lin, S., Zheng, X., Zhang, B., Yang, F., Ji, R.: Pams: Quantized superresolution via parameterized max scale. In: ECCV (2020)", + "34. Li, Y., Gu, S., Zhang, K., Gool, L.V., Timofte, R.: Dhp: Differentiable meta pruning via hypernetworks. In: ECCV (2020)", + "35. Li, Y., Li, W., Danelljan, M., Zhang, K., Gu, S., Van Gool, L., Timofte, R.: The heterogeneity hypothesis: Finding layer-wise differentiated network architectures. In: CVPR (2021)", + "36. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021)", + "37. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017)", + "38. Liu, J., Wang, Q., Zhang, D., Shen, L.: Super-resolution model quantized in multi-precision. Electronics 10(17), 2176 (2021)", + "39. Lou, Q., Guo, F., Liu, L., Kim, M., Jiang, L.: AutoQ: Automated kernel-wise neural network quantization. In: ICLR (2020)", + "40. Ma, Y., Xiong, H., Hu, Z., Ma, L.: Efficient super resolution using binarized neural network. In: CVPR Workshops (2019)", + "41. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001)", + "42. Oh, J., Kim, H., Nah, S., Hong, C., Choi, J., Lee, K.M.: Attentive fine-grained structured sparsity for image restoration. In: CVPR (2022)", + "43. Park, J., Son, S., Lee, K.M.: Content-aware local gan for photo-realistic super-resolution. In: ICCV (2023)", + "44. Qin, H., Zhang, Y., Ding, Y., Liu, X., Danelljan, M., Yu, F.: Quantsr: Accurate low-bit quantization for efficient image super-resolution. In: NeurIPS (2024)", + "45. Son, S., Kim, J., Lai, W.S., Yang, M.H., Lee, K.M.: Toward real-world super-resolution via adaptive downsampling models. IEEE TPAMI 44(11), 8657-8670 (2021)", + "46. Song, D., Xu, C., Jia, X., Chen, Y., Xu, C., Wang, Y.: Efficient residual dense block search for image super-resolution. In: AAAI (2020)", + "47. Tian, S., Lu, M., Liu, J., Guo, Y., Chen, Y., Zhang, S.: Cabm: Content-aware bit mapping for single image super-resolution network with large input. In: CVPR (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Hong and K.M. Lee", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wang, H., Chen, P., Zhuang, B., Shen, C.: Fully quantized image super-resolution networks. In: ACMMM (2021)", + "49. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: CVPR (2019)", + "50. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., et al.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)", + "51. Xin, J., Wang, N., Jiang, X., Li, J., Huang, H., Gao, X.: Binarized neural network for single image super resolution. In: ECCV (2020)", + "52. Yang, L., Jin, Q.: Fracbits: Mixed precision quantization via fractional bit-widths. In: AAAI (2021)", + "53. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: CVPR (2021)", + "54. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018)", + "55. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image superresolution. In: CVPR (2018)", + "56. Zhang, Y., Wang, H., Qin, C., Fu, Y.: Learning efficient image super-resolution networks via structure-regularized pruning. In: ICLR (2022)", + "57. Zhong, Y., Lin, M., Li, X., Li, K., Shen, Y., Chao, F., Wu, Y., Ji, R.: Dynamic dual trainable bounds for ultra-low precision super-resolution networks. In: ECCV (2022)", + "58. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: DoReFa-Net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016)", + "59. Zhuang, B., Shen, C., Tan, M., Liu, L., Reid, I.: Towards effective low-bitwidth convolutional neural networks. In: CVPR (2018)" + ], + "bbox": [ + 212, + 146, + 784, + 492 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ODM", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_model.json b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7dced836dfd19a080a1ec193166da9a02925a3c9 --- /dev/null +++ b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_model.json @@ -0,0 +1,2929 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.256, + 0.142, + 0.75, + 0.187 + ], + "angle": 0, + "content": "Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.213, + 0.641, + 0.228 + ], + "angle": 0, + "content": "Cheeun Hong1 and Kyoung Mu Lee1,2" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.24, + 0.717, + 0.254 + ], + "angle": 0, + "content": "Dept. of ECE & ASRI, {cheeun914, kyoungmu}@snu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.254, + 0.604, + 0.268 + ], + "angle": 0, + "content": "\\(^{2}\\) IPAI, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.31, + 0.744, + 0.601 + ], + "angle": 0, + "content": "Abstract. Although quantization has emerged as a promising approach to reducing computational complexity across various high-level vision tasks, it inevitably leads to accuracy loss in image super-resolution (SR) networks. This is due to the significantly divergent feature distributions across different channels and input images of the SR networks, which complicates the selection of a fixed quantization range. Existing works address this distribution mismatch problem by dynamically adapting quantization ranges to the varying distributions during test time. However, such a dynamic adaptation incurs additional computational costs during inference. In contrast, we propose a new quantization-aware training scheme that effectively Overcomes the Distribution Mismatch problem in SR networks without the need for dynamic adaptation. Intuitively, this mismatch can be mitigated by regularizing the distance between the feature and a fixed quantization range. However, we observe that such regularization can conflict with the reconstruction loss during training, negatively impacting SR accuracy. Therefore, we opt to regularize the mismatch only when the gradients of the regularization are aligned with those of the reconstruction loss. Additionally, we introduce a layer-wise weight clipping correction scheme to determine a more suitable quantization range for layer-wise weights. Experimental results demonstrate that our framework effectively reduces the distribution mismatch and achieves state-of-the-art performance with minimal computational overhead. Codes are available at https://github.com/Cheeun/ODM." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.615, + 0.744, + 0.645 + ], + "angle": 0, + "content": "Keywords: Image super-resolution \\(\\cdot\\) Network quantization \\(\\cdot\\) Quantization-aware training" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.671, + 0.359, + 0.687 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Image super-resolution (SR) is a core low-level vision task aimed at reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. Recent advances in deep learning [7, 11, 31, 36, 37, 43, 45, 54, 55] have led to astonishing achievements in producing high-fidelity images. However, the remarkable performance is based on heavy network architectures that incur significant computational costs, limiting practical viability, such as mobile deployment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.808, + 0.84 + ], + "angle": 0, + "content": "To mitigate the computational complexity of neural networks, quantization has emerged as a promising avenue. Network quantization has proven effective in reducing computational costs without significant accuracy loss, particularly in high-level vision tasks," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantization methods on SR. For performance, existing methods rely on channel-wise quantization or input-adaptive module, which incur computational overhead. Our method achieves high accuracy without utilizing channel-wise quantization or input-adaptive modules." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.199, + 0.684, + 0.276 + ], + "angle": 0, + "content": "
MethodChannel-wise QInput-adaptive ModulesPSNRSSIM
PAMS [33]XX29.510.835
DDTB [57]X✓30.970.876
DAQ [18]✓✓31.010.871
OursXX31.500.882
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.307, + 0.786, + 0.352 + ], + "angle": 0, + "content": "such as image classification [8, 20, 58]. However, when it comes to quantizing SR networks to lower bit-widths, substantial performance degradation [24] often occurs, posing a persistent and challenging problem to be addressed." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.353, + 0.787, + 0.459 + ], + "angle": 0, + "content": "This degradation can be attributed to the significant variance in the feature (activation) distributions of SR networks. The feature distribution of a layer exhibits substantial discrepancies across different channels and images, which makes it difficult to determine a single quantization range for a layer. Early approaches to SR quantization [33] adopt a training scheme to learn the quantization range of each layer. However, despite careful selection, the quantization ranges do not align with the varied values within the channel and image dimensions, which we refer to as distribution mismatch in features." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.567 + ], + "angle": 0, + "content": "Recent approaches aim to address this issue by incorporating dynamic adaptation modules that accommodate the varying distributions. For example, the quantization range is dynamically adjusted by directly leveraging the distribution mean and variance at test time [18] or by employing input-adaptive dynamic modules [57]. Although adapting the quantization function to each image during inference might handle variable distributions, the dynamic modules introduce considerable computational overhead, potentially compromising the computational benefits of quantization." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.568, + 0.787, + 0.748 + ], + "angle": 0, + "content": "In this study, we propose a novel quantization-aware training framework that addresses the distribution mismatch problem with a loss term that regulates the mismatch in distributions. Although directly minimizing the feature mismatch presents the potential for quantization-friendliness, whether it preserves reconstruction accuracy is questionable. We observe that concurrent optimization with the mismatch regularization and the original reconstruction loss can disrupt the image reconstruction process. Therefore, we introduce a cooperative mismatch regularization strategy, where the mismatch is regulated only when it collaborates harmoniously with the reconstruction loss. To determine the cooperative behavior, we assess the cosine similarity of the gradients from each loss, then we weigh the gradients of mismatch regularization based on this similarity. Consequently, we effectively update the SR network to hold both quantization-friendliness and reconstruction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Furthermore, we identify the distribution mismatch among the weights of different layers. We discover that employing a fixed policy to determine the layer-wise weight quantization range [17,18,33,47] can be suboptimal and that a further precise range can be obtained by considering both the current distribution of weights and the distinct tendencies of each layer. Therefore, we additionally incorporate layer-specific variations using a correction parameter for each layer. This strategy allows us to accurately find" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.785, + 0.126 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "the quantization range for weights while incurring only a minimal overhead (0.01% additional storage size and no additional bitOPs)—a significantly smaller impact compared to methods that use dynamic modules. Overall, the contributions of our work include the following:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.217, + 0.784, + 0.275 + ], + "angle": 0, + "content": "- We introduce the first quantization framework to address the distribution mismatch problem in SR networks without dynamic modules, as compared in Table 1. Our framework updates the SR network to be quantization-friendly and accurate simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.277, + 0.784, + 0.32 + ], + "angle": 0, + "content": "- Based on the observations of the distribution mismatch in SR networks, we effectively reduce the mismatch by introducing a cooperative mismatch regularization term and a weight clipping correction term." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.322, + 0.784, + 0.351 + ], + "angle": 0, + "content": "- Compared to existing approaches to SR quantization, ours achieves state-of-the-art performance with similar or fewer computations." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.217, + 0.784, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.375, + 0.375, + 0.39 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.406, + 0.426, + 0.421 + ], + "angle": 0, + "content": "2.1 Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.431, + 0.787, + 0.566 + ], + "angle": 0, + "content": "Convolutional Neural Network (CNN)-based approaches have demonstrated remarkable advancements in the image super-resolution (SR) task [32, 37], but at the cost of substantial computational resources. The intensive computations required by SR networks have spurred interest in developing lightweight SR architectures [10, 22, 23, 28, 54]. Furthermore, various strategies for lightweight SR networks have been explored, including neural architecture search [9, 30, 34, 35, 46], knowledge distillation [22, 23, 53], and pruning [26, 42, 56]. While these methods predominantly focus on reducing network depth or the number of channels, our work specifically aims to lower the precision of floating-point operations through network quantization." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.588, + 0.414, + 0.603 + ], + "angle": 0, + "content": "2.2 Network Quantization" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.613, + 0.787, + 0.703 + ], + "angle": 0, + "content": "By mapping 32-bit floating point values of features and weights in convolutional layers to lower-bit representations, network quantization provides a dramatic reduction in computational resources [5,8,14,29,58,59]. Recent works have successfully quantized various networks to low bit-widths with minimal compromise in accuracy [6,12,16, 27,39,49,52]. However, these efforts primarily target high-level vision tasks, whereas networks for low-level vision tasks remain vulnerable to low-bit quantization." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.726, + 0.525, + 0.74 + ], + "angle": 0, + "content": "2.3 Quantized Super-Resolution Networks" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In contrast to high-level vision tasks, SR poses different challenges due to its inherent high sensitivity to quantization [24, 40, 48, 51]. Some works have attempted to recover accuracy by modifying the network architecture [2, 25, 51] or by assigning different bits for each image [17, 19, 47] or network stage [38]. However, the primary challenge in quantizing SR networks lies in the vastly distinct feature distributions. To address this, Li et al. [33] adopted a learnable quantization range for different layers. Subsequently," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "recognizing that the distributions vary not only by layer, but also by channel and input, Hong et al. [18] introduced a channel-wise dynamic quantization function. Additionally, Zhong et al. [57] utilized an input-adaptive dynamic module to tailor quantization ranges for each specific input image. However, such dynamic adaptations of quantization functions during test-time incur non-negligible computational overheads. Instead of relying on test-time adaptive modules, our approach focuses on mitigating the feature mismatch before quantization. Our framework reduces the inherent distribution mismatch in SR networks with minimal overhead, accurately quantizing SR networks without dynamic modules. More recently, Qin et al. [44] introduced additional transformation functions in both the forward and backward processes. However, performance degradation is still evident in ultra-low bit (e.g., 2-bit) scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.335, + 0.4, + 0.353 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.457 + ], + "angle": 0, + "content": "In this section, after a brief introduction to network quantization (Section 3.1), we first analyze the mismatch in the features of SR networks (Section 3.2). Subsequently, we propose a solution to reduce this feature mismatch during training (Section 3.3). Additionally, we examine the mismatch in weights across SR networks (Section 3.4) and introduce a quantization range selection scheme that addresses the weight mismatch (Section 3.5). The overall training process is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.478, + 0.353, + 0.492 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.593 + ], + "angle": 0, + "content": "To reduce the heavy computations of convolutional and linear layers in neural networks, the input features (activations) and weights of each convolution/linear layer are quantized to low-bit values [5,8,15,29]. Given the input feature of the \\(i\\)-th convolution/linear layer \\(\\mathbf{X}_i \\in \\mathbb{R}^{B \\times C \\times H \\times W}\\), where \\(B, C, H\\), and \\(W\\) represent the dimensions of the input batch, channel, height, and width, respectively, a quantization operator \\(q(\\cdot)\\) quantizes the feature \\(\\mathbf{X}_i\\) with bit-width \\(b\\):" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.601, + 0.787, + 0.631 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {X} _ {i}; l, u\\right) = \\operatorname {I n t} \\left(\\frac {\\operatorname {c l i p} \\left(\\boldsymbol {X} _ {i} , l , u\\right) - l}{s}\\right) \\cdot s + l, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.787, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathrm{clip}(\\cdot ,l,u)\\) truncates the input into the quantization range of \\([l,u]\\) and \\(s = \\frac{u - l}{2^b - 1}\\). After truncation, the truncated feature is scaled to \\([0,2^b -1]\\), then rounded to integer values with \\(\\mathrm{Int}(\\cdot)\\), and rescaled to range \\([l,u]\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For activation quantization, the quantization range is defined by \\([l_a, u_a]\\). To obtain better quantization ranges in SR networks, the clipping parameters \\(l_a, u_a\\) for each layer are typically learned through quantization-aware training [33, 57]. Since the rounding function is not differentiable, a straight-through estimator (STE) [3] is used to train the clipping parameters in an end-to-end manner. Following [57], we initialize \\(l_a\\) and \\(u_a\\) as the \\((100-j)\\)-th and \\(j\\)-th percentile values of the feature, averaged among the training data, where \\(j\\) is set to 99 in our experiments to avoid outliers that corrupt the quantization range. For weights, as their distributions tend to be symmetric and the mean approximates zero, we utilize a symmetric quantization function. Thus, for the weight \\(W\\) of each convolution/linear layer, the quantization range is defined by \\([-u_w, u_w]\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.151, + 0.396, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.266, + 0.374, + 0.276 + ], + "angle": 0, + "content": "(a) Channel-wise features" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.151, + 0.574, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.266, + 0.542, + 0.276 + ], + "angle": 0, + "content": "(b) Image-wise features" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.157, + 0.775, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.297, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Figure 1: Distribution mismatch in SR networks. Compared to a classification network (e.g., ResNet-18), an SR network (e.g., EDSR) exhibits significant mismatches within the feature distributions across channel and image dimensions. The large distribution mismatch complicates the selection of an appropriate quantization range. Channels and images from the 2nd layer are randomly selected for visualization. Additional results are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.395, + 0.53, + 0.409 + ], + "angle": 0, + "content": "3.2 Distribution Mismatch in SR Networks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.789, + 0.678 + ], + "angle": 0, + "content": "The unfriendliness to quantization in SR networks arises from diverse feature (activation) distributions, as reported in previous studies [18, 33, 57], primarily due to the absence of batch normalization layers in SR networks. As illustrated in Figure 1, where there are notable discrepancies between the channel and image distributions, quantization grids are unnecessarily allocated to regions with minimal feature density. Early SR quantization methods tackled this issue by employing learnable quantization range parameters [33] for each feature. However, even though the quantization-aware training process strives to find the optimal range for each feature, it fails to account for the channel-wise and input-wise variance in distributions. This mismatch results in a large quantization error that can impair SR performance. To deal with these distribution mismatches, existing methods adopt different quantization ranges for each channel [18] or input image [18, 57]. Nevertheless, the test-time adaptation modules used to determine these ranges introduce unwanted computational overhead during inference. Thus, our straightforward solution is to pre-adjust the distributions to be quantization-friendly, thereby eliminating the need for additional adaptations at inference. The following sections will introduce a new quantization-aware training scheme designed to resolve the distribution mismatch problem." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.699, + 0.524, + 0.714 + ], + "angle": 0, + "content": "3.3 Cooperative Mismatch Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Instead of trying to identify a quantization range capable of accommodating diverse feature distributions, our approach aims to regularize the distribution mismatch beforehand. Obtaining an appropriate quantization range for a feature with high image- and channel-wise variance is difficult, as a certain number of channels or images will invariably be distant from the selected quantization range. In this work, we refer to the total distance of each feature from the selected quantization grid as the mismatch," + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.825, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\nM \\left(\\boldsymbol {X} _ {i}\\right) = \\left| \\left| \\boldsymbol {X} _ {i} - q \\left(\\boldsymbol {X} _ {i}; l _ {a}, u _ {a}\\right) \\right| \\right| _ {2}, \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.146, + 0.413, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.246, + 0.386, + 0.258 + ], + "angle": 0, + "content": "(a) Gradient conflict [13]" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.146, + 0.586, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.246, + 0.589, + 0.259 + ], + "angle": 0, + "content": "(b) Cosine similarity of two gradients" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.147, + 0.761, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.246, + 0.734, + 0.258 + ], + "angle": 0, + "content": "(c) Gradient conflict ratio" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.269, + 0.789, + 0.368 + ], + "angle": 0, + "content": "Figure 2: Conflict between mismatch regularization and reconstruction loss. Mismatch regularization updates a number of parameters in the contradictory direction to the reconstruction loss, which we refer to as gradient conflict. (b) When the two losses are jointly used, gradient conflict consistently occurs during training, outputting a negative cosine similarity value. (c) We plot the ratio of conflicted gradients during training. Nearly half of the parameters undergo gradient conflict, which indicates that merely combining mismatch regularization with the reconstruction loss can impair SR accuracy. Visualizations are done on EDSR." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.395, + 0.789, + 0.456 + ], + "angle": 0, + "content": "where \\( ||\\cdot||_2 \\) calculates the Frobenius norm. Further analyses of the definition of mismatch are provided in the supplementary material. We can reduce the overall mismatch by directly regularizing the mismatch of each feature to be quantized. The mismatch regularization loss is obtained by summing the mismatch over all quantized features:" + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.464, + 0.787, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M} = \\sum_ {i} ^ {\\# \\text {l a y e r s}} M \\left(\\boldsymbol {X} _ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.513, + 0.788, + 0.545 + ], + "angle": 0, + "content": "The mismatch regularization loss can be used in line with the original reconstruction loss typically used in the general quantization-aware training pipeline for SR networks:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.559, + 0.787, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\mathcal {L} _ {1} \\left(\\mathcal {Q} \\left(\\boldsymbol {I} _ {L R}\\right), \\boldsymbol {I} _ {H R}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.58, + 0.788, + 0.625 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_1\\) loss indicates the \\(l_{1}\\) distance between the reconstructed image using the quantized network \\(\\mathcal{Q}\\) and the ground-truth HR image \\(I_{HR}\\). Then, the optimization of the parameter \\(\\theta^t\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.633, + 0.787, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\theta^ {t + 1} = \\theta^ {t} - \\beta^ {t} \\cdot \\left(\\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right) + \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.804, + 0.719 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\theta}\\mathcal{L}_R(\\theta^t)\\) denotes the gradient from the original reconstruction loss and \\(\\nabla_{\\theta}\\mathcal{L}_M(\\theta^t)\\) is the gradient from mismatch regularization loss, and \\(\\beta^{t}\\) refers to the learning rate. Updating the network to minimize the mismatch regularization loss will reduce the overall error from feature quantization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "However, then a question arises: does reducing the quantization error of each feature lead to improved reconstruction accuracy? The answer is, according to our observation in Figure 2, not necessarily. During the training process, the mismatch regularization loss can collide with the original reconstruction loss. That is, for some parameters, the direction of the gradient from reconstruction loss and that of the mismatch regularization are opposing, referred to as gradient conflict [13]. As in Figure 2b, the cosine similarity of two gradients oscillates between positive and negative values during training, indicating that the directions of two gradients do not converge and the gradient" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.146, + 0.411, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.242, + 0.413, + 0.253 + ], + "angle": 0, + "content": "(a) Layer-wise QE using max policy" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.147, + 0.583, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.242, + 0.563, + 0.254 + ], + "angle": 0, + "content": "(b) 4-bit quantization grid" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.147, + 0.756, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.242, + 0.736, + 0.253 + ], + "angle": 0, + "content": "(c) 2-bit quantization grid" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.265, + 0.789, + 0.364 + ], + "angle": 0, + "content": "Figure 3: Layer-wise variation in error from weight quantization. (a) Quantization error (QE) varies across different layers when a fixed global policy (i.e., max) is used to determine the quantization range, particularly for low bits. For some layers, using max does not effectively serve as a proper policy for quantization range selection. (b) Outliers often dominate the quantization range, leading to quantization grids being wasted on low-density areas. (c) At low bits, quantization grids fail to cover high-density regions adequately. Therefore, the quantization range should be adjusted for certain layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.789, + 0.435 + ], + "angle": 0, + "content": "occasionally conflicts. Furthermore, as shown in Figure 2c, the proportion of parameters undergoing gradient conflict is not minor, implying that the regularization loss can severely hinder the reconstruction loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.789, + 0.542 + ], + "angle": 0, + "content": "We aim to avoid the conflict between these two losses, in other words, to minimize the mismatch as long as it does not hinder the reconstruction loss. Thus, we dismiss the mismatch regularization term when it is not cooperative with reconstruction loss and make more use of it when it is cooperative. Specifically, we determine whether the two losses are cooperative by examining the cosine similarity of the gradients of each loss and then simply weigh the gradient of mismatch regularization by the gradient similarity. Our cooperative mismatch regularization can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.551, + 0.788, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\theta^ {t + 1} = \\theta^ {t} - \\beta^ {t} \\cdot \\left(\\lambda_ {R} \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right) + \\lambda_ {M} \\cdot s i m \\left(\\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right), \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right) \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.735 + ], + "angle": 0, + "content": "where the underlined term, gradient similarity, is defined as \\( \\text{sim}(\\pmb{v}_a, \\pmb{v}_b) = \\frac{\\cos(\\pmb{v}_a, \\pmb{v}_b) + 1}{2} \\) and \\( \\cos(\\cdot, \\cdot) \\in [-1, 1] \\) calculates the cosine similarity between two vectors. \\( \\lambda_R, \\lambda_M \\) are hyper-parameters to balance the two gradients. If the directions of two gradients are similar (i.e., smaller than \\( 90^\\circ \\) and closer to \\( 0^\\circ \\)), the gradient similarity is a large value, then the parameter is updated in the direction where mismatch regularization is also substantially considered. On the contrary, if the two gradients point in the other direction (i.e., larger than \\( 90^\\circ \\) and closer to \\( 180^\\circ \\)), the two losses restrain each other, and the gradient similarity is close to 0. In this case, we follow the gradient of reconstruction loss. This allows the network to reduce the quantization error cooperatively with the reconstruction error. Details on gradient similarity are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.496, + 0.772 + ], + "angle": 0, + "content": "3.4 Distribution Mismatch in Weights" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "The weight distributions of SR networks have remained relatively unexplored in previous literature. This is because the weights in SR networks typically exhibit bell-shaped distributions, which are considered easier to quantize compared to the long-tailed, input-wise and channel-wise distinct activation distributions. Consequently, many" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.607, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Quantization-aware training process of ODM" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.164, + 0.787, + 0.387 + ], + "angle": 0, + "content": "Input: Pre-trained 32-bit network \\(\\mathcal{P}\\) \nOutput: Quantized network \\(\\mathcal{Q}\\) \nfor \\(t = 1,\\dots ,\\) # iters do for \\(i = 1,\\dots ,\\) # layers do if \\(t = 1\\) then Initialize activation quantization range \\([l_a,u_a]\\) Initialize weight quantization range \\([-u_w,u_w]\\) Given quantization range, obtain \\(q(\\pmb {X}_i;l_a,u_a)\\) using Eq. (1) Given \\(\\pmb {X}_i\\) and \\(q(\\pmb {X}_i;l_a,u_a)\\) , obtain mismatch using Eq. (2) Adjust weight quantization range \\([-u_w,u_w]\\) using Eq. (9) Given quantization range, obtain \\(q(\\pmb {W}_i; - u_w,u_w)\\) using Eq. (1) Replace \\(\\pmb {X}_i\\) \\(\\pmb {W}_i\\) in \\(\\mathcal{P}\\) with \\(q(\\pmb {X}_i),q(\\pmb {W}_i)\\) Calculate mismatch regularization loss (Eq. (3)) and reconstruction loss (Eq. (4)) Update parameters of \\(\\mathcal{Q}\\) with two losses cooperatively using Eq. (6)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.79, + 0.628 + ], + "angle": 0, + "content": "studies [17, 33, 47] simply adopt max quantization for weights, setting the quantization range with the maximum value of the current weight distribution. However, we notice that this is suboptimal and may contribute to low performance when SR networks are quantized to ultra-low bits (e.g., 2-bit). The issue arises because, in some layers, the outliers are far from the distribution mean, as visualized in Figures 3b and 3c. Thus, when the maximum value is used to determine the quantization range for such distributions, the quantization range is dominated by outliers, with quantization grids not allocated to high-density regions (e.g., near 0). This leads to substantial quantization errors that can accumulate and degrade restoration performance. In the case of 4-bit quantization (Figure 3b), although grids still cover high-density regions to an extent, a number of grids are wasted on low-density areas. The problem intensifies with low-bit (2-bit) quantization (Figure 3c), where quantization errors become significantly larger. This observation underscores the need for careful selection of the quantization range for layer-wise weights, particularly in low-bit quantization scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.455, + 0.663 + ], + "angle": 0, + "content": "3.5 Weight Clipping Correction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.789, + 0.777 + ], + "angle": 0, + "content": "Also, we notice that the error from weight quantization varies across different layers, as shown in Figure 3a. For instance, employing the maximum value as the quantization range can be an effective policy for certain layers, yet this policy proves inadequate for others (e.g., the layer shown in Figure 3c). Given the unique tendency of each layer, applying a uniform global policy for selecting the quantization range across all layers is suboptimal. Existing methods [17, 33, 47] utilize a fixed global policy throughout training to set the quantization range clipping parameter \\( u_{w} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.784, + 0.787, + 0.803 + ], + "angle": 0, + "content": "\\[\nu _ {w} ^ {t} = f \\left(\\boldsymbol {W} ^ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where the global policy \\( f(\\cdot) \\) is the same function for all layers (e.g., \\( \\max(\\cdot) \\)). A simple solution to accommodate layer-specific variations is to make the clipping parameter \\( u_w \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.694, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.331, + 0.145, + 0.674, + 0.16 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons on EDSR of scale 4" + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.171, + 0.747, + 0.384 + ], + "angle": 0, + "content": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
EDSR3232.100.89428.580.78127.560.73626.040.785
EDSR-PAMS431.590.88528.200.77327.320.72825.320.762
EDSR-DAQ431.850.88728.380.77627.420.73225.730.772
EDSR-DDTB431.850.88928.390.77727.440.73225.690.774
EDSR-ODM (Ours)432.000.89128.470.77927.510.73525.800.778
EDSR-PAMS327.250.78025.240.67325.380.64422.760.641
EDSR-DAQ331.660.88428.190.77127.280.72825.400.762
EDSR-DDTB331.520.88328.180.77127.300.72725.330.761
EDSR-ODM (Ours)331.850.88828.380.77627.430.73225.590.771
EDSR-PAMS229.510.83526.790.73426.450.69623.720.688
EDSR-DAQ†231.010.87127.890.76227.090.71924.880.740
EDSR-DDTB230.970.87627.870.76427.090.71924.820.742
EDSR-ODM (Ours)231.500.88228.140.77027.270.72625.170.755
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.385, + 0.786, + 0.409 + ], + "angle": 0, + "content": "† We note that the reported results of DAQ [18] are obtained using EDSR of 32 residual blocks. For a fair comparison, we reproduce EDSR-DAQ using EDSR of 16 residual blocks (i.e., EDSR-baseline)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.437, + 0.49, + 0.452 + ], + "angle": 0, + "content": "for each layer a learnable parameter [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.462, + 0.787, + 0.481 + ], + "angle": 0, + "content": "\\[\nu _ {w} ^ {t + 1} = u _ {w} ^ {t} - \\beta^ {t} \\cdot \\nabla_ {u _ {w}} \\mathcal {L} _ {R} \\left(u _ {w} ^ {t}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.597 + ], + "angle": 0, + "content": "where \\(\\beta^t\\) denotes the learning rate. This process determines the clipping parameter \\(u_w^t\\) to quantize \\(\\mathbf{W}^t\\) based on the weight of the previous iteration, \\(\\mathbf{W}^{t-1}\\). However, since the weight is also updated at iteration step \\(t\\) (\\(\\mathbf{W}^{t-1} \\to \\mathbf{W}^t\\)), a mismatch occurs between the current weight and the weight quantization range derived from the previous weight. To address this, we first obtain the quantization range by applying the global policy to the current weight, then adjust the range with a learnable parameter that accounts for layer-specific tendencies. Our clipping parameter is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.606, + 0.787, + 0.626 + ], + "angle": 0, + "content": "\\[\nu _ {w} ^ {t + 1} = f \\left(\\boldsymbol {W} ^ {t + 1}\\right) \\cdot \\left(\\gamma_ {w} ^ {t} - \\beta^ {t} \\cdot \\nabla_ {\\gamma_ {w}} \\mathcal {L} _ {R} \\left(\\gamma_ {w} ^ {t}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.713 + ], + "angle": 0, + "content": "where \\(\\gamma_w\\) is the learnable parameter for each layer representing the layer-wise adjustment. Each \\(\\gamma_w\\) is initially set to 1. To prevent outliers from dominating the initial quantization range, we set the global policy \\(f(\\cdot)\\) as the \\(j\\)-th percentile function. Our clipping correction scheme introduces only one additional parameter per layer; thus, the overall computational overhead is minimal. For further details, please refer to Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.733, + 0.357, + 0.751 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The efficacy and adaptability of the proposed quantization framework, ODM, are assessed through its application across several SR networks. The experimental settings are described (Section 4.1), and quantitative (Section 4.2), qualitative (Section 4.4), and complexity (Section 4.3) evaluations are conducted on various SR networks. Ablation studies are conducted to examine each component of the framework (Section 4.5)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.22, + 0.117, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.41, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.335, + 0.145, + 0.67, + 0.16 + ], + "angle": 0, + "content": "Table 3: Quantitative comparisons on RDN of scale 4" + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.172, + 0.742, + 0.384 + ], + "angle": 0, + "content": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
RDN3232.240.89628.670.78427.630.73826.290.792
RDN-PAMS430.440.86227.540.75326.870.71024.520.726
RDN-DAQ431.910.88928.380.77527.380.73325.810.779
RDN-DDTB431.970.89128.490.78027.490.73525.900.783
RDN-ODM (Ours)432.060.89228.490.78027.520.73625.880.783
RDN-PAMS329.540.83826.820.73426.470.69623.830.692
RDN-DAQ331.570.88328.180.77127.270.72825.470.765
RDN-DDTB331.490.88328.170.77227.300.72825.350.764
RDN-ODM (Ours)331.790.88728.330.77627.420.73225.510.770
RDN-PAMS229.730.84326.960.73926.570.70023.870.696
RDN-DAQ230.710.86627.610.75526.930.71524.710.731
RDN-DDTB230.570.86727.560.75726.910.71424.500.728
RDN-ODM (Ours)231.370.88028.080.77027.240.72725.090.755
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.413, + 0.421, + 0.429 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.438, + 0.788, + 0.65 + ], + "angle": 0, + "content": "Models and Training. The proposed framework is applied directly to existing representative SR networks that produce satisfactory SR results, but involve heavy computations: EDSR (baseline) [37] and RDN [55]. Furthermore, we apply our method to the Transformer-based SR model, SwinIR-S [36]. Following prior works on SR quantization [17, 18, 33, 40, 51, 57], weights and activations of the high-level feature extraction module are quantized, which is the most computationally demanding. Training and validation are conducted using the DIV2K [1] dataset. ODM trains the network for 60K iterations with a batch size of 8. The weights are updated with an initial learning rate of \\(\\beta^0 = 10^{-4}\\). For cooperative update, we update clipping parameters with \\(10\\cdot \\beta^0\\) initial learning rate. The learning rates are halved every 15K iteration. The hyperparameter for the percentile is set to \\(j = 99\\), and to balance the gradient of the loss terms, we set \\(\\lambda_{R} = 1\\) and \\(\\lambda_{M} = 10^{-5}\\). Specially, we set \\(\\lambda_{M} = 10^{-6}\\) for RDN whose overall mismatch is large. Ablation studies on the hyperparameters are in the supplementary material. All our experiments are implemented using PyTorch and run on an RTX 2080Ti GPU." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.672, + 0.797, + 0.763 + ], + "angle": 0, + "content": "Evaluation. We evaluate our framework on the standard benchmark (Set5 [4], Set14 [32], BSD100 [41], and Urban100 [21]) by measuring the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM [50]). To assess the computational complexity of our framework, we measure bitOPs and storage size. BitOPs refers to the number of operations weighted by the bit-widths of the two operands. Storage size is calculated as the number of stored parameters weighted by the precision of each parameter value." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.785, + 0.4, + 0.8 + ], + "angle": 0, + "content": "4.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed scheme, we compare the results with existing SR quantization works using their official codes: PAMS [33], DAQ [18], and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.328, + 0.145, + 0.677, + 0.159 + ], + "angle": 0, + "content": "Table 4: Quantitative comparisons on SwinIR of scale 4" + }, + { + "type": "table", + "bbox": [ + 0.255, + 0.172, + 0.747, + 0.384 + ], + "angle": 0, + "content": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-PAMS431.990.89028.500.77927.490.73525.860.780
SwinIR-DAQ431.820.88728.340.77527.370.73025.680.772
SwinIR-DDTB432.090.89128.550.78027.540.73526.010.783
SwinIR-ODM (Ours)432.170.89228.590.78127.560.73626.060.785
SwinIR-PAMS331.620.88428.230.77127.310.72825.380.762
SwinIR-DAQ331.500.88227.990.77027.120.72725.290.761
SwinIR-DDTB331.800.88728.340.77527.400.73125.630.771
SwinIR-ODM (Ours)331.940.88928.390.77727.450.73325.720.775
SwinIR-PAMS229.480.83426.790.73326.460.69823.720.688
SwinIR-DAQ229.100.82426.550.72526.300.69123.510.678
SwinIR-DDTB231.010.87327.800.76227.040.71924.790.739
SwinIR-ODM (Ours)231.440.88028.060.76927.230.72525.140.754
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.435, + 0.787, + 0.496 + ], + "angle": 0, + "content": "DDTB [57]. For a fair comparison, we reproduce other methods using the same training iterations, 60K iterations. The supplementary materials provide additional experiments that further demonstrate the applicability of ODM, including results on 300K iterations, scale 2, and fully quantized settings." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.506, + 0.787, + 0.581 + ], + "angle": 0, + "content": "EDSR. As shown in Table 2, ODM outperforms other methods in the 4, 3, and 2-bit settings, and notably, the improvement is significant for 2-bit, achieving a gain of more than \\(0.49\\mathrm{dB}\\) for Set5. We notice that 4-bit EDSR-ODM achieves closer accuracy to 32-bit EDSR, with a marginal difference of \\(0.1\\mathrm{dB}\\) for Set5. This indicates that ODM can effectively bridge the gap between the quantized network and the 32-bit network." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.592, + 0.787, + 0.653 + ], + "angle": 0, + "content": "RDN. Similarly, Table 3 compares the results on RDN, whose computational complexity is more burdensome than EDSR. The results show that ODM consistently achieves superior performance on 4, 3, and 2-bit quantization. The gain over existing methods is especially large for the 2-bit setting, where it exceeds 0.66 dB for Set5." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.663, + 0.787, + 0.724 + ], + "angle": 0, + "content": "SwinIR. Furthermore, we evaluate our framework on the Transformer-based architecture, SwinIR. The linear and convolutional layers of SwinIR are quantized. According to Table 4, ODM is also proven effective in quantizing SwinIR across all bit settings, where the improvement is most notable in the 2-bit setting (0.43 dB)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.734, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Comparison with QuantSR. We also compare our method with the concurrent work, QuantSR [44]. As the training code of QuantSR has not been released, we base our comparison on the reported performance. For a fair comparison with QuantSR's reported performance, we also train our model for 300K iterations on SRResNet [32] and SwinIR [36]. In Table 5, the results demonstrate that our method achieves better results than QuantSR; compared to QuantSR, ours shows a gain of 0.51 dB on 2-bit SRResNet and a gain of 0.14 dB on 2-bit SwinIR for Set5." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 5: Quantitative comparisons with QuantSR on SRResNet and SwinIR of scale 4. For a fair comparison, our model (ODM*) is trained for 300K iterations following QuantSR." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.185, + 0.758, + 0.321 + ], + "angle": 0, + "content": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SRResNet3232.070.89328.500.78027.520.73525.860.779
SRResNet-QuantSR231.300.88228.080.76927.230.72525.130.754
SRResNet-ODM*(Ours)231.810.88828.320.77427.380.73025.540.767
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-QuantSR231.530.88528.160.77227.280.72725.260.761
SwinIR-ODM*(Ours)231.670.88528.230.77227.330.72825.360.762
" + }, + { + "type": "table_caption", + "bbox": [ + 0.26, + 0.335, + 0.744, + 0.35 + ], + "angle": 0, + "content": "Table 6: Computational complexity comparison with SR quantization methods" + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.362, + 0.508, + 0.435 + ], + "angle": 0, + "content": "
ModelBitStorage sizeBitOPsPSNRSSIM
EDSR321517.6K527.1T32.100.894
EDSR-PAMS2411.7K215.1T29.510.835
EDSR-DAQ2411.7K215.6T31.010.871
EDSR-DDTB2413.6K215.1T30.970.876
EDSR-ODM (Ours)2411.8K215.1T31.500.882
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.341, + 0.438, + 0.388, + 0.447 + ], + "angle": 0, + "content": "(a) EDSR" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.362, + 0.798, + 0.435 + ], + "angle": 0, + "content": "
ModelBitStorage sizeBitOPsPSNRSSIM
RDN3222271.1K6032.9T32.240.896
RDN-PAMS21715.9K236.6T29.730.843
RDN-DAQ21715.9K287.7T30.710.866
RDN-DDTB21761.6K236.6T30.570.867
RDN-ODM (Ours)21716.1K236.6T31.370.880
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.634, + 0.438, + 0.677, + 0.447 + ], + "angle": 0, + "content": "(b) RDN" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.403, + 0.507 + ], + "angle": 0, + "content": "4.3 Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.79, + 0.761 + ], + "angle": 0, + "content": "Along with the accuracy of SR, we also evaluate the computational complexity of our framework in Table 6. We measure the storage size for the model weights and the bitOPs required for generating a \\(1920 \\times 1080\\) image with \\(\\times 4\\) SR network. Overall, our framework, ODM, achieves higher restoration accuracy with similar or fewer computational resources. Specifically, our weight-clipping correction involves an additional storage size of \\(0.06\\mathrm{K} / 0.15\\mathrm{K}\\) for EDSR/RDN. Since the correction process can be predetermined before test time, no extra bitOPs are required. Compared to existing methods, our method achieves \\(\\times 31.7\\) smaller storage size overhead than DDTB and \\(0.5\\mathrm{T}\\) fewer bitOPs than DAQ on EDSR. For RDN, the gap is larger; our method's overhead is \\(\\times 304.7\\) smaller in storage size than DDTB and \\(51.1\\mathrm{T}\\) fewer in bitOPs than DAQ. Moreover, we note that DAQ adopts a channel-wise dynamic quantization function and DDTB an asymmetric weight quantization function, which are not favorably supported by hardware. Despite incurring a minor additional storage size of \\(0.06\\mathrm{K} / 0.15\\mathrm{K}\\) over PAMS, the accuracy improvement over PAMS is significant \\((+1.99\\mathrm{dB} / +1.64\\mathrm{dB})\\). The results prove that our method achieves significant accuracy gain with minimal or no computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.391, + 0.799 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Figure 4 provides qualitative results and comparisons with the output images of quantized EDSR, RDN, and SwinIR. Our method, ODM, produces visually cleaner output" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.332, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.232, + 0.312, + 0.242 + ], + "angle": 0, + "content": "GT (img052)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.144, + 0.443, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.232, + 0.424, + 0.241 + ], + "angle": 0, + "content": "EDSR-PAMS" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.144, + 0.555, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.232, + 0.531, + 0.241 + ], + "angle": 0, + "content": "EDSR-DAQ" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.144, + 0.666, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.232, + 0.645, + 0.241 + ], + "angle": 0, + "content": "EDSR-DDTB" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.144, + 0.777, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.232, + 0.771, + 0.241 + ], + "angle": 0, + "content": "EDSR-ODM (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.242, + 0.332, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.33, + 0.311, + 0.34 + ], + "angle": 0, + "content": "GT (img060)" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.242, + 0.444, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.33, + 0.426, + 0.34 + ], + "angle": 0, + "content": "SwinIR-PAMS" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.242, + 0.555, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.33, + 0.534, + 0.34 + ], + "angle": 0, + "content": "SwinIR-DAQ" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.242, + 0.665, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.33, + 0.648, + 0.34 + ], + "angle": 0, + "content": "SwinIR-DDTB" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.242, + 0.777, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.33, + 0.774, + 0.34 + ], + "angle": 0, + "content": "SwinIR-ODM (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.341, + 0.332, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.428, + 0.311, + 0.439 + ], + "angle": 0, + "content": "GT (img024)" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.341, + 0.444, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.428, + 0.421, + 0.438 + ], + "angle": 0, + "content": "RDN-PAMS" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.341, + 0.555, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.428, + 0.529, + 0.438 + ], + "angle": 0, + "content": "RDN-DAQ" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.341, + 0.666, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.428, + 0.643, + 0.438 + ], + "angle": 0, + "content": "RDN-DDTB" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.341, + 0.777, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.428, + 0.768, + 0.438 + ], + "angle": 0, + "content": "RDN-ODM (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.452, + 0.766, + 0.466 + ], + "angle": 0, + "content": "Figure 4: Qualitative results on Urban100 with EDSR, SwinIR, and RDN-based models" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.508, + 0.788, + 0.555 + ], + "angle": 0, + "content": "images compared to existing methods. In contrast, existing methods, especially PAMS, suffer from blurred lines or artifacts. These qualitative results stress the importance of alleviating the distribution mismatch problem in SR networks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.364, + 0.59 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In Table 7, we verify the importance of each attribute of our framework: cooperative mismatch regularization and weight clipping correction. According to the results, each attribute individually improves baseline accuracy. Weight clipping correction improves the baseline by \\(+1.18\\) dB for Set5, indicating that considering both the layer-wise trend and the current weight distribution is important for weight quantization. Also, while directly integrating mismatch regularization with the reconstruction loss (Model (d)) rather degrades performance by \\(-0.13\\) dB, our cooperative scheme (Model (e)) improves the SR accuracy by \\(+0.51\\) dB. This shows that reducing the mismatch in both activations and weights is important for accurately quantizing SR networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Furthermore, we visualize the feature distributions in Figure 5 to validate the importance of our cooperative regularization term. After training only with the reconstruction loss, the outliers remain far from the quantization grid. When mismatch regularization loss is simply added to the original reconstruction loss, the activation distribution falls into a narrower range and resembles a multi-modal distribution near the quantization grids. Although this distribution is quantization-friendly, it significantly deviates from the original activation of the 32-bit network and removes originally dense values (near" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 7: Ablation study on each attribute of our framework. WCC refers to weight clipping correction and MR refers to mismatch regularization. Cooperative denotes whether the mismatch regularization is cooperatively used with the reconstruction loss. Non-cooperative MR denotes that the two losses are simply used together." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.213, + 0.737, + 0.304 + ], + "angle": 0, + "content": "
ModelWCCCooperativeMRSet5 (PSNR / SSIM)Urban100 (PSNR / SSIM)
(a)---29.94 / 0.84823.99 / 0.703
(b)-✓✓30.34 / 0.85924.27 / 0.715
(c)✓--31.12 / 0.87624.91 / 0.746
(d)✓-✓30.99 / 0.87124.79 / 0.735
(e)✓✓✓31.50 / 0.88225.17 / 0.755
" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.317, + 0.371, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.396, + 0.345, + 0.407 + ], + "angle": 0, + "content": "32-bit network" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.318, + 0.5, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.396, + 0.447, + 0.408 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_R\\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.318, + 0.628, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.396, + 0.595, + 0.408 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_R + \\mathcal{L}_M\\)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.318, + 0.757, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.396, + 0.735, + 0.408 + ], + "angle": 0, + "content": "Cooperative MR" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.421, + 0.789, + 0.464 + ], + "angle": 0, + "content": "Figure 5: Distribution of activations before quantization. Using our cooperative mismatch regularization results in distributions more robust to low-bit quantization. 8th conv layer of EDSR-ODM (2-bit) on 'baby' (Set5) are visualized." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.789, + 0.551 + ], + "angle": 0, + "content": "0). This can lead to an accuracy drop, as demonstrated in Table 7 (d). In contrast, our cooperative mismatch regularization results in quantization-friendly distribution while largely preserving the activations in the dense region of the original 32-bit network." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.573, + 0.345, + 0.589 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.789, + 0.727 + ], + "angle": 0, + "content": "SR networks suffer accuracy loss from quantization due to the inherent mismatch in feature distributions. Instead of employing resource-demanding dynamic modules to handle distinct distributions during test time, we introduce a new quantization-aware training technique that alleviates this mismatch through distribution optimization. We utilize cooperative mismatch regularization to update the SR network to be quantization-friendly and accurate. Additionally, to address the mismatch in layer-wise weights, we propose a weight-clipping correction strategy. These straightforward solutions effectively reduce the distribution mismatch with minimal computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.749, + 0.367, + 0.765 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A9E4080782] funded by the Korean government (MSIT)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.694, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image super-resolution: Dataset and study. In: CVPR Workshops (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Ayazoglu, M.: Extremely lightweight quantization robust real-time single-image super resolution for mobile devices. In: CVPR Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.26 + ], + "angle": 0, + "content": "3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.287 + ], + "angle": 0, + "content": "4. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.315 + ], + "angle": 0, + "content": "5. Cai, Z., He, X., Sun, J., Vasconcelos, N.: Deep learning with low precision by half-wave gaussian quantization. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.342 + ], + "angle": 0, + "content": "6. Cai, Z., Vasconcelos, N.: Rethinking differentiable search for mixed-precision neural networks. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "8. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "9. Chu, X., Zhang, B., Ma, H., Xu, R., Li, Q.: Fast, accurate and lightweight super-resolution with neural architecture search. In: ICPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.467 + ], + "angle": 0, + "content": "10. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "1. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE TPAMI 38(2), 295-307 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "2. Dong, Z., Yao, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Hawq: Hessian aware quantization of neural networks with mixed-precision. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "3. Du, Y., Czarnecki, W.M., Jayakumar, S.M., Farajtabar, M., Pascanu, R., Lakshminarayanan, B.: Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "4. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "5. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "6. Habi, H.V., Jennings, R.H., Netzer, A.: Hmq: Hardware friendly mixed precision quantization block for cnns. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "7. Hong, C., Baik, S., Kim, H., Nah, S., Lee, K.M.: Cadyq: Content-aware dynamic quantization for image super-resolution. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "8. Hong, C., Kim, H., Baik, S., Oh, J., Lee, K.M.: Daq: Channel-wise distribution-aware quantization for deep image super-resolution networks. In: WACV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "9. Hong, C., Lee, K.M.: Adabm: On-the-fly adaptive bit mapping for image super-resolution. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.765, + 0.757 + ], + "angle": 0, + "content": "20. Hou, L., Kwok, J.T.: Loss-aware weight quantization of deep networks. In: ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "21. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "22. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACMMM (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "23. Hui, Z., Wang, X., Gao, X.: Fast and accurate single image super-resolution via information distillation network. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "C. Hong and K.M. Lee" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "24. Ignatov, A., Timofte, R., Denna, M., Younes, A.: Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In: CVPR Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "25. Jiang, X., Wang, N., Xin, J., Li, K., Yang, X., Gao, X.: Training binary neural network without batch normalization for image super-resolution. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "26. Jiang, X., Wang, N., Xin, J., Xia, X., Yang, X., Gao, X.: Learning lightweight superresolution networks with weight pruning. Neural Networks 144, 21-32 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "27. Jin, Q., Yang, L., Liao, Z.: Adabits: Neural network quantization with adaptive bit-widths. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.283 + ], + "angle": 0, + "content": "28. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "29. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.351 + ], + "angle": 0, + "content": "30. Kim, H., Hong, S., Han, B., Myeong, H., Lee, K.M.: Fine-grained neural architecture search for image super-resolution. Journal of Visual Communication and Image Representation 89, 103654 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.378 + ], + "angle": 0, + "content": "31. Kim, J., Lee, J., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "32. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super-resolution using a generative adversarial network. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.446 + ], + "angle": 0, + "content": "33. Li, H., Yan, C., Lin, S., Zheng, X., Zhang, B., Yang, F., Ji, R.: Pams: Quantized superresolution via parameterized max scale. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "34. Li, Y., Gu, S., Zhang, K., Gool, L.V., Timofte, R.: Dhp: Differentiable meta pruning via hypernetworks. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.501 + ], + "angle": 0, + "content": "35. Li, Y., Li, W., Danelljan, M., Zhang, K., Gu, S., Van Gool, L., Timofte, R.: The heterogeneity hypothesis: Finding layer-wise differentiated network architectures. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.528 + ], + "angle": 0, + "content": "36. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.555 + ], + "angle": 0, + "content": "37. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "38. Liu, J., Wang, Q., Zhang, D., Shen, L.: Super-resolution model quantized in multi-precision. Electronics 10(17), 2176 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "39. Lou, Q., Guo, F., Liu, L., Kim, M., Jiang, L.: AutoQ: Automated kernel-wise neural network quantization. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.636 + ], + "angle": 0, + "content": "40. Ma, Y., Xiong, H., Hu, Z., Ma, L.: Efficient super resolution using binarized neural network. In: CVPR Workshops (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "41. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "42. Oh, J., Kim, H., Nah, S., Hong, C., Choi, J., Lee, K.M.: Attentive fine-grained structured sparsity for image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "43. Park, J., Son, S., Lee, K.M.: Content-aware local gan for photo-realistic super-resolution. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.758 + ], + "angle": 0, + "content": "44. Qin, H., Zhang, Y., Ding, Y., Liu, X., Danelljan, M., Yu, F.: Quantsr: Accurate low-bit quantization for efficient image super-resolution. In: NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "45. Son, S., Kim, J., Lai, W.S., Yang, M.H., Lee, K.M.: Toward real-world super-resolution via adaptive downsampling models. IEEE TPAMI 44(11), 8657-8670 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "46. Song, D., Xu, C., Jia, X., Chen, Y., Xu, C., Wang, Y.: Efficient residual dense block search for image super-resolution. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Tian, S., Lu, M., Liu, J., Guo, Y., Chen, Y., Zhang, S.: Cabm: Content-aware bit mapping for single image super-resolution network with large input. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "ODM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "48. Wang, H., Chen, P., Zhuang, B., Shen, C.: Fully quantized image super-resolution networks. In: ACMMM (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "49. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.203, + 0.785, + 0.231 + ], + "angle": 0, + "content": "50. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., et al.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.231, + 0.785, + 0.259 + ], + "angle": 0, + "content": "51. Xin, J., Wang, N., Jiang, X., Li, J., Huang, H., Gao, X.: Binarized neural network for single image super resolution. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.259, + 0.785, + 0.285 + ], + "angle": 0, + "content": "52. Yang, L., Jin, Q.: Fracbits: Mixed precision quantization via fractional bit-widths. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.286, + 0.785, + 0.314 + ], + "angle": 0, + "content": "53. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "54. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "55. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image superresolution. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "56. Zhang, Y., Wang, H., Qin, C., Fu, Y.: Learning efficient image super-resolution networks via structure-regularized pruning. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.397, + 0.785, + 0.424 + ], + "angle": 0, + "content": "57. Zhong, Y., Lin, M., Li, X., Li, K., Shen, Y., Chao, F., Wu, Y., Ji, R.: Dynamic dual trainable bounds for ultra-low precision super-resolution networks. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.424, + 0.785, + 0.466 + ], + "angle": 0, + "content": "58. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: DoReFa-Net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.466, + 0.785, + 0.493 + ], + "angle": 0, + "content": "59. Zhuang, B., Shen, C., Tan, M., Liu, L., Reid, I.: Towards effective low-bitwidth convolutional neural networks. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.493 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_origin.pdf b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8e499418c1575ddc7adb6f9147585ad9cd5c762 --- /dev/null +++ b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/a2bbb741-3d1f-428c-b0a6-24b0918a6f98_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0265502b39e54a0c3597a509b131ca4c34d34a5a84d02c3659d0919abe74a41 +size 6067851 diff --git a/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/full.md b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5499724f629353459d3a0a77a38ca286d42d2fbb --- /dev/null +++ b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/full.md @@ -0,0 +1,382 @@ +# Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks + +Cheeun Hong1 and Kyoung Mu Lee1,2 + +Dept. of ECE & ASRI, {cheeun914, kyoungmu}@snu.ac.kr + +$^{2}$ IPAI, Seoul National University + +Abstract. Although quantization has emerged as a promising approach to reducing computational complexity across various high-level vision tasks, it inevitably leads to accuracy loss in image super-resolution (SR) networks. This is due to the significantly divergent feature distributions across different channels and input images of the SR networks, which complicates the selection of a fixed quantization range. Existing works address this distribution mismatch problem by dynamically adapting quantization ranges to the varying distributions during test time. However, such a dynamic adaptation incurs additional computational costs during inference. In contrast, we propose a new quantization-aware training scheme that effectively Overcomes the Distribution Mismatch problem in SR networks without the need for dynamic adaptation. Intuitively, this mismatch can be mitigated by regularizing the distance between the feature and a fixed quantization range. However, we observe that such regularization can conflict with the reconstruction loss during training, negatively impacting SR accuracy. Therefore, we opt to regularize the mismatch only when the gradients of the regularization are aligned with those of the reconstruction loss. Additionally, we introduce a layer-wise weight clipping correction scheme to determine a more suitable quantization range for layer-wise weights. Experimental results demonstrate that our framework effectively reduces the distribution mismatch and achieves state-of-the-art performance with minimal computational overhead. Codes are available at https://github.com/Cheeun/ODM. + +Keywords: Image super-resolution $\cdot$ Network quantization $\cdot$ Quantization-aware training + +# 1 Introduction + +Image super-resolution (SR) is a core low-level vision task aimed at reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. Recent advances in deep learning [7, 11, 31, 36, 37, 43, 45, 54, 55] have led to astonishing achievements in producing high-fidelity images. However, the remarkable performance is based on heavy network architectures that incur significant computational costs, limiting practical viability, such as mobile deployment. + +To mitigate the computational complexity of neural networks, quantization has emerged as a promising avenue. Network quantization has proven effective in reducing computational costs without significant accuracy loss, particularly in high-level vision tasks, + +Table 1: Quantization methods on SR. For performance, existing methods rely on channel-wise quantization or input-adaptive module, which incur computational overhead. Our method achieves high accuracy without utilizing channel-wise quantization or input-adaptive modules. + +
MethodChannel-wise QInput-adaptive ModulesPSNRSSIM
PAMS [33]XX29.510.835
DDTB [57]X✓30.970.876
DAQ [18]✓✓31.010.871
OursXX31.500.882
+ +such as image classification [8, 20, 58]. However, when it comes to quantizing SR networks to lower bit-widths, substantial performance degradation [24] often occurs, posing a persistent and challenging problem to be addressed. + +This degradation can be attributed to the significant variance in the feature (activation) distributions of SR networks. The feature distribution of a layer exhibits substantial discrepancies across different channels and images, which makes it difficult to determine a single quantization range for a layer. Early approaches to SR quantization [33] adopt a training scheme to learn the quantization range of each layer. However, despite careful selection, the quantization ranges do not align with the varied values within the channel and image dimensions, which we refer to as distribution mismatch in features. + +Recent approaches aim to address this issue by incorporating dynamic adaptation modules that accommodate the varying distributions. For example, the quantization range is dynamically adjusted by directly leveraging the distribution mean and variance at test time [18] or by employing input-adaptive dynamic modules [57]. Although adapting the quantization function to each image during inference might handle variable distributions, the dynamic modules introduce considerable computational overhead, potentially compromising the computational benefits of quantization. + +In this study, we propose a novel quantization-aware training framework that addresses the distribution mismatch problem with a loss term that regulates the mismatch in distributions. Although directly minimizing the feature mismatch presents the potential for quantization-friendliness, whether it preserves reconstruction accuracy is questionable. We observe that concurrent optimization with the mismatch regularization and the original reconstruction loss can disrupt the image reconstruction process. Therefore, we introduce a cooperative mismatch regularization strategy, where the mismatch is regulated only when it collaborates harmoniously with the reconstruction loss. To determine the cooperative behavior, we assess the cosine similarity of the gradients from each loss, then we weigh the gradients of mismatch regularization based on this similarity. Consequently, we effectively update the SR network to hold both quantization-friendliness and reconstruction accuracy. + +Furthermore, we identify the distribution mismatch among the weights of different layers. We discover that employing a fixed policy to determine the layer-wise weight quantization range [17,18,33,47] can be suboptimal and that a further precise range can be obtained by considering both the current distribution of weights and the distinct tendencies of each layer. Therefore, we additionally incorporate layer-specific variations using a correction parameter for each layer. This strategy allows us to accurately find + +the quantization range for weights while incurring only a minimal overhead (0.01% additional storage size and no additional bitOPs)—a significantly smaller impact compared to methods that use dynamic modules. Overall, the contributions of our work include the following: + +- We introduce the first quantization framework to address the distribution mismatch problem in SR networks without dynamic modules, as compared in Table 1. Our framework updates the SR network to be quantization-friendly and accurate simultaneously. +- Based on the observations of the distribution mismatch in SR networks, we effectively reduce the mismatch by introducing a cooperative mismatch regularization term and a weight clipping correction term. +- Compared to existing approaches to SR quantization, ours achieves state-of-the-art performance with similar or fewer computations. + +# 2 Related Works + +# 2.1 Image Super-Resolution + +Convolutional Neural Network (CNN)-based approaches have demonstrated remarkable advancements in the image super-resolution (SR) task [32, 37], but at the cost of substantial computational resources. The intensive computations required by SR networks have spurred interest in developing lightweight SR architectures [10, 22, 23, 28, 54]. Furthermore, various strategies for lightweight SR networks have been explored, including neural architecture search [9, 30, 34, 35, 46], knowledge distillation [22, 23, 53], and pruning [26, 42, 56]. While these methods predominantly focus on reducing network depth or the number of channels, our work specifically aims to lower the precision of floating-point operations through network quantization. + +# 2.2 Network Quantization + +By mapping 32-bit floating point values of features and weights in convolutional layers to lower-bit representations, network quantization provides a dramatic reduction in computational resources [5,8,14,29,58,59]. Recent works have successfully quantized various networks to low bit-widths with minimal compromise in accuracy [6,12,16, 27,39,49,52]. However, these efforts primarily target high-level vision tasks, whereas networks for low-level vision tasks remain vulnerable to low-bit quantization. + +# 2.3 Quantized Super-Resolution Networks + +In contrast to high-level vision tasks, SR poses different challenges due to its inherent high sensitivity to quantization [24, 40, 48, 51]. Some works have attempted to recover accuracy by modifying the network architecture [2, 25, 51] or by assigning different bits for each image [17, 19, 47] or network stage [38]. However, the primary challenge in quantizing SR networks lies in the vastly distinct feature distributions. To address this, Li et al. [33] adopted a learnable quantization range for different layers. Subsequently, + +recognizing that the distributions vary not only by layer, but also by channel and input, Hong et al. [18] introduced a channel-wise dynamic quantization function. Additionally, Zhong et al. [57] utilized an input-adaptive dynamic module to tailor quantization ranges for each specific input image. However, such dynamic adaptations of quantization functions during test-time incur non-negligible computational overheads. Instead of relying on test-time adaptive modules, our approach focuses on mitigating the feature mismatch before quantization. Our framework reduces the inherent distribution mismatch in SR networks with minimal overhead, accurately quantizing SR networks without dynamic modules. More recently, Qin et al. [44] introduced additional transformation functions in both the forward and backward processes. However, performance degradation is still evident in ultra-low bit (e.g., 2-bit) scenarios. + +# 3 Proposed Method + +In this section, after a brief introduction to network quantization (Section 3.1), we first analyze the mismatch in the features of SR networks (Section 3.2). Subsequently, we propose a solution to reduce this feature mismatch during training (Section 3.3). Additionally, we examine the mismatch in weights across SR networks (Section 3.4) and introduce a quantization range selection scheme that addresses the weight mismatch (Section 3.5). The overall training process is summarized in Algorithm 1. + +# 3.1 Preliminaries + +To reduce the heavy computations of convolutional and linear layers in neural networks, the input features (activations) and weights of each convolution/linear layer are quantized to low-bit values [5,8,15,29]. Given the input feature of the $i$ -th convolution/linear layer $\mathbf{X}_i \in \mathbb{R}^{B \times C \times H \times W}$ , where $B, C, H$ , and $W$ represent the dimensions of the input batch, channel, height, and width, respectively, a quantization operator $q(\cdot)$ quantizes the feature $\mathbf{X}_i$ with bit-width $b$ : + +$$ +q \left(\boldsymbol {X} _ {i}; l, u\right) = \operatorname {I n t} \left(\frac {\operatorname {c l i p} \left(\boldsymbol {X} _ {i} , l , u\right) - l}{s}\right) \cdot s + l, \tag {1} +$$ + +where $\mathrm{clip}(\cdot ,l,u)$ truncates the input into the quantization range of $[l,u]$ and $s = \frac{u - l}{2^b - 1}$ . After truncation, the truncated feature is scaled to $[0,2^b -1]$ , then rounded to integer values with $\mathrm{Int}(\cdot)$ , and rescaled to range $[l,u]$ . + +For activation quantization, the quantization range is defined by $[l_a, u_a]$ . To obtain better quantization ranges in SR networks, the clipping parameters $l_a, u_a$ for each layer are typically learned through quantization-aware training [33, 57]. Since the rounding function is not differentiable, a straight-through estimator (STE) [3] is used to train the clipping parameters in an end-to-end manner. Following [57], we initialize $l_a$ and $u_a$ as the $(100-j)$ -th and $j$ -th percentile values of the feature, averaged among the training data, where $j$ is set to 99 in our experiments to avoid outliers that corrupt the quantization range. For weights, as their distributions tend to be symmetric and the mean approximates zero, we utilize a symmetric quantization function. Thus, for the weight $W$ of each convolution/linear layer, the quantization range is defined by $[-u_w, u_w]$ . + +![](images/d92441ffc2ec62642ec83b5e93c005e8a1a3b5ea3320f7b650a5d30bf62c2413.jpg) +(a) Channel-wise features + +![](images/4fb8a124108b508df44192265acd63ee451d89a09dca2232802e0d91f15135cd.jpg) +(b) Image-wise features + +![](images/0caab5668e1ca1407062e3e85501a6508542407dee4bbca90d549c4f25eea568.jpg) +Figure 1: Distribution mismatch in SR networks. Compared to a classification network (e.g., ResNet-18), an SR network (e.g., EDSR) exhibits significant mismatches within the feature distributions across channel and image dimensions. The large distribution mismatch complicates the selection of an appropriate quantization range. Channels and images from the 2nd layer are randomly selected for visualization. Additional results are available in the supplementary material. + +# 3.2 Distribution Mismatch in SR Networks + +The unfriendliness to quantization in SR networks arises from diverse feature (activation) distributions, as reported in previous studies [18, 33, 57], primarily due to the absence of batch normalization layers in SR networks. As illustrated in Figure 1, where there are notable discrepancies between the channel and image distributions, quantization grids are unnecessarily allocated to regions with minimal feature density. Early SR quantization methods tackled this issue by employing learnable quantization range parameters [33] for each feature. However, even though the quantization-aware training process strives to find the optimal range for each feature, it fails to account for the channel-wise and input-wise variance in distributions. This mismatch results in a large quantization error that can impair SR performance. To deal with these distribution mismatches, existing methods adopt different quantization ranges for each channel [18] or input image [18, 57]. Nevertheless, the test-time adaptation modules used to determine these ranges introduce unwanted computational overhead during inference. Thus, our straightforward solution is to pre-adjust the distributions to be quantization-friendly, thereby eliminating the need for additional adaptations at inference. The following sections will introduce a new quantization-aware training scheme designed to resolve the distribution mismatch problem. + +# 3.3 Cooperative Mismatch Regularization + +Instead of trying to identify a quantization range capable of accommodating diverse feature distributions, our approach aims to regularize the distribution mismatch beforehand. Obtaining an appropriate quantization range for a feature with high image- and channel-wise variance is difficult, as a certain number of channels or images will invariably be distant from the selected quantization range. In this work, we refer to the total distance of each feature from the selected quantization grid as the mismatch, + +$$ +M \left(\boldsymbol {X} _ {i}\right) = \left| \left| \boldsymbol {X} _ {i} - q \left(\boldsymbol {X} _ {i}; l _ {a}, u _ {a}\right) \right| \right| _ {2}, \tag {2} +$$ + +![](images/4202b0a395949f8651e631f3770f7db57f4318e0e989692d6c27c8c40890ff9e.jpg) +(a) Gradient conflict [13] + +![](images/0c48f56800665b41efc99d9f91fc9d9a3ea69907a9ecd99714b4a3c7ab8dcdef.jpg) +Figure 2: Conflict between mismatch regularization and reconstruction loss. Mismatch regularization updates a number of parameters in the contradictory direction to the reconstruction loss, which we refer to as gradient conflict. (b) When the two losses are jointly used, gradient conflict consistently occurs during training, outputting a negative cosine similarity value. (c) We plot the ratio of conflicted gradients during training. Nearly half of the parameters undergo gradient conflict, which indicates that merely combining mismatch regularization with the reconstruction loss can impair SR accuracy. Visualizations are done on EDSR. + +![](images/577d111d740facb7db319f01807c78e6686e30d2d7571d279f6ed47a1d1f8f5d.jpg) +(b) Cosine similarity of two gradients +(c) Gradient conflict ratio + +where $||\cdot||_2$ calculates the Frobenius norm. Further analyses of the definition of mismatch are provided in the supplementary material. We can reduce the overall mismatch by directly regularizing the mismatch of each feature to be quantized. The mismatch regularization loss is obtained by summing the mismatch over all quantized features: + +$$ +\mathcal {L} _ {M} = \sum_ {i} ^ {\# \text {l a y e r s}} M \left(\boldsymbol {X} _ {i}\right). \tag {3} +$$ + +The mismatch regularization loss can be used in line with the original reconstruction loss typically used in the general quantization-aware training pipeline for SR networks: + +$$ +\mathcal {L} _ {R} = \mathcal {L} _ {1} \left(\mathcal {Q} \left(\boldsymbol {I} _ {L R}\right), \boldsymbol {I} _ {H R}\right), \tag {4} +$$ + +where $\mathcal{L}_1$ loss indicates the $l_{1}$ distance between the reconstructed image using the quantized network $\mathcal{Q}$ and the ground-truth HR image $I_{HR}$ . Then, the optimization of the parameter $\theta^t$ is formulated as: + +$$ +\theta^ {t + 1} = \theta^ {t} - \beta^ {t} \cdot \left(\nabla_ {\theta} \mathcal {L} _ {R} \left(\theta^ {t}\right) + \nabla_ {\theta} \mathcal {L} _ {M} \left(\theta^ {t}\right)\right), \tag {5} +$$ + +where $\nabla_{\theta}\mathcal{L}_R(\theta^t)$ denotes the gradient from the original reconstruction loss and $\nabla_{\theta}\mathcal{L}_M(\theta^t)$ is the gradient from mismatch regularization loss, and $\beta^{t}$ refers to the learning rate. Updating the network to minimize the mismatch regularization loss will reduce the overall error from feature quantization. + +However, then a question arises: does reducing the quantization error of each feature lead to improved reconstruction accuracy? The answer is, according to our observation in Figure 2, not necessarily. During the training process, the mismatch regularization loss can collide with the original reconstruction loss. That is, for some parameters, the direction of the gradient from reconstruction loss and that of the mismatch regularization are opposing, referred to as gradient conflict [13]. As in Figure 2b, the cosine similarity of two gradients oscillates between positive and negative values during training, indicating that the directions of two gradients do not converge and the gradient + +![](images/64f1105e59ced2ec4a37d7b67bd46c76c604069032719521e323cb524ff0eb57.jpg) +(a) Layer-wise QE using max policy + +![](images/761c64155d31b35d0453355c3a91ea0cb50014020ce0fa63629d7ab7f6973068.jpg) +(b) 4-bit quantization grid +Figure 3: Layer-wise variation in error from weight quantization. (a) Quantization error (QE) varies across different layers when a fixed global policy (i.e., max) is used to determine the quantization range, particularly for low bits. For some layers, using max does not effectively serve as a proper policy for quantization range selection. (b) Outliers often dominate the quantization range, leading to quantization grids being wasted on low-density areas. (c) At low bits, quantization grids fail to cover high-density regions adequately. Therefore, the quantization range should be adjusted for certain layers. + +![](images/505d8e18b256eb5946f6611527585c4ba7a19fcdb11468ceab4808cef71b9080.jpg) +(c) 2-bit quantization grid + +occasionally conflicts. Furthermore, as shown in Figure 2c, the proportion of parameters undergoing gradient conflict is not minor, implying that the regularization loss can severely hinder the reconstruction loss. + +We aim to avoid the conflict between these two losses, in other words, to minimize the mismatch as long as it does not hinder the reconstruction loss. Thus, we dismiss the mismatch regularization term when it is not cooperative with reconstruction loss and make more use of it when it is cooperative. Specifically, we determine whether the two losses are cooperative by examining the cosine similarity of the gradients of each loss and then simply weigh the gradient of mismatch regularization by the gradient similarity. Our cooperative mismatch regularization can be formulated as follows: + +$$ +\theta^ {t + 1} = \theta^ {t} - \beta^ {t} \cdot \left(\lambda_ {R} \cdot \nabla_ {\theta} \mathcal {L} _ {R} \left(\theta^ {t}\right) + \lambda_ {M} \cdot s i m \left(\nabla_ {\theta} \mathcal {L} _ {R} \left(\theta^ {t}\right), \nabla_ {\theta} \mathcal {L} _ {M} \left(\theta^ {t}\right)\right) \cdot \nabla_ {\theta} \mathcal {L} _ {M} \left(\theta^ {t}\right)\right), \tag {6} +$$ + +where the underlined term, gradient similarity, is defined as $\text{sim}(\pmb{v}_a, \pmb{v}_b) = \frac{\cos(\pmb{v}_a, \pmb{v}_b) + 1}{2}$ and $\cos(\cdot, \cdot) \in [-1, 1]$ calculates the cosine similarity between two vectors. $\lambda_R, \lambda_M$ are hyper-parameters to balance the two gradients. If the directions of two gradients are similar (i.e., smaller than $90^\circ$ and closer to $0^\circ$ ), the gradient similarity is a large value, then the parameter is updated in the direction where mismatch regularization is also substantially considered. On the contrary, if the two gradients point in the other direction (i.e., larger than $90^\circ$ and closer to $180^\circ$ ), the two losses restrain each other, and the gradient similarity is close to 0. In this case, we follow the gradient of reconstruction loss. This allows the network to reduce the quantization error cooperatively with the reconstruction error. Details on gradient similarity are in the supplementary material. + +# 3.4 Distribution Mismatch in Weights + +The weight distributions of SR networks have remained relatively unexplored in previous literature. This is because the weights in SR networks typically exhibit bell-shaped distributions, which are considered easier to quantize compared to the long-tailed, input-wise and channel-wise distinct activation distributions. Consequently, many + +Algorithm 1 Quantization-aware training process of ODM +Input: Pre-trained 32-bit network $\mathcal{P}$ +Output: Quantized network $\mathcal{Q}$ +for $t = 1,\dots ,$ # iters do for $i = 1,\dots ,$ # layers do if $t = 1$ then Initialize activation quantization range $[l_a,u_a]$ Initialize weight quantization range $[-u_w,u_w]$ Given quantization range, obtain $q(\pmb {X}_i;l_a,u_a)$ using Eq. (1) Given $\pmb {X}_i$ and $q(\pmb {X}_i;l_a,u_a)$ , obtain mismatch using Eq. (2) Adjust weight quantization range $[-u_w,u_w]$ using Eq. (9) Given quantization range, obtain $q(\pmb {W}_i; - u_w,u_w)$ using Eq. (1) Replace $\pmb {X}_i$ $\pmb {W}_i$ in $\mathcal{P}$ with $q(\pmb {X}_i),q(\pmb {W}_i)$ Calculate mismatch regularization loss (Eq. (3)) and reconstruction loss (Eq. (4)) Update parameters of $\mathcal{Q}$ with two losses cooperatively using Eq. (6) + +studies [17, 33, 47] simply adopt max quantization for weights, setting the quantization range with the maximum value of the current weight distribution. However, we notice that this is suboptimal and may contribute to low performance when SR networks are quantized to ultra-low bits (e.g., 2-bit). The issue arises because, in some layers, the outliers are far from the distribution mean, as visualized in Figures 3b and 3c. Thus, when the maximum value is used to determine the quantization range for such distributions, the quantization range is dominated by outliers, with quantization grids not allocated to high-density regions (e.g., near 0). This leads to substantial quantization errors that can accumulate and degrade restoration performance. In the case of 4-bit quantization (Figure 3b), although grids still cover high-density regions to an extent, a number of grids are wasted on low-density areas. The problem intensifies with low-bit (2-bit) quantization (Figure 3c), where quantization errors become significantly larger. This observation underscores the need for careful selection of the quantization range for layer-wise weights, particularly in low-bit quantization scenarios. + +# 3.5 Weight Clipping Correction + +Also, we notice that the error from weight quantization varies across different layers, as shown in Figure 3a. For instance, employing the maximum value as the quantization range can be an effective policy for certain layers, yet this policy proves inadequate for others (e.g., the layer shown in Figure 3c). Given the unique tendency of each layer, applying a uniform global policy for selecting the quantization range across all layers is suboptimal. Existing methods [17, 33, 47] utilize a fixed global policy throughout training to set the quantization range clipping parameter $u_{w}$ as follows: + +$$ +u _ {w} ^ {t} = f \left(\boldsymbol {W} ^ {t}\right), \tag {7} +$$ + +where the global policy $f(\cdot)$ is the same function for all layers (e.g., $\max(\cdot)$ ). A simple solution to accommodate layer-specific variations is to make the clipping parameter $u_w$ + +Table 2: Quantitative comparisons on EDSR of scale 4 + +
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
EDSR3232.100.89428.580.78127.560.73626.040.785
EDSR-PAMS431.590.88528.200.77327.320.72825.320.762
EDSR-DAQ431.850.88728.380.77627.420.73225.730.772
EDSR-DDTB431.850.88928.390.77727.440.73225.690.774
EDSR-ODM (Ours)432.000.89128.470.77927.510.73525.800.778
EDSR-PAMS327.250.78025.240.67325.380.64422.760.641
EDSR-DAQ331.660.88428.190.77127.280.72825.400.762
EDSR-DDTB331.520.88328.180.77127.300.72725.330.761
EDSR-ODM (Ours)331.850.88828.380.77627.430.73225.590.771
EDSR-PAMS229.510.83526.790.73426.450.69623.720.688
EDSR-DAQ†231.010.87127.890.76227.090.71924.880.740
EDSR-DDTB230.970.87627.870.76427.090.71924.820.742
EDSR-ODM (Ours)231.500.88228.140.77027.270.72625.170.755
+ +† We note that the reported results of DAQ [18] are obtained using EDSR of 32 residual blocks. For a fair comparison, we reproduce EDSR-DAQ using EDSR of 16 residual blocks (i.e., EDSR-baseline). + +for each layer a learnable parameter [14]: + +$$ +u _ {w} ^ {t + 1} = u _ {w} ^ {t} - \beta^ {t} \cdot \nabla_ {u _ {w}} \mathcal {L} _ {R} \left(u _ {w} ^ {t}\right), \tag {8} +$$ + +where $\beta^t$ denotes the learning rate. This process determines the clipping parameter $u_w^t$ to quantize $\mathbf{W}^t$ based on the weight of the previous iteration, $\mathbf{W}^{t-1}$ . However, since the weight is also updated at iteration step $t$ ( $\mathbf{W}^{t-1} \to \mathbf{W}^t$ ), a mismatch occurs between the current weight and the weight quantization range derived from the previous weight. To address this, we first obtain the quantization range by applying the global policy to the current weight, then adjust the range with a learnable parameter that accounts for layer-specific tendencies. Our clipping parameter is formulated as follows: + +$$ +u _ {w} ^ {t + 1} = f \left(\boldsymbol {W} ^ {t + 1}\right) \cdot \left(\gamma_ {w} ^ {t} - \beta^ {t} \cdot \nabla_ {\gamma_ {w}} \mathcal {L} _ {R} \left(\gamma_ {w} ^ {t}\right)\right), \tag {9} +$$ + +where $\gamma_w$ is the learnable parameter for each layer representing the layer-wise adjustment. Each $\gamma_w$ is initially set to 1. To prevent outliers from dominating the initial quantization range, we set the global policy $f(\cdot)$ as the $j$ -th percentile function. Our clipping correction scheme introduces only one additional parameter per layer; thus, the overall computational overhead is minimal. For further details, please refer to Section 4.3. + +# 4 Experiments + +The efficacy and adaptability of the proposed quantization framework, ODM, are assessed through its application across several SR networks. The experimental settings are described (Section 4.1), and quantitative (Section 4.2), qualitative (Section 4.4), and complexity (Section 4.3) evaluations are conducted on various SR networks. Ablation studies are conducted to examine each component of the framework (Section 4.5). + +Table 3: Quantitative comparisons on RDN of scale 4 + +
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
RDN3232.240.89628.670.78427.630.73826.290.792
RDN-PAMS430.440.86227.540.75326.870.71024.520.726
RDN-DAQ431.910.88928.380.77527.380.73325.810.779
RDN-DDTB431.970.89128.490.78027.490.73525.900.783
RDN-ODM (Ours)432.060.89228.490.78027.520.73625.880.783
RDN-PAMS329.540.83826.820.73426.470.69623.830.692
RDN-DAQ331.570.88328.180.77127.270.72825.470.765
RDN-DDTB331.490.88328.170.77227.300.72825.350.764
RDN-ODM (Ours)331.790.88728.330.77627.420.73225.510.770
RDN-PAMS229.730.84326.960.73926.570.70023.870.696
RDN-DAQ230.710.86627.610.75526.930.71524.710.731
RDN-DDTB230.570.86727.560.75726.910.71424.500.728
RDN-ODM (Ours)231.370.88028.080.77027.240.72725.090.755
+ +# 4.1 Implementation Details + +Models and Training. The proposed framework is applied directly to existing representative SR networks that produce satisfactory SR results, but involve heavy computations: EDSR (baseline) [37] and RDN [55]. Furthermore, we apply our method to the Transformer-based SR model, SwinIR-S [36]. Following prior works on SR quantization [17, 18, 33, 40, 51, 57], weights and activations of the high-level feature extraction module are quantized, which is the most computationally demanding. Training and validation are conducted using the DIV2K [1] dataset. ODM trains the network for 60K iterations with a batch size of 8. The weights are updated with an initial learning rate of $\beta^0 = 10^{-4}$ . For cooperative update, we update clipping parameters with $10\cdot \beta^0$ initial learning rate. The learning rates are halved every 15K iteration. The hyperparameter for the percentile is set to $j = 99$ , and to balance the gradient of the loss terms, we set $\lambda_{R} = 1$ and $\lambda_{M} = 10^{-5}$ . Specially, we set $\lambda_{M} = 10^{-6}$ for RDN whose overall mismatch is large. Ablation studies on the hyperparameters are in the supplementary material. All our experiments are implemented using PyTorch and run on an RTX 2080Ti GPU. + +Evaluation. We evaluate our framework on the standard benchmark (Set5 [4], Set14 [32], BSD100 [41], and Urban100 [21]) by measuring the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM [50]). To assess the computational complexity of our framework, we measure bitOPs and storage size. BitOPs refers to the number of operations weighted by the bit-widths of the two operands. Storage size is calculated as the number of stored parameters weighted by the precision of each parameter value. + +# 4.2 Quantitative Results + +To evaluate the effectiveness of our proposed scheme, we compare the results with existing SR quantization works using their official codes: PAMS [33], DAQ [18], and + +Table 4: Quantitative comparisons on SwinIR of scale 4 + +
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-PAMS431.990.89028.500.77927.490.73525.860.780
SwinIR-DAQ431.820.88728.340.77527.370.73025.680.772
SwinIR-DDTB432.090.89128.550.78027.540.73526.010.783
SwinIR-ODM (Ours)432.170.89228.590.78127.560.73626.060.785
SwinIR-PAMS331.620.88428.230.77127.310.72825.380.762
SwinIR-DAQ331.500.88227.990.77027.120.72725.290.761
SwinIR-DDTB331.800.88728.340.77527.400.73125.630.771
SwinIR-ODM (Ours)331.940.88928.390.77727.450.73325.720.775
SwinIR-PAMS229.480.83426.790.73326.460.69823.720.688
SwinIR-DAQ229.100.82426.550.72526.300.69123.510.678
SwinIR-DDTB231.010.87327.800.76227.040.71924.790.739
SwinIR-ODM (Ours)231.440.88028.060.76927.230.72525.140.754
+ +DDTB [57]. For a fair comparison, we reproduce other methods using the same training iterations, 60K iterations. The supplementary materials provide additional experiments that further demonstrate the applicability of ODM, including results on 300K iterations, scale 2, and fully quantized settings. + +EDSR. As shown in Table 2, ODM outperforms other methods in the 4, 3, and 2-bit settings, and notably, the improvement is significant for 2-bit, achieving a gain of more than $0.49\mathrm{dB}$ for Set5. We notice that 4-bit EDSR-ODM achieves closer accuracy to 32-bit EDSR, with a marginal difference of $0.1\mathrm{dB}$ for Set5. This indicates that ODM can effectively bridge the gap between the quantized network and the 32-bit network. + +RDN. Similarly, Table 3 compares the results on RDN, whose computational complexity is more burdensome than EDSR. The results show that ODM consistently achieves superior performance on 4, 3, and 2-bit quantization. The gain over existing methods is especially large for the 2-bit setting, where it exceeds 0.66 dB for Set5. + +SwinIR. Furthermore, we evaluate our framework on the Transformer-based architecture, SwinIR. The linear and convolutional layers of SwinIR are quantized. According to Table 4, ODM is also proven effective in quantizing SwinIR across all bit settings, where the improvement is most notable in the 2-bit setting (0.43 dB). + +Comparison with QuantSR. We also compare our method with the concurrent work, QuantSR [44]. As the training code of QuantSR has not been released, we base our comparison on the reported performance. For a fair comparison with QuantSR's reported performance, we also train our model for 300K iterations on SRResNet [32] and SwinIR [36]. In Table 5, the results demonstrate that our method achieves better results than QuantSR; compared to QuantSR, ours shows a gain of 0.51 dB on 2-bit SRResNet and a gain of 0.14 dB on 2-bit SwinIR for Set5. + +Table 5: Quantitative comparisons with QuantSR on SRResNet and SwinIR of scale 4. For a fair comparison, our model (ODM*) is trained for 300K iterations following QuantSR. + +
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SRResNet3232.070.89328.500.78027.520.73525.860.779
SRResNet-QuantSR231.300.88228.080.76927.230.72525.130.754
SRResNet-ODM*(Ours)231.810.88828.320.77427.380.73025.540.767
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-QuantSR231.530.88528.160.77227.280.72725.260.761
SwinIR-ODM*(Ours)231.670.88528.230.77227.330.72825.360.762
+ +Table 6: Computational complexity comparison with SR quantization methods + +
ModelBitStorage sizeBitOPsPSNRSSIM
EDSR321517.6K527.1T32.100.894
EDSR-PAMS2411.7K215.1T29.510.835
EDSR-DAQ2411.7K215.6T31.010.871
EDSR-DDTB2413.6K215.1T30.970.876
EDSR-ODM (Ours)2411.8K215.1T31.500.882
+ +(a) EDSR + +
ModelBitStorage sizeBitOPsPSNRSSIM
RDN3222271.1K6032.9T32.240.896
RDN-PAMS21715.9K236.6T29.730.843
RDN-DAQ21715.9K287.7T30.710.866
RDN-DDTB21761.6K236.6T30.570.867
RDN-ODM (Ours)21716.1K236.6T31.370.880
+ +(b) RDN + +# 4.3 Complexity Analysis + +Along with the accuracy of SR, we also evaluate the computational complexity of our framework in Table 6. We measure the storage size for the model weights and the bitOPs required for generating a $1920 \times 1080$ image with $\times 4$ SR network. Overall, our framework, ODM, achieves higher restoration accuracy with similar or fewer computational resources. Specifically, our weight-clipping correction involves an additional storage size of $0.06\mathrm{K} / 0.15\mathrm{K}$ for EDSR/RDN. Since the correction process can be predetermined before test time, no extra bitOPs are required. Compared to existing methods, our method achieves $\times 31.7$ smaller storage size overhead than DDTB and $0.5\mathrm{T}$ fewer bitOPs than DAQ on EDSR. For RDN, the gap is larger; our method's overhead is $\times 304.7$ smaller in storage size than DDTB and $51.1\mathrm{T}$ fewer in bitOPs than DAQ. Moreover, we note that DAQ adopts a channel-wise dynamic quantization function and DDTB an asymmetric weight quantization function, which are not favorably supported by hardware. Despite incurring a minor additional storage size of $0.06\mathrm{K} / 0.15\mathrm{K}$ over PAMS, the accuracy improvement over PAMS is significant $(+1.99\mathrm{dB} / +1.64\mathrm{dB})$ . The results prove that our method achieves significant accuracy gain with minimal or no computational overhead. + +# 4.4 Qualitative Results + +Figure 4 provides qualitative results and comparisons with the output images of quantized EDSR, RDN, and SwinIR. Our method, ODM, produces visually cleaner output + +![](images/48f31ccff2149a662116a591268ea1d85671aa8092085a36e6c3288afb605d3d.jpg) + +![](images/bd9fc4437f955de8ad27668a738c32de34abbc2716ae7ac98327e46e63b3e1ed.jpg) + +![](images/6bd5970ab69c00ebe20b77452ec46237d89a94deeb9d0fec53fc720c48e58c1b.jpg) + +![](images/e6ee8476d336b356eb837369dd161a8185082c8dd1eed02580f940ada0041802.jpg) + +![](images/d3d8666199992305fd506fe9f9c341f5ed64987eafbac29a6e2056e9082d5c02.jpg) + +![](images/9a50e368c941ce706fb8ce96038b518d158e5424ab6a3b527191de290e3da13a.jpg) +GT (img052) + +![](images/94ac52e687adaec83ce44802358a1af6b281543c2eb634705d46b9e899b6822d.jpg) +EDSR-PAMS + +![](images/f5bb72cfabcf01540dafd10e4205225b8ccd6f774e898e3769955e95ea515ce5.jpg) +EDSR-DAQ + +![](images/87009efb5c6a912e285bad0cf882d8fc15bbf626bfb9c69c4e74699fd96d39f4.jpg) +EDSR-DDTB + +![](images/42374a7692ca11d70464905bf91d12d9cb58183504d900c0fed7489e1c5c9607.jpg) +EDSR-ODM (Ours) + +![](images/8148fe4b70330995e0f2858bc25e553e1867d20f02ab11318201ba2ed4260740.jpg) +GT (img060) +GT (img024) +Figure 4: Qualitative results on Urban100 with EDSR, SwinIR, and RDN-based models + +![](images/957c19027a7689b2252fbfebd423a03b41af2f32a44ce9408fe80b793fa4e098.jpg) +SwinIR-PAMS +RDN-PAMS + +![](images/102846df4210d3a34ea014f5e39104b5e8821d2b5229dc78f9eb68be84bde719.jpg) +SwinIR-DAQ +RDN-DAQ + +![](images/31c11b324d8fd9c3c087792ad483b28c3f2a8928111d655ff31b99b40854ba3e.jpg) +SwinIR-DDTB +RDN-DDTB + +![](images/a9377a151552375a06ebd3bc5b814aac447ba04db828c248af4eb8b8f3f0b28a.jpg) +SwinIR-ODM (Ours) +RDN-ODM (Ours) + +images compared to existing methods. In contrast, existing methods, especially PAMS, suffer from blurred lines or artifacts. These qualitative results stress the importance of alleviating the distribution mismatch problem in SR networks. + +# 4.5 Ablation Study + +In Table 7, we verify the importance of each attribute of our framework: cooperative mismatch regularization and weight clipping correction. According to the results, each attribute individually improves baseline accuracy. Weight clipping correction improves the baseline by $+1.18$ dB for Set5, indicating that considering both the layer-wise trend and the current weight distribution is important for weight quantization. Also, while directly integrating mismatch regularization with the reconstruction loss (Model (d)) rather degrades performance by $-0.13$ dB, our cooperative scheme (Model (e)) improves the SR accuracy by $+0.51$ dB. This shows that reducing the mismatch in both activations and weights is important for accurately quantizing SR networks. + +Furthermore, we visualize the feature distributions in Figure 5 to validate the importance of our cooperative regularization term. After training only with the reconstruction loss, the outliers remain far from the quantization grid. When mismatch regularization loss is simply added to the original reconstruction loss, the activation distribution falls into a narrower range and resembles a multi-modal distribution near the quantization grids. Although this distribution is quantization-friendly, it significantly deviates from the original activation of the 32-bit network and removes originally dense values (near + +Table 7: Ablation study on each attribute of our framework. WCC refers to weight clipping correction and MR refers to mismatch regularization. Cooperative denotes whether the mismatch regularization is cooperatively used with the reconstruction loss. Non-cooperative MR denotes that the two losses are simply used together. + +
ModelWCCCooperativeMRSet5 (PSNR / SSIM)Urban100 (PSNR / SSIM)
(a)---29.94 / 0.84823.99 / 0.703
(b)-✓✓30.34 / 0.85924.27 / 0.715
(c)✓--31.12 / 0.87624.91 / 0.746
(d)✓-✓30.99 / 0.87124.79 / 0.735
(e)✓✓✓31.50 / 0.88225.17 / 0.755
+ +![](images/cb1d982ff5ed9d4b32ea978e8f62584334404d66cc635ced9f85cddf219e85dd.jpg) +32-bit network + +![](images/9421d72fd81ed07469ad22e1c1dd5c21da9645ca017142de6d76d72e53001947.jpg) +$\mathcal{L}_R$ +Figure 5: Distribution of activations before quantization. Using our cooperative mismatch regularization results in distributions more robust to low-bit quantization. 8th conv layer of EDSR-ODM (2-bit) on 'baby' (Set5) are visualized. + +![](images/3fb76831ab786402541c5738c45d160c5883dc0554c15cc9181a0eff6f1bedaa.jpg) +$\mathcal{L}_R + \mathcal{L}_M$ + +![](images/57b31d47aa8a7e73a362d9f0e8c69193b3cda3e294d574c690803cd669c0c670.jpg) +Cooperative MR + +0). This can lead to an accuracy drop, as demonstrated in Table 7 (d). In contrast, our cooperative mismatch regularization results in quantization-friendly distribution while largely preserving the activations in the dense region of the original 32-bit network. + +# 5 Conclusion + +SR networks suffer accuracy loss from quantization due to the inherent mismatch in feature distributions. Instead of employing resource-demanding dynamic modules to handle distinct distributions during test time, we introduce a new quantization-aware training technique that alleviates this mismatch through distribution optimization. We utilize cooperative mismatch regularization to update the SR network to be quantization-friendly and accurate. Additionally, to address the mismatch in layer-wise weights, we propose a weight-clipping correction strategy. These straightforward solutions effectively reduce the distribution mismatch with minimal computational overhead. + +# Acknowledgment + +This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A9E4080782] funded by the Korean government (MSIT). + +# References + +1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image super-resolution: Dataset and study. In: CVPR Workshops (2017) +2. Ayazoglu, M.: Extremely lightweight quantization robust real-time single-image super resolution for mobile devices. In: CVPR Workshops (2021) +3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013) +4. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012) +5. Cai, Z., He, X., Sun, J., Vasconcelos, N.: Deep learning with low precision by half-wave gaussian quantization. In: CVPR (2017) +6. Cai, Z., Vasconcelos, N.: Rethinking differentiable search for mixed-precision neural networks. In: CVPR (2020) +7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023) +8. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018) +9. Chu, X., Zhang, B., Ma, H., Xu, R., Li, Q.: Fast, accurate and lightweight super-resolution with neural architecture search. In: ICPR (2021) +10. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014) +1. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE TPAMI 38(2), 295-307 (2015) +2. Dong, Z., Yao, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Hawq: Hessian aware quantization of neural networks with mixed-precision. In: ICCV (2019) +3. Du, Y., Czarnecki, W.M., Jayakumar, S.M., Farajtabar, M., Pascanu, R., Lakshminarayanan, B.: Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224 (2018) +4. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: ICLR (2020) +5. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022) +6. Habi, H.V., Jennings, R.H., Netzer, A.: Hmq: Hardware friendly mixed precision quantization block for cnns. In: ECCV (2020) +7. Hong, C., Baik, S., Kim, H., Nah, S., Lee, K.M.: Cadyq: Content-aware dynamic quantization for image super-resolution. In: ECCV (2022) +8. Hong, C., Kim, H., Baik, S., Oh, J., Lee, K.M.: Daq: Channel-wise distribution-aware quantization for deep image super-resolution networks. In: WACV (2022) +9. Hong, C., Lee, K.M.: Adabm: On-the-fly adaptive bit mapping for image super-resolution. In: CVPR (2024) +20. Hou, L., Kwok, J.T.: Loss-aware weight quantization of deep networks. In: ICLR (2018) +21. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015) +22. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACMMM (2019) +23. Hui, Z., Wang, X., Gao, X.: Fast and accurate single image super-resolution via information distillation network. In: CVPR (2018) + +24. Ignatov, A., Timofte, R., Denna, M., Younes, A.: Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In: CVPR Workshops (2021) +25. Jiang, X., Wang, N., Xin, J., Li, K., Yang, X., Gao, X.: Training binary neural network without batch normalization for image super-resolution. In: AAAI (2021) +26. Jiang, X., Wang, N., Xin, J., Xia, X., Yang, X., Gao, X.: Learning lightweight superresolution networks with weight pruning. Neural Networks 144, 21-32 (2021) +27. Jin, Q., Yang, L., Liao, Z.: Adabits: Neural network quantization with adaptive bit-widths. In: CVPR (2020) +28. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: CVPR (2021) +29. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: CVPR (2019) +30. Kim, H., Hong, S., Han, B., Myeong, H., Lee, K.M.: Fine-grained neural architecture search for image super-resolution. Journal of Visual Communication and Image Representation 89, 103654 (2022) +31. Kim, J., Lee, J., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016) +32. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super-resolution using a generative adversarial network. In: CVPR (2017) +33. Li, H., Yan, C., Lin, S., Zheng, X., Zhang, B., Yang, F., Ji, R.: Pams: Quantized superresolution via parameterized max scale. In: ECCV (2020) +34. Li, Y., Gu, S., Zhang, K., Gool, L.V., Timofte, R.: Dhp: Differentiable meta pruning via hypernetworks. In: ECCV (2020) +35. Li, Y., Li, W., Danelljan, M., Zhang, K., Gu, S., Van Gool, L., Timofte, R.: The heterogeneity hypothesis: Finding layer-wise differentiated network architectures. In: CVPR (2021) +36. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021) +37. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017) +38. Liu, J., Wang, Q., Zhang, D., Shen, L.: Super-resolution model quantized in multi-precision. Electronics 10(17), 2176 (2021) +39. Lou, Q., Guo, F., Liu, L., Kim, M., Jiang, L.: AutoQ: Automated kernel-wise neural network quantization. In: ICLR (2020) +40. Ma, Y., Xiong, H., Hu, Z., Ma, L.: Efficient super resolution using binarized neural network. In: CVPR Workshops (2019) +41. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) +42. Oh, J., Kim, H., Nah, S., Hong, C., Choi, J., Lee, K.M.: Attentive fine-grained structured sparsity for image restoration. In: CVPR (2022) +43. Park, J., Son, S., Lee, K.M.: Content-aware local gan for photo-realistic super-resolution. In: ICCV (2023) +44. Qin, H., Zhang, Y., Ding, Y., Liu, X., Danelljan, M., Yu, F.: Quantsr: Accurate low-bit quantization for efficient image super-resolution. In: NeurIPS (2024) +45. Son, S., Kim, J., Lai, W.S., Yang, M.H., Lee, K.M.: Toward real-world super-resolution via adaptive downsampling models. IEEE TPAMI 44(11), 8657-8670 (2021) +46. Song, D., Xu, C., Jia, X., Chen, Y., Xu, C., Wang, Y.: Efficient residual dense block search for image super-resolution. In: AAAI (2020) +47. Tian, S., Lu, M., Liu, J., Guo, Y., Chen, Y., Zhang, S.: Cabm: Content-aware bit mapping for single image super-resolution network with large input. In: CVPR (2023) + +48. Wang, H., Chen, P., Zhuang, B., Shen, C.: Fully quantized image super-resolution networks. In: ACMMM (2021) +49. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: CVPR (2019) +50. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., et al.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) +51. Xin, J., Wang, N., Jiang, X., Li, J., Huang, H., Gao, X.: Binarized neural network for single image super resolution. In: ECCV (2020) +52. Yang, L., Jin, Q.: Fracbits: Mixed precision quantization via fractional bit-widths. In: AAAI (2021) +53. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: CVPR (2021) +54. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018) +55. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image superresolution. In: CVPR (2018) +56. Zhang, Y., Wang, H., Qin, C., Fu, Y.: Learning efficient image super-resolution networks via structure-regularized pruning. In: ICLR (2022) +57. Zhong, Y., Lin, M., Li, X., Li, K., Shen, Y., Chao, F., Wu, Y., Ji, R.: Dynamic dual trainable bounds for ultra-low precision super-resolution networks. In: ECCV (2022) +58. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: DoReFa-Net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016) +59. Zhuang, B., Shen, C., Tan, M., Liu, L., Reid, I.: Towards effective low-bitwidth convolutional neural networks. In: CVPR (2018) \ No newline at end of file diff --git a/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/images.zip b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..83f82d0a59f345df4b958718beb698e12bf71902 --- /dev/null +++ b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a95a0f189e541ba001ca0792115f8c08abc0cc51a0904d49634920709fabee77 +size 652624 diff --git a/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/layout.json b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fc9e2bc6aef37d54a300f31d977186d7efdd28f7 --- /dev/null +++ b/2024/Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks/layout.json @@ -0,0 +1,11067 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 112, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 112, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 112, + 459, + 148 + ], + "type": "text", + "content": "Overcoming Distribution Mismatch in Quantizing Image Super-Resolution Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 220, + 168, + 392, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 168, + 392, + 180 + ], + "spans": [ + { + "bbox": [ + 220, + 168, + 392, + 180 + ], + "type": "text", + "content": "Cheeun Hong1 and Kyoung Mu Lee1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 190, + 438, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 190, + 438, + 201 + ], + "spans": [ + { + "bbox": [ + 175, + 190, + 438, + 201 + ], + "type": "text", + "content": "Dept. of ECE & ASRI, {cheeun914, kyoungmu}@snu.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 201, + 369, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 201, + 369, + 212 + ], + "spans": [ + { + "bbox": [ + 244, + 201, + 369, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 244, + 201, + 369, + 212 + ], + "type": "text", + "content": " IPAI, Seoul National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 245, + 455, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 245, + 455, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 245, + 455, + 475 + ], + "type": "text", + "content": "Abstract. Although quantization has emerged as a promising approach to reducing computational complexity across various high-level vision tasks, it inevitably leads to accuracy loss in image super-resolution (SR) networks. This is due to the significantly divergent feature distributions across different channels and input images of the SR networks, which complicates the selection of a fixed quantization range. Existing works address this distribution mismatch problem by dynamically adapting quantization ranges to the varying distributions during test time. However, such a dynamic adaptation incurs additional computational costs during inference. In contrast, we propose a new quantization-aware training scheme that effectively Overcomes the Distribution Mismatch problem in SR networks without the need for dynamic adaptation. Intuitively, this mismatch can be mitigated by regularizing the distance between the feature and a fixed quantization range. However, we observe that such regularization can conflict with the reconstruction loss during training, negatively impacting SR accuracy. Therefore, we opt to regularize the mismatch only when the gradients of the regularization are aligned with those of the reconstruction loss. Additionally, we introduce a layer-wise weight clipping correction scheme to determine a more suitable quantization range for layer-wise weights. Experimental results demonstrate that our framework effectively reduces the distribution mismatch and achieves state-of-the-art performance with minimal computational overhead. Codes are available at https://github.com/Cheeun/ODM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 487, + 455, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 487, + 455, + 510 + ], + "spans": [ + { + "bbox": [ + 160, + 487, + 455, + 510 + ], + "type": "text", + "content": "Keywords: Image super-resolution " + }, + { + "bbox": [ + 160, + 487, + 455, + 510 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 487, + 455, + 510 + ], + "type": "text", + "content": " Network quantization " + }, + { + "bbox": [ + 160, + 487, + 455, + 510 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 487, + 455, + 510 + ], + "type": "text", + "content": " Quantization-aware training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 531, + 219, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 219, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 219, + 544 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "Image super-resolution (SR) is a core low-level vision task aimed at reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. Recent advances in deep learning [7, 11, 31, 36, 37, 43, 45, 54, 55] have led to astonishing achievements in producing high-fidelity images. However, the remarkable performance is based on heavy network architectures that incur significant computational costs, limiting practical viability, such as mobile deployment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 494, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 494, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 494, + 665 + ], + "type": "text", + "content": "To mitigate the computational complexity of neural networks, quantization has emerged as a promising avenue. Network quantization has proven effective in reducing computational costs without significant accuracy loss, particularly in high-level vision tasks," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 194, + 157, + 418, + 218 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantization methods on SR. For performance, existing methods rely on channel-wise quantization or input-adaptive module, which incur computational overhead. Our method achieves high accuracy without utilizing channel-wise quantization or input-adaptive modules." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 157, + 418, + 218 + ], + "lines": [ + { + "bbox": [ + 194, + 157, + 418, + 218 + ], + "spans": [ + { + "bbox": [ + 194, + 157, + 418, + 218 + ], + "type": "table", + "html": "
MethodChannel-wise QInput-adaptive ModulesPSNRSSIM
PAMS [33]XX29.510.835
DDTB [57]X✓30.970.876
DAQ [18]✓✓31.010.871
OursXX31.500.882
", + "image_path": "4f6bfbf9bc8581da0579778a9a898eb0a74f74974949eb8bbe82a50b5b528b16.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 481, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 481, + 278 + ], + "type": "text", + "content": "such as image classification [8, 20, 58]. However, when it comes to quantizing SR networks to lower bit-widths, substantial performance degradation [24] often occurs, posing a persistent and challenging problem to be addressed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 279, + 481, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 481, + 363 + ], + "type": "text", + "content": "This degradation can be attributed to the significant variance in the feature (activation) distributions of SR networks. The feature distribution of a layer exhibits substantial discrepancies across different channels and images, which makes it difficult to determine a single quantization range for a layer. Early approaches to SR quantization [33] adopt a training scheme to learn the quantization range of each layer. However, despite careful selection, the quantization ranges do not align with the varied values within the channel and image dimensions, which we refer to as distribution mismatch in features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 364, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 449 + ], + "type": "text", + "content": "Recent approaches aim to address this issue by incorporating dynamic adaptation modules that accommodate the varying distributions. For example, the quantization range is dynamically adjusted by directly leveraging the distribution mean and variance at test time [18] or by employing input-adaptive dynamic modules [57]. Although adapting the quantization function to each image during inference might handle variable distributions, the dynamic modules introduce considerable computational overhead, potentially compromising the computational benefits of quantization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 449, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 481, + 592 + ], + "type": "text", + "content": "In this study, we propose a novel quantization-aware training framework that addresses the distribution mismatch problem with a loss term that regulates the mismatch in distributions. Although directly minimizing the feature mismatch presents the potential for quantization-friendliness, whether it preserves reconstruction accuracy is questionable. We observe that concurrent optimization with the mismatch regularization and the original reconstruction loss can disrupt the image reconstruction process. Therefore, we introduce a cooperative mismatch regularization strategy, where the mismatch is regulated only when it collaborates harmoniously with the reconstruction loss. To determine the cooperative behavior, we assess the cosine similarity of the gradients from each loss, then we weigh the gradients of mismatch regularization based on this similarity. Consequently, we effectively update the SR network to hold both quantization-friendliness and reconstruction accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": "Furthermore, we identify the distribution mismatch among the weights of different layers. We discover that employing a fixed policy to determine the layer-wise weight quantization range [17,18,33,47] can be suboptimal and that a further precise range can be obtained by considering both the current distribution of weights and the distinct tendencies of each layer. Therefore, we additionally incorporate layer-specific variations using a correction parameter for each layer. This strategy allows us to accurately find" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 481, + 163 + ], + "type": "text", + "content": "the quantization range for weights while incurring only a minimal overhead (0.01% additional storage size and no additional bitOPs)—a significantly smaller impact compared to methods that use dynamic modules. Overall, the contributions of our work include the following:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 171, + 479, + 277 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 171, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 479, + 217 + ], + "type": "text", + "content": "- We introduce the first quantization framework to address the distribution mismatch problem in SR networks without dynamic modules, as compared in Table 1. Our framework updates the SR network to be quantization-friendly and accurate simultaneously." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 219, + 479, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 219, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 479, + 253 + ], + "type": "text", + "content": "- Based on the observations of the distribution mismatch in SR networks, we effectively reduce the mismatch by introducing a cooperative mismatch regularization term and a weight clipping correction term." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 255, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 255, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 139, + 255, + 479, + 277 + ], + "type": "text", + "content": "- Compared to existing approaches to SR quantization, ours achieves state-of-the-art performance with similar or fewer computations." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 297, + 229, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 297, + 229, + 308 + ], + "spans": [ + { + "bbox": [ + 133, + 297, + 229, + 308 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 321, + 260, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 321, + 260, + 333 + ], + "spans": [ + { + "bbox": [ + 133, + 321, + 260, + 333 + ], + "type": "text", + "content": "2.1 Image Super-Resolution" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 341, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 341, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 133, + 341, + 481, + 448 + ], + "type": "text", + "content": "Convolutional Neural Network (CNN)-based approaches have demonstrated remarkable advancements in the image super-resolution (SR) task [32, 37], but at the cost of substantial computational resources. The intensive computations required by SR networks have spurred interest in developing lightweight SR architectures [10, 22, 23, 28, 54]. Furthermore, various strategies for lightweight SR networks have been explored, including neural architecture search [9, 30, 34, 35, 46], knowledge distillation [22, 23, 53], and pruning [26, 42, 56]. While these methods predominantly focus on reducing network depth or the number of channels, our work specifically aims to lower the precision of floating-point operations through network quantization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 465, + 253, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 465, + 253, + 477 + ], + "spans": [ + { + "bbox": [ + 133, + 465, + 253, + 477 + ], + "type": "text", + "content": "2.2 Network Quantization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 485, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 485, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 133, + 485, + 481, + 556 + ], + "type": "text", + "content": "By mapping 32-bit floating point values of features and weights in convolutional layers to lower-bit representations, network quantization provides a dramatic reduction in computational resources [5,8,14,29,58,59]. Recent works have successfully quantized various networks to low bit-widths with minimal compromise in accuracy [6,12,16, 27,39,49,52]. However, these efforts primarily target high-level vision tasks, whereas networks for low-level vision tasks remain vulnerable to low-bit quantization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 574, + 321, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 574, + 321, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 574, + 321, + 586 + ], + "type": "text", + "content": "2.3 Quantized Super-Resolution Networks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 594, + 481, + 665 + ], + "type": "text", + "content": "In contrast to high-level vision tasks, SR poses different challenges due to its inherent high sensitivity to quantization [24, 40, 48, 51]. Some works have attempted to recover accuracy by modifying the network architecture [2, 25, 51] or by assigning different bits for each image [17, 19, 47] or network stage [38]. However, the primary challenge in quantizing SR networks lies in the vastly distinct feature distributions. To address this, Li et al. [33] adopted a learnable quantization range for different layers. Subsequently," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 99 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "recognizing that the distributions vary not only by layer, but also by channel and input, Hong et al. [18] introduced a channel-wise dynamic quantization function. Additionally, Zhong et al. [57] utilized an input-adaptive dynamic module to tailor quantization ranges for each specific input image. However, such dynamic adaptations of quantization functions during test-time incur non-negligible computational overheads. Instead of relying on test-time adaptive modules, our approach focuses on mitigating the feature mismatch before quantization. Our framework reduces the inherent distribution mismatch in SR networks with minimal overhead, accurately quantizing SR networks without dynamic modules. More recently, Qin et al. [44] introduced additional transformation functions in both the forward and backward processes. However, performance degradation is still evident in ultra-low bit (e.g., 2-bit) scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 265, + 244, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 265, + 244, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 265, + 244, + 279 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 361 + ], + "type": "text", + "content": "In this section, after a brief introduction to network quantization (Section 3.1), we first analyze the mismatch in the features of SR networks (Section 3.2). Subsequently, we propose a solution to reduce this feature mismatch during training (Section 3.3). Additionally, we examine the mismatch in weights across SR networks (Section 3.4) and introduce a quantization range selection scheme that addresses the weight mismatch (Section 3.5). The overall training process is summarized in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 378, + 216, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 216, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 216, + 389 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": "To reduce the heavy computations of convolutional and linear layers in neural networks, the input features (activations) and weights of each convolution/linear layer are quantized to low-bit values [5,8,15,29]. Given the input feature of the " + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": "-th convolution/linear layer " + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i \\in \\mathbb{R}^{B \\times C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "inline_equation", + "content": "B, C, H" + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": " represent the dimensions of the input batch, channel, height, and width, respectively, a quantization operator " + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": " quantizes the feature " + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": " with bit-width " + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 397, + 482, + 469 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 475, + 481, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 475, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 216, + 475, + 481, + 499 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {X} _ {i}; l, u\\right) = \\operatorname {I n t} \\left(\\frac {\\operatorname {c l i p} \\left(\\boldsymbol {X} _ {i} , l , u\\right) - l}{s}\\right) \\cdot s + l, \\tag {1}", + "image_path": "48f33432a24d72472df8e97b6993cc9db8cf6654780cce8793125bddeac2133e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{clip}(\\cdot ,l,u)" + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "content": " truncates the input into the quantization range of " + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "inline_equation", + "content": "[l,u]" + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "inline_equation", + "content": "s = \\frac{u - l}{2^b - 1}" + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "content": ". After truncation, the truncated feature is scaled to " + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "inline_equation", + "content": "[0,2^b -1]" + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "content": ", then rounded to integer values with " + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{Int}(\\cdot)" + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "content": ", and rescaled to range " + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "inline_equation", + "content": "[l,u]" + }, + { + "bbox": [ + 130, + 507, + 481, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "For activation quantization, the quantization range is defined by " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "[l_a, u_a]" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". To obtain better quantization ranges in SR networks, the clipping parameters " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_a, u_a" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " for each layer are typically learned through quantization-aware training [33, 57]. Since the rounding function is not differentiable, a straight-through estimator (STE) [3] is used to train the clipping parameters in an end-to-end manner. Following [57], we initialize " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_a" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_a" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "(100-j)" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "-th percentile values of the feature, averaged among the training data, where " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " is set to 99 in our experiments to avoid outliers that corrupt the quantization range. For weights, as their distributions tend to be symmetric and the mean approximates zero, we utilize a symmetric quantization function. Thus, for the weight " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " of each convolution/linear layer, the quantization range is defined by " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "[-u_w, u_w]" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 119, + 242, + 205 + ], + "blocks": [ + { + "bbox": [ + 140, + 119, + 242, + 205 + ], + "lines": [ + { + "bbox": [ + 140, + 119, + 242, + 205 + ], + "spans": [ + { + "bbox": [ + 140, + 119, + 242, + 205 + ], + "type": "image", + "image_path": "d92441ffc2ec62642ec83b5e93c005e8a1a3b5ea3320f7b650a5d30bf62c2413.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 210, + 228, + 218 + ], + "lines": [ + { + "bbox": [ + 160, + 210, + 228, + 218 + ], + "spans": [ + { + "bbox": [ + 160, + 210, + 228, + 218 + ], + "type": "text", + "content": "(a) Channel-wise features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 119, + 351, + 204 + ], + "blocks": [ + { + "bbox": [ + 244, + 119, + 351, + 204 + ], + "lines": [ + { + "bbox": [ + 244, + 119, + 351, + 204 + ], + "spans": [ + { + "bbox": [ + 244, + 119, + 351, + 204 + ], + "type": "image", + "image_path": "4fb8a124108b508df44192265acd63ee451d89a09dca2232802e0d91f15135cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 210, + 331, + 218 + ], + "lines": [ + { + "bbox": [ + 269, + 210, + 331, + 218 + ], + "spans": [ + { + "bbox": [ + 269, + 210, + 331, + 218 + ], + "type": "text", + "content": "(b) Image-wise features" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 353, + 124, + 474, + 222 + ], + "blocks": [ + { + "bbox": [ + 353, + 124, + 474, + 222 + ], + "lines": [ + { + "bbox": [ + 353, + 124, + 474, + 222 + ], + "spans": [ + { + "bbox": [ + 353, + 124, + 474, + 222 + ], + "type": "image", + "image_path": "0caab5668e1ca1407062e3e85501a6508542407dee4bbca90d549c4f25eea568.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 235, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 235, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 482, + 289 + ], + "type": "text", + "content": "Figure 1: Distribution mismatch in SR networks. Compared to a classification network (e.g., ResNet-18), an SR network (e.g., EDSR) exhibits significant mismatches within the feature distributions across channel and image dimensions. The large distribution mismatch complicates the selection of an appropriate quantization range. Channels and images from the 2nd layer are randomly selected for visualization. Additional results are available in the supplementary material." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 312, + 324, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 324, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 324, + 323 + ], + "type": "text", + "content": "3.2 Distribution Mismatch in SR Networks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 332, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 536 + ], + "type": "text", + "content": "The unfriendliness to quantization in SR networks arises from diverse feature (activation) distributions, as reported in previous studies [18, 33, 57], primarily due to the absence of batch normalization layers in SR networks. As illustrated in Figure 1, where there are notable discrepancies between the channel and image distributions, quantization grids are unnecessarily allocated to regions with minimal feature density. Early SR quantization methods tackled this issue by employing learnable quantization range parameters [33] for each feature. However, even though the quantization-aware training process strives to find the optimal range for each feature, it fails to account for the channel-wise and input-wise variance in distributions. This mismatch results in a large quantization error that can impair SR performance. To deal with these distribution mismatches, existing methods adopt different quantization ranges for each channel [18] or input image [18, 57]. Nevertheless, the test-time adaptation modules used to determine these ranges introduce unwanted computational overhead during inference. Thus, our straightforward solution is to pre-adjust the distributions to be quantization-friendly, thereby eliminating the need for additional adaptations at inference. The following sections will introduce a new quantization-aware training scheme designed to resolve the distribution mismatch problem." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 553, + 320, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 320, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 320, + 565 + ], + "type": "text", + "content": "3.3 Cooperative Mismatch Regularization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 572, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 643 + ], + "type": "text", + "content": "Instead of trying to identify a quantization range capable of accommodating diverse feature distributions, our approach aims to regularize the distribution mismatch beforehand. Obtaining an appropriate quantization range for a feature with high image- and channel-wise variance is difficult, as a certain number of channels or images will invariably be distant from the selected quantization range. In this work, we refer to the total distance of each feature from the selected quantization grid as the mismatch," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 235, + 653, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 235, + 653, + 481, + 666 + ], + "type": "interline_equation", + "content": "M \\left(\\boldsymbol {X} _ {i}\\right) = \\left| \\left| \\boldsymbol {X} _ {i} - q \\left(\\boldsymbol {X} _ {i}; l _ {a}, u _ {a}\\right) \\right| \\right| _ {2}, \\tag {2}", + "image_path": "8f7ce3b82e195edb5832f371fb62db5f45b9944021eca43444ad506cd57697db.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 115, + 252, + 190 + ], + "blocks": [ + { + "bbox": [ + 150, + 115, + 252, + 190 + ], + "lines": [ + { + "bbox": [ + 150, + 115, + 252, + 190 + ], + "spans": [ + { + "bbox": [ + 150, + 115, + 252, + 190 + ], + "type": "image", + "image_path": "4202b0a395949f8651e631f3770f7db57f4318e0e989692d6c27c8c40890ff9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 194, + 236, + 204 + ], + "lines": [ + { + "bbox": [ + 164, + 194, + 236, + 204 + ], + "spans": [ + { + "bbox": [ + 164, + 194, + 236, + 204 + ], + "type": "text", + "content": "(a) Gradient conflict [13]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 115, + 358, + 191 + ], + "blocks": [ + { + "bbox": [ + 255, + 115, + 358, + 191 + ], + "lines": [ + { + "bbox": [ + 255, + 115, + 358, + 191 + ], + "spans": [ + { + "bbox": [ + 255, + 115, + 358, + 191 + ], + "type": "image", + "image_path": "0c48f56800665b41efc99d9f91fc9d9a3ea69907a9ecd99714b4a3c7ab8dcdef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 291 + ], + "type": "text", + "content": "Figure 2: Conflict between mismatch regularization and reconstruction loss. Mismatch regularization updates a number of parameters in the contradictory direction to the reconstruction loss, which we refer to as gradient conflict. (b) When the two losses are jointly used, gradient conflict consistently occurs during training, outputting a negative cosine similarity value. (c) We plot the ratio of conflicted gradients during training. Nearly half of the parameters undergo gradient conflict, which indicates that merely combining mismatch regularization with the reconstruction loss can impair SR accuracy. Visualizations are done on EDSR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 116, + 465, + 191 + ], + "blocks": [ + { + "bbox": [ + 254, + 194, + 360, + 205 + ], + "lines": [ + { + "bbox": [ + 254, + 194, + 360, + 205 + ], + "spans": [ + { + "bbox": [ + 254, + 194, + 360, + 205 + ], + "type": "text", + "content": "(b) Cosine similarity of two gradients" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 116, + 465, + 191 + ], + "lines": [ + { + "bbox": [ + 362, + 116, + 465, + 191 + ], + "spans": [ + { + "bbox": [ + 362, + 116, + 465, + 191 + ], + "type": "image", + "image_path": "577d111d740facb7db319f01807c78e6686e30d2d7571d279f6ed47a1d1f8f5d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 194, + 449, + 204 + ], + "lines": [ + { + "bbox": [ + 376, + 194, + 449, + 204 + ], + "spans": [ + { + "bbox": [ + 376, + 194, + 449, + 204 + ], + "type": "text", + "content": "(c) Gradient conflict ratio" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "||\\cdot||_2" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " calculates the Frobenius norm. Further analyses of the definition of mismatch are provided in the supplementary material. We can reduce the overall mismatch by directly regularizing the mismatch of each feature to be quantized. The mismatch regularization loss is obtained by summing the mismatch over all quantized features:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 260, + 367, + 481, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 367, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 260, + 367, + 481, + 400 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M} = \\sum_ {i} ^ {\\# \\text {l a y e r s}} M \\left(\\boldsymbol {X} _ {i}\\right). \\tag {3}", + "image_path": "dec9c6ee91f77225a3d63a482784d00e64e2b28fb13ee84fca4559665573d111.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 406, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 482, + 431 + ], + "type": "text", + "content": "The mismatch regularization loss can be used in line with the original reconstruction loss typically used in the general quantization-aware training pipeline for SR networks:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 252, + 442, + 481, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 442, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 252, + 442, + 481, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\mathcal {L} _ {1} \\left(\\mathcal {Q} \\left(\\boldsymbol {I} _ {L R}\\right), \\boldsymbol {I} _ {H R}\\right), \\tag {4}", + "image_path": "4ab5f4e0289c5d272b05db2eac74ec9833c94cc10d7acd10437fcbd02d16cb9e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "text", + "content": " loss indicates the " + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "text", + "content": " distance between the reconstructed image using the quantized network " + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "text", + "content": " and the ground-truth HR image " + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "inline_equation", + "content": "I_{HR}" + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "text", + "content": ". Then, the optimization of the parameter " + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\theta^t" + }, + { + "bbox": [ + 131, + 459, + 482, + 495 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 214, + 501, + 481, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 501, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 214, + 501, + 481, + 515 + ], + "type": "interline_equation", + "content": "\\theta^ {t + 1} = \\theta^ {t} - \\beta^ {t} \\cdot \\left(\\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right) + \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right), \\tag {5}", + "image_path": "4a984ba7ed3b6b96492bcc184e988c974e54e1627f309aaed9edfb6ac10ac5eb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_R(\\theta^t)" + }, + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "text", + "content": " denotes the gradient from the original reconstruction loss and " + }, + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_M(\\theta^t)" + }, + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "text", + "content": " is the gradient from mismatch regularization loss, and " + }, + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "inline_equation", + "content": "\\beta^{t}" + }, + { + "bbox": [ + 130, + 521, + 492, + 569 + ], + "type": "text", + "content": " refers to the learning rate. Updating the network to minimize the mismatch regularization loss will reduce the overall error from feature quantization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "However, then a question arises: does reducing the quantization error of each feature lead to improved reconstruction accuracy? The answer is, according to our observation in Figure 2, not necessarily. During the training process, the mismatch regularization loss can collide with the original reconstruction loss. That is, for some parameters, the direction of the gradient from reconstruction loss and that of the mismatch regularization are opposing, referred to as gradient conflict [13]. As in Figure 2b, the cosine similarity of two gradients oscillates between positive and negative values during training, indicating that the directions of two gradients do not converge and the gradient" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 115, + 251, + 188 + ], + "blocks": [ + { + "bbox": [ + 152, + 115, + 251, + 188 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 251, + 188 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 251, + 188 + ], + "type": "image", + "image_path": "64f1105e59ced2ec4a37d7b67bd46c76c604069032719521e323cb524ff0eb57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 191, + 252, + 200 + ], + "lines": [ + { + "bbox": [ + 149, + 191, + 252, + 200 + ], + "spans": [ + { + "bbox": [ + 149, + 191, + 252, + 200 + ], + "type": "text", + "content": "(a) Layer-wise QE using max policy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 116, + 356, + 188 + ], + "blocks": [ + { + "bbox": [ + 258, + 116, + 356, + 188 + ], + "lines": [ + { + "bbox": [ + 258, + 116, + 356, + 188 + ], + "spans": [ + { + "bbox": [ + 258, + 116, + 356, + 188 + ], + "type": "image", + "image_path": "761c64155d31b35d0453355c3a91ea0cb50014020ce0fa63629d7ab7f6973068.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 191, + 344, + 201 + ], + "lines": [ + { + "bbox": [ + 269, + 191, + 344, + 201 + ], + "spans": [ + { + "bbox": [ + 269, + 191, + 344, + 201 + ], + "type": "text", + "content": "(b) 4-bit quantization grid" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 209, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 131, + 209, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 482, + 288 + ], + "type": "text", + "content": "Figure 3: Layer-wise variation in error from weight quantization. (a) Quantization error (QE) varies across different layers when a fixed global policy (i.e., max) is used to determine the quantization range, particularly for low bits. For some layers, using max does not effectively serve as a proper policy for quantization range selection. (b) Outliers often dominate the quantization range, leading to quantization grids being wasted on low-density areas. (c) At low bits, quantization grids fail to cover high-density regions adequately. Therefore, the quantization range should be adjusted for certain layers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 364, + 116, + 462, + 188 + ], + "blocks": [ + { + "bbox": [ + 364, + 116, + 462, + 188 + ], + "lines": [ + { + "bbox": [ + 364, + 116, + 462, + 188 + ], + "spans": [ + { + "bbox": [ + 364, + 116, + 462, + 188 + ], + "type": "image", + "image_path": "505d8e18b256eb5946f6611527585c4ba7a19fcdb11468ceab4808cef71b9080.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 191, + 450, + 200 + ], + "lines": [ + { + "bbox": [ + 376, + 191, + 450, + 200 + ], + "spans": [ + { + "bbox": [ + 376, + 191, + 450, + 200 + ], + "type": "text", + "content": "(c) 2-bit quantization grid" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 309, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 344 + ], + "type": "text", + "content": "occasionally conflicts. Furthermore, as shown in Figure 2c, the proportion of parameters undergoing gradient conflict is not minor, implying that the regularization loss can severely hinder the reconstruction loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 346, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 429 + ], + "type": "text", + "content": "We aim to avoid the conflict between these two losses, in other words, to minimize the mismatch as long as it does not hinder the reconstruction loss. Thus, we dismiss the mismatch regularization term when it is not cooperative with reconstruction loss and make more use of it when it is cooperative. Specifically, we determine whether the two losses are cooperative by examining the cosine similarity of the gradients of each loss and then simply weigh the gradient of mismatch regularization by the gradient similarity. Our cooperative mismatch regularization can be formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 436, + 482, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 436, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 135, + 436, + 482, + 452 + ], + "type": "interline_equation", + "content": "\\theta^ {t + 1} = \\theta^ {t} - \\beta^ {t} \\cdot \\left(\\lambda_ {R} \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right) + \\lambda_ {M} \\cdot s i m \\left(\\nabla_ {\\theta} \\mathcal {L} _ {R} \\left(\\theta^ {t}\\right), \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right) \\cdot \\nabla_ {\\theta} \\mathcal {L} _ {M} \\left(\\theta^ {t}\\right)\\right), \\tag {6}", + "image_path": "d6fa004ba18016556239de843faf130b10f6e54a9a3c5449a88af71120577d2b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": "where the underlined term, gradient similarity, is defined as " + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\text{sim}(\\pmb{v}_a, \\pmb{v}_b) = \\frac{\\cos(\\pmb{v}_a, \\pmb{v}_b) + 1}{2}" + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\cos(\\cdot, \\cdot) \\in [-1, 1]" + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": " calculates the cosine similarity between two vectors. " + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_R, \\lambda_M" + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": " are hyper-parameters to balance the two gradients. If the directions of two gradients are similar (i.e., smaller than " + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "inline_equation", + "content": "90^\\circ" + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": " and closer to " + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "inline_equation", + "content": "0^\\circ" + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": "), the gradient similarity is a large value, then the parameter is updated in the direction where mismatch regularization is also substantially considered. On the contrary, if the two gradients point in the other direction (i.e., larger than " + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "inline_equation", + "content": "90^\\circ" + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": " and closer to " + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "inline_equation", + "content": "180^\\circ" + }, + { + "bbox": [ + 130, + 460, + 482, + 582 + ], + "type": "text", + "content": "), the two losses restrain each other, and the gradient similarity is close to 0. In this case, we follow the gradient of reconstruction loss. This allows the network to reduce the quantization error cooperatively with the reconstruction error. Details on gradient similarity are in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 598, + 303, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 303, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 303, + 611 + ], + "type": "text", + "content": "3.4 Distribution Mismatch in Weights" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "The weight distributions of SR networks have remained relatively unexplored in previous literature. This is because the weights in SR networks typically exhibit bell-shaped distributions, which are considered easier to quantize compared to the long-tailed, input-wise and channel-wise distinct activation distributions. Consequently, many" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 129, + 481, + 306 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 371, + 129 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 371, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 371, + 129 + ], + "type": "text", + "content": "Algorithm 1 Quantization-aware training process of ODM" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": "Input: Pre-trained 32-bit network " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " \nOutput: Quantized network " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ," + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " # iters do for " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ," + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " # layers do if " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " then Initialize activation quantization range " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "[l_a,u_a]" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " Initialize weight quantization range " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "[-u_w,u_w]" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " Given quantization range, obtain " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "q(\\pmb {X}_i;l_a,u_a)" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " using Eq. (1) Given " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\pmb {X}_i" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "q(\\pmb {X}_i;l_a,u_a)" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " , obtain mismatch using Eq. (2) Adjust weight quantization range " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "[-u_w,u_w]" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " using Eq. (9) Given quantization range, obtain " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "q(\\pmb {W}_i; - u_w,u_w)" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " using Eq. (1) Replace " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\pmb {X}_i" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\pmb {W}_i" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "q(\\pmb {X}_i),q(\\pmb {W}_i)" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " Calculate mismatch regularization loss (Eq. (3)) and reconstruction loss (Eq. (4)) Update parameters of " + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 132, + 129, + 481, + 306 + ], + "type": "text", + "content": " with two losses cooperatively using Eq. (6)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 329, + 483, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 483, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 483, + 497 + ], + "type": "text", + "content": "studies [17, 33, 47] simply adopt max quantization for weights, setting the quantization range with the maximum value of the current weight distribution. However, we notice that this is suboptimal and may contribute to low performance when SR networks are quantized to ultra-low bits (e.g., 2-bit). The issue arises because, in some layers, the outliers are far from the distribution mean, as visualized in Figures 3b and 3c. Thus, when the maximum value is used to determine the quantization range for such distributions, the quantization range is dominated by outliers, with quantization grids not allocated to high-density regions (e.g., near 0). This leads to substantial quantization errors that can accumulate and degrade restoration performance. In the case of 4-bit quantization (Figure 3b), although grids still cover high-density regions to an extent, a number of grids are wasted on low-density areas. The problem intensifies with low-bit (2-bit) quantization (Figure 3c), where quantization errors become significantly larger. This observation underscores the need for careful selection of the quantization range for layer-wise weights, particularly in low-bit quantization scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 512, + 278, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 278, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 278, + 525 + ], + "type": "text", + "content": "3.5 Weight Clipping Correction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 530, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 615 + ], + "type": "text", + "content": "Also, we notice that the error from weight quantization varies across different layers, as shown in Figure 3a. For instance, employing the maximum value as the quantization range can be an effective policy for certain layers, yet this policy proves inadequate for others (e.g., the layer shown in Figure 3c). Given the unique tendency of each layer, applying a uniform global policy for selecting the quantization range across all layers is suboptimal. Existing methods [17, 33, 47] utilize a fixed global policy throughout training to set the quantization range clipping parameter " + }, + { + "bbox": [ + 130, + 530, + 482, + 615 + ], + "type": "inline_equation", + "content": "u_{w}" + }, + { + "bbox": [ + 130, + 530, + 482, + 615 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 276, + 620, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 620, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 276, + 620, + 481, + 635 + ], + "type": "interline_equation", + "content": "u _ {w} ^ {t} = f \\left(\\boldsymbol {W} ^ {t}\\right), \\tag {7}", + "image_path": "541f5a1c4560277cfb60606f70a9615b7b2626f08c8322156117ebae01600761.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "where the global policy " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the same function for all layers (e.g., " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\max(\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "). A simple solution to accommodate layer-specific variations is to make the clipping parameter " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_w" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 135, + 457, + 304 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 412, + 126 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 412, + 126 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 412, + 126 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons on EDSR of scale 4" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 135, + 457, + 304 + ], + "lines": [ + { + "bbox": [ + 157, + 135, + 457, + 304 + ], + "spans": [ + { + "bbox": [ + 157, + 135, + 457, + 304 + ], + "type": "table", + "html": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
EDSR3232.100.89428.580.78127.560.73626.040.785
EDSR-PAMS431.590.88528.200.77327.320.72825.320.762
EDSR-DAQ431.850.88728.380.77627.420.73225.730.772
EDSR-DDTB431.850.88928.390.77727.440.73225.690.774
EDSR-ODM (Ours)432.000.89128.470.77927.510.73525.800.778
EDSR-PAMS327.250.78025.240.67325.380.64422.760.641
EDSR-DAQ331.660.88428.190.77127.280.72825.400.762
EDSR-DDTB331.520.88328.180.77127.300.72725.330.761
EDSR-ODM (Ours)331.850.88828.380.77627.430.73225.590.771
EDSR-PAMS229.510.83526.790.73426.450.69623.720.688
EDSR-DAQ†231.010.87127.890.76227.090.71924.880.740
EDSR-DDTB230.970.87627.870.76427.090.71924.820.742
EDSR-ODM (Ours)231.500.88228.140.77027.270.72625.170.755
", + "image_path": "601fb6c7c1f802781b422fa3656a700b51a560506fcbca99c9b16709a9c21be2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 304, + 481, + 323 + ], + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 323 + ], + "type": "text", + "content": "† We note that the reported results of DAQ [18] are obtained using EDSR of 32 residual blocks. For a fair comparison, we reproduce EDSR-DAQ using EDSR of 16 residual blocks (i.e., EDSR-baseline)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 346, + 299, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 299, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 299, + 357 + ], + "type": "text", + "content": "for each layer a learnable parameter [14]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 365, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 365, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 239, + 365, + 481, + 380 + ], + "type": "interline_equation", + "content": "u _ {w} ^ {t + 1} = u _ {w} ^ {t} - \\beta^ {t} \\cdot \\nabla_ {u _ {w}} \\mathcal {L} _ {R} \\left(u _ {w} ^ {t}\\right), \\tag {8}", + "image_path": "cec646eb427b65d801b2afc2f6f5215317223375106930c70356e259122dd428.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\beta^t" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " denotes the learning rate. This process determines the clipping parameter " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "u_w^t" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " to quantize " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^t" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " based on the weight of the previous iteration, " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{t-1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": ". However, since the weight is also updated at iteration step " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{t-1} \\to \\mathbf{W}^t" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": "), a mismatch occurs between the current weight and the weight quantization range derived from the previous weight. To address this, we first obtain the quantization range by applying the global policy to the current weight, then adjust the range with a learnable parameter that accounts for layer-specific tendencies. Our clipping parameter is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 213, + 479, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 479, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 213, + 479, + 481, + 495 + ], + "type": "interline_equation", + "content": "u _ {w} ^ {t + 1} = f \\left(\\boldsymbol {W} ^ {t + 1}\\right) \\cdot \\left(\\gamma_ {w} ^ {t} - \\beta^ {t} \\cdot \\nabla_ {\\gamma_ {w}} \\mathcal {L} _ {R} \\left(\\gamma_ {w} ^ {t}\\right)\\right), \\tag {9}", + "image_path": "667d763028e4b5a27595ddbba61769fffecb51789d5c4f8f32c1c50a20298404.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\gamma_w" + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "text", + "content": " is the learnable parameter for each layer representing the layer-wise adjustment. Each " + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\gamma_w" + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "text", + "content": " is initially set to 1. To prevent outliers from dominating the initial quantization range, we set the global policy " + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 502, + 482, + 564 + ], + "type": "text", + "content": "-th percentile function. Our clipping correction scheme introduces only one additional parameter per layer; thus, the overall computational overhead is minimal. For further details, please refer to Section 4.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 580, + 218, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 218, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 218, + 594 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The efficacy and adaptability of the proposed quantization framework, ODM, are assessed through its application across several SR networks. The experimental settings are described (Section 4.1), and quantitative (Section 4.2), qualitative (Section 4.4), and complexity (Section 4.3) evaluations are conducted on various SR networks. Ablation studies are conducted to examine each component of the framework (Section 4.5)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 136, + 454, + 304 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 410, + 126 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 410, + 126 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 410, + 126 + ], + "type": "text", + "content": "Table 3: Quantitative comparisons on RDN of scale 4" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 136, + 454, + 304 + ], + "lines": [ + { + "bbox": [ + 159, + 136, + 454, + 304 + ], + "spans": [ + { + "bbox": [ + 159, + 136, + 454, + 304 + ], + "type": "table", + "html": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
RDN3232.240.89628.670.78427.630.73826.290.792
RDN-PAMS430.440.86227.540.75326.870.71024.520.726
RDN-DAQ431.910.88928.380.77527.380.73325.810.779
RDN-DDTB431.970.89128.490.78027.490.73525.900.783
RDN-ODM (Ours)432.060.89228.490.78027.520.73625.880.783
RDN-PAMS329.540.83826.820.73426.470.69623.830.692
RDN-DAQ331.570.88328.180.77127.270.72825.470.765
RDN-DDTB331.490.88328.170.77227.300.72825.350.764
RDN-ODM (Ours)331.790.88728.330.77627.420.73225.510.770
RDN-PAMS229.730.84326.960.73926.570.70023.870.696
RDN-DAQ230.710.86627.610.75526.930.71524.710.731
RDN-DDTB230.570.86727.560.75726.910.71424.500.728
RDN-ODM (Ours)231.370.88028.080.77027.240.72725.090.755
", + "image_path": "3313582ce5ea6b7f65e274bddcd4b112ab128e24d7f42db284e859a9de38f75c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 257, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 257, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 257, + 339 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "content": "Models and Training. The proposed framework is applied directly to existing representative SR networks that produce satisfactory SR results, but involve heavy computations: EDSR (baseline) [37] and RDN [55]. Furthermore, we apply our method to the Transformer-based SR model, SwinIR-S [36]. Following prior works on SR quantization [17, 18, 33, 40, 51, 57], weights and activations of the high-level feature extraction module are quantized, which is the most computationally demanding. Training and validation are conducted using the DIV2K [1] dataset. ODM trains the network for 60K iterations with a batch size of 8. The weights are updated with an initial learning rate of " + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\beta^0 = 10^{-4}" + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "content": ". For cooperative update, we update clipping parameters with " + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "inline_equation", + "content": "10\\cdot \\beta^0" + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "content": " initial learning rate. The learning rates are halved every 15K iteration. The hyperparameter for the percentile is set to " + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "inline_equation", + "content": "j = 99" + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "content": ", and to balance the gradient of the loss terms, we set " + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{R} = 1" + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{M} = 10^{-5}" + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "content": ". Specially, we set " + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{M} = 10^{-6}" + }, + { + "bbox": [ + 132, + 346, + 482, + 514 + ], + "type": "text", + "content": " for RDN whose overall mismatch is large. Ablation studies on the hyperparameters are in the supplementary material. All our experiments are implemented using PyTorch and run on an RTX 2080Ti GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 532, + 487, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 487, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 487, + 604 + ], + "type": "text", + "content": "Evaluation. We evaluate our framework on the standard benchmark (Set5 [4], Set14 [32], BSD100 [41], and Urban100 [21]) by measuring the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM [50]). To assess the computational complexity of our framework, we measure bitOPs and storage size. BitOPs refers to the number of operations weighted by the bit-widths of the two operands. Storage size is calculated as the number of stored parameters weighted by the precision of each parameter value." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 621, + 244, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 244, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 244, + 633 + ], + "type": "text", + "content": "4.2 Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed scheme, we compare the results with existing SR quantization works using their official codes: PAMS [33], DAQ [18], and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 92, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 92, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 92, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 136, + 457, + 304 + ], + "blocks": [ + { + "bbox": [ + 200, + 114, + 414, + 125 + ], + "lines": [ + { + "bbox": [ + 200, + 114, + 414, + 125 + ], + "spans": [ + { + "bbox": [ + 200, + 114, + 414, + 125 + ], + "type": "text", + "content": "Table 4: Quantitative comparisons on SwinIR of scale 4" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 136, + 457, + 304 + ], + "lines": [ + { + "bbox": [ + 156, + 136, + 457, + 304 + ], + "spans": [ + { + "bbox": [ + 156, + 136, + 457, + 304 + ], + "type": "table", + "html": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-PAMS431.990.89028.500.77927.490.73525.860.780
SwinIR-DAQ431.820.88728.340.77527.370.73025.680.772
SwinIR-DDTB432.090.89128.550.78027.540.73526.010.783
SwinIR-ODM (Ours)432.170.89228.590.78127.560.73626.060.785
SwinIR-PAMS331.620.88428.230.77127.310.72825.380.762
SwinIR-DAQ331.500.88227.990.77027.120.72725.290.761
SwinIR-DDTB331.800.88728.340.77527.400.73125.630.771
SwinIR-ODM (Ours)331.940.88928.390.77727.450.73325.720.775
SwinIR-PAMS229.480.83426.790.73326.460.69823.720.688
SwinIR-DAQ229.100.82426.550.72526.300.69123.510.678
SwinIR-DDTB231.010.87327.800.76227.040.71924.790.739
SwinIR-ODM (Ours)231.440.88028.060.76927.230.72525.140.754
", + "image_path": "3a692a0cd6455135f82ba27e37f82d65a2c22d714dcbaf6d97f76202130d4b89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 344, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 344, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 344, + 481, + 392 + ], + "type": "text", + "content": "DDTB [57]. For a fair comparison, we reproduce other methods using the same training iterations, 60K iterations. The supplementary materials provide additional experiments that further demonstrate the applicability of ODM, including results on 300K iterations, scale 2, and fully quantized settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 400, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 400, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 400, + 481, + 460 + ], + "type": "text", + "content": "EDSR. As shown in Table 2, ODM outperforms other methods in the 4, 3, and 2-bit settings, and notably, the improvement is significant for 2-bit, achieving a gain of more than " + }, + { + "bbox": [ + 133, + 400, + 481, + 460 + ], + "type": "inline_equation", + "content": "0.49\\mathrm{dB}" + }, + { + "bbox": [ + 133, + 400, + 481, + 460 + ], + "type": "text", + "content": " for Set5. We notice that 4-bit EDSR-ODM achieves closer accuracy to 32-bit EDSR, with a marginal difference of " + }, + { + "bbox": [ + 133, + 400, + 481, + 460 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{dB}" + }, + { + "bbox": [ + 133, + 400, + 481, + 460 + ], + "type": "text", + "content": " for Set5. This indicates that ODM can effectively bridge the gap between the quantized network and the 32-bit network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 468, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 481, + 517 + ], + "type": "text", + "content": "RDN. Similarly, Table 3 compares the results on RDN, whose computational complexity is more burdensome than EDSR. The results show that ODM consistently achieves superior performance on 4, 3, and 2-bit quantization. The gain over existing methods is especially large for the 2-bit setting, where it exceeds 0.66 dB for Set5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 525, + 481, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 481, + 573 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 481, + 573 + ], + "type": "text", + "content": "SwinIR. Furthermore, we evaluate our framework on the Transformer-based architecture, SwinIR. The linear and convolutional layers of SwinIR are quantized. According to Table 4, ODM is also proven effective in quantizing SwinIR across all bit settings, where the improvement is most notable in the 2-bit setting (0.43 dB)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 581, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 581, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 581, + 481, + 665 + ], + "type": "text", + "content": "Comparison with QuantSR. We also compare our method with the concurrent work, QuantSR [44]. As the training code of QuantSR has not been released, we base our comparison on the reported performance. For a fair comparison with QuantSR's reported performance, we also train our model for 300K iterations on SRResNet [32] and SwinIR [36]. In Table 5, the results demonstrate that our method achieves better results than QuantSR; compared to QuantSR, ours shows a gain of 0.51 dB on 2-bit SRResNet and a gain of 0.14 dB on 2-bit SwinIR for Set5." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 146, + 463, + 254 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 5: Quantitative comparisons with QuantSR on SRResNet and SwinIR of scale 4. For a fair comparison, our model (ODM*) is trained for 300K iterations following QuantSR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 146, + 463, + 254 + ], + "lines": [ + { + "bbox": [ + 149, + 146, + 463, + 254 + ], + "spans": [ + { + "bbox": [ + 149, + 146, + 463, + 254 + ], + "type": "table", + "html": "
ModelBitSet5Set14B100Urban100
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SRResNet3232.070.89328.500.78027.520.73525.860.779
SRResNet-QuantSR231.300.88228.080.76927.230.72525.130.754
SRResNet-ODM*(Ours)231.810.88828.320.77427.380.73025.540.767
SwinIR3232.440.89828.770.78627.690.74126.470.798
SwinIR-QuantSR231.530.88528.160.77227.280.72725.260.761
SwinIR-ODM*(Ours)231.670.88528.230.77227.330.72825.360.762
", + "image_path": "c4e523073a7c9532acb32136b83aed3d88406966bafc75a98546239cc1e5eaa8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 286, + 310, + 344 + ], + "blocks": [ + { + "bbox": [ + 159, + 265, + 455, + 277 + ], + "lines": [ + { + "bbox": [ + 159, + 265, + 455, + 277 + ], + "spans": [ + { + "bbox": [ + 159, + 265, + 455, + 277 + ], + "type": "text", + "content": "Table 6: Computational complexity comparison with SR quantization methods" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 286, + 310, + 344 + ], + "lines": [ + { + "bbox": [ + 138, + 286, + 310, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 310, + 344 + ], + "type": "table", + "html": "
ModelBitStorage sizeBitOPsPSNRSSIM
EDSR321517.6K527.1T32.100.894
EDSR-PAMS2411.7K215.1T29.510.835
EDSR-DAQ2411.7K215.6T31.010.871
EDSR-DDTB2413.6K215.1T30.970.876
EDSR-ODM (Ours)2411.8K215.1T31.500.882
", + "image_path": "d52149543286d58e35f3522e09e57386533aab09bf120df26eac15a8f9725d44.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 208, + 346, + 237, + 354 + ], + "lines": [ + { + "bbox": [ + 208, + 346, + 237, + 354 + ], + "spans": [ + { + "bbox": [ + 208, + 346, + 237, + 354 + ], + "type": "text", + "content": "(a) EDSR" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 286, + 488, + 344 + ], + "blocks": [ + { + "bbox": [ + 316, + 286, + 488, + 344 + ], + "lines": [ + { + "bbox": [ + 316, + 286, + 488, + 344 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 488, + 344 + ], + "type": "table", + "html": "
ModelBitStorage sizeBitOPsPSNRSSIM
RDN3222271.1K6032.9T32.240.896
RDN-PAMS21715.9K236.6T29.730.843
RDN-DAQ21715.9K287.7T30.710.866
RDN-DDTB21761.6K236.6T30.570.867
RDN-ODM (Ours)21716.1K236.6T31.370.880
", + "image_path": "8c25a971270eeb087a805d8743f5b7c54f6963edcbe149932dcadd91753dd3d1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 388, + 346, + 414, + 354 + ], + "lines": [ + { + "bbox": [ + 388, + 346, + 414, + 354 + ], + "spans": [ + { + "bbox": [ + 388, + 346, + 414, + 354 + ], + "type": "text", + "content": "(b) RDN" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 388, + 246, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 246, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 246, + 401 + ], + "type": "text", + "content": "4.3 Complexity Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": "Along with the accuracy of SR, we also evaluate the computational complexity of our framework in Table 6. We measure the storage size for the model weights and the bitOPs required for generating a " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " image with " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " SR network. Overall, our framework, ODM, achieves higher restoration accuracy with similar or fewer computational resources. Specifically, our weight-clipping correction involves an additional storage size of " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "0.06\\mathrm{K} / 0.15\\mathrm{K}" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " for EDSR/RDN. Since the correction process can be predetermined before test time, no extra bitOPs are required. Compared to existing methods, our method achieves " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "\\times 31.7" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " smaller storage size overhead than DDTB and " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{T}" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " fewer bitOPs than DAQ on EDSR. For RDN, the gap is larger; our method's overhead is " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "\\times 304.7" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " smaller in storage size than DDTB and " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "51.1\\mathrm{T}" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " fewer in bitOPs than DAQ. Moreover, we note that DAQ adopts a channel-wise dynamic quantization function and DDTB an asymmetric weight quantization function, which are not favorably supported by hardware. Despite incurring a minor additional storage size of " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "0.06\\mathrm{K} / 0.15\\mathrm{K}" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": " over PAMS, the accuracy improvement over PAMS is significant " + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "inline_equation", + "content": "(+1.99\\mathrm{dB} / +1.64\\mathrm{dB})" + }, + { + "bbox": [ + 130, + 410, + 483, + 602 + ], + "type": "text", + "content": ". The results prove that our method achieves significant accuracy gain with minimal or no computational overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 620, + 239, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 239, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 239, + 632 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Figure 4 provides qualitative results and comparisons with the output images of quantized EDSR, RDN, and SwinIR. Our method, ODM, produces visually cleaner output" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 203, + 178 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 203, + 178 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 203, + 178 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 203, + 178 + ], + "type": "image", + "image_path": "48f31ccff2149a662116a591268ea1d85671aa8092085a36e6c3288afb605d3d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 114, + 271, + 178 + ], + "blocks": [ + { + "bbox": [ + 206, + 114, + 271, + 178 + ], + "lines": [ + { + "bbox": [ + 206, + 114, + 271, + 178 + ], + "spans": [ + { + "bbox": [ + 206, + 114, + 271, + 178 + ], + "type": "image", + "image_path": "bd9fc4437f955de8ad27668a738c32de34abbc2716ae7ac98327e46e63b3e1ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 275, + 114, + 339, + 178 + ], + "blocks": [ + { + "bbox": [ + 275, + 114, + 339, + 178 + ], + "lines": [ + { + "bbox": [ + 275, + 114, + 339, + 178 + ], + "spans": [ + { + "bbox": [ + 275, + 114, + 339, + 178 + ], + "type": "image", + "image_path": "6bd5970ab69c00ebe20b77452ec46237d89a94deeb9d0fec53fc720c48e58c1b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 407, + 178 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 407, + 178 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 407, + 178 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 407, + 178 + ], + "type": "image", + "image_path": "e6ee8476d336b356eb837369dd161a8185082c8dd1eed02580f940ada0041802.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 475, + 178 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 475, + 178 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 475, + 178 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 475, + 178 + ], + "type": "image", + "image_path": "d3d8666199992305fd506fe9f9c341f5ed64987eafbac29a6e2056e9082d5c02.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 138, + 191, + 203, + 256 + ], + "blocks": [ + { + "bbox": [ + 151, + 183, + 190, + 191 + ], + "lines": [ + { + "bbox": [ + 151, + 183, + 190, + 191 + ], + "spans": [ + { + "bbox": [ + 151, + 183, + 190, + 191 + ], + "type": "text", + "content": "GT (img052)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 191, + 203, + 256 + ], + "lines": [ + { + "bbox": [ + 138, + 191, + 203, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 203, + 256 + ], + "type": "image", + "image_path": "9a50e368c941ce706fb8ce96038b518d158e5424ab6a3b527191de290e3da13a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 207, + 191, + 271, + 255 + ], + "blocks": [ + { + "bbox": [ + 219, + 183, + 259, + 190 + ], + "lines": [ + { + "bbox": [ + 219, + 183, + 259, + 190 + ], + "spans": [ + { + "bbox": [ + 219, + 183, + 259, + 190 + ], + "type": "text", + "content": "EDSR-PAMS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 207, + 191, + 271, + 255 + ], + "lines": [ + { + "bbox": [ + 207, + 191, + 271, + 255 + ], + "spans": [ + { + "bbox": [ + 207, + 191, + 271, + 255 + ], + "type": "image", + "image_path": "94ac52e687adaec83ce44802358a1af6b281543c2eb634705d46b9e899b6822d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 275, + 191, + 339, + 255 + ], + "blocks": [ + { + "bbox": [ + 289, + 183, + 324, + 190 + ], + "lines": [ + { + "bbox": [ + 289, + 183, + 324, + 190 + ], + "spans": [ + { + "bbox": [ + 289, + 183, + 324, + 190 + ], + "type": "text", + "content": "EDSR-DAQ" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 275, + 191, + 339, + 255 + ], + "lines": [ + { + "bbox": [ + 275, + 191, + 339, + 255 + ], + "spans": [ + { + "bbox": [ + 275, + 191, + 339, + 255 + ], + "type": "image", + "image_path": "f5bb72cfabcf01540dafd10e4205225b8ccd6f774e898e3769955e95ea515ce5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 342, + 191, + 406, + 255 + ], + "blocks": [ + { + "bbox": [ + 354, + 183, + 394, + 190 + ], + "lines": [ + { + "bbox": [ + 354, + 183, + 394, + 190 + ], + "spans": [ + { + "bbox": [ + 354, + 183, + 394, + 190 + ], + "type": "text", + "content": "EDSR-DDTB" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 191, + 406, + 255 + ], + "lines": [ + { + "bbox": [ + 342, + 191, + 406, + 255 + ], + "spans": [ + { + "bbox": [ + 342, + 191, + 406, + 255 + ], + "type": "image", + "image_path": "87009efb5c6a912e285bad0cf882d8fc15bbf626bfb9c69c4e74699fd96d39f4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 410, + 191, + 475, + 255 + ], + "blocks": [ + { + "bbox": [ + 414, + 183, + 471, + 190 + ], + "lines": [ + { + "bbox": [ + 414, + 183, + 471, + 190 + ], + "spans": [ + { + "bbox": [ + 414, + 183, + 471, + 190 + ], + "type": "text", + "content": "EDSR-ODM (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 191, + 475, + 255 + ], + "lines": [ + { + "bbox": [ + 410, + 191, + 475, + 255 + ], + "spans": [ + { + "bbox": [ + 410, + 191, + 475, + 255 + ], + "type": "image", + "image_path": "42374a7692ca11d70464905bf91d12d9cb58183504d900c0fed7489e1c5c9607.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 138, + 270, + 203, + 334 + ], + "blocks": [ + { + "bbox": [ + 151, + 261, + 190, + 269 + ], + "lines": [ + { + "bbox": [ + 151, + 261, + 190, + 269 + ], + "spans": [ + { + "bbox": [ + 151, + 261, + 190, + 269 + ], + "type": "text", + "content": "GT (img060)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 270, + 203, + 334 + ], + "lines": [ + { + "bbox": [ + 138, + 270, + 203, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 270, + 203, + 334 + ], + "type": "image", + "image_path": "8148fe4b70330995e0f2858bc25e553e1867d20f02ab11318201ba2ed4260740.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 338, + 190, + 347 + ], + "lines": [ + { + "bbox": [ + 151, + 338, + 190, + 347 + ], + "spans": [ + { + "bbox": [ + 151, + 338, + 190, + 347 + ], + "type": "text", + "content": "GT (img024)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 144, + 357, + 468, + 369 + ], + "lines": [ + { + "bbox": [ + 144, + 357, + 468, + 369 + ], + "spans": [ + { + "bbox": [ + 144, + 357, + 468, + 369 + ], + "type": "text", + "content": "Figure 4: Qualitative results on Urban100 with EDSR, SwinIR, and RDN-based models" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 207, + 270, + 271, + 334 + ], + "blocks": [ + { + "bbox": [ + 217, + 261, + 260, + 269 + ], + "lines": [ + { + "bbox": [ + 217, + 261, + 260, + 269 + ], + "spans": [ + { + "bbox": [ + 217, + 261, + 260, + 269 + ], + "type": "text", + "content": "SwinIR-PAMS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 207, + 270, + 271, + 334 + ], + "lines": [ + { + "bbox": [ + 207, + 270, + 271, + 334 + ], + "spans": [ + { + "bbox": [ + 207, + 270, + 271, + 334 + ], + "type": "image", + "image_path": "957c19027a7689b2252fbfebd423a03b41af2f32a44ce9408fe80b793fa4e098.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 338, + 257, + 346 + ], + "lines": [ + { + "bbox": [ + 220, + 338, + 257, + 346 + ], + "spans": [ + { + "bbox": [ + 220, + 338, + 257, + 346 + ], + "type": "text", + "content": "RDN-PAMS" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 275, + 270, + 339, + 334 + ], + "blocks": [ + { + "bbox": [ + 287, + 261, + 326, + 269 + ], + "lines": [ + { + "bbox": [ + 287, + 261, + 326, + 269 + ], + "spans": [ + { + "bbox": [ + 287, + 261, + 326, + 269 + ], + "type": "text", + "content": "SwinIR-DAQ" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 275, + 270, + 339, + 334 + ], + "lines": [ + { + "bbox": [ + 275, + 270, + 339, + 334 + ], + "spans": [ + { + "bbox": [ + 275, + 270, + 339, + 334 + ], + "type": "image", + "image_path": "102846df4210d3a34ea014f5e39104b5e8821d2b5229dc78f9eb68be84bde719.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 338, + 323, + 346 + ], + "lines": [ + { + "bbox": [ + 290, + 338, + 323, + 346 + ], + "spans": [ + { + "bbox": [ + 290, + 338, + 323, + 346 + ], + "type": "text", + "content": "RDN-DAQ" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 342, + 270, + 407, + 334 + ], + "blocks": [ + { + "bbox": [ + 353, + 261, + 396, + 269 + ], + "lines": [ + { + "bbox": [ + 353, + 261, + 396, + 269 + ], + "spans": [ + { + "bbox": [ + 353, + 261, + 396, + 269 + ], + "type": "text", + "content": "SwinIR-DDTB" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 270, + 407, + 334 + ], + "lines": [ + { + "bbox": [ + 342, + 270, + 407, + 334 + ], + "spans": [ + { + "bbox": [ + 342, + 270, + 407, + 334 + ], + "type": "image", + "image_path": "31c11b324d8fd9c3c087792ad483b28c3f2a8928111d655ff31b99b40854ba3e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 338, + 393, + 346 + ], + "lines": [ + { + "bbox": [ + 356, + 338, + 393, + 346 + ], + "spans": [ + { + "bbox": [ + 356, + 338, + 393, + 346 + ], + "type": "text", + "content": "RDN-DDTB" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 410, + 270, + 475, + 334 + ], + "blocks": [ + { + "bbox": [ + 412, + 261, + 473, + 269 + ], + "lines": [ + { + "bbox": [ + 412, + 261, + 473, + 269 + ], + "spans": [ + { + "bbox": [ + 412, + 261, + 473, + 269 + ], + "type": "text", + "content": "SwinIR-ODM (Ours)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 270, + 475, + 334 + ], + "lines": [ + { + "bbox": [ + 410, + 270, + 475, + 334 + ], + "spans": [ + { + "bbox": [ + 410, + 270, + 475, + 334 + ], + "type": "image", + "image_path": "a9377a151552375a06ebd3bc5b814aac447ba04db828c248af4eb8b8f3f0b28a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 338, + 470, + 346 + ], + "lines": [ + { + "bbox": [ + 415, + 338, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 415, + 338, + 470, + 346 + ], + "type": "text", + "content": "RDN-ODM (Ours)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "content": "images compared to existing methods. In contrast, existing methods, especially PAMS, suffer from blurred lines or artifacts. These qualitative results stress the importance of alleviating the distribution mismatch problem in SR networks." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 131, + 455, + 222, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 222, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 222, + 467 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": "In Table 7, we verify the importance of each attribute of our framework: cooperative mismatch regularization and weight clipping correction. According to the results, each attribute individually improves baseline accuracy. Weight clipping correction improves the baseline by " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "+1.18" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " dB for Set5, indicating that considering both the layer-wise trend and the current weight distribution is important for weight quantization. Also, while directly integrating mismatch regularization with the reconstruction loss (Model (d)) rather degrades performance by " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "-0.13" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " dB, our cooperative scheme (Model (e)) improves the SR accuracy by " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "+0.51" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " dB. This shows that reducing the mismatch in both activations and weights is important for accurately quantizing SR networks." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, we visualize the feature distributions in Figure 5 to validate the importance of our cooperative regularization term. After training only with the reconstruction loss, the outliers remain far from the quantization grid. When mismatch regularization loss is simply added to the original reconstruction loss, the activation distribution falls into a narrower range and resembles a multi-modal distribution near the quantization grids. Although this distribution is quantization-friendly, it significantly deviates from the original activation of the 32-bit network and removes originally dense values (near" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 168, + 451, + 240 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 7: Ablation study on each attribute of our framework. WCC refers to weight clipping correction and MR refers to mismatch regularization. Cooperative denotes whether the mismatch regularization is cooperatively used with the reconstruction loss. Non-cooperative MR denotes that the two losses are simply used together." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 168, + 451, + 240 + ], + "lines": [ + { + "bbox": [ + 160, + 168, + 451, + 240 + ], + "spans": [ + { + "bbox": [ + 160, + 168, + 451, + 240 + ], + "type": "table", + "html": "
ModelWCCCooperativeMRSet5 (PSNR / SSIM)Urban100 (PSNR / SSIM)
(a)---29.94 / 0.84823.99 / 0.703
(b)-✓✓30.34 / 0.85924.27 / 0.715
(c)✓--31.12 / 0.87624.91 / 0.746
(d)✓-✓30.99 / 0.87124.79 / 0.735
(e)✓✓✓31.50 / 0.88225.17 / 0.755
", + "image_path": "b69669853f569d6450d28da44ef5f78ddd9a49f301e4c968a0a66a4f6b63812c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 251, + 227, + 307 + ], + "blocks": [ + { + "bbox": [ + 149, + 251, + 227, + 307 + ], + "lines": [ + { + "bbox": [ + 149, + 251, + 227, + 307 + ], + "spans": [ + { + "bbox": [ + 149, + 251, + 227, + 307 + ], + "type": "image", + "image_path": "cb1d982ff5ed9d4b32ea978e8f62584334404d66cc635ced9f85cddf219e85dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 313, + 211, + 322 + ], + "lines": [ + { + "bbox": [ + 167, + 313, + 211, + 322 + ], + "spans": [ + { + "bbox": [ + 167, + 313, + 211, + 322 + ], + "type": "text", + "content": "32-bit network" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 229, + 251, + 306, + 306 + ], + "blocks": [ + { + "bbox": [ + 229, + 251, + 306, + 306 + ], + "lines": [ + { + "bbox": [ + 229, + 251, + 306, + 306 + ], + "spans": [ + { + "bbox": [ + 229, + 251, + 306, + 306 + ], + "type": "image", + "image_path": "9421d72fd81ed07469ad22e1c1dd5c21da9645ca017142de6d76d72e53001947.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 313, + 273, + 323 + ], + "lines": [ + { + "bbox": [ + 261, + 313, + 273, + 323 + ], + "spans": [ + { + "bbox": [ + 261, + 313, + 273, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 333, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 367 + ], + "type": "text", + "content": "Figure 5: Distribution of activations before quantization. Using our cooperative mismatch regularization results in distributions more robust to low-bit quantization. 8th conv layer of EDSR-ODM (2-bit) on 'baby' (Set5) are visualized." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 251, + 384, + 306 + ], + "blocks": [ + { + "bbox": [ + 309, + 251, + 384, + 306 + ], + "lines": [ + { + "bbox": [ + 309, + 251, + 384, + 306 + ], + "spans": [ + { + "bbox": [ + 309, + 251, + 384, + 306 + ], + "type": "image", + "image_path": "3fb76831ab786402541c5738c45d160c5883dc0554c15cc9181a0eff6f1bedaa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 313, + 364, + 323 + ], + "lines": [ + { + "bbox": [ + 329, + 313, + 364, + 323 + ], + "spans": [ + { + "bbox": [ + 329, + 313, + 364, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R + \\mathcal{L}_M" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 388, + 251, + 463, + 306 + ], + "blocks": [ + { + "bbox": [ + 388, + 251, + 463, + 306 + ], + "lines": [ + { + "bbox": [ + 388, + 251, + 463, + 306 + ], + "spans": [ + { + "bbox": [ + 388, + 251, + 463, + 306 + ], + "type": "image", + "image_path": "57b31d47aa8a7e73a362d9f0e8c69193b3cda3e294d574c690803cd669c0c670.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 313, + 449, + 323 + ], + "lines": [ + { + "bbox": [ + 401, + 313, + 449, + 323 + ], + "spans": [ + { + "bbox": [ + 401, + 313, + 449, + 323 + ], + "type": "text", + "content": "Cooperative MR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 399, + 482, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 436 + ], + "type": "text", + "content": "0). This can lead to an accuracy drop, as demonstrated in Table 7 (d). In contrast, our cooperative mismatch regularization results in quantization-friendly distribution while largely preserving the activations in the dense region of the original 32-bit network." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 453, + 211, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 211, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 211, + 466 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "text", + "content": "SR networks suffer accuracy loss from quantization due to the inherent mismatch in feature distributions. Instead of employing resource-demanding dynamic modules to handle distinct distributions during test time, we introduce a new quantization-aware training technique that alleviates this mismatch through distribution optimization. We utilize cooperative mismatch regularization to update the SR network to be quantization-friendly and accurate. Additionally, to address the mismatch in layer-wise weights, we propose a weight-clipping correction strategy. These straightforward solutions effectively reduce the distribution mismatch with minimal computational overhead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 593, + 224, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 224, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 224, + 605 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A9E4080782] funded by the Korean government (MSIT)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image super-resolution: Dataset and study. In: CVPR Workshops (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Ayazoglu, M.: Extremely lightweight quantization robust real-time single-image super resolution for mobile devices. In: CVPR Workshops (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "text", + "content": "3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "text", + "content": "4. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "5. Cai, Z., He, X., Sun, J., Vasconcelos, N.: Deep learning with low precision by half-wave gaussian quantization. In: CVPR (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "text", + "content": "6. Cai, Z., Vasconcelos, N.: Rethinking differentiable search for mixed-precision neural networks. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "7. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image superresolution transformer. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "8. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "9. Chu, X., Zhang, B., Ma, H., Xu, R., Li, Q.: Fast, accurate and lightweight super-resolution with neural architecture search. In: ICPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "text", + "content": "10. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "text", + "content": "1. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE TPAMI 38(2), 295-307 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "2. Dong, Z., Yao, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Hawq: Hessian aware quantization of neural networks with mixed-precision. In: ICCV (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "text", + "content": "3. Du, Y., Czarnecki, W.M., Jayakumar, S.M., Farajtabar, M., Pascanu, R., Lakshminarayanan, B.: Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "4. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: ICLR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "5. Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A survey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291-326. Chapman and Hall/CRC (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "text", + "content": "6. Habi, H.V., Jennings, R.H., Netzer, A.: Hmq: Hardware friendly mixed precision quantization block for cnns. In: ECCV (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "7. Hong, C., Baik, S., Kim, H., Nah, S., Lee, K.M.: Cadyq: Content-aware dynamic quantization for image super-resolution. In: ECCV (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "8. Hong, C., Kim, H., Baik, S., Oh, J., Lee, K.M.: Daq: Channel-wise distribution-aware quantization for deep image super-resolution networks. In: WACV (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "9. Hong, C., Lee, K.M.: Adabm: On-the-fly adaptive bit mapping for image super-resolution. In: CVPR (2024)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 589, + 468, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 468, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 468, + 599 + ], + "type": "text", + "content": "20. Hou, L., Kwok, J.T.: Loss-aware weight quantization of deep networks. In: ICLR (2018)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "type": "text", + "content": "21. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "22. Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACMMM (2019)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "23. Hui, Z., Wang, X., Gao, X.: Fast and accurate single image super-resolution via information distillation network. In: CVPR (2018)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "24. Ignatov, A., Timofte, R., Denna, M., Younes, A.: Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In: CVPR Workshops (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "25. Jiang, X., Wang, N., Xin, J., Li, K., Yang, X., Gao, X.: Training binary neural network without batch normalization for image super-resolution. In: AAAI (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "26. Jiang, X., Wang, N., Xin, J., Xia, X., Yang, X., Gao, X.: Learning lightweight superresolution networks with weight pruning. Neural Networks 144, 21-32 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "27. Jin, Q., Yang, L., Liao, Z.: Adabits: Neural network quantization with adaptive bit-widths. In: CVPR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "28. Jo, Y., Kim, S.J.: Practical single-image super-resolution using look-up table. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "29. Jung, S., Son, C., Lee, S., Son, J., Han, J.J., Kwak, Y., Hwang, S.J., Choi, C.: Learning to quantize deep networks by optimizing quantization intervals with task loss. In: CVPR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 247, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 277 + ], + "type": "text", + "content": "30. Kim, H., Hong, S., Han, B., Myeong, H., Lee, K.M.: Fine-grained neural architecture search for image super-resolution. Journal of Visual Communication and Image Representation 89, 103654 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 299 + ], + "type": "text", + "content": "31. Kim, J., Lee, J., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "32. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super-resolution using a generative adversarial network. In: CVPR (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 333, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 353 + ], + "type": "text", + "content": "33. Li, H., Yan, C., Lin, S., Zheng, X., Zhang, B., Yang, F., Ji, R.: Pams: Quantized superresolution via parameterized max scale. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "34. Li, Y., Gu, S., Zhang, K., Gool, L.V., Timofte, R.: Dhp: Differentiable meta pruning via hypernetworks. In: ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "35. Li, Y., Li, W., Danelljan, M., Zhang, K., Gu, S., Van Gool, L., Timofte, R.: The heterogeneity hypothesis: Finding layer-wise differentiated network architectures. In: CVPR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "36. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "37. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: CVPR Workshops (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "38. Liu, J., Wang, Q., Zhang, D., Shen, L.: Super-resolution model quantized in multi-precision. Electronics 10(17), 2176 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "39. Lou, Q., Guo, F., Liu, L., Kim, M., Jiang, L.: AutoQ: Automated kernel-wise neural network quantization. In: ICLR (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "text", + "content": "40. Ma, Y., Xiong, H., Hu, Z., Ma, L.: Efficient super resolution using binarized neural network. In: CVPR Workshops (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "41. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "42. Oh, J., Kim, H., Nah, S., Hong, C., Choi, J., Lee, K.M.: Attentive fine-grained structured sparsity for image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "text", + "content": "43. Park, J., Son, S., Lee, K.M.: Content-aware local gan for photo-realistic super-resolution. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 579, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 600 + ], + "type": "text", + "content": "44. Qin, H., Zhang, Y., Ding, Y., Liu, X., Danelljan, M., Yu, F.: Quantsr: Accurate low-bit quantization for efficient image super-resolution. In: NeurIPS (2024)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "text", + "content": "45. Son, S., Kim, J., Lai, W.S., Yang, M.H., Lee, K.M.: Toward real-world super-resolution via adaptive downsampling models. IEEE TPAMI 44(11), 8657-8670 (2021)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "46. Song, D., Xu, C., Jia, X., Chen, Y., Xu, C., Wang, Y.: Efficient residual dense block search for image super-resolution. In: AAAI (2020)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "47. Tian, S., Lu, M., Liu, J., Guo, Y., Chen, Y., Zhang, S.: Cabm: Content-aware bit mapping for single image super-resolution network with large input. In: CVPR (2023)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "C. Hong and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 390 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "48. Wang, H., Chen, P., Zhuang, B., Shen, C.: Fully quantized image super-resolution networks. In: ACMMM (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "text", + "content": "49. Wang, K., Liu, Z., Lin, Y., Lin, J., Han, S.: Haq: Hardware-aware automated quantization with mixed precision. In: CVPR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "50. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., et al.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "51. Xin, J., Wang, N., Jiang, X., Li, J., Huang, H., Gao, X.: Binarized neural network for single image super resolution. In: ECCV (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "text", + "content": "52. Yang, L., Jin, Q.: Fracbits: Mixed precision quantization via fractional bit-widths. In: AAAI (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "text", + "content": "53. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "text", + "content": "54. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "55. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image superresolution. In: CVPR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "56. Zhang, Y., Wang, H., Qin, C., Fu, Y.: Learning efficient image super-resolution networks via structure-regularized pruning. In: ICLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "text", + "content": "57. Zhong, Y., Lin, M., Li, X., Li, K., Shen, Y., Chao, F., Wu, Y., Ji, R.: Dynamic dual trainable bounds for ultra-low precision super-resolution networks. In: ECCV (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "text", + "content": "58. Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., Zou, Y.: DoReFa-Net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "text", + "content": "59. Zhuang, B., Shen, C., Tan, M., Liu, L., Reid, I.: Towards effective low-bitwidth convolutional neural networks. In: CVPR (2018)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "ODM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_content_list.json b/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c6309babd32c11fc12a2ad935e8ab00341195c53 --- /dev/null +++ b/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_content_list.json @@ -0,0 +1,2062 @@ +[ + { + "type": "text", + "text": "PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments", + "text_level": 1, + "bbox": [ + 269, + 141, + 733, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang You $^{1\\star}$ Kai Xiong $^{2}$ Zhening Yang $^{3}$ Zhengxiang Huang $^{2}$", + "bbox": [ + 264, + 210, + 738, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junwei Zhou² Ruoxi Shi⁴ Zhou Fang²", + "bbox": [ + 354, + 227, + 647, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adam W. Harley $^{1}$ Leonidas Guibas $^{1}$ Cewu Lu $^{2\\star}$", + "bbox": [ + 305, + 243, + 694, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Stanford University", + "bbox": [ + 426, + 268, + 575, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shanghai Jiao Tong University", + "bbox": [ + 390, + 284, + 611, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Horizon Robotics Inc.", + "bbox": [ + 421, + 297, + 581, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "UC San Diego", + "bbox": [ + 444, + 311, + 557, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/709319356a84f83ce0ded0ee55b0d9d33032ba0e8c81825c61ca5ce0bb79cb41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 357, + 330, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/582e21ec12be91a48836679f23f4bc568f7da8b23de0c6cee9c97e0e29dabacc.jpg", + "image_caption": [ + "Fig. 1: We propose PACE: a large-scale object pose dataset, with diverse objects, complex scenes, and various types of occlusions, reflecting real-world challenges." + ], + "image_footnote": [], + "bbox": [ + 218, + 409, + 328, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3da2842c4be018ab84f33186bbe51cd77c96cf30cd79c2e846242e9e50ee0397.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 358, + 442, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad871cf0ebf1b0fa973cc0a89380acaa8bfcd8ab8309b7a747d4ba7d919677bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 410, + 442, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01c4e6e6a4833910bed75a900a139974904b68085bf98d60b01b67bcfce7ddb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 358, + 555, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d4c03167f2156c6ccf163d36d6d548c006270afaa1dbb0d799395da2d4ff3e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 410, + 555, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02b4d783904db49da959e93fac7a08b08792064619b79a2b98b334bb39c3e59c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 358, + 669, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b1bc824ed10b2d585c91f30ad951f2f616004511f789de50be521995e6c3610.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 410, + 669, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dec61119fa07d701745b1d54a945cc6bee9fea09d6d4550f6400b3d3dabea163.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 358, + 782, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e9e474a3d2d627184d9d2a02db6866a1c7b5e62fddcd350623d85924a20a37d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 410, + 782, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce PACE (Pose Annotations in Cluttered Environments), a large-scale benchmark designed to advance the development and evaluation of pose estimation methods in cluttered scenarios. PACE provides a large-scale real-world benchmark for both instance-level and category-level settings. The benchmark consists of 55K frames with 258K annotations across 300 videos, covering 238 objects from 43 categories and featuring a mix of rigid and articulated items in cluttered scenes. To annotate the real-world data efficiently, we develop an innovative annotation system with a calibrated 3-camera setup. Additionally, we offer PACE-Sim, which contains 100K photo-realistic simulated frames with 2.4M annotations across 931 objects. We test state-of-the-art algorithms in PACE along two tracks: pose estimation, and object pose tracking, revealing the benchmark's challenges and research opportunities. Our benchmark code and data is available on https://github.com/qq456cvb/PACE.", + "bbox": [ + 259, + 545, + 743, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 760, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of 3D object pose estimation is integral to a myriad of applications, particularly within robotic manipulation. Recent advancements in both instance and", + "bbox": [ + 212, + 787, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors. Email: yangyou@stanford.edu, lucewu@sjtu.edu.cn", + "bbox": [ + 217, + 825, + 746, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "category-level pose estimation have been bolstered by deep learning approaches, and perhaps more importantly, by data.", + "bbox": [ + 212, + 146, + 785, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PoseCNN [38] advanced pose estimation into the deep learning era, and simultaneously introduced the influential YCB-Video dataset. This benchmark has catalyzed methodological development and offered a consistent evaluation platform. Besides, the Benchmark for 6D Object Pose Estimation (BOP) challenges [14] have consolidated datasets metrics for instance-level pose estimation.", + "bbox": [ + 212, + 176, + 787, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In parallel, NOCS [34] has addressed category-level pose estimation, albeit with a smaller real-world dataset for validation and testing. Despite these strides, the field grapples with a fundamental challenge: state-of-the-art pose estimation models still perform poorly in real-world settings, and existing evaluation datasets are too constrained to thoroughly reveal this fact. The NOCS-REAL275 dataset, for instance, spans only six categories and includes a mere 18 videos. This limitation has led to an array of published models which perform near-perfectly on NOCS but do not generalize to new data, as demonstrated in Figure 2.", + "bbox": [ + 212, + 252, + 787, + 373 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c20085b68f4ec9a69863169cd08c64499d50d73275af2b321dfbd7855a3b3f31.jpg", + "image_caption": [ + "Fig. 2: While current state-of-the-art methods yield satisfactory outcomes on the NOCS-REAL275 dataset, their models' performance significantly deteriorates when transferred to previously unseen datasets such as PACE. Left: Qualitative visualizations of various models' pose predictions for a mug in REAL275 vs. a mug in PACE. Right: The performance of state-of-the-art methods markedly declines on PACE, even when evaluating on categories that exist in both datasets." + ], + "image_footnote": [], + "bbox": [ + 217, + 398, + 503, + 494 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/57796ce431ca4423fb9a27c602025c1c641f5b941092e428f786298afb86b7bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodAP@15°5cm (%)↑
bottlebowlcanmug
REAL275 [34]HS-Pose [43]99.899.899.586.2
DualPoseNet [22]97.599.797.768.7
SARNet [21]98.298.297.470.6
PACE (Ours)HS-Pose [43]0.40.00.87.5
DualPoseNet [22]0.32.00.51.6
SARNet [21]2.10.00.30.0
", + "bbox": [ + 524, + 397, + 784, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce PACE, a benchmark for pose estimation, and present a comprehensive study evaluating a wide range of pose estimation and tracking methods. Our contributions are threefold:", + "bbox": [ + 212, + 619, + 785, + 664 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The PACE dataset: This dataset includes 238 objects across 43 categories, captured in 300 video clips within diverse scenes. With an average of 183 frames per clip, the dataset consists of 55K frames and 258K annotations, providing a large-scale benchmark for pose estimation. We also provide PACESim, a photo-realistic training dataset which contains 100K frames with 2.4M annotations across 931 objects.", + "- Our evaluation study: To the best of our knowledge, we are the first to analyze and report the performance of state-of-the-art (SOTA) pose estimation methods in a large-scale cluttered setting. These results provide valuable insights on the scalability and generalizability of SOTA methods, making it clear that they are far from reliable in general." + ], + "bbox": [ + 223, + 674, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our annotation pipeline: We will open-source our annotation pipeline, which uses a calibrated 3-camera system, enhancing the precision and scalability of annotating poses in real data. This tool significantly mitigates human error and reduces the effort required for annotating 3D poses, providing a solution to one of the major bottlenecks in pose dataset creation.", + "bbox": [ + 223, + 146, + 787, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our work aims to support the development of more robust and generalizable pose estimation techniques, thereby facilitating progress towards successful pose estimation in the real world.", + "bbox": [ + 212, + 233, + 785, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 301, + 395, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The field of 3D object pose estimation has seen substantial progress over the past few years. This progress has been facilitated by the introduction of standardized datasets and the development of innovative algorithms.", + "bbox": [ + 212, + 332, + 785, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Object Pose Datasets", + "text_level": 1, + "bbox": [ + 215, + 398, + 439, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance-Level Pose Datasets YCB-Video dataset [38] is a comprehensive resource for 6D object pose estimation, containing a large number of video frames with accurate pose annotations for 21 objects. LINEMOD-Occluded [3] offers a challenging setting for pose estimation with piled multiple objects in occluded scenes. NAVI dataset [15] presents casually captured images of objects with high-quality 3D scans and precise 2D-3D alignments for 3D reconstruction tasks. Recently, StereoObj-1M [24] is proposed to address challenging cases such as object transparency, in addition to common challenges of occlusion and symmetry.", + "bbox": [ + 212, + 424, + 787, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Category-Level Pose Datasets NOCS-REAL275 [34] and Wild6D [40] both annotate a limited number of categories with upright poses. Objectron [1] is a collection of short, single-object-centric video clips with no clutters. The Scan2CAD [2] dataset aligns 14225 CAD models from ShapeNet to 1506 ScanNet scans, promoting CAD model alignment in RGB-D scans. Pix3D [29] is a benchmark with image-shape pairs and pixel-level 2D-3D alignment, aiding in shape reconstruction and retrieval. HANDAL [11] focuses on pose estimation and affordance prediction for robotics-ready manipulable objects. HouseCat6D [17] encompasses 194 diverse objects across 10 household categories with additional grasp annotations. ROPE [41] provides a large-scale evaluation benchmark covering a wide spectrum of categories. However, it does not include challenging scenarios such as moving and piled objects. Table 1 presents a comparative comparison of these datasets, alongside our new dataset PACE.", + "bbox": [ + 212, + 566, + 787, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Pose Estimation Methods", + "text_level": 1, + "bbox": [ + 215, + 785, + 473, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance-level Pose Estimation PPF (Point Pair Features) [8] is a pre-deep learning standard for instance-level pose estimation, using local geometric features", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 379, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/7d253db0637a9e39f0f65d0929326c3f8e7c76710b689df2c5070459d7137d96.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Comparison of object pose datasets. From left to right, the table captures input modality, number of categories, number of instances, number of videos, number of images, number of total annotations, whether 3D CAD models are provided, whether videos include still and/or moving objects, whether objects are occluded in some frames, whether images contain artificial markers, whether poses for each part of articulated objects are provided, and whether multiple objects are piled in some frames. Compared with most other datasets, PACE contains moving and articulated objects." + ], + "table_body": "
ModalityCat.Obj.Vid.Img.Anno.CADMov.Occ.Marker-freeArtic.Piled
YCB-Video [38]RGBD-211220K99K✓X✓✓X✓
LINEMOD-O [3]RGBD-811.2K9.2K✓X✓XX✓
NAVI [15]RGBD-3632410K10k✓XX✓XX
StereoObj-1M [24]RGBD-18182393K1.5M✓X✓XX✓
NOCS-REAL275 [34]RGBD642188K✓X✓XXX
Wild6D [40]RGBD5172251661.1M1.1MXXX✓XX
Objectron [1]RGB917k14k4M4MXXX✓XX
Scan2CAD [2]RGBD93K1506-14K✓X✓✓XX
Pix3D [29]RGBD9395-10K10K✓XX✓XX
HANDAL [11]RGB172122K308K308K✓✓✓✓XX
HouseCat6D [17]RGBD101924124K160K✓X✓✓X✓
ROPE [41]RGBD149581363332K1.5M✓X✓✓XX
PACE (Ours)RGBD4323830055K258K✓✓✓✓✓✓
", + "bbox": [ + 218, + 143, + 787, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from point clouds. PoseCNN [38] is the first model trained end-to-end for the task. CosyPose [19] integrates a global refinement strategy in its end-to-end pipeline. SurfEmb [12] leverages surface embeddings for correspondence matching, and GDRNPP [33] uses geometry-guided regression.", + "bbox": [ + 212, + 460, + 789, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category-level Pose Estimation Category-level pose estimation extends the challenge to generic object categories. NOCS [34] introduces a unified coordinate space for all objects, predicting object NOCS maps from RGB images. SGPA [4] aims to adapt the structure-guided prior in the pose estimation process, while SAR-Net [21] uses shape alignment and symmetric correspondence to estimate a coarse 3D object shape and facilitate object center and size estimation. Recently, HS-Pose [43] proposes a network structure with a HS-layer that extends 3D graph convolution to extract hybrid scope latent features from point clouds for category-level object pose estimation. GenPose [42] incorporates probabilistic stable diffusion models into pose estimation, achieving strong performance.", + "bbox": [ + 212, + 541, + 789, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Pose Tracking Methods", + "text_level": 1, + "bbox": [ + 215, + 712, + 457, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-level Pose Tracking Methods like RBOT [30] use RGB data and 3D models to track multiple objects, employing color histograms in their cost function. PoseRBPF [5] separates rotation and translation, using an autoencoder for rotation feature embeddings. ICG [28] iteratively refines pose using geometric cues and is effective for textureless objects, with extensions incorporating visual data [27]. The first deep learning tracker, D6DT [9], and se(3)-TrackNet [36] predict frame-to-frame relative poses, using a render-and-compare strategy.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category-level Pose Tracking 6-PACK [31] performs category-level tracking, using DenseFusion [32] features and an attention mechanism for unsupervised keypoint ordering and interframe motion via keypoint matching. BundleTrack [35] performs pose tracking without relying on 3D models, instead using video segmentation and pose graph optimization. CAPTRA [37] tracks 9DoF poses for rigid and articulated objects, with heads for rotation regression and normalized coordinate prediction, from which 3D size and translation can be computed.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Construction of PACE", + "text_level": 1, + "bbox": [ + 215, + 282, + 472, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key contribution of this work is the establishment of a scalable and reliable annotation framework, enabling the collection of large-scale and accurate pose annotations. An overview of the pipeline is depicted in Figure 3.", + "bbox": [ + 212, + 321, + 787, + 367 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d213f454beaf986d901978d1b98d1f1fd795646c7b6dbfdb5d3b5b5e3e7620f.jpg", + "image_caption": [ + "Fig. 3: Overview of the PACE annotation pipeline." + ], + "image_footnote": [], + "bbox": [ + 246, + 393, + 426, + 570 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a451200f8d3953ddf29530ce271117558c74735c7d539ec6663f1d06d5bfb53a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 393, + 756, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Acquisition of 3D Scans of Common Objects", + "text_level": 1, + "bbox": [ + 214, + 656, + 630, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We began by digitizing an extensive collection of commonplace objects. These items were categorized into 43 distinct categories, as represented in Figure 5. The Einscan Pro 2X was utilized for rapid scanning of all objects, typically completing within 5 to 10 minutes per object. To expedite this process, we employed a rotatable platform to acquire multiple viewpoints of objects. After scanning, objects were manually aligned to a standard pose within a uniform coordinate system. For each object, we centered its axis-aligned bounding box at the origin and then aligned the bounding box orientation along a common axis within the category. We additionally annotated rotational symmetries. High-resolution meshes were simplified to lower-resolution ones for a smoother annotation workflow.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 379, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Articulated Objects Diverging from many prior pose estimation datasets, our collection encompasses a very wide set of objects, including articulated objects from the AKB48 [23] dataset, namely: scissors, cutters, clips, and boxes. We adopt the alignment methodology from the original AKB48 dataset, without modification. These objects are segmented into multiple parts with hierarchical relationships, presenting a difficult challenge for pose estimation.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 RGB-D Sequence Acquisition", + "text_level": 1, + "bbox": [ + 214, + 262, + 504, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implemented a 3-camera system to aid in data acquisition and annotation, comprising three Intel Realsense D415 RGB-D cameras affixed to a metal framework, as illustrated in the bottom of Figure 3. The advantages of this setup include: (1) tripling the data yield; (2) reducing ambiguity in pose annotation, especially regarding translation along the depth dimension, by enforcing annotation consistency across all 3 views; (3) enhancing tracking accuracy of still objects with ArUco markers [10], making PnP more stable.", + "bbox": [ + 212, + 290, + 789, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Calibration of Multi-Camera Extrinsic Parameters We calibrated this multicamera system through a semi-automatic process. ArUco markers [10] alone proved insufficient for high-accuracy rotation estimation. Hence, we resorted to trifocal tensor estimation, i.e. TFT [16]. The process begins with feature extraction and matching, followed by bundle adjustment to refine the positions of the 3D landmarks and camera poses. For reliable feature matching, we employed the SuperPoint [7] descriptor and SuperGlue [26] matcher, using a stringent threshold for matching. We observed that rotational component of the resulting extrinsic parameters is precise, but the translation aspect suffers from scale ambiguity inherent in Structure-from-Motion approaches. We corrected this by calibrating the scale against markers, applying the following formula to obtain the optimal scale factor, by setting gradient of $\\| s\\hat{\\mathbf{t}} - \\mathbf{t}'\\|_2^2$ to zero:", + "bbox": [ + 212, + 412, + 789, + 595 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {1 \\rightarrow 2} = \\frac {\\hat {\\mathbf {t}} _ {1 \\rightarrow 2} \\cdot \\mathbf {t} _ {1 \\rightarrow 2} ^ {\\prime}}{\\hat {\\mathbf {t}} _ {1 \\rightarrow 2} \\cdot \\hat {\\mathbf {t}} _ {1 \\rightarrow 2}}, s _ {1 \\rightarrow 3} = \\frac {\\hat {\\mathbf {t}} _ {1 \\rightarrow 3} \\cdot \\mathbf {t} _ {1 \\rightarrow 3} ^ {\\prime}}{\\hat {\\mathbf {t}} _ {1 \\rightarrow 3} \\cdot \\hat {\\mathbf {t}} _ {1 \\rightarrow 3}},\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 608, + 651, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{t}}_{i\\rightarrow j}$ is the TFT predicted translation (up to a scale) from camera $i$ to camera $j$ , $\\mathbf{t}_{i\\rightarrow j}^{\\prime}$ is the marker-calibrated translation in real metric scale. We set the extrinsics of the first camera to be the identity matrix. We use Intel's dynamic calibration tool to calibrate intrinsics. In cases of repetitive textures, where SuperPoint+SuperGlue were unreliable, we made correspondences manually.", + "bbox": [ + 212, + 650, + 787, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Annotation of Pose Ground-Truths", + "text_level": 1, + "bbox": [ + 214, + 751, + 552, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous works have used ArUco markers to automate pose estimation through Perspective-n-Points, and we do this also. Specifically, we used ArUco markers to automate the annotation of still object poses (i.e. not moving), following LINEMOD [3] and NOCS [34]. We note that this step has two limitations:", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Markers in the scene detract from realism and compromise dataset integrity: training on marker-augmented imagery may result in overfitting to these artificial patterns.", + "2. Marker-based annotations are inapplicable to moving objects (since the markers are typically placed on the background), thus limiting the setup to annotating still objects." + ], + "bbox": [ + 222, + 146, + 785, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compensate for both of these issues: we remove marker appearances from the dataset (after using them for annotation), and we introduce additional tools to annotate the moving objects. We describe these compensation steps next.", + "bbox": [ + 212, + 250, + 782, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Marker removal We removed the marker appearances from the dataset, using a marker inpainting strategy, detailed as follows. In step 1, we place a marker (Marker 1) somewhere within the camera's field of view. We then record a short video with this marker in view. In step 2, we place a second marker (Marker 2) at a chosen distance from the first, and record another video. In step 3, we remove Marker 1, and begin the actual object capture process (with only Marker 2 present). After this process, we end up with: (1) frames with Marker 1 only, which clearly depict the surface where Marker 2 will later appear; (2) frames capturing both markers, providing helpful calibration cues; (3) frames with Marker 2 only, which represent our main capture. We leverage Marker 2 for automated pose tracking, and manually correct the tracking every 40 frames in case of drift. We use the frames from the first two steps to seamlessly inpaint [25] Marker 2's area for the final dataset, as depicted in Figure 4.", + "bbox": [ + 212, + 310, + 787, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f34949e0cd9adf9c22dc9d477a3416bbc3c61039963a8655e9afb27a0d19f5a4.jpg", + "image_caption": [ + "Fig. 4: Illustration of the marker inpainting process." + ], + "image_footnote": [], + "bbox": [ + 218, + 532, + 784, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotation of Moving Object Positions For moving objects, the traditional marker-based tracking approach is insufficient, as the scene markers do not move in tandem with the objects. In such instances, we harnessed the capabilities of BundleTrack [35], an advanced RGB image-based tracking algorithm. BundleTrack conducts feature correspondence analysis between successive frames to estimate poses, complemented by a bundle adjustment algorithm to optimize keyframes globally and minimize tracking errors. Despite BundleTrack's proficiency in approximating poses, it is prone to drift, necessitating the manual adjustment of poses every ten frames to ensure precision. We adopted a rigorous manual checking process, examining every tenth frame for 3D pointcloud alignment and 2D multi-view agreement. We detail the annotation tool and its", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 379, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "graphical interface in the supplementary. This delicate process represents the most labor-intensive aspect of our annotation pipeline.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to verify that our annotation process gives accurate poses, we conducted an experiment where we used our setup to manually annotate the poses of Scene 1 from NOCS-REAL275 [34] dataset, and compare the annotations to the provided ground truths. We recorded an average pose annotation error of $0.9^{\\circ}$ for rotation and $2.3\\mathrm{mm}$ for translation.", + "bbox": [ + 215, + 176, + 785, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generation of Segmentation Masks Upon successful pose annotation, we generated occlusion-aware segmentation masks, via depth rendering. In cases where hand interactions are involved, we employed the segmentation model SAM [18] to estimate hand masks, and subsequently subtracted these from the previously computed object masks to obtain the occlusion-aware segmentation.", + "bbox": [ + 215, + 266, + 785, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Dataset Statistics", + "text_level": 1, + "bbox": [ + 215, + 364, + 423, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In pursuit of diversity, we placed the objects in ten disparate environments, each featuring varying levels and configurations of occlusion. We captured 10 video sequences per scene, each sequence including 1 to 7 objects. We captured RGB and depth data using an Intel RealSense D415 camera, with a resolution of $1280 \\times 720$ . The capture process spans distances ranging from $0.5\\mathrm{m}$ to $1.5\\mathrm{m}$ .", + "bbox": [ + 215, + 396, + 785, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Annotation Distribution and Object Diversity", + "text_level": 1, + "bbox": [ + 215, + 494, + 640, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The comprehensive distribution of pose annotation counts is depicted in Figure 5, showcasing a diverse array including both rigid and articulated objects. As demonstrated in Figure 7, our object database encompasses a broad spectrum of object sizes, with the majority measuring near $0.2\\mathrm{m}$ along the bounding box diagonal. It also includes larger items such as storage bins, adding to the heterogeneity of the dataset.", + "bbox": [ + 215, + 518, + 785, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Variability in Occlusion, Pose, and Environmental Context", + "text_level": 1, + "bbox": [ + 215, + 631, + 743, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The dataset has rich variability in occlusion levels and object poses, as illustrated in Figure 6, which outlines the statistical distribution of both visibility ratios and azimuth/elevation angles, indicating a comprehensive coverage reflecting real-world scenarios. These statistics can be helpful for assessing the robustness of pose estimation algorithms. Besides, the ten distinct capturing environments are shown in Figure 8, illustrating diverse backgrounds and lighting conditions.", + "bbox": [ + 215, + 656, + 785, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Simulated Training Data and Dataset Split", + "text_level": 1, + "bbox": [ + 215, + 768, + 614, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To facilitate a comprehensive and equitable evaluation of pose estimation methodologies, we partitioned our entire real-world dataset into validation and test subsets following a $20/80$ ratio. The test subset is used for both instance-level and", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/032379076a8233035dd5607347c922c1ad5bc5145350c1aa942ed736f32959f1.jpg", + "image_caption": [ + "Fig. 5: Distribution of pose annotation counts across different object categories." + ], + "image_footnote": [], + "bbox": [ + 250, + 148, + 751, + 335 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "category-level pose estimation evaluation tasks. The real-world validation subset is used for instance-level pose estimation only. We render another photo-realistic validation subset for category-level pose estimation, in order to cover enough category-level objects that are different from the test set.", + "bbox": [ + 217, + 398, + 784, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Additionally, to support research necessitating extensive training datasets, we generated a large synthetic dataset called PACE-Sim that contains two different training subsets, for instance-level and category-level pose estimation, respectively. We use a physically based renderer [6] to ensure the images are highly photo-realistic. The synthetic training sets consist of 100K images, each featuring dozens of objects, yielding 2.4M annotations. The distribution of object instances in both PACE and PACE-Sim can be found in supplementary.", + "bbox": [ + 217, + 459, + 785, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Evaluation Benchmarks", + "text_level": 1, + "bbox": [ + 217, + 589, + 478, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We assess both pose estimation and tracking. State-of-the-art (SOTA) pose estimation techniques generally involve two phases: first detecting or segmenting the object, and then estimating its pose within the identified mask. For a fair comparison, we evaluate pose estimation assuming perfect object detection. Detailed SOTA detection performance is available in the supplementary materials but is beyond this dataset's primary scope.", + "bbox": [ + 217, + 622, + 785, + 713 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For pose tracking, we provide ground-truth data in the initial frame and require methods to accurately track the object's rotations and translations through the video sequence. Both model-free and model-based trackers are assessed.", + "bbox": [ + 217, + 713, + 784, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Pose Estimation Benchmark", + "text_level": 1, + "bbox": [ + 217, + 782, + 493, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This benchmark is divided into instance-level and category-level pose estimation. The former concerns object instances seen during training (in different scenes),", + "bbox": [ + 217, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 380, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5f77a48e05238c0c1bd36df306fae288707a230de6618eddf3046e79edad0f55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 143, + 519, + 234 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f6a81fe12a22692ed36cb230cefc7be97e55e0ef936f4248533b98d639d44f96.jpg", + "image_caption": [ + "Fig. 6: Distribution of occlusions, azimuth and elevation viewing angles." + ], + "image_footnote": [], + "bbox": [ + 222, + 239, + 398, + 339 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/128b405aacc842d5b5dacb25428b3949349e0dadcdce381532c2f0b8eeaa7a04.jpg", + "image_caption": [ + "Fig. 7: Distribution of object sizes." + ], + "image_footnote": [], + "bbox": [ + 400, + 244, + 550, + 339 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e68397aeb3544284f961a7e34c973659c7dd2e127272bbf378c282a35ac53607.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 181, + 784, + 316 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/77172050f348784b0bccf84914bccbe018dcfb8fdf6dafbecd53b082d8355988.jpg", + "image_caption": [ + "(a) basket" + ], + "image_footnote": [], + "bbox": [ + 217, + 395, + 326, + 444 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6ab287369884ee13589e1c34d5694c61f4f0e1c8f2241ca629ff07c29da2fea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 396, + 441, + 444 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/db28857fbb2811f8ace59a39ac602e70164ac4d36c63ee3c0fd676e8208b1141.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 395, + 555, + 444 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3989cae8268163364a378d396a4b1bacdbf1b2d45547e274267ed89dfc434967.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 395, + 669, + 444 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/69b6ae908aa0d229ee469777bf56494c15eda96ed4b5a8987c52d352dbc25102.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 395, + 785, + 444 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7577a49491983fb703e7a9633429086acfdacba2a52f38814b53a710d4b3c056.jpg", + "image_caption": [ + "(f) china floor" + ], + "image_footnote": [], + "bbox": [ + 217, + 458, + 326, + 506 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ad7a091475671192b9d3591a1dea85bd027f7bde6db1b590ae6a829380f4fda4.jpg", + "image_caption": [ + "(b) box top", + "(g) desk1", + "Fig. 8: Sample images from the 10 environments." + ], + "image_footnote": [], + "bbox": [ + 331, + 458, + 441, + 506 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d255b6ee5f955337beeb5175ced43327c2ef62e8bc01c1cdb9cf31a5fa326519.jpg", + "image_caption": [ + "(c) cabinet", + "(h) desk2" + ], + "image_footnote": [], + "bbox": [ + 446, + 458, + 555, + 506 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d461684a4c36d427002b7ec2c948581676618f306e7c186b5755ba770c98fdfd.jpg", + "image_caption": [ + "(d) carpet", + "(i) stairs" + ], + "image_footnote": [], + "bbox": [ + 558, + 458, + 669, + 506 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/414f78905a4253b1f53d17be0e720ea3fdd274f2bb2c0161c685f34ee5b4772b.jpg", + "image_caption": [ + "(e) chair", + "(j) wooden floor" + ], + "image_footnote": [], + "bbox": [ + 674, + 458, + 785, + 506 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "while the latter concerns instances which were not seen but come from the same category. We will also give an in-depth analysis of SOTA's failure modes.", + "bbox": [ + 212, + 577, + 787, + 607 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Instance-Level Pose Estimation This task demands the prediction of rotation and translation for known instances from the training set.", + "bbox": [ + 212, + 633, + 784, + 666 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics: Adhering to the BOP challenge protocol [14], we use Average Recall (AR) of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) as our metrics. Detailed computation of these metrics is described in [14]. Results are averaged across all instances.", + "bbox": [ + 212, + 684, + 787, + 760 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines: We assessed four baselines: PPF [8], CosyPose [19], SurfEmb [12], and GDRNPP [33]. While PPF does not require training data, the others are SOTA methods dependent on additional training on the PBR dataset. To focus on pose estimation performance, we assume that ground-truth detections are available.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Result Analysis: The quantitative analysis is presented in Table 2. Notably, while SOTA methods excel on the BOP benchmarks, they do not perform as well as PPF, which relies on local geometric feature matching in point clouds. The relative success of PPF highlights the potential for improvement in SOTA methods, particularly regarding sim-to-real ability and scalability to handle a large set of instances. Qualitative results can be found in the supplementary.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/00ae53e713ed9349b2d0c00d1fe8457e9e074c3bc848534e5677a561e02fabfe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modality DetectionARVSD↑ARMSSD↑ARMSPD↑AR↑
PPF [8]DG.T.53.448.155.252.2
CosyPose [19]RGBG.T.1.40.311.54.4
SurfEmb [12]RGBG.T.6.23.017.89.0
GDRNPP [33]RGB-DG.T.3.62.115.47.0
", + "bbox": [ + 305, + 262, + 697, + 333 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e2bddca684d5d3684e08c686bb19a5fa15d2376c07705db3e4bc2f649f351a3a.jpg", + "table_caption": [ + "Table 2: Instance-level pose estimation results showcasing the robustness of PPF and the potential for improvement in SOTA deep learning-based methods." + ], + "table_footnote": [], + "table_body": "
DetectionIoU25↑IoU50↑AP
0:20°↑0:60°↑0:5cm↑0:15cm↑0:20°↑0:60°↑
NOCS [34]Mask-RCNN0.2/0.00.0/0.01.2/0.04.3/0.017.0/0.033.5/0.01.1/0.04.2/0.0
HS-Pose [43]G.T.36.6/0.02.7/0.05.3/0.08.6/0.348.7/34.181.0/75.63.9/0.08.1/0.3
SGPA [4]G.T.2.6/0.01.2/0.05.6/1.111.4/7.316.1/6.850.7/17.53.3/0.910.1/7.2
DualPoseNet [22]G.T.24.2/0.30.1/0.05.5/0.08.1/0.018.6/27.363.4/65.21.8/0.06.5/0.0
SAR-Net [21]G.T.22.8/0.10.2/0.05.3/0.08.8/0.848.6/38.579.8/69.73.7/0.08.2/0.8
CPPF++ [39]G.T.44.5/3.64.4/0.015.2/1.727.3/6.235.3/31.174.0/66.79.9/1.124.9/5.9
ANCSH [20]G.T.- / 0.0- / 0.0- / 0.0- / 0.2- / 18.6- / 50.4- / 0.0- / 0.2
", + "bbox": [ + 220, + 441, + 782, + 568 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Category-level pose estimation benchmark combining the performance metrics for both rigid and articulated object pose estimation, separated by slash.", + "bbox": [ + 212, + 578, + 787, + 608 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Category-Level Pose Estimation Category-level pose estimation tasks involve predicting the 3D bounding box dimensions, rotation, and translation of target instances, given only the category information a priori.", + "bbox": [ + 212, + 672, + 787, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics: We adopted from NOCS [34], calculating mean Average Precision (AP) across predefined angle and translation thresholds. Specifically, AP@0:20° represents AP averaged from $0^{\\circ}$ to $20^{\\circ}$ at $1^{\\circ}$ intervals; AP@0:5cm is averaged from 0cm to 5cm at 0.25cm intervals; AP@0:20°,0:5cm combines these angle and translation thresholds. IoU $_{25}$ and IoU $_{50}$ denote AP for 3D bounding box matches at IoU thresholds of 25 and 50, respectively. Results are separately reported for both rigid and articulated objects.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 379, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines: We evaluated seven recent category-level pose estimation methods: NOCS [34], HS-Pose [43], SGPA [4], DualPoseNet [22], SAR-Net [21], CPPF++ [39] and ANCSH [20], the latter specifically designed for articulated objects. For adapting rigid-object-focused methods to articulated objects, each movable part is considered a distinct category. Ground-truth detections are presumed for each method, except for NOCS, which outputs both instance masks and pose estimations using a unified network.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Result Analysis: Quantitative comparisons are listed in Table 3. On rigid objects, HS-Pose and SAR-Net show fair performance on AP metrics for translation, but falls short on rotation AP metrics compared to CPPF++, suggesting a proficiency in size and translation prediction but not rotation. CPPF++ shows the best performance but is weaker on translation estimation. For articulated objects, all methods face challenges due to the movable parts, with ANCSH underperforming on the large-scale real-world dataset, indicative of a significant sim-to-real gap. NOCS shows instability, likely due to its inaccurate mask prediction. Qualitative results are available in supplementary.", + "bbox": [ + 212, + 268, + 789, + 405 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Failure Mode Analysis This section delves into the analysis of two critical questions that underline the failure of SOTA deep learning-based methods in pose estimation: 1) Is the failure attributed to the sim-to-real gap, considering the exclusive use of synthetic RGB-D images for training? 2) Does the failure stem from the inherent large-scale complexity of the PACE dataset, which poses a significant challenge for baseline models?", + "bbox": [ + 212, + 429, + 787, + 521 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Setup: To explore these questions, we introduced a new $80/10/10$ train/validation/test split for the real-world data, diverging from the $0/20/80$ split described in Section 4.3. We then examined the following settings:", + "bbox": [ + 212, + 537, + 787, + 583 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Real2Real Single Instance/Category Fitting: This basic setup involves training on a single instance or category selected (we randomly select the category can and sample a can with object ID 57) from the $80 / 10 / 10$ split of real data.", + "- Sim2Real Single Instance/Category Fitting: This scenario differs from the first by utilizing all simulated data for training and validation, with evaluation on the same real-world test split.", + "- Real2Real All Instance/Category Fitting: Unlike the first setting, this involves training, validation, and testing across all instances or categories." + ], + "bbox": [ + 223, + 595, + 785, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Result Analysis and Discussions: Our analysis in Table 4 reveals that baseline models effectively handle single instance or category fitting under the Real2Real Single Instance/Category Fitting setting, provided the training and test distributions align. However, the sim-to-real gap, especially noticeable in depth-input methods like DualPoseNet, HS-Pose, and SGPA, highlights significant challenges, as these methods struggle with the depth noise in real world.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c0acd45070132becb0738d5a8307486c159f0733e11dfbb55eb26c18f292d32c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training SettingInstance-Level (AR%)†Category-Level (mean AP%@0:60°, 0:15cm)†
CosyPoseSurfEmbGDRNPPNOCSHS-PoseSGPADualPoseNetSAR-NetCPPF++
Real2Real Single Inst./Cat.76.486.788.276.981.286.758.659.383.4
Sim2Real Single Inst./Cat.67.1 (-9.3↓)73.7 (-13.0↓)78.1 (-10.1↓)55.8 (-21.1↓)1.1 (-80.1↓)0.3 (-86.4↓)1.4 (-57.2↓)33.2 (-26.1↓)55.8 (-27.6↓)
Real2Real All Inst./Cat.8.2 (-68.2↓)11.2 (-75.5↓)9.7 (-78.5↓)0.3 (-76.6↓)5.6 (-75.6↓)9.8 (-76.9↓)1.2 (-57.4↓)2.1 (-57.2↓)13.7 (-69.7↓)
", + "bbox": [ + 217, + 143, + 782, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Additionally, learning across all instances/categories leads to a notable performance drop, as shown in Table 4. This drop, together with suboptimal training data performance and loss plateauing, suggests underfitting and highlights the complexity posed by the PACE dataset's diversity. Innovative approaches or larger foundational models might be necessary to address these challenges and effectively bridge the sim-to-real gap.", + "bbox": [ + 212, + 310, + 787, + 401 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Pose Tracking Benchmark", + "text_level": 1, + "bbox": [ + 214, + 420, + 477, + 435 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We distinguish SOTA pose tracking techniques into model-based, utilizing a 3D CAD model, and model-free, which only require an initial pose from first frame.", + "bbox": [ + 212, + 443, + 787, + 473 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Metrics: For model-based tracking, we measure the area under curve (AUC) for ADD, ADD-S [38], and ADD(-S) metrics, with a maximum threshold of $0.1\\mathrm{m}$ [32]. The ADD metric, introduced in [13], computes the average per-point distance between point clouds transformed by the predicted and ground-truth poses. ADD-S accounts for point correspondence ambiguities in symmetric objects.", + "bbox": [ + 212, + 484, + 787, + 560 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For model-free tracking, we adopt four metrics from prior studies [31]: 1) $5^{\\circ}5\\mathrm{cm}$ , indicating the percentage of predictions within $5^{\\circ}$ rotation and $5\\mathrm{cm}$ translation errors. 2) IoU25, the percentage of instances where the intersection over union of the 3D bounding boxes exceeds $25\\%$ . 3) $\\mathbf{R}_{\\mathrm{err}}$ and 4) $\\mathbf{T}_{\\mathrm{err}}$ , the mean rotational and translational errors in degrees and centimeters, respectively, only considering instances with IoU $>25\\%$ . We exclude objects with visibility under $10\\%$ as per the BOP convention [14]. Results are averaged separately across rigid and articulated object categories.", + "bbox": [ + 212, + 561, + 787, + 681 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/619fff9e33bbff2cefe1fe5ad704fe19c3a9cadaa27e90b71f69168b0ed8596f.jpg", + "table_caption": [ + "Table 4: Ablation Study: This study illustrates the challenges models face when transitioning from real-to-real settings. It showcases performance disparities when scaling from single instances/categories to the entire dataset, emphasizing the scalability challenge from PACE and the profound impact of sim-to-real discrepancies on depth-dependent methodologies." + ], + "table_footnote": [], + "table_body": "
ModalityADD↑ADD-S↑ADD(-S)↑
RBOT [30]RGB7.1/0.510.3/0.87.4/0.5
ICG [28]RGB-D35.6/10.148.1/15.238.1/10.1
", + "bbox": [ + 341, + 705, + 663, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Model-Based Pose Tracking. We report the area under curve (AUC) with respect to ADD, ADD-S and ADD(-S). The higher value, the better performance.", + "bbox": [ + 212, + 767, + 787, + 796 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 379, + 114, + 732, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3b6be880248e3fd23dcb4794656c55cbc8c3c7ae7c4fdba3702bea810fbecc41.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training-Free Modality5°5cm↑IoU25↑Rerr↓Terr↓
BundleTrack [35]✓RGB6.4/11.29.1/14.13.2/5.52.6/0.8
CAPTRA [37]✗D12.9/4.445.8/20.619.2/40.92.2/1.5
6-PACK [31]✗RGB-D9.2/3.923.1/16.717.7/33.62.1/1.2
", + "bbox": [ + 277, + 143, + 727, + 204 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: Model-Free Pose Tracking. Both rigid and articulated results are reported. For $5^{\\circ}5\\mathrm{cm}$ and IoU25, the higher value means better performance while for $\\mathbf{R}_{err}$ and $\\mathrm{T_{err}}$ , the situation is reversed. Note that $\\mathrm{R}_{err}$ and $\\mathrm{T_{err}}$ are respect to IoU25 since objects with IoU $\\leq 25\\%$ are not counted.", + "bbox": [ + 212, + 215, + 787, + 272 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Baselines: For tracking methods using object masks, we applied the ground-truth masks directly. Except for CAPTRA, which also predicts the 3D bounding box size, all methods estimate only rotation and translation; however, we used ground-truth bounding box sizes for consistency in evaluations. Unlike other methods that treat each part of an articulated object independently, CAPTRA considers the entire object as one entity, trained on our synthetic PBR data.", + "bbox": [ + 212, + 311, + 787, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Result and Failure Mode Analysis: Tables 5 and 6 display the performance of state-of-the-art (SOTA) tracking methods on the PACE dataset, highlighting both rigid and articulated object results. The highest achieved ADD(-S) for model-based methods is only $38.1\\%$ for rigid and $10.1\\%$ for articulated objects by ICG. Furthermore, Table 6 reveals severe limitations of model-free methods, with the best $5^{\\circ}5\\mathrm{cm}$ accuracy under $13\\%$ , posing major tracking challenges.", + "bbox": [ + 212, + 414, + 787, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The convention followed by 6-PACK [31] and BundleTrack [35] disregards objects with $\\mathrm{IoU} \\leq 25\\%$ when computing rotational and translational errors. However, this approach could overestimate the performance on $\\mathrm{R}_{err}$ and $\\mathrm{T}_{err}$ . BundleTrack [35] reports lower performance in $5^{\\circ}5\\mathrm{cm}$ and IoU25 metrics yet shows seemingly better results for $\\mathrm{R}_{err}$ and $\\mathrm{T}_{err}$ , suggesting that it can only track well within a threshold but cannot recover once the object is lost. Qualitative results can be found in the supplementary.", + "bbox": [ + 212, + 505, + 789, + 611 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 214, + 631, + 529, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present PACE, a new large-scale benchmark for 3D object pose estimation and tracking, designed with various occlusion levels across ten diverse environments. This benchmark both reflects prior progress and urges further research to tackle real-world complexities. Our analysis indicates significant advances in pose estimation, yet underscores a notable real-world performance gap, particularly with articulated objects and in terms of robustness and scalability. Future efforts could focus on advanced sim-to-real techniques and developing larger models that capture a broader spectrum of object features and environments.", + "bbox": [ + 212, + 661, + 787, + 782 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations Due to resource limitations, PACE lacks large and valuable objects like tables, cameras, and laptops, which may affect its breadth in representing real-world object scenarios.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 434, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Project of China (No. 2022ZD0160102), National Key Research and Development Project of China (No. 2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants. Yang You is also supported in part by the Outstanding Doctoral Graduates Development Scholarship of Shanghai Jiao Tong University.", + "bbox": [ + 212, + 175, + 787, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 273, + 323, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7822-7831 (2021)", + "2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2614-2623 (2019)", + "3. Brachmann, E., Krull, A., Michel, F., Gumhold, S., Shotton, J., Rother, C.: Learning 6d object pose estimation using 3d object coordinates. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 536-551. Springer (2014)", + "4. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021)", + "5. Deng, X., Mousavian, A., Xiang, Y., Xia, F., Bretl, T., Fox, D.: Poserbpf: A raobackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics 37(5), 1328-1342 (2021)", + "6. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901", + "7. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 224-236 (2018)", + "8. Drost, B., Ulrich, M., Navab, N., Ilic, S.: Model globally, match locally: Efficient and robust 3d object recognition. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 998-1005. IEEE (2010)", + "9. Garon, M., Lalonde, J.F.: Deep 6-dof tracking. IEEE transactions on visualization and computer graphics 23(11), 2410-2418 (2017)", + "0. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition 47(6), 2280-2292 (2014)", + "1. Guo, A., Wen, B., Yuan, J., OTHERS: Handal: A dataset of real-world manipulable object categories with pose annotations, affordances, and reconstructions (2023), https://arxiv.org/pdf/2308.01477.pdf", + "2. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022)" + ], + "bbox": [ + 225, + 301, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 379, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. pp. 548-562. Springer (2013)", + "14. Hodaţ, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 577-594. Springer (2020)", + "15. Jampani, V., Maninis, K.K., Engelhardt, A., Karpur, A., Truong, K., Sargent, K., Popov, S., Araujo, A., Martin-Brualla, R., Patel, K., et al.: Navi: Category-agnostic image collections with high-quality 3d shape and pose annotations. arXiv preprint arXiv:2306.09109 (2023)", + "16. Julià, L.F., Monasse, P.: A critical review of the trifocal tensor estimation. In: Image and Video Technology: 8th Pacific-Rim Symposium, PSIVT 2017, Wuhan, China, November 20-24, 2017, Revised Selected Papers 8. pp. 337-349. Springer (2018)", + "17. Jung, H., Wu, S.C., Ruhkamp, P., Zhai, G., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., et al.: Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22498-22508 (2024)", + "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "19. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: Cosypose: Consistent multi-view multi-object 6d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 574-591. Springer (2020)", + "20. Li, X., Wang, H., Yi, L., Guibas, L., Abbott, A.L., Song, S.: Category-level articulated object pose estimation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020)", + "21. Lin, H., Liu, Z., Cheang, C., Fu, Y., Guo, G., Xue, X.: Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6707-6717 (2022)", + "22. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3560-3569 (2021)", + "23. Liu, L., Xu, W., Fu, H., Qian, S., Yu, Q., Han, Y., Lu, C.: Akb-48: A real-world articulated object knowledge base. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14809-14818 (2022)", + "24. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10870-10879 (2021)", + "25. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023)", + "26. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4938-4947 (2020)" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Stoiber, M., Elsayed, M., Reichert, A.E., Steidle, F., Lee, D., Triebel, R.: Fusing visual appearance and geometry for multi-modality 6dof object tracking. arXiv preprint arXiv:2302.11458 (2023)", + "28. Stoiber, M., Sundermeyer, M., Triebel, R.: Iterative corresponding geometry: Fusing region and depth for highly efficient 3d tracking of textureless objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6855-6865 (2022)", + "29. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2974-2983 (2018)", + "30. Tjaden, H., Schwanecke, U., Schömer, E., Cremers, D.: A region-based gauss-newton approach to real-time monocular multiple object tracking. IEEE transactions on pattern analysis and machine intelligence 41(8), 1797-1812 (2018)", + "31. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 10059-10066. IEEE (2020)", + "32. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019)", + "33. Wang, G., Manhardt, F., Tombari, F., Ji, X.: GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16611-16621 (June 2021)", + "34. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2642-2651 (2019)", + "35. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8067-8074. IEEE (2021)", + "36. Wen, B., Mitash, C., Ren, B., Bekris, K.E.: se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10367-10373. IEEE (2020)", + "37. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13209-13218 (2021)", + "38. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017)", + "39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. arXiv preprint arXiv:2211.13398 (2022)", + "40. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022)" + ], + "bbox": [ + 212, + 146, + 789, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PACE: Pose Annotations in Cluttered Environments", + "bbox": [ + 379, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Zhang, J., Huang, W., Peng, B., Wu, M., Hu, F., Chen, Z., Zhao, B., Dong, H.: Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. arXiv preprint arXiv:2406.04316 (2024)", + "42. Zhang, J., Wu, M., Dong, H.: Genpose: Generative category-level object pose estimation via diffusion models. arXiv preprint arXiv:2306.10531 (2023)", + "43. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-pose: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17163-17173 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 272 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. You et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_model.json b/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72e7ce9efd75255e007b929f21ee5bc32457a54e --- /dev/null +++ b/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_model.json @@ -0,0 +1,2678 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.271, + 0.142, + 0.734, + 0.185 + ], + "angle": 0, + "content": "PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.212, + 0.739, + 0.228 + ], + "angle": 0, + "content": "Yang You\\(^{1\\star}\\) Kai Xiong\\(^{2}\\) Zhening Yang\\(^{3}\\) Zhengxiang Huang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.228, + 0.648, + 0.244 + ], + "angle": 0, + "content": "Junwei Zhou² Ruoxi Shi⁴ Zhou Fang²" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.244, + 0.696, + 0.258 + ], + "angle": 0, + "content": "Adam W. Harley\\(^{1}\\) Leonidas Guibas\\(^{1}\\) Cewu Lu\\(^{2\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.27, + 0.576, + 0.285 + ], + "angle": 0, + "content": "1 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.285, + 0.612, + 0.299 + ], + "angle": 0, + "content": "2 Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.299, + 0.583, + 0.311 + ], + "angle": 0, + "content": "3 Horizon Robotics Inc." + }, + { + "type": "text", + "bbox": [ + 0.446, + 0.312, + 0.558, + 0.326 + ], + "angle": 0, + "content": "UC San Diego" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.358, + 0.331, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.41, + 0.33, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.359, + 0.443, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.411, + 0.443, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.359, + 0.557, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.411, + 0.556, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.359, + 0.671, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.411, + 0.671, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.359, + 0.784, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.411, + 0.784, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.472, + 0.789, + 0.501 + ], + "angle": 0, + "content": "Fig. 1: We propose PACE: a large-scale object pose dataset, with diverse objects, complex scenes, and various types of occlusions, reflecting real-world challenges." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.546, + 0.744, + 0.742 + ], + "angle": 0, + "content": "Abstract. We introduce PACE (Pose Annotations in Cluttered Environments), a large-scale benchmark designed to advance the development and evaluation of pose estimation methods in cluttered scenarios. PACE provides a large-scale real-world benchmark for both instance-level and category-level settings. The benchmark consists of 55K frames with 258K annotations across 300 videos, covering 238 objects from 43 categories and featuring a mix of rigid and articulated items in cluttered scenes. To annotate the real-world data efficiently, we develop an innovative annotation system with a calibrated 3-camera setup. Additionally, we offer PACE-Sim, which contains 100K photo-realistic simulated frames with 2.4M annotations across 931 objects. We test state-of-the-art algorithms in PACE along two tracks: pose estimation, and object pose tracking, revealing the benchmark's challenges and research opportunities. Our benchmark code and data is available on https://github.com/qq456cvb/PACE." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.789, + 0.789, + 0.821 + ], + "angle": 0, + "content": "The field of 3D object pose estimation is integral to a myriad of applications, particularly within robotic manipulation. Recent advancements in both instance and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.748, + 0.841 + ], + "angle": 0, + "content": "* Corresponding authors. Email: yangyou@stanford.edu, lucewu@sjtu.edu.cn" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.177 + ], + "angle": 0, + "content": "category-level pose estimation have been bolstered by deep learning approaches, and perhaps more importantly, by data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.253 + ], + "angle": 0, + "content": "PoseCNN [38] advanced pose estimation into the deep learning era, and simultaneously introduced the influential YCB-Video dataset. This benchmark has catalyzed methodological development and offered a consistent evaluation platform. Besides, the Benchmark for 6D Object Pose Estimation (BOP) challenges [14] have consolidated datasets metrics for instance-level pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.374 + ], + "angle": 0, + "content": "In parallel, NOCS [34] has addressed category-level pose estimation, albeit with a smaller real-world dataset for validation and testing. Despite these strides, the field grapples with a fundamental challenge: state-of-the-art pose estimation models still perform poorly in real-world settings, and existing evaluation datasets are too constrained to thoroughly reveal this fact. The NOCS-REAL275 dataset, for instance, spans only six categories and includes a mere 18 videos. This limitation has led to an array of published models which perform near-perfectly on NOCS but do not generalize to new data, as demonstrated in Figure 2." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.4, + 0.504, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.398, + 0.785, + 0.497 + ], + "angle": 0, + "content": "
DatasetMethodAP@15°5cm (%)↑
bottlebowlcanmug
REAL275 [34]HS-Pose [43]99.899.899.586.2
DualPoseNet [22]97.599.797.768.7
SARNet [21]98.298.297.470.6
PACE (Ours)HS-Pose [43]0.40.00.87.5
DualPoseNet [22]0.32.00.51.6
SARNet [21]2.10.00.30.0
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Fig. 2: While current state-of-the-art methods yield satisfactory outcomes on the NOCS-REAL275 dataset, their models' performance significantly deteriorates when transferred to previously unseen datasets such as PACE. Left: Qualitative visualizations of various models' pose predictions for a mug in REAL275 vs. a mug in PACE. Right: The performance of state-of-the-art methods markedly declines on PACE, even when evaluating on categories that exist in both datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.787, + 0.665 + ], + "angle": 0, + "content": "In this work, we introduce PACE, a benchmark for pose estimation, and present a comprehensive study evaluating a wide range of pose estimation and tracking methods. Our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.675, + 0.788, + 0.765 + ], + "angle": 0, + "content": "- The PACE dataset: This dataset includes 238 objects across 43 categories, captured in 300 video clips within diverse scenes. With an average of 183 frames per clip, the dataset consists of 55K frames and 258K annotations, providing a large-scale benchmark for pose estimation. We also provide PACESim, a photo-realistic training dataset which contains 100K frames with 2.4M annotations across 931 objects." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "- Our evaluation study: To the best of our knowledge, we are the first to analyze and report the performance of state-of-the-art (SOTA) pose estimation methods in a large-scale cluttered setting. These results provide valuable insights on the scalability and generalizability of SOTA methods, making it clear that they are far from reliable in general." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.675, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.38, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "- Our annotation pipeline: We will open-source our annotation pipeline, which uses a calibrated 3-camera system, enhancing the precision and scalability of annotating poses in real data. This tool significantly mitigates human error and reduces the effort required for annotating 3D poses, providing a solution to one of the major bottlenecks in pose dataset creation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.234, + 0.786, + 0.28 + ], + "angle": 0, + "content": "Overall, our work aims to support the development of more robust and generalizable pose estimation techniques, thereby facilitating progress towards successful pose estimation in the real world." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.302, + 0.396, + 0.318 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.786, + 0.379 + ], + "angle": 0, + "content": "The field of 3D object pose estimation has seen substantial progress over the past few years. This progress has been facilitated by the introduction of standardized datasets and the development of innovative algorithms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.4, + 0.44, + 0.416 + ], + "angle": 0, + "content": "2.1 Object Pose Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Instance-Level Pose Datasets YCB-Video dataset [38] is a comprehensive resource for 6D object pose estimation, containing a large number of video frames with accurate pose annotations for 21 objects. LINEMOD-Occluded [3] offers a challenging setting for pose estimation with piled multiple objects in occluded scenes. NAVI dataset [15] presents casually captured images of objects with high-quality 3D scans and precise 2D-3D alignments for 3D reconstruction tasks. Recently, StereoObj-1M [24] is proposed to address challenging cases such as object transparency, in addition to common challenges of occlusion and symmetry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Category-Level Pose Datasets NOCS-REAL275 [34] and Wild6D [40] both annotate a limited number of categories with upright poses. Objectron [1] is a collection of short, single-object-centric video clips with no clutters. The Scan2CAD [2] dataset aligns 14225 CAD models from ShapeNet to 1506 ScanNet scans, promoting CAD model alignment in RGB-D scans. Pix3D [29] is a benchmark with image-shape pairs and pixel-level 2D-3D alignment, aiding in shape reconstruction and retrieval. HANDAL [11] focuses on pose estimation and affordance prediction for robotics-ready manipulable objects. HouseCat6D [17] encompasses 194 diverse objects across 10 household categories with additional grasp annotations. ROPE [41] provides a large-scale evaluation benchmark covering a wide spectrum of categories. However, it does not include challenging scenarios such as moving and piled objects. Table 1 presents a comparative comparison of these datasets, alongside our new dataset PACE." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.475, + 0.8 + ], + "angle": 0, + "content": "2.2 Pose Estimation Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Instance-level Pose Estimation PPF (Point Pair Features) [8] is a pre-deep learning standard for instance-level pose estimation, using local geometric features" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.788, + 0.325 + ], + "angle": 0, + "content": "
ModalityCat.Obj.Vid.Img.Anno.CADMov.Occ.Marker-freeArtic.Piled
YCB-Video [38]RGBD-211220K99K✓X✓✓X✓
LINEMOD-O [3]RGBD-811.2K9.2K✓X✓XX✓
NAVI [15]RGBD-3632410K10k✓XX✓XX
StereoObj-1M [24]RGBD-18182393K1.5M✓X✓XX✓
NOCS-REAL275 [34]RGBD642188K✓X✓XXX
Wild6D [40]RGBD5172251661.1M1.1MXXX✓XX
Objectron [1]RGB917k14k4M4MXXX✓XX
Scan2CAD [2]RGBD93K1506-14K✓X✓✓XX
Pix3D [29]RGBD9395-10K10K✓XX✓XX
HANDAL [11]RGB172122K308K308K✓✓✓✓XX
HouseCat6D [17]RGBD101924124K160K✓X✓✓X✓
ROPE [41]RGBD149581363332K1.5M✓X✓✓XX
PACE (Ours)RGBD4323830055K258K✓✓✓✓✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Table 1: Comparison of object pose datasets. From left to right, the table captures input modality, number of categories, number of instances, number of videos, number of images, number of total annotations, whether 3D CAD models are provided, whether videos include still and/or moving objects, whether objects are occluded in some frames, whether images contain artificial markers, whether poses for each part of articulated objects are provided, and whether multiple objects are piled in some frames. Compared with most other datasets, PACE contains moving and articulated objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.79, + 0.524 + ], + "angle": 0, + "content": "from point clouds. PoseCNN [38] is the first model trained end-to-end for the task. CosyPose [19] integrates a global refinement strategy in its end-to-end pipeline. SurfEmb [12] leverages surface embeddings for correspondence matching, and GDRNPP [33] uses geometry-guided regression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.79, + 0.695 + ], + "angle": 0, + "content": "Category-level Pose Estimation Category-level pose estimation extends the challenge to generic object categories. NOCS [34] introduces a unified coordinate space for all objects, predicting object NOCS maps from RGB images. SGPA [4] aims to adapt the structure-guided prior in the pose estimation process, while SAR-Net [21] uses shape alignment and symmetric correspondence to estimate a coarse 3D object shape and facilitate object center and size estimation. Recently, HS-Pose [43] proposes a network structure with a HS-layer that extends 3D graph convolution to extract hybrid scope latent features from point clouds for category-level object pose estimation. GenPose [42] incorporates probabilistic stable diffusion models into pose estimation, achieving strong performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.713, + 0.458, + 0.728 + ], + "angle": 0, + "content": "2.3 Pose Tracking Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Instance-level Pose Tracking Methods like RBOT [30] use RGB data and 3D models to track multiple objects, employing color histograms in their cost function. PoseRBPF [5] separates rotation and translation, using an autoencoder for rotation feature embeddings. ICG [28] iteratively refines pose using geometric cues and is effective for textureless objects, with extensions incorporating visual data [27]. The first deep learning tracker, D6DT [9], and se(3)-TrackNet [36] predict frame-to-frame relative poses, using a render-and-compare strategy." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.38, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Category-level Pose Tracking 6-PACK [31] performs category-level tracking, using DenseFusion [32] features and an attention mechanism for unsupervised keypoint ordering and interframe motion via keypoint matching. BundleTrack [35] performs pose tracking without relying on 3D models, instead using video segmentation and pose graph optimization. CAPTRA [37] tracks 9DoF poses for rigid and articulated objects, with heads for rotation regression and normalized coordinate prediction, from which 3D size and translation can be computed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.283, + 0.473, + 0.299 + ], + "angle": 0, + "content": "3 Construction of PACE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.368 + ], + "angle": 0, + "content": "A key contribution of this work is the establishment of a scalable and reliable annotation framework, enabling the collection of large-scale and accurate pose annotations. An overview of the pipeline is depicted in Figure 3." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.395, + 0.427, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.395, + 0.758, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.58, + 0.673, + 0.596 + ], + "angle": 0, + "content": "Fig. 3: Overview of the PACE annotation pipeline." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.657, + 0.631, + 0.673 + ], + "angle": 0, + "content": "3.1 Acquisition of 3D Scans of Common Objects" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We began by digitizing an extensive collection of commonplace objects. These items were categorized into 43 distinct categories, as represented in Figure 5. The Einscan Pro 2X was utilized for rapid scanning of all objects, typically completing within 5 to 10 minutes per object. To expedite this process, we employed a rotatable platform to acquire multiple viewpoints of objects. After scanning, objects were manually aligned to a standard pose within a uniform coordinate system. For each object, we centered its axis-aligned bounding box at the origin and then aligned the bounding box orientation along a common axis within the category. We additionally annotated rotational symmetries. High-resolution meshes were simplified to lower-resolution ones for a smoother annotation workflow." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Articulated Objects Diverging from many prior pose estimation datasets, our collection encompasses a very wide set of objects, including articulated objects from the AKB48 [23] dataset, namely: scissors, cutters, clips, and boxes. We adopt the alignment methodology from the original AKB48 dataset, without modification. These objects are segmented into multiple parts with hierarchical relationships, presenting a difficult challenge for pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.263, + 0.506, + 0.279 + ], + "angle": 0, + "content": "3.2 RGB-D Sequence Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.79, + 0.398 + ], + "angle": 0, + "content": "We implemented a 3-camera system to aid in data acquisition and annotation, comprising three Intel Realsense D415 RGB-D cameras affixed to a metal framework, as illustrated in the bottom of Figure 3. The advantages of this setup include: (1) tripling the data yield; (2) reducing ambiguity in pose annotation, especially regarding translation along the depth dimension, by enforcing annotation consistency across all 3 views; (3) enhancing tracking accuracy of still objects with ArUco markers [10], making PnP more stable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.79, + 0.597 + ], + "angle": 0, + "content": "Calibration of Multi-Camera Extrinsic Parameters We calibrated this multicamera system through a semi-automatic process. ArUco markers [10] alone proved insufficient for high-accuracy rotation estimation. Hence, we resorted to trifocal tensor estimation, i.e. TFT [16]. The process begins with feature extraction and matching, followed by bundle adjustment to refine the positions of the 3D landmarks and camera poses. For reliable feature matching, we employed the SuperPoint [7] descriptor and SuperGlue [26] matcher, using a stringent threshold for matching. We observed that rotational component of the resulting extrinsic parameters is precise, but the translation aspect suffers from scale ambiguity inherent in Structure-from-Motion approaches. We corrected this by calibrating the scale against markers, applying the following formula to obtain the optimal scale factor, by setting gradient of \\(\\| s\\hat{\\mathbf{t}} - \\mathbf{t}'\\|_2^2\\) to zero:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.609, + 0.652, + 0.644 + ], + "angle": 0, + "content": "\\[\ns _ {1 \\rightarrow 2} = \\frac {\\hat {\\mathbf {t}} _ {1 \\rightarrow 2} \\cdot \\mathbf {t} _ {1 \\rightarrow 2} ^ {\\prime}}{\\hat {\\mathbf {t}} _ {1 \\rightarrow 2} \\cdot \\hat {\\mathbf {t}} _ {1 \\rightarrow 2}}, s _ {1 \\rightarrow 3} = \\frac {\\hat {\\mathbf {t}} _ {1 \\rightarrow 3} \\cdot \\mathbf {t} _ {1 \\rightarrow 3} ^ {\\prime}}{\\hat {\\mathbf {t}} _ {1 \\rightarrow 3} \\cdot \\hat {\\mathbf {t}} _ {1 \\rightarrow 3}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.728 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{t}}_{i\\rightarrow j}\\) is the TFT predicted translation (up to a scale) from camera \\(i\\) to camera \\(j\\), \\(\\mathbf{t}_{i\\rightarrow j}^{\\prime}\\) is the marker-calibrated translation in real metric scale. We set the extrinsics of the first camera to be the identity matrix. We use Intel's dynamic calibration tool to calibrate intrinsics. In cases of repetitive textures, where SuperPoint+SuperGlue were unreliable, we made correspondences manually." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.752, + 0.553, + 0.767 + ], + "angle": 0, + "content": "3.3 Annotation of Pose Ground-Truths" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Previous works have used ArUco markers to automate pose estimation through Perspective-n-Points, and we do this also. Specifically, we used ArUco markers to automate the annotation of still object poses (i.e. not moving), following LINEMOD [3] and NOCS [34]. We note that this step has two limitations:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.38, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.147, + 0.786, + 0.191 + ], + "angle": 0, + "content": "1. Markers in the scene detract from realism and compromise dataset integrity: training on marker-augmented imagery may result in overfitting to these artificial patterns." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.192, + 0.786, + 0.238 + ], + "angle": 0, + "content": "2. Marker-based annotations are inapplicable to moving objects (since the markers are typically placed on the background), thus limiting the setup to annotating still objects." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.147, + 0.786, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.784, + 0.296 + ], + "angle": 0, + "content": "We compensate for both of these issues: we remove marker appearances from the dataset (after using them for annotation), and we introduce additional tools to annotate the moving objects. We describe these compensation steps next." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.508 + ], + "angle": 0, + "content": "Marker removal We removed the marker appearances from the dataset, using a marker inpainting strategy, detailed as follows. In step 1, we place a marker (Marker 1) somewhere within the camera's field of view. We then record a short video with this marker in view. In step 2, we place a second marker (Marker 2) at a chosen distance from the first, and record another video. In step 3, we remove Marker 1, and begin the actual object capture process (with only Marker 2 present). After this process, we end up with: (1) frames with Marker 1 only, which clearly depict the surface where Marker 2 will later appear; (2) frames capturing both markers, providing helpful calibration cues; (3) frames with Marker 2 only, which represent our main capture. We leverage Marker 2 for automated pose tracking, and manually correct the tracking every 40 frames in case of drift. We use the frames from the first two steps to seamlessly inpaint [25] Marker 2's area for the final dataset, as depicted in Figure 4." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.534, + 0.785, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.615, + 0.677, + 0.63 + ], + "angle": 0, + "content": "Fig. 4: Illustration of the marker inpainting process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Annotation of Moving Object Positions For moving objects, the traditional marker-based tracking approach is insufficient, as the scene markers do not move in tandem with the objects. In such instances, we harnessed the capabilities of BundleTrack [35], an advanced RGB image-based tracking algorithm. BundleTrack conducts feature correspondence analysis between successive frames to estimate poses, complemented by a bundle adjustment algorithm to optimize keyframes globally and minimize tracking errors. Despite BundleTrack's proficiency in approximating poses, it is prone to drift, necessitating the manual adjustment of poses every ten frames to ensure precision. We adopted a rigorous manual checking process, examining every tenth frame for 3D pointcloud alignment and 2D multi-view agreement. We detail the annotation tool and its" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "graphical interface in the supplementary. This delicate process represents the most labor-intensive aspect of our annotation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.787, + 0.252 + ], + "angle": 0, + "content": "In order to verify that our annotation process gives accurate poses, we conducted an experiment where we used our setup to manually annotate the poses of Scene 1 from NOCS-REAL275 [34] dataset, and compare the annotations to the provided ground truths. We recorded an average pose annotation error of \\(0.9^{\\circ}\\) for rotation and \\(2.3\\mathrm{mm}\\) for translation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.267, + 0.787, + 0.343 + ], + "angle": 0, + "content": "Generation of Segmentation Masks Upon successful pose annotation, we generated occlusion-aware segmentation masks, via depth rendering. In cases where hand interactions are involved, we employed the segmentation model SAM [18] to estimate hand masks, and subsequently subtracted these from the previously computed object masks to obtain the occlusion-aware segmentation." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.366, + 0.424, + 0.383 + ], + "angle": 0, + "content": "4 Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.397, + 0.787, + 0.473 + ], + "angle": 0, + "content": "In pursuit of diversity, we placed the objects in ten disparate environments, each featuring varying levels and configurations of occlusion. We captured 10 video sequences per scene, each sequence including 1 to 7 objects. We captured RGB and depth data using an Intel RealSense D415 camera, with a resolution of \\(1280 \\times 720\\). The capture process spans distances ranging from \\(0.5\\mathrm{m}\\) to \\(1.5\\mathrm{m}\\)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.495, + 0.641, + 0.511 + ], + "angle": 0, + "content": "4.1 Annotation Distribution and Object Diversity" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.52, + 0.787, + 0.61 + ], + "angle": 0, + "content": "The comprehensive distribution of pose annotation counts is depicted in Figure 5, showcasing a diverse array including both rigid and articulated objects. As demonstrated in Figure 7, our object database encompasses a broad spectrum of object sizes, with the majority measuring near \\(0.2\\mathrm{m}\\) along the bounding box diagonal. It also includes larger items such as storage bins, adding to the heterogeneity of the dataset." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.632, + 0.745, + 0.648 + ], + "angle": 0, + "content": "4.2 Variability in Occlusion, Pose, and Environmental Context" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.748 + ], + "angle": 0, + "content": "The dataset has rich variability in occlusion levels and object poses, as illustrated in Figure 6, which outlines the statistical distribution of both visibility ratios and azimuth/elevation angles, indicating a comprehensive coverage reflecting real-world scenarios. These statistics can be helpful for assessing the robustness of pose estimation algorithms. Besides, the ten distinct capturing environments are shown in Figure 8, illustrating diverse backgrounds and lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.77, + 0.616, + 0.786 + ], + "angle": 0, + "content": "4.3 Simulated Training Data and Dataset Split" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To facilitate a comprehensive and equitable evaluation of pose estimation methodologies, we partitioned our entire real-world dataset into validation and test subsets following a \\(20/80\\) ratio. The test subset is used for both instance-level and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.381, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.149, + 0.753, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.354, + 0.768, + 0.369 + ], + "angle": 0, + "content": "Fig. 5: Distribution of pose annotation counts across different object categories." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.399, + 0.785, + 0.459 + ], + "angle": 0, + "content": "category-level pose estimation evaluation tasks. The real-world validation subset is used for instance-level pose estimation only. We render another photo-realistic validation subset for category-level pose estimation, in order to cover enough category-level objects that are different from the test set." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.46, + 0.787, + 0.566 + ], + "angle": 0, + "content": "Additionally, to support research necessitating extensive training datasets, we generated a large synthetic dataset called PACE-Sim that contains two different training subsets, for instance-level and category-level pose estimation, respectively. We use a physically based renderer [6] to ensure the images are highly photo-realistic. The synthetic training sets consist of 100K images, each featuring dozens of objects, yielding 2.4M annotations. The distribution of object instances in both PACE and PACE-Sim can be found in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.59, + 0.48, + 0.606 + ], + "angle": 0, + "content": "5 Evaluation Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.623, + 0.787, + 0.714 + ], + "angle": 0, + "content": "We assess both pose estimation and tracking. State-of-the-art (SOTA) pose estimation techniques generally involve two phases: first detecting or segmenting the object, and then estimating its pose within the identified mask. For a fair comparison, we evaluate pose estimation assuming perfect object detection. Detailed SOTA detection performance is available in the supplementary materials but is beyond this dataset's primary scope." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.714, + 0.785, + 0.759 + ], + "angle": 0, + "content": "For pose tracking, we provide ground-truth data in the initial frame and require methods to accurately track the object's rotations and translations through the video sequence. Both model-free and model-based trackers are assessed." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.784, + 0.495, + 0.798 + ], + "angle": 0, + "content": "5.1 Pose Estimation Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This benchmark is divided into instance-level and category-level pose estimation. The former concerns object instances seen during training (in different scenes)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.145, + 0.52, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.241, + 0.4, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.245, + 0.551, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.183, + 0.785, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.327, + 0.782, + 0.342 + ], + "angle": 0, + "content": "Fig. 7: Distribution of object sizes." + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.352, + 0.553, + 0.38 + ], + "angle": 0, + "content": "Fig. 6: Distribution of occlusions, azimuth and elevation viewing angles." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.396, + 0.328, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.447, + 0.303, + 0.458 + ], + "angle": 0, + "content": "(a) basket" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.397, + 0.442, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.447, + 0.422, + 0.459 + ], + "angle": 0, + "content": "(b) box top" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.396, + 0.556, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.447, + 0.534, + 0.458 + ], + "angle": 0, + "content": "(c) cabinet" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.396, + 0.671, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.447, + 0.646, + 0.458 + ], + "angle": 0, + "content": "(d) carpet" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.396, + 0.786, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.447, + 0.756, + 0.458 + ], + "angle": 0, + "content": "(e) chair" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.459, + 0.328, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.509, + 0.315, + 0.52 + ], + "angle": 0, + "content": "(f) china floor" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.459, + 0.442, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.509, + 0.415, + 0.52 + ], + "angle": 0, + "content": "(g) desk1" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.459, + 0.556, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.509, + 0.531, + 0.52 + ], + "angle": 0, + "content": "(h) desk2" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.459, + 0.671, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.509, + 0.642, + 0.52 + ], + "angle": 0, + "content": "(i) stairs" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.459, + 0.786, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.509, + 0.778, + 0.52 + ], + "angle": 0, + "content": "(j) wooden floor" + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.532, + 0.667, + 0.547 + ], + "angle": 0, + "content": "Fig. 8: Sample images from the 10 environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.608 + ], + "angle": 0, + "content": "while the latter concerns instances which were not seen but come from the same category. We will also give an in-depth analysis of SOTA's failure modes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.785, + 0.667 + ], + "angle": 0, + "content": "Instance-Level Pose Estimation This task demands the prediction of rotation and translation for known instances from the training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.761 + ], + "angle": 0, + "content": "Metrics: Adhering to the BOP challenge protocol [14], we use Average Recall (AR) of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) as our metrics. Detailed computation of these metrics is described in [14]. Results are averaged across all instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baselines: We assessed four baselines: PPF [8], CosyPose [19], SurfEmb [12], and GDRNPP [33]. While PPF does not require training data, the others are SOTA methods dependent on additional training on the PBR dataset. To focus on pose estimation performance, we assume that ground-truth detections are available." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.38, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Result Analysis: The quantitative analysis is presented in Table 2. Notably, while SOTA methods excel on the BOP benchmarks, they do not perform as well as PPF, which relies on local geometric feature matching in point clouds. The relative success of PPF highlights the potential for improvement in SOTA methods, particularly regarding sim-to-real ability and scalability to handle a large set of instances. Qualitative results can be found in the supplementary." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.263, + 0.699, + 0.334 + ], + "angle": 0, + "content": "
Modality DetectionARVSD↑ARMSSD↑ARMSPD↑AR↑
PPF [8]DG.T.53.448.155.252.2
CosyPose [19]RGBG.T.1.40.311.54.4
SurfEmb [12]RGBG.T.6.23.017.89.0
GDRNPP [33]RGB-DG.T.3.62.115.47.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.345, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Table 2: Instance-level pose estimation results showcasing the robustness of PPF and the potential for improvement in SOTA deep learning-based methods." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.442, + 0.784, + 0.569 + ], + "angle": 0, + "content": "
DetectionIoU25↑IoU50↑AP
0:20°↑0:60°↑0:5cm↑0:15cm↑0:20°↑0:60°↑
NOCS [34]Mask-RCNN0.2/0.00.0/0.01.2/0.04.3/0.017.0/0.033.5/0.01.1/0.04.2/0.0
HS-Pose [43]G.T.36.6/0.02.7/0.05.3/0.08.6/0.348.7/34.181.0/75.63.9/0.08.1/0.3
SGPA [4]G.T.2.6/0.01.2/0.05.6/1.111.4/7.316.1/6.850.7/17.53.3/0.910.1/7.2
DualPoseNet [22]G.T.24.2/0.30.1/0.05.5/0.08.1/0.018.6/27.363.4/65.21.8/0.06.5/0.0
SAR-Net [21]G.T.22.8/0.10.2/0.05.3/0.08.8/0.848.6/38.579.8/69.73.7/0.08.2/0.8
CPPF++ [39]G.T.44.5/3.64.4/0.015.2/1.727.3/6.235.3/31.174.0/66.79.9/1.124.9/5.9
ANCSH [20]G.T.- / 0.0- / 0.0- / 0.0- / 0.2- / 18.6- / 50.4- / 0.0- / 0.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Table 3: Category-level pose estimation benchmark combining the performance metrics for both rigid and articulated object pose estimation, separated by slash." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Category-Level Pose Estimation Category-level pose estimation tasks involve predicting the 3D bounding box dimensions, rotation, and translation of target instances, given only the category information a priori." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Metrics: We adopted from NOCS [34], calculating mean Average Precision (AP) across predefined angle and translation thresholds. Specifically, AP@0:20° represents AP averaged from \\(0^{\\circ}\\) to \\(20^{\\circ}\\) at \\(1^{\\circ}\\) intervals; AP@0:5cm is averaged from 0cm to 5cm at 0.25cm intervals; AP@0:20°,0:5cm combines these angle and translation thresholds. IoU\\(_{25}\\) and IoU\\(_{50}\\) denote AP for 3D bounding box matches at IoU thresholds of 25 and 50, respectively. Results are separately reported for both rigid and articulated objects." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Baselines: We evaluated seven recent category-level pose estimation methods: NOCS [34], HS-Pose [43], SGPA [4], DualPoseNet [22], SAR-Net [21], CPPF++ [39] and ANCSH [20], the latter specifically designed for articulated objects. For adapting rigid-object-focused methods to articulated objects, each movable part is considered a distinct category. Ground-truth detections are presumed for each method, except for NOCS, which outputs both instance masks and pose estimations using a unified network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.79, + 0.406 + ], + "angle": 0, + "content": "Result Analysis: Quantitative comparisons are listed in Table 3. On rigid objects, HS-Pose and SAR-Net show fair performance on AP metrics for translation, but falls short on rotation AP metrics compared to CPPF++, suggesting a proficiency in size and translation prediction but not rotation. CPPF++ shows the best performance but is weaker on translation estimation. For articulated objects, all methods face challenges due to the movable parts, with ANCSH underperforming on the large-scale real-world dataset, indicative of a significant sim-to-real gap. NOCS shows instability, likely due to its inaccurate mask prediction. Qualitative results are available in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Failure Mode Analysis This section delves into the analysis of two critical questions that underline the failure of SOTA deep learning-based methods in pose estimation: 1) Is the failure attributed to the sim-to-real gap, considering the exclusive use of synthetic RGB-D images for training? 2) Does the failure stem from the inherent large-scale complexity of the PACE dataset, which poses a significant challenge for baseline models?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Ablation Setup: To explore these questions, we introduced a new \\(80/10/10\\) train/validation/test split for the real-world data, diverging from the \\(0/20/80\\) split described in Section 4.3. We then examined the following settings:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.597, + 0.786, + 0.641 + ], + "angle": 0, + "content": "Real2Real Single Instance/Category Fitting: This basic setup involves training on a single instance or category selected (we randomly select the category can and sample a can with object ID 57) from the \\(80 / 10 / 10\\) split of real data." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "- Sim2Real Single Instance/Category Fitting: This scenario differs from the first by utilizing all simulated data for training and validation, with evaluation on the same real-world test split." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.688, + 0.785, + 0.717 + ], + "angle": 0, + "content": "- Real2Real All Instance/Category Fitting: Unlike the first setting, this involves training, validation, and testing across all instances or categories." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.597, + 0.786, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Result Analysis and Discussions: Our analysis in Table 4 reveals that baseline models effectively handle single instance or category fitting under the Real2Real Single Instance/Category Fitting setting, provided the training and test distributions align. However, the sim-to-real gap, especially noticeable in depth-input methods like DualPoseNet, HS-Pose, and SGPA, highlights significant challenges, as these methods struggle with the depth noise in real world." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.38, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.201 + ], + "angle": 0, + "content": "
Training SettingInstance-Level (AR%)†Category-Level (mean AP%@0:60°, 0:15cm)†
CosyPoseSurfEmbGDRNPPNOCSHS-PoseSGPADualPoseNetSAR-NetCPPF++
Real2Real Single Inst./Cat.76.486.788.276.981.286.758.659.383.4
Sim2Real Single Inst./Cat.67.1 (-9.3↓)73.7 (-13.0↓)78.1 (-10.1↓)55.8 (-21.1↓)1.1 (-80.1↓)0.3 (-86.4↓)1.4 (-57.2↓)33.2 (-26.1↓)55.8 (-27.6↓)
Real2Real All Inst./Cat.8.2 (-68.2↓)11.2 (-75.5↓)9.7 (-78.5↓)0.3 (-76.6↓)5.6 (-75.6↓)9.8 (-76.9↓)1.2 (-57.4↓)2.1 (-57.2↓)13.7 (-69.7↓)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.202, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Table 4: Ablation Study: This study illustrates the challenges models face when transitioning from real-to-real settings. It showcases performance disparities when scaling from single instances/categories to the entire dataset, emphasizing the scalability challenge from PACE and the profound impact of sim-to-real discrepancies on depth-dependent methodologies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.402 + ], + "angle": 0, + "content": "Additionally, learning across all instances/categories leads to a notable performance drop, as shown in Table 4. This drop, together with suboptimal training data performance and loss plateauing, suggests underfitting and highlights the complexity posed by the PACE dataset's diversity. Innovative approaches or larger foundational models might be necessary to address these challenges and effectively bridge the sim-to-real gap." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.421, + 0.478, + 0.436 + ], + "angle": 0, + "content": "5.2 Pose Tracking Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.474 + ], + "angle": 0, + "content": "We distinguish SOTA pose tracking techniques into model-based, utilizing a 3D CAD model, and model-free, which only require an initial pose from first frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.789, + 0.561 + ], + "angle": 0, + "content": "Metrics: For model-based tracking, we measure the area under curve (AUC) for ADD, ADD-S [38], and ADD(-S) metrics, with a maximum threshold of \\(0.1\\mathrm{m}\\) [32]. The ADD metric, introduced in [13], computes the average per-point distance between point clouds transformed by the predicted and ground-truth poses. ADD-S accounts for point correspondence ambiguities in symmetric objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.789, + 0.683 + ], + "angle": 0, + "content": "For model-free tracking, we adopt four metrics from prior studies [31]: 1) \\(5^{\\circ}5\\mathrm{cm}\\), indicating the percentage of predictions within \\(5^{\\circ}\\) rotation and \\(5\\mathrm{cm}\\) translation errors. 2) IoU25, the percentage of instances where the intersection over union of the 3D bounding boxes exceeds \\(25\\%\\). 3) \\(\\mathbf{R}_{\\mathrm{err}}\\) and 4) \\(\\mathbf{T}_{\\mathrm{err}}\\), the mean rotational and translational errors in degrees and centimeters, respectively, only considering instances with IoU \\(>25\\%\\). We exclude objects with visibility under \\(10\\%\\) as per the BOP convention [14]. Results are averaged separately across rigid and articulated object categories." + }, + { + "type": "table", + "bbox": [ + 0.342, + 0.706, + 0.664, + 0.759 + ], + "angle": 0, + "content": "
ModalityADD↑ADD-S↑ADD(-S)↑
RBOT [30]RGB7.1/0.510.3/0.87.4/0.5
ICG [28]RGB-D35.6/10.148.1/15.238.1/10.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.768, + 0.788, + 0.797 + ], + "angle": 0, + "content": "Table 5: Model-Based Pose Tracking. We report the area under curve (AUC) with respect to ADD, ADD-S and ADD(-S). The higher value, the better performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.144, + 0.728, + 0.205 + ], + "angle": 0, + "content": "
Training-Free Modality5°5cm↑IoU25↑Rerr↓Terr↓
BundleTrack [35]✓RGB6.4/11.29.1/14.13.2/5.52.6/0.8
CAPTRA [37]✗D12.9/4.445.8/20.619.2/40.92.2/1.5
6-PACK [31]✗RGB-D9.2/3.923.1/16.717.7/33.62.1/1.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.216, + 0.789, + 0.273 + ], + "angle": 0, + "content": "Table 6: Model-Free Pose Tracking. Both rigid and articulated results are reported. For \\(5^{\\circ}5\\mathrm{cm}\\) and IoU25, the higher value means better performance while for \\(\\mathbf{R}_{err}\\) and \\(\\mathrm{T_{err}}\\) , the situation is reversed. Note that \\(\\mathrm{R}_{err}\\) and \\(\\mathrm{T_{err}}\\) are respect to IoU25 since objects with IoU \\(\\leq 25\\%\\) are not counted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.404 + ], + "angle": 0, + "content": "Baselines: For tracking methods using object masks, we applied the ground-truth masks directly. Except for CAPTRA, which also predicts the 3D bounding box size, all methods estimate only rotation and translation; however, we used ground-truth bounding box sizes for consistency in evaluations. Unlike other methods that treat each part of an articulated object independently, CAPTRA considers the entire object as one entity, trained on our synthetic PBR data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.505 + ], + "angle": 0, + "content": "Result and Failure Mode Analysis: Tables 5 and 6 display the performance of state-of-the-art (SOTA) tracking methods on the PACE dataset, highlighting both rigid and articulated object results. The highest achieved ADD(-S) for model-based methods is only \\(38.1\\%\\) for rigid and \\(10.1\\%\\) for articulated objects by ICG. Furthermore, Table 6 reveals severe limitations of model-free methods, with the best \\(5^{\\circ}5\\mathrm{cm}\\) accuracy under \\(13\\%\\), posing major tracking challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.79, + 0.612 + ], + "angle": 0, + "content": "The convention followed by 6-PACK [31] and BundleTrack [35] disregards objects with \\(\\mathrm{IoU} \\leq 25\\%\\) when computing rotational and translational errors. However, this approach could overestimate the performance on \\(\\mathrm{R}_{err}\\) and \\(\\mathrm{T}_{err}\\). BundleTrack [35] reports lower performance in \\(5^{\\circ}5\\mathrm{cm}\\) and IoU25 metrics yet shows seemingly better results for \\(\\mathrm{R}_{err}\\) and \\(\\mathrm{T}_{err}\\), suggesting that it can only track well within a threshold but cannot recover once the object is lost. Qualitative results can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.632, + 0.53, + 0.649 + ], + "angle": 0, + "content": "6 Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.789, + 0.783 + ], + "angle": 0, + "content": "We present PACE, a new large-scale benchmark for 3D object pose estimation and tracking, designed with various occlusion levels across ten diverse environments. This benchmark both reflects prior progress and urges further research to tackle real-world complexities. Our analysis indicates significant advances in pose estimation, yet underscores a notable real-world performance gap, particularly with articulated objects and in terms of robustness and scalability. Future efforts could focus on advanced sim-to-real techniques and developing larger models that capture a broader spectrum of object features and environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Limitations Due to resource limitations, PACE lacks large and valuable objects like tables, cameras, and laptops, which may affect its breadth in representing real-world object scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.38, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.436, + 0.163 + ], + "angle": 0, + "content": "7 Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.252 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Project of China (No. 2022ZD0160102), National Key Research and Development Project of China (No. 2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants. Yang You is also supported in part by the Outstanding Doctoral Graduates Development Scholarship of Shanghai Jiao Tong University." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.324, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.787, + 0.359 + ], + "angle": 0, + "content": "1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7822-7831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.787, + 0.401 + ], + "angle": 0, + "content": "2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2614-2623 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.787, + 0.456 + ], + "angle": 0, + "content": "3. Brachmann, E., Krull, A., Michel, F., Gumhold, S., Shotton, J., Rother, C.: Learning 6d object pose estimation using 3d object coordinates. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 536-551. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.787, + 0.497 + ], + "angle": 0, + "content": "4. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.787, + 0.538 + ], + "angle": 0, + "content": "5. Deng, X., Mousavian, A., Xiang, Y., Xia, F., Bretl, T., Fox, D.: Poserbpf: A raobackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics 37(5), 1328-1342 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.787, + 0.594 + ], + "angle": 0, + "content": "6. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.787, + 0.635 + ], + "angle": 0, + "content": "7. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 224-236 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.787, + 0.676 + ], + "angle": 0, + "content": "8. Drost, B., Ulrich, M., Navab, N., Ilic, S.: Model globally, match locally: Efficient and robust 3d object recognition. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 998-1005. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.787, + 0.703 + ], + "angle": 0, + "content": "9. Garon, M., Lalonde, J.F.: Deep 6-dof tracking. IEEE transactions on visualization and computer graphics 23(11), 2410-2418 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.787, + 0.744 + ], + "angle": 0, + "content": "0. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition 47(6), 2280-2292 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.786 + ], + "angle": 0, + "content": "1. Guo, A., Wen, B., Yuan, J., OTHERS: Handal: A dataset of real-world manipulable object categories with pose annotations, affordances, and reconstructions (2023), https://arxiv.org/pdf/2308.01477.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "2. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.303, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.127 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.217 + ], + "angle": 0, + "content": "13. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. pp. 548-562. Springer (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.788, + 0.272 + ], + "angle": 0, + "content": "14. Hodaţ, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 577-594. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.272, + 0.788, + 0.325 + ], + "angle": 0, + "content": "15. Jampani, V., Maninis, K.K., Engelhardt, A., Karpur, A., Truong, K., Sargent, K., Popov, S., Araujo, A., Martin-Brualla, R., Patel, K., et al.: Navi: Category-agnostic image collections with high-quality 3d shape and pose annotations. arXiv preprint arXiv:2306.09109 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.788, + 0.379 + ], + "angle": 0, + "content": "16. Julià, L.F., Monasse, P.: A critical review of the trifocal tensor estimation. In: Image and Video Technology: 8th Pacific-Rim Symposium, PSIVT 2017, Wuhan, China, November 20-24, 2017, Revised Selected Papers 8. pp. 337-349. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.788, + 0.448 + ], + "angle": 0, + "content": "17. Jung, H., Wu, S.C., Ruhkamp, P., Zhai, G., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., et al.: Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22498-22508 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.448, + 0.788, + 0.488 + ], + "angle": 0, + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.788, + 0.543 + ], + "angle": 0, + "content": "19. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: Cosypose: Consistent multi-view multi-object 6d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 574-591. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.543, + 0.788, + 0.583 + ], + "angle": 0, + "content": "20. Li, X., Wang, H., Yi, L., Guibas, L., Abbott, A.L., Song, S.: Category-level articulated object pose estimation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.583, + 0.788, + 0.638 + ], + "angle": 0, + "content": "21. Lin, H., Liu, Z., Cheang, C., Fu, Y., Guo, G., Xue, X.: Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6707-6717 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.638, + 0.788, + 0.692 + ], + "angle": 0, + "content": "22. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3560-3569 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.692, + 0.788, + 0.733 + ], + "angle": 0, + "content": "23. Liu, L., Xu, W., Fu, H., Qian, S., Yu, Q., Han, Y., Lu, C.: Akb-48: A real-world articulated object knowledge base. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14809-14818 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.733, + 0.788, + 0.772 + ], + "angle": 0, + "content": "24. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10870-10879 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "25. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "26. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4938-4947 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.38, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PACE: Pose Annotations in Cluttered Environments" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "27. Stoiber, M., Elsayed, M., Reichert, A.E., Steidle, F., Lee, D., Triebel, R.: Fusing visual appearance and geometry for multi-modality 6dof object tracking. arXiv preprint arXiv:2302.11458 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.79, + 0.245 + ], + "angle": 0, + "content": "28. Stoiber, M., Sundermeyer, M., Triebel, R.: Iterative corresponding geometry: Fusing region and depth for highly efficient 3d tracking of textureless objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6855-6865 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.79, + 0.3 + ], + "angle": 0, + "content": "29. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2974-2983 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.789, + 0.342 + ], + "angle": 0, + "content": "30. Tjaden, H., Schwanecke, U., Schömer, E., Cremers, D.: A region-based gauss-newton approach to real-time monocular multiple object tracking. IEEE transactions on pattern analysis and machine intelligence 41(8), 1797-1812 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.79, + 0.397 + ], + "angle": 0, + "content": "31. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 10059-10066. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.79, + 0.452 + ], + "angle": 0, + "content": "32. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.789, + 0.508 + ], + "angle": 0, + "content": "33. Wang, G., Manhardt, F., Tombari, F., Ji, X.: GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16611-16621 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.789, + 0.564 + ], + "angle": 0, + "content": "34. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2642-2651 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.789, + 0.605 + ], + "angle": 0, + "content": "35. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8067-8074. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.789, + 0.66 + ], + "angle": 0, + "content": "36. Wen, B., Mitash, C., Ren, B., Bekris, K.E.: se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10367-10373. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.789, + 0.716 + ], + "angle": 0, + "content": "37. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13209-13218 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.789, + 0.757 + ], + "angle": 0, + "content": "38. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.789, + 0.799 + ], + "angle": 0, + "content": "39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. arXiv preprint arXiv:2211.13398 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.789, + 0.841 + ], + "angle": 0, + "content": "40. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.79, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. You et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "41. Zhang, J., Huang, W., Peng, B., Wu, M., Hu, F., Chen, Z., Zhao, B., Dong, H.: Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. arXiv preprint arXiv:2406.04316 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "42. Zhang, J., Wu, M., Dong, H.: Genpose: Generative category-level object pose estimation via diffusion models. arXiv preprint arXiv:2306.10531 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.273 + ], + "angle": 0, + "content": "43. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-pose: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17163-17173 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_origin.pdf b/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bba01014ec6b23ae76f7802bb7d67c7f9c4d5fa7 --- /dev/null +++ b/2024/PACE_ Pose Annotations in Cluttered Environments/13a98d0a-b6bf-4769-9157-c4d8353816c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b6ca253512ea85c3edd56b11e3aafb09740b52122d02cd4c797f2d3a63ed448 +size 3885622 diff --git a/2024/PACE_ Pose Annotations in Cluttered Environments/full.md b/2024/PACE_ Pose Annotations in Cluttered Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..11df2c30827e735e834771d23f6a8376c47b7b5f --- /dev/null +++ b/2024/PACE_ Pose Annotations in Cluttered Environments/full.md @@ -0,0 +1,336 @@ +# PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments + +Yang You $^{1\star}$ Kai Xiong $^{2}$ Zhening Yang $^{3}$ Zhengxiang Huang $^{2}$ + +Junwei Zhou² Ruoxi Shi⁴ Zhou Fang² + +Adam W. Harley $^{1}$ Leonidas Guibas $^{1}$ Cewu Lu $^{2\star}$ + +1 Stanford University + +2 Shanghai Jiao Tong University + +3 Horizon Robotics Inc. + +UC San Diego + +![](images/709319356a84f83ce0ded0ee55b0d9d33032ba0e8c81825c61ca5ce0bb79cb41.jpg) + +![](images/582e21ec12be91a48836679f23f4bc568f7da8b23de0c6cee9c97e0e29dabacc.jpg) +Fig. 1: We propose PACE: a large-scale object pose dataset, with diverse objects, complex scenes, and various types of occlusions, reflecting real-world challenges. + +![](images/3da2842c4be018ab84f33186bbe51cd77c96cf30cd79c2e846242e9e50ee0397.jpg) + +![](images/ad871cf0ebf1b0fa973cc0a89380acaa8bfcd8ab8309b7a747d4ba7d919677bb.jpg) + +![](images/01c4e6e6a4833910bed75a900a139974904b68085bf98d60b01b67bcfce7ddb4.jpg) + +![](images/5d4c03167f2156c6ccf163d36d6d548c006270afaa1dbb0d799395da2d4ff3e8.jpg) + +![](images/02b4d783904db49da959e93fac7a08b08792064619b79a2b98b334bb39c3e59c.jpg) + +![](images/2b1bc824ed10b2d585c91f30ad951f2f616004511f789de50be521995e6c3610.jpg) + +![](images/dec61119fa07d701745b1d54a945cc6bee9fea09d6d4550f6400b3d3dabea163.jpg) + +![](images/5e9e474a3d2d627184d9d2a02db6866a1c7b5e62fddcd350623d85924a20a37d.jpg) + +Abstract. We introduce PACE (Pose Annotations in Cluttered Environments), a large-scale benchmark designed to advance the development and evaluation of pose estimation methods in cluttered scenarios. PACE provides a large-scale real-world benchmark for both instance-level and category-level settings. The benchmark consists of 55K frames with 258K annotations across 300 videos, covering 238 objects from 43 categories and featuring a mix of rigid and articulated items in cluttered scenes. To annotate the real-world data efficiently, we develop an innovative annotation system with a calibrated 3-camera setup. Additionally, we offer PACE-Sim, which contains 100K photo-realistic simulated frames with 2.4M annotations across 931 objects. We test state-of-the-art algorithms in PACE along two tracks: pose estimation, and object pose tracking, revealing the benchmark's challenges and research opportunities. Our benchmark code and data is available on https://github.com/qq456cvb/PACE. + +# 1 Introduction + +The field of 3D object pose estimation is integral to a myriad of applications, particularly within robotic manipulation. Recent advancements in both instance and + +category-level pose estimation have been bolstered by deep learning approaches, and perhaps more importantly, by data. + +PoseCNN [38] advanced pose estimation into the deep learning era, and simultaneously introduced the influential YCB-Video dataset. This benchmark has catalyzed methodological development and offered a consistent evaluation platform. Besides, the Benchmark for 6D Object Pose Estimation (BOP) challenges [14] have consolidated datasets metrics for instance-level pose estimation. + +In parallel, NOCS [34] has addressed category-level pose estimation, albeit with a smaller real-world dataset for validation and testing. Despite these strides, the field grapples with a fundamental challenge: state-of-the-art pose estimation models still perform poorly in real-world settings, and existing evaluation datasets are too constrained to thoroughly reveal this fact. The NOCS-REAL275 dataset, for instance, spans only six categories and includes a mere 18 videos. This limitation has led to an array of published models which perform near-perfectly on NOCS but do not generalize to new data, as demonstrated in Figure 2. + +![](images/c20085b68f4ec9a69863169cd08c64499d50d73275af2b321dfbd7855a3b3f31.jpg) +Fig. 2: While current state-of-the-art methods yield satisfactory outcomes on the NOCS-REAL275 dataset, their models' performance significantly deteriorates when transferred to previously unseen datasets such as PACE. Left: Qualitative visualizations of various models' pose predictions for a mug in REAL275 vs. a mug in PACE. Right: The performance of state-of-the-art methods markedly declines on PACE, even when evaluating on categories that exist in both datasets. + +
DatasetMethodAP@15°5cm (%)↑
bottlebowlcanmug
REAL275 [34]HS-Pose [43]99.899.899.586.2
DualPoseNet [22]97.599.797.768.7
SARNet [21]98.298.297.470.6
PACE (Ours)HS-Pose [43]0.40.00.87.5
DualPoseNet [22]0.32.00.51.6
SARNet [21]2.10.00.30.0
+ +In this work, we introduce PACE, a benchmark for pose estimation, and present a comprehensive study evaluating a wide range of pose estimation and tracking methods. Our contributions are threefold: + +- The PACE dataset: This dataset includes 238 objects across 43 categories, captured in 300 video clips within diverse scenes. With an average of 183 frames per clip, the dataset consists of 55K frames and 258K annotations, providing a large-scale benchmark for pose estimation. We also provide PACESim, a photo-realistic training dataset which contains 100K frames with 2.4M annotations across 931 objects. +- Our evaluation study: To the best of our knowledge, we are the first to analyze and report the performance of state-of-the-art (SOTA) pose estimation methods in a large-scale cluttered setting. These results provide valuable insights on the scalability and generalizability of SOTA methods, making it clear that they are far from reliable in general. + +- Our annotation pipeline: We will open-source our annotation pipeline, which uses a calibrated 3-camera system, enhancing the precision and scalability of annotating poses in real data. This tool significantly mitigates human error and reduces the effort required for annotating 3D poses, providing a solution to one of the major bottlenecks in pose dataset creation. + +Overall, our work aims to support the development of more robust and generalizable pose estimation techniques, thereby facilitating progress towards successful pose estimation in the real world. + +# 2 Related Works + +The field of 3D object pose estimation has seen substantial progress over the past few years. This progress has been facilitated by the introduction of standardized datasets and the development of innovative algorithms. + +# 2.1 Object Pose Datasets + +Instance-Level Pose Datasets YCB-Video dataset [38] is a comprehensive resource for 6D object pose estimation, containing a large number of video frames with accurate pose annotations for 21 objects. LINEMOD-Occluded [3] offers a challenging setting for pose estimation with piled multiple objects in occluded scenes. NAVI dataset [15] presents casually captured images of objects with high-quality 3D scans and precise 2D-3D alignments for 3D reconstruction tasks. Recently, StereoObj-1M [24] is proposed to address challenging cases such as object transparency, in addition to common challenges of occlusion and symmetry. + +Category-Level Pose Datasets NOCS-REAL275 [34] and Wild6D [40] both annotate a limited number of categories with upright poses. Objectron [1] is a collection of short, single-object-centric video clips with no clutters. The Scan2CAD [2] dataset aligns 14225 CAD models from ShapeNet to 1506 ScanNet scans, promoting CAD model alignment in RGB-D scans. Pix3D [29] is a benchmark with image-shape pairs and pixel-level 2D-3D alignment, aiding in shape reconstruction and retrieval. HANDAL [11] focuses on pose estimation and affordance prediction for robotics-ready manipulable objects. HouseCat6D [17] encompasses 194 diverse objects across 10 household categories with additional grasp annotations. ROPE [41] provides a large-scale evaluation benchmark covering a wide spectrum of categories. However, it does not include challenging scenarios such as moving and piled objects. Table 1 presents a comparative comparison of these datasets, alongside our new dataset PACE. + +# 2.2 Pose Estimation Methods + +Instance-level Pose Estimation PPF (Point Pair Features) [8] is a pre-deep learning standard for instance-level pose estimation, using local geometric features + +
ModalityCat.Obj.Vid.Img.Anno.CADMov.Occ.Marker-freeArtic.Piled
YCB-Video [38]RGBD-211220K99K✓X✓✓X✓
LINEMOD-O [3]RGBD-811.2K9.2K✓X✓XX✓
NAVI [15]RGBD-3632410K10k✓XX✓XX
StereoObj-1M [24]RGBD-18182393K1.5M✓X✓XX✓
NOCS-REAL275 [34]RGBD642188K✓X✓XXX
Wild6D [40]RGBD5172251661.1M1.1MXXX✓XX
Objectron [1]RGB917k14k4M4MXXX✓XX
Scan2CAD [2]RGBD93K1506-14K✓X✓✓XX
Pix3D [29]RGBD9395-10K10K✓XX✓XX
HANDAL [11]RGB172122K308K308K✓✓✓✓XX
HouseCat6D [17]RGBD101924124K160K✓X✓✓X✓
ROPE [41]RGBD149581363332K1.5M✓X✓✓XX
PACE (Ours)RGBD4323830055K258K✓✓✓✓✓✓
+ +Table 1: Comparison of object pose datasets. From left to right, the table captures input modality, number of categories, number of instances, number of videos, number of images, number of total annotations, whether 3D CAD models are provided, whether videos include still and/or moving objects, whether objects are occluded in some frames, whether images contain artificial markers, whether poses for each part of articulated objects are provided, and whether multiple objects are piled in some frames. Compared with most other datasets, PACE contains moving and articulated objects. + +from point clouds. PoseCNN [38] is the first model trained end-to-end for the task. CosyPose [19] integrates a global refinement strategy in its end-to-end pipeline. SurfEmb [12] leverages surface embeddings for correspondence matching, and GDRNPP [33] uses geometry-guided regression. + +Category-level Pose Estimation Category-level pose estimation extends the challenge to generic object categories. NOCS [34] introduces a unified coordinate space for all objects, predicting object NOCS maps from RGB images. SGPA [4] aims to adapt the structure-guided prior in the pose estimation process, while SAR-Net [21] uses shape alignment and symmetric correspondence to estimate a coarse 3D object shape and facilitate object center and size estimation. Recently, HS-Pose [43] proposes a network structure with a HS-layer that extends 3D graph convolution to extract hybrid scope latent features from point clouds for category-level object pose estimation. GenPose [42] incorporates probabilistic stable diffusion models into pose estimation, achieving strong performance. + +# 2.3 Pose Tracking Methods + +Instance-level Pose Tracking Methods like RBOT [30] use RGB data and 3D models to track multiple objects, employing color histograms in their cost function. PoseRBPF [5] separates rotation and translation, using an autoencoder for rotation feature embeddings. ICG [28] iteratively refines pose using geometric cues and is effective for textureless objects, with extensions incorporating visual data [27]. The first deep learning tracker, D6DT [9], and se(3)-TrackNet [36] predict frame-to-frame relative poses, using a render-and-compare strategy. + +Category-level Pose Tracking 6-PACK [31] performs category-level tracking, using DenseFusion [32] features and an attention mechanism for unsupervised keypoint ordering and interframe motion via keypoint matching. BundleTrack [35] performs pose tracking without relying on 3D models, instead using video segmentation and pose graph optimization. CAPTRA [37] tracks 9DoF poses for rigid and articulated objects, with heads for rotation regression and normalized coordinate prediction, from which 3D size and translation can be computed. + +# 3 Construction of PACE + +A key contribution of this work is the establishment of a scalable and reliable annotation framework, enabling the collection of large-scale and accurate pose annotations. An overview of the pipeline is depicted in Figure 3. + +![](images/8d213f454beaf986d901978d1b98d1f1fd795646c7b6dbfdb5d3b5b5e3e7620f.jpg) +Fig. 3: Overview of the PACE annotation pipeline. + +![](images/a451200f8d3953ddf29530ce271117558c74735c7d539ec6663f1d06d5bfb53a.jpg) + +# 3.1 Acquisition of 3D Scans of Common Objects + +We began by digitizing an extensive collection of commonplace objects. These items were categorized into 43 distinct categories, as represented in Figure 5. The Einscan Pro 2X was utilized for rapid scanning of all objects, typically completing within 5 to 10 minutes per object. To expedite this process, we employed a rotatable platform to acquire multiple viewpoints of objects. After scanning, objects were manually aligned to a standard pose within a uniform coordinate system. For each object, we centered its axis-aligned bounding box at the origin and then aligned the bounding box orientation along a common axis within the category. We additionally annotated rotational symmetries. High-resolution meshes were simplified to lower-resolution ones for a smoother annotation workflow. + +Articulated Objects Diverging from many prior pose estimation datasets, our collection encompasses a very wide set of objects, including articulated objects from the AKB48 [23] dataset, namely: scissors, cutters, clips, and boxes. We adopt the alignment methodology from the original AKB48 dataset, without modification. These objects are segmented into multiple parts with hierarchical relationships, presenting a difficult challenge for pose estimation. + +# 3.2 RGB-D Sequence Acquisition + +We implemented a 3-camera system to aid in data acquisition and annotation, comprising three Intel Realsense D415 RGB-D cameras affixed to a metal framework, as illustrated in the bottom of Figure 3. The advantages of this setup include: (1) tripling the data yield; (2) reducing ambiguity in pose annotation, especially regarding translation along the depth dimension, by enforcing annotation consistency across all 3 views; (3) enhancing tracking accuracy of still objects with ArUco markers [10], making PnP more stable. + +Calibration of Multi-Camera Extrinsic Parameters We calibrated this multicamera system through a semi-automatic process. ArUco markers [10] alone proved insufficient for high-accuracy rotation estimation. Hence, we resorted to trifocal tensor estimation, i.e. TFT [16]. The process begins with feature extraction and matching, followed by bundle adjustment to refine the positions of the 3D landmarks and camera poses. For reliable feature matching, we employed the SuperPoint [7] descriptor and SuperGlue [26] matcher, using a stringent threshold for matching. We observed that rotational component of the resulting extrinsic parameters is precise, but the translation aspect suffers from scale ambiguity inherent in Structure-from-Motion approaches. We corrected this by calibrating the scale against markers, applying the following formula to obtain the optimal scale factor, by setting gradient of $\| s\hat{\mathbf{t}} - \mathbf{t}'\|_2^2$ to zero: + +$$ +s _ {1 \rightarrow 2} = \frac {\hat {\mathbf {t}} _ {1 \rightarrow 2} \cdot \mathbf {t} _ {1 \rightarrow 2} ^ {\prime}}{\hat {\mathbf {t}} _ {1 \rightarrow 2} \cdot \hat {\mathbf {t}} _ {1 \rightarrow 2}}, s _ {1 \rightarrow 3} = \frac {\hat {\mathbf {t}} _ {1 \rightarrow 3} \cdot \mathbf {t} _ {1 \rightarrow 3} ^ {\prime}}{\hat {\mathbf {t}} _ {1 \rightarrow 3} \cdot \hat {\mathbf {t}} _ {1 \rightarrow 3}}, +$$ + +where $\hat{\mathbf{t}}_{i\rightarrow j}$ is the TFT predicted translation (up to a scale) from camera $i$ to camera $j$ , $\mathbf{t}_{i\rightarrow j}^{\prime}$ is the marker-calibrated translation in real metric scale. We set the extrinsics of the first camera to be the identity matrix. We use Intel's dynamic calibration tool to calibrate intrinsics. In cases of repetitive textures, where SuperPoint+SuperGlue were unreliable, we made correspondences manually. + +# 3.3 Annotation of Pose Ground-Truths + +Previous works have used ArUco markers to automate pose estimation through Perspective-n-Points, and we do this also. Specifically, we used ArUco markers to automate the annotation of still object poses (i.e. not moving), following LINEMOD [3] and NOCS [34]. We note that this step has two limitations: + +1. Markers in the scene detract from realism and compromise dataset integrity: training on marker-augmented imagery may result in overfitting to these artificial patterns. +2. Marker-based annotations are inapplicable to moving objects (since the markers are typically placed on the background), thus limiting the setup to annotating still objects. + +We compensate for both of these issues: we remove marker appearances from the dataset (after using them for annotation), and we introduce additional tools to annotate the moving objects. We describe these compensation steps next. + +Marker removal We removed the marker appearances from the dataset, using a marker inpainting strategy, detailed as follows. In step 1, we place a marker (Marker 1) somewhere within the camera's field of view. We then record a short video with this marker in view. In step 2, we place a second marker (Marker 2) at a chosen distance from the first, and record another video. In step 3, we remove Marker 1, and begin the actual object capture process (with only Marker 2 present). After this process, we end up with: (1) frames with Marker 1 only, which clearly depict the surface where Marker 2 will later appear; (2) frames capturing both markers, providing helpful calibration cues; (3) frames with Marker 2 only, which represent our main capture. We leverage Marker 2 for automated pose tracking, and manually correct the tracking every 40 frames in case of drift. We use the frames from the first two steps to seamlessly inpaint [25] Marker 2's area for the final dataset, as depicted in Figure 4. + +![](images/f34949e0cd9adf9c22dc9d477a3416bbc3c61039963a8655e9afb27a0d19f5a4.jpg) +Fig. 4: Illustration of the marker inpainting process. + +Annotation of Moving Object Positions For moving objects, the traditional marker-based tracking approach is insufficient, as the scene markers do not move in tandem with the objects. In such instances, we harnessed the capabilities of BundleTrack [35], an advanced RGB image-based tracking algorithm. BundleTrack conducts feature correspondence analysis between successive frames to estimate poses, complemented by a bundle adjustment algorithm to optimize keyframes globally and minimize tracking errors. Despite BundleTrack's proficiency in approximating poses, it is prone to drift, necessitating the manual adjustment of poses every ten frames to ensure precision. We adopted a rigorous manual checking process, examining every tenth frame for 3D pointcloud alignment and 2D multi-view agreement. We detail the annotation tool and its + +graphical interface in the supplementary. This delicate process represents the most labor-intensive aspect of our annotation pipeline. + +In order to verify that our annotation process gives accurate poses, we conducted an experiment where we used our setup to manually annotate the poses of Scene 1 from NOCS-REAL275 [34] dataset, and compare the annotations to the provided ground truths. We recorded an average pose annotation error of $0.9^{\circ}$ for rotation and $2.3\mathrm{mm}$ for translation. + +Generation of Segmentation Masks Upon successful pose annotation, we generated occlusion-aware segmentation masks, via depth rendering. In cases where hand interactions are involved, we employed the segmentation model SAM [18] to estimate hand masks, and subsequently subtracted these from the previously computed object masks to obtain the occlusion-aware segmentation. + +# 4 Dataset Statistics + +In pursuit of diversity, we placed the objects in ten disparate environments, each featuring varying levels and configurations of occlusion. We captured 10 video sequences per scene, each sequence including 1 to 7 objects. We captured RGB and depth data using an Intel RealSense D415 camera, with a resolution of $1280 \times 720$ . The capture process spans distances ranging from $0.5\mathrm{m}$ to $1.5\mathrm{m}$ . + +# 4.1 Annotation Distribution and Object Diversity + +The comprehensive distribution of pose annotation counts is depicted in Figure 5, showcasing a diverse array including both rigid and articulated objects. As demonstrated in Figure 7, our object database encompasses a broad spectrum of object sizes, with the majority measuring near $0.2\mathrm{m}$ along the bounding box diagonal. It also includes larger items such as storage bins, adding to the heterogeneity of the dataset. + +# 4.2 Variability in Occlusion, Pose, and Environmental Context + +The dataset has rich variability in occlusion levels and object poses, as illustrated in Figure 6, which outlines the statistical distribution of both visibility ratios and azimuth/elevation angles, indicating a comprehensive coverage reflecting real-world scenarios. These statistics can be helpful for assessing the robustness of pose estimation algorithms. Besides, the ten distinct capturing environments are shown in Figure 8, illustrating diverse backgrounds and lighting conditions. + +# 4.3 Simulated Training Data and Dataset Split + +To facilitate a comprehensive and equitable evaluation of pose estimation methodologies, we partitioned our entire real-world dataset into validation and test subsets following a $20/80$ ratio. The test subset is used for both instance-level and + +![](images/032379076a8233035dd5607347c922c1ad5bc5145350c1aa942ed736f32959f1.jpg) +Fig. 5: Distribution of pose annotation counts across different object categories. + +category-level pose estimation evaluation tasks. The real-world validation subset is used for instance-level pose estimation only. We render another photo-realistic validation subset for category-level pose estimation, in order to cover enough category-level objects that are different from the test set. + +Additionally, to support research necessitating extensive training datasets, we generated a large synthetic dataset called PACE-Sim that contains two different training subsets, for instance-level and category-level pose estimation, respectively. We use a physically based renderer [6] to ensure the images are highly photo-realistic. The synthetic training sets consist of 100K images, each featuring dozens of objects, yielding 2.4M annotations. The distribution of object instances in both PACE and PACE-Sim can be found in supplementary. + +# 5 Evaluation Benchmarks + +We assess both pose estimation and tracking. State-of-the-art (SOTA) pose estimation techniques generally involve two phases: first detecting or segmenting the object, and then estimating its pose within the identified mask. For a fair comparison, we evaluate pose estimation assuming perfect object detection. Detailed SOTA detection performance is available in the supplementary materials but is beyond this dataset's primary scope. + +For pose tracking, we provide ground-truth data in the initial frame and require methods to accurately track the object's rotations and translations through the video sequence. Both model-free and model-based trackers are assessed. + +# 5.1 Pose Estimation Benchmark + +This benchmark is divided into instance-level and category-level pose estimation. The former concerns object instances seen during training (in different scenes), + +![](images/5f77a48e05238c0c1bd36df306fae288707a230de6618eddf3046e79edad0f55.jpg) + +![](images/f6a81fe12a22692ed36cb230cefc7be97e55e0ef936f4248533b98d639d44f96.jpg) +Fig. 6: Distribution of occlusions, azimuth and elevation viewing angles. + +![](images/128b405aacc842d5b5dacb25428b3949349e0dadcdce381532c2f0b8eeaa7a04.jpg) +Fig. 7: Distribution of object sizes. + +![](images/e68397aeb3544284f961a7e34c973659c7dd2e127272bbf378c282a35ac53607.jpg) + +![](images/77172050f348784b0bccf84914bccbe018dcfb8fdf6dafbecd53b082d8355988.jpg) +(a) basket + +![](images/6ab287369884ee13589e1c34d5694c61f4f0e1c8f2241ca629ff07c29da2fea8.jpg) + +![](images/db28857fbb2811f8ace59a39ac602e70164ac4d36c63ee3c0fd676e8208b1141.jpg) + +![](images/3989cae8268163364a378d396a4b1bacdbf1b2d45547e274267ed89dfc434967.jpg) + +![](images/69b6ae908aa0d229ee469777bf56494c15eda96ed4b5a8987c52d352dbc25102.jpg) + +![](images/7577a49491983fb703e7a9633429086acfdacba2a52f38814b53a710d4b3c056.jpg) +(f) china floor + +![](images/ad7a091475671192b9d3591a1dea85bd027f7bde6db1b590ae6a829380f4fda4.jpg) +(b) box top +(g) desk1 +Fig. 8: Sample images from the 10 environments. + +![](images/d255b6ee5f955337beeb5175ced43327c2ef62e8bc01c1cdb9cf31a5fa326519.jpg) +(c) cabinet +(h) desk2 + +![](images/d461684a4c36d427002b7ec2c948581676618f306e7c186b5755ba770c98fdfd.jpg) +(d) carpet +(i) stairs + +![](images/414f78905a4253b1f53d17be0e720ea3fdd274f2bb2c0161c685f34ee5b4772b.jpg) +(e) chair +(j) wooden floor + +while the latter concerns instances which were not seen but come from the same category. We will also give an in-depth analysis of SOTA's failure modes. + +Instance-Level Pose Estimation This task demands the prediction of rotation and translation for known instances from the training set. + +Metrics: Adhering to the BOP challenge protocol [14], we use Average Recall (AR) of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) as our metrics. Detailed computation of these metrics is described in [14]. Results are averaged across all instances. + +Baselines: We assessed four baselines: PPF [8], CosyPose [19], SurfEmb [12], and GDRNPP [33]. While PPF does not require training data, the others are SOTA methods dependent on additional training on the PBR dataset. To focus on pose estimation performance, we assume that ground-truth detections are available. + +Result Analysis: The quantitative analysis is presented in Table 2. Notably, while SOTA methods excel on the BOP benchmarks, they do not perform as well as PPF, which relies on local geometric feature matching in point clouds. The relative success of PPF highlights the potential for improvement in SOTA methods, particularly regarding sim-to-real ability and scalability to handle a large set of instances. Qualitative results can be found in the supplementary. + +
Modality DetectionARVSD↑ARMSSD↑ARMSPD↑AR↑
PPF [8]DG.T.53.448.155.252.2
CosyPose [19]RGBG.T.1.40.311.54.4
SurfEmb [12]RGBG.T.6.23.017.89.0
GDRNPP [33]RGB-DG.T.3.62.115.47.0
+ +Table 2: Instance-level pose estimation results showcasing the robustness of PPF and the potential for improvement in SOTA deep learning-based methods. + +
DetectionIoU25↑IoU50↑AP
0:20°↑0:60°↑0:5cm↑0:15cm↑0:20°↑0:60°↑
NOCS [34]Mask-RCNN0.2/0.00.0/0.01.2/0.04.3/0.017.0/0.033.5/0.01.1/0.04.2/0.0
HS-Pose [43]G.T.36.6/0.02.7/0.05.3/0.08.6/0.348.7/34.181.0/75.63.9/0.08.1/0.3
SGPA [4]G.T.2.6/0.01.2/0.05.6/1.111.4/7.316.1/6.850.7/17.53.3/0.910.1/7.2
DualPoseNet [22]G.T.24.2/0.30.1/0.05.5/0.08.1/0.018.6/27.363.4/65.21.8/0.06.5/0.0
SAR-Net [21]G.T.22.8/0.10.2/0.05.3/0.08.8/0.848.6/38.579.8/69.73.7/0.08.2/0.8
CPPF++ [39]G.T.44.5/3.64.4/0.015.2/1.727.3/6.235.3/31.174.0/66.79.9/1.124.9/5.9
ANCSH [20]G.T.- / 0.0- / 0.0- / 0.0- / 0.2- / 18.6- / 50.4- / 0.0- / 0.2
+ +Table 3: Category-level pose estimation benchmark combining the performance metrics for both rigid and articulated object pose estimation, separated by slash. + +Category-Level Pose Estimation Category-level pose estimation tasks involve predicting the 3D bounding box dimensions, rotation, and translation of target instances, given only the category information a priori. + +Metrics: We adopted from NOCS [34], calculating mean Average Precision (AP) across predefined angle and translation thresholds. Specifically, AP@0:20° represents AP averaged from $0^{\circ}$ to $20^{\circ}$ at $1^{\circ}$ intervals; AP@0:5cm is averaged from 0cm to 5cm at 0.25cm intervals; AP@0:20°,0:5cm combines these angle and translation thresholds. IoU $_{25}$ and IoU $_{50}$ denote AP for 3D bounding box matches at IoU thresholds of 25 and 50, respectively. Results are separately reported for both rigid and articulated objects. + +Baselines: We evaluated seven recent category-level pose estimation methods: NOCS [34], HS-Pose [43], SGPA [4], DualPoseNet [22], SAR-Net [21], CPPF++ [39] and ANCSH [20], the latter specifically designed for articulated objects. For adapting rigid-object-focused methods to articulated objects, each movable part is considered a distinct category. Ground-truth detections are presumed for each method, except for NOCS, which outputs both instance masks and pose estimations using a unified network. + +Result Analysis: Quantitative comparisons are listed in Table 3. On rigid objects, HS-Pose and SAR-Net show fair performance on AP metrics for translation, but falls short on rotation AP metrics compared to CPPF++, suggesting a proficiency in size and translation prediction but not rotation. CPPF++ shows the best performance but is weaker on translation estimation. For articulated objects, all methods face challenges due to the movable parts, with ANCSH underperforming on the large-scale real-world dataset, indicative of a significant sim-to-real gap. NOCS shows instability, likely due to its inaccurate mask prediction. Qualitative results are available in supplementary. + +Failure Mode Analysis This section delves into the analysis of two critical questions that underline the failure of SOTA deep learning-based methods in pose estimation: 1) Is the failure attributed to the sim-to-real gap, considering the exclusive use of synthetic RGB-D images for training? 2) Does the failure stem from the inherent large-scale complexity of the PACE dataset, which poses a significant challenge for baseline models? + +Ablation Setup: To explore these questions, we introduced a new $80/10/10$ train/validation/test split for the real-world data, diverging from the $0/20/80$ split described in Section 4.3. We then examined the following settings: + +Real2Real Single Instance/Category Fitting: This basic setup involves training on a single instance or category selected (we randomly select the category can and sample a can with object ID 57) from the $80 / 10 / 10$ split of real data. +- Sim2Real Single Instance/Category Fitting: This scenario differs from the first by utilizing all simulated data for training and validation, with evaluation on the same real-world test split. +- Real2Real All Instance/Category Fitting: Unlike the first setting, this involves training, validation, and testing across all instances or categories. + +Result Analysis and Discussions: Our analysis in Table 4 reveals that baseline models effectively handle single instance or category fitting under the Real2Real Single Instance/Category Fitting setting, provided the training and test distributions align. However, the sim-to-real gap, especially noticeable in depth-input methods like DualPoseNet, HS-Pose, and SGPA, highlights significant challenges, as these methods struggle with the depth noise in real world. + +
Training SettingInstance-Level (AR%)†Category-Level (mean AP%@0:60°, 0:15cm)†
CosyPoseSurfEmbGDRNPPNOCSHS-PoseSGPADualPoseNetSAR-NetCPPF++
Real2Real Single Inst./Cat.76.486.788.276.981.286.758.659.383.4
Sim2Real Single Inst./Cat.67.1 (-9.3↓)73.7 (-13.0↓)78.1 (-10.1↓)55.8 (-21.1↓)1.1 (-80.1↓)0.3 (-86.4↓)1.4 (-57.2↓)33.2 (-26.1↓)55.8 (-27.6↓)
Real2Real All Inst./Cat.8.2 (-68.2↓)11.2 (-75.5↓)9.7 (-78.5↓)0.3 (-76.6↓)5.6 (-75.6↓)9.8 (-76.9↓)1.2 (-57.4↓)2.1 (-57.2↓)13.7 (-69.7↓)
+ +Additionally, learning across all instances/categories leads to a notable performance drop, as shown in Table 4. This drop, together with suboptimal training data performance and loss plateauing, suggests underfitting and highlights the complexity posed by the PACE dataset's diversity. Innovative approaches or larger foundational models might be necessary to address these challenges and effectively bridge the sim-to-real gap. + +# 5.2 Pose Tracking Benchmark + +We distinguish SOTA pose tracking techniques into model-based, utilizing a 3D CAD model, and model-free, which only require an initial pose from first frame. + +Metrics: For model-based tracking, we measure the area under curve (AUC) for ADD, ADD-S [38], and ADD(-S) metrics, with a maximum threshold of $0.1\mathrm{m}$ [32]. The ADD metric, introduced in [13], computes the average per-point distance between point clouds transformed by the predicted and ground-truth poses. ADD-S accounts for point correspondence ambiguities in symmetric objects. + +For model-free tracking, we adopt four metrics from prior studies [31]: 1) $5^{\circ}5\mathrm{cm}$ , indicating the percentage of predictions within $5^{\circ}$ rotation and $5\mathrm{cm}$ translation errors. 2) IoU25, the percentage of instances where the intersection over union of the 3D bounding boxes exceeds $25\%$ . 3) $\mathbf{R}_{\mathrm{err}}$ and 4) $\mathbf{T}_{\mathrm{err}}$ , the mean rotational and translational errors in degrees and centimeters, respectively, only considering instances with IoU $>25\%$ . We exclude objects with visibility under $10\%$ as per the BOP convention [14]. Results are averaged separately across rigid and articulated object categories. + +Table 4: Ablation Study: This study illustrates the challenges models face when transitioning from real-to-real settings. It showcases performance disparities when scaling from single instances/categories to the entire dataset, emphasizing the scalability challenge from PACE and the profound impact of sim-to-real discrepancies on depth-dependent methodologies. + +
ModalityADD↑ADD-S↑ADD(-S)↑
RBOT [30]RGB7.1/0.510.3/0.87.4/0.5
ICG [28]RGB-D35.6/10.148.1/15.238.1/10.1
+ +Table 5: Model-Based Pose Tracking. We report the area under curve (AUC) with respect to ADD, ADD-S and ADD(-S). The higher value, the better performance. + +
Training-Free Modality5°5cm↑IoU25↑Rerr↓Terr↓
BundleTrack [35]✓RGB6.4/11.29.1/14.13.2/5.52.6/0.8
CAPTRA [37]✗D12.9/4.445.8/20.619.2/40.92.2/1.5
6-PACK [31]✗RGB-D9.2/3.923.1/16.717.7/33.62.1/1.2
+ +Table 6: Model-Free Pose Tracking. Both rigid and articulated results are reported. For $5^{\circ}5\mathrm{cm}$ and IoU25, the higher value means better performance while for $\mathbf{R}_{err}$ and $\mathrm{T_{err}}$ , the situation is reversed. Note that $\mathrm{R}_{err}$ and $\mathrm{T_{err}}$ are respect to IoU25 since objects with IoU $\leq 25\%$ are not counted. + +Baselines: For tracking methods using object masks, we applied the ground-truth masks directly. Except for CAPTRA, which also predicts the 3D bounding box size, all methods estimate only rotation and translation; however, we used ground-truth bounding box sizes for consistency in evaluations. Unlike other methods that treat each part of an articulated object independently, CAPTRA considers the entire object as one entity, trained on our synthetic PBR data. + +Result and Failure Mode Analysis: Tables 5 and 6 display the performance of state-of-the-art (SOTA) tracking methods on the PACE dataset, highlighting both rigid and articulated object results. The highest achieved ADD(-S) for model-based methods is only $38.1\%$ for rigid and $10.1\%$ for articulated objects by ICG. Furthermore, Table 6 reveals severe limitations of model-free methods, with the best $5^{\circ}5\mathrm{cm}$ accuracy under $13\%$ , posing major tracking challenges. + +The convention followed by 6-PACK [31] and BundleTrack [35] disregards objects with $\mathrm{IoU} \leq 25\%$ when computing rotational and translational errors. However, this approach could overestimate the performance on $\mathrm{R}_{err}$ and $\mathrm{T}_{err}$ . BundleTrack [35] reports lower performance in $5^{\circ}5\mathrm{cm}$ and IoU25 metrics yet shows seemingly better results for $\mathrm{R}_{err}$ and $\mathrm{T}_{err}$ , suggesting that it can only track well within a threshold but cannot recover once the object is lost. Qualitative results can be found in the supplementary. + +# 6 Conclusions and Limitations + +We present PACE, a new large-scale benchmark for 3D object pose estimation and tracking, designed with various occlusion levels across ten diverse environments. This benchmark both reflects prior progress and urges further research to tackle real-world complexities. Our analysis indicates significant advances in pose estimation, yet underscores a notable real-world performance gap, particularly with articulated objects and in terms of robustness and scalability. Future efforts could focus on advanced sim-to-real techniques and developing larger models that capture a broader spectrum of object features and environments. + +Limitations Due to resource limitations, PACE lacks large and valuable objects like tables, cameras, and laptops, which may affect its breadth in representing real-world object scenarios. + +# 7 Acknowledgements + +This work is supported by the National Key Research and Development Project of China (No. 2022ZD0160102), National Key Research and Development Project of China (No. 2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants. Yang You is also supported in part by the Outstanding Doctoral Graduates Development Scholarship of Shanghai Jiao Tong University. + +# References + +1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7822-7831 (2021) +2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2614-2623 (2019) +3. Brachmann, E., Krull, A., Michel, F., Gumhold, S., Shotton, J., Rother, C.: Learning 6d object pose estimation using 3d object coordinates. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 536-551. Springer (2014) +4. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021) +5. Deng, X., Mousavian, A., Xiang, Y., Xia, F., Bretl, T., Fox, D.: Poserbpf: A raobackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics 37(5), 1328-1342 (2021) +6. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901 +7. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 224-236 (2018) +8. Drost, B., Ulrich, M., Navab, N., Ilic, S.: Model globally, match locally: Efficient and robust 3d object recognition. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 998-1005. IEEE (2010) +9. Garon, M., Lalonde, J.F.: Deep 6-dof tracking. IEEE transactions on visualization and computer graphics 23(11), 2410-2418 (2017) +0. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition 47(6), 2280-2292 (2014) +1. Guo, A., Wen, B., Yuan, J., OTHERS: Handal: A dataset of real-world manipulable object categories with pose annotations, affordances, and reconstructions (2023), https://arxiv.org/pdf/2308.01477.pdf +2. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022) + +13. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. pp. 548-562. Springer (2013) +14. Hodaţ, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 577-594. Springer (2020) +15. Jampani, V., Maninis, K.K., Engelhardt, A., Karpur, A., Truong, K., Sargent, K., Popov, S., Araujo, A., Martin-Brualla, R., Patel, K., et al.: Navi: Category-agnostic image collections with high-quality 3d shape and pose annotations. arXiv preprint arXiv:2306.09109 (2023) +16. Julià, L.F., Monasse, P.: A critical review of the trifocal tensor estimation. In: Image and Video Technology: 8th Pacific-Rim Symposium, PSIVT 2017, Wuhan, China, November 20-24, 2017, Revised Selected Papers 8. pp. 337-349. Springer (2018) +17. Jung, H., Wu, S.C., Ruhkamp, P., Zhai, G., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., et al.: Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22498-22508 (2024) +18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +19. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: Cosypose: Consistent multi-view multi-object 6d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 574-591. Springer (2020) +20. Li, X., Wang, H., Yi, L., Guibas, L., Abbott, A.L., Song, S.: Category-level articulated object pose estimation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020) +21. Lin, H., Liu, Z., Cheang, C., Fu, Y., Guo, G., Xue, X.: Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6707-6717 (2022) +22. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3560-3569 (2021) +23. Liu, L., Xu, W., Fu, H., Qian, S., Yu, Q., Han, Y., Lu, C.: Akb-48: A real-world articulated object knowledge base. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14809-14818 (2022) +24. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10870-10879 (2021) +25. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023) +26. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4938-4947 (2020) + +27. Stoiber, M., Elsayed, M., Reichert, A.E., Steidle, F., Lee, D., Triebel, R.: Fusing visual appearance and geometry for multi-modality 6dof object tracking. arXiv preprint arXiv:2302.11458 (2023) +28. Stoiber, M., Sundermeyer, M., Triebel, R.: Iterative corresponding geometry: Fusing region and depth for highly efficient 3d tracking of textureless objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6855-6865 (2022) +29. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2974-2983 (2018) +30. Tjaden, H., Schwanecke, U., Schömer, E., Cremers, D.: A region-based gauss-newton approach to real-time monocular multiple object tracking. IEEE transactions on pattern analysis and machine intelligence 41(8), 1797-1812 (2018) +31. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 10059-10066. IEEE (2020) +32. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019) +33. Wang, G., Manhardt, F., Tombari, F., Ji, X.: GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16611-16621 (June 2021) +34. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2642-2651 (2019) +35. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8067-8074. IEEE (2021) +36. Wen, B., Mitash, C., Ren, B., Bekris, K.E.: se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10367-10373. IEEE (2020) +37. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13209-13218 (2021) +38. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017) +39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. arXiv preprint arXiv:2211.13398 (2022) +40. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022) + +41. Zhang, J., Huang, W., Peng, B., Wu, M., Hu, F., Chen, Z., Zhao, B., Dong, H.: Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. arXiv preprint arXiv:2406.04316 (2024) +42. Zhang, J., Wu, M., Dong, H.: Genpose: Generative category-level object pose estimation via diffusion models. arXiv preprint arXiv:2306.10531 (2023) +43. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-pose: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17163-17173 (2023) \ No newline at end of file diff --git a/2024/PACE_ Pose Annotations in Cluttered Environments/images.zip b/2024/PACE_ Pose Annotations in Cluttered Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9043f13484cbc191d43ecaa2b5be199831f00ccd --- /dev/null +++ b/2024/PACE_ Pose Annotations in Cluttered Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a29d7465470a3558af61a00930c129858ea591edaecc63db37c891ab59a7ace5 +size 494058 diff --git a/2024/PACE_ Pose Annotations in Cluttered Environments/layout.json b/2024/PACE_ Pose Annotations in Cluttered Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f8e8a871b2f9526aa40dd3af66ae162caa1212db --- /dev/null +++ b/2024/PACE_ Pose Annotations in Cluttered Environments/layout.json @@ -0,0 +1,9451 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 112, + 449, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 112, + 449, + 146 + ], + "spans": [ + { + "bbox": [ + 165, + 112, + 449, + 146 + ], + "type": "text", + "content": "PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "spans": [ + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "text", + "content": "Yang You" + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "text", + "content": " Kai Xiong" + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "text", + "content": " Zhening Yang" + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "text", + "content": " Zhengxiang Huang" + }, + { + "bbox": [ + 162, + 167, + 452, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 217, + 180, + 396, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 180, + 396, + 193 + ], + "spans": [ + { + "bbox": [ + 217, + 180, + 396, + 193 + ], + "type": "text", + "content": "Junwei Zhou² Ruoxi Shi⁴ Zhou Fang²" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "spans": [ + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "type": "text", + "content": "Adam W. Harley" + }, + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "type": "text", + "content": " Leonidas Guibas" + }, + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "type": "text", + "content": " Cewu Lu" + }, + { + "bbox": [ + 187, + 193, + 425, + 204 + ], + "type": "inline_equation", + "content": "^{2\\star}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 213, + 352, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 213, + 352, + 225 + ], + "spans": [ + { + "bbox": [ + 261, + 213, + 352, + 225 + ], + "type": "text", + "content": "1 Stanford University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 225, + 374, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 225, + 374, + 236 + ], + "spans": [ + { + "bbox": [ + 239, + 225, + 374, + 236 + ], + "type": "text", + "content": "2 Shanghai Jiao Tong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 258, + 236, + 356, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 236, + 356, + 246 + ], + "spans": [ + { + "bbox": [ + 258, + 236, + 356, + 246 + ], + "type": "text", + "content": "3 Horizon Robotics Inc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 272, + 247, + 341, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 247, + 341, + 258 + ], + "spans": [ + { + "bbox": [ + 272, + 247, + 341, + 258 + ], + "type": "text", + "content": "UC San Diego" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 134, + 283, + 202, + 323 + ], + "blocks": [ + { + "bbox": [ + 134, + 283, + 202, + 323 + ], + "lines": [ + { + "bbox": [ + 134, + 283, + 202, + 323 + ], + "spans": [ + { + "bbox": [ + 134, + 283, + 202, + 323 + ], + "type": "image", + "image_path": "709319356a84f83ce0ded0ee55b0d9d33032ba0e8c81825c61ca5ce0bb79cb41.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 134, + 324, + 201, + 364 + ], + "blocks": [ + { + "bbox": [ + 134, + 324, + 201, + 364 + ], + "lines": [ + { + "bbox": [ + 134, + 324, + 201, + 364 + ], + "spans": [ + { + "bbox": [ + 134, + 324, + 201, + 364 + ], + "type": "image", + "image_path": "582e21ec12be91a48836679f23f4bc568f7da8b23de0c6cee9c97e0e29dabacc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 373, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 131, + 373, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 482, + 396 + ], + "type": "text", + "content": "Fig. 1: We propose PACE: a large-scale object pose dataset, with diverse objects, complex scenes, and various types of occlusions, reflecting real-world challenges." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 203, + 284, + 271, + 323 + ], + "blocks": [ + { + "bbox": [ + 203, + 284, + 271, + 323 + ], + "lines": [ + { + "bbox": [ + 203, + 284, + 271, + 323 + ], + "spans": [ + { + "bbox": [ + 203, + 284, + 271, + 323 + ], + "type": "image", + "image_path": "3da2842c4be018ab84f33186bbe51cd77c96cf30cd79c2e846242e9e50ee0397.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 203, + 325, + 271, + 364 + ], + "blocks": [ + { + "bbox": [ + 203, + 325, + 271, + 364 + ], + "lines": [ + { + "bbox": [ + 203, + 325, + 271, + 364 + ], + "spans": [ + { + "bbox": [ + 203, + 325, + 271, + 364 + ], + "type": "image", + "image_path": "ad871cf0ebf1b0fa973cc0a89380acaa8bfcd8ab8309b7a747d4ba7d919677bb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 272, + 284, + 340, + 323 + ], + "blocks": [ + { + "bbox": [ + 272, + 284, + 340, + 323 + ], + "lines": [ + { + "bbox": [ + 272, + 284, + 340, + 323 + ], + "spans": [ + { + "bbox": [ + 272, + 284, + 340, + 323 + ], + "type": "image", + "image_path": "01c4e6e6a4833910bed75a900a139974904b68085bf98d60b01b67bcfce7ddb4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 272, + 325, + 340, + 364 + ], + "blocks": [ + { + "bbox": [ + 272, + 325, + 340, + 364 + ], + "lines": [ + { + "bbox": [ + 272, + 325, + 340, + 364 + ], + "spans": [ + { + "bbox": [ + 272, + 325, + 340, + 364 + ], + "type": "image", + "image_path": "5d4c03167f2156c6ccf163d36d6d548c006270afaa1dbb0d799395da2d4ff3e8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 342, + 284, + 410, + 323 + ], + "blocks": [ + { + "bbox": [ + 342, + 284, + 410, + 323 + ], + "lines": [ + { + "bbox": [ + 342, + 284, + 410, + 323 + ], + "spans": [ + { + "bbox": [ + 342, + 284, + 410, + 323 + ], + "type": "image", + "image_path": "02b4d783904db49da959e93fac7a08b08792064619b79a2b98b334bb39c3e59c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 342, + 325, + 410, + 364 + ], + "blocks": [ + { + "bbox": [ + 342, + 325, + 410, + 364 + ], + "lines": [ + { + "bbox": [ + 342, + 325, + 410, + 364 + ], + "spans": [ + { + "bbox": [ + 342, + 325, + 410, + 364 + ], + "type": "image", + "image_path": "2b1bc824ed10b2d585c91f30ad951f2f616004511f789de50be521995e6c3610.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 411, + 284, + 479, + 323 + ], + "blocks": [ + { + "bbox": [ + 411, + 284, + 479, + 323 + ], + "lines": [ + { + "bbox": [ + 411, + 284, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 411, + 284, + 479, + 323 + ], + "type": "image", + "image_path": "dec61119fa07d701745b1d54a945cc6bee9fea09d6d4550f6400b3d3dabea163.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 411, + 325, + 479, + 364 + ], + "blocks": [ + { + "bbox": [ + 411, + 325, + 479, + 364 + ], + "lines": [ + { + "bbox": [ + 411, + 325, + 479, + 364 + ], + "spans": [ + { + "bbox": [ + 411, + 325, + 479, + 364 + ], + "type": "image", + "image_path": "5e9e474a3d2d627184d9d2a02db6866a1c7b5e62fddcd350623d85924a20a37d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 159, + 432, + 455, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 432, + 455, + 587 + ], + "spans": [ + { + "bbox": [ + 159, + 432, + 455, + 587 + ], + "type": "text", + "content": "Abstract. We introduce PACE (Pose Annotations in Cluttered Environments), a large-scale benchmark designed to advance the development and evaluation of pose estimation methods in cluttered scenarios. PACE provides a large-scale real-world benchmark for both instance-level and category-level settings. The benchmark consists of 55K frames with 258K annotations across 300 videos, covering 238 objects from 43 categories and featuring a mix of rigid and articulated items in cluttered scenes. To annotate the real-world data efficiently, we develop an innovative annotation system with a calibrated 3-camera setup. Additionally, we offer PACE-Sim, which contains 100K photo-realistic simulated frames with 2.4M annotations across 931 objects. We test state-of-the-art algorithms in PACE along two tracks: pose estimation, and object pose tracking, revealing the benchmark's challenges and research opportunities. Our benchmark code and data is available on https://github.com/qq456cvb/PACE." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "type": "text", + "content": "The field of 3D object pose estimation is integral to a myriad of applications, particularly within robotic manipulation. Recent advancements in both instance and" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 457, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 457, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 457, + 666 + ], + "type": "text", + "content": "* Corresponding authors. Email: yangyou@stanford.edu, lucewu@sjtu.edu.cn" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 140 + ], + "type": "text", + "content": "category-level pose estimation have been bolstered by deep learning approaches, and perhaps more importantly, by data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 200 + ], + "type": "text", + "content": "PoseCNN [38] advanced pose estimation into the deep learning era, and simultaneously introduced the influential YCB-Video dataset. This benchmark has catalyzed methodological development and offered a consistent evaluation platform. Besides, the Benchmark for 6D Object Pose Estimation (BOP) challenges [14] have consolidated datasets metrics for instance-level pose estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "type": "text", + "content": "In parallel, NOCS [34] has addressed category-level pose estimation, albeit with a smaller real-world dataset for validation and testing. Despite these strides, the field grapples with a fundamental challenge: state-of-the-art pose estimation models still perform poorly in real-world settings, and existing evaluation datasets are too constrained to thoroughly reveal this fact. The NOCS-REAL275 dataset, for instance, spans only six categories and includes a mere 18 videos. This limitation has led to an array of published models which perform near-perfectly on NOCS but do not generalize to new data, as demonstrated in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 316, + 308, + 392 + ], + "blocks": [ + { + "bbox": [ + 133, + 316, + 308, + 392 + ], + "lines": [ + { + "bbox": [ + 133, + 316, + 308, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 316, + 308, + 392 + ], + "type": "image", + "image_path": "c20085b68f4ec9a69863169cd08c64499d50d73275af2b321dfbd7855a3b3f31.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 401, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 468 + ], + "type": "text", + "content": "Fig. 2: While current state-of-the-art methods yield satisfactory outcomes on the NOCS-REAL275 dataset, their models' performance significantly deteriorates when transferred to previously unseen datasets such as PACE. Left: Qualitative visualizations of various models' pose predictions for a mug in REAL275 vs. a mug in PACE. Right: The performance of state-of-the-art methods markedly declines on PACE, even when evaluating on categories that exist in both datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 315, + 480, + 393 + ], + "blocks": [ + { + "bbox": [ + 321, + 315, + 480, + 393 + ], + "lines": [ + { + "bbox": [ + 321, + 315, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 321, + 315, + 480, + 393 + ], + "type": "table", + "html": "
DatasetMethodAP@15°5cm (%)↑
bottlebowlcanmug
REAL275 [34]HS-Pose [43]99.899.899.586.2
DualPoseNet [22]97.599.797.768.7
SARNet [21]98.298.297.470.6
PACE (Ours)HS-Pose [43]0.40.00.87.5
DualPoseNet [22]0.32.00.51.6
SARNet [21]2.10.00.30.0
", + "image_path": "57796ce431ca4423fb9a27c602025c1c641f5b941092e428f786298afb86b7bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 481, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 526 + ], + "type": "text", + "content": "In this work, we introduce PACE, a benchmark for pose estimation, and present a comprehensive study evaluating a wide range of pose estimation and tracking methods. Our contributions are threefold:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 534, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 137, + 534, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 534, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 137, + 534, + 482, + 605 + ], + "type": "text", + "content": "- The PACE dataset: This dataset includes 238 objects across 43 categories, captured in 300 video clips within diverse scenes. With an average of 183 frames per clip, the dataset consists of 55K frames and 258K annotations, providing a large-scale benchmark for pose estimation. We also provide PACESim, a photo-realistic training dataset which contains 100K frames with 2.4M annotations across 931 objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 605, + 482, + 665 + ], + "type": "text", + "content": "- Our evaluation study: To the best of our knowledge, we are the first to analyze and report the performance of state-of-the-art (SOTA) pose estimation methods in a large-scale cluttered setting. These results provide valuable insights on the scalability and generalizability of SOTA methods, making it clear that they are far from reliable in general." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 482, + 175 + ], + "type": "text", + "content": "- Our annotation pipeline: We will open-source our annotation pipeline, which uses a calibrated 3-camera system, enhancing the precision and scalability of annotating poses in real data. This tool significantly mitigates human error and reduces the effort required for annotating 3D poses, providing a solution to one of the major bottlenecks in pose dataset creation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 185, + 481, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 481, + 221 + ], + "type": "text", + "content": "Overall, our work aims to support the development of more robust and generalizable pose estimation techniques, thereby facilitating progress towards successful pose estimation in the real world." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 239, + 242, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 242, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 242, + 251 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 263, + 481, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 481, + 300 + ], + "type": "text", + "content": "The field of 3D object pose estimation has seen substantial progress over the past few years. This progress has been facilitated by the introduction of standardized datasets and the development of innovative algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 316, + 269, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 269, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 269, + 329 + ], + "type": "text", + "content": "2.1 Object Pose Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 336, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 433 + ], + "type": "text", + "content": "Instance-Level Pose Datasets YCB-Video dataset [38] is a comprehensive resource for 6D object pose estimation, containing a large number of video frames with accurate pose annotations for 21 objects. LINEMOD-Occluded [3] offers a challenging setting for pose estimation with piled multiple objects in occluded scenes. NAVI dataset [15] presents casually captured images of objects with high-quality 3D scans and precise 2D-3D alignments for 3D reconstruction tasks. Recently, StereoObj-1M [24] is proposed to address challenging cases such as object transparency, in addition to common challenges of occlusion and symmetry." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": "Category-Level Pose Datasets NOCS-REAL275 [34] and Wild6D [40] both annotate a limited number of categories with upright poses. Objectron [1] is a collection of short, single-object-centric video clips with no clutters. The Scan2CAD [2] dataset aligns 14225 CAD models from ShapeNet to 1506 ScanNet scans, promoting CAD model alignment in RGB-D scans. Pix3D [29] is a benchmark with image-shape pairs and pixel-level 2D-3D alignment, aiding in shape reconstruction and retrieval. HANDAL [11] focuses on pose estimation and affordance prediction for robotics-ready manipulable objects. HouseCat6D [17] encompasses 194 diverse objects across 10 household categories with additional grasp annotations. ROPE [41] provides a large-scale evaluation benchmark covering a wide spectrum of categories. However, it does not include challenging scenarios such as moving and piled objects. Table 1 presents a comparative comparison of these datasets, alongside our new dataset PACE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 622, + 290, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 290, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 290, + 633 + ], + "type": "text", + "content": "2.2 Pose Estimation Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Instance-level Pose Estimation PPF (Point Pair Features) [8] is a pre-deep learning standard for instance-level pose estimation, using local geometric features" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 232, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 232, + 91, + 447, + 102 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 482, + 257 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 257 + ], + "type": "table", + "html": "
ModalityCat.Obj.Vid.Img.Anno.CADMov.Occ.Marker-freeArtic.Piled
YCB-Video [38]RGBD-211220K99K✓X✓✓X✓
LINEMOD-O [3]RGBD-811.2K9.2K✓X✓XX✓
NAVI [15]RGBD-3632410K10k✓XX✓XX
StereoObj-1M [24]RGBD-18182393K1.5M✓X✓XX✓
NOCS-REAL275 [34]RGBD642188K✓X✓XXX
Wild6D [40]RGBD5172251661.1M1.1MXXX✓XX
Objectron [1]RGB917k14k4M4MXXX✓XX
Scan2CAD [2]RGBD93K1506-14K✓X✓✓XX
Pix3D [29]RGBD9395-10K10K✓XX✓XX
HANDAL [11]RGB172122K308K308K✓✓✓✓XX
HouseCat6D [17]RGBD101924124K160K✓X✓✓X✓
ROPE [41]RGBD149581363332K1.5M✓X✓✓XX
PACE (Ours)RGBD4323830055K258K✓✓✓✓✓✓
", + "image_path": "7d253db0637a9e39f0f65d0929326c3f8e7c76710b689df2c5070459d7137d96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 335 + ], + "type": "text", + "content": "Table 1: Comparison of object pose datasets. From left to right, the table captures input modality, number of categories, number of instances, number of videos, number of images, number of total annotations, whether 3D CAD models are provided, whether videos include still and/or moving objects, whether objects are occluded in some frames, whether images contain artificial markers, whether poses for each part of articulated objects are provided, and whether multiple objects are piled in some frames. Compared with most other datasets, PACE contains moving and articulated objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 483, + 415 + ], + "type": "text", + "content": "from point clouds. PoseCNN [38] is the first model trained end-to-end for the task. CosyPose [19] integrates a global refinement strategy in its end-to-end pipeline. SurfEmb [12] leverages surface embeddings for correspondence matching, and GDRNPP [33] uses geometry-guided regression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 429, + 483, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 483, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 483, + 550 + ], + "type": "text", + "content": "Category-level Pose Estimation Category-level pose estimation extends the challenge to generic object categories. NOCS [34] introduces a unified coordinate space for all objects, predicting object NOCS maps from RGB images. SGPA [4] aims to adapt the structure-guided prior in the pose estimation process, while SAR-Net [21] uses shape alignment and symmetric correspondence to estimate a coarse 3D object shape and facilitate object center and size estimation. Recently, HS-Pose [43] proposes a network structure with a HS-layer that extends 3D graph convolution to extract hybrid scope latent features from point clouds for category-level object pose estimation. GenPose [42] incorporates probabilistic stable diffusion models into pose estimation, achieving strong performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 564, + 280, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 280, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 280, + 576 + ], + "type": "text", + "content": "2.3 Pose Tracking Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Instance-level Pose Tracking Methods like RBOT [30] use RGB data and 3D models to track multiple objects, employing color histograms in their cost function. PoseRBPF [5] separates rotation and translation, using an autoencoder for rotation feature embeddings. ICG [28] iteratively refines pose using geometric cues and is effective for textureless objects, with extensions incorporating visual data [27]. The first deep learning tracker, D6DT [9], and se(3)-TrackNet [36] predict frame-to-frame relative poses, using a render-and-compare strategy." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Category-level Pose Tracking 6-PACK [31] performs category-level tracking, using DenseFusion [32] features and an attention mechanism for unsupervised keypoint ordering and interframe motion via keypoint matching. BundleTrack [35] performs pose tracking without relying on 3D models, instead using video segmentation and pose graph optimization. CAPTRA [37] tracks 9DoF poses for rigid and articulated objects, with heads for rotation regression and normalized coordinate prediction, from which 3D size and translation can be computed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 224, + 289, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 289, + 236 + ], + "type": "text", + "content": "3 Construction of PACE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 291 + ], + "type": "text", + "content": "A key contribution of this work is the establishment of a scalable and reliable annotation framework, enabling the collection of large-scale and accurate pose annotations. An overview of the pipeline is depicted in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 151, + 312, + 261, + 452 + ], + "blocks": [ + { + "bbox": [ + 151, + 312, + 261, + 452 + ], + "lines": [ + { + "bbox": [ + 151, + 312, + 261, + 452 + ], + "spans": [ + { + "bbox": [ + 151, + 312, + 261, + 452 + ], + "type": "image", + "image_path": "8d213f454beaf986d901978d1b98d1f1fd795646c7b6dbfdb5d3b5b5e3e7620f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 459, + 411, + 472 + ], + "lines": [ + { + "bbox": [ + 201, + 459, + 411, + 472 + ], + "spans": [ + { + "bbox": [ + 201, + 459, + 411, + 472 + ], + "type": "text", + "content": "Fig. 3: Overview of the PACE annotation pipeline." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 262, + 312, + 463, + 451 + ], + "blocks": [ + { + "bbox": [ + 262, + 312, + 463, + 451 + ], + "lines": [ + { + "bbox": [ + 262, + 312, + 463, + 451 + ], + "spans": [ + { + "bbox": [ + 262, + 312, + 463, + 451 + ], + "type": "image", + "image_path": "a451200f8d3953ddf29530ce271117558c74735c7d539ec6663f1d06d5bfb53a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 520, + 386, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 520, + 386, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 520, + 386, + 533 + ], + "type": "text", + "content": "3.1 Acquisition of 3D Scans of Common Objects" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We began by digitizing an extensive collection of commonplace objects. These items were categorized into 43 distinct categories, as represented in Figure 5. The Einscan Pro 2X was utilized for rapid scanning of all objects, typically completing within 5 to 10 minutes per object. To expedite this process, we employed a rotatable platform to acquire multiple viewpoints of objects. After scanning, objects were manually aligned to a standard pose within a uniform coordinate system. For each object, we centered its axis-aligned bounding box at the origin and then aligned the bounding box orientation along a common axis within the category. We additionally annotated rotational symmetries. High-resolution meshes were simplified to lower-resolution ones for a smoother annotation workflow." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 232, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 232, + 91, + 448, + 102 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Articulated Objects Diverging from many prior pose estimation datasets, our collection encompasses a very wide set of objects, including articulated objects from the AKB48 [23] dataset, namely: scissors, cutters, clips, and boxes. We adopt the alignment methodology from the original AKB48 dataset, without modification. These objects are segmented into multiple parts with hierarchical relationships, presenting a difficult challenge for pose estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 208, + 309, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 208, + 309, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 208, + 309, + 220 + ], + "type": "text", + "content": "3.2 RGB-D Sequence Acquisition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 230, + 483, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 483, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 483, + 315 + ], + "type": "text", + "content": "We implemented a 3-camera system to aid in data acquisition and annotation, comprising three Intel Realsense D415 RGB-D cameras affixed to a metal framework, as illustrated in the bottom of Figure 3. The advantages of this setup include: (1) tripling the data yield; (2) reducing ambiguity in pose annotation, especially regarding translation along the depth dimension, by enforcing annotation consistency across all 3 views; (3) enhancing tracking accuracy of still objects with ArUco markers [10], making PnP more stable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 483, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 483, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 483, + 472 + ], + "type": "text", + "content": "Calibration of Multi-Camera Extrinsic Parameters We calibrated this multicamera system through a semi-automatic process. ArUco markers [10] alone proved insufficient for high-accuracy rotation estimation. Hence, we resorted to trifocal tensor estimation, i.e. TFT [16]. The process begins with feature extraction and matching, followed by bundle adjustment to refine the positions of the 3D landmarks and camera poses. For reliable feature matching, we employed the SuperPoint [7] descriptor and SuperGlue [26] matcher, using a stringent threshold for matching. We observed that rotational component of the resulting extrinsic parameters is precise, but the translation aspect suffers from scale ambiguity inherent in Structure-from-Motion approaches. We corrected this by calibrating the scale against markers, applying the following formula to obtain the optimal scale factor, by setting gradient of " + }, + { + "bbox": [ + 130, + 327, + 483, + 472 + ], + "type": "inline_equation", + "content": "\\| s\\hat{\\mathbf{t}} - \\mathbf{t}'\\|_2^2" + }, + { + "bbox": [ + 130, + 327, + 483, + 472 + ], + "type": "text", + "content": " to zero:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 217, + 482, + 399, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 482, + 399, + 510 + ], + "spans": [ + { + "bbox": [ + 217, + 482, + 399, + 510 + ], + "type": "interline_equation", + "content": "s _ {1 \\rightarrow 2} = \\frac {\\hat {\\mathbf {t}} _ {1 \\rightarrow 2} \\cdot \\mathbf {t} _ {1 \\rightarrow 2} ^ {\\prime}}{\\hat {\\mathbf {t}} _ {1 \\rightarrow 2} \\cdot \\hat {\\mathbf {t}} _ {1 \\rightarrow 2}}, s _ {1 \\rightarrow 3} = \\frac {\\hat {\\mathbf {t}} _ {1 \\rightarrow 3} \\cdot \\mathbf {t} _ {1 \\rightarrow 3} ^ {\\prime}}{\\hat {\\mathbf {t}} _ {1 \\rightarrow 3} \\cdot \\hat {\\mathbf {t}} _ {1 \\rightarrow 3}},", + "image_path": "6b17fc63799bec2ebc270468a1829d395faa5a55fb6a384fad453c94cf838827.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_{i\\rightarrow j}" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": " is the TFT predicted translation (up to a scale) from camera " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": " to camera " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{i\\rightarrow j}^{\\prime}" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": " is the marker-calibrated translation in real metric scale. We set the extrinsics of the first camera to be the identity matrix. We use Intel's dynamic calibration tool to calibrate intrinsics. In cases of repetitive textures, where SuperPoint+SuperGlue were unreliable, we made correspondences manually." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 595, + 338, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 595, + 338, + 607 + ], + "spans": [ + { + "bbox": [ + 131, + 595, + 338, + 607 + ], + "type": "text", + "content": "3.3 Annotation of Pose Ground-Truths" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Previous works have used ArUco markers to automate pose estimation through Perspective-n-Points, and we do this also. Specifically, we used ArUco markers to automate the annotation of still object poses (i.e. not moving), following LINEMOD [3] and NOCS [34]. We note that this step has two limitations:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 188 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 151 + ], + "type": "text", + "content": "1. Markers in the scene detract from realism and compromise dataset integrity: training on marker-augmented imagery may result in overfitting to these artificial patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 152, + 481, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 152, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 136, + 152, + 481, + 188 + ], + "type": "text", + "content": "2. Marker-based annotations are inapplicable to moving objects (since the markers are typically placed on the background), thus limiting the setup to annotating still objects." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 198, + 479, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 479, + 234 + ], + "type": "text", + "content": "We compensate for both of these issues: we remove marker appearances from the dataset (after using them for annotation), and we introduce additional tools to annotate the moving objects. We describe these compensation steps next." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 402 + ], + "type": "text", + "content": "Marker removal We removed the marker appearances from the dataset, using a marker inpainting strategy, detailed as follows. In step 1, we place a marker (Marker 1) somewhere within the camera's field of view. We then record a short video with this marker in view. In step 2, we place a second marker (Marker 2) at a chosen distance from the first, and record another video. In step 3, we remove Marker 1, and begin the actual object capture process (with only Marker 2 present). After this process, we end up with: (1) frames with Marker 1 only, which clearly depict the surface where Marker 2 will later appear; (2) frames capturing both markers, providing helpful calibration cues; (3) frames with Marker 2 only, which represent our main capture. We leverage Marker 2 for automated pose tracking, and manually correct the tracking every 40 frames in case of drift. We use the frames from the first two steps to seamlessly inpaint [25] Marker 2's area for the final dataset, as depicted in Figure 4." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 134, + 422, + 480, + 478 + ], + "blocks": [ + { + "bbox": [ + 134, + 422, + 480, + 478 + ], + "lines": [ + { + "bbox": [ + 134, + 422, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 134, + 422, + 480, + 478 + ], + "type": "image", + "image_path": "f34949e0cd9adf9c22dc9d477a3416bbc3c61039963a8655e9afb27a0d19f5a4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 487, + 414, + 498 + ], + "lines": [ + { + "bbox": [ + 198, + 487, + 414, + 498 + ], + "spans": [ + { + "bbox": [ + 198, + 487, + 414, + 498 + ], + "type": "text", + "content": "Fig. 4: Illustration of the marker inpainting process." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Annotation of Moving Object Positions For moving objects, the traditional marker-based tracking approach is insufficient, as the scene markers do not move in tandem with the objects. In such instances, we harnessed the capabilities of BundleTrack [35], an advanced RGB image-based tracking algorithm. BundleTrack conducts feature correspondence analysis between successive frames to estimate poses, complemented by a bundle adjustment algorithm to optimize keyframes globally and minimize tracking errors. Despite BundleTrack's proficiency in approximating poses, it is prone to drift, necessitating the manual adjustment of poses every ten frames to ensure precision. We adopted a rigorous manual checking process, examining every tenth frame for 3D pointcloud alignment and 2D multi-view agreement. We detail the annotation tool and its" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 232, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 232, + 91, + 448, + 101 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "graphical interface in the supplementary. This delicate process represents the most labor-intensive aspect of our annotation pipeline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 199 + ], + "type": "text", + "content": "In order to verify that our annotation process gives accurate poses, we conducted an experiment where we used our setup to manually annotate the poses of Scene 1 from NOCS-REAL275 [34] dataset, and compare the annotations to the provided ground truths. We recorded an average pose annotation error of " + }, + { + "bbox": [ + 132, + 140, + 481, + 199 + ], + "type": "inline_equation", + "content": "0.9^{\\circ}" + }, + { + "bbox": [ + 132, + 140, + 481, + 199 + ], + "type": "text", + "content": " for rotation and " + }, + { + "bbox": [ + 132, + 140, + 481, + 199 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{mm}" + }, + { + "bbox": [ + 132, + 140, + 481, + 199 + ], + "type": "text", + "content": " for translation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 211, + 481, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 211, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 481, + 271 + ], + "type": "text", + "content": "Generation of Segmentation Masks Upon successful pose annotation, we generated occlusion-aware segmentation masks, via depth rendering. In cases where hand interactions are involved, we employed the segmentation model SAM [18] to estimate hand masks, and subsequently subtracted these from the previously computed object masks to obtain the occlusion-aware segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 289, + 259, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 259, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 259, + 303 + ], + "type": "text", + "content": "4 Dataset Statistics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "text", + "content": "In pursuit of diversity, we placed the objects in ten disparate environments, each featuring varying levels and configurations of occlusion. We captured 10 video sequences per scene, each sequence including 1 to 7 objects. We captured RGB and depth data using an Intel RealSense D415 camera, with a resolution of " + }, + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "text", + "content": ". The capture process spans distances ranging from " + }, + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 132, + 314, + 481, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 392, + 392, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 392, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 392, + 404 + ], + "type": "text", + "content": "4.1 Annotation Distribution and Object Diversity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 411, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 481, + 483 + ], + "type": "text", + "content": "The comprehensive distribution of pose annotation counts is depicted in Figure 5, showcasing a diverse array including both rigid and articulated objects. As demonstrated in Figure 7, our object database encompasses a broad spectrum of object sizes, with the majority measuring near " + }, + { + "bbox": [ + 132, + 411, + 481, + 483 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 132, + 411, + 481, + 483 + ], + "type": "text", + "content": " along the bounding box diagonal. It also includes larger items such as storage bins, adding to the heterogeneity of the dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 500, + 455, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 455, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 455, + 513 + ], + "type": "text", + "content": "4.2 Variability in Occlusion, Pose, and Environmental Context" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 520, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 592 + ], + "type": "text", + "content": "The dataset has rich variability in occlusion levels and object poses, as illustrated in Figure 6, which outlines the statistical distribution of both visibility ratios and azimuth/elevation angles, indicating a comprehensive coverage reflecting real-world scenarios. These statistics can be helpful for assessing the robustness of pose estimation algorithms. Besides, the ten distinct capturing environments are shown in Figure 8, illustrating diverse backgrounds and lighting conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 609, + 376, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 376, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 376, + 622 + ], + "type": "text", + "content": "4.3 Simulated Training Data and Dataset Split" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "To facilitate a comprehensive and equitable evaluation of pose estimation methodologies, we partitioned our entire real-world dataset into validation and test subsets following a " + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "20/80" + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": " ratio. The test subset is used for both instance-level and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 118, + 460, + 266 + ], + "blocks": [ + { + "bbox": [ + 153, + 118, + 460, + 266 + ], + "lines": [ + { + "bbox": [ + 153, + 118, + 460, + 266 + ], + "spans": [ + { + "bbox": [ + 153, + 118, + 460, + 266 + ], + "type": "image", + "image_path": "032379076a8233035dd5607347c922c1ad5bc5145350c1aa942ed736f32959f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 280, + 470, + 292 + ], + "lines": [ + { + "bbox": [ + 143, + 280, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 143, + 280, + 470, + 292 + ], + "type": "text", + "content": "Fig. 5: Distribution of pose annotation counts across different object categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 316, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 316, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 133, + 316, + 480, + 363 + ], + "type": "text", + "content": "category-level pose estimation evaluation tasks. The real-world validation subset is used for instance-level pose estimation only. We render another photo-realistic validation subset for category-level pose estimation, in order to cover enough category-level objects that are different from the test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 364, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 481, + 448 + ], + "type": "text", + "content": "Additionally, to support research necessitating extensive training datasets, we generated a large synthetic dataset called PACE-Sim that contains two different training subsets, for instance-level and category-level pose estimation, respectively. We use a physically based renderer [6] to ensure the images are highly photo-realistic. The synthetic training sets consist of 100K images, each featuring dozens of objects, yielding 2.4M annotations. The distribution of object instances in both PACE and PACE-Sim can be found in supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 467, + 293, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 467, + 293, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 467, + 293, + 479 + ], + "type": "text", + "content": "5 Evaluation Benchmarks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 493, + 481, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 493, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 133, + 493, + 481, + 565 + ], + "type": "text", + "content": "We assess both pose estimation and tracking. State-of-the-art (SOTA) pose estimation techniques generally involve two phases: first detecting or segmenting the object, and then estimating its pose within the identified mask. For a fair comparison, we evaluate pose estimation assuming perfect object detection. Detailed SOTA detection performance is available in the supplementary materials but is beyond this dataset's primary scope." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 565, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 480, + 601 + ], + "type": "text", + "content": "For pose tracking, we provide ground-truth data in the initial frame and require methods to accurately track the object's rotations and translations through the video sequence. Both model-free and model-based trackers are assessed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 620, + 302, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 620, + 302, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 620, + 302, + 632 + ], + "type": "text", + "content": "5.1 Pose Estimation Benchmark" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "text", + "content": "This benchmark is divided into instance-level and category-level pose estimation. The former concerns object instances seen during training (in different scenes)," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 101 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 114, + 318, + 186 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 318, + 186 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 318, + 186 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 318, + 186 + ], + "type": "image", + "image_path": "5f77a48e05238c0c1bd36df306fae288707a230de6618eddf3046e79edad0f55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 190, + 244, + 269 + ], + "blocks": [ + { + "bbox": [ + 136, + 190, + 244, + 269 + ], + "lines": [ + { + "bbox": [ + 136, + 190, + 244, + 269 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 244, + 269 + ], + "type": "image", + "image_path": "f6a81fe12a22692ed36cb230cefc7be97e55e0ef936f4248533b98d639d44f96.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 278, + 338, + 300 + ], + "lines": [ + { + "bbox": [ + 133, + 278, + 338, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 338, + 300 + ], + "type": "text", + "content": "Fig. 6: Distribution of occlusions, azimuth and elevation viewing angles." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 245, + 194, + 337, + 269 + ], + "blocks": [ + { + "bbox": [ + 245, + 194, + 337, + 269 + ], + "lines": [ + { + "bbox": [ + 245, + 194, + 337, + 269 + ], + "spans": [ + { + "bbox": [ + 245, + 194, + 337, + 269 + ], + "type": "image", + "image_path": "128b405aacc842d5b5dacb25428b3949349e0dadcdce381532c2f0b8eeaa7a04.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 258, + 478, + 270 + ], + "lines": [ + { + "bbox": [ + 338, + 258, + 478, + 270 + ], + "spans": [ + { + "bbox": [ + 338, + 258, + 478, + 270 + ], + "type": "text", + "content": "Fig. 7: Distribution of object sizes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 339, + 144, + 480, + 251 + ], + "blocks": [ + { + "bbox": [ + 339, + 144, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 339, + 144, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 339, + 144, + 480, + 251 + ], + "type": "image", + "image_path": "e68397aeb3544284f961a7e34c973659c7dd2e127272bbf378c282a35ac53607.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 313, + 200, + 352 + ], + "blocks": [ + { + "bbox": [ + 133, + 313, + 200, + 352 + ], + "lines": [ + { + "bbox": [ + 133, + 313, + 200, + 352 + ], + "spans": [ + { + "bbox": [ + 133, + 313, + 200, + 352 + ], + "type": "image", + "image_path": "77172050f348784b0bccf84914bccbe018dcfb8fdf6dafbecd53b082d8355988.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 354, + 185, + 362 + ], + "lines": [ + { + "bbox": [ + 147, + 354, + 185, + 362 + ], + "spans": [ + { + "bbox": [ + 147, + 354, + 185, + 362 + ], + "type": "text", + "content": "(a) basket" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 203, + 314, + 270, + 352 + ], + "blocks": [ + { + "bbox": [ + 203, + 314, + 270, + 352 + ], + "lines": [ + { + "bbox": [ + 203, + 314, + 270, + 352 + ], + "spans": [ + { + "bbox": [ + 203, + 314, + 270, + 352 + ], + "type": "image", + "image_path": "6ab287369884ee13589e1c34d5694c61f4f0e1c8f2241ca629ff07c29da2fea8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 273, + 313, + 340, + 352 + ], + "blocks": [ + { + "bbox": [ + 273, + 313, + 340, + 352 + ], + "lines": [ + { + "bbox": [ + 273, + 313, + 340, + 352 + ], + "spans": [ + { + "bbox": [ + 273, + 313, + 340, + 352 + ], + "type": "image", + "image_path": "db28857fbb2811f8ace59a39ac602e70164ac4d36c63ee3c0fd676e8208b1141.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 343, + 313, + 410, + 352 + ], + "blocks": [ + { + "bbox": [ + 343, + 313, + 410, + 352 + ], + "lines": [ + { + "bbox": [ + 343, + 313, + 410, + 352 + ], + "spans": [ + { + "bbox": [ + 343, + 313, + 410, + 352 + ], + "type": "image", + "image_path": "3989cae8268163364a378d396a4b1bacdbf1b2d45547e274267ed89dfc434967.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 413, + 313, + 481, + 352 + ], + "blocks": [ + { + "bbox": [ + 413, + 313, + 481, + 352 + ], + "lines": [ + { + "bbox": [ + 413, + 313, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 413, + 313, + 481, + 352 + ], + "type": "image", + "image_path": "69b6ae908aa0d229ee469777bf56494c15eda96ed4b5a8987c52d352dbc25102.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 133, + 363, + 200, + 401 + ], + "blocks": [ + { + "bbox": [ + 133, + 363, + 200, + 401 + ], + "lines": [ + { + "bbox": [ + 133, + 363, + 200, + 401 + ], + "spans": [ + { + "bbox": [ + 133, + 363, + 200, + 401 + ], + "type": "image", + "image_path": "7577a49491983fb703e7a9633429086acfdacba2a52f38814b53a710d4b3c056.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 403, + 192, + 411 + ], + "lines": [ + { + "bbox": [ + 140, + 403, + 192, + 411 + ], + "spans": [ + { + "bbox": [ + 140, + 403, + 192, + 411 + ], + "type": "text", + "content": "(f) china floor" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 203, + 363, + 270, + 401 + ], + "blocks": [ + { + "bbox": [ + 215, + 354, + 258, + 363 + ], + "lines": [ + { + "bbox": [ + 215, + 354, + 258, + 363 + ], + "spans": [ + { + "bbox": [ + 215, + 354, + 258, + 363 + ], + "type": "text", + "content": "(b) box top" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 363, + 270, + 401 + ], + "lines": [ + { + "bbox": [ + 203, + 363, + 270, + 401 + ], + "spans": [ + { + "bbox": [ + 203, + 363, + 270, + 401 + ], + "type": "image", + "image_path": "ad7a091475671192b9d3591a1dea85bd027f7bde6db1b590ae6a829380f4fda4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 403, + 253, + 411 + ], + "lines": [ + { + "bbox": [ + 219, + 403, + 253, + 411 + ], + "spans": [ + { + "bbox": [ + 219, + 403, + 253, + 411 + ], + "type": "text", + "content": "(g) desk1" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 421, + 408, + 433 + ], + "lines": [ + { + "bbox": [ + 205, + 421, + 408, + 433 + ], + "spans": [ + { + "bbox": [ + 205, + 421, + 408, + 433 + ], + "type": "text", + "content": "Fig. 8: Sample images from the 10 environments." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 273, + 363, + 340, + 401 + ], + "blocks": [ + { + "bbox": [ + 287, + 354, + 326, + 362 + ], + "lines": [ + { + "bbox": [ + 287, + 354, + 326, + 362 + ], + "spans": [ + { + "bbox": [ + 287, + 354, + 326, + 362 + ], + "type": "text", + "content": "(c) cabinet" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 273, + 363, + 340, + 401 + ], + "lines": [ + { + "bbox": [ + 273, + 363, + 340, + 401 + ], + "spans": [ + { + "bbox": [ + 273, + 363, + 340, + 401 + ], + "type": "image", + "image_path": "d255b6ee5f955337beeb5175ced43327c2ef62e8bc01c1cdb9cf31a5fa326519.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 403, + 324, + 411 + ], + "lines": [ + { + "bbox": [ + 289, + 403, + 324, + 411 + ], + "spans": [ + { + "bbox": [ + 289, + 403, + 324, + 411 + ], + "type": "text", + "content": "(h) desk2" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 342, + 363, + 410, + 401 + ], + "blocks": [ + { + "bbox": [ + 358, + 354, + 395, + 362 + ], + "lines": [ + { + "bbox": [ + 358, + 354, + 395, + 362 + ], + "spans": [ + { + "bbox": [ + 358, + 354, + 395, + 362 + ], + "type": "text", + "content": "(d) carpet" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 363, + 410, + 401 + ], + "lines": [ + { + "bbox": [ + 342, + 363, + 410, + 401 + ], + "spans": [ + { + "bbox": [ + 342, + 363, + 410, + 401 + ], + "type": "image", + "image_path": "d461684a4c36d427002b7ec2c948581676618f306e7c186b5755ba770c98fdfd.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 403, + 392, + 411 + ], + "lines": [ + { + "bbox": [ + 361, + 403, + 392, + 411 + ], + "spans": [ + { + "bbox": [ + 361, + 403, + 392, + 411 + ], + "type": "text", + "content": "(i) stairs" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 413, + 363, + 481, + 401 + ], + "blocks": [ + { + "bbox": [ + 431, + 354, + 462, + 362 + ], + "lines": [ + { + "bbox": [ + 431, + 354, + 462, + 362 + ], + "spans": [ + { + "bbox": [ + 431, + 354, + 462, + 362 + ], + "type": "text", + "content": "(e) chair" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 363, + 481, + 401 + ], + "lines": [ + { + "bbox": [ + 413, + 363, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 413, + 363, + 481, + 401 + ], + "type": "image", + "image_path": "414f78905a4253b1f53d17be0e720ea3fdd274f2bb2c0161c685f34ee5b4772b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 403, + 476, + 411 + ], + "lines": [ + { + "bbox": [ + 417, + 403, + 476, + 411 + ], + "spans": [ + { + "bbox": [ + 417, + 403, + 476, + 411 + ], + "type": "text", + "content": "(j) wooden floor" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 457, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 481 + ], + "type": "text", + "content": "while the latter concerns instances which were not seen but come from the same category. We will also give an in-depth analysis of SOTA's failure modes." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 502, + 480, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 480, + 528 + ], + "type": "text", + "content": "Instance-Level Pose Estimation This task demands the prediction of rotation and translation for known instances from the training set." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 542, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 602 + ], + "type": "text", + "content": "Metrics: Adhering to the BOP challenge protocol [14], we use Average Recall (AR) of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) as our metrics. Detailed computation of these metrics is described in [14]. Results are averaged across all instances." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Baselines: We assessed four baselines: PPF [8], CosyPose [19], SurfEmb [12], and GDRNPP [33]. While PPF does not require training data, the others are SOTA methods dependent on additional training on the PBR dataset. To focus on pose estimation performance, we assume that ground-truth detections are available." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Result Analysis: The quantitative analysis is presented in Table 2. Notably, while SOTA methods excel on the BOP benchmarks, they do not perform as well as PPF, which relies on local geometric feature matching in point clouds. The relative success of PPF highlights the potential for improvement in SOTA methods, particularly regarding sim-to-real ability and scalability to handle a large set of instances. Qualitative results can be found in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 187, + 208, + 427, + 264 + ], + "blocks": [ + { + "bbox": [ + 187, + 208, + 427, + 264 + ], + "lines": [ + { + "bbox": [ + 187, + 208, + 427, + 264 + ], + "spans": [ + { + "bbox": [ + 187, + 208, + 427, + 264 + ], + "type": "table", + "html": "
Modality DetectionARVSD↑ARMSSD↑ARMSPD↑AR↑
PPF [8]DG.T.53.448.155.252.2
CosyPose [19]RGBG.T.1.40.311.54.4
SurfEmb [12]RGBG.T.6.23.017.89.0
GDRNPP [33]RGB-DG.T.3.62.115.47.0
", + "image_path": "00ae53e713ed9349b2d0c00d1fe8457e9e074c3bc848534e5677a561e02fabfe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 350, + 479, + 450 + ], + "blocks": [ + { + "bbox": [ + 130, + 273, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 296 + ], + "type": "text", + "content": "Table 2: Instance-level pose estimation results showcasing the robustness of PPF and the potential for improvement in SOTA deep learning-based methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 350, + 479, + 450 + ], + "lines": [ + { + "bbox": [ + 135, + 350, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 135, + 350, + 479, + 450 + ], + "type": "table", + "html": "
DetectionIoU25↑IoU50↑AP
0:20°↑0:60°↑0:5cm↑0:15cm↑0:20°↑0:60°↑
NOCS [34]Mask-RCNN0.2/0.00.0/0.01.2/0.04.3/0.017.0/0.033.5/0.01.1/0.04.2/0.0
HS-Pose [43]G.T.36.6/0.02.7/0.05.3/0.08.6/0.348.7/34.181.0/75.63.9/0.08.1/0.3
SGPA [4]G.T.2.6/0.01.2/0.05.6/1.111.4/7.316.1/6.850.7/17.53.3/0.910.1/7.2
DualPoseNet [22]G.T.24.2/0.30.1/0.05.5/0.08.1/0.018.6/27.363.4/65.21.8/0.06.5/0.0
SAR-Net [21]G.T.22.8/0.10.2/0.05.3/0.08.8/0.848.6/38.579.8/69.73.7/0.08.2/0.8
CPPF++ [39]G.T.44.5/3.64.4/0.015.2/1.727.3/6.235.3/31.174.0/66.79.9/1.124.9/5.9
ANCSH [20]G.T.- / 0.0- / 0.0- / 0.0- / 0.2- / 18.6- / 50.4- / 0.0- / 0.2
", + "image_path": "e2bddca684d5d3684e08c686bb19a5fa15d2376c07705db3e4bc2f649f351a3a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 458, + 482, + 482 + ], + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 482 + ], + "type": "text", + "content": "Table 3: Category-level pose estimation benchmark combining the performance metrics for both rigid and articulated object pose estimation, separated by slash." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 533, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 571 + ], + "type": "text", + "content": "Category-Level Pose Estimation Category-level pose estimation tasks involve predicting the 3D bounding box dimensions, rotation, and translation of target instances, given only the category information a priori." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Metrics: We adopted from NOCS [34], calculating mean Average Precision (AP) across predefined angle and translation thresholds. Specifically, AP@0:20° represents AP averaged from " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "1^{\\circ}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " intervals; AP@0:5cm is averaged from 0cm to 5cm at 0.25cm intervals; AP@0:20°,0:5cm combines these angle and translation thresholds. IoU" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "_{25}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and IoU" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "_{50}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " denote AP for 3D bounding box matches at IoU thresholds of 25 and 50, respectively. Results are separately reported for both rigid and articulated objects." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 232, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 232, + 91, + 447, + 102 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Baselines: We evaluated seven recent category-level pose estimation methods: NOCS [34], HS-Pose [43], SGPA [4], DualPoseNet [22], SAR-Net [21], CPPF++ [39] and ANCSH [20], the latter specifically designed for articulated objects. For adapting rigid-object-focused methods to articulated objects, each movable part is considered a distinct category. Ground-truth detections are presumed for each method, except for NOCS, which outputs both instance masks and pose estimations using a unified network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 483, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 483, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 483, + 321 + ], + "type": "text", + "content": "Result Analysis: Quantitative comparisons are listed in Table 3. On rigid objects, HS-Pose and SAR-Net show fair performance on AP metrics for translation, but falls short on rotation AP metrics compared to CPPF++, suggesting a proficiency in size and translation prediction but not rotation. CPPF++ shows the best performance but is weaker on translation estimation. For articulated objects, all methods face challenges due to the movable parts, with ANCSH underperforming on the large-scale real-world dataset, indicative of a significant sim-to-real gap. NOCS shows instability, likely due to its inaccurate mask prediction. Qualitative results are available in supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "content": "Failure Mode Analysis This section delves into the analysis of two critical questions that underline the failure of SOTA deep learning-based methods in pose estimation: 1) Is the failure attributed to the sim-to-real gap, considering the exclusive use of synthetic RGB-D images for training? 2) Does the failure stem from the inherent large-scale complexity of the PACE dataset, which poses a significant challenge for baseline models?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 462 + ], + "type": "text", + "content": "Ablation Setup: To explore these questions, we introduced a new " + }, + { + "bbox": [ + 130, + 426, + 482, + 462 + ], + "type": "inline_equation", + "content": "80/10/10" + }, + { + "bbox": [ + 130, + 426, + 482, + 462 + ], + "type": "text", + "content": " train/validation/test split for the real-world data, diverging from the " + }, + { + "bbox": [ + 130, + 426, + 482, + 462 + ], + "type": "inline_equation", + "content": "0/20/80" + }, + { + "bbox": [ + 130, + 426, + 482, + 462 + ], + "type": "text", + "content": " split described in Section 4.3. We then examined the following settings:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 472, + 481, + 567 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 472, + 481, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 507 + ], + "type": "text", + "content": "Real2Real Single Instance/Category Fitting: This basic setup involves training on a single instance or category selected (we randomly select the category can and sample a can with object ID 57) from the " + }, + { + "bbox": [ + 138, + 472, + 481, + 507 + ], + "type": "inline_equation", + "content": "80 / 10 / 10" + }, + { + "bbox": [ + 138, + 472, + 481, + 507 + ], + "type": "text", + "content": " split of real data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 509, + 480, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 137, + 509, + 480, + 542 + ], + "type": "text", + "content": "- Sim2Real Single Instance/Category Fitting: This scenario differs from the first by utilizing all simulated data for training and validation, with evaluation on the same real-world test split." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 544, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 137, + 544, + 480, + 567 + ], + "type": "text", + "content": "- Real2Real All Instance/Category Fitting: Unlike the first setting, this involves training, validation, and testing across all instances or categories." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Result Analysis and Discussions: Our analysis in Table 4 reveals that baseline models effectively handle single instance or category fitting under the Real2Real Single Instance/Category Fitting setting, provided the training and test distributions align. However, the sim-to-real gap, especially noticeable in depth-input methods like DualPoseNet, HS-Pose, and SGPA, highlights significant challenges, as these methods struggle with the depth noise in real world." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 479, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 159 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 159 + ], + "type": "table", + "html": "
Training SettingInstance-Level (AR%)†Category-Level (mean AP%@0:60°, 0:15cm)†
CosyPoseSurfEmbGDRNPPNOCSHS-PoseSGPADualPoseNetSAR-NetCPPF++
Real2Real Single Inst./Cat.76.486.788.276.981.286.758.659.383.4
Sim2Real Single Inst./Cat.67.1 (-9.3↓)73.7 (-13.0↓)78.1 (-10.1↓)55.8 (-21.1↓)1.1 (-80.1↓)0.3 (-86.4↓)1.4 (-57.2↓)33.2 (-26.1↓)55.8 (-27.6↓)
Real2Real All Inst./Cat.8.2 (-68.2↓)11.2 (-75.5↓)9.7 (-78.5↓)0.3 (-76.6↓)5.6 (-75.6↓)9.8 (-76.9↓)1.2 (-57.4↓)2.1 (-57.2↓)13.7 (-69.7↓)
", + "image_path": "c0acd45070132becb0738d5a8307486c159f0733e11dfbb55eb26c18f292d32c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 246, + 482, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 318 + ], + "type": "text", + "content": "Additionally, learning across all instances/categories leads to a notable performance drop, as shown in Table 4. This drop, together with suboptimal training data performance and loss plateauing, suggests underfitting and highlights the complexity posed by the PACE dataset's diversity. Innovative approaches or larger foundational models might be necessary to address these challenges and effectively bridge the sim-to-real gap." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 333, + 292, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 333, + 292, + 345 + ], + "spans": [ + { + "bbox": [ + 131, + 333, + 292, + 345 + ], + "type": "text", + "content": "5.2 Pose Tracking Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 375 + ], + "type": "text", + "content": "We distinguish SOTA pose tracking techniques into model-based, utilizing a 3D CAD model, and model-free, which only require an initial pose from first frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "type": "text", + "content": "Metrics: For model-based tracking, we measure the area under curve (AUC) for ADD, ADD-S [38], and ADD(-S) metrics, with a maximum threshold of " + }, + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 130, + 384, + 482, + 444 + ], + "type": "text", + "content": " [32]. The ADD metric, introduced in [13], computes the average per-point distance between point clouds transformed by the predicted and ground-truth poses. ADD-S accounts for point correspondence ambiguities in symmetric objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": "For model-free tracking, we adopt four metrics from prior studies [31]: 1) " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": ", indicating the percentage of predictions within " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": " rotation and " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": " translation errors. 2) IoU25, the percentage of instances where the intersection over union of the 3D bounding boxes exceeds " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": ". 3) " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{err}}" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": " and 4) " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{err}}" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": ", the mean rotational and translational errors in degrees and centimeters, respectively, only considering instances with IoU " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": ">25\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": ". We exclude objects with visibility under " + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 540 + ], + "type": "text", + "content": " as per the BOP convention [14]. Results are averaged separately across rigid and articulated object categories." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 209, + 559, + 406, + 601 + ], + "blocks": [ + { + "bbox": [ + 130, + 159, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 214 + ], + "type": "text", + "content": "Table 4: Ablation Study: This study illustrates the challenges models face when transitioning from real-to-real settings. It showcases performance disparities when scaling from single instances/categories to the entire dataset, emphasizing the scalability challenge from PACE and the profound impact of sim-to-real discrepancies on depth-dependent methodologies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 209, + 559, + 406, + 601 + ], + "lines": [ + { + "bbox": [ + 209, + 559, + 406, + 601 + ], + "spans": [ + { + "bbox": [ + 209, + 559, + 406, + 601 + ], + "type": "table", + "html": "
ModalityADD↑ADD-S↑ADD(-S)↑
RBOT [30]RGB7.1/0.510.3/0.87.4/0.5
ICG [28]RGB-D35.6/10.148.1/15.238.1/10.1
", + "image_path": "619fff9e33bbff2cefe1fe5ad704fe19c3a9cadaa27e90b71f69168b0ed8596f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 608, + 482, + 631 + ], + "lines": [ + { + "bbox": [ + 130, + 608, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 482, + 631 + ], + "type": "text", + "content": "Table 5: Model-Based Pose Tracking. We report the area under curve (AUC) with respect to ADD, ADD-S and ADD(-S). The higher value, the better performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 232, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 232, + 91, + 448, + 101 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 445, + 162 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 445, + 162 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 445, + 162 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 445, + 162 + ], + "type": "table", + "html": "
Training-Free Modality5°5cm↑IoU25↑Rerr↓Terr↓
BundleTrack [35]✓RGB6.4/11.29.1/14.13.2/5.52.6/0.8
CAPTRA [37]✗D12.9/4.445.8/20.619.2/40.92.2/1.5
6-PACK [31]✗RGB-D9.2/3.923.1/16.717.7/33.62.1/1.2
", + "image_path": "3b6be880248e3fd23dcb4794656c55cbc8c3c7ae7c4fdba3702bea810fbecc41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "text", + "content": "Table 6: Model-Free Pose Tracking. Both rigid and articulated results are reported. For " + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "text", + "content": " and IoU25, the higher value means better performance while for " + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{err}" + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{err}}" + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "text", + "content": " , the situation is reversed. Note that " + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathrm{R}_{err}" + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{err}}" + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "text", + "content": " are respect to IoU25 since objects with IoU " + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\leq 25\\%" + }, + { + "bbox": [ + 130, + 171, + 482, + 216 + ], + "type": "text", + "content": " are not counted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 247, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 319 + ], + "type": "text", + "content": "Baselines: For tracking methods using object masks, we applied the ground-truth masks directly. Except for CAPTRA, which also predicts the 3D bounding box size, all methods estimate only rotation and translation; however, we used ground-truth bounding box sizes for consistency in evaluations. Unlike other methods that treat each part of an articulated object independently, CAPTRA considers the entire object as one entity, trained on our synthetic PBR data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "content": "Result and Failure Mode Analysis: Tables 5 and 6 display the performance of state-of-the-art (SOTA) tracking methods on the PACE dataset, highlighting both rigid and articulated object results. The highest achieved ADD(-S) for model-based methods is only " + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "inline_equation", + "content": "38.1\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "content": " for rigid and " + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "inline_equation", + "content": "10.1\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "content": " for articulated objects by ICG. Furthermore, Table 6 reveals severe limitations of model-free methods, with the best " + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "content": " accuracy under " + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "content": ", posing major tracking challenges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "content": "The convention followed by 6-PACK [31] and BundleTrack [35] disregards objects with " + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU} \\leq 25\\%" + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "content": " when computing rotational and translational errors. However, this approach could overestimate the performance on " + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{R}_{err}" + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{err}" + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "content": ". BundleTrack [35] reports lower performance in " + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "content": " and IoU25 metrics yet shows seemingly better results for " + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{R}_{err}" + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{err}" + }, + { + "bbox": [ + 130, + 400, + 483, + 484 + ], + "type": "text", + "content": ", suggesting that it can only track well within a threshold but cannot recover once the object is lost. Qualitative results can be found in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 500, + 324, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 324, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 324, + 514 + ], + "type": "text", + "content": "6 Conclusions and Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 524, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 482, + 620 + ], + "type": "text", + "content": "We present PACE, a new large-scale benchmark for 3D object pose estimation and tracking, designed with various occlusion levels across ten diverse environments. This benchmark both reflects prior progress and urges further research to tackle real-world complexities. Our analysis indicates significant advances in pose estimation, yet underscores a notable real-world performance gap, particularly with articulated objects and in terms of robustness and scalability. Future efforts could focus on advanced sim-to-real techniques and developing larger models that capture a broader spectrum of object features and environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Limitations Due to resource limitations, PACE lacks large and valuable objects like tables, cameras, and laptops, which may affect its breadth in representing real-world object scenarios." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 266, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 266, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 266, + 129 + ], + "type": "text", + "content": "7 Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Project of China (No. 2022ZD0160102), National Key Research and Development Project of China (No. 2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants. Yang You is also supported in part by the Outstanding Doctoral Graduates Development Scholarship of Shanghai Jiao Tong University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 198, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 198, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 198, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 239, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 239, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 284 + ], + "type": "text", + "content": "1. Ahmadyan, A., Zhang, L., Ablavatski, A., Wei, J., Grundmann, M.: Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7822-7831 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 284, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 481, + 317 + ], + "type": "text", + "content": "2. Avetisyan, A., Dahnert, M., Dai, A., Savva, M., Chang, A.X., Nießner, M.: Scan2cad: Learning cad model alignment in rgb-d scans. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2614-2623 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 317, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 481, + 361 + ], + "type": "text", + "content": "3. Brachmann, E., Krull, A., Michel, F., Gumhold, S., Shotton, J., Rother, C.: Learning 6d object pose estimation using 3d object coordinates. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13. pp. 536-551. Springer (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 393 + ], + "type": "text", + "content": "4. Chen, K., Dou, Q.: Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2773-2782 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 393, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 481, + 426 + ], + "type": "text", + "content": "5. Deng, X., Mousavian, A., Xiang, Y., Xia, F., Bretl, T., Fox, D.: Poserbpf: A raobackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics 37(5), 1328-1342 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 426, + 481, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 481, + 470 + ], + "type": "text", + "content": "6. Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R.: Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8(82), 4901 (2023). https://doi.org/10.21105/joss.04901, https://doi.org/10.21105/joss.04901" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "type": "text", + "content": "7. DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 224-236 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 502, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 535 + ], + "type": "text", + "content": "8. Drost, B., Ulrich, M., Navab, N., Ilic, S.: Model globally, match locally: Efficient and robust 3d object recognition. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 998-1005. IEEE (2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "type": "text", + "content": "9. Garon, M., Lalonde, J.F.: Deep 6-dof tracking. IEEE transactions on visualization and computer graphics 23(11), 2410-2418 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 589 + ], + "type": "text", + "content": "0. Garrido-Jurado, S., Muñoz-Salinas, R., Madrid-Cuevas, F.J., Marín-Jiménez, M.J.: Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition 47(6), 2280-2292 (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "type": "text", + "content": "1. Guo, A., Wen, B., Yuan, J., OTHERS: Handal: A dataset of real-world manipulable object categories with pose annotations, affordances, and reconstructions (2023), https://arxiv.org/pdf/2308.01477.pdf" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "2. Haugaard, R.L., Buch, A.G.: Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6749-6758 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 232, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 232, + 91, + 447, + 101 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 171 + ], + "type": "text", + "content": "13. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N.: Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. pp. 548-562. Springer (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 171, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 482, + 215 + ], + "type": "text", + "content": "14. Hodaţ, T., Sundermeyer, M., Drost, B., Labbé, Y., Brachmann, E., Michel, F., Rother, C., Matas, J.: Bop challenge 2020 on 6d object localization. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 577-594. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 215, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 215, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 215, + 482, + 257 + ], + "type": "text", + "content": "15. Jampani, V., Maninis, K.K., Engelhardt, A., Karpur, A., Truong, K., Sargent, K., Popov, S., Araujo, A., Martin-Brualla, R., Patel, K., et al.: Navi: Category-agnostic image collections with high-quality 3d shape and pose annotations. arXiv preprint arXiv:2306.09109 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 257, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 482, + 300 + ], + "type": "text", + "content": "16. Julià, L.F., Monasse, P.: A critical review of the trifocal tensor estimation. In: Image and Video Technology: 8th Pacific-Rim Symposium, PSIVT 2017, Wuhan, China, November 20-24, 2017, Revised Selected Papers 8. pp. 337-349. Springer (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 300, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 482, + 354 + ], + "type": "text", + "content": "17. Jung, H., Wu, S.C., Ruhkamp, P., Zhai, G., Schieber, H., Rizzoli, G., Wang, P., Zhao, H., Garattoni, L., Meier, S., et al.: Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22498-22508 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 482, + 386 + ], + "type": "text", + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 386, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 482, + 430 + ], + "type": "text", + "content": "19. Labbe, Y., Carpentier, J., Aubry, M., Sivic, J.: Cosypose: Consistent multi-view multi-object 6d pose estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 574-591. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 430, + 482, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 430, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 133, + 430, + 482, + 461 + ], + "type": "text", + "content": "20. Li, X., Wang, H., Yi, L., Guibas, L., Abbott, A.L., Song, S.: Category-level articulated object pose estimation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 461, + 482, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 461, + 482, + 505 + ], + "spans": [ + { + "bbox": [ + 133, + 461, + 482, + 505 + ], + "type": "text", + "content": "21. Lin, H., Liu, Z., Cheang, C., Fu, Y., Guo, G., Xue, X.: Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6707-6717 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 505, + 482, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 505, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 133, + 505, + 482, + 548 + ], + "type": "text", + "content": "22. Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., Li, Y.: Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3560-3569 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 548, + 482, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 548, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 133, + 548, + 482, + 580 + ], + "type": "text", + "content": "23. Liu, L., Xu, W., Fu, H., Qian, S., Yu, Q., Han, Y., Lu, C.: Akb-48: A real-world articulated object knowledge base. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14809-14818 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 580, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 580, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 580, + 482, + 611 + ], + "type": "text", + "content": "24. Liu, X., Iwase, S., Kitani, K.M.: Stereobj-1m: Large-scale stereo image dataset for 6d object pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10870-10879 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "25. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 482, + 666 + ], + "type": "text", + "content": "26. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4938-4947 (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 483, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "27. Stoiber, M., Elsayed, M., Reichert, A.E., Steidle, F., Lee, D., Triebel, R.: Fusing visual appearance and geometry for multi-modality 6dof object tracking. arXiv preprint arXiv:2302.11458 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 483, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 483, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 483, + 194 + ], + "type": "text", + "content": "28. Stoiber, M., Sundermeyer, M., Triebel, R.: Iterative corresponding geometry: Fusing region and depth for highly efficient 3d tracking of textureless objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6855-6865 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 483, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 483, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 483, + 237 + ], + "type": "text", + "content": "29. Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2974-2983 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "text", + "content": "30. Tjaden, H., Schwanecke, U., Schömer, E., Cremers, D.: A region-based gauss-newton approach to real-time monocular multiple object tracking. IEEE transactions on pattern analysis and machine intelligence 41(8), 1797-1812 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 483, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 483, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 483, + 314 + ], + "type": "text", + "content": "31. Wang, C., Martin-Martin, R., Xu, D., Lv, J., Lu, C., Fei-Fei, L., Savarese, S., Zhu, Y.: 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 10059-10066. IEEE (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 483, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 483, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 483, + 357 + ], + "type": "text", + "content": "32. Wang, C., Xu, D., Zhu, Y., Martin-Martín, R., Lu, C., Fei-Fei, L., Savarese, S.: Densefusion: 6d object pose estimation by iterative dense fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3343-3352 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 482, + 402 + ], + "type": "text", + "content": "33. Wang, G., Manhardt, F., Tombari, F., Ji, X.: GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16611-16621 (June 2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 402, + 482, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 482, + 446 + ], + "type": "text", + "content": "34. Wang, H., Sridhar, S., Huang, J., Valentin, J., Song, S., Guibas, L.J.: Normalized object coordinate space for category-level 6d object pose and size estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2642-2651 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 447, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 482, + 479 + ], + "type": "text", + "content": "35. Wen, B., Bekris, K.: Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8067-8074. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 482, + 522 + ], + "type": "text", + "content": "36. Wen, B., Mitash, C., Ren, B., Bekris, K.E.: se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10367-10373. IEEE (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 522, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 482, + 567 + ], + "type": "text", + "content": "37. Weng, Y., Wang, H., Zhou, Q., Qin, Y., Duan, Y., Fan, Q., Chen, B., Su, H., Guibas, L.J.: Captra: Category-level pose tracking for rigid and articulated objects from point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13209-13218 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 567, + 482, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 482, + 599 + ], + "type": "text", + "content": "38. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D.: PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 600, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 482, + 632 + ], + "type": "text", + "content": "39. You, Y., He, W., Liu, J., Xiong, H., Wang, W., Lu, C.: Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation. arXiv preprint arXiv:2211.13398 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "type": "text", + "content": "40. Ze, Y., Wang, X.: Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems 35, 27469-27483 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 232, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 232, + 91, + 447, + 101 + ], + "type": "text", + "content": "PACE: Pose Annotations in Cluttered Environments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "41. Zhang, J., Huang, W., Peng, B., Wu, M., Hu, F., Chen, Z., Zhao, B., Dong, H.: Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. arXiv preprint arXiv:2406.04316 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "42. Zhang, J., Wu, M., Dong, H.: Genpose: Generative category-level object pose estimation via diffusion models. arXiv preprint arXiv:2306.10531 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 216 + ], + "type": "text", + "content": "43. Zheng, L., Wang, C., Sun, Y., Dasgupta, E., Chen, H., Leonardis, A., Zhang, W., Chang, H.J.: Hs-pose: Hybrid scope feature extraction for category-level object pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17163-17173 (2023)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. You et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_content_list.json b/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..800f0561c3bd637826a6e9ff093ce86eb1a271b9 --- /dev/null +++ b/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_content_list.json @@ -0,0 +1,2150 @@ +[ + { + "type": "text", + "text": "PALM: Predicting Actions through Language Models", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanghwan Kim $^{1}$ , Daoji Huang $^{1}$ , Yongqin Xian $^{2}$ , Otmar Hilliges $^{1}$ , Luc Van Gool $^{1,3,4}$ , and Xi Wang $^{1}$", + "bbox": [ + 316, + 212, + 681, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 ETH Zürich", + "bbox": [ + 450, + 253, + 550, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Google", + "bbox": [ + 468, + 268, + 532, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 KU Leuven", + "bbox": [ + 454, + 282, + 547, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 INSAIT, Sofia", + "bbox": [ + 444, + 296, + 557, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Understanding human activity is a crucial yet intricate task in egocentric vision, a field that focuses on capturing visual perspectives from the camera wearer's viewpoint. Traditional methods heavily rely on representation learning that is trained on a large amount of video data. However, a major challenge arises from the difficulty of obtaining effective video representation. This difficulty stems from the complex and variable nature of human activities, which contrasts with the limited availability of data. In this study, we introduce PALM, an approach that tackles the task of long-term action anticipation, which aims to forecast forthcoming sequences of actions over an extended period. Our method PALM incorporates an action recognition model to track previous action sequences and a vision-language model to articulate relevant environmental details. By leveraging the context provided by these past events, we devise a prompting strategy for action anticipation using large language models (LLMs). Moreover, we implement maximal marginal relevance for example selection to facilitate in-context learning of the LLMs. Our experimental results demonstrate that PALM surpasses the state-of-the-art methods in the task of long-term action anticipation on the Ego4D benchmark. We further validate PALM on two additional benchmarks, affirming its capacity for generalization across intricate activities with different sets of taxonomies.", + "bbox": [ + 261, + 353, + 738, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: egocentric vision $\\cdot$ video understanding $\\cdot$ action anticipation $\\cdot$ large language model", + "bbox": [ + 261, + 657, + 738, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 714, + 374, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As wearable technology, such as smart glasses and body-worn cameras, becomes more prevalent, the availability of egocentric videos has grown significantly. These videos capture activities from the camera wearer's viewpoint, providing a unique perspective that features individual behaviors in various contexts. Egocentric vision holds immense promise for understanding how humans execute specific tasks and activities. It enables numerous applications such as real-time", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/445e7a530ebaf75d625e14375332c35d3b460b207f5ecb592f89e2b6c5cf9ecb.jpg", + "image_caption": [ + "Fig. 1: PALM for long-term action anticipation. Given a video input, PALM uses a vision-language model (VLM) and an action recognition model (ARM) to compile a text prompt that describes the past video content. A large-language model (LLM) then takes the prompt and predicts future actions." + ], + "image_footnote": [], + "bbox": [ + 344, + 146, + 656, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "systems that anticipate and respond to user needs [50, 51, 58], monitoring patient activities [37, 39, 61], and providing personalized instructions and procedures [27, 42, 43].", + "bbox": [ + 212, + 339, + 782, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing works on understanding human activities mostly rely on learning representations from large-scale video datasets [15, 46, 49, 56]. However, these approaches face two critical challenges. First, the complexity and variability of human activities make it challenging to obtain comprehensive representations, particularly for longer videos. Second, reliance on large datasets limits the model's capability to generalize to less common, long-tail classes and unseen scenarios. In this context, there is a pressing need for a more versatile approach to human activity understanding, adept at handling the complicated nature of human activities without overly relying on large-scale video data. Therefore, recent works have started to exploit the procedure knowledge and generalization ability of large language models (LLMs) [1, 21, 32, 44, 45].", + "bbox": [ + 212, + 386, + 784, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most prior works based on LLMs have been largely confined to generating instructions and plans given a high-level task description (e.g., \"prepare the dinner table\") and have not been tailored for intricate human activity understanding from videos. Indeed, there is a significant gap in the literature when it comes to the use of LLMs for grounded activity understanding, especially in challenges involving understanding long-term action sequences. In this work, we efficiently boost the versatile ability of LLMs in long-term action anticipation [12, 18, 20] by providing sufficient visual context in discrete text.", + "bbox": [ + 212, + 551, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, we introduce a simple yet effective framework called PALM that combines descriptions of past action sequences and corresponding narrations to bridge the gap between visual and textual domains. In detail, PALM takes an input video with annotated action periods and leverages a vision-language model and an action recognition model to describe past events. We then design a prompt strategy for large language models (LLMs) to anticipate future actions given the context of these past events. To facilitate the predictive capabilities of LLMs, we use maximal marginal relevance (MMR) [7, 59] to retrieve similar and diverse examples from the training set, enabling in-context learning for the LLMs. Through this prompting strategy, our system can tackle intricate tasks in understanding human activities, benefiting from existing foun", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dition models. Experimental results demonstrate that our method PALM significantly improves performance in long-term action anticipation, as evaluated using the Ego4D benchmark [18]. Further evaluations on two additional benchmarks, EPIC-KITCHEN [12] and EGTEA [33], reveal the generalization ability of PALM across different contexts and datasets. Our contributions can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an effective textural representation of past events, consisting of both action sequences and accompanying narrations that provide additional context and environmental details.", + "- We design a novel prompting strategy that enables efficient in-context learning for large language models.", + "- PALM outperforms existing state-of-the-art methods in the task of long-term action anticipation, which is part of the Ego4D benchmark." + ], + "bbox": [ + 225, + 239, + 785, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 364, + 387, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Foundation Model", + "text_level": 1, + "bbox": [ + 215, + 393, + 413, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation models serve as the cornerstone for various downstream tasks across different data modalities [5,64]. Here we consider two types of foundation models: vision-language models (VLMs) to generate the textual narrations of video and large language models (LLMs) to predict future action given visual context.", + "bbox": [ + 212, + 417, + 785, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language model. VLMs utilize complementary characteristics of image and text datasets to perform multimodal tasks such as image captioning, visual question-answering, and visual reasoning [31, 35, 47]. Various works [10, 57, 65] attempt to unify vision-language tasks with a single model, overcoming the limited utility of previous models. In this paper, the image or video captioning ability of VLMs is utilized to directly transform visual information into narrative text descriptions. We empirically found that BLIP family [11, 29, 30] provides concise and informative captions. For instance, BLIP [30] proposes a novel encoder-decoder architecture specialized for multimodal pertaining, and BLIP-2 [29] significantly reduces the trainable parameters by using a querying transformer to bridge the modality gap between frozen vision models and language models. InstructBLIP [11] further performs instruction tuning on BLIP-2, enabling general-purpose models with a unified natural language interface.", + "bbox": [ + 212, + 478, + 787, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large language model. Language foundation models [3,6,48,53,54,62] have had a significant impact on the NLP field, demonstrating surprising ability to generalize across unseen tasks. With their extensive training data and large parameter size, LLMs have demonstrated the ability to learn from examples provided in input prompts, a concept known as in-context learning [6]. LLMs have also shown success in the tasks of decision-making and planning. For example, SayCan [24] combines a language model with learned low-level skills for grounded planning, while Socratic Models [60] formulate action prediction as text completion problems from captioned images. In this work, we explore how such procedure knowledge can be exploited in the complex action anticipation task and showcase the effectiveness of in-context learning for this purpose.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Long-term Action Anticipation", + "text_level": 1, + "bbox": [ + 215, + 146, + 516, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input video, the long-term action anticipation (LTA) task aims to predict the next actions in the future. In this work, we explore various action recognition models to estimate accurate action sequences and extensively compare our method with previous action anticipation models.", + "bbox": [ + 212, + 174, + 782, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action recognition. SlowFast [16] proposes to fuse two channels of input that consist of a low frame and a high frame rate video, using 3D convolutional kernels. Similar to Slowfast, StillFast [49] adopts a specific architecture with two branches to effectively process high-resolution still images and low-resolution video frames. On the other hand, MViT [15] employs several channel-resolution scale stages to obtain a multiscale pyramid of features, based on vision transformers. Compared to previous works solely optimized on video inputs, EgoVLP [34] and EgoVLPv2 [46] pre-train video and text encoders with contrastive loss specialized for the egocentric domain. Each encoder provides useful feature representations for downstream video tasks such as natural language queries, video question-answering, and video summarization.", + "bbox": [ + 212, + 234, + 782, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action anticipation. Baselines provided by Ego4D [18] employ the SlowFast [16] video encoder to extract features from sampled clips, followed by transformer layers and classification heads to predict the verb-noun pair of future action sequence. HierVL [2] utilizes a hierarchical contrastive training objective that encourages text-visual alignment at both the clip and video levels. ICVAE [36] proposes using a predicted scenario, which summarizes the activity depicted in the video, as a condition for action prediction. A few concurrent approaches [9, 55, 63] leverage the generalization abilities and abundant knowledge of LLMs to predict future actions. Compared to these LLM-based methods, PALM differentiates itself by utilizing vision-language model for effective extraction of visual context and employing in-context learning in combination with MMR.", + "bbox": [ + 212, + 401, + 782, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 608, + 328, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce PALM, a comprehensive framework designed for video understanding and action anticipation. PALM comprises three key modules, as illustrated in Fig. 2: an image captioning module (Sec. 3.2), an action recognition module (Sec. 3.3), and an action anticipation module based on LLMs (Sec. 3.4).", + "bbox": [ + 212, + 642, + 782, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an untrimmed input video, the action recognition module generates labels for past actions in the form of verb and noun pairs (eight past actions in Fig. 2). The action recognition model is trained to produce valid verb and noun pairs as predefined by each dataset. Simultaneously, the image captioning module generates captions from the input video, providing additional visual information relevant to the events, termed visual context. The prompt is then constructed using the visual context and past actions, passed to LLMs to predict future actions. Additionally, we incorporate exemplars from the training set to facilitate in-context learning for LLMs. We demonstrate the effectiveness of our", + "bbox": [ + 212, + 704, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5795b18eb1b49b39c88317000ad4c8ea310959ecea7496a448affae6437be336.jpg", + "image_caption": [ + "Fig. 2: Overview of PALM. Given an input video, an image captioning model and an action recognition model convert visual information of past activities into text format (visual context and past actions). Then, LLM utilizes visual context and past actions to predict the sequence of future actions. The image captioning model is frozen and in-context learning is adapted for LLM prompting." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 776, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "approach on the long-term action anticipation (LTA) task, achieving state-of-the-art (SOTA) performance on Ego4D benchmarks.", + "bbox": [ + 212, + 455, + 784, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Task Description", + "text_level": 1, + "bbox": [ + 214, + 506, + 401, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The LTA task involves predicting a chronological sequence of future actions given an untrimmed video $V$ . The input video $V$ comprises multiple action segments with variable lengths, where each action segment corresponds to a single action (verb and noun pair). The segment boundaries for each action in the videos are provided as part of the input. Formally, after observing the video until time $t$ ( $e.g., V_{:t}$ ), the LTA model sequentially predicts $Z$ future actions that the camerawearer is likely to perform:", + "bbox": [ + 212, + 530, + 787, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\hat {n} _ {z}, \\hat {v} _ {z} \\right\\} _ {z = 1} ^ {Z} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 459, + 635, + 784, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{n}_z\\in \\mathcal{N}$ and $\\hat{v}_z\\in \\mathcal{V}$ , given the predefined noun and verb taxonomy $\\mathcal{N}$ and $\\mathcal{V}$ . The input video $V_{:t}$ contains $P$ action segments.", + "bbox": [ + 212, + 657, + 787, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Image Captioning Module", + "text_level": 1, + "bbox": [ + 214, + 709, + 478, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The past action labels do not capture all the visual information and context presented in the video (e.g., the location of the person or the background objects). Our idea is to use an image captioning model to capture the action context, providing more informative prompts related to the action performer for LLMs.", + "bbox": [ + 212, + 734, + 785, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the descriptive power of captioning models, we formulate the image captioning task as a question-answering problem. Fig. 3 displays a set of questions related to six key concepts that the models should aim to answer given", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "Question Dictionary $= \\{$ Location: Where is the person? Detection: What objects are in this image? Action: What is the person doing in this image? Prediction: What will the person do later in this image? Interaction: What is the person interacting with in this image? Intention: What does the person want to do in this image?", + "guess_lang": "txt", + "bbox": [ + 302, + 147, + 697, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "an input image: location, detection, action, prediction, interaction, and intention. Given a question $Q$ , the conditional text is formed as \"Question: $< Q >$ ? Answer: \" where the answer part will be filled by the captioning model. After extensive experiments, we select intention-based question for conditional generation and VideoBLIP [25] as our image captioning model (see supplementary).", + "bbox": [ + 212, + 290, + 782, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By posing these questions to the captioning models, we anticipate extracting useful descriptions beneficial for LLMs to predict future actions. In practice, the conditional caption generation yields consistent and accurate descriptions of the action context for later prediction. For image captioning, we use the middle frame of each clip segment to generate the captions, based on the assumption that the middle frame is highly likely to be the moment when the action is happening.", + "bbox": [ + 212, + 366, + 787, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Action Recognition Module", + "text_level": 1, + "bbox": [ + 214, + 497, + 491, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cf9fd26667fd9cafbe07949498e02a6febba0d53c62f33862516d35fa84b929.jpg", + "image_caption": [ + "Fig. 3: List of questions asking for information related to key concepts of egocentric image.", + "Fig. 4: Action recognition network. A frozen video encoder generates visual features from $n = 4$ continuous action segments, followed by transformer layers and classification heads." + ], + "image_footnote": [], + "bbox": [ + 334, + 553, + 669, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The accurate past action descriptions are crucial, as they provide essential contextual information about past events in a given video input. Our model recognizes actions from the observed video clips, taking inspiration from the baseline model used in the Ego4D paper [18]. Visual features are extracted from a video encoder given $n$ consecutive action segments ( $n = 4$ in Fig. 4). We adopt", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the off-the-shelf video encoder of EgoVLP [34], a vision-language model pretrained on Ego4D. Next, transformer layers are used to combine multiple visual features into an aggregated feature, followed by $n$ classification heads to recognize the verbs and nouns independently. The video encoder is frozen, and the rest of the model is trained with cross-entropy loss summed over $n$ classification heads. We also try to exploit a text encoder by formulating the action recognition task as matching a proper action label with an input video, which does not improve the accuracy (See supplementary).", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We empirically find that recognizing actions in isolation falls short in effectiveness compared to recognizing $n$ actions within a continuous sequence. This is because actions in neighboring segments are often highly correlated, reflecting the natural continuity embedded in human activities. Thus, we use a sliding input video window that includes $n = 4$ segments with the window striding of one. The top-1 action is determined by identifying the most common action from the classifications across different windows. See supplementary for more details and ablation results.", + "bbox": [ + 212, + 268, + 787, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Action Anticipation Module", + "text_level": 1, + "bbox": [ + 214, + 415, + 493, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we discuss the details of action anticipation with LLMs, from how the prompt is formatted to the details of inference and post-processing of raw outputs. Llama 2 (7B) [54], an open-source LLM, is employed in our framework. Prompt Design. Following common practice, we design the prompt with a few examples to enable in-context learning (ICL), maximizing references given to LLMs. Fig. 5 shows the template of our prompts, which consists of an instruction, a few examples, and a query. The instruction includes the task description and defines the input and output for LLMs. The examples for ICL, selected from the training set, include past and future actions (ground truth) and corresponding textual descriptions generated by the captioning model, which we call narrations. The query includes recognized past actions and generated narrations, and future actions are left empty for LLMs to complete.", + "bbox": [ + 212, + 443, + 787, + 626 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3bae9eac5b178ad79edeacef38b7c564131452287cbde34e626c8ff22efd80b.jpg", + "image_caption": [ + "Fig. 5: Prompt structure for LLMs. The prompt is composed of an instruction, examples retrieved from the training set, and the query related to the prediction video. Each video description consists of narrations, past actions, and future actions." + ], + "image_footnote": [], + "bbox": [ + 228, + 650, + 777, + 801 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Example Selection. Incorporating some examples that demonstrate proper generation will greatly improve the performance of LLMs by providing additional context and references [6, 40]. Considering the diversity of egocentric videos, especially within the Ego4D dataset, randomly selected examples are unlikely to generalize well on various scenarios within the test set.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, we utilize a maximal marginal relevance (MMR) [7, 59] based exemplar selection strategy. The main idea behind MMR is to select exemplars that are relevant to the query while exhibiting sufficient variety to provide robust generalization across the test set. To be specific, we iteratively select a set of examples $p_i \\in T$ from the training set $D$ that are semantically close to the query prompt $q$ but also diverse enough to provide additional information.", + "bbox": [ + 212, + 220, + 784, + 313 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {j} = \\underset {p _ {j} \\in D / T} {\\arg \\max } \\left[ \\lambda S (q, p _ {j}) - (1 - \\lambda) \\underset {p _ {i} \\in T} {\\max } S \\left(p _ {i}, p _ {j}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 321, + 784, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $S$ measures the semantic similarity between exemplars and the parameter $\\lambda$ strikes a balance between similarity and diversity. One exemplar includes past and future action sequences and corresponding captions in natural language. We use MPNet [52] to extract text embeddings and calculate cosine similarity to measure semantic similarity. We compare various exemplar selection strategies in Sec. 4.3.", + "bbox": [ + 212, + 363, + 784, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning LLM. To facilitate LLMs to generate proper action sequence, we fine-tune the LLMs based on the training set of each egocentric dataset. As fine-tuning whole parameters of the LLMs can be computationally inefficient and may lead to catastrophic forgetting [26], parameter-efficient fine-tuning is leveraged. Specifically, we utilize LoRA [19] with 8-bit quantization using a batch size of 8 and a learning rate of $5 \\times 10^{-4}$ for 10 epochs. The training set for fine-tuning consists of input prompts in the format of Fig. 5 with different exemplars and ground truth future actions as targets to be generated by the LLMs.", + "bbox": [ + 212, + 454, + 784, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference of LLM and Post-processing. Following the evaluation protocol of Ego4d [18], we generate multiple predictions for each test video. To reduce the variability in predictions, we prompt LLMs multiple times using the same prompt structure, as shown in Fig. 5, but with different exemplars. After inference, we map the raw output of LLMs to the label space of predefined verbs and nouns. One should note that the output of LLMs may not always fall within the given taxonomy, despite all input prompts containing verbs and nouns from the predefined domain. For simplicity, we adopt a rule-based mapping (see details in supplementary).", + "bbox": [ + 212, + 575, + 784, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate PALM on the long-term action anticipation (LTA) task, shedding light on the various design aspects. We conduct a comprehensive evaluation using the Ego4D dataset, which stands as the largest repository of egocentric videos. We report the final performance metrics on the Ego4D [18], EPIC-KITCHEN-55 [12], and EGTEA Gaze+ [33] benchmarks.", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ace61cc0a2ed5d6ef6d7090d3ae3693362750a927004a389120f98965641eae.jpg", + "table_caption": [ + "Table 1: Comparison to the SOTA on the Ego4D test set. PALM outperforms all baselines on the Ego4D LTA benchmark in respect of edit distance (ED $\\downarrow$ )." + ], + "table_footnote": [], + "table_body": "
MethodVersionVerb EDNoun EDAction ED
Slowfast [18]v10.73890.78000.9432
VCLIP [14]v10.73890.76880.9412
ICVAE [36]v10.74100.73960.9304
HierVL [2]v10.72390.73500.9276
AntGPT [63]v10.65840.65460.8814
PALM (ours)v10.65590.64010.8613
Last actionv20.83290.72580.9438
Repeat actionv20.74910.70730.9155
Retrieve actionv20.74170.70410.9127
Slowfast [18]v20.71690.73590.9253
VideoLLM [9]v20.7210.7250.921
AntGPT [63]v20.65030.64980.8770
PALM (ours)v20.64710.61170.8503
", + "bbox": [ + 303, + 184, + 697, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 431, + 334, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ego4D [18]. With 243 hours of egocentric video content encompassing 53 distinct scenarios, Ego4D offers a rich variety of data with a predefined annotation of 117 verbs and 521 nouns for the LTA task. Ego4D is available in two versions, v1 and v2, where Ego4D v1 is a subset of Ego4D v2 with fewer videos and a smaller verb and noun set. We use both versions of the dataset to enable extensive comparison with previous methods and adhere to the identical training, validation, and test splits as specified in the benchmark.", + "bbox": [ + 212, + 455, + 787, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "EPIC-KITCHENS-55 (EK-55) [12]. This dataset comprises 55 hours of egocentric videos centered around cooking scenarios, featuring 125 annotated verbs and 352 nouns. We follow the same train and test splits as EGO-TOPO [38].", + "bbox": [ + 212, + 561, + 787, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "EGTEA Gaze+ (EGTEA) [33]. With approximately 26 hours of egocentric cooking videos, EGTEA includes 19 verbs and 53 nouns. Similar to EK-55, we employ the same train and test splits as described in EGO-TOPO [38].", + "bbox": [ + 212, + 607, + 787, + 654 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparison to the SOTA", + "text_level": 1, + "bbox": [ + 215, + 675, + 473, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We include three baselines:", + "bbox": [ + 215, + 698, + 500, + 712 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Last action outputs the last action of past action sequence as $Z$ future actions.", + "- Repeat action repeats the given past $P$ actions to fill $Z$ future actions.", + "- Retrieve action retrieves the most semantically similar example in the training set (as in the example selection step) and uses their future actions as the prediction." + ], + "bbox": [ + 223, + 723, + 785, + 814 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also compare to three previous SOTA methods:", + "bbox": [ + 215, + 824, + 586, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Slowfast [18] is the baseline LTA model given by Ego4d.", + "- VideoLLM [9] converts multimodal data into a unified token sequence and utilizes fine-tuned LLMs", + "- AntGPT [63] leveraged action recognition model and fine-tuned LLMs to predict future actions." + ], + "bbox": [ + 223, + 146, + 784, + 220 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "① Single Object Scenario (1 object)", + "text_level": 1, + "bbox": [ + 233, + 260, + 367, + 270 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GT Past", + "text_level": 1, + "bbox": [ + 233, + 313, + 267, + 320 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f2534dc7abb416e8231d85550615eef7afb248d0c96c520f3e7ae3285eb7adfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 273, + 764, + 310 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GT Future ", + "text_level": 1, + "bbox": [ + 233, + 324, + 269, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick).", + "bbox": [ + 274, + 311, + 617, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(0.55)", + "text_level": 1, + "bbox": [ + 233, + 349, + 259, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (take, brick).", + "bbox": [ + 274, + 323, + 750, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(0.55)", + "text_level": 1, + "bbox": [ + 233, + 349, + 259, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(put, brick), (put, brick), (move, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick).", + "bbox": [ + 274, + 332, + 658, + 340 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SlowFast", + "text_level": 1, + "bbox": [ + 233, + 345, + 269, + 352 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "put, brick), (put, brick), (put, brick), (adjust, brick), (take, brick), (take, brick), (take, brick), (take, brick), (put, brick)", + "bbox": [ + 274, + 343, + 743, + 352 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(0.55)", + "text_level": 1, + "bbox": [ + 233, + 354, + 259, + 359 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(take, brick), (adjust, brick), (put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick).", + "bbox": [ + 274, + 353, + 656, + 361 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "PALM", + "text_level": 1, + "bbox": [ + 233, + 364, + 259, + 372 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick), (take, brick),", + "bbox": [ + 274, + 364, + 730, + 372 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(0.4)", + "text_level": 1, + "bbox": [ + 233, + 375, + 259, + 380 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick).", + "bbox": [ + 274, + 373, + 638, + 380 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "② Few Object Scenario (2~5 objects)", + "text_level": 1, + "bbox": [ + 233, + 388, + 370, + 398 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6d539fde5b9c097e01d4f2f0a50e8e85fde79cfa9fd4526e662ee6c05dad0c76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 400, + 766, + 436 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GTPast", + "text_level": 1, + "bbox": [ + 233, + 440, + 267, + 448 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(dip, paintbrush), (put, container), (take, paintbrush), (put, paintbrush), (adjust, paintbrush), (take, container), (dip, container), (paint, wall).", + "bbox": [ + 274, + 440, + 736, + 449 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GT Futur", + "text_level": 1, + "bbox": [ + 233, + 452, + 269, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(paint, wall), (dip, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (dip, container), (paint, wall)", + "bbox": [ + 274, + 452, + 759, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(paint, paintbrush), (paint, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush).", + "bbox": [ + 274, + 462, + 725, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SlowFast", + "text_level": 1, + "bbox": [ + 233, + 479, + 269, + 486 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "take, container), (put, container), (hold, container), (put, string), (open, brush), (turn, container), (touch, nail gun), (hold, container), (throw, soap),", + "bbox": [ + 274, + 479, + 763, + 488 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "take, cereal), move, sheft), (push, wire), (break, cloth), (break, wire), (hold, circuit), (hold, container), (hold, garbage can), (remove, switch), (put, camera). (hold, cloth).", + "bbox": [ + 274, + 489, + 746, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "PALM", + "text_level": 1, + "bbox": [ + 233, + 506, + 259, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "put paintbrush). (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall),", + "bbox": [ + 274, + 505, + 741, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(0.6)", + "text_level": 1, + "bbox": [ + 233, + 515, + 259, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(dip, paintbrush), (paint, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall).", + "bbox": [ + 274, + 515, + 705, + 522 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "③ Many Object Scenario (>5 objects)", + "text_level": 1, + "bbox": [ + 233, + 529, + 370, + 537 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/49f1a59419bf3967b1417737113621f4b758ae98cd718115618972a64b38d786.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 540, + 764, + 577 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GT Past", + "text_level": 1, + "bbox": [ + 233, + 580, + 267, + 587 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "put, screwdriver), (put, bench), (take, wire), (attach, wire), (adjust, wire), (attach, wire), (hold, wire), (hold, wire)", + "bbox": [ + 274, + 580, + 648, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GT Future", + "text_level": 1, + "bbox": [ + 233, + 592, + 269, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(squeeze, wire), (adjust, wire), (hold, wire), (move, stool), (sit, stool), (take, screwdriver), (serve, mower), (put, screwdriver), (take, nut),", + "bbox": [ + 274, + 592, + 728, + 601 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(hit, mower), (take, screwdriver), (move, screwdriver), (repair, mower), (put, screwdriver), (repair, mower), (repair, mower), (hold, screwdriver), (take, screwdriver), (repair, mower), (move, screwdriver).", + "bbox": [ + 274, + 602, + 750, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SlowFast", + "text_level": 1, + "bbox": [ + 233, + 619, + 269, + 626 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(attach, wheel), (turn, pipe), (take, valve), (put, wheel), (put, sticker), (take, drawer), (put, wheel), (repair, torch), (hold, wheel), (repair, plier),", + "bbox": [ + 274, + 619, + 746, + 627 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(1.0)", + "text_level": 1, + "bbox": [ + 233, + 628, + 259, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(put, screwdriver), (attach, wheel), (repair, container), (put, wheel), (put, wheel), (attach, seasoning), (take, wheel), (take, cutter), (repair, screw)", + "bbox": [ + 274, + 628, + 748, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "PALM", + "text_level": 1, + "bbox": [ + 233, + 646, + 259, + 652 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "move, screwdriver), (move, screwdriver), (take, wrench), (put, screwdriver), (move, screwdriver), (take, screwdriver), (move, screw", + "(move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver)" + ], + "bbox": [ + 274, + 645, + 743, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fig. 6: Qualitative results of predictions on Ego4D. We show the comparison between the predictions from the Ego4D baseline using SlowFast [18] and the predictions from our model PALM respectively. For each of the three examples shown here, we include the middle frames of the previous eight action segments with the corresponding ground truth past actions GT Past and future actions GT Future alongside the predicted future actions. The green words represent correct predictions while the red words denote incorrect predictions. The edit distance between the predicted actions and ground truth actions is shown in parenthesis.", + "bbox": [ + 212, + 684, + 787, + 795 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation. We conform to the evaluation procedures set forth by the Ego4D [18] and calculate the edit distance (ED) using the Damerau-Levenshtein distance", + "bbox": [ + 212, + 809, + 789, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 218, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "metric [13, 28]. ED is computed separately for verbs, nouns, and the entire action sequences. Given $P = 8$ action segments of the input video, we report the minimum edit distance among $K = 5$ predicted sequences, each with a length of $Z = 20$ actions.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Tab. 1 shows that PALM improves existing SOTA methods and achieves the best performance on the test set of the Ego4D v1 and v2 benchmark. In general, LLM-based approaches (VideoLLM, AntGPT, and PALM) predict more accurate future actions compared to other approaches based on representation learning. Notably, repeat and retrieve action baselines in Ego4D v2 perform better than Slowfast with respect to noun and action ED. These observations clearly indicate the advantages of leveraging foundation models and the limitations of feature learning methods on the LTA task. Moreover, PALM outperforms other LLM-based methods, which demonstrates the effectiveness of our approach to provide ample visual context to LLMs, equipped with action recognition and image captioning models.", + "bbox": [ + 212, + 205, + 789, + 372 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 6, we perform a qualitative comparison between the Ego4D baseline using SlowFast [18] and PALM. The first example exemplifies a scenario where the person mainly interacts with a single object, in this case, a brick. Both methods can accurately predict the future noun \"brick\". In the second scenario, the person is drawing a painting on a wall, interacting with multiple objects such as a paintbrush, a container, and the wall. PALM predicts the repeated sequence of actions, unlike SlowFast which incorrectly anticipates unrelated actions such as putting a string, touching a nail gun, or throwing a soap, none of which are present in the input video. This also demonstrates the importance of previous actions in the anticipation of future actions. The final scenario portrays the person adjusting a wire in a repair shop, a task involving numerous items. While SlowFast and PALM recognize central objects such as a screwdriver, they both fail to predict a coherent sequence of repair actions. We hypothesize that a language model with expert knowledge can produce more meaningful anticipations, particularly in the context of repairing mechanics, where it is crucial to execute actions in a precise order. More qualitative results are shown in supplementary.", + "bbox": [ + 212, + 373, + 787, + 614 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablations", + "text_level": 1, + "bbox": [ + 215, + 635, + 344, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We perform various ablation studies regarding action recognition, example selection methods, LLM models, and prompt contents. We evaluate on the validation set of Ego4D v2.", + "bbox": [ + 212, + 657, + 785, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Action Recognition. The action recognition module is an essential part of our framework PALM as we utilize LLMs for predicting future actions and the accuracy of these predictions is highly influenced by the sequence of previous actions. We evaluate various video encoders trained on egocentric videos, including MViT [15], SlowFast [16], StillFast [49], EgoVLP [34], and EgoVLPv2 [46]. It is important to note that we employ a frozen video encoder without any finetuning, as depicted in Fig. 4.", + "bbox": [ + 212, + 704, + 787, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 2 presents the Top-1 Accuracy $(\\%)$ and the edit distance of each video encoder. EgoVLP emerges as the top performer, achieving the best accuracy", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/87a3f883b0c447225f0e67207e4fea2101db399cdfb168271c1d15c08c8de82d.jpg", + "table_caption": [ + "Table 2: Ablation on action recognition module. We compare different pretrained video encoders by reporting the Top-1 accuracy (Acc. ↑) and edit distance (ED ↓) for verbs, nouns, and actions. For comparative purposes, we also include the ED when ground truth actions (GT action) are used in the prompts. The best Acc. and ED are highlighted in bold except those from GT action." + ], + "table_footnote": [], + "table_body": "
MethodVerb Acc.Noun Acc.Action Acc.Verb EDNoun EDAction ED
MViT [15]19.872.550.510.72910.85970.9630
SlowFast [16]19.4214.653.120.72520.74720.9278
StillFast [49]19.1219.414.060.73560.73690.9241
EgoVLPv2 [46]33.8440.6316.310.71820.66510.8937
EgoVLP [34]40.3245.5320.630.71110.64650.8819
GT action---0.66630.58870.8319
", + "bbox": [ + 225, + 226, + 767, + 340 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and edit distance among all encoders. EgoVLP utilizes a custom contrastive loss to align egocentric video frames with narrations, effectively establishing a connection between visual and textual embeddings. However, EgoVLPv2 utilizes task-specific loss for question-answering tasks built on the EgoVLP setting, which results in slightly less informative video features for action recognition. Unlike EgoVLP and EgoVLPv2, other video encoders are solely trained on video frames, making it challenging to draw robust connections between their video features and textual information (verbs and nouns). Based on these results, we select EgoVLP as our video encoder.", + "bbox": [ + 212, + 354, + 787, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "It is expected that better accuracy in action recognition leads to a reduced edit distance, as accurate estimation of past actions is crucial for predicting future actions. Notably, the use of ground truth (GT) actions significantly outperforms all video encoders, which highlights the enhancement of action recognition as the foremost challenge within our framework. We believe that fine-tuning the video encoder on specific egocentric datasets would likely enhance recognition accuracy.", + "bbox": [ + 212, + 489, + 787, + 597 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a8f6eae55a9cf4d763a210428e7cb29fe20b6373fa8275da64a3e4419ddcba21.jpg", + "table_caption": [ + "Table 3: Ablation on example selection strategies exhibits that MMR outperforms other selection methods." + ], + "table_footnote": [], + "table_body": "
MethodVerb EDNoun EDAction ED
Random0.73620.67470.9002
Similarity0.72040.66610.8923
MMR0.71790.66470.8908
", + "bbox": [ + 359, + 662, + 638, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Example Selection. We compare different example selection strategies for in-context learning (ICL), including random, similarity, and maximal marginal relevance (MMR) [59]. The random method uniformly selects examples from the training set, whereas the similarity method selects the most semantically similar example to the query. MMR considers both the similarity and diversity of", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "examples, striving for a balance between them. Tab. 3 demonstrates that MMR outperforms the other selection methods.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a1c2c07655192ee2a55a80e05b622c9a32718ef4dcb0ffd973c89ce4d79a18fd.jpg", + "table_caption": [ + "Table 4: Ablation on LLMs shows that larger LLMs perform better on the LTA task." + ], + "table_footnote": [], + "table_body": "
MethodVerb EDNoun EDAction ED
GPT-Neo (125M)0.76570.68670.9119
GPT-Neo (1.3B)0.73140.67590.8970
GPT-Neo (2.7B)0.71780.67470.8943
Llama 2 (7B)0.71790.66470.8908
Llama 2 (13B)0.69580.66340.8867
", + "bbox": [ + 323, + 247, + 674, + 347 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "LLM Ablation. We ablate the impact of LLMs by comparing different open-source models, primarily GPT-Neo [4] and Llama 2 [54] families. Tab. 4 indicates that larger language models are more capable of predicting future actions with their improved capacity for generalization across novel scenarios. This observation aligns with the performance of LLMs on general NLP benchmarks. Taking into account the computational resources at our disposal, we select Llama 2 with 7B parameters.", + "bbox": [ + 212, + 378, + 784, + 484 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8d150a7ea9f45aedd850192c1737fabada4f7ca974ed3f3d8141345bf3da53d5.jpg", + "table_caption": [ + "Table 5: Ablation on the content of the prompt. We use abbreviations for each content in the input prompt: $\\mathcal{A}$ - action sequence, $\\mathcal{C}$ - captions" + ], + "table_footnote": [], + "table_body": "
Prompt ContentVerb EDNoun EDAction ED
A0.72030.66880.8947
C0.70880.69010.9064
A+C0.71790.66470.8908
", + "bbox": [ + 338, + 554, + 658, + 627 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Prompt Contents. We examine different content included in the input prompt to validate the contribution of each module to final performance. PALM utilizes recognized past actions and generated image captions to compose prompts. The action sequence describes the temporal flow of activity in (verb, noun) format, while captions narrate visual information relevant to human. In Tab. 5, the action sequence with captions $(\\mathcal{A} + \\mathcal{C})$ achieves the lowest action ED, which thereby is used in our framework. Notably, caption-only prompt $(\\mathcal{C})$ reaches the lowest verb ED, showing that narrations are capable of capturing dynamic motions of human activity, while action-only prompt $(\\mathcal{A})$ effectively detects key objects interacted with human supported by relatively lower noun ED. This indicates that action sequence and narrations provide a beneficial complement, thereby supporting our decision to incorporate both in the prompts.", + "bbox": [ + 212, + 657, + 784, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a8b3765327af2b14f1ee32366ec58a9b3915e488b302e2ac5795d293b09edc21.jpg", + "table_caption": [ + "Table 6: Comparison to the SOTA on EK-55 and EGTEA. Performance is evaluated using the mAP metric $(\\uparrow)$ ." + ], + "table_footnote": [], + "table_body": "
MethodEK-55EGTEA
ALLFREQRAREALLFREQRARE
I3D [8]32.753.323.072.179.353.3
ActionVLAD [17]29.853.518.673.379.058.6
Timeception [22]35.655.926.174.179.759.7
VideoGraph [23]22.549.414.067.777.147.2
EGO-TOPO [38]38.056.929.273.580.754.7
Anticipatr [41]39.158.129.176.883.355.1
AntGPT [63]40.158.831.980.284.872.9
PALM (ours)40.459.330.380.785.073.5
", + "bbox": [ + 287, + 184, + 709, + 334 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Evaluation on EK-55 and EGTEA", + "text_level": 1, + "bbox": [ + 215, + 359, + 542, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Evaluation. For EK-55 [12] and EGTEA [33], we adhere to the experimental setup established by EGO-TOPO [38] and employ mean average precision (mAP) for multi-label classification as the evaluation metric. The initial $R\\%$ ( $R = [25\\%, 50\\%, 75\\%]$ ) of each video serves as input, with the objective of predicting the set of actions occurring in the remaining $(100 - R)\\%$ of the video. Notably, actions here are confined to verb predictions only without noun, following previous conventions. We calculate the mAP on the validation set averaged over all $R$ , examining various performance aspects, such as all target actions (ALL), frequently occurring actions (FREQ), and rarely occurring actions (RARE).", + "bbox": [ + 212, + 397, + 785, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 6, our method achieves SOTA results in the ALL and FREQ section while exhibiting comparable performance in the RARE section. The results demonstrate the effectiveness of PALM by capturing additional visual information with image captioning models. This aligns with the ablation study regarding prompt contents in Sec. 4.3 where captions contribute better on verb predictions. We observe that verb in RARE section are seldom represented in the captions, which results in relatively lower mAP in RARE section. Thus, we believe fine-tuning image captioning model or employing better video captioning model will further improve the performance accordingly.", + "bbox": [ + 212, + 534, + 785, + 671 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 689, + 359, + 705 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present PALM, a framework that leverages vision-language and large language models for long-term action anticipation. PALM achieves the best performance on the Ego4D benchmark, demonstrating the effectiveness of performing the LTA task in the language domain and the benefits of using foundation models. Our model also shows competitive results on two other benchmarks, EK-55 and EGTEA. As a result, we are optimistic that our work will spur further research in using commonsense knowledge extracted from LLMs for video understanding.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022) 2", + "2. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023) 4, 9", + "3. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. If you use this software, please cite it using these metadata 58 (2021) 3", + "4. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. https://doi.org/10.5281/zenodo.5297715 (2021). https://doi.org/10.5281/ zenodo.5297715, https://doi.org/10.5281/zenodo.5297715, If you use this software, please cite it using these metadata. 13", + "5. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021) 3", + "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 3, 8", + "7. Carbonell, J., Goldstein, J.: The use of mmr, diversity-based reranking for reordering documents and producing summaries. In: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval. pp. 335-336 (1998) 2, 8", + "8. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 14", + "9. Chen, G., Zheng, Y.D., Wang, J., Xu, J., Huang, Y., Pan, J., Wang, Y., Wang, Y., Qiao, Y., Lu, T., et al.: Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292 (2023) 4, 9, 10", + "0. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021) 3", + "1. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) 3", + "2. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(11), 4125-4141 (2020) 2, 3, 8, 9, 14", + "3. Damerau, F.J.: A technique for computer detection and correction of spelling errors. Communications of the ACM 7(3), 171-176 (1964) 11", + "4. Das, S., Ryoo, M.S.: Video+ clip baseline for ego4d long-term action anticipation. arXiv preprint arXiv:2207.00579 (2022) 9", + "5. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 2, 4, 11, 12" + ], + "bbox": [ + 225, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) 4, 11, 12", + "17. Girdhar, R., Ramanan, D., Gupta, A., Sivic, J., Russell, B.: Actionvlad: Learning spatio-temporal aggregation for action classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 971-980 (2017) 14", + "18. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022) 2, 3, 4, 6, 8, 9, 10, 11", + "19. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) 8", + "20. Huang, D., Hilliges, O., Van Gool, L., Wang, X.: Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545 (2023) 2", + "21. Huang, W., Abbeel, P., Pathak, D., Mordatch, I.: Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In: International Conference on Machine Learning. pp. 9118-9147. PMLR (2022) 2", + "22. Hussein, N., Gavves, E., Smeulders, A.W.: Timeception for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 254-263 (2019) 14", + "23. Hussein, N., Gavves, E., Smeulders, A.W.: Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143 (2019) 14", + "24. Ichter, B., Brohan, A., Chebotar, Y., Finn, C., Hausman, K., Herzog, A., Ho, D., Ibarz, J., Irpan, A., Jang, E., et al.: Do as i can, not as i say: Grounding language in robotic affordances. In: 6th Annual Conference on Robot Learning (2022) 3", + "25. Keunwoo Peter, Y.: Videooblip is a large vision-language model based on blip-2 that can generate texts conditioned on videos. https://github.com/yukw777/VideoBLIP (2021). https://doi.org/https://orcid.org/0000-0002-4387-2410, https://github.com/yukw777/VideoBLIP, If you use this software, please cite it using these metadata. 6", + "26. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 8", + "27. Leo, M., Medioni, G., Trivedi, M., Kanade, T., Farinella, G.M.: Computer vision for assistive technologies. Computer Vision and Image Understanding 154, 1-15 (2017) 2", + "28. Levenshtein, V.I., et al.: Binary codes capable of correcting deletions, insertions, and reversals. In: Soviet physics doklady. vol. 10, pp. 707-710. Soviet Union (1966) 11", + "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 3", + "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 3" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019) 3", + "32. Li, S., Puig, X., Paxton, C., Du, Y., Wang, C., Fan, L., Chen, T., Huang, D.A., Akyurek, E., Anandkumar, A., et al.: Pre-trained language models for interactive decision-making. Advances in Neural Information Processing Systems 35, 31199-31212 (2022) 2", + "33. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018) 3, 8, 9, 14", + "34. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022) 4, 7, 11, 12", + "35. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 3", + "36. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023) 4, 9", + "37. Meditskos, G., Plans, P.M., Stavropoulos, T.G., Benois-Pineau, J., Buso, V., Kompatsiaris, I.: Multi-modal activity recognition from egocentric vision, semantic enrichment and lifelogging applications for the care of dementia. Journal of Visual Communication and Image Representation 51, 169-190 (2018) 2", + "38. Nagarajan, T., Li, Y., Feichtenhofer, C., Grauman, K.: Ego-topo: Environment affordances from egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 163-172 (2020) 9, 14", + "39. Nakazawa, A., Honda, M.: First-person camera system to evaluate tender dementia-care skill. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) 2", + "40. Naveed, H., Khan, A.U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Barnes, N., Mian, A.: A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 (2023) 8", + "41. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022) 14", + "42. Neumann, L., Zisserman, A., Vedaldi, A.: Future event prediction: If and when. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) 2", + "43. OhnBar, E., Kitani, K., Asakawa, C.: Personalized dynamics models for adaptive assistive navigation systems. In: Conference on Robot Learning. pp. 16-39. PMLR (2018) 2", + "44. Pasca, R.G., Gavryushin, A., Hamza, M., Kuo, Y.L., Mo, K., Van Gool, L., Hilliges, O., Wang, X.: Summarize the past to predict the future: Natural language descriptions of context boost multimodal object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18286-18296 (2024) 2", + "45. Patel, D., Eghbalzadeh, H., Kamra, N., Iuzzolino, M.L., Jain, U., Desai, R.: Pre-trained language models as visual planners for human assistance. arXiv preprint arXiv:2304.09179 (2023) 2" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023) 2, 4, 11, 12", + "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 3", + "48. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners (2019) 3", + "49. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023) 2, 4, 11, 12", + "50. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021) 2", + "51. Ryoo, M.S., Fuchs, T.J., Xia, L., Aggarwal, J.K., Matthies, L.: Robot-centric activity prediction from first-person videos: What will they do to me? In: Proceedings of the tenth annual ACM/IEEE international conference on human-robot interaction. pp. 295–302 (2015) 2", + "52. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020) 8", + "53. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 3", + "54. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 3, 7, 13", + "55. Wang, S., Zhao, Q., Do, M.Q., Agarwal, N., Lee, K., Sun, C.: Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627 (2023) 4", + "56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 2", + "57. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. In: International Conference on Learning Representations (2021) 3", + "58. Yao, Y., Xu, M., Choi, C., Crandall, D.J., Atkins, E.M., Dariush, B.: Egocentric vision-based future vehicle localization for intelligent driving assistance systems. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9711-9717. IEEE (2019) 2", + "59. Ye, X., Iyer, S., Celikyilmaz, A., Stoyanov, V., Durrett, G., Pasunuru, R.: Complementary explanations for effective in-context learning. arXiv preprint arXiv:2211.13892 (2022) 2, 8, 12", + "60. Zeng, A., Attarian, M., Choromanski, K.M., Wong, A., Welker, S., Tombari, F., Purohit, A., Ryoo, M.S., Sindhwani, V., Lee, J., et al.: Socratic models: Composing zero-shot multimodal reasoning with language. In: The Eleventh International Conference on Learning Representations (2022) 3" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Zhan, K., Faux, S., Ramos, F.: Multi-scale conditional random fields for first-person activity recognition. In: 2014 IEEE international conference on pervasive computing and communications (PerCom). pp. 51-59. IEEE (2014) 2", + "62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 3", + "63. Zhao, Q., Zhang, C., Wang, S., Fu, C., Agarwal, N., Lee, K., Sun, C.: Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368 (2023) 4, 9, 10, 14", + "64. Zhou, C., Li, Q., Li, C., Yu, J., Liu, Y., Wang, G., Zhang, K., Ji, C., Yan, Q., He, L., et al.: A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419 (2023) 3", + "65. Zhou, L., Palangi, H., Zhang, L., Hu, H., Corso, J., Gao, J.: Unified vision-language pre-training for image captioning and vqa. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13041-13049 (2020) 3" + ], + "bbox": [ + 212, + 146, + 787, + 354 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PALM: Predicting Actions through Language Models", + "bbox": [ + 375, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_model.json b/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e5a00030d499859563b0aab216f245ee39cd2386 --- /dev/null +++ b/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_model.json @@ -0,0 +1,2955 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.243, + 0.141, + 0.761, + 0.184 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.213, + 0.682, + 0.244 + ], + "angle": 0, + "content": "Sanghwan Kim\\(^{1}\\), Daoji Huang\\(^{1}\\), Yongqin Xian\\(^{2}\\), Otmar Hilliges\\(^{1}\\), Luc Van Gool\\(^{1,3,4}\\), and Xi Wang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.452, + 0.255, + 0.551, + 0.269 + ], + "angle": 0, + "content": "1 ETH Zürich" + }, + { + "type": "text", + "bbox": [ + 0.47, + 0.27, + 0.534, + 0.283 + ], + "angle": 0, + "content": "2 Google" + }, + { + "type": "text", + "bbox": [ + 0.455, + 0.284, + 0.548, + 0.297 + ], + "angle": 0, + "content": "3 KU Leuven" + }, + { + "type": "text", + "bbox": [ + 0.446, + 0.297, + 0.558, + 0.311 + ], + "angle": 0, + "content": "4 INSAIT, Sofia" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.354, + 0.74, + 0.644 + ], + "angle": 0, + "content": "Abstract. Understanding human activity is a crucial yet intricate task in egocentric vision, a field that focuses on capturing visual perspectives from the camera wearer's viewpoint. Traditional methods heavily rely on representation learning that is trained on a large amount of video data. However, a major challenge arises from the difficulty of obtaining effective video representation. This difficulty stems from the complex and variable nature of human activities, which contrasts with the limited availability of data. In this study, we introduce PALM, an approach that tackles the task of long-term action anticipation, which aims to forecast forthcoming sequences of actions over an extended period. Our method PALM incorporates an action recognition model to track previous action sequences and a vision-language model to articulate relevant environmental details. By leveraging the context provided by these past events, we devise a prompting strategy for action anticipation using large language models (LLMs). Moreover, we implement maximal marginal relevance for example selection to facilitate in-context learning of the LLMs. Our experimental results demonstrate that PALM surpasses the state-of-the-art methods in the task of long-term action anticipation on the Ego4D benchmark. We further validate PALM on two additional benchmarks, affirming its capacity for generalization across intricate activities with different sets of taxonomies." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.659, + 0.74, + 0.687 + ], + "angle": 0, + "content": "Keywords: egocentric vision \\(\\cdot\\) video understanding \\(\\cdot\\) action anticipation \\(\\cdot\\) large language model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.715, + 0.375, + 0.731 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As wearable technology, such as smart glasses and body-worn cameras, becomes more prevalent, the availability of egocentric videos has grown significantly. These videos capture activities from the camera wearer's viewpoint, providing a unique perspective that features individual behaviors in various contexts. Egocentric vision holds immense promise for understanding how humans execute specific tasks and activities. It enables numerous applications such as real-time" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.147, + 0.658, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 1: PALM for long-term action anticipation. Given a video input, PALM uses a vision-language model (VLM) and an action recognition model (ARM) to compile a text prompt that describes the past video content. A large-language model (LLM) then takes the prompt and predicts future actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.385 + ], + "angle": 0, + "content": "systems that anticipate and respond to user needs [50, 51, 58], monitoring patient activities [37, 39, 61], and providing personalized instructions and procedures [27, 42, 43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.552 + ], + "angle": 0, + "content": "Existing works on understanding human activities mostly rely on learning representations from large-scale video datasets [15, 46, 49, 56]. However, these approaches face two critical challenges. First, the complexity and variability of human activities make it challenging to obtain comprehensive representations, particularly for longer videos. Second, reliance on large datasets limits the model's capability to generalize to less common, long-tail classes and unseen scenarios. In this context, there is a pressing need for a more versatile approach to human activity understanding, adept at handling the complicated nature of human activities without overly relying on large-scale video data. Therefore, recent works have started to exploit the procedure knowledge and generalization ability of large language models (LLMs) [1, 21, 32, 44, 45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Most prior works based on LLMs have been largely confined to generating instructions and plans given a high-level task description (e.g., \"prepare the dinner table\") and have not been tailored for intricate human activity understanding from videos. Indeed, there is a significant gap in the literature when it comes to the use of LLMs for grounded activity understanding, especially in challenges involving understanding long-term action sequences. In this work, we efficiently boost the versatile ability of LLMs in long-term action anticipation [12, 18, 20] by providing sufficient visual context in discrete text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 1, we introduce a simple yet effective framework called PALM that combines descriptions of past action sequences and corresponding narrations to bridge the gap between visual and textual domains. In detail, PALM takes an input video with annotated action periods and leverages a vision-language model and an action recognition model to describe past events. We then design a prompt strategy for large language models (LLMs) to anticipate future actions given the context of these past events. To facilitate the predictive capabilities of LLMs, we use maximal marginal relevance (MMR) [7, 59] to retrieve similar and diverse examples from the training set, enabling in-context learning for the LLMs. Through this prompting strategy, our system can tackle intricate tasks in understanding human activities, benefiting from existing foun" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.376, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "dition models. Experimental results demonstrate that our method PALM significantly improves performance in long-term action anticipation, as evaluated using the Ego4D benchmark [18]. Further evaluations on two additional benchmarks, EPIC-KITCHEN [12] and EGTEA [33], reveal the generalization ability of PALM across different contexts and datasets. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.24, + 0.787, + 0.284 + ], + "angle": 0, + "content": "- We propose an effective textural representation of past events, consisting of both action sequences and accompanying narrations that provide additional context and environmental details." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.285, + 0.787, + 0.314 + ], + "angle": 0, + "content": "- We design a novel prompting strategy that enables efficient in-context learning for large language models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.315, + 0.787, + 0.345 + ], + "angle": 0, + "content": "- PALM outperforms existing state-of-the-art methods in the task of long-term action anticipation, which is part of the Ego4D benchmark." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.24, + 0.787, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.366, + 0.388, + 0.381 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.395, + 0.414, + 0.409 + ], + "angle": 0, + "content": "2.1 Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.787, + 0.478 + ], + "angle": 0, + "content": "Foundation models serve as the cornerstone for various downstream tasks across different data modalities [5,64]. Here we consider two types of foundation models: vision-language models (VLMs) to generate the textual narrations of video and large language models (LLMs) to predict future action given visual context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Vision-language model. VLMs utilize complementary characteristics of image and text datasets to perform multimodal tasks such as image captioning, visual question-answering, and visual reasoning [31, 35, 47]. Various works [10, 57, 65] attempt to unify vision-language tasks with a single model, overcoming the limited utility of previous models. In this paper, the image or video captioning ability of VLMs is utilized to directly transform visual information into narrative text descriptions. We empirically found that BLIP family [11, 29, 30] provides concise and informative captions. For instance, BLIP [30] proposes a novel encoder-decoder architecture specialized for multimodal pertaining, and BLIP-2 [29] significantly reduces the trainable parameters by using a querying transformer to bridge the modality gap between frozen vision models and language models. InstructBLIP [11] further performs instruction tuning on BLIP-2, enabling general-purpose models with a unified natural language interface." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Large language model. Language foundation models [3,6,48,53,54,62] have had a significant impact on the NLP field, demonstrating surprising ability to generalize across unseen tasks. With their extensive training data and large parameter size, LLMs have demonstrated the ability to learn from examples provided in input prompts, a concept known as in-context learning [6]. LLMs have also shown success in the tasks of decision-making and planning. For example, SayCan [24] combines a language model with learned low-level skills for grounded planning, while Socratic Models [60] formulate action prediction as text completion problems from captioned images. In this work, we explore how such procedure knowledge can be exploited in the complex action anticipation task and showcase the effectiveness of in-context learning for this purpose." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.517, + 0.163 + ], + "angle": 0, + "content": "2.2 Long-term Action Anticipation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.783, + 0.234 + ], + "angle": 0, + "content": "Given an input video, the long-term action anticipation (LTA) task aims to predict the next actions in the future. In this work, we explore various action recognition models to estimate accurate action sequences and extensively compare our method with previous action anticipation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.784, + 0.4 + ], + "angle": 0, + "content": "Action recognition. SlowFast [16] proposes to fuse two channels of input that consist of a low frame and a high frame rate video, using 3D convolutional kernels. Similar to Slowfast, StillFast [49] adopts a specific architecture with two branches to effectively process high-resolution still images and low-resolution video frames. On the other hand, MViT [15] employs several channel-resolution scale stages to obtain a multiscale pyramid of features, based on vision transformers. Compared to previous works solely optimized on video inputs, EgoVLP [34] and EgoVLPv2 [46] pre-train video and text encoders with contrastive loss specialized for the egocentric domain. Each encoder provides useful feature representations for downstream video tasks such as natural language queries, video question-answering, and video summarization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.581 + ], + "angle": 0, + "content": "Action anticipation. Baselines provided by Ego4D [18] employ the SlowFast [16] video encoder to extract features from sampled clips, followed by transformer layers and classification heads to predict the verb-noun pair of future action sequence. HierVL [2] utilizes a hierarchical contrastive training objective that encourages text-visual alignment at both the clip and video levels. ICVAE [36] proposes using a predicted scenario, which summarizes the activity depicted in the video, as a condition for action prediction. A few concurrent approaches [9, 55, 63] leverage the generalization abilities and abundant knowledge of LLMs to predict future actions. Compared to these LLM-based methods, PALM differentiates itself by utilizing vision-language model for effective extraction of visual context and employing in-context learning in combination with MMR." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.33, + 0.625 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.783, + 0.703 + ], + "angle": 0, + "content": "We introduce PALM, a comprehensive framework designed for video understanding and action anticipation. PALM comprises three key modules, as illustrated in Fig. 2: an image captioning module (Sec. 3.2), an action recognition module (Sec. 3.3), and an action anticipation module based on LLMs (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Given an untrimmed input video, the action recognition module generates labels for past actions in the form of verb and noun pairs (eight past actions in Fig. 2). The action recognition model is trained to produce valid verb and noun pairs as predefined by each dataset. Simultaneously, the image captioning module generates captions from the input video, providing additional visual information relevant to the events, termed visual context. The prompt is then constructed using the visual context and past actions, passed to LLMs to predict future actions. Additionally, we incorporate exemplars from the training set to facilitate in-context learning for LLMs. We demonstrate the effectiveness of our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.376, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.144, + 0.777, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 2: Overview of PALM. Given an input video, an image captioning model and an action recognition model convert visual information of past activities into text format (visual context and past actions). Then, LLM utilizes visual context and past actions to predict the sequence of future actions. The image captioning model is frozen and in-context learning is adapted for LLM prompting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.785, + 0.486 + ], + "angle": 0, + "content": "approach on the long-term action anticipation (LTA) task, achieving state-of-the-art (SOTA) performance on Ego4D benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.403, + 0.523 + ], + "angle": 0, + "content": "3.1 Task Description" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.636 + ], + "angle": 0, + "content": "The LTA task involves predicting a chronological sequence of future actions given an untrimmed video \\( V \\). The input video \\( V \\) comprises multiple action segments with variable lengths, where each action segment corresponds to a single action (verb and noun pair). The segment boundaries for each action in the videos are provided as part of the input. Formally, after observing the video until time \\( t \\) (\\( e.g., V_{:t} \\)), the LTA model sequentially predicts \\( Z \\) future actions that the camerawearer is likely to perform:" + }, + { + "type": "equation", + "bbox": [ + 0.46, + 0.636, + 0.785, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\hat {n} _ {z}, \\hat {v} _ {z} \\right\\} _ {z = 1} ^ {Z} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.69 + ], + "angle": 0, + "content": "where \\(\\hat{n}_z\\in \\mathcal{N}\\) and \\(\\hat{v}_z\\in \\mathcal{V}\\), given the predefined noun and verb taxonomy \\(\\mathcal{N}\\) and \\(\\mathcal{V}\\). The input video \\(V_{:t}\\) contains \\(P\\) action segments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.48, + 0.726 + ], + "angle": 0, + "content": "3.2 Image Captioning Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.795 + ], + "angle": 0, + "content": "The past action labels do not capture all the visual information and context presented in the video (e.g., the location of the person or the background objects). Our idea is to use an image captioning model to capture the action context, providing more informative prompts related to the action performer for LLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To enhance the descriptive power of captioning models, we formulate the image captioning task as a question-answering problem. Fig. 3 displays a set of questions related to six key concepts that the models should aim to answer given" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "code", + "bbox": [ + 0.303, + 0.148, + 0.699, + 0.249 + ], + "angle": 0, + "content": "Question Dictionary \\(= \\{ \\) Location: Where is the person? Detection: What objects are in this image? Action: What is the person doing in this image? Prediction: What will the person do later in this image? Interaction: What is the person interacting with in this image? Intention: What does the person want to do in this image?" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.281 + ], + "angle": 0, + "content": "Fig. 3: List of questions asking for information related to key concepts of egocentric image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.784, + 0.366 + ], + "angle": 0, + "content": "an input image: location, detection, action, prediction, interaction, and intention. Given a question \\( Q \\), the conditional text is formed as \"Question: \\( < Q > \\)? Answer: \" where the answer part will be filled by the captioning model. After extensive experiments, we select intention-based question for conditional generation and VideoBLIP [25] as our image captioning model (see supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.473 + ], + "angle": 0, + "content": "By posing these questions to the captioning models, we anticipate extracting useful descriptions beneficial for LLMs to predict future actions. In practice, the conditional caption generation yields consistent and accurate descriptions of the action context for later prediction. For image captioning, we use the middle frame of each clip segment to generate the captions, based on the assumption that the middle frame is highly likely to be the moment when the action is happening." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.492, + 0.514 + ], + "angle": 0, + "content": "3.3 Action Recognition Module" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.554, + 0.67, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.757 + ], + "angle": 0, + "content": "Fig. 4: Action recognition network. A frozen video encoder generates visual features from \\( n = 4 \\) continuous action segments, followed by transformer layers and classification heads." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The accurate past action descriptions are crucial, as they provide essential contextual information about past events in a given video input. Our model recognizes actions from the observed video clips, taking inspiration from the baseline model used in the Ego4D paper [18]. Visual features are extracted from a video encoder given \\( n \\) consecutive action segments (\\( n = 4 \\) in Fig. 4). We adopt" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.376, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "the off-the-shelf video encoder of EgoVLP [34], a vision-language model pretrained on Ego4D. Next, transformer layers are used to combine multiple visual features into an aggregated feature, followed by \\( n \\) classification heads to recognize the verbs and nouns independently. The video encoder is frozen, and the rest of the model is trained with cross-entropy loss summed over \\( n \\) classification heads. We also try to exploit a text encoder by formulating the action recognition task as matching a proper action label with an input video, which does not improve the accuracy (See supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.389 + ], + "angle": 0, + "content": "We empirically find that recognizing actions in isolation falls short in effectiveness compared to recognizing \\( n \\) actions within a continuous sequence. This is because actions in neighboring segments are often highly correlated, reflecting the natural continuity embedded in human activities. Thus, we use a sliding input video window that includes \\( n = 4 \\) segments with the window striding of one. The top-1 action is determined by identifying the most common action from the classifications across different windows. See supplementary for more details and ablation results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.416, + 0.495, + 0.431 + ], + "angle": 0, + "content": "3.4 Action Anticipation Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.627 + ], + "angle": 0, + "content": "In this section, we discuss the details of action anticipation with LLMs, from how the prompt is formatted to the details of inference and post-processing of raw outputs. Llama 2 (7B) [54], an open-source LLM, is employed in our framework. Prompt Design. Following common practice, we design the prompt with a few examples to enable in-context learning (ICL), maximizing references given to LLMs. Fig. 5 shows the template of our prompts, which consists of an instruction, a few examples, and a query. The instruction includes the task description and defines the input and output for LLMs. The examples for ICL, selected from the training set, include past and future actions (ground truth) and corresponding textual descriptions generated by the captioning model, which we call narrations. The query includes recognized past actions and generated narrations, and future actions are left empty for LLMs to complete." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.651, + 0.778, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.812, + 0.788, + 0.855 + ], + "angle": 0, + "content": "Fig. 5: Prompt structure for LLMs. The prompt is composed of an instruction, examples retrieved from the training set, and the query related to the prediction video. Each video description consists of narrations, past actions, and future actions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "Example Selection. Incorporating some examples that demonstrate proper generation will greatly improve the performance of LLMs by providing additional context and references [6, 40]. Considering the diversity of egocentric videos, especially within the Ego4D dataset, randomly selected examples are unlikely to generalize well on various scenarios within the test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Therefore, we utilize a maximal marginal relevance (MMR) [7, 59] based exemplar selection strategy. The main idea behind MMR is to select exemplars that are relevant to the query while exhibiting sufficient variety to provide robust generalization across the test set. To be specific, we iteratively select a set of examples \\( p_i \\in T \\) from the training set \\( D \\) that are semantically close to the query prompt \\( q \\) but also diverse enough to provide additional information." + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.323, + 0.785, + 0.358 + ], + "angle": 0, + "content": "\\[\np _ {j} = \\underset {p _ {j} \\in D / T} {\\arg \\max } \\left[ \\lambda S (q, p _ {j}) - (1 - \\lambda) \\underset {p _ {i} \\in T} {\\max } S \\left(p _ {i}, p _ {j}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.785, + 0.453 + ], + "angle": 0, + "content": "where \\( S \\) measures the semantic similarity between exemplars and the parameter \\( \\lambda \\) strikes a balance between similarity and diversity. One exemplar includes past and future action sequences and corresponding captions in natural language. We use MPNet [52] to extract text embeddings and calculate cosine similarity to measure semantic similarity. We compare various exemplar selection strategies in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.785, + 0.575 + ], + "angle": 0, + "content": "Fine-tuning LLM. To facilitate LLMs to generate proper action sequence, we fine-tune the LLMs based on the training set of each egocentric dataset. As fine-tuning whole parameters of the LLMs can be computationally inefficient and may lead to catastrophic forgetting [26], parameter-efficient fine-tuning is leveraged. Specifically, we utilize LoRA [19] with 8-bit quantization using a batch size of 8 and a learning rate of \\(5 \\times 10^{-4}\\) for 10 epochs. The training set for fine-tuning consists of input prompts in the format of Fig. 5 with different exemplars and ground truth future actions as targets to be generated by the LLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.785, + 0.712 + ], + "angle": 0, + "content": "Inference of LLM and Post-processing. Following the evaluation protocol of Ego4d [18], we generate multiple predictions for each test video. To reduce the variability in predictions, we prompt LLMs multiple times using the same prompt structure, as shown in Fig. 5, but with different exemplars. After inference, we map the raw output of LLMs to the label space of predefined verbs and nouns. One should note that the output of LLMs may not always fall within the given taxonomy, despite all input prompts containing verbs and nouns from the predefined domain. For simplicity, we adopt a rule-based mapping (see details in supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.376, + 0.751 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "In this section, we evaluate PALM on the long-term action anticipation (LTA) task, shedding light on the various design aspects. We conduct a comprehensive evaluation using the Ego4D dataset, which stands as the largest repository of egocentric videos. We report the final performance metrics on the Ego4D [18], EPIC-KITCHEN-55 [12], and EGTEA Gaze+ [33] benchmarks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.376, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Comparison to the SOTA on the Ego4D test set. PALM outperforms all baselines on the Ego4D LTA benchmark in respect of edit distance (ED \\(\\downarrow\\))." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.185, + 0.699, + 0.404 + ], + "angle": 0, + "content": "
MethodVersionVerb EDNoun EDAction ED
Slowfast [18]v10.73890.78000.9432
VCLIP [14]v10.73890.76880.9412
ICVAE [36]v10.74100.73960.9304
HierVL [2]v10.72390.73500.9276
AntGPT [63]v10.65840.65460.8814
PALM (ours)v10.65590.64010.8613
Last actionv20.83290.72580.9438
Repeat actionv20.74910.70730.9155
Retrieve actionv20.74170.70410.9127
Slowfast [18]v20.71690.73590.9253
VideoLLM [9]v20.7210.7250.921
AntGPT [63]v20.65030.64980.8770
PALM (ours)v20.64710.61170.8503
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.432, + 0.335, + 0.445 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Ego4D [18]. With 243 hours of egocentric video content encompassing 53 distinct scenarios, Ego4D offers a rich variety of data with a predefined annotation of 117 verbs and 521 nouns for the LTA task. Ego4D is available in two versions, v1 and v2, where Ego4D v1 is a subset of Ego4D v2 with fewer videos and a smaller verb and noun set. We use both versions of the dataset to enable extensive comparison with previous methods and adhere to the identical training, validation, and test splits as specified in the benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.607 + ], + "angle": 0, + "content": "EPIC-KITCHENS-55 (EK-55) [12]. This dataset comprises 55 hours of egocentric videos centered around cooking scenarios, featuring 125 annotated verbs and 352 nouns. We follow the same train and test splits as EGO-TOPO [38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.655 + ], + "angle": 0, + "content": "EGTEA Gaze+ (EGTEA) [33]. With approximately 26 hours of egocentric cooking videos, EGTEA includes 19 verbs and 53 nouns. Similar to EK-55, we employ the same train and test splits as described in EGO-TOPO [38]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.475, + 0.691 + ], + "angle": 0, + "content": "4.2 Comparison to the SOTA" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.699, + 0.501, + 0.713 + ], + "angle": 0, + "content": "Baselines. We include three baselines:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.724, + 0.786, + 0.754 + ], + "angle": 0, + "content": "- Last action outputs the last action of past action sequence as \\( Z \\) future actions." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.755, + 0.768, + 0.77 + ], + "angle": 0, + "content": "- Repeat action repeats the given past \\( P \\) actions to fill \\( Z \\) future actions." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.77, + 0.787, + 0.815 + ], + "angle": 0, + "content": "- Retrieve action retrieves the most semantically similar example in the training set (as in the example selection step) and uses their future actions as the prediction." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.724, + 0.787, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.588, + 0.84 + ], + "angle": 0, + "content": "We also compare to three previous SOTA methods:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.219, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.657, + 0.162 + ], + "angle": 0, + "content": "- Slowfast [18] is the baseline LTA model given by Ego4d." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.162, + 0.784, + 0.19 + ], + "angle": 0, + "content": "- VideoLLM [9] converts multimodal data into a unified token sequence and utilizes fine-tuned LLMs" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.785, + 0.222 + ], + "angle": 0, + "content": "- AntGPT [63] leveraged action recognition model and fine-tuned LLMs to predict future actions." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.261, + 0.368, + 0.271 + ], + "angle": 0, + "content": "① Single Object Scenario (1 object)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.314, + 0.268, + 0.321 + ], + "angle": 0, + "content": "GT Past" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.274, + 0.766, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.325, + 0.27, + 0.332 + ], + "angle": 0, + "content": "GT Future " + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.313, + 0.619, + 0.322 + ], + "angle": 0, + "content": "(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.35, + 0.26, + 0.357 + ], + "angle": 0, + "content": "(0.55)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.324, + 0.75, + 0.332 + ], + "angle": 0, + "content": "(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (take, brick)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.35, + 0.26, + 0.357 + ], + "angle": 0, + "content": "(0.55)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.333, + 0.66, + 0.342 + ], + "angle": 0, + "content": "(put, brick), (put, brick), (move, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.346, + 0.27, + 0.353 + ], + "angle": 0, + "content": "SlowFast" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.344, + 0.745, + 0.353 + ], + "angle": 0, + "content": "put, brick), (put, brick), (put, brick), (adjust, brick), (take, brick), (take, brick), (take, brick), (take, brick), (put, brick)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.355, + 0.26, + 0.361 + ], + "angle": 0, + "content": "(0.55)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.354, + 0.657, + 0.362 + ], + "angle": 0, + "content": "(take, brick), (adjust, brick), (put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.366, + 0.26, + 0.373 + ], + "angle": 0, + "content": "PALM" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.365, + 0.731, + 0.373 + ], + "angle": 0, + "content": "(put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick), (take, brick)," + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.375, + 0.26, + 0.381 + ], + "angle": 0, + "content": "(0.4)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.374, + 0.64, + 0.381 + ], + "angle": 0, + "content": "(put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.39, + 0.371, + 0.399 + ], + "angle": 0, + "content": "② Few Object Scenario (2~5 objects)" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.401, + 0.767, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.441, + 0.268, + 0.449 + ], + "angle": 0, + "content": "GTPast" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.441, + 0.738, + 0.45 + ], + "angle": 0, + "content": "(dip, paintbrush), (put, container), (take, paintbrush), (put, paintbrush), (adjust, paintbrush), (take, container), (dip, container), (paint, wall)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.453, + 0.27, + 0.461 + ], + "angle": 0, + "content": "GT Futur" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.453, + 0.76, + 0.462 + ], + "angle": 0, + "content": "(paint, wall), (dip, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (dip, container), (paint, wall)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.463, + 0.726, + 0.473 + ], + "angle": 0, + "content": "(paint, paintbrush), (paint, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.48, + 0.27, + 0.487 + ], + "angle": 0, + "content": "SlowFast" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.48, + 0.764, + 0.489 + ], + "angle": 0, + "content": "take, container), (put, container), (hold, container), (put, string), (open, brush), (turn, container), (touch, nail gun), (hold, container), (throw, soap)," + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.49, + 0.748, + 0.5 + ], + "angle": 0, + "content": "take, cereal), move, sheft), (push, wire), (break, cloth), (break, wire), (hold, circuit), (hold, container), (hold, garbage can), (remove, switch), (put, camera). (hold, cloth)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.507, + 0.26, + 0.514 + ], + "angle": 0, + "content": "PALM" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.506, + 0.743, + 0.515 + ], + "angle": 0, + "content": "put paintbrush). (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall)," + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.516, + 0.26, + 0.522 + ], + "angle": 0, + "content": "(0.6)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.516, + 0.707, + 0.523 + ], + "angle": 0, + "content": "(dip, paintbrush), (paint, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.53, + 0.371, + 0.538 + ], + "angle": 0, + "content": "③ Many Object Scenario (>5 objects)" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.541, + 0.766, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.581, + 0.268, + 0.588 + ], + "angle": 0, + "content": "GT Past" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.581, + 0.65, + 0.589 + ], + "angle": 0, + "content": "put, screwdriver), (put, bench), (take, wire), (attach, wire), (adjust, wire), (attach, wire), (hold, wire), (hold, wire)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.593, + 0.27, + 0.6 + ], + "angle": 0, + "content": "GT Future" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.593, + 0.73, + 0.602 + ], + "angle": 0, + "content": "(squeeze, wire), (adjust, wire), (hold, wire), (move, stool), (sit, stool), (take, screwdriver), (serve, mower), (put, screwdriver), (take, nut)," + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.603, + 0.751, + 0.615 + ], + "angle": 0, + "content": "(hit, mower), (take, screwdriver), (move, screwdriver), (repair, mower), (put, screwdriver), (repair, mower), (repair, mower), (hold, screwdriver), (take, screwdriver), (repair, mower), (move, screwdriver)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.62, + 0.27, + 0.627 + ], + "angle": 0, + "content": "SlowFast" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.62, + 0.748, + 0.628 + ], + "angle": 0, + "content": "(attach, wheel), (turn, pipe), (take, valve), (put, wheel), (put, sticker), (take, drawer), (put, wheel), (repair, torch), (hold, wheel), (repair, plier)," + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.63, + 0.26, + 0.636 + ], + "angle": 0, + "content": "(1.0)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.63, + 0.749, + 0.639 + ], + "angle": 0, + "content": "(put, screwdriver), (attach, wheel), (repair, container), (put, wheel), (put, wheel), (attach, seasoning), (take, wheel), (take, cutter), (repair, screw)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.647, + 0.26, + 0.654 + ], + "angle": 0, + "content": "PALM" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.646, + 0.721, + 0.655 + ], + "angle": 0, + "content": "move, screwdriver), (move, screwdriver), (take, wrench), (put, screwdriver), (move, screwdriver), (take, screwdriver), (move, screw" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.656, + 0.745, + 0.664 + ], + "angle": 0, + "content": "(move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver)" + }, + { + "type": "list", + "bbox": [ + 0.276, + 0.646, + 0.745, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.796 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results of predictions on Ego4D. We show the comparison between the predictions from the Ego4D baseline using SlowFast [18] and the predictions from our model PALM respectively. For each of the three examples shown here, we include the middle frames of the previous eight action segments with the corresponding ground truth past actions GT Past and future actions GT Future alongside the predicted future actions. The green words represent correct predictions while the red words denote incorrect predictions. The edit distance between the predicted actions and ground truth actions is shown in parenthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.79, + 0.84 + ], + "angle": 0, + "content": "Evaluation. We conform to the evaluation procedures set forth by the Ego4D [18] and calculate the edit distance (ED) using the Damerau-Levenshtein distance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.376, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.206 + ], + "angle": 0, + "content": "metric [13, 28]. ED is computed separately for verbs, nouns, and the entire action sequences. Given \\( P = 8 \\) action segments of the input video, we report the minimum edit distance among \\( K = 5 \\) predicted sequences, each with a length of \\( Z = 20 \\) actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.79, + 0.373 + ], + "angle": 0, + "content": "Results. Tab. 1 shows that PALM improves existing SOTA methods and achieves the best performance on the test set of the Ego4D v1 and v2 benchmark. In general, LLM-based approaches (VideoLLM, AntGPT, and PALM) predict more accurate future actions compared to other approaches based on representation learning. Notably, repeat and retrieve action baselines in Ego4D v2 perform better than Slowfast with respect to noun and action ED. These observations clearly indicate the advantages of leveraging foundation models and the limitations of feature learning methods on the LTA task. Moreover, PALM outperforms other LLM-based methods, which demonstrates the effectiveness of our approach to provide ample visual context to LLMs, equipped with action recognition and image captioning models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.789, + 0.616 + ], + "angle": 0, + "content": "In Fig. 6, we perform a qualitative comparison between the Ego4D baseline using SlowFast [18] and PALM. The first example exemplifies a scenario where the person mainly interacts with a single object, in this case, a brick. Both methods can accurately predict the future noun \"brick\". In the second scenario, the person is drawing a painting on a wall, interacting with multiple objects such as a paintbrush, a container, and the wall. PALM predicts the repeated sequence of actions, unlike SlowFast which incorrectly anticipates unrelated actions such as putting a string, touching a nail gun, or throwing a soap, none of which are present in the input video. This also demonstrates the importance of previous actions in the anticipation of future actions. The final scenario portrays the person adjusting a wire in a repair shop, a task involving numerous items. While SlowFast and PALM recognize central objects such as a screwdriver, they both fail to predict a coherent sequence of repair actions. We hypothesize that a language model with expert knowledge can produce more meaningful anticipations, particularly in the context of repairing mechanics, where it is crucial to execute actions in a precise order. More qualitative results are shown in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.345, + 0.65 + ], + "angle": 0, + "content": "4.3 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.703 + ], + "angle": 0, + "content": "We perform various ablation studies regarding action recognition, example selection methods, LLM models, and prompt contents. We evaluate on the validation set of Ego4D v2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Action Recognition. The action recognition module is an essential part of our framework PALM as we utilize LLMs for predicting future actions and the accuracy of these predictions is highly influenced by the sequence of previous actions. We evaluate various video encoders trained on egocentric videos, including MViT [15], SlowFast [16], StillFast [49], EgoVLP [34], and EgoVLPv2 [46]. It is important to note that we employ a frozen video encoder without any finetuning, as depicted in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 presents the Top-1 Accuracy \\((\\%)\\) and the edit distance of each video encoder. EgoVLP emerges as the top performer, achieving the best accuracy" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Ablation on action recognition module. We compare different pretrained video encoders by reporting the Top-1 accuracy (Acc. ↑) and edit distance (ED ↓) for verbs, nouns, and actions. For comparative purposes, we also include the ED when ground truth actions (GT action) are used in the prompts. The best Acc. and ED are highlighted in bold except those from GT action." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.227, + 0.768, + 0.342 + ], + "angle": 0, + "content": "
MethodVerb Acc.Noun Acc.Action Acc.Verb EDNoun EDAction ED
MViT [15]19.872.550.510.72910.85970.9630
SlowFast [16]19.4214.653.120.72520.74720.9278
StillFast [49]19.1219.414.060.73560.73690.9241
EgoVLPv2 [46]33.8440.6316.310.71820.66510.8937
EgoVLP [34]40.3245.5320.630.71110.64650.8819
GT action---0.66630.58870.8319
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.49 + ], + "angle": 0, + "content": "and edit distance among all encoders. EgoVLP utilizes a custom contrastive loss to align egocentric video frames with narrations, effectively establishing a connection between visual and textual embeddings. However, EgoVLPv2 utilizes task-specific loss for question-answering tasks built on the EgoVLP setting, which results in slightly less informative video features for action recognition. Unlike EgoVLP and EgoVLPv2, other video encoders are solely trained on video frames, making it challenging to draw robust connections between their video features and textual information (verbs and nouns). Based on these results, we select EgoVLP as our video encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.598 + ], + "angle": 0, + "content": "It is expected that better accuracy in action recognition leads to a reduced edit distance, as accurate estimation of past actions is crucial for predicting future actions. Notably, the use of ground truth (GT) actions significantly outperforms all video encoders, which highlights the enhancement of action recognition as the foremost challenge within our framework. We believe that fine-tuning the video encoder on specific egocentric datasets would likely enhance recognition accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.623, + 0.787, + 0.651 + ], + "angle": 0, + "content": "Table 3: Ablation on example selection strategies exhibits that MMR outperforms other selection methods." + }, + { + "type": "table", + "bbox": [ + 0.36, + 0.664, + 0.639, + 0.737 + ], + "angle": 0, + "content": "
MethodVerb EDNoun EDAction ED
Random0.73620.67470.9002
Similarity0.72040.66610.8923
MMR0.71790.66470.8908
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Example Selection. We compare different example selection strategies for in-context learning (ICL), including random, similarity, and maximal marginal relevance (MMR) [59]. The random method uniformly selects examples from the training set, whereas the similarity method selects the most semantically similar example to the query. MMR considers both the similarity and diversity of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.376, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "examples, striving for a balance between them. Tab. 3 demonstrates that MMR outperforms the other selection methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.206, + 0.784, + 0.234 + ], + "angle": 0, + "content": "Table 4: Ablation on LLMs shows that larger LLMs perform better on the LTA task." + }, + { + "type": "table", + "bbox": [ + 0.325, + 0.248, + 0.675, + 0.348 + ], + "angle": 0, + "content": "
MethodVerb EDNoun EDAction ED
GPT-Neo (125M)0.76570.68670.9119
GPT-Neo (1.3B)0.73140.67590.8970
GPT-Neo (2.7B)0.71780.67470.8943
Llama 2 (7B)0.71790.66470.8908
Llama 2 (13B)0.69580.66340.8867
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.785, + 0.485 + ], + "angle": 0, + "content": "LLM Ablation. We ablate the impact of LLMs by comparing different open-source models, primarily GPT-Neo [4] and Llama 2 [54] families. Tab. 4 indicates that larger language models are more capable of predicting future actions with their improved capacity for generalization across novel scenarios. This observation aligns with the performance of LLMs on general NLP benchmarks. Taking into account the computational resources at our disposal, we select Llama 2 with 7B parameters." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.514, + 0.784, + 0.543 + ], + "angle": 0, + "content": "Table 5: Ablation on the content of the prompt. We use abbreviations for each content in the input prompt: \\(\\mathcal{A}\\) - action sequence, \\(\\mathcal{C}\\) - captions" + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.555, + 0.66, + 0.628 + ], + "angle": 0, + "content": "
Prompt ContentVerb EDNoun EDAction ED
A0.72030.66880.8947
C0.70880.69010.9064
A+C0.71790.66470.8908
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Prompt Contents. We examine different content included in the input prompt to validate the contribution of each module to final performance. PALM utilizes recognized past actions and generated image captions to compose prompts. The action sequence describes the temporal flow of activity in (verb, noun) format, while captions narrate visual information relevant to human. In Tab. 5, the action sequence with captions \\((\\mathcal{A} + \\mathcal{C})\\) achieves the lowest action ED, which thereby is used in our framework. Notably, caption-only prompt \\((\\mathcal{C})\\) reaches the lowest verb ED, showing that narrations are capable of capturing dynamic motions of human activity, while action-only prompt \\((\\mathcal{A})\\) effectively detects key objects interacted with human supported by relatively lower noun ED. This indicates that action sequence and narrations provide a beneficial complement, thereby supporting our decision to incorporate both in the prompts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 6: Comparison to the SOTA on EK-55 and EGTEA. Performance is evaluated using the mAP metric \\((\\uparrow)\\)." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.185, + 0.71, + 0.335 + ], + "angle": 0, + "content": "
MethodEK-55EGTEA
ALLFREQRAREALLFREQRARE
I3D [8]32.753.323.072.179.353.3
ActionVLAD [17]29.853.518.673.379.058.6
Timeception [22]35.655.926.174.179.759.7
VideoGraph [23]22.549.414.067.777.147.2
EGO-TOPO [38]38.056.929.273.580.754.7
Anticipatr [41]39.158.129.176.883.355.1
AntGPT [63]40.158.831.980.284.872.9
PALM (ours)40.459.330.380.785.073.5
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.361, + 0.544, + 0.375 + ], + "angle": 0, + "content": "4.4 Evaluation on EK-55 and EGTEA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.787, + 0.534 + ], + "angle": 0, + "content": "Evaluation. For EK-55 [12] and EGTEA [33], we adhere to the experimental setup established by EGO-TOPO [38] and employ mean average precision (mAP) for multi-label classification as the evaluation metric. The initial \\( R\\% \\) (\\( R = [25\\%, 50\\%, 75\\%] \\)) of each video serves as input, with the objective of predicting the set of actions occurring in the remaining \\( (100 - R)\\% \\) of the video. Notably, actions here are confined to verb predictions only without noun, following previous conventions. We calculate the mAP on the validation set averaged over all \\( R \\), examining various performance aspects, such as all target actions (ALL), frequently occurring actions (FREQ), and rarely occurring actions (RARE)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.787, + 0.672 + ], + "angle": 0, + "content": "Results. As shown in Tab. 6, our method achieves SOTA results in the ALL and FREQ section while exhibiting comparable performance in the RARE section. The results demonstrate the effectiveness of PALM by capturing additional visual information with image captioning models. This aligns with the ablation study regarding prompt contents in Sec. 4.3 where captions contribute better on verb predictions. We observe that verb in RARE section are seldom represented in the captions, which results in relatively lower mAP in RARE section. Thus, we believe fine-tuning image captioning model or employing better video captioning model will further improve the performance accordingly." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.36, + 0.707 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We present PALM, a framework that leverages vision-language and large language models for long-term action anticipation. PALM achieves the best performance on the Ego4D benchmark, demonstrating the effectiveness of performing the LTA task in the language domain and the benefits of using foundation models. Our model also shows competitive results on two other benchmarks, EK-55 and EGTEA. As a result, we are optimistic that our work will spur further research in using commonsense knowledge extracted from LLMs for video understanding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.26 + ], + "angle": 0, + "content": "2. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.787, + 0.3 + ], + "angle": 0, + "content": "3. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. If you use this software, please cite it using these metadata 58 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.787, + 0.37 + ], + "angle": 0, + "content": "4. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. https://doi.org/10.5281/zenodo.5297715 (2021). https://doi.org/10.5281/ zenodo.5297715, https://doi.org/10.5281/zenodo.5297715, If you use this software, please cite it using these metadata. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "5. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.787, + 0.466 + ], + "angle": 0, + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "7. Carbonell, J., Goldstein, J.: The use of mmr, diversity-based reranking for reordering documents and producing summaries. In: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval. pp. 335-336 (1998) 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.787, + 0.564 + ], + "angle": 0, + "content": "8. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "9. Chen, G., Zheng, Y.D., Wang, J., Xu, J., Huang, Y., Pan, J., Wang, Y., Wang, Y., Qiao, Y., Lu, T., et al.: Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292 (2023) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "0. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "1. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "2. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(11), 4125-4141 (2020) 2, 3, 8, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.772 + ], + "angle": 0, + "content": "3. Damerau, F.J.: A technique for computer detection and correction of spelling errors. Communications of the ACM 7(3), 171-176 (1964) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "4. Das, S., Ryoo, M.S.: Video+ clip baseline for ego4d long-term action anticipation. arXiv preprint arXiv:2207.00579 (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 2, 4, 11, 12" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "17. Girdhar, R., Ramanan, D., Gupta, A., Sivic, J., Russell, B.: Actionvlad: Learning spatio-temporal aggregation for action classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 971-980 (2017) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.3 + ], + "angle": 0, + "content": "18. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022) 2, 3, 4, 6, 8, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.302, + 0.787, + 0.345 + ], + "angle": 0, + "content": "19. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.387 + ], + "angle": 0, + "content": "20. Huang, D., Hilliges, O., Van Gool, L., Wang, X.: Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.787, + 0.43 + ], + "angle": 0, + "content": "21. Huang, W., Abbeel, P., Pathak, D., Mordatch, I.: Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In: International Conference on Machine Learning. pp. 9118-9147. PMLR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "22. Hussein, N., Gavves, E., Smeulders, A.W.: Timeception for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 254-263 (2019) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.5 + ], + "angle": 0, + "content": "23. Hussein, N., Gavves, E., Smeulders, A.W.: Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143 (2019) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "24. Ichter, B., Brohan, A., Chebotar, Y., Finn, C., Hausman, K., Herzog, A., Ho, D., Ibarz, J., Irpan, A., Jang, E., et al.: Do as i can, not as i say: Grounding language in robotic affordances. In: 6th Annual Conference on Robot Learning (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.613 + ], + "angle": 0, + "content": "25. Keunwoo Peter, Y.: Videooblip is a large vision-language model based on blip-2 that can generate texts conditioned on videos. https://github.com/yukw777/VideoBLIP (2021). https://doi.org/https://orcid.org/0000-0002-4387-2410, https://github.com/yukw777/VideoBLIP, If you use this software, please cite it using these metadata. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.67 + ], + "angle": 0, + "content": "26. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.712 + ], + "angle": 0, + "content": "27. Leo, M., Medioni, G., Trivedi, M., Kanade, T., Farinella, G.M.: Computer vision for assistive technologies. Computer Vision and Image Understanding 154, 1-15 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.753 + ], + "angle": 0, + "content": "28. Levenshtein, V.I., et al.: Binary codes capable of correcting deletions, insertions, and reversals. In: Soviet physics doklady. vol. 10, pp. 707-710. Soviet Union (1966) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.797 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.247 + ], + "angle": 0, + "content": "32. Li, S., Puig, X., Paxton, C., Du, Y., Wang, C., Fan, L., Chen, T., Huang, D.A., Akyurek, E., Anandkumar, A., et al.: Pre-trained language models for interactive decision-making. Advances in Neural Information Processing Systems 35, 31199-31212 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.248, + 0.788, + 0.288 + ], + "angle": 0, + "content": "33. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018) 3, 8, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.29, + 0.788, + 0.331 + ], + "angle": 0, + "content": "34. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022) 4, 7, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.333, + 0.788, + 0.375 + ], + "angle": 0, + "content": "35. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.376, + 0.788, + 0.416 + ], + "angle": 0, + "content": "36. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.418, + 0.788, + 0.473 + ], + "angle": 0, + "content": "37. Meditskos, G., Plans, P.M., Stavropoulos, T.G., Benois-Pineau, J., Buso, V., Kompatsiaris, I.: Multi-modal activity recognition from egocentric vision, semantic enrichment and lifelogging applications for the care of dementia. Journal of Visual Communication and Image Representation 51, 169-190 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.475, + 0.788, + 0.515 + ], + "angle": 0, + "content": "38. Nagarajan, T., Li, Y., Feichtenhofer, C., Grauman, K.: Ego-topo: Environment affordances from egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 163-172 (2020) 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.517, + 0.788, + 0.558 + ], + "angle": 0, + "content": "39. Nakazawa, A., Honda, M.: First-person camera system to evaluate tender dementia-care skill. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.559, + 0.788, + 0.6 + ], + "angle": 0, + "content": "40. Naveed, H., Khan, A.U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Barnes, N., Mian, A.: A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.602, + 0.788, + 0.642 + ], + "angle": 0, + "content": "41. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.644, + 0.788, + 0.685 + ], + "angle": 0, + "content": "42. Neumann, L., Zisserman, A., Vedaldi, A.: Future event prediction: If and when. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.687, + 0.788, + 0.728 + ], + "angle": 0, + "content": "43. OhnBar, E., Kitani, K., Asakawa, C.: Personalized dynamics models for adaptive assistive navigation systems. In: Conference on Robot Learning. pp. 16-39. PMLR (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.788, + 0.797 + ], + "angle": 0, + "content": "44. Pasca, R.G., Gavryushin, A., Hamza, M., Kuo, Y.L., Mo, K., Van Gool, L., Hilliges, O., Wang, X.: Summarize the past to predict the future: Natural language descriptions of context boost multimodal object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18286-18296 (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "45. Patel, D., Eghbalzadeh, H., Kamra, N., Iuzzolino, M.L., Jain, U., Desai, R.: Pre-trained language models as visual planners for human assistance. arXiv preprint arXiv:2304.09179 (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "S. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "46. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023) 2, 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.259 + ], + "angle": 0, + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.288 + ], + "angle": 0, + "content": "48. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.343 + ], + "angle": 0, + "content": "49. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023) 2, 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.387 + ], + "angle": 0, + "content": "50. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.443 + ], + "angle": 0, + "content": "51. Ryoo, M.S., Fuchs, T.J., Xia, L., Aggarwal, J.K., Matthies, L.: Robot-centric activity prediction from first-person videos: What will they do to me? In: Proceedings of the tenth annual ACM/IEEE international conference on human-robot interaction. pp. 295–302 (2015) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.486 + ], + "angle": 0, + "content": "52. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "53. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.572 + ], + "angle": 0, + "content": "54. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 3, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.6 + ], + "angle": 0, + "content": "55. Wang, S., Zhao, Q., Do, M.Q., Agarwal, N., Lee, K., Sun, C.: Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.787, + 0.642 + ], + "angle": 0, + "content": "56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.685 + ], + "angle": 0, + "content": "57. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. In: International Conference on Learning Representations (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.787, + 0.741 + ], + "angle": 0, + "content": "58. Yao, Y., Xu, M., Choi, C., Crandall, D.J., Atkins, E.M., Dariush, B.: Egocentric vision-based future vehicle localization for intelligent driving assistance systems. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9711-9717. IEEE (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.783 + ], + "angle": 0, + "content": "59. Ye, X., Iyer, S., Celikyilmaz, A., Stoyanov, V., Durrett, G., Pasunuru, R.: Complementary explanations for effective in-context learning. arXiv preprint arXiv:2211.13892 (2022) 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "60. Zeng, A., Attarian, M., Choromanski, K.M., Wong, A., Welker, S., Tombari, F., Purohit, A., Ryoo, M.S., Sindhwani, V., Lee, J., et al.: Socratic models: Composing zero-shot multimodal reasoning with language. In: The Eleventh International Conference on Learning Representations (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.377, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PALM: Predicting Actions through Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "61. Zhan, K., Faux, S., Ramos, F.: Multi-scale conditional random fields for first-person activity recognition. In: 2014 IEEE international conference on pervasive computing and communications (PerCom). pp. 51-59. IEEE (2014) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "63. Zhao, Q., Zhang, C., Wang, S., Fu, C., Agarwal, N., Lee, K., Sun, C.: Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368 (2023) 4, 9, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "64. Zhou, C., Li, Q., Li, C., Yu, J., Liu, Y., Wang, G., Zhang, K., Ji, C., Yan, Q., He, L., et al.: A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.356 + ], + "angle": 0, + "content": "65. Zhou, L., Palangi, H., Zhang, L., Hu, H., Corso, J., Gao, J.: Unified vision-language pre-training for image captioning and vqa. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13041-13049 (2020) 3" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_origin.pdf b/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aebd22deb0a7cfab9fa3e671bd0b7ffdb4ba74ea --- /dev/null +++ b/2024/PALM_ Predicting Actions through Language Models/26a4ce29-0668-4f26-8004-06463aacfbd9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e00841eb60c762d20a72c76d3f5eec1fe37a410073834ad98caebba4d0a21e9 +size 5245726 diff --git a/2024/PALM_ Predicting Actions through Language Models/full.md b/2024/PALM_ Predicting Actions through Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..972e683b5a64675483f0cee60234b52ff0e2b040 --- /dev/null +++ b/2024/PALM_ Predicting Actions through Language Models/full.md @@ -0,0 +1,373 @@ +# PALM: Predicting Actions through Language Models + +Sanghwan Kim $^{1}$ , Daoji Huang $^{1}$ , Yongqin Xian $^{2}$ , Otmar Hilliges $^{1}$ , Luc Van Gool $^{1,3,4}$ , and Xi Wang $^{1}$ + +1 ETH Zürich + +2 Google + +3 KU Leuven + +4 INSAIT, Sofia + +Abstract. Understanding human activity is a crucial yet intricate task in egocentric vision, a field that focuses on capturing visual perspectives from the camera wearer's viewpoint. Traditional methods heavily rely on representation learning that is trained on a large amount of video data. However, a major challenge arises from the difficulty of obtaining effective video representation. This difficulty stems from the complex and variable nature of human activities, which contrasts with the limited availability of data. In this study, we introduce PALM, an approach that tackles the task of long-term action anticipation, which aims to forecast forthcoming sequences of actions over an extended period. Our method PALM incorporates an action recognition model to track previous action sequences and a vision-language model to articulate relevant environmental details. By leveraging the context provided by these past events, we devise a prompting strategy for action anticipation using large language models (LLMs). Moreover, we implement maximal marginal relevance for example selection to facilitate in-context learning of the LLMs. Our experimental results demonstrate that PALM surpasses the state-of-the-art methods in the task of long-term action anticipation on the Ego4D benchmark. We further validate PALM on two additional benchmarks, affirming its capacity for generalization across intricate activities with different sets of taxonomies. + +Keywords: egocentric vision $\cdot$ video understanding $\cdot$ action anticipation $\cdot$ large language model + +# 1 Introduction + +As wearable technology, such as smart glasses and body-worn cameras, becomes more prevalent, the availability of egocentric videos has grown significantly. These videos capture activities from the camera wearer's viewpoint, providing a unique perspective that features individual behaviors in various contexts. Egocentric vision holds immense promise for understanding how humans execute specific tasks and activities. It enables numerous applications such as real-time + +![](images/445e7a530ebaf75d625e14375332c35d3b460b207f5ecb592f89e2b6c5cf9ecb.jpg) +Fig. 1: PALM for long-term action anticipation. Given a video input, PALM uses a vision-language model (VLM) and an action recognition model (ARM) to compile a text prompt that describes the past video content. A large-language model (LLM) then takes the prompt and predicts future actions. + +systems that anticipate and respond to user needs [50, 51, 58], monitoring patient activities [37, 39, 61], and providing personalized instructions and procedures [27, 42, 43]. + +Existing works on understanding human activities mostly rely on learning representations from large-scale video datasets [15, 46, 49, 56]. However, these approaches face two critical challenges. First, the complexity and variability of human activities make it challenging to obtain comprehensive representations, particularly for longer videos. Second, reliance on large datasets limits the model's capability to generalize to less common, long-tail classes and unseen scenarios. In this context, there is a pressing need for a more versatile approach to human activity understanding, adept at handling the complicated nature of human activities without overly relying on large-scale video data. Therefore, recent works have started to exploit the procedure knowledge and generalization ability of large language models (LLMs) [1, 21, 32, 44, 45]. + +Most prior works based on LLMs have been largely confined to generating instructions and plans given a high-level task description (e.g., "prepare the dinner table") and have not been tailored for intricate human activity understanding from videos. Indeed, there is a significant gap in the literature when it comes to the use of LLMs for grounded activity understanding, especially in challenges involving understanding long-term action sequences. In this work, we efficiently boost the versatile ability of LLMs in long-term action anticipation [12, 18, 20] by providing sufficient visual context in discrete text. + +As shown in Fig. 1, we introduce a simple yet effective framework called PALM that combines descriptions of past action sequences and corresponding narrations to bridge the gap between visual and textual domains. In detail, PALM takes an input video with annotated action periods and leverages a vision-language model and an action recognition model to describe past events. We then design a prompt strategy for large language models (LLMs) to anticipate future actions given the context of these past events. To facilitate the predictive capabilities of LLMs, we use maximal marginal relevance (MMR) [7, 59] to retrieve similar and diverse examples from the training set, enabling in-context learning for the LLMs. Through this prompting strategy, our system can tackle intricate tasks in understanding human activities, benefiting from existing foun + +dition models. Experimental results demonstrate that our method PALM significantly improves performance in long-term action anticipation, as evaluated using the Ego4D benchmark [18]. Further evaluations on two additional benchmarks, EPIC-KITCHEN [12] and EGTEA [33], reveal the generalization ability of PALM across different contexts and datasets. Our contributions can be summarized as follows: + +- We propose an effective textural representation of past events, consisting of both action sequences and accompanying narrations that provide additional context and environmental details. +- We design a novel prompting strategy that enables efficient in-context learning for large language models. +- PALM outperforms existing state-of-the-art methods in the task of long-term action anticipation, which is part of the Ego4D benchmark. + +# 2 Related Work + +# 2.1 Foundation Model + +Foundation models serve as the cornerstone for various downstream tasks across different data modalities [5,64]. Here we consider two types of foundation models: vision-language models (VLMs) to generate the textual narrations of video and large language models (LLMs) to predict future action given visual context. + +Vision-language model. VLMs utilize complementary characteristics of image and text datasets to perform multimodal tasks such as image captioning, visual question-answering, and visual reasoning [31, 35, 47]. Various works [10, 57, 65] attempt to unify vision-language tasks with a single model, overcoming the limited utility of previous models. In this paper, the image or video captioning ability of VLMs is utilized to directly transform visual information into narrative text descriptions. We empirically found that BLIP family [11, 29, 30] provides concise and informative captions. For instance, BLIP [30] proposes a novel encoder-decoder architecture specialized for multimodal pertaining, and BLIP-2 [29] significantly reduces the trainable parameters by using a querying transformer to bridge the modality gap between frozen vision models and language models. InstructBLIP [11] further performs instruction tuning on BLIP-2, enabling general-purpose models with a unified natural language interface. + +Large language model. Language foundation models [3,6,48,53,54,62] have had a significant impact on the NLP field, demonstrating surprising ability to generalize across unseen tasks. With their extensive training data and large parameter size, LLMs have demonstrated the ability to learn from examples provided in input prompts, a concept known as in-context learning [6]. LLMs have also shown success in the tasks of decision-making and planning. For example, SayCan [24] combines a language model with learned low-level skills for grounded planning, while Socratic Models [60] formulate action prediction as text completion problems from captioned images. In this work, we explore how such procedure knowledge can be exploited in the complex action anticipation task and showcase the effectiveness of in-context learning for this purpose. + +# 2.2 Long-term Action Anticipation + +Given an input video, the long-term action anticipation (LTA) task aims to predict the next actions in the future. In this work, we explore various action recognition models to estimate accurate action sequences and extensively compare our method with previous action anticipation models. + +Action recognition. SlowFast [16] proposes to fuse two channels of input that consist of a low frame and a high frame rate video, using 3D convolutional kernels. Similar to Slowfast, StillFast [49] adopts a specific architecture with two branches to effectively process high-resolution still images and low-resolution video frames. On the other hand, MViT [15] employs several channel-resolution scale stages to obtain a multiscale pyramid of features, based on vision transformers. Compared to previous works solely optimized on video inputs, EgoVLP [34] and EgoVLPv2 [46] pre-train video and text encoders with contrastive loss specialized for the egocentric domain. Each encoder provides useful feature representations for downstream video tasks such as natural language queries, video question-answering, and video summarization. + +Action anticipation. Baselines provided by Ego4D [18] employ the SlowFast [16] video encoder to extract features from sampled clips, followed by transformer layers and classification heads to predict the verb-noun pair of future action sequence. HierVL [2] utilizes a hierarchical contrastive training objective that encourages text-visual alignment at both the clip and video levels. ICVAE [36] proposes using a predicted scenario, which summarizes the activity depicted in the video, as a condition for action prediction. A few concurrent approaches [9, 55, 63] leverage the generalization abilities and abundant knowledge of LLMs to predict future actions. Compared to these LLM-based methods, PALM differentiates itself by utilizing vision-language model for effective extraction of visual context and employing in-context learning in combination with MMR. + +# 3 Method + +We introduce PALM, a comprehensive framework designed for video understanding and action anticipation. PALM comprises three key modules, as illustrated in Fig. 2: an image captioning module (Sec. 3.2), an action recognition module (Sec. 3.3), and an action anticipation module based on LLMs (Sec. 3.4). + +Given an untrimmed input video, the action recognition module generates labels for past actions in the form of verb and noun pairs (eight past actions in Fig. 2). The action recognition model is trained to produce valid verb and noun pairs as predefined by each dataset. Simultaneously, the image captioning module generates captions from the input video, providing additional visual information relevant to the events, termed visual context. The prompt is then constructed using the visual context and past actions, passed to LLMs to predict future actions. Additionally, we incorporate exemplars from the training set to facilitate in-context learning for LLMs. We demonstrate the effectiveness of our + +![](images/5795b18eb1b49b39c88317000ad4c8ea310959ecea7496a448affae6437be336.jpg) +Fig. 2: Overview of PALM. Given an input video, an image captioning model and an action recognition model convert visual information of past activities into text format (visual context and past actions). Then, LLM utilizes visual context and past actions to predict the sequence of future actions. The image captioning model is frozen and in-context learning is adapted for LLM prompting. + +approach on the long-term action anticipation (LTA) task, achieving state-of-the-art (SOTA) performance on Ego4D benchmarks. + +# 3.1 Task Description + +The LTA task involves predicting a chronological sequence of future actions given an untrimmed video $V$ . The input video $V$ comprises multiple action segments with variable lengths, where each action segment corresponds to a single action (verb and noun pair). The segment boundaries for each action in the videos are provided as part of the input. Formally, after observing the video until time $t$ ( $e.g., V_{:t}$ ), the LTA model sequentially predicts $Z$ future actions that the camerawearer is likely to perform: + +$$ +\left\{\hat {n} _ {z}, \hat {v} _ {z} \right\} _ {z = 1} ^ {Z} \tag {1} +$$ + +where $\hat{n}_z\in \mathcal{N}$ and $\hat{v}_z\in \mathcal{V}$ , given the predefined noun and verb taxonomy $\mathcal{N}$ and $\mathcal{V}$ . The input video $V_{:t}$ contains $P$ action segments. + +# 3.2 Image Captioning Module + +The past action labels do not capture all the visual information and context presented in the video (e.g., the location of the person or the background objects). Our idea is to use an image captioning model to capture the action context, providing more informative prompts related to the action performer for LLMs. + +To enhance the descriptive power of captioning models, we formulate the image captioning task as a question-answering problem. Fig. 3 displays a set of questions related to six key concepts that the models should aim to answer given + +```txt +Question Dictionary $= \{$ Location: Where is the person? Detection: What objects are in this image? Action: What is the person doing in this image? Prediction: What will the person do later in this image? Interaction: What is the person interacting with in this image? Intention: What does the person want to do in this image? +``` + +an input image: location, detection, action, prediction, interaction, and intention. Given a question $Q$ , the conditional text is formed as "Question: $< Q >$ ? Answer: " where the answer part will be filled by the captioning model. After extensive experiments, we select intention-based question for conditional generation and VideoBLIP [25] as our image captioning model (see supplementary). + +By posing these questions to the captioning models, we anticipate extracting useful descriptions beneficial for LLMs to predict future actions. In practice, the conditional caption generation yields consistent and accurate descriptions of the action context for later prediction. For image captioning, we use the middle frame of each clip segment to generate the captions, based on the assumption that the middle frame is highly likely to be the moment when the action is happening. + +# 3.3 Action Recognition Module + +![](images/7cf9fd26667fd9cafbe07949498e02a6febba0d53c62f33862516d35fa84b929.jpg) +Fig. 3: List of questions asking for information related to key concepts of egocentric image. +Fig. 4: Action recognition network. A frozen video encoder generates visual features from $n = 4$ continuous action segments, followed by transformer layers and classification heads. + +The accurate past action descriptions are crucial, as they provide essential contextual information about past events in a given video input. Our model recognizes actions from the observed video clips, taking inspiration from the baseline model used in the Ego4D paper [18]. Visual features are extracted from a video encoder given $n$ consecutive action segments ( $n = 4$ in Fig. 4). We adopt + +the off-the-shelf video encoder of EgoVLP [34], a vision-language model pretrained on Ego4D. Next, transformer layers are used to combine multiple visual features into an aggregated feature, followed by $n$ classification heads to recognize the verbs and nouns independently. The video encoder is frozen, and the rest of the model is trained with cross-entropy loss summed over $n$ classification heads. We also try to exploit a text encoder by formulating the action recognition task as matching a proper action label with an input video, which does not improve the accuracy (See supplementary). + +We empirically find that recognizing actions in isolation falls short in effectiveness compared to recognizing $n$ actions within a continuous sequence. This is because actions in neighboring segments are often highly correlated, reflecting the natural continuity embedded in human activities. Thus, we use a sliding input video window that includes $n = 4$ segments with the window striding of one. The top-1 action is determined by identifying the most common action from the classifications across different windows. See supplementary for more details and ablation results. + +# 3.4 Action Anticipation Module + +In this section, we discuss the details of action anticipation with LLMs, from how the prompt is formatted to the details of inference and post-processing of raw outputs. Llama 2 (7B) [54], an open-source LLM, is employed in our framework. Prompt Design. Following common practice, we design the prompt with a few examples to enable in-context learning (ICL), maximizing references given to LLMs. Fig. 5 shows the template of our prompts, which consists of an instruction, a few examples, and a query. The instruction includes the task description and defines the input and output for LLMs. The examples for ICL, selected from the training set, include past and future actions (ground truth) and corresponding textual descriptions generated by the captioning model, which we call narrations. The query includes recognized past actions and generated narrations, and future actions are left empty for LLMs to complete. + +![](images/d3bae9eac5b178ad79edeacef38b7c564131452287cbde34e626c8ff22efd80b.jpg) +Fig. 5: Prompt structure for LLMs. The prompt is composed of an instruction, examples retrieved from the training set, and the query related to the prediction video. Each video description consists of narrations, past actions, and future actions. + +Example Selection. Incorporating some examples that demonstrate proper generation will greatly improve the performance of LLMs by providing additional context and references [6, 40]. Considering the diversity of egocentric videos, especially within the Ego4D dataset, randomly selected examples are unlikely to generalize well on various scenarios within the test set. + +Therefore, we utilize a maximal marginal relevance (MMR) [7, 59] based exemplar selection strategy. The main idea behind MMR is to select exemplars that are relevant to the query while exhibiting sufficient variety to provide robust generalization across the test set. To be specific, we iteratively select a set of examples $p_i \in T$ from the training set $D$ that are semantically close to the query prompt $q$ but also diverse enough to provide additional information. + +$$ +p _ {j} = \underset {p _ {j} \in D / T} {\arg \max } \left[ \lambda S (q, p _ {j}) - (1 - \lambda) \underset {p _ {i} \in T} {\max } S \left(p _ {i}, p _ {j}\right) \right], \tag {2} +$$ + +where $S$ measures the semantic similarity between exemplars and the parameter $\lambda$ strikes a balance between similarity and diversity. One exemplar includes past and future action sequences and corresponding captions in natural language. We use MPNet [52] to extract text embeddings and calculate cosine similarity to measure semantic similarity. We compare various exemplar selection strategies in Sec. 4.3. + +Fine-tuning LLM. To facilitate LLMs to generate proper action sequence, we fine-tune the LLMs based on the training set of each egocentric dataset. As fine-tuning whole parameters of the LLMs can be computationally inefficient and may lead to catastrophic forgetting [26], parameter-efficient fine-tuning is leveraged. Specifically, we utilize LoRA [19] with 8-bit quantization using a batch size of 8 and a learning rate of $5 \times 10^{-4}$ for 10 epochs. The training set for fine-tuning consists of input prompts in the format of Fig. 5 with different exemplars and ground truth future actions as targets to be generated by the LLMs. + +Inference of LLM and Post-processing. Following the evaluation protocol of Ego4d [18], we generate multiple predictions for each test video. To reduce the variability in predictions, we prompt LLMs multiple times using the same prompt structure, as shown in Fig. 5, but with different exemplars. After inference, we map the raw output of LLMs to the label space of predefined verbs and nouns. One should note that the output of LLMs may not always fall within the given taxonomy, despite all input prompts containing verbs and nouns from the predefined domain. For simplicity, we adopt a rule-based mapping (see details in supplementary). + +# 4 Experiments + +In this section, we evaluate PALM on the long-term action anticipation (LTA) task, shedding light on the various design aspects. We conduct a comprehensive evaluation using the Ego4D dataset, which stands as the largest repository of egocentric videos. We report the final performance metrics on the Ego4D [18], EPIC-KITCHEN-55 [12], and EGTEA Gaze+ [33] benchmarks. + +Table 1: Comparison to the SOTA on the Ego4D test set. PALM outperforms all baselines on the Ego4D LTA benchmark in respect of edit distance (ED $\downarrow$ ). + +
MethodVersionVerb EDNoun EDAction ED
Slowfast [18]v10.73890.78000.9432
VCLIP [14]v10.73890.76880.9412
ICVAE [36]v10.74100.73960.9304
HierVL [2]v10.72390.73500.9276
AntGPT [63]v10.65840.65460.8814
PALM (ours)v10.65590.64010.8613
Last actionv20.83290.72580.9438
Repeat actionv20.74910.70730.9155
Retrieve actionv20.74170.70410.9127
Slowfast [18]v20.71690.73590.9253
VideoLLM [9]v20.7210.7250.921
AntGPT [63]v20.65030.64980.8770
PALM (ours)v20.64710.61170.8503
+ +# 4.1 Datasets + +Ego4D [18]. With 243 hours of egocentric video content encompassing 53 distinct scenarios, Ego4D offers a rich variety of data with a predefined annotation of 117 verbs and 521 nouns for the LTA task. Ego4D is available in two versions, v1 and v2, where Ego4D v1 is a subset of Ego4D v2 with fewer videos and a smaller verb and noun set. We use both versions of the dataset to enable extensive comparison with previous methods and adhere to the identical training, validation, and test splits as specified in the benchmark. + +EPIC-KITCHENS-55 (EK-55) [12]. This dataset comprises 55 hours of egocentric videos centered around cooking scenarios, featuring 125 annotated verbs and 352 nouns. We follow the same train and test splits as EGO-TOPO [38]. + +EGTEA Gaze+ (EGTEA) [33]. With approximately 26 hours of egocentric cooking videos, EGTEA includes 19 verbs and 53 nouns. Similar to EK-55, we employ the same train and test splits as described in EGO-TOPO [38]. + +# 4.2 Comparison to the SOTA + +Baselines. We include three baselines: + +- Last action outputs the last action of past action sequence as $Z$ future actions. +- Repeat action repeats the given past $P$ actions to fill $Z$ future actions. +- Retrieve action retrieves the most semantically similar example in the training set (as in the example selection step) and uses their future actions as the prediction. + +We also compare to three previous SOTA methods: + +- Slowfast [18] is the baseline LTA model given by Ego4d. +- VideoLLM [9] converts multimodal data into a unified token sequence and utilizes fine-tuned LLMs +- AntGPT [63] leveraged action recognition model and fine-tuned LLMs to predict future actions. + +# ① Single Object Scenario (1 object) + +# GT Past + +![](images/f2534dc7abb416e8231d85550615eef7afb248d0c96c520f3e7ae3285eb7adfe.jpg) + +# GT Future + +(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick). + +# (0.55) + +(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (take, brick). + +# (0.55) + +(put, brick), (put, brick), (move, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick). + +# SlowFast + +put, brick), (put, brick), (put, brick), (adjust, brick), (take, brick), (take, brick), (take, brick), (take, brick), (put, brick) + +# (0.55) + +(take, brick), (adjust, brick), (put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick). + +# PALM + +(put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick), (take, brick), + +# (0.4) + +(put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick). + +# ② Few Object Scenario (2~5 objects) + +![](images/6d539fde5b9c097e01d4f2f0a50e8e85fde79cfa9fd4526e662ee6c05dad0c76.jpg) + +# GTPast + +(dip, paintbrush), (put, container), (take, paintbrush), (put, paintbrush), (adjust, paintbrush), (take, container), (dip, container), (paint, wall). + +# GT Futur + +(paint, wall), (dip, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (dip, container), (paint, wall) + +(paint, paintbrush), (paint, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush). + +# SlowFast + +take, container), (put, container), (hold, container), (put, string), (open, brush), (turn, container), (touch, nail gun), (hold, container), (throw, soap), + +take, cereal), move, sheft), (push, wire), (break, cloth), (break, wire), (hold, circuit), (hold, container), (hold, garbage can), (remove, switch), (put, camera). (hold, cloth). + +# PALM + +put paintbrush). (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), + +# (0.6) + +(dip, paintbrush), (paint, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall). + +# ③ Many Object Scenario (>5 objects) + +![](images/49f1a59419bf3967b1417737113621f4b758ae98cd718115618972a64b38d786.jpg) + +# GT Past + +put, screwdriver), (put, bench), (take, wire), (attach, wire), (adjust, wire), (attach, wire), (hold, wire), (hold, wire) + +# GT Future + +(squeeze, wire), (adjust, wire), (hold, wire), (move, stool), (sit, stool), (take, screwdriver), (serve, mower), (put, screwdriver), (take, nut), + +(hit, mower), (take, screwdriver), (move, screwdriver), (repair, mower), (put, screwdriver), (repair, mower), (repair, mower), (hold, screwdriver), (take, screwdriver), (repair, mower), (move, screwdriver). + +# SlowFast + +(attach, wheel), (turn, pipe), (take, valve), (put, wheel), (put, sticker), (take, drawer), (put, wheel), (repair, torch), (hold, wheel), (repair, plier), + +# (1.0) + +(put, screwdriver), (attach, wheel), (repair, container), (put, wheel), (put, wheel), (attach, seasoning), (take, wheel), (take, cutter), (repair, screw) + +# PALM + +move, screwdriver), (move, screwdriver), (take, wrench), (put, screwdriver), (move, screwdriver), (take, screwdriver), (move, screw +(move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver) + +Fig. 6: Qualitative results of predictions on Ego4D. We show the comparison between the predictions from the Ego4D baseline using SlowFast [18] and the predictions from our model PALM respectively. For each of the three examples shown here, we include the middle frames of the previous eight action segments with the corresponding ground truth past actions GT Past and future actions GT Future alongside the predicted future actions. The green words represent correct predictions while the red words denote incorrect predictions. The edit distance between the predicted actions and ground truth actions is shown in parenthesis. + +Evaluation. We conform to the evaluation procedures set forth by the Ego4D [18] and calculate the edit distance (ED) using the Damerau-Levenshtein distance + +metric [13, 28]. ED is computed separately for verbs, nouns, and the entire action sequences. Given $P = 8$ action segments of the input video, we report the minimum edit distance among $K = 5$ predicted sequences, each with a length of $Z = 20$ actions. + +Results. Tab. 1 shows that PALM improves existing SOTA methods and achieves the best performance on the test set of the Ego4D v1 and v2 benchmark. In general, LLM-based approaches (VideoLLM, AntGPT, and PALM) predict more accurate future actions compared to other approaches based on representation learning. Notably, repeat and retrieve action baselines in Ego4D v2 perform better than Slowfast with respect to noun and action ED. These observations clearly indicate the advantages of leveraging foundation models and the limitations of feature learning methods on the LTA task. Moreover, PALM outperforms other LLM-based methods, which demonstrates the effectiveness of our approach to provide ample visual context to LLMs, equipped with action recognition and image captioning models. + +In Fig. 6, we perform a qualitative comparison between the Ego4D baseline using SlowFast [18] and PALM. The first example exemplifies a scenario where the person mainly interacts with a single object, in this case, a brick. Both methods can accurately predict the future noun "brick". In the second scenario, the person is drawing a painting on a wall, interacting with multiple objects such as a paintbrush, a container, and the wall. PALM predicts the repeated sequence of actions, unlike SlowFast which incorrectly anticipates unrelated actions such as putting a string, touching a nail gun, or throwing a soap, none of which are present in the input video. This also demonstrates the importance of previous actions in the anticipation of future actions. The final scenario portrays the person adjusting a wire in a repair shop, a task involving numerous items. While SlowFast and PALM recognize central objects such as a screwdriver, they both fail to predict a coherent sequence of repair actions. We hypothesize that a language model with expert knowledge can produce more meaningful anticipations, particularly in the context of repairing mechanics, where it is crucial to execute actions in a precise order. More qualitative results are shown in supplementary. + +# 4.3 Ablations + +We perform various ablation studies regarding action recognition, example selection methods, LLM models, and prompt contents. We evaluate on the validation set of Ego4D v2. + +Action Recognition. The action recognition module is an essential part of our framework PALM as we utilize LLMs for predicting future actions and the accuracy of these predictions is highly influenced by the sequence of previous actions. We evaluate various video encoders trained on egocentric videos, including MViT [15], SlowFast [16], StillFast [49], EgoVLP [34], and EgoVLPv2 [46]. It is important to note that we employ a frozen video encoder without any finetuning, as depicted in Fig. 4. + +Tab. 2 presents the Top-1 Accuracy $(\%)$ and the edit distance of each video encoder. EgoVLP emerges as the top performer, achieving the best accuracy + +Table 2: Ablation on action recognition module. We compare different pretrained video encoders by reporting the Top-1 accuracy (Acc. ↑) and edit distance (ED ↓) for verbs, nouns, and actions. For comparative purposes, we also include the ED when ground truth actions (GT action) are used in the prompts. The best Acc. and ED are highlighted in bold except those from GT action. + +
MethodVerb Acc.Noun Acc.Action Acc.Verb EDNoun EDAction ED
MViT [15]19.872.550.510.72910.85970.9630
SlowFast [16]19.4214.653.120.72520.74720.9278
StillFast [49]19.1219.414.060.73560.73690.9241
EgoVLPv2 [46]33.8440.6316.310.71820.66510.8937
EgoVLP [34]40.3245.5320.630.71110.64650.8819
GT action---0.66630.58870.8319
+ +and edit distance among all encoders. EgoVLP utilizes a custom contrastive loss to align egocentric video frames with narrations, effectively establishing a connection between visual and textual embeddings. However, EgoVLPv2 utilizes task-specific loss for question-answering tasks built on the EgoVLP setting, which results in slightly less informative video features for action recognition. Unlike EgoVLP and EgoVLPv2, other video encoders are solely trained on video frames, making it challenging to draw robust connections between their video features and textual information (verbs and nouns). Based on these results, we select EgoVLP as our video encoder. + +It is expected that better accuracy in action recognition leads to a reduced edit distance, as accurate estimation of past actions is crucial for predicting future actions. Notably, the use of ground truth (GT) actions significantly outperforms all video encoders, which highlights the enhancement of action recognition as the foremost challenge within our framework. We believe that fine-tuning the video encoder on specific egocentric datasets would likely enhance recognition accuracy. + +Table 3: Ablation on example selection strategies exhibits that MMR outperforms other selection methods. + +
MethodVerb EDNoun EDAction ED
Random0.73620.67470.9002
Similarity0.72040.66610.8923
MMR0.71790.66470.8908
+ +Example Selection. We compare different example selection strategies for in-context learning (ICL), including random, similarity, and maximal marginal relevance (MMR) [59]. The random method uniformly selects examples from the training set, whereas the similarity method selects the most semantically similar example to the query. MMR considers both the similarity and diversity of + +examples, striving for a balance between them. Tab. 3 demonstrates that MMR outperforms the other selection methods. + +Table 4: Ablation on LLMs shows that larger LLMs perform better on the LTA task. + +
MethodVerb EDNoun EDAction ED
GPT-Neo (125M)0.76570.68670.9119
GPT-Neo (1.3B)0.73140.67590.8970
GPT-Neo (2.7B)0.71780.67470.8943
Llama 2 (7B)0.71790.66470.8908
Llama 2 (13B)0.69580.66340.8867
+ +LLM Ablation. We ablate the impact of LLMs by comparing different open-source models, primarily GPT-Neo [4] and Llama 2 [54] families. Tab. 4 indicates that larger language models are more capable of predicting future actions with their improved capacity for generalization across novel scenarios. This observation aligns with the performance of LLMs on general NLP benchmarks. Taking into account the computational resources at our disposal, we select Llama 2 with 7B parameters. + +Table 5: Ablation on the content of the prompt. We use abbreviations for each content in the input prompt: $\mathcal{A}$ - action sequence, $\mathcal{C}$ - captions + +
Prompt ContentVerb EDNoun EDAction ED
A0.72030.66880.8947
C0.70880.69010.9064
A+C0.71790.66470.8908
+ +Prompt Contents. We examine different content included in the input prompt to validate the contribution of each module to final performance. PALM utilizes recognized past actions and generated image captions to compose prompts. The action sequence describes the temporal flow of activity in (verb, noun) format, while captions narrate visual information relevant to human. In Tab. 5, the action sequence with captions $(\mathcal{A} + \mathcal{C})$ achieves the lowest action ED, which thereby is used in our framework. Notably, caption-only prompt $(\mathcal{C})$ reaches the lowest verb ED, showing that narrations are capable of capturing dynamic motions of human activity, while action-only prompt $(\mathcal{A})$ effectively detects key objects interacted with human supported by relatively lower noun ED. This indicates that action sequence and narrations provide a beneficial complement, thereby supporting our decision to incorporate both in the prompts. + +Table 6: Comparison to the SOTA on EK-55 and EGTEA. Performance is evaluated using the mAP metric $(\uparrow)$ . + +
MethodEK-55EGTEA
ALLFREQRAREALLFREQRARE
I3D [8]32.753.323.072.179.353.3
ActionVLAD [17]29.853.518.673.379.058.6
Timeception [22]35.655.926.174.179.759.7
VideoGraph [23]22.549.414.067.777.147.2
EGO-TOPO [38]38.056.929.273.580.754.7
Anticipatr [41]39.158.129.176.883.355.1
AntGPT [63]40.158.831.980.284.872.9
PALM (ours)40.459.330.380.785.073.5
+ +# 4.4 Evaluation on EK-55 and EGTEA + +Evaluation. For EK-55 [12] and EGTEA [33], we adhere to the experimental setup established by EGO-TOPO [38] and employ mean average precision (mAP) for multi-label classification as the evaluation metric. The initial $R\%$ ( $R = [25\%, 50\%, 75\%]$ ) of each video serves as input, with the objective of predicting the set of actions occurring in the remaining $(100 - R)\%$ of the video. Notably, actions here are confined to verb predictions only without noun, following previous conventions. We calculate the mAP on the validation set averaged over all $R$ , examining various performance aspects, such as all target actions (ALL), frequently occurring actions (FREQ), and rarely occurring actions (RARE). + +Results. As shown in Tab. 6, our method achieves SOTA results in the ALL and FREQ section while exhibiting comparable performance in the RARE section. The results demonstrate the effectiveness of PALM by capturing additional visual information with image captioning models. This aligns with the ablation study regarding prompt contents in Sec. 4.3 where captions contribute better on verb predictions. We observe that verb in RARE section are seldom represented in the captions, which results in relatively lower mAP in RARE section. Thus, we believe fine-tuning image captioning model or employing better video captioning model will further improve the performance accordingly. + +# 5 Conclusion + +We present PALM, a framework that leverages vision-language and large language models for long-term action anticipation. PALM achieves the best performance on the Ego4D benchmark, demonstrating the effectiveness of performing the LTA task in the language domain and the benefits of using foundation models. Our model also shows competitive results on two other benchmarks, EK-55 and EGTEA. As a result, we are optimistic that our work will spur further research in using commonsense knowledge extracted from LLMs for video understanding. + +# References + +1. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022) 2 +2. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023) 4, 9 +3. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. If you use this software, please cite it using these metadata 58 (2021) 3 +4. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. https://doi.org/10.5281/zenodo.5297715 (2021). https://doi.org/10.5281/ zenodo.5297715, https://doi.org/10.5281/zenodo.5297715, If you use this software, please cite it using these metadata. 13 +5. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021) 3 +6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 3, 8 +7. Carbonell, J., Goldstein, J.: The use of mmr, diversity-based reranking for reordering documents and producing summaries. In: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval. pp. 335-336 (1998) 2, 8 +8. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 14 +9. Chen, G., Zheng, Y.D., Wang, J., Xu, J., Huang, Y., Pan, J., Wang, Y., Wang, Y., Qiao, Y., Lu, T., et al.: Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292 (2023) 4, 9, 10 +0. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021) 3 +1. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) 3 +2. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(11), 4125-4141 (2020) 2, 3, 8, 9, 14 +3. Damerau, F.J.: A technique for computer detection and correction of spelling errors. Communications of the ACM 7(3), 171-176 (1964) 11 +4. Das, S., Ryoo, M.S.: Video+ clip baseline for ego4d long-term action anticipation. arXiv preprint arXiv:2207.00579 (2022) 9 +5. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 2, 4, 11, 12 + +16. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) 4, 11, 12 +17. Girdhar, R., Ramanan, D., Gupta, A., Sivic, J., Russell, B.: Actionvlad: Learning spatio-temporal aggregation for action classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 971-980 (2017) 14 +18. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022) 2, 3, 4, 6, 8, 9, 10, 11 +19. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) 8 +20. Huang, D., Hilliges, O., Van Gool, L., Wang, X.: Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545 (2023) 2 +21. Huang, W., Abbeel, P., Pathak, D., Mordatch, I.: Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In: International Conference on Machine Learning. pp. 9118-9147. PMLR (2022) 2 +22. Hussein, N., Gavves, E., Smeulders, A.W.: Timeception for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 254-263 (2019) 14 +23. Hussein, N., Gavves, E., Smeulders, A.W.: Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143 (2019) 14 +24. Ichter, B., Brohan, A., Chebotar, Y., Finn, C., Hausman, K., Herzog, A., Ho, D., Ibarz, J., Irpan, A., Jang, E., et al.: Do as i can, not as i say: Grounding language in robotic affordances. In: 6th Annual Conference on Robot Learning (2022) 3 +25. Keunwoo Peter, Y.: Videooblip is a large vision-language model based on blip-2 that can generate texts conditioned on videos. https://github.com/yukw777/VideoBLIP (2021). https://doi.org/https://orcid.org/0000-0002-4387-2410, https://github.com/yukw777/VideoBLIP, If you use this software, please cite it using these metadata. 6 +26. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 8 +27. Leo, M., Medioni, G., Trivedi, M., Kanade, T., Farinella, G.M.: Computer vision for assistive technologies. Computer Vision and Image Understanding 154, 1-15 (2017) 2 +28. Levenshtein, V.I., et al.: Binary codes capable of correcting deletions, insertions, and reversals. In: Soviet physics doklady. vol. 10, pp. 707-710. Soviet Union (1966) 11 +29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 3 +30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 3 + +31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019) 3 +32. Li, S., Puig, X., Paxton, C., Du, Y., Wang, C., Fan, L., Chen, T., Huang, D.A., Akyurek, E., Anandkumar, A., et al.: Pre-trained language models for interactive decision-making. Advances in Neural Information Processing Systems 35, 31199-31212 (2022) 2 +33. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018) 3, 8, 9, 14 +34. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022) 4, 7, 11, 12 +35. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 3 +36. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023) 4, 9 +37. Meditskos, G., Plans, P.M., Stavropoulos, T.G., Benois-Pineau, J., Buso, V., Kompatsiaris, I.: Multi-modal activity recognition from egocentric vision, semantic enrichment and lifelogging applications for the care of dementia. Journal of Visual Communication and Image Representation 51, 169-190 (2018) 2 +38. Nagarajan, T., Li, Y., Feichtenhofer, C., Grauman, K.: Ego-topo: Environment affordances from egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 163-172 (2020) 9, 14 +39. Nakazawa, A., Honda, M.: First-person camera system to evaluate tender dementia-care skill. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) 2 +40. Naveed, H., Khan, A.U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Barnes, N., Mian, A.: A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 (2023) 8 +41. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022) 14 +42. Neumann, L., Zisserman, A., Vedaldi, A.: Future event prediction: If and when. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) 2 +43. OhnBar, E., Kitani, K., Asakawa, C.: Personalized dynamics models for adaptive assistive navigation systems. In: Conference on Robot Learning. pp. 16-39. PMLR (2018) 2 +44. Pasca, R.G., Gavryushin, A., Hamza, M., Kuo, Y.L., Mo, K., Van Gool, L., Hilliges, O., Wang, X.: Summarize the past to predict the future: Natural language descriptions of context boost multimodal object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18286-18296 (2024) 2 +45. Patel, D., Eghbalzadeh, H., Kamra, N., Iuzzolino, M.L., Jain, U., Desai, R.: Pre-trained language models as visual planners for human assistance. arXiv preprint arXiv:2304.09179 (2023) 2 + +46. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023) 2, 4, 11, 12 +47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 3 +48. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners (2019) 3 +49. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023) 2, 4, 11, 12 +50. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021) 2 +51. Ryoo, M.S., Fuchs, T.J., Xia, L., Aggarwal, J.K., Matthies, L.: Robot-centric activity prediction from first-person videos: What will they do to me? In: Proceedings of the tenth annual ACM/IEEE international conference on human-robot interaction. pp. 295–302 (2015) 2 +52. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020) 8 +53. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 3 +54. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 3, 7, 13 +55. Wang, S., Zhao, Q., Do, M.Q., Agarwal, N., Lee, K., Sun, C.: Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627 (2023) 4 +56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 2 +57. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. In: International Conference on Learning Representations (2021) 3 +58. Yao, Y., Xu, M., Choi, C., Crandall, D.J., Atkins, E.M., Dariush, B.: Egocentric vision-based future vehicle localization for intelligent driving assistance systems. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9711-9717. IEEE (2019) 2 +59. Ye, X., Iyer, S., Celikyilmaz, A., Stoyanov, V., Durrett, G., Pasunuru, R.: Complementary explanations for effective in-context learning. arXiv preprint arXiv:2211.13892 (2022) 2, 8, 12 +60. Zeng, A., Attarian, M., Choromanski, K.M., Wong, A., Welker, S., Tombari, F., Purohit, A., Ryoo, M.S., Sindhwani, V., Lee, J., et al.: Socratic models: Composing zero-shot multimodal reasoning with language. In: The Eleventh International Conference on Learning Representations (2022) 3 + +61. Zhan, K., Faux, S., Ramos, F.: Multi-scale conditional random fields for first-person activity recognition. In: 2014 IEEE international conference on pervasive computing and communications (PerCom). pp. 51-59. IEEE (2014) 2 +62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 3 +63. Zhao, Q., Zhang, C., Wang, S., Fu, C., Agarwal, N., Lee, K., Sun, C.: Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368 (2023) 4, 9, 10, 14 +64. Zhou, C., Li, Q., Li, C., Yu, J., Liu, Y., Wang, G., Zhang, K., Ji, C., Yan, Q., He, L., et al.: A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419 (2023) 3 +65. Zhou, L., Palangi, H., Zhang, L., Hu, H., Corso, J., Gao, J.: Unified vision-language pre-training for image captioning and vqa. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13041-13049 (2020) 3 \ No newline at end of file diff --git a/2024/PALM_ Predicting Actions through Language Models/images.zip b/2024/PALM_ Predicting Actions through Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a30d2bbfd78d00b65c8e6267b7d196e27b481280 --- /dev/null +++ b/2024/PALM_ Predicting Actions through Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1c2df9457adc3cfe52ab24b7f43209f59aaf7839c9bf7881329f0f7f6e7b95b +size 479106 diff --git a/2024/PALM_ Predicting Actions through Language Models/layout.json b/2024/PALM_ Predicting Actions through Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..73bfb6e740fb7c128262812a718d209033f8dfe8 --- /dev/null +++ b/2024/PALM_ Predicting Actions through Language Models/layout.json @@ -0,0 +1,10036 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 145 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 145 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "spans": [ + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "text", + "content": "Sanghwan Kim" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "text", + "content": ", Daoji Huang" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "text", + "content": ", Yongqin Xian" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "text", + "content": ", Otmar Hilliges" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "text", + "content": ", Luc Van Gool" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "inline_equation", + "content": "^{1,3,4}" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "text", + "content": ", and Xi Wang" + }, + { + "bbox": [ + 194, + 168, + 417, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 276, + 201, + 337, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 201, + 337, + 213 + ], + "spans": [ + { + "bbox": [ + 276, + 201, + 337, + 213 + ], + "type": "text", + "content": "1 ETH Zürich" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 287, + 213, + 326, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 213, + 326, + 224 + ], + "spans": [ + { + "bbox": [ + 287, + 213, + 326, + 224 + ], + "type": "text", + "content": "2 Google" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 278, + 224, + 335, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 224, + 335, + 235 + ], + "spans": [ + { + "bbox": [ + 278, + 224, + 335, + 235 + ], + "type": "text", + "content": "3 KU Leuven" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 272, + 235, + 341, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 235, + 341, + 246 + ], + "spans": [ + { + "bbox": [ + 272, + 235, + 341, + 246 + ], + "type": "text", + "content": "4 INSAIT, Sofia" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 280, + 452, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 280, + 452, + 510 + ], + "spans": [ + { + "bbox": [ + 160, + 280, + 452, + 510 + ], + "type": "text", + "content": "Abstract. Understanding human activity is a crucial yet intricate task in egocentric vision, a field that focuses on capturing visual perspectives from the camera wearer's viewpoint. Traditional methods heavily rely on representation learning that is trained on a large amount of video data. However, a major challenge arises from the difficulty of obtaining effective video representation. This difficulty stems from the complex and variable nature of human activities, which contrasts with the limited availability of data. In this study, we introduce PALM, an approach that tackles the task of long-term action anticipation, which aims to forecast forthcoming sequences of actions over an extended period. Our method PALM incorporates an action recognition model to track previous action sequences and a vision-language model to articulate relevant environmental details. By leveraging the context provided by these past events, we devise a prompting strategy for action anticipation using large language models (LLMs). Moreover, we implement maximal marginal relevance for example selection to facilitate in-context learning of the LLMs. Our experimental results demonstrate that PALM surpasses the state-of-the-art methods in the task of long-term action anticipation on the Ego4D benchmark. We further validate PALM on two additional benchmarks, affirming its capacity for generalization across intricate activities with different sets of taxonomies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "spans": [ + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "text", + "content": "Keywords: egocentric vision " + }, + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "text", + "content": " video understanding " + }, + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "text", + "content": " action anticipation " + }, + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 452, + 544 + ], + "type": "text", + "content": " large language model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 566, + 229, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 229, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 229, + 578 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "As wearable technology, such as smart glasses and body-worn cameras, becomes more prevalent, the availability of egocentric videos has grown significantly. These videos capture activities from the camera wearer's viewpoint, providing a unique perspective that features individual behaviors in various contexts. Egocentric vision holds immense promise for understanding how humans execute specific tasks and activities. It enables numerous applications such as real-time" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 211, + 116, + 402, + 206 + ], + "blocks": [ + { + "bbox": [ + 211, + 116, + 402, + 206 + ], + "lines": [ + { + "bbox": [ + 211, + 116, + 402, + 206 + ], + "spans": [ + { + "bbox": [ + 211, + 116, + 402, + 206 + ], + "type": "image", + "image_path": "445e7a530ebaf75d625e14375332c35d3b460b207f5ecb592f89e2b6c5cf9ecb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 258 + ], + "type": "text", + "content": "Fig. 1: PALM for long-term action anticipation. Given a video input, PALM uses a vision-language model (VLM) and an action recognition model (ARM) to compile a text prompt that describes the past video content. A large-language model (LLM) then takes the prompt and predicts future actions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "content": "systems that anticipate and respond to user needs [50, 51, 58], monitoring patient activities [37, 39, 61], and providing personalized instructions and procedures [27, 42, 43]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 437 + ], + "type": "text", + "content": "Existing works on understanding human activities mostly rely on learning representations from large-scale video datasets [15, 46, 49, 56]. However, these approaches face two critical challenges. First, the complexity and variability of human activities make it challenging to obtain comprehensive representations, particularly for longer videos. Second, reliance on large datasets limits the model's capability to generalize to less common, long-tail classes and unseen scenarios. In this context, there is a pressing need for a more versatile approach to human activity understanding, adept at handling the complicated nature of human activities without overly relying on large-scale video data. Therefore, recent works have started to exploit the procedure knowledge and generalization ability of large language models (LLMs) [1, 21, 32, 44, 45]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "Most prior works based on LLMs have been largely confined to generating instructions and plans given a high-level task description (e.g., \"prepare the dinner table\") and have not been tailored for intricate human activity understanding from videos. Indeed, there is a significant gap in the literature when it comes to the use of LLMs for grounded activity understanding, especially in challenges involving understanding long-term action sequences. In this work, we efficiently boost the versatile ability of LLMs in long-term action anticipation [12, 18, 20] by providing sufficient visual context in discrete text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 1, we introduce a simple yet effective framework called PALM that combines descriptions of past action sequences and corresponding narrations to bridge the gap between visual and textual domains. In detail, PALM takes an input video with annotated action periods and leverages a vision-language model and an action recognition model to describe past events. We then design a prompt strategy for large language models (LLMs) to anticipate future actions given the context of these past events. To facilitate the predictive capabilities of LLMs, we use maximal marginal relevance (MMR) [7, 59] to retrieve similar and diverse examples from the training set, enabling in-context learning for the LLMs. Through this prompting strategy, our system can tackle intricate tasks in understanding human activities, benefiting from existing foun" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "dition models. Experimental results demonstrate that our method PALM significantly improves performance in long-term action anticipation, as evaluated using the Ego4D benchmark [18]. Further evaluations on two additional benchmarks, EPIC-KITCHEN [12] and EGTEA [33], reveal the generalization ability of PALM across different contexts and datasets. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 190, + 481, + 273 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 190, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 481, + 224 + ], + "type": "text", + "content": "- We propose an effective textural representation of past events, consisting of both action sequences and accompanying narrations that provide additional context and environmental details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 225, + 481, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 225, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 225, + 481, + 248 + ], + "type": "text", + "content": "- We design a novel prompting strategy that enables efficient in-context learning for large language models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 249, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 273 + ], + "type": "text", + "content": "- PALM outperforms existing state-of-the-art methods in the task of long-term action anticipation, which is part of the Ego4D benchmark." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 289, + 237, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 237, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 237, + 301 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 253, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 253, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 253, + 323 + ], + "type": "text", + "content": "2.1 Foundation Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 331, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 378 + ], + "type": "text", + "content": "Foundation models serve as the cornerstone for various downstream tasks across different data modalities [5,64]. Here we consider two types of foundation models: vision-language models (VLMs) to generate the textual narrations of video and large language models (LLMs) to predict future action given visual context." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 533 + ], + "type": "text", + "content": "Vision-language model. VLMs utilize complementary characteristics of image and text datasets to perform multimodal tasks such as image captioning, visual question-answering, and visual reasoning [31, 35, 47]. Various works [10, 57, 65] attempt to unify vision-language tasks with a single model, overcoming the limited utility of previous models. In this paper, the image or video captioning ability of VLMs is utilized to directly transform visual information into narrative text descriptions. We empirically found that BLIP family [11, 29, 30] provides concise and informative captions. For instance, BLIP [30] proposes a novel encoder-decoder architecture specialized for multimodal pertaining, and BLIP-2 [29] significantly reduces the trainable parameters by using a querying transformer to bridge the modality gap between frozen vision models and language models. InstructBLIP [11] further performs instruction tuning on BLIP-2, enabling general-purpose models with a unified natural language interface." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Large language model. Language foundation models [3,6,48,53,54,62] have had a significant impact on the NLP field, demonstrating surprising ability to generalize across unseen tasks. With their extensive training data and large parameter size, LLMs have demonstrated the ability to learn from examples provided in input prompts, a concept known as in-context learning [6]. LLMs have also shown success in the tasks of decision-making and planning. For example, SayCan [24] combines a language model with learned low-level skills for grounded planning, while Socratic Models [60] formulate action prediction as text completion problems from captioned images. In this work, we explore how such procedure knowledge can be exploited in the complex action anticipation task and showcase the effectiveness of in-context learning for this purpose." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 316, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 316, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 316, + 129 + ], + "type": "text", + "content": "2.2 Long-term Action Anticipation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 479, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 479, + 185 + ], + "type": "text", + "content": "Given an input video, the long-term action anticipation (LTA) task aims to predict the next actions in the future. In this work, we explore various action recognition models to estimate accurate action sequences and extensively compare our method with previous action anticipation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 316 + ], + "type": "text", + "content": "Action recognition. SlowFast [16] proposes to fuse two channels of input that consist of a low frame and a high frame rate video, using 3D convolutional kernels. Similar to Slowfast, StillFast [49] adopts a specific architecture with two branches to effectively process high-resolution still images and low-resolution video frames. On the other hand, MViT [15] employs several channel-resolution scale stages to obtain a multiscale pyramid of features, based on vision transformers. Compared to previous works solely optimized on video inputs, EgoVLP [34] and EgoVLPv2 [46] pre-train video and text encoders with contrastive loss specialized for the egocentric domain. Each encoder provides useful feature representations for downstream video tasks such as natural language queries, video question-answering, and video summarization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 460 + ], + "type": "text", + "content": "Action anticipation. Baselines provided by Ego4D [18] employ the SlowFast [16] video encoder to extract features from sampled clips, followed by transformer layers and classification heads to predict the verb-noun pair of future action sequence. HierVL [2] utilizes a hierarchical contrastive training objective that encourages text-visual alignment at both the clip and video levels. ICVAE [36] proposes using a predicted scenario, which summarizes the activity depicted in the video, as a condition for action prediction. A few concurrent approaches [9, 55, 63] leverage the generalization abilities and abundant knowledge of LLMs to predict future actions. Compared to these LLM-based methods, PALM differentiates itself by utilizing vision-language model for effective extraction of visual context and employing in-context learning in combination with MMR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 482, + 201, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 201, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 201, + 495 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 479, + 556 + ], + "type": "text", + "content": "We introduce PALM, a comprehensive framework designed for video understanding and action anticipation. PALM comprises three key modules, as illustrated in Fig. 2: an image captioning module (Sec. 3.2), an action recognition module (Sec. 3.3), and an action anticipation module based on LLMs (Sec. 3.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 479, + 665 + ], + "type": "text", + "content": "Given an untrimmed input video, the action recognition module generates labels for past actions in the form of verb and noun pairs (eight past actions in Fig. 2). The action recognition model is trained to produce valid verb and noun pairs as predefined by each dataset. Simultaneously, the image captioning module generates captions from the input video, providing additional visual information relevant to the events, termed visual context. The prompt is then constructed using the visual context and past actions, passed to LLMs to predict future actions. Additionally, we incorporate exemplars from the training set to facilitate in-context learning for LLMs. We demonstrate the effectiveness of our" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 475, + 288 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 475, + 288 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 475, + 288 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 475, + 288 + ], + "type": "image", + "image_path": "5795b18eb1b49b39c88317000ad4c8ea310959ecea7496a448affae6437be336.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 296, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 352 + ], + "type": "text", + "content": "Fig. 2: Overview of PALM. Given an input video, an image captioning model and an action recognition model convert visual information of past activities into text format (visual context and past actions). Then, LLM utilizes visual context and past actions to predict the sequence of future actions. The image captioning model is frozen and in-context learning is adapted for LLM prompting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 361, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 384 + ], + "type": "text", + "content": "approach on the long-term action anticipation (LTA) task, achieving state-of-the-art (SOTA) performance on Ego4D benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 401, + 246, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 246, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 246, + 414 + ], + "type": "text", + "content": "3.1 Task Description" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "text", + "content": "The LTA task involves predicting a chronological sequence of future actions given an untrimmed video " + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "text", + "content": ". The input video " + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "text", + "content": " comprises multiple action segments with variable lengths, where each action segment corresponds to a single action (verb and noun pair). The segment boundaries for each action in the videos are provided as part of the input. Formally, after observing the video until time " + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "inline_equation", + "content": "e.g., V_{:t}" + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "text", + "content": "), the LTA model sequentially predicts " + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 420, + 482, + 503 + ], + "type": "text", + "content": " future actions that the camerawearer is likely to perform:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 281, + 503, + 480, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 503, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 281, + 503, + 480, + 517 + ], + "type": "interline_equation", + "content": "\\left\\{\\hat {n} _ {z}, \\hat {v} _ {z} \\right\\} _ {z = 1} ^ {Z} \\tag {1}", + "image_path": "348d59ab440fe7731ed628a4621f1ee489a033bc8cce944046df650c72ac4e49.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\hat{n}_z\\in \\mathcal{N}" + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\hat{v}_z\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "content": ", given the predefined noun and verb taxonomy " + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "content": ". The input video " + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "inline_equation", + "content": "V_{:t}" + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 521, + 482, + 546 + ], + "type": "text", + "content": " action segments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 562, + 293, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 293, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 293, + 574 + ], + "type": "text", + "content": "3.2 Image Captioning Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "type": "text", + "content": "The past action labels do not capture all the visual information and context presented in the video (e.g., the location of the person or the background objects). Our idea is to use an image captioning model to capture the action context, providing more informative prompts related to the action performer for LLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "To enhance the descriptive power of captioning models, we formulate the image captioning task as a question-answering problem. Fig. 3 displays a set of questions related to six key concepts that the models should aim to answer given" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 185, + 117, + 427, + 197 + ], + "blocks": [ + { + "bbox": [ + 185, + 117, + 427, + 197 + ], + "lines": [ + { + "bbox": [ + 185, + 117, + 427, + 197 + ], + "spans": [ + { + "bbox": [ + 185, + 117, + 427, + 197 + ], + "type": "text", + "content": "Question Dictionary " + }, + { + "bbox": [ + 185, + 117, + 427, + 197 + ], + "type": "inline_equation", + "content": "= \\{" + }, + { + "bbox": [ + 185, + 117, + 427, + 197 + ], + "type": "text", + "content": " Location: Where is the person? Detection: What objects are in this image? Action: What is the person doing in this image? Prediction: What will the person do later in this image? Interaction: What is the person interacting with in this image? Intention: What does the person want to do in this image?" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 130, + 230, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 479, + 289 + ], + "type": "text", + "content": "an input image: location, detection, action, prediction, interaction, and intention. Given a question " + }, + { + "bbox": [ + 130, + 230, + 479, + 289 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 230, + 479, + 289 + ], + "type": "text", + "content": ", the conditional text is formed as \"Question: " + }, + { + "bbox": [ + 130, + 230, + 479, + 289 + ], + "type": "inline_equation", + "content": "< Q >" + }, + { + "bbox": [ + 130, + 230, + 479, + 289 + ], + "type": "text", + "content": "? Answer: \" where the answer part will be filled by the captioning model. After extensive experiments, we select intention-based question for conditional generation and VideoBLIP [25] as our image captioning model (see supplementary)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 374 + ], + "type": "text", + "content": "By posing these questions to the captioning models, we anticipate extracting useful descriptions beneficial for LLMs to predict future actions. In practice, the conditional caption generation yields consistent and accurate descriptions of the action context for later prediction. For image captioning, we use the middle frame of each clip segment to generate the captions, based on the assumption that the middle frame is highly likely to be the moment when the action is happening." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 394, + 301, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 301, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 301, + 407 + ], + "type": "text", + "content": "3.3 Action Recognition Module" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 205, + 438, + 410, + 557 + ], + "blocks": [ + { + "bbox": [ + 130, + 201, + 482, + 222 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 222 + ], + "type": "text", + "content": "Fig. 3: List of questions asking for information related to key concepts of egocentric image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 438, + 410, + 557 + ], + "lines": [ + { + "bbox": [ + 205, + 438, + 410, + 557 + ], + "spans": [ + { + "bbox": [ + 205, + 438, + 410, + 557 + ], + "type": "image", + "image_path": "7cf9fd26667fd9cafbe07949498e02a6febba0d53c62f33862516d35fa84b929.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "type": "text", + "content": "Fig. 4: Action recognition network. A frozen video encoder generates visual features from " + }, + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "type": "text", + "content": " continuous action segments, followed by transformer layers and classification heads." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "The accurate past action descriptions are crucial, as they provide essential contextual information about past events in a given video input. Our model recognizes actions from the observed video clips, taking inspiration from the baseline model used in the Ego4D paper [18]. Visual features are extracted from a video encoder given " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " consecutive action segments (" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " in Fig. 4). We adopt" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "the off-the-shelf video encoder of EgoVLP [34], a vision-language model pretrained on Ego4D. Next, transformer layers are used to combine multiple visual features into an aggregated feature, followed by " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " classification heads to recognize the verbs and nouns independently. The video encoder is frozen, and the rest of the model is trained with cross-entropy loss summed over " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " classification heads. We also try to exploit a text encoder by formulating the action recognition task as matching a proper action label with an input video, which does not improve the accuracy (See supplementary)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": "We empirically find that recognizing actions in isolation falls short in effectiveness compared to recognizing " + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": " actions within a continuous sequence. This is because actions in neighboring segments are often highly correlated, reflecting the natural continuity embedded in human activities. Thus, we use a sliding input video window that includes " + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": " segments with the window striding of one. The top-1 action is determined by identifying the most common action from the classifications across different windows. See supplementary for more details and ablation results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 329, + 302, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 329, + 302, + 341 + ], + "spans": [ + { + "bbox": [ + 131, + 329, + 302, + 341 + ], + "type": "text", + "content": "3.4 Action Anticipation Module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 496 + ], + "type": "text", + "content": "In this section, we discuss the details of action anticipation with LLMs, from how the prompt is formatted to the details of inference and post-processing of raw outputs. Llama 2 (7B) [54], an open-source LLM, is employed in our framework. Prompt Design. Following common practice, we design the prompt with a few examples to enable in-context learning (ICL), maximizing references given to LLMs. Fig. 5 shows the template of our prompts, which consists of an instruction, a few examples, and a query. The instruction includes the task description and defines the input and output for LLMs. The examples for ICL, selected from the training set, include past and future actions (ground truth) and corresponding textual descriptions generated by the captioning model, which we call narrations. The query includes recognized past actions and generated narrations, and future actions are left empty for LLMs to complete." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 140, + 515, + 476, + 635 + ], + "blocks": [ + { + "bbox": [ + 140, + 515, + 476, + 635 + ], + "lines": [ + { + "bbox": [ + 140, + 515, + 476, + 635 + ], + "spans": [ + { + "bbox": [ + 140, + 515, + 476, + 635 + ], + "type": "image", + "image_path": "d3bae9eac5b178ad79edeacef38b7c564131452287cbde34e626c8ff22efd80b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 643, + 482, + 677 + ], + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 677 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 677 + ], + "type": "text", + "content": "Fig. 5: Prompt structure for LLMs. The prompt is composed of an instruction, examples retrieved from the training set, and the query related to the prediction video. Each video description consists of narrations, past actions, and future actions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Example Selection. Incorporating some examples that demonstrate proper generation will greatly improve the performance of LLMs by providing additional context and references [6, 40]. Considering the diversity of egocentric videos, especially within the Ego4D dataset, randomly selected examples are unlikely to generalize well on various scenarios within the test set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "text", + "content": "Therefore, we utilize a maximal marginal relevance (MMR) [7, 59] based exemplar selection strategy. The main idea behind MMR is to select exemplars that are relevant to the query while exhibiting sufficient variety to provide robust generalization across the test set. To be specific, we iteratively select a set of examples " + }, + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "inline_equation", + "content": "p_i \\in T" + }, + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "text", + "content": " from the training set " + }, + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "text", + "content": " that are semantically close to the query prompt " + }, + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 175, + 480, + 248 + ], + "type": "text", + "content": " but also diverse enough to provide additional information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 255, + 480, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 255, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 200, + 255, + 480, + 283 + ], + "type": "interline_equation", + "content": "p _ {j} = \\underset {p _ {j} \\in D / T} {\\arg \\max } \\left[ \\lambda S (q, p _ {j}) - (1 - \\lambda) \\underset {p _ {i} \\in T} {\\max } S \\left(p _ {i}, p _ {j}\\right) \\right], \\tag {2}", + "image_path": "6b077b6fe28d0f479b277985e6d2f1da921086af48833beade7d8c771d3acf78.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 480, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 480, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 288, + 480, + 358 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 288, + 480, + 358 + ], + "type": "text", + "content": " measures the semantic similarity between exemplars and the parameter " + }, + { + "bbox": [ + 130, + 288, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 288, + 480, + 358 + ], + "type": "text", + "content": " strikes a balance between similarity and diversity. One exemplar includes past and future action sequences and corresponding captions in natural language. We use MPNet [52] to extract text embeddings and calculate cosine similarity to measure semantic similarity. We compare various exemplar selection strategies in Sec. 4.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 480, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 480, + 455 + ], + "type": "text", + "content": "Fine-tuning LLM. To facilitate LLMs to generate proper action sequence, we fine-tune the LLMs based on the training set of each egocentric dataset. As fine-tuning whole parameters of the LLMs can be computationally inefficient and may lead to catastrophic forgetting [26], parameter-efficient fine-tuning is leveraged. Specifically, we utilize LoRA [19] with 8-bit quantization using a batch size of 8 and a learning rate of " + }, + { + "bbox": [ + 130, + 360, + 480, + 455 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 360, + 480, + 455 + ], + "type": "text", + "content": " for 10 epochs. The training set for fine-tuning consists of input prompts in the format of Fig. 5 with different exemplars and ground truth future actions as targets to be generated by the LLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 480, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 480, + 563 + ], + "type": "text", + "content": "Inference of LLM and Post-processing. Following the evaluation protocol of Ego4d [18], we generate multiple predictions for each test video. To reduce the variability in predictions, we prompt LLMs multiple times using the same prompt structure, as shown in Fig. 5, but with different exemplars. After inference, we map the raw output of LLMs to the label space of predefined verbs and nouns. One should note that the output of LLMs may not always fall within the given taxonomy, despite all input prompts containing verbs and nouns from the predefined domain. For simplicity, we adopt a rule-based mapping (see details in supplementary)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "In this section, we evaluate PALM on the long-term action anticipation (LTA) task, shedding light on the various design aspects. We conduct a comprehensive evaluation using the Ego4D dataset, which stands as the largest repository of egocentric videos. We report the final performance metrics on the Ego4D [18], EPIC-KITCHEN-55 [12], and EGTEA Gaze+ [33] benchmarks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 146, + 427, + 319 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Comparison to the SOTA on the Ego4D test set. PALM outperforms all baselines on the Ego4D LTA benchmark in respect of edit distance (ED " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 146, + 427, + 319 + ], + "lines": [ + { + "bbox": [ + 186, + 146, + 427, + 319 + ], + "spans": [ + { + "bbox": [ + 186, + 146, + 427, + 319 + ], + "type": "table", + "html": "
MethodVersionVerb EDNoun EDAction ED
Slowfast [18]v10.73890.78000.9432
VCLIP [14]v10.73890.76880.9412
ICVAE [36]v10.74100.73960.9304
HierVL [2]v10.72390.73500.9276
AntGPT [63]v10.65840.65460.8814
PALM (ours)v10.65590.64010.8613
Last actionv20.83290.72580.9438
Repeat actionv20.74910.70730.9155
Retrieve actionv20.74170.70410.9127
Slowfast [18]v20.71690.73590.9253
VideoLLM [9]v20.7210.7250.921
AntGPT [63]v20.65030.64980.8770
PALM (ours)v20.64710.61170.8503
", + "image_path": "6ace61cc0a2ed5d6ef6d7090d3ae3693362750a927004a389120f98965641eae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 342, + 205, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 205, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 205, + 352 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": "Ego4D [18]. With 243 hours of egocentric video content encompassing 53 distinct scenarios, Ego4D offers a rich variety of data with a predefined annotation of 117 verbs and 521 nouns for the LTA task. Ego4D is available in two versions, v1 and v2, where Ego4D v1 is a subset of Ego4D v2 with fewer videos and a smaller verb and noun set. We use both versions of the dataset to enable extensive comparison with previous methods and adhere to the identical training, validation, and test splits as specified in the benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 445, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 480 + ], + "type": "text", + "content": "EPIC-KITCHENS-55 (EK-55) [12]. This dataset comprises 55 hours of egocentric videos centered around cooking scenarios, featuring 125 annotated verbs and 352 nouns. We follow the same train and test splits as EGO-TOPO [38]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "content": "EGTEA Gaze+ (EGTEA) [33]. With approximately 26 hours of egocentric cooking videos, EGTEA includes 19 verbs and 53 nouns. Similar to EK-55, we employ the same train and test splits as described in EGO-TOPO [38]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 535, + 290, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 290, + 547 + ], + "type": "text", + "content": "4.2 Comparison to the SOTA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 553, + 306, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 306, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 306, + 564 + ], + "type": "text", + "content": "Baselines. We include three baselines:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 573, + 481, + 645 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 137, + 573, + 481, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 573, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 137, + 573, + 481, + 597 + ], + "type": "text", + "content": "- Last action outputs the last action of past action sequence as " + }, + { + "bbox": [ + 137, + 573, + 481, + 597 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 137, + 573, + 481, + 597 + ], + "type": "text", + "content": " future actions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 597, + 470, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 597, + 470, + 609 + ], + "spans": [ + { + "bbox": [ + 137, + 597, + 470, + 609 + ], + "type": "text", + "content": "- Repeat action repeats the given past " + }, + { + "bbox": [ + 137, + 597, + 470, + 609 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 137, + 597, + 470, + 609 + ], + "type": "text", + "content": " actions to fill " + }, + { + "bbox": [ + 137, + 597, + 470, + 609 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 137, + 597, + 470, + 609 + ], + "type": "text", + "content": " future actions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 609, + 481, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 609, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 137, + 609, + 481, + 645 + ], + "type": "text", + "content": "- Retrieve action retrieves the most semantically similar example in the training set (as in the example selection step) and uses their future actions as the prediction." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 653, + 359, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 359, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 359, + 665 + ], + "type": "text", + "content": "We also compare to three previous SOTA methods:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 175 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 402, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 402, + 128 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 402, + 128 + ], + "type": "text", + "content": "- Slowfast [18] is the baseline LTA model given by Ego4d." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 128, + 479, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 128, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 137, + 128, + 479, + 150 + ], + "type": "text", + "content": "- VideoLLM [9] converts multimodal data into a unified token sequence and utilizes fine-tuned LLMs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 152, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 175 + ], + "type": "text", + "content": "- AntGPT [63] leveraged action recognition model and fine-tuned LLMs to predict future actions." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 143, + 206, + 225, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 206, + 225, + 214 + ], + "spans": [ + { + "bbox": [ + 143, + 206, + 225, + 214 + ], + "type": "text", + "content": "① Single Object Scenario (1 object)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 248, + 164, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 248, + 164, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 248, + 164, + 254 + ], + "type": "text", + "content": "GT Past" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 168, + 217, + 468, + 246 + ], + "blocks": [ + { + "bbox": [ + 168, + 217, + 468, + 246 + ], + "lines": [ + { + "bbox": [ + 168, + 217, + 468, + 246 + ], + "spans": [ + { + "bbox": [ + 168, + 217, + 468, + 246 + ], + "type": "image", + "image_path": "f2534dc7abb416e8231d85550615eef7afb248d0c96c520f3e7ae3285eb7adfe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 257, + 165, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 257, + 165, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 257, + 165, + 262 + ], + "type": "text", + "content": "GT Future " + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 168, + 247, + 378, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 247, + 378, + 255 + ], + "spans": [ + { + "bbox": [ + 168, + 247, + 378, + 255 + ], + "type": "text", + "content": "(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 277, + 159, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 277, + 159, + 282 + ], + "spans": [ + { + "bbox": [ + 143, + 277, + 159, + 282 + ], + "type": "text", + "content": "(0.55)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 168, + 256, + 459, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 256, + 459, + 262 + ], + "spans": [ + { + "bbox": [ + 168, + 256, + 459, + 262 + ], + "type": "text", + "content": "(put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick), (move, brick), (take, brick)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 277, + 159, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 277, + 159, + 282 + ], + "spans": [ + { + "bbox": [ + 143, + 277, + 159, + 282 + ], + "type": "text", + "content": "(0.55)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 168, + 263, + 403, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 263, + 403, + 270 + ], + "spans": [ + { + "bbox": [ + 168, + 263, + 403, + 270 + ], + "type": "text", + "content": "(put, brick), (put, brick), (move, brick), (take, brick), (put, brick), (move, brick), (put, brick), (take, brick), (put, brick)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 274, + 165, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 274, + 165, + 279 + ], + "spans": [ + { + "bbox": [ + 143, + 274, + 165, + 279 + ], + "type": "text", + "content": "SlowFast" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 168, + 272, + 455, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 272, + 455, + 279 + ], + "spans": [ + { + "bbox": [ + 168, + 272, + 455, + 279 + ], + "type": "text", + "content": "put, brick), (put, brick), (put, brick), (adjust, brick), (take, brick), (take, brick), (take, brick), (take, brick), (put, brick)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 281, + 159, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 281, + 159, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 281, + 159, + 285 + ], + "type": "text", + "content": "(0.55)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 168, + 280, + 402, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 280, + 402, + 286 + ], + "spans": [ + { + "bbox": [ + 168, + 280, + 402, + 286 + ], + "type": "text", + "content": "(take, brick), (adjust, brick), (put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 289, + 159, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 289, + 159, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 289, + 159, + 295 + ], + "type": "text", + "content": "PALM" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 168, + 289, + 447, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 289, + 447, + 295 + ], + "spans": [ + { + "bbox": [ + 168, + 289, + 447, + 295 + ], + "type": "text", + "content": "(put, brick), (take, brick), (put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick), (take, brick)," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 143, + 297, + 159, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 297, + 159, + 301 + ], + "spans": [ + { + "bbox": [ + 143, + 297, + 159, + 301 + ], + "type": "text", + "content": "(0.4)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 168, + 296, + 391, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 296, + 391, + 301 + ], + "spans": [ + { + "bbox": [ + 168, + 296, + 391, + 301 + ], + "type": "text", + "content": "(put, brick), (put, brick), (take, brick), (put, brick), (put, brick), (put, brick), (put, brick)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 143, + 308, + 227, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 308, + 227, + 316 + ], + "spans": [ + { + "bbox": [ + 143, + 308, + 227, + 316 + ], + "type": "text", + "content": "② Few Object Scenario (2~5 objects)" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 168, + 317, + 469, + 346 + ], + "blocks": [ + { + "bbox": [ + 168, + 317, + 469, + 346 + ], + "lines": [ + { + "bbox": [ + 168, + 317, + 469, + 346 + ], + "spans": [ + { + "bbox": [ + 168, + 317, + 469, + 346 + ], + "type": "image", + "image_path": "6d539fde5b9c097e01d4f2f0a50e8e85fde79cfa9fd4526e662ee6c05dad0c76.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 143, + 349, + 164, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 349, + 164, + 355 + ], + "spans": [ + { + "bbox": [ + 143, + 349, + 164, + 355 + ], + "type": "text", + "content": "GTPast" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 168, + 349, + 451, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 349, + 451, + 356 + ], + "spans": [ + { + "bbox": [ + 168, + 349, + 451, + 356 + ], + "type": "text", + "content": "(dip, paintbrush), (put, container), (take, paintbrush), (put, paintbrush), (adjust, paintbrush), (take, container), (dip, container), (paint, wall)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 143, + 358, + 165, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 358, + 165, + 365 + ], + "spans": [ + { + "bbox": [ + 143, + 358, + 165, + 365 + ], + "type": "text", + "content": "GT Futur" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 168, + 358, + 465, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 358, + 465, + 365 + ], + "spans": [ + { + "bbox": [ + 168, + 358, + 465, + 365 + ], + "type": "text", + "content": "(paint, wall), (dip, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (dip, container), (paint, wall)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 168, + 366, + 444, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 366, + 444, + 374 + ], + "spans": [ + { + "bbox": [ + 168, + 366, + 444, + 374 + ], + "type": "text", + "content": "(paint, paintbrush), (paint, wall), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush)." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 143, + 380, + 165, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 380, + 165, + 385 + ], + "spans": [ + { + "bbox": [ + 143, + 380, + 165, + 385 + ], + "type": "text", + "content": "SlowFast" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 168, + 380, + 467, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 380, + 467, + 387 + ], + "spans": [ + { + "bbox": [ + 168, + 380, + 467, + 387 + ], + "type": "text", + "content": "take, container), (put, container), (hold, container), (put, string), (open, brush), (turn, container), (touch, nail gun), (hold, container), (throw, soap)," + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 168, + 388, + 457, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 388, + 457, + 396 + ], + "spans": [ + { + "bbox": [ + 168, + 388, + 457, + 396 + ], + "type": "text", + "content": "take, cereal), move, sheft), (push, wire), (break, cloth), (break, wire), (hold, circuit), (hold, container), (hold, garbage can), (remove, switch), (put, camera). (hold, cloth)." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 143, + 401, + 159, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 401, + 159, + 407 + ], + "spans": [ + { + "bbox": [ + 143, + 401, + 159, + 407 + ], + "type": "text", + "content": "PALM" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 168, + 400, + 454, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 400, + 454, + 407 + ], + "spans": [ + { + "bbox": [ + 168, + 400, + 454, + 407 + ], + "type": "text", + "content": "put paintbrush). (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall), (dip, paintbrush), (paint, wall)," + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 143, + 408, + 159, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 408, + 159, + 413 + ], + "spans": [ + { + "bbox": [ + 143, + 408, + 159, + 413 + ], + "type": "text", + "content": "(0.6)" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 168, + 408, + 432, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 408, + 432, + 414 + ], + "spans": [ + { + "bbox": [ + 168, + 408, + 432, + 414 + ], + "type": "text", + "content": "(dip, paintbrush), (paint, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall), (dip, wall)." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 143, + 419, + 227, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 419, + 227, + 426 + ], + "spans": [ + { + "bbox": [ + 143, + 419, + 227, + 426 + ], + "type": "text", + "content": "③ Many Object Scenario (>5 objects)" + } + ] + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 168, + 428, + 468, + 457 + ], + "blocks": [ + { + "bbox": [ + 168, + 428, + 468, + 457 + ], + "lines": [ + { + "bbox": [ + 168, + 428, + 468, + 457 + ], + "spans": [ + { + "bbox": [ + 168, + 428, + 468, + 457 + ], + "type": "image", + "image_path": "49f1a59419bf3967b1417737113621f4b758ae98cd718115618972a64b38d786.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 143, + 460, + 164, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 460, + 164, + 465 + ], + "spans": [ + { + "bbox": [ + 143, + 460, + 164, + 465 + ], + "type": "text", + "content": "GT Past" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 168, + 460, + 397, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 460, + 397, + 466 + ], + "spans": [ + { + "bbox": [ + 168, + 460, + 397, + 466 + ], + "type": "text", + "content": "put, screwdriver), (put, bench), (take, wire), (attach, wire), (adjust, wire), (attach, wire), (hold, wire), (hold, wire)" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 143, + 469, + 165, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 469, + 165, + 475 + ], + "spans": [ + { + "bbox": [ + 143, + 469, + 165, + 475 + ], + "type": "text", + "content": "GT Future" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 168, + 469, + 446, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 469, + 446, + 476 + ], + "spans": [ + { + "bbox": [ + 168, + 469, + 446, + 476 + ], + "type": "text", + "content": "(squeeze, wire), (adjust, wire), (hold, wire), (move, stool), (sit, stool), (take, screwdriver), (serve, mower), (put, screwdriver), (take, nut)," + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 168, + 477, + 459, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 477, + 459, + 487 + ], + "spans": [ + { + "bbox": [ + 168, + 477, + 459, + 487 + ], + "type": "text", + "content": "(hit, mower), (take, screwdriver), (move, screwdriver), (repair, mower), (put, screwdriver), (repair, mower), (repair, mower), (hold, screwdriver), (take, screwdriver), (repair, mower), (move, screwdriver)." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 143, + 491, + 165, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 491, + 165, + 496 + ], + "spans": [ + { + "bbox": [ + 143, + 491, + 165, + 496 + ], + "type": "text", + "content": "SlowFast" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 168, + 491, + 457, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 491, + 457, + 497 + ], + "spans": [ + { + "bbox": [ + 168, + 491, + 457, + 497 + ], + "type": "text", + "content": "(attach, wheel), (turn, pipe), (take, valve), (put, wheel), (put, sticker), (take, drawer), (put, wheel), (repair, torch), (hold, wheel), (repair, plier)," + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 143, + 498, + 159, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 498, + 159, + 503 + ], + "spans": [ + { + "bbox": [ + 143, + 498, + 159, + 503 + ], + "type": "text", + "content": "(1.0)" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 168, + 498, + 458, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 498, + 458, + 506 + ], + "spans": [ + { + "bbox": [ + 168, + 498, + 458, + 506 + ], + "type": "text", + "content": "(put, screwdriver), (attach, wheel), (repair, container), (put, wheel), (put, wheel), (attach, seasoning), (take, wheel), (take, cutter), (repair, screw)" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 143, + 512, + 159, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 512, + 159, + 517 + ], + "spans": [ + { + "bbox": [ + 143, + 512, + 159, + 517 + ], + "type": "text", + "content": "PALM" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 168, + 511, + 455, + 525 + ], + "type": "list", + "angle": 0, + "index": 51, + "blocks": [ + { + "bbox": [ + 168, + 511, + 441, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 511, + 441, + 518 + ], + "spans": [ + { + "bbox": [ + 168, + 511, + 441, + 518 + ], + "type": "text", + "content": "move, screwdriver), (move, screwdriver), (take, wrench), (put, screwdriver), (move, screwdriver), (take, screwdriver), (move, screw" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 168, + 519, + 455, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 519, + 455, + 525 + ], + "spans": [ + { + "bbox": [ + 168, + 519, + 455, + 525 + ], + "type": "text", + "content": "(move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver), (move, screwdriver)" + } + ] + } + ], + "index": 50 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 542, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 630 + ], + "type": "text", + "content": "Fig. 6: Qualitative results of predictions on Ego4D. We show the comparison between the predictions from the Ego4D baseline using SlowFast [18] and the predictions from our model PALM respectively. For each of the three examples shown here, we include the middle frames of the previous eight action segments with the corresponding ground truth past actions GT Past and future actions GT Future alongside the predicted future actions. The green words represent correct predictions while the red words denote incorrect predictions. The edit distance between the predicted actions and ground truth actions is shown in parenthesis." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 130, + 641, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 483, + 665 + ], + "type": "text", + "content": "Evaluation. We conform to the evaluation procedures set forth by the Ego4D [18] and calculate the edit distance (ED) using the Damerau-Levenshtein distance" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "metric [13, 28]. ED is computed separately for verbs, nouns, and the entire action sequences. Given " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "P = 8" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " action segments of the input video, we report the minimum edit distance among " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " predicted sequences, each with a length of " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "Z = 20" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " actions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 483, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 483, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 483, + 295 + ], + "type": "text", + "content": "Results. Tab. 1 shows that PALM improves existing SOTA methods and achieves the best performance on the test set of the Ego4D v1 and v2 benchmark. In general, LLM-based approaches (VideoLLM, AntGPT, and PALM) predict more accurate future actions compared to other approaches based on representation learning. Notably, repeat and retrieve action baselines in Ego4D v2 perform better than Slowfast with respect to noun and action ED. These observations clearly indicate the advantages of leveraging foundation models and the limitations of feature learning methods on the LTA task. Moreover, PALM outperforms other LLM-based methods, which demonstrates the effectiveness of our approach to provide ample visual context to LLMs, equipped with action recognition and image captioning models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 487 + ], + "type": "text", + "content": "In Fig. 6, we perform a qualitative comparison between the Ego4D baseline using SlowFast [18] and PALM. The first example exemplifies a scenario where the person mainly interacts with a single object, in this case, a brick. Both methods can accurately predict the future noun \"brick\". In the second scenario, the person is drawing a painting on a wall, interacting with multiple objects such as a paintbrush, a container, and the wall. PALM predicts the repeated sequence of actions, unlike SlowFast which incorrectly anticipates unrelated actions such as putting a string, touching a nail gun, or throwing a soap, none of which are present in the input video. This also demonstrates the importance of previous actions in the anticipation of future actions. The final scenario portrays the person adjusting a wire in a repair shop, a task involving numerous items. While SlowFast and PALM recognize central objects such as a screwdriver, they both fail to predict a coherent sequence of repair actions. We hypothesize that a language model with expert knowledge can produce more meaningful anticipations, particularly in the context of repairing mechanics, where it is crucial to execute actions in a precise order. More qualitative results are shown in supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 503, + 211, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 211, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 211, + 514 + ], + "type": "text", + "content": "4.3 Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 556 + ], + "type": "text", + "content": "We perform various ablation studies regarding action recognition, example selection methods, LLM models, and prompt contents. We evaluate on the validation set of Ego4D v2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 640 + ], + "type": "text", + "content": "Action Recognition. The action recognition module is an essential part of our framework PALM as we utilize LLMs for predicting future actions and the accuracy of these predictions is highly influenced by the sequence of previous actions. We evaluate various video encoders trained on egocentric videos, including MViT [15], SlowFast [16], StillFast [49], EgoVLP [34], and EgoVLPv2 [46]. It is important to note that we employ a frozen video encoder without any finetuning, as depicted in Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Tab. 2 presents the Top-1 Accuracy " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " and the edit distance of each video encoder. EgoVLP emerges as the top performer, achieving the best accuracy" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 179, + 470, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Ablation on action recognition module. We compare different pretrained video encoders by reporting the Top-1 accuracy (Acc. ↑) and edit distance (ED ↓) for verbs, nouns, and actions. For comparative purposes, we also include the ED when ground truth actions (GT action) are used in the prompts. The best Acc. and ED are highlighted in bold except those from GT action." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 179, + 470, + 270 + ], + "lines": [ + { + "bbox": [ + 138, + 179, + 470, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 179, + 470, + 270 + ], + "type": "table", + "html": "
MethodVerb Acc.Noun Acc.Action Acc.Verb EDNoun EDAction ED
MViT [15]19.872.550.510.72910.85970.9630
SlowFast [16]19.4214.653.120.72520.74720.9278
StillFast [49]19.1219.414.060.73560.73690.9241
EgoVLPv2 [46]33.8440.6316.310.71820.66510.8937
EgoVLP [34]40.3245.5320.630.71110.64650.8819
GT action---0.66630.58870.8319
", + "image_path": "87a3f883b0c447225f0e67207e4fea2101db399cdfb168271c1d15c08c8de82d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 388 + ], + "type": "text", + "content": "and edit distance among all encoders. EgoVLP utilizes a custom contrastive loss to align egocentric video frames with narrations, effectively establishing a connection between visual and textual embeddings. However, EgoVLPv2 utilizes task-specific loss for question-answering tasks built on the EgoVLP setting, which results in slightly less informative video features for action recognition. Unlike EgoVLP and EgoVLPv2, other video encoders are solely trained on video frames, making it challenging to draw robust connections between their video features and textual information (verbs and nouns). Based on these results, we select EgoVLP as our video encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 473 + ], + "type": "text", + "content": "It is expected that better accuracy in action recognition leads to a reduced edit distance, as accurate estimation of past actions is crucial for predicting future actions. Notably, the use of ground truth (GT) actions significantly outperforms all video encoders, which highlights the enhancement of action recognition as the foremost challenge within our framework. We believe that fine-tuning the video encoder on specific egocentric datasets would likely enhance recognition accuracy." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 220, + 525, + 391, + 583 + ], + "blocks": [ + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "type": "text", + "content": "Table 3: Ablation on example selection strategies exhibits that MMR outperforms other selection methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 220, + 525, + 391, + 583 + ], + "lines": [ + { + "bbox": [ + 220, + 525, + 391, + 583 + ], + "spans": [ + { + "bbox": [ + 220, + 525, + 391, + 583 + ], + "type": "table", + "html": "
MethodVerb EDNoun EDAction ED
Random0.73620.67470.9002
Similarity0.72040.66610.8923
MMR0.71790.66470.8908
", + "image_path": "a8f6eae55a9cf4d763a210428e7cb29fe20b6373fa8275da64a3e4419ddcba21.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Example Selection. We compare different example selection strategies for in-context learning (ICL), including random, similarity, and maximal marginal relevance (MMR) [59]. The random method uniformly selects examples from the training set, whereas the similarity method selects the most semantically similar example to the query. MMR considers both the similarity and diversity of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "examples, striving for a balance between them. Tab. 3 demonstrates that MMR outperforms the other selection methods." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 198, + 196, + 413, + 275 + ], + "blocks": [ + { + "bbox": [ + 130, + 163, + 479, + 185 + ], + "lines": [ + { + "bbox": [ + 130, + 163, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 479, + 185 + ], + "type": "text", + "content": "Table 4: Ablation on LLMs shows that larger LLMs perform better on the LTA task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 198, + 196, + 413, + 275 + ], + "lines": [ + { + "bbox": [ + 198, + 196, + 413, + 275 + ], + "spans": [ + { + "bbox": [ + 198, + 196, + 413, + 275 + ], + "type": "table", + "html": "
MethodVerb EDNoun EDAction ED
GPT-Neo (125M)0.76570.68670.9119
GPT-Neo (1.3B)0.73140.67590.8970
GPT-Neo (2.7B)0.71780.67470.8943
Llama 2 (7B)0.71790.66470.8908
Llama 2 (13B)0.69580.66340.8867
", + "image_path": "a1c2c07655192ee2a55a80e05b622c9a32718ef4dcb0ffd973c89ce4d79a18fd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 384 + ], + "type": "text", + "content": "LLM Ablation. We ablate the impact of LLMs by comparing different open-source models, primarily GPT-Neo [4] and Llama 2 [54] families. Tab. 4 indicates that larger language models are more capable of predicting future actions with their improved capacity for generalization across novel scenarios. This observation aligns with the performance of LLMs on general NLP benchmarks. Taking into account the computational resources at our disposal, we select Llama 2 with 7B parameters." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 207, + 439, + 403, + 497 + ], + "blocks": [ + { + "bbox": [ + 130, + 407, + 479, + 430 + ], + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 430 + ], + "type": "text", + "content": "Table 5: Ablation on the content of the prompt. We use abbreviations for each content in the input prompt: " + }, + { + "bbox": [ + 130, + 407, + 479, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 407, + 479, + 430 + ], + "type": "text", + "content": " - action sequence, " + }, + { + "bbox": [ + 130, + 407, + 479, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 407, + 479, + 430 + ], + "type": "text", + "content": " - captions" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 207, + 439, + 403, + 497 + ], + "lines": [ + { + "bbox": [ + 207, + 439, + 403, + 497 + ], + "spans": [ + { + "bbox": [ + 207, + 439, + 403, + 497 + ], + "type": "table", + "html": "
Prompt ContentVerb EDNoun EDAction ED
A0.72030.66880.8947
C0.70880.69010.9064
A+C0.71790.66470.8908
", + "image_path": "8d150a7ea9f45aedd850192c1737fabada4f7ca974ed3f3d8141345bf3da53d5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "text", + "content": "Prompt Contents. We examine different content included in the input prompt to validate the contribution of each module to final performance. PALM utilizes recognized past actions and generated image captions to compose prompts. The action sequence describes the temporal flow of activity in (verb, noun) format, while captions narrate visual information relevant to human. In Tab. 5, the action sequence with captions " + }, + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{A} + \\mathcal{C})" + }, + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "text", + "content": " achieves the lowest action ED, which thereby is used in our framework. Notably, caption-only prompt " + }, + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{C})" + }, + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "text", + "content": " reaches the lowest verb ED, showing that narrations are capable of capturing dynamic motions of human activity, while action-only prompt " + }, + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{A})" + }, + { + "bbox": [ + 130, + 521, + 480, + 666 + ], + "type": "text", + "content": " effectively detects key objects interacted with human supported by relatively lower noun ED. This indicates that action sequence and narrations provide a beneficial complement, thereby supporting our decision to incorporate both in the prompts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 146, + 434, + 265 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 6: Comparison to the SOTA on EK-55 and EGTEA. Performance is evaluated using the mAP metric " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 146, + 434, + 265 + ], + "lines": [ + { + "bbox": [ + 176, + 146, + 434, + 265 + ], + "spans": [ + { + "bbox": [ + 176, + 146, + 434, + 265 + ], + "type": "table", + "html": "
MethodEK-55EGTEA
ALLFREQRAREALLFREQRARE
I3D [8]32.753.323.072.179.353.3
ActionVLAD [17]29.853.518.673.379.058.6
Timeception [22]35.655.926.174.179.759.7
VideoGraph [23]22.549.414.067.777.147.2
EGO-TOPO [38]38.056.929.273.580.754.7
Anticipatr [41]39.158.129.176.883.355.1
AntGPT [63]40.158.831.980.284.872.9
PALM (ours)40.459.330.380.785.073.5
", + "image_path": "a8b3765327af2b14f1ee32366ec58a9b3915e488b302e2ac5795d293b09edc21.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 285, + 332, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 332, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 332, + 297 + ], + "type": "text", + "content": "4.4 Evaluation on EK-55 and EGTEA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "text", + "content": "Evaluation. For EK-55 [12] and EGTEA [33], we adhere to the experimental setup established by EGO-TOPO [38] and employ mean average precision (mAP) for multi-label classification as the evaluation metric. The initial " + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "inline_equation", + "content": "R = [25\\%, 50\\%, 75\\%]" + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "text", + "content": ") of each video serves as input, with the objective of predicting the set of actions occurring in the remaining " + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "inline_equation", + "content": "(100 - R)\\%" + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "text", + "content": " of the video. Notably, actions here are confined to verb predictions only without noun, following previous conventions. We calculate the mAP on the validation set averaged over all " + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 315, + 481, + 422 + ], + "type": "text", + "content": ", examining various performance aspects, such as all target actions (ALL), frequently occurring actions (FREQ), and rarely occurring actions (RARE)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 532 + ], + "type": "text", + "content": "Results. As shown in Tab. 6, our method achieves SOTA results in the ALL and FREQ section while exhibiting comparable performance in the RARE section. The results demonstrate the effectiveness of PALM by capturing additional visual information with image captioning models. This aligns with the ablation study regarding prompt contents in Sec. 4.3 where captions contribute better on verb predictions. We observe that verb in RARE section are seldom represented in the captions, which results in relatively lower mAP in RARE section. Thus, we believe fine-tuning image captioning model or employing better video captioning model will further improve the performance accordingly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 546, + 220, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 220, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 220, + 559 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "We present PALM, a framework that leverages vision-language and large language models for long-term action anticipation. PALM achieves the best performance on the Ego4D benchmark, demonstrating the effectiveness of performing the LTA task in the language domain and the benefits of using foundation models. Our model also shows competitive results on two other benchmarks, EK-55 and EGTEA. As a result, we are optimistic that our work will spur further research in using commonsense knowledge extracted from LLMs for video understanding." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al.: Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 (2022) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "text", + "content": "2. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023) 4, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 237 + ], + "type": "text", + "content": "3. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. If you use this software, please cite it using these metadata 58 (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 238, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 481, + 293 + ], + "type": "text", + "content": "4. Black, S., Gao, L., Wang, P., Leahy, C., Biderman, S.: Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. https://doi.org/10.5281/zenodo.5297715 (2021). https://doi.org/10.5281/ zenodo.5297715, https://doi.org/10.5281/zenodo.5297715, If you use this software, please cite it using these metadata. 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 326 + ], + "type": "text", + "content": "5. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 369 + ], + "type": "text", + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 481, + 413 + ], + "type": "text", + "content": "7. Carbonell, J., Goldstein, J.: The use of mmr, diversity-based reranking for reordering documents and producing summaries. In: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval. pp. 335-336 (1998) 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "text", + "content": "8. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 479 + ], + "type": "text", + "content": "9. Chen, G., Zheng, Y.D., Wang, J., Xu, J., Huang, Y., Pan, J., Wang, Y., Wang, Y., Qiao, Y., Lu, T., et al.: Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292 (2023) 4, 9, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "text", + "content": "0. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "text", + "content": "1. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 588 + ], + "type": "text", + "content": "2. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(11), 4125-4141 (2020) 2, 3, 8, 9, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 589, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 611 + ], + "type": "text", + "content": "3. Damerau, F.J.: A technique for computer detection and correction of spelling errors. Communications of the ACM 7(3), 171-176 (1964) 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "4. Das, S., Ryoo, M.S.: Video+ clip baseline for ego4d long-term action anticipation. arXiv preprint arXiv:2207.00579 (2022) 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824-6835 (2021) 2, 4, 11, 12" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) 4, 11, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "17. Girdhar, R., Ramanan, D., Gupta, A., Sivic, J., Russell, B.: Actionvlad: Learning spatio-temporal aggregation for action classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 971-980 (2017) 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 237 + ], + "type": "text", + "content": "18. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022) 2, 3, 4, 6, 8, 9, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 481, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 273 + ], + "type": "text", + "content": "19. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "text", + "content": "20. Huang, D., Hilliges, O., Van Gool, L., Wang, X.: Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545 (2023) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "text", + "content": "21. Huang, W., Abbeel, P., Pathak, D., Mordatch, I.: Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In: International Conference on Machine Learning. pp. 9118-9147. PMLR (2022) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "22. Hussein, N., Gavves, E., Smeulders, A.W.: Timeception for complex action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 254-263 (2019) 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "text", + "content": "23. Hussein, N., Gavves, E., Smeulders, A.W.: Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143 (2019) 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "24. Ichter, B., Brohan, A., Chebotar, Y., Finn, C., Hausman, K., Herzog, A., Ho, D., Ibarz, J., Irpan, A., Jang, E., et al.: Do as i can, not as i say: Grounding language in robotic affordances. In: 6th Annual Conference on Robot Learning (2022) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 485 + ], + "type": "text", + "content": "25. Keunwoo Peter, Y.: Videooblip is a large vision-language model based on blip-2 that can generate texts conditioned on videos. https://github.com/yukw777/VideoBLIP (2021). https://doi.org/https://orcid.org/0000-0002-4387-2410, https://github.com/yukw777/VideoBLIP, If you use this software, please cite it using these metadata. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 530 + ], + "type": "text", + "content": "26. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "text", + "content": "27. Leo, M., Medioni, G., Trivedi, M., Kanade, T., Farinella, G.M.: Computer vision for assistive technologies. Computer Vision and Image Understanding 154, 1-15 (2017) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 481, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 596 + ], + "type": "text", + "content": "28. Levenshtein, V.I., et al.: Binary codes capable of correcting deletions, insertions, and reversals. In: Soviet physics doklady. vol. 10, pp. 707-710. Soviet Union (1966) 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 598, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 631 + ], + "type": "text", + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 195 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 195 + ], + "type": "text", + "content": "32. Li, S., Puig, X., Paxton, C., Du, Y., Wang, C., Fan, L., Chen, T., Huang, D.A., Akyurek, E., Anandkumar, A., et al.: Pre-trained language models for interactive decision-making. Advances in Neural Information Processing Systems 35, 31199-31212 (2022) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 482, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 482, + 228 + ], + "type": "text", + "content": "33. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018) 3, 8, 9, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 482, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 482, + 262 + ], + "type": "text", + "content": "34. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022) 4, 7, 11, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 482, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 482, + 297 + ], + "type": "text", + "content": "35. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 482, + 329 + ], + "type": "text", + "content": "36. Mascaró, E.V., Ahn, H., Lee, D.: Intention-conditioned long-term human egocentric action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6048-6057 (2023) 4, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 331, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 482, + 374 + ], + "type": "text", + "content": "37. Meditskos, G., Plans, P.M., Stavropoulos, T.G., Benois-Pineau, J., Buso, V., Kompatsiaris, I.: Multi-modal activity recognition from egocentric vision, semantic enrichment and lifelogging applications for the care of dementia. Journal of Visual Communication and Image Representation 51, 169-190 (2018) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 376, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 482, + 407 + ], + "type": "text", + "content": "38. Nagarajan, T., Li, Y., Feichtenhofer, C., Grauman, K.: Ego-topo: Environment affordances from egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 163-172 (2020) 9, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 409, + 482, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 482, + 441 + ], + "type": "text", + "content": "39. Nakazawa, A., Honda, M.: First-person camera system to evaluate tender dementia-care skill. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 482, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 482, + 475 + ], + "type": "text", + "content": "40. Naveed, H., Khan, A.U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Barnes, N., Mian, A.: A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 (2023) 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 476, + 482, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 482, + 508 + ], + "type": "text", + "content": "41. Nawhal, M., Jyothi, A.A., Mori, G.: Rethinking learning approaches for long-term action anticipation. In: European Conference on Computer Vision. pp. 558-576. Springer (2022) 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 510, + 482, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 482, + 542 + ], + "type": "text", + "content": "42. Neumann, L., Zisserman, A., Vedaldi, A.: Future event prediction: If and when. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 482, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 482, + 576 + ], + "type": "text", + "content": "43. OhnBar, E., Kitani, K., Asakawa, C.: Personalized dynamics models for adaptive assistive navigation systems. In: Conference on Robot Learning. pp. 16-39. PMLR (2018) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 577, + 482, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 482, + 631 + ], + "type": "text", + "content": "44. Pasca, R.G., Gavryushin, A., Hamza, M., Kuo, Y.L., Mo, K., Van Gool, L., Hilliges, O., Wang, X.: Summarize the past to predict the future: Natural language descriptions of context boost multimodal object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18286-18296 (2024) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "text", + "content": "45. Patel, D., Eghbalzadeh, H., Kamra, N., Iuzzolino, M.L., Jain, U., Desai, R.: Pre-trained language models as visual planners for human assistance. arXiv preprint arXiv:2304.09179 (2023) 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "46. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023) 2, 4, 11, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 205 + ], + "type": "text", + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "48. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners (2019) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 271 + ], + "type": "text", + "content": "49. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023) 2, 4, 11, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "text", + "content": "50. Rodin, I., Furnari, A., Mavroeidis, D., Farinella, G.M.: Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding 211, 103252 (2021) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 350 + ], + "type": "text", + "content": "51. Ryoo, M.S., Fuchs, T.J., Xia, L., Aggarwal, J.K., Matthies, L.: Robot-centric activity prediction from first-person videos: What will they do to me? In: Proceedings of the tenth annual ACM/IEEE international conference on human-robot interaction. pp. 295–302 (2015) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 352, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 384 + ], + "type": "text", + "content": "52. Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems 33, 16857-16867 (2020) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "53. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "text", + "content": "54. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 3, 7, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "text", + "content": "55. Wang, S., Zhao, Q., Do, M.Q., Agarwal, N., Lee, K., Sun, C.: Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 476, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 481, + 508 + ], + "type": "text", + "content": "56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 542 + ], + "type": "text", + "content": "57. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. In: International Conference on Learning Representations (2021) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 586 + ], + "type": "text", + "content": "58. Yao, Y., Xu, M., Choi, C., Crandall, D.J., Atkins, E.M., Dariush, B.: Egocentric vision-based future vehicle localization for intelligent driving assistance systems. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 9711-9717. IEEE (2019) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 620 + ], + "type": "text", + "content": "59. Ye, X., Iyer, S., Celikyilmaz, A., Stoyanov, V., Durrett, G., Pasunuru, R.: Complementary explanations for effective in-context learning. arXiv preprint arXiv:2211.13892 (2022) 2, 8, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "60. Zeng, A., Attarian, M., Choromanski, K.M., Wong, A., Welker, S., Tombari, F., Purohit, A., Ryoo, M.S., Sindhwani, V., Lee, J., et al.: Socratic models: Composing zero-shot multimodal reasoning with language. In: The Eleventh International Conference on Learning Representations (2022) 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 281 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "61. Zhan, K., Faux, S., Ramos, F.: Multi-scale conditional random fields for first-person activity recognition. In: 2014 IEEE international conference on pervasive computing and communications (PerCom). pp. 51-59. IEEE (2014) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "63. Zhao, Q., Zhang, C., Wang, S., Fu, C., Agarwal, N., Lee, K., Sun, C.: Antgpt: Can large language models help long-term action anticipation from videos? arXiv preprint arXiv:2307.16368 (2023) 4, 9, 10, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "64. Zhou, C., Li, Q., Li, C., Yu, J., Liu, Y., Wang, G., Zhang, K., Ji, C., Yan, Q., He, L., et al.: A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419 (2023) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "text", + "content": "65. Zhou, L., Palangi, H., Zhang, L., Hu, H., Corso, J., Gao, J.: Unified vision-language pre-training for image captioning and vqa. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13041-13049 (2020) 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 447, + 102 + ], + "type": "text", + "content": "PALM: Predicting Actions through Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_content_list.json b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aaf8b869ed14a89c33cb04906faeb881ca069be6 --- /dev/null +++ b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_content_list.json @@ -0,0 +1,1695 @@ +[ + { + "type": "text", + "text": "PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration", + "text_level": 1, + "bbox": [ + 222, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang", + "bbox": [ + 230, + 212, + 771, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, China", + "bbox": [ + 220, + 255, + 781, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Learning rotation-invariant distinctive features is a fundamental requirement for point cloud registration. Existing methods often use rotation-sensitive networks to extract features, while employing rotation augmentation to learn an approximate invariant mapping rudely. This makes networks fragile to rotations, overweight, and hinders the distinctiveness of features. To tackle these problems, we propose a novel position-aware rotation-equivariant network, for efficient, light-weighted, and robust registration. The network can provide a strong model inductive bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned limitations. To further improve the distinctiveness of descriptors, we propose a position-aware convolution, which can better learn spatial information of local structures. Moreover, we also propose a feature-based hypothesis proposer. It leverages rotation-equivariant features that encode fine-grained structure orientations to generate reliable model hypotheses. Each correspondence can generate a hypothesis, thus it is more efficient than classic estimators that require multiple reliable correspondences. Accordingly, a contrastive rotation loss is presented to enhance the robustness of rotation-equivariant features against data degradation. Extensive experiments on indoor and outdoor datasets demonstrate that our method significantly outperforms the SOTA methods in terms of registration recall while being lightweight and keeping a fast speed. Moreover, experiments on rotated datasets demonstrate its robustness against rotation variations. Code is available at https://github.com/yaorz97/PARENet.", + "bbox": [ + 261, + 333, + 738, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Point cloud registration $\\cdot$ Rotation invariance/equivalence $\\cdot$ Robust transformation estimation", + "bbox": [ + 261, + 679, + 740, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud registration, estimating a rigid transformation to align two point clouds that are partially overlapped, is fundamental research in 3D computer vision. It is widely applied in autonomous driving [20,21], robot localization [23], and 3D reconstruction [38]. Feature-based point cloud registration framework is widely studied, which mainly involves descriptor extraction [4, 15, 25, 34] and", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "robust transformation estimation [3,6,32]. One inherent contradiction in point cloud registration is the pose variation of point clouds and the desired invariance of their descriptors. To extract such rotation-invariant descriptors, plenty of approaches have been proposed that can be roughly categorized into patch-wise and scene-wise feature extractors.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/980f9b4fc9a18938bb508e16c4db13c257158c71adcea63ce949e1c50cf98bc1.jpg", + "image_caption": [ + "Fig. 1: Experimental results on 3DMatch and 3DLoMatch. Our method significantly outperforms state-of-the-art methods w.r.t. registration recall (RR) while maintaining fast speed and lightweight." + ], + "image_footnote": [], + "bbox": [ + 218, + 246, + 498, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1891bf84bac650bb834c2d014ffc40dd64adc48fdd6e9f5c59347f3fcf1ca2dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 246, + 784, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Patch-wise feature extractors [2, 9, 10] learn geometric features from local patches of point clouds and typically employ data preprocessing techniques to achieve rotation-invariance. Two common approaches include: 1) extracting rotation invariant statistics as initial representation from local patches, such as distance, angle, and point pair features (PPF [9, 10]); and 2) estimating a local reference frame (LRF [2]) to normalize poses of local patches. The former approach suffers from information loss, while the latter approach is susceptible to data occlusion and noise, lacking robustness. Moreover, these extractors are generally computationally expensive because all patches of a point cloud are processed individually, without sharing intermediate representations [7].", + "bbox": [ + 212, + 401, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene-wise feature extractors [4,7,15,25,34] learn dense feature descriptors for every point in the point cloud, and most of them employ extensive rotation augmentation in the training process to rudely enforce inherently rotation-sensitive networks (e.g., KPCov [26]) to learn rotation-invariant mappings. These methods have three shortcomings. Firstly, models are fragile to unseen rotations because it is difficult to cover the SO(3) space in the training process [35]. Secondly, learning rotation-invariant mappings requires more network parameters. Thirdly, these networks are limited to generating discriminative feature descriptors as they cannot focus solely on learning structural distinctiveness.", + "bbox": [ + 212, + 551, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel registration network, PARE-Net, whose core is to fully exploit the advantages of rotation-equivariant networks in both feature extraction and transformation estimation, thereby achieving a lightweight, efficient, and robust registration method, as shown in Fig. 1.", + "bbox": [ + 212, + 688, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First of all, we propose a Position-Aware Rotation-Equivariant Convolution (PARE-Conv) that enhances the representational power of networks for better distinctive feature learning. Vector Neuron (VN) [8] builds a general rotation-equivariant network framework. However, its representative power is limited as it ignores the importance of spatial localization property of local points and utilizes multi-layer perceptrons (MLPs) to crudely learn spatial encodings. In", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stead, PARE-Conv is proposed to explicitly leverage the spatial information of point clouds while guaranteeing the rotation equivariant property. Specifically, we first define a set of shadow kernel points equipped with learnable weights. Then, we introduce a correlation network that predicts rotation-invariant correlations between points and shadow kernels according to rotation-equivariant spatial features of points. Therefore, the geometric structures of point clouds are well captured. Finally, the correlation scores are subsequently leveraged to dynamically assemble kernel weights for rotation-equivariant convolution. As a result, PARE-Conv inherently provides a strong model bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned problems raised by rotation-sensitive networks. Moreover, it can learn the spatial structure of point clouds better, facilitating to extraction of more distinctive descriptors.", + "bbox": [ + 212, + 146, + 797, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, we propose a feature-based hypothesis proposer that can efficiently generate reliable solution hypotheses. Once obtaining correspondences by feature matching, rotation-equivariant features encoding fine-grained structure information are used to explicitly generate multiple hypotheses. This way can alleviate the orientation ambiguity stemming from optimizing distance errors of correspondences. Moreover, to make the equivariant features more robust against data degradation, we present a contrastive rotation loss inspired by metric learning [7, 16], facilitating more robust pose estimation. As shown in Fig. 2, our method can achieve significantly higher registration", + "bbox": [ + 212, + 329, + 599, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "recall compared to RANSAC [11] and LGR (presented in GeoTrans [25]) when varying the number of hypotheses, demonstrating our method is more robust than classical estimators. Because they require multiple correspondences and the probability of sampling all correct correspondences decreases rapidly as the number of used correspondences increases. Overall, our main contributions are:", + "bbox": [ + 212, + 523, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dfa6aaa7c60440553160fd36a304b8e1660a41206ad6af8814a41a69eee1d1f0.jpg", + "image_caption": [ + "Fig.2: Comparison of our feature-based hypothesis proposer, RANSAC and LGR on 3DLoMatch." + ], + "image_footnote": [], + "bbox": [ + 612, + 358, + 785, + 426 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A lightweight, efficient, and robust point cloud registration method that is achieved by fully exploiting the advantages of rotation-equivariant networks in feature extraction and pose estimation.", + "- A position-aware rotation-equivariant convolution that can better leverage spatial information, enabling the extraction of discriminative descriptors.", + "- A feature-based hypothesis proposer that can efficiently generate multiple reliable hypotheses and a contrastive rotation loss makes the rotation-equivariant features more robust to data degradation." + ], + "bbox": [ + 223, + 612, + 784, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 760, + 387, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patch-wise feature extractors. In the early stage, most learning-based feature descriptors [2, 9, 10, 13, 17, 37] were patch-wise extractors, primarily due to limitations in point cloud convolution techniques, which could not handle", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "large-scale point clouds effectively. During this period, the works drew inspiration from traditional feature descriptors [14] and focused on designing features with rotation-invariance. PPFNet [10] and PPFoldNet [9] extracted rotation-invariant PPFs as initial features and further learned high-dimensional features. PerfectMatch [13] and SpinNet [2] utilized LRF to normalize the pose of local patches and then employed networks to learn the density/structural features of patches. These methods are primarily limited by their computational efficiency.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene-wise feature extractors [4, 7, 15, 19, 25, 33-36] have become the mainstream research in recent years, due to their high efficiency. These methods mainly achieved rotation invariance through data augmentation and aim to tackle the low-overlapped problem. Predator [15] designed an overlap-attention module to detect overlap regions between two point clouds. CoFiNet [34] and GeoTrans [25] presented a coarse-to-fine matching scheme, and GeoTrans introduced a geometric transformer module that significantly increased the matching accuracy by injecting rotation-invariant position information. PEAL [36] presented an explicit one-way attention module that iteratively leveraged overlap prior to relieving feature ambiguity while bringing expensive computation costs. Moreover, some methods with rotation invariant/equivariant designs were proposed. YOHO [27] and RoReg [28] exploited the group convolution technique to extract rotation-invariant/equivariant group features, which are extremely computation expensive. Because they need to extract point features under 60 different poses and perform convolutions on massive point features. RoITR [35] used PPF to design an invariant convolution and presented a rotation invariant transformer to extract distinctive features. BUFFER [1] combined scene-wise and patch-wise feature extractors to achieve faster alignment with great generalization ability. It first leveraged a rotation-equivariant network built on vanilla VN to extract invariant and equivariant features for keypoints and orientations prediction. Then, it also used SpinNet to extract patch-wise features due to the limited representative ability of vanilla VN.", + "bbox": [ + 212, + 252, + 787, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robust transformation estimators. Since the estimated correspondences are usually contaminated by outliers, a robust transformation estimator is necessary. RANSAC [11] is the most widely used method due to its simplicity, but it typically exhibits lower accuracy and requires plenty of iterations for convergence. Recently, some learning-based methods [3,6,18,32] are proposed, whose core idea is to design networks to learn spatial consistency between correspondences, enabling efficient inlier identification and reliable hypothesis generation. We compare our approach with these methods in Appendix F.1.", + "bbox": [ + 212, + 585, + 787, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 729, + 330, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given two partially overlapped point clouds $\\mathcal{P} = \\{\\mathbf{p}_i\\in \\mathbb{R}^3 |i = 1,2,\\dots ,N\\}$ and $\\mathcal{Q} = \\{\\mathbf{q}_j\\in \\mathbb{R}^3 |j = 1,2,\\dots ,M\\}$ , we aim to estimate a transformation $\\mathbf{T} = (\\mathbf{R}\\in SO(3),\\mathbf{t}\\in \\mathbb{R}^3)$ that aligns the source point cloud to the target one. To address this problem, we propose a registration framework based on rotation-equivariant network $f(\\mathbf{X})$ , whose output is equivariant to the SO(3) action $\\mathbf{R}$", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9666065ae66c08d2e36b35726224420c6174f7e3a96a14e18983c6af1aa633f8.jpg", + "image_caption": [ + "Fig. 3: The framework of our method. Given point clouds $\\mathcal{P}$ and $\\mathcal{Q}$ , a hierarchical backbone based on PARE-Conv is used to extract patch-level and point-level features. Then, we follow a coarse-to-fine scheme to obtain point correspondences using rotation-invariant features. Finally, the feature-based hypothesis proposer leverages the matched rotation-equivariant features $\\tilde{\\mathbf{F}}_{x_{jn}}^{\\mathcal{P}}$ and $\\tilde{\\mathbf{F}}_{y_{jn}}^{\\mathcal{Q}}$ , which encode fine-grained structure orientations, to generate multiple reliable hypotheses. The best hypothesis is selected and refined as the final solution $(\\mathbf{R}^{*},\\mathbf{t}^{*})$ ." + ], + "image_footnote": [], + "bbox": [ + 232, + 142, + 772, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "applied on the input point cloud $\\mathbf{X}$ , i.e., $f(\\mathbf{R} \\circ \\mathbf{X}) = \\mathbf{R} \\circ f(\\mathbf{X})$ . As we can see, rotation-equivariant features $f(\\mathbf{R} \\circ \\mathbf{X})$ couple the structure information $f(\\mathbf{X})$ and orientation information $\\mathbf{R}$ of point clouds, which can be further decoupled for feature matching and transformation estimation, respectively.", + "bbox": [ + 212, + 503, + 784, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The framework of PARE-Net is depicted in Fig. 3. We build a hierarchical backbone based on PARE-Conv to extract both patch-level and point-level rotation invariant/equivariant features. Correspondences are estimated in a coarse-to-fine fashion by using rotation-invariant features. After that, the feature-based hypothesis proposer employs rotation-equivariant features to generate multiple reliable hypotheses, and the best one is selected as the final output $(\\mathbf{R}^{*},\\mathbf{t}^{*})$ .", + "bbox": [ + 212, + 565, + 787, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Position-Aware Rotation-Equivariant Network", + "text_level": 1, + "bbox": [ + 214, + 678, + 643, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PARE-Conv. Recent point cloud registration approaches [15,25,34,36] mainly exploited KPCov [26] to build backbones. However, KPCov is inherently rotation-sensitive and needs data augmentation to learn an approximate invariant mapping. As a result, models trained on limited samples of continuous SO(3) space are fragile to rotation variations. Moreover, enforcing the network to learn a rotation-invariant mapping inevitably requires more model parameters and also hinders the distinctiveness of features. Inspired by dynamic and conditional convolutions [30,31], we propose position-aware rotation-equivariant convolution to learn distinctive rotation-invariant/equivariant features, as shown in Fig. 3.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a point cloud $\\mathcal{P} = \\{\\mathbf{p}_i\\in \\mathbb{R}^3|i = 1,2,\\dots ,N\\}$ and attached feature map $\\mathbf{F} = \\{\\mathbf{F}_i\\in \\mathbb{R}^{C\\times 3}\\}$ , the general convolution of $\\mathbf{F}$ by a kernel $g$ at a point $\\mathbf{p}_i$ is defined as:", + "bbox": [ + 212, + 145, + 782, + 189 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {F} * g\\right) \\left(\\mathbf {p} _ {i}\\right) = \\sum_ {\\mathbf {p} _ {j} \\in \\mathcal {N} _ {i}} g \\left(\\mathbf {p} _ {j} - \\mathbf {p} _ {i}\\right) \\mathbf {F} _ {j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 191, + 785, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_i$ is the K-nearest neighbor set of the point $\\mathbf{p}_i$ . The core problem is the definition of kernel function $g$ . In vanilla VN [8], it omits the position information and simply uses a linear layer, $\\mathbf{W}\\mathbf{F}_i$ , followed by a VN non-linear layer and a VN pooling layer to aggregate local information, where $\\mathbf{W} \\in \\mathbb{R}^{C' \\times C}$ is a learnable weight matrix. Instead, we define a set of shadow kernel points, equipped with learnable weight matrices $\\{\\mathbf{W}_k | k = 1, 2, \\dots, K\\}$ . In image convolution, each pixel corresponds to a convolutional kernel, and the fixed spatial arrangement of the kernels enables the learning of the distribution information of image pixels. Due to the irregularity of point clouds, establishing such point-to-kernel correspondences is intractable. We follow KPConv [26] to establish soft assignments between data points and kernel points and the kernel function $g$ is defined as:", + "bbox": [ + 212, + 229, + 787, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(\\mathbf {p} _ {i j}\\right) = \\sum_ {k = 1} ^ {K} \\operatorname {s o f t m a x} (\\gamma \\left(\\mathbf {p} _ {i j}, k\\right)) \\mathbf {W} _ {k}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 410, + 785, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{ij} = \\mathbf{p}_j - \\mathbf{p}_i$ is the relative coordinate of $\\mathbf{p}_j$ and $\\gamma (\\cdot)$ is a correlation function. In KPConv [26], it uses a linear correlation $\\gamma (\\mathbf{p}_{ij},k) = \\max (0,1-\\|\\mathbf{p}_{ij} - \\tilde{\\mathbf{x}}_k\\| / \\sigma)$ , where $\\tilde{\\mathbf{x}}_k$ is the 3D coordinate of the $k$ -th kernel point and $\\sigma$ is a distance influence factor. However, kernel-to-point distances are sensitive to rotations thus cannot guarantee a rotation-equivariant convolution. Therefore, the key challenge lies in ensuring the rotation equivariance of the convolution while well capturing the spatial information of patches. Inspired by PACConv [30], we use a mini-network that consumes rotation-equivariant spatial features and predicts rotation-invariant correlation scores between data points and kernel points. Note that predicting rotation-invariant scores is necessary to guarantee the rotation-equivariance of the convolution (see Appendix A for the proof). As shown in Fig. 3, we implement this function with 1) a VN-MLP block, learning geometric features $\\mathbf{X}_j \\in \\mathbb{R}^{d \\times 3}$ from rotation-equivariant spatial statistics, such as $\\mathbf{p}_{ij}$ , $\\frac{1}{|\\mathcal{N}_i|} \\sum_j \\mathbf{p}_{ij}$ and their cross product; 2) a L2-Norm layer, computing the magnitude of the vector neurons $\\tilde{\\mathbf{X}}_j = \\| \\mathbf{X}_j\\|_2 \\in \\mathbb{R}^d$ to transform the feature into a rotation-invariant form; and 3) a MLP block, mapping this feature vector into K dimension to obtain the correlation vector, $\\gamma_j = \\mathrm{MLP}(\\tilde{\\mathbf{X}}_j) \\in \\mathbb{R}^K$ . Finally, a softmax operation is leveraged to output correlation scores. As a result, our convolution is formulated as:", + "bbox": [ + 212, + 462, + 787, + 753 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {F} * g\\right) \\left(\\mathbf {p} _ {i}\\right) = \\sum_ {\\mathbf {p} _ {j} \\in \\mathcal {N} _ {i}} \\sum_ {k} \\gamma_ {j k} \\mathbf {W} _ {k} \\mathbf {F} _ {j}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 765, + 785, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to KPConv using a linear correlation, our correlation network can learn more complex relationships. Moreover, PARE-Conv predicts correlation", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scores depending on local spatial property, thus can leverage the spatial information better and can provide stronger expressive power to extract more discriminative descriptors. Moreover, it is compatible with DGCNN [29], allowing replacing the node feature $\\mathbf{F}_j$ with the edge feature $[\\mathbf{F}_j - \\mathbf{F}_i,\\mathbf{F}_j]$ . This can further enhance the expressive power, as discussed in the ablation study 4.3.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fully Convolutional Network. As shown in Fig. 3, we develop a hierarchical convolution network based on PARE-Conv. The original point clouds $\\mathcal{P}$ and $\\mathcal{Q}$ are downsampled three times to yield sparse superpoints $\\hat{\\mathcal{P}}$ and $\\hat{\\mathcal{Q}}$ , whose resolution is $\\frac{1}{2^3}$ of the original point clouds. Consequently, three convolution blocks are leveraged to extract multi-level features. The bottleneck rotation-equivariant features are denoted as $\\hat{\\mathbf{F}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\hat{\\mathcal{P}}| \\times \\hat{d} \\times 3}$ and $\\hat{\\mathbf{F}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\hat{\\mathcal{Q}}| \\times \\hat{d} \\times 3}$ , and a VN-invariant layer [8] is used to obtain the rotation-invariant features $\\hat{\\mathbf{X}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\hat{\\mathcal{P}}| \\times 3\\hat{d}}$ and $\\hat{\\mathbf{X}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\hat{\\mathcal{Q}}| \\times 3\\hat{d}}$ . For the decoder, two nearest upsampling blocks are used to get the rotation-equivariant features $\\tilde{\\mathbf{F}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{P}}| \\times \\tilde{d} \\times 3}$ and $\\tilde{\\mathbf{F}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{Q}}| \\times \\tilde{d} \\times 3}$ of the first-level downsampled points $\\tilde{\\mathcal{P}}$ and $\\tilde{\\mathcal{Q}}$ . Another VN-invariant layer is leveraged to get their rotation-invariant features $\\tilde{\\mathbf{X}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{P}}| \\times 3\\tilde{d}}$ and $\\tilde{\\mathbf{X}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{Q}}| \\times 3\\tilde{d}}$ . For more details on the convolutional block, please refer to Appendix B.", + "bbox": [ + 212, + 222, + 784, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following GeoTrans [25], we exploit the point-to-node grouping strategy to assign the points $\\hat{\\mathcal{P}}$ to the superpoints $\\hat{\\mathcal{P}}$ . The subset of $\\hat{\\mathcal{P}}$ belongs to the superpoint $\\hat{\\mathcal{P}}_i$ is denoted as $\\mathcal{G}_i^{\\mathcal{P}}$ and the associated rotation-equivariant and invariant feature matrices of subset points are denoted as $\\tilde{\\mathbf{F}}_i^{\\mathcal{P}}$ and $\\tilde{\\mathbf{X}}_i^{\\mathcal{P}}$ . For point cloud $\\tilde{\\mathcal{Q}}$ , the points are grouped as $\\{\\mathcal{G}_i^{\\mathcal{Q}}\\}$ , and the feature matrices are denoted as $\\tilde{\\mathbf{F}}_i^{\\mathcal{Q}}$ and $\\tilde{\\mathbf{X}}_i^{\\mathcal{Q}}$ in a same way.", + "bbox": [ + 212, + 412, + 784, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Coarse-to-Fine Matching", + "text_level": 1, + "bbox": [ + 215, + 526, + 468, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our backbone outputs both patch-level and point-level features, we use a coarse-to-fine matching strategy [25,34,35] to filter out non-overlapped regions and estimate more accurate correspondences.", + "bbox": [ + 212, + 551, + 782, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Superpoint Matching. Direct matching superpoints by nearest neighbor search lacks robustness to repeated patterns and low-overlapped problems. Therefore, we utilize the Geometric Transformer module [25] presented by Qin et al. to reason the global context of two point clouds, which iteratively uses self and cross-attention to capture intra and inter-point-cloud features.", + "bbox": [ + 212, + 598, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We feed the rotation-invariant features $\\hat{\\mathbf{X}}^{\\mathcal{P}}$ and $\\hat{\\mathbf{X}}^{\\mathcal{Q}}$ and the attached position information $\\hat{\\mathcal{P}}$ and $\\hat{\\mathcal{Q}}$ into this module to get more discriminative features $\\hat{\\mathbf{H}}^{\\mathcal{P}}$ and $\\hat{\\mathbf{H}}^{\\mathcal{Q}}$ . Then, a Gaussian correlation matrix is computed, followed by a dual normalization operation. The superpoint correspondences are obtained by selecting the top- $k$ reliable correspondences $\\hat{\\mathcal{C}} = \\{(\\hat{\\mathbf{p}}_{x_i},\\hat{\\mathbf{q}}_{y_i})|(x_i,y_i)\\in \\arg \\max_{x,y}\\mathbf{S}_{x,y}\\}$ , where $\\mathbf{S}$ is the similarity matrix.", + "bbox": [ + 212, + 672, + 784, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Improvement. Here, we find this module is sensitive to the overlap distribution of the dataset due to its cross-attention layers. This could lead to a significant performance decrease when there is a substantial difference between the overlap distributions of the test set and the training set. To tackle this issue, we present a simple but effective data augmentation technique, RandomCrop,", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where point clouds are randomly cropped to make the model more robust to low-overlapped point cloud pairs. We present more details in the Appendix C.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Point Matching. When the correspondences $\\hat{\\mathcal{C}}_i = (\\hat{\\mathbf{p}}_{x_i},\\hat{\\mathbf{q}}_{y_i})$ of superpoints are established, we search point-level correspondences within $\\mathcal{G}_{x_i}^{\\mathcal{P}}$ and $\\mathcal{G}_{y_i}^{\\mathcal{Q}}$ . We follow [22] that disentangles the similarity and saliency of features by using a matchability head $\\mathbf{W}_m:\\mathbb{R}^{3\\tilde{d}}\\to \\mathbb{R}^{3\\tilde{d}}$ and a saliency head $\\mathbf{W}_s:\\mathbb{R}^{3\\tilde{d}}\\to \\mathbb{R}$ . Specifically, we compute a matching matrix:", + "bbox": [ + 212, + 176, + 784, + 253 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {x _ {i}, y _ {i}} = \\left(\\mathbf {W} _ {m} \\mathbf {X} _ {x _ {i}} ^ {\\mathcal {P}}\\right) ^ {\\mathrm {T}} \\mathbf {W} _ {m} \\mathbf {X} _ {y _ {i}} ^ {\\mathcal {Q}} / \\sqrt {3 \\tilde {d}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 263, + 785, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and then the salient scores are predicted:", + "bbox": [ + 214, + 292, + 511, + 306 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {x _ {i}} ^ {\\mathcal {P}} = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {s} \\mathbf {X} _ {x _ {i}} ^ {\\mathcal {P}}\\right), \\sigma_ {y _ {i}} ^ {\\mathcal {Q}} = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {s} \\mathbf {X} _ {y _ {i}} ^ {\\mathcal {Q}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 314, + 785, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, a soft assignment matrix $\\mathbf{Z}_i$ is computed by:", + "bbox": [ + 214, + 340, + 596, + 356 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {i} = \\sigma_ {x _ {i}} ^ {\\mathcal {P}} \\sigma_ {y _ {i}} ^ {\\mathcal {Q}} \\varphi \\left(\\mathbf {M} _ {x _ {i}, y _ {i}}\\right) _ {x _ {i}} \\varphi \\left(\\mathbf {M} _ {x _ {i}, y _ {i}}\\right) _ {y _ {i}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 364, + 784, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varphi$ is the softmax operation. As a result, we select $N_{f}$ point-level correspondences $\\mathcal{C} = \\{(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})\\}$ across all patch correspondence whose assignment scores are the highest. The impact of $N_{f}$ is analyzed in Fig. 2.", + "bbox": [ + 212, + 391, + 782, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Feature-based hypothesis Proposer", + "text_level": 1, + "bbox": [ + 214, + 455, + 549, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since our rotation-equivariant features encode orientation information of local structures, we derive the pose transformation from matched features.", + "bbox": [ + 212, + 479, + 570, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hypothesis Generation and Selection. Once obtaining point-wise correspondence set $\\mathcal{C} = \\{(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})\\}$ , we leverage their corresponding rotation-equivariant feature maps $\\{(\\tilde{\\mathbf{F}}_{x_j}^{\\mathcal{P}},\\tilde{\\mathbf{F}}_{y_j}^{\\mathcal{Q}})\\}$ to generate multiple model hypotheses. For each correspondence $(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})$ , a rotation can be estimated by solving:", + "bbox": [ + 212, + 523, + 571, + 632 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {R}} _ {j} = \\underset {\\mathbf {R}} {\\arg \\min } \\| \\mathbf {R} \\left(\\tilde {\\mathbf {F}} _ {x _ {j}} ^ {\\mathcal {P}}\\right) ^ {\\mathrm {T}} - \\left(\\tilde {\\mathbf {F}} _ {y _ {j}} ^ {\\mathcal {Q}}\\right) ^ {\\mathrm {T}} \\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 638, + 570, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This can be efficiently solved by SVD [5] in closed form. Subsequently, the translation can be solved by:", + "bbox": [ + 214, + 672, + 568, + 717 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {t}} _ {j} = \\tilde {\\mathbf {q}} _ {y _ {j}} - \\tilde {\\mathbf {R}} _ {j} \\tilde {\\mathbf {p}} _ {x _ {j}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 717, + 570, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a result, we can get a set of model hypotheses $\\mathcal{H}_c = \\{(\\tilde{\\mathbf{R}}_i,\\tilde{\\mathbf{t}}_i)\\}$ . The hypothesis that has the most inlier support will be selected as the alignment solution:", + "bbox": [ + 214, + 739, + 571, + 799 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/872be62db4fc44d150e006ad7bfdfe42a597a8ef8651879900fad93a58f72a6b.jpg", + "image_caption": [ + "Source point cloud Target point cloud" + ], + "image_footnote": [], + "bbox": [ + 588, + 508, + 771, + 583 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56e8e0b2304dcdfcefe6b980b269d476da04bbe5d1505373cb327ed56ec70f6c.jpg", + "image_caption": [ + "Symmetric alignment Correct alignment" + ], + "image_footnote": [], + "bbox": [ + 609, + 601, + 756, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig.4: An illustration of incorrect alignment caused by the ambiguity of using coordinates to align. This problem can be solved by using rotation-equivariant features that contain fine-grained local orientation information.", + "bbox": [ + 581, + 683, + 784, + 792 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {R} ^ {*}, \\mathbf {t} ^ {*}\\right) = \\underset {\\tilde {\\mathbf {R}} _ {j}, \\tilde {\\mathbf {t}} _ {j}} {\\arg \\max } \\sum_ {\\left(\\tilde {\\mathbf {p}} _ {x _ {j}}, \\tilde {\\mathbf {q}} _ {y _ {j}}\\right) \\in \\mathcal {C}} \\mathbb {1} \\left(\\| \\tilde {\\mathbf {R}} _ {j} \\tilde {\\mathbf {p}} _ {x _ {j}} + \\tilde {\\mathbf {t}} _ {j} - \\tilde {\\mathbf {q}} _ {y _ {j}} \\| _ {2} ^ {2} < \\tau_ {d}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 804, + 784, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tau_{d}$ is the acceptance radius. The best model hypothesis will be iteratively refined by its inlier correspondences for a more accurate solution, as in [25].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Discussion. Traditional estimators, such as RANSAC [11] randomly sample multiple pairs of correspondences and use their coordinates to estimate the transformation. Only when all the sampled correspondences belong to inliers, a reliable transformation can be estimated. This approach is inefficient and requires multiple iterations. In contrast, we directly utilize rotation-equivariant features to estimate the pose of the point cloud, and each pair of correspondences generates a hypothesis. This approach is more efficient, as depicted in the Fig. 2. Our method achieves a significantly high registration recall by generating only 1000 hypotheses. In contrast, RANSAC is still inferior to our method even with 50,000 iterations. Moreover, using coordinates to solve the transformation may falsely align the point clouds with opposite orientations, due to ignoring the orientation information, as shown in Fig. 4. This problem can be easily addressed by our method as the rotation-equivariant features encode orientations of local structures. In addition, some methods [1, 28] use regressors to implicitly estimate transformations from equivariant features. Instead, our method can directly solve pose estimation through feature alignment, which is conceptually simpler and computationally efficient.", + "bbox": [ + 212, + 178, + 787, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 460, + 375, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The network is supervised by three losses $\\mathcal{L} = \\mathcal{L}_c + \\mathcal{L}_f + \\mathcal{L}_r$ , including superpoint matching loss $\\mathcal{L}_c$ , point matching loss $\\mathcal{L}_f$ , and contrastive rotation loss $\\mathcal{L}_r$ .", + "bbox": [ + 212, + 491, + 782, + 520 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Point Matching Loss $\\mathcal{L}_f$ . Since point matching is performed within the scopes of matched superpoints, unreliable superpoint matching may lead to sparse supervision signals for point matching. Therefore, we use positive superpoint correspondences computed by ground-truth transformation during the training process. For two grouped point sets $\\mathcal{G}_{x_i}^{\\mathcal{P}}$ and $\\mathcal{G}_{y_i}^{\\mathcal{Q}}$ , the positive correspondences $\\mathcal{C}_i^*$ , having distance errors smaller than a threshold $d_p$ , are identified by the groundtruth. The unmatched points are denoted as $\\mathcal{I}_i$ , $\\mathcal{J}_i$ . We compute the log-likelihood loss on soft assignment matrix $\\mathbf{Z}_i$ and the salient scores $\\sigma_{x_i}$ , $\\sigma_{y_i}$ :", + "bbox": [ + 212, + 522, + 787, + 645 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {f _ {i}} = - \\frac {1}{| \\mathcal {C} _ {i} ^ {*} |} \\sum_ {(x _ {i}, y _ {i}) \\in \\mathcal {C} _ {i} ^ {*}} \\log \\mathbf {Z} _ {x _ {i}, y _ {i}} - \\frac {1}{2 | \\mathcal {I} _ {i} |} \\sum_ {x _ {i} \\in \\mathcal {I} _ {i}} \\log \\left(1 - \\sigma_ {x _ {i}} ^ {\\mathcal {P}}\\right) \\\\ - \\frac {1}{2 \\left| \\mathcal {J} _ {i} \\right|} \\sum_ {y _ {i} \\in \\mathcal {J} _ {i}} \\log \\left(1 - \\sigma_ {y _ {i}} ^ {Q}\\right). \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 657, + 784, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall point matching loss is a sum of $\\mathcal{L}_f = \\sum_i\\mathcal{L}_{f_i}$", + "bbox": [ + 214, + 748, + 627, + 765 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Contrastive Rotation Loss $\\mathcal{L}_r$ . Rotation-equivariant feature inherently encodes pose information of local geometric structure. However, such features of a positive correspondence inevitably be contaminated by data degradation, such as occlusion, noise and down-sampling errors. To make the features more robust to these troubles, we propose a contrastive rotation loss, to supervise point-level", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "correspondences:", + "bbox": [ + 215, + 146, + 339, + 161 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {r _ {i}} = \\frac {1}{| \\mathcal {C} _ {i} ^ {*} | \\cdot \\tilde {d}} \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in \\mathcal {C} _ {i} ^ {*}} \\sum_ {c = 1} ^ {\\tilde {d}} [ \\| \\tilde {\\mathbf {F}} _ {x _ {i}, c} ^ {\\mathcal {P}} \\mathbf {R} _ {g t} ^ {\\mathrm {T}} - \\tilde {\\mathbf {F}} _ {y _ {i}, c} ^ {\\mathcal {Q}} \\| _ {2} ^ {2} - \\alpha ] _ {+} \\tag {11} \\\\ + \\frac {1}{| \\bar {\\mathcal {C}} _ {i} | \\cdot \\tilde {d}} \\sum_ {(x _ {i}, y _ {i}) \\in \\bar {\\mathcal {C}} _ {i}} \\sum_ {c = 1} ^ {\\tilde {d}} [ \\beta - \\| \\tilde {\\mathbf {F}} _ {x _ {i}, c} ^ {\\mathcal {P}} \\mathbf {R} _ {g t} ^ {\\mathrm {T}} - \\tilde {\\mathbf {F}} _ {y _ {i}, c} ^ {\\mathcal {Q}} \\| _ {2} ^ {2} ] _ {+}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 175, + 784, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_i^*$ is the aforementioned positive correspondence set, $\\bar{\\mathcal{C}}_i$ is the negative correspondence set whose point pair has spatial distance greater than $d_{n}$ , $\\alpha = 0.1$ and $\\beta = 1.4$ are positive and negative margins, and $[\\cdot]_{+}$ is the ReLU operation.", + "bbox": [ + 212, + 281, + 785, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Superpoint Matching Loss $\\mathcal{L}_c$ . To supervise the rotation-invariant features of superpoints, we utilize the overlap-aware circle loss that considers the overlap ratio of two superpoints to compute more accurate gradients. For more details, please refer to GeoTrans [25].", + "bbox": [ + 212, + 327, + 785, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 412, + 375, + 429 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To evaluate the performance of our method, we compare our method with state-of-the-art approaches and perform extensive experiments on 3DMatch [37], 3DLoMatch [15], and KITTI Odometry [12]. Finally, ablation studies are conducted to analyze the components of our method. All the experiments are conducted with a RTX 3090 GPU with Intel (R) Xeon (R) Silver 4314 CPU. More implementation details are reported in Appendix D.", + "bbox": [ + 212, + 444, + 787, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Indoor Dataset: 3DMatch & 3DLoMatch", + "text_level": 1, + "bbox": [ + 215, + 556, + 599, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "DataSet. 3DMatch [37] and 3DLoMatch [15] are collected from indoor RGB-D datasets. 3DMatch contains 62 scenes and there are $46/8/8$ scenes for training/validation/testing. It consists of point cloud pairs with overlap ratios greater than $30\\%$ . 3DLoMatch is only used for testing that contains point cloud pairs with overlap ratios from $10\\%$ to $30\\%$ . We follow the protocols [15, 25, 28] to perform all the experiments.", + "bbox": [ + 212, + 582, + 785, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baseline. We compare our method with strong baselines: 1) FCGF [7], Predator [15], GeoTrans [25], and PEAL [36] are advanced scene-level feature extractors, and some of them [25,36] follow a coarse-to-fine manner to extract robust correspondences. PEAL requires overlap prior as input, we follow the office implementation that leverages 3D overlap prior predicted by GeoTrans for a fair comparison. 2) YOHO [27], RoReg [28], and RoITR [35] are also scene-level feature extractors but focus on rotation-invariant/equivariant network designs. 3) SpinNet [2] and BUFFER [1] leverage patch-level feature extractors with elaborate designs for rotation-invariance.", + "bbox": [ + 212, + 672, + 785, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. We follow [15,25] that use three evaluation metrics, including Inlier Ratio (IR), Feature Matching Recall (FMR), and Registration Recall (RR). To", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e805df25ddd1cbbb824633e98e8bf6506ff488033a20bf6246c62bd84ce1ebc1.jpg", + "table_caption": [ + "Table 1: Evaluation results on 3DMatch and 3DLoMatch. Methods with the RANSAC estimator are marked by $\\diamond$ , which exploit 5000 points to establish correspondences." + ], + "table_footnote": [], + "table_body": "
MethodSize (MB)3DMatch3DLoMatch
FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)
FCGF° [7]8.7694.731.182.80.1259.49.838.00.13
SpinNet° [2]1.4197.647.588.69.8575.320.559.89.03
YOHO [27]12.3898.264.490.82.8178.925.966.02.62
Predator° [15]7.4396.658.089.00.6478.226.764.40.47
GeoTrans [25]9.8398.170.992.40.1887.443.574.30.17
RoReg [28]12.7198.281.693.02.2782.339.670.12.10
BUFFER [1]0.92--92.10.22--70.30.21
RoITR° [35]10.1098.082.491.90.3689.254.674.10.34
PEAL [36]9.8398.471.094.21.4688.346.078.81.19
Ours3.8498.576.995.00.1788.347.580.50.17
", + "bbox": [ + 220, + 185, + 785, + 339 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "evaluate the performance of different methods on rotated datasets, RoITR [35] uses two ways to compute the RMSE on the original and rotated datasets, which cannot show the robustness of methods against pose variations. To tackle this issue, we follow [28] to report Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) on both original and rotated datasets. Please refer to the Appendix E for more details about these evaluation metrics.", + "bbox": [ + 217, + 369, + 784, + 459 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative results. We report the experimental results in Table 1. Compared with SOTA methods, our approach achieves a significantly high RR on both 3DMatch and 3DLoMatch, while simultaneously maintaining lightweight and computation efficient. PEAL [36] is a formidable competitor that employs an iterative manner for alignment due to the requirements of overlap prior, resulting in significant time overhead. In contrast, our method achieves a speed improvement of 7 times compared to PEAL, while also outperforming it by $1.7\\%$ in terms of RR on 3DLoMatch. GeoTrans [25], RoITR [35], and our method all employ a coarse-to-fine matching framework. Despite using a more lightweight feature extractor, our RR is significantly higher than theirs, particularly on 3DLoMatch, where our method outperforms them by $6.2\\%$ and $6.4\\%$ respectively. This can be attributed to the ability of our rotation-equivariant network to focus more on learning discriminative structural information. YOHO [27] and RoReg [28] are extremely time-consuming due to the group feature extractor and embedder. Our efficiency advantage stems from two aspects. Firstly, we utilize a lightweight scene-wise feature extractor, which only needs a forward process to extract dense rotation-equivariant features. Secondly, our feature-based hypothesis proposer can simultaneously generate multiple valid hypotheses by using matched rotation-equivariant features. Thus, it is more efficient compared to RANSAC [11] with many approaches [2, 7, 15, 35] used.", + "bbox": [ + 217, + 460, + 785, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experiments on the rotated dataset. To test the robustness against rotations, we follow previous works [28, 35] that evaluate all the methods on the rotated 3DLoMatch, where arbitrary rotations are applied to point clouds. The experimental results are shown in Table 2. As we can see, PEAL experiences a decrease in TR by $2.5\\%$ on the Rotated 3DLoMatch dataset, while our", + "bbox": [ + 217, + 763, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1cf4c04b061ed6ae1bba0a5e8f424b59f85b6a736d0a1eda34bcf4afb9995f18.jpg", + "table_caption": [ + "Table 2: Evaluation results on 3DLoMatch and Rotated 3DLoMatch. For easy comparison of the impact of rotation, the numerical changes w.r.t. TR on the rotated dataset are annotated in the top right corner of each value. Methods with rotation invariant/equivariant designs are marked by * symbol." + ], + "table_footnote": [], + "table_body": "
MethodSize (MB)3DLoMatchRotated 3DLoMatch
RE (°↓)TE (cm ↓)TR (% ↑)RE (°↓)TE (cm ↓)TR (% ↑)
FCGF [7]8.764.8412.8739.64.7413.3924.5-15.1
PREDATOR [15]7.433.6110.6565.63.5510.3064.0-1.6
GeoTrans [25]9.832.918.7175.42.948.8572.6-2.8
PEAL [36]9.832.848.6481.22.868.5378.7-2.5
YOHO* [27]12.383.5410.3466.63.6110.1667.1+0.5
RoReg* [28]12.713.019.2671.33.039.2871.0-0.3
BUFFER* [1]0.923.039.8674.43.029.9974.7+0.3
RoITR* [35]10.102.959.0375.12.979.0875.5+0.4
Ours*3.842.878.8381.32.848.7181.8+0.5
", + "bbox": [ + 220, + 212, + 785, + 356 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "method demonstrates an increase of $0.5\\%$ . This result validates the robustness of our approach against rotations, illustrating its ability to handle such variations. Methods without rotation-invariant design, such as FCGF [7], Predator [15], and GeoTrans [25] also show performance degradation, indicating that networks trained with rotation augmentation are vulnerable to rotation variations. On the other hand, methods such as YOHO [27], RoReg [28], RoITR [35], and BUFFER [1], which incorporate rotation equivariant or invariant designs, demonstrate better robustness against data rotations. However, our method significantly outperforms these methods by $6\\%$ w.r.t TR on both 3DLoMatch and Rotated 3DLoMatch. This is because our backbone can capture spatial geometric information better, thus extracting more distinctive descriptors.", + "bbox": [ + 212, + 406, + 787, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative results. We show some registration results in Fig. 5, which demonstrate that our method can estimate more accurate patch-level and point-level correspondences and can successfully align point cloud pairs with low-overlapped ratios. Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method suc", + "bbox": [ + 212, + 583, + 431, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "cessfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity. Please refer to Appendix G for more visualization results.", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/57e7c06b187c6799d5c32b4d99473c80cc496d629a69e79dac984c417b94811f.jpg", + "table_caption": [ + "Table 3: Evaluation results on KITTI Odometry." + ], + "table_footnote": [], + "table_body": "
MethodSizeRE (°)TE (cm)TR (%)Time (s)
FCGF [7]8.760.309.596.6-
D3Feat [4]14.080.307.299.8-
Predator [15]22.770.276.899.80.77
SpinNet [2]1.410.479.999.116.24
CoFiNet [34]5.480.418.299.80.59
GeoTrans [25]25.500.236.299.80.26
BUFFER [1]0.920.267.199.80.27
Ours2.080.234.999.80.21
", + "bbox": [ + 446, + 641, + 785, + 768 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a0e48852e9221594f77b1c4ba1dc3b744b58a355ef96c4fadfef1ae120c4b1d1.jpg", + "image_caption": [ + "Fig. 5: Qualitative results. Our method can successfully align low-overlapped pairs with higher IR and PIR (Patch IR). Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method successfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Outdoor Dataset: KITTI Odometry", + "text_level": 1, + "bbox": [ + 215, + 412, + 555, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Dataset. We follow previous methods [7,25] to evaluate our method on KITTI Odometry [12]. It comprises 11 urban scene sequences captured by a LiDAR sensor. We use 0-5/6-7/8-10 sequences for training/validation/testing, respectively.", + "bbox": [ + 212, + 439, + 782, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Metrics. We use the Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) to evaluate our method. Following [7,25], the TR is computed with rotation threshold $5^{\\circ}$ and translation threshold $2\\mathrm{m}$ .", + "bbox": [ + 212, + 484, + 784, + 530 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Experimental results. The experimental results are shown in Table 3. On KITTI dataset, many methods have reached performance saturation, while our approach achieves the highest accuracy when achieving the highest TR. Our method also shows advantages w.r.t. model size and running speed.", + "bbox": [ + 212, + 531, + 784, + 590 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 614, + 387, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To better understand the designs of our method, we conduct comprehensive ablation experiments and the evaluated results are reported in Table 4.", + "bbox": [ + 212, + 642, + 782, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of backbone. To investigate the impact of different convolution methods, we compare our PARE-Conv with VN [8]. We leverage both node features [24] and edge features [29] for evaluation. The experimental results indicate that our approach achieves significant improvement compared to VN whether using node features or edge features. This demonstrates the enhanced representative power of PARE-Conv to learn more distinctive features. We also compare PARE-Conv with KPCnv [26], refer to the Appendix F.5 for details.", + "bbox": [ + 212, + 672, + 784, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of pose estimator. To demonstrate the superiority of our pose estimator, we compared it with RANSAC [11] and LGR [25]. RANSAC randomly selects a few pairs of correspondences to generate hypotheses, while LGR generates a hypothesis using the correspondences within a matched patch pair.", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In contrast, our approach directly generates a hypothesis using the equivariant features of a pair of matched points. As shown in Table 4, our method achieves the highest RR, outperforming RANSAC and LGR by $3\\%$ and $1.6\\%$ respectively on 3DLoMatch. This validates that estimating poses using rotation-equivariant features based on local structures is more robust than using point-to-point correspondences. Additionally, the results of FMR and IR are identical as we used the same correspondences for different pose estimators.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/8413e2cd8d17377a7b45784952ce9c23fef5bfb53bdfe559e1b2ec9fe2b04ba8.jpg", + "table_caption": [ + "Table 4: Ablation experiments of our method." + ], + "table_footnote": [], + "table_body": "
Component Method3DMatch3DLoMatch
FMR (% ↑) IR (% ↑) RR (% ↑)FMR (% ↑) IR (% ↑) RR (% ↑)
BackboneVN (Node) [8]98.671.093.487.841.476.6
PARE-Conv (Node)98.974.194.5 +1.188.344.078.8 +2.2
VN (Edge) [8]98.772.494.188.843.577.8
PARE-Conv (Edge)98.576.995.0 +0.988.147.580.5 +2.7
Pose EstimatorRANSAC98.576.993.788.347.577.5
LGR [25]98.576.994.0 +0.388.347.578.9 +1.4
Ours98.576.995.0 +1.388.347.580.5 +3.0
Rotation LossWithout98.374.493.388.745.579.3
L298.665.494.2 +0.985.935.075.4 -3.9
Contrastive98.576.995.0 +1.788.347.580.5 +1.2
", + "bbox": [ + 220, + 305, + 787, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation of rotation loss. The contrastive rotation loss is designed to improve the robustness of rotation-equivariant features to data occlusion, noise, and partial overlap. When the rotation-equivariant features are not supervised by a loss function (Indeed, it is indirectly supervised by the loss function of rotation-invariant features, as the invariant features are derived from the equivariant features), the method shows decent performance on 3DLoMatch but shows poor performance on 3DMatch. When using the L2 loss to supervise equivariant features of positive point pairs, there is a slight improvement in RR on 3DMatch but shows performance degradation on 3DLoMatch. However, when utilizing our contrastive rotation loss, the performance of the method is enhanced on both 3DMatch and 3DLoMatch, demonstrating its superiority.", + "bbox": [ + 212, + 493, + 787, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 684, + 359, + 700 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose PARE-Net, a lightweight network for fast and robust point cloud registration. We introduce position-aware rotation-equivariant convolution to learn spatial information effectively, thus extracting more distinctive descriptors. We also propose an efficient feature-based hypothesis proposer to generate reliable model hypotheses. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art approaches and shows great robustness to rotation variations. The limitations of our method and future works are discussed in Appendix H.", + "bbox": [ + 212, + 718, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Science and Technology Major Project of China under Grant No. 2020AAA0108102, the National Natural Science Foundation of China under Grant Nos. 62327808, 62073257 and 62088102, and the Fundamental Research Funds for the Central Universities under Grant No. xzy022024007.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 247, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ao, S., Hu, Q., Wang, H., Xu, K., Guo, Y.: BUFFER: Balancing accuracy, efficiency, and generalizability in point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1255-1264 (2023)", + "2. Ao, S., Hu, Q., Yang, B., Markham, A., Guo, Y.: SpinNet: Learning a general surface descriptor for 3D point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11753-11762 (2021)", + "3. Bai, X., Luo, Z., Zhou, L., Chen, H., Li, L., Hu, Z., Fu, H., Tai, C.L.: PointDSC: Robust point cloud registration using deep spatial consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15859-15869 (2021)", + "4. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3Feat: Joint learning of dense detection and description of 3D local features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6359-6367 (2020)", + "5. Besl, P.J., McKay, N.D.: Method for registration of 3-D shapes. In: Sensor Fusion IV: Control Paradigms and Data Structures. vol. 1611, pp. 586-606 (1992)", + "6. Chen, Z., Sun, K., Yang, F., Tao, W.: SC²-PCR: A second order spatial compatibility for efficient and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13221-13231 (2022)", + "7. Choy, C., Park, J., Koltun, V.: Fully convolutional geometric features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8958-8966 (2019)", + "8. Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., Guibas, L.J.: Vector Neurons: A general framework for SO(3)-equivariant networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12200-12209 (2021)", + "9. Deng, H., Birdal, T., Ilic, S.: PPF-FoldNet: Unsupervised learning of rotation invariant 3D local descriptors. In: Proceedings of the European Conference on Computer Vision. pp. 602-618 (2018)", + "10. Deng, H., Birdal, T., Ilic, S.: PPFNet: Global context aware local features for robust 3D point matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 195-205 (2018)", + "11. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. vol. 24, pp. 381-395. ACM New York (1981)", + "12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012)" + ], + "bbox": [ + 218, + 280, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Gojcic, Z., Zhou, C., Wegner, J.D., Wieser, A.: The perfect match: 3D point cloud matching with smoothed densities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5545-5554 (2019)", + "14. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3D local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016)", + "15. Huang, S., Gojcic, Z., Usvyatsov, M., Wieser, A., Schindler, K.: PREDATOR: Registration of 3D point clouds with low overlap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4267-4276 (2021)", + "16. Kaya, M., Bilge, H.S.: Deep metric learning: A survey. Symmetry 11(9), 1066-1092 (2019)", + "17. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 153-161 (2017)", + "18. Lee, J., Kim, S., Cho, M., Park, J.: Deep hough voting for robust global registration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15994-16003 (2021)", + "19. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022)", + "20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems 32(8), 3412-3432 (2020)", + "21. Lu, W., Zhou, Y., Wan, G., Hou, S., Song, S.: L3-Net: Towards learning based LiDAR localization for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6389-6398 (2019)", + "22. Philipp, L., Paul-Edouard, S., Marc, P.: LightGlue: Local Feature Matching at Light Speed. In: Proceedings of the IEEE International Conference on Computer Vision (2023)", + "23. Pomerleau, F., Colas, F., Siegwart, R.: A review of point cloud registration algorithms for mobile robotics. Foundations and Trends in Robotics 4(1), 1-104 (2015)", + "24. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)", + "25. Qin, Z., Yu, H., Wang, C., Guo, Y., Peng, Y., Xu, K.: Geometric transformer for fast and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11143-11152 (June 2022)", + "26. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPConv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6411-6420 (2019)", + "27. Wang, H., Liu, Y., Dong, Z., Wang, W.: You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1630-1641 (2022)", + "28. Wang, H., Liu, Y., Hu, Q., Wang, B., Chen, J., Dong, Z., Guo, Y., Wang, W., Yang, B.: RoReg: Pairwise point cloud registration with oriented descriptors and local rotations. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 10376-10393 (2023). https://doi.org/10.1109/TPAMI.2023.3244951" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Yao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph CNN for learning on point clouds. ACM Transactions On Graphics (ToG) 38(5), 1-12 (2019)", + "30. Xu, M., Ding, R., Zhao, H., Qi, X.: PACConv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3173-3182 (2021)", + "31. Yang, B., Bender, G., Le, Q.V., Ngiam, J.: CondConv: Conditionally parameterized convolutions for efficient inference. In: Proceedings of Advances in Neural Information Processing Systems. vol. 32 (2019)", + "32. Yao, R., Du, S., Cui, W., Ye, A., Wen, F., Zhang, H., Tian, Z., Gao, Y.: Hunter: Exploring high-order consistency for point cloud registration with severe outliers. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14760-14776 (2023). https://doi.org/10.1109/TPAMI.2023.3312592", + "33. Yew, Z.J., Lee, G.H.: REGTR: End-to-end point cloud correspondences with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6677-6686 (2022)", + "34. Yu, H., Li, F., Saleh, M., Busam, B., Ilic, S.: CoFiNet: Reliable coarse-to-fine correspondences for robust point cloud registration. In: Proceedings of Advances in Neural Information Processing Systems. pp. 23872-23884 (2021)", + "35. Yu, H., Qin, Z., Hou, J., Saleh, M., Li, D., Busam, B., Ilic, S.: Rotation-invariant transformer for point cloud matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5384-5393 (2023)", + "36. Yu, J., Ren, L., Zhang, Y., Zhou, W., Lin, L., Dai, G.: PEAL: Prior-embedded explicit attention learning for low-overlap point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17702-17711 (2023)", + "37. Zeng, A., Song, S., Nießner, M., Fisher, M., Xiao, J., Funkhouser, T.: 3DMatch: Learning local geometric descriptors from RGB-D reconstructions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1802-1811 (2017)", + "38. Zhang, Z., Dai, Y., Sun, J.: Deep learning based point cloud registration: An overview. Virtual Reality and Intelligent Hardware 2(3), 222-246 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 590 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PARE-Net", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_model.json b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1034e4efb2d1322fffc9d333530fe958d49c40ce --- /dev/null +++ b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_model.json @@ -0,0 +1,2170 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.213, + 0.772, + 0.244 + ], + "angle": 0, + "content": "Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.256, + 0.782, + 0.298 + ], + "angle": 0, + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, China" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.334, + 0.74, + 0.667 + ], + "angle": 0, + "content": "Abstract. Learning rotation-invariant distinctive features is a fundamental requirement for point cloud registration. Existing methods often use rotation-sensitive networks to extract features, while employing rotation augmentation to learn an approximate invariant mapping rudely. This makes networks fragile to rotations, overweight, and hinders the distinctiveness of features. To tackle these problems, we propose a novel position-aware rotation-equivariant network, for efficient, light-weighted, and robust registration. The network can provide a strong model inductive bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned limitations. To further improve the distinctiveness of descriptors, we propose a position-aware convolution, which can better learn spatial information of local structures. Moreover, we also propose a feature-based hypothesis proposer. It leverages rotation-equivariant features that encode fine-grained structure orientations to generate reliable model hypotheses. Each correspondence can generate a hypothesis, thus it is more efficient than classic estimators that require multiple reliable correspondences. Accordingly, a contrastive rotation loss is presented to enhance the robustness of rotation-equivariant features against data degradation. Extensive experiments on indoor and outdoor datasets demonstrate that our method significantly outperforms the SOTA methods in terms of registration recall while being lightweight and keeping a fast speed. Moreover, experiments on rotated datasets demonstrate its robustness against rotation variations. Code is available at https://github.com/yaorz97/PARENet." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.68, + 0.741, + 0.707 + ], + "angle": 0, + "content": "Keywords: Point cloud registration \\(\\cdot\\) Rotation invariance/equivalence \\(\\cdot\\) Robust transformation estimation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.734, + 0.377, + 0.75 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Point cloud registration, estimating a rigid transformation to align two point clouds that are partially overlapped, is fundamental research in 3D computer vision. It is widely applied in autonomous driving [20,21], robot localization [23], and 3D reconstruction [38]. Feature-based point cloud registration framework is widely studied, which mainly involves descriptor extraction [4, 15, 25, 34] and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "robust transformation estimation [3,6,32]. One inherent contradiction in point cloud registration is the pose variation of point clouds and the desired invariance of their descriptors. To extract such rotation-invariant descriptors, plenty of approaches have been proposed that can be roughly categorized into patch-wise and scene-wise feature extractors." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.247, + 0.499, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.247, + 0.785, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.369 + ], + "angle": 0, + "content": "Fig. 1: Experimental results on 3DMatch and 3DLoMatch. Our method significantly outperforms state-of-the-art methods w.r.t. registration recall (RR) while maintaining fast speed and lightweight." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.552 + ], + "angle": 0, + "content": "Patch-wise feature extractors [2, 9, 10] learn geometric features from local patches of point clouds and typically employ data preprocessing techniques to achieve rotation-invariance. Two common approaches include: 1) extracting rotation invariant statistics as initial representation from local patches, such as distance, angle, and point pair features (PPF [9, 10]); and 2) estimating a local reference frame (LRF [2]) to normalize poses of local patches. The former approach suffers from information loss, while the latter approach is susceptible to data occlusion and noise, lacking robustness. Moreover, these extractors are generally computationally expensive because all patches of a point cloud are processed individually, without sharing intermediate representations [7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Scene-wise feature extractors [4,7,15,25,34] learn dense feature descriptors for every point in the point cloud, and most of them employ extensive rotation augmentation in the training process to rudely enforce inherently rotation-sensitive networks (e.g., KPCov [26]) to learn rotation-invariant mappings. These methods have three shortcomings. Firstly, models are fragile to unseen rotations because it is difficult to cover the SO(3) space in the training process [35]. Secondly, learning rotation-invariant mappings requires more network parameters. Thirdly, these networks are limited to generating discriminative feature descriptors as they cannot focus solely on learning structural distinctiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.749 + ], + "angle": 0, + "content": "In this paper, we propose a novel registration network, PARE-Net, whose core is to fully exploit the advantages of rotation-equivariant networks in both feature extraction and transformation estimation, thereby achieving a lightweight, efficient, and robust registration method, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.84 + ], + "angle": 0, + "content": "First of all, we propose a Position-Aware Rotation-Equivariant Convolution (PARE-Conv) that enhances the representational power of networks for better distinctive feature learning. Vector Neuron (VN) [8] builds a general rotation-equivariant network framework. However, its representative power is limited as it ignores the importance of spatial localization property of local points and utilizes multi-layer perceptrons (MLPs) to crudely learn spatial encodings. In" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.799, + 0.328 + ], + "angle": 0, + "content": "stead, PARE-Conv is proposed to explicitly leverage the spatial information of point clouds while guaranteeing the rotation equivariant property. Specifically, we first define a set of shadow kernel points equipped with learnable weights. Then, we introduce a correlation network that predicts rotation-invariant correlations between points and shadow kernels according to rotation-equivariant spatial features of points. Therefore, the geometric structures of point clouds are well captured. Finally, the correlation scores are subsequently leveraged to dynamically assemble kernel weights for rotation-equivariant convolution. As a result, PARE-Conv inherently provides a strong model bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned problems raised by rotation-sensitive networks. Moreover, it can learn the spatial structure of point clouds better, facilitating to extraction of more distinctive descriptors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.601, + 0.525 + ], + "angle": 0, + "content": "Secondly, we propose a feature-based hypothesis proposer that can efficiently generate reliable solution hypotheses. Once obtaining correspondences by feature matching, rotation-equivariant features encoding fine-grained structure information are used to explicitly generate multiple hypotheses. This way can alleviate the orientation ambiguity stemming from optimizing distance errors of correspondences. Moreover, to make the equivariant features more robust against data degradation, we present a contrastive rotation loss inspired by metric learning [7, 16], facilitating more robust pose estimation. As shown in Fig. 2, our method can achieve significantly higher registration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.601 + ], + "angle": 0, + "content": "recall compared to RANSAC [11] and LGR (presented in GeoTrans [25]) when varying the number of hypotheses, demonstrating our method is more robust than classical estimators. Because they require multiple correspondences and the probability of sampling all correct correspondences decreases rapidly as the number of used correspondences increases. Overall, our main contributions are:" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.359, + 0.787, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.439, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig.2: Comparison of our feature-based hypothesis proposer, RANSAC and LGR on 3DLoMatch." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.613, + 0.785, + 0.657 + ], + "angle": 0, + "content": "- A lightweight, efficient, and robust point cloud registration method that is achieved by fully exploiting the advantages of rotation-equivariant networks in feature extraction and pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.658, + 0.785, + 0.688 + ], + "angle": 0, + "content": "- A position-aware rotation-equivariant convolution that can better leverage spatial information, enabling the extraction of discriminative descriptors." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.689, + 0.785, + 0.734 + ], + "angle": 0, + "content": "- A feature-based hypothesis proposer that can efficiently generate multiple reliable hypotheses and a contrastive rotation loss makes the rotation-equivariant features more robust to data degradation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.613, + 0.785, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.388, + 0.777 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Patch-wise feature extractors. In the early stage, most learning-based feature descriptors [2, 9, 10, 13, 17, 37] were patch-wise extractors, primarily due to limitations in point cloud convolution techniques, which could not handle" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "large-scale point clouds effectively. During this period, the works drew inspiration from traditional feature descriptors [14] and focused on designing features with rotation-invariance. PPFNet [10] and PPFoldNet [9] extracted rotation-invariant PPFs as initial features and further learned high-dimensional features. PerfectMatch [13] and SpinNet [2] utilized LRF to normalize the pose of local patches and then employed networks to learn the density/structural features of patches. These methods are primarily limited by their computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Scene-wise feature extractors [4, 7, 15, 19, 25, 33-36] have become the mainstream research in recent years, due to their high efficiency. These methods mainly achieved rotation invariance through data augmentation and aim to tackle the low-overlapped problem. Predator [15] designed an overlap-attention module to detect overlap regions between two point clouds. CoFiNet [34] and GeoTrans [25] presented a coarse-to-fine matching scheme, and GeoTrans introduced a geometric transformer module that significantly increased the matching accuracy by injecting rotation-invariant position information. PEAL [36] presented an explicit one-way attention module that iteratively leveraged overlap prior to relieving feature ambiguity while bringing expensive computation costs. Moreover, some methods with rotation invariant/equivariant designs were proposed. YOHO [27] and RoReg [28] exploited the group convolution technique to extract rotation-invariant/equivariant group features, which are extremely computation expensive. Because they need to extract point features under 60 different poses and perform convolutions on massive point features. RoITR [35] used PPF to design an invariant convolution and presented a rotation invariant transformer to extract distinctive features. BUFFER [1] combined scene-wise and patch-wise feature extractors to achieve faster alignment with great generalization ability. It first leveraged a rotation-equivariant network built on vanilla VN to extract invariant and equivariant features for keypoints and orientations prediction. Then, it also used SpinNet to extract patch-wise features due to the limited representative ability of vanilla VN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Robust transformation estimators. Since the estimated correspondences are usually contaminated by outliers, a robust transformation estimator is necessary. RANSAC [11] is the most widely used method due to its simplicity, but it typically exhibits lower accuracy and requires plenty of iterations for convergence. Recently, some learning-based methods [3,6,18,32] are proposed, whose core idea is to design networks to learn spatial consistency between correspondences, enabling efficient inlier identification and reliable hypothesis generation. We compare our approach with these methods in Appendix F.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.331, + 0.748 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Given two partially overlapped point clouds \\(\\mathcal{P} = \\{\\mathbf{p}_i\\in \\mathbb{R}^3 |i = 1,2,\\dots ,N\\}\\) and \\(\\mathcal{Q} = \\{\\mathbf{q}_j\\in \\mathbb{R}^3 |j = 1,2,\\dots ,M\\}\\), we aim to estimate a transformation \\(\\mathbf{T} = (\\mathbf{R}\\in SO(3),\\mathbf{t}\\in \\mathbb{R}^3)\\) that aligns the source point cloud to the target one. To address this problem, we propose a registration framework based on rotation-equivariant network \\(f(\\mathbf{X})\\), whose output is equivariant to the SO(3) action \\(\\mathbf{R}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.143, + 0.773, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig. 3: The framework of our method. Given point clouds \\(\\mathcal{P}\\) and \\(\\mathcal{Q}\\), a hierarchical backbone based on PARE-Conv is used to extract patch-level and point-level features. Then, we follow a coarse-to-fine scheme to obtain point correspondences using rotation-invariant features. Finally, the feature-based hypothesis proposer leverages the matched rotation-equivariant features \\(\\tilde{\\mathbf{F}}_{x_{jn}}^{\\mathcal{P}}\\) and \\(\\tilde{\\mathbf{F}}_{y_{jn}}^{\\mathcal{Q}}\\), which encode fine-grained structure orientations, to generate multiple reliable hypotheses. The best hypothesis is selected and refined as the final solution \\((\\mathbf{R}^{*},\\mathbf{t}^{*})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.566 + ], + "angle": 0, + "content": "applied on the input point cloud \\(\\mathbf{X}\\), i.e., \\(f(\\mathbf{R} \\circ \\mathbf{X}) = \\mathbf{R} \\circ f(\\mathbf{X})\\). As we can see, rotation-equivariant features \\(f(\\mathbf{R} \\circ \\mathbf{X})\\) couple the structure information \\(f(\\mathbf{X})\\) and orientation information \\(\\mathbf{R}\\) of point clouds, which can be further decoupled for feature matching and transformation estimation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.657 + ], + "angle": 0, + "content": "The framework of PARE-Net is depicted in Fig. 3. We build a hierarchical backbone based on PARE-Conv to extract both patch-level and point-level rotation invariant/equivariant features. Correspondences are estimated in a coarse-to-fine fashion by using rotation-invariant features. After that, the feature-based hypothesis proposer employs rotation-equivariant features to generate multiple reliable hypotheses, and the best one is selected as the final output \\((\\mathbf{R}^{*},\\mathbf{t}^{*})\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.679, + 0.644, + 0.695 + ], + "angle": 0, + "content": "3.1 Position-Aware Rotation-Equivariant Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "PARE-Conv. Recent point cloud registration approaches [15,25,34,36] mainly exploited KPCov [26] to build backbones. However, KPCov is inherently rotation-sensitive and needs data augmentation to learn an approximate invariant mapping. As a result, models trained on limited samples of continuous SO(3) space are fragile to rotation variations. Moreover, enforcing the network to learn a rotation-invariant mapping inevitably requires more model parameters and also hinders the distinctiveness of features. Inspired by dynamic and conditional convolutions [30,31], we propose position-aware rotation-equivariant convolution to learn distinctive rotation-invariant/equivariant features, as shown in Fig. 3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.19 + ], + "angle": 0, + "content": "Given a point cloud \\(\\mathcal{P} = \\{\\mathbf{p}_i\\in \\mathbb{R}^3|i = 1,2,\\dots ,N\\}\\) and attached feature map \\(\\mathbf{F} = \\{\\mathbf{F}_i\\in \\mathbb{R}^{C\\times 3}\\}\\), the general convolution of \\(\\mathbf{F}\\) by a kernel \\(g\\) at a point \\(\\mathbf{p}_i\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.192, + 0.786, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {F} * g\\right) \\left(\\mathbf {p} _ {i}\\right) = \\sum_ {\\mathbf {p} _ {j} \\in \\mathcal {N} _ {i}} g \\left(\\mathbf {p} _ {j} - \\mathbf {p} _ {i}\\right) \\mathbf {F} _ {j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.399 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_i\\) is the K-nearest neighbor set of the point \\(\\mathbf{p}_i\\). The core problem is the definition of kernel function \\(g\\). In vanilla VN [8], it omits the position information and simply uses a linear layer, \\(\\mathbf{W}\\mathbf{F}_i\\), followed by a VN non-linear layer and a VN pooling layer to aggregate local information, where \\(\\mathbf{W} \\in \\mathbb{R}^{C' \\times C}\\) is a learnable weight matrix. Instead, we define a set of shadow kernel points, equipped with learnable weight matrices \\(\\{\\mathbf{W}_k | k = 1, 2, \\dots, K\\}\\). In image convolution, each pixel corresponds to a convolutional kernel, and the fixed spatial arrangement of the kernels enables the learning of the distribution information of image pixels. Due to the irregularity of point clouds, establishing such point-to-kernel correspondences is intractable. We follow KPConv [26] to establish soft assignments between data points and kernel points and the kernel function \\(g\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.411, + 0.786, + 0.451 + ], + "angle": 0, + "content": "\\[\ng \\left(\\mathbf {p} _ {i j}\\right) = \\sum_ {k = 1} ^ {K} \\operatorname {s o f t m a x} (\\gamma \\left(\\mathbf {p} _ {i j}, k\\right)) \\mathbf {W} _ {k}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.755 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{ij} = \\mathbf{p}_j - \\mathbf{p}_i\\) is the relative coordinate of \\(\\mathbf{p}_j\\) and \\(\\gamma (\\cdot)\\) is a correlation function. In KPConv [26], it uses a linear correlation \\(\\gamma (\\mathbf{p}_{ij},k) = \\max (0,1-\\|\\mathbf{p}_{ij} - \\tilde{\\mathbf{x}}_k\\| / \\sigma)\\), where \\(\\tilde{\\mathbf{x}}_k\\) is the 3D coordinate of the \\(k\\)-th kernel point and \\(\\sigma\\) is a distance influence factor. However, kernel-to-point distances are sensitive to rotations thus cannot guarantee a rotation-equivariant convolution. Therefore, the key challenge lies in ensuring the rotation equivariance of the convolution while well capturing the spatial information of patches. Inspired by PACConv [30], we use a mini-network that consumes rotation-equivariant spatial features and predicts rotation-invariant correlation scores between data points and kernel points. Note that predicting rotation-invariant scores is necessary to guarantee the rotation-equivariance of the convolution (see Appendix A for the proof). As shown in Fig. 3, we implement this function with 1) a VN-MLP block, learning geometric features \\(\\mathbf{X}_j \\in \\mathbb{R}^{d \\times 3}\\) from rotation-equivariant spatial statistics, such as \\(\\mathbf{p}_{ij}\\), \\(\\frac{1}{|\\mathcal{N}_i|} \\sum_j \\mathbf{p}_{ij}\\) and their cross product; 2) a L2-Norm layer, computing the magnitude of the vector neurons \\(\\tilde{\\mathbf{X}}_j = \\| \\mathbf{X}_j\\|_2 \\in \\mathbb{R}^d\\) to transform the feature into a rotation-invariant form; and 3) a MLP block, mapping this feature vector into K dimension to obtain the correlation vector, \\(\\gamma_j = \\mathrm{MLP}(\\tilde{\\mathbf{X}}_j) \\in \\mathbb{R}^K\\). Finally, a softmax operation is leveraged to output correlation scores. As a result, our convolution is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.766, + 0.786, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {F} * g\\right) \\left(\\mathbf {p} _ {i}\\right) = \\sum_ {\\mathbf {p} _ {j} \\in \\mathcal {N} _ {i}} \\sum_ {k} \\gamma_ {j k} \\mathbf {W} _ {k} \\mathbf {F} _ {j}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Compared to KPConv using a linear correlation, our correlation network can learn more complex relationships. Moreover, PARE-Conv predicts correlation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "scores depending on local spatial property, thus can leverage the spatial information better and can provide stronger expressive power to extract more discriminative descriptors. Moreover, it is compatible with DGCNN [29], allowing replacing the node feature \\(\\mathbf{F}_j\\) with the edge feature \\([\\mathbf{F}_j - \\mathbf{F}_i,\\mathbf{F}_j]\\). This can further enhance the expressive power, as discussed in the ablation study 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Fully Convolutional Network. As shown in Fig. 3, we develop a hierarchical convolution network based on PARE-Conv. The original point clouds \\(\\mathcal{P}\\) and \\(\\mathcal{Q}\\) are downsampled three times to yield sparse superpoints \\(\\hat{\\mathcal{P}}\\) and \\(\\hat{\\mathcal{Q}}\\), whose resolution is \\(\\frac{1}{2^3}\\) of the original point clouds. Consequently, three convolution blocks are leveraged to extract multi-level features. The bottleneck rotation-equivariant features are denoted as \\(\\hat{\\mathbf{F}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\hat{\\mathcal{P}}| \\times \\hat{d} \\times 3}\\) and \\(\\hat{\\mathbf{F}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\hat{\\mathcal{Q}}| \\times \\hat{d} \\times 3}\\), and a VN-invariant layer [8] is used to obtain the rotation-invariant features \\(\\hat{\\mathbf{X}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\hat{\\mathcal{P}}| \\times 3\\hat{d}}\\) and \\(\\hat{\\mathbf{X}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\hat{\\mathcal{Q}}| \\times 3\\hat{d}}\\). For the decoder, two nearest upsampling blocks are used to get the rotation-equivariant features \\(\\tilde{\\mathbf{F}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{P}}| \\times \\tilde{d} \\times 3}\\) and \\(\\tilde{\\mathbf{F}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{Q}}| \\times \\tilde{d} \\times 3}\\) of the first-level downsampled points \\(\\tilde{\\mathcal{P}}\\) and \\(\\tilde{\\mathcal{Q}}\\). Another VN-invariant layer is leveraged to get their rotation-invariant features \\(\\tilde{\\mathbf{X}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{P}}| \\times 3\\tilde{d}}\\) and \\(\\tilde{\\mathbf{X}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{Q}}| \\times 3\\tilde{d}}\\). For more details on the convolutional block, please refer to Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.504 + ], + "angle": 0, + "content": "Following GeoTrans [25], we exploit the point-to-node grouping strategy to assign the points \\(\\hat{\\mathcal{P}}\\) to the superpoints \\(\\hat{\\mathcal{P}}\\). The subset of \\(\\hat{\\mathcal{P}}\\) belongs to the superpoint \\(\\hat{\\mathcal{P}}_i\\) is denoted as \\(\\mathcal{G}_i^{\\mathcal{P}}\\) and the associated rotation-equivariant and invariant feature matrices of subset points are denoted as \\(\\tilde{\\mathbf{F}}_i^{\\mathcal{P}}\\) and \\(\\tilde{\\mathbf{X}}_i^{\\mathcal{P}}\\). For point cloud \\(\\tilde{\\mathcal{Q}}\\), the points are grouped as \\(\\{\\mathcal{G}_i^{\\mathcal{Q}}\\}\\), and the feature matrices are denoted as \\(\\tilde{\\mathbf{F}}_i^{\\mathcal{Q}}\\) and \\(\\tilde{\\mathbf{X}}_i^{\\mathcal{Q}}\\) in a same way." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.527, + 0.469, + 0.543 + ], + "angle": 0, + "content": "3.2 Coarse-to-Fine Matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.784, + 0.598 + ], + "angle": 0, + "content": "Since our backbone outputs both patch-level and point-level features, we use a coarse-to-fine matching strategy [25,34,35] to filter out non-overlapped regions and estimate more accurate correspondences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.673 + ], + "angle": 0, + "content": "Superpoint Matching. Direct matching superpoints by nearest neighbor search lacks robustness to repeated patterns and low-overlapped problems. Therefore, we utilize the Geometric Transformer module [25] presented by Qin et al. to reason the global context of two point clouds, which iteratively uses self and cross-attention to capture intra and inter-point-cloud features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.785, + 0.764 + ], + "angle": 0, + "content": "We feed the rotation-invariant features \\(\\hat{\\mathbf{X}}^{\\mathcal{P}}\\) and \\(\\hat{\\mathbf{X}}^{\\mathcal{Q}}\\) and the attached position information \\(\\hat{\\mathcal{P}}\\) and \\(\\hat{\\mathcal{Q}}\\) into this module to get more discriminative features \\(\\hat{\\mathbf{H}}^{\\mathcal{P}}\\) and \\(\\hat{\\mathbf{H}}^{\\mathcal{Q}}\\). Then, a Gaussian correlation matrix is computed, followed by a dual normalization operation. The superpoint correspondences are obtained by selecting the top-\\(k\\) reliable correspondences \\(\\hat{\\mathcal{C}} = \\{(\\hat{\\mathbf{p}}_{x_i},\\hat{\\mathbf{q}}_{y_i})|(x_i,y_i)\\in \\arg \\max_{x,y}\\mathbf{S}_{x,y}\\}\\), where \\(\\mathbf{S}\\) is the similarity matrix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Improvement. Here, we find this module is sensitive to the overlap distribution of the dataset due to its cross-attention layers. This could lead to a significant performance decrease when there is a substantial difference between the overlap distributions of the test set and the training set. To tackle this issue, we present a simple but effective data augmentation technique, RandomCrop," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where point clouds are randomly cropped to make the model more robust to low-overlapped point cloud pairs. We present more details in the Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.255 + ], + "angle": 0, + "content": "Point Matching. When the correspondences \\(\\hat{\\mathcal{C}}_i = (\\hat{\\mathbf{p}}_{x_i},\\hat{\\mathbf{q}}_{y_i})\\) of superpoints are established, we search point-level correspondences within \\(\\mathcal{G}_{x_i}^{\\mathcal{P}}\\) and \\(\\mathcal{G}_{y_i}^{\\mathcal{Q}}\\). We follow [22] that disentangles the similarity and saliency of features by using a matchability head \\(\\mathbf{W}_m:\\mathbb{R}^{3\\tilde{d}}\\to \\mathbb{R}^{3\\tilde{d}}\\) and a saliency head \\(\\mathbf{W}_s:\\mathbb{R}^{3\\tilde{d}}\\to \\mathbb{R}\\). Specifically, we compute a matching matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.264, + 0.786, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {x _ {i}, y _ {i}} = \\left(\\mathbf {W} _ {m} \\mathbf {X} _ {x _ {i}} ^ {\\mathcal {P}}\\right) ^ {\\mathrm {T}} \\mathbf {W} _ {m} \\mathbf {X} _ {y _ {i}} ^ {\\mathcal {Q}} / \\sqrt {3 \\tilde {d}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.293, + 0.513, + 0.307 + ], + "angle": 0, + "content": "and then the salient scores are predicted:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.315, + 0.786, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {x _ {i}} ^ {\\mathcal {P}} = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {s} \\mathbf {X} _ {x _ {i}} ^ {\\mathcal {P}}\\right), \\sigma_ {y _ {i}} ^ {\\mathcal {Q}} = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {s} \\mathbf {X} _ {y _ {i}} ^ {\\mathcal {Q}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.342, + 0.597, + 0.357 + ], + "angle": 0, + "content": "Finally, a soft assignment matrix \\(\\mathbf{Z}_i\\) is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.365, + 0.785, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {i} = \\sigma_ {x _ {i}} ^ {\\mathcal {P}} \\sigma_ {y _ {i}} ^ {\\mathcal {Q}} \\varphi \\left(\\mathbf {M} _ {x _ {i}, y _ {i}}\\right) _ {x _ {i}} \\varphi \\left(\\mathbf {M} _ {x _ {i}, y _ {i}}\\right) _ {y _ {i}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.784, + 0.437 + ], + "angle": 0, + "content": "where \\(\\varphi\\) is the softmax operation. As a result, we select \\(N_{f}\\) point-level correspondences \\(\\mathcal{C} = \\{(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})\\}\\) across all patch correspondence whose assignment scores are the highest. The impact of \\(N_{f}\\) is analyzed in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.457, + 0.55, + 0.472 + ], + "angle": 0, + "content": "3.3 Feature-based hypothesis Proposer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.571, + 0.524 + ], + "angle": 0, + "content": "Since our rotation-equivariant features encode orientation information of local structures, we derive the pose transformation from matched features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.572, + 0.633 + ], + "angle": 0, + "content": "Hypothesis Generation and Selection. Once obtaining point-wise correspondence set \\(\\mathcal{C} = \\{(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})\\}\\), we leverage their corresponding rotation-equivariant feature maps \\(\\{(\\tilde{\\mathbf{F}}_{x_j}^{\\mathcal{P}},\\tilde{\\mathbf{F}}_{y_j}^{\\mathcal{Q}})\\}\\) to generate multiple model hypotheses. For each correspondence \\((\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})\\), a rotation can be estimated by solving:" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.639, + 0.571, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {R}} _ {j} = \\underset {\\mathbf {R}} {\\arg \\min } \\| \\mathbf {R} \\left(\\tilde {\\mathbf {F}} _ {x _ {j}} ^ {\\mathcal {P}}\\right) ^ {\\mathrm {T}} - \\left(\\tilde {\\mathbf {F}} _ {y _ {j}} ^ {\\mathcal {Q}}\\right) ^ {\\mathrm {T}} \\| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.57, + 0.718 + ], + "angle": 0, + "content": "This can be efficiently solved by SVD [5] in closed form. Subsequently, the translation can be solved by:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.718, + 0.571, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {t}} _ {j} = \\tilde {\\mathbf {q}} _ {y _ {j}} - \\tilde {\\mathbf {R}} _ {j} \\tilde {\\mathbf {p}} _ {x _ {j}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.74, + 0.572, + 0.8 + ], + "angle": 0, + "content": "As a result, we can get a set of model hypotheses \\(\\mathcal{H}_c = \\{(\\tilde{\\mathbf{R}}_i,\\tilde{\\mathbf{t}}_i)\\}\\). The hypothesis that has the most inlier support will be selected as the alignment solution:" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.509, + 0.772, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.586, + 0.771, + 0.596 + ], + "angle": 0, + "content": "Source point cloud Target point cloud" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.602, + 0.757, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.66, + 0.771, + 0.672 + ], + "angle": 0, + "content": "Symmetric alignment Correct alignment" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.684, + 0.785, + 0.793 + ], + "angle": 0, + "content": "Fig.4: An illustration of incorrect alignment caused by the ambiguity of using coordinates to align. This problem can be solved by using rotation-equivariant features that contain fine-grained local orientation information." + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.805, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {R} ^ {*}, \\mathbf {t} ^ {*}\\right) = \\underset {\\tilde {\\mathbf {R}} _ {j}, \\tilde {\\mathbf {t}} _ {j}} {\\arg \\max } \\sum_ {\\left(\\tilde {\\mathbf {p}} _ {x _ {j}}, \\tilde {\\mathbf {q}} _ {y _ {j}}\\right) \\in \\mathcal {C}} \\mathbb {1} \\left(\\| \\tilde {\\mathbf {R}} _ {j} \\tilde {\\mathbf {p}} _ {x _ {j}} + \\tilde {\\mathbf {t}} _ {j} - \\tilde {\\mathbf {q}} _ {y _ {j}} \\| _ {2} ^ {2} < \\tau_ {d}\\right), \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(\\tau_{d}\\) is the acceptance radius. The best model hypothesis will be iteratively refined by its inlier correspondences for a more accurate solution, as in [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Discussion. Traditional estimators, such as RANSAC [11] randomly sample multiple pairs of correspondences and use their coordinates to estimate the transformation. Only when all the sampled correspondences belong to inliers, a reliable transformation can be estimated. This approach is inefficient and requires multiple iterations. In contrast, we directly utilize rotation-equivariant features to estimate the pose of the point cloud, and each pair of correspondences generates a hypothesis. This approach is more efficient, as depicted in the Fig. 2. Our method achieves a significantly high registration recall by generating only 1000 hypotheses. In contrast, RANSAC is still inferior to our method even with 50,000 iterations. Moreover, using coordinates to solve the transformation may falsely align the point clouds with opposite orientations, due to ignoring the orientation information, as shown in Fig. 4. This problem can be easily addressed by our method as the rotation-equivariant features encode orientations of local structures. In addition, some methods [1, 28] use regressors to implicitly estimate transformations from equivariant features. Instead, our method can directly solve pose estimation through feature alignment, which is conceptually simpler and computationally efficient." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.462, + 0.376, + 0.475 + ], + "angle": 0, + "content": "3.4 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.784, + 0.521 + ], + "angle": 0, + "content": "The network is supervised by three losses \\(\\mathcal{L} = \\mathcal{L}_c + \\mathcal{L}_f + \\mathcal{L}_r\\), including superpoint matching loss \\(\\mathcal{L}_c\\), point matching loss \\(\\mathcal{L}_f\\), and contrastive rotation loss \\(\\mathcal{L}_r\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Point Matching Loss \\(\\mathcal{L}_f\\). Since point matching is performed within the scopes of matched superpoints, unreliable superpoint matching may lead to sparse supervision signals for point matching. Therefore, we use positive superpoint correspondences computed by ground-truth transformation during the training process. For two grouped point sets \\(\\mathcal{G}_{x_i}^{\\mathcal{P}}\\) and \\(\\mathcal{G}_{y_i}^{\\mathcal{Q}}\\), the positive correspondences \\(\\mathcal{C}_i^*\\), having distance errors smaller than a threshold \\(d_p\\), are identified by the groundtruth. The unmatched points are denoted as \\(\\mathcal{I}_i\\), \\(\\mathcal{J}_i\\). We compute the log-likelihood loss on soft assignment matrix \\(\\mathbf{Z}_i\\) and the salient scores \\(\\sigma_{x_i}\\), \\(\\sigma_{y_i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.658, + 0.785, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {f _ {i}} = - \\frac {1}{| \\mathcal {C} _ {i} ^ {*} |} \\sum_ {(x _ {i}, y _ {i}) \\in \\mathcal {C} _ {i} ^ {*}} \\log \\mathbf {Z} _ {x _ {i}, y _ {i}} - \\frac {1}{2 | \\mathcal {I} _ {i} |} \\sum_ {x _ {i} \\in \\mathcal {I} _ {i}} \\log \\left(1 - \\sigma_ {x _ {i}} ^ {\\mathcal {P}}\\right) \\\\ - \\frac {1}{2 \\left| \\mathcal {J} _ {i} \\right|} \\sum_ {y _ {i} \\in \\mathcal {J} _ {i}} \\log \\left(1 - \\sigma_ {y _ {i}} ^ {Q}\\right). \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.749, + 0.628, + 0.766 + ], + "angle": 0, + "content": "The overall point matching loss is a sum of \\(\\mathcal{L}_f = \\sum_i\\mathcal{L}_{f_i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Contrastive Rotation Loss \\(\\mathcal{L}_r\\). Rotation-equivariant feature inherently encodes pose information of local geometric structure. However, such features of a positive correspondence inevitably be contaminated by data degradation, such as occlusion, noise and down-sampling errors. To make the features more robust to these troubles, we propose a contrastive rotation loss, to supervise point-level" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.341, + 0.162 + ], + "angle": 0, + "content": "correspondences:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.176, + 0.785, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {r _ {i}} = \\frac {1}{| \\mathcal {C} _ {i} ^ {*} | \\cdot \\tilde {d}} \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in \\mathcal {C} _ {i} ^ {*}} \\sum_ {c = 1} ^ {\\tilde {d}} [ \\| \\tilde {\\mathbf {F}} _ {x _ {i}, c} ^ {\\mathcal {P}} \\mathbf {R} _ {g t} ^ {\\mathrm {T}} - \\tilde {\\mathbf {F}} _ {y _ {i}, c} ^ {\\mathcal {Q}} \\| _ {2} ^ {2} - \\alpha ] _ {+} \\tag {11} \\\\ + \\frac {1}{| \\bar {\\mathcal {C}} _ {i} | \\cdot \\tilde {d}} \\sum_ {(x _ {i}, y _ {i}) \\in \\bar {\\mathcal {C}} _ {i}} \\sum_ {c = 1} ^ {\\tilde {d}} [ \\beta - \\| \\tilde {\\mathbf {F}} _ {x _ {i}, c} ^ {\\mathcal {P}} \\mathbf {R} _ {g t} ^ {\\mathrm {T}} - \\tilde {\\mathbf {F}} _ {y _ {i}, c} ^ {\\mathcal {Q}} \\| _ {2} ^ {2} ] _ {+}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.786, + 0.328 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_i^*\\) is the aforementioned positive correspondence set, \\(\\bar{\\mathcal{C}}_i\\) is the negative correspondence set whose point pair has spatial distance greater than \\(d_{n}\\), \\(\\alpha = 0.1\\) and \\(\\beta = 1.4\\) are positive and negative margins, and \\([\\cdot]_{+}\\) is the ReLU operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.39 + ], + "angle": 0, + "content": "Superpoint Matching Loss \\(\\mathcal{L}_c\\). To supervise the rotation-invariant features of superpoints, we utilize the overlap-aware circle loss that considers the overlap ratio of two superpoints to compute more accurate gradients. For more details, please refer to GeoTrans [25]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.376, + 0.43 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.537 + ], + "angle": 0, + "content": "To evaluate the performance of our method, we compare our method with state-of-the-art approaches and perform extensive experiments on 3DMatch [37], 3DLoMatch [15], and KITTI Odometry [12]. Finally, ablation studies are conducted to analyze the components of our method. All the experiments are conducted with a RTX 3090 GPU with Intel (R) Xeon (R) Silver 4314 CPU. More implementation details are reported in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.6, + 0.573 + ], + "angle": 0, + "content": "4.1 Indoor Dataset: 3DMatch & 3DLoMatch" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.786, + 0.674 + ], + "angle": 0, + "content": "DataSet. 3DMatch [37] and 3DLoMatch [15] are collected from indoor RGB-D datasets. 3DMatch contains 62 scenes and there are \\(46/8/8\\) scenes for training/validation/testing. It consists of point cloud pairs with overlap ratios greater than \\(30\\%\\). 3DLoMatch is only used for testing that contains point cloud pairs with overlap ratios from \\(10\\%\\) to \\(30\\%\\). We follow the protocols [15, 25, 28] to perform all the experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Baseline. We compare our method with strong baselines: 1) FCGF [7], Predator [15], GeoTrans [25], and PEAL [36] are advanced scene-level feature extractors, and some of them [25,36] follow a coarse-to-fine manner to extract robust correspondences. PEAL requires overlap prior as input, we follow the office implementation that leverages 3D overlap prior predicted by GeoTrans for a fair comparison. 2) YOHO [27], RoReg [28], and RoITR [35] are also scene-level feature extractors but focus on rotation-invariant/equivariant network designs. 3) SpinNet [2] and BUFFER [1] leverage patch-level feature extractors with elaborate designs for rotation-invariance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Metrics. We follow [15,25] that use three evaluation metrics, including Inlier Ratio (IR), Feature Matching Recall (FMR), and Registration Recall (RR). To" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Evaluation results on 3DMatch and 3DLoMatch. Methods with the RANSAC estimator are marked by \\(\\diamond\\), which exploit 5000 points to establish correspondences." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.186, + 0.787, + 0.34 + ], + "angle": 0, + "content": "
MethodSize (MB)3DMatch3DLoMatch
FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)
FCGF° [7]8.7694.731.182.80.1259.49.838.00.13
SpinNet° [2]1.4197.647.588.69.8575.320.559.89.03
YOHO [27]12.3898.264.490.82.8178.925.966.02.62
Predator° [15]7.4396.658.089.00.6478.226.764.40.47
GeoTrans [25]9.8398.170.992.40.1887.443.574.30.17
RoReg [28]12.7198.281.693.02.2782.339.670.12.10
BUFFER [1]0.92--92.10.22--70.30.21
RoITR° [35]10.1098.082.491.90.3689.254.674.10.34
PEAL [36]9.8398.471.094.21.4688.346.078.81.19
Ours3.8498.576.995.00.1788.347.580.50.17
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.37, + 0.785, + 0.46 + ], + "angle": 0, + "content": "evaluate the performance of different methods on rotated datasets, RoITR [35] uses two ways to compute the RMSE on the original and rotated datasets, which cannot show the robustness of methods against pose variations. To tackle this issue, we follow [28] to report Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) on both original and rotated datasets. Please refer to the Appendix E for more details about these evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.462, + 0.786, + 0.763 + ], + "angle": 0, + "content": "Quantitative results. We report the experimental results in Table 1. Compared with SOTA methods, our approach achieves a significantly high RR on both 3DMatch and 3DLoMatch, while simultaneously maintaining lightweight and computation efficient. PEAL [36] is a formidable competitor that employs an iterative manner for alignment due to the requirements of overlap prior, resulting in significant time overhead. In contrast, our method achieves a speed improvement of 7 times compared to PEAL, while also outperforming it by \\(1.7\\%\\) in terms of RR on 3DLoMatch. GeoTrans [25], RoITR [35], and our method all employ a coarse-to-fine matching framework. Despite using a more lightweight feature extractor, our RR is significantly higher than theirs, particularly on 3DLoMatch, where our method outperforms them by \\(6.2\\%\\) and \\(6.4\\%\\) respectively. This can be attributed to the ability of our rotation-equivariant network to focus more on learning discriminative structural information. YOHO [27] and RoReg [28] are extremely time-consuming due to the group feature extractor and embedder. Our efficiency advantage stems from two aspects. Firstly, we utilize a lightweight scene-wise feature extractor, which only needs a forward process to extract dense rotation-equivariant features. Secondly, our feature-based hypothesis proposer can simultaneously generate multiple valid hypotheses by using matched rotation-equivariant features. Thus, it is more efficient compared to RANSAC [11] with many approaches [2, 7, 15, 35] used." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Experiments on the rotated dataset. To test the robustness against rotations, we follow previous works [28, 35] that evaluate all the methods on the rotated 3DLoMatch, where arbitrary rotations are applied to point clouds. The experimental results are shown in Table 2. As we can see, PEAL experiences a decrease in TR by \\(2.5\\%\\) on the Rotated 3DLoMatch dataset, while our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Evaluation results on 3DLoMatch and Rotated 3DLoMatch. For easy comparison of the impact of rotation, the numerical changes w.r.t. TR on the rotated dataset are annotated in the top right corner of each value. Methods with rotation invariant/equivariant designs are marked by * symbol." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.787, + 0.357 + ], + "angle": 0, + "content": "
MethodSize (MB)3DLoMatchRotated 3DLoMatch
RE (°↓)TE (cm ↓)TR (% ↑)RE (°↓)TE (cm ↓)TR (% ↑)
FCGF [7]8.764.8412.8739.64.7413.3924.5-15.1
PREDATOR [15]7.433.6110.6565.63.5510.3064.0-1.6
GeoTrans [25]9.832.918.7175.42.948.8572.6-2.8
PEAL [36]9.832.848.6481.22.868.5378.7-2.5
YOHO* [27]12.383.5410.3466.63.6110.1667.1+0.5
RoReg* [28]12.713.019.2671.33.039.2871.0-0.3
BUFFER* [1]0.923.039.8674.43.029.9974.7+0.3
RoITR* [35]10.102.959.0375.12.979.0875.5+0.4
Ours*3.842.878.8381.32.848.7181.8+0.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.573 + ], + "angle": 0, + "content": "method demonstrates an increase of \\(0.5\\%\\). This result validates the robustness of our approach against rotations, illustrating its ability to handle such variations. Methods without rotation-invariant design, such as FCGF [7], Predator [15], and GeoTrans [25] also show performance degradation, indicating that networks trained with rotation augmentation are vulnerable to rotation variations. On the other hand, methods such as YOHO [27], RoReg [28], RoITR [35], and BUFFER [1], which incorporate rotation equivariant or invariant designs, demonstrate better robustness against data rotations. However, our method significantly outperforms these methods by \\(6\\%\\) w.r.t TR on both 3DLoMatch and Rotated 3DLoMatch. This is because our backbone can capture spatial geometric information better, thus extracting more distinctive descriptors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.432, + 0.794 + ], + "angle": 0, + "content": "Qualitative results. We show some registration results in Fig. 5, which demonstrate that our method can estimate more accurate patch-level and point-level correspondences and can successfully align point cloud pairs with low-overlapped ratios. Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method suc" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "cessfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity. Please refer to Appendix G for more visualization results." + }, + { + "type": "table_caption", + "bbox": [ + 0.447, + 0.628, + 0.782, + 0.641 + ], + "angle": 0, + "content": "Table 3: Evaluation results on KITTI Odometry." + }, + { + "type": "table", + "bbox": [ + 0.447, + 0.642, + 0.787, + 0.769 + ], + "angle": 0, + "content": "
MethodSizeRE (°)TE (cm)TR (%)Time (s)
FCGF [7]8.760.309.596.6-
D3Feat [4]14.080.307.299.8-
Predator [15]22.770.276.899.80.77
SpinNet [2]1.410.479.999.116.24
CoFiNet [34]5.480.418.299.80.59
GeoTrans [25]25.500.236.299.80.26
BUFFER [1]0.920.267.199.80.27
Ours2.080.234.999.80.21
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results. Our method can successfully align low-overlapped pairs with higher IR and PIR (Patch IR). Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method successfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.557, + 0.429 + ], + "angle": 0, + "content": "4.2 Outdoor Dataset: KITTI Odometry" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.784, + 0.485 + ], + "angle": 0, + "content": "Dataset. We follow previous methods [7,25] to evaluate our method on KITTI Odometry [12]. It comprises 11 urban scene sequences captured by a LiDAR sensor. We use 0-5/6-7/8-10 sequences for training/validation/testing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.785, + 0.531 + ], + "angle": 0, + "content": "Metrics. We use the Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) to evaluate our method. Following [7,25], the TR is computed with rotation threshold \\( 5^{\\circ} \\) and translation threshold \\( 2\\mathrm{m} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.785, + 0.591 + ], + "angle": 0, + "content": "Experimental results. The experimental results are shown in Table 3. On KITTI dataset, many methods have reached performance saturation, while our approach achieves the highest accuracy when achieving the highest TR. Our method also shows advantages w.r.t. model size and running speed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.388, + 0.631 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.784, + 0.673 + ], + "angle": 0, + "content": "To better understand the designs of our method, we conduct comprehensive ablation experiments and the evaluated results are reported in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Ablation of backbone. To investigate the impact of different convolution methods, we compare our PARE-Conv with VN [8]. We leverage both node features [24] and edge features [29] for evaluation. The experimental results indicate that our approach achieves significant improvement compared to VN whether using node features or edge features. This demonstrates the enhanced representative power of PARE-Conv to learn more distinctive features. We also compare PARE-Conv with KPCnv [26], refer to the Appendix F.5 for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Ablation of pose estimator. To demonstrate the superiority of our pose estimator, we compared it with RANSAC [11] and LGR [25]. RANSAC randomly selects a few pairs of correspondences to generate hypotheses, while LGR generates a hypothesis using the correspondences within a matched patch pair." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "In contrast, our approach directly generates a hypothesis using the equivariant features of a pair of matched points. As shown in Table 4, our method achieves the highest RR, outperforming RANSAC and LGR by \\(3\\%\\) and \\(1.6\\%\\) respectively on 3DLoMatch. This validates that estimating poses using rotation-equivariant features based on local structures is more robust than using point-to-point correspondences. Additionally, the results of FMR and IR are identical as we used the same correspondences for different pose estimators." + }, + { + "type": "table_caption", + "bbox": [ + 0.344, + 0.28, + 0.658, + 0.294 + ], + "angle": 0, + "content": "Table 4: Ablation experiments of our method." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.306, + 0.789, + 0.466 + ], + "angle": 0, + "content": "
Component Method3DMatch3DLoMatch
FMR (% ↑) IR (% ↑) RR (% ↑)FMR (% ↑) IR (% ↑) RR (% ↑)
BackboneVN (Node) [8]98.671.093.487.841.476.6
PARE-Conv (Node)98.974.194.5 +1.188.344.078.8 +2.2
VN (Edge) [8]98.772.494.188.843.577.8
PARE-Conv (Edge)98.576.995.0 +0.988.147.580.5 +2.7
Pose EstimatorRANSAC98.576.993.788.347.577.5
LGR [25]98.576.994.0 +0.388.347.578.9 +1.4
Ours98.576.995.0 +1.388.347.580.5 +3.0
Rotation LossWithout98.374.493.388.745.579.3
L298.665.494.2 +0.985.935.075.4 -3.9
Contrastive98.576.995.0 +1.788.347.580.5 +1.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.661 + ], + "angle": 0, + "content": "Ablation of rotation loss. The contrastive rotation loss is designed to improve the robustness of rotation-equivariant features to data occlusion, noise, and partial overlap. When the rotation-equivariant features are not supervised by a loss function (Indeed, it is indirectly supervised by the loss function of rotation-invariant features, as the invariant features are derived from the equivariant features), the method shows decent performance on 3DLoMatch but shows poor performance on 3DMatch. When using the L2 loss to supervise equivariant features of positive point pairs, there is a slight improvement in RR on 3DMatch but shows performance degradation on 3DLoMatch. However, when utilizing our contrastive rotation loss, the performance of the method is enhanced on both 3DMatch and 3DLoMatch, demonstrating its superiority." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.36, + 0.701 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose PARE-Net, a lightweight network for fast and robust point cloud registration. We introduce position-aware rotation-equivariant convolution to learn spatial information effectively, thus extracting more distinctive descriptors. We also propose an efficient feature-based hypothesis proposer to generate reliable model hypotheses. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art approaches and shows great robustness to rotation variations. The limitations of our method and future works are discussed in Appendix H." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Science and Technology Major Project of China under Grant No. 2020AAA0108102, the National Natural Science Foundation of China under Grant Nos. 62327808, 62073257 and 62088102, and the Fundamental Research Funds for the Central Universities under Grant No. xzy022024007." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.248, + 0.323, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.786, + 0.337 + ], + "angle": 0, + "content": "1. Ao, S., Hu, Q., Wang, H., Xu, K., Guo, Y.: BUFFER: Balancing accuracy, efficiency, and generalizability in point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1255-1264 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "2. Ao, S., Hu, Q., Yang, B., Markham, A., Guo, Y.: SpinNet: Learning a general surface descriptor for 3D point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11753-11762 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.786, + 0.434 + ], + "angle": 0, + "content": "3. Bai, X., Luo, Z., Zhou, L., Chen, H., Li, L., Hu, Z., Fu, H., Tai, C.L.: PointDSC: Robust point cloud registration using deep spatial consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15859-15869 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.436, + 0.786, + 0.49 + ], + "angle": 0, + "content": "4. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3Feat: Joint learning of dense detection and description of 3D local features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6359-6367 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.786, + 0.519 + ], + "angle": 0, + "content": "5. Besl, P.J., McKay, N.D.: Method for registration of 3-D shapes. In: Sensor Fusion IV: Control Paradigms and Data Structures. vol. 1611, pp. 586-606 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.52, + 0.786, + 0.574 + ], + "angle": 0, + "content": "6. Chen, Z., Sun, K., Yang, F., Tao, W.: SC²-PCR: A second order spatial compatibility for efficient and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13221-13231 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.786, + 0.616 + ], + "angle": 0, + "content": "7. Choy, C., Park, J., Koltun, V.: Fully convolutional geometric features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8958-8966 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.786, + 0.672 + ], + "angle": 0, + "content": "8. Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., Guibas, L.J.: Vector Neurons: A general framework for SO(3)-equivariant networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12200-12209 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.786, + 0.714 + ], + "angle": 0, + "content": "9. Deng, H., Birdal, T., Ilic, S.: PPF-FoldNet: Unsupervised learning of rotation invariant 3D local descriptors. In: Proceedings of the European Conference on Computer Vision. pp. 602-618 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.715, + 0.786, + 0.757 + ], + "angle": 0, + "content": "10. Deng, H., Birdal, T., Ilic, S.: PPFNet: Global context aware local features for robust 3D point matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 195-205 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.798 + ], + "angle": 0, + "content": "11. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. vol. 24, pp. 381-395. ACM New York (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.281, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Yao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Gojcic, Z., Zhou, C., Wegner, J.D., Wieser, A.: The perfect match: 3D point cloud matching with smoothed densities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5545-5554 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "14. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3D local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.274 + ], + "angle": 0, + "content": "15. Huang, S., Gojcic, Z., Usvyatsov, M., Wieser, A., Schindler, K.: PREDATOR: Registration of 3D point clouds with low overlap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4267-4276 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.786, + 0.303 + ], + "angle": 0, + "content": "16. Kaya, M., Bilge, H.S.: Deep metric learning: A survey. Symmetry 11(9), 1066-1092 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.786, + 0.346 + ], + "angle": 0, + "content": "17. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 153-161 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.786, + 0.389 + ], + "angle": 0, + "content": "18. Lee, J., Kim, S., Cho, M., Park, J.: Deep hough voting for robust global registration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15994-16003 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.786, + 0.431 + ], + "angle": 0, + "content": "19. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.786, + 0.473 + ], + "angle": 0, + "content": "20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems 32(8), 3412-3432 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.786, + 0.516 + ], + "angle": 0, + "content": "21. Lu, W., Zhou, Y., Wan, G., Hou, S., Song, S.: L3-Net: Towards learning based LiDAR localization for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6389-6398 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.786, + 0.557 + ], + "angle": 0, + "content": "22. Philipp, L., Paul-Edouard, S., Marc, P.: LightGlue: Local Feature Matching at Light Speed. In: Proceedings of the IEEE International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.786, + 0.587 + ], + "angle": 0, + "content": "23. Pomerleau, F., Colas, F., Siegwart, R.: A review of point cloud registration algorithms for mobile robotics. Foundations and Trends in Robotics 4(1), 1-104 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.786, + 0.629 + ], + "angle": 0, + "content": "24. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.786, + 0.685 + ], + "angle": 0, + "content": "25. Qin, Z., Yu, H., Wang, C., Guo, Y., Peng, Y., Xu, K.: Geometric transformer for fast and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11143-11152 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.786, + 0.741 + ], + "angle": 0, + "content": "26. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPConv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6411-6420 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.786, + 0.784 + ], + "angle": 0, + "content": "27. Wang, H., Liu, Y., Dong, Z., Wang, W.: You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1630-1641 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "28. Wang, H., Liu, Y., Hu, Q., Wang, B., Chen, J., Dong, Z., Guo, Y., Wang, W., Yang, B.: RoReg: Pairwise point cloud registration with oriented descriptors and local rotations. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 10376-10393 (2023). https://doi.org/10.1109/TPAMI.2023.3244951" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PARE-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph CNN for learning on point clouds. ACM Transactions On Graphics (ToG) 38(5), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "30. Xu, M., Ding, R., Zhao, H., Qi, X.: PACConv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3173-3182 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "31. Yang, B., Bender, G., Le, Q.V., Ngiam, J.: CondConv: Conditionally parameterized convolutions for efficient inference. In: Proceedings of Advances in Neural Information Processing Systems. vol. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "32. Yao, R., Du, S., Cui, W., Ye, A., Wen, F., Zhang, H., Tian, Z., Gao, Y.: Hunter: Exploring high-order consistency for point cloud registration with severe outliers. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14760-14776 (2023). https://doi.org/10.1109/TPAMI.2023.3312592" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "33. Yew, Z.J., Lee, G.H.: REGTR: End-to-end point cloud correspondences with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6677-6686 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "34. Yu, H., Li, F., Saleh, M., Busam, B., Ilic, S.: CoFiNet: Reliable coarse-to-fine correspondences for robust point cloud registration. In: Proceedings of Advances in Neural Information Processing Systems. pp. 23872-23884 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "35. Yu, H., Qin, Z., Hou, J., Saleh, M., Li, D., Busam, B., Ilic, S.: Rotation-invariant transformer for point cloud matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5384-5393 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.787, + 0.507 + ], + "angle": 0, + "content": "36. Yu, J., Ren, L., Zhang, Y., Zhou, W., Lin, L., Dai, G.: PEAL: Prior-embedded explicit attention learning for low-overlap point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17702-17711 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.563 + ], + "angle": 0, + "content": "37. Zeng, A., Song, S., Nießner, M., Fisher, M., Xiao, J., Funkhouser, T.: 3DMatch: Learning local geometric descriptors from RGB-D reconstructions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1802-1811 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.591 + ], + "angle": 0, + "content": "38. Zhang, Z., Dai, Y., Sun, J.: Deep learning based point cloud registration: An overview. Virtual Reality and Intelligent Hardware 2(3), 222-246 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_origin.pdf b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..650827f0dd2d15eb0651a8850b871f02691fc2c5 --- /dev/null +++ b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70770b58102031cb63fa3829e94c9c0e29adb786e553b2ec574807eb3fbee310 +size 3108195 diff --git a/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/full.md b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6446d8282d281a91f6414cfcd518895e228b56fa --- /dev/null +++ b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/full.md @@ -0,0 +1,294 @@ +# PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration + +Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang + +National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, China + +Abstract. Learning rotation-invariant distinctive features is a fundamental requirement for point cloud registration. Existing methods often use rotation-sensitive networks to extract features, while employing rotation augmentation to learn an approximate invariant mapping rudely. This makes networks fragile to rotations, overweight, and hinders the distinctiveness of features. To tackle these problems, we propose a novel position-aware rotation-equivariant network, for efficient, light-weighted, and robust registration. The network can provide a strong model inductive bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned limitations. To further improve the distinctiveness of descriptors, we propose a position-aware convolution, which can better learn spatial information of local structures. Moreover, we also propose a feature-based hypothesis proposer. It leverages rotation-equivariant features that encode fine-grained structure orientations to generate reliable model hypotheses. Each correspondence can generate a hypothesis, thus it is more efficient than classic estimators that require multiple reliable correspondences. Accordingly, a contrastive rotation loss is presented to enhance the robustness of rotation-equivariant features against data degradation. Extensive experiments on indoor and outdoor datasets demonstrate that our method significantly outperforms the SOTA methods in terms of registration recall while being lightweight and keeping a fast speed. Moreover, experiments on rotated datasets demonstrate its robustness against rotation variations. Code is available at https://github.com/yaorz97/PARENet. + +Keywords: Point cloud registration $\cdot$ Rotation invariance/equivalence $\cdot$ Robust transformation estimation + +# 1 Introduction + +Point cloud registration, estimating a rigid transformation to align two point clouds that are partially overlapped, is fundamental research in 3D computer vision. It is widely applied in autonomous driving [20,21], robot localization [23], and 3D reconstruction [38]. Feature-based point cloud registration framework is widely studied, which mainly involves descriptor extraction [4, 15, 25, 34] and + +robust transformation estimation [3,6,32]. One inherent contradiction in point cloud registration is the pose variation of point clouds and the desired invariance of their descriptors. To extract such rotation-invariant descriptors, plenty of approaches have been proposed that can be roughly categorized into patch-wise and scene-wise feature extractors. + +![](images/980f9b4fc9a18938bb508e16c4db13c257158c71adcea63ce949e1c50cf98bc1.jpg) +Fig. 1: Experimental results on 3DMatch and 3DLoMatch. Our method significantly outperforms state-of-the-art methods w.r.t. registration recall (RR) while maintaining fast speed and lightweight. + +![](images/1891bf84bac650bb834c2d014ffc40dd64adc48fdd6e9f5c59347f3fcf1ca2dd.jpg) + +Patch-wise feature extractors [2, 9, 10] learn geometric features from local patches of point clouds and typically employ data preprocessing techniques to achieve rotation-invariance. Two common approaches include: 1) extracting rotation invariant statistics as initial representation from local patches, such as distance, angle, and point pair features (PPF [9, 10]); and 2) estimating a local reference frame (LRF [2]) to normalize poses of local patches. The former approach suffers from information loss, while the latter approach is susceptible to data occlusion and noise, lacking robustness. Moreover, these extractors are generally computationally expensive because all patches of a point cloud are processed individually, without sharing intermediate representations [7]. + +Scene-wise feature extractors [4,7,15,25,34] learn dense feature descriptors for every point in the point cloud, and most of them employ extensive rotation augmentation in the training process to rudely enforce inherently rotation-sensitive networks (e.g., KPCov [26]) to learn rotation-invariant mappings. These methods have three shortcomings. Firstly, models are fragile to unseen rotations because it is difficult to cover the SO(3) space in the training process [35]. Secondly, learning rotation-invariant mappings requires more network parameters. Thirdly, these networks are limited to generating discriminative feature descriptors as they cannot focus solely on learning structural distinctiveness. + +In this paper, we propose a novel registration network, PARE-Net, whose core is to fully exploit the advantages of rotation-equivariant networks in both feature extraction and transformation estimation, thereby achieving a lightweight, efficient, and robust registration method, as shown in Fig. 1. + +First of all, we propose a Position-Aware Rotation-Equivariant Convolution (PARE-Conv) that enhances the representational power of networks for better distinctive feature learning. Vector Neuron (VN) [8] builds a general rotation-equivariant network framework. However, its representative power is limited as it ignores the importance of spatial localization property of local points and utilizes multi-layer perceptrons (MLPs) to crudely learn spatial encodings. In + +stead, PARE-Conv is proposed to explicitly leverage the spatial information of point clouds while guaranteeing the rotation equivariant property. Specifically, we first define a set of shadow kernel points equipped with learnable weights. Then, we introduce a correlation network that predicts rotation-invariant correlations between points and shadow kernels according to rotation-equivariant spatial features of points. Therefore, the geometric structures of point clouds are well captured. Finally, the correlation scores are subsequently leveraged to dynamically assemble kernel weights for rotation-equivariant convolution. As a result, PARE-Conv inherently provides a strong model bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned problems raised by rotation-sensitive networks. Moreover, it can learn the spatial structure of point clouds better, facilitating to extraction of more distinctive descriptors. + +Secondly, we propose a feature-based hypothesis proposer that can efficiently generate reliable solution hypotheses. Once obtaining correspondences by feature matching, rotation-equivariant features encoding fine-grained structure information are used to explicitly generate multiple hypotheses. This way can alleviate the orientation ambiguity stemming from optimizing distance errors of correspondences. Moreover, to make the equivariant features more robust against data degradation, we present a contrastive rotation loss inspired by metric learning [7, 16], facilitating more robust pose estimation. As shown in Fig. 2, our method can achieve significantly higher registration + +recall compared to RANSAC [11] and LGR (presented in GeoTrans [25]) when varying the number of hypotheses, demonstrating our method is more robust than classical estimators. Because they require multiple correspondences and the probability of sampling all correct correspondences decreases rapidly as the number of used correspondences increases. Overall, our main contributions are: + +![](images/dfa6aaa7c60440553160fd36a304b8e1660a41206ad6af8814a41a69eee1d1f0.jpg) +Fig.2: Comparison of our feature-based hypothesis proposer, RANSAC and LGR on 3DLoMatch. + +- A lightweight, efficient, and robust point cloud registration method that is achieved by fully exploiting the advantages of rotation-equivariant networks in feature extraction and pose estimation. +- A position-aware rotation-equivariant convolution that can better leverage spatial information, enabling the extraction of discriminative descriptors. +- A feature-based hypothesis proposer that can efficiently generate multiple reliable hypotheses and a contrastive rotation loss makes the rotation-equivariant features more robust to data degradation. + +# 2 Related Work + +Patch-wise feature extractors. In the early stage, most learning-based feature descriptors [2, 9, 10, 13, 17, 37] were patch-wise extractors, primarily due to limitations in point cloud convolution techniques, which could not handle + +large-scale point clouds effectively. During this period, the works drew inspiration from traditional feature descriptors [14] and focused on designing features with rotation-invariance. PPFNet [10] and PPFoldNet [9] extracted rotation-invariant PPFs as initial features and further learned high-dimensional features. PerfectMatch [13] and SpinNet [2] utilized LRF to normalize the pose of local patches and then employed networks to learn the density/structural features of patches. These methods are primarily limited by their computational efficiency. + +Scene-wise feature extractors [4, 7, 15, 19, 25, 33-36] have become the mainstream research in recent years, due to their high efficiency. These methods mainly achieved rotation invariance through data augmentation and aim to tackle the low-overlapped problem. Predator [15] designed an overlap-attention module to detect overlap regions between two point clouds. CoFiNet [34] and GeoTrans [25] presented a coarse-to-fine matching scheme, and GeoTrans introduced a geometric transformer module that significantly increased the matching accuracy by injecting rotation-invariant position information. PEAL [36] presented an explicit one-way attention module that iteratively leveraged overlap prior to relieving feature ambiguity while bringing expensive computation costs. Moreover, some methods with rotation invariant/equivariant designs were proposed. YOHO [27] and RoReg [28] exploited the group convolution technique to extract rotation-invariant/equivariant group features, which are extremely computation expensive. Because they need to extract point features under 60 different poses and perform convolutions on massive point features. RoITR [35] used PPF to design an invariant convolution and presented a rotation invariant transformer to extract distinctive features. BUFFER [1] combined scene-wise and patch-wise feature extractors to achieve faster alignment with great generalization ability. It first leveraged a rotation-equivariant network built on vanilla VN to extract invariant and equivariant features for keypoints and orientations prediction. Then, it also used SpinNet to extract patch-wise features due to the limited representative ability of vanilla VN. + +Robust transformation estimators. Since the estimated correspondences are usually contaminated by outliers, a robust transformation estimator is necessary. RANSAC [11] is the most widely used method due to its simplicity, but it typically exhibits lower accuracy and requires plenty of iterations for convergence. Recently, some learning-based methods [3,6,18,32] are proposed, whose core idea is to design networks to learn spatial consistency between correspondences, enabling efficient inlier identification and reliable hypothesis generation. We compare our approach with these methods in Appendix F.1. + +# 3 Method + +Given two partially overlapped point clouds $\mathcal{P} = \{\mathbf{p}_i\in \mathbb{R}^3 |i = 1,2,\dots ,N\}$ and $\mathcal{Q} = \{\mathbf{q}_j\in \mathbb{R}^3 |j = 1,2,\dots ,M\}$ , we aim to estimate a transformation $\mathbf{T} = (\mathbf{R}\in SO(3),\mathbf{t}\in \mathbb{R}^3)$ that aligns the source point cloud to the target one. To address this problem, we propose a registration framework based on rotation-equivariant network $f(\mathbf{X})$ , whose output is equivariant to the SO(3) action $\mathbf{R}$ + +![](images/9666065ae66c08d2e36b35726224420c6174f7e3a96a14e18983c6af1aa633f8.jpg) +Fig. 3: The framework of our method. Given point clouds $\mathcal{P}$ and $\mathcal{Q}$ , a hierarchical backbone based on PARE-Conv is used to extract patch-level and point-level features. Then, we follow a coarse-to-fine scheme to obtain point correspondences using rotation-invariant features. Finally, the feature-based hypothesis proposer leverages the matched rotation-equivariant features $\tilde{\mathbf{F}}_{x_{jn}}^{\mathcal{P}}$ and $\tilde{\mathbf{F}}_{y_{jn}}^{\mathcal{Q}}$ , which encode fine-grained structure orientations, to generate multiple reliable hypotheses. The best hypothesis is selected and refined as the final solution $(\mathbf{R}^{*},\mathbf{t}^{*})$ . + +applied on the input point cloud $\mathbf{X}$ , i.e., $f(\mathbf{R} \circ \mathbf{X}) = \mathbf{R} \circ f(\mathbf{X})$ . As we can see, rotation-equivariant features $f(\mathbf{R} \circ \mathbf{X})$ couple the structure information $f(\mathbf{X})$ and orientation information $\mathbf{R}$ of point clouds, which can be further decoupled for feature matching and transformation estimation, respectively. + +The framework of PARE-Net is depicted in Fig. 3. We build a hierarchical backbone based on PARE-Conv to extract both patch-level and point-level rotation invariant/equivariant features. Correspondences are estimated in a coarse-to-fine fashion by using rotation-invariant features. After that, the feature-based hypothesis proposer employs rotation-equivariant features to generate multiple reliable hypotheses, and the best one is selected as the final output $(\mathbf{R}^{*},\mathbf{t}^{*})$ . + +# 3.1 Position-Aware Rotation-Equivariant Network + +PARE-Conv. Recent point cloud registration approaches [15,25,34,36] mainly exploited KPCov [26] to build backbones. However, KPCov is inherently rotation-sensitive and needs data augmentation to learn an approximate invariant mapping. As a result, models trained on limited samples of continuous SO(3) space are fragile to rotation variations. Moreover, enforcing the network to learn a rotation-invariant mapping inevitably requires more model parameters and also hinders the distinctiveness of features. Inspired by dynamic and conditional convolutions [30,31], we propose position-aware rotation-equivariant convolution to learn distinctive rotation-invariant/equivariant features, as shown in Fig. 3. + +Given a point cloud $\mathcal{P} = \{\mathbf{p}_i\in \mathbb{R}^3|i = 1,2,\dots ,N\}$ and attached feature map $\mathbf{F} = \{\mathbf{F}_i\in \mathbb{R}^{C\times 3}\}$ , the general convolution of $\mathbf{F}$ by a kernel $g$ at a point $\mathbf{p}_i$ is defined as: + +$$ +\left(\mathbf {F} * g\right) \left(\mathbf {p} _ {i}\right) = \sum_ {\mathbf {p} _ {j} \in \mathcal {N} _ {i}} g \left(\mathbf {p} _ {j} - \mathbf {p} _ {i}\right) \mathbf {F} _ {j}, \tag {1} +$$ + +where $\mathcal{N}_i$ is the K-nearest neighbor set of the point $\mathbf{p}_i$ . The core problem is the definition of kernel function $g$ . In vanilla VN [8], it omits the position information and simply uses a linear layer, $\mathbf{W}\mathbf{F}_i$ , followed by a VN non-linear layer and a VN pooling layer to aggregate local information, where $\mathbf{W} \in \mathbb{R}^{C' \times C}$ is a learnable weight matrix. Instead, we define a set of shadow kernel points, equipped with learnable weight matrices $\{\mathbf{W}_k | k = 1, 2, \dots, K\}$ . In image convolution, each pixel corresponds to a convolutional kernel, and the fixed spatial arrangement of the kernels enables the learning of the distribution information of image pixels. Due to the irregularity of point clouds, establishing such point-to-kernel correspondences is intractable. We follow KPConv [26] to establish soft assignments between data points and kernel points and the kernel function $g$ is defined as: + +$$ +g \left(\mathbf {p} _ {i j}\right) = \sum_ {k = 1} ^ {K} \operatorname {s o f t m a x} (\gamma \left(\mathbf {p} _ {i j}, k\right)) \mathbf {W} _ {k}, \tag {2} +$$ + +where $\mathbf{p}_{ij} = \mathbf{p}_j - \mathbf{p}_i$ is the relative coordinate of $\mathbf{p}_j$ and $\gamma (\cdot)$ is a correlation function. In KPConv [26], it uses a linear correlation $\gamma (\mathbf{p}_{ij},k) = \max (0,1-\|\mathbf{p}_{ij} - \tilde{\mathbf{x}}_k\| / \sigma)$ , where $\tilde{\mathbf{x}}_k$ is the 3D coordinate of the $k$ -th kernel point and $\sigma$ is a distance influence factor. However, kernel-to-point distances are sensitive to rotations thus cannot guarantee a rotation-equivariant convolution. Therefore, the key challenge lies in ensuring the rotation equivariance of the convolution while well capturing the spatial information of patches. Inspired by PACConv [30], we use a mini-network that consumes rotation-equivariant spatial features and predicts rotation-invariant correlation scores between data points and kernel points. Note that predicting rotation-invariant scores is necessary to guarantee the rotation-equivariance of the convolution (see Appendix A for the proof). As shown in Fig. 3, we implement this function with 1) a VN-MLP block, learning geometric features $\mathbf{X}_j \in \mathbb{R}^{d \times 3}$ from rotation-equivariant spatial statistics, such as $\mathbf{p}_{ij}$ , $\frac{1}{|\mathcal{N}_i|} \sum_j \mathbf{p}_{ij}$ and their cross product; 2) a L2-Norm layer, computing the magnitude of the vector neurons $\tilde{\mathbf{X}}_j = \| \mathbf{X}_j\|_2 \in \mathbb{R}^d$ to transform the feature into a rotation-invariant form; and 3) a MLP block, mapping this feature vector into K dimension to obtain the correlation vector, $\gamma_j = \mathrm{MLP}(\tilde{\mathbf{X}}_j) \in \mathbb{R}^K$ . Finally, a softmax operation is leveraged to output correlation scores. As a result, our convolution is formulated as: + +$$ +\left(\mathbf {F} * g\right) \left(\mathbf {p} _ {i}\right) = \sum_ {\mathbf {p} _ {j} \in \mathcal {N} _ {i}} \sum_ {k} \gamma_ {j k} \mathbf {W} _ {k} \mathbf {F} _ {j}. \tag {3} +$$ + +Compared to KPConv using a linear correlation, our correlation network can learn more complex relationships. Moreover, PARE-Conv predicts correlation + +scores depending on local spatial property, thus can leverage the spatial information better and can provide stronger expressive power to extract more discriminative descriptors. Moreover, it is compatible with DGCNN [29], allowing replacing the node feature $\mathbf{F}_j$ with the edge feature $[\mathbf{F}_j - \mathbf{F}_i,\mathbf{F}_j]$ . This can further enhance the expressive power, as discussed in the ablation study 4.3. + +Fully Convolutional Network. As shown in Fig. 3, we develop a hierarchical convolution network based on PARE-Conv. The original point clouds $\mathcal{P}$ and $\mathcal{Q}$ are downsampled three times to yield sparse superpoints $\hat{\mathcal{P}}$ and $\hat{\mathcal{Q}}$ , whose resolution is $\frac{1}{2^3}$ of the original point clouds. Consequently, three convolution blocks are leveraged to extract multi-level features. The bottleneck rotation-equivariant features are denoted as $\hat{\mathbf{F}}^{\mathcal{P}} \in \mathbb{R}^{|\hat{\mathcal{P}}| \times \hat{d} \times 3}$ and $\hat{\mathbf{F}}^{\mathcal{Q}} \in \mathbb{R}^{|\hat{\mathcal{Q}}| \times \hat{d} \times 3}$ , and a VN-invariant layer [8] is used to obtain the rotation-invariant features $\hat{\mathbf{X}}^{\mathcal{P}} \in \mathbb{R}^{|\hat{\mathcal{P}}| \times 3\hat{d}}$ and $\hat{\mathbf{X}}^{\mathcal{Q}} \in \mathbb{R}^{|\hat{\mathcal{Q}}| \times 3\hat{d}}$ . For the decoder, two nearest upsampling blocks are used to get the rotation-equivariant features $\tilde{\mathbf{F}}^{\mathcal{P}} \in \mathbb{R}^{|\tilde{\mathcal{P}}| \times \tilde{d} \times 3}$ and $\tilde{\mathbf{F}}^{\mathcal{Q}} \in \mathbb{R}^{|\tilde{\mathcal{Q}}| \times \tilde{d} \times 3}$ of the first-level downsampled points $\tilde{\mathcal{P}}$ and $\tilde{\mathcal{Q}}$ . Another VN-invariant layer is leveraged to get their rotation-invariant features $\tilde{\mathbf{X}}^{\mathcal{P}} \in \mathbb{R}^{|\tilde{\mathcal{P}}| \times 3\tilde{d}}$ and $\tilde{\mathbf{X}}^{\mathcal{Q}} \in \mathbb{R}^{|\tilde{\mathcal{Q}}| \times 3\tilde{d}}$ . For more details on the convolutional block, please refer to Appendix B. + +Following GeoTrans [25], we exploit the point-to-node grouping strategy to assign the points $\hat{\mathcal{P}}$ to the superpoints $\hat{\mathcal{P}}$ . The subset of $\hat{\mathcal{P}}$ belongs to the superpoint $\hat{\mathcal{P}}_i$ is denoted as $\mathcal{G}_i^{\mathcal{P}}$ and the associated rotation-equivariant and invariant feature matrices of subset points are denoted as $\tilde{\mathbf{F}}_i^{\mathcal{P}}$ and $\tilde{\mathbf{X}}_i^{\mathcal{P}}$ . For point cloud $\tilde{\mathcal{Q}}$ , the points are grouped as $\{\mathcal{G}_i^{\mathcal{Q}}\}$ , and the feature matrices are denoted as $\tilde{\mathbf{F}}_i^{\mathcal{Q}}$ and $\tilde{\mathbf{X}}_i^{\mathcal{Q}}$ in a same way. + +# 3.2 Coarse-to-Fine Matching + +Since our backbone outputs both patch-level and point-level features, we use a coarse-to-fine matching strategy [25,34,35] to filter out non-overlapped regions and estimate more accurate correspondences. + +Superpoint Matching. Direct matching superpoints by nearest neighbor search lacks robustness to repeated patterns and low-overlapped problems. Therefore, we utilize the Geometric Transformer module [25] presented by Qin et al. to reason the global context of two point clouds, which iteratively uses self and cross-attention to capture intra and inter-point-cloud features. + +We feed the rotation-invariant features $\hat{\mathbf{X}}^{\mathcal{P}}$ and $\hat{\mathbf{X}}^{\mathcal{Q}}$ and the attached position information $\hat{\mathcal{P}}$ and $\hat{\mathcal{Q}}$ into this module to get more discriminative features $\hat{\mathbf{H}}^{\mathcal{P}}$ and $\hat{\mathbf{H}}^{\mathcal{Q}}$ . Then, a Gaussian correlation matrix is computed, followed by a dual normalization operation. The superpoint correspondences are obtained by selecting the top- $k$ reliable correspondences $\hat{\mathcal{C}} = \{(\hat{\mathbf{p}}_{x_i},\hat{\mathbf{q}}_{y_i})|(x_i,y_i)\in \arg \max_{x,y}\mathbf{S}_{x,y}\}$ , where $\mathbf{S}$ is the similarity matrix. + +Improvement. Here, we find this module is sensitive to the overlap distribution of the dataset due to its cross-attention layers. This could lead to a significant performance decrease when there is a substantial difference between the overlap distributions of the test set and the training set. To tackle this issue, we present a simple but effective data augmentation technique, RandomCrop, + +where point clouds are randomly cropped to make the model more robust to low-overlapped point cloud pairs. We present more details in the Appendix C. + +Point Matching. When the correspondences $\hat{\mathcal{C}}_i = (\hat{\mathbf{p}}_{x_i},\hat{\mathbf{q}}_{y_i})$ of superpoints are established, we search point-level correspondences within $\mathcal{G}_{x_i}^{\mathcal{P}}$ and $\mathcal{G}_{y_i}^{\mathcal{Q}}$ . We follow [22] that disentangles the similarity and saliency of features by using a matchability head $\mathbf{W}_m:\mathbb{R}^{3\tilde{d}}\to \mathbb{R}^{3\tilde{d}}$ and a saliency head $\mathbf{W}_s:\mathbb{R}^{3\tilde{d}}\to \mathbb{R}$ . Specifically, we compute a matching matrix: + +$$ +\mathbf {M} _ {x _ {i}, y _ {i}} = \left(\mathbf {W} _ {m} \mathbf {X} _ {x _ {i}} ^ {\mathcal {P}}\right) ^ {\mathrm {T}} \mathbf {W} _ {m} \mathbf {X} _ {y _ {i}} ^ {\mathcal {Q}} / \sqrt {3 \tilde {d}}, \tag {4} +$$ + +and then the salient scores are predicted: + +$$ +\sigma_ {x _ {i}} ^ {\mathcal {P}} = \operatorname {S i g m o i d} \left(\mathbf {W} _ {s} \mathbf {X} _ {x _ {i}} ^ {\mathcal {P}}\right), \sigma_ {y _ {i}} ^ {\mathcal {Q}} = \operatorname {S i g m o i d} \left(\mathbf {W} _ {s} \mathbf {X} _ {y _ {i}} ^ {\mathcal {Q}}\right). \tag {5} +$$ + +Finally, a soft assignment matrix $\mathbf{Z}_i$ is computed by: + +$$ +\mathbf {Z} _ {i} = \sigma_ {x _ {i}} ^ {\mathcal {P}} \sigma_ {y _ {i}} ^ {\mathcal {Q}} \varphi \left(\mathbf {M} _ {x _ {i}, y _ {i}}\right) _ {x _ {i}} \varphi \left(\mathbf {M} _ {x _ {i}, y _ {i}}\right) _ {y _ {i}}, \tag {6} +$$ + +where $\varphi$ is the softmax operation. As a result, we select $N_{f}$ point-level correspondences $\mathcal{C} = \{(\tilde{\mathbf{p}}_{x_j},\tilde{\mathbf{q}}_{y_j})\}$ across all patch correspondence whose assignment scores are the highest. The impact of $N_{f}$ is analyzed in Fig. 2. + +# 3.3 Feature-based hypothesis Proposer + +Since our rotation-equivariant features encode orientation information of local structures, we derive the pose transformation from matched features. + +Hypothesis Generation and Selection. Once obtaining point-wise correspondence set $\mathcal{C} = \{(\tilde{\mathbf{p}}_{x_j},\tilde{\mathbf{q}}_{y_j})\}$ , we leverage their corresponding rotation-equivariant feature maps $\{(\tilde{\mathbf{F}}_{x_j}^{\mathcal{P}},\tilde{\mathbf{F}}_{y_j}^{\mathcal{Q}})\}$ to generate multiple model hypotheses. For each correspondence $(\tilde{\mathbf{p}}_{x_j},\tilde{\mathbf{q}}_{y_j})$ , a rotation can be estimated by solving: + +$$ +\tilde {\mathbf {R}} _ {j} = \underset {\mathbf {R}} {\arg \min } \| \mathbf {R} \left(\tilde {\mathbf {F}} _ {x _ {j}} ^ {\mathcal {P}}\right) ^ {\mathrm {T}} - \left(\tilde {\mathbf {F}} _ {y _ {j}} ^ {\mathcal {Q}}\right) ^ {\mathrm {T}} \| _ {2} ^ {2}. \tag {7} +$$ + +This can be efficiently solved by SVD [5] in closed form. Subsequently, the translation can be solved by: + +$$ +\tilde {\mathbf {t}} _ {j} = \tilde {\mathbf {q}} _ {y _ {j}} - \tilde {\mathbf {R}} _ {j} \tilde {\mathbf {p}} _ {x _ {j}}. \tag {8} +$$ + +As a result, we can get a set of model hypotheses $\mathcal{H}_c = \{(\tilde{\mathbf{R}}_i,\tilde{\mathbf{t}}_i)\}$ . The hypothesis that has the most inlier support will be selected as the alignment solution: + +![](images/872be62db4fc44d150e006ad7bfdfe42a597a8ef8651879900fad93a58f72a6b.jpg) +Source point cloud Target point cloud + +![](images/56e8e0b2304dcdfcefe6b980b269d476da04bbe5d1505373cb327ed56ec70f6c.jpg) +Symmetric alignment Correct alignment + +Fig.4: An illustration of incorrect alignment caused by the ambiguity of using coordinates to align. This problem can be solved by using rotation-equivariant features that contain fine-grained local orientation information. + +$$ +\left(\mathbf {R} ^ {*}, \mathbf {t} ^ {*}\right) = \underset {\tilde {\mathbf {R}} _ {j}, \tilde {\mathbf {t}} _ {j}} {\arg \max } \sum_ {\left(\tilde {\mathbf {p}} _ {x _ {j}}, \tilde {\mathbf {q}} _ {y _ {j}}\right) \in \mathcal {C}} \mathbb {1} \left(\| \tilde {\mathbf {R}} _ {j} \tilde {\mathbf {p}} _ {x _ {j}} + \tilde {\mathbf {t}} _ {j} - \tilde {\mathbf {q}} _ {y _ {j}} \| _ {2} ^ {2} < \tau_ {d}\right), \tag {9} +$$ + +where $\tau_{d}$ is the acceptance radius. The best model hypothesis will be iteratively refined by its inlier correspondences for a more accurate solution, as in [25]. + +Discussion. Traditional estimators, such as RANSAC [11] randomly sample multiple pairs of correspondences and use their coordinates to estimate the transformation. Only when all the sampled correspondences belong to inliers, a reliable transformation can be estimated. This approach is inefficient and requires multiple iterations. In contrast, we directly utilize rotation-equivariant features to estimate the pose of the point cloud, and each pair of correspondences generates a hypothesis. This approach is more efficient, as depicted in the Fig. 2. Our method achieves a significantly high registration recall by generating only 1000 hypotheses. In contrast, RANSAC is still inferior to our method even with 50,000 iterations. Moreover, using coordinates to solve the transformation may falsely align the point clouds with opposite orientations, due to ignoring the orientation information, as shown in Fig. 4. This problem can be easily addressed by our method as the rotation-equivariant features encode orientations of local structures. In addition, some methods [1, 28] use regressors to implicitly estimate transformations from equivariant features. Instead, our method can directly solve pose estimation through feature alignment, which is conceptually simpler and computationally efficient. + +# 3.4 Loss Function + +The network is supervised by three losses $\mathcal{L} = \mathcal{L}_c + \mathcal{L}_f + \mathcal{L}_r$ , including superpoint matching loss $\mathcal{L}_c$ , point matching loss $\mathcal{L}_f$ , and contrastive rotation loss $\mathcal{L}_r$ . + +Point Matching Loss $\mathcal{L}_f$ . Since point matching is performed within the scopes of matched superpoints, unreliable superpoint matching may lead to sparse supervision signals for point matching. Therefore, we use positive superpoint correspondences computed by ground-truth transformation during the training process. For two grouped point sets $\mathcal{G}_{x_i}^{\mathcal{P}}$ and $\mathcal{G}_{y_i}^{\mathcal{Q}}$ , the positive correspondences $\mathcal{C}_i^*$ , having distance errors smaller than a threshold $d_p$ , are identified by the groundtruth. The unmatched points are denoted as $\mathcal{I}_i$ , $\mathcal{J}_i$ . We compute the log-likelihood loss on soft assignment matrix $\mathbf{Z}_i$ and the salient scores $\sigma_{x_i}$ , $\sigma_{y_i}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {f _ {i}} = - \frac {1}{| \mathcal {C} _ {i} ^ {*} |} \sum_ {(x _ {i}, y _ {i}) \in \mathcal {C} _ {i} ^ {*}} \log \mathbf {Z} _ {x _ {i}, y _ {i}} - \frac {1}{2 | \mathcal {I} _ {i} |} \sum_ {x _ {i} \in \mathcal {I} _ {i}} \log \left(1 - \sigma_ {x _ {i}} ^ {\mathcal {P}}\right) \\ - \frac {1}{2 \left| \mathcal {J} _ {i} \right|} \sum_ {y _ {i} \in \mathcal {J} _ {i}} \log \left(1 - \sigma_ {y _ {i}} ^ {Q}\right). \tag {10} \\ \end{array} +$$ + +The overall point matching loss is a sum of $\mathcal{L}_f = \sum_i\mathcal{L}_{f_i}$ + +Contrastive Rotation Loss $\mathcal{L}_r$ . Rotation-equivariant feature inherently encodes pose information of local geometric structure. However, such features of a positive correspondence inevitably be contaminated by data degradation, such as occlusion, noise and down-sampling errors. To make the features more robust to these troubles, we propose a contrastive rotation loss, to supervise point-level + +correspondences: + +$$ +\begin{array}{l} \mathcal {L} _ {r _ {i}} = \frac {1}{| \mathcal {C} _ {i} ^ {*} | \cdot \tilde {d}} \sum_ {\left(x _ {i}, y _ {i}\right) \in \mathcal {C} _ {i} ^ {*}} \sum_ {c = 1} ^ {\tilde {d}} [ \| \tilde {\mathbf {F}} _ {x _ {i}, c} ^ {\mathcal {P}} \mathbf {R} _ {g t} ^ {\mathrm {T}} - \tilde {\mathbf {F}} _ {y _ {i}, c} ^ {\mathcal {Q}} \| _ {2} ^ {2} - \alpha ] _ {+} \tag {11} \\ + \frac {1}{| \bar {\mathcal {C}} _ {i} | \cdot \tilde {d}} \sum_ {(x _ {i}, y _ {i}) \in \bar {\mathcal {C}} _ {i}} \sum_ {c = 1} ^ {\tilde {d}} [ \beta - \| \tilde {\mathbf {F}} _ {x _ {i}, c} ^ {\mathcal {P}} \mathbf {R} _ {g t} ^ {\mathrm {T}} - \tilde {\mathbf {F}} _ {y _ {i}, c} ^ {\mathcal {Q}} \| _ {2} ^ {2} ] _ {+}, \\ \end{array} +$$ + +where $\mathcal{C}_i^*$ is the aforementioned positive correspondence set, $\bar{\mathcal{C}}_i$ is the negative correspondence set whose point pair has spatial distance greater than $d_{n}$ , $\alpha = 0.1$ and $\beta = 1.4$ are positive and negative margins, and $[\cdot]_{+}$ is the ReLU operation. + +Superpoint Matching Loss $\mathcal{L}_c$ . To supervise the rotation-invariant features of superpoints, we utilize the overlap-aware circle loss that considers the overlap ratio of two superpoints to compute more accurate gradients. For more details, please refer to GeoTrans [25]. + +# 4 Experiments + +To evaluate the performance of our method, we compare our method with state-of-the-art approaches and perform extensive experiments on 3DMatch [37], 3DLoMatch [15], and KITTI Odometry [12]. Finally, ablation studies are conducted to analyze the components of our method. All the experiments are conducted with a RTX 3090 GPU with Intel (R) Xeon (R) Silver 4314 CPU. More implementation details are reported in Appendix D. + +# 4.1 Indoor Dataset: 3DMatch & 3DLoMatch + +DataSet. 3DMatch [37] and 3DLoMatch [15] are collected from indoor RGB-D datasets. 3DMatch contains 62 scenes and there are $46/8/8$ scenes for training/validation/testing. It consists of point cloud pairs with overlap ratios greater than $30\%$ . 3DLoMatch is only used for testing that contains point cloud pairs with overlap ratios from $10\%$ to $30\%$ . We follow the protocols [15, 25, 28] to perform all the experiments. + +Baseline. We compare our method with strong baselines: 1) FCGF [7], Predator [15], GeoTrans [25], and PEAL [36] are advanced scene-level feature extractors, and some of them [25,36] follow a coarse-to-fine manner to extract robust correspondences. PEAL requires overlap prior as input, we follow the office implementation that leverages 3D overlap prior predicted by GeoTrans for a fair comparison. 2) YOHO [27], RoReg [28], and RoITR [35] are also scene-level feature extractors but focus on rotation-invariant/equivariant network designs. 3) SpinNet [2] and BUFFER [1] leverage patch-level feature extractors with elaborate designs for rotation-invariance. + +Metrics. We follow [15,25] that use three evaluation metrics, including Inlier Ratio (IR), Feature Matching Recall (FMR), and Registration Recall (RR). To + +Table 1: Evaluation results on 3DMatch and 3DLoMatch. Methods with the RANSAC estimator are marked by $\diamond$ , which exploit 5000 points to establish correspondences. + +
MethodSize (MB)3DMatch3DLoMatch
FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)
FCGF° [7]8.7694.731.182.80.1259.49.838.00.13
SpinNet° [2]1.4197.647.588.69.8575.320.559.89.03
YOHO [27]12.3898.264.490.82.8178.925.966.02.62
Predator° [15]7.4396.658.089.00.6478.226.764.40.47
GeoTrans [25]9.8398.170.992.40.1887.443.574.30.17
RoReg [28]12.7198.281.693.02.2782.339.670.12.10
BUFFER [1]0.92--92.10.22--70.30.21
RoITR° [35]10.1098.082.491.90.3689.254.674.10.34
PEAL [36]9.8398.471.094.21.4688.346.078.81.19
Ours3.8498.576.995.00.1788.347.580.50.17
+ +evaluate the performance of different methods on rotated datasets, RoITR [35] uses two ways to compute the RMSE on the original and rotated datasets, which cannot show the robustness of methods against pose variations. To tackle this issue, we follow [28] to report Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) on both original and rotated datasets. Please refer to the Appendix E for more details about these evaluation metrics. + +Quantitative results. We report the experimental results in Table 1. Compared with SOTA methods, our approach achieves a significantly high RR on both 3DMatch and 3DLoMatch, while simultaneously maintaining lightweight and computation efficient. PEAL [36] is a formidable competitor that employs an iterative manner for alignment due to the requirements of overlap prior, resulting in significant time overhead. In contrast, our method achieves a speed improvement of 7 times compared to PEAL, while also outperforming it by $1.7\%$ in terms of RR on 3DLoMatch. GeoTrans [25], RoITR [35], and our method all employ a coarse-to-fine matching framework. Despite using a more lightweight feature extractor, our RR is significantly higher than theirs, particularly on 3DLoMatch, where our method outperforms them by $6.2\%$ and $6.4\%$ respectively. This can be attributed to the ability of our rotation-equivariant network to focus more on learning discriminative structural information. YOHO [27] and RoReg [28] are extremely time-consuming due to the group feature extractor and embedder. Our efficiency advantage stems from two aspects. Firstly, we utilize a lightweight scene-wise feature extractor, which only needs a forward process to extract dense rotation-equivariant features. Secondly, our feature-based hypothesis proposer can simultaneously generate multiple valid hypotheses by using matched rotation-equivariant features. Thus, it is more efficient compared to RANSAC [11] with many approaches [2, 7, 15, 35] used. + +Experiments on the rotated dataset. To test the robustness against rotations, we follow previous works [28, 35] that evaluate all the methods on the rotated 3DLoMatch, where arbitrary rotations are applied to point clouds. The experimental results are shown in Table 2. As we can see, PEAL experiences a decrease in TR by $2.5\%$ on the Rotated 3DLoMatch dataset, while our + +Table 2: Evaluation results on 3DLoMatch and Rotated 3DLoMatch. For easy comparison of the impact of rotation, the numerical changes w.r.t. TR on the rotated dataset are annotated in the top right corner of each value. Methods with rotation invariant/equivariant designs are marked by * symbol. + +
MethodSize (MB)3DLoMatchRotated 3DLoMatch
RE (°↓)TE (cm ↓)TR (% ↑)RE (°↓)TE (cm ↓)TR (% ↑)
FCGF [7]8.764.8412.8739.64.7413.3924.5-15.1
PREDATOR [15]7.433.6110.6565.63.5510.3064.0-1.6
GeoTrans [25]9.832.918.7175.42.948.8572.6-2.8
PEAL [36]9.832.848.6481.22.868.5378.7-2.5
YOHO* [27]12.383.5410.3466.63.6110.1667.1+0.5
RoReg* [28]12.713.019.2671.33.039.2871.0-0.3
BUFFER* [1]0.923.039.8674.43.029.9974.7+0.3
RoITR* [35]10.102.959.0375.12.979.0875.5+0.4
Ours*3.842.878.8381.32.848.7181.8+0.5
+ +method demonstrates an increase of $0.5\%$ . This result validates the robustness of our approach against rotations, illustrating its ability to handle such variations. Methods without rotation-invariant design, such as FCGF [7], Predator [15], and GeoTrans [25] also show performance degradation, indicating that networks trained with rotation augmentation are vulnerable to rotation variations. On the other hand, methods such as YOHO [27], RoReg [28], RoITR [35], and BUFFER [1], which incorporate rotation equivariant or invariant designs, demonstrate better robustness against data rotations. However, our method significantly outperforms these methods by $6\%$ w.r.t TR on both 3DLoMatch and Rotated 3DLoMatch. This is because our backbone can capture spatial geometric information better, thus extracting more distinctive descriptors. + +Qualitative results. We show some registration results in Fig. 5, which demonstrate that our method can estimate more accurate patch-level and point-level correspondences and can successfully align point cloud pairs with low-overlapped ratios. Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method suc + +cessfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity. Please refer to Appendix G for more visualization results. + +Table 3: Evaluation results on KITTI Odometry. + +
MethodSizeRE (°)TE (cm)TR (%)Time (s)
FCGF [7]8.760.309.596.6-
D3Feat [4]14.080.307.299.8-
Predator [15]22.770.276.899.80.77
SpinNet [2]1.410.479.999.116.24
CoFiNet [34]5.480.418.299.80.59
GeoTrans [25]25.500.236.299.80.26
BUFFER [1]0.920.267.199.80.27
Ours2.080.234.999.80.21
+ +![](images/a0e48852e9221594f77b1c4ba1dc3b744b58a355ef96c4fadfef1ae120c4b1d1.jpg) +Fig. 5: Qualitative results. Our method can successfully align low-overlapped pairs with higher IR and PIR (Patch IR). Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method successfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity. + +# 4.2 Outdoor Dataset: KITTI Odometry + +Dataset. We follow previous methods [7,25] to evaluate our method on KITTI Odometry [12]. It comprises 11 urban scene sequences captured by a LiDAR sensor. We use 0-5/6-7/8-10 sequences for training/validation/testing, respectively. + +Metrics. We use the Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) to evaluate our method. Following [7,25], the TR is computed with rotation threshold $5^{\circ}$ and translation threshold $2\mathrm{m}$ . + +Experimental results. The experimental results are shown in Table 3. On KITTI dataset, many methods have reached performance saturation, while our approach achieves the highest accuracy when achieving the highest TR. Our method also shows advantages w.r.t. model size and running speed. + +# 4.3 Ablation Study + +To better understand the designs of our method, we conduct comprehensive ablation experiments and the evaluated results are reported in Table 4. + +Ablation of backbone. To investigate the impact of different convolution methods, we compare our PARE-Conv with VN [8]. We leverage both node features [24] and edge features [29] for evaluation. The experimental results indicate that our approach achieves significant improvement compared to VN whether using node features or edge features. This demonstrates the enhanced representative power of PARE-Conv to learn more distinctive features. We also compare PARE-Conv with KPCnv [26], refer to the Appendix F.5 for details. + +Ablation of pose estimator. To demonstrate the superiority of our pose estimator, we compared it with RANSAC [11] and LGR [25]. RANSAC randomly selects a few pairs of correspondences to generate hypotheses, while LGR generates a hypothesis using the correspondences within a matched patch pair. + +In contrast, our approach directly generates a hypothesis using the equivariant features of a pair of matched points. As shown in Table 4, our method achieves the highest RR, outperforming RANSAC and LGR by $3\%$ and $1.6\%$ respectively on 3DLoMatch. This validates that estimating poses using rotation-equivariant features based on local structures is more robust than using point-to-point correspondences. Additionally, the results of FMR and IR are identical as we used the same correspondences for different pose estimators. + +Table 4: Ablation experiments of our method. + +
Component Method3DMatch3DLoMatch
FMR (% ↑) IR (% ↑) RR (% ↑)FMR (% ↑) IR (% ↑) RR (% ↑)
BackboneVN (Node) [8]98.671.093.487.841.476.6
PARE-Conv (Node)98.974.194.5 +1.188.344.078.8 +2.2
VN (Edge) [8]98.772.494.188.843.577.8
PARE-Conv (Edge)98.576.995.0 +0.988.147.580.5 +2.7
Pose EstimatorRANSAC98.576.993.788.347.577.5
LGR [25]98.576.994.0 +0.388.347.578.9 +1.4
Ours98.576.995.0 +1.388.347.580.5 +3.0
Rotation LossWithout98.374.493.388.745.579.3
L298.665.494.2 +0.985.935.075.4 -3.9
Contrastive98.576.995.0 +1.788.347.580.5 +1.2
+ +Ablation of rotation loss. The contrastive rotation loss is designed to improve the robustness of rotation-equivariant features to data occlusion, noise, and partial overlap. When the rotation-equivariant features are not supervised by a loss function (Indeed, it is indirectly supervised by the loss function of rotation-invariant features, as the invariant features are derived from the equivariant features), the method shows decent performance on 3DLoMatch but shows poor performance on 3DMatch. When using the L2 loss to supervise equivariant features of positive point pairs, there is a slight improvement in RR on 3DMatch but shows performance degradation on 3DLoMatch. However, when utilizing our contrastive rotation loss, the performance of the method is enhanced on both 3DMatch and 3DLoMatch, demonstrating its superiority. + +# 5 Conclusion + +In this paper, we propose PARE-Net, a lightweight network for fast and robust point cloud registration. We introduce position-aware rotation-equivariant convolution to learn spatial information effectively, thus extracting more distinctive descriptors. We also propose an efficient feature-based hypothesis proposer to generate reliable model hypotheses. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art approaches and shows great robustness to rotation variations. The limitations of our method and future works are discussed in Appendix H. + +Acknowledgements. This work was supported by the National Science and Technology Major Project of China under Grant No. 2020AAA0108102, the National Natural Science Foundation of China under Grant Nos. 62327808, 62073257 and 62088102, and the Fundamental Research Funds for the Central Universities under Grant No. xzy022024007. + +# References + +1. Ao, S., Hu, Q., Wang, H., Xu, K., Guo, Y.: BUFFER: Balancing accuracy, efficiency, and generalizability in point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1255-1264 (2023) +2. Ao, S., Hu, Q., Yang, B., Markham, A., Guo, Y.: SpinNet: Learning a general surface descriptor for 3D point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11753-11762 (2021) +3. Bai, X., Luo, Z., Zhou, L., Chen, H., Li, L., Hu, Z., Fu, H., Tai, C.L.: PointDSC: Robust point cloud registration using deep spatial consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15859-15869 (2021) +4. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3Feat: Joint learning of dense detection and description of 3D local features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6359-6367 (2020) +5. Besl, P.J., McKay, N.D.: Method for registration of 3-D shapes. In: Sensor Fusion IV: Control Paradigms and Data Structures. vol. 1611, pp. 586-606 (1992) +6. Chen, Z., Sun, K., Yang, F., Tao, W.: SC²-PCR: A second order spatial compatibility for efficient and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13221-13231 (2022) +7. Choy, C., Park, J., Koltun, V.: Fully convolutional geometric features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8958-8966 (2019) +8. Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., Guibas, L.J.: Vector Neurons: A general framework for SO(3)-equivariant networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12200-12209 (2021) +9. Deng, H., Birdal, T., Ilic, S.: PPF-FoldNet: Unsupervised learning of rotation invariant 3D local descriptors. In: Proceedings of the European Conference on Computer Vision. pp. 602-618 (2018) +10. Deng, H., Birdal, T., Ilic, S.: PPFNet: Global context aware local features for robust 3D point matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 195-205 (2018) +11. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. vol. 24, pp. 381-395. ACM New York (1981) +12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012) + +13. Gojcic, Z., Zhou, C., Wegner, J.D., Wieser, A.: The perfect match: 3D point cloud matching with smoothed densities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5545-5554 (2019) +14. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3D local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016) +15. Huang, S., Gojcic, Z., Usvyatsov, M., Wieser, A., Schindler, K.: PREDATOR: Registration of 3D point clouds with low overlap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4267-4276 (2021) +16. Kaya, M., Bilge, H.S.: Deep metric learning: A survey. Symmetry 11(9), 1066-1092 (2019) +17. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 153-161 (2017) +18. Lee, J., Kim, S., Cho, M., Park, J.: Deep hough voting for robust global registration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15994-16003 (2021) +19. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022) +20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems 32(8), 3412-3432 (2020) +21. Lu, W., Zhou, Y., Wan, G., Hou, S., Song, S.: L3-Net: Towards learning based LiDAR localization for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6389-6398 (2019) +22. Philipp, L., Paul-Edouard, S., Marc, P.: LightGlue: Local Feature Matching at Light Speed. In: Proceedings of the IEEE International Conference on Computer Vision (2023) +23. Pomerleau, F., Colas, F., Siegwart, R.: A review of point cloud registration algorithms for mobile robotics. Foundations and Trends in Robotics 4(1), 1-104 (2015) +24. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017) +25. Qin, Z., Yu, H., Wang, C., Guo, Y., Peng, Y., Xu, K.: Geometric transformer for fast and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11143-11152 (June 2022) +26. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPConv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6411-6420 (2019) +27. Wang, H., Liu, Y., Dong, Z., Wang, W.: You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1630-1641 (2022) +28. Wang, H., Liu, Y., Hu, Q., Wang, B., Chen, J., Dong, Z., Guo, Y., Wang, W., Yang, B.: RoReg: Pairwise point cloud registration with oriented descriptors and local rotations. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 10376-10393 (2023). https://doi.org/10.1109/TPAMI.2023.3244951 + +29. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph CNN for learning on point clouds. ACM Transactions On Graphics (ToG) 38(5), 1-12 (2019) +30. Xu, M., Ding, R., Zhao, H., Qi, X.: PACConv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3173-3182 (2021) +31. Yang, B., Bender, G., Le, Q.V., Ngiam, J.: CondConv: Conditionally parameterized convolutions for efficient inference. In: Proceedings of Advances in Neural Information Processing Systems. vol. 32 (2019) +32. Yao, R., Du, S., Cui, W., Ye, A., Wen, F., Zhang, H., Tian, Z., Gao, Y.: Hunter: Exploring high-order consistency for point cloud registration with severe outliers. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14760-14776 (2023). https://doi.org/10.1109/TPAMI.2023.3312592 +33. Yew, Z.J., Lee, G.H.: REGTR: End-to-end point cloud correspondences with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6677-6686 (2022) +34. Yu, H., Li, F., Saleh, M., Busam, B., Ilic, S.: CoFiNet: Reliable coarse-to-fine correspondences for robust point cloud registration. In: Proceedings of Advances in Neural Information Processing Systems. pp. 23872-23884 (2021) +35. Yu, H., Qin, Z., Hou, J., Saleh, M., Li, D., Busam, B., Ilic, S.: Rotation-invariant transformer for point cloud matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5384-5393 (2023) +36. Yu, J., Ren, L., Zhang, Y., Zhou, W., Lin, L., Dai, G.: PEAL: Prior-embedded explicit attention learning for low-overlap point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17702-17711 (2023) +37. Zeng, A., Song, S., Nießner, M., Fisher, M., Xiao, J., Funkhouser, T.: 3DMatch: Learning local geometric descriptors from RGB-D reconstructions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1802-1811 (2017) +38. Zhang, Z., Dai, Y., Sun, J.: Deep learning based point cloud registration: An overview. Virtual Reality and Intelligent Hardware 2(3), 222-246 (2020) \ No newline at end of file diff --git a/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/images.zip b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c68b24b1c9804a873c4390bbfe6e57966b5a1763 --- /dev/null +++ b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dd572ce1fba407c170759c6a17b7228ddd11c6296fefb1d59dca58b3d230613 +size 520847 diff --git a/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/layout.json b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5759f825540964120e8ab6e6a90e3dc5ff439577 --- /dev/null +++ b/2024/PARE-Net_ Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration/layout.json @@ -0,0 +1,9285 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "text", + "content": "PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 168, + 472, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 168, + 472, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 168, + 472, + 193 + ], + "type": "text", + "content": "Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 202, + 478, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 202, + 478, + 236 + ], + "spans": [ + { + "bbox": [ + 135, + 202, + 478, + 236 + ], + "type": "text", + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 264, + 452, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 264, + 452, + 528 + ], + "spans": [ + { + "bbox": [ + 160, + 264, + 452, + 528 + ], + "type": "text", + "content": "Abstract. Learning rotation-invariant distinctive features is a fundamental requirement for point cloud registration. Existing methods often use rotation-sensitive networks to extract features, while employing rotation augmentation to learn an approximate invariant mapping rudely. This makes networks fragile to rotations, overweight, and hinders the distinctiveness of features. To tackle these problems, we propose a novel position-aware rotation-equivariant network, for efficient, light-weighted, and robust registration. The network can provide a strong model inductive bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned limitations. To further improve the distinctiveness of descriptors, we propose a position-aware convolution, which can better learn spatial information of local structures. Moreover, we also propose a feature-based hypothesis proposer. It leverages rotation-equivariant features that encode fine-grained structure orientations to generate reliable model hypotheses. Each correspondence can generate a hypothesis, thus it is more efficient than classic estimators that require multiple reliable correspondences. Accordingly, a contrastive rotation loss is presented to enhance the robustness of rotation-equivariant features against data degradation. Extensive experiments on indoor and outdoor datasets demonstrate that our method significantly outperforms the SOTA methods in terms of registration recall while being lightweight and keeping a fast speed. Moreover, experiments on rotated datasets demonstrate its robustness against rotation variations. Code is available at https://github.com/yaorz97/PARENet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 538, + 453, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 538, + 453, + 559 + ], + "spans": [ + { + "bbox": [ + 160, + 538, + 453, + 559 + ], + "type": "text", + "content": "Keywords: Point cloud registration " + }, + { + "bbox": [ + 160, + 538, + 453, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 453, + 559 + ], + "type": "text", + "content": " Rotation invariance/equivalence " + }, + { + "bbox": [ + 160, + 538, + 453, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 538, + 453, + 559 + ], + "type": "text", + "content": " Robust transformation estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Point cloud registration, estimating a rigid transformation to align two point clouds that are partially overlapped, is fundamental research in 3D computer vision. It is widely applied in autonomous driving [20,21], robot localization [23], and 3D reconstruction [38]. Feature-based point cloud registration framework is widely studied, which mainly involves descriptor extraction [4, 15, 25, 34] and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "robust transformation estimation [3,6,32]. One inherent contradiction in point cloud registration is the pose variation of point clouds and the desired invariance of their descriptors. To extract such rotation-invariant descriptors, plenty of approaches have been proposed that can be roughly categorized into patch-wise and scene-wise feature extractors." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 195, + 305, + 250 + ], + "blocks": [ + { + "bbox": [ + 134, + 195, + 305, + 250 + ], + "lines": [ + { + "bbox": [ + 134, + 195, + 305, + 250 + ], + "spans": [ + { + "bbox": [ + 134, + 195, + 305, + 250 + ], + "type": "image", + "image_path": "980f9b4fc9a18938bb508e16c4db13c257158c71adcea63ce949e1c50cf98bc1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 292 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 292 + ], + "type": "text", + "content": "Fig. 1: Experimental results on 3DMatch and 3DLoMatch. Our method significantly outperforms state-of-the-art methods w.r.t. registration recall (RR) while maintaining fast speed and lightweight." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 195, + 480, + 249 + ], + "blocks": [ + { + "bbox": [ + 309, + 195, + 480, + 249 + ], + "lines": [ + { + "bbox": [ + 309, + 195, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 309, + 195, + 480, + 249 + ], + "type": "image", + "image_path": "1891bf84bac650bb834c2d014ffc40dd64adc48fdd6e9f5c59347f3fcf1ca2dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 437 + ], + "type": "text", + "content": "Patch-wise feature extractors [2, 9, 10] learn geometric features from local patches of point clouds and typically employ data preprocessing techniques to achieve rotation-invariance. Two common approaches include: 1) extracting rotation invariant statistics as initial representation from local patches, such as distance, angle, and point pair features (PPF [9, 10]); and 2) estimating a local reference frame (LRF [2]) to normalize poses of local patches. The former approach suffers from information loss, while the latter approach is susceptible to data occlusion and noise, lacking robustness. Moreover, these extractors are generally computationally expensive because all patches of a point cloud are processed individually, without sharing intermediate representations [7]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "text", + "content": "Scene-wise feature extractors [4,7,15,25,34] learn dense feature descriptors for every point in the point cloud, and most of them employ extensive rotation augmentation in the training process to rudely enforce inherently rotation-sensitive networks (e.g., KPCov [26]) to learn rotation-invariant mappings. These methods have three shortcomings. Firstly, models are fragile to unseen rotations because it is difficult to cover the SO(3) space in the training process [35]. Secondly, learning rotation-invariant mappings requires more network parameters. Thirdly, these networks are limited to generating discriminative feature descriptors as they cannot focus solely on learning structural distinctiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "content": "In this paper, we propose a novel registration network, PARE-Net, whose core is to fully exploit the advantages of rotation-equivariant networks in both feature extraction and transformation estimation, thereby achieving a lightweight, efficient, and robust registration method, as shown in Fig. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "First of all, we propose a Position-Aware Rotation-Equivariant Convolution (PARE-Conv) that enhances the representational power of networks for better distinctive feature learning. Vector Neuron (VN) [8] builds a general rotation-equivariant network framework. However, its representative power is limited as it ignores the importance of spatial localization property of local points and utilizes multi-layer perceptrons (MLPs) to crudely learn spatial encodings. In" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 488, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 488, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 488, + 259 + ], + "type": "text", + "content": "stead, PARE-Conv is proposed to explicitly leverage the spatial information of point clouds while guaranteeing the rotation equivariant property. Specifically, we first define a set of shadow kernel points equipped with learnable weights. Then, we introduce a correlation network that predicts rotation-invariant correlations between points and shadow kernels according to rotation-equivariant spatial features of points. Therefore, the geometric structures of point clouds are well captured. Finally, the correlation scores are subsequently leveraged to dynamically assemble kernel weights for rotation-equivariant convolution. As a result, PARE-Conv inherently provides a strong model bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned problems raised by rotation-sensitive networks. Moreover, it can learn the spatial structure of point clouds better, facilitating to extraction of more distinctive descriptors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 367, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 367, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 367, + 415 + ], + "type": "text", + "content": "Secondly, we propose a feature-based hypothesis proposer that can efficiently generate reliable solution hypotheses. Once obtaining correspondences by feature matching, rotation-equivariant features encoding fine-grained structure information are used to explicitly generate multiple hypotheses. This way can alleviate the orientation ambiguity stemming from optimizing distance errors of correspondences. Moreover, to make the equivariant features more robust against data degradation, we present a contrastive rotation loss inspired by metric learning [7, 16], facilitating more robust pose estimation. As shown in Fig. 2, our method can achieve significantly higher registration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 475 + ], + "type": "text", + "content": "recall compared to RANSAC [11] and LGR (presented in GeoTrans [25]) when varying the number of hypotheses, demonstrating our method is more robust than classical estimators. Because they require multiple correspondences and the probability of sampling all correct correspondences decreases rapidly as the number of used correspondences increases. Overall, our main contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 375, + 284, + 481, + 338 + ], + "blocks": [ + { + "bbox": [ + 375, + 284, + 481, + 338 + ], + "lines": [ + { + "bbox": [ + 375, + 284, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 375, + 284, + 481, + 338 + ], + "type": "image", + "image_path": "dfa6aaa7c60440553160fd36a304b8e1660a41206ad6af8814a41a69eee1d1f0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 347, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 374, + 347, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 374, + 347, + 482, + 390 + ], + "type": "text", + "content": "Fig.2: Comparison of our feature-based hypothesis proposer, RANSAC and LGR on 3DLoMatch." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 485, + 480, + 581 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 137, + 485, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 485, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 137, + 485, + 480, + 520 + ], + "type": "text", + "content": "- A lightweight, efficient, and robust point cloud registration method that is achieved by fully exploiting the advantages of rotation-equivariant networks in feature extraction and pose estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 521, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 521, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 137, + 521, + 480, + 544 + ], + "type": "text", + "content": "- A position-aware rotation-equivariant convolution that can better leverage spatial information, enabling the extraction of discriminative descriptors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 545, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 545, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 137, + 545, + 480, + 581 + ], + "type": "text", + "content": "- A feature-based hypothesis proposer that can efficiently generate multiple reliable hypotheses and a contrastive rotation loss makes the rotation-equivariant features more robust to data degradation." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Patch-wise feature extractors. In the early stage, most learning-based feature descriptors [2, 9, 10, 13, 17, 37] were patch-wise extractors, primarily due to limitations in point cloud convolution techniques, which could not handle" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "large-scale point clouds effectively. During this period, the works drew inspiration from traditional feature descriptors [14] and focused on designing features with rotation-invariance. PPFNet [10] and PPFoldNet [9] extracted rotation-invariant PPFs as initial features and further learned high-dimensional features. PerfectMatch [13] and SpinNet [2] utilized LRF to normalize the pose of local patches and then employed networks to learn the density/structural features of patches. These methods are primarily limited by their computational efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 463 + ], + "type": "text", + "content": "Scene-wise feature extractors [4, 7, 15, 19, 25, 33-36] have become the mainstream research in recent years, due to their high efficiency. These methods mainly achieved rotation invariance through data augmentation and aim to tackle the low-overlapped problem. Predator [15] designed an overlap-attention module to detect overlap regions between two point clouds. CoFiNet [34] and GeoTrans [25] presented a coarse-to-fine matching scheme, and GeoTrans introduced a geometric transformer module that significantly increased the matching accuracy by injecting rotation-invariant position information. PEAL [36] presented an explicit one-way attention module that iteratively leveraged overlap prior to relieving feature ambiguity while bringing expensive computation costs. Moreover, some methods with rotation invariant/equivariant designs were proposed. YOHO [27] and RoReg [28] exploited the group convolution technique to extract rotation-invariant/equivariant group features, which are extremely computation expensive. Because they need to extract point features under 60 different poses and perform convolutions on massive point features. RoITR [35] used PPF to design an invariant convolution and presented a rotation invariant transformer to extract distinctive features. BUFFER [1] combined scene-wise and patch-wise feature extractors to achieve faster alignment with great generalization ability. It first leveraged a rotation-equivariant network built on vanilla VN to extract invariant and equivariant features for keypoints and orientations prediction. Then, it also used SpinNet to extract patch-wise features due to the limited representative ability of vanilla VN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": "Robust transformation estimators. Since the estimated correspondences are usually contaminated by outliers, a robust transformation estimator is necessary. RANSAC [11] is the most widely used method due to its simplicity, but it typically exhibits lower accuracy and requires plenty of iterations for convergence. Recently, some learning-based methods [3,6,18,32] are proposed, whose core idea is to design networks to learn spatial consistency between correspondences, enabling efficient inlier identification and reliable hypothesis generation. We compare our approach with these methods in Appendix F.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 578, + 202, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 202, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 202, + 592 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Given two partially overlapped point clouds " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathbf{p}_i\\in \\mathbb{R}^3 |i = 1,2,\\dots ,N\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{\\mathbf{q}_j\\in \\mathbb{R}^3 |j = 1,2,\\dots ,M\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", we aim to estimate a transformation " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = (\\mathbf{R}\\in SO(3),\\mathbf{t}\\in \\mathbb{R}^3)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " that aligns the source point cloud to the target one. To address this problem, we propose a registration framework based on rotation-equivariant network " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "f(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", whose output is equivariant to the SO(3) action " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 113, + 473, + 290 + ], + "blocks": [ + { + "bbox": [ + 142, + 113, + 473, + 290 + ], + "lines": [ + { + "bbox": [ + 142, + 113, + 473, + 290 + ], + "spans": [ + { + "bbox": [ + 142, + 113, + 473, + 290 + ], + "type": "image", + "image_path": "9666065ae66c08d2e36b35726224420c6174f7e3a96a14e18983c6af1aa633f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "text", + "content": "Fig. 3: The framework of our method. Given point clouds " + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "text", + "content": ", a hierarchical backbone based on PARE-Conv is used to extract patch-level and point-level features. Then, we follow a coarse-to-fine scheme to obtain point correspondences using rotation-invariant features. Finally, the feature-based hypothesis proposer leverages the matched rotation-equivariant features " + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{x_{jn}}^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{y_{jn}}^{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "text", + "content": ", which encode fine-grained structure orientations, to generate multiple reliable hypotheses. The best hypothesis is selected and refined as the final solution " + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}^{*},\\mathbf{t}^{*})" + }, + { + "bbox": [ + 130, + 299, + 482, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": "applied on the input point cloud " + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "inline_equation", + "content": "f(\\mathbf{R} \\circ \\mathbf{X}) = \\mathbf{R} \\circ f(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": ". As we can see, rotation-equivariant features " + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "inline_equation", + "content": "f(\\mathbf{R} \\circ \\mathbf{X})" + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": " couple the structure information " + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "inline_equation", + "content": "f(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": " and orientation information " + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": " of point clouds, which can be further decoupled for feature matching and transformation estimation, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 520 + ], + "type": "text", + "content": "The framework of PARE-Net is depicted in Fig. 3. We build a hierarchical backbone based on PARE-Conv to extract both patch-level and point-level rotation invariant/equivariant features. Correspondences are estimated in a coarse-to-fine fashion by using rotation-invariant features. After that, the feature-based hypothesis proposer employs rotation-equivariant features to generate multiple reliable hypotheses, and the best one is selected as the final output " + }, + { + "bbox": [ + 130, + 448, + 482, + 520 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}^{*},\\mathbf{t}^{*})" + }, + { + "bbox": [ + 130, + 448, + 482, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 537, + 394, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 537, + 394, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 537, + 394, + 550 + ], + "type": "text", + "content": "3.1 Position-Aware Rotation-Equivariant Network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "PARE-Conv. Recent point cloud registration approaches [15,25,34,36] mainly exploited KPCov [26] to build backbones. However, KPCov is inherently rotation-sensitive and needs data augmentation to learn an approximate invariant mapping. As a result, models trained on limited samples of continuous SO(3) space are fragile to rotation variations. Moreover, enforcing the network to learn a rotation-invariant mapping inevitably requires more model parameters and also hinders the distinctiveness of features. Inspired by dynamic and conditional convolutions [30,31], we propose position-aware rotation-equivariant convolution to learn distinctive rotation-invariant/equivariant features, as shown in Fig. 3." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "text", + "content": "Given a point cloud " + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathbf{p}_i\\in \\mathbb{R}^3|i = 1,2,\\dots ,N\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "text", + "content": " and attached feature map " + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{\\mathbf{F}_i\\in \\mathbb{R}^{C\\times 3}\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "text", + "content": ", the general convolution of " + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "text", + "content": " by a kernel " + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "text", + "content": " at a point " + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 115, + 479, + 150 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 152, + 481, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 152, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 235, + 152, + 481, + 177 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {F} * g\\right) \\left(\\mathbf {p} _ {i}\\right) = \\sum_ {\\mathbf {p} _ {j} \\in \\mathcal {N} _ {i}} g \\left(\\mathbf {p} _ {j} - \\mathbf {p} _ {i}\\right) \\mathbf {F} _ {j}, \\tag {1}", + "image_path": "e924b87dd3d96eae5cd01fc2122dc7907a3908673be48621edfc8a2af72a7fc2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": " is the K-nearest neighbor set of the point " + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": ". The core problem is the definition of kernel function " + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": ". In vanilla VN [8], it omits the position information and simply uses a linear layer, " + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{F}_i" + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": ", followed by a VN non-linear layer and a VN pooling layer to aggregate local information, where " + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{C' \\times C}" + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": " is a learnable weight matrix. Instead, we define a set of shadow kernel points, equipped with learnable weight matrices " + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{W}_k | k = 1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": ". In image convolution, each pixel corresponds to a convolutional kernel, and the fixed spatial arrangement of the kernels enables the learning of the distribution information of image pixels. Due to the irregularity of point clouds, establishing such point-to-kernel correspondences is intractable. We follow KPConv [26] to establish soft assignments between data points and kernel points and the kernel function " + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 182, + 482, + 316 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 325, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 230, + 325, + 481, + 357 + ], + "type": "interline_equation", + "content": "g \\left(\\mathbf {p} _ {i j}\\right) = \\sum_ {k = 1} ^ {K} \\operatorname {s o f t m a x} (\\gamma \\left(\\mathbf {p} _ {i j}, k\\right)) \\mathbf {W} _ {k}, \\tag {2}", + "image_path": "c1b2e87b8a97d78e072597b80fb4b54833dfef78c0a3c8007bbdce34e1a132d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij} = \\mathbf{p}_j - \\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " is the relative coordinate of " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " is a correlation function. In KPConv [26], it uses a linear correlation " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{p}_{ij},k) = \\max (0,1-\\|\\mathbf{p}_{ij} - \\tilde{\\mathbf{x}}_k\\| / \\sigma)" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_k" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " is the 3D coordinate of the " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": "-th kernel point and " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " is a distance influence factor. However, kernel-to-point distances are sensitive to rotations thus cannot guarantee a rotation-equivariant convolution. Therefore, the key challenge lies in ensuring the rotation equivariance of the convolution while well capturing the spatial information of patches. Inspired by PACConv [30], we use a mini-network that consumes rotation-equivariant spatial features and predicts rotation-invariant correlation scores between data points and kernel points. Note that predicting rotation-invariant scores is necessary to guarantee the rotation-equivariance of the convolution (see Appendix A for the proof). As shown in Fig. 3, we implement this function with 1) a VN-MLP block, learning geometric features " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j \\in \\mathbb{R}^{d \\times 3}" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " from rotation-equivariant spatial statistics, such as " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij}" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\frac{1}{|\\mathcal{N}_i|} \\sum_j \\mathbf{p}_{ij}" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " and their cross product; 2) a L2-Norm layer, computing the magnitude of the vector neurons " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}_j = \\| \\mathbf{X}_j\\|_2 \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": " to transform the feature into a rotation-invariant form; and 3) a MLP block, mapping this feature vector into K dimension to obtain the correlation vector, " + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\gamma_j = \\mathrm{MLP}(\\tilde{\\mathbf{X}}_j) \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 130, + 366, + 482, + 597 + ], + "type": "text", + "content": ". Finally, a softmax operation is leveraged to output correlation scores. As a result, our convolution is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 234, + 606, + 481, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 606, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 234, + 606, + 481, + 632 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {F} * g\\right) \\left(\\mathbf {p} _ {i}\\right) = \\sum_ {\\mathbf {p} _ {j} \\in \\mathcal {N} _ {i}} \\sum_ {k} \\gamma_ {j k} \\mathbf {W} _ {k} \\mathbf {F} _ {j}. \\tag {3}", + "image_path": "d4d41f85eb938a35e8bda5bcf0d1c1dc1eed2e988ffbb440b3d91479b36eb8d3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Compared to KPConv using a linear correlation, our correlation network can learn more complex relationships. Moreover, PARE-Conv predicts correlation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "scores depending on local spatial property, thus can leverage the spatial information better and can provide stronger expressive power to extract more discriminative descriptors. Moreover, it is compatible with DGCNN [29], allowing replacing the node feature " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " with the edge feature " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "[\\mathbf{F}_j - \\mathbf{F}_i,\\mathbf{F}_j]" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". This can further enhance the expressive power, as discussed in the ablation study 4.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": "Fully Convolutional Network. As shown in Fig. 3, we develop a hierarchical convolution network based on PARE-Conv. The original point clouds " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " are downsampled three times to yield sparse superpoints " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": ", whose resolution is " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\frac{1}{2^3}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " of the original point clouds. Consequently, three convolution blocks are leveraged to extract multi-level features. The bottleneck rotation-equivariant features are denoted as " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\hat{\\mathcal{P}}| \\times \\hat{d} \\times 3}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\hat{\\mathcal{Q}}| \\times \\hat{d} \\times 3}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": ", and a VN-invariant layer [8] is used to obtain the rotation-invariant features " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\hat{\\mathcal{P}}| \\times 3\\hat{d}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\hat{\\mathcal{Q}}| \\times 3\\hat{d}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": ". For the decoder, two nearest upsampling blocks are used to get the rotation-equivariant features " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{P}}| \\times \\tilde{d} \\times 3}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{Q}}| \\times \\tilde{d} \\times 3}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " of the first-level downsampled points " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": ". Another VN-invariant layer is leveraged to get their rotation-invariant features " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}^{\\mathcal{P}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{P}}| \\times 3\\tilde{d}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}^{\\mathcal{Q}} \\in \\mathbb{R}^{|\\tilde{\\mathcal{Q}}| \\times 3\\tilde{d}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 326 + ], + "type": "text", + "content": ". For more details on the convolutional block, please refer to Appendix B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": "Following GeoTrans [25], we exploit the point-to-node grouping strategy to assign the points " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " to the superpoints " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": ". The subset of " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " belongs to the superpoint " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}_i" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " and the associated rotation-equivariant and invariant feature matrices of subset points are denoted as " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}_i^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": ". For point cloud " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": ", the points are grouped as " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_i^{\\mathcal{Q}}\\}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": ", and the feature matrices are denoted as " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}_i^{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " in a same way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 417, + 287, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 287, + 430 + ], + "type": "text", + "content": "3.2 Coarse-to-Fine Matching" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 479, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 473 + ], + "type": "text", + "content": "Since our backbone outputs both patch-level and point-level features, we use a coarse-to-fine matching strategy [25,34,35] to filter out non-overlapped regions and estimate more accurate correspondences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 533 + ], + "type": "text", + "content": "Superpoint Matching. Direct matching superpoints by nearest neighbor search lacks robustness to repeated patterns and low-overlapped problems. Therefore, we utilize the Geometric Transformer module [25] presented by Qin et al. to reason the global context of two point clouds, which iteratively uses self and cross-attention to capture intra and inter-point-cloud features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": "We feed the rotation-invariant features " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " and the attached position information " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " into this module to get more discriminative features " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}^{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": ". Then, a Gaussian correlation matrix is computed, followed by a dual normalization operation. The superpoint correspondences are obtained by selecting the top-" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " reliable correspondences " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}} = \\{(\\hat{\\mathbf{p}}_{x_i},\\hat{\\mathbf{q}}_{y_i})|(x_i,y_i)\\in \\arg \\max_{x,y}\\mathbf{S}_{x,y}\\}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 533, + 480, + 605 + ], + "type": "text", + "content": " is the similarity matrix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Improvement. Here, we find this module is sensitive to the overlap distribution of the dataset due to its cross-attention layers. This could lead to a significant performance decrease when there is a substantial difference between the overlap distributions of the test set and the training set. To tackle this issue, we present a simple but effective data augmentation technique, RandomCrop," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where point clouds are randomly cropped to make the model more robust to low-overlapped point cloud pairs. We present more details in the Appendix C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": "Point Matching. When the correspondences " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}_i = (\\hat{\\mathbf{p}}_{x_i},\\hat{\\mathbf{q}}_{y_i})" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " of superpoints are established, we search point-level correspondences within " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{x_i}^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{y_i}^{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": ". We follow [22] that disentangles the similarity and saliency of features by using a matchability head " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_m:\\mathbb{R}^{3\\tilde{d}}\\to \\mathbb{R}^{3\\tilde{d}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " and a saliency head " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s:\\mathbb{R}^{3\\tilde{d}}\\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": ". Specifically, we compute a matching matrix:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 226, + 209, + 481, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 209, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 226, + 209, + 481, + 227 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {x _ {i}, y _ {i}} = \\left(\\mathbf {W} _ {m} \\mathbf {X} _ {x _ {i}} ^ {\\mathcal {P}}\\right) ^ {\\mathrm {T}} \\mathbf {W} _ {m} \\mathbf {X} _ {y _ {i}} ^ {\\mathcal {Q}} / \\sqrt {3 \\tilde {d}}, \\tag {4}", + "image_path": "cc7a2d6b6783878e3143544fa4e10dd6869d67756954506d5a6c4e729320ab94.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 232, + 313, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 313, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 313, + 243 + ], + "type": "text", + "content": "and then the salient scores are predicted:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 249, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 249, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 194, + 249, + 481, + 266 + ], + "type": "interline_equation", + "content": "\\sigma_ {x _ {i}} ^ {\\mathcal {P}} = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {s} \\mathbf {X} _ {x _ {i}} ^ {\\mathcal {P}}\\right), \\sigma_ {y _ {i}} ^ {\\mathcal {Q}} = \\operatorname {S i g m o i d} \\left(\\mathbf {W} _ {s} \\mathbf {X} _ {y _ {i}} ^ {\\mathcal {Q}}\\right). \\tag {5}", + "image_path": "de835d271831ca61af38977c3c3c2147b28d7ccf11de6f82937c18c7f782287b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 270, + 365, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 270, + 365, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 365, + 282 + ], + "type": "text", + "content": "Finally, a soft assignment matrix " + }, + { + "bbox": [ + 131, + 270, + 365, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i" + }, + { + "bbox": [ + 131, + 270, + 365, + 282 + ], + "type": "text", + "content": " is computed by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 289, + 480, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 289, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 227, + 289, + 480, + 304 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {i} = \\sigma_ {x _ {i}} ^ {\\mathcal {P}} \\sigma_ {y _ {i}} ^ {\\mathcal {Q}} \\varphi \\left(\\mathbf {M} _ {x _ {i}, y _ {i}}\\right) _ {x _ {i}} \\varphi \\left(\\mathbf {M} _ {x _ {i}, y _ {i}}\\right) _ {y _ {i}}, \\tag {6}", + "image_path": "e6a38c639e8026ba13a03b9473e20f870a0116fb6bea4a1b91a43187bc9a8c7f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "text", + "content": " is the softmax operation. As a result, we select " + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "text", + "content": " point-level correspondences " + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})\\}" + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "text", + "content": " across all patch correspondence whose assignment scores are the highest. The impact of " + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 130, + 310, + 479, + 346 + ], + "type": "text", + "content": " is analyzed in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 361, + 336, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 336, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 336, + 373 + ], + "type": "text", + "content": "3.3 Feature-based hypothesis Proposer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 349, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 349, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 349, + 415 + ], + "type": "text", + "content": "Since our rotation-equivariant features encode orientation information of local structures, we derive the pose transformation from matched features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "text", + "content": "Hypothesis Generation and Selection. Once obtaining point-wise correspondence set " + }, + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})\\}" + }, + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "text", + "content": ", we leverage their corresponding rotation-equivariant feature maps " + }, + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "inline_equation", + "content": "\\{(\\tilde{\\mathbf{F}}_{x_j}^{\\mathcal{P}},\\tilde{\\mathbf{F}}_{y_j}^{\\mathcal{Q}})\\}" + }, + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "text", + "content": " to generate multiple model hypotheses. For each correspondence " + }, + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{p}}_{x_j},\\tilde{\\mathbf{q}}_{y_j})" + }, + { + "bbox": [ + 130, + 415, + 350, + 501 + ], + "type": "text", + "content": ", a rotation can be estimated by solving:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 160, + 506, + 349, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 506, + 349, + 527 + ], + "spans": [ + { + "bbox": [ + 160, + 506, + 349, + 527 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {R}} _ {j} = \\underset {\\mathbf {R}} {\\arg \\min } \\| \\mathbf {R} \\left(\\tilde {\\mathbf {F}} _ {x _ {j}} ^ {\\mathcal {P}}\\right) ^ {\\mathrm {T}} - \\left(\\tilde {\\mathbf {F}} _ {y _ {j}} ^ {\\mathcal {Q}}\\right) ^ {\\mathrm {T}} \\| _ {2} ^ {2}. \\tag {7}", + "image_path": "43e54ab135432ac7655851d0475734a11fb98d62b062403f16506575e3b5cb71.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 533, + 348, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 348, + 568 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 348, + 568 + ], + "type": "text", + "content": "This can be efficiently solved by SVD [5] in closed form. Subsequently, the translation can be solved by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 199, + 568, + 349, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 568, + 349, + 583 + ], + "spans": [ + { + "bbox": [ + 199, + 568, + 349, + 583 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {t}} _ {j} = \\tilde {\\mathbf {q}} _ {y _ {j}} - \\tilde {\\mathbf {R}} _ {j} \\tilde {\\mathbf {p}} _ {x _ {j}}. \\tag {8}", + "image_path": "b6bf0298bd962793e086119fae38f9f2f57b978dede98e40edff514c39f4ae6f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 586, + 350, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 350, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 350, + 633 + ], + "type": "text", + "content": "As a result, we can get a set of model hypotheses " + }, + { + "bbox": [ + 131, + 586, + 350, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_c = \\{(\\tilde{\\mathbf{R}}_i,\\tilde{\\mathbf{t}}_i)\\}" + }, + { + "bbox": [ + 131, + 586, + 350, + 633 + ], + "type": "text", + "content": ". The hypothesis that has the most inlier support will be selected as the alignment solution:" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 360, + 403, + 472, + 462 + ], + "blocks": [ + { + "bbox": [ + 360, + 403, + 472, + 462 + ], + "lines": [ + { + "bbox": [ + 360, + 403, + 472, + 462 + ], + "spans": [ + { + "bbox": [ + 360, + 403, + 472, + 462 + ], + "type": "image", + "image_path": "872be62db4fc44d150e006ad7bfdfe42a597a8ef8651879900fad93a58f72a6b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 464, + 471, + 472 + ], + "lines": [ + { + "bbox": [ + 360, + 464, + 471, + 472 + ], + "spans": [ + { + "bbox": [ + 360, + 464, + 471, + 472 + ], + "type": "text", + "content": "Source point cloud Target point cloud" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 373, + 476, + 463, + 518 + ], + "blocks": [ + { + "bbox": [ + 373, + 476, + 463, + 518 + ], + "lines": [ + { + "bbox": [ + 373, + 476, + 463, + 518 + ], + "spans": [ + { + "bbox": [ + 373, + 476, + 463, + 518 + ], + "type": "image", + "image_path": "56e8e0b2304dcdfcefe6b980b269d476da04bbe5d1505373cb327ed56ec70f6c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 522, + 471, + 532 + ], + "lines": [ + { + "bbox": [ + 360, + 522, + 471, + 532 + ], + "spans": [ + { + "bbox": [ + 360, + 522, + 471, + 532 + ], + "type": "text", + "content": "Symmetric alignment Correct alignment" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 356, + 541, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 541, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 356, + 541, + 480, + 628 + ], + "type": "text", + "content": "Fig.4: An illustration of incorrect alignment caused by the ambiguity of using coordinates to align. This problem can be solved by using rotation-equivariant features that contain fine-grained local orientation information." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 171, + 637, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 637, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 171, + 637, + 480, + 668 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {R} ^ {*}, \\mathbf {t} ^ {*}\\right) = \\underset {\\tilde {\\mathbf {R}} _ {j}, \\tilde {\\mathbf {t}} _ {j}} {\\arg \\max } \\sum_ {\\left(\\tilde {\\mathbf {p}} _ {x _ {j}}, \\tilde {\\mathbf {q}} _ {y _ {j}}\\right) \\in \\mathcal {C}} \\mathbb {1} \\left(\\| \\tilde {\\mathbf {R}} _ {j} \\tilde {\\mathbf {p}} _ {x _ {j}} + \\tilde {\\mathbf {t}} _ {j} - \\tilde {\\mathbf {q}} _ {y _ {j}} \\| _ {2} ^ {2} < \\tau_ {d}\\right), \\tag {9}", + "image_path": "c8f65a134637bcc54610ff38a6fc3ecabad02f23e334dc67ac976ec9ae42999b.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\tau_{d}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the acceptance radius. The best model hypothesis will be iteratively refined by its inlier correspondences for a more accurate solution, as in [25]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 344 + ], + "type": "text", + "content": "Discussion. Traditional estimators, such as RANSAC [11] randomly sample multiple pairs of correspondences and use their coordinates to estimate the transformation. Only when all the sampled correspondences belong to inliers, a reliable transformation can be estimated. This approach is inefficient and requires multiple iterations. In contrast, we directly utilize rotation-equivariant features to estimate the pose of the point cloud, and each pair of correspondences generates a hypothesis. This approach is more efficient, as depicted in the Fig. 2. Our method achieves a significantly high registration recall by generating only 1000 hypotheses. In contrast, RANSAC is still inferior to our method even with 50,000 iterations. Moreover, using coordinates to solve the transformation may falsely align the point clouds with opposite orientations, due to ignoring the orientation information, as shown in Fig. 4. This problem can be easily addressed by our method as the rotation-equivariant features encode orientations of local structures. In addition, some methods [1, 28] use regressors to implicitly estimate transformations from equivariant features. Instead, our method can directly solve pose estimation through feature alignment, which is conceptually simpler and computationally efficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 365, + 230, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 230, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 230, + 376 + ], + "type": "text", + "content": "3.4 Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "text", + "content": "The network is supervised by three losses " + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_c + \\mathcal{L}_f + \\mathcal{L}_r" + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "text", + "content": ", including superpoint matching loss " + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "text", + "content": ", point matching loss " + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "text", + "content": ", and contrastive rotation loss " + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 130, + 389, + 479, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": "Point Matching Loss " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ". Since point matching is performed within the scopes of matched superpoints, unreliable superpoint matching may lead to sparse supervision signals for point matching. Therefore, we use positive superpoint correspondences computed by ground-truth transformation during the training process. For two grouped point sets " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{x_i}^{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{y_i}^{\\mathcal{Q}}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", the positive correspondences " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i^*" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", having distance errors smaller than a threshold " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "d_p" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", are identified by the groundtruth. The unmatched points are denoted as " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_i" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ". We compute the log-likelihood loss on soft assignment matrix " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " and the salient scores " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\sigma_{x_i}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\sigma_{y_i}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 184, + 521, + 480, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 521, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 184, + 521, + 480, + 582 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {f _ {i}} = - \\frac {1}{| \\mathcal {C} _ {i} ^ {*} |} \\sum_ {(x _ {i}, y _ {i}) \\in \\mathcal {C} _ {i} ^ {*}} \\log \\mathbf {Z} _ {x _ {i}, y _ {i}} - \\frac {1}{2 | \\mathcal {I} _ {i} |} \\sum_ {x _ {i} \\in \\mathcal {I} _ {i}} \\log \\left(1 - \\sigma_ {x _ {i}} ^ {\\mathcal {P}}\\right) \\\\ - \\frac {1}{2 \\left| \\mathcal {J} _ {i} \\right|} \\sum_ {y _ {i} \\in \\mathcal {J} _ {i}} \\log \\left(1 - \\sigma_ {y _ {i}} ^ {Q}\\right). \\tag {10} \\\\ \\end{array}", + "image_path": "fed304f11b1ef2b5a66013ff098b543a2e795873d172af3bb2b10b6339f71b8a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 593, + 384, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 384, + 606 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 384, + 606 + ], + "type": "text", + "content": "The overall point matching loss is a sum of " + }, + { + "bbox": [ + 131, + 593, + 384, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f = \\sum_i\\mathcal{L}_{f_i}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Contrastive Rotation Loss " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ". Rotation-equivariant feature inherently encodes pose information of local geometric structure. However, such features of a positive correspondence inevitably be contaminated by data degradation, such as occlusion, noise and down-sampling errors. To make the features more robust to these troubles, we propose a contrastive rotation loss, to supervise point-level" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 208, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 208, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 208, + 128 + ], + "type": "text", + "content": "correspondences:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 139, + 480, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 139, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 192, + 139, + 480, + 213 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {r _ {i}} = \\frac {1}{| \\mathcal {C} _ {i} ^ {*} | \\cdot \\tilde {d}} \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in \\mathcal {C} _ {i} ^ {*}} \\sum_ {c = 1} ^ {\\tilde {d}} [ \\| \\tilde {\\mathbf {F}} _ {x _ {i}, c} ^ {\\mathcal {P}} \\mathbf {R} _ {g t} ^ {\\mathrm {T}} - \\tilde {\\mathbf {F}} _ {y _ {i}, c} ^ {\\mathcal {Q}} \\| _ {2} ^ {2} - \\alpha ] _ {+} \\tag {11} \\\\ + \\frac {1}{| \\bar {\\mathcal {C}} _ {i} | \\cdot \\tilde {d}} \\sum_ {(x _ {i}, y _ {i}) \\in \\bar {\\mathcal {C}} _ {i}} \\sum_ {c = 1} ^ {\\tilde {d}} [ \\beta - \\| \\tilde {\\mathbf {F}} _ {x _ {i}, c} ^ {\\mathcal {P}} \\mathbf {R} _ {g t} ^ {\\mathrm {T}} - \\tilde {\\mathbf {F}} _ {y _ {i}, c} ^ {\\mathcal {Q}} \\| _ {2} ^ {2} ] _ {+}, \\\\ \\end{array}", + "image_path": "41ef2161842907de65df2872de860e5884ac3114b4edab67731efcee76a6c421.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i^*" + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "content": " is the aforementioned positive correspondence set, " + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{C}}_i" + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "content": " is the negative correspondence set whose point pair has spatial distance greater than " + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "inline_equation", + "content": "d_{n}" + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "inline_equation", + "content": "\\beta = 1.4" + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "content": " are positive and negative margins, and " + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "inline_equation", + "content": "[\\cdot]_{+}" + }, + { + "bbox": [ + 130, + 223, + 481, + 259 + ], + "type": "text", + "content": " is the ReLU operation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "text", + "content": "Superpoint Matching Loss " + }, + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "text", + "content": ". To supervise the rotation-invariant features of superpoints, we utilize the overlap-aware circle loss that considers the overlap ratio of two superpoints to compute more accurate gradients. For more details, please refer to GeoTrans [25]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 327, + 230, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 230, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 230, + 340 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 352, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 425 + ], + "type": "text", + "content": "To evaluate the performance of our method, we compare our method with state-of-the-art approaches and perform extensive experiments on 3DMatch [37], 3DLoMatch [15], and KITTI Odometry [12]. Finally, ablation studies are conducted to analyze the components of our method. All the experiments are conducted with a RTX 3090 GPU with Intel (R) Xeon (R) Silver 4314 CPU. More implementation details are reported in Appendix D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 441, + 367, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 367, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 367, + 453 + ], + "type": "text", + "content": "4.1 Indoor Dataset: 3DMatch & 3DLoMatch" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "content": "DataSet. 3DMatch [37] and 3DLoMatch [15] are collected from indoor RGB-D datasets. 3DMatch contains 62 scenes and there are " + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "inline_equation", + "content": "46/8/8" + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "content": " scenes for training/validation/testing. It consists of point cloud pairs with overlap ratios greater than " + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "content": ". 3DLoMatch is only used for testing that contains point cloud pairs with overlap ratios from " + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 461, + 481, + 533 + ], + "type": "text", + "content": ". We follow the protocols [15, 25, 28] to perform all the experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "content": "Baseline. We compare our method with strong baselines: 1) FCGF [7], Predator [15], GeoTrans [25], and PEAL [36] are advanced scene-level feature extractors, and some of them [25,36] follow a coarse-to-fine manner to extract robust correspondences. PEAL requires overlap prior as input, we follow the office implementation that leverages 3D overlap prior predicted by GeoTrans for a fair comparison. 2) YOHO [27], RoReg [28], and RoITR [35] are also scene-level feature extractors but focus on rotation-invariant/equivariant network designs. 3) SpinNet [2] and BUFFER [1] leverage patch-level feature extractors with elaborate designs for rotation-invariance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Metrics. We follow [15,25] that use three evaluation metrics, including Inlier Ratio (IR), Feature Matching Recall (FMR), and Registration Recall (RR). To" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 147, + 481, + 269 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Evaluation results on 3DMatch and 3DLoMatch. Methods with the RANSAC estimator are marked by " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ", which exploit 5000 points to establish correspondences." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 147, + 481, + 269 + ], + "lines": [ + { + "bbox": [ + 135, + 147, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 135, + 147, + 481, + 269 + ], + "type": "table", + "html": "
MethodSize (MB)3DMatch3DLoMatch
FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)FMR (% ↑)IR (% ↑)RR (% ↑)Time (s ↓)
FCGF° [7]8.7694.731.182.80.1259.49.838.00.13
SpinNet° [2]1.4197.647.588.69.8575.320.559.89.03
YOHO [27]12.3898.264.490.82.8178.925.966.02.62
Predator° [15]7.4396.658.089.00.6478.226.764.40.47
GeoTrans [25]9.8398.170.992.40.1887.443.574.30.17
RoReg [28]12.7198.281.693.02.2782.339.670.12.10
BUFFER [1]0.92--92.10.22--70.30.21
RoITR° [35]10.1098.082.491.90.3689.254.674.10.34
PEAL [36]9.8398.471.094.21.4688.346.078.81.19
Ours3.8498.576.995.00.1788.347.580.50.17
", + "image_path": "e805df25ddd1cbbb824633e98e8bf6506ff488033a20bf6246c62bd84ce1ebc1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 293, + 480, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 364 + ], + "type": "text", + "content": "evaluate the performance of different methods on rotated datasets, RoITR [35] uses two ways to compute the RMSE on the original and rotated datasets, which cannot show the robustness of methods against pose variations. To tackle this issue, we follow [28] to report Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) on both original and rotated datasets. Please refer to the Appendix E for more details about these evaluation metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "text", + "content": "Quantitative results. We report the experimental results in Table 1. Compared with SOTA methods, our approach achieves a significantly high RR on both 3DMatch and 3DLoMatch, while simultaneously maintaining lightweight and computation efficient. PEAL [36] is a formidable competitor that employs an iterative manner for alignment due to the requirements of overlap prior, resulting in significant time overhead. In contrast, our method achieves a speed improvement of 7 times compared to PEAL, while also outperforming it by " + }, + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "text", + "content": " in terms of RR on 3DLoMatch. GeoTrans [25], RoITR [35], and our method all employ a coarse-to-fine matching framework. Despite using a more lightweight feature extractor, our RR is significantly higher than theirs, particularly on 3DLoMatch, where our method outperforms them by " + }, + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 133, + 365, + 481, + 604 + ], + "type": "text", + "content": " respectively. This can be attributed to the ability of our rotation-equivariant network to focus more on learning discriminative structural information. YOHO [27] and RoReg [28] are extremely time-consuming due to the group feature extractor and embedder. Our efficiency advantage stems from two aspects. Firstly, we utilize a lightweight scene-wise feature extractor, which only needs a forward process to extract dense rotation-equivariant features. Secondly, our feature-based hypothesis proposer can simultaneously generate multiple valid hypotheses by using matched rotation-equivariant features. Thus, it is more efficient compared to RANSAC [11] with many approaches [2, 7, 15, 35] used." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 605, + 480, + 665 + ], + "type": "text", + "content": "Experiments on the rotated dataset. To test the robustness against rotations, we follow previous works [28, 35] that evaluate all the methods on the rotated 3DLoMatch, where arbitrary rotations are applied to point clouds. The experimental results are shown in Table 2. As we can see, PEAL experiences a decrease in TR by " + }, + { + "bbox": [ + 133, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 133, + 605, + 480, + 665 + ], + "type": "text", + "content": " on the Rotated 3DLoMatch dataset, while our" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 481, + 282 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Evaluation results on 3DLoMatch and Rotated 3DLoMatch. For easy comparison of the impact of rotation, the numerical changes w.r.t. TR on the rotated dataset are annotated in the top right corner of each value. Methods with rotation invariant/equivariant designs are marked by * symbol." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 481, + 282 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 481, + 282 + ], + "type": "table", + "html": "
MethodSize (MB)3DLoMatchRotated 3DLoMatch
RE (°↓)TE (cm ↓)TR (% ↑)RE (°↓)TE (cm ↓)TR (% ↑)
FCGF [7]8.764.8412.8739.64.7413.3924.5-15.1
PREDATOR [15]7.433.6110.6565.63.5510.3064.0-1.6
GeoTrans [25]9.832.918.7175.42.948.8572.6-2.8
PEAL [36]9.832.848.6481.22.868.5378.7-2.5
YOHO* [27]12.383.5410.3466.63.6110.1667.1+0.5
RoReg* [28]12.713.019.2671.33.039.2871.0-0.3
BUFFER* [1]0.923.039.8674.43.029.9974.7+0.3
RoITR* [35]10.102.959.0375.12.979.0875.5+0.4
Ours*3.842.878.8381.32.848.7181.8+0.5
", + "image_path": "1cf4c04b061ed6ae1bba0a5e8f424b59f85b6a736d0a1eda34bcf4afb9995f18.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "text", + "content": "method demonstrates an increase of " + }, + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "text", + "content": ". This result validates the robustness of our approach against rotations, illustrating its ability to handle such variations. Methods without rotation-invariant design, such as FCGF [7], Predator [15], and GeoTrans [25] also show performance degradation, indicating that networks trained with rotation augmentation are vulnerable to rotation variations. On the other hand, methods such as YOHO [27], RoReg [28], RoITR [35], and BUFFER [1], which incorporate rotation equivariant or invariant designs, demonstrate better robustness against data rotations. However, our method significantly outperforms these methods by " + }, + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "text", + "content": " w.r.t TR on both 3DLoMatch and Rotated 3DLoMatch. This is because our backbone can capture spatial geometric information better, thus extracting more distinctive descriptors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 264, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 264, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 264, + 628 + ], + "type": "text", + "content": "Qualitative results. We show some registration results in Fig. 5, which demonstrate that our method can estimate more accurate patch-level and point-level correspondences and can successfully align point cloud pairs with low-overlapped ratios. Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method suc" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "cessfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity. Please refer to Appendix G for more visualization results." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 273, + 508, + 481, + 609 + ], + "blocks": [ + { + "bbox": [ + 273, + 497, + 478, + 507 + ], + "lines": [ + { + "bbox": [ + 273, + 497, + 478, + 507 + ], + "spans": [ + { + "bbox": [ + 273, + 497, + 478, + 507 + ], + "type": "text", + "content": "Table 3: Evaluation results on KITTI Odometry." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 273, + 508, + 481, + 609 + ], + "lines": [ + { + "bbox": [ + 273, + 508, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 273, + 508, + 481, + 609 + ], + "type": "table", + "html": "
MethodSizeRE (°)TE (cm)TR (%)Time (s)
FCGF [7]8.760.309.596.6-
D3Feat [4]14.080.307.299.8-
Predator [15]22.770.276.899.80.77
SpinNet [2]1.410.479.999.116.24
CoFiNet [34]5.480.418.299.80.59
GeoTrans [25]25.500.236.299.80.26
BUFFER [1]0.920.267.199.80.27
Ours2.080.234.999.80.21
", + "image_path": "57e7c06b187c6799d5c32b4d99473c80cc496d629a69e79dac984c417b94811f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 239 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 239 + ], + "type": "image", + "image_path": "a0e48852e9221594f77b1c4ba1dc3b744b58a355ef96c4fadfef1ae120c4b1d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 301 + ], + "type": "text", + "content": "Fig. 5: Qualitative results. Our method can successfully align low-overlapped pairs with higher IR and PIR (Patch IR). Moreover, GeoTrans [25] and PEAL [36] incorrectly align point clouds with opposite orientations (the first and second rows), our method successfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 327, + 340, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 340, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 340, + 339 + ], + "type": "text", + "content": "4.2 Outdoor Dataset: KITTI Odometry" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 479, + 384 + ], + "type": "text", + "content": "Dataset. We follow previous methods [7,25] to evaluate our method on KITTI Odometry [12]. It comprises 11 urban scene sequences captured by a LiDAR sensor. We use 0-5/6-7/8-10 sequences for training/validation/testing, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 420 + ], + "type": "text", + "content": "Metrics. We use the Rotation Error (RE), Translation Error (TE), and Transformation Recall (TR) to evaluate our method. Following [7,25], the TR is computed with rotation threshold " + }, + { + "bbox": [ + 130, + 384, + 480, + 420 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 130, + 384, + 480, + 420 + ], + "type": "text", + "content": " and translation threshold " + }, + { + "bbox": [ + 130, + 384, + 480, + 420 + ], + "type": "inline_equation", + "content": "2\\mathrm{m}" + }, + { + "bbox": [ + 130, + 384, + 480, + 420 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 421, + 480, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 480, + 468 + ], + "type": "text", + "content": "Experimental results. The experimental results are shown in Table 3. On KITTI dataset, many methods have reached performance saturation, while our approach achieves the highest accuracy when achieving the highest TR. Our method also shows advantages w.r.t. model size and running speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 487, + 237, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 237, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 237, + 499 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "type": "text", + "content": "To better understand the designs of our method, we conduct comprehensive ablation experiments and the evaluated results are reported in Table 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "type": "text", + "content": "Ablation of backbone. To investigate the impact of different convolution methods, we compare our PARE-Conv with VN [8]. We leverage both node features [24] and edge features [29] for evaluation. The experimental results indicate that our approach achieves significant improvement compared to VN whether using node features or edge features. This demonstrates the enhanced representative power of PARE-Conv to learn more distinctive features. We also compare PARE-Conv with KPCnv [26], refer to the Appendix F.5 for details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Ablation of pose estimator. To demonstrate the superiority of our pose estimator, we compared it with RANSAC [11] and LGR [25]. RANSAC randomly selects a few pairs of correspondences to generate hypotheses, while LGR generates a hypothesis using the correspondences within a matched patch pair." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "In contrast, our approach directly generates a hypothesis using the equivariant features of a pair of matched points. As shown in Table 4, our method achieves the highest RR, outperforming RANSAC and LGR by " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " respectively on 3DLoMatch. This validates that estimating poses using rotation-equivariant features based on local structures is more robust than using point-to-point correspondences. Additionally, the results of FMR and IR are identical as we used the same correspondences for different pose estimators." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 242, + 482, + 369 + ], + "blocks": [ + { + "bbox": [ + 210, + 221, + 402, + 232 + ], + "lines": [ + { + "bbox": [ + 210, + 221, + 402, + 232 + ], + "spans": [ + { + "bbox": [ + 210, + 221, + 402, + 232 + ], + "type": "text", + "content": "Table 4: Ablation experiments of our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 242, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 135, + 242, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 135, + 242, + 482, + 369 + ], + "type": "table", + "html": "
Component Method3DMatch3DLoMatch
FMR (% ↑) IR (% ↑) RR (% ↑)FMR (% ↑) IR (% ↑) RR (% ↑)
BackboneVN (Node) [8]98.671.093.487.841.476.6
PARE-Conv (Node)98.974.194.5 +1.188.344.078.8 +2.2
VN (Edge) [8]98.772.494.188.843.577.8
PARE-Conv (Edge)98.576.995.0 +0.988.147.580.5 +2.7
Pose EstimatorRANSAC98.576.993.788.347.577.5
LGR [25]98.576.994.0 +0.388.347.578.9 +1.4
Ours98.576.995.0 +1.388.347.580.5 +3.0
Rotation LossWithout98.374.493.388.745.579.3
L298.665.494.2 +0.985.935.075.4 -3.9
Contrastive98.576.995.0 +1.788.347.580.5 +1.2
", + "image_path": "8413e2cd8d17377a7b45784952ce9c23fef5bfb53bdfe559e1b2ec9fe2b04ba8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 391, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 523 + ], + "type": "text", + "content": "Ablation of rotation loss. The contrastive rotation loss is designed to improve the robustness of rotation-equivariant features to data occlusion, noise, and partial overlap. When the rotation-equivariant features are not supervised by a loss function (Indeed, it is indirectly supervised by the loss function of rotation-invariant features, as the invariant features are derived from the equivariant features), the method shows decent performance on 3DLoMatch but shows poor performance on 3DMatch. When using the L2 loss to supervise equivariant features of positive point pairs, there is a slight improvement in RR on 3DMatch but shows performance degradation on 3DLoMatch. However, when utilizing our contrastive rotation loss, the performance of the method is enhanced on both 3DMatch and 3DLoMatch, demonstrating its superiority." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 542, + 220, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 220, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 220, + 555 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we propose PARE-Net, a lightweight network for fast and robust point cloud registration. We introduce position-aware rotation-equivariant convolution to learn spatial information effectively, thus extracting more distinctive descriptors. We also propose an efficient feature-based hypothesis proposer to generate reliable model hypotheses. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art approaches and shows great robustness to rotation variations. The limitations of our method and future works are discussed in Appendix H." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Science and Technology Major Project of China under Grant No. 2020AAA0108102, the National Natural Science Foundation of China under Grant Nos. 62327808, 62073257 and 62088102, and the Fundamental Research Funds for the Central Universities under Grant No. xzy022024007." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 222, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 222, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 481, + 266 + ], + "type": "text", + "content": "1. Ao, S., Hu, Q., Wang, H., Xu, K., Guo, Y.: BUFFER: Balancing accuracy, efficiency, and generalizability in point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1255-1264 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "text", + "content": "2. Ao, S., Hu, Q., Yang, B., Markham, A., Guo, Y.: SpinNet: Learning a general surface descriptor for 3D point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11753-11762 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 301, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 343 + ], + "type": "text", + "content": "3. Bai, X., Luo, Z., Zhou, L., Chen, H., Li, L., Hu, Z., Fu, H., Tai, C.L.: PointDSC: Robust point cloud registration using deep spatial consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15859-15869 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 345, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 481, + 388 + ], + "type": "text", + "content": "4. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3Feat: Joint learning of dense detection and description of 3D local features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6359-6367 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 389, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 481, + 411 + ], + "type": "text", + "content": "5. Besl, P.J., McKay, N.D.: Method for registration of 3-D shapes. In: Sensor Fusion IV: Control Paradigms and Data Structures. vol. 1611, pp. 586-606 (1992)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 411, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 454 + ], + "type": "text", + "content": "6. Chen, Z., Sun, K., Yang, F., Tao, W.: SC²-PCR: A second order spatial compatibility for efficient and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13221-13231 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 456, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 487 + ], + "type": "text", + "content": "7. Choy, C., Park, J., Koltun, V.: Fully convolutional geometric features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8958-8966 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 488, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 532 + ], + "type": "text", + "content": "8. Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., Guibas, L.J.: Vector Neurons: A general framework for SO(3)-equivariant networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12200-12209 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "type": "text", + "content": "9. Deng, H., Birdal, T., Ilic, S.: PPF-FoldNet: Unsupervised learning of rotation invariant 3D local descriptors. In: Proceedings of the European Conference on Computer Vision. pp. 602-618 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 566, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 566, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 566, + 481, + 599 + ], + "type": "text", + "content": "10. Deng, H., Birdal, T., Ilic, S.: PPFNet: Global context aware local features for robust 3D point matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 195-205 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "11. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. vol. 24, pp. 381-395. ACM New York (1981)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (2012)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Gojcic, Z., Zhou, C., Wegner, J.D., Wieser, A.: The perfect match: 3D point cloud matching with smoothed densities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5545-5554 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "14. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3D local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "15. Huang, S., Gojcic, Z., Usvyatsov, M., Wieser, A., Schindler, K.: PREDATOR: Registration of 3D point clouds with low overlap. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4267-4276 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "16. Kaya, M., Bilge, H.S.: Deep metric learning: A survey. Symmetry 11(9), 1066-1092 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "17. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 153-161 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "18. Lee, J., Kim, S., Cho, M., Park, J.: Deep hough voting for robust global registration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15994-16003 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "19. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "20. Li, Y., Ma, L., Zhong, Z., Liu, F., Chapman, M.A., Cao, D., Li, J.: Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems 32(8), 3412-3432 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 408 + ], + "type": "text", + "content": "21. Lu, W., Zhou, Y., Wan, G., Hou, S., Song, S.: L3-Net: Towards learning based LiDAR localization for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6389-6398 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "22. Philipp, L., Paul-Edouard, S., Marc, P.: LightGlue: Local Feature Matching at Light Speed. In: Proceedings of the IEEE International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "text", + "content": "23. Pomerleau, F., Colas, F., Siegwart, R.: A review of point cloud registration algorithms for mobile robotics. Foundations and Trends in Robotics 4(1), 1-104 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 481, + 498 + ], + "type": "text", + "content": "24. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "text", + "content": "25. Qin, Z., Yu, H., Wang, C., Guo, Y., Peng, Y., Xu, K.: Geometric transformer for fast and robust point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11143-11152 (June 2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "text", + "content": "26. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPConv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6411-6420 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 620 + ], + "type": "text", + "content": "27. Wang, H., Liu, Y., Dong, Z., Wang, W.: You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1630-1641 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "28. Wang, H., Liu, Y., Hu, Q., Wang, B., Chen, J., Dong, Z., Guo, Y., Wang, W., Yang, B.: RoReg: Pairwise point cloud registration with oriented descriptors and local rotations. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 10376-10393 (2023). https://doi.org/10.1109/TPAMI.2023.3244951" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Yao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 468 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph CNN for learning on point clouds. ACM Transactions On Graphics (ToG) 38(5), 1-12 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "30. Xu, M., Ding, R., Zhao, H., Qi, X.: PACConv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3173-3182 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "type": "text", + "content": "31. Yang, B., Bender, G., Le, Q.V., Ngiam, J.: CondConv: Conditionally parameterized convolutions for efficient inference. In: Proceedings of Advances in Neural Information Processing Systems. vol. 32 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 258 + ], + "type": "text", + "content": "32. Yao, R., Du, S., Cui, W., Ye, A., Wen, F., Zhang, H., Tian, Z., Gao, Y.: Hunter: Exploring high-order consistency for point cloud registration with severe outliers. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14760-14776 (2023). https://doi.org/10.1109/TPAMI.2023.3312592" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "text", + "content": "33. Yew, Z.J., Lee, G.H.: REGTR: End-to-end point cloud correspondences with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6677-6686 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "type": "text", + "content": "34. Yu, H., Li, F., Saleh, M., Busam, B., Ilic, S.: CoFiNet: Reliable coarse-to-fine correspondences for robust point cloud registration. In: Proceedings of Advances in Neural Information Processing Systems. pp. 23872-23884 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "type": "text", + "content": "35. Yu, H., Qin, Z., Hou, J., Saleh, M., Li, D., Busam, B., Ilic, S.: Rotation-invariant transformer for point cloud matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5384-5393 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 481, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 481, + 401 + ], + "type": "text", + "content": "36. Yu, J., Ren, L., Zhang, Y., Zhou, W., Lin, L., Dai, G.: PEAL: Prior-embedded explicit attention learning for low-overlap point cloud registration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17702-17711 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 445 + ], + "type": "text", + "content": "37. Zeng, A., Song, S., Nießner, M., Fisher, M., Xiao, J., Funkhouser, T.: 3DMatch: Learning local geometric descriptors from RGB-D reconstructions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1802-1811 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "text", + "content": "38. Zhang, Z., Dai, Y., Sun, J.: Deep learning based point cloud registration: An overview. Virtual Reality and Intelligent Hardware 2(3), 222-246 (2020)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PARE-Net" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_content_list.json b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f52a1b1e0de50a7c306d33d146d7782a095233f5 --- /dev/null +++ b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "PARIS3D: Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amrin Kareem1, Jean Lahoud1, and Hisham Cholakkal1", + "bbox": [ + 279, + 212, + 720, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE", + "bbox": [ + 251, + 239, + 750, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advancements in 3D perception systems have significantly improved their ability to perform visual recognition tasks such as segmentation. However, these systems still heavily rely on explicit human instruction to identify target objects or categories, lacking the capability to actively reason and comprehend implicit user intentions. We introduce a novel segmentation task known as reasoning part segmentation for 3D objects, aiming to output a segmentation mask based on complex and implicit textual queries about specific parts of a 3D object. To facilitate evaluation and benchmarking, we present a large 3D dataset comprising over 60k instructions paired with corresponding ground-truth part segmentation annotations specifically curated for reasoning-based 3D part segmentation. We propose a model that is capable of segmenting parts of 3D objects based on implicit textual queries and generating natural language explanations corresponding to 3D object segmentation requests. Experiments show that our method achieves competitive performance to models that use explicit queries, with the additional abilities to identify part concepts, reason about them, and complement them with world knowledge. Our source code, dataset, and trained models are accessible at this url.", + "bbox": [ + 261, + 292, + 738, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D $\\cdot$ Vision-Language Models $\\cdot$ Reasoning", + "bbox": [ + 261, + 569, + 630, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 611, + 375, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancements in 3D data capture technologies, including LIDARs and RGB-D cameras, have led to a growing demand for automated analysis of 3D point clouds. 3D semantic segmentation, the process of automatically assigning predefined semantic labels to each point in a cloud, is crucial for enabling complex tasks such as scene understanding. Similarly, 3D part segmentation involves further segmenting object instances into their components, such as identifying the handle of a pot or the lid of a bottle. These tasks find applications in various fields including autonomous vehicles, mobile robotics, industrial automation, augmented reality, and medical imagery analysis. While recent advancements in pre-trained 3D representations and the introduction of various 3D datasets have significantly improved 3D perception, the capacity for nuanced reasoning in 3D contexts remains limited. This limitation is primarily due to the lack of comprehensive datasets for reasoning and describing 3D scenes and objects.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4711b544402ea5357b6492736fd8756fbdd1ba22bb757f8c91df9071d0161ac2.jpg", + "image_caption": [ + "Fig. 1: Capabilities of PARIS3D. Parts of 3D objects are segmented based on reasoning, shape, location, material, colour, and concept instructions. Additionally, for the segmentations, PARIS3D can explain why it chose that region, or describe 3D objects with respect to their parts. The original point clouds are on the left. The segmented parts are shown to the right, highlighted in golden colour." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 782, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, 2D images accompanied by textual descriptions have contributed to significant progress in large-scale image language models. Recent methods have emerged, performing segmentation based on 2D images projected from 3D scenes. However, they lack the ability to reason about object concepts and their 3D properties, functions, or parts in a conversational manner.", + "bbox": [ + 212, + 470, + 787, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we define a model as having reasoning capabilities when (i) it can understand implicit instructions, such as referencing properties of an object or its parts without explicit articulation, and (ii) explain or justify its responses, whether generated text or predicted segmentation masks. Such reasoning ability is one of the fundamental cognitive skills possessed by humans, essential for daily activities ranging from locating items to manipulating tools. Previous 3D vision methods are capable of exploring the semantics of a given environment based on instructions containing pre-defined category names. These explicit categories are limiting and do not address cases where instructions can be implicit and may contain complex queries. For example, an automated system might need to find the part of the kettle that can be opened to pour water, without knowing the explicit category name \"lid\". To this end, we introduce the 3D reasoning part segmentation task, which aims at segmenting object parts based on general queries, including color, 3D information, and implicit functionalities.", + "bbox": [ + 212, + 549, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a novel task termed reasoning part segmentation in 3D. This task involves generating a part segmentation mask for a 3D object based on implicit textual queries requiring complex reasoning. These queries go beyond simple references and encompass intricate descriptions, demanding sophisticated reasoning or worldly knowledge. For example, knowing where to hold a kettle or", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which part of a bottle to open in a fine-grained manner. To address this challenging task, we propose PARIS3D (ReasonIng-based 3D PArt Segmentation), a multi-modal Large Language Model (LLM) capable of reasoning on user input text, predicting 3D part segmentation masks, and providing explanations for the model's response. As depicted in Figure 1, PARIS3D adeptly navigates diverse scenarios, encompassing complex 3D reasoning, material, color, shape, and location-based knowledge, providing explanatory responses and detailed descriptions. Additionally, to validate the effectiveness of PARIS3D and support future research, we establish an evaluation benchmark and a dataset named RPSeg3D for reasoning 3D part segmentation. Our RPSeg3D dataset comprises 2624 3D objects and over 60k instructions, providing persuasive evaluation metrics for the task.", + "bbox": [ + 212, + 146, + 787, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our key contributions are the following:", + "bbox": [ + 238, + 328, + 620, + 343 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the reasoning part segmentation task for 3D objects, emphasizing the necessity of reasoning capabilities which is crucial for the development of intelligent perception systems.", + "- We provide a comprehensive dataset named RPSeg3D for reasoning-based 3D part segmentation, comprising over 2624 3D objects and 60k instructions, serving as a useful resource for future research.", + "- We present PARIS3D, a novel approach for 3D part segmentation, and further improve its capabilities by fine-tuning on our RPSeg3D dataset." + ], + "bbox": [ + 225, + 349, + 782, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 492, + 387, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Semantic Segmentation. The challenge of understanding and reasoning within 3D environments has been an ongoing research focus. The goal of 3D semantic segmentation is to acquire semantic predictions for each point in a cloud. Notable contributions include point-based approaches [19, 38], methods which incorporate intricately crafted point convolution methods [46, 50], voxel-based strategies [5, 10], including those that employ 3D sparse convolutions [11] for generating point-wise segmentation outcomes, as well as transformer-based techniques [23]. Multi-view semantic segmentation methods such as DeepViewAgg [41], Diffuser [22, 34], 3D-CG [15], 3D-CLR [16] in 3D vision concentrate on improving representation learning by generating 2D renderings from 3D under multiple view points. These works have shown the effectiveness of multi-view representations in enhancing the performance and robustness of various 3D tasks. Nonetheless, these methods rely on a predefined set of semantic labels, whereas we focus in our proposed method on responding to complex reasoning-based queries and explaining them.", + "bbox": [ + 212, + 522, + 787, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Multimodal Models. Extensive research on Large Language Models (LLMs) has demonstrated reasoning capabilities, prompting an exploration into extending these skills into the visual domain through Large Multimodal Models. LMMs are a highly adaptable and versatile means to perform tasks requiring language and vision capabilities. Prominent models such as BLIP-2 [26], LLaVA [29] and MiniGPT-4 [55] generally employ a dual-phase training process, aligning", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/986ea139b338629237f7acb1203f732929f3b5ad5b322a04537989679ad1bf9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputTask
Point CloudReasoning QueryConversationSegmentationExplanation
SQA3D [33], 3D-VisTA [56]✓✓XXX
ViewRefer [13], Point-Bind [12]✓XXXX
3D-OVS [30]XXX✓X
OpenMask3D [45]✓XX✓X
PLA [6], OpenScene [37]✓XX✓X
Chat-3D [49]✓✓✓X✓
M3DBench [27]✓X✓XX
LLM-Grounder [52]✓X✓X✓
3D-LLM [17]✓✓✓X✓
LL3DA [4], PointLLM [51]✓X✓X✓
PARIS3D✓✓✓✓✓
", + "bbox": [ + 238, + 143, + 763, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Comparison of recent 3D segmentation models and Large Multi-modal Models (LMMs) emphasizing their capabilities for 3D reasoning and conversations. Reasoning query means the model is asked to self-reason a task and either output text or perform an action. Segmentation highlights models that can respond with 3D segmentation masks, and Conversation represents models that can provide a conversation-style answer to the user. Among these, our proposed PARIS3D stands out with comprehensive 3D understanding and reasoning, segmentation in response to natural language queries, and conversational capabilities.", + "bbox": [ + 212, + 315, + 787, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visual representations with the linguistic embeddings of LLMs through extensive image-text and video-text datasets [2,3,28,35,42-44]. Recent efforts have focused on the convergence of multi-modal LLMs with vision tasks, where VisionLLM [48] provides a versatile interaction interface for a spectrum of vision-centric tasks through instruction tuning. Yet, it does not fully leverage the complex reasoning potential of LLMs. GPT4RoI [53] integrates spatial boxes as inputs and trains models on region-text pairs, showcasing a novel approach. Our method aims to benefit from these advances in the LMM space by merging the vision-language abilities of LMMs and the reasoning of LLMs in a novel 3D perception task.", + "bbox": [ + 212, + 474, + 787, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Instructed 3D Tasks. The integration of point clouds with natural language processing has widespread implications, drawing considerable interest in the realms of 3D scene understanding. This fast-growing field promises enhancements in human-robot interaction, metaverse, robotics, and embodied intelligence. Central to the dialogue systems designed for 3D environments are two critical capabilities: perception within three-dimensional spaces and reasoning. Recently, there has been a rise in the number of tasks uniting 3D scenes and language, such as 3D captioning, 3D question answering, 3D situated question answering, embodied Q and A, planning, navigation, 3D assisted multi-turn dialogue, 3D object detection, and scene description. We divide the 3D perception task models into 3 categories (see Table 1. Separated by dotted lines.). The first encompasses models that perform tasks like 3D captioning, situated question answering, and visual grounding [33, 56], visual grounding [14]. These models are capable of providing a word or phrase as its text output. The second category has 3D semantic segmentation models that output 3D segmentation masks. 3D", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OVS [30], Openmask3D [45], OpenScene [37], PLA [7], perform open-vocabulary semantic segmentation for 3D scenes. These methods, on the other hand, cannot provide a conversational output to a user query or explain the reasoning for their tasks. The third comprises models that employ an LLM and perform visual perception tasks such as captioning, scene understanding, and visual grounding, providing conversational outputs [4, 12, 17, 27, 49, 51, 52]. However, they do not perform fine-grained semantic segmentation or reasoning-based 3D vision tasks. As such, we identify a gap in performing 3D segmentation in response to complex natural language prompts instead of a single phrase or word. Distinct from the existing works in this domain, our research is committed to: (1) streamlining the integration of 3D segmentation capabilities into multi-modal LLMs, and (2) enhancing current perception systems with 3D reasoning-based segmentation and explanation abilities, thereby pushing the boundaries of what is achievable in the intersection of language and visual understanding.", + "bbox": [ + 212, + 145, + 787, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Reasoning 3D Part Segmentation", + "text_level": 1, + "bbox": [ + 215, + 381, + 576, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic segmentation involves assigning a semantic label to each geometric primitive, such as points [38], voxels [9], or superpoints [25]. In part segmentation, object instances are decomposed into their components. Given a coloured point cloud P, the goal of a 3D segmentation model is to predict its label for each point. However, in our reasoning segmentation task, we go further to output a 3D segmentation mask M, given an input point cloud P and an implicit query text instruction $x_{txt}$ . The task shares a similar formulation with the referring segmentation task [20], with an additional challenge for the model to reason about the fine-grained parts in response to implicit queries and output the corresponding segmentation mask. The complexity of the query text in reasoning part segmentation is a key differentiator. Instead of providing the names of the parts, the query text may include more intricate expressions that involve an understanding of structures, geometries, and semantics of 3D objects. By introducing this task, we aim to bridge the gap between user intent and system response, enabling more intuitive and dynamic interactions in 3D object perception.", + "bbox": [ + 212, + 414, + 787, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Our RPSeg3D Dataset", + "text_level": 1, + "bbox": [ + 215, + 662, + 450, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the unavailability of established datasets and evaluation benchmarks in the literature, we introduce a dataset, named RPSeg3D, specifically designed for the reasoning 3D part segmentation task. Our dataset comprises 2624 3D objects and over 60k instructions. We use 718 objects and their corresponding instructions as the train set, and the remaining 1906 objects along with their instructions are used for testing. For reliable and fair assessment, we have aligned the 3D objects with those from PartNet-Ensemble [31], annotating them with implicit text instructions and using ground truth labels to generate high-quality target masks. To generate the text instructions corresponding to each 3D object part, we prepare a set of templates, for example, \"Which part of this object", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 253, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13fa119c38da0917bdff7b22fe309d1bce23dd4c0ba4652f6cad6095142ce70a.jpg", + "image_caption": [ + "Normal Query: \"In this toaster, which part can be slid up and down for temperature adjustment?\"" + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 395, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Query: \"What is the gray rectangular part in the middle of the Toaster that you can slide up and down to set the level of toasting?\"", + "bbox": [ + 223, + 207, + 392, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/298e393bbfcde62fc3d36675d0fd6eee35c1d11a8bd896a660ca9b185e28e873.jpg", + "image_caption": [ + "Normal Query: \"Which part of this kettle does the water come out of?\"" + ], + "image_footnote": [], + "bbox": [ + 243, + 237, + 366, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Query: \"How long is the long cylindrical pipe-like brown part of this kettle that liquid comes out of compared to its whole length?\"", + "bbox": [ + 228, + 328, + 397, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e61b493c240b4d2cfb1ddabceb47b73c26a89894af611cde6ec0b09b311d9697.jpg", + "image_caption": [ + "Normal Query: \"Which part of this keyboard used to type information?\"" + ], + "image_footnote": [], + "bbox": [ + 403, + 157, + 488, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Query: \"Can you identify the white rectangular part of the keyboard that makes up about a tenth of its height and is located in the middle?\"", + "bbox": [ + 408, + 207, + 583, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2ffd37fc7c354aa46fbb3abd3024d54e3e4d3dbf5424fad98dc470b8d98310a2.jpg", + "image_caption": [ + "Normal Query: \"Which part of this Bucket provides a grip for carrying it?\"" + ], + "image_footnote": [], + "bbox": [ + 424, + 243, + 478, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Query: \"What is the shape of the gray part of this Bucket located in front of and on the side of it? What is it used for?\"", + "bbox": [ + 408, + 328, + 576, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/987d35ac10ca66fd06153ec7e92d191ab2c68f5db44bcd77ca3addc8cada906f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 157, + 589, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e366f35541b633c3a95203b013b090da930a607a47ae18e81dcad8410b0c0c35.jpg", + "image_caption": [ + "Normal Query: \"Which part of this storage furniture is drawn out and used for storage?\"" + ], + "image_footnote": [], + "bbox": [ + 596, + 152, + 679, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Query: \"In this StorageFurniture, which is the rectangular part used for storage that constitutes about a quarter of its total height?\"", + "bbox": [ + 612, + 207, + 767, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97ba835333ffc25140ea1d4685080dbcb2d8830664ed04d870e4de2affd2b593.jpg", + "image_caption": [ + "Normal Query: \"Which part of this door surrounds the entrance?\"" + ], + "image_footnote": [], + "bbox": [ + 617, + 232, + 665, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Query: \"What is the shape of the brown part of this Door located on top of and on the sides of it? What is it used for?\"", + "bbox": [ + 612, + 328, + 774, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f927279a59d0691d64d076c84329ef5ec182c7f6bf77657851fa278df3d78856.jpg", + "image_caption": [ + "Fig. 2: Examples of the annotated object-instruction pairs for training with two types of queries. On the left is one view of the rendered image from the original point cloud. On the right is the corresponding ground truth segmentation mask, shown in green.", + "Fig. 3: Preparing the instructions of RPSeg3D. Simple templates are provided to GPT-3.5, which populates them with part information-related segmentation instructions. In parallel, colour, shape, location, and dimension-related data is extracted from 3D point clouds. Enriching the instructions with this information and manually checking them for inaccuracies, we obtain the RPSeg3D dataset for part segmentation." + ], + "image_footnote": [], + "bbox": [ + 217, + 419, + 787, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "?\" We leverage GPT-3.5 [36] for building instructions using these templates by supplying information about the part and rephrasing. We also extract 3D information from the point cloud, thus designing the instructions to cover relations, dimensions (length, height), comparisons, colour, texture features, object concepts, and functions. This was further verified manually to avoid inaccurate prompts at test time. The steps are illustrated in Figure 3. To cover 3D object segmentation tasks effectively, our text instructions consist of two types: 1) normal queries; 2) 3D queries; as illustrated in Figure 2. The dataset is partitioned into train and test splits, containing 718 and 1906 3D objects and over 16k and 47k instructions, respectively. As the primary purpose of the dataset is evaluation, the testing set includes a larger number of instructions.", + "bbox": [ + 212, + 585, + 789, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Reasoning 3D Part Segmentation Architecture", + "text_level": 1, + "bbox": [ + 214, + 787, + 648, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method takes as input a dense and coloured 3D point cloud of an object. One of the common methods of 3D analysis is predominantly using point clouds", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to represent 3D data. However, this contrasts with human spatial reasoning processes. Humans typically engage with their surroundings through active exploration, synthesizing perspectives from multiple vantages to form an integrated 3D understanding, rather than processing a 3D environment at one go. Our approach advocates for 3D reasoning derived from multi-view imagery. This approach also benefits from the large-scale 2D pretraining available in vision-language models, similar to previous methods that have taken advantage of pre-trained vision-language models for 3D vision tasks. Thus, we render multiple images $x_{img}$ from K predefined camera poses by rasterization. The camera poses cover all parts of the object since they are uniformly distributed around the input point cloud. Given a complex text instruction $x_{txt}$ along with the images, we feed them into a multimodal LLM, denoted by $F$ , and the visual backbone, $F_{enc}$ . $F$ outputs a text response $\\hat{y}_{txt}$ . In parallel, $F_{enc}$ extracts the visual embeddings, f, from each of the input images $x_{img}$ . The formulation is as follows:", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t x t} = F \\left(x _ {i m g}, x _ {t x t}\\right), f = F _ {e n c} \\left(x _ {i m g}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 372, + 785, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Whenever the LLM is expected to yield a binary segmentation mask, the resultant text output, denoted as $\\hat{y}_{txt}$ is required to an extra token, which sends a request for the segmentation output. Following this, the embedding at the last layer, $\\hat{h}_{seg}$ , corresponding to the additional segmentation token is extracted. This embedding is subsequently processed through an MLP (Multilayer Perceptron) projection layer, represented as $\\gamma$ , to derive $h_{seg}$ and then input to the mask decoder $F_{dec}$ as its prompt embedding, thereby tightly coupling the text output from $F$ and the mask predicted by the decoder. Hence, the mask decoder takes the visual features $f$ along with $\\langle \\mathrm{SEG} \\rangle$ based prompt embeddings to perform segmentation, generating the final segmentation mask, $\\hat{M}$ similar to [24]. The process is as follows:", + "bbox": [ + 212, + 398, + 787, + 566 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh _ {s e g} = \\gamma (\\hat {h} _ {s e g}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 580, + 784, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The architectural specifics of the decoder, $F_{dec}$ , are in accordance with [21].", + "bbox": [ + 220, + 604, + 767, + 619 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} = F _ {\\text {d e c}} \\left(h _ {\\text {s e g}}, f\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 635, + 784, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once the K segmentations are obtained from each camera angle, a 3D semantic voting module computes scores and assigns the semantic labels for each part. We have $K$ semantic segmentation masks $M_{k}$ , where $k$ is the view from which the image was rendered from the point cloud. We aggregate the masks from multiple views and lift them to obtain a three-dimensional semantic segmentation of the original point cloud. To achieve this, we segment the input point cloud $\\mathbf{P}$ into a set of superpoints $SP_{i}$ , similar to [25]. Superpoints refer to a way of representing large 3D point clouds as a collection of interconnected geometrically simple shapes. This representation is advantageous because (i) it considers complete object parts as a whole, making them easier to identify, (ii) this method can provide a detailed description of the relationship between adjacent objects, and (iii) the size of superpoints is determined by the number of simple structures in a", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 253, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfcc961beadcc49a08be7acda431c29681d340f37872c49c648b7baa2c3a8e99.jpg", + "image_caption": [ + "Fig. 4: Overview of the proposed reasoning-based 3D part segmentation approach named PARIS3D. It comprises four subsequent steps: (i) The 3D point cloud is rendered into K multi-view images $x_{img}$ using a renderer. (ii) These images are passed through a frozen vision backbone ( $F_{enc}$ ) and multi-modal large language model ( $F$ ) of the reasoning module. $F$ also accepts the text query $x_{txt}$ , and produces text outputs corresponding to each view. (iii) The decoder decodes the final layer embedding which contains the extra token, thus producing K segmentation masks. (iv) Finally, a mask to 3D segmentation algorithm lifts the projections back into 3D and a view-guided scoring module is used to obtain the final text response." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scene, rather than the total number of points, making the representation several orders of magnitude smaller, thus more efficient to work with. Each superpoint contains points with similar normals and colours, suggesting they belong to the same instance. This superpoint-based part labelling not only conserves computational resources but also potentially improves performance by leveraging 3D priors. In the 3D semantic voting step, we assign semantic labels for each superpoint through a voting mechanism similar to [31]. This technique leverages information from multiple views and the superpoints, so that even if one mask covers irrelevant points, the aggregation of masks from all views counters the effect of such errors. For a superpoint $SP_{i}$ and a part category $j$ , a score $s_{i,j}$ is computed reflecting the proportion of visible points present in the predicted segmentation masks of category $j$ in each view:", + "bbox": [ + 212, + 438, + 787, + 619 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = \\frac {\\sum_ {k} \\sum_ {p \\in S P _ {i}} \\left[ V _ {k} (p) \\right] \\left[ \\exists b \\in M _ {k} ^ {j} : I _ {m} (p) \\right]}{\\sum_ {k} \\sum_ {p \\in S P _ {i}} \\left[ V _ {k} (p) \\right]} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 627, + 785, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $[\\cdot ]$ is the Iverson bracket, which turns true predicates into 1 and false into 0. $M_k^j$ , represents the predicted mask of category j in view k. $V_{k}(p)$ indicates the visibility of 3D point p in the view k. $I_{m}(p)$ checks if the projection of point p in view k falls within the mask m. The superpoint is then assigned the semantic label of the part category with the highest score.", + "bbox": [ + 212, + 672, + 787, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To provide the most relevant text explanation corresponding to the 3D masks, we propose a view-guided scoring module. It assigns a score to each view based on its correspondence with the final semantic labels. The best text explanation is chosen based on the highest score. Specifically, for each view, we compute the intersection over union (IoU) of the mask with the projections of the final 3D semantic labels output by the model, and the corresponding text explanation is", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "saved. This is called the view-guided score of the explanation. At the end of the score computation for all the masks, we choose the text explanation with the highest score, and in turn, the most correspondence with the output labels.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Training the PARIS3D Architecture", + "text_level": 1, + "bbox": [ + 214, + 209, + 562, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Data Formulation The training data consists of 718 3D objects rendered into multiview images, resulting in over 16k image-instruction pairs. Out of the 718, 360 objects with their instructions are used for training the model and 358 objects with their instructions are used for validation. Each image is provided with an annotation file, which has instructions corresponding to the 3D point cloud that the image came from, the name of the image, and its ground truth mask. Generation of instructions follows the same steps as Section 3.1 illustrated in Figure 3. Images in the training set may have more than one instruction in its \"instruction\" field. This is helpful to introduce diversity as users may randomly select one as the reasoning query during training, thus obtaining a better model.", + "bbox": [ + 212, + 229, + 784, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Distillation-Based Explanation Refining Training the multiview model requires explanatory data as well as training image-instruction pairs. To build the explanatory data, we built an annotation pipeline using a distillation approach. Using the multi-view images as input to the teacher model [24], we generate explanations for each part of the object. These explanations serve as our pseudolabels which we use as ground truth explanations of the student model. We augment these annotations with 3D features extracted from the 3D point clouds from which the images were rendered. The 3D feature-augmented explanations contain appropriate responses and critical elements differentiating object parts from each other such as location, size, shape, material, and colour.", + "bbox": [ + 212, + 412, + 784, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Objective We leverage LoRA [18] to perform efficient fine-tuning of the pretrained multi-modal LLM $F$ [29] to retain its generalization capability. We completely freeze the vision backbone $F_{enc}$ . The decoder $F_{dec}$ is fully fine-tuned. The word embeddings of the LLM and the projection layer of $\\gamma$ are also trainable, allowing the model to learn the specific meanings and semantic concepts of parts. The model is trained end-to-end using the text generation loss $L_{txt}$ and the segmentation mask loss $L_{mask}$ . The overall objective is given as:", + "bbox": [ + 212, + 580, + 784, + 686 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {t x t} L _ {t x t} + \\lambda_ {m a s k} L _ {m a s k} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 695, + 784, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $L_{mask}$ encourages the model to produce high-quality segmentation results. It is a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights $\\lambda_{bce}$ and $\\lambda_{dice}$ , given by:", + "bbox": [ + 212, + 715, + 782, + 760 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m a s k} = \\lambda_ {b c e} B C E (\\hat {M}, M) + \\lambda_ {d i c e} D I C E (\\hat {M}, M) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 766, + 784, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $y_{txt}$ and $M$ are the ground-truth targets. $L_{txt}$ is the auto-regressive cross-entropy loss for text generation. It is computed as:", + "bbox": [ + 212, + 787, + 782, + 818 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t x t} = C E \\left(\\hat {y} _ {t x t}, y _ {t x t}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 825, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 253, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We perform quantitative and qualitative evaluation of PARIS3D on our dataset for reasoning-based semantic segmentation.", + "bbox": [ + 215, + 176, + 784, + 207 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details and Metrics For the experiments, we follow [24] where the multi-modal LLM F is LLaVA-13B-v1-1 [29] and the vision backbone $F_{enc}$ is the ViT-H SAM. The projection layer of $\\gamma$ is an MLP with channels of [256, 4096, 4096]. We use our dataset RPSeg3D which contains coloured point clouds and rendered 2D images of them. Using Pytorch3D [40], each input point cloud is rendered into $\\mathrm{K} = 10$ colour images. The fine-tuning scripts for the LLaVA and SAM architecture are based on DeepSpeed [1] engine. We adopt the settings of [24] for the optimizer (AdamW [32]) and its learning rate (0.0003). Similar steps are followed for the learning rate scheduler (WarmupDecayLR), text generation loss $\\lambda_{txtgen}$ weight (1.0) and the mask loss $\\lambda_{mask}$ weight (1.0), the BCE loss $\\lambda_{bce}$ (2.0), dice loss $\\lambda_{dice}$ (0.5), batch size per device (2), and the gradient accumulation step (10). The semantic segmentation metric used is category mIoU, following [31]. It is calculated as follows: first, mIoU for each part category is calculated for all test objects. Then, part mIoUs that belong to each object category are averaged to compute the object category mIoU.", + "bbox": [ + 215, + 229, + 787, + 457 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Reasoning Part Segmentation Results", + "text_level": 1, + "bbox": [ + 215, + 479, + 573, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We establish the reasoning part segmentation task on our dataset RPSeg3D. Table 2 shows the results of reasoning part segmentation. We observe that without any fine-tuning, the model's performance is low compared to its fine-tuned counterparts. The general observation is that 3D queries help the model to output better masks. We compare our method to two baselines. The first baseline is LISA [24] applied to the multiple views without any finetuning. The second baseline consists of LISA finetuned on few-shot part segmentation data. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts.", + "bbox": [ + 215, + 503, + 787, + 640 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 3D Semantic Segmentation Comparison with Existing Models", + "text_level": 1, + "bbox": [ + 215, + 662, + 769, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 3 shows the results of semantic segmentation compared to existing methods. Our method has better performance than all the fully supervised baselines and achieves competitive results with [31]. For this baseline, we observe that the model with prompt tuning done separately on each category achieves impressive few-shot performance but when unified into a single model for all categories, the performance is only slightly better than its zero-shot performance $(27.2\\%)$ . This is attributed to the redundant part names across object categories, hindering the model's learning of the semantic meanings of part names. This results in performance to drop significantly from when only one category is learnt per model. To prompt for segmentation, the baseline models [38,39,54] are provided with the", + "bbox": [ + 215, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/718045743a50b104b3a246e96408e3e19f5929d490123c10e06c10de5368c7e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodValTest
Normal Query3D QueryOverallNormal Query3D QueryOverall
LISA-MV16.6020.1618.3817.6020.5719.08
LISA-MV (ft)50.4350.2850.3550.7550.8150.78
PARIS3D55.3355.5055.4255.9457.6056.77
", + "bbox": [ + 272, + 143, + 728, + 215 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5790435ce1b42ee27cac1f0d110118b9f49056ec08481571be46c827fcfe78c9.jpg", + "table_caption": [ + "Table 2: Results of reasoning part segmentation. LISA-MV [24] is LISA in multi-view setting without fine-tuning. LISA-MV (ft) is the experiment in which it has been fine-tuned on few-shot part segmentation data. Our proposed PARIS3D method has been fine-tuned with 3D queries and explanations. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts. Here Test is the test set of 1906 shapes and Val is the validation set of 358 3D shapes with their instructions." + ], + "table_footnote": [], + "table_body": "
3D DataMethodOverlapping CategoriesNon-overlapping CategoriesOverall(45)
Bottle ChairDisplay DoorKnife LampStorage-FurnitureTableOverall(17)CameraCartDis-PenserKettleKitchenPotOvenSuit-caseToasterOverall(28)
Extra data(45x8+28k)PointNet++ [38]48.884.778.445.735.468.046.963.755.66.56.412.120.915.834.340.614.725.436.8
PointNeXt [39]68.491.889.443.858.764.968.552.158.533.236.326.045.157.037.813.58.345.150.2
SoftGroup [47]41.488.362.153.131.382.260.254.850.223.623.918.957.445.513.618.326.430.738.1
Few-shot(45x8)PartSLIP* [31]83.485.384.840.865.266.053.642.456.358.388.173.777.069.673.570.460.061.359.4
PointNet++ [38]27.042.230.220.522.210.58.47.318.19.711.67.028.631.719.43.30.021.820.4
PointNeXt [39]67.665.153.746.359.755.420.622.139.226.047.722.660.566.036.814.50.041.540.6
SoftGroup [47]20.880.539.716.338.338.318.924.932.828.640.842.960.754.835.629.814.841.138.0
ACD [8]22.439.029.218.939.613.77.613.519.210.131.519.440.251.88.913.20.025.623.2
Prototype [54]60.170.867.333.450.438.230.225.741.132.036.853.462.763.336.535.510.146.344.3
PartSLIP+64.869.559.524.534.537.132.040.135.325.575.715.630.558.431.149.46.626.729.9
Ours84.081.070.168.447.261.239.445.155.129.371.740.159.378.859.161.624.959.157.6
", + "bbox": [ + 217, + 338, + 785, + 459 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Comparison to previous 3D part segmentation methods. Object category mIoU(%) is shown. In the $45 \\times 8 + 28k$ setting, baseline models use an additional $28k$ training shapes for 17 overlapping object categories. These are categories present in common with PartNet dataset. For the remaining 28 non-overlapping object categories, there are only 8 shapes per object category during training. PartSLIP* indicates that one model has been trained for each category. + shows our implementation of PartSLIP where one model is trained for all the categories together.", + "bbox": [ + 212, + 460, + 787, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "class IDs of the parts to be segmented (e.g: 0 to 103 including the background). In [31], the model is provided with the names of the parts to be segmented (e.g. \"seat\", \"arm\", \"back\" of a chair). Our method, PARIS3D, uses language instructions to prompt segmentation. For a fair comparison, we use hand-crafted prompts with short instructions containing generic concept and location-based clues about the part to be segmented.", + "bbox": [ + 212, + 598, + 784, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 710, + 388, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Instruction Rephrasing. We perform additional experiments with varying instruction types to study the merits of using rephrased prompts in training the model (Table 4 here) compared to using only category names and manually generated prompts for each category. These experimental results on the test set show that GPT-generated prompts provide superior performance, both without and with color and 3D information. In both cases, performance improves from using category names to manual prompts to GPT prompts, indicating the strong", + "bbox": [ + 212, + 734, + 784, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 254, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ac5ab5acba333b66e1e8a5cb61a47bda707edcfb258d110a6284e5907a7a48d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 145, + 410, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0f56d92919fbc3932df4a802f31d19c15fd6833471ea0cf2d590b80d0507578f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 146, + 782, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1a5802da97d5113cb9abba88298ae92f1d2536859b17bd1b759cca8ab4591baa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 227, + 392, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1ca4b3e964b5a6dc61f92d73bd65c63cf5a35e3087d8bb638f6dca8bffc4ff37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 228, + 539, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/55bcaa69480aaf021c4dc74292ae372faa0329104848855a490f3e59fdea417e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 228, + 782, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/543cbfc730632ebead5d5c98ce5fa329cc302306772b6087ad824e2e6ae43be2.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of PARIS3D's performance. We showcase examples from three tasks: reasoning 3D object part segmentation, object description, and reasoning question-answering, demonstrating its capabilities in offering in-depth reasoning, 3D understanding, part segmentation, and conversational abilities." + ], + "image_footnote": [], + "bbox": [ + 233, + 299, + 387, + 369 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0be43e828d54c1968801690d851e0786415948b72579b18cad191020617ebdac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 299, + 549, + 368 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e9a017b4ea1cd2dae4e60beb7941e7bd55908b4f20a915e9dffde443b08ce804.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 299, + 782, + 369 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/54a7566db3918461817c020de6a41db47e77a69caa1047600da7d95ec8b4d971.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Instruction TypeCategory NameManually GeneratedGPT-generatedCategory NameManually generatedGPT-generated
Colour + 3DXXX✓✓✓
Overall mIoU38.041.546.743.052.457.6
", + "bbox": [ + 250, + 448, + 759, + 511 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: A study on the effectiveness of rephrased prompts on the overall model performance. Here, the ability of the model to accurately segment an object part based on an implicit (reasoning-based) description is evaluated for different types of training prompts. The results show the superior performance of GPT-generated prompts over manual prompts, both with and without colour and 3D information.", + "bbox": [ + 212, + 512, + 787, + 582 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "generalization capabilities introduced by GPT-based prompts. For instance, using color and 3D information prompts along with 'manual prompts' resulted in an overall mIoU of $52.4\\%$ . However, using GPT to expand simple templates into varied prompts for describing object part categories resulted in about a $5.2\\%$ improvement over the manually curated training templates (last vs second-last column).", + "bbox": [ + 212, + 626, + 787, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "One Model vs Multiple Models. One of the baseline models [31] trained a model for each of the 45 categories, loading one at a time to evaluate each category. When we replicate this setting for PARIS3D, there are significant jumps in performance as shown in Table 5. However, one-model-for-one-category does not offer a generalizable solution in a real-world problem setting, where multiple object categories and their parts need to be analysed.", + "bbox": [ + 212, + 719, + 787, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Number of Rendered Views. The effect of incrementing the number of views rendered from the input point cloud is shown in Table 6. With only 1 view in", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a8d260ebf372b042f542b6c6f2f017bdfe43f65899f3efaf92f17529c30c71be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodChair DispenserKeyboard EyeglassesesBucket
PartSLIP [31]85.373.753.688.336.5
PARIS3D86.575.688.292.283.9
", + "bbox": [ + 299, + 143, + 702, + 195 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "formation provided, the segmentations are not meaningful. With 5 views, the segmentation quality improves, and improves further with an increase in the number of views. For each point cloud, we use 10 rendered views in our main experiments.", + "bbox": [ + 212, + 320, + 782, + 380 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Single prompt vs Multiple prompts. In Table 7, we show ablation experiments regarding the number of training prompts. Single prompt refers to one concept-based reasoning training instruction that follows a template. In further experiments, we use the template to generate multiple (about 5-6) rephrased instructions for each part category using [36]. Training with these multiple training prompts per part category results in a better-performing model than with a single prompt.", + "bbox": [ + 212, + 380, + 784, + 484 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2ec45b4d75a2f6a56d1bc9ea375fc1139d27234dcf000b72993376d400f576c1.jpg", + "table_caption": [ + "Table 5: Results of PARIS3D on one model trained for each of 5 categories. In [31], 45 models were trained and each point cloud was tested by loading its corresponding model to perform evaluation. Repeating this exact setting by training a model on each category for PARIS3D, we easily gain $+1.2\\%$ , $+1.9\\%$ , $+34.6\\%$ , $+3.9\\%$ , and $+47.4\\%$ improvements on the tested categories." + ], + "table_footnote": [], + "table_body": "
No. of viewsBottleBucketCamera
11.11.11.2
523.740.112.1
1084.884.929.3
", + "bbox": [ + 225, + 527, + 467, + 595 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5d55b7cced23566582289d728d1c3fa4283c01be08aa473100c07aef470a9725.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No. of promptsBottleKnifeClockChair
173.844.345.569.5
5-683.952.751.378.7
", + "bbox": [ + 475, + 529, + 751, + 580 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Ablation experiments on the number of prompts provided in train-number of rendered views. ing data.", + "bbox": [ + 215, + 595, + 733, + 623 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 700, + 419, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Figure 5, we provide examples of PARIS3D output for 3D object part segmentation tasks. PARIS3D is capable of segmenting fine-grained details in categories like buttons of a remote or legs of an eyeglass from implicit queries that expect it to do self-reasoning to generate the right answer, and then segment the part. Moreover, it can perform tasks like object description and question-answering with regard to object parts and give detailed answers. This demonstrates its conversational, reasoning, and part segmentation capabilities.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 253, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8540031978837ec9a5e228819188bef2cdbdecf1064807b26ae5dcb6fc535c09.jpg", + "image_caption": [ + "Fig. 6: Qualitative results on real-world examples of point clouds. The input point cloud is shown on the left and the segmentation mask is in yellow on the right. Our model generalizes to point clouds with a significant domain gap from those it was trained on and displays impressive results with segmenting small object parts such as the furniture's knob in this example." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 787, + 260 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Generalizability to Real-World Examples", + "text_level": 1, + "bbox": [ + 215, + 359, + 602, + 376 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Most existing 3D tasks and their corresponding models can usually only handle cases within the same distribution of the training sets without generalization, since they are sensitive to the format of the input and significant domain gap between synthetic experiments and real-world examples. Thus, it can be difficult to use them in use cases involving real point clouds from an everyday setting. To demonstrate the generalizability of PARIS3D to data derived from the real world, we perform our 3D segmentation on real point clouds shot using a smartphone's LiDAR sensor, as suggested by [31]. In Figure 6, we show qualitative examples of passing the fused point clouds through the PARIS3D architecture to obtain part segmentation labels as in the previous experiments without much drop in performance.", + "bbox": [ + 212, + 382, + 787, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 569, + 359, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce a novel challenge within 3D segmentation, reasoning-based part segmentation. This task requires models to infer, reason, and explain based on implicit user instructions, making it considerably more complex than the regular 3D referring segmentation task. We introduce a dataset for this task, RPSeg3D, to enable effective evaluation. We believe this dataset will play a crucial role in fostering the growth of technologies in this area. Additionally, we outline a pipeline that integrates 3D segmentation capabilities into multimodal Large Language Models (LLMs), showcasing our model, PARIS3D, which exhibits competitive performance. It additionally demonstrates the ability to identify part concepts, reason about them, and complement them with world knowledge. However, we identify limitations: Although the model has improved generalizability, there is a possibility that the model may occasionally result in inaccurate predictions while being prompted with inappropriate, incorrect, or biased text inputs. Also, the model in its current form cannot perform instance segmentation. This is a direction for future research as we expand the dataset to accommodate such tasks.", + "bbox": [ + 212, + 598, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aminabadi, R.Y., Rajbhandari, S., Zhang, M., Awan, A.A., Li, C., Li, D., Zheng, E., Rasley, J., Smith, S., Ruwase, O., He, Y.: Deepspeed inference: Enabling efficient inference of transformer models at unprecedented scale (2022)", + "2. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval (2022)", + "3. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts (2021)", + "4. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: L13da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning (2023)", + "5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3070-3079. IEEE Computer Society, Los Alamitos, CA, USA (jun 2019). https://doi.org/10.1109/CVPR.2019.00319, https://doi.ieeecomputersociety.org/10.1109/CVPR.2019.00319", + "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "7. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7010-7019 (June 2023)", + "8. Gadelha, M., RoyChowdhury, A., Sharma, G., Kalogerakis, E., Cao, L., Learned-Miller, E., Wang, R., Maji, S.: Label-efficient learning on point clouds using approximate convex decompositions. In: European Conference on Computer Vision (ECCV) (2020)", + "9. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks (2017)", + "10. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9224-9232 (2018). https://doi.org/10.1109/CVPR.2018.00961", + "11. Graham, B., van der Maaten, L.: Submanifold sparse convolutional networks (2017)", + "12. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-lm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following (2023)", + "13. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15372-15383 (October 2023)", + "14. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance (2023)", + "15. Hong, Y., Du, Y., Lin, C., Tenenbaum, J.B., Gan, C.: 3d concept grounding on neural fields. NeurIPS (2022)", + "16. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + ], + "bbox": [ + 218, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. NeurIPS (2023)", + "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9", + "19. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2018). https://doi.org/10.1109/CVPR.2018.00278", + "20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 787-798. Association for Computational Linguistics, Doha, Qatar (Oct 2014). https://doi.org/10.3115/v1/D14-1086, https://aclanthology.org/D14-1086", + "21. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything (2023)", + "22. Kundu, A., Yin, X., Fathi, A., Ross, D., Brewington, B., Funkhouser, T., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16. pp. 518-535. Springer (2020)", + "23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8500-8509 (2022)", + "24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)", + "25. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs (2018)", + "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models (2023)", + "27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts (2023)", + "28. Lin, T.Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P., Ramanan, D., Zitnick, C.L., Dollar, P.: Microsoft coco: Common objects in context (2015)", + "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)", + "30. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 53433-53456. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/a76b693f36916a5ed84d6e5b39a0dc03-Paper-Conference.pdf", + "31. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21736-21746 (June 2023)", + "32. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019)", + "33. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=IDJx97BC38" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13589-13595 (2021). https://doi.org/10.1109/ICRA48506.2021.9561801", + "35. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips (2019)", + "36. OpenAI: Gpt-4 technical report (2023)", + "37. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "38. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation (2017)", + "39. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "40. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d (2020)", + "41. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022)", + "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)", + "43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs (2021)", + "44. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)", + "45. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)", + "46. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. Proceedings of the IEEE International Conference on Computer Vision (2019)", + "47. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)", + "48. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., Dai, J.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks (2023)", + "49. Wang, Z., Huang, H., Zhao, Y., Zhang, Z., Zhao, Z.: Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769 (2023)", + "50. Xu, M., Ding, R., Zhao, H., Qi, X.: Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021)", + "51. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent (2023)", + "53. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Liu, Y., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest (2023)", + "54. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)", + "55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023)", + "56. Ziyu, Z., Xiaojian, M., Yixin, C., Zhidong, D., Siyuan, H., Qing, L.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 327 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Kareem et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_model.json b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a0528359c5a9ee5dfe331bcd718a16d250a9d957 --- /dev/null +++ b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_model.json @@ -0,0 +1,2414 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "PARIS3D: Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.213, + 0.721, + 0.228 + ], + "angle": 0, + "content": "Amrin Kareem1, Jean Lahoud1, and Hisham Cholakkal1" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.24, + 0.751, + 0.255 + ], + "angle": 0, + "content": "Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.293, + 0.74, + 0.555 + ], + "angle": 0, + "content": "Abstract. Recent advancements in 3D perception systems have significantly improved their ability to perform visual recognition tasks such as segmentation. However, these systems still heavily rely on explicit human instruction to identify target objects or categories, lacking the capability to actively reason and comprehend implicit user intentions. We introduce a novel segmentation task known as reasoning part segmentation for 3D objects, aiming to output a segmentation mask based on complex and implicit textual queries about specific parts of a 3D object. To facilitate evaluation and benchmarking, we present a large 3D dataset comprising over 60k instructions paired with corresponding ground-truth part segmentation annotations specifically curated for reasoning-based 3D part segmentation. We propose a model that is capable of segmenting parts of 3D objects based on implicit textual queries and generating natural language explanations corresponding to 3D object segmentation requests. Experiments show that our method achieves competitive performance to models that use explicit queries, with the additional abilities to identify part concepts, reason about them, and complement them with world knowledge. Our source code, dataset, and trained models are accessible at this url." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.57, + 0.631, + 0.585 + ], + "angle": 0, + "content": "Keywords: 3D \\(\\cdot\\) Vision-Language Models \\(\\cdot\\) Reasoning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.612, + 0.377, + 0.627 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The rapid advancements in 3D data capture technologies, including LIDARs and RGB-D cameras, have led to a growing demand for automated analysis of 3D point clouds. 3D semantic segmentation, the process of automatically assigning predefined semantic labels to each point in a cloud, is crucial for enabling complex tasks such as scene understanding. Similarly, 3D part segmentation involves further segmenting object instances into their components, such as identifying the handle of a pot or the lid of a bottle. These tasks find applications in various fields including autonomous vehicles, mobile robotics, industrial automation, augmented reality, and medical imagery analysis. While recent advancements in pre-trained 3D representations and the introduction of various 3D datasets have significantly improved 3D perception, the capacity for nuanced reasoning in 3D contexts remains limited. This limitation is primarily due to the lack of comprehensive datasets for reasoning and describing 3D scenes and objects." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.128 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.784, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.789, + 0.442 + ], + "angle": 0, + "content": "Fig. 1: Capabilities of PARIS3D. Parts of 3D objects are segmented based on reasoning, shape, location, material, colour, and concept instructions. Additionally, for the segmentations, PARIS3D can explain why it chose that region, or describe 3D objects with respect to their parts. The original point clouds are on the left. The segmented parts are shown to the right, highlighted in golden colour." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.789, + 0.547 + ], + "angle": 0, + "content": "In contrast, 2D images accompanied by textual descriptions have contributed to significant progress in large-scale image language models. Recent methods have emerged, performing segmentation based on 2D images projected from 3D scenes. However, they lack the ability to reason about object concepts and their 3D properties, functions, or parts in a conversational manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.789, + 0.763 + ], + "angle": 0, + "content": "Here, we define a model as having reasoning capabilities when (i) it can understand implicit instructions, such as referencing properties of an object or its parts without explicit articulation, and (ii) explain or justify its responses, whether generated text or predicted segmentation masks. Such reasoning ability is one of the fundamental cognitive skills possessed by humans, essential for daily activities ranging from locating items to manipulating tools. Previous 3D vision methods are capable of exploring the semantics of a given environment based on instructions containing pre-defined category names. These explicit categories are limiting and do not address cases where instructions can be implicit and may contain complex queries. For example, an automated system might need to find the part of the kettle that can be opened to pour water, without knowing the explicit category name \"lid\". To this end, we introduce the 3D reasoning part segmentation task, which aims at segmenting object parts based on general queries, including color, 3D information, and implicit functionalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we introduce a novel task termed reasoning part segmentation in 3D. This task involves generating a part segmentation mask for a 3D object based on implicit textual queries requiring complex reasoning. These queries go beyond simple references and encompass intricate descriptions, demanding sophisticated reasoning or worldly knowledge. For example, knowing where to hold a kettle or" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.326 + ], + "angle": 0, + "content": "which part of a bottle to open in a fine-grained manner. To address this challenging task, we propose PARIS3D (ReasonIng-based 3D PArt Segmentation), a multi-modal Large Language Model (LLM) capable of reasoning on user input text, predicting 3D part segmentation masks, and providing explanations for the model's response. As depicted in Figure 1, PARIS3D adeptly navigates diverse scenarios, encompassing complex 3D reasoning, material, color, shape, and location-based knowledge, providing explanatory responses and detailed descriptions. Additionally, to validate the effectiveness of PARIS3D and support future research, we establish an evaluation benchmark and a dataset named RPSeg3D for reasoning 3D part segmentation. Our RPSeg3D dataset comprises 2624 3D objects and over 60k instructions, providing persuasive evaluation metrics for the task." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.329, + 0.621, + 0.344 + ], + "angle": 0, + "content": "In summary, our key contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.351, + 0.784, + 0.396 + ], + "angle": 0, + "content": "- We introduce the reasoning part segmentation task for 3D objects, emphasizing the necessity of reasoning capabilities which is crucial for the development of intelligent perception systems." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.397, + 0.784, + 0.44 + ], + "angle": 0, + "content": "- We provide a comprehensive dataset named RPSeg3D for reasoning-based 3D part segmentation, comprising over 2624 3D objects and 60k instructions, serving as a useful resource for future research." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.441, + 0.784, + 0.471 + ], + "angle": 0, + "content": "- We present PARIS3D, a novel approach for 3D part segmentation, and further improve its capabilities by fine-tuning on our RPSeg3D dataset." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.351, + 0.784, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.388, + 0.509 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.749 + ], + "angle": 0, + "content": "3D Semantic Segmentation. The challenge of understanding and reasoning within 3D environments has been an ongoing research focus. The goal of 3D semantic segmentation is to acquire semantic predictions for each point in a cloud. Notable contributions include point-based approaches [19, 38], methods which incorporate intricately crafted point convolution methods [46, 50], voxel-based strategies [5, 10], including those that employ 3D sparse convolutions [11] for generating point-wise segmentation outcomes, as well as transformer-based techniques [23]. Multi-view semantic segmentation methods such as DeepViewAgg [41], Diffuser [22, 34], 3D-CG [15], 3D-CLR [16] in 3D vision concentrate on improving representation learning by generating 2D renderings from 3D under multiple view points. These works have shown the effectiveness of multi-view representations in enhancing the performance and robustness of various 3D tasks. Nonetheless, these methods rely on a predefined set of semantic labels, whereas we focus in our proposed method on responding to complex reasoning-based queries and explaining them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Large Multimodal Models. Extensive research on Large Language Models (LLMs) has demonstrated reasoning capabilities, prompting an exploration into extending these skills into the visual domain through Large Multimodal Models. LMMs are a highly adaptable and versatile means to perform tasks requiring language and vision capabilities. Prominent models such as BLIP-2 [26], LLaVA [29] and MiniGPT-4 [55] generally employ a dual-phase training process, aligning" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.128 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.144, + 0.764, + 0.315 + ], + "angle": 0, + "content": "
MethodInputTask
Point CloudReasoning QueryConversationSegmentationExplanation
SQA3D [33], 3D-VisTA [56]✓✓XXX
ViewRefer [13], Point-Bind [12]✓XXXX
3D-OVS [30]XXX✓X
OpenMask3D [45]✓XX✓X
PLA [6], OpenScene [37]✓XX✓X
Chat-3D [49]✓✓✓X✓
M3DBench [27]✓X✓XX
LLM-Grounder [52]✓X✓X✓
3D-LLM [17]✓✓✓X✓
LL3DA [4], PointLLM [51]✓X✓X✓
PARIS3D✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.316, + 0.788, + 0.428 + ], + "angle": 0, + "content": "Table 1: Comparison of recent 3D segmentation models and Large Multi-modal Models (LMMs) emphasizing their capabilities for 3D reasoning and conversations. Reasoning query means the model is asked to self-reason a task and either output text or perform an action. Segmentation highlights models that can respond with 3D segmentation masks, and Conversation represents models that can provide a conversation-style answer to the user. Among these, our proposed PARIS3D stands out with comprehensive 3D understanding and reasoning, segmentation in response to natural language queries, and conversational capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.612 + ], + "angle": 0, + "content": "visual representations with the linguistic embeddings of LLMs through extensive image-text and video-text datasets [2,3,28,35,42-44]. Recent efforts have focused on the convergence of multi-modal LLMs with vision tasks, where VisionLLM [48] provides a versatile interaction interface for a spectrum of vision-centric tasks through instruction tuning. Yet, it does not fully leverage the complex reasoning potential of LLMs. GPT4RoI [53] integrates spatial boxes as inputs and trains models on region-text pairs, showcasing a novel approach. Our method aims to benefit from these advances in the LMM space by merging the vision-language abilities of LMMs and the reasoning of LLMs in a novel 3D perception task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Language Instructed 3D Tasks. The integration of point clouds with natural language processing has widespread implications, drawing considerable interest in the realms of 3D scene understanding. This fast-growing field promises enhancements in human-robot interaction, metaverse, robotics, and embodied intelligence. Central to the dialogue systems designed for 3D environments are two critical capabilities: perception within three-dimensional spaces and reasoning. Recently, there has been a rise in the number of tasks uniting 3D scenes and language, such as 3D captioning, 3D question answering, 3D situated question answering, embodied Q and A, planning, navigation, 3D assisted multi-turn dialogue, 3D object detection, and scene description. We divide the 3D perception task models into 3 categories (see Table 1. Separated by dotted lines.). The first encompasses models that perform tasks like 3D captioning, situated question answering, and visual grounding [33, 56], visual grounding [14]. These models are capable of providing a word or phrase as its text output. The second category has 3D semantic segmentation models that output 3D segmentation masks. 3D" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.36 + ], + "angle": 0, + "content": "OVS [30], Openmask3D [45], OpenScene [37], PLA [7], perform open-vocabulary semantic segmentation for 3D scenes. These methods, on the other hand, cannot provide a conversational output to a user query or explain the reasoning for their tasks. The third comprises models that employ an LLM and perform visual perception tasks such as captioning, scene understanding, and visual grounding, providing conversational outputs [4, 12, 17, 27, 49, 51, 52]. However, they do not perform fine-grained semantic segmentation or reasoning-based 3D vision tasks. As such, we identify a gap in performing 3D segmentation in response to complex natural language prompts instead of a single phrase or word. Distinct from the existing works in this domain, our research is committed to: (1) streamlining the integration of 3D segmentation capabilities into multi-modal LLMs, and (2) enhancing current perception systems with 3D reasoning-based segmentation and explanation abilities, thereby pushing the boundaries of what is achievable in the intersection of language and visual understanding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.382, + 0.577, + 0.4 + ], + "angle": 0, + "content": "3 Reasoning 3D Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.642 + ], + "angle": 0, + "content": "Semantic segmentation involves assigning a semantic label to each geometric primitive, such as points [38], voxels [9], or superpoints [25]. In part segmentation, object instances are decomposed into their components. Given a coloured point cloud P, the goal of a 3D segmentation model is to predict its label for each point. However, in our reasoning segmentation task, we go further to output a 3D segmentation mask M, given an input point cloud P and an implicit query text instruction \\( x_{txt} \\). The task shares a similar formulation with the referring segmentation task [20], with an additional challenge for the model to reason about the fine-grained parts in response to implicit queries and output the corresponding segmentation mask. The complexity of the query text in reasoning part segmentation is a key differentiator. Instead of providing the names of the parts, the query text may include more intricate expressions that involve an understanding of structures, geometries, and semantics of 3D objects. By introducing this task, we aim to bridge the gap between user intent and system response, enabling more intuitive and dynamic interactions in 3D object perception." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.452, + 0.679 + ], + "angle": 0, + "content": "3.1 Our RPSeg3D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Considering the unavailability of established datasets and evaluation benchmarks in the literature, we introduce a dataset, named RPSeg3D, specifically designed for the reasoning 3D part segmentation task. Our dataset comprises 2624 3D objects and over 60k instructions. We use 718 objects and their corresponding instructions as the train set, and the remaining 1906 objects along with their instructions are used for testing. For reliable and fair assessment, we have aligned the 3D objects with those from PartNet-Ensemble [31], annotating them with implicit text instructions and using ground truth labels to generate high-quality target masks. To generate the text instructions corresponding to each 3D object part, we prepare a set of templates, for example, \"Which part of this object" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.128 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.15, + 0.396, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.188, + 0.394, + 0.204 + ], + "angle": 0, + "content": "Normal Query: \"In this toaster, which part can be slid up and down for temperature adjustment?\"" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.208, + 0.393, + 0.23 + ], + "angle": 0, + "content": "3D Query: \"What is the gray rectangular part in the middle of the Toaster that you can slide up and down to set the level of toasting?\"" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.238, + 0.367, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.309, + 0.397, + 0.325 + ], + "angle": 0, + "content": "Normal Query: \"Which part of this kettle does the water come out of?\"" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.329, + 0.398, + 0.352 + ], + "angle": 0, + "content": "3D Query: \"How long is the long cylindrical pipe-like brown part of this kettle that liquid comes out of compared to its whole length?\"" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.158, + 0.489, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.189, + 0.582, + 0.203 + ], + "angle": 0, + "content": "Normal Query: \"Which part of this keyboard used to type information?\"" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.208, + 0.584, + 0.23 + ], + "angle": 0, + "content": "3D Query: \"Can you identify the white rectangular part of the keyboard that makes up about a tenth of its height and is located in the middle?\"" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.244, + 0.48, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.309, + 0.579, + 0.325 + ], + "angle": 0, + "content": "Normal Query: \"Which part of this Bucket provides a grip for carrying it?\"" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.329, + 0.578, + 0.351 + ], + "angle": 0, + "content": "3D Query: \"What is the shape of the gray part of this Bucket located in front of and on the side of it? What is it used for?\"" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.159, + 0.591, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.154, + 0.681, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.189, + 0.776, + 0.203 + ], + "angle": 0, + "content": "Normal Query: \"Which part of this storage furniture is drawn out and used for storage?\"" + }, + { + "type": "text", + "bbox": [ + 0.613, + 0.208, + 0.769, + 0.23 + ], + "angle": 0, + "content": "3D Query: \"In this StorageFurniture, which is the rectangular part used for storage that constitutes about a quarter of its total height?\"" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.233, + 0.666, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.309, + 0.78, + 0.325 + ], + "angle": 0, + "content": "Normal Query: \"Which part of this door surrounds the entrance?\"" + }, + { + "type": "text", + "bbox": [ + 0.613, + 0.329, + 0.775, + 0.351 + ], + "angle": 0, + "content": "3D Query: \"What is the shape of the brown part of this Door located on top of and on the sides of it? What is it used for?\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.368, + 0.789, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: Examples of the annotated object-instruction pairs for training with two types of queries. On the left is one view of the rendered image from the original point cloud. On the right is the corresponding ground truth segmentation mask, shown in green." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.42, + 0.789, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.496, + 0.789, + 0.567 + ], + "angle": 0, + "content": "Fig. 3: Preparing the instructions of RPSeg3D. Simple templates are provided to GPT-3.5, which populates them with part information-related segmentation instructions. In parallel, colour, shape, location, and dimension-related data is extracted from 3D point clouds. Enriching the instructions with this information and manually checking them for inaccuracies, we obtain the RPSeg3D dataset for part segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.79, + 0.768 + ], + "angle": 0, + "content": "?\" We leverage GPT-3.5 [36] for building instructions using these templates by supplying information about the part and rephrasing. We also extract 3D information from the point cloud, thus designing the instructions to cover relations, dimensions (length, height), comparisons, colour, texture features, object concepts, and functions. This was further verified manually to avoid inaccurate prompts at test time. The steps are illustrated in Figure 3. To cover 3D object segmentation tasks effectively, our text instructions consist of two types: 1) normal queries; 2) 3D queries; as illustrated in Figure 2. The dataset is partitioned into train and test splits, containing 718 and 1906 3D objects and over 16k and 47k instructions, respectively. As the primary purpose of the dataset is evaluation, the testing set includes a larger number of instructions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.649, + 0.804 + ], + "angle": 0, + "content": "3.2 Reasoning 3D Part Segmentation Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our method takes as input a dense and coloured 3D point cloud of an object. One of the common methods of 3D analysis is predominantly using point clouds" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "to represent 3D data. However, this contrasts with human spatial reasoning processes. Humans typically engage with their surroundings through active exploration, synthesizing perspectives from multiple vantages to form an integrated 3D understanding, rather than processing a 3D environment at one go. Our approach advocates for 3D reasoning derived from multi-view imagery. This approach also benefits from the large-scale 2D pretraining available in vision-language models, similar to previous methods that have taken advantage of pre-trained vision-language models for 3D vision tasks. Thus, we render multiple images \\( x_{img} \\) from K predefined camera poses by rasterization. The camera poses cover all parts of the object since they are uniformly distributed around the input point cloud. Given a complex text instruction \\( x_{txt} \\) along with the images, we feed them into a multimodal LLM, denoted by \\( F \\), and the visual backbone, \\( F_{enc} \\). \\( F \\) outputs a text response \\( \\hat{y}_{txt} \\). In parallel, \\( F_{enc} \\) extracts the visual embeddings, f, from each of the input images \\( x_{img} \\). The formulation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.373, + 0.786, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t x t} = F \\left(x _ {i m g}, x _ {t x t}\\right), f = F _ {e n c} \\left(x _ {i m g}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.568 + ], + "angle": 0, + "content": "Whenever the LLM is expected to yield a binary segmentation mask, the resultant text output, denoted as \\(\\hat{y}_{txt}\\) is required to an extra token, which sends a request for the segmentation output. Following this, the embedding at the last layer, \\(\\hat{h}_{seg}\\), corresponding to the additional segmentation token is extracted. This embedding is subsequently processed through an MLP (Multilayer Perceptron) projection layer, represented as \\(\\gamma\\), to derive \\(h_{seg}\\) and then input to the mask decoder \\(F_{dec}\\) as its prompt embedding, thereby tightly coupling the text output from \\(F\\) and the mask predicted by the decoder. Hence, the mask decoder takes the visual features \\(f\\) along with \\(\\langle \\mathrm{SEG} \\rangle\\) based prompt embeddings to perform segmentation, generating the final segmentation mask, \\(\\hat{M}\\) similar to [24]. The process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.581, + 0.785, + 0.598 + ], + "angle": 0, + "content": "\\[\nh _ {s e g} = \\gamma (\\hat {h} _ {s e g}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.605, + 0.768, + 0.621 + ], + "angle": 0, + "content": "The architectural specifics of the decoder, \\( F_{dec} \\), are in accordance with [21]." + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.636, + 0.785, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\hat {M} = F _ {\\text {d e c}} \\left(h _ {\\text {s e g}}, f\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Once the K segmentations are obtained from each camera angle, a 3D semantic voting module computes scores and assigns the semantic labels for each part. We have \\(K\\) semantic segmentation masks \\(M_{k}\\), where \\(k\\) is the view from which the image was rendered from the point cloud. We aggregate the masks from multiple views and lift them to obtain a three-dimensional semantic segmentation of the original point cloud. To achieve this, we segment the input point cloud \\(\\mathbf{P}\\) into a set of superpoints \\(SP_{i}\\), similar to [25]. Superpoints refer to a way of representing large 3D point clouds as a collection of interconnected geometrically simple shapes. This representation is advantageous because (i) it considers complete object parts as a whole, making them easier to identify, (ii) this method can provide a detailed description of the relationship between adjacent objects, and (iii) the size of superpoints is determined by the number of simple structures in a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.293, + 0.789, + 0.419 + ], + "angle": 0, + "content": "Fig. 4: Overview of the proposed reasoning-based 3D part segmentation approach named PARIS3D. It comprises four subsequent steps: (i) The 3D point cloud is rendered into K multi-view images \\( x_{img} \\) using a renderer. (ii) These images are passed through a frozen vision backbone (\\( F_{enc} \\)) and multi-modal large language model (\\( F \\)) of the reasoning module. \\( F \\) also accepts the text query \\( x_{txt} \\), and produces text outputs corresponding to each view. (iii) The decoder decodes the final layer embedding which contains the extra token, thus producing K segmentation masks. (iv) Finally, a mask to 3D segmentation algorithm lifts the projections back into 3D and a view-guided scoring module is used to obtain the final text response." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.789, + 0.621 + ], + "angle": 0, + "content": "scene, rather than the total number of points, making the representation several orders of magnitude smaller, thus more efficient to work with. Each superpoint contains points with similar normals and colours, suggesting they belong to the same instance. This superpoint-based part labelling not only conserves computational resources but also potentially improves performance by leveraging 3D priors. In the 3D semantic voting step, we assign semantic labels for each superpoint through a voting mechanism similar to [31]. This technique leverages information from multiple views and the superpoints, so that even if one mask covers irrelevant points, the aggregation of masks from all views counters the effect of such errors. For a superpoint \\( SP_{i} \\) and a part category \\( j \\), a score \\( s_{i,j} \\) is computed reflecting the proportion of visible points present in the predicted segmentation masks of category \\( j \\) in each view:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.628, + 0.787, + 0.667 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = \\frac {\\sum_ {k} \\sum_ {p \\in S P _ {i}} \\left[ V _ {k} (p) \\right] \\left[ \\exists b \\in M _ {k} ^ {j} : I _ {m} (p) \\right]}{\\sum_ {k} \\sum_ {p \\in S P _ {i}} \\left[ V _ {k} (p) \\right]} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Here, \\([\\cdot ]\\) is the Iverson bracket, which turns true predicates into 1 and false into 0. \\(M_k^j\\), represents the predicted mask of category j in view k. \\(V_{k}(p)\\) indicates the visibility of 3D point p in the view k. \\(I_{m}(p)\\) checks if the projection of point p in view k falls within the mask m. The superpoint is then assigned the semantic label of the part category with the highest score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To provide the most relevant text explanation corresponding to the 3D masks, we propose a view-guided scoring module. It assigns a score to each view based on its correspondence with the final semantic labels. The best text explanation is chosen based on the highest score. Specifically, for each view, we compute the intersection over union (IoU) of the mask with the projections of the final 3D semantic labels output by the model, and the corresponding text explanation is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "saved. This is called the view-guided score of the explanation. At the end of the score computation for all the masks, we choose the text explanation with the highest score, and in turn, the most correspondence with the output labels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.21, + 0.563, + 0.226 + ], + "angle": 0, + "content": "3.3 Training the PARIS3D Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.785, + 0.395 + ], + "angle": 0, + "content": "Training Data Formulation The training data consists of 718 3D objects rendered into multiview images, resulting in over 16k image-instruction pairs. Out of the 718, 360 objects with their instructions are used for training the model and 358 objects with their instructions are used for validation. Each image is provided with an annotation file, which has instructions corresponding to the 3D point cloud that the image came from, the name of the image, and its ground truth mask. Generation of instructions follows the same steps as Section 3.1 illustrated in Figure 3. Images in the training set may have more than one instruction in its \"instruction\" field. This is helpful to introduce diversity as users may randomly select one as the reasoning query during training, thus obtaining a better model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.785, + 0.565 + ], + "angle": 0, + "content": "Distillation-Based Explanation Refining Training the multiview model requires explanatory data as well as training image-instruction pairs. To build the explanatory data, we built an annotation pipeline using a distillation approach. Using the multi-view images as input to the teacher model [24], we generate explanations for each part of the object. These explanations serve as our pseudolabels which we use as ground truth explanations of the student model. We augment these annotations with 3D features extracted from the 3D point clouds from which the images were rendered. The 3D feature-augmented explanations contain appropriate responses and critical elements differentiating object parts from each other such as location, size, shape, material, and colour." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.785, + 0.688 + ], + "angle": 0, + "content": "Objective We leverage LoRA [18] to perform efficient fine-tuning of the pretrained multi-modal LLM \\( F \\) [29] to retain its generalization capability. We completely freeze the vision backbone \\( F_{enc} \\). The decoder \\( F_{dec} \\) is fully fine-tuned. The word embeddings of the LLM and the projection layer of \\( \\gamma \\) are also trainable, allowing the model to learn the specific meanings and semantic concepts of parts. The model is trained end-to-end using the text generation loss \\( L_{txt} \\) and the segmentation mask loss \\( L_{mask} \\). The overall objective is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.696, + 0.785, + 0.71 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {t x t} L _ {t x t} + \\lambda_ {m a s k} L _ {m a s k} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.784, + 0.761 + ], + "angle": 0, + "content": "Here, \\( L_{mask} \\) encourages the model to produce high-quality segmentation results. It is a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights \\( \\lambda_{bce} \\) and \\( \\lambda_{dice} \\), given by:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.767, + 0.785, + 0.783 + ], + "angle": 0, + "content": "\\[\nL _ {m a s k} = \\lambda_ {b c e} B C E (\\hat {M}, M) + \\lambda_ {d i c e} D I C E (\\hat {M}, M) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.789, + 0.784, + 0.819 + ], + "angle": 0, + "content": "where \\( y_{txt} \\) and \\( M \\) are the ground-truth targets. \\( L_{txt} \\) is the auto-regressive cross-entropy loss for text generation. It is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.827, + 0.785, + 0.841 + ], + "angle": 0, + "content": "\\[\nL _ {t x t} = C E \\left(\\hat {y} _ {t x t}, y _ {t x t}\\right) \\tag {7}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.177, + 0.785, + 0.208 + ], + "angle": 0, + "content": "We perform quantitative and qualitative evaluation of PARIS3D on our dataset for reasoning-based semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.23, + 0.788, + 0.458 + ], + "angle": 0, + "content": "Implementation Details and Metrics For the experiments, we follow [24] where the multi-modal LLM F is LLaVA-13B-v1-1 [29] and the vision backbone \\(F_{enc}\\) is the ViT-H SAM. The projection layer of \\(\\gamma\\) is an MLP with channels of [256, 4096, 4096]. We use our dataset RPSeg3D which contains coloured point clouds and rendered 2D images of them. Using Pytorch3D [40], each input point cloud is rendered into \\(\\mathrm{K} = 10\\) colour images. The fine-tuning scripts for the LLaVA and SAM architecture are based on DeepSpeed [1] engine. We adopt the settings of [24] for the optimizer (AdamW [32]) and its learning rate (0.0003). Similar steps are followed for the learning rate scheduler (WarmupDecayLR), text generation loss \\(\\lambda_{txtgen}\\) weight (1.0) and the mask loss \\(\\lambda_{mask}\\) weight (1.0), the BCE loss \\(\\lambda_{bce}\\) (2.0), dice loss \\(\\lambda_{dice}\\) (0.5), batch size per device (2), and the gradient accumulation step (10). The semantic segmentation metric used is category mIoU, following [31]. It is calculated as follows: first, mIoU for each part category is calculated for all test objects. Then, part mIoUs that belong to each object category are averaged to compute the object category mIoU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.48, + 0.575, + 0.496 + ], + "angle": 0, + "content": "4.1 Reasoning Part Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.505, + 0.788, + 0.641 + ], + "angle": 0, + "content": "We establish the reasoning part segmentation task on our dataset RPSeg3D. Table 2 shows the results of reasoning part segmentation. We observe that without any fine-tuning, the model's performance is low compared to its fine-tuned counterparts. The general observation is that 3D queries help the model to output better masks. We compare our method to two baselines. The first baseline is LISA [24] applied to the multiple views without any finetuning. The second baseline consists of LISA finetuned on few-shot part segmentation data. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.771, + 0.68 + ], + "angle": 0, + "content": "4.2 3D Semantic Segmentation Comparison with Existing Models" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Table 3 shows the results of semantic segmentation compared to existing methods. Our method has better performance than all the fully supervised baselines and achieves competitive results with [31]. For this baseline, we observe that the model with prompt tuning done separately on each category achieves impressive few-shot performance but when unified into a single model for all categories, the performance is only slightly better than its zero-shot performance \\((27.2\\%)\\). This is attributed to the redundant part names across object categories, hindering the model's learning of the semantic meanings of part names. This results in performance to drop significantly from when only one category is learnt per model. To prompt for segmentation, the baseline models [38,39,54] are provided with the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.255, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.144, + 0.73, + 0.216 + ], + "angle": 0, + "content": "
MethodValTest
Normal Query3D QueryOverallNormal Query3D QueryOverall
LISA-MV16.6020.1618.3817.6020.5719.08
LISA-MV (ft)50.4350.2850.3550.7550.8150.78
PARIS3D55.3355.5055.4255.9457.6056.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.216, + 0.785, + 0.311 + ], + "angle": 0, + "content": "Table 2: Results of reasoning part segmentation. LISA-MV [24] is LISA in multi-view setting without fine-tuning. LISA-MV (ft) is the experiment in which it has been fine-tuned on few-shot part segmentation data. Our proposed PARIS3D method has been fine-tuned with 3D queries and explanations. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts. Here Test is the test set of 1906 shapes and Val is the validation set of 358 3D shapes with their instructions." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.339, + 0.787, + 0.46 + ], + "angle": 0, + "content": "
3D DataMethodOverlapping CategoriesNon-overlapping CategoriesOverall(45)
Bottle ChairDisplay DoorKnife LampStorage-FurnitureTableOverall(17)CameraCartDis-PenserKettleKitchenPotOvenSuit-caseToasterOverall(28)
Extra data(45x8+28k)PointNet++ [38]48.884.778.445.735.468.046.963.755.66.56.412.120.915.834.340.614.725.436.8
PointNeXt [39]68.491.889.443.858.764.968.552.158.533.236.326.045.157.037.813.58.345.150.2
SoftGroup [47]41.488.362.153.131.382.260.254.850.223.623.918.957.445.513.618.326.430.738.1
Few-shot(45x8)PartSLIP* [31]83.485.384.840.865.266.053.642.456.358.388.173.777.069.673.570.460.061.359.4
PointNet++ [38]27.042.230.220.522.210.58.47.318.19.711.67.028.631.719.43.30.021.820.4
PointNeXt [39]67.665.153.746.359.755.420.622.139.226.047.722.660.566.036.814.50.041.540.6
SoftGroup [47]20.880.539.716.338.338.318.924.932.828.640.842.960.754.835.629.814.841.138.0
ACD [8]22.439.029.218.939.613.77.613.519.210.131.519.440.251.88.913.20.025.623.2
Prototype [54]60.170.867.333.450.438.230.225.741.132.036.853.462.763.336.535.510.146.344.3
PartSLIP+64.869.559.524.534.537.132.040.135.325.575.715.630.558.431.149.46.626.729.9
Ours84.081.070.168.447.261.239.445.155.129.371.740.159.378.859.161.624.959.157.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Table 3: Comparison to previous 3D part segmentation methods. Object category mIoU(%) is shown. In the \\(45 \\times 8 + 28k\\) setting, baseline models use an additional \\(28k\\) training shapes for 17 overlapping object categories. These are categories present in common with PartNet dataset. For the remaining 28 non-overlapping object categories, there are only 8 shapes per object category during training. PartSLIP* indicates that one model has been trained for each category. + shows our implementation of PartSLIP where one model is trained for all the categories together." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.689 + ], + "angle": 0, + "content": "class IDs of the parts to be segmented (e.g: 0 to 103 including the background). In [31], the model is provided with the names of the parts to be segmented (e.g. \"seat\", \"arm\", \"back\" of a chair). Our method, PARIS3D, uses language instructions to prompt segmentation. For a fair comparison, we use hand-crafted prompts with short instructions containing generic concept and location-based clues about the part to be segmented." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.389, + 0.726 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Instruction Rephrasing. We perform additional experiments with varying instruction types to study the merits of using rephrased prompts in training the model (Table 4 here) compared to using only category names and manually generated prompts for each category. These experimental results on the test set show that GPT-generated prompts provide superior performance, both without and with color and 3D information. In both cases, performance improves from using category names to manual prompts to GPT prompts, indicating the strong" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.128 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.146, + 0.411, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.147, + 0.784, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.228, + 0.393, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.229, + 0.54, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.229, + 0.784, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.3, + 0.388, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.3, + 0.55, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.3, + 0.784, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.789, + 0.44 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of PARIS3D's performance. We showcase examples from three tasks: reasoning 3D object part segmentation, object description, and reasoning question-answering, demonstrating its capabilities in offering in-depth reasoning, 3D understanding, part segmentation, and conversational abilities." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.449, + 0.76, + 0.512 + ], + "angle": 0, + "content": "
Instruction TypeCategory NameManually GeneratedGPT-generatedCategory NameManually generatedGPT-generated
Colour + 3DXXX✓✓✓
Overall mIoU38.041.546.743.052.457.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.513, + 0.789, + 0.583 + ], + "angle": 0, + "content": "Table 4: A study on the effectiveness of rephrased prompts on the overall model performance. Here, the ability of the model to accurately segment an object part based on an implicit (reasoning-based) description is evaluated for different types of training prompts. The results show the superior performance of GPT-generated prompts over manual prompts, both with and without colour and 3D information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.718 + ], + "angle": 0, + "content": "generalization capabilities introduced by GPT-based prompts. For instance, using color and 3D information prompts along with 'manual prompts' resulted in an overall mIoU of \\(52.4\\%\\). However, using GPT to expand simple templates into varied prompts for describing object part categories resulted in about a \\(5.2\\%\\) improvement over the manually curated training templates (last vs second-last column)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "One Model vs Multiple Models. One of the baseline models [31] trained a model for each of the 45 categories, loading one at a time to evaluate each category. When we replicate this setting for PARIS3D, there are significant jumps in performance as shown in Table 5. However, one-model-for-one-category does not offer a generalizable solution in a real-world problem setting, where multiple object categories and their parts need to be analysed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Number of Rendered Views. The effect of incrementing the number of views rendered from the input point cloud is shown in Table 6. With only 1 view in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.144, + 0.703, + 0.196 + ], + "angle": 0, + "content": "
MethodChair DispenserKeyboard EyeglassesesBucket
PartSLIP [31]85.373.753.688.336.5
PARIS3D86.575.688.292.283.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.267 + ], + "angle": 0, + "content": "Table 5: Results of PARIS3D on one model trained for each of 5 categories. In [31], 45 models were trained and each point cloud was tested by loading its corresponding model to perform evaluation. Repeating this exact setting by training a model on each category for PARIS3D, we easily gain \\(+1.2\\%\\), \\(+1.9\\%\\), \\(+34.6\\%\\), \\(+3.9\\%\\), and \\(+47.4\\%\\) improvements on the tested categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.784, + 0.381 + ], + "angle": 0, + "content": "formation provided, the segmentations are not meaningful. With 5 views, the segmentation quality improves, and improves further with an increase in the number of views. For each point cloud, we use 10 rendered views in our main experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.486 + ], + "angle": 0, + "content": "Single prompt vs Multiple prompts. In Table 7, we show ablation experiments regarding the number of training prompts. Single prompt refers to one concept-based reasoning training instruction that follows a template. In further experiments, we use the template to generate multiple (about 5-6) rephrased instructions for each part category using [36]. Training with these multiple training prompts per part category results in a better-performing model than with a single prompt." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.529, + 0.468, + 0.596 + ], + "angle": 0, + "content": "
No. of viewsBottleBucketCamera
11.11.11.2
523.740.112.1
1084.884.929.3
" + }, + { + "type": "table", + "bbox": [ + 0.476, + 0.53, + 0.753, + 0.582 + ], + "angle": 0, + "content": "
No. of promptsBottleKnifeClockChair
173.844.345.569.5
5-683.952.751.378.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.597, + 0.734, + 0.624 + ], + "angle": 0, + "content": "Table 6: Ablation experiments on the number of prompts provided in train-number of rendered views. ing data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.702, + 0.421, + 0.717 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Figure 5, we provide examples of PARIS3D output for 3D object part segmentation tasks. PARIS3D is capable of segmenting fine-grained details in categories like buttons of a remote or legs of an eyeglass from implicit queries that expect it to do self-reasoning to generate the right answer, and then segment the part. Moreover, it can perform tasks like object description and question-answering with regard to object parts and give detailed answers. This demonstrates its conversational, reasoning, and part segmentation capabilities." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.128 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.788, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.789, + 0.34 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results on real-world examples of point clouds. The input point cloud is shown on the left and the segmentation mask is in yellow on the right. Our model generalizes to point clouds with a significant domain gap from those it was trained on and displays impressive results with segmenting small object parts such as the furniture's knob in this example." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.36, + 0.603, + 0.377 + ], + "angle": 0, + "content": "4.5 Generalizability to Real-World Examples" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.789, + 0.549 + ], + "angle": 0, + "content": "Most existing 3D tasks and their corresponding models can usually only handle cases within the same distribution of the training sets without generalization, since they are sensitive to the format of the input and significant domain gap between synthetic experiments and real-world examples. Thus, it can be difficult to use them in use cases involving real point clouds from an everyday setting. To demonstrate the generalizability of PARIS3D to data derived from the real world, we perform our 3D segmentation on real point clouds shot using a smartphone's LiDAR sensor, as suggested by [31]. In Figure 6, we show qualitative examples of passing the fused point clouds through the PARIS3D architecture to obtain part segmentation labels as in the previous experiments without much drop in performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.36, + 0.586 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this work, we introduce a novel challenge within 3D segmentation, reasoning-based part segmentation. This task requires models to infer, reason, and explain based on implicit user instructions, making it considerably more complex than the regular 3D referring segmentation task. We introduce a dataset for this task, RPSeg3D, to enable effective evaluation. We believe this dataset will play a crucial role in fostering the growth of technologies in this area. Additionally, we outline a pipeline that integrates 3D segmentation capabilities into multimodal Large Language Models (LLMs), showcasing our model, PARIS3D, which exhibits competitive performance. It additionally demonstrates the ability to identify part concepts, reason about them, and complement them with world knowledge. However, we identify limitations: Although the model has improved generalizability, there is a possibility that the model may occasionally result in inaccurate predictions while being prompted with inappropriate, incorrect, or biased text inputs. Also, the model in its current form cannot perform instance segmentation. This is a direction for future research as we expand the dataset to accommodate such tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Aminabadi, R.Y., Rajbhandari, S., Zhang, M., Awan, A.A., Li, C., Li, D., Zheng, E., Rasley, J., Smith, S., Ruwase, O., He, Y.: Deepspeed inference: Enabling efficient inference of transformer models at unprecedented scale (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.278 + ], + "angle": 0, + "content": "3. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.32 + ], + "angle": 0, + "content": "4. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: L13da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.322, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3070-3079. IEEE Computer Society, Los Alamitos, CA, USA (jun 2019). https://doi.org/10.1109/CVPR.2019.00319, https://doi.ieeecomputersociety.org/10.1109/CVPR.2019.00319" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.432 + ], + "angle": 0, + "content": "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.474 + ], + "angle": 0, + "content": "7. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7010-7019 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.53 + ], + "angle": 0, + "content": "8. Gadelha, M., RoyChowdhury, A., Sharma, G., Kalogerakis, E., Cao, L., Learned-Miller, E., Wang, R., Maji, S.: Label-efficient learning on point clouds using approximate convex decompositions. In: European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.559 + ], + "angle": 0, + "content": "9. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.561, + 0.785, + 0.613 + ], + "angle": 0, + "content": "10. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9224-9232 (2018). https://doi.org/10.1109/CVPR.2018.00961" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.616, + 0.785, + 0.629 + ], + "angle": 0, + "content": "11. Graham, B., van der Maaten, L.: Submanifold sparse convolutional networks (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.631, + 0.785, + 0.671 + ], + "angle": 0, + "content": "12. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-lm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.673, + 0.785, + 0.727 + ], + "angle": 0, + "content": "13. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15372-15383 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.785, + 0.769 + ], + "angle": 0, + "content": "14. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "15. Hong, Y., Du, Y., Lin, C., Tenenbaum, J.B., Gan, C.: 3d concept grounding on neural fields. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.228 + ], + "angle": 0, + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.786, + 0.271 + ], + "angle": 0, + "content": "19. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2018). https://doi.org/10.1109/CVPR.2018.00278" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.352 + ], + "angle": 0, + "content": "20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 787-798. Association for Computational Linguistics, Doha, Qatar (Oct 2014). https://doi.org/10.3115/v1/D14-1086, https://aclanthology.org/D14-1086" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.786, + 0.393 + ], + "angle": 0, + "content": "21. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.786, + 0.448 + ], + "angle": 0, + "content": "22. Kundu, A., Yin, X., Fathi, A., Ross, D., Brewington, B., Funkhouser, T., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16. pp. 518-535. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.786, + 0.489 + ], + "angle": 0, + "content": "23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8500-8509 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.786, + 0.516 + ], + "angle": 0, + "content": "24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.786, + 0.543 + ], + "angle": 0, + "content": "25. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.786, + 0.569 + ], + "angle": 0, + "content": "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.786, + 0.596 + ], + "angle": 0, + "content": "27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.786, + 0.637 + ], + "angle": 0, + "content": "28. Lin, T.Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P., Ramanan, D., Zitnick, C.L., Dollar, P.: Microsoft coco: Common objects in context (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.689, + 0.65 + ], + "angle": 0, + "content": "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.786, + 0.732 + ], + "angle": 0, + "content": "30. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 53433-53456. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/a76b693f36916a5ed84d6e5b39a0dc03-Paper-Conference.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.786, + 0.787 + ], + "angle": 0, + "content": "31. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21736-21746 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.722, + 0.799 + ], + "angle": 0, + "content": "32. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "33. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=IDJx97BC38" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "34. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13589-13595 (2021). https://doi.org/10.1109/ICRA48506.2021.9561801" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "35. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.503, + 0.257 + ], + "angle": 0, + "content": "36. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.787, + 0.311 + ], + "angle": 0, + "content": "37. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.787, + 0.338 + ], + "angle": 0, + "content": "38. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.379 + ], + "angle": 0, + "content": "39. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.405 + ], + "angle": 0, + "content": "40. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.787, + 0.446 + ], + "angle": 0, + "content": "41. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.5 + ], + "angle": 0, + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.787, + 0.597 + ], + "angle": 0, + "content": "44. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.637 + ], + "angle": 0, + "content": "45. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.787, + 0.678 + ], + "angle": 0, + "content": "46. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. Proceedings of the IEEE International Conference on Computer Vision (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.679, + 0.787, + 0.704 + ], + "angle": 0, + "content": "47. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.745 + ], + "angle": 0, + "content": "48. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., Dai, J.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.787, + 0.786 + ], + "angle": 0, + "content": "49. Wang, Z., Huang, H., Zhao, Y., Zhang, Z., Zhao, Z.: Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.787, + 0.813 + ], + "angle": 0, + "content": "50. Xu, M., Ding, R., Zhao, H., Qi, X.: Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "51. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "A. Kareem et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "52. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.231 + ], + "angle": 0, + "content": "53. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Liu, Y., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "54. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "56. Ziyu, Z., Xiaojian, M., Yixin, C., Zhidong, D., Siyuan, H., Qing, L.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_origin.pdf b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0011f5eca99ffeb77644eab366fb9f1dd6208bc9 --- /dev/null +++ b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/ee219343-d433-449b-a58a-eb679a19a121_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdf72d2b0948e1a2c8c1628dba93c4836e2ec2e817c26d2cc6967c897808bb8c +size 10557391 diff --git a/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/full.md b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..069175928908435c73d815a1c520f48ac5d09af4 --- /dev/null +++ b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/full.md @@ -0,0 +1,301 @@ +# PARIS3D: Reasoning-based 3D Part Segmentation Using Large Multimodal Model + +Amrin Kareem1, Jean Lahoud1, and Hisham Cholakkal1 + +Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE + +Abstract. Recent advancements in 3D perception systems have significantly improved their ability to perform visual recognition tasks such as segmentation. However, these systems still heavily rely on explicit human instruction to identify target objects or categories, lacking the capability to actively reason and comprehend implicit user intentions. We introduce a novel segmentation task known as reasoning part segmentation for 3D objects, aiming to output a segmentation mask based on complex and implicit textual queries about specific parts of a 3D object. To facilitate evaluation and benchmarking, we present a large 3D dataset comprising over 60k instructions paired with corresponding ground-truth part segmentation annotations specifically curated for reasoning-based 3D part segmentation. We propose a model that is capable of segmenting parts of 3D objects based on implicit textual queries and generating natural language explanations corresponding to 3D object segmentation requests. Experiments show that our method achieves competitive performance to models that use explicit queries, with the additional abilities to identify part concepts, reason about them, and complement them with world knowledge. Our source code, dataset, and trained models are accessible at this url. + +Keywords: 3D $\cdot$ Vision-Language Models $\cdot$ Reasoning + +# 1 Introduction + +The rapid advancements in 3D data capture technologies, including LIDARs and RGB-D cameras, have led to a growing demand for automated analysis of 3D point clouds. 3D semantic segmentation, the process of automatically assigning predefined semantic labels to each point in a cloud, is crucial for enabling complex tasks such as scene understanding. Similarly, 3D part segmentation involves further segmenting object instances into their components, such as identifying the handle of a pot or the lid of a bottle. These tasks find applications in various fields including autonomous vehicles, mobile robotics, industrial automation, augmented reality, and medical imagery analysis. While recent advancements in pre-trained 3D representations and the introduction of various 3D datasets have significantly improved 3D perception, the capacity for nuanced reasoning in 3D contexts remains limited. This limitation is primarily due to the lack of comprehensive datasets for reasoning and describing 3D scenes and objects. + +![](images/4711b544402ea5357b6492736fd8756fbdd1ba22bb757f8c91df9071d0161ac2.jpg) +Fig. 1: Capabilities of PARIS3D. Parts of 3D objects are segmented based on reasoning, shape, location, material, colour, and concept instructions. Additionally, for the segmentations, PARIS3D can explain why it chose that region, or describe 3D objects with respect to their parts. The original point clouds are on the left. The segmented parts are shown to the right, highlighted in golden colour. + +In contrast, 2D images accompanied by textual descriptions have contributed to significant progress in large-scale image language models. Recent methods have emerged, performing segmentation based on 2D images projected from 3D scenes. However, they lack the ability to reason about object concepts and their 3D properties, functions, or parts in a conversational manner. + +Here, we define a model as having reasoning capabilities when (i) it can understand implicit instructions, such as referencing properties of an object or its parts without explicit articulation, and (ii) explain or justify its responses, whether generated text or predicted segmentation masks. Such reasoning ability is one of the fundamental cognitive skills possessed by humans, essential for daily activities ranging from locating items to manipulating tools. Previous 3D vision methods are capable of exploring the semantics of a given environment based on instructions containing pre-defined category names. These explicit categories are limiting and do not address cases where instructions can be implicit and may contain complex queries. For example, an automated system might need to find the part of the kettle that can be opened to pour water, without knowing the explicit category name "lid". To this end, we introduce the 3D reasoning part segmentation task, which aims at segmenting object parts based on general queries, including color, 3D information, and implicit functionalities. + +In this work, we introduce a novel task termed reasoning part segmentation in 3D. This task involves generating a part segmentation mask for a 3D object based on implicit textual queries requiring complex reasoning. These queries go beyond simple references and encompass intricate descriptions, demanding sophisticated reasoning or worldly knowledge. For example, knowing where to hold a kettle or + +which part of a bottle to open in a fine-grained manner. To address this challenging task, we propose PARIS3D (ReasonIng-based 3D PArt Segmentation), a multi-modal Large Language Model (LLM) capable of reasoning on user input text, predicting 3D part segmentation masks, and providing explanations for the model's response. As depicted in Figure 1, PARIS3D adeptly navigates diverse scenarios, encompassing complex 3D reasoning, material, color, shape, and location-based knowledge, providing explanatory responses and detailed descriptions. Additionally, to validate the effectiveness of PARIS3D and support future research, we establish an evaluation benchmark and a dataset named RPSeg3D for reasoning 3D part segmentation. Our RPSeg3D dataset comprises 2624 3D objects and over 60k instructions, providing persuasive evaluation metrics for the task. + +In summary, our key contributions are the following: + +- We introduce the reasoning part segmentation task for 3D objects, emphasizing the necessity of reasoning capabilities which is crucial for the development of intelligent perception systems. +- We provide a comprehensive dataset named RPSeg3D for reasoning-based 3D part segmentation, comprising over 2624 3D objects and 60k instructions, serving as a useful resource for future research. +- We present PARIS3D, a novel approach for 3D part segmentation, and further improve its capabilities by fine-tuning on our RPSeg3D dataset. + +# 2 Related Work + +3D Semantic Segmentation. The challenge of understanding and reasoning within 3D environments has been an ongoing research focus. The goal of 3D semantic segmentation is to acquire semantic predictions for each point in a cloud. Notable contributions include point-based approaches [19, 38], methods which incorporate intricately crafted point convolution methods [46, 50], voxel-based strategies [5, 10], including those that employ 3D sparse convolutions [11] for generating point-wise segmentation outcomes, as well as transformer-based techniques [23]. Multi-view semantic segmentation methods such as DeepViewAgg [41], Diffuser [22, 34], 3D-CG [15], 3D-CLR [16] in 3D vision concentrate on improving representation learning by generating 2D renderings from 3D under multiple view points. These works have shown the effectiveness of multi-view representations in enhancing the performance and robustness of various 3D tasks. Nonetheless, these methods rely on a predefined set of semantic labels, whereas we focus in our proposed method on responding to complex reasoning-based queries and explaining them. + +Large Multimodal Models. Extensive research on Large Language Models (LLMs) has demonstrated reasoning capabilities, prompting an exploration into extending these skills into the visual domain through Large Multimodal Models. LMMs are a highly adaptable and versatile means to perform tasks requiring language and vision capabilities. Prominent models such as BLIP-2 [26], LLaVA [29] and MiniGPT-4 [55] generally employ a dual-phase training process, aligning + +
MethodInputTask
Point CloudReasoning QueryConversationSegmentationExplanation
SQA3D [33], 3D-VisTA [56]✓✓XXX
ViewRefer [13], Point-Bind [12]✓XXXX
3D-OVS [30]XXX✓X
OpenMask3D [45]✓XX✓X
PLA [6], OpenScene [37]✓XX✓X
Chat-3D [49]✓✓✓X✓
M3DBench [27]✓X✓XX
LLM-Grounder [52]✓X✓X✓
3D-LLM [17]✓✓✓X✓
LL3DA [4], PointLLM [51]✓X✓X✓
PARIS3D✓✓✓✓✓
+ +Table 1: Comparison of recent 3D segmentation models and Large Multi-modal Models (LMMs) emphasizing their capabilities for 3D reasoning and conversations. Reasoning query means the model is asked to self-reason a task and either output text or perform an action. Segmentation highlights models that can respond with 3D segmentation masks, and Conversation represents models that can provide a conversation-style answer to the user. Among these, our proposed PARIS3D stands out with comprehensive 3D understanding and reasoning, segmentation in response to natural language queries, and conversational capabilities. + +visual representations with the linguistic embeddings of LLMs through extensive image-text and video-text datasets [2,3,28,35,42-44]. Recent efforts have focused on the convergence of multi-modal LLMs with vision tasks, where VisionLLM [48] provides a versatile interaction interface for a spectrum of vision-centric tasks through instruction tuning. Yet, it does not fully leverage the complex reasoning potential of LLMs. GPT4RoI [53] integrates spatial boxes as inputs and trains models on region-text pairs, showcasing a novel approach. Our method aims to benefit from these advances in the LMM space by merging the vision-language abilities of LMMs and the reasoning of LLMs in a novel 3D perception task. + +Language Instructed 3D Tasks. The integration of point clouds with natural language processing has widespread implications, drawing considerable interest in the realms of 3D scene understanding. This fast-growing field promises enhancements in human-robot interaction, metaverse, robotics, and embodied intelligence. Central to the dialogue systems designed for 3D environments are two critical capabilities: perception within three-dimensional spaces and reasoning. Recently, there has been a rise in the number of tasks uniting 3D scenes and language, such as 3D captioning, 3D question answering, 3D situated question answering, embodied Q and A, planning, navigation, 3D assisted multi-turn dialogue, 3D object detection, and scene description. We divide the 3D perception task models into 3 categories (see Table 1. Separated by dotted lines.). The first encompasses models that perform tasks like 3D captioning, situated question answering, and visual grounding [33, 56], visual grounding [14]. These models are capable of providing a word or phrase as its text output. The second category has 3D semantic segmentation models that output 3D segmentation masks. 3D + +OVS [30], Openmask3D [45], OpenScene [37], PLA [7], perform open-vocabulary semantic segmentation for 3D scenes. These methods, on the other hand, cannot provide a conversational output to a user query or explain the reasoning for their tasks. The third comprises models that employ an LLM and perform visual perception tasks such as captioning, scene understanding, and visual grounding, providing conversational outputs [4, 12, 17, 27, 49, 51, 52]. However, they do not perform fine-grained semantic segmentation or reasoning-based 3D vision tasks. As such, we identify a gap in performing 3D segmentation in response to complex natural language prompts instead of a single phrase or word. Distinct from the existing works in this domain, our research is committed to: (1) streamlining the integration of 3D segmentation capabilities into multi-modal LLMs, and (2) enhancing current perception systems with 3D reasoning-based segmentation and explanation abilities, thereby pushing the boundaries of what is achievable in the intersection of language and visual understanding. + +# 3 Reasoning 3D Part Segmentation + +Semantic segmentation involves assigning a semantic label to each geometric primitive, such as points [38], voxels [9], or superpoints [25]. In part segmentation, object instances are decomposed into their components. Given a coloured point cloud P, the goal of a 3D segmentation model is to predict its label for each point. However, in our reasoning segmentation task, we go further to output a 3D segmentation mask M, given an input point cloud P and an implicit query text instruction $x_{txt}$ . The task shares a similar formulation with the referring segmentation task [20], with an additional challenge for the model to reason about the fine-grained parts in response to implicit queries and output the corresponding segmentation mask. The complexity of the query text in reasoning part segmentation is a key differentiator. Instead of providing the names of the parts, the query text may include more intricate expressions that involve an understanding of structures, geometries, and semantics of 3D objects. By introducing this task, we aim to bridge the gap between user intent and system response, enabling more intuitive and dynamic interactions in 3D object perception. + +# 3.1 Our RPSeg3D Dataset + +Considering the unavailability of established datasets and evaluation benchmarks in the literature, we introduce a dataset, named RPSeg3D, specifically designed for the reasoning 3D part segmentation task. Our dataset comprises 2624 3D objects and over 60k instructions. We use 718 objects and their corresponding instructions as the train set, and the remaining 1906 objects along with their instructions are used for testing. For reliable and fair assessment, we have aligned the 3D objects with those from PartNet-Ensemble [31], annotating them with implicit text instructions and using ground truth labels to generate high-quality target masks. To generate the text instructions corresponding to each 3D object part, we prepare a set of templates, for example, "Which part of this object + +![](images/13fa119c38da0917bdff7b22fe309d1bce23dd4c0ba4652f6cad6095142ce70a.jpg) +Normal Query: "In this toaster, which part can be slid up and down for temperature adjustment?" + +3D Query: "What is the gray rectangular part in the middle of the Toaster that you can slide up and down to set the level of toasting?" + +![](images/298e393bbfcde62fc3d36675d0fd6eee35c1d11a8bd896a660ca9b185e28e873.jpg) +Normal Query: "Which part of this kettle does the water come out of?" + +3D Query: "How long is the long cylindrical pipe-like brown part of this kettle that liquid comes out of compared to its whole length?" + +![](images/e61b493c240b4d2cfb1ddabceb47b73c26a89894af611cde6ec0b09b311d9697.jpg) +Normal Query: "Which part of this keyboard used to type information?" + +3D Query: "Can you identify the white rectangular part of the keyboard that makes up about a tenth of its height and is located in the middle?" + +![](images/2ffd37fc7c354aa46fbb3abd3024d54e3e4d3dbf5424fad98dc470b8d98310a2.jpg) +Normal Query: "Which part of this Bucket provides a grip for carrying it?" + +3D Query: "What is the shape of the gray part of this Bucket located in front of and on the side of it? What is it used for?" + +![](images/987d35ac10ca66fd06153ec7e92d191ab2c68f5db44bcd77ca3addc8cada906f.jpg) + +![](images/e366f35541b633c3a95203b013b090da930a607a47ae18e81dcad8410b0c0c35.jpg) +Normal Query: "Which part of this storage furniture is drawn out and used for storage?" + +3D Query: "In this StorageFurniture, which is the rectangular part used for storage that constitutes about a quarter of its total height?" + +![](images/97ba835333ffc25140ea1d4685080dbcb2d8830664ed04d870e4de2affd2b593.jpg) +Normal Query: "Which part of this door surrounds the entrance?" + +3D Query: "What is the shape of the brown part of this Door located on top of and on the sides of it? What is it used for?" + +![](images/f927279a59d0691d64d076c84329ef5ec182c7f6bf77657851fa278df3d78856.jpg) +Fig. 2: Examples of the annotated object-instruction pairs for training with two types of queries. On the left is one view of the rendered image from the original point cloud. On the right is the corresponding ground truth segmentation mask, shown in green. +Fig. 3: Preparing the instructions of RPSeg3D. Simple templates are provided to GPT-3.5, which populates them with part information-related segmentation instructions. In parallel, colour, shape, location, and dimension-related data is extracted from 3D point clouds. Enriching the instructions with this information and manually checking them for inaccuracies, we obtain the RPSeg3D dataset for part segmentation. + +?" We leverage GPT-3.5 [36] for building instructions using these templates by supplying information about the part and rephrasing. We also extract 3D information from the point cloud, thus designing the instructions to cover relations, dimensions (length, height), comparisons, colour, texture features, object concepts, and functions. This was further verified manually to avoid inaccurate prompts at test time. The steps are illustrated in Figure 3. To cover 3D object segmentation tasks effectively, our text instructions consist of two types: 1) normal queries; 2) 3D queries; as illustrated in Figure 2. The dataset is partitioned into train and test splits, containing 718 and 1906 3D objects and over 16k and 47k instructions, respectively. As the primary purpose of the dataset is evaluation, the testing set includes a larger number of instructions. + +# 3.2 Reasoning 3D Part Segmentation Architecture + +Our method takes as input a dense and coloured 3D point cloud of an object. One of the common methods of 3D analysis is predominantly using point clouds + +to represent 3D data. However, this contrasts with human spatial reasoning processes. Humans typically engage with their surroundings through active exploration, synthesizing perspectives from multiple vantages to form an integrated 3D understanding, rather than processing a 3D environment at one go. Our approach advocates for 3D reasoning derived from multi-view imagery. This approach also benefits from the large-scale 2D pretraining available in vision-language models, similar to previous methods that have taken advantage of pre-trained vision-language models for 3D vision tasks. Thus, we render multiple images $x_{img}$ from K predefined camera poses by rasterization. The camera poses cover all parts of the object since they are uniformly distributed around the input point cloud. Given a complex text instruction $x_{txt}$ along with the images, we feed them into a multimodal LLM, denoted by $F$ , and the visual backbone, $F_{enc}$ . $F$ outputs a text response $\hat{y}_{txt}$ . In parallel, $F_{enc}$ extracts the visual embeddings, f, from each of the input images $x_{img}$ . The formulation is as follows: + +$$ +\hat {y} _ {t x t} = F \left(x _ {i m g}, x _ {t x t}\right), f = F _ {e n c} \left(x _ {i m g}\right) \tag {1} +$$ + +Whenever the LLM is expected to yield a binary segmentation mask, the resultant text output, denoted as $\hat{y}_{txt}$ is required to an extra token, which sends a request for the segmentation output. Following this, the embedding at the last layer, $\hat{h}_{seg}$ , corresponding to the additional segmentation token is extracted. This embedding is subsequently processed through an MLP (Multilayer Perceptron) projection layer, represented as $\gamma$ , to derive $h_{seg}$ and then input to the mask decoder $F_{dec}$ as its prompt embedding, thereby tightly coupling the text output from $F$ and the mask predicted by the decoder. Hence, the mask decoder takes the visual features $f$ along with $\langle \mathrm{SEG} \rangle$ based prompt embeddings to perform segmentation, generating the final segmentation mask, $\hat{M}$ similar to [24]. The process is as follows: + +$$ +h _ {s e g} = \gamma (\hat {h} _ {s e g}) \tag {2} +$$ + +The architectural specifics of the decoder, $F_{dec}$ , are in accordance with [21]. + +$$ +\hat {M} = F _ {\text {d e c}} \left(h _ {\text {s e g}}, f\right) \tag {3} +$$ + +Once the K segmentations are obtained from each camera angle, a 3D semantic voting module computes scores and assigns the semantic labels for each part. We have $K$ semantic segmentation masks $M_{k}$ , where $k$ is the view from which the image was rendered from the point cloud. We aggregate the masks from multiple views and lift them to obtain a three-dimensional semantic segmentation of the original point cloud. To achieve this, we segment the input point cloud $\mathbf{P}$ into a set of superpoints $SP_{i}$ , similar to [25]. Superpoints refer to a way of representing large 3D point clouds as a collection of interconnected geometrically simple shapes. This representation is advantageous because (i) it considers complete object parts as a whole, making them easier to identify, (ii) this method can provide a detailed description of the relationship between adjacent objects, and (iii) the size of superpoints is determined by the number of simple structures in a + +![](images/cfcc961beadcc49a08be7acda431c29681d340f37872c49c648b7baa2c3a8e99.jpg) +Fig. 4: Overview of the proposed reasoning-based 3D part segmentation approach named PARIS3D. It comprises four subsequent steps: (i) The 3D point cloud is rendered into K multi-view images $x_{img}$ using a renderer. (ii) These images are passed through a frozen vision backbone ( $F_{enc}$ ) and multi-modal large language model ( $F$ ) of the reasoning module. $F$ also accepts the text query $x_{txt}$ , and produces text outputs corresponding to each view. (iii) The decoder decodes the final layer embedding which contains the extra token, thus producing K segmentation masks. (iv) Finally, a mask to 3D segmentation algorithm lifts the projections back into 3D and a view-guided scoring module is used to obtain the final text response. + +scene, rather than the total number of points, making the representation several orders of magnitude smaller, thus more efficient to work with. Each superpoint contains points with similar normals and colours, suggesting they belong to the same instance. This superpoint-based part labelling not only conserves computational resources but also potentially improves performance by leveraging 3D priors. In the 3D semantic voting step, we assign semantic labels for each superpoint through a voting mechanism similar to [31]. This technique leverages information from multiple views and the superpoints, so that even if one mask covers irrelevant points, the aggregation of masks from all views counters the effect of such errors. For a superpoint $SP_{i}$ and a part category $j$ , a score $s_{i,j}$ is computed reflecting the proportion of visible points present in the predicted segmentation masks of category $j$ in each view: + +$$ +s _ {i, j} = \frac {\sum_ {k} \sum_ {p \in S P _ {i}} \left[ V _ {k} (p) \right] \left[ \exists b \in M _ {k} ^ {j} : I _ {m} (p) \right]}{\sum_ {k} \sum_ {p \in S P _ {i}} \left[ V _ {k} (p) \right]} \tag {4} +$$ + +Here, $[\cdot ]$ is the Iverson bracket, which turns true predicates into 1 and false into 0. $M_k^j$ , represents the predicted mask of category j in view k. $V_{k}(p)$ indicates the visibility of 3D point p in the view k. $I_{m}(p)$ checks if the projection of point p in view k falls within the mask m. The superpoint is then assigned the semantic label of the part category with the highest score. + +To provide the most relevant text explanation corresponding to the 3D masks, we propose a view-guided scoring module. It assigns a score to each view based on its correspondence with the final semantic labels. The best text explanation is chosen based on the highest score. Specifically, for each view, we compute the intersection over union (IoU) of the mask with the projections of the final 3D semantic labels output by the model, and the corresponding text explanation is + +saved. This is called the view-guided score of the explanation. At the end of the score computation for all the masks, we choose the text explanation with the highest score, and in turn, the most correspondence with the output labels. + +# 3.3 Training the PARIS3D Architecture + +Training Data Formulation The training data consists of 718 3D objects rendered into multiview images, resulting in over 16k image-instruction pairs. Out of the 718, 360 objects with their instructions are used for training the model and 358 objects with their instructions are used for validation. Each image is provided with an annotation file, which has instructions corresponding to the 3D point cloud that the image came from, the name of the image, and its ground truth mask. Generation of instructions follows the same steps as Section 3.1 illustrated in Figure 3. Images in the training set may have more than one instruction in its "instruction" field. This is helpful to introduce diversity as users may randomly select one as the reasoning query during training, thus obtaining a better model. + +Distillation-Based Explanation Refining Training the multiview model requires explanatory data as well as training image-instruction pairs. To build the explanatory data, we built an annotation pipeline using a distillation approach. Using the multi-view images as input to the teacher model [24], we generate explanations for each part of the object. These explanations serve as our pseudolabels which we use as ground truth explanations of the student model. We augment these annotations with 3D features extracted from the 3D point clouds from which the images were rendered. The 3D feature-augmented explanations contain appropriate responses and critical elements differentiating object parts from each other such as location, size, shape, material, and colour. + +Objective We leverage LoRA [18] to perform efficient fine-tuning of the pretrained multi-modal LLM $F$ [29] to retain its generalization capability. We completely freeze the vision backbone $F_{enc}$ . The decoder $F_{dec}$ is fully fine-tuned. The word embeddings of the LLM and the projection layer of $\gamma$ are also trainable, allowing the model to learn the specific meanings and semantic concepts of parts. The model is trained end-to-end using the text generation loss $L_{txt}$ and the segmentation mask loss $L_{mask}$ . The overall objective is given as: + +$$ +L = \lambda_ {t x t} L _ {t x t} + \lambda_ {m a s k} L _ {m a s k} \tag {5} +$$ + +Here, $L_{mask}$ encourages the model to produce high-quality segmentation results. It is a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights $\lambda_{bce}$ and $\lambda_{dice}$ , given by: + +$$ +L _ {m a s k} = \lambda_ {b c e} B C E (\hat {M}, M) + \lambda_ {d i c e} D I C E (\hat {M}, M) \tag {6} +$$ + +where $y_{txt}$ and $M$ are the ground-truth targets. $L_{txt}$ is the auto-regressive cross-entropy loss for text generation. It is computed as: + +$$ +L _ {t x t} = C E \left(\hat {y} _ {t x t}, y _ {t x t}\right) \tag {7} +$$ + +# 4 Experiments + +We perform quantitative and qualitative evaluation of PARIS3D on our dataset for reasoning-based semantic segmentation. + +Implementation Details and Metrics For the experiments, we follow [24] where the multi-modal LLM F is LLaVA-13B-v1-1 [29] and the vision backbone $F_{enc}$ is the ViT-H SAM. The projection layer of $\gamma$ is an MLP with channels of [256, 4096, 4096]. We use our dataset RPSeg3D which contains coloured point clouds and rendered 2D images of them. Using Pytorch3D [40], each input point cloud is rendered into $\mathrm{K} = 10$ colour images. The fine-tuning scripts for the LLaVA and SAM architecture are based on DeepSpeed [1] engine. We adopt the settings of [24] for the optimizer (AdamW [32]) and its learning rate (0.0003). Similar steps are followed for the learning rate scheduler (WarmupDecayLR), text generation loss $\lambda_{txtgen}$ weight (1.0) and the mask loss $\lambda_{mask}$ weight (1.0), the BCE loss $\lambda_{bce}$ (2.0), dice loss $\lambda_{dice}$ (0.5), batch size per device (2), and the gradient accumulation step (10). The semantic segmentation metric used is category mIoU, following [31]. It is calculated as follows: first, mIoU for each part category is calculated for all test objects. Then, part mIoUs that belong to each object category are averaged to compute the object category mIoU. + +# 4.1 Reasoning Part Segmentation Results + +We establish the reasoning part segmentation task on our dataset RPSeg3D. Table 2 shows the results of reasoning part segmentation. We observe that without any fine-tuning, the model's performance is low compared to its fine-tuned counterparts. The general observation is that 3D queries help the model to output better masks. We compare our method to two baselines. The first baseline is LISA [24] applied to the multiple views without any finetuning. The second baseline consists of LISA finetuned on few-shot part segmentation data. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts. + +# 4.2 3D Semantic Segmentation Comparison with Existing Models + +Table 3 shows the results of semantic segmentation compared to existing methods. Our method has better performance than all the fully supervised baselines and achieves competitive results with [31]. For this baseline, we observe that the model with prompt tuning done separately on each category achieves impressive few-shot performance but when unified into a single model for all categories, the performance is only slightly better than its zero-shot performance $(27.2\%)$ . This is attributed to the redundant part names across object categories, hindering the model's learning of the semantic meanings of part names. This results in performance to drop significantly from when only one category is learnt per model. To prompt for segmentation, the baseline models [38,39,54] are provided with the + +
MethodValTest
Normal Query3D QueryOverallNormal Query3D QueryOverall
LISA-MV16.6020.1618.3817.6020.5719.08
LISA-MV (ft)50.4350.2850.3550.7550.8150.78
PARIS3D55.3355.5055.4255.9457.6056.77
+ +Table 2: Results of reasoning part segmentation. LISA-MV [24] is LISA in multi-view setting without fine-tuning. LISA-MV (ft) is the experiment in which it has been fine-tuned on few-shot part segmentation data. Our proposed PARIS3D method has been fine-tuned with 3D queries and explanations. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts. Here Test is the test set of 1906 shapes and Val is the validation set of 358 3D shapes with their instructions. + +
3D DataMethodOverlapping CategoriesNon-overlapping CategoriesOverall(45)
Bottle ChairDisplay DoorKnife LampStorage-FurnitureTableOverall(17)CameraCartDis-PenserKettleKitchenPotOvenSuit-caseToasterOverall(28)
Extra data(45x8+28k)PointNet++ [38]48.884.778.445.735.468.046.963.755.66.56.412.120.915.834.340.614.725.436.8
PointNeXt [39]68.491.889.443.858.764.968.552.158.533.236.326.045.157.037.813.58.345.150.2
SoftGroup [47]41.488.362.153.131.382.260.254.850.223.623.918.957.445.513.618.326.430.738.1
Few-shot(45x8)PartSLIP* [31]83.485.384.840.865.266.053.642.456.358.388.173.777.069.673.570.460.061.359.4
PointNet++ [38]27.042.230.220.522.210.58.47.318.19.711.67.028.631.719.43.30.021.820.4
PointNeXt [39]67.665.153.746.359.755.420.622.139.226.047.722.660.566.036.814.50.041.540.6
SoftGroup [47]20.880.539.716.338.338.318.924.932.828.640.842.960.754.835.629.814.841.138.0
ACD [8]22.439.029.218.939.613.77.613.519.210.131.519.440.251.88.913.20.025.623.2
Prototype [54]60.170.867.333.450.438.230.225.741.132.036.853.462.763.336.535.510.146.344.3
PartSLIP+64.869.559.524.534.537.132.040.135.325.575.715.630.558.431.149.46.626.729.9
Ours84.081.070.168.447.261.239.445.155.129.371.740.159.378.859.161.624.959.157.6
+ +Table 3: Comparison to previous 3D part segmentation methods. Object category mIoU(%) is shown. In the $45 \times 8 + 28k$ setting, baseline models use an additional $28k$ training shapes for 17 overlapping object categories. These are categories present in common with PartNet dataset. For the remaining 28 non-overlapping object categories, there are only 8 shapes per object category during training. PartSLIP* indicates that one model has been trained for each category. + shows our implementation of PartSLIP where one model is trained for all the categories together. + +class IDs of the parts to be segmented (e.g: 0 to 103 including the background). In [31], the model is provided with the names of the parts to be segmented (e.g. "seat", "arm", "back" of a chair). Our method, PARIS3D, uses language instructions to prompt segmentation. For a fair comparison, we use hand-crafted prompts with short instructions containing generic concept and location-based clues about the part to be segmented. + +# 4.3 Ablation Study + +Instruction Rephrasing. We perform additional experiments with varying instruction types to study the merits of using rephrased prompts in training the model (Table 4 here) compared to using only category names and manually generated prompts for each category. These experimental results on the test set show that GPT-generated prompts provide superior performance, both without and with color and 3D information. In both cases, performance improves from using category names to manual prompts to GPT prompts, indicating the strong + +![](images/ac5ab5acba333b66e1e8a5cb61a47bda707edcfb258d110a6284e5907a7a48d0.jpg) + +![](images/0f56d92919fbc3932df4a802f31d19c15fd6833471ea0cf2d590b80d0507578f.jpg) + +![](images/1a5802da97d5113cb9abba88298ae92f1d2536859b17bd1b759cca8ab4591baa.jpg) + +![](images/1ca4b3e964b5a6dc61f92d73bd65c63cf5a35e3087d8bb638f6dca8bffc4ff37.jpg) + +![](images/55bcaa69480aaf021c4dc74292ae372faa0329104848855a490f3e59fdea417e.jpg) + +![](images/543cbfc730632ebead5d5c98ce5fa329cc302306772b6087ad824e2e6ae43be2.jpg) +Fig. 5: Qualitative results of PARIS3D's performance. We showcase examples from three tasks: reasoning 3D object part segmentation, object description, and reasoning question-answering, demonstrating its capabilities in offering in-depth reasoning, 3D understanding, part segmentation, and conversational abilities. + +![](images/0be43e828d54c1968801690d851e0786415948b72579b18cad191020617ebdac.jpg) + +![](images/e9a017b4ea1cd2dae4e60beb7941e7bd55908b4f20a915e9dffde443b08ce804.jpg) + +
Instruction TypeCategory NameManually GeneratedGPT-generatedCategory NameManually generatedGPT-generated
Colour + 3DXXX✓✓✓
Overall mIoU38.041.546.743.052.457.6
+ +Table 4: A study on the effectiveness of rephrased prompts on the overall model performance. Here, the ability of the model to accurately segment an object part based on an implicit (reasoning-based) description is evaluated for different types of training prompts. The results show the superior performance of GPT-generated prompts over manual prompts, both with and without colour and 3D information. + +generalization capabilities introduced by GPT-based prompts. For instance, using color and 3D information prompts along with 'manual prompts' resulted in an overall mIoU of $52.4\%$ . However, using GPT to expand simple templates into varied prompts for describing object part categories resulted in about a $5.2\%$ improvement over the manually curated training templates (last vs second-last column). + +One Model vs Multiple Models. One of the baseline models [31] trained a model for each of the 45 categories, loading one at a time to evaluate each category. When we replicate this setting for PARIS3D, there are significant jumps in performance as shown in Table 5. However, one-model-for-one-category does not offer a generalizable solution in a real-world problem setting, where multiple object categories and their parts need to be analysed. + +Number of Rendered Views. The effect of incrementing the number of views rendered from the input point cloud is shown in Table 6. With only 1 view in + +
MethodChair DispenserKeyboard EyeglassesesBucket
PartSLIP [31]85.373.753.688.336.5
PARIS3D86.575.688.292.283.9
+ +formation provided, the segmentations are not meaningful. With 5 views, the segmentation quality improves, and improves further with an increase in the number of views. For each point cloud, we use 10 rendered views in our main experiments. + +Single prompt vs Multiple prompts. In Table 7, we show ablation experiments regarding the number of training prompts. Single prompt refers to one concept-based reasoning training instruction that follows a template. In further experiments, we use the template to generate multiple (about 5-6) rephrased instructions for each part category using [36]. Training with these multiple training prompts per part category results in a better-performing model than with a single prompt. + +Table 5: Results of PARIS3D on one model trained for each of 5 categories. In [31], 45 models were trained and each point cloud was tested by loading its corresponding model to perform evaluation. Repeating this exact setting by training a model on each category for PARIS3D, we easily gain $+1.2\%$ , $+1.9\%$ , $+34.6\%$ , $+3.9\%$ , and $+47.4\%$ improvements on the tested categories. + +
No. of viewsBottleBucketCamera
11.11.11.2
523.740.112.1
1084.884.929.3
+ +
No. of promptsBottleKnifeClockChair
173.844.345.569.5
5-683.952.751.378.7
+ +Table 6: Ablation experiments on the number of prompts provided in train-number of rendered views. ing data. + +# 4.4 Qualitative Results + +In Figure 5, we provide examples of PARIS3D output for 3D object part segmentation tasks. PARIS3D is capable of segmenting fine-grained details in categories like buttons of a remote or legs of an eyeglass from implicit queries that expect it to do self-reasoning to generate the right answer, and then segment the part. Moreover, it can perform tasks like object description and question-answering with regard to object parts and give detailed answers. This demonstrates its conversational, reasoning, and part segmentation capabilities. + +![](images/8540031978837ec9a5e228819188bef2cdbdecf1064807b26ae5dcb6fc535c09.jpg) +Fig. 6: Qualitative results on real-world examples of point clouds. The input point cloud is shown on the left and the segmentation mask is in yellow on the right. Our model generalizes to point clouds with a significant domain gap from those it was trained on and displays impressive results with segmenting small object parts such as the furniture's knob in this example. + +# 4.5 Generalizability to Real-World Examples + +Most existing 3D tasks and their corresponding models can usually only handle cases within the same distribution of the training sets without generalization, since they are sensitive to the format of the input and significant domain gap between synthetic experiments and real-world examples. Thus, it can be difficult to use them in use cases involving real point clouds from an everyday setting. To demonstrate the generalizability of PARIS3D to data derived from the real world, we perform our 3D segmentation on real point clouds shot using a smartphone's LiDAR sensor, as suggested by [31]. In Figure 6, we show qualitative examples of passing the fused point clouds through the PARIS3D architecture to obtain part segmentation labels as in the previous experiments without much drop in performance. + +# 5 Conclusion + +In this work, we introduce a novel challenge within 3D segmentation, reasoning-based part segmentation. This task requires models to infer, reason, and explain based on implicit user instructions, making it considerably more complex than the regular 3D referring segmentation task. We introduce a dataset for this task, RPSeg3D, to enable effective evaluation. We believe this dataset will play a crucial role in fostering the growth of technologies in this area. Additionally, we outline a pipeline that integrates 3D segmentation capabilities into multimodal Large Language Models (LLMs), showcasing our model, PARIS3D, which exhibits competitive performance. It additionally demonstrates the ability to identify part concepts, reason about them, and complement them with world knowledge. However, we identify limitations: Although the model has improved generalizability, there is a possibility that the model may occasionally result in inaccurate predictions while being prompted with inappropriate, incorrect, or biased text inputs. Also, the model in its current form cannot perform instance segmentation. This is a direction for future research as we expand the dataset to accommodate such tasks. + +# References + +1. Aminabadi, R.Y., Rajbhandari, S., Zhang, M., Awan, A.A., Li, C., Li, D., Zheng, E., Rasley, J., Smith, S., Ruwase, O., He, Y.: Deepspeed inference: Enabling efficient inference of transformer models at unprecedented scale (2022) +2. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval (2022) +3. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts (2021) +4. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: L13da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning (2023) +5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3070-3079. IEEE Computer Society, Los Alamitos, CA, USA (jun 2019). https://doi.org/10.1109/CVPR.2019.00319, https://doi.ieeecomputersociety.org/10.1109/CVPR.2019.00319 +6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +7. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7010-7019 (June 2023) +8. Gadelha, M., RoyChowdhury, A., Sharma, G., Kalogerakis, E., Cao, L., Learned-Miller, E., Wang, R., Maji, S.: Label-efficient learning on point clouds using approximate convex decompositions. In: European Conference on Computer Vision (ECCV) (2020) +9. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks (2017) +10. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9224-9232 (2018). https://doi.org/10.1109/CVPR.2018.00961 +11. Graham, B., van der Maaten, L.: Submanifold sparse convolutional networks (2017) +12. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-lm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following (2023) +13. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15372-15383 (October 2023) +14. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance (2023) +15. Hong, Y., Du, Y., Lin, C., Tenenbaum, J.B., Gan, C.: 3d concept grounding on neural fields. NeurIPS (2022) +16. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) + +17. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. NeurIPS (2023) +18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9 +19. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2018). https://doi.org/10.1109/CVPR.2018.00278 +20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 787-798. Association for Computational Linguistics, Doha, Qatar (Oct 2014). https://doi.org/10.3115/v1/D14-1086, https://aclanthology.org/D14-1086 +21. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything (2023) +22. Kundu, A., Yin, X., Fathi, A., Ross, D., Brewington, B., Funkhouser, T., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16. pp. 518-535. Springer (2020) +23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8500-8509 (2022) +24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023) +25. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs (2018) +26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models (2023) +27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts (2023) +28. Lin, T.Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P., Ramanan, D., Zitnick, C.L., Dollar, P.: Microsoft coco: Common objects in context (2015) +29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023) +30. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 53433-53456. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/a76b693f36916a5ed84d6e5b39a0dc03-Paper-Conference.pdf +31. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21736-21746 (June 2023) +32. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019) +33. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=IDJx97BC38 + +34. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13589-13595 (2021). https://doi.org/10.1109/ICRA48506.2021.9561801 +35. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips (2019) +36. OpenAI: Gpt-4 technical report (2023) +37. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +38. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation (2017) +39. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) +40. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d (2020) +41. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022) +42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022) +43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs (2021) +44. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018) +45. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) +46. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. Proceedings of the IEEE International Conference on Computer Vision (2019) +47. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022) +48. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., Dai, J.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks (2023) +49. Wang, Z., Huang, H., Zhao, Y., Zhang, Z., Zhao, Z.: Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769 (2023) +50. Xu, M., Ding, R., Zhao, H., Qi, X.: Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021) +51. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds (2023) + +52. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent (2023) +53. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Liu, Y., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest (2023) +54. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) +55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023) +56. Ziyu, Z., Xiaojian, M., Yixin, C., Zhidong, D., Siyuan, H., Qing, L.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023) \ No newline at end of file diff --git a/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/images.zip b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af8c6587b1136ad965ad9193cc8e32604cca2fc7 --- /dev/null +++ b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1487c1138a6f4fa2850c74110cb87f3f5cbc0b0f99bda27f3474f47d229da5e +size 549053 diff --git a/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/layout.json b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ecab24b19c594c46b5804a8f7209a7f0fd2fef63 --- /dev/null +++ b/2024/PARIS3D_ Reasoning-based 3D Part Segmentation Using Large Multimodal Model/layout.json @@ -0,0 +1,8922 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "text", + "content": "PARIS3D: Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 171, + 168, + 441, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 168, + 441, + 180 + ], + "spans": [ + { + "bbox": [ + 171, + 168, + 441, + 180 + ], + "type": "text", + "content": "Amrin Kareem1, Jean Lahoud1, and Hisham Cholakkal1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 154, + 190, + 459, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 190, + 459, + 201 + ], + "spans": [ + { + "bbox": [ + 154, + 190, + 459, + 201 + ], + "type": "text", + "content": "Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 232, + 452, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 232, + 452, + 439 + ], + "spans": [ + { + "bbox": [ + 160, + 232, + 452, + 439 + ], + "type": "text", + "content": "Abstract. Recent advancements in 3D perception systems have significantly improved their ability to perform visual recognition tasks such as segmentation. However, these systems still heavily rely on explicit human instruction to identify target objects or categories, lacking the capability to actively reason and comprehend implicit user intentions. We introduce a novel segmentation task known as reasoning part segmentation for 3D objects, aiming to output a segmentation mask based on complex and implicit textual queries about specific parts of a 3D object. To facilitate evaluation and benchmarking, we present a large 3D dataset comprising over 60k instructions paired with corresponding ground-truth part segmentation annotations specifically curated for reasoning-based 3D part segmentation. We propose a model that is capable of segmenting parts of 3D objects based on implicit textual queries and generating natural language explanations corresponding to 3D object segmentation requests. Experiments show that our method achieves competitive performance to models that use explicit queries, with the additional abilities to identify part concepts, reason about them, and complement them with world knowledge. Our source code, dataset, and trained models are accessible at this url." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 451, + 386, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 451, + 386, + 463 + ], + "spans": [ + { + "bbox": [ + 160, + 451, + 386, + 463 + ], + "type": "text", + "content": "Keywords: 3D " + }, + { + "bbox": [ + 160, + 451, + 386, + 463 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 451, + 386, + 463 + ], + "type": "text", + "content": " Vision-Language Models " + }, + { + "bbox": [ + 160, + 451, + 386, + 463 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 451, + 386, + 463 + ], + "type": "text", + "content": " Reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 484, + 230, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 484, + 230, + 496 + ], + "spans": [ + { + "bbox": [ + 133, + 484, + 230, + 496 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "The rapid advancements in 3D data capture technologies, including LIDARs and RGB-D cameras, have led to a growing demand for automated analysis of 3D point clouds. 3D semantic segmentation, the process of automatically assigning predefined semantic labels to each point in a cloud, is crucial for enabling complex tasks such as scene understanding. Similarly, 3D part segmentation involves further segmenting object instances into their components, such as identifying the handle of a pot or the lid of a bottle. These tasks find applications in various fields including autonomous vehicles, mobile robotics, industrial automation, augmented reality, and medical imagery analysis. While recent advancements in pre-trained 3D representations and the introduction of various 3D datasets have significantly improved 3D perception, the capacity for nuanced reasoning in 3D contexts remains limited. This limitation is primarily due to the lack of comprehensive datasets for reasoning and describing 3D scenes and objects." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 479, + 285 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 479, + 285 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 479, + 285 + ], + "type": "image", + "image_path": "4711b544402ea5357b6492736fd8756fbdd1ba22bb757f8c91df9071d0161ac2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 350 + ], + "type": "text", + "content": "Fig. 1: Capabilities of PARIS3D. Parts of 3D objects are segmented based on reasoning, shape, location, material, colour, and concept instructions. Additionally, for the segmentations, PARIS3D can explain why it chose that region, or describe 3D objects with respect to their parts. The original point clouds are on the left. The segmented parts are shown to the right, highlighted in golden colour." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 373, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 433 + ], + "type": "text", + "content": "In contrast, 2D images accompanied by textual descriptions have contributed to significant progress in large-scale image language models. Recent methods have emerged, performing segmentation based on 2D images projected from 3D scenes. However, they lack the ability to reason about object concepts and their 3D properties, functions, or parts in a conversational manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 435, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 604 + ], + "type": "text", + "content": "Here, we define a model as having reasoning capabilities when (i) it can understand implicit instructions, such as referencing properties of an object or its parts without explicit articulation, and (ii) explain or justify its responses, whether generated text or predicted segmentation masks. Such reasoning ability is one of the fundamental cognitive skills possessed by humans, essential for daily activities ranging from locating items to manipulating tools. Previous 3D vision methods are capable of exploring the semantics of a given environment based on instructions containing pre-defined category names. These explicit categories are limiting and do not address cases where instructions can be implicit and may contain complex queries. For example, an automated system might need to find the part of the kettle that can be opened to pour water, without knowing the explicit category name \"lid\". To this end, we introduce the 3D reasoning part segmentation task, which aims at segmenting object parts based on general queries, including color, 3D information, and implicit functionalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce a novel task termed reasoning part segmentation in 3D. This task involves generating a part segmentation mask for a 3D object based on implicit textual queries requiring complex reasoning. These queries go beyond simple references and encompass intricate descriptions, demanding sophisticated reasoning or worldly knowledge. For example, knowing where to hold a kettle or" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": "which part of a bottle to open in a fine-grained manner. To address this challenging task, we propose PARIS3D (ReasonIng-based 3D PArt Segmentation), a multi-modal Large Language Model (LLM) capable of reasoning on user input text, predicting 3D part segmentation masks, and providing explanations for the model's response. As depicted in Figure 1, PARIS3D adeptly navigates diverse scenarios, encompassing complex 3D reasoning, material, color, shape, and location-based knowledge, providing explanatory responses and detailed descriptions. Additionally, to validate the effectiveness of PARIS3D and support future research, we establish an evaluation benchmark and a dataset named RPSeg3D for reasoning 3D part segmentation. Our RPSeg3D dataset comprises 2624 3D objects and over 60k instructions, providing persuasive evaluation metrics for the task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 260, + 380, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 260, + 380, + 272 + ], + "spans": [ + { + "bbox": [ + 146, + 260, + 380, + 272 + ], + "type": "text", + "content": "In summary, our key contributions are the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 277, + 479, + 373 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 277, + 479, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 479, + 313 + ], + "type": "text", + "content": "- We introduce the reasoning part segmentation task for 3D objects, emphasizing the necessity of reasoning capabilities which is crucial for the development of intelligent perception systems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 314, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 314, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 314, + 479, + 348 + ], + "type": "text", + "content": "- We provide a comprehensive dataset named RPSeg3D for reasoning-based 3D part segmentation, comprising over 2624 3D objects and 60k instructions, serving as a useful resource for future research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 349, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 479, + 373 + ], + "type": "text", + "content": "- We present PARIS3D, a novel approach for 3D part segmentation, and further improve its capabilities by fine-tuning on our RPSeg3D dataset." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 390, + 237, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 237, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 237, + 403 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 593 + ], + "type": "text", + "content": "3D Semantic Segmentation. The challenge of understanding and reasoning within 3D environments has been an ongoing research focus. The goal of 3D semantic segmentation is to acquire semantic predictions for each point in a cloud. Notable contributions include point-based approaches [19, 38], methods which incorporate intricately crafted point convolution methods [46, 50], voxel-based strategies [5, 10], including those that employ 3D sparse convolutions [11] for generating point-wise segmentation outcomes, as well as transformer-based techniques [23]. Multi-view semantic segmentation methods such as DeepViewAgg [41], Diffuser [22, 34], 3D-CG [15], 3D-CLR [16] in 3D vision concentrate on improving representation learning by generating 2D renderings from 3D under multiple view points. These works have shown the effectiveness of multi-view representations in enhancing the performance and robustness of various 3D tasks. Nonetheless, these methods rely on a predefined set of semantic labels, whereas we focus in our proposed method on responding to complex reasoning-based queries and explaining them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Large Multimodal Models. Extensive research on Large Language Models (LLMs) has demonstrated reasoning capabilities, prompting an exploration into extending these skills into the visual domain through Large Multimodal Models. LMMs are a highly adaptable and versatile means to perform tasks requiring language and vision capabilities. Prominent models such as BLIP-2 [26], LLaVA [29] and MiniGPT-4 [55] generally employ a dual-phase training process, aligning" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 114, + 467, + 249 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 467, + 249 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 467, + 249 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 467, + 249 + ], + "type": "table", + "html": "
MethodInputTask
Point CloudReasoning QueryConversationSegmentationExplanation
SQA3D [33], 3D-VisTA [56]✓✓XXX
ViewRefer [13], Point-Bind [12]✓XXXX
3D-OVS [30]XXX✓X
OpenMask3D [45]✓XX✓X
PLA [6], OpenScene [37]✓XX✓X
Chat-3D [49]✓✓✓X✓
M3DBench [27]✓X✓XX
LLM-Grounder [52]✓X✓X✓
3D-LLM [17]✓✓✓X✓
LL3DA [4], PointLLM [51]✓X✓X✓
PARIS3D✓✓✓✓✓
", + "image_path": "986ea139b338629237f7acb1203f732929f3b5ad5b322a04537989679ad1bf9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 250, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 338 + ], + "type": "text", + "content": "Table 1: Comparison of recent 3D segmentation models and Large Multi-modal Models (LMMs) emphasizing their capabilities for 3D reasoning and conversations. Reasoning query means the model is asked to self-reason a task and either output text or perform an action. Segmentation highlights models that can respond with 3D segmentation masks, and Conversation represents models that can provide a conversation-style answer to the user. Among these, our proposed PARIS3D stands out with comprehensive 3D understanding and reasoning, segmentation in response to natural language queries, and conversational capabilities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 484 + ], + "type": "text", + "content": "visual representations with the linguistic embeddings of LLMs through extensive image-text and video-text datasets [2,3,28,35,42-44]. Recent efforts have focused on the convergence of multi-modal LLMs with vision tasks, where VisionLLM [48] provides a versatile interaction interface for a spectrum of vision-centric tasks through instruction tuning. Yet, it does not fully leverage the complex reasoning potential of LLMs. GPT4RoI [53] integrates spatial boxes as inputs and trains models on region-text pairs, showcasing a novel approach. Our method aims to benefit from these advances in the LMM space by merging the vision-language abilities of LMMs and the reasoning of LLMs in a novel 3D perception task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Language Instructed 3D Tasks. The integration of point clouds with natural language processing has widespread implications, drawing considerable interest in the realms of 3D scene understanding. This fast-growing field promises enhancements in human-robot interaction, metaverse, robotics, and embodied intelligence. Central to the dialogue systems designed for 3D environments are two critical capabilities: perception within three-dimensional spaces and reasoning. Recently, there has been a rise in the number of tasks uniting 3D scenes and language, such as 3D captioning, 3D question answering, 3D situated question answering, embodied Q and A, planning, navigation, 3D assisted multi-turn dialogue, 3D object detection, and scene description. We divide the 3D perception task models into 3 categories (see Table 1. Separated by dotted lines.). The first encompasses models that perform tasks like 3D captioning, situated question answering, and visual grounding [33, 56], visual grounding [14]. These models are capable of providing a word or phrase as its text output. The second category has 3D semantic segmentation models that output 3D segmentation masks. 3D" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 285 + ], + "type": "text", + "content": "OVS [30], Openmask3D [45], OpenScene [37], PLA [7], perform open-vocabulary semantic segmentation for 3D scenes. These methods, on the other hand, cannot provide a conversational output to a user query or explain the reasoning for their tasks. The third comprises models that employ an LLM and perform visual perception tasks such as captioning, scene understanding, and visual grounding, providing conversational outputs [4, 12, 17, 27, 49, 51, 52]. However, they do not perform fine-grained semantic segmentation or reasoning-based 3D vision tasks. As such, we identify a gap in performing 3D segmentation in response to complex natural language prompts instead of a single phrase or word. Distinct from the existing works in this domain, our research is committed to: (1) streamlining the integration of 3D segmentation capabilities into multi-modal LLMs, and (2) enhancing current perception systems with 3D reasoning-based segmentation and explanation abilities, thereby pushing the boundaries of what is achievable in the intersection of language and visual understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 302, + 353, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 302, + 353, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 353, + 316 + ], + "type": "text", + "content": "3 Reasoning 3D Part Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 328, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 508 + ], + "type": "text", + "content": "Semantic segmentation involves assigning a semantic label to each geometric primitive, such as points [38], voxels [9], or superpoints [25]. In part segmentation, object instances are decomposed into their components. Given a coloured point cloud P, the goal of a 3D segmentation model is to predict its label for each point. However, in our reasoning segmentation task, we go further to output a 3D segmentation mask M, given an input point cloud P and an implicit query text instruction " + }, + { + "bbox": [ + 130, + 328, + 482, + 508 + ], + "type": "inline_equation", + "content": "x_{txt}" + }, + { + "bbox": [ + 130, + 328, + 482, + 508 + ], + "type": "text", + "content": ". The task shares a similar formulation with the referring segmentation task [20], with an additional challenge for the model to reason about the fine-grained parts in response to implicit queries and output the corresponding segmentation mask. The complexity of the query text in reasoning part segmentation is a key differentiator. Instead of providing the names of the parts, the query text may include more intricate expressions that involve an understanding of structures, geometries, and semantics of 3D objects. By introducing this task, we aim to bridge the gap between user intent and system response, enabling more intuitive and dynamic interactions in 3D object perception." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 525, + 276, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 276, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 276, + 537 + ], + "type": "text", + "content": "3.1 Our RPSeg3D Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Considering the unavailability of established datasets and evaluation benchmarks in the literature, we introduce a dataset, named RPSeg3D, specifically designed for the reasoning 3D part segmentation task. Our dataset comprises 2624 3D objects and over 60k instructions. We use 718 objects and their corresponding instructions as the train set, and the remaining 1906 objects along with their instructions are used for testing. For reliable and fair assessment, we have aligned the 3D objects with those from PartNet-Ensemble [31], annotating them with implicit text instructions and using ground truth labels to generate high-quality target masks. To generate the text instructions corresponding to each 3D object part, we prepare a set of templates, for example, \"Which part of this object" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 242, + 145 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 242, + 145 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 242, + 145 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 242, + 145 + ], + "type": "image", + "image_path": "13fa119c38da0917bdff7b22fe309d1bce23dd4c0ba4652f6cad6095142ce70a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 148, + 241, + 161 + ], + "lines": [ + { + "bbox": [ + 136, + 148, + 241, + 161 + ], + "spans": [ + { + "bbox": [ + 136, + 148, + 241, + 161 + ], + "type": "text", + "content": "Normal Query: \"In this toaster, which part can be slid up and down for temperature adjustment?\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 164, + 240, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 164, + 240, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 164, + 240, + 182 + ], + "type": "text", + "content": "3D Query: \"What is the gray rectangular part in the middle of the Toaster that you can slide up and down to set the level of toasting?\"" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 188, + 224, + 240 + ], + "blocks": [ + { + "bbox": [ + 149, + 188, + 224, + 240 + ], + "lines": [ + { + "bbox": [ + 149, + 188, + 224, + 240 + ], + "spans": [ + { + "bbox": [ + 149, + 188, + 224, + 240 + ], + "type": "image", + "image_path": "298e393bbfcde62fc3d36675d0fd6eee35c1d11a8bd896a660ca9b185e28e873.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 244, + 242, + 257 + ], + "lines": [ + { + "bbox": [ + 140, + 244, + 242, + 257 + ], + "spans": [ + { + "bbox": [ + 140, + 244, + 242, + 257 + ], + "type": "text", + "content": "Normal Query: \"Which part of this kettle does the water come out of?\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 260, + 243, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 260, + 243, + 278 + ], + "spans": [ + { + "bbox": [ + 140, + 260, + 243, + 278 + ], + "type": "text", + "content": "3D Query: \"How long is the long cylindrical pipe-like brown part of this kettle that liquid comes out of compared to its whole length?\"" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 247, + 125, + 299, + 148 + ], + "blocks": [ + { + "bbox": [ + 247, + 125, + 299, + 148 + ], + "lines": [ + { + "bbox": [ + 247, + 125, + 299, + 148 + ], + "spans": [ + { + "bbox": [ + 247, + 125, + 299, + 148 + ], + "type": "image", + "image_path": "e61b493c240b4d2cfb1ddabceb47b73c26a89894af611cde6ec0b09b311d9697.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 149, + 356, + 160 + ], + "lines": [ + { + "bbox": [ + 252, + 149, + 356, + 160 + ], + "spans": [ + { + "bbox": [ + 252, + 149, + 356, + 160 + ], + "type": "text", + "content": "Normal Query: \"Which part of this keyboard used to type information?\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 250, + 164, + 357, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 164, + 357, + 182 + ], + "spans": [ + { + "bbox": [ + 250, + 164, + 357, + 182 + ], + "type": "text", + "content": "3D Query: \"Can you identify the white rectangular part of the keyboard that makes up about a tenth of its height and is located in the middle?\"" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 260, + 193, + 293, + 241 + ], + "blocks": [ + { + "bbox": [ + 260, + 193, + 293, + 241 + ], + "lines": [ + { + "bbox": [ + 260, + 193, + 293, + 241 + ], + "spans": [ + { + "bbox": [ + 260, + 193, + 293, + 241 + ], + "type": "image", + "image_path": "2ffd37fc7c354aa46fbb3abd3024d54e3e4d3dbf5424fad98dc470b8d98310a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 244, + 354, + 257 + ], + "lines": [ + { + "bbox": [ + 250, + 244, + 354, + 257 + ], + "spans": [ + { + "bbox": [ + 250, + 244, + 354, + 257 + ], + "type": "text", + "content": "Normal Query: \"Which part of this Bucket provides a grip for carrying it?\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 250, + 260, + 353, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 260, + 353, + 277 + ], + "spans": [ + { + "bbox": [ + 250, + 260, + 353, + 277 + ], + "type": "text", + "content": "3D Query: \"What is the shape of the gray part of this Bucket located in front of and on the side of it? What is it used for?\"" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 125, + 361, + 148 + ], + "blocks": [ + { + "bbox": [ + 309, + 125, + 361, + 148 + ], + "lines": [ + { + "bbox": [ + 309, + 125, + 361, + 148 + ], + "spans": [ + { + "bbox": [ + 309, + 125, + 361, + 148 + ], + "type": "image", + "image_path": "987d35ac10ca66fd06153ec7e92d191ab2c68f5db44bcd77ca3addc8cada906f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 365, + 121, + 416, + 148 + ], + "blocks": [ + { + "bbox": [ + 365, + 121, + 416, + 148 + ], + "lines": [ + { + "bbox": [ + 365, + 121, + 416, + 148 + ], + "spans": [ + { + "bbox": [ + 365, + 121, + 416, + 148 + ], + "type": "image", + "image_path": "e366f35541b633c3a95203b013b090da930a607a47ae18e81dcad8410b0c0c35.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 149, + 474, + 160 + ], + "lines": [ + { + "bbox": [ + 369, + 149, + 474, + 160 + ], + "spans": [ + { + "bbox": [ + 369, + 149, + 474, + 160 + ], + "type": "text", + "content": "Normal Query: \"Which part of this storage furniture is drawn out and used for storage?\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 164, + 470, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 164, + 470, + 182 + ], + "spans": [ + { + "bbox": [ + 375, + 164, + 470, + 182 + ], + "type": "text", + "content": "3D Query: \"In this StorageFurniture, which is the rectangular part used for storage that constitutes about a quarter of its total height?\"" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 378, + 184, + 407, + 243 + ], + "blocks": [ + { + "bbox": [ + 378, + 184, + 407, + 243 + ], + "lines": [ + { + "bbox": [ + 378, + 184, + 407, + 243 + ], + "spans": [ + { + "bbox": [ + 378, + 184, + 407, + 243 + ], + "type": "image", + "image_path": "97ba835333ffc25140ea1d4685080dbcb2d8830664ed04d870e4de2affd2b593.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 244, + 477, + 257 + ], + "lines": [ + { + "bbox": [ + 371, + 244, + 477, + 257 + ], + "spans": [ + { + "bbox": [ + 371, + 244, + 477, + 257 + ], + "type": "text", + "content": "Normal Query: \"Which part of this door surrounds the entrance?\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 375, + 260, + 474, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 260, + 474, + 277 + ], + "spans": [ + { + "bbox": [ + 375, + 260, + 474, + 277 + ], + "type": "text", + "content": "3D Query: \"What is the shape of the brown part of this Door located on top of and on the sides of it? What is it used for?\"" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 133, + 332, + 482, + 385 + ], + "blocks": [ + { + "bbox": [ + 131, + 291, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 131, + 291, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 482, + 326 + ], + "type": "text", + "content": "Fig. 2: Examples of the annotated object-instruction pairs for training with two types of queries. On the left is one view of the rendered image from the original point cloud. On the right is the corresponding ground truth segmentation mask, shown in green." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 332, + 482, + 385 + ], + "lines": [ + { + "bbox": [ + 133, + 332, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 482, + 385 + ], + "type": "image", + "image_path": "f927279a59d0691d64d076c84329ef5ec182c7f6bf77657851fa278df3d78856.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 392, + 482, + 449 + ], + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 449 + ], + "type": "text", + "content": "Fig. 3: Preparing the instructions of RPSeg3D. Simple templates are provided to GPT-3.5, which populates them with part information-related segmentation instructions. In parallel, colour, shape, location, and dimension-related data is extracted from 3D point clouds. Enriching the instructions with this information and manually checking them for inaccuracies, we obtain the RPSeg3D dataset for part segmentation." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 464, + 483, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 483, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 483, + 608 + ], + "type": "text", + "content": "?\" We leverage GPT-3.5 [36] for building instructions using these templates by supplying information about the part and rephrasing. We also extract 3D information from the point cloud, thus designing the instructions to cover relations, dimensions (length, height), comparisons, colour, texture features, object concepts, and functions. This was further verified manually to avoid inaccurate prompts at test time. The steps are illustrated in Figure 3. To cover 3D object segmentation tasks effectively, our text instructions consist of two types: 1) normal queries; 2) 3D queries; as illustrated in Figure 2. The dataset is partitioned into train and test splits, containing 718 and 1906 3D objects and over 16k and 47k instructions, respectively. As the primary purpose of the dataset is evaluation, the testing set includes a larger number of instructions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 624, + 397, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 397, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 397, + 636 + ], + "type": "text", + "content": "3.2 Reasoning 3D Part Segmentation Architecture" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Our method takes as input a dense and coloured 3D point cloud of an object. One of the common methods of 3D analysis is predominantly using point clouds" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "to represent 3D data. However, this contrasts with human spatial reasoning processes. Humans typically engage with their surroundings through active exploration, synthesizing perspectives from multiple vantages to form an integrated 3D understanding, rather than processing a 3D environment at one go. Our approach advocates for 3D reasoning derived from multi-view imagery. This approach also benefits from the large-scale 2D pretraining available in vision-language models, similar to previous methods that have taken advantage of pre-trained vision-language models for 3D vision tasks. Thus, we render multiple images " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "x_{img}" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " from K predefined camera poses by rasterization. The camera poses cover all parts of the object since they are uniformly distributed around the input point cloud. Given a complex text instruction " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "x_{txt}" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " along with the images, we feed them into a multimodal LLM, denoted by " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ", and the visual backbone, " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "F_{enc}" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " outputs a text response " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\hat{y}_{txt}" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ". In parallel, " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "F_{enc}" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " extracts the visual embeddings, f, from each of the input images " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "x_{img}" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ". The formulation is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 295, + 481, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 295, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 228, + 295, + 481, + 307 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t x t} = F \\left(x _ {i m g}, x _ {t x t}\\right), f = F _ {e n c} \\left(x _ {i m g}\\right) \\tag {1}", + "image_path": "56840b801b5f204aa7ab54addf26283bdfbbb99ef0c7f2c218194714278a7cbb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": "Whenever the LLM is expected to yield a binary segmentation mask, the resultant text output, denoted as " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\hat{y}_{txt}" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": " is required to an extra token, which sends a request for the segmentation output. Following this, the embedding at the last layer, " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\hat{h}_{seg}" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": ", corresponding to the additional segmentation token is extracted. This embedding is subsequently processed through an MLP (Multilayer Perceptron) projection layer, represented as " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": ", to derive " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "h_{seg}" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": " and then input to the mask decoder " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "F_{dec}" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": " as its prompt embedding, thereby tightly coupling the text output from " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": " and the mask predicted by the decoder. Hence, the mask decoder takes the visual features " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SEG} \\rangle" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": " based prompt embeddings to perform segmentation, generating the final segmentation mask, " + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 130, + 316, + 482, + 449 + ], + "type": "text", + "content": " similar to [24]. The process is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 276, + 460, + 480, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 460, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 276, + 460, + 480, + 473 + ], + "type": "interline_equation", + "content": "h _ {s e g} = \\gamma (\\hat {h} _ {s e g}) \\tag {2}", + "image_path": "380040880de6adb75094bfbbf3641469ef1aab5a7a08c8b4caf9260181b2bd54.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 479, + 470, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 479, + 470, + 491 + ], + "spans": [ + { + "bbox": [ + 135, + 479, + 470, + 491 + ], + "type": "text", + "content": "The architectural specifics of the decoder, " + }, + { + "bbox": [ + 135, + 479, + 470, + 491 + ], + "type": "inline_equation", + "content": "F_{dec}" + }, + { + "bbox": [ + 135, + 479, + 470, + 491 + ], + "type": "text", + "content": ", are in accordance with [21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 268, + 503, + 480, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 503, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 268, + 503, + 480, + 516 + ], + "type": "interline_equation", + "content": "\\hat {M} = F _ {\\text {d e c}} \\left(h _ {\\text {s e g}}, f\\right) \\tag {3}", + "image_path": "b7c5e4470075b610287337386477067cf06b501b38d25dc44784b9c611e67864.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Once the K segmentations are obtained from each camera angle, a 3D semantic voting module computes scores and assigns the semantic labels for each part. We have " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " semantic segmentation masks " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{k}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is the view from which the image was rendered from the point cloud. We aggregate the masks from multiple views and lift them to obtain a three-dimensional semantic segmentation of the original point cloud. To achieve this, we segment the input point cloud " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " into a set of superpoints " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "SP_{i}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", similar to [25]. Superpoints refer to a way of representing large 3D point clouds as a collection of interconnected geometrically simple shapes. This representation is advantageous because (i) it considers complete object parts as a whole, making them easier to identify, (ii) this method can provide a detailed description of the relationship between adjacent objects, and (iii) the size of superpoints is determined by the number of simple structures in a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 224 + ], + "type": "image", + "image_path": "cfcc961beadcc49a08be7acda431c29681d340f37872c49c648b7baa2c3a8e99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "text", + "content": "Fig. 4: Overview of the proposed reasoning-based 3D part segmentation approach named PARIS3D. It comprises four subsequent steps: (i) The 3D point cloud is rendered into K multi-view images " + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "x_{img}" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "text", + "content": " using a renderer. (ii) These images are passed through a frozen vision backbone (" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "F_{enc}" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "text", + "content": ") and multi-modal large language model (" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "text", + "content": ") of the reasoning module. " + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "text", + "content": " also accepts the text query " + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "x_{txt}" + }, + { + "bbox": [ + 130, + 232, + 482, + 331 + ], + "type": "text", + "content": ", and produces text outputs corresponding to each view. (iii) The decoder decodes the final layer embedding which contains the extra token, thus producing K segmentation masks. (iv) Finally, a mask to 3D segmentation algorithm lifts the projections back into 3D and a view-guided scoring module is used to obtain the final text response." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "text", + "content": "scene, rather than the total number of points, making the representation several orders of magnitude smaller, thus more efficient to work with. Each superpoint contains points with similar normals and colours, suggesting they belong to the same instance. This superpoint-based part labelling not only conserves computational resources but also potentially improves performance by leveraging 3D priors. In the 3D semantic voting step, we assign semantic labels for each superpoint through a voting mechanism similar to [31]. This technique leverages information from multiple views and the superpoints, so that even if one mask covers irrelevant points, the aggregation of masks from all views counters the effect of such errors. For a superpoint " + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "inline_equation", + "content": "SP_{i}" + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "text", + "content": " and a part category " + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "text", + "content": ", a score " + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "inline_equation", + "content": "s_{i,j}" + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "text", + "content": " is computed reflecting the proportion of visible points present in the predicted segmentation masks of category " + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 347, + 482, + 491 + ], + "type": "text", + "content": " in each view:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 497, + 481, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 497, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 212, + 497, + 481, + 528 + ], + "type": "interline_equation", + "content": "s _ {i, j} = \\frac {\\sum_ {k} \\sum_ {p \\in S P _ {i}} \\left[ V _ {k} (p) \\right] \\left[ \\exists b \\in M _ {k} ^ {j} : I _ {m} (p) \\right]}{\\sum_ {k} \\sum_ {p \\in S P _ {i}} \\left[ V _ {k} (p) \\right]} \\tag {4}", + "image_path": "32b361617ffb7257b2c317ef4084c3ae74bf05e456caa27ce7802f60267ab33b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " is the Iverson bracket, which turns true predicates into 1 and false into 0. " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "M_k^j" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": ", represents the predicted mask of category j in view k. " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "V_{k}(p)" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " indicates the visibility of 3D point p in the view k. " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{m}(p)" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " checks if the projection of point p in view k falls within the mask m. The superpoint is then assigned the semantic label of the part category with the highest score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To provide the most relevant text explanation corresponding to the 3D masks, we propose a view-guided scoring module. It assigns a score to each view based on its correspondence with the final semantic labels. The best text explanation is chosen based on the highest score. Specifically, for each view, we compute the intersection over union (IoU) of the mask with the projections of the final 3D semantic labels output by the model, and the corresponding text explanation is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "saved. This is called the view-guided score of the explanation. At the end of the score computation for all the masks, we choose the text explanation with the highest score, and in turn, the most correspondence with the output labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 166, + 344, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 166, + 344, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 166, + 344, + 178 + ], + "type": "text", + "content": "3.3 Training the PARIS3D Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 312 + ], + "type": "text", + "content": "Training Data Formulation The training data consists of 718 3D objects rendered into multiview images, resulting in over 16k image-instruction pairs. Out of the 718, 360 objects with their instructions are used for training the model and 358 objects with their instructions are used for validation. Each image is provided with an annotation file, which has instructions corresponding to the 3D point cloud that the image came from, the name of the image, and its ground truth mask. Generation of instructions follows the same steps as Section 3.1 illustrated in Figure 3. Images in the training set may have more than one instruction in its \"instruction\" field. This is helpful to introduce diversity as users may randomly select one as the reasoning query during training, thus obtaining a better model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 480, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 447 + ], + "type": "text", + "content": "Distillation-Based Explanation Refining Training the multiview model requires explanatory data as well as training image-instruction pairs. To build the explanatory data, we built an annotation pipeline using a distillation approach. Using the multi-view images as input to the teacher model [24], we generate explanations for each part of the object. These explanations serve as our pseudolabels which we use as ground truth explanations of the student model. We augment these annotations with 3D features extracted from the 3D point clouds from which the images were rendered. The 3D feature-augmented explanations contain appropriate responses and critical elements differentiating object parts from each other such as location, size, shape, material, and colour." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "content": "Objective We leverage LoRA [18] to perform efficient fine-tuning of the pretrained multi-modal LLM " + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "content": " [29] to retain its generalization capability. We completely freeze the vision backbone " + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "inline_equation", + "content": "F_{enc}" + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "content": ". The decoder " + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "inline_equation", + "content": "F_{dec}" + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "content": " is fully fine-tuned. The word embeddings of the LLM and the projection layer of " + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "content": " are also trainable, allowing the model to learn the specific meanings and semantic concepts of parts. The model is trained end-to-end using the text generation loss " + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "inline_equation", + "content": "L_{txt}" + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "content": " and the segmentation mask loss " + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "inline_equation", + "content": "L_{mask}" + }, + { + "bbox": [ + 130, + 460, + 480, + 544 + ], + "type": "text", + "content": ". The overall objective is given as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 551, + 480, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 551, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 247, + 551, + 480, + 562 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {t x t} L _ {t x t} + \\lambda_ {m a s k} L _ {m a s k} \\tag {5}", + "image_path": "cd9d43f8100a6ca35c38c0e406da0e4d5df4b47ebe96c8737a3e0c74a3f9a190.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "inline_equation", + "content": "L_{mask}" + }, + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "text", + "content": " encourages the model to produce high-quality segmentation results. It is a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights " + }, + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 130, + 567, + 479, + 602 + ], + "type": "text", + "content": ", given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 202, + 607, + 480, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 607, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 202, + 607, + 480, + 620 + ], + "type": "interline_equation", + "content": "L _ {m a s k} = \\lambda_ {b c e} B C E (\\hat {M}, M) + \\lambda_ {d i c e} D I C E (\\hat {M}, M) \\tag {6}", + "image_path": "3b6032438c5da4c940a2d40bc977f42c92b48d17438c37489d0363eaf39ddee5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "inline_equation", + "content": "y_{txt}" + }, + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "text", + "content": " are the ground-truth targets. " + }, + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "inline_equation", + "content": "L_{txt}" + }, + { + "bbox": [ + 130, + 624, + 479, + 648 + ], + "type": "text", + "content": " is the auto-regressive cross-entropy loss for text generation. It is computed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 261, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 261, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "L _ {t x t} = C E \\left(\\hat {y} _ {t x t}, y _ {t x t}\\right) \\tag {7}", + "image_path": "44125a0713a8b01cd48a55ad46895ce1fc3127ead2e995006a6b27c05d9a834c.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 164 + ], + "type": "text", + "content": "We perform quantitative and qualitative evaluation of PARIS3D on our dataset for reasoning-based semantic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": "Implementation Details and Metrics For the experiments, we follow [24] where the multi-modal LLM F is LLaVA-13B-v1-1 [29] and the vision backbone " + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "inline_equation", + "content": "F_{enc}" + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": " is the ViT-H SAM. The projection layer of " + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": " is an MLP with channels of [256, 4096, 4096]. We use our dataset RPSeg3D which contains coloured point clouds and rendered 2D images of them. Using Pytorch3D [40], each input point cloud is rendered into " + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 10" + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": " colour images. The fine-tuning scripts for the LLaVA and SAM architecture are based on DeepSpeed [1] engine. We adopt the settings of [24] for the optimizer (AdamW [32]) and its learning rate (0.0003). Similar steps are followed for the learning rate scheduler (WarmupDecayLR), text generation loss " + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\lambda_{txtgen}" + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": " weight (1.0) and the mask loss " + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": " weight (1.0), the BCE loss " + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": " (2.0), dice loss " + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 132, + 182, + 482, + 362 + ], + "type": "text", + "content": " (0.5), batch size per device (2), and the gradient accumulation step (10). The semantic segmentation metric used is category mIoU, following [31]. It is calculated as follows: first, mIoU for each part category is calculated for all test objects. Then, part mIoUs that belong to each object category are averaged to compute the object category mIoU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 380, + 351, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 351, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 351, + 392 + ], + "type": "text", + "content": "4.1 Reasoning Part Segmentation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 399, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 482, + 507 + ], + "type": "text", + "content": "We establish the reasoning part segmentation task on our dataset RPSeg3D. Table 2 shows the results of reasoning part segmentation. We observe that without any fine-tuning, the model's performance is low compared to its fine-tuned counterparts. The general observation is that 3D queries help the model to output better masks. We compare our method to two baselines. The first baseline is LISA [24] applied to the multiple views without any finetuning. The second baseline consists of LISA finetuned on few-shot part segmentation data. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 525, + 471, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 471, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 471, + 538 + ], + "type": "text", + "content": "4.2 3D Semantic Segmentation Comparison with Existing Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": "Table 3 shows the results of semantic segmentation compared to existing methods. Our method has better performance than all the fully supervised baselines and achieves competitive results with [31]. For this baseline, we observe that the model with prompt tuning done separately on each category achieves impressive few-shot performance but when unified into a single model for all categories, the performance is only slightly better than its zero-shot performance " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(27.2\\%)" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": ". This is attributed to the redundant part names across object categories, hindering the model's learning of the semantic meanings of part names. This results in performance to drop significantly from when only one category is learnt per model. To prompt for segmentation, the baseline models [38,39,54] are provided with the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 114, + 446, + 171 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 446, + 171 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 446, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 446, + 171 + ], + "type": "table", + "html": "
MethodValTest
Normal Query3D QueryOverallNormal Query3D QueryOverall
LISA-MV16.6020.1618.3817.6020.5719.08
LISA-MV (ft)50.4350.2850.3550.7550.8150.78
PARIS3D55.3355.5055.4255.9457.6056.77
", + "image_path": "718045743a50b104b3a246e96408e3e19f5929d490123c10e06c10de5368c7e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 268, + 481, + 364 + ], + "blocks": [ + { + "bbox": [ + 130, + 171, + 480, + 246 + ], + "lines": [ + { + "bbox": [ + 130, + 171, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 480, + 246 + ], + "type": "text", + "content": "Table 2: Results of reasoning part segmentation. LISA-MV [24] is LISA in multi-view setting without fine-tuning. LISA-MV (ft) is the experiment in which it has been fine-tuned on few-shot part segmentation data. Our proposed PARIS3D method has been fine-tuned with 3D queries and explanations. When fine-tuned with 3D information, our model performs better than the baselines for normal and 3D prompts. Here Test is the test set of 1906 shapes and Val is the validation set of 358 3D shapes with their instructions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 268, + 481, + 364 + ], + "lines": [ + { + "bbox": [ + 133, + 268, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 481, + 364 + ], + "type": "table", + "html": "
3D DataMethodOverlapping CategoriesNon-overlapping CategoriesOverall(45)
Bottle ChairDisplay DoorKnife LampStorage-FurnitureTableOverall(17)CameraCartDis-PenserKettleKitchenPotOvenSuit-caseToasterOverall(28)
Extra data(45x8+28k)PointNet++ [38]48.884.778.445.735.468.046.963.755.66.56.412.120.915.834.340.614.725.436.8
PointNeXt [39]68.491.889.443.858.764.968.552.158.533.236.326.045.157.037.813.58.345.150.2
SoftGroup [47]41.488.362.153.131.382.260.254.850.223.623.918.957.445.513.618.326.430.738.1
Few-shot(45x8)PartSLIP* [31]83.485.384.840.865.266.053.642.456.358.388.173.777.069.673.570.460.061.359.4
PointNet++ [38]27.042.230.220.522.210.58.47.318.19.711.67.028.631.719.43.30.021.820.4
PointNeXt [39]67.665.153.746.359.755.420.622.139.226.047.722.660.566.036.814.50.041.540.6
SoftGroup [47]20.880.539.716.338.338.318.924.932.828.640.842.960.754.835.629.814.841.138.0
ACD [8]22.439.029.218.939.613.77.613.519.210.131.519.440.251.88.913.20.025.623.2
Prototype [54]60.170.867.333.450.438.230.225.741.132.036.853.462.763.336.535.510.146.344.3
PartSLIP+64.869.559.524.534.537.132.040.135.325.575.715.630.558.431.149.46.626.729.9
Ours84.081.070.168.447.261.239.445.155.129.371.740.159.378.859.161.624.959.157.6
", + "image_path": "5790435ce1b42ee27cac1f0d110118b9f49056ec08481571be46c827fcfe78c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 441 + ], + "type": "text", + "content": "Table 3: Comparison to previous 3D part segmentation methods. Object category mIoU(%) is shown. In the " + }, + { + "bbox": [ + 130, + 365, + 482, + 441 + ], + "type": "inline_equation", + "content": "45 \\times 8 + 28k" + }, + { + "bbox": [ + 130, + 365, + 482, + 441 + ], + "type": "text", + "content": " setting, baseline models use an additional " + }, + { + "bbox": [ + 130, + 365, + 482, + 441 + ], + "type": "inline_equation", + "content": "28k" + }, + { + "bbox": [ + 130, + 365, + 482, + 441 + ], + "type": "text", + "content": " training shapes for 17 overlapping object categories. These are categories present in common with PartNet dataset. For the remaining 28 non-overlapping object categories, there are only 8 shapes per object category during training. PartSLIP* indicates that one model has been trained for each category. + shows our implementation of PartSLIP where one model is trained for all the categories together." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "content": "class IDs of the parts to be segmented (e.g: 0 to 103 including the background). In [31], the model is provided with the names of the parts to be segmented (e.g. \"seat\", \"arm\", \"back\" of a chair). Our method, PARIS3D, uses language instructions to prompt segmentation. For a fair comparison, we use hand-crafted prompts with short instructions containing generic concept and location-based clues about the part to be segmented." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 563, + 238, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 238, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 238, + 574 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 666 + ], + "type": "text", + "content": "Instruction Rephrasing. We perform additional experiments with varying instruction types to study the merits of using rephrased prompts in training the model (Table 4 here) compared to using only category names and manually generated prompts for each category. These experimental results on the test set show that GPT-generated prompts provide superior performance, both without and with color and 3D information. In both cases, performance improves from using category names to manual prompts to GPT prompts, indicating the strong" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 251, + 178 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 251, + 178 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 251, + 178 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 251, + 178 + ], + "type": "image", + "image_path": "ac5ab5acba333b66e1e8a5cb61a47bda707edcfb258d110a6284e5907a7a48d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 116, + 479, + 178 + ], + "blocks": [ + { + "bbox": [ + 256, + 116, + 479, + 178 + ], + "lines": [ + { + "bbox": [ + 256, + 116, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 256, + 116, + 479, + 178 + ], + "type": "image", + "image_path": "0f56d92919fbc3932df4a802f31d19c15fd6833471ea0cf2d590b80d0507578f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 180, + 240, + 236 + ], + "blocks": [ + { + "bbox": [ + 143, + 180, + 240, + 236 + ], + "lines": [ + { + "bbox": [ + 143, + 180, + 240, + 236 + ], + "spans": [ + { + "bbox": [ + 143, + 180, + 240, + 236 + ], + "type": "image", + "image_path": "1a5802da97d5113cb9abba88298ae92f1d2536859b17bd1b759cca8ab4591baa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 181, + 330, + 236 + ], + "blocks": [ + { + "bbox": [ + 242, + 181, + 330, + 236 + ], + "lines": [ + { + "bbox": [ + 242, + 181, + 330, + 236 + ], + "spans": [ + { + "bbox": [ + 242, + 181, + 330, + 236 + ], + "type": "image", + "image_path": "1ca4b3e964b5a6dc61f92d73bd65c63cf5a35e3087d8bb638f6dca8bffc4ff37.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 181, + 479, + 236 + ], + "blocks": [ + { + "bbox": [ + 331, + 181, + 479, + 236 + ], + "lines": [ + { + "bbox": [ + 331, + 181, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 331, + 181, + 479, + 236 + ], + "type": "image", + "image_path": "55bcaa69480aaf021c4dc74292ae372faa0329104848855a490f3e59fdea417e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 237, + 237, + 293 + ], + "blocks": [ + { + "bbox": [ + 143, + 237, + 237, + 293 + ], + "lines": [ + { + "bbox": [ + 143, + 237, + 237, + 293 + ], + "spans": [ + { + "bbox": [ + 143, + 237, + 237, + 293 + ], + "type": "image", + "image_path": "543cbfc730632ebead5d5c98ce5fa329cc302306772b6087ad824e2e6ae43be2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 348 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 348 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of PARIS3D's performance. We showcase examples from three tasks: reasoning 3D object part segmentation, object description, and reasoning question-answering, demonstrating its capabilities in offering in-depth reasoning, 3D understanding, part segmentation, and conversational abilities." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 241, + 237, + 336, + 292 + ], + "blocks": [ + { + "bbox": [ + 241, + 237, + 336, + 292 + ], + "lines": [ + { + "bbox": [ + 241, + 237, + 336, + 292 + ], + "spans": [ + { + "bbox": [ + 241, + 237, + 336, + 292 + ], + "type": "image", + "image_path": "0be43e828d54c1968801690d851e0786415948b72579b18cad191020617ebdac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 336, + 237, + 479, + 293 + ], + "blocks": [ + { + "bbox": [ + 336, + 237, + 479, + 293 + ], + "lines": [ + { + "bbox": [ + 336, + 237, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 336, + 237, + 479, + 293 + ], + "type": "image", + "image_path": "e9a017b4ea1cd2dae4e60beb7941e7bd55908b4f20a915e9dffde443b08ce804.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 153, + 355, + 465, + 405 + ], + "blocks": [ + { + "bbox": [ + 153, + 355, + 465, + 405 + ], + "lines": [ + { + "bbox": [ + 153, + 355, + 465, + 405 + ], + "spans": [ + { + "bbox": [ + 153, + 355, + 465, + 405 + ], + "type": "table", + "html": "
Instruction TypeCategory NameManually GeneratedGPT-generatedCategory NameManually generatedGPT-generated
Colour + 3DXXX✓✓✓
Overall mIoU38.041.546.743.052.457.6
", + "image_path": "54a7566db3918461817c020de6a41db47e77a69caa1047600da7d95ec8b4d971.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 406, + 482, + 461 + ], + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 461 + ], + "type": "text", + "content": "Table 4: A study on the effectiveness of rephrased prompts on the overall model performance. Here, the ability of the model to accurately segment an object part based on an implicit (reasoning-based) description is evaluated for different types of training prompts. The results show the superior performance of GPT-generated prompts over manual prompts, both with and without colour and 3D information." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "content": "generalization capabilities introduced by GPT-based prompts. For instance, using color and 3D information prompts along with 'manual prompts' resulted in an overall mIoU of " + }, + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "inline_equation", + "content": "52.4\\%" + }, + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "content": ". However, using GPT to expand simple templates into varied prompts for describing object part categories resulted in about a " + }, + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "content": " improvement over the manually curated training templates (last vs second-last column)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "One Model vs Multiple Models. One of the baseline models [31] trained a model for each of the 45 categories, loading one at a time to evaluate each category. When we replicate this setting for PARIS3D, there are significant jumps in performance as shown in Table 5. However, one-model-for-one-category does not offer a generalizable solution in a real-world problem setting, where multiple object categories and their parts need to be analysed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Number of Rendered Views. The effect of incrementing the number of views rendered from the input point cloud is shown in Table 6. With only 1 view in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 183, + 114, + 430, + 155 + ], + "blocks": [ + { + "bbox": [ + 183, + 114, + 430, + 155 + ], + "lines": [ + { + "bbox": [ + 183, + 114, + 430, + 155 + ], + "spans": [ + { + "bbox": [ + 183, + 114, + 430, + 155 + ], + "type": "table", + "html": "
MethodChair DispenserKeyboard EyeglassesesBucket
PartSLIP [31]85.373.753.688.336.5
PARIS3D86.575.688.292.283.9
", + "image_path": "a8d260ebf372b042f542b6c6f2f017bdfe43f65899f3efaf92f17529c30c71be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 254, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 479, + 301 + ], + "type": "text", + "content": "formation provided, the segmentations are not meaningful. With 5 views, the segmentation quality improves, and improves further with an increase in the number of views. For each point cloud, we use 10 rendered views in our main experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 384 + ], + "type": "text", + "content": "Single prompt vs Multiple prompts. In Table 7, we show ablation experiments regarding the number of training prompts. Single prompt refers to one concept-based reasoning training instruction that follows a template. In further experiments, we use the template to generate multiple (about 5-6) rephrased instructions for each part category using [36]. Training with these multiple training prompts per part category results in a better-performing model than with a single prompt." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 138, + 418, + 286, + 472 + ], + "blocks": [ + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "text", + "content": "Table 5: Results of PARIS3D on one model trained for each of 5 categories. In [31], 45 models were trained and each point cloud was tested by loading its corresponding model to perform evaluation. Repeating this exact setting by training a model on each category for PARIS3D, we easily gain " + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "inline_equation", + "content": "+1.9\\%" + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "inline_equation", + "content": "+34.6\\%" + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "inline_equation", + "content": "+3.9\\%" + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "inline_equation", + "content": "+47.4\\%" + }, + { + "bbox": [ + 130, + 156, + 482, + 211 + ], + "type": "text", + "content": " improvements on the tested categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 418, + 286, + 472 + ], + "lines": [ + { + "bbox": [ + 138, + 418, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 286, + 472 + ], + "type": "table", + "html": "
No. of viewsBottleBucketCamera
11.11.11.2
523.740.112.1
1084.884.929.3
", + "image_path": "2ec45b4d75a2f6a56d1bc9ea375fc1139d27234dcf000b72993376d400f576c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 291, + 419, + 460, + 460 + ], + "blocks": [ + { + "bbox": [ + 291, + 419, + 460, + 460 + ], + "lines": [ + { + "bbox": [ + 291, + 419, + 460, + 460 + ], + "spans": [ + { + "bbox": [ + 291, + 419, + 460, + 460 + ], + "type": "table", + "html": "
No. of promptsBottleKnifeClockChair
173.844.345.569.5
5-683.952.751.378.7
", + "image_path": "5d55b7cced23566582289d728d1c3fa4283c01be08aa473100c07aef470a9725.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 472, + 449, + 494 + ], + "lines": [ + { + "bbox": [ + 132, + 472, + 449, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 449, + 494 + ], + "type": "text", + "content": "Table 6: Ablation experiments on the number of prompts provided in train-number of rendered views. ing data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 555, + 257, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 257, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 257, + 567 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "In Figure 5, we provide examples of PARIS3D output for 3D object part segmentation tasks. PARIS3D is capable of segmenting fine-grained details in categories like buttons of a remote or legs of an eyeglass from implicit queries that expect it to do self-reasoning to generate the right answer, and then segment the part. Moreover, it can perform tasks like object description and question-answering with regard to object parts and give detailed answers. This demonstrates its conversational, reasoning, and part segmentation capabilities." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 90, + 447, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 482, + 206 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 482, + 206 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 482, + 206 + ], + "type": "image", + "image_path": "8540031978837ec9a5e228819188bef2cdbdecf1064807b26ae5dcb6fc535c09.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": "Fig. 6: Qualitative results on real-world examples of point clouds. The input point cloud is shown on the left and the segmentation mask is in yellow on the right. Our model generalizes to point clouds with a significant domain gap from those it was trained on and displays impressive results with segmenting small object parts such as the furniture's knob in this example." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 285, + 369, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 369, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 369, + 298 + ], + "type": "text", + "content": "4.5 Generalizability to Real-World Examples" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 303, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 434 + ], + "type": "text", + "content": "Most existing 3D tasks and their corresponding models can usually only handle cases within the same distribution of the training sets without generalization, since they are sensitive to the format of the input and significant domain gap between synthetic experiments and real-world examples. Thus, it can be difficult to use them in use cases involving real point clouds from an everyday setting. To demonstrate the generalizability of PARIS3D to data derived from the real world, we perform our 3D segmentation on real point clouds shot using a smartphone's LiDAR sensor, as suggested by [31]. In Figure 6, we show qualitative examples of passing the fused point clouds through the PARIS3D architecture to obtain part segmentation labels as in the previous experiments without much drop in performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": "In this work, we introduce a novel challenge within 3D segmentation, reasoning-based part segmentation. This task requires models to infer, reason, and explain based on implicit user instructions, making it considerably more complex than the regular 3D referring segmentation task. We introduce a dataset for this task, RPSeg3D, to enable effective evaluation. We believe this dataset will play a crucial role in fostering the growth of technologies in this area. Additionally, we outline a pipeline that integrates 3D segmentation capabilities into multimodal Large Language Models (LLMs), showcasing our model, PARIS3D, which exhibits competitive performance. It additionally demonstrates the ability to identify part concepts, reason about them, and complement them with world knowledge. However, we identify limitations: Although the model has improved generalizability, there is a possibility that the model may occasionally result in inaccurate predictions while being prompted with inappropriate, incorrect, or biased text inputs. Also, the model in its current form cannot perform instance segmentation. This is a direction for future research as we expand the dataset to accommodate such tasks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Aminabadi, R.Y., Rajbhandari, S., Zhang, M., Awan, A.A., Li, C., Li, D., Zheng, E., Rasley, J., Smith, S., Ruwase, O., He, Y.: Deepspeed inference: Enabling efficient inference of transformer models at unprecedented scale (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "text", + "content": "2. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "text", + "content": "3. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 220, + 480, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 253 + ], + "type": "text", + "content": "4. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: L13da: Visual interactive instruction tuning for omni-3d understanding, reasoning, and planning (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 308 + ], + "type": "text", + "content": "5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3070-3079. IEEE Computer Society, Los Alamitos, CA, USA (jun 2019). https://doi.org/10.1109/CVPR.2019.00319, https://doi.ieeecomputersociety.org/10.1109/CVPR.2019.00319" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "text", + "content": "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 375 + ], + "type": "text", + "content": "7. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7010-7019 (June 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "type": "text", + "content": "8. Gadelha, M., RoyChowdhury, A., Sharma, G., Kalogerakis, E., Cao, L., Learned-Miller, E., Wang, R., Maji, S.: Label-efficient learning on point clouds using approximate convex decompositions. In: European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 421, + 480, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 442 + ], + "type": "text", + "content": "9. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 444, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 444, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 135, + 444, + 480, + 485 + ], + "type": "text", + "content": "10. Graham, B., Engelcke, M., Maaten, L.v.d.: 3d semantic segmentation with submanifold sparse convolutional networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9224-9232 (2018). https://doi.org/10.1109/CVPR.2018.00961" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 487, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 498 + ], + "type": "text", + "content": "11. Graham, B., van der Maaten, L.: Submanifold sparse convolutional networks (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 499, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 499, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 134, + 499, + 480, + 531 + ], + "type": "text", + "content": "12. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., Heng, P.A.: Point-bind & point-lm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 533, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 480, + 575 + ], + "type": "text", + "content": "13. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15372-15383 (October 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "text", + "content": "14. Guo, Z., Tang, Y., Zhang, R., Wang, D., Wang, Z., Zhao, B., Li, X.: Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 480, + 632 + ], + "type": "text", + "content": "15. Hong, Y., Du, Y., Lin, C., Tenenbaum, J.B., Gan, C.: 3d concept grounding on neural fields. NeurIPS (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "16. Hong, Y., Lin, C., Du, Y., Chen, Z., Tenenbaum, J.B., Gan, C.: 3d concept learning and reasoning from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "17. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. NeurIPS (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 180 + ], + "type": "text", + "content": "18. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 481, + 214 + ], + "type": "text", + "content": "19. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2018). https://doi.org/10.1109/CVPR.2018.00278" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 278 + ], + "type": "text", + "content": "20. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: ReferItGame: Referring to objects in photographs of natural scenes. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 787-798. Association for Computational Linguistics, Doha, Qatar (Oct 2014). https://doi.org/10.3115/v1/D14-1086, https://aclanthology.org/D14-1086" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 279, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 481, + 311 + ], + "type": "text", + "content": "21. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 312, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 481, + 354 + ], + "type": "text", + "content": "22. Kundu, A., Yin, X., Fathi, A., Ross, D., Brewington, B., Funkhouser, T., Pantofaru, C.: Virtual multi-view fusion for 3d semantic segmentation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16. pp. 518-535. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 354, + 481, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 387 + ], + "type": "text", + "content": "23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8500-8509 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 387, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 481, + 408 + ], + "type": "text", + "content": "24. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 430 + ], + "type": "text", + "content": "25. Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 450 + ], + "type": "text", + "content": "26. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 472 + ], + "type": "text", + "content": "27. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "28. Lin, T.Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P., Ramanan, D., Zitnick, C.L., Dollar, P.: Microsoft coco: Common objects in context (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 421, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 421, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 421, + 514 + ], + "type": "text", + "content": "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 579 + ], + "type": "text", + "content": "30. Liu, K., Zhan, F., Zhang, J., Xu, M., Yu, Y., El Saddik, A., Theobalt, C., Xing, E., Lu, S.: Weakly supervised 3d open-vocabulary segmentation. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 53433-53456. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/a76b693f36916a5ed84d6e5b39a0dc03-Paper-Conference.pdf" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 579, + 481, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 623 + ], + "type": "text", + "content": "31. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21736-21746 (June 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 623, + 441, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 441, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 441, + 632 + ], + "type": "text", + "content": "32. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "33. Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Situated question answering in 3d scenes. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=IDJx97BC38" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "type": "text", + "content": "34. Mascaro, R., Teixeira, L., Chli, M.: Diffuser: Multi-view 2d-to-3d label diffusion for semantic scene segmentation. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 13589-13595 (2021). https://doi.org/10.1109/ICRA48506.2021.9561801" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "text", + "content": "35. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million narrated video clips (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 307, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 307, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 307, + 203 + ], + "type": "text", + "content": "36. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 204, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 481, + 246 + ], + "type": "text", + "content": "37. Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "type": "text", + "content": "38. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "text", + "content": "39. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "text", + "content": "40. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 353 + ], + "type": "text", + "content": "41. Robert, D., Vallet, B., Landrieu, L.: Learning multi-view aggregation in the wild for large-scale 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5575-5584 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 396 + ], + "type": "text", + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 472 + ], + "type": "text", + "content": "44. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "45. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 505, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 481, + 536 + ], + "type": "text", + "content": "46. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. Proceedings of the IEEE International Conference on Computer Vision (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "type": "text", + "content": "47. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "text", + "content": "48. Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., Dai, J.: Visionllm: Large language model is also an open-ended decoder for vision-centric tasks (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "49. Wang, Z., Huang, H., Zhao, Y., Zhang, Z., Zhao, Z.: Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 623, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 481, + 643 + ], + "type": "text", + "content": "50. Xu, M., Ding, R., Zhao, H., Qi, X.: Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "51. Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reasoning-based 3D Part Segmentation Using Large Multimodal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 259 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "52. Yang, J., Chen, X., Qian, S., Madaan, N., Iyengar, M., Fouhey, D.F., Chai, J.: Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "53. Zhang, S., Sun, P., Chen, S., Xiao, M., Shao, W., Zhang, W., Liu, Y., Chen, K., Luo, P.: Gpt4roi: Instruction tuning large language model on region-of-interest (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "54. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "55. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "56. Ziyu, Z., Xiaojian, M., Yixin, C., Zhidong, D., Siyuan, H., Qing, L.: 3d-vista: Pre-trained transformer for 3d vision and text alignment. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "A. Kareem et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file